1 /* 2 * Copyright (c) 2001, 2015, 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/systemDictionary.hpp" 29 #include "code/codeCache.hpp" 30 #include "gc/cms/cmsCollectorPolicy.hpp" 31 #include "gc/cms/cmsOopClosures.inline.hpp" 32 #include "gc/cms/compactibleFreeListSpace.hpp" 33 #include "gc/cms/concurrentMarkSweepGeneration.inline.hpp" 34 #include "gc/cms/concurrentMarkSweepThread.hpp" 35 #include "gc/cms/parNewGeneration.hpp" 36 #include "gc/cms/vmCMSOperations.hpp" 37 #include "gc/serial/genMarkSweep.hpp" 38 #include "gc/serial/tenuredGeneration.hpp" 39 #include "gc/shared/adaptiveSizePolicy.hpp" 40 #include "gc/shared/cardGeneration.inline.hpp" 41 #include "gc/shared/cardTableRS.hpp" 42 #include "gc/shared/collectedHeap.inline.hpp" 43 #include "gc/shared/collectorCounters.hpp" 44 #include "gc/shared/collectorPolicy.hpp" 45 #include "gc/shared/gcLocker.inline.hpp" 46 #include "gc/shared/gcPolicyCounters.hpp" 47 #include "gc/shared/gcTimer.hpp" 48 #include "gc/shared/gcTrace.hpp" 49 #include "gc/shared/gcTraceTime.inline.hpp" 50 #include "gc/shared/genCollectedHeap.hpp" 51 #include "gc/shared/genOopClosures.inline.hpp" 52 #include "gc/shared/isGCActiveMark.hpp" 53 #include "gc/shared/referencePolicy.hpp" 54 #include "gc/shared/strongRootsScope.hpp" 55 #include "gc/shared/taskqueue.inline.hpp" 56 #include "logging/log.hpp" 57 #include "memory/allocation.hpp" 58 #include "memory/iterator.inline.hpp" 59 #include "memory/padded.hpp" 60 #include "memory/resourceArea.hpp" 61 #include "oops/oop.inline.hpp" 62 #include "prims/jvmtiExport.hpp" 63 #include "runtime/atomic.inline.hpp" 64 #include "runtime/globals_extension.hpp" 65 #include "runtime/handles.inline.hpp" 66 #include "runtime/java.hpp" 67 #include "runtime/orderAccess.inline.hpp" 68 #include "runtime/timer.hpp" 69 #include "runtime/vmThread.hpp" 70 #include "services/memoryService.hpp" 71 #include "services/runtimeService.hpp" 72 #include "utilities/stack.inline.hpp" 73 74 // statics 75 CMSCollector* ConcurrentMarkSweepGeneration::_collector = NULL; 76 bool CMSCollector::_full_gc_requested = false; 77 GCCause::Cause CMSCollector::_full_gc_cause = GCCause::_no_gc; 78 79 ////////////////////////////////////////////////////////////////// 80 // In support of CMS/VM thread synchronization 81 ////////////////////////////////////////////////////////////////// 82 // We split use of the CGC_lock into 2 "levels". 83 // The low-level locking is of the usual CGC_lock monitor. We introduce 84 // a higher level "token" (hereafter "CMS token") built on top of the 85 // low level monitor (hereafter "CGC lock"). 86 // The token-passing protocol gives priority to the VM thread. The 87 // CMS-lock doesn't provide any fairness guarantees, but clients 88 // should ensure that it is only held for very short, bounded 89 // durations. 90 // 91 // When either of the CMS thread or the VM thread is involved in 92 // collection operations during which it does not want the other 93 // thread to interfere, it obtains the CMS token. 94 // 95 // If either thread tries to get the token while the other has 96 // it, that thread waits. However, if the VM thread and CMS thread 97 // both want the token, then the VM thread gets priority while the 98 // CMS thread waits. This ensures, for instance, that the "concurrent" 99 // phases of the CMS thread's work do not block out the VM thread 100 // for long periods of time as the CMS thread continues to hog 101 // the token. (See bug 4616232). 102 // 103 // The baton-passing functions are, however, controlled by the 104 // flags _foregroundGCShouldWait and _foregroundGCIsActive, 105 // and here the low-level CMS lock, not the high level token, 106 // ensures mutual exclusion. 107 // 108 // Two important conditions that we have to satisfy: 109 // 1. if a thread does a low-level wait on the CMS lock, then it 110 // relinquishes the CMS token if it were holding that token 111 // when it acquired the low-level CMS lock. 112 // 2. any low-level notifications on the low-level lock 113 // should only be sent when a thread has relinquished the token. 114 // 115 // In the absence of either property, we'd have potential deadlock. 116 // 117 // We protect each of the CMS (concurrent and sequential) phases 118 // with the CMS _token_, not the CMS _lock_. 119 // 120 // The only code protected by CMS lock is the token acquisition code 121 // itself, see ConcurrentMarkSweepThread::[de]synchronize(), and the 122 // baton-passing code. 123 // 124 // Unfortunately, i couldn't come up with a good abstraction to factor and 125 // hide the naked CGC_lock manipulation in the baton-passing code 126 // further below. That's something we should try to do. Also, the proof 127 // of correctness of this 2-level locking scheme is far from obvious, 128 // and potentially quite slippery. We have an uneasy suspicion, for instance, 129 // that there may be a theoretical possibility of delay/starvation in the 130 // low-level lock/wait/notify scheme used for the baton-passing because of 131 // potential interference with the priority scheme embodied in the 132 // CMS-token-passing protocol. See related comments at a CGC_lock->wait() 133 // invocation further below and marked with "XXX 20011219YSR". 134 // Indeed, as we note elsewhere, this may become yet more slippery 135 // in the presence of multiple CMS and/or multiple VM threads. XXX 136 137 class CMSTokenSync: public StackObj { 138 private: 139 bool _is_cms_thread; 140 public: 141 CMSTokenSync(bool is_cms_thread): 142 _is_cms_thread(is_cms_thread) { 143 assert(is_cms_thread == Thread::current()->is_ConcurrentGC_thread(), 144 "Incorrect argument to constructor"); 145 ConcurrentMarkSweepThread::synchronize(_is_cms_thread); 146 } 147 148 ~CMSTokenSync() { 149 assert(_is_cms_thread ? 150 ConcurrentMarkSweepThread::cms_thread_has_cms_token() : 151 ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 152 "Incorrect state"); 153 ConcurrentMarkSweepThread::desynchronize(_is_cms_thread); 154 } 155 }; 156 157 // Convenience class that does a CMSTokenSync, and then acquires 158 // upto three locks. 159 class CMSTokenSyncWithLocks: public CMSTokenSync { 160 private: 161 // Note: locks are acquired in textual declaration order 162 // and released in the opposite order 163 MutexLockerEx _locker1, _locker2, _locker3; 164 public: 165 CMSTokenSyncWithLocks(bool is_cms_thread, Mutex* mutex1, 166 Mutex* mutex2 = NULL, Mutex* mutex3 = NULL): 167 CMSTokenSync(is_cms_thread), 168 _locker1(mutex1, Mutex::_no_safepoint_check_flag), 169 _locker2(mutex2, Mutex::_no_safepoint_check_flag), 170 _locker3(mutex3, Mutex::_no_safepoint_check_flag) 171 { } 172 }; 173 174 175 ////////////////////////////////////////////////////////////////// 176 // Concurrent Mark-Sweep Generation ///////////////////////////// 177 ////////////////////////////////////////////////////////////////// 178 179 NOT_PRODUCT(CompactibleFreeListSpace* debug_cms_space;) 180 181 // This struct contains per-thread things necessary to support parallel 182 // young-gen collection. 183 class CMSParGCThreadState: public CHeapObj<mtGC> { 184 public: 185 CFLS_LAB lab; 186 PromotionInfo promo; 187 188 // Constructor. 189 CMSParGCThreadState(CompactibleFreeListSpace* cfls) : lab(cfls) { 190 promo.setSpace(cfls); 191 } 192 }; 193 194 ConcurrentMarkSweepGeneration::ConcurrentMarkSweepGeneration( 195 ReservedSpace rs, size_t initial_byte_size, CardTableRS* ct) : 196 CardGeneration(rs, initial_byte_size, ct), 197 _dilatation_factor(((double)MinChunkSize)/((double)(CollectedHeap::min_fill_size()))), 198 _did_compact(false) 199 { 200 HeapWord* bottom = (HeapWord*) _virtual_space.low(); 201 HeapWord* end = (HeapWord*) _virtual_space.high(); 202 203 _direct_allocated_words = 0; 204 NOT_PRODUCT( 205 _numObjectsPromoted = 0; 206 _numWordsPromoted = 0; 207 _numObjectsAllocated = 0; 208 _numWordsAllocated = 0; 209 ) 210 211 _cmsSpace = new CompactibleFreeListSpace(_bts, MemRegion(bottom, end)); 212 NOT_PRODUCT(debug_cms_space = _cmsSpace;) 213 _cmsSpace->_old_gen = this; 214 215 _gc_stats = new CMSGCStats(); 216 217 // Verify the assumption that FreeChunk::_prev and OopDesc::_klass 218 // offsets match. The ability to tell free chunks from objects 219 // depends on this property. 220 debug_only( 221 FreeChunk* junk = NULL; 222 assert(UseCompressedClassPointers || 223 junk->prev_addr() == (void*)(oop(junk)->klass_addr()), 224 "Offset of FreeChunk::_prev within FreeChunk must match" 225 " that of OopDesc::_klass within OopDesc"); 226 ) 227 228 _par_gc_thread_states = NEW_C_HEAP_ARRAY(CMSParGCThreadState*, ParallelGCThreads, mtGC); 229 for (uint i = 0; i < ParallelGCThreads; i++) { 230 _par_gc_thread_states[i] = new CMSParGCThreadState(cmsSpace()); 231 } 232 233 _incremental_collection_failed = false; 234 // The "dilatation_factor" is the expansion that can occur on 235 // account of the fact that the minimum object size in the CMS 236 // generation may be larger than that in, say, a contiguous young 237 // generation. 238 // Ideally, in the calculation below, we'd compute the dilatation 239 // factor as: MinChunkSize/(promoting_gen's min object size) 240 // Since we do not have such a general query interface for the 241 // promoting generation, we'll instead just use the minimum 242 // object size (which today is a header's worth of space); 243 // note that all arithmetic is in units of HeapWords. 244 assert(MinChunkSize >= CollectedHeap::min_fill_size(), "just checking"); 245 assert(_dilatation_factor >= 1.0, "from previous assert"); 246 } 247 248 249 // The field "_initiating_occupancy" represents the occupancy percentage 250 // at which we trigger a new collection cycle. Unless explicitly specified 251 // via CMSInitiatingOccupancyFraction (argument "io" below), it 252 // is calculated by: 253 // 254 // Let "f" be MinHeapFreeRatio in 255 // 256 // _initiating_occupancy = 100-f + 257 // f * (CMSTriggerRatio/100) 258 // where CMSTriggerRatio is the argument "tr" below. 259 // 260 // That is, if we assume the heap is at its desired maximum occupancy at the 261 // end of a collection, we let CMSTriggerRatio of the (purported) free 262 // space be allocated before initiating a new collection cycle. 263 // 264 void ConcurrentMarkSweepGeneration::init_initiating_occupancy(intx io, uintx tr) { 265 assert(io <= 100 && tr <= 100, "Check the arguments"); 266 if (io >= 0) { 267 _initiating_occupancy = (double)io / 100.0; 268 } else { 269 _initiating_occupancy = ((100 - MinHeapFreeRatio) + 270 (double)(tr * MinHeapFreeRatio) / 100.0) 271 / 100.0; 272 } 273 } 274 275 void ConcurrentMarkSweepGeneration::ref_processor_init() { 276 assert(collector() != NULL, "no collector"); 277 collector()->ref_processor_init(); 278 } 279 280 void CMSCollector::ref_processor_init() { 281 if (_ref_processor == NULL) { 282 // Allocate and initialize a reference processor 283 _ref_processor = 284 new ReferenceProcessor(_span, // span 285 (ParallelGCThreads > 1) && ParallelRefProcEnabled, // mt processing 286 ParallelGCThreads, // mt processing degree 287 _cmsGen->refs_discovery_is_mt(), // mt discovery 288 MAX2(ConcGCThreads, ParallelGCThreads), // mt discovery degree 289 _cmsGen->refs_discovery_is_atomic(), // discovery is not atomic 290 &_is_alive_closure); // closure for liveness info 291 // Initialize the _ref_processor field of CMSGen 292 _cmsGen->set_ref_processor(_ref_processor); 293 294 } 295 } 296 297 AdaptiveSizePolicy* CMSCollector::size_policy() { 298 GenCollectedHeap* gch = GenCollectedHeap::heap(); 299 return gch->gen_policy()->size_policy(); 300 } 301 302 void ConcurrentMarkSweepGeneration::initialize_performance_counters() { 303 304 const char* gen_name = "old"; 305 GenCollectorPolicy* gcp = GenCollectedHeap::heap()->gen_policy(); 306 // Generation Counters - generation 1, 1 subspace 307 _gen_counters = new GenerationCounters(gen_name, 1, 1, 308 gcp->min_old_size(), gcp->max_old_size(), &_virtual_space); 309 310 _space_counters = new GSpaceCounters(gen_name, 0, 311 _virtual_space.reserved_size(), 312 this, _gen_counters); 313 } 314 315 CMSStats::CMSStats(ConcurrentMarkSweepGeneration* cms_gen, unsigned int alpha): 316 _cms_gen(cms_gen) 317 { 318 assert(alpha <= 100, "bad value"); 319 _saved_alpha = alpha; 320 321 // Initialize the alphas to the bootstrap value of 100. 322 _gc0_alpha = _cms_alpha = 100; 323 324 _cms_begin_time.update(); 325 _cms_end_time.update(); 326 327 _gc0_duration = 0.0; 328 _gc0_period = 0.0; 329 _gc0_promoted = 0; 330 331 _cms_duration = 0.0; 332 _cms_period = 0.0; 333 _cms_allocated = 0; 334 335 _cms_used_at_gc0_begin = 0; 336 _cms_used_at_gc0_end = 0; 337 _allow_duty_cycle_reduction = false; 338 _valid_bits = 0; 339 } 340 341 double CMSStats::cms_free_adjustment_factor(size_t free) const { 342 // TBD: CR 6909490 343 return 1.0; 344 } 345 346 void CMSStats::adjust_cms_free_adjustment_factor(bool fail, size_t free) { 347 } 348 349 // If promotion failure handling is on use 350 // the padded average size of the promotion for each 351 // young generation collection. 352 double CMSStats::time_until_cms_gen_full() const { 353 size_t cms_free = _cms_gen->cmsSpace()->free(); 354 GenCollectedHeap* gch = GenCollectedHeap::heap(); 355 size_t expected_promotion = MIN2(gch->young_gen()->capacity(), 356 (size_t) _cms_gen->gc_stats()->avg_promoted()->padded_average()); 357 if (cms_free > expected_promotion) { 358 // Start a cms collection if there isn't enough space to promote 359 // for the next young collection. Use the padded average as 360 // a safety factor. 361 cms_free -= expected_promotion; 362 363 // Adjust by the safety factor. 364 double cms_free_dbl = (double)cms_free; 365 double cms_adjustment = (100.0 - CMSIncrementalSafetyFactor) / 100.0; 366 // Apply a further correction factor which tries to adjust 367 // for recent occurance of concurrent mode failures. 368 cms_adjustment = cms_adjustment * cms_free_adjustment_factor(cms_free); 369 cms_free_dbl = cms_free_dbl * cms_adjustment; 370 371 log_trace(gc)("CMSStats::time_until_cms_gen_full: cms_free " SIZE_FORMAT " expected_promotion " SIZE_FORMAT, 372 cms_free, expected_promotion); 373 log_trace(gc)(" cms_free_dbl %f cms_consumption_rate %f", cms_free_dbl, cms_consumption_rate() + 1.0); 374 // Add 1 in case the consumption rate goes to zero. 375 return cms_free_dbl / (cms_consumption_rate() + 1.0); 376 } 377 return 0.0; 378 } 379 380 // Compare the duration of the cms collection to the 381 // time remaining before the cms generation is empty. 382 // Note that the time from the start of the cms collection 383 // to the start of the cms sweep (less than the total 384 // duration of the cms collection) can be used. This 385 // has been tried and some applications experienced 386 // promotion failures early in execution. This was 387 // possibly because the averages were not accurate 388 // enough at the beginning. 389 double CMSStats::time_until_cms_start() const { 390 // We add "gc0_period" to the "work" calculation 391 // below because this query is done (mostly) at the 392 // end of a scavenge, so we need to conservatively 393 // account for that much possible delay 394 // in the query so as to avoid concurrent mode failures 395 // due to starting the collection just a wee bit too 396 // late. 397 double work = cms_duration() + gc0_period(); 398 double deadline = time_until_cms_gen_full(); 399 // If a concurrent mode failure occurred recently, we want to be 400 // more conservative and halve our expected time_until_cms_gen_full() 401 if (work > deadline) { 402 log_develop_trace(gc)("CMSCollector: collect because of anticipated promotion before full %3.7f + %3.7f > %3.7f ", 403 cms_duration(), gc0_period(), time_until_cms_gen_full()); 404 return 0.0; 405 } 406 return work - deadline; 407 } 408 409 #ifndef PRODUCT 410 void CMSStats::print_on(outputStream *st) const { 411 st->print(" gc0_alpha=%d,cms_alpha=%d", _gc0_alpha, _cms_alpha); 412 st->print(",gc0_dur=%g,gc0_per=%g,gc0_promo=" SIZE_FORMAT, 413 gc0_duration(), gc0_period(), gc0_promoted()); 414 st->print(",cms_dur=%g,cms_per=%g,cms_alloc=" SIZE_FORMAT, 415 cms_duration(), cms_period(), cms_allocated()); 416 st->print(",cms_since_beg=%g,cms_since_end=%g", 417 cms_time_since_begin(), cms_time_since_end()); 418 st->print(",cms_used_beg=" SIZE_FORMAT ",cms_used_end=" SIZE_FORMAT, 419 _cms_used_at_gc0_begin, _cms_used_at_gc0_end); 420 421 if (valid()) { 422 st->print(",promo_rate=%g,cms_alloc_rate=%g", 423 promotion_rate(), cms_allocation_rate()); 424 st->print(",cms_consumption_rate=%g,time_until_full=%g", 425 cms_consumption_rate(), time_until_cms_gen_full()); 426 } 427 st->print(" "); 428 } 429 #endif // #ifndef PRODUCT 430 431 CMSCollector::CollectorState CMSCollector::_collectorState = 432 CMSCollector::Idling; 433 bool CMSCollector::_foregroundGCIsActive = false; 434 bool CMSCollector::_foregroundGCShouldWait = false; 435 436 CMSCollector::CMSCollector(ConcurrentMarkSweepGeneration* cmsGen, 437 CardTableRS* ct, 438 ConcurrentMarkSweepPolicy* cp): 439 _cmsGen(cmsGen), 440 _ct(ct), 441 _ref_processor(NULL), // will be set later 442 _conc_workers(NULL), // may be set later 443 _abort_preclean(false), 444 _start_sampling(false), 445 _between_prologue_and_epilogue(false), 446 _markBitMap(0, Mutex::leaf + 1, "CMS_markBitMap_lock"), 447 _modUnionTable((CardTableModRefBS::card_shift - LogHeapWordSize), 448 -1 /* lock-free */, "No_lock" /* dummy */), 449 _modUnionClosurePar(&_modUnionTable), 450 // Adjust my span to cover old (cms) gen 451 _span(cmsGen->reserved()), 452 // Construct the is_alive_closure with _span & markBitMap 453 _is_alive_closure(_span, &_markBitMap), 454 _restart_addr(NULL), 455 _overflow_list(NULL), 456 _stats(cmsGen), 457 _eden_chunk_lock(new Mutex(Mutex::leaf + 1, "CMS_eden_chunk_lock", true, 458 //verify that this lock should be acquired with safepoint check. 459 Monitor::_safepoint_check_sometimes)), 460 _eden_chunk_array(NULL), // may be set in ctor body 461 _eden_chunk_capacity(0), // -- ditto -- 462 _eden_chunk_index(0), // -- ditto -- 463 _survivor_plab_array(NULL), // -- ditto -- 464 _survivor_chunk_array(NULL), // -- ditto -- 465 _survivor_chunk_capacity(0), // -- ditto -- 466 _survivor_chunk_index(0), // -- ditto -- 467 _ser_pmc_preclean_ovflw(0), 468 _ser_kac_preclean_ovflw(0), 469 _ser_pmc_remark_ovflw(0), 470 _par_pmc_remark_ovflw(0), 471 _ser_kac_ovflw(0), 472 _par_kac_ovflw(0), 473 #ifndef PRODUCT 474 _num_par_pushes(0), 475 #endif 476 _collection_count_start(0), 477 _verifying(false), 478 _verification_mark_bm(0, Mutex::leaf + 1, "CMS_verification_mark_bm_lock"), 479 _completed_initialization(false), 480 _collector_policy(cp), 481 _should_unload_classes(CMSClassUnloadingEnabled), 482 _concurrent_cycles_since_last_unload(0), 483 _roots_scanning_options(GenCollectedHeap::SO_None), 484 _inter_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding), 485 _intra_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding), 486 _gc_tracer_cm(new (ResourceObj::C_HEAP, mtGC) CMSTracer()), 487 _gc_timer_cm(new (ResourceObj::C_HEAP, mtGC) ConcurrentGCTimer()), 488 _cms_start_registered(false) 489 { 490 if (ExplicitGCInvokesConcurrentAndUnloadsClasses) { 491 ExplicitGCInvokesConcurrent = true; 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 warning("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 warning("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 warning("GC/CMS: _conc_workers allocation failure: " 530 "forcing -CMSConcurrentMTEnabled"); 531 CMSConcurrentMTEnabled = false; 532 } else { 533 _conc_workers->initialize_workers(); 534 } 535 } else { 536 CMSConcurrentMTEnabled = false; 537 } 538 } 539 if (!CMSConcurrentMTEnabled) { 540 ConcGCThreads = 0; 541 } else { 542 // Turn off CMSCleanOnEnter optimization temporarily for 543 // the MT case where it's not fixed yet; see 6178663. 544 CMSCleanOnEnter = false; 545 } 546 assert((_conc_workers != NULL) == (ConcGCThreads > 1), 547 "Inconsistency"); 548 549 // Parallel task queues; these are shared for the 550 // concurrent and stop-world phases of CMS, but 551 // are not shared with parallel scavenge (ParNew). 552 { 553 uint i; 554 uint num_queues = MAX2(ParallelGCThreads, ConcGCThreads); 555 556 if ((CMSParallelRemarkEnabled || CMSConcurrentMTEnabled 557 || ParallelRefProcEnabled) 558 && num_queues > 0) { 559 _task_queues = new OopTaskQueueSet(num_queues); 560 if (_task_queues == NULL) { 561 warning("task_queues allocation failure."); 562 return; 563 } 564 _hash_seed = NEW_C_HEAP_ARRAY(int, num_queues, mtGC); 565 typedef Padded<OopTaskQueue> PaddedOopTaskQueue; 566 for (i = 0; i < num_queues; i++) { 567 PaddedOopTaskQueue *q = new PaddedOopTaskQueue(); 568 if (q == NULL) { 569 warning("work_queue allocation failure."); 570 return; 571 } 572 _task_queues->register_queue(i, q); 573 } 574 for (i = 0; i < num_queues; i++) { 575 _task_queues->queue(i)->initialize(); 576 _hash_seed[i] = 17; // copied from ParNew 577 } 578 } 579 } 580 581 _cmsGen ->init_initiating_occupancy(CMSInitiatingOccupancyFraction, CMSTriggerRatio); 582 583 // Clip CMSBootstrapOccupancy between 0 and 100. 584 _bootstrap_occupancy = CMSBootstrapOccupancy / 100.0; 585 586 // Now tell CMS generations the identity of their collector 587 ConcurrentMarkSweepGeneration::set_collector(this); 588 589 // Create & start a CMS thread for this CMS collector 590 _cmsThread = ConcurrentMarkSweepThread::start(this); 591 assert(cmsThread() != NULL, "CMS Thread should have been created"); 592 assert(cmsThread()->collector() == this, 593 "CMS Thread should refer to this gen"); 594 assert(CGC_lock != NULL, "Where's the CGC_lock?"); 595 596 // Support for parallelizing young gen rescan 597 GenCollectedHeap* gch = GenCollectedHeap::heap(); 598 assert(gch->young_gen()->kind() == Generation::ParNew, "CMS can only be used with ParNew"); 599 _young_gen = (ParNewGeneration*)gch->young_gen(); 600 if (gch->supports_inline_contig_alloc()) { 601 _top_addr = gch->top_addr(); 602 _end_addr = gch->end_addr(); 603 assert(_young_gen != NULL, "no _young_gen"); 604 _eden_chunk_index = 0; 605 _eden_chunk_capacity = (_young_gen->max_capacity() + CMSSamplingGrain) / CMSSamplingGrain; 606 _eden_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, _eden_chunk_capacity, mtGC); 607 } 608 609 // Support for parallelizing survivor space rescan 610 if ((CMSParallelRemarkEnabled && CMSParallelSurvivorRemarkEnabled) || CMSParallelInitialMarkEnabled) { 611 const size_t max_plab_samples = 612 _young_gen->max_survivor_size() / (PLAB::min_size() * HeapWordSize); 613 614 _survivor_plab_array = NEW_C_HEAP_ARRAY(ChunkArray, ParallelGCThreads, mtGC); 615 _survivor_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples, mtGC); 616 _cursor = NEW_C_HEAP_ARRAY(size_t, ParallelGCThreads, mtGC); 617 _survivor_chunk_capacity = max_plab_samples; 618 for (uint i = 0; i < ParallelGCThreads; i++) { 619 HeapWord** vec = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples, mtGC); 620 ChunkArray* cur = ::new (&_survivor_plab_array[i]) ChunkArray(vec, max_plab_samples); 621 assert(cur->end() == 0, "Should be 0"); 622 assert(cur->array() == vec, "Should be vec"); 623 assert(cur->capacity() == max_plab_samples, "Error"); 624 } 625 } 626 627 NOT_PRODUCT(_overflow_counter = CMSMarkStackOverflowInterval;) 628 _gc_counters = new CollectorCounters("CMS", 1); 629 _completed_initialization = true; 630 _inter_sweep_timer.start(); // start of time 631 } 632 633 const char* ConcurrentMarkSweepGeneration::name() const { 634 return "concurrent mark-sweep generation"; 635 } 636 void ConcurrentMarkSweepGeneration::update_counters() { 637 if (UsePerfData) { 638 _space_counters->update_all(); 639 _gen_counters->update_all(); 640 } 641 } 642 643 // this is an optimized version of update_counters(). it takes the 644 // used value as a parameter rather than computing it. 645 // 646 void ConcurrentMarkSweepGeneration::update_counters(size_t used) { 647 if (UsePerfData) { 648 _space_counters->update_used(used); 649 _space_counters->update_capacity(); 650 _gen_counters->update_all(); 651 } 652 } 653 654 void ConcurrentMarkSweepGeneration::print() const { 655 Generation::print(); 656 cmsSpace()->print(); 657 } 658 659 #ifndef PRODUCT 660 void ConcurrentMarkSweepGeneration::print_statistics() { 661 cmsSpace()->printFLCensus(0); 662 } 663 #endif 664 665 size_t 666 ConcurrentMarkSweepGeneration::contiguous_available() const { 667 // dld proposes an improvement in precision here. If the committed 668 // part of the space ends in a free block we should add that to 669 // uncommitted size in the calculation below. Will make this 670 // change later, staying with the approximation below for the 671 // time being. -- ysr. 672 return MAX2(_virtual_space.uncommitted_size(), unsafe_max_alloc_nogc()); 673 } 674 675 size_t 676 ConcurrentMarkSweepGeneration::unsafe_max_alloc_nogc() const { 677 return _cmsSpace->max_alloc_in_words() * HeapWordSize; 678 } 679 680 size_t ConcurrentMarkSweepGeneration::max_available() const { 681 return free() + _virtual_space.uncommitted_size(); 682 } 683 684 bool ConcurrentMarkSweepGeneration::promotion_attempt_is_safe(size_t max_promotion_in_bytes) const { 685 size_t available = max_available(); 686 size_t av_promo = (size_t)gc_stats()->avg_promoted()->padded_average(); 687 bool res = (available >= av_promo) || (available >= max_promotion_in_bytes); 688 log_trace(gc, promotion)("CMS: promo attempt is%s safe: available(" SIZE_FORMAT ") %s av_promo(" SIZE_FORMAT "), max_promo(" SIZE_FORMAT ")", 689 res? "":" not", available, res? ">=":"<", av_promo, max_promotion_in_bytes); 690 return res; 691 } 692 693 // At a promotion failure dump information on block layout in heap 694 // (cms old generation). 695 void ConcurrentMarkSweepGeneration::promotion_failure_occurred() { 696 LogHandle(gc, promotion) log; 697 if (log.is_trace()) { 698 ResourceMark rm; 699 cmsSpace()->dump_at_safepoint_with_locks(collector(), log.trace_stream()); 700 } 701 } 702 703 void ConcurrentMarkSweepGeneration::reset_after_compaction() { 704 // Clear the promotion information. These pointers can be adjusted 705 // along with all the other pointers into the heap but 706 // compaction is expected to be a rare event with 707 // a heap using cms so don't do it without seeing the need. 708 for (uint i = 0; i < ParallelGCThreads; i++) { 709 _par_gc_thread_states[i]->promo.reset(); 710 } 711 } 712 713 void ConcurrentMarkSweepGeneration::compute_new_size() { 714 assert_locked_or_safepoint(Heap_lock); 715 716 // If incremental collection failed, we just want to expand 717 // to the limit. 718 if (incremental_collection_failed()) { 719 clear_incremental_collection_failed(); 720 grow_to_reserved(); 721 return; 722 } 723 724 // The heap has been compacted but not reset yet. 725 // Any metric such as free() or used() will be incorrect. 726 727 CardGeneration::compute_new_size(); 728 729 // Reset again after a possible resizing 730 if (did_compact()) { 731 cmsSpace()->reset_after_compaction(); 732 } 733 } 734 735 void ConcurrentMarkSweepGeneration::compute_new_size_free_list() { 736 assert_locked_or_safepoint(Heap_lock); 737 738 // If incremental collection failed, we just want to expand 739 // to the limit. 740 if (incremental_collection_failed()) { 741 clear_incremental_collection_failed(); 742 grow_to_reserved(); 743 return; 744 } 745 746 double free_percentage = ((double) free()) / capacity(); 747 double desired_free_percentage = (double) MinHeapFreeRatio / 100; 748 double maximum_free_percentage = (double) MaxHeapFreeRatio / 100; 749 750 // compute expansion delta needed for reaching desired free percentage 751 if (free_percentage < desired_free_percentage) { 752 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 753 assert(desired_capacity >= capacity(), "invalid expansion size"); 754 size_t expand_bytes = MAX2(desired_capacity - capacity(), MinHeapDeltaBytes); 755 LogHandle(gc) log; 756 if (log.is_trace()) { 757 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 758 log.trace("From compute_new_size: "); 759 log.trace(" Free fraction %f", free_percentage); 760 log.trace(" Desired free fraction %f", desired_free_percentage); 761 log.trace(" Maximum free fraction %f", maximum_free_percentage); 762 log.trace(" Capacity " SIZE_FORMAT, capacity() / 1000); 763 log.trace(" Desired capacity " SIZE_FORMAT, desired_capacity / 1000); 764 GenCollectedHeap* gch = GenCollectedHeap::heap(); 765 assert(gch->is_old_gen(this), "The CMS generation should always be the old generation"); 766 size_t young_size = gch->young_gen()->capacity(); 767 log.trace(" Young gen size " SIZE_FORMAT, young_size / 1000); 768 log.trace(" unsafe_max_alloc_nogc " SIZE_FORMAT, unsafe_max_alloc_nogc() / 1000); 769 log.trace(" contiguous available " SIZE_FORMAT, contiguous_available() / 1000); 770 log.trace(" Expand by " SIZE_FORMAT " (bytes)", expand_bytes); 771 } 772 // safe if expansion fails 773 expand_for_gc_cause(expand_bytes, 0, CMSExpansionCause::_satisfy_free_ratio); 774 log.trace(" Expanded free fraction %f", ((double) free()) / capacity()); 775 } else { 776 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 777 assert(desired_capacity <= capacity(), "invalid expansion size"); 778 size_t shrink_bytes = capacity() - desired_capacity; 779 // Don't shrink unless the delta is greater than the minimum shrink we want 780 if (shrink_bytes >= MinHeapDeltaBytes) { 781 shrink_free_list_by(shrink_bytes); 782 } 783 } 784 } 785 786 Mutex* ConcurrentMarkSweepGeneration::freelistLock() const { 787 return cmsSpace()->freelistLock(); 788 } 789 790 HeapWord* ConcurrentMarkSweepGeneration::allocate(size_t size, bool tlab) { 791 CMSSynchronousYieldRequest yr; 792 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 793 return have_lock_and_allocate(size, tlab); 794 } 795 796 HeapWord* ConcurrentMarkSweepGeneration::have_lock_and_allocate(size_t size, 797 bool tlab /* ignored */) { 798 assert_lock_strong(freelistLock()); 799 size_t adjustedSize = CompactibleFreeListSpace::adjustObjectSize(size); 800 HeapWord* res = cmsSpace()->allocate(adjustedSize); 801 // Allocate the object live (grey) if the background collector has 802 // started marking. This is necessary because the marker may 803 // have passed this address and consequently this object will 804 // not otherwise be greyed and would be incorrectly swept up. 805 // Note that if this object contains references, the writing 806 // of those references will dirty the card containing this object 807 // allowing the object to be blackened (and its references scanned) 808 // either during a preclean phase or at the final checkpoint. 809 if (res != NULL) { 810 // We may block here with an uninitialized object with 811 // its mark-bit or P-bits not yet set. Such objects need 812 // to be safely navigable by block_start(). 813 assert(oop(res)->klass_or_null() == NULL, "Object should be uninitialized here."); 814 assert(!((FreeChunk*)res)->is_free(), "Error, block will look free but show wrong size"); 815 collector()->direct_allocated(res, adjustedSize); 816 _direct_allocated_words += adjustedSize; 817 // allocation counters 818 NOT_PRODUCT( 819 _numObjectsAllocated++; 820 _numWordsAllocated += (int)adjustedSize; 821 ) 822 } 823 return res; 824 } 825 826 // In the case of direct allocation by mutators in a generation that 827 // is being concurrently collected, the object must be allocated 828 // live (grey) if the background collector has started marking. 829 // This is necessary because the marker may 830 // have passed this address and consequently this object will 831 // not otherwise be greyed and would be incorrectly swept up. 832 // Note that if this object contains references, the writing 833 // of those references will dirty the card containing this object 834 // allowing the object to be blackened (and its references scanned) 835 // either during a preclean phase or at the final checkpoint. 836 void CMSCollector::direct_allocated(HeapWord* start, size_t size) { 837 assert(_markBitMap.covers(start, size), "Out of bounds"); 838 if (_collectorState >= Marking) { 839 MutexLockerEx y(_markBitMap.lock(), 840 Mutex::_no_safepoint_check_flag); 841 // [see comments preceding SweepClosure::do_blk() below for details] 842 // 843 // Can the P-bits be deleted now? JJJ 844 // 845 // 1. need to mark the object as live so it isn't collected 846 // 2. need to mark the 2nd bit to indicate the object may be uninitialized 847 // 3. need to mark the end of the object so marking, precleaning or sweeping 848 // can skip over uninitialized or unparsable objects. An allocated 849 // object is considered uninitialized for our purposes as long as 850 // its klass word is NULL. All old gen objects are parsable 851 // as soon as they are initialized.) 852 _markBitMap.mark(start); // object is live 853 _markBitMap.mark(start + 1); // object is potentially uninitialized? 854 _markBitMap.mark(start + size - 1); 855 // mark end of object 856 } 857 // check that oop looks uninitialized 858 assert(oop(start)->klass_or_null() == NULL, "_klass should be NULL"); 859 } 860 861 void CMSCollector::promoted(bool par, HeapWord* start, 862 bool is_obj_array, size_t obj_size) { 863 assert(_markBitMap.covers(start), "Out of bounds"); 864 // See comment in direct_allocated() about when objects should 865 // be allocated live. 866 if (_collectorState >= Marking) { 867 // we already hold the marking bit map lock, taken in 868 // the prologue 869 if (par) { 870 _markBitMap.par_mark(start); 871 } else { 872 _markBitMap.mark(start); 873 } 874 // We don't need to mark the object as uninitialized (as 875 // in direct_allocated above) because this is being done with the 876 // world stopped and the object will be initialized by the 877 // time the marking, precleaning or sweeping get to look at it. 878 // But see the code for copying objects into the CMS generation, 879 // where we need to ensure that concurrent readers of the 880 // block offset table are able to safely navigate a block that 881 // is in flux from being free to being allocated (and in 882 // transition while being copied into) and subsequently 883 // becoming a bona-fide object when the copy/promotion is complete. 884 assert(SafepointSynchronize::is_at_safepoint(), 885 "expect promotion only at safepoints"); 886 887 if (_collectorState < Sweeping) { 888 // Mark the appropriate cards in the modUnionTable, so that 889 // this object gets scanned before the sweep. If this is 890 // not done, CMS generation references in the object might 891 // not get marked. 892 // For the case of arrays, which are otherwise precisely 893 // marked, we need to dirty the entire array, not just its head. 894 if (is_obj_array) { 895 // The [par_]mark_range() method expects mr.end() below to 896 // be aligned to the granularity of a bit's representation 897 // in the heap. In the case of the MUT below, that's a 898 // card size. 899 MemRegion mr(start, 900 (HeapWord*)round_to((intptr_t)(start + obj_size), 901 CardTableModRefBS::card_size /* bytes */)); 902 if (par) { 903 _modUnionTable.par_mark_range(mr); 904 } else { 905 _modUnionTable.mark_range(mr); 906 } 907 } else { // not an obj array; we can just mark the head 908 if (par) { 909 _modUnionTable.par_mark(start); 910 } else { 911 _modUnionTable.mark(start); 912 } 913 } 914 } 915 } 916 } 917 918 oop ConcurrentMarkSweepGeneration::promote(oop obj, size_t obj_size) { 919 assert(obj_size == (size_t)obj->size(), "bad obj_size passed in"); 920 // allocate, copy and if necessary update promoinfo -- 921 // delegate to underlying space. 922 assert_lock_strong(freelistLock()); 923 924 #ifndef PRODUCT 925 if (GenCollectedHeap::heap()->promotion_should_fail()) { 926 return NULL; 927 } 928 #endif // #ifndef PRODUCT 929 930 oop res = _cmsSpace->promote(obj, obj_size); 931 if (res == NULL) { 932 // expand and retry 933 size_t s = _cmsSpace->expansionSpaceRequired(obj_size); // HeapWords 934 expand_for_gc_cause(s*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_satisfy_promotion); 935 // Since this is the old generation, we don't try to promote 936 // into a more senior generation. 937 res = _cmsSpace->promote(obj, obj_size); 938 } 939 if (res != NULL) { 940 // See comment in allocate() about when objects should 941 // be allocated live. 942 assert(obj->is_oop(), "Will dereference klass pointer below"); 943 collector()->promoted(false, // Not parallel 944 (HeapWord*)res, obj->is_objArray(), obj_size); 945 // promotion counters 946 NOT_PRODUCT( 947 _numObjectsPromoted++; 948 _numWordsPromoted += 949 (int)(CompactibleFreeListSpace::adjustObjectSize(obj->size())); 950 ) 951 } 952 return res; 953 } 954 955 956 // IMPORTANT: Notes on object size recognition in CMS. 957 // --------------------------------------------------- 958 // A block of storage in the CMS generation is always in 959 // one of three states. A free block (FREE), an allocated 960 // object (OBJECT) whose size() method reports the correct size, 961 // and an intermediate state (TRANSIENT) in which its size cannot 962 // be accurately determined. 963 // STATE IDENTIFICATION: (32 bit and 64 bit w/o COOPS) 964 // ----------------------------------------------------- 965 // FREE: klass_word & 1 == 1; mark_word holds block size 966 // 967 // OBJECT: klass_word installed; klass_word != 0 && klass_word & 1 == 0; 968 // obj->size() computes correct size 969 // 970 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT 971 // 972 // STATE IDENTIFICATION: (64 bit+COOPS) 973 // ------------------------------------ 974 // FREE: mark_word & CMS_FREE_BIT == 1; mark_word & ~CMS_FREE_BIT gives block_size 975 // 976 // OBJECT: klass_word installed; klass_word != 0; 977 // obj->size() computes correct size 978 // 979 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT 980 // 981 // 982 // STATE TRANSITION DIAGRAM 983 // 984 // mut / parnew mut / parnew 985 // FREE --------------------> TRANSIENT ---------------------> OBJECT --| 986 // ^ | 987 // |------------------------ DEAD <------------------------------------| 988 // sweep mut 989 // 990 // While a block is in TRANSIENT state its size cannot be determined 991 // so readers will either need to come back later or stall until 992 // the size can be determined. Note that for the case of direct 993 // allocation, P-bits, when available, may be used to determine the 994 // size of an object that may not yet have been initialized. 995 996 // Things to support parallel young-gen collection. 997 oop 998 ConcurrentMarkSweepGeneration::par_promote(int thread_num, 999 oop old, markOop m, 1000 size_t word_sz) { 1001 #ifndef PRODUCT 1002 if (GenCollectedHeap::heap()->promotion_should_fail()) { 1003 return NULL; 1004 } 1005 #endif // #ifndef PRODUCT 1006 1007 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1008 PromotionInfo* promoInfo = &ps->promo; 1009 // if we are tracking promotions, then first ensure space for 1010 // promotion (including spooling space for saving header if necessary). 1011 // then allocate and copy, then track promoted info if needed. 1012 // When tracking (see PromotionInfo::track()), the mark word may 1013 // be displaced and in this case restoration of the mark word 1014 // occurs in the (oop_since_save_marks_)iterate phase. 1015 if (promoInfo->tracking() && !promoInfo->ensure_spooling_space()) { 1016 // Out of space for allocating spooling buffers; 1017 // try expanding and allocating spooling buffers. 1018 if (!expand_and_ensure_spooling_space(promoInfo)) { 1019 return NULL; 1020 } 1021 } 1022 assert(promoInfo->has_spooling_space(), "Control point invariant"); 1023 const size_t alloc_sz = CompactibleFreeListSpace::adjustObjectSize(word_sz); 1024 HeapWord* obj_ptr = ps->lab.alloc(alloc_sz); 1025 if (obj_ptr == NULL) { 1026 obj_ptr = expand_and_par_lab_allocate(ps, alloc_sz); 1027 if (obj_ptr == NULL) { 1028 return NULL; 1029 } 1030 } 1031 oop obj = oop(obj_ptr); 1032 OrderAccess::storestore(); 1033 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1034 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1035 // IMPORTANT: See note on object initialization for CMS above. 1036 // Otherwise, copy the object. Here we must be careful to insert the 1037 // klass pointer last, since this marks the block as an allocated object. 1038 // Except with compressed oops it's the mark word. 1039 HeapWord* old_ptr = (HeapWord*)old; 1040 // Restore the mark word copied above. 1041 obj->set_mark(m); 1042 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1043 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1044 OrderAccess::storestore(); 1045 1046 if (UseCompressedClassPointers) { 1047 // Copy gap missed by (aligned) header size calculation below 1048 obj->set_klass_gap(old->klass_gap()); 1049 } 1050 if (word_sz > (size_t)oopDesc::header_size()) { 1051 Copy::aligned_disjoint_words(old_ptr + oopDesc::header_size(), 1052 obj_ptr + oopDesc::header_size(), 1053 word_sz - oopDesc::header_size()); 1054 } 1055 1056 // Now we can track the promoted object, if necessary. We take care 1057 // to delay the transition from uninitialized to full object 1058 // (i.e., insertion of klass pointer) until after, so that it 1059 // atomically becomes a promoted object. 