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/symbolTable.hpp" 29 #include "classfile/systemDictionary.hpp" 30 #include "code/codeCache.hpp" 31 #include "gc/cms/cmsCollectorPolicy.hpp" 32 #include "gc/cms/cmsOopClosures.inline.hpp" 33 #include "gc/cms/compactibleFreeListSpace.hpp" 34 #include "gc/cms/concurrentMarkSweepGeneration.inline.hpp" 35 #include "gc/cms/concurrentMarkSweepThread.hpp" 36 #include "gc/cms/parNewGeneration.hpp" 37 #include "gc/cms/vmCMSOperations.hpp" 38 #include "gc/serial/genMarkSweep.hpp" 39 #include "gc/serial/tenuredGeneration.hpp" 40 #include "gc/shared/adaptiveSizePolicy.hpp" 41 #include "gc/shared/cardGeneration.inline.hpp" 42 #include "gc/shared/cardTableRS.hpp" 43 #include "gc/shared/collectedHeap.inline.hpp" 44 #include "gc/shared/collectorCounters.hpp" 45 #include "gc/shared/collectorPolicy.hpp" 46 #include "gc/shared/gcLocker.inline.hpp" 47 #include "gc/shared/gcPolicyCounters.hpp" 48 #include "gc/shared/gcTimer.hpp" 49 #include "gc/shared/gcTrace.hpp" 50 #include "gc/shared/gcTraceTime.inline.hpp" 51 #include "gc/shared/genCollectedHeap.hpp" 52 #include "gc/shared/genOopClosures.inline.hpp" 53 #include "gc/shared/isGCActiveMark.hpp" 54 #include "gc/shared/referencePolicy.hpp" 55 #include "gc/shared/strongRootsScope.hpp" 56 #include "gc/shared/taskqueue.inline.hpp" 57 #include "logging/log.hpp" 58 #include "memory/allocation.hpp" 59 #include "memory/iterator.inline.hpp" 60 #include "memory/padded.hpp" 61 #include "memory/resourceArea.hpp" 62 #include "oops/oop.inline.hpp" 63 #include "prims/jvmtiExport.hpp" 64 #include "runtime/atomic.inline.hpp" 65 #include "runtime/globals_extension.hpp" 66 #include "runtime/handles.inline.hpp" 67 #include "runtime/java.hpp" 68 #include "runtime/orderAccess.inline.hpp" 69 #include "runtime/timer.hpp" 70 #include "runtime/vmThread.hpp" 71 #include "services/memoryService.hpp" 72 #include "services/runtimeService.hpp" 73 #include "utilities/stack.inline.hpp" 74 75 // statics 76 CMSCollector* ConcurrentMarkSweepGeneration::_collector = NULL; 77 bool CMSCollector::_full_gc_requested = false; 78 GCCause::Cause CMSCollector::_full_gc_cause = GCCause::_no_gc; 79 80 ////////////////////////////////////////////////////////////////// 81 // In support of CMS/VM thread synchronization 82 ////////////////////////////////////////////////////////////////// 83 // We split use of the CGC_lock into 2 "levels". 84 // The low-level locking is of the usual CGC_lock monitor. We introduce 85 // a higher level "token" (hereafter "CMS token") built on top of the 86 // low level monitor (hereafter "CGC lock"). 87 // The token-passing protocol gives priority to the VM thread. The 88 // CMS-lock doesn't provide any fairness guarantees, but clients 89 // should ensure that it is only held for very short, bounded 90 // durations. 91 // 92 // When either of the CMS thread or the VM thread is involved in 93 // collection operations during which it does not want the other 94 // thread to interfere, it obtains the CMS token. 95 // 96 // If either thread tries to get the token while the other has 97 // it, that thread waits. However, if the VM thread and CMS thread 98 // both want the token, then the VM thread gets priority while the 99 // CMS thread waits. This ensures, for instance, that the "concurrent" 100 // phases of the CMS thread's work do not block out the VM thread 101 // for long periods of time as the CMS thread continues to hog 102 // the token. (See bug 4616232). 103 // 104 // The baton-passing functions are, however, controlled by the 105 // flags _foregroundGCShouldWait and _foregroundGCIsActive, 106 // and here the low-level CMS lock, not the high level token, 107 // ensures mutual exclusion. 108 // 109 // Two important conditions that we have to satisfy: 110 // 1. if a thread does a low-level wait on the CMS lock, then it 111 // relinquishes the CMS token if it were holding that token 112 // when it acquired the low-level CMS lock. 113 // 2. any low-level notifications on the low-level lock 114 // should only be sent when a thread has relinquished the token. 115 // 116 // In the absence of either property, we'd have potential deadlock. 117 // 118 // We protect each of the CMS (concurrent and sequential) phases 119 // with the CMS _token_, not the CMS _lock_. 120 // 121 // The only code protected by CMS lock is the token acquisition code 122 // itself, see ConcurrentMarkSweepThread::[de]synchronize(), and the 123 // baton-passing code. 124 // 125 // Unfortunately, i couldn't come up with a good abstraction to factor and 126 // hide the naked CGC_lock manipulation in the baton-passing code 127 // further below. That's something we should try to do. Also, the proof 128 // of correctness of this 2-level locking scheme is far from obvious, 129 // and potentially quite slippery. We have an uneasy suspicion, for instance, 130 // that there may be a theoretical possibility of delay/starvation in the 131 // low-level lock/wait/notify scheme used for the baton-passing because of 132 // potential interference with the priority scheme embodied in the 133 // CMS-token-passing protocol. See related comments at a CGC_lock->wait() 134 // invocation further below and marked with "XXX 20011219YSR". 135 // Indeed, as we note elsewhere, this may become yet more slippery 136 // in the presence of multiple CMS and/or multiple VM threads. XXX 137 138 class CMSTokenSync: public StackObj { 139 private: 140 bool _is_cms_thread; 141 public: 142 CMSTokenSync(bool is_cms_thread): 143 _is_cms_thread(is_cms_thread) { 144 assert(is_cms_thread == Thread::current()->is_ConcurrentGC_thread(), 145 "Incorrect argument to constructor"); 146 ConcurrentMarkSweepThread::synchronize(_is_cms_thread); 147 } 148 149 ~CMSTokenSync() { 150 assert(_is_cms_thread ? 151 ConcurrentMarkSweepThread::cms_thread_has_cms_token() : 152 ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 153 "Incorrect state"); 154 ConcurrentMarkSweepThread::desynchronize(_is_cms_thread); 155 } 156 }; 157 158 // Convenience class that does a CMSTokenSync, and then acquires 159 // upto three locks. 160 class CMSTokenSyncWithLocks: public CMSTokenSync { 161 private: 162 // Note: locks are acquired in textual declaration order 163 // and released in the opposite order 164 MutexLockerEx _locker1, _locker2, _locker3; 165 public: 166 CMSTokenSyncWithLocks(bool is_cms_thread, Mutex* mutex1, 167 Mutex* mutex2 = NULL, Mutex* mutex3 = NULL): 168 CMSTokenSync(is_cms_thread), 169 _locker1(mutex1, Mutex::_no_safepoint_check_flag), 170 _locker2(mutex2, Mutex::_no_safepoint_check_flag), 171 _locker3(mutex3, Mutex::_no_safepoint_check_flag) 172 { } 173 }; 174 175 176 ////////////////////////////////////////////////////////////////// 177 // Concurrent Mark-Sweep Generation ///////////////////////////// 178 ////////////////////////////////////////////////////////////////// 179 180 NOT_PRODUCT(CompactibleFreeListSpace* debug_cms_space;) 181 182 // This struct contains per-thread things necessary to support parallel 183 // young-gen collection. 184 class CMSParGCThreadState: public CHeapObj<mtGC> { 185 public: 186 CFLS_LAB lab; 187 PromotionInfo promo; 188 189 // Constructor. 190 CMSParGCThreadState(CompactibleFreeListSpace* cfls) : lab(cfls) { 191 promo.setSpace(cfls); 192 } 193 }; 194 195 ConcurrentMarkSweepGeneration::ConcurrentMarkSweepGeneration( 196 ReservedSpace rs, size_t initial_byte_size, CardTableRS* ct) : 197 CardGeneration(rs, initial_byte_size, ct), 198 _dilatation_factor(((double)MinChunkSize)/((double)(CollectedHeap::min_fill_size()))), 199 _did_compact(false) 200 { 201 HeapWord* bottom = (HeapWord*) _virtual_space.low(); 202 HeapWord* end = (HeapWord*) _virtual_space.high(); 203 204 _direct_allocated_words = 0; 205 NOT_PRODUCT( 206 _numObjectsPromoted = 0; 207 _numWordsPromoted = 0; 208 _numObjectsAllocated = 0; 209 _numWordsAllocated = 0; 210 ) 211 212 _cmsSpace = new CompactibleFreeListSpace(_bts, MemRegion(bottom, end)); 213 NOT_PRODUCT(debug_cms_space = _cmsSpace;) 214 _cmsSpace->_old_gen = this; 215 216 _gc_stats = new CMSGCStats(); 217 218 // Verify the assumption that FreeChunk::_prev and OopDesc::_klass 219 // offsets match. The ability to tell free chunks from objects 220 // depends on this property. 221 debug_only( 222 FreeChunk* junk = NULL; 223 assert(UseCompressedClassPointers || 224 junk->prev_addr() == (void*)(oop(junk)->klass_addr()), 225 "Offset of FreeChunk::_prev within FreeChunk must match" 226 " that of OopDesc::_klass within OopDesc"); 227 ) 228 229 _par_gc_thread_states = NEW_C_HEAP_ARRAY(CMSParGCThreadState*, ParallelGCThreads, mtGC); 230 for (uint i = 0; i < ParallelGCThreads; i++) { 231 _par_gc_thread_states[i] = new CMSParGCThreadState(cmsSpace()); 232 } 233 234 _incremental_collection_failed = false; 235 // The "dilatation_factor" is the expansion that can occur on 236 // account of the fact that the minimum object size in the CMS 237 // generation may be larger than that in, say, a contiguous young 238 // generation. 239 // Ideally, in the calculation below, we'd compute the dilatation 240 // factor as: MinChunkSize/(promoting_gen's min object size) 241 // Since we do not have such a general query interface for the 242 // promoting generation, we'll instead just use the minimum 243 // object size (which today is a header's worth of space); 244 // note that all arithmetic is in units of HeapWords. 245 assert(MinChunkSize >= CollectedHeap::min_fill_size(), "just checking"); 246 assert(_dilatation_factor >= 1.0, "from previous assert"); 247 } 248 249 250 // The field "_initiating_occupancy" represents the occupancy percentage 251 // at which we trigger a new collection cycle. Unless explicitly specified 252 // via CMSInitiatingOccupancyFraction (argument "io" below), it 253 // is calculated by: 254 // 255 // Let "f" be MinHeapFreeRatio in 256 // 257 // _initiating_occupancy = 100-f + 258 // f * (CMSTriggerRatio/100) 259 // where CMSTriggerRatio is the argument "tr" below. 260 // 261 // That is, if we assume the heap is at its desired maximum occupancy at the 262 // end of a collection, we let CMSTriggerRatio of the (purported) free 263 // space be allocated before initiating a new collection cycle. 264 // 265 void ConcurrentMarkSweepGeneration::init_initiating_occupancy(intx io, uintx tr) { 266 assert(io <= 100 && tr <= 100, "Check the arguments"); 267 if (io >= 0) { 268 _initiating_occupancy = (double)io / 100.0; 269 } else { 270 _initiating_occupancy = ((100 - MinHeapFreeRatio) + 271 (double)(tr * MinHeapFreeRatio) / 100.0) 272 / 100.0; 273 } 274 } 275 276 void ConcurrentMarkSweepGeneration::ref_processor_init() { 277 assert(collector() != NULL, "no collector"); 278 collector()->ref_processor_init(); 279 } 280 281 void CMSCollector::ref_processor_init() { 282 if (_ref_processor == NULL) { 283 // Allocate and initialize a reference processor 284 _ref_processor = 285 new ReferenceProcessor(_span, // span 286 (ParallelGCThreads > 1) && ParallelRefProcEnabled, // mt processing 287 ParallelGCThreads, // mt processing degree 288 _cmsGen->refs_discovery_is_mt(), // mt discovery 289 MAX2(ConcGCThreads, ParallelGCThreads), // mt discovery degree 290 _cmsGen->refs_discovery_is_atomic(), // discovery is not atomic 291 &_is_alive_closure); // closure for liveness info 292 // Initialize the _ref_processor field of CMSGen 293 _cmsGen->set_ref_processor(_ref_processor); 294 295 } 296 } 297 298 AdaptiveSizePolicy* CMSCollector::size_policy() { 299 GenCollectedHeap* gch = GenCollectedHeap::heap(); 300 return gch->gen_policy()->size_policy(); 301 } 302 303 void ConcurrentMarkSweepGeneration::initialize_performance_counters() { 304 305 const char* gen_name = "old"; 306 GenCollectorPolicy* gcp = GenCollectedHeap::heap()->gen_policy(); 307 // Generation Counters - generation 1, 1 subspace 308 _gen_counters = new GenerationCounters(gen_name, 1, 1, 309 gcp->min_old_size(), gcp->max_old_size(), &_virtual_space); 310 311 _space_counters = new GSpaceCounters(gen_name, 0, 312 _virtual_space.reserved_size(), 313 this, _gen_counters); 314 } 315 316 CMSStats::CMSStats(ConcurrentMarkSweepGeneration* cms_gen, unsigned int alpha): 317 _cms_gen(cms_gen) 318 { 319 assert(alpha <= 100, "bad value"); 320 _saved_alpha = alpha; 321 322 // Initialize the alphas to the bootstrap value of 100. 323 _gc0_alpha = _cms_alpha = 100; 324 325 _cms_begin_time.update(); 326 _cms_end_time.update(); 327 328 _gc0_duration = 0.0; 329 _gc0_period = 0.0; 330 _gc0_promoted = 0; 331 332 _cms_duration = 0.0; 333 _cms_period = 0.0; 334 _cms_allocated = 0; 335 336 _cms_used_at_gc0_begin = 0; 337 _cms_used_at_gc0_end = 0; 338 _allow_duty_cycle_reduction = false; 339 _valid_bits = 0; 340 } 341 342 double CMSStats::cms_free_adjustment_factor(size_t free) const { 343 // TBD: CR 6909490 344 return 1.0; 345 } 346 347 void CMSStats::adjust_cms_free_adjustment_factor(bool fail, size_t free) { 348 } 349 350 // If promotion failure handling is on use 351 // the padded average size of the promotion for each 352 // young generation collection. 353 double CMSStats::time_until_cms_gen_full() const { 354 size_t cms_free = _cms_gen->cmsSpace()->free(); 355 GenCollectedHeap* gch = GenCollectedHeap::heap(); 356 size_t expected_promotion = MIN2(gch->young_gen()->capacity(), 357 (size_t) _cms_gen->gc_stats()->avg_promoted()->padded_average()); 358 if (cms_free > expected_promotion) { 359 // Start a cms collection if there isn't enough space to promote 360 // for the next young collection. Use the padded average as 361 // a safety factor. 362 cms_free -= expected_promotion; 363 364 // Adjust by the safety factor. 365 double cms_free_dbl = (double)cms_free; 366 double cms_adjustment = (100.0 - CMSIncrementalSafetyFactor) / 100.0; 367 // Apply a further correction factor which tries to adjust 368 // for recent occurance of concurrent mode failures. 369 cms_adjustment = cms_adjustment * cms_free_adjustment_factor(cms_free); 370 cms_free_dbl = cms_free_dbl * cms_adjustment; 371 372 log_trace(gc)("CMSStats::time_until_cms_gen_full: cms_free " SIZE_FORMAT " expected_promotion " SIZE_FORMAT, 373 cms_free, expected_promotion); 374 log_trace(gc)(" cms_free_dbl %f cms_consumption_rate %f", cms_free_dbl, cms_consumption_rate() + 1.0); 375 // Add 1 in case the consumption rate goes to zero. 376 return cms_free_dbl / (cms_consumption_rate() + 1.0); 377 } 378 return 0.0; 379 } 380 381 // Compare the duration of the cms collection to the 382 // time remaining before the cms generation is empty. 383 // Note that the time from the start of the cms collection 384 // to the start of the cms sweep (less than the total 385 // duration of the cms collection) can be used. This 386 // has been tried and some applications experienced 387 // promotion failures early in execution. This was 388 // possibly because the averages were not accurate 389 // enough at the beginning. 390 double CMSStats::time_until_cms_start() const { 391 // We add "gc0_period" to the "work" calculation 392 // below because this query is done (mostly) at the 393 // end of a scavenge, so we need to conservatively 394 // account for that much possible delay 395 // in the query so as to avoid concurrent mode failures 396 // due to starting the collection just a wee bit too 397 // late. 398 double work = cms_duration() + gc0_period(); 399 double deadline = time_until_cms_gen_full(); 400 // If a concurrent mode failure occurred recently, we want to be 401 // more conservative and halve our expected time_until_cms_gen_full() 402 if (work > deadline) { 403 log_develop_trace(gc)("CMSCollector: collect because of anticipated promotion before full %3.7f + %3.7f > %3.7f ", 404 cms_duration(), gc0_period(), time_until_cms_gen_full()); 405 return 0.0; 406 } 407 return work - deadline; 408 } 409 410 #ifndef PRODUCT 411 void CMSStats::print_on(outputStream *st) const { 412 st->print(" gc0_alpha=%d,cms_alpha=%d", _gc0_alpha, _cms_alpha); 413 st->print(",gc0_dur=%g,gc0_per=%g,gc0_promo=" SIZE_FORMAT, 414 gc0_duration(), gc0_period(), gc0_promoted()); 415 st->print(",cms_dur=%g,cms_per=%g,cms_alloc=" SIZE_FORMAT, 416 cms_duration(), cms_period(), cms_allocated()); 417 st->print(",cms_since_beg=%g,cms_since_end=%g", 418 cms_time_since_begin(), cms_time_since_end()); 419 st->print(",cms_used_beg=" SIZE_FORMAT ",cms_used_end=" SIZE_FORMAT, 420 _cms_used_at_gc0_begin, _cms_used_at_gc0_end); 421 422 if (valid()) { 423 st->print(",promo_rate=%g,cms_alloc_rate=%g", 424 promotion_rate(), cms_allocation_rate()); 425 st->print(",cms_consumption_rate=%g,time_until_full=%g", 426 cms_consumption_rate(), time_until_cms_gen_full()); 427 } 428 st->print(" "); 429 } 430 #endif // #ifndef PRODUCT 431 432 CMSCollector::CollectorState CMSCollector::_collectorState = 433 CMSCollector::Idling; 434 bool CMSCollector::_foregroundGCIsActive = false; 435 bool CMSCollector::_foregroundGCShouldWait = false; 436 437 CMSCollector::CMSCollector(ConcurrentMarkSweepGeneration* cmsGen, 438 CardTableRS* ct, 439 ConcurrentMarkSweepPolicy* cp): 440 _cmsGen(cmsGen), 441 _ct(ct), 442 _ref_processor(NULL), // will be set later 443 _conc_workers(NULL), // may be set later 444 _abort_preclean(false), 445 _start_sampling(false), 446 _between_prologue_and_epilogue(false), 447 _markBitMap(0, Mutex::leaf + 1, "CMS_markBitMap_lock"), 448 _modUnionTable((CardTableModRefBS::card_shift - LogHeapWordSize), 449 -1 /* lock-free */, "No_lock" /* dummy */), 450 _modUnionClosurePar(&_modUnionTable), 451 // Adjust my span to cover old (cms) gen 452 _span(cmsGen->reserved()), 453 // Construct the is_alive_closure with _span & markBitMap 454 _is_alive_closure(_span, &_markBitMap), 455 _restart_addr(NULL), 456 _overflow_list(NULL), 457 _stats(cmsGen), 458 _eden_chunk_lock(new Mutex(Mutex::leaf + 1, "CMS_eden_chunk_lock", true, 459 //verify that this lock should be acquired with safepoint check. 460 Monitor::_safepoint_check_sometimes)), 461 _eden_chunk_array(NULL), // may be set in ctor body 462 _eden_chunk_capacity(0), // -- ditto -- 463 _eden_chunk_index(0), // -- ditto -- 464 _survivor_plab_array(NULL), // -- ditto -- 465 _survivor_chunk_array(NULL), // -- ditto -- 466 _survivor_chunk_capacity(0), // -- ditto -- 467 _survivor_chunk_index(0), // -- ditto -- 468 _ser_pmc_preclean_ovflw(0), 469 _ser_kac_preclean_ovflw(0), 470 _ser_pmc_remark_ovflw(0), 471 _par_pmc_remark_ovflw(0), 472 _ser_kac_ovflw(0), 473 _par_kac_ovflw(0), 474 #ifndef PRODUCT 475 _num_par_pushes(0), 476 #endif 477 _collection_count_start(0), 478 _verifying(false), 479 _verification_mark_bm(0, Mutex::leaf + 1, "CMS_verification_mark_bm_lock"), 480 _completed_initialization(false), 481 _collector_policy(cp), 482 _should_unload_classes(CMSClassUnloadingEnabled), 483 _concurrent_cycles_since_last_unload(0), 484 _roots_scanning_options(GenCollectedHeap::SO_None), 485 _inter_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding), 486 _intra_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding), 487 _gc_tracer_cm(new (ResourceObj::C_HEAP, mtGC) CMSTracer()), 488 _gc_timer_cm(new (ResourceObj::C_HEAP, mtGC) ConcurrentGCTimer()), 489 _cms_start_registered(false) 490 { 491 if (ExplicitGCInvokesConcurrentAndUnloadsClasses) { 492 ExplicitGCInvokesConcurrent = true; 493 } 494 // Now expand the span and allocate the collection support structures 495 // (MUT, marking bit map etc.) to cover both generations subject to 496 // collection. 497 498 // For use by dirty card to oop closures. 499 _cmsGen->cmsSpace()->set_collector(this); 500 501 // Allocate MUT and marking bit map 502 { 503 MutexLockerEx x(_markBitMap.lock(), Mutex::_no_safepoint_check_flag); 504 if (!_markBitMap.allocate(_span)) { 505 warning("Failed to allocate CMS Bit Map"); 506 return; 507 } 508 assert(_markBitMap.covers(_span), "_markBitMap inconsistency?"); 509 } 510 { 511 _modUnionTable.allocate(_span); 512 assert(_modUnionTable.covers(_span), "_modUnionTable inconsistency?"); 513 } 514 515 if (!_markStack.allocate(MarkStackSize)) { 516 warning("Failed to allocate CMS Marking Stack"); 517 return; 518 } 519 520 // Support for multi-threaded concurrent phases 521 if (CMSConcurrentMTEnabled) { 522 if (FLAG_IS_DEFAULT(ConcGCThreads)) { 523 // just for now 524 FLAG_SET_DEFAULT(ConcGCThreads, (ParallelGCThreads + 3) / 4); 525 } 526 if (ConcGCThreads > 1) { 527 _conc_workers = new YieldingFlexibleWorkGang("CMS Thread", 528 ConcGCThreads, true); 529 if (_conc_workers == NULL) { 530 warning("GC/CMS: _conc_workers allocation failure: " 531 "forcing -CMSConcurrentMTEnabled"); 532 CMSConcurrentMTEnabled = false; 533 } else { 534 _conc_workers->initialize_workers(); 535 } 536 } else { 537 CMSConcurrentMTEnabled = false; 538 } 539 } 540 if (!CMSConcurrentMTEnabled) { 541 ConcGCThreads = 0; 542 } else { 543 // Turn off CMSCleanOnEnter optimization temporarily for 544 // the MT case where it's not fixed yet; see 6178663. 545 CMSCleanOnEnter = false; 546 } 547 assert((_conc_workers != NULL) == (ConcGCThreads > 1), 548 "Inconsistency"); 549 550 // Parallel task queues; these are shared for the 551 // concurrent and stop-world phases of CMS, but 552 // are not shared with parallel scavenge (ParNew). 553 { 554 uint i; 555 uint num_queues = MAX2(ParallelGCThreads, ConcGCThreads); 556 557 if ((CMSParallelRemarkEnabled || CMSConcurrentMTEnabled 558 || ParallelRefProcEnabled) 559 && num_queues > 0) { 560 _task_queues = new OopTaskQueueSet(num_queues); 561 if (_task_queues == NULL) { 562 warning("task_queues allocation failure."); 563 return; 564 } 565 _hash_seed = NEW_C_HEAP_ARRAY(int, num_queues, mtGC); 566 typedef Padded<OopTaskQueue> PaddedOopTaskQueue; 567 for (i = 0; i < num_queues; i++) { 568 PaddedOopTaskQueue *q = new PaddedOopTaskQueue(); 569 if (q == NULL) { 570 warning("work_queue allocation failure."); 571 return; 572 } 573 _task_queues->register_queue(i, q); 574 } 575 for (i = 0; i < num_queues; i++) { 576 _task_queues->queue(i)->initialize(); 577 _hash_seed[i] = 17; // copied from ParNew 578 } 579 } 580 } 581 582 _cmsGen ->init_initiating_occupancy(CMSInitiatingOccupancyFraction, CMSTriggerRatio); 583 584 // Clip CMSBootstrapOccupancy between 0 and 100. 585 _bootstrap_occupancy = CMSBootstrapOccupancy / 100.0; 586 587 // Now tell CMS generations the identity of their collector 588 ConcurrentMarkSweepGeneration::set_collector(this); 589 590 // Create & start a CMS thread for this CMS collector 591 _cmsThread = ConcurrentMarkSweepThread::start(this); 592 assert(cmsThread() != NULL, "CMS Thread should have been created"); 593 assert(cmsThread()->collector() == this, 594 "CMS Thread should refer to this gen"); 595 assert(CGC_lock != NULL, "Where's the CGC_lock?"); 596 597 // Support for parallelizing young gen rescan 598 GenCollectedHeap* gch = GenCollectedHeap::heap(); 599 assert(gch->young_gen()->kind() == Generation::ParNew, "CMS can only be used with ParNew"); 600 _young_gen = (ParNewGeneration*)gch->young_gen(); 601 if (gch->supports_inline_contig_alloc()) { 602 _top_addr = gch->top_addr(); 603 _end_addr = gch->end_addr(); 604 assert(_young_gen != NULL, "no _young_gen"); 605 _eden_chunk_index = 0; 606 _eden_chunk_capacity = (_young_gen->max_capacity() + CMSSamplingGrain) / CMSSamplingGrain; 607 _eden_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, _eden_chunk_capacity, mtGC); 608 } 609 610 // Support for parallelizing survivor space rescan 611 if ((CMSParallelRemarkEnabled && CMSParallelSurvivorRemarkEnabled) || CMSParallelInitialMarkEnabled) { 612 const size_t max_plab_samples = 613 _young_gen->max_survivor_size() / (PLAB::min_size() * HeapWordSize); 614 615 _survivor_plab_array = NEW_C_HEAP_ARRAY(ChunkArray, ParallelGCThreads, mtGC); 616 _survivor_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples, mtGC); 617 _cursor = NEW_C_HEAP_ARRAY(size_t, ParallelGCThreads, mtGC); 618 _survivor_chunk_capacity = max_plab_samples; 619 for (uint i = 0; i < ParallelGCThreads; i++) { 620 HeapWord** vec = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples, mtGC); 621 ChunkArray* cur = ::new (&_survivor_plab_array[i]) ChunkArray(vec, max_plab_samples); 622 assert(cur->end() == 0, "Should be 0"); 623 assert(cur->array() == vec, "Should be vec"); 624 assert(cur->capacity() == max_plab_samples, "Error"); 625 } 626 } 627 628 NOT_PRODUCT(_overflow_counter = CMSMarkStackOverflowInterval;) 629 _gc_counters = new CollectorCounters("CMS", 1); 630 _completed_initialization = true; 631 _inter_sweep_timer.start(); // start of time 632 } 633 634 const char* ConcurrentMarkSweepGeneration::name() const { 635 return "concurrent mark-sweep generation"; 636 } 637 void ConcurrentMarkSweepGeneration::update_counters() { 638 if (UsePerfData) { 639 _space_counters->update_all(); 640 _gen_counters->update_all(); 641 } 642 } 643 644 // this is an optimized version of update_counters(). it takes the 645 // used value as a parameter rather than computing it. 646 // 647 void ConcurrentMarkSweepGeneration::update_counters(size_t used) { 648 if (UsePerfData) { 649 _space_counters->update_used(used); 650 _space_counters->update_capacity(); 651 _gen_counters->update_all(); 652 } 653 } 654 655 void ConcurrentMarkSweepGeneration::print() const { 656 Generation::print(); 657 cmsSpace()->print(); 658 } 659 660 #ifndef PRODUCT 661 void ConcurrentMarkSweepGeneration::print_statistics() { 662 cmsSpace()->printFLCensus(0); 663 } 664 #endif 665 666 size_t 667 ConcurrentMarkSweepGeneration::contiguous_available() const { 668 // dld proposes an improvement in precision here. If the committed 669 // part of the space ends in a free block we should add that to 670 // uncommitted size in the calculation below. Will make this 671 // change later, staying with the approximation below for the 672 // time being. -- ysr. 673 return MAX2(_virtual_space.uncommitted_size(), unsafe_max_alloc_nogc()); 674 } 675 676 size_t 677 ConcurrentMarkSweepGeneration::unsafe_max_alloc_nogc() const { 678 return _cmsSpace->max_alloc_in_words() * HeapWordSize; 679 } 680 681 size_t ConcurrentMarkSweepGeneration::max_available() const { 682 return free() + _virtual_space.uncommitted_size(); 683 } 684 685 bool ConcurrentMarkSweepGeneration::promotion_attempt_is_safe(size_t max_promotion_in_bytes) const { 686 size_t available = max_available(); 687 size_t av_promo = (size_t)gc_stats()->avg_promoted()->padded_average(); 688 bool res = (available >= av_promo) || (available >= max_promotion_in_bytes); 689 log_trace(gc, promotion)("CMS: promo attempt is%s safe: available(" SIZE_FORMAT ") %s av_promo(" SIZE_FORMAT "), max_promo(" SIZE_FORMAT ")", 690 res? "":" not", available, res? ">=":"<", av_promo, max_promotion_in_bytes); 691 return res; 692 } 693 694 // At a promotion failure dump information on block layout in heap 695 // (cms old generation). 696 void ConcurrentMarkSweepGeneration::promotion_failure_occurred() { 697 LogHandle(gc, promotion) log; 698 if (log.is_trace()) { 699 ResourceMark rm; 700 cmsSpace()->dump_at_safepoint_with_locks(collector(), log.trace_stream()); 701 } 702 } 703 704 void ConcurrentMarkSweepGeneration::reset_after_compaction() { 705 // Clear the promotion information. These pointers can be adjusted 706 // along with all the other pointers into the heap but 707 // compaction is expected to be a rare event with 708 // a heap using cms so don't do it without seeing the need. 709 for (uint i = 0; i < ParallelGCThreads; i++) { 710 _par_gc_thread_states[i]->promo.reset(); 711 } 712 } 713 714 void ConcurrentMarkSweepGeneration::compute_new_size() { 715 assert_locked_or_safepoint(Heap_lock); 716 717 // If incremental collection failed, we just want to expand 718 // to the limit. 719 if (incremental_collection_failed()) { 720 clear_incremental_collection_failed(); 721 grow_to_reserved(); 722 return; 723 } 724 725 // The heap has been compacted but not reset yet. 726 // Any metric such as free() or used() will be incorrect. 727 728 CardGeneration::compute_new_size(); 729 730 // Reset again after a possible resizing 731 if (did_compact()) { 732 cmsSpace()->reset_after_compaction(); 733 } 734 } 735 736 void ConcurrentMarkSweepGeneration::compute_new_size_free_list() { 737 assert_locked_or_safepoint(Heap_lock); 738 739 // If incremental collection failed, we just want to expand 740 // to the limit. 741 if (incremental_collection_failed()) { 742 clear_incremental_collection_failed(); 743 grow_to_reserved(); 744 return; 745 } 746 747 double free_percentage = ((double) free()) / capacity(); 748 double desired_free_percentage = (double) MinHeapFreeRatio / 100; 749 double maximum_free_percentage = (double) MaxHeapFreeRatio / 100; 750 751 // compute expansion delta needed for reaching desired free percentage 752 if (free_percentage < desired_free_percentage) { 753 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 754 assert(desired_capacity >= capacity(), "invalid expansion size"); 755 size_t expand_bytes = MAX2(desired_capacity - capacity(), MinHeapDeltaBytes); 756 LogHandle(gc) log; 757 if (log.is_trace()) { 758 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 759 log.trace("From compute_new_size: "); 760 log.trace(" Free fraction %f", free_percentage); 761 log.trace(" Desired free fraction %f", desired_free_percentage); 762 log.trace(" Maximum free fraction %f", maximum_free_percentage); 763 log.trace(" Capacity " SIZE_FORMAT, capacity() / 1000); 764 log.trace(" Desired capacity " SIZE_FORMAT, desired_capacity / 1000); 765 GenCollectedHeap* gch = GenCollectedHeap::heap(); 766 assert(gch->is_old_gen(this), "The CMS generation should always be the old generation"); 767 size_t young_size = gch->young_gen()->capacity(); 768 log.trace(" Young gen size " SIZE_FORMAT, young_size / 1000); 769 log.trace(" unsafe_max_alloc_nogc " SIZE_FORMAT, unsafe_max_alloc_nogc() / 1000); 770 log.trace(" contiguous available " SIZE_FORMAT, contiguous_available() / 1000); 771 log.trace(" Expand by " SIZE_FORMAT " (bytes)", expand_bytes); 772 } 773 // safe if expansion fails 774 expand_for_gc_cause(expand_bytes, 0, CMSExpansionCause::_satisfy_free_ratio); 775 log.trace(" Expanded free fraction %f", ((double) free()) / capacity()); 776 } else { 777 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 778 assert(desired_capacity <= capacity(), "invalid expansion size"); 779 size_t shrink_bytes = capacity() - desired_capacity; 780 // Don't shrink unless the delta is greater than the minimum shrink we want 781 if (shrink_bytes >= MinHeapDeltaBytes) { 782 shrink_free_list_by(shrink_bytes); 783 } 784 } 785 } 786 787 Mutex* ConcurrentMarkSweepGeneration::freelistLock() const { 788 return cmsSpace()->freelistLock(); 789 } 790 791 HeapWord* ConcurrentMarkSweepGeneration::allocate(size_t size, bool tlab) { 792 CMSSynchronousYieldRequest yr; 793 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 794 return have_lock_and_allocate(size, tlab); 795 } 796 797 HeapWord* ConcurrentMarkSweepGeneration::have_lock_and_allocate(size_t size, 798 bool tlab /* ignored */) { 799 assert_lock_strong(freelistLock()); 800 size_t adjustedSize = CompactibleFreeListSpace::adjustObjectSize(size); 801 HeapWord* res = cmsSpace()->allocate(adjustedSize); 802 // Allocate the object live (grey) if the background collector has 803 // started marking. This is necessary because the marker may 804 // have passed this address and consequently this object will 805 // not otherwise be greyed and would be incorrectly swept up. 806 // Note that if this object contains references, the writing 807 // of those references will dirty the card containing this object 808 // allowing the object to be blackened (and its references scanned) 809 // either during a preclean phase or at the final checkpoint. 810 if (res != NULL) { 811 // We may block here with an uninitialized object with 812 // its mark-bit or P-bits not yet set. Such objects need 813 // to be safely navigable by block_start(). 814 assert(oop(res)->klass_or_null() == NULL, "Object should be uninitialized here."); 815 assert(!((FreeChunk*)res)->is_free(), "Error, block will look free but show wrong size"); 816 collector()->direct_allocated(res, adjustedSize); 817 _direct_allocated_words += adjustedSize; 818 // allocation counters 819 NOT_PRODUCT( 820 _numObjectsAllocated++; 821 _numWordsAllocated += (int)adjustedSize; 822 ) 823 } 824 return res; 825 } 826 827 // In the case of direct allocation by mutators in a generation that 828 // is being concurrently collected, the object must be allocated 829 // live (grey) if the background collector has started marking. 830 // This is necessary because the marker may 831 // have passed this address and consequently this object will 832 // not otherwise be greyed and would be incorrectly swept up. 833 // Note that if this object contains references, the writing 834 // of those references will dirty the card containing this object 835 // allowing the object to be blackened (and its references scanned) 836 // either during a preclean phase or at the final checkpoint. 837 void CMSCollector::direct_allocated(HeapWord* start, size_t size) { 838 assert(_markBitMap.covers(start, size), "Out of bounds"); 839 if (_collectorState >= Marking) { 840 MutexLockerEx y(_markBitMap.lock(), 841 Mutex::_no_safepoint_check_flag); 842 // [see comments preceding SweepClosure::do_blk() below for details] 843 // 844 // Can the P-bits be deleted now? JJJ 845 // 846 // 1. need to mark the object as live so it isn't collected 847 // 2. need to mark the 2nd bit to indicate the object may be uninitialized 848 // 3. need to mark the end of the object so marking, precleaning or sweeping 849 // can skip over uninitialized or unparsable objects. An allocated 850 // object is considered uninitialized for our purposes as long as 851 // its klass word is NULL. All old gen objects are parsable 852 // as soon as they are initialized.) 853 _markBitMap.mark(start); // object is live 854 _markBitMap.mark(start + 1); // object is potentially uninitialized? 855 _markBitMap.mark(start + size - 1); 856 // mark end of object 857 } 858 // check that oop looks uninitialized 859 assert(oop(start)->klass_or_null() == NULL, "_klass should be NULL"); 860 } 861 862 void CMSCollector::promoted(bool par, HeapWord* start, 863 bool is_obj_array, size_t obj_size) { 864 assert(_markBitMap.covers(start), "Out of bounds"); 865 // See comment in direct_allocated() about when objects should 866 // be allocated live. 867 if (_collectorState >= Marking) { 868 // we already hold the marking bit map lock, taken in 869 // the prologue 870 if (par) { 871 _markBitMap.par_mark(start); 872 } else { 873 _markBitMap.mark(start); 874 } 875 // We don't need to mark the object as uninitialized (as 876 // in direct_allocated above) because this is being done with the 877 // world stopped and the object will be initialized by the 878 // time the marking, precleaning or sweeping get to look at it. 879 // But see the code for copying objects into the CMS generation, 880 // where we need to ensure that concurrent readers of the 881 // block offset table are able to safely navigate a block that 882 // is in flux from being free to being allocated (and in 883 // transition while being copied into) and subsequently 884 // becoming a bona-fide object when the copy/promotion is complete. 885 assert(SafepointSynchronize::is_at_safepoint(), 886 "expect promotion only at safepoints"); 887 888 if (_collectorState < Sweeping) { 889 // Mark the appropriate cards in the modUnionTable, so that 890 // this object gets scanned before the sweep. If this is 891 // not done, CMS generation references in the object might 892 // not get marked. 893 // For the case of arrays, which are otherwise precisely 894 // marked, we need to dirty the entire array, not just its head. 895 if (is_obj_array) { 896 // The [par_]mark_range() method expects mr.end() below to 897 // be aligned to the granularity of a bit's representation 898 // in the heap. In the case of the MUT below, that's a 899 // card size. 900 MemRegion mr(start, 901 (HeapWord*)round_to((intptr_t)(start + obj_size), 902 CardTableModRefBS::card_size /* bytes */)); 903 if (par) { 904 _modUnionTable.par_mark_range(mr); 905 } else { 906 _modUnionTable.mark_range(mr); 907 } 908 } else { // not an obj array; we can just mark the head 909 if (par) { 910 _modUnionTable.par_mark(start); 911 } else { 912 _modUnionTable.mark(start); 913 } 914 } 915 } 916 } 917 } 918 919 oop ConcurrentMarkSweepGeneration::promote(oop obj, size_t obj_size) { 920 assert(obj_size == (size_t)obj->size(), "bad obj_size passed in"); 921 // allocate, copy and if necessary update promoinfo -- 922 // delegate to underlying space. 923 assert_lock_strong(freelistLock()); 924 925 #ifndef PRODUCT 926 if (GenCollectedHeap::heap()->promotion_should_fail()) { 927 return NULL; 928 } 929 #endif // #ifndef PRODUCT 930 931 oop res = _cmsSpace->promote(obj, obj_size); 932 if (res == NULL) { 933 // expand and retry 934 size_t s = _cmsSpace->expansionSpaceRequired(obj_size); // HeapWords 935 expand_for_gc_cause(s*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_satisfy_promotion); 936 // Since this is the old generation, we don't try to promote 937 // into a more senior generation. 938 res = _cmsSpace->promote(obj, obj_size); 939 } 940 if (res != NULL) { 941 // See comment in allocate() about when objects should 942 // be allocated live. 943 assert(obj->is_oop(), "Will dereference klass pointer below"); 944 collector()->promoted(false, // Not parallel 945 (HeapWord*)res, obj->is_objArray(), obj_size); 946 // promotion counters 947 NOT_PRODUCT( 948 _numObjectsPromoted++; 949 _numWordsPromoted += 950 (int)(CompactibleFreeListSpace::adjustObjectSize(obj->size())); 951 ) 952 } 953 return res; 954 } 955 956 957 // IMPORTANT: Notes on object size recognition in CMS. 958 // --------------------------------------------------- 959 // A block of storage in the CMS generation is always in 960 // one of three states. A free block (FREE), an allocated 961 // object (OBJECT) whose size() method reports the correct size, 962 // and an intermediate state (TRANSIENT) in which its size cannot 963 // be accurately determined. 964 // STATE IDENTIFICATION: (32 bit and 64 bit w/o COOPS) 965 // ----------------------------------------------------- 966 // FREE: klass_word & 1 == 1; mark_word holds block size 967 // 968 // OBJECT: klass_word installed; klass_word != 0 && klass_word & 1 == 0; 969 // obj->size() computes correct size 970 // 971 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT 972 // 973 // STATE IDENTIFICATION: (64 bit+COOPS) 974 // ------------------------------------ 975 // FREE: mark_word & CMS_FREE_BIT == 1; mark_word & ~CMS_FREE_BIT gives block_size 976 // 977 // OBJECT: klass_word installed; klass_word != 0; 978 // obj->size() computes correct size 979 // 980 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT 981 // 982 // 983 // STATE TRANSITION DIAGRAM 984 // 985 // mut / parnew mut / parnew 986 // FREE --------------------> TRANSIENT ---------------------> OBJECT --| 987 // ^ | 988 // |------------------------ DEAD <------------------------------------| 989 // sweep mut 990 // 991 // While a block is in TRANSIENT state its size cannot be determined 992 // so readers will either need to come back later or stall until 993 // the size can be determined. Note that for the case of direct 994 // allocation, P-bits, when available, may be used to determine the 995 // size of an object that may not yet have been initialized. 996 997 // Things to support parallel young-gen collection. 998 oop 999 ConcurrentMarkSweepGeneration::par_promote(int thread_num, 1000 oop old, markOop m, 1001 size_t word_sz) { 1002 #ifndef PRODUCT 1003 if (GenCollectedHeap::heap()->promotion_should_fail()) { 1004 return NULL; 1005 } 1006 #endif // #ifndef PRODUCT 1007 1008 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1009 PromotionInfo* promoInfo = &ps->promo; 1010 // if we are tracking promotions, then first ensure space for 1011 // promotion (including spooling space for saving header if necessary). 1012 // then allocate and copy, then track promoted info if needed. 1013 // When tracking (see PromotionInfo::track()), the mark word may 1014 // be displaced and in this case restoration of the mark word 1015 // occurs in the (oop_since_save_marks_)iterate phase. 1016 if (promoInfo->tracking() && !promoInfo->ensure_spooling_space()) { 1017 // Out of space for allocating spooling buffers; 1018 // try expanding and allocating spooling buffers. 1019 if (!expand_and_ensure_spooling_space(promoInfo)) { 1020 return NULL; 1021 } 1022 } 1023 assert(promoInfo->has_spooling_space(), "Control point invariant"); 1024 const size_t alloc_sz = CompactibleFreeListSpace::adjustObjectSize(word_sz); 1025 HeapWord* obj_ptr = ps->lab.alloc(alloc_sz); 1026 if (obj_ptr == NULL) { 1027 obj_ptr = expand_and_par_lab_allocate(ps, alloc_sz); 1028 if (obj_ptr == NULL) { 1029 return NULL; 1030 } 1031 } 1032 oop obj = oop(obj_ptr); 1033 OrderAccess::storestore(); 1034 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1035 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1036 // IMPORTANT: See note on object initialization for CMS above. 1037 // Otherwise, copy the object. Here we must be careful to insert the 1038 // klass pointer last, since this marks the block as an allocated object. 