1 /* 2 * Copyright (c) 2001, 2017, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "classfile/classLoaderData.hpp" 27 #include "classfile/stringTable.hpp" 28 #include "classfile/symbolTable.hpp" 29 #include "classfile/systemDictionary.hpp" 30 #include "code/codeCache.hpp" 31 #include "gc/cms/cmsCollectorPolicy.hpp" 32 #include "gc/cms/cmsOopClosures.inline.hpp" 33 #include "gc/cms/compactibleFreeListSpace.hpp" 34 #include "gc/cms/concurrentMarkSweepGeneration.inline.hpp" 35 #include "gc/cms/concurrentMarkSweepThread.hpp" 36 #include "gc/cms/parNewGeneration.hpp" 37 #include "gc/cms/vmCMSOperations.hpp" 38 #include "gc/serial/genMarkSweep.hpp" 39 #include "gc/serial/tenuredGeneration.hpp" 40 #include "gc/shared/adaptiveSizePolicy.hpp" 41 #include "gc/shared/cardGeneration.inline.hpp" 42 #include "gc/shared/cardTableRS.hpp" 43 #include "gc/shared/collectedHeap.inline.hpp" 44 #include "gc/shared/collectorCounters.hpp" 45 #include "gc/shared/collectorPolicy.hpp" 46 #include "gc/shared/gcLocker.inline.hpp" 47 #include "gc/shared/gcPolicyCounters.hpp" 48 #include "gc/shared/gcTimer.hpp" 49 #include "gc/shared/gcTrace.hpp" 50 #include "gc/shared/gcTraceTime.inline.hpp" 51 #include "gc/shared/genCollectedHeap.hpp" 52 #include "gc/shared/genOopClosures.inline.hpp" 53 #include "gc/shared/isGCActiveMark.hpp" 54 #include "gc/shared/referencePolicy.hpp" 55 #include "gc/shared/strongRootsScope.hpp" 56 #include "gc/shared/taskqueue.inline.hpp" 57 #include "logging/log.hpp" 58 #include "logging/logStream.hpp" 59 #include "memory/allocation.hpp" 60 #include "memory/iterator.inline.hpp" 61 #include "memory/padded.hpp" 62 #include "memory/resourceArea.hpp" 63 #include "oops/oop.inline.hpp" 64 #include "prims/jvmtiExport.hpp" 65 #include "runtime/atomic.hpp" 66 #include "runtime/globals_extension.hpp" 67 #include "runtime/handles.inline.hpp" 68 #include "runtime/java.hpp" 69 #include "runtime/orderAccess.inline.hpp" 70 #include "runtime/timer.hpp" 71 #include "runtime/vmThread.hpp" 72 #include "services/memoryService.hpp" 73 #include "services/runtimeService.hpp" 74 #include "utilities/align.hpp" 75 #include "utilities/stack.inline.hpp" 76 77 // statics 78 CMSCollector* ConcurrentMarkSweepGeneration::_collector = NULL; 79 bool CMSCollector::_full_gc_requested = false; 80 GCCause::Cause CMSCollector::_full_gc_cause = GCCause::_no_gc; 81 82 ////////////////////////////////////////////////////////////////// 83 // In support of CMS/VM thread synchronization 84 ////////////////////////////////////////////////////////////////// 85 // We split use of the CGC_lock into 2 "levels". 86 // The low-level locking is of the usual CGC_lock monitor. We introduce 87 // a higher level "token" (hereafter "CMS token") built on top of the 88 // low level monitor (hereafter "CGC lock"). 89 // The token-passing protocol gives priority to the VM thread. The 90 // CMS-lock doesn't provide any fairness guarantees, but clients 91 // should ensure that it is only held for very short, bounded 92 // durations. 93 // 94 // When either of the CMS thread or the VM thread is involved in 95 // collection operations during which it does not want the other 96 // thread to interfere, it obtains the CMS token. 97 // 98 // If either thread tries to get the token while the other has 99 // it, that thread waits. However, if the VM thread and CMS thread 100 // both want the token, then the VM thread gets priority while the 101 // CMS thread waits. This ensures, for instance, that the "concurrent" 102 // phases of the CMS thread's work do not block out the VM thread 103 // for long periods of time as the CMS thread continues to hog 104 // the token. (See bug 4616232). 105 // 106 // The baton-passing functions are, however, controlled by the 107 // flags _foregroundGCShouldWait and _foregroundGCIsActive, 108 // and here the low-level CMS lock, not the high level token, 109 // ensures mutual exclusion. 110 // 111 // Two important conditions that we have to satisfy: 112 // 1. if a thread does a low-level wait on the CMS lock, then it 113 // relinquishes the CMS token if it were holding that token 114 // when it acquired the low-level CMS lock. 115 // 2. any low-level notifications on the low-level lock 116 // should only be sent when a thread has relinquished the token. 117 // 118 // In the absence of either property, we'd have potential deadlock. 119 // 120 // We protect each of the CMS (concurrent and sequential) phases 121 // with the CMS _token_, not the CMS _lock_. 122 // 123 // The only code protected by CMS lock is the token acquisition code 124 // itself, see ConcurrentMarkSweepThread::[de]synchronize(), and the 125 // baton-passing code. 126 // 127 // Unfortunately, i couldn't come up with a good abstraction to factor and 128 // hide the naked CGC_lock manipulation in the baton-passing code 129 // further below. That's something we should try to do. Also, the proof 130 // of correctness of this 2-level locking scheme is far from obvious, 131 // and potentially quite slippery. We have an uneasy suspicion, for instance, 132 // that there may be a theoretical possibility of delay/starvation in the 133 // low-level lock/wait/notify scheme used for the baton-passing because of 134 // potential interference with the priority scheme embodied in the 135 // CMS-token-passing protocol. See related comments at a CGC_lock->wait() 136 // invocation further below and marked with "XXX 20011219YSR". 137 // Indeed, as we note elsewhere, this may become yet more slippery 138 // in the presence of multiple CMS and/or multiple VM threads. XXX 139 140 class CMSTokenSync: public StackObj { 141 private: 142 bool _is_cms_thread; 143 public: 144 CMSTokenSync(bool is_cms_thread): 145 _is_cms_thread(is_cms_thread) { 146 assert(is_cms_thread == Thread::current()->is_ConcurrentGC_thread(), 147 "Incorrect argument to constructor"); 148 ConcurrentMarkSweepThread::synchronize(_is_cms_thread); 149 } 150 151 ~CMSTokenSync() { 152 assert(_is_cms_thread ? 153 ConcurrentMarkSweepThread::cms_thread_has_cms_token() : 154 ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 155 "Incorrect state"); 156 ConcurrentMarkSweepThread::desynchronize(_is_cms_thread); 157 } 158 }; 159 160 // Convenience class that does a CMSTokenSync, and then acquires 161 // upto three locks. 162 class CMSTokenSyncWithLocks: public CMSTokenSync { 163 private: 164 // Note: locks are acquired in textual declaration order 165 // and released in the opposite order 166 MutexLockerEx _locker1, _locker2, _locker3; 167 public: 168 CMSTokenSyncWithLocks(bool is_cms_thread, Mutex* mutex1, 169 Mutex* mutex2 = NULL, Mutex* mutex3 = NULL): 170 CMSTokenSync(is_cms_thread), 171 _locker1(mutex1, Mutex::_no_safepoint_check_flag), 172 _locker2(mutex2, Mutex::_no_safepoint_check_flag), 173 _locker3(mutex3, Mutex::_no_safepoint_check_flag) 174 { } 175 }; 176 177 178 ////////////////////////////////////////////////////////////////// 179 // Concurrent Mark-Sweep Generation ///////////////////////////// 180 ////////////////////////////////////////////////////////////////// 181 182 NOT_PRODUCT(CompactibleFreeListSpace* debug_cms_space;) 183 184 // This struct contains per-thread things necessary to support parallel 185 // young-gen collection. 186 class CMSParGCThreadState: public CHeapObj<mtGC> { 187 public: 188 CompactibleFreeListSpaceLAB lab; 189 PromotionInfo promo; 190 191 // Constructor. 192 CMSParGCThreadState(CompactibleFreeListSpace* cfls) : lab(cfls) { 193 promo.setSpace(cfls); 194 } 195 }; 196 197 ConcurrentMarkSweepGeneration::ConcurrentMarkSweepGeneration( 198 ReservedSpace rs, size_t initial_byte_size, CardTableRS* ct) : 199 CardGeneration(rs, initial_byte_size, ct), 200 _dilatation_factor(((double)MinChunkSize)/((double)(CollectedHeap::min_fill_size()))), 201 _did_compact(false) 202 { 203 HeapWord* bottom = (HeapWord*) _virtual_space.low(); 204 HeapWord* end = (HeapWord*) _virtual_space.high(); 205 206 _direct_allocated_words = 0; 207 NOT_PRODUCT( 208 _numObjectsPromoted = 0; 209 _numWordsPromoted = 0; 210 _numObjectsAllocated = 0; 211 _numWordsAllocated = 0; 212 ) 213 214 _cmsSpace = new CompactibleFreeListSpace(_bts, MemRegion(bottom, end)); 215 NOT_PRODUCT(debug_cms_space = _cmsSpace;) 216 _cmsSpace->_old_gen = this; 217 218 _gc_stats = new CMSGCStats(); 219 220 // Verify the assumption that FreeChunk::_prev and OopDesc::_klass 221 // offsets match. The ability to tell free chunks from objects 222 // depends on this property. 223 debug_only( 224 FreeChunk* junk = NULL; 225 assert(UseCompressedClassPointers || 226 junk->prev_addr() == (void*)(oop(junk)->klass_addr()), 227 "Offset of FreeChunk::_prev within FreeChunk must match" 228 " that of OopDesc::_klass within OopDesc"); 229 ) 230 231 _par_gc_thread_states = NEW_C_HEAP_ARRAY(CMSParGCThreadState*, ParallelGCThreads, mtGC); 232 for (uint i = 0; i < ParallelGCThreads; i++) { 233 _par_gc_thread_states[i] = new CMSParGCThreadState(cmsSpace()); 234 } 235 236 _incremental_collection_failed = false; 237 // The "dilatation_factor" is the expansion that can occur on 238 // account of the fact that the minimum object size in the CMS 239 // generation may be larger than that in, say, a contiguous young 240 // generation. 241 // Ideally, in the calculation below, we'd compute the dilatation 242 // factor as: MinChunkSize/(promoting_gen's min object size) 243 // Since we do not have such a general query interface for the 244 // promoting generation, we'll instead just use the minimum 245 // object size (which today is a header's worth of space); 246 // note that all arithmetic is in units of HeapWords. 247 assert(MinChunkSize >= CollectedHeap::min_fill_size(), "just checking"); 248 assert(_dilatation_factor >= 1.0, "from previous assert"); 249 } 250 251 252 // The field "_initiating_occupancy" represents the occupancy percentage 253 // at which we trigger a new collection cycle. Unless explicitly specified 254 // via CMSInitiatingOccupancyFraction (argument "io" below), it 255 // is calculated by: 256 // 257 // Let "f" be MinHeapFreeRatio in 258 // 259 // _initiating_occupancy = 100-f + 260 // f * (CMSTriggerRatio/100) 261 // where CMSTriggerRatio is the argument "tr" below. 262 // 263 // That is, if we assume the heap is at its desired maximum occupancy at the 264 // end of a collection, we let CMSTriggerRatio of the (purported) free 265 // space be allocated before initiating a new collection cycle. 266 // 267 void ConcurrentMarkSweepGeneration::init_initiating_occupancy(intx io, uintx tr) { 268 assert(io <= 100 && tr <= 100, "Check the arguments"); 269 if (io >= 0) { 270 _initiating_occupancy = (double)io / 100.0; 271 } else { 272 _initiating_occupancy = ((100 - MinHeapFreeRatio) + 273 (double)(tr * MinHeapFreeRatio) / 100.0) 274 / 100.0; 275 } 276 } 277 278 void ConcurrentMarkSweepGeneration::ref_processor_init() { 279 assert(collector() != NULL, "no collector"); 280 collector()->ref_processor_init(); 281 } 282 283 void CMSCollector::ref_processor_init() { 284 if (_ref_processor == NULL) { 285 // Allocate and initialize a reference processor 286 _ref_processor = 287 new ReferenceProcessor(_span, // span 288 (ParallelGCThreads > 1) && ParallelRefProcEnabled, // mt processing 289 ParallelGCThreads, // mt processing degree 290 _cmsGen->refs_discovery_is_mt(), // mt discovery 291 MAX2(ConcGCThreads, ParallelGCThreads), // mt discovery degree 292 _cmsGen->refs_discovery_is_atomic(), // discovery is not atomic 293 &_is_alive_closure); // closure for liveness info 294 // Initialize the _ref_processor field of CMSGen 295 _cmsGen->set_ref_processor(_ref_processor); 296 297 } 298 } 299 300 AdaptiveSizePolicy* CMSCollector::size_policy() { 301 GenCollectedHeap* gch = GenCollectedHeap::heap(); 302 return gch->gen_policy()->size_policy(); 303 } 304 305 void ConcurrentMarkSweepGeneration::initialize_performance_counters() { 306 307 const char* gen_name = "old"; 308 GenCollectorPolicy* gcp = GenCollectedHeap::heap()->gen_policy(); 309 // Generation Counters - generation 1, 1 subspace 310 _gen_counters = new GenerationCounters(gen_name, 1, 1, 311 gcp->min_old_size(), gcp->max_old_size(), &_virtual_space); 312 313 _space_counters = new GSpaceCounters(gen_name, 0, 314 _virtual_space.reserved_size(), 315 this, _gen_counters); 316 } 317 318 CMSStats::CMSStats(ConcurrentMarkSweepGeneration* cms_gen, unsigned int alpha): 319 _cms_gen(cms_gen) 320 { 321 assert(alpha <= 100, "bad value"); 322 _saved_alpha = alpha; 323 324 // Initialize the alphas to the bootstrap value of 100. 325 _gc0_alpha = _cms_alpha = 100; 326 327 _cms_begin_time.update(); 328 _cms_end_time.update(); 329 330 _gc0_duration = 0.0; 331 _gc0_period = 0.0; 332 _gc0_promoted = 0; 333 334 _cms_duration = 0.0; 335 _cms_period = 0.0; 336 _cms_allocated = 0; 337 338 _cms_used_at_gc0_begin = 0; 339 _cms_used_at_gc0_end = 0; 340 _allow_duty_cycle_reduction = false; 341 _valid_bits = 0; 342 } 343 344 double CMSStats::cms_free_adjustment_factor(size_t free) const { 345 // TBD: CR 6909490 346 return 1.0; 347 } 348 349 void CMSStats::adjust_cms_free_adjustment_factor(bool fail, size_t free) { 350 } 351 352 // If promotion failure handling is on use 353 // the padded average size of the promotion for each 354 // young generation collection. 355 double CMSStats::time_until_cms_gen_full() const { 356 size_t cms_free = _cms_gen->cmsSpace()->free(); 357 GenCollectedHeap* gch = GenCollectedHeap::heap(); 358 size_t expected_promotion = MIN2(gch->young_gen()->capacity(), 359 (size_t) _cms_gen->gc_stats()->avg_promoted()->padded_average()); 360 if (cms_free > expected_promotion) { 361 // Start a cms collection if there isn't enough space to promote 362 // for the next young collection. Use the padded average as 363 // a safety factor. 364 cms_free -= expected_promotion; 365 366 // Adjust by the safety factor. 367 double cms_free_dbl = (double)cms_free; 368 double cms_adjustment = (100.0 - CMSIncrementalSafetyFactor) / 100.0; 369 // Apply a further correction factor which tries to adjust 370 // for recent occurance of concurrent mode failures. 371 cms_adjustment = cms_adjustment * cms_free_adjustment_factor(cms_free); 372 cms_free_dbl = cms_free_dbl * cms_adjustment; 373 374 log_trace(gc)("CMSStats::time_until_cms_gen_full: cms_free " SIZE_FORMAT " expected_promotion " SIZE_FORMAT, 375 cms_free, expected_promotion); 376 log_trace(gc)(" cms_free_dbl %f cms_consumption_rate %f", cms_free_dbl, cms_consumption_rate() + 1.0); 377 // Add 1 in case the consumption rate goes to zero. 378 return cms_free_dbl / (cms_consumption_rate() + 1.0); 379 } 380 return 0.0; 381 } 382 383 // Compare the duration of the cms collection to the 384 // time remaining before the cms generation is empty. 385 // Note that the time from the start of the cms collection 386 // to the start of the cms sweep (less than the total 387 // duration of the cms collection) can be used. This 388 // has been tried and some applications experienced 389 // promotion failures early in execution. This was 390 // possibly because the averages were not accurate 391 // enough at the beginning. 392 double CMSStats::time_until_cms_start() const { 393 // We add "gc0_period" to the "work" calculation 394 // below because this query is done (mostly) at the 395 // end of a scavenge, so we need to conservatively 396 // account for that much possible delay 397 // in the query so as to avoid concurrent mode failures 398 // due to starting the collection just a wee bit too 399 // late. 400 double work = cms_duration() + gc0_period(); 401 double deadline = time_until_cms_gen_full(); 402 // If a concurrent mode failure occurred recently, we want to be 403 // more conservative and halve our expected time_until_cms_gen_full() 404 if (work > deadline) { 405 log_develop_trace(gc)("CMSCollector: collect because of anticipated promotion before full %3.7f + %3.7f > %3.7f ", 406 cms_duration(), gc0_period(), time_until_cms_gen_full()); 407 return 0.0; 408 } 409 return work - deadline; 410 } 411 412 #ifndef PRODUCT 413 void CMSStats::print_on(outputStream *st) const { 414 st->print(" gc0_alpha=%d,cms_alpha=%d", _gc0_alpha, _cms_alpha); 415 st->print(",gc0_dur=%g,gc0_per=%g,gc0_promo=" SIZE_FORMAT, 416 gc0_duration(), gc0_period(), gc0_promoted()); 417 st->print(",cms_dur=%g,cms_per=%g,cms_alloc=" SIZE_FORMAT, 418 cms_duration(), cms_period(), cms_allocated()); 419 st->print(",cms_since_beg=%g,cms_since_end=%g", 420 cms_time_since_begin(), cms_time_since_end()); 421 st->print(",cms_used_beg=" SIZE_FORMAT ",cms_used_end=" SIZE_FORMAT, 422 _cms_used_at_gc0_begin, _cms_used_at_gc0_end); 423 424 if (valid()) { 425 st->print(",promo_rate=%g,cms_alloc_rate=%g", 426 promotion_rate(), cms_allocation_rate()); 427 st->print(",cms_consumption_rate=%g,time_until_full=%g", 428 cms_consumption_rate(), time_until_cms_gen_full()); 429 } 430 st->cr(); 431 } 432 #endif // #ifndef PRODUCT 433 434 CMSCollector::CollectorState CMSCollector::_collectorState = 435 CMSCollector::Idling; 436 bool CMSCollector::_foregroundGCIsActive = false; 437 bool CMSCollector::_foregroundGCShouldWait = false; 438 439 CMSCollector::CMSCollector(ConcurrentMarkSweepGeneration* cmsGen, 440 CardTableRS* ct, 441 ConcurrentMarkSweepPolicy* cp): 442 _cmsGen(cmsGen), 443 _ct(ct), 444 _ref_processor(NULL), // will be set later 445 _conc_workers(NULL), // may be set later 446 _abort_preclean(false), 447 _start_sampling(false), 448 _between_prologue_and_epilogue(false), 449 _markBitMap(0, Mutex::leaf + 1, "CMS_markBitMap_lock"), 450 _modUnionTable((CardTableModRefBS::card_shift - LogHeapWordSize), 451 -1 /* lock-free */, "No_lock" /* dummy */), 452 _modUnionClosurePar(&_modUnionTable), 453 // Adjust my span to cover old (cms) gen 454 _span(cmsGen->reserved()), 455 // Construct the is_alive_closure with _span & markBitMap 456 _is_alive_closure(_span, &_markBitMap), 457 _restart_addr(NULL), 458 _overflow_list(NULL), 459 _stats(cmsGen), 460 _eden_chunk_lock(new Mutex(Mutex::leaf + 1, "CMS_eden_chunk_lock", true, 461 //verify that this lock should be acquired with safepoint check. 462 Monitor::_safepoint_check_sometimes)), 463 _eden_chunk_array(NULL), // may be set in ctor body 464 _eden_chunk_capacity(0), // -- ditto -- 465 _eden_chunk_index(0), // -- ditto -- 466 _survivor_plab_array(NULL), // -- ditto -- 467 _survivor_chunk_array(NULL), // -- ditto -- 468 _survivor_chunk_capacity(0), // -- ditto -- 469 _survivor_chunk_index(0), // -- ditto -- 470 _ser_pmc_preclean_ovflw(0), 471 _ser_kac_preclean_ovflw(0), 472 _ser_pmc_remark_ovflw(0), 473 _par_pmc_remark_ovflw(0), 474 _ser_kac_ovflw(0), 475 _par_kac_ovflw(0), 476 #ifndef PRODUCT 477 _num_par_pushes(0), 478 #endif 479 _collection_count_start(0), 480 _verifying(false), 481 _verification_mark_bm(0, Mutex::leaf + 1, "CMS_verification_mark_bm_lock"), 482 _completed_initialization(false), 483 _collector_policy(cp), 484 _should_unload_classes(CMSClassUnloadingEnabled), 485 _concurrent_cycles_since_last_unload(0), 486 _roots_scanning_options(GenCollectedHeap::SO_None), 487 _inter_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding), 488 _intra_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding), 489 _gc_tracer_cm(new (ResourceObj::C_HEAP, mtGC) CMSTracer()), 490 _gc_timer_cm(new (ResourceObj::C_HEAP, mtGC) ConcurrentGCTimer()), 491 _cms_start_registered(false) 492 { 493 // Now expand the span and allocate the collection support structures 494 // (MUT, marking bit map etc.) to cover both generations subject to 495 // collection. 496 497 // For use by dirty card to oop closures. 498 _cmsGen->cmsSpace()->set_collector(this); 499 500 // Allocate MUT and marking bit map 501 { 502 MutexLockerEx x(_markBitMap.lock(), Mutex::_no_safepoint_check_flag); 503 if (!_markBitMap.allocate(_span)) { 504 log_warning(gc)("Failed to allocate CMS Bit Map"); 505 return; 506 } 507 assert(_markBitMap.covers(_span), "_markBitMap inconsistency?"); 508 } 509 { 510 _modUnionTable.allocate(_span); 511 assert(_modUnionTable.covers(_span), "_modUnionTable inconsistency?"); 512 } 513 514 if (!_markStack.allocate(MarkStackSize)) { 515 log_warning(gc)("Failed to allocate CMS Marking Stack"); 516 return; 517 } 518 519 // Support for multi-threaded concurrent phases 520 if (CMSConcurrentMTEnabled) { 521 if (FLAG_IS_DEFAULT(ConcGCThreads)) { 522 // just for now 523 FLAG_SET_DEFAULT(ConcGCThreads, (ParallelGCThreads + 3) / 4); 524 } 525 if (ConcGCThreads > 1) { 526 _conc_workers = new YieldingFlexibleWorkGang("CMS Thread", 527 ConcGCThreads, true); 528 if (_conc_workers == NULL) { 529 log_warning(gc)("GC/CMS: _conc_workers allocation failure: forcing -CMSConcurrentMTEnabled"); 530 CMSConcurrentMTEnabled = false; 531 } else { 532 _conc_workers->initialize_workers(); 533 } 534 } else { 535 CMSConcurrentMTEnabled = false; 536 } 537 } 538 if (!CMSConcurrentMTEnabled) { 539 ConcGCThreads = 0; 540 } else { 541 // Turn off CMSCleanOnEnter optimization temporarily for 542 // the MT case where it's not fixed yet; see 6178663. 543 CMSCleanOnEnter = false; 544 } 545 assert((_conc_workers != NULL) == (ConcGCThreads > 1), 546 "Inconsistency"); 547 log_debug(gc)("ConcGCThreads: %u", ConcGCThreads); 548 log_debug(gc)("ParallelGCThreads: %u", ParallelGCThreads); 549 550 // Parallel task queues; these are shared for the 551 // concurrent and stop-world phases of CMS, but 552 // are not shared with parallel scavenge (ParNew). 553 { 554 uint i; 555 uint num_queues = MAX2(ParallelGCThreads, ConcGCThreads); 556 557 if ((CMSParallelRemarkEnabled || CMSConcurrentMTEnabled 558 || ParallelRefProcEnabled) 559 && num_queues > 0) { 560 _task_queues = new OopTaskQueueSet(num_queues); 561 if (_task_queues == NULL) { 562 log_warning(gc)("task_queues allocation failure."); 563 return; 564 } 565 _hash_seed = NEW_C_HEAP_ARRAY(int, num_queues, mtGC); 566 typedef Padded<OopTaskQueue> PaddedOopTaskQueue; 567 for (i = 0; i < num_queues; i++) { 568 PaddedOopTaskQueue *q = new PaddedOopTaskQueue(); 569 if (q == NULL) { 570 log_warning(gc)("work_queue allocation failure."); 571 return; 572 } 573 _task_queues->register_queue(i, q); 574 } 575 for (i = 0; i < num_queues; i++) { 576 _task_queues->queue(i)->initialize(); 577 _hash_seed[i] = 17; // copied from ParNew 578 } 579 } 580 } 581 582 _cmsGen ->init_initiating_occupancy(CMSInitiatingOccupancyFraction, CMSTriggerRatio); 583 584 // Clip CMSBootstrapOccupancy between 0 and 100. 585 _bootstrap_occupancy = CMSBootstrapOccupancy / 100.0; 586 587 // Now tell CMS generations the identity of their collector 588 ConcurrentMarkSweepGeneration::set_collector(this); 589 590 // Create & start a CMS thread for this CMS collector 591 _cmsThread = ConcurrentMarkSweepThread::start(this); 592 assert(cmsThread() != NULL, "CMS Thread should have been created"); 593 assert(cmsThread()->collector() == this, 594 "CMS Thread should refer to this gen"); 595 assert(CGC_lock != NULL, "Where's the CGC_lock?"); 596 597 // Support for parallelizing young gen rescan 598 GenCollectedHeap* gch = GenCollectedHeap::heap(); 599 assert(gch->young_gen()->kind() == Generation::ParNew, "CMS can only be used with ParNew"); 600 _young_gen = (ParNewGeneration*)gch->young_gen(); 601 if (gch->supports_inline_contig_alloc()) { 602 _top_addr = gch->top_addr(); 603 _end_addr = gch->end_addr(); 604 assert(_young_gen != NULL, "no _young_gen"); 605 _eden_chunk_index = 0; 606 _eden_chunk_capacity = (_young_gen->max_capacity() + CMSSamplingGrain) / CMSSamplingGrain; 607 _eden_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, _eden_chunk_capacity, mtGC); 608 } 609 610 // Support for parallelizing survivor space rescan 611 if ((CMSParallelRemarkEnabled && CMSParallelSurvivorRemarkEnabled) || CMSParallelInitialMarkEnabled) { 612 const size_t max_plab_samples = 613 _young_gen->max_survivor_size() / (PLAB::min_size() * HeapWordSize); 614 615 _survivor_plab_array = NEW_C_HEAP_ARRAY(ChunkArray, ParallelGCThreads, mtGC); 616 _survivor_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples, mtGC); 617 _cursor = NEW_C_HEAP_ARRAY(size_t, ParallelGCThreads, mtGC); 618 _survivor_chunk_capacity = max_plab_samples; 619 for (uint i = 0; i < ParallelGCThreads; i++) { 620 HeapWord** vec = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples, mtGC); 621 ChunkArray* cur = ::new (&_survivor_plab_array[i]) ChunkArray(vec, max_plab_samples); 622 assert(cur->end() == 0, "Should be 0"); 623 assert(cur->array() == vec, "Should be vec"); 624 assert(cur->capacity() == max_plab_samples, "Error"); 625 } 626 } 627 628 NOT_PRODUCT(_overflow_counter = CMSMarkStackOverflowInterval;) 629 _gc_counters = new CollectorCounters("CMS", 1); 630 _completed_initialization = true; 631 _inter_sweep_timer.start(); // start of time 632 } 633 634 const char* ConcurrentMarkSweepGeneration::name() const { 635 return "concurrent mark-sweep generation"; 636 } 637 void ConcurrentMarkSweepGeneration::update_counters() { 638 if (UsePerfData) { 639 _space_counters->update_all(); 640 _gen_counters->update_all(); 641 } 642 } 643 644 // this is an optimized version of update_counters(). it takes the 645 // used value as a parameter rather than computing it. 646 // 647 void ConcurrentMarkSweepGeneration::update_counters(size_t used) { 648 if (UsePerfData) { 649 _space_counters->update_used(used); 650 _space_counters->update_capacity(); 651 _gen_counters->update_all(); 652 } 653 } 654 655 void ConcurrentMarkSweepGeneration::print() const { 656 Generation::print(); 657 cmsSpace()->print(); 658 } 659 660 #ifndef PRODUCT 661 void ConcurrentMarkSweepGeneration::print_statistics() { 662 cmsSpace()->printFLCensus(0); 663 } 664 #endif 665 666 size_t 667 ConcurrentMarkSweepGeneration::contiguous_available() const { 668 // dld proposes an improvement in precision here. If the committed 669 // part of the space ends in a free block we should add that to 670 // uncommitted size in the calculation below. Will make this 671 // change later, staying with the approximation below for the 672 // time being. -- ysr. 673 return MAX2(_virtual_space.uncommitted_size(), unsafe_max_alloc_nogc()); 674 } 675 676 size_t 677 ConcurrentMarkSweepGeneration::unsafe_max_alloc_nogc() const { 678 return _cmsSpace->max_alloc_in_words() * HeapWordSize; 679 } 680 681 size_t ConcurrentMarkSweepGeneration::max_available() const { 682 return free() + _virtual_space.uncommitted_size(); 683 } 684 685 bool ConcurrentMarkSweepGeneration::promotion_attempt_is_safe(size_t max_promotion_in_bytes) const { 686 size_t available = max_available(); 687 size_t av_promo = (size_t)gc_stats()->avg_promoted()->padded_average(); 688 bool res = (available >= av_promo) || (available >= max_promotion_in_bytes); 689 log_trace(gc, promotion)("CMS: promo attempt is%s safe: available(" SIZE_FORMAT ") %s av_promo(" SIZE_FORMAT "), max_promo(" SIZE_FORMAT ")", 690 res? "":" not", available, res? ">=":"<", av_promo, max_promotion_in_bytes); 691 return res; 692 } 693 694 // At a promotion failure dump information on block layout in heap 695 // (cms old generation). 696 void ConcurrentMarkSweepGeneration::promotion_failure_occurred() { 697 Log(gc, promotion) log; 698 if (log.is_trace()) { 699 LogStream ls(log.trace()); 700 cmsSpace()->dump_at_safepoint_with_locks(collector(), &ls); 701 } 702 } 703 704 void ConcurrentMarkSweepGeneration::reset_after_compaction() { 705 // Clear the promotion information. These pointers can be adjusted 706 // along with all the other pointers into the heap but 707 // compaction is expected to be a rare event with 708 // a heap using cms so don't do it without seeing the need. 709 for (uint i = 0; i < ParallelGCThreads; i++) { 710 _par_gc_thread_states[i]->promo.reset(); 711 } 712 } 713 714 void ConcurrentMarkSweepGeneration::compute_new_size() { 715 assert_locked_or_safepoint(Heap_lock); 716 717 // If incremental collection failed, we just want to expand 718 // to the limit. 719 if (incremental_collection_failed()) { 720 clear_incremental_collection_failed(); 721 grow_to_reserved(); 722 return; 723 } 724 725 // The heap has been compacted but not reset yet. 726 // Any metric such as free() or used() will be incorrect. 727 728 CardGeneration::compute_new_size(); 729 730 // Reset again after a possible resizing 731 if (did_compact()) { 732 cmsSpace()->reset_after_compaction(); 733 } 734 } 735 736 void ConcurrentMarkSweepGeneration::compute_new_size_free_list() { 737 assert_locked_or_safepoint(Heap_lock); 738 739 // If incremental collection failed, we just want to expand 740 // to the limit. 741 if (incremental_collection_failed()) { 742 clear_incremental_collection_failed(); 743 grow_to_reserved(); 744 return; 745 } 746 747 double free_percentage = ((double) free()) / capacity(); 748 double desired_free_percentage = (double) MinHeapFreeRatio / 100; 749 double maximum_free_percentage = (double) MaxHeapFreeRatio / 100; 750 751 // compute expansion delta needed for reaching desired free percentage 752 if (free_percentage < desired_free_percentage) { 753 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 754 assert(desired_capacity >= capacity(), "invalid expansion size"); 755 size_t expand_bytes = MAX2(desired_capacity - capacity(), MinHeapDeltaBytes); 756 Log(gc) log; 757 if (log.is_trace()) { 758 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 759 log.trace("From compute_new_size: "); 760 log.trace(" Free fraction %f", free_percentage); 761 log.trace(" Desired free fraction %f", desired_free_percentage); 762 log.trace(" Maximum free fraction %f", maximum_free_percentage); 763 log.trace(" Capacity " SIZE_FORMAT, capacity() / 1000); 764 log.trace(" Desired capacity " SIZE_FORMAT, desired_capacity / 1000); 765 GenCollectedHeap* gch = GenCollectedHeap::heap(); 766 assert(gch->is_old_gen(this), "The CMS generation should always be the old generation"); 767 size_t young_size = gch->young_gen()->capacity(); 768 log.trace(" Young gen size " SIZE_FORMAT, young_size / 1000); 769 log.trace(" unsafe_max_alloc_nogc " SIZE_FORMAT, unsafe_max_alloc_nogc() / 1000); 770 log.trace(" contiguous available " SIZE_FORMAT, contiguous_available() / 1000); 771 log.trace(" Expand by " SIZE_FORMAT " (bytes)", expand_bytes); 772 } 773 // safe if expansion fails 774 expand_for_gc_cause(expand_bytes, 0, CMSExpansionCause::_satisfy_free_ratio); 775 log.trace(" Expanded free fraction %f", ((double) free()) / capacity()); 776 } else { 777 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 778 assert(desired_capacity <= capacity(), "invalid expansion size"); 779 size_t shrink_bytes = capacity() - desired_capacity; 780 // Don't shrink unless the delta is greater than the minimum shrink we want 781 if (shrink_bytes >= MinHeapDeltaBytes) { 782 shrink_free_list_by(shrink_bytes); 783 } 784 } 785 } 786 787 Mutex* ConcurrentMarkSweepGeneration::freelistLock() const { 788 return cmsSpace()->freelistLock(); 789 } 790 791 HeapWord* ConcurrentMarkSweepGeneration::allocate(size_t size, bool tlab) { 792 CMSSynchronousYieldRequest yr; 793 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 794 return have_lock_and_allocate(size, tlab); 795 } 796 797 HeapWord* ConcurrentMarkSweepGeneration::have_lock_and_allocate(size_t size, 798 bool tlab /* ignored */) { 799 assert_lock_strong(freelistLock()); 800 size_t adjustedSize = CompactibleFreeListSpace::adjustObjectSize(size); 801 HeapWord* res = cmsSpace()->allocate(adjustedSize); 802 // Allocate the object live (grey) if the background collector has 803 // started marking. This is necessary because the marker may 804 // have passed this address and consequently this object will 805 // not otherwise be greyed and would be incorrectly swept up. 806 // Note that if this object contains references, the writing 807 // of those references will dirty the card containing this object 808 // allowing the object to be blackened (and its references scanned) 809 // either during a preclean phase or at the final checkpoint. 810 if (res != NULL) { 811 // We may block here with an uninitialized object with 812 // its mark-bit or P-bits not yet set. Such objects need 813 // to be safely navigable by block_start(). 814 assert(oop(res)->klass_or_null() == NULL, "Object should be uninitialized here."); 815 assert(!((FreeChunk*)res)->is_free(), "Error, block will look free but show wrong size"); 816 collector()->direct_allocated(res, adjustedSize); 817 _direct_allocated_words += adjustedSize; 818 // allocation counters 819 NOT_PRODUCT( 820 _numObjectsAllocated++; 821 _numWordsAllocated += (int)adjustedSize; 822 ) 823 } 824 return res; 825 } 826 827 // In the case of direct allocation by mutators in a generation that 828 // is being concurrently collected, the object must be allocated 829 // live (grey) if the background collector has started marking. 830 // This is necessary because the marker may 831 // have passed this address and consequently this object will 832 // not otherwise be greyed and would be incorrectly swept up. 833 // Note that if this object contains references, the writing 834 // of those references will dirty the card containing this object 835 // allowing the object to be blackened (and its references scanned) 836 // either during a preclean phase or at the final checkpoint. 837 void CMSCollector::direct_allocated(HeapWord* start, size_t size) { 838 assert(_markBitMap.covers(start, size), "Out of bounds"); 839 if (_collectorState >= Marking) { 840 MutexLockerEx y(_markBitMap.lock(), 841 Mutex::_no_safepoint_check_flag); 842 // [see comments preceding SweepClosure::do_blk() below for details] 843 // 844 // Can the P-bits be deleted now? JJJ 845 // 846 // 1. need to mark the object as live so it isn't collected 847 // 2. need to mark the 2nd bit to indicate the object may be uninitialized 848 // 3. need to mark the end of the object so marking, precleaning or sweeping 849 // can skip over uninitialized or unparsable objects. An allocated 850 // object is considered uninitialized for our purposes as long as 851 // its klass word is NULL. All old gen objects are parsable 852 // as soon as they are initialized.) 853 _markBitMap.mark(start); // object is live 854 _markBitMap.mark(start + 1); // object is potentially uninitialized? 855 _markBitMap.mark(start + size - 1); 856 // mark end of object 857 } 858 // check that oop looks uninitialized 859 assert(oop(start)->klass_or_null() == NULL, "_klass should be NULL"); 860 } 861 862 void CMSCollector::promoted(bool par, HeapWord* start, 863 bool is_obj_array, size_t obj_size) { 864 assert(_markBitMap.covers(start), "Out of bounds"); 865 // See comment in direct_allocated() about when objects should 866 // be allocated live. 867 if (_collectorState >= Marking) { 868 // we already hold the marking bit map lock, taken in 869 // the prologue 870 if (par) { 871 _markBitMap.par_mark(start); 872 } else { 873 _markBitMap.mark(start); 874 } 875 // We don't need to mark the object as uninitialized (as 876 // in direct_allocated above) because this is being done with the 877 // world stopped and the object will be initialized by the 878 // time the marking, precleaning or sweeping get to look at it. 879 // But see the code for copying objects into the CMS generation, 880 // where we need to ensure that concurrent readers of the 881 // block offset table are able to safely navigate a block that 882 // is in flux from being free to being allocated (and in 883 // transition while being copied into) and subsequently 884 // becoming a bona-fide object when the copy/promotion is complete. 885 assert(SafepointSynchronize::is_at_safepoint(), 886 "expect promotion only at safepoints"); 887 888 if (_collectorState < Sweeping) { 889 // Mark the appropriate cards in the modUnionTable, so that 890 // this object gets scanned before the sweep. If this is 891 // not done, CMS generation references in the object might 892 // not get marked. 893 // For the case of arrays, which are otherwise precisely 894 // marked, we need to dirty the entire array, not just its head. 895 if (is_obj_array) { 896 // The [par_]mark_range() method expects mr.end() below to 897 // be aligned to the granularity of a bit's representation 898 // in the heap. In the case of the MUT below, that's a 899 // card size. 900 MemRegion mr(start, 901 align_up(start + obj_size, 902 CardTableModRefBS::card_size /* bytes */)); 903 if (par) { 904 _modUnionTable.par_mark_range(mr); 905 } else { 906 _modUnionTable.mark_range(mr); 907 } 908 } else { // not an obj array; we can just mark the head 909 if (par) { 910 _modUnionTable.par_mark(start); 911 } else { 912 _modUnionTable.mark(start); 913 } 914 } 915 } 916 } 917 } 918 919 oop ConcurrentMarkSweepGeneration::promote(oop obj, size_t obj_size) { 920 assert(obj_size == (size_t)obj->size(), "bad obj_size passed in"); 921 // allocate, copy and if necessary update promoinfo -- 922 // delegate to underlying space. 923 assert_lock_strong(freelistLock()); 924 925 #ifndef PRODUCT 926 if (GenCollectedHeap::heap()->promotion_should_fail()) { 927 return NULL; 928 } 929 #endif // #ifndef PRODUCT 930 931 oop res = _cmsSpace->promote(obj, obj_size); 932 if (res == NULL) { 933 // expand and retry 934 size_t s = _cmsSpace->expansionSpaceRequired(obj_size); // HeapWords 935 expand_for_gc_cause(s*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_satisfy_promotion); 936 // Since this is the old generation, we don't try to promote 937 // into a more senior generation. 938 res = _cmsSpace->promote(obj, obj_size); 939 } 940 if (res != NULL) { 941 // See comment in allocate() about when objects should 942 // be allocated live. 943 assert(obj->is_oop(), "Will dereference klass pointer below"); 944 collector()->promoted(false, // Not parallel 945 (HeapWord*)res, obj->is_objArray(), obj_size); 946 // promotion counters 947 NOT_PRODUCT( 948 _numObjectsPromoted++; 949 _numWordsPromoted += 950 (int)(CompactibleFreeListSpace::adjustObjectSize(obj->size())); 951 ) 952 } 953 return res; 954 } 955 956 957 // IMPORTANT: Notes on object size recognition in CMS. 958 // --------------------------------------------------- 959 // A block of storage in the CMS generation is always in 960 // one of three states. A free block (FREE), an allocated 961 // object (OBJECT) whose size() method reports the correct size, 962 // and an intermediate state (TRANSIENT) in which its size cannot 963 // be accurately determined. 964 // STATE IDENTIFICATION: (32 bit and 64 bit w/o COOPS) 965 // ----------------------------------------------------- 966 // FREE: klass_word & 1 == 1; mark_word holds block size 967 // 968 // OBJECT: klass_word installed; klass_word != 0 && klass_word & 1 == 0; 969 // obj->size() computes correct size 970 // 971 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT 972 // 973 // STATE IDENTIFICATION: (64 bit+COOPS) 974 // ------------------------------------ 975 // FREE: mark_word & CMS_FREE_BIT == 1; mark_word & ~CMS_FREE_BIT gives block_size 976 // 977 // OBJECT: klass_word installed; klass_word != 0; 978 // obj->size() computes correct size 979 // 980 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT 981 // 982 // 983 // STATE TRANSITION DIAGRAM 984 // 985 // mut / parnew mut / parnew 986 // FREE --------------------> TRANSIENT ---------------------> OBJECT --| 987 // ^ | 988 // |------------------------ DEAD <------------------------------------| 989 // sweep mut 990 // 991 // While a block is in TRANSIENT state its size cannot be determined 992 // so readers will either need to come back later or stall until 993 // the size can be determined. Note that for the case of direct 994 // allocation, P-bits, when available, may be used to determine the 995 // size of an object that may not yet have been initialized. 996 997 // Things to support parallel young-gen collection. 998 oop 999 ConcurrentMarkSweepGeneration::par_promote(int thread_num, 1000 oop old, markOop m, 1001 size_t word_sz) { 1002 #ifndef PRODUCT 1003 if (GenCollectedHeap::heap()->promotion_should_fail()) { 1004 return NULL; 1005 } 1006 #endif // #ifndef PRODUCT 1007 1008 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1009 PromotionInfo* promoInfo = &ps->promo; 1010 // if we are tracking promotions, then first ensure space for 1011 // promotion (including spooling space for saving header if necessary). 1012 // then allocate and copy, then track promoted info if needed. 1013 // When tracking (see PromotionInfo::track()), the mark word may 1014 // be displaced and in this case restoration of the mark word 1015 // occurs in the (oop_since_save_marks_)iterate phase. 