1060 if (promoInfo->tracking()) { 1061 promoInfo->track((PromotedObject*)obj, old->klass()); 1062 } 1063 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1064 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1065 assert(old->is_oop(), "Will use and dereference old klass ptr below"); 1066 1067 // Finally, install the klass pointer (this should be volatile). 1068 OrderAccess::storestore(); 1069 obj->set_klass(old->klass()); 1070 // We should now be able to calculate the right size for this object 1071 assert(obj->is_oop() && obj->size() == (int)word_sz, "Error, incorrect size computed for promoted object"); 1072 1073 collector()->promoted(true, // parallel 1074 obj_ptr, old->is_objArray(), word_sz); 1075 1076 NOT_PRODUCT( 1077 Atomic::inc_ptr(&_numObjectsPromoted); 1078 Atomic::add_ptr(alloc_sz, &_numWordsPromoted); 1079 ) 1080 1081 return obj; 1082 } 1083 1084 void 1085 ConcurrentMarkSweepGeneration:: 1086 par_promote_alloc_done(int thread_num) { 1087 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1088 ps->lab.retire(thread_num); 1089 } 1090 1091 void 1092 ConcurrentMarkSweepGeneration:: 1093 par_oop_since_save_marks_iterate_done(int thread_num) { 1094 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1095 ParScanWithoutBarrierClosure* dummy_cl = NULL; 1096 ps->promo.promoted_oops_iterate_nv(dummy_cl); 1097 } 1098 1099 bool ConcurrentMarkSweepGeneration::should_collect(bool full, 1100 size_t size, 1101 bool tlab) 1102 { 1103 // We allow a STW collection only if a full 1104 // collection was requested. 1105 return full || should_allocate(size, tlab); // FIX ME !!! 1106 // This and promotion failure handling are connected at the 1107 // hip and should be fixed by untying them. 1108 } 1109 1110 bool CMSCollector::shouldConcurrentCollect() { 1111 if (_full_gc_requested) { 1112 log_trace(gc)("CMSCollector: collect because of explicit gc request (or gc_locker)"); 1113 return true; 1114 } 1115 1116 FreelistLocker x(this); 1117 // ------------------------------------------------------------------ 1118 // Print out lots of information which affects the initiation of 1119 // a collection. 1120 LogHandle(gc) log; 1121 if (log.is_trace() && stats().valid()) { 1122 log.trace("CMSCollector shouldConcurrentCollect: "); 1123 ResourceMark rm; 1124 stats().print_on(log.debug_stream()); 1125 log.trace("time_until_cms_gen_full %3.7f", stats().time_until_cms_gen_full()); 1126 log.trace("free=" SIZE_FORMAT, _cmsGen->free()); 1127 log.trace("contiguous_available=" SIZE_FORMAT, _cmsGen->contiguous_available()); 1128 log.trace("promotion_rate=%g", stats().promotion_rate()); 1129 log.trace("cms_allocation_rate=%g", stats().cms_allocation_rate()); 1130 log.trace("occupancy=%3.7f", _cmsGen->occupancy()); 1131 log.trace("initiatingOccupancy=%3.7f", _cmsGen->initiating_occupancy()); 1132 log.trace("cms_time_since_begin=%3.7f", stats().cms_time_since_begin()); 1133 log.trace("cms_time_since_end=%3.7f", stats().cms_time_since_end()); 1134 log.trace("metadata initialized %d", MetaspaceGC::should_concurrent_collect()); 1135 } 1136 // ------------------------------------------------------------------ 1137 1138 // If the estimated time to complete a cms collection (cms_duration()) 1139 // is less than the estimated time remaining until the cms generation 1140 // is full, start a collection. 1141 if (!UseCMSInitiatingOccupancyOnly) { 1142 if (stats().valid()) { 1143 if (stats().time_until_cms_start() == 0.0) { 1144 return true; 1145 } 1146 } else { 1147 // We want to conservatively collect somewhat early in order 1148 // to try and "bootstrap" our CMS/promotion statistics; 1149 // this branch will not fire after the first successful CMS 1150 // collection because the stats should then be valid. 1151 if (_cmsGen->occupancy() >= _bootstrap_occupancy) { 1152 log_trace(gc)(" CMSCollector: collect for bootstrapping statistics: occupancy = %f, boot occupancy = %f", 1153 _cmsGen->occupancy(), _bootstrap_occupancy); 1154 return true; 1155 } 1156 } 1157 } 1158 1159 // Otherwise, we start a collection cycle if 1160 // old gen want a collection cycle started. Each may use 1161 // an appropriate criterion for making this decision. 1162 // XXX We need to make sure that the gen expansion 1163 // criterion dovetails well with this. XXX NEED TO FIX THIS 1164 if (_cmsGen->should_concurrent_collect()) { 1165 log_trace(gc)("CMS old gen initiated"); 1166 return true; 1167 } 1168 1169 // We start a collection if we believe an incremental collection may fail; 1170 // this is not likely to be productive in practice because it's probably too 1171 // late anyway. 1172 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1173 assert(gch->collector_policy()->is_generation_policy(), 1174 "You may want to check the correctness of the following"); 1175 if (gch->incremental_collection_will_fail(true /* consult_young */)) { 1176 log_trace(gc)("CMSCollector: collect because incremental collection will fail "); 1177 return true; 1178 } 1179 1180 if (MetaspaceGC::should_concurrent_collect()) { 1181 log_trace(gc)("CMSCollector: collect for metadata allocation "); 1182 return true; 1183 } 1184 1185 // CMSTriggerInterval starts a CMS cycle if enough time has passed. 1186 if (CMSTriggerInterval >= 0) { 1187 if (CMSTriggerInterval == 0) { 1188 // Trigger always 1189 return true; 1190 } 1191 1192 // Check the CMS time since begin (we do not check the stats validity 1193 // as we want to be able to trigger the first CMS cycle as well) 1194 if (stats().cms_time_since_begin() >= (CMSTriggerInterval / ((double) MILLIUNITS))) { 1195 if (stats().valid()) { 1196 log_trace(gc)("CMSCollector: collect because of trigger interval (time since last begin %3.7f secs)", 1197 stats().cms_time_since_begin()); 1198 } else { 1199 log_trace(gc)("CMSCollector: collect because of trigger interval (first collection)"); 1200 } 1201 return true; 1202 } 1203 } 1204 1205 return false; 1206 } 1207 1208 void CMSCollector::set_did_compact(bool v) { _cmsGen->set_did_compact(v); } 1209 1210 // Clear _expansion_cause fields of constituent generations 1211 void CMSCollector::clear_expansion_cause() { 1212 _cmsGen->clear_expansion_cause(); 1213 } 1214 1215 // We should be conservative in starting a collection cycle. To 1216 // start too eagerly runs the risk of collecting too often in the 1217 // extreme. To collect too rarely falls back on full collections, 1218 // which works, even if not optimum in terms of concurrent work. 1219 // As a work around for too eagerly collecting, use the flag 1220 // UseCMSInitiatingOccupancyOnly. This also has the advantage of 1221 // giving the user an easily understandable way of controlling the 1222 // collections. 1223 // We want to start a new collection cycle if any of the following 1224 // conditions hold: 1225 // . our current occupancy exceeds the configured initiating occupancy 1226 // for this generation, or 1227 // . we recently needed to expand this space and have not, since that 1228 // expansion, done a collection of this generation, or 1229 // . the underlying space believes that it may be a good idea to initiate 1230 // a concurrent collection (this may be based on criteria such as the 1231 // following: the space uses linear allocation and linear allocation is 1232 // going to fail, or there is believed to be excessive fragmentation in 1233 // the generation, etc... or ... 1234 // [.(currently done by CMSCollector::shouldConcurrentCollect() only for 1235 // the case of the old generation; see CR 6543076): 1236 // we may be approaching a point at which allocation requests may fail because 1237 // we will be out of sufficient free space given allocation rate estimates.] 1238 bool ConcurrentMarkSweepGeneration::should_concurrent_collect() const { 1239 1240 assert_lock_strong(freelistLock()); 1241 if (occupancy() > initiating_occupancy()) { 1242 log_trace(gc)(" %s: collect because of occupancy %f / %f ", 1243 short_name(), occupancy(), initiating_occupancy()); 1244 return true; 1245 } 1246 if (UseCMSInitiatingOccupancyOnly) { 1247 return false; 1248 } 1249 if (expansion_cause() == CMSExpansionCause::_satisfy_allocation) { 1250 log_trace(gc)(" %s: collect because expanded for allocation ", short_name()); 1251 return true; 1252 } 1253 return false; 1254 } 1255 1256 void ConcurrentMarkSweepGeneration::collect(bool full, 1257 bool clear_all_soft_refs, 1258 size_t size, 1259 bool tlab) 1260 { 1261 collector()->collect(full, clear_all_soft_refs, size, tlab); 1262 } 1263 1264 void CMSCollector::collect(bool full, 1265 bool clear_all_soft_refs, 1266 size_t size, 1267 bool tlab) 1268 { 1269 // The following "if" branch is present for defensive reasons. 1270 // In the current uses of this interface, it can be replaced with: 1271 // assert(!GC_locker.is_active(), "Can't be called otherwise"); 1272 // But I am not placing that assert here to allow future 1273 // generality in invoking this interface. 1274 if (GC_locker::is_active()) { 1275 // A consistency test for GC_locker 1276 assert(GC_locker::needs_gc(), "Should have been set already"); 1277 // Skip this foreground collection, instead 1278 // expanding the heap if necessary. 1279 // Need the free list locks for the call to free() in compute_new_size() 1280 compute_new_size(); 1281 return; 1282 } 1283 acquire_control_and_collect(full, clear_all_soft_refs); 1284 } 1285 1286 void CMSCollector::request_full_gc(unsigned int full_gc_count, GCCause::Cause cause) { 1287 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1288 unsigned int gc_count = gch->total_full_collections(); 1289 if (gc_count == full_gc_count) { 1290 MutexLockerEx y(CGC_lock, Mutex::_no_safepoint_check_flag); 1291 _full_gc_requested = true; 1292 _full_gc_cause = cause; 1293 CGC_lock->notify(); // nudge CMS thread 1294 } else { 1295 assert(gc_count > full_gc_count, "Error: causal loop"); 1296 } 1297 } 1298 1299 bool CMSCollector::is_external_interruption() { 1300 GCCause::Cause cause = GenCollectedHeap::heap()->gc_cause(); 1301 return GCCause::is_user_requested_gc(cause) || 1302 GCCause::is_serviceability_requested_gc(cause); 1303 } 1304 1305 void CMSCollector::report_concurrent_mode_interruption() { 1306 if (is_external_interruption()) { 1307 log_debug(gc)("Concurrent mode interrupted"); 1308 } else { 1309 log_debug(gc)("Concurrent mode failure"); 1310 _gc_tracer_cm->report_concurrent_mode_failure(); 1311 } 1312 } 1313 1314 1315 // The foreground and background collectors need to coordinate in order 1316 // to make sure that they do not mutually interfere with CMS collections. 1317 // When a background collection is active, 1318 // the foreground collector may need to take over (preempt) and 1319 // synchronously complete an ongoing collection. Depending on the 1320 // frequency of the background collections and the heap usage 1321 // of the application, this preemption can be seldom or frequent. 1322 // There are only certain 1323 // points in the background collection that the "collection-baton" 1324 // can be passed to the foreground collector. 1325 // 1326 // The foreground collector will wait for the baton before 1327 // starting any part of the collection. The foreground collector 1328 // will only wait at one location. 1329 // 1330 // The background collector will yield the baton before starting a new 1331 // phase of the collection (e.g., before initial marking, marking from roots, 1332 // precleaning, final re-mark, sweep etc.) This is normally done at the head 1333 // of the loop which switches the phases. The background collector does some 1334 // of the phases (initial mark, final re-mark) with the world stopped. 1335 // Because of locking involved in stopping the world, 1336 // the foreground collector should not block waiting for the background 1337 // collector when it is doing a stop-the-world phase. The background 1338 // collector will yield the baton at an additional point just before 1339 // it enters a stop-the-world phase. Once the world is stopped, the 1340 // background collector checks the phase of the collection. If the 1341 // phase has not changed, it proceeds with the collection. If the 1342 // phase has changed, it skips that phase of the collection. See 1343 // the comments on the use of the Heap_lock in collect_in_background(). 1344 // 1345 // Variable used in baton passing. 1346 // _foregroundGCIsActive - Set to true by the foreground collector when 1347 // it wants the baton. The foreground clears it when it has finished 1348 // the collection. 1349 // _foregroundGCShouldWait - Set to true by the background collector 1350 // when it is running. The foreground collector waits while 1351 // _foregroundGCShouldWait is true. 1352 // CGC_lock - monitor used to protect access to the above variables 1353 // and to notify the foreground and background collectors. 1354 // _collectorState - current state of the CMS collection. 1355 // 1356 // The foreground collector 1357 // acquires the CGC_lock 1358 // sets _foregroundGCIsActive 1359 // waits on the CGC_lock for _foregroundGCShouldWait to be false 1360 // various locks acquired in preparation for the collection 1361 // are released so as not to block the background collector 1362 // that is in the midst of a collection 1363 // proceeds with the collection 1364 // clears _foregroundGCIsActive 1365 // returns 1366 // 1367 // The background collector in a loop iterating on the phases of the 1368 // collection 1369 // acquires the CGC_lock 1370 // sets _foregroundGCShouldWait 1371 // if _foregroundGCIsActive is set 1372 // clears _foregroundGCShouldWait, notifies _CGC_lock 1373 // waits on _CGC_lock for _foregroundGCIsActive to become false 1374 // and exits the loop. 1375 // otherwise 1376 // proceed with that phase of the collection 1377 // if the phase is a stop-the-world phase, 1378 // yield the baton once more just before enqueueing 1379 // the stop-world CMS operation (executed by the VM thread). 1380 // returns after all phases of the collection are done 1381 // 1382 1383 void CMSCollector::acquire_control_and_collect(bool full, 1384 bool clear_all_soft_refs) { 1385 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint"); 1386 assert(!Thread::current()->is_ConcurrentGC_thread(), 1387 "shouldn't try to acquire control from self!"); 1388 1389 // Start the protocol for acquiring control of the 1390 // collection from the background collector (aka CMS thread). 1391 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 1392 "VM thread should have CMS token"); 1393 // Remember the possibly interrupted state of an ongoing 1394 // concurrent collection 1395 CollectorState first_state = _collectorState; 1396 1397 // Signal to a possibly ongoing concurrent collection that 1398 // we want to do a foreground collection. 1399 _foregroundGCIsActive = true; 1400 1401 // release locks and wait for a notify from the background collector 1402 // releasing the locks in only necessary for phases which 1403 // do yields to improve the granularity of the collection. 1404 assert_lock_strong(bitMapLock()); 1405 // We need to lock the Free list lock for the space that we are 1406 // currently collecting. 1407 assert(haveFreelistLocks(), "Must be holding free list locks"); 1408 bitMapLock()->unlock(); 1409 releaseFreelistLocks(); 1410 { 1411 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1412 if (_foregroundGCShouldWait) { 1413 // We are going to be waiting for action for the CMS thread; 1414 // it had better not be gone (for instance at shutdown)! 1415 assert(ConcurrentMarkSweepThread::cmst() != NULL, 1416 "CMS thread must be running"); 1417 // Wait here until the background collector gives us the go-ahead 1418 ConcurrentMarkSweepThread::clear_CMS_flag( 1419 ConcurrentMarkSweepThread::CMS_vm_has_token); // release token 1420 // Get a possibly blocked CMS thread going: 1421 // Note that we set _foregroundGCIsActive true above, 1422 // without protection of the CGC_lock. 1423 CGC_lock->notify(); 1424 assert(!ConcurrentMarkSweepThread::vm_thread_wants_cms_token(), 1425 "Possible deadlock"); 1426 while (_foregroundGCShouldWait) { 1427 // wait for notification 1428 CGC_lock->wait(Mutex::_no_safepoint_check_flag); 1429 // Possibility of delay/starvation here, since CMS token does 1430 // not know to give priority to VM thread? Actually, i think 1431 // there wouldn't be any delay/starvation, but the proof of 1432 // that "fact" (?) appears non-trivial. XXX 20011219YSR 1433 } 1434 ConcurrentMarkSweepThread::set_CMS_flag( 1435 ConcurrentMarkSweepThread::CMS_vm_has_token); 1436 } 1437 } 1438 // The CMS_token is already held. Get back the other locks. 1439 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 1440 "VM thread should have CMS token"); 1441 getFreelistLocks(); 1442 bitMapLock()->lock_without_safepoint_check(); 1443 log_debug(gc, state)("CMS foreground collector has asked for control " INTPTR_FORMAT " with first state %d", 1444 p2i(Thread::current()), first_state); 1445 log_debug(gc, state)(" gets control with state %d", _collectorState); 1446 1447 // Inform cms gen if this was due to partial collection failing. 1448 // The CMS gen may use this fact to determine its expansion policy. 1449 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1450 if (gch->incremental_collection_will_fail(false /* don't consult_young */)) { 1451 assert(!_cmsGen->incremental_collection_failed(), 1452 "Should have been noticed, reacted to and cleared"); 1453 _cmsGen->set_incremental_collection_failed(); 1454 } 1455 1456 if (first_state > Idling) { 1457 report_concurrent_mode_interruption(); 1458 } 1459 1460 set_did_compact(true); 1461 1462 // If the collection is being acquired from the background 1463 // collector, there may be references on the discovered 1464 // references lists. Abandon those references, since some 1465 // of them may have become unreachable after concurrent 1466 // discovery; the STW compacting collector will redo discovery 1467 // more precisely, without being subject to floating garbage. 1468 // Leaving otherwise unreachable references in the discovered 1469 // lists would require special handling. 1470 ref_processor()->disable_discovery(); 1471 ref_processor()->abandon_partial_discovery(); 1472 ref_processor()->verify_no_references_recorded(); 1473 1474 if (first_state > Idling) { 1475 save_heap_summary(); 1476 } 1477 1478 do_compaction_work(clear_all_soft_refs); 1479 1480 // Has the GC time limit been exceeded? 1481 size_t max_eden_size = _young_gen->max_eden_size(); 1482 GCCause::Cause gc_cause = gch->gc_cause(); 1483 size_policy()->check_gc_overhead_limit(_young_gen->used(), 1484 _young_gen->eden()->used(), 1485 _cmsGen->max_capacity(), 1486 max_eden_size, 1487 full, 1488 gc_cause, 1489 gch->collector_policy()); 1490 1491 // Reset the expansion cause, now that we just completed 1492 // a collection cycle. 1493 clear_expansion_cause(); 1494 _foregroundGCIsActive = false; 1495 return; 1496 } 1497 1498 // Resize the tenured generation 1499 // after obtaining the free list locks for the 1500 // two generations. 1501 void CMSCollector::compute_new_size() { 1502 assert_locked_or_safepoint(Heap_lock); 1503 FreelistLocker z(this); 1504 MetaspaceGC::compute_new_size(); 1505 _cmsGen->compute_new_size_free_list(); 1506 } 1507 1508 // A work method used by the foreground collector to do 1509 // a mark-sweep-compact. 1510 void CMSCollector::do_compaction_work(bool clear_all_soft_refs) { 1511 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1512 1513 STWGCTimer* gc_timer = GenMarkSweep::gc_timer(); 1514 gc_timer->register_gc_start(); 1515 1516 SerialOldTracer* gc_tracer = GenMarkSweep::gc_tracer(); 1517 gc_tracer->report_gc_start(gch->gc_cause(), gc_timer->gc_start()); 1518 1519 GCTraceTime(Trace, gc) t("CMS:MSC"); 1520 1521 // Temporarily widen the span of the weak reference processing to 1522 // the entire heap. 1523 MemRegion new_span(GenCollectedHeap::heap()->reserved_region()); 1524 ReferenceProcessorSpanMutator rp_mut_span(ref_processor(), new_span); 1525 // Temporarily, clear the "is_alive_non_header" field of the 1526 // reference processor. 1527 ReferenceProcessorIsAliveMutator rp_mut_closure(ref_processor(), NULL); 1528 // Temporarily make reference _processing_ single threaded (non-MT). 1529 ReferenceProcessorMTProcMutator rp_mut_mt_processing(ref_processor(), false); 1530 // Temporarily make refs discovery atomic 1531 ReferenceProcessorAtomicMutator rp_mut_atomic(ref_processor(), true); 1532 // Temporarily make reference _discovery_ single threaded (non-MT) 1533 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false); 1534 1535 ref_processor()->set_enqueuing_is_done(false); 1536 ref_processor()->enable_discovery(); 1537 ref_processor()->setup_policy(clear_all_soft_refs); 1538 // If an asynchronous collection finishes, the _modUnionTable is 1539 // all clear. If we are assuming the collection from an asynchronous 1540 // collection, clear the _modUnionTable. 1541 assert(_collectorState != Idling || _modUnionTable.isAllClear(), 1542 "_modUnionTable should be clear if the baton was not passed"); 1543 _modUnionTable.clear_all(); 1544 assert(_collectorState != Idling || _ct->klass_rem_set()->mod_union_is_clear(), 1545 "mod union for klasses should be clear if the baton was passed"); 1546 _ct->klass_rem_set()->clear_mod_union(); 1547 1548 // We must adjust the allocation statistics being maintained 1549 // in the free list space. We do so by reading and clearing 1550 // the sweep timer and updating the block flux rate estimates below. 1551 assert(!_intra_sweep_timer.is_active(), "_intra_sweep_timer should be inactive"); 1552 if (_inter_sweep_timer.is_active()) { 1553 _inter_sweep_timer.stop(); 1554 // Note that we do not use this sample to update the _inter_sweep_estimate. 1555 _cmsGen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()), 1556 _inter_sweep_estimate.padded_average(), 1557 _intra_sweep_estimate.padded_average()); 1558 } 1559 1560 GenMarkSweep::invoke_at_safepoint(ref_processor(), clear_all_soft_refs); 1561 #ifdef ASSERT 1562 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 1563 size_t free_size = cms_space->free(); 1564 assert(free_size == 1565 pointer_delta(cms_space->end(), cms_space->compaction_top()) 1566 * HeapWordSize, 1567 "All the free space should be compacted into one chunk at top"); 1568 assert(cms_space->dictionary()->total_chunk_size( 1569 debug_only(cms_space->freelistLock())) == 0 || 1570 cms_space->totalSizeInIndexedFreeLists() == 0, 1571 "All the free space should be in a single chunk"); 1572 size_t num = cms_space->totalCount(); 1573 assert((free_size == 0 && num == 0) || 1574 (free_size > 0 && (num == 1 || num == 2)), 1575 "There should be at most 2 free chunks after compaction"); 1576 #endif // ASSERT 1577 _collectorState = Resetting; 1578 assert(_restart_addr == NULL, 1579 "Should have been NULL'd before baton was passed"); 1580 reset_stw(); 1581 _cmsGen->reset_after_compaction(); 1582 _concurrent_cycles_since_last_unload = 0; 1583 1584 // Clear any data recorded in the PLAB chunk arrays. 1585 if (_survivor_plab_array != NULL) { 1586 reset_survivor_plab_arrays(); 1587 } 1588 1589 // Adjust the per-size allocation stats for the next epoch. 1590 _cmsGen->cmsSpace()->endSweepFLCensus(sweep_count() /* fake */); 1591 // Restart the "inter sweep timer" for the next epoch. 1592 _inter_sweep_timer.reset(); 1593 _inter_sweep_timer.start(); 1594 1595 gc_timer->register_gc_end(); 1596 1597 gc_tracer->report_gc_end(gc_timer->gc_end(), gc_timer->time_partitions()); 1598 1599 // For a mark-sweep-compact, compute_new_size() will be called 1600 // in the heap's do_collection() method. 1601 } 1602 1603 void CMSCollector::print_eden_and_survivor_chunk_arrays() { 1604 LogHandle(gc, heap) log; 1605 if (!log.is_trace()) { 1606 return; 1607 } 1608 1609 ContiguousSpace* eden_space = _young_gen->eden(); 1610 ContiguousSpace* from_space = _young_gen->from(); 1611 ContiguousSpace* to_space = _young_gen->to(); 1612 // Eden 1613 if (_eden_chunk_array != NULL) { 1614 log.trace("eden " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")", 1615 p2i(eden_space->bottom()), p2i(eden_space->top()), 1616 p2i(eden_space->end()), eden_space->capacity()); 1617 log.trace("_eden_chunk_index=" SIZE_FORMAT ", _eden_chunk_capacity=" SIZE_FORMAT, 1618 _eden_chunk_index, _eden_chunk_capacity); 1619 for (size_t i = 0; i < _eden_chunk_index; i++) { 1620 log.trace("_eden_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT, i, p2i(_eden_chunk_array[i])); 1621 } 1622 } 1623 // Survivor 1624 if (_survivor_chunk_array != NULL) { 1625 log.trace("survivor " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")", 1626 p2i(from_space->bottom()), p2i(from_space->top()), 1627 p2i(from_space->end()), from_space->capacity()); 1628 log.trace("_survivor_chunk_index=" SIZE_FORMAT ", _survivor_chunk_capacity=" SIZE_FORMAT, 1629 _survivor_chunk_index, _survivor_chunk_capacity); 1630 for (size_t i = 0; i < _survivor_chunk_index; i++) { 1631 log.trace("_survivor_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT, i, p2i(_survivor_chunk_array[i])); 1632 } 1633 } 1634 } 1635 1636 void CMSCollector::getFreelistLocks() const { 1637 // Get locks for all free lists in all generations that this 1638 // collector is responsible for 1639 _cmsGen->freelistLock()->lock_without_safepoint_check(); 1640 } 1641 1642 void CMSCollector::releaseFreelistLocks() const { 1643 // Release locks for all free lists in all generations that this 1644 // collector is responsible for 1645 _cmsGen->freelistLock()->unlock(); 1646 } 1647 1648 bool CMSCollector::haveFreelistLocks() const { 1649 // Check locks for all free lists in all generations that this 1650 // collector is responsible for 1651 assert_lock_strong(_cmsGen->freelistLock()); 1652 PRODUCT_ONLY(ShouldNotReachHere()); 1653 return true; 1654 } 1655 1656 // A utility class that is used by the CMS collector to 1657 // temporarily "release" the foreground collector from its 1658 // usual obligation to wait for the background collector to 1659 // complete an ongoing phase before proceeding. 1660 class ReleaseForegroundGC: public StackObj { 1661 private: 1662 CMSCollector* _c; 1663 public: 1664 ReleaseForegroundGC(CMSCollector* c) : _c(c) { 1665 assert(_c->_foregroundGCShouldWait, "Else should not need to call"); 1666 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1667 // allow a potentially blocked foreground collector to proceed 1668 _c->_foregroundGCShouldWait = false; 1669 if (_c->_foregroundGCIsActive) { 1670 CGC_lock->notify(); 1671 } 1672 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 1673 "Possible deadlock"); 1674 } 1675 1676 ~ReleaseForegroundGC() { 1677 assert(!_c->_foregroundGCShouldWait, "Usage protocol violation?"); 1678 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1679 _c->_foregroundGCShouldWait = true; 1680 } 1681 }; 1682 1683 void CMSCollector::collect_in_background(GCCause::Cause cause) { 1684 assert(Thread::current()->is_ConcurrentGC_thread(), 1685 "A CMS asynchronous collection is only allowed on a CMS thread."); 1686 1687 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1688 { 1689 bool safepoint_check = Mutex::_no_safepoint_check_flag; 1690 MutexLockerEx hl(Heap_lock, safepoint_check); 1691 FreelistLocker fll(this); 1692 MutexLockerEx x(CGC_lock, safepoint_check); 1693 if (_foregroundGCIsActive) { 1694 // The foreground collector is. Skip this 1695 // background collection. 1696 assert(!_foregroundGCShouldWait, "Should be clear"); 1697 return; 1698 } else { 1699 assert(_collectorState == Idling, "Should be idling before start."); 1700 _collectorState = InitialMarking; 1701 register_gc_start(cause); 1702 // Reset the expansion cause, now that we are about to begin 1703 // a new cycle. 1704 clear_expansion_cause(); 1705 1706 // Clear the MetaspaceGC flag since a concurrent collection 1707 // is starting but also clear it after the collection. 1708 MetaspaceGC::set_should_concurrent_collect(false); 1709 } 1710 // Decide if we want to enable class unloading as part of the 1711 // ensuing concurrent GC cycle. 1712 update_should_unload_classes(); 1713 _full_gc_requested = false; // acks all outstanding full gc requests 1714 _full_gc_cause = GCCause::_no_gc; 1715 // Signal that we are about to start a collection 1716 gch->increment_total_full_collections(); // ... starting a collection cycle 1717 _collection_count_start = gch->total_full_collections(); 1718 } 1719 1720 size_t prev_used = _cmsGen->used(); 1721 1722 // The change of the collection state is normally done at this level; 1723 // the exceptions are phases that are executed while the world is 1724 // stopped. For those phases the change of state is done while the 1725 // world is stopped. For baton passing purposes this allows the 1726 // background collector to finish the phase and change state atomically. 1727 // The foreground collector cannot wait on a phase that is done 1728 // while the world is stopped because the foreground collector already 1729 // has the world stopped and would deadlock. 1730 while (_collectorState != Idling) { 1731 log_debug(gc, state)("Thread " INTPTR_FORMAT " in CMS state %d", 1732 p2i(Thread::current()), _collectorState); 1733 // The foreground collector 1734 // holds the Heap_lock throughout its collection. 1735 // holds the CMS token (but not the lock) 1736 // except while it is waiting for the background collector to yield. 1737 // 1738 // The foreground collector should be blocked (not for long) 1739 // if the background collector is about to start a phase 1740 // executed with world stopped. If the background 1741 // collector has already started such a phase, the 1742 // foreground collector is blocked waiting for the 1743 // Heap_lock. The stop-world phases (InitialMarking and FinalMarking) 1744 // are executed in the VM thread. 1745 // 1746 // The locking order is 1747 // PendingListLock (PLL) -- if applicable (FinalMarking) 1748 // Heap_lock (both this & PLL locked in VM_CMS_Operation::prologue()) 1749 // CMS token (claimed in 1750 // stop_world_and_do() --> 1751 // safepoint_synchronize() --> 1752 // CMSThread::synchronize()) 1753 1754 { 1755 // Check if the FG collector wants us to yield. 1756 CMSTokenSync x(true); // is cms thread 1757 if (waitForForegroundGC()) { 1758 // We yielded to a foreground GC, nothing more to be 1759 // done this round. 1760 assert(_foregroundGCShouldWait == false, "We set it to false in " 1761 "waitForForegroundGC()"); 1762 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " exiting collection CMS state %d", 1763 p2i(Thread::current()), _collectorState); 1764 return; 1765 } else { 1766 // The background collector can run but check to see if the 1767 // foreground collector has done a collection while the 1768 // background collector was waiting to get the CGC_lock 1769 // above. If yes, break so that _foregroundGCShouldWait 1770 // is cleared before returning. 1771 if (_collectorState == Idling) { 1772 break; 1773 } 1774 } 1775 } 1776 1777 assert(_foregroundGCShouldWait, "Foreground collector, if active, " 1778 "should be waiting"); 1779 1780 switch (_collectorState) { 1781 case InitialMarking: 1782 { 1783 ReleaseForegroundGC x(this); 1784 stats().record_cms_begin(); 1785 VM_CMS_Initial_Mark initial_mark_op(this); 1786 VMThread::execute(&initial_mark_op); 1787 } 1788 // The collector state may be any legal state at this point 1789 // since the background collector may have yielded to the 1790 // foreground collector. 1791 break; 1792 case Marking: 1793 // initial marking in checkpointRootsInitialWork has been completed 1794 if (markFromRoots()) { // we were successful 1795 assert(_collectorState == Precleaning, "Collector state should " 1796 "have changed"); 1797 } else { 1798 assert(_foregroundGCIsActive, "Internal state inconsistency"); 1799 } 1800 break; 1801 case Precleaning: 1802 // marking from roots in markFromRoots has been completed 1803 preclean(); 1804 assert(_collectorState == AbortablePreclean || 1805 _collectorState == FinalMarking, 1806 "Collector state should have changed"); 1807 break; 1808 case AbortablePreclean: 1809 abortable_preclean(); 1810 assert(_collectorState == FinalMarking, "Collector state should " 1811 "have changed"); 1812 break; 1813 case FinalMarking: 1814 { 1815 ReleaseForegroundGC x(this); 1816 1817 VM_CMS_Final_Remark final_remark_op(this); 1818 VMThread::execute(&final_remark_op); 1819 } 1820 assert(_foregroundGCShouldWait, "block post-condition"); 1821 break; 1822 case Sweeping: 1823 // final marking in checkpointRootsFinal has been completed 1824 sweep(); 1825 assert(_collectorState == Resizing, "Collector state change " 1826 "to Resizing must be done under the free_list_lock"); 1827 1828 case Resizing: { 1829 // Sweeping has been completed... 1830 // At this point the background collection has completed. 1831 // Don't move the call to compute_new_size() down 1832 // into code that might be executed if the background 1833 // collection was preempted. 1834 { 1835 ReleaseForegroundGC x(this); // unblock FG collection 1836 MutexLockerEx y(Heap_lock, Mutex::_no_safepoint_check_flag); 1837 CMSTokenSync z(true); // not strictly needed. 1838 if (_collectorState == Resizing) { 1839 compute_new_size(); 1840 save_heap_summary(); 1841 _collectorState = Resetting; 1842 } else { 1843 assert(_collectorState == Idling, "The state should only change" 1844 " because the foreground collector has finished the collection"); 1845 } 1846 } 1847 break; 1848 } 1849 case Resetting: 1850 // CMS heap resizing has been completed 1851 reset_concurrent(); 1852 assert(_collectorState == Idling, "Collector state should " 1853 "have changed"); 1854 1855 MetaspaceGC::set_should_concurrent_collect(false); 1856 1857 stats().record_cms_end(); 1858 // Don't move the concurrent_phases_end() and compute_new_size() 1859 // calls to here because a preempted background collection 1860 // has it's state set to "Resetting". 1861 break; 1862 case Idling: 1863 default: 1864 ShouldNotReachHere(); 1865 break; 1866 } 1867 log_debug(gc, state)(" Thread " INTPTR_FORMAT " done - next CMS state %d", 1868 p2i(Thread::current()), _collectorState); 1869 assert(_foregroundGCShouldWait, "block post-condition"); 1870 } 1871 1872 // Should this be in gc_epilogue? 1873 collector_policy()->counters()->update_counters(); 1874 1875 { 1876 // Clear _foregroundGCShouldWait and, in the event that the 1877 // foreground collector is waiting, notify it, before 1878 // returning. 1879 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1880 _foregroundGCShouldWait = false; 1881 if (_foregroundGCIsActive) { 1882 CGC_lock->notify(); 1883 } 1884 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 1885 "Possible deadlock"); 1886 } 1887 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " exiting collection CMS state %d", 1888 p2i(Thread::current()), _collectorState); 1889 log_info(gc, heap)("Old: " SIZE_FORMAT "K->" SIZE_FORMAT "K(" SIZE_FORMAT "K)", 1890 prev_used / K, _cmsGen->used()/K, _cmsGen->capacity() /K); 1891 } 1892 1893 void CMSCollector::register_gc_start(GCCause::Cause cause) { 1894 _cms_start_registered = true; 1895 _gc_timer_cm->register_gc_start(); 1896 _gc_tracer_cm->report_gc_start(cause, _gc_timer_cm->gc_start()); 1897 } 1898 1899 void CMSCollector::register_gc_end() { 1900 if (_cms_start_registered) { 1901 report_heap_summary(GCWhen::AfterGC); 1902 1903 _gc_timer_cm->register_gc_end(); 1904 _gc_tracer_cm->report_gc_end(_gc_timer_cm->gc_end(), _gc_timer_cm->time_partitions()); 1905 _cms_start_registered = false; 1906 } 1907 } 1908 1909 void CMSCollector::save_heap_summary() { 1910 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1911 _last_heap_summary = gch->create_heap_summary(); 1912 _last_metaspace_summary = gch->create_metaspace_summary(); 1913 } 1914 1915 void CMSCollector::report_heap_summary(GCWhen::Type when) { 1916 _gc_tracer_cm->report_gc_heap_summary(when, _last_heap_summary); 1917 _gc_tracer_cm->report_metaspace_summary(when, _last_metaspace_summary); 1918 } 1919 1920 bool CMSCollector::waitForForegroundGC() { 1921 bool res = false; 1922 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 1923 "CMS thread should have CMS token"); 1924 // Block the foreground collector until the 1925 // background collectors decides whether to 1926 // yield. 1927 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1928 _foregroundGCShouldWait = true; 1929 if (_foregroundGCIsActive) { 1930 // The background collector yields to the 1931 // foreground collector and returns a value 1932 // indicating that it has yielded. The foreground 1933 // collector can proceed. 1934 res = true; 1935 _foregroundGCShouldWait = false; 1936 ConcurrentMarkSweepThread::clear_CMS_flag( 1937 ConcurrentMarkSweepThread::CMS_cms_has_token); 1938 ConcurrentMarkSweepThread::set_CMS_flag( 1939 ConcurrentMarkSweepThread::CMS_cms_wants_token); 1940 // Get a possibly blocked foreground thread going 1941 CGC_lock->notify(); 1942 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " waiting at CMS state %d", 1943 p2i(Thread::current()), _collectorState); 1944 while (_foregroundGCIsActive) { 1945 CGC_lock->wait(Mutex::_no_safepoint_check_flag); 1946 } 1947 ConcurrentMarkSweepThread::set_CMS_flag( 1948 ConcurrentMarkSweepThread::CMS_cms_has_token); 1949 ConcurrentMarkSweepThread::clear_CMS_flag( 1950 ConcurrentMarkSweepThread::CMS_cms_wants_token); 1951 } 1952 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " continuing at CMS state %d", 1953 p2i(Thread::current()), _collectorState); 1954 return res; 1955 } 1956 1957 // Because of the need to lock the free lists and other structures in 1958 // the collector, common to all the generations that the collector is 1959 // collecting, we need the gc_prologues of individual CMS generations 1960 // delegate to their collector. It may have been simpler had the 1961 // current infrastructure allowed one to call a prologue on a 1962 // collector. In the absence of that we have the generation's 1963 // prologue delegate to the collector, which delegates back 1964 // some "local" work to a worker method in the individual generations 1965 // that it's responsible for collecting, while itself doing any 1966 // work common to all generations it's responsible for. A similar 1967 // comment applies to the gc_epilogue()'s. 1968 // The role of the variable _between_prologue_and_epilogue is to 1969 // enforce the invocation protocol. 1970 void CMSCollector::gc_prologue(bool full) { 1971 // Call gc_prologue_work() for the CMSGen 1972 // we are responsible for. 1973 1974 // The following locking discipline assumes that we are only called 1975 // when the world is stopped. 1976 assert(SafepointSynchronize::is_at_safepoint(), "world is stopped assumption"); 1977 1978 // The CMSCollector prologue must call the gc_prologues for the 1979 // "generations" that it's responsible 1980 // for. 1981 1982 assert( Thread::current()->is_VM_thread() 1983 || ( CMSScavengeBeforeRemark 1984 && Thread::current()->is_ConcurrentGC_thread()), 1985 "Incorrect thread type for prologue execution"); 1986 1987 if (_between_prologue_and_epilogue) { 1988 // We have already been invoked; this is a gc_prologue delegation 1989 // from yet another CMS generation that we are responsible for, just 1990 // ignore it since all relevant work has already been done. 1991 return; 1992 } 1993 1994 // set a bit saying prologue has been called; cleared in epilogue 1995 _between_prologue_and_epilogue = true; 1996 // Claim locks for common data structures, then call gc_prologue_work() 1997 // for each CMSGen. 1998 1999 getFreelistLocks(); // gets free list locks on constituent spaces 2000 bitMapLock()->lock_without_safepoint_check(); 2001 2002 // Should call gc_prologue_work() for all cms gens we are responsible for 2003 bool duringMarking = _collectorState >= Marking 2004 && _collectorState < Sweeping; 2005 2006 // The young collections clear the modified oops state, which tells if 2007 // there are any modified oops in the class. The remark phase also needs 2008 // that information. Tell the young collection to save the union of all 2009 // modified klasses. 2010 if (duringMarking) { 2011 _ct->klass_rem_set()->set_accumulate_modified_oops(true); 2012 } 2013 2014 bool registerClosure = duringMarking; 2015 2016 _cmsGen->gc_prologue_work(full, registerClosure, &_modUnionClosurePar); 2017 2018 if (!full) { 2019 stats().record_gc0_begin(); 2020 } 2021 } 2022 2023 void ConcurrentMarkSweepGeneration::gc_prologue(bool full) { 2024 2025 _capacity_at_prologue = capacity(); 2026 _used_at_prologue = used(); 2027 2028 // Delegate to CMScollector which knows how to coordinate between 2029 // this and any other CMS generations that it is responsible for 2030 // collecting. 2031 collector()->gc_prologue(full); 2032 } 2033 2034 // This is a "private" interface for use by this generation's CMSCollector. 2035 // Not to be called directly by any other entity (for instance, 2036 // GenCollectedHeap, which calls the "public" gc_prologue method above). 