1039 // Except with compressed oops it's the mark word. 1040 HeapWord* old_ptr = (HeapWord*)old; 1041 // Restore the mark word copied above. 1042 obj->set_mark(m); 1043 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1044 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1045 OrderAccess::storestore(); 1046 1047 if (UseCompressedClassPointers) { 1048 // Copy gap missed by (aligned) header size calculation below 1049 obj->set_klass_gap(old->klass_gap()); 1050 } 1051 if (word_sz > (size_t)oopDesc::header_size()) { 1052 Copy::aligned_disjoint_words(old_ptr + oopDesc::header_size(), 1053 obj_ptr + oopDesc::header_size(), 1054 word_sz - oopDesc::header_size()); 1055 } 1056 1057 // Now we can track the promoted object, if necessary. We take care 1058 // to delay the transition from uninitialized to full object 1059 // (i.e., insertion of klass pointer) until after, so that it 1060 // atomically becomes a promoted object. 1061 if (promoInfo->tracking()) { 1062 promoInfo->track((PromotedObject*)obj, old->klass()); 1063 } 1064 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1065 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1066 assert(old->is_oop(), "Will use and dereference old klass ptr below"); 1067 1068 // Finally, install the klass pointer (this should be volatile). 1069 OrderAccess::storestore(); 1070 obj->set_klass(old->klass()); 1071 // We should now be able to calculate the right size for this object 1072 assert(obj->is_oop() && obj->size() == (int)word_sz, "Error, incorrect size computed for promoted object"); 1073 1074 collector()->promoted(true, // parallel 1075 obj_ptr, old->is_objArray(), word_sz); 1076 1077 NOT_PRODUCT( 1078 Atomic::inc_ptr(&_numObjectsPromoted); 1079 Atomic::add_ptr(alloc_sz, &_numWordsPromoted); 1080 ) 1081 1082 return obj; 1083 } 1084 1085 void 1086 ConcurrentMarkSweepGeneration:: 1087 par_promote_alloc_done(int thread_num) { 1088 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1089 ps->lab.retire(thread_num); 1090 } 1091 1092 void 1093 ConcurrentMarkSweepGeneration:: 1094 par_oop_since_save_marks_iterate_done(int thread_num) { 1095 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1096 ParScanWithoutBarrierClosure* dummy_cl = NULL; 1097 ps->promo.promoted_oops_iterate_nv(dummy_cl); 1098 } 1099 1100 bool ConcurrentMarkSweepGeneration::should_collect(bool full, 1101 size_t size, 1102 bool tlab) 1103 { 1104 // We allow a STW collection only if a full 1105 // collection was requested. 1106 return full || should_allocate(size, tlab); // FIX ME !!! 1107 // This and promotion failure handling are connected at the 1108 // hip and should be fixed by untying them. 1109 } 1110 1111 bool CMSCollector::shouldConcurrentCollect() { 1112 if (_full_gc_requested) { 1113 log_trace(gc)("CMSCollector: collect because of explicit gc request (or gc_locker)"); 1114 return true; 1115 } 1116 1117 FreelistLocker x(this); 1118 // ------------------------------------------------------------------ 1119 // Print out lots of information which affects the initiation of 1120 // a collection. 1121 LogHandle(gc) log; 1122 if (log.is_trace() && stats().valid()) { 1123 log.trace("CMSCollector shouldConcurrentCollect: "); 1124 ResourceMark rm; 1125 stats().print_on(log.debug_stream()); 1126 log.trace("time_until_cms_gen_full %3.7f", stats().time_until_cms_gen_full()); 1127 log.trace("free=" SIZE_FORMAT, _cmsGen->free()); 1128 log.trace("contiguous_available=" SIZE_FORMAT, _cmsGen->contiguous_available()); 1129 log.trace("promotion_rate=%g", stats().promotion_rate()); 1130 log.trace("cms_allocation_rate=%g", stats().cms_allocation_rate()); 1131 log.trace("occupancy=%3.7f", _cmsGen->occupancy()); 1132 log.trace("initiatingOccupancy=%3.7f", _cmsGen->initiating_occupancy()); 1133 log.trace("cms_time_since_begin=%3.7f", stats().cms_time_since_begin()); 1134 log.trace("cms_time_since_end=%3.7f", stats().cms_time_since_end()); 1135 log.trace("metadata initialized %d", MetaspaceGC::should_concurrent_collect()); 1136 } 1137 // ------------------------------------------------------------------ 1138 1139 // If the estimated time to complete a cms collection (cms_duration()) 1140 // is less than the estimated time remaining until the cms generation 1141 // is full, start a collection. 1142 if (!UseCMSInitiatingOccupancyOnly) { 1143 if (stats().valid()) { 1144 if (stats().time_until_cms_start() == 0.0) { 1145 return true; 1146 } 1147 } else { 1148 // We want to conservatively collect somewhat early in order 1149 // to try and "bootstrap" our CMS/promotion statistics; 1150 // this branch will not fire after the first successful CMS 1151 // collection because the stats should then be valid. 1152 if (_cmsGen->occupancy() >= _bootstrap_occupancy) { 1153 log_trace(gc)(" CMSCollector: collect for bootstrapping statistics: occupancy = %f, boot occupancy = %f", 1154 _cmsGen->occupancy(), _bootstrap_occupancy); 1155 return true; 1156 } 1157 } 1158 } 1159 1160 // Otherwise, we start a collection cycle if 1161 // old gen want a collection cycle started. Each may use 1162 // an appropriate criterion for making this decision. 1163 // XXX We need to make sure that the gen expansion 1164 // criterion dovetails well with this. XXX NEED TO FIX THIS 1165 if (_cmsGen->should_concurrent_collect()) { 1166 log_trace(gc)("CMS old gen initiated"); 1167 return true; 1168 } 1169 1170 // We start a collection if we believe an incremental collection may fail; 1171 // this is not likely to be productive in practice because it's probably too 1172 // late anyway. 1173 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1174 assert(gch->collector_policy()->is_generation_policy(), 1175 "You may want to check the correctness of the following"); 1176 if (gch->incremental_collection_will_fail(true /* consult_young */)) { 1177 log_trace(gc)("CMSCollector: collect because incremental collection will fail "); 1178 return true; 1179 } 1180 1181 if (MetaspaceGC::should_concurrent_collect()) { 1182 log_trace(gc)("CMSCollector: collect for metadata allocation "); 1183 return true; 1184 } 1185 1186 // CMSTriggerInterval starts a CMS cycle if enough time has passed. 1187 if (CMSTriggerInterval >= 0) { 1188 if (CMSTriggerInterval == 0) { 1189 // Trigger always 1190 return true; 1191 } 1192 1193 // Check the CMS time since begin (we do not check the stats validity 1194 // as we want to be able to trigger the first CMS cycle as well) 1195 if (stats().cms_time_since_begin() >= (CMSTriggerInterval / ((double) MILLIUNITS))) { 1196 if (stats().valid()) { 1197 log_trace(gc)("CMSCollector: collect because of trigger interval (time since last begin %3.7f secs)", 1198 stats().cms_time_since_begin()); 1199 } else { 1200 log_trace(gc)("CMSCollector: collect because of trigger interval (first collection)"); 1201 } 1202 return true; 1203 } 1204 } 1205 1206 return false; 1207 } 1208 1209 void CMSCollector::set_did_compact(bool v) { _cmsGen->set_did_compact(v); } 1210 1211 // Clear _expansion_cause fields of constituent generations 1212 void CMSCollector::clear_expansion_cause() { 1213 _cmsGen->clear_expansion_cause(); 1214 } 1215 1216 // We should be conservative in starting a collection cycle. To 1217 // start too eagerly runs the risk of collecting too often in the 1218 // extreme. To collect too rarely falls back on full collections, 1219 // which works, even if not optimum in terms of concurrent work. 1220 // As a work around for too eagerly collecting, use the flag 1221 // UseCMSInitiatingOccupancyOnly. This also has the advantage of 1222 // giving the user an easily understandable way of controlling the 1223 // collections. 1224 // We want to start a new collection cycle if any of the following 1225 // conditions hold: 1226 // . our current occupancy exceeds the configured initiating occupancy 1227 // for this generation, or 1228 // . we recently needed to expand this space and have not, since that 1229 // expansion, done a collection of this generation, or 1230 // . the underlying space believes that it may be a good idea to initiate 1231 // a concurrent collection (this may be based on criteria such as the 1232 // following: the space uses linear allocation and linear allocation is 1233 // going to fail, or there is believed to be excessive fragmentation in 1234 // the generation, etc... or ... 1235 // [.(currently done by CMSCollector::shouldConcurrentCollect() only for 1236 // the case of the old generation; see CR 6543076): 1237 // we may be approaching a point at which allocation requests may fail because 1238 // we will be out of sufficient free space given allocation rate estimates.] 1239 bool ConcurrentMarkSweepGeneration::should_concurrent_collect() const { 1240 1241 assert_lock_strong(freelistLock()); 1242 if (occupancy() > initiating_occupancy()) { 1243 log_trace(gc)(" %s: collect because of occupancy %f / %f ", 1244 short_name(), occupancy(), initiating_occupancy()); 1245 return true; 1246 } 1247 if (UseCMSInitiatingOccupancyOnly) { 1248 return false; 1249 } 1250 if (expansion_cause() == CMSExpansionCause::_satisfy_allocation) { 1251 log_trace(gc)(" %s: collect because expanded for allocation ", short_name()); 1252 return true; 1253 } 1254 return false; 1255 } 1256 1257 void ConcurrentMarkSweepGeneration::collect(bool full, 1258 bool clear_all_soft_refs, 1259 size_t size, 1260 bool tlab) 1261 { 1262 collector()->collect(full, clear_all_soft_refs, size, tlab); 1263 } 1264 1265 void CMSCollector::collect(bool full, 1266 bool clear_all_soft_refs, 1267 size_t size, 1268 bool tlab) 1269 { 1270 // The following "if" branch is present for defensive reasons. 1271 // In the current uses of this interface, it can be replaced with: 1272 // assert(!GC_locker.is_active(), "Can't be called otherwise"); 1273 // But I am not placing that assert here to allow future 1274 // generality in invoking this interface. 1275 if (GC_locker::is_active()) { 1276 // A consistency test for GC_locker 1277 assert(GC_locker::needs_gc(), "Should have been set already"); 1278 // Skip this foreground collection, instead 1279 // expanding the heap if necessary. 1280 // Need the free list locks for the call to free() in compute_new_size() 1281 compute_new_size(); 1282 return; 1283 } 1284 acquire_control_and_collect(full, clear_all_soft_refs); 1285 } 1286 1287 void CMSCollector::request_full_gc(unsigned int full_gc_count, GCCause::Cause cause) { 1288 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1289 unsigned int gc_count = gch->total_full_collections(); 1290 if (gc_count == full_gc_count) { 1291 MutexLockerEx y(CGC_lock, Mutex::_no_safepoint_check_flag); 1292 _full_gc_requested = true; 1293 _full_gc_cause = cause; 1294 CGC_lock->notify(); // nudge CMS thread 1295 } else { 1296 assert(gc_count > full_gc_count, "Error: causal loop"); 1297 } 1298 } 1299 1300 bool CMSCollector::is_external_interruption() { 1301 GCCause::Cause cause = GenCollectedHeap::heap()->gc_cause(); 1302 return GCCause::is_user_requested_gc(cause) || 1303 GCCause::is_serviceability_requested_gc(cause); 1304 } 1305 1306 void CMSCollector::report_concurrent_mode_interruption() { 1307 if (is_external_interruption()) { 1308 log_debug(gc)("Concurrent mode interrupted"); 1309 } else { 1310 log_debug(gc)("Concurrent mode failure"); 1311 _gc_tracer_cm->report_concurrent_mode_failure(); 1312 } 1313 } 1314 1315 1316 // The foreground and background collectors need to coordinate in order 1317 // to make sure that they do not mutually interfere with CMS collections. 1318 // When a background collection is active, 1319 // the foreground collector may need to take over (preempt) and 1320 // synchronously complete an ongoing collection. Depending on the 1321 // frequency of the background collections and the heap usage 1322 // of the application, this preemption can be seldom or frequent. 1323 // There are only certain 1324 // points in the background collection that the "collection-baton" 1325 // can be passed to the foreground collector. 1326 // 1327 // The foreground collector will wait for the baton before 1328 // starting any part of the collection. The foreground collector 1329 // will only wait at one location. 1330 // 1331 // The background collector will yield the baton before starting a new 1332 // phase of the collection (e.g., before initial marking, marking from roots, 1333 // precleaning, final re-mark, sweep etc.) This is normally done at the head 1334 // of the loop which switches the phases. The background collector does some 1335 // of the phases (initial mark, final re-mark) with the world stopped. 1336 // Because of locking involved in stopping the world, 1337 // the foreground collector should not block waiting for the background 1338 // collector when it is doing a stop-the-world phase. The background 1339 // collector will yield the baton at an additional point just before 1340 // it enters a stop-the-world phase. Once the world is stopped, the 1341 // background collector checks the phase of the collection. If the 1342 // phase has not changed, it proceeds with the collection. If the 1343 // phase has changed, it skips that phase of the collection. See 1344 // the comments on the use of the Heap_lock in collect_in_background(). 1345 // 1346 // Variable used in baton passing. 1347 // _foregroundGCIsActive - Set to true by the foreground collector when 1348 // it wants the baton. The foreground clears it when it has finished 1349 // the collection. 1350 // _foregroundGCShouldWait - Set to true by the background collector 1351 // when it is running. The foreground collector waits while 1352 // _foregroundGCShouldWait is true. 1353 // CGC_lock - monitor used to protect access to the above variables 1354 // and to notify the foreground and background collectors. 1355 // _collectorState - current state of the CMS collection. 1356 // 1357 // The foreground collector 1358 // acquires the CGC_lock 1359 // sets _foregroundGCIsActive 1360 // waits on the CGC_lock for _foregroundGCShouldWait to be false 1361 // various locks acquired in preparation for the collection 1362 // are released so as not to block the background collector 1363 // that is in the midst of a collection 1364 // proceeds with the collection 1365 // clears _foregroundGCIsActive 1366 // returns 1367 // 1368 // The background collector in a loop iterating on the phases of the 1369 // collection 1370 // acquires the CGC_lock 1371 // sets _foregroundGCShouldWait 1372 // if _foregroundGCIsActive is set 1373 // clears _foregroundGCShouldWait, notifies _CGC_lock 1374 // waits on _CGC_lock for _foregroundGCIsActive to become false 1375 // and exits the loop. 1376 // otherwise 1377 // proceed with that phase of the collection 1378 // if the phase is a stop-the-world phase, 1379 // yield the baton once more just before enqueueing 1380 // the stop-world CMS operation (executed by the VM thread). 1381 // returns after all phases of the collection are done 1382 // 1383 1384 void CMSCollector::acquire_control_and_collect(bool full, 1385 bool clear_all_soft_refs) { 1386 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint"); 1387 assert(!Thread::current()->is_ConcurrentGC_thread(), 1388 "shouldn't try to acquire control from self!"); 1389 1390 // Start the protocol for acquiring control of the 1391 // collection from the background collector (aka CMS thread). 1392 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 1393 "VM thread should have CMS token"); 1394 // Remember the possibly interrupted state of an ongoing 1395 // concurrent collection 1396 CollectorState first_state = _collectorState; 1397 1398 // Signal to a possibly ongoing concurrent collection that 1399 // we want to do a foreground collection. 1400 _foregroundGCIsActive = true; 1401 1402 // release locks and wait for a notify from the background collector 1403 // releasing the locks in only necessary for phases which 1404 // do yields to improve the granularity of the collection. 1405 assert_lock_strong(bitMapLock()); 1406 // We need to lock the Free list lock for the space that we are 1407 // currently collecting. 1408 assert(haveFreelistLocks(), "Must be holding free list locks"); 1409 bitMapLock()->unlock(); 1410 releaseFreelistLocks(); 1411 { 1412 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1413 if (_foregroundGCShouldWait) { 1414 // We are going to be waiting for action for the CMS thread; 1415 // it had better not be gone (for instance at shutdown)! 1416 assert(ConcurrentMarkSweepThread::cmst() != NULL, 1417 "CMS thread must be running"); 1418 // Wait here until the background collector gives us the go-ahead 1419 ConcurrentMarkSweepThread::clear_CMS_flag( 1420 ConcurrentMarkSweepThread::CMS_vm_has_token); // release token 1421 // Get a possibly blocked CMS thread going: 1422 // Note that we set _foregroundGCIsActive true above, 1423 // without protection of the CGC_lock. 1424 CGC_lock->notify(); 1425 assert(!ConcurrentMarkSweepThread::vm_thread_wants_cms_token(), 1426 "Possible deadlock"); 1427 while (_foregroundGCShouldWait) { 1428 // wait for notification 1429 CGC_lock->wait(Mutex::_no_safepoint_check_flag); 1430 // Possibility of delay/starvation here, since CMS token does 1431 // not know to give priority to VM thread? Actually, i think 1432 // there wouldn't be any delay/starvation, but the proof of 1433 // that "fact" (?) appears non-trivial. XXX 20011219YSR 1434 } 1435 ConcurrentMarkSweepThread::set_CMS_flag( 1436 ConcurrentMarkSweepThread::CMS_vm_has_token); 1437 } 1438 } 1439 // The CMS_token is already held. Get back the other locks. 1440 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 1441 "VM thread should have CMS token"); 1442 getFreelistLocks(); 1443 bitMapLock()->lock_without_safepoint_check(); 1444 log_debug(gc, state)("CMS foreground collector has asked for control " INTPTR_FORMAT " with first state %d", 1445 p2i(Thread::current()), first_state); 1446 log_debug(gc, state)(" gets control with state %d", _collectorState); 1447 1448 // Inform cms gen if this was due to partial collection failing. 1449 // The CMS gen may use this fact to determine its expansion policy. 1450 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1451 if (gch->incremental_collection_will_fail(false /* don't consult_young */)) { 1452 assert(!_cmsGen->incremental_collection_failed(), 1453 "Should have been noticed, reacted to and cleared"); 1454 _cmsGen->set_incremental_collection_failed(); 1455 } 1456 1457 if (first_state > Idling) { 1458 report_concurrent_mode_interruption(); 1459 } 1460 1461 set_did_compact(true); 1462 1463 // If the collection is being acquired from the background 1464 // collector, there may be references on the discovered 1465 // references lists. Abandon those references, since some 1466 // of them may have become unreachable after concurrent 1467 // discovery; the STW compacting collector will redo discovery 1468 // more precisely, without being subject to floating garbage. 1469 // Leaving otherwise unreachable references in the discovered 1470 // lists would require special handling. 1471 ref_processor()->disable_discovery(); 1472 ref_processor()->abandon_partial_discovery(); 1473 ref_processor()->verify_no_references_recorded(); 1474 1475 if (first_state > Idling) { 1476 save_heap_summary(); 1477 } 1478 1479 do_compaction_work(clear_all_soft_refs); 1480 1481 // Has the GC time limit been exceeded? 1482 size_t max_eden_size = _young_gen->max_eden_size(); 1483 GCCause::Cause gc_cause = gch->gc_cause(); 1484 size_policy()->check_gc_overhead_limit(_young_gen->used(), 1485 _young_gen->eden()->used(), 1486 _cmsGen->max_capacity(), 1487 max_eden_size, 1488 full, 1489 gc_cause, 1490 gch->collector_policy()); 1491 1492 // Reset the expansion cause, now that we just completed 1493 // a collection cycle. 1494 clear_expansion_cause(); 1495 _foregroundGCIsActive = false; 1496 return; 1497 } 1498 1499 // Resize the tenured generation 1500 // after obtaining the free list locks for the 1501 // two generations. 1502 void CMSCollector::compute_new_size() { 1503 assert_locked_or_safepoint(Heap_lock); 1504 FreelistLocker z(this); 1505 MetaspaceGC::compute_new_size(); 1506 _cmsGen->compute_new_size_free_list(); 1507 } 1508 1509 // A work method used by the foreground collector to do 1510 // a mark-sweep-compact. 1511 void CMSCollector::do_compaction_work(bool clear_all_soft_refs) { 1512 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1513 1514 STWGCTimer* gc_timer = GenMarkSweep::gc_timer(); 1515 gc_timer->register_gc_start(); 1516 1517 SerialOldTracer* gc_tracer = GenMarkSweep::gc_tracer(); 1518 gc_tracer->report_gc_start(gch->gc_cause(), gc_timer->gc_start()); 1519 1520 GCTraceTime(Trace, gc) t("CMS:MSC"); 1521 1522 // Temporarily widen the span of the weak reference processing to 1523 // the entire heap. 1524 MemRegion new_span(GenCollectedHeap::heap()->reserved_region()); 1525 ReferenceProcessorSpanMutator rp_mut_span(ref_processor(), new_span); 1526 // Temporarily, clear the "is_alive_non_header" field of the 1527 // reference processor. 1528 ReferenceProcessorIsAliveMutator rp_mut_closure(ref_processor(), NULL); 1529 // Temporarily make reference _processing_ single threaded (non-MT). 1530 ReferenceProcessorMTProcMutator rp_mut_mt_processing(ref_processor(), false); 1531 // Temporarily make refs discovery atomic 1532 ReferenceProcessorAtomicMutator rp_mut_atomic(ref_processor(), true); 1533 // Temporarily make reference _discovery_ single threaded (non-MT) 1534 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false); 1535 1536 ref_processor()->set_enqueuing_is_done(false); 1537 ref_processor()->enable_discovery(); 1538 ref_processor()->setup_policy(clear_all_soft_refs); 1539 // If an asynchronous collection finishes, the _modUnionTable is 1540 // all clear. If we are assuming the collection from an asynchronous 1541 // collection, clear the _modUnionTable. 1542 assert(_collectorState != Idling || _modUnionTable.isAllClear(), 1543 "_modUnionTable should be clear if the baton was not passed"); 1544 _modUnionTable.clear_all(); 1545 assert(_collectorState != Idling || _ct->klass_rem_set()->mod_union_is_clear(), 1546 "mod union for klasses should be clear if the baton was passed"); 1547 _ct->klass_rem_set()->clear_mod_union(); 1548 1549 // We must adjust the allocation statistics being maintained 1550 // in the free list space. We do so by reading and clearing 1551 // the sweep timer and updating the block flux rate estimates below. 1552 assert(!_intra_sweep_timer.is_active(), "_intra_sweep_timer should be inactive"); 1553 if (_inter_sweep_timer.is_active()) { 1554 _inter_sweep_timer.stop(); 1555 // Note that we do not use this sample to update the _inter_sweep_estimate. 1556 _cmsGen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()), 1557 _inter_sweep_estimate.padded_average(), 1558 _intra_sweep_estimate.padded_average()); 1559 } 1560 1561 GenMarkSweep::invoke_at_safepoint(ref_processor(), clear_all_soft_refs); 1562 #ifdef ASSERT 1563 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 1564 size_t free_size = cms_space->free(); 1565 assert(free_size == 1566 pointer_delta(cms_space->end(), cms_space->compaction_top()) 1567 * HeapWordSize, 1568 "All the free space should be compacted into one chunk at top"); 1569 assert(cms_space->dictionary()->total_chunk_size( 1570 debug_only(cms_space->freelistLock())) == 0 || 1571 cms_space->totalSizeInIndexedFreeLists() == 0, 1572 "All the free space should be in a single chunk"); 1573 size_t num = cms_space->totalCount(); 1574 assert((free_size == 0 && num == 0) || 1575 (free_size > 0 && (num == 1 || num == 2)), 1576 "There should be at most 2 free chunks after compaction"); 1577 #endif // ASSERT 1578 _collectorState = Resetting; 1579 assert(_restart_addr == NULL, 1580 "Should have been NULL'd before baton was passed"); 1581 reset_stw(); 1582 _cmsGen->reset_after_compaction(); 1583 _concurrent_cycles_since_last_unload = 0; 1584 1585 // Clear any data recorded in the PLAB chunk arrays. 1586 if (_survivor_plab_array != NULL) { 1587 reset_survivor_plab_arrays(); 1588 } 1589 1590 // Adjust the per-size allocation stats for the next epoch. 1591 _cmsGen->cmsSpace()->endSweepFLCensus(sweep_count() /* fake */); 1592 // Restart the "inter sweep timer" for the next epoch. 1593 _inter_sweep_timer.reset(); 1594 _inter_sweep_timer.start(); 1595 1596 gc_timer->register_gc_end(); 1597 1598 gc_tracer->report_gc_end(gc_timer->gc_end(), gc_timer->time_partitions()); 1599 1600 // For a mark-sweep-compact, compute_new_size() will be called 1601 // in the heap's do_collection() method. 1602 } 1603 1604 void CMSCollector::print_eden_and_survivor_chunk_arrays() { 1605 LogHandle(gc, heap) log; 1606 if (!log.is_trace()) { 1607 return; 1608 } 1609 1610 ContiguousSpace* eden_space = _young_gen->eden(); 1611 ContiguousSpace* from_space = _young_gen->from(); 1612 ContiguousSpace* to_space = _young_gen->to(); 1613 // Eden 1614 if (_eden_chunk_array != NULL) { 1615 log.trace("eden " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")", 1616 p2i(eden_space->bottom()), p2i(eden_space->top()), 1617 p2i(eden_space->end()), eden_space->capacity()); 1618 log.trace("_eden_chunk_index=" SIZE_FORMAT ", _eden_chunk_capacity=" SIZE_FORMAT, 1619 _eden_chunk_index, _eden_chunk_capacity); 1620 for (size_t i = 0; i < _eden_chunk_index; i++) { 1621 log.trace("_eden_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT, i, p2i(_eden_chunk_array[i])); 1622 } 1623 } 1624 // Survivor 1625 if (_survivor_chunk_array != NULL) { 1626 log.trace("survivor " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")", 1627 p2i(from_space->bottom()), p2i(from_space->top()), 1628 p2i(from_space->end()), from_space->capacity()); 1629 log.trace("_survivor_chunk_index=" SIZE_FORMAT ", _survivor_chunk_capacity=" SIZE_FORMAT, 1630 _survivor_chunk_index, _survivor_chunk_capacity); 1631 for (size_t i = 0; i < _survivor_chunk_index; i++) { 1632 log.trace("_survivor_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT, i, p2i(_survivor_chunk_array[i])); 1633 } 1634 } 1635 } 1636 1637 void CMSCollector::getFreelistLocks() const { 1638 // Get locks for all free lists in all generations that this 1639 // collector is responsible for 1640 _cmsGen->freelistLock()->lock_without_safepoint_check(); 1641 } 1642 1643 void CMSCollector::releaseFreelistLocks() const { 1644 // Release locks for all free lists in all generations that this 1645 // collector is responsible for 1646 _cmsGen->freelistLock()->unlock(); 1647 } 1648 1649 bool CMSCollector::haveFreelistLocks() const { 1650 // Check locks for all free lists in all generations that this 1651 // collector is responsible for 1652 assert_lock_strong(_cmsGen->freelistLock()); 1653 PRODUCT_ONLY(ShouldNotReachHere()); 1654 return true; 1655 } 1656 1657 // A utility class that is used by the CMS collector to 1658 // temporarily "release" the foreground collector from its 1659 // usual obligation to wait for the background collector to 1660 // complete an ongoing phase before proceeding. 1661 class ReleaseForegroundGC: public StackObj { 1662 private: 1663 CMSCollector* _c; 1664 public: 1665 ReleaseForegroundGC(CMSCollector* c) : _c(c) { 1666 assert(_c->_foregroundGCShouldWait, "Else should not need to call"); 1667 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1668 // allow a potentially blocked foreground collector to proceed 1669 _c->_foregroundGCShouldWait = false; 1670 if (_c->_foregroundGCIsActive) { 1671 CGC_lock->notify(); 1672 } 1673 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 1674 "Possible deadlock"); 1675 } 1676 1677 ~ReleaseForegroundGC() { 1678 assert(!_c->_foregroundGCShouldWait, "Usage protocol violation?"); 1679 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1680 _c->_foregroundGCShouldWait = true; 1681 } 1682 }; 1683 1684 void CMSCollector::collect_in_background(GCCause::Cause cause) { 1685 assert(Thread::current()->is_ConcurrentGC_thread(), 1686 "A CMS asynchronous collection is only allowed on a CMS thread."); 1687 1688 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1689 { 1690 bool safepoint_check = Mutex::_no_safepoint_check_flag; 1691 MutexLockerEx hl(Heap_lock, safepoint_check); 1692 FreelistLocker fll(this); 1693 MutexLockerEx x(CGC_lock, safepoint_check); 1694 if (_foregroundGCIsActive) { 1695 // The foreground collector is. Skip this 1696 // background collection. 1697 assert(!_foregroundGCShouldWait, "Should be clear"); 1698 return; 1699 } else { 1700 assert(_collectorState == Idling, "Should be idling before start."); 1701 _collectorState = InitialMarking; 1702 register_gc_start(cause); 1703 // Reset the expansion cause, now that we are about to begin 1704 // a new cycle. 1705 clear_expansion_cause(); 1706 1707 // Clear the MetaspaceGC flag since a concurrent collection 1708 // is starting but also clear it after the collection. 1709 MetaspaceGC::set_should_concurrent_collect(false); 1710 } 1711 // Decide if we want to enable class unloading as part of the 1712 // ensuing concurrent GC cycle. 1713 update_should_unload_classes(); 1714 _full_gc_requested = false; // acks all outstanding full gc requests 1715 _full_gc_cause = GCCause::_no_gc; 1716 // Signal that we are about to start a collection 1717 gch->increment_total_full_collections(); // ... starting a collection cycle 1718 _collection_count_start = gch->total_full_collections(); 1719 } 1720 1721 size_t prev_used = _cmsGen->used(); 1722 1723 // The change of the collection state is normally done at this level; 1724 // the exceptions are phases that are executed while the world is 1725 // stopped. For those phases the change of state is done while the 1726 // world is stopped. For baton passing purposes this allows the 1727 // background collector to finish the phase and change state atomically. 1728 // The foreground collector cannot wait on a phase that is done 1729 // while the world is stopped because the foreground collector already 1730 // has the world stopped and would deadlock. 1731 while (_collectorState != Idling) { 1732 log_debug(gc, state)("Thread " INTPTR_FORMAT " in CMS state %d", 1733 p2i(Thread::current()), _collectorState); 1734 // The foreground collector 1735 // holds the Heap_lock throughout its collection. 1736 // holds the CMS token (but not the lock) 1737 // except while it is waiting for the background collector to yield. 1738 // 1739 // The foreground collector should be blocked (not for long) 1740 // if the background collector is about to start a phase 1741 // executed with world stopped. If the background 1742 // collector has already started such a phase, the 1743 // foreground collector is blocked waiting for the 1744 // Heap_lock. The stop-world phases (InitialMarking and FinalMarking) 1745 // are executed in the VM thread. 1746 // 1747 // The locking order is 1748 // PendingListLock (PLL) -- if applicable (FinalMarking) 1749 // Heap_lock (both this & PLL locked in VM_CMS_Operation::prologue()) 1750 // CMS token (claimed in 1751 // stop_world_and_do() --> 1752 // safepoint_synchronize() --> 1753 // CMSThread::synchronize()) 1754 1755 { 1756 // Check if the FG collector wants us to yield. 1757 CMSTokenSync x(true); // is cms thread 1758 if (waitForForegroundGC()) { 1759 // We yielded to a foreground GC, nothing more to be 1760 // done this round. 1761 assert(_foregroundGCShouldWait == false, "We set it to false in " 1762 "waitForForegroundGC()"); 1763 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " exiting collection CMS state %d", 1764 p2i(Thread::current()), _collectorState); 1765 return; 1766 } else { 1767 // The background collector can run but check to see if the 1768 // foreground collector has done a collection while the 1769 // background collector was waiting to get the CGC_lock 1770 // above. If yes, break so that _foregroundGCShouldWait 1771 // is cleared before returning. 1772 if (_collectorState == Idling) { 1773 break; 1774 } 1775 } 1776 } 1777 1778 assert(_foregroundGCShouldWait, "Foreground collector, if active, " 1779 "should be waiting"); 1780 1781 switch (_collectorState) { 1782 case InitialMarking: 1783 { 1784 ReleaseForegroundGC x(this); 1785 stats().record_cms_begin(); 1786 VM_CMS_Initial_Mark initial_mark_op(this); 1787 VMThread::execute(&initial_mark_op); 1788 } 1789 // The collector state may be any legal state at this point 1790 // since the background collector may have yielded to the 1791 // foreground collector. 1792 break; 1793 case Marking: 1794 // initial marking in checkpointRootsInitialWork has been completed 1795 if (markFromRoots()) { // we were successful 1796 assert(_collectorState == Precleaning, "Collector state should " 1797 "have changed"); 1798 } else { 1799 assert(_foregroundGCIsActive, "Internal state inconsistency"); 1800 } 1801 break; 1802 case Precleaning: 1803 // marking from roots in markFromRoots has been completed 1804 preclean(); 1805 assert(_collectorState == AbortablePreclean || 1806 _collectorState == FinalMarking, 1807 "Collector state should have changed"); 1808 break; 1809 case AbortablePreclean: 1810 abortable_preclean(); 1811 assert(_collectorState == FinalMarking, "Collector state should " 1812 "have changed"); 1813 break; 1814 case FinalMarking: 1815 { 1816 ReleaseForegroundGC x(this); 1817 1818 VM_CMS_Final_Remark final_remark_op(this); 1819 VMThread::execute(&final_remark_op); 1820 } 1821 assert(_foregroundGCShouldWait, "block post-condition"); 1822 break; 1823 case Sweeping: 1824 // final marking in checkpointRootsFinal has been completed 1825 sweep(); 1826 assert(_collectorState == Resizing, "Collector state change " 1827 "to Resizing must be done under the free_list_lock"); 1828 1829 case Resizing: { 1830 // Sweeping has been completed... 1831 // At this point the background collection has completed. 1832 // Don't move the call to compute_new_size() down 1833 // into code that might be executed if the background 1834 // collection was preempted. 1835 { 1836 ReleaseForegroundGC x(this); // unblock FG collection 1837 MutexLockerEx y(Heap_lock, Mutex::_no_safepoint_check_flag); 1838 CMSTokenSync z(true); // not strictly needed. 1839 if (_collectorState == Resizing) { 1840 compute_new_size(); 1841 save_heap_summary(); 1842 _collectorState = Resetting; 1843 } else { 1844 assert(_collectorState == Idling, "The state should only change" 1845 " because the foreground collector has finished the collection"); 1846 } 1847 } 1848 break; 1849 } 1850 case Resetting: 1851 // CMS heap resizing has been completed 1852 reset_concurrent(); 1853 assert(_collectorState == Idling, "Collector state should " 1854 "have changed"); 1855 1856 MetaspaceGC::set_should_concurrent_collect(false); 1857 1858 stats().record_cms_end(); 1859 // Don't move the concurrent_phases_end() and compute_new_size() 1860 // calls to here because a preempted background collection 1861 // has it's state set to "Resetting". 1862 break; 1863 case Idling: 1864 default: 1865 ShouldNotReachHere(); 1866 break; 1867 } 1868 log_debug(gc, state)(" Thread " INTPTR_FORMAT " done - next CMS state %d", 1869 p2i(Thread::current()), _collectorState); 1870 assert(_foregroundGCShouldWait, "block post-condition"); 1871 } 1872 1873 // Should this be in gc_epilogue? 1874 collector_policy()->counters()->update_counters(); 1875 1876 { 1877 // Clear _foregroundGCShouldWait and, in the event that the 1878 // foreground collector is waiting, notify it, before 1879 // returning. 1880 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1881 _foregroundGCShouldWait = false; 1882 if (_foregroundGCIsActive) { 1883 CGC_lock->notify(); 1884 } 1885 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 1886 "Possible deadlock"); 1887 } 1888 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " exiting collection CMS state %d", 1889 p2i(Thread::current()), _collectorState); 1890 log_info(gc, heap)("Old: " SIZE_FORMAT "K->" SIZE_FORMAT "K(" SIZE_FORMAT "K)", 1891 prev_used / K, _cmsGen->used()/K, _cmsGen->capacity() /K); 1892 } 1893 1894 void CMSCollector::register_gc_start(GCCause::Cause cause) { 1895 _cms_start_registered = true; 1896 _gc_timer_cm->register_gc_start(); 1897 _gc_tracer_cm->report_gc_start(cause, _gc_timer_cm->gc_start()); 1898 } 1899 1900 void CMSCollector::register_gc_end() { 1901 if (_cms_start_registered) { 1902 report_heap_summary(GCWhen::AfterGC); 1903 1904 _gc_timer_cm->register_gc_end(); 1905 _gc_tracer_cm->report_gc_end(_gc_timer_cm->gc_end(), _gc_timer_cm->time_partitions()); 1906 _cms_start_registered = false; 1907 } 1908 } 1909 1910 void CMSCollector::save_heap_summary() { 1911 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1912 _last_heap_summary = gch->create_heap_summary(); 1913 _last_metaspace_summary = gch->create_metaspace_summary(); 1914 } 1915 1916 void CMSCollector::report_heap_summary(GCWhen::Type when) { 1917 _gc_tracer_cm->report_gc_heap_summary(when, _last_heap_summary); 1918 _gc_tracer_cm->report_metaspace_summary(when, _last_metaspace_summary); 1919 } 1920 1921 bool CMSCollector::waitForForegroundGC() { 1922 bool res = false; 1923 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 1924 "CMS thread should have CMS token"); 1925 // Block the foreground collector until the 1926 // background collectors decides whether to 1927 // yield. 1928 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1929 _foregroundGCShouldWait = true; 1930 if (_foregroundGCIsActive) { 1931 // The background collector yields to the 1932 // foreground collector and returns a value 1933 // indicating that it has yielded. The foreground 1934 // collector can proceed. 1935 res = true; 1936 _foregroundGCShouldWait = false; 1937 ConcurrentMarkSweepThread::clear_CMS_flag( 1938 ConcurrentMarkSweepThread::CMS_cms_has_token); 1939 ConcurrentMarkSweepThread::set_CMS_flag( 1940 ConcurrentMarkSweepThread::CMS_cms_wants_token); 1941 // Get a possibly blocked foreground thread going 1942 CGC_lock->notify(); 1943 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " waiting at CMS state %d", 1944 p2i(Thread::current()), _collectorState); 1945 while (_foregroundGCIsActive) { 1946 CGC_lock->wait(Mutex::_no_safepoint_check_flag); 1947 } 1948 ConcurrentMarkSweepThread::set_CMS_flag( 1949 ConcurrentMarkSweepThread::CMS_cms_has_token); 1950 ConcurrentMarkSweepThread::clear_CMS_flag( 1951 ConcurrentMarkSweepThread::CMS_cms_wants_token); 1952 } 1953 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " continuing at CMS state %d", 1954 p2i(Thread::current()), _collectorState); 1955 return res; 1956 } 1957 1958 // Because of the need to lock the free lists and other structures in 1959 // the collector, common to all the generations that the collector is 1960 // collecting, we need the gc_prologues of individual CMS generations 1961 // delegate to their collector. It may have been simpler had the 1962 // current infrastructure allowed one to call a prologue on a 1963 // collector. In the absence of that we have the generation's 1964 // prologue delegate to the collector, which delegates back 1965 // some "local" work to a worker method in the individual generations 1966 // that it's responsible for collecting, while itself doing any 1967 // work common to all generations it's responsible for. A similar 1968 // comment applies to the gc_epilogue()'s. 1969 // The role of the variable _between_prologue_and_epilogue is to 1970 // enforce the invocation protocol. 1971 void CMSCollector::gc_prologue(bool full) { 1972 // Call gc_prologue_work() for the CMSGen 1973 // we are responsible for. 1974 1975 // The following locking discipline assumes that we are only called 1976 // when the world is stopped. 1977 assert(SafepointSynchronize::is_at_safepoint(), "world is stopped assumption"); 1978 1979 // The CMSCollector prologue must call the gc_prologues for the 1980 // "generations" that it's responsible 1981 // for. 1982 1983 assert( Thread::current()->is_VM_thread() 1984 || ( CMSScavengeBeforeRemark 1985 && Thread::current()->is_ConcurrentGC_thread()), 1986 "Incorrect thread type for prologue execution"); 1987 1988 if (_between_prologue_and_epilogue) { 1989 // We have already been invoked; this is a gc_prologue delegation 1990 // from yet another CMS generation that we are responsible for, just 1991 // ignore it since all relevant work has already been done. 1992 return; 1993 } 1994 1995 // set a bit saying prologue has been called; cleared in epilogue 1996 _between_prologue_and_epilogue = true; 1997 // Claim locks for common data structures, then call gc_prologue_work() 1998 // for each CMSGen. 1999 2000 getFreelistLocks(); // gets free list locks on constituent spaces 2001 bitMapLock()->lock_without_safepoint_check(); 2002 2003 // Should call gc_prologue_work() for all cms gens we are responsible for 2004 bool duringMarking = _collectorState >= Marking 2005 && _collectorState < Sweeping; 2006 2007 // The young collections clear the modified oops state, which tells if 2008 // there are any modified oops in the class. The remark phase also needs 2009 // that information. Tell the young collection to save the union of all 2010 // modified klasses. 