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 jlong 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 } 2746 2747 CMSPhaseAccounting::~CMSPhaseAccounting() { 2748 _collector->stopTimer(); 2749 log_debug(gc)("Concurrent active time: %.3fms", TimeHelper::counter_to_seconds(_collector->timerTicks())); 2750 log_trace(gc)(" (CMS %s yielded %d times)", _title, _collector->yields()); 2751 } 2752 2753 // CMS work 2754 2755 // The common parts of CMSParInitialMarkTask and CMSParRemarkTask. 2756 class CMSParMarkTask : public AbstractGangTask { 2757 protected: 2758 CMSCollector* _collector; 2759 uint _n_workers; 2760 CMSParMarkTask(const char* name, CMSCollector* collector, uint n_workers) : 2761 AbstractGangTask(name), 2762 _collector(collector), 2763 _n_workers(n_workers) {} 2764 // Work method in support of parallel rescan ... of young gen spaces 2765 void do_young_space_rescan(uint worker_id, OopsInGenClosure* cl, 2766 ContiguousSpace* space, 2767 HeapWord** chunk_array, size_t chunk_top); 2768 void work_on_young_gen_roots(uint worker_id, OopsInGenClosure* cl); 2769 }; 2770 2771 // Parallel initial mark task 2772 class CMSParInitialMarkTask: public CMSParMarkTask { 2773 StrongRootsScope* _strong_roots_scope; 2774 public: 2775 CMSParInitialMarkTask(CMSCollector* collector, StrongRootsScope* strong_roots_scope, uint n_workers) : 2776 CMSParMarkTask("Scan roots and young gen for initial mark in parallel", collector, n_workers), 2777 _strong_roots_scope(strong_roots_scope) {} 2778 void work(uint worker_id); 2779 }; 2780 2781 // Checkpoint the roots into this generation from outside 2782 // this generation. [Note this initial checkpoint need only 2783 // be approximate -- we'll do a catch up phase subsequently.] 2784 void CMSCollector::checkpointRootsInitial() { 2785 assert(_collectorState == InitialMarking, "Wrong collector state"); 2786 check_correct_thread_executing(); 2787 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 2788 2789 save_heap_summary(); 2790 report_heap_summary(GCWhen::BeforeGC); 2791 2792 ReferenceProcessor* rp = ref_processor(); 2793 assert(_restart_addr == NULL, "Control point invariant"); 2794 { 2795 // acquire locks for subsequent manipulations 2796 MutexLockerEx x(bitMapLock(), 2797 Mutex::_no_safepoint_check_flag); 2798 checkpointRootsInitialWork(); 2799 // enable ("weak") refs discovery 2800 rp->enable_discovery(); 2801 _collectorState = Marking; 2802 } 2803 } 2804 2805 void CMSCollector::checkpointRootsInitialWork() { 2806 assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped"); 2807 assert(_collectorState == InitialMarking, "just checking"); 2808 2809 // Already have locks. 2810 assert_lock_strong(bitMapLock()); 2811 assert(_markBitMap.isAllClear(), "was reset at end of previous cycle"); 2812 2813 // Setup the verification and class unloading state for this 2814 // CMS collection cycle. 2815 setup_cms_unloading_and_verification_state(); 2816 2817 GCTraceTime(Trace, gc) ts("checkpointRootsInitialWork", _gc_timer_cm); 2818 2819 // Reset all the PLAB chunk arrays if necessary. 2820 if (_survivor_plab_array != NULL && !CMSPLABRecordAlways) { 2821 reset_survivor_plab_arrays(); 2822 } 2823 2824 ResourceMark rm; 2825 HandleMark hm; 2826 2827 MarkRefsIntoClosure notOlder(_span, &_markBitMap); 2828 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2829 2830 verify_work_stacks_empty(); 2831 verify_overflow_empty(); 2832 2833 gch->ensure_parsability(false); // fill TLABs, but no need to retire them 2834 // Update the saved marks which may affect the root scans. 2835 gch->save_marks(); 2836 2837 // weak reference processing has not started yet. 2838 ref_processor()->set_enqueuing_is_done(false); 2839 2840 // Need to remember all newly created CLDs, 2841 // so that we can guarantee that the remark finds them. 2842 ClassLoaderDataGraph::remember_new_clds(true); 2843 2844 // Whenever a CLD is found, it will be claimed before proceeding to mark 2845 // the klasses. The claimed marks need to be cleared before marking starts. 2846 ClassLoaderDataGraph::clear_claimed_marks(); 2847 2848 print_eden_and_survivor_chunk_arrays(); 2849 2850 { 2851 #if defined(COMPILER2) || INCLUDE_JVMCI 2852 DerivedPointerTableDeactivate dpt_deact; 2853 #endif 2854 if (CMSParallelInitialMarkEnabled) { 2855 // The parallel version. 2856 WorkGang* workers = gch->workers(); 2857 assert(workers != NULL, "Need parallel worker threads."); 2858 uint n_workers = workers->active_workers(); 2859 2860 StrongRootsScope srs(n_workers); 2861 2862 CMSParInitialMarkTask tsk(this, &srs, n_workers); 2863 initialize_sequential_subtasks_for_young_gen_rescan(n_workers); 2864 if (n_workers > 1) { 2865 workers->run_task(&tsk); 2866 } else { 2867 tsk.work(0); 2868 } 2869 } else { 2870 // The serial version. 2871 CLDToOopClosure cld_closure(¬Older, true); 2872 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2873 2874 StrongRootsScope srs(1); 2875 2876 gch->gen_process_roots(&srs, 2877 GenCollectedHeap::OldGen, 2878 true, // young gen as roots 2879 GenCollectedHeap::ScanningOption(roots_scanning_options()), 2880 should_unload_classes(), 2881 ¬Older, 2882 NULL, 2883 &cld_closure); 2884 } 2885 } 2886 2887 // Clear mod-union table; it will be dirtied in the prologue of 2888 // CMS generation per each young generation collection. 2889 2890 assert(_modUnionTable.isAllClear(), 2891 "Was cleared in most recent final checkpoint phase" 2892 " or no bits are set in the gc_prologue before the start of the next " 2893 "subsequent marking phase."); 2894 2895 assert(_ct->klass_rem_set()->mod_union_is_clear(), "Must be"); 2896 2897 // Save the end of the used_region of the constituent generations 2898 // to be used to limit the extent of sweep in each generation. 2899 save_sweep_limits(); 2900 verify_overflow_empty(); 2901 } 2902 2903 bool CMSCollector::markFromRoots() { 2904 // we might be tempted to assert that: 2905 // assert(!SafepointSynchronize::is_at_safepoint(), 2906 // "inconsistent argument?"); 2907 // However that wouldn't be right, because it's possible that 2908 // a safepoint is indeed in progress as a young generation 2909 // stop-the-world GC happens even as we mark in this generation. 2910 assert(_collectorState == Marking, "inconsistent state?"); 2911 check_correct_thread_executing(); 2912 verify_overflow_empty(); 2913 2914 // Weak ref discovery note: We may be discovering weak 2915 // refs in this generation concurrent (but interleaved) with 2916 // weak ref discovery by the young generation collector. 2917 2918 CMSTokenSyncWithLocks ts(true, bitMapLock()); 2919 GCTraceCPUTime tcpu; 2920 CMSPhaseAccounting pa(this, "Concrurrent Mark"); 2921 bool res = markFromRootsWork(); 2922 if (res) { 2923 _collectorState = Precleaning; 2924 } else { // We failed and a foreground collection wants to take over 2925 assert(_foregroundGCIsActive, "internal state inconsistency"); 2926 assert(_restart_addr == NULL, "foreground will restart from scratch"); 2927 log_debug(gc)("bailing out to foreground collection"); 2928 } 2929 verify_overflow_empty(); 2930 return res; 2931 } 2932 2933 bool CMSCollector::markFromRootsWork() { 2934 // iterate over marked bits in bit map, doing a full scan and mark 2935 // from these roots using the following algorithm: 2936 // . if oop is to the right of the current scan pointer, 2937 // mark corresponding bit (we'll process it later) 2938 // . else (oop is to left of current scan pointer) 2939 // push oop on marking stack 2940 // . drain the marking stack 2941 2942 // Note that when we do a marking step we need to hold the 2943 // bit map lock -- recall that direct allocation (by mutators) 2944 // and promotion (by the young generation collector) is also 2945 // marking the bit map. [the so-called allocate live policy.] 2946 // Because the implementation of bit map marking is not 2947 // robust wrt simultaneous marking of bits in the same word, 2948 // we need to make sure that there is no such interference 2949 // between concurrent such updates. 2950 2951 // already have locks 2952 assert_lock_strong(bitMapLock()); 2953 2954 verify_work_stacks_empty(); 2955 verify_overflow_empty(); 2956 bool result = false; 2957 if (CMSConcurrentMTEnabled && ConcGCThreads > 0) { 2958 result = do_marking_mt(); 2959 } else { 2960 result = do_marking_st(); 2961 } 2962 return result; 2963 } 2964 2965 // Forward decl 2966 class CMSConcMarkingTask; 2967 2968 class CMSConcMarkingTerminator: public ParallelTaskTerminator { 2969 CMSCollector* _collector; 2970 CMSConcMarkingTask* _task; 2971 public: 2972 virtual void yield(); 2973 2974 // "n_threads" is the number of threads to be terminated. 2975 // "queue_set" is a set of work queues of other threads. 2976 // "collector" is the CMS collector associated with this task terminator. 2977 // "yield" indicates whether we need the gang as a whole to yield. 2978 CMSConcMarkingTerminator(int n_threads, TaskQueueSetSuper* queue_set, CMSCollector* collector) : 2979 ParallelTaskTerminator(n_threads, queue_set), 2980 _collector(collector) { } 2981 2982 void set_task(CMSConcMarkingTask* task) { 2983 _task = task; 2984 } 2985 }; 2986 2987 class CMSConcMarkingTerminatorTerminator: public TerminatorTerminator { 2988 CMSConcMarkingTask* _task; 2989 public: 2990 bool should_exit_termination(); 2991 void set_task(CMSConcMarkingTask* task) { 2992 _task = task; 2993 } 2994 }; 2995 2996 // MT Concurrent Marking Task 2997 class CMSConcMarkingTask: public YieldingFlexibleGangTask { 2998 CMSCollector* _collector; 2999 uint _n_workers; // requested/desired # workers 3000 bool _result; 3001 CompactibleFreeListSpace* _cms_space; 3002 char _pad_front[64]; // padding to ... 3003 HeapWord* _global_finger; // ... avoid sharing cache line 3004 char _pad_back[64]; 3005 HeapWord* _restart_addr; 3006 3007 // Exposed here for yielding support 3008 Mutex* const _bit_map_lock; 3009 3010 // The per thread work queues, available here for stealing 3011 OopTaskQueueSet* _task_queues; 3012 3013 // Termination (and yielding) support 3014 CMSConcMarkingTerminator _term; 3015 CMSConcMarkingTerminatorTerminator _term_term; 3016 3017 public: 3018 CMSConcMarkingTask(CMSCollector* collector, 3019 CompactibleFreeListSpace* cms_space, 3020 YieldingFlexibleWorkGang* workers, 3021 OopTaskQueueSet* task_queues): 3022 YieldingFlexibleGangTask("Concurrent marking done multi-threaded"), 3023 _collector(collector), 3024 _cms_space(cms_space), 3025 _n_workers(0), _result(true), 3026 _task_queues(task_queues), 3027 _term(_n_workers, task_queues, _collector), 3028 _bit_map_lock(collector->bitMapLock()) 3029 { 3030 _requested_size = _n_workers; 3031 _term.set_task(this); 3032 _term_term.set_task(this); 3033 _restart_addr = _global_finger = _cms_space->bottom(); 3034 } 3035 3036 3037 OopTaskQueueSet* task_queues() { return _task_queues; } 3038 3039 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 3040 3041 HeapWord** global_finger_addr() { return &_global_finger; } 3042 3043 CMSConcMarkingTerminator* terminator() { return &_term; } 3044 3045 virtual void set_for_termination(uint active_workers) { 3046 terminator()->reset_for_reuse(active_workers); 3047 } 3048 3049 void work(uint worker_id); 3050 bool should_yield() { 3051 return ConcurrentMarkSweepThread::should_yield() 3052 && !_collector->foregroundGCIsActive(); 3053 } 3054 3055 virtual void coordinator_yield(); // stuff done by coordinator 3056 bool result() { return _result; } 3057 3058 void reset(HeapWord* ra) { 3059 assert(_global_finger >= _cms_space->end(), "Postcondition of ::work(i)"); 3060 _restart_addr = _global_finger = ra; 3061 _term.reset_for_reuse(); 3062 } 3063 3064 static bool get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 3065 OopTaskQueue* work_q); 3066 3067 private: 3068 void do_scan_and_mark(int i, CompactibleFreeListSpace* sp); 3069 void do_work_steal(int i); 3070 void bump_global_finger(HeapWord* f); 3071 }; 3072 3073 bool CMSConcMarkingTerminatorTerminator::should_exit_termination() { 3074 assert(_task != NULL, "Error"); 3075 return _task->yielding(); 3076 // Note that we do not need the disjunct || _task->should_yield() above 3077 // because we want terminating threads to yield only if the task 3078 // is already in the midst of yielding, which happens only after at least one 3079 // thread has yielded. 3080 } 3081 3082 void CMSConcMarkingTerminator::yield() { 3083 if (_task->should_yield()) { 3084 _task->yield(); 3085 } else { 3086 ParallelTaskTerminator::yield(); 3087 } 3088 } 3089 3090 //////////////////////////////////////////////////////////////// 3091 // Concurrent Marking Algorithm Sketch 3092 //////////////////////////////////////////////////////////////// 3093 // Until all tasks exhausted (both spaces): 3094 // -- claim next available chunk 3095 // -- bump global finger via CAS 3096 // -- find first object that starts in this chunk 3097 // and start scanning bitmap from that position 3098 // -- scan marked objects for oops 3099 // -- CAS-mark target, and if successful: 3100 // . if target oop is above global finger (volatile read) 3101 // nothing to do 3102 // . if target oop is in chunk and above local finger 3103 // then nothing to do 3104 // . else push on work-queue 3105 // -- Deal with possible overflow issues: 3106 // . local work-queue overflow causes stuff to be pushed on 3107 // global (common) overflow queue 3108 // . always first empty local work queue 3109 // . then get a batch of oops from global work queue if any 3110 // . then do work stealing 3111 // -- When all tasks claimed (both spaces) 3112 // and local work queue empty, 3113 // then in a loop do: 3114 // . check global overflow stack; steal a batch of oops and trace 3115 // . try to steal from other threads oif GOS is empty 3116 // . if neither is available, offer termination 3117 // -- Terminate and return result 3118 // 3119 void CMSConcMarkingTask::work(uint worker_id) { 3120 elapsedTimer _timer; 3121 ResourceMark rm; 3122 HandleMark hm; 3123 3124 DEBUG_ONLY(_collector->verify_overflow_empty();) 3125 3126 // Before we begin work, our work queue should be empty 3127 assert(work_queue(worker_id)->size() == 0, "Expected to be empty"); 3128 // Scan the bitmap covering _cms_space, tracing through grey objects. 3129 _timer.start(); 3130 do_scan_and_mark(worker_id, _cms_space); 3131 _timer.stop(); 3132 log_trace(gc, task)("Finished cms space scanning in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 3133 3134 // ... do work stealing 3135 _timer.reset(); 3136 _timer.start(); 3137 do_work_steal(worker_id); 3138 _timer.stop(); 3139 log_trace(gc, task)("Finished work stealing in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 3140 assert(_collector->_markStack.isEmpty(), "Should have been emptied"); 3141 assert(work_queue(worker_id)->size() == 0, "Should have been emptied"); 3142 // Note that under the current task protocol, the 3143 // following assertion is true even of the spaces 3144 // expanded since the completion of the concurrent 3145 // marking. XXX This will likely change under a strict 3146 // ABORT semantics. 3147 // After perm removal the comparison was changed to 3148 // greater than or equal to from strictly greater than. 3149 // Before perm removal the highest address sweep would 3150 // have been at the end of perm gen but now is at the 3151 // end of the tenured gen. 3152 assert(_global_finger >= _cms_space->end(), 3153 "All tasks have been completed"); 3154 DEBUG_ONLY(_collector->verify_overflow_empty();) 3155 } 3156 3157 void CMSConcMarkingTask::bump_global_finger(HeapWord* f) { 3158 HeapWord* read = _global_finger; 3159 HeapWord* cur = read; 3160 while (f > read) { 3161 cur = read; 3162 read = (HeapWord*) Atomic::cmpxchg_ptr(f, &_global_finger, cur); 3163 if (cur == read) { 3164 // our cas succeeded 3165 assert(_global_finger >= f, "protocol consistency"); 3166 break; 3167 } 3168 } 3169 } 3170 3171 // This is really inefficient, and should be redone by 3172 // using (not yet available) block-read and -write interfaces to the 3173 // stack and the work_queue. XXX FIX ME !!! 3174 bool CMSConcMarkingTask::get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 3175 OopTaskQueue* work_q) { 3176 // Fast lock-free check 3177 if (ovflw_stk->length() == 0) { 3178 return false; 3179 } 3180 assert(work_q->size() == 0, "Shouldn't steal"); 3181 MutexLockerEx ml(ovflw_stk->par_lock(), 3182 Mutex::_no_safepoint_check_flag); 3183 // Grab up to 1/4 the size of the work queue 3184 size_t num = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 3185 (size_t)ParGCDesiredObjsFromOverflowList); 3186 num = MIN2(num, ovflw_stk->length()); 3187 for (int i = (int) num; i > 0; i--) { 3188 oop cur = ovflw_stk->pop(); 3189 assert(cur != NULL, "Counted wrong?"); 3190 work_q->push(cur); 3191 } 3192 return num > 0; 3193 } 3194 3195 void CMSConcMarkingTask::do_scan_and_mark(int i, CompactibleFreeListSpace* sp) { 3196 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 3197 int n_tasks = pst->n_tasks(); 3198 // We allow that there may be no tasks to do here because 3199 // we are restarting after a stack overflow. 3200 assert(pst->valid() || n_tasks == 0, "Uninitialized use?"); 3201 uint nth_task = 0; 3202 3203 HeapWord* aligned_start = sp->bottom(); 3204 if (sp->used_region().contains(_restart_addr)) { 3205 // Align down to a card boundary for the start of 0th task 3206 // for this space. 3207 aligned_start = 3208 (HeapWord*)align_size_down((uintptr_t)_restart_addr, 3209 CardTableModRefBS::card_size); 3210 } 3211 3212 size_t chunk_size = sp->marking_task_size(); 3213 while (!pst->is_task_claimed(/* reference */ nth_task)) { 3214 // Having claimed the nth task in this space, 3215 // compute the chunk that it corresponds to: 3216 MemRegion span = MemRegion(aligned_start + nth_task*chunk_size, 3217 aligned_start + (nth_task+1)*chunk_size); 3218 // Try and bump the global finger via a CAS; 3219 // note that we need to do the global finger bump 3220 // _before_ taking the intersection below, because 3221 // the task corresponding to that region will be 3222 // deemed done even if the used_region() expands 3223 // because of allocation -- as it almost certainly will 3224 // during start-up while the threads yield in the 3225 // closure below. 3226 HeapWord* finger = span.end(); 3227 bump_global_finger(finger); // atomically 3228 // There are null tasks here corresponding to chunks 3229 // beyond the "top" address of the space. 3230 span = span.intersection(sp->used_region()); 3231 if (!span.is_empty()) { // Non-null task 3232 HeapWord* prev_obj; 3233 assert(!span.contains(_restart_addr) || nth_task == 0, 3234 "Inconsistency"); 3235 if (nth_task == 0) { 3236 // For the 0th task, we'll not need to compute a block_start. 3237 if (span.contains(_restart_addr)) { 3238 // In the case of a restart because of stack overflow, 3239 // we might additionally skip a chunk prefix. 3240 prev_obj = _restart_addr; 3241 } else { 3242 prev_obj = span.start(); 3243 } 3244 } else { 3245 // We want to skip the first object because 3246 // the protocol is to scan any object in its entirety 3247 // that _starts_ in this span; a fortiori, any 3248 // object starting in an earlier span is scanned 3249 // as part of an earlier claimed task. 3250 // Below we use the "careful" version of block_start 3251 // so we do not try to navigate uninitialized objects. 3252 prev_obj = sp->block_start_careful(span.start()); 3253 // Below we use a variant of block_size that uses the 3254 // Printezis bits to avoid waiting for allocated 3255 // objects to become initialized/parsable. 3256 while (prev_obj < span.start()) { 3257 size_t sz = sp->block_size_no_stall(prev_obj, _collector); 3258 if (sz > 0) { 3259 prev_obj += sz; 3260 } else { 3261 // In this case we may end up doing a bit of redundant 3262 // scanning, but that appears unavoidable, short of 3263 // locking the free list locks; see bug 6324141. 3264 break; 3265 } 3266 } 3267 } 3268 if (prev_obj < span.end()) { 3269 MemRegion my_span = MemRegion(prev_obj, span.end()); 3270 // Do the marking work within a non-empty span -- 3271 // the last argument to the constructor indicates whether the 3272 // iteration should be incremental with periodic yields. 