2037 void ConcurrentMarkSweepGeneration::gc_prologue_work(bool full, 2038 bool registerClosure, ModUnionClosure* modUnionClosure) { 2039 assert(!incremental_collection_failed(), "Shouldn't be set yet"); 2040 assert(cmsSpace()->preconsumptionDirtyCardClosure() == NULL, 2041 "Should be NULL"); 2042 if (registerClosure) { 2043 cmsSpace()->setPreconsumptionDirtyCardClosure(modUnionClosure); 2044 } 2045 cmsSpace()->gc_prologue(); 2046 // Clear stat counters 2047 NOT_PRODUCT( 2048 assert(_numObjectsPromoted == 0, "check"); 2049 assert(_numWordsPromoted == 0, "check"); 2050 log_develop_trace(gc, alloc)("Allocated " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes concurrently", 2051 _numObjectsAllocated, _numWordsAllocated*sizeof(HeapWord)); 2052 _numObjectsAllocated = 0; 2053 _numWordsAllocated = 0; 2054 ) 2055 } 2056 2057 void CMSCollector::gc_epilogue(bool full) { 2058 // The following locking discipline assumes that we are only called 2059 // when the world is stopped. 2060 assert(SafepointSynchronize::is_at_safepoint(), 2061 "world is stopped assumption"); 2062 2063 // Currently the CMS epilogue (see CompactibleFreeListSpace) merely checks 2064 // if linear allocation blocks need to be appropriately marked to allow the 2065 // the blocks to be parsable. We also check here whether we need to nudge the 2066 // CMS collector thread to start a new cycle (if it's not already active). 2067 assert( Thread::current()->is_VM_thread() 2068 || ( CMSScavengeBeforeRemark 2069 && Thread::current()->is_ConcurrentGC_thread()), 2070 "Incorrect thread type for epilogue execution"); 2071 2072 if (!_between_prologue_and_epilogue) { 2073 // We have already been invoked; this is a gc_epilogue delegation 2074 // from yet another CMS generation that we are responsible for, just 2075 // ignore it since all relevant work has already been done. 2076 return; 2077 } 2078 assert(haveFreelistLocks(), "must have freelist locks"); 2079 assert_lock_strong(bitMapLock()); 2080 2081 _ct->klass_rem_set()->set_accumulate_modified_oops(false); 2082 2083 _cmsGen->gc_epilogue_work(full); 2084 2085 if (_collectorState == AbortablePreclean || _collectorState == Precleaning) { 2086 // in case sampling was not already enabled, enable it 2087 _start_sampling = true; 2088 } 2089 // reset _eden_chunk_array so sampling starts afresh 2090 _eden_chunk_index = 0; 2091 2092 size_t cms_used = _cmsGen->cmsSpace()->used(); 2093 2094 // update performance counters - this uses a special version of 2095 // update_counters() that allows the utilization to be passed as a 2096 // parameter, avoiding multiple calls to used(). 2097 // 2098 _cmsGen->update_counters(cms_used); 2099 2100 bitMapLock()->unlock(); 2101 releaseFreelistLocks(); 2102 2103 if (!CleanChunkPoolAsync) { 2104 Chunk::clean_chunk_pool(); 2105 } 2106 2107 set_did_compact(false); 2108 _between_prologue_and_epilogue = false; // ready for next cycle 2109 } 2110 2111 void ConcurrentMarkSweepGeneration::gc_epilogue(bool full) { 2112 collector()->gc_epilogue(full); 2113 2114 // Also reset promotion tracking in par gc thread states. 2115 for (uint i = 0; i < ParallelGCThreads; i++) { 2116 _par_gc_thread_states[i]->promo.stopTrackingPromotions(i); 2117 } 2118 } 2119 2120 void ConcurrentMarkSweepGeneration::gc_epilogue_work(bool full) { 2121 assert(!incremental_collection_failed(), "Should have been cleared"); 2122 cmsSpace()->setPreconsumptionDirtyCardClosure(NULL); 2123 cmsSpace()->gc_epilogue(); 2124 // Print stat counters 2125 NOT_PRODUCT( 2126 assert(_numObjectsAllocated == 0, "check"); 2127 assert(_numWordsAllocated == 0, "check"); 2128 log_develop_trace(gc, promotion)("Promoted " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes", 2129 _numObjectsPromoted, _numWordsPromoted*sizeof(HeapWord)); 2130 _numObjectsPromoted = 0; 2131 _numWordsPromoted = 0; 2132 ) 2133 2134 // Call down the chain in contiguous_available needs the freelistLock 2135 // so print this out before releasing the freeListLock. 2136 log_develop_trace(gc)(" Contiguous available " SIZE_FORMAT " bytes ", contiguous_available()); 2137 } 2138 2139 #ifndef PRODUCT 2140 bool CMSCollector::have_cms_token() { 2141 Thread* thr = Thread::current(); 2142 if (thr->is_VM_thread()) { 2143 return ConcurrentMarkSweepThread::vm_thread_has_cms_token(); 2144 } else if (thr->is_ConcurrentGC_thread()) { 2145 return ConcurrentMarkSweepThread::cms_thread_has_cms_token(); 2146 } else if (thr->is_GC_task_thread()) { 2147 return ConcurrentMarkSweepThread::vm_thread_has_cms_token() && 2148 ParGCRareEvent_lock->owned_by_self(); 2149 } 2150 return false; 2151 } 2152 2153 // Check reachability of the given heap address in CMS generation, 2154 // treating all other generations as roots. 2155 bool CMSCollector::is_cms_reachable(HeapWord* addr) { 2156 // We could "guarantee" below, rather than assert, but I'll 2157 // leave these as "asserts" so that an adventurous debugger 2158 // could try this in the product build provided some subset of 2159 // the conditions were met, provided they were interested in the 2160 // results and knew that the computation below wouldn't interfere 2161 // with other concurrent computations mutating the structures 2162 // being read or written. 2163 assert(SafepointSynchronize::is_at_safepoint(), 2164 "Else mutations in object graph will make answer suspect"); 2165 assert(have_cms_token(), "Should hold cms token"); 2166 assert(haveFreelistLocks(), "must hold free list locks"); 2167 assert_lock_strong(bitMapLock()); 2168 2169 // Clear the marking bit map array before starting, but, just 2170 // for kicks, first report if the given address is already marked 2171 tty->print_cr("Start: Address " PTR_FORMAT " is%s marked", p2i(addr), 2172 _markBitMap.isMarked(addr) ? "" : " not"); 2173 2174 if (verify_after_remark()) { 2175 MutexLockerEx x(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag); 2176 bool result = verification_mark_bm()->isMarked(addr); 2177 tty->print_cr("TransitiveMark: Address " PTR_FORMAT " %s marked", p2i(addr), 2178 result ? "IS" : "is NOT"); 2179 return result; 2180 } else { 2181 tty->print_cr("Could not compute result"); 2182 return false; 2183 } 2184 } 2185 #endif 2186 2187 void 2188 CMSCollector::print_on_error(outputStream* st) { 2189 CMSCollector* collector = ConcurrentMarkSweepGeneration::_collector; 2190 if (collector != NULL) { 2191 CMSBitMap* bitmap = &collector->_markBitMap; 2192 st->print_cr("Marking Bits: (CMSBitMap*) " PTR_FORMAT, p2i(bitmap)); 2193 bitmap->print_on_error(st, " Bits: "); 2194 2195 st->cr(); 2196 2197 CMSBitMap* mut_bitmap = &collector->_modUnionTable; 2198 st->print_cr("Mod Union Table: (CMSBitMap*) " PTR_FORMAT, p2i(mut_bitmap)); 2199 mut_bitmap->print_on_error(st, " Bits: "); 2200 } 2201 } 2202 2203 //////////////////////////////////////////////////////// 2204 // CMS Verification Support 2205 //////////////////////////////////////////////////////// 2206 // Following the remark phase, the following invariant 2207 // should hold -- each object in the CMS heap which is 2208 // marked in markBitMap() should be marked in the verification_mark_bm(). 2209 2210 class VerifyMarkedClosure: public BitMapClosure { 2211 CMSBitMap* _marks; 2212 bool _failed; 2213 2214 public: 2215 VerifyMarkedClosure(CMSBitMap* bm): _marks(bm), _failed(false) {} 2216 2217 bool do_bit(size_t offset) { 2218 HeapWord* addr = _marks->offsetToHeapWord(offset); 2219 if (!_marks->isMarked(addr)) { 2220 LogHandle(gc, verify) log; 2221 ResourceMark rm; 2222 oop(addr)->print_on(log.info_stream()); 2223 log.info(" (" INTPTR_FORMAT " should have been marked)", p2i(addr)); 2224 _failed = true; 2225 } 2226 return true; 2227 } 2228 2229 bool failed() { return _failed; } 2230 }; 2231 2232 bool CMSCollector::verify_after_remark() { 2233 GCTraceTime(Info, gc, verify) tm("Verifying CMS Marking."); 2234 MutexLockerEx ml(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag); 2235 static bool init = false; 2236 2237 assert(SafepointSynchronize::is_at_safepoint(), 2238 "Else mutations in object graph will make answer suspect"); 2239 assert(have_cms_token(), 2240 "Else there may be mutual interference in use of " 2241 " verification data structures"); 2242 assert(_collectorState > Marking && _collectorState <= Sweeping, 2243 "Else marking info checked here may be obsolete"); 2244 assert(haveFreelistLocks(), "must hold free list locks"); 2245 assert_lock_strong(bitMapLock()); 2246 2247 2248 // Allocate marking bit map if not already allocated 2249 if (!init) { // first time 2250 if (!verification_mark_bm()->allocate(_span)) { 2251 return false; 2252 } 2253 init = true; 2254 } 2255 2256 assert(verification_mark_stack()->isEmpty(), "Should be empty"); 2257 2258 // Turn off refs discovery -- so we will be tracing through refs. 2259 // This is as intended, because by this time 2260 // GC must already have cleared any refs that need to be cleared, 2261 // and traced those that need to be marked; moreover, 2262 // the marking done here is not going to interfere in any 2263 // way with the marking information used by GC. 2264 NoRefDiscovery no_discovery(ref_processor()); 2265 2266 #if defined(COMPILER2) || INCLUDE_JVMCI 2267 DerivedPointerTableDeactivate dpt_deact; 2268 #endif 2269 2270 // Clear any marks from a previous round 2271 verification_mark_bm()->clear_all(); 2272 assert(verification_mark_stack()->isEmpty(), "markStack should be empty"); 2273 verify_work_stacks_empty(); 2274 2275 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2276 gch->ensure_parsability(false); // fill TLABs, but no need to retire them 2277 // Update the saved marks which may affect the root scans. 2278 gch->save_marks(); 2279 2280 if (CMSRemarkVerifyVariant == 1) { 2281 // In this first variant of verification, we complete 2282 // all marking, then check if the new marks-vector is 2283 // a subset of the CMS marks-vector. 2284 verify_after_remark_work_1(); 2285 } else if (CMSRemarkVerifyVariant == 2) { 2286 // In this second variant of verification, we flag an error 2287 // (i.e. an object reachable in the new marks-vector not reachable 2288 // in the CMS marks-vector) immediately, also indicating the 2289 // identify of an object (A) that references the unmarked object (B) -- 2290 // presumably, a mutation to A failed to be picked up by preclean/remark? 2291 verify_after_remark_work_2(); 2292 } else { 2293 warning("Unrecognized value " UINTX_FORMAT " for CMSRemarkVerifyVariant", 2294 CMSRemarkVerifyVariant); 2295 } 2296 return true; 2297 } 2298 2299 void CMSCollector::verify_after_remark_work_1() { 2300 ResourceMark rm; 2301 HandleMark hm; 2302 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2303 2304 // Get a clear set of claim bits for the roots processing to work with. 2305 ClassLoaderDataGraph::clear_claimed_marks(); 2306 2307 // Mark from roots one level into CMS 2308 MarkRefsIntoClosure notOlder(_span, verification_mark_bm()); 2309 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2310 2311 { 2312 StrongRootsScope srs(1); 2313 2314 gch->gen_process_roots(&srs, 2315 GenCollectedHeap::OldGen, 2316 true, // young gen as roots 2317 GenCollectedHeap::ScanningOption(roots_scanning_options()), 2318 should_unload_classes(), 2319 ¬Older, 2320 NULL, 2321 NULL); 2322 } 2323 2324 // Now mark from the roots 2325 MarkFromRootsClosure markFromRootsClosure(this, _span, 2326 verification_mark_bm(), verification_mark_stack(), 2327 false /* don't yield */, true /* verifying */); 2328 assert(_restart_addr == NULL, "Expected pre-condition"); 2329 verification_mark_bm()->iterate(&markFromRootsClosure); 2330 while (_restart_addr != NULL) { 2331 // Deal with stack overflow: by restarting at the indicated 2332 // address. 2333 HeapWord* ra = _restart_addr; 2334 markFromRootsClosure.reset(ra); 2335 _restart_addr = NULL; 2336 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end()); 2337 } 2338 assert(verification_mark_stack()->isEmpty(), "Should have been drained"); 2339 verify_work_stacks_empty(); 2340 2341 // Marking completed -- now verify that each bit marked in 2342 // verification_mark_bm() is also marked in markBitMap(); flag all 2343 // errors by printing corresponding objects. 2344 VerifyMarkedClosure vcl(markBitMap()); 2345 verification_mark_bm()->iterate(&vcl); 2346 if (vcl.failed()) { 2347 LogHandle(gc, verify) log; 2348 log.info("Verification failed"); 2349 ResourceMark rm; 2350 gch->print_on(log.info_stream()); 2351 fatal("CMS: failed marking verification after remark"); 2352 } 2353 } 2354 2355 class VerifyKlassOopsKlassClosure : public KlassClosure { 2356 class VerifyKlassOopsClosure : public OopClosure { 2357 CMSBitMap* _bitmap; 2358 public: 2359 VerifyKlassOopsClosure(CMSBitMap* bitmap) : _bitmap(bitmap) { } 2360 void do_oop(oop* p) { guarantee(*p == NULL || _bitmap->isMarked((HeapWord*) *p), "Should be marked"); } 2361 void do_oop(narrowOop* p) { ShouldNotReachHere(); } 2362 } _oop_closure; 2363 public: 2364 VerifyKlassOopsKlassClosure(CMSBitMap* bitmap) : _oop_closure(bitmap) {} 2365 void do_klass(Klass* k) { 2366 k->oops_do(&_oop_closure); 2367 } 2368 }; 2369 2370 void CMSCollector::verify_after_remark_work_2() { 2371 ResourceMark rm; 2372 HandleMark hm; 2373 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2374 2375 // Get a clear set of claim bits for the roots processing to work with. 2376 ClassLoaderDataGraph::clear_claimed_marks(); 2377 2378 // Mark from roots one level into CMS 2379 MarkRefsIntoVerifyClosure notOlder(_span, verification_mark_bm(), 2380 markBitMap()); 2381 CLDToOopClosure cld_closure(¬Older, true); 2382 2383 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2384 2385 { 2386 StrongRootsScope srs(1); 2387 2388 gch->gen_process_roots(&srs, 2389 GenCollectedHeap::OldGen, 2390 true, // young gen as roots 2391 GenCollectedHeap::ScanningOption(roots_scanning_options()), 2392 should_unload_classes(), 2393 ¬Older, 2394 NULL, 2395 &cld_closure); 2396 } 2397 2398 // Now mark from the roots 2399 MarkFromRootsVerifyClosure markFromRootsClosure(this, _span, 2400 verification_mark_bm(), markBitMap(), verification_mark_stack()); 2401 assert(_restart_addr == NULL, "Expected pre-condition"); 2402 verification_mark_bm()->iterate(&markFromRootsClosure); 2403 while (_restart_addr != NULL) { 2404 // Deal with stack overflow: by restarting at the indicated 2405 // address. 2406 HeapWord* ra = _restart_addr; 2407 markFromRootsClosure.reset(ra); 2408 _restart_addr = NULL; 2409 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end()); 2410 } 2411 assert(verification_mark_stack()->isEmpty(), "Should have been drained"); 2412 verify_work_stacks_empty(); 2413 2414 VerifyKlassOopsKlassClosure verify_klass_oops(verification_mark_bm()); 2415 ClassLoaderDataGraph::classes_do(&verify_klass_oops); 2416 2417 // Marking completed -- now verify that each bit marked in 2418 // verification_mark_bm() is also marked in markBitMap(); flag all 2419 // errors by printing corresponding objects. 2420 VerifyMarkedClosure vcl(markBitMap()); 2421 verification_mark_bm()->iterate(&vcl); 2422 assert(!vcl.failed(), "Else verification above should not have succeeded"); 2423 } 2424 2425 void ConcurrentMarkSweepGeneration::save_marks() { 2426 // delegate to CMS space 2427 cmsSpace()->save_marks(); 2428 for (uint i = 0; i < ParallelGCThreads; i++) { 2429 _par_gc_thread_states[i]->promo.startTrackingPromotions(); 2430 } 2431 } 2432 2433 bool ConcurrentMarkSweepGeneration::no_allocs_since_save_marks() { 2434 return cmsSpace()->no_allocs_since_save_marks(); 2435 } 2436 2437 #define CMS_SINCE_SAVE_MARKS_DEFN(OopClosureType, nv_suffix) \ 2438 \ 2439 void ConcurrentMarkSweepGeneration:: \ 2440 oop_since_save_marks_iterate##nv_suffix(OopClosureType* cl) { \ 2441 cl->set_generation(this); \ 2442 cmsSpace()->oop_since_save_marks_iterate##nv_suffix(cl); \ 2443 cl->reset_generation(); \ 2444 save_marks(); \ 2445 } 2446 2447 ALL_SINCE_SAVE_MARKS_CLOSURES(CMS_SINCE_SAVE_MARKS_DEFN) 2448 2449 void 2450 ConcurrentMarkSweepGeneration::oop_iterate(ExtendedOopClosure* cl) { 2451 if (freelistLock()->owned_by_self()) { 2452 Generation::oop_iterate(cl); 2453 } else { 2454 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2455 Generation::oop_iterate(cl); 2456 } 2457 } 2458 2459 void 2460 ConcurrentMarkSweepGeneration::object_iterate(ObjectClosure* cl) { 2461 if (freelistLock()->owned_by_self()) { 2462 Generation::object_iterate(cl); 2463 } else { 2464 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2465 Generation::object_iterate(cl); 2466 } 2467 } 2468 2469 void 2470 ConcurrentMarkSweepGeneration::safe_object_iterate(ObjectClosure* cl) { 2471 if (freelistLock()->owned_by_self()) { 2472 Generation::safe_object_iterate(cl); 2473 } else { 2474 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2475 Generation::safe_object_iterate(cl); 2476 } 2477 } 2478 2479 void 2480 ConcurrentMarkSweepGeneration::post_compact() { 2481 } 2482 2483 void 2484 ConcurrentMarkSweepGeneration::prepare_for_verify() { 2485 // Fix the linear allocation blocks to look like free blocks. 2486 2487 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those 2488 // are not called when the heap is verified during universe initialization and 2489 // at vm shutdown. 2490 if (freelistLock()->owned_by_self()) { 2491 cmsSpace()->prepare_for_verify(); 2492 } else { 2493 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag); 2494 cmsSpace()->prepare_for_verify(); 2495 } 2496 } 2497 2498 void 2499 ConcurrentMarkSweepGeneration::verify() { 2500 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those 2501 // are not called when the heap is verified during universe initialization and 2502 // at vm shutdown. 2503 if (freelistLock()->owned_by_self()) { 2504 cmsSpace()->verify(); 2505 } else { 2506 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag); 2507 cmsSpace()->verify(); 2508 } 2509 } 2510 2511 void CMSCollector::verify() { 2512 _cmsGen->verify(); 2513 } 2514 2515 #ifndef PRODUCT 2516 bool CMSCollector::overflow_list_is_empty() const { 2517 assert(_num_par_pushes >= 0, "Inconsistency"); 2518 if (_overflow_list == NULL) { 2519 assert(_num_par_pushes == 0, "Inconsistency"); 2520 } 2521 return _overflow_list == NULL; 2522 } 2523 2524 // The methods verify_work_stacks_empty() and verify_overflow_empty() 2525 // merely consolidate assertion checks that appear to occur together frequently. 2526 void CMSCollector::verify_work_stacks_empty() const { 2527 assert(_markStack.isEmpty(), "Marking stack should be empty"); 2528 assert(overflow_list_is_empty(), "Overflow list should be empty"); 2529 } 2530 2531 void CMSCollector::verify_overflow_empty() const { 2532 assert(overflow_list_is_empty(), "Overflow list should be empty"); 2533 assert(no_preserved_marks(), "No preserved marks"); 2534 } 2535 #endif // PRODUCT 2536 2537 // Decide if we want to enable class unloading as part of the 2538 // ensuing concurrent GC cycle. We will collect and 2539 // unload classes if it's the case that: 2540 // (1) an explicit gc request has been made and the flag 2541 // ExplicitGCInvokesConcurrentAndUnloadsClasses is set, OR 2542 // (2) (a) class unloading is enabled at the command line, and 2543 // (b) old gen is getting really full 2544 // NOTE: Provided there is no change in the state of the heap between 2545 // calls to this method, it should have idempotent results. Moreover, 2546 // its results should be monotonically increasing (i.e. going from 0 to 1, 2547 // but not 1 to 0) between successive calls between which the heap was 2548 // not collected. For the implementation below, it must thus rely on 2549 // the property that concurrent_cycles_since_last_unload() 2550 // will not decrease unless a collection cycle happened and that 2551 // _cmsGen->is_too_full() are 2552 // themselves also monotonic in that sense. See check_monotonicity() 2553 // below. 2554 void CMSCollector::update_should_unload_classes() { 2555 _should_unload_classes = false; 2556 // Condition 1 above 2557 if (_full_gc_requested && ExplicitGCInvokesConcurrentAndUnloadsClasses) { 2558 _should_unload_classes = true; 2559 } else if (CMSClassUnloadingEnabled) { // Condition 2.a above 2560 // Disjuncts 2.b.(i,ii,iii) above 2561 _should_unload_classes = (concurrent_cycles_since_last_unload() >= 2562 CMSClassUnloadingMaxInterval) 2563 || _cmsGen->is_too_full(); 2564 } 2565 } 2566 2567 bool ConcurrentMarkSweepGeneration::is_too_full() const { 2568 bool res = should_concurrent_collect(); 2569 res = res && (occupancy() > (double)CMSIsTooFullPercentage/100.0); 2570 return res; 2571 } 2572 2573 void CMSCollector::setup_cms_unloading_and_verification_state() { 2574 const bool should_verify = VerifyBeforeGC || VerifyAfterGC || VerifyDuringGC 2575 || VerifyBeforeExit; 2576 const int rso = GenCollectedHeap::SO_AllCodeCache; 2577 2578 // We set the proper root for this CMS cycle here. 2579 if (should_unload_classes()) { // Should unload classes this cycle 2580 remove_root_scanning_option(rso); // Shrink the root set appropriately 2581 set_verifying(should_verify); // Set verification state for this cycle 2582 return; // Nothing else needs to be done at this time 2583 } 2584 2585 // Not unloading classes this cycle 2586 assert(!should_unload_classes(), "Inconsistency!"); 2587 2588 // If we are not unloading classes then add SO_AllCodeCache to root 2589 // scanning options. 2590 add_root_scanning_option(rso); 2591 2592 if ((!verifying() || unloaded_classes_last_cycle()) && should_verify) { 2593 set_verifying(true); 2594 } else if (verifying() && !should_verify) { 2595 // We were verifying, but some verification flags got disabled. 2596 set_verifying(false); 2597 // Exclude symbols, strings and code cache elements from root scanning to 2598 // reduce IM and RM pauses. 2599 remove_root_scanning_option(rso); 2600 } 2601 } 2602 2603 2604 #ifndef PRODUCT 2605 HeapWord* CMSCollector::block_start(const void* p) const { 2606 const HeapWord* addr = (HeapWord*)p; 2607 if (_span.contains(p)) { 2608 if (_cmsGen->cmsSpace()->is_in_reserved(addr)) { 2609 return _cmsGen->cmsSpace()->block_start(p); 2610 } 2611 } 2612 return NULL; 2613 } 2614 #endif 2615 2616 HeapWord* 2617 ConcurrentMarkSweepGeneration::expand_and_allocate(size_t word_size, 2618 bool tlab, 2619 bool parallel) { 2620 CMSSynchronousYieldRequest yr; 2621 assert(!tlab, "Can't deal with TLAB allocation"); 2622 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2623 expand_for_gc_cause(word_size*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_satisfy_allocation); 2624 if (GCExpandToAllocateDelayMillis > 0) { 2625 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 2626 } 2627 return have_lock_and_allocate(word_size, tlab); 2628 } 2629 2630 void ConcurrentMarkSweepGeneration::expand_for_gc_cause( 2631 size_t bytes, 2632 size_t expand_bytes, 2633 CMSExpansionCause::Cause cause) 2634 { 2635 2636 bool success = expand(bytes, expand_bytes); 2637 2638 // remember why we expanded; this information is used 2639 // by shouldConcurrentCollect() when making decisions on whether to start 2640 // a new CMS cycle. 2641 if (success) { 2642 set_expansion_cause(cause); 2643 log_trace(gc)("Expanded CMS gen for %s", CMSExpansionCause::to_string(cause)); 2644 } 2645 } 2646 2647 HeapWord* ConcurrentMarkSweepGeneration::expand_and_par_lab_allocate(CMSParGCThreadState* ps, size_t word_sz) { 2648 HeapWord* res = NULL; 2649 MutexLocker x(ParGCRareEvent_lock); 2650 while (true) { 2651 // Expansion by some other thread might make alloc OK now: 2652 res = ps->lab.alloc(word_sz); 2653 if (res != NULL) return res; 2654 // If there's not enough expansion space available, give up. 2655 if (_virtual_space.uncommitted_size() < (word_sz * HeapWordSize)) { 2656 return NULL; 2657 } 2658 // Otherwise, we try expansion. 2659 expand_for_gc_cause(word_sz*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_allocate_par_lab); 2660 // Now go around the loop and try alloc again; 2661 // A competing par_promote might beat us to the expansion space, 2662 // so we may go around the loop again if promotion fails again. 2663 if (GCExpandToAllocateDelayMillis > 0) { 2664 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 2665 } 2666 } 2667 } 2668 2669 2670 bool ConcurrentMarkSweepGeneration::expand_and_ensure_spooling_space( 2671 PromotionInfo* promo) { 2672 MutexLocker x(ParGCRareEvent_lock); 2673 size_t refill_size_bytes = promo->refillSize() * HeapWordSize; 2674 while (true) { 2675 // Expansion by some other thread might make alloc OK now: 2676 if (promo->ensure_spooling_space()) { 2677 assert(promo->has_spooling_space(), 2678 "Post-condition of successful ensure_spooling_space()"); 2679 return true; 2680 } 2681 // If there's not enough expansion space available, give up. 2682 if (_virtual_space.uncommitted_size() < refill_size_bytes) { 2683 return false; 2684 } 2685 // Otherwise, we try expansion. 2686 expand_for_gc_cause(refill_size_bytes, MinHeapDeltaBytes, CMSExpansionCause::_allocate_par_spooling_space); 2687 // Now go around the loop and try alloc again; 2688 // A competing allocation might beat us to the expansion space, 2689 // so we may go around the loop again if allocation fails again. 2690 if (GCExpandToAllocateDelayMillis > 0) { 2691 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 2692 } 2693 } 2694 } 2695 2696 void ConcurrentMarkSweepGeneration::shrink(size_t bytes) { 2697 // Only shrink if a compaction was done so that all the free space 2698 // in the generation is in a contiguous block at the end. 2699 if (did_compact()) { 2700 CardGeneration::shrink(bytes); 2701 } 2702 } 2703 2704 void ConcurrentMarkSweepGeneration::assert_correct_size_change_locking() { 2705 assert_locked_or_safepoint(Heap_lock); 2706 } 2707 2708 void ConcurrentMarkSweepGeneration::shrink_free_list_by(size_t bytes) { 2709 assert_locked_or_safepoint(Heap_lock); 2710 assert_lock_strong(freelistLock()); 2711 log_trace(gc)("Shrinking of CMS not yet implemented"); 2712 return; 2713 } 2714 2715 2716 // Simple ctor/dtor wrapper for accounting & timer chores around concurrent 2717 // phases. 2718 class CMSPhaseAccounting: public StackObj { 2719 public: 2720 CMSPhaseAccounting(CMSCollector *collector, 2721 const char *title); 2722 ~CMSPhaseAccounting(); 2723 2724 private: 2725 CMSCollector *_collector; 2726 const char *_title; 2727 GCTraceConcTime(Info, gc) _trace_time; 2728 2729 public: 2730 // Not MT-safe; so do not pass around these StackObj's 2731 // where they may be accessed by other threads. 2732 jlong wallclock_millis() { 2733 return TimeHelper::counter_to_millis(os::elapsed_counter() - _trace_time.start_time()); 2734 } 2735 }; 2736 2737 CMSPhaseAccounting::CMSPhaseAccounting(CMSCollector *collector, 2738 const char *title) : 2739 _collector(collector), _title(title), _trace_time(title) { 2740 2741 _collector->resetYields(); 2742 _collector->resetTimer(); 2743 _collector->startTimer(); 2744 } 2745 2746 CMSPhaseAccounting::~CMSPhaseAccounting() { 2747 _collector->stopTimer(); 2748 log_debug(gc)("Concurrent active time: %.3fms", TimeHelper::counter_to_seconds(_collector->timerTicks())); 2749 log_trace(gc)(" (CMS %s yielded %d times)", _title, _collector->yields()); 2750 } 2751 2752 // CMS work 2753 2754 // The common parts of CMSParInitialMarkTask and CMSParRemarkTask. 2755 class CMSParMarkTask : public AbstractGangTask { 2756 protected: 2757 CMSCollector* _collector; 2758 uint _n_workers; 2759 CMSParMarkTask(const char* name, CMSCollector* collector, uint n_workers) : 2760 AbstractGangTask(name), 2761 _collector(collector), 2762 _n_workers(n_workers) {} 2763 // Work method in support of parallel rescan ... of young gen spaces 2764 void do_young_space_rescan(uint worker_id, OopsInGenClosure* cl, 2765 ContiguousSpace* space, 2766 HeapWord** chunk_array, size_t chunk_top); 2767 void work_on_young_gen_roots(uint worker_id, OopsInGenClosure* cl); 2768 }; 2769 2770 // Parallel initial mark task 2771 class CMSParInitialMarkTask: public CMSParMarkTask { 2772 StrongRootsScope* _strong_roots_scope; 2773 public: 2774 CMSParInitialMarkTask(CMSCollector* collector, StrongRootsScope* strong_roots_scope, uint n_workers) : 2775 CMSParMarkTask("Scan roots and young gen for initial mark in parallel", collector, n_workers), 2776 _strong_roots_scope(strong_roots_scope) {} 2777 void work(uint worker_id); 2778 }; 2779 2780 // Checkpoint the roots into this generation from outside 2781 // this generation. [Note this initial checkpoint need only 2782 // be approximate -- we'll do a catch up phase subsequently.] 2783 void CMSCollector::checkpointRootsInitial() { 2784 assert(_collectorState == InitialMarking, "Wrong collector state"); 2785 check_correct_thread_executing(); 2786 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 2787 2788 save_heap_summary(); 2789 report_heap_summary(GCWhen::BeforeGC); 2790 2791 ReferenceProcessor* rp = ref_processor(); 2792 assert(_restart_addr == NULL, "Control point invariant"); 2793 { 2794 // acquire locks for subsequent manipulations 2795 MutexLockerEx x(bitMapLock(), 2796 Mutex::_no_safepoint_check_flag); 2797 checkpointRootsInitialWork(); 2798 // enable ("weak") refs discovery 2799 rp->enable_discovery(); 2800 _collectorState = Marking; 2801 } 2802 } 2803 2804 void CMSCollector::checkpointRootsInitialWork() { 2805 assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped"); 2806 assert(_collectorState == InitialMarking, "just checking"); 2807 2808 // Already have locks. 2809 assert_lock_strong(bitMapLock()); 2810 assert(_markBitMap.isAllClear(), "was reset at end of previous cycle"); 2811 2812 // Setup the verification and class unloading state for this 2813 // CMS collection cycle. 2814 setup_cms_unloading_and_verification_state(); 2815 2816 GCTraceTime(Trace, gc) ts("checkpointRootsInitialWork", _gc_timer_cm); 2817 2818 // Reset all the PLAB chunk arrays if necessary. 2819 if (_survivor_plab_array != NULL && !CMSPLABRecordAlways) { 2820 reset_survivor_plab_arrays(); 2821 } 2822 2823 ResourceMark rm; 2824 HandleMark hm; 2825 2826 MarkRefsIntoClosure notOlder(_span, &_markBitMap); 2827 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2828 2829 verify_work_stacks_empty(); 2830 verify_overflow_empty(); 2831 2832 gch->ensure_parsability(false); // fill TLABs, but no need to retire them 2833 // Update the saved marks which may affect the root scans. 2834 gch->save_marks(); 2835 2836 // weak reference processing has not started yet. 2837 ref_processor()->set_enqueuing_is_done(false); 2838 2839 // Need to remember all newly created CLDs, 2840 // so that we can guarantee that the remark finds them. 2841 ClassLoaderDataGraph::remember_new_clds(true); 2842 2843 // Whenever a CLD is found, it will be claimed before proceeding to mark 2844 // the klasses. The claimed marks need to be cleared before marking starts. 2845 ClassLoaderDataGraph::clear_claimed_marks(); 2846 2847 print_eden_and_survivor_chunk_arrays(); 2848 2849 { 2850 #if defined(COMPILER2) || INCLUDE_JVMCI 2851 DerivedPointerTableDeactivate dpt_deact; 2852 #endif 2853 if (CMSParallelInitialMarkEnabled) { 2854 // The parallel version. 2855 WorkGang* workers = gch->workers(); 2856 assert(workers != NULL, "Need parallel worker threads."); 2857 uint n_workers = workers->active_workers(); 2858 2859 StrongRootsScope srs(n_workers); 2860 2861 CMSParInitialMarkTask tsk(this, &srs, n_workers); 2862 initialize_sequential_subtasks_for_young_gen_rescan(n_workers); 2863 if (n_workers > 1) { 2864 workers->run_task(&tsk); 2865 } else { 2866 tsk.work(0); 2867 } 2868 } else { 2869 // The serial version. 2870 CLDToOopClosure cld_closure(¬Older, true); 2871 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2872 2873 StrongRootsScope srs(1); 2874 2875 gch->gen_process_roots(&srs, 2876 GenCollectedHeap::OldGen, 2877 true, // young gen as roots 2878 GenCollectedHeap::ScanningOption(roots_scanning_options()), 2879 should_unload_classes(), 2880 ¬Older, 2881 NULL, 2882 &cld_closure); 2883 } 2884 } 2885 2886 // Clear mod-union table; it will be dirtied in the prologue of 2887 // CMS generation per each young generation collection. 2888 2889 assert(_modUnionTable.isAllClear(), 2890 "Was cleared in most recent final checkpoint phase" 2891 " or no bits are set in the gc_prologue before the start of the next " 2892 "subsequent marking phase."); 2893 2894 assert(_ct->klass_rem_set()->mod_union_is_clear(), "Must be"); 2895 2896 // Save the end of the used_region of the constituent generations 2897 // to be used to limit the extent of sweep in each generation. 2898 save_sweep_limits(); 2899 verify_overflow_empty(); 2900 } 2901 2902 bool CMSCollector::markFromRoots() { 2903 // we might be tempted to assert that: 2904 // assert(!SafepointSynchronize::is_at_safepoint(), 2905 // "inconsistent argument?"); 2906 // However that wouldn't be right, because it's possible that 2907 // a safepoint is indeed in progress as a young generation 2908 // stop-the-world GC happens even as we mark in this generation. 2909 assert(_collectorState == Marking, "inconsistent state?"); 2910 check_correct_thread_executing(); 2911 verify_overflow_empty(); 2912 2913 // Weak ref discovery note: We may be discovering weak 2914 // refs in this generation concurrent (but interleaved) with 2915 // weak ref discovery by the young generation collector. 2916 2917 CMSTokenSyncWithLocks ts(true, bitMapLock()); 2918 GCTraceCPUTime tcpu; 2919 CMSPhaseAccounting pa(this, "Concrurrent Mark"); 2920 bool res = markFromRootsWork(); 2921 if (res) { 2922 _collectorState = Precleaning; 2923 } else { // We failed and a foreground collection wants to take over 2924 assert(_foregroundGCIsActive, "internal state inconsistency"); 2925 assert(_restart_addr == NULL, "foreground will restart from scratch"); 2926 log_debug(gc)("bailing out to foreground collection"); 2927 } 2928 verify_overflow_empty(); 2929 return res; 2930 } 2931 2932 bool CMSCollector::markFromRootsWork() { 2933 // iterate over marked bits in bit map, doing a full scan and mark 2934 // from these roots using the following algorithm: 2935 // . if oop is to the right of the current scan pointer, 2936 // mark corresponding bit (we'll process it later) 2937 // . else (oop is to left of current scan pointer) 2938 // push oop on marking stack 2939 // . drain the marking stack 2940 2941 // Note that when we do a marking step we need to hold the 2942 // bit map lock -- recall that direct allocation (by mutators) 2943 // and promotion (by the young generation collector) is also 2944 // marking the bit map. [the so-called allocate live policy.] 2945 // Because the implementation of bit map marking is not 2946 // robust wrt simultaneous marking of bits in the same word, 2947 // we need to make sure that there is no such interference 2948 // between concurrent such updates. 2949 2950 // already have locks 2951 assert_lock_strong(bitMapLock()); 2952 2953 verify_work_stacks_empty(); 2954 verify_overflow_empty(); 2955 bool result = false; 2956 if (CMSConcurrentMTEnabled && ConcGCThreads > 0) { 2957 result = do_marking_mt(); 2958 } else { 2959 result = do_marking_st(); 2960 } 2961 return result; 2962 } 2963 2964 // Forward decl 2965 class CMSConcMarkingTask; 2966 2967 class CMSConcMarkingTerminator: public ParallelTaskTerminator { 2968 CMSCollector* _collector; 2969 CMSConcMarkingTask* _task; 2970 public: 2971 virtual void yield(); 2972 2973 // "n_threads" is the number of threads to be terminated. 2974 // "queue_set" is a set of work queues of other threads. 2975 // "collector" is the CMS collector associated with this task terminator. 2976 // "yield" indicates whether we need the gang as a whole to yield. 2977 CMSConcMarkingTerminator(int n_threads, TaskQueueSetSuper* queue_set, CMSCollector* collector) : 2978 ParallelTaskTerminator(n_threads, queue_set), 2979 _collector(collector) { } 2980 2981 void set_task(CMSConcMarkingTask* task) { 2982 _task = task; 2983 } 2984 }; 2985 2986 class CMSConcMarkingTerminatorTerminator: public TerminatorTerminator { 2987 CMSConcMarkingTask* _task; 2988 public: 2989 bool should_exit_termination(); 2990 void set_task(CMSConcMarkingTask* task) { 2991 _task = task; 2992 } 2993 }; 2994 2995 // MT Concurrent Marking Task 2996 class CMSConcMarkingTask: public YieldingFlexibleGangTask { 2997 CMSCollector* _collector; 2998 uint _n_workers; // requested/desired # workers 2999 bool _result; 3000 CompactibleFreeListSpace* _cms_space; 3001 char _pad_front[64]; // padding to ... 3002 HeapWord* _global_finger; // ... avoid sharing cache line 3003 char _pad_back[64]; 3004 HeapWord* _restart_addr; 3005 3006 // Exposed here for yielding support 3007 Mutex* const _bit_map_lock; 3008 3009 // The per thread work queues, available here for stealing 3010 OopTaskQueueSet* _task_queues; 3011 3012 // Termination (and yielding) support 3013 CMSConcMarkingTerminator _term; 3014 CMSConcMarkingTerminatorTerminator _term_term; 3015 3016 public: 3017 CMSConcMarkingTask(CMSCollector* collector, 3018 CompactibleFreeListSpace* cms_space, 3019 YieldingFlexibleWorkGang* workers, 3020 OopTaskQueueSet* task_queues): 3021 YieldingFlexibleGangTask("Concurrent marking done multi-threaded"), 3022 _collector(collector), 3023 _cms_space(cms_space), 3024 _n_workers(0), _result(true), 3025 _task_queues(task_queues), 3026 _term(_n_workers, task_queues, _collector), 3027 _bit_map_lock(collector->bitMapLock()) 3028 { 3029 _requested_size = _n_workers; 3030 _term.set_task(this); 3031 _term_term.set_task(this); 3032 _restart_addr = _global_finger = _cms_space->bottom(); 3033 } 3034 3035 3036 OopTaskQueueSet* task_queues() { return _task_queues; } 3037 3038 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 3039 3040 HeapWord** global_finger_addr() { return &_global_finger; } 3041 3042 CMSConcMarkingTerminator* terminator() { return &_term; } 3043 3044 virtual void set_for_termination(uint active_workers) { 3045 terminator()->reset_for_reuse(active_workers); 3046 } 3047 3048 void work(uint worker_id); 3049 bool should_yield() { 3050 return ConcurrentMarkSweepThread::should_yield() 3051 && !_collector->foregroundGCIsActive(); 3052 } 3053 3054 virtual void coordinator_yield(); // stuff done by coordinator 3055 bool result() { return _result; } 3056 3057 void reset(HeapWord* ra) { 3058 assert(_global_finger >= _cms_space->end(), "Postcondition of ::work(i)"); 3059 _restart_addr = _global_finger = ra; 3060 _term.reset_for_reuse(); 3061 } 3062 3063 static bool get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 3064 OopTaskQueue* work_q); 3065 3066 private: 3067 void do_scan_and_mark(int i, CompactibleFreeListSpace* sp); 3068 void do_work_steal(int i); 3069 void bump_global_finger(HeapWord* f); 3070 }; 3071 3072 bool CMSConcMarkingTerminatorTerminator::should_exit_termination() { 3073 assert(_task != NULL, "Error"); 3074 return _task->yielding(); 3075 // Note that we do not need the disjunct || _task->should_yield() above 3076 // because we want terminating threads to yield only if the task 3077 // is already in the midst of yielding, which happens only after at least one 3078 // thread has yielded. 3079 } 3080 3081 void CMSConcMarkingTerminator::yield() { 3082 if (_task->should_yield()) { 3083 _task->yield(); 3084 } else { 3085 ParallelTaskTerminator::yield(); 3086 } 3087 } 3088 3089 //////////////////////////////////////////////////////////////// 3090 // Concurrent Marking Algorithm Sketch 3091 //////////////////////////////////////////////////////////////// 3092 // Until all tasks exhausted (both spaces): 3093 // -- claim next available chunk 3094 // -- bump global finger via CAS 3095 // -- find first object that starts in this chunk 3096 // and start scanning bitmap from that position 3097 // -- scan marked objects for oops 3098 // -- CAS-mark target, and if successful: 3099 // . if target oop is above global finger (volatile read) 3100 // nothing to do 3101 // . if target oop is in chunk and above local finger 3102 // then nothing to do 3103 // . else push on work-queue 3104 // -- Deal with possible overflow issues: 3105 // . local work-queue overflow causes stuff to be pushed on 3106 // global (common) overflow queue 3107 // . always first empty local work queue 3108 // . then get a batch of oops from global work queue if any 3109 // . then do work stealing 3110 // -- When all tasks claimed (both spaces) 3111 // and local work queue empty, 3112 // then in a loop do: 3113 // . check global overflow stack; steal a batch of oops and trace 3114 // . try to steal from other threads oif GOS is empty 3115 // . if neither is available, offer termination 3116 // -- Terminate and return result 3117 // 3118 void CMSConcMarkingTask::work(uint worker_id) { 3119 elapsedTimer _timer; 3120 ResourceMark rm; 3121 HandleMark hm; 3122 3123 DEBUG_ONLY(_collector->verify_overflow_empty();) 3124 3125 // Before we begin work, our work queue should be empty 3126 assert(work_queue(worker_id)->size() == 0, "Expected to be empty"); 3127 // Scan the bitmap covering _cms_space, tracing through grey objects. 