2011 if (duringMarking) { 2012 _ct->klass_rem_set()->set_accumulate_modified_oops(true); 2013 } 2014 2015 bool registerClosure = duringMarking; 2016 2017 _cmsGen->gc_prologue_work(full, registerClosure, &_modUnionClosurePar); 2018 2019 if (!full) { 2020 stats().record_gc0_begin(); 2021 } 2022 } 2023 2024 void ConcurrentMarkSweepGeneration::gc_prologue(bool full) { 2025 2026 _capacity_at_prologue = capacity(); 2027 _used_at_prologue = used(); 2028 2029 // Delegate to CMScollector which knows how to coordinate between 2030 // this and any other CMS generations that it is responsible for 2031 // collecting. 2032 collector()->gc_prologue(full); 2033 } 2034 2035 // This is a "private" interface for use by this generation's CMSCollector. 2036 // Not to be called directly by any other entity (for instance, 2037 // GenCollectedHeap, which calls the "public" gc_prologue method above). 2038 void ConcurrentMarkSweepGeneration::gc_prologue_work(bool full, 2039 bool registerClosure, ModUnionClosure* modUnionClosure) { 2040 assert(!incremental_collection_failed(), "Shouldn't be set yet"); 2041 assert(cmsSpace()->preconsumptionDirtyCardClosure() == NULL, 2042 "Should be NULL"); 2043 if (registerClosure) { 2044 cmsSpace()->setPreconsumptionDirtyCardClosure(modUnionClosure); 2045 } 2046 cmsSpace()->gc_prologue(); 2047 // Clear stat counters 2048 NOT_PRODUCT( 2049 assert(_numObjectsPromoted == 0, "check"); 2050 assert(_numWordsPromoted == 0, "check"); 2051 log_develop_trace(gc, alloc)("Allocated " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes concurrently", 2052 _numObjectsAllocated, _numWordsAllocated*sizeof(HeapWord)); 2053 _numObjectsAllocated = 0; 2054 _numWordsAllocated = 0; 2055 ) 2056 } 2057 2058 void CMSCollector::gc_epilogue(bool full) { 2059 // The following locking discipline assumes that we are only called 2060 // when the world is stopped. 2061 assert(SafepointSynchronize::is_at_safepoint(), 2062 "world is stopped assumption"); 2063 2064 // Currently the CMS epilogue (see CompactibleFreeListSpace) merely checks 2065 // if linear allocation blocks need to be appropriately marked to allow the 2066 // the blocks to be parsable. We also check here whether we need to nudge the 2067 // CMS collector thread to start a new cycle (if it's not already active). 2068 assert( Thread::current()->is_VM_thread() 2069 || ( CMSScavengeBeforeRemark 2070 && Thread::current()->is_ConcurrentGC_thread()), 2071 "Incorrect thread type for epilogue execution"); 2072 2073 if (!_between_prologue_and_epilogue) { 2074 // We have already been invoked; this is a gc_epilogue delegation 2075 // from yet another CMS generation that we are responsible for, just 2076 // ignore it since all relevant work has already been done. 2077 return; 2078 } 2079 assert(haveFreelistLocks(), "must have freelist locks"); 2080 assert_lock_strong(bitMapLock()); 2081 2082 _ct->klass_rem_set()->set_accumulate_modified_oops(false); 2083 2084 _cmsGen->gc_epilogue_work(full); 2085 2086 if (_collectorState == AbortablePreclean || _collectorState == Precleaning) { 2087 // in case sampling was not already enabled, enable it 2088 _start_sampling = true; 2089 } 2090 // reset _eden_chunk_array so sampling starts afresh 2091 _eden_chunk_index = 0; 2092 2093 size_t cms_used = _cmsGen->cmsSpace()->used(); 2094 2095 // update performance counters - this uses a special version of 2096 // update_counters() that allows the utilization to be passed as a 2097 // parameter, avoiding multiple calls to used(). 2098 // 2099 _cmsGen->update_counters(cms_used); 2100 2101 bitMapLock()->unlock(); 2102 releaseFreelistLocks(); 2103 2104 if (!CleanChunkPoolAsync) { 2105 Chunk::clean_chunk_pool(); 2106 } 2107 2108 set_did_compact(false); 2109 _between_prologue_and_epilogue = false; // ready for next cycle 2110 } 2111 2112 void ConcurrentMarkSweepGeneration::gc_epilogue(bool full) { 2113 collector()->gc_epilogue(full); 2114 2115 // Also reset promotion tracking in par gc thread states. 2116 for (uint i = 0; i < ParallelGCThreads; i++) { 2117 _par_gc_thread_states[i]->promo.stopTrackingPromotions(i); 2118 } 2119 } 2120 2121 void ConcurrentMarkSweepGeneration::gc_epilogue_work(bool full) { 2122 assert(!incremental_collection_failed(), "Should have been cleared"); 2123 cmsSpace()->setPreconsumptionDirtyCardClosure(NULL); 2124 cmsSpace()->gc_epilogue(); 2125 // Print stat counters 2126 NOT_PRODUCT( 2127 assert(_numObjectsAllocated == 0, "check"); 2128 assert(_numWordsAllocated == 0, "check"); 2129 log_develop_trace(gc, promotion)("Promoted " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes", 2130 _numObjectsPromoted, _numWordsPromoted*sizeof(HeapWord)); 2131 _numObjectsPromoted = 0; 2132 _numWordsPromoted = 0; 2133 ) 2134 2135 // Call down the chain in contiguous_available needs the freelistLock 2136 // so print this out before releasing the freeListLock. 2137 log_develop_trace(gc)(" Contiguous available " SIZE_FORMAT " bytes ", contiguous_available()); 2138 } 2139 2140 #ifndef PRODUCT 2141 bool CMSCollector::have_cms_token() { 2142 Thread* thr = Thread::current(); 2143 if (thr->is_VM_thread()) { 2144 return ConcurrentMarkSweepThread::vm_thread_has_cms_token(); 2145 } else if (thr->is_ConcurrentGC_thread()) { 2146 return ConcurrentMarkSweepThread::cms_thread_has_cms_token(); 2147 } else if (thr->is_GC_task_thread()) { 2148 return ConcurrentMarkSweepThread::vm_thread_has_cms_token() && 2149 ParGCRareEvent_lock->owned_by_self(); 2150 } 2151 return false; 2152 } 2153 2154 // Check reachability of the given heap address in CMS generation, 2155 // treating all other generations as roots. 2156 bool CMSCollector::is_cms_reachable(HeapWord* addr) { 2157 // We could "guarantee" below, rather than assert, but I'll 2158 // leave these as "asserts" so that an adventurous debugger 2159 // could try this in the product build provided some subset of 2160 // the conditions were met, provided they were interested in the 2161 // results and knew that the computation below wouldn't interfere 2162 // with other concurrent computations mutating the structures 2163 // being read or written. 2164 assert(SafepointSynchronize::is_at_safepoint(), 2165 "Else mutations in object graph will make answer suspect"); 2166 assert(have_cms_token(), "Should hold cms token"); 2167 assert(haveFreelistLocks(), "must hold free list locks"); 2168 assert_lock_strong(bitMapLock()); 2169 2170 // Clear the marking bit map array before starting, but, just 2171 // for kicks, first report if the given address is already marked 2172 tty->print_cr("Start: Address " PTR_FORMAT " is%s marked", p2i(addr), 2173 _markBitMap.isMarked(addr) ? "" : " not"); 2174 2175 if (verify_after_remark()) { 2176 MutexLockerEx x(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag); 2177 bool result = verification_mark_bm()->isMarked(addr); 2178 tty->print_cr("TransitiveMark: Address " PTR_FORMAT " %s marked", p2i(addr), 2179 result ? "IS" : "is NOT"); 2180 return result; 2181 } else { 2182 tty->print_cr("Could not compute result"); 2183 return false; 2184 } 2185 } 2186 #endif 2187 2188 void 2189 CMSCollector::print_on_error(outputStream* st) { 2190 CMSCollector* collector = ConcurrentMarkSweepGeneration::_collector; 2191 if (collector != NULL) { 2192 CMSBitMap* bitmap = &collector->_markBitMap; 2193 st->print_cr("Marking Bits: (CMSBitMap*) " PTR_FORMAT, p2i(bitmap)); 2194 bitmap->print_on_error(st, " Bits: "); 2195 2196 st->cr(); 2197 2198 CMSBitMap* mut_bitmap = &collector->_modUnionTable; 2199 st->print_cr("Mod Union Table: (CMSBitMap*) " PTR_FORMAT, p2i(mut_bitmap)); 2200 mut_bitmap->print_on_error(st, " Bits: "); 2201 } 2202 } 2203 2204 //////////////////////////////////////////////////////// 2205 // CMS Verification Support 2206 //////////////////////////////////////////////////////// 2207 // Following the remark phase, the following invariant 2208 // should hold -- each object in the CMS heap which is 2209 // marked in markBitMap() should be marked in the verification_mark_bm(). 2210 2211 class VerifyMarkedClosure: public BitMapClosure { 2212 CMSBitMap* _marks; 2213 bool _failed; 2214 2215 public: 2216 VerifyMarkedClosure(CMSBitMap* bm): _marks(bm), _failed(false) {} 2217 2218 bool do_bit(size_t offset) { 2219 HeapWord* addr = _marks->offsetToHeapWord(offset); 2220 if (!_marks->isMarked(addr)) { 2221 LogHandle(gc, verify) log; 2222 ResourceMark rm; 2223 oop(addr)->print_on(log.info_stream()); 2224 log.info(" (" INTPTR_FORMAT " should have been marked)", p2i(addr)); 2225 _failed = true; 2226 } 2227 return true; 2228 } 2229 2230 bool failed() { return _failed; } 2231 }; 2232 2233 bool CMSCollector::verify_after_remark() { 2234 GCTraceTime(Info, gc, verify) tm("Verifying CMS Marking."); 2235 MutexLockerEx ml(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag); 2236 static bool init = false; 2237 2238 assert(SafepointSynchronize::is_at_safepoint(), 2239 "Else mutations in object graph will make answer suspect"); 2240 assert(have_cms_token(), 2241 "Else there may be mutual interference in use of " 2242 " verification data structures"); 2243 assert(_collectorState > Marking && _collectorState <= Sweeping, 2244 "Else marking info checked here may be obsolete"); 2245 assert(haveFreelistLocks(), "must hold free list locks"); 2246 assert_lock_strong(bitMapLock()); 2247 2248 2249 // Allocate marking bit map if not already allocated 2250 if (!init) { // first time 2251 if (!verification_mark_bm()->allocate(_span)) { 2252 return false; 2253 } 2254 init = true; 2255 } 2256 2257 assert(verification_mark_stack()->isEmpty(), "Should be empty"); 2258 2259 // Turn off refs discovery -- so we will be tracing through refs. 2260 // This is as intended, because by this time 2261 // GC must already have cleared any refs that need to be cleared, 2262 // and traced those that need to be marked; moreover, 2263 // the marking done here is not going to interfere in any 2264 // way with the marking information used by GC. 2265 NoRefDiscovery no_discovery(ref_processor()); 2266 2267 #if defined(COMPILER2) || INCLUDE_JVMCI 2268 DerivedPointerTableDeactivate dpt_deact; 2269 #endif 2270 2271 // Clear any marks from a previous round 2272 verification_mark_bm()->clear_all(); 2273 assert(verification_mark_stack()->isEmpty(), "markStack should be empty"); 2274 verify_work_stacks_empty(); 2275 2276 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2277 gch->ensure_parsability(false); // fill TLABs, but no need to retire them 2278 // Update the saved marks which may affect the root scans. 2279 gch->save_marks(); 2280 2281 if (CMSRemarkVerifyVariant == 1) { 2282 // In this first variant of verification, we complete 2283 // all marking, then check if the new marks-vector is 2284 // a subset of the CMS marks-vector. 2285 verify_after_remark_work_1(); 2286 } else if (CMSRemarkVerifyVariant == 2) { 2287 // In this second variant of verification, we flag an error 2288 // (i.e. an object reachable in the new marks-vector not reachable 2289 // in the CMS marks-vector) immediately, also indicating the 2290 // identify of an object (A) that references the unmarked object (B) -- 2291 // presumably, a mutation to A failed to be picked up by preclean/remark? 2292 verify_after_remark_work_2(); 2293 } else { 2294 warning("Unrecognized value " UINTX_FORMAT " for CMSRemarkVerifyVariant", 2295 CMSRemarkVerifyVariant); 2296 } 2297 return true; 2298 } 2299 2300 void CMSCollector::verify_after_remark_work_1() { 2301 ResourceMark rm; 2302 HandleMark hm; 2303 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2304 2305 // Get a clear set of claim bits for the roots processing to work with. 2306 ClassLoaderDataGraph::clear_claimed_marks(); 2307 2308 // Mark from roots one level into CMS 2309 MarkRefsIntoClosure notOlder(_span, verification_mark_bm()); 2310 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2311 2312 { 2313 StrongRootsScope srs(1); 2314 2315 gch->gen_process_roots(&srs, 2316 GenCollectedHeap::OldGen, 2317 true, // young gen as roots 2318 GenCollectedHeap::ScanningOption(roots_scanning_options()), 2319 should_unload_classes(), 2320 ¬Older, 2321 NULL, 2322 NULL); 2323 } 2324 2325 // Now mark from the roots 2326 MarkFromRootsClosure markFromRootsClosure(this, _span, 2327 verification_mark_bm(), verification_mark_stack(), 2328 false /* don't yield */, true /* verifying */); 2329 assert(_restart_addr == NULL, "Expected pre-condition"); 2330 verification_mark_bm()->iterate(&markFromRootsClosure); 2331 while (_restart_addr != NULL) { 2332 // Deal with stack overflow: by restarting at the indicated 2333 // address. 2334 HeapWord* ra = _restart_addr; 2335 markFromRootsClosure.reset(ra); 2336 _restart_addr = NULL; 2337 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end()); 2338 } 2339 assert(verification_mark_stack()->isEmpty(), "Should have been drained"); 2340 verify_work_stacks_empty(); 2341 2342 // Marking completed -- now verify that each bit marked in 2343 // verification_mark_bm() is also marked in markBitMap(); flag all 2344 // errors by printing corresponding objects. 2345 VerifyMarkedClosure vcl(markBitMap()); 2346 verification_mark_bm()->iterate(&vcl); 2347 if (vcl.failed()) { 2348 LogHandle(gc, verify) log; 2349 log.info("Verification failed"); 2350 ResourceMark rm; 2351 gch->print_on(log.info_stream()); 2352 fatal("CMS: failed marking verification after remark"); 2353 } 2354 } 2355 2356 class VerifyKlassOopsKlassClosure : public KlassClosure { 2357 class VerifyKlassOopsClosure : public OopClosure { 2358 CMSBitMap* _bitmap; 2359 public: 2360 VerifyKlassOopsClosure(CMSBitMap* bitmap) : _bitmap(bitmap) { } 2361 void do_oop(oop* p) { guarantee(*p == NULL || _bitmap->isMarked((HeapWord*) *p), "Should be marked"); } 2362 void do_oop(narrowOop* p) { ShouldNotReachHere(); } 2363 } _oop_closure; 2364 public: 2365 VerifyKlassOopsKlassClosure(CMSBitMap* bitmap) : _oop_closure(bitmap) {} 2366 void do_klass(Klass* k) { 2367 k->oops_do(&_oop_closure); 2368 } 2369 }; 2370 2371 void CMSCollector::verify_after_remark_work_2() { 2372 ResourceMark rm; 2373 HandleMark hm; 2374 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2375 2376 // Get a clear set of claim bits for the roots processing to work with. 2377 ClassLoaderDataGraph::clear_claimed_marks(); 2378 2379 // Mark from roots one level into CMS 2380 MarkRefsIntoVerifyClosure notOlder(_span, verification_mark_bm(), 2381 markBitMap()); 2382 CLDToOopClosure cld_closure(¬Older, true); 2383 2384 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2385 2386 { 2387 StrongRootsScope srs(1); 2388 2389 gch->gen_process_roots(&srs, 2390 GenCollectedHeap::OldGen, 2391 true, // young gen as roots 2392 GenCollectedHeap::ScanningOption(roots_scanning_options()), 2393 should_unload_classes(), 2394 ¬Older, 2395 NULL, 2396 &cld_closure); 2397 } 2398 2399 // Now mark from the roots 2400 MarkFromRootsVerifyClosure markFromRootsClosure(this, _span, 2401 verification_mark_bm(), markBitMap(), verification_mark_stack()); 2402 assert(_restart_addr == NULL, "Expected pre-condition"); 2403 verification_mark_bm()->iterate(&markFromRootsClosure); 2404 while (_restart_addr != NULL) { 2405 // Deal with stack overflow: by restarting at the indicated 2406 // address. 2407 HeapWord* ra = _restart_addr; 2408 markFromRootsClosure.reset(ra); 2409 _restart_addr = NULL; 2410 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end()); 2411 } 2412 assert(verification_mark_stack()->isEmpty(), "Should have been drained"); 2413 verify_work_stacks_empty(); 2414 2415 VerifyKlassOopsKlassClosure verify_klass_oops(verification_mark_bm()); 2416 ClassLoaderDataGraph::classes_do(&verify_klass_oops); 2417 2418 // Marking completed -- now verify that each bit marked in 2419 // verification_mark_bm() is also marked in markBitMap(); flag all 2420 // errors by printing corresponding objects. 2421 VerifyMarkedClosure vcl(markBitMap()); 2422 verification_mark_bm()->iterate(&vcl); 2423 assert(!vcl.failed(), "Else verification above should not have succeeded"); 2424 } 2425 2426 void ConcurrentMarkSweepGeneration::save_marks() { 2427 // delegate to CMS space 2428 cmsSpace()->save_marks(); 2429 for (uint i = 0; i < ParallelGCThreads; i++) { 2430 _par_gc_thread_states[i]->promo.startTrackingPromotions(); 2431 } 2432 } 2433 2434 bool ConcurrentMarkSweepGeneration::no_allocs_since_save_marks() { 2435 return cmsSpace()->no_allocs_since_save_marks(); 2436 } 2437 2438 #define CMS_SINCE_SAVE_MARKS_DEFN(OopClosureType, nv_suffix) \ 2439 \ 2440 void ConcurrentMarkSweepGeneration:: \ 2441 oop_since_save_marks_iterate##nv_suffix(OopClosureType* cl) { \ 2442 cl->set_generation(this); \ 2443 cmsSpace()->oop_since_save_marks_iterate##nv_suffix(cl); \ 2444 cl->reset_generation(); \ 2445 save_marks(); \ 2446 } 2447 2448 ALL_SINCE_SAVE_MARKS_CLOSURES(CMS_SINCE_SAVE_MARKS_DEFN) 2449 2450 void 2451 ConcurrentMarkSweepGeneration::oop_iterate(ExtendedOopClosure* cl) { 2452 if (freelistLock()->owned_by_self()) { 2453 Generation::oop_iterate(cl); 2454 } else { 2455 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2456 Generation::oop_iterate(cl); 2457 } 2458 } 2459 2460 void 2461 ConcurrentMarkSweepGeneration::object_iterate(ObjectClosure* cl) { 2462 if (freelistLock()->owned_by_self()) { 2463 Generation::object_iterate(cl); 2464 } else { 2465 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2466 Generation::object_iterate(cl); 2467 } 2468 } 2469 2470 void 2471 ConcurrentMarkSweepGeneration::safe_object_iterate(ObjectClosure* cl) { 2472 if (freelistLock()->owned_by_self()) { 2473 Generation::safe_object_iterate(cl); 2474 } else { 2475 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2476 Generation::safe_object_iterate(cl); 2477 } 2478 } 2479 2480 void 2481 ConcurrentMarkSweepGeneration::post_compact() { 2482 } 2483 2484 void 2485 ConcurrentMarkSweepGeneration::prepare_for_verify() { 2486 // Fix the linear allocation blocks to look like free blocks. 2487 2488 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those 2489 // are not called when the heap is verified during universe initialization and 2490 // at vm shutdown. 2491 if (freelistLock()->owned_by_self()) { 2492 cmsSpace()->prepare_for_verify(); 2493 } else { 2494 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag); 2495 cmsSpace()->prepare_for_verify(); 2496 } 2497 } 2498 2499 void 2500 ConcurrentMarkSweepGeneration::verify() { 2501 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those 2502 // are not called when the heap is verified during universe initialization and 2503 // at vm shutdown. 2504 if (freelistLock()->owned_by_self()) { 2505 cmsSpace()->verify(); 2506 } else { 2507 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag); 2508 cmsSpace()->verify(); 2509 } 2510 } 2511 2512 void CMSCollector::verify() { 2513 _cmsGen->verify(); 2514 } 2515 2516 #ifndef PRODUCT 2517 bool CMSCollector::overflow_list_is_empty() const { 2518 assert(_num_par_pushes >= 0, "Inconsistency"); 2519 if (_overflow_list == NULL) { 2520 assert(_num_par_pushes == 0, "Inconsistency"); 2521 } 2522 return _overflow_list == NULL; 2523 } 2524 2525 // The methods verify_work_stacks_empty() and verify_overflow_empty() 2526 // merely consolidate assertion checks that appear to occur together frequently. 2527 void CMSCollector::verify_work_stacks_empty() const { 2528 assert(_markStack.isEmpty(), "Marking stack should be empty"); 2529 assert(overflow_list_is_empty(), "Overflow list should be empty"); 2530 } 2531 2532 void CMSCollector::verify_overflow_empty() const { 2533 assert(overflow_list_is_empty(), "Overflow list should be empty"); 2534 assert(no_preserved_marks(), "No preserved marks"); 2535 } 2536 #endif // PRODUCT 2537 2538 // Decide if we want to enable class unloading as part of the 2539 // ensuing concurrent GC cycle. We will collect and 2540 // unload classes if it's the case that: 2541 // (1) an explicit gc request has been made and the flag 2542 // ExplicitGCInvokesConcurrentAndUnloadsClasses is set, OR 2543 // (2) (a) class unloading is enabled at the command line, and 2544 // (b) old gen is getting really full 2545 // NOTE: Provided there is no change in the state of the heap between 2546 // calls to this method, it should have idempotent results. Moreover, 2547 // its results should be monotonically increasing (i.e. going from 0 to 1, 2548 // but not 1 to 0) between successive calls between which the heap was 2549 // not collected. For the implementation below, it must thus rely on 2550 // the property that concurrent_cycles_since_last_unload() 2551 // will not decrease unless a collection cycle happened and that 2552 // _cmsGen->is_too_full() are 2553 // themselves also monotonic in that sense. See check_monotonicity() 2554 // below. 2555 void CMSCollector::update_should_unload_classes() { 2556 _should_unload_classes = false; 2557 // Condition 1 above 2558 if (_full_gc_requested && ExplicitGCInvokesConcurrentAndUnloadsClasses) { 2559 _should_unload_classes = true; 2560 } else if (CMSClassUnloadingEnabled) { // Condition 2.a above 2561 // Disjuncts 2.b.(i,ii,iii) above 2562 _should_unload_classes = (concurrent_cycles_since_last_unload() >= 2563 CMSClassUnloadingMaxInterval) 2564 || _cmsGen->is_too_full(); 2565 } 2566 } 2567 2568 bool ConcurrentMarkSweepGeneration::is_too_full() const { 2569 bool res = should_concurrent_collect(); 2570 res = res && (occupancy() > (double)CMSIsTooFullPercentage/100.0); 2571 return res; 2572 } 2573 2574 void CMSCollector::setup_cms_unloading_and_verification_state() { 2575 const bool should_verify = VerifyBeforeGC || VerifyAfterGC || VerifyDuringGC 2576 || VerifyBeforeExit; 2577 const int rso = GenCollectedHeap::SO_AllCodeCache; 2578 2579 // We set the proper root for this CMS cycle here. 2580 if (should_unload_classes()) { // Should unload classes this cycle 2581 remove_root_scanning_option(rso); // Shrink the root set appropriately 2582 set_verifying(should_verify); // Set verification state for this cycle 2583 return; // Nothing else needs to be done at this time 2584 } 2585 2586 // Not unloading classes this cycle 2587 assert(!should_unload_classes(), "Inconsistency!"); 2588 2589 // If we are not unloading classes then add SO_AllCodeCache to root 2590 // scanning options. 2591 add_root_scanning_option(rso); 2592 2593 if ((!verifying() || unloaded_classes_last_cycle()) && should_verify) { 2594 set_verifying(true); 2595 } else if (verifying() && !should_verify) { 2596 // We were verifying, but some verification flags got disabled. 2597 set_verifying(false); 2598 // Exclude symbols, strings and code cache elements from root scanning to 2599 // reduce IM and RM pauses. 2600 remove_root_scanning_option(rso); 2601 } 2602 } 2603 2604 2605 #ifndef PRODUCT 2606 HeapWord* CMSCollector::block_start(const void* p) const { 2607 const HeapWord* addr = (HeapWord*)p; 2608 if (_span.contains(p)) { 2609 if (_cmsGen->cmsSpace()->is_in_reserved(addr)) { 2610 return _cmsGen->cmsSpace()->block_start(p); 2611 } 2612 } 2613 return NULL; 2614 } 2615 #endif 2616 2617 HeapWord* 2618 ConcurrentMarkSweepGeneration::expand_and_allocate(size_t word_size, 2619 bool tlab, 2620 bool parallel) { 2621 CMSSynchronousYieldRequest yr; 2622 assert(!tlab, "Can't deal with TLAB allocation"); 2623 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2624 expand_for_gc_cause(word_size*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_satisfy_allocation); 2625 if (GCExpandToAllocateDelayMillis > 0) { 2626 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 2627 } 2628 return have_lock_and_allocate(word_size, tlab); 2629 } 2630 2631 void ConcurrentMarkSweepGeneration::expand_for_gc_cause( 2632 size_t bytes, 2633 size_t expand_bytes, 2634 CMSExpansionCause::Cause cause) 2635 { 2636 2637 bool success = expand(bytes, expand_bytes); 2638 2639 // remember why we expanded; this information is used 2640 // by shouldConcurrentCollect() when making decisions on whether to start 2641 // a new CMS cycle. 2642 if (success) { 2643 set_expansion_cause(cause); 2644 log_trace(gc)("Expanded CMS gen for %s", CMSExpansionCause::to_string(cause)); 2645 } 2646 } 2647 2648 HeapWord* ConcurrentMarkSweepGeneration::expand_and_par_lab_allocate(CMSParGCThreadState* ps, size_t word_sz) { 2649 HeapWord* res = NULL; 2650 MutexLocker x(ParGCRareEvent_lock); 2651 while (true) { 2652 // Expansion by some other thread might make alloc OK now: 2653 res = ps->lab.alloc(word_sz); 2654 if (res != NULL) return res; 2655 // If there's not enough expansion space available, give up. 2656 if (_virtual_space.uncommitted_size() < (word_sz * HeapWordSize)) { 2657 return NULL; 2658 } 2659 // Otherwise, we try expansion. 2660 expand_for_gc_cause(word_sz*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_allocate_par_lab); 2661 // Now go around the loop and try alloc again; 2662 // A competing par_promote might beat us to the expansion space, 2663 // so we may go around the loop again if promotion fails again. 2664 if (GCExpandToAllocateDelayMillis > 0) { 2665 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 2666 } 2667 } 2668 } 2669 2670 2671 bool ConcurrentMarkSweepGeneration::expand_and_ensure_spooling_space( 2672 PromotionInfo* promo) { 2673 MutexLocker x(ParGCRareEvent_lock); 2674 size_t refill_size_bytes = promo->refillSize() * HeapWordSize; 2675 while (true) { 2676 // Expansion by some other thread might make alloc OK now: 2677 if (promo->ensure_spooling_space()) { 2678 assert(promo->has_spooling_space(), 2679 "Post-condition of successful ensure_spooling_space()"); 2680 return true; 2681 } 2682 // If there's not enough expansion space available, give up. 2683 if (_virtual_space.uncommitted_size() < refill_size_bytes) { 2684 return false; 2685 } 2686 // Otherwise, we try expansion. 2687 expand_for_gc_cause(refill_size_bytes, MinHeapDeltaBytes, CMSExpansionCause::_allocate_par_spooling_space); 2688 // Now go around the loop and try alloc again; 2689 // A competing allocation might beat us to the expansion space, 2690 // so we may go around the loop again if allocation fails again. 2691 if (GCExpandToAllocateDelayMillis > 0) { 2692 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 2693 } 2694 } 2695 } 2696 2697 void ConcurrentMarkSweepGeneration::shrink(size_t bytes) { 2698 // Only shrink if a compaction was done so that all the free space 2699 // in the generation is in a contiguous block at the end. 2700 if (did_compact()) { 2701 CardGeneration::shrink(bytes); 2702 } 2703 } 2704 2705 void ConcurrentMarkSweepGeneration::assert_correct_size_change_locking() { 2706 assert_locked_or_safepoint(Heap_lock); 2707 } 2708 2709 void ConcurrentMarkSweepGeneration::shrink_free_list_by(size_t bytes) { 2710 assert_locked_or_safepoint(Heap_lock); 2711 assert_lock_strong(freelistLock()); 2712 log_trace(gc)("Shrinking of CMS not yet implemented"); 2713 return; 2714 } 2715 2716 2717 // Simple ctor/dtor wrapper for accounting & timer chores around concurrent 2718 // phases. 2719 class CMSPhaseAccounting: public StackObj { 2720 public: 2721 CMSPhaseAccounting(CMSCollector *collector, 2722 const char *title); 2723 ~CMSPhaseAccounting(); 2724 2725 private: 2726 CMSCollector *_collector; 2727 const char *_title; 2728 GCTraceConcTime(Info, gc) _trace_time; 2729 2730 public: 2731 // Not MT-safe; so do not pass around these StackObj's 2732 // where they may be accessed by other threads. 2733 double wallclock_millis() { 2734 return TimeHelper::counter_to_millis(os::elapsed_counter() - _trace_time.start_time()); 2735 } 2736 }; 2737 2738 CMSPhaseAccounting::CMSPhaseAccounting(CMSCollector *collector, 2739 const char *title) : 2740 _collector(collector), _title(title), _trace_time(title) { 2741 2742 _collector->resetYields(); 2743 _collector->resetTimer(); 2744 _collector->startTimer(); 2745 _collector->gc_timer_cm()->register_gc_concurrent_start(title); 2746 } 2747 2748 CMSPhaseAccounting::~CMSPhaseAccounting() { 2749 _collector->gc_timer_cm()->register_gc_concurrent_end(); 2750 _collector->stopTimer(); 2751 log_debug(gc)("Concurrent active time: %.3fms", TimeHelper::counter_to_seconds(_collector->timerTicks())); 2752 log_trace(gc)(" (CMS %s yielded %d times)", _title, _collector->yields()); 2753 } 2754 2755 // CMS work 2756 2757 // The common parts of CMSParInitialMarkTask and CMSParRemarkTask. 2758 class CMSParMarkTask : public AbstractGangTask { 2759 protected: 2760 CMSCollector* _collector; 2761 uint _n_workers; 2762 CMSParMarkTask(const char* name, CMSCollector* collector, uint n_workers) : 2763 AbstractGangTask(name), 2764 _collector(collector), 2765 _n_workers(n_workers) {} 2766 // Work method in support of parallel rescan ... of young gen spaces 2767 void do_young_space_rescan(uint worker_id, OopsInGenClosure* cl, 2768 ContiguousSpace* space, 2769 HeapWord** chunk_array, size_t chunk_top); 2770 void work_on_young_gen_roots(uint worker_id, OopsInGenClosure* cl); 2771 }; 2772 2773 // Parallel initial mark task 2774 class CMSParInitialMarkTask: public CMSParMarkTask { 2775 StrongRootsScope* _strong_roots_scope; 2776 public: 2777 CMSParInitialMarkTask(CMSCollector* collector, StrongRootsScope* strong_roots_scope, uint n_workers) : 2778 CMSParMarkTask("Scan roots and young gen for initial mark in parallel", collector, n_workers), 2779 _strong_roots_scope(strong_roots_scope) {} 2780 void work(uint worker_id); 2781 }; 2782 2783 // Checkpoint the roots into this generation from outside 2784 // this generation. [Note this initial checkpoint need only 2785 // be approximate -- we'll do a catch up phase subsequently.] 2786 void CMSCollector::checkpointRootsInitial() { 2787 assert(_collectorState == InitialMarking, "Wrong collector state"); 2788 check_correct_thread_executing(); 2789 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 2790 2791 save_heap_summary(); 2792 report_heap_summary(GCWhen::BeforeGC); 2793 2794 ReferenceProcessor* rp = ref_processor(); 2795 assert(_restart_addr == NULL, "Control point invariant"); 2796 { 2797 // acquire locks for subsequent manipulations 2798 MutexLockerEx x(bitMapLock(), 2799 Mutex::_no_safepoint_check_flag); 2800 checkpointRootsInitialWork(); 2801 // enable ("weak") refs discovery 2802 rp->enable_discovery(); 2803 _collectorState = Marking; 2804 } 2805 } 2806 2807 void CMSCollector::checkpointRootsInitialWork() { 2808 assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped"); 2809 assert(_collectorState == InitialMarking, "just checking"); 2810 2811 // Already have locks. 2812 assert_lock_strong(bitMapLock()); 2813 assert(_markBitMap.isAllClear(), "was reset at end of previous cycle"); 2814 2815 // Setup the verification and class unloading state for this 2816 // CMS collection cycle. 2817 setup_cms_unloading_and_verification_state(); 2818 2819 GCTraceTime(Trace, gc) ts("checkpointRootsInitialWork", _gc_timer_cm); 2820 2821 // Reset all the PLAB chunk arrays if necessary. 2822 if (_survivor_plab_array != NULL && !CMSPLABRecordAlways) { 2823 reset_survivor_plab_arrays(); 2824 } 2825 2826 ResourceMark rm; 2827 HandleMark hm; 2828 2829 MarkRefsIntoClosure notOlder(_span, &_markBitMap); 2830 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2831 2832 verify_work_stacks_empty(); 2833 verify_overflow_empty(); 2834 2835 gch->ensure_parsability(false); // fill TLABs, but no need to retire them 2836 // Update the saved marks which may affect the root scans. 2837 gch->save_marks(); 2838 2839 // weak reference processing has not started yet. 2840 ref_processor()->set_enqueuing_is_done(false); 2841 2842 // Need to remember all newly created CLDs, 2843 // so that we can guarantee that the remark finds them. 2844 ClassLoaderDataGraph::remember_new_clds(true); 2845 2846 // Whenever a CLD is found, it will be claimed before proceeding to mark 2847 // the klasses. The claimed marks need to be cleared before marking starts. 2848 ClassLoaderDataGraph::clear_claimed_marks(); 2849 2850 print_eden_and_survivor_chunk_arrays(); 2851 2852 { 2853 #if defined(COMPILER2) || INCLUDE_JVMCI 2854 DerivedPointerTableDeactivate dpt_deact; 2855 #endif 2856 if (CMSParallelInitialMarkEnabled) { 2857 // The parallel version. 2858 WorkGang* workers = gch->workers(); 2859 assert(workers != NULL, "Need parallel worker threads."); 2860 uint n_workers = workers->active_workers(); 2861 2862 StrongRootsScope srs(n_workers); 2863 2864 CMSParInitialMarkTask tsk(this, &srs, n_workers); 2865 initialize_sequential_subtasks_for_young_gen_rescan(n_workers); 2866 if (n_workers > 1) { 2867 workers->run_task(&tsk); 2868 } else { 2869 tsk.work(0); 2870 } 2871 } else { 2872 // The serial version. 2873 CLDToOopClosure cld_closure(¬Older, true); 2874 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2875 2876 StrongRootsScope srs(1); 2877 2878 gch->gen_process_roots(&srs, 2879 GenCollectedHeap::OldGen, 2880 true, // young gen as roots 2881 GenCollectedHeap::ScanningOption(roots_scanning_options()), 2882 should_unload_classes(), 2883 ¬Older, 2884 NULL, 2885 &cld_closure); 2886 } 2887 } 2888 2889 // Clear mod-union table; it will be dirtied in the prologue of 2890 // CMS generation per each young generation collection. 2891 2892 assert(_modUnionTable.isAllClear(), 2893 "Was cleared in most recent final checkpoint phase" 2894 " or no bits are set in the gc_prologue before the start of the next " 2895 "subsequent marking phase."); 2896 2897 assert(_ct->klass_rem_set()->mod_union_is_clear(), "Must be"); 2898 2899 // Save the end of the used_region of the constituent generations 2900 // to be used to limit the extent of sweep in each generation. 2901 save_sweep_limits(); 2902 verify_overflow_empty(); 2903 } 2904 2905 bool CMSCollector::markFromRoots() { 2906 // we might be tempted to assert that: 2907 // assert(!SafepointSynchronize::is_at_safepoint(), 2908 // "inconsistent argument?"); 2909 // However that wouldn't be right, because it's possible that 2910 // a safepoint is indeed in progress as a young generation 2911 // stop-the-world GC happens even as we mark in this generation. 2912 assert(_collectorState == Marking, "inconsistent state?"); 2913 check_correct_thread_executing(); 2914 verify_overflow_empty(); 2915 2916 // Weak ref discovery note: We may be discovering weak 2917 // refs in this generation concurrent (but interleaved) with 2918 // weak ref discovery by the young generation collector. 2919 2920 CMSTokenSyncWithLocks ts(true, bitMapLock()); 2921 GCTraceCPUTime tcpu; 2922 CMSPhaseAccounting pa(this, "Concrurrent Mark"); 2923 bool res = markFromRootsWork(); 2924 if (res) { 2925 _collectorState = Precleaning; 2926 } else { // We failed and a foreground collection wants to take over 2927 assert(_foregroundGCIsActive, "internal state inconsistency"); 2928 assert(_restart_addr == NULL, "foreground will restart from scratch"); 2929 log_debug(gc)("bailing out to foreground collection"); 2930 } 2931 verify_overflow_empty(); 2932 return res; 2933 } 2934 2935 bool CMSCollector::markFromRootsWork() { 2936 // iterate over marked bits in bit map, doing a full scan and mark 2937 // from these roots using the following algorithm: 2938 // . if oop is to the right of the current scan pointer, 2939 // mark corresponding bit (we'll process it later) 2940 // . else (oop is to left of current scan pointer) 2941 // push oop on marking stack 2942 // . drain the marking stack 2943 2944 // Note that when we do a marking step we need to hold the 2945 // bit map lock -- recall that direct allocation (by mutators) 2946 // and promotion (by the young generation collector) is also 2947 // marking the bit map. [the so-called allocate live policy.] 2948 // Because the implementation of bit map marking is not 2949 // robust wrt simultaneous marking of bits in the same word, 2950 // we need to make sure that there is no such interference 2951 // between concurrent such updates. 2952 2953 // already have locks 2954 assert_lock_strong(bitMapLock()); 2955 2956 verify_work_stacks_empty(); 2957 verify_overflow_empty(); 2958 bool result = false; 2959 if (CMSConcurrentMTEnabled && ConcGCThreads > 0) { 2960 result = do_marking_mt(); 2961 } else { 2962 result = do_marking_st(); 2963 } 2964 return result; 2965 } 2966 2967 // Forward decl 2968 class CMSConcMarkingTask; 2969 2970 class CMSConcMarkingTerminator: public ParallelTaskTerminator { 2971 CMSCollector* _collector; 2972 CMSConcMarkingTask* _task; 2973 public: 2974 virtual void yield(); 2975 2976 // "n_threads" is the number of threads to be terminated. 2977 // "queue_set" is a set of work queues of other threads. 2978 // "collector" is the CMS collector associated with this task terminator. 2979 // "yield" indicates whether we need the gang as a whole to yield. 2980 CMSConcMarkingTerminator(int n_threads, TaskQueueSetSuper* queue_set, CMSCollector* collector) : 2981 ParallelTaskTerminator(n_threads, queue_set), 2982 _collector(collector) { } 2983 2984 void set_task(CMSConcMarkingTask* task) { 2985 _task = task; 2986 } 2987 }; 2988 2989 class CMSConcMarkingTerminatorTerminator: public TerminatorTerminator { 2990 CMSConcMarkingTask* _task; 2991 public: 2992 bool should_exit_termination(); 2993 void set_task(CMSConcMarkingTask* task) { 2994 _task = task; 2995 } 2996 }; 2997 2998 // MT Concurrent Marking Task 2999 class CMSConcMarkingTask: public YieldingFlexibleGangTask { 3000 CMSCollector* _collector; 3001 uint _n_workers; // requested/desired # workers 3002 bool _result; 3003 CompactibleFreeListSpace* _cms_space; 3004 char _pad_front[64]; // padding to ... 3005 HeapWord* _global_finger; // ... avoid sharing cache line 3006 char _pad_back[64]; 3007 HeapWord* _restart_addr; 3008 3009 // Exposed here for yielding support 3010 Mutex* const _bit_map_lock; 3011 3012 // The per thread work queues, available here for stealing 3013 OopTaskQueueSet* _task_queues; 3014 3015 // Termination (and yielding) support 3016 CMSConcMarkingTerminator _term; 3017 CMSConcMarkingTerminatorTerminator _term_term; 3018 3019 public: 3020 CMSConcMarkingTask(CMSCollector* collector, 3021 CompactibleFreeListSpace* cms_space, 3022 YieldingFlexibleWorkGang* workers, 3023 OopTaskQueueSet* task_queues): 3024 YieldingFlexibleGangTask("Concurrent marking done multi-threaded"), 3025 _collector(collector), 3026 _cms_space(cms_space), 3027 _n_workers(0), _result(true), 3028 _task_queues(task_queues), 3029 _term(_n_workers, task_queues, _collector), 3030 _bit_map_lock(collector->bitMapLock()) 3031 { 3032 _requested_size = _n_workers; 3033 _term.set_task(this); 3034 _term_term.set_task(this); 3035 _restart_addr = _global_finger = _cms_space->bottom(); 3036 } 3037 3038 3039 OopTaskQueueSet* task_queues() { return _task_queues; } 3040 3041 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 3042 3043 HeapWord** global_finger_addr() { return &_global_finger; } 3044 3045 CMSConcMarkingTerminator* terminator() { return &_term; } 3046 3047 virtual void set_for_termination(uint active_workers) { 3048 terminator()->reset_for_reuse(active_workers); 3049 } 3050 3051 void work(uint worker_id); 3052 bool should_yield() { 3053 return ConcurrentMarkSweepThread::should_yield() 3054 && !_collector->foregroundGCIsActive(); 3055 } 3056 3057 virtual void coordinator_yield(); // stuff done by coordinator 3058 bool result() { return _result; } 3059 3060 void reset(HeapWord* ra) { 3061 assert(_global_finger >= _cms_space->end(), "Postcondition of ::work(i)"); 3062 _restart_addr = _global_finger = ra; 3063 _term.reset_for_reuse(); 3064 } 3065 3066 static bool get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 3067 OopTaskQueue* work_q); 3068 3069 private: 3070 void do_scan_and_mark(int i, CompactibleFreeListSpace* sp); 3071 void do_work_steal(int i); 3072 void bump_global_finger(HeapWord* f); 3073 }; 3074 3075 bool CMSConcMarkingTerminatorTerminator::should_exit_termination() { 3076 assert(_task != NULL, "Error"); 3077 return _task->yielding(); 3078 // Note that we do not need the disjunct || _task->should_yield() above 3079 // because we want terminating threads to yield only if the task 3080 // is already in the midst of yielding, which happens only after at least one 3081 // thread has yielded. 3082 } 3083 3084 void CMSConcMarkingTerminator::yield() { 3085 if (_task->should_yield()) { 3086 _task->yield(); 3087 } else { 3088 ParallelTaskTerminator::yield(); 3089 } 3090 } 3091 3092 //////////////////////////////////////////////////////////////// 3093 // Concurrent Marking Algorithm Sketch 3094 //////////////////////////////////////////////////////////////// 3095 // Until all tasks exhausted (both spaces): 3096 // -- claim next available chunk 3097 // -- bump global finger via CAS 3098 // -- find first object that starts in this chunk 3099 // and start scanning bitmap from that position 3100 // -- scan marked objects for oops 3101 // -- CAS-mark target, and if successful: 3102 // . if target oop is above global finger (volatile read) 3103 // nothing to do 3104 // . if target oop is in chunk and above local finger 3105 // then nothing to do 3106 // . else push on work-queue 3107 // -- Deal with possible overflow issues: 3108 // . local work-queue overflow causes stuff to be pushed on 3109 // global (common) overflow queue 3110 // . always first empty local work queue 3111 // . then get a batch of oops from global work queue if any 3112 // . then do work stealing 3113 // -- When all tasks claimed (both spaces) 3114 // and local work queue empty, 3115 // then in a loop do: 3116 // . check global overflow stack; steal a batch of oops and trace 3117 // . try to steal from other threads oif GOS is empty 3118 // . if neither is available, offer termination 3119 // -- Terminate and return result 3120 // 3121 void CMSConcMarkingTask::work(uint worker_id) { 3122 elapsedTimer _timer; 3123 ResourceMark rm; 3124 HandleMark hm; 3125 3126 DEBUG_ONLY(_collector->verify_overflow_empty();) 3127 3128 // Before we begin work, our work queue should be empty 3129 assert(work_queue(worker_id)->size() == 0, "Expected to be empty"); 3130 // Scan the bitmap covering _cms_space, tracing through grey objects. 