3273 Par_MarkFromRootsClosure cl(this, _collector, my_span, 3274 &_collector->_markBitMap, 3275 work_queue(i), 3276 &_collector->_markStack); 3277 _collector->_markBitMap.iterate(&cl, my_span.start(), my_span.end()); 3278 } // else nothing to do for this task 3279 } // else nothing to do for this task 3280 } 3281 // We'd be tempted to assert here that since there are no 3282 // more tasks left to claim in this space, the global_finger 3283 // must exceed space->top() and a fortiori space->end(). However, 3284 // that would not quite be correct because the bumping of 3285 // global_finger occurs strictly after the claiming of a task, 3286 // so by the time we reach here the global finger may not yet 3287 // have been bumped up by the thread that claimed the last 3288 // task. 3289 pst->all_tasks_completed(); 3290 } 3291 3292 class Par_ConcMarkingClosure: public MetadataAwareOopClosure { 3293 private: 3294 CMSCollector* _collector; 3295 CMSConcMarkingTask* _task; 3296 MemRegion _span; 3297 CMSBitMap* _bit_map; 3298 CMSMarkStack* _overflow_stack; 3299 OopTaskQueue* _work_queue; 3300 protected: 3301 DO_OOP_WORK_DEFN 3302 public: 3303 Par_ConcMarkingClosure(CMSCollector* collector, CMSConcMarkingTask* task, OopTaskQueue* work_queue, 3304 CMSBitMap* bit_map, CMSMarkStack* overflow_stack): 3305 MetadataAwareOopClosure(collector->ref_processor()), 3306 _collector(collector), 3307 _task(task), 3308 _span(collector->_span), 3309 _work_queue(work_queue), 3310 _bit_map(bit_map), 3311 _overflow_stack(overflow_stack) 3312 { } 3313 virtual void do_oop(oop* p); 3314 virtual void do_oop(narrowOop* p); 3315 3316 void trim_queue(size_t max); 3317 void handle_stack_overflow(HeapWord* lost); 3318 void do_yield_check() { 3319 if (_task->should_yield()) { 3320 _task->yield(); 3321 } 3322 } 3323 }; 3324 3325 // Grey object scanning during work stealing phase -- 3326 // the salient assumption here is that any references 3327 // that are in these stolen objects being scanned must 3328 // already have been initialized (else they would not have 3329 // been published), so we do not need to check for 3330 // uninitialized objects before pushing here. 3331 void Par_ConcMarkingClosure::do_oop(oop obj) { 3332 assert(obj->is_oop_or_null(true), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 3333 HeapWord* addr = (HeapWord*)obj; 3334 // Check if oop points into the CMS generation 3335 // and is not marked 3336 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 3337 // a white object ... 3338 // If we manage to "claim" the object, by being the 3339 // first thread to mark it, then we push it on our 3340 // marking stack 3341 if (_bit_map->par_mark(addr)) { // ... now grey 3342 // push on work queue (grey set) 3343 bool simulate_overflow = false; 3344 NOT_PRODUCT( 3345 if (CMSMarkStackOverflowALot && 3346 _collector->simulate_overflow()) { 3347 // simulate a stack overflow 3348 simulate_overflow = true; 3349 } 3350 ) 3351 if (simulate_overflow || 3352 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) { 3353 // stack overflow 3354 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _overflow_stack->capacity()); 3355 // We cannot assert that the overflow stack is full because 3356 // it may have been emptied since. 3357 assert(simulate_overflow || 3358 _work_queue->size() == _work_queue->max_elems(), 3359 "Else push should have succeeded"); 3360 handle_stack_overflow(addr); 3361 } 3362 } // Else, some other thread got there first 3363 do_yield_check(); 3364 } 3365 } 3366 3367 void Par_ConcMarkingClosure::do_oop(oop* p) { Par_ConcMarkingClosure::do_oop_work(p); } 3368 void Par_ConcMarkingClosure::do_oop(narrowOop* p) { Par_ConcMarkingClosure::do_oop_work(p); } 3369 3370 void Par_ConcMarkingClosure::trim_queue(size_t max) { 3371 while (_work_queue->size() > max) { 3372 oop new_oop; 3373 if (_work_queue->pop_local(new_oop)) { 3374 assert(new_oop->is_oop(), "Should be an oop"); 3375 assert(_bit_map->isMarked((HeapWord*)new_oop), "Grey object"); 3376 assert(_span.contains((HeapWord*)new_oop), "Not in span"); 3377 new_oop->oop_iterate(this); // do_oop() above 3378 do_yield_check(); 3379 } 3380 } 3381 } 3382 3383 // Upon stack overflow, we discard (part of) the stack, 3384 // remembering the least address amongst those discarded 3385 // in CMSCollector's _restart_address. 3386 void Par_ConcMarkingClosure::handle_stack_overflow(HeapWord* lost) { 3387 // We need to do this under a mutex to prevent other 3388 // workers from interfering with the work done below. 3389 MutexLockerEx ml(_overflow_stack->par_lock(), 3390 Mutex::_no_safepoint_check_flag); 3391 // Remember the least grey address discarded 3392 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 3393 _collector->lower_restart_addr(ra); 3394 _overflow_stack->reset(); // discard stack contents 3395 _overflow_stack->expand(); // expand the stack if possible 3396 } 3397 3398 3399 void CMSConcMarkingTask::do_work_steal(int i) { 3400 OopTaskQueue* work_q = work_queue(i); 3401 oop obj_to_scan; 3402 CMSBitMap* bm = &(_collector->_markBitMap); 3403 CMSMarkStack* ovflw = &(_collector->_markStack); 3404 int* seed = _collector->hash_seed(i); 3405 Par_ConcMarkingClosure cl(_collector, this, work_q, bm, ovflw); 3406 while (true) { 3407 cl.trim_queue(0); 3408 assert(work_q->size() == 0, "Should have been emptied above"); 3409 if (get_work_from_overflow_stack(ovflw, work_q)) { 3410 // Can't assert below because the work obtained from the 3411 // overflow stack may already have been stolen from us. 3412 // assert(work_q->size() > 0, "Work from overflow stack"); 3413 continue; 3414 } else if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 3415 assert(obj_to_scan->is_oop(), "Should be an oop"); 3416 assert(bm->isMarked((HeapWord*)obj_to_scan), "Grey object"); 3417 obj_to_scan->oop_iterate(&cl); 3418 } else if (terminator()->offer_termination(&_term_term)) { 3419 assert(work_q->size() == 0, "Impossible!"); 3420 break; 3421 } else if (yielding() || should_yield()) { 3422 yield(); 3423 } 3424 } 3425 } 3426 3427 // This is run by the CMS (coordinator) thread. 3428 void CMSConcMarkingTask::coordinator_yield() { 3429 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 3430 "CMS thread should hold CMS token"); 3431 // First give up the locks, then yield, then re-lock 3432 // We should probably use a constructor/destructor idiom to 3433 // do this unlock/lock or modify the MutexUnlocker class to 3434 // serve our purpose. XXX 3435 assert_lock_strong(_bit_map_lock); 3436 _bit_map_lock->unlock(); 3437 ConcurrentMarkSweepThread::desynchronize(true); 3438 _collector->stopTimer(); 3439 _collector->incrementYields(); 3440 3441 // It is possible for whichever thread initiated the yield request 3442 // not to get a chance to wake up and take the bitmap lock between 3443 // this thread releasing it and reacquiring it. So, while the 3444 // should_yield() flag is on, let's sleep for a bit to give the 3445 // other thread a chance to wake up. The limit imposed on the number 3446 // of iterations is defensive, to avoid any unforseen circumstances 3447 // putting us into an infinite loop. Since it's always been this 3448 // (coordinator_yield()) method that was observed to cause the 3449 // problem, we are using a parameter (CMSCoordinatorYieldSleepCount) 3450 // which is by default non-zero. For the other seven methods that 3451 // also perform the yield operation, as are using a different 3452 // parameter (CMSYieldSleepCount) which is by default zero. This way we 3453 // can enable the sleeping for those methods too, if necessary. 3454 // See 6442774. 3455 // 3456 // We really need to reconsider the synchronization between the GC 3457 // thread and the yield-requesting threads in the future and we 3458 // should really use wait/notify, which is the recommended 3459 // way of doing this type of interaction. Additionally, we should 3460 // consolidate the eight methods that do the yield operation and they 3461 // are almost identical into one for better maintainability and 3462 // readability. See 6445193. 3463 // 3464 // Tony 2006.06.29 3465 for (unsigned i = 0; i < CMSCoordinatorYieldSleepCount && 3466 ConcurrentMarkSweepThread::should_yield() && 3467 !CMSCollector::foregroundGCIsActive(); ++i) { 3468 os::sleep(Thread::current(), 1, false); 3469 } 3470 3471 ConcurrentMarkSweepThread::synchronize(true); 3472 _bit_map_lock->lock_without_safepoint_check(); 3473 _collector->startTimer(); 3474 } 3475 3476 bool CMSCollector::do_marking_mt() { 3477 assert(ConcGCThreads > 0 && conc_workers() != NULL, "precondition"); 3478 uint num_workers = AdaptiveSizePolicy::calc_active_conc_workers(conc_workers()->total_workers(), 3479 conc_workers()->active_workers(), 3480 Threads::number_of_non_daemon_threads()); 3481 conc_workers()->set_active_workers(num_workers); 3482 3483 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 3484 3485 CMSConcMarkingTask tsk(this, 3486 cms_space, 3487 conc_workers(), 3488 task_queues()); 3489 3490 // Since the actual number of workers we get may be different 3491 // from the number we requested above, do we need to do anything different 3492 // below? In particular, may be we need to subclass the SequantialSubTasksDone 3493 // class?? XXX 3494 cms_space ->initialize_sequential_subtasks_for_marking(num_workers); 3495 3496 // Refs discovery is already non-atomic. 3497 assert(!ref_processor()->discovery_is_atomic(), "Should be non-atomic"); 3498 assert(ref_processor()->discovery_is_mt(), "Discovery should be MT"); 3499 conc_workers()->start_task(&tsk); 3500 while (tsk.yielded()) { 3501 tsk.coordinator_yield(); 3502 conc_workers()->continue_task(&tsk); 3503 } 3504 // If the task was aborted, _restart_addr will be non-NULL 3505 assert(tsk.completed() || _restart_addr != NULL, "Inconsistency"); 3506 while (_restart_addr != NULL) { 3507 // XXX For now we do not make use of ABORTED state and have not 3508 // yet implemented the right abort semantics (even in the original 3509 // single-threaded CMS case). That needs some more investigation 3510 // and is deferred for now; see CR# TBF. 07252005YSR. XXX 3511 assert(!CMSAbortSemantics || tsk.aborted(), "Inconsistency"); 3512 // If _restart_addr is non-NULL, a marking stack overflow 3513 // occurred; we need to do a fresh marking iteration from the 3514 // indicated restart address. 3515 if (_foregroundGCIsActive) { 3516 // We may be running into repeated stack overflows, having 3517 // reached the limit of the stack size, while making very 3518 // slow forward progress. It may be best to bail out and 3519 // let the foreground collector do its job. 3520 // Clear _restart_addr, so that foreground GC 3521 // works from scratch. This avoids the headache of 3522 // a "rescan" which would otherwise be needed because 3523 // of the dirty mod union table & card table. 3524 _restart_addr = NULL; 3525 return false; 3526 } 3527 // Adjust the task to restart from _restart_addr 3528 tsk.reset(_restart_addr); 3529 cms_space ->initialize_sequential_subtasks_for_marking(num_workers, 3530 _restart_addr); 3531 _restart_addr = NULL; 3532 // Get the workers going again 3533 conc_workers()->start_task(&tsk); 3534 while (tsk.yielded()) { 3535 tsk.coordinator_yield(); 3536 conc_workers()->continue_task(&tsk); 3537 } 3538 } 3539 assert(tsk.completed(), "Inconsistency"); 3540 assert(tsk.result() == true, "Inconsistency"); 3541 return true; 3542 } 3543 3544 bool CMSCollector::do_marking_st() { 3545 ResourceMark rm; 3546 HandleMark hm; 3547 3548 // Temporarily make refs discovery single threaded (non-MT) 3549 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false); 3550 MarkFromRootsClosure markFromRootsClosure(this, _span, &_markBitMap, 3551 &_markStack, CMSYield); 3552 // the last argument to iterate indicates whether the iteration 3553 // should be incremental with periodic yields. 3554 _markBitMap.iterate(&markFromRootsClosure); 3555 // If _restart_addr is non-NULL, a marking stack overflow 3556 // occurred; we need to do a fresh iteration from the 3557 // indicated restart address. 3558 while (_restart_addr != NULL) { 3559 if (_foregroundGCIsActive) { 3560 // We may be running into repeated stack overflows, having 3561 // reached the limit of the stack size, while making very 3562 // slow forward progress. It may be best to bail out and 3563 // let the foreground collector do its job. 3564 // Clear _restart_addr, so that foreground GC 3565 // works from scratch. This avoids the headache of 3566 // a "rescan" which would otherwise be needed because 3567 // of the dirty mod union table & card table. 3568 _restart_addr = NULL; 3569 return false; // indicating failure to complete marking 3570 } 3571 // Deal with stack overflow: 3572 // we restart marking from _restart_addr 3573 HeapWord* ra = _restart_addr; 3574 markFromRootsClosure.reset(ra); 3575 _restart_addr = NULL; 3576 _markBitMap.iterate(&markFromRootsClosure, ra, _span.end()); 3577 } 3578 return true; 3579 } 3580 3581 void CMSCollector::preclean() { 3582 check_correct_thread_executing(); 3583 assert(Thread::current()->is_ConcurrentGC_thread(), "Wrong thread"); 3584 verify_work_stacks_empty(); 3585 verify_overflow_empty(); 3586 _abort_preclean = false; 3587 if (CMSPrecleaningEnabled) { 3588 if (!CMSEdenChunksRecordAlways) { 3589 _eden_chunk_index = 0; 3590 } 3591 size_t used = get_eden_used(); 3592 size_t capacity = get_eden_capacity(); 3593 // Don't start sampling unless we will get sufficiently 3594 // many samples. 3595 if (used < (capacity/(CMSScheduleRemarkSamplingRatio * 100) 3596 * CMSScheduleRemarkEdenPenetration)) { 3597 _start_sampling = true; 3598 } else { 3599 _start_sampling = false; 3600 } 3601 GCTraceCPUTime tcpu; 3602 CMSPhaseAccounting pa(this, "Concurrent Preclean"); 3603 preclean_work(CMSPrecleanRefLists1, CMSPrecleanSurvivors1); 3604 } 3605 CMSTokenSync x(true); // is cms thread 3606 if (CMSPrecleaningEnabled) { 3607 sample_eden(); 3608 _collectorState = AbortablePreclean; 3609 } else { 3610 _collectorState = FinalMarking; 3611 } 3612 verify_work_stacks_empty(); 3613 verify_overflow_empty(); 3614 } 3615 3616 // Try and schedule the remark such that young gen 3617 // occupancy is CMSScheduleRemarkEdenPenetration %. 3618 void CMSCollector::abortable_preclean() { 3619 check_correct_thread_executing(); 3620 assert(CMSPrecleaningEnabled, "Inconsistent control state"); 3621 assert(_collectorState == AbortablePreclean, "Inconsistent control state"); 3622 3623 // If Eden's current occupancy is below this threshold, 3624 // immediately schedule the remark; else preclean 3625 // past the next scavenge in an effort to 3626 // schedule the pause as described above. By choosing 3627 // CMSScheduleRemarkEdenSizeThreshold >= max eden size 3628 // we will never do an actual abortable preclean cycle. 3629 if (get_eden_used() > CMSScheduleRemarkEdenSizeThreshold) { 3630 GCTraceCPUTime tcpu; 3631 CMSPhaseAccounting pa(this, "Concurrent Abortable Preclean"); 3632 // We need more smarts in the abortable preclean 3633 // loop below to deal with cases where allocation 3634 // in young gen is very very slow, and our precleaning 3635 // is running a losing race against a horde of 3636 // mutators intent on flooding us with CMS updates 3637 // (dirty cards). 3638 // One, admittedly dumb, strategy is to give up 3639 // after a certain number of abortable precleaning loops 3640 // or after a certain maximum time. We want to make 3641 // this smarter in the next iteration. 3642 // XXX FIX ME!!! YSR 3643 size_t loops = 0, workdone = 0, cumworkdone = 0, waited = 0; 3644 while (!(should_abort_preclean() || 3645 ConcurrentMarkSweepThread::should_terminate())) { 3646 workdone = preclean_work(CMSPrecleanRefLists2, CMSPrecleanSurvivors2); 3647 cumworkdone += workdone; 3648 loops++; 3649 // Voluntarily terminate abortable preclean phase if we have 3650 // been at it for too long. 3651 if ((CMSMaxAbortablePrecleanLoops != 0) && 3652 loops >= CMSMaxAbortablePrecleanLoops) { 3653 log_debug(gc)(" CMS: abort preclean due to loops "); 3654 break; 3655 } 3656 if (pa.wallclock_millis() > CMSMaxAbortablePrecleanTime) { 3657 log_debug(gc)(" CMS: abort preclean due to time "); 3658 break; 3659 } 3660 // If we are doing little work each iteration, we should 3661 // take a short break. 3662 if (workdone < CMSAbortablePrecleanMinWorkPerIteration) { 3663 // Sleep for some time, waiting for work to accumulate 3664 stopTimer(); 3665 cmsThread()->wait_on_cms_lock(CMSAbortablePrecleanWaitMillis); 3666 startTimer(); 3667 waited++; 3668 } 3669 } 3670 log_trace(gc)(" [" SIZE_FORMAT " iterations, " SIZE_FORMAT " waits, " SIZE_FORMAT " cards)] ", 3671 loops, waited, cumworkdone); 3672 } 3673 CMSTokenSync x(true); // is cms thread 3674 if (_collectorState != Idling) { 3675 assert(_collectorState == AbortablePreclean, 3676 "Spontaneous state transition?"); 3677 _collectorState = FinalMarking; 3678 } // Else, a foreground collection completed this CMS cycle. 3679 return; 3680 } 3681 3682 // Respond to an Eden sampling opportunity 3683 void CMSCollector::sample_eden() { 3684 // Make sure a young gc cannot sneak in between our 3685 // reading and recording of a sample. 3686 assert(Thread::current()->is_ConcurrentGC_thread(), 3687 "Only the cms thread may collect Eden samples"); 3688 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 3689 "Should collect samples while holding CMS token"); 3690 if (!_start_sampling) { 3691 return; 3692 } 3693 // When CMSEdenChunksRecordAlways is true, the eden chunk array 3694 // is populated by the young generation. 3695 if (_eden_chunk_array != NULL && !CMSEdenChunksRecordAlways) { 3696 if (_eden_chunk_index < _eden_chunk_capacity) { 3697 _eden_chunk_array[_eden_chunk_index] = *_top_addr; // take sample 3698 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr, 3699 "Unexpected state of Eden"); 3700 // We'd like to check that what we just sampled is an oop-start address; 3701 // however, we cannot do that here since the object may not yet have been 3702 // initialized. So we'll instead do the check when we _use_ this sample 3703 // later. 3704 if (_eden_chunk_index == 0 || 3705 (pointer_delta(_eden_chunk_array[_eden_chunk_index], 3706 _eden_chunk_array[_eden_chunk_index-1]) 3707 >= CMSSamplingGrain)) { 3708 _eden_chunk_index++; // commit sample 3709 } 3710 } 3711 } 3712 if ((_collectorState == AbortablePreclean) && !_abort_preclean) { 3713 size_t used = get_eden_used(); 3714 size_t capacity = get_eden_capacity(); 3715 assert(used <= capacity, "Unexpected state of Eden"); 3716 if (used > (capacity/100 * CMSScheduleRemarkEdenPenetration)) { 3717 _abort_preclean = true; 3718 } 3719 } 3720 } 3721 3722 3723 size_t CMSCollector::preclean_work(bool clean_refs, bool clean_survivor) { 3724 assert(_collectorState == Precleaning || 3725 _collectorState == AbortablePreclean, "incorrect state"); 3726 ResourceMark rm; 3727 HandleMark hm; 3728 3729 // Precleaning is currently not MT but the reference processor 3730 // may be set for MT. Disable it temporarily here. 3731 ReferenceProcessor* rp = ref_processor(); 3732 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(rp, false); 3733 3734 // Do one pass of scrubbing the discovered reference lists 3735 // to remove any reference objects with strongly-reachable 3736 // referents. 3737 if (clean_refs) { 3738 CMSPrecleanRefsYieldClosure yield_cl(this); 3739 assert(rp->span().equals(_span), "Spans should be equal"); 3740 CMSKeepAliveClosure keep_alive(this, _span, &_markBitMap, 3741 &_markStack, true /* preclean */); 3742 CMSDrainMarkingStackClosure complete_trace(this, 3743 _span, &_markBitMap, &_markStack, 3744 &keep_alive, true /* preclean */); 3745 3746 // We don't want this step to interfere with a young 3747 // collection because we don't want to take CPU 3748 // or memory bandwidth away from the young GC threads 3749 // (which may be as many as there are CPUs). 3750 // Note that we don't need to protect ourselves from 3751 // interference with mutators because they can't 3752 // manipulate the discovered reference lists nor affect 3753 // the computed reachability of the referents, the 3754 // only properties manipulated by the precleaning 3755 // of these reference lists. 3756 stopTimer(); 3757 CMSTokenSyncWithLocks x(true /* is cms thread */, 3758 bitMapLock()); 3759 startTimer(); 3760 sample_eden(); 3761 3762 // The following will yield to allow foreground 3763 // collection to proceed promptly. XXX YSR: 3764 // The code in this method may need further 3765 // tweaking for better performance and some restructuring 3766 // for cleaner interfaces. 3767 GCTimer *gc_timer = NULL; // Currently not tracing concurrent phases 3768 rp->preclean_discovered_references( 3769 rp->is_alive_non_header(), &keep_alive, &complete_trace, &yield_cl, 3770 gc_timer); 3771 } 3772 3773 if (clean_survivor) { // preclean the active survivor space(s) 3774 PushAndMarkClosure pam_cl(this, _span, ref_processor(), 3775 &_markBitMap, &_modUnionTable, 3776 &_markStack, true /* precleaning phase */); 3777 stopTimer(); 3778 CMSTokenSyncWithLocks ts(true /* is cms thread */, 3779 bitMapLock()); 3780 startTimer(); 3781 unsigned int before_count = 3782 GenCollectedHeap::heap()->total_collections(); 3783 SurvivorSpacePrecleanClosure 3784 sss_cl(this, _span, &_markBitMap, &_markStack, 3785 &pam_cl, before_count, CMSYield); 3786 _young_gen->from()->object_iterate_careful(&sss_cl); 3787 _young_gen->to()->object_iterate_careful(&sss_cl); 3788 } 3789 MarkRefsIntoAndScanClosure 3790 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable, 3791 &_markStack, this, CMSYield, 3792 true /* precleaning phase */); 3793 // CAUTION: The following closure has persistent state that may need to 3794 // be reset upon a decrease in the sequence of addresses it 3795 // processes. 3796 ScanMarkedObjectsAgainCarefullyClosure 3797 smoac_cl(this, _span, 3798 &_markBitMap, &_markStack, &mrias_cl, CMSYield); 3799 3800 // Preclean dirty cards in ModUnionTable and CardTable using 3801 // appropriate convergence criterion; 3802 // repeat CMSPrecleanIter times unless we find that 3803 // we are losing. 