1016 if (promoInfo->tracking() && !promoInfo->ensure_spooling_space()) { 1017 // Out of space for allocating spooling buffers; 1018 // try expanding and allocating spooling buffers. 1019 if (!expand_and_ensure_spooling_space(promoInfo)) { 1020 return NULL; 1021 } 1022 } 1023 assert(!promoInfo->tracking() || promoInfo->has_spooling_space(), "Control point invariant"); 1024 const size_t alloc_sz = CompactibleFreeListSpace::adjustObjectSize(word_sz); 1025 HeapWord* obj_ptr = ps->lab.alloc(alloc_sz); 1026 if (obj_ptr == NULL) { 1027 obj_ptr = expand_and_par_lab_allocate(ps, alloc_sz); 1028 if (obj_ptr == NULL) { 1029 return NULL; 1030 } 1031 } 1032 oop obj = oop(obj_ptr); 1033 OrderAccess::storestore(); 1034 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1035 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1036 // IMPORTANT: See note on object initialization for CMS above. 1037 // Otherwise, copy the object. Here we must be careful to insert the 1038 // klass pointer last, since this marks the block as an allocated object. 1039 // Except with compressed oops it's the mark word. 1040 HeapWord* old_ptr = (HeapWord*)old; 1041 // Restore the mark word copied above. 1042 obj->set_mark(m); 1043 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1044 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1045 OrderAccess::storestore(); 1046 1047 if (UseCompressedClassPointers) { 1048 // Copy gap missed by (aligned) header size calculation below 1049 obj->set_klass_gap(old->klass_gap()); 1050 } 1051 if (word_sz > (size_t)oopDesc::header_size()) { 1052 Copy::aligned_disjoint_words(old_ptr + oopDesc::header_size(), 1053 obj_ptr + oopDesc::header_size(), 1054 word_sz - oopDesc::header_size()); 1055 } 1056 1057 // Now we can track the promoted object, if necessary. We take care 1058 // to delay the transition from uninitialized to full object 1059 // (i.e., insertion of klass pointer) until after, so that it 1060 // atomically becomes a promoted object. 1061 if (promoInfo->tracking()) { 1062 promoInfo->track((PromotedObject*)obj, old->klass()); 1063 } 1064 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1065 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1066 assert(old->is_oop(), "Will use and dereference old klass ptr below"); 1067 1068 // Finally, install the klass pointer (this should be volatile). 1069 OrderAccess::storestore(); 1070 obj->set_klass(old->klass()); 1071 // We should now be able to calculate the right size for this object 1072 assert(obj->is_oop() && obj->size() == (int)word_sz, "Error, incorrect size computed for promoted object"); 1073 1074 collector()->promoted(true, // parallel 1075 obj_ptr, old->is_objArray(), word_sz); 1076 1077 NOT_PRODUCT( 1078 Atomic::inc_ptr(&_numObjectsPromoted); 1079 Atomic::add_ptr(alloc_sz, &_numWordsPromoted); 1080 ) 1081 1082 return obj; 1083 } 1084 1085 void 1086 ConcurrentMarkSweepGeneration:: 1087 par_promote_alloc_done(int thread_num) { 1088 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1089 ps->lab.retire(thread_num); 1090 } 1091 1092 void 1093 ConcurrentMarkSweepGeneration:: 1094 par_oop_since_save_marks_iterate_done(int thread_num) { 1095 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1096 ParScanWithoutBarrierClosure* dummy_cl = NULL; 1097 ps->promo.promoted_oops_iterate_nv(dummy_cl); 1098 1099 // Because card-scanning has been completed, subsequent phases 1100 // (e.g., reference processing) will not need to recognize which 1101 // objects have been promoted during this GC. So, we can now disable 1102 // promotion tracking. 1103 ps->promo.stopTrackingPromotions(); 1104 } 1105 1106 bool ConcurrentMarkSweepGeneration::should_collect(bool full, 1107 size_t size, 1108 bool tlab) 1109 { 1110 // We allow a STW collection only if a full 1111 // collection was requested. 1112 return full || should_allocate(size, tlab); // FIX ME !!! 1113 // This and promotion failure handling are connected at the 1114 // hip and should be fixed by untying them. 1115 } 1116 1117 bool CMSCollector::shouldConcurrentCollect() { 1118 LogTarget(Trace, gc) log; 1119 1120 if (_full_gc_requested) { 1121 log.print("CMSCollector: collect because of explicit gc request (or GCLocker)"); 1122 return true; 1123 } 1124 1125 FreelistLocker x(this); 1126 // ------------------------------------------------------------------ 1127 // Print out lots of information which affects the initiation of 1128 // a collection. 1129 if (log.is_enabled() && stats().valid()) { 1130 log.print("CMSCollector shouldConcurrentCollect: "); 1131 1132 LogStream out(log); 1133 stats().print_on(&out); 1134 1135 log.print("time_until_cms_gen_full %3.7f", stats().time_until_cms_gen_full()); 1136 log.print("free=" SIZE_FORMAT, _cmsGen->free()); 1137 log.print("contiguous_available=" SIZE_FORMAT, _cmsGen->contiguous_available()); 1138 log.print("promotion_rate=%g", stats().promotion_rate()); 1139 log.print("cms_allocation_rate=%g", stats().cms_allocation_rate()); 1140 log.print("occupancy=%3.7f", _cmsGen->occupancy()); 1141 log.print("initiatingOccupancy=%3.7f", _cmsGen->initiating_occupancy()); 1142 log.print("cms_time_since_begin=%3.7f", stats().cms_time_since_begin()); 1143 log.print("cms_time_since_end=%3.7f", stats().cms_time_since_end()); 1144 log.print("metadata initialized %d", MetaspaceGC::should_concurrent_collect()); 1145 } 1146 // ------------------------------------------------------------------ 1147 1148 // If the estimated time to complete a cms collection (cms_duration()) 1149 // is less than the estimated time remaining until the cms generation 1150 // is full, start a collection. 1151 if (!UseCMSInitiatingOccupancyOnly) { 1152 if (stats().valid()) { 1153 if (stats().time_until_cms_start() == 0.0) { 1154 return true; 1155 } 1156 } else { 1157 // We want to conservatively collect somewhat early in order 1158 // to try and "bootstrap" our CMS/promotion statistics; 1159 // this branch will not fire after the first successful CMS 1160 // collection because the stats should then be valid. 1161 if (_cmsGen->occupancy() >= _bootstrap_occupancy) { 1162 log.print(" CMSCollector: collect for bootstrapping statistics: occupancy = %f, boot occupancy = %f", 1163 _cmsGen->occupancy(), _bootstrap_occupancy); 1164 return true; 1165 } 1166 } 1167 } 1168 1169 // Otherwise, we start a collection cycle if 1170 // old gen want a collection cycle started. Each may use 1171 // an appropriate criterion for making this decision. 1172 // XXX We need to make sure that the gen expansion 1173 // criterion dovetails well with this. XXX NEED TO FIX THIS 1174 if (_cmsGen->should_concurrent_collect()) { 1175 log.print("CMS old gen initiated"); 1176 return true; 1177 } 1178 1179 // We start a collection if we believe an incremental collection may fail; 1180 // this is not likely to be productive in practice because it's probably too 1181 // late anyway. 1182 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1183 assert(gch->collector_policy()->is_generation_policy(), 1184 "You may want to check the correctness of the following"); 1185 if (gch->incremental_collection_will_fail(true /* consult_young */)) { 1186 log.print("CMSCollector: collect because incremental collection will fail "); 1187 return true; 1188 } 1189 1190 if (MetaspaceGC::should_concurrent_collect()) { 1191 log.print("CMSCollector: collect for metadata allocation "); 1192 return true; 1193 } 1194 1195 // CMSTriggerInterval starts a CMS cycle if enough time has passed. 1196 if (CMSTriggerInterval >= 0) { 1197 if (CMSTriggerInterval == 0) { 1198 // Trigger always 1199 return true; 1200 } 1201 1202 // Check the CMS time since begin (we do not check the stats validity 1203 // as we want to be able to trigger the first CMS cycle as well) 1204 if (stats().cms_time_since_begin() >= (CMSTriggerInterval / ((double) MILLIUNITS))) { 1205 if (stats().valid()) { 1206 log.print("CMSCollector: collect because of trigger interval (time since last begin %3.7f secs)", 1207 stats().cms_time_since_begin()); 1208 } else { 1209 log.print("CMSCollector: collect because of trigger interval (first collection)"); 1210 } 1211 return true; 1212 } 1213 } 1214 1215 return false; 1216 } 1217 1218 void CMSCollector::set_did_compact(bool v) { _cmsGen->set_did_compact(v); } 1219 1220 // Clear _expansion_cause fields of constituent generations 1221 void CMSCollector::clear_expansion_cause() { 1222 _cmsGen->clear_expansion_cause(); 1223 } 1224 1225 // We should be conservative in starting a collection cycle. To 1226 // start too eagerly runs the risk of collecting too often in the 1227 // extreme. To collect too rarely falls back on full collections, 1228 // which works, even if not optimum in terms of concurrent work. 1229 // As a work around for too eagerly collecting, use the flag 1230 // UseCMSInitiatingOccupancyOnly. This also has the advantage of 1231 // giving the user an easily understandable way of controlling the 1232 // collections. 1233 // We want to start a new collection cycle if any of the following 1234 // conditions hold: 1235 // . our current occupancy exceeds the configured initiating occupancy 1236 // for this generation, or 1237 // . we recently needed to expand this space and have not, since that 1238 // expansion, done a collection of this generation, or 1239 // . the underlying space believes that it may be a good idea to initiate 1240 // a concurrent collection (this may be based on criteria such as the 1241 // following: the space uses linear allocation and linear allocation is 1242 // going to fail, or there is believed to be excessive fragmentation in 1243 // the generation, etc... or ... 1244 // [.(currently done by CMSCollector::shouldConcurrentCollect() only for 1245 // the case of the old generation; see CR 6543076): 1246 // we may be approaching a point at which allocation requests may fail because 1247 // we will be out of sufficient free space given allocation rate estimates.] 1248 bool ConcurrentMarkSweepGeneration::should_concurrent_collect() const { 1249 1250 assert_lock_strong(freelistLock()); 1251 if (occupancy() > initiating_occupancy()) { 1252 log_trace(gc)(" %s: collect because of occupancy %f / %f ", 1253 short_name(), occupancy(), initiating_occupancy()); 1254 return true; 1255 } 1256 if (UseCMSInitiatingOccupancyOnly) { 1257 return false; 1258 } 1259 if (expansion_cause() == CMSExpansionCause::_satisfy_allocation) { 1260 log_trace(gc)(" %s: collect because expanded for allocation ", short_name()); 1261 return true; 1262 } 1263 return false; 1264 } 1265 1266 void ConcurrentMarkSweepGeneration::collect(bool full, 1267 bool clear_all_soft_refs, 1268 size_t size, 1269 bool tlab) 1270 { 1271 collector()->collect(full, clear_all_soft_refs, size, tlab); 1272 } 1273 1274 void CMSCollector::collect(bool full, 1275 bool clear_all_soft_refs, 1276 size_t size, 1277 bool tlab) 1278 { 1279 // The following "if" branch is present for defensive reasons. 1280 // In the current uses of this interface, it can be replaced with: 1281 // assert(!GCLocker.is_active(), "Can't be called otherwise"); 1282 // But I am not placing that assert here to allow future 1283 // generality in invoking this interface. 1284 if (GCLocker::is_active()) { 1285 // A consistency test for GCLocker 1286 assert(GCLocker::needs_gc(), "Should have been set already"); 1287 // Skip this foreground collection, instead 1288 // expanding the heap if necessary. 1289 // Need the free list locks for the call to free() in compute_new_size() 1290 compute_new_size(); 1291 return; 1292 } 1293 acquire_control_and_collect(full, clear_all_soft_refs); 1294 } 1295 1296 void CMSCollector::request_full_gc(unsigned int full_gc_count, GCCause::Cause cause) { 1297 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1298 unsigned int gc_count = gch->total_full_collections(); 1299 if (gc_count == full_gc_count) { 1300 MutexLockerEx y(CGC_lock, Mutex::_no_safepoint_check_flag); 1301 _full_gc_requested = true; 1302 _full_gc_cause = cause; 1303 CGC_lock->notify(); // nudge CMS thread 1304 } else { 1305 assert(gc_count > full_gc_count, "Error: causal loop"); 1306 } 1307 } 1308 1309 bool CMSCollector::is_external_interruption() { 1310 GCCause::Cause cause = GenCollectedHeap::heap()->gc_cause(); 1311 return GCCause::is_user_requested_gc(cause) || 1312 GCCause::is_serviceability_requested_gc(cause); 1313 } 1314 1315 void CMSCollector::report_concurrent_mode_interruption() { 1316 if (is_external_interruption()) { 1317 log_debug(gc)("Concurrent mode interrupted"); 1318 } else { 1319 log_debug(gc)("Concurrent mode failure"); 1320 _gc_tracer_cm->report_concurrent_mode_failure(); 1321 } 1322 } 1323 1324 1325 // The foreground and background collectors need to coordinate in order 1326 // to make sure that they do not mutually interfere with CMS collections. 1327 // When a background collection is active, 1328 // the foreground collector may need to take over (preempt) and 1329 // synchronously complete an ongoing collection. Depending on the 1330 // frequency of the background collections and the heap usage 1331 // of the application, this preemption can be seldom or frequent. 1332 // There are only certain 1333 // points in the background collection that the "collection-baton" 1334 // can be passed to the foreground collector. 1335 // 1336 // The foreground collector will wait for the baton before 1337 // starting any part of the collection. The foreground collector 1338 // will only wait at one location. 1339 // 1340 // The background collector will yield the baton before starting a new 1341 // phase of the collection (e.g., before initial marking, marking from roots, 1342 // precleaning, final re-mark, sweep etc.) This is normally done at the head 1343 // of the loop which switches the phases. The background collector does some 1344 // of the phases (initial mark, final re-mark) with the world stopped. 1345 // Because of locking involved in stopping the world, 1346 // the foreground collector should not block waiting for the background 1347 // collector when it is doing a stop-the-world phase. The background 1348 // collector will yield the baton at an additional point just before 1349 // it enters a stop-the-world phase. Once the world is stopped, the 1350 // background collector checks the phase of the collection. If the 1351 // phase has not changed, it proceeds with the collection. If the 1352 // phase has changed, it skips that phase of the collection. See 1353 // the comments on the use of the Heap_lock in collect_in_background(). 1354 // 1355 // Variable used in baton passing. 1356 // _foregroundGCIsActive - Set to true by the foreground collector when 1357 // it wants the baton. The foreground clears it when it has finished 1358 // the collection. 1359 // _foregroundGCShouldWait - Set to true by the background collector 1360 // when it is running. The foreground collector waits while 1361 // _foregroundGCShouldWait is true. 1362 // CGC_lock - monitor used to protect access to the above variables 1363 // and to notify the foreground and background collectors. 1364 // _collectorState - current state of the CMS collection. 1365 // 1366 // The foreground collector 1367 // acquires the CGC_lock 1368 // sets _foregroundGCIsActive 1369 // waits on the CGC_lock for _foregroundGCShouldWait to be false 1370 // various locks acquired in preparation for the collection 1371 // are released so as not to block the background collector 1372 // that is in the midst of a collection 1373 // proceeds with the collection 1374 // clears _foregroundGCIsActive 1375 // returns 1376 // 1377 // The background collector in a loop iterating on the phases of the 1378 // collection 1379 // acquires the CGC_lock 1380 // sets _foregroundGCShouldWait 1381 // if _foregroundGCIsActive is set 1382 // clears _foregroundGCShouldWait, notifies _CGC_lock 1383 // waits on _CGC_lock for _foregroundGCIsActive to become false 1384 // and exits the loop. 1385 // otherwise 1386 // proceed with that phase of the collection 1387 // if the phase is a stop-the-world phase, 1388 // yield the baton once more just before enqueueing 1389 // the stop-world CMS operation (executed by the VM thread). 1390 // returns after all phases of the collection are done 1391 // 1392 1393 void CMSCollector::acquire_control_and_collect(bool full, 1394 bool clear_all_soft_refs) { 1395 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint"); 1396 assert(!Thread::current()->is_ConcurrentGC_thread(), 1397 "shouldn't try to acquire control from self!"); 1398 1399 // Start the protocol for acquiring control of the 1400 // collection from the background collector (aka CMS thread). 1401 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 1402 "VM thread should have CMS token"); 1403 // Remember the possibly interrupted state of an ongoing 1404 // concurrent collection 1405 CollectorState first_state = _collectorState; 1406 1407 // Signal to a possibly ongoing concurrent collection that 1408 // we want to do a foreground collection. 1409 _foregroundGCIsActive = true; 1410 1411 // release locks and wait for a notify from the background collector 1412 // releasing the locks in only necessary for phases which 1413 // do yields to improve the granularity of the collection. 1414 assert_lock_strong(bitMapLock()); 1415 // We need to lock the Free list lock for the space that we are 1416 // currently collecting. 1417 assert(haveFreelistLocks(), "Must be holding free list locks"); 1418 bitMapLock()->unlock(); 1419 releaseFreelistLocks(); 1420 { 1421 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1422 if (_foregroundGCShouldWait) { 1423 // We are going to be waiting for action for the CMS thread; 1424 // it had better not be gone (for instance at shutdown)! 1425 assert(ConcurrentMarkSweepThread::cmst() != NULL && !ConcurrentMarkSweepThread::cmst()->has_terminated(), 1426 "CMS thread must be running"); 1427 // Wait here until the background collector gives us the go-ahead 1428 ConcurrentMarkSweepThread::clear_CMS_flag( 1429 ConcurrentMarkSweepThread::CMS_vm_has_token); // release token 1430 // Get a possibly blocked CMS thread going: 1431 // Note that we set _foregroundGCIsActive true above, 1432 // without protection of the CGC_lock. 1433 CGC_lock->notify(); 1434 assert(!ConcurrentMarkSweepThread::vm_thread_wants_cms_token(), 1435 "Possible deadlock"); 1436 while (_foregroundGCShouldWait) { 1437 // wait for notification 1438 CGC_lock->wait(Mutex::_no_safepoint_check_flag); 1439 // Possibility of delay/starvation here, since CMS token does 1440 // not know to give priority to VM thread? Actually, i think 1441 // there wouldn't be any delay/starvation, but the proof of 1442 // that "fact" (?) appears non-trivial. XXX 20011219YSR 1443 } 1444 ConcurrentMarkSweepThread::set_CMS_flag( 1445 ConcurrentMarkSweepThread::CMS_vm_has_token); 1446 } 1447 } 1448 // The CMS_token is already held. Get back the other locks. 1449 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 1450 "VM thread should have CMS token"); 1451 getFreelistLocks(); 1452 bitMapLock()->lock_without_safepoint_check(); 1453 log_debug(gc, state)("CMS foreground collector has asked for control " INTPTR_FORMAT " with first state %d", 1454 p2i(Thread::current()), first_state); 1455 log_debug(gc, state)(" gets control with state %d", _collectorState); 1456 1457 // Inform cms gen if this was due to partial collection failing. 1458 // The CMS gen may use this fact to determine its expansion policy. 1459 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1460 if (gch->incremental_collection_will_fail(false /* don't consult_young */)) { 1461 assert(!_cmsGen->incremental_collection_failed(), 1462 "Should have been noticed, reacted to and cleared"); 1463 _cmsGen->set_incremental_collection_failed(); 1464 } 1465 1466 if (first_state > Idling) { 1467 report_concurrent_mode_interruption(); 1468 } 1469 1470 set_did_compact(true); 1471 1472 // If the collection is being acquired from the background 1473 // collector, there may be references on the discovered 1474 // references lists. Abandon those references, since some 1475 // of them may have become unreachable after concurrent 1476 // discovery; the STW compacting collector will redo discovery 1477 // more precisely, without being subject to floating garbage. 1478 // Leaving otherwise unreachable references in the discovered 1479 // lists would require special handling. 1480 ref_processor()->disable_discovery(); 1481 ref_processor()->abandon_partial_discovery(); 1482 ref_processor()->verify_no_references_recorded(); 1483 1484 if (first_state > Idling) { 1485 save_heap_summary(); 1486 } 1487 1488 do_compaction_work(clear_all_soft_refs); 1489 1490 // Has the GC time limit been exceeded? 1491 size_t max_eden_size = _young_gen->max_eden_size(); 1492 GCCause::Cause gc_cause = gch->gc_cause(); 1493 size_policy()->check_gc_overhead_limit(_young_gen->used(), 1494 _young_gen->eden()->used(), 1495 _cmsGen->max_capacity(), 1496 max_eden_size, 1497 full, 1498 gc_cause, 1499 gch->collector_policy()); 1500 1501 // Reset the expansion cause, now that we just completed 1502 // a collection cycle. 1503 clear_expansion_cause(); 1504 _foregroundGCIsActive = false; 1505 return; 1506 } 1507 1508 // Resize the tenured generation 1509 // after obtaining the free list locks for the 1510 // two generations. 1511 void CMSCollector::compute_new_size() { 1512 assert_locked_or_safepoint(Heap_lock); 1513 FreelistLocker z(this); 1514 MetaspaceGC::compute_new_size(); 1515 _cmsGen->compute_new_size_free_list(); 1516 } 1517 1518 // A work method used by the foreground collector to do 1519 // a mark-sweep-compact. 1520 void CMSCollector::do_compaction_work(bool clear_all_soft_refs) { 1521 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1522 1523 STWGCTimer* gc_timer = GenMarkSweep::gc_timer(); 1524 gc_timer->register_gc_start(); 1525 1526 SerialOldTracer* gc_tracer = GenMarkSweep::gc_tracer(); 1527 gc_tracer->report_gc_start(gch->gc_cause(), gc_timer->gc_start()); 1528 1529 gch->pre_full_gc_dump(gc_timer); 1530 1531 GCTraceTime(Trace, gc, phases) t("CMS:MSC"); 1532 1533 // Temporarily widen the span of the weak reference processing to 1534 // the entire heap. 1535 MemRegion new_span(GenCollectedHeap::heap()->reserved_region()); 1536 ReferenceProcessorSpanMutator rp_mut_span(ref_processor(), new_span); 1537 // Temporarily, clear the "is_alive_non_header" field of the 1538 // reference processor. 1539 ReferenceProcessorIsAliveMutator rp_mut_closure(ref_processor(), NULL); 1540 // Temporarily make reference _processing_ single threaded (non-MT). 1541 ReferenceProcessorMTProcMutator rp_mut_mt_processing(ref_processor(), false); 1542 // Temporarily make refs discovery atomic 1543 ReferenceProcessorAtomicMutator rp_mut_atomic(ref_processor(), true); 1544 // Temporarily make reference _discovery_ single threaded (non-MT) 1545 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false); 1546 1547 ref_processor()->set_enqueuing_is_done(false); 1548 ref_processor()->enable_discovery(); 1549 ref_processor()->setup_policy(clear_all_soft_refs); 1550 // If an asynchronous collection finishes, the _modUnionTable is 1551 // all clear. If we are assuming the collection from an asynchronous 1552 // collection, clear the _modUnionTable. 1553 assert(_collectorState != Idling || _modUnionTable.isAllClear(), 1554 "_modUnionTable should be clear if the baton was not passed"); 1555 _modUnionTable.clear_all(); 1556 assert(_collectorState != Idling || _ct->klass_rem_set()->mod_union_is_clear(), 1557 "mod union for klasses should be clear if the baton was passed"); 1558 _ct->klass_rem_set()->clear_mod_union(); 1559 1560 // We must adjust the allocation statistics being maintained 1561 // in the free list space. We do so by reading and clearing 1562 // the sweep timer and updating the block flux rate estimates below. 1563 assert(!_intra_sweep_timer.is_active(), "_intra_sweep_timer should be inactive"); 1564 if (_inter_sweep_timer.is_active()) { 1565 _inter_sweep_timer.stop(); 1566 // Note that we do not use this sample to update the _inter_sweep_estimate. 1567 _cmsGen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()), 1568 _inter_sweep_estimate.padded_average(), 1569 _intra_sweep_estimate.padded_average()); 1570 } 1571 1572 GenMarkSweep::invoke_at_safepoint(ref_processor(), clear_all_soft_refs); 1573 #ifdef ASSERT 1574 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 1575 size_t free_size = cms_space->free(); 1576 assert(free_size == 1577 pointer_delta(cms_space->end(), cms_space->compaction_top()) 1578 * HeapWordSize, 1579 "All the free space should be compacted into one chunk at top"); 1580 assert(cms_space->dictionary()->total_chunk_size( 1581 debug_only(cms_space->freelistLock())) == 0 || 1582 cms_space->totalSizeInIndexedFreeLists() == 0, 1583 "All the free space should be in a single chunk"); 1584 size_t num = cms_space->totalCount(); 1585 assert((free_size == 0 && num == 0) || 1586 (free_size > 0 && (num == 1 || num == 2)), 1587 "There should be at most 2 free chunks after compaction"); 1588 #endif // ASSERT 1589 _collectorState = Resetting; 1590 assert(_restart_addr == NULL, 1591 "Should have been NULL'd before baton was passed"); 1592 reset_stw(); 1593 _cmsGen->reset_after_compaction(); 1594 _concurrent_cycles_since_last_unload = 0; 1595 1596 // Clear any data recorded in the PLAB chunk arrays. 1597 if (_survivor_plab_array != NULL) { 1598 reset_survivor_plab_arrays(); 1599 } 1600 1601 // Adjust the per-size allocation stats for the next epoch. 1602 _cmsGen->cmsSpace()->endSweepFLCensus(sweep_count() /* fake */); 1603 // Restart the "inter sweep timer" for the next epoch. 1604 _inter_sweep_timer.reset(); 1605 _inter_sweep_timer.start(); 1606 1607 // No longer a need to do a concurrent collection for Metaspace. 1608 MetaspaceGC::set_should_concurrent_collect(false); 1609 1610 gch->post_full_gc_dump(gc_timer); 1611 1612 gc_timer->register_gc_end(); 1613 1614 gc_tracer->report_gc_end(gc_timer->gc_end(), gc_timer->time_partitions()); 1615 1616 // For a mark-sweep-compact, compute_new_size() will be called 1617 // in the heap's do_collection() method. 1618 } 1619 1620 void CMSCollector::print_eden_and_survivor_chunk_arrays() { 1621 Log(gc, heap) log; 1622 if (!log.is_trace()) { 1623 return; 1624 } 1625 1626 ContiguousSpace* eden_space = _young_gen->eden(); 1627 ContiguousSpace* from_space = _young_gen->from(); 1628 ContiguousSpace* to_space = _young_gen->to(); 1629 // Eden 1630 if (_eden_chunk_array != NULL) { 1631 log.trace("eden " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")", 1632 p2i(eden_space->bottom()), p2i(eden_space->top()), 1633 p2i(eden_space->end()), eden_space->capacity()); 1634 log.trace("_eden_chunk_index=" SIZE_FORMAT ", _eden_chunk_capacity=" SIZE_FORMAT, 1635 _eden_chunk_index, _eden_chunk_capacity); 1636 for (size_t i = 0; i < _eden_chunk_index; i++) { 1637 log.trace("_eden_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT, i, p2i(_eden_chunk_array[i])); 1638 } 1639 } 1640 // Survivor 1641 if (_survivor_chunk_array != NULL) { 1642 log.trace("survivor " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")", 1643 p2i(from_space->bottom()), p2i(from_space->top()), 1644 p2i(from_space->end()), from_space->capacity()); 1645 log.trace("_survivor_chunk_index=" SIZE_FORMAT ", _survivor_chunk_capacity=" SIZE_FORMAT, 1646 _survivor_chunk_index, _survivor_chunk_capacity); 1647 for (size_t i = 0; i < _survivor_chunk_index; i++) { 1648 log.trace("_survivor_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT, i, p2i(_survivor_chunk_array[i])); 1649 } 1650 } 1651 } 1652 1653 void CMSCollector::getFreelistLocks() const { 1654 // Get locks for all free lists in all generations that this 1655 // collector is responsible for 1656 _cmsGen->freelistLock()->lock_without_safepoint_check(); 1657 } 1658 1659 void CMSCollector::releaseFreelistLocks() const { 1660 // Release locks for all free lists in all generations that this 1661 // collector is responsible for 1662 _cmsGen->freelistLock()->unlock(); 1663 } 1664 1665 bool CMSCollector::haveFreelistLocks() const { 1666 // Check locks for all free lists in all generations that this 1667 // collector is responsible for 1668 assert_lock_strong(_cmsGen->freelistLock()); 1669 PRODUCT_ONLY(ShouldNotReachHere()); 1670 return true; 1671 } 1672 1673 // A utility class that is used by the CMS collector to 1674 // temporarily "release" the foreground collector from its 1675 // usual obligation to wait for the background collector to 1676 // complete an ongoing phase before proceeding. 1677 class ReleaseForegroundGC: public StackObj { 1678 private: 1679 CMSCollector* _c; 1680 public: 1681 ReleaseForegroundGC(CMSCollector* c) : _c(c) { 1682 assert(_c->_foregroundGCShouldWait, "Else should not need to call"); 1683 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1684 // allow a potentially blocked foreground collector to proceed 1685 _c->_foregroundGCShouldWait = false; 1686 if (_c->_foregroundGCIsActive) { 1687 CGC_lock->notify(); 1688 } 1689 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 1690 "Possible deadlock"); 1691 } 1692 1693 ~ReleaseForegroundGC() { 1694 assert(!_c->_foregroundGCShouldWait, "Usage protocol violation?"); 1695 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1696 _c->_foregroundGCShouldWait = true; 1697 } 1698 }; 1699 1700 void CMSCollector::collect_in_background(GCCause::Cause cause) { 1701 assert(Thread::current()->is_ConcurrentGC_thread(), 1702 "A CMS asynchronous collection is only allowed on a CMS thread."); 1703 1704 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1705 { 1706 bool safepoint_check = Mutex::_no_safepoint_check_flag; 1707 MutexLockerEx hl(Heap_lock, safepoint_check); 1708 FreelistLocker fll(this); 1709 MutexLockerEx x(CGC_lock, safepoint_check); 1710 if (_foregroundGCIsActive) { 1711 // The foreground collector is. Skip this 1712 // background collection. 1713 assert(!_foregroundGCShouldWait, "Should be clear"); 1714 return; 1715 } else { 1716 assert(_collectorState == Idling, "Should be idling before start."); 1717 _collectorState = InitialMarking; 1718 register_gc_start(cause); 1719 // Reset the expansion cause, now that we are about to begin 1720 // a new cycle. 1721 clear_expansion_cause(); 1722 1723 // Clear the MetaspaceGC flag since a concurrent collection 1724 // is starting but also clear it after the collection. 1725 MetaspaceGC::set_should_concurrent_collect(false); 1726 } 1727 // Decide if we want to enable class unloading as part of the 1728 // ensuing concurrent GC cycle. 1729 update_should_unload_classes(); 1730 _full_gc_requested = false; // acks all outstanding full gc requests 1731 _full_gc_cause = GCCause::_no_gc; 1732 // Signal that we are about to start a collection 1733 gch->increment_total_full_collections(); // ... starting a collection cycle 1734 _collection_count_start = gch->total_full_collections(); 1735 } 1736 1737 size_t prev_used = _cmsGen->used(); 1738 1739 // The change of the collection state is normally done at this level; 1740 // the exceptions are phases that are executed while the world is 1741 // stopped. For those phases the change of state is done while the 1742 // world is stopped. For baton passing purposes this allows the 1743 // background collector to finish the phase and change state atomically. 1744 // The foreground collector cannot wait on a phase that is done 1745 // while the world is stopped because the foreground collector already 1746 // has the world stopped and would deadlock. 1747 while (_collectorState != Idling) { 1748 log_debug(gc, state)("Thread " INTPTR_FORMAT " in CMS state %d", 1749 p2i(Thread::current()), _collectorState); 1750 // The foreground collector 1751 // holds the Heap_lock throughout its collection. 1752 // holds the CMS token (but not the lock) 1753 // except while it is waiting for the background collector to yield. 1754 // 1755 // The foreground collector should be blocked (not for long) 1756 // if the background collector is about to start a phase 1757 // executed with world stopped. If the background 1758 // collector has already started such a phase, the 1759 // foreground collector is blocked waiting for the 1760 // Heap_lock. The stop-world phases (InitialMarking and FinalMarking) 1761 // are executed in the VM thread. 1762 // 1763 // The locking order is 1764 // PendingListLock (PLL) -- if applicable (FinalMarking) 1765 // Heap_lock (both this & PLL locked in VM_CMS_Operation::prologue()) 1766 // CMS token (claimed in 1767 // stop_world_and_do() --> 1768 // safepoint_synchronize() --> 1769 // CMSThread::synchronize()) 1770 1771 { 1772 // Check if the FG collector wants us to yield. 1773 CMSTokenSync x(true); // is cms thread 1774 if (waitForForegroundGC()) { 1775 // We yielded to a foreground GC, nothing more to be 1776 // done this round. 1777 assert(_foregroundGCShouldWait == false, "We set it to false in " 1778 "waitForForegroundGC()"); 1779 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " exiting collection CMS state %d", 1780 p2i(Thread::current()), _collectorState); 1781 return; 1782 } else { 1783 // The background collector can run but check to see if the 1784 // foreground collector has done a collection while the 1785 // background collector was waiting to get the CGC_lock 1786 // above. If yes, break so that _foregroundGCShouldWait 1787 // is cleared before returning. 1788 if (_collectorState == Idling) { 1789 break; 1790 } 1791 } 1792 } 1793 1794 assert(_foregroundGCShouldWait, "Foreground collector, if active, " 1795 "should be waiting"); 1796 1797 switch (_collectorState) { 1798 case InitialMarking: 1799 { 1800 ReleaseForegroundGC x(this); 1801 stats().record_cms_begin(); 1802 VM_CMS_Initial_Mark initial_mark_op(this); 1803 VMThread::execute(&initial_mark_op); 1804 } 1805 // The collector state may be any legal state at this point 1806 // since the background collector may have yielded to the 1807 // foreground collector. 1808 break; 1809 case Marking: 1810 // initial marking in checkpointRootsInitialWork has been completed 1811 if (markFromRoots()) { // we were successful 1812 assert(_collectorState == Precleaning, "Collector state should " 1813 "have changed"); 1814 } else { 1815 assert(_foregroundGCIsActive, "Internal state inconsistency"); 1816 } 1817 break; 1818 case Precleaning: 1819 // marking from roots in markFromRoots has been completed 1820 preclean(); 1821 assert(_collectorState == AbortablePreclean || 1822 _collectorState == FinalMarking, 1823 "Collector state should have changed"); 1824 break; 1825 case AbortablePreclean: 1826 abortable_preclean(); 1827 assert(_collectorState == FinalMarking, "Collector state should " 1828 "have changed"); 1829 break; 1830 case FinalMarking: 1831 { 1832 ReleaseForegroundGC x(this); 1833 1834 VM_CMS_Final_Remark final_remark_op(this); 1835 VMThread::execute(&final_remark_op); 1836 } 1837 assert(_foregroundGCShouldWait, "block post-condition"); 1838 break; 1839 case Sweeping: 1840 // final marking in checkpointRootsFinal has been completed 1841 sweep(); 1842 assert(_collectorState == Resizing, "Collector state change " 1843 "to Resizing must be done under the free_list_lock"); 1844 1845 case Resizing: { 1846 // Sweeping has been completed... 1847 // At this point the background collection has completed. 1848 // Don't move the call to compute_new_size() down 1849 // into code that might be executed if the background 1850 // collection was preempted. 1851 { 1852 ReleaseForegroundGC x(this); // unblock FG collection 1853 MutexLockerEx y(Heap_lock, Mutex::_no_safepoint_check_flag); 1854 CMSTokenSync z(true); // not strictly needed. 1855 if (_collectorState == Resizing) { 1856 compute_new_size(); 1857 save_heap_summary(); 1858 _collectorState = Resetting; 1859 } else { 1860 assert(_collectorState == Idling, "The state should only change" 1861 " because the foreground collector has finished the collection"); 1862 } 1863 } 1864 break; 1865 } 1866 case Resetting: 1867 // CMS heap resizing has been completed 1868 reset_concurrent(); 1869 assert(_collectorState == Idling, "Collector state should " 1870 "have changed"); 1871 1872 MetaspaceGC::set_should_concurrent_collect(false); 1873 1874 stats().record_cms_end(); 1875 // Don't move the concurrent_phases_end() and compute_new_size() 1876 // calls to here because a preempted background collection 1877 // has it's state set to "Resetting". 1878 break; 1879 case Idling: 1880 default: 1881 ShouldNotReachHere(); 1882 break; 1883 } 1884 log_debug(gc, state)(" Thread " INTPTR_FORMAT " done - next CMS state %d", 1885 p2i(Thread::current()), _collectorState); 1886 assert(_foregroundGCShouldWait, "block post-condition"); 1887 } 1888 1889 // Should this be in gc_epilogue? 1890 collector_policy()->counters()->update_counters(); 1891 1892 { 1893 // Clear _foregroundGCShouldWait and, in the event that the 1894 // foreground collector is waiting, notify it, before 1895 // returning. 1896 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1897 _foregroundGCShouldWait = false; 1898 if (_foregroundGCIsActive) { 1899 CGC_lock->notify(); 1900 } 1901 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 1902 "Possible deadlock"); 1903 } 1904 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " exiting collection CMS state %d", 1905 p2i(Thread::current()), _collectorState); 1906 log_info(gc, heap)("Old: " SIZE_FORMAT "K->" SIZE_FORMAT "K(" SIZE_FORMAT "K)", 1907 prev_used / K, _cmsGen->used()/K, _cmsGen->capacity() /K); 1908 } 1909 1910 void CMSCollector::register_gc_start(GCCause::Cause cause) { 1911 _cms_start_registered = true; 1912 _gc_timer_cm->register_gc_start(); 1913 _gc_tracer_cm->report_gc_start(cause, _gc_timer_cm->gc_start()); 1914 } 1915 1916 void CMSCollector::register_gc_end() { 1917 if (_cms_start_registered) { 1918 report_heap_summary(GCWhen::AfterGC); 1919 1920 _gc_timer_cm->register_gc_end(); 1921 _gc_tracer_cm->report_gc_end(_gc_timer_cm->gc_end(), _gc_timer_cm->time_partitions()); 1922 _cms_start_registered = false; 1923 } 1924 } 1925 1926 void CMSCollector::save_heap_summary() { 1927 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1928 _last_heap_summary = gch->create_heap_summary(); 1929 _last_metaspace_summary = gch->create_metaspace_summary(); 1930 } 1931 1932 void CMSCollector::report_heap_summary(GCWhen::Type when) { 1933 _gc_tracer_cm->report_gc_heap_summary(when, _last_heap_summary); 1934 _gc_tracer_cm->report_metaspace_summary(when, _last_metaspace_summary); 1935 } 1936 1937 bool CMSCollector::waitForForegroundGC() { 1938 bool res = false; 1939 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 1940 "CMS thread should have CMS token"); 1941 // Block the foreground collector until the 1942 // background collectors decides whether to 1943 // yield. 1944 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1945 _foregroundGCShouldWait = true; 1946 if (_foregroundGCIsActive) { 1947 // The background collector yields to the 1948 // foreground collector and returns a value 1949 // indicating that it has yielded. The foreground 1950 // collector can proceed. 1951 res = true; 1952 _foregroundGCShouldWait = false; 1953 ConcurrentMarkSweepThread::clear_CMS_flag( 1954 ConcurrentMarkSweepThread::CMS_cms_has_token); 1955 ConcurrentMarkSweepThread::set_CMS_flag( 1956 ConcurrentMarkSweepThread::CMS_cms_wants_token); 1957 // Get a possibly blocked foreground thread going 1958 CGC_lock->notify(); 1959 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " waiting at CMS state %d", 1960 p2i(Thread::current()), _collectorState); 1961 while (_foregroundGCIsActive) { 1962 CGC_lock->wait(Mutex::_no_safepoint_check_flag); 1963 } 1964 ConcurrentMarkSweepThread::set_CMS_flag( 1965 ConcurrentMarkSweepThread::CMS_cms_has_token); 1966 ConcurrentMarkSweepThread::clear_CMS_flag( 1967 ConcurrentMarkSweepThread::CMS_cms_wants_token); 1968 } 1969 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " continuing at CMS state %d", 1970 p2i(Thread::current()), _collectorState); 1971 return res; 1972 } 1973 1974 // Because of the need to lock the free lists and other structures in 1975 // the collector, common to all the generations that the collector is 1976 // collecting, we need the gc_prologues of individual CMS generations 1977 // delegate to their collector. It may have been simpler had the 1978 // current infrastructure allowed one to call a prologue on a 1979 // collector. In the absence of that we have the generation's 1980 // prologue delegate to the collector, which delegates back 1981 // some "local" work to a worker method in the individual generations 1982 // that it's responsible for collecting, while itself doing any 1983 // work common to all generations it's responsible for. A similar 1984 // comment applies to the gc_epilogue()'s. 1985 // The role of the variable _between_prologue_and_epilogue is to 1986 // enforce the invocation protocol. 