1 /* 2 * Copyright (c) 2001, 2017, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "classfile/classLoaderData.hpp" 27 #include "classfile/stringTable.hpp" 28 #include "classfile/symbolTable.hpp" 29 #include "classfile/systemDictionary.hpp" 30 #include "code/codeCache.hpp" 31 #include "gc/cms/cmsCollectorPolicy.hpp" 32 #include "gc/cms/cmsOopClosures.inline.hpp" 33 #include "gc/cms/compactibleFreeListSpace.hpp" 34 #include "gc/cms/concurrentMarkSweepGeneration.inline.hpp" 35 #include "gc/cms/concurrentMarkSweepThread.hpp" 36 #include "gc/cms/parNewGeneration.hpp" 37 #include "gc/cms/vmCMSOperations.hpp" 38 #include "gc/serial/genMarkSweep.hpp" 39 #include "gc/serial/tenuredGeneration.hpp" 40 #include "gc/shared/adaptiveSizePolicy.hpp" 41 #include "gc/shared/cardGeneration.inline.hpp" 42 #include "gc/shared/cardTableRS.hpp" 43 #include "gc/shared/collectedHeap.inline.hpp" 44 #include "gc/shared/collectorCounters.hpp" 45 #include "gc/shared/collectorPolicy.hpp" 46 #include "gc/shared/gcLocker.inline.hpp" 47 #include "gc/shared/gcPolicyCounters.hpp" 48 #include "gc/shared/gcTimer.hpp" 49 #include "gc/shared/gcTrace.hpp" 50 #include "gc/shared/gcTraceTime.inline.hpp" 51 #include "gc/shared/genCollectedHeap.hpp" 52 #include "gc/shared/genOopClosures.inline.hpp" 53 #include "gc/shared/isGCActiveMark.hpp" 54 #include "gc/shared/referencePolicy.hpp" 55 #include "gc/shared/strongRootsScope.hpp" 56 #include "gc/shared/taskqueue.inline.hpp" 57 #include "logging/log.hpp" 58 #include "memory/allocation.hpp" 59 #include "memory/iterator.inline.hpp" 60 #include "memory/padded.hpp" 61 #include "memory/resourceArea.hpp" 62 #include "oops/oop.inline.hpp" 63 #include "prims/jvmtiExport.hpp" 64 #include "runtime/atomic.hpp" 65 #include "runtime/globals_extension.hpp" 66 #include "runtime/handles.inline.hpp" 67 #include "runtime/java.hpp" 68 #include "runtime/orderAccess.inline.hpp" 69 #include "runtime/timer.hpp" 70 #include "runtime/vmThread.hpp" 71 #include "services/memoryService.hpp" 72 #include "services/runtimeService.hpp" 73 #include "utilities/stack.inline.hpp" 74 75 // statics 76 CMSCollector* ConcurrentMarkSweepGeneration::_collector = NULL; 77 bool CMSCollector::_full_gc_requested = false; 78 GCCause::Cause CMSCollector::_full_gc_cause = GCCause::_no_gc; 79 80 ////////////////////////////////////////////////////////////////// 81 // In support of CMS/VM thread synchronization 82 ////////////////////////////////////////////////////////////////// 83 // We split use of the CGC_lock into 2 "levels". 84 // The low-level locking is of the usual CGC_lock monitor. We introduce 85 // a higher level "token" (hereafter "CMS token") built on top of the 86 // low level monitor (hereafter "CGC lock"). 87 // The token-passing protocol gives priority to the VM thread. The 88 // CMS-lock doesn't provide any fairness guarantees, but clients 89 // should ensure that it is only held for very short, bounded 90 // durations. 91 // 92 // When either of the CMS thread or the VM thread is involved in 93 // collection operations during which it does not want the other 94 // thread to interfere, it obtains the CMS token. 95 // 96 // If either thread tries to get the token while the other has 97 // it, that thread waits. However, if the VM thread and CMS thread 98 // both want the token, then the VM thread gets priority while the 99 // CMS thread waits. This ensures, for instance, that the "concurrent" 100 // phases of the CMS thread's work do not block out the VM thread 101 // for long periods of time as the CMS thread continues to hog 102 // the token. (See bug 4616232). 103 // 104 // The baton-passing functions are, however, controlled by the 105 // flags _foregroundGCShouldWait and _foregroundGCIsActive, 106 // and here the low-level CMS lock, not the high level token, 107 // ensures mutual exclusion. 108 // 109 // Two important conditions that we have to satisfy: 110 // 1. if a thread does a low-level wait on the CMS lock, then it 111 // relinquishes the CMS token if it were holding that token 112 // when it acquired the low-level CMS lock. 113 // 2. any low-level notifications on the low-level lock 114 // should only be sent when a thread has relinquished the token. 115 // 116 // In the absence of either property, we'd have potential deadlock. 117 // 118 // We protect each of the CMS (concurrent and sequential) phases 119 // with the CMS _token_, not the CMS _lock_. 120 // 121 // The only code protected by CMS lock is the token acquisition code 122 // itself, see ConcurrentMarkSweepThread::[de]synchronize(), and the 123 // baton-passing code. 124 // 125 // Unfortunately, i couldn't come up with a good abstraction to factor and 126 // hide the naked CGC_lock manipulation in the baton-passing code 127 // further below. That's something we should try to do. Also, the proof 128 // of correctness of this 2-level locking scheme is far from obvious, 129 // and potentially quite slippery. We have an uneasy suspicion, for instance, 130 // that there may be a theoretical possibility of delay/starvation in the 131 // low-level lock/wait/notify scheme used for the baton-passing because of 132 // potential interference with the priority scheme embodied in the 133 // CMS-token-passing protocol. See related comments at a CGC_lock->wait() 134 // invocation further below and marked with "XXX 20011219YSR". 135 // Indeed, as we note elsewhere, this may become yet more slippery 136 // in the presence of multiple CMS and/or multiple VM threads. XXX 137 138 class CMSTokenSync: public StackObj { 139 private: 140 bool _is_cms_thread; 141 public: 142 CMSTokenSync(bool is_cms_thread): 143 _is_cms_thread(is_cms_thread) { 144 assert(is_cms_thread == Thread::current()->is_ConcurrentGC_thread(), 145 "Incorrect argument to constructor"); 146 ConcurrentMarkSweepThread::synchronize(_is_cms_thread); 147 } 148 149 ~CMSTokenSync() { 150 assert(_is_cms_thread ? 151 ConcurrentMarkSweepThread::cms_thread_has_cms_token() : 152 ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 153 "Incorrect state"); 154 ConcurrentMarkSweepThread::desynchronize(_is_cms_thread); 155 } 156 }; 157 158 // Convenience class that does a CMSTokenSync, and then acquires 159 // upto three locks. 160 class CMSTokenSyncWithLocks: public CMSTokenSync { 161 private: 162 // Note: locks are acquired in textual declaration order 163 // and released in the opposite order 164 MutexLockerEx _locker1, _locker2, _locker3; 165 public: 166 CMSTokenSyncWithLocks(bool is_cms_thread, Mutex* mutex1, 167 Mutex* mutex2 = NULL, Mutex* mutex3 = NULL): 168 CMSTokenSync(is_cms_thread), 169 _locker1(mutex1, Mutex::_no_safepoint_check_flag), 170 _locker2(mutex2, Mutex::_no_safepoint_check_flag), 171 _locker3(mutex3, Mutex::_no_safepoint_check_flag) 172 { } 173 }; 174 175 176 ////////////////////////////////////////////////////////////////// 177 // Concurrent Mark-Sweep Generation ///////////////////////////// 178 ////////////////////////////////////////////////////////////////// 179 180 NOT_PRODUCT(CompactibleFreeListSpace* debug_cms_space;) 181 182 // This struct contains per-thread things necessary to support parallel 183 // young-gen collection. 184 class CMSParGCThreadState: public CHeapObj<mtGC> { 185 public: 186 CompactibleFreeListSpaceLAB lab; 187 PromotionInfo promo; 188 189 // Constructor. 190 CMSParGCThreadState(CompactibleFreeListSpace* cfls) : lab(cfls) { 191 promo.setSpace(cfls); 192 } 193 }; 194 195 ConcurrentMarkSweepGeneration::ConcurrentMarkSweepGeneration( 196 ReservedSpace rs, size_t initial_byte_size, CardTableRS* ct) : 197 CardGeneration(rs, initial_byte_size, ct), 198 _dilatation_factor(((double)MinChunkSize)/((double)(CollectedHeap::min_fill_size()))), 199 _did_compact(false) 200 { 201 HeapWord* bottom = (HeapWord*) _virtual_space.low(); 202 HeapWord* end = (HeapWord*) _virtual_space.high(); 203 204 _direct_allocated_words = 0; 205 NOT_PRODUCT( 206 _numObjectsPromoted = 0; 207 _numWordsPromoted = 0; 208 _numObjectsAllocated = 0; 209 _numWordsAllocated = 0; 210 ) 211 212 _cmsSpace = new CompactibleFreeListSpace(_bts, MemRegion(bottom, end)); 213 NOT_PRODUCT(debug_cms_space = _cmsSpace;) 214 _cmsSpace->_old_gen = this; 215 216 _gc_stats = new CMSGCStats(); 217 218 // Verify the assumption that FreeChunk::_prev and OopDesc::_klass 219 // offsets match. The ability to tell free chunks from objects 220 // depends on this property. 221 debug_only( 222 FreeChunk* junk = NULL; 223 assert(UseCompressedClassPointers || 224 junk->prev_addr() == (void*)(oop(junk)->klass_addr()), 225 "Offset of FreeChunk::_prev within FreeChunk must match" 226 " that of OopDesc::_klass within OopDesc"); 227 ) 228 229 _par_gc_thread_states = NEW_C_HEAP_ARRAY(CMSParGCThreadState*, ParallelGCThreads, mtGC); 230 for (uint i = 0; i < ParallelGCThreads; i++) { 231 _par_gc_thread_states[i] = new CMSParGCThreadState(cmsSpace()); 232 } 233 234 _incremental_collection_failed = false; 235 // The "dilatation_factor" is the expansion that can occur on 236 // account of the fact that the minimum object size in the CMS 237 // generation may be larger than that in, say, a contiguous young 238 // generation. 239 // Ideally, in the calculation below, we'd compute the dilatation 240 // factor as: MinChunkSize/(promoting_gen's min object size) 241 // Since we do not have such a general query interface for the 242 // promoting generation, we'll instead just use the minimum 243 // object size (which today is a header's worth of space); 244 // note that all arithmetic is in units of HeapWords. 245 assert(MinChunkSize >= CollectedHeap::min_fill_size(), "just checking"); 246 assert(_dilatation_factor >= 1.0, "from previous assert"); 247 } 248 249 250 // The field "_initiating_occupancy" represents the occupancy percentage 251 // at which we trigger a new collection cycle. Unless explicitly specified 252 // via CMSInitiatingOccupancyFraction (argument "io" below), it 253 // is calculated by: 254 // 255 // Let "f" be MinHeapFreeRatio in 256 // 257 // _initiating_occupancy = 100-f + 258 // f * (CMSTriggerRatio/100) 259 // where CMSTriggerRatio is the argument "tr" below. 260 // 261 // That is, if we assume the heap is at its desired maximum occupancy at the 262 // end of a collection, we let CMSTriggerRatio of the (purported) free 263 // space be allocated before initiating a new collection cycle. 264 // 265 void ConcurrentMarkSweepGeneration::init_initiating_occupancy(intx io, uintx tr) { 266 assert(io <= 100 && tr <= 100, "Check the arguments"); 267 if (io >= 0) { 268 _initiating_occupancy = (double)io / 100.0; 269 } else { 270 _initiating_occupancy = ((100 - MinHeapFreeRatio) + 271 (double)(tr * MinHeapFreeRatio) / 100.0) 272 / 100.0; 273 } 274 } 275 276 void ConcurrentMarkSweepGeneration::ref_processor_init() { 277 assert(collector() != NULL, "no collector"); 278 collector()->ref_processor_init(); 279 } 280 281 void CMSCollector::ref_processor_init() { 282 if (_ref_processor == NULL) { 283 // Allocate and initialize a reference processor 284 _ref_processor = 285 new ReferenceProcessor(_span, // span 286 (ParallelGCThreads > 1) && ParallelRefProcEnabled, // mt processing 287 ParallelGCThreads, // mt processing degree 288 _cmsGen->refs_discovery_is_mt(), // mt discovery 289 MAX2(ConcGCThreads, ParallelGCThreads), // mt discovery degree 290 _cmsGen->refs_discovery_is_atomic(), // discovery is not atomic 291 &_is_alive_closure); // closure for liveness info 292 // Initialize the _ref_processor field of CMSGen 293 _cmsGen->set_ref_processor(_ref_processor); 294 295 } 296 } 297 298 AdaptiveSizePolicy* CMSCollector::size_policy() { 299 GenCollectedHeap* gch = GenCollectedHeap::heap(); 300 return gch->gen_policy()->size_policy(); 301 } 302 303 void ConcurrentMarkSweepGeneration::initialize_performance_counters() { 304 305 const char* gen_name = "old"; 306 GenCollectorPolicy* gcp = GenCollectedHeap::heap()->gen_policy(); 307 // Generation Counters - generation 1, 1 subspace 308 _gen_counters = new GenerationCounters(gen_name, 1, 1, 309 gcp->min_old_size(), gcp->max_old_size(), &_virtual_space); 310 311 _space_counters = new GSpaceCounters(gen_name, 0, 312 _virtual_space.reserved_size(), 313 this, _gen_counters); 314 } 315 316 CMSStats::CMSStats(ConcurrentMarkSweepGeneration* cms_gen, unsigned int alpha): 317 _cms_gen(cms_gen) 318 { 319 assert(alpha <= 100, "bad value"); 320 _saved_alpha = alpha; 321 322 // Initialize the alphas to the bootstrap value of 100. 323 _gc0_alpha = _cms_alpha = 100; 324 325 _cms_begin_time.update(); 326 _cms_end_time.update(); 327 328 _gc0_duration = 0.0; 329 _gc0_period = 0.0; 330 _gc0_promoted = 0; 331 332 _cms_duration = 0.0; 333 _cms_period = 0.0; 334 _cms_allocated = 0; 335 336 _cms_used_at_gc0_begin = 0; 337 _cms_used_at_gc0_end = 0; 338 _allow_duty_cycle_reduction = false; 339 _valid_bits = 0; 340 } 341 342 double CMSStats::cms_free_adjustment_factor(size_t free) const { 343 // TBD: CR 6909490 344 return 1.0; 345 } 346 347 void CMSStats::adjust_cms_free_adjustment_factor(bool fail, size_t free) { 348 } 349 350 // If promotion failure handling is on use 351 // the padded average size of the promotion for each 352 // young generation collection. 353 double CMSStats::time_until_cms_gen_full() const { 354 size_t cms_free = _cms_gen->cmsSpace()->free(); 355 GenCollectedHeap* gch = GenCollectedHeap::heap(); 356 size_t expected_promotion = MIN2(gch->young_gen()->capacity(), 357 (size_t) _cms_gen->gc_stats()->avg_promoted()->padded_average()); 358 if (cms_free > expected_promotion) { 359 // Start a cms collection if there isn't enough space to promote 360 // for the next young collection. Use the padded average as 361 // a safety factor. 362 cms_free -= expected_promotion; 363 364 // Adjust by the safety factor. 365 double cms_free_dbl = (double)cms_free; 366 double cms_adjustment = (100.0 - CMSIncrementalSafetyFactor) / 100.0; 367 // Apply a further correction factor which tries to adjust 368 // for recent occurance of concurrent mode failures. 369 cms_adjustment = cms_adjustment * cms_free_adjustment_factor(cms_free); 370 cms_free_dbl = cms_free_dbl * cms_adjustment; 371 372 log_trace(gc)("CMSStats::time_until_cms_gen_full: cms_free " SIZE_FORMAT " expected_promotion " SIZE_FORMAT, 373 cms_free, expected_promotion); 374 log_trace(gc)(" cms_free_dbl %f cms_consumption_rate %f", cms_free_dbl, cms_consumption_rate() + 1.0); 375 // Add 1 in case the consumption rate goes to zero. 376 return cms_free_dbl / (cms_consumption_rate() + 1.0); 377 } 378 return 0.0; 379 } 380 381 // Compare the duration of the cms collection to the 382 // time remaining before the cms generation is empty. 383 // Note that the time from the start of the cms collection 384 // to the start of the cms sweep (less than the total 385 // duration of the cms collection) can be used. This 386 // has been tried and some applications experienced 387 // promotion failures early in execution. This was 388 // possibly because the averages were not accurate 389 // enough at the beginning. 390 double CMSStats::time_until_cms_start() const { 391 // We add "gc0_period" to the "work" calculation 392 // below because this query is done (mostly) at the 393 // end of a scavenge, so we need to conservatively 394 // account for that much possible delay 395 // in the query so as to avoid concurrent mode failures 396 // due to starting the collection just a wee bit too 397 // late. 398 double work = cms_duration() + gc0_period(); 399 double deadline = time_until_cms_gen_full(); 400 // If a concurrent mode failure occurred recently, we want to be 401 // more conservative and halve our expected time_until_cms_gen_full() 402 if (work > deadline) { 403 log_develop_trace(gc)("CMSCollector: collect because of anticipated promotion before full %3.7f + %3.7f > %3.7f ", 404 cms_duration(), gc0_period(), time_until_cms_gen_full()); 405 return 0.0; 406 } 407 return work - deadline; 408 } 409 410 #ifndef PRODUCT 411 void CMSStats::print_on(outputStream *st) const { 412 st->print(" gc0_alpha=%d,cms_alpha=%d", _gc0_alpha, _cms_alpha); 413 st->print(",gc0_dur=%g,gc0_per=%g,gc0_promo=" SIZE_FORMAT, 414 gc0_duration(), gc0_period(), gc0_promoted()); 415 st->print(",cms_dur=%g,cms_per=%g,cms_alloc=" SIZE_FORMAT, 416 cms_duration(), cms_period(), cms_allocated()); 417 st->print(",cms_since_beg=%g,cms_since_end=%g", 418 cms_time_since_begin(), cms_time_since_end()); 419 st->print(",cms_used_beg=" SIZE_FORMAT ",cms_used_end=" SIZE_FORMAT, 420 _cms_used_at_gc0_begin, _cms_used_at_gc0_end); 421 422 if (valid()) { 423 st->print(",promo_rate=%g,cms_alloc_rate=%g", 424 promotion_rate(), cms_allocation_rate()); 425 st->print(",cms_consumption_rate=%g,time_until_full=%g", 426 cms_consumption_rate(), time_until_cms_gen_full()); 427 } 428 st->cr(); 429 } 430 #endif // #ifndef PRODUCT 431 432 CMSCollector::CollectorState CMSCollector::_collectorState = 433 CMSCollector::Idling; 434 bool CMSCollector::_foregroundGCIsActive = false; 435 bool CMSCollector::_foregroundGCShouldWait = false; 436 437 CMSCollector::CMSCollector(ConcurrentMarkSweepGeneration* cmsGen, 438 CardTableRS* ct, 439 ConcurrentMarkSweepPolicy* cp): 440 _cmsGen(cmsGen), 441 _ct(ct), 442 _ref_processor(NULL), // will be set later 443 _conc_workers(NULL), // may be set later 444 _abort_preclean(false), 445 _start_sampling(false), 446 _between_prologue_and_epilogue(false), 447 _markBitMap(0, Mutex::leaf + 1, "CMS_markBitMap_lock"), 448 _modUnionTable((CardTable::card_shift - LogHeapWordSize), 449 -1 /* lock-free */, "No_lock" /* dummy */), 450 _modUnionClosurePar(&_modUnionTable), 451 // Adjust my span to cover old (cms) gen 452 _span(cmsGen->reserved()), 453 // Construct the is_alive_closure with _span & markBitMap 454 _is_alive_closure(_span, &_markBitMap), 455 _restart_addr(NULL), 456 _overflow_list(NULL), 457 _stats(cmsGen), 458 _eden_chunk_lock(new Mutex(Mutex::leaf + 1, "CMS_eden_chunk_lock", true, 459 //verify that this lock should be acquired with safepoint check. 460 Monitor::_safepoint_check_sometimes)), 461 _eden_chunk_array(NULL), // may be set in ctor body 462 _eden_chunk_capacity(0), // -- ditto -- 463 _eden_chunk_index(0), // -- ditto -- 464 _survivor_plab_array(NULL), // -- ditto -- 465 _survivor_chunk_array(NULL), // -- ditto -- 466 _survivor_chunk_capacity(0), // -- ditto -- 467 _survivor_chunk_index(0), // -- ditto -- 468 _ser_pmc_preclean_ovflw(0), 469 _ser_kac_preclean_ovflw(0), 470 _ser_pmc_remark_ovflw(0), 471 _par_pmc_remark_ovflw(0), 472 _ser_kac_ovflw(0), 473 _par_kac_ovflw(0), 474 #ifndef PRODUCT 475 _num_par_pushes(0), 476 #endif 477 _collection_count_start(0), 478 _verifying(false), 479 _verification_mark_bm(0, Mutex::leaf + 1, "CMS_verification_mark_bm_lock"), 480 _completed_initialization(false), 481 _collector_policy(cp), 482 _should_unload_classes(CMSClassUnloadingEnabled), 483 _concurrent_cycles_since_last_unload(0), 484 _roots_scanning_options(GenCollectedHeap::SO_None), 485 _inter_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding), 486 _intra_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding), 487 _gc_tracer_cm(new (ResourceObj::C_HEAP, mtGC) CMSTracer()), 488 _gc_timer_cm(new (ResourceObj::C_HEAP, mtGC) ConcurrentGCTimer()), 489 _cms_start_registered(false) 490 { 491 // Now expand the span and allocate the collection support structures 492 // (MUT, marking bit map etc.) to cover both generations subject to 493 // collection. 494 495 // For use by dirty card to oop closures. 496 _cmsGen->cmsSpace()->set_collector(this); 497 498 // Allocate MUT and marking bit map 499 { 500 MutexLockerEx x(_markBitMap.lock(), Mutex::_no_safepoint_check_flag); 501 if (!_markBitMap.allocate(_span)) { 502 log_warning(gc)("Failed to allocate CMS Bit Map"); 503 return; 504 } 505 assert(_markBitMap.covers(_span), "_markBitMap inconsistency?"); 506 } 507 { 508 _modUnionTable.allocate(_span); 509 assert(_modUnionTable.covers(_span), "_modUnionTable inconsistency?"); 510 } 511 512 if (!_markStack.allocate(MarkStackSize)) { 513 log_warning(gc)("Failed to allocate CMS Marking Stack"); 514 return; 515 } 516 517 // Support for multi-threaded concurrent phases 518 if (CMSConcurrentMTEnabled) { 519 if (FLAG_IS_DEFAULT(ConcGCThreads)) { 520 // just for now 521 FLAG_SET_DEFAULT(ConcGCThreads, (ParallelGCThreads + 3) / 4); 522 } 523 if (ConcGCThreads > 1) { 524 _conc_workers = new YieldingFlexibleWorkGang("CMS Thread", 525 ConcGCThreads, true); 526 if (_conc_workers == NULL) { 527 log_warning(gc)("GC/CMS: _conc_workers allocation failure: forcing -CMSConcurrentMTEnabled"); 528 CMSConcurrentMTEnabled = false; 529 } else { 530 _conc_workers->initialize_workers(); 531 } 532 } else { 533 CMSConcurrentMTEnabled = false; 534 } 535 } 536 if (!CMSConcurrentMTEnabled) { 537 ConcGCThreads = 0; 538 } else { 539 // Turn off CMSCleanOnEnter optimization temporarily for 540 // the MT case where it's not fixed yet; see 6178663. 541 CMSCleanOnEnter = false; 542 } 543 assert((_conc_workers != NULL) == (ConcGCThreads > 1), 544 "Inconsistency"); 545 log_debug(gc)("ConcGCThreads: %u", ConcGCThreads); 546 log_debug(gc)("ParallelGCThreads: %u", ParallelGCThreads); 547 548 // Parallel task queues; these are shared for the 549 // concurrent and stop-world phases of CMS, but 550 // are not shared with parallel scavenge (ParNew). 551 { 552 uint i; 553 uint num_queues = MAX2(ParallelGCThreads, ConcGCThreads); 554 555 if ((CMSParallelRemarkEnabled || CMSConcurrentMTEnabled 556 || ParallelRefProcEnabled) 557 && num_queues > 0) { 558 _task_queues = new OopTaskQueueSet(num_queues); 559 if (_task_queues == NULL) { 560 log_warning(gc)("task_queues allocation failure."); 561 return; 562 } 563 _hash_seed = NEW_C_HEAP_ARRAY(int, num_queues, mtGC); 564 typedef Padded<OopTaskQueue> PaddedOopTaskQueue; 565 for (i = 0; i < num_queues; i++) { 566 PaddedOopTaskQueue *q = new PaddedOopTaskQueue(); 567 if (q == NULL) { 568 log_warning(gc)("work_queue allocation failure."); 569 return; 570 } 571 _task_queues->register_queue(i, q); 572 } 573 for (i = 0; i < num_queues; i++) { 574 _task_queues->queue(i)->initialize(); 575 _hash_seed[i] = 17; // copied from ParNew 576 } 577 } 578 } 579 580 _cmsGen ->init_initiating_occupancy(CMSInitiatingOccupancyFraction, CMSTriggerRatio); 581 582 // Clip CMSBootstrapOccupancy between 0 and 100. 583 _bootstrap_occupancy = CMSBootstrapOccupancy / 100.0; 584 585 // Now tell CMS generations the identity of their collector 586 ConcurrentMarkSweepGeneration::set_collector(this); 587 588 // Create & start a CMS thread for this CMS collector 589 _cmsThread = ConcurrentMarkSweepThread::start(this); 590 assert(cmsThread() != NULL, "CMS Thread should have been created"); 591 assert(cmsThread()->collector() == this, 592 "CMS Thread should refer to this gen"); 593 assert(CGC_lock != NULL, "Where's the CGC_lock?"); 594 595 // Support for parallelizing young gen rescan 596 GenCollectedHeap* gch = GenCollectedHeap::heap(); 597 assert(gch->young_gen()->kind() == Generation::ParNew, "CMS can only be used with ParNew"); 598 _young_gen = (ParNewGeneration*)gch->young_gen(); 599 if (gch->supports_inline_contig_alloc()) { 600 _top_addr = gch->top_addr(); 601 _end_addr = gch->end_addr(); 602 assert(_young_gen != NULL, "no _young_gen"); 603 _eden_chunk_index = 0; 604 _eden_chunk_capacity = (_young_gen->max_capacity() + CMSSamplingGrain) / CMSSamplingGrain; 605 _eden_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, _eden_chunk_capacity, mtGC); 606 } 607 608 // Support for parallelizing survivor space rescan 609 if ((CMSParallelRemarkEnabled && CMSParallelSurvivorRemarkEnabled) || CMSParallelInitialMarkEnabled) { 610 const size_t max_plab_samples = 611 _young_gen->max_survivor_size() / (PLAB::min_size() * HeapWordSize); 612 613 _survivor_plab_array = NEW_C_HEAP_ARRAY(ChunkArray, ParallelGCThreads, mtGC); 614 _survivor_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples, mtGC); 615 _cursor = NEW_C_HEAP_ARRAY(size_t, ParallelGCThreads, mtGC); 616 _survivor_chunk_capacity = max_plab_samples; 617 for (uint i = 0; i < ParallelGCThreads; i++) { 618 HeapWord** vec = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples, mtGC); 619 ChunkArray* cur = ::new (&_survivor_plab_array[i]) ChunkArray(vec, max_plab_samples); 620 assert(cur->end() == 0, "Should be 0"); 621 assert(cur->array() == vec, "Should be vec"); 622 assert(cur->capacity() == max_plab_samples, "Error"); 623 } 624 } 625 626 NOT_PRODUCT(_overflow_counter = CMSMarkStackOverflowInterval;) 627 _gc_counters = new CollectorCounters("CMS", 1); 628 _completed_initialization = true; 629 _inter_sweep_timer.start(); // start of time 630 } 631 632 const char* ConcurrentMarkSweepGeneration::name() const { 633 return "concurrent mark-sweep generation"; 634 } 635 void ConcurrentMarkSweepGeneration::update_counters() { 636 if (UsePerfData) { 637 _space_counters->update_all(); 638 _gen_counters->update_all(); 639 } 640 } 641 642 // this is an optimized version of update_counters(). it takes the 643 // used value as a parameter rather than computing it. 644 // 645 void ConcurrentMarkSweepGeneration::update_counters(size_t used) { 646 if (UsePerfData) { 647 _space_counters->update_used(used); 648 _space_counters->update_capacity(); 649 _gen_counters->update_all(); 650 } 651 } 652 653 void ConcurrentMarkSweepGeneration::print() const { 654 Generation::print(); 655 cmsSpace()->print(); 656 } 657 658 #ifndef PRODUCT 659 void ConcurrentMarkSweepGeneration::print_statistics() { 660 cmsSpace()->printFLCensus(0); 661 } 662 #endif 663 664 size_t 665 ConcurrentMarkSweepGeneration::contiguous_available() const { 666 // dld proposes an improvement in precision here. If the committed 667 // part of the space ends in a free block we should add that to 668 // uncommitted size in the calculation below. Will make this 669 // change later, staying with the approximation below for the 670 // time being. -- ysr. 671 return MAX2(_virtual_space.uncommitted_size(), unsafe_max_alloc_nogc()); 672 } 673 674 size_t 675 ConcurrentMarkSweepGeneration::unsafe_max_alloc_nogc() const { 676 return _cmsSpace->max_alloc_in_words() * HeapWordSize; 677 } 678 679 size_t ConcurrentMarkSweepGeneration::max_available() const { 680 return free() + _virtual_space.uncommitted_size(); 681 } 682 683 bool ConcurrentMarkSweepGeneration::promotion_attempt_is_safe(size_t max_promotion_in_bytes) const { 684 size_t available = max_available(); 685 size_t av_promo = (size_t)gc_stats()->avg_promoted()->padded_average(); 686 bool res = (available >= av_promo) || (available >= max_promotion_in_bytes); 687 log_trace(gc, promotion)("CMS: promo attempt is%s safe: available(" SIZE_FORMAT ") %s av_promo(" SIZE_FORMAT "), max_promo(" SIZE_FORMAT ")", 688 res? "":" not", available, res? ">=":"<", av_promo, max_promotion_in_bytes); 689 return res; 690 } 691 692 // At a promotion failure dump information on block layout in heap 693 // (cms old generation). 694 void ConcurrentMarkSweepGeneration::promotion_failure_occurred() { 695 Log(gc, promotion) log; 696 if (log.is_trace()) { 697 ResourceMark rm; 698 cmsSpace()->dump_at_safepoint_with_locks(collector(), log.trace_stream()); 699 } 700 } 701 702 void ConcurrentMarkSweepGeneration::reset_after_compaction() { 703 // Clear the promotion information. These pointers can be adjusted 704 // along with all the other pointers into the heap but 705 // compaction is expected to be a rare event with 706 // a heap using cms so don't do it without seeing the need. 707 for (uint i = 0; i < ParallelGCThreads; i++) { 708 _par_gc_thread_states[i]->promo.reset(); 709 } 710 } 711 712 void ConcurrentMarkSweepGeneration::compute_new_size() { 713 assert_locked_or_safepoint(Heap_lock); 714 715 // If incremental collection failed, we just want to expand 716 // to the limit. 717 if (incremental_collection_failed()) { 718 clear_incremental_collection_failed(); 719 grow_to_reserved(); 720 return; 721 } 722 723 // The heap has been compacted but not reset yet. 724 // Any metric such as free() or used() will be incorrect. 725 726 CardGeneration::compute_new_size(); 727 728 // Reset again after a possible resizing 729 if (did_compact()) { 730 cmsSpace()->reset_after_compaction(); 731 } 732 } 733 734 void ConcurrentMarkSweepGeneration::compute_new_size_free_list() { 735 assert_locked_or_safepoint(Heap_lock); 736 737 // If incremental collection failed, we just want to expand 738 // to the limit. 739 if (incremental_collection_failed()) { 740 clear_incremental_collection_failed(); 741 grow_to_reserved(); 742 return; 743 } 744 745 double free_percentage = ((double) free()) / capacity(); 746 double desired_free_percentage = (double) MinHeapFreeRatio / 100; 747 double maximum_free_percentage = (double) MaxHeapFreeRatio / 100; 748 749 // compute expansion delta needed for reaching desired free percentage 750 if (free_percentage < desired_free_percentage) { 751 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 752 assert(desired_capacity >= capacity(), "invalid expansion size"); 753 size_t expand_bytes = MAX2(desired_capacity - capacity(), MinHeapDeltaBytes); 754 Log(gc) log; 755 if (log.is_trace()) { 756 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 757 log.trace("From compute_new_size: "); 758 log.trace(" Free fraction %f", free_percentage); 759 log.trace(" Desired free fraction %f", desired_free_percentage); 760 log.trace(" Maximum free fraction %f", maximum_free_percentage); 761 log.trace(" Capacity " SIZE_FORMAT, capacity() / 1000); 762 log.trace(" Desired capacity " SIZE_FORMAT, desired_capacity / 1000); 763 GenCollectedHeap* gch = GenCollectedHeap::heap(); 764 assert(gch->is_old_gen(this), "The CMS generation should always be the old generation"); 765 size_t young_size = gch->young_gen()->capacity(); 766 log.trace(" Young gen size " SIZE_FORMAT, young_size / 1000); 767 log.trace(" unsafe_max_alloc_nogc " SIZE_FORMAT, unsafe_max_alloc_nogc() / 1000); 768 log.trace(" contiguous available " SIZE_FORMAT, contiguous_available() / 1000); 769 log.trace(" Expand by " SIZE_FORMAT " (bytes)", expand_bytes); 770 } 771 // safe if expansion fails 772 expand_for_gc_cause(expand_bytes, 0, CMSExpansionCause::_satisfy_free_ratio); 773 log.trace(" Expanded free fraction %f", ((double) free()) / capacity()); 774 } else { 775 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 776 assert(desired_capacity <= capacity(), "invalid expansion size"); 777 size_t shrink_bytes = capacity() - desired_capacity; 778 // Don't shrink unless the delta is greater than the minimum shrink we want 779 if (shrink_bytes >= MinHeapDeltaBytes) { 780 shrink_free_list_by(shrink_bytes); 781 } 782 } 783 } 784 785 Mutex* ConcurrentMarkSweepGeneration::freelistLock() const { 786 return cmsSpace()->freelistLock(); 787 } 788 789 HeapWord* ConcurrentMarkSweepGeneration::allocate(size_t size, bool tlab) { 790 CMSSynchronousYieldRequest yr; 791 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 792 return have_lock_and_allocate(size, tlab); 793 } 794 795 HeapWord* ConcurrentMarkSweepGeneration::have_lock_and_allocate(size_t size, 796 bool tlab /* ignored */) { 797 assert_lock_strong(freelistLock()); 798 size_t adjustedSize = CompactibleFreeListSpace::adjustObjectSize(size); 799 HeapWord* res = cmsSpace()->allocate(adjustedSize); 800 // Allocate the object live (grey) if the background collector has 801 // started marking. This is necessary because the marker may 802 // have passed this address and consequently this object will 803 // not otherwise be greyed and would be incorrectly swept up. 804 // Note that if this object contains references, the writing 805 // of those references will dirty the card containing this object 806 // allowing the object to be blackened (and its references scanned) 807 // either during a preclean phase or at the final checkpoint. 808 if (res != NULL) { 809 // We may block here with an uninitialized object with 810 // its mark-bit or P-bits not yet set. Such objects need 811 // to be safely navigable by block_start(). 812 assert(oop(res)->klass_or_null() == NULL, "Object should be uninitialized here."); 813 assert(!((FreeChunk*)res)->is_free(), "Error, block will look free but show wrong size"); 814 collector()->direct_allocated(res, adjustedSize); 815 _direct_allocated_words += adjustedSize; 816 // allocation counters 817 NOT_PRODUCT( 818 _numObjectsAllocated++; 819 _numWordsAllocated += (int)adjustedSize; 820 ) 821 } 822 return res; 823 } 824 825 // In the case of direct allocation by mutators in a generation that 826 // is being concurrently collected, the object must be allocated 827 // live (grey) if the background collector has started marking. 828 // This is necessary because the marker may 829 // have passed this address and consequently this object will 830 // not otherwise be greyed and would be incorrectly swept up. 831 // Note that if this object contains references, the writing 832 // of those references will dirty the card containing this object 833 // allowing the object to be blackened (and its references scanned) 834 // either during a preclean phase or at the final checkpoint. 835 void CMSCollector::direct_allocated(HeapWord* start, size_t size) { 836 assert(_markBitMap.covers(start, size), "Out of bounds"); 837 if (_collectorState >= Marking) { 838 MutexLockerEx y(_markBitMap.lock(), 839 Mutex::_no_safepoint_check_flag); 840 // [see comments preceding SweepClosure::do_blk() below for details] 841 // 842 // Can the P-bits be deleted now? JJJ 843 // 844 // 1. need to mark the object as live so it isn't collected 845 // 2. need to mark the 2nd bit to indicate the object may be uninitialized 846 // 3. need to mark the end of the object so marking, precleaning or sweeping 847 // can skip over uninitialized or unparsable objects. An allocated 848 // object is considered uninitialized for our purposes as long as 849 // its klass word is NULL. All old gen objects are parsable 850 // as soon as they are initialized.) 851 _markBitMap.mark(start); // object is live 852 _markBitMap.mark(start + 1); // object is potentially uninitialized? 853 _markBitMap.mark(start + size - 1); 854 // mark end of object 855 } 856 // check that oop looks uninitialized 857 assert(oop(start)->klass_or_null() == NULL, "_klass should be NULL"); 858 } 859 860 void CMSCollector::promoted(bool par, HeapWord* start, 861 bool is_obj_array, size_t obj_size) { 862 assert(_markBitMap.covers(start), "Out of bounds"); 863 // See comment in direct_allocated() about when objects should 864 // be allocated live. 865 if (_collectorState >= Marking) { 866 // we already hold the marking bit map lock, taken in 867 // the prologue 868 if (par) { 869 _markBitMap.par_mark(start); 870 } else { 871 _markBitMap.mark(start); 872 } 873 // We don't need to mark the object as uninitialized (as 874 // in direct_allocated above) because this is being done with the 875 // world stopped and the object will be initialized by the 876 // time the marking, precleaning or sweeping get to look at it. 877 // But see the code for copying objects into the CMS generation, 878 // where we need to ensure that concurrent readers of the 879 // block offset table are able to safely navigate a block that 880 // is in flux from being free to being allocated (and in 881 // transition while being copied into) and subsequently 882 // becoming a bona-fide object when the copy/promotion is complete. 883 assert(SafepointSynchronize::is_at_safepoint(), 884 "expect promotion only at safepoints"); 885 886 if (_collectorState < Sweeping) { 887 // Mark the appropriate cards in the modUnionTable, so that 888 // this object gets scanned before the sweep. If this is 889 // not done, CMS generation references in the object might 890 // not get marked. 891 // For the case of arrays, which are otherwise precisely 892 // marked, we need to dirty the entire array, not just its head. 893 if (is_obj_array) { 894 // The [par_]mark_range() method expects mr.end() below to 895 // be aligned to the granularity of a bit's representation 896 // in the heap. In the case of the MUT below, that's a 897 // card size. 898 MemRegion mr(start, 899 (HeapWord*)round_to((intptr_t)(start + obj_size), 900 CardTable::card_size /* bytes */)); 901 if (par) { 902 _modUnionTable.par_mark_range(mr); 903 } else { 904 _modUnionTable.mark_range(mr); 905 } 906 } else { // not an obj array; we can just mark the head 907 if (par) { 908 _modUnionTable.par_mark(start); 909 } else { 910 _modUnionTable.mark(start); 911 } 912 } 913 } 914 } 915 } 916 917 oop ConcurrentMarkSweepGeneration::promote(oop obj, size_t obj_size) { 918 assert(obj_size == (size_t)obj->size(), "bad obj_size passed in"); 919 // allocate, copy and if necessary update promoinfo -- 920 // delegate to underlying space. 921 assert_lock_strong(freelistLock()); 922 923 #ifndef PRODUCT 924 if (GenCollectedHeap::heap()->promotion_should_fail()) { 925 return NULL; 926 } 927 #endif // #ifndef PRODUCT 928 929 oop res = _cmsSpace->promote(obj, obj_size); 930 if (res == NULL) { 931 // expand and retry 932 size_t s = _cmsSpace->expansionSpaceRequired(obj_size); // HeapWords 933 expand_for_gc_cause(s*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_satisfy_promotion); 934 // Since this is the old generation, we don't try to promote 935 // into a more senior generation. 936 res = _cmsSpace->promote(obj, obj_size); 937 } 938 if (res != NULL) { 939 // See comment in allocate() about when objects should 940 // be allocated live. 941 assert(obj->is_oop(), "Will dereference klass pointer below"); 942 collector()->promoted(false, // Not parallel 943 (HeapWord*)res, obj->is_objArray(), obj_size); 944 // promotion counters 945 NOT_PRODUCT( 946 _numObjectsPromoted++; 947 _numWordsPromoted += 948 (int)(CompactibleFreeListSpace::adjustObjectSize(obj->size())); 949 ) 950 } 951 return res; 952 } 953 954 955 // IMPORTANT: Notes on object size recognition in CMS. 956 // --------------------------------------------------- 957 // A block of storage in the CMS generation is always in 958 // one of three states. A free block (FREE), an allocated 959 // object (OBJECT) whose size() method reports the correct size, 960 // and an intermediate state (TRANSIENT) in which its size cannot 961 // be accurately determined. 962 // STATE IDENTIFICATION: (32 bit and 64 bit w/o COOPS) 963 // ----------------------------------------------------- 964 // FREE: klass_word & 1 == 1; mark_word holds block size 965 // 966 // OBJECT: klass_word installed; klass_word != 0 && klass_word & 1 == 0; 967 // obj->size() computes correct size 968 // 969 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT 970 // 971 // STATE IDENTIFICATION: (64 bit+COOPS) 972 // ------------------------------------ 973 // FREE: mark_word & CMS_FREE_BIT == 1; mark_word & ~CMS_FREE_BIT gives block_size 974 // 975 // OBJECT: klass_word installed; klass_word != 0; 976 // obj->size() computes correct size 977 // 978 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT 979 // 980 // 981 // STATE TRANSITION DIAGRAM 982 // 983 // mut / parnew mut / parnew 984 // FREE --------------------> TRANSIENT ---------------------> OBJECT --| 985 // ^ | 986 // |------------------------ DEAD <------------------------------------| 987 // sweep mut 988 // 989 // While a block is in TRANSIENT state its size cannot be determined 990 // so readers will either need to come back later or stall until 991 // the size can be determined. Note that for the case of direct 992 // allocation, P-bits, when available, may be used to determine the 993 // size of an object that may not yet have been initialized. 994 995 // Things to support parallel young-gen collection. 996 oop 997 ConcurrentMarkSweepGeneration::par_promote(int thread_num, 998 oop old, markOop m, 999 size_t word_sz) { 1000 #ifndef PRODUCT 1001 if (GenCollectedHeap::heap()->promotion_should_fail()) { 1002 return NULL; 1003 } 1004 #endif // #ifndef PRODUCT 1005 1006 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1007 PromotionInfo* promoInfo = &ps->promo; 1008 // if we are tracking promotions, then first ensure space for 1009 // promotion (including spooling space for saving header if necessary). 1010 // then allocate and copy, then track promoted info if needed. 1011 // When tracking (see PromotionInfo::track()), the mark word may 1012 // be displaced and in this case restoration of the mark word 1013 // occurs in the (oop_since_save_marks_)iterate phase. 1014 if (promoInfo->tracking() && !promoInfo->ensure_spooling_space()) { 1015 // Out of space for allocating spooling buffers; 1016 // try expanding and allocating spooling buffers. 