3804 assert(CMSPrecleanIter < 10, "CMSPrecleanIter is too large"); 3805 assert(CMSPrecleanNumerator < CMSPrecleanDenominator, 3806 "Bad convergence multiplier"); 3807 assert(CMSPrecleanThreshold >= 100, 3808 "Unreasonably low CMSPrecleanThreshold"); 3809 3810 size_t numIter, cumNumCards, lastNumCards, curNumCards; 3811 for (numIter = 0, cumNumCards = lastNumCards = curNumCards = 0; 3812 numIter < CMSPrecleanIter; 3813 numIter++, lastNumCards = curNumCards, cumNumCards += curNumCards) { 3814 curNumCards = preclean_mod_union_table(_cmsGen, &smoac_cl); 3815 log_trace(gc)(" (modUnionTable: " SIZE_FORMAT " cards)", curNumCards); 3816 // Either there are very few dirty cards, so re-mark 3817 // pause will be small anyway, or our pre-cleaning isn't 3818 // that much faster than the rate at which cards are being 3819 // dirtied, so we might as well stop and re-mark since 3820 // precleaning won't improve our re-mark time by much. 3821 if (curNumCards <= CMSPrecleanThreshold || 3822 (numIter > 0 && 3823 (curNumCards * CMSPrecleanDenominator > 3824 lastNumCards * CMSPrecleanNumerator))) { 3825 numIter++; 3826 cumNumCards += curNumCards; 3827 break; 3828 } 3829 } 3830 3831 preclean_klasses(&mrias_cl, _cmsGen->freelistLock()); 3832 3833 curNumCards = preclean_card_table(_cmsGen, &smoac_cl); 3834 cumNumCards += curNumCards; 3835 log_trace(gc)(" (cardTable: " SIZE_FORMAT " cards, re-scanned " SIZE_FORMAT " cards, " SIZE_FORMAT " iterations)", 3836 curNumCards, cumNumCards, numIter); 3837 return cumNumCards; // as a measure of useful work done 3838 } 3839 3840 // PRECLEANING NOTES: 3841 // Precleaning involves: 3842 // . reading the bits of the modUnionTable and clearing the set bits. 3843 // . For the cards corresponding to the set bits, we scan the 3844 // objects on those cards. This means we need the free_list_lock 3845 // so that we can safely iterate over the CMS space when scanning 3846 // for oops. 3847 // . When we scan the objects, we'll be both reading and setting 3848 // marks in the marking bit map, so we'll need the marking bit map. 3849 // . For protecting _collector_state transitions, we take the CGC_lock. 3850 // Note that any races in the reading of of card table entries by the 3851 // CMS thread on the one hand and the clearing of those entries by the 3852 // VM thread or the setting of those entries by the mutator threads on the 3853 // other are quite benign. However, for efficiency it makes sense to keep 3854 // the VM thread from racing with the CMS thread while the latter is 3855 // dirty card info to the modUnionTable. We therefore also use the 3856 // CGC_lock to protect the reading of the card table and the mod union 3857 // table by the CM thread. 3858 // . We run concurrently with mutator updates, so scanning 3859 // needs to be done carefully -- we should not try to scan 3860 // potentially uninitialized objects. 3861 // 3862 // Locking strategy: While holding the CGC_lock, we scan over and 3863 // reset a maximal dirty range of the mod union / card tables, then lock 3864 // the free_list_lock and bitmap lock to do a full marking, then 3865 // release these locks; and repeat the cycle. This allows for a 3866 // certain amount of fairness in the sharing of these locks between 3867 // the CMS collector on the one hand, and the VM thread and the 3868 // mutators on the other. 3869 3870 // NOTE: preclean_mod_union_table() and preclean_card_table() 3871 // further below are largely identical; if you need to modify 3872 // one of these methods, please check the other method too. 3873 3874 size_t CMSCollector::preclean_mod_union_table( 3875 ConcurrentMarkSweepGeneration* old_gen, 3876 ScanMarkedObjectsAgainCarefullyClosure* cl) { 3877 verify_work_stacks_empty(); 3878 verify_overflow_empty(); 3879 3880 // strategy: starting with the first card, accumulate contiguous 3881 // ranges of dirty cards; clear these cards, then scan the region 3882 // covered by these cards. 3883 3884 // Since all of the MUT is committed ahead, we can just use 3885 // that, in case the generations expand while we are precleaning. 3886 // It might also be fine to just use the committed part of the 3887 // generation, but we might potentially miss cards when the 3888 // generation is rapidly expanding while we are in the midst 3889 // of precleaning. 3890 HeapWord* startAddr = old_gen->reserved().start(); 3891 HeapWord* endAddr = old_gen->reserved().end(); 3892 3893 cl->setFreelistLock(old_gen->freelistLock()); // needed for yielding 3894 3895 size_t numDirtyCards, cumNumDirtyCards; 3896 HeapWord *nextAddr, *lastAddr; 3897 for (cumNumDirtyCards = numDirtyCards = 0, 3898 nextAddr = lastAddr = startAddr; 3899 nextAddr < endAddr; 3900 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 3901 3902 ResourceMark rm; 3903 HandleMark hm; 3904 3905 MemRegion dirtyRegion; 3906 { 3907 stopTimer(); 3908 // Potential yield point 3909 CMSTokenSync ts(true); 3910 startTimer(); 3911 sample_eden(); 3912 // Get dirty region starting at nextOffset (inclusive), 3913 // simultaneously clearing it. 3914 dirtyRegion = 3915 _modUnionTable.getAndClearMarkedRegion(nextAddr, endAddr); 3916 assert(dirtyRegion.start() >= nextAddr, 3917 "returned region inconsistent?"); 3918 } 3919 // Remember where the next search should begin. 3920 // The returned region (if non-empty) is a right open interval, 3921 // so lastOffset is obtained from the right end of that 3922 // interval. 3923 lastAddr = dirtyRegion.end(); 3924 // Should do something more transparent and less hacky XXX 3925 numDirtyCards = 3926 _modUnionTable.heapWordDiffToOffsetDiff(dirtyRegion.word_size()); 3927 3928 // We'll scan the cards in the dirty region (with periodic 3929 // yields for foreground GC as needed). 3930 if (!dirtyRegion.is_empty()) { 3931 assert(numDirtyCards > 0, "consistency check"); 3932 HeapWord* stop_point = NULL; 3933 stopTimer(); 3934 // Potential yield point 3935 CMSTokenSyncWithLocks ts(true, old_gen->freelistLock(), 3936 bitMapLock()); 3937 startTimer(); 3938 { 3939 verify_work_stacks_empty(); 3940 verify_overflow_empty(); 3941 sample_eden(); 3942 stop_point = 3943 old_gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 3944 } 3945 if (stop_point != NULL) { 3946 // The careful iteration stopped early either because it found an 3947 // uninitialized object, or because we were in the midst of an 3948 // "abortable preclean", which should now be aborted. Redirty 3949 // the bits corresponding to the partially-scanned or unscanned 3950 // cards. We'll either restart at the next block boundary or 3951 // abort the preclean. 3952 assert((_collectorState == AbortablePreclean && should_abort_preclean()), 3953 "Should only be AbortablePreclean."); 3954 _modUnionTable.mark_range(MemRegion(stop_point, dirtyRegion.end())); 3955 if (should_abort_preclean()) { 3956 break; // out of preclean loop 3957 } else { 3958 // Compute the next address at which preclean should pick up; 3959 // might need bitMapLock in order to read P-bits. 3960 lastAddr = next_card_start_after_block(stop_point); 3961 } 3962 } 3963 } else { 3964 assert(lastAddr == endAddr, "consistency check"); 3965 assert(numDirtyCards == 0, "consistency check"); 3966 break; 3967 } 3968 } 3969 verify_work_stacks_empty(); 3970 verify_overflow_empty(); 3971 return cumNumDirtyCards; 3972 } 3973 3974 // NOTE: preclean_mod_union_table() above and preclean_card_table() 3975 // below are largely identical; if you need to modify 3976 // one of these methods, please check the other method too. 3977 3978 size_t CMSCollector::preclean_card_table(ConcurrentMarkSweepGeneration* old_gen, 3979 ScanMarkedObjectsAgainCarefullyClosure* cl) { 3980 // strategy: it's similar to precleamModUnionTable above, in that 3981 // we accumulate contiguous ranges of dirty cards, mark these cards 3982 // precleaned, then scan the region covered by these cards. 3983 HeapWord* endAddr = (HeapWord*)(old_gen->_virtual_space.high()); 3984 HeapWord* startAddr = (HeapWord*)(old_gen->_virtual_space.low()); 3985 3986 cl->setFreelistLock(old_gen->freelistLock()); // needed for yielding 3987 3988 size_t numDirtyCards, cumNumDirtyCards; 3989 HeapWord *lastAddr, *nextAddr; 3990 3991 for (cumNumDirtyCards = numDirtyCards = 0, 3992 nextAddr = lastAddr = startAddr; 3993 nextAddr < endAddr; 3994 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 3995 3996 ResourceMark rm; 3997 HandleMark hm; 3998 3999 MemRegion dirtyRegion; 4000 { 4001 // See comments in "Precleaning notes" above on why we 4002 // do this locking. XXX Could the locking overheads be 4003 // too high when dirty cards are sparse? [I don't think so.] 4004 stopTimer(); 4005 CMSTokenSync x(true); // is cms thread 4006 startTimer(); 4007 sample_eden(); 4008 // Get and clear dirty region from card table 4009 dirtyRegion = _ct->ct_bs()->dirty_card_range_after_reset( 4010 MemRegion(nextAddr, endAddr), 4011 true, 4012 CardTableModRefBS::precleaned_card_val()); 4013 4014 assert(dirtyRegion.start() >= nextAddr, 4015 "returned region inconsistent?"); 4016 } 4017 lastAddr = dirtyRegion.end(); 4018 numDirtyCards = 4019 dirtyRegion.word_size()/CardTableModRefBS::card_size_in_words; 4020 4021 if (!dirtyRegion.is_empty()) { 4022 stopTimer(); 4023 CMSTokenSyncWithLocks ts(true, old_gen->freelistLock(), bitMapLock()); 4024 startTimer(); 4025 sample_eden(); 4026 verify_work_stacks_empty(); 4027 verify_overflow_empty(); 4028 HeapWord* stop_point = 4029 old_gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 4030 if (stop_point != NULL) { 4031 assert((_collectorState == AbortablePreclean && should_abort_preclean()), 4032 "Should only be AbortablePreclean."); 4033 _ct->ct_bs()->invalidate(MemRegion(stop_point, dirtyRegion.end())); 4034 if (should_abort_preclean()) { 4035 break; // out of preclean loop 4036 } else { 4037 // Compute the next address at which preclean should pick up. 4038 lastAddr = next_card_start_after_block(stop_point); 4039 } 4040 } 4041 } else { 4042 break; 4043 } 4044 } 4045 verify_work_stacks_empty(); 4046 verify_overflow_empty(); 4047 return cumNumDirtyCards; 4048 } 4049 4050 class PrecleanKlassClosure : public KlassClosure { 4051 KlassToOopClosure _cm_klass_closure; 4052 public: 4053 PrecleanKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {} 4054 void do_klass(Klass* k) { 4055 if (k->has_accumulated_modified_oops()) { 4056 k->clear_accumulated_modified_oops(); 4057 4058 _cm_klass_closure.do_klass(k); 4059 } 4060 } 4061 }; 4062 4063 // The freelist lock is needed to prevent asserts, is it really needed? 4064 void CMSCollector::preclean_klasses(MarkRefsIntoAndScanClosure* cl, Mutex* freelistLock) { 4065 4066 cl->set_freelistLock(freelistLock); 4067 4068 CMSTokenSyncWithLocks ts(true, freelistLock, bitMapLock()); 4069 4070 // SSS: Add equivalent to ScanMarkedObjectsAgainCarefullyClosure::do_yield_check and should_abort_preclean? 4071 // SSS: We should probably check if precleaning should be aborted, at suitable intervals? 4072 PrecleanKlassClosure preclean_klass_closure(cl); 4073 ClassLoaderDataGraph::classes_do(&preclean_klass_closure); 4074 4075 verify_work_stacks_empty(); 4076 verify_overflow_empty(); 4077 } 4078 4079 void CMSCollector::checkpointRootsFinal() { 4080 assert(_collectorState == FinalMarking, "incorrect state transition?"); 4081 check_correct_thread_executing(); 4082 // world is stopped at this checkpoint 4083 assert(SafepointSynchronize::is_at_safepoint(), 4084 "world should be stopped"); 4085 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 4086 4087 verify_work_stacks_empty(); 4088 verify_overflow_empty(); 4089 4090 log_debug(gc)("YG occupancy: " SIZE_FORMAT " K (" SIZE_FORMAT " K)", 4091 _young_gen->used() / K, _young_gen->capacity() / K); 4092 { 4093 if (CMSScavengeBeforeRemark) { 4094 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4095 // Temporarily set flag to false, GCH->do_collection will 4096 // expect it to be false and set to true 4097 FlagSetting fl(gch->_is_gc_active, false); 4098 4099 GCTraceTime(Trace, gc) tm("Pause Scavenge Before Remark", _gc_timer_cm); 4100 4101 gch->do_collection(true, // full (i.e. force, see below) 4102 false, // !clear_all_soft_refs 4103 0, // size 4104 false, // is_tlab 4105 GenCollectedHeap::YoungGen // type 4106 ); 4107 } 4108 FreelistLocker x(this); 4109 MutexLockerEx y(bitMapLock(), 4110 Mutex::_no_safepoint_check_flag); 4111 checkpointRootsFinalWork(); 4112 } 4113 verify_work_stacks_empty(); 4114 verify_overflow_empty(); 4115 } 4116 4117 void CMSCollector::checkpointRootsFinalWork() { 4118 GCTraceTime(Trace, gc) tm("checkpointRootsFinalWork", _gc_timer_cm); 4119 4120 assert(haveFreelistLocks(), "must have free list locks"); 4121 assert_lock_strong(bitMapLock()); 4122 4123 ResourceMark rm; 4124 HandleMark hm; 4125 4126 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4127 4128 if (should_unload_classes()) { 4129 CodeCache::gc_prologue(); 4130 } 4131 assert(haveFreelistLocks(), "must have free list locks"); 4132 assert_lock_strong(bitMapLock()); 4133 4134 // We might assume that we need not fill TLAB's when 4135 // CMSScavengeBeforeRemark is set, because we may have just done 4136 // a scavenge which would have filled all TLAB's -- and besides 4137 // Eden would be empty. This however may not always be the case -- 4138 // for instance although we asked for a scavenge, it may not have 4139 // happened because of a JNI critical section. We probably need 4140 // a policy for deciding whether we can in that case wait until 4141 // the critical section releases and then do the remark following 4142 // the scavenge, and skip it here. In the absence of that policy, 4143 // or of an indication of whether the scavenge did indeed occur, 4144 // we cannot rely on TLAB's having been filled and must do 4145 // so here just in case a scavenge did not happen. 4146 gch->ensure_parsability(false); // fill TLAB's, but no need to retire them 4147 // Update the saved marks which may affect the root scans. 4148 gch->save_marks(); 4149 4150 print_eden_and_survivor_chunk_arrays(); 4151 4152 { 4153 #if defined(COMPILER2) || INCLUDE_JVMCI 4154 DerivedPointerTableDeactivate dpt_deact; 4155 #endif 4156 4157 // Note on the role of the mod union table: 4158 // Since the marker in "markFromRoots" marks concurrently with 4159 // mutators, it is possible for some reachable objects not to have been 4160 // scanned. For instance, an only reference to an object A was 4161 // placed in object B after the marker scanned B. Unless B is rescanned, 4162 // A would be collected. Such updates to references in marked objects 4163 // are detected via the mod union table which is the set of all cards 4164 // dirtied since the first checkpoint in this GC cycle and prior to 4165 // the most recent young generation GC, minus those cleaned up by the 4166 // concurrent precleaning. 4167 if (CMSParallelRemarkEnabled) { 4168 GCTraceTime(Debug, gc) t("Rescan (parallel)", _gc_timer_cm); 4169 do_remark_parallel(); 4170 } else { 4171 GCTraceTime(Debug, gc) t("Rescan (non-parallel)", _gc_timer_cm); 4172 do_remark_non_parallel(); 4173 } 4174 } 4175 verify_work_stacks_empty(); 4176 verify_overflow_empty(); 4177 4178 { 4179 GCTraceTime(Trace, gc) ts("refProcessingWork", _gc_timer_cm); 4180 refProcessingWork(); 4181 } 4182 verify_work_stacks_empty(); 4183 verify_overflow_empty(); 4184 4185 if (should_unload_classes()) { 4186 CodeCache::gc_epilogue(); 4187 } 4188 JvmtiExport::gc_epilogue(); 4189 4190 // If we encountered any (marking stack / work queue) overflow 4191 // events during the current CMS cycle, take appropriate 4192 // remedial measures, where possible, so as to try and avoid 4193 // recurrence of that condition. 4194 assert(_markStack.isEmpty(), "No grey objects"); 4195 size_t ser_ovflw = _ser_pmc_remark_ovflw + _ser_pmc_preclean_ovflw + 4196 _ser_kac_ovflw + _ser_kac_preclean_ovflw; 4197 if (ser_ovflw > 0) { 4198 log_trace(gc)("Marking stack overflow (benign) (pmc_pc=" SIZE_FORMAT ", pmc_rm=" SIZE_FORMAT ", kac=" SIZE_FORMAT ", kac_preclean=" SIZE_FORMAT ")", 4199 _ser_pmc_preclean_ovflw, _ser_pmc_remark_ovflw, _ser_kac_ovflw, _ser_kac_preclean_ovflw); 4200 _markStack.expand(); 4201 _ser_pmc_remark_ovflw = 0; 4202 _ser_pmc_preclean_ovflw = 0; 4203 _ser_kac_preclean_ovflw = 0; 4204 _ser_kac_ovflw = 0; 4205 } 4206 if (_par_pmc_remark_ovflw > 0 || _par_kac_ovflw > 0) { 4207 log_trace(gc)("Work queue overflow (benign) (pmc_rm=" SIZE_FORMAT ", kac=" SIZE_FORMAT ")", 4208 _par_pmc_remark_ovflw, _par_kac_ovflw); 4209 _par_pmc_remark_ovflw = 0; 4210 _par_kac_ovflw = 0; 4211 } 4212 if (_markStack._hit_limit > 0) { 4213 log_trace(gc)(" (benign) Hit max stack size limit (" SIZE_FORMAT ")", 4214 _markStack._hit_limit); 4215 } 4216 if (_markStack._failed_double > 0) { 4217 log_trace(gc)(" (benign) Failed stack doubling (" SIZE_FORMAT "), current capacity " SIZE_FORMAT, 4218 _markStack._failed_double, _markStack.capacity()); 4219 } 4220 _markStack._hit_limit = 0; 4221 _markStack._failed_double = 0; 4222 4223 if ((VerifyAfterGC || VerifyDuringGC) && 4224 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 4225 verify_after_remark(); 4226 } 4227 4228 _gc_tracer_cm->report_object_count_after_gc(&_is_alive_closure); 4229 4230 // Change under the freelistLocks. 4231 _collectorState = Sweeping; 4232 // Call isAllClear() under bitMapLock 4233 assert(_modUnionTable.isAllClear(), 4234 "Should be clear by end of the final marking"); 4235 assert(_ct->klass_rem_set()->mod_union_is_clear(), 4236 "Should be clear by end of the final marking"); 4237 } 4238 4239 void CMSParInitialMarkTask::work(uint worker_id) { 4240 elapsedTimer _timer; 4241 ResourceMark rm; 4242 HandleMark hm; 4243 4244 // ---------- scan from roots -------------- 4245 _timer.start(); 4246 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4247 Par_MarkRefsIntoClosure par_mri_cl(_collector->_span, &(_collector->_markBitMap)); 4248 4249 // ---------- young gen roots -------------- 4250 { 4251 work_on_young_gen_roots(worker_id, &par_mri_cl); 4252 _timer.stop(); 4253 log_trace(gc, task)("Finished young gen initial mark scan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4254 } 4255 4256 // ---------- remaining roots -------------- 4257 _timer.reset(); 4258 _timer.start(); 4259 4260 CLDToOopClosure cld_closure(&par_mri_cl, true); 4261 4262 gch->gen_process_roots(_strong_roots_scope, 4263 GenCollectedHeap::OldGen, 4264 false, // yg was scanned above 4265 GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()), 4266 _collector->should_unload_classes(), 4267 &par_mri_cl, 4268 NULL, 4269 &cld_closure); 4270 assert(_collector->should_unload_classes() 4271 || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache), 4272 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 4273 _timer.stop(); 4274 log_trace(gc, task)("Finished remaining root initial mark scan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4275 } 4276 4277 // Parallel remark task 4278 class CMSParRemarkTask: public CMSParMarkTask { 4279 CompactibleFreeListSpace* _cms_space; 4280 4281 // The per-thread work queues, available here for stealing. 4282 OopTaskQueueSet* _task_queues; 4283 ParallelTaskTerminator _term; 4284 StrongRootsScope* _strong_roots_scope; 4285 4286 public: 4287 // A value of 0 passed to n_workers will cause the number of 4288 // workers to be taken from the active workers in the work gang. 4289 CMSParRemarkTask(CMSCollector* collector, 4290 CompactibleFreeListSpace* cms_space, 4291 uint n_workers, WorkGang* workers, 4292 OopTaskQueueSet* task_queues, 4293 StrongRootsScope* strong_roots_scope): 4294 CMSParMarkTask("Rescan roots and grey objects in parallel", 4295 collector, n_workers), 4296 _cms_space(cms_space), 4297 _task_queues(task_queues), 4298 _term(n_workers, task_queues), 4299 _strong_roots_scope(strong_roots_scope) { } 4300 4301 OopTaskQueueSet* task_queues() { return _task_queues; } 4302 4303 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 4304 4305 ParallelTaskTerminator* terminator() { return &_term; } 4306 uint n_workers() { return _n_workers; } 4307 4308 void work(uint worker_id); 4309 4310 private: 4311 // ... of dirty cards in old space 4312 void do_dirty_card_rescan_tasks(CompactibleFreeListSpace* sp, int i, 4313 Par_MarkRefsIntoAndScanClosure* cl); 4314 4315 // ... work stealing for the above 4316 void do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, int* seed); 4317 }; 4318 4319 class RemarkKlassClosure : public KlassClosure { 4320 KlassToOopClosure _cm_klass_closure; 4321 public: 4322 RemarkKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {} 4323 void do_klass(Klass* k) { 4324 // Check if we have modified any oops in the Klass during the concurrent marking. 4325 if (k->has_accumulated_modified_oops()) { 4326 k->clear_accumulated_modified_oops(); 4327 4328 // We could have transfered the current modified marks to the accumulated marks, 4329 // like we do with the Card Table to Mod Union Table. But it's not really necessary. 4330 } else if (k->has_modified_oops()) { 4331 // Don't clear anything, this info is needed by the next young collection. 4332 } else { 4333 // No modified oops in the Klass. 4334 return; 4335 } 4336 4337 // The klass has modified fields, need to scan the klass. 4338 _cm_klass_closure.do_klass(k); 4339 } 4340 }; 4341 4342 void CMSParMarkTask::work_on_young_gen_roots(uint worker_id, OopsInGenClosure* cl) { 4343 ParNewGeneration* young_gen = _collector->_young_gen; 4344 ContiguousSpace* eden_space = young_gen->eden(); 4345 ContiguousSpace* from_space = young_gen->from(); 4346 ContiguousSpace* to_space = young_gen->to(); 4347 4348 HeapWord** eca = _collector->_eden_chunk_array; 4349 size_t ect = _collector->_eden_chunk_index; 4350 HeapWord** sca = _collector->_survivor_chunk_array; 4351 size_t sct = _collector->_survivor_chunk_index; 4352 4353 assert(ect <= _collector->_eden_chunk_capacity, "out of bounds"); 4354 assert(sct <= _collector->_survivor_chunk_capacity, "out of bounds"); 4355 4356 do_young_space_rescan(worker_id, cl, to_space, NULL, 0); 4357 do_young_space_rescan(worker_id, cl, from_space, sca, sct); 4358 do_young_space_rescan(worker_id, cl, eden_space, eca, ect); 4359 } 4360 4361 // work_queue(i) is passed to the closure 4362 // Par_MarkRefsIntoAndScanClosure. The "i" parameter 4363 // also is passed to do_dirty_card_rescan_tasks() and to 4364 // do_work_steal() to select the i-th task_queue. 