3128 _timer.start(); 3129 do_scan_and_mark(worker_id, _cms_space); 3130 _timer.stop(); 3131 log_trace(gc, task)("Finished cms space scanning in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 3132 3133 // ... do work stealing 3134 _timer.reset(); 3135 _timer.start(); 3136 do_work_steal(worker_id); 3137 _timer.stop(); 3138 log_trace(gc, task)("Finished work stealing in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 3139 assert(_collector->_markStack.isEmpty(), "Should have been emptied"); 3140 assert(work_queue(worker_id)->size() == 0, "Should have been emptied"); 3141 // Note that under the current task protocol, the 3142 // following assertion is true even of the spaces 3143 // expanded since the completion of the concurrent 3144 // marking. XXX This will likely change under a strict 3145 // ABORT semantics. 3146 // After perm removal the comparison was changed to 3147 // greater than or equal to from strictly greater than. 3148 // Before perm removal the highest address sweep would 3149 // have been at the end of perm gen but now is at the 3150 // end of the tenured gen. 3151 assert(_global_finger >= _cms_space->end(), 3152 "All tasks have been completed"); 3153 DEBUG_ONLY(_collector->verify_overflow_empty();) 3154 } 3155 3156 void CMSConcMarkingTask::bump_global_finger(HeapWord* f) { 3157 HeapWord* read = _global_finger; 3158 HeapWord* cur = read; 3159 while (f > read) { 3160 cur = read; 3161 read = (HeapWord*) Atomic::cmpxchg_ptr(f, &_global_finger, cur); 3162 if (cur == read) { 3163 // our cas succeeded 3164 assert(_global_finger >= f, "protocol consistency"); 3165 break; 3166 } 3167 } 3168 } 3169 3170 // This is really inefficient, and should be redone by 3171 // using (not yet available) block-read and -write interfaces to the 3172 // stack and the work_queue. XXX FIX ME !!! 3173 bool CMSConcMarkingTask::get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 3174 OopTaskQueue* work_q) { 3175 // Fast lock-free check 3176 if (ovflw_stk->length() == 0) { 3177 return false; 3178 } 3179 assert(work_q->size() == 0, "Shouldn't steal"); 3180 MutexLockerEx ml(ovflw_stk->par_lock(), 3181 Mutex::_no_safepoint_check_flag); 3182 // Grab up to 1/4 the size of the work queue 3183 size_t num = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 3184 (size_t)ParGCDesiredObjsFromOverflowList); 3185 num = MIN2(num, ovflw_stk->length()); 3186 for (int i = (int) num; i > 0; i--) { 3187 oop cur = ovflw_stk->pop(); 3188 assert(cur != NULL, "Counted wrong?"); 3189 work_q->push(cur); 3190 } 3191 return num > 0; 3192 } 3193 3194 void CMSConcMarkingTask::do_scan_and_mark(int i, CompactibleFreeListSpace* sp) { 3195 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 3196 int n_tasks = pst->n_tasks(); 3197 // We allow that there may be no tasks to do here because 3198 // we are restarting after a stack overflow. 3199 assert(pst->valid() || n_tasks == 0, "Uninitialized use?"); 3200 uint nth_task = 0; 3201 3202 HeapWord* aligned_start = sp->bottom(); 3203 if (sp->used_region().contains(_restart_addr)) { 3204 // Align down to a card boundary for the start of 0th task 3205 // for this space. 3206 aligned_start = 3207 (HeapWord*)align_size_down((uintptr_t)_restart_addr, 3208 CardTableModRefBS::card_size); 3209 } 3210 3211 size_t chunk_size = sp->marking_task_size(); 3212 while (!pst->is_task_claimed(/* reference */ nth_task)) { 3213 // Having claimed the nth task in this space, 3214 // compute the chunk that it corresponds to: 3215 MemRegion span = MemRegion(aligned_start + nth_task*chunk_size, 3216 aligned_start + (nth_task+1)*chunk_size); 3217 // Try and bump the global finger via a CAS; 3218 // note that we need to do the global finger bump 3219 // _before_ taking the intersection below, because 3220 // the task corresponding to that region will be 3221 // deemed done even if the used_region() expands 3222 // because of allocation -- as it almost certainly will 3223 // during start-up while the threads yield in the 3224 // closure below. 3225 HeapWord* finger = span.end(); 3226 bump_global_finger(finger); // atomically 3227 // There are null tasks here corresponding to chunks 3228 // beyond the "top" address of the space. 3229 span = span.intersection(sp->used_region()); 3230 if (!span.is_empty()) { // Non-null task 3231 HeapWord* prev_obj; 3232 assert(!span.contains(_restart_addr) || nth_task == 0, 3233 "Inconsistency"); 3234 if (nth_task == 0) { 3235 // For the 0th task, we'll not need to compute a block_start. 3236 if (span.contains(_restart_addr)) { 3237 // In the case of a restart because of stack overflow, 3238 // we might additionally skip a chunk prefix. 3239 prev_obj = _restart_addr; 3240 } else { 3241 prev_obj = span.start(); 3242 } 3243 } else { 3244 // We want to skip the first object because 3245 // the protocol is to scan any object in its entirety 3246 // that _starts_ in this span; a fortiori, any 3247 // object starting in an earlier span is scanned 3248 // as part of an earlier claimed task. 3249 // Below we use the "careful" version of block_start 3250 // so we do not try to navigate uninitialized objects. 3251 prev_obj = sp->block_start_careful(span.start()); 3252 // Below we use a variant of block_size that uses the 3253 // Printezis bits to avoid waiting for allocated 3254 // objects to become initialized/parsable. 3255 while (prev_obj < span.start()) { 3256 size_t sz = sp->block_size_no_stall(prev_obj, _collector); 3257 if (sz > 0) { 3258 prev_obj += sz; 3259 } else { 3260 // In this case we may end up doing a bit of redundant 3261 // scanning, but that appears unavoidable, short of 3262 // locking the free list locks; see bug 6324141. 3263 break; 3264 } 3265 } 3266 } 3267 if (prev_obj < span.end()) { 3268 MemRegion my_span = MemRegion(prev_obj, span.end()); 3269 // Do the marking work within a non-empty span -- 3270 // the last argument to the constructor indicates whether the 3271 // iteration should be incremental with periodic yields. 3272 Par_MarkFromRootsClosure cl(this, _collector, my_span, 3273 &_collector->_markBitMap, 3274 work_queue(i), 3275 &_collector->_markStack); 3276 _collector->_markBitMap.iterate(&cl, my_span.start(), my_span.end()); 3277 } // else nothing to do for this task 3278 } // else nothing to do for this task 3279 } 3280 // We'd be tempted to assert here that since there are no 3281 // more tasks left to claim in this space, the global_finger 3282 // must exceed space->top() and a fortiori space->end(). However, 3283 // that would not quite be correct because the bumping of 3284 // global_finger occurs strictly after the claiming of a task, 3285 // so by the time we reach here the global finger may not yet 3286 // have been bumped up by the thread that claimed the last 3287 // task. 3288 pst->all_tasks_completed(); 3289 } 3290 3291 class Par_ConcMarkingClosure: public MetadataAwareOopClosure { 3292 private: 3293 CMSCollector* _collector; 3294 CMSConcMarkingTask* _task; 3295 MemRegion _span; 3296 CMSBitMap* _bit_map; 3297 CMSMarkStack* _overflow_stack; 3298 OopTaskQueue* _work_queue; 3299 protected: 3300 DO_OOP_WORK_DEFN 3301 public: 3302 Par_ConcMarkingClosure(CMSCollector* collector, CMSConcMarkingTask* task, OopTaskQueue* work_queue, 3303 CMSBitMap* bit_map, CMSMarkStack* overflow_stack): 3304 MetadataAwareOopClosure(collector->ref_processor()), 3305 _collector(collector), 3306 _task(task), 3307 _span(collector->_span), 3308 _work_queue(work_queue), 3309 _bit_map(bit_map), 3310 _overflow_stack(overflow_stack) 3311 { } 3312 virtual void do_oop(oop* p); 3313 virtual void do_oop(narrowOop* p); 3314 3315 void trim_queue(size_t max); 3316 void handle_stack_overflow(HeapWord* lost); 3317 void do_yield_check() { 3318 if (_task->should_yield()) { 3319 _task->yield(); 3320 } 3321 } 3322 }; 3323 3324 // Grey object scanning during work stealing phase -- 3325 // the salient assumption here is that any references 3326 // that are in these stolen objects being scanned must 3327 // already have been initialized (else they would not have 3328 // been published), so we do not need to check for 3329 // uninitialized objects before pushing here. 3330 void Par_ConcMarkingClosure::do_oop(oop obj) { 3331 assert(obj->is_oop_or_null(true), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 3332 HeapWord* addr = (HeapWord*)obj; 3333 // Check if oop points into the CMS generation 3334 // and is not marked 3335 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 3336 // a white object ... 3337 // If we manage to "claim" the object, by being the 3338 // first thread to mark it, then we push it on our 3339 // marking stack 3340 if (_bit_map->par_mark(addr)) { // ... now grey 3341 // push on work queue (grey set) 3342 bool simulate_overflow = false; 3343 NOT_PRODUCT( 3344 if (CMSMarkStackOverflowALot && 3345 _collector->simulate_overflow()) { 3346 // simulate a stack overflow 3347 simulate_overflow = true; 3348 } 3349 ) 3350 if (simulate_overflow || 3351 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) { 3352 // stack overflow 3353 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _overflow_stack->capacity()); 3354 // We cannot assert that the overflow stack is full because 3355 // it may have been emptied since. 3356 assert(simulate_overflow || 3357 _work_queue->size() == _work_queue->max_elems(), 3358 "Else push should have succeeded"); 3359 handle_stack_overflow(addr); 3360 } 3361 } // Else, some other thread got there first 3362 do_yield_check(); 3363 } 3364 } 3365 3366 void Par_ConcMarkingClosure::do_oop(oop* p) { Par_ConcMarkingClosure::do_oop_work(p); } 3367 void Par_ConcMarkingClosure::do_oop(narrowOop* p) { Par_ConcMarkingClosure::do_oop_work(p); } 3368 3369 void Par_ConcMarkingClosure::trim_queue(size_t max) { 3370 while (_work_queue->size() > max) { 3371 oop new_oop; 3372 if (_work_queue->pop_local(new_oop)) { 3373 assert(new_oop->is_oop(), "Should be an oop"); 3374 assert(_bit_map->isMarked((HeapWord*)new_oop), "Grey object"); 3375 assert(_span.contains((HeapWord*)new_oop), "Not in span"); 3376 new_oop->oop_iterate(this); // do_oop() above 3377 do_yield_check(); 3378 } 3379 } 3380 } 3381 3382 // Upon stack overflow, we discard (part of) the stack, 3383 // remembering the least address amongst those discarded 3384 // in CMSCollector's _restart_address. 3385 void Par_ConcMarkingClosure::handle_stack_overflow(HeapWord* lost) { 3386 // We need to do this under a mutex to prevent other 3387 // workers from interfering with the work done below. 3388 MutexLockerEx ml(_overflow_stack->par_lock(), 3389 Mutex::_no_safepoint_check_flag); 3390 // Remember the least grey address discarded 3391 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 3392 _collector->lower_restart_addr(ra); 3393 _overflow_stack->reset(); // discard stack contents 3394 _overflow_stack->expand(); // expand the stack if possible 3395 } 3396 3397 3398 void CMSConcMarkingTask::do_work_steal(int i) { 3399 OopTaskQueue* work_q = work_queue(i); 3400 oop obj_to_scan; 3401 CMSBitMap* bm = &(_collector->_markBitMap); 3402 CMSMarkStack* ovflw = &(_collector->_markStack); 3403 int* seed = _collector->hash_seed(i); 3404 Par_ConcMarkingClosure cl(_collector, this, work_q, bm, ovflw); 3405 while (true) { 3406 cl.trim_queue(0); 3407 assert(work_q->size() == 0, "Should have been emptied above"); 3408 if (get_work_from_overflow_stack(ovflw, work_q)) { 3409 // Can't assert below because the work obtained from the 3410 // overflow stack may already have been stolen from us. 3411 // assert(work_q->size() > 0, "Work from overflow stack"); 3412 continue; 3413 } else if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 3414 assert(obj_to_scan->is_oop(), "Should be an oop"); 3415 assert(bm->isMarked((HeapWord*)obj_to_scan), "Grey object"); 3416 obj_to_scan->oop_iterate(&cl); 3417 } else if (terminator()->offer_termination(&_term_term)) { 3418 assert(work_q->size() == 0, "Impossible!"); 3419 break; 3420 } else if (yielding() || should_yield()) { 3421 yield(); 3422 } 3423 } 3424 } 3425 3426 // This is run by the CMS (coordinator) thread. 3427 void CMSConcMarkingTask::coordinator_yield() { 3428 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 3429 "CMS thread should hold CMS token"); 3430 // First give up the locks, then yield, then re-lock 3431 // We should probably use a constructor/destructor idiom to 3432 // do this unlock/lock or modify the MutexUnlocker class to 3433 // serve our purpose. XXX 3434 assert_lock_strong(_bit_map_lock); 3435 _bit_map_lock->unlock(); 3436 ConcurrentMarkSweepThread::desynchronize(true); 3437 _collector->stopTimer(); 3438 _collector->incrementYields(); 3439 3440 // It is possible for whichever thread initiated the yield request 3441 // not to get a chance to wake up and take the bitmap lock between 3442 // this thread releasing it and reacquiring it. So, while the 3443 // should_yield() flag is on, let's sleep for a bit to give the 3444 // other thread a chance to wake up. The limit imposed on the number 3445 // of iterations is defensive, to avoid any unforseen circumstances 3446 // putting us into an infinite loop. Since it's always been this 3447 // (coordinator_yield()) method that was observed to cause the 3448 // problem, we are using a parameter (CMSCoordinatorYieldSleepCount) 3449 // which is by default non-zero. For the other seven methods that 3450 // also perform the yield operation, as are using a different 3451 // parameter (CMSYieldSleepCount) which is by default zero. This way we 3452 // can enable the sleeping for those methods too, if necessary. 3453 // See 6442774. 3454 // 3455 // We really need to reconsider the synchronization between the GC 3456 // thread and the yield-requesting threads in the future and we 3457 // should really use wait/notify, which is the recommended 3458 // way of doing this type of interaction. Additionally, we should 3459 // consolidate the eight methods that do the yield operation and they 3460 // are almost identical into one for better maintainability and 3461 // readability. See 6445193. 3462 // 3463 // Tony 2006.06.29 3464 for (unsigned i = 0; i < CMSCoordinatorYieldSleepCount && 3465 ConcurrentMarkSweepThread::should_yield() && 3466 !CMSCollector::foregroundGCIsActive(); ++i) { 3467 os::sleep(Thread::current(), 1, false); 3468 } 3469 3470 ConcurrentMarkSweepThread::synchronize(true); 3471 _bit_map_lock->lock_without_safepoint_check(); 3472 _collector->startTimer(); 3473 } 3474 3475 bool CMSCollector::do_marking_mt() { 3476 assert(ConcGCThreads > 0 && conc_workers() != NULL, "precondition"); 3477 uint num_workers = AdaptiveSizePolicy::calc_active_conc_workers(conc_workers()->total_workers(), 3478 conc_workers()->active_workers(), 3479 Threads::number_of_non_daemon_threads()); 3480 conc_workers()->set_active_workers(num_workers); 3481 3482 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 3483 3484 CMSConcMarkingTask tsk(this, 3485 cms_space, 3486 conc_workers(), 3487 task_queues()); 3488 3489 // Since the actual number of workers we get may be different 3490 // from the number we requested above, do we need to do anything different 3491 // below? In particular, may be we need to subclass the SequantialSubTasksDone 3492 // class?? XXX 3493 cms_space ->initialize_sequential_subtasks_for_marking(num_workers); 3494 3495 // Refs discovery is already non-atomic. 3496 assert(!ref_processor()->discovery_is_atomic(), "Should be non-atomic"); 3497 assert(ref_processor()->discovery_is_mt(), "Discovery should be MT"); 3498 conc_workers()->start_task(&tsk); 3499 while (tsk.yielded()) { 3500 tsk.coordinator_yield(); 3501 conc_workers()->continue_task(&tsk); 3502 } 3503 // If the task was aborted, _restart_addr will be non-NULL 3504 assert(tsk.completed() || _restart_addr != NULL, "Inconsistency"); 3505 while (_restart_addr != NULL) { 3506 // XXX For now we do not make use of ABORTED state and have not 3507 // yet implemented the right abort semantics (even in the original 3508 // single-threaded CMS case). That needs some more investigation 3509 // and is deferred for now; see CR# TBF. 07252005YSR. XXX 3510 assert(!CMSAbortSemantics || tsk.aborted(), "Inconsistency"); 3511 // If _restart_addr is non-NULL, a marking stack overflow 3512 // occurred; we need to do a fresh marking iteration from the 3513 // indicated restart address. 3514 if (_foregroundGCIsActive) { 3515 // We may be running into repeated stack overflows, having 3516 // reached the limit of the stack size, while making very 3517 // slow forward progress. It may be best to bail out and 3518 // let the foreground collector do its job. 3519 // Clear _restart_addr, so that foreground GC 3520 // works from scratch. This avoids the headache of 3521 // a "rescan" which would otherwise be needed because 3522 // of the dirty mod union table & card table. 3523 _restart_addr = NULL; 3524 return false; 3525 } 3526 // Adjust the task to restart from _restart_addr 3527 tsk.reset(_restart_addr); 3528 cms_space ->initialize_sequential_subtasks_for_marking(num_workers, 3529 _restart_addr); 3530 _restart_addr = NULL; 3531 // Get the workers going again 3532 conc_workers()->start_task(&tsk); 3533 while (tsk.yielded()) { 3534 tsk.coordinator_yield(); 3535 conc_workers()->continue_task(&tsk); 3536 } 3537 } 3538 assert(tsk.completed(), "Inconsistency"); 3539 assert(tsk.result() == true, "Inconsistency"); 3540 return true; 3541 } 3542 3543 bool CMSCollector::do_marking_st() { 3544 ResourceMark rm; 3545 HandleMark hm; 3546 3547 // Temporarily make refs discovery single threaded (non-MT) 3548 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false); 3549 MarkFromRootsClosure markFromRootsClosure(this, _span, &_markBitMap, 3550 &_markStack, CMSYield); 3551 // the last argument to iterate indicates whether the iteration 3552 // should be incremental with periodic yields. 3553 _markBitMap.iterate(&markFromRootsClosure); 3554 // If _restart_addr is non-NULL, a marking stack overflow 3555 // occurred; we need to do a fresh iteration from the 3556 // indicated restart address. 3557 while (_restart_addr != NULL) { 3558 if (_foregroundGCIsActive) { 3559 // We may be running into repeated stack overflows, having 3560 // reached the limit of the stack size, while making very 3561 // slow forward progress. It may be best to bail out and 3562 // let the foreground collector do its job. 3563 // Clear _restart_addr, so that foreground GC 3564 // works from scratch. This avoids the headache of 3565 // a "rescan" which would otherwise be needed because 3566 // of the dirty mod union table & card table. 3567 _restart_addr = NULL; 3568 return false; // indicating failure to complete marking 3569 } 3570 // Deal with stack overflow: 3571 // we restart marking from _restart_addr 3572 HeapWord* ra = _restart_addr; 3573 markFromRootsClosure.reset(ra); 3574 _restart_addr = NULL; 3575 _markBitMap.iterate(&markFromRootsClosure, ra, _span.end()); 3576 } 3577 return true; 3578 } 3579 3580 void CMSCollector::preclean() { 3581 check_correct_thread_executing(); 3582 assert(Thread::current()->is_ConcurrentGC_thread(), "Wrong thread"); 3583 verify_work_stacks_empty(); 3584 verify_overflow_empty(); 3585 _abort_preclean = false; 3586 if (CMSPrecleaningEnabled) { 3587 if (!CMSEdenChunksRecordAlways) { 3588 _eden_chunk_index = 0; 3589 } 3590 size_t used = get_eden_used(); 3591 size_t capacity = get_eden_capacity(); 3592 // Don't start sampling unless we will get sufficiently 3593 // many samples. 3594 if (used < (capacity/(CMSScheduleRemarkSamplingRatio * 100) 3595 * CMSScheduleRemarkEdenPenetration)) { 3596 _start_sampling = true; 3597 } else { 3598 _start_sampling = false; 3599 } 3600 GCTraceCPUTime tcpu; 3601 CMSPhaseAccounting pa(this, "Concurrent Preclean"); 3602 preclean_work(CMSPrecleanRefLists1, CMSPrecleanSurvivors1); 3603 } 3604 CMSTokenSync x(true); // is cms thread 3605 if (CMSPrecleaningEnabled) { 3606 sample_eden(); 3607 _collectorState = AbortablePreclean; 3608 } else { 3609 _collectorState = FinalMarking; 3610 } 3611 verify_work_stacks_empty(); 3612 verify_overflow_empty(); 3613 } 3614 3615 // Try and schedule the remark such that young gen 3616 // occupancy is CMSScheduleRemarkEdenPenetration %. 3617 void CMSCollector::abortable_preclean() { 3618 check_correct_thread_executing(); 3619 assert(CMSPrecleaningEnabled, "Inconsistent control state"); 3620 assert(_collectorState == AbortablePreclean, "Inconsistent control state"); 3621 3622 // If Eden's current occupancy is below this threshold, 3623 // immediately schedule the remark; else preclean 3624 // past the next scavenge in an effort to 3625 // schedule the pause as described above. By choosing 3626 // CMSScheduleRemarkEdenSizeThreshold >= max eden size 3627 // we will never do an actual abortable preclean cycle. 3628 if (get_eden_used() > CMSScheduleRemarkEdenSizeThreshold) { 3629 GCTraceCPUTime tcpu; 3630 CMSPhaseAccounting pa(this, "Concurrent Abortable Preclean"); 3631 // We need more smarts in the abortable preclean 3632 // loop below to deal with cases where allocation 3633 // in young gen is very very slow, and our precleaning 3634 // is running a losing race against a horde of 3635 // mutators intent on flooding us with CMS updates 3636 // (dirty cards). 3637 // One, admittedly dumb, strategy is to give up 3638 // after a certain number of abortable precleaning loops 3639 // or after a certain maximum time. We want to make 3640 // this smarter in the next iteration. 3641 // XXX FIX ME!!! YSR 3642 size_t loops = 0, workdone = 0, cumworkdone = 0, waited = 0; 3643 while (!(should_abort_preclean() || 3644 ConcurrentMarkSweepThread::should_terminate())) { 3645 workdone = preclean_work(CMSPrecleanRefLists2, CMSPrecleanSurvivors2); 3646 cumworkdone += workdone; 3647 loops++; 3648 // Voluntarily terminate abortable preclean phase if we have 3649 // been at it for too long. 3650 if ((CMSMaxAbortablePrecleanLoops != 0) && 3651 loops >= CMSMaxAbortablePrecleanLoops) { 3652 log_debug(gc)(" CMS: abort preclean due to loops "); 3653 break; 3654 } 3655 if (pa.wallclock_millis() > CMSMaxAbortablePrecleanTime) { 3656 log_debug(gc)(" CMS: abort preclean due to time "); 3657 break; 3658 } 3659 // If we are doing little work each iteration, we should 3660 // take a short break. 3661 if (workdone < CMSAbortablePrecleanMinWorkPerIteration) { 3662 // Sleep for some time, waiting for work to accumulate 3663 stopTimer(); 3664 cmsThread()->wait_on_cms_lock(CMSAbortablePrecleanWaitMillis); 3665 startTimer(); 3666 waited++; 3667 } 3668 } 3669 log_trace(gc)(" [" SIZE_FORMAT " iterations, " SIZE_FORMAT " waits, " SIZE_FORMAT " cards)] ", 3670 loops, waited, cumworkdone); 3671 } 3672 CMSTokenSync x(true); // is cms thread 3673 if (_collectorState != Idling) { 3674 assert(_collectorState == AbortablePreclean, 3675 "Spontaneous state transition?"); 3676 _collectorState = FinalMarking; 3677 } // Else, a foreground collection completed this CMS cycle. 3678 return; 3679 } 3680 3681 // Respond to an Eden sampling opportunity 3682 void CMSCollector::sample_eden() { 3683 // Make sure a young gc cannot sneak in between our 3684 // reading and recording of a sample. 3685 assert(Thread::current()->is_ConcurrentGC_thread(), 3686 "Only the cms thread may collect Eden samples"); 3687 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 3688 "Should collect samples while holding CMS token"); 3689 if (!_start_sampling) { 3690 return; 3691 } 3692 // When CMSEdenChunksRecordAlways is true, the eden chunk array 3693 // is populated by the young generation. 3694 if (_eden_chunk_array != NULL && !CMSEdenChunksRecordAlways) { 3695 if (_eden_chunk_index < _eden_chunk_capacity) { 3696 _eden_chunk_array[_eden_chunk_index] = *_top_addr; // take sample 3697 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr, 3698 "Unexpected state of Eden"); 3699 // We'd like to check that what we just sampled is an oop-start address; 3700 // however, we cannot do that here since the object may not yet have been 3701 // initialized. So we'll instead do the check when we _use_ this sample 3702 // later. 3703 if (_eden_chunk_index == 0 || 3704 (pointer_delta(_eden_chunk_array[_eden_chunk_index], 3705 _eden_chunk_array[_eden_chunk_index-1]) 3706 >= CMSSamplingGrain)) { 3707 _eden_chunk_index++; // commit sample 3708 } 3709 } 3710 } 3711 if ((_collectorState == AbortablePreclean) && !_abort_preclean) { 3712 size_t used = get_eden_used(); 3713 size_t capacity = get_eden_capacity(); 3714 assert(used <= capacity, "Unexpected state of Eden"); 3715 if (used > (capacity/100 * CMSScheduleRemarkEdenPenetration)) { 3716 _abort_preclean = true; 3717 } 3718 } 3719 } 3720 3721 3722 size_t CMSCollector::preclean_work(bool clean_refs, bool clean_survivor) { 3723 assert(_collectorState == Precleaning || 3724 _collectorState == AbortablePreclean, "incorrect state"); 3725 ResourceMark rm; 3726 HandleMark hm; 3727 3728 // Precleaning is currently not MT but the reference processor 3729 // may be set for MT. Disable it temporarily here. 3730 ReferenceProcessor* rp = ref_processor(); 3731 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(rp, false); 3732 3733 // Do one pass of scrubbing the discovered reference lists 3734 // to remove any reference objects with strongly-reachable 3735 // referents. 3736 if (clean_refs) { 3737 CMSPrecleanRefsYieldClosure yield_cl(this); 3738 assert(rp->span().equals(_span), "Spans should be equal"); 3739 CMSKeepAliveClosure keep_alive(this, _span, &_markBitMap, 3740 &_markStack, true /* preclean */); 3741 CMSDrainMarkingStackClosure complete_trace(this, 3742 _span, &_markBitMap, &_markStack, 3743 &keep_alive, true /* preclean */); 3744 3745 // We don't want this step to interfere with a young 3746 // collection because we don't want to take CPU 3747 // or memory bandwidth away from the young GC threads 3748 // (which may be as many as there are CPUs). 3749 // Note that we don't need to protect ourselves from 3750 // interference with mutators because they can't 3751 // manipulate the discovered reference lists nor affect 3752 // the computed reachability of the referents, the 3753 // only properties manipulated by the precleaning 3754 // of these reference lists. 3755 stopTimer(); 3756 CMSTokenSyncWithLocks x(true /* is cms thread */, 3757 bitMapLock()); 3758 startTimer(); 3759 sample_eden(); 3760 3761 // The following will yield to allow foreground 3762 // collection to proceed promptly. XXX YSR: 3763 // The code in this method may need further 3764 // tweaking for better performance and some restructuring 3765 // for cleaner interfaces. 3766 GCTimer *gc_timer = NULL; // Currently not tracing concurrent phases 3767 rp->preclean_discovered_references( 3768 rp->is_alive_non_header(), &keep_alive, &complete_trace, &yield_cl, 3769 gc_timer); 3770 } 3771 3772 if (clean_survivor) { // preclean the active survivor space(s) 3773 PushAndMarkClosure pam_cl(this, _span, ref_processor(), 3774 &_markBitMap, &_modUnionTable, 3775 &_markStack, true /* precleaning phase */); 3776 stopTimer(); 3777 CMSTokenSyncWithLocks ts(true /* is cms thread */, 3778 bitMapLock()); 3779 startTimer(); 3780 unsigned int before_count = 3781 GenCollectedHeap::heap()->total_collections(); 3782 SurvivorSpacePrecleanClosure 3783 sss_cl(this, _span, &_markBitMap, &_markStack, 3784 &pam_cl, before_count, CMSYield); 3785 _young_gen->from()->object_iterate_careful(&sss_cl); 3786 _young_gen->to()->object_iterate_careful(&sss_cl); 3787 } 3788 MarkRefsIntoAndScanClosure 3789 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable, 3790 &_markStack, this, CMSYield, 3791 true /* precleaning phase */); 3792 // CAUTION: The following closure has persistent state that may need to 3793 // be reset upon a decrease in the sequence of addresses it 3794 // processes. 3795 ScanMarkedObjectsAgainCarefullyClosure 3796 smoac_cl(this, _span, 3797 &_markBitMap, &_markStack, &mrias_cl, CMSYield); 3798 3799 // Preclean dirty cards in ModUnionTable and CardTable using 3800 // appropriate convergence criterion; 3801 // repeat CMSPrecleanIter times unless we find that 3802 // we are losing. 3803 assert(CMSPrecleanIter < 10, "CMSPrecleanIter is too large"); 3804 assert(CMSPrecleanNumerator < CMSPrecleanDenominator, 3805 "Bad convergence multiplier"); 3806 assert(CMSPrecleanThreshold >= 100, 3807 "Unreasonably low CMSPrecleanThreshold"); 3808 3809 size_t numIter, cumNumCards, lastNumCards, curNumCards; 3810 for (numIter = 0, cumNumCards = lastNumCards = curNumCards = 0; 3811 numIter < CMSPrecleanIter; 3812 numIter++, lastNumCards = curNumCards, cumNumCards += curNumCards) { 3813 curNumCards = preclean_mod_union_table(_cmsGen, &smoac_cl); 3814 log_trace(gc)(" (modUnionTable: " SIZE_FORMAT " cards)", curNumCards); 3815 // Either there are very few dirty cards, so re-mark 3816 // pause will be small anyway, or our pre-cleaning isn't 3817 // that much faster than the rate at which cards are being 3818 // dirtied, so we might as well stop and re-mark since 3819 // precleaning won't improve our re-mark time by much. 3820 if (curNumCards <= CMSPrecleanThreshold || 3821 (numIter > 0 && 3822 (curNumCards * CMSPrecleanDenominator > 3823 lastNumCards * CMSPrecleanNumerator))) { 3824 numIter++; 3825 cumNumCards += curNumCards; 3826 break; 3827 } 3828 } 3829 3830 preclean_klasses(&mrias_cl, _cmsGen->freelistLock()); 3831 3832 curNumCards = preclean_card_table(_cmsGen, &smoac_cl); 3833 cumNumCards += curNumCards; 3834 log_trace(gc)(" (cardTable: " SIZE_FORMAT " cards, re-scanned " SIZE_FORMAT " cards, " SIZE_FORMAT " iterations)", 3835 curNumCards, cumNumCards, numIter); 3836 return cumNumCards; // as a measure of useful work done 3837 } 3838 3839 // PRECLEANING NOTES: 3840 // Precleaning involves: 3841 // . reading the bits of the modUnionTable and clearing the set bits. 3842 // . For the cards corresponding to the set bits, we scan the 3843 // objects on those cards. This means we need the free_list_lock 3844 // so that we can safely iterate over the CMS space when scanning 3845 // for oops. 3846 // . When we scan the objects, we'll be both reading and setting 3847 // marks in the marking bit map, so we'll need the marking bit map. 3848 // . For protecting _collector_state transitions, we take the CGC_lock. 3849 // Note that any races in the reading of of card table entries by the 3850 // CMS thread on the one hand and the clearing of those entries by the 3851 // VM thread or the setting of those entries by the mutator threads on the 3852 // other are quite benign. However, for efficiency it makes sense to keep 3853 // the VM thread from racing with the CMS thread while the latter is 3854 // dirty card info to the modUnionTable. We therefore also use the 3855 // CGC_lock to protect the reading of the card table and the mod union 3856 // table by the CM thread. 3857 // . We run concurrently with mutator updates, so scanning 3858 // needs to be done carefully -- we should not try to scan 3859 // potentially uninitialized objects. 3860 // 3861 // Locking strategy: While holding the CGC_lock, we scan over and 3862 // reset a maximal dirty range of the mod union / card tables, then lock 3863 // the free_list_lock and bitmap lock to do a full marking, then 3864 // release these locks; and repeat the cycle. This allows for a 3865 // certain amount of fairness in the sharing of these locks between 3866 // the CMS collector on the one hand, and the VM thread and the 3867 // mutators on the other. 3868 3869 // NOTE: preclean_mod_union_table() and preclean_card_table() 3870 // further below are largely identical; if you need to modify 3871 // one of these methods, please check the other method too. 3872 3873 size_t CMSCollector::preclean_mod_union_table( 3874 ConcurrentMarkSweepGeneration* old_gen, 3875 ScanMarkedObjectsAgainCarefullyClosure* cl) { 3876 verify_work_stacks_empty(); 3877 verify_overflow_empty(); 3878 3879 // strategy: starting with the first card, accumulate contiguous 3880 // ranges of dirty cards; clear these cards, then scan the region 3881 // covered by these cards. 3882 3883 // Since all of the MUT is committed ahead, we can just use 3884 // that, in case the generations expand while we are precleaning. 3885 // It might also be fine to just use the committed part of the 3886 // generation, but we might potentially miss cards when the 3887 // generation is rapidly expanding while we are in the midst 3888 // of precleaning. 3889 HeapWord* startAddr = old_gen->reserved().start(); 3890 HeapWord* endAddr = old_gen->reserved().end(); 3891 3892 cl->setFreelistLock(old_gen->freelistLock()); // needed for yielding 3893 3894 size_t numDirtyCards, cumNumDirtyCards; 3895 HeapWord *nextAddr, *lastAddr; 3896 for (cumNumDirtyCards = numDirtyCards = 0, 3897 nextAddr = lastAddr = startAddr; 3898 nextAddr < endAddr; 3899 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 3900 3901 ResourceMark rm; 3902 HandleMark hm; 3903 3904 MemRegion dirtyRegion; 3905 { 3906 stopTimer(); 3907 // Potential yield point 3908 CMSTokenSync ts(true); 3909 startTimer(); 3910 sample_eden(); 3911 // Get dirty region starting at nextOffset (inclusive), 3912 // simultaneously clearing it. 3913 dirtyRegion = 3914 _modUnionTable.getAndClearMarkedRegion(nextAddr, endAddr); 3915 assert(dirtyRegion.start() >= nextAddr, 3916 "returned region inconsistent?"); 3917 } 3918 // Remember where the next search should begin. 3919 // The returned region (if non-empty) is a right open interval, 3920 // so lastOffset is obtained from the right end of that 3921 // interval. 3922 lastAddr = dirtyRegion.end(); 3923 // Should do something more transparent and less hacky XXX 3924 numDirtyCards = 3925 _modUnionTable.heapWordDiffToOffsetDiff(dirtyRegion.word_size()); 3926 3927 // We'll scan the cards in the dirty region (with periodic 3928 // yields for foreground GC as needed). 3929 if (!dirtyRegion.is_empty()) { 3930 assert(numDirtyCards > 0, "consistency check"); 3931 HeapWord* stop_point = NULL; 3932 stopTimer(); 3933 // Potential yield point 3934 CMSTokenSyncWithLocks ts(true, old_gen->freelistLock(), 3935 bitMapLock()); 3936 startTimer(); 3937 { 3938 verify_work_stacks_empty(); 3939 verify_overflow_empty(); 3940 sample_eden(); 3941 stop_point = 3942 old_gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 3943 } 3944 if (stop_point != NULL) { 3945 // The careful iteration stopped early either because it found an 3946 // uninitialized object, or because we were in the midst of an 3947 // "abortable preclean", which should now be aborted. Redirty 3948 // the bits corresponding to the partially-scanned or unscanned 3949 // cards. We'll either restart at the next block boundary or 3950 // abort the preclean. 3951 assert((_collectorState == AbortablePreclean && should_abort_preclean()), 3952 "Should only be AbortablePreclean."); 3953 _modUnionTable.mark_range(MemRegion(stop_point, dirtyRegion.end())); 3954 if (should_abort_preclean()) { 3955 break; // out of preclean loop 3956 } else { 3957 // Compute the next address at which preclean should pick up; 3958 // might need bitMapLock in order to read P-bits. 3959 lastAddr = next_card_start_after_block(stop_point); 3960 } 3961 } 3962 } else { 3963 assert(lastAddr == endAddr, "consistency check"); 3964 assert(numDirtyCards == 0, "consistency check"); 3965 break; 3966 } 3967 } 3968 verify_work_stacks_empty(); 3969 verify_overflow_empty(); 3970 return cumNumDirtyCards; 3971 } 3972 3973 // NOTE: preclean_mod_union_table() above and preclean_card_table() 3974 // below are largely identical; if you need to modify 3975 // one of these methods, please check the other method too. 3976 3977 size_t CMSCollector::preclean_card_table(ConcurrentMarkSweepGeneration* old_gen, 3978 ScanMarkedObjectsAgainCarefullyClosure* cl) { 3979 // strategy: it's similar to precleamModUnionTable above, in that 3980 // we accumulate contiguous ranges of dirty cards, mark these cards 3981 // precleaned, then scan the region covered by these cards. 3982 HeapWord* endAddr = (HeapWord*)(old_gen->_virtual_space.high()); 3983 HeapWord* startAddr = (HeapWord*)(old_gen->_virtual_space.low()); 3984 3985 cl->setFreelistLock(old_gen->freelistLock()); // needed for yielding 3986 3987 size_t numDirtyCards, cumNumDirtyCards; 3988 HeapWord *lastAddr, *nextAddr; 3989 3990 for (cumNumDirtyCards = numDirtyCards = 0, 3991 nextAddr = lastAddr = startAddr; 3992 nextAddr < endAddr; 3993 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 3994 3995 ResourceMark rm; 3996 HandleMark hm; 3997 3998 MemRegion dirtyRegion; 3999 { 4000 // See comments in "Precleaning notes" above on why we 4001 // do this locking. XXX Could the locking overheads be 4002 // too high when dirty cards are sparse? [I don't think so.] 4003 stopTimer(); 4004 CMSTokenSync x(true); // is cms thread 4005 startTimer(); 4006 sample_eden(); 4007 // Get and clear dirty region from card table 4008 dirtyRegion = _ct->ct_bs()->dirty_card_range_after_reset( 4009 MemRegion(nextAddr, endAddr), 4010 true, 4011 CardTableModRefBS::precleaned_card_val()); 4012 4013 assert(dirtyRegion.start() >= nextAddr, 4014 "returned region inconsistent?"); 4015 } 4016 lastAddr = dirtyRegion.end(); 4017 numDirtyCards = 4018 dirtyRegion.word_size()/CardTableModRefBS::card_size_in_words; 4019 4020 if (!dirtyRegion.is_empty()) { 4021 stopTimer(); 4022 CMSTokenSyncWithLocks ts(true, old_gen->freelistLock(), bitMapLock()); 4023 startTimer(); 4024 sample_eden(); 4025 verify_work_stacks_empty(); 4026 verify_overflow_empty(); 4027 HeapWord* stop_point = 4028 old_gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 4029 if (stop_point != NULL) { 4030 assert((_collectorState == AbortablePreclean && should_abort_preclean()), 4031 "Should only be AbortablePreclean."); 4032 _ct->ct_bs()->invalidate(MemRegion(stop_point, dirtyRegion.end())); 4033 if (should_abort_preclean()) { 4034 break; // out of preclean loop 4035 } else { 4036 // Compute the next address at which preclean should pick up. 4037 lastAddr = next_card_start_after_block(stop_point); 4038 } 4039 } 4040 } else { 4041 break; 4042 } 4043 } 4044 verify_work_stacks_empty(); 4045 verify_overflow_empty(); 4046 return cumNumDirtyCards; 4047 } 4048 4049 class PrecleanKlassClosure : public KlassClosure { 4050 KlassToOopClosure _cm_klass_closure; 4051 public: 4052 PrecleanKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {} 4053 void do_klass(Klass* k) { 4054 if (k->has_accumulated_modified_oops()) { 4055 k->clear_accumulated_modified_oops(); 4056 4057 _cm_klass_closure.do_klass(k); 4058 } 4059 } 4060 }; 4061 4062 // The freelist lock is needed to prevent asserts, is it really needed? 4063 void CMSCollector::preclean_klasses(MarkRefsIntoAndScanClosure* cl, Mutex* freelistLock) { 4064 4065 cl->set_freelistLock(freelistLock); 4066 4067 CMSTokenSyncWithLocks ts(true, freelistLock, bitMapLock()); 4068 4069 // SSS: Add equivalent to ScanMarkedObjectsAgainCarefullyClosure::do_yield_check and should_abort_preclean? 4070 // SSS: We should probably check if precleaning should be aborted, at suitable intervals? 4071 PrecleanKlassClosure preclean_klass_closure(cl); 4072 ClassLoaderDataGraph::classes_do(&preclean_klass_closure); 4073 4074 verify_work_stacks_empty(); 4075 verify_overflow_empty(); 4076 } 4077 4078 void CMSCollector::checkpointRootsFinal() { 4079 assert(_collectorState == FinalMarking, "incorrect state transition?"); 4080 check_correct_thread_executing(); 4081 // world is stopped at this checkpoint 4082 assert(SafepointSynchronize::is_at_safepoint(), 4083 "world should be stopped"); 4084 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 4085 4086 verify_work_stacks_empty(); 4087 verify_overflow_empty(); 4088 4089 log_debug(gc)("YG occupancy: " SIZE_FORMAT " K (" SIZE_FORMAT " K)", 4090 _young_gen->used() / K, _young_gen->capacity() / K); 4091 { 4092 if (CMSScavengeBeforeRemark) { 4093 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4094 // Temporarily set flag to false, GCH->do_collection will 4095 // expect it to be false and set to true 4096 FlagSetting fl(gch->_is_gc_active, false); 4097 4098 GCTraceTime(Trace, gc) tm("Pause Scavenge Before Remark", _gc_timer_cm); 4099 4100 gch->do_collection(true, // full (i.e. force, see below) 4101 false, // !clear_all_soft_refs 4102 0, // size 4103 false, // is_tlab 4104 GenCollectedHeap::YoungGen // type 4105 ); 4106 } 4107 FreelistLocker x(this); 4108 MutexLockerEx y(bitMapLock(), 4109 Mutex::_no_safepoint_check_flag); 4110 checkpointRootsFinalWork(); 4111 } 4112 verify_work_stacks_empty(); 4113 verify_overflow_empty(); 4114 } 4115 4116 void CMSCollector::checkpointRootsFinalWork() { 4117 GCTraceTime(Trace, gc) tm("checkpointRootsFinalWork", _gc_timer_cm); 4118 4119 assert(haveFreelistLocks(), "must have free list locks"); 4120 assert_lock_strong(bitMapLock()); 4121 4122 ResourceMark rm; 4123 HandleMark hm; 4124 4125 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4126 4127 if (should_unload_classes()) { 4128 CodeCache::gc_prologue(); 4129 } 4130 assert(haveFreelistLocks(), "must have free list locks"); 4131 assert_lock_strong(bitMapLock()); 4132 4133 // We might assume that we need not fill TLAB's when 4134 // CMSScavengeBeforeRemark is set, because we may have just done 4135 // a scavenge which would have filled all TLAB's -- and besides 4136 // Eden would be empty. This however may not always be the case -- 4137 // for instance although we asked for a scavenge, it may not have 4138 // happened because of a JNI critical section. We probably need 4139 // a policy for deciding whether we can in that case wait until 4140 // the critical section releases and then do the remark following 4141 // the scavenge, and skip it here. In the absence of that policy, 4142 // or of an indication of whether the scavenge did indeed occur, 4143 // we cannot rely on TLAB's having been filled and must do 4144 // so here just in case a scavenge did not happen. 