3131 _timer.start(); 3132 do_scan_and_mark(worker_id, _cms_space); 3133 _timer.stop(); 3134 log_trace(gc, task)("Finished cms space scanning in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 3135 3136 // ... do work stealing 3137 _timer.reset(); 3138 _timer.start(); 3139 do_work_steal(worker_id); 3140 _timer.stop(); 3141 log_trace(gc, task)("Finished work stealing in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 3142 assert(_collector->_markStack.isEmpty(), "Should have been emptied"); 3143 assert(work_queue(worker_id)->size() == 0, "Should have been emptied"); 3144 // Note that under the current task protocol, the 3145 // following assertion is true even of the spaces 3146 // expanded since the completion of the concurrent 3147 // marking. XXX This will likely change under a strict 3148 // ABORT semantics. 3149 // After perm removal the comparison was changed to 3150 // greater than or equal to from strictly greater than. 3151 // Before perm removal the highest address sweep would 3152 // have been at the end of perm gen but now is at the 3153 // end of the tenured gen. 3154 assert(_global_finger >= _cms_space->end(), 3155 "All tasks have been completed"); 3156 DEBUG_ONLY(_collector->verify_overflow_empty();) 3157 } 3158 3159 void CMSConcMarkingTask::bump_global_finger(HeapWord* f) { 3160 HeapWord* read = _global_finger; 3161 HeapWord* cur = read; 3162 while (f > read) { 3163 cur = read; 3164 read = (HeapWord*) Atomic::cmpxchg_ptr(f, &_global_finger, cur); 3165 if (cur == read) { 3166 // our cas succeeded 3167 assert(_global_finger >= f, "protocol consistency"); 3168 break; 3169 } 3170 } 3171 } 3172 3173 // This is really inefficient, and should be redone by 3174 // using (not yet available) block-read and -write interfaces to the 3175 // stack and the work_queue. XXX FIX ME !!! 3176 bool CMSConcMarkingTask::get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 3177 OopTaskQueue* work_q) { 3178 // Fast lock-free check 3179 if (ovflw_stk->length() == 0) { 3180 return false; 3181 } 3182 assert(work_q->size() == 0, "Shouldn't steal"); 3183 MutexLockerEx ml(ovflw_stk->par_lock(), 3184 Mutex::_no_safepoint_check_flag); 3185 // Grab up to 1/4 the size of the work queue 3186 size_t num = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 3187 (size_t)ParGCDesiredObjsFromOverflowList); 3188 num = MIN2(num, ovflw_stk->length()); 3189 for (int i = (int) num; i > 0; i--) { 3190 oop cur = ovflw_stk->pop(); 3191 assert(cur != NULL, "Counted wrong?"); 3192 work_q->push(cur); 3193 } 3194 return num > 0; 3195 } 3196 3197 void CMSConcMarkingTask::do_scan_and_mark(int i, CompactibleFreeListSpace* sp) { 3198 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 3199 int n_tasks = pst->n_tasks(); 3200 // We allow that there may be no tasks to do here because 3201 // we are restarting after a stack overflow. 3202 assert(pst->valid() || n_tasks == 0, "Uninitialized use?"); 3203 uint nth_task = 0; 3204 3205 HeapWord* aligned_start = sp->bottom(); 3206 if (sp->used_region().contains(_restart_addr)) { 3207 // Align down to a card boundary for the start of 0th task 3208 // for this space. 3209 aligned_start = 3210 (HeapWord*)align_size_down((uintptr_t)_restart_addr, 3211 CardTableModRefBS::card_size); 3212 } 3213 3214 size_t chunk_size = sp->marking_task_size(); 3215 while (!pst->is_task_claimed(/* reference */ nth_task)) { 3216 // Having claimed the nth task in this space, 3217 // compute the chunk that it corresponds to: 3218 MemRegion span = MemRegion(aligned_start + nth_task*chunk_size, 3219 aligned_start + (nth_task+1)*chunk_size); 3220 // Try and bump the global finger via a CAS; 3221 // note that we need to do the global finger bump 3222 // _before_ taking the intersection below, because 3223 // the task corresponding to that region will be 3224 // deemed done even if the used_region() expands 3225 // because of allocation -- as it almost certainly will 3226 // during start-up while the threads yield in the 3227 // closure below. 3228 HeapWord* finger = span.end(); 3229 bump_global_finger(finger); // atomically 3230 // There are null tasks here corresponding to chunks 3231 // beyond the "top" address of the space. 3232 span = span.intersection(sp->used_region()); 3233 if (!span.is_empty()) { // Non-null task 3234 HeapWord* prev_obj; 3235 assert(!span.contains(_restart_addr) || nth_task == 0, 3236 "Inconsistency"); 3237 if (nth_task == 0) { 3238 // For the 0th task, we'll not need to compute a block_start. 3239 if (span.contains(_restart_addr)) { 3240 // In the case of a restart because of stack overflow, 3241 // we might additionally skip a chunk prefix. 3242 prev_obj = _restart_addr; 3243 } else { 3244 prev_obj = span.start(); 3245 } 3246 } else { 3247 // We want to skip the first object because 3248 // the protocol is to scan any object in its entirety 3249 // that _starts_ in this span; a fortiori, any 3250 // object starting in an earlier span is scanned 3251 // as part of an earlier claimed task. 3252 // Below we use the "careful" version of block_start 3253 // so we do not try to navigate uninitialized objects. 3254 prev_obj = sp->block_start_careful(span.start()); 3255 // Below we use a variant of block_size that uses the 3256 // Printezis bits to avoid waiting for allocated 3257 // objects to become initialized/parsable. 3258 while (prev_obj < span.start()) { 3259 size_t sz = sp->block_size_no_stall(prev_obj, _collector); 3260 if (sz > 0) { 3261 prev_obj += sz; 3262 } else { 3263 // In this case we may end up doing a bit of redundant 3264 // scanning, but that appears unavoidable, short of 3265 // locking the free list locks; see bug 6324141. 3266 break; 3267 } 3268 } 3269 } 3270 if (prev_obj < span.end()) { 3271 MemRegion my_span = MemRegion(prev_obj, span.end()); 3272 // Do the marking work within a non-empty span -- 3273 // the last argument to the constructor indicates whether the 3274 // iteration should be incremental with periodic yields. 3275 Par_MarkFromRootsClosure cl(this, _collector, my_span, 3276 &_collector->_markBitMap, 3277 work_queue(i), 3278 &_collector->_markStack); 3279 _collector->_markBitMap.iterate(&cl, my_span.start(), my_span.end()); 3280 } // else nothing to do for this task 3281 } // else nothing to do for this task 3282 } 3283 // We'd be tempted to assert here that since there are no 3284 // more tasks left to claim in this space, the global_finger 3285 // must exceed space->top() and a fortiori space->end(). However, 3286 // that would not quite be correct because the bumping of 3287 // global_finger occurs strictly after the claiming of a task, 3288 // so by the time we reach here the global finger may not yet 3289 // have been bumped up by the thread that claimed the last 3290 // task. 3291 pst->all_tasks_completed(); 3292 } 3293 3294 class Par_ConcMarkingClosure: public MetadataAwareOopClosure { 3295 private: 3296 CMSCollector* _collector; 3297 CMSConcMarkingTask* _task; 3298 MemRegion _span; 3299 CMSBitMap* _bit_map; 3300 CMSMarkStack* _overflow_stack; 3301 OopTaskQueue* _work_queue; 3302 protected: 3303 DO_OOP_WORK_DEFN 3304 public: 3305 Par_ConcMarkingClosure(CMSCollector* collector, CMSConcMarkingTask* task, OopTaskQueue* work_queue, 3306 CMSBitMap* bit_map, CMSMarkStack* overflow_stack): 3307 MetadataAwareOopClosure(collector->ref_processor()), 3308 _collector(collector), 3309 _task(task), 3310 _span(collector->_span), 3311 _work_queue(work_queue), 3312 _bit_map(bit_map), 3313 _overflow_stack(overflow_stack) 3314 { } 3315 virtual void do_oop(oop* p); 3316 virtual void do_oop(narrowOop* p); 3317 3318 void trim_queue(size_t max); 3319 void handle_stack_overflow(HeapWord* lost); 3320 void do_yield_check() { 3321 if (_task->should_yield()) { 3322 _task->yield(); 3323 } 3324 } 3325 }; 3326 3327 // Grey object scanning during work stealing phase -- 3328 // the salient assumption here is that any references 3329 // that are in these stolen objects being scanned must 3330 // already have been initialized (else they would not have 3331 // been published), so we do not need to check for 3332 // uninitialized objects before pushing here. 3333 void Par_ConcMarkingClosure::do_oop(oop obj) { 3334 assert(obj->is_oop_or_null(true), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 3335 HeapWord* addr = (HeapWord*)obj; 3336 // Check if oop points into the CMS generation 3337 // and is not marked 3338 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 3339 // a white object ... 3340 // If we manage to "claim" the object, by being the 3341 // first thread to mark it, then we push it on our 3342 // marking stack 3343 if (_bit_map->par_mark(addr)) { // ... now grey 3344 // push on work queue (grey set) 3345 bool simulate_overflow = false; 3346 NOT_PRODUCT( 3347 if (CMSMarkStackOverflowALot && 3348 _collector->simulate_overflow()) { 3349 // simulate a stack overflow 3350 simulate_overflow = true; 3351 } 3352 ) 3353 if (simulate_overflow || 3354 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) { 3355 // stack overflow 3356 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _overflow_stack->capacity()); 3357 // We cannot assert that the overflow stack is full because 3358 // it may have been emptied since. 3359 assert(simulate_overflow || 3360 _work_queue->size() == _work_queue->max_elems(), 3361 "Else push should have succeeded"); 3362 handle_stack_overflow(addr); 3363 } 3364 } // Else, some other thread got there first 3365 do_yield_check(); 3366 } 3367 } 3368 3369 void Par_ConcMarkingClosure::do_oop(oop* p) { Par_ConcMarkingClosure::do_oop_work(p); } 3370 void Par_ConcMarkingClosure::do_oop(narrowOop* p) { Par_ConcMarkingClosure::do_oop_work(p); } 3371 3372 void Par_ConcMarkingClosure::trim_queue(size_t max) { 3373 while (_work_queue->size() > max) { 3374 oop new_oop; 3375 if (_work_queue->pop_local(new_oop)) { 3376 assert(new_oop->is_oop(), "Should be an oop"); 3377 assert(_bit_map->isMarked((HeapWord*)new_oop), "Grey object"); 3378 assert(_span.contains((HeapWord*)new_oop), "Not in span"); 3379 new_oop->oop_iterate(this); // do_oop() above 3380 do_yield_check(); 3381 } 3382 } 3383 } 3384 3385 // Upon stack overflow, we discard (part of) the stack, 3386 // remembering the least address amongst those discarded 3387 // in CMSCollector's _restart_address. 3388 void Par_ConcMarkingClosure::handle_stack_overflow(HeapWord* lost) { 3389 // We need to do this under a mutex to prevent other 3390 // workers from interfering with the work done below. 3391 MutexLockerEx ml(_overflow_stack->par_lock(), 3392 Mutex::_no_safepoint_check_flag); 3393 // Remember the least grey address discarded 3394 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 3395 _collector->lower_restart_addr(ra); 3396 _overflow_stack->reset(); // discard stack contents 3397 _overflow_stack->expand(); // expand the stack if possible 3398 } 3399 3400 3401 void CMSConcMarkingTask::do_work_steal(int i) { 3402 OopTaskQueue* work_q = work_queue(i); 3403 oop obj_to_scan; 3404 CMSBitMap* bm = &(_collector->_markBitMap); 3405 CMSMarkStack* ovflw = &(_collector->_markStack); 3406 int* seed = _collector->hash_seed(i); 3407 Par_ConcMarkingClosure cl(_collector, this, work_q, bm, ovflw); 3408 while (true) { 3409 cl.trim_queue(0); 3410 assert(work_q->size() == 0, "Should have been emptied above"); 3411 if (get_work_from_overflow_stack(ovflw, work_q)) { 3412 // Can't assert below because the work obtained from the 3413 // overflow stack may already have been stolen from us. 3414 // assert(work_q->size() > 0, "Work from overflow stack"); 3415 continue; 3416 } else if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 3417 assert(obj_to_scan->is_oop(), "Should be an oop"); 3418 assert(bm->isMarked((HeapWord*)obj_to_scan), "Grey object"); 3419 obj_to_scan->oop_iterate(&cl); 3420 } else if (terminator()->offer_termination(&_term_term)) { 3421 assert(work_q->size() == 0, "Impossible!"); 3422 break; 3423 } else if (yielding() || should_yield()) { 3424 yield(); 3425 } 3426 } 3427 } 3428 3429 // This is run by the CMS (coordinator) thread. 3430 void CMSConcMarkingTask::coordinator_yield() { 3431 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 3432 "CMS thread should hold CMS token"); 3433 // First give up the locks, then yield, then re-lock 3434 // We should probably use a constructor/destructor idiom to 3435 // do this unlock/lock or modify the MutexUnlocker class to 3436 // serve our purpose. XXX 3437 assert_lock_strong(_bit_map_lock); 3438 _bit_map_lock->unlock(); 3439 ConcurrentMarkSweepThread::desynchronize(true); 3440 _collector->stopTimer(); 3441 _collector->incrementYields(); 3442 3443 // It is possible for whichever thread initiated the yield request 3444 // not to get a chance to wake up and take the bitmap lock between 3445 // this thread releasing it and reacquiring it. So, while the 3446 // should_yield() flag is on, let's sleep for a bit to give the 3447 // other thread a chance to wake up. The limit imposed on the number 3448 // of iterations is defensive, to avoid any unforseen circumstances 3449 // putting us into an infinite loop. Since it's always been this 3450 // (coordinator_yield()) method that was observed to cause the 3451 // problem, we are using a parameter (CMSCoordinatorYieldSleepCount) 3452 // which is by default non-zero. For the other seven methods that 3453 // also perform the yield operation, as are using a different 3454 // parameter (CMSYieldSleepCount) which is by default zero. This way we 3455 // can enable the sleeping for those methods too, if necessary. 3456 // See 6442774. 3457 // 3458 // We really need to reconsider the synchronization between the GC 3459 // thread and the yield-requesting threads in the future and we 3460 // should really use wait/notify, which is the recommended 3461 // way of doing this type of interaction. Additionally, we should 3462 // consolidate the eight methods that do the yield operation and they 3463 // are almost identical into one for better maintainability and 3464 // readability. See 6445193. 3465 // 3466 // Tony 2006.06.29 3467 for (unsigned i = 0; i < CMSCoordinatorYieldSleepCount && 3468 ConcurrentMarkSweepThread::should_yield() && 3469 !CMSCollector::foregroundGCIsActive(); ++i) { 3470 os::sleep(Thread::current(), 1, false); 3471 } 3472 3473 ConcurrentMarkSweepThread::synchronize(true); 3474 _bit_map_lock->lock_without_safepoint_check(); 3475 _collector->startTimer(); 3476 } 3477 3478 bool CMSCollector::do_marking_mt() { 3479 assert(ConcGCThreads > 0 && conc_workers() != NULL, "precondition"); 3480 uint num_workers = AdaptiveSizePolicy::calc_active_conc_workers(conc_workers()->total_workers(), 3481 conc_workers()->active_workers(), 3482 Threads::number_of_non_daemon_threads()); 3483 conc_workers()->set_active_workers(num_workers); 3484 3485 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 3486 3487 CMSConcMarkingTask tsk(this, 3488 cms_space, 3489 conc_workers(), 3490 task_queues()); 3491 3492 // Since the actual number of workers we get may be different 3493 // from the number we requested above, do we need to do anything different 3494 // below? In particular, may be we need to subclass the SequantialSubTasksDone 3495 // class?? XXX 3496 cms_space ->initialize_sequential_subtasks_for_marking(num_workers); 3497 3498 // Refs discovery is already non-atomic. 3499 assert(!ref_processor()->discovery_is_atomic(), "Should be non-atomic"); 3500 assert(ref_processor()->discovery_is_mt(), "Discovery should be MT"); 3501 conc_workers()->start_task(&tsk); 3502 while (tsk.yielded()) { 3503 tsk.coordinator_yield(); 3504 conc_workers()->continue_task(&tsk); 3505 } 3506 // If the task was aborted, _restart_addr will be non-NULL 3507 assert(tsk.completed() || _restart_addr != NULL, "Inconsistency"); 3508 while (_restart_addr != NULL) { 3509 // XXX For now we do not make use of ABORTED state and have not 3510 // yet implemented the right abort semantics (even in the original 3511 // single-threaded CMS case). That needs some more investigation 3512 // and is deferred for now; see CR# TBF. 07252005YSR. XXX 3513 assert(!CMSAbortSemantics || tsk.aborted(), "Inconsistency"); 3514 // If _restart_addr is non-NULL, a marking stack overflow 3515 // occurred; we need to do a fresh marking iteration from the 3516 // indicated restart address. 3517 if (_foregroundGCIsActive) { 3518 // We may be running into repeated stack overflows, having 3519 // reached the limit of the stack size, while making very 3520 // slow forward progress. It may be best to bail out and 3521 // let the foreground collector do its job. 3522 // Clear _restart_addr, so that foreground GC 3523 // works from scratch. This avoids the headache of 3524 // a "rescan" which would otherwise be needed because 3525 // of the dirty mod union table & card table. 3526 _restart_addr = NULL; 3527 return false; 3528 } 3529 // Adjust the task to restart from _restart_addr 3530 tsk.reset(_restart_addr); 3531 cms_space ->initialize_sequential_subtasks_for_marking(num_workers, 3532 _restart_addr); 3533 _restart_addr = NULL; 3534 // Get the workers going again 3535 conc_workers()->start_task(&tsk); 3536 while (tsk.yielded()) { 3537 tsk.coordinator_yield(); 3538 conc_workers()->continue_task(&tsk); 3539 } 3540 } 3541 assert(tsk.completed(), "Inconsistency"); 3542 assert(tsk.result() == true, "Inconsistency"); 3543 return true; 3544 } 3545 3546 bool CMSCollector::do_marking_st() { 3547 ResourceMark rm; 3548 HandleMark hm; 3549 3550 // Temporarily make refs discovery single threaded (non-MT) 3551 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false); 3552 MarkFromRootsClosure markFromRootsClosure(this, _span, &_markBitMap, 3553 &_markStack, CMSYield); 3554 // the last argument to iterate indicates whether the iteration 3555 // should be incremental with periodic yields. 3556 _markBitMap.iterate(&markFromRootsClosure); 3557 // If _restart_addr is non-NULL, a marking stack overflow 3558 // occurred; we need to do a fresh iteration from the 3559 // indicated restart address. 3560 while (_restart_addr != NULL) { 3561 if (_foregroundGCIsActive) { 3562 // We may be running into repeated stack overflows, having 3563 // reached the limit of the stack size, while making very 3564 // slow forward progress. It may be best to bail out and 3565 // let the foreground collector do its job. 3566 // Clear _restart_addr, so that foreground GC 3567 // works from scratch. This avoids the headache of 3568 // a "rescan" which would otherwise be needed because 3569 // of the dirty mod union table & card table. 3570 _restart_addr = NULL; 3571 return false; // indicating failure to complete marking 3572 } 3573 // Deal with stack overflow: 3574 // we restart marking from _restart_addr 3575 HeapWord* ra = _restart_addr; 3576 markFromRootsClosure.reset(ra); 3577 _restart_addr = NULL; 3578 _markBitMap.iterate(&markFromRootsClosure, ra, _span.end()); 3579 } 3580 return true; 3581 } 3582 3583 void CMSCollector::preclean() { 3584 check_correct_thread_executing(); 3585 assert(Thread::current()->is_ConcurrentGC_thread(), "Wrong thread"); 3586 verify_work_stacks_empty(); 3587 verify_overflow_empty(); 3588 _abort_preclean = false; 3589 if (CMSPrecleaningEnabled) { 3590 if (!CMSEdenChunksRecordAlways) { 3591 _eden_chunk_index = 0; 3592 } 3593 size_t used = get_eden_used(); 3594 size_t capacity = get_eden_capacity(); 3595 // Don't start sampling unless we will get sufficiently 3596 // many samples. 3597 if (used < (capacity/(CMSScheduleRemarkSamplingRatio * 100) 3598 * CMSScheduleRemarkEdenPenetration)) { 3599 _start_sampling = true; 3600 } else { 3601 _start_sampling = false; 3602 } 3603 GCTraceCPUTime tcpu; 3604 CMSPhaseAccounting pa(this, "Concurrent Preclean"); 3605 preclean_work(CMSPrecleanRefLists1, CMSPrecleanSurvivors1); 3606 } 3607 CMSTokenSync x(true); // is cms thread 3608 if (CMSPrecleaningEnabled) { 3609 sample_eden(); 3610 _collectorState = AbortablePreclean; 3611 } else { 3612 _collectorState = FinalMarking; 3613 } 3614 verify_work_stacks_empty(); 3615 verify_overflow_empty(); 3616 } 3617 3618 // Try and schedule the remark such that young gen 3619 // occupancy is CMSScheduleRemarkEdenPenetration %. 3620 void CMSCollector::abortable_preclean() { 3621 check_correct_thread_executing(); 3622 assert(CMSPrecleaningEnabled, "Inconsistent control state"); 3623 assert(_collectorState == AbortablePreclean, "Inconsistent control state"); 3624 3625 // If Eden's current occupancy is below this threshold, 3626 // immediately schedule the remark; else preclean 3627 // past the next scavenge in an effort to 3628 // schedule the pause as described above. By choosing 3629 // CMSScheduleRemarkEdenSizeThreshold >= max eden size 3630 // we will never do an actual abortable preclean cycle. 3631 if (get_eden_used() > CMSScheduleRemarkEdenSizeThreshold) { 3632 GCTraceCPUTime tcpu; 3633 CMSPhaseAccounting pa(this, "Concurrent Abortable Preclean"); 3634 // We need more smarts in the abortable preclean 3635 // loop below to deal with cases where allocation 3636 // in young gen is very very slow, and our precleaning 3637 // is running a losing race against a horde of 3638 // mutators intent on flooding us with CMS updates 3639 // (dirty cards). 3640 // One, admittedly dumb, strategy is to give up 3641 // after a certain number of abortable precleaning loops 3642 // or after a certain maximum time. We want to make 3643 // this smarter in the next iteration. 3644 // XXX FIX ME!!! YSR 3645 size_t loops = 0, workdone = 0, cumworkdone = 0, waited = 0; 3646 while (!(should_abort_preclean() || 3647 ConcurrentMarkSweepThread::should_terminate())) { 3648 workdone = preclean_work(CMSPrecleanRefLists2, CMSPrecleanSurvivors2); 3649 cumworkdone += workdone; 3650 loops++; 3651 // Voluntarily terminate abortable preclean phase if we have 3652 // been at it for too long. 3653 if ((CMSMaxAbortablePrecleanLoops != 0) && 3654 loops >= CMSMaxAbortablePrecleanLoops) { 3655 log_debug(gc)(" CMS: abort preclean due to loops "); 3656 break; 3657 } 3658 if (pa.wallclock_millis() > CMSMaxAbortablePrecleanTime) { 3659 log_debug(gc)(" CMS: abort preclean due to time "); 3660 break; 3661 } 3662 // If we are doing little work each iteration, we should 3663 // take a short break. 3664 if (workdone < CMSAbortablePrecleanMinWorkPerIteration) { 3665 // Sleep for some time, waiting for work to accumulate 3666 stopTimer(); 3667 cmsThread()->wait_on_cms_lock(CMSAbortablePrecleanWaitMillis); 3668 startTimer(); 3669 waited++; 3670 } 3671 } 3672 log_trace(gc)(" [" SIZE_FORMAT " iterations, " SIZE_FORMAT " waits, " SIZE_FORMAT " cards)] ", 3673 loops, waited, cumworkdone); 3674 } 3675 CMSTokenSync x(true); // is cms thread 3676 if (_collectorState != Idling) { 3677 assert(_collectorState == AbortablePreclean, 3678 "Spontaneous state transition?"); 3679 _collectorState = FinalMarking; 3680 } // Else, a foreground collection completed this CMS cycle. 3681 return; 3682 } 3683 3684 // Respond to an Eden sampling opportunity 3685 void CMSCollector::sample_eden() { 3686 // Make sure a young gc cannot sneak in between our 3687 // reading and recording of a sample. 3688 assert(Thread::current()->is_ConcurrentGC_thread(), 3689 "Only the cms thread may collect Eden samples"); 3690 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 3691 "Should collect samples while holding CMS token"); 3692 if (!_start_sampling) { 3693 return; 3694 } 3695 // When CMSEdenChunksRecordAlways is true, the eden chunk array 3696 // is populated by the young generation. 3697 if (_eden_chunk_array != NULL && !CMSEdenChunksRecordAlways) { 3698 if (_eden_chunk_index < _eden_chunk_capacity) { 3699 _eden_chunk_array[_eden_chunk_index] = *_top_addr; // take sample 3700 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr, 3701 "Unexpected state of Eden"); 3702 // We'd like to check that what we just sampled is an oop-start address; 3703 // however, we cannot do that here since the object may not yet have been 3704 // initialized. So we'll instead do the check when we _use_ this sample 3705 // later. 3706 if (_eden_chunk_index == 0 || 3707 (pointer_delta(_eden_chunk_array[_eden_chunk_index], 3708 _eden_chunk_array[_eden_chunk_index-1]) 3709 >= CMSSamplingGrain)) { 3710 _eden_chunk_index++; // commit sample 3711 } 3712 } 3713 } 3714 if ((_collectorState == AbortablePreclean) && !_abort_preclean) { 3715 size_t used = get_eden_used(); 3716 size_t capacity = get_eden_capacity(); 3717 assert(used <= capacity, "Unexpected state of Eden"); 3718 if (used > (capacity/100 * CMSScheduleRemarkEdenPenetration)) { 3719 _abort_preclean = true; 3720 } 3721 } 3722 } 3723 3724 3725 size_t CMSCollector::preclean_work(bool clean_refs, bool clean_survivor) { 3726 assert(_collectorState == Precleaning || 3727 _collectorState == AbortablePreclean, "incorrect state"); 3728 ResourceMark rm; 3729 HandleMark hm; 3730 3731 // Precleaning is currently not MT but the reference processor 3732 // may be set for MT. Disable it temporarily here. 3733 ReferenceProcessor* rp = ref_processor(); 3734 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(rp, false); 3735 3736 // Do one pass of scrubbing the discovered reference lists 3737 // to remove any reference objects with strongly-reachable 3738 // referents. 3739 if (clean_refs) { 3740 CMSPrecleanRefsYieldClosure yield_cl(this); 3741 assert(rp->span().equals(_span), "Spans should be equal"); 3742 CMSKeepAliveClosure keep_alive(this, _span, &_markBitMap, 3743 &_markStack, true /* preclean */); 3744 CMSDrainMarkingStackClosure complete_trace(this, 3745 _span, &_markBitMap, &_markStack, 3746 &keep_alive, true /* preclean */); 3747 3748 // We don't want this step to interfere with a young 3749 // collection because we don't want to take CPU 3750 // or memory bandwidth away from the young GC threads 3751 // (which may be as many as there are CPUs). 3752 // Note that we don't need to protect ourselves from 3753 // interference with mutators because they can't 3754 // manipulate the discovered reference lists nor affect 3755 // the computed reachability of the referents, the 3756 // only properties manipulated by the precleaning 3757 // of these reference lists. 3758 stopTimer(); 3759 CMSTokenSyncWithLocks x(true /* is cms thread */, 3760 bitMapLock()); 3761 startTimer(); 3762 sample_eden(); 3763 3764 // The following will yield to allow foreground 3765 // collection to proceed promptly. XXX YSR: 3766 // The code in this method may need further 3767 // tweaking for better performance and some restructuring 3768 // for cleaner interfaces. 3769 GCTimer *gc_timer = NULL; // Currently not tracing concurrent phases 3770 rp->preclean_discovered_references( 3771 rp->is_alive_non_header(), &keep_alive, &complete_trace, &yield_cl, 3772 gc_timer); 3773 } 3774 3775 if (clean_survivor) { // preclean the active survivor space(s) 3776 PushAndMarkClosure pam_cl(this, _span, ref_processor(), 3777 &_markBitMap, &_modUnionTable, 3778 &_markStack, true /* precleaning phase */); 3779 stopTimer(); 3780 CMSTokenSyncWithLocks ts(true /* is cms thread */, 3781 bitMapLock()); 3782 startTimer(); 3783 unsigned int before_count = 3784 GenCollectedHeap::heap()->total_collections(); 3785 SurvivorSpacePrecleanClosure 3786 sss_cl(this, _span, &_markBitMap, &_markStack, 3787 &pam_cl, before_count, CMSYield); 3788 _young_gen->from()->object_iterate_careful(&sss_cl); 3789 _young_gen->to()->object_iterate_careful(&sss_cl); 3790 } 3791 MarkRefsIntoAndScanClosure 3792 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable, 3793 &_markStack, this, CMSYield, 3794 true /* precleaning phase */); 3795 // CAUTION: The following closure has persistent state that may need to 3796 // be reset upon a decrease in the sequence of addresses it 3797 // processes. 3798 ScanMarkedObjectsAgainCarefullyClosure 3799 smoac_cl(this, _span, 3800 &_markBitMap, &_markStack, &mrias_cl, CMSYield); 3801 3802 // Preclean dirty cards in ModUnionTable and CardTable using 3803 // appropriate convergence criterion; 3804 // repeat CMSPrecleanIter times unless we find that 3805 // we are losing. 3806 assert(CMSPrecleanIter < 10, "CMSPrecleanIter is too large"); 3807 assert(CMSPrecleanNumerator < CMSPrecleanDenominator, 3808 "Bad convergence multiplier"); 3809 assert(CMSPrecleanThreshold >= 100, 3810 "Unreasonably low CMSPrecleanThreshold"); 3811 3812 size_t numIter, cumNumCards, lastNumCards, curNumCards; 3813 for (numIter = 0, cumNumCards = lastNumCards = curNumCards = 0; 3814 numIter < CMSPrecleanIter; 3815 numIter++, lastNumCards = curNumCards, cumNumCards += curNumCards) { 3816 curNumCards = preclean_mod_union_table(_cmsGen, &smoac_cl); 3817 log_trace(gc)(" (modUnionTable: " SIZE_FORMAT " cards)", curNumCards); 3818 // Either there are very few dirty cards, so re-mark 3819 // pause will be small anyway, or our pre-cleaning isn't 3820 // that much faster than the rate at which cards are being 3821 // dirtied, so we might as well stop and re-mark since 3822 // precleaning won't improve our re-mark time by much. 3823 if (curNumCards <= CMSPrecleanThreshold || 3824 (numIter > 0 && 3825 (curNumCards * CMSPrecleanDenominator > 3826 lastNumCards * CMSPrecleanNumerator))) { 3827 numIter++; 3828 cumNumCards += curNumCards; 3829 break; 3830 } 3831 } 3832 3833 preclean_klasses(&mrias_cl, _cmsGen->freelistLock()); 3834 3835 curNumCards = preclean_card_table(_cmsGen, &smoac_cl); 3836 cumNumCards += curNumCards; 3837 log_trace(gc)(" (cardTable: " SIZE_FORMAT " cards, re-scanned " SIZE_FORMAT " cards, " SIZE_FORMAT " iterations)", 3838 curNumCards, cumNumCards, numIter); 3839 return cumNumCards; // as a measure of useful work done 3840 } 3841 3842 // PRECLEANING NOTES: 3843 // Precleaning involves: 3844 // . reading the bits of the modUnionTable and clearing the set bits. 3845 // . For the cards corresponding to the set bits, we scan the 3846 // objects on those cards. This means we need the free_list_lock 3847 // so that we can safely iterate over the CMS space when scanning 3848 // for oops. 3849 // . When we scan the objects, we'll be both reading and setting 3850 // marks in the marking bit map, so we'll need the marking bit map. 3851 // . For protecting _collector_state transitions, we take the CGC_lock. 3852 // Note that any races in the reading of of card table entries by the 3853 // CMS thread on the one hand and the clearing of those entries by the 3854 // VM thread or the setting of those entries by the mutator threads on the 3855 // other are quite benign. However, for efficiency it makes sense to keep 3856 // the VM thread from racing with the CMS thread while the latter is 3857 // dirty card info to the modUnionTable. We therefore also use the 3858 // CGC_lock to protect the reading of the card table and the mod union 3859 // table by the CM thread. 3860 // . We run concurrently with mutator updates, so scanning 3861 // needs to be done carefully -- we should not try to scan 3862 // potentially uninitialized objects. 3863 // 3864 // Locking strategy: While holding the CGC_lock, we scan over and 3865 // reset a maximal dirty range of the mod union / card tables, then lock 3866 // the free_list_lock and bitmap lock to do a full marking, then 3867 // release these locks; and repeat the cycle. This allows for a 3868 // certain amount of fairness in the sharing of these locks between 3869 // the CMS collector on the one hand, and the VM thread and the 3870 // mutators on the other. 3871 3872 // NOTE: preclean_mod_union_table() and preclean_card_table() 3873 // further below are largely identical; if you need to modify 3874 // one of these methods, please check the other method too. 3875 3876 size_t CMSCollector::preclean_mod_union_table( 3877 ConcurrentMarkSweepGeneration* old_gen, 3878 ScanMarkedObjectsAgainCarefullyClosure* cl) { 3879 verify_work_stacks_empty(); 3880 verify_overflow_empty(); 3881 3882 // strategy: starting with the first card, accumulate contiguous 3883 // ranges of dirty cards; clear these cards, then scan the region 3884 // covered by these cards. 3885 3886 // Since all of the MUT is committed ahead, we can just use 3887 // that, in case the generations expand while we are precleaning. 3888 // It might also be fine to just use the committed part of the 3889 // generation, but we might potentially miss cards when the 3890 // generation is rapidly expanding while we are in the midst 3891 // of precleaning. 3892 HeapWord* startAddr = old_gen->reserved().start(); 3893 HeapWord* endAddr = old_gen->reserved().end(); 3894 3895 cl->setFreelistLock(old_gen->freelistLock()); // needed for yielding 3896 3897 size_t numDirtyCards, cumNumDirtyCards; 3898 HeapWord *nextAddr, *lastAddr; 3899 for (cumNumDirtyCards = numDirtyCards = 0, 3900 nextAddr = lastAddr = startAddr; 3901 nextAddr < endAddr; 3902 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 3903 3904 ResourceMark rm; 3905 HandleMark hm; 3906 3907 MemRegion dirtyRegion; 3908 { 3909 stopTimer(); 3910 // Potential yield point 3911 CMSTokenSync ts(true); 3912 startTimer(); 3913 sample_eden(); 3914 // Get dirty region starting at nextOffset (inclusive), 3915 // simultaneously clearing it. 3916 dirtyRegion = 3917 _modUnionTable.getAndClearMarkedRegion(nextAddr, endAddr); 3918 assert(dirtyRegion.start() >= nextAddr, 3919 "returned region inconsistent?"); 3920 } 3921 // Remember where the next search should begin. 3922 // The returned region (if non-empty) is a right open interval, 3923 // so lastOffset is obtained from the right end of that 3924 // interval. 3925 lastAddr = dirtyRegion.end(); 3926 // Should do something more transparent and less hacky XXX 3927 numDirtyCards = 3928 _modUnionTable.heapWordDiffToOffsetDiff(dirtyRegion.word_size()); 3929 3930 // We'll scan the cards in the dirty region (with periodic 3931 // yields for foreground GC as needed). 3932 if (!dirtyRegion.is_empty()) { 3933 assert(numDirtyCards > 0, "consistency check"); 3934 HeapWord* stop_point = NULL; 3935 stopTimer(); 3936 // Potential yield point 3937 CMSTokenSyncWithLocks ts(true, old_gen->freelistLock(), 3938 bitMapLock()); 3939 startTimer(); 3940 { 3941 verify_work_stacks_empty(); 3942 verify_overflow_empty(); 3943 sample_eden(); 3944 stop_point = 3945 old_gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 3946 } 3947 if (stop_point != NULL) { 3948 // The careful iteration stopped early either because it found an 3949 // uninitialized object, or because we were in the midst of an 3950 // "abortable preclean", which should now be aborted. Redirty 3951 // the bits corresponding to the partially-scanned or unscanned 3952 // cards. We'll either restart at the next block boundary or 3953 // abort the preclean. 3954 assert((_collectorState == AbortablePreclean && should_abort_preclean()), 3955 "Should only be AbortablePreclean."); 3956 _modUnionTable.mark_range(MemRegion(stop_point, dirtyRegion.end())); 3957 if (should_abort_preclean()) { 3958 break; // out of preclean loop 3959 } else { 3960 // Compute the next address at which preclean should pick up; 3961 // might need bitMapLock in order to read P-bits. 3962 lastAddr = next_card_start_after_block(stop_point); 3963 } 3964 } 3965 } else { 3966 assert(lastAddr == endAddr, "consistency check"); 3967 assert(numDirtyCards == 0, "consistency check"); 3968 break; 3969 } 3970 } 3971 verify_work_stacks_empty(); 3972 verify_overflow_empty(); 3973 return cumNumDirtyCards; 3974 } 3975 3976 // NOTE: preclean_mod_union_table() above and preclean_card_table() 3977 // below are largely identical; if you need to modify 3978 // one of these methods, please check the other method too. 3979 3980 size_t CMSCollector::preclean_card_table(ConcurrentMarkSweepGeneration* old_gen, 3981 ScanMarkedObjectsAgainCarefullyClosure* cl) { 3982 // strategy: it's similar to precleamModUnionTable above, in that 3983 // we accumulate contiguous ranges of dirty cards, mark these cards 3984 // precleaned, then scan the region covered by these cards. 3985 HeapWord* endAddr = (HeapWord*)(old_gen->_virtual_space.high()); 3986 HeapWord* startAddr = (HeapWord*)(old_gen->_virtual_space.low()); 3987 3988 cl->setFreelistLock(old_gen->freelistLock()); // needed for yielding 3989 3990 size_t numDirtyCards, cumNumDirtyCards; 3991 HeapWord *lastAddr, *nextAddr; 3992 3993 for (cumNumDirtyCards = numDirtyCards = 0, 3994 nextAddr = lastAddr = startAddr; 3995 nextAddr < endAddr; 3996 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 3997 3998 ResourceMark rm; 3999 HandleMark hm; 4000 4001 MemRegion dirtyRegion; 4002 { 4003 // See comments in "Precleaning notes" above on why we 4004 // do this locking. XXX Could the locking overheads be 4005 // too high when dirty cards are sparse? [I don't think so.] 4006 stopTimer(); 4007 CMSTokenSync x(true); // is cms thread 4008 startTimer(); 4009 sample_eden(); 4010 // Get and clear dirty region from card table 4011 dirtyRegion = _ct->ct_bs()->dirty_card_range_after_reset( 4012 MemRegion(nextAddr, endAddr), 4013 true, 4014 CardTableModRefBS::precleaned_card_val()); 4015 4016 assert(dirtyRegion.start() >= nextAddr, 4017 "returned region inconsistent?"); 4018 } 4019 lastAddr = dirtyRegion.end(); 4020 numDirtyCards = 4021 dirtyRegion.word_size()/CardTableModRefBS::card_size_in_words; 4022 4023 if (!dirtyRegion.is_empty()) { 4024 stopTimer(); 4025 CMSTokenSyncWithLocks ts(true, old_gen->freelistLock(), bitMapLock()); 4026 startTimer(); 4027 sample_eden(); 4028 verify_work_stacks_empty(); 4029 verify_overflow_empty(); 4030 HeapWord* stop_point = 4031 old_gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 4032 if (stop_point != NULL) { 4033 assert((_collectorState == AbortablePreclean && should_abort_preclean()), 4034 "Should only be AbortablePreclean."); 4035 _ct->ct_bs()->invalidate(MemRegion(stop_point, dirtyRegion.end())); 4036 if (should_abort_preclean()) { 4037 break; // out of preclean loop 4038 } else { 4039 // Compute the next address at which preclean should pick up. 4040 lastAddr = next_card_start_after_block(stop_point); 4041 } 4042 } 4043 } else { 4044 break; 4045 } 4046 } 4047 verify_work_stacks_empty(); 4048 verify_overflow_empty(); 4049 return cumNumDirtyCards; 4050 } 4051 4052 class PrecleanKlassClosure : public KlassClosure { 4053 KlassToOopClosure _cm_klass_closure; 4054 public: 4055 PrecleanKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {} 4056 void do_klass(Klass* k) { 4057 if (k->has_accumulated_modified_oops()) { 4058 k->clear_accumulated_modified_oops(); 4059 4060 _cm_klass_closure.do_klass(k); 4061 } 4062 } 4063 }; 4064 4065 // The freelist lock is needed to prevent asserts, is it really needed? 4066 void CMSCollector::preclean_klasses(MarkRefsIntoAndScanClosure* cl, Mutex* freelistLock) { 4067 4068 cl->set_freelistLock(freelistLock); 4069 4070 CMSTokenSyncWithLocks ts(true, freelistLock, bitMapLock()); 4071 4072 // SSS: Add equivalent to ScanMarkedObjectsAgainCarefullyClosure::do_yield_check and should_abort_preclean? 4073 // SSS: We should probably check if precleaning should be aborted, at suitable intervals? 4074 PrecleanKlassClosure preclean_klass_closure(cl); 4075 ClassLoaderDataGraph::classes_do(&preclean_klass_closure); 4076 4077 verify_work_stacks_empty(); 4078 verify_overflow_empty(); 4079 } 4080 4081 void CMSCollector::checkpointRootsFinal() { 4082 assert(_collectorState == FinalMarking, "incorrect state transition?"); 4083 check_correct_thread_executing(); 4084 // world is stopped at this checkpoint 4085 assert(SafepointSynchronize::is_at_safepoint(), 4086 "world should be stopped"); 4087 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 4088 4089 verify_work_stacks_empty(); 4090 verify_overflow_empty(); 4091 4092 log_debug(gc)("YG occupancy: " SIZE_FORMAT " K (" SIZE_FORMAT " K)", 4093 _young_gen->used() / K, _young_gen->capacity() / K); 4094 { 4095 if (CMSScavengeBeforeRemark) { 4096 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4097 // Temporarily set flag to false, GCH->do_collection will 4098 // expect it to be false and set to true 4099 FlagSetting fl(gch->_is_gc_active, false); 4100 4101 GCTraceTime(Trace, gc) tm("Pause Scavenge Before Remark", _gc_timer_cm); 4102 4103 gch->do_collection(true, // full (i.e. force, see below) 4104 false, // !clear_all_soft_refs 4105 0, // size 4106 false, // is_tlab 4107 GenCollectedHeap::YoungGen // type 4108 ); 4109 } 4110 FreelistLocker x(this); 4111 MutexLockerEx y(bitMapLock(), 4112 Mutex::_no_safepoint_check_flag); 4113 checkpointRootsFinalWork(); 4114 } 4115 verify_work_stacks_empty(); 4116 verify_overflow_empty(); 4117 } 4118 4119 void CMSCollector::checkpointRootsFinalWork() { 4120 GCTraceTime(Trace, gc) tm("checkpointRootsFinalWork", _gc_timer_cm); 4121 4122 assert(haveFreelistLocks(), "must have free list locks"); 4123 assert_lock_strong(bitMapLock()); 4124 4125 ResourceMark rm; 4126 HandleMark hm; 4127 4128 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4129 4130 if (should_unload_classes()) { 4131 CodeCache::gc_prologue(); 4132 } 4133 assert(haveFreelistLocks(), "must have free list locks"); 4134 assert_lock_strong(bitMapLock()); 4135 4136 // We might assume that we need not fill TLAB's when 4137 // CMSScavengeBeforeRemark is set, because we may have just done 4138 // a scavenge which would have filled all TLAB's -- and besides 4139 // Eden would be empty. This however may not always be the case -- 4140 // for instance although we asked for a scavenge, it may not have 4141 // happened because of a JNI critical section. We probably need 4142 // a policy for deciding whether we can in that case wait until 4143 // the critical section releases and then do the remark following 4144 // the scavenge, and skip it here. In the absence of that policy, 4145 // or of an indication of whether the scavenge did indeed occur, 4146 // we cannot rely on TLAB's having been filled and must do 4147 // so here just in case a scavenge did not happen. 