4365 4366 void CMSParRemarkTask::work(uint worker_id) { 4367 elapsedTimer _timer; 4368 ResourceMark rm; 4369 HandleMark hm; 4370 4371 // ---------- rescan from roots -------------- 4372 _timer.start(); 4373 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4374 Par_MarkRefsIntoAndScanClosure par_mrias_cl(_collector, 4375 _collector->_span, _collector->ref_processor(), 4376 &(_collector->_markBitMap), 4377 work_queue(worker_id)); 4378 4379 // Rescan young gen roots first since these are likely 4380 // coarsely partitioned and may, on that account, constitute 4381 // the critical path; thus, it's best to start off that 4382 // work first. 4383 // ---------- young gen roots -------------- 4384 { 4385 work_on_young_gen_roots(worker_id, &par_mrias_cl); 4386 _timer.stop(); 4387 log_trace(gc, task)("Finished young gen rescan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4388 } 4389 4390 // ---------- remaining roots -------------- 4391 _timer.reset(); 4392 _timer.start(); 4393 gch->gen_process_roots(_strong_roots_scope, 4394 GenCollectedHeap::OldGen, 4395 false, // yg was scanned above 4396 GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()), 4397 _collector->should_unload_classes(), 4398 &par_mrias_cl, 4399 NULL, 4400 NULL); // The dirty klasses will be handled below 4401 4402 assert(_collector->should_unload_classes() 4403 || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache), 4404 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 4405 _timer.stop(); 4406 log_trace(gc, task)("Finished remaining root rescan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4407 4408 // ---------- unhandled CLD scanning ---------- 4409 if (worker_id == 0) { // Single threaded at the moment. 4410 _timer.reset(); 4411 _timer.start(); 4412 4413 // Scan all new class loader data objects and new dependencies that were 4414 // introduced during concurrent marking. 4415 ResourceMark rm; 4416 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds(); 4417 for (int i = 0; i < array->length(); i++) { 4418 par_mrias_cl.do_cld_nv(array->at(i)); 4419 } 4420 4421 // We don't need to keep track of new CLDs anymore. 4422 ClassLoaderDataGraph::remember_new_clds(false); 4423 4424 _timer.stop(); 4425 log_trace(gc, task)("Finished unhandled CLD scanning work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4426 } 4427 4428 // ---------- dirty klass scanning ---------- 4429 if (worker_id == 0) { // Single threaded at the moment. 4430 _timer.reset(); 4431 _timer.start(); 4432 4433 // Scan all classes that was dirtied during the concurrent marking phase. 4434 RemarkKlassClosure remark_klass_closure(&par_mrias_cl); 4435 ClassLoaderDataGraph::classes_do(&remark_klass_closure); 4436 4437 _timer.stop(); 4438 log_trace(gc, task)("Finished dirty klass scanning work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4439 } 4440 4441 // We might have added oops to ClassLoaderData::_handles during the 4442 // concurrent marking phase. These oops point to newly allocated objects 4443 // that are guaranteed to be kept alive. Either by the direct allocation 4444 // code, or when the young collector processes the roots. Hence, 4445 // we don't have to revisit the _handles block during the remark phase. 4446 4447 // ---------- rescan dirty cards ------------ 4448 _timer.reset(); 4449 _timer.start(); 4450 4451 // Do the rescan tasks for each of the two spaces 4452 // (cms_space) in turn. 4453 // "worker_id" is passed to select the task_queue for "worker_id" 4454 do_dirty_card_rescan_tasks(_cms_space, worker_id, &par_mrias_cl); 4455 _timer.stop(); 4456 log_trace(gc, task)("Finished dirty card rescan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4457 4458 // ---------- steal work from other threads ... 4459 // ---------- ... and drain overflow list. 4460 _timer.reset(); 4461 _timer.start(); 4462 do_work_steal(worker_id, &par_mrias_cl, _collector->hash_seed(worker_id)); 4463 _timer.stop(); 4464 log_trace(gc, task)("Finished work stealing in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4465 } 4466 4467 // Note that parameter "i" is not used. 4468 void 4469 CMSParMarkTask::do_young_space_rescan(uint worker_id, 4470 OopsInGenClosure* cl, ContiguousSpace* space, 4471 HeapWord** chunk_array, size_t chunk_top) { 4472 // Until all tasks completed: 4473 // . claim an unclaimed task 4474 // . compute region boundaries corresponding to task claimed 4475 // using chunk_array 4476 // . par_oop_iterate(cl) over that region 4477 4478 ResourceMark rm; 4479 HandleMark hm; 4480 4481 SequentialSubTasksDone* pst = space->par_seq_tasks(); 4482 4483 uint nth_task = 0; 4484 uint n_tasks = pst->n_tasks(); 4485 4486 if (n_tasks > 0) { 4487 assert(pst->valid(), "Uninitialized use?"); 4488 HeapWord *start, *end; 4489 while (!pst->is_task_claimed(/* reference */ nth_task)) { 4490 // We claimed task # nth_task; compute its boundaries. 4491 if (chunk_top == 0) { // no samples were taken 4492 assert(nth_task == 0 && n_tasks == 1, "Can have only 1 eden task"); 4493 start = space->bottom(); 4494 end = space->top(); 4495 } else if (nth_task == 0) { 4496 start = space->bottom(); 4497 end = chunk_array[nth_task]; 4498 } else if (nth_task < (uint)chunk_top) { 4499 assert(nth_task >= 1, "Control point invariant"); 4500 start = chunk_array[nth_task - 1]; 4501 end = chunk_array[nth_task]; 4502 } else { 4503 assert(nth_task == (uint)chunk_top, "Control point invariant"); 4504 start = chunk_array[chunk_top - 1]; 4505 end = space->top(); 4506 } 4507 MemRegion mr(start, end); 4508 // Verify that mr is in space 4509 assert(mr.is_empty() || space->used_region().contains(mr), 4510 "Should be in space"); 4511 // Verify that "start" is an object boundary 4512 assert(mr.is_empty() || oop(mr.start())->is_oop(), 4513 "Should be an oop"); 4514 space->par_oop_iterate(mr, cl); 4515 } 4516 pst->all_tasks_completed(); 4517 } 4518 } 4519 4520 void 4521 CMSParRemarkTask::do_dirty_card_rescan_tasks( 4522 CompactibleFreeListSpace* sp, int i, 4523 Par_MarkRefsIntoAndScanClosure* cl) { 4524 // Until all tasks completed: 4525 // . claim an unclaimed task 4526 // . compute region boundaries corresponding to task claimed 4527 // . transfer dirty bits ct->mut for that region 4528 // . apply rescanclosure to dirty mut bits for that region 4529 4530 ResourceMark rm; 4531 HandleMark hm; 4532 4533 OopTaskQueue* work_q = work_queue(i); 4534 ModUnionClosure modUnionClosure(&(_collector->_modUnionTable)); 4535 // CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! 4536 // CAUTION: This closure has state that persists across calls to 4537 // the work method dirty_range_iterate_clear() in that it has 4538 // embedded in it a (subtype of) UpwardsObjectClosure. The 4539 // use of that state in the embedded UpwardsObjectClosure instance 4540 // assumes that the cards are always iterated (even if in parallel 4541 // by several threads) in monotonically increasing order per each 4542 // thread. This is true of the implementation below which picks 4543 // card ranges (chunks) in monotonically increasing order globally 4544 // and, a-fortiori, in monotonically increasing order per thread 4545 // (the latter order being a subsequence of the former). 4546 // If the work code below is ever reorganized into a more chaotic 4547 // work-partitioning form than the current "sequential tasks" 4548 // paradigm, the use of that persistent state will have to be 4549 // revisited and modified appropriately. See also related 4550 // bug 4756801 work on which should examine this code to make 4551 // sure that the changes there do not run counter to the 4552 // assumptions made here and necessary for correctness and 4553 // efficiency. Note also that this code might yield inefficient 4554 // behavior in the case of very large objects that span one or 4555 // more work chunks. Such objects would potentially be scanned 4556 // several times redundantly. Work on 4756801 should try and 4557 // address that performance anomaly if at all possible. XXX 4558 MemRegion full_span = _collector->_span; 4559 CMSBitMap* bm = &(_collector->_markBitMap); // shared 4560 MarkFromDirtyCardsClosure 4561 greyRescanClosure(_collector, full_span, // entire span of interest 4562 sp, bm, work_q, cl); 4563 4564 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 4565 assert(pst->valid(), "Uninitialized use?"); 4566 uint nth_task = 0; 4567 const int alignment = CardTableModRefBS::card_size * BitsPerWord; 4568 MemRegion span = sp->used_region(); 4569 HeapWord* start_addr = span.start(); 4570 HeapWord* end_addr = (HeapWord*)round_to((intptr_t)span.end(), 4571 alignment); 4572 const size_t chunk_size = sp->rescan_task_size(); // in HeapWord units 4573 assert((HeapWord*)round_to((intptr_t)start_addr, alignment) == 4574 start_addr, "Check alignment"); 4575 assert((size_t)round_to((intptr_t)chunk_size, alignment) == 4576 chunk_size, "Check alignment"); 4577 4578 while (!pst->is_task_claimed(/* reference */ nth_task)) { 4579 // Having claimed the nth_task, compute corresponding mem-region, 4580 // which is a-fortiori aligned correctly (i.e. at a MUT boundary). 4581 // The alignment restriction ensures that we do not need any 4582 // synchronization with other gang-workers while setting or 4583 // clearing bits in thus chunk of the MUT. 4584 MemRegion this_span = MemRegion(start_addr + nth_task*chunk_size, 4585 start_addr + (nth_task+1)*chunk_size); 4586 // The last chunk's end might be way beyond end of the 4587 // used region. In that case pull back appropriately. 4588 if (this_span.end() > end_addr) { 4589 this_span.set_end(end_addr); 4590 assert(!this_span.is_empty(), "Program logic (calculation of n_tasks)"); 4591 } 4592 // Iterate over the dirty cards covering this chunk, marking them 4593 // precleaned, and setting the corresponding bits in the mod union 4594 // table. Since we have been careful to partition at Card and MUT-word 4595 // boundaries no synchronization is needed between parallel threads. 4596 _collector->_ct->ct_bs()->dirty_card_iterate(this_span, 4597 &modUnionClosure); 4598 4599 // Having transferred these marks into the modUnionTable, 4600 // rescan the marked objects on the dirty cards in the modUnionTable. 4601 // Even if this is at a synchronous collection, the initial marking 4602 // may have been done during an asynchronous collection so there 4603 // may be dirty bits in the mod-union table. 4604 _collector->_modUnionTable.dirty_range_iterate_clear( 4605 this_span, &greyRescanClosure); 4606 _collector->_modUnionTable.verifyNoOneBitsInRange( 4607 this_span.start(), 4608 this_span.end()); 4609 } 4610 pst->all_tasks_completed(); // declare that i am done 4611 } 4612 4613 // . see if we can share work_queues with ParNew? XXX 4614 void 4615 CMSParRemarkTask::do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, 4616 int* seed) { 4617 OopTaskQueue* work_q = work_queue(i); 4618 NOT_PRODUCT(int num_steals = 0;) 4619 oop obj_to_scan; 4620 CMSBitMap* bm = &(_collector->_markBitMap); 4621 4622 while (true) { 4623 // Completely finish any left over work from (an) earlier round(s) 4624 cl->trim_queue(0); 4625 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 4626 (size_t)ParGCDesiredObjsFromOverflowList); 4627 // Now check if there's any work in the overflow list 4628 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads, 4629 // only affects the number of attempts made to get work from the 4630 // overflow list and does not affect the number of workers. Just 4631 // pass ParallelGCThreads so this behavior is unchanged. 4632 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 4633 work_q, 4634 ParallelGCThreads)) { 4635 // found something in global overflow list; 4636 // not yet ready to go stealing work from others. 4637 // We'd like to assert(work_q->size() != 0, ...) 4638 // because we just took work from the overflow list, 4639 // but of course we can't since all of that could have 4640 // been already stolen from us. 4641 // "He giveth and He taketh away." 4642 continue; 4643 } 4644 // Verify that we have no work before we resort to stealing 4645 assert(work_q->size() == 0, "Have work, shouldn't steal"); 4646 // Try to steal from other queues that have work 4647 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 4648 NOT_PRODUCT(num_steals++;) 4649 assert(obj_to_scan->is_oop(), "Oops, not an oop!"); 4650 assert(bm->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 4651 // Do scanning work 4652 obj_to_scan->oop_iterate(cl); 4653 // Loop around, finish this work, and try to steal some more 4654 } else if (terminator()->offer_termination()) { 4655 break; // nirvana from the infinite cycle 4656 } 4657 } 4658 log_develop_trace(gc, task)("\t(%d: stole %d oops)", i, num_steals); 4659 assert(work_q->size() == 0 && _collector->overflow_list_is_empty(), 4660 "Else our work is not yet done"); 4661 } 4662 4663 // Record object boundaries in _eden_chunk_array by sampling the eden 4664 // top in the slow-path eden object allocation code path and record 4665 // the boundaries, if CMSEdenChunksRecordAlways is true. If 4666 // CMSEdenChunksRecordAlways is false, we use the other asynchronous 4667 // sampling in sample_eden() that activates during the part of the 4668 // preclean phase. 4669 void CMSCollector::sample_eden_chunk() { 4670 if (CMSEdenChunksRecordAlways && _eden_chunk_array != NULL) { 4671 if (_eden_chunk_lock->try_lock()) { 4672 // Record a sample. This is the critical section. The contents 4673 // of the _eden_chunk_array have to be non-decreasing in the 4674 // address order. 4675 _eden_chunk_array[_eden_chunk_index] = *_top_addr; 4676 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr, 4677 "Unexpected state of Eden"); 4678 if (_eden_chunk_index == 0 || 4679 ((_eden_chunk_array[_eden_chunk_index] > _eden_chunk_array[_eden_chunk_index-1]) && 4680 (pointer_delta(_eden_chunk_array[_eden_chunk_index], 4681 _eden_chunk_array[_eden_chunk_index-1]) >= CMSSamplingGrain))) { 4682 _eden_chunk_index++; // commit sample 4683 } 4684 _eden_chunk_lock->unlock(); 4685 } 4686 } 4687 } 4688 4689 // Return a thread-local PLAB recording array, as appropriate. 4690 void* CMSCollector::get_data_recorder(int thr_num) { 4691 if (_survivor_plab_array != NULL && 4692 (CMSPLABRecordAlways || 4693 (_collectorState > Marking && _collectorState < FinalMarking))) { 4694 assert(thr_num < (int)ParallelGCThreads, "thr_num is out of bounds"); 4695 ChunkArray* ca = &_survivor_plab_array[thr_num]; 4696 ca->reset(); // clear it so that fresh data is recorded 4697 return (void*) ca; 4698 } else { 4699 return NULL; 4700 } 4701 } 4702 4703 // Reset all the thread-local PLAB recording arrays 4704 void CMSCollector::reset_survivor_plab_arrays() { 4705 for (uint i = 0; i < ParallelGCThreads; i++) { 4706 _survivor_plab_array[i].reset(); 4707 } 4708 } 4709 4710 // Merge the per-thread plab arrays into the global survivor chunk 4711 // array which will provide the partitioning of the survivor space 4712 // for CMS initial scan and rescan. 4713 void CMSCollector::merge_survivor_plab_arrays(ContiguousSpace* surv, 4714 int no_of_gc_threads) { 4715 assert(_survivor_plab_array != NULL, "Error"); 4716 assert(_survivor_chunk_array != NULL, "Error"); 4717 assert(_collectorState == FinalMarking || 4718 (CMSParallelInitialMarkEnabled && _collectorState == InitialMarking), "Error"); 4719 for (int j = 0; j < no_of_gc_threads; j++) { 4720 _cursor[j] = 0; 4721 } 4722 HeapWord* top = surv->top(); 4723 size_t i; 4724 for (i = 0; i < _survivor_chunk_capacity; i++) { // all sca entries 4725 HeapWord* min_val = top; // Higher than any PLAB address 4726 uint min_tid = 0; // position of min_val this round 4727 for (int j = 0; j < no_of_gc_threads; j++) { 4728 ChunkArray* cur_sca = &_survivor_plab_array[j]; 4729 if (_cursor[j] == cur_sca->end()) { 4730 continue; 4731 } 4732 assert(_cursor[j] < cur_sca->end(), "ctl pt invariant"); 4733 HeapWord* cur_val = cur_sca->nth(_cursor[j]); 4734 assert(surv->used_region().contains(cur_val), "Out of bounds value"); 4735 if (cur_val < min_val) { 4736 min_tid = j; 4737 min_val = cur_val; 4738 } else { 4739 assert(cur_val < top, "All recorded addresses should be less"); 4740 } 4741 } 4742 // At this point min_val and min_tid are respectively 4743 // the least address in _survivor_plab_array[j]->nth(_cursor[j]) 4744 // and the thread (j) that witnesses that address. 4745 // We record this address in the _survivor_chunk_array[i] 4746 // and increment _cursor[min_tid] prior to the next round i. 4747 if (min_val == top) { 4748 break; 4749 } 4750 _survivor_chunk_array[i] = min_val; 4751 _cursor[min_tid]++; 4752 } 4753 // We are all done; record the size of the _survivor_chunk_array 4754 _survivor_chunk_index = i; // exclusive: [0, i) 4755 log_trace(gc, survivor)(" (Survivor:" SIZE_FORMAT "chunks) ", i); 4756 // Verify that we used up all the recorded entries 4757 #ifdef ASSERT 4758 size_t total = 0; 4759 for (int j = 0; j < no_of_gc_threads; j++) { 4760 assert(_cursor[j] == _survivor_plab_array[j].end(), "Ctl pt invariant"); 4761 total += _cursor[j]; 4762 } 4763 assert(total == _survivor_chunk_index, "Ctl Pt Invariant"); 4764 // Check that the merged array is in sorted order 4765 if (total > 0) { 4766 for (size_t i = 0; i < total - 1; i++) { 4767 log_develop_trace(gc, survivor)(" (chunk" SIZE_FORMAT ":" INTPTR_FORMAT ") ", 4768 i, p2i(_survivor_chunk_array[i])); 4769 assert(_survivor_chunk_array[i] < _survivor_chunk_array[i+1], 4770 "Not sorted"); 4771 } 4772 } 4773 #endif // ASSERT 4774 } 4775 4776 // Set up the space's par_seq_tasks structure for work claiming 4777 // for parallel initial scan and rescan of young gen. 4778 // See ParRescanTask where this is currently used. 4779 void 4780 CMSCollector:: 4781 initialize_sequential_subtasks_for_young_gen_rescan(int n_threads) { 4782 assert(n_threads > 0, "Unexpected n_threads argument"); 4783 4784 // Eden space 4785 if (!_young_gen->eden()->is_empty()) { 4786 SequentialSubTasksDone* pst = _young_gen->eden()->par_seq_tasks(); 4787 assert(!pst->valid(), "Clobbering existing data?"); 4788 // Each valid entry in [0, _eden_chunk_index) represents a task. 4789 size_t n_tasks = _eden_chunk_index + 1; 4790 assert(n_tasks == 1 || _eden_chunk_array != NULL, "Error"); 4791 // Sets the condition for completion of the subtask (how many threads 4792 // need to finish in order to be done). 4793 pst->set_n_threads(n_threads); 4794 pst->set_n_tasks((int)n_tasks); 4795 } 4796 4797 // Merge the survivor plab arrays into _survivor_chunk_array 4798 if (_survivor_plab_array != NULL) { 4799 merge_survivor_plab_arrays(_young_gen->from(), n_threads); 4800 } else { 4801 assert(_survivor_chunk_index == 0, "Error"); 4802 } 4803 4804 // To space 4805 { 4806 SequentialSubTasksDone* pst = _young_gen->to()->par_seq_tasks(); 4807 assert(!pst->valid(), "Clobbering existing data?"); 4808 // Sets the condition for completion of the subtask (how many threads 4809 // need to finish in order to be done). 4810 pst->set_n_threads(n_threads); 4811 pst->set_n_tasks(1); 4812 assert(pst->valid(), "Error"); 4813 } 4814 4815 // From space 4816 { 4817 SequentialSubTasksDone* pst = _young_gen->from()->par_seq_tasks(); 4818 assert(!pst->valid(), "Clobbering existing data?"); 4819 size_t n_tasks = _survivor_chunk_index + 1; 4820 assert(n_tasks == 1 || _survivor_chunk_array != NULL, "Error"); 4821 // Sets the condition for completion of the subtask (how many threads 4822 // need to finish in order to be done). 4823 pst->set_n_threads(n_threads); 4824 pst->set_n_tasks((int)n_tasks); 4825 assert(pst->valid(), "Error"); 4826 } 4827 } 4828 4829 // Parallel version of remark 4830 void CMSCollector::do_remark_parallel() { 4831 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4832 WorkGang* workers = gch->workers(); 4833 assert(workers != NULL, "Need parallel worker threads."); 4834 // Choose to use the number of GC workers most recently set 4835 // into "active_workers". 4836 uint n_workers = workers->active_workers(); 4837 4838 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 4839 4840 StrongRootsScope srs(n_workers); 4841 4842 CMSParRemarkTask tsk(this, cms_space, n_workers, workers, task_queues(), &srs); 4843 4844 // We won't be iterating over the cards in the card table updating 4845 // the younger_gen cards, so we shouldn't call the following else 4846 // the verification code as well as subsequent younger_refs_iterate 4847 // code would get confused. XXX 4848 // gch->rem_set()->prepare_for_younger_refs_iterate(true); // parallel 4849 4850 // The young gen rescan work will not be done as part of 4851 // process_roots (which currently doesn't know how to 4852 // parallelize such a scan), but rather will be broken up into 4853 // a set of parallel tasks (via the sampling that the [abortable] 4854 // preclean phase did of eden, plus the [two] tasks of 4855 // scanning the [two] survivor spaces. Further fine-grain 4856 // parallelization of the scanning of the survivor spaces 4857 // themselves, and of precleaning of the young gen itself 4858 // is deferred to the future. 