1987 void CMSCollector::gc_prologue(bool full) { 1988 // Call gc_prologue_work() for the CMSGen 1989 // we are responsible for. 1990 1991 // The following locking discipline assumes that we are only called 1992 // when the world is stopped. 1993 assert(SafepointSynchronize::is_at_safepoint(), "world is stopped assumption"); 1994 1995 // The CMSCollector prologue must call the gc_prologues for the 1996 // "generations" that it's responsible 1997 // for. 1998 1999 assert( Thread::current()->is_VM_thread() 2000 || ( CMSScavengeBeforeRemark 2001 && Thread::current()->is_ConcurrentGC_thread()), 2002 "Incorrect thread type for prologue execution"); 2003 2004 if (_between_prologue_and_epilogue) { 2005 // We have already been invoked; this is a gc_prologue delegation 2006 // from yet another CMS generation that we are responsible for, just 2007 // ignore it since all relevant work has already been done. 2008 return; 2009 } 2010 2011 // set a bit saying prologue has been called; cleared in epilogue 2012 _between_prologue_and_epilogue = true; 2013 // Claim locks for common data structures, then call gc_prologue_work() 2014 // for each CMSGen. 2015 2016 getFreelistLocks(); // gets free list locks on constituent spaces 2017 bitMapLock()->lock_without_safepoint_check(); 2018 2019 // Should call gc_prologue_work() for all cms gens we are responsible for 2020 bool duringMarking = _collectorState >= Marking 2021 && _collectorState < Sweeping; 2022 2023 // The young collections clear the modified oops state, which tells if 2024 // there are any modified oops in the class. The remark phase also needs 2025 // that information. Tell the young collection to save the union of all 2026 // modified klasses. 2027 if (duringMarking) { 2028 _ct->klass_rem_set()->set_accumulate_modified_oops(true); 2029 } 2030 2031 bool registerClosure = duringMarking; 2032 2033 _cmsGen->gc_prologue_work(full, registerClosure, &_modUnionClosurePar); 2034 2035 if (!full) { 2036 stats().record_gc0_begin(); 2037 } 2038 } 2039 2040 void ConcurrentMarkSweepGeneration::gc_prologue(bool full) { 2041 2042 _capacity_at_prologue = capacity(); 2043 _used_at_prologue = used(); 2044 2045 // We enable promotion tracking so that card-scanning can recognize 2046 // which objects have been promoted during this GC and skip them. 2047 for (uint i = 0; i < ParallelGCThreads; i++) { 2048 _par_gc_thread_states[i]->promo.startTrackingPromotions(); 2049 } 2050 2051 // Delegate to CMScollector which knows how to coordinate between 2052 // this and any other CMS generations that it is responsible for 2053 // collecting. 2054 collector()->gc_prologue(full); 2055 } 2056 2057 // This is a "private" interface for use by this generation's CMSCollector. 2058 // Not to be called directly by any other entity (for instance, 2059 // GenCollectedHeap, which calls the "public" gc_prologue method above). 2060 void ConcurrentMarkSweepGeneration::gc_prologue_work(bool full, 2061 bool registerClosure, ModUnionClosure* modUnionClosure) { 2062 assert(!incremental_collection_failed(), "Shouldn't be set yet"); 2063 assert(cmsSpace()->preconsumptionDirtyCardClosure() == NULL, 2064 "Should be NULL"); 2065 if (registerClosure) { 2066 cmsSpace()->setPreconsumptionDirtyCardClosure(modUnionClosure); 2067 } 2068 cmsSpace()->gc_prologue(); 2069 // Clear stat counters 2070 NOT_PRODUCT( 2071 assert(_numObjectsPromoted == 0, "check"); 2072 assert(_numWordsPromoted == 0, "check"); 2073 log_develop_trace(gc, alloc)("Allocated " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes concurrently", 2074 _numObjectsAllocated, _numWordsAllocated*sizeof(HeapWord)); 2075 _numObjectsAllocated = 0; 2076 _numWordsAllocated = 0; 2077 ) 2078 } 2079 2080 void CMSCollector::gc_epilogue(bool full) { 2081 // The following locking discipline assumes that we are only called 2082 // when the world is stopped. 2083 assert(SafepointSynchronize::is_at_safepoint(), 2084 "world is stopped assumption"); 2085 2086 // Currently the CMS epilogue (see CompactibleFreeListSpace) merely checks 2087 // if linear allocation blocks need to be appropriately marked to allow the 2088 // the blocks to be parsable. We also check here whether we need to nudge the 2089 // CMS collector thread to start a new cycle (if it's not already active). 2090 assert( Thread::current()->is_VM_thread() 2091 || ( CMSScavengeBeforeRemark 2092 && Thread::current()->is_ConcurrentGC_thread()), 2093 "Incorrect thread type for epilogue execution"); 2094 2095 if (!_between_prologue_and_epilogue) { 2096 // We have already been invoked; this is a gc_epilogue delegation 2097 // from yet another CMS generation that we are responsible for, just 2098 // ignore it since all relevant work has already been done. 2099 return; 2100 } 2101 assert(haveFreelistLocks(), "must have freelist locks"); 2102 assert_lock_strong(bitMapLock()); 2103 2104 _ct->klass_rem_set()->set_accumulate_modified_oops(false); 2105 2106 _cmsGen->gc_epilogue_work(full); 2107 2108 if (_collectorState == AbortablePreclean || _collectorState == Precleaning) { 2109 // in case sampling was not already enabled, enable it 2110 _start_sampling = true; 2111 } 2112 // reset _eden_chunk_array so sampling starts afresh 2113 _eden_chunk_index = 0; 2114 2115 size_t cms_used = _cmsGen->cmsSpace()->used(); 2116 2117 // update performance counters - this uses a special version of 2118 // update_counters() that allows the utilization to be passed as a 2119 // parameter, avoiding multiple calls to used(). 2120 // 2121 _cmsGen->update_counters(cms_used); 2122 2123 bitMapLock()->unlock(); 2124 releaseFreelistLocks(); 2125 2126 if (!CleanChunkPoolAsync) { 2127 Chunk::clean_chunk_pool(); 2128 } 2129 2130 set_did_compact(false); 2131 _between_prologue_and_epilogue = false; // ready for next cycle 2132 } 2133 2134 void ConcurrentMarkSweepGeneration::gc_epilogue(bool full) { 2135 collector()->gc_epilogue(full); 2136 2137 // When using ParNew, promotion tracking should have already been 2138 // disabled. However, the prologue (which enables promotion 2139 // tracking) and epilogue are called irrespective of the type of 2140 // GC. So they will also be called before and after Full GCs, during 2141 // which promotion tracking will not be explicitly disabled. So, 2142 // it's safer to also disable it here too (to be symmetric with 2143 // enabling it in the prologue). 2144 for (uint i = 0; i < ParallelGCThreads; i++) { 2145 _par_gc_thread_states[i]->promo.stopTrackingPromotions(); 2146 } 2147 } 2148 2149 void ConcurrentMarkSweepGeneration::gc_epilogue_work(bool full) { 2150 assert(!incremental_collection_failed(), "Should have been cleared"); 2151 cmsSpace()->setPreconsumptionDirtyCardClosure(NULL); 2152 cmsSpace()->gc_epilogue(); 2153 // Print stat counters 2154 NOT_PRODUCT( 2155 assert(_numObjectsAllocated == 0, "check"); 2156 assert(_numWordsAllocated == 0, "check"); 2157 log_develop_trace(gc, promotion)("Promoted " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes", 2158 _numObjectsPromoted, _numWordsPromoted*sizeof(HeapWord)); 2159 _numObjectsPromoted = 0; 2160 _numWordsPromoted = 0; 2161 ) 2162 2163 // Call down the chain in contiguous_available needs the freelistLock 2164 // so print this out before releasing the freeListLock. 2165 log_develop_trace(gc)(" Contiguous available " SIZE_FORMAT " bytes ", contiguous_available()); 2166 } 2167 2168 #ifndef PRODUCT 2169 bool CMSCollector::have_cms_token() { 2170 Thread* thr = Thread::current(); 2171 if (thr->is_VM_thread()) { 2172 return ConcurrentMarkSweepThread::vm_thread_has_cms_token(); 2173 } else if (thr->is_ConcurrentGC_thread()) { 2174 return ConcurrentMarkSweepThread::cms_thread_has_cms_token(); 2175 } else if (thr->is_GC_task_thread()) { 2176 return ConcurrentMarkSweepThread::vm_thread_has_cms_token() && 2177 ParGCRareEvent_lock->owned_by_self(); 2178 } 2179 return false; 2180 } 2181 2182 // Check reachability of the given heap address in CMS generation, 2183 // treating all other generations as roots. 2184 bool CMSCollector::is_cms_reachable(HeapWord* addr) { 2185 // We could "guarantee" below, rather than assert, but I'll 2186 // leave these as "asserts" so that an adventurous debugger 2187 // could try this in the product build provided some subset of 2188 // the conditions were met, provided they were interested in the 2189 // results and knew that the computation below wouldn't interfere 2190 // with other concurrent computations mutating the structures 2191 // being read or written. 2192 assert(SafepointSynchronize::is_at_safepoint(), 2193 "Else mutations in object graph will make answer suspect"); 2194 assert(have_cms_token(), "Should hold cms token"); 2195 assert(haveFreelistLocks(), "must hold free list locks"); 2196 assert_lock_strong(bitMapLock()); 2197 2198 // Clear the marking bit map array before starting, but, just 2199 // for kicks, first report if the given address is already marked 2200 tty->print_cr("Start: Address " PTR_FORMAT " is%s marked", p2i(addr), 2201 _markBitMap.isMarked(addr) ? "" : " not"); 2202 2203 if (verify_after_remark()) { 2204 MutexLockerEx x(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag); 2205 bool result = verification_mark_bm()->isMarked(addr); 2206 tty->print_cr("TransitiveMark: Address " PTR_FORMAT " %s marked", p2i(addr), 2207 result ? "IS" : "is NOT"); 2208 return result; 2209 } else { 2210 tty->print_cr("Could not compute result"); 2211 return false; 2212 } 2213 } 2214 #endif 2215 2216 void 2217 CMSCollector::print_on_error(outputStream* st) { 2218 CMSCollector* collector = ConcurrentMarkSweepGeneration::_collector; 2219 if (collector != NULL) { 2220 CMSBitMap* bitmap = &collector->_markBitMap; 2221 st->print_cr("Marking Bits: (CMSBitMap*) " PTR_FORMAT, p2i(bitmap)); 2222 bitmap->print_on_error(st, " Bits: "); 2223 2224 st->cr(); 2225 2226 CMSBitMap* mut_bitmap = &collector->_modUnionTable; 2227 st->print_cr("Mod Union Table: (CMSBitMap*) " PTR_FORMAT, p2i(mut_bitmap)); 2228 mut_bitmap->print_on_error(st, " Bits: "); 2229 } 2230 } 2231 2232 //////////////////////////////////////////////////////// 2233 // CMS Verification Support 2234 //////////////////////////////////////////////////////// 2235 // Following the remark phase, the following invariant 2236 // should hold -- each object in the CMS heap which is 2237 // marked in markBitMap() should be marked in the verification_mark_bm(). 2238 2239 class VerifyMarkedClosure: public BitMapClosure { 2240 CMSBitMap* _marks; 2241 bool _failed; 2242 2243 public: 2244 VerifyMarkedClosure(CMSBitMap* bm): _marks(bm), _failed(false) {} 2245 2246 bool do_bit(size_t offset) { 2247 HeapWord* addr = _marks->offsetToHeapWord(offset); 2248 if (!_marks->isMarked(addr)) { 2249 Log(gc, verify) log; 2250 ResourceMark rm; 2251 LogStream ls(log.error()); 2252 oop(addr)->print_on(&ls); 2253 log.error(" (" INTPTR_FORMAT " should have been marked)", p2i(addr)); 2254 _failed = true; 2255 } 2256 return true; 2257 } 2258 2259 bool failed() { return _failed; } 2260 }; 2261 2262 bool CMSCollector::verify_after_remark() { 2263 GCTraceTime(Info, gc, phases, verify) tm("Verifying CMS Marking."); 2264 MutexLockerEx ml(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag); 2265 static bool init = false; 2266 2267 assert(SafepointSynchronize::is_at_safepoint(), 2268 "Else mutations in object graph will make answer suspect"); 2269 assert(have_cms_token(), 2270 "Else there may be mutual interference in use of " 2271 " verification data structures"); 2272 assert(_collectorState > Marking && _collectorState <= Sweeping, 2273 "Else marking info checked here may be obsolete"); 2274 assert(haveFreelistLocks(), "must hold free list locks"); 2275 assert_lock_strong(bitMapLock()); 2276 2277 2278 // Allocate marking bit map if not already allocated 2279 if (!init) { // first time 2280 if (!verification_mark_bm()->allocate(_span)) { 2281 return false; 2282 } 2283 init = true; 2284 } 2285 2286 assert(verification_mark_stack()->isEmpty(), "Should be empty"); 2287 2288 // Turn off refs discovery -- so we will be tracing through refs. 2289 // This is as intended, because by this time 2290 // GC must already have cleared any refs that need to be cleared, 2291 // and traced those that need to be marked; moreover, 2292 // the marking done here is not going to interfere in any 2293 // way with the marking information used by GC. 2294 NoRefDiscovery no_discovery(ref_processor()); 2295 2296 #if defined(COMPILER2) || INCLUDE_JVMCI 2297 DerivedPointerTableDeactivate dpt_deact; 2298 #endif 2299 2300 // Clear any marks from a previous round 2301 verification_mark_bm()->clear_all(); 2302 assert(verification_mark_stack()->isEmpty(), "markStack should be empty"); 2303 verify_work_stacks_empty(); 2304 2305 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2306 gch->ensure_parsability(false); // fill TLABs, but no need to retire them 2307 // Update the saved marks which may affect the root scans. 2308 gch->save_marks(); 2309 2310 if (CMSRemarkVerifyVariant == 1) { 2311 // In this first variant of verification, we complete 2312 // all marking, then check if the new marks-vector is 2313 // a subset of the CMS marks-vector. 2314 verify_after_remark_work_1(); 2315 } else { 2316 guarantee(CMSRemarkVerifyVariant == 2, "Range checking for CMSRemarkVerifyVariant should guarantee 1 or 2"); 2317 // In this second variant of verification, we flag an error 2318 // (i.e. an object reachable in the new marks-vector not reachable 2319 // in the CMS marks-vector) immediately, also indicating the 2320 // identify of an object (A) that references the unmarked object (B) -- 2321 // presumably, a mutation to A failed to be picked up by preclean/remark? 2322 verify_after_remark_work_2(); 2323 } 2324 2325 return true; 2326 } 2327 2328 void CMSCollector::verify_after_remark_work_1() { 2329 ResourceMark rm; 2330 HandleMark hm; 2331 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2332 2333 // Get a clear set of claim bits for the roots processing to work with. 2334 ClassLoaderDataGraph::clear_claimed_marks(); 2335 2336 // Mark from roots one level into CMS 2337 MarkRefsIntoClosure notOlder(_span, verification_mark_bm()); 2338 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2339 2340 { 2341 StrongRootsScope srs(1); 2342 2343 gch->cms_process_roots(&srs, 2344 true, // young gen as roots 2345 GenCollectedHeap::ScanningOption(roots_scanning_options()), 2346 should_unload_classes(), 2347 ¬Older, 2348 NULL); 2349 } 2350 2351 // Now mark from the roots 2352 MarkFromRootsClosure markFromRootsClosure(this, _span, 2353 verification_mark_bm(), verification_mark_stack(), 2354 false /* don't yield */, true /* verifying */); 2355 assert(_restart_addr == NULL, "Expected pre-condition"); 2356 verification_mark_bm()->iterate(&markFromRootsClosure); 2357 while (_restart_addr != NULL) { 2358 // Deal with stack overflow: by restarting at the indicated 2359 // address. 2360 HeapWord* ra = _restart_addr; 2361 markFromRootsClosure.reset(ra); 2362 _restart_addr = NULL; 2363 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end()); 2364 } 2365 assert(verification_mark_stack()->isEmpty(), "Should have been drained"); 2366 verify_work_stacks_empty(); 2367 2368 // Marking completed -- now verify that each bit marked in 2369 // verification_mark_bm() is also marked in markBitMap(); flag all 2370 // errors by printing corresponding objects. 2371 VerifyMarkedClosure vcl(markBitMap()); 2372 verification_mark_bm()->iterate(&vcl); 2373 if (vcl.failed()) { 2374 Log(gc, verify) log; 2375 log.error("Failed marking verification after remark"); 2376 ResourceMark rm; 2377 LogStream ls(log.error()); 2378 gch->print_on(&ls); 2379 fatal("CMS: failed marking verification after remark"); 2380 } 2381 } 2382 2383 class VerifyKlassOopsKlassClosure : public KlassClosure { 2384 class VerifyKlassOopsClosure : public OopClosure { 2385 CMSBitMap* _bitmap; 2386 public: 2387 VerifyKlassOopsClosure(CMSBitMap* bitmap) : _bitmap(bitmap) { } 2388 void do_oop(oop* p) { guarantee(*p == NULL || _bitmap->isMarked((HeapWord*) *p), "Should be marked"); } 2389 void do_oop(narrowOop* p) { ShouldNotReachHere(); } 2390 } _oop_closure; 2391 public: 2392 VerifyKlassOopsKlassClosure(CMSBitMap* bitmap) : _oop_closure(bitmap) {} 2393 void do_klass(Klass* k) { 2394 k->oops_do(&_oop_closure); 2395 } 2396 }; 2397 2398 void CMSCollector::verify_after_remark_work_2() { 2399 ResourceMark rm; 2400 HandleMark hm; 2401 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2402 2403 // Get a clear set of claim bits for the roots processing to work with. 2404 ClassLoaderDataGraph::clear_claimed_marks(); 2405 2406 // Mark from roots one level into CMS 2407 MarkRefsIntoVerifyClosure notOlder(_span, verification_mark_bm(), 2408 markBitMap()); 2409 CLDToOopClosure cld_closure(¬Older, true); 2410 2411 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2412 2413 { 2414 StrongRootsScope srs(1); 2415 2416 gch->cms_process_roots(&srs, 2417 true, // young gen as roots 2418 GenCollectedHeap::ScanningOption(roots_scanning_options()), 2419 should_unload_classes(), 2420 ¬Older, 2421 &cld_closure); 2422 } 2423 2424 // Now mark from the roots 2425 MarkFromRootsVerifyClosure markFromRootsClosure(this, _span, 2426 verification_mark_bm(), markBitMap(), verification_mark_stack()); 2427 assert(_restart_addr == NULL, "Expected pre-condition"); 2428 verification_mark_bm()->iterate(&markFromRootsClosure); 2429 while (_restart_addr != NULL) { 2430 // Deal with stack overflow: by restarting at the indicated 2431 // address. 2432 HeapWord* ra = _restart_addr; 2433 markFromRootsClosure.reset(ra); 2434 _restart_addr = NULL; 2435 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end()); 2436 } 2437 assert(verification_mark_stack()->isEmpty(), "Should have been drained"); 2438 verify_work_stacks_empty(); 2439 2440 VerifyKlassOopsKlassClosure verify_klass_oops(verification_mark_bm()); 2441 ClassLoaderDataGraph::classes_do(&verify_klass_oops); 2442 2443 // Marking completed -- now verify that each bit marked in 2444 // verification_mark_bm() is also marked in markBitMap(); flag all 2445 // errors by printing corresponding objects. 2446 VerifyMarkedClosure vcl(markBitMap()); 2447 verification_mark_bm()->iterate(&vcl); 2448 assert(!vcl.failed(), "Else verification above should not have succeeded"); 2449 } 2450 2451 void ConcurrentMarkSweepGeneration::save_marks() { 2452 // delegate to CMS space 2453 cmsSpace()->save_marks(); 2454 } 2455 2456 bool ConcurrentMarkSweepGeneration::no_allocs_since_save_marks() { 2457 return cmsSpace()->no_allocs_since_save_marks(); 2458 } 2459 2460 #define CMS_SINCE_SAVE_MARKS_DEFN(OopClosureType, nv_suffix) \ 2461 \ 2462 void ConcurrentMarkSweepGeneration:: \ 2463 oop_since_save_marks_iterate##nv_suffix(OopClosureType* cl) { \ 2464 cl->set_generation(this); \ 2465 cmsSpace()->oop_since_save_marks_iterate##nv_suffix(cl); \ 2466 cl->reset_generation(); \ 2467 save_marks(); \ 2468 } 2469 2470 ALL_SINCE_SAVE_MARKS_CLOSURES(CMS_SINCE_SAVE_MARKS_DEFN) 2471 2472 void 2473 ConcurrentMarkSweepGeneration::oop_iterate(ExtendedOopClosure* cl) { 2474 if (freelistLock()->owned_by_self()) { 2475 Generation::oop_iterate(cl); 2476 } else { 2477 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2478 Generation::oop_iterate(cl); 2479 } 2480 } 2481 2482 void 2483 ConcurrentMarkSweepGeneration::object_iterate(ObjectClosure* cl) { 2484 if (freelistLock()->owned_by_self()) { 2485 Generation::object_iterate(cl); 2486 } else { 2487 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2488 Generation::object_iterate(cl); 2489 } 2490 } 2491 2492 void 2493 ConcurrentMarkSweepGeneration::safe_object_iterate(ObjectClosure* cl) { 2494 if (freelistLock()->owned_by_self()) { 2495 Generation::safe_object_iterate(cl); 2496 } else { 2497 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2498 Generation::safe_object_iterate(cl); 2499 } 2500 } 2501 2502 void 2503 ConcurrentMarkSweepGeneration::post_compact() { 2504 } 2505 2506 void 2507 ConcurrentMarkSweepGeneration::prepare_for_verify() { 2508 // Fix the linear allocation blocks to look like free blocks. 2509 2510 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those 2511 // are not called when the heap is verified during universe initialization and 2512 // at vm shutdown. 2513 if (freelistLock()->owned_by_self()) { 2514 cmsSpace()->prepare_for_verify(); 2515 } else { 2516 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag); 2517 cmsSpace()->prepare_for_verify(); 2518 } 2519 } 2520 2521 void 2522 ConcurrentMarkSweepGeneration::verify() { 2523 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those 2524 // are not called when the heap is verified during universe initialization and 2525 // at vm shutdown. 2526 if (freelistLock()->owned_by_self()) { 2527 cmsSpace()->verify(); 2528 } else { 2529 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag); 2530 cmsSpace()->verify(); 2531 } 2532 } 2533 2534 void CMSCollector::verify() { 2535 _cmsGen->verify(); 2536 } 2537 2538 #ifndef PRODUCT 2539 bool CMSCollector::overflow_list_is_empty() const { 2540 assert(_num_par_pushes >= 0, "Inconsistency"); 2541 if (_overflow_list == NULL) { 2542 assert(_num_par_pushes == 0, "Inconsistency"); 2543 } 2544 return _overflow_list == NULL; 2545 } 2546 2547 // The methods verify_work_stacks_empty() and verify_overflow_empty() 2548 // merely consolidate assertion checks that appear to occur together frequently. 2549 void CMSCollector::verify_work_stacks_empty() const { 2550 assert(_markStack.isEmpty(), "Marking stack should be empty"); 2551 assert(overflow_list_is_empty(), "Overflow list should be empty"); 2552 } 2553 2554 void CMSCollector::verify_overflow_empty() const { 2555 assert(overflow_list_is_empty(), "Overflow list should be empty"); 2556 assert(no_preserved_marks(), "No preserved marks"); 2557 } 2558 #endif // PRODUCT 2559 2560 // Decide if we want to enable class unloading as part of the 2561 // ensuing concurrent GC cycle. We will collect and 2562 // unload classes if it's the case that: 2563 // (a) class unloading is enabled at the command line, and 2564 // (b) old gen is getting really full 2565 // NOTE: Provided there is no change in the state of the heap between 2566 // calls to this method, it should have idempotent results. Moreover, 2567 // its results should be monotonically increasing (i.e. going from 0 to 1, 2568 // but not 1 to 0) between successive calls between which the heap was 2569 // not collected. For the implementation below, it must thus rely on 2570 // the property that concurrent_cycles_since_last_unload() 2571 // will not decrease unless a collection cycle happened and that 2572 // _cmsGen->is_too_full() are 2573 // themselves also monotonic in that sense. See check_monotonicity() 2574 // below. 2575 void CMSCollector::update_should_unload_classes() { 2576 _should_unload_classes = false; 2577 if (CMSClassUnloadingEnabled) { 2578 _should_unload_classes = (concurrent_cycles_since_last_unload() >= 2579 CMSClassUnloadingMaxInterval) 2580 || _cmsGen->is_too_full(); 2581 } 2582 } 2583 2584 bool ConcurrentMarkSweepGeneration::is_too_full() const { 2585 bool res = should_concurrent_collect(); 2586 res = res && (occupancy() > (double)CMSIsTooFullPercentage/100.0); 2587 return res; 2588 } 2589 2590 void CMSCollector::setup_cms_unloading_and_verification_state() { 2591 const bool should_verify = VerifyBeforeGC || VerifyAfterGC || VerifyDuringGC 2592 || VerifyBeforeExit; 2593 const int rso = GenCollectedHeap::SO_AllCodeCache; 2594 2595 // We set the proper root for this CMS cycle here. 2596 if (should_unload_classes()) { // Should unload classes this cycle 2597 remove_root_scanning_option(rso); // Shrink the root set appropriately 2598 set_verifying(should_verify); // Set verification state for this cycle 2599 return; // Nothing else needs to be done at this time 2600 } 2601 2602 // Not unloading classes this cycle 2603 assert(!should_unload_classes(), "Inconsistency!"); 2604 2605 // If we are not unloading classes then add SO_AllCodeCache to root 2606 // scanning options. 2607 add_root_scanning_option(rso); 2608 2609 if ((!verifying() || unloaded_classes_last_cycle()) && should_verify) { 2610 set_verifying(true); 2611 } else if (verifying() && !should_verify) { 2612 // We were verifying, but some verification flags got disabled. 2613 set_verifying(false); 2614 // Exclude symbols, strings and code cache elements from root scanning to 2615 // reduce IM and RM pauses. 2616 remove_root_scanning_option(rso); 2617 } 2618 } 2619 2620 2621 #ifndef PRODUCT 2622 HeapWord* CMSCollector::block_start(const void* p) const { 2623 const HeapWord* addr = (HeapWord*)p; 2624 if (_span.contains(p)) { 2625 if (_cmsGen->cmsSpace()->is_in_reserved(addr)) { 2626 return _cmsGen->cmsSpace()->block_start(p); 2627 } 2628 } 2629 return NULL; 2630 } 2631 #endif 2632 2633 HeapWord* 2634 ConcurrentMarkSweepGeneration::expand_and_allocate(size_t word_size, 2635 bool tlab, 2636 bool parallel) { 2637 CMSSynchronousYieldRequest yr; 2638 assert(!tlab, "Can't deal with TLAB allocation"); 2639 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2640 expand_for_gc_cause(word_size*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_satisfy_allocation); 2641 if (GCExpandToAllocateDelayMillis > 0) { 2642 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 2643 } 2644 return have_lock_and_allocate(word_size, tlab); 2645 } 2646 2647 void ConcurrentMarkSweepGeneration::expand_for_gc_cause( 2648 size_t bytes, 2649 size_t expand_bytes, 2650 CMSExpansionCause::Cause cause) 2651 { 2652 2653 bool success = expand(bytes, expand_bytes); 2654 2655 // remember why we expanded; this information is used 2656 // by shouldConcurrentCollect() when making decisions on whether to start 2657 // a new CMS cycle. 2658 if (success) { 2659 set_expansion_cause(cause); 2660 log_trace(gc)("Expanded CMS gen for %s", CMSExpansionCause::to_string(cause)); 2661 } 2662 } 2663 2664 HeapWord* ConcurrentMarkSweepGeneration::expand_and_par_lab_allocate(CMSParGCThreadState* ps, size_t word_sz) { 2665 HeapWord* res = NULL; 2666 MutexLocker x(ParGCRareEvent_lock); 2667 while (true) { 2668 // Expansion by some other thread might make alloc OK now: 2669 res = ps->lab.alloc(word_sz); 2670 if (res != NULL) return res; 2671 // If there's not enough expansion space available, give up. 2672 if (_virtual_space.uncommitted_size() < (word_sz * HeapWordSize)) { 2673 return NULL; 2674 } 2675 // Otherwise, we try expansion. 2676 expand_for_gc_cause(word_sz*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_allocate_par_lab); 2677 // Now go around the loop and try alloc again; 2678 // A competing par_promote might beat us to the expansion space, 2679 // so we may go around the loop again if promotion fails again. 2680 if (GCExpandToAllocateDelayMillis > 0) { 2681 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 2682 } 2683 } 2684 } 2685 2686 2687 bool ConcurrentMarkSweepGeneration::expand_and_ensure_spooling_space( 2688 PromotionInfo* promo) { 2689 MutexLocker x(ParGCRareEvent_lock); 2690 size_t refill_size_bytes = promo->refillSize() * HeapWordSize; 2691 while (true) { 2692 // Expansion by some other thread might make alloc OK now: 2693 if (promo->ensure_spooling_space()) { 2694 assert(promo->has_spooling_space(), 2695 "Post-condition of successful ensure_spooling_space()"); 2696 return true; 2697 } 2698 // If there's not enough expansion space available, give up. 2699 if (_virtual_space.uncommitted_size() < refill_size_bytes) { 2700 return false; 2701 } 2702 // Otherwise, we try expansion. 2703 expand_for_gc_cause(refill_size_bytes, MinHeapDeltaBytes, CMSExpansionCause::_allocate_par_spooling_space); 2704 // Now go around the loop and try alloc again; 2705 // A competing allocation might beat us to the expansion space, 2706 // so we may go around the loop again if allocation fails again. 2707 if (GCExpandToAllocateDelayMillis > 0) { 2708 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 2709 } 2710 } 2711 } 2712 2713 void ConcurrentMarkSweepGeneration::shrink(size_t bytes) { 2714 // Only shrink if a compaction was done so that all the free space 2715 // in the generation is in a contiguous block at the end. 2716 if (did_compact()) { 2717 CardGeneration::shrink(bytes); 2718 } 2719 } 2720 2721 void ConcurrentMarkSweepGeneration::assert_correct_size_change_locking() { 2722 assert_locked_or_safepoint(Heap_lock); 2723 } 2724 2725 void ConcurrentMarkSweepGeneration::shrink_free_list_by(size_t bytes) { 2726 assert_locked_or_safepoint(Heap_lock); 2727 assert_lock_strong(freelistLock()); 2728 log_trace(gc)("Shrinking of CMS not yet implemented"); 2729 return; 2730 } 2731 2732 2733 // Simple ctor/dtor wrapper for accounting & timer chores around concurrent 2734 // phases. 2735 class CMSPhaseAccounting: public StackObj { 2736 public: 2737 CMSPhaseAccounting(CMSCollector *collector, 2738 const char *title); 2739 ~CMSPhaseAccounting(); 2740 2741 private: 2742 CMSCollector *_collector; 2743 const char *_title; 2744 GCTraceConcTime(Info, gc) _trace_time; 2745 2746 public: 2747 // Not MT-safe; so do not pass around these StackObj's 2748 // where they may be accessed by other threads. 2749 double wallclock_millis() { 2750 return TimeHelper::counter_to_millis(os::elapsed_counter() - _trace_time.start_time()); 2751 } 2752 }; 2753 2754 CMSPhaseAccounting::CMSPhaseAccounting(CMSCollector *collector, 2755 const char *title) : 2756 _collector(collector), _title(title), _trace_time(title) { 2757 2758 _collector->resetYields(); 2759 _collector->resetTimer(); 2760 _collector->startTimer(); 2761 _collector->gc_timer_cm()->register_gc_concurrent_start(title); 2762 } 2763 2764 CMSPhaseAccounting::~CMSPhaseAccounting() { 2765 _collector->gc_timer_cm()->register_gc_concurrent_end(); 2766 _collector->stopTimer(); 2767 log_debug(gc)("Concurrent active time: %.3fms", TimeHelper::counter_to_seconds(_collector->timerTicks())); 2768 log_trace(gc)(" (CMS %s yielded %d times)", _title, _collector->yields()); 2769 } 2770 2771 // CMS work 2772 2773 // The common parts of CMSParInitialMarkTask and CMSParRemarkTask. 2774 class CMSParMarkTask : public AbstractGangTask { 2775 protected: 2776 CMSCollector* _collector; 2777 uint _n_workers; 2778 CMSParMarkTask(const char* name, CMSCollector* collector, uint n_workers) : 2779 AbstractGangTask(name), 2780 _collector(collector), 2781 _n_workers(n_workers) {} 2782 // Work method in support of parallel rescan ... of young gen spaces 2783 void do_young_space_rescan(OopsInGenClosure* cl, 2784 ContiguousSpace* space, 2785 HeapWord** chunk_array, size_t chunk_top); 2786 void work_on_young_gen_roots(OopsInGenClosure* cl); 2787 }; 2788 2789 // Parallel initial mark task 2790 class CMSParInitialMarkTask: public CMSParMarkTask { 2791 StrongRootsScope* _strong_roots_scope; 2792 public: 2793 CMSParInitialMarkTask(CMSCollector* collector, StrongRootsScope* strong_roots_scope, uint n_workers) : 2794 CMSParMarkTask("Scan roots and young gen for initial mark in parallel", collector, n_workers), 2795 _strong_roots_scope(strong_roots_scope) {} 2796 void work(uint worker_id); 2797 }; 2798 2799 // Checkpoint the roots into this generation from outside 2800 // this generation. [Note this initial checkpoint need only 2801 // be approximate -- we'll do a catch up phase subsequently.] 2802 void CMSCollector::checkpointRootsInitial() { 2803 assert(_collectorState == InitialMarking, "Wrong collector state"); 2804 check_correct_thread_executing(); 2805 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 2806 2807 save_heap_summary(); 2808 report_heap_summary(GCWhen::BeforeGC); 2809 2810 ReferenceProcessor* rp = ref_processor(); 2811 assert(_restart_addr == NULL, "Control point invariant"); 2812 { 2813 // acquire locks for subsequent manipulations 2814 MutexLockerEx x(bitMapLock(), 2815 Mutex::_no_safepoint_check_flag); 2816 checkpointRootsInitialWork(); 2817 // enable ("weak") refs discovery 2818 rp->enable_discovery(); 2819 _collectorState = Marking; 2820 } 2821 } 2822 2823 void CMSCollector::checkpointRootsInitialWork() { 2824 assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped"); 2825 assert(_collectorState == InitialMarking, "just checking"); 2826 2827 // Already have locks. 2828 assert_lock_strong(bitMapLock()); 2829 assert(_markBitMap.isAllClear(), "was reset at end of previous cycle"); 2830 2831 // Setup the verification and class unloading state for this 2832 // CMS collection cycle. 2833 setup_cms_unloading_and_verification_state(); 2834 2835 GCTraceTime(Trace, gc, phases) ts("checkpointRootsInitialWork", _gc_timer_cm); 2836 2837 // Reset all the PLAB chunk arrays if necessary. 2838 if (_survivor_plab_array != NULL && !CMSPLABRecordAlways) { 2839 reset_survivor_plab_arrays(); 2840 } 2841 2842 ResourceMark rm; 2843 HandleMark hm; 2844 2845 MarkRefsIntoClosure notOlder(_span, &_markBitMap); 2846 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2847 2848 verify_work_stacks_empty(); 2849 verify_overflow_empty(); 2850 2851 gch->ensure_parsability(false); // fill TLABs, but no need to retire them 2852 // Update the saved marks which may affect the root scans. 2853 gch->save_marks(); 2854 2855 // weak reference processing has not started yet. 2856 ref_processor()->set_enqueuing_is_done(false); 2857 2858 // Need to remember all newly created CLDs, 2859 // so that we can guarantee that the remark finds them. 2860 ClassLoaderDataGraph::remember_new_clds(true); 2861 2862 // Whenever a CLD is found, it will be claimed before proceeding to mark 2863 // the klasses. The claimed marks need to be cleared before marking starts. 2864 ClassLoaderDataGraph::clear_claimed_marks(); 2865 2866 print_eden_and_survivor_chunk_arrays(); 2867 2868 { 2869 #if defined(COMPILER2) || INCLUDE_JVMCI 2870 DerivedPointerTableDeactivate dpt_deact; 2871 #endif 2872 if (CMSParallelInitialMarkEnabled) { 2873 // The parallel version. 2874 WorkGang* workers = gch->workers(); 2875 assert(workers != NULL, "Need parallel worker threads."); 2876 uint n_workers = workers->active_workers(); 2877 2878 StrongRootsScope srs(n_workers); 2879 2880 CMSParInitialMarkTask tsk(this, &srs, n_workers); 2881 initialize_sequential_subtasks_for_young_gen_rescan(n_workers); 2882 // If the total workers is greater than 1, then multiple workers 2883 // may be used at some time and the initialization has been set 2884 // such that the single threaded path cannot be used. 2885 if (workers->total_workers() > 1) { 2886 workers->run_task(&tsk); 2887 } else { 2888 tsk.work(0); 2889 } 2890 } else { 2891 // The serial version. 2892 CLDToOopClosure cld_closure(¬Older, true); 2893 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2894 2895 StrongRootsScope srs(1); 2896 2897 gch->cms_process_roots(&srs, 2898 true, // young gen as roots 2899 GenCollectedHeap::ScanningOption(roots_scanning_options()), 2900 should_unload_classes(), 2901 ¬Older, 2902 &cld_closure); 2903 } 2904 } 2905 2906 // Clear mod-union table; it will be dirtied in the prologue of 2907 // CMS generation per each young generation collection. 2908 2909 assert(_modUnionTable.isAllClear(), 2910 "Was cleared in most recent final checkpoint phase" 2911 " or no bits are set in the gc_prologue before the start of the next " 2912 "subsequent marking phase."); 2913 2914 assert(_ct->klass_rem_set()->mod_union_is_clear(), "Must be"); 2915 2916 // Save the end of the used_region of the constituent generations 2917 // to be used to limit the extent of sweep in each generation. 2918 save_sweep_limits(); 2919 verify_overflow_empty(); 2920 } 2921 2922 bool CMSCollector::markFromRoots() { 2923 // we might be tempted to assert that: 2924 // assert(!SafepointSynchronize::is_at_safepoint(), 2925 // "inconsistent argument?"); 2926 // However that wouldn't be right, because it's possible that 2927 // a safepoint is indeed in progress as a young generation 2928 // stop-the-world GC happens even as we mark in this generation. 2929 assert(_collectorState == Marking, "inconsistent state?"); 2930 check_correct_thread_executing(); 2931 verify_overflow_empty(); 2932 2933 // Weak ref discovery note: We may be discovering weak 2934 // refs in this generation concurrent (but interleaved) with 2935 // weak ref discovery by the young generation collector. 2936 2937 CMSTokenSyncWithLocks ts(true, bitMapLock()); 2938 GCTraceCPUTime tcpu; 2939 CMSPhaseAccounting pa(this, "Concurrent Mark"); 2940 bool res = markFromRootsWork(); 2941 if (res) { 2942 _collectorState = Precleaning; 2943 } else { // We failed and a foreground collection wants to take over 2944 assert(_foregroundGCIsActive, "internal state inconsistency"); 2945 assert(_restart_addr == NULL, "foreground will restart from scratch"); 2946 log_debug(gc)("bailing out to foreground collection"); 2947 } 2948 verify_overflow_empty(); 2949 return res; 2950 } 2951 2952 bool CMSCollector::markFromRootsWork() { 2953 // iterate over marked bits in bit map, doing a full scan and mark 2954 // from these roots using the following algorithm: 2955 // . if oop is to the right of the current scan pointer, 2956 // mark corresponding bit (we'll process it later) 2957 // . else (oop is to left of current scan pointer) 2958 // push oop on marking stack 2959 // . drain the marking stack 2960 2961 // Note that when we do a marking step we need to hold the 2962 // bit map lock -- recall that direct allocation (by mutators) 2963 // and promotion (by the young generation collector) is also 2964 // marking the bit map. [the so-called allocate live policy.] 2965 // Because the implementation of bit map marking is not 2966 // robust wrt simultaneous marking of bits in the same word, 2967 // we need to make sure that there is no such interference 2968 // between concurrent such updates. 2969 2970 // already have locks 2971 assert_lock_strong(bitMapLock()); 2972 2973 verify_work_stacks_empty(); 2974 verify_overflow_empty(); 2975 bool result = false; 2976 if (CMSConcurrentMTEnabled && ConcGCThreads > 0) { 2977 result = do_marking_mt(); 2978 } else { 2979 result = do_marking_st(); 2980 } 2981 return result; 2982 } 2983 2984 // Forward decl 2985 class CMSConcMarkingTask; 2986 2987 class CMSConcMarkingTerminator: public ParallelTaskTerminator { 2988 CMSCollector* _collector; 2989 CMSConcMarkingTask* _task; 2990 public: 2991 virtual void yield(); 2992 2993 // "n_threads" is the number of threads to be terminated. 2994 // "queue_set" is a set of work queues of other threads. 2995 // "collector" is the CMS collector associated with this task terminator. 2996 // "yield" indicates whether we need the gang as a whole to yield. 2997 CMSConcMarkingTerminator(int n_threads, TaskQueueSetSuper* queue_set, CMSCollector* collector) : 2998 ParallelTaskTerminator(n_threads, queue_set), 2999 _collector(collector) { } 3000 3001 void set_task(CMSConcMarkingTask* task) { 3002 _task = task; 3003 } 3004 }; 3005 3006 class CMSConcMarkingTerminatorTerminator: public TerminatorTerminator { 3007 CMSConcMarkingTask* _task; 3008 public: 3009 bool should_exit_termination(); 3010 void set_task(CMSConcMarkingTask* task) { 3011 _task = task; 3012 } 3013 }; 3014 3015 // MT Concurrent Marking Task 3016 class CMSConcMarkingTask: public YieldingFlexibleGangTask { 3017 CMSCollector* _collector; 3018 uint _n_workers; // requested/desired # workers 3019 bool _result; 3020 CompactibleFreeListSpace* _cms_space; 3021 char _pad_front[64]; // padding to ... 3022 HeapWord* volatile _global_finger; // ... avoid sharing cache line 3023 char _pad_back[64]; 3024 HeapWord* _restart_addr; 3025 3026 // Exposed here for yielding support 3027 Mutex* const _bit_map_lock; 3028 3029 // The per thread work queues, available here for stealing 3030 OopTaskQueueSet* _task_queues; 3031 3032 // Termination (and yielding) support 3033 CMSConcMarkingTerminator _term; 3034 CMSConcMarkingTerminatorTerminator _term_term; 3035 3036 public: 3037 CMSConcMarkingTask(CMSCollector* collector, 3038 CompactibleFreeListSpace* cms_space, 3039 YieldingFlexibleWorkGang* workers, 3040 OopTaskQueueSet* task_queues): 3041 YieldingFlexibleGangTask("Concurrent marking done multi-threaded"), 3042 _collector(collector), 3043 _cms_space(cms_space), 3044 _n_workers(0), _result(true), 3045 _task_queues(task_queues), 3046 _term(_n_workers, task_queues, _collector), 3047 _bit_map_lock(collector->bitMapLock()) 3048 { 3049 _requested_size = _n_workers; 3050 _term.set_task(this); 3051 _term_term.set_task(this); 3052 _restart_addr = _global_finger = _cms_space->bottom(); 3053 } 3054 3055 3056 OopTaskQueueSet* task_queues() { return _task_queues; } 3057 3058 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 3059 3060 HeapWord* volatile* global_finger_addr() { return &_global_finger; } 3061 3062 CMSConcMarkingTerminator* terminator() { return &_term; } 3063 3064 virtual void set_for_termination(uint active_workers) { 3065 terminator()->reset_for_reuse(active_workers); 3066 } 3067 3068 void work(uint worker_id); 3069 bool should_yield() { 3070 return ConcurrentMarkSweepThread::should_yield() 3071 && !