4145 gch->ensure_parsability(false); // fill TLAB's, but no need to retire them 4146 // Update the saved marks which may affect the root scans. 4147 gch->save_marks(); 4148 4149 print_eden_and_survivor_chunk_arrays(); 4150 4151 { 4152 #if defined(COMPILER2) || INCLUDE_JVMCI 4153 DerivedPointerTableDeactivate dpt_deact; 4154 #endif 4155 4156 // Note on the role of the mod union table: 4157 // Since the marker in "markFromRoots" marks concurrently with 4158 // mutators, it is possible for some reachable objects not to have been 4159 // scanned. For instance, an only reference to an object A was 4160 // placed in object B after the marker scanned B. Unless B is rescanned, 4161 // A would be collected. Such updates to references in marked objects 4162 // are detected via the mod union table which is the set of all cards 4163 // dirtied since the first checkpoint in this GC cycle and prior to 4164 // the most recent young generation GC, minus those cleaned up by the 4165 // concurrent precleaning. 4166 if (CMSParallelRemarkEnabled) { 4167 GCTraceTime(Debug, gc) t("Rescan (parallel)", _gc_timer_cm); 4168 do_remark_parallel(); 4169 } else { 4170 GCTraceTime(Debug, gc) t("Rescan (non-parallel)", _gc_timer_cm); 4171 do_remark_non_parallel(); 4172 } 4173 } 4174 verify_work_stacks_empty(); 4175 verify_overflow_empty(); 4176 4177 { 4178 GCTraceTime(Trace, gc) ts("refProcessingWork", _gc_timer_cm); 4179 refProcessingWork(); 4180 } 4181 verify_work_stacks_empty(); 4182 verify_overflow_empty(); 4183 4184 if (should_unload_classes()) { 4185 CodeCache::gc_epilogue(); 4186 } 4187 JvmtiExport::gc_epilogue(); 4188 4189 // If we encountered any (marking stack / work queue) overflow 4190 // events during the current CMS cycle, take appropriate 4191 // remedial measures, where possible, so as to try and avoid 4192 // recurrence of that condition. 4193 assert(_markStack.isEmpty(), "No grey objects"); 4194 size_t ser_ovflw = _ser_pmc_remark_ovflw + _ser_pmc_preclean_ovflw + 4195 _ser_kac_ovflw + _ser_kac_preclean_ovflw; 4196 if (ser_ovflw > 0) { 4197 log_trace(gc)("Marking stack overflow (benign) (pmc_pc=" SIZE_FORMAT ", pmc_rm=" SIZE_FORMAT ", kac=" SIZE_FORMAT ", kac_preclean=" SIZE_FORMAT ")", 4198 _ser_pmc_preclean_ovflw, _ser_pmc_remark_ovflw, _ser_kac_ovflw, _ser_kac_preclean_ovflw); 4199 _markStack.expand(); 4200 _ser_pmc_remark_ovflw = 0; 4201 _ser_pmc_preclean_ovflw = 0; 4202 _ser_kac_preclean_ovflw = 0; 4203 _ser_kac_ovflw = 0; 4204 } 4205 if (_par_pmc_remark_ovflw > 0 || _par_kac_ovflw > 0) { 4206 log_trace(gc)("Work queue overflow (benign) (pmc_rm=" SIZE_FORMAT ", kac=" SIZE_FORMAT ")", 4207 _par_pmc_remark_ovflw, _par_kac_ovflw); 4208 _par_pmc_remark_ovflw = 0; 4209 _par_kac_ovflw = 0; 4210 } 4211 if (_markStack._hit_limit > 0) { 4212 log_trace(gc)(" (benign) Hit max stack size limit (" SIZE_FORMAT ")", 4213 _markStack._hit_limit); 4214 } 4215 if (_markStack._failed_double > 0) { 4216 log_trace(gc)(" (benign) Failed stack doubling (" SIZE_FORMAT "), current capacity " SIZE_FORMAT, 4217 _markStack._failed_double, _markStack.capacity()); 4218 } 4219 _markStack._hit_limit = 0; 4220 _markStack._failed_double = 0; 4221 4222 if ((VerifyAfterGC || VerifyDuringGC) && 4223 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 4224 verify_after_remark(); 4225 } 4226 4227 _gc_tracer_cm->report_object_count_after_gc(&_is_alive_closure); 4228 4229 // Change under the freelistLocks. 4230 _collectorState = Sweeping; 4231 // Call isAllClear() under bitMapLock 4232 assert(_modUnionTable.isAllClear(), 4233 "Should be clear by end of the final marking"); 4234 assert(_ct->klass_rem_set()->mod_union_is_clear(), 4235 "Should be clear by end of the final marking"); 4236 } 4237 4238 void CMSParInitialMarkTask::work(uint worker_id) { 4239 elapsedTimer _timer; 4240 ResourceMark rm; 4241 HandleMark hm; 4242 4243 // ---------- scan from roots -------------- 4244 _timer.start(); 4245 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4246 Par_MarkRefsIntoClosure par_mri_cl(_collector->_span, &(_collector->_markBitMap)); 4247 4248 // ---------- young gen roots -------------- 4249 { 4250 work_on_young_gen_roots(worker_id, &par_mri_cl); 4251 _timer.stop(); 4252 log_trace(gc, task)("Finished young gen initial mark scan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4253 } 4254 4255 // ---------- remaining roots -------------- 4256 _timer.reset(); 4257 _timer.start(); 4258 4259 CLDToOopClosure cld_closure(&par_mri_cl, true); 4260 4261 gch->gen_process_roots(_strong_roots_scope, 4262 GenCollectedHeap::OldGen, 4263 false, // yg was scanned above 4264 GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()), 4265 _collector->should_unload_classes(), 4266 &par_mri_cl, 4267 NULL, 4268 &cld_closure); 4269 assert(_collector->should_unload_classes() 4270 || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache), 4271 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 4272 _timer.stop(); 4273 log_trace(gc, task)("Finished remaining root initial mark scan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4274 } 4275 4276 // Parallel remark task 4277 class CMSParRemarkTask: public CMSParMarkTask { 4278 CompactibleFreeListSpace* _cms_space; 4279 4280 // The per-thread work queues, available here for stealing. 4281 OopTaskQueueSet* _task_queues; 4282 ParallelTaskTerminator _term; 4283 StrongRootsScope* _strong_roots_scope; 4284 4285 public: 4286 // A value of 0 passed to n_workers will cause the number of 4287 // workers to be taken from the active workers in the work gang. 4288 CMSParRemarkTask(CMSCollector* collector, 4289 CompactibleFreeListSpace* cms_space, 4290 uint n_workers, WorkGang* workers, 4291 OopTaskQueueSet* task_queues, 4292 StrongRootsScope* strong_roots_scope): 4293 CMSParMarkTask("Rescan roots and grey objects in parallel", 4294 collector, n_workers), 4295 _cms_space(cms_space), 4296 _task_queues(task_queues), 4297 _term(n_workers, task_queues), 4298 _strong_roots_scope(strong_roots_scope) { } 4299 4300 OopTaskQueueSet* task_queues() { return _task_queues; } 4301 4302 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 4303 4304 ParallelTaskTerminator* terminator() { return &_term; } 4305 uint n_workers() { return _n_workers; } 4306 4307 void work(uint worker_id); 4308 4309 private: 4310 // ... of dirty cards in old space 4311 void do_dirty_card_rescan_tasks(CompactibleFreeListSpace* sp, int i, 4312 Par_MarkRefsIntoAndScanClosure* cl); 4313 4314 // ... work stealing for the above 4315 void do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, int* seed); 4316 }; 4317 4318 class RemarkKlassClosure : public KlassClosure { 4319 KlassToOopClosure _cm_klass_closure; 4320 public: 4321 RemarkKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {} 4322 void do_klass(Klass* k) { 4323 // Check if we have modified any oops in the Klass during the concurrent marking. 4324 if (k->has_accumulated_modified_oops()) { 4325 k->clear_accumulated_modified_oops(); 4326 4327 // We could have transfered the current modified marks to the accumulated marks, 4328 // like we do with the Card Table to Mod Union Table. But it's not really necessary. 4329 } else if (k->has_modified_oops()) { 4330 // Don't clear anything, this info is needed by the next young collection. 4331 } else { 4332 // No modified oops in the Klass. 4333 return; 4334 } 4335 4336 // The klass has modified fields, need to scan the klass. 4337 _cm_klass_closure.do_klass(k); 4338 } 4339 }; 4340 4341 void CMSParMarkTask::work_on_young_gen_roots(uint worker_id, OopsInGenClosure* cl) { 4342 ParNewGeneration* young_gen = _collector->_young_gen; 4343 ContiguousSpace* eden_space = young_gen->eden(); 4344 ContiguousSpace* from_space = young_gen->from(); 4345 ContiguousSpace* to_space = young_gen->to(); 4346 4347 HeapWord** eca = _collector->_eden_chunk_array; 4348 size_t ect = _collector->_eden_chunk_index; 4349 HeapWord** sca = _collector->_survivor_chunk_array; 4350 size_t sct = _collector->_survivor_chunk_index; 4351 4352 assert(ect <= _collector->_eden_chunk_capacity, "out of bounds"); 4353 assert(sct <= _collector->_survivor_chunk_capacity, "out of bounds"); 4354 4355 do_young_space_rescan(worker_id, cl, to_space, NULL, 0); 4356 do_young_space_rescan(worker_id, cl, from_space, sca, sct); 4357 do_young_space_rescan(worker_id, cl, eden_space, eca, ect); 4358 } 4359 4360 // work_queue(i) is passed to the closure 4361 // Par_MarkRefsIntoAndScanClosure. The "i" parameter 4362 // also is passed to do_dirty_card_rescan_tasks() and to 4363 // do_work_steal() to select the i-th task_queue. 4364 4365 void CMSParRemarkTask::work(uint worker_id) { 4366 elapsedTimer _timer; 4367 ResourceMark rm; 4368 HandleMark hm; 4369 4370 // ---------- rescan from roots -------------- 4371 _timer.start(); 4372 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4373 Par_MarkRefsIntoAndScanClosure par_mrias_cl(_collector, 4374 _collector->_span, _collector->ref_processor(), 4375 &(_collector->_markBitMap), 4376 work_queue(worker_id)); 4377 4378 // Rescan young gen roots first since these are likely 4379 // coarsely partitioned and may, on that account, constitute 4380 // the critical path; thus, it's best to start off that 4381 // work first. 4382 // ---------- young gen roots -------------- 4383 { 4384 work_on_young_gen_roots(worker_id, &par_mrias_cl); 4385 _timer.stop(); 4386 log_trace(gc, task)("Finished young gen rescan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4387 } 4388 4389 // ---------- remaining roots -------------- 4390 _timer.reset(); 4391 _timer.start(); 4392 gch->gen_process_roots(_strong_roots_scope, 4393 GenCollectedHeap::OldGen, 4394 false, // yg was scanned above 4395 GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()), 4396 _collector->should_unload_classes(), 4397 &par_mrias_cl, 4398 NULL, 4399 NULL); // The dirty klasses will be handled below 4400 4401 assert(_collector->should_unload_classes() 4402 || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache), 4403 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 4404 _timer.stop(); 4405 log_trace(gc, task)("Finished remaining root rescan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4406 4407 // ---------- unhandled CLD scanning ---------- 4408 if (worker_id == 0) { // Single threaded at the moment. 4409 _timer.reset(); 4410 _timer.start(); 4411 4412 // Scan all new class loader data objects and new dependencies that were 4413 // introduced during concurrent marking. 4414 ResourceMark rm; 4415 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds(); 4416 for (int i = 0; i < array->length(); i++) { 4417 par_mrias_cl.do_cld_nv(array->at(i)); 4418 } 4419 4420 // We don't need to keep track of new CLDs anymore. 4421 ClassLoaderDataGraph::remember_new_clds(false); 4422 4423 _timer.stop(); 4424 log_trace(gc, task)("Finished unhandled CLD scanning work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4425 } 4426 4427 // ---------- dirty klass scanning ---------- 4428 if (worker_id == 0) { // Single threaded at the moment. 4429 _timer.reset(); 4430 _timer.start(); 4431 4432 // Scan all classes that was dirtied during the concurrent marking phase. 4433 RemarkKlassClosure remark_klass_closure(&par_mrias_cl); 4434 ClassLoaderDataGraph::classes_do(&remark_klass_closure); 4435 4436 _timer.stop(); 4437 log_trace(gc, task)("Finished dirty klass scanning work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4438 } 4439 4440 // We might have added oops to ClassLoaderData::_handles during the 4441 // concurrent marking phase. These oops point to newly allocated objects 4442 // that are guaranteed to be kept alive. Either by the direct allocation 4443 // code, or when the young collector processes the roots. Hence, 4444 // we don't have to revisit the _handles block during the remark phase. 4445 4446 // ---------- rescan dirty cards ------------ 4447 _timer.reset(); 4448 _timer.start(); 4449 4450 // Do the rescan tasks for each of the two spaces 4451 // (cms_space) in turn. 4452 // "worker_id" is passed to select the task_queue for "worker_id" 4453 do_dirty_card_rescan_tasks(_cms_space, worker_id, &par_mrias_cl); 4454 _timer.stop(); 4455 log_trace(gc, task)("Finished dirty card rescan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4456 4457 // ---------- steal work from other threads ... 4458 // ---------- ... and drain overflow list. 4459 _timer.reset(); 4460 _timer.start(); 4461 do_work_steal(worker_id, &par_mrias_cl, _collector->hash_seed(worker_id)); 4462 _timer.stop(); 4463 log_trace(gc, task)("Finished work stealing in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4464 } 4465 4466 // Note that parameter "i" is not used. 4467 void 4468 CMSParMarkTask::do_young_space_rescan(uint worker_id, 4469 OopsInGenClosure* cl, ContiguousSpace* space, 4470 HeapWord** chunk_array, size_t chunk_top) { 4471 // Until all tasks completed: 4472 // . claim an unclaimed task 4473 // . compute region boundaries corresponding to task claimed 4474 // using chunk_array 4475 // . par_oop_iterate(cl) over that region 4476 4477 ResourceMark rm; 4478 HandleMark hm; 4479 4480 SequentialSubTasksDone* pst = space->par_seq_tasks(); 4481 4482 uint nth_task = 0; 4483 uint n_tasks = pst->n_tasks(); 4484 4485 if (n_tasks > 0) { 4486 assert(pst->valid(), "Uninitialized use?"); 4487 HeapWord *start, *end; 4488 while (!pst->is_task_claimed(/* reference */ nth_task)) { 4489 // We claimed task # nth_task; compute its boundaries. 4490 if (chunk_top == 0) { // no samples were taken 4491 assert(nth_task == 0 && n_tasks == 1, "Can have only 1 eden task"); 4492 start = space->bottom(); 4493 end = space->top(); 4494 } else if (nth_task == 0) { 4495 start = space->bottom(); 4496 end = chunk_array[nth_task]; 4497 } else if (nth_task < (uint)chunk_top) { 4498 assert(nth_task >= 1, "Control point invariant"); 4499 start = chunk_array[nth_task - 1]; 4500 end = chunk_array[nth_task]; 4501 } else { 4502 assert(nth_task == (uint)chunk_top, "Control point invariant"); 4503 start = chunk_array[chunk_top - 1]; 4504 end = space->top(); 4505 } 4506 MemRegion mr(start, end); 4507 // Verify that mr is in space 4508 assert(mr.is_empty() || space->used_region().contains(mr), 4509 "Should be in space"); 4510 // Verify that "start" is an object boundary 4511 assert(mr.is_empty() || oop(mr.start())->is_oop(), 4512 "Should be an oop"); 4513 space->par_oop_iterate(mr, cl); 4514 } 4515 pst->all_tasks_completed(); 4516 } 4517 } 4518 4519 void 4520 CMSParRemarkTask::do_dirty_card_rescan_tasks( 4521 CompactibleFreeListSpace* sp, int i, 4522 Par_MarkRefsIntoAndScanClosure* cl) { 4523 // Until all tasks completed: 4524 // . claim an unclaimed task 4525 // . compute region boundaries corresponding to task claimed 4526 // . transfer dirty bits ct->mut for that region 4527 // . apply rescanclosure to dirty mut bits for that region 4528 4529 ResourceMark rm; 4530 HandleMark hm; 4531 4532 OopTaskQueue* work_q = work_queue(i); 4533 ModUnionClosure modUnionClosure(&(_collector->_modUnionTable)); 4534 // CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! 4535 // CAUTION: This closure has state that persists across calls to 4536 // the work method dirty_range_iterate_clear() in that it has 4537 // embedded in it a (subtype of) UpwardsObjectClosure. The 4538 // use of that state in the embedded UpwardsObjectClosure instance 4539 // assumes that the cards are always iterated (even if in parallel 4540 // by several threads) in monotonically increasing order per each 4541 // thread. This is true of the implementation below which picks 4542 // card ranges (chunks) in monotonically increasing order globally 4543 // and, a-fortiori, in monotonically increasing order per thread 4544 // (the latter order being a subsequence of the former). 4545 // If the work code below is ever reorganized into a more chaotic 4546 // work-partitioning form than the current "sequential tasks" 4547 // paradigm, the use of that persistent state will have to be 4548 // revisited and modified appropriately. See also related 4549 // bug 4756801 work on which should examine this code to make 4550 // sure that the changes there do not run counter to the 4551 // assumptions made here and necessary for correctness and 4552 // efficiency. Note also that this code might yield inefficient 4553 // behavior in the case of very large objects that span one or 4554 // more work chunks. Such objects would potentially be scanned 4555 // several times redundantly. Work on 4756801 should try and 4556 // address that performance anomaly if at all possible. XXX 4557 MemRegion full_span = _collector->_span; 4558 CMSBitMap* bm = &(_collector->_markBitMap); // shared 4559 MarkFromDirtyCardsClosure 4560 greyRescanClosure(_collector, full_span, // entire span of interest 4561 sp, bm, work_q, cl); 4562 4563 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 4564 assert(pst->valid(), "Uninitialized use?"); 4565 uint nth_task = 0; 4566 const int alignment = CardTableModRefBS::card_size * BitsPerWord; 4567 MemRegion span = sp->used_region(); 4568 HeapWord* start_addr = span.start(); 4569 HeapWord* end_addr = (HeapWord*)round_to((intptr_t)span.end(), 4570 alignment); 4571 const size_t chunk_size = sp->rescan_task_size(); // in HeapWord units 4572 assert((HeapWord*)round_to((intptr_t)start_addr, alignment) == 4573 start_addr, "Check alignment"); 4574 assert((size_t)round_to((intptr_t)chunk_size, alignment) == 4575 chunk_size, "Check alignment"); 4576 4577 while (!pst->is_task_claimed(/* reference */ nth_task)) { 4578 // Having claimed the nth_task, compute corresponding mem-region, 4579 // which is a-fortiori aligned correctly (i.e. at a MUT boundary). 4580 // The alignment restriction ensures that we do not need any 4581 // synchronization with other gang-workers while setting or 4582 // clearing bits in thus chunk of the MUT. 4583 MemRegion this_span = MemRegion(start_addr + nth_task*chunk_size, 4584 start_addr + (nth_task+1)*chunk_size); 4585 // The last chunk's end might be way beyond end of the 4586 // used region. In that case pull back appropriately. 4587 if (this_span.end() > end_addr) { 4588 this_span.set_end(end_addr); 4589 assert(!this_span.is_empty(), "Program logic (calculation of n_tasks)"); 4590 } 4591 // Iterate over the dirty cards covering this chunk, marking them 4592 // precleaned, and setting the corresponding bits in the mod union 4593 // table. Since we have been careful to partition at Card and MUT-word 4594 // boundaries no synchronization is needed between parallel threads. 4595 _collector->_ct->ct_bs()->dirty_card_iterate(this_span, 4596 &modUnionClosure); 4597 4598 // Having transferred these marks into the modUnionTable, 4599 // rescan the marked objects on the dirty cards in the modUnionTable. 4600 // Even if this is at a synchronous collection, the initial marking 4601 // may have been done during an asynchronous collection so there 4602 // may be dirty bits in the mod-union table. 4603 _collector->_modUnionTable.dirty_range_iterate_clear( 4604 this_span, &greyRescanClosure); 4605 _collector->_modUnionTable.verifyNoOneBitsInRange( 4606 this_span.start(), 4607 this_span.end()); 4608 } 4609 pst->all_tasks_completed(); // declare that i am done 4610 } 4611 4612 // . see if we can share work_queues with ParNew? XXX 4613 void 4614 CMSParRemarkTask::do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, 4615 int* seed) { 4616 OopTaskQueue* work_q = work_queue(i); 4617 NOT_PRODUCT(int num_steals = 0;) 4618 oop obj_to_scan; 4619 CMSBitMap* bm = &(_collector->_markBitMap); 4620 4621 while (true) { 4622 // Completely finish any left over work from (an) earlier round(s) 4623 cl->trim_queue(0); 4624 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 4625 (size_t)ParGCDesiredObjsFromOverflowList); 4626 // Now check if there's any work in the overflow list 4627 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads, 4628 // only affects the number of attempts made to get work from the 4629 // overflow list and does not affect the number of workers. Just 4630 // pass ParallelGCThreads so this behavior is unchanged. 4631 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 4632 work_q, 4633 ParallelGCThreads)) { 4634 // found something in global overflow list; 4635 // not yet ready to go stealing work from others. 4636 // We'd like to assert(work_q->size() != 0, ...) 4637 // because we just took work from the overflow list, 4638 // but of course we can't since all of that could have 4639 // been already stolen from us. 4640 // "He giveth and He taketh away." 4641 continue; 4642 } 4643 // Verify that we have no work before we resort to stealing 4644 assert(work_q->size() == 0, "Have work, shouldn't steal"); 4645 // Try to steal from other queues that have work 4646 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 4647 NOT_PRODUCT(num_steals++;) 4648 assert(obj_to_scan->is_oop(), "Oops, not an oop!"); 4649 assert(bm->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 4650 // Do scanning work 4651 obj_to_scan->oop_iterate(cl); 4652 // Loop around, finish this work, and try to steal some more 4653 } else if (terminator()->offer_termination()) { 4654 break; // nirvana from the infinite cycle 4655 } 4656 } 4657 log_develop_trace(gc, task)("\t(%d: stole %d oops)", i, num_steals); 4658 assert(work_q->size() == 0 && _collector->overflow_list_is_empty(), 4659 "Else our work is not yet done"); 4660 } 4661 4662 // Record object boundaries in _eden_chunk_array by sampling the eden 4663 // top in the slow-path eden object allocation code path and record 4664 // the boundaries, if CMSEdenChunksRecordAlways is true. If 4665 // CMSEdenChunksRecordAlways is false, we use the other asynchronous 4666 // sampling in sample_eden() that activates during the part of the 4667 // preclean phase. 4668 void CMSCollector::sample_eden_chunk() { 4669 if (CMSEdenChunksRecordAlways && _eden_chunk_array != NULL) { 4670 if (_eden_chunk_lock->try_lock()) { 4671 // Record a sample. This is the critical section. The contents 4672 // of the _eden_chunk_array have to be non-decreasing in the 4673 // address order. 4674 _eden_chunk_array[_eden_chunk_index] = *_top_addr; 4675 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr, 4676 "Unexpected state of Eden"); 4677 if (_eden_chunk_index == 0 || 4678 ((_eden_chunk_array[_eden_chunk_index] > _eden_chunk_array[_eden_chunk_index-1]) && 4679 (pointer_delta(_eden_chunk_array[_eden_chunk_index], 4680 _eden_chunk_array[_eden_chunk_index-1]) >= CMSSamplingGrain))) { 4681 _eden_chunk_index++; // commit sample 4682 } 4683 _eden_chunk_lock->unlock(); 4684 } 4685 } 4686 } 4687 4688 // Return a thread-local PLAB recording array, as appropriate. 4689 void* CMSCollector::get_data_recorder(int thr_num) { 4690 if (_survivor_plab_array != NULL && 4691 (CMSPLABRecordAlways || 4692 (_collectorState > Marking && _collectorState < FinalMarking))) { 4693 assert(thr_num < (int)ParallelGCThreads, "thr_num is out of bounds"); 4694 ChunkArray* ca = &_survivor_plab_array[thr_num]; 4695 ca->reset(); // clear it so that fresh data is recorded 4696 return (void*) ca; 4697 } else { 4698 return NULL; 4699 } 4700 } 4701 4702 // Reset all the thread-local PLAB recording arrays 4703 void CMSCollector::reset_survivor_plab_arrays() { 4704 for (uint i = 0; i < ParallelGCThreads; i++) { 4705 _survivor_plab_array[i].reset(); 4706 } 4707 } 4708 4709 // Merge the per-thread plab arrays into the global survivor chunk 4710 // array which will provide the partitioning of the survivor space 4711 // for CMS initial scan and rescan. 4712 void CMSCollector::merge_survivor_plab_arrays(ContiguousSpace* surv, 4713 int no_of_gc_threads) { 4714 assert(_survivor_plab_array != NULL, "Error"); 4715 assert(_survivor_chunk_array != NULL, "Error"); 4716 assert(_collectorState == FinalMarking || 4717 (CMSParallelInitialMarkEnabled && _collectorState == InitialMarking), "Error"); 4718 for (int j = 0; j < no_of_gc_threads; j++) { 4719 _cursor[j] = 0; 4720 } 4721 HeapWord* top = surv->top(); 4722 size_t i; 4723 for (i = 0; i < _survivor_chunk_capacity; i++) { // all sca entries 4724 HeapWord* min_val = top; // Higher than any PLAB address 4725 uint min_tid = 0; // position of min_val this round 4726 for (int j = 0; j < no_of_gc_threads; j++) { 4727 ChunkArray* cur_sca = &_survivor_plab_array[j]; 4728 if (_cursor[j] == cur_sca->end()) { 4729 continue; 4730 } 4731 assert(_cursor[j] < cur_sca->end(), "ctl pt invariant"); 4732 HeapWord* cur_val = cur_sca->nth(_cursor[j]); 4733 assert(surv->used_region().contains(cur_val), "Out of bounds value"); 4734 if (cur_val < min_val) { 4735 min_tid = j; 4736 min_val = cur_val; 4737 } else { 4738 assert(cur_val < top, "All recorded addresses should be less"); 4739 } 4740 } 4741 // At this point min_val and min_tid are respectively 4742 // the least address in _survivor_plab_array[j]->nth(_cursor[j]) 4743 // and the thread (j) that witnesses that address. 4744 // We record this address in the _survivor_chunk_array[i] 4745 // and increment _cursor[min_tid] prior to the next round i. 4746 if (min_val == top) { 4747 break; 4748 } 4749 _survivor_chunk_array[i] = min_val; 4750 _cursor[min_tid]++; 4751 } 4752 // We are all done; record the size of the _survivor_chunk_array 4753 _survivor_chunk_index = i; // exclusive: [0, i) 4754 log_trace(gc, survivor)(" (Survivor:" SIZE_FORMAT "chunks) ", i); 4755 // Verify that we used up all the recorded entries 4756 #ifdef ASSERT 4757 size_t total = 0; 4758 for (int j = 0; j < no_of_gc_threads; j++) { 4759 assert(_cursor[j] == _survivor_plab_array[j].end(), "Ctl pt invariant"); 4760 total += _cursor[j]; 4761 } 4762 assert(total == _survivor_chunk_index, "Ctl Pt Invariant"); 4763 // Check that the merged array is in sorted order 4764 if (total > 0) { 4765 for (size_t i = 0; i < total - 1; i++) { 4766 log_develop_trace(gc, survivor)(" (chunk" SIZE_FORMAT ":" INTPTR_FORMAT ") ", 4767 i, p2i(_survivor_chunk_array[i])); 4768 assert(_survivor_chunk_array[i] < _survivor_chunk_array[i+1], 4769 "Not sorted"); 4770 } 4771 } 4772 #endif // ASSERT 4773 } 4774 4775 // Set up the space's par_seq_tasks structure for work claiming 4776 // for parallel initial scan and rescan of young gen. 4777 // See ParRescanTask where this is currently used. 4778 void 4779 CMSCollector:: 4780 initialize_sequential_subtasks_for_young_gen_rescan(int n_threads) { 4781 assert(n_threads > 0, "Unexpected n_threads argument"); 4782 4783 // Eden space 4784 if (!_young_gen->eden()->is_empty()) { 4785 SequentialSubTasksDone* pst = _young_gen->eden()->par_seq_tasks(); 4786 assert(!pst->valid(), "Clobbering existing data?"); 4787 // Each valid entry in [0, _eden_chunk_index) represents a task. 4788 size_t n_tasks = _eden_chunk_index + 1; 4789 assert(n_tasks == 1 || _eden_chunk_array != NULL, "Error"); 4790 // Sets the condition for completion of the subtask (how many threads 4791 // need to finish in order to be done). 4792 pst->set_n_threads(n_threads); 4793 pst->set_n_tasks((int)n_tasks); 4794 } 4795 4796 // Merge the survivor plab arrays into _survivor_chunk_array 4797 if (_survivor_plab_array != NULL) { 4798 merge_survivor_plab_arrays(_young_gen->from(), n_threads); 4799 } else { 4800 assert(_survivor_chunk_index == 0, "Error"); 4801 } 4802 4803 // To space 4804 { 4805 SequentialSubTasksDone* pst = _young_gen->to()->par_seq_tasks(); 4806 assert(!pst->valid(), "Clobbering existing data?"); 4807 // Sets the condition for completion of the subtask (how many threads 4808 // need to finish in order to be done). 4809 pst->set_n_threads(n_threads); 4810 pst->set_n_tasks(1); 4811 assert(pst->valid(), "Error"); 4812 } 4813 4814 // From space 4815 { 4816 SequentialSubTasksDone* pst = _young_gen->from()->par_seq_tasks(); 4817 assert(!pst->valid(), "Clobbering existing data?"); 4818 size_t n_tasks = _survivor_chunk_index + 1; 4819 assert(n_tasks == 1 || _survivor_chunk_array != NULL, "Error"); 4820 // Sets the condition for completion of the subtask (how many threads 4821 // need to finish in order to be done). 4822 pst->set_n_threads(n_threads); 4823 pst->set_n_tasks((int)n_tasks); 4824 assert(pst->valid(), "Error"); 4825 } 4826 } 4827 4828 // Parallel version of remark 4829 void CMSCollector::do_remark_parallel() { 4830 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4831 WorkGang* workers = gch->workers(); 4832 assert(workers != NULL, "Need parallel worker threads."); 4833 // Choose to use the number of GC workers most recently set 4834 // into "active_workers". 4835 uint n_workers = workers->active_workers(); 4836 4837 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 4838 4839 StrongRootsScope srs(n_workers); 4840 4841 CMSParRemarkTask tsk(this, cms_space, n_workers, workers, task_queues(), &srs); 4842 4843 // We won't be iterating over the cards in the card table updating 4844 // the younger_gen cards, so we shouldn't call the following else 4845 // the verification code as well as subsequent younger_refs_iterate 4846 // code would get confused. XXX 4847 // gch->rem_set()->prepare_for_younger_refs_iterate(true); // parallel 4848 4849 // The young gen rescan work will not be done as part of 4850 // process_roots (which currently doesn't know how to 4851 // parallelize such a scan), but rather will be broken up into 4852 // a set of parallel tasks (via the sampling that the [abortable] 4853 // preclean phase did of eden, plus the [two] tasks of 4854 // scanning the [two] survivor spaces. Further fine-grain 4855 // parallelization of the scanning of the survivor spaces 4856 // themselves, and of precleaning of the young gen itself 4857 // is deferred to the future. 4858 initialize_sequential_subtasks_for_young_gen_rescan(n_workers); 4859 4860 // The dirty card rescan work is broken up into a "sequence" 4861 // of parallel tasks (per constituent space) that are dynamically 4862 // claimed by the parallel threads. 4863 cms_space->initialize_sequential_subtasks_for_rescan(n_workers); 4864 4865 // It turns out that even when we're using 1 thread, doing the work in a 4866 // separate thread causes wide variance in run times. We can't help this 4867 // in the multi-threaded case, but we special-case n=1 here to get 4868 // repeatable measurements of the 1-thread overhead of the parallel code. 4869 if (n_workers > 1) { 4870 // Make refs discovery MT-safe, if it isn't already: it may not 4871 // necessarily be so, since it's possible that we are doing 4872 // ST marking. 4873 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), true); 4874 workers->run_task(&tsk); 4875 } else { 4876 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false); 4877 tsk.work(0); 4878 } 4879 4880 // restore, single-threaded for now, any preserved marks 4881 // as a result of work_q overflow 4882 restore_preserved_marks_if_any(); 4883 } 4884 4885 // Non-parallel version of remark 4886 void CMSCollector::do_remark_non_parallel() { 4887 ResourceMark rm; 4888 HandleMark hm; 4889 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4890 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false); 4891 4892 MarkRefsIntoAndScanClosure 4893 mrias_cl(_span, ref_processor(), &_markBitMap, NULL /* not precleaning */, 4894 &_markStack, this, 4895 false /* should_yield */, false /* not precleaning */); 4896 MarkFromDirtyCardsClosure 4897 markFromDirtyCardsClosure(this, _span, 4898 NULL, // space is set further below 4899 &_markBitMap, &_markStack, &mrias_cl); 4900 { 4901 GCTraceTime(Trace, gc) t("Grey Object Rescan", _gc_timer_cm); 4902 // Iterate over the dirty cards, setting the corresponding bits in the 4903 // mod union table. 4904 { 4905 ModUnionClosure modUnionClosure(&_modUnionTable); 4906 _ct->ct_bs()->dirty_card_iterate( 4907 _cmsGen->used_region(), 4908 &modUnionClosure); 4909 } 4910 // Having transferred these marks into the modUnionTable, we just need 4911 // to rescan the marked objects on the dirty cards in the modUnionTable. 4912 // The initial marking may have been done during an asynchronous 4913 // collection so there may be dirty bits in the mod-union table. 4914 const int alignment = 4915 CardTableModRefBS::card_size * BitsPerWord; 4916 { 4917 // ... First handle dirty cards in CMS gen 4918 markFromDirtyCardsClosure.set_space(_cmsGen->cmsSpace()); 4919 MemRegion ur = _cmsGen->used_region(); 4920 HeapWord* lb = ur.start(); 4921 HeapWord* ub = (HeapWord*)round_to((intptr_t)ur.end(), alignment); 4922 MemRegion cms_span(lb, ub); 4923 _modUnionTable.dirty_range_iterate_clear(cms_span, 4924 &markFromDirtyCardsClosure); 4925 verify_work_stacks_empty(); 4926 log_trace(gc)(" (re-scanned " SIZE_FORMAT " dirty cards in cms gen) ", markFromDirtyCardsClosure.num_dirty_cards()); 4927 } 4928 } 4929 if (VerifyDuringGC && 4930 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 4931 HandleMark hm; // Discard invalid handles created during verification 4932 Universe::verify(); 4933 } 4934 { 4935 GCTraceTime(Trace, gc) t("Root Rescan", _gc_timer_cm); 4936 4937 verify_work_stacks_empty(); 4938 4939 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 4940 StrongRootsScope srs(1); 4941 4942 gch->gen_process_roots(&srs, 4943 GenCollectedHeap::OldGen, 4944 true, // young gen as roots 4945 GenCollectedHeap::ScanningOption(roots_scanning_options()), 4946 should_unload_classes(), 4947 &mrias_cl, 4948 NULL, 4949 NULL); // The dirty klasses will be handled below 4950 4951 assert(should_unload_classes() 4952 || (roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache), 4953 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 4954 } 4955 4956 { 4957 GCTraceTime(Trace, gc) t("Visit Unhandled CLDs", _gc_timer_cm); 4958 4959 verify_work_stacks_empty(); 4960 4961 // Scan all class loader data objects that might have been introduced 4962 // during concurrent marking. 4963 ResourceMark rm; 4964 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds(); 4965 for (int i = 0; i < array->length(); i++) { 4966 mrias_cl.do_cld_nv(array->at(i)); 4967 } 4968 4969 // We don't need to keep track of new CLDs anymore. 4970 ClassLoaderDataGraph::remember_new_clds(false); 4971 4972 verify_work_stacks_empty(); 4973 } 4974 4975 { 4976 GCTraceTime(Trace, gc) t("Dirty Klass Scan", _gc_timer_cm); 4977 4978 verify_work_stacks_empty(); 4979 4980 RemarkKlassClosure remark_klass_closure(&mrias_cl); 4981 ClassLoaderDataGraph::classes_do(&remark_klass_closure); 4982 4983 verify_work_stacks_empty(); 4984 } 4985 4986 // We might have added oops to ClassLoaderData::_handles during the 4987 // concurrent marking phase. These oops point to newly allocated objects 4988 // that are guaranteed to be kept alive. Either by the direct allocation 4989 // code, or when the young collector processes the roots. Hence, 4990 // we don't have to revisit the _handles block during the remark phase. 4991 4992 verify_work_stacks_empty(); 4993 // Restore evacuated mark words, if any, used for overflow list links 4994 restore_preserved_marks_if_any(); 4995 4996 verify_overflow_empty(); 4997 } 4998 4999 //////////////////////////////////////////////////////// 5000 // Parallel Reference Processing Task Proxy Class 5001 //////////////////////////////////////////////////////// 5002 class AbstractGangTaskWOopQueues : public AbstractGangTask { 5003 OopTaskQueueSet* _queues; 5004 ParallelTaskTerminator _terminator; 5005 public: 5006 AbstractGangTaskWOopQueues(const char* name, OopTaskQueueSet* queues, uint n_threads) : 5007 AbstractGangTask(name), _queues(queues), _terminator(n_threads, _queues) {} 5008 ParallelTaskTerminator* terminator() { return &_terminator; } 5009 OopTaskQueueSet* queues() { return _queues; } 5010 }; 5011 5012 class CMSRefProcTaskProxy: public AbstractGangTaskWOopQueues { 5013 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask; 5014 CMSCollector* _collector; 5015 CMSBitMap* _mark_bit_map; 5016 const MemRegion _span; 5017 ProcessTask& _task; 5018 5019 public: 5020 CMSRefProcTaskProxy(ProcessTask& task, 5021 CMSCollector* collector, 5022 const MemRegion& span, 5023 CMSBitMap* mark_bit_map, 5024 AbstractWorkGang* workers, 5025 OopTaskQueueSet* task_queues): 5026 AbstractGangTaskWOopQueues("Process referents by policy in parallel", 5027 task_queues, 5028 workers->active_workers()), 5029 _task(task), 5030 _collector(collector), _span(span), _mark_bit_map(mark_bit_map) 5031 { 5032 assert(_collector->_span.equals(_span) && !_span.is_empty(), 5033 "Inconsistency in _span"); 5034 } 5035 5036 OopTaskQueueSet* task_queues() { return queues(); } 5037 5038 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 5039 5040 void do_work_steal(int i, 5041 CMSParDrainMarkingStackClosure* drain, 5042 CMSParKeepAliveClosure* keep_alive, 5043 int* seed); 5044 5045 virtual void work(uint worker_id); 5046 }; 5047 5048 void CMSRefProcTaskProxy::work(uint worker_id) { 5049 ResourceMark rm; 5050 HandleMark hm; 5051 assert(_collector->_span.equals(_span), "Inconsistency in _span"); 5052 CMSParKeepAliveClosure par_keep_alive(_collector, _span, 5053 _mark_bit_map, 5054 work_queue(worker_id)); 5055 CMSParDrainMarkingStackClosure par_drain_stack(_collector, _span, 5056 _mark_bit_map, 5057 work_queue(worker_id)); 5058 CMSIsAliveClosure is_alive_closure(_span, _mark_bit_map); 5059 _task.work(worker_id, is_alive_closure, par_keep_alive, par_drain_stack); 5060 if (_task.marks_oops_alive()) { 5061 do_work_steal(worker_id, &par_drain_stack, &par_keep_alive, 5062 _collector->hash_seed(worker_id)); 5063 } 5064 assert(work_queue(worker_id)->size() == 0, "work_queue should be empty"); 5065 assert(_collector->_overflow_list == NULL, "non-empty _overflow_list"); 5066 } 5067 5068 class CMSRefEnqueueTaskProxy: public AbstractGangTask { 5069 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask; 5070 EnqueueTask& _task; 5071 5072 public: 5073 CMSRefEnqueueTaskProxy(EnqueueTask& task) 5074 : AbstractGangTask("Enqueue reference objects in parallel"), 5075 _task(task) 5076 { } 5077 5078 virtual void work(uint worker_id) 5079 { 5080 _task.work(worker_id); 5081 } 5082 }; 5083 5084 CMSParKeepAliveClosure::CMSParKeepAliveClosure(CMSCollector* collector, 5085 MemRegion span, CMSBitMap* bit_map, OopTaskQueue* work_queue): 5086 _span(span), 5087 _bit_map(bit_map), 5088 _work_queue(work_queue), 5089 _mark_and_push(collector, span, bit_map, work_queue), 5090 _low_water_mark(MIN2((work_queue->max_elems()/4), 5091 ((uint)CMSWorkQueueDrainThreshold * ParallelGCThreads))) 5092 { } 5093 5094 // . see if we can share work_queues with ParNew? XXX 5095 void CMSRefProcTaskProxy::do_work_steal(int i, 5096 CMSParDrainMarkingStackClosure* drain, 5097 CMSParKeepAliveClosure* keep_alive, 5098 int* seed) { 5099 OopTaskQueue* work_q = work_queue(i); 5100 NOT_PRODUCT(int num_steals = 0;) 5101 oop obj_to_scan; 5102 5103 while (true) { 5104 // Completely finish any left over work from (an) earlier round(s) 5105 drain->trim_queue(0); 5106 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 5107 (size_t)ParGCDesiredObjsFromOverflowList); 5108 // Now check if there's any work in the overflow list 5109 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads, 5110 // only affects the number of attempts made to get work from the 5111 // overflow list and does not affect the number of workers. Just 5112 // pass ParallelGCThreads so this behavior is unchanged. 5113 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 5114 work_q, 5115 ParallelGCThreads)) { 5116 // Found something in global overflow list; 5117 // not yet ready to go stealing work from others. 5118 // We'd like to assert(work_q->size() != 0, ...) 