4148 gch->ensure_parsability(false); // fill TLAB's, but no need to retire them 4149 // Update the saved marks which may affect the root scans. 4150 gch->save_marks(); 4151 4152 print_eden_and_survivor_chunk_arrays(); 4153 4154 { 4155 #if defined(COMPILER2) || INCLUDE_JVMCI 4156 DerivedPointerTableDeactivate dpt_deact; 4157 #endif 4158 4159 // Note on the role of the mod union table: 4160 // Since the marker in "markFromRoots" marks concurrently with 4161 // mutators, it is possible for some reachable objects not to have been 4162 // scanned. For instance, an only reference to an object A was 4163 // placed in object B after the marker scanned B. Unless B is rescanned, 4164 // A would be collected. Such updates to references in marked objects 4165 // are detected via the mod union table which is the set of all cards 4166 // dirtied since the first checkpoint in this GC cycle and prior to 4167 // the most recent young generation GC, minus those cleaned up by the 4168 // concurrent precleaning. 4169 if (CMSParallelRemarkEnabled) { 4170 GCTraceTime(Debug, gc) t("Rescan (parallel)", _gc_timer_cm); 4171 do_remark_parallel(); 4172 } else { 4173 GCTraceTime(Debug, gc) t("Rescan (non-parallel)", _gc_timer_cm); 4174 do_remark_non_parallel(); 4175 } 4176 } 4177 verify_work_stacks_empty(); 4178 verify_overflow_empty(); 4179 4180 { 4181 GCTraceTime(Trace, gc) ts("refProcessingWork", _gc_timer_cm); 4182 refProcessingWork(); 4183 } 4184 verify_work_stacks_empty(); 4185 verify_overflow_empty(); 4186 4187 if (should_unload_classes()) { 4188 CodeCache::gc_epilogue(); 4189 } 4190 JvmtiExport::gc_epilogue(); 4191 4192 // If we encountered any (marking stack / work queue) overflow 4193 // events during the current CMS cycle, take appropriate 4194 // remedial measures, where possible, so as to try and avoid 4195 // recurrence of that condition. 4196 assert(_markStack.isEmpty(), "No grey objects"); 4197 size_t ser_ovflw = _ser_pmc_remark_ovflw + _ser_pmc_preclean_ovflw + 4198 _ser_kac_ovflw + _ser_kac_preclean_ovflw; 4199 if (ser_ovflw > 0) { 4200 log_trace(gc)("Marking stack overflow (benign) (pmc_pc=" SIZE_FORMAT ", pmc_rm=" SIZE_FORMAT ", kac=" SIZE_FORMAT ", kac_preclean=" SIZE_FORMAT ")", 4201 _ser_pmc_preclean_ovflw, _ser_pmc_remark_ovflw, _ser_kac_ovflw, _ser_kac_preclean_ovflw); 4202 _markStack.expand(); 4203 _ser_pmc_remark_ovflw = 0; 4204 _ser_pmc_preclean_ovflw = 0; 4205 _ser_kac_preclean_ovflw = 0; 4206 _ser_kac_ovflw = 0; 4207 } 4208 if (_par_pmc_remark_ovflw > 0 || _par_kac_ovflw > 0) { 4209 log_trace(gc)("Work queue overflow (benign) (pmc_rm=" SIZE_FORMAT ", kac=" SIZE_FORMAT ")", 4210 _par_pmc_remark_ovflw, _par_kac_ovflw); 4211 _par_pmc_remark_ovflw = 0; 4212 _par_kac_ovflw = 0; 4213 } 4214 if (_markStack._hit_limit > 0) { 4215 log_trace(gc)(" (benign) Hit max stack size limit (" SIZE_FORMAT ")", 4216 _markStack._hit_limit); 4217 } 4218 if (_markStack._failed_double > 0) { 4219 log_trace(gc)(" (benign) Failed stack doubling (" SIZE_FORMAT "), current capacity " SIZE_FORMAT, 4220 _markStack._failed_double, _markStack.capacity()); 4221 } 4222 _markStack._hit_limit = 0; 4223 _markStack._failed_double = 0; 4224 4225 if ((VerifyAfterGC || VerifyDuringGC) && 4226 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 4227 verify_after_remark(); 4228 } 4229 4230 _gc_tracer_cm->report_object_count_after_gc(&_is_alive_closure); 4231 4232 // Change under the freelistLocks. 4233 _collectorState = Sweeping; 4234 // Call isAllClear() under bitMapLock 4235 assert(_modUnionTable.isAllClear(), 4236 "Should be clear by end of the final marking"); 4237 assert(_ct->klass_rem_set()->mod_union_is_clear(), 4238 "Should be clear by end of the final marking"); 4239 } 4240 4241 void CMSParInitialMarkTask::work(uint worker_id) { 4242 elapsedTimer _timer; 4243 ResourceMark rm; 4244 HandleMark hm; 4245 4246 // ---------- scan from roots -------------- 4247 _timer.start(); 4248 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4249 Par_MarkRefsIntoClosure par_mri_cl(_collector->_span, &(_collector->_markBitMap)); 4250 4251 // ---------- young gen roots -------------- 4252 { 4253 work_on_young_gen_roots(worker_id, &par_mri_cl); 4254 _timer.stop(); 4255 log_trace(gc, task)("Finished young gen initial mark scan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4256 } 4257 4258 // ---------- remaining roots -------------- 4259 _timer.reset(); 4260 _timer.start(); 4261 4262 CLDToOopClosure cld_closure(&par_mri_cl, true); 4263 4264 gch->gen_process_roots(_strong_roots_scope, 4265 GenCollectedHeap::OldGen, 4266 false, // yg was scanned above 4267 GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()), 4268 _collector->should_unload_classes(), 4269 &par_mri_cl, 4270 NULL, 4271 &cld_closure); 4272 assert(_collector->should_unload_classes() 4273 || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache), 4274 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 4275 _timer.stop(); 4276 log_trace(gc, task)("Finished remaining root initial mark scan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4277 } 4278 4279 // Parallel remark task 4280 class CMSParRemarkTask: public CMSParMarkTask { 4281 CompactibleFreeListSpace* _cms_space; 4282 4283 // The per-thread work queues, available here for stealing. 4284 OopTaskQueueSet* _task_queues; 4285 ParallelTaskTerminator _term; 4286 StrongRootsScope* _strong_roots_scope; 4287 4288 public: 4289 // A value of 0 passed to n_workers will cause the number of 4290 // workers to be taken from the active workers in the work gang. 4291 CMSParRemarkTask(CMSCollector* collector, 4292 CompactibleFreeListSpace* cms_space, 4293 uint n_workers, WorkGang* workers, 4294 OopTaskQueueSet* task_queues, 4295 StrongRootsScope* strong_roots_scope): 4296 CMSParMarkTask("Rescan roots and grey objects in parallel", 4297 collector, n_workers), 4298 _cms_space(cms_space), 4299 _task_queues(task_queues), 4300 _term(n_workers, task_queues), 4301 _strong_roots_scope(strong_roots_scope) { } 4302 4303 OopTaskQueueSet* task_queues() { return _task_queues; } 4304 4305 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 4306 4307 ParallelTaskTerminator* terminator() { return &_term; } 4308 uint n_workers() { return _n_workers; } 4309 4310 void work(uint worker_id); 4311 4312 private: 4313 // ... of dirty cards in old space 4314 void do_dirty_card_rescan_tasks(CompactibleFreeListSpace* sp, int i, 4315 Par_MarkRefsIntoAndScanClosure* cl); 4316 4317 // ... work stealing for the above 4318 void do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, int* seed); 4319 }; 4320 4321 class RemarkKlassClosure : public KlassClosure { 4322 KlassToOopClosure _cm_klass_closure; 4323 public: 4324 RemarkKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {} 4325 void do_klass(Klass* k) { 4326 // Check if we have modified any oops in the Klass during the concurrent marking. 4327 if (k->has_accumulated_modified_oops()) { 4328 k->clear_accumulated_modified_oops(); 4329 4330 // We could have transfered the current modified marks to the accumulated marks, 4331 // like we do with the Card Table to Mod Union Table. But it's not really necessary. 4332 } else if (k->has_modified_oops()) { 4333 // Don't clear anything, this info is needed by the next young collection. 4334 } else { 4335 // No modified oops in the Klass. 4336 return; 4337 } 4338 4339 // The klass has modified fields, need to scan the klass. 4340 _cm_klass_closure.do_klass(k); 4341 } 4342 }; 4343 4344 void CMSParMarkTask::work_on_young_gen_roots(uint worker_id, OopsInGenClosure* cl) { 4345 ParNewGeneration* young_gen = _collector->_young_gen; 4346 ContiguousSpace* eden_space = young_gen->eden(); 4347 ContiguousSpace* from_space = young_gen->from(); 4348 ContiguousSpace* to_space = young_gen->to(); 4349 4350 HeapWord** eca = _collector->_eden_chunk_array; 4351 size_t ect = _collector->_eden_chunk_index; 4352 HeapWord** sca = _collector->_survivor_chunk_array; 4353 size_t sct = _collector->_survivor_chunk_index; 4354 4355 assert(ect <= _collector->_eden_chunk_capacity, "out of bounds"); 4356 assert(sct <= _collector->_survivor_chunk_capacity, "out of bounds"); 4357 4358 do_young_space_rescan(worker_id, cl, to_space, NULL, 0); 4359 do_young_space_rescan(worker_id, cl, from_space, sca, sct); 4360 do_young_space_rescan(worker_id, cl, eden_space, eca, ect); 4361 } 4362 4363 // work_queue(i) is passed to the closure 4364 // Par_MarkRefsIntoAndScanClosure. The "i" parameter 4365 // also is passed to do_dirty_card_rescan_tasks() and to 4366 // do_work_steal() to select the i-th task_queue. 4367 4368 void CMSParRemarkTask::work(uint worker_id) { 4369 elapsedTimer _timer; 4370 ResourceMark rm; 4371 HandleMark hm; 4372 4373 // ---------- rescan from roots -------------- 4374 _timer.start(); 4375 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4376 Par_MarkRefsIntoAndScanClosure par_mrias_cl(_collector, 4377 _collector->_span, _collector->ref_processor(), 4378 &(_collector->_markBitMap), 4379 work_queue(worker_id)); 4380 4381 // Rescan young gen roots first since these are likely 4382 // coarsely partitioned and may, on that account, constitute 4383 // the critical path; thus, it's best to start off that 4384 // work first. 4385 // ---------- young gen roots -------------- 4386 { 4387 work_on_young_gen_roots(worker_id, &par_mrias_cl); 4388 _timer.stop(); 4389 log_trace(gc, task)("Finished young gen rescan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4390 } 4391 4392 // ---------- remaining roots -------------- 4393 _timer.reset(); 4394 _timer.start(); 4395 gch->gen_process_roots(_strong_roots_scope, 4396 GenCollectedHeap::OldGen, 4397 false, // yg was scanned above 4398 GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()), 4399 _collector->should_unload_classes(), 4400 &par_mrias_cl, 4401 NULL, 4402 NULL); // The dirty klasses will be handled below 4403 4404 assert(_collector->should_unload_classes() 4405 || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache), 4406 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 4407 _timer.stop(); 4408 log_trace(gc, task)("Finished remaining root rescan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4409 4410 // ---------- unhandled CLD scanning ---------- 4411 if (worker_id == 0) { // Single threaded at the moment. 4412 _timer.reset(); 4413 _timer.start(); 4414 4415 // Scan all new class loader data objects and new dependencies that were 4416 // introduced during concurrent marking. 4417 ResourceMark rm; 4418 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds(); 4419 for (int i = 0; i < array->length(); i++) { 4420 par_mrias_cl.do_cld_nv(array->at(i)); 4421 } 4422 4423 // We don't need to keep track of new CLDs anymore. 4424 ClassLoaderDataGraph::remember_new_clds(false); 4425 4426 _timer.stop(); 4427 log_trace(gc, task)("Finished unhandled CLD scanning work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4428 } 4429 4430 // ---------- dirty klass scanning ---------- 4431 if (worker_id == 0) { // Single threaded at the moment. 4432 _timer.reset(); 4433 _timer.start(); 4434 4435 // Scan all classes that was dirtied during the concurrent marking phase. 4436 RemarkKlassClosure remark_klass_closure(&par_mrias_cl); 4437 ClassLoaderDataGraph::classes_do(&remark_klass_closure); 4438 4439 _timer.stop(); 4440 log_trace(gc, task)("Finished dirty klass scanning work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4441 } 4442 4443 // We might have added oops to ClassLoaderData::_handles during the 4444 // concurrent marking phase. These oops point to newly allocated objects 4445 // that are guaranteed to be kept alive. Either by the direct allocation 4446 // code, or when the young collector processes the roots. Hence, 4447 // we don't have to revisit the _handles block during the remark phase. 4448 4449 // ---------- rescan dirty cards ------------ 4450 _timer.reset(); 4451 _timer.start(); 4452 4453 // Do the rescan tasks for each of the two spaces 4454 // (cms_space) in turn. 4455 // "worker_id" is passed to select the task_queue for "worker_id" 4456 do_dirty_card_rescan_tasks(_cms_space, worker_id, &par_mrias_cl); 4457 _timer.stop(); 4458 log_trace(gc, task)("Finished dirty card rescan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4459 4460 // ---------- steal work from other threads ... 4461 // ---------- ... and drain overflow list. 4462 _timer.reset(); 4463 _timer.start(); 4464 do_work_steal(worker_id, &par_mrias_cl, _collector->hash_seed(worker_id)); 4465 _timer.stop(); 4466 log_trace(gc, task)("Finished work stealing in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4467 } 4468 4469 // Note that parameter "i" is not used. 4470 void 4471 CMSParMarkTask::do_young_space_rescan(uint worker_id, 4472 OopsInGenClosure* cl, ContiguousSpace* space, 4473 HeapWord** chunk_array, size_t chunk_top) { 4474 // Until all tasks completed: 4475 // . claim an unclaimed task 4476 // . compute region boundaries corresponding to task claimed 4477 // using chunk_array 4478 // . par_oop_iterate(cl) over that region 4479 4480 ResourceMark rm; 4481 HandleMark hm; 4482 4483 SequentialSubTasksDone* pst = space->par_seq_tasks(); 4484 4485 uint nth_task = 0; 4486 uint n_tasks = pst->n_tasks(); 4487 4488 if (n_tasks > 0) { 4489 assert(pst->valid(), "Uninitialized use?"); 4490 HeapWord *start, *end; 4491 while (!pst->is_task_claimed(/* reference */ nth_task)) { 4492 // We claimed task # nth_task; compute its boundaries. 4493 if (chunk_top == 0) { // no samples were taken 4494 assert(nth_task == 0 && n_tasks == 1, "Can have only 1 eden task"); 4495 start = space->bottom(); 4496 end = space->top(); 4497 } else if (nth_task == 0) { 4498 start = space->bottom(); 4499 end = chunk_array[nth_task]; 4500 } else if (nth_task < (uint)chunk_top) { 4501 assert(nth_task >= 1, "Control point invariant"); 4502 start = chunk_array[nth_task - 1]; 4503 end = chunk_array[nth_task]; 4504 } else { 4505 assert(nth_task == (uint)chunk_top, "Control point invariant"); 4506 start = chunk_array[chunk_top - 1]; 4507 end = space->top(); 4508 } 4509 MemRegion mr(start, end); 4510 // Verify that mr is in space 4511 assert(mr.is_empty() || space->used_region().contains(mr), 4512 "Should be in space"); 4513 // Verify that "start" is an object boundary 4514 assert(mr.is_empty() || oop(mr.start())->is_oop(), 4515 "Should be an oop"); 4516 space->par_oop_iterate(mr, cl); 4517 } 4518 pst->all_tasks_completed(); 4519 } 4520 } 4521 4522 void 4523 CMSParRemarkTask::do_dirty_card_rescan_tasks( 4524 CompactibleFreeListSpace* sp, int i, 4525 Par_MarkRefsIntoAndScanClosure* cl) { 4526 // Until all tasks completed: 4527 // . claim an unclaimed task 4528 // . compute region boundaries corresponding to task claimed 4529 // . transfer dirty bits ct->mut for that region 4530 // . apply rescanclosure to dirty mut bits for that region 4531 4532 ResourceMark rm; 4533 HandleMark hm; 4534 4535 OopTaskQueue* work_q = work_queue(i); 4536 ModUnionClosure modUnionClosure(&(_collector->_modUnionTable)); 4537 // CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! 4538 // CAUTION: This closure has state that persists across calls to 4539 // the work method dirty_range_iterate_clear() in that it has 4540 // embedded in it a (subtype of) UpwardsObjectClosure. The 4541 // use of that state in the embedded UpwardsObjectClosure instance 4542 // assumes that the cards are always iterated (even if in parallel 4543 // by several threads) in monotonically increasing order per each 4544 // thread. This is true of the implementation below which picks 4545 // card ranges (chunks) in monotonically increasing order globally 4546 // and, a-fortiori, in monotonically increasing order per thread 4547 // (the latter order being a subsequence of the former). 4548 // If the work code below is ever reorganized into a more chaotic 4549 // work-partitioning form than the current "sequential tasks" 4550 // paradigm, the use of that persistent state will have to be 4551 // revisited and modified appropriately. See also related 4552 // bug 4756801 work on which should examine this code to make 4553 // sure that the changes there do not run counter to the 4554 // assumptions made here and necessary for correctness and 4555 // efficiency. Note also that this code might yield inefficient 4556 // behavior in the case of very large objects that span one or 4557 // more work chunks. Such objects would potentially be scanned 4558 // several times redundantly. Work on 4756801 should try and 4559 // address that performance anomaly if at all possible. XXX 4560 MemRegion full_span = _collector->_span; 4561 CMSBitMap* bm = &(_collector->_markBitMap); // shared 4562 MarkFromDirtyCardsClosure 4563 greyRescanClosure(_collector, full_span, // entire span of interest 4564 sp, bm, work_q, cl); 4565 4566 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 4567 assert(pst->valid(), "Uninitialized use?"); 4568 uint nth_task = 0; 4569 const int alignment = CardTableModRefBS::card_size * BitsPerWord; 4570 MemRegion span = sp->used_region(); 4571 HeapWord* start_addr = span.start(); 4572 HeapWord* end_addr = (HeapWord*)round_to((intptr_t)span.end(), 4573 alignment); 4574 const size_t chunk_size = sp->rescan_task_size(); // in HeapWord units 4575 assert((HeapWord*)round_to((intptr_t)start_addr, alignment) == 4576 start_addr, "Check alignment"); 4577 assert((size_t)round_to((intptr_t)chunk_size, alignment) == 4578 chunk_size, "Check alignment"); 4579 4580 while (!pst->is_task_claimed(/* reference */ nth_task)) { 4581 // Having claimed the nth_task, compute corresponding mem-region, 4582 // which is a-fortiori aligned correctly (i.e. at a MUT boundary). 4583 // The alignment restriction ensures that we do not need any 4584 // synchronization with other gang-workers while setting or 4585 // clearing bits in thus chunk of the MUT. 4586 MemRegion this_span = MemRegion(start_addr + nth_task*chunk_size, 4587 start_addr + (nth_task+1)*chunk_size); 4588 // The last chunk's end might be way beyond end of the 4589 // used region. In that case pull back appropriately. 4590 if (this_span.end() > end_addr) { 4591 this_span.set_end(end_addr); 4592 assert(!this_span.is_empty(), "Program logic (calculation of n_tasks)"); 4593 } 4594 // Iterate over the dirty cards covering this chunk, marking them 4595 // precleaned, and setting the corresponding bits in the mod union 4596 // table. Since we have been careful to partition at Card and MUT-word 4597 // boundaries no synchronization is needed between parallel threads. 4598 _collector->_ct->ct_bs()->dirty_card_iterate(this_span, 4599 &modUnionClosure); 4600 4601 // Having transferred these marks into the modUnionTable, 4602 // rescan the marked objects on the dirty cards in the modUnionTable. 4603 // Even if this is at a synchronous collection, the initial marking 4604 // may have been done during an asynchronous collection so there 4605 // may be dirty bits in the mod-union table. 4606 _collector->_modUnionTable.dirty_range_iterate_clear( 4607 this_span, &greyRescanClosure); 4608 _collector->_modUnionTable.verifyNoOneBitsInRange( 4609 this_span.start(), 4610 this_span.end()); 4611 } 4612 pst->all_tasks_completed(); // declare that i am done 4613 } 4614 4615 // . see if we can share work_queues with ParNew? XXX 4616 void 4617 CMSParRemarkTask::do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, 4618 int* seed) { 4619 OopTaskQueue* work_q = work_queue(i); 4620 NOT_PRODUCT(int num_steals = 0;) 4621 oop obj_to_scan; 4622 CMSBitMap* bm = &(_collector->_markBitMap); 4623 4624 while (true) { 4625 // Completely finish any left over work from (an) earlier round(s) 4626 cl->trim_queue(0); 4627 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 4628 (size_t)ParGCDesiredObjsFromOverflowList); 4629 // Now check if there's any work in the overflow list 4630 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads, 4631 // only affects the number of attempts made to get work from the 4632 // overflow list and does not affect the number of workers. Just 4633 // pass ParallelGCThreads so this behavior is unchanged. 4634 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 4635 work_q, 4636 ParallelGCThreads)) { 4637 // found something in global overflow list; 4638 // not yet ready to go stealing work from others. 4639 // We'd like to assert(work_q->size() != 0, ...) 4640 // because we just took work from the overflow list, 4641 // but of course we can't since all of that could have 4642 // been already stolen from us. 4643 // "He giveth and He taketh away." 4644 continue; 4645 } 4646 // Verify that we have no work before we resort to stealing 4647 assert(work_q->size() == 0, "Have work, shouldn't steal"); 4648 // Try to steal from other queues that have work 4649 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 4650 NOT_PRODUCT(num_steals++;) 4651 assert(obj_to_scan->is_oop(), "Oops, not an oop!"); 4652 assert(bm->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 4653 // Do scanning work 4654 obj_to_scan->oop_iterate(cl); 4655 // Loop around, finish this work, and try to steal some more 4656 } else if (terminator()->offer_termination()) { 4657 break; // nirvana from the infinite cycle 4658 } 4659 } 4660 log_develop_trace(gc, task)("\t(%d: stole %d oops)", i, num_steals); 4661 assert(work_q->size() == 0 && _collector->overflow_list_is_empty(), 4662 "Else our work is not yet done"); 4663 } 4664 4665 // Record object boundaries in _eden_chunk_array by sampling the eden 4666 // top in the slow-path eden object allocation code path and record 4667 // the boundaries, if CMSEdenChunksRecordAlways is true. If 4668 // CMSEdenChunksRecordAlways is false, we use the other asynchronous 4669 // sampling in sample_eden() that activates during the part of the 4670 // preclean phase. 4671 void CMSCollector::sample_eden_chunk() { 4672 if (CMSEdenChunksRecordAlways && _eden_chunk_array != NULL) { 4673 if (_eden_chunk_lock->try_lock()) { 4674 // Record a sample. This is the critical section. The contents 4675 // of the _eden_chunk_array have to be non-decreasing in the 4676 // address order. 4677 _eden_chunk_array[_eden_chunk_index] = *_top_addr; 4678 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr, 4679 "Unexpected state of Eden"); 4680 if (_eden_chunk_index == 0 || 4681 ((_eden_chunk_array[_eden_chunk_index] > _eden_chunk_array[_eden_chunk_index-1]) && 4682 (pointer_delta(_eden_chunk_array[_eden_chunk_index], 4683 _eden_chunk_array[_eden_chunk_index-1]) >= CMSSamplingGrain))) { 4684 _eden_chunk_index++; // commit sample 4685 } 4686 _eden_chunk_lock->unlock(); 4687 } 4688 } 4689 } 4690 4691 // Return a thread-local PLAB recording array, as appropriate. 4692 void* CMSCollector::get_data_recorder(int thr_num) { 4693 if (_survivor_plab_array != NULL && 4694 (CMSPLABRecordAlways || 4695 (_collectorState > Marking && _collectorState < FinalMarking))) { 4696 assert(thr_num < (int)ParallelGCThreads, "thr_num is out of bounds"); 4697 ChunkArray* ca = &_survivor_plab_array[thr_num]; 4698 ca->reset(); // clear it so that fresh data is recorded 4699 return (void*) ca; 4700 } else { 4701 return NULL; 4702 } 4703 } 4704 4705 // Reset all the thread-local PLAB recording arrays 4706 void CMSCollector::reset_survivor_plab_arrays() { 4707 for (uint i = 0; i < ParallelGCThreads; i++) { 4708 _survivor_plab_array[i].reset(); 4709 } 4710 } 4711 4712 // Merge the per-thread plab arrays into the global survivor chunk 4713 // array which will provide the partitioning of the survivor space 4714 // for CMS initial scan and rescan. 4715 void CMSCollector::merge_survivor_plab_arrays(ContiguousSpace* surv, 4716 int no_of_gc_threads) { 4717 assert(_survivor_plab_array != NULL, "Error"); 4718 assert(_survivor_chunk_array != NULL, "Error"); 4719 assert(_collectorState == FinalMarking || 4720 (CMSParallelInitialMarkEnabled && _collectorState == InitialMarking), "Error"); 4721 for (int j = 0; j < no_of_gc_threads; j++) { 4722 _cursor[j] = 0; 4723 } 4724 HeapWord* top = surv->top(); 4725 size_t i; 4726 for (i = 0; i < _survivor_chunk_capacity; i++) { // all sca entries 4727 HeapWord* min_val = top; // Higher than any PLAB address 4728 uint min_tid = 0; // position of min_val this round 4729 for (int j = 0; j < no_of_gc_threads; j++) { 4730 ChunkArray* cur_sca = &_survivor_plab_array[j]; 4731 if (_cursor[j] == cur_sca->end()) { 4732 continue; 4733 } 4734 assert(_cursor[j] < cur_sca->end(), "ctl pt invariant"); 4735 HeapWord* cur_val = cur_sca->nth(_cursor[j]); 4736 assert(surv->used_region().contains(cur_val), "Out of bounds value"); 4737 if (cur_val < min_val) { 4738 min_tid = j; 4739 min_val = cur_val; 4740 } else { 4741 assert(cur_val < top, "All recorded addresses should be less"); 4742 } 4743 } 4744 // At this point min_val and min_tid are respectively 4745 // the least address in _survivor_plab_array[j]->nth(_cursor[j]) 4746 // and the thread (j) that witnesses that address. 4747 // We record this address in the _survivor_chunk_array[i] 4748 // and increment _cursor[min_tid] prior to the next round i. 4749 if (min_val == top) { 4750 break; 4751 } 4752 _survivor_chunk_array[i] = min_val; 4753 _cursor[min_tid]++; 4754 } 4755 // We are all done; record the size of the _survivor_chunk_array 4756 _survivor_chunk_index = i; // exclusive: [0, i) 4757 log_trace(gc, survivor)(" (Survivor:" SIZE_FORMAT "chunks) ", i); 4758 // Verify that we used up all the recorded entries 4759 #ifdef ASSERT 4760 size_t total = 0; 4761 for (int j = 0; j < no_of_gc_threads; j++) { 4762 assert(_cursor[j] == _survivor_plab_array[j].end(), "Ctl pt invariant"); 4763 total += _cursor[j]; 4764 } 4765 assert(total == _survivor_chunk_index, "Ctl Pt Invariant"); 4766 // Check that the merged array is in sorted order 4767 if (total > 0) { 4768 for (size_t i = 0; i < total - 1; i++) { 4769 log_develop_trace(gc, survivor)(" (chunk" SIZE_FORMAT ":" INTPTR_FORMAT ") ", 4770 i, p2i(_survivor_chunk_array[i])); 4771 assert(_survivor_chunk_array[i] < _survivor_chunk_array[i+1], 4772 "Not sorted"); 4773 } 4774 } 4775 #endif // ASSERT 4776 } 4777 4778 // Set up the space's par_seq_tasks structure for work claiming 4779 // for parallel initial scan and rescan of young gen. 4780 // See ParRescanTask where this is currently used. 4781 void 4782 CMSCollector:: 4783 initialize_sequential_subtasks_for_young_gen_rescan(int n_threads) { 4784 assert(n_threads > 0, "Unexpected n_threads argument"); 4785 4786 // Eden space 4787 if (!_young_gen->eden()->is_empty()) { 4788 SequentialSubTasksDone* pst = _young_gen->eden()->par_seq_tasks(); 4789 assert(!pst->valid(), "Clobbering existing data?"); 4790 // Each valid entry in [0, _eden_chunk_index) represents a task. 4791 size_t n_tasks = _eden_chunk_index + 1; 4792 assert(n_tasks == 1 || _eden_chunk_array != NULL, "Error"); 4793 // Sets the condition for completion of the subtask (how many threads 4794 // need to finish in order to be done). 4795 pst->set_n_threads(n_threads); 4796 pst->set_n_tasks((int)n_tasks); 4797 } 4798 4799 // Merge the survivor plab arrays into _survivor_chunk_array 4800 if (_survivor_plab_array != NULL) { 4801 merge_survivor_plab_arrays(_young_gen->from(), n_threads); 4802 } else { 4803 assert(_survivor_chunk_index == 0, "Error"); 4804 } 4805 4806 // To space 4807 { 4808 SequentialSubTasksDone* pst = _young_gen->to()->par_seq_tasks(); 4809 assert(!pst->valid(), "Clobbering existing data?"); 4810 // Sets the condition for completion of the subtask (how many threads 4811 // need to finish in order to be done). 4812 pst->set_n_threads(n_threads); 4813 pst->set_n_tasks(1); 4814 assert(pst->valid(), "Error"); 4815 } 4816 4817 // From space 4818 { 4819 SequentialSubTasksDone* pst = _young_gen->from()->par_seq_tasks(); 4820 assert(!pst->valid(), "Clobbering existing data?"); 4821 size_t n_tasks = _survivor_chunk_index + 1; 4822 assert(n_tasks == 1 || _survivor_chunk_array != NULL, "Error"); 4823 // Sets the condition for completion of the subtask (how many threads 4824 // need to finish in order to be done). 4825 pst->set_n_threads(n_threads); 4826 pst->set_n_tasks((int)n_tasks); 4827 assert(pst->valid(), "Error"); 4828 } 4829 } 4830 4831 // Parallel version of remark 4832 void CMSCollector::do_remark_parallel() { 4833 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4834 WorkGang* workers = gch->workers(); 4835 assert(workers != NULL, "Need parallel worker threads."); 4836 // Choose to use the number of GC workers most recently set 4837 // into "active_workers". 4838 uint n_workers = workers->active_workers(); 4839 4840 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 4841 4842 StrongRootsScope srs(n_workers); 4843 4844 CMSParRemarkTask tsk(this, cms_space, n_workers, workers, task_queues(), &srs); 4845 4846 // We won't be iterating over the cards in the card table updating 4847 // the younger_gen cards, so we shouldn't call the following else 4848 // the verification code as well as subsequent younger_refs_iterate 4849 // code would get confused. XXX 4850 // gch->rem_set()->prepare_for_younger_refs_iterate(true); // parallel 4851 4852 // The young gen rescan work will not be done as part of 4853 // process_roots (which currently doesn't know how to 4854 // parallelize such a scan), but rather will be broken up into 4855 // a set of parallel tasks (via the sampling that the [abortable] 4856 // preclean phase did of eden, plus the [two] tasks of 4857 // scanning the [two] survivor spaces. Further fine-grain 4858 // parallelization of the scanning of the survivor spaces 4859 // themselves, and of precleaning of the young gen itself 4860 // is deferred to the future. 4861 initialize_sequential_subtasks_for_young_gen_rescan(n_workers); 4862 4863 // The dirty card rescan work is broken up into a "sequence" 4864 // of parallel tasks (per constituent space) that are dynamically 4865 // claimed by the parallel threads. 4866 cms_space->initialize_sequential_subtasks_for_rescan(n_workers); 4867 4868 // It turns out that even when we're using 1 thread, doing the work in a 4869 // separate thread causes wide variance in run times. We can't help this 4870 // in the multi-threaded case, but we special-case n=1 here to get 4871 // repeatable measurements of the 1-thread overhead of the parallel code. 4872 if (n_workers > 1) { 4873 // Make refs discovery MT-safe, if it isn't already: it may not 4874 // necessarily be so, since it's possible that we are doing 4875 // ST marking. 4876 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), true); 4877 workers->run_task(&tsk); 4878 } else { 4879 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false); 4880 tsk.work(0); 4881 } 4882 4883 // restore, single-threaded for now, any preserved marks 4884 // as a result of work_q overflow 4885 restore_preserved_marks_if_any(); 4886 } 4887 4888 // Non-parallel version of remark 4889 void CMSCollector::do_remark_non_parallel() { 4890 ResourceMark rm; 4891 HandleMark hm; 4892 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4893 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false); 4894 4895 MarkRefsIntoAndScanClosure 4896 mrias_cl(_span, ref_processor(), &_markBitMap, NULL /* not precleaning */, 4897 &_markStack, this, 4898 false /* should_yield */, false /* not precleaning */); 4899 MarkFromDirtyCardsClosure 4900 markFromDirtyCardsClosure(this, _span, 4901 NULL, // space is set further below 4902 &_markBitMap, &_markStack, &mrias_cl); 4903 { 4904 GCTraceTime(Trace, gc) t("Grey Object Rescan", _gc_timer_cm); 4905 // Iterate over the dirty cards, setting the corresponding bits in the 4906 // mod union table. 4907 { 4908 ModUnionClosure modUnionClosure(&_modUnionTable); 4909 _ct->ct_bs()->dirty_card_iterate( 4910 _cmsGen->used_region(), 4911 &modUnionClosure); 4912 } 4913 // Having transferred these marks into the modUnionTable, we just need 4914 // to rescan the marked objects on the dirty cards in the modUnionTable. 4915 // The initial marking may have been done during an asynchronous 4916 // collection so there may be dirty bits in the mod-union table. 4917 const int alignment = 4918 CardTableModRefBS::card_size * BitsPerWord; 4919 { 4920 // ... First handle dirty cards in CMS gen 4921 markFromDirtyCardsClosure.set_space(_cmsGen->cmsSpace()); 4922 MemRegion ur = _cmsGen->used_region(); 4923 HeapWord* lb = ur.start(); 4924 HeapWord* ub = (HeapWord*)round_to((intptr_t)ur.end(), alignment); 4925 MemRegion cms_span(lb, ub); 4926 _modUnionTable.dirty_range_iterate_clear(cms_span, 4927 &markFromDirtyCardsClosure); 4928 verify_work_stacks_empty(); 4929 log_trace(gc)(" (re-scanned " SIZE_FORMAT " dirty cards in cms gen) ", markFromDirtyCardsClosure.num_dirty_cards()); 4930 } 4931 } 4932 if (VerifyDuringGC && 4933 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 4934 HandleMark hm; // Discard invalid handles created during verification 4935 Universe::verify(); 4936 } 4937 { 4938 GCTraceTime(Trace, gc) t("Root Rescan", _gc_timer_cm); 4939 4940 verify_work_stacks_empty(); 4941 4942 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 4943 StrongRootsScope srs(1); 4944 4945 gch->gen_process_roots(&srs, 4946 GenCollectedHeap::OldGen, 4947 true, // young gen as roots 4948 GenCollectedHeap::ScanningOption(roots_scanning_options()), 4949 should_unload_classes(), 4950 &mrias_cl, 4951 NULL, 4952 NULL); // The dirty klasses will be handled below 4953 4954 assert(should_unload_classes() 4955 || (roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache), 4956 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 4957 } 4958 4959 { 4960 GCTraceTime(Trace, gc) t("Visit Unhandled CLDs", _gc_timer_cm); 4961 4962 verify_work_stacks_empty(); 4963 4964 // Scan all class loader data objects that might have been introduced 4965 // during concurrent marking. 4966 ResourceMark rm; 4967 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds(); 4968 for (int i = 0; i < array->length(); i++) { 4969 mrias_cl.do_cld_nv(array->at(i)); 4970 } 4971 4972 // We don't need to keep track of new CLDs anymore. 4973 ClassLoaderDataGraph::remember_new_clds(false); 4974 4975 verify_work_stacks_empty(); 4976 } 4977 4978 { 4979 GCTraceTime(Trace, gc) t("Dirty Klass Scan", _gc_timer_cm); 4980 4981 verify_work_stacks_empty(); 4982 4983 RemarkKlassClosure remark_klass_closure(&mrias_cl); 4984 ClassLoaderDataGraph::classes_do(&remark_klass_closure); 4985 4986 verify_work_stacks_empty(); 4987 } 4988 4989 // We might have added oops to ClassLoaderData::_handles during the 4990 // concurrent marking phase. These oops point to newly allocated objects 4991 // that are guaranteed to be kept alive. Either by the direct allocation 4992 // code, or when the young collector processes the roots. Hence, 4993 // we don't have to revisit the _handles block during the remark phase. 4994 4995 verify_work_stacks_empty(); 4996 // Restore evacuated mark words, if any, used for overflow list links 4997 restore_preserved_marks_if_any(); 4998 4999 verify_overflow_empty(); 5000 } 5001 5002 //////////////////////////////////////////////////////// 5003 // Parallel Reference Processing Task Proxy Class 5004 //////////////////////////////////////////////////////// 5005 class AbstractGangTaskWOopQueues : public AbstractGangTask { 5006 OopTaskQueueSet* _queues; 5007 ParallelTaskTerminator _terminator; 5008 public: 5009 AbstractGangTaskWOopQueues(const char* name, OopTaskQueueSet* queues, uint n_threads) : 5010 AbstractGangTask(name), _queues(queues), _terminator(n_threads, _queues) {} 5011 ParallelTaskTerminator* terminator() { return &_terminator; } 5012 OopTaskQueueSet* queues() { return _queues; } 5013 }; 5014 5015 class CMSRefProcTaskProxy: public AbstractGangTaskWOopQueues { 5016 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask; 5017 CMSCollector* _collector; 5018 CMSBitMap* _mark_bit_map; 5019 const MemRegion _span; 5020 ProcessTask& _task; 5021 5022 public: 5023 CMSRefProcTaskProxy(ProcessTask& task, 5024 CMSCollector* collector, 5025 const MemRegion& span, 5026 CMSBitMap* mark_bit_map, 5027 AbstractWorkGang* workers, 5028 OopTaskQueueSet* task_queues): 5029 AbstractGangTaskWOopQueues("Process referents by policy in parallel", 5030 task_queues, 5031 workers->active_workers()), 5032 _task(task), 5033 _collector(collector), _span(span), _mark_bit_map(mark_bit_map) 5034 { 5035 assert(_collector->_span.equals(_span) && !_span.is_empty(), 5036 "Inconsistency in _span"); 5037 } 5038 5039 OopTaskQueueSet* task_queues() { return queues(); } 5040 5041 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 5042 5043 void do_work_steal(int i, 5044 CMSParDrainMarkingStackClosure* drain, 5045 CMSParKeepAliveClosure* keep_alive, 5046 int* seed); 5047 5048 virtual void work(uint worker_id); 5049 }; 5050 5051 void CMSRefProcTaskProxy::work(uint worker_id) { 5052 ResourceMark rm; 5053 HandleMark hm; 5054 assert(_collector->_span.equals(_span), "Inconsistency in _span"); 5055 CMSParKeepAliveClosure par_keep_alive(_collector, _span, 5056 _mark_bit_map, 5057 work_queue(worker_id)); 5058 CMSParDrainMarkingStackClosure par_drain_stack(_collector, _span, 5059 _mark_bit_map, 5060 work_queue(worker_id)); 5061 CMSIsAliveClosure is_alive_closure(_span, _mark_bit_map); 5062 _task.work(worker_id, is_alive_closure, par_keep_alive, par_drain_stack); 5063 if (_task.marks_oops_alive()) { 5064 do_work_steal(worker_id, &par_drain_stack, &par_keep_alive, 5065 _collector->hash_seed(worker_id)); 5066 } 5067 assert(work_queue(worker_id)->size() == 0, "work_queue should be empty"); 5068 assert(_collector->_overflow_list == NULL, "non-empty _overflow_list"); 5069 } 5070 5071 class CMSRefEnqueueTaskProxy: public AbstractGangTask { 5072 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask; 5073 EnqueueTask& _task; 5074 5075 public: 5076 CMSRefEnqueueTaskProxy(EnqueueTask& task) 5077 : AbstractGangTask("Enqueue reference objects in parallel"), 5078 _task(task) 5079 { } 5080 5081 virtual void work(uint worker_id) 5082 { 5083 _task.work(worker_id); 5084 } 5085 }; 5086 5087 CMSParKeepAliveClosure::CMSParKeepAliveClosure(CMSCollector* collector, 5088 MemRegion span, CMSBitMap* bit_map, OopTaskQueue* work_queue): 5089 _span(span), 5090 _bit_map(bit_map), 5091 _work_queue(work_queue), 5092 _mark_and_push(collector, span, bit_map, work_queue), 5093 _low_water_mark(MIN2((work_queue->max_elems()/4), 5094 ((uint)CMSWorkQueueDrainThreshold * ParallelGCThreads))) 5095 { } 5096 5097 // . see if we can share work_queues with ParNew? XXX 5098 void CMSRefProcTaskProxy::do_work_steal(int i, 5099 CMSParDrainMarkingStackClosure* drain, 5100 CMSParKeepAliveClosure* keep_alive, 5101 int* seed) { 5102 OopTaskQueue* work_q = work_queue(i); 5103 NOT_PRODUCT(int num_steals = 0;) 5104 oop obj_to_scan; 5105 5106 while (true) { 5107 // Completely finish any left over work from (an) earlier round(s) 5108 drain->trim_queue(0); 5109 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 5110 (size_t)ParGCDesiredObjsFromOverflowList); 5111 // Now check if there's any work in the overflow list 5112 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads, 5113 // only affects the number of attempts made to get work from the 5114 // overflow list and does not affect the number of workers. Just 5115 // pass ParallelGCThreads so this behavior is unchanged. 5116 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 5117 work_q, 5118 ParallelGCThreads)) { 5119 // Found something in global overflow list; 5120 // not yet ready to go stealing work from others. 5121 // We'd like to assert(work_q->size() != 0, ...) 