_collector->foregroundGCIsActive(); 3072 } 3073 3074 virtual void coordinator_yield(); // stuff done by coordinator 3075 bool result() { return _result; } 3076 3077 void reset(HeapWord* ra) { 3078 assert(_global_finger >= _cms_space->end(), "Postcondition of ::work(i)"); 3079 _restart_addr = _global_finger = ra; 3080 _term.reset_for_reuse(); 3081 } 3082 3083 static bool get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 3084 OopTaskQueue* work_q); 3085 3086 private: 3087 void do_scan_and_mark(int i, CompactibleFreeListSpace* sp); 3088 void do_work_steal(int i); 3089 void bump_global_finger(HeapWord* f); 3090 }; 3091 3092 bool CMSConcMarkingTerminatorTerminator::should_exit_termination() { 3093 assert(_task != NULL, "Error"); 3094 return _task->yielding(); 3095 // Note that we do not need the disjunct || _task->should_yield() above 3096 // because we want terminating threads to yield only if the task 3097 // is already in the midst of yielding, which happens only after at least one 3098 // thread has yielded. 3099 } 3100 3101 void CMSConcMarkingTerminator::yield() { 3102 if (_task->should_yield()) { 3103 _task->yield(); 3104 } else { 3105 ParallelTaskTerminator::yield(); 3106 } 3107 } 3108 3109 //////////////////////////////////////////////////////////////// 3110 // Concurrent Marking Algorithm Sketch 3111 //////////////////////////////////////////////////////////////// 3112 // Until all tasks exhausted (both spaces): 3113 // -- claim next available chunk 3114 // -- bump global finger via CAS 3115 // -- find first object that starts in this chunk 3116 // and start scanning bitmap from that position 3117 // -- scan marked objects for oops 3118 // -- CAS-mark target, and if successful: 3119 // . if target oop is above global finger (volatile read) 3120 // nothing to do 3121 // . if target oop is in chunk and above local finger 3122 // then nothing to do 3123 // . else push on work-queue 3124 // -- Deal with possible overflow issues: 3125 // . local work-queue overflow causes stuff to be pushed on 3126 // global (common) overflow queue 3127 // . always first empty local work queue 3128 // . then get a batch of oops from global work queue if any 3129 // . then do work stealing 3130 // -- When all tasks claimed (both spaces) 3131 // and local work queue empty, 3132 // then in a loop do: 3133 // . check global overflow stack; steal a batch of oops and trace 3134 // . try to steal from other threads oif GOS is empty 3135 // . if neither is available, offer termination 3136 // -- Terminate and return result 3137 // 3138 void CMSConcMarkingTask::work(uint worker_id) { 3139 elapsedTimer _timer; 3140 ResourceMark rm; 3141 HandleMark hm; 3142 3143 DEBUG_ONLY(_collector->verify_overflow_empty();) 3144 3145 // Before we begin work, our work queue should be empty 3146 assert(work_queue(worker_id)->size() == 0, "Expected to be empty"); 3147 // Scan the bitmap covering _cms_space, tracing through grey objects. 3148 _timer.start(); 3149 do_scan_and_mark(worker_id, _cms_space); 3150 _timer.stop(); 3151 log_trace(gc, task)("Finished cms space scanning in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 3152 3153 // ... do work stealing 3154 _timer.reset(); 3155 _timer.start(); 3156 do_work_steal(worker_id); 3157 _timer.stop(); 3158 log_trace(gc, task)("Finished work stealing in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 3159 assert(_collector->_markStack.isEmpty(), "Should have been emptied"); 3160 assert(work_queue(worker_id)->size() == 0, "Should have been emptied"); 3161 // Note that under the current task protocol, the 3162 // following assertion is true even of the spaces 3163 // expanded since the completion of the concurrent 3164 // marking. XXX This will likely change under a strict 3165 // ABORT semantics. 3166 // After perm removal the comparison was changed to 3167 // greater than or equal to from strictly greater than. 3168 // Before perm removal the highest address sweep would 3169 // have been at the end of perm gen but now is at the 3170 // end of the tenured gen. 3171 assert(_global_finger >= _cms_space->end(), 3172 "All tasks have been completed"); 3173 DEBUG_ONLY(_collector->verify_overflow_empty();) 3174 } 3175 3176 void CMSConcMarkingTask::bump_global_finger(HeapWord* f) { 3177 HeapWord* read = _global_finger; 3178 HeapWord* cur = read; 3179 while (f > read) { 3180 cur = read; 3181 read = (HeapWord*) Atomic::cmpxchg_ptr(f, &_global_finger, cur); 3182 if (cur == read) { 3183 // our cas succeeded 3184 assert(_global_finger >= f, "protocol consistency"); 3185 break; 3186 } 3187 } 3188 } 3189 3190 // This is really inefficient, and should be redone by 3191 // using (not yet available) block-read and -write interfaces to the 3192 // stack and the work_queue. XXX FIX ME !!! 3193 bool CMSConcMarkingTask::get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 3194 OopTaskQueue* work_q) { 3195 // Fast lock-free check 3196 if (ovflw_stk->length() == 0) { 3197 return false; 3198 } 3199 assert(work_q->size() == 0, "Shouldn't steal"); 3200 MutexLockerEx ml(ovflw_stk->par_lock(), 3201 Mutex::_no_safepoint_check_flag); 3202 // Grab up to 1/4 the size of the work queue 3203 size_t num = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 3204 (size_t)ParGCDesiredObjsFromOverflowList); 3205 num = MIN2(num, ovflw_stk->length()); 3206 for (int i = (int) num; i > 0; i--) { 3207 oop cur = ovflw_stk->pop(); 3208 assert(cur != NULL, "Counted wrong?"); 3209 work_q->push(cur); 3210 } 3211 return num > 0; 3212 } 3213 3214 void CMSConcMarkingTask::do_scan_and_mark(int i, CompactibleFreeListSpace* sp) { 3215 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 3216 int n_tasks = pst->n_tasks(); 3217 // We allow that there may be no tasks to do here because 3218 // we are restarting after a stack overflow. 3219 assert(pst->valid() || n_tasks == 0, "Uninitialized use?"); 3220 uint nth_task = 0; 3221 3222 HeapWord* aligned_start = sp->bottom(); 3223 if (sp->used_region().contains(_restart_addr)) { 3224 // Align down to a card boundary for the start of 0th task 3225 // for this space. 3226 aligned_start = align_down(_restart_addr, CardTableModRefBS::card_size); 3227 } 3228 3229 size_t chunk_size = sp->marking_task_size(); 3230 while (!pst->is_task_claimed(/* reference */ nth_task)) { 3231 // Having claimed the nth task in this space, 3232 // compute the chunk that it corresponds to: 3233 MemRegion span = MemRegion(aligned_start + nth_task*chunk_size, 3234 aligned_start + (nth_task+1)*chunk_size); 3235 // Try and bump the global finger via a CAS; 3236 // note that we need to do the global finger bump 3237 // _before_ taking the intersection below, because 3238 // the task corresponding to that region will be 3239 // deemed done even if the used_region() expands 3240 // because of allocation -- as it almost certainly will 3241 // during start-up while the threads yield in the 3242 // closure below. 3243 HeapWord* finger = span.end(); 3244 bump_global_finger(finger); // atomically 3245 // There are null tasks here corresponding to chunks 3246 // beyond the "top" address of the space. 3247 span = span.intersection(sp->used_region()); 3248 if (!span.is_empty()) { // Non-null task 3249 HeapWord* prev_obj; 3250 assert(!span.contains(_restart_addr) || nth_task == 0, 3251 "Inconsistency"); 3252 if (nth_task == 0) { 3253 // For the 0th task, we'll not need to compute a block_start. 3254 if (span.contains(_restart_addr)) { 3255 // In the case of a restart because of stack overflow, 3256 // we might additionally skip a chunk prefix. 3257 prev_obj = _restart_addr; 3258 } else { 3259 prev_obj = span.start(); 3260 } 3261 } else { 3262 // We want to skip the first object because 3263 // the protocol is to scan any object in its entirety 3264 // that _starts_ in this span; a fortiori, any 3265 // object starting in an earlier span is scanned 3266 // as part of an earlier claimed task. 3267 // Below we use the "careful" version of block_start 3268 // so we do not try to navigate uninitialized objects. 3269 prev_obj = sp->block_start_careful(span.start()); 3270 // Below we use a variant of block_size that uses the 3271 // Printezis bits to avoid waiting for allocated 3272 // objects to become initialized/parsable. 3273 while (prev_obj < span.start()) { 3274 size_t sz = sp->block_size_no_stall(prev_obj, _collector); 3275 if (sz > 0) { 3276 prev_obj += sz; 3277 } else { 3278 // In this case we may end up doing a bit of redundant 3279 // scanning, but that appears unavoidable, short of 3280 // locking the free list locks; see bug 6324141. 3281 break; 3282 } 3283 } 3284 } 3285 if (prev_obj < span.end()) { 3286 MemRegion my_span = MemRegion(prev_obj, span.end()); 3287 // Do the marking work within a non-empty span -- 3288 // the last argument to the constructor indicates whether the 3289 // iteration should be incremental with periodic yields. 3290 ParMarkFromRootsClosure cl(this, _collector, my_span, 3291 &_collector->_markBitMap, 3292 work_queue(i), 3293 &_collector->_markStack); 3294 _collector->_markBitMap.iterate(&cl, my_span.start(), my_span.end()); 3295 } // else nothing to do for this task 3296 } // else nothing to do for this task 3297 } 3298 // We'd be tempted to assert here that since there are no 3299 // more tasks left to claim in this space, the global_finger 3300 // must exceed space->top() and a fortiori space->end(). However, 3301 // that would not quite be correct because the bumping of 3302 // global_finger occurs strictly after the claiming of a task, 3303 // so by the time we reach here the global finger may not yet 3304 // have been bumped up by the thread that claimed the last 3305 // task. 3306 pst->all_tasks_completed(); 3307 } 3308 3309 class ParConcMarkingClosure: public MetadataAwareOopClosure { 3310 private: 3311 CMSCollector* _collector; 3312 CMSConcMarkingTask* _task; 3313 MemRegion _span; 3314 CMSBitMap* _bit_map; 3315 CMSMarkStack* _overflow_stack; 3316 OopTaskQueue* _work_queue; 3317 protected: 3318 DO_OOP_WORK_DEFN 3319 public: 3320 ParConcMarkingClosure(CMSCollector* collector, CMSConcMarkingTask* task, OopTaskQueue* work_queue, 3321 CMSBitMap* bit_map, CMSMarkStack* overflow_stack): 3322 MetadataAwareOopClosure(collector->ref_processor()), 3323 _collector(collector), 3324 _task(task), 3325 _span(collector->_span), 3326 _work_queue(work_queue), 3327 _bit_map(bit_map), 3328 _overflow_stack(overflow_stack) 3329 { } 3330 virtual void do_oop(oop* p); 3331 virtual void do_oop(narrowOop* p); 3332 3333 void trim_queue(size_t max); 3334 void handle_stack_overflow(HeapWord* lost); 3335 void do_yield_check() { 3336 if (_task->should_yield()) { 3337 _task->yield(); 3338 } 3339 } 3340 }; 3341 3342 DO_OOP_WORK_IMPL(ParConcMarkingClosure) 3343 3344 // Grey object scanning during work stealing phase -- 3345 // the salient assumption here is that any references 3346 // that are in these stolen objects being scanned must 3347 // already have been initialized (else they would not have 3348 // been published), so we do not need to check for 3349 // uninitialized objects before pushing here. 3350 void ParConcMarkingClosure::do_oop(oop obj) { 3351 assert(obj->is_oop_or_null(true), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 3352 HeapWord* addr = (HeapWord*)obj; 3353 // Check if oop points into the CMS generation 3354 // and is not marked 3355 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 3356 // a white object ... 3357 // If we manage to "claim" the object, by being the 3358 // first thread to mark it, then we push it on our 3359 // marking stack 3360 if (_bit_map->par_mark(addr)) { // ... now grey 3361 // push on work queue (grey set) 3362 bool simulate_overflow = false; 3363 NOT_PRODUCT( 3364 if (CMSMarkStackOverflowALot && 3365 _collector->simulate_overflow()) { 3366 // simulate a stack overflow 3367 simulate_overflow = true; 3368 } 3369 ) 3370 if (simulate_overflow || 3371 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) { 3372 // stack overflow 3373 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _overflow_stack->capacity()); 3374 // We cannot assert that the overflow stack is full because 3375 // it may have been emptied since. 3376 assert(simulate_overflow || 3377 _work_queue->size() == _work_queue->max_elems(), 3378 "Else push should have succeeded"); 3379 handle_stack_overflow(addr); 3380 } 3381 } // Else, some other thread got there first 3382 do_yield_check(); 3383 } 3384 } 3385 3386 void ParConcMarkingClosure::do_oop(oop* p) { ParConcMarkingClosure::do_oop_work(p); } 3387 void ParConcMarkingClosure::do_oop(narrowOop* p) { ParConcMarkingClosure::do_oop_work(p); } 3388 3389 void ParConcMarkingClosure::trim_queue(size_t max) { 3390 while (_work_queue->size() > max) { 3391 oop new_oop; 3392 if (_work_queue->pop_local(new_oop)) { 3393 assert(new_oop->is_oop(), "Should be an oop"); 3394 assert(_bit_map->isMarked((HeapWord*)new_oop), "Grey object"); 3395 assert(_span.contains((HeapWord*)new_oop), "Not in span"); 3396 new_oop->oop_iterate(this); // do_oop() above 3397 do_yield_check(); 3398 } 3399 } 3400 } 3401 3402 // Upon stack overflow, we discard (part of) the stack, 3403 // remembering the least address amongst those discarded 3404 // in CMSCollector's _restart_address. 3405 void ParConcMarkingClosure::handle_stack_overflow(HeapWord* lost) { 3406 // We need to do this under a mutex to prevent other 3407 // workers from interfering with the work done below. 3408 MutexLockerEx ml(_overflow_stack->par_lock(), 3409 Mutex::_no_safepoint_check_flag); 3410 // Remember the least grey address discarded 3411 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 3412 _collector->lower_restart_addr(ra); 3413 _overflow_stack->reset(); // discard stack contents 3414 _overflow_stack->expand(); // expand the stack if possible 3415 } 3416 3417 3418 void CMSConcMarkingTask::do_work_steal(int i) { 3419 OopTaskQueue* work_q = work_queue(i); 3420 oop obj_to_scan; 3421 CMSBitMap* bm = &(_collector->_markBitMap); 3422 CMSMarkStack* ovflw = &(_collector->_markStack); 3423 int* seed = _collector->hash_seed(i); 3424 ParConcMarkingClosure cl(_collector, this, work_q, bm, ovflw); 3425 while (true) { 3426 cl.trim_queue(0); 3427 assert(work_q->size() == 0, "Should have been emptied above"); 3428 if (get_work_from_overflow_stack(ovflw, work_q)) { 3429 // Can't assert below because the work obtained from the 3430 // overflow stack may already have been stolen from us. 3431 // assert(work_q->size() > 0, "Work from overflow stack"); 3432 continue; 3433 } else if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 3434 assert(obj_to_scan->is_oop(), "Should be an oop"); 3435 assert(bm->isMarked((HeapWord*)obj_to_scan), "Grey object"); 3436 obj_to_scan->oop_iterate(&cl); 3437 } else if (terminator()->offer_termination(&_term_term)) { 3438 assert(work_q->size() == 0, "Impossible!"); 3439 break; 3440 } else if (yielding() || should_yield()) { 3441 yield(); 3442 } 3443 } 3444 } 3445 3446 // This is run by the CMS (coordinator) thread. 3447 void CMSConcMarkingTask::coordinator_yield() { 3448 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 3449 "CMS thread should hold CMS token"); 3450 // First give up the locks, then yield, then re-lock 3451 // We should probably use a constructor/destructor idiom to 3452 // do this unlock/lock or modify the MutexUnlocker class to 3453 // serve our purpose. XXX 3454 assert_lock_strong(_bit_map_lock); 3455 _bit_map_lock->unlock(); 3456 ConcurrentMarkSweepThread::desynchronize(true); 3457 _collector->stopTimer(); 3458 _collector->incrementYields(); 3459 3460 // It is possible for whichever thread initiated the yield request 3461 // not to get a chance to wake up and take the bitmap lock between 3462 // this thread releasing it and reacquiring it. So, while the 3463 // should_yield() flag is on, let's sleep for a bit to give the 3464 // other thread a chance to wake up. The limit imposed on the number 3465 // of iterations is defensive, to avoid any unforseen circumstances 3466 // putting us into an infinite loop. Since it's always been this 3467 // (coordinator_yield()) method that was observed to cause the 3468 // problem, we are using a parameter (CMSCoordinatorYieldSleepCount) 3469 // which is by default non-zero. For the other seven methods that 3470 // also perform the yield operation, as are using a different 3471 // parameter (CMSYieldSleepCount) which is by default zero. This way we 3472 // can enable the sleeping for those methods too, if necessary. 3473 // See 6442774. 3474 // 3475 // We really need to reconsider the synchronization between the GC 3476 // thread and the yield-requesting threads in the future and we 3477 // should really use wait/notify, which is the recommended 3478 // way of doing this type of interaction. Additionally, we should 3479 // consolidate the eight methods that do the yield operation and they 3480 // are almost identical into one for better maintainability and 3481 // readability. See 6445193. 3482 // 3483 // Tony 2006.06.29 3484 for (unsigned i = 0; i < CMSCoordinatorYieldSleepCount && 3485 ConcurrentMarkSweepThread::should_yield() && 3486 !CMSCollector::foregroundGCIsActive(); ++i) { 3487 os::sleep(Thread::current(), 1, false); 3488 } 3489 3490 ConcurrentMarkSweepThread::synchronize(true); 3491 _bit_map_lock->lock_without_safepoint_check(); 3492 _collector->startTimer(); 3493 } 3494 3495 bool CMSCollector::do_marking_mt() { 3496 assert(ConcGCThreads > 0 && conc_workers() != NULL, "precondition"); 3497 uint num_workers = AdaptiveSizePolicy::calc_active_conc_workers(conc_workers()->total_workers(), 3498 conc_workers()->active_workers(), 3499 Threads::number_of_non_daemon_threads()); 3500 num_workers = conc_workers()->update_active_workers(num_workers); 3501 log_info(gc,task)("Using %u workers of %u for marking", num_workers, conc_workers()->total_workers()); 3502 3503 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 3504 3505 CMSConcMarkingTask tsk(this, 3506 cms_space, 3507 conc_workers(), 3508 task_queues()); 3509 3510 // Since the actual number of workers we get may be different 3511 // from the number we requested above, do we need to do anything different 3512 // below? In particular, may be we need to subclass the SequantialSubTasksDone 3513 // class?? XXX 3514 cms_space ->initialize_sequential_subtasks_for_marking(num_workers); 3515 3516 // Refs discovery is already non-atomic. 3517 assert(!ref_processor()->discovery_is_atomic(), "Should be non-atomic"); 3518 assert(ref_processor()->discovery_is_mt(), "Discovery should be MT"); 3519 conc_workers()->start_task(&tsk); 3520 while (tsk.yielded()) { 3521 tsk.coordinator_yield(); 3522 conc_workers()->continue_task(&tsk); 3523 } 3524 // If the task was aborted, _restart_addr will be non-NULL 3525 assert(tsk.completed() || _restart_addr != NULL, "Inconsistency"); 3526 while (_restart_addr != NULL) { 3527 // XXX For now we do not make use of ABORTED state and have not 3528 // yet implemented the right abort semantics (even in the original 3529 // single-threaded CMS case). That needs some more investigation 3530 // and is deferred for now; see CR# TBF. 07252005YSR. XXX 3531 assert(!CMSAbortSemantics || tsk.aborted(), "Inconsistency"); 3532 // If _restart_addr is non-NULL, a marking stack overflow 3533 // occurred; we need to do a fresh marking iteration from the 3534 // indicated restart address. 3535 if (_foregroundGCIsActive) { 3536 // We may be running into repeated stack overflows, having 3537 // reached the limit of the stack size, while making very 3538 // slow forward progress. It may be best to bail out and 3539 // let the foreground collector do its job. 3540 // Clear _restart_addr, so that foreground GC 3541 // works from scratch. This avoids the headache of 3542 // a "rescan" which would otherwise be needed because 3543 // of the dirty mod union table & card table. 3544 _restart_addr = NULL; 3545 return false; 3546 } 3547 // Adjust the task to restart from _restart_addr 3548 tsk.reset(_restart_addr); 3549 cms_space ->initialize_sequential_subtasks_for_marking(num_workers, 3550 _restart_addr); 3551 _restart_addr = NULL; 3552 // Get the workers going again 3553 conc_workers()->start_task(&tsk); 3554 while (tsk.yielded()) { 3555 tsk.coordinator_yield(); 3556 conc_workers()->continue_task(&tsk); 3557 } 3558 } 3559 assert(tsk.completed(), "Inconsistency"); 3560 assert(tsk.result() == true, "Inconsistency"); 3561 return true; 3562 } 3563 3564 bool CMSCollector::do_marking_st() { 3565 ResourceMark rm; 3566 HandleMark hm; 3567 3568 // Temporarily make refs discovery single threaded (non-MT) 3569 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false); 3570 MarkFromRootsClosure markFromRootsClosure(this, _span, &_markBitMap, 3571 &_markStack, CMSYield); 3572 // the last argument to iterate indicates whether the iteration 3573 // should be incremental with periodic yields. 3574 _markBitMap.iterate(&markFromRootsClosure); 3575 // If _restart_addr is non-NULL, a marking stack overflow 3576 // occurred; we need to do a fresh iteration from the 3577 // indicated restart address. 3578 while (_restart_addr != NULL) { 3579 if (_foregroundGCIsActive) { 3580 // We may be running into repeated stack overflows, having 3581 // reached the limit of the stack size, while making very 3582 // slow forward progress. It may be best to bail out and 3583 // let the foreground collector do its job. 3584 // Clear _restart_addr, so that foreground GC 3585 // works from scratch. This avoids the headache of 3586 // a "rescan" which would otherwise be needed because 3587 // of the dirty mod union table & card table. 3588 _restart_addr = NULL; 3589 return false; // indicating failure to complete marking 3590 } 3591 // Deal with stack overflow: 3592 // we restart marking from _restart_addr 3593 HeapWord* ra = _restart_addr; 3594 markFromRootsClosure.reset(ra); 3595 _restart_addr = NULL; 3596 _markBitMap.iterate(&markFromRootsClosure, ra, _span.end()); 3597 } 3598 return true; 3599 } 3600 3601 void CMSCollector::preclean() { 3602 check_correct_thread_executing(); 3603 assert(Thread::current()->is_ConcurrentGC_thread(), "Wrong thread"); 3604 verify_work_stacks_empty(); 3605 verify_overflow_empty(); 3606 _abort_preclean = false; 3607 if (CMSPrecleaningEnabled) { 3608 if (!CMSEdenChunksRecordAlways) { 3609 _eden_chunk_index = 0; 3610 } 3611 size_t used = get_eden_used(); 3612 size_t capacity = get_eden_capacity(); 3613 // Don't start sampling unless we will get sufficiently 3614 // many samples. 3615 if (used < (((capacity / CMSScheduleRemarkSamplingRatio) / 100) 3616 * CMSScheduleRemarkEdenPenetration)) { 3617 _start_sampling = true; 3618 } else { 3619 _start_sampling = false; 3620 } 3621 GCTraceCPUTime tcpu; 3622 CMSPhaseAccounting pa(this, "Concurrent Preclean"); 3623 preclean_work(CMSPrecleanRefLists1, CMSPrecleanSurvivors1); 3624 } 3625 CMSTokenSync x(true); // is cms thread 3626 if (CMSPrecleaningEnabled) { 3627 sample_eden(); 3628 _collectorState = AbortablePreclean; 3629 } else { 3630 _collectorState = FinalMarking; 3631 } 3632 verify_work_stacks_empty(); 3633 verify_overflow_empty(); 3634 } 3635 3636 // Try and schedule the remark such that young gen 3637 // occupancy is CMSScheduleRemarkEdenPenetration %. 3638 void CMSCollector::abortable_preclean() { 3639 check_correct_thread_executing(); 3640 assert(CMSPrecleaningEnabled, "Inconsistent control state"); 3641 assert(_collectorState == AbortablePreclean, "Inconsistent control state"); 3642 3643 // If Eden's current occupancy is below this threshold, 3644 // immediately schedule the remark; else preclean 3645 // past the next scavenge in an effort to 3646 // schedule the pause as described above. By choosing 3647 // CMSScheduleRemarkEdenSizeThreshold >= max eden size 3648 // we will never do an actual abortable preclean cycle. 3649 if (get_eden_used() > CMSScheduleRemarkEdenSizeThreshold) { 3650 GCTraceCPUTime tcpu; 3651 CMSPhaseAccounting pa(this, "Concurrent Abortable Preclean"); 3652 // We need more smarts in the abortable preclean 3653 // loop below to deal with cases where allocation 3654 // in young gen is very very slow, and our precleaning 3655 // is running a losing race against a horde of 3656 // mutators intent on flooding us with CMS updates 3657 // (dirty cards). 3658 // One, admittedly dumb, strategy is to give up 3659 // after a certain number of abortable precleaning loops 3660 // or after a certain maximum time. We want to make 3661 // this smarter in the next iteration. 3662 // XXX FIX ME!!! YSR 3663 size_t loops = 0, workdone = 0, cumworkdone = 0, waited = 0; 3664 while (!(should_abort_preclean() || 3665 ConcurrentMarkSweepThread::cmst()->should_terminate())) { 3666 workdone = preclean_work(CMSPrecleanRefLists2, CMSPrecleanSurvivors2); 3667 cumworkdone += workdone; 3668 loops++; 3669 // Voluntarily terminate abortable preclean phase if we have 3670 // been at it for too long. 3671 if ((CMSMaxAbortablePrecleanLoops != 0) && 3672 loops >= CMSMaxAbortablePrecleanLoops) { 3673 log_debug(gc)(" CMS: abort preclean due to loops "); 3674 break; 3675 } 3676 if (pa.wallclock_millis() > CMSMaxAbortablePrecleanTime) { 3677 log_debug(gc)(" CMS: abort preclean due to time "); 3678 break; 3679 } 3680 // If we are doing little work each iteration, we should 3681 // take a short break. 3682 if (workdone < CMSAbortablePrecleanMinWorkPerIteration) { 3683 // Sleep for some time, waiting for work to accumulate 3684 stopTimer(); 3685 cmsThread()->wait_on_cms_lock(CMSAbortablePrecleanWaitMillis); 3686 startTimer(); 3687 waited++; 3688 } 3689 } 3690 log_trace(gc)(" [" SIZE_FORMAT " iterations, " SIZE_FORMAT " waits, " SIZE_FORMAT " cards)] ", 3691 loops, waited, cumworkdone); 3692 } 3693 CMSTokenSync x(true); // is cms thread 3694 if (_collectorState != Idling) { 3695 assert(_collectorState == AbortablePreclean, 3696 "Spontaneous state transition?"); 3697 _collectorState = FinalMarking; 3698 } // Else, a foreground collection completed this CMS cycle. 3699 return; 3700 } 3701 3702 // Respond to an Eden sampling opportunity 3703 void CMSCollector::sample_eden() { 3704 // Make sure a young gc cannot sneak in between our 3705 // reading and recording of a sample. 3706 assert(Thread::current()->is_ConcurrentGC_thread(), 3707 "Only the cms thread may collect Eden samples"); 3708 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 3709 "Should collect samples while holding CMS token"); 3710 if (!_start_sampling) { 3711 return; 3712 } 3713 // When CMSEdenChunksRecordAlways is true, the eden chunk array 3714 // is populated by the young generation. 3715 if (_eden_chunk_array != NULL && !CMSEdenChunksRecordAlways) { 3716 if (_eden_chunk_index < _eden_chunk_capacity) { 3717 _eden_chunk_array[_eden_chunk_index] = *_top_addr; // take sample 3718 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr, 3719 "Unexpected state of Eden"); 3720 // We'd like to check that what we just sampled is an oop-start address; 3721 // however, we cannot do that here since the object may not yet have been 3722 // initialized. So we'll instead do the check when we _use_ this sample 3723 // later. 3724 if (_eden_chunk_index == 0 || 3725 (pointer_delta(_eden_chunk_array[_eden_chunk_index], 3726 _eden_chunk_array[_eden_chunk_index-1]) 3727 >= CMSSamplingGrain)) { 3728 _eden_chunk_index++; // commit sample 3729 } 3730 } 3731 } 3732 if ((_collectorState == AbortablePreclean) && !_abort_preclean) { 3733 size_t used = get_eden_used(); 3734 size_t capacity = get_eden_capacity(); 3735 assert(used <= capacity, "Unexpected state of Eden"); 3736 if (used > (capacity/100 * CMSScheduleRemarkEdenPenetration)) { 3737 _abort_preclean = true; 3738 } 3739 } 3740 } 3741 3742 3743 size_t CMSCollector::preclean_work(bool clean_refs, bool clean_survivor) { 3744 assert(_collectorState == Precleaning || 3745 _collectorState == AbortablePreclean, "incorrect state"); 3746 ResourceMark rm; 3747 HandleMark hm; 3748 3749 // Precleaning is currently not MT but the reference processor 3750 // may be set for MT. Disable it temporarily here. 3751 ReferenceProcessor* rp = ref_processor(); 3752 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(rp, false); 3753 3754 // Do one pass of scrubbing the discovered reference lists 3755 // to remove any reference objects with strongly-reachable 3756 // referents. 3757 if (clean_refs) { 3758 CMSPrecleanRefsYieldClosure yield_cl(this); 3759 assert(rp->span().equals(_span), "Spans should be equal"); 3760 CMSKeepAliveClosure keep_alive(this, _span, &_markBitMap, 3761 &_markStack, true /* preclean */); 3762 CMSDrainMarkingStackClosure complete_trace(this, 3763 _span, &_markBitMap, &_markStack, 3764 &keep_alive, true /* preclean */); 3765 3766 // We don't want this step to interfere with a young 3767 // collection because we don't want to take CPU 3768 // or memory bandwidth away from the young GC threads 3769 // (which may be as many as there are CPUs). 3770 // Note that we don't need to protect ourselves from 3771 // interference with mutators because they can't 3772 // manipulate the discovered reference lists nor affect 3773 // the computed reachability of the referents, the 3774 // only properties manipulated by the precleaning 3775 // of these reference lists. 3776 stopTimer(); 3777 CMSTokenSyncWithLocks x(true /* is cms thread */, 3778 bitMapLock()); 3779 startTimer(); 3780 sample_eden(); 3781 3782 // The following will yield to allow foreground 3783 // collection to proceed promptly. XXX YSR: 3784 // The code in this method may need further 3785 // tweaking for better performance and some restructuring 3786 // for cleaner interfaces. 3787 GCTimer *gc_timer = NULL; // Currently not tracing concurrent phases 3788 rp->preclean_discovered_references( 3789 rp->is_alive_non_header(), &keep_alive, &complete_trace, &yield_cl, 3790 gc_timer); 3791 } 3792 3793 if (clean_survivor) { // preclean the active survivor space(s) 3794 PushAndMarkClosure pam_cl(this, _span, ref_processor(), 3795 &_markBitMap, &_modUnionTable, 3796 &_markStack, true /* precleaning phase */); 3797 stopTimer(); 3798 CMSTokenSyncWithLocks ts(true /* is cms thread */, 3799 bitMapLock()); 3800 startTimer(); 3801 unsigned int before_count = 3802 GenCollectedHeap::heap()->total_collections(); 3803 SurvivorSpacePrecleanClosure 3804 sss_cl(this, _span, &_markBitMap, &_markStack, 3805 &pam_cl, before_count, CMSYield); 3806 _young_gen->from()->object_iterate_careful(&sss_cl); 3807 _young_gen->to()->object_iterate_careful(&sss_cl); 3808 } 3809 MarkRefsIntoAndScanClosure 3810 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable, 3811 &_markStack, this, CMSYield, 3812 true /* precleaning phase */); 3813 // CAUTION: The following closure has persistent state that may need to 3814 // be reset upon a decrease in the sequence of addresses it 3815 // processes. 3816 ScanMarkedObjectsAgainCarefullyClosure 3817 smoac_cl(this, _span, 3818 &_markBitMap, &_markStack, &mrias_cl, CMSYield); 3819 3820 // Preclean dirty cards in ModUnionTable and CardTable using 3821 // appropriate convergence criterion; 3822 // repeat CMSPrecleanIter times unless we find that 3823 // we are losing. 3824 assert(CMSPrecleanIter < 10, "CMSPrecleanIter is too large"); 3825 assert(CMSPrecleanNumerator < CMSPrecleanDenominator, 3826 "Bad convergence multiplier"); 3827 assert(CMSPrecleanThreshold >= 100, 3828 "Unreasonably low CMSPrecleanThreshold"); 3829 3830 size_t numIter, cumNumCards, lastNumCards, curNumCards; 3831 for (numIter = 0, cumNumCards = lastNumCards = curNumCards = 0; 3832 numIter < CMSPrecleanIter; 3833 numIter++, lastNumCards = curNumCards, cumNumCards += curNumCards) { 3834 curNumCards = preclean_mod_union_table(_cmsGen, &smoac_cl); 3835 log_trace(gc)(" (modUnionTable: " SIZE_FORMAT " cards)", curNumCards); 3836 // Either there are very few dirty cards, so re-mark 3837 // pause will be small anyway, or our pre-cleaning isn't 3838 // that much faster than the rate at which cards are being 3839 // dirtied, so we might as well stop and re-mark since 3840 // precleaning won't improve our re-mark time by much. 3841 if (curNumCards <= CMSPrecleanThreshold || 3842 (numIter > 0 && 3843 (curNumCards * CMSPrecleanDenominator > 3844 lastNumCards * CMSPrecleanNumerator))) { 3845 numIter++; 3846 cumNumCards += curNumCards; 3847 break; 3848 } 3849 } 3850 3851 preclean_klasses(&mrias_cl, _cmsGen->freelistLock()); 3852 3853 curNumCards = preclean_card_table(_cmsGen, &smoac_cl); 3854 cumNumCards += curNumCards; 3855 log_trace(gc)(" (cardTable: " SIZE_FORMAT " cards, re-scanned " SIZE_FORMAT " cards, " SIZE_FORMAT " iterations)", 3856 curNumCards, cumNumCards, numIter); 3857 return cumNumCards; // as a measure of useful work done 3858 } 3859 3860 // PRECLEANING NOTES: 3861 // Precleaning involves: 3862 // . reading the bits of the modUnionTable and clearing the set bits. 3863 // . For the cards corresponding to the set bits, we scan the 3864 // objects on those cards. This means we need the free_list_lock 3865 // so that we can safely iterate over the CMS space when scanning 3866 // for oops. 3867 // . When we scan the objects, we'll be both reading and setting 3868 // marks in the marking bit map, so we'll need the marking bit map. 3869 // . For protecting _collector_state transitions, we take the CGC_lock. 3870 // Note that any races in the reading of of card table entries by the 3871 // CMS thread on the one hand and the clearing of those entries by the 3872 // VM thread or the setting of those entries by the mutator threads on the 3873 // other are quite benign. However, for efficiency it makes sense to keep 3874 // the VM thread from racing with the CMS thread while the latter is 3875 // dirty card info to the modUnionTable. We therefore also use the 3876 // CGC_lock to protect the reading of the card table and the mod union 3877 // table by the CM thread. 3878 // . We run concurrently with mutator updates, so scanning 3879 // needs to be done carefully -- we should not try to scan 3880 // potentially uninitialized objects. 3881 // 3882 // Locking strategy: While holding the CGC_lock, we scan over and 3883 // reset a maximal dirty range of the mod union / card tables, then lock 3884 // the free_list_lock and bitmap lock to do a full marking, then 3885 // release these locks; and repeat the cycle. This allows for a 3886 // certain amount of fairness in the sharing of these locks between 3887 // the CMS collector on the one hand, and the VM thread and the 3888 // mutators on the other. 3889 3890 // NOTE: preclean_mod_union_table() and preclean_card_table() 3891 // further below are largely identical; if you need to modify 3892 // one of these methods, please check the other method too. 3893 3894 size_t CMSCollector::preclean_mod_union_table( 3895 ConcurrentMarkSweepGeneration* old_gen, 3896 ScanMarkedObjectsAgainCarefullyClosure* cl) { 3897 verify_work_stacks_empty(); 3898 verify_overflow_empty(); 3899 3900 // strategy: starting with the first card, accumulate contiguous 3901 // ranges of dirty cards; clear these cards, then scan the region 3902 // covered by these cards. 3903 3904 // Since all of the MUT is committed ahead, we can just use 3905 // that, in case the generations expand while we are precleaning. 3906 // It might also be fine to just use the committed part of the 3907 // generation, but we might potentially miss cards when the 3908 // generation is rapidly expanding while we are in the midst 3909 // of precleaning. 3910 HeapWord* startAddr = old_gen->reserved().start(); 3911 HeapWord* endAddr = old_gen->reserved().end(); 3912 3913 cl->setFreelistLock(old_gen->freelistLock()); // needed for yielding 3914 3915 size_t numDirtyCards, cumNumDirtyCards; 3916 HeapWord *nextAddr, *lastAddr; 3917 for (cumNumDirtyCards = numDirtyCards = 0, 3918 nextAddr = lastAddr = startAddr; 3919 nextAddr < endAddr; 3920 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 3921 3922 ResourceMark rm; 3923 HandleMark hm; 3924 3925 MemRegion dirtyRegion; 3926 { 3927 stopTimer(); 3928 // Potential yield point 3929 CMSTokenSync ts(true); 3930 startTimer(); 3931 sample_eden(); 3932 // Get dirty region starting at nextOffset (inclusive), 3933 // simultaneously clearing it. 3934 dirtyRegion = 3935 _modUnionTable.getAndClearMarkedRegion(nextAddr, endAddr); 3936 assert(dirtyRegion.start() >= nextAddr, 3937 "returned region inconsistent?"); 3938 } 3939 // Remember where the next search should begin. 3940 // The returned region (if non-empty) is a right open interval, 3941 // so lastOffset is obtained from the right end of that 3942 // interval. 3943 lastAddr = dirtyRegion.end(); 3944 // Should do something more transparent and less hacky XXX 3945 numDirtyCards = 3946 _modUnionTable.heapWordDiffToOffsetDiff(dirtyRegion.word_size()); 3947 3948 // We'll scan the cards in the dirty region (with periodic 3949 // yields for foreground GC as needed). 3950 if (!dirtyRegion.is_empty()) { 3951 assert(numDirtyCards > 0, "consistency check"); 3952 HeapWord* stop_point = NULL; 3953 stopTimer(); 3954 // Potential yield point 3955 CMSTokenSyncWithLocks ts(true, old_gen->freelistLock(), 3956 bitMapLock()); 3957 startTimer(); 3958 { 3959 verify_work_stacks_empty(); 3960 verify_overflow_empty(); 3961 sample_eden(); 3962 stop_point = 3963 old_gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 3964 } 3965 if (stop_point != NULL) { 3966 // The careful iteration stopped early either because it found an 3967 // uninitialized object, or because we were in the midst of an 3968 // "abortable preclean", which should now be aborted. Redirty 3969 // the bits corresponding to the partially-scanned or unscanned 3970 // cards. We'll either restart at the next block boundary or 3971 // abort the preclean. 3972 assert((_collectorState == AbortablePreclean && should_abort_preclean()), 3973 "Should only be AbortablePreclean."); 3974 _modUnionTable.mark_range(MemRegion(stop_point, dirtyRegion.end())); 3975 if (should_abort_preclean()) { 3976 break; // out of preclean loop 3977 } else { 3978 // Compute the next address at which preclean should pick up; 3979 // might need bitMapLock in order to read P-bits. 3980 lastAddr = next_card_start_after_block(stop_point); 3981 } 3982 } 3983 } else { 3984 assert(lastAddr == endAddr, "consistency check"); 3985 assert(numDirtyCards == 0, "consistency check"); 3986 break; 3987 } 3988 } 3989 verify_work_stacks_empty(); 3990 verify_overflow_empty(); 3991 return cumNumDirtyCards; 3992 } 3993 3994 // NOTE: preclean_mod_union_table() above and preclean_card_table() 3995 // below are largely identical; if you need to modify 3996 // one of these methods, please check the other method too. 3997 3998 size_t CMSCollector::preclean_card_table(ConcurrentMarkSweepGeneration* old_gen, 3999 ScanMarkedObjectsAgainCarefullyClosure* cl) { 4000 // strategy: it's similar to precleamModUnionTable above, in that 4001 // we accumulate contiguous ranges of dirty cards, mark these cards 4002 // precleaned, then scan the region covered by these cards. 