4859 initialize_sequential_subtasks_for_young_gen_rescan(n_workers); 4860 4861 // The dirty card rescan work is broken up into a "sequence" 4862 // of parallel tasks (per constituent space) that are dynamically 4863 // claimed by the parallel threads. 4864 cms_space->initialize_sequential_subtasks_for_rescan(n_workers); 4865 4866 // It turns out that even when we're using 1 thread, doing the work in a 4867 // separate thread causes wide variance in run times. We can't help this 4868 // in the multi-threaded case, but we special-case n=1 here to get 4869 // repeatable measurements of the 1-thread overhead of the parallel code. 4870 if (n_workers > 1) { 4871 // Make refs discovery MT-safe, if it isn't already: it may not 4872 // necessarily be so, since it's possible that we are doing 4873 // ST marking. 4874 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), true); 4875 workers->run_task(&tsk); 4876 } else { 4877 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false); 4878 tsk.work(0); 4879 } 4880 4881 // restore, single-threaded for now, any preserved marks 4882 // as a result of work_q overflow 4883 restore_preserved_marks_if_any(); 4884 } 4885 4886 // Non-parallel version of remark 4887 void CMSCollector::do_remark_non_parallel() { 4888 ResourceMark rm; 4889 HandleMark hm; 4890 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4891 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false); 4892 4893 MarkRefsIntoAndScanClosure 4894 mrias_cl(_span, ref_processor(), &_markBitMap, NULL /* not precleaning */, 4895 &_markStack, this, 4896 false /* should_yield */, false /* not precleaning */); 4897 MarkFromDirtyCardsClosure 4898 markFromDirtyCardsClosure(this, _span, 4899 NULL, // space is set further below 4900 &_markBitMap, &_markStack, &mrias_cl); 4901 { 4902 GCTraceTime(Trace, gc) t("Grey Object Rescan", _gc_timer_cm); 4903 // Iterate over the dirty cards, setting the corresponding bits in the 4904 // mod union table. 4905 { 4906 ModUnionClosure modUnionClosure(&_modUnionTable); 4907 _ct->ct_bs()->dirty_card_iterate( 4908 _cmsGen->used_region(), 4909 &modUnionClosure); 4910 } 4911 // Having transferred these marks into the modUnionTable, we just need 4912 // to rescan the marked objects on the dirty cards in the modUnionTable. 4913 // The initial marking may have been done during an asynchronous 4914 // collection so there may be dirty bits in the mod-union table. 4915 const int alignment = 4916 CardTableModRefBS::card_size * BitsPerWord; 4917 { 4918 // ... First handle dirty cards in CMS gen 4919 markFromDirtyCardsClosure.set_space(_cmsGen->cmsSpace()); 4920 MemRegion ur = _cmsGen->used_region(); 4921 HeapWord* lb = ur.start(); 4922 HeapWord* ub = (HeapWord*)round_to((intptr_t)ur.end(), alignment); 4923 MemRegion cms_span(lb, ub); 4924 _modUnionTable.dirty_range_iterate_clear(cms_span, 4925 &markFromDirtyCardsClosure); 4926 verify_work_stacks_empty(); 4927 log_trace(gc)(" (re-scanned " SIZE_FORMAT " dirty cards in cms gen) ", markFromDirtyCardsClosure.num_dirty_cards()); 4928 } 4929 } 4930 if (VerifyDuringGC && 4931 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 4932 HandleMark hm; // Discard invalid handles created during verification 4933 Universe::verify(); 4934 } 4935 { 4936 GCTraceTime(Trace, gc) t("Root Rescan", _gc_timer_cm); 4937 4938 verify_work_stacks_empty(); 4939 4940 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 4941 StrongRootsScope srs(1); 4942 4943 gch->gen_process_roots(&srs, 4944 GenCollectedHeap::OldGen, 4945 true, // young gen as roots 4946 GenCollectedHeap::ScanningOption(roots_scanning_options()), 4947 should_unload_classes(), 4948 &mrias_cl, 4949 NULL, 4950 NULL); // The dirty klasses will be handled below 4951 4952 assert(should_unload_classes() 4953 || (roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache), 4954 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 4955 } 4956 4957 { 4958 GCTraceTime(Trace, gc) t("Visit Unhandled CLDs", _gc_timer_cm); 4959 4960 verify_work_stacks_empty(); 4961 4962 // Scan all class loader data objects that might have been introduced 4963 // during concurrent marking. 4964 ResourceMark rm; 4965 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds(); 4966 for (int i = 0; i < array->length(); i++) { 4967 mrias_cl.do_cld_nv(array->at(i)); 4968 } 4969 4970 // We don't need to keep track of new CLDs anymore. 4971 ClassLoaderDataGraph::remember_new_clds(false); 4972 4973 verify_work_stacks_empty(); 4974 } 4975 4976 { 4977 GCTraceTime(Trace, gc) t("Dirty Klass Scan", _gc_timer_cm); 4978 4979 verify_work_stacks_empty(); 4980 4981 RemarkKlassClosure remark_klass_closure(&mrias_cl); 4982 ClassLoaderDataGraph::classes_do(&remark_klass_closure); 4983 4984 verify_work_stacks_empty(); 4985 } 4986 4987 // We might have added oops to ClassLoaderData::_handles during the 4988 // concurrent marking phase. These oops point to newly allocated objects 4989 // that are guaranteed to be kept alive. Either by the direct allocation 4990 // code, or when the young collector processes the roots. Hence, 4991 // we don't have to revisit the _handles block during the remark phase. 4992 4993 verify_work_stacks_empty(); 4994 // Restore evacuated mark words, if any, used for overflow list links 4995 restore_preserved_marks_if_any(); 4996 4997 verify_overflow_empty(); 4998 } 4999 5000 //////////////////////////////////////////////////////// 5001 // Parallel Reference Processing Task Proxy Class 5002 //////////////////////////////////////////////////////// 5003 class AbstractGangTaskWOopQueues : public AbstractGangTask { 5004 OopTaskQueueSet* _queues; 5005 ParallelTaskTerminator _terminator; 5006 public: 5007 AbstractGangTaskWOopQueues(const char* name, OopTaskQueueSet* queues, uint n_threads) : 5008 AbstractGangTask(name), _queues(queues), _terminator(n_threads, _queues) {} 5009 ParallelTaskTerminator* terminator() { return &_terminator; } 5010 OopTaskQueueSet* queues() { return _queues; } 5011 }; 5012 5013 class CMSRefProcTaskProxy: public AbstractGangTaskWOopQueues { 5014 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask; 5015 CMSCollector* _collector; 5016 CMSBitMap* _mark_bit_map; 5017 const MemRegion _span; 5018 ProcessTask& _task; 5019 5020 public: 5021 CMSRefProcTaskProxy(ProcessTask& task, 5022 CMSCollector* collector, 5023 const MemRegion& span, 5024 CMSBitMap* mark_bit_map, 5025 AbstractWorkGang* workers, 5026 OopTaskQueueSet* task_queues): 5027 AbstractGangTaskWOopQueues("Process referents by policy in parallel", 5028 task_queues, 5029 workers->active_workers()), 5030 _task(task), 5031 _collector(collector), _span(span), _mark_bit_map(mark_bit_map) 5032 { 5033 assert(_collector->_span.equals(_span) && !_span.is_empty(), 5034 "Inconsistency in _span"); 5035 } 5036 5037 OopTaskQueueSet* task_queues() { return queues(); } 5038 5039 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 5040 5041 void do_work_steal(int i, 5042 CMSParDrainMarkingStackClosure* drain, 5043 CMSParKeepAliveClosure* keep_alive, 5044 int* seed); 5045 5046 virtual void work(uint worker_id); 5047 }; 5048 5049 void CMSRefProcTaskProxy::work(uint worker_id) { 5050 ResourceMark rm; 5051 HandleMark hm; 5052 assert(_collector->_span.equals(_span), "Inconsistency in _span"); 5053 CMSParKeepAliveClosure par_keep_alive(_collector, _span, 5054 _mark_bit_map, 5055 work_queue(worker_id)); 5056 CMSParDrainMarkingStackClosure par_drain_stack(_collector, _span, 5057 _mark_bit_map, 5058 work_queue(worker_id)); 5059 CMSIsAliveClosure is_alive_closure(_span, _mark_bit_map); 5060 _task.work(worker_id, is_alive_closure, par_keep_alive, par_drain_stack); 5061 if (_task.marks_oops_alive()) { 5062 do_work_steal(worker_id, &par_drain_stack, &par_keep_alive, 5063 _collector->hash_seed(worker_id)); 5064 } 5065 assert(work_queue(worker_id)->size() == 0, "work_queue should be empty"); 5066 assert(_collector->_overflow_list == NULL, "non-empty _overflow_list"); 5067 } 5068 5069 class CMSRefEnqueueTaskProxy: public AbstractGangTask { 5070 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask; 5071 EnqueueTask& _task; 5072 5073 public: 5074 CMSRefEnqueueTaskProxy(EnqueueTask& task) 5075 : AbstractGangTask("Enqueue reference objects in parallel"), 5076 _task(task) 5077 { } 5078 5079 virtual void work(uint worker_id) 5080 { 5081 _task.work(worker_id); 5082 } 5083 }; 5084 5085 CMSParKeepAliveClosure::CMSParKeepAliveClosure(CMSCollector* collector, 5086 MemRegion span, CMSBitMap* bit_map, OopTaskQueue* work_queue): 5087 _span(span), 5088 _bit_map(bit_map), 5089 _work_queue(work_queue), 5090 _mark_and_push(collector, span, bit_map, work_queue), 5091 _low_water_mark(MIN2((work_queue->max_elems()/4), 5092 ((uint)CMSWorkQueueDrainThreshold * ParallelGCThreads))) 5093 { } 5094 5095 // . see if we can share work_queues with ParNew? XXX 5096 void CMSRefProcTaskProxy::do_work_steal(int i, 5097 CMSParDrainMarkingStackClosure* drain, 5098 CMSParKeepAliveClosure* keep_alive, 5099 int* seed) { 5100 OopTaskQueue* work_q = work_queue(i); 5101 NOT_PRODUCT(int num_steals = 0;) 5102 oop obj_to_scan; 5103 5104 while (true) { 5105 // Completely finish any left over work from (an) earlier round(s) 5106 drain->trim_queue(0); 5107 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 5108 (size_t)ParGCDesiredObjsFromOverflowList); 5109 // Now check if there's any work in the overflow list 5110 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads, 5111 // only affects the number of attempts made to get work from the 5112 // overflow list and does not affect the number of workers. Just 5113 // pass ParallelGCThreads so this behavior is unchanged. 5114 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 5115 work_q, 5116 ParallelGCThreads)) { 5117 // Found something in global overflow list; 5118 // not yet ready to go stealing work from others. 5119 // We'd like to assert(work_q->size() != 0, ...) 5120 // because we just took work from the overflow list, 5121 // but of course we can't, since all of that might have 5122 // been already stolen from us. 5123 continue; 5124 } 5125 // Verify that we have no work before we resort to stealing 5126 assert(work_q->size() == 0, "Have work, shouldn't steal"); 5127 // Try to steal from other queues that have work 5128 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 5129 NOT_PRODUCT(num_steals++;) 5130 assert(obj_to_scan->is_oop(), "Oops, not an oop!"); 5131 assert(_mark_bit_map->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 5132 // Do scanning work 5133 obj_to_scan->oop_iterate(keep_alive); 5134 // Loop around, finish this work, and try to steal some more 5135 } else if (terminator()->offer_termination()) { 5136 break; // nirvana from the infinite cycle 5137 } 5138 } 5139 log_develop_trace(gc, task)("\t(%d: stole %d oops)", i, num_steals); 5140 } 5141 5142 void CMSRefProcTaskExecutor::execute(ProcessTask& task) 5143 { 5144 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5145 WorkGang* workers = gch->workers(); 5146 assert(workers != NULL, "Need parallel worker threads."); 5147 CMSRefProcTaskProxy rp_task(task, &_collector, 5148 _collector.ref_processor()->span(), 5149 _collector.markBitMap(), 5150 workers, _collector.task_queues()); 5151 workers->run_task(&rp_task); 5152 } 5153 5154 void CMSRefProcTaskExecutor::execute(EnqueueTask& task) 5155 { 5156 5157 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5158 WorkGang* workers = gch->workers(); 5159 assert(workers != NULL, "Need parallel worker threads."); 5160 CMSRefEnqueueTaskProxy enq_task(task); 5161 workers->run_task(&enq_task); 5162 } 5163 5164 void CMSCollector::refProcessingWork() { 5165 ResourceMark rm; 5166 HandleMark hm; 5167 5168 ReferenceProcessor* rp = ref_processor(); 5169 assert(rp->span().equals(_span), "Spans should be equal"); 5170 assert(!rp->enqueuing_is_done(), "Enqueuing should not be complete"); 5171 // Process weak references. 5172 rp->setup_policy(false); 5173 verify_work_stacks_empty(); 5174 5175 CMSKeepAliveClosure cmsKeepAliveClosure(this, _span, &_markBitMap, 5176 &_markStack, false /* !preclean */); 5177 CMSDrainMarkingStackClosure cmsDrainMarkingStackClosure(this, 5178 _span, &_markBitMap, &_markStack, 5179 &cmsKeepAliveClosure, false /* !preclean */); 5180 { 5181 GCTraceTime(Debug, gc) t("Weak Refs Processing", _gc_timer_cm); 5182 5183 ReferenceProcessorStats stats; 5184 if (rp->processing_is_mt()) { 5185 // Set the degree of MT here. If the discovery is done MT, there 5186 // may have been a different number of threads doing the discovery 5187 // and a different number of discovered lists may have Ref objects. 5188 // That is OK as long as the Reference lists are balanced (see 5189 // balance_all_queues() and balance_queues()). 5190 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5191 uint active_workers = ParallelGCThreads; 5192 WorkGang* workers = gch->workers(); 5193 if (workers != NULL) { 5194 active_workers = workers->active_workers(); 5195 // The expectation is that active_workers will have already 5196 // been set to a reasonable value. If it has not been set, 5197 // investigate. 5198 assert(active_workers > 0, "Should have been set during scavenge"); 5199 } 5200 rp->set_active_mt_degree(active_workers); 5201 CMSRefProcTaskExecutor task_executor(*this); 5202 stats = rp->process_discovered_references(&_is_alive_closure, 5203 &cmsKeepAliveClosure, 5204 &cmsDrainMarkingStackClosure, 5205 &task_executor, 5206 _gc_timer_cm); 5207 } else { 5208 stats = rp->process_discovered_references(&_is_alive_closure, 5209 &cmsKeepAliveClosure, 5210 &cmsDrainMarkingStackClosure, 5211 NULL, 5212 _gc_timer_cm); 5213 } 5214 _gc_tracer_cm->report_gc_reference_stats(stats); 5215 5216 } 5217 5218 // This is the point where the entire marking should have completed. 5219 verify_work_stacks_empty(); 5220 5221 if (should_unload_classes()) { 5222 { 5223 GCTraceTime(Debug, gc) t("Class Unloading", _gc_timer_cm); 5224 5225 // Unload classes and purge the SystemDictionary. 5226 bool purged_class = SystemDictionary::do_unloading(&_is_alive_closure); 5227 5228 // Unload nmethods. 5229 CodeCache::do_unloading(&_is_alive_closure, purged_class); 5230 5231 // Prune dead klasses from subklass/sibling/implementor lists. 5232 Klass::clean_weak_klass_links(&_is_alive_closure); 5233 } 5234 5235 { 5236 GCTraceTime(Debug, gc) t("Scrub Symbol Table", _gc_timer_cm); 5237 // Clean up unreferenced symbols in symbol table. 5238 SymbolTable::unlink(); 5239 } 5240 5241 { 5242 GCTraceTime(Debug, gc) t("Scrub String Table", _gc_timer_cm); 5243 // Delete entries for dead interned strings. 5244 StringTable::unlink(&_is_alive_closure); 5245 } 5246 } 5247 5248 5249 // Restore any preserved marks as a result of mark stack or 5250 // work queue overflow 5251 restore_preserved_marks_if_any(); // done single-threaded for now 5252 5253 rp->set_enqueuing_is_done(true); 5254 if (rp->processing_is_mt()) { 5255 rp->balance_all_queues(); 5256 CMSRefProcTaskExecutor task_executor(*this); 5257 rp->enqueue_discovered_references(&task_executor); 5258 } else { 5259 rp->enqueue_discovered_references(NULL); 5260 } 5261 rp->verify_no_references_recorded(); 5262 assert(!rp->discovery_enabled(), "should have been disabled"); 5263 } 5264 5265 #ifndef PRODUCT 5266 void CMSCollector::check_correct_thread_executing() { 5267 Thread* t = Thread::current(); 5268 // Only the VM thread or the CMS thread should be here. 5269 assert(t->is_ConcurrentGC_thread() || t->is_VM_thread(), 5270 "Unexpected thread type"); 5271 // If this is the vm thread, the foreground process 5272 // should not be waiting. Note that _foregroundGCIsActive is 5273 // true while the foreground collector is waiting. 5274 if (_foregroundGCShouldWait) { 5275 // We cannot be the VM thread 5276 assert(t->is_ConcurrentGC_thread(), 5277 "Should be CMS thread"); 5278 } else { 5279 // We can be the CMS thread only if we are in a stop-world 5280 // phase of CMS collection. 5281 if (t->is_ConcurrentGC_thread()) { 5282 assert(_collectorState == InitialMarking || 5283 _collectorState == FinalMarking, 5284 "Should be a stop-world phase"); 5285 // The CMS thread should be holding the CMS_token. 5286 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 5287 "Potential interference with concurrently " 5288 "executing VM thread"); 5289 } 5290 } 5291 } 5292 #endif 5293 5294 void CMSCollector::sweep() { 5295 assert(_collectorState == Sweeping, "just checking"); 5296 check_correct_thread_executing(); 5297 verify_work_stacks_empty(); 5298 verify_overflow_empty(); 5299 increment_sweep_count(); 5300 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 5301 5302 _inter_sweep_timer.stop(); 5303 _inter_sweep_estimate.sample(_inter_sweep_timer.seconds()); 5304 5305 assert(!_intra_sweep_timer.is_active(), "Should not be active"); 5306 _intra_sweep_timer.reset(); 5307 _intra_sweep_timer.start(); 5308 { 5309 GCTraceCPUTime tcpu; 5310 CMSPhaseAccounting pa(this, "Concurrent Sweep"); 5311 // First sweep the old gen 5312 { 5313 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(), 5314 bitMapLock()); 5315 sweepWork(_cmsGen); 5316 } 5317 5318 // Update Universe::_heap_*_at_gc figures. 5319 // We need all the free list locks to make the abstract state 5320 // transition from Sweeping to Resetting. See detailed note 5321 // further below. 5322 { 5323 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock()); 5324 // Update heap occupancy information which is used as 5325 // input to soft ref clearing policy at the next gc. 5326 Universe::update_heap_info_at_gc(); 5327 _collectorState = Resizing; 5328 } 5329 } 5330 verify_work_stacks_empty(); 5331 verify_overflow_empty(); 5332 5333 if (should_unload_classes()) { 5334 // Delay purge to the beginning of the next safepoint. Metaspace::contains 5335 // requires that the virtual spaces are stable and not deleted. 5336 ClassLoaderDataGraph::set_should_purge(true); 5337 } 5338 5339 _intra_sweep_timer.stop(); 5340 _intra_sweep_estimate.sample(_intra_sweep_timer.seconds()); 5341 5342 _inter_sweep_timer.reset(); 5343 _inter_sweep_timer.start(); 5344 5345 // We need to use a monotonically non-decreasing time in ms 5346 // or we will see time-warp warnings and os::javaTimeMillis() 5347 // does not guarantee monotonicity. 5348 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC; 5349 update_time_of_last_gc(now); 5350 5351 // NOTE on abstract state transitions: 5352 // Mutators allocate-live and/or mark the mod-union table dirty 5353 // based on the state of the collection. The former is done in 5354 // the interval [Marking, Sweeping] and the latter in the interval 5355 // [Marking, Sweeping). Thus the transitions into the Marking state 5356 // and out of the Sweeping state must be synchronously visible 5357 // globally to the mutators. 5358 // The transition into the Marking state happens with the world 5359 // stopped so the mutators will globally see it. Sweeping is 5360 // done asynchronously by the background collector so the transition 5361 // from the Sweeping state to the Resizing state must be done 5362 // under the freelistLock (as is the check for whether to 5363 // allocate-live and whether to dirty the mod-union table). 5364 assert(_collectorState == Resizing, "Change of collector state to" 5365 " Resizing must be done under the freelistLocks (plural)"); 5366 5367 // Now that sweeping has been completed, we clear 5368 // the incremental_collection_failed flag, 5369 // thus inviting a younger gen collection to promote into 5370 // this generation. If such a promotion may still fail, 5371 // the flag will be set again when a young collection is 5372 // attempted. 5373 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5374 gch->clear_incremental_collection_failed(); // Worth retrying as fresh space may have been freed up 5375 gch->update_full_collections_completed(_collection_count_start); 5376 } 5377 5378 // FIX ME!!! Looks like this belongs in CFLSpace, with 5379 // CMSGen merely delegating to it. 5380 void ConcurrentMarkSweepGeneration::setNearLargestChunk() { 5381 double nearLargestPercent = FLSLargestBlockCoalesceProximity; 5382 HeapWord* minAddr = _cmsSpace->bottom(); 5383 HeapWord* largestAddr = 5384 (HeapWord*) _cmsSpace->dictionary()->find_largest_dict(); 5385 if (largestAddr == NULL) { 5386 // The dictionary appears to be empty. In this case 5387 // try to coalesce at the end of the heap. 5388 largestAddr = _cmsSpace->end(); 5389 } 5390 size_t largestOffset = pointer_delta(largestAddr, minAddr); 5391 size_t nearLargestOffset = 5392 (size_t)((double)largestOffset * nearLargestPercent) - MinChunkSize; 5393 log_debug(gc, freelist)("CMS: Large Block: " PTR_FORMAT "; Proximity: " PTR_FORMAT " -> " PTR_FORMAT, 5394 p2i(largestAddr), p2i(_cmsSpace->nearLargestChunk()), p2i(minAddr + nearLargestOffset)); 5395 _cmsSpace->set_nearLargestChunk(minAddr + nearLargestOffset); 5396 } 5397 5398 bool ConcurrentMarkSweepGeneration::isNearLargestChunk(HeapWord* addr) { 5399 return addr >= _cmsSpace->nearLargestChunk(); 5400 } 5401 5402 FreeChunk* ConcurrentMarkSweepGeneration::find_chunk_at_end() { 5403 return _cmsSpace->find_chunk_at_end(); 5404 } 5405 5406 void ConcurrentMarkSweepGeneration::update_gc_stats(Generation* current_generation, 5407 bool full) { 5408 // If the young generation has been collected, gather any statistics 5409 // that are of interest at this point. 