1017 if (!expand_and_ensure_spooling_space(promoInfo)) { 1018 return NULL; 1019 } 1020 } 1021 assert(!promoInfo->tracking() || promoInfo->has_spooling_space(), "Control point invariant"); 1022 const size_t alloc_sz = CompactibleFreeListSpace::adjustObjectSize(word_sz); 1023 HeapWord* obj_ptr = ps->lab.alloc(alloc_sz); 1024 if (obj_ptr == NULL) { 1025 obj_ptr = expand_and_par_lab_allocate(ps, alloc_sz); 1026 if (obj_ptr == NULL) { 1027 return NULL; 1028 } 1029 } 1030 oop obj = oop(obj_ptr); 1031 OrderAccess::storestore(); 1032 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1033 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1034 // IMPORTANT: See note on object initialization for CMS above. 1035 // Otherwise, copy the object. Here we must be careful to insert the 1036 // klass pointer last, since this marks the block as an allocated object. 1037 // Except with compressed oops it's the mark word. 1038 HeapWord* old_ptr = (HeapWord*)old; 1039 // Restore the mark word copied above. 1040 obj->set_mark(m); 1041 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1042 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1043 OrderAccess::storestore(); 1044 1045 if (UseCompressedClassPointers) { 1046 // Copy gap missed by (aligned) header size calculation below 1047 obj->set_klass_gap(old->klass_gap()); 1048 } 1049 if (word_sz > (size_t)oopDesc::header_size()) { 1050 Copy::aligned_disjoint_words(old_ptr + oopDesc::header_size(), 1051 obj_ptr + oopDesc::header_size(), 1052 word_sz - oopDesc::header_size()); 1053 } 1054 1055 // Now we can track the promoted object, if necessary. We take care 1056 // to delay the transition from uninitialized to full object 1057 // (i.e., insertion of klass pointer) until after, so that it 1058 // atomically becomes a promoted object. 1059 if (promoInfo->tracking()) { 1060 promoInfo->track((PromotedObject*)obj, old->klass()); 1061 } 1062 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1063 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1064 assert(old->is_oop(), "Will use and dereference old klass ptr below"); 1065 1066 // Finally, install the klass pointer (this should be volatile). 1067 OrderAccess::storestore(); 1068 obj->set_klass(old->klass()); 1069 // We should now be able to calculate the right size for this object 1070 assert(obj->is_oop() && obj->size() == (int)word_sz, "Error, incorrect size computed for promoted object"); 1071 1072 collector()->promoted(true, // parallel 1073 obj_ptr, old->is_objArray(), word_sz); 1074 1075 NOT_PRODUCT( 1076 Atomic::inc_ptr(&_numObjectsPromoted); 1077 Atomic::add_ptr(alloc_sz, &_numWordsPromoted); 1078 ) 1079 1080 return obj; 1081 } 1082 1083 void 1084 ConcurrentMarkSweepGeneration:: 1085 par_promote_alloc_done(int thread_num) { 1086 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1087 ps->lab.retire(thread_num); 1088 } 1089 1090 void 1091 ConcurrentMarkSweepGeneration:: 1092 par_oop_since_save_marks_iterate_done(int thread_num) { 1093 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1094 ParScanWithoutBarrierClosure* dummy_cl = NULL; 1095 ps->promo.promoted_oops_iterate_nv(dummy_cl); 1096 1097 // Because card-scanning has been completed, subsequent phases 1098 // (e.g., reference processing) will not need to recognize which 1099 // objects have been promoted during this GC. So, we can now disable 1100 // promotion tracking. 1101 ps->promo.stopTrackingPromotions(); 1102 } 1103 1104 bool ConcurrentMarkSweepGeneration::should_collect(bool full, 1105 size_t size, 1106 bool tlab) 1107 { 1108 // We allow a STW collection only if a full 1109 // collection was requested. 1110 return full || should_allocate(size, tlab); // FIX ME !!! 1111 // This and promotion failure handling are connected at the 1112 // hip and should be fixed by untying them. 1113 } 1114 1115 bool CMSCollector::shouldConcurrentCollect() { 1116 LogTarget(Trace, gc) log; 1117 1118 if (_full_gc_requested) { 1119 log.print("CMSCollector: collect because of explicit gc request (or GCLocker)"); 1120 return true; 1121 } 1122 1123 FreelistLocker x(this); 1124 // ------------------------------------------------------------------ 1125 // Print out lots of information which affects the initiation of 1126 // a collection. 1127 if (log.is_enabled() && stats().valid()) { 1128 log.print("CMSCollector shouldConcurrentCollect: "); 1129 1130 LogStream out(log); 1131 stats().print_on(&out); 1132 1133 log.print("time_until_cms_gen_full %3.7f", stats().time_until_cms_gen_full()); 1134 log.print("free=" SIZE_FORMAT, _cmsGen->free()); 1135 log.print("contiguous_available=" SIZE_FORMAT, _cmsGen->contiguous_available()); 1136 log.print("promotion_rate=%g", stats().promotion_rate()); 1137 log.print("cms_allocation_rate=%g", stats().cms_allocation_rate()); 1138 log.print("occupancy=%3.7f", _cmsGen->occupancy()); 1139 log.print("initiatingOccupancy=%3.7f", _cmsGen->initiating_occupancy()); 1140 log.print("cms_time_since_begin=%3.7f", stats().cms_time_since_begin()); 1141 log.print("cms_time_since_end=%3.7f", stats().cms_time_since_end()); 1142 log.print("metadata initialized %d", MetaspaceGC::should_concurrent_collect()); 1143 } 1144 // ------------------------------------------------------------------ 1145 1146 // If the estimated time to complete a cms collection (cms_duration()) 1147 // is less than the estimated time remaining until the cms generation 1148 // is full, start a collection. 1149 if (!UseCMSInitiatingOccupancyOnly) { 1150 if (stats().valid()) { 1151 if (stats().time_until_cms_start() == 0.0) { 1152 return true; 1153 } 1154 } else { 1155 // We want to conservatively collect somewhat early in order 1156 // to try and "bootstrap" our CMS/promotion statistics; 1157 // this branch will not fire after the first successful CMS 1158 // collection because the stats should then be valid. 1159 if (_cmsGen->occupancy() >= _bootstrap_occupancy) { 1160 log.print(" CMSCollector: collect for bootstrapping statistics: occupancy = %f, boot occupancy = %f", 1161 _cmsGen->occupancy(), _bootstrap_occupancy); 1162 return true; 1163 } 1164 } 1165 } 1166 1167 // Otherwise, we start a collection cycle if 1168 // old gen want a collection cycle started. Each may use 1169 // an appropriate criterion for making this decision. 1170 // XXX We need to make sure that the gen expansion 1171 // criterion dovetails well with this. XXX NEED TO FIX THIS 1172 if (_cmsGen->should_concurrent_collect()) { 1173 log.print("CMS old gen initiated"); 1174 return true; 1175 } 1176 1177 // We start a collection if we believe an incremental collection may fail; 1178 // this is not likely to be productive in practice because it's probably too 1179 // late anyway. 1180 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1181 assert(gch->collector_policy()->is_generation_policy(), 1182 "You may want to check the correctness of the following"); 1183 if (gch->incremental_collection_will_fail(true /* consult_young */)) { 1184 log.print("CMSCollector: collect because incremental collection will fail "); 1185 return true; 1186 } 1187 1188 if (MetaspaceGC::should_concurrent_collect()) { 1189 log.print("CMSCollector: collect for metadata allocation "); 1190 return true; 1191 } 1192 1193 // CMSTriggerInterval starts a CMS cycle if enough time has passed. 1194 if (CMSTriggerInterval >= 0) { 1195 if (CMSTriggerInterval == 0) { 1196 // Trigger always 1197 return true; 1198 } 1199 1200 // Check the CMS time since begin (we do not check the stats validity 1201 // as we want to be able to trigger the first CMS cycle as well) 1202 if (stats().cms_time_since_begin() >= (CMSTriggerInterval / ((double) MILLIUNITS))) { 1203 if (stats().valid()) { 1204 log.print("CMSCollector: collect because of trigger interval (time since last begin %3.7f secs)", 1205 stats().cms_time_since_begin()); 1206 } else { 1207 log.print("CMSCollector: collect because of trigger interval (first collection)"); 1208 } 1209 return true; 1210 } 1211 } 1212 1213 return false; 1214 } 1215 1216 void CMSCollector::set_did_compact(bool v) { _cmsGen->set_did_compact(v); } 1217 1218 // Clear _expansion_cause fields of constituent generations 1219 void CMSCollector::clear_expansion_cause() { 1220 _cmsGen->clear_expansion_cause(); 1221 } 1222 1223 // We should be conservative in starting a collection cycle. To 1224 // start too eagerly runs the risk of collecting too often in the 1225 // extreme. To collect too rarely falls back on full collections, 1226 // which works, even if not optimum in terms of concurrent work. 1227 // As a work around for too eagerly collecting, use the flag 1228 // UseCMSInitiatingOccupancyOnly. This also has the advantage of 1229 // giving the user an easily understandable way of controlling the 1230 // collections. 1231 // We want to start a new collection cycle if any of the following 1232 // conditions hold: 1233 // . our current occupancy exceeds the configured initiating occupancy 1234 // for this generation, or 1235 // . we recently needed to expand this space and have not, since that 1236 // expansion, done a collection of this generation, or 1237 // . the underlying space believes that it may be a good idea to initiate 1238 // a concurrent collection (this may be based on criteria such as the 1239 // following: the space uses linear allocation and linear allocation is 1240 // going to fail, or there is believed to be excessive fragmentation in 1241 // the generation, etc... or ... 1242 // [.(currently done by CMSCollector::shouldConcurrentCollect() only for 1243 // the case of the old generation; see CR 6543076): 1244 // we may be approaching a point at which allocation requests may fail because 1245 // we will be out of sufficient free space given allocation rate estimates.] 1246 bool ConcurrentMarkSweepGeneration::should_concurrent_collect() const { 1247 1248 assert_lock_strong(freelistLock()); 1249 if (occupancy() > initiating_occupancy()) { 1250 log_trace(gc)(" %s: collect because of occupancy %f / %f ", 1251 short_name(), occupancy(), initiating_occupancy()); 1252 return true; 1253 } 1254 if (UseCMSInitiatingOccupancyOnly) { 1255 return false; 1256 } 1257 if (expansion_cause() == CMSExpansionCause::_satisfy_allocation) { 1258 log_trace(gc)(" %s: collect because expanded for allocation ", short_name()); 1259 return true; 1260 } 1261 return false; 1262 } 1263 1264 void ConcurrentMarkSweepGeneration::collect(bool full, 1265 bool clear_all_soft_refs, 1266 size_t size, 1267 bool tlab) 1268 { 1269 collector()->collect(full, clear_all_soft_refs, size, tlab); 1270 } 1271 1272 void CMSCollector::collect(bool full, 1273 bool clear_all_soft_refs, 1274 size_t size, 1275 bool tlab) 1276 { 1277 // The following "if" branch is present for defensive reasons. 1278 // In the current uses of this interface, it can be replaced with: 1279 // assert(!GCLocker.is_active(), "Can't be called otherwise"); 1280 // But I am not placing that assert here to allow future 1281 // generality in invoking this interface. 1282 if (GCLocker::is_active()) { 1283 // A consistency test for GCLocker 1284 assert(GCLocker::needs_gc(), "Should have been set already"); 1285 // Skip this foreground collection, instead 1286 // expanding the heap if necessary. 1287 // Need the free list locks for the call to free() in compute_new_size() 1288 compute_new_size(); 1289 return; 1290 } 1291 acquire_control_and_collect(full, clear_all_soft_refs); 1292 } 1293 1294 void CMSCollector::request_full_gc(unsigned int full_gc_count, GCCause::Cause cause) { 1295 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1296 unsigned int gc_count = gch->total_full_collections(); 1297 if (gc_count == full_gc_count) { 1298 MutexLockerEx y(CGC_lock, Mutex::_no_safepoint_check_flag); 1299 _full_gc_requested = true; 1300 _full_gc_cause = cause; 1301 CGC_lock->notify(); // nudge CMS thread 1302 } else { 1303 assert(gc_count > full_gc_count, "Error: causal loop"); 1304 } 1305 } 1306 1307 bool CMSCollector::is_external_interruption() { 1308 GCCause::Cause cause = GenCollectedHeap::heap()->gc_cause(); 1309 return GCCause::is_user_requested_gc(cause) || 1310 GCCause::is_serviceability_requested_gc(cause); 1311 } 1312 1313 void CMSCollector::report_concurrent_mode_interruption() { 1314 if (is_external_interruption()) { 1315 log_debug(gc)("Concurrent mode interrupted"); 1316 } else { 1317 log_debug(gc)("Concurrent mode failure"); 1318 _gc_tracer_cm->report_concurrent_mode_failure(); 1319 } 1320 } 1321 1322 1323 // The foreground and background collectors need to coordinate in order 1324 // to make sure that they do not mutually interfere with CMS collections. 1325 // When a background collection is active, 1326 // the foreground collector may need to take over (preempt) and 1327 // synchronously complete an ongoing collection. Depending on the 1328 // frequency of the background collections and the heap usage 1329 // of the application, this preemption can be seldom or frequent. 1330 // There are only certain 1331 // points in the background collection that the "collection-baton" 1332 // can be passed to the foreground collector. 1333 // 1334 // The foreground collector will wait for the baton before 1335 // starting any part of the collection. The foreground collector 1336 // will only wait at one location. 1337 // 1338 // The background collector will yield the baton before starting a new 1339 // phase of the collection (e.g., before initial marking, marking from roots, 1340 // precleaning, final re-mark, sweep etc.) This is normally done at the head 1341 // of the loop which switches the phases. The background collector does some 1342 // of the phases (initial mark, final re-mark) with the world stopped. 1343 // Because of locking involved in stopping the world, 1344 // the foreground collector should not block waiting for the background 1345 // collector when it is doing a stop-the-world phase. The background 1346 // collector will yield the baton at an additional point just before 1347 // it enters a stop-the-world phase. Once the world is stopped, the 1348 // background collector checks the phase of the collection. If the 1349 // phase has not changed, it proceeds with the collection. If the 1350 // phase has changed, it skips that phase of the collection. See 1351 // the comments on the use of the Heap_lock in collect_in_background(). 1352 // 1353 // Variable used in baton passing. 1354 // _foregroundGCIsActive - Set to true by the foreground collector when 1355 // it wants the baton. The foreground clears it when it has finished 1356 // the collection. 1357 // _foregroundGCShouldWait - Set to true by the background collector 1358 // when it is running. The foreground collector waits while 1359 // _foregroundGCShouldWait is true. 1360 // CGC_lock - monitor used to protect access to the above variables 1361 // and to notify the foreground and background collectors. 1362 // _collectorState - current state of the CMS collection. 1363 // 1364 // The foreground collector 1365 // acquires the CGC_lock 1366 // sets _foregroundGCIsActive 1367 // waits on the CGC_lock for _foregroundGCShouldWait to be false 1368 // various locks acquired in preparation for the collection 1369 // are released so as not to block the background collector 1370 // that is in the midst of a collection 1371 // proceeds with the collection 1372 // clears _foregroundGCIsActive 1373 // returns 1374 // 1375 // The background collector in a loop iterating on the phases of the 1376 // collection 1377 // acquires the CGC_lock 1378 // sets _foregroundGCShouldWait 1379 // if _foregroundGCIsActive is set 1380 // clears _foregroundGCShouldWait, notifies _CGC_lock 1381 // waits on _CGC_lock for _foregroundGCIsActive to become false 1382 // and exits the loop. 1383 // otherwise 1384 // proceed with that phase of the collection 1385 // if the phase is a stop-the-world phase, 1386 // yield the baton once more just before enqueueing 1387 // the stop-world CMS operation (executed by the VM thread). 1388 // returns after all phases of the collection are done 1389 // 1390 1391 void CMSCollector::acquire_control_and_collect(bool full, 1392 bool clear_all_soft_refs) { 1393 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint"); 1394 assert(!Thread::current()->is_ConcurrentGC_thread(), 1395 "shouldn't try to acquire control from self!"); 1396 1397 // Start the protocol for acquiring control of the 1398 // collection from the background collector (aka CMS thread). 1399 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 1400 "VM thread should have CMS token"); 1401 // Remember the possibly interrupted state of an ongoing 1402 // concurrent collection 1403 CollectorState first_state = _collectorState; 1404 1405 // Signal to a possibly ongoing concurrent collection that 1406 // we want to do a foreground collection. 1407 _foregroundGCIsActive = true; 1408 1409 // release locks and wait for a notify from the background collector 1410 // releasing the locks in only necessary for phases which 1411 // do yields to improve the granularity of the collection. 1412 assert_lock_strong(bitMapLock()); 1413 // We need to lock the Free list lock for the space that we are 1414 // currently collecting. 1415 assert(haveFreelistLocks(), "Must be holding free list locks"); 1416 bitMapLock()->unlock(); 1417 releaseFreelistLocks(); 1418 { 1419 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1420 if (_foregroundGCShouldWait) { 1421 // We are going to be waiting for action for the CMS thread; 1422 // it had better not be gone (for instance at shutdown)! 1423 assert(ConcurrentMarkSweepThread::cmst() != NULL && !ConcurrentMarkSweepThread::cmst()->has_terminated(), 1424 "CMS thread must be running"); 1425 // Wait here until the background collector gives us the go-ahead 1426 ConcurrentMarkSweepThread::clear_CMS_flag( 1427 ConcurrentMarkSweepThread::CMS_vm_has_token); // release token 1428 // Get a possibly blocked CMS thread going: 1429 // Note that we set _foregroundGCIsActive true above, 1430 // without protection of the CGC_lock. 1431 CGC_lock->notify(); 1432 assert(!ConcurrentMarkSweepThread::vm_thread_wants_cms_token(), 1433 "Possible deadlock"); 1434 while (_foregroundGCShouldWait) { 1435 // wait for notification 1436 CGC_lock->wait(Mutex::_no_safepoint_check_flag); 1437 // Possibility of delay/starvation here, since CMS token does 1438 // not know to give priority to VM thread? Actually, i think 1439 // there wouldn't be any delay/starvation, but the proof of 1440 // that "fact" (?) appears non-trivial. XXX 20011219YSR 1441 } 1442 ConcurrentMarkSweepThread::set_CMS_flag( 1443 ConcurrentMarkSweepThread::CMS_vm_has_token); 1444 } 1445 } 1446 // The CMS_token is already held. Get back the other locks. 1447 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 1448 "VM thread should have CMS token"); 1449 getFreelistLocks(); 1450 bitMapLock()->lock_without_safepoint_check(); 1451 log_debug(gc, state)("CMS foreground collector has asked for control " INTPTR_FORMAT " with first state %d", 1452 p2i(Thread::current()), first_state); 1453 log_debug(gc, state)(" gets control with state %d", _collectorState); 1454 1455 // Inform cms gen if this was due to partial collection failing. 1456 // The CMS gen may use this fact to determine its expansion policy. 1457 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1458 if (gch->incremental_collection_will_fail(false /* don't consult_young */)) { 1459 assert(!_cmsGen->incremental_collection_failed(), 1460 "Should have been noticed, reacted to and cleared"); 1461 _cmsGen->set_incremental_collection_failed(); 1462 } 1463 1464 if (first_state > Idling) { 1465 report_concurrent_mode_interruption(); 1466 } 1467 1468 set_did_compact(true); 1469 1470 // If the collection is being acquired from the background 1471 // collector, there may be references on the discovered 1472 // references lists. Abandon those references, since some 1473 // of them may have become unreachable after concurrent 1474 // discovery; the STW compacting collector will redo discovery 1475 // more precisely, without being subject to floating garbage. 1476 // Leaving otherwise unreachable references in the discovered 1477 // lists would require special handling. 1478 ref_processor()->disable_discovery(); 1479 ref_processor()->abandon_partial_discovery(); 1480 ref_processor()->verify_no_references_recorded(); 1481 1482 if (first_state > Idling) { 1483 save_heap_summary(); 1484 } 1485 1486 do_compaction_work(clear_all_soft_refs); 1487 1488 // Has the GC time limit been exceeded? 1489 size_t max_eden_size = _young_gen->max_eden_size(); 1490 GCCause::Cause gc_cause = gch->gc_cause(); 1491 size_policy()->check_gc_overhead_limit(_young_gen->used(), 1492 _young_gen->eden()->used(), 1493 _cmsGen->max_capacity(), 1494 max_eden_size, 1495 full, 1496 gc_cause, 1497 gch->collector_policy()); 1498 1499 // Reset the expansion cause, now that we just completed 1500 // a collection cycle. 1501 clear_expansion_cause(); 1502 _foregroundGCIsActive = false; 1503 return; 1504 } 1505 1506 // Resize the tenured generation 1507 // after obtaining the free list locks for the 1508 // two generations. 1509 void CMSCollector::compute_new_size() { 1510 assert_locked_or_safepoint(Heap_lock); 1511 FreelistLocker z(this); 1512 MetaspaceGC::compute_new_size(); 1513 _cmsGen->compute_new_size_free_list(); 1514 } 1515 1516 // A work method used by the foreground collector to do 1517 // a mark-sweep-compact. 1518 void CMSCollector::do_compaction_work(bool clear_all_soft_refs) { 1519 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1520 1521 STWGCTimer* gc_timer = GenMarkSweep::gc_timer(); 1522 gc_timer->register_gc_start(); 1523 1524 SerialOldTracer* gc_tracer = GenMarkSweep::gc_tracer(); 1525 gc_tracer->report_gc_start(gch->gc_cause(), gc_timer->gc_start()); 1526 1527 gch->pre_full_gc_dump(gc_timer); 1528 1529 GCTraceTime(Trace, gc, phases) t("CMS:MSC"); 1530 1531 // Temporarily widen the span of the weak reference processing to 1532 // the entire heap. 1533 MemRegion new_span(GenCollectedHeap::heap()->reserved_region()); 1534 ReferenceProcessorSpanMutator rp_mut_span(ref_processor(), new_span); 1535 // Temporarily, clear the "is_alive_non_header" field of the 1536 // reference processor. 1537 ReferenceProcessorIsAliveMutator rp_mut_closure(ref_processor(), NULL); 1538 // Temporarily make reference _processing_ single threaded (non-MT). 1539 ReferenceProcessorMTProcMutator rp_mut_mt_processing(ref_processor(), false); 1540 // Temporarily make refs discovery atomic 1541 ReferenceProcessorAtomicMutator rp_mut_atomic(ref_processor(), true); 1542 // Temporarily make reference _discovery_ single threaded (non-MT) 1543 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false); 1544 1545 ref_processor()->set_enqueuing_is_done(false); 1546 ref_processor()->enable_discovery(); 1547 ref_processor()->setup_policy(clear_all_soft_refs); 1548 // If an asynchronous collection finishes, the _modUnionTable is 1549 // all clear. If we are assuming the collection from an asynchronous 1550 // collection, clear the _modUnionTable. 1551 assert(_collectorState != Idling || _modUnionTable.isAllClear(), 1552 "_modUnionTable should be clear if the baton was not passed"); 1553 _modUnionTable.clear_all(); 1554 assert(_collectorState != Idling || _ct->klass_rem_set()->mod_union_is_clear(), 1555 "mod union for klasses should be clear if the baton was passed"); 1556 _ct->klass_rem_set()->clear_mod_union(); 1557 1558 // We must adjust the allocation statistics being maintained 1559 // in the free list space. We do so by reading and clearing 1560 // the sweep timer and updating the block flux rate estimates below. 1561 assert(!_intra_sweep_timer.is_active(), "_intra_sweep_timer should be inactive"); 1562 if (_inter_sweep_timer.is_active()) { 1563 _inter_sweep_timer.stop(); 1564 // Note that we do not use this sample to update the _inter_sweep_estimate. 1565 _cmsGen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()), 1566 _inter_sweep_estimate.padded_average(), 1567 _intra_sweep_estimate.padded_average()); 1568 } 1569 1570 GenMarkSweep::invoke_at_safepoint(ref_processor(), clear_all_soft_refs); 1571 #ifdef ASSERT 1572 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 1573 size_t free_size = cms_space->free(); 1574 assert(free_size == 1575 pointer_delta(cms_space->end(), cms_space->compaction_top()) 1576 * HeapWordSize, 1577 "All the free space should be compacted into one chunk at top"); 1578 assert(cms_space->dictionary()->total_chunk_size( 1579 debug_only(cms_space->freelistLock())) == 0 || 1580 cms_space->totalSizeInIndexedFreeLists() == 0, 1581 "All the free space should be in a single chunk"); 1582 size_t num = cms_space->totalCount(); 1583 assert((free_size == 0 && num == 0) || 1584 (free_size > 0 && (num == 1 || num == 2)), 1585 "There should be at most 2 free chunks after compaction"); 1586 #endif // ASSERT 1587 _collectorState = Resetting; 1588 assert(_restart_addr == NULL, 1589 "Should have been NULL'd before baton was passed"); 1590 reset_stw(); 1591 _cmsGen->reset_after_compaction(); 1592 _concurrent_cycles_since_last_unload = 0; 1593 1594 // Clear any data recorded in the PLAB chunk arrays. 1595 if (_survivor_plab_array != NULL) { 1596 reset_survivor_plab_arrays(); 1597 } 1598 1599 // Adjust the per-size allocation stats for the next epoch. 1600 _cmsGen->cmsSpace()->endSweepFLCensus(sweep_count() /* fake */); 1601 // Restart the "inter sweep timer" for the next epoch. 1602 _inter_sweep_timer.reset(); 1603 _inter_sweep_timer.start(); 1604 1605 // No longer a need to do a concurrent collection for Metaspace. 1606 MetaspaceGC::set_should_concurrent_collect(false); 1607 1608 gch->post_full_gc_dump(gc_timer); 1609 1610 gc_timer->register_gc_end(); 1611 1612 gc_tracer->report_gc_end(gc_timer->gc_end(), gc_timer->time_partitions()); 1613 1614 // For a mark-sweep-compact, compute_new_size() will be called 1615 // in the heap's do_collection() method. 1616 } 1617 1618 void CMSCollector::print_eden_and_survivor_chunk_arrays() { 1619 Log(gc, heap) log; 1620 if (!log.is_trace()) { 1621 return; 1622 } 1623 1624 ContiguousSpace* eden_space = _young_gen->eden(); 1625 ContiguousSpace* from_space = _young_gen->from(); 1626 ContiguousSpace* to_space = _young_gen->to(); 1627 // Eden 1628 if (_eden_chunk_array != NULL) { 1629 log.trace("eden " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")", 1630 p2i(eden_space->bottom()), p2i(eden_space->top()), 1631 p2i(eden_space->end()), eden_space->capacity()); 1632 log.trace("_eden_chunk_index=" SIZE_FORMAT ", _eden_chunk_capacity=" SIZE_FORMAT, 1633 _eden_chunk_index, _eden_chunk_capacity); 1634 for (size_t i = 0; i < _eden_chunk_index; i++) { 1635 log.trace("_eden_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT, i, p2i(_eden_chunk_array[i])); 1636 } 1637 } 1638 // Survivor 1639 if (_survivor_chunk_array != NULL) { 1640 log.trace("survivor " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")", 1641 p2i(from_space->bottom()), p2i(from_space->top()), 1642 p2i(from_space->end()), from_space->capacity()); 1643 log.trace("_survivor_chunk_index=" SIZE_FORMAT ", _survivor_chunk_capacity=" SIZE_FORMAT, 1644 _survivor_chunk_index, _survivor_chunk_capacity); 1645 for (size_t i = 0; i < _survivor_chunk_index; i++) { 1646 log.trace("_survivor_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT, i, p2i(_survivor_chunk_array[i])); 1647 } 1648 } 1649 } 1650 1651 void CMSCollector::getFreelistLocks() const { 1652 // Get locks for all free lists in all generations that this 1653 // collector is responsible for 1654 _cmsGen->freelistLock()->lock_without_safepoint_check(); 1655 } 1656 1657 void CMSCollector::releaseFreelistLocks() const { 1658 // Release locks for all free lists in all generations that this 1659 // collector is responsible for 1660 _cmsGen->freelistLock()->unlock(); 1661 } 1662 1663 bool CMSCollector::haveFreelistLocks() const { 1664 // Check locks for all free lists in all generations that this 1665 // collector is responsible for 1666 assert_lock_strong(_cmsGen->freelistLock()); 1667 PRODUCT_ONLY(ShouldNotReachHere()); 1668 return true; 1669 } 1670 1671 // A utility class that is used by the CMS collector to 1672 // temporarily "release" the foreground collector from its 1673 // usual obligation to wait for the background collector to 1674 // complete an ongoing phase before proceeding. 1675 class ReleaseForegroundGC: public StackObj { 1676 private: 1677 CMSCollector* _c; 1678 public: 1679 ReleaseForegroundGC(CMSCollector* c) : _c(c) { 1680 assert(_c->_foregroundGCShouldWait, "Else should not need to call"); 1681 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1682 // allow a potentially blocked foreground collector to proceed 1683 _c->_foregroundGCShouldWait = false; 1684 if (_c->_foregroundGCIsActive) { 1685 CGC_lock->notify(); 1686 } 1687 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 1688 "Possible deadlock"); 1689 } 1690 1691 ~ReleaseForegroundGC() { 1692 assert(!_c->_foregroundGCShouldWait, "Usage protocol violation?"); 1693 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1694 _c->_foregroundGCShouldWait = true; 1695 } 1696 }; 1697 1698 void CMSCollector::collect_in_background(GCCause::Cause cause) { 1699 assert(Thread::current()->is_ConcurrentGC_thread(), 1700 "A CMS asynchronous collection is only allowed on a CMS thread."); 1701 1702 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1703 { 1704 bool safepoint_check = Mutex::_no_safepoint_check_flag; 1705 MutexLockerEx hl(Heap_lock, safepoint_check); 1706 FreelistLocker fll(this); 1707 MutexLockerEx x(CGC_lock, safepoint_check); 1708 if (_foregroundGCIsActive) { 1709 // The foreground collector is. Skip this 1710 // background collection. 1711 assert(!_foregroundGCShouldWait, "Should be clear"); 1712 return; 1713 } else { 1714 assert(_collectorState == Idling, "Should be idling before start."); 1715 _collectorState = InitialMarking; 1716 register_gc_start(cause); 1717 // Reset the expansion cause, now that we are about to begin 1718 // a new cycle. 1719 clear_expansion_cause(); 1720 1721 // Clear the MetaspaceGC flag since a concurrent collection 1722 // is starting but also clear it after the collection. 1723 MetaspaceGC::set_should_concurrent_collect(false); 1724 } 1725 // Decide if we want to enable class unloading as part of the 1726 // ensuing concurrent GC cycle. 1727 update_should_unload_classes(); 1728 _full_gc_requested = false; // acks all outstanding full gc requests 1729 _full_gc_cause = GCCause::_no_gc; 1730 // Signal that we are about to start a collection 1731 gch->increment_total_full_collections(); // ... starting a collection cycle 1732 _collection_count_start = gch->total_full_collections(); 1733 } 1734 1735 size_t prev_used = _cmsGen->used(); 1736 1737 // The change of the collection state is normally done at this level; 1738 // the exceptions are phases that are executed while the world is 1739 // stopped. For those phases the change of state is done while the 1740 // world is stopped. For baton passing purposes this allows the 1741 // background collector to finish the phase and change state atomically. 1742 // The foreground collector cannot wait on a phase that is done 1743 // while the world is stopped because the foreground collector already 1744 // has the world stopped and would deadlock. 1745 while (_collectorState != Idling) { 1746 log_debug(gc, state)("Thread " INTPTR_FORMAT " in CMS state %d", 1747 p2i(Thread::current()), _collectorState); 1748 // The foreground collector 1749 // holds the Heap_lock throughout its collection. 1750 // holds the CMS token (but not the lock) 1751 // except while it is waiting for the background collector to yield. 1752 // 1753 // The foreground collector should be blocked (not for long) 1754 // if the background collector is about to start a phase 1755 // executed with world stopped. If the background 1756 // collector has already started such a phase, the 1757 // foreground collector is blocked waiting for the 1758 // Heap_lock. The stop-world phases (InitialMarking and FinalMarking) 1759 // are executed in the VM thread. 1760 // 1761 // The locking order is 1762 // PendingListLock (PLL) -- if applicable (FinalMarking) 1763 // Heap_lock (both this & PLL locked in VM_CMS_Operation::prologue()) 1764 // CMS token (claimed in 1765 // stop_world_and_do() --> 1766 // safepoint_synchronize() --> 1767 // CMSThread::synchronize()) 1768 1769 { 1770 // Check if the FG collector wants us to yield. 1771 CMSTokenSync x(true); // is cms thread 1772 if (waitForForegroundGC()) { 1773 // We yielded to a foreground GC, nothing more to be 1774 // done this round. 1775 assert(_foregroundGCShouldWait == false, "We set it to false in " 1776 "waitForForegroundGC()"); 1777 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " exiting collection CMS state %d", 1778 p2i(Thread::current()), _collectorState); 1779 return; 1780 } else { 1781 // The background collector can run but check to see if the 1782 // foreground collector has done a collection while the 1783 // background collector was waiting to get the CGC_lock 1784 // above. If yes, break so that _foregroundGCShouldWait 1785 // is cleared before returning. 1786 if (_collectorState == Idling) { 1787 break; 1788 } 1789 } 1790 } 1791 1792 assert(_foregroundGCShouldWait, "Foreground collector, if active, " 1793 "should be waiting"); 1794 1795 switch (_collectorState) { 1796 case InitialMarking: 1797 { 1798 ReleaseForegroundGC x(this); 1799 stats().record_cms_begin(); 1800 VM_CMS_Initial_Mark initial_mark_op(this); 1801 VMThread::execute(&initial_mark_op); 1802 } 1803 // The collector state may be any legal state at this point 1804 // since the background collector may have yielded to the 1805 // foreground collector. 1806 break; 1807 case Marking: 1808 // initial marking in checkpointRootsInitialWork has been completed 1809 if (markFromRoots()) { // we were successful 1810 assert(_collectorState == Precleaning, "Collector state should " 1811 "have changed"); 1812 } else { 1813 assert(_foregroundGCIsActive, "Internal state inconsistency"); 1814 } 1815 break; 1816 case Precleaning: 1817 // marking from roots in markFromRoots has been completed 1818 preclean(); 1819 assert(_collectorState == AbortablePreclean || 1820 _collectorState == FinalMarking, 1821 "Collector state should have changed"); 1822 break; 1823 case AbortablePreclean: 1824 abortable_preclean(); 1825 assert(_collectorState == FinalMarking, "Collector state should " 1826 "have changed"); 1827 break; 1828 case FinalMarking: 1829 { 1830 ReleaseForegroundGC x(this); 1831 1832 VM_CMS_Final_Remark final_remark_op(this); 1833 VMThread::execute(&final_remark_op); 1834 } 1835 assert(_foregroundGCShouldWait, "block post-condition"); 1836 break; 1837 case Sweeping: 1838 // final marking in checkpointRootsFinal has been completed 1839 sweep(); 1840 assert(_collectorState == Resizing, "Collector state change " 1841 "to Resizing must be done under the free_list_lock"); 1842 1843 case Resizing: { 1844 // Sweeping has been completed... 1845 // At this point the background collection has completed. 1846 // Don't move the call to compute_new_size() down 1847 // into code that might be executed if the background 1848 // collection was preempted. 1849 { 1850 ReleaseForegroundGC x(this); // unblock FG collection 1851 MutexLockerEx y(Heap_lock, Mutex::_no_safepoint_check_flag); 1852 CMSTokenSync z(true); // not strictly needed. 1853 if (_collectorState == Resizing) { 1854 compute_new_size(); 1855 save_heap_summary(); 1856 _collectorState = Resetting; 1857 } else { 1858 assert(_collectorState == Idling, "The state should only change" 1859 " because the foreground collector has finished the collection"); 1860 } 1861 } 1862 break; 1863 } 1864 case Resetting: 1865 // CMS heap resizing has been completed 1866 reset_concurrent(); 1867 assert(_collectorState == Idling, "Collector state should " 1868 "have changed"); 1869 1870 MetaspaceGC::set_should_concurrent_collect(false); 1871 1872 stats().record_cms_end(); 1873 // Don't move the concurrent_phases_end() and compute_new_size() 1874 // calls to here because a preempted background collection 1875 // has it's state set to "Resetting". 1876 break; 1877 case Idling: 1878 default: 1879 ShouldNotReachHere(); 1880 break; 1881 } 1882 log_debug(gc, state)(" Thread " INTPTR_FORMAT " done - next CMS state %d", 1883 p2i(Thread::current()), _collectorState); 1884 assert(_foregroundGCShouldWait, "block post-condition"); 1885 } 1886 1887 // Should this be in gc_epilogue? 1888 collector_policy()->counters()->update_counters(); 1889 1890 { 1891 // Clear _foregroundGCShouldWait and, in the event that the 1892 // foreground collector is waiting, notify it, before 1893 // returning. 1894 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1895 _foregroundGCShouldWait = false; 1896 if (_foregroundGCIsActive) { 1897 CGC_lock->notify(); 1898 } 1899 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 1900 "Possible deadlock"); 1901 } 1902 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " exiting collection CMS state %d", 1903 p2i(Thread::current()), _collectorState); 1904 log_info(gc, heap)("Old: " SIZE_FORMAT "K->" SIZE_FORMAT "K(" SIZE_FORMAT "K)", 1905 prev_used / K, _cmsGen->used()/K, _cmsGen->capacity() /K); 1906 } 1907 1908 void CMSCollector::register_gc_start(GCCause::Cause cause) { 1909 _cms_start_registered = true; 1910 _gc_timer_cm->register_gc_start(); 1911 _gc_tracer_cm->report_gc_start(cause, _gc_timer_cm->gc_start()); 1912 } 1913 1914 void CMSCollector::register_gc_end() { 1915 if (_cms_start_registered) { 1916 report_heap_summary(GCWhen::AfterGC); 1917 1918 _gc_timer_cm->register_gc_end(); 1919 _gc_tracer_cm->report_gc_end(_gc_timer_cm->gc_end(), _gc_timer_cm->time_partitions()); 1920 _cms_start_registered = false; 1921 } 1922 } 1923 1924 void CMSCollector::save_heap_summary() { 1925 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1926 _last_heap_summary = gch->create_heap_summary(); 1927 _last_metaspace_summary = gch->create_metaspace_summary(); 1928 } 1929 1930 void CMSCollector::report_heap_summary(GCWhen::Type when) { 1931 _gc_tracer_cm->report_gc_heap_summary(when, _last_heap_summary); 1932 _gc_tracer_cm->report_metaspace_summary(when, _last_metaspace_summary); 1933 } 1934 1935 bool CMSCollector::waitForForegroundGC() { 1936 bool res = false; 1937 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 1938 "CMS thread should have CMS token"); 1939 // Block the foreground collector until the 1940 // background collectors decides whether to 1941 // yield. 1942 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1943 _foregroundGCShouldWait = true; 1944 if (_foregroundGCIsActive) { 1945 // The background collector yields to the 1946 // foreground collector and returns a value 1947 // indicating that it has yielded. The foreground 1948 // collector can proceed. 1949 res = true; 1950 _foregroundGCShouldWait = false; 1951 ConcurrentMarkSweepThread::clear_CMS_flag( 1952 ConcurrentMarkSweepThread::CMS_cms_has_token); 1953 ConcurrentMarkSweepThread::set_CMS_flag( 1954 ConcurrentMarkSweepThread::CMS_cms_wants_token); 1955 // Get a possibly blocked foreground thread going 1956 CGC_lock->notify(); 1957 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " waiting at CMS state %d", 1958 p2i(Thread::current()), _collectorState); 1959 while (_foregroundGCIsActive) { 1960 CGC_lock->wait(Mutex::_no_safepoint_check_flag); 1961 } 1962 ConcurrentMarkSweepThread::set_CMS_flag( 1963 ConcurrentMarkSweepThread::CMS_cms_has_token); 1964 ConcurrentMarkSweepThread::clear_CMS_flag( 1965 ConcurrentMarkSweepThread::CMS_cms_wants_token); 1966 } 1967 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " continuing at CMS state %d", 1968 p2i(Thread::current()), _collectorState); 1969 return res; 1970 } 1971 1972 // Because of the need to lock the free lists and other structures in 1973 // the collector, common to all the generations that the collector is 1974 // collecting, we need the gc_prologues of individual CMS generations 1975 // delegate to their collector. It may have been simpler had the 1976 // current infrastructure allowed one to call a prologue on a 1977 // collector. In the absence of that we have the generation's 1978 // prologue delegate to the collector, which delegates back 1979 // some "local" work to a worker method in the individual generations 1980 // that it's responsible for collecting, while itself doing any 1981 // work common to all generations it's responsible for. A similar 1982 // comment applies to the gc_epilogue()'s. 1983 // The role of the variable _between_prologue_and_epilogue is to 1984 // enforce the invocation protocol. 