5119 // because we just took work from the overflow list, 5120 // but of course we can't, since all of that might have 5121 // been already stolen from us. 5122 continue; 5123 } 5124 // Verify that we have no work before we resort to stealing 5125 assert(work_q->size() == 0, "Have work, shouldn't steal"); 5126 // Try to steal from other queues that have work 5127 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 5128 NOT_PRODUCT(num_steals++;) 5129 assert(obj_to_scan->is_oop(), "Oops, not an oop!"); 5130 assert(_mark_bit_map->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 5131 // Do scanning work 5132 obj_to_scan->oop_iterate(keep_alive); 5133 // Loop around, finish this work, and try to steal some more 5134 } else if (terminator()->offer_termination()) { 5135 break; // nirvana from the infinite cycle 5136 } 5137 } 5138 log_develop_trace(gc, task)("\t(%d: stole %d oops)", i, num_steals); 5139 } 5140 5141 void CMSRefProcTaskExecutor::execute(ProcessTask& task) 5142 { 5143 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5144 WorkGang* workers = gch->workers(); 5145 assert(workers != NULL, "Need parallel worker threads."); 5146 CMSRefProcTaskProxy rp_task(task, &_collector, 5147 _collector.ref_processor()->span(), 5148 _collector.markBitMap(), 5149 workers, _collector.task_queues()); 5150 workers->run_task(&rp_task); 5151 } 5152 5153 void CMSRefProcTaskExecutor::execute(EnqueueTask& task) 5154 { 5155 5156 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5157 WorkGang* workers = gch->workers(); 5158 assert(workers != NULL, "Need parallel worker threads."); 5159 CMSRefEnqueueTaskProxy enq_task(task); 5160 workers->run_task(&enq_task); 5161 } 5162 5163 void CMSCollector::refProcessingWork() { 5164 ResourceMark rm; 5165 HandleMark hm; 5166 5167 ReferenceProcessor* rp = ref_processor(); 5168 assert(rp->span().equals(_span), "Spans should be equal"); 5169 assert(!rp->enqueuing_is_done(), "Enqueuing should not be complete"); 5170 // Process weak references. 5171 rp->setup_policy(false); 5172 verify_work_stacks_empty(); 5173 5174 CMSKeepAliveClosure cmsKeepAliveClosure(this, _span, &_markBitMap, 5175 &_markStack, false /* !preclean */); 5176 CMSDrainMarkingStackClosure cmsDrainMarkingStackClosure(this, 5177 _span, &_markBitMap, &_markStack, 5178 &cmsKeepAliveClosure, false /* !preclean */); 5179 { 5180 GCTraceTime(Debug, gc) t("Weak Refs Processing", _gc_timer_cm); 5181 5182 ReferenceProcessorStats stats; 5183 if (rp->processing_is_mt()) { 5184 // Set the degree of MT here. If the discovery is done MT, there 5185 // may have been a different number of threads doing the discovery 5186 // and a different number of discovered lists may have Ref objects. 5187 // That is OK as long as the Reference lists are balanced (see 5188 // balance_all_queues() and balance_queues()). 5189 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5190 uint active_workers = ParallelGCThreads; 5191 WorkGang* workers = gch->workers(); 5192 if (workers != NULL) { 5193 active_workers = workers->active_workers(); 5194 // The expectation is that active_workers will have already 5195 // been set to a reasonable value. If it has not been set, 5196 // investigate. 5197 assert(active_workers > 0, "Should have been set during scavenge"); 5198 } 5199 rp->set_active_mt_degree(active_workers); 5200 CMSRefProcTaskExecutor task_executor(*this); 5201 stats = rp->process_discovered_references(&_is_alive_closure, 5202 &cmsKeepAliveClosure, 5203 &cmsDrainMarkingStackClosure, 5204 &task_executor, 5205 _gc_timer_cm); 5206 } else { 5207 stats = rp->process_discovered_references(&_is_alive_closure, 5208 &cmsKeepAliveClosure, 5209 &cmsDrainMarkingStackClosure, 5210 NULL, 5211 _gc_timer_cm); 5212 } 5213 _gc_tracer_cm->report_gc_reference_stats(stats); 5214 5215 } 5216 5217 // This is the point where the entire marking should have completed. 5218 verify_work_stacks_empty(); 5219 5220 if (should_unload_classes()) { 5221 { 5222 GCTraceTime(Debug, gc) t("Class Unloading", _gc_timer_cm); 5223 5224 // Unload classes and purge the SystemDictionary. 5225 bool purged_class = SystemDictionary::do_unloading(&_is_alive_closure); 5226 5227 // Unload nmethods. 5228 CodeCache::do_unloading(&_is_alive_closure, purged_class); 5229 5230 // Prune dead klasses from subklass/sibling/implementor lists. 5231 Klass::clean_weak_klass_links(&_is_alive_closure); 5232 } 5233 5234 { 5235 GCTraceTime(Debug, gc) t("Scrub Symbol Table", _gc_timer_cm); 5236 // Clean up unreferenced symbols in symbol table. 5237 SymbolTable::unlink(); 5238 } 5239 5240 { 5241 GCTraceTime(Debug, gc) t("Scrub String Table", _gc_timer_cm); 5242 // Delete entries for dead interned strings. 5243 StringTable::unlink(&_is_alive_closure); 5244 } 5245 } 5246 5247 5248 // Restore any preserved marks as a result of mark stack or 5249 // work queue overflow 5250 restore_preserved_marks_if_any(); // done single-threaded for now 5251 5252 rp->set_enqueuing_is_done(true); 5253 if (rp->processing_is_mt()) { 5254 rp->balance_all_queues(); 5255 CMSRefProcTaskExecutor task_executor(*this); 5256 rp->enqueue_discovered_references(&task_executor); 5257 } else { 5258 rp->enqueue_discovered_references(NULL); 5259 } 5260 rp->verify_no_references_recorded(); 5261 assert(!rp->discovery_enabled(), "should have been disabled"); 5262 } 5263 5264 #ifndef PRODUCT 5265 void CMSCollector::check_correct_thread_executing() { 5266 Thread* t = Thread::current(); 5267 // Only the VM thread or the CMS thread should be here. 5268 assert(t->is_ConcurrentGC_thread() || t->is_VM_thread(), 5269 "Unexpected thread type"); 5270 // If this is the vm thread, the foreground process 5271 // should not be waiting. Note that _foregroundGCIsActive is 5272 // true while the foreground collector is waiting. 5273 if (_foregroundGCShouldWait) { 5274 // We cannot be the VM thread 5275 assert(t->is_ConcurrentGC_thread(), 5276 "Should be CMS thread"); 5277 } else { 5278 // We can be the CMS thread only if we are in a stop-world 5279 // phase of CMS collection. 5280 if (t->is_ConcurrentGC_thread()) { 5281 assert(_collectorState == InitialMarking || 5282 _collectorState == FinalMarking, 5283 "Should be a stop-world phase"); 5284 // The CMS thread should be holding the CMS_token. 5285 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 5286 "Potential interference with concurrently " 5287 "executing VM thread"); 5288 } 5289 } 5290 } 5291 #endif 5292 5293 void CMSCollector::sweep() { 5294 assert(_collectorState == Sweeping, "just checking"); 5295 check_correct_thread_executing(); 5296 verify_work_stacks_empty(); 5297 verify_overflow_empty(); 5298 increment_sweep_count(); 5299 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 5300 5301 _inter_sweep_timer.stop(); 5302 _inter_sweep_estimate.sample(_inter_sweep_timer.seconds()); 5303 5304 assert(!_intra_sweep_timer.is_active(), "Should not be active"); 5305 _intra_sweep_timer.reset(); 5306 _intra_sweep_timer.start(); 5307 { 5308 GCTraceCPUTime tcpu; 5309 CMSPhaseAccounting pa(this, "Concurrent Sweep"); 5310 // First sweep the old gen 5311 { 5312 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(), 5313 bitMapLock()); 5314 sweepWork(_cmsGen); 5315 } 5316 5317 // Update Universe::_heap_*_at_gc figures. 5318 // We need all the free list locks to make the abstract state 5319 // transition from Sweeping to Resetting. See detailed note 5320 // further below. 5321 { 5322 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock()); 5323 // Update heap occupancy information which is used as 5324 // input to soft ref clearing policy at the next gc. 5325 Universe::update_heap_info_at_gc(); 5326 _collectorState = Resizing; 5327 } 5328 } 5329 verify_work_stacks_empty(); 5330 verify_overflow_empty(); 5331 5332 if (should_unload_classes()) { 5333 // Delay purge to the beginning of the next safepoint. Metaspace::contains 5334 // requires that the virtual spaces are stable and not deleted. 5335 ClassLoaderDataGraph::set_should_purge(true); 5336 } 5337 5338 _intra_sweep_timer.stop(); 5339 _intra_sweep_estimate.sample(_intra_sweep_timer.seconds()); 5340 5341 _inter_sweep_timer.reset(); 5342 _inter_sweep_timer.start(); 5343 5344 // We need to use a monotonically non-decreasing time in ms 5345 // or we will see time-warp warnings and os::javaTimeMillis() 5346 // does not guarantee monotonicity. 5347 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC; 5348 update_time_of_last_gc(now); 5349 5350 // NOTE on abstract state transitions: 5351 // Mutators allocate-live and/or mark the mod-union table dirty 5352 // based on the state of the collection. The former is done in 5353 // the interval [Marking, Sweeping] and the latter in the interval 5354 // [Marking, Sweeping). Thus the transitions into the Marking state 5355 // and out of the Sweeping state must be synchronously visible 5356 // globally to the mutators. 5357 // The transition into the Marking state happens with the world 5358 // stopped so the mutators will globally see it. Sweeping is 5359 // done asynchronously by the background collector so the transition 5360 // from the Sweeping state to the Resizing state must be done 5361 // under the freelistLock (as is the check for whether to 5362 // allocate-live and whether to dirty the mod-union table). 5363 assert(_collectorState == Resizing, "Change of collector state to" 5364 " Resizing must be done under the freelistLocks (plural)"); 5365 5366 // Now that sweeping has been completed, we clear 5367 // the incremental_collection_failed flag, 5368 // thus inviting a younger gen collection to promote into 5369 // this generation. If such a promotion may still fail, 5370 // the flag will be set again when a young collection is 5371 // attempted. 5372 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5373 gch->clear_incremental_collection_failed(); // Worth retrying as fresh space may have been freed up 5374 gch->update_full_collections_completed(_collection_count_start); 5375 } 5376 5377 // FIX ME!!! Looks like this belongs in CFLSpace, with 5378 // CMSGen merely delegating to it. 5379 void ConcurrentMarkSweepGeneration::setNearLargestChunk() { 5380 double nearLargestPercent = FLSLargestBlockCoalesceProximity; 5381 HeapWord* minAddr = _cmsSpace->bottom(); 5382 HeapWord* largestAddr = 5383 (HeapWord*) _cmsSpace->dictionary()->find_largest_dict(); 5384 if (largestAddr == NULL) { 5385 // The dictionary appears to be empty. In this case 5386 // try to coalesce at the end of the heap. 5387 largestAddr = _cmsSpace->end(); 5388 } 5389 size_t largestOffset = pointer_delta(largestAddr, minAddr); 5390 size_t nearLargestOffset = 5391 (size_t)((double)largestOffset * nearLargestPercent) - MinChunkSize; 5392 log_debug(gc, freelist)("CMS: Large Block: " PTR_FORMAT "; Proximity: " PTR_FORMAT " -> " PTR_FORMAT, 5393 p2i(largestAddr), p2i(_cmsSpace->nearLargestChunk()), p2i(minAddr + nearLargestOffset)); 5394 _cmsSpace->set_nearLargestChunk(minAddr + nearLargestOffset); 5395 } 5396 5397 bool ConcurrentMarkSweepGeneration::isNearLargestChunk(HeapWord* addr) { 5398 return addr >= _cmsSpace->nearLargestChunk(); 5399 } 5400 5401 FreeChunk* ConcurrentMarkSweepGeneration::find_chunk_at_end() { 5402 return _cmsSpace->find_chunk_at_end(); 5403 } 5404 5405 void ConcurrentMarkSweepGeneration::update_gc_stats(Generation* current_generation, 5406 bool full) { 5407 // If the young generation has been collected, gather any statistics 5408 // that are of interest at this point. 5409 bool current_is_young = GenCollectedHeap::heap()->is_young_gen(current_generation); 5410 if (!full && current_is_young) { 5411 // Gather statistics on the young generation collection. 5412 collector()->stats().record_gc0_end(used()); 5413 } 5414 } 5415 5416 void CMSCollector::sweepWork(ConcurrentMarkSweepGeneration* old_gen) { 5417 // We iterate over the space(s) underlying this generation, 5418 // checking the mark bit map to see if the bits corresponding 5419 // to specific blocks are marked or not. Blocks that are 5420 // marked are live and are not swept up. All remaining blocks 5421 // are swept up, with coalescing on-the-fly as we sweep up 5422 // contiguous free and/or garbage blocks: 5423 // We need to ensure that the sweeper synchronizes with allocators 5424 // and stop-the-world collectors. In particular, the following 5425 // locks are used: 5426 // . CMS token: if this is held, a stop the world collection cannot occur 5427 // . freelistLock: if this is held no allocation can occur from this 5428 // generation by another thread 5429 // . bitMapLock: if this is held, no other thread can access or update 5430 // 5431 5432 // Note that we need to hold the freelistLock if we use 5433 // block iterate below; else the iterator might go awry if 5434 // a mutator (or promotion) causes block contents to change 5435 // (for instance if the allocator divvies up a block). 5436 // If we hold the free list lock, for all practical purposes 5437 // young generation GC's can't occur (they'll usually need to 5438 // promote), so we might as well prevent all young generation 5439 // GC's while we do a sweeping step. For the same reason, we might 5440 // as well take the bit map lock for the entire duration 5441 5442 // check that we hold the requisite locks 5443 assert(have_cms_token(), "Should hold cms token"); 5444 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), "Should possess CMS token to sweep"); 5445 assert_lock_strong(old_gen->freelistLock()); 5446 assert_lock_strong(bitMapLock()); 5447 5448 assert(!_inter_sweep_timer.is_active(), "Was switched off in an outer context"); 5449 assert(_intra_sweep_timer.is_active(), "Was switched on in an outer context"); 5450 old_gen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()), 5451 _inter_sweep_estimate.padded_average(), 5452 _intra_sweep_estimate.padded_average()); 5453 old_gen->setNearLargestChunk(); 5454 5455 { 5456 SweepClosure sweepClosure(this, old_gen, &_markBitMap, CMSYield); 5457 old_gen->cmsSpace()->blk_iterate_careful(&sweepClosure); 5458 // We need to free-up/coalesce garbage/blocks from a 5459 // co-terminal free run. This is done in the SweepClosure 5460 // destructor; so, do not remove this scope, else the 5461 // end-of-sweep-census below will be off by a little bit. 5462 } 5463 old_gen->cmsSpace()->sweep_completed(); 5464 old_gen->cmsSpace()->endSweepFLCensus(sweep_count()); 5465 if (should_unload_classes()) { // unloaded classes this cycle, 5466 _concurrent_cycles_since_last_unload = 0; // ... reset count 5467 } else { // did not unload classes, 5468 _concurrent_cycles_since_last_unload++; // ... increment count 5469 } 5470 } 5471 5472 // Reset CMS data structures (for now just the marking bit map) 5473 // preparatory for the next cycle. 5474 void CMSCollector::reset_concurrent() { 5475 CMSTokenSyncWithLocks ts(true, bitMapLock()); 5476 5477 // If the state is not "Resetting", the foreground thread 5478 // has done a collection and the resetting. 5479 if (_collectorState != Resetting) { 5480 assert(_collectorState == Idling, "The state should only change" 5481 " because the foreground collector has finished the collection"); 5482 return; 5483 } 5484 5485 // Clear the mark bitmap (no grey objects to start with) 5486 // for the next cycle. 5487 GCTraceCPUTime tcpu; 5488 CMSPhaseAccounting cmspa(this, "Concurrent Reset"); 5489 5490 HeapWord* curAddr = _markBitMap.startWord(); 5491 while (curAddr < _markBitMap.endWord()) { 5492 size_t remaining = pointer_delta(_markBitMap.endWord(), curAddr); 5493 MemRegion chunk(curAddr, MIN2(CMSBitMapYieldQuantum, remaining)); 5494 _markBitMap.clear_large_range(chunk); 5495 if (ConcurrentMarkSweepThread::should_yield() && 5496 !foregroundGCIsActive() && 5497 CMSYield) { 5498 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 5499 "CMS thread should hold CMS token"); 5500 assert_lock_strong(bitMapLock()); 5501 bitMapLock()->unlock(); 5502 ConcurrentMarkSweepThread::desynchronize(true); 5503 stopTimer(); 5504 incrementYields(); 5505 5506 // See the comment in coordinator_yield() 5507 for (unsigned i = 0; i < CMSYieldSleepCount && 5508 ConcurrentMarkSweepThread::should_yield() && 5509 !CMSCollector::foregroundGCIsActive(); ++i) { 5510 os::sleep(Thread::current(), 1, false); 5511 } 5512 5513 ConcurrentMarkSweepThread::synchronize(true); 5514 bitMapLock()->lock_without_safepoint_check(); 5515 startTimer(); 5516 } 5517 curAddr = chunk.end(); 5518 } 5519 // A successful mostly concurrent collection has been done. 5520 // Because only the full (i.e., concurrent mode failure) collections 5521 // are being measured for gc overhead limits, clean the "near" flag 5522 // and count. 5523 size_policy()->reset_gc_overhead_limit_count(); 5524 _collectorState = Idling; 5525 5526 register_gc_end(); 5527 } 5528 5529 // Same as above but for STW paths 5530 void CMSCollector::reset_stw() { 5531 // already have the lock 5532 assert(_collectorState == Resetting, "just checking"); 5533 assert_lock_strong(bitMapLock()); 5534 GCIdMarkAndRestore gc_id_mark(_cmsThread->gc_id()); 5535 _markBitMap.clear_all(); 5536 _collectorState = Idling; 5537 register_gc_end(); 5538 } 5539 5540 void CMSCollector::do_CMS_operation(CMS_op_type op, GCCause::Cause gc_cause) { 5541 GCTraceCPUTime tcpu; 5542 TraceCollectorStats tcs(counters()); 5543 5544 switch (op) { 5545 case CMS_op_checkpointRootsInitial: { 5546 GCTraceTime(Info, gc) t("Pause Initial Mark", NULL, GCCause::_no_gc, true); 5547 SvcGCMarker sgcm(SvcGCMarker::OTHER); 5548 checkpointRootsInitial(); 5549 break; 5550 } 5551 case CMS_op_checkpointRootsFinal: { 5552 GCTraceTime(Info, gc) t("Pause Remark", NULL, GCCause::_no_gc, true); 5553 SvcGCMarker sgcm(SvcGCMarker::OTHER); 5554 checkpointRootsFinal(); 5555 break; 5556 } 5557 default: 5558 fatal("No such CMS_op"); 5559 } 5560 } 5561 5562 #ifndef PRODUCT 5563 size_t const CMSCollector::skip_header_HeapWords() { 5564 return FreeChunk::header_size(); 5565 } 5566 5567 // Try and collect here conditions that should hold when 5568 // CMS thread is exiting. The idea is that the foreground GC 5569 // thread should not be blocked if it wants to terminate 5570 // the CMS thread and yet continue to run the VM for a while 5571 // after that. 5572 void CMSCollector::verify_ok_to_terminate() const { 5573 assert(Thread::current()->is_ConcurrentGC_thread(), 5574 "should be called by CMS thread"); 5575 assert(!_foregroundGCShouldWait, "should be false"); 5576 // We could check here that all the various low-level locks 5577 // are not held by the CMS thread, but that is overkill; see 5578 // also CMSThread::verify_ok_to_terminate() where the CGC_lock 5579 // is checked. 5580 } 5581 #endif 5582 5583 size_t CMSCollector::block_size_using_printezis_bits(HeapWord* addr) const { 5584 assert(_markBitMap.isMarked(addr) && _markBitMap.isMarked(addr + 1), 5585 "missing Printezis mark?"); 5586 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 5587 size_t size = pointer_delta(nextOneAddr + 1, addr); 5588 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 5589 "alignment problem"); 5590 assert(size >= 3, "Necessary for Printezis marks to work"); 5591 return size; 5592 } 5593 5594 // A variant of the above (block_size_using_printezis_bits()) except 5595 // that we return 0 if the P-bits are not yet set. 5596 size_t CMSCollector::block_size_if_printezis_bits(HeapWord* addr) const { 5597 if (_markBitMap.isMarked(addr + 1)) { 5598 assert(_markBitMap.isMarked(addr), "P-bit can be set only for marked objects"); 5599 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 5600 size_t size = pointer_delta(nextOneAddr + 1, addr); 5601 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 5602 "alignment problem"); 5603 assert(size >= 3, "Necessary for Printezis marks to work"); 5604 return size; 5605 } 5606 return 0; 5607 } 5608 5609 HeapWord* CMSCollector::next_card_start_after_block(HeapWord* addr) const { 5610 size_t sz = 0; 5611 oop p = (oop)addr; 5612 if (p->klass_or_null() != NULL) { 5613 sz = CompactibleFreeListSpace::adjustObjectSize(p->size()); 5614 } else { 5615 sz = block_size_using_printezis_bits(addr); 5616 } 5617 assert(sz > 0, "size must be nonzero"); 5618 HeapWord* next_block = addr + sz; 5619 HeapWord* next_card = (HeapWord*)round_to((uintptr_t)next_block, 5620 CardTableModRefBS::card_size); 5621 assert(round_down((uintptr_t)addr, CardTableModRefBS::card_size) < 5622 round_down((uintptr_t)next_card, CardTableModRefBS::card_size), 5623 "must be different cards"); 5624 return next_card; 5625 } 5626 5627 5628 // CMS Bit Map Wrapper ///////////////////////////////////////// 5629 5630 // Construct a CMS bit map infrastructure, but don't create the 5631 // bit vector itself. That is done by a separate call CMSBitMap::allocate() 5632 // further below. 5633 CMSBitMap::CMSBitMap(int shifter, int mutex_rank, const char* mutex_name): 5634 _bm(), 5635 _shifter(shifter), 5636 _lock(mutex_rank >= 0 ? new Mutex(mutex_rank, mutex_name, true, 5637 Monitor::_safepoint_check_sometimes) : NULL) 5638 { 5639 _bmStartWord = 0; 5640 _bmWordSize = 0; 5641 } 5642 5643 bool CMSBitMap::allocate(MemRegion mr) { 5644 _bmStartWord = mr.start(); 5645 _bmWordSize = mr.word_size(); 5646 ReservedSpace brs(ReservedSpace::allocation_align_size_up( 5647 (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1)); 5648 if (!brs.is_reserved()) { 5649 warning("CMS bit map allocation failure"); 5650 return false; 5651 } 5652 // For now we'll just commit all of the bit map up front. 5653 // Later on we'll try to be more parsimonious with swap. 5654 if (!_virtual_space.initialize(brs, brs.size())) { 5655 warning("CMS bit map backing store failure"); 5656 return false; 5657 } 5658 assert(_virtual_space.committed_size() == brs.size(), 5659 "didn't reserve backing store for all of CMS bit map?"); 5660 _bm.set_map((BitMap::bm_word_t*)_virtual_space.low()); 5661 assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >= 5662 _bmWordSize, "inconsistency in bit map sizing"); 5663 _bm.set_size(_bmWordSize >> _shifter); 5664 5665 // bm.clear(); // can we rely on getting zero'd memory? verify below 5666 assert(isAllClear(), 5667 "Expected zero'd memory from ReservedSpace constructor"); 5668 assert(_bm.size() == heapWordDiffToOffsetDiff(sizeInWords()), 5669 "consistency check"); 5670 return true; 5671 } 5672 5673 void CMSBitMap::dirty_range_iterate_clear(MemRegion mr, MemRegionClosure* cl) { 5674 HeapWord *next_addr, *end_addr, *last_addr; 5675 assert_locked(); 5676 assert(covers(mr), "out-of-range error"); 5677 // XXX assert that start and end are appropriately aligned 5678 for (next_addr = mr.start(), end_addr = mr.end(); 5679 next_addr < end_addr; next_addr = last_addr) { 5680 MemRegion dirty_region = getAndClearMarkedRegion(next_addr, end_addr); 5681 last_addr = dirty_region.end(); 5682 if (!dirty_region.is_empty()) { 5683 cl->do_MemRegion(dirty_region); 5684 } else { 5685 assert(last_addr == end_addr, "program logic"); 5686 return; 5687 } 5688 } 5689 } 5690 5691 void CMSBitMap::print_on_error(outputStream* st, const char* prefix) const { 5692 _bm.print_on_error(st, prefix); 5693 } 5694 5695 #ifndef PRODUCT 5696 void CMSBitMap::assert_locked() const { 5697 CMSLockVerifier::assert_locked(lock()); 5698 } 5699 5700 bool CMSBitMap::covers(MemRegion mr) const { 5701 // assert(_bm.map() == _virtual_space.low(), "map inconsistency"); 5702 assert((size_t)_bm.size() == (_bmWordSize >> _shifter), 5703 "size inconsistency"); 5704 return (mr.start() >= _bmStartWord) && 5705 (mr.end() <= endWord()); 5706 } 5707 5708 bool CMSBitMap::covers(HeapWord* start, size_t size) const { 5709 return (start >= _bmStartWord && (start + size) <= endWord()); 5710 } 5711 5712 void CMSBitMap::verifyNoOneBitsInRange(HeapWord* left, HeapWord* right) { 5713 // verify that there are no 1 bits in the interval [left, right) 5714 FalseBitMapClosure falseBitMapClosure; 5715 iterate(&falseBitMapClosure, left, right); 5716 } 5717 5718 void CMSBitMap::region_invariant(MemRegion mr) 5719 { 5720 assert_locked(); 5721 // mr = mr.intersection(MemRegion(_bmStartWord, _bmWordSize)); 5722 assert(!mr.is_empty(), "unexpected empty region"); 5723 assert(covers(mr), "mr should be covered by bit map"); 5724 // convert address range into offset range 5725 size_t start_ofs = heapWordToOffset(mr.start()); 5726 // Make sure that end() is appropriately aligned 5727 assert(mr.end() == (HeapWord*)round_to((intptr_t)mr.end(), 5728 (1 << (_shifter+LogHeapWordSize))), 5729 "Misaligned mr.end()"); 5730 size_t end_ofs = heapWordToOffset(mr.end()); 5731 assert(end_ofs > start_ofs, "Should mark at least one bit"); 5732 } 5733 5734 #endif 5735 5736 bool CMSMarkStack::allocate(size_t size) { 5737 // allocate a stack of the requisite depth 5738 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 5739 size * sizeof(oop))); 5740 if (!rs.is_reserved()) { 5741 warning("CMSMarkStack allocation failure"); 5742 return false; 5743 } 5744 if (!_virtual_space.initialize(rs, rs.size())) { 5745 warning("CMSMarkStack backing store failure"); 5746 return false; 5747 } 5748 assert(_virtual_space.committed_size() == rs.size(), 5749 "didn't reserve backing store for all of CMS stack?"); 5750 _base = (oop*)(_virtual_space.low()); 5751 _index = 0; 5752 _capacity = size; 5753 NOT_PRODUCT(_max_depth = 0); 5754 return true; 5755 } 5756 5757 // XXX FIX ME !!! In the MT case we come in here holding a 5758 // leaf lock. For printing we need to take a further lock 5759 // which has lower rank. We need to recalibrate the two 5760 // lock-ranks involved in order to be able to print the 5761 // messages below. (Or defer the printing to the caller. 5762 // For now we take the expedient path of just disabling the 5763 // messages for the problematic case.) 5764 void CMSMarkStack::expand() { 5765 assert(_capacity <= MarkStackSizeMax, "stack bigger than permitted"); 5766 if (_capacity == MarkStackSizeMax) { 5767 if (_hit_limit++ == 0 && !CMSConcurrentMTEnabled) { 5768 // We print a warning message only once per CMS cycle. 5769 log_debug(gc)(" (benign) Hit CMSMarkStack max size limit"); 5770 } 5771 return; 5772 } 5773 // Double capacity if possible 5774 size_t new_capacity = MIN2(_capacity*2, MarkStackSizeMax); 5775 // Do not give up existing stack until we have managed to 5776 // get the double capacity that we desired. 5777 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 5778 new_capacity * sizeof(oop))); 5779 if (rs.is_reserved()) { 5780 // Release the backing store associated with old stack 5781 _virtual_space.release(); 5782 // Reinitialize virtual space for new stack 5783 if (!_virtual_space.initialize(rs, rs.size())) { 5784 fatal("Not enough swap for expanded marking stack"); 5785 } 5786 _base = (oop*)(_virtual_space.low()); 5787 _index = 0; 5788 _capacity = new_capacity; 5789 } else if (_failed_double++ == 0 && !CMSConcurrentMTEnabled) { 5790 // Failed to double capacity, continue; 5791 // we print a detail message only once per CMS cycle. 5792 log_debug(gc)(" (benign) Failed to expand marking stack from " SIZE_FORMAT "K to " SIZE_FORMAT "K", 5793 _capacity / K, new_capacity / K); 5794 } 5795 } 5796 5797 5798 // Closures 5799 // XXX: there seems to be a lot of code duplication here; 5800 // should refactor and consolidate common code. 5801 5802 // This closure is used to mark refs into the CMS generation in 5803 // the CMS bit map. Called at the first checkpoint. This closure 5804 // assumes that we do not need to re-mark dirty cards; if the CMS 5805 // generation on which this is used is not an oldest 5806 // generation then this will lose younger_gen cards! 5807 5808 MarkRefsIntoClosure::MarkRefsIntoClosure( 5809 MemRegion span, CMSBitMap* bitMap): 5810 _span(span), 5811 _bitMap(bitMap) 5812 { 5813 assert(ref_processor() == NULL, "deliberately left NULL"); 5814 assert(_bitMap->covers(_span), "_bitMap/_span mismatch"); 5815 } 5816 5817 void MarkRefsIntoClosure::do_oop(oop obj) { 5818 // if p points into _span, then mark corresponding bit in _markBitMap 5819 assert(obj->is_oop(), "expected an oop"); 5820 HeapWord* addr = (HeapWord*)obj; 5821 if (_span.contains(addr)) { 5822 // this should be made more efficient 5823 _bitMap->mark(addr); 5824 } 5825 } 5826 5827 void MarkRefsIntoClosure::do_oop(oop* p) { MarkRefsIntoClosure::do_oop_work(p); } 5828 void MarkRefsIntoClosure::do_oop(narrowOop* p) { MarkRefsIntoClosure::do_oop_work(p); } 5829 5830 Par_MarkRefsIntoClosure::Par_MarkRefsIntoClosure( 5831 MemRegion span, CMSBitMap* bitMap): 5832 _span(span), 5833 _bitMap(bitMap) 5834 { 5835 assert(ref_processor() == NULL, "deliberately left NULL"); 5836 assert(_bitMap->covers(_span), "_bitMap/_span mismatch"); 5837 } 5838 5839 void Par_MarkRefsIntoClosure::do_oop(oop obj) { 5840 // if p points into _span, then mark corresponding bit in _markBitMap 5841 assert(obj->is_oop(), "expected an oop"); 5842 HeapWord* addr = (HeapWord*)obj; 5843 if (_span.contains(addr)) { 5844 // this should be made more efficient 5845 _bitMap->par_mark(addr); 5846 } 5847 } 5848 5849 void Par_MarkRefsIntoClosure::do_oop(oop* p) { Par_MarkRefsIntoClosure::do_oop_work(p); } 5850 void Par_MarkRefsIntoClosure::do_oop(narrowOop* p) { Par_MarkRefsIntoClosure::do_oop_work(p); } 5851 5852 // A variant of the above, used for CMS marking verification. 5853 MarkRefsIntoVerifyClosure::MarkRefsIntoVerifyClosure( 5854 MemRegion span, CMSBitMap* verification_bm, CMSBitMap* cms_bm): 5855 _span(span), 5856 _verification_bm(verification_bm), 5857 _cms_bm(cms_bm) 5858 { 5859 assert(ref_processor() == NULL, "deliberately left NULL"); 5860 assert(_verification_bm->covers(_span), "_verification_bm/_span mismatch"); 5861 } 5862 5863 void MarkRefsIntoVerifyClosure::do_oop(oop obj) { 5864 // if p points into _span, then mark corresponding bit in _markBitMap 5865 assert(obj->is_oop(), "expected an oop"); 5866 HeapWord* addr = (HeapWord*)obj; 5867 if (_span.contains(addr)) { 5868 _verification_bm->mark(addr); 5869 if (!_cms_bm->isMarked(addr)) { 5870 LogHandle(gc, verify) log; 5871 ResourceMark rm; 5872 oop(addr)->print_on(log.info_stream()); 5873 log.info(" (" INTPTR_FORMAT " should have been marked)", p2i(addr)); 5874 fatal("... aborting"); 5875 } 5876 } 5877 } 5878 5879 void MarkRefsIntoVerifyClosure::do_oop(oop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); } 5880 void MarkRefsIntoVerifyClosure::do_oop(narrowOop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); } 5881 5882 ////////////////////////////////////////////////// 5883 // MarkRefsIntoAndScanClosure 5884 ////////////////////////////////////////////////// 5885 5886 MarkRefsIntoAndScanClosure::MarkRefsIntoAndScanClosure(MemRegion span, 5887 ReferenceProcessor* rp, 5888 CMSBitMap* bit_map, 5889 CMSBitMap* mod_union_table, 5890 CMSMarkStack* mark_stack, 5891 CMSCollector* collector, 5892 bool should_yield, 5893 bool concurrent_precleaning): 5894 _collector(collector), 5895 _span(span), 5896 _bit_map(bit_map), 5897 _mark_stack(mark_stack), 5898 _pushAndMarkClosure(collector, span, rp, bit_map, mod_union_table, 5899 mark_stack, concurrent_precleaning), 5900 _yield(should_yield), 5901 _concurrent_precleaning(concurrent_precleaning), 5902 _freelistLock(NULL) 5903 { 5904 // FIXME: Should initialize in base class constructor. 5905 assert(rp != NULL, "ref_processor shouldn't be NULL"); 5906 set_ref_processor_internal(rp); 5907 } 5908 5909 // This closure is used to mark refs into the CMS generation at the 5910 // second (final) checkpoint, and to scan and transitively follow 5911 // the unmarked oops. It is also used during the concurrent precleaning 5912 // phase while scanning objects on dirty cards in the CMS generation. 5913 // The marks are made in the marking bit map and the marking stack is 5914 // used for keeping the (newly) grey objects during the scan. 5915 // The parallel version (Par_...) appears further below. 5916 void MarkRefsIntoAndScanClosure::do_oop(oop obj) { 5917 if (obj != NULL) { 5918 assert(obj->is_oop(), "expected an oop"); 5919 HeapWord* addr = (HeapWord*)obj; 5920 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 5921 assert(_collector->overflow_list_is_empty(), 5922 "overflow list should be empty"); 5923 if (_span.contains(addr) && 5924 !_bit_map->isMarked(addr)) { 5925 // mark bit map (object is now grey) 5926 _bit_map->mark(addr); 5927 // push on marking stack (stack should be empty), and drain the 5928 // stack by applying this closure to the oops in the oops popped 5929 // from the stack (i.e. blacken the grey objects) 5930 bool res = _mark_stack->push(obj); 5931 assert(res, "Should have space to push on empty stack"); 5932 do { 5933 oop new_oop = _mark_stack->pop(); 5934 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 5935 assert(_bit_map->isMarked((HeapWord*)new_oop), 5936 "only grey objects on this stack"); 5937 // iterate over the oops in this oop, marking and pushing 5938 // the ones in CMS heap (i.e. in _span). 5939 new_oop->oop_iterate(&_pushAndMarkClosure); 5940 // check if it's time to yield 5941 do_yield_check(); 5942 } while (!_mark_stack->isEmpty() || 5943 (!_concurrent_precleaning && take_from_overflow_list())); 5944 // if marking stack is empty, and we are not doing this 5945 // during precleaning, then check the overflow list 5946 } 5947 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 5948 assert(_collector->overflow_list_is_empty(), 5949 "overflow list was drained above"); 5950 5951 assert(_collector->no_preserved_marks(), 5952 "All preserved marks should have been restored above"); 5953 } 5954 } 5955 5956 void MarkRefsIntoAndScanClosure::do_oop(oop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); } 5957 void MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); } 5958 5959 void MarkRefsIntoAndScanClosure::do_yield_work() { 5960 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 5961 "CMS thread should hold CMS token"); 5962 assert_lock_strong(_freelistLock); 5963 assert_lock_strong(_bit_map->lock()); 5964 // relinquish the free_list_lock and bitMaplock() 5965 _bit_map->lock()->unlock(); 5966 _freelistLock->unlock(); 5967 ConcurrentMarkSweepThread::desynchronize(true); 5968 _collector->stopTimer(); 5969 _collector->incrementYields(); 5970 5971 // See the comment in coordinator_yield() 5972 for (unsigned i = 0; 5973 i < CMSYieldSleepCount && 5974 ConcurrentMarkSweepThread::should_yield() && 5975 !CMSCollector::foregroundGCIsActive(); 5976 ++i) { 5977 os::sleep(Thread::current(), 1, false); 5978 } 5979 5980 ConcurrentMarkSweepThread::synchronize(true); 5981 _freelistLock->lock_without_safepoint_check(); 5982 _bit_map->lock()->lock_without_safepoint_check(); 5983 _collector->startTimer(); 5984 } 5985 5986 /////////////////////////////////////////////////////////// 5987 // Par_MarkRefsIntoAndScanClosure: a parallel version of 5988 // MarkRefsIntoAndScanClosure 5989 /////////////////////////////////////////////////////////// 5990 Par_MarkRefsIntoAndScanClosure::Par_MarkRefsIntoAndScanClosure( 5991 CMSCollector* collector, MemRegion span, ReferenceProcessor* rp, 5992 CMSBitMap* bit_map, OopTaskQueue* work_queue): 5993 _span(span), 5994 _bit_map(bit_map), 5995 _work_queue(work_queue), 5996 _low_water_mark(MIN2((work_queue->max_elems()/4), 5997 ((uint)CMSWorkQueueDrainThreshold * ParallelGCThreads))), 5998 _par_pushAndMarkClosure(collector, span, rp, bit_map, work_queue) 5999 { 6000 // FIXME: Should initialize in base class constructor. 6001 assert(rp != NULL, "ref_processor shouldn't be NULL"); 6002 set_ref_processor_internal(rp); 6003 } 6004 6005 // This closure is used to mark refs into the CMS generation at the 6006 // second (final) checkpoint, and to scan and transitively follow 6007 // the unmarked oops. The marks are made in the marking bit map and 6008 // the work_queue is used for keeping the (newly) grey objects during 6009 // the scan phase whence they are also available for stealing by parallel 6010 // threads. Since the marking bit map is shared, updates are 6011 // synchronized (via CAS). 6012 void Par_MarkRefsIntoAndScanClosure::do_oop(oop obj) { 6013 if (obj != NULL) { 6014 // Ignore mark word because this could be an already marked oop 6015 // that may be chained at the end of the overflow list. 6016 assert(obj->is_oop(true), "expected an oop"); 6017 HeapWord* addr = (HeapWord*)obj; 6018 if (_span.contains(addr) && 6019 !_bit_map->isMarked(addr)) { 6020 // mark bit map (object will become grey): 6021 // It is possible for several threads to be 6022 // trying to "claim" this object concurrently; 6023 // the unique thread that succeeds in marking the 6024 // object first will do the subsequent push on 6025 // to the work queue (or overflow list). 6026 if (_bit_map->par_mark(addr)) { 6027 // push on work_queue (which may not be empty), and trim the 6028 // queue to an appropriate length by applying this closure to 6029 // the oops in the oops popped from the stack (i.e. blacken the 6030 // grey objects) 6031 bool res = _work_queue->push(obj); 6032 assert(res, "Low water mark should be less than capacity?"); 6033 trim_queue(_low_water_mark); 6034 } // Else, another thread claimed the object 6035 } 6036 } 6037 } 6038 6039 void Par_MarkRefsIntoAndScanClosure::do_oop(oop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); } 6040 void Par_MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); } 6041 6042 // This closure is used to rescan the marked objects on the dirty cards 6043 // in the mod union table and the card table proper. 6044 size_t ScanMarkedObjectsAgainCarefullyClosure::do_object_careful_m( 6045 oop p, MemRegion mr) { 6046 6047 size_t size = 0; 6048 HeapWord* addr = (HeapWord*)p; 6049 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6050 assert(_span.contains(addr), "we are scanning the CMS generation"); 6051 // check if it's time to yield 6052 if (do_yield_check()) { 6053 // We yielded for some foreground stop-world work, 6054 // and we have been asked to abort this ongoing preclean cycle. 6055 return 0; 6056 } 6057 if (_bitMap->isMarked(addr)) { 6058 // it's marked; is it potentially uninitialized? 6059 if (p->klass_or_null() != NULL) { 6060 // an initialized object; ignore mark word in verification below 6061 // since we are running concurrent with mutators 6062 assert(p->is_oop(true), "should be an oop"); 6063 if (p->is_objArray()) { 6064 // objArrays are precisely marked; restrict scanning 6065 // to dirty cards only. 6066 size = CompactibleFreeListSpace::adjustObjectSize( 6067 p->oop_iterate_size(_scanningClosure, mr)); 6068 } else { 6069 // A non-array may have been imprecisely marked; we need 6070 // to scan object in its entirety. 6071 size = CompactibleFreeListSpace::adjustObjectSize( 6072 p->oop_iterate_size(_scanningClosure)); 6073 } 6074 #ifdef ASSERT 6075 size_t direct_size = 6076 CompactibleFreeListSpace::adjustObjectSize(p->size()); 6077 assert(size == direct_size, "Inconsistency in size"); 6078 assert(size >= 3, "Necessary for Printezis marks to work"); 6079 if (!_bitMap->isMarked(addr+1)) { 6080 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size); 6081 } else { 6082 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size-1); 6083 assert(_bitMap->isMarked(addr+size-1), 6084 "inconsistent Printezis mark"); 6085 } 6086 #endif // ASSERT 6087 } else { 6088 // An uninitialized object. 6089 assert(_bitMap->isMarked(addr+1), "missing Printezis mark?"); 6090 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 6091 size = pointer_delta(nextOneAddr + 1, addr); 6092 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 6093 "alignment problem"); 6094 // Note that pre-cleaning needn't redirty the card. OopDesc::set_klass() 6095 // will dirty the card when the klass pointer is installed in the 6096 // object (signaling the completion of initialization). 6097 } 6098 } else { 6099 // Either a not yet marked object or an uninitialized object 6100 if (p->klass_or_null() == NULL) { 6101 // An uninitialized object, skip to the next card, since 6102 // we may not be able to read its P-bits yet. 6103 assert(size == 0, "Initial value"); 6104 } else { 6105 // An object not (yet) reached by marking: we merely need to 6106 // compute its size so as to go look at the next block. 6107 assert(p->is_oop(true), "should be an oop"); 6108 size = CompactibleFreeListSpace::adjustObjectSize(p->size()); 6109 } 6110 } 6111 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6112 return size; 6113 } 6114 6115 void ScanMarkedObjectsAgainCarefullyClosure::do_yield_work() { 6116 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6117 "CMS thread should hold CMS token"); 6118 assert_lock_strong(_freelistLock); 6119 assert_lock_strong(_bitMap->lock()); 6120 // relinquish the free_list_lock and bitMaplock() 6121 _bitMap->lock()->unlock(); 6122 _freelistLock->unlock(); 6123 ConcurrentMarkSweepThread::desynchronize(true); 6124 _collector->stopTimer(); 6125 _collector->incrementYields(); 6126 6127 // See the comment in coordinator_yield() 6128 for (unsigned i = 0; i < CMSYieldSleepCount && 6129 ConcurrentMarkSweepThread::should_yield() && 6130 !CMSCollector::foregroundGCIsActive(); ++i) { 6131 os::sleep(Thread::current(), 1, false); 6132 } 6133 6134 ConcurrentMarkSweepThread::synchronize(true); 6135 _freelistLock->lock_without_safepoint_check(); 6136 _bitMap->lock()->lock_without_safepoint_check(); 6137 _collector->startTimer(); 6138 } 6139 6140 6141 ////////////////////////////////////////////////////////////////// 6142 // SurvivorSpacePrecleanClosure 6143 ////////////////////////////////////////////////////////////////// 6144 // This (single-threaded) closure is used to preclean the oops in 6145 // the survivor spaces. 6146 size_t SurvivorSpacePrecleanClosure::do_object_careful(oop p) { 6147 6148 HeapWord* addr = (HeapWord*)p; 6149 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6150 assert(!_span.contains(addr), "we are scanning the survivor spaces"); 6151 assert(p->klass_or_null() != NULL, "object should be initialized"); 6152 // an initialized object; ignore mark word in verification below 6153 // since we are running concurrent with mutators 6154 assert(p->is_oop(true), "should be an oop"); 6155 // Note that we do not yield while we iterate over 6156 // the interior oops of p, pushing the relevant ones 6157 // on our marking stack. 