5410 bool current_is_young = GenCollectedHeap::heap()->is_young_gen(current_generation); 5411 if (!full && current_is_young) { 5412 // Gather statistics on the young generation collection. 5413 collector()->stats().record_gc0_end(used()); 5414 } 5415 } 5416 5417 void CMSCollector::sweepWork(ConcurrentMarkSweepGeneration* old_gen) { 5418 // We iterate over the space(s) underlying this generation, 5419 // checking the mark bit map to see if the bits corresponding 5420 // to specific blocks are marked or not. Blocks that are 5421 // marked are live and are not swept up. All remaining blocks 5422 // are swept up, with coalescing on-the-fly as we sweep up 5423 // contiguous free and/or garbage blocks: 5424 // We need to ensure that the sweeper synchronizes with allocators 5425 // and stop-the-world collectors. In particular, the following 5426 // locks are used: 5427 // . CMS token: if this is held, a stop the world collection cannot occur 5428 // . freelistLock: if this is held no allocation can occur from this 5429 // generation by another thread 5430 // . bitMapLock: if this is held, no other thread can access or update 5431 // 5432 5433 // Note that we need to hold the freelistLock if we use 5434 // block iterate below; else the iterator might go awry if 5435 // a mutator (or promotion) causes block contents to change 5436 // (for instance if the allocator divvies up a block). 5437 // If we hold the free list lock, for all practical purposes 5438 // young generation GC's can't occur (they'll usually need to 5439 // promote), so we might as well prevent all young generation 5440 // GC's while we do a sweeping step. For the same reason, we might 5441 // as well take the bit map lock for the entire duration 5442 5443 // check that we hold the requisite locks 5444 assert(have_cms_token(), "Should hold cms token"); 5445 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), "Should possess CMS token to sweep"); 5446 assert_lock_strong(old_gen->freelistLock()); 5447 assert_lock_strong(bitMapLock()); 5448 5449 assert(!_inter_sweep_timer.is_active(), "Was switched off in an outer context"); 5450 assert(_intra_sweep_timer.is_active(), "Was switched on in an outer context"); 5451 old_gen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()), 5452 _inter_sweep_estimate.padded_average(), 5453 _intra_sweep_estimate.padded_average()); 5454 old_gen->setNearLargestChunk(); 5455 5456 { 5457 SweepClosure sweepClosure(this, old_gen, &_markBitMap, CMSYield); 5458 old_gen->cmsSpace()->blk_iterate_careful(&sweepClosure); 5459 // We need to free-up/coalesce garbage/blocks from a 5460 // co-terminal free run. This is done in the SweepClosure 5461 // destructor; so, do not remove this scope, else the 5462 // end-of-sweep-census below will be off by a little bit. 5463 } 5464 old_gen->cmsSpace()->sweep_completed(); 5465 old_gen->cmsSpace()->endSweepFLCensus(sweep_count()); 5466 if (should_unload_classes()) { // unloaded classes this cycle, 5467 _concurrent_cycles_since_last_unload = 0; // ... reset count 5468 } else { // did not unload classes, 5469 _concurrent_cycles_since_last_unload++; // ... increment count 5470 } 5471 } 5472 5473 // Reset CMS data structures (for now just the marking bit map) 5474 // preparatory for the next cycle. 5475 void CMSCollector::reset_concurrent() { 5476 CMSTokenSyncWithLocks ts(true, bitMapLock()); 5477 5478 // If the state is not "Resetting", the foreground thread 5479 // has done a collection and the resetting. 5480 if (_collectorState != Resetting) { 5481 assert(_collectorState == Idling, "The state should only change" 5482 " because the foreground collector has finished the collection"); 5483 return; 5484 } 5485 5486 // Clear the mark bitmap (no grey objects to start with) 5487 // for the next cycle. 5488 GCTraceCPUTime tcpu; 5489 CMSPhaseAccounting cmspa(this, "Concurrent Reset"); 5490 5491 HeapWord* curAddr = _markBitMap.startWord(); 5492 while (curAddr < _markBitMap.endWord()) { 5493 size_t remaining = pointer_delta(_markBitMap.endWord(), curAddr); 5494 MemRegion chunk(curAddr, MIN2(CMSBitMapYieldQuantum, remaining)); 5495 _markBitMap.clear_large_range(chunk); 5496 if (ConcurrentMarkSweepThread::should_yield() && 5497 !foregroundGCIsActive() && 5498 CMSYield) { 5499 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 5500 "CMS thread should hold CMS token"); 5501 assert_lock_strong(bitMapLock()); 5502 bitMapLock()->unlock(); 5503 ConcurrentMarkSweepThread::desynchronize(true); 5504 stopTimer(); 5505 incrementYields(); 5506 5507 // See the comment in coordinator_yield() 5508 for (unsigned i = 0; i < CMSYieldSleepCount && 5509 ConcurrentMarkSweepThread::should_yield() && 5510 !CMSCollector::foregroundGCIsActive(); ++i) { 5511 os::sleep(Thread::current(), 1, false); 5512 } 5513 5514 ConcurrentMarkSweepThread::synchronize(true); 5515 bitMapLock()->lock_without_safepoint_check(); 5516 startTimer(); 5517 } 5518 curAddr = chunk.end(); 5519 } 5520 // A successful mostly concurrent collection has been done. 5521 // Because only the full (i.e., concurrent mode failure) collections 5522 // are being measured for gc overhead limits, clean the "near" flag 5523 // and count. 5524 size_policy()->reset_gc_overhead_limit_count(); 5525 _collectorState = Idling; 5526 5527 register_gc_end(); 5528 } 5529 5530 // Same as above but for STW paths 5531 void CMSCollector::reset_stw() { 5532 // already have the lock 5533 assert(_collectorState == Resetting, "just checking"); 5534 assert_lock_strong(bitMapLock()); 5535 GCIdMarkAndRestore gc_id_mark(_cmsThread->gc_id()); 5536 _markBitMap.clear_all(); 5537 _collectorState = Idling; 5538 register_gc_end(); 5539 } 5540 5541 void CMSCollector::do_CMS_operation(CMS_op_type op, GCCause::Cause gc_cause) { 5542 GCTraceCPUTime tcpu; 5543 TraceCollectorStats tcs(counters()); 5544 5545 switch (op) { 5546 case CMS_op_checkpointRootsInitial: { 5547 GCTraceTime(Info, gc) t("Pause Initial Mark", NULL, GCCause::_no_gc, true); 5548 SvcGCMarker sgcm(SvcGCMarker::OTHER); 5549 checkpointRootsInitial(); 5550 break; 5551 } 5552 case CMS_op_checkpointRootsFinal: { 5553 GCTraceTime(Info, gc) t("Pause Remark", NULL, GCCause::_no_gc, true); 5554 SvcGCMarker sgcm(SvcGCMarker::OTHER); 5555 checkpointRootsFinal(); 5556 break; 5557 } 5558 default: 5559 fatal("No such CMS_op"); 5560 } 5561 } 5562 5563 #ifndef PRODUCT 5564 size_t const CMSCollector::skip_header_HeapWords() { 5565 return FreeChunk::header_size(); 5566 } 5567 5568 // Try and collect here conditions that should hold when 5569 // CMS thread is exiting. The idea is that the foreground GC 5570 // thread should not be blocked if it wants to terminate 5571 // the CMS thread and yet continue to run the VM for a while 5572 // after that. 5573 void CMSCollector::verify_ok_to_terminate() const { 5574 assert(Thread::current()->is_ConcurrentGC_thread(), 5575 "should be called by CMS thread"); 5576 assert(!_foregroundGCShouldWait, "should be false"); 5577 // We could check here that all the various low-level locks 5578 // are not held by the CMS thread, but that is overkill; see 5579 // also CMSThread::verify_ok_to_terminate() where the CGC_lock 5580 // is checked. 5581 } 5582 #endif 5583 5584 size_t CMSCollector::block_size_using_printezis_bits(HeapWord* addr) const { 5585 assert(_markBitMap.isMarked(addr) && _markBitMap.isMarked(addr + 1), 5586 "missing Printezis mark?"); 5587 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 5588 size_t size = pointer_delta(nextOneAddr + 1, addr); 5589 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 5590 "alignment problem"); 5591 assert(size >= 3, "Necessary for Printezis marks to work"); 5592 return size; 5593 } 5594 5595 // A variant of the above (block_size_using_printezis_bits()) except 5596 // that we return 0 if the P-bits are not yet set. 5597 size_t CMSCollector::block_size_if_printezis_bits(HeapWord* addr) const { 5598 if (_markBitMap.isMarked(addr + 1)) { 5599 assert(_markBitMap.isMarked(addr), "P-bit can be set only for marked objects"); 5600 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 5601 size_t size = pointer_delta(nextOneAddr + 1, addr); 5602 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 5603 "alignment problem"); 5604 assert(size >= 3, "Necessary for Printezis marks to work"); 5605 return size; 5606 } 5607 return 0; 5608 } 5609 5610 HeapWord* CMSCollector::next_card_start_after_block(HeapWord* addr) const { 5611 size_t sz = 0; 5612 oop p = (oop)addr; 5613 if (p->klass_or_null() != NULL) { 5614 sz = CompactibleFreeListSpace::adjustObjectSize(p->size()); 5615 } else { 5616 sz = block_size_using_printezis_bits(addr); 5617 } 5618 assert(sz > 0, "size must be nonzero"); 5619 HeapWord* next_block = addr + sz; 5620 HeapWord* next_card = (HeapWord*)round_to((uintptr_t)next_block, 5621 CardTableModRefBS::card_size); 5622 assert(round_down((uintptr_t)addr, CardTableModRefBS::card_size) < 5623 round_down((uintptr_t)next_card, CardTableModRefBS::card_size), 5624 "must be different cards"); 5625 return next_card; 5626 } 5627 5628 5629 // CMS Bit Map Wrapper ///////////////////////////////////////// 5630 5631 // Construct a CMS bit map infrastructure, but don't create the 5632 // bit vector itself. That is done by a separate call CMSBitMap::allocate() 5633 // further below. 5634 CMSBitMap::CMSBitMap(int shifter, int mutex_rank, const char* mutex_name): 5635 _bm(), 5636 _shifter(shifter), 5637 _lock(mutex_rank >= 0 ? new Mutex(mutex_rank, mutex_name, true, 5638 Monitor::_safepoint_check_sometimes) : NULL) 5639 { 5640 _bmStartWord = 0; 5641 _bmWordSize = 0; 5642 } 5643 5644 bool CMSBitMap::allocate(MemRegion mr) { 5645 _bmStartWord = mr.start(); 5646 _bmWordSize = mr.word_size(); 5647 ReservedSpace brs(ReservedSpace::allocation_align_size_up( 5648 (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1)); 5649 if (!brs.is_reserved()) { 5650 warning("CMS bit map allocation failure"); 5651 return false; 5652 } 5653 // For now we'll just commit all of the bit map up front. 5654 // Later on we'll try to be more parsimonious with swap. 5655 if (!_virtual_space.initialize(brs, brs.size())) { 5656 warning("CMS bit map backing store failure"); 5657 return false; 5658 } 5659 assert(_virtual_space.committed_size() == brs.size(), 5660 "didn't reserve backing store for all of CMS bit map?"); 5661 _bm.set_map((BitMap::bm_word_t*)_virtual_space.low()); 5662 assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >= 5663 _bmWordSize, "inconsistency in bit map sizing"); 5664 _bm.set_size(_bmWordSize >> _shifter); 5665 5666 // bm.clear(); // can we rely on getting zero'd memory? verify below 5667 assert(isAllClear(), 5668 "Expected zero'd memory from ReservedSpace constructor"); 5669 assert(_bm.size() == heapWordDiffToOffsetDiff(sizeInWords()), 5670 "consistency check"); 5671 return true; 5672 } 5673 5674 void CMSBitMap::dirty_range_iterate_clear(MemRegion mr, MemRegionClosure* cl) { 5675 HeapWord *next_addr, *end_addr, *last_addr; 5676 assert_locked(); 5677 assert(covers(mr), "out-of-range error"); 5678 // XXX assert that start and end are appropriately aligned 5679 for (next_addr = mr.start(), end_addr = mr.end(); 5680 next_addr < end_addr; next_addr = last_addr) { 5681 MemRegion dirty_region = getAndClearMarkedRegion(next_addr, end_addr); 5682 last_addr = dirty_region.end(); 5683 if (!dirty_region.is_empty()) { 5684 cl->do_MemRegion(dirty_region); 5685 } else { 5686 assert(last_addr == end_addr, "program logic"); 5687 return; 5688 } 5689 } 5690 } 5691 5692 void CMSBitMap::print_on_error(outputStream* st, const char* prefix) const { 5693 _bm.print_on_error(st, prefix); 5694 } 5695 5696 #ifndef PRODUCT 5697 void CMSBitMap::assert_locked() const { 5698 CMSLockVerifier::assert_locked(lock()); 5699 } 5700 5701 bool CMSBitMap::covers(MemRegion mr) const { 5702 // assert(_bm.map() == _virtual_space.low(), "map inconsistency"); 5703 assert((size_t)_bm.size() == (_bmWordSize >> _shifter), 5704 "size inconsistency"); 5705 return (mr.start() >= _bmStartWord) && 5706 (mr.end() <= endWord()); 5707 } 5708 5709 bool CMSBitMap::covers(HeapWord* start, size_t size) const { 5710 return (start >= _bmStartWord && (start + size) <= endWord()); 5711 } 5712 5713 void CMSBitMap::verifyNoOneBitsInRange(HeapWord* left, HeapWord* right) { 5714 // verify that there are no 1 bits in the interval [left, right) 5715 FalseBitMapClosure falseBitMapClosure; 5716 iterate(&falseBitMapClosure, left, right); 5717 } 5718 5719 void CMSBitMap::region_invariant(MemRegion mr) 5720 { 5721 assert_locked(); 5722 // mr = mr.intersection(MemRegion(_bmStartWord, _bmWordSize)); 5723 assert(!mr.is_empty(), "unexpected empty region"); 5724 assert(covers(mr), "mr should be covered by bit map"); 5725 // convert address range into offset range 5726 size_t start_ofs = heapWordToOffset(mr.start()); 5727 // Make sure that end() is appropriately aligned 5728 assert(mr.end() == (HeapWord*)round_to((intptr_t)mr.end(), 5729 (1 << (_shifter+LogHeapWordSize))), 5730 "Misaligned mr.end()"); 5731 size_t end_ofs = heapWordToOffset(mr.end()); 5732 assert(end_ofs > start_ofs, "Should mark at least one bit"); 5733 } 5734 5735 #endif 5736 5737 bool CMSMarkStack::allocate(size_t size) { 5738 // allocate a stack of the requisite depth 5739 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 5740 size * sizeof(oop))); 5741 if (!rs.is_reserved()) { 5742 warning("CMSMarkStack allocation failure"); 5743 return false; 5744 } 5745 if (!_virtual_space.initialize(rs, rs.size())) { 5746 warning("CMSMarkStack backing store failure"); 5747 return false; 5748 } 5749 assert(_virtual_space.committed_size() == rs.size(), 5750 "didn't reserve backing store for all of CMS stack?"); 5751 _base = (oop*)(_virtual_space.low()); 5752 _index = 0; 5753 _capacity = size; 5754 NOT_PRODUCT(_max_depth = 0); 5755 return true; 5756 } 5757 5758 // XXX FIX ME !!! In the MT case we come in here holding a 5759 // leaf lock. For printing we need to take a further lock 5760 // which has lower rank. We need to recalibrate the two 5761 // lock-ranks involved in order to be able to print the 5762 // messages below. (Or defer the printing to the caller. 5763 // For now we take the expedient path of just disabling the 5764 // messages for the problematic case.) 5765 void CMSMarkStack::expand() { 5766 assert(_capacity <= MarkStackSizeMax, "stack bigger than permitted"); 5767 if (_capacity == MarkStackSizeMax) { 5768 if (_hit_limit++ == 0 && !CMSConcurrentMTEnabled) { 5769 // We print a warning message only once per CMS cycle. 5770 log_debug(gc)(" (benign) Hit CMSMarkStack max size limit"); 5771 } 5772 return; 5773 } 5774 // Double capacity if possible 5775 size_t new_capacity = MIN2(_capacity*2, MarkStackSizeMax); 5776 // Do not give up existing stack until we have managed to 5777 // get the double capacity that we desired. 5778 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 5779 new_capacity * sizeof(oop))); 5780 if (rs.is_reserved()) { 5781 // Release the backing store associated with old stack 5782 _virtual_space.release(); 5783 // Reinitialize virtual space for new stack 5784 if (!_virtual_space.initialize(rs, rs.size())) { 5785 fatal("Not enough swap for expanded marking stack"); 5786 } 5787 _base = (oop*)(_virtual_space.low()); 5788 _index = 0; 5789 _capacity = new_capacity; 5790 } else if (_failed_double++ == 0 && !CMSConcurrentMTEnabled) { 5791 // Failed to double capacity, continue; 5792 // we print a detail message only once per CMS cycle. 5793 log_debug(gc)(" (benign) Failed to expand marking stack from " SIZE_FORMAT "K to " SIZE_FORMAT "K", 5794 _capacity / K, new_capacity / K); 5795 } 5796 } 5797 5798 5799 // Closures 5800 // XXX: there seems to be a lot of code duplication here; 5801 // should refactor and consolidate common code. 5802 5803 // This closure is used to mark refs into the CMS generation in 5804 // the CMS bit map. Called at the first checkpoint. This closure 5805 // assumes that we do not need to re-mark dirty cards; if the CMS 5806 // generation on which this is used is not an oldest 5807 // generation then this will lose younger_gen cards! 5808 5809 MarkRefsIntoClosure::MarkRefsIntoClosure( 5810 MemRegion span, CMSBitMap* bitMap): 5811 _span(span), 5812 _bitMap(bitMap) 5813 { 5814 assert(ref_processor() == NULL, "deliberately left NULL"); 5815 assert(_bitMap->covers(_span), "_bitMap/_span mismatch"); 5816 } 5817 5818 void MarkRefsIntoClosure::do_oop(oop obj) { 5819 // if p points into _span, then mark corresponding bit in _markBitMap 5820 assert(obj->is_oop(), "expected an oop"); 5821 HeapWord* addr = (HeapWord*)obj; 5822 if (_span.contains(addr)) { 5823 // this should be made more efficient 5824 _bitMap->mark(addr); 5825 } 5826 } 5827 5828 void MarkRefsIntoClosure::do_oop(oop* p) { MarkRefsIntoClosure::do_oop_work(p); } 5829 void MarkRefsIntoClosure::do_oop(narrowOop* p) { MarkRefsIntoClosure::do_oop_work(p); } 5830 5831 Par_MarkRefsIntoClosure::Par_MarkRefsIntoClosure( 5832 MemRegion span, CMSBitMap* bitMap): 5833 _span(span), 5834 _bitMap(bitMap) 5835 { 5836 assert(ref_processor() == NULL, "deliberately left NULL"); 5837 assert(_bitMap->covers(_span), "_bitMap/_span mismatch"); 5838 } 5839 5840 void Par_MarkRefsIntoClosure::do_oop(oop obj) { 5841 // if p points into _span, then mark corresponding bit in _markBitMap 5842 assert(obj->is_oop(), "expected an oop"); 5843 HeapWord* addr = (HeapWord*)obj; 5844 if (_span.contains(addr)) { 5845 // this should be made more efficient 5846 _bitMap->par_mark(addr); 5847 } 5848 } 5849 5850 void Par_MarkRefsIntoClosure::do_oop(oop* p) { Par_MarkRefsIntoClosure::do_oop_work(p); } 5851 void Par_MarkRefsIntoClosure::do_oop(narrowOop* p) { Par_MarkRefsIntoClosure::do_oop_work(p); } 5852 5853 // A variant of the above, used for CMS marking verification. 5854 MarkRefsIntoVerifyClosure::MarkRefsIntoVerifyClosure( 5855 MemRegion span, CMSBitMap* verification_bm, CMSBitMap* cms_bm): 5856 _span(span), 5857 _verification_bm(verification_bm), 5858 _cms_bm(cms_bm) 5859 { 5860 assert(ref_processor() == NULL, "deliberately left NULL"); 5861 assert(_verification_bm->covers(_span), "_verification_bm/_span mismatch"); 5862 } 5863 5864 void MarkRefsIntoVerifyClosure::do_oop(oop obj) { 5865 // if p points into _span, then mark corresponding bit in _markBitMap 5866 assert(obj->is_oop(), "expected an oop"); 5867 HeapWord* addr = (HeapWord*)obj; 5868 if (_span.contains(addr)) { 5869 _verification_bm->mark(addr); 5870 if (!_cms_bm->isMarked(addr)) { 5871 LogHandle(gc, verify) log; 5872 ResourceMark rm; 5873 oop(addr)->print_on(log.info_stream()); 5874 log.info(" (" INTPTR_FORMAT " should have been marked)", p2i(addr)); 5875 fatal("... aborting"); 5876 } 5877 } 5878 } 5879 5880 void MarkRefsIntoVerifyClosure::do_oop(oop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); } 5881 void MarkRefsIntoVerifyClosure::do_oop(narrowOop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); } 5882 5883 ////////////////////////////////////////////////// 5884 // MarkRefsIntoAndScanClosure 5885 ////////////////////////////////////////////////// 5886 5887 MarkRefsIntoAndScanClosure::MarkRefsIntoAndScanClosure(MemRegion span, 5888 ReferenceProcessor* rp, 5889 CMSBitMap* bit_map, 5890 CMSBitMap* mod_union_table, 5891 CMSMarkStack* mark_stack, 5892 CMSCollector* collector, 5893 bool should_yield, 5894 bool concurrent_precleaning): 5895 _collector(collector), 5896 _span(span), 5897 _bit_map(bit_map), 5898 _mark_stack(mark_stack), 5899 _pushAndMarkClosure(collector, span, rp, bit_map, mod_union_table, 5900 mark_stack, concurrent_precleaning), 5901 _yield(should_yield), 5902 _concurrent_precleaning(concurrent_precleaning), 5903 _freelistLock(NULL) 5904 { 5905 // FIXME: Should initialize in base class constructor. 5906 assert(rp != NULL, "ref_processor shouldn't be NULL"); 5907 set_ref_processor_internal(rp); 5908 } 5909 5910 // This closure is used to mark refs into the CMS generation at the 5911 // second (final) checkpoint, and to scan and transitively follow 5912 // the unmarked oops. It is also used during the concurrent precleaning 5913 // phase while scanning objects on dirty cards in the CMS generation. 5914 // The marks are made in the marking bit map and the marking stack is 5915 // used for keeping the (newly) grey objects during the scan. 5916 // The parallel version (Par_...) appears further below. 5917 void MarkRefsIntoAndScanClosure::do_oop(oop obj) { 5918 if (obj != NULL) { 5919 assert(obj->is_oop(), "expected an oop"); 5920 HeapWord* addr = (HeapWord*)obj; 5921 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 5922 assert(_collector->overflow_list_is_empty(), 5923 "overflow list should be empty"); 5924 if (_span.contains(addr) && 5925 !_bit_map->isMarked(addr)) { 5926 // mark bit map (object is now grey) 5927 _bit_map->mark(addr); 5928 // push on marking stack (stack should be empty), and drain the 5929 // stack by applying this closure to the oops in the oops popped 5930 // from the stack (i.e. blacken the grey objects) 5931 bool res = _mark_stack->push(obj); 5932 assert(res, "Should have space to push on empty stack"); 5933 do { 5934 oop new_oop = _mark_stack->pop(); 5935 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 5936 assert(_bit_map->isMarked((HeapWord*)new_oop), 5937 "only grey objects on this stack"); 5938 // iterate over the oops in this oop, marking and pushing 5939 // the ones in CMS heap (i.e. in _span). 