1985 void CMSCollector::gc_prologue(bool full) { 1986 // Call gc_prologue_work() for the CMSGen 1987 // we are responsible for. 1988 1989 // The following locking discipline assumes that we are only called 1990 // when the world is stopped. 1991 assert(SafepointSynchronize::is_at_safepoint(), "world is stopped assumption"); 1992 1993 // The CMSCollector prologue must call the gc_prologues for the 1994 // "generations" that it's responsible 1995 // for. 1996 1997 assert( Thread::current()->is_VM_thread() 1998 || ( CMSScavengeBeforeRemark 1999 && Thread::current()->is_ConcurrentGC_thread()), 2000 "Incorrect thread type for prologue execution"); 2001 2002 if (_between_prologue_and_epilogue) { 2003 // We have already been invoked; this is a gc_prologue delegation 2004 // from yet another CMS generation that we are responsible for, just 2005 // ignore it since all relevant work has already been done. 2006 return; 2007 } 2008 2009 // set a bit saying prologue has been called; cleared in epilogue 2010 _between_prologue_and_epilogue = true; 2011 // Claim locks for common data structures, then call gc_prologue_work() 2012 // for each CMSGen. 2013 2014 getFreelistLocks(); // gets free list locks on constituent spaces 2015 bitMapLock()->lock_without_safepoint_check(); 2016 2017 // Should call gc_prologue_work() for all cms gens we are responsible for 2018 bool duringMarking = _collectorState >= Marking 2019 && _collectorState < Sweeping; 2020 2021 // The young collections clear the modified oops state, which tells if 2022 // there are any modified oops in the class. The remark phase also needs 2023 // that information. Tell the young collection to save the union of all 2024 // modified klasses. 2025 if (duringMarking) { 2026 _ct->klass_rem_set()->set_accumulate_modified_oops(true); 2027 } 2028 2029 bool registerClosure = duringMarking; 2030 2031 _cmsGen->gc_prologue_work(full, registerClosure, &_modUnionClosurePar); 2032 2033 if (!full) { 2034 stats().record_gc0_begin(); 2035 } 2036 } 2037 2038 void ConcurrentMarkSweepGeneration::gc_prologue(bool full) { 2039 2040 _capacity_at_prologue = capacity(); 2041 _used_at_prologue = used(); 2042 2043 // We enable promotion tracking so that card-scanning can recognize 2044 // which objects have been promoted during this GC and skip them. 2045 for (uint i = 0; i < ParallelGCThreads; i++) { 2046 _par_gc_thread_states[i]->promo.startTrackingPromotions(); 2047 } 2048 2049 // Delegate to CMScollector which knows how to coordinate between 2050 // this and any other CMS generations that it is responsible for 2051 // collecting. 2052 collector()->gc_prologue(full); 2053 } 2054 2055 // This is a "private" interface for use by this generation's CMSCollector. 2056 // Not to be called directly by any other entity (for instance, 2057 // GenCollectedHeap, which calls the "public" gc_prologue method above). 2058 void ConcurrentMarkSweepGeneration::gc_prologue_work(bool full, 2059 bool registerClosure, ModUnionClosure* modUnionClosure) { 2060 assert(!incremental_collection_failed(), "Shouldn't be set yet"); 2061 assert(cmsSpace()->preconsumptionDirtyCardClosure() == NULL, 2062 "Should be NULL"); 2063 if (registerClosure) { 2064 cmsSpace()->setPreconsumptionDirtyCardClosure(modUnionClosure); 2065 } 2066 cmsSpace()->gc_prologue(); 2067 // Clear stat counters 2068 NOT_PRODUCT( 2069 assert(_numObjectsPromoted == 0, "check"); 2070 assert(_numWordsPromoted == 0, "check"); 2071 log_develop_trace(gc, alloc)("Allocated " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes concurrently", 2072 _numObjectsAllocated, _numWordsAllocated*sizeof(HeapWord)); 2073 _numObjectsAllocated = 0; 2074 _numWordsAllocated = 0; 2075 ) 2076 } 2077 2078 void CMSCollector::gc_epilogue(bool full) { 2079 // The following locking discipline assumes that we are only called 2080 // when the world is stopped. 2081 assert(SafepointSynchronize::is_at_safepoint(), 2082 "world is stopped assumption"); 2083 2084 // Currently the CMS epilogue (see CompactibleFreeListSpace) merely checks 2085 // if linear allocation blocks need to be appropriately marked to allow the 2086 // the blocks to be parsable. We also check here whether we need to nudge the 2087 // CMS collector thread to start a new cycle (if it's not already active). 2088 assert( Thread::current()->is_VM_thread() 2089 || ( CMSScavengeBeforeRemark 2090 && Thread::current()->is_ConcurrentGC_thread()), 2091 "Incorrect thread type for epilogue execution"); 2092 2093 if (!_between_prologue_and_epilogue) { 2094 // We have already been invoked; this is a gc_epilogue delegation 2095 // from yet another CMS generation that we are responsible for, just 2096 // ignore it since all relevant work has already been done. 2097 return; 2098 } 2099 assert(haveFreelistLocks(), "must have freelist locks"); 2100 assert_lock_strong(bitMapLock()); 2101 2102 _ct->klass_rem_set()->set_accumulate_modified_oops(false); 2103 2104 _cmsGen->gc_epilogue_work(full); 2105 2106 if (_collectorState == AbortablePreclean || _collectorState == Precleaning) { 2107 // in case sampling was not already enabled, enable it 2108 _start_sampling = true; 2109 } 2110 // reset _eden_chunk_array so sampling starts afresh 2111 _eden_chunk_index = 0; 2112 2113 size_t cms_used = _cmsGen->cmsSpace()->used(); 2114 2115 // update performance counters - this uses a special version of 2116 // update_counters() that allows the utilization to be passed as a 2117 // parameter, avoiding multiple calls to used(). 2118 // 2119 _cmsGen->update_counters(cms_used); 2120 2121 bitMapLock()->unlock(); 2122 releaseFreelistLocks(); 2123 2124 if (!CleanChunkPoolAsync) { 2125 Chunk::clean_chunk_pool(); 2126 } 2127 2128 set_did_compact(false); 2129 _between_prologue_and_epilogue = false; // ready for next cycle 2130 } 2131 2132 void ConcurrentMarkSweepGeneration::gc_epilogue(bool full) { 2133 collector()->gc_epilogue(full); 2134 2135 // When using ParNew, promotion tracking should have already been 2136 // disabled. However, the prologue (which enables promotion 2137 // tracking) and epilogue are called irrespective of the type of 2138 // GC. So they will also be called before and after Full GCs, during 2139 // which promotion tracking will not be explicitly disabled. So, 2140 // it's safer to also disable it here too (to be symmetric with 2141 // enabling it in the prologue). 2142 for (uint i = 0; i < ParallelGCThreads; i++) { 2143 _par_gc_thread_states[i]->promo.stopTrackingPromotions(); 2144 } 2145 } 2146 2147 void ConcurrentMarkSweepGeneration::gc_epilogue_work(bool full) { 2148 assert(!incremental_collection_failed(), "Should have been cleared"); 2149 cmsSpace()->setPreconsumptionDirtyCardClosure(NULL); 2150 cmsSpace()->gc_epilogue(); 2151 // Print stat counters 2152 NOT_PRODUCT( 2153 assert(_numObjectsAllocated == 0, "check"); 2154 assert(_numWordsAllocated == 0, "check"); 2155 log_develop_trace(gc, promotion)("Promoted " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes", 2156 _numObjectsPromoted, _numWordsPromoted*sizeof(HeapWord)); 2157 _numObjectsPromoted = 0; 2158 _numWordsPromoted = 0; 2159 ) 2160 2161 // Call down the chain in contiguous_available needs the freelistLock 2162 // so print this out before releasing the freeListLock. 2163 log_develop_trace(gc)(" Contiguous available " SIZE_FORMAT " bytes ", contiguous_available()); 2164 } 2165 2166 #ifndef PRODUCT 2167 bool CMSCollector::have_cms_token() { 2168 Thread* thr = Thread::current(); 2169 if (thr->is_VM_thread()) { 2170 return ConcurrentMarkSweepThread::vm_thread_has_cms_token(); 2171 } else if (thr->is_ConcurrentGC_thread()) { 2172 return ConcurrentMarkSweepThread::cms_thread_has_cms_token(); 2173 } else if (thr->is_GC_task_thread()) { 2174 return ConcurrentMarkSweepThread::vm_thread_has_cms_token() && 2175 ParGCRareEvent_lock->owned_by_self(); 2176 } 2177 return false; 2178 } 2179 2180 // Check reachability of the given heap address in CMS generation, 2181 // treating all other generations as roots. 2182 bool CMSCollector::is_cms_reachable(HeapWord* addr) { 2183 // We could "guarantee" below, rather than assert, but I'll 2184 // leave these as "asserts" so that an adventurous debugger 2185 // could try this in the product build provided some subset of 2186 // the conditions were met, provided they were interested in the 2187 // results and knew that the computation below wouldn't interfere 2188 // with other concurrent computations mutating the structures 2189 // being read or written. 2190 assert(SafepointSynchronize::is_at_safepoint(), 2191 "Else mutations in object graph will make answer suspect"); 2192 assert(have_cms_token(), "Should hold cms token"); 2193 assert(haveFreelistLocks(), "must hold free list locks"); 2194 assert_lock_strong(bitMapLock()); 2195 2196 // Clear the marking bit map array before starting, but, just 2197 // for kicks, first report if the given address is already marked 2198 tty->print_cr("Start: Address " PTR_FORMAT " is%s marked", p2i(addr), 2199 _markBitMap.isMarked(addr) ? "" : " not"); 2200 2201 if (verify_after_remark()) { 2202 MutexLockerEx x(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag); 2203 bool result = verification_mark_bm()->isMarked(addr); 2204 tty->print_cr("TransitiveMark: Address " PTR_FORMAT " %s marked", p2i(addr), 2205 result ? "IS" : "is NOT"); 2206 return result; 2207 } else { 2208 tty->print_cr("Could not compute result"); 2209 return false; 2210 } 2211 } 2212 #endif 2213 2214 void 2215 CMSCollector::print_on_error(outputStream* st) { 2216 CMSCollector* collector = ConcurrentMarkSweepGeneration::_collector; 2217 if (collector != NULL) { 2218 CMSBitMap* bitmap = &collector->_markBitMap; 2219 st->print_cr("Marking Bits: (CMSBitMap*) " PTR_FORMAT, p2i(bitmap)); 2220 bitmap->print_on_error(st, " Bits: "); 2221 2222 st->cr(); 2223 2224 CMSBitMap* mut_bitmap = &collector->_modUnionTable; 2225 st->print_cr("Mod Union Table: (CMSBitMap*) " PTR_FORMAT, p2i(mut_bitmap)); 2226 mut_bitmap->print_on_error(st, " Bits: "); 2227 } 2228 } 2229 2230 //////////////////////////////////////////////////////// 2231 // CMS Verification Support 2232 //////////////////////////////////////////////////////// 2233 // Following the remark phase, the following invariant 2234 // should hold -- each object in the CMS heap which is 2235 // marked in markBitMap() should be marked in the verification_mark_bm(). 2236 2237 class VerifyMarkedClosure: public BitMapClosure { 2238 CMSBitMap* _marks; 2239 bool _failed; 2240 2241 public: 2242 VerifyMarkedClosure(CMSBitMap* bm): _marks(bm), _failed(false) {} 2243 2244 bool do_bit(size_t offset) { 2245 HeapWord* addr = _marks->offsetToHeapWord(offset); 2246 if (!_marks->isMarked(addr)) { 2247 Log(gc, verify) log; 2248 ResourceMark rm; 2249 oop(addr)->print_on(log.error_stream()); 2250 log.error(" (" INTPTR_FORMAT " should have been marked)", p2i(addr)); 2251 _failed = true; 2252 } 2253 return true; 2254 } 2255 2256 bool failed() { return _failed; } 2257 }; 2258 2259 bool CMSCollector::verify_after_remark() { 2260 GCTraceTime(Info, gc, phases, verify) tm("Verifying CMS Marking."); 2261 MutexLockerEx ml(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag); 2262 static bool init = false; 2263 2264 assert(SafepointSynchronize::is_at_safepoint(), 2265 "Else mutations in object graph will make answer suspect"); 2266 assert(have_cms_token(), 2267 "Else there may be mutual interference in use of " 2268 " verification data structures"); 2269 assert(_collectorState > Marking && _collectorState <= Sweeping, 2270 "Else marking info checked here may be obsolete"); 2271 assert(haveFreelistLocks(), "must hold free list locks"); 2272 assert_lock_strong(bitMapLock()); 2273 2274 2275 // Allocate marking bit map if not already allocated 2276 if (!init) { // first time 2277 if (!verification_mark_bm()->allocate(_span)) { 2278 return false; 2279 } 2280 init = true; 2281 } 2282 2283 assert(verification_mark_stack()->isEmpty(), "Should be empty"); 2284 2285 // Turn off refs discovery -- so we will be tracing through refs. 2286 // This is as intended, because by this time 2287 // GC must already have cleared any refs that need to be cleared, 2288 // and traced those that need to be marked; moreover, 2289 // the marking done here is not going to interfere in any 2290 // way with the marking information used by GC. 2291 NoRefDiscovery no_discovery(ref_processor()); 2292 2293 #if defined(COMPILER2) || INCLUDE_JVMCI 2294 DerivedPointerTableDeactivate dpt_deact; 2295 #endif 2296 2297 // Clear any marks from a previous round 2298 verification_mark_bm()->clear_all(); 2299 assert(verification_mark_stack()->isEmpty(), "markStack should be empty"); 2300 verify_work_stacks_empty(); 2301 2302 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2303 gch->ensure_parsability(false); // fill TLABs, but no need to retire them 2304 // Update the saved marks which may affect the root scans. 2305 gch->save_marks(); 2306 2307 if (CMSRemarkVerifyVariant == 1) { 2308 // In this first variant of verification, we complete 2309 // all marking, then check if the new marks-vector is 2310 // a subset of the CMS marks-vector. 2311 verify_after_remark_work_1(); 2312 } else { 2313 guarantee(CMSRemarkVerifyVariant == 2, "Range checking for CMSRemarkVerifyVariant should guarantee 1 or 2"); 2314 // In this second variant of verification, we flag an error 2315 // (i.e. an object reachable in the new marks-vector not reachable 2316 // in the CMS marks-vector) immediately, also indicating the 2317 // identify of an object (A) that references the unmarked object (B) -- 2318 // presumably, a mutation to A failed to be picked up by preclean/remark? 2319 verify_after_remark_work_2(); 2320 } 2321 2322 return true; 2323 } 2324 2325 void CMSCollector::verify_after_remark_work_1() { 2326 ResourceMark rm; 2327 HandleMark hm; 2328 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2329 2330 // Get a clear set of claim bits for the roots processing to work with. 2331 ClassLoaderDataGraph::clear_claimed_marks(); 2332 2333 // Mark from roots one level into CMS 2334 MarkRefsIntoClosure notOlder(_span, verification_mark_bm()); 2335 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2336 2337 { 2338 StrongRootsScope srs(1); 2339 2340 gch->cms_process_roots(&srs, 2341 true, // young gen as roots 2342 GenCollectedHeap::ScanningOption(roots_scanning_options()), 2343 should_unload_classes(), 2344 ¬Older, 2345 NULL); 2346 } 2347 2348 // Now mark from the roots 2349 MarkFromRootsClosure markFromRootsClosure(this, _span, 2350 verification_mark_bm(), verification_mark_stack(), 2351 false /* don't yield */, true /* verifying */); 2352 assert(_restart_addr == NULL, "Expected pre-condition"); 2353 verification_mark_bm()->iterate(&markFromRootsClosure); 2354 while (_restart_addr != NULL) { 2355 // Deal with stack overflow: by restarting at the indicated 2356 // address. 2357 HeapWord* ra = _restart_addr; 2358 markFromRootsClosure.reset(ra); 2359 _restart_addr = NULL; 2360 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end()); 2361 } 2362 assert(verification_mark_stack()->isEmpty(), "Should have been drained"); 2363 verify_work_stacks_empty(); 2364 2365 // Marking completed -- now verify that each bit marked in 2366 // verification_mark_bm() is also marked in markBitMap(); flag all 2367 // errors by printing corresponding objects. 2368 VerifyMarkedClosure vcl(markBitMap()); 2369 verification_mark_bm()->iterate(&vcl); 2370 if (vcl.failed()) { 2371 Log(gc, verify) log; 2372 log.error("Failed marking verification after remark"); 2373 ResourceMark rm; 2374 gch->print_on(log.error_stream()); 2375 fatal("CMS: failed marking verification after remark"); 2376 } 2377 } 2378 2379 class VerifyKlassOopsKlassClosure : public KlassClosure { 2380 class VerifyKlassOopsClosure : public OopClosure { 2381 CMSBitMap* _bitmap; 2382 public: 2383 VerifyKlassOopsClosure(CMSBitMap* bitmap) : _bitmap(bitmap) { } 2384 void do_oop(oop* p) { guarantee(*p == NULL || _bitmap->isMarked((HeapWord*) *p), "Should be marked"); } 2385 void do_oop(narrowOop* p) { ShouldNotReachHere(); } 2386 } _oop_closure; 2387 public: 2388 VerifyKlassOopsKlassClosure(CMSBitMap* bitmap) : _oop_closure(bitmap) {} 2389 void do_klass(Klass* k) { 2390 k->oops_do(&_oop_closure); 2391 } 2392 }; 2393 2394 void CMSCollector::verify_after_remark_work_2() { 2395 ResourceMark rm; 2396 HandleMark hm; 2397 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2398 2399 // Get a clear set of claim bits for the roots processing to work with. 2400 ClassLoaderDataGraph::clear_claimed_marks(); 2401 2402 // Mark from roots one level into CMS 2403 MarkRefsIntoVerifyClosure notOlder(_span, verification_mark_bm(), 2404 markBitMap()); 2405 CLDToOopClosure cld_closure(¬Older, true); 2406 2407 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2408 2409 { 2410 StrongRootsScope srs(1); 2411 2412 gch->cms_process_roots(&srs, 2413 true, // young gen as roots 2414 GenCollectedHeap::ScanningOption(roots_scanning_options()), 2415 should_unload_classes(), 2416 ¬Older, 2417 &cld_closure); 2418 } 2419 2420 // Now mark from the roots 2421 MarkFromRootsVerifyClosure markFromRootsClosure(this, _span, 2422 verification_mark_bm(), markBitMap(), verification_mark_stack()); 2423 assert(_restart_addr == NULL, "Expected pre-condition"); 2424 verification_mark_bm()->iterate(&markFromRootsClosure); 2425 while (_restart_addr != NULL) { 2426 // Deal with stack overflow: by restarting at the indicated 2427 // address. 2428 HeapWord* ra = _restart_addr; 2429 markFromRootsClosure.reset(ra); 2430 _restart_addr = NULL; 2431 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end()); 2432 } 2433 assert(verification_mark_stack()->isEmpty(), "Should have been drained"); 2434 verify_work_stacks_empty(); 2435 2436 VerifyKlassOopsKlassClosure verify_klass_oops(verification_mark_bm()); 2437 ClassLoaderDataGraph::classes_do(&verify_klass_oops); 2438 2439 // Marking completed -- now verify that each bit marked in 2440 // verification_mark_bm() is also marked in markBitMap(); flag all 2441 // errors by printing corresponding objects. 2442 VerifyMarkedClosure vcl(markBitMap()); 2443 verification_mark_bm()->iterate(&vcl); 2444 assert(!vcl.failed(), "Else verification above should not have succeeded"); 2445 } 2446 2447 void ConcurrentMarkSweepGeneration::save_marks() { 2448 // delegate to CMS space 2449 cmsSpace()->save_marks(); 2450 } 2451 2452 bool ConcurrentMarkSweepGeneration::no_allocs_since_save_marks() { 2453 return cmsSpace()->no_allocs_since_save_marks(); 2454 } 2455 2456 #define CMS_SINCE_SAVE_MARKS_DEFN(OopClosureType, nv_suffix) \ 2457 \ 2458 void ConcurrentMarkSweepGeneration:: \ 2459 oop_since_save_marks_iterate##nv_suffix(OopClosureType* cl) { \ 2460 cl->set_generation(this); \ 2461 cmsSpace()->oop_since_save_marks_iterate##nv_suffix(cl); \ 2462 cl->reset_generation(); \ 2463 save_marks(); \ 2464 } 2465 2466 ALL_SINCE_SAVE_MARKS_CLOSURES(CMS_SINCE_SAVE_MARKS_DEFN) 2467 2468 void 2469 ConcurrentMarkSweepGeneration::oop_iterate(ExtendedOopClosure* cl) { 2470 if (freelistLock()->owned_by_self()) { 2471 Generation::oop_iterate(cl); 2472 } else { 2473 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2474 Generation::oop_iterate(cl); 2475 } 2476 } 2477 2478 void 2479 ConcurrentMarkSweepGeneration::object_iterate(ObjectClosure* cl) { 2480 if (freelistLock()->owned_by_self()) { 2481 Generation::object_iterate(cl); 2482 } else { 2483 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2484 Generation::object_iterate(cl); 2485 } 2486 } 2487 2488 void 2489 ConcurrentMarkSweepGeneration::safe_object_iterate(ObjectClosure* cl) { 2490 if (freelistLock()->owned_by_self()) { 2491 Generation::safe_object_iterate(cl); 2492 } else { 2493 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2494 Generation::safe_object_iterate(cl); 2495 } 2496 } 2497 2498 void 2499 ConcurrentMarkSweepGeneration::post_compact() { 2500 } 2501 2502 void 2503 ConcurrentMarkSweepGeneration::prepare_for_verify() { 2504 // Fix the linear allocation blocks to look like free blocks. 2505 2506 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those 2507 // are not called when the heap is verified during universe initialization and 2508 // at vm shutdown. 2509 if (freelistLock()->owned_by_self()) { 2510 cmsSpace()->prepare_for_verify(); 2511 } else { 2512 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag); 2513 cmsSpace()->prepare_for_verify(); 2514 } 2515 } 2516 2517 void 2518 ConcurrentMarkSweepGeneration::verify() { 2519 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those 2520 // are not called when the heap is verified during universe initialization and 2521 // at vm shutdown. 2522 if (freelistLock()->owned_by_self()) { 2523 cmsSpace()->verify(); 2524 } else { 2525 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag); 2526 cmsSpace()->verify(); 2527 } 2528 } 2529 2530 void CMSCollector::verify() { 2531 _cmsGen->verify(); 2532 } 2533 2534 #ifndef PRODUCT 2535 bool CMSCollector::overflow_list_is_empty() const { 2536 assert(_num_par_pushes >= 0, "Inconsistency"); 2537 if (_overflow_list == NULL) { 2538 assert(_num_par_pushes == 0, "Inconsistency"); 2539 } 2540 return _overflow_list == NULL; 2541 } 2542 2543 // The methods verify_work_stacks_empty() and verify_overflow_empty() 2544 // merely consolidate assertion checks that appear to occur together frequently. 2545 void CMSCollector::verify_work_stacks_empty() const { 2546 assert(_markStack.isEmpty(), "Marking stack should be empty"); 2547 assert(overflow_list_is_empty(), "Overflow list should be empty"); 2548 } 2549 2550 void CMSCollector::verify_overflow_empty() const { 2551 assert(overflow_list_is_empty(), "Overflow list should be empty"); 2552 assert(no_preserved_marks(), "No preserved marks"); 2553 } 2554 #endif // PRODUCT 2555 2556 // Decide if we want to enable class unloading as part of the 2557 // ensuing concurrent GC cycle. We will collect and 2558 // unload classes if it's the case that: 2559 // (a) class unloading is enabled at the command line, and 2560 // (b) old gen is getting really full 2561 // NOTE: Provided there is no change in the state of the heap between 2562 // calls to this method, it should have idempotent results. Moreover, 2563 // its results should be monotonically increasing (i.e. going from 0 to 1, 2564 // but not 1 to 0) between successive calls between which the heap was 2565 // not collected. For the implementation below, it must thus rely on 2566 // the property that concurrent_cycles_since_last_unload() 2567 // will not decrease unless a collection cycle happened and that 2568 // _cmsGen->is_too_full() are 2569 // themselves also monotonic in that sense. See check_monotonicity() 2570 // below. 2571 void CMSCollector::update_should_unload_classes() { 2572 _should_unload_classes = false; 2573 if (CMSClassUnloadingEnabled) { 2574 _should_unload_classes = (concurrent_cycles_since_last_unload() >= 2575 CMSClassUnloadingMaxInterval) 2576 || _cmsGen->is_too_full(); 2577 } 2578 } 2579 2580 bool ConcurrentMarkSweepGeneration::is_too_full() const { 2581 bool res = should_concurrent_collect(); 2582 res = res && (occupancy() > (double)CMSIsTooFullPercentage/100.0); 2583 return res; 2584 } 2585 2586 void CMSCollector::setup_cms_unloading_and_verification_state() { 2587 const bool should_verify = VerifyBeforeGC || VerifyAfterGC || VerifyDuringGC 2588 || VerifyBeforeExit; 2589 const int rso = GenCollectedHeap::SO_AllCodeCache; 2590 2591 // We set the proper root for this CMS cycle here. 2592 if (should_unload_classes()) { // Should unload classes this cycle 2593 remove_root_scanning_option(rso); // Shrink the root set appropriately 2594 set_verifying(should_verify); // Set verification state for this cycle 2595 return; // Nothing else needs to be done at this time 2596 } 2597 2598 // Not unloading classes this cycle 2599 assert(!should_unload_classes(), "Inconsistency!"); 2600 2601 // If we are not unloading classes then add SO_AllCodeCache to root 2602 // scanning options. 2603 add_root_scanning_option(rso); 2604 2605 if ((!verifying() || unloaded_classes_last_cycle()) && should_verify) { 2606 set_verifying(true); 2607 } else if (verifying() && !should_verify) { 2608 // We were verifying, but some verification flags got disabled. 2609 set_verifying(false); 2610 // Exclude symbols, strings and code cache elements from root scanning to 2611 // reduce IM and RM pauses. 2612 remove_root_scanning_option(rso); 2613 } 2614 } 2615 2616 2617 #ifndef PRODUCT 2618 HeapWord* CMSCollector::block_start(const void* p) const { 2619 const HeapWord* addr = (HeapWord*)p; 2620 if (_span.contains(p)) { 2621 if (_cmsGen->cmsSpace()->is_in_reserved(addr)) { 2622 return _cmsGen->cmsSpace()->block_start(p); 2623 } 2624 } 2625 return NULL; 2626 } 2627 #endif 2628 2629 HeapWord* 2630 ConcurrentMarkSweepGeneration::expand_and_allocate(size_t word_size, 2631 bool tlab, 2632 bool parallel) { 2633 CMSSynchronousYieldRequest yr; 2634 assert(!tlab, "Can't deal with TLAB allocation"); 2635 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2636 expand_for_gc_cause(word_size*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_satisfy_allocation); 2637 if (GCExpandToAllocateDelayMillis > 0) { 2638 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 2639 } 2640 return have_lock_and_allocate(word_size, tlab); 2641 } 2642 2643 void ConcurrentMarkSweepGeneration::expand_for_gc_cause( 2644 size_t bytes, 2645 size_t expand_bytes, 2646 CMSExpansionCause::Cause cause) 2647 { 2648 2649 bool success = expand(bytes, expand_bytes); 2650 2651 // remember why we expanded; this information is used 2652 // by shouldConcurrentCollect() when making decisions on whether to start 2653 // a new CMS cycle. 2654 if (success) { 2655 set_expansion_cause(cause); 2656 log_trace(gc)("Expanded CMS gen for %s", CMSExpansionCause::to_string(cause)); 2657 } 2658 } 2659 2660 HeapWord* ConcurrentMarkSweepGeneration::expand_and_par_lab_allocate(CMSParGCThreadState* ps, size_t word_sz) { 2661 HeapWord* res = NULL; 2662 MutexLocker x(ParGCRareEvent_lock); 2663 while (true) { 2664 // Expansion by some other thread might make alloc OK now: 2665 res = ps->lab.alloc(word_sz); 2666 if (res != NULL) return res; 2667 // If there's not enough expansion space available, give up. 2668 if (_virtual_space.uncommitted_size() < (word_sz * HeapWordSize)) { 2669 return NULL; 2670 } 2671 // Otherwise, we try expansion. 2672 expand_for_gc_cause(word_sz*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_allocate_par_lab); 2673 // Now go around the loop and try alloc again; 2674 // A competing par_promote might beat us to the expansion space, 2675 // so we may go around the loop again if promotion fails again. 2676 if (GCExpandToAllocateDelayMillis > 0) { 2677 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 2678 } 2679 } 2680 } 2681 2682 2683 bool ConcurrentMarkSweepGeneration::expand_and_ensure_spooling_space( 2684 PromotionInfo* promo) { 2685 MutexLocker x(ParGCRareEvent_lock); 2686 size_t refill_size_bytes = promo->refillSize() * HeapWordSize; 2687 while (true) { 2688 // Expansion by some other thread might make alloc OK now: 2689 if (promo->ensure_spooling_space()) { 2690 assert(promo->has_spooling_space(), 2691 "Post-condition of successful ensure_spooling_space()"); 2692 return true; 2693 } 2694 // If there's not enough expansion space available, give up. 2695 if (_virtual_space.uncommitted_size() < refill_size_bytes) { 2696 return false; 2697 } 2698 // Otherwise, we try expansion. 2699 expand_for_gc_cause(refill_size_bytes, MinHeapDeltaBytes, CMSExpansionCause::_allocate_par_spooling_space); 2700 // Now go around the loop and try alloc again; 2701 // A competing allocation might beat us to the expansion space, 2702 // so we may go around the loop again if allocation fails again. 2703 if (GCExpandToAllocateDelayMillis > 0) { 2704 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 2705 } 2706 } 2707 } 2708 2709 void ConcurrentMarkSweepGeneration::shrink(size_t bytes) { 2710 // Only shrink if a compaction was done so that all the free space 2711 // in the generation is in a contiguous block at the end. 2712 if (did_compact()) { 2713 CardGeneration::shrink(bytes); 2714 } 2715 } 2716 2717 void ConcurrentMarkSweepGeneration::assert_correct_size_change_locking() { 2718 assert_locked_or_safepoint(Heap_lock); 2719 } 2720 2721 void ConcurrentMarkSweepGeneration::shrink_free_list_by(size_t bytes) { 2722 assert_locked_or_safepoint(Heap_lock); 2723 assert_lock_strong(freelistLock()); 2724 log_trace(gc)("Shrinking of CMS not yet implemented"); 2725 return; 2726 } 2727 2728 2729 // Simple ctor/dtor wrapper for accounting & timer chores around concurrent 2730 // phases. 2731 class CMSPhaseAccounting: public StackObj { 2732 public: 2733 CMSPhaseAccounting(CMSCollector *collector, 2734 const char *title); 2735 ~CMSPhaseAccounting(); 2736 2737 private: 2738 CMSCollector *_collector; 2739 const char *_title; 2740 GCTraceConcTime(Info, gc) _trace_time; 2741 2742 public: 2743 // Not MT-safe; so do not pass around these StackObj's 2744 // where they may be accessed by other threads. 2745 double wallclock_millis() { 2746 return TimeHelper::counter_to_millis(os::elapsed_counter() - _trace_time.start_time()); 2747 } 2748 }; 2749 2750 CMSPhaseAccounting::CMSPhaseAccounting(CMSCollector *collector, 2751 const char *title) : 2752 _collector(collector), _title(title), _trace_time(title) { 2753 2754 _collector->resetYields(); 2755 _collector->resetTimer(); 2756 _collector->startTimer(); 2757 _collector->gc_timer_cm()->register_gc_concurrent_start(title); 2758 } 2759 2760 CMSPhaseAccounting::~CMSPhaseAccounting() { 2761 _collector->gc_timer_cm()->register_gc_concurrent_end(); 2762 _collector->stopTimer(); 2763 log_debug(gc)("Concurrent active time: %.3fms", TimeHelper::counter_to_seconds(_collector->timerTicks())); 2764 log_trace(gc)(" (CMS %s yielded %d times)", _title, _collector->yields()); 2765 } 2766 2767 // CMS work 2768 2769 // The common parts of CMSParInitialMarkTask and CMSParRemarkTask. 2770 class CMSParMarkTask : public AbstractGangTask { 2771 protected: 2772 CMSCollector* _collector; 2773 uint _n_workers; 2774 CMSParMarkTask(const char* name, CMSCollector* collector, uint n_workers) : 2775 AbstractGangTask(name), 2776 _collector(collector), 2777 _n_workers(n_workers) {} 2778 // Work method in support of parallel rescan ... of young gen spaces 2779 void do_young_space_rescan(OopsInGenClosure* cl, 2780 ContiguousSpace* space, 2781 HeapWord** chunk_array, size_t chunk_top); 2782 void work_on_young_gen_roots(OopsInGenClosure* cl); 2783 }; 2784 2785 // Parallel initial mark task 2786 class CMSParInitialMarkTask: public CMSParMarkTask { 2787 StrongRootsScope* _strong_roots_scope; 2788 public: 2789 CMSParInitialMarkTask(CMSCollector* collector, StrongRootsScope* strong_roots_scope, uint n_workers) : 2790 CMSParMarkTask("Scan roots and young gen for initial mark in parallel", collector, n_workers), 2791 _strong_roots_scope(strong_roots_scope) {} 2792 void work(uint worker_id); 2793 }; 2794 2795 // Checkpoint the roots into this generation from outside 2796 // this generation. [Note this initial checkpoint need only 2797 // be approximate -- we'll do a catch up phase subsequently.] 2798 void CMSCollector::checkpointRootsInitial() { 2799 assert(_collectorState == InitialMarking, "Wrong collector state"); 2800 check_correct_thread_executing(); 2801 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 2802 2803 save_heap_summary(); 2804 report_heap_summary(GCWhen::BeforeGC); 2805 2806 ReferenceProcessor* rp = ref_processor(); 2807 assert(_restart_addr == NULL, "Control point invariant"); 2808 { 2809 // acquire locks for subsequent manipulations 2810 MutexLockerEx x(bitMapLock(), 2811 Mutex::_no_safepoint_check_flag); 2812 checkpointRootsInitialWork(); 2813 // enable ("weak") refs discovery 2814 rp->enable_discovery(); 2815 _collectorState = Marking; 2816 } 2817 } 2818 2819 void CMSCollector::checkpointRootsInitialWork() { 2820 assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped"); 2821 assert(_collectorState == InitialMarking, "just checking"); 2822 2823 // Already have locks. 2824 assert_lock_strong(bitMapLock()); 2825 assert(_markBitMap.isAllClear(), "was reset at end of previous cycle"); 2826 2827 // Setup the verification and class unloading state for this 2828 // CMS collection cycle. 2829 setup_cms_unloading_and_verification_state(); 2830 2831 GCTraceTime(Trace, gc, phases) ts("checkpointRootsInitialWork", _gc_timer_cm); 2832 2833 // Reset all the PLAB chunk arrays if necessary. 2834 if (_survivor_plab_array != NULL && !CMSPLABRecordAlways) { 2835 reset_survivor_plab_arrays(); 2836 } 2837 2838 ResourceMark rm; 2839 HandleMark hm; 2840 2841 MarkRefsIntoClosure notOlder(_span, &_markBitMap); 2842 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2843 2844 verify_work_stacks_empty(); 2845 verify_overflow_empty(); 2846 2847 gch->ensure_parsability(false); // fill TLABs, but no need to retire them 2848 // Update the saved marks which may affect the root scans. 2849 gch->save_marks(); 2850 2851 // weak reference processing has not started yet. 2852 ref_processor()->set_enqueuing_is_done(false); 2853 2854 // Need to remember all newly created CLDs, 2855 // so that we can guarantee that the remark finds them. 2856 ClassLoaderDataGraph::remember_new_clds(true); 2857 2858 // Whenever a CLD is found, it will be claimed before proceeding to mark 2859 // the klasses. The claimed marks need to be cleared before marking starts. 2860 ClassLoaderDataGraph::clear_claimed_marks(); 2861 2862 print_eden_and_survivor_chunk_arrays(); 2863 2864 { 2865 #if defined(COMPILER2) || INCLUDE_JVMCI 2866 DerivedPointerTableDeactivate dpt_deact; 2867 #endif 2868 if (CMSParallelInitialMarkEnabled) { 2869 // The parallel version. 2870 WorkGang* workers = gch->workers(); 2871 assert(workers != NULL, "Need parallel worker threads."); 2872 uint n_workers = workers->active_workers(); 2873 2874 StrongRootsScope srs(n_workers); 2875 2876 CMSParInitialMarkTask tsk(this, &srs, n_workers); 2877 initialize_sequential_subtasks_for_young_gen_rescan(n_workers); 2878 // If the total workers is greater than 1, then multiple workers 2879 // may be used at some time and the initialization has been set 2880 // such that the single threaded path cannot be used. 2881 if (workers->total_workers() > 1) { 2882 workers->run_task(&tsk); 2883 } else { 2884 tsk.work(0); 2885 } 2886 } else { 2887 // The serial version. 2888 CLDToOopClosure cld_closure(¬Older, true); 2889 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2890 2891 StrongRootsScope srs(1); 2892 2893 gch->cms_process_roots(&srs, 2894 true, // young gen as roots 2895 GenCollectedHeap::ScanningOption(roots_scanning_options()), 2896 should_unload_classes(), 2897 ¬Older, 2898 &cld_closure); 2899 } 2900 } 2901 2902 // Clear mod-union table; it will be dirtied in the prologue of 2903 // CMS generation per each young generation collection. 2904 2905 assert(_modUnionTable.isAllClear(), 2906 "Was cleared in most recent final checkpoint phase" 2907 " or no bits are set in the gc_prologue before the start of the next " 2908 "subsequent marking phase."); 2909 2910 assert(_ct->klass_rem_set()->mod_union_is_clear(), "Must be"); 2911 2912 // Save the end of the used_region of the constituent generations 2913 // to be used to limit the extent of sweep in each generation. 2914 save_sweep_limits(); 2915 verify_overflow_empty(); 2916 } 2917 2918 bool CMSCollector::markFromRoots() { 2919 // we might be tempted to assert that: 2920 // assert(!SafepointSynchronize::is_at_safepoint(), 2921 // "inconsistent argument?"); 2922 // However that wouldn't be right, because it's possible that 2923 // a safepoint is indeed in progress as a young generation 2924 // stop-the-world GC happens even as we mark in this generation. 2925 assert(_collectorState == Marking, "inconsistent state?"); 2926 check_correct_thread_executing(); 2927 verify_overflow_empty(); 2928 2929 // Weak ref discovery note: We may be discovering weak 2930 // refs in this generation concurrent (but interleaved) with 2931 // weak ref discovery by the young generation collector. 2932 2933 CMSTokenSyncWithLocks ts(true, bitMapLock()); 2934 GCTraceCPUTime tcpu; 2935 CMSPhaseAccounting pa(this, "Concurrent Mark"); 2936 bool res = markFromRootsWork(); 2937 if (res) { 2938 _collectorState = Precleaning; 2939 } else { // We failed and a foreground collection wants to take over 2940 assert(_foregroundGCIsActive, "internal state inconsistency"); 2941 assert(_restart_addr == NULL, "foreground will restart from scratch"); 2942 log_debug(gc)("bailing out to foreground collection"); 2943 } 2944 verify_overflow_empty(); 2945 return res; 2946 } 2947 2948 bool CMSCollector::markFromRootsWork() { 2949 // iterate over marked bits in bit map, doing a full scan and mark 2950 // from these roots using the following algorithm: 2951 // . if oop is to the right of the current scan pointer, 2952 // mark corresponding bit (we'll process it later) 2953 // . else (oop is to left of current scan pointer) 2954 // push oop on marking stack 2955 // . drain the marking stack 2956 2957 // Note that when we do a marking step we need to hold the 2958 // bit map lock -- recall that direct allocation (by mutators) 2959 // and promotion (by the young generation collector) is also 2960 // marking the bit map. [the so-called allocate live policy.] 2961 // Because the implementation of bit map marking is not 2962 // robust wrt simultaneous marking of bits in the same word, 2963 // we need to make sure that there is no such interference 2964 // between concurrent such updates. 2965 2966 // already have locks 2967 assert_lock_strong(bitMapLock()); 2968 2969 verify_work_stacks_empty(); 2970 verify_overflow_empty(); 2971 bool result = false; 2972 if (CMSConcurrentMTEnabled && ConcGCThreads > 0) { 2973 result = do_marking_mt(); 2974 } else { 2975 result = do_marking_st(); 2976 } 2977 return result; 2978 } 2979 2980 // Forward decl 2981 class CMSConcMarkingTask; 2982 2983 class CMSConcMarkingTerminator: public ParallelTaskTerminator { 2984 CMSCollector* _collector; 2985 CMSConcMarkingTask* _task; 2986 public: 2987 virtual void yield(); 2988 2989 // "n_threads" is the number of threads to be terminated. 2990 // "queue_set" is a set of work queues of other threads. 2991 // "collector" is the CMS collector associated with this task terminator. 2992 // "yield" indicates whether we need the gang as a whole to yield. 2993 CMSConcMarkingTerminator(int n_threads, TaskQueueSetSuper* queue_set, CMSCollector* collector) : 2994 ParallelTaskTerminator(n_threads, queue_set), 2995 _collector(collector) { } 2996 2997 void set_task(CMSConcMarkingTask* task) { 2998 _task = task; 2999 } 3000 }; 3001 3002 class CMSConcMarkingTerminatorTerminator: public TerminatorTerminator { 3003 CMSConcMarkingTask* _task; 3004 public: 3005 bool should_exit_termination(); 3006 void set_task(CMSConcMarkingTask* task) { 3007 _task = task; 3008 } 3009 }; 3010 3011 // MT Concurrent Marking Task 3012 class CMSConcMarkingTask: public YieldingFlexibleGangTask { 3013 CMSCollector* _collector; 3014 uint _n_workers; // requested/desired # workers 3015 bool _result; 3016 CompactibleFreeListSpace* _cms_space; 3017 char _pad_front[64]; // padding to ... 3018 HeapWord* volatile _global_finger; // ... avoid sharing cache line 3019 char _pad_back[64]; 3020 HeapWord* _restart_addr; 3021 3022 // Exposed here for yielding support 3023 Mutex* const _bit_map_lock; 3024 3025 // The per thread work queues, available here for stealing 3026 OopTaskQueueSet* _task_queues; 3027 3028 // Termination (and yielding) support 3029 CMSConcMarkingTerminator _term; 3030 CMSConcMarkingTerminatorTerminator _term_term; 3031 3032 public: 3033 CMSConcMarkingTask(CMSCollector* collector, 3034 CompactibleFreeListSpace* cms_space, 3035 YieldingFlexibleWorkGang* workers, 3036 OopTaskQueueSet* task_queues): 3037 YieldingFlexibleGangTask("Concurrent marking done multi-threaded"), 3038 _collector(collector), 3039 _cms_space(cms_space), 3040 _n_workers(0), _result(true), 3041 _task_queues(task_queues), 3042 _term(_n_workers, task_queues, _collector), 3043 _bit_map_lock(collector->bitMapLock()) 3044 { 3045 _requested_size = _n_workers; 3046 _term.set_task(this); 3047 _term_term.set_task(this); 3048 _restart_addr = _global_finger = _cms_space->bottom(); 3049 } 3050 3051 3052 OopTaskQueueSet* task_queues() { return _task_queues; } 3053 3054 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 3055 3056 HeapWord* volatile* global_finger_addr() { return &_global_finger; } 3057 3058 CMSConcMarkingTerminator* terminator() { return &_term; } 3059 3060 virtual void set_for_termination(uint active_workers) { 3061 terminator()->reset_for_reuse(active_workers); 3062 } 3063 3064 void work(uint worker_id); 3065 bool should_yield() { 3066 return ConcurrentMarkSweepThread::should_yield() 3067 && !