6158 size_t size = p->oop_iterate_size(_scanning_closure); 6159 do_yield_check(); 6160 // Observe that below, we do not abandon the preclean 6161 // phase as soon as we should; rather we empty the 6162 // marking stack before returning. This is to satisfy 6163 // some existing assertions. In general, it may be a 6164 // good idea to abort immediately and complete the marking 6165 // from the grey objects at a later time. 6166 while (!_mark_stack->isEmpty()) { 6167 oop new_oop = _mark_stack->pop(); 6168 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 6169 assert(_bit_map->isMarked((HeapWord*)new_oop), 6170 "only grey objects on this stack"); 6171 // iterate over the oops in this oop, marking and pushing 6172 // the ones in CMS heap (i.e. in _span). 6173 new_oop->oop_iterate(_scanning_closure); 6174 // check if it's time to yield 6175 do_yield_check(); 6176 } 6177 unsigned int after_count = 6178 GenCollectedHeap::heap()->total_collections(); 6179 bool abort = (_before_count != after_count) || 6180 _collector->should_abort_preclean(); 6181 return abort ? 0 : size; 6182 } 6183 6184 void SurvivorSpacePrecleanClosure::do_yield_work() { 6185 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6186 "CMS thread should hold CMS token"); 6187 assert_lock_strong(_bit_map->lock()); 6188 // Relinquish the bit map lock 6189 _bit_map->lock()->unlock(); 6190 ConcurrentMarkSweepThread::desynchronize(true); 6191 _collector->stopTimer(); 6192 _collector->incrementYields(); 6193 6194 // See the comment in coordinator_yield() 6195 for (unsigned i = 0; i < CMSYieldSleepCount && 6196 ConcurrentMarkSweepThread::should_yield() && 6197 !CMSCollector::foregroundGCIsActive(); ++i) { 6198 os::sleep(Thread::current(), 1, false); 6199 } 6200 6201 ConcurrentMarkSweepThread::synchronize(true); 6202 _bit_map->lock()->lock_without_safepoint_check(); 6203 _collector->startTimer(); 6204 } 6205 6206 // This closure is used to rescan the marked objects on the dirty cards 6207 // in the mod union table and the card table proper. In the parallel 6208 // case, although the bitMap is shared, we do a single read so the 6209 // isMarked() query is "safe". 6210 bool ScanMarkedObjectsAgainClosure::do_object_bm(oop p, MemRegion mr) { 6211 // Ignore mark word because we are running concurrent with mutators 6212 assert(p->is_oop_or_null(true), "Expected an oop or NULL at " PTR_FORMAT, p2i(p)); 6213 HeapWord* addr = (HeapWord*)p; 6214 assert(_span.contains(addr), "we are scanning the CMS generation"); 6215 bool is_obj_array = false; 6216 #ifdef ASSERT 6217 if (!_parallel) { 6218 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 6219 assert(_collector->overflow_list_is_empty(), 6220 "overflow list should be empty"); 6221 6222 } 6223 #endif // ASSERT 6224 if (_bit_map->isMarked(addr)) { 6225 // Obj arrays are precisely marked, non-arrays are not; 6226 // so we scan objArrays precisely and non-arrays in their 6227 // entirety. 6228 if (p->is_objArray()) { 6229 is_obj_array = true; 6230 if (_parallel) { 6231 p->oop_iterate(_par_scan_closure, mr); 6232 } else { 6233 p->oop_iterate(_scan_closure, mr); 6234 } 6235 } else { 6236 if (_parallel) { 6237 p->oop_iterate(_par_scan_closure); 6238 } else { 6239 p->oop_iterate(_scan_closure); 6240 } 6241 } 6242 } 6243 #ifdef ASSERT 6244 if (!_parallel) { 6245 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 6246 assert(_collector->overflow_list_is_empty(), 6247 "overflow list should be empty"); 6248 6249 } 6250 #endif // ASSERT 6251 return is_obj_array; 6252 } 6253 6254 MarkFromRootsClosure::MarkFromRootsClosure(CMSCollector* collector, 6255 MemRegion span, 6256 CMSBitMap* bitMap, CMSMarkStack* markStack, 6257 bool should_yield, bool verifying): 6258 _collector(collector), 6259 _span(span), 6260 _bitMap(bitMap), 6261 _mut(&collector->_modUnionTable), 6262 _markStack(markStack), 6263 _yield(should_yield), 6264 _skipBits(0) 6265 { 6266 assert(_markStack->isEmpty(), "stack should be empty"); 6267 _finger = _bitMap->startWord(); 6268 _threshold = _finger; 6269 assert(_collector->_restart_addr == NULL, "Sanity check"); 6270 assert(_span.contains(_finger), "Out of bounds _finger?"); 6271 DEBUG_ONLY(_verifying = verifying;) 6272 } 6273 6274 void MarkFromRootsClosure::reset(HeapWord* addr) { 6275 assert(_markStack->isEmpty(), "would cause duplicates on stack"); 6276 assert(_span.contains(addr), "Out of bounds _finger?"); 6277 _finger = addr; 6278 _threshold = (HeapWord*)round_to( 6279 (intptr_t)_finger, CardTableModRefBS::card_size); 6280 } 6281 6282 // Should revisit to see if this should be restructured for 6283 // greater efficiency. 6284 bool MarkFromRootsClosure::do_bit(size_t offset) { 6285 if (_skipBits > 0) { 6286 _skipBits--; 6287 return true; 6288 } 6289 // convert offset into a HeapWord* 6290 HeapWord* addr = _bitMap->startWord() + offset; 6291 assert(_bitMap->endWord() && addr < _bitMap->endWord(), 6292 "address out of range"); 6293 assert(_bitMap->isMarked(addr), "tautology"); 6294 if (_bitMap->isMarked(addr+1)) { 6295 // this is an allocated but not yet initialized object 6296 assert(_skipBits == 0, "tautology"); 6297 _skipBits = 2; // skip next two marked bits ("Printezis-marks") 6298 oop p = oop(addr); 6299 if (p->klass_or_null() == NULL) { 6300 DEBUG_ONLY(if (!_verifying) {) 6301 // We re-dirty the cards on which this object lies and increase 6302 // the _threshold so that we'll come back to scan this object 6303 // during the preclean or remark phase. (CMSCleanOnEnter) 6304 if (CMSCleanOnEnter) { 6305 size_t sz = _collector->block_size_using_printezis_bits(addr); 6306 HeapWord* end_card_addr = (HeapWord*)round_to( 6307 (intptr_t)(addr+sz), CardTableModRefBS::card_size); 6308 MemRegion redirty_range = MemRegion(addr, end_card_addr); 6309 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 6310 // Bump _threshold to end_card_addr; note that 6311 // _threshold cannot possibly exceed end_card_addr, anyhow. 6312 // This prevents future clearing of the card as the scan proceeds 6313 // to the right. 6314 assert(_threshold <= end_card_addr, 6315 "Because we are just scanning into this object"); 6316 if (_threshold < end_card_addr) { 6317 _threshold = end_card_addr; 6318 } 6319 if (p->klass_or_null() != NULL) { 6320 // Redirty the range of cards... 6321 _mut->mark_range(redirty_range); 6322 } // ...else the setting of klass will dirty the card anyway. 6323 } 6324 DEBUG_ONLY(}) 6325 return true; 6326 } 6327 } 6328 scanOopsInOop(addr); 6329 return true; 6330 } 6331 6332 // We take a break if we've been at this for a while, 6333 // so as to avoid monopolizing the locks involved. 6334 void MarkFromRootsClosure::do_yield_work() { 6335 // First give up the locks, then yield, then re-lock 6336 // We should probably use a constructor/destructor idiom to 6337 // do this unlock/lock or modify the MutexUnlocker class to 6338 // serve our purpose. XXX 6339 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6340 "CMS thread should hold CMS token"); 6341 assert_lock_strong(_bitMap->lock()); 6342 _bitMap->lock()->unlock(); 6343 ConcurrentMarkSweepThread::desynchronize(true); 6344 _collector->stopTimer(); 6345 _collector->incrementYields(); 6346 6347 // See the comment in coordinator_yield() 6348 for (unsigned i = 0; i < CMSYieldSleepCount && 6349 ConcurrentMarkSweepThread::should_yield() && 6350 !CMSCollector::foregroundGCIsActive(); ++i) { 6351 os::sleep(Thread::current(), 1, false); 6352 } 6353 6354 ConcurrentMarkSweepThread::synchronize(true); 6355 _bitMap->lock()->lock_without_safepoint_check(); 6356 _collector->startTimer(); 6357 } 6358 6359 void MarkFromRootsClosure::scanOopsInOop(HeapWord* ptr) { 6360 assert(_bitMap->isMarked(ptr), "expected bit to be set"); 6361 assert(_markStack->isEmpty(), 6362 "should drain stack to limit stack usage"); 6363 // convert ptr to an oop preparatory to scanning 6364 oop obj = oop(ptr); 6365 // Ignore mark word in verification below, since we 6366 // may be running concurrent with mutators. 6367 assert(obj->is_oop(true), "should be an oop"); 6368 assert(_finger <= ptr, "_finger runneth ahead"); 6369 // advance the finger to right end of this object 6370 _finger = ptr + obj->size(); 6371 assert(_finger > ptr, "we just incremented it above"); 6372 // On large heaps, it may take us some time to get through 6373 // the marking phase. During 6374 // this time it's possible that a lot of mutations have 6375 // accumulated in the card table and the mod union table -- 6376 // these mutation records are redundant until we have 6377 // actually traced into the corresponding card. 6378 // Here, we check whether advancing the finger would make 6379 // us cross into a new card, and if so clear corresponding 6380 // cards in the MUT (preclean them in the card-table in the 6381 // future). 6382 6383 DEBUG_ONLY(if (!_verifying) {) 6384 // The clean-on-enter optimization is disabled by default, 6385 // until we fix 6178663. 6386 if (CMSCleanOnEnter && (_finger > _threshold)) { 6387 // [_threshold, _finger) represents the interval 6388 // of cards to be cleared in MUT (or precleaned in card table). 6389 // The set of cards to be cleared is all those that overlap 6390 // with the interval [_threshold, _finger); note that 6391 // _threshold is always kept card-aligned but _finger isn't 6392 // always card-aligned. 6393 HeapWord* old_threshold = _threshold; 6394 assert(old_threshold == (HeapWord*)round_to( 6395 (intptr_t)old_threshold, CardTableModRefBS::card_size), 6396 "_threshold should always be card-aligned"); 6397 _threshold = (HeapWord*)round_to( 6398 (intptr_t)_finger, CardTableModRefBS::card_size); 6399 MemRegion mr(old_threshold, _threshold); 6400 assert(!mr.is_empty(), "Control point invariant"); 6401 assert(_span.contains(mr), "Should clear within span"); 6402 _mut->clear_range(mr); 6403 } 6404 DEBUG_ONLY(}) 6405 // Note: the finger doesn't advance while we drain 6406 // the stack below. 6407 PushOrMarkClosure pushOrMarkClosure(_collector, 6408 _span, _bitMap, _markStack, 6409 _finger, this); 6410 bool res = _markStack->push(obj); 6411 assert(res, "Empty non-zero size stack should have space for single push"); 6412 while (!_markStack->isEmpty()) { 6413 oop new_oop = _markStack->pop(); 6414 // Skip verifying header mark word below because we are 6415 // running concurrent with mutators. 6416 assert(new_oop->is_oop(true), "Oops! expected to pop an oop"); 6417 // now scan this oop's oops 6418 new_oop->oop_iterate(&pushOrMarkClosure); 6419 do_yield_check(); 6420 } 6421 assert(_markStack->isEmpty(), "tautology, emphasizing post-condition"); 6422 } 6423 6424 Par_MarkFromRootsClosure::Par_MarkFromRootsClosure(CMSConcMarkingTask* task, 6425 CMSCollector* collector, MemRegion span, 6426 CMSBitMap* bit_map, 6427 OopTaskQueue* work_queue, 6428 CMSMarkStack* overflow_stack): 6429 _collector(collector), 6430 _whole_span(collector->_span), 6431 _span(span), 6432 _bit_map(bit_map), 6433 _mut(&collector->_modUnionTable), 6434 _work_queue(work_queue), 6435 _overflow_stack(overflow_stack), 6436 _skip_bits(0), 6437 _task(task) 6438 { 6439 assert(_work_queue->size() == 0, "work_queue should be empty"); 6440 _finger = span.start(); 6441 _threshold = _finger; // XXX Defer clear-on-enter optimization for now 6442 assert(_span.contains(_finger), "Out of bounds _finger?"); 6443 } 6444 6445 // Should revisit to see if this should be restructured for 6446 // greater efficiency. 6447 bool Par_MarkFromRootsClosure::do_bit(size_t offset) { 6448 if (_skip_bits > 0) { 6449 _skip_bits--; 6450 return true; 6451 } 6452 // convert offset into a HeapWord* 6453 HeapWord* addr = _bit_map->startWord() + offset; 6454 assert(_bit_map->endWord() && addr < _bit_map->endWord(), 6455 "address out of range"); 6456 assert(_bit_map->isMarked(addr), "tautology"); 6457 if (_bit_map->isMarked(addr+1)) { 6458 // this is an allocated object that might not yet be initialized 6459 assert(_skip_bits == 0, "tautology"); 6460 _skip_bits = 2; // skip next two marked bits ("Printezis-marks") 6461 oop p = oop(addr); 6462 if (p->klass_or_null() == NULL) { 6463 // in the case of Clean-on-Enter optimization, redirty card 6464 // and avoid clearing card by increasing the threshold. 6465 return true; 6466 } 6467 } 6468 scan_oops_in_oop(addr); 6469 return true; 6470 } 6471 6472 void Par_MarkFromRootsClosure::scan_oops_in_oop(HeapWord* ptr) { 6473 assert(_bit_map->isMarked(ptr), "expected bit to be set"); 6474 // Should we assert that our work queue is empty or 6475 // below some drain limit? 6476 assert(_work_queue->size() == 0, 6477 "should drain stack to limit stack usage"); 6478 // convert ptr to an oop preparatory to scanning 6479 oop obj = oop(ptr); 6480 // Ignore mark word in verification below, since we 6481 // may be running concurrent with mutators. 6482 assert(obj->is_oop(true), "should be an oop"); 6483 assert(_finger <= ptr, "_finger runneth ahead"); 6484 // advance the finger to right end of this object 6485 _finger = ptr + obj->size(); 6486 assert(_finger > ptr, "we just incremented it above"); 6487 // On large heaps, it may take us some time to get through 6488 // the marking phase. During 6489 // this time it's possible that a lot of mutations have 6490 // accumulated in the card table and the mod union table -- 6491 // these mutation records are redundant until we have 6492 // actually traced into the corresponding card. 6493 // Here, we check whether advancing the finger would make 6494 // us cross into a new card, and if so clear corresponding 6495 // cards in the MUT (preclean them in the card-table in the 6496 // future). 6497 6498 // The clean-on-enter optimization is disabled by default, 6499 // until we fix 6178663. 6500 if (CMSCleanOnEnter && (_finger > _threshold)) { 6501 // [_threshold, _finger) represents the interval 6502 // of cards to be cleared in MUT (or precleaned in card table). 6503 // The set of cards to be cleared is all those that overlap 6504 // with the interval [_threshold, _finger); note that 6505 // _threshold is always kept card-aligned but _finger isn't 6506 // always card-aligned. 6507 HeapWord* old_threshold = _threshold; 6508 assert(old_threshold == (HeapWord*)round_to( 6509 (intptr_t)old_threshold, CardTableModRefBS::card_size), 6510 "_threshold should always be card-aligned"); 6511 _threshold = (HeapWord*)round_to( 6512 (intptr_t)_finger, CardTableModRefBS::card_size); 6513 MemRegion mr(old_threshold, _threshold); 6514 assert(!mr.is_empty(), "Control point invariant"); 6515 assert(_span.contains(mr), "Should clear within span"); // _whole_span ?? 6516 _mut->clear_range(mr); 6517 } 6518 6519 // Note: the local finger doesn't advance while we drain 6520 // the stack below, but the global finger sure can and will. 6521 HeapWord** gfa = _task->global_finger_addr(); 6522 Par_PushOrMarkClosure pushOrMarkClosure(_collector, 6523 _span, _bit_map, 6524 _work_queue, 6525 _overflow_stack, 6526 _finger, 6527 gfa, this); 6528 bool res = _work_queue->push(obj); // overflow could occur here 6529 assert(res, "Will hold once we use workqueues"); 6530 while (true) { 6531 oop new_oop; 6532 if (!_work_queue->pop_local(new_oop)) { 6533 // We emptied our work_queue; check if there's stuff that can 6534 // be gotten from the overflow stack. 6535 if (CMSConcMarkingTask::get_work_from_overflow_stack( 6536 _overflow_stack, _work_queue)) { 6537 do_yield_check(); 6538 continue; 6539 } else { // done 6540 break; 6541 } 6542 } 6543 // Skip verifying header mark word below because we are 6544 // running concurrent with mutators. 6545 assert(new_oop->is_oop(true), "Oops! expected to pop an oop"); 6546 // now scan this oop's oops 6547 new_oop->oop_iterate(&pushOrMarkClosure); 6548 do_yield_check(); 6549 } 6550 assert(_work_queue->size() == 0, "tautology, emphasizing post-condition"); 6551 } 6552 6553 // Yield in response to a request from VM Thread or 6554 // from mutators. 6555 void Par_MarkFromRootsClosure::do_yield_work() { 6556 assert(_task != NULL, "sanity"); 6557 _task->yield(); 6558 } 6559 6560 // A variant of the above used for verifying CMS marking work. 6561 MarkFromRootsVerifyClosure::MarkFromRootsVerifyClosure(CMSCollector* collector, 6562 MemRegion span, 6563 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 6564 CMSMarkStack* mark_stack): 6565 _collector(collector), 6566 _span(span), 6567 _verification_bm(verification_bm), 6568 _cms_bm(cms_bm), 6569 _mark_stack(mark_stack), 6570 _pam_verify_closure(collector, span, verification_bm, cms_bm, 6571 mark_stack) 6572 { 6573 assert(_mark_stack->isEmpty(), "stack should be empty"); 6574 _finger = _verification_bm->startWord(); 6575 assert(_collector->_restart_addr == NULL, "Sanity check"); 6576 assert(_span.contains(_finger), "Out of bounds _finger?"); 6577 } 6578 6579 void MarkFromRootsVerifyClosure::reset(HeapWord* addr) { 6580 assert(_mark_stack->isEmpty(), "would cause duplicates on stack"); 6581 assert(_span.contains(addr), "Out of bounds _finger?"); 6582 _finger = addr; 6583 } 6584 6585 // Should revisit to see if this should be restructured for 6586 // greater efficiency. 6587 bool MarkFromRootsVerifyClosure::do_bit(size_t offset) { 6588 // convert offset into a HeapWord* 6589 HeapWord* addr = _verification_bm->startWord() + offset; 6590 assert(_verification_bm->endWord() && addr < _verification_bm->endWord(), 6591 "address out of range"); 6592 assert(_verification_bm->isMarked(addr), "tautology"); 6593 assert(_cms_bm->isMarked(addr), "tautology"); 6594 6595 assert(_mark_stack->isEmpty(), 6596 "should drain stack to limit stack usage"); 6597 // convert addr to an oop preparatory to scanning 6598 oop obj = oop(addr); 6599 assert(obj->is_oop(), "should be an oop"); 6600 assert(_finger <= addr, "_finger runneth ahead"); 6601 // advance the finger to right end of this object 6602 _finger = addr + obj->size(); 6603 assert(_finger > addr, "we just incremented it above"); 6604 // Note: the finger doesn't advance while we drain 6605 // the stack below. 6606 bool res = _mark_stack->push(obj); 6607 assert(res, "Empty non-zero size stack should have space for single push"); 6608 while (!_mark_stack->isEmpty()) { 6609 oop new_oop = _mark_stack->pop(); 6610 assert(new_oop->is_oop(), "Oops! expected to pop an oop"); 6611 // now scan this oop's oops 6612 new_oop->oop_iterate(&_pam_verify_closure); 6613 } 6614 assert(_mark_stack->isEmpty(), "tautology, emphasizing post-condition"); 6615 return true; 6616 } 6617 6618 PushAndMarkVerifyClosure::PushAndMarkVerifyClosure( 6619 CMSCollector* collector, MemRegion span, 6620 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 6621 CMSMarkStack* mark_stack): 6622 MetadataAwareOopClosure(collector->ref_processor()), 6623 _collector(collector), 6624 _span(span), 6625 _verification_bm(verification_bm), 6626 _cms_bm(cms_bm), 6627 _mark_stack(mark_stack) 6628 { } 6629 6630 void PushAndMarkVerifyClosure::do_oop(oop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 6631 void PushAndMarkVerifyClosure::do_oop(narrowOop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 6632 6633 // Upon stack overflow, we discard (part of) the stack, 6634 // remembering the least address amongst those discarded 6635 // in CMSCollector's _restart_address. 6636 void PushAndMarkVerifyClosure::handle_stack_overflow(HeapWord* lost) { 6637 // Remember the least grey address discarded 6638 HeapWord* ra = (HeapWord*)_mark_stack->least_value(lost); 6639 _collector->lower_restart_addr(ra); 6640 _mark_stack->reset(); // discard stack contents 6641 _mark_stack->expand(); // expand the stack if possible 6642 } 6643 6644 void PushAndMarkVerifyClosure::do_oop(oop obj) { 6645 assert(obj->is_oop_or_null(), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6646 HeapWord* addr = (HeapWord*)obj; 6647 if (_span.contains(addr) && !_verification_bm->isMarked(addr)) { 6648 // Oop lies in _span and isn't yet grey or black 6649 _verification_bm->mark(addr); // now grey 6650 if (!_cms_bm->isMarked(addr)) { 6651 LogHandle(gc, verify) log; 6652 ResourceMark rm; 6653 oop(addr)->print_on(log.info_stream()); 6654 log.info(" (" INTPTR_FORMAT " should have been marked)", p2i(addr)); 6655 fatal("... aborting"); 6656 } 6657 6658 if (!_mark_stack->push(obj)) { // stack overflow 6659 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _mark_stack->capacity()); 6660 assert(_mark_stack->isFull(), "Else push should have succeeded"); 6661 handle_stack_overflow(addr); 6662 } 6663 // anything including and to the right of _finger 6664 // will be scanned as we iterate over the remainder of the 6665 // bit map 6666 } 6667 } 6668 6669 PushOrMarkClosure::PushOrMarkClosure(CMSCollector* collector, 6670 MemRegion span, 6671 CMSBitMap* bitMap, CMSMarkStack* markStack, 6672 HeapWord* finger, MarkFromRootsClosure* parent) : 6673 MetadataAwareOopClosure(collector->ref_processor()), 6674 _collector(collector), 6675 _span(span), 6676 _bitMap(bitMap), 6677 _markStack(markStack), 6678 _finger(finger), 6679 _parent(parent) 6680 { } 6681 6682 Par_PushOrMarkClosure::Par_PushOrMarkClosure(CMSCollector* collector, 6683 MemRegion span, 6684 CMSBitMap* bit_map, 6685 OopTaskQueue* work_queue, 6686 CMSMarkStack* overflow_stack, 6687 HeapWord* finger, 6688 HeapWord** global_finger_addr, 6689 Par_MarkFromRootsClosure* parent) : 6690 MetadataAwareOopClosure(collector->ref_processor()), 6691 _collector(collector), 6692 _whole_span(collector->_span), 6693 _span(span), 6694 _bit_map(bit_map), 6695 _work_queue(work_queue), 6696 _overflow_stack(overflow_stack), 6697 _finger(finger), 6698 _global_finger_addr(global_finger_addr), 6699 _parent(parent) 6700 { } 6701 6702 // Assumes thread-safe access by callers, who are 6703 // responsible for mutual exclusion. 6704 void CMSCollector::lower_restart_addr(HeapWord* low) { 6705 assert(_span.contains(low), "Out of bounds addr"); 6706 if (_restart_addr == NULL) { 6707 _restart_addr = low; 6708 } else { 6709 _restart_addr = MIN2(_restart_addr, low); 6710 } 6711 } 6712 6713 // Upon stack overflow, we discard (part of) the stack, 6714 // remembering the least address amongst those discarded 6715 // in CMSCollector's _restart_address. 6716 void PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 6717 // Remember the least grey address discarded 6718 HeapWord* ra = (HeapWord*)_markStack->least_value(lost); 6719 _collector->lower_restart_addr(ra); 6720 _markStack->reset(); // discard stack contents 6721 _markStack->expand(); // expand the stack if possible 6722 } 6723 6724 // Upon stack overflow, we discard (part of) the stack, 6725 // remembering the least address amongst those discarded 6726 // in CMSCollector's _restart_address. 6727 void Par_PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 6728 // We need to do this under a mutex to prevent other 6729 // workers from interfering with the work done below. 6730 MutexLockerEx ml(_overflow_stack->par_lock(), 6731 Mutex::_no_safepoint_check_flag); 6732 // Remember the least grey address discarded 6733 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 6734 _collector->lower_restart_addr(ra); 6735 _overflow_stack->reset(); // discard stack contents 6736 _overflow_stack->expand(); // expand the stack if possible 6737 } 6738 6739 void PushOrMarkClosure::do_oop(oop obj) { 6740 // Ignore mark word because we are running concurrent with mutators. 6741 assert(obj->is_oop_or_null(true), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6742 HeapWord* addr = (HeapWord*)obj; 6743 if (_span.contains(addr) && !_bitMap->isMarked(addr)) { 6744 // Oop lies in _span and isn't yet grey or black 6745 _bitMap->mark(addr); // now grey 6746 if (addr < _finger) { 6747 // the bit map iteration has already either passed, or 6748 // sampled, this bit in the bit map; we'll need to 6749 // use the marking stack to scan this oop's oops. 6750 bool simulate_overflow = false; 6751 NOT_PRODUCT( 6752 if (CMSMarkStackOverflowALot && 6753 _collector->simulate_overflow()) { 6754 // simulate a stack overflow 6755 simulate_overflow = true; 6756 } 6757 ) 6758 if (simulate_overflow || !_markStack->push(obj)) { // stack overflow 6759 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _markStack->capacity()); 6760 assert(simulate_overflow || _markStack->isFull(), "Else push should have succeeded"); 6761 handle_stack_overflow(addr); 6762 } 6763 } 6764 // anything including and to the right of _finger 6765 // will be scanned as we iterate over the remainder of the 6766 // bit map 6767 do_yield_check(); 6768 } 6769 } 6770 6771 void PushOrMarkClosure::do_oop(oop* p) { PushOrMarkClosure::do_oop_work(p); } 6772 void PushOrMarkClosure::do_oop(narrowOop* p) { PushOrMarkClosure::do_oop_work(p); } 6773 6774 void Par_PushOrMarkClosure::do_oop(oop obj) { 6775 // Ignore mark word because we are running concurrent with mutators. 6776 assert(obj->is_oop_or_null(true), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6777 HeapWord* addr = (HeapWord*)obj; 6778 if (_whole_span.contains(addr) && !_bit_map->isMarked(addr)) { 6779 // Oop lies in _span and isn't yet grey or black 6780 // We read the global_finger (volatile read) strictly after marking oop 6781 bool res = _bit_map->par_mark(addr); // now grey 6782 volatile HeapWord** gfa = (volatile HeapWord**)_global_finger_addr; 6783 // Should we push this marked oop on our stack? 6784 // -- if someone else marked it, nothing to do 6785 // -- if target oop is above global finger nothing to do 6786 // -- if target oop is in chunk and above local finger 6787 // then nothing to do 6788 // -- else push on work queue 6789 if ( !res // someone else marked it, they will deal with it 6790 || (addr >= *gfa) // will be scanned in a later task 6791 || (_span.contains(addr) && addr >= _finger)) { // later in this chunk 6792 return; 6793 } 6794 // the bit map iteration has already either passed, or 6795 // sampled, this bit in the bit map; we'll need to 6796 // use the marking stack to scan this oop's oops. 6797 bool simulate_overflow = false; 6798 NOT_PRODUCT( 6799 if (CMSMarkStackOverflowALot && 6800 _collector->simulate_overflow()) { 6801 // simulate a stack overflow 6802 simulate_overflow = true; 6803 } 6804 ) 6805 if (simulate_overflow || 6806 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) { 6807 // stack overflow 6808 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _overflow_stack->capacity()); 6809 // We cannot assert that the overflow stack is full because 6810 // it may have been emptied since. 6811 assert(simulate_overflow || 6812 _work_queue->size() == _work_queue->max_elems(), 6813 "Else push should have succeeded"); 6814 handle_stack_overflow(addr); 6815 } 6816 do_yield_check(); 6817 } 6818 } 6819 6820 void Par_PushOrMarkClosure::do_oop(oop* p) { Par_PushOrMarkClosure::do_oop_work(p); } 6821 void Par_PushOrMarkClosure::do_oop(narrowOop* p) { Par_PushOrMarkClosure::do_oop_work(p); } 6822 6823 PushAndMarkClosure::PushAndMarkClosure(CMSCollector* collector, 6824 MemRegion span, 6825 ReferenceProcessor* rp, 6826 CMSBitMap* bit_map, 6827 CMSBitMap* mod_union_table, 6828 CMSMarkStack* mark_stack, 6829 bool concurrent_precleaning): 6830 MetadataAwareOopClosure(rp), 6831 _collector(collector), 6832 _span(span), 6833 _bit_map(bit_map), 6834 _mod_union_table(mod_union_table), 6835 _mark_stack(mark_stack), 6836 _concurrent_precleaning(concurrent_precleaning) 6837 { 6838 assert(ref_processor() != NULL, "ref_processor shouldn't be NULL"); 6839 } 6840 6841 // Grey object rescan during pre-cleaning and second checkpoint phases -- 6842 // the non-parallel version (the parallel version appears further below.) 6843 void PushAndMarkClosure::do_oop(oop obj) { 6844 // Ignore mark word verification. If during concurrent precleaning, 6845 // the object monitor may be locked. If during the checkpoint 6846 // phases, the object may already have been reached by a different 6847 // path and may be at the end of the global overflow list (so 6848 // the mark word may be NULL). 6849 assert(obj->is_oop_or_null(true /* ignore mark word */), 6850 "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6851 HeapWord* addr = (HeapWord*)obj; 6852 // Check if oop points into the CMS generation 6853 // and is not marked 6854 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 6855 // a white object ... 6856 _bit_map->mark(addr); // ... now grey 6857 // push on the marking stack (grey set) 6858 bool simulate_overflow = false; 6859 NOT_PRODUCT( 6860 if (CMSMarkStackOverflowALot && 6861 _collector->simulate_overflow()) { 6862 // simulate a stack overflow 6863 simulate_overflow = true; 6864 } 6865 ) 6866 if (simulate_overflow || !_mark_stack->push(obj)) { 6867 if (_concurrent_precleaning) { 6868 // During precleaning we can just dirty the appropriate card(s) 6869 // in the mod union table, thus ensuring that the object remains 6870 // in the grey set and continue. In the case of object arrays 6871 // we need to dirty all of the cards that the object spans, 6872 // since the rescan of object arrays will be limited to the 6873 // dirty cards. 6874 // Note that no one can be interfering with us in this action 6875 // of dirtying the mod union table, so no locking or atomics 6876 // are required. 6877 if (obj->is_objArray()) { 6878 size_t sz = obj->size(); 6879 HeapWord* end_card_addr = (HeapWord*)round_to( 6880 (intptr_t)(addr+sz), CardTableModRefBS::card_size); 6881 MemRegion redirty_range = MemRegion(addr, end_card_addr); 6882 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 6883 _mod_union_table->mark_range(redirty_range); 6884 } else { 6885 _mod_union_table->mark(addr); 6886 } 6887 _collector->_ser_pmc_preclean_ovflw++; 6888 } else { 6889 // During the remark phase, we need to remember this oop 6890 // in the overflow list. 6891 _collector->push_on_overflow_list(obj); 6892 _collector->_ser_pmc_remark_ovflw++; 6893 } 6894 } 6895 } 6896 } 6897 6898 Par_PushAndMarkClosure::Par_PushAndMarkClosure(CMSCollector* collector, 6899 MemRegion span, 6900 ReferenceProcessor* rp, 6901 CMSBitMap* bit_map, 6902 OopTaskQueue* work_queue): 6903 MetadataAwareOopClosure(rp), 6904 _collector(collector), 6905 _span(span), 6906 _bit_map(bit_map), 6907 _work_queue(work_queue) 6908 { 6909 assert(ref_processor() != NULL, "ref_processor shouldn't be NULL"); 6910 } 6911 6912 void PushAndMarkClosure::do_oop(oop* p) { PushAndMarkClosure::do_oop_work(p); } 6913 void PushAndMarkClosure::do_oop(narrowOop* p) { PushAndMarkClosure::do_oop_work(p); } 6914 6915 // Grey object rescan during second checkpoint phase -- 6916 // the parallel version. 6917 void Par_PushAndMarkClosure::do_oop(oop obj) { 6918 // In the assert below, we ignore the mark word because 6919 // this oop may point to an already visited object that is 6920 // on the overflow stack (in which case the mark word has 6921 // been hijacked for chaining into the overflow stack -- 6922 // if this is the last object in the overflow stack then 6923 // its mark word will be NULL). Because this object may 6924 // have been subsequently popped off the global overflow 6925 // stack, and the mark word possibly restored to the prototypical 6926 // value, by the time we get to examined this failing assert in 6927 // the debugger, is_oop_or_null(false) may subsequently start 6928 // to hold. 6929 assert(obj->is_oop_or_null(true), 6930 "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6931 HeapWord* addr = (HeapWord*)obj; 6932 // Check if oop points into the CMS generation 6933 // and is not marked 6934 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 6935 // a white object ... 6936 // If we manage to "claim" the object, by being the 6937 // first thread to mark it, then we push it on our 6938 // marking stack 6939 if (_bit_map->par_mark(addr)) { // ... now grey 6940 // push on work queue (grey set) 6941 bool simulate_overflow = false; 6942 NOT_PRODUCT( 6943 if (CMSMarkStackOverflowALot && 6944 _collector->par_simulate_overflow()) { 6945 // simulate a stack overflow 6946 simulate_overflow = true; 6947 } 6948 ) 6949 if (simulate_overflow || !_work_queue->push(obj)) { 6950 _collector->par_push_on_overflow_list(obj); 6951 _collector->_par_pmc_remark_ovflw++; // imprecise OK: no need to CAS 6952 } 6953 } // Else, some other thread got there first 6954 } 6955 } 6956 6957 void Par_PushAndMarkClosure::do_oop(oop* p) { Par_PushAndMarkClosure::do_oop_work(p); } 6958 void Par_PushAndMarkClosure::do_oop(narrowOop* p) { Par_PushAndMarkClosure::do_oop_work(p); } 6959 6960 void CMSPrecleanRefsYieldClosure::do_yield_work() { 6961 Mutex* bml = _collector->bitMapLock(); 6962 assert_lock_strong(bml); 6963 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6964 "CMS thread should hold CMS token"); 6965 6966 bml->unlock(); 6967 ConcurrentMarkSweepThread::desynchronize(true); 6968 6969 _collector->stopTimer(); 6970 _collector->incrementYields(); 6971 6972 // See the comment in coordinator_yield() 6973 for (unsigned i = 0; i < CMSYieldSleepCount && 6974 ConcurrentMarkSweepThread::should_yield() && 6975 !CMSCollector::foregroundGCIsActive(); ++i) { 6976 os::sleep(Thread::current(), 1, false); 6977 } 6978 6979 ConcurrentMarkSweepThread::synchronize(true); 6980 bml->lock(); 6981 6982 _collector->startTimer(); 6983 } 6984 6985 bool CMSPrecleanRefsYieldClosure::should_return() { 6986 if (ConcurrentMarkSweepThread::should_yield()) { 6987 do_yield_work(); 6988 } 6989 return _collector->foregroundGCIsActive(); 6990 } 6991 6992 void MarkFromDirtyCardsClosure::do_MemRegion(MemRegion mr) { 6993 assert(((size_t)mr.start())%CardTableModRefBS::card_size_in_words == 0, 6994 "mr should be aligned to start at a card boundary"); 6995 // We'd like to assert: 6996 // assert(mr.word_size()%CardTableModRefBS::card_size_in_words == 0, 6997 // "mr should be a range of cards"); 6998 // However, that would be too strong in one case -- the last 6999 // partition ends at _unallocated_block which, in general, can be 7000 // an arbitrary boundary, not necessarily card aligned. 7001 _num_dirty_cards += mr.word_size()/CardTableModRefBS::card_size_in_words; 7002 _space->object_iterate_mem(mr, &_scan_cl); 7003 } 7004 7005 SweepClosure::SweepClosure(CMSCollector* collector, 7006 ConcurrentMarkSweepGeneration* g, 7007 CMSBitMap* bitMap, bool should_yield) : 7008 _collector(collector), 7009 _g(g), 7010 _sp(g->cmsSpace()), 7011 _limit(_sp->sweep_limit()), 7012 _freelistLock(_sp->freelistLock()), 7013 _bitMap(bitMap), 7014 _yield(should_yield), 7015 _inFreeRange(false), // No free range at beginning of sweep 7016 _freeRangeInFreeLists(false), // No free range at beginning of sweep 7017 _lastFreeRangeCoalesced(false), 7018 _freeFinger(g->used_region().start()) 7019 { 7020 NOT_PRODUCT( 7021 _numObjectsFreed = 0; 7022 _numWordsFreed = 0; 7023 _numObjectsLive = 0; 7024 _numWordsLive = 0; 7025 _numObjectsAlreadyFree = 0; 7026 _numWordsAlreadyFree = 0; 7027 _last_fc = NULL; 7028 7029 _sp->initializeIndexedFreeListArrayReturnedBytes(); 7030 _sp->dictionary()->initialize_dict_returned_bytes(); 7031 ) 7032 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7033 "sweep _limit out of bounds"); 7034 log_develop_trace(gc, sweep)("===================="); 7035 log_develop_trace(gc, sweep)("Starting new sweep with limit " PTR_FORMAT, p2i(_limit)); 7036 } 7037 7038 void SweepClosure::print_on(outputStream* st) const { 7039 tty->print_cr("_sp = [" PTR_FORMAT "," PTR_FORMAT ")", 7040 p2i(_sp->bottom()), p2i(_sp->end())); 7041 tty->print_cr("_limit = " PTR_FORMAT, p2i(_limit)); 7042 tty->print_cr("_freeFinger = " PTR_FORMAT, p2i(_freeFinger)); 7043 NOT_PRODUCT(tty->print_cr("_last_fc = " PTR_FORMAT, p2i(_last_fc));) 7044 tty->print_cr("_inFreeRange = %d, _freeRangeInFreeLists = %d, _lastFreeRangeCoalesced = %d", 7045 _inFreeRange, _freeRangeInFreeLists, _lastFreeRangeCoalesced); 7046 } 7047 7048 #ifndef PRODUCT 7049 // Assertion checking only: no useful work in product mode -- 7050 // however, if any of the flags below become product flags, 7051 // you may need to review this code to see if it needs to be 7052 // enabled in product mode. 7053 SweepClosure::~SweepClosure() { 7054 assert_lock_strong(_freelistLock); 7055 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7056 "sweep _limit out of bounds"); 7057 if (inFreeRange()) { 7058 warning("inFreeRange() should have been reset; dumping state of SweepClosure"); 7059 print(); 7060 ShouldNotReachHere(); 7061 } 7062 7063 if (log_is_enabled(Debug, gc, sweep)) { 7064 log_debug(gc, sweep)("Collected " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes", 7065 _numObjectsFreed, _numWordsFreed*sizeof(HeapWord)); 7066 log_debug(gc, sweep)("Live " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes Already free " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes", 7067 _numObjectsLive, _numWordsLive*sizeof(HeapWord), _numObjectsAlreadyFree, _numWordsAlreadyFree*sizeof(HeapWord)); 7068 size_t totalBytes = (_numWordsFreed + _numWordsLive + _numWordsAlreadyFree) * sizeof(HeapWord); 7069 log_debug(gc, sweep)("Total sweep: " SIZE_FORMAT " bytes", totalBytes); 7070 } 7071 7072 if (log_is_enabled(Trace, gc, sweep) && CMSVerifyReturnedBytes) { 7073 size_t indexListReturnedBytes = _sp->sumIndexedFreeListArrayReturnedBytes(); 7074 size_t dict_returned_bytes = _sp->dictionary()->sum_dict_returned_bytes(); 7075 size_t returned_bytes = indexListReturnedBytes + dict_returned_bytes; 7076 log_trace(gc, sweep)("Returned " SIZE_FORMAT " bytes Indexed List Returned " SIZE_FORMAT " bytes Dictionary Returned " SIZE_FORMAT " bytes", 7077 returned_bytes, indexListReturnedBytes, dict_returned_bytes); 7078 } 7079 log_develop_trace(gc, sweep)("end of sweep with _limit = " PTR_FORMAT, p2i(_limit)); 7080 log_develop_trace(gc, sweep)("================"); 7081 } 7082 #endif // PRODUCT 7083 7084 void SweepClosure::initialize_free_range(HeapWord* freeFinger, 7085 bool freeRangeInFreeLists) { 7086 log_develop_trace(gc, sweep)("---- Start free range at " PTR_FORMAT " with free block (%d)", 7087 p2i(freeFinger), freeRangeInFreeLists); 7088 assert(!inFreeRange(), "Trampling existing free range"); 7089 set_inFreeRange(true); 7090 set_lastFreeRangeCoalesced(false); 7091 7092 set_freeFinger(freeFinger); 7093 set_freeRangeInFreeLists(freeRangeInFreeLists); 7094 if (CMSTestInFreeList) { 7095 if (freeRangeInFreeLists) { 7096 FreeChunk* fc = (FreeChunk*) freeFinger; 7097 assert(fc->is_free(), "A chunk on the free list should be free."); 7098 assert(fc->size() > 0, "Free range should have a size"); 7099 assert(_sp->verify_chunk_in_free_list(fc), "Chunk is not in free lists"); 7100 } 7101 } 7102 } 7103 7104 // Note that the sweeper runs concurrently with mutators. Thus, 7105 // it is possible for direct allocation in this generation to happen 7106 // in the middle of the sweep. Note that the sweeper also coalesces 7107 // contiguous free blocks. Thus, unless the sweeper and the allocator 7108 // synchronize appropriately freshly allocated blocks may get swept up. 7109 // This is accomplished by the sweeper locking the free lists while 7110 // it is sweeping. Thus blocks that are determined to be free are 7111 // indeed free. There is however one additional complication: 7112 // blocks that have been allocated since the final checkpoint and 7113 // mark, will not have been marked and so would be treated as 7114 // unreachable and swept up. To prevent this, the allocator marks 7115 // the bit map when allocating during the sweep phase. This leads, 7116 // however, to a further complication -- objects may have been allocated 7117 // but not yet initialized -- in the sense that the header isn't yet 7118 // installed. The sweeper can not then determine the size of the block 7119 // in order to skip over it. To deal with this case, we use a technique 7120 // (due to Printezis) to encode such uninitialized block sizes in the 7121 // bit map. Since the bit map uses a bit per every HeapWord, but the 7122 // CMS generation has a minimum object size of 3 HeapWords, it follows 7123 // that "normal marks" won't be adjacent in the bit map (there will 7124 // always be at least two 0 bits between successive 1 bits). We make use 7125 // of these "unused" bits to represent uninitialized blocks -- the bit 7126 // corresponding to the start of the uninitialized object and the next 7127 // bit are both set. Finally, a 1 bit marks the end of the object that 7128 // started with the two consecutive 1 bits to indicate its potentially 7129 // uninitialized state. 7130 7131 size_t SweepClosure::do_blk_careful(HeapWord* addr) { 7132 FreeChunk* fc = (FreeChunk*)addr; 7133 size_t res; 7134 7135 // Check if we are done sweeping. Below we check "addr >= _limit" rather 7136 // than "addr == _limit" because although _limit was a block boundary when 7137 // we started the sweep, it may no longer be one because heap expansion 7138 // may have caused us to coalesce the block ending at the address _limit 7139 // with a newly expanded chunk (this happens when _limit was set to the 7140 // previous _end of the space), so we may have stepped past _limit: 7141 // see the following Zeno-like trail of CRs 6977970, 7008136, 7042740. 