4003 HeapWord* endAddr = (HeapWord*)(old_gen->_virtual_space.high()); 4004 HeapWord* startAddr = (HeapWord*)(old_gen->_virtual_space.low()); 4005 4006 cl->setFreelistLock(old_gen->freelistLock()); // needed for yielding 4007 4008 size_t numDirtyCards, cumNumDirtyCards; 4009 HeapWord *lastAddr, *nextAddr; 4010 4011 for (cumNumDirtyCards = numDirtyCards = 0, 4012 nextAddr = lastAddr = startAddr; 4013 nextAddr < endAddr; 4014 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 4015 4016 ResourceMark rm; 4017 HandleMark hm; 4018 4019 MemRegion dirtyRegion; 4020 { 4021 // See comments in "Precleaning notes" above on why we 4022 // do this locking. XXX Could the locking overheads be 4023 // too high when dirty cards are sparse? [I don't think so.] 4024 stopTimer(); 4025 CMSTokenSync x(true); // is cms thread 4026 startTimer(); 4027 sample_eden(); 4028 // Get and clear dirty region from card table 4029 dirtyRegion = _ct->ct_bs()->dirty_card_range_after_reset( 4030 MemRegion(nextAddr, endAddr), 4031 true, 4032 CardTableModRefBS::precleaned_card_val()); 4033 4034 assert(dirtyRegion.start() >= nextAddr, 4035 "returned region inconsistent?"); 4036 } 4037 lastAddr = dirtyRegion.end(); 4038 numDirtyCards = 4039 dirtyRegion.word_size()/CardTableModRefBS::card_size_in_words; 4040 4041 if (!dirtyRegion.is_empty()) { 4042 stopTimer(); 4043 CMSTokenSyncWithLocks ts(true, old_gen->freelistLock(), bitMapLock()); 4044 startTimer(); 4045 sample_eden(); 4046 verify_work_stacks_empty(); 4047 verify_overflow_empty(); 4048 HeapWord* stop_point = 4049 old_gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 4050 if (stop_point != NULL) { 4051 assert((_collectorState == AbortablePreclean && should_abort_preclean()), 4052 "Should only be AbortablePreclean."); 4053 _ct->ct_bs()->invalidate(MemRegion(stop_point, dirtyRegion.end())); 4054 if (should_abort_preclean()) { 4055 break; // out of preclean loop 4056 } else { 4057 // Compute the next address at which preclean should pick up. 4058 lastAddr = next_card_start_after_block(stop_point); 4059 } 4060 } 4061 } else { 4062 break; 4063 } 4064 } 4065 verify_work_stacks_empty(); 4066 verify_overflow_empty(); 4067 return cumNumDirtyCards; 4068 } 4069 4070 class PrecleanKlassClosure : public KlassClosure { 4071 KlassToOopClosure _cm_klass_closure; 4072 public: 4073 PrecleanKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {} 4074 void do_klass(Klass* k) { 4075 if (k->has_accumulated_modified_oops()) { 4076 k->clear_accumulated_modified_oops(); 4077 4078 _cm_klass_closure.do_klass(k); 4079 } 4080 } 4081 }; 4082 4083 // The freelist lock is needed to prevent asserts, is it really needed? 4084 void CMSCollector::preclean_klasses(MarkRefsIntoAndScanClosure* cl, Mutex* freelistLock) { 4085 4086 cl->set_freelistLock(freelistLock); 4087 4088 CMSTokenSyncWithLocks ts(true, freelistLock, bitMapLock()); 4089 4090 // SSS: Add equivalent to ScanMarkedObjectsAgainCarefullyClosure::do_yield_check and should_abort_preclean? 4091 // SSS: We should probably check if precleaning should be aborted, at suitable intervals? 4092 PrecleanKlassClosure preclean_klass_closure(cl); 4093 ClassLoaderDataGraph::classes_do(&preclean_klass_closure); 4094 4095 verify_work_stacks_empty(); 4096 verify_overflow_empty(); 4097 } 4098 4099 void CMSCollector::checkpointRootsFinal() { 4100 assert(_collectorState == FinalMarking, "incorrect state transition?"); 4101 check_correct_thread_executing(); 4102 // world is stopped at this checkpoint 4103 assert(SafepointSynchronize::is_at_safepoint(), 4104 "world should be stopped"); 4105 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 4106 4107 verify_work_stacks_empty(); 4108 verify_overflow_empty(); 4109 4110 log_debug(gc)("YG occupancy: " SIZE_FORMAT " K (" SIZE_FORMAT " K)", 4111 _young_gen->used() / K, _young_gen->capacity() / K); 4112 { 4113 if (CMSScavengeBeforeRemark) { 4114 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4115 // Temporarily set flag to false, GCH->do_collection will 4116 // expect it to be false and set to true 4117 FlagSetting fl(gch->_is_gc_active, false); 4118 4119 gch->do_collection(true, // full (i.e. force, see below) 4120 false, // !clear_all_soft_refs 4121 0, // size 4122 false, // is_tlab 4123 GenCollectedHeap::YoungGen // type 4124 ); 4125 } 4126 FreelistLocker x(this); 4127 MutexLockerEx y(bitMapLock(), 4128 Mutex::_no_safepoint_check_flag); 4129 checkpointRootsFinalWork(); 4130 } 4131 verify_work_stacks_empty(); 4132 verify_overflow_empty(); 4133 } 4134 4135 void CMSCollector::checkpointRootsFinalWork() { 4136 GCTraceTime(Trace, gc, phases) tm("checkpointRootsFinalWork", _gc_timer_cm); 4137 4138 assert(haveFreelistLocks(), "must have free list locks"); 4139 assert_lock_strong(bitMapLock()); 4140 4141 ResourceMark rm; 4142 HandleMark hm; 4143 4144 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4145 4146 if (should_unload_classes()) { 4147 CodeCache::gc_prologue(); 4148 } 4149 assert(haveFreelistLocks(), "must have free list locks"); 4150 assert_lock_strong(bitMapLock()); 4151 4152 // We might assume that we need not fill TLAB's when 4153 // CMSScavengeBeforeRemark is set, because we may have just done 4154 // a scavenge which would have filled all TLAB's -- and besides 4155 // Eden would be empty. This however may not always be the case -- 4156 // for instance although we asked for a scavenge, it may not have 4157 // happened because of a JNI critical section. We probably need 4158 // a policy for deciding whether we can in that case wait until 4159 // the critical section releases and then do the remark following 4160 // the scavenge, and skip it here. In the absence of that policy, 4161 // or of an indication of whether the scavenge did indeed occur, 4162 // we cannot rely on TLAB's having been filled and must do 4163 // so here just in case a scavenge did not happen. 4164 gch->ensure_parsability(false); // fill TLAB's, but no need to retire them 4165 // Update the saved marks which may affect the root scans. 4166 gch->save_marks(); 4167 4168 print_eden_and_survivor_chunk_arrays(); 4169 4170 { 4171 #if defined(COMPILER2) || INCLUDE_JVMCI 4172 DerivedPointerTableDeactivate dpt_deact; 4173 #endif 4174 4175 // Note on the role of the mod union table: 4176 // Since the marker in "markFromRoots" marks concurrently with 4177 // mutators, it is possible for some reachable objects not to have been 4178 // scanned. For instance, an only reference to an object A was 4179 // placed in object B after the marker scanned B. Unless B is rescanned, 4180 // A would be collected. Such updates to references in marked objects 4181 // are detected via the mod union table which is the set of all cards 4182 // dirtied since the first checkpoint in this GC cycle and prior to 4183 // the most recent young generation GC, minus those cleaned up by the 4184 // concurrent precleaning. 4185 if (CMSParallelRemarkEnabled) { 4186 GCTraceTime(Debug, gc, phases) t("Rescan (parallel)", _gc_timer_cm); 4187 do_remark_parallel(); 4188 } else { 4189 GCTraceTime(Debug, gc, phases) t("Rescan (non-parallel)", _gc_timer_cm); 4190 do_remark_non_parallel(); 4191 } 4192 } 4193 verify_work_stacks_empty(); 4194 verify_overflow_empty(); 4195 4196 { 4197 GCTraceTime(Trace, gc, phases) ts("refProcessingWork", _gc_timer_cm); 4198 refProcessingWork(); 4199 } 4200 verify_work_stacks_empty(); 4201 verify_overflow_empty(); 4202 4203 if (should_unload_classes()) { 4204 CodeCache::gc_epilogue(); 4205 } 4206 JvmtiExport::gc_epilogue(); 4207 4208 // If we encountered any (marking stack / work queue) overflow 4209 // events during the current CMS cycle, take appropriate 4210 // remedial measures, where possible, so as to try and avoid 4211 // recurrence of that condition. 4212 assert(_markStack.isEmpty(), "No grey objects"); 4213 size_t ser_ovflw = _ser_pmc_remark_ovflw + _ser_pmc_preclean_ovflw + 4214 _ser_kac_ovflw + _ser_kac_preclean_ovflw; 4215 if (ser_ovflw > 0) { 4216 log_trace(gc)("Marking stack overflow (benign) (pmc_pc=" SIZE_FORMAT ", pmc_rm=" SIZE_FORMAT ", kac=" SIZE_FORMAT ", kac_preclean=" SIZE_FORMAT ")", 4217 _ser_pmc_preclean_ovflw, _ser_pmc_remark_ovflw, _ser_kac_ovflw, _ser_kac_preclean_ovflw); 4218 _markStack.expand(); 4219 _ser_pmc_remark_ovflw = 0; 4220 _ser_pmc_preclean_ovflw = 0; 4221 _ser_kac_preclean_ovflw = 0; 4222 _ser_kac_ovflw = 0; 4223 } 4224 if (_par_pmc_remark_ovflw > 0 || _par_kac_ovflw > 0) { 4225 log_trace(gc)("Work queue overflow (benign) (pmc_rm=" SIZE_FORMAT ", kac=" SIZE_FORMAT ")", 4226 _par_pmc_remark_ovflw, _par_kac_ovflw); 4227 _par_pmc_remark_ovflw = 0; 4228 _par_kac_ovflw = 0; 4229 } 4230 if (_markStack._hit_limit > 0) { 4231 log_trace(gc)(" (benign) Hit max stack size limit (" SIZE_FORMAT ")", 4232 _markStack._hit_limit); 4233 } 4234 if (_markStack._failed_double > 0) { 4235 log_trace(gc)(" (benign) Failed stack doubling (" SIZE_FORMAT "), current capacity " SIZE_FORMAT, 4236 _markStack._failed_double, _markStack.capacity()); 4237 } 4238 _markStack._hit_limit = 0; 4239 _markStack._failed_double = 0; 4240 4241 if ((VerifyAfterGC || VerifyDuringGC) && 4242 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 4243 verify_after_remark(); 4244 } 4245 4246 _gc_tracer_cm->report_object_count_after_gc(&_is_alive_closure); 4247 4248 // Change under the freelistLocks. 4249 _collectorState = Sweeping; 4250 // Call isAllClear() under bitMapLock 4251 assert(_modUnionTable.isAllClear(), 4252 "Should be clear by end of the final marking"); 4253 assert(_ct->klass_rem_set()->mod_union_is_clear(), 4254 "Should be clear by end of the final marking"); 4255 } 4256 4257 void CMSParInitialMarkTask::work(uint worker_id) { 4258 elapsedTimer _timer; 4259 ResourceMark rm; 4260 HandleMark hm; 4261 4262 // ---------- scan from roots -------------- 4263 _timer.start(); 4264 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4265 ParMarkRefsIntoClosure par_mri_cl(_collector->_span, &(_collector->_markBitMap)); 4266 4267 // ---------- young gen roots -------------- 4268 { 4269 work_on_young_gen_roots(&par_mri_cl); 4270 _timer.stop(); 4271 log_trace(gc, task)("Finished young gen initial mark scan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4272 } 4273 4274 // ---------- remaining roots -------------- 4275 _timer.reset(); 4276 _timer.start(); 4277 4278 CLDToOopClosure cld_closure(&par_mri_cl, true); 4279 4280 gch->cms_process_roots(_strong_roots_scope, 4281 false, // yg was scanned above 4282 GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()), 4283 _collector->should_unload_classes(), 4284 &par_mri_cl, 4285 &cld_closure); 4286 assert(_collector->should_unload_classes() 4287 || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache), 4288 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 4289 _timer.stop(); 4290 log_trace(gc, task)("Finished remaining root initial mark scan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4291 } 4292 4293 // Parallel remark task 4294 class CMSParRemarkTask: public CMSParMarkTask { 4295 CompactibleFreeListSpace* _cms_space; 4296 4297 // The per-thread work queues, available here for stealing. 4298 OopTaskQueueSet* _task_queues; 4299 ParallelTaskTerminator _term; 4300 StrongRootsScope* _strong_roots_scope; 4301 4302 public: 4303 // A value of 0 passed to n_workers will cause the number of 4304 // workers to be taken from the active workers in the work gang. 4305 CMSParRemarkTask(CMSCollector* collector, 4306 CompactibleFreeListSpace* cms_space, 4307 uint n_workers, WorkGang* workers, 4308 OopTaskQueueSet* task_queues, 4309 StrongRootsScope* strong_roots_scope): 4310 CMSParMarkTask("Rescan roots and grey objects in parallel", 4311 collector, n_workers), 4312 _cms_space(cms_space), 4313 _task_queues(task_queues), 4314 _term(n_workers, task_queues), 4315 _strong_roots_scope(strong_roots_scope) { } 4316 4317 OopTaskQueueSet* task_queues() { return _task_queues; } 4318 4319 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 4320 4321 ParallelTaskTerminator* terminator() { return &_term; } 4322 uint n_workers() { return _n_workers; } 4323 4324 void work(uint worker_id); 4325 4326 private: 4327 // ... of dirty cards in old space 4328 void do_dirty_card_rescan_tasks(CompactibleFreeListSpace* sp, int i, 4329 ParMarkRefsIntoAndScanClosure* cl); 4330 4331 // ... work stealing for the above 4332 void do_work_steal(int i, ParMarkRefsIntoAndScanClosure* cl, int* seed); 4333 }; 4334 4335 class RemarkKlassClosure : public KlassClosure { 4336 KlassToOopClosure _cm_klass_closure; 4337 public: 4338 RemarkKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {} 4339 void do_klass(Klass* k) { 4340 // Check if we have modified any oops in the Klass during the concurrent marking. 4341 if (k->has_accumulated_modified_oops()) { 4342 k->clear_accumulated_modified_oops(); 4343 4344 // We could have transfered the current modified marks to the accumulated marks, 4345 // like we do with the Card Table to Mod Union Table. But it's not really necessary. 4346 } else if (k->has_modified_oops()) { 4347 // Don't clear anything, this info is needed by the next young collection. 4348 } else { 4349 // No modified oops in the Klass. 4350 return; 4351 } 4352 4353 // The klass has modified fields, need to scan the klass. 4354 _cm_klass_closure.do_klass(k); 4355 } 4356 }; 4357 4358 void CMSParMarkTask::work_on_young_gen_roots(OopsInGenClosure* cl) { 4359 ParNewGeneration* young_gen = _collector->_young_gen; 4360 ContiguousSpace* eden_space = young_gen->eden(); 4361 ContiguousSpace* from_space = young_gen->from(); 4362 ContiguousSpace* to_space = young_gen->to(); 4363 4364 HeapWord** eca = _collector->_eden_chunk_array; 4365 size_t ect = _collector->_eden_chunk_index; 4366 HeapWord** sca = _collector->_survivor_chunk_array; 4367 size_t sct = _collector->_survivor_chunk_index; 4368 4369 assert(ect <= _collector->_eden_chunk_capacity, "out of bounds"); 4370 assert(sct <= _collector->_survivor_chunk_capacity, "out of bounds"); 4371 4372 do_young_space_rescan(cl, to_space, NULL, 0); 4373 do_young_space_rescan(cl, from_space, sca, sct); 4374 do_young_space_rescan(cl, eden_space, eca, ect); 4375 } 4376 4377 // work_queue(i) is passed to the closure 4378 // ParMarkRefsIntoAndScanClosure. The "i" parameter 4379 // also is passed to do_dirty_card_rescan_tasks() and to 4380 // do_work_steal() to select the i-th task_queue. 4381 4382 void CMSParRemarkTask::work(uint worker_id) { 4383 elapsedTimer _timer; 4384 ResourceMark rm; 4385 HandleMark hm; 4386 4387 // ---------- rescan from roots -------------- 4388 _timer.start(); 4389 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4390 ParMarkRefsIntoAndScanClosure par_mrias_cl(_collector, 4391 _collector->_span, _collector->ref_processor(), 4392 &(_collector->_markBitMap), 4393 work_queue(worker_id)); 4394 4395 // Rescan young gen roots first since these are likely 4396 // coarsely partitioned and may, on that account, constitute 4397 // the critical path; thus, it's best to start off that 4398 // work first. 4399 // ---------- young gen roots -------------- 4400 { 4401 work_on_young_gen_roots(&par_mrias_cl); 4402 _timer.stop(); 4403 log_trace(gc, task)("Finished young gen rescan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4404 } 4405 4406 // ---------- remaining roots -------------- 4407 _timer.reset(); 4408 _timer.start(); 4409 gch->cms_process_roots(_strong_roots_scope, 4410 false, // yg was scanned above 4411 GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()), 4412 _collector->should_unload_classes(), 4413 &par_mrias_cl, 4414 NULL); // The dirty klasses will be handled below 4415 4416 assert(_collector->should_unload_classes() 4417 || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache), 4418 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 4419 _timer.stop(); 4420 log_trace(gc, task)("Finished remaining root rescan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4421 4422 // ---------- unhandled CLD scanning ---------- 4423 if (worker_id == 0) { // Single threaded at the moment. 4424 _timer.reset(); 4425 _timer.start(); 4426 4427 // Scan all new class loader data objects and new dependencies that were 4428 // introduced during concurrent marking. 4429 ResourceMark rm; 4430 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds(); 4431 for (int i = 0; i < array->length(); i++) { 4432 par_mrias_cl.do_cld_nv(array->at(i)); 4433 } 4434 4435 // We don't need to keep track of new CLDs anymore. 4436 ClassLoaderDataGraph::remember_new_clds(false); 4437 4438 _timer.stop(); 4439 log_trace(gc, task)("Finished unhandled CLD scanning work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4440 } 4441 4442 // ---------- dirty klass scanning ---------- 4443 if (worker_id == 0) { // Single threaded at the moment. 4444 _timer.reset(); 4445 _timer.start(); 4446 4447 // Scan all classes that was dirtied during the concurrent marking phase. 4448 RemarkKlassClosure remark_klass_closure(&par_mrias_cl); 4449 ClassLoaderDataGraph::classes_do(&remark_klass_closure); 4450 4451 _timer.stop(); 4452 log_trace(gc, task)("Finished dirty klass scanning work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4453 } 4454 4455 // We might have added oops to ClassLoaderData::_handles during the 4456 // concurrent marking phase. These oops point to newly allocated objects 4457 // that are guaranteed to be kept alive. Either by the direct allocation 4458 // code, or when the young collector processes the roots. Hence, 4459 // we don't have to revisit the _handles block during the remark phase. 4460 4461 // ---------- rescan dirty cards ------------ 4462 _timer.reset(); 4463 _timer.start(); 4464 4465 // Do the rescan tasks for each of the two spaces 4466 // (cms_space) in turn. 4467 // "worker_id" is passed to select the task_queue for "worker_id" 4468 do_dirty_card_rescan_tasks(_cms_space, worker_id, &par_mrias_cl); 4469 _timer.stop(); 4470 log_trace(gc, task)("Finished dirty card rescan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4471 4472 // ---------- steal work from other threads ... 4473 // ---------- ... and drain overflow list. 4474 _timer.reset(); 4475 _timer.start(); 4476 do_work_steal(worker_id, &par_mrias_cl, _collector->hash_seed(worker_id)); 4477 _timer.stop(); 4478 log_trace(gc, task)("Finished work stealing in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4479 } 4480 4481 void 4482 CMSParMarkTask::do_young_space_rescan( 4483 OopsInGenClosure* cl, ContiguousSpace* space, 4484 HeapWord** chunk_array, size_t chunk_top) { 4485 // Until all tasks completed: 4486 // . claim an unclaimed task 4487 // . compute region boundaries corresponding to task claimed 4488 // using chunk_array 4489 // . par_oop_iterate(cl) over that region 4490 4491 ResourceMark rm; 4492 HandleMark hm; 4493 4494 SequentialSubTasksDone* pst = space->par_seq_tasks(); 4495 4496 uint nth_task = 0; 4497 uint n_tasks = pst->n_tasks(); 4498 4499 if (n_tasks > 0) { 4500 assert(pst->valid(), "Uninitialized use?"); 4501 HeapWord *start, *end; 4502 while (!pst->is_task_claimed(/* reference */ nth_task)) { 4503 // We claimed task # nth_task; compute its boundaries. 4504 if (chunk_top == 0) { // no samples were taken 4505 assert(nth_task == 0 && n_tasks == 1, "Can have only 1 eden task"); 4506 start = space->bottom(); 4507 end = space->top(); 4508 } else if (nth_task == 0) { 4509 start = space->bottom(); 4510 end = chunk_array[nth_task]; 4511 } else if (nth_task < (uint)chunk_top) { 4512 assert(nth_task >= 1, "Control point invariant"); 4513 start = chunk_array[nth_task - 1]; 4514 end = chunk_array[nth_task]; 4515 } else { 4516 assert(nth_task == (uint)chunk_top, "Control point invariant"); 4517 start = chunk_array[chunk_top - 1]; 4518 end = space->top(); 4519 } 4520 MemRegion mr(start, end); 4521 // Verify that mr is in space 4522 assert(mr.is_empty() || space->used_region().contains(mr), 4523 "Should be in space"); 4524 // Verify that "start" is an object boundary 4525 assert(mr.is_empty() || oop(mr.start())->is_oop(), 4526 "Should be an oop"); 4527 space->par_oop_iterate(mr, cl); 4528 } 4529 pst->all_tasks_completed(); 4530 } 4531 } 4532 4533 void 4534 CMSParRemarkTask::do_dirty_card_rescan_tasks( 4535 CompactibleFreeListSpace* sp, int i, 4536 ParMarkRefsIntoAndScanClosure* cl) { 4537 // Until all tasks completed: 4538 // . claim an unclaimed task 4539 // . compute region boundaries corresponding to task claimed 4540 // . transfer dirty bits ct->mut for that region 4541 // . apply rescanclosure to dirty mut bits for that region 4542 4543 ResourceMark rm; 4544 HandleMark hm; 4545 4546 OopTaskQueue* work_q = work_queue(i); 4547 ModUnionClosure modUnionClosure(&(_collector->_modUnionTable)); 4548 // CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! 4549 // CAUTION: This closure has state that persists across calls to 4550 // the work method dirty_range_iterate_clear() in that it has 4551 // embedded in it a (subtype of) UpwardsObjectClosure. The 4552 // use of that state in the embedded UpwardsObjectClosure instance 4553 // assumes that the cards are always iterated (even if in parallel 4554 // by several threads) in monotonically increasing order per each 4555 // thread. This is true of the implementation below which picks 4556 // card ranges (chunks) in monotonically increasing order globally 4557 // and, a-fortiori, in monotonically increasing order per thread 4558 // (the latter order being a subsequence of the former). 4559 // If the work code below is ever reorganized into a more chaotic 4560 // work-partitioning form than the current "sequential tasks" 4561 // paradigm, the use of that persistent state will have to be 4562 // revisited and modified appropriately. See also related 4563 // bug 4756801 work on which should examine this code to make 4564 // sure that the changes there do not run counter to the 4565 // assumptions made here and necessary for correctness and 4566 // efficiency. Note also that this code might yield inefficient 4567 // behavior in the case of very large objects that span one or 4568 // more work chunks. Such objects would potentially be scanned 4569 // several times redundantly. Work on 4756801 should try and 4570 // address that performance anomaly if at all possible. XXX 4571 MemRegion full_span = _collector->_span; 4572 CMSBitMap* bm = &(_collector->_markBitMap); // shared 4573 MarkFromDirtyCardsClosure 4574 greyRescanClosure(_collector, full_span, // entire span of interest 4575 sp, bm, work_q, cl); 4576 4577 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 4578 assert(pst->valid(), "Uninitialized use?"); 4579 uint nth_task = 0; 4580 const int alignment = CardTableModRefBS::card_size * BitsPerWord; 4581 MemRegion span = sp->used_region(); 4582 HeapWord* start_addr = span.start(); 4583 HeapWord* end_addr = align_up(span.end(), alignment); 4584 const size_t chunk_size = sp->rescan_task_size(); // in HeapWord units 4585 assert(is_aligned(start_addr, alignment), "Check alignment"); 4586 assert(is_aligned(chunk_size, alignment), "Check alignment"); 4587 4588 while (!pst->is_task_claimed(/* reference */ nth_task)) { 4589 // Having claimed the nth_task, compute corresponding mem-region, 4590 // which is a-fortiori aligned correctly (i.e. at a MUT boundary). 4591 // The alignment restriction ensures that we do not need any 4592 // synchronization with other gang-workers while setting or 4593 // clearing bits in thus chunk of the MUT. 4594 MemRegion this_span = MemRegion(start_addr + nth_task*chunk_size, 4595 start_addr + (nth_task+1)*chunk_size); 4596 // The last chunk's end might be way beyond end of the 4597 // used region. In that case pull back appropriately. 4598 if (this_span.end() > end_addr) { 4599 this_span.set_end(end_addr); 4600 assert(!this_span.is_empty(), "Program logic (calculation of n_tasks)"); 4601 } 4602 // Iterate over the dirty cards covering this chunk, marking them 4603 // precleaned, and setting the corresponding bits in the mod union 4604 // table. Since we have been careful to partition at Card and MUT-word 4605 // boundaries no synchronization is needed between parallel threads. 4606 _collector->_ct->ct_bs()->dirty_card_iterate(this_span, 4607 &modUnionClosure); 4608 4609 // Having transferred these marks into the modUnionTable, 4610 // rescan the marked objects on the dirty cards in the modUnionTable. 4611 // Even if this is at a synchronous collection, the initial marking 4612 // may have been done during an asynchronous collection so there 4613 // may be dirty bits in the mod-union table. 4614 _collector->_modUnionTable.dirty_range_iterate_clear( 4615 this_span, &greyRescanClosure); 4616 _collector->_modUnionTable.verifyNoOneBitsInRange( 4617 this_span.start(), 4618 this_span.end()); 4619 } 4620 pst->all_tasks_completed(); // declare that i am done 4621 } 4622 4623 // . see if we can share work_queues with ParNew? XXX 4624 void 4625 CMSParRemarkTask::do_work_steal(int i, ParMarkRefsIntoAndScanClosure* cl, 4626 int* seed) { 4627 OopTaskQueue* work_q = work_queue(i); 4628 NOT_PRODUCT(int num_steals = 0;) 4629 oop obj_to_scan; 4630 CMSBitMap* bm = &(_collector->_markBitMap); 4631 4632 while (true) { 4633 // Completely finish any left over work from (an) earlier round(s) 4634 cl->trim_queue(0); 4635 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 4636 (size_t)ParGCDesiredObjsFromOverflowList); 4637 // Now check if there's any work in the overflow list 4638 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads, 4639 // only affects the number of attempts made to get work from the 4640 // overflow list and does not affect the number of workers. Just 4641 // pass ParallelGCThreads so this behavior is unchanged. 4642 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 4643 work_q, 4644 ParallelGCThreads)) { 4645 // found something in global overflow list; 4646 // not yet ready to go stealing work from others. 4647 // We'd like to assert(work_q->size() != 0, ...) 4648 // because we just took work from the overflow list, 4649 // but of course we can't since all of that could have 4650 // been already stolen from us. 4651 // "He giveth and He taketh away." 4652 continue; 4653 } 4654 // Verify that we have no work before we resort to stealing 4655 assert(work_q->size() == 0, "Have work, shouldn't steal"); 4656 // Try to steal from other queues that have work 4657 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 4658 NOT_PRODUCT(num_steals++;) 4659 assert(obj_to_scan->is_oop(), "Oops, not an oop!"); 4660 assert(bm->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 4661 // Do scanning work 4662 obj_to_scan->oop_iterate(cl); 4663 // Loop around, finish this work, and try to steal some more 4664 } else if (terminator()->offer_termination()) { 4665 break; // nirvana from the infinite cycle 4666 } 4667 } 4668 log_develop_trace(gc, task)("\t(%d: stole %d oops)", i, num_steals); 4669 assert(work_q->size() == 0 && _collector->overflow_list_is_empty(), 4670 "Else our work is not yet done"); 4671 } 4672 4673 // Record object boundaries in _eden_chunk_array by sampling the eden 4674 // top in the slow-path eden object allocation code path and record 4675 // the boundaries, if CMSEdenChunksRecordAlways is true. If 4676 // CMSEdenChunksRecordAlways is false, we use the other asynchronous 4677 // sampling in sample_eden() that activates during the part of the 4678 // preclean phase. 4679 void CMSCollector::sample_eden_chunk() { 4680 if (CMSEdenChunksRecordAlways && _eden_chunk_array != NULL) { 4681 if (_eden_chunk_lock->try_lock()) { 4682 // Record a sample. This is the critical section. The contents 4683 // of the _eden_chunk_array have to be non-decreasing in the 4684 // address order. 4685 _eden_chunk_array[_eden_chunk_index] = *_top_addr; 4686 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr, 4687 "Unexpected state of Eden"); 4688 if (_eden_chunk_index == 0 || 4689 ((_eden_chunk_array[_eden_chunk_index] > _eden_chunk_array[_eden_chunk_index-1]) && 4690 (pointer_delta(_eden_chunk_array[_eden_chunk_index], 4691 _eden_chunk_array[_eden_chunk_index-1]) >= CMSSamplingGrain))) { 4692 _eden_chunk_index++; // commit sample 4693 } 4694 _eden_chunk_lock->unlock(); 4695 } 4696 } 4697 } 4698 4699 // Return a thread-local PLAB recording array, as appropriate. 4700 void* CMSCollector::get_data_recorder(int thr_num) { 4701 if (_survivor_plab_array != NULL && 4702 (CMSPLABRecordAlways || 4703 (_collectorState > Marking && _collectorState < FinalMarking))) { 4704 assert(thr_num < (int)ParallelGCThreads, "thr_num is out of bounds"); 4705 ChunkArray* ca = &_survivor_plab_array[thr_num]; 4706 ca->reset(); // clear it so that fresh data is recorded 4707 return (void*) ca; 4708 } else { 4709 return NULL; 4710 } 4711 } 4712 4713 // Reset all the thread-local PLAB recording arrays 4714 void CMSCollector::reset_survivor_plab_arrays() { 4715 for (uint i = 0; i < ParallelGCThreads; i++) { 4716 _survivor_plab_array[i].reset(); 4717 } 4718 } 4719 4720 // Merge the per-thread plab arrays into the global survivor chunk 4721 // array which will provide the partitioning of the survivor space 4722 // for CMS initial scan and rescan. 4723 void CMSCollector::merge_survivor_plab_arrays(ContiguousSpace* surv, 4724 int no_of_gc_threads) { 4725 assert(_survivor_plab_array != NULL, "Error"); 4726 assert(_survivor_chunk_array != NULL, "Error"); 4727 assert(_collectorState == FinalMarking || 4728 (CMSParallelInitialMarkEnabled && _collectorState == InitialMarking), "Error"); 4729 for (int j = 0; j < no_of_gc_threads; j++) { 4730 _cursor[j] = 0; 4731 } 4732 HeapWord* top = surv->top(); 4733 size_t i; 4734 for (i = 0; i < _survivor_chunk_capacity; i++) { // all sca entries 4735 HeapWord* min_val = top; // Higher than any PLAB address 4736 uint min_tid = 0; // position of min_val this round 4737 for (int j = 0; j < no_of_gc_threads; j++) { 4738 ChunkArray* cur_sca = &_survivor_plab_array[j]; 4739 if (_cursor[j] == cur_sca->end()) { 4740 continue; 4741 } 4742 assert(_cursor[j] < cur_sca->end(), "ctl pt invariant"); 4743 HeapWord* cur_val = cur_sca->nth(_cursor[j]); 4744 assert(surv->used_region().contains(cur_val), "Out of bounds value"); 4745 if (cur_val < min_val) { 4746 min_tid = j; 4747 min_val = cur_val; 4748 } else { 4749 assert(cur_val < top, "All recorded addresses should be less"); 4750 } 4751 } 4752 // At this point min_val and min_tid are respectively 4753 // the least address in _survivor_plab_array[j]->nth(_cursor[j]) 4754 // and the thread (j) that witnesses that address. 4755 // We record this address in the _survivor_chunk_array[i] 4756 // and increment _cursor[min_tid] prior to the next round i. 4757 if (min_val == top) { 4758 break; 4759 } 4760 _survivor_chunk_array[i] = min_val; 4761 _cursor[min_tid]++; 4762 } 4763 // We are all done; record the size of the _survivor_chunk_array 4764 _survivor_chunk_index = i; // exclusive: [0, i) 4765 log_trace(gc, survivor)(" (Survivor:" SIZE_FORMAT "chunks) ", i); 4766 // Verify that we used up all the recorded entries 4767 #ifdef ASSERT 4768 size_t total = 0; 4769 for (int j = 0; j < no_of_gc_threads; j++) { 4770 assert(_cursor[j] == _survivor_plab_array[j].end(), "Ctl pt invariant"); 4771 total += _cursor[j]; 4772 } 4773 assert(total == _survivor_chunk_index, "Ctl Pt Invariant"); 4774 // Check that the merged array is in sorted order 4775 if (total > 0) { 4776 for (size_t i = 0; i < total - 1; i++) { 4777 log_develop_trace(gc, survivor)(" (chunk" SIZE_FORMAT ":" INTPTR_FORMAT ") ", 4778 i, p2i(_survivor_chunk_array[i])); 4779 assert(_survivor_chunk_array[i] < _survivor_chunk_array[i+1], 4780 "Not sorted"); 4781 } 4782 } 4783 #endif // ASSERT 4784 } 4785 4786 // Set up the space's par_seq_tasks structure for work claiming 4787 // for parallel initial scan and rescan of young gen. 4788 // See ParRescanTask where this is currently used. 4789 void 4790 CMSCollector:: 4791 initialize_sequential_subtasks_for_young_gen_rescan(int n_threads) { 4792 assert(n_threads > 0, "Unexpected n_threads argument"); 4793 4794 // Eden space 4795 if (!_young_gen->eden()->is_empty()) { 4796 SequentialSubTasksDone* pst = _young_gen->eden()->par_seq_tasks(); 4797 assert(!pst->valid(), "Clobbering existing data?"); 4798 // Each valid entry in [0, _eden_chunk_index) represents a task. 4799 size_t n_tasks = _eden_chunk_index + 1; 4800 assert(n_tasks == 1 || _eden_chunk_array != NULL, "Error"); 4801 // Sets the condition for completion of the subtask (how many threads 4802 // need to finish in order to be done). 4803 pst->set_n_threads(n_threads); 4804 pst->set_n_tasks((int)n_tasks); 4805 } 4806 4807 // Merge the survivor plab arrays into _survivor_chunk_array 4808 if (_survivor_plab_array != NULL) { 4809 merge_survivor_plab_arrays(_young_gen->from(), n_threads); 4810 } else { 4811 assert(_survivor_chunk_index == 0, "Error"); 4812 } 4813 4814 // To space 4815 { 4816 SequentialSubTasksDone* pst = _young_gen->to()->par_seq_tasks(); 4817 assert(!pst->valid(), "Clobbering existing data?"); 4818 // Sets the condition for completion of the subtask (how many threads 4819 // need to finish in order to be done). 4820 pst->set_n_threads(n_threads); 4821 pst->set_n_tasks(1); 4822 assert(pst->valid(), "Error"); 4823 } 4824 4825 // From space 4826 { 4827 SequentialSubTasksDone* pst = _young_gen->from()->par_seq_tasks(); 4828 assert(!pst->valid(), "Clobbering existing data?"); 4829 size_t n_tasks = _survivor_chunk_index + 1; 4830 assert(n_tasks == 1 || _survivor_chunk_array != NULL, "Error"); 4831 // Sets the condition for completion of the subtask (how many threads 4832 // need to finish in order to be done). 4833 pst->set_n_threads(n_threads); 4834 pst->set_n_tasks((int)n_tasks); 4835 assert(pst->valid(), "Error"); 4836 } 4837 } 4838 4839 // Parallel version of remark 4840 void CMSCollector::do_remark_parallel() { 4841 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4842 WorkGang* workers = gch->workers(); 4843 assert(workers != NULL, "Need parallel worker threads."); 4844 // Choose to use the number of GC workers most recently set 4845 // into "active_workers". 4846 uint n_workers = workers->active_workers(); 4847 4848 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 4849 4850 StrongRootsScope srs(n_workers); 4851 4852 CMSParRemarkTask tsk(this, cms_space, n_workers, workers, task_queues(), &srs); 4853 4854 // We won't be iterating over the cards in the card table updating 4855 // the younger_gen cards, so we shouldn't call the following else 4856 // the verification code as well as subsequent younger_refs_iterate 4857 // code would get confused. XXX 4858 // gch->rem_set()->prepare_for_younger_refs_iterate(true); // parallel 4859 4860 // The young gen rescan work will not be done as part of 4861 // process_roots (which currently doesn't know how to 4862 // parallelize such a scan), but rather will be broken up into 4863 // a set of parallel tasks (via the sampling that the [abortable] 4864 // preclean phase did of eden, plus the [two] tasks of 4865 // scanning the [two] survivor spaces. Further fine-grain 4866 // parallelization of the scanning of the survivor spaces 4867 // themselves, and of precleaning of the young gen itself 4868 // is deferred to the future. 4869 initialize_sequential_subtasks_for_young_gen_rescan(n_workers); 4870 4871 // The dirty card rescan work is broken up into a "sequence" 4872 // of parallel tasks (per constituent space) that are dynamically 4873 // claimed by the parallel threads. 4874 cms_space->initialize_sequential_subtasks_for_rescan(n_workers); 4875 4876 // It turns out that even when we're using 1 thread, doing the work in a 4877 // separate thread causes wide variance in run times. We can't help this 4878 // in the multi-threaded case, but we special-case n=1 here to get 4879 // repeatable measurements of the 1-thread overhead of the parallel code. 4880 if (n_workers > 1) { 4881 // Make refs discovery MT-safe, if it isn't already: it may not 4882 // necessarily be so, since it's possible that we are doing 4883 // ST marking. 4884 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), true); 4885 workers->run_task(&tsk); 4886 } else { 4887 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false); 4888 tsk.work(0); 4889 } 4890 4891 // restore, single-threaded for now, any preserved marks 4892 // as a result of work_q overflow 4893 restore_preserved_marks_if_any(); 4894 } 4895 4896 // Non-parallel version of remark 4897 void CMSCollector::do_remark_non_parallel() { 4898 ResourceMark rm; 4899 HandleMark hm; 4900 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4901 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false); 4902 4903 MarkRefsIntoAndScanClosure 4904 mrias_cl(_span, ref_processor(), &_markBitMap, NULL /* not precleaning */, 4905 &_markStack, this, 4906 false /* should_yield */, false /* not precleaning */); 4907 MarkFromDirtyCardsClosure 4908 markFromDirtyCardsClosure(this, _span, 4909 NULL, // space is set further below 4910 &_markBitMap, &_markStack, &mrias_cl); 4911 { 4912 GCTraceTime(Trace, gc, phases) t("Grey Object Rescan", _gc_timer_cm); 4913 // Iterate over the dirty cards, setting the corresponding bits in the 4914 // mod union table. 4915 { 4916 ModUnionClosure modUnionClosure(&_modUnionTable); 4917 _ct->ct_bs()->dirty_card_iterate( 4918 _cmsGen->used_region(), 4919 &modUnionClosure); 4920 } 4921 // Having transferred these marks into the modUnionTable, we just need 4922 // to rescan the marked objects on the dirty cards in the modUnionTable. 4923 // The initial marking may have been done during an asynchronous 4924 // collection so there may be dirty bits in the mod-union table. 4925 const int alignment = 4926 CardTableModRefBS::card_size * BitsPerWord; 4927 { 4928 // ... First handle dirty cards in CMS gen 4929 markFromDirtyCardsClosure.set_space(_cmsGen->cmsSpace()); 4930 MemRegion ur = _cmsGen->used_region(); 4931 HeapWord* lb = ur.start(); 4932 HeapWord* ub = align_up(ur.end(), alignment); 4933 MemRegion cms_span(lb, ub); 4934 _modUnionTable.dirty_range_iterate_clear(cms_span, 4935 &markFromDirtyCardsClosure); 4936 verify_work_stacks_empty(); 4937 log_trace(gc)(" (re-scanned " SIZE_FORMAT " dirty cards in cms gen) ", markFromDirtyCardsClosure.num_dirty_cards()); 4938 } 4939 } 4940 if (VerifyDuringGC && 4941 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 4942 HandleMark hm; // Discard invalid handles created during verification 4943 Universe::verify(); 4944 } 4945 { 4946 GCTraceTime(Trace, gc, phases) t("Root Rescan", _gc_timer_cm); 4947 4948 verify_work_stacks_empty(); 4949 4950 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 4951 StrongRootsScope srs(1); 4952 4953 gch->cms_process_roots(&srs, 4954 true, // young gen as roots 4955 GenCollectedHeap::ScanningOption(roots_scanning_options()), 4956 should_unload_classes(), 4957 &mrias_cl, 4958 NULL); // The dirty klasses will be handled below 4959 4960 assert(should_unload_classes() 4961 || (roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache), 4962 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 4963 } 4964 4965 { 4966 GCTraceTime(Trace, gc, phases) t("Visit Unhandled CLDs", _gc_timer_cm); 4967 4968 verify_work_stacks_empty(); 4969 4970 // Scan all class loader data objects that might have been introduced 4971 // during concurrent marking. 4972 ResourceMark rm; 4973 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds(); 4974 for (int i = 0; i < array->length(); i++) { 4975 mrias_cl.do_cld_nv(array->at(i)); 4976 } 4977 4978 // We don't need to keep track of new CLDs anymore. 4979 ClassLoaderDataGraph::remember_new_clds(false); 4980 4981 verify_work_stacks_empty(); 4982 } 4983 4984 { 4985 GCTraceTime(Trace, gc, phases) t("Dirty Klass Scan", _gc_timer_cm); 4986 4987 verify_work_stacks_empty(); 4988 4989 RemarkKlassClosure remark_klass_closure(&mrias_cl); 4990 ClassLoaderDataGraph::classes_do(&remark_klass_closure); 4991 4992 verify_work_stacks_empty(); 4993 } 4994 4995 // We might have added oops to ClassLoaderData::_handles during the 4996 // concurrent marking phase. These oops point to newly allocated objects 4997 // that are guaranteed to be kept alive. Either by the direct allocation 4998 // code, or when the young collector processes the roots. Hence, 4999 // we don't have to revisit the _handles block during the remark phase. 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(obj_to_scan->is_oop(), "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 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5153 WorkGang* workers = gch->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 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5166 WorkGang* workers = gch->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 { 5189 GCTraceTime(Debug, gc, phases) t("Reference Processing", _gc_timer_cm); 5190 5191 ReferenceProcessorStats stats; 5192 if (rp->processing_is_mt()) { 5193 // Set the degree of MT here. If the discovery is done MT, there 5194 // may have been a different number of threads doing the discovery 5195 // and a different number of discovered lists may have Ref objects. 5196 // That is OK as long as the Reference lists are balanced (see 5197 // balance_all_queues() and balance_queues()). 