_collector->foregroundGCIsActive(); 3068 } 3069 3070 virtual void coordinator_yield(); // stuff done by coordinator 3071 bool result() { return _result; } 3072 3073 void reset(HeapWord* ra) { 3074 assert(_global_finger >= _cms_space->end(), "Postcondition of ::work(i)"); 3075 _restart_addr = _global_finger = ra; 3076 _term.reset_for_reuse(); 3077 } 3078 3079 static bool get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 3080 OopTaskQueue* work_q); 3081 3082 private: 3083 void do_scan_and_mark(int i, CompactibleFreeListSpace* sp); 3084 void do_work_steal(int i); 3085 void bump_global_finger(HeapWord* f); 3086 }; 3087 3088 bool CMSConcMarkingTerminatorTerminator::should_exit_termination() { 3089 assert(_task != NULL, "Error"); 3090 return _task->yielding(); 3091 // Note that we do not need the disjunct || _task->should_yield() above 3092 // because we want terminating threads to yield only if the task 3093 // is already in the midst of yielding, which happens only after at least one 3094 // thread has yielded. 3095 } 3096 3097 void CMSConcMarkingTerminator::yield() { 3098 if (_task->should_yield()) { 3099 _task->yield(); 3100 } else { 3101 ParallelTaskTerminator::yield(); 3102 } 3103 } 3104 3105 //////////////////////////////////////////////////////////////// 3106 // Concurrent Marking Algorithm Sketch 3107 //////////////////////////////////////////////////////////////// 3108 // Until all tasks exhausted (both spaces): 3109 // -- claim next available chunk 3110 // -- bump global finger via CAS 3111 // -- find first object that starts in this chunk 3112 // and start scanning bitmap from that position 3113 // -- scan marked objects for oops 3114 // -- CAS-mark target, and if successful: 3115 // . if target oop is above global finger (volatile read) 3116 // nothing to do 3117 // . if target oop is in chunk and above local finger 3118 // then nothing to do 3119 // . else push on work-queue 3120 // -- Deal with possible overflow issues: 3121 // . local work-queue overflow causes stuff to be pushed on 3122 // global (common) overflow queue 3123 // . always first empty local work queue 3124 // . then get a batch of oops from global work queue if any 3125 // . then do work stealing 3126 // -- When all tasks claimed (both spaces) 3127 // and local work queue empty, 3128 // then in a loop do: 3129 // . check global overflow stack; steal a batch of oops and trace 3130 // . try to steal from other threads oif GOS is empty 3131 // . if neither is available, offer termination 3132 // -- Terminate and return result 3133 // 3134 void CMSConcMarkingTask::work(uint worker_id) { 3135 elapsedTimer _timer; 3136 ResourceMark rm; 3137 HandleMark hm; 3138 3139 DEBUG_ONLY(_collector->verify_overflow_empty();) 3140 3141 // Before we begin work, our work queue should be empty 3142 assert(work_queue(worker_id)->size() == 0, "Expected to be empty"); 3143 // Scan the bitmap covering _cms_space, tracing through grey objects. 3144 _timer.start(); 3145 do_scan_and_mark(worker_id, _cms_space); 3146 _timer.stop(); 3147 log_trace(gc, task)("Finished cms space scanning in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 3148 3149 // ... do work stealing 3150 _timer.reset(); 3151 _timer.start(); 3152 do_work_steal(worker_id); 3153 _timer.stop(); 3154 log_trace(gc, task)("Finished work stealing in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 3155 assert(_collector->_markStack.isEmpty(), "Should have been emptied"); 3156 assert(work_queue(worker_id)->size() == 0, "Should have been emptied"); 3157 // Note that under the current task protocol, the 3158 // following assertion is true even of the spaces 3159 // expanded since the completion of the concurrent 3160 // marking. XXX This will likely change under a strict 3161 // ABORT semantics. 3162 // After perm removal the comparison was changed to 3163 // greater than or equal to from strictly greater than. 3164 // Before perm removal the highest address sweep would 3165 // have been at the end of perm gen but now is at the 3166 // end of the tenured gen. 3167 assert(_global_finger >= _cms_space->end(), 3168 "All tasks have been completed"); 3169 DEBUG_ONLY(_collector->verify_overflow_empty();) 3170 } 3171 3172 void CMSConcMarkingTask::bump_global_finger(HeapWord* f) { 3173 HeapWord* read = _global_finger; 3174 HeapWord* cur = read; 3175 while (f > read) { 3176 cur = read; 3177 read = (HeapWord*) Atomic::cmpxchg_ptr(f, &_global_finger, cur); 3178 if (cur == read) { 3179 // our cas succeeded 3180 assert(_global_finger >= f, "protocol consistency"); 3181 break; 3182 } 3183 } 3184 } 3185 3186 // This is really inefficient, and should be redone by 3187 // using (not yet available) block-read and -write interfaces to the 3188 // stack and the work_queue. XXX FIX ME !!! 3189 bool CMSConcMarkingTask::get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 3190 OopTaskQueue* work_q) { 3191 // Fast lock-free check 3192 if (ovflw_stk->length() == 0) { 3193 return false; 3194 } 3195 assert(work_q->size() == 0, "Shouldn't steal"); 3196 MutexLockerEx ml(ovflw_stk->par_lock(), 3197 Mutex::_no_safepoint_check_flag); 3198 // Grab up to 1/4 the size of the work queue 3199 size_t num = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 3200 (size_t)ParGCDesiredObjsFromOverflowList); 3201 num = MIN2(num, ovflw_stk->length()); 3202 for (int i = (int) num; i > 0; i--) { 3203 oop cur = ovflw_stk->pop(); 3204 assert(cur != NULL, "Counted wrong?"); 3205 work_q->push(cur); 3206 } 3207 return num > 0; 3208 } 3209 3210 void CMSConcMarkingTask::do_scan_and_mark(int i, CompactibleFreeListSpace* sp) { 3211 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 3212 int n_tasks = pst->n_tasks(); 3213 // We allow that there may be no tasks to do here because 3214 // we are restarting after a stack overflow. 3215 assert(pst->valid() || n_tasks == 0, "Uninitialized use?"); 3216 uint nth_task = 0; 3217 3218 HeapWord* aligned_start = sp->bottom(); 3219 if (sp->used_region().contains(_restart_addr)) { 3220 // Align down to a card boundary for the start of 0th task 3221 // for this space. 3222 aligned_start = 3223 (HeapWord*)align_size_down((uintptr_t)_restart_addr, 3224 CardTable::card_size); 3225 } 3226 3227 size_t chunk_size = sp->marking_task_size(); 3228 while (!pst->is_task_claimed(/* reference */ nth_task)) { 3229 // Having claimed the nth task in this space, 3230 // compute the chunk that it corresponds to: 3231 MemRegion span = MemRegion(aligned_start + nth_task*chunk_size, 3232 aligned_start + (nth_task+1)*chunk_size); 3233 // Try and bump the global finger via a CAS; 3234 // note that we need to do the global finger bump 3235 // _before_ taking the intersection below, because 3236 // the task corresponding to that region will be 3237 // deemed done even if the used_region() expands 3238 // because of allocation -- as it almost certainly will 3239 // during start-up while the threads yield in the 3240 // closure below. 3241 HeapWord* finger = span.end(); 3242 bump_global_finger(finger); // atomically 3243 // There are null tasks here corresponding to chunks 3244 // beyond the "top" address of the space. 3245 span = span.intersection(sp->used_region()); 3246 if (!span.is_empty()) { // Non-null task 3247 HeapWord* prev_obj; 3248 assert(!span.contains(_restart_addr) || nth_task == 0, 3249 "Inconsistency"); 3250 if (nth_task == 0) { 3251 // For the 0th task, we'll not need to compute a block_start. 3252 if (span.contains(_restart_addr)) { 3253 // In the case of a restart because of stack overflow, 3254 // we might additionally skip a chunk prefix. 3255 prev_obj = _restart_addr; 3256 } else { 3257 prev_obj = span.start(); 3258 } 3259 } else { 3260 // We want to skip the first object because 3261 // the protocol is to scan any object in its entirety 3262 // that _starts_ in this span; a fortiori, any 3263 // object starting in an earlier span is scanned 3264 // as part of an earlier claimed task. 3265 // Below we use the "careful" version of block_start 3266 // so we do not try to navigate uninitialized objects. 3267 prev_obj = sp->block_start_careful(span.start()); 3268 // Below we use a variant of block_size that uses the 3269 // Printezis bits to avoid waiting for allocated 3270 // objects to become initialized/parsable. 3271 while (prev_obj < span.start()) { 3272 size_t sz = sp->block_size_no_stall(prev_obj, _collector); 3273 if (sz > 0) { 3274 prev_obj += sz; 3275 } else { 3276 // In this case we may end up doing a bit of redundant 3277 // scanning, but that appears unavoidable, short of 3278 // locking the free list locks; see bug 6324141. 3279 break; 3280 } 3281 } 3282 } 3283 if (prev_obj < span.end()) { 3284 MemRegion my_span = MemRegion(prev_obj, span.end()); 3285 // Do the marking work within a non-empty span -- 3286 // the last argument to the constructor indicates whether the 3287 // iteration should be incremental with periodic yields. 3288 ParMarkFromRootsClosure cl(this, _collector, my_span, 3289 &_collector->_markBitMap, 3290 work_queue(i), 3291 &_collector->_markStack); 3292 _collector->_markBitMap.iterate(&cl, my_span.start(), my_span.end()); 3293 } // else nothing to do for this task 3294 } // else nothing to do for this task 3295 } 3296 // We'd be tempted to assert here that since there are no 3297 // more tasks left to claim in this space, the global_finger 3298 // must exceed space->top() and a fortiori space->end(). However, 3299 // that would not quite be correct because the bumping of 3300 // global_finger occurs strictly after the claiming of a task, 3301 // so by the time we reach here the global finger may not yet 3302 // have been bumped up by the thread that claimed the last 3303 // task. 3304 pst->all_tasks_completed(); 3305 } 3306 3307 class ParConcMarkingClosure: public MetadataAwareOopClosure { 3308 private: 3309 CMSCollector* _collector; 3310 CMSConcMarkingTask* _task; 3311 MemRegion _span; 3312 CMSBitMap* _bit_map; 3313 CMSMarkStack* _overflow_stack; 3314 OopTaskQueue* _work_queue; 3315 protected: 3316 DO_OOP_WORK_DEFN 3317 public: 3318 ParConcMarkingClosure(CMSCollector* collector, CMSConcMarkingTask* task, OopTaskQueue* work_queue, 3319 CMSBitMap* bit_map, CMSMarkStack* overflow_stack): 3320 MetadataAwareOopClosure(collector->ref_processor()), 3321 _collector(collector), 3322 _task(task), 3323 _span(collector->_span), 3324 _work_queue(work_queue), 3325 _bit_map(bit_map), 3326 _overflow_stack(overflow_stack) 3327 { } 3328 virtual void do_oop(oop* p); 3329 virtual void do_oop(narrowOop* p); 3330 3331 void trim_queue(size_t max); 3332 void handle_stack_overflow(HeapWord* lost); 3333 void do_yield_check() { 3334 if (_task->should_yield()) { 3335 _task->yield(); 3336 } 3337 } 3338 }; 3339 3340 DO_OOP_WORK_IMPL(ParConcMarkingClosure) 3341 3342 // Grey object scanning during work stealing phase -- 3343 // the salient assumption here is that any references 3344 // that are in these stolen objects being scanned must 3345 // already have been initialized (else they would not have 3346 // been published), so we do not need to check for 3347 // uninitialized objects before pushing here. 3348 void ParConcMarkingClosure::do_oop(oop obj) { 3349 assert(obj->is_oop_or_null(true), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 3350 HeapWord* addr = (HeapWord*)obj; 3351 // Check if oop points into the CMS generation 3352 // and is not marked 3353 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 3354 // a white object ... 3355 // If we manage to "claim" the object, by being the 3356 // first thread to mark it, then we push it on our 3357 // marking stack 3358 if (_bit_map->par_mark(addr)) { // ... now grey 3359 // push on work queue (grey set) 3360 bool simulate_overflow = false; 3361 NOT_PRODUCT( 3362 if (CMSMarkStackOverflowALot && 3363 _collector->simulate_overflow()) { 3364 // simulate a stack overflow 3365 simulate_overflow = true; 3366 } 3367 ) 3368 if (simulate_overflow || 3369 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) { 3370 // stack overflow 3371 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _overflow_stack->capacity()); 3372 // We cannot assert that the overflow stack is full because 3373 // it may have been emptied since. 3374 assert(simulate_overflow || 3375 _work_queue->size() == _work_queue->max_elems(), 3376 "Else push should have succeeded"); 3377 handle_stack_overflow(addr); 3378 } 3379 } // Else, some other thread got there first 3380 do_yield_check(); 3381 } 3382 } 3383 3384 void ParConcMarkingClosure::do_oop(oop* p) { ParConcMarkingClosure::do_oop_work(p); } 3385 void ParConcMarkingClosure::do_oop(narrowOop* p) { ParConcMarkingClosure::do_oop_work(p); } 3386 3387 void ParConcMarkingClosure::trim_queue(size_t max) { 3388 while (_work_queue->size() > max) { 3389 oop new_oop; 3390 if (_work_queue->pop_local(new_oop)) { 3391 assert(new_oop->is_oop(), "Should be an oop"); 3392 assert(_bit_map->isMarked((HeapWord*)new_oop), "Grey object"); 3393 assert(_span.contains((HeapWord*)new_oop), "Not in span"); 3394 new_oop->oop_iterate(this); // do_oop() above 3395 do_yield_check(); 3396 } 3397 } 3398 } 3399 3400 // Upon stack overflow, we discard (part of) the stack, 3401 // remembering the least address amongst those discarded 3402 // in CMSCollector's _restart_address. 3403 void ParConcMarkingClosure::handle_stack_overflow(HeapWord* lost) { 3404 // We need to do this under a mutex to prevent other 3405 // workers from interfering with the work done below. 3406 MutexLockerEx ml(_overflow_stack->par_lock(), 3407 Mutex::_no_safepoint_check_flag); 3408 // Remember the least grey address discarded 3409 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 3410 _collector->lower_restart_addr(ra); 3411 _overflow_stack->reset(); // discard stack contents 3412 _overflow_stack->expand(); // expand the stack if possible 3413 } 3414 3415 3416 void CMSConcMarkingTask::do_work_steal(int i) { 3417 OopTaskQueue* work_q = work_queue(i); 3418 oop obj_to_scan; 3419 CMSBitMap* bm = &(_collector->_markBitMap); 3420 CMSMarkStack* ovflw = &(_collector->_markStack); 3421 int* seed = _collector->hash_seed(i); 3422 ParConcMarkingClosure cl(_collector, this, work_q, bm, ovflw); 3423 while (true) { 3424 cl.trim_queue(0); 3425 assert(work_q->size() == 0, "Should have been emptied above"); 3426 if (get_work_from_overflow_stack(ovflw, work_q)) { 3427 // Can't assert below because the work obtained from the 3428 // overflow stack may already have been stolen from us. 3429 // assert(work_q->size() > 0, "Work from overflow stack"); 3430 continue; 3431 } else if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 3432 assert(obj_to_scan->is_oop(), "Should be an oop"); 3433 assert(bm->isMarked((HeapWord*)obj_to_scan), "Grey object"); 3434 obj_to_scan->oop_iterate(&cl); 3435 } else if (terminator()->offer_termination(&_term_term)) { 3436 assert(work_q->size() == 0, "Impossible!"); 3437 break; 3438 } else if (yielding() || should_yield()) { 3439 yield(); 3440 } 3441 } 3442 } 3443 3444 // This is run by the CMS (coordinator) thread. 3445 void CMSConcMarkingTask::coordinator_yield() { 3446 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 3447 "CMS thread should hold CMS token"); 3448 // First give up the locks, then yield, then re-lock 3449 // We should probably use a constructor/destructor idiom to 3450 // do this unlock/lock or modify the MutexUnlocker class to 3451 // serve our purpose. XXX 3452 assert_lock_strong(_bit_map_lock); 3453 _bit_map_lock->unlock(); 3454 ConcurrentMarkSweepThread::desynchronize(true); 3455 _collector->stopTimer(); 3456 _collector->incrementYields(); 3457 3458 // It is possible for whichever thread initiated the yield request 3459 // not to get a chance to wake up and take the bitmap lock between 3460 // this thread releasing it and reacquiring it. So, while the 3461 // should_yield() flag is on, let's sleep for a bit to give the 3462 // other thread a chance to wake up. The limit imposed on the number 3463 // of iterations is defensive, to avoid any unforseen circumstances 3464 // putting us into an infinite loop. Since it's always been this 3465 // (coordinator_yield()) method that was observed to cause the 3466 // problem, we are using a parameter (CMSCoordinatorYieldSleepCount) 3467 // which is by default non-zero. For the other seven methods that 3468 // also perform the yield operation, as are using a different 3469 // parameter (CMSYieldSleepCount) which is by default zero. This way we 3470 // can enable the sleeping for those methods too, if necessary. 3471 // See 6442774. 3472 // 3473 // We really need to reconsider the synchronization between the GC 3474 // thread and the yield-requesting threads in the future and we 3475 // should really use wait/notify, which is the recommended 3476 // way of doing this type of interaction. Additionally, we should 3477 // consolidate the eight methods that do the yield operation and they 3478 // are almost identical into one for better maintainability and 3479 // readability. See 6445193. 3480 // 3481 // Tony 2006.06.29 3482 for (unsigned i = 0; i < CMSCoordinatorYieldSleepCount && 3483 ConcurrentMarkSweepThread::should_yield() && 3484 !CMSCollector::foregroundGCIsActive(); ++i) { 3485 os::sleep(Thread::current(), 1, false); 3486 } 3487 3488 ConcurrentMarkSweepThread::synchronize(true); 3489 _bit_map_lock->lock_without_safepoint_check(); 3490 _collector->startTimer(); 3491 } 3492 3493 bool CMSCollector::do_marking_mt() { 3494 assert(ConcGCThreads > 0 && conc_workers() != NULL, "precondition"); 3495 uint num_workers = AdaptiveSizePolicy::calc_active_conc_workers(conc_workers()->total_workers(), 3496 conc_workers()->active_workers(), 3497 Threads::number_of_non_daemon_threads()); 3498 num_workers = conc_workers()->update_active_workers(num_workers); 3499 log_info(gc,task)("Using %u workers of %u for marking", num_workers, conc_workers()->total_workers()); 3500 3501 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 3502 3503 CMSConcMarkingTask tsk(this, 3504 cms_space, 3505 conc_workers(), 3506 task_queues()); 3507 3508 // Since the actual number of workers we get may be different 3509 // from the number we requested above, do we need to do anything different 3510 // below? In particular, may be we need to subclass the SequantialSubTasksDone 3511 // class?? XXX 3512 cms_space ->initialize_sequential_subtasks_for_marking(num_workers); 3513 3514 // Refs discovery is already non-atomic. 3515 assert(!ref_processor()->discovery_is_atomic(), "Should be non-atomic"); 3516 assert(ref_processor()->discovery_is_mt(), "Discovery should be MT"); 3517 conc_workers()->start_task(&tsk); 3518 while (tsk.yielded()) { 3519 tsk.coordinator_yield(); 3520 conc_workers()->continue_task(&tsk); 3521 } 3522 // If the task was aborted, _restart_addr will be non-NULL 3523 assert(tsk.completed() || _restart_addr != NULL, "Inconsistency"); 3524 while (_restart_addr != NULL) { 3525 // XXX For now we do not make use of ABORTED state and have not 3526 // yet implemented the right abort semantics (even in the original 3527 // single-threaded CMS case). That needs some more investigation 3528 // and is deferred for now; see CR# TBF. 07252005YSR. XXX 3529 assert(!CMSAbortSemantics || tsk.aborted(), "Inconsistency"); 3530 // If _restart_addr is non-NULL, a marking stack overflow 3531 // occurred; we need to do a fresh marking iteration from the 3532 // indicated restart address. 3533 if (_foregroundGCIsActive) { 3534 // We may be running into repeated stack overflows, having 3535 // reached the limit of the stack size, while making very 3536 // slow forward progress. It may be best to bail out and 3537 // let the foreground collector do its job. 3538 // Clear _restart_addr, so that foreground GC 3539 // works from scratch. This avoids the headache of 3540 // a "rescan" which would otherwise be needed because 3541 // of the dirty mod union table & card table. 3542 _restart_addr = NULL; 3543 return false; 3544 } 3545 // Adjust the task to restart from _restart_addr 3546 tsk.reset(_restart_addr); 3547 cms_space ->initialize_sequential_subtasks_for_marking(num_workers, 3548 _restart_addr); 3549 _restart_addr = NULL; 3550 // Get the workers going again 3551 conc_workers()->start_task(&tsk); 3552 while (tsk.yielded()) { 3553 tsk.coordinator_yield(); 3554 conc_workers()->continue_task(&tsk); 3555 } 3556 } 3557 assert(tsk.completed(), "Inconsistency"); 3558 assert(tsk.result() == true, "Inconsistency"); 3559 return true; 3560 } 3561 3562 bool CMSCollector::do_marking_st() { 3563 ResourceMark rm; 3564 HandleMark hm; 3565 3566 // Temporarily make refs discovery single threaded (non-MT) 3567 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false); 3568 MarkFromRootsClosure markFromRootsClosure(this, _span, &_markBitMap, 3569 &_markStack, CMSYield); 3570 // the last argument to iterate indicates whether the iteration 3571 // should be incremental with periodic yields. 3572 _markBitMap.iterate(&markFromRootsClosure); 3573 // If _restart_addr is non-NULL, a marking stack overflow 3574 // occurred; we need to do a fresh iteration from the 3575 // indicated restart address. 3576 while (_restart_addr != NULL) { 3577 if (_foregroundGCIsActive) { 3578 // We may be running into repeated stack overflows, having 3579 // reached the limit of the stack size, while making very 3580 // slow forward progress. It may be best to bail out and 3581 // let the foreground collector do its job. 3582 // Clear _restart_addr, so that foreground GC 3583 // works from scratch. This avoids the headache of 3584 // a "rescan" which would otherwise be needed because 3585 // of the dirty mod union table & card table. 3586 _restart_addr = NULL; 3587 return false; // indicating failure to complete marking 3588 } 3589 // Deal with stack overflow: 3590 // we restart marking from _restart_addr 3591 HeapWord* ra = _restart_addr; 3592 markFromRootsClosure.reset(ra); 3593 _restart_addr = NULL; 3594 _markBitMap.iterate(&markFromRootsClosure, ra, _span.end()); 3595 } 3596 return true; 3597 } 3598 3599 void CMSCollector::preclean() { 3600 check_correct_thread_executing(); 3601 assert(Thread::current()->is_ConcurrentGC_thread(), "Wrong thread"); 3602 verify_work_stacks_empty(); 3603 verify_overflow_empty(); 3604 _abort_preclean = false; 3605 if (CMSPrecleaningEnabled) { 3606 if (!CMSEdenChunksRecordAlways) { 3607 _eden_chunk_index = 0; 3608 } 3609 size_t used = get_eden_used(); 3610 size_t capacity = get_eden_capacity(); 3611 // Don't start sampling unless we will get sufficiently 3612 // many samples. 3613 if (used < (((capacity / CMSScheduleRemarkSamplingRatio) / 100) 3614 * CMSScheduleRemarkEdenPenetration)) { 3615 _start_sampling = true; 3616 } else { 3617 _start_sampling = false; 3618 } 3619 GCTraceCPUTime tcpu; 3620 CMSPhaseAccounting pa(this, "Concurrent Preclean"); 3621 preclean_work(CMSPrecleanRefLists1, CMSPrecleanSurvivors1); 3622 } 3623 CMSTokenSync x(true); // is cms thread 3624 if (CMSPrecleaningEnabled) { 3625 sample_eden(); 3626 _collectorState = AbortablePreclean; 3627 } else { 3628 _collectorState = FinalMarking; 3629 } 3630 verify_work_stacks_empty(); 3631 verify_overflow_empty(); 3632 } 3633 3634 // Try and schedule the remark such that young gen 3635 // occupancy is CMSScheduleRemarkEdenPenetration %. 3636 void CMSCollector::abortable_preclean() { 3637 check_correct_thread_executing(); 3638 assert(CMSPrecleaningEnabled, "Inconsistent control state"); 3639 assert(_collectorState == AbortablePreclean, "Inconsistent control state"); 3640 3641 // If Eden's current occupancy is below this threshold, 3642 // immediately schedule the remark; else preclean 3643 // past the next scavenge in an effort to 3644 // schedule the pause as described above. By choosing 3645 // CMSScheduleRemarkEdenSizeThreshold >= max eden size 3646 // we will never do an actual abortable preclean cycle. 3647 if (get_eden_used() > CMSScheduleRemarkEdenSizeThreshold) { 3648 GCTraceCPUTime tcpu; 3649 CMSPhaseAccounting pa(this, "Concurrent Abortable Preclean"); 3650 // We need more smarts in the abortable preclean 3651 // loop below to deal with cases where allocation 3652 // in young gen is very very slow, and our precleaning 3653 // is running a losing race against a horde of 3654 // mutators intent on flooding us with CMS updates 3655 // (dirty cards). 3656 // One, admittedly dumb, strategy is to give up 3657 // after a certain number of abortable precleaning loops 3658 // or after a certain maximum time. We want to make 3659 // this smarter in the next iteration. 3660 // XXX FIX ME!!! YSR 3661 size_t loops = 0, workdone = 0, cumworkdone = 0, waited = 0; 3662 while (!(should_abort_preclean() || 3663 ConcurrentMarkSweepThread::cmst()->should_terminate())) { 3664 workdone = preclean_work(CMSPrecleanRefLists2, CMSPrecleanSurvivors2); 3665 cumworkdone += workdone; 3666 loops++; 3667 // Voluntarily terminate abortable preclean phase if we have 3668 // been at it for too long. 3669 if ((CMSMaxAbortablePrecleanLoops != 0) && 3670 loops >= CMSMaxAbortablePrecleanLoops) { 3671 log_debug(gc)(" CMS: abort preclean due to loops "); 3672 break; 3673 } 3674 if (pa.wallclock_millis() > CMSMaxAbortablePrecleanTime) { 3675 log_debug(gc)(" CMS: abort preclean due to time "); 3676 break; 3677 } 3678 // If we are doing little work each iteration, we should 3679 // take a short break. 3680 if (workdone < CMSAbortablePrecleanMinWorkPerIteration) { 3681 // Sleep for some time, waiting for work to accumulate 3682 stopTimer(); 3683 cmsThread()->wait_on_cms_lock(CMSAbortablePrecleanWaitMillis); 3684 startTimer(); 3685 waited++; 3686 } 3687 } 3688 log_trace(gc)(" [" SIZE_FORMAT " iterations, " SIZE_FORMAT " waits, " SIZE_FORMAT " cards)] ", 3689 loops, waited, cumworkdone); 3690 } 3691 CMSTokenSync x(true); // is cms thread 3692 if (_collectorState != Idling) { 3693 assert(_collectorState == AbortablePreclean, 3694 "Spontaneous state transition?"); 3695 _collectorState = FinalMarking; 3696 } // Else, a foreground collection completed this CMS cycle. 3697 return; 3698 } 3699 3700 // Respond to an Eden sampling opportunity 3701 void CMSCollector::sample_eden() { 3702 // Make sure a young gc cannot sneak in between our 3703 // reading and recording of a sample. 3704 assert(Thread::current()->is_ConcurrentGC_thread(), 3705 "Only the cms thread may collect Eden samples"); 3706 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 3707 "Should collect samples while holding CMS token"); 3708 if (!_start_sampling) { 3709 return; 3710 } 3711 // When CMSEdenChunksRecordAlways is true, the eden chunk array 3712 // is populated by the young generation. 3713 if (_eden_chunk_array != NULL && !CMSEdenChunksRecordAlways) { 3714 if (_eden_chunk_index < _eden_chunk_capacity) { 3715 _eden_chunk_array[_eden_chunk_index] = *_top_addr; // take sample 3716 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr, 3717 "Unexpected state of Eden"); 3718 // We'd like to check that what we just sampled is an oop-start address; 3719 // however, we cannot do that here since the object may not yet have been 3720 // initialized. So we'll instead do the check when we _use_ this sample 3721 // later. 3722 if (_eden_chunk_index == 0 || 3723 (pointer_delta(_eden_chunk_array[_eden_chunk_index], 3724 _eden_chunk_array[_eden_chunk_index-1]) 3725 >= CMSSamplingGrain)) { 3726 _eden_chunk_index++; // commit sample 3727 } 3728 } 3729 } 3730 if ((_collectorState == AbortablePreclean) && !_abort_preclean) { 3731 size_t used = get_eden_used(); 3732 size_t capacity = get_eden_capacity(); 3733 assert(used <= capacity, "Unexpected state of Eden"); 3734 if (used > (capacity/100 * CMSScheduleRemarkEdenPenetration)) { 3735 _abort_preclean = true; 3736 } 3737 } 3738 } 3739 3740 3741 size_t CMSCollector::preclean_work(bool clean_refs, bool clean_survivor) { 3742 assert(_collectorState == Precleaning || 3743 _collectorState == AbortablePreclean, "incorrect state"); 3744 ResourceMark rm; 3745 HandleMark hm; 3746 3747 // Precleaning is currently not MT but the reference processor 3748 // may be set for MT. Disable it temporarily here. 3749 ReferenceProcessor* rp = ref_processor(); 3750 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(rp, false); 3751 3752 // Do one pass of scrubbing the discovered reference lists 3753 // to remove any reference objects with strongly-reachable 3754 // referents. 3755 if (clean_refs) { 3756 CMSPrecleanRefsYieldClosure yield_cl(this); 3757 assert(rp->span().equals(_span), "Spans should be equal"); 3758 CMSKeepAliveClosure keep_alive(this, _span, &_markBitMap, 3759 &_markStack, true /* preclean */); 3760 CMSDrainMarkingStackClosure complete_trace(this, 3761 _span, &_markBitMap, &_markStack, 3762 &keep_alive, true /* preclean */); 3763 3764 // We don't want this step to interfere with a young 3765 // collection because we don't want to take CPU 3766 // or memory bandwidth away from the young GC threads 3767 // (which may be as many as there are CPUs). 3768 // Note that we don't need to protect ourselves from 3769 // interference with mutators because they can't 3770 // manipulate the discovered reference lists nor affect 3771 // the computed reachability of the referents, the 3772 // only properties manipulated by the precleaning 3773 // of these reference lists. 3774 stopTimer(); 3775 CMSTokenSyncWithLocks x(true /* is cms thread */, 3776 bitMapLock()); 3777 startTimer(); 3778 sample_eden(); 3779 3780 // The following will yield to allow foreground 3781 // collection to proceed promptly. XXX YSR: 3782 // The code in this method may need further 3783 // tweaking for better performance and some restructuring 3784 // for cleaner interfaces. 3785 GCTimer *gc_timer = NULL; // Currently not tracing concurrent phases 3786 rp->preclean_discovered_references( 3787 rp->is_alive_non_header(), &keep_alive, &complete_trace, &yield_cl, 3788 gc_timer); 3789 } 3790 3791 if (clean_survivor) { // preclean the active survivor space(s) 3792 PushAndMarkClosure pam_cl(this, _span, ref_processor(), 3793 &_markBitMap, &_modUnionTable, 3794 &_markStack, true /* precleaning phase */); 3795 stopTimer(); 3796 CMSTokenSyncWithLocks ts(true /* is cms thread */, 3797 bitMapLock()); 3798 startTimer(); 3799 unsigned int before_count = 3800 GenCollectedHeap::heap()->total_collections(); 3801 SurvivorSpacePrecleanClosure 3802 sss_cl(this, _span, &_markBitMap, &_markStack, 3803 &pam_cl, before_count, CMSYield); 3804 _young_gen->from()->object_iterate_careful(&sss_cl); 3805 _young_gen->to()->object_iterate_careful(&sss_cl); 3806 } 3807 MarkRefsIntoAndScanClosure 3808 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable, 3809 &_markStack, this, CMSYield, 3810 true /* precleaning phase */); 3811 // CAUTION: The following closure has persistent state that may need to 3812 // be reset upon a decrease in the sequence of addresses it 3813 // processes. 3814 ScanMarkedObjectsAgainCarefullyClosure 3815 smoac_cl(this, _span, 3816 &_markBitMap, &_markStack, &mrias_cl, CMSYield); 3817 3818 // Preclean dirty cards in ModUnionTable and CardTable using 3819 // appropriate convergence criterion; 3820 // repeat CMSPrecleanIter times unless we find that 3821 // we are losing. 3822 assert(CMSPrecleanIter < 10, "CMSPrecleanIter is too large"); 3823 assert(CMSPrecleanNumerator < CMSPrecleanDenominator, 3824 "Bad convergence multiplier"); 3825 assert(CMSPrecleanThreshold >= 100, 3826 "Unreasonably low CMSPrecleanThreshold"); 3827 3828 size_t numIter, cumNumCards, lastNumCards, curNumCards; 3829 for (numIter = 0, cumNumCards = lastNumCards = curNumCards = 0; 3830 numIter < CMSPrecleanIter; 3831 numIter++, lastNumCards = curNumCards, cumNumCards += curNumCards) { 3832 curNumCards = preclean_mod_union_table(_cmsGen, &smoac_cl); 3833 log_trace(gc)(" (modUnionTable: " SIZE_FORMAT " cards)", curNumCards); 3834 // Either there are very few dirty cards, so re-mark 3835 // pause will be small anyway, or our pre-cleaning isn't 3836 // that much faster than the rate at which cards are being 3837 // dirtied, so we might as well stop and re-mark since 3838 // precleaning won't improve our re-mark time by much. 3839 if (curNumCards <= CMSPrecleanThreshold || 3840 (numIter > 0 && 3841 (curNumCards * CMSPrecleanDenominator > 3842 lastNumCards * CMSPrecleanNumerator))) { 3843 numIter++; 3844 cumNumCards += curNumCards; 3845 break; 3846 } 3847 } 3848 3849 preclean_klasses(&mrias_cl, _cmsGen->freelistLock()); 3850 3851 curNumCards = preclean_card_table(_cmsGen, &smoac_cl); 3852 cumNumCards += curNumCards; 3853 log_trace(gc)(" (cardTable: " SIZE_FORMAT " cards, re-scanned " SIZE_FORMAT " cards, " SIZE_FORMAT " iterations)", 3854 curNumCards, cumNumCards, numIter); 3855 return cumNumCards; // as a measure of useful work done 3856 } 3857 3858 // PRECLEANING NOTES: 3859 // Precleaning involves: 3860 // . reading the bits of the modUnionTable and clearing the set bits. 3861 // . For the cards corresponding to the set bits, we scan the 3862 // objects on those cards. This means we need the free_list_lock 3863 // so that we can safely iterate over the CMS space when scanning 3864 // for oops. 3865 // . When we scan the objects, we'll be both reading and setting 3866 // marks in the marking bit map, so we'll need the marking bit map. 3867 // . For protecting _collector_state transitions, we take the CGC_lock. 3868 // Note that any races in the reading of of card table entries by the 3869 // CMS thread on the one hand and the clearing of those entries by the 3870 // VM thread or the setting of those entries by the mutator threads on the 3871 // other are quite benign. However, for efficiency it makes sense to keep 3872 // the VM thread from racing with the CMS thread while the latter is 3873 // dirty card info to the modUnionTable. We therefore also use the 3874 // CGC_lock to protect the reading of the card table and the mod union 3875 // table by the CM thread. 3876 // . We run concurrently with mutator updates, so scanning 3877 // needs to be done carefully -- we should not try to scan 3878 // potentially uninitialized objects. 3879 // 3880 // Locking strategy: While holding the CGC_lock, we scan over and 3881 // reset a maximal dirty range of the mod union / card tables, then lock 3882 // the free_list_lock and bitmap lock to do a full marking, then 3883 // release these locks; and repeat the cycle. This allows for a 3884 // certain amount of fairness in the sharing of these locks between 3885 // the CMS collector on the one hand, and the VM thread and the 3886 // mutators on the other. 3887 3888 // NOTE: preclean_mod_union_table() and preclean_card_table() 3889 // further below are largely identical; if you need to modify 3890 // one of these methods, please check the other method too. 3891 3892 size_t CMSCollector::preclean_mod_union_table( 3893 ConcurrentMarkSweepGeneration* old_gen, 3894 ScanMarkedObjectsAgainCarefullyClosure* cl) { 3895 verify_work_stacks_empty(); 3896 verify_overflow_empty(); 3897 3898 // strategy: starting with the first card, accumulate contiguous 3899 // ranges of dirty cards; clear these cards, then scan the region 3900 // covered by these cards. 3901 3902 // Since all of the MUT is committed ahead, we can just use 3903 // that, in case the generations expand while we are precleaning. 3904 // It might also be fine to just use the committed part of the 3905 // generation, but we might potentially miss cards when the 3906 // generation is rapidly expanding while we are in the midst 3907 // of precleaning. 3908 HeapWord* startAddr = old_gen->reserved().start(); 3909 HeapWord* endAddr = old_gen->reserved().end(); 3910 3911 cl->setFreelistLock(old_gen->freelistLock()); // needed for yielding 3912 3913 size_t numDirtyCards, cumNumDirtyCards; 3914 HeapWord *nextAddr, *lastAddr; 3915 for (cumNumDirtyCards = numDirtyCards = 0, 3916 nextAddr = lastAddr = startAddr; 3917 nextAddr < endAddr; 3918 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 3919 3920 ResourceMark rm; 3921 HandleMark hm; 3922 3923 MemRegion dirtyRegion; 3924 { 3925 stopTimer(); 3926 // Potential yield point 3927 CMSTokenSync ts(true); 3928 startTimer(); 3929 sample_eden(); 3930 // Get dirty region starting at nextOffset (inclusive), 3931 // simultaneously clearing it. 3932 dirtyRegion = 3933 _modUnionTable.getAndClearMarkedRegion(nextAddr, endAddr); 3934 assert(dirtyRegion.start() >= nextAddr, 3935 "returned region inconsistent?"); 3936 } 3937 // Remember where the next search should begin. 3938 // The returned region (if non-empty) is a right open interval, 3939 // so lastOffset is obtained from the right end of that 3940 // interval. 3941 lastAddr = dirtyRegion.end(); 3942 // Should do something more transparent and less hacky XXX 3943 numDirtyCards = 3944 _modUnionTable.heapWordDiffToOffsetDiff(dirtyRegion.word_size()); 3945 3946 // We'll scan the cards in the dirty region (with periodic 3947 // yields for foreground GC as needed). 3948 if (!dirtyRegion.is_empty()) { 3949 assert(numDirtyCards > 0, "consistency check"); 3950 HeapWord* stop_point = NULL; 3951 stopTimer(); 3952 // Potential yield point 3953 CMSTokenSyncWithLocks ts(true, old_gen->freelistLock(), 3954 bitMapLock()); 3955 startTimer(); 3956 { 3957 verify_work_stacks_empty(); 3958 verify_overflow_empty(); 3959 sample_eden(); 3960 stop_point = 3961 old_gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 3962 } 3963 if (stop_point != NULL) { 3964 // The careful iteration stopped early either because it found an 3965 // uninitialized object, or because we were in the midst of an 3966 // "abortable preclean", which should now be aborted. Redirty 3967 // the bits corresponding to the partially-scanned or unscanned 3968 // cards. We'll either restart at the next block boundary or 3969 // abort the preclean. 3970 assert((_collectorState == AbortablePreclean && should_abort_preclean()), 3971 "Should only be AbortablePreclean."); 3972 _modUnionTable.mark_range(MemRegion(stop_point, dirtyRegion.end())); 3973 if (should_abort_preclean()) { 3974 break; // out of preclean loop 3975 } else { 3976 // Compute the next address at which preclean should pick up; 3977 // might need bitMapLock in order to read P-bits. 3978 lastAddr = next_card_start_after_block(stop_point); 3979 } 3980 } 3981 } else { 3982 assert(lastAddr == endAddr, "consistency check"); 3983 assert(numDirtyCards == 0, "consistency check"); 3984 break; 3985 } 3986 } 3987 verify_work_stacks_empty(); 3988 verify_overflow_empty(); 3989 return cumNumDirtyCards; 3990 } 3991 3992 // NOTE: preclean_mod_union_table() above and preclean_card_table() 3993 // below are largely identical; if you need to modify 3994 // one of these methods, please check the other method too. 3995 3996 size_t CMSCollector::preclean_card_table(ConcurrentMarkSweepGeneration* old_gen, 3997 ScanMarkedObjectsAgainCarefullyClosure* cl) { 3998 // strategy: it's similar to precleamModUnionTable above, in that 3999 // we accumulate contiguous ranges of dirty cards, mark these cards 4000 // precleaned, then scan the region covered by these cards. 