5122 // because we just took work from the overflow list, 5123 // but of course we can't, since all of that might have 5124 // been already stolen from us. 5125 continue; 5126 } 5127 // Verify that we have no work before we resort to stealing 5128 assert(work_q->size() == 0, "Have work, shouldn't steal"); 5129 // Try to steal from other queues that have work 5130 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 5131 NOT_PRODUCT(num_steals++;) 5132 assert(obj_to_scan->is_oop(), "Oops, not an oop!"); 5133 assert(_mark_bit_map->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 5134 // Do scanning work 5135 obj_to_scan->oop_iterate(keep_alive); 5136 // Loop around, finish this work, and try to steal some more 5137 } else if (terminator()->offer_termination()) { 5138 break; // nirvana from the infinite cycle 5139 } 5140 } 5141 log_develop_trace(gc, task)("\t(%d: stole %d oops)", i, num_steals); 5142 } 5143 5144 void CMSRefProcTaskExecutor::execute(ProcessTask& task) 5145 { 5146 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5147 WorkGang* workers = gch->workers(); 5148 assert(workers != NULL, "Need parallel worker threads."); 5149 CMSRefProcTaskProxy rp_task(task, &_collector, 5150 _collector.ref_processor()->span(), 5151 _collector.markBitMap(), 5152 workers, _collector.task_queues()); 5153 workers->run_task(&rp_task); 5154 } 5155 5156 void CMSRefProcTaskExecutor::execute(EnqueueTask& task) 5157 { 5158 5159 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5160 WorkGang* workers = gch->workers(); 5161 assert(workers != NULL, "Need parallel worker threads."); 5162 CMSRefEnqueueTaskProxy enq_task(task); 5163 workers->run_task(&enq_task); 5164 } 5165 5166 void CMSCollector::refProcessingWork() { 5167 ResourceMark rm; 5168 HandleMark hm; 5169 5170 ReferenceProcessor* rp = ref_processor(); 5171 assert(rp->span().equals(_span), "Spans should be equal"); 5172 assert(!rp->enqueuing_is_done(), "Enqueuing should not be complete"); 5173 // Process weak references. 5174 rp->setup_policy(false); 5175 verify_work_stacks_empty(); 5176 5177 CMSKeepAliveClosure cmsKeepAliveClosure(this, _span, &_markBitMap, 5178 &_markStack, false /* !preclean */); 5179 CMSDrainMarkingStackClosure cmsDrainMarkingStackClosure(this, 5180 _span, &_markBitMap, &_markStack, 5181 &cmsKeepAliveClosure, false /* !preclean */); 5182 { 5183 GCTraceTime(Debug, gc) t("Weak Refs Processing", _gc_timer_cm); 5184 5185 ReferenceProcessorStats stats; 5186 if (rp->processing_is_mt()) { 5187 // Set the degree of MT here. If the discovery is done MT, there 5188 // may have been a different number of threads doing the discovery 5189 // and a different number of discovered lists may have Ref objects. 5190 // That is OK as long as the Reference lists are balanced (see 5191 // balance_all_queues() and balance_queues()). 5192 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5193 uint active_workers = ParallelGCThreads; 5194 WorkGang* workers = gch->workers(); 5195 if (workers != NULL) { 5196 active_workers = workers->active_workers(); 5197 // The expectation is that active_workers will have already 5198 // been set to a reasonable value. If it has not been set, 5199 // investigate. 5200 assert(active_workers > 0, "Should have been set during scavenge"); 5201 } 5202 rp->set_active_mt_degree(active_workers); 5203 CMSRefProcTaskExecutor task_executor(*this); 5204 stats = rp->process_discovered_references(&_is_alive_closure, 5205 &cmsKeepAliveClosure, 5206 &cmsDrainMarkingStackClosure, 5207 &task_executor, 5208 _gc_timer_cm); 5209 } else { 5210 stats = rp->process_discovered_references(&_is_alive_closure, 5211 &cmsKeepAliveClosure, 5212 &cmsDrainMarkingStackClosure, 5213 NULL, 5214 _gc_timer_cm); 5215 } 5216 _gc_tracer_cm->report_gc_reference_stats(stats); 5217 5218 } 5219 5220 // This is the point where the entire marking should have completed. 5221 verify_work_stacks_empty(); 5222 5223 if (should_unload_classes()) { 5224 { 5225 GCTraceTime(Debug, gc) t("Class Unloading", _gc_timer_cm); 5226 5227 // Unload classes and purge the SystemDictionary. 5228 bool purged_class = SystemDictionary::do_unloading(&_is_alive_closure); 5229 5230 // Unload nmethods. 5231 CodeCache::do_unloading(&_is_alive_closure, purged_class); 5232 5233 // Prune dead klasses from subklass/sibling/implementor lists. 5234 Klass::clean_weak_klass_links(&_is_alive_closure); 5235 } 5236 5237 { 5238 GCTraceTime(Debug, gc) t("Scrub Symbol Table", _gc_timer_cm); 5239 // Clean up unreferenced symbols in symbol table. 5240 SymbolTable::unlink(); 5241 } 5242 5243 { 5244 GCTraceTime(Debug, gc) t("Scrub String Table", _gc_timer_cm); 5245 // Delete entries for dead interned strings. 5246 StringTable::unlink(&_is_alive_closure); 5247 } 5248 } 5249 5250 5251 // Restore any preserved marks as a result of mark stack or 5252 // work queue overflow 5253 restore_preserved_marks_if_any(); // done single-threaded for now 5254 5255 rp->set_enqueuing_is_done(true); 5256 if (rp->processing_is_mt()) { 5257 rp->balance_all_queues(); 5258 CMSRefProcTaskExecutor task_executor(*this); 5259 rp->enqueue_discovered_references(&task_executor); 5260 } else { 5261 rp->enqueue_discovered_references(NULL); 5262 } 5263 rp->verify_no_references_recorded(); 5264 assert(!rp->discovery_enabled(), "should have been disabled"); 5265 } 5266 5267 #ifndef PRODUCT 5268 void CMSCollector::check_correct_thread_executing() { 5269 Thread* t = Thread::current(); 5270 // Only the VM thread or the CMS thread should be here. 5271 assert(t->is_ConcurrentGC_thread() || t->is_VM_thread(), 5272 "Unexpected thread type"); 5273 // If this is the vm thread, the foreground process 5274 // should not be waiting. Note that _foregroundGCIsActive is 5275 // true while the foreground collector is waiting. 5276 if (_foregroundGCShouldWait) { 5277 // We cannot be the VM thread 5278 assert(t->is_ConcurrentGC_thread(), 5279 "Should be CMS thread"); 5280 } else { 5281 // We can be the CMS thread only if we are in a stop-world 5282 // phase of CMS collection. 5283 if (t->is_ConcurrentGC_thread()) { 5284 assert(_collectorState == InitialMarking || 5285 _collectorState == FinalMarking, 5286 "Should be a stop-world phase"); 5287 // The CMS thread should be holding the CMS_token. 5288 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 5289 "Potential interference with concurrently " 5290 "executing VM thread"); 5291 } 5292 } 5293 } 5294 #endif 5295 5296 void CMSCollector::sweep() { 5297 assert(_collectorState == Sweeping, "just checking"); 5298 check_correct_thread_executing(); 5299 verify_work_stacks_empty(); 5300 verify_overflow_empty(); 5301 increment_sweep_count(); 5302 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 5303 5304 _inter_sweep_timer.stop(); 5305 _inter_sweep_estimate.sample(_inter_sweep_timer.seconds()); 5306 5307 assert(!_intra_sweep_timer.is_active(), "Should not be active"); 5308 _intra_sweep_timer.reset(); 5309 _intra_sweep_timer.start(); 5310 { 5311 GCTraceCPUTime tcpu; 5312 CMSPhaseAccounting pa(this, "Concurrent Sweep"); 5313 // First sweep the old gen 5314 { 5315 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(), 5316 bitMapLock()); 5317 sweepWork(_cmsGen); 5318 } 5319 5320 // Update Universe::_heap_*_at_gc figures. 5321 // We need all the free list locks to make the abstract state 5322 // transition from Sweeping to Resetting. See detailed note 5323 // further below. 5324 { 5325 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock()); 5326 // Update heap occupancy information which is used as 5327 // input to soft ref clearing policy at the next gc. 5328 Universe::update_heap_info_at_gc(); 5329 _collectorState = Resizing; 5330 } 5331 } 5332 verify_work_stacks_empty(); 5333 verify_overflow_empty(); 5334 5335 if (should_unload_classes()) { 5336 // Delay purge to the beginning of the next safepoint. Metaspace::contains 5337 // requires that the virtual spaces are stable and not deleted. 5338 ClassLoaderDataGraph::set_should_purge(true); 5339 } 5340 5341 _intra_sweep_timer.stop(); 5342 _intra_sweep_estimate.sample(_intra_sweep_timer.seconds()); 5343 5344 _inter_sweep_timer.reset(); 5345 _inter_sweep_timer.start(); 5346 5347 // We need to use a monotonically non-decreasing time in ms 5348 // or we will see time-warp warnings and os::javaTimeMillis() 5349 // does not guarantee monotonicity. 5350 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC; 5351 update_time_of_last_gc(now); 5352 5353 // NOTE on abstract state transitions: 5354 // Mutators allocate-live and/or mark the mod-union table dirty 5355 // based on the state of the collection. The former is done in 5356 // the interval [Marking, Sweeping] and the latter in the interval 5357 // [Marking, Sweeping). Thus the transitions into the Marking state 5358 // and out of the Sweeping state must be synchronously visible 5359 // globally to the mutators. 5360 // The transition into the Marking state happens with the world 5361 // stopped so the mutators will globally see it. Sweeping is 5362 // done asynchronously by the background collector so the transition 5363 // from the Sweeping state to the Resizing state must be done 5364 // under the freelistLock (as is the check for whether to 5365 // allocate-live and whether to dirty the mod-union table). 5366 assert(_collectorState == Resizing, "Change of collector state to" 5367 " Resizing must be done under the freelistLocks (plural)"); 5368 5369 // Now that sweeping has been completed, we clear 5370 // the incremental_collection_failed flag, 5371 // thus inviting a younger gen collection to promote into 5372 // this generation. If such a promotion may still fail, 5373 // the flag will be set again when a young collection is 5374 // attempted. 5375 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5376 gch->clear_incremental_collection_failed(); // Worth retrying as fresh space may have been freed up 5377 gch->update_full_collections_completed(_collection_count_start); 5378 } 5379 5380 // FIX ME!!! Looks like this belongs in CFLSpace, with 5381 // CMSGen merely delegating to it. 5382 void ConcurrentMarkSweepGeneration::setNearLargestChunk() { 5383 double nearLargestPercent = FLSLargestBlockCoalesceProximity; 5384 HeapWord* minAddr = _cmsSpace->bottom(); 5385 HeapWord* largestAddr = 5386 (HeapWord*) _cmsSpace->dictionary()->find_largest_dict(); 5387 if (largestAddr == NULL) { 5388 // The dictionary appears to be empty. In this case 5389 // try to coalesce at the end of the heap. 5390 largestAddr = _cmsSpace->end(); 5391 } 5392 size_t largestOffset = pointer_delta(largestAddr, minAddr); 5393 size_t nearLargestOffset = 5394 (size_t)((double)largestOffset * nearLargestPercent) - MinChunkSize; 5395 log_debug(gc, freelist)("CMS: Large Block: " PTR_FORMAT "; Proximity: " PTR_FORMAT " -> " PTR_FORMAT, 5396 p2i(largestAddr), p2i(_cmsSpace->nearLargestChunk()), p2i(minAddr + nearLargestOffset)); 5397 _cmsSpace->set_nearLargestChunk(minAddr + nearLargestOffset); 5398 } 5399 5400 bool ConcurrentMarkSweepGeneration::isNearLargestChunk(HeapWord* addr) { 5401 return addr >= _cmsSpace->nearLargestChunk(); 5402 } 5403 5404 FreeChunk* ConcurrentMarkSweepGeneration::find_chunk_at_end() { 5405 return _cmsSpace->find_chunk_at_end(); 5406 } 5407 5408 void ConcurrentMarkSweepGeneration::update_gc_stats(Generation* current_generation, 5409 bool full) { 5410 // If the young generation has been collected, gather any statistics 5411 // that are of interest at this point. 5412 bool current_is_young = GenCollectedHeap::heap()->is_young_gen(current_generation); 5413 if (!full && current_is_young) { 5414 // Gather statistics on the young generation collection. 5415 collector()->stats().record_gc0_end(used()); 5416 } 5417 } 5418 5419 void CMSCollector::sweepWork(ConcurrentMarkSweepGeneration* old_gen) { 5420 // We iterate over the space(s) underlying this generation, 5421 // checking the mark bit map to see if the bits corresponding 5422 // to specific blocks are marked or not. Blocks that are 5423 // marked are live and are not swept up. All remaining blocks 5424 // are swept up, with coalescing on-the-fly as we sweep up 5425 // contiguous free and/or garbage blocks: 5426 // We need to ensure that the sweeper synchronizes with allocators 5427 // and stop-the-world collectors. In particular, the following 5428 // locks are used: 5429 // . CMS token: if this is held, a stop the world collection cannot occur 5430 // . freelistLock: if this is held no allocation can occur from this 5431 // generation by another thread 5432 // . bitMapLock: if this is held, no other thread can access or update 5433 // 5434 5435 // Note that we need to hold the freelistLock if we use 5436 // block iterate below; else the iterator might go awry if 5437 // a mutator (or promotion) causes block contents to change 5438 // (for instance if the allocator divvies up a block). 5439 // If we hold the free list lock, for all practical purposes 5440 // young generation GC's can't occur (they'll usually need to 5441 // promote), so we might as well prevent all young generation 5442 // GC's while we do a sweeping step. For the same reason, we might 5443 // as well take the bit map lock for the entire duration 5444 5445 // check that we hold the requisite locks 5446 assert(have_cms_token(), "Should hold cms token"); 5447 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), "Should possess CMS token to sweep"); 5448 assert_lock_strong(old_gen->freelistLock()); 5449 assert_lock_strong(bitMapLock()); 5450 5451 assert(!_inter_sweep_timer.is_active(), "Was switched off in an outer context"); 5452 assert(_intra_sweep_timer.is_active(), "Was switched on in an outer context"); 5453 old_gen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()), 5454 _inter_sweep_estimate.padded_average(), 5455 _intra_sweep_estimate.padded_average()); 5456 old_gen->setNearLargestChunk(); 5457 5458 { 5459 SweepClosure sweepClosure(this, old_gen, &_markBitMap, CMSYield); 5460 old_gen->cmsSpace()->blk_iterate_careful(&sweepClosure); 5461 // We need to free-up/coalesce garbage/blocks from a 5462 // co-terminal free run. This is done in the SweepClosure 5463 // destructor; so, do not remove this scope, else the 5464 // end-of-sweep-census below will be off by a little bit. 5465 } 5466 old_gen->cmsSpace()->sweep_completed(); 5467 old_gen->cmsSpace()->endSweepFLCensus(sweep_count()); 5468 if (should_unload_classes()) { // unloaded classes this cycle, 5469 _concurrent_cycles_since_last_unload = 0; // ... reset count 5470 } else { // did not unload classes, 5471 _concurrent_cycles_since_last_unload++; // ... increment count 5472 } 5473 } 5474 5475 // Reset CMS data structures (for now just the marking bit map) 5476 // preparatory for the next cycle. 5477 void CMSCollector::reset_concurrent() { 5478 CMSTokenSyncWithLocks ts(true, bitMapLock()); 5479 5480 // If the state is not "Resetting", the foreground thread 5481 // has done a collection and the resetting. 5482 if (_collectorState != Resetting) { 5483 assert(_collectorState == Idling, "The state should only change" 5484 " because the foreground collector has finished the collection"); 5485 return; 5486 } 5487 5488 { 5489 // Clear the mark bitmap (no grey objects to start with) 5490 // for the next cycle. 5491 GCTraceCPUTime tcpu; 5492 CMSPhaseAccounting cmspa(this, "Concurrent Reset"); 5493 5494 HeapWord* curAddr = _markBitMap.startWord(); 5495 while (curAddr < _markBitMap.endWord()) { 5496 size_t remaining = pointer_delta(_markBitMap.endWord(), curAddr); 5497 MemRegion chunk(curAddr, MIN2(CMSBitMapYieldQuantum, remaining)); 5498 _markBitMap.clear_large_range(chunk); 5499 if (ConcurrentMarkSweepThread::should_yield() && 5500 !foregroundGCIsActive() && 5501 CMSYield) { 5502 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 5503 "CMS thread should hold CMS token"); 5504 assert_lock_strong(bitMapLock()); 5505 bitMapLock()->unlock(); 5506 ConcurrentMarkSweepThread::desynchronize(true); 5507 stopTimer(); 5508 incrementYields(); 5509 5510 // See the comment in coordinator_yield() 5511 for (unsigned i = 0; i < CMSYieldSleepCount && 5512 ConcurrentMarkSweepThread::should_yield() && 5513 !CMSCollector::foregroundGCIsActive(); ++i) { 5514 os::sleep(Thread::current(), 1, false); 5515 } 5516 5517 ConcurrentMarkSweepThread::synchronize(true); 5518 bitMapLock()->lock_without_safepoint_check(); 5519 startTimer(); 5520 } 5521 curAddr = chunk.end(); 5522 } 5523 // A successful mostly concurrent collection has been done. 5524 // Because only the full (i.e., concurrent mode failure) collections 5525 // are being measured for gc overhead limits, clean the "near" flag 5526 // and count. 5527 size_policy()->reset_gc_overhead_limit_count(); 5528 _collectorState = Idling; 5529 } 5530 5531 register_gc_end(); 5532 } 5533 5534 // Same as above but for STW paths 5535 void CMSCollector::reset_stw() { 5536 // already have the lock 5537 assert(_collectorState == Resetting, "just checking"); 5538 assert_lock_strong(bitMapLock()); 5539 GCIdMarkAndRestore gc_id_mark(_cmsThread->gc_id()); 5540 _markBitMap.clear_all(); 5541 _collectorState = Idling; 5542 register_gc_end(); 5543 } 5544 5545 void CMSCollector::do_CMS_operation(CMS_op_type op, GCCause::Cause gc_cause) { 5546 GCTraceCPUTime tcpu; 5547 TraceCollectorStats tcs(counters()); 5548 5549 switch (op) { 5550 case CMS_op_checkpointRootsInitial: { 5551 GCTraceTime(Info, gc) t("Pause Initial Mark", NULL, GCCause::_no_gc, true); 5552 SvcGCMarker sgcm(SvcGCMarker::OTHER); 5553 checkpointRootsInitial(); 5554 break; 5555 } 5556 case CMS_op_checkpointRootsFinal: { 5557 GCTraceTime(Info, gc) t("Pause Remark", NULL, GCCause::_no_gc, true); 5558 SvcGCMarker sgcm(SvcGCMarker::OTHER); 5559 checkpointRootsFinal(); 5560 break; 5561 } 5562 default: 5563 fatal("No such CMS_op"); 5564 } 5565 } 5566 5567 #ifndef PRODUCT 5568 size_t const CMSCollector::skip_header_HeapWords() { 5569 return FreeChunk::header_size(); 5570 } 5571 5572 // Try and collect here conditions that should hold when 5573 // CMS thread is exiting. The idea is that the foreground GC 5574 // thread should not be blocked if it wants to terminate 5575 // the CMS thread and yet continue to run the VM for a while 5576 // after that. 5577 void CMSCollector::verify_ok_to_terminate() const { 5578 assert(Thread::current()->is_ConcurrentGC_thread(), 5579 "should be called by CMS thread"); 5580 assert(!_foregroundGCShouldWait, "should be false"); 5581 // We could check here that all the various low-level locks 5582 // are not held by the CMS thread, but that is overkill; see 5583 // also CMSThread::verify_ok_to_terminate() where the CGC_lock 5584 // is checked. 5585 } 5586 #endif 5587 5588 size_t CMSCollector::block_size_using_printezis_bits(HeapWord* addr) const { 5589 assert(_markBitMap.isMarked(addr) && _markBitMap.isMarked(addr + 1), 5590 "missing Printezis mark?"); 5591 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 5592 size_t size = pointer_delta(nextOneAddr + 1, addr); 5593 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 5594 "alignment problem"); 5595 assert(size >= 3, "Necessary for Printezis marks to work"); 5596 return size; 5597 } 5598 5599 // A variant of the above (block_size_using_printezis_bits()) except 5600 // that we return 0 if the P-bits are not yet set. 5601 size_t CMSCollector::block_size_if_printezis_bits(HeapWord* addr) const { 5602 if (_markBitMap.isMarked(addr + 1)) { 5603 assert(_markBitMap.isMarked(addr), "P-bit can be set only for marked objects"); 5604 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 5605 size_t size = pointer_delta(nextOneAddr + 1, addr); 5606 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 5607 "alignment problem"); 5608 assert(size >= 3, "Necessary for Printezis marks to work"); 5609 return size; 5610 } 5611 return 0; 5612 } 5613 5614 HeapWord* CMSCollector::next_card_start_after_block(HeapWord* addr) const { 5615 size_t sz = 0; 5616 oop p = (oop)addr; 5617 if (p->klass_or_null() != NULL) { 5618 sz = CompactibleFreeListSpace::adjustObjectSize(p->size()); 5619 } else { 5620 sz = block_size_using_printezis_bits(addr); 5621 } 5622 assert(sz > 0, "size must be nonzero"); 5623 HeapWord* next_block = addr + sz; 5624 HeapWord* next_card = (HeapWord*)round_to((uintptr_t)next_block, 5625 CardTableModRefBS::card_size); 5626 assert(round_down((uintptr_t)addr, CardTableModRefBS::card_size) < 5627 round_down((uintptr_t)next_card, CardTableModRefBS::card_size), 5628 "must be different cards"); 5629 return next_card; 5630 } 5631 5632 5633 // CMS Bit Map Wrapper ///////////////////////////////////////// 5634 5635 // Construct a CMS bit map infrastructure, but don't create the 5636 // bit vector itself. That is done by a separate call CMSBitMap::allocate() 5637 // further below. 5638 CMSBitMap::CMSBitMap(int shifter, int mutex_rank, const char* mutex_name): 5639 _bm(), 5640 _shifter(shifter), 5641 _lock(mutex_rank >= 0 ? new Mutex(mutex_rank, mutex_name, true, 5642 Monitor::_safepoint_check_sometimes) : NULL) 5643 { 5644 _bmStartWord = 0; 5645 _bmWordSize = 0; 5646 } 5647 5648 bool CMSBitMap::allocate(MemRegion mr) { 5649 _bmStartWord = mr.start(); 5650 _bmWordSize = mr.word_size(); 5651 ReservedSpace brs(ReservedSpace::allocation_align_size_up( 5652 (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1)); 5653 if (!brs.is_reserved()) { 5654 warning("CMS bit map allocation failure"); 5655 return false; 5656 } 5657 // For now we'll just commit all of the bit map up front. 5658 // Later on we'll try to be more parsimonious with swap. 5659 if (!_virtual_space.initialize(brs, brs.size())) { 5660 warning("CMS bit map backing store failure"); 5661 return false; 5662 } 5663 assert(_virtual_space.committed_size() == brs.size(), 5664 "didn't reserve backing store for all of CMS bit map?"); 5665 _bm.set_map((BitMap::bm_word_t*)_virtual_space.low()); 5666 assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >= 5667 _bmWordSize, "inconsistency in bit map sizing"); 5668 _bm.set_size(_bmWordSize >> _shifter); 5669 5670 // bm.clear(); // can we rely on getting zero'd memory? verify below 5671 assert(isAllClear(), 5672 "Expected zero'd memory from ReservedSpace constructor"); 5673 assert(_bm.size() == heapWordDiffToOffsetDiff(sizeInWords()), 5674 "consistency check"); 5675 return true; 5676 } 5677 5678 void CMSBitMap::dirty_range_iterate_clear(MemRegion mr, MemRegionClosure* cl) { 5679 HeapWord *next_addr, *end_addr, *last_addr; 5680 assert_locked(); 5681 assert(covers(mr), "out-of-range error"); 5682 // XXX assert that start and end are appropriately aligned 5683 for (next_addr = mr.start(), end_addr = mr.end(); 5684 next_addr < end_addr; next_addr = last_addr) { 5685 MemRegion dirty_region = getAndClearMarkedRegion(next_addr, end_addr); 5686 last_addr = dirty_region.end(); 5687 if (!dirty_region.is_empty()) { 5688 cl->do_MemRegion(dirty_region); 5689 } else { 5690 assert(last_addr == end_addr, "program logic"); 5691 return; 5692 } 5693 } 5694 } 5695 5696 void CMSBitMap::print_on_error(outputStream* st, const char* prefix) const { 5697 _bm.print_on_error(st, prefix); 5698 } 5699 5700 #ifndef PRODUCT 5701 void CMSBitMap::assert_locked() const { 5702 CMSLockVerifier::assert_locked(lock()); 5703 } 5704 5705 bool CMSBitMap::covers(MemRegion mr) const { 5706 // assert(_bm.map() == _virtual_space.low(), "map inconsistency"); 5707 assert((size_t)_bm.size() == (_bmWordSize >> _shifter), 5708 "size inconsistency"); 5709 return (mr.start() >= _bmStartWord) && 5710 (mr.end() <= endWord()); 5711 } 5712 5713 bool CMSBitMap::covers(HeapWord* start, size_t size) const { 5714 return (start >= _bmStartWord && (start + size) <= endWord()); 5715 } 5716 5717 void CMSBitMap::verifyNoOneBitsInRange(HeapWord* left, HeapWord* right) { 5718 // verify that there are no 1 bits in the interval [left, right) 5719 FalseBitMapClosure falseBitMapClosure; 5720 iterate(&falseBitMapClosure, left, right); 5721 } 5722 5723 void CMSBitMap::region_invariant(MemRegion mr) 5724 { 5725 assert_locked(); 5726 // mr = mr.intersection(MemRegion(_bmStartWord, _bmWordSize)); 5727 assert(!mr.is_empty(), "unexpected empty region"); 5728 assert(covers(mr), "mr should be covered by bit map"); 5729 // convert address range into offset range 5730 size_t start_ofs = heapWordToOffset(mr.start()); 5731 // Make sure that end() is appropriately aligned 5732 assert(mr.end() == (HeapWord*)round_to((intptr_t)mr.end(), 5733 (1 << (_shifter+LogHeapWordSize))), 5734 "Misaligned mr.end()"); 5735 size_t end_ofs = heapWordToOffset(mr.end()); 5736 assert(end_ofs > start_ofs, "Should mark at least one bit"); 5737 } 5738 5739 #endif 5740 5741 bool CMSMarkStack::allocate(size_t size) { 5742 // allocate a stack of the requisite depth 5743 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 5744 size * sizeof(oop))); 5745 if (!rs.is_reserved()) { 5746 warning("CMSMarkStack allocation failure"); 5747 return false; 5748 } 5749 if (!_virtual_space.initialize(rs, rs.size())) { 5750 warning("CMSMarkStack backing store failure"); 5751 return false; 5752 } 5753 assert(_virtual_space.committed_size() == rs.size(), 5754 "didn't reserve backing store for all of CMS stack?"); 5755 _base = (oop*)(_virtual_space.low()); 5756 _index = 0; 5757 _capacity = size; 5758 NOT_PRODUCT(_max_depth = 0); 5759 return true; 5760 } 5761 5762 // XXX FIX ME !!! In the MT case we come in here holding a 5763 // leaf lock. For printing we need to take a further lock 5764 // which has lower rank. We need to recalibrate the two 5765 // lock-ranks involved in order to be able to print the 5766 // messages below. (Or defer the printing to the caller. 5767 // For now we take the expedient path of just disabling the 5768 // messages for the problematic case.) 5769 void CMSMarkStack::expand() { 5770 assert(_capacity <= MarkStackSizeMax, "stack bigger than permitted"); 5771 if (_capacity == MarkStackSizeMax) { 5772 if (_hit_limit++ == 0 && !CMSConcurrentMTEnabled) { 5773 // We print a warning message only once per CMS cycle. 5774 log_debug(gc)(" (benign) Hit CMSMarkStack max size limit"); 5775 } 5776 return; 5777 } 5778 // Double capacity if possible 5779 size_t new_capacity = MIN2(_capacity*2, MarkStackSizeMax); 5780 // Do not give up existing stack until we have managed to 5781 // get the double capacity that we desired. 5782 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 5783 new_capacity * sizeof(oop))); 5784 if (rs.is_reserved()) { 5785 // Release the backing store associated with old stack 5786 _virtual_space.release(); 5787 // Reinitialize virtual space for new stack 5788 if (!_virtual_space.initialize(rs, rs.size())) { 5789 fatal("Not enough swap for expanded marking stack"); 5790 } 5791 _base = (oop*)(_virtual_space.low()); 5792 _index = 0; 5793 _capacity = new_capacity; 5794 } else if (_failed_double++ == 0 && !CMSConcurrentMTEnabled) { 5795 // Failed to double capacity, continue; 5796 // we print a detail message only once per CMS cycle. 5797 log_debug(gc)(" (benign) Failed to expand marking stack from " SIZE_FORMAT "K to " SIZE_FORMAT "K", 5798 _capacity / K, new_capacity / K); 5799 } 5800 } 5801 5802 5803 // Closures 5804 // XXX: there seems to be a lot of code duplication here; 5805 // should refactor and consolidate common code. 5806 5807 // This closure is used to mark refs into the CMS generation in 5808 // the CMS bit map. Called at the first checkpoint. This closure 5809 // assumes that we do not need to re-mark dirty cards; if the CMS 5810 // generation on which this is used is not an oldest 5811 // generation then this will lose younger_gen cards! 5812 5813 MarkRefsIntoClosure::MarkRefsIntoClosure( 5814 MemRegion span, CMSBitMap* bitMap): 5815 _span(span), 5816 _bitMap(bitMap) 5817 { 5818 assert(ref_processor() == NULL, "deliberately left NULL"); 5819 assert(_bitMap->covers(_span), "_bitMap/_span mismatch"); 5820 } 5821 5822 void MarkRefsIntoClosure::do_oop(oop obj) { 5823 // if p points into _span, then mark corresponding bit in _markBitMap 5824 assert(obj->is_oop(), "expected an oop"); 5825 HeapWord* addr = (HeapWord*)obj; 5826 if (_span.contains(addr)) { 5827 // this should be made more efficient 5828 _bitMap->mark(addr); 5829 } 5830 } 5831 5832 void MarkRefsIntoClosure::do_oop(oop* p) { MarkRefsIntoClosure::do_oop_work(p); } 5833 void MarkRefsIntoClosure::do_oop(narrowOop* p) { MarkRefsIntoClosure::do_oop_work(p); } 5834 5835 Par_MarkRefsIntoClosure::Par_MarkRefsIntoClosure( 5836 MemRegion span, CMSBitMap* bitMap): 5837 _span(span), 5838 _bitMap(bitMap) 5839 { 5840 assert(ref_processor() == NULL, "deliberately left NULL"); 5841 assert(_bitMap->covers(_span), "_bitMap/_span mismatch"); 5842 } 5843 5844 void Par_MarkRefsIntoClosure::do_oop(oop obj) { 5845 // if p points into _span, then mark corresponding bit in _markBitMap 5846 assert(obj->is_oop(), "expected an oop"); 5847 HeapWord* addr = (HeapWord*)obj; 5848 if (_span.contains(addr)) { 5849 // this should be made more efficient 5850 _bitMap->par_mark(addr); 5851 } 5852 } 5853 5854 void Par_MarkRefsIntoClosure::do_oop(oop* p) { Par_MarkRefsIntoClosure::do_oop_work(p); } 5855 void Par_MarkRefsIntoClosure::do_oop(narrowOop* p) { Par_MarkRefsIntoClosure::do_oop_work(p); } 5856 5857 // A variant of the above, used for CMS marking verification. 5858 MarkRefsIntoVerifyClosure::MarkRefsIntoVerifyClosure( 5859 MemRegion span, CMSBitMap* verification_bm, CMSBitMap* cms_bm): 5860 _span(span), 5861 _verification_bm(verification_bm), 5862 _cms_bm(cms_bm) 5863 { 5864 assert(ref_processor() == NULL, "deliberately left NULL"); 5865 assert(_verification_bm->covers(_span), "_verification_bm/_span mismatch"); 5866 } 5867 5868 void MarkRefsIntoVerifyClosure::do_oop(oop obj) { 5869 // if p points into _span, then mark corresponding bit in _markBitMap 5870 assert(obj->is_oop(), "expected an oop"); 5871 HeapWord* addr = (HeapWord*)obj; 5872 if (_span.contains(addr)) { 5873 _verification_bm->mark(addr); 5874 if (!_cms_bm->isMarked(addr)) { 5875 LogHandle(gc, verify) log; 5876 ResourceMark rm; 5877 oop(addr)->print_on(log.info_stream()); 5878 log.info(" (" INTPTR_FORMAT " should have been marked)", p2i(addr)); 5879 fatal("... aborting"); 5880 } 5881 } 5882 } 5883 5884 void MarkRefsIntoVerifyClosure::do_oop(oop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); } 5885 void MarkRefsIntoVerifyClosure::do_oop(narrowOop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); } 5886 5887 ////////////////////////////////////////////////// 5888 // MarkRefsIntoAndScanClosure 5889 ////////////////////////////////////////////////// 5890 5891 MarkRefsIntoAndScanClosure::MarkRefsIntoAndScanClosure(MemRegion span, 5892 ReferenceProcessor* rp, 5893 CMSBitMap* bit_map, 5894 CMSBitMap* mod_union_table, 5895 CMSMarkStack* mark_stack, 5896 CMSCollector* collector, 5897 bool should_yield, 5898 bool concurrent_precleaning): 5899 _collector(collector), 5900 _span(span), 5901 _bit_map(bit_map), 5902 _mark_stack(mark_stack), 5903 _pushAndMarkClosure(collector, span, rp, bit_map, mod_union_table, 5904 mark_stack, concurrent_precleaning), 5905 _yield(should_yield), 5906 _concurrent_precleaning(concurrent_precleaning), 5907 _freelistLock(NULL) 5908 { 5909 // FIXME: Should initialize in base class constructor. 5910 assert(rp != NULL, "ref_processor shouldn't be NULL"); 5911 set_ref_processor_internal(rp); 5912 } 5913 5914 // This closure is used to mark refs into the CMS generation at the 5915 // second (final) checkpoint, and to scan and transitively follow 5916 // the unmarked oops. It is also used during the concurrent precleaning 5917 // phase while scanning objects on dirty cards in the CMS generation. 5918 // The marks are made in the marking bit map and the marking stack is 5919 // used for keeping the (newly) grey objects during the scan. 5920 // The parallel version (Par_...) appears further below. 5921 void MarkRefsIntoAndScanClosure::do_oop(oop obj) { 5922 if (obj != NULL) { 5923 assert(obj->is_oop(), "expected an oop"); 5924 HeapWord* addr = (HeapWord*)obj; 5925 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 5926 assert(_collector->overflow_list_is_empty(), 5927 "overflow list should be empty"); 5928 if (_span.contains(addr) && 5929 !_bit_map->isMarked(addr)) { 5930 // mark bit map (object is now grey) 5931 _bit_map->mark(addr); 5932 // push on marking stack (stack should be empty), and drain the 5933 // stack by applying this closure to the oops in the oops popped 5934 // from the stack (i.e. blacken the grey objects) 5935 bool res = _mark_stack->push(obj); 5936 assert(res, "Should have space to push on empty stack"); 5937 do { 5938 oop new_oop = _mark_stack->pop(); 5939 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 5940 assert(_bit_map->isMarked((HeapWord*)new_oop), 5941 "only grey objects on this stack"); 5942 // iterate over the oops in this oop, marking and pushing 5943 // the ones in CMS heap (i.e. in _span). 5944 new_oop->oop_iterate(&_pushAndMarkClosure); 5945 // check if it's time to yield 5946 do_yield_check(); 5947 } while (!_mark_stack->isEmpty() || 5948 (!_concurrent_precleaning && take_from_overflow_list())); 5949 // if marking stack is empty, and we are not doing this 5950 // during precleaning, then check the overflow list 5951 } 5952 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 5953 assert(_collector->overflow_list_is_empty(), 5954 "overflow list was drained above"); 5955 5956 assert(_collector->no_preserved_marks(), 5957 "All preserved marks should have been restored above"); 5958 } 5959 } 5960 5961 void MarkRefsIntoAndScanClosure::do_oop(oop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); } 5962 void MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); } 5963 5964 void MarkRefsIntoAndScanClosure::do_yield_work() { 5965 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 5966 "CMS thread should hold CMS token"); 5967 assert_lock_strong(_freelistLock); 5968 assert_lock_strong(_bit_map->lock()); 5969 // relinquish the free_list_lock and bitMaplock() 5970 _bit_map->lock()->unlock(); 5971 _freelistLock->unlock(); 5972 ConcurrentMarkSweepThread::desynchronize(true); 5973 _collector->stopTimer(); 5974 _collector->incrementYields(); 5975 5976 // See the comment in coordinator_yield() 5977 for (unsigned i = 0; 5978 i < CMSYieldSleepCount && 5979 ConcurrentMarkSweepThread::should_yield() && 5980 !CMSCollector::foregroundGCIsActive(); 5981 ++i) { 5982 os::sleep(Thread::current(), 1, false); 5983 } 5984 5985 ConcurrentMarkSweepThread::synchronize(true); 5986 _freelistLock->lock_without_safepoint_check(); 5987 _bit_map->lock()->lock_without_safepoint_check(); 5988 _collector->startTimer(); 5989 } 5990 5991 /////////////////////////////////////////////////////////// 5992 // Par_MarkRefsIntoAndScanClosure: a parallel version of 5993 // MarkRefsIntoAndScanClosure 5994 /////////////////////////////////////////////////////////// 5995 Par_MarkRefsIntoAndScanClosure::Par_MarkRefsIntoAndScanClosure( 5996 CMSCollector* collector, MemRegion span, ReferenceProcessor* rp, 5997 CMSBitMap* bit_map, OopTaskQueue* work_queue): 5998 _span(span), 5999 _bit_map(bit_map), 6000 _work_queue(work_queue), 6001 _low_water_mark(MIN2((work_queue->max_elems()/4), 6002 ((uint)CMSWorkQueueDrainThreshold * ParallelGCThreads))), 6003 _par_pushAndMarkClosure(collector, span, rp, bit_map, work_queue) 6004 { 6005 // FIXME: Should initialize in base class constructor. 6006 assert(rp != NULL, "ref_processor shouldn't be NULL"); 6007 set_ref_processor_internal(rp); 6008 } 6009 6010 // This closure is used to mark refs into the CMS generation at the 6011 // second (final) checkpoint, and to scan and transitively follow 6012 // the unmarked oops. The marks are made in the marking bit map and 6013 // the work_queue is used for keeping the (newly) grey objects during 6014 // the scan phase whence they are also available for stealing by parallel 6015 // threads. Since the marking bit map is shared, updates are 6016 // synchronized (via CAS). 6017 void Par_MarkRefsIntoAndScanClosure::do_oop(oop obj) { 6018 if (obj != NULL) { 6019 // Ignore mark word because this could be an already marked oop 6020 // that may be chained at the end of the overflow list. 6021 assert(obj->is_oop(true), "expected an oop"); 6022 HeapWord* addr = (HeapWord*)obj; 6023 if (_span.contains(addr) && 6024 !_bit_map->isMarked(addr)) { 6025 // mark bit map (object will become grey): 6026 // It is possible for several threads to be 6027 // trying to "claim" this object concurrently; 6028 // the unique thread that succeeds in marking the 6029 // object first will do the subsequent push on 6030 // to the work queue (or overflow list). 6031 if (_bit_map->par_mark(addr)) { 6032 // push on work_queue (which may not be empty), and trim the 6033 // queue to an appropriate length by applying this closure to 6034 // the oops in the oops popped from the stack (i.e. blacken the 6035 // grey objects) 6036 bool res = _work_queue->push(obj); 6037 assert(res, "Low water mark should be less than capacity?"); 6038 trim_queue(_low_water_mark); 6039 } // Else, another thread claimed the object 6040 } 6041 } 6042 } 6043 6044 void Par_MarkRefsIntoAndScanClosure::do_oop(oop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); } 6045 void Par_MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); } 6046 6047 // This closure is used to rescan the marked objects on the dirty cards 6048 // in the mod union table and the card table proper. 6049 size_t ScanMarkedObjectsAgainCarefullyClosure::do_object_careful_m( 6050 oop p, MemRegion mr) { 6051 6052 size_t size = 0; 6053 HeapWord* addr = (HeapWord*)p; 6054 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6055 assert(_span.contains(addr), "we are scanning the CMS generation"); 6056 // check if it's time to yield 6057 if (do_yield_check()) { 6058 // We yielded for some foreground stop-world work, 6059 // and we have been asked to abort this ongoing preclean cycle. 6060 return 0; 6061 } 6062 if (_bitMap->isMarked(addr)) { 6063 // it's marked; is it potentially uninitialized? 6064 if (p->klass_or_null() != NULL) { 6065 // an initialized object; ignore mark word in verification below 6066 // since we are running concurrent with mutators 6067 assert(p->is_oop(true), "should be an oop"); 6068 if (p->is_objArray()) { 6069 // objArrays are precisely marked; restrict scanning 6070 // to dirty cards only. 6071 size = CompactibleFreeListSpace::adjustObjectSize( 6072 p->oop_iterate_size(_scanningClosure, mr)); 6073 } else { 6074 // A non-array may have been imprecisely marked; we need 6075 // to scan object in its entirety. 6076 size = CompactibleFreeListSpace::adjustObjectSize( 6077 p->oop_iterate_size(_scanningClosure)); 6078 } 6079 #ifdef ASSERT 6080 size_t direct_size = 6081 CompactibleFreeListSpace::adjustObjectSize(p->size()); 6082 assert(size == direct_size, "Inconsistency in size"); 6083 assert(size >= 3, "Necessary for Printezis marks to work"); 6084 if (!_bitMap->isMarked(addr+1)) { 6085 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size); 6086 } else { 6087 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size-1); 6088 assert(_bitMap->isMarked(addr+size-1), 6089 "inconsistent Printezis mark"); 6090 } 6091 #endif // ASSERT 6092 } else { 6093 // An uninitialized object. 6094 assert(_bitMap->isMarked(addr+1), "missing Printezis mark?"); 6095 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 6096 size = pointer_delta(nextOneAddr + 1, addr); 6097 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 6098 "alignment problem"); 6099 // Note that pre-cleaning needn't redirty the card. OopDesc::set_klass() 6100 // will dirty the card when the klass pointer is installed in the 6101 // object (signaling the completion of initialization). 6102 } 6103 } else { 6104 // Either a not yet marked object or an uninitialized object 6105 if (p->klass_or_null() == NULL) { 6106 // An uninitialized object, skip to the next card, since 6107 // we may not be able to read its P-bits yet. 6108 assert(size == 0, "Initial value"); 6109 } else { 6110 // An object not (yet) reached by marking: we merely need to 6111 // compute its size so as to go look at the next block. 6112 assert(p->is_oop(true), "should be an oop"); 6113 size = CompactibleFreeListSpace::adjustObjectSize(p->size()); 6114 } 6115 } 6116 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6117 return size; 6118 } 6119 6120 void ScanMarkedObjectsAgainCarefullyClosure::do_yield_work() { 6121 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6122 "CMS thread should hold CMS token"); 6123 assert_lock_strong(_freelistLock); 6124 assert_lock_strong(_bitMap->lock()); 6125 // relinquish the free_list_lock and bitMaplock() 6126 _bitMap->lock()->unlock(); 6127 _freelistLock->unlock(); 6128 ConcurrentMarkSweepThread::desynchronize(true); 6129 _collector->stopTimer(); 6130 _collector->incrementYields(); 6131 6132 // See the comment in coordinator_yield() 6133 for (unsigned i = 0; i < CMSYieldSleepCount && 6134 ConcurrentMarkSweepThread::should_yield() && 6135 !CMSCollector::foregroundGCIsActive(); ++i) { 6136 os::sleep(Thread::current(), 1, false); 6137 } 6138 6139 ConcurrentMarkSweepThread::synchronize(true); 6140 _freelistLock->lock_without_safepoint_check(); 6141 _bitMap->lock()->lock_without_safepoint_check(); 6142 _collector->startTimer(); 6143 } 6144 6145 6146 ////////////////////////////////////////////////////////////////// 6147 // SurvivorSpacePrecleanClosure 6148 ////////////////////////////////////////////////////////////////// 6149 // This (single-threaded) closure is used to preclean the oops in 6150 // the survivor spaces. 6151 size_t SurvivorSpacePrecleanClosure::do_object_careful(oop p) { 6152 6153 HeapWord* addr = (HeapWord*)p; 6154 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6155 assert(!