5940 new_oop->oop_iterate(&_pushAndMarkClosure); 5941 // check if it's time to yield 5942 do_yield_check(); 5943 } while (!_mark_stack->isEmpty() || 5944 (!_concurrent_precleaning && take_from_overflow_list())); 5945 // if marking stack is empty, and we are not doing this 5946 // during precleaning, then check the overflow list 5947 } 5948 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 5949 assert(_collector->overflow_list_is_empty(), 5950 "overflow list was drained above"); 5951 5952 assert(_collector->no_preserved_marks(), 5953 "All preserved marks should have been restored above"); 5954 } 5955 } 5956 5957 void MarkRefsIntoAndScanClosure::do_oop(oop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); } 5958 void MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); } 5959 5960 void MarkRefsIntoAndScanClosure::do_yield_work() { 5961 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 5962 "CMS thread should hold CMS token"); 5963 assert_lock_strong(_freelistLock); 5964 assert_lock_strong(_bit_map->lock()); 5965 // relinquish the free_list_lock and bitMaplock() 5966 _bit_map->lock()->unlock(); 5967 _freelistLock->unlock(); 5968 ConcurrentMarkSweepThread::desynchronize(true); 5969 _collector->stopTimer(); 5970 _collector->incrementYields(); 5971 5972 // See the comment in coordinator_yield() 5973 for (unsigned i = 0; 5974 i < CMSYieldSleepCount && 5975 ConcurrentMarkSweepThread::should_yield() && 5976 !CMSCollector::foregroundGCIsActive(); 5977 ++i) { 5978 os::sleep(Thread::current(), 1, false); 5979 } 5980 5981 ConcurrentMarkSweepThread::synchronize(true); 5982 _freelistLock->lock_without_safepoint_check(); 5983 _bit_map->lock()->lock_without_safepoint_check(); 5984 _collector->startTimer(); 5985 } 5986 5987 /////////////////////////////////////////////////////////// 5988 // Par_MarkRefsIntoAndScanClosure: a parallel version of 5989 // MarkRefsIntoAndScanClosure 5990 /////////////////////////////////////////////////////////// 5991 Par_MarkRefsIntoAndScanClosure::Par_MarkRefsIntoAndScanClosure( 5992 CMSCollector* collector, MemRegion span, ReferenceProcessor* rp, 5993 CMSBitMap* bit_map, OopTaskQueue* work_queue): 5994 _span(span), 5995 _bit_map(bit_map), 5996 _work_queue(work_queue), 5997 _low_water_mark(MIN2((work_queue->max_elems()/4), 5998 ((uint)CMSWorkQueueDrainThreshold * ParallelGCThreads))), 5999 _par_pushAndMarkClosure(collector, span, rp, bit_map, work_queue) 6000 { 6001 // FIXME: Should initialize in base class constructor. 6002 assert(rp != NULL, "ref_processor shouldn't be NULL"); 6003 set_ref_processor_internal(rp); 6004 } 6005 6006 // This closure is used to mark refs into the CMS generation at the 6007 // second (final) checkpoint, and to scan and transitively follow 6008 // the unmarked oops. The marks are made in the marking bit map and 6009 // the work_queue is used for keeping the (newly) grey objects during 6010 // the scan phase whence they are also available for stealing by parallel 6011 // threads. Since the marking bit map is shared, updates are 6012 // synchronized (via CAS). 6013 void Par_MarkRefsIntoAndScanClosure::do_oop(oop obj) { 6014 if (obj != NULL) { 6015 // Ignore mark word because this could be an already marked oop 6016 // that may be chained at the end of the overflow list. 6017 assert(obj->is_oop(true), "expected an oop"); 6018 HeapWord* addr = (HeapWord*)obj; 6019 if (_span.contains(addr) && 6020 !_bit_map->isMarked(addr)) { 6021 // mark bit map (object will become grey): 6022 // It is possible for several threads to be 6023 // trying to "claim" this object concurrently; 6024 // the unique thread that succeeds in marking the 6025 // object first will do the subsequent push on 6026 // to the work queue (or overflow list). 6027 if (_bit_map->par_mark(addr)) { 6028 // push on work_queue (which may not be empty), and trim the 6029 // queue to an appropriate length by applying this closure to 6030 // the oops in the oops popped from the stack (i.e. blacken the 6031 // grey objects) 6032 bool res = _work_queue->push(obj); 6033 assert(res, "Low water mark should be less than capacity?"); 6034 trim_queue(_low_water_mark); 6035 } // Else, another thread claimed the object 6036 } 6037 } 6038 } 6039 6040 void Par_MarkRefsIntoAndScanClosure::do_oop(oop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); } 6041 void Par_MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); } 6042 6043 // This closure is used to rescan the marked objects on the dirty cards 6044 // in the mod union table and the card table proper. 6045 size_t ScanMarkedObjectsAgainCarefullyClosure::do_object_careful_m( 6046 oop p, MemRegion mr) { 6047 6048 size_t size = 0; 6049 HeapWord* addr = (HeapWord*)p; 6050 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6051 assert(_span.contains(addr), "we are scanning the CMS generation"); 6052 // check if it's time to yield 6053 if (do_yield_check()) { 6054 // We yielded for some foreground stop-world work, 6055 // and we have been asked to abort this ongoing preclean cycle. 6056 return 0; 6057 } 6058 if (_bitMap->isMarked(addr)) { 6059 // it's marked; is it potentially uninitialized? 6060 if (p->klass_or_null() != NULL) { 6061 // an initialized object; ignore mark word in verification below 6062 // since we are running concurrent with mutators 6063 assert(p->is_oop(true), "should be an oop"); 6064 if (p->is_objArray()) { 6065 // objArrays are precisely marked; restrict scanning 6066 // to dirty cards only. 6067 size = CompactibleFreeListSpace::adjustObjectSize( 6068 p->oop_iterate_size(_scanningClosure, mr)); 6069 } else { 6070 // A non-array may have been imprecisely marked; we need 6071 // to scan object in its entirety. 6072 size = CompactibleFreeListSpace::adjustObjectSize( 6073 p->oop_iterate_size(_scanningClosure)); 6074 } 6075 #ifdef ASSERT 6076 size_t direct_size = 6077 CompactibleFreeListSpace::adjustObjectSize(p->size()); 6078 assert(size == direct_size, "Inconsistency in size"); 6079 assert(size >= 3, "Necessary for Printezis marks to work"); 6080 if (!_bitMap->isMarked(addr+1)) { 6081 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size); 6082 } else { 6083 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size-1); 6084 assert(_bitMap->isMarked(addr+size-1), 6085 "inconsistent Printezis mark"); 6086 } 6087 #endif // ASSERT 6088 } else { 6089 // An uninitialized object. 6090 assert(_bitMap->isMarked(addr+1), "missing Printezis mark?"); 6091 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 6092 size = pointer_delta(nextOneAddr + 1, addr); 6093 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 6094 "alignment problem"); 6095 // Note that pre-cleaning needn't redirty the card. OopDesc::set_klass() 6096 // will dirty the card when the klass pointer is installed in the 6097 // object (signaling the completion of initialization). 6098 } 6099 } else { 6100 // Either a not yet marked object or an uninitialized object 6101 if (p->klass_or_null() == NULL) { 6102 // An uninitialized object, skip to the next card, since 6103 // we may not be able to read its P-bits yet. 6104 assert(size == 0, "Initial value"); 6105 } else { 6106 // An object not (yet) reached by marking: we merely need to 6107 // compute its size so as to go look at the next block. 6108 assert(p->is_oop(true), "should be an oop"); 6109 size = CompactibleFreeListSpace::adjustObjectSize(p->size()); 6110 } 6111 } 6112 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6113 return size; 6114 } 6115 6116 void ScanMarkedObjectsAgainCarefullyClosure::do_yield_work() { 6117 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6118 "CMS thread should hold CMS token"); 6119 assert_lock_strong(_freelistLock); 6120 assert_lock_strong(_bitMap->lock()); 6121 // relinquish the free_list_lock and bitMaplock() 6122 _bitMap->lock()->unlock(); 6123 _freelistLock->unlock(); 6124 ConcurrentMarkSweepThread::desynchronize(true); 6125 _collector->stopTimer(); 6126 _collector->incrementYields(); 6127 6128 // See the comment in coordinator_yield() 6129 for (unsigned i = 0; i < CMSYieldSleepCount && 6130 ConcurrentMarkSweepThread::should_yield() && 6131 !CMSCollector::foregroundGCIsActive(); ++i) { 6132 os::sleep(Thread::current(), 1, false); 6133 } 6134 6135 ConcurrentMarkSweepThread::synchronize(true); 6136 _freelistLock->lock_without_safepoint_check(); 6137 _bitMap->lock()->lock_without_safepoint_check(); 6138 _collector->startTimer(); 6139 } 6140 6141 6142 ////////////////////////////////////////////////////////////////// 6143 // SurvivorSpacePrecleanClosure 6144 ////////////////////////////////////////////////////////////////// 6145 // This (single-threaded) closure is used to preclean the oops in 6146 // the survivor spaces. 6147 size_t SurvivorSpacePrecleanClosure::do_object_careful(oop p) { 6148 6149 HeapWord* addr = (HeapWord*)p; 6150 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6151 assert(!_span.contains(addr), "we are scanning the survivor spaces"); 6152 assert(p->klass_or_null() != NULL, "object should be initialized"); 6153 // an initialized object; ignore mark word in verification below 6154 // since we are running concurrent with mutators 6155 assert(p->is_oop(true), "should be an oop"); 6156 // Note that we do not yield while we iterate over 6157 // the interior oops of p, pushing the relevant ones 6158 // on our marking stack. 6159 size_t size = p->oop_iterate_size(_scanning_closure); 6160 do_yield_check(); 6161 // Observe that below, we do not abandon the preclean 6162 // phase as soon as we should; rather we empty the 6163 // marking stack before returning. This is to satisfy 6164 // some existing assertions. In general, it may be a 6165 // good idea to abort immediately and complete the marking 6166 // from the grey objects at a later time. 6167 while (!_mark_stack->isEmpty()) { 6168 oop new_oop = _mark_stack->pop(); 6169 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 6170 assert(_bit_map->isMarked((HeapWord*)new_oop), 6171 "only grey objects on this stack"); 6172 // iterate over the oops in this oop, marking and pushing 6173 // the ones in CMS heap (i.e. in _span). 6174 new_oop->oop_iterate(_scanning_closure); 6175 // check if it's time to yield 6176 do_yield_check(); 6177 } 6178 unsigned int after_count = 6179 GenCollectedHeap::heap()->total_collections(); 6180 bool abort = (_before_count != after_count) || 6181 _collector->should_abort_preclean(); 6182 return abort ? 0 : size; 6183 } 6184 6185 void SurvivorSpacePrecleanClosure::do_yield_work() { 6186 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6187 "CMS thread should hold CMS token"); 6188 assert_lock_strong(_bit_map->lock()); 6189 // Relinquish the bit map lock 6190 _bit_map->lock()->unlock(); 6191 ConcurrentMarkSweepThread::desynchronize(true); 6192 _collector->stopTimer(); 6193 _collector->incrementYields(); 6194 6195 // See the comment in coordinator_yield() 6196 for (unsigned i = 0; i < CMSYieldSleepCount && 6197 ConcurrentMarkSweepThread::should_yield() && 6198 !CMSCollector::foregroundGCIsActive(); ++i) { 6199 os::sleep(Thread::current(), 1, false); 6200 } 6201 6202 ConcurrentMarkSweepThread::synchronize(true); 6203 _bit_map->lock()->lock_without_safepoint_check(); 6204 _collector->startTimer(); 6205 } 6206 6207 // This closure is used to rescan the marked objects on the dirty cards 6208 // in the mod union table and the card table proper. In the parallel 6209 // case, although the bitMap is shared, we do a single read so the 6210 // isMarked() query is "safe". 6211 bool ScanMarkedObjectsAgainClosure::do_object_bm(oop p, MemRegion mr) { 6212 // Ignore mark word because we are running concurrent with mutators 6213 assert(p->is_oop_or_null(true), "Expected an oop or NULL at " PTR_FORMAT, p2i(p)); 6214 HeapWord* addr = (HeapWord*)p; 6215 assert(_span.contains(addr), "we are scanning the CMS generation"); 6216 bool is_obj_array = false; 6217 #ifdef ASSERT 6218 if (!_parallel) { 6219 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 6220 assert(_collector->overflow_list_is_empty(), 6221 "overflow list should be empty"); 6222 6223 } 6224 #endif // ASSERT 6225 if (_bit_map->isMarked(addr)) { 6226 // Obj arrays are precisely marked, non-arrays are not; 6227 // so we scan objArrays precisely and non-arrays in their 6228 // entirety. 6229 if (p->is_objArray()) { 6230 is_obj_array = true; 6231 if (_parallel) { 6232 p->oop_iterate(_par_scan_closure, mr); 6233 } else { 6234 p->oop_iterate(_scan_closure, mr); 6235 } 6236 } else { 6237 if (_parallel) { 6238 p->oop_iterate(_par_scan_closure); 6239 } else { 6240 p->oop_iterate(_scan_closure); 6241 } 6242 } 6243 } 6244 #ifdef ASSERT 6245 if (!_parallel) { 6246 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 6247 assert(_collector->overflow_list_is_empty(), 6248 "overflow list should be empty"); 6249 6250 } 6251 #endif // ASSERT 6252 return is_obj_array; 6253 } 6254 6255 MarkFromRootsClosure::MarkFromRootsClosure(CMSCollector* collector, 6256 MemRegion span, 6257 CMSBitMap* bitMap, CMSMarkStack* markStack, 6258 bool should_yield, bool verifying): 6259 _collector(collector), 6260 _span(span), 6261 _bitMap(bitMap), 6262 _mut(&collector->_modUnionTable), 6263 _markStack(markStack), 6264 _yield(should_yield), 6265 _skipBits(0) 6266 { 6267 assert(_markStack->isEmpty(), "stack should be empty"); 6268 _finger = _bitMap->startWord(); 6269 _threshold = _finger; 6270 assert(_collector->_restart_addr == NULL, "Sanity check"); 6271 assert(_span.contains(_finger), "Out of bounds _finger?"); 6272 DEBUG_ONLY(_verifying = verifying;) 6273 } 6274 6275 void MarkFromRootsClosure::reset(HeapWord* addr) { 6276 assert(_markStack->isEmpty(), "would cause duplicates on stack"); 6277 assert(_span.contains(addr), "Out of bounds _finger?"); 6278 _finger = addr; 6279 _threshold = (HeapWord*)round_to( 6280 (intptr_t)_finger, CardTableModRefBS::card_size); 6281 } 6282 6283 // Should revisit to see if this should be restructured for 6284 // greater efficiency. 6285 bool MarkFromRootsClosure::do_bit(size_t offset) { 6286 if (_skipBits > 0) { 6287 _skipBits--; 6288 return true; 6289 } 6290 // convert offset into a HeapWord* 6291 HeapWord* addr = _bitMap->startWord() + offset; 6292 assert(_bitMap->endWord() && addr < _bitMap->endWord(), 6293 "address out of range"); 6294 assert(_bitMap->isMarked(addr), "tautology"); 6295 if (_bitMap->isMarked(addr+1)) { 6296 // this is an allocated but not yet initialized object 6297 assert(_skipBits == 0, "tautology"); 6298 _skipBits = 2; // skip next two marked bits ("Printezis-marks") 6299 oop p = oop(addr); 6300 if (p->klass_or_null() == NULL) { 6301 DEBUG_ONLY(if (!_verifying) {) 6302 // We re-dirty the cards on which this object lies and increase 6303 // the _threshold so that we'll come back to scan this object 6304 // during the preclean or remark phase. (CMSCleanOnEnter) 6305 if (CMSCleanOnEnter) { 6306 size_t sz = _collector->block_size_using_printezis_bits(addr); 6307 HeapWord* end_card_addr = (HeapWord*)round_to( 6308 (intptr_t)(addr+sz), CardTableModRefBS::card_size); 6309 MemRegion redirty_range = MemRegion(addr, end_card_addr); 6310 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 6311 // Bump _threshold to end_card_addr; note that 6312 // _threshold cannot possibly exceed end_card_addr, anyhow. 6313 // This prevents future clearing of the card as the scan proceeds 6314 // to the right. 6315 assert(_threshold <= end_card_addr, 6316 "Because we are just scanning into this object"); 6317 if (_threshold < end_card_addr) { 6318 _threshold = end_card_addr; 6319 } 6320 if (p->klass_or_null() != NULL) { 6321 // Redirty the range of cards... 6322 _mut->mark_range(redirty_range); 6323 } // ...else the setting of klass will dirty the card anyway. 6324 } 6325 DEBUG_ONLY(}) 6326 return true; 6327 } 6328 } 6329 scanOopsInOop(addr); 6330 return true; 6331 } 6332 6333 // We take a break if we've been at this for a while, 6334 // so as to avoid monopolizing the locks involved. 6335 void MarkFromRootsClosure::do_yield_work() { 6336 // First give up the locks, then yield, then re-lock 6337 // We should probably use a constructor/destructor idiom to 6338 // do this unlock/lock or modify the MutexUnlocker class to 6339 // serve our purpose. XXX 6340 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6341 "CMS thread should hold CMS token"); 6342 assert_lock_strong(_bitMap->lock()); 6343 _bitMap->lock()->unlock(); 6344 ConcurrentMarkSweepThread::desynchronize(true); 6345 _collector->stopTimer(); 6346 _collector->incrementYields(); 6347 6348 // See the comment in coordinator_yield() 6349 for (unsigned i = 0; i < CMSYieldSleepCount && 6350 ConcurrentMarkSweepThread::should_yield() && 6351 !CMSCollector::foregroundGCIsActive(); ++i) { 6352 os::sleep(Thread::current(), 1, false); 6353 } 6354 6355 ConcurrentMarkSweepThread::synchronize(true); 6356 _bitMap->lock()->lock_without_safepoint_check(); 6357 _collector->startTimer(); 6358 } 6359 6360 void MarkFromRootsClosure::scanOopsInOop(HeapWord* ptr) { 6361 assert(_bitMap->isMarked(ptr), "expected bit to be set"); 6362 assert(_markStack->isEmpty(), 6363 "should drain stack to limit stack usage"); 6364 // convert ptr to an oop preparatory to scanning 6365 oop obj = oop(ptr); 6366 // Ignore mark word in verification below, since we 6367 // may be running concurrent with mutators. 6368 assert(obj->is_oop(true), "should be an oop"); 6369 assert(_finger <= ptr, "_finger runneth ahead"); 6370 // advance the finger to right end of this object 6371 _finger = ptr + obj->size(); 6372 assert(_finger > ptr, "we just incremented it above"); 6373 // On large heaps, it may take us some time to get through 6374 // the marking phase. During 6375 // this time it's possible that a lot of mutations have 6376 // accumulated in the card table and the mod union table -- 6377 // these mutation records are redundant until we have 6378 // actually traced into the corresponding card. 6379 // Here, we check whether advancing the finger would make 6380 // us cross into a new card, and if so clear corresponding 6381 // cards in the MUT (preclean them in the card-table in the 6382 // future). 6383 6384 DEBUG_ONLY(if (!_verifying) {) 6385 // The clean-on-enter optimization is disabled by default, 6386 // until we fix 6178663. 6387 if (CMSCleanOnEnter && (_finger > _threshold)) { 6388 // [_threshold, _finger) represents the interval 6389 // of cards to be cleared in MUT (or precleaned in card table). 6390 // The set of cards to be cleared is all those that overlap 6391 // with the interval [_threshold, _finger); note that 6392 // _threshold is always kept card-aligned but _finger isn't 6393 // always card-aligned. 6394 HeapWord* old_threshold = _threshold; 6395 assert(old_threshold == (HeapWord*)round_to( 6396 (intptr_t)old_threshold, CardTableModRefBS::card_size), 6397 "_threshold should always be card-aligned"); 6398 _threshold = (HeapWord*)round_to( 6399 (intptr_t)_finger, CardTableModRefBS::card_size); 6400 MemRegion mr(old_threshold, _threshold); 6401 assert(!mr.is_empty(), "Control point invariant"); 6402 assert(_span.contains(mr), "Should clear within span"); 6403 _mut->clear_range(mr); 6404 } 6405 DEBUG_ONLY(}) 6406 // Note: the finger doesn't advance while we drain 6407 // the stack below. 6408 PushOrMarkClosure pushOrMarkClosure(_collector, 6409 _span, _bitMap, _markStack, 6410 _finger, this); 6411 bool res = _markStack->push(obj); 6412 assert(res, "Empty non-zero size stack should have space for single push"); 6413 while (!_markStack->isEmpty()) { 6414 oop new_oop = _markStack->pop(); 6415 // Skip verifying header mark word below because we are 6416 // running concurrent with mutators. 6417 assert(new_oop->is_oop(true), "Oops! expected to pop an oop"); 6418 // now scan this oop's oops 6419 new_oop->oop_iterate(&pushOrMarkClosure); 6420 do_yield_check(); 6421 } 6422 assert(_markStack->isEmpty(), "tautology, emphasizing post-condition"); 6423 } 6424 6425 Par_MarkFromRootsClosure::Par_MarkFromRootsClosure(CMSConcMarkingTask* task, 6426 CMSCollector* collector, MemRegion span, 6427 CMSBitMap* bit_map, 6428 OopTaskQueue* work_queue, 6429 CMSMarkStack* overflow_stack): 6430 _collector(collector), 6431 _whole_span(collector->_span), 6432 _span(span), 6433 _bit_map(bit_map), 6434 _mut(&collector->_modUnionTable), 6435 _work_queue(work_queue), 6436 _overflow_stack(overflow_stack), 6437 _skip_bits(0), 6438 _task(task) 6439 { 6440 assert(_work_queue->size() == 0, "work_queue should be empty"); 6441 _finger = span.start(); 6442 _threshold = _finger; // XXX Defer clear-on-enter optimization for now 6443 assert(_span.contains(_finger), "Out of bounds _finger?"); 6444 } 6445 6446 // Should revisit to see if this should be restructured for 6447 // greater efficiency. 6448 bool Par_MarkFromRootsClosure::do_bit(size_t offset) { 6449 if (_skip_bits > 0) { 6450 _skip_bits--; 6451 return true; 6452 } 6453 // convert offset into a HeapWord* 6454 HeapWord* addr = _bit_map->startWord() + offset; 6455 assert(_bit_map->endWord() && addr < _bit_map->endWord(), 6456 "address out of range"); 6457 assert(_bit_map->isMarked(addr), "tautology"); 6458 if (_bit_map->isMarked(addr+1)) { 6459 // this is an allocated object that might not yet be initialized 6460 assert(_skip_bits == 0, "tautology"); 6461 _skip_bits = 2; // skip next two marked bits ("Printezis-marks") 6462 oop p = oop(addr); 6463 if (p->klass_or_null() == NULL) { 6464 // in the case of Clean-on-Enter optimization, redirty card 6465 // and avoid clearing card by increasing the threshold. 