4001 HeapWord* endAddr = (HeapWord*)(old_gen->_virtual_space.high()); 4002 HeapWord* startAddr = (HeapWord*)(old_gen->_virtual_space.low()); 4003 4004 cl->setFreelistLock(old_gen->freelistLock()); // needed for yielding 4005 4006 size_t numDirtyCards, cumNumDirtyCards; 4007 HeapWord *lastAddr, *nextAddr; 4008 4009 for (cumNumDirtyCards = numDirtyCards = 0, 4010 nextAddr = lastAddr = startAddr; 4011 nextAddr < endAddr; 4012 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 4013 4014 ResourceMark rm; 4015 HandleMark hm; 4016 4017 MemRegion dirtyRegion; 4018 { 4019 // See comments in "Precleaning notes" above on why we 4020 // do this locking. XXX Could the locking overheads be 4021 // too high when dirty cards are sparse? [I don't think so.] 4022 stopTimer(); 4023 CMSTokenSync x(true); // is cms thread 4024 startTimer(); 4025 sample_eden(); 4026 // Get and clear dirty region from card table 4027 dirtyRegion = _ct->dirty_card_range_after_reset(MemRegion(nextAddr, endAddr), 4028 true, 4029 CardTable::precleaned_card_val()); 4030 4031 assert(dirtyRegion.start() >= nextAddr, 4032 "returned region inconsistent?"); 4033 } 4034 lastAddr = dirtyRegion.end(); 4035 numDirtyCards = 4036 dirtyRegion.word_size()/CardTable::card_size_in_words; 4037 4038 if (!dirtyRegion.is_empty()) { 4039 stopTimer(); 4040 CMSTokenSyncWithLocks ts(true, old_gen->freelistLock(), bitMapLock()); 4041 startTimer(); 4042 sample_eden(); 4043 verify_work_stacks_empty(); 4044 verify_overflow_empty(); 4045 HeapWord* stop_point = 4046 old_gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 4047 if (stop_point != NULL) { 4048 assert((_collectorState == AbortablePreclean && should_abort_preclean()), 4049 "Should only be AbortablePreclean."); 4050 _ct->invalidate(MemRegion(stop_point, dirtyRegion.end())); 4051 if (should_abort_preclean()) { 4052 break; // out of preclean loop 4053 } else { 4054 // Compute the next address at which preclean should pick up. 4055 lastAddr = next_card_start_after_block(stop_point); 4056 } 4057 } 4058 } else { 4059 break; 4060 } 4061 } 4062 verify_work_stacks_empty(); 4063 verify_overflow_empty(); 4064 return cumNumDirtyCards; 4065 } 4066 4067 class PrecleanKlassClosure : public KlassClosure { 4068 KlassToOopClosure _cm_klass_closure; 4069 public: 4070 PrecleanKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {} 4071 void do_klass(Klass* k) { 4072 if (k->has_accumulated_modified_oops()) { 4073 k->clear_accumulated_modified_oops(); 4074 4075 _cm_klass_closure.do_klass(k); 4076 } 4077 } 4078 }; 4079 4080 // The freelist lock is needed to prevent asserts, is it really needed? 4081 void CMSCollector::preclean_klasses(MarkRefsIntoAndScanClosure* cl, Mutex* freelistLock) { 4082 4083 cl->set_freelistLock(freelistLock); 4084 4085 CMSTokenSyncWithLocks ts(true, freelistLock, bitMapLock()); 4086 4087 // SSS: Add equivalent to ScanMarkedObjectsAgainCarefullyClosure::do_yield_check and should_abort_preclean? 4088 // SSS: We should probably check if precleaning should be aborted, at suitable intervals? 4089 PrecleanKlassClosure preclean_klass_closure(cl); 4090 ClassLoaderDataGraph::classes_do(&preclean_klass_closure); 4091 4092 verify_work_stacks_empty(); 4093 verify_overflow_empty(); 4094 } 4095 4096 void CMSCollector::checkpointRootsFinal() { 4097 assert(_collectorState == FinalMarking, "incorrect state transition?"); 4098 check_correct_thread_executing(); 4099 // world is stopped at this checkpoint 4100 assert(SafepointSynchronize::is_at_safepoint(), 4101 "world should be stopped"); 4102 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 4103 4104 verify_work_stacks_empty(); 4105 verify_overflow_empty(); 4106 4107 log_debug(gc)("YG occupancy: " SIZE_FORMAT " K (" SIZE_FORMAT " K)", 4108 _young_gen->used() / K, _young_gen->capacity() / K); 4109 { 4110 if (CMSScavengeBeforeRemark) { 4111 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4112 // Temporarily set flag to false, GCH->do_collection will 4113 // expect it to be false and set to true 4114 FlagSetting fl(gch->_is_gc_active, false); 4115 4116 gch->do_collection(true, // full (i.e. force, see below) 4117 false, // !clear_all_soft_refs 4118 0, // size 4119 false, // is_tlab 4120 GenCollectedHeap::YoungGen // type 4121 ); 4122 } 4123 FreelistLocker x(this); 4124 MutexLockerEx y(bitMapLock(), 4125 Mutex::_no_safepoint_check_flag); 4126 checkpointRootsFinalWork(); 4127 } 4128 verify_work_stacks_empty(); 4129 verify_overflow_empty(); 4130 } 4131 4132 void CMSCollector::checkpointRootsFinalWork() { 4133 GCTraceTime(Trace, gc, phases) tm("checkpointRootsFinalWork", _gc_timer_cm); 4134 4135 assert(haveFreelistLocks(), "must have free list locks"); 4136 assert_lock_strong(bitMapLock()); 4137 4138 ResourceMark rm; 4139 HandleMark hm; 4140 4141 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4142 4143 if (should_unload_classes()) { 4144 CodeCache::gc_prologue(); 4145 } 4146 assert(haveFreelistLocks(), "must have free list locks"); 4147 assert_lock_strong(bitMapLock()); 4148 4149 // We might assume that we need not fill TLAB's when 4150 // CMSScavengeBeforeRemark is set, because we may have just done 4151 // a scavenge which would have filled all TLAB's -- and besides 4152 // Eden would be empty. This however may not always be the case -- 4153 // for instance although we asked for a scavenge, it may not have 4154 // happened because of a JNI critical section. We probably need 4155 // a policy for deciding whether we can in that case wait until 4156 // the critical section releases and then do the remark following 4157 // the scavenge, and skip it here. In the absence of that policy, 4158 // or of an indication of whether the scavenge did indeed occur, 4159 // we cannot rely on TLAB's having been filled and must do 4160 // so here just in case a scavenge did not happen. 4161 gch->ensure_parsability(false); // fill TLAB's, but no need to retire them 4162 // Update the saved marks which may affect the root scans. 4163 gch->save_marks(); 4164 4165 print_eden_and_survivor_chunk_arrays(); 4166 4167 { 4168 #if defined(COMPILER2) || INCLUDE_JVMCI 4169 DerivedPointerTableDeactivate dpt_deact; 4170 #endif 4171 4172 // Note on the role of the mod union table: 4173 // Since the marker in "markFromRoots" marks concurrently with 4174 // mutators, it is possible for some reachable objects not to have been 4175 // scanned. For instance, an only reference to an object A was 4176 // placed in object B after the marker scanned B. Unless B is rescanned, 4177 // A would be collected. Such updates to references in marked objects 4178 // are detected via the mod union table which is the set of all cards 4179 // dirtied since the first checkpoint in this GC cycle and prior to 4180 // the most recent young generation GC, minus those cleaned up by the 4181 // concurrent precleaning. 4182 if (CMSParallelRemarkEnabled) { 4183 GCTraceTime(Debug, gc, phases) t("Rescan (parallel)", _gc_timer_cm); 4184 do_remark_parallel(); 4185 } else { 4186 GCTraceTime(Debug, gc, phases) t("Rescan (non-parallel)", _gc_timer_cm); 4187 do_remark_non_parallel(); 4188 } 4189 } 4190 verify_work_stacks_empty(); 4191 verify_overflow_empty(); 4192 4193 { 4194 GCTraceTime(Trace, gc, phases) ts("refProcessingWork", _gc_timer_cm); 4195 refProcessingWork(); 4196 } 4197 verify_work_stacks_empty(); 4198 verify_overflow_empty(); 4199 4200 if (should_unload_classes()) { 4201 CodeCache::gc_epilogue(); 4202 } 4203 JvmtiExport::gc_epilogue(); 4204 4205 // If we encountered any (marking stack / work queue) overflow 4206 // events during the current CMS cycle, take appropriate 4207 // remedial measures, where possible, so as to try and avoid 4208 // recurrence of that condition. 4209 assert(_markStack.isEmpty(), "No grey objects"); 4210 size_t ser_ovflw = _ser_pmc_remark_ovflw + _ser_pmc_preclean_ovflw + 4211 _ser_kac_ovflw + _ser_kac_preclean_ovflw; 4212 if (ser_ovflw > 0) { 4213 log_trace(gc)("Marking stack overflow (benign) (pmc_pc=" SIZE_FORMAT ", pmc_rm=" SIZE_FORMAT ", kac=" SIZE_FORMAT ", kac_preclean=" SIZE_FORMAT ")", 4214 _ser_pmc_preclean_ovflw, _ser_pmc_remark_ovflw, _ser_kac_ovflw, _ser_kac_preclean_ovflw); 4215 _markStack.expand(); 4216 _ser_pmc_remark_ovflw = 0; 4217 _ser_pmc_preclean_ovflw = 0; 4218 _ser_kac_preclean_ovflw = 0; 4219 _ser_kac_ovflw = 0; 4220 } 4221 if (_par_pmc_remark_ovflw > 0 || _par_kac_ovflw > 0) { 4222 log_trace(gc)("Work queue overflow (benign) (pmc_rm=" SIZE_FORMAT ", kac=" SIZE_FORMAT ")", 4223 _par_pmc_remark_ovflw, _par_kac_ovflw); 4224 _par_pmc_remark_ovflw = 0; 4225 _par_kac_ovflw = 0; 4226 } 4227 if (_markStack._hit_limit > 0) { 4228 log_trace(gc)(" (benign) Hit max stack size limit (" SIZE_FORMAT ")", 4229 _markStack._hit_limit); 4230 } 4231 if (_markStack._failed_double > 0) { 4232 log_trace(gc)(" (benign) Failed stack doubling (" SIZE_FORMAT "), current capacity " SIZE_FORMAT, 4233 _markStack._failed_double, _markStack.capacity()); 4234 } 4235 _markStack._hit_limit = 0; 4236 _markStack._failed_double = 0; 4237 4238 if ((VerifyAfterGC || VerifyDuringGC) && 4239 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 4240 verify_after_remark(); 4241 } 4242 4243 _gc_tracer_cm->report_object_count_after_gc(&_is_alive_closure); 4244 4245 // Change under the freelistLocks. 4246 _collectorState = Sweeping; 4247 // Call isAllClear() under bitMapLock 4248 assert(_modUnionTable.isAllClear(), 4249 "Should be clear by end of the final marking"); 4250 assert(_ct->klass_rem_set()->mod_union_is_clear(), 4251 "Should be clear by end of the final marking"); 4252 } 4253 4254 void CMSParInitialMarkTask::work(uint worker_id) { 4255 elapsedTimer _timer; 4256 ResourceMark rm; 4257 HandleMark hm; 4258 4259 // ---------- scan from roots -------------- 4260 _timer.start(); 4261 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4262 ParMarkRefsIntoClosure par_mri_cl(_collector->_span, &(_collector->_markBitMap)); 4263 4264 // ---------- young gen roots -------------- 4265 { 4266 work_on_young_gen_roots(&par_mri_cl); 4267 _timer.stop(); 4268 log_trace(gc, task)("Finished young gen initial mark scan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4269 } 4270 4271 // ---------- remaining roots -------------- 4272 _timer.reset(); 4273 _timer.start(); 4274 4275 CLDToOopClosure cld_closure(&par_mri_cl, true); 4276 4277 gch->cms_process_roots(_strong_roots_scope, 4278 false, // yg was scanned above 4279 GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()), 4280 _collector->should_unload_classes(), 4281 &par_mri_cl, 4282 &cld_closure); 4283 assert(_collector->should_unload_classes() 4284 || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache), 4285 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 4286 _timer.stop(); 4287 log_trace(gc, task)("Finished remaining root initial mark scan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4288 } 4289 4290 // Parallel remark task 4291 class CMSParRemarkTask: public CMSParMarkTask { 4292 CompactibleFreeListSpace* _cms_space; 4293 4294 // The per-thread work queues, available here for stealing. 4295 OopTaskQueueSet* _task_queues; 4296 ParallelTaskTerminator _term; 4297 StrongRootsScope* _strong_roots_scope; 4298 4299 public: 4300 // A value of 0 passed to n_workers will cause the number of 4301 // workers to be taken from the active workers in the work gang. 4302 CMSParRemarkTask(CMSCollector* collector, 4303 CompactibleFreeListSpace* cms_space, 4304 uint n_workers, WorkGang* workers, 4305 OopTaskQueueSet* task_queues, 4306 StrongRootsScope* strong_roots_scope): 4307 CMSParMarkTask("Rescan roots and grey objects in parallel", 4308 collector, n_workers), 4309 _cms_space(cms_space), 4310 _task_queues(task_queues), 4311 _term(n_workers, task_queues), 4312 _strong_roots_scope(strong_roots_scope) { } 4313 4314 OopTaskQueueSet* task_queues() { return _task_queues; } 4315 4316 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 4317 4318 ParallelTaskTerminator* terminator() { return &_term; } 4319 uint n_workers() { return _n_workers; } 4320 4321 void work(uint worker_id); 4322 4323 private: 4324 // ... of dirty cards in old space 4325 void do_dirty_card_rescan_tasks(CompactibleFreeListSpace* sp, int i, 4326 ParMarkRefsIntoAndScanClosure* cl); 4327 4328 // ... work stealing for the above 4329 void do_work_steal(int i, ParMarkRefsIntoAndScanClosure* cl, int* seed); 4330 }; 4331 4332 class RemarkKlassClosure : public KlassClosure { 4333 KlassToOopClosure _cm_klass_closure; 4334 public: 4335 RemarkKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {} 4336 void do_klass(Klass* k) { 4337 // Check if we have modified any oops in the Klass during the concurrent marking. 4338 if (k->has_accumulated_modified_oops()) { 4339 k->clear_accumulated_modified_oops(); 4340 4341 // We could have transfered the current modified marks to the accumulated marks, 4342 // like we do with the Card Table to Mod Union Table. But it's not really necessary. 4343 } else if (k->has_modified_oops()) { 4344 // Don't clear anything, this info is needed by the next young collection. 4345 } else { 4346 // No modified oops in the Klass. 4347 return; 4348 } 4349 4350 // The klass has modified fields, need to scan the klass. 4351 _cm_klass_closure.do_klass(k); 4352 } 4353 }; 4354 4355 void CMSParMarkTask::work_on_young_gen_roots(OopsInGenClosure* cl) { 4356 ParNewGeneration* young_gen = _collector->_young_gen; 4357 ContiguousSpace* eden_space = young_gen->eden(); 4358 ContiguousSpace* from_space = young_gen->from(); 4359 ContiguousSpace* to_space = young_gen->to(); 4360 4361 HeapWord** eca = _collector->_eden_chunk_array; 4362 size_t ect = _collector->_eden_chunk_index; 4363 HeapWord** sca = _collector->_survivor_chunk_array; 4364 size_t sct = _collector->_survivor_chunk_index; 4365 4366 assert(ect <= _collector->_eden_chunk_capacity, "out of bounds"); 4367 assert(sct <= _collector->_survivor_chunk_capacity, "out of bounds"); 4368 4369 do_young_space_rescan(cl, to_space, NULL, 0); 4370 do_young_space_rescan(cl, from_space, sca, sct); 4371 do_young_space_rescan(cl, eden_space, eca, ect); 4372 } 4373 4374 // work_queue(i) is passed to the closure 4375 // ParMarkRefsIntoAndScanClosure. The "i" parameter 4376 // also is passed to do_dirty_card_rescan_tasks() and to 4377 // do_work_steal() to select the i-th task_queue. 4378 4379 void CMSParRemarkTask::work(uint worker_id) { 4380 elapsedTimer _timer; 4381 ResourceMark rm; 4382 HandleMark hm; 4383 4384 // ---------- rescan from roots -------------- 4385 _timer.start(); 4386 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4387 ParMarkRefsIntoAndScanClosure par_mrias_cl(_collector, 4388 _collector->_span, _collector->ref_processor(), 4389 &(_collector->_markBitMap), 4390 work_queue(worker_id)); 4391 4392 // Rescan young gen roots first since these are likely 4393 // coarsely partitioned and may, on that account, constitute 4394 // the critical path; thus, it's best to start off that 4395 // work first. 4396 // ---------- young gen roots -------------- 4397 { 4398 work_on_young_gen_roots(&par_mrias_cl); 4399 _timer.stop(); 4400 log_trace(gc, task)("Finished young gen rescan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4401 } 4402 4403 // ---------- remaining roots -------------- 4404 _timer.reset(); 4405 _timer.start(); 4406 gch->cms_process_roots(_strong_roots_scope, 4407 false, // yg was scanned above 4408 GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()), 4409 _collector->should_unload_classes(), 4410 &par_mrias_cl, 4411 NULL); // The dirty klasses will be handled below 4412 4413 assert(_collector->should_unload_classes() 4414 || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache), 4415 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 4416 _timer.stop(); 4417 log_trace(gc, task)("Finished remaining root rescan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4418 4419 // ---------- unhandled CLD scanning ---------- 4420 if (worker_id == 0) { // Single threaded at the moment. 4421 _timer.reset(); 4422 _timer.start(); 4423 4424 // Scan all new class loader data objects and new dependencies that were 4425 // introduced during concurrent marking. 4426 ResourceMark rm; 4427 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds(); 4428 for (int i = 0; i < array->length(); i++) { 4429 par_mrias_cl.do_cld_nv(array->at(i)); 4430 } 4431 4432 // We don't need to keep track of new CLDs anymore. 4433 ClassLoaderDataGraph::remember_new_clds(false); 4434 4435 _timer.stop(); 4436 log_trace(gc, task)("Finished unhandled CLD scanning work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4437 } 4438 4439 // ---------- dirty klass scanning ---------- 4440 if (worker_id == 0) { // Single threaded at the moment. 4441 _timer.reset(); 4442 _timer.start(); 4443 4444 // Scan all classes that was dirtied during the concurrent marking phase. 4445 RemarkKlassClosure remark_klass_closure(&par_mrias_cl); 4446 ClassLoaderDataGraph::classes_do(&remark_klass_closure); 4447 4448 _timer.stop(); 4449 log_trace(gc, task)("Finished dirty klass scanning work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4450 } 4451 4452 // We might have added oops to ClassLoaderData::_handles during the 4453 // concurrent marking phase. These oops point to newly allocated objects 4454 // that are guaranteed to be kept alive. Either by the direct allocation 4455 // code, or when the young collector processes the roots. Hence, 4456 // we don't have to revisit the _handles block during the remark phase. 4457 4458 // ---------- rescan dirty cards ------------ 4459 _timer.reset(); 4460 _timer.start(); 4461 4462 // Do the rescan tasks for each of the two spaces 4463 // (cms_space) in turn. 4464 // "worker_id" is passed to select the task_queue for "worker_id" 4465 do_dirty_card_rescan_tasks(_cms_space, worker_id, &par_mrias_cl); 4466 _timer.stop(); 4467 log_trace(gc, task)("Finished dirty card rescan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4468 4469 // ---------- steal work from other threads ... 4470 // ---------- ... and drain overflow list. 4471 _timer.reset(); 4472 _timer.start(); 4473 do_work_steal(worker_id, &par_mrias_cl, _collector->hash_seed(worker_id)); 4474 _timer.stop(); 4475 log_trace(gc, task)("Finished work stealing in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4476 } 4477 4478 void 4479 CMSParMarkTask::do_young_space_rescan( 4480 OopsInGenClosure* cl, ContiguousSpace* space, 4481 HeapWord** chunk_array, size_t chunk_top) { 4482 // Until all tasks completed: 4483 // . claim an unclaimed task 4484 // . compute region boundaries corresponding to task claimed 4485 // using chunk_array 4486 // . par_oop_iterate(cl) over that region 4487 4488 ResourceMark rm; 4489 HandleMark hm; 4490 4491 SequentialSubTasksDone* pst = space->par_seq_tasks(); 4492 4493 uint nth_task = 0; 4494 uint n_tasks = pst->n_tasks(); 4495 4496 if (n_tasks > 0) { 4497 assert(pst->valid(), "Uninitialized use?"); 4498 HeapWord *start, *end; 4499 while (!pst->is_task_claimed(/* reference */ nth_task)) { 4500 // We claimed task # nth_task; compute its boundaries. 4501 if (chunk_top == 0) { // no samples were taken 4502 assert(nth_task == 0 && n_tasks == 1, "Can have only 1 eden task"); 4503 start = space->bottom(); 4504 end = space->top(); 4505 } else if (nth_task == 0) { 4506 start = space->bottom(); 4507 end = chunk_array[nth_task]; 4508 } else if (nth_task < (uint)chunk_top) { 4509 assert(nth_task >= 1, "Control point invariant"); 4510 start = chunk_array[nth_task - 1]; 4511 end = chunk_array[nth_task]; 4512 } else { 4513 assert(nth_task == (uint)chunk_top, "Control point invariant"); 4514 start = chunk_array[chunk_top - 1]; 4515 end = space->top(); 4516 } 4517 MemRegion mr(start, end); 4518 // Verify that mr is in space 4519 assert(mr.is_empty() || space->used_region().contains(mr), 4520 "Should be in space"); 4521 // Verify that "start" is an object boundary 4522 assert(mr.is_empty() || oop(mr.start())->is_oop(), 4523 "Should be an oop"); 4524 space->par_oop_iterate(mr, cl); 4525 } 4526 pst->all_tasks_completed(); 4527 } 4528 } 4529 4530 void 4531 CMSParRemarkTask::do_dirty_card_rescan_tasks( 4532 CompactibleFreeListSpace* sp, int i, 4533 ParMarkRefsIntoAndScanClosure* cl) { 4534 // Until all tasks completed: 4535 // . claim an unclaimed task 4536 // . compute region boundaries corresponding to task claimed 4537 // . transfer dirty bits ct->mut for that region 4538 // . apply rescanclosure to dirty mut bits for that region 4539 4540 ResourceMark rm; 4541 HandleMark hm; 4542 4543 OopTaskQueue* work_q = work_queue(i); 4544 ModUnionClosure modUnionClosure(&(_collector->_modUnionTable)); 4545 // CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! 4546 // CAUTION: This closure has state that persists across calls to 4547 // the work method dirty_range_iterate_clear() in that it has 4548 // embedded in it a (subtype of) UpwardsObjectClosure. The 4549 // use of that state in the embedded UpwardsObjectClosure instance 4550 // assumes that the cards are always iterated (even if in parallel 4551 // by several threads) in monotonically increasing order per each 4552 // thread. This is true of the implementation below which picks 4553 // card ranges (chunks) in monotonically increasing order globally 4554 // and, a-fortiori, in monotonically increasing order per thread 4555 // (the latter order being a subsequence of the former). 4556 // If the work code below is ever reorganized into a more chaotic 4557 // work-partitioning form than the current "sequential tasks" 4558 // paradigm, the use of that persistent state will have to be 4559 // revisited and modified appropriately. See also related 4560 // bug 4756801 work on which should examine this code to make 4561 // sure that the changes there do not run counter to the 4562 // assumptions made here and necessary for correctness and 4563 // efficiency. Note also that this code might yield inefficient 4564 // behavior in the case of very large objects that span one or 4565 // more work chunks. Such objects would potentially be scanned 4566 // several times redundantly. Work on 4756801 should try and 4567 // address that performance anomaly if at all possible. XXX 4568 MemRegion full_span = _collector->_span; 4569 CMSBitMap* bm = &(_collector->_markBitMap); // shared 4570 MarkFromDirtyCardsClosure 4571 greyRescanClosure(_collector, full_span, // entire span of interest 4572 sp, bm, work_q, cl); 4573 4574 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 4575 assert(pst->valid(), "Uninitialized use?"); 4576 uint nth_task = 0; 4577 const int alignment = CardTable::card_size * BitsPerWord; 4578 MemRegion span = sp->used_region(); 4579 HeapWord* start_addr = span.start(); 4580 HeapWord* end_addr = (HeapWord*)round_to((intptr_t)span.end(), 4581 alignment); 4582 const size_t chunk_size = sp->rescan_task_size(); // in HeapWord units 4583 assert((HeapWord*)round_to((intptr_t)start_addr, alignment) == 4584 start_addr, "Check alignment"); 4585 assert((size_t)round_to((intptr_t)chunk_size, alignment) == 4586 chunk_size, "Check alignment"); 4587 4588 while (!pst->is_task_claimed(/* reference */ nth_task)) { 4589 // Having claimed the nth_task, compute corresponding mem-region, 4590 // which is a-fortiori aligned correctly (i.e. at a MUT boundary). 4591 // The alignment restriction ensures that we do not need any 4592 // synchronization with other gang-workers while setting or 4593 // clearing bits in thus chunk of the MUT. 4594 MemRegion this_span = MemRegion(start_addr + nth_task*chunk_size, 4595 start_addr + (nth_task+1)*chunk_size); 4596 // The last chunk's end might be way beyond end of the 4597 // used region. In that case pull back appropriately. 4598 if (this_span.end() > end_addr) { 4599 this_span.set_end(end_addr); 4600 assert(!this_span.is_empty(), "Program logic (calculation of n_tasks)"); 4601 } 4602 // Iterate over the dirty cards covering this chunk, marking them 4603 // precleaned, and setting the corresponding bits in the mod union 4604 // table. Since we have been careful to partition at Card and MUT-word 4605 // boundaries no synchronization is needed between parallel threads. 4606 _collector->_ct->dirty_card_iterate(this_span, 4607 &modUnionClosure); 4608 4609 // Having transferred these marks into the modUnionTable, 4610 // rescan the marked objects on the dirty cards in the modUnionTable. 4611 // Even if this is at a synchronous collection, the initial marking 4612 // may have been done during an asynchronous collection so there 4613 // may be dirty bits in the mod-union table. 4614 _collector->_modUnionTable.dirty_range_iterate_clear( 4615 this_span, &greyRescanClosure); 4616 _collector->_modUnionTable.verifyNoOneBitsInRange( 4617 this_span.start(), 4618 this_span.end()); 4619 } 4620 pst->all_tasks_completed(); // declare that i am done 4621 } 4622 4623 // . see if we can share work_queues with ParNew? XXX 4624 void 4625 CMSParRemarkTask::do_work_steal(int i, ParMarkRefsIntoAndScanClosure* cl, 4626 int* seed) { 4627 OopTaskQueue* work_q = work_queue(i); 4628 NOT_PRODUCT(int num_steals = 0;) 4629 oop obj_to_scan; 4630 CMSBitMap* bm = &(_collector->_markBitMap); 4631 4632 while (true) { 4633 // Completely finish any left over work from (an) earlier round(s) 4634 cl->trim_queue(0); 4635 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 4636 (size_t)ParGCDesiredObjsFromOverflowList); 4637 // Now check if there's any work in the overflow list 4638 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads, 4639 // only affects the number of attempts made to get work from the 4640 // overflow list and does not affect the number of workers. Just 4641 // pass ParallelGCThreads so this behavior is unchanged. 4642 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 4643 work_q, 4644 ParallelGCThreads)) { 4645 // found something in global overflow list; 4646 // not yet ready to go stealing work from others. 4647 // We'd like to assert(work_q->size() != 0, ...) 4648 // because we just took work from the overflow list, 4649 // but of course we can't since all of that could have 4650 // been already stolen from us. 4651 // "He giveth and He taketh away." 4652 continue; 4653 } 4654 // Verify that we have no work before we resort to stealing 4655 assert(work_q->size() == 0, "Have work, shouldn't steal"); 4656 // Try to steal from other queues that have work 4657 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 4658 NOT_PRODUCT(num_steals++;) 4659 assert(obj_to_scan->is_oop(), "Oops, not an oop!"); 4660 assert(bm->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 4661 // Do scanning work 4662 obj_to_scan->oop_iterate(cl); 4663 // Loop around, finish this work, and try to steal some more 4664 } else if (terminator()->offer_termination()) { 4665 break; // nirvana from the infinite cycle 4666 } 4667 } 4668 log_develop_trace(gc, task)("\t(%d: stole %d oops)", i, num_steals); 4669 assert(work_q->size() == 0 && _collector->overflow_list_is_empty(), 4670 "Else our work is not yet done"); 4671 } 4672 4673 // Record object boundaries in _eden_chunk_array by sampling the eden 4674 // top in the slow-path eden object allocation code path and record 4675 // the boundaries, if CMSEdenChunksRecordAlways is true. If 4676 // CMSEdenChunksRecordAlways is false, we use the other asynchronous 4677 // sampling in sample_eden() that activates during the part of the 4678 // preclean phase. 4679 void CMSCollector::sample_eden_chunk() { 4680 if (CMSEdenChunksRecordAlways && _eden_chunk_array != NULL) { 4681 if (_eden_chunk_lock->try_lock()) { 4682 // Record a sample. This is the critical section. The contents 4683 // of the _eden_chunk_array have to be non-decreasing in the 4684 // address order. 4685 _eden_chunk_array[_eden_chunk_index] = *_top_addr; 4686 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr, 4687 "Unexpected state of Eden"); 4688 if (_eden_chunk_index == 0 || 4689 ((_eden_chunk_array[_eden_chunk_index] > _eden_chunk_array[_eden_chunk_index-1]) && 4690 (pointer_delta(_eden_chunk_array[_eden_chunk_index], 4691 _eden_chunk_array[_eden_chunk_index-1]) >= CMSSamplingGrain))) { 4692 _eden_chunk_index++; // commit sample 4693 } 4694 _eden_chunk_lock->unlock(); 4695 } 4696 } 4697 } 4698 4699 // Return a thread-local PLAB recording array, as appropriate. 4700 void* CMSCollector::get_data_recorder(int thr_num) { 4701 if (_survivor_plab_array != NULL && 4702 (CMSPLABRecordAlways || 4703 (_collectorState > Marking && _collectorState < FinalMarking))) { 4704 assert(thr_num < (int)ParallelGCThreads, "thr_num is out of bounds"); 4705 ChunkArray* ca = &_survivor_plab_array[thr_num]; 4706 ca->reset(); // clear it so that fresh data is recorded 4707 return (void*) ca; 4708 } else { 4709 return NULL; 4710 } 4711 } 4712 4713 // Reset all the thread-local PLAB recording arrays 4714 void CMSCollector::reset_survivor_plab_arrays() { 4715 for (uint i = 0; i < ParallelGCThreads; i++) { 4716 _survivor_plab_array[i].reset(); 4717 } 4718 } 4719 4720 // Merge the per-thread plab arrays into the global survivor chunk 4721 // array which will provide the partitioning of the survivor space 4722 // for CMS initial scan and rescan. 4723 void CMSCollector::merge_survivor_plab_arrays(ContiguousSpace* surv, 4724 int no_of_gc_threads) { 4725 assert(_survivor_plab_array != NULL, "Error"); 4726 assert(_survivor_chunk_array != NULL, "Error"); 4727 assert(_collectorState == FinalMarking || 4728 (CMSParallelInitialMarkEnabled && _collectorState == InitialMarking), "Error"); 4729 for (int j = 0; j < no_of_gc_threads; j++) { 4730 _cursor[j] = 0; 4731 } 4732 HeapWord* top = surv->top(); 4733 size_t i; 4734 for (i = 0; i < _survivor_chunk_capacity; i++) { // all sca entries 4735 HeapWord* min_val = top; // Higher than any PLAB address 4736 uint min_tid = 0; // position of min_val this round 4737 for (int j = 0; j < no_of_gc_threads; j++) { 4738 ChunkArray* cur_sca = &_survivor_plab_array[j]; 4739 if (_cursor[j] == cur_sca->end()) { 4740 continue; 4741 } 4742 assert(_cursor[j] < cur_sca->end(), "ctl pt invariant"); 4743 HeapWord* cur_val = cur_sca->nth(_cursor[j]); 4744 assert(surv->used_region().contains(cur_val), "Out of bounds value"); 4745 if (cur_val < min_val) { 4746 min_tid = j; 4747 min_val = cur_val; 4748 } else { 4749 assert(cur_val < top, "All recorded addresses should be less"); 4750 } 4751 } 4752 // At this point min_val and min_tid are respectively 4753 // the least address in _survivor_plab_array[j]->nth(_cursor[j]) 4754 // and the thread (j) that witnesses that address. 4755 // We record this address in the _survivor_chunk_array[i] 4756 // and increment _cursor[min_tid] prior to the next round i. 4757 if (min_val == top) { 4758 break; 4759 } 4760 _survivor_chunk_array[i] = min_val; 4761 _cursor[min_tid]++; 4762 } 4763 // We are all done; record the size of the _survivor_chunk_array 4764 _survivor_chunk_index = i; // exclusive: [0, i) 4765 log_trace(gc, survivor)(" (Survivor:" SIZE_FORMAT "chunks) ", i); 4766 // Verify that we used up all the recorded entries 4767 #ifdef ASSERT 4768 size_t total = 0; 4769 for (int j = 0; j < no_of_gc_threads; j++) { 4770 assert(_cursor[j] == _survivor_plab_array[j].end(), "Ctl pt invariant"); 4771 total += _cursor[j]; 4772 } 4773 assert(total == _survivor_chunk_index, "Ctl Pt Invariant"); 4774 // Check that the merged array is in sorted order 4775 if (total > 0) { 4776 for (size_t i = 0; i < total - 1; i++) { 4777 log_develop_trace(gc, survivor)(" (chunk" SIZE_FORMAT ":" INTPTR_FORMAT ") ", 4778 i, p2i(_survivor_chunk_array[i])); 4779 assert(_survivor_chunk_array[i] < _survivor_chunk_array[i+1], 4780 "Not sorted"); 4781 } 4782 } 4783 #endif // ASSERT 4784 } 4785 4786 // Set up the space's par_seq_tasks structure for work claiming 4787 // for parallel initial scan and rescan of young gen. 4788 // See ParRescanTask where this is currently used. 4789 void 4790 CMSCollector:: 4791 initialize_sequential_subtasks_for_young_gen_rescan(int n_threads) { 4792 assert(n_threads > 0, "Unexpected n_threads argument"); 4793 4794 // Eden space 4795 if (!_young_gen->eden()->is_empty()) { 4796 SequentialSubTasksDone* pst = _young_gen->eden()->par_seq_tasks(); 4797 assert(!pst->valid(), "Clobbering existing data?"); 4798 // Each valid entry in [0, _eden_chunk_index) represents a task. 4799 size_t n_tasks = _eden_chunk_index + 1; 4800 assert(n_tasks == 1 || _eden_chunk_array != NULL, "Error"); 4801 // Sets the condition for completion of the subtask (how many threads 4802 // need to finish in order to be done). 4803 pst->set_n_threads(n_threads); 4804 pst->set_n_tasks((int)n_tasks); 4805 } 4806 4807 // Merge the survivor plab arrays into _survivor_chunk_array 4808 if (_survivor_plab_array != NULL) { 4809 merge_survivor_plab_arrays(_young_gen->from(), n_threads); 4810 } else { 4811 assert(_survivor_chunk_index == 0, "Error"); 4812 } 4813 4814 // To space 4815 { 4816 SequentialSubTasksDone* pst = _young_gen->to()->par_seq_tasks(); 4817 assert(!pst->valid(), "Clobbering existing data?"); 4818 // Sets the condition for completion of the subtask (how many threads 4819 // need to finish in order to be done). 4820 pst->set_n_threads(n_threads); 4821 pst->set_n_tasks(1); 4822 assert(pst->valid(), "Error"); 4823 } 4824 4825 // From space 4826 { 4827 SequentialSubTasksDone* pst = _young_gen->from()->par_seq_tasks(); 4828 assert(!pst->valid(), "Clobbering existing data?"); 4829 size_t n_tasks = _survivor_chunk_index + 1; 4830 assert(n_tasks == 1 || _survivor_chunk_array != NULL, "Error"); 4831 // Sets the condition for completion of the subtask (how many threads 4832 // need to finish in order to be done). 4833 pst->set_n_threads(n_threads); 4834 pst->set_n_tasks((int)n_tasks); 4835 assert(pst->valid(), "Error"); 4836 } 4837 } 4838 4839 // Parallel version of remark 4840 void CMSCollector::do_remark_parallel() { 4841 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4842 WorkGang* workers = gch->workers(); 4843 assert(workers != NULL, "Need parallel worker threads."); 4844 // Choose to use the number of GC workers most recently set 4845 // into "active_workers". 4846 uint n_workers = workers->active_workers(); 4847 4848 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 4849 4850 StrongRootsScope srs(n_workers); 4851 4852 CMSParRemarkTask tsk(this, cms_space, n_workers, workers, task_queues(), &srs); 4853 4854 // We won't be iterating over the cards in the card table updating 4855 // the younger_gen cards, so we shouldn't call the following else 4856 // the verification code as well as subsequent younger_refs_iterate 4857 // code would get confused. XXX 4858 // gch->rem_set()->prepare_for_younger_refs_iterate(true); // parallel 4859 4860 // The young gen rescan work will not be done as part of 4861 // process_roots (which currently doesn't know how to 4862 // parallelize such a scan), but rather will be broken up into 4863 // a set of parallel tasks (via the sampling that the [abortable] 4864 // preclean phase did of eden, plus the [two] tasks of 4865 // scanning the [two] survivor spaces. Further fine-grain 4866 // parallelization of the scanning of the survivor spaces 4867 // themselves, and of precleaning of the young gen itself 4868 // is deferred to the future. 4869 initialize_sequential_subtasks_for_young_gen_rescan(n_workers); 4870 4871 // The dirty card rescan work is broken up into a "sequence" 4872 // of parallel tasks (per constituent space) that are dynamically 4873 // claimed by the parallel threads. 4874 cms_space->initialize_sequential_subtasks_for_rescan(n_workers); 4875 4876 // It turns out that even when we're using 1 thread, doing the work in a 4877 // separate thread causes wide variance in run times. We can't help this 4878 // in the multi-threaded case, but we special-case n=1 here to get 4879 // repeatable measurements of the 1-thread overhead of the parallel code. 4880 if (n_workers > 1) { 4881 // Make refs discovery MT-safe, if it isn't already: it may not 4882 // necessarily be so, since it's possible that we are doing 4883 // ST marking. 4884 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), true); 4885 workers->run_task(&tsk); 4886 } else { 4887 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false); 4888 tsk.work(0); 4889 } 4890 4891 // restore, single-threaded for now, any preserved marks 4892 // as a result of work_q overflow 4893 restore_preserved_marks_if_any(); 4894 } 4895 4896 // Non-parallel version of remark 4897 void CMSCollector::do_remark_non_parallel() { 4898 ResourceMark rm; 4899 HandleMark hm; 4900 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4901 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false); 4902 4903 MarkRefsIntoAndScanClosure 4904 mrias_cl(_span, ref_processor(), &_markBitMap, NULL /* not precleaning */, 4905 &_markStack, this, 4906 false /* should_yield */, false /* not precleaning */); 4907 MarkFromDirtyCardsClosure 4908 markFromDirtyCardsClosure(this, _span, 4909 NULL, // space is set further below 4910 &_markBitMap, &_markStack, &mrias_cl); 4911 { 4912 GCTraceTime(Trace, gc, phases) t("Grey Object Rescan", _gc_timer_cm); 4913 // Iterate over the dirty cards, setting the corresponding bits in the 4914 // mod union table. 4915 { 4916 ModUnionClosure modUnionClosure(&_modUnionTable); 4917 _ct->dirty_card_iterate(_cmsGen->used_region(), 4918 &modUnionClosure); 4919 } 4920 // Having transferred these marks into the modUnionTable, we just need 4921 // to rescan the marked objects on the dirty cards in the modUnionTable. 4922 // The initial marking may have been done during an asynchronous 4923 // collection so there may be dirty bits in the mod-union table. 4924 const int alignment = CardTable::card_size * BitsPerWord; 4925 { 4926 // ... First handle dirty cards in CMS gen 4927 markFromDirtyCardsClosure.set_space(_cmsGen->cmsSpace()); 4928 MemRegion ur = _cmsGen->used_region(); 4929 HeapWord* lb = ur.start(); 4930 HeapWord* ub = (HeapWord*)round_to((intptr_t)ur.end(), alignment); 4931 MemRegion cms_span(lb, ub); 4932 _modUnionTable.dirty_range_iterate_clear(cms_span, 4933 &markFromDirtyCardsClosure); 4934 verify_work_stacks_empty(); 4935 log_trace(gc)(" (re-scanned " SIZE_FORMAT " dirty cards in cms gen) ", markFromDirtyCardsClosure.num_dirty_cards()); 4936 } 4937 } 4938 if (VerifyDuringGC && 4939 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 4940 HandleMark hm; // Discard invalid handles created during verification 4941 Universe::verify(); 4942 } 4943 { 4944 GCTraceTime(Trace, gc, phases) t("Root Rescan", _gc_timer_cm); 4945 4946 verify_work_stacks_empty(); 4947 4948 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 4949 StrongRootsScope srs(1); 4950 4951 gch->cms_process_roots(&srs, 4952 true, // young gen as roots 4953 GenCollectedHeap::ScanningOption(roots_scanning_options()), 4954 should_unload_classes(), 4955 &mrias_cl, 4956 NULL); // The dirty klasses will be handled below 4957 4958 assert(should_unload_classes() 4959 || (roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache), 4960 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 4961 } 4962 4963 { 4964 GCTraceTime(Trace, gc, phases) t("Visit Unhandled CLDs", _gc_timer_cm); 4965 4966 verify_work_stacks_empty(); 4967 4968 // Scan all class loader data objects that might have been introduced 4969 // during concurrent marking. 4970 ResourceMark rm; 4971 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds(); 4972 for (int i = 0; i < array->length(); i++) { 4973 mrias_cl.do_cld_nv(array->at(i)); 4974 } 4975 4976 // We don't need to keep track of new CLDs anymore. 4977 ClassLoaderDataGraph::remember_new_clds(false); 4978 4979 verify_work_stacks_empty(); 4980 } 4981 4982 { 4983 GCTraceTime(Trace, gc, phases) t("Dirty Klass Scan", _gc_timer_cm); 4984 4985 verify_work_stacks_empty(); 4986 4987 RemarkKlassClosure remark_klass_closure(&mrias_cl); 4988 ClassLoaderDataGraph::classes_do(&remark_klass_closure); 4989 4990 verify_work_stacks_empty(); 4991 } 4992 4993 // We might have added oops to ClassLoaderData::_handles during the 4994 // concurrent marking phase. These oops point to newly allocated objects 4995 // that are guaranteed to be kept alive. Either by the direct allocation 4996 // code, or when the young collector processes the roots. Hence, 4997 // we don't have to revisit the _handles block during the remark phase. 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(obj_to_scan->is_oop(), "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 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5151 WorkGang* workers = gch->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 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5164 WorkGang* workers = gch->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 CMSKeepAliveClosure cmsKeepAliveClosure(this, _span, &_markBitMap, 5182 &_markStack, false /* !preclean */); 5183 CMSDrainMarkingStackClosure cmsDrainMarkingStackClosure(this, 5184 _span, &_markBitMap, &_markStack, 5185 &cmsKeepAliveClosure, false /* !preclean */); 5186 { 5187 GCTraceTime(Debug, gc, phases) t("Reference Processing", _gc_timer_cm); 5188 5189 ReferenceProcessorStats stats; 5190 if (rp->processing_is_mt()) { 5191 // Set the degree of MT here. If the discovery is done MT, there 5192 // may have been a different number of threads doing the discovery 5193 // and a different number of discovered lists may have Ref objects. 5194 // That is OK as long as the Reference lists are balanced (see 5195 // balance_all_queues() and balance_queues()). 