_span.contains(addr), "we are scanning the survivor spaces"); 6156 assert(p->klass_or_null() != NULL, "object should be initialized"); 6157 // an initialized object; ignore mark word in verification below 6158 // since we are running concurrent with mutators 6159 assert(p->is_oop(true), "should be an oop"); 6160 // Note that we do not yield while we iterate over 6161 // the interior oops of p, pushing the relevant ones 6162 // on our marking stack. 6163 size_t size = p->oop_iterate_size(_scanning_closure); 6164 do_yield_check(); 6165 // Observe that below, we do not abandon the preclean 6166 // phase as soon as we should; rather we empty the 6167 // marking stack before returning. This is to satisfy 6168 // some existing assertions. In general, it may be a 6169 // good idea to abort immediately and complete the marking 6170 // from the grey objects at a later time. 6171 while (!_mark_stack->isEmpty()) { 6172 oop new_oop = _mark_stack->pop(); 6173 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 6174 assert(_bit_map->isMarked((HeapWord*)new_oop), 6175 "only grey objects on this stack"); 6176 // iterate over the oops in this oop, marking and pushing 6177 // the ones in CMS heap (i.e. in _span). 6178 new_oop->oop_iterate(_scanning_closure); 6179 // check if it's time to yield 6180 do_yield_check(); 6181 } 6182 unsigned int after_count = 6183 GenCollectedHeap::heap()->total_collections(); 6184 bool abort = (_before_count != after_count) || 6185 _collector->should_abort_preclean(); 6186 return abort ? 0 : size; 6187 } 6188 6189 void SurvivorSpacePrecleanClosure::do_yield_work() { 6190 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6191 "CMS thread should hold CMS token"); 6192 assert_lock_strong(_bit_map->lock()); 6193 // Relinquish the bit map lock 6194 _bit_map->lock()->unlock(); 6195 ConcurrentMarkSweepThread::desynchronize(true); 6196 _collector->stopTimer(); 6197 _collector->incrementYields(); 6198 6199 // See the comment in coordinator_yield() 6200 for (unsigned i = 0; i < CMSYieldSleepCount && 6201 ConcurrentMarkSweepThread::should_yield() && 6202 !CMSCollector::foregroundGCIsActive(); ++i) { 6203 os::sleep(Thread::current(), 1, false); 6204 } 6205 6206 ConcurrentMarkSweepThread::synchronize(true); 6207 _bit_map->lock()->lock_without_safepoint_check(); 6208 _collector->startTimer(); 6209 } 6210 6211 // This closure is used to rescan the marked objects on the dirty cards 6212 // in the mod union table and the card table proper. In the parallel 6213 // case, although the bitMap is shared, we do a single read so the 6214 // isMarked() query is "safe". 6215 bool ScanMarkedObjectsAgainClosure::do_object_bm(oop p, MemRegion mr) { 6216 // Ignore mark word because we are running concurrent with mutators 6217 assert(p->is_oop_or_null(true), "Expected an oop or NULL at " PTR_FORMAT, p2i(p)); 6218 HeapWord* addr = (HeapWord*)p; 6219 assert(_span.contains(addr), "we are scanning the CMS generation"); 6220 bool is_obj_array = false; 6221 #ifdef ASSERT 6222 if (!_parallel) { 6223 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 6224 assert(_collector->overflow_list_is_empty(), 6225 "overflow list should be empty"); 6226 6227 } 6228 #endif // ASSERT 6229 if (_bit_map->isMarked(addr)) { 6230 // Obj arrays are precisely marked, non-arrays are not; 6231 // so we scan objArrays precisely and non-arrays in their 6232 // entirety. 6233 if (p->is_objArray()) { 6234 is_obj_array = true; 6235 if (_parallel) { 6236 p->oop_iterate(_par_scan_closure, mr); 6237 } else { 6238 p->oop_iterate(_scan_closure, mr); 6239 } 6240 } else { 6241 if (_parallel) { 6242 p->oop_iterate(_par_scan_closure); 6243 } else { 6244 p->oop_iterate(_scan_closure); 6245 } 6246 } 6247 } 6248 #ifdef ASSERT 6249 if (!_parallel) { 6250 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 6251 assert(_collector->overflow_list_is_empty(), 6252 "overflow list should be empty"); 6253 6254 } 6255 #endif // ASSERT 6256 return is_obj_array; 6257 } 6258 6259 MarkFromRootsClosure::MarkFromRootsClosure(CMSCollector* collector, 6260 MemRegion span, 6261 CMSBitMap* bitMap, CMSMarkStack* markStack, 6262 bool should_yield, bool verifying): 6263 _collector(collector), 6264 _span(span), 6265 _bitMap(bitMap), 6266 _mut(&collector->_modUnionTable), 6267 _markStack(markStack), 6268 _yield(should_yield), 6269 _skipBits(0) 6270 { 6271 assert(_markStack->isEmpty(), "stack should be empty"); 6272 _finger = _bitMap->startWord(); 6273 _threshold = _finger; 6274 assert(_collector->_restart_addr == NULL, "Sanity check"); 6275 assert(_span.contains(_finger), "Out of bounds _finger?"); 6276 DEBUG_ONLY(_verifying = verifying;) 6277 } 6278 6279 void MarkFromRootsClosure::reset(HeapWord* addr) { 6280 assert(_markStack->isEmpty(), "would cause duplicates on stack"); 6281 assert(_span.contains(addr), "Out of bounds _finger?"); 6282 _finger = addr; 6283 _threshold = (HeapWord*)round_to( 6284 (intptr_t)_finger, CardTableModRefBS::card_size); 6285 } 6286 6287 // Should revisit to see if this should be restructured for 6288 // greater efficiency. 6289 bool MarkFromRootsClosure::do_bit(size_t offset) { 6290 if (_skipBits > 0) { 6291 _skipBits--; 6292 return true; 6293 } 6294 // convert offset into a HeapWord* 6295 HeapWord* addr = _bitMap->startWord() + offset; 6296 assert(_bitMap->endWord() && addr < _bitMap->endWord(), 6297 "address out of range"); 6298 assert(_bitMap->isMarked(addr), "tautology"); 6299 if (_bitMap->isMarked(addr+1)) { 6300 // this is an allocated but not yet initialized object 6301 assert(_skipBits == 0, "tautology"); 6302 _skipBits = 2; // skip next two marked bits ("Printezis-marks") 6303 oop p = oop(addr); 6304 if (p->klass_or_null() == NULL) { 6305 DEBUG_ONLY(if (!_verifying) {) 6306 // We re-dirty the cards on which this object lies and increase 6307 // the _threshold so that we'll come back to scan this object 6308 // during the preclean or remark phase. (CMSCleanOnEnter) 6309 if (CMSCleanOnEnter) { 6310 size_t sz = _collector->block_size_using_printezis_bits(addr); 6311 HeapWord* end_card_addr = (HeapWord*)round_to( 6312 (intptr_t)(addr+sz), CardTableModRefBS::card_size); 6313 MemRegion redirty_range = MemRegion(addr, end_card_addr); 6314 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 6315 // Bump _threshold to end_card_addr; note that 6316 // _threshold cannot possibly exceed end_card_addr, anyhow. 6317 // This prevents future clearing of the card as the scan proceeds 6318 // to the right. 6319 assert(_threshold <= end_card_addr, 6320 "Because we are just scanning into this object"); 6321 if (_threshold < end_card_addr) { 6322 _threshold = end_card_addr; 6323 } 6324 if (p->klass_or_null() != NULL) { 6325 // Redirty the range of cards... 6326 _mut->mark_range(redirty_range); 6327 } // ...else the setting of klass will dirty the card anyway. 6328 } 6329 DEBUG_ONLY(}) 6330 return true; 6331 } 6332 } 6333 scanOopsInOop(addr); 6334 return true; 6335 } 6336 6337 // We take a break if we've been at this for a while, 6338 // so as to avoid monopolizing the locks involved. 6339 void MarkFromRootsClosure::do_yield_work() { 6340 // First give up the locks, then yield, then re-lock 6341 // We should probably use a constructor/destructor idiom to 6342 // do this unlock/lock or modify the MutexUnlocker class to 6343 // serve our purpose. XXX 6344 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6345 "CMS thread should hold CMS token"); 6346 assert_lock_strong(_bitMap->lock()); 6347 _bitMap->lock()->unlock(); 6348 ConcurrentMarkSweepThread::desynchronize(true); 6349 _collector->stopTimer(); 6350 _collector->incrementYields(); 6351 6352 // See the comment in coordinator_yield() 6353 for (unsigned i = 0; i < CMSYieldSleepCount && 6354 ConcurrentMarkSweepThread::should_yield() && 6355 !CMSCollector::foregroundGCIsActive(); ++i) { 6356 os::sleep(Thread::current(), 1, false); 6357 } 6358 6359 ConcurrentMarkSweepThread::synchronize(true); 6360 _bitMap->lock()->lock_without_safepoint_check(); 6361 _collector->startTimer(); 6362 } 6363 6364 void MarkFromRootsClosure::scanOopsInOop(HeapWord* ptr) { 6365 assert(_bitMap->isMarked(ptr), "expected bit to be set"); 6366 assert(_markStack->isEmpty(), 6367 "should drain stack to limit stack usage"); 6368 // convert ptr to an oop preparatory to scanning 6369 oop obj = oop(ptr); 6370 // Ignore mark word in verification below, since we 6371 // may be running concurrent with mutators. 6372 assert(obj->is_oop(true), "should be an oop"); 6373 assert(_finger <= ptr, "_finger runneth ahead"); 6374 // advance the finger to right end of this object 6375 _finger = ptr + obj->size(); 6376 assert(_finger > ptr, "we just incremented it above"); 6377 // On large heaps, it may take us some time to get through 6378 // the marking phase. During 6379 // this time it's possible that a lot of mutations have 6380 // accumulated in the card table and the mod union table -- 6381 // these mutation records are redundant until we have 6382 // actually traced into the corresponding card. 6383 // Here, we check whether advancing the finger would make 6384 // us cross into a new card, and if so clear corresponding 6385 // cards in the MUT (preclean them in the card-table in the 6386 // future). 6387 6388 DEBUG_ONLY(if (!_verifying) {) 6389 // The clean-on-enter optimization is disabled by default, 6390 // until we fix 6178663. 6391 if (CMSCleanOnEnter && (_finger > _threshold)) { 6392 // [_threshold, _finger) represents the interval 6393 // of cards to be cleared in MUT (or precleaned in card table). 6394 // The set of cards to be cleared is all those that overlap 6395 // with the interval [_threshold, _finger); note that 6396 // _threshold is always kept card-aligned but _finger isn't 6397 // always card-aligned. 6398 HeapWord* old_threshold = _threshold; 6399 assert(old_threshold == (HeapWord*)round_to( 6400 (intptr_t)old_threshold, CardTableModRefBS::card_size), 6401 "_threshold should always be card-aligned"); 6402 _threshold = (HeapWord*)round_to( 6403 (intptr_t)_finger, CardTableModRefBS::card_size); 6404 MemRegion mr(old_threshold, _threshold); 6405 assert(!mr.is_empty(), "Control point invariant"); 6406 assert(_span.contains(mr), "Should clear within span"); 6407 _mut->clear_range(mr); 6408 } 6409 DEBUG_ONLY(}) 6410 // Note: the finger doesn't advance while we drain 6411 // the stack below. 6412 PushOrMarkClosure pushOrMarkClosure(_collector, 6413 _span, _bitMap, _markStack, 6414 _finger, this); 6415 bool res = _markStack->push(obj); 6416 assert(res, "Empty non-zero size stack should have space for single push"); 6417 while (!_markStack->isEmpty()) { 6418 oop new_oop = _markStack->pop(); 6419 // Skip verifying header mark word below because we are 6420 // running concurrent with mutators. 6421 assert(new_oop->is_oop(true), "Oops! expected to pop an oop"); 6422 // now scan this oop's oops 6423 new_oop->oop_iterate(&pushOrMarkClosure); 6424 do_yield_check(); 6425 } 6426 assert(_markStack->isEmpty(), "tautology, emphasizing post-condition"); 6427 } 6428 6429 Par_MarkFromRootsClosure::Par_MarkFromRootsClosure(CMSConcMarkingTask* task, 6430 CMSCollector* collector, MemRegion span, 6431 CMSBitMap* bit_map, 6432 OopTaskQueue* work_queue, 6433 CMSMarkStack* overflow_stack): 6434 _collector(collector), 6435 _whole_span(collector->_span), 6436 _span(span), 6437 _bit_map(bit_map), 6438 _mut(&collector->_modUnionTable), 6439 _work_queue(work_queue), 6440 _overflow_stack(overflow_stack), 6441 _skip_bits(0), 6442 _task(task) 6443 { 6444 assert(_work_queue->size() == 0, "work_queue should be empty"); 6445 _finger = span.start(); 6446 _threshold = _finger; // XXX Defer clear-on-enter optimization for now 6447 assert(_span.contains(_finger), "Out of bounds _finger?"); 6448 } 6449 6450 // Should revisit to see if this should be restructured for 6451 // greater efficiency. 6452 bool Par_MarkFromRootsClosure::do_bit(size_t offset) { 6453 if (_skip_bits > 0) { 6454 _skip_bits--; 6455 return true; 6456 } 6457 // convert offset into a HeapWord* 6458 HeapWord* addr = _bit_map->startWord() + offset; 6459 assert(_bit_map->endWord() && addr < _bit_map->endWord(), 6460 "address out of range"); 6461 assert(_bit_map->isMarked(addr), "tautology"); 6462 if (_bit_map->isMarked(addr+1)) { 6463 // this is an allocated object that might not yet be initialized 6464 assert(_skip_bits == 0, "tautology"); 6465 _skip_bits = 2; // skip next two marked bits ("Printezis-marks") 6466 oop p = oop(addr); 6467 if (p->klass_or_null() == NULL) { 6468 // in the case of Clean-on-Enter optimization, redirty card 6469 // and avoid clearing card by increasing the threshold. 6470 return true; 6471 } 6472 } 6473 scan_oops_in_oop(addr); 6474 return true; 6475 } 6476 6477 void Par_MarkFromRootsClosure::scan_oops_in_oop(HeapWord* ptr) { 6478 assert(_bit_map->isMarked(ptr), "expected bit to be set"); 6479 // Should we assert that our work queue is empty or 6480 // below some drain limit? 6481 assert(_work_queue->size() == 0, 6482 "should drain stack to limit stack usage"); 6483 // convert ptr to an oop preparatory to scanning 6484 oop obj = oop(ptr); 6485 // Ignore mark word in verification below, since we 6486 // may be running concurrent with mutators. 6487 assert(obj->is_oop(true), "should be an oop"); 6488 assert(_finger <= ptr, "_finger runneth ahead"); 6489 // advance the finger to right end of this object 6490 _finger = ptr + obj->size(); 6491 assert(_finger > ptr, "we just incremented it above"); 6492 // On large heaps, it may take us some time to get through 6493 // the marking phase. During 6494 // this time it's possible that a lot of mutations have 6495 // accumulated in the card table and the mod union table -- 6496 // these mutation records are redundant until we have 6497 // actually traced into the corresponding card. 6498 // Here, we check whether advancing the finger would make 6499 // us cross into a new card, and if so clear corresponding 6500 // cards in the MUT (preclean them in the card-table in the 6501 // future). 6502 6503 // The clean-on-enter optimization is disabled by default, 6504 // until we fix 6178663. 6505 if (CMSCleanOnEnter && (_finger > _threshold)) { 6506 // [_threshold, _finger) represents the interval 6507 // of cards to be cleared in MUT (or precleaned in card table). 6508 // The set of cards to be cleared is all those that overlap 6509 // with the interval [_threshold, _finger); note that 6510 // _threshold is always kept card-aligned but _finger isn't 6511 // always card-aligned. 6512 HeapWord* old_threshold = _threshold; 6513 assert(old_threshold == (HeapWord*)round_to( 6514 (intptr_t)old_threshold, CardTableModRefBS::card_size), 6515 "_threshold should always be card-aligned"); 6516 _threshold = (HeapWord*)round_to( 6517 (intptr_t)_finger, CardTableModRefBS::card_size); 6518 MemRegion mr(old_threshold, _threshold); 6519 assert(!mr.is_empty(), "Control point invariant"); 6520 assert(_span.contains(mr), "Should clear within span"); // _whole_span ?? 6521 _mut->clear_range(mr); 6522 } 6523 6524 // Note: the local finger doesn't advance while we drain 6525 // the stack below, but the global finger sure can and will. 6526 HeapWord** gfa = _task->global_finger_addr(); 6527 Par_PushOrMarkClosure pushOrMarkClosure(_collector, 6528 _span, _bit_map, 6529 _work_queue, 6530 _overflow_stack, 6531 _finger, 6532 gfa, this); 6533 bool res = _work_queue->push(obj); // overflow could occur here 6534 assert(res, "Will hold once we use workqueues"); 6535 while (true) { 6536 oop new_oop; 6537 if (!_work_queue->pop_local(new_oop)) { 6538 // We emptied our work_queue; check if there's stuff that can 6539 // be gotten from the overflow stack. 6540 if (CMSConcMarkingTask::get_work_from_overflow_stack( 6541 _overflow_stack, _work_queue)) { 6542 do_yield_check(); 6543 continue; 6544 } else { // done 6545 break; 6546 } 6547 } 6548 // Skip verifying header mark word below because we are 6549 // running concurrent with mutators. 6550 assert(new_oop->is_oop(true), "Oops! expected to pop an oop"); 6551 // now scan this oop's oops 6552 new_oop->oop_iterate(&pushOrMarkClosure); 6553 do_yield_check(); 6554 } 6555 assert(_work_queue->size() == 0, "tautology, emphasizing post-condition"); 6556 } 6557 6558 // Yield in response to a request from VM Thread or 6559 // from mutators. 6560 void Par_MarkFromRootsClosure::do_yield_work() { 6561 assert(_task != NULL, "sanity"); 6562 _task->yield(); 6563 } 6564 6565 // A variant of the above used for verifying CMS marking work. 6566 MarkFromRootsVerifyClosure::MarkFromRootsVerifyClosure(CMSCollector* collector, 6567 MemRegion span, 6568 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 6569 CMSMarkStack* mark_stack): 6570 _collector(collector), 6571 _span(span), 6572 _verification_bm(verification_bm), 6573 _cms_bm(cms_bm), 6574 _mark_stack(mark_stack), 6575 _pam_verify_closure(collector, span, verification_bm, cms_bm, 6576 mark_stack) 6577 { 6578 assert(_mark_stack->isEmpty(), "stack should be empty"); 6579 _finger = _verification_bm->startWord(); 6580 assert(_collector->_restart_addr == NULL, "Sanity check"); 6581 assert(_span.contains(_finger), "Out of bounds _finger?"); 6582 } 6583 6584 void MarkFromRootsVerifyClosure::reset(HeapWord* addr) { 6585 assert(_mark_stack->isEmpty(), "would cause duplicates on stack"); 6586 assert(_span.contains(addr), "Out of bounds _finger?"); 6587 _finger = addr; 6588 } 6589 6590 // Should revisit to see if this should be restructured for 6591 // greater efficiency. 6592 bool MarkFromRootsVerifyClosure::do_bit(size_t offset) { 6593 // convert offset into a HeapWord* 6594 HeapWord* addr = _verification_bm->startWord() + offset; 6595 assert(_verification_bm->endWord() && addr < _verification_bm->endWord(), 6596 "address out of range"); 6597 assert(_verification_bm->isMarked(addr), "tautology"); 6598 assert(_cms_bm->isMarked(addr), "tautology"); 6599 6600 assert(_mark_stack->isEmpty(), 6601 "should drain stack to limit stack usage"); 6602 // convert addr to an oop preparatory to scanning 6603 oop obj = oop(addr); 6604 assert(obj->is_oop(), "should be an oop"); 6605 assert(_finger <= addr, "_finger runneth ahead"); 6606 // advance the finger to right end of this object 6607 _finger = addr + obj->size(); 6608 assert(_finger > addr, "we just incremented it above"); 6609 // Note: the finger doesn't advance while we drain 6610 // the stack below. 6611 bool res = _mark_stack->push(obj); 6612 assert(res, "Empty non-zero size stack should have space for single push"); 6613 while (!_mark_stack->isEmpty()) { 6614 oop new_oop = _mark_stack->pop(); 6615 assert(new_oop->is_oop(), "Oops! expected to pop an oop"); 6616 // now scan this oop's oops 6617 new_oop->oop_iterate(&_pam_verify_closure); 6618 } 6619 assert(_mark_stack->isEmpty(), "tautology, emphasizing post-condition"); 6620 return true; 6621 } 6622 6623 PushAndMarkVerifyClosure::PushAndMarkVerifyClosure( 6624 CMSCollector* collector, MemRegion span, 6625 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 6626 CMSMarkStack* mark_stack): 6627 MetadataAwareOopClosure(collector->ref_processor()), 6628 _collector(collector), 6629 _span(span), 6630 _verification_bm(verification_bm), 6631 _cms_bm(cms_bm), 6632 _mark_stack(mark_stack) 6633 { } 6634 6635 void PushAndMarkVerifyClosure::do_oop(oop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 6636 void PushAndMarkVerifyClosure::do_oop(narrowOop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 6637 6638 // Upon stack overflow, we discard (part of) the stack, 6639 // remembering the least address amongst those discarded 6640 // in CMSCollector's _restart_address. 6641 void PushAndMarkVerifyClosure::handle_stack_overflow(HeapWord* lost) { 6642 // Remember the least grey address discarded 6643 HeapWord* ra = (HeapWord*)_mark_stack->least_value(lost); 6644 _collector->lower_restart_addr(ra); 6645 _mark_stack->reset(); // discard stack contents 6646 _mark_stack->expand(); // expand the stack if possible 6647 } 6648 6649 void PushAndMarkVerifyClosure::do_oop(oop obj) { 6650 assert(obj->is_oop_or_null(), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6651 HeapWord* addr = (HeapWord*)obj; 6652 if (_span.contains(addr) && !_verification_bm->isMarked(addr)) { 6653 // Oop lies in _span and isn't yet grey or black 6654 _verification_bm->mark(addr); // now grey 6655 if (!_cms_bm->isMarked(addr)) { 6656 LogHandle(gc, verify) log; 6657 ResourceMark rm; 6658 oop(addr)->print_on(log.info_stream()); 6659 log.info(" (" INTPTR_FORMAT " should have been marked)", p2i(addr)); 6660 fatal("... aborting"); 6661 } 6662 6663 if (!_mark_stack->push(obj)) { // stack overflow 6664 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _mark_stack->capacity()); 6665 assert(_mark_stack->isFull(), "Else push should have succeeded"); 6666 handle_stack_overflow(addr); 6667 } 6668 // anything including and to the right of _finger 6669 // will be scanned as we iterate over the remainder of the 6670 // bit map 6671 } 6672 } 6673 6674 PushOrMarkClosure::PushOrMarkClosure(CMSCollector* collector, 6675 MemRegion span, 6676 CMSBitMap* bitMap, CMSMarkStack* markStack, 6677 HeapWord* finger, MarkFromRootsClosure* parent) : 6678 MetadataAwareOopClosure(collector->ref_processor()), 6679 _collector(collector), 6680 _span(span), 6681 _bitMap(bitMap), 6682 _markStack(markStack), 6683 _finger(finger), 6684 _parent(parent) 6685 { } 6686 6687 Par_PushOrMarkClosure::Par_PushOrMarkClosure(CMSCollector* collector, 6688 MemRegion span, 6689 CMSBitMap* bit_map, 6690 OopTaskQueue* work_queue, 6691 CMSMarkStack* overflow_stack, 6692 HeapWord* finger, 6693 HeapWord** global_finger_addr, 6694 Par_MarkFromRootsClosure* parent) : 6695 MetadataAwareOopClosure(collector->ref_processor()), 6696 _collector(collector), 6697 _whole_span(collector->_span), 6698 _span(span), 6699 _bit_map(bit_map), 6700 _work_queue(work_queue), 6701 _overflow_stack(overflow_stack), 6702 _finger(finger), 6703 _global_finger_addr(global_finger_addr), 6704 _parent(parent) 6705 { } 6706 6707 // Assumes thread-safe access by callers, who are 6708 // responsible for mutual exclusion. 6709 void CMSCollector::lower_restart_addr(HeapWord* low) { 6710 assert(_span.contains(low), "Out of bounds addr"); 6711 if (_restart_addr == NULL) { 6712 _restart_addr = low; 6713 } else { 6714 _restart_addr = MIN2(_restart_addr, low); 6715 } 6716 } 6717 6718 // Upon stack overflow, we discard (part of) the stack, 6719 // remembering the least address amongst those discarded 6720 // in CMSCollector's _restart_address. 6721 void PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 6722 // Remember the least grey address discarded 6723 HeapWord* ra = (HeapWord*)_markStack->least_value(lost); 6724 _collector->lower_restart_addr(ra); 6725 _markStack->reset(); // discard stack contents 6726 _markStack->expand(); // expand the stack if possible 6727 } 6728 6729 // Upon stack overflow, we discard (part of) the stack, 6730 // remembering the least address amongst those discarded 6731 // in CMSCollector's _restart_address. 6732 void Par_PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 6733 // We need to do this under a mutex to prevent other 6734 // workers from interfering with the work done below. 6735 MutexLockerEx ml(_overflow_stack->par_lock(), 6736 Mutex::_no_safepoint_check_flag); 6737 // Remember the least grey address discarded 6738 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 6739 _collector->lower_restart_addr(ra); 6740 _overflow_stack->reset(); // discard stack contents 6741 _overflow_stack->expand(); // expand the stack if possible 6742 } 6743 6744 void PushOrMarkClosure::do_oop(oop obj) { 6745 // Ignore mark word because we are running concurrent with mutators. 6746 assert(obj->is_oop_or_null(true), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6747 HeapWord* addr = (HeapWord*)obj; 6748 if (_span.contains(addr) && !_bitMap->isMarked(addr)) { 6749 // Oop lies in _span and isn't yet grey or black 6750 _bitMap->mark(addr); // now grey 6751 if (addr < _finger) { 6752 // the bit map iteration has already either passed, or 6753 // sampled, this bit in the bit map; we'll need to 6754 // use the marking stack to scan this oop's oops. 6755 bool simulate_overflow = false; 6756 NOT_PRODUCT( 6757 if (CMSMarkStackOverflowALot && 6758 _collector->simulate_overflow()) { 6759 // simulate a stack overflow 6760 simulate_overflow = true; 6761 } 6762 ) 6763 if (simulate_overflow || !_markStack->push(obj)) { // stack overflow 6764 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _markStack->capacity()); 6765 assert(simulate_overflow || _markStack->isFull(), "Else push should have succeeded"); 6766 handle_stack_overflow(addr); 6767 } 6768 } 6769 // anything including and to the right of _finger 6770 // will be scanned as we iterate over the remainder of the 6771 // bit map 6772 do_yield_check(); 6773 } 6774 } 6775 6776 void PushOrMarkClosure::do_oop(oop* p) { PushOrMarkClosure::do_oop_work(p); } 6777 void PushOrMarkClosure::do_oop(narrowOop* p) { PushOrMarkClosure::do_oop_work(p); } 6778 6779 void Par_PushOrMarkClosure::do_oop(oop obj) { 6780 // Ignore mark word because we are running concurrent with mutators. 6781 assert(obj->is_oop_or_null(true), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6782 HeapWord* addr = (HeapWord*)obj; 6783 if (_whole_span.contains(addr) && !_bit_map->isMarked(addr)) { 6784 // Oop lies in _span and isn't yet grey or black 6785 // We read the global_finger (volatile read) strictly after marking oop 6786 bool res = _bit_map->par_mark(addr); // now grey 6787 volatile HeapWord** gfa = (volatile HeapWord**)_global_finger_addr; 6788 // Should we push this marked oop on our stack? 6789 // -- if someone else marked it, nothing to do 6790 // -- if target oop is above global finger nothing to do 6791 // -- if target oop is in chunk and above local finger 6792 // then nothing to do 6793 // -- else push on work queue 6794 if ( !res // someone else marked it, they will deal with it 6795 || (addr >= *gfa) // will be scanned in a later task 6796 || (_span.contains(addr) && addr >= _finger)) { // later in this chunk 6797 return; 6798 } 6799 // the bit map iteration has already either passed, or 6800 // sampled, this bit in the bit map; we'll need to 6801 // use the marking stack to scan this oop's oops. 6802 bool simulate_overflow = false; 6803 NOT_PRODUCT( 6804 if (CMSMarkStackOverflowALot && 6805 _collector->simulate_overflow()) { 6806 // simulate a stack overflow 6807 simulate_overflow = true; 6808 } 6809 ) 6810 if (simulate_overflow || 6811 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) { 6812 // stack overflow 6813 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _overflow_stack->capacity()); 6814 // We cannot assert that the overflow stack is full because 6815 // it may have been emptied since. 6816 assert(simulate_overflow || 6817 _work_queue->size() == _work_queue->max_elems(), 6818 "Else push should have succeeded"); 6819 handle_stack_overflow(addr); 6820 } 6821 do_yield_check(); 6822 } 6823 } 6824 6825 void Par_PushOrMarkClosure::do_oop(oop* p) { Par_PushOrMarkClosure::do_oop_work(p); } 6826 void Par_PushOrMarkClosure::do_oop(narrowOop* p) { Par_PushOrMarkClosure::do_oop_work(p); } 6827 6828 PushAndMarkClosure::PushAndMarkClosure(CMSCollector* collector, 6829 MemRegion span, 6830 ReferenceProcessor* rp, 6831 CMSBitMap* bit_map, 6832 CMSBitMap* mod_union_table, 6833 CMSMarkStack* mark_stack, 6834 bool concurrent_precleaning): 6835 MetadataAwareOopClosure(rp), 6836 _collector(collector), 6837 _span(span), 6838 _bit_map(bit_map), 6839 _mod_union_table(mod_union_table), 6840 _mark_stack(mark_stack), 6841 _concurrent_precleaning(concurrent_precleaning) 6842 { 6843 assert(ref_processor() != NULL, "ref_processor shouldn't be NULL"); 6844 } 6845 6846 // Grey object rescan during pre-cleaning and second checkpoint phases -- 6847 // the non-parallel version (the parallel version appears further below.) 6848 void PushAndMarkClosure::do_oop(oop obj) { 6849 // Ignore mark word verification. If during concurrent precleaning, 6850 // the object monitor may be locked. If during the checkpoint 6851 // phases, the object may already have been reached by a different 6852 // path and may be at the end of the global overflow list (so 6853 // the mark word may be NULL). 6854 assert(obj->is_oop_or_null(true /* ignore mark word */), 6855 "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6856 HeapWord* addr = (HeapWord*)obj; 6857 // Check if oop points into the CMS generation 6858 // and is not marked 6859 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 6860 // a white object ... 6861 _bit_map->mark(addr); // ... now grey 6862 // push on the marking stack (grey set) 6863 bool simulate_overflow = false; 6864 NOT_PRODUCT( 6865 if (CMSMarkStackOverflowALot && 6866 _collector->simulate_overflow()) { 6867 // simulate a stack overflow 6868 simulate_overflow = true; 6869 } 6870 ) 6871 if (simulate_overflow || !_mark_stack->push(obj)) { 6872 if (_concurrent_precleaning) { 6873 // During precleaning we can just dirty the appropriate card(s) 6874 // in the mod union table, thus ensuring that the object remains 6875 // in the grey set and continue. In the case of object arrays 6876 // we need to dirty all of the cards that the object spans, 6877 // since the rescan of object arrays will be limited to the 6878 // dirty cards. 6879 // Note that no one can be interfering with us in this action 6880 // of dirtying the mod union table, so no locking or atomics 6881 // are required. 6882 if (obj->is_objArray()) { 6883 size_t sz = obj->size(); 6884 HeapWord* end_card_addr = (HeapWord*)round_to( 6885 (intptr_t)(addr+sz), CardTableModRefBS::card_size); 6886 MemRegion redirty_range = MemRegion(addr, end_card_addr); 6887 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 6888 _mod_union_table->mark_range(redirty_range); 6889 } else { 6890 _mod_union_table->mark(addr); 6891 } 6892 _collector->_ser_pmc_preclean_ovflw++; 6893 } else { 6894 // During the remark phase, we need to remember this oop 6895 // in the overflow list. 6896 _collector->push_on_overflow_list(obj); 6897 _collector->_ser_pmc_remark_ovflw++; 6898 } 6899 } 6900 } 6901 } 6902 6903 Par_PushAndMarkClosure::Par_PushAndMarkClosure(CMSCollector* collector, 6904 MemRegion span, 6905 ReferenceProcessor* rp, 6906 CMSBitMap* bit_map, 6907 OopTaskQueue* work_queue): 6908 MetadataAwareOopClosure(rp), 6909 _collector(collector), 6910 _span(span), 6911 _bit_map(bit_map), 6912 _work_queue(work_queue) 6913 { 6914 assert(ref_processor() != NULL, "ref_processor shouldn't be NULL"); 6915 } 6916 6917 void PushAndMarkClosure::do_oop(oop* p) { PushAndMarkClosure::do_oop_work(p); } 6918 void PushAndMarkClosure::do_oop(narrowOop* p) { PushAndMarkClosure::do_oop_work(p); } 6919 6920 // Grey object rescan during second checkpoint phase -- 6921 // the parallel version. 6922 void Par_PushAndMarkClosure::do_oop(oop obj) { 6923 // In the assert below, we ignore the mark word because 6924 // this oop may point to an already visited object that is 6925 // on the overflow stack (in which case the mark word has 6926 // been hijacked for chaining into the overflow stack -- 6927 // if this is the last object in the overflow stack then 6928 // its mark word will be NULL). Because this object may 6929 // have been subsequently popped off the global overflow 6930 // stack, and the mark word possibly restored to the prototypical 6931 // value, by the time we get to examined this failing assert in 6932 // the debugger, is_oop_or_null(false) may subsequently start 6933 // to hold. 6934 assert(obj->is_oop_or_null(true), 6935 "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6936 HeapWord* addr = (HeapWord*)obj; 6937 // Check if oop points into the CMS generation 6938 // and is not marked 6939 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 6940 // a white object ... 6941 // If we manage to "claim" the object, by being the 6942 // first thread to mark it, then we push it on our 6943 // marking stack 6944 if (_bit_map->par_mark(addr)) { // ... now grey 6945 // push on work queue (grey set) 6946 bool simulate_overflow = false; 6947 NOT_PRODUCT( 6948 if (CMSMarkStackOverflowALot && 6949 _collector->par_simulate_overflow()) { 6950 // simulate a stack overflow 6951 simulate_overflow = true; 6952 } 6953 ) 6954 if (simulate_overflow || !_work_queue->push(obj)) { 6955 _collector->par_push_on_overflow_list(obj); 6956 _collector->_par_pmc_remark_ovflw++; // imprecise OK: no need to CAS 6957 } 6958 } // Else, some other thread got there first 6959 } 6960 } 6961 6962 void Par_PushAndMarkClosure::do_oop(oop* p) { Par_PushAndMarkClosure::do_oop_work(p); } 6963 void Par_PushAndMarkClosure::do_oop(narrowOop* p) { Par_PushAndMarkClosure::do_oop_work(p); } 6964 6965 void CMSPrecleanRefsYieldClosure::do_yield_work() { 6966 Mutex* bml = _collector->bitMapLock(); 6967 assert_lock_strong(bml); 6968 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6969 "CMS thread should hold CMS token"); 6970 6971 bml->unlock(); 6972 ConcurrentMarkSweepThread::desynchronize(true); 6973 6974 _collector->stopTimer(); 6975 _collector->incrementYields(); 6976 6977 // See the comment in coordinator_yield() 6978 for (unsigned i = 0; i < CMSYieldSleepCount && 6979 ConcurrentMarkSweepThread::should_yield() && 6980 !CMSCollector::foregroundGCIsActive(); ++i) { 6981 os::sleep(Thread::current(), 1, false); 6982 } 6983 6984 ConcurrentMarkSweepThread::synchronize(true); 6985 bml->lock(); 6986 6987 _collector->startTimer(); 6988 } 6989 6990 bool CMSPrecleanRefsYieldClosure::should_return() { 6991 if (ConcurrentMarkSweepThread::should_yield()) { 6992 do_yield_work(); 6993 } 6994 return _collector->foregroundGCIsActive(); 6995 } 6996 6997 void MarkFromDirtyCardsClosure::do_MemRegion(MemRegion mr) { 6998 assert(((size_t)mr.start())%CardTableModRefBS::card_size_in_words == 0, 6999 "mr should be aligned to start at a card boundary"); 7000 // We'd like to assert: 7001 // assert(mr.word_size()%CardTableModRefBS::card_size_in_words == 0, 7002 // "mr should be a range of cards"); 7003 // However, that would be too strong in one case -- the last 7004 // partition ends at _unallocated_block which, in general, can be 7005 // an arbitrary boundary, not necessarily card aligned. 7006 _num_dirty_cards += mr.word_size()/CardTableModRefBS::card_size_in_words; 7007 _space->object_iterate_mem(mr, &_scan_cl); 7008 } 7009 7010 SweepClosure::SweepClosure(CMSCollector* collector, 7011 ConcurrentMarkSweepGeneration* g, 7012 CMSBitMap* bitMap, bool should_yield) : 7013 _collector(collector), 7014 _g(g), 7015 _sp(g->cmsSpace()), 7016 _limit(_sp->sweep_limit()), 7017 _freelistLock(_sp->freelistLock()), 7018 _bitMap(bitMap), 7019 _yield(should_yield), 7020 _inFreeRange(false), // No free range at beginning of sweep 7021 _freeRangeInFreeLists(false), // No free range at beginning of sweep 7022 _lastFreeRangeCoalesced(false), 7023 _freeFinger(g->used_region().start()) 7024 { 7025 NOT_PRODUCT( 7026 _numObjectsFreed = 0; 7027 _numWordsFreed = 0; 7028 _numObjectsLive = 0; 7029 _numWordsLive = 0; 7030 _numObjectsAlreadyFree = 0; 7031 _numWordsAlreadyFree = 0; 7032 _last_fc = NULL; 7033 7034 _sp->initializeIndexedFreeListArrayReturnedBytes(); 7035 _sp->dictionary()->initialize_dict_returned_bytes(); 7036 ) 7037 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7038 "sweep _limit out of bounds"); 7039 log_develop_trace(gc, sweep)("===================="); 7040 log_develop_trace(gc, sweep)("Starting new sweep with limit " PTR_FORMAT, p2i(_limit)); 7041 } 7042 7043 void SweepClosure::print_on(outputStream* st) const { 7044 tty->print_cr("_sp = [" PTR_FORMAT "," PTR_FORMAT ")", 7045 p2i(_sp->bottom()), p2i(_sp->end())); 7046 tty->print_cr("_limit = " PTR_FORMAT, p2i(_limit)); 7047 tty->print_cr("_freeFinger = " PTR_FORMAT, p2i(_freeFinger)); 7048 NOT_PRODUCT(tty->print_cr("_last_fc = " PTR_FORMAT, p2i(_last_fc));) 7049 tty->print_cr("_inFreeRange = %d, _freeRangeInFreeLists = %d, _lastFreeRangeCoalesced = %d", 7050 _inFreeRange, _freeRangeInFreeLists, _lastFreeRangeCoalesced); 7051 } 7052 7053 #ifndef PRODUCT 7054 // Assertion checking only: no useful work in product mode -- 7055 // however, if any of the flags below become product flags, 7056 // you may need to review this code to see if it needs to be 7057 // enabled in product mode. 7058 SweepClosure::~SweepClosure() { 7059 assert_lock_strong(_freelistLock); 7060 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7061 "sweep _limit out of bounds"); 7062 if (inFreeRange()) { 7063 warning("inFreeRange() should have been reset; dumping state of SweepClosure"); 7064 print(); 7065 ShouldNotReachHere(); 7066 } 7067 7068 if (log_is_enabled(Debug, gc, sweep)) { 7069 log_debug(gc, sweep)("Collected " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes", 7070 _numObjectsFreed, _numWordsFreed*sizeof(HeapWord)); 7071 log_debug(gc, sweep)("Live " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes Already free " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes", 7072 _numObjectsLive, _numWordsLive*sizeof(HeapWord), _numObjectsAlreadyFree, _numWordsAlreadyFree*sizeof(HeapWord)); 7073 size_t totalBytes = (_numWordsFreed + _numWordsLive + _numWordsAlreadyFree) * sizeof(HeapWord); 7074 log_debug(gc, sweep)("Total sweep: " SIZE_FORMAT " bytes", totalBytes); 7075 } 7076 7077 if (log_is_enabled(Trace, gc, sweep) && CMSVerifyReturnedBytes) { 7078 size_t indexListReturnedBytes = _sp->sumIndexedFreeListArrayReturnedBytes(); 7079 size_t dict_returned_bytes = _sp->dictionary()->sum_dict_returned_bytes(); 7080 size_t returned_bytes = indexListReturnedBytes + dict_returned_bytes; 7081 log_trace(gc, sweep)("Returned " SIZE_FORMAT " bytes Indexed List Returned " SIZE_FORMAT " bytes Dictionary Returned " SIZE_FORMAT " bytes", 7082 returned_bytes, indexListReturnedBytes, dict_returned_bytes); 7083 } 7084 log_develop_trace(gc, sweep)("end of sweep with _limit = " PTR_FORMAT, p2i(_limit)); 7085 log_develop_trace(gc, sweep)("================"); 7086 } 7087 #endif // PRODUCT 7088 7089 void SweepClosure::initialize_free_range(HeapWord* freeFinger, 7090 bool freeRangeInFreeLists) { 7091 log_develop_trace(gc, sweep)("---- Start free range at " PTR_FORMAT " with free block (%d)", 7092 p2i(freeFinger), freeRangeInFreeLists); 7093 assert(!inFreeRange(), "Trampling existing free range"); 7094 set_inFreeRange(true); 7095 set_lastFreeRangeCoalesced(false); 7096 7097 set_freeFinger(freeFinger); 7098 set_freeRangeInFreeLists(freeRangeInFreeLists); 7099 if (CMSTestInFreeList) { 7100 if (freeRangeInFreeLists) { 7101 FreeChunk* fc = (FreeChunk*) freeFinger; 7102 assert(fc->is_free(), "A chunk on the free list should be free."); 7103 assert(fc->size() > 0, "Free range should have a size"); 7104 assert(_sp->verify_chunk_in_free_list(fc), "Chunk is not in free lists"); 7105 } 7106 } 7107 } 7108 7109 // Note that the sweeper runs concurrently with mutators. Thus, 7110 // it is possible for direct allocation in this generation to happen 7111 // in the middle of the sweep. Note that the sweeper also coalesces 7112 // contiguous free blocks. Thus, unless the sweeper and the allocator 7113 // synchronize appropriately freshly allocated blocks may get swept up. 7114 // This is accomplished by the sweeper locking the free lists while 7115 // it is sweeping. Thus blocks that are determined to be free are 7116 // indeed free. There is however one additional complication: 7117 // blocks that have been allocated since the final checkpoint and 7118 // mark, will not have been marked and so would be treated as 7119 // unreachable and swept up. To prevent this, the allocator marks 7120 // the bit map when allocating during the sweep phase. This leads, 7121 // however, to a further complication -- objects may have been allocated 7122 // but not yet initialized -- in the sense that the header isn't yet 7123 // installed. The sweeper can not then determine the size of the block 7124 // in order to skip over it. To deal with this case, we use a technique 7125 // (due to Printezis) to encode such uninitialized block sizes in the 7126 // bit map. Since the bit map uses a bit per every HeapWord, but the 7127 // CMS generation has a minimum object size of 3 HeapWords, it follows 7128 // that "normal marks" won't be adjacent in the bit map (there will 7129 // always be at least two 0 bits between successive 1 bits). We make use 7130 // of these "unused" bits to represent uninitialized blocks -- the bit 7131 // corresponding to the start of the uninitialized object and the next 7132 // bit are both set. Finally, a 1 bit marks the end of the object that 7133 // started with the two consecutive 1 bits to indicate its potentially 7134 // uninitialized state. 7135 7136 size_t SweepClosure::do_blk_careful(HeapWord* addr) { 7137 FreeChunk* fc = (FreeChunk*)addr; 7138 size_t res; 7139 7140 // Check if we are done sweeping. Below we check "addr >= _limit" rather 7141 // than "addr == _limit" because although _limit was a block boundary when 7142 // we started the sweep, it may no longer be one because heap expansion 7143 // may have caused us to coalesce the block ending at the address _limit 7144 // with a newly expanded chunk (this happens when _limit was set to the 7145 // previous _end of the space), so we may have stepped past _limit: 7146 // see the following Zeno-like trail of CRs 6977970, 7008136, 7042740. 