5198 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5199 uint active_workers = ParallelGCThreads; 5200 WorkGang* workers = gch->workers(); 5201 if (workers != NULL) { 5202 active_workers = workers->active_workers(); 5203 // The expectation is that active_workers will have already 5204 // been set to a reasonable value. If it has not been set, 5205 // investigate. 5206 assert(active_workers > 0, "Should have been set during scavenge"); 5207 } 5208 rp->set_active_mt_degree(active_workers); 5209 CMSRefProcTaskExecutor task_executor(*this); 5210 stats = rp->process_discovered_references(&_is_alive_closure, 5211 &cmsKeepAliveClosure, 5212 &cmsDrainMarkingStackClosure, 5213 &task_executor, 5214 _gc_timer_cm); 5215 } else { 5216 stats = rp->process_discovered_references(&_is_alive_closure, 5217 &cmsKeepAliveClosure, 5218 &cmsDrainMarkingStackClosure, 5219 NULL, 5220 _gc_timer_cm); 5221 } 5222 _gc_tracer_cm->report_gc_reference_stats(stats); 5223 5224 } 5225 5226 // This is the point where the entire marking should have completed. 5227 verify_work_stacks_empty(); 5228 5229 if (should_unload_classes()) { 5230 { 5231 GCTraceTime(Debug, gc, phases) t("Class Unloading", _gc_timer_cm); 5232 5233 // Unload classes and purge the SystemDictionary. 5234 bool purged_class = SystemDictionary::do_unloading(&_is_alive_closure, _gc_timer_cm); 5235 5236 // Unload nmethods. 5237 CodeCache::do_unloading(&_is_alive_closure, purged_class); 5238 5239 // Prune dead klasses from subklass/sibling/implementor lists. 5240 Klass::clean_weak_klass_links(&_is_alive_closure); 5241 } 5242 5243 { 5244 GCTraceTime(Debug, gc, phases) t("Scrub Symbol Table", _gc_timer_cm); 5245 // Clean up unreferenced symbols in symbol table. 5246 SymbolTable::unlink(); 5247 } 5248 5249 { 5250 GCTraceTime(Debug, gc, phases) t("Scrub String Table", _gc_timer_cm); 5251 // Delete entries for dead interned strings. 5252 StringTable::unlink(&_is_alive_closure); 5253 } 5254 } 5255 5256 // Restore any preserved marks as a result of mark stack or 5257 // work queue overflow 5258 restore_preserved_marks_if_any(); // done single-threaded for now 5259 5260 rp->set_enqueuing_is_done(true); 5261 if (rp->processing_is_mt()) { 5262 rp->balance_all_queues(); 5263 CMSRefProcTaskExecutor task_executor(*this); 5264 rp->enqueue_discovered_references(&task_executor); 5265 } else { 5266 rp->enqueue_discovered_references(NULL); 5267 } 5268 rp->verify_no_references_recorded(); 5269 assert(!rp->discovery_enabled(), "should have been disabled"); 5270 } 5271 5272 #ifndef PRODUCT 5273 void CMSCollector::check_correct_thread_executing() { 5274 Thread* t = Thread::current(); 5275 // Only the VM thread or the CMS thread should be here. 5276 assert(t->is_ConcurrentGC_thread() || t->is_VM_thread(), 5277 "Unexpected thread type"); 5278 // If this is the vm thread, the foreground process 5279 // should not be waiting. Note that _foregroundGCIsActive is 5280 // true while the foreground collector is waiting. 5281 if (_foregroundGCShouldWait) { 5282 // We cannot be the VM thread 5283 assert(t->is_ConcurrentGC_thread(), 5284 "Should be CMS thread"); 5285 } else { 5286 // We can be the CMS thread only if we are in a stop-world 5287 // phase of CMS collection. 5288 if (t->is_ConcurrentGC_thread()) { 5289 assert(_collectorState == InitialMarking || 5290 _collectorState == FinalMarking, 5291 "Should be a stop-world phase"); 5292 // The CMS thread should be holding the CMS_token. 5293 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 5294 "Potential interference with concurrently " 5295 "executing VM thread"); 5296 } 5297 } 5298 } 5299 #endif 5300 5301 void CMSCollector::sweep() { 5302 assert(_collectorState == Sweeping, "just checking"); 5303 check_correct_thread_executing(); 5304 verify_work_stacks_empty(); 5305 verify_overflow_empty(); 5306 increment_sweep_count(); 5307 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 5308 5309 _inter_sweep_timer.stop(); 5310 _inter_sweep_estimate.sample(_inter_sweep_timer.seconds()); 5311 5312 assert(!_intra_sweep_timer.is_active(), "Should not be active"); 5313 _intra_sweep_timer.reset(); 5314 _intra_sweep_timer.start(); 5315 { 5316 GCTraceCPUTime tcpu; 5317 CMSPhaseAccounting pa(this, "Concurrent Sweep"); 5318 // First sweep the old gen 5319 { 5320 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(), 5321 bitMapLock()); 5322 sweepWork(_cmsGen); 5323 } 5324 5325 // Update Universe::_heap_*_at_gc figures. 5326 // We need all the free list locks to make the abstract state 5327 // transition from Sweeping to Resetting. See detailed note 5328 // further below. 5329 { 5330 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock()); 5331 // Update heap occupancy information which is used as 5332 // input to soft ref clearing policy at the next gc. 5333 Universe::update_heap_info_at_gc(); 5334 _collectorState = Resizing; 5335 } 5336 } 5337 verify_work_stacks_empty(); 5338 verify_overflow_empty(); 5339 5340 if (should_unload_classes()) { 5341 // Delay purge to the beginning of the next safepoint. Metaspace::contains 5342 // requires that the virtual spaces are stable and not deleted. 5343 ClassLoaderDataGraph::set_should_purge(true); 5344 } 5345 5346 _intra_sweep_timer.stop(); 5347 _intra_sweep_estimate.sample(_intra_sweep_timer.seconds()); 5348 5349 _inter_sweep_timer.reset(); 5350 _inter_sweep_timer.start(); 5351 5352 // We need to use a monotonically non-decreasing time in ms 5353 // or we will see time-warp warnings and os::javaTimeMillis() 5354 // does not guarantee monotonicity. 5355 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC; 5356 update_time_of_last_gc(now); 5357 5358 // NOTE on abstract state transitions: 5359 // Mutators allocate-live and/or mark the mod-union table dirty 5360 // based on the state of the collection. The former is done in 5361 // the interval [Marking, Sweeping] and the latter in the interval 5362 // [Marking, Sweeping). Thus the transitions into the Marking state 5363 // and out of the Sweeping state must be synchronously visible 5364 // globally to the mutators. 5365 // The transition into the Marking state happens with the world 5366 // stopped so the mutators will globally see it. Sweeping is 5367 // done asynchronously by the background collector so the transition 5368 // from the Sweeping state to the Resizing state must be done 5369 // under the freelistLock (as is the check for whether to 5370 // allocate-live and whether to dirty the mod-union table). 5371 assert(_collectorState == Resizing, "Change of collector state to" 5372 " Resizing must be done under the freelistLocks (plural)"); 5373 5374 // Now that sweeping has been completed, we clear 5375 // the incremental_collection_failed flag, 5376 // thus inviting a younger gen collection to promote into 5377 // this generation. If such a promotion may still fail, 5378 // the flag will be set again when a young collection is 5379 // attempted. 5380 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5381 gch->clear_incremental_collection_failed(); // Worth retrying as fresh space may have been freed up 5382 gch->update_full_collections_completed(_collection_count_start); 5383 } 5384 5385 // FIX ME!!! Looks like this belongs in CFLSpace, with 5386 // CMSGen merely delegating to it. 5387 void ConcurrentMarkSweepGeneration::setNearLargestChunk() { 5388 double nearLargestPercent = FLSLargestBlockCoalesceProximity; 5389 HeapWord* minAddr = _cmsSpace->bottom(); 5390 HeapWord* largestAddr = 5391 (HeapWord*) _cmsSpace->dictionary()->find_largest_dict(); 5392 if (largestAddr == NULL) { 5393 // The dictionary appears to be empty. In this case 5394 // try to coalesce at the end of the heap. 5395 largestAddr = _cmsSpace->end(); 5396 } 5397 size_t largestOffset = pointer_delta(largestAddr, minAddr); 5398 size_t nearLargestOffset = 5399 (size_t)((double)largestOffset * nearLargestPercent) - MinChunkSize; 5400 log_debug(gc, freelist)("CMS: Large Block: " PTR_FORMAT "; Proximity: " PTR_FORMAT " -> " PTR_FORMAT, 5401 p2i(largestAddr), p2i(_cmsSpace->nearLargestChunk()), p2i(minAddr + nearLargestOffset)); 5402 _cmsSpace->set_nearLargestChunk(minAddr + nearLargestOffset); 5403 } 5404 5405 bool ConcurrentMarkSweepGeneration::isNearLargestChunk(HeapWord* addr) { 5406 return addr >= _cmsSpace->nearLargestChunk(); 5407 } 5408 5409 FreeChunk* ConcurrentMarkSweepGeneration::find_chunk_at_end() { 5410 return _cmsSpace->find_chunk_at_end(); 5411 } 5412 5413 void ConcurrentMarkSweepGeneration::update_gc_stats(Generation* current_generation, 5414 bool full) { 5415 // If the young generation has been collected, gather any statistics 5416 // that are of interest at this point. 5417 bool current_is_young = GenCollectedHeap::heap()->is_young_gen(current_generation); 5418 if (!full && current_is_young) { 5419 // Gather statistics on the young generation collection. 5420 collector()->stats().record_gc0_end(used()); 5421 } 5422 } 5423 5424 void CMSCollector::sweepWork(ConcurrentMarkSweepGeneration* old_gen) { 5425 // We iterate over the space(s) underlying this generation, 5426 // checking the mark bit map to see if the bits corresponding 5427 // to specific blocks are marked or not. Blocks that are 5428 // marked are live and are not swept up. All remaining blocks 5429 // are swept up, with coalescing on-the-fly as we sweep up 5430 // contiguous free and/or garbage blocks: 5431 // We need to ensure that the sweeper synchronizes with allocators 5432 // and stop-the-world collectors. In particular, the following 5433 // locks are used: 5434 // . CMS token: if this is held, a stop the world collection cannot occur 5435 // . freelistLock: if this is held no allocation can occur from this 5436 // generation by another thread 5437 // . bitMapLock: if this is held, no other thread can access or update 5438 // 5439 5440 // Note that we need to hold the freelistLock if we use 5441 // block iterate below; else the iterator might go awry if 5442 // a mutator (or promotion) causes block contents to change 5443 // (for instance if the allocator divvies up a block). 5444 // If we hold the free list lock, for all practical purposes 5445 // young generation GC's can't occur (they'll usually need to 5446 // promote), so we might as well prevent all young generation 5447 // GC's while we do a sweeping step. For the same reason, we might 5448 // as well take the bit map lock for the entire duration 5449 5450 // check that we hold the requisite locks 5451 assert(have_cms_token(), "Should hold cms token"); 5452 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), "Should possess CMS token to sweep"); 5453 assert_lock_strong(old_gen->freelistLock()); 5454 assert_lock_strong(bitMapLock()); 5455 5456 assert(!_inter_sweep_timer.is_active(), "Was switched off in an outer context"); 5457 assert(_intra_sweep_timer.is_active(), "Was switched on in an outer context"); 5458 old_gen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()), 5459 _inter_sweep_estimate.padded_average(), 5460 _intra_sweep_estimate.padded_average()); 5461 old_gen->setNearLargestChunk(); 5462 5463 { 5464 SweepClosure sweepClosure(this, old_gen, &_markBitMap, CMSYield); 5465 old_gen->cmsSpace()->blk_iterate_careful(&sweepClosure); 5466 // We need to free-up/coalesce garbage/blocks from a 5467 // co-terminal free run. This is done in the SweepClosure 5468 // destructor; so, do not remove this scope, else the 5469 // end-of-sweep-census below will be off by a little bit. 5470 } 5471 old_gen->cmsSpace()->sweep_completed(); 5472 old_gen->cmsSpace()->endSweepFLCensus(sweep_count()); 5473 if (should_unload_classes()) { // unloaded classes this cycle, 5474 _concurrent_cycles_since_last_unload = 0; // ... reset count 5475 } else { // did not unload classes, 5476 _concurrent_cycles_since_last_unload++; // ... increment count 5477 } 5478 } 5479 5480 // Reset CMS data structures (for now just the marking bit map) 5481 // preparatory for the next cycle. 5482 void CMSCollector::reset_concurrent() { 5483 CMSTokenSyncWithLocks ts(true, bitMapLock()); 5484 5485 // If the state is not "Resetting", the foreground thread 5486 // has done a collection and the resetting. 5487 if (_collectorState != Resetting) { 5488 assert(_collectorState == Idling, "The state should only change" 5489 " because the foreground collector has finished the collection"); 5490 return; 5491 } 5492 5493 { 5494 // Clear the mark bitmap (no grey objects to start with) 5495 // for the next cycle. 5496 GCTraceCPUTime tcpu; 5497 CMSPhaseAccounting cmspa(this, "Concurrent Reset"); 5498 5499 HeapWord* curAddr = _markBitMap.startWord(); 5500 while (curAddr < _markBitMap.endWord()) { 5501 size_t remaining = pointer_delta(_markBitMap.endWord(), curAddr); 5502 MemRegion chunk(curAddr, MIN2(CMSBitMapYieldQuantum, remaining)); 5503 _markBitMap.clear_large_range(chunk); 5504 if (ConcurrentMarkSweepThread::should_yield() && 5505 !foregroundGCIsActive() && 5506 CMSYield) { 5507 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 5508 "CMS thread should hold CMS token"); 5509 assert_lock_strong(bitMapLock()); 5510 bitMapLock()->unlock(); 5511 ConcurrentMarkSweepThread::desynchronize(true); 5512 stopTimer(); 5513 incrementYields(); 5514 5515 // See the comment in coordinator_yield() 5516 for (unsigned i = 0; i < CMSYieldSleepCount && 5517 ConcurrentMarkSweepThread::should_yield() && 5518 !CMSCollector::foregroundGCIsActive(); ++i) { 5519 os::sleep(Thread::current(), 1, false); 5520 } 5521 5522 ConcurrentMarkSweepThread::synchronize(true); 5523 bitMapLock()->lock_without_safepoint_check(); 5524 startTimer(); 5525 } 5526 curAddr = chunk.end(); 5527 } 5528 // A successful mostly concurrent collection has been done. 5529 // Because only the full (i.e., concurrent mode failure) collections 5530 // are being measured for gc overhead limits, clean the "near" flag 5531 // and count. 5532 size_policy()->reset_gc_overhead_limit_count(); 5533 _collectorState = Idling; 5534 } 5535 5536 register_gc_end(); 5537 } 5538 5539 // Same as above but for STW paths 5540 void CMSCollector::reset_stw() { 5541 // already have the lock 5542 assert(_collectorState == Resetting, "just checking"); 5543 assert_lock_strong(bitMapLock()); 5544 GCIdMarkAndRestore gc_id_mark(_cmsThread->gc_id()); 5545 _markBitMap.clear_all(); 5546 _collectorState = Idling; 5547 register_gc_end(); 5548 } 5549 5550 void CMSCollector::do_CMS_operation(CMS_op_type op, GCCause::Cause gc_cause) { 5551 GCTraceCPUTime tcpu; 5552 TraceCollectorStats tcs(counters()); 5553 5554 switch (op) { 5555 case CMS_op_checkpointRootsInitial: { 5556 GCTraceTime(Info, gc) t("Pause Initial Mark", NULL, GCCause::_no_gc, true); 5557 SvcGCMarker sgcm(SvcGCMarker::OTHER); 5558 checkpointRootsInitial(); 5559 break; 5560 } 5561 case CMS_op_checkpointRootsFinal: { 5562 GCTraceTime(Info, gc) t("Pause Remark", NULL, GCCause::_no_gc, true); 5563 SvcGCMarker sgcm(SvcGCMarker::OTHER); 5564 checkpointRootsFinal(); 5565 break; 5566 } 5567 default: 5568 fatal("No such CMS_op"); 5569 } 5570 } 5571 5572 #ifndef PRODUCT 5573 size_t const CMSCollector::skip_header_HeapWords() { 5574 return FreeChunk::header_size(); 5575 } 5576 5577 // Try and collect here conditions that should hold when 5578 // CMS thread is exiting. The idea is that the foreground GC 5579 // thread should not be blocked if it wants to terminate 5580 // the CMS thread and yet continue to run the VM for a while 5581 // after that. 5582 void CMSCollector::verify_ok_to_terminate() const { 5583 assert(Thread::current()->is_ConcurrentGC_thread(), 5584 "should be called by CMS thread"); 5585 assert(!_foregroundGCShouldWait, "should be false"); 5586 // We could check here that all the various low-level locks 5587 // are not held by the CMS thread, but that is overkill; see 5588 // also CMSThread::verify_ok_to_terminate() where the CGC_lock 5589 // is checked. 5590 } 5591 #endif 5592 5593 size_t CMSCollector::block_size_using_printezis_bits(HeapWord* addr) const { 5594 assert(_markBitMap.isMarked(addr) && _markBitMap.isMarked(addr + 1), 5595 "missing Printezis mark?"); 5596 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 5597 size_t size = pointer_delta(nextOneAddr + 1, addr); 5598 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 5599 "alignment problem"); 5600 assert(size >= 3, "Necessary for Printezis marks to work"); 5601 return size; 5602 } 5603 5604 // A variant of the above (block_size_using_printezis_bits()) except 5605 // that we return 0 if the P-bits are not yet set. 5606 size_t CMSCollector::block_size_if_printezis_bits(HeapWord* addr) const { 5607 if (_markBitMap.isMarked(addr + 1)) { 5608 assert(_markBitMap.isMarked(addr), "P-bit can be set only for marked objects"); 5609 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 5610 size_t size = pointer_delta(nextOneAddr + 1, addr); 5611 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 5612 "alignment problem"); 5613 assert(size >= 3, "Necessary for Printezis marks to work"); 5614 return size; 5615 } 5616 return 0; 5617 } 5618 5619 HeapWord* CMSCollector::next_card_start_after_block(HeapWord* addr) const { 5620 size_t sz = 0; 5621 oop p = (oop)addr; 5622 if (p->klass_or_null_acquire() != NULL) { 5623 sz = CompactibleFreeListSpace::adjustObjectSize(p->size()); 5624 } else { 5625 sz = block_size_using_printezis_bits(addr); 5626 } 5627 assert(sz > 0, "size must be nonzero"); 5628 HeapWord* next_block = addr + sz; 5629 HeapWord* next_card = align_up(next_block, CardTableModRefBS::card_size); 5630 assert(align_down((uintptr_t)addr, CardTableModRefBS::card_size) < 5631 align_down((uintptr_t)next_card, CardTableModRefBS::card_size), 5632 "must be different cards"); 5633 return next_card; 5634 } 5635 5636 5637 // CMS Bit Map Wrapper ///////////////////////////////////////// 5638 5639 // Construct a CMS bit map infrastructure, but don't create the 5640 // bit vector itself. That is done by a separate call CMSBitMap::allocate() 5641 // further below. 5642 CMSBitMap::CMSBitMap(int shifter, int mutex_rank, const char* mutex_name): 5643 _bm(), 5644 _shifter(shifter), 5645 _lock(mutex_rank >= 0 ? new Mutex(mutex_rank, mutex_name, true, 5646 Monitor::_safepoint_check_sometimes) : NULL) 5647 { 5648 _bmStartWord = 0; 5649 _bmWordSize = 0; 5650 } 5651 5652 bool CMSBitMap::allocate(MemRegion mr) { 5653 _bmStartWord = mr.start(); 5654 _bmWordSize = mr.word_size(); 5655 ReservedSpace brs(ReservedSpace::allocation_align_size_up( 5656 (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1)); 5657 if (!brs.is_reserved()) { 5658 log_warning(gc)("CMS bit map allocation failure"); 5659 return false; 5660 } 5661 // For now we'll just commit all of the bit map up front. 5662 // Later on we'll try to be more parsimonious with swap. 5663 if (!_virtual_space.initialize(brs, brs.size())) { 5664 log_warning(gc)("CMS bit map backing store failure"); 5665 return false; 5666 } 5667 assert(_virtual_space.committed_size() == brs.size(), 5668 "didn't reserve backing store for all of CMS bit map?"); 5669 assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >= 5670 _bmWordSize, "inconsistency in bit map sizing"); 5671 _bm = BitMapView((BitMap::bm_word_t*)_virtual_space.low(), _bmWordSize >> _shifter); 5672 5673 // bm.clear(); // can we rely on getting zero'd memory? verify below 5674 assert(isAllClear(), 5675 "Expected zero'd memory from ReservedSpace constructor"); 5676 assert(_bm.size() == heapWordDiffToOffsetDiff(sizeInWords()), 5677 "consistency check"); 5678 return true; 5679 } 5680 5681 void CMSBitMap::dirty_range_iterate_clear(MemRegion mr, MemRegionClosure* cl) { 5682 HeapWord *next_addr, *end_addr, *last_addr; 5683 assert_locked(); 5684 assert(covers(mr), "out-of-range error"); 5685 // XXX assert that start and end are appropriately aligned 5686 for (next_addr = mr.start(), end_addr = mr.end(); 5687 next_addr < end_addr; next_addr = last_addr) { 5688 MemRegion dirty_region = getAndClearMarkedRegion(next_addr, end_addr); 5689 last_addr = dirty_region.end(); 5690 if (!dirty_region.is_empty()) { 5691 cl->do_MemRegion(dirty_region); 5692 } else { 5693 assert(last_addr == end_addr, "program logic"); 5694 return; 5695 } 5696 } 5697 } 5698 5699 void CMSBitMap::print_on_error(outputStream* st, const char* prefix) const { 5700 _bm.print_on_error(st, prefix); 5701 } 5702 5703 #ifndef PRODUCT 5704 void CMSBitMap::assert_locked() const { 5705 CMSLockVerifier::assert_locked(lock()); 5706 } 5707 5708 bool CMSBitMap::covers(MemRegion mr) const { 5709 // assert(_bm.map() == _virtual_space.low(), "map inconsistency"); 5710 assert((size_t)_bm.size() == (_bmWordSize >> _shifter), 5711 "size inconsistency"); 5712 return (mr.start() >= _bmStartWord) && 5713 (mr.end() <= endWord()); 5714 } 5715 5716 bool CMSBitMap::covers(HeapWord* start, size_t size) const { 5717 return (start >= _bmStartWord && (start + size) <= endWord()); 5718 } 5719 5720 void CMSBitMap::verifyNoOneBitsInRange(HeapWord* left, HeapWord* right) { 5721 // verify that there are no 1 bits in the interval [left, right) 5722 FalseBitMapClosure falseBitMapClosure; 5723 iterate(&falseBitMapClosure, left, right); 5724 } 5725 5726 void CMSBitMap::region_invariant(MemRegion mr) 5727 { 5728 assert_locked(); 5729 // mr = mr.intersection(MemRegion(_bmStartWord, _bmWordSize)); 5730 assert(!mr.is_empty(), "unexpected empty region"); 5731 assert(covers(mr), "mr should be covered by bit map"); 5732 // convert address range into offset range 5733 size_t start_ofs = heapWordToOffset(mr.start()); 5734 // Make sure that end() is appropriately aligned 5735 assert(mr.end() == align_up(mr.end(), (1 << (_shifter+LogHeapWordSize))), 5736 "Misaligned mr.end()"); 5737 size_t end_ofs = heapWordToOffset(mr.end()); 5738 assert(end_ofs > start_ofs, "Should mark at least one bit"); 5739 } 5740 5741 #endif 5742 5743 bool CMSMarkStack::allocate(size_t size) { 5744 // allocate a stack of the requisite depth 5745 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 5746 size * sizeof(oop))); 5747 if (!rs.is_reserved()) { 5748 log_warning(gc)("CMSMarkStack allocation failure"); 5749 return false; 5750 } 5751 if (!_virtual_space.initialize(rs, rs.size())) { 5752 log_warning(gc)("CMSMarkStack backing store failure"); 5753 return false; 5754 } 5755 assert(_virtual_space.committed_size() == rs.size(), 5756 "didn't reserve backing store for all of CMS stack?"); 5757 _base = (oop*)(_virtual_space.low()); 5758 _index = 0; 5759 _capacity = size; 5760 NOT_PRODUCT(_max_depth = 0); 5761 return true; 5762 } 5763 5764 // XXX FIX ME !!! In the MT case we come in here holding a 5765 // leaf lock. For printing we need to take a further lock 5766 // which has lower rank. We need to recalibrate the two 5767 // lock-ranks involved in order to be able to print the 5768 // messages below. (Or defer the printing to the caller. 5769 // For now we take the expedient path of just disabling the 5770 // messages for the problematic case.) 5771 void CMSMarkStack::expand() { 5772 assert(_capacity <= MarkStackSizeMax, "stack bigger than permitted"); 5773 if (_capacity == MarkStackSizeMax) { 5774 if (_hit_limit++ == 0 && !CMSConcurrentMTEnabled) { 5775 // We print a warning message only once per CMS cycle. 5776 log_debug(gc)(" (benign) Hit CMSMarkStack max size limit"); 5777 } 5778 return; 5779 } 5780 // Double capacity if possible 5781 size_t new_capacity = MIN2(_capacity*2, MarkStackSizeMax); 5782 // Do not give up existing stack until we have managed to 5783 // get the double capacity that we desired. 5784 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 5785 new_capacity * sizeof(oop))); 5786 if (rs.is_reserved()) { 5787 // Release the backing store associated with old stack 5788 _virtual_space.release(); 5789 // Reinitialize virtual space for new stack 5790 if (!_virtual_space.initialize(rs, rs.size())) { 5791 fatal("Not enough swap for expanded marking stack"); 5792 } 5793 _base = (oop*)(_virtual_space.low()); 5794 _index = 0; 5795 _capacity = new_capacity; 5796 } else if (_failed_double++ == 0 && !CMSConcurrentMTEnabled) { 5797 // Failed to double capacity, continue; 5798 // we print a detail message only once per CMS cycle. 5799 log_debug(gc)(" (benign) Failed to expand marking stack from " SIZE_FORMAT "K to " SIZE_FORMAT "K", 5800 _capacity / K, new_capacity / K); 5801 } 5802 } 5803 5804 5805 // Closures 5806 // XXX: there seems to be a lot of code duplication here; 5807 // should refactor and consolidate common code. 5808 5809 // This closure is used to mark refs into the CMS generation in 5810 // the CMS bit map. Called at the first checkpoint. This closure 5811 // assumes that we do not need to re-mark dirty cards; if the CMS 5812 // generation on which this is used is not an oldest 5813 // generation then this will lose younger_gen cards! 5814 5815 MarkRefsIntoClosure::MarkRefsIntoClosure( 5816 MemRegion span, CMSBitMap* bitMap): 5817 _span(span), 5818 _bitMap(bitMap) 5819 { 5820 assert(ref_processor() == NULL, "deliberately left NULL"); 5821 assert(_bitMap->covers(_span), "_bitMap/_span mismatch"); 5822 } 5823 5824 void MarkRefsIntoClosure::do_oop(oop obj) { 5825 // if p points into _span, then mark corresponding bit in _markBitMap 5826 assert(obj->is_oop(), "expected an oop"); 5827 HeapWord* addr = (HeapWord*)obj; 5828 if (_span.contains(addr)) { 5829 // this should be made more efficient 5830 _bitMap->mark(addr); 5831 } 5832 } 5833 5834 void MarkRefsIntoClosure::do_oop(oop* p) { MarkRefsIntoClosure::do_oop_work(p); } 5835 void MarkRefsIntoClosure::do_oop(narrowOop* p) { MarkRefsIntoClosure::do_oop_work(p); } 5836 5837 ParMarkRefsIntoClosure::ParMarkRefsIntoClosure( 5838 MemRegion span, CMSBitMap* bitMap): 5839 _span(span), 5840 _bitMap(bitMap) 5841 { 5842 assert(ref_processor() == NULL, "deliberately left NULL"); 5843 assert(_bitMap->covers(_span), "_bitMap/_span mismatch"); 5844 } 5845 5846 void ParMarkRefsIntoClosure::do_oop(oop obj) { 5847 // if p points into _span, then mark corresponding bit in _markBitMap 5848 assert(obj->is_oop(), "expected an oop"); 5849 HeapWord* addr = (HeapWord*)obj; 5850 if (_span.contains(addr)) { 5851 // this should be made more efficient 5852 _bitMap->par_mark(addr); 5853 } 5854 } 5855 5856 void ParMarkRefsIntoClosure::do_oop(oop* p) { ParMarkRefsIntoClosure::do_oop_work(p); } 5857 void ParMarkRefsIntoClosure::do_oop(narrowOop* p) { ParMarkRefsIntoClosure::do_oop_work(p); } 5858 5859 // A variant of the above, used for CMS marking verification. 5860 MarkRefsIntoVerifyClosure::MarkRefsIntoVerifyClosure( 5861 MemRegion span, CMSBitMap* verification_bm, CMSBitMap* cms_bm): 5862 _span(span), 5863 _verification_bm(verification_bm), 5864 _cms_bm(cms_bm) 5865 { 5866 assert(ref_processor() == NULL, "deliberately left NULL"); 5867 assert(_verification_bm->covers(_span), "_verification_bm/_span mismatch"); 5868 } 5869 5870 void MarkRefsIntoVerifyClosure::do_oop(oop obj) { 5871 // if p points into _span, then mark corresponding bit in _markBitMap 5872 assert(obj->is_oop(), "expected an oop"); 5873 HeapWord* addr = (HeapWord*)obj; 5874 if (_span.contains(addr)) { 5875 _verification_bm->mark(addr); 5876 if (!_cms_bm->isMarked(addr)) { 5877 Log(gc, verify) log; 5878 ResourceMark rm; 5879 LogStream ls(log.error()); 5880 oop(addr)->print_on(&ls); 5881 log.error(" (" INTPTR_FORMAT " should have been marked)", p2i(addr)); 5882 fatal("... aborting"); 5883 } 5884 } 5885 } 5886 5887 void MarkRefsIntoVerifyClosure::do_oop(oop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); } 5888 void MarkRefsIntoVerifyClosure::do_oop(narrowOop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); } 5889 5890 ////////////////////////////////////////////////// 5891 // MarkRefsIntoAndScanClosure 5892 ////////////////////////////////////////////////// 5893 5894 MarkRefsIntoAndScanClosure::MarkRefsIntoAndScanClosure(MemRegion span, 5895 ReferenceProcessor* rp, 5896 CMSBitMap* bit_map, 5897 CMSBitMap* mod_union_table, 5898 CMSMarkStack* mark_stack, 5899 CMSCollector* collector, 5900 bool should_yield, 5901 bool concurrent_precleaning): 5902 _collector(collector), 5903 _span(span), 5904 _bit_map(bit_map), 5905 _mark_stack(mark_stack), 5906 _pushAndMarkClosure(collector, span, rp, bit_map, mod_union_table, 5907 mark_stack, concurrent_precleaning), 5908 _yield(should_yield), 5909 _concurrent_precleaning(concurrent_precleaning), 5910 _freelistLock(NULL) 5911 { 5912 // FIXME: Should initialize in base class constructor. 5913 assert(rp != NULL, "ref_processor shouldn't be NULL"); 5914 set_ref_processor_internal(rp); 5915 } 5916 5917 // This closure is used to mark refs into the CMS generation at the 5918 // second (final) checkpoint, and to scan and transitively follow 5919 // the unmarked oops. It is also used during the concurrent precleaning 5920 // phase while scanning objects on dirty cards in the CMS generation. 5921 // The marks are made in the marking bit map and the marking stack is 5922 // used for keeping the (newly) grey objects during the scan. 5923 // The parallel version (Par_...) appears further below. 5924 void MarkRefsIntoAndScanClosure::do_oop(oop obj) { 5925 if (obj != NULL) { 5926 assert(obj->is_oop(), "expected an oop"); 5927 HeapWord* addr = (HeapWord*)obj; 5928 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 5929 assert(_collector->overflow_list_is_empty(), 5930 "overflow list should be empty"); 5931 if (_span.contains(addr) && 5932 !_bit_map->isMarked(addr)) { 5933 // mark bit map (object is now grey) 5934 _bit_map->mark(addr); 5935 // push on marking stack (stack should be empty), and drain the 5936 // stack by applying this closure to the oops in the oops popped 5937 // from the stack (i.e. blacken the grey objects) 5938 bool res = _mark_stack->push(obj); 5939 assert(res, "Should have space to push on empty stack"); 5940 do { 5941 oop new_oop = _mark_stack->pop(); 5942 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 5943 assert(_bit_map->isMarked((HeapWord*)new_oop), 5944 "only grey objects on this stack"); 5945 // iterate over the oops in this oop, marking and pushing 5946 // the ones in CMS heap (i.e. in _span). 5947 new_oop->oop_iterate(&_pushAndMarkClosure); 5948 // check if it's time to yield 5949 do_yield_check(); 5950 } while (!_mark_stack->isEmpty() || 5951 (!_concurrent_precleaning && take_from_overflow_list())); 5952 // if marking stack is empty, and we are not doing this 5953 // during precleaning, then check the overflow list 5954 } 5955 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 5956 assert(_collector->overflow_list_is_empty(), 5957 "overflow list was drained above"); 5958 5959 assert(_collector->no_preserved_marks(), 5960 "All preserved marks should have been restored above"); 5961 } 5962 } 5963 5964 void MarkRefsIntoAndScanClosure::do_oop(oop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); } 5965 void MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); } 5966 5967 void MarkRefsIntoAndScanClosure::do_yield_work() { 5968 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 5969 "CMS thread should hold CMS token"); 5970 assert_lock_strong(_freelistLock); 5971 assert_lock_strong(_bit_map->lock()); 5972 // relinquish the free_list_lock and bitMaplock() 5973 _bit_map->lock()->unlock(); 5974 _freelistLock->unlock(); 5975 ConcurrentMarkSweepThread::desynchronize(true); 5976 _collector->stopTimer(); 5977 _collector->incrementYields(); 5978 5979 // See the comment in coordinator_yield() 5980 for (unsigned i = 0; 5981 i < CMSYieldSleepCount && 5982 ConcurrentMarkSweepThread::should_yield() && 5983 !CMSCollector::foregroundGCIsActive(); 5984 ++i) { 5985 os::sleep(Thread::current(), 1, false); 5986 } 5987 5988 ConcurrentMarkSweepThread::synchronize(true); 5989 _freelistLock->lock_without_safepoint_check(); 5990 _bit_map->lock()->lock_without_safepoint_check(); 5991 _collector->startTimer(); 5992 } 5993 5994 /////////////////////////////////////////////////////////// 5995 // ParMarkRefsIntoAndScanClosure: a parallel version of 5996 // MarkRefsIntoAndScanClosure 5997 /////////////////////////////////////////////////////////// 5998 ParMarkRefsIntoAndScanClosure::ParMarkRefsIntoAndScanClosure( 5999 CMSCollector* collector, MemRegion span, ReferenceProcessor* rp, 6000 CMSBitMap* bit_map, OopTaskQueue* work_queue): 6001 _span(span), 6002 _bit_map(bit_map), 6003 _work_queue(work_queue), 6004 _low_water_mark(MIN2((work_queue->max_elems()/4), 6005 ((uint)CMSWorkQueueDrainThreshold * ParallelGCThreads))), 6006 _parPushAndMarkClosure(collector, span, rp, bit_map, work_queue) 6007 { 6008 // FIXME: Should initialize in base class constructor. 6009 assert(rp != NULL, "ref_processor shouldn't be NULL"); 6010 set_ref_processor_internal(rp); 6011 } 6012 6013 // This closure is used to mark refs into the CMS generation at the 6014 // second (final) checkpoint, and to scan and transitively follow 6015 // the unmarked oops. The marks are made in the marking bit map and 6016 // the work_queue is used for keeping the (newly) grey objects during 6017 // the scan phase whence they are also available for stealing by parallel 6018 // threads. Since the marking bit map is shared, updates are 6019 // synchronized (via CAS). 6020 void ParMarkRefsIntoAndScanClosure::do_oop(oop obj) { 6021 if (obj != NULL) { 6022 // Ignore mark word because this could be an already marked oop 6023 // that may be chained at the end of the overflow list. 6024 assert(obj->is_oop(true), "expected an oop"); 6025 HeapWord* addr = (HeapWord*)obj; 6026 if (_span.contains(addr) && 6027 !_bit_map->isMarked(addr)) { 6028 // mark bit map (object will become grey): 6029 // It is possible for several threads to be 6030 // trying to "claim" this object concurrently; 6031 // the unique thread that succeeds in marking the 6032 // object first will do the subsequent push on 6033 // to the work queue (or overflow list). 6034 if (_bit_map->par_mark(addr)) { 6035 // push on work_queue (which may not be empty), and trim the 6036 // queue to an appropriate length by applying this closure to 6037 // the oops in the oops popped from the stack (i.e. blacken the 6038 // grey objects) 6039 bool res = _work_queue->push(obj); 6040 assert(res, "Low water mark should be less than capacity?"); 6041 trim_queue(_low_water_mark); 6042 } // Else, another thread claimed the object 6043 } 6044 } 6045 } 6046 6047 void ParMarkRefsIntoAndScanClosure::do_oop(oop* p) { ParMarkRefsIntoAndScanClosure::do_oop_work(p); } 6048 void ParMarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { ParMarkRefsIntoAndScanClosure::do_oop_work(p); } 6049 6050 // This closure is used to rescan the marked objects on the dirty cards 6051 // in the mod union table and the card table proper. 6052 size_t ScanMarkedObjectsAgainCarefullyClosure::do_object_careful_m( 6053 oop p, MemRegion mr) { 6054 6055 size_t size = 0; 6056 HeapWord* addr = (HeapWord*)p; 6057 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6058 assert(_span.contains(addr), "we are scanning the CMS generation"); 6059 // check if it's time to yield 6060 if (do_yield_check()) { 6061 // We yielded for some foreground stop-world work, 6062 // and we have been asked to abort this ongoing preclean cycle. 6063 return 0; 6064 } 6065 if (_bitMap->isMarked(addr)) { 6066 // it's marked; is it potentially uninitialized? 6067 if (p->klass_or_null_acquire() != NULL) { 6068 // an initialized object; ignore mark word in verification below 6069 // since we are running concurrent with mutators 6070 assert(p->is_oop(true), "should be an oop"); 6071 if (p->is_objArray()) { 6072 // objArrays are precisely marked; restrict scanning 6073 // to dirty cards only. 6074 size = CompactibleFreeListSpace::adjustObjectSize( 6075 p->oop_iterate_size(_scanningClosure, mr)); 6076 } else { 6077 // A non-array may have been imprecisely marked; we need 6078 // to scan object in its entirety. 6079 size = CompactibleFreeListSpace::adjustObjectSize( 6080 p->oop_iterate_size(_scanningClosure)); 6081 } 6082 #ifdef ASSERT 6083 size_t direct_size = 6084 CompactibleFreeListSpace::adjustObjectSize(p->size()); 6085 assert(size == direct_size, "Inconsistency in size"); 6086 assert(size >= 3, "Necessary for Printezis marks to work"); 6087 HeapWord* start_pbit = addr + 1; 6088 HeapWord* end_pbit = addr + size - 1; 6089 assert(_bitMap->isMarked(start_pbit) == _bitMap->isMarked(end_pbit), 6090 "inconsistent Printezis mark"); 6091 // Verify inner mark bits (between Printezis bits) are clear, 6092 // but don't repeat if there are multiple dirty regions for 6093 // the same object, to avoid potential O(N^2) performance. 6094 if (addr != _last_scanned_object) { 6095 _bitMap->verifyNoOneBitsInRange(start_pbit + 1, end_pbit); 6096 _last_scanned_object = addr; 6097 } 6098 #endif // ASSERT 6099 } else { 6100 // An uninitialized object. 6101 assert(_bitMap->isMarked(addr+1), "missing Printezis mark?"); 6102 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 6103 size = pointer_delta(nextOneAddr + 1, addr); 6104 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 6105 "alignment problem"); 6106 // Note that pre-cleaning needn't redirty the card. OopDesc::set_klass() 6107 // will dirty the card when the klass pointer is installed in the 6108 // object (signaling the completion of initialization). 6109 } 6110 } else { 6111 // Either a not yet marked object or an uninitialized object 6112 if (p->klass_or_null_acquire() == NULL) { 6113 // An uninitialized object, skip to the next card, since 6114 // we may not be able to read its P-bits yet. 6115 assert(size == 0, "Initial value"); 6116 } else { 6117 // An object not (yet) reached by marking: we merely need to 6118 // compute its size so as to go look at the next block. 6119 assert(p->is_oop(true), "should be an oop"); 6120 size = CompactibleFreeListSpace::adjustObjectSize(p->size()); 6121 } 6122 } 6123 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6124 return size; 6125 } 6126 6127 void ScanMarkedObjectsAgainCarefullyClosure::do_yield_work() { 6128 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6129 "CMS thread should hold CMS token"); 6130 assert_lock_strong(_freelistLock); 6131 assert_lock_strong(_bitMap->lock()); 6132 // relinquish the free_list_lock and bitMaplock() 6133 _bitMap->lock()->unlock(); 6134 _freelistLock->unlock(); 6135 ConcurrentMarkSweepThread::desynchronize(true); 6136 _collector->stopTimer(); 6137 _collector->incrementYields(); 6138 6139 // See the comment in coordinator_yield() 6140 for (unsigned i = 0; i < CMSYieldSleepCount && 6141 ConcurrentMarkSweepThread::should_yield() && 6142 !CMSCollector::foregroundGCIsActive(); ++i) { 6143 os::sleep(Thread::current(), 1, false); 6144 } 6145 6146 ConcurrentMarkSweepThread::synchronize(true); 6147 _freelistLock->lock_without_safepoint_check(); 6148 _bitMap->lock()->lock_without_safepoint_check(); 6149 _collector->startTimer(); 6150 } 6151 6152 6153 ////////////////////////////////////////////////////////////////// 6154 // SurvivorSpacePrecleanClosure 6155 ////////////////////////////////////////////////////////////////// 6156 // This (single-threaded) closure is used to preclean the oops in 6157 // the survivor spaces. 6158 size_t SurvivorSpacePrecleanClosure::do_object_careful(oop p) { 6159 6160 HeapWord* addr = (HeapWord*)p; 6161 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6162 assert(!_span.contains(addr), "we are scanning the survivor spaces"); 6163 assert(p->klass_or_null() != NULL, "object should be initialized"); 6164 // an initialized object; ignore mark word in verification below 6165 // since we are running concurrent with mutators 6166 assert(p->is_oop(true), "should be an oop"); 6167 // Note that we do not yield while we iterate over 6168 // the interior oops of p, pushing the relevant ones 6169 // on our marking stack. 6170 size_t size = p->oop_iterate_size(_scanning_closure); 6171 do_yield_check(); 6172 // Observe that below, we do not abandon the preclean 6173 // phase as soon as we should; rather we empty the 6174 // marking stack before returning. This is to satisfy 6175 // some existing assertions. In general, it may be a 6176 // good idea to abort immediately and complete the marking 6177 // from the grey objects at a later time. 6178 while (!_mark_stack->isEmpty()) { 6179 oop new_oop = _mark_stack->pop(); 6180 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 6181 assert(_bit_map->isMarked((HeapWord*)new_oop), 6182 "only grey objects on this stack"); 6183 // iterate over the oops in this oop, marking and pushing 6184 // the ones in CMS heap (i.e. in _span). 