5196 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5197 uint active_workers = ParallelGCThreads; 5198 WorkGang* workers = gch->workers(); 5199 if (workers != NULL) { 5200 active_workers = workers->active_workers(); 5201 // The expectation is that active_workers will have already 5202 // been set to a reasonable value. If it has not been set, 5203 // investigate. 5204 assert(active_workers > 0, "Should have been set during scavenge"); 5205 } 5206 rp->set_active_mt_degree(active_workers); 5207 CMSRefProcTaskExecutor task_executor(*this); 5208 stats = rp->process_discovered_references(&_is_alive_closure, 5209 &cmsKeepAliveClosure, 5210 &cmsDrainMarkingStackClosure, 5211 &task_executor, 5212 _gc_timer_cm); 5213 } else { 5214 stats = rp->process_discovered_references(&_is_alive_closure, 5215 &cmsKeepAliveClosure, 5216 &cmsDrainMarkingStackClosure, 5217 NULL, 5218 _gc_timer_cm); 5219 } 5220 _gc_tracer_cm->report_gc_reference_stats(stats); 5221 5222 } 5223 5224 // This is the point where the entire marking should have completed. 5225 verify_work_stacks_empty(); 5226 5227 if (should_unload_classes()) { 5228 { 5229 GCTraceTime(Debug, gc, phases) t("Class Unloading", _gc_timer_cm); 5230 5231 // Unload classes and purge the SystemDictionary. 5232 bool purged_class = SystemDictionary::do_unloading(&_is_alive_closure); 5233 5234 // Unload nmethods. 5235 CodeCache::do_unloading(&_is_alive_closure, purged_class); 5236 5237 // Prune dead klasses from subklass/sibling/implementor lists. 5238 Klass::clean_weak_klass_links(&_is_alive_closure); 5239 } 5240 5241 { 5242 GCTraceTime(Debug, gc, phases) t("Scrub Symbol Table", _gc_timer_cm); 5243 // Clean up unreferenced symbols in symbol table. 5244 SymbolTable::unlink(); 5245 } 5246 5247 { 5248 GCTraceTime(Debug, gc, phases) t("Scrub String Table", _gc_timer_cm); 5249 // Delete entries for dead interned strings. 5250 StringTable::unlink(&_is_alive_closure); 5251 } 5252 } 5253 5254 5255 // Restore any preserved marks as a result of mark stack or 5256 // work queue overflow 5257 restore_preserved_marks_if_any(); // done single-threaded for now 5258 5259 rp->set_enqueuing_is_done(true); 5260 if (rp->processing_is_mt()) { 5261 rp->balance_all_queues(); 5262 CMSRefProcTaskExecutor task_executor(*this); 5263 rp->enqueue_discovered_references(&task_executor); 5264 } else { 5265 rp->enqueue_discovered_references(NULL); 5266 } 5267 rp->verify_no_references_recorded(); 5268 assert(!rp->discovery_enabled(), "should have been disabled"); 5269 } 5270 5271 #ifndef PRODUCT 5272 void CMSCollector::check_correct_thread_executing() { 5273 Thread* t = Thread::current(); 5274 // Only the VM thread or the CMS thread should be here. 5275 assert(t->is_ConcurrentGC_thread() || t->is_VM_thread(), 5276 "Unexpected thread type"); 5277 // If this is the vm thread, the foreground process 5278 // should not be waiting. Note that _foregroundGCIsActive is 5279 // true while the foreground collector is waiting. 5280 if (_foregroundGCShouldWait) { 5281 // We cannot be the VM thread 5282 assert(t->is_ConcurrentGC_thread(), 5283 "Should be CMS thread"); 5284 } else { 5285 // We can be the CMS thread only if we are in a stop-world 5286 // phase of CMS collection. 5287 if (t->is_ConcurrentGC_thread()) { 5288 assert(_collectorState == InitialMarking || 5289 _collectorState == FinalMarking, 5290 "Should be a stop-world phase"); 5291 // The CMS thread should be holding the CMS_token. 5292 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 5293 "Potential interference with concurrently " 5294 "executing VM thread"); 5295 } 5296 } 5297 } 5298 #endif 5299 5300 void CMSCollector::sweep() { 5301 assert(_collectorState == Sweeping, "just checking"); 5302 check_correct_thread_executing(); 5303 verify_work_stacks_empty(); 5304 verify_overflow_empty(); 5305 increment_sweep_count(); 5306 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 5307 5308 _inter_sweep_timer.stop(); 5309 _inter_sweep_estimate.sample(_inter_sweep_timer.seconds()); 5310 5311 assert(!_intra_sweep_timer.is_active(), "Should not be active"); 5312 _intra_sweep_timer.reset(); 5313 _intra_sweep_timer.start(); 5314 { 5315 GCTraceCPUTime tcpu; 5316 CMSPhaseAccounting pa(this, "Concurrent Sweep"); 5317 // First sweep the old gen 5318 { 5319 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(), 5320 bitMapLock()); 5321 sweepWork(_cmsGen); 5322 } 5323 5324 // Update Universe::_heap_*_at_gc figures. 5325 // We need all the free list locks to make the abstract state 5326 // transition from Sweeping to Resetting. See detailed note 5327 // further below. 5328 { 5329 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock()); 5330 // Update heap occupancy information which is used as 5331 // input to soft ref clearing policy at the next gc. 5332 Universe::update_heap_info_at_gc(); 5333 _collectorState = Resizing; 5334 } 5335 } 5336 verify_work_stacks_empty(); 5337 verify_overflow_empty(); 5338 5339 if (should_unload_classes()) { 5340 // Delay purge to the beginning of the next safepoint. Metaspace::contains 5341 // requires that the virtual spaces are stable and not deleted. 5342 ClassLoaderDataGraph::set_should_purge(true); 5343 } 5344 5345 _intra_sweep_timer.stop(); 5346 _intra_sweep_estimate.sample(_intra_sweep_timer.seconds()); 5347 5348 _inter_sweep_timer.reset(); 5349 _inter_sweep_timer.start(); 5350 5351 // We need to use a monotonically non-decreasing time in ms 5352 // or we will see time-warp warnings and os::javaTimeMillis() 5353 // does not guarantee monotonicity. 5354 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC; 5355 update_time_of_last_gc(now); 5356 5357 // NOTE on abstract state transitions: 5358 // Mutators allocate-live and/or mark the mod-union table dirty 5359 // based on the state of the collection. The former is done in 5360 // the interval [Marking, Sweeping] and the latter in the interval 5361 // [Marking, Sweeping). Thus the transitions into the Marking state 5362 // and out of the Sweeping state must be synchronously visible 5363 // globally to the mutators. 5364 // The transition into the Marking state happens with the world 5365 // stopped so the mutators will globally see it. Sweeping is 5366 // done asynchronously by the background collector so the transition 5367 // from the Sweeping state to the Resizing state must be done 5368 // under the freelistLock (as is the check for whether to 5369 // allocate-live and whether to dirty the mod-union table). 5370 assert(_collectorState == Resizing, "Change of collector state to" 5371 " Resizing must be done under the freelistLocks (plural)"); 5372 5373 // Now that sweeping has been completed, we clear 5374 // the incremental_collection_failed flag, 5375 // thus inviting a younger gen collection to promote into 5376 // this generation. If such a promotion may still fail, 5377 // the flag will be set again when a young collection is 5378 // attempted. 5379 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5380 gch->clear_incremental_collection_failed(); // Worth retrying as fresh space may have been freed up 5381 gch->update_full_collections_completed(_collection_count_start); 5382 } 5383 5384 // FIX ME!!! Looks like this belongs in CFLSpace, with 5385 // CMSGen merely delegating to it. 5386 void ConcurrentMarkSweepGeneration::setNearLargestChunk() { 5387 double nearLargestPercent = FLSLargestBlockCoalesceProximity; 5388 HeapWord* minAddr = _cmsSpace->bottom(); 5389 HeapWord* largestAddr = 5390 (HeapWord*) _cmsSpace->dictionary()->find_largest_dict(); 5391 if (largestAddr == NULL) { 5392 // The dictionary appears to be empty. In this case 5393 // try to coalesce at the end of the heap. 5394 largestAddr = _cmsSpace->end(); 5395 } 5396 size_t largestOffset = pointer_delta(largestAddr, minAddr); 5397 size_t nearLargestOffset = 5398 (size_t)((double)largestOffset * nearLargestPercent) - MinChunkSize; 5399 log_debug(gc, freelist)("CMS: Large Block: " PTR_FORMAT "; Proximity: " PTR_FORMAT " -> " PTR_FORMAT, 5400 p2i(largestAddr), p2i(_cmsSpace->nearLargestChunk()), p2i(minAddr + nearLargestOffset)); 5401 _cmsSpace->set_nearLargestChunk(minAddr + nearLargestOffset); 5402 } 5403 5404 bool ConcurrentMarkSweepGeneration::isNearLargestChunk(HeapWord* addr) { 5405 return addr >= _cmsSpace->nearLargestChunk(); 5406 } 5407 5408 FreeChunk* ConcurrentMarkSweepGeneration::find_chunk_at_end() { 5409 return _cmsSpace->find_chunk_at_end(); 5410 } 5411 5412 void ConcurrentMarkSweepGeneration::update_gc_stats(Generation* current_generation, 5413 bool full) { 5414 // If the young generation has been collected, gather any statistics 5415 // that are of interest at this point. 5416 bool current_is_young = GenCollectedHeap::heap()->is_young_gen(current_generation); 5417 if (!full && current_is_young) { 5418 // Gather statistics on the young generation collection. 5419 collector()->stats().record_gc0_end(used()); 5420 } 5421 } 5422 5423 void CMSCollector::sweepWork(ConcurrentMarkSweepGeneration* old_gen) { 5424 // We iterate over the space(s) underlying this generation, 5425 // checking the mark bit map to see if the bits corresponding 5426 // to specific blocks are marked or not. Blocks that are 5427 // marked are live and are not swept up. All remaining blocks 5428 // are swept up, with coalescing on-the-fly as we sweep up 5429 // contiguous free and/or garbage blocks: 5430 // We need to ensure that the sweeper synchronizes with allocators 5431 // and stop-the-world collectors. In particular, the following 5432 // locks are used: 5433 // . CMS token: if this is held, a stop the world collection cannot occur 5434 // . freelistLock: if this is held no allocation can occur from this 5435 // generation by another thread 5436 // . bitMapLock: if this is held, no other thread can access or update 5437 // 5438 5439 // Note that we need to hold the freelistLock if we use 5440 // block iterate below; else the iterator might go awry if 5441 // a mutator (or promotion) causes block contents to change 5442 // (for instance if the allocator divvies up a block). 5443 // If we hold the free list lock, for all practical purposes 5444 // young generation GC's can't occur (they'll usually need to 5445 // promote), so we might as well prevent all young generation 5446 // GC's while we do a sweeping step. For the same reason, we might 5447 // as well take the bit map lock for the entire duration 5448 5449 // check that we hold the requisite locks 5450 assert(have_cms_token(), "Should hold cms token"); 5451 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), "Should possess CMS token to sweep"); 5452 assert_lock_strong(old_gen->freelistLock()); 5453 assert_lock_strong(bitMapLock()); 5454 5455 assert(!_inter_sweep_timer.is_active(), "Was switched off in an outer context"); 5456 assert(_intra_sweep_timer.is_active(), "Was switched on in an outer context"); 5457 old_gen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()), 5458 _inter_sweep_estimate.padded_average(), 5459 _intra_sweep_estimate.padded_average()); 5460 old_gen->setNearLargestChunk(); 5461 5462 { 5463 SweepClosure sweepClosure(this, old_gen, &_markBitMap, CMSYield); 5464 old_gen->cmsSpace()->blk_iterate_careful(&sweepClosure); 5465 // We need to free-up/coalesce garbage/blocks from a 5466 // co-terminal free run. This is done in the SweepClosure 5467 // destructor; so, do not remove this scope, else the 5468 // end-of-sweep-census below will be off by a little bit. 5469 } 5470 old_gen->cmsSpace()->sweep_completed(); 5471 old_gen->cmsSpace()->endSweepFLCensus(sweep_count()); 5472 if (should_unload_classes()) { // unloaded classes this cycle, 5473 _concurrent_cycles_since_last_unload = 0; // ... reset count 5474 } else { // did not unload classes, 5475 _concurrent_cycles_since_last_unload++; // ... increment count 5476 } 5477 } 5478 5479 // Reset CMS data structures (for now just the marking bit map) 5480 // preparatory for the next cycle. 5481 void CMSCollector::reset_concurrent() { 5482 CMSTokenSyncWithLocks ts(true, bitMapLock()); 5483 5484 // If the state is not "Resetting", the foreground thread 5485 // has done a collection and the resetting. 5486 if (_collectorState != Resetting) { 5487 assert(_collectorState == Idling, "The state should only change" 5488 " because the foreground collector has finished the collection"); 5489 return; 5490 } 5491 5492 { 5493 // Clear the mark bitmap (no grey objects to start with) 5494 // for the next cycle. 5495 GCTraceCPUTime tcpu; 5496 CMSPhaseAccounting cmspa(this, "Concurrent Reset"); 5497 5498 HeapWord* curAddr = _markBitMap.startWord(); 5499 while (curAddr < _markBitMap.endWord()) { 5500 size_t remaining = pointer_delta(_markBitMap.endWord(), curAddr); 5501 MemRegion chunk(curAddr, MIN2(CMSBitMapYieldQuantum, remaining)); 5502 _markBitMap.clear_large_range(chunk); 5503 if (ConcurrentMarkSweepThread::should_yield() && 5504 !foregroundGCIsActive() && 5505 CMSYield) { 5506 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 5507 "CMS thread should hold CMS token"); 5508 assert_lock_strong(bitMapLock()); 5509 bitMapLock()->unlock(); 5510 ConcurrentMarkSweepThread::desynchronize(true); 5511 stopTimer(); 5512 incrementYields(); 5513 5514 // See the comment in coordinator_yield() 5515 for (unsigned i = 0; i < CMSYieldSleepCount && 5516 ConcurrentMarkSweepThread::should_yield() && 5517 !CMSCollector::foregroundGCIsActive(); ++i) { 5518 os::sleep(Thread::current(), 1, false); 5519 } 5520 5521 ConcurrentMarkSweepThread::synchronize(true); 5522 bitMapLock()->lock_without_safepoint_check(); 5523 startTimer(); 5524 } 5525 curAddr = chunk.end(); 5526 } 5527 // A successful mostly concurrent collection has been done. 5528 // Because only the full (i.e., concurrent mode failure) collections 5529 // are being measured for gc overhead limits, clean the "near" flag 5530 // and count. 5531 size_policy()->reset_gc_overhead_limit_count(); 5532 _collectorState = Idling; 5533 } 5534 5535 register_gc_end(); 5536 } 5537 5538 // Same as above but for STW paths 5539 void CMSCollector::reset_stw() { 5540 // already have the lock 5541 assert(_collectorState == Resetting, "just checking"); 5542 assert_lock_strong(bitMapLock()); 5543 GCIdMarkAndRestore gc_id_mark(_cmsThread->gc_id()); 5544 _markBitMap.clear_all(); 5545 _collectorState = Idling; 5546 register_gc_end(); 5547 } 5548 5549 void CMSCollector::do_CMS_operation(CMS_op_type op, GCCause::Cause gc_cause) { 5550 GCTraceCPUTime tcpu; 5551 TraceCollectorStats tcs(counters()); 5552 5553 switch (op) { 5554 case CMS_op_checkpointRootsInitial: { 5555 GCTraceTime(Info, gc) t("Pause Initial Mark", NULL, GCCause::_no_gc, true); 5556 SvcGCMarker sgcm(SvcGCMarker::OTHER); 5557 checkpointRootsInitial(); 5558 break; 5559 } 5560 case CMS_op_checkpointRootsFinal: { 5561 GCTraceTime(Info, gc) t("Pause Remark", NULL, GCCause::_no_gc, true); 5562 SvcGCMarker sgcm(SvcGCMarker::OTHER); 5563 checkpointRootsFinal(); 5564 break; 5565 } 5566 default: 5567 fatal("No such CMS_op"); 5568 } 5569 } 5570 5571 #ifndef PRODUCT 5572 size_t const CMSCollector::skip_header_HeapWords() { 5573 return FreeChunk::header_size(); 5574 } 5575 5576 // Try and collect here conditions that should hold when 5577 // CMS thread is exiting. The idea is that the foreground GC 5578 // thread should not be blocked if it wants to terminate 5579 // the CMS thread and yet continue to run the VM for a while 5580 // after that. 5581 void CMSCollector::verify_ok_to_terminate() const { 5582 assert(Thread::current()->is_ConcurrentGC_thread(), 5583 "should be called by CMS thread"); 5584 assert(!_foregroundGCShouldWait, "should be false"); 5585 // We could check here that all the various low-level locks 5586 // are not held by the CMS thread, but that is overkill; see 5587 // also CMSThread::verify_ok_to_terminate() where the CGC_lock 5588 // is checked. 5589 } 5590 #endif 5591 5592 size_t CMSCollector::block_size_using_printezis_bits(HeapWord* addr) const { 5593 assert(_markBitMap.isMarked(addr) && _markBitMap.isMarked(addr + 1), 5594 "missing Printezis mark?"); 5595 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 5596 size_t size = pointer_delta(nextOneAddr + 1, addr); 5597 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 5598 "alignment problem"); 5599 assert(size >= 3, "Necessary for Printezis marks to work"); 5600 return size; 5601 } 5602 5603 // A variant of the above (block_size_using_printezis_bits()) except 5604 // that we return 0 if the P-bits are not yet set. 5605 size_t CMSCollector::block_size_if_printezis_bits(HeapWord* addr) const { 5606 if (_markBitMap.isMarked(addr + 1)) { 5607 assert(_markBitMap.isMarked(addr), "P-bit can be set only for marked objects"); 5608 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 5609 size_t size = pointer_delta(nextOneAddr + 1, addr); 5610 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 5611 "alignment problem"); 5612 assert(size >= 3, "Necessary for Printezis marks to work"); 5613 return size; 5614 } 5615 return 0; 5616 } 5617 5618 HeapWord* CMSCollector::next_card_start_after_block(HeapWord* addr) const { 5619 size_t sz = 0; 5620 oop p = (oop)addr; 5621 if (p->klass_or_null_acquire() != NULL) { 5622 sz = CompactibleFreeListSpace::adjustObjectSize(p->size()); 5623 } else { 5624 sz = block_size_using_printezis_bits(addr); 5625 } 5626 assert(sz > 0, "size must be nonzero"); 5627 HeapWord* next_block = addr + sz; 5628 HeapWord* next_card = (HeapWord*)round_to((uintptr_t)next_block, 5629 CardTable::card_size); 5630 assert(round_down((uintptr_t)addr, CardTable::card_size) < 5631 round_down((uintptr_t)next_card, CardTable::card_size), 5632 "must be different cards"); 5633 return next_card; 5634 } 5635 5636 5637 // CMS Bit Map Wrapper ///////////////////////////////////////// 5638 5639 // Construct a CMS bit map infrastructure, but don't create the 5640 // bit vector itself. That is done by a separate call CMSBitMap::allocate() 5641 // further below. 5642 CMSBitMap::CMSBitMap(int shifter, int mutex_rank, const char* mutex_name): 5643 _bm(), 5644 _shifter(shifter), 5645 _lock(mutex_rank >= 0 ? new Mutex(mutex_rank, mutex_name, true, 5646 Monitor::_safepoint_check_sometimes) : NULL) 5647 { 5648 _bmStartWord = 0; 5649 _bmWordSize = 0; 5650 } 5651 5652 bool CMSBitMap::allocate(MemRegion mr) { 5653 _bmStartWord = mr.start(); 5654 _bmWordSize = mr.word_size(); 5655 ReservedSpace brs(ReservedSpace::allocation_align_size_up( 5656 (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1)); 5657 if (!brs.is_reserved()) { 5658 log_warning(gc)("CMS bit map allocation failure"); 5659 return false; 5660 } 5661 // For now we'll just commit all of the bit map up front. 5662 // Later on we'll try to be more parsimonious with swap. 5663 if (!_virtual_space.initialize(brs, brs.size())) { 5664 log_warning(gc)("CMS bit map backing store failure"); 5665 return false; 5666 } 5667 assert(_virtual_space.committed_size() == brs.size(), 5668 "didn't reserve backing store for all of CMS bit map?"); 5669 assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >= 5670 _bmWordSize, "inconsistency in bit map sizing"); 5671 _bm = BitMapView((BitMap::bm_word_t*)_virtual_space.low(), _bmWordSize >> _shifter); 5672 5673 // bm.clear(); // can we rely on getting zero'd memory? verify below 5674 assert(isAllClear(), 5675 "Expected zero'd memory from ReservedSpace constructor"); 5676 assert(_bm.size() == heapWordDiffToOffsetDiff(sizeInWords()), 5677 "consistency check"); 5678 return true; 5679 } 5680 5681 void CMSBitMap::dirty_range_iterate_clear(MemRegion mr, MemRegionClosure* cl) { 5682 HeapWord *next_addr, *end_addr, *last_addr; 5683 assert_locked(); 5684 assert(covers(mr), "out-of-range error"); 5685 // XXX assert that start and end are appropriately aligned 5686 for (next_addr = mr.start(), end_addr = mr.end(); 5687 next_addr < end_addr; next_addr = last_addr) { 5688 MemRegion dirty_region = getAndClearMarkedRegion(next_addr, end_addr); 5689 last_addr = dirty_region.end(); 5690 if (!dirty_region.is_empty()) { 5691 cl->do_MemRegion(dirty_region); 5692 } else { 5693 assert(last_addr == end_addr, "program logic"); 5694 return; 5695 } 5696 } 5697 } 5698 5699 void CMSBitMap::print_on_error(outputStream* st, const char* prefix) const { 5700 _bm.print_on_error(st, prefix); 5701 } 5702 5703 #ifndef PRODUCT 5704 void CMSBitMap::assert_locked() const { 5705 CMSLockVerifier::assert_locked(lock()); 5706 } 5707 5708 bool CMSBitMap::covers(MemRegion mr) const { 5709 // assert(_bm.map() == _virtual_space.low(), "map inconsistency"); 5710 assert((size_t)_bm.size() == (_bmWordSize >> _shifter), 5711 "size inconsistency"); 5712 return (mr.start() >= _bmStartWord) && 5713 (mr.end() <= endWord()); 5714 } 5715 5716 bool CMSBitMap::covers(HeapWord* start, size_t size) const { 5717 return (start >= _bmStartWord && (start + size) <= endWord()); 5718 } 5719 5720 void CMSBitMap::verifyNoOneBitsInRange(HeapWord* left, HeapWord* right) { 5721 // verify that there are no 1 bits in the interval [left, right) 5722 FalseBitMapClosure falseBitMapClosure; 5723 iterate(&falseBitMapClosure, left, right); 5724 } 5725 5726 void CMSBitMap::region_invariant(MemRegion mr) 5727 { 5728 assert_locked(); 5729 // mr = mr.intersection(MemRegion(_bmStartWord, _bmWordSize)); 5730 assert(!mr.is_empty(), "unexpected empty region"); 5731 assert(covers(mr), "mr should be covered by bit map"); 5732 // convert address range into offset range 5733 size_t start_ofs = heapWordToOffset(mr.start()); 5734 // Make sure that end() is appropriately aligned 5735 assert(mr.end() == (HeapWord*)round_to((intptr_t)mr.end(), 5736 (1 << (_shifter+LogHeapWordSize))), 5737 "Misaligned mr.end()"); 5738 size_t end_ofs = heapWordToOffset(mr.end()); 5739 assert(end_ofs > start_ofs, "Should mark at least one bit"); 5740 } 5741 5742 #endif 5743 5744 bool CMSMarkStack::allocate(size_t size) { 5745 // allocate a stack of the requisite depth 5746 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 5747 size * sizeof(oop))); 5748 if (!rs.is_reserved()) { 5749 log_warning(gc)("CMSMarkStack allocation failure"); 5750 return false; 5751 } 5752 if (!_virtual_space.initialize(rs, rs.size())) { 5753 log_warning(gc)("CMSMarkStack backing store failure"); 5754 return false; 5755 } 5756 assert(_virtual_space.committed_size() == rs.size(), 5757 "didn't reserve backing store for all of CMS stack?"); 5758 _base = (oop*)(_virtual_space.low()); 5759 _index = 0; 5760 _capacity = size; 5761 NOT_PRODUCT(_max_depth = 0); 5762 return true; 5763 } 5764 5765 // XXX FIX ME !!! In the MT case we come in here holding a 5766 // leaf lock. For printing we need to take a further lock 5767 // which has lower rank. We need to recalibrate the two 5768 // lock-ranks involved in order to be able to print the 5769 // messages below. (Or defer the printing to the caller. 5770 // For now we take the expedient path of just disabling the 5771 // messages for the problematic case.) 5772 void CMSMarkStack::expand() { 5773 assert(_capacity <= MarkStackSizeMax, "stack bigger than permitted"); 5774 if (_capacity == MarkStackSizeMax) { 5775 if (_hit_limit++ == 0 && !CMSConcurrentMTEnabled) { 5776 // We print a warning message only once per CMS cycle. 5777 log_debug(gc)(" (benign) Hit CMSMarkStack max size limit"); 5778 } 5779 return; 5780 } 5781 // Double capacity if possible 5782 size_t new_capacity = MIN2(_capacity*2, MarkStackSizeMax); 5783 // Do not give up existing stack until we have managed to 5784 // get the double capacity that we desired. 5785 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 5786 new_capacity * sizeof(oop))); 5787 if (rs.is_reserved()) { 5788 // Release the backing store associated with old stack 5789 _virtual_space.release(); 5790 // Reinitialize virtual space for new stack 5791 if (!_virtual_space.initialize(rs, rs.size())) { 5792 fatal("Not enough swap for expanded marking stack"); 5793 } 5794 _base = (oop*)(_virtual_space.low()); 5795 _index = 0; 5796 _capacity = new_capacity; 5797 } else if (_failed_double++ == 0 && !CMSConcurrentMTEnabled) { 5798 // Failed to double capacity, continue; 5799 // we print a detail message only once per CMS cycle. 5800 log_debug(gc)(" (benign) Failed to expand marking stack from " SIZE_FORMAT "K to " SIZE_FORMAT "K", 5801 _capacity / K, new_capacity / K); 5802 } 5803 } 5804 5805 5806 // Closures 5807 // XXX: there seems to be a lot of code duplication here; 5808 // should refactor and consolidate common code. 5809 5810 // This closure is used to mark refs into the CMS generation in 5811 // the CMS bit map. Called at the first checkpoint. This closure 5812 // assumes that we do not need to re-mark dirty cards; if the CMS 5813 // generation on which this is used is not an oldest 5814 // generation then this will lose younger_gen cards! 5815 5816 MarkRefsIntoClosure::MarkRefsIntoClosure( 5817 MemRegion span, CMSBitMap* bitMap): 5818 _span(span), 5819 _bitMap(bitMap) 5820 { 5821 assert(ref_processor() == NULL, "deliberately left NULL"); 5822 assert(_bitMap->covers(_span), "_bitMap/_span mismatch"); 5823 } 5824 5825 void MarkRefsIntoClosure::do_oop(oop obj) { 5826 // if p points into _span, then mark corresponding bit in _markBitMap 5827 assert(obj->is_oop(), "expected an oop"); 5828 HeapWord* addr = (HeapWord*)obj; 5829 if (_span.contains(addr)) { 5830 // this should be made more efficient 5831 _bitMap->mark(addr); 5832 } 5833 } 5834 5835 void MarkRefsIntoClosure::do_oop(oop* p) { MarkRefsIntoClosure::do_oop_work(p); } 5836 void MarkRefsIntoClosure::do_oop(narrowOop* p) { MarkRefsIntoClosure::do_oop_work(p); } 5837 5838 ParMarkRefsIntoClosure::ParMarkRefsIntoClosure( 5839 MemRegion span, CMSBitMap* bitMap): 5840 _span(span), 5841 _bitMap(bitMap) 5842 { 5843 assert(ref_processor() == NULL, "deliberately left NULL"); 5844 assert(_bitMap->covers(_span), "_bitMap/_span mismatch"); 5845 } 5846 5847 void ParMarkRefsIntoClosure::do_oop(oop obj) { 5848 // if p points into _span, then mark corresponding bit in _markBitMap 5849 assert(obj->is_oop(), "expected an oop"); 5850 HeapWord* addr = (HeapWord*)obj; 5851 if (_span.contains(addr)) { 5852 // this should be made more efficient 5853 _bitMap->par_mark(addr); 5854 } 5855 } 5856 5857 void ParMarkRefsIntoClosure::do_oop(oop* p) { ParMarkRefsIntoClosure::do_oop_work(p); } 5858 void ParMarkRefsIntoClosure::do_oop(narrowOop* p) { ParMarkRefsIntoClosure::do_oop_work(p); } 5859 5860 // A variant of the above, used for CMS marking verification. 5861 MarkRefsIntoVerifyClosure::MarkRefsIntoVerifyClosure( 5862 MemRegion span, CMSBitMap* verification_bm, CMSBitMap* cms_bm): 5863 _span(span), 5864 _verification_bm(verification_bm), 5865 _cms_bm(cms_bm) 5866 { 5867 assert(ref_processor() == NULL, "deliberately left NULL"); 5868 assert(_verification_bm->covers(_span), "_verification_bm/_span mismatch"); 5869 } 5870 5871 void MarkRefsIntoVerifyClosure::do_oop(oop obj) { 5872 // if p points into _span, then mark corresponding bit in _markBitMap 5873 assert(obj->is_oop(), "expected an oop"); 5874 HeapWord* addr = (HeapWord*)obj; 5875 if (_span.contains(addr)) { 5876 _verification_bm->mark(addr); 5877 if (!_cms_bm->isMarked(addr)) { 5878 Log(gc, verify) log; 5879 ResourceMark rm; 5880 oop(addr)->print_on(log.error_stream()); 5881 log.error(" (" INTPTR_FORMAT " should have been marked)", p2i(addr)); 5882 fatal("... aborting"); 5883 } 5884 } 5885 } 5886 5887 void MarkRefsIntoVerifyClosure::do_oop(oop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); } 5888 void MarkRefsIntoVerifyClosure::do_oop(narrowOop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); } 5889 5890 ////////////////////////////////////////////////// 5891 // MarkRefsIntoAndScanClosure 5892 ////////////////////////////////////////////////// 5893 5894 MarkRefsIntoAndScanClosure::MarkRefsIntoAndScanClosure(MemRegion span, 5895 ReferenceProcessor* rp, 5896 CMSBitMap* bit_map, 5897 CMSBitMap* mod_union_table, 5898 CMSMarkStack* mark_stack, 5899 CMSCollector* collector, 5900 bool should_yield, 5901 bool concurrent_precleaning): 5902 _collector(collector), 5903 _span(span), 5904 _bit_map(bit_map), 5905 _mark_stack(mark_stack), 5906 _pushAndMarkClosure(collector, span, rp, bit_map, mod_union_table, 5907 mark_stack, concurrent_precleaning), 5908 _yield(should_yield), 5909 _concurrent_precleaning(concurrent_precleaning), 5910 _freelistLock(NULL) 5911 { 5912 // FIXME: Should initialize in base class constructor. 5913 assert(rp != NULL, "ref_processor shouldn't be NULL"); 5914 set_ref_processor_internal(rp); 5915 } 5916 5917 // This closure is used to mark refs into the CMS generation at the 5918 // second (final) checkpoint, and to scan and transitively follow 5919 // the unmarked oops. It is also used during the concurrent precleaning 5920 // phase while scanning objects on dirty cards in the CMS generation. 5921 // The marks are made in the marking bit map and the marking stack is 5922 // used for keeping the (newly) grey objects during the scan. 5923 // The parallel version (Par_...) appears further below. 5924 void MarkRefsIntoAndScanClosure::do_oop(oop obj) { 5925 if (obj != NULL) { 5926 assert(obj->is_oop(), "expected an oop"); 5927 HeapWord* addr = (HeapWord*)obj; 5928 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 5929 assert(_collector->overflow_list_is_empty(), 5930 "overflow list should be empty"); 5931 if (_span.contains(addr) && 5932 !_bit_map->isMarked(addr)) { 5933 // mark bit map (object is now grey) 5934 _bit_map->mark(addr); 5935 // push on marking stack (stack should be empty), and drain the 5936 // stack by applying this closure to the oops in the oops popped 5937 // from the stack (i.e. blacken the grey objects) 5938 bool res = _mark_stack->push(obj); 5939 assert(res, "Should have space to push on empty stack"); 5940 do { 5941 oop new_oop = _mark_stack->pop(); 5942 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 5943 assert(_bit_map->isMarked((HeapWord*)new_oop), 5944 "only grey objects on this stack"); 5945 // iterate over the oops in this oop, marking and pushing 5946 // the ones in CMS heap (i.e. in _span). 5947 new_oop->oop_iterate(&_pushAndMarkClosure); 5948 // check if it's time to yield 5949 do_yield_check(); 5950 } while (!_mark_stack->isEmpty() || 5951 (!_concurrent_precleaning && take_from_overflow_list())); 5952 // if marking stack is empty, and we are not doing this 5953 // during precleaning, then check the overflow list 5954 } 5955 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 5956 assert(_collector->overflow_list_is_empty(), 5957 "overflow list was drained above"); 5958 5959 assert(_collector->no_preserved_marks(), 5960 "All preserved marks should have been restored above"); 5961 } 5962 } 5963 5964 void MarkRefsIntoAndScanClosure::do_oop(oop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); } 5965 void MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); } 5966 5967 void MarkRefsIntoAndScanClosure::do_yield_work() { 5968 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 5969 "CMS thread should hold CMS token"); 5970 assert_lock_strong(_freelistLock); 5971 assert_lock_strong(_bit_map->lock()); 5972 // relinquish the free_list_lock and bitMaplock() 5973 _bit_map->lock()->unlock(); 5974 _freelistLock->unlock(); 5975 ConcurrentMarkSweepThread::desynchronize(true); 5976 _collector->stopTimer(); 5977 _collector->incrementYields(); 5978 5979 // See the comment in coordinator_yield() 5980 for (unsigned i = 0; 5981 i < CMSYieldSleepCount && 5982 ConcurrentMarkSweepThread::should_yield() && 5983 !CMSCollector::foregroundGCIsActive(); 5984 ++i) { 5985 os::sleep(Thread::current(), 1, false); 5986 } 5987 5988 ConcurrentMarkSweepThread::synchronize(true); 5989 _freelistLock->lock_without_safepoint_check(); 5990 _bit_map->lock()->lock_without_safepoint_check(); 5991 _collector->startTimer(); 5992 } 5993 5994 /////////////////////////////////////////////////////////// 5995 // ParMarkRefsIntoAndScanClosure: a parallel version of 5996 // MarkRefsIntoAndScanClosure 5997 /////////////////////////////////////////////////////////// 5998 ParMarkRefsIntoAndScanClosure::ParMarkRefsIntoAndScanClosure( 5999 CMSCollector* collector, MemRegion span, ReferenceProcessor* rp, 6000 CMSBitMap* bit_map, OopTaskQueue* work_queue): 6001 _span(span), 6002 _bit_map(bit_map), 6003 _work_queue(work_queue), 6004 _low_water_mark(MIN2((work_queue->max_elems()/4), 6005 ((uint)CMSWorkQueueDrainThreshold * ParallelGCThreads))), 6006 _parPushAndMarkClosure(collector, span, rp, bit_map, work_queue) 6007 { 6008 // FIXME: Should initialize in base class constructor. 6009 assert(rp != NULL, "ref_processor shouldn't be NULL"); 6010 set_ref_processor_internal(rp); 6011 } 6012 6013 // This closure is used to mark refs into the CMS generation at the 6014 // second (final) checkpoint, and to scan and transitively follow 6015 // the unmarked oops. The marks are made in the marking bit map and 6016 // the work_queue is used for keeping the (newly) grey objects during 6017 // the scan phase whence they are also available for stealing by parallel 6018 // threads. Since the marking bit map is shared, updates are 6019 // synchronized (via CAS). 6020 void ParMarkRefsIntoAndScanClosure::do_oop(oop obj) { 6021 if (obj != NULL) { 6022 // Ignore mark word because this could be an already marked oop 6023 // that may be chained at the end of the overflow list. 6024 assert(obj->is_oop(true), "expected an oop"); 6025 HeapWord* addr = (HeapWord*)obj; 6026 if (_span.contains(addr) && 6027 !_bit_map->isMarked(addr)) { 6028 // mark bit map (object will become grey): 6029 // It is possible for several threads to be 6030 // trying to "claim" this object concurrently; 6031 // the unique thread that succeeds in marking the 6032 // object first will do the subsequent push on 6033 // to the work queue (or overflow list). 6034 if (_bit_map->par_mark(addr)) { 6035 // push on work_queue (which may not be empty), and trim the 6036 // queue to an appropriate length by applying this closure to 6037 // the oops in the oops popped from the stack (i.e. blacken the 6038 // grey objects) 6039 bool res = _work_queue->push(obj); 6040 assert(res, "Low water mark should be less than capacity?"); 6041 trim_queue(_low_water_mark); 6042 } // Else, another thread claimed the object 6043 } 6044 } 6045 } 6046 6047 void ParMarkRefsIntoAndScanClosure::do_oop(oop* p) { ParMarkRefsIntoAndScanClosure::do_oop_work(p); } 6048 void ParMarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { ParMarkRefsIntoAndScanClosure::do_oop_work(p); } 6049 6050 // This closure is used to rescan the marked objects on the dirty cards 6051 // in the mod union table and the card table proper. 6052 size_t ScanMarkedObjectsAgainCarefullyClosure::do_object_careful_m( 6053 oop p, MemRegion mr) { 6054 6055 size_t size = 0; 6056 HeapWord* addr = (HeapWord*)p; 6057 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6058 assert(_span.contains(addr), "we are scanning the CMS generation"); 6059 // check if it's time to yield 6060 if (do_yield_check()) { 6061 // We yielded for some foreground stop-world work, 6062 // and we have been asked to abort this ongoing preclean cycle. 6063 return 0; 6064 } 6065 if (_bitMap->isMarked(addr)) { 6066 // it's marked; is it potentially uninitialized? 6067 if (p->klass_or_null_acquire() != NULL) { 6068 // an initialized object; ignore mark word in verification below 6069 // since we are running concurrent with mutators 6070 assert(p->is_oop(true), "should be an oop"); 6071 if (p->is_objArray()) { 6072 // objArrays are precisely marked; restrict scanning 6073 // to dirty cards only. 6074 size = CompactibleFreeListSpace::adjustObjectSize( 6075 p->oop_iterate_size(_scanningClosure, mr)); 6076 } else { 6077 // A non-array may have been imprecisely marked; we need 6078 // to scan object in its entirety. 6079 size = CompactibleFreeListSpace::adjustObjectSize( 6080 p->oop_iterate_size(_scanningClosure)); 6081 } 6082 #ifdef ASSERT 6083 size_t direct_size = 6084 CompactibleFreeListSpace::adjustObjectSize(p->size()); 6085 assert(size == direct_size, "Inconsistency in size"); 6086 assert(size >= 3, "Necessary for Printezis marks to work"); 6087 HeapWord* start_pbit = addr + 1; 6088 HeapWord* end_pbit = addr + size - 1; 6089 assert(_bitMap->isMarked(start_pbit) == _bitMap->isMarked(end_pbit), 6090 "inconsistent Printezis mark"); 6091 // Verify inner mark bits (between Printezis bits) are clear, 6092 // but don't repeat if there are multiple dirty regions for 6093 // the same object, to avoid potential O(N^2) performance. 6094 if (addr != _last_scanned_object) { 6095 _bitMap->verifyNoOneBitsInRange(start_pbit + 1, end_pbit); 6096 _last_scanned_object = addr; 6097 } 6098 #endif // ASSERT 6099 } else { 6100 // An uninitialized object. 6101 assert(_bitMap->isMarked(addr+1), "missing Printezis mark?"); 6102 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 6103 size = pointer_delta(nextOneAddr + 1, addr); 6104 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 6105 "alignment problem"); 6106 // Note that pre-cleaning needn't redirty the card. OopDesc::set_klass() 6107 // will dirty the card when the klass pointer is installed in the 6108 // object (signaling the completion of initialization). 6109 } 6110 } else { 6111 // Either a not yet marked object or an uninitialized object 6112 if (p->klass_or_null_acquire() == NULL) { 6113 // An uninitialized object, skip to the next card, since 6114 // we may not be able to read its P-bits yet. 6115 assert(size == 0, "Initial value"); 6116 } else { 6117 // An object not (yet) reached by marking: we merely need to 6118 // compute its size so as to go look at the next block. 6119 assert(p->is_oop(true), "should be an oop"); 6120 size = CompactibleFreeListSpace::adjustObjectSize(p->size()); 6121 } 6122 } 6123 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6124 return size; 6125 } 6126 6127 void ScanMarkedObjectsAgainCarefullyClosure::do_yield_work() { 6128 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6129 "CMS thread should hold CMS token"); 6130 assert_lock_strong(_freelistLock); 6131 assert_lock_strong(_bitMap->lock()); 6132 // relinquish the free_list_lock and bitMaplock() 6133 _bitMap->lock()->unlock(); 6134 _freelistLock->unlock(); 6135 ConcurrentMarkSweepThread::desynchronize(true); 6136 _collector->stopTimer(); 6137 _collector->incrementYields(); 6138 6139 // See the comment in coordinator_yield() 6140 for (unsigned i = 0; i < CMSYieldSleepCount && 6141 ConcurrentMarkSweepThread::should_yield() && 6142 !CMSCollector::foregroundGCIsActive(); ++i) { 6143 os::sleep(Thread::current(), 1, false); 6144 } 6145 6146 ConcurrentMarkSweepThread::synchronize(true); 6147 _freelistLock->lock_without_safepoint_check(); 6148 _bitMap->lock()->lock_without_safepoint_check(); 6149 _collector->startTimer(); 6150 } 6151 6152 6153 ////////////////////////////////////////////////////////////////// 6154 // SurvivorSpacePrecleanClosure 6155 ////////////////////////////////////////////////////////////////// 6156 // This (single-threaded) closure is used to preclean the oops in 6157 // the survivor spaces. 6158 size_t SurvivorSpacePrecleanClosure::do_object_careful(oop p) { 6159 6160 HeapWord* addr = (HeapWord*)p; 6161 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6162 assert(!_span.contains(addr), "we are scanning the survivor spaces"); 6163 assert(p->klass_or_null() != NULL, "object should be initialized"); 6164 // an initialized object; ignore mark word in verification below 6165 // since we are running concurrent with mutators 6166 assert(p->is_oop(true), "should be an oop"); 6167 // Note that we do not yield while we iterate over 6168 // the interior oops of p, pushing the relevant ones 6169 // on our marking stack. 6170 size_t size = p->oop_iterate_size(_scanning_closure); 6171 do_yield_check(); 6172 // Observe that below, we do not abandon the preclean 6173 // phase as soon as we should; rather we empty the 6174 // marking stack before returning. This is to satisfy 6175 // some existing assertions. In general, it may be a 6176 // good idea to abort immediately and complete the marking 6177 // from the grey objects at a later time. 6178 while (!_mark_stack->isEmpty()) { 6179 oop new_oop = _mark_stack->pop(); 6180 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 6181 assert(_bit_map->isMarked((HeapWord*)new_oop), 6182 "only grey objects on this stack"); 6183 // iterate over the oops in this oop, marking and pushing 6184 // the ones in CMS heap (i.e. in _span). 