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