7142 if (addr >= _limit) { // we have swept up to or past the limit: finish up 7143 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7144 "sweep _limit out of bounds"); 7145 assert(addr < _sp->end(), "addr out of bounds"); 7146 // Flush any free range we might be holding as a single 7147 // coalesced chunk to the appropriate free list. 7148 if (inFreeRange()) { 7149 assert(freeFinger() >= _sp->bottom() && freeFinger() < _limit, 7150 "freeFinger() " PTR_FORMAT " is out-of-bounds", p2i(freeFinger())); 7151 flush_cur_free_chunk(freeFinger(), 7152 pointer_delta(addr, freeFinger())); 7153 log_develop_trace(gc, sweep)("Sweep: last chunk: put_free_blk " PTR_FORMAT " (" SIZE_FORMAT ") [coalesced:%d]", 7154 p2i(freeFinger()), pointer_delta(addr, freeFinger()), 7155 lastFreeRangeCoalesced() ? 1 : 0); 7156 } 7157 7158 // help the iterator loop finish 7159 return pointer_delta(_sp->end(), addr); 7160 } 7161 7162 assert(addr < _limit, "sweep invariant"); 7163 // check if we should yield 7164 do_yield_check(addr); 7165 if (fc->is_free()) { 7166 // Chunk that is already free 7167 res = fc->size(); 7168 do_already_free_chunk(fc); 7169 debug_only(_sp->verifyFreeLists()); 7170 // If we flush the chunk at hand in lookahead_and_flush() 7171 // and it's coalesced with a preceding chunk, then the 7172 // process of "mangling" the payload of the coalesced block 7173 // will cause erasure of the size information from the 7174 // (erstwhile) header of all the coalesced blocks but the 7175 // first, so the first disjunct in the assert will not hold 7176 // in that specific case (in which case the second disjunct 7177 // will hold). 7178 assert(res == fc->size() || ((HeapWord*)fc) + res >= _limit, 7179 "Otherwise the size info doesn't change at this step"); 7180 NOT_PRODUCT( 7181 _numObjectsAlreadyFree++; 7182 _numWordsAlreadyFree += res; 7183 ) 7184 NOT_PRODUCT(_last_fc = fc;) 7185 } else if (!_bitMap->isMarked(addr)) { 7186 // Chunk is fresh garbage 7187 res = do_garbage_chunk(fc); 7188 debug_only(_sp->verifyFreeLists()); 7189 NOT_PRODUCT( 7190 _numObjectsFreed++; 7191 _numWordsFreed += res; 7192 ) 7193 } else { 7194 // Chunk that is alive. 7195 res = do_live_chunk(fc); 7196 debug_only(_sp->verifyFreeLists()); 7197 NOT_PRODUCT( 7198 _numObjectsLive++; 7199 _numWordsLive += res; 7200 ) 7201 } 7202 return res; 7203 } 7204 7205 // For the smart allocation, record following 7206 // split deaths - a free chunk is removed from its free list because 7207 // it is being split into two or more chunks. 7208 // split birth - a free chunk is being added to its free list because 7209 // a larger free chunk has been split and resulted in this free chunk. 7210 // coal death - a free chunk is being removed from its free list because 7211 // it is being coalesced into a large free chunk. 7212 // coal birth - a free chunk is being added to its free list because 7213 // it was created when two or more free chunks where coalesced into 7214 // this free chunk. 7215 // 7216 // These statistics are used to determine the desired number of free 7217 // chunks of a given size. The desired number is chosen to be relative 7218 // to the end of a CMS sweep. The desired number at the end of a sweep 7219 // is the 7220 // count-at-end-of-previous-sweep (an amount that was enough) 7221 // - count-at-beginning-of-current-sweep (the excess) 7222 // + split-births (gains in this size during interval) 7223 // - split-deaths (demands on this size during interval) 7224 // where the interval is from the end of one sweep to the end of the 7225 // next. 7226 // 7227 // When sweeping the sweeper maintains an accumulated chunk which is 7228 // the chunk that is made up of chunks that have been coalesced. That 7229 // will be termed the left-hand chunk. A new chunk of garbage that 7230 // is being considered for coalescing will be referred to as the 7231 // right-hand chunk. 7232 // 7233 // When making a decision on whether to coalesce a right-hand chunk with 7234 // the current left-hand chunk, the current count vs. the desired count 7235 // of the left-hand chunk is considered. Also if the right-hand chunk 7236 // is near the large chunk at the end of the heap (see 7237 // ConcurrentMarkSweepGeneration::isNearLargestChunk()), then the 7238 // left-hand chunk is coalesced. 7239 // 7240 // When making a decision about whether to split a chunk, the desired count 7241 // vs. the current count of the candidate to be split is also considered. 7242 // If the candidate is underpopulated (currently fewer chunks than desired) 7243 // a chunk of an overpopulated (currently more chunks than desired) size may 7244 // be chosen. The "hint" associated with a free list, if non-null, points 7245 // to a free list which may be overpopulated. 7246 // 7247 7248 void SweepClosure::do_already_free_chunk(FreeChunk* fc) { 7249 const size_t size = fc->size(); 7250 // Chunks that cannot be coalesced are not in the 7251 // free lists. 7252 if (CMSTestInFreeList && !fc->cantCoalesce()) { 7253 assert(_sp->verify_chunk_in_free_list(fc), 7254 "free chunk should be in free lists"); 7255 } 7256 // a chunk that is already free, should not have been 7257 // marked in the bit map 7258 HeapWord* const addr = (HeapWord*) fc; 7259 assert(!_bitMap->isMarked(addr), "free chunk should be unmarked"); 7260 // Verify that the bit map has no bits marked between 7261 // addr and purported end of this block. 7262 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 7263 7264 // Some chunks cannot be coalesced under any circumstances. 7265 // See the definition of cantCoalesce(). 7266 if (!fc->cantCoalesce()) { 7267 // This chunk can potentially be coalesced. 7268 // All the work is done in 7269 do_post_free_or_garbage_chunk(fc, size); 7270 // Note that if the chunk is not coalescable (the else arm 7271 // below), we unconditionally flush, without needing to do 7272 // a "lookahead," as we do below. 7273 if (inFreeRange()) lookahead_and_flush(fc, size); 7274 } else { 7275 // Code path common to both original and adaptive free lists. 7276 7277 // cant coalesce with previous block; this should be treated 7278 // as the end of a free run if any 7279 if (inFreeRange()) { 7280 // we kicked some butt; time to pick up the garbage 7281 assert(freeFinger() < addr, "freeFinger points too high"); 7282 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 7283 } 7284 // else, nothing to do, just continue 7285 } 7286 } 7287 7288 size_t SweepClosure::do_garbage_chunk(FreeChunk* fc) { 7289 // This is a chunk of garbage. It is not in any free list. 7290 // Add it to a free list or let it possibly be coalesced into 7291 // a larger chunk. 7292 HeapWord* const addr = (HeapWord*) fc; 7293 const size_t size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 7294 7295 // Verify that the bit map has no bits marked between 7296 // addr and purported end of just dead object. 7297 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 7298 do_post_free_or_garbage_chunk(fc, size); 7299 7300 assert(_limit >= addr + size, 7301 "A freshly garbage chunk can't possibly straddle over _limit"); 7302 if (inFreeRange()) lookahead_and_flush(fc, size); 7303 return size; 7304 } 7305 7306 size_t SweepClosure::do_live_chunk(FreeChunk* fc) { 7307 HeapWord* addr = (HeapWord*) fc; 7308 // The sweeper has just found a live object. Return any accumulated 7309 // left hand chunk to the free lists. 7310 if (inFreeRange()) { 7311 assert(freeFinger() < addr, "freeFinger points too high"); 7312 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 7313 } 7314 7315 // This object is live: we'd normally expect this to be 7316 // an oop, and like to assert the following: 7317 // assert(oop(addr)->is_oop(), "live block should be an oop"); 7318 // However, as we commented above, this may be an object whose 7319 // header hasn't yet been initialized. 7320 size_t size; 7321 assert(_bitMap->isMarked(addr), "Tautology for this control point"); 7322 if (_bitMap->isMarked(addr + 1)) { 7323 // Determine the size from the bit map, rather than trying to 7324 // compute it from the object header. 7325 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 7326 size = pointer_delta(nextOneAddr + 1, addr); 7327 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 7328 "alignment problem"); 7329 7330 #ifdef ASSERT 7331 if (oop(addr)->klass_or_null() != NULL) { 7332 // Ignore mark word because we are running concurrent with mutators 7333 assert(oop(addr)->is_oop(true), "live block should be an oop"); 7334 assert(size == 7335 CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()), 7336 "P-mark and computed size do not agree"); 7337 } 7338 #endif 7339 7340 } else { 7341 // This should be an initialized object that's alive. 7342 assert(oop(addr)->klass_or_null() != NULL, 7343 "Should be an initialized object"); 7344 // Ignore mark word because we are running concurrent with mutators 7345 assert(oop(addr)->is_oop(true), "live block should be an oop"); 7346 // Verify that the bit map has no bits marked between 7347 // addr and purported end of this block. 7348 size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 7349 assert(size >= 3, "Necessary for Printezis marks to work"); 7350 assert(!_bitMap->isMarked(addr+1), "Tautology for this control point"); 7351 DEBUG_ONLY(_bitMap->verifyNoOneBitsInRange(addr+2, addr+size);) 7352 } 7353 return size; 7354 } 7355 7356 void SweepClosure::do_post_free_or_garbage_chunk(FreeChunk* fc, 7357 size_t chunkSize) { 7358 // do_post_free_or_garbage_chunk() should only be called in the case 7359 // of the adaptive free list allocator. 7360 const bool fcInFreeLists = fc->is_free(); 7361 assert((HeapWord*)fc <= _limit, "sweep invariant"); 7362 if (CMSTestInFreeList && fcInFreeLists) { 7363 assert(_sp->verify_chunk_in_free_list(fc), "free chunk is not in free lists"); 7364 } 7365 7366 log_develop_trace(gc, sweep)(" -- pick up another chunk at " PTR_FORMAT " (" SIZE_FORMAT ")", p2i(fc), chunkSize); 7367 7368 HeapWord* const fc_addr = (HeapWord*) fc; 7369 7370 bool coalesce = false; 7371 const size_t left = pointer_delta(fc_addr, freeFinger()); 7372 const size_t right = chunkSize; 7373 switch (FLSCoalescePolicy) { 7374 // numeric value forms a coalition aggressiveness metric 7375 case 0: { // never coalesce 7376 coalesce = false; 7377 break; 7378 } 7379 case 1: { // coalesce if left & right chunks on overpopulated lists 7380 coalesce = _sp->coalOverPopulated(left) && 7381 _sp->coalOverPopulated(right); 7382 break; 7383 } 7384 case 2: { // coalesce if left chunk on overpopulated list (default) 7385 coalesce = _sp->coalOverPopulated(left); 7386 break; 7387 } 7388 case 3: { // coalesce if left OR right chunk on overpopulated list 7389 coalesce = _sp->coalOverPopulated(left) || 7390 _sp->coalOverPopulated(right); 7391 break; 7392 } 7393 case 4: { // always coalesce 7394 coalesce = true; 7395 break; 7396 } 7397 default: 7398 ShouldNotReachHere(); 7399 } 7400 7401 // Should the current free range be coalesced? 7402 // If the chunk is in a free range and either we decided to coalesce above 7403 // or the chunk is near the large block at the end of the heap 7404 // (isNearLargestChunk() returns true), then coalesce this chunk. 7405 const bool doCoalesce = inFreeRange() 7406 && (coalesce || _g->isNearLargestChunk(fc_addr)); 7407 if (doCoalesce) { 7408 // Coalesce the current free range on the left with the new 7409 // chunk on the right. If either is on a free list, 7410 // it must be removed from the list and stashed in the closure. 7411 if (freeRangeInFreeLists()) { 7412 FreeChunk* const ffc = (FreeChunk*)freeFinger(); 7413 assert(ffc->size() == pointer_delta(fc_addr, freeFinger()), 7414 "Size of free range is inconsistent with chunk size."); 7415 if (CMSTestInFreeList) { 7416 assert(_sp->verify_chunk_in_free_list(ffc), 7417 "Chunk is not in free lists"); 7418 } 7419 _sp->coalDeath(ffc->size()); 7420 _sp->removeFreeChunkFromFreeLists(ffc); 7421 set_freeRangeInFreeLists(false); 7422 } 7423 if (fcInFreeLists) { 7424 _sp->coalDeath(chunkSize); 7425 assert(fc->size() == chunkSize, 7426 "The chunk has the wrong size or is not in the free lists"); 7427 _sp->removeFreeChunkFromFreeLists(fc); 7428 } 7429 set_lastFreeRangeCoalesced(true); 7430 print_free_block_coalesced(fc); 7431 } else { // not in a free range and/or should not coalesce 7432 // Return the current free range and start a new one. 7433 if (inFreeRange()) { 7434 // In a free range but cannot coalesce with the right hand chunk. 7435 // Put the current free range into the free lists. 7436 flush_cur_free_chunk(freeFinger(), 7437 pointer_delta(fc_addr, freeFinger())); 7438 } 7439 // Set up for new free range. Pass along whether the right hand 7440 // chunk is in the free lists. 7441 initialize_free_range((HeapWord*)fc, fcInFreeLists); 7442 } 7443 } 7444 7445 // Lookahead flush: 7446 // If we are tracking a free range, and this is the last chunk that 7447 // we'll look at because its end crosses past _limit, we'll preemptively 7448 // flush it along with any free range we may be holding on to. Note that 7449 // this can be the case only for an already free or freshly garbage 7450 // chunk. If this block is an object, it can never straddle 7451 // over _limit. The "straddling" occurs when _limit is set at 7452 // the previous end of the space when this cycle started, and 7453 // a subsequent heap expansion caused the previously co-terminal 7454 // free block to be coalesced with the newly expanded portion, 7455 // thus rendering _limit a non-block-boundary making it dangerous 7456 // for the sweeper to step over and examine. 7457 void SweepClosure::lookahead_and_flush(FreeChunk* fc, size_t chunk_size) { 7458 assert(inFreeRange(), "Should only be called if currently in a free range."); 7459 HeapWord* const eob = ((HeapWord*)fc) + chunk_size; 7460 assert(_sp->used_region().contains(eob - 1), 7461 "eob = " PTR_FORMAT " eob-1 = " PTR_FORMAT " _limit = " PTR_FORMAT 7462 " out of bounds wrt _sp = [" PTR_FORMAT "," PTR_FORMAT ")" 7463 " when examining fc = " PTR_FORMAT "(" SIZE_FORMAT ")", 7464 p2i(eob), p2i(eob-1), p2i(_limit), p2i(_sp->bottom()), p2i(_sp->end()), p2i(fc), chunk_size); 7465 if (eob >= _limit) { 7466 assert(eob == _limit || fc->is_free(), "Only a free chunk should allow us to cross over the limit"); 7467 log_develop_trace(gc, sweep)("_limit " PTR_FORMAT " reached or crossed by block " 7468 "[" PTR_FORMAT "," PTR_FORMAT ") in space " 7469 "[" PTR_FORMAT "," PTR_FORMAT ")", 7470 p2i(_limit), p2i(fc), p2i(eob), p2i(_sp->bottom()), p2i(_sp->end())); 7471 // Return the storage we are tracking back into the free lists. 7472 log_develop_trace(gc, sweep)("Flushing ... "); 7473 assert(freeFinger() < eob, "Error"); 7474 flush_cur_free_chunk( freeFinger(), pointer_delta(eob, freeFinger())); 7475 } 7476 } 7477 7478 void SweepClosure::flush_cur_free_chunk(HeapWord* chunk, size_t size) { 7479 assert(inFreeRange(), "Should only be called if currently in a free range."); 7480 assert(size > 0, 7481 "A zero sized chunk cannot be added to the free lists."); 7482 if (!freeRangeInFreeLists()) { 7483 if (CMSTestInFreeList) { 7484 FreeChunk* fc = (FreeChunk*) chunk; 7485 fc->set_size(size); 7486 assert(!_sp->verify_chunk_in_free_list(fc), 7487 "chunk should not be in free lists yet"); 7488 } 7489 log_develop_trace(gc, sweep)(" -- add free block " PTR_FORMAT " (" SIZE_FORMAT ") to free lists", p2i(chunk), size); 7490 // A new free range is going to be starting. The current 7491 // free range has not been added to the free lists yet or 7492 // was removed so add it back. 7493 // If the current free range was coalesced, then the death 7494 // of the free range was recorded. Record a birth now. 7495 if (lastFreeRangeCoalesced()) { 7496 _sp->coalBirth(size); 7497 } 7498 _sp->addChunkAndRepairOffsetTable(chunk, size, 7499 lastFreeRangeCoalesced()); 7500 } else { 7501 log_develop_trace(gc, sweep)("Already in free list: nothing to flush"); 7502 } 7503 set_inFreeRange(false); 7504 set_freeRangeInFreeLists(false); 7505 } 7506 7507 // We take a break if we've been at this for a while, 7508 // so as to avoid monopolizing the locks involved. 7509 void SweepClosure::do_yield_work(HeapWord* addr) { 7510 // Return current free chunk being used for coalescing (if any) 7511 // to the appropriate freelist. After yielding, the next 7512 // free block encountered will start a coalescing range of 7513 // free blocks. If the next free block is adjacent to the 7514 // chunk just flushed, they will need to wait for the next 7515 // sweep to be coalesced. 7516 if (inFreeRange()) { 7517 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 7518 } 7519 7520 // First give up the locks, then yield, then re-lock. 7521 // We should probably use a constructor/destructor idiom to 7522 // do this unlock/lock or modify the MutexUnlocker class to 7523 // serve our purpose. XXX 7524 assert_lock_strong(_bitMap->lock()); 7525 assert_lock_strong(_freelistLock); 7526 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 7527 "CMS thread should hold CMS token"); 7528 _bitMap->lock()->unlock(); 7529 _freelistLock->unlock(); 7530 ConcurrentMarkSweepThread::desynchronize(true); 7531 _collector->stopTimer(); 7532 _collector->incrementYields(); 7533 7534 // See the comment in coordinator_yield() 7535 for (unsigned i = 0; i < CMSYieldSleepCount && 7536 ConcurrentMarkSweepThread::should_yield() && 7537 !CMSCollector::foregroundGCIsActive(); ++i) { 7538 os::sleep(Thread::current(), 1, false); 7539 } 7540 7541 ConcurrentMarkSweepThread::synchronize(true); 7542 _freelistLock->lock(); 7543 _bitMap->lock()->lock_without_safepoint_check(); 7544 _collector->startTimer(); 7545 } 7546 7547 #ifndef PRODUCT 7548 // This is actually very useful in a product build if it can 7549 // be called from the debugger. Compile it into the product 7550 // as needed. 7551 bool debug_verify_chunk_in_free_list(FreeChunk* fc) { 7552 return debug_cms_space->verify_chunk_in_free_list(fc); 7553 } 7554 #endif 7555 7556 void SweepClosure::print_free_block_coalesced(FreeChunk* fc) const { 7557 log_develop_trace(gc, sweep)("Sweep:coal_free_blk " PTR_FORMAT " (" SIZE_FORMAT ")", 7558 p2i(fc), fc->size()); 7559 } 7560 7561 // CMSIsAliveClosure 7562 bool CMSIsAliveClosure::do_object_b(oop obj) { 7563 HeapWord* addr = (HeapWord*)obj; 7564 return addr != NULL && 7565 (!_span.contains(addr) || _bit_map->isMarked(addr)); 7566 } 7567 7568 7569 CMSKeepAliveClosure::CMSKeepAliveClosure( CMSCollector* collector, 7570 MemRegion span, 7571 CMSBitMap* bit_map, CMSMarkStack* mark_stack, 7572 bool cpc): 7573 _collector(collector), 7574 _span(span), 7575 _bit_map(bit_map), 7576 _mark_stack(mark_stack), 7577 _concurrent_precleaning(cpc) { 7578 assert(!_span.is_empty(), "Empty span could spell trouble"); 7579 } 7580 7581 7582 // CMSKeepAliveClosure: the serial version 7583 void CMSKeepAliveClosure::do_oop(oop obj) { 7584 HeapWord* addr = (HeapWord*)obj; 7585 if (_span.contains(addr) && 7586 !_bit_map->isMarked(addr)) { 7587 _bit_map->mark(addr); 7588 bool simulate_overflow = false; 7589 NOT_PRODUCT( 7590 if (CMSMarkStackOverflowALot && 7591 _collector->simulate_overflow()) { 7592 // simulate a stack overflow 7593 simulate_overflow = true; 7594 } 7595 ) 7596 if (simulate_overflow || !_mark_stack->push(obj)) { 7597 if (_concurrent_precleaning) { 7598 // We dirty the overflown object and let the remark 7599 // phase deal with it. 7600 assert(_collector->overflow_list_is_empty(), "Error"); 7601 // In the case of object arrays, we need to dirty all of 7602 // the cards that the object spans. No locking or atomics 7603 // are needed since no one else can be mutating the mod union 7604 // table. 7605 if (obj->is_objArray()) { 7606 size_t sz = obj->size(); 7607 HeapWord* end_card_addr = 7608 (HeapWord*)round_to((intptr_t)(addr+sz), CardTableModRefBS::card_size); 7609 MemRegion redirty_range = MemRegion(addr, end_card_addr); 7610 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 7611 _collector->_modUnionTable.mark_range(redirty_range); 7612 } else { 7613 _collector->_modUnionTable.mark(addr); 7614 } 7615 _collector->_ser_kac_preclean_ovflw++; 7616 } else { 7617 _collector->push_on_overflow_list(obj); 7618 _collector->_ser_kac_ovflw++; 7619 } 7620 } 7621 } 7622 } 7623 7624 void CMSKeepAliveClosure::do_oop(oop* p) { CMSKeepAliveClosure::do_oop_work(p); } 7625 void CMSKeepAliveClosure::do_oop(narrowOop* p) { CMSKeepAliveClosure::do_oop_work(p); } 7626 7627 // CMSParKeepAliveClosure: a parallel version of the above. 7628 // The work queues are private to each closure (thread), 7629 // but (may be) available for stealing by other threads. 7630 void CMSParKeepAliveClosure::do_oop(oop obj) { 7631 HeapWord* addr = (HeapWord*)obj; 7632 if (_span.contains(addr) && 7633 !_bit_map->isMarked(addr)) { 7634 // In general, during recursive tracing, several threads 7635 // may be concurrently getting here; the first one to 7636 // "tag" it, claims it. 7637 if (_bit_map->par_mark(addr)) { 7638 bool res = _work_queue->push(obj); 7639 assert(res, "Low water mark should be much less than capacity"); 7640 // Do a recursive trim in the hope that this will keep 7641 // stack usage lower, but leave some oops for potential stealers 7642 trim_queue(_low_water_mark); 7643 } // Else, another thread got there first 7644 } 7645 } 7646 7647 void CMSParKeepAliveClosure::do_oop(oop* p) { CMSParKeepAliveClosure::do_oop_work(p); } 7648 void CMSParKeepAliveClosure::do_oop(narrowOop* p) { CMSParKeepAliveClosure::do_oop_work(p); } 7649 7650 void CMSParKeepAliveClosure::trim_queue(uint max) { 7651 while (_work_queue->size() > max) { 7652 oop new_oop; 7653 if (_work_queue->pop_local(new_oop)) { 7654 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 7655 assert(_bit_map->isMarked((HeapWord*)new_oop), 7656 "no white objects on this stack!"); 7657 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 7658 // iterate over the oops in this oop, marking and pushing 7659 // the ones in CMS heap (i.e. in _span). 7660 new_oop->oop_iterate(&_mark_and_push); 7661 } 7662 } 7663 } 7664 7665 CMSInnerParMarkAndPushClosure::CMSInnerParMarkAndPushClosure( 7666 CMSCollector* collector, 7667 MemRegion span, CMSBitMap* bit_map, 7668 OopTaskQueue* work_queue): 7669 _collector(collector), 7670 _span(span), 7671 _bit_map(bit_map), 7672 _work_queue(work_queue) { } 7673 7674 void CMSInnerParMarkAndPushClosure::do_oop(oop obj) { 7675 HeapWord* addr = (HeapWord*)obj; 7676 if (_span.contains(addr) && 7677 !_bit_map->isMarked(addr)) { 7678 if (_bit_map->par_mark(addr)) { 7679 bool simulate_overflow = false; 7680 NOT_PRODUCT( 7681 if (CMSMarkStackOverflowALot && 7682 _collector->par_simulate_overflow()) { 7683 // simulate a stack overflow 7684 simulate_overflow = true; 7685 } 7686 ) 7687 if (simulate_overflow || !_work_queue->push(obj)) { 7688 _collector->par_push_on_overflow_list(obj); 7689 _collector->_par_kac_ovflw++; 7690 } 7691 } // Else another thread got there already 7692 } 7693 } 7694 7695 void CMSInnerParMarkAndPushClosure::do_oop(oop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); } 7696 void CMSInnerParMarkAndPushClosure::do_oop(narrowOop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); } 7697 7698 ////////////////////////////////////////////////////////////////// 7699 // CMSExpansionCause ///////////////////////////// 7700 ////////////////////////////////////////////////////////////////// 7701 const char* CMSExpansionCause::to_string(CMSExpansionCause::Cause cause) { 7702 switch (cause) { 7703 case _no_expansion: 7704 return "No expansion"; 7705 case _satisfy_free_ratio: 7706 return "Free ratio"; 7707 case _satisfy_promotion: 7708 return "Satisfy promotion"; 7709 case _satisfy_allocation: 7710 return "allocation"; 7711 case _allocate_par_lab: 7712 return "Par LAB"; 7713 case _allocate_par_spooling_space: 7714 return "Par Spooling Space"; 7715 case _adaptive_size_policy: 7716 return "Ergonomics"; 7717 default: 7718 return "unknown"; 7719 } 7720 } 7721 7722 void CMSDrainMarkingStackClosure::do_void() { 7723 // the max number to take from overflow list at a time 7724 const size_t num = _mark_stack->capacity()/4; 7725 assert(!_concurrent_precleaning || _collector->overflow_list_is_empty(), 7726 "Overflow list should be NULL during concurrent phases"); 7727 while (!_mark_stack->isEmpty() || 7728 // if stack is empty, check the overflow list 7729 _collector->take_from_overflow_list(num, _mark_stack)) { 7730 oop obj = _mark_stack->pop(); 7731 HeapWord* addr = (HeapWord*)obj; 7732 assert(_span.contains(addr), "Should be within span"); 7733 assert(_bit_map->isMarked(addr), "Should be marked"); 7734 assert(obj->is_oop(), "Should be an oop"); 7735 obj->oop_iterate(_keep_alive); 7736 } 7737 } 7738 7739 void CMSParDrainMarkingStackClosure::do_void() { 7740 // drain queue 7741 trim_queue(0); 7742 } 7743 7744 // Trim our work_queue so its length is below max at return 7745 void CMSParDrainMarkingStackClosure::trim_queue(uint max) { 7746 while (_work_queue->size() > max) { 7747 oop new_oop; 7748 if (_work_queue->pop_local(new_oop)) { 7749 assert(new_oop->is_oop(), "Expected an oop"); 7750 assert(_bit_map->isMarked((HeapWord*)new_oop), 7751 "no white objects on this stack!"); 7752 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 7753 // iterate over the oops in this oop, marking and pushing 7754 // the ones in CMS heap (i.e. in _span). 7755 new_oop->oop_iterate(&_mark_and_push); 7756 } 7757 } 7758 } 7759 7760 //////////////////////////////////////////////////////////////////// 7761 // Support for Marking Stack Overflow list handling and related code 7762 //////////////////////////////////////////////////////////////////// 7763 // Much of the following code is similar in shape and spirit to the 7764 // code used in ParNewGC. We should try and share that code 7765 // as much as possible in the future. 7766 7767 #ifndef PRODUCT 7768 // Debugging support for CMSStackOverflowALot 7769 7770 // It's OK to call this multi-threaded; the worst thing 7771 // that can happen is that we'll get a bunch of closely 7772 // spaced simulated overflows, but that's OK, in fact 7773 // probably good as it would exercise the overflow code 7774 // under contention. 7775 bool CMSCollector::simulate_overflow() { 7776 if (_overflow_counter-- <= 0) { // just being defensive 7777 _overflow_counter = CMSMarkStackOverflowInterval; 7778 return true; 7779 } else { 7780 return false; 7781 } 7782 } 7783 7784 bool CMSCollector::par_simulate_overflow() { 7785 return simulate_overflow(); 7786 } 7787 #endif 7788 7789 // Single-threaded 7790 bool CMSCollector::take_from_overflow_list(size_t num, CMSMarkStack* stack) { 7791 assert(stack->isEmpty(), "Expected precondition"); 7792 assert(stack->capacity() > num, "Shouldn't bite more than can chew"); 7793 size_t i = num; 7794 oop cur = _overflow_list; 7795 const markOop proto = markOopDesc::prototype(); 7796 NOT_PRODUCT(ssize_t n = 0;) 7797 for (oop next; i > 0 && cur != NULL; cur = next, i--) { 7798 next = oop(cur->mark()); 7799 cur->set_mark(proto); // until proven otherwise 7800 assert(cur->is_oop(), "Should be an oop"); 7801 bool res = stack->push(cur); 7802 assert(res, "Bit off more than can chew?"); 7803 NOT_PRODUCT(n++;) 7804 } 7805 _overflow_list = cur; 7806 #ifndef PRODUCT 7807 assert(_num_par_pushes >= n, "Too many pops?"); 7808 _num_par_pushes -=n; 7809 #endif 7810 return !stack->isEmpty(); 7811 } 7812 7813 #define BUSY (cast_to_oop<intptr_t>(0x1aff1aff)) 7814 // (MT-safe) Get a prefix of at most "num" from the list. 7815 // The overflow list is chained through the mark word of 7816 // each object in the list. We fetch the entire list, 7817 // break off a prefix of the right size and return the 7818 // remainder. If other threads try to take objects from 7819 // the overflow list at that time, they will wait for 7820 // some time to see if data becomes available. If (and 7821 // only if) another thread places one or more object(s) 7822 // on the global list before we have returned the suffix 7823 // to the global list, we will walk down our local list 7824 // to find its end and append the global list to 7825 // our suffix before returning it. This suffix walk can 7826 // prove to be expensive (quadratic in the amount of traffic) 7827 // when there are many objects in the overflow list and 7828 // there is much producer-consumer contention on the list. 7829 // *NOTE*: The overflow list manipulation code here and 7830 // in ParNewGeneration:: are very similar in shape, 7831 // except that in the ParNew case we use the old (from/eden) 7832 // copy of the object to thread the list via its klass word. 7833 // Because of the common code, if you make any changes in 7834 // the code below, please check the ParNew version to see if 7835 // similar changes might be needed. 7836 // CR 6797058 has been filed to consolidate the common code. 7837 bool CMSCollector::par_take_from_overflow_list(size_t num, 7838 OopTaskQueue* work_q, 7839 int no_of_gc_threads) { 7840 assert(work_q->size() == 0, "First empty local work queue"); 7841 assert(num < work_q->max_elems(), "Can't bite more than we can chew"); 7842 if (_overflow_list == NULL) { 7843 return false; 7844 } 7845 // Grab the entire list; we'll put back a suffix 7846 oop prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list)); 7847 Thread* tid = Thread::current(); 7848 // Before "no_of_gc_threads" was introduced CMSOverflowSpinCount was 7849 // set to ParallelGCThreads. 7850 size_t CMSOverflowSpinCount = (size_t) no_of_gc_threads; // was ParallelGCThreads; 7851 size_t sleep_time_millis = MAX2((size_t)1, num/100); 7852 // If the list is busy, we spin for a short while, 7853 // sleeping between attempts to get the list. 7854 for (size_t spin = 0; prefix == BUSY && spin < CMSOverflowSpinCount; spin++) { 7855 os::sleep(tid, sleep_time_millis, false); 7856 if (_overflow_list == NULL) { 7857 // Nothing left to take 7858 return false; 7859 } else if (_overflow_list != BUSY) { 7860 // Try and grab the prefix 7861 prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list)); 7862 } 7863 } 7864 // If the list was found to be empty, or we spun long 7865 // enough, we give up and return empty-handed. If we leave 7866 // the list in the BUSY state below, it must be the case that 7867 // some other thread holds the overflow list and will set it 7868 // to a non-BUSY state in the future. 7869 if (prefix == NULL || prefix == BUSY) { 7870 // Nothing to take or waited long enough 7871 if (prefix == NULL) { 7872 // Write back the NULL in case we overwrote it with BUSY above 7873 // and it is still the same value. 7874 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY); 7875 } 7876 return false; 7877 } 7878 assert(prefix != NULL && prefix != BUSY, "Error"); 7879 size_t i = num; 7880 oop cur = prefix; 7881 // Walk down the first "num" objects, unless we reach the end. 7882 for (; i > 1 && cur->mark() != NULL; cur = oop(cur->mark()), i--); 7883 if (cur->mark() == NULL) { 7884 // We have "num" or fewer elements in the list, so there 7885 // is nothing to return to the global list. 7886 // Write back the NULL in lieu of the BUSY we wrote 7887 // above, if it is still the same value. 7888 if (_overflow_list == BUSY) { 7889 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY); 7890 } 7891 } else { 7892 // Chop off the suffix and return it to the global list. 7893 assert(cur->mark() != BUSY, "Error"); 7894 oop suffix_head = cur->mark(); // suffix will be put back on global list 7895 cur->set_mark(NULL); // break off suffix 7896 // It's possible that the list is still in the empty(busy) state 7897 // we left it in a short while ago; in that case we may be 7898 // able to place back the suffix without incurring the cost 7899 // of a walk down the list. 7900 oop observed_overflow_list = _overflow_list; 7901 oop cur_overflow_list = observed_overflow_list; 7902 bool attached = false; 7903 while (observed_overflow_list == BUSY || observed_overflow_list == NULL) { 7904 observed_overflow_list = 7905 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list); 7906 if (cur_overflow_list == observed_overflow_list) { 7907 attached = true; 7908 break; 7909 } else cur_overflow_list = observed_overflow_list; 7910 } 7911 if (!attached) { 7912 // Too bad, someone else sneaked in (at least) an element; we'll need 7913 // to do a splice. Find tail of suffix so we can prepend suffix to global 7914 // list. 7915 for (cur = suffix_head; cur->mark() != NULL; cur = (oop)(cur->mark())); 7916 oop suffix_tail = cur; 7917 assert(suffix_tail != NULL && suffix_tail->mark() == NULL, 7918 "Tautology"); 7919 observed_overflow_list = _overflow_list; 7920 do { 7921 cur_overflow_list = observed_overflow_list; 7922 if (cur_overflow_list != BUSY) { 7923 // Do the splice ... 7924 suffix_tail->set_mark(markOop(cur_overflow_list)); 7925 } else { // cur_overflow_list == BUSY 7926 suffix_tail->set_mark(NULL); 7927 } 7928 // ... and try to place spliced list back on overflow_list ... 7929 observed_overflow_list = 7930 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list); 7931 } while (cur_overflow_list != observed_overflow_list); 7932 // ... until we have succeeded in doing so. 7933 } 7934 } 7935 7936 // Push the prefix elements on work_q 7937 assert(prefix != NULL, "control point invariant"); 7938 const markOop proto = markOopDesc::prototype(); 7939 oop next; 7940 NOT_PRODUCT(ssize_t n = 0;) 7941 for (cur = prefix; cur != NULL; cur = next) { 7942 next = oop(cur->mark()); 7943 cur->set_mark(proto); // until proven otherwise 7944 assert(cur->is_oop(), "Should be an oop"); 7945 bool res = work_q->push(cur); 7946 assert(res, "Bit off more than we can chew?"); 7947 NOT_PRODUCT(n++;) 7948 } 7949 #ifndef PRODUCT 7950 assert(_num_par_pushes >= n, "Too many pops?"); 7951 Atomic::add_ptr(-(intptr_t)n, &_num_par_pushes); 7952 #endif 7953 return true; 7954 } 7955 7956 // Single-threaded 7957 void CMSCollector::push_on_overflow_list(oop p) { 7958 NOT_PRODUCT(_num_par_pushes++;) 7959 assert(p->is_oop(), "Not an oop"); 7960 preserve_mark_if_necessary(p); 7961 p->set_mark((markOop)_overflow_list); 7962 _overflow_list = p; 7963 } 7964 7965 // Multi-threaded; use CAS to prepend to overflow list 7966 void CMSCollector::par_push_on_overflow_list(oop p) { 7967 NOT_PRODUCT(Atomic::inc_ptr(&_num_par_pushes);) 7968 assert(p->is_oop(), "Not an oop"); 7969 par_preserve_mark_if_necessary(p); 7970 oop observed_overflow_list = _overflow_list; 7971 oop cur_overflow_list; 7972 do { 7973 cur_overflow_list = observed_overflow_list; 7974 if (cur_overflow_list != BUSY) { 7975 p->set_mark(markOop(cur_overflow_list)); 7976 } else { 7977 p->set_mark(NULL); 7978 } 7979 observed_overflow_list = 7980 (oop) Atomic::cmpxchg_ptr(p, &_overflow_list, cur_overflow_list); 7981 } while (cur_overflow_list != observed_overflow_list); 7982 } 7983 #undef BUSY 7984 7985 // Single threaded 7986 // General Note on GrowableArray: pushes may silently fail 7987 // because we are (temporarily) out of C-heap for expanding 7988 // the stack. The problem is quite ubiquitous and affects 7989 // a lot of code in the JVM. The prudent thing for GrowableArray 7990 // to do (for now) is to exit with an error. However, that may 7991 // be too draconian in some cases because the caller may be 7992 // able to recover without much harm. For such cases, we 7993 // should probably introduce a "soft_push" method which returns 7994 // an indication of success or failure with the assumption that 7995 // the caller may be able to recover from a failure; code in 7996 // the VM can then be changed, incrementally, to deal with such 7997 // failures where possible, thus, incrementally hardening the VM 7998 // in such low resource situations. 7999 void CMSCollector::preserve_mark_work(oop p, markOop m) { 8000 _preserved_oop_stack.push(p); 8001 _preserved_mark_stack.push(m); 8002 assert(m == p->mark(), "Mark word changed"); 8003 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), 8004 "bijection"); 8005 } 8006 8007 // Single threaded 8008 void CMSCollector::preserve_mark_if_necessary(oop p) { 8009 markOop m = p->mark(); 8010 if (m->must_be_preserved(p)) { 8011 preserve_mark_work(p, m); 8012 } 8013 } 8014 8015 void CMSCollector::par_preserve_mark_if_necessary(oop p) { 8016 markOop m = p->mark(); 8017 if (m->must_be_preserved(p)) { 8018 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag); 8019 // Even though we read the mark word without holding 8020 // the lock, we are assured that it will not change 8021 // because we "own" this oop, so no other thread can 8022 // be trying to push it on the overflow list; see 8023 // the assertion in preserve_mark_work() that checks 8024 // that m == p->mark(). 8025 preserve_mark_work(p, m); 8026 } 8027 } 8028 8029 // We should be able to do this multi-threaded, 8030 // a chunk of stack being a task (this is 8031 // correct because each oop only ever appears 8032 // once in the overflow list. However, it's 8033 // not very easy to completely overlap this with 8034 // other operations, so will generally not be done 8035 // until all work's been completed. Because we 8036 // expect the preserved oop stack (set) to be small, 8037 // it's probably fine to do this single-threaded. 8038 // We can explore cleverer concurrent/overlapped/parallel 8039 // processing of preserved marks if we feel the 8040 // need for this in the future. Stack overflow should 8041 // be so rare in practice and, when it happens, its 8042 // effect on performance so great that this will 8043 // likely just be in the noise anyway. 8044 void CMSCollector::restore_preserved_marks_if_any() { 8045 assert(SafepointSynchronize::is_at_safepoint(), 8046 "world should be stopped"); 8047 assert(Thread::current()->is_ConcurrentGC_thread() || 8048 Thread::current()->is_VM_thread(), 8049 "should be single-threaded"); 8050 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), 8051 "bijection"); 8052 8053 while (!_preserved_oop_stack.is_empty()) { 8054 oop p = _preserved_oop_stack.pop(); 8055 assert(p->is_oop(), "Should be an oop"); 8056 assert(_span.contains(p), "oop should be in _span"); 8057 assert(p->mark() == markOopDesc::prototype(), 8058 "Set when taken from overflow list"); 8059 markOop m = _preserved_mark_stack.pop(); 8060 p->set_mark(m); 8061 } 8062 assert(_preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(), 8063 "stacks were cleared above"); 8064 } 8065 8066 #ifndef PRODUCT 8067 bool CMSCollector::no_preserved_marks() const { 8068 return _preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(); 8069 } 8070 #endif 8071 8072 // Transfer some number of overflown objects to usual marking 8073 // stack. Return true if some objects were transferred. 8074 bool MarkRefsIntoAndScanClosure::take_from_overflow_list() { 8075 size_t num = MIN2((size_t)(_mark_stack->capacity() - _mark_stack->length())/4, 8076 (size_t)ParGCDesiredObjsFromOverflowList); 8077 8078 bool res = _collector->take_from_overflow_list(num, _mark_stack); 8079 assert(_collector->overflow_list_is_empty() || res, 8080 "If list is not empty, we should have taken something"); 8081 assert(!res || !_mark_stack->isEmpty(), 8082 "If we took something, it should now be on our stack"); 8083 return res; 8084 } 8085 8086 size_t MarkDeadObjectsClosure::do_blk(HeapWord* addr) { 8087 size_t res = _sp->block_size_no_stall(addr, _collector); 8088 if (_sp->block_is_obj(addr)) { 8089 if (_live_bit_map->isMarked(addr)) { 8090 // It can't have been dead in a previous cycle 8091 guarantee(!_dead_bit_map->isMarked(addr), "No resurrection!"); 8092 } else { 8093 _dead_bit_map->mark(addr); // mark the dead object 8094 } 8095 } 8096 // Could be 0, if the block size could not be computed without stalling. 8097 return res; 8098 } 8099 8100 TraceCMSMemoryManagerStats::TraceCMSMemoryManagerStats(CMSCollector::CollectorState phase, GCCause::Cause cause): TraceMemoryManagerStats() { 8101 8102 switch (phase) { 8103 case CMSCollector::InitialMarking: 8104 initialize(true /* fullGC */ , 8105 cause /* cause of the GC */, 8106 true /* recordGCBeginTime */, 8107 true /* recordPreGCUsage */, 8108 false /* recordPeakUsage */, 8109 false /* recordPostGCusage */, 8110 true /* recordAccumulatedGCTime */, 8111 false /* recordGCEndTime */, 8112 false /* countCollection */ ); 8113 break; 8114 8115 case CMSCollector::FinalMarking: 8116 initialize(true /* fullGC */ , 8117 cause /* cause of the GC */, 8118 false /* recordGCBeginTime */, 8119 false /* recordPreGCUsage */, 8120 false /* recordPeakUsage */, 8121 false /* recordPostGCusage */, 8122 true /* recordAccumulatedGCTime */, 8123 false /* recordGCEndTime */, 8124 false /* countCollection */ ); 8125 break; 8126 8127 case CMSCollector::Sweeping: 8128 initialize(true /* fullGC */ , 8129 cause /* cause of the GC */, 8130 false /* recordGCBeginTime */, 8131 false /* recordPreGCUsage */, 8132 true /* recordPeakUsage */, 8133 true /* recordPostGCusage */, 8134 false /* recordAccumulatedGCTime */, 8135 true /* recordGCEndTime */, 8136 true /* countCollection */ ); 8137 break; 8138 8139 default: 8140 ShouldNotReachHere(); 8141 } 8142 }