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