6185 new_oop->oop_iterate(_scanning_closure); 6186 // check if it's time to yield 6187 do_yield_check(); 6188 } 6189 unsigned int after_count = 6190 GenCollectedHeap::heap()->total_collections(); 6191 bool abort = (_before_count != after_count) || 6192 _collector->should_abort_preclean(); 6193 return abort ? 0 : size; 6194 } 6195 6196 void SurvivorSpacePrecleanClosure::do_yield_work() { 6197 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6198 "CMS thread should hold CMS token"); 6199 assert_lock_strong(_bit_map->lock()); 6200 // Relinquish the bit map lock 6201 _bit_map->lock()->unlock(); 6202 ConcurrentMarkSweepThread::desynchronize(true); 6203 _collector->stopTimer(); 6204 _collector->incrementYields(); 6205 6206 // See the comment in coordinator_yield() 6207 for (unsigned i = 0; i < CMSYieldSleepCount && 6208 ConcurrentMarkSweepThread::should_yield() && 6209 !CMSCollector::foregroundGCIsActive(); ++i) { 6210 os::sleep(Thread::current(), 1, false); 6211 } 6212 6213 ConcurrentMarkSweepThread::synchronize(true); 6214 _bit_map->lock()->lock_without_safepoint_check(); 6215 _collector->startTimer(); 6216 } 6217 6218 // This closure is used to rescan the marked objects on the dirty cards 6219 // in the mod union table and the card table proper. In the parallel 6220 // case, although the bitMap is shared, we do a single read so the 6221 // isMarked() query is "safe". 6222 bool ScanMarkedObjectsAgainClosure::do_object_bm(oop p, MemRegion mr) { 6223 // Ignore mark word because we are running concurrent with mutators 6224 assert(p->is_oop_or_null(true), "Expected an oop or NULL at " PTR_FORMAT, p2i(p)); 6225 HeapWord* addr = (HeapWord*)p; 6226 assert(_span.contains(addr), "we are scanning the CMS generation"); 6227 bool is_obj_array = false; 6228 #ifdef ASSERT 6229 if (!_parallel) { 6230 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 6231 assert(_collector->overflow_list_is_empty(), 6232 "overflow list should be empty"); 6233 6234 } 6235 #endif // ASSERT 6236 if (_bit_map->isMarked(addr)) { 6237 // Obj arrays are precisely marked, non-arrays are not; 6238 // so we scan objArrays precisely and non-arrays in their 6239 // entirety. 6240 if (p->is_objArray()) { 6241 is_obj_array = true; 6242 if (_parallel) { 6243 p->oop_iterate(_par_scan_closure, mr); 6244 } else { 6245 p->oop_iterate(_scan_closure, mr); 6246 } 6247 } else { 6248 if (_parallel) { 6249 p->oop_iterate(_par_scan_closure); 6250 } else { 6251 p->oop_iterate(_scan_closure); 6252 } 6253 } 6254 } 6255 #ifdef ASSERT 6256 if (!_parallel) { 6257 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 6258 assert(_collector->overflow_list_is_empty(), 6259 "overflow list should be empty"); 6260 6261 } 6262 #endif // ASSERT 6263 return is_obj_array; 6264 } 6265 6266 MarkFromRootsClosure::MarkFromRootsClosure(CMSCollector* collector, 6267 MemRegion span, 6268 CMSBitMap* bitMap, CMSMarkStack* markStack, 6269 bool should_yield, bool verifying): 6270 _collector(collector), 6271 _span(span), 6272 _bitMap(bitMap), 6273 _mut(&collector->_modUnionTable), 6274 _markStack(markStack), 6275 _yield(should_yield), 6276 _skipBits(0) 6277 { 6278 assert(_markStack->isEmpty(), "stack should be empty"); 6279 _finger = _bitMap->startWord(); 6280 _threshold = _finger; 6281 assert(_collector->_restart_addr == NULL, "Sanity check"); 6282 assert(_span.contains(_finger), "Out of bounds _finger?"); 6283 DEBUG_ONLY(_verifying = verifying;) 6284 } 6285 6286 void MarkFromRootsClosure::reset(HeapWord* addr) { 6287 assert(_markStack->isEmpty(), "would cause duplicates on stack"); 6288 assert(_span.contains(addr), "Out of bounds _finger?"); 6289 _finger = addr; 6290 _threshold = (HeapWord*)round_to( 6291 (intptr_t)_finger, CardTable::card_size); 6292 } 6293 6294 // Should revisit to see if this should be restructured for 6295 // greater efficiency. 6296 bool MarkFromRootsClosure::do_bit(size_t offset) { 6297 if (_skipBits > 0) { 6298 _skipBits--; 6299 return true; 6300 } 6301 // convert offset into a HeapWord* 6302 HeapWord* addr = _bitMap->startWord() + offset; 6303 assert(_bitMap->endWord() && addr < _bitMap->endWord(), 6304 "address out of range"); 6305 assert(_bitMap->isMarked(addr), "tautology"); 6306 if (_bitMap->isMarked(addr+1)) { 6307 // this is an allocated but not yet initialized object 6308 assert(_skipBits == 0, "tautology"); 6309 _skipBits = 2; // skip next two marked bits ("Printezis-marks") 6310 oop p = oop(addr); 6311 if (p->klass_or_null_acquire() == NULL) { 6312 DEBUG_ONLY(if (!_verifying) {) 6313 // We re-dirty the cards on which this object lies and increase 6314 // the _threshold so that we'll come back to scan this object 6315 // during the preclean or remark phase. (CMSCleanOnEnter) 6316 if (CMSCleanOnEnter) { 6317 size_t sz = _collector->block_size_using_printezis_bits(addr); 6318 HeapWord* end_card_addr = (HeapWord*)round_to( 6319 (intptr_t)(addr+sz), CardTable::card_size); 6320 MemRegion redirty_range = MemRegion(addr, end_card_addr); 6321 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 6322 // Bump _threshold to end_card_addr; note that 6323 // _threshold cannot possibly exceed end_card_addr, anyhow. 6324 // This prevents future clearing of the card as the scan proceeds 6325 // to the right. 6326 assert(_threshold <= end_card_addr, 6327 "Because we are just scanning into this object"); 6328 if (_threshold < end_card_addr) { 6329 _threshold = end_card_addr; 6330 } 6331 if (p->klass_or_null_acquire() != NULL) { 6332 // Redirty the range of cards... 6333 _mut->mark_range(redirty_range); 6334 } // ...else the setting of klass will dirty the card anyway. 6335 } 6336 DEBUG_ONLY(}) 6337 return true; 6338 } 6339 } 6340 scanOopsInOop(addr); 6341 return true; 6342 } 6343 6344 // We take a break if we've been at this for a while, 6345 // so as to avoid monopolizing the locks involved. 6346 void MarkFromRootsClosure::do_yield_work() { 6347 // First give up the locks, then yield, then re-lock 6348 // We should probably use a constructor/destructor idiom to 6349 // do this unlock/lock or modify the MutexUnlocker class to 6350 // serve our purpose. XXX 6351 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6352 "CMS thread should hold CMS token"); 6353 assert_lock_strong(_bitMap->lock()); 6354 _bitMap->lock()->unlock(); 6355 ConcurrentMarkSweepThread::desynchronize(true); 6356 _collector->stopTimer(); 6357 _collector->incrementYields(); 6358 6359 // See the comment in coordinator_yield() 6360 for (unsigned i = 0; i < CMSYieldSleepCount && 6361 ConcurrentMarkSweepThread::should_yield() && 6362 !CMSCollector::foregroundGCIsActive(); ++i) { 6363 os::sleep(Thread::current(), 1, false); 6364 } 6365 6366 ConcurrentMarkSweepThread::synchronize(true); 6367 _bitMap->lock()->lock_without_safepoint_check(); 6368 _collector->startTimer(); 6369 } 6370 6371 void MarkFromRootsClosure::scanOopsInOop(HeapWord* ptr) { 6372 assert(_bitMap->isMarked(ptr), "expected bit to be set"); 6373 assert(_markStack->isEmpty(), 6374 "should drain stack to limit stack usage"); 6375 // convert ptr to an oop preparatory to scanning 6376 oop obj = oop(ptr); 6377 // Ignore mark word in verification below, since we 6378 // may be running concurrent with mutators. 6379 assert(obj->is_oop(true), "should be an oop"); 6380 assert(_finger <= ptr, "_finger runneth ahead"); 6381 // advance the finger to right end of this object 6382 _finger = ptr + obj->size(); 6383 assert(_finger > ptr, "we just incremented it above"); 6384 // On large heaps, it may take us some time to get through 6385 // the marking phase. During 6386 // this time it's possible that a lot of mutations have 6387 // accumulated in the card table and the mod union table -- 6388 // these mutation records are redundant until we have 6389 // actually traced into the corresponding card. 6390 // Here, we check whether advancing the finger would make 6391 // us cross into a new card, and if so clear corresponding 6392 // cards in the MUT (preclean them in the card-table in the 6393 // future). 6394 6395 DEBUG_ONLY(if (!_verifying) {) 6396 // The clean-on-enter optimization is disabled by default, 6397 // until we fix 6178663. 6398 if (CMSCleanOnEnter && (_finger > _threshold)) { 6399 // [_threshold, _finger) represents the interval 6400 // of cards to be cleared in MUT (or precleaned in card table). 6401 // The set of cards to be cleared is all those that overlap 6402 // with the interval [_threshold, _finger); note that 6403 // _threshold is always kept card-aligned but _finger isn't 6404 // always card-aligned. 6405 HeapWord* old_threshold = _threshold; 6406 assert(old_threshold == (HeapWord*)round_to( 6407 (intptr_t)old_threshold, CardTable::card_size), 6408 "_threshold should always be card-aligned"); 6409 _threshold = (HeapWord*)round_to( 6410 (intptr_t)_finger, CardTable::card_size); 6411 MemRegion mr(old_threshold, _threshold); 6412 assert(!mr.is_empty(), "Control point invariant"); 6413 assert(_span.contains(mr), "Should clear within span"); 6414 _mut->clear_range(mr); 6415 } 6416 DEBUG_ONLY(}) 6417 // Note: the finger doesn't advance while we drain 6418 // the stack below. 6419 PushOrMarkClosure pushOrMarkClosure(_collector, 6420 _span, _bitMap, _markStack, 6421 _finger, this); 6422 bool res = _markStack->push(obj); 6423 assert(res, "Empty non-zero size stack should have space for single push"); 6424 while (!_markStack->isEmpty()) { 6425 oop new_oop = _markStack->pop(); 6426 // Skip verifying header mark word below because we are 6427 // running concurrent with mutators. 6428 assert(new_oop->is_oop(true), "Oops! expected to pop an oop"); 6429 // now scan this oop's oops 6430 new_oop->oop_iterate(&pushOrMarkClosure); 6431 do_yield_check(); 6432 } 6433 assert(_markStack->isEmpty(), "tautology, emphasizing post-condition"); 6434 } 6435 6436 ParMarkFromRootsClosure::ParMarkFromRootsClosure(CMSConcMarkingTask* task, 6437 CMSCollector* collector, MemRegion span, 6438 CMSBitMap* bit_map, 6439 OopTaskQueue* work_queue, 6440 CMSMarkStack* overflow_stack): 6441 _collector(collector), 6442 _whole_span(collector->_span), 6443 _span(span), 6444 _bit_map(bit_map), 6445 _mut(&collector->_modUnionTable), 6446 _work_queue(work_queue), 6447 _overflow_stack(overflow_stack), 6448 _skip_bits(0), 6449 _task(task) 6450 { 6451 assert(_work_queue->size() == 0, "work_queue should be empty"); 6452 _finger = span.start(); 6453 _threshold = _finger; // XXX Defer clear-on-enter optimization for now 6454 assert(_span.contains(_finger), "Out of bounds _finger?"); 6455 } 6456 6457 // Should revisit to see if this should be restructured for 6458 // greater efficiency. 6459 bool ParMarkFromRootsClosure::do_bit(size_t offset) { 6460 if (_skip_bits > 0) { 6461 _skip_bits--; 6462 return true; 6463 } 6464 // convert offset into a HeapWord* 6465 HeapWord* addr = _bit_map->startWord() + offset; 6466 assert(_bit_map->endWord() && addr < _bit_map->endWord(), 6467 "address out of range"); 6468 assert(_bit_map->isMarked(addr), "tautology"); 6469 if (_bit_map->isMarked(addr+1)) { 6470 // this is an allocated object that might not yet be initialized 6471 assert(_skip_bits == 0, "tautology"); 6472 _skip_bits = 2; // skip next two marked bits ("Printezis-marks") 6473 oop p = oop(addr); 6474 if (p->klass_or_null_acquire() == NULL) { 6475 // in the case of Clean-on-Enter optimization, redirty card 6476 // and avoid clearing card by increasing the threshold. 6477 return true; 6478 } 6479 } 6480 scan_oops_in_oop(addr); 6481 return true; 6482 } 6483 6484 void ParMarkFromRootsClosure::scan_oops_in_oop(HeapWord* ptr) { 6485 assert(_bit_map->isMarked(ptr), "expected bit to be set"); 6486 // Should we assert that our work queue is empty or 6487 // below some drain limit? 6488 assert(_work_queue->size() == 0, 6489 "should drain stack to limit stack usage"); 6490 // convert ptr to an oop preparatory to scanning 6491 oop obj = oop(ptr); 6492 // Ignore mark word in verification below, since we 6493 // may be running concurrent with mutators. 6494 assert(obj->is_oop(true), "should be an oop"); 6495 assert(_finger <= ptr, "_finger runneth ahead"); 6496 // advance the finger to right end of this object 6497 _finger = ptr + obj->size(); 6498 assert(_finger > ptr, "we just incremented it above"); 6499 // On large heaps, it may take us some time to get through 6500 // the marking phase. During 6501 // this time it's possible that a lot of mutations have 6502 // accumulated in the card table and the mod union table -- 6503 // these mutation records are redundant until we have 6504 // actually traced into the corresponding card. 6505 // Here, we check whether advancing the finger would make 6506 // us cross into a new card, and if so clear corresponding 6507 // cards in the MUT (preclean them in the card-table in the 6508 // future). 6509 6510 // The clean-on-enter optimization is disabled by default, 6511 // until we fix 6178663. 6512 if (CMSCleanOnEnter && (_finger > _threshold)) { 6513 // [_threshold, _finger) represents the interval 6514 // of cards to be cleared in MUT (or precleaned in card table). 6515 // The set of cards to be cleared is all those that overlap 6516 // with the interval [_threshold, _finger); note that 6517 // _threshold is always kept card-aligned but _finger isn't 6518 // always card-aligned. 6519 HeapWord* old_threshold = _threshold; 6520 assert(old_threshold == (HeapWord*)round_to( 6521 (intptr_t)old_threshold, CardTable::card_size), 6522 "_threshold should always be card-aligned"); 6523 _threshold = (HeapWord*)round_to( 6524 (intptr_t)_finger, CardTable::card_size); 6525 MemRegion mr(old_threshold, _threshold); 6526 assert(!mr.is_empty(), "Control point invariant"); 6527 assert(_span.contains(mr), "Should clear within span"); // _whole_span ?? 6528 _mut->clear_range(mr); 6529 } 6530 6531 // Note: the local finger doesn't advance while we drain 6532 // the stack below, but the global finger sure can and will. 6533 HeapWord* volatile* gfa = _task->global_finger_addr(); 6534 ParPushOrMarkClosure pushOrMarkClosure(_collector, 6535 _span, _bit_map, 6536 _work_queue, 6537 _overflow_stack, 6538 _finger, 6539 gfa, this); 6540 bool res = _work_queue->push(obj); // overflow could occur here 6541 assert(res, "Will hold once we use workqueues"); 6542 while (true) { 6543 oop new_oop; 6544 if (!_work_queue->pop_local(new_oop)) { 6545 // We emptied our work_queue; check if there's stuff that can 6546 // be gotten from the overflow stack. 6547 if (CMSConcMarkingTask::get_work_from_overflow_stack( 6548 _overflow_stack, _work_queue)) { 6549 do_yield_check(); 6550 continue; 6551 } else { // done 6552 break; 6553 } 6554 } 6555 // Skip verifying header mark word below because we are 6556 // running concurrent with mutators. 6557 assert(new_oop->is_oop(true), "Oops! expected to pop an oop"); 6558 // now scan this oop's oops 6559 new_oop->oop_iterate(&pushOrMarkClosure); 6560 do_yield_check(); 6561 } 6562 assert(_work_queue->size() == 0, "tautology, emphasizing post-condition"); 6563 } 6564 6565 // Yield in response to a request from VM Thread or 6566 // from mutators. 6567 void ParMarkFromRootsClosure::do_yield_work() { 6568 assert(_task != NULL, "sanity"); 6569 _task->yield(); 6570 } 6571 6572 // A variant of the above used for verifying CMS marking work. 6573 MarkFromRootsVerifyClosure::MarkFromRootsVerifyClosure(CMSCollector* collector, 6574 MemRegion span, 6575 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 6576 CMSMarkStack* mark_stack): 6577 _collector(collector), 6578 _span(span), 6579 _verification_bm(verification_bm), 6580 _cms_bm(cms_bm), 6581 _mark_stack(mark_stack), 6582 _pam_verify_closure(collector, span, verification_bm, cms_bm, 6583 mark_stack) 6584 { 6585 assert(_mark_stack->isEmpty(), "stack should be empty"); 6586 _finger = _verification_bm->startWord(); 6587 assert(_collector->_restart_addr == NULL, "Sanity check"); 6588 assert(_span.contains(_finger), "Out of bounds _finger?"); 6589 } 6590 6591 void MarkFromRootsVerifyClosure::reset(HeapWord* addr) { 6592 assert(_mark_stack->isEmpty(), "would cause duplicates on stack"); 6593 assert(_span.contains(addr), "Out of bounds _finger?"); 6594 _finger = addr; 6595 } 6596 6597 // Should revisit to see if this should be restructured for 6598 // greater efficiency. 6599 bool MarkFromRootsVerifyClosure::do_bit(size_t offset) { 6600 // convert offset into a HeapWord* 6601 HeapWord* addr = _verification_bm->startWord() + offset; 6602 assert(_verification_bm->endWord() && addr < _verification_bm->endWord(), 6603 "address out of range"); 6604 assert(_verification_bm->isMarked(addr), "tautology"); 6605 assert(_cms_bm->isMarked(addr), "tautology"); 6606 6607 assert(_mark_stack->isEmpty(), 6608 "should drain stack to limit stack usage"); 6609 // convert addr to an oop preparatory to scanning 6610 oop obj = oop(addr); 6611 assert(obj->is_oop(), "should be an oop"); 6612 assert(_finger <= addr, "_finger runneth ahead"); 6613 // advance the finger to right end of this object 6614 _finger = addr + obj->size(); 6615 assert(_finger > addr, "we just incremented it above"); 6616 // Note: the finger doesn't advance while we drain 6617 // the stack below. 6618 bool res = _mark_stack->push(obj); 6619 assert(res, "Empty non-zero size stack should have space for single push"); 6620 while (!_mark_stack->isEmpty()) { 6621 oop new_oop = _mark_stack->pop(); 6622 assert(new_oop->is_oop(), "Oops! expected to pop an oop"); 6623 // now scan this oop's oops 6624 new_oop->oop_iterate(&_pam_verify_closure); 6625 } 6626 assert(_mark_stack->isEmpty(), "tautology, emphasizing post-condition"); 6627 return true; 6628 } 6629 6630 PushAndMarkVerifyClosure::PushAndMarkVerifyClosure( 6631 CMSCollector* collector, MemRegion span, 6632 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 6633 CMSMarkStack* mark_stack): 6634 MetadataAwareOopClosure(collector->ref_processor()), 6635 _collector(collector), 6636 _span(span), 6637 _verification_bm(verification_bm), 6638 _cms_bm(cms_bm), 6639 _mark_stack(mark_stack) 6640 { } 6641 6642 void PushAndMarkVerifyClosure::do_oop(oop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 6643 void PushAndMarkVerifyClosure::do_oop(narrowOop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 6644 6645 // Upon stack overflow, we discard (part of) the stack, 6646 // remembering the least address amongst those discarded 6647 // in CMSCollector's _restart_address. 6648 void PushAndMarkVerifyClosure::handle_stack_overflow(HeapWord* lost) { 6649 // Remember the least grey address discarded 6650 HeapWord* ra = (HeapWord*)_mark_stack->least_value(lost); 6651 _collector->lower_restart_addr(ra); 6652 _mark_stack->reset(); // discard stack contents 6653 _mark_stack->expand(); // expand the stack if possible 6654 } 6655 6656 void PushAndMarkVerifyClosure::do_oop(oop obj) { 6657 assert(obj->is_oop_or_null(), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6658 HeapWord* addr = (HeapWord*)obj; 6659 if (_span.contains(addr) && !_verification_bm->isMarked(addr)) { 6660 // Oop lies in _span and isn't yet grey or black 6661 _verification_bm->mark(addr); // now grey 6662 if (!_cms_bm->isMarked(addr)) { 6663 Log(gc, verify) log; 6664 ResourceMark rm; 6665 oop(addr)->print_on(log.error_stream()); 6666 log.error(" (" INTPTR_FORMAT " should have been marked)", p2i(addr)); 6667 fatal("... aborting"); 6668 } 6669 6670 if (!_mark_stack->push(obj)) { // stack overflow 6671 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _mark_stack->capacity()); 6672 assert(_mark_stack->isFull(), "Else push should have succeeded"); 6673 handle_stack_overflow(addr); 6674 } 6675 // anything including and to the right of _finger 6676 // will be scanned as we iterate over the remainder of the 6677 // bit map 6678 } 6679 } 6680 6681 PushOrMarkClosure::PushOrMarkClosure(CMSCollector* collector, 6682 MemRegion span, 6683 CMSBitMap* bitMap, CMSMarkStack* markStack, 6684 HeapWord* finger, MarkFromRootsClosure* parent) : 6685 MetadataAwareOopClosure(collector->ref_processor()), 6686 _collector(collector), 6687 _span(span), 6688 _bitMap(bitMap), 6689 _markStack(markStack), 6690 _finger(finger), 6691 _parent(parent) 6692 { } 6693 6694 ParPushOrMarkClosure::ParPushOrMarkClosure(CMSCollector* collector, 6695 MemRegion span, 6696 CMSBitMap* bit_map, 6697 OopTaskQueue* work_queue, 6698 CMSMarkStack* overflow_stack, 6699 HeapWord* finger, 6700 HeapWord* volatile* global_finger_addr, 6701 ParMarkFromRootsClosure* parent) : 6702 MetadataAwareOopClosure(collector->ref_processor()), 6703 _collector(collector), 6704 _whole_span(collector->_span), 6705 _span(span), 6706 _bit_map(bit_map), 6707 _work_queue(work_queue), 6708 _overflow_stack(overflow_stack), 6709 _finger(finger), 6710 _global_finger_addr(global_finger_addr), 6711 _parent(parent) 6712 { } 6713 6714 // Assumes thread-safe access by callers, who are 6715 // responsible for mutual exclusion. 6716 void CMSCollector::lower_restart_addr(HeapWord* low) { 6717 assert(_span.contains(low), "Out of bounds addr"); 6718 if (_restart_addr == NULL) { 6719 _restart_addr = low; 6720 } else { 6721 _restart_addr = MIN2(_restart_addr, low); 6722 } 6723 } 6724 6725 // Upon stack overflow, we discard (part of) the stack, 6726 // remembering the least address amongst those discarded 6727 // in CMSCollector's _restart_address. 6728 void PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 6729 // Remember the least grey address discarded 6730 HeapWord* ra = (HeapWord*)_markStack->least_value(lost); 6731 _collector->lower_restart_addr(ra); 6732 _markStack->reset(); // discard stack contents 6733 _markStack->expand(); // expand the stack if possible 6734 } 6735 6736 // Upon stack overflow, we discard (part of) the stack, 6737 // remembering the least address amongst those discarded 6738 // in CMSCollector's _restart_address. 6739 void ParPushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 6740 // We need to do this under a mutex to prevent other 6741 // workers from interfering with the work done below. 6742 MutexLockerEx ml(_overflow_stack->par_lock(), 6743 Mutex::_no_safepoint_check_flag); 6744 // Remember the least grey address discarded 6745 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 6746 _collector->lower_restart_addr(ra); 6747 _overflow_stack->reset(); // discard stack contents 6748 _overflow_stack->expand(); // expand the stack if possible 6749 } 6750 6751 void PushOrMarkClosure::do_oop(oop obj) { 6752 // Ignore mark word because we are running concurrent with mutators. 6753 assert(obj->is_oop_or_null(true), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6754 HeapWord* addr = (HeapWord*)obj; 6755 if (_span.contains(addr) && !_bitMap->isMarked(addr)) { 6756 // Oop lies in _span and isn't yet grey or black 6757 _bitMap->mark(addr); // now grey 6758 if (addr < _finger) { 6759 // the bit map iteration has already either passed, or 6760 // sampled, this bit in the bit map; we'll need to 6761 // use the marking stack to scan this oop's oops. 6762 bool simulate_overflow = false; 6763 NOT_PRODUCT( 6764 if (CMSMarkStackOverflowALot && 6765 _collector->simulate_overflow()) { 6766 // simulate a stack overflow 6767 simulate_overflow = true; 6768 } 6769 ) 6770 if (simulate_overflow || !_markStack->push(obj)) { // stack overflow 6771 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _markStack->capacity()); 6772 assert(simulate_overflow || _markStack->isFull(), "Else push should have succeeded"); 6773 handle_stack_overflow(addr); 6774 } 6775 } 6776 // anything including and to the right of _finger 6777 // will be scanned as we iterate over the remainder of the 6778 // bit map 6779 do_yield_check(); 6780 } 6781 } 6782 6783 void PushOrMarkClosure::do_oop(oop* p) { PushOrMarkClosure::do_oop_work(p); } 6784 void PushOrMarkClosure::do_oop(narrowOop* p) { PushOrMarkClosure::do_oop_work(p); } 6785 6786 void ParPushOrMarkClosure::do_oop(oop obj) { 6787 // Ignore mark word because we are running concurrent with mutators. 6788 assert(obj->is_oop_or_null(true), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6789 HeapWord* addr = (HeapWord*)obj; 6790 if (_whole_span.contains(addr) && !_bit_map->isMarked(addr)) { 6791 // Oop lies in _span and isn't yet grey or black 6792 // We read the global_finger (volatile read) strictly after marking oop 6793 bool res = _bit_map->par_mark(addr); // now grey 6794 volatile HeapWord** gfa = (volatile HeapWord**)_global_finger_addr; 6795 // Should we push this marked oop on our stack? 6796 // -- if someone else marked it, nothing to do 6797 // -- if target oop is above global finger nothing to do 6798 // -- if target oop is in chunk and above local finger 6799 // then nothing to do 6800 // -- else push on work queue 6801 if ( !res // someone else marked it, they will deal with it 6802 || (addr >= *gfa) // will be scanned in a later task 6803 || (_span.contains(addr) && addr >= _finger)) { // later in this chunk 6804 return; 6805 } 6806 // the bit map iteration has already either passed, or 6807 // sampled, this bit in the bit map; we'll need to 6808 // use the marking stack to scan this oop's oops. 6809 bool simulate_overflow = false; 6810 NOT_PRODUCT( 6811 if (CMSMarkStackOverflowALot && 6812 _collector->simulate_overflow()) { 6813 // simulate a stack overflow 6814 simulate_overflow = true; 6815 } 6816 ) 6817 if (simulate_overflow || 6818 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) { 6819 // stack overflow 6820 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _overflow_stack->capacity()); 6821 // We cannot assert that the overflow stack is full because 6822 // it may have been emptied since. 6823 assert(simulate_overflow || 6824 _work_queue->size() == _work_queue->max_elems(), 6825 "Else push should have succeeded"); 6826 handle_stack_overflow(addr); 6827 } 6828 do_yield_check(); 6829 } 6830 } 6831 6832 void ParPushOrMarkClosure::do_oop(oop* p) { ParPushOrMarkClosure::do_oop_work(p); } 6833 void ParPushOrMarkClosure::do_oop(narrowOop* p) { ParPushOrMarkClosure::do_oop_work(p); } 6834 6835 PushAndMarkClosure::PushAndMarkClosure(CMSCollector* collector, 6836 MemRegion span, 6837 ReferenceProcessor* rp, 6838 CMSBitMap* bit_map, 6839 CMSBitMap* mod_union_table, 6840 CMSMarkStack* mark_stack, 6841 bool concurrent_precleaning): 6842 MetadataAwareOopClosure(rp), 6843 _collector(collector), 6844 _span(span), 6845 _bit_map(bit_map), 6846 _mod_union_table(mod_union_table), 6847 _mark_stack(mark_stack), 6848 _concurrent_precleaning(concurrent_precleaning) 6849 { 6850 assert(ref_processor() != NULL, "ref_processor shouldn't be NULL"); 6851 } 6852 6853 // Grey object rescan during pre-cleaning and second checkpoint phases -- 6854 // the non-parallel version (the parallel version appears further below.) 6855 void PushAndMarkClosure::do_oop(oop obj) { 6856 // Ignore mark word verification. If during concurrent precleaning, 6857 // the object monitor may be locked. If during the checkpoint 6858 // phases, the object may already have been reached by a different 6859 // path and may be at the end of the global overflow list (so 6860 // the mark word may be NULL). 6861 assert(obj->is_oop_or_null(true /* ignore mark word */), 6862 "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6863 HeapWord* addr = (HeapWord*)obj; 6864 // Check if oop points into the CMS generation 6865 // and is not marked 6866 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 6867 // a white object ... 6868 _bit_map->mark(addr); // ... now grey 6869 // push on the marking stack (grey set) 6870 bool simulate_overflow = false; 6871 NOT_PRODUCT( 6872 if (CMSMarkStackOverflowALot && 6873 _collector->simulate_overflow()) { 6874 // simulate a stack overflow 6875 simulate_overflow = true; 6876 } 6877 ) 6878 if (simulate_overflow || !_mark_stack->push(obj)) { 6879 if (_concurrent_precleaning) { 6880 // During precleaning we can just dirty the appropriate card(s) 6881 // in the mod union table, thus ensuring that the object remains 6882 // in the grey set and continue. In the case of object arrays 6883 // we need to dirty all of the cards that the object spans, 6884 // since the rescan of object arrays will be limited to the 6885 // dirty cards. 6886 // Note that no one can be interfering with us in this action 6887 // of dirtying the mod union table, so no locking or atomics 6888 // are required. 6889 if (obj->is_objArray()) { 6890 size_t sz = obj->size(); 6891 HeapWord* end_card_addr = (HeapWord*)round_to( 6892 (intptr_t)(addr+sz), CardTable::card_size); 6893 MemRegion redirty_range = MemRegion(addr, end_card_addr); 6894 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 6895 _mod_union_table->mark_range(redirty_range); 6896 } else { 6897 _mod_union_table->mark(addr); 6898 } 6899 _collector->_ser_pmc_preclean_ovflw++; 6900 } else { 6901 // During the remark phase, we need to remember this oop 6902 // in the overflow list. 6903 _collector->push_on_overflow_list(obj); 6904 _collector->_ser_pmc_remark_ovflw++; 6905 } 6906 } 6907 } 6908 } 6909 6910 ParPushAndMarkClosure::ParPushAndMarkClosure(CMSCollector* collector, 6911 MemRegion span, 6912 ReferenceProcessor* rp, 6913 CMSBitMap* bit_map, 6914 OopTaskQueue* work_queue): 6915 MetadataAwareOopClosure(rp), 6916 _collector(collector), 6917 _span(span), 6918 _bit_map(bit_map), 6919 _work_queue(work_queue) 6920 { 6921 assert(ref_processor() != NULL, "ref_processor shouldn't be NULL"); 6922 } 6923 6924 void PushAndMarkClosure::do_oop(oop* p) { PushAndMarkClosure::do_oop_work(p); } 6925 void PushAndMarkClosure::do_oop(narrowOop* p) { PushAndMarkClosure::do_oop_work(p); } 6926 6927 // Grey object rescan during second checkpoint phase -- 6928 // the parallel version. 6929 void ParPushAndMarkClosure::do_oop(oop obj) { 6930 // In the assert below, we ignore the mark word because 6931 // this oop may point to an already visited object that is 6932 // on the overflow stack (in which case the mark word has 6933 // been hijacked for chaining into the overflow stack -- 6934 // if this is the last object in the overflow stack then 6935 // its mark word will be NULL). Because this object may 6936 // have been subsequently popped off the global overflow 6937 // stack, and the mark word possibly restored to the prototypical 6938 // value, by the time we get to examined this failing assert in 6939 // the debugger, is_oop_or_null(false) may subsequently start 6940 // to hold. 6941 assert(obj->is_oop_or_null(true), 6942 "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6943 HeapWord* addr = (HeapWord*)obj; 6944 // Check if oop points into the CMS generation 6945 // and is not marked 6946 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 6947 // a white object ... 6948 // If we manage to "claim" the object, by being the 6949 // first thread to mark it, then we push it on our 6950 // marking stack 6951 if (_bit_map->par_mark(addr)) { // ... now grey 6952 // push on work queue (grey set) 6953 bool simulate_overflow = false; 6954 NOT_PRODUCT( 6955 if (CMSMarkStackOverflowALot && 6956 _collector->par_simulate_overflow()) { 6957 // simulate a stack overflow 6958 simulate_overflow = true; 6959 } 6960 ) 6961 if (simulate_overflow || !_work_queue->push(obj)) { 6962 _collector->par_push_on_overflow_list(obj); 6963 _collector->_par_pmc_remark_ovflw++; // imprecise OK: no need to CAS 6964 } 6965 } // Else, some other thread got there first 6966 } 6967 } 6968 6969 void ParPushAndMarkClosure::do_oop(oop* p) { ParPushAndMarkClosure::do_oop_work(p); } 6970 void ParPushAndMarkClosure::do_oop(narrowOop* p) { ParPushAndMarkClosure::do_oop_work(p); } 6971 6972 void CMSPrecleanRefsYieldClosure::do_yield_work() { 6973 Mutex* bml = _collector->bitMapLock(); 6974 assert_lock_strong(bml); 6975 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6976 "CMS thread should hold CMS token"); 6977 6978 bml->unlock(); 6979 ConcurrentMarkSweepThread::desynchronize(true); 6980 6981 _collector->stopTimer(); 6982 _collector->incrementYields(); 6983 6984 // See the comment in coordinator_yield() 6985 for (unsigned i = 0; i < CMSYieldSleepCount && 6986 ConcurrentMarkSweepThread::should_yield() && 6987 !CMSCollector::foregroundGCIsActive(); ++i) { 6988 os::sleep(Thread::current(), 1, false); 6989 } 6990 6991 ConcurrentMarkSweepThread::synchronize(true); 6992 bml->lock(); 6993 6994 _collector->startTimer(); 6995 } 6996 6997 bool CMSPrecleanRefsYieldClosure::should_return() { 6998 if (ConcurrentMarkSweepThread::should_yield()) { 6999 do_yield_work(); 7000 } 7001 return _collector->foregroundGCIsActive(); 7002 } 7003 7004 void MarkFromDirtyCardsClosure::do_MemRegion(MemRegion mr) { 7005 assert(((size_t)mr.start())%CardTable::card_size_in_words == 0, 7006 "mr should be aligned to start at a card boundary"); 7007 // We'd like to assert: 7008 // assert(mr.word_size()%CardTable::card_size_in_words == 0, 7009 // "mr should be a range of cards"); 7010 // However, that would be too strong in one case -- the last 7011 // partition ends at _unallocated_block which, in general, can be 7012 // an arbitrary boundary, not necessarily card aligned. 7013 _num_dirty_cards += mr.word_size()/CardTable::card_size_in_words; 7014 _space->object_iterate_mem(mr, &_scan_cl); 7015 } 7016 7017 SweepClosure::SweepClosure(CMSCollector* collector, 7018 ConcurrentMarkSweepGeneration* g, 7019 CMSBitMap* bitMap, bool should_yield) : 7020 _collector(collector), 7021 _g(g), 7022 _sp(g->cmsSpace()), 7023 _limit(_sp->sweep_limit()), 7024 _freelistLock(_sp->freelistLock()), 7025 _bitMap(bitMap), 7026 _yield(should_yield), 7027 _inFreeRange(false), // No free range at beginning of sweep 7028 _freeRangeInFreeLists(false), // No free range at beginning of sweep 7029 _lastFreeRangeCoalesced(false), 7030 _freeFinger(g->used_region().start()) 7031 { 7032 NOT_PRODUCT( 7033 _numObjectsFreed = 0; 7034 _numWordsFreed = 0; 7035 _numObjectsLive = 0; 7036 _numWordsLive = 0; 7037 _numObjectsAlreadyFree = 0; 7038 _numWordsAlreadyFree = 0; 7039 _last_fc = NULL; 7040 7041 _sp->initializeIndexedFreeListArrayReturnedBytes(); 7042 _sp->dictionary()->initialize_dict_returned_bytes(); 7043 ) 7044 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7045 "sweep _limit out of bounds"); 7046 log_develop_trace(gc, sweep)("===================="); 7047 log_develop_trace(gc, sweep)("Starting new sweep with limit " PTR_FORMAT, p2i(_limit)); 7048 } 7049 7050 void SweepClosure::print_on(outputStream* st) const { 7051 st->print_cr("_sp = [" PTR_FORMAT "," PTR_FORMAT ")", 7052 p2i(_sp->bottom()), p2i(_sp->end())); 7053 st->print_cr("_limit = " PTR_FORMAT, p2i(_limit)); 7054 st->print_cr("_freeFinger = " PTR_FORMAT, p2i(_freeFinger)); 7055 NOT_PRODUCT(st->print_cr("_last_fc = " PTR_FORMAT, p2i(_last_fc));) 7056 st->print_cr("_inFreeRange = %d, _freeRangeInFreeLists = %d, _lastFreeRangeCoalesced = %d", 7057 _inFreeRange, _freeRangeInFreeLists, _lastFreeRangeCoalesced); 7058 } 7059 7060 #ifndef PRODUCT 7061 // Assertion checking only: no useful work in product mode -- 7062 // however, if any of the flags below become product flags, 7063 // you may need to review this code to see if it needs to be 7064 // enabled in product mode. 7065 SweepClosure::~SweepClosure() { 7066 assert_lock_strong(_freelistLock); 7067 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7068 "sweep _limit out of bounds"); 7069 if (inFreeRange()) { 7070 Log(gc, sweep) log; 7071 log.error("inFreeRange() should have been reset; dumping state of SweepClosure"); 7072 ResourceMark rm; 7073 print_on(log.error_stream()); 7074 ShouldNotReachHere(); 7075 } 7076 7077 if (log_is_enabled(Debug, gc, sweep)) { 7078 log_debug(gc, sweep)("Collected " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes", 7079 _numObjectsFreed, _numWordsFreed*sizeof(HeapWord)); 7080 log_debug(gc, sweep)("Live " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes Already free " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes", 7081 _numObjectsLive, _numWordsLive*sizeof(HeapWord), _numObjectsAlreadyFree, _numWordsAlreadyFree*sizeof(HeapWord)); 7082 size_t totalBytes = (_numWordsFreed + _numWordsLive + _numWordsAlreadyFree) * sizeof(HeapWord); 7083 log_debug(gc, sweep)("Total sweep: " SIZE_FORMAT " bytes", totalBytes); 7084 } 7085 7086 if (log_is_enabled(Trace, gc, sweep) && CMSVerifyReturnedBytes) { 7087 size_t indexListReturnedBytes = _sp->sumIndexedFreeListArrayReturnedBytes(); 7088 size_t dict_returned_bytes = _sp->dictionary()->sum_dict_returned_bytes(); 7089 size_t returned_bytes = indexListReturnedBytes + dict_returned_bytes; 7090 log_trace(gc, sweep)("Returned " SIZE_FORMAT " bytes Indexed List Returned " SIZE_FORMAT " bytes Dictionary Returned " SIZE_FORMAT " bytes", 7091 returned_bytes, indexListReturnedBytes, dict_returned_bytes); 7092 } 7093 log_develop_trace(gc, sweep)("end of sweep with _limit = " PTR_FORMAT, p2i(_limit)); 7094 log_develop_trace(gc, sweep)("================"); 7095 } 7096 #endif // PRODUCT 7097 7098 void SweepClosure::initialize_free_range(HeapWord* freeFinger, 7099 bool freeRangeInFreeLists) { 7100 log_develop_trace(gc, sweep)("---- Start free range at " PTR_FORMAT " with free block (%d)", 7101 p2i(freeFinger), freeRangeInFreeLists); 7102 assert(!inFreeRange(), "Trampling existing free range"); 7103 set_inFreeRange(true); 7104 set_lastFreeRangeCoalesced(false); 7105 7106 set_freeFinger(freeFinger); 7107 set_freeRangeInFreeLists(freeRangeInFreeLists); 7108 if (CMSTestInFreeList) { 7109 if (freeRangeInFreeLists) { 7110 FreeChunk* fc = (FreeChunk*) freeFinger; 7111 assert(fc->is_free(), "A chunk on the free list should be free."); 7112 assert(fc->size() > 0, "Free range should have a size"); 7113 assert(_sp->verify_chunk_in_free_list(fc), "Chunk is not in free lists"); 7114 } 7115 } 7116 } 7117 7118 // Note that the sweeper runs concurrently with mutators. Thus, 7119 // it is possible for direct allocation in this generation to happen 7120 // in the middle of the sweep. Note that the sweeper also coalesces 7121 // contiguous free blocks. Thus, unless the sweeper and the allocator 7122 // synchronize appropriately freshly allocated blocks may get swept up. 7123 // This is accomplished by the sweeper locking the free lists while 7124 // it is sweeping. Thus blocks that are determined to be free are 7125 // indeed free. There is however one additional complication: 7126 // blocks that have been allocated since the final checkpoint and 7127 // mark, will not have been marked and so would be treated as 7128 // unreachable and swept up. To prevent this, the allocator marks 7129 // the bit map when allocating during the sweep phase. This leads, 7130 // however, to a further complication -- objects may have been allocated 7131 // but not yet initialized -- in the sense that the header isn't yet 7132 // installed. The sweeper can not then determine the size of the block 7133 // in order to skip over it. To deal with this case, we use a technique 7134 // (due to Printezis) to encode such uninitialized block sizes in the 7135 // bit map. Since the bit map uses a bit per every HeapWord, but the 7136 // CMS generation has a minimum object size of 3 HeapWords, it follows 7137 // that "normal marks" won't be adjacent in the bit map (there will 7138 // always be at least two 0 bits between successive 1 bits). We make use 7139 // of these "unused" bits to represent uninitialized blocks -- the bit 7140 // corresponding to the start of the uninitialized object and the next 7141 // bit are both set. Finally, a 1 bit marks the end of the object that 7142 // started with the two consecutive 1 bits to indicate its potentially 7143 // uninitialized state. 7144 7145 size_t SweepClosure::do_blk_careful(HeapWord* addr) { 7146 FreeChunk* fc = (FreeChunk*)addr; 7147 size_t res; 7148 7149 // Check if we are done sweeping. Below we check "addr >= _limit" rather 7150 // than "addr == _limit" because although _limit was a block boundary when 7151 // we started the sweep, it may no longer be one because heap expansion 7152 // may have caused us to coalesce the block ending at the address _limit 7153 // with a newly expanded chunk (this happens when _limit was set to the 7154 // previous _end of the space), so we may have stepped past _limit: 7155 // see the following Zeno-like trail of CRs 6977970, 7008136, 7042740. 7156 if (addr >= _limit) { // we have swept up to or past the limit: finish up 7157 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7158 "sweep _limit out of bounds"); 7159 assert(addr < _sp->end(), "addr out of bounds"); 7160 // Flush any free range we might be holding as a single 7161 // coalesced chunk to the appropriate free list. 7162 if (inFreeRange()) { 7163 assert(freeFinger() >= _sp->bottom() && freeFinger() < _limit, 7164 "freeFinger() " PTR_FORMAT " is out-of-bounds", p2i(freeFinger())); 7165 flush_cur_free_chunk(freeFinger(), 7166 pointer_delta(addr, freeFinger())); 7167 log_develop_trace(gc, sweep)("Sweep: last chunk: put_free_blk " PTR_FORMAT " (" SIZE_FORMAT ") [coalesced:%d]", 7168 p2i(freeFinger()), pointer_delta(addr, freeFinger()), 7169 lastFreeRangeCoalesced() ? 