7147 if (addr >= _limit) { // we have swept up to or past the limit: finish up 7148 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7149 "sweep _limit out of bounds"); 7150 assert(addr < _sp->end(), "addr out of bounds"); 7151 // Flush any free range we might be holding as a single 7152 // coalesced chunk to the appropriate free list. 7153 if (inFreeRange()) { 7154 assert(freeFinger() >= _sp->bottom() && freeFinger() < _limit, 7155 "freeFinger() " PTR_FORMAT " is out-of-bounds", p2i(freeFinger())); 7156 flush_cur_free_chunk(freeFinger(), 7157 pointer_delta(addr, freeFinger())); 7158 log_develop_trace(gc, sweep)("Sweep: last chunk: put_free_blk " PTR_FORMAT " (" SIZE_FORMAT ") [coalesced:%d]", 7159 p2i(freeFinger()), pointer_delta(addr, freeFinger()), 7160 lastFreeRangeCoalesced() ? 1 : 0); 7161 } 7162 7163 // help the iterator loop finish 7164 return pointer_delta(_sp->end(), addr); 7165 } 7166 7167 assert(addr < _limit, "sweep invariant"); 7168 // check if we should yield 7169 do_yield_check(addr); 7170 if (fc->is_free()) { 7171 // Chunk that is already free 7172 res = fc->size(); 7173 do_already_free_chunk(fc); 7174 debug_only(_sp->verifyFreeLists()); 7175 // If we flush the chunk at hand in lookahead_and_flush() 7176 // and it's coalesced with a preceding chunk, then the 7177 // process of "mangling" the payload of the coalesced block 7178 // will cause erasure of the size information from the 7179 // (erstwhile) header of all the coalesced blocks but the 7180 // first, so the first disjunct in the assert will not hold 7181 // in that specific case (in which case the second disjunct 7182 // will hold). 7183 assert(res == fc->size() || ((HeapWord*)fc) + res >= _limit, 7184 "Otherwise the size info doesn't change at this step"); 7185 NOT_PRODUCT( 7186 _numObjectsAlreadyFree++; 7187 _numWordsAlreadyFree += res; 7188 ) 7189 NOT_PRODUCT(_last_fc = fc;) 7190 } else if (!_bitMap->isMarked(addr)) { 7191 // Chunk is fresh garbage 7192 res = do_garbage_chunk(fc); 7193 debug_only(_sp->verifyFreeLists()); 7194 NOT_PRODUCT( 7195 _numObjectsFreed++; 7196 _numWordsFreed += res; 7197 ) 7198 } else { 7199 // Chunk that is alive. 7200 res = do_live_chunk(fc); 7201 debug_only(_sp->verifyFreeLists()); 7202 NOT_PRODUCT( 7203 _numObjectsLive++; 7204 _numWordsLive += res; 7205 ) 7206 } 7207 return res; 7208 } 7209 7210 // For the smart allocation, record following 7211 // split deaths - a free chunk is removed from its free list because 7212 // it is being split into two or more chunks. 7213 // split birth - a free chunk is being added to its free list because 7214 // a larger free chunk has been split and resulted in this free chunk. 7215 // coal death - a free chunk is being removed from its free list because 7216 // it is being coalesced into a large free chunk. 7217 // coal birth - a free chunk is being added to its free list because 7218 // it was created when two or more free chunks where coalesced into 7219 // this free chunk. 7220 // 7221 // These statistics are used to determine the desired number of free 7222 // chunks of a given size. The desired number is chosen to be relative 7223 // to the end of a CMS sweep. The desired number at the end of a sweep 7224 // is the 7225 // count-at-end-of-previous-sweep (an amount that was enough) 7226 // - count-at-beginning-of-current-sweep (the excess) 7227 // + split-births (gains in this size during interval) 7228 // - split-deaths (demands on this size during interval) 7229 // where the interval is from the end of one sweep to the end of the 7230 // next. 7231 // 7232 // When sweeping the sweeper maintains an accumulated chunk which is 7233 // the chunk that is made up of chunks that have been coalesced. That 7234 // will be termed the left-hand chunk. A new chunk of garbage that 7235 // is being considered for coalescing will be referred to as the 7236 // right-hand chunk. 7237 // 7238 // When making a decision on whether to coalesce a right-hand chunk with 7239 // the current left-hand chunk, the current count vs. the desired count 7240 // of the left-hand chunk is considered. Also if the right-hand chunk 7241 // is near the large chunk at the end of the heap (see 7242 // ConcurrentMarkSweepGeneration::isNearLargestChunk()), then the 7243 // left-hand chunk is coalesced. 7244 // 7245 // When making a decision about whether to split a chunk, the desired count 7246 // vs. the current count of the candidate to be split is also considered. 7247 // If the candidate is underpopulated (currently fewer chunks than desired) 7248 // a chunk of an overpopulated (currently more chunks than desired) size may 7249 // be chosen. The "hint" associated with a free list, if non-null, points 7250 // to a free list which may be overpopulated. 7251 // 7252 7253 void SweepClosure::do_already_free_chunk(FreeChunk* fc) { 7254 const size_t size = fc->size(); 7255 // Chunks that cannot be coalesced are not in the 7256 // free lists. 7257 if (CMSTestInFreeList && !fc->cantCoalesce()) { 7258 assert(_sp->verify_chunk_in_free_list(fc), 7259 "free chunk should be in free lists"); 7260 } 7261 // a chunk that is already free, should not have been 7262 // marked in the bit map 7263 HeapWord* const addr = (HeapWord*) fc; 7264 assert(!_bitMap->isMarked(addr), "free chunk should be unmarked"); 7265 // Verify that the bit map has no bits marked between 7266 // addr and purported end of this block. 7267 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 7268 7269 // Some chunks cannot be coalesced under any circumstances. 7270 // See the definition of cantCoalesce(). 7271 if (!fc->cantCoalesce()) { 7272 // This chunk can potentially be coalesced. 7273 // All the work is done in 7274 do_post_free_or_garbage_chunk(fc, size); 7275 // Note that if the chunk is not coalescable (the else arm 7276 // below), we unconditionally flush, without needing to do 7277 // a "lookahead," as we do below. 7278 if (inFreeRange()) lookahead_and_flush(fc, size); 7279 } else { 7280 // Code path common to both original and adaptive free lists. 7281 7282 // cant coalesce with previous block; this should be treated 7283 // as the end of a free run if any 7284 if (inFreeRange()) { 7285 // we kicked some butt; time to pick up the garbage 7286 assert(freeFinger() < addr, "freeFinger points too high"); 7287 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 7288 } 7289 // else, nothing to do, just continue 7290 } 7291 } 7292 7293 size_t SweepClosure::do_garbage_chunk(FreeChunk* fc) { 7294 // This is a chunk of garbage. It is not in any free list. 7295 // Add it to a free list or let it possibly be coalesced into 7296 // a larger chunk. 7297 HeapWord* const addr = (HeapWord*) fc; 7298 const size_t size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 7299 7300 // Verify that the bit map has no bits marked between 7301 // addr and purported end of just dead object. 7302 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 7303 do_post_free_or_garbage_chunk(fc, size); 7304 7305 assert(_limit >= addr + size, 7306 "A freshly garbage chunk can't possibly straddle over _limit"); 7307 if (inFreeRange()) lookahead_and_flush(fc, size); 7308 return size; 7309 } 7310 7311 size_t SweepClosure::do_live_chunk(FreeChunk* fc) { 7312 HeapWord* addr = (HeapWord*) fc; 7313 // The sweeper has just found a live object. Return any accumulated 7314 // left hand chunk to the free lists. 7315 if (inFreeRange()) { 7316 assert(freeFinger() < addr, "freeFinger points too high"); 7317 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 7318 } 7319 7320 // This object is live: we'd normally expect this to be 7321 // an oop, and like to assert the following: 7322 // assert(oop(addr)->is_oop(), "live block should be an oop"); 7323 // However, as we commented above, this may be an object whose 7324 // header hasn't yet been initialized. 7325 size_t size; 7326 assert(_bitMap->isMarked(addr), "Tautology for this control point"); 7327 if (_bitMap->isMarked(addr + 1)) { 7328 // Determine the size from the bit map, rather than trying to 7329 // compute it from the object header. 7330 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 7331 size = pointer_delta(nextOneAddr + 1, addr); 7332 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 7333 "alignment problem"); 7334 7335 #ifdef ASSERT 7336 if (oop(addr)->klass_or_null() != NULL) { 7337 // Ignore mark word because we are running concurrent with mutators 7338 assert(oop(addr)->is_oop(true), "live block should be an oop"); 7339 assert(size == 7340 CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()), 7341 "P-mark and computed size do not agree"); 7342 } 7343 #endif 7344 7345 } else { 7346 // This should be an initialized object that's alive. 7347 assert(oop(addr)->klass_or_null() != NULL, 7348 "Should be an initialized object"); 7349 // Ignore mark word because we are running concurrent with mutators 7350 assert(oop(addr)->is_oop(true), "live block should be an oop"); 7351 // Verify that the bit map has no bits marked between 7352 // addr and purported end of this block. 7353 size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 7354 assert(size >= 3, "Necessary for Printezis marks to work"); 7355 assert(!_bitMap->isMarked(addr+1), "Tautology for this control point"); 7356 DEBUG_ONLY(_bitMap->verifyNoOneBitsInRange(addr+2, addr+size);) 7357 } 7358 return size; 7359 } 7360 7361 void SweepClosure::do_post_free_or_garbage_chunk(FreeChunk* fc, 7362 size_t chunkSize) { 7363 // do_post_free_or_garbage_chunk() should only be called in the case 7364 // of the adaptive free list allocator. 7365 const bool fcInFreeLists = fc->is_free(); 7366 assert((HeapWord*)fc <= _limit, "sweep invariant"); 7367 if (CMSTestInFreeList && fcInFreeLists) { 7368 assert(_sp->verify_chunk_in_free_list(fc), "free chunk is not in free lists"); 7369 } 7370 7371 log_develop_trace(gc, sweep)(" -- pick up another chunk at " PTR_FORMAT " (" SIZE_FORMAT ")", p2i(fc), chunkSize); 7372 7373 HeapWord* const fc_addr = (HeapWord*) fc; 7374 7375 bool coalesce = false; 7376 const size_t left = pointer_delta(fc_addr, freeFinger()); 7377 const size_t right = chunkSize; 7378 switch (FLSCoalescePolicy) { 7379 // numeric value forms a coalition aggressiveness metric 7380 case 0: { // never coalesce 7381 coalesce = false; 7382 break; 7383 } 7384 case 1: { // coalesce if left & right chunks on overpopulated lists 7385 coalesce = _sp->coalOverPopulated(left) && 7386 _sp->coalOverPopulated(right); 7387 break; 7388 } 7389 case 2: { // coalesce if left chunk on overpopulated list (default) 7390 coalesce = _sp->coalOverPopulated(left); 7391 break; 7392 } 7393 case 3: { // coalesce if left OR right chunk on overpopulated list 7394 coalesce = _sp->coalOverPopulated(left) || 7395 _sp->coalOverPopulated(right); 7396 break; 7397 } 7398 case 4: { // always coalesce 7399 coalesce = true; 7400 break; 7401 } 7402 default: 7403 ShouldNotReachHere(); 7404 } 7405 7406 // Should the current free range be coalesced? 7407 // If the chunk is in a free range and either we decided to coalesce above 7408 // or the chunk is near the large block at the end of the heap 7409 // (isNearLargestChunk() returns true), then coalesce this chunk. 7410 const bool doCoalesce = inFreeRange() 7411 && (coalesce || _g->isNearLargestChunk(fc_addr)); 7412 if (doCoalesce) { 7413 // Coalesce the current free range on the left with the new 7414 // chunk on the right. If either is on a free list, 7415 // it must be removed from the list and stashed in the closure. 7416 if (freeRangeInFreeLists()) { 7417 FreeChunk* const ffc = (FreeChunk*)freeFinger(); 7418 assert(ffc->size() == pointer_delta(fc_addr, freeFinger()), 7419 "Size of free range is inconsistent with chunk size."); 7420 if (CMSTestInFreeList) { 7421 assert(_sp->verify_chunk_in_free_list(ffc), 7422 "Chunk is not in free lists"); 7423 } 7424 _sp->coalDeath(ffc->size()); 7425 _sp->removeFreeChunkFromFreeLists(ffc); 7426 set_freeRangeInFreeLists(false); 7427 } 7428 if (fcInFreeLists) { 7429 _sp->coalDeath(chunkSize); 7430 assert(fc->size() == chunkSize, 7431 "The chunk has the wrong size or is not in the free lists"); 7432 _sp->removeFreeChunkFromFreeLists(fc); 7433 } 7434 set_lastFreeRangeCoalesced(true); 7435 print_free_block_coalesced(fc); 7436 } else { // not in a free range and/or should not coalesce 7437 // Return the current free range and start a new one. 7438 if (inFreeRange()) { 7439 // In a free range but cannot coalesce with the right hand chunk. 7440 // Put the current free range into the free lists. 7441 flush_cur_free_chunk(freeFinger(), 7442 pointer_delta(fc_addr, freeFinger())); 7443 } 7444 // Set up for new free range. Pass along whether the right hand 7445 // chunk is in the free lists. 7446 initialize_free_range((HeapWord*)fc, fcInFreeLists); 7447 } 7448 } 7449 7450 // Lookahead flush: 7451 // If we are tracking a free range, and this is the last chunk that 7452 // we'll look at because its end crosses past _limit, we'll preemptively 7453 // flush it along with any free range we may be holding on to. Note that 7454 // this can be the case only for an already free or freshly garbage 7455 // chunk. If this block is an object, it can never straddle 7456 // over _limit. The "straddling" occurs when _limit is set at 7457 // the previous end of the space when this cycle started, and 7458 // a subsequent heap expansion caused the previously co-terminal 7459 // free block to be coalesced with the newly expanded portion, 7460 // thus rendering _limit a non-block-boundary making it dangerous 7461 // for the sweeper to step over and examine. 7462 void SweepClosure::lookahead_and_flush(FreeChunk* fc, size_t chunk_size) { 7463 assert(inFreeRange(), "Should only be called if currently in a free range."); 7464 HeapWord* const eob = ((HeapWord*)fc) + chunk_size; 7465 assert(_sp->used_region().contains(eob - 1), 7466 "eob = " PTR_FORMAT " eob-1 = " PTR_FORMAT " _limit = " PTR_FORMAT 7467 " out of bounds wrt _sp = [" PTR_FORMAT "," PTR_FORMAT ")" 7468 " when examining fc = " PTR_FORMAT "(" SIZE_FORMAT ")", 7469 p2i(eob), p2i(eob-1), p2i(_limit), p2i(_sp->bottom()), p2i(_sp->end()), p2i(fc), chunk_size); 7470 if (eob >= _limit) { 7471 assert(eob == _limit || fc->is_free(), "Only a free chunk should allow us to cross over the limit"); 7472 log_develop_trace(gc, sweep)("_limit " PTR_FORMAT " reached or crossed by block " 7473 "[" PTR_FORMAT "," PTR_FORMAT ") in space " 7474 "[" PTR_FORMAT "," PTR_FORMAT ")", 7475 p2i(_limit), p2i(fc), p2i(eob), p2i(_sp->bottom()), p2i(_sp->end())); 7476 // Return the storage we are tracking back into the free lists. 7477 log_develop_trace(gc, sweep)("Flushing ... "); 7478 assert(freeFinger() < eob, "Error"); 7479 flush_cur_free_chunk( freeFinger(), pointer_delta(eob, freeFinger())); 7480 } 7481 } 7482 7483 void SweepClosure::flush_cur_free_chunk(HeapWord* chunk, size_t size) { 7484 assert(inFreeRange(), "Should only be called if currently in a free range."); 7485 assert(size > 0, 7486 "A zero sized chunk cannot be added to the free lists."); 7487 if (!freeRangeInFreeLists()) { 7488 if (CMSTestInFreeList) { 7489 FreeChunk* fc = (FreeChunk*) chunk; 7490 fc->set_size(size); 7491 assert(!_sp->verify_chunk_in_free_list(fc), 7492 "chunk should not be in free lists yet"); 7493 } 7494 log_develop_trace(gc, sweep)(" -- add free block " PTR_FORMAT " (" SIZE_FORMAT ") to free lists", p2i(chunk), size); 7495 // A new free range is going to be starting. The current 7496 // free range has not been added to the free lists yet or 7497 // was removed so add it back. 7498 // If the current free range was coalesced, then the death 7499 // of the free range was recorded. Record a birth now. 7500 if (lastFreeRangeCoalesced()) { 7501 _sp->coalBirth(size); 7502 } 7503 _sp->addChunkAndRepairOffsetTable(chunk, size, 7504 lastFreeRangeCoalesced()); 7505 } else { 7506 log_develop_trace(gc, sweep)("Already in free list: nothing to flush"); 7507 } 7508 set_inFreeRange(false); 7509 set_freeRangeInFreeLists(false); 7510 } 7511 7512 // We take a break if we've been at this for a while, 7513 // so as to avoid monopolizing the locks involved. 7514 void SweepClosure::do_yield_work(HeapWord* addr) { 7515 // Return current free chunk being used for coalescing (if any) 7516 // to the appropriate freelist. After yielding, the next 7517 // free block encountered will start a coalescing range of 7518 // free blocks. If the next free block is adjacent to the 7519 // chunk just flushed, they will need to wait for the next 7520 // sweep to be coalesced. 7521 if (inFreeRange()) { 7522 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 7523 } 7524 7525 // First give up the locks, then yield, then re-lock. 7526 // We should probably use a constructor/destructor idiom to 7527 // do this unlock/lock or modify the MutexUnlocker class to 7528 // serve our purpose. XXX 7529 assert_lock_strong(_bitMap->lock()); 7530 assert_lock_strong(_freelistLock); 7531 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 7532 "CMS thread should hold CMS token"); 7533 _bitMap->lock()->unlock(); 7534 _freelistLock->unlock(); 7535 ConcurrentMarkSweepThread::desynchronize(true); 7536 _collector->stopTimer(); 7537 _collector->incrementYields(); 7538 7539 // See the comment in coordinator_yield() 7540 for (unsigned i = 0; i < CMSYieldSleepCount && 7541 ConcurrentMarkSweepThread::should_yield() && 7542 !CMSCollector::foregroundGCIsActive(); ++i) { 7543 os::sleep(Thread::current(), 1, false); 7544 } 7545 7546 ConcurrentMarkSweepThread::synchronize(true); 7547 _freelistLock->lock(); 7548 _bitMap->lock()->lock_without_safepoint_check(); 7549 _collector->startTimer(); 7550 } 7551 7552 #ifndef PRODUCT 7553 // This is actually very useful in a product build if it can 7554 // be called from the debugger. Compile it into the product 7555 // as needed. 7556 bool debug_verify_chunk_in_free_list(FreeChunk* fc) { 7557 return debug_cms_space->verify_chunk_in_free_list(fc); 7558 } 7559 #endif 7560 7561 void SweepClosure::print_free_block_coalesced(FreeChunk* fc) const { 7562 log_develop_trace(gc, sweep)("Sweep:coal_free_blk " PTR_FORMAT " (" SIZE_FORMAT ")", 7563 p2i(fc), fc->size()); 7564 } 7565 7566 // CMSIsAliveClosure 7567 bool CMSIsAliveClosure::do_object_b(oop obj) { 7568 HeapWord* addr = (HeapWord*)obj; 7569 return addr != NULL && 7570 (!_span.contains(addr) || _bit_map->isMarked(addr)); 7571 } 7572 7573 7574 CMSKeepAliveClosure::CMSKeepAliveClosure( CMSCollector* collector, 7575 MemRegion span, 7576 CMSBitMap* bit_map, CMSMarkStack* mark_stack, 7577 bool cpc): 7578 _collector(collector), 7579 _span(span), 7580 _bit_map(bit_map), 7581 _mark_stack(mark_stack), 7582 _concurrent_precleaning(cpc) { 7583 assert(!_span.is_empty(), "Empty span could spell trouble"); 7584 } 7585 7586 7587 // CMSKeepAliveClosure: the serial version 7588 void CMSKeepAliveClosure::do_oop(oop obj) { 7589 HeapWord* addr = (HeapWord*)obj; 7590 if (_span.contains(addr) && 7591 !_bit_map->isMarked(addr)) { 7592 _bit_map->mark(addr); 7593 bool simulate_overflow = false; 7594 NOT_PRODUCT( 7595 if (CMSMarkStackOverflowALot && 7596 _collector->simulate_overflow()) { 7597 // simulate a stack overflow 7598 simulate_overflow = true; 7599 } 7600 ) 7601 if (simulate_overflow || !_mark_stack->push(obj)) { 7602 if (_concurrent_precleaning) { 7603 // We dirty the overflown object and let the remark 7604 // phase deal with it. 7605 assert(_collector->overflow_list_is_empty(), "Error"); 7606 // In the case of object arrays, we need to dirty all of 7607 // the cards that the object spans. No locking or atomics 7608 // are needed since no one else can be mutating the mod union 7609 // table. 7610 if (obj->is_objArray()) { 7611 size_t sz = obj->size(); 7612 HeapWord* end_card_addr = 7613 (HeapWord*)round_to((intptr_t)(addr+sz), CardTableModRefBS::card_size); 7614 MemRegion redirty_range = MemRegion(addr, end_card_addr); 7615 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 7616 _collector->_modUnionTable.mark_range(redirty_range); 7617 } else { 7618 _collector->_modUnionTable.mark(addr); 7619 } 7620 _collector->_ser_kac_preclean_ovflw++; 7621 } else { 7622 _collector->push_on_overflow_list(obj); 7623 _collector->_ser_kac_ovflw++; 7624 } 7625 } 7626 } 7627 } 7628 7629 void CMSKeepAliveClosure::do_oop(oop* p) { CMSKeepAliveClosure::do_oop_work(p); } 7630 void CMSKeepAliveClosure::do_oop(narrowOop* p) { CMSKeepAliveClosure::do_oop_work(p); } 7631 7632 // CMSParKeepAliveClosure: a parallel version of the above. 7633 // The work queues are private to each closure (thread), 7634 // but (may be) available for stealing by other threads. 7635 void CMSParKeepAliveClosure::do_oop(oop obj) { 7636 HeapWord* addr = (HeapWord*)obj; 7637 if (_span.contains(addr) && 7638 !_bit_map->isMarked(addr)) { 7639 // In general, during recursive tracing, several threads 7640 // may be concurrently getting here; the first one to 7641 // "tag" it, claims it. 7642 if (_bit_map->par_mark(addr)) { 7643 bool res = _work_queue->push(obj); 7644 assert(res, "Low water mark should be much less than capacity"); 7645 // Do a recursive trim in the hope that this will keep 7646 // stack usage lower, but leave some oops for potential stealers 7647 trim_queue(_low_water_mark); 7648 } // Else, another thread got there first 7649 } 7650 } 7651 7652 void CMSParKeepAliveClosure::do_oop(oop* p) { CMSParKeepAliveClosure::do_oop_work(p); } 7653 void CMSParKeepAliveClosure::do_oop(narrowOop* p) { CMSParKeepAliveClosure::do_oop_work(p); } 7654 7655 void CMSParKeepAliveClosure::trim_queue(uint max) { 7656 while (_work_queue->size() > max) { 7657 oop new_oop; 7658 if (_work_queue->pop_local(new_oop)) { 7659 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 7660 assert(_bit_map->isMarked((HeapWord*)new_oop), 7661 "no white objects on this stack!"); 7662 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 7663 // iterate over the oops in this oop, marking and pushing 7664 // the ones in CMS heap (i.e. in _span). 7665 new_oop->oop_iterate(&_mark_and_push); 7666 } 7667 } 7668 } 7669 7670 CMSInnerParMarkAndPushClosure::CMSInnerParMarkAndPushClosure( 7671 CMSCollector* collector, 7672 MemRegion span, CMSBitMap* bit_map, 7673 OopTaskQueue* work_queue): 7674 _collector(collector), 7675 _span(span), 7676 _bit_map(bit_map), 7677 _work_queue(work_queue) { } 7678 7679 void CMSInnerParMarkAndPushClosure::do_oop(oop obj) { 7680 HeapWord* addr = (HeapWord*)obj; 7681 if (_span.contains(addr) && 7682 !_bit_map->isMarked(addr)) { 7683 if (_bit_map->par_mark(addr)) { 7684 bool simulate_overflow = false; 7685 NOT_PRODUCT( 7686 if (CMSMarkStackOverflowALot && 7687 _collector->par_simulate_overflow()) { 7688 // simulate a stack overflow 7689 simulate_overflow = true; 7690 } 7691 ) 7692 if (simulate_overflow || !_work_queue->push(obj)) { 7693 _collector->par_push_on_overflow_list(obj); 7694 _collector->_par_kac_ovflw++; 7695 } 7696 } // Else another thread got there already 7697 } 7698 } 7699 7700 void CMSInnerParMarkAndPushClosure::do_oop(oop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); } 7701 void CMSInnerParMarkAndPushClosure::do_oop(narrowOop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); } 7702 7703 ////////////////////////////////////////////////////////////////// 7704 // CMSExpansionCause ///////////////////////////// 7705 ////////////////////////////////////////////////////////////////// 7706 const char* CMSExpansionCause::to_string(CMSExpansionCause::Cause cause) { 7707 switch (cause) { 7708 case _no_expansion: 7709 return "No expansion"; 7710 case _satisfy_free_ratio: 7711 return "Free ratio"; 7712 case _satisfy_promotion: 7713 return "Satisfy promotion"; 7714 case _satisfy_allocation: 7715 return "allocation"; 7716 case _allocate_par_lab: 7717 return "Par LAB"; 7718 case _allocate_par_spooling_space: 7719 return "Par Spooling Space"; 7720 case _adaptive_size_policy: 7721 return "Ergonomics"; 7722 default: 7723 return "unknown"; 7724 } 7725 } 7726 7727 void CMSDrainMarkingStackClosure::do_void() { 7728 // the max number to take from overflow list at a time 7729 const size_t num = _mark_stack->capacity()/4; 7730 assert(!_concurrent_precleaning || _collector->overflow_list_is_empty(), 7731 "Overflow list should be NULL during concurrent phases"); 7732 while (!_mark_stack->isEmpty() || 7733 // if stack is empty, check the overflow list 7734 _collector->take_from_overflow_list(num, _mark_stack)) { 7735 oop obj = _mark_stack->pop(); 7736 HeapWord* addr = (HeapWord*)obj; 7737 assert(_span.contains(addr), "Should be within span"); 7738 assert(_bit_map->isMarked(addr), "Should be marked"); 7739 assert(obj->is_oop(), "Should be an oop"); 7740 obj->oop_iterate(_keep_alive); 7741 } 7742 } 7743 7744 void CMSParDrainMarkingStackClosure::do_void() { 7745 // drain queue 7746 trim_queue(0); 7747 } 7748 7749 // Trim our work_queue so its length is below max at return 7750 void CMSParDrainMarkingStackClosure::trim_queue(uint max) { 7751 while (_work_queue->size() > max) { 7752 oop new_oop; 7753 if (_work_queue->pop_local(new_oop)) { 7754 assert(new_oop->is_oop(), "Expected an oop"); 7755 assert(_bit_map->isMarked((HeapWord*)new_oop), 7756 "no white objects on this stack!"); 7757 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 7758 // iterate over the oops in this oop, marking and pushing 7759 // the ones in CMS heap (i.e. in _span). 7760 new_oop->oop_iterate(&_mark_and_push); 7761 } 7762 } 7763 } 7764 7765 //////////////////////////////////////////////////////////////////// 7766 // Support for Marking Stack Overflow list handling and related code 7767 //////////////////////////////////////////////////////////////////// 7768 // Much of the following code is similar in shape and spirit to the 7769 // code used in ParNewGC. We should try and share that code 7770 // as much as possible in the future. 7771 7772 #ifndef PRODUCT 7773 // Debugging support for CMSStackOverflowALot 7774 7775 // It's OK to call this multi-threaded; the worst thing 7776 // that can happen is that we'll get a bunch of closely 7777 // spaced simulated overflows, but that's OK, in fact 7778 // probably good as it would exercise the overflow code 7779 // under contention. 7780 bool CMSCollector::simulate_overflow() { 7781 if (_overflow_counter-- <= 0) { // just being defensive 7782 _overflow_counter = CMSMarkStackOverflowInterval; 7783 return true; 7784 } else { 7785 return false; 7786 } 7787 } 7788 7789 bool CMSCollector::par_simulate_overflow() { 7790 return simulate_overflow(); 7791 } 7792 #endif 7793 7794 // Single-threaded 7795 bool CMSCollector::take_from_overflow_list(size_t num, CMSMarkStack* stack) { 7796 assert(stack->isEmpty(), "Expected precondition"); 7797 assert(stack->capacity() > num, "Shouldn't bite more than can chew"); 7798 size_t i = num; 7799 oop cur = _overflow_list; 7800 const markOop proto = markOopDesc::prototype(); 7801 NOT_PRODUCT(ssize_t n = 0;) 7802 for (oop next; i > 0 && cur != NULL; cur = next, i--) { 7803 next = oop(cur->mark()); 7804 cur->set_mark(proto); // until proven otherwise 7805 assert(cur->is_oop(), "Should be an oop"); 7806 bool res = stack->push(cur); 7807 assert(res, "Bit off more than can chew?"); 7808 NOT_PRODUCT(n++;) 7809 } 7810 _overflow_list = cur; 7811 #ifndef PRODUCT 7812 assert(_num_par_pushes >= n, "Too many pops?"); 7813 _num_par_pushes -=n; 7814 #endif 7815 return !stack->isEmpty(); 7816 } 7817 7818 #define BUSY (cast_to_oop<intptr_t>(0x1aff1aff)) 7819 // (MT-safe) Get a prefix of at most "num" from the list. 7820 // The overflow list is chained through the mark word of 7821 // each object in the list. We fetch the entire list, 7822 // break off a prefix of the right size and return the 7823 // remainder. If other threads try to take objects from 7824 // the overflow list at that time, they will wait for 7825 // some time to see if data becomes available. If (and 7826 // only if) another thread places one or more object(s) 7827 // on the global list before we have returned the suffix 7828 // to the global list, we will walk down our local list 7829 // to find its end and append the global list to 7830 // our suffix before returning it. This suffix walk can 7831 // prove to be expensive (quadratic in the amount of traffic) 7832 // when there are many objects in the overflow list and 7833 // there is much producer-consumer contention on the list. 7834 // *NOTE*: The overflow list manipulation code here and 7835 // in ParNewGeneration:: are very similar in shape, 7836 // except that in the ParNew case we use the old (from/eden) 7837 // copy of the object to thread the list via its klass word. 7838 // Because of the common code, if you make any changes in 7839 // the code below, please check the ParNew version to see if 7840 // similar changes might be needed. 7841 // CR 6797058 has been filed to consolidate the common code. 7842 bool CMSCollector::par_take_from_overflow_list(size_t num, 7843 OopTaskQueue* work_q, 7844 int no_of_gc_threads) { 7845 assert(work_q->size() == 0, "First empty local work queue"); 7846 assert(num < work_q->max_elems(), "Can't bite more than we can chew"); 7847 if (_overflow_list == NULL) { 7848 return false; 7849 } 7850 // Grab the entire list; we'll put back a suffix 7851 oop prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list)); 7852 Thread* tid = Thread::current(); 7853 // Before "no_of_gc_threads" was introduced CMSOverflowSpinCount was 7854 // set to ParallelGCThreads. 7855 size_t CMSOverflowSpinCount = (size_t) no_of_gc_threads; // was ParallelGCThreads; 7856 size_t sleep_time_millis = MAX2((size_t)1, num/100); 7857 // If the list is busy, we spin for a short while, 7858 // sleeping between attempts to get the list. 7859 for (size_t spin = 0; prefix == BUSY && spin < CMSOverflowSpinCount; spin++) { 7860 os::sleep(tid, sleep_time_millis, false); 7861 if (_overflow_list == NULL) { 7862 // Nothing left to take 7863 return false; 7864 } else if (_overflow_list != BUSY) { 7865 // Try and grab the prefix 7866 prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list)); 7867 } 7868 } 7869 // If the list was found to be empty, or we spun long 7870 // enough, we give up and return empty-handed. If we leave 7871 // the list in the BUSY state below, it must be the case that 7872 // some other thread holds the overflow list and will set it 7873 // to a non-BUSY state in the future. 7874 if (prefix == NULL || prefix == BUSY) { 7875 // Nothing to take or waited long enough 7876 if (prefix == NULL) { 7877 // Write back the NULL in case we overwrote it with BUSY above 7878 // and it is still the same value. 7879 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY); 7880 } 7881 return false; 7882 } 7883 assert(prefix != NULL && prefix != BUSY, "Error"); 7884 size_t i = num; 7885 oop cur = prefix; 7886 // Walk down the first "num" objects, unless we reach the end. 7887 for (; i > 1 && cur->mark() != NULL; cur = oop(cur->mark()), i--); 7888 if (cur->mark() == NULL) { 7889 // We have "num" or fewer elements in the list, so there 7890 // is nothing to return to the global list. 7891 // Write back the NULL in lieu of the BUSY we wrote 7892 // above, if it is still the same value. 7893 if (_overflow_list == BUSY) { 7894 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY); 7895 } 7896 } else { 7897 // Chop off the suffix and return it to the global list. 7898 assert(cur->mark() != BUSY, "Error"); 7899 oop suffix_head = cur->mark(); // suffix will be put back on global list 7900 cur->set_mark(NULL); // break off suffix 7901 // It's possible that the list is still in the empty(busy) state 7902 // we left it in a short while ago; in that case we may be 7903 // able to place back the suffix without incurring the cost 7904 // of a walk down the list. 7905 oop observed_overflow_list = _overflow_list; 7906 oop cur_overflow_list = observed_overflow_list; 7907 bool attached = false; 7908 while (observed_overflow_list == BUSY || observed_overflow_list == NULL) { 7909 observed_overflow_list = 7910 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list); 7911 if (cur_overflow_list == observed_overflow_list) { 7912 attached = true; 7913 break; 7914 } else cur_overflow_list = observed_overflow_list; 7915 } 7916 if (!attached) { 7917 // Too bad, someone else sneaked in (at least) an element; we'll need 7918 // to do a splice. Find tail of suffix so we can prepend suffix to global 7919 // list. 7920 for (cur = suffix_head; cur->mark() != NULL; cur = (oop)(cur->mark())); 7921 oop suffix_tail = cur; 7922 assert(suffix_tail != NULL && suffix_tail->mark() == NULL, 7923 "Tautology"); 7924 observed_overflow_list = _overflow_list; 7925 do { 7926 cur_overflow_list = observed_overflow_list; 7927 if (cur_overflow_list != BUSY) { 7928 // Do the splice ... 7929 suffix_tail->set_mark(markOop(cur_overflow_list)); 7930 } else { // cur_overflow_list == BUSY 7931 suffix_tail->set_mark(NULL); 7932 } 7933 // ... and try to place spliced list back on overflow_list ... 7934 observed_overflow_list = 7935 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list); 7936 } while (cur_overflow_list != observed_overflow_list); 7937 // ... until we have succeeded in doing so. 7938 } 7939 } 7940 7941 // Push the prefix elements on work_q 7942 assert(prefix != NULL, "control point invariant"); 7943 const markOop proto = markOopDesc::prototype(); 7944 oop next; 7945 NOT_PRODUCT(ssize_t n = 0;) 7946 for (cur = prefix; cur != NULL; cur = next) { 7947 next = oop(cur->mark()); 7948 cur->set_mark(proto); // until proven otherwise 7949 assert(cur->is_oop(), "Should be an oop"); 7950 bool res = work_q->push(cur); 7951 assert(res, "Bit off more than we can chew?"); 7952 NOT_PRODUCT(n++;) 7953 } 7954 #ifndef PRODUCT 7955 assert(_num_par_pushes >= n, "Too many pops?"); 7956 Atomic::add_ptr(-(intptr_t)n, &_num_par_pushes); 7957 #endif 7958 return true; 7959 } 7960 7961 // Single-threaded 7962 void CMSCollector::push_on_overflow_list(oop p) { 7963 NOT_PRODUCT(_num_par_pushes++;) 7964 assert(p->is_oop(), "Not an oop"); 7965 preserve_mark_if_necessary(p); 7966 p->set_mark((markOop)_overflow_list); 7967 _overflow_list = p; 7968 } 7969 7970 // Multi-threaded; use CAS to prepend to overflow list 7971 void CMSCollector::par_push_on_overflow_list(oop p) { 7972 NOT_PRODUCT(Atomic::inc_ptr(&_num_par_pushes);) 7973 assert(p->is_oop(), "Not an oop"); 7974 par_preserve_mark_if_necessary(p); 7975 oop observed_overflow_list = _overflow_list; 7976 oop cur_overflow_list; 7977 do { 7978 cur_overflow_list = observed_overflow_list; 7979 if (cur_overflow_list != BUSY) { 7980 p->set_mark(markOop(cur_overflow_list)); 7981 } else { 7982 p->set_mark(NULL); 7983 } 7984 observed_overflow_list = 7985 (oop) Atomic::cmpxchg_ptr(p, &_overflow_list, cur_overflow_list); 7986 } while (cur_overflow_list != observed_overflow_list); 7987 } 7988 #undef BUSY 7989 7990 // Single threaded 7991 // General Note on GrowableArray: pushes may silently fail 7992 // because we are (temporarily) out of C-heap for expanding 7993 // the stack. The problem is quite ubiquitous and affects 7994 // a lot of code in the JVM. The prudent thing for GrowableArray 7995 // to do (for now) is to exit with an error. However, that may 7996 // be too draconian in some cases because the caller may be 7997 // able to recover without much harm. For such cases, we 7998 // should probably introduce a "soft_push" method which returns 7999 // an indication of success or failure with the assumption that 8000 // the caller may be able to recover from a failure; code in 8001 // the VM can then be changed, incrementally, to deal with such 8002 // failures where possible, thus, incrementally hardening the VM 8003 // in such low resource situations. 8004 void CMSCollector::preserve_mark_work(oop p, markOop m) { 8005 _preserved_oop_stack.push(p); 8006 _preserved_mark_stack.push(m); 8007 assert(m == p->mark(), "Mark word changed"); 8008 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), 8009 "bijection"); 8010 } 8011 8012 // Single threaded 8013 void CMSCollector::preserve_mark_if_necessary(oop p) { 8014 markOop m = p->mark(); 8015 if (m->must_be_preserved(p)) { 8016 preserve_mark_work(p, m); 8017 } 8018 } 8019 8020 void CMSCollector::par_preserve_mark_if_necessary(oop p) { 8021 markOop m = p->mark(); 8022 if (m->must_be_preserved(p)) { 8023 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag); 8024 // Even though we read the mark word without holding 8025 // the lock, we are assured that it will not change 8026 // because we "own" this oop, so no other thread can 8027 // be trying to push it on the overflow list; see 8028 // the assertion in preserve_mark_work() that checks 8029 // that m == p->mark(). 8030 preserve_mark_work(p, m); 8031 } 8032 } 8033 8034 // We should be able to do this multi-threaded, 8035 // a chunk of stack being a task (this is 8036 // correct because each oop only ever appears 8037 // once in the overflow list. However, it's 8038 // not very easy to completely overlap this with 8039 // other operations, so will generally not be done 8040 // until all work's been completed. Because we 8041 // expect the preserved oop stack (set) to be small, 8042 // it's probably fine to do this single-threaded. 8043 // We can explore cleverer concurrent/overlapped/parallel 8044 // processing of preserved marks if we feel the 8045 // need for this in the future. Stack overflow should 8046 // be so rare in practice and, when it happens, its 8047 // effect on performance so great that this will 8048 // likely just be in the noise anyway. 8049 void CMSCollector::restore_preserved_marks_if_any() { 8050 assert(SafepointSynchronize::is_at_safepoint(), 8051 "world should be stopped"); 8052 assert(Thread::current()->is_ConcurrentGC_thread() || 8053 Thread::current()->is_VM_thread(), 8054 "should be single-threaded"); 8055 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), 8056 "bijection"); 8057 8058 while (!_preserved_oop_stack.is_empty()) { 8059 oop p = _preserved_oop_stack.pop(); 8060 assert(p->is_oop(), "Should be an oop"); 8061 assert(_span.contains(p), "oop should be in _span"); 8062 assert(p->mark() == markOopDesc::prototype(), 8063 "Set when taken from overflow list"); 8064 markOop m = _preserved_mark_stack.pop(); 8065 p->set_mark(m); 8066 } 8067 assert(_preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(), 8068 "stacks were cleared above"); 8069 } 8070 8071 #ifndef PRODUCT 8072 bool CMSCollector::no_preserved_marks() const { 8073 return _preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(); 8074 } 8075 #endif 8076 8077 // Transfer some number of overflown objects to usual marking 8078 // stack. Return true if some objects were transferred. 8079 bool MarkRefsIntoAndScanClosure::take_from_overflow_list() { 8080 size_t num = MIN2((size_t)(_mark_stack->capacity() - _mark_stack->length())/4, 8081 (size_t)ParGCDesiredObjsFromOverflowList); 8082 8083 bool res = _collector->take_from_overflow_list(num, _mark_stack); 8084 assert(_collector->overflow_list_is_empty() || res, 8085 "If list is not empty, we should have taken something"); 8086 assert(!res || !_mark_stack->isEmpty(), 8087 "If we took something, it should now be on our stack"); 8088 return res; 8089 } 8090 8091 size_t MarkDeadObjectsClosure::do_blk(HeapWord* addr) { 8092 size_t res = _sp->block_size_no_stall(addr, _collector); 8093 if (_sp->block_is_obj(addr)) { 8094 if (_live_bit_map->isMarked(addr)) { 8095 // It can't have been dead in a previous cycle 8096 guarantee(!_dead_bit_map->isMarked(addr), "No resurrection!"); 8097 } else { 8098 _dead_bit_map->mark(addr); // mark the dead object 8099 } 8100 } 8101 // Could be 0, if the block size could not be computed without stalling. 8102 return res; 8103 } 8104 8105 TraceCMSMemoryManagerStats::TraceCMSMemoryManagerStats(CMSCollector::CollectorState phase, GCCause::Cause cause): TraceMemoryManagerStats() { 8106 8107 switch (phase) { 8108 case CMSCollector::InitialMarking: 8109 initialize(true /* fullGC */ , 8110 cause /* cause of the GC */, 8111 true /* recordGCBeginTime */, 8112 true /* recordPreGCUsage */, 8113 false /* recordPeakUsage */, 8114 false /* recordPostGCusage */, 8115 true /* recordAccumulatedGCTime */, 8116 false /* recordGCEndTime */, 8117 false /* countCollection */ ); 8118 break; 8119 8120 case CMSCollector::FinalMarking: 8121 initialize(true /* fullGC */ , 8122 cause /* cause of the GC */, 8123 false /* recordGCBeginTime */, 8124 false /* recordPreGCUsage */, 8125 false /* recordPeakUsage */, 8126 false /* recordPostGCusage */, 8127 true /* recordAccumulatedGCTime */, 8128 false /* recordGCEndTime */, 8129 false /* countCollection */ ); 8130 break; 8131 8132 case CMSCollector::Sweeping: 8133 initialize(true /* fullGC */ , 8134 cause /* cause of the GC */, 8135 false /* recordGCBeginTime */, 8136 false /* recordPreGCUsage */, 8137 true /* recordPeakUsage */, 8138 true /* recordPostGCusage */, 8139 false /* recordAccumulatedGCTime */, 8140 true /* recordGCEndTime */, 8141 true /* countCollection */ ); 8142 break; 8143 8144 default: 8145 ShouldNotReachHere(); 8146 } 8147 }