6185 new_oop->oop_iterate(_scanning_closure); 6186 // check if it's time to yield 6187 do_yield_check(); 6188 } 6189 unsigned int after_count = 6190 GenCollectedHeap::heap()->total_collections(); 6191 bool abort = (_before_count != after_count) || 6192 _collector->should_abort_preclean(); 6193 return abort ? 0 : size; 6194 } 6195 6196 void SurvivorSpacePrecleanClosure::do_yield_work() { 6197 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6198 "CMS thread should hold CMS token"); 6199 assert_lock_strong(_bit_map->lock()); 6200 // Relinquish the bit map lock 6201 _bit_map->lock()->unlock(); 6202 ConcurrentMarkSweepThread::desynchronize(true); 6203 _collector->stopTimer(); 6204 _collector->incrementYields(); 6205 6206 // See the comment in coordinator_yield() 6207 for (unsigned i = 0; i < CMSYieldSleepCount && 6208 ConcurrentMarkSweepThread::should_yield() && 6209 !CMSCollector::foregroundGCIsActive(); ++i) { 6210 os::sleep(Thread::current(), 1, false); 6211 } 6212 6213 ConcurrentMarkSweepThread::synchronize(true); 6214 _bit_map->lock()->lock_without_safepoint_check(); 6215 _collector->startTimer(); 6216 } 6217 6218 // This closure is used to rescan the marked objects on the dirty cards 6219 // in the mod union table and the card table proper. In the parallel 6220 // case, although the bitMap is shared, we do a single read so the 6221 // isMarked() query is "safe". 6222 bool ScanMarkedObjectsAgainClosure::do_object_bm(oop p, MemRegion mr) { 6223 // Ignore mark word because we are running concurrent with mutators 6224 assert(p->is_oop_or_null(true), "Expected an oop or NULL at " PTR_FORMAT, p2i(p)); 6225 HeapWord* addr = (HeapWord*)p; 6226 assert(_span.contains(addr), "we are scanning the CMS generation"); 6227 bool is_obj_array = false; 6228 #ifdef ASSERT 6229 if (!_parallel) { 6230 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 6231 assert(_collector->overflow_list_is_empty(), 6232 "overflow list should be empty"); 6233 6234 } 6235 #endif // ASSERT 6236 if (_bit_map->isMarked(addr)) { 6237 // Obj arrays are precisely marked, non-arrays are not; 6238 // so we scan objArrays precisely and non-arrays in their 6239 // entirety. 6240 if (p->is_objArray()) { 6241 is_obj_array = true; 6242 if (_parallel) { 6243 p->oop_iterate(_par_scan_closure, mr); 6244 } else { 6245 p->oop_iterate(_scan_closure, mr); 6246 } 6247 } else { 6248 if (_parallel) { 6249 p->oop_iterate(_par_scan_closure); 6250 } else { 6251 p->oop_iterate(_scan_closure); 6252 } 6253 } 6254 } 6255 #ifdef ASSERT 6256 if (!_parallel) { 6257 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 6258 assert(_collector->overflow_list_is_empty(), 6259 "overflow list should be empty"); 6260 6261 } 6262 #endif // ASSERT 6263 return is_obj_array; 6264 } 6265 6266 MarkFromRootsClosure::MarkFromRootsClosure(CMSCollector* collector, 6267 MemRegion span, 6268 CMSBitMap* bitMap, CMSMarkStack* markStack, 6269 bool should_yield, bool verifying): 6270 _collector(collector), 6271 _span(span), 6272 _bitMap(bitMap), 6273 _mut(&collector->_modUnionTable), 6274 _markStack(markStack), 6275 _yield(should_yield), 6276 _skipBits(0) 6277 { 6278 assert(_markStack->isEmpty(), "stack should be empty"); 6279 _finger = _bitMap->startWord(); 6280 _threshold = _finger; 6281 assert(_collector->_restart_addr == NULL, "Sanity check"); 6282 assert(_span.contains(_finger), "Out of bounds _finger?"); 6283 DEBUG_ONLY(_verifying = verifying;) 6284 } 6285 6286 void MarkFromRootsClosure::reset(HeapWord* addr) { 6287 assert(_markStack->isEmpty(), "would cause duplicates on stack"); 6288 assert(_span.contains(addr), "Out of bounds _finger?"); 6289 _finger = addr; 6290 _threshold = align_up(_finger, CardTableModRefBS::card_size); 6291 } 6292 6293 // Should revisit to see if this should be restructured for 6294 // greater efficiency. 6295 bool MarkFromRootsClosure::do_bit(size_t offset) { 6296 if (_skipBits > 0) { 6297 _skipBits--; 6298 return true; 6299 } 6300 // convert offset into a HeapWord* 6301 HeapWord* addr = _bitMap->startWord() + offset; 6302 assert(_bitMap->endWord() && addr < _bitMap->endWord(), 6303 "address out of range"); 6304 assert(_bitMap->isMarked(addr), "tautology"); 6305 if (_bitMap->isMarked(addr+1)) { 6306 // this is an allocated but not yet initialized object 6307 assert(_skipBits == 0, "tautology"); 6308 _skipBits = 2; // skip next two marked bits ("Printezis-marks") 6309 oop p = oop(addr); 6310 if (p->klass_or_null_acquire() == NULL) { 6311 DEBUG_ONLY(if (!_verifying) {) 6312 // We re-dirty the cards on which this object lies and increase 6313 // the _threshold so that we'll come back to scan this object 6314 // during the preclean or remark phase. (CMSCleanOnEnter) 6315 if (CMSCleanOnEnter) { 6316 size_t sz = _collector->block_size_using_printezis_bits(addr); 6317 HeapWord* end_card_addr = align_up(addr + sz, CardTableModRefBS::card_size); 6318 MemRegion redirty_range = MemRegion(addr, end_card_addr); 6319 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 6320 // Bump _threshold to end_card_addr; note that 6321 // _threshold cannot possibly exceed end_card_addr, anyhow. 6322 // This prevents future clearing of the card as the scan proceeds 6323 // to the right. 6324 assert(_threshold <= end_card_addr, 6325 "Because we are just scanning into this object"); 6326 if (_threshold < end_card_addr) { 6327 _threshold = end_card_addr; 6328 } 6329 if (p->klass_or_null_acquire() != NULL) { 6330 // Redirty the range of cards... 6331 _mut->mark_range(redirty_range); 6332 } // ...else the setting of klass will dirty the card anyway. 6333 } 6334 DEBUG_ONLY(}) 6335 return true; 6336 } 6337 } 6338 scanOopsInOop(addr); 6339 return true; 6340 } 6341 6342 // We take a break if we've been at this for a while, 6343 // so as to avoid monopolizing the locks involved. 6344 void MarkFromRootsClosure::do_yield_work() { 6345 // First give up the locks, then yield, then re-lock 6346 // We should probably use a constructor/destructor idiom to 6347 // do this unlock/lock or modify the MutexUnlocker class to 6348 // serve our purpose. XXX 6349 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6350 "CMS thread should hold CMS token"); 6351 assert_lock_strong(_bitMap->lock()); 6352 _bitMap->lock()->unlock(); 6353 ConcurrentMarkSweepThread::desynchronize(true); 6354 _collector->stopTimer(); 6355 _collector->incrementYields(); 6356 6357 // See the comment in coordinator_yield() 6358 for (unsigned i = 0; i < CMSYieldSleepCount && 6359 ConcurrentMarkSweepThread::should_yield() && 6360 !CMSCollector::foregroundGCIsActive(); ++i) { 6361 os::sleep(Thread::current(), 1, false); 6362 } 6363 6364 ConcurrentMarkSweepThread::synchronize(true); 6365 _bitMap->lock()->lock_without_safepoint_check(); 6366 _collector->startTimer(); 6367 } 6368 6369 void MarkFromRootsClosure::scanOopsInOop(HeapWord* ptr) { 6370 assert(_bitMap->isMarked(ptr), "expected bit to be set"); 6371 assert(_markStack->isEmpty(), 6372 "should drain stack to limit stack usage"); 6373 // convert ptr to an oop preparatory to scanning 6374 oop obj = oop(ptr); 6375 // Ignore mark word in verification below, since we 6376 // may be running concurrent with mutators. 6377 assert(obj->is_oop(true), "should be an oop"); 6378 assert(_finger <= ptr, "_finger runneth ahead"); 6379 // advance the finger to right end of this object 6380 _finger = ptr + obj->size(); 6381 assert(_finger > ptr, "we just incremented it above"); 6382 // On large heaps, it may take us some time to get through 6383 // the marking phase. During 6384 // this time it's possible that a lot of mutations have 6385 // accumulated in the card table and the mod union table -- 6386 // these mutation records are redundant until we have 6387 // actually traced into the corresponding card. 6388 // Here, we check whether advancing the finger would make 6389 // us cross into a new card, and if so clear corresponding 6390 // cards in the MUT (preclean them in the card-table in the 6391 // future). 6392 6393 DEBUG_ONLY(if (!_verifying) {) 6394 // The clean-on-enter optimization is disabled by default, 6395 // until we fix 6178663. 6396 if (CMSCleanOnEnter && (_finger > _threshold)) { 6397 // [_threshold, _finger) represents the interval 6398 // of cards to be cleared in MUT (or precleaned in card table). 6399 // The set of cards to be cleared is all those that overlap 6400 // with the interval [_threshold, _finger); note that 6401 // _threshold is always kept card-aligned but _finger isn't 6402 // always card-aligned. 6403 HeapWord* old_threshold = _threshold; 6404 assert(is_aligned(old_threshold, CardTableModRefBS::card_size), 6405 "_threshold should always be card-aligned"); 6406 _threshold = align_up(_finger, CardTableModRefBS::card_size); 6407 MemRegion mr(old_threshold, _threshold); 6408 assert(!mr.is_empty(), "Control point invariant"); 6409 assert(_span.contains(mr), "Should clear within span"); 6410 _mut->clear_range(mr); 6411 } 6412 DEBUG_ONLY(}) 6413 // Note: the finger doesn't advance while we drain 6414 // the stack below. 6415 PushOrMarkClosure pushOrMarkClosure(_collector, 6416 _span, _bitMap, _markStack, 6417 _finger, this); 6418 bool res = _markStack->push(obj); 6419 assert(res, "Empty non-zero size stack should have space for single push"); 6420 while (!_markStack->isEmpty()) { 6421 oop new_oop = _markStack->pop(); 6422 // Skip verifying header mark word below because we are 6423 // running concurrent with mutators. 6424 assert(new_oop->is_oop(true), "Oops! expected to pop an oop"); 6425 // now scan this oop's oops 6426 new_oop->oop_iterate(&pushOrMarkClosure); 6427 do_yield_check(); 6428 } 6429 assert(_markStack->isEmpty(), "tautology, emphasizing post-condition"); 6430 } 6431 6432 ParMarkFromRootsClosure::ParMarkFromRootsClosure(CMSConcMarkingTask* task, 6433 CMSCollector* collector, MemRegion span, 6434 CMSBitMap* bit_map, 6435 OopTaskQueue* work_queue, 6436 CMSMarkStack* overflow_stack): 6437 _collector(collector), 6438 _whole_span(collector->_span), 6439 _span(span), 6440 _bit_map(bit_map), 6441 _mut(&collector->_modUnionTable), 6442 _work_queue(work_queue), 6443 _overflow_stack(overflow_stack), 6444 _skip_bits(0), 6445 _task(task) 6446 { 6447 assert(_work_queue->size() == 0, "work_queue should be empty"); 6448 _finger = span.start(); 6449 _threshold = _finger; // XXX Defer clear-on-enter optimization for now 6450 assert(_span.contains(_finger), "Out of bounds _finger?"); 6451 } 6452 6453 // Should revisit to see if this should be restructured for 6454 // greater efficiency. 6455 bool ParMarkFromRootsClosure::do_bit(size_t offset) { 6456 if (_skip_bits > 0) { 6457 _skip_bits--; 6458 return true; 6459 } 6460 // convert offset into a HeapWord* 6461 HeapWord* addr = _bit_map->startWord() + offset; 6462 assert(_bit_map->endWord() && addr < _bit_map->endWord(), 6463 "address out of range"); 6464 assert(_bit_map->isMarked(addr), "tautology"); 6465 if (_bit_map->isMarked(addr+1)) { 6466 // this is an allocated object that might not yet be initialized 6467 assert(_skip_bits == 0, "tautology"); 6468 _skip_bits = 2; // skip next two marked bits ("Printezis-marks") 6469 oop p = oop(addr); 6470 if (p->klass_or_null_acquire() == NULL) { 6471 // in the case of Clean-on-Enter optimization, redirty card 6472 // and avoid clearing card by increasing the threshold. 6473 return true; 6474 } 6475 } 6476 scan_oops_in_oop(addr); 6477 return true; 6478 } 6479 6480 void ParMarkFromRootsClosure::scan_oops_in_oop(HeapWord* ptr) { 6481 assert(_bit_map->isMarked(ptr), "expected bit to be set"); 6482 // Should we assert that our work queue is empty or 6483 // below some drain limit? 6484 assert(_work_queue->size() == 0, 6485 "should drain stack to limit stack usage"); 6486 // convert ptr to an oop preparatory to scanning 6487 oop obj = oop(ptr); 6488 // Ignore mark word in verification below, since we 6489 // may be running concurrent with mutators. 6490 assert(obj->is_oop(true), "should be an oop"); 6491 assert(_finger <= ptr, "_finger runneth ahead"); 6492 // advance the finger to right end of this object 6493 _finger = ptr + obj->size(); 6494 assert(_finger > ptr, "we just incremented it above"); 6495 // On large heaps, it may take us some time to get through 6496 // the marking phase. During 6497 // this time it's possible that a lot of mutations have 6498 // accumulated in the card table and the mod union table -- 6499 // these mutation records are redundant until we have 6500 // actually traced into the corresponding card. 6501 // Here, we check whether advancing the finger would make 6502 // us cross into a new card, and if so clear corresponding 6503 // cards in the MUT (preclean them in the card-table in the 6504 // future). 6505 6506 // The clean-on-enter optimization is disabled by default, 6507 // until we fix 6178663. 6508 if (CMSCleanOnEnter && (_finger > _threshold)) { 6509 // [_threshold, _finger) represents the interval 6510 // of cards to be cleared in MUT (or precleaned in card table). 6511 // The set of cards to be cleared is all those that overlap 6512 // with the interval [_threshold, _finger); note that 6513 // _threshold is always kept card-aligned but _finger isn't 6514 // always card-aligned. 6515 HeapWord* old_threshold = _threshold; 6516 assert(is_aligned(old_threshold, CardTableModRefBS::card_size), 6517 "_threshold should always be card-aligned"); 6518 _threshold = align_up(_finger, CardTableModRefBS::card_size); 6519 MemRegion mr(old_threshold, _threshold); 6520 assert(!mr.is_empty(), "Control point invariant"); 6521 assert(_span.contains(mr), "Should clear within span"); // _whole_span ?? 6522 _mut->clear_range(mr); 6523 } 6524 6525 // Note: the local finger doesn't advance while we drain 6526 // the stack below, but the global finger sure can and will. 6527 HeapWord* volatile* gfa = _task->global_finger_addr(); 6528 ParPushOrMarkClosure pushOrMarkClosure(_collector, 6529 _span, _bit_map, 6530 _work_queue, 6531 _overflow_stack, 6532 _finger, 6533 gfa, this); 6534 bool res = _work_queue->push(obj); // overflow could occur here 6535 assert(res, "Will hold once we use workqueues"); 6536 while (true) { 6537 oop new_oop; 6538 if (!_work_queue->pop_local(new_oop)) { 6539 // We emptied our work_queue; check if there's stuff that can 6540 // be gotten from the overflow stack. 6541 if (CMSConcMarkingTask::get_work_from_overflow_stack( 6542 _overflow_stack, _work_queue)) { 6543 do_yield_check(); 6544 continue; 6545 } else { // done 6546 break; 6547 } 6548 } 6549 // Skip verifying header mark word below because we are 6550 // running concurrent with mutators. 6551 assert(new_oop->is_oop(true), "Oops! expected to pop an oop"); 6552 // now scan this oop's oops 6553 new_oop->oop_iterate(&pushOrMarkClosure); 6554 do_yield_check(); 6555 } 6556 assert(_work_queue->size() == 0, "tautology, emphasizing post-condition"); 6557 } 6558 6559 // Yield in response to a request from VM Thread or 6560 // from mutators. 6561 void ParMarkFromRootsClosure::do_yield_work() { 6562 assert(_task != NULL, "sanity"); 6563 _task->yield(); 6564 } 6565 6566 // A variant of the above used for verifying CMS marking work. 6567 MarkFromRootsVerifyClosure::MarkFromRootsVerifyClosure(CMSCollector* collector, 6568 MemRegion span, 6569 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 6570 CMSMarkStack* mark_stack): 6571 _collector(collector), 6572 _span(span), 6573 _verification_bm(verification_bm), 6574 _cms_bm(cms_bm), 6575 _mark_stack(mark_stack), 6576 _pam_verify_closure(collector, span, verification_bm, cms_bm, 6577 mark_stack) 6578 { 6579 assert(_mark_stack->isEmpty(), "stack should be empty"); 6580 _finger = _verification_bm->startWord(); 6581 assert(_collector->_restart_addr == NULL, "Sanity check"); 6582 assert(_span.contains(_finger), "Out of bounds _finger?"); 6583 } 6584 6585 void MarkFromRootsVerifyClosure::reset(HeapWord* addr) { 6586 assert(_mark_stack->isEmpty(), "would cause duplicates on stack"); 6587 assert(_span.contains(addr), "Out of bounds _finger?"); 6588 _finger = addr; 6589 } 6590 6591 // Should revisit to see if this should be restructured for 6592 // greater efficiency. 6593 bool MarkFromRootsVerifyClosure::do_bit(size_t offset) { 6594 // convert offset into a HeapWord* 6595 HeapWord* addr = _verification_bm->startWord() + offset; 6596 assert(_verification_bm->endWord() && addr < _verification_bm->endWord(), 6597 "address out of range"); 6598 assert(_verification_bm->isMarked(addr), "tautology"); 6599 assert(_cms_bm->isMarked(addr), "tautology"); 6600 6601 assert(_mark_stack->isEmpty(), 6602 "should drain stack to limit stack usage"); 6603 // convert addr to an oop preparatory to scanning 6604 oop obj = oop(addr); 6605 assert(obj->is_oop(), "should be an oop"); 6606 assert(_finger <= addr, "_finger runneth ahead"); 6607 // advance the finger to right end of this object 6608 _finger = addr + obj->size(); 6609 assert(_finger > addr, "we just incremented it above"); 6610 // Note: the finger doesn't advance while we drain 6611 // the stack below. 6612 bool res = _mark_stack->push(obj); 6613 assert(res, "Empty non-zero size stack should have space for single push"); 6614 while (!_mark_stack->isEmpty()) { 6615 oop new_oop = _mark_stack->pop(); 6616 assert(new_oop->is_oop(), "Oops! expected to pop an oop"); 6617 // now scan this oop's oops 6618 new_oop->oop_iterate(&_pam_verify_closure); 6619 } 6620 assert(_mark_stack->isEmpty(), "tautology, emphasizing post-condition"); 6621 return true; 6622 } 6623 6624 PushAndMarkVerifyClosure::PushAndMarkVerifyClosure( 6625 CMSCollector* collector, MemRegion span, 6626 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 6627 CMSMarkStack* mark_stack): 6628 MetadataAwareOopClosure(collector->ref_processor()), 6629 _collector(collector), 6630 _span(span), 6631 _verification_bm(verification_bm), 6632 _cms_bm(cms_bm), 6633 _mark_stack(mark_stack) 6634 { } 6635 6636 void PushAndMarkVerifyClosure::do_oop(oop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 6637 void PushAndMarkVerifyClosure::do_oop(narrowOop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 6638 6639 // Upon stack overflow, we discard (part of) the stack, 6640 // remembering the least address amongst those discarded 6641 // in CMSCollector's _restart_address. 6642 void PushAndMarkVerifyClosure::handle_stack_overflow(HeapWord* lost) { 6643 // Remember the least grey address discarded 6644 HeapWord* ra = (HeapWord*)_mark_stack->least_value(lost); 6645 _collector->lower_restart_addr(ra); 6646 _mark_stack->reset(); // discard stack contents 6647 _mark_stack->expand(); // expand the stack if possible 6648 } 6649 6650 void PushAndMarkVerifyClosure::do_oop(oop obj) { 6651 assert(obj->is_oop_or_null(), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6652 HeapWord* addr = (HeapWord*)obj; 6653 if (_span.contains(addr) && !_verification_bm->isMarked(addr)) { 6654 // Oop lies in _span and isn't yet grey or black 6655 _verification_bm->mark(addr); // now grey 6656 if (!_cms_bm->isMarked(addr)) { 6657 Log(gc, verify) log; 6658 ResourceMark rm; 6659 LogStream ls(log.error()); 6660 oop(addr)->print_on(&ls); 6661 log.error(" (" INTPTR_FORMAT " should have been marked)", p2i(addr)); 6662 fatal("... aborting"); 6663 } 6664 6665 if (!_mark_stack->push(obj)) { // stack overflow 6666 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _mark_stack->capacity()); 6667 assert(_mark_stack->isFull(), "Else push should have succeeded"); 6668 handle_stack_overflow(addr); 6669 } 6670 // anything including and to the right of _finger 6671 // will be scanned as we iterate over the remainder of the 6672 // bit map 6673 } 6674 } 6675 6676 PushOrMarkClosure::PushOrMarkClosure(CMSCollector* collector, 6677 MemRegion span, 6678 CMSBitMap* bitMap, CMSMarkStack* markStack, 6679 HeapWord* finger, MarkFromRootsClosure* parent) : 6680 MetadataAwareOopClosure(collector->ref_processor()), 6681 _collector(collector), 6682 _span(span), 6683 _bitMap(bitMap), 6684 _markStack(markStack), 6685 _finger(finger), 6686 _parent(parent) 6687 { } 6688 6689 ParPushOrMarkClosure::ParPushOrMarkClosure(CMSCollector* collector, 6690 MemRegion span, 6691 CMSBitMap* bit_map, 6692 OopTaskQueue* work_queue, 6693 CMSMarkStack* overflow_stack, 6694 HeapWord* finger, 6695 HeapWord* volatile* global_finger_addr, 6696 ParMarkFromRootsClosure* parent) : 6697 MetadataAwareOopClosure(collector->ref_processor()), 6698 _collector(collector), 6699 _whole_span(collector->_span), 6700 _span(span), 6701 _bit_map(bit_map), 6702 _work_queue(work_queue), 6703 _overflow_stack(overflow_stack), 6704 _finger(finger), 6705 _global_finger_addr(global_finger_addr), 6706 _parent(parent) 6707 { } 6708 6709 // Assumes thread-safe access by callers, who are 6710 // responsible for mutual exclusion. 6711 void CMSCollector::lower_restart_addr(HeapWord* low) { 6712 assert(_span.contains(low), "Out of bounds addr"); 6713 if (_restart_addr == NULL) { 6714 _restart_addr = low; 6715 } else { 6716 _restart_addr = MIN2(_restart_addr, low); 6717 } 6718 } 6719 6720 // Upon stack overflow, we discard (part of) the stack, 6721 // remembering the least address amongst those discarded 6722 // in CMSCollector's _restart_address. 6723 void PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 6724 // Remember the least grey address discarded 6725 HeapWord* ra = (HeapWord*)_markStack->least_value(lost); 6726 _collector->lower_restart_addr(ra); 6727 _markStack->reset(); // discard stack contents 6728 _markStack->expand(); // expand the stack if possible 6729 } 6730 6731 // Upon stack overflow, we discard (part of) the stack, 6732 // remembering the least address amongst those discarded 6733 // in CMSCollector's _restart_address. 6734 void ParPushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 6735 // We need to do this under a mutex to prevent other 6736 // workers from interfering with the work done below. 6737 MutexLockerEx ml(_overflow_stack->par_lock(), 6738 Mutex::_no_safepoint_check_flag); 6739 // Remember the least grey address discarded 6740 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 6741 _collector->lower_restart_addr(ra); 6742 _overflow_stack->reset(); // discard stack contents 6743 _overflow_stack->expand(); // expand the stack if possible 6744 } 6745 6746 void PushOrMarkClosure::do_oop(oop obj) { 6747 // Ignore mark word because we are running concurrent with mutators. 6748 assert(obj->is_oop_or_null(true), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6749 HeapWord* addr = (HeapWord*)obj; 6750 if (_span.contains(addr) && !_bitMap->isMarked(addr)) { 6751 // Oop lies in _span and isn't yet grey or black 6752 _bitMap->mark(addr); // now grey 6753 if (addr < _finger) { 6754 // the bit map iteration has already either passed, or 6755 // sampled, this bit in the bit map; we'll need to 6756 // use the marking stack to scan this oop's oops. 6757 bool simulate_overflow = false; 6758 NOT_PRODUCT( 6759 if (CMSMarkStackOverflowALot && 6760 _collector->simulate_overflow()) { 6761 // simulate a stack overflow 6762 simulate_overflow = true; 6763 } 6764 ) 6765 if (simulate_overflow || !_markStack->push(obj)) { // stack overflow 6766 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _markStack->capacity()); 6767 assert(simulate_overflow || _markStack->isFull(), "Else push should have succeeded"); 6768 handle_stack_overflow(addr); 6769 } 6770 } 6771 // anything including and to the right of _finger 6772 // will be scanned as we iterate over the remainder of the 6773 // bit map 6774 do_yield_check(); 6775 } 6776 } 6777 6778 void PushOrMarkClosure::do_oop(oop* p) { PushOrMarkClosure::do_oop_work(p); } 6779 void PushOrMarkClosure::do_oop(narrowOop* p) { PushOrMarkClosure::do_oop_work(p); } 6780 6781 void ParPushOrMarkClosure::do_oop(oop obj) { 6782 // Ignore mark word because we are running concurrent with mutators. 6783 assert(obj->is_oop_or_null(true), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6784 HeapWord* addr = (HeapWord*)obj; 6785 if (_whole_span.contains(addr) && !_bit_map->isMarked(addr)) { 6786 // Oop lies in _span and isn't yet grey or black 6787 // We read the global_finger (volatile read) strictly after marking oop 6788 bool res = _bit_map->par_mark(addr); // now grey 6789 volatile HeapWord** gfa = (volatile HeapWord**)_global_finger_addr; 6790 // Should we push this marked oop on our stack? 6791 // -- if someone else marked it, nothing to do 6792 // -- if target oop is above global finger nothing to do 6793 // -- if target oop is in chunk and above local finger 6794 // then nothing to do 6795 // -- else push on work queue 6796 if ( !res // someone else marked it, they will deal with it 6797 || (addr >= *gfa) // will be scanned in a later task 6798 || (_span.contains(addr) && addr >= _finger)) { // later in this chunk 6799 return; 6800 } 6801 // the bit map iteration has already either passed, or 6802 // sampled, this bit in the bit map; we'll need to 6803 // use the marking stack to scan this oop's oops. 6804 bool simulate_overflow = false; 6805 NOT_PRODUCT( 6806 if (CMSMarkStackOverflowALot && 6807 _collector->simulate_overflow()) { 6808 // simulate a stack overflow 6809 simulate_overflow = true; 6810 } 6811 ) 6812 if (simulate_overflow || 6813 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) { 6814 // stack overflow 6815 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _overflow_stack->capacity()); 6816 // We cannot assert that the overflow stack is full because 6817 // it may have been emptied since. 6818 assert(simulate_overflow || 6819 _work_queue->size() == _work_queue->max_elems(), 6820 "Else push should have succeeded"); 6821 handle_stack_overflow(addr); 6822 } 6823 do_yield_check(); 6824 } 6825 } 6826 6827 void ParPushOrMarkClosure::do_oop(oop* p) { ParPushOrMarkClosure::do_oop_work(p); } 6828 void ParPushOrMarkClosure::do_oop(narrowOop* p) { ParPushOrMarkClosure::do_oop_work(p); } 6829 6830 PushAndMarkClosure::PushAndMarkClosure(CMSCollector* collector, 6831 MemRegion span, 6832 ReferenceProcessor* rp, 6833 CMSBitMap* bit_map, 6834 CMSBitMap* mod_union_table, 6835 CMSMarkStack* mark_stack, 6836 bool concurrent_precleaning): 6837 MetadataAwareOopClosure(rp), 6838 _collector(collector), 6839 _span(span), 6840 _bit_map(bit_map), 6841 _mod_union_table(mod_union_table), 6842 _mark_stack(mark_stack), 6843 _concurrent_precleaning(concurrent_precleaning) 6844 { 6845 assert(ref_processor() != NULL, "ref_processor shouldn't be NULL"); 6846 } 6847 6848 // Grey object rescan during pre-cleaning and second checkpoint phases -- 6849 // the non-parallel version (the parallel version appears further below.) 6850 void PushAndMarkClosure::do_oop(oop obj) { 6851 // Ignore mark word verification. If during concurrent precleaning, 6852 // the object monitor may be locked. If during the checkpoint 6853 // phases, the object may already have been reached by a different 6854 // path and may be at the end of the global overflow list (so 6855 // the mark word may be NULL). 6856 assert(obj->is_oop_or_null(true /* ignore mark word */), 6857 "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6858 HeapWord* addr = (HeapWord*)obj; 6859 // Check if oop points into the CMS generation 6860 // and is not marked 6861 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 6862 // a white object ... 6863 _bit_map->mark(addr); // ... now grey 6864 // push on the marking stack (grey set) 6865 bool simulate_overflow = false; 6866 NOT_PRODUCT( 6867 if (CMSMarkStackOverflowALot && 6868 _collector->simulate_overflow()) { 6869 // simulate a stack overflow 6870 simulate_overflow = true; 6871 } 6872 ) 6873 if (simulate_overflow || !_mark_stack->push(obj)) { 6874 if (_concurrent_precleaning) { 6875 // During precleaning we can just dirty the appropriate card(s) 6876 // in the mod union table, thus ensuring that the object remains 6877 // in the grey set and continue. In the case of object arrays 6878 // we need to dirty all of the cards that the object spans, 6879 // since the rescan of object arrays will be limited to the 6880 // dirty cards. 6881 // Note that no one can be interfering with us in this action 6882 // of dirtying the mod union table, so no locking or atomics 6883 // are required. 6884 if (obj->is_objArray()) { 6885 size_t sz = obj->size(); 6886 HeapWord* end_card_addr = align_up(addr + sz, CardTableModRefBS::card_size); 6887 MemRegion redirty_range = MemRegion(addr, end_card_addr); 6888 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 6889 _mod_union_table->mark_range(redirty_range); 6890 } else { 6891 _mod_union_table->mark(addr); 6892 } 6893 _collector->_ser_pmc_preclean_ovflw++; 6894 } else { 6895 // During the remark phase, we need to remember this oop 6896 // in the overflow list. 6897 _collector->push_on_overflow_list(obj); 6898 _collector->_ser_pmc_remark_ovflw++; 6899 } 6900 } 6901 } 6902 } 6903 6904 ParPushAndMarkClosure::ParPushAndMarkClosure(CMSCollector* collector, 6905 MemRegion span, 6906 ReferenceProcessor* rp, 6907 CMSBitMap* bit_map, 6908 OopTaskQueue* work_queue): 6909 MetadataAwareOopClosure(rp), 6910 _collector(collector), 6911 _span(span), 6912 _bit_map(bit_map), 6913 _work_queue(work_queue) 6914 { 6915 assert(ref_processor() != NULL, "ref_processor shouldn't be NULL"); 6916 } 6917 6918 void PushAndMarkClosure::do_oop(oop* p) { PushAndMarkClosure::do_oop_work(p); } 6919 void PushAndMarkClosure::do_oop(narrowOop* p) { PushAndMarkClosure::do_oop_work(p); } 6920 6921 // Grey object rescan during second checkpoint phase -- 6922 // the parallel version. 6923 void ParPushAndMarkClosure::do_oop(oop obj) { 6924 // In the assert below, we ignore the mark word because 6925 // this oop may point to an already visited object that is 6926 // on the overflow stack (in which case the mark word has 6927 // been hijacked for chaining into the overflow stack -- 6928 // if this is the last object in the overflow stack then 6929 // its mark word will be NULL). Because this object may 6930 // have been subsequently popped off the global overflow 6931 // stack, and the mark word possibly restored to the prototypical 6932 // value, by the time we get to examined this failing assert in 6933 // the debugger, is_oop_or_null(false) may subsequently start 6934 // to hold. 6935 assert(obj->is_oop_or_null(true), 6936 "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6937 HeapWord* addr = (HeapWord*)obj; 6938 // Check if oop points into the CMS generation 6939 // and is not marked 6940 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 6941 // a white object ... 6942 // If we manage to "claim" the object, by being the 6943 // first thread to mark it, then we push it on our 6944 // marking stack 6945 if (_bit_map->par_mark(addr)) { // ... now grey 6946 // push on work queue (grey set) 6947 bool simulate_overflow = false; 6948 NOT_PRODUCT( 6949 if (CMSMarkStackOverflowALot && 6950 _collector->par_simulate_overflow()) { 6951 // simulate a stack overflow 6952 simulate_overflow = true; 6953 } 6954 ) 6955 if (simulate_overflow || !_work_queue->push(obj)) { 6956 _collector->par_push_on_overflow_list(obj); 6957 _collector->_par_pmc_remark_ovflw++; // imprecise OK: no need to CAS 6958 } 6959 } // Else, some other thread got there first 6960 } 6961 } 6962 6963 void ParPushAndMarkClosure::do_oop(oop* p) { ParPushAndMarkClosure::do_oop_work(p); } 6964 void ParPushAndMarkClosure::do_oop(narrowOop* p) { ParPushAndMarkClosure::do_oop_work(p); } 6965 6966 void CMSPrecleanRefsYieldClosure::do_yield_work() { 6967 Mutex* bml = _collector->bitMapLock(); 6968 assert_lock_strong(bml); 6969 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6970 "CMS thread should hold CMS token"); 6971 6972 bml->unlock(); 6973 ConcurrentMarkSweepThread::desynchronize(true); 6974 6975 _collector->stopTimer(); 6976 _collector->incrementYields(); 6977 6978 // See the comment in coordinator_yield() 6979 for (unsigned i = 0; i < CMSYieldSleepCount && 6980 ConcurrentMarkSweepThread::should_yield() && 6981 !CMSCollector::foregroundGCIsActive(); ++i) { 6982 os::sleep(Thread::current(), 1, false); 6983 } 6984 6985 ConcurrentMarkSweepThread::synchronize(true); 6986 bml->lock(); 6987 6988 _collector->startTimer(); 6989 } 6990 6991 bool CMSPrecleanRefsYieldClosure::should_return() { 6992 if (ConcurrentMarkSweepThread::should_yield()) { 6993 do_yield_work(); 6994 } 6995 return _collector->foregroundGCIsActive(); 6996 } 6997 6998 void MarkFromDirtyCardsClosure::do_MemRegion(MemRegion mr) { 6999 assert(((size_t)mr.start())%CardTableModRefBS::card_size_in_words == 0, 7000 "mr should be aligned to start at a card boundary"); 7001 // We'd like to assert: 7002 // assert(mr.word_size()%CardTableModRefBS::card_size_in_words == 0, 7003 // "mr should be a range of cards"); 7004 // However, that would be too strong in one case -- the last 7005 // partition ends at _unallocated_block which, in general, can be 7006 // an arbitrary boundary, not necessarily card aligned. 7007 _num_dirty_cards += mr.word_size()/CardTableModRefBS::card_size_in_words; 7008 _space->object_iterate_mem(mr, &_scan_cl); 7009 } 7010 7011 SweepClosure::SweepClosure(CMSCollector* collector, 7012 ConcurrentMarkSweepGeneration* g, 7013 CMSBitMap* bitMap, bool should_yield) : 7014 _collector(collector), 7015 _g(g), 7016 _sp(g->cmsSpace()), 7017 _limit(_sp->sweep_limit()), 7018 _freelistLock(_sp->freelistLock()), 7019 _bitMap(bitMap), 7020 _yield(should_yield), 7021 _inFreeRange(false), // No free range at beginning of sweep 7022 _freeRangeInFreeLists(false), // No free range at beginning of sweep 7023 _lastFreeRangeCoalesced(false), 7024 _freeFinger(g->used_region().start()) 7025 { 7026 NOT_PRODUCT( 7027 _numObjectsFreed = 0; 7028 _numWordsFreed = 0; 7029 _numObjectsLive = 0; 7030 _numWordsLive = 0; 7031 _numObjectsAlreadyFree = 0; 7032 _numWordsAlreadyFree = 0; 7033 _last_fc = NULL; 7034 7035 _sp->initializeIndexedFreeListArrayReturnedBytes(); 7036 _sp->dictionary()->initialize_dict_returned_bytes(); 7037 ) 7038 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7039 "sweep _limit out of bounds"); 7040 log_develop_trace(gc, sweep)("===================="); 7041 log_develop_trace(gc, sweep)("Starting new sweep with limit " PTR_FORMAT, p2i(_limit)); 7042 } 7043 7044 void SweepClosure::print_on(outputStream* st) const { 7045 st->print_cr("_sp = [" PTR_FORMAT "," PTR_FORMAT ")", 7046 p2i(_sp->bottom()), p2i(_sp->end())); 7047 st->print_cr("_limit = " PTR_FORMAT, p2i(_limit)); 7048 st->print_cr("_freeFinger = " PTR_FORMAT, p2i(_freeFinger)); 7049 NOT_PRODUCT(st->print_cr("_last_fc = " PTR_FORMAT, p2i(_last_fc));) 7050 st->print_cr("_inFreeRange = %d, _freeRangeInFreeLists = %d, _lastFreeRangeCoalesced = %d", 7051 _inFreeRange, _freeRangeInFreeLists, _lastFreeRangeCoalesced); 7052 } 7053 7054 #ifndef PRODUCT 7055 // Assertion checking only: no useful work in product mode -- 7056 // however, if any of the flags below become product flags, 7057 // you may need to review this code to see if it needs to be 7058 // enabled in product mode. 7059 SweepClosure::~SweepClosure() { 7060 assert_lock_strong(_freelistLock); 7061 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7062 "sweep _limit out of bounds"); 7063 if (inFreeRange()) { 7064 Log(gc, sweep) log; 7065 log.error("inFreeRange() should have been reset; dumping state of SweepClosure"); 7066 ResourceMark rm; 7067 LogStream ls(log.error()); 7068 print_on(&ls); 7069 ShouldNotReachHere(); 7070 } 7071 7072 if (log_is_enabled(Debug, gc, sweep)) { 7073 log_debug(gc, sweep)("Collected " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes", 7074 _numObjectsFreed, _numWordsFreed*sizeof(HeapWord)); 7075 log_debug(gc, sweep)("Live " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes Already free " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes", 7076 _numObjectsLive, _numWordsLive*sizeof(HeapWord), _numObjectsAlreadyFree, _numWordsAlreadyFree*sizeof(HeapWord)); 7077 size_t totalBytes = (_numWordsFreed + _numWordsLive + _numWordsAlreadyFree) * sizeof(HeapWord); 7078 log_debug(gc, sweep)("Total sweep: " SIZE_FORMAT " bytes", totalBytes); 7079 } 7080 7081 if (log_is_enabled(Trace, gc, sweep) && CMSVerifyReturnedBytes) { 7082 size_t indexListReturnedBytes = _sp->sumIndexedFreeListArrayReturnedBytes(); 7083 size_t dict_returned_bytes = _sp->dictionary()->sum_dict_returned_bytes(); 7084 size_t returned_bytes = indexListReturnedBytes + dict_returned_bytes; 7085 log_trace(gc, sweep)("Returned " SIZE_FORMAT " bytes Indexed List Returned " SIZE_FORMAT " bytes Dictionary Returned " SIZE_FORMAT " bytes", 7086 returned_bytes, indexListReturnedBytes, dict_returned_bytes); 7087 } 7088 log_develop_trace(gc, sweep)("end of sweep with _limit = " PTR_FORMAT, p2i(_limit)); 7089 log_develop_trace(gc, sweep)("================"); 7090 } 7091 #endif // PRODUCT 7092 7093 void SweepClosure::initialize_free_range(HeapWord* freeFinger, 7094 bool freeRangeInFreeLists) { 7095 log_develop_trace(gc, sweep)("---- Start free range at " PTR_FORMAT " with free block (%d)", 7096 p2i(freeFinger), freeRangeInFreeLists); 7097 assert(!inFreeRange(), "Trampling existing free range"); 7098 set_inFreeRange(true); 7099 set_lastFreeRangeCoalesced(false); 7100 7101 set_freeFinger(freeFinger); 7102 set_freeRangeInFreeLists(freeRangeInFreeLists); 7103 if (CMSTestInFreeList) { 7104 if (freeRangeInFreeLists) { 7105 FreeChunk* fc = (FreeChunk*) freeFinger; 7106 assert(fc->is_free(), "A chunk on the free list should be free."); 7107 assert(fc->size() > 0, "Free range should have a size"); 7108 assert(_sp->verify_chunk_in_free_list(fc), "Chunk is not in free lists"); 7109 } 7110 } 7111 } 7112 7113 // Note that the sweeper runs concurrently with mutators. Thus, 7114 // it is possible for direct allocation in this generation to happen 7115 // in the middle of the sweep. Note that the sweeper also coalesces 7116 // contiguous free blocks. Thus, unless the sweeper and the allocator 7117 // synchronize appropriately freshly allocated blocks may get swept up. 7118 // This is accomplished by the sweeper locking the free lists while 7119 // it is sweeping. Thus blocks that are determined to be free are 7120 // indeed free. There is however one additional complication: 7121 // blocks that have been allocated since the final checkpoint and 7122 // mark, will not have been marked and so would be treated as 7123 // unreachable and swept up. To prevent this, the allocator marks 7124 // the bit map when allocating during the sweep phase. This leads, 7125 // however, to a further complication -- objects may have been allocated 7126 // but not yet initialized -- in the sense that the header isn't yet 7127 // installed. The sweeper can not then determine the size of the block 7128 // in order to skip over it. To deal with this case, we use a technique 7129 // (due to Printezis) to encode such uninitialized block sizes in the 7130 // bit map. Since the bit map uses a bit per every HeapWord, but the 7131 // CMS generation has a minimum object size of 3 HeapWords, it follows 7132 // that "normal marks" won't be adjacent in the bit map (there will 7133 // always be at least two 0 bits between successive 1 bits). We make use 7134 // of these "unused" bits to represent uninitialized blocks -- the bit 7135 // corresponding to the start of the uninitialized object and the next 7136 // bit are both set. Finally, a 1 bit marks the end of the object that 7137 // started with the two consecutive 1 bits to indicate its potentially 7138 // uninitialized state. 7139 7140 size_t SweepClosure::do_blk_careful(HeapWord* addr) { 7141 FreeChunk* fc = (FreeChunk*)addr; 7142 size_t res; 7143 7144 // Check if we are done sweeping. Below we check "addr >= _limit" rather 7145 // than "addr == _limit" because although _limit was a block boundary when 7146 // we started the sweep, it may no longer be one because heap expansion 7147 // may have caused us to coalesce the block ending at the address _limit 7148 // with a newly expanded chunk (this happens when _limit was set to the 7149 // previous _end of the space), so we may have stepped past _limit: 7150 // see the following Zeno-like trail of CRs 6977970, 7008136, 7042740. 7151 if (addr >= _limit) { // we have swept up to or past the limit: finish up 7152 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7153 "sweep _limit out of bounds"); 7154 assert(addr < _sp->end(), "addr out of bounds"); 7155 // Flush any free range we might be holding as a single 7156 // coalesced chunk to the appropriate free list. 7157 if (inFreeRange()) { 7158 assert(freeFinger() >= _sp->bottom() && freeFinger() < _limit, 7159 "freeFinger() " PTR_FORMAT " is out-of-bounds", p2i(freeFinger())); 7160 flush_cur_free_chunk(freeFinger(), 7161 pointer_delta(addr, freeFinger())); 7162 log_develop_trace(gc, sweep)("Sweep: last chunk: put_free_blk " PTR_FORMAT " (" SIZE_FORMAT ") [coalesced:%d]", 7163 p2i(freeFinger()), pointer_delta(addr, freeFinger()), 7164 lastFreeRangeCoalesced() ? 