1 : 0); 7170 } 7171 7172 // help the iterator loop finish 7173 return pointer_delta(_sp->end(), addr); 7174 } 7175 7176 assert(addr < _limit, "sweep invariant"); 7177 // check if we should yield 7178 do_yield_check(addr); 7179 if (fc->is_free()) { 7180 // Chunk that is already free 7181 res = fc->size(); 7182 do_already_free_chunk(fc); 7183 debug_only(_sp->verifyFreeLists()); 7184 // If we flush the chunk at hand in lookahead_and_flush() 7185 // and it's coalesced with a preceding chunk, then the 7186 // process of "mangling" the payload of the coalesced block 7187 // will cause erasure of the size information from the 7188 // (erstwhile) header of all the coalesced blocks but the 7189 // first, so the first disjunct in the assert will not hold 7190 // in that specific case (in which case the second disjunct 7191 // will hold). 7192 assert(res == fc->size() || ((HeapWord*)fc) + res >= _limit, 7193 "Otherwise the size info doesn't change at this step"); 7194 NOT_PRODUCT( 7195 _numObjectsAlreadyFree++; 7196 _numWordsAlreadyFree += res; 7197 ) 7198 NOT_PRODUCT(_last_fc = fc;) 7199 } else if (!_bitMap->isMarked(addr)) { 7200 // Chunk is fresh garbage 7201 res = do_garbage_chunk(fc); 7202 debug_only(_sp->verifyFreeLists()); 7203 NOT_PRODUCT( 7204 _numObjectsFreed++; 7205 _numWordsFreed += res; 7206 ) 7207 } else { 7208 // Chunk that is alive. 7209 res = do_live_chunk(fc); 7210 debug_only(_sp->verifyFreeLists()); 7211 NOT_PRODUCT( 7212 _numObjectsLive++; 7213 _numWordsLive += res; 7214 ) 7215 } 7216 return res; 7217 } 7218 7219 // For the smart allocation, record following 7220 // split deaths - a free chunk is removed from its free list because 7221 // it is being split into two or more chunks. 7222 // split birth - a free chunk is being added to its free list because 7223 // a larger free chunk has been split and resulted in this free chunk. 7224 // coal death - a free chunk is being removed from its free list because 7225 // it is being coalesced into a large free chunk. 7226 // coal birth - a free chunk is being added to its free list because 7227 // it was created when two or more free chunks where coalesced into 7228 // this free chunk. 7229 // 7230 // These statistics are used to determine the desired number of free 7231 // chunks of a given size. The desired number is chosen to be relative 7232 // to the end of a CMS sweep. The desired number at the end of a sweep 7233 // is the 7234 // count-at-end-of-previous-sweep (an amount that was enough) 7235 // - count-at-beginning-of-current-sweep (the excess) 7236 // + split-births (gains in this size during interval) 7237 // - split-deaths (demands on this size during interval) 7238 // where the interval is from the end of one sweep to the end of the 7239 // next. 7240 // 7241 // When sweeping the sweeper maintains an accumulated chunk which is 7242 // the chunk that is made up of chunks that have been coalesced. That 7243 // will be termed the left-hand chunk. A new chunk of garbage that 7244 // is being considered for coalescing will be referred to as the 7245 // right-hand chunk. 7246 // 7247 // When making a decision on whether to coalesce a right-hand chunk with 7248 // the current left-hand chunk, the current count vs. the desired count 7249 // of the left-hand chunk is considered. Also if the right-hand chunk 7250 // is near the large chunk at the end of the heap (see 7251 // ConcurrentMarkSweepGeneration::isNearLargestChunk()), then the 7252 // left-hand chunk is coalesced. 7253 // 7254 // When making a decision about whether to split a chunk, the desired count 7255 // vs. the current count of the candidate to be split is also considered. 7256 // If the candidate is underpopulated (currently fewer chunks than desired) 7257 // a chunk of an overpopulated (currently more chunks than desired) size may 7258 // be chosen. The "hint" associated with a free list, if non-null, points 7259 // to a free list which may be overpopulated. 7260 // 7261 7262 void SweepClosure::do_already_free_chunk(FreeChunk* fc) { 7263 const size_t size = fc->size(); 7264 // Chunks that cannot be coalesced are not in the 7265 // free lists. 7266 if (CMSTestInFreeList && !fc->cantCoalesce()) { 7267 assert(_sp->verify_chunk_in_free_list(fc), 7268 "free chunk should be in free lists"); 7269 } 7270 // a chunk that is already free, should not have been 7271 // marked in the bit map 7272 HeapWord* const addr = (HeapWord*) fc; 7273 assert(!_bitMap->isMarked(addr), "free chunk should be unmarked"); 7274 // Verify that the bit map has no bits marked between 7275 // addr and purported end of this block. 7276 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 7277 7278 // Some chunks cannot be coalesced under any circumstances. 7279 // See the definition of cantCoalesce(). 7280 if (!fc->cantCoalesce()) { 7281 // This chunk can potentially be coalesced. 7282 // All the work is done in 7283 do_post_free_or_garbage_chunk(fc, size); 7284 // Note that if the chunk is not coalescable (the else arm 7285 // below), we unconditionally flush, without needing to do 7286 // a "lookahead," as we do below. 7287 if (inFreeRange()) lookahead_and_flush(fc, size); 7288 } else { 7289 // Code path common to both original and adaptive free lists. 7290 7291 // cant coalesce with previous block; this should be treated 7292 // as the end of a free run if any 7293 if (inFreeRange()) { 7294 // we kicked some butt; time to pick up the garbage 7295 assert(freeFinger() < addr, "freeFinger points too high"); 7296 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 7297 } 7298 // else, nothing to do, just continue 7299 } 7300 } 7301 7302 size_t SweepClosure::do_garbage_chunk(FreeChunk* fc) { 7303 // This is a chunk of garbage. It is not in any free list. 7304 // Add it to a free list or let it possibly be coalesced into 7305 // a larger chunk. 7306 HeapWord* const addr = (HeapWord*) fc; 7307 const size_t size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 7308 7309 // Verify that the bit map has no bits marked between 7310 // addr and purported end of just dead object. 7311 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 7312 do_post_free_or_garbage_chunk(fc, size); 7313 7314 assert(_limit >= addr + size, 7315 "A freshly garbage chunk can't possibly straddle over _limit"); 7316 if (inFreeRange()) lookahead_and_flush(fc, size); 7317 return size; 7318 } 7319 7320 size_t SweepClosure::do_live_chunk(FreeChunk* fc) { 7321 HeapWord* addr = (HeapWord*) fc; 7322 // The sweeper has just found a live object. Return any accumulated 7323 // left hand chunk to the free lists. 7324 if (inFreeRange()) { 7325 assert(freeFinger() < addr, "freeFinger points too high"); 7326 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 7327 } 7328 7329 // This object is live: we'd normally expect this to be 7330 // an oop, and like to assert the following: 7331 // assert(oop(addr)->is_oop(), "live block should be an oop"); 7332 // However, as we commented above, this may be an object whose 7333 // header hasn't yet been initialized. 7334 size_t size; 7335 assert(_bitMap->isMarked(addr), "Tautology for this control point"); 7336 if (_bitMap->isMarked(addr + 1)) { 7337 // Determine the size from the bit map, rather than trying to 7338 // compute it from the object header. 7339 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 7340 size = pointer_delta(nextOneAddr + 1, addr); 7341 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 7342 "alignment problem"); 7343 7344 #ifdef ASSERT 7345 if (oop(addr)->klass_or_null_acquire() != NULL) { 7346 // Ignore mark word because we are running concurrent with mutators 7347 assert(oop(addr)->is_oop(true), "live block should be an oop"); 7348 assert(size == 7349 CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()), 7350 "P-mark and computed size do not agree"); 7351 } 7352 #endif 7353 7354 } else { 7355 // This should be an initialized object that's alive. 7356 assert(oop(addr)->klass_or_null_acquire() != NULL, 7357 "Should be an initialized object"); 7358 // Ignore mark word because we are running concurrent with mutators 7359 assert(oop(addr)->is_oop(true), "live block should be an oop"); 7360 // Verify that the bit map has no bits marked between 7361 // addr and purported end of this block. 7362 size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 7363 assert(size >= 3, "Necessary for Printezis marks to work"); 7364 assert(!_bitMap->isMarked(addr+1), "Tautology for this control point"); 7365 DEBUG_ONLY(_bitMap->verifyNoOneBitsInRange(addr+2, addr+size);) 7366 } 7367 return size; 7368 } 7369 7370 void SweepClosure::do_post_free_or_garbage_chunk(FreeChunk* fc, 7371 size_t chunkSize) { 7372 // do_post_free_or_garbage_chunk() should only be called in the case 7373 // of the adaptive free list allocator. 7374 const bool fcInFreeLists = fc->is_free(); 7375 assert((HeapWord*)fc <= _limit, "sweep invariant"); 7376 if (CMSTestInFreeList && fcInFreeLists) { 7377 assert(_sp->verify_chunk_in_free_list(fc), "free chunk is not in free lists"); 7378 } 7379 7380 log_develop_trace(gc, sweep)(" -- pick up another chunk at " PTR_FORMAT " (" SIZE_FORMAT ")", p2i(fc), chunkSize); 7381 7382 HeapWord* const fc_addr = (HeapWord*) fc; 7383 7384 bool coalesce = false; 7385 const size_t left = pointer_delta(fc_addr, freeFinger()); 7386 const size_t right = chunkSize; 7387 switch (FLSCoalescePolicy) { 7388 // numeric value forms a coalition aggressiveness metric 7389 case 0: { // never coalesce 7390 coalesce = false; 7391 break; 7392 } 7393 case 1: { // coalesce if left & right chunks on overpopulated lists 7394 coalesce = _sp->coalOverPopulated(left) && 7395 _sp->coalOverPopulated(right); 7396 break; 7397 } 7398 case 2: { // coalesce if left chunk on overpopulated list (default) 7399 coalesce = _sp->coalOverPopulated(left); 7400 break; 7401 } 7402 case 3: { // coalesce if left OR right chunk on overpopulated list 7403 coalesce = _sp->coalOverPopulated(left) || 7404 _sp->coalOverPopulated(right); 7405 break; 7406 } 7407 case 4: { // always coalesce 7408 coalesce = true; 7409 break; 7410 } 7411 default: 7412 ShouldNotReachHere(); 7413 } 7414 7415 // Should the current free range be coalesced? 7416 // If the chunk is in a free range and either we decided to coalesce above 7417 // or the chunk is near the large block at the end of the heap 7418 // (isNearLargestChunk() returns true), then coalesce this chunk. 7419 const bool doCoalesce = inFreeRange() 7420 && (coalesce || _g->isNearLargestChunk(fc_addr)); 7421 if (doCoalesce) { 7422 // Coalesce the current free range on the left with the new 7423 // chunk on the right. If either is on a free list, 7424 // it must be removed from the list and stashed in the closure. 7425 if (freeRangeInFreeLists()) { 7426 FreeChunk* const ffc = (FreeChunk*)freeFinger(); 7427 assert(ffc->size() == pointer_delta(fc_addr, freeFinger()), 7428 "Size of free range is inconsistent with chunk size."); 7429 if (CMSTestInFreeList) { 7430 assert(_sp->verify_chunk_in_free_list(ffc), 7431 "Chunk is not in free lists"); 7432 } 7433 _sp->coalDeath(ffc->size()); 7434 _sp->removeFreeChunkFromFreeLists(ffc); 7435 set_freeRangeInFreeLists(false); 7436 } 7437 if (fcInFreeLists) { 7438 _sp->coalDeath(chunkSize); 7439 assert(fc->size() == chunkSize, 7440 "The chunk has the wrong size or is not in the free lists"); 7441 _sp->removeFreeChunkFromFreeLists(fc); 7442 } 7443 set_lastFreeRangeCoalesced(true); 7444 print_free_block_coalesced(fc); 7445 } else { // not in a free range and/or should not coalesce 7446 // Return the current free range and start a new one. 7447 if (inFreeRange()) { 7448 // In a free range but cannot coalesce with the right hand chunk. 7449 // Put the current free range into the free lists. 7450 flush_cur_free_chunk(freeFinger(), 7451 pointer_delta(fc_addr, freeFinger())); 7452 } 7453 // Set up for new free range. Pass along whether the right hand 7454 // chunk is in the free lists. 7455 initialize_free_range((HeapWord*)fc, fcInFreeLists); 7456 } 7457 } 7458 7459 // Lookahead flush: 7460 // If we are tracking a free range, and this is the last chunk that 7461 // we'll look at because its end crosses past _limit, we'll preemptively 7462 // flush it along with any free range we may be holding on to. Note that 7463 // this can be the case only for an already free or freshly garbage 7464 // chunk. If this block is an object, it can never straddle 7465 // over _limit. The "straddling" occurs when _limit is set at 7466 // the previous end of the space when this cycle started, and 7467 // a subsequent heap expansion caused the previously co-terminal 7468 // free block to be coalesced with the newly expanded portion, 7469 // thus rendering _limit a non-block-boundary making it dangerous 7470 // for the sweeper to step over and examine. 7471 void SweepClosure::lookahead_and_flush(FreeChunk* fc, size_t chunk_size) { 7472 assert(inFreeRange(), "Should only be called if currently in a free range."); 7473 HeapWord* const eob = ((HeapWord*)fc) + chunk_size; 7474 assert(_sp->used_region().contains(eob - 1), 7475 "eob = " PTR_FORMAT " eob-1 = " PTR_FORMAT " _limit = " PTR_FORMAT 7476 " out of bounds wrt _sp = [" PTR_FORMAT "," PTR_FORMAT ")" 7477 " when examining fc = " PTR_FORMAT "(" SIZE_FORMAT ")", 7478 p2i(eob), p2i(eob-1), p2i(_limit), p2i(_sp->bottom()), p2i(_sp->end()), p2i(fc), chunk_size); 7479 if (eob >= _limit) { 7480 assert(eob == _limit || fc->is_free(), "Only a free chunk should allow us to cross over the limit"); 7481 log_develop_trace(gc, sweep)("_limit " PTR_FORMAT " reached or crossed by block " 7482 "[" PTR_FORMAT "," PTR_FORMAT ") in space " 7483 "[" PTR_FORMAT "," PTR_FORMAT ")", 7484 p2i(_limit), p2i(fc), p2i(eob), p2i(_sp->bottom()), p2i(_sp->end())); 7485 // Return the storage we are tracking back into the free lists. 7486 log_develop_trace(gc, sweep)("Flushing ... "); 7487 assert(freeFinger() < eob, "Error"); 7488 flush_cur_free_chunk( freeFinger(), pointer_delta(eob, freeFinger())); 7489 } 7490 } 7491 7492 void SweepClosure::flush_cur_free_chunk(HeapWord* chunk, size_t size) { 7493 assert(inFreeRange(), "Should only be called if currently in a free range."); 7494 assert(size > 0, 7495 "A zero sized chunk cannot be added to the free lists."); 7496 if (!freeRangeInFreeLists()) { 7497 if (CMSTestInFreeList) { 7498 FreeChunk* fc = (FreeChunk*) chunk; 7499 fc->set_size(size); 7500 assert(!_sp->verify_chunk_in_free_list(fc), 7501 "chunk should not be in free lists yet"); 7502 } 7503 log_develop_trace(gc, sweep)(" -- add free block " PTR_FORMAT " (" SIZE_FORMAT ") to free lists", p2i(chunk), size); 7504 // A new free range is going to be starting. The current 7505 // free range has not been added to the free lists yet or 7506 // was removed so add it back. 7507 // If the current free range was coalesced, then the death 7508 // of the free range was recorded. Record a birth now. 7509 if (lastFreeRangeCoalesced()) { 7510 _sp->coalBirth(size); 7511 } 7512 _sp->addChunkAndRepairOffsetTable(chunk, size, 7513 lastFreeRangeCoalesced()); 7514 } else { 7515 log_develop_trace(gc, sweep)("Already in free list: nothing to flush"); 7516 } 7517 set_inFreeRange(false); 7518 set_freeRangeInFreeLists(false); 7519 } 7520 7521 // We take a break if we've been at this for a while, 7522 // so as to avoid monopolizing the locks involved. 7523 void SweepClosure::do_yield_work(HeapWord* addr) { 7524 // Return current free chunk being used for coalescing (if any) 7525 // to the appropriate freelist. After yielding, the next 7526 // free block encountered will start a coalescing range of 7527 // free blocks. If the next free block is adjacent to the 7528 // chunk just flushed, they will need to wait for the next 7529 // sweep to be coalesced. 7530 if (inFreeRange()) { 7531 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 7532 } 7533 7534 // First give up the locks, then yield, then re-lock. 7535 // We should probably use a constructor/destructor idiom to 7536 // do this unlock/lock or modify the MutexUnlocker class to 7537 // serve our purpose. XXX 7538 assert_lock_strong(_bitMap->lock()); 7539 assert_lock_strong(_freelistLock); 7540 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 7541 "CMS thread should hold CMS token"); 7542 _bitMap->lock()->unlock(); 7543 _freelistLock->unlock(); 7544 ConcurrentMarkSweepThread::desynchronize(true); 7545 _collector->stopTimer(); 7546 _collector->incrementYields(); 7547 7548 // See the comment in coordinator_yield() 7549 for (unsigned i = 0; i < CMSYieldSleepCount && 7550 ConcurrentMarkSweepThread::should_yield() && 7551 !CMSCollector::foregroundGCIsActive(); ++i) { 7552 os::sleep(Thread::current(), 1, false); 7553 } 7554 7555 ConcurrentMarkSweepThread::synchronize(true); 7556 _freelistLock->lock(); 7557 _bitMap->lock()->lock_without_safepoint_check(); 7558 _collector->startTimer(); 7559 } 7560 7561 #ifndef PRODUCT 7562 // This is actually very useful in a product build if it can 7563 // be called from the debugger. Compile it into the product 7564 // as needed. 7565 bool debug_verify_chunk_in_free_list(FreeChunk* fc) { 7566 return debug_cms_space->verify_chunk_in_free_list(fc); 7567 } 7568 #endif 7569 7570 void SweepClosure::print_free_block_coalesced(FreeChunk* fc) const { 7571 log_develop_trace(gc, sweep)("Sweep:coal_free_blk " PTR_FORMAT " (" SIZE_FORMAT ")", 7572 p2i(fc), fc->size()); 7573 } 7574 7575 // CMSIsAliveClosure 7576 bool CMSIsAliveClosure::do_object_b(oop obj) { 7577 HeapWord* addr = (HeapWord*)obj; 7578 return addr != NULL && 7579 (!_span.contains(addr) || _bit_map->isMarked(addr)); 7580 } 7581 7582 7583 CMSKeepAliveClosure::CMSKeepAliveClosure( CMSCollector* collector, 7584 MemRegion span, 7585 CMSBitMap* bit_map, CMSMarkStack* mark_stack, 7586 bool cpc): 7587 _collector(collector), 7588 _span(span), 7589 _bit_map(bit_map), 7590 _mark_stack(mark_stack), 7591 _concurrent_precleaning(cpc) { 7592 assert(!_span.is_empty(), "Empty span could spell trouble"); 7593 } 7594 7595 7596 // CMSKeepAliveClosure: the serial version 7597 void CMSKeepAliveClosure::do_oop(oop obj) { 7598 HeapWord* addr = (HeapWord*)obj; 7599 if (_span.contains(addr) && 7600 !_bit_map->isMarked(addr)) { 7601 _bit_map->mark(addr); 7602 bool simulate_overflow = false; 7603 NOT_PRODUCT( 7604 if (CMSMarkStackOverflowALot && 7605 _collector->simulate_overflow()) { 7606 // simulate a stack overflow 7607 simulate_overflow = true; 7608 } 7609 ) 7610 if (simulate_overflow || !_mark_stack->push(obj)) { 7611 if (_concurrent_precleaning) { 7612 // We dirty the overflown object and let the remark 7613 // phase deal with it. 7614 assert(_collector->overflow_list_is_empty(), "Error"); 7615 // In the case of object arrays, we need to dirty all of 7616 // the cards that the object spans. No locking or atomics 7617 // are needed since no one else can be mutating the mod union 7618 // table. 7619 if (obj->is_objArray()) { 7620 size_t sz = obj->size(); 7621 HeapWord* end_card_addr = 7622 (HeapWord*)round_to((intptr_t)(addr+sz), CardTable::card_size); 7623 MemRegion redirty_range = MemRegion(addr, end_card_addr); 7624 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 7625 _collector->_modUnionTable.mark_range(redirty_range); 7626 } else { 7627 _collector->_modUnionTable.mark(addr); 7628 } 7629 _collector->_ser_kac_preclean_ovflw++; 7630 } else { 7631 _collector->push_on_overflow_list(obj); 7632 _collector->_ser_kac_ovflw++; 7633 } 7634 } 7635 } 7636 } 7637 7638 void CMSKeepAliveClosure::do_oop(oop* p) { CMSKeepAliveClosure::do_oop_work(p); } 7639 void CMSKeepAliveClosure::do_oop(narrowOop* p) { CMSKeepAliveClosure::do_oop_work(p); } 7640 7641 // CMSParKeepAliveClosure: a parallel version of the above. 7642 // The work queues are private to each closure (thread), 7643 // but (may be) available for stealing by other threads. 7644 void CMSParKeepAliveClosure::do_oop(oop obj) { 7645 HeapWord* addr = (HeapWord*)obj; 7646 if (_span.contains(addr) && 7647 !_bit_map->isMarked(addr)) { 7648 // In general, during recursive tracing, several threads 7649 // may be concurrently getting here; the first one to 7650 // "tag" it, claims it. 7651 if (_bit_map->par_mark(addr)) { 7652 bool res = _work_queue->push(obj); 7653 assert(res, "Low water mark should be much less than capacity"); 7654 // Do a recursive trim in the hope that this will keep 7655 // stack usage lower, but leave some oops for potential stealers 7656 trim_queue(_low_water_mark); 7657 } // Else, another thread got there first 7658 } 7659 } 7660 7661 void CMSParKeepAliveClosure::do_oop(oop* p) { CMSParKeepAliveClosure::do_oop_work(p); } 7662 void CMSParKeepAliveClosure::do_oop(narrowOop* p) { CMSParKeepAliveClosure::do_oop_work(p); } 7663 7664 void CMSParKeepAliveClosure::trim_queue(uint max) { 7665 while (_work_queue->size() > max) { 7666 oop new_oop; 7667 if (_work_queue->pop_local(new_oop)) { 7668 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 7669 assert(_bit_map->isMarked((HeapWord*)new_oop), 7670 "no white objects on this stack!"); 7671 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 7672 // iterate over the oops in this oop, marking and pushing 7673 // the ones in CMS heap (i.e. in _span). 7674 new_oop->oop_iterate(&_mark_and_push); 7675 } 7676 } 7677 } 7678 7679 CMSInnerParMarkAndPushClosure::CMSInnerParMarkAndPushClosure( 7680 CMSCollector* collector, 7681 MemRegion span, CMSBitMap* bit_map, 7682 OopTaskQueue* work_queue): 7683 _collector(collector), 7684 _span(span), 7685 _bit_map(bit_map), 7686 _work_queue(work_queue) { } 7687 7688 void CMSInnerParMarkAndPushClosure::do_oop(oop obj) { 7689 HeapWord* addr = (HeapWord*)obj; 7690 if (_span.contains(addr) && 7691 !_bit_map->isMarked(addr)) { 7692 if (_bit_map->par_mark(addr)) { 7693 bool simulate_overflow = false; 7694 NOT_PRODUCT( 7695 if (CMSMarkStackOverflowALot && 7696 _collector->par_simulate_overflow()) { 7697 // simulate a stack overflow 7698 simulate_overflow = true; 7699 } 7700 ) 7701 if (simulate_overflow || !_work_queue->push(obj)) { 7702 _collector->par_push_on_overflow_list(obj); 7703 _collector->_par_kac_ovflw++; 7704 } 7705 } // Else another thread got there already 7706 } 7707 } 7708 7709 void CMSInnerParMarkAndPushClosure::do_oop(oop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); } 7710 void CMSInnerParMarkAndPushClosure::do_oop(narrowOop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); } 7711 7712 ////////////////////////////////////////////////////////////////// 7713 // CMSExpansionCause ///////////////////////////// 7714 ////////////////////////////////////////////////////////////////// 7715 const char* CMSExpansionCause::to_string(CMSExpansionCause::Cause cause) { 7716 switch (cause) { 7717 case _no_expansion: 7718 return "No expansion"; 7719 case _satisfy_free_ratio: 7720 return "Free ratio"; 7721 case _satisfy_promotion: 7722 return "Satisfy promotion"; 7723 case _satisfy_allocation: 7724 return "allocation"; 7725 case _allocate_par_lab: 7726 return "Par LAB"; 7727 case _allocate_par_spooling_space: 7728 return "Par Spooling Space"; 7729 case _adaptive_size_policy: 7730 return "Ergonomics"; 7731 default: 7732 return "unknown"; 7733 } 7734 } 7735 7736 void CMSDrainMarkingStackClosure::do_void() { 7737 // the max number to take from overflow list at a time 7738 const size_t num = _mark_stack->capacity()/4; 7739 assert(!_concurrent_precleaning || _collector->overflow_list_is_empty(), 7740 "Overflow list should be NULL during concurrent phases"); 7741 while (!_mark_stack->isEmpty() || 7742 // if stack is empty, check the overflow list 7743 _collector->take_from_overflow_list(num, _mark_stack)) { 7744 oop obj = _mark_stack->pop(); 7745 HeapWord* addr = (HeapWord*)obj; 7746 assert(_span.contains(addr), "Should be within span"); 7747 assert(_bit_map->isMarked(addr), "Should be marked"); 7748 assert(obj->is_oop(), "Should be an oop"); 7749 obj->oop_iterate(_keep_alive); 7750 } 7751 } 7752 7753 void CMSParDrainMarkingStackClosure::do_void() { 7754 // drain queue 7755 trim_queue(0); 7756 } 7757 7758 // Trim our work_queue so its length is below max at return 7759 void CMSParDrainMarkingStackClosure::trim_queue(uint max) { 7760 while (_work_queue->size() > max) { 7761 oop new_oop; 7762 if (_work_queue->pop_local(new_oop)) { 7763 assert(new_oop->is_oop(), "Expected an oop"); 7764 assert(_bit_map->isMarked((HeapWord*)new_oop), 7765 "no white objects on this stack!"); 7766 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 7767 // iterate over the oops in this oop, marking and pushing 7768 // the ones in CMS heap (i.e. in _span). 7769 new_oop->oop_iterate(&_mark_and_push); 7770 } 7771 } 7772 } 7773 7774 //////////////////////////////////////////////////////////////////// 7775 // Support for Marking Stack Overflow list handling and related code 7776 //////////////////////////////////////////////////////////////////// 7777 // Much of the following code is similar in shape and spirit to the 7778 // code used in ParNewGC. We should try and share that code 7779 // as much as possible in the future. 7780 7781 #ifndef PRODUCT 7782 // Debugging support for CMSStackOverflowALot 7783 7784 // It's OK to call this multi-threaded; the worst thing 7785 // that can happen is that we'll get a bunch of closely 7786 // spaced simulated overflows, but that's OK, in fact 7787 // probably good as it would exercise the overflow code 7788 // under contention. 7789 bool CMSCollector::simulate_overflow() { 7790 if (_overflow_counter-- <= 0) { // just being defensive 7791 _overflow_counter = CMSMarkStackOverflowInterval; 7792 return true; 7793 } else { 7794 return false; 7795 } 7796 } 7797 7798 bool CMSCollector::par_simulate_overflow() { 7799 return simulate_overflow(); 7800 } 7801 #endif 7802 7803 // Single-threaded 7804 bool CMSCollector::take_from_overflow_list(size_t num, CMSMarkStack* stack) { 7805 assert(stack->isEmpty(), "Expected precondition"); 7806 assert(stack->capacity() > num, "Shouldn't bite more than can chew"); 7807 size_t i = num; 7808 oop cur = _overflow_list; 7809 const markOop proto = markOopDesc::prototype(); 7810 NOT_PRODUCT(ssize_t n = 0;) 7811 for (oop next; i > 0 && cur != NULL; cur = next, i--) { 7812 next = oop(cur->mark()); 7813 cur->set_mark(proto); // until proven otherwise 7814 assert(cur->is_oop(), "Should be an oop"); 7815 bool res = stack->push(cur); 7816 assert(res, "Bit off more than can chew?"); 7817 NOT_PRODUCT(n++;) 7818 } 7819 _overflow_list = cur; 7820 #ifndef PRODUCT 7821 assert(_num_par_pushes >= n, "Too many pops?"); 7822 _num_par_pushes -=n; 7823 #endif 7824 return !stack->isEmpty(); 7825 } 7826 7827 #define BUSY (cast_to_oop<intptr_t>(0x1aff1aff)) 7828 // (MT-safe) Get a prefix of at most "num" from the list. 7829 // The overflow list is chained through the mark word of 7830 // each object in the list. We fetch the entire list, 7831 // break off a prefix of the right size and return the 7832 // remainder. If other threads try to take objects from 7833 // the overflow list at that time, they will wait for 7834 // some time to see if data becomes available. If (and 7835 // only if) another thread places one or more object(s) 7836 // on the global list before we have returned the suffix 7837 // to the global list, we will walk down our local list 7838 // to find its end and append the global list to 7839 // our suffix before returning it. This suffix walk can 7840 // prove to be expensive (quadratic in the amount of traffic) 7841 // when there are many objects in the overflow list and 7842 // there is much producer-consumer contention on the list. 7843 // *NOTE*: The overflow list manipulation code here and 7844 // in ParNewGeneration:: are very similar in shape, 7845 // except that in the ParNew case we use the old (from/eden) 7846 // copy of the object to thread the list via its klass word. 7847 // Because of the common code, if you make any changes in 7848 // the code below, please check the ParNew version to see if 7849 // similar changes might be needed. 7850 // CR 6797058 has been filed to consolidate the common code. 7851 bool CMSCollector::par_take_from_overflow_list(size_t num, 7852 OopTaskQueue* work_q, 7853 int no_of_gc_threads) { 7854 assert(work_q->size() == 0, "First empty local work queue"); 7855 assert(num < work_q->max_elems(), "Can't bite more than we can chew"); 7856 if (_overflow_list == NULL) { 7857 return false; 7858 } 7859 // Grab the entire list; we'll put back a suffix 7860 oop prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list)); 7861 Thread* tid = Thread::current(); 7862 // Before "no_of_gc_threads" was introduced CMSOverflowSpinCount was 7863 // set to ParallelGCThreads. 7864 size_t CMSOverflowSpinCount = (size_t) no_of_gc_threads; // was ParallelGCThreads; 7865 size_t sleep_time_millis = MAX2((size_t)1, num/100); 7866 // If the list is busy, we spin for a short while, 7867 // sleeping between attempts to get the list. 7868 for (size_t spin = 0; prefix == BUSY && spin < CMSOverflowSpinCount; spin++) { 7869 os::sleep(tid, sleep_time_millis, false); 7870 if (_overflow_list == NULL) { 7871 // Nothing left to take 7872 return false; 7873 } else if (_overflow_list != BUSY) { 7874 // Try and grab the prefix 7875 prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list)); 7876 } 7877 } 7878 // If the list was found to be empty, or we spun long 7879 // enough, we give up and return empty-handed. If we leave 7880 // the list in the BUSY state below, it must be the case that 7881 // some other thread holds the overflow list and will set it 7882 // to a non-BUSY state in the future. 7883 if (prefix == NULL || prefix == BUSY) { 7884 // Nothing to take or waited long enough 7885 if (prefix == NULL) { 7886 // Write back the NULL in case we overwrote it with BUSY above 7887 // and it is still the same value. 7888 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY); 7889 } 7890 return false; 7891 } 7892 assert(prefix != NULL && prefix != BUSY, "Error"); 7893 size_t i = num; 7894 oop cur = prefix; 7895 // Walk down the first "num" objects, unless we reach the end. 7896 for (; i > 1 && cur->mark() != NULL; cur = oop(cur->mark()), i--); 7897 if (cur->mark() == NULL) { 7898 // We have "num" or fewer elements in the list, so there 7899 // is nothing to return to the global list. 7900 // Write back the NULL in lieu of the BUSY we wrote 7901 // above, if it is still the same value. 7902 if (_overflow_list == BUSY) { 7903 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY); 7904 } 7905 } else { 7906 // Chop off the suffix and return it to the global list. 7907 assert(cur->mark() != BUSY, "Error"); 7908 oop suffix_head = cur->mark(); // suffix will be put back on global list 7909 cur->set_mark(NULL); // break off suffix 7910 // It's possible that the list is still in the empty(busy) state 7911 // we left it in a short while ago; in that case we may be 7912 // able to place back the suffix without incurring the cost 7913 // of a walk down the list. 7914 oop observed_overflow_list = _overflow_list; 7915 oop cur_overflow_list = observed_overflow_list; 7916 bool attached = false; 7917 while (observed_overflow_list == BUSY || observed_overflow_list == NULL) { 7918 observed_overflow_list = 7919 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list); 7920 if (cur_overflow_list == observed_overflow_list) { 7921 attached = true; 7922 break; 7923 } else cur_overflow_list = observed_overflow_list; 7924 } 7925 if (!attached) { 7926 // Too bad, someone else sneaked in (at least) an element; we'll need 7927 // to do a splice. Find tail of suffix so we can prepend suffix to global 7928 // list. 7929 for (cur = suffix_head; cur->mark() != NULL; cur = (oop)(cur->mark())); 7930 oop suffix_tail = cur; 7931 assert(suffix_tail != NULL && suffix_tail->mark() == NULL, 7932 "Tautology"); 7933 observed_overflow_list = _overflow_list; 7934 do { 7935 cur_overflow_list = observed_overflow_list; 7936 if (cur_overflow_list != BUSY) { 7937 // Do the splice ... 7938 suffix_tail->set_mark(markOop(cur_overflow_list)); 7939 } else { // cur_overflow_list == BUSY 7940 suffix_tail->set_mark(NULL); 7941 } 7942 // ... and try to place spliced list back on overflow_list ... 7943 observed_overflow_list = 7944 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list); 7945 } while (cur_overflow_list != observed_overflow_list); 7946 // ... until we have succeeded in doing so. 7947 } 7948 } 7949 7950 // Push the prefix elements on work_q 7951 assert(prefix != NULL, "control point invariant"); 7952 const markOop proto = markOopDesc::prototype(); 7953 oop next; 7954 NOT_PRODUCT(ssize_t n = 0;) 7955 for (cur = prefix; cur != NULL; cur = next) { 7956 next = oop(cur->mark()); 7957 cur->set_mark(proto); // until proven otherwise 7958 assert(cur->is_oop(), "Should be an oop"); 7959 bool res = work_q->push(cur); 7960 assert(res, "Bit off more than we can chew?"); 7961 NOT_PRODUCT(n++;) 7962 } 7963 #ifndef PRODUCT 7964 assert(_num_par_pushes >= n, "Too many pops?"); 7965 Atomic::add_ptr(-(intptr_t)n, &_num_par_pushes); 7966 #endif 7967 return true; 7968 } 7969 7970 // Single-threaded 7971 void CMSCollector::push_on_overflow_list(oop p) { 7972 NOT_PRODUCT(_num_par_pushes++;) 7973 assert(p->is_oop(), "Not an oop"); 7974 preserve_mark_if_necessary(p); 7975 p->set_mark((markOop)_overflow_list); 7976 _overflow_list = p; 7977 } 7978 7979 // Multi-threaded; use CAS to prepend to overflow list 7980 void CMSCollector::par_push_on_overflow_list(oop p) { 7981 NOT_PRODUCT(Atomic::inc_ptr(&_num_par_pushes);) 7982 assert(p->is_oop(), "Not an oop"); 7983 par_preserve_mark_if_necessary(p); 7984 oop observed_overflow_list = _overflow_list; 7985 oop cur_overflow_list; 7986 do { 7987 cur_overflow_list = observed_overflow_list; 7988 if (cur_overflow_list != BUSY) { 7989 p->set_mark(markOop(cur_overflow_list)); 7990 } else { 7991 p->set_mark(NULL); 7992 } 7993 observed_overflow_list = 7994 (oop) Atomic::cmpxchg_ptr(p, &_overflow_list, cur_overflow_list); 7995 } while (cur_overflow_list != observed_overflow_list); 7996 } 7997 #undef BUSY 7998 7999 // Single threaded 8000 // General Note on GrowableArray: pushes may silently fail 8001 // because we are (temporarily) out of C-heap for expanding 8002 // the stack. The problem is quite ubiquitous and affects 8003 // a lot of code in the JVM. The prudent thing for GrowableArray 8004 // to do (for now) is to exit with an error. However, that may 8005 // be too draconian in some cases because the caller may be 8006 // able to recover without much harm. For such cases, we 8007 // should probably introduce a "soft_push" method which returns 8008 // an indication of success or failure with the assumption that 8009 // the caller may be able to recover from a failure; code in 8010 // the VM can then be changed, incrementally, to deal with such 8011 // failures where possible, thus, incrementally hardening the VM 8012 // in such low resource situations. 8013 void CMSCollector::preserve_mark_work(oop p, markOop m) { 8014 _preserved_oop_stack.push(p); 8015 _preserved_mark_stack.push(m); 8016 assert(m == p->mark(), "Mark word changed"); 8017 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), 8018 "bijection"); 8019 } 8020 8021 // Single threaded 8022 void CMSCollector::preserve_mark_if_necessary(oop p) { 8023 markOop m = p->mark(); 8024 if (m->must_be_preserved(p)) { 8025 preserve_mark_work(p, m); 8026 } 8027 } 8028 8029 void CMSCollector::par_preserve_mark_if_necessary(oop p) { 8030 markOop m = p->mark(); 8031 if (m->must_be_preserved(p)) { 8032 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag); 8033 // Even though we read the mark word without holding 8034 // the lock, we are assured that it will not change 8035 // because we "own" this oop, so no other thread can 8036 // be trying to push it on the overflow list; see 8037 // the assertion in preserve_mark_work() that checks 8038 // that m == p->mark(). 8039 preserve_mark_work(p, m); 8040 } 8041 } 8042 8043 // We should be able to do this multi-threaded, 8044 // a chunk of stack being a task (this is 8045 // correct because each oop only ever appears 8046 // once in the overflow list. However, it's 8047 // not very easy to completely overlap this with 8048 // other operations, so will generally not be done 8049 // until all work's been completed. Because we 8050 // expect the preserved oop stack (set) to be small, 8051 // it's probably fine to do this single-threaded. 8052 // We can explore cleverer concurrent/overlapped/parallel 8053 // processing of preserved marks if we feel the 8054 // need for this in the future. Stack overflow should 8055 // be so rare in practice and, when it happens, its 8056 // effect on performance so great that this will 8057 // likely just be in the noise anyway. 8058 void CMSCollector::restore_preserved_marks_if_any() { 8059 assert(SafepointSynchronize::is_at_safepoint(), 8060 "world should be stopped"); 8061 assert(Thread::current()->is_ConcurrentGC_thread() || 8062 Thread::current()->is_VM_thread(), 8063 "should be single-threaded"); 8064 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), 8065 "bijection"); 8066 8067 while (!_preserved_oop_stack.is_empty()) { 8068 oop p = _preserved_oop_stack.pop(); 8069 assert(p->is_oop(), "Should be an oop"); 8070 assert(_span.contains(p), "oop should be in _span"); 8071 assert(p->mark() == markOopDesc::prototype(), 8072 "Set when taken from overflow list"); 8073 markOop m = _preserved_mark_stack.pop(); 8074 p->set_mark(m); 8075 } 8076 assert(_preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(), 8077 "stacks were cleared above"); 8078 } 8079 8080 #ifndef PRODUCT 8081 bool CMSCollector::no_preserved_marks() const { 8082 return _preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(); 8083 } 8084 #endif 8085 8086 // Transfer some number of overflown objects to usual marking 8087 // stack. Return true if some objects were transferred. 8088 bool MarkRefsIntoAndScanClosure::take_from_overflow_list() { 8089 size_t num = MIN2((size_t)(_mark_stack->capacity() - _mark_stack->length())/4, 8090 (size_t)ParGCDesiredObjsFromOverflowList); 8091 8092 bool res = _collector->take_from_overflow_list(num, _mark_stack); 8093 assert(_collector->overflow_list_is_empty() || res, 8094 "If list is not empty, we should have taken something"); 8095 assert(!res || !_mark_stack->isEmpty(), 8096 "If we took something, it should now be on our stack"); 8097 return res; 8098 } 8099 8100 size_t MarkDeadObjectsClosure::do_blk(HeapWord* addr) { 8101 size_t res = _sp->block_size_no_stall(addr, _collector); 8102 if (_sp->block_is_obj(addr)) { 8103 if (_live_bit_map->isMarked(addr)) { 8104 // It can't have been dead in a previous cycle 8105 guarantee(!_dead_bit_map->isMarked(addr), "No resurrection!"); 8106 } else { 8107 _dead_bit_map->mark(addr); // mark the dead object 8108 } 8109 } 8110 // Could be 0, if the block size could not be computed without stalling. 8111 return res; 8112 } 8113 8114 TraceCMSMemoryManagerStats::TraceCMSMemoryManagerStats(CMSCollector::CollectorState phase, GCCause::Cause cause): TraceMemoryManagerStats() { 8115 8116 switch (phase) { 8117 case CMSCollector::InitialMarking: 8118 initialize(true /* fullGC */ , 8119 cause /* cause of the GC */, 8120 true /* recordGCBeginTime */, 8121 true /* recordPreGCUsage */, 8122 false /* recordPeakUsage */, 8123 false /* recordPostGCusage */, 8124 true /* recordAccumulatedGCTime */, 8125 false /* recordGCEndTime */, 8126 false /* countCollection */ ); 8127 break; 8128 8129 case CMSCollector::FinalMarking: 8130 initialize(true /* fullGC */ , 8131 cause /* cause of the GC */, 8132 false /* recordGCBeginTime */, 8133 false /* recordPreGCUsage */, 8134 false /* recordPeakUsage */, 8135 false /* recordPostGCusage */, 8136 true /* recordAccumulatedGCTime */, 8137 false /* recordGCEndTime */, 8138 false /* countCollection */ ); 8139 break; 8140 8141 case CMSCollector::Sweeping: 8142 initialize(true /* fullGC */ , 8143 cause /* cause of the GC */, 8144 false /* recordGCBeginTime */, 8145 false /* recordPreGCUsage */, 8146 true /* recordPeakUsage */, 8147 true /* recordPostGCusage */, 8148 false /* recordAccumulatedGCTime */, 8149 true /* recordGCEndTime */, 8150 true /* countCollection */ ); 8151 break; 8152 8153 default: 8154 ShouldNotReachHere(); 8155 } 8156 }