1 : 0); 7165 } 7166 7167 // help the iterator loop finish 7168 return pointer_delta(_sp->end(), addr); 7169 } 7170 7171 assert(addr < _limit, "sweep invariant"); 7172 // check if we should yield 7173 do_yield_check(addr); 7174 if (fc->is_free()) { 7175 // Chunk that is already free 7176 res = fc->size(); 7177 do_already_free_chunk(fc); 7178 debug_only(_sp->verifyFreeLists()); 7179 // If we flush the chunk at hand in lookahead_and_flush() 7180 // and it's coalesced with a preceding chunk, then the 7181 // process of "mangling" the payload of the coalesced block 7182 // will cause erasure of the size information from the 7183 // (erstwhile) header of all the coalesced blocks but the 7184 // first, so the first disjunct in the assert will not hold 7185 // in that specific case (in which case the second disjunct 7186 // will hold). 7187 assert(res == fc->size() || ((HeapWord*)fc) + res >= _limit, 7188 "Otherwise the size info doesn't change at this step"); 7189 NOT_PRODUCT( 7190 _numObjectsAlreadyFree++; 7191 _numWordsAlreadyFree += res; 7192 ) 7193 NOT_PRODUCT(_last_fc = fc;) 7194 } else if (!_bitMap->isMarked(addr)) { 7195 // Chunk is fresh garbage 7196 res = do_garbage_chunk(fc); 7197 debug_only(_sp->verifyFreeLists()); 7198 NOT_PRODUCT( 7199 _numObjectsFreed++; 7200 _numWordsFreed += res; 7201 ) 7202 } else { 7203 // Chunk that is alive. 7204 res = do_live_chunk(fc); 7205 debug_only(_sp->verifyFreeLists()); 7206 NOT_PRODUCT( 7207 _numObjectsLive++; 7208 _numWordsLive += res; 7209 ) 7210 } 7211 return res; 7212 } 7213 7214 // For the smart allocation, record following 7215 // split deaths - a free chunk is removed from its free list because 7216 // it is being split into two or more chunks. 7217 // split birth - a free chunk is being added to its free list because 7218 // a larger free chunk has been split and resulted in this free chunk. 7219 // coal death - a free chunk is being removed from its free list because 7220 // it is being coalesced into a large free chunk. 7221 // coal birth - a free chunk is being added to its free list because 7222 // it was created when two or more free chunks where coalesced into 7223 // this free chunk. 7224 // 7225 // These statistics are used to determine the desired number of free 7226 // chunks of a given size. The desired number is chosen to be relative 7227 // to the end of a CMS sweep. The desired number at the end of a sweep 7228 // is the 7229 // count-at-end-of-previous-sweep (an amount that was enough) 7230 // - count-at-beginning-of-current-sweep (the excess) 7231 // + split-births (gains in this size during interval) 7232 // - split-deaths (demands on this size during interval) 7233 // where the interval is from the end of one sweep to the end of the 7234 // next. 7235 // 7236 // When sweeping the sweeper maintains an accumulated chunk which is 7237 // the chunk that is made up of chunks that have been coalesced. That 7238 // will be termed the left-hand chunk. A new chunk of garbage that 7239 // is being considered for coalescing will be referred to as the 7240 // right-hand chunk. 7241 // 7242 // When making a decision on whether to coalesce a right-hand chunk with 7243 // the current left-hand chunk, the current count vs. the desired count 7244 // of the left-hand chunk is considered. Also if the right-hand chunk 7245 // is near the large chunk at the end of the heap (see 7246 // ConcurrentMarkSweepGeneration::isNearLargestChunk()), then the 7247 // left-hand chunk is coalesced. 7248 // 7249 // When making a decision about whether to split a chunk, the desired count 7250 // vs. the current count of the candidate to be split is also considered. 7251 // If the candidate is underpopulated (currently fewer chunks than desired) 7252 // a chunk of an overpopulated (currently more chunks than desired) size may 7253 // be chosen. The "hint" associated with a free list, if non-null, points 7254 // to a free list which may be overpopulated. 7255 // 7256 7257 void SweepClosure::do_already_free_chunk(FreeChunk* fc) { 7258 const size_t size = fc->size(); 7259 // Chunks that cannot be coalesced are not in the 7260 // free lists. 7261 if (CMSTestInFreeList && !fc->cantCoalesce()) { 7262 assert(_sp->verify_chunk_in_free_list(fc), 7263 "free chunk should be in free lists"); 7264 } 7265 // a chunk that is already free, should not have been 7266 // marked in the bit map 7267 HeapWord* const addr = (HeapWord*) fc; 7268 assert(!_bitMap->isMarked(addr), "free chunk should be unmarked"); 7269 // Verify that the bit map has no bits marked between 7270 // addr and purported end of this block. 7271 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 7272 7273 // Some chunks cannot be coalesced under any circumstances. 7274 // See the definition of cantCoalesce(). 7275 if (!fc->cantCoalesce()) { 7276 // This chunk can potentially be coalesced. 7277 // All the work is done in 7278 do_post_free_or_garbage_chunk(fc, size); 7279 // Note that if the chunk is not coalescable (the else arm 7280 // below), we unconditionally flush, without needing to do 7281 // a "lookahead," as we do below. 7282 if (inFreeRange()) lookahead_and_flush(fc, size); 7283 } else { 7284 // Code path common to both original and adaptive free lists. 7285 7286 // cant coalesce with previous block; this should be treated 7287 // as the end of a free run if any 7288 if (inFreeRange()) { 7289 // we kicked some butt; time to pick up the garbage 7290 assert(freeFinger() < addr, "freeFinger points too high"); 7291 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 7292 } 7293 // else, nothing to do, just continue 7294 } 7295 } 7296 7297 size_t SweepClosure::do_garbage_chunk(FreeChunk* fc) { 7298 // This is a chunk of garbage. It is not in any free list. 7299 // Add it to a free list or let it possibly be coalesced into 7300 // a larger chunk. 7301 HeapWord* const addr = (HeapWord*) fc; 7302 const size_t size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 7303 7304 // Verify that the bit map has no bits marked between 7305 // addr and purported end of just dead object. 7306 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 7307 do_post_free_or_garbage_chunk(fc, size); 7308 7309 assert(_limit >= addr + size, 7310 "A freshly garbage chunk can't possibly straddle over _limit"); 7311 if (inFreeRange()) lookahead_and_flush(fc, size); 7312 return size; 7313 } 7314 7315 size_t SweepClosure::do_live_chunk(FreeChunk* fc) { 7316 HeapWord* addr = (HeapWord*) fc; 7317 // The sweeper has just found a live object. Return any accumulated 7318 // left hand chunk to the free lists. 7319 if (inFreeRange()) { 7320 assert(freeFinger() < addr, "freeFinger points too high"); 7321 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 7322 } 7323 7324 // This object is live: we'd normally expect this to be 7325 // an oop, and like to assert the following: 7326 // assert(oop(addr)->is_oop(), "live block should be an oop"); 7327 // However, as we commented above, this may be an object whose 7328 // header hasn't yet been initialized. 7329 size_t size; 7330 assert(_bitMap->isMarked(addr), "Tautology for this control point"); 7331 if (_bitMap->isMarked(addr + 1)) { 7332 // Determine the size from the bit map, rather than trying to 7333 // compute it from the object header. 7334 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 7335 size = pointer_delta(nextOneAddr + 1, addr); 7336 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 7337 "alignment problem"); 7338 7339 #ifdef ASSERT 7340 if (oop(addr)->klass_or_null_acquire() != NULL) { 7341 // Ignore mark word because we are running concurrent with mutators 7342 assert(oop(addr)->is_oop(true), "live block should be an oop"); 7343 assert(size == 7344 CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()), 7345 "P-mark and computed size do not agree"); 7346 } 7347 #endif 7348 7349 } else { 7350 // This should be an initialized object that's alive. 7351 assert(oop(addr)->klass_or_null_acquire() != NULL, 7352 "Should be an initialized object"); 7353 // Ignore mark word because we are running concurrent with mutators 7354 assert(oop(addr)->is_oop(true), "live block should be an oop"); 7355 // Verify that the bit map has no bits marked between 7356 // addr and purported end of this block. 7357 size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 7358 assert(size >= 3, "Necessary for Printezis marks to work"); 7359 assert(!_bitMap->isMarked(addr+1), "Tautology for this control point"); 7360 DEBUG_ONLY(_bitMap->verifyNoOneBitsInRange(addr+2, addr+size);) 7361 } 7362 return size; 7363 } 7364 7365 void SweepClosure::do_post_free_or_garbage_chunk(FreeChunk* fc, 7366 size_t chunkSize) { 7367 // do_post_free_or_garbage_chunk() should only be called in the case 7368 // of the adaptive free list allocator. 7369 const bool fcInFreeLists = fc->is_free(); 7370 assert((HeapWord*)fc <= _limit, "sweep invariant"); 7371 if (CMSTestInFreeList && fcInFreeLists) { 7372 assert(_sp->verify_chunk_in_free_list(fc), "free chunk is not in free lists"); 7373 } 7374 7375 log_develop_trace(gc, sweep)(" -- pick up another chunk at " PTR_FORMAT " (" SIZE_FORMAT ")", p2i(fc), chunkSize); 7376 7377 HeapWord* const fc_addr = (HeapWord*) fc; 7378 7379 bool coalesce = false; 7380 const size_t left = pointer_delta(fc_addr, freeFinger()); 7381 const size_t right = chunkSize; 7382 switch (FLSCoalescePolicy) { 7383 // numeric value forms a coalition aggressiveness metric 7384 case 0: { // never coalesce 7385 coalesce = false; 7386 break; 7387 } 7388 case 1: { // coalesce if left & right chunks on overpopulated lists 7389 coalesce = _sp->coalOverPopulated(left) && 7390 _sp->coalOverPopulated(right); 7391 break; 7392 } 7393 case 2: { // coalesce if left chunk on overpopulated list (default) 7394 coalesce = _sp->coalOverPopulated(left); 7395 break; 7396 } 7397 case 3: { // coalesce if left OR right chunk on overpopulated list 7398 coalesce = _sp->coalOverPopulated(left) || 7399 _sp->coalOverPopulated(right); 7400 break; 7401 } 7402 case 4: { // always coalesce 7403 coalesce = true; 7404 break; 7405 } 7406 default: 7407 ShouldNotReachHere(); 7408 } 7409 7410 // Should the current free range be coalesced? 7411 // If the chunk is in a free range and either we decided to coalesce above 7412 // or the chunk is near the large block at the end of the heap 7413 // (isNearLargestChunk() returns true), then coalesce this chunk. 7414 const bool doCoalesce = inFreeRange() 7415 && (coalesce || _g->isNearLargestChunk(fc_addr)); 7416 if (doCoalesce) { 7417 // Coalesce the current free range on the left with the new 7418 // chunk on the right. If either is on a free list, 7419 // it must be removed from the list and stashed in the closure. 7420 if (freeRangeInFreeLists()) { 7421 FreeChunk* const ffc = (FreeChunk*)freeFinger(); 7422 assert(ffc->size() == pointer_delta(fc_addr, freeFinger()), 7423 "Size of free range is inconsistent with chunk size."); 7424 if (CMSTestInFreeList) { 7425 assert(_sp->verify_chunk_in_free_list(ffc), 7426 "Chunk is not in free lists"); 7427 } 7428 _sp->coalDeath(ffc->size()); 7429 _sp->removeFreeChunkFromFreeLists(ffc); 7430 set_freeRangeInFreeLists(false); 7431 } 7432 if (fcInFreeLists) { 7433 _sp->coalDeath(chunkSize); 7434 assert(fc->size() == chunkSize, 7435 "The chunk has the wrong size or is not in the free lists"); 7436 _sp->removeFreeChunkFromFreeLists(fc); 7437 } 7438 set_lastFreeRangeCoalesced(true); 7439 print_free_block_coalesced(fc); 7440 } else { // not in a free range and/or should not coalesce 7441 // Return the current free range and start a new one. 7442 if (inFreeRange()) { 7443 // In a free range but cannot coalesce with the right hand chunk. 7444 // Put the current free range into the free lists. 7445 flush_cur_free_chunk(freeFinger(), 7446 pointer_delta(fc_addr, freeFinger())); 7447 } 7448 // Set up for new free range. Pass along whether the right hand 7449 // chunk is in the free lists. 7450 initialize_free_range((HeapWord*)fc, fcInFreeLists); 7451 } 7452 } 7453 7454 // Lookahead flush: 7455 // If we are tracking a free range, and this is the last chunk that 7456 // we'll look at because its end crosses past _limit, we'll preemptively 7457 // flush it along with any free range we may be holding on to. Note that 7458 // this can be the case only for an already free or freshly garbage 7459 // chunk. If this block is an object, it can never straddle 7460 // over _limit. The "straddling" occurs when _limit is set at 7461 // the previous end of the space when this cycle started, and 7462 // a subsequent heap expansion caused the previously co-terminal 7463 // free block to be coalesced with the newly expanded portion, 7464 // thus rendering _limit a non-block-boundary making it dangerous 7465 // for the sweeper to step over and examine. 7466 void SweepClosure::lookahead_and_flush(FreeChunk* fc, size_t chunk_size) { 7467 assert(inFreeRange(), "Should only be called if currently in a free range."); 7468 HeapWord* const eob = ((HeapWord*)fc) + chunk_size; 7469 assert(_sp->used_region().contains(eob - 1), 7470 "eob = " PTR_FORMAT " eob-1 = " PTR_FORMAT " _limit = " PTR_FORMAT 7471 " out of bounds wrt _sp = [" PTR_FORMAT "," PTR_FORMAT ")" 7472 " when examining fc = " PTR_FORMAT "(" SIZE_FORMAT ")", 7473 p2i(eob), p2i(eob-1), p2i(_limit), p2i(_sp->bottom()), p2i(_sp->end()), p2i(fc), chunk_size); 7474 if (eob >= _limit) { 7475 assert(eob == _limit || fc->is_free(), "Only a free chunk should allow us to cross over the limit"); 7476 log_develop_trace(gc, sweep)("_limit " PTR_FORMAT " reached or crossed by block " 7477 "[" PTR_FORMAT "," PTR_FORMAT ") in space " 7478 "[" PTR_FORMAT "," PTR_FORMAT ")", 7479 p2i(_limit), p2i(fc), p2i(eob), p2i(_sp->bottom()), p2i(_sp->end())); 7480 // Return the storage we are tracking back into the free lists. 7481 log_develop_trace(gc, sweep)("Flushing ... "); 7482 assert(freeFinger() < eob, "Error"); 7483 flush_cur_free_chunk( freeFinger(), pointer_delta(eob, freeFinger())); 7484 } 7485 } 7486 7487 void SweepClosure::flush_cur_free_chunk(HeapWord* chunk, size_t size) { 7488 assert(inFreeRange(), "Should only be called if currently in a free range."); 7489 assert(size > 0, 7490 "A zero sized chunk cannot be added to the free lists."); 7491 if (!freeRangeInFreeLists()) { 7492 if (CMSTestInFreeList) { 7493 FreeChunk* fc = (FreeChunk*) chunk; 7494 fc->set_size(size); 7495 assert(!_sp->verify_chunk_in_free_list(fc), 7496 "chunk should not be in free lists yet"); 7497 } 7498 log_develop_trace(gc, sweep)(" -- add free block " PTR_FORMAT " (" SIZE_FORMAT ") to free lists", p2i(chunk), size); 7499 // A new free range is going to be starting. The current 7500 // free range has not been added to the free lists yet or 7501 // was removed so add it back. 7502 // If the current free range was coalesced, then the death 7503 // of the free range was recorded. Record a birth now. 7504 if (lastFreeRangeCoalesced()) { 7505 _sp->coalBirth(size); 7506 } 7507 _sp->addChunkAndRepairOffsetTable(chunk, size, 7508 lastFreeRangeCoalesced()); 7509 } else { 7510 log_develop_trace(gc, sweep)("Already in free list: nothing to flush"); 7511 } 7512 set_inFreeRange(false); 7513 set_freeRangeInFreeLists(false); 7514 } 7515 7516 // We take a break if we've been at this for a while, 7517 // so as to avoid monopolizing the locks involved. 7518 void SweepClosure::do_yield_work(HeapWord* addr) { 7519 // Return current free chunk being used for coalescing (if any) 7520 // to the appropriate freelist. After yielding, the next 7521 // free block encountered will start a coalescing range of 7522 // free blocks. If the next free block is adjacent to the 7523 // chunk just flushed, they will need to wait for the next 7524 // sweep to be coalesced. 7525 if (inFreeRange()) { 7526 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 7527 } 7528 7529 // First give up the locks, then yield, then re-lock. 7530 // We should probably use a constructor/destructor idiom to 7531 // do this unlock/lock or modify the MutexUnlocker class to 7532 // serve our purpose. XXX 7533 assert_lock_strong(_bitMap->lock()); 7534 assert_lock_strong(_freelistLock); 7535 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 7536 "CMS thread should hold CMS token"); 7537 _bitMap->lock()->unlock(); 7538 _freelistLock->unlock(); 7539 ConcurrentMarkSweepThread::desynchronize(true); 7540 _collector->stopTimer(); 7541 _collector->incrementYields(); 7542 7543 // See the comment in coordinator_yield() 7544 for (unsigned i = 0; i < CMSYieldSleepCount && 7545 ConcurrentMarkSweepThread::should_yield() && 7546 !CMSCollector::foregroundGCIsActive(); ++i) { 7547 os::sleep(Thread::current(), 1, false); 7548 } 7549 7550 ConcurrentMarkSweepThread::synchronize(true); 7551 _freelistLock->lock(); 7552 _bitMap->lock()->lock_without_safepoint_check(); 7553 _collector->startTimer(); 7554 } 7555 7556 #ifndef PRODUCT 7557 // This is actually very useful in a product build if it can 7558 // be called from the debugger. Compile it into the product 7559 // as needed. 7560 bool debug_verify_chunk_in_free_list(FreeChunk* fc) { 7561 return debug_cms_space->verify_chunk_in_free_list(fc); 7562 } 7563 #endif 7564 7565 void SweepClosure::print_free_block_coalesced(FreeChunk* fc) const { 7566 log_develop_trace(gc, sweep)("Sweep:coal_free_blk " PTR_FORMAT " (" SIZE_FORMAT ")", 7567 p2i(fc), fc->size()); 7568 } 7569 7570 // CMSIsAliveClosure 7571 bool CMSIsAliveClosure::do_object_b(oop obj) { 7572 HeapWord* addr = (HeapWord*)obj; 7573 return addr != NULL && 7574 (!_span.contains(addr) || _bit_map->isMarked(addr)); 7575 } 7576 7577 7578 CMSKeepAliveClosure::CMSKeepAliveClosure( CMSCollector* collector, 7579 MemRegion span, 7580 CMSBitMap* bit_map, CMSMarkStack* mark_stack, 7581 bool cpc): 7582 _collector(collector), 7583 _span(span), 7584 _bit_map(bit_map), 7585 _mark_stack(mark_stack), 7586 _concurrent_precleaning(cpc) { 7587 assert(!_span.is_empty(), "Empty span could spell trouble"); 7588 } 7589 7590 7591 // CMSKeepAliveClosure: the serial version 7592 void CMSKeepAliveClosure::do_oop(oop obj) { 7593 HeapWord* addr = (HeapWord*)obj; 7594 if (_span.contains(addr) && 7595 !_bit_map->isMarked(addr)) { 7596 _bit_map->mark(addr); 7597 bool simulate_overflow = false; 7598 NOT_PRODUCT( 7599 if (CMSMarkStackOverflowALot && 7600 _collector->simulate_overflow()) { 7601 // simulate a stack overflow 7602 simulate_overflow = true; 7603 } 7604 ) 7605 if (simulate_overflow || !_mark_stack->push(obj)) { 7606 if (_concurrent_precleaning) { 7607 // We dirty the overflown object and let the remark 7608 // phase deal with it. 7609 assert(_collector->overflow_list_is_empty(), "Error"); 7610 // In the case of object arrays, we need to dirty all of 7611 // the cards that the object spans. No locking or atomics 7612 // are needed since no one else can be mutating the mod union 7613 // table. 7614 if (obj->is_objArray()) { 7615 size_t sz = obj->size(); 7616 HeapWord* end_card_addr = align_up(addr + sz, CardTableModRefBS::card_size); 7617 MemRegion redirty_range = MemRegion(addr, end_card_addr); 7618 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 7619 _collector->_modUnionTable.mark_range(redirty_range); 7620 } else { 7621 _collector->_modUnionTable.mark(addr); 7622 } 7623 _collector->_ser_kac_preclean_ovflw++; 7624 } else { 7625 _collector->push_on_overflow_list(obj); 7626 _collector->_ser_kac_ovflw++; 7627 } 7628 } 7629 } 7630 } 7631 7632 void CMSKeepAliveClosure::do_oop(oop* p) { CMSKeepAliveClosure::do_oop_work(p); } 7633 void CMSKeepAliveClosure::do_oop(narrowOop* p) { CMSKeepAliveClosure::do_oop_work(p); } 7634 7635 // CMSParKeepAliveClosure: a parallel version of the above. 7636 // The work queues are private to each closure (thread), 7637 // but (may be) available for stealing by other threads. 7638 void CMSParKeepAliveClosure::do_oop(oop obj) { 7639 HeapWord* addr = (HeapWord*)obj; 7640 if (_span.contains(addr) && 7641 !_bit_map->isMarked(addr)) { 7642 // In general, during recursive tracing, several threads 7643 // may be concurrently getting here; the first one to 7644 // "tag" it, claims it. 7645 if (_bit_map->par_mark(addr)) { 7646 bool res = _work_queue->push(obj); 7647 assert(res, "Low water mark should be much less than capacity"); 7648 // Do a recursive trim in the hope that this will keep 7649 // stack usage lower, but leave some oops for potential stealers 7650 trim_queue(_low_water_mark); 7651 } // Else, another thread got there first 7652 } 7653 } 7654 7655 void CMSParKeepAliveClosure::do_oop(oop* p) { CMSParKeepAliveClosure::do_oop_work(p); } 7656 void CMSParKeepAliveClosure::do_oop(narrowOop* p) { CMSParKeepAliveClosure::do_oop_work(p); } 7657 7658 void CMSParKeepAliveClosure::trim_queue(uint max) { 7659 while (_work_queue->size() > max) { 7660 oop new_oop; 7661 if (_work_queue->pop_local(new_oop)) { 7662 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 7663 assert(_bit_map->isMarked((HeapWord*)new_oop), 7664 "no white objects on this stack!"); 7665 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 7666 // iterate over the oops in this oop, marking and pushing 7667 // the ones in CMS heap (i.e. in _span). 7668 new_oop->oop_iterate(&_mark_and_push); 7669 } 7670 } 7671 } 7672 7673 CMSInnerParMarkAndPushClosure::CMSInnerParMarkAndPushClosure( 7674 CMSCollector* collector, 7675 MemRegion span, CMSBitMap* bit_map, 7676 OopTaskQueue* work_queue): 7677 _collector(collector), 7678 _span(span), 7679 _bit_map(bit_map), 7680 _work_queue(work_queue) { } 7681 7682 void CMSInnerParMarkAndPushClosure::do_oop(oop obj) { 7683 HeapWord* addr = (HeapWord*)obj; 7684 if (_span.contains(addr) && 7685 !_bit_map->isMarked(addr)) { 7686 if (_bit_map->par_mark(addr)) { 7687 bool simulate_overflow = false; 7688 NOT_PRODUCT( 7689 if (CMSMarkStackOverflowALot && 7690 _collector->par_simulate_overflow()) { 7691 // simulate a stack overflow 7692 simulate_overflow = true; 7693 } 7694 ) 7695 if (simulate_overflow || !_work_queue->push(obj)) { 7696 _collector->par_push_on_overflow_list(obj); 7697 _collector->_par_kac_ovflw++; 7698 } 7699 } // Else another thread got there already 7700 } 7701 } 7702 7703 void CMSInnerParMarkAndPushClosure::do_oop(oop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); } 7704 void CMSInnerParMarkAndPushClosure::do_oop(narrowOop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); } 7705 7706 ////////////////////////////////////////////////////////////////// 7707 // CMSExpansionCause ///////////////////////////// 7708 ////////////////////////////////////////////////////////////////// 7709 const char* CMSExpansionCause::to_string(CMSExpansionCause::Cause cause) { 7710 switch (cause) { 7711 case _no_expansion: 7712 return "No expansion"; 7713 case _satisfy_free_ratio: 7714 return "Free ratio"; 7715 case _satisfy_promotion: 7716 return "Satisfy promotion"; 7717 case _satisfy_allocation: 7718 return "allocation"; 7719 case _allocate_par_lab: 7720 return "Par LAB"; 7721 case _allocate_par_spooling_space: 7722 return "Par Spooling Space"; 7723 case _adaptive_size_policy: 7724 return "Ergonomics"; 7725 default: 7726 return "unknown"; 7727 } 7728 } 7729 7730 void CMSDrainMarkingStackClosure::do_void() { 7731 // the max number to take from overflow list at a time 7732 const size_t num = _mark_stack->capacity()/4; 7733 assert(!_concurrent_precleaning || _collector->overflow_list_is_empty(), 7734 "Overflow list should be NULL during concurrent phases"); 7735 while (!_mark_stack->isEmpty() || 7736 // if stack is empty, check the overflow list 7737 _collector->take_from_overflow_list(num, _mark_stack)) { 7738 oop obj = _mark_stack->pop(); 7739 HeapWord* addr = (HeapWord*)obj; 7740 assert(_span.contains(addr), "Should be within span"); 7741 assert(_bit_map->isMarked(addr), "Should be marked"); 7742 assert(obj->is_oop(), "Should be an oop"); 7743 obj->oop_iterate(_keep_alive); 7744 } 7745 } 7746 7747 void CMSParDrainMarkingStackClosure::do_void() { 7748 // drain queue 7749 trim_queue(0); 7750 } 7751 7752 // Trim our work_queue so its length is below max at return 7753 void CMSParDrainMarkingStackClosure::trim_queue(uint max) { 7754 while (_work_queue->size() > max) { 7755 oop new_oop; 7756 if (_work_queue->pop_local(new_oop)) { 7757 assert(new_oop->is_oop(), "Expected an oop"); 7758 assert(_bit_map->isMarked((HeapWord*)new_oop), 7759 "no white objects on this stack!"); 7760 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 7761 // iterate over the oops in this oop, marking and pushing 7762 // the ones in CMS heap (i.e. in _span). 7763 new_oop->oop_iterate(&_mark_and_push); 7764 } 7765 } 7766 } 7767 7768 //////////////////////////////////////////////////////////////////// 7769 // Support for Marking Stack Overflow list handling and related code 7770 //////////////////////////////////////////////////////////////////// 7771 // Much of the following code is similar in shape and spirit to the 7772 // code used in ParNewGC. We should try and share that code 7773 // as much as possible in the future. 7774 7775 #ifndef PRODUCT 7776 // Debugging support for CMSStackOverflowALot 7777 7778 // It's OK to call this multi-threaded; the worst thing 7779 // that can happen is that we'll get a bunch of closely 7780 // spaced simulated overflows, but that's OK, in fact 7781 // probably good as it would exercise the overflow code 7782 // under contention. 7783 bool CMSCollector::simulate_overflow() { 7784 if (_overflow_counter-- <= 0) { // just being defensive 7785 _overflow_counter = CMSMarkStackOverflowInterval; 7786 return true; 7787 } else { 7788 return false; 7789 } 7790 } 7791 7792 bool CMSCollector::par_simulate_overflow() { 7793 return simulate_overflow(); 7794 } 7795 #endif 7796 7797 // Single-threaded 7798 bool CMSCollector::take_from_overflow_list(size_t num, CMSMarkStack* stack) { 7799 assert(stack->isEmpty(), "Expected precondition"); 7800 assert(stack->capacity() > num, "Shouldn't bite more than can chew"); 7801 size_t i = num; 7802 oop cur = _overflow_list; 7803 const markOop proto = markOopDesc::prototype(); 7804 NOT_PRODUCT(ssize_t n = 0;) 7805 for (oop next; i > 0 && cur != NULL; cur = next, i--) { 7806 next = oop(cur->mark()); 7807 cur->set_mark(proto); // until proven otherwise 7808 assert(cur->is_oop(), "Should be an oop"); 7809 bool res = stack->push(cur); 7810 assert(res, "Bit off more than can chew?"); 7811 NOT_PRODUCT(n++;) 7812 } 7813 _overflow_list = cur; 7814 #ifndef PRODUCT 7815 assert(_num_par_pushes >= n, "Too many pops?"); 7816 _num_par_pushes -=n; 7817 #endif 7818 return !stack->isEmpty(); 7819 } 7820 7821 #define BUSY (cast_to_oop<intptr_t>(0x1aff1aff)) 7822 // (MT-safe) Get a prefix of at most "num" from the list. 7823 // The overflow list is chained through the mark word of 7824 // each object in the list. We fetch the entire list, 7825 // break off a prefix of the right size and return the 7826 // remainder. If other threads try to take objects from 7827 // the overflow list at that time, they will wait for 7828 // some time to see if data becomes available. If (and 7829 // only if) another thread places one or more object(s) 7830 // on the global list before we have returned the suffix 7831 // to the global list, we will walk down our local list 7832 // to find its end and append the global list to 7833 // our suffix before returning it. This suffix walk can 7834 // prove to be expensive (quadratic in the amount of traffic) 7835 // when there are many objects in the overflow list and 7836 // there is much producer-consumer contention on the list. 7837 // *NOTE*: The overflow list manipulation code here and 7838 // in ParNewGeneration:: are very similar in shape, 7839 // except that in the ParNew case we use the old (from/eden) 7840 // copy of the object to thread the list via its klass word. 7841 // Because of the common code, if you make any changes in 7842 // the code below, please check the ParNew version to see if 7843 // similar changes might be needed. 7844 // CR 6797058 has been filed to consolidate the common code. 7845 bool CMSCollector::par_take_from_overflow_list(size_t num, 7846 OopTaskQueue* work_q, 7847 int no_of_gc_threads) { 7848 assert(work_q->size() == 0, "First empty local work queue"); 7849 assert(num < work_q->max_elems(), "Can't bite more than we can chew"); 7850 if (_overflow_list == NULL) { 7851 return false; 7852 } 7853 // Grab the entire list; we'll put back a suffix 7854 oop prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list)); 7855 Thread* tid = Thread::current(); 7856 // Before "no_of_gc_threads" was introduced CMSOverflowSpinCount was 7857 // set to ParallelGCThreads. 7858 size_t CMSOverflowSpinCount = (size_t) no_of_gc_threads; // was ParallelGCThreads; 7859 size_t sleep_time_millis = MAX2((size_t)1, num/100); 7860 // If the list is busy, we spin for a short while, 7861 // sleeping between attempts to get the list. 7862 for (size_t spin = 0; prefix == BUSY && spin < CMSOverflowSpinCount; spin++) { 7863 os::sleep(tid, sleep_time_millis, false); 7864 if (_overflow_list == NULL) { 7865 // Nothing left to take 7866 return false; 7867 } else if (_overflow_list != BUSY) { 7868 // Try and grab the prefix 7869 prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list)); 7870 } 7871 } 7872 // If the list was found to be empty, or we spun long 7873 // enough, we give up and return empty-handed. If we leave 7874 // the list in the BUSY state below, it must be the case that 7875 // some other thread holds the overflow list and will set it 7876 // to a non-BUSY state in the future. 7877 if (prefix == NULL || prefix == BUSY) { 7878 // Nothing to take or waited long enough 7879 if (prefix == NULL) { 7880 // Write back the NULL in case we overwrote it with BUSY above 7881 // and it is still the same value. 7882 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY); 7883 } 7884 return false; 7885 } 7886 assert(prefix != NULL && prefix != BUSY, "Error"); 7887 size_t i = num; 7888 oop cur = prefix; 7889 // Walk down the first "num" objects, unless we reach the end. 7890 for (; i > 1 && cur->mark() != NULL; cur = oop(cur->mark()), i--); 7891 if (cur->mark() == NULL) { 7892 // We have "num" or fewer elements in the list, so there 7893 // is nothing to return to the global list. 7894 // Write back the NULL in lieu of the BUSY we wrote 7895 // above, if it is still the same value. 7896 if (_overflow_list == BUSY) { 7897 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY); 7898 } 7899 } else { 7900 // Chop off the suffix and return it to the global list. 7901 assert(cur->mark() != BUSY, "Error"); 7902 oop suffix_head = cur->mark(); // suffix will be put back on global list 7903 cur->set_mark(NULL); // break off suffix 7904 // It's possible that the list is still in the empty(busy) state 7905 // we left it in a short while ago; in that case we may be 7906 // able to place back the suffix without incurring the cost 7907 // of a walk down the list. 7908 oop observed_overflow_list = _overflow_list; 7909 oop cur_overflow_list = observed_overflow_list; 7910 bool attached = false; 7911 while (observed_overflow_list == BUSY || observed_overflow_list == NULL) { 7912 observed_overflow_list = 7913 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list); 7914 if (cur_overflow_list == observed_overflow_list) { 7915 attached = true; 7916 break; 7917 } else cur_overflow_list = observed_overflow_list; 7918 } 7919 if (!attached) { 7920 // Too bad, someone else sneaked in (at least) an element; we'll need 7921 // to do a splice. Find tail of suffix so we can prepend suffix to global 7922 // list. 7923 for (cur = suffix_head; cur->mark() != NULL; cur = (oop)(cur->mark())); 7924 oop suffix_tail = cur; 7925 assert(suffix_tail != NULL && suffix_tail->mark() == NULL, 7926 "Tautology"); 7927 observed_overflow_list = _overflow_list; 7928 do { 7929 cur_overflow_list = observed_overflow_list; 7930 if (cur_overflow_list != BUSY) { 7931 // Do the splice ... 7932 suffix_tail->set_mark(markOop(cur_overflow_list)); 7933 } else { // cur_overflow_list == BUSY 7934 suffix_tail->set_mark(NULL); 7935 } 7936 // ... and try to place spliced list back on overflow_list ... 7937 observed_overflow_list = 7938 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list); 7939 } while (cur_overflow_list != observed_overflow_list); 7940 // ... until we have succeeded in doing so. 7941 } 7942 } 7943 7944 // Push the prefix elements on work_q 7945 assert(prefix != NULL, "control point invariant"); 7946 const markOop proto = markOopDesc::prototype(); 7947 oop next; 7948 NOT_PRODUCT(ssize_t n = 0;) 7949 for (cur = prefix; cur != NULL; cur = next) { 7950 next = oop(cur->mark()); 7951 cur->set_mark(proto); // until proven otherwise 7952 assert(cur->is_oop(), "Should be an oop"); 7953 bool res = work_q->push(cur); 7954 assert(res, "Bit off more than we can chew?"); 7955 NOT_PRODUCT(n++;) 7956 } 7957 #ifndef PRODUCT 7958 assert(_num_par_pushes >= n, "Too many pops?"); 7959 Atomic::add_ptr(-(intptr_t)n, &_num_par_pushes); 7960 #endif 7961 return true; 7962 } 7963 7964 // Single-threaded 7965 void CMSCollector::push_on_overflow_list(oop p) { 7966 NOT_PRODUCT(_num_par_pushes++;) 7967 assert(p->is_oop(), "Not an oop"); 7968 preserve_mark_if_necessary(p); 7969 p->set_mark((markOop)_overflow_list); 7970 _overflow_list = p; 7971 } 7972 7973 // Multi-threaded; use CAS to prepend to overflow list 7974 void CMSCollector::par_push_on_overflow_list(oop p) { 7975 NOT_PRODUCT(Atomic::inc_ptr(&_num_par_pushes);) 7976 assert(p->is_oop(), "Not an oop"); 7977 par_preserve_mark_if_necessary(p); 7978 oop observed_overflow_list = _overflow_list; 7979 oop cur_overflow_list; 7980 do { 7981 cur_overflow_list = observed_overflow_list; 7982 if (cur_overflow_list != BUSY) { 7983 p->set_mark(markOop(cur_overflow_list)); 7984 } else { 7985 p->set_mark(NULL); 7986 } 7987 observed_overflow_list = 7988 (oop) Atomic::cmpxchg_ptr(p, &_overflow_list, cur_overflow_list); 7989 } while (cur_overflow_list != observed_overflow_list); 7990 } 7991 #undef BUSY 7992 7993 // Single threaded 7994 // General Note on GrowableArray: pushes may silently fail 7995 // because we are (temporarily) out of C-heap for expanding 7996 // the stack. The problem is quite ubiquitous and affects 7997 // a lot of code in the JVM. The prudent thing for GrowableArray 7998 // to do (for now) is to exit with an error. However, that may 7999 // be too draconian in some cases because the caller may be 8000 // able to recover without much harm. For such cases, we 8001 // should probably introduce a "soft_push" method which returns 8002 // an indication of success or failure with the assumption that 8003 // the caller may be able to recover from a failure; code in 8004 // the VM can then be changed, incrementally, to deal with such 8005 // failures where possible, thus, incrementally hardening the VM 8006 // in such low resource situations. 8007 void CMSCollector::preserve_mark_work(oop p, markOop m) { 8008 _preserved_oop_stack.push(p); 8009 _preserved_mark_stack.push(m); 8010 assert(m == p->mark(), "Mark word changed"); 8011 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), 8012 "bijection"); 8013 } 8014 8015 // Single threaded 8016 void CMSCollector::preserve_mark_if_necessary(oop p) { 8017 markOop m = p->mark(); 8018 if (m->must_be_preserved(p)) { 8019 preserve_mark_work(p, m); 8020 } 8021 } 8022 8023 void CMSCollector::par_preserve_mark_if_necessary(oop p) { 8024 markOop m = p->mark(); 8025 if (m->must_be_preserved(p)) { 8026 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag); 8027 // Even though we read the mark word without holding 8028 // the lock, we are assured that it will not change 8029 // because we "own" this oop, so no other thread can 8030 // be trying to push it on the overflow list; see 8031 // the assertion in preserve_mark_work() that checks 8032 // that m == p->mark(). 8033 preserve_mark_work(p, m); 8034 } 8035 } 8036 8037 // We should be able to do this multi-threaded, 8038 // a chunk of stack being a task (this is 8039 // correct because each oop only ever appears 8040 // once in the overflow list. However, it's 8041 // not very easy to completely overlap this with 8042 // other operations, so will generally not be done 8043 // until all work's been completed. Because we 8044 // expect the preserved oop stack (set) to be small, 8045 // it's probably fine to do this single-threaded. 8046 // We can explore cleverer concurrent/overlapped/parallel 8047 // processing of preserved marks if we feel the 8048 // need for this in the future. Stack overflow should 8049 // be so rare in practice and, when it happens, its 8050 // effect on performance so great that this will 8051 // likely just be in the noise anyway. 8052 void CMSCollector::restore_preserved_marks_if_any() { 8053 assert(SafepointSynchronize::is_at_safepoint(), 8054 "world should be stopped"); 8055 assert(Thread::current()->is_ConcurrentGC_thread() || 8056 Thread::current()->is_VM_thread(), 8057 "should be single-threaded"); 8058 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), 8059 "bijection"); 8060 8061 while (!_preserved_oop_stack.is_empty()) { 8062 oop p = _preserved_oop_stack.pop(); 8063 assert(p->is_oop(), "Should be an oop"); 8064 assert(_span.contains(p), "oop should be in _span"); 8065 assert(p->mark() == markOopDesc::prototype(), 8066 "Set when taken from overflow list"); 8067 markOop m = _preserved_mark_stack.pop(); 8068 p->set_mark(m); 8069 } 8070 assert(_preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(), 8071 "stacks were cleared above"); 8072 } 8073 8074 #ifndef PRODUCT 8075 bool CMSCollector::no_preserved_marks() const { 8076 return _preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(); 8077 } 8078 #endif 8079 8080 // Transfer some number of overflown objects to usual marking 8081 // stack. Return true if some objects were transferred. 8082 bool MarkRefsIntoAndScanClosure::take_from_overflow_list() { 8083 size_t num = MIN2((size_t)(_mark_stack->capacity() - _mark_stack->length())/4, 8084 (size_t)ParGCDesiredObjsFromOverflowList); 8085 8086 bool res = _collector->take_from_overflow_list(num, _mark_stack); 8087 assert(_collector->overflow_list_is_empty() || res, 8088 "If list is not empty, we should have taken something"); 8089 assert(!res || !_mark_stack->isEmpty(), 8090 "If we took something, it should now be on our stack"); 8091 return res; 8092 } 8093 8094 size_t MarkDeadObjectsClosure::do_blk(HeapWord* addr) { 8095 size_t res = _sp->block_size_no_stall(addr, _collector); 8096 if (_sp->block_is_obj(addr)) { 8097 if (_live_bit_map->isMarked(addr)) { 8098 // It can't have been dead in a previous cycle 8099 guarantee(!_dead_bit_map->isMarked(addr), "No resurrection!"); 8100 } else { 8101 _dead_bit_map->mark(addr); // mark the dead object 8102 } 8103 } 8104 // Could be 0, if the block size could not be computed without stalling. 8105 return res; 8106 } 8107 8108 TraceCMSMemoryManagerStats::TraceCMSMemoryManagerStats(CMSCollector::CollectorState phase, GCCause::Cause cause): TraceMemoryManagerStats() { 8109 8110 switch (phase) { 8111 case CMSCollector::InitialMarking: 8112 initialize(true /* fullGC */ , 8113 cause /* cause of the GC */, 8114 true /* recordGCBeginTime */, 8115 true /* recordPreGCUsage */, 8116 false /* recordPeakUsage */, 8117 false /* recordPostGCusage */, 8118 true /* recordAccumulatedGCTime */, 8119 false /* recordGCEndTime */, 8120 false /* countCollection */ ); 8121 break; 8122 8123 case CMSCollector::FinalMarking: 8124 initialize(true /* fullGC */ , 8125 cause /* cause of the GC */, 8126 false /* recordGCBeginTime */, 8127 false /* recordPreGCUsage */, 8128 false /* recordPeakUsage */, 8129 false /* recordPostGCusage */, 8130 true /* recordAccumulatedGCTime */, 8131 false /* recordGCEndTime */, 8132 false /* countCollection */ ); 8133 break; 8134 8135 case CMSCollector::Sweeping: 8136 initialize(true /* fullGC */ , 8137 cause /* cause of the GC */, 8138 false /* recordGCBeginTime */, 8139 false /* recordPreGCUsage */, 8140 true /* recordPeakUsage */, 8141 true /* recordPostGCusage */, 8142 false /* recordAccumulatedGCTime */, 8143 true /* recordGCEndTime */, 8144 true /* countCollection */ ); 8145 break; 8146 8147 default: 8148 ShouldNotReachHere(); 8149 } 8150 }