6466 return true; 6467 } 6468 } 6469 scan_oops_in_oop(addr); 6470 return true; 6471 } 6472 6473 void Par_MarkFromRootsClosure::scan_oops_in_oop(HeapWord* ptr) { 6474 assert(_bit_map->isMarked(ptr), "expected bit to be set"); 6475 // Should we assert that our work queue is empty or 6476 // below some drain limit? 6477 assert(_work_queue->size() == 0, 6478 "should drain stack to limit stack usage"); 6479 // convert ptr to an oop preparatory to scanning 6480 oop obj = oop(ptr); 6481 // Ignore mark word in verification below, since we 6482 // may be running concurrent with mutators. 6483 assert(obj->is_oop(true), "should be an oop"); 6484 assert(_finger <= ptr, "_finger runneth ahead"); 6485 // advance the finger to right end of this object 6486 _finger = ptr + obj->size(); 6487 assert(_finger > ptr, "we just incremented it above"); 6488 // On large heaps, it may take us some time to get through 6489 // the marking phase. During 6490 // this time it's possible that a lot of mutations have 6491 // accumulated in the card table and the mod union table -- 6492 // these mutation records are redundant until we have 6493 // actually traced into the corresponding card. 6494 // Here, we check whether advancing the finger would make 6495 // us cross into a new card, and if so clear corresponding 6496 // cards in the MUT (preclean them in the card-table in the 6497 // future). 6498 6499 // The clean-on-enter optimization is disabled by default, 6500 // until we fix 6178663. 6501 if (CMSCleanOnEnter && (_finger > _threshold)) { 6502 // [_threshold, _finger) represents the interval 6503 // of cards to be cleared in MUT (or precleaned in card table). 6504 // The set of cards to be cleared is all those that overlap 6505 // with the interval [_threshold, _finger); note that 6506 // _threshold is always kept card-aligned but _finger isn't 6507 // always card-aligned. 6508 HeapWord* old_threshold = _threshold; 6509 assert(old_threshold == (HeapWord*)round_to( 6510 (intptr_t)old_threshold, CardTableModRefBS::card_size), 6511 "_threshold should always be card-aligned"); 6512 _threshold = (HeapWord*)round_to( 6513 (intptr_t)_finger, CardTableModRefBS::card_size); 6514 MemRegion mr(old_threshold, _threshold); 6515 assert(!mr.is_empty(), "Control point invariant"); 6516 assert(_span.contains(mr), "Should clear within span"); // _whole_span ?? 6517 _mut->clear_range(mr); 6518 } 6519 6520 // Note: the local finger doesn't advance while we drain 6521 // the stack below, but the global finger sure can and will. 6522 HeapWord** gfa = _task->global_finger_addr(); 6523 Par_PushOrMarkClosure pushOrMarkClosure(_collector, 6524 _span, _bit_map, 6525 _work_queue, 6526 _overflow_stack, 6527 _finger, 6528 gfa, this); 6529 bool res = _work_queue->push(obj); // overflow could occur here 6530 assert(res, "Will hold once we use workqueues"); 6531 while (true) { 6532 oop new_oop; 6533 if (!_work_queue->pop_local(new_oop)) { 6534 // We emptied our work_queue; check if there's stuff that can 6535 // be gotten from the overflow stack. 6536 if (CMSConcMarkingTask::get_work_from_overflow_stack( 6537 _overflow_stack, _work_queue)) { 6538 do_yield_check(); 6539 continue; 6540 } else { // done 6541 break; 6542 } 6543 } 6544 // Skip verifying header mark word below because we are 6545 // running concurrent with mutators. 6546 assert(new_oop->is_oop(true), "Oops! expected to pop an oop"); 6547 // now scan this oop's oops 6548 new_oop->oop_iterate(&pushOrMarkClosure); 6549 do_yield_check(); 6550 } 6551 assert(_work_queue->size() == 0, "tautology, emphasizing post-condition"); 6552 } 6553 6554 // Yield in response to a request from VM Thread or 6555 // from mutators. 6556 void Par_MarkFromRootsClosure::do_yield_work() { 6557 assert(_task != NULL, "sanity"); 6558 _task->yield(); 6559 } 6560 6561 // A variant of the above used for verifying CMS marking work. 6562 MarkFromRootsVerifyClosure::MarkFromRootsVerifyClosure(CMSCollector* collector, 6563 MemRegion span, 6564 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 6565 CMSMarkStack* mark_stack): 6566 _collector(collector), 6567 _span(span), 6568 _verification_bm(verification_bm), 6569 _cms_bm(cms_bm), 6570 _mark_stack(mark_stack), 6571 _pam_verify_closure(collector, span, verification_bm, cms_bm, 6572 mark_stack) 6573 { 6574 assert(_mark_stack->isEmpty(), "stack should be empty"); 6575 _finger = _verification_bm->startWord(); 6576 assert(_collector->_restart_addr == NULL, "Sanity check"); 6577 assert(_span.contains(_finger), "Out of bounds _finger?"); 6578 } 6579 6580 void MarkFromRootsVerifyClosure::reset(HeapWord* addr) { 6581 assert(_mark_stack->isEmpty(), "would cause duplicates on stack"); 6582 assert(_span.contains(addr), "Out of bounds _finger?"); 6583 _finger = addr; 6584 } 6585 6586 // Should revisit to see if this should be restructured for 6587 // greater efficiency. 6588 bool MarkFromRootsVerifyClosure::do_bit(size_t offset) { 6589 // convert offset into a HeapWord* 6590 HeapWord* addr = _verification_bm->startWord() + offset; 6591 assert(_verification_bm->endWord() && addr < _verification_bm->endWord(), 6592 "address out of range"); 6593 assert(_verification_bm->isMarked(addr), "tautology"); 6594 assert(_cms_bm->isMarked(addr), "tautology"); 6595 6596 assert(_mark_stack->isEmpty(), 6597 "should drain stack to limit stack usage"); 6598 // convert addr to an oop preparatory to scanning 6599 oop obj = oop(addr); 6600 assert(obj->is_oop(), "should be an oop"); 6601 assert(_finger <= addr, "_finger runneth ahead"); 6602 // advance the finger to right end of this object 6603 _finger = addr + obj->size(); 6604 assert(_finger > addr, "we just incremented it above"); 6605 // Note: the finger doesn't advance while we drain 6606 // the stack below. 6607 bool res = _mark_stack->push(obj); 6608 assert(res, "Empty non-zero size stack should have space for single push"); 6609 while (!_mark_stack->isEmpty()) { 6610 oop new_oop = _mark_stack->pop(); 6611 assert(new_oop->is_oop(), "Oops! expected to pop an oop"); 6612 // now scan this oop's oops 6613 new_oop->oop_iterate(&_pam_verify_closure); 6614 } 6615 assert(_mark_stack->isEmpty(), "tautology, emphasizing post-condition"); 6616 return true; 6617 } 6618 6619 PushAndMarkVerifyClosure::PushAndMarkVerifyClosure( 6620 CMSCollector* collector, MemRegion span, 6621 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 6622 CMSMarkStack* mark_stack): 6623 MetadataAwareOopClosure(collector->ref_processor()), 6624 _collector(collector), 6625 _span(span), 6626 _verification_bm(verification_bm), 6627 _cms_bm(cms_bm), 6628 _mark_stack(mark_stack) 6629 { } 6630 6631 void PushAndMarkVerifyClosure::do_oop(oop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 6632 void PushAndMarkVerifyClosure::do_oop(narrowOop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 6633 6634 // Upon stack overflow, we discard (part of) the stack, 6635 // remembering the least address amongst those discarded 6636 // in CMSCollector's _restart_address. 6637 void PushAndMarkVerifyClosure::handle_stack_overflow(HeapWord* lost) { 6638 // Remember the least grey address discarded 6639 HeapWord* ra = (HeapWord*)_mark_stack->least_value(lost); 6640 _collector->lower_restart_addr(ra); 6641 _mark_stack->reset(); // discard stack contents 6642 _mark_stack->expand(); // expand the stack if possible 6643 } 6644 6645 void PushAndMarkVerifyClosure::do_oop(oop obj) { 6646 assert(obj->is_oop_or_null(), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6647 HeapWord* addr = (HeapWord*)obj; 6648 if (_span.contains(addr) && !_verification_bm->isMarked(addr)) { 6649 // Oop lies in _span and isn't yet grey or black 6650 _verification_bm->mark(addr); // now grey 6651 if (!_cms_bm->isMarked(addr)) { 6652 LogHandle(gc, verify) log; 6653 ResourceMark rm; 6654 oop(addr)->print_on(log.info_stream()); 6655 log.info(" (" INTPTR_FORMAT " should have been marked)", p2i(addr)); 6656 fatal("... aborting"); 6657 } 6658 6659 if (!_mark_stack->push(obj)) { // stack overflow 6660 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _mark_stack->capacity()); 6661 assert(_mark_stack->isFull(), "Else push should have succeeded"); 6662 handle_stack_overflow(addr); 6663 } 6664 // anything including and to the right of _finger 6665 // will be scanned as we iterate over the remainder of the 6666 // bit map 6667 } 6668 } 6669 6670 PushOrMarkClosure::PushOrMarkClosure(CMSCollector* collector, 6671 MemRegion span, 6672 CMSBitMap* bitMap, CMSMarkStack* markStack, 6673 HeapWord* finger, MarkFromRootsClosure* parent) : 6674 MetadataAwareOopClosure(collector->ref_processor()), 6675 _collector(collector), 6676 _span(span), 6677 _bitMap(bitMap), 6678 _markStack(markStack), 6679 _finger(finger), 6680 _parent(parent) 6681 { } 6682 6683 Par_PushOrMarkClosure::Par_PushOrMarkClosure(CMSCollector* collector, 6684 MemRegion span, 6685 CMSBitMap* bit_map, 6686 OopTaskQueue* work_queue, 6687 CMSMarkStack* overflow_stack, 6688 HeapWord* finger, 6689 HeapWord** global_finger_addr, 6690 Par_MarkFromRootsClosure* parent) : 6691 MetadataAwareOopClosure(collector->ref_processor()), 6692 _collector(collector), 6693 _whole_span(collector->_span), 6694 _span(span), 6695 _bit_map(bit_map), 6696 _work_queue(work_queue), 6697 _overflow_stack(overflow_stack), 6698 _finger(finger), 6699 _global_finger_addr(global_finger_addr), 6700 _parent(parent) 6701 { } 6702 6703 // Assumes thread-safe access by callers, who are 6704 // responsible for mutual exclusion. 6705 void CMSCollector::lower_restart_addr(HeapWord* low) { 6706 assert(_span.contains(low), "Out of bounds addr"); 6707 if (_restart_addr == NULL) { 6708 _restart_addr = low; 6709 } else { 6710 _restart_addr = MIN2(_restart_addr, low); 6711 } 6712 } 6713 6714 // Upon stack overflow, we discard (part of) the stack, 6715 // remembering the least address amongst those discarded 6716 // in CMSCollector's _restart_address. 6717 void PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 6718 // Remember the least grey address discarded 6719 HeapWord* ra = (HeapWord*)_markStack->least_value(lost); 6720 _collector->lower_restart_addr(ra); 6721 _markStack->reset(); // discard stack contents 6722 _markStack->expand(); // expand the stack if possible 6723 } 6724 6725 // Upon stack overflow, we discard (part of) the stack, 6726 // remembering the least address amongst those discarded 6727 // in CMSCollector's _restart_address. 6728 void Par_PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 6729 // We need to do this under a mutex to prevent other 6730 // workers from interfering with the work done below. 6731 MutexLockerEx ml(_overflow_stack->par_lock(), 6732 Mutex::_no_safepoint_check_flag); 6733 // Remember the least grey address discarded 6734 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 6735 _collector->lower_restart_addr(ra); 6736 _overflow_stack->reset(); // discard stack contents 6737 _overflow_stack->expand(); // expand the stack if possible 6738 } 6739 6740 void PushOrMarkClosure::do_oop(oop obj) { 6741 // Ignore mark word because we are running concurrent with mutators. 6742 assert(obj->is_oop_or_null(true), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6743 HeapWord* addr = (HeapWord*)obj; 6744 if (_span.contains(addr) && !_bitMap->isMarked(addr)) { 6745 // Oop lies in _span and isn't yet grey or black 6746 _bitMap->mark(addr); // now grey 6747 if (addr < _finger) { 6748 // the bit map iteration has already either passed, or 6749 // sampled, this bit in the bit map; we'll need to 6750 // use the marking stack to scan this oop's oops. 6751 bool simulate_overflow = false; 6752 NOT_PRODUCT( 6753 if (CMSMarkStackOverflowALot && 6754 _collector->simulate_overflow()) { 6755 // simulate a stack overflow 6756 simulate_overflow = true; 6757 } 6758 ) 6759 if (simulate_overflow || !_markStack->push(obj)) { // stack overflow 6760 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _markStack->capacity()); 6761 assert(simulate_overflow || _markStack->isFull(), "Else push should have succeeded"); 6762 handle_stack_overflow(addr); 6763 } 6764 } 6765 // anything including and to the right of _finger 6766 // will be scanned as we iterate over the remainder of the 6767 // bit map 6768 do_yield_check(); 6769 } 6770 } 6771 6772 void PushOrMarkClosure::do_oop(oop* p) { PushOrMarkClosure::do_oop_work(p); } 6773 void PushOrMarkClosure::do_oop(narrowOop* p) { PushOrMarkClosure::do_oop_work(p); } 6774 6775 void Par_PushOrMarkClosure::do_oop(oop obj) { 6776 // Ignore mark word because we are running concurrent with mutators. 6777 assert(obj->is_oop_or_null(true), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6778 HeapWord* addr = (HeapWord*)obj; 6779 if (_whole_span.contains(addr) && !_bit_map->isMarked(addr)) { 6780 // Oop lies in _span and isn't yet grey or black 6781 // We read the global_finger (volatile read) strictly after marking oop 6782 bool res = _bit_map->par_mark(addr); // now grey 6783 volatile HeapWord** gfa = (volatile HeapWord**)_global_finger_addr; 6784 // Should we push this marked oop on our stack? 6785 // -- if someone else marked it, nothing to do 6786 // -- if target oop is above global finger nothing to do 6787 // -- if target oop is in chunk and above local finger 6788 // then nothing to do 6789 // -- else push on work queue 6790 if ( !res // someone else marked it, they will deal with it 6791 || (addr >= *gfa) // will be scanned in a later task 6792 || (_span.contains(addr) && addr >= _finger)) { // later in this chunk 6793 return; 6794 } 6795 // the bit map iteration has already either passed, or 6796 // sampled, this bit in the bit map; we'll need to 6797 // use the marking stack to scan this oop's oops. 6798 bool simulate_overflow = false; 6799 NOT_PRODUCT( 6800 if (CMSMarkStackOverflowALot && 6801 _collector->simulate_overflow()) { 6802 // simulate a stack overflow 6803 simulate_overflow = true; 6804 } 6805 ) 6806 if (simulate_overflow || 6807 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) { 6808 // stack overflow 6809 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _overflow_stack->capacity()); 6810 // We cannot assert that the overflow stack is full because 6811 // it may have been emptied since. 6812 assert(simulate_overflow || 6813 _work_queue->size() == _work_queue->max_elems(), 6814 "Else push should have succeeded"); 6815 handle_stack_overflow(addr); 6816 } 6817 do_yield_check(); 6818 } 6819 } 6820 6821 void Par_PushOrMarkClosure::do_oop(oop* p) { Par_PushOrMarkClosure::do_oop_work(p); } 6822 void Par_PushOrMarkClosure::do_oop(narrowOop* p) { Par_PushOrMarkClosure::do_oop_work(p); } 6823 6824 PushAndMarkClosure::PushAndMarkClosure(CMSCollector* collector, 6825 MemRegion span, 6826 ReferenceProcessor* rp, 6827 CMSBitMap* bit_map, 6828 CMSBitMap* mod_union_table, 6829 CMSMarkStack* mark_stack, 6830 bool concurrent_precleaning): 6831 MetadataAwareOopClosure(rp), 6832 _collector(collector), 6833 _span(span), 6834 _bit_map(bit_map), 6835 _mod_union_table(mod_union_table), 6836 _mark_stack(mark_stack), 6837 _concurrent_precleaning(concurrent_precleaning) 6838 { 6839 assert(ref_processor() != NULL, "ref_processor shouldn't be NULL"); 6840 } 6841 6842 // Grey object rescan during pre-cleaning and second checkpoint phases -- 6843 // the non-parallel version (the parallel version appears further below.) 6844 void PushAndMarkClosure::do_oop(oop obj) { 6845 // Ignore mark word verification. If during concurrent precleaning, 6846 // the object monitor may be locked. If during the checkpoint 6847 // phases, the object may already have been reached by a different 6848 // path and may be at the end of the global overflow list (so 6849 // the mark word may be NULL). 6850 assert(obj->is_oop_or_null(true /* ignore mark word */), 6851 "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6852 HeapWord* addr = (HeapWord*)obj; 6853 // Check if oop points into the CMS generation 6854 // and is not marked 6855 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 6856 // a white object ... 6857 _bit_map->mark(addr); // ... now grey 6858 // push on the marking stack (grey set) 6859 bool simulate_overflow = false; 6860 NOT_PRODUCT( 6861 if (CMSMarkStackOverflowALot && 6862 _collector->simulate_overflow()) { 6863 // simulate a stack overflow 6864 simulate_overflow = true; 6865 } 6866 ) 6867 if (simulate_overflow || !_mark_stack->push(obj)) { 6868 if (_concurrent_precleaning) { 6869 // During precleaning we can just dirty the appropriate card(s) 6870 // in the mod union table, thus ensuring that the object remains 6871 // in the grey set and continue. In the case of object arrays 6872 // we need to dirty all of the cards that the object spans, 6873 // since the rescan of object arrays will be limited to the 6874 // dirty cards. 6875 // Note that no one can be interfering with us in this action 6876 // of dirtying the mod union table, so no locking or atomics 6877 // are required. 6878 if (obj->is_objArray()) { 6879 size_t sz = obj->size(); 6880 HeapWord* end_card_addr = (HeapWord*)round_to( 6881 (intptr_t)(addr+sz), CardTableModRefBS::card_size); 6882 MemRegion redirty_range = MemRegion(addr, end_card_addr); 6883 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 6884 _mod_union_table->mark_range(redirty_range); 6885 } else { 6886 _mod_union_table->mark(addr); 6887 } 6888 _collector->_ser_pmc_preclean_ovflw++; 6889 } else { 6890 // During the remark phase, we need to remember this oop 6891 // in the overflow list. 6892 _collector->push_on_overflow_list(obj); 6893 _collector->_ser_pmc_remark_ovflw++; 6894 } 6895 } 6896 } 6897 } 6898 6899 Par_PushAndMarkClosure::Par_PushAndMarkClosure(CMSCollector* collector, 6900 MemRegion span, 6901 ReferenceProcessor* rp, 6902 CMSBitMap* bit_map, 6903 OopTaskQueue* work_queue): 6904 MetadataAwareOopClosure(rp), 6905 _collector(collector), 6906 _span(span), 6907 _bit_map(bit_map), 6908 _work_queue(work_queue) 6909 { 6910 assert(ref_processor() != NULL, "ref_processor shouldn't be NULL"); 6911 } 6912 6913 void PushAndMarkClosure::do_oop(oop* p) { PushAndMarkClosure::do_oop_work(p); } 6914 void PushAndMarkClosure::do_oop(narrowOop* p) { PushAndMarkClosure::do_oop_work(p); } 6915 6916 // Grey object rescan during second checkpoint phase -- 6917 // the parallel version. 6918 void Par_PushAndMarkClosure::do_oop(oop obj) { 6919 // In the assert below, we ignore the mark word because 6920 // this oop may point to an already visited object that is 6921 // on the overflow stack (in which case the mark word has 6922 // been hijacked for chaining into the overflow stack -- 6923 // if this is the last object in the overflow stack then 6924 // its mark word will be NULL). Because this object may 6925 // have been subsequently popped off the global overflow 6926 // stack, and the mark word possibly restored to the prototypical 6927 // value, by the time we get to examined this failing assert in 6928 // the debugger, is_oop_or_null(false) may subsequently start 6929 // to hold. 6930 assert(obj->is_oop_or_null(true), 6931 "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6932 HeapWord* addr = (HeapWord*)obj; 6933 // Check if oop points into the CMS generation 6934 // and is not marked 6935 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 6936 // a white object ... 6937 // If we manage to "claim" the object, by being the 6938 // first thread to mark it, then we push it on our 6939 // marking stack 6940 if (_bit_map->par_mark(addr)) { // ... now grey 6941 // push on work queue (grey set) 6942 bool simulate_overflow = false; 6943 NOT_PRODUCT( 6944 if (CMSMarkStackOverflowALot && 6945 _collector->par_simulate_overflow()) { 6946 // simulate a stack overflow 6947 simulate_overflow = true; 6948 } 6949 ) 6950 if (simulate_overflow || !_work_queue->push(obj)) { 6951 _collector->par_push_on_overflow_list(obj); 6952 _collector->_par_pmc_remark_ovflw++; // imprecise OK: no need to CAS 6953 } 6954 } // Else, some other thread got there first 6955 } 6956 } 6957 6958 void Par_PushAndMarkClosure::do_oop(oop* p) { Par_PushAndMarkClosure::do_oop_work(p); } 6959 void Par_PushAndMarkClosure::do_oop(narrowOop* p) { Par_PushAndMarkClosure::do_oop_work(p); } 6960 6961 void CMSPrecleanRefsYieldClosure::do_yield_work() { 6962 Mutex* bml = _collector->bitMapLock(); 6963 assert_lock_strong(bml); 6964 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6965 "CMS thread should hold CMS token"); 6966 6967 bml->unlock(); 6968 ConcurrentMarkSweepThread::desynchronize(true); 6969 6970 _collector->stopTimer(); 6971 _collector->incrementYields(); 6972 6973 // See the comment in coordinator_yield() 6974 for (unsigned i = 0; i < CMSYieldSleepCount && 6975 ConcurrentMarkSweepThread::should_yield() && 6976 !CMSCollector::foregroundGCIsActive(); ++i) { 6977 os::sleep(Thread::current(), 1, false); 6978 } 6979 6980 ConcurrentMarkSweepThread::synchronize(true); 6981 bml->lock(); 6982 6983 _collector->startTimer(); 6984 } 6985 6986 bool CMSPrecleanRefsYieldClosure::should_return() { 6987 if (ConcurrentMarkSweepThread::should_yield()) { 6988 do_yield_work(); 6989 } 6990 return _collector->foregroundGCIsActive(); 6991 } 6992 6993 void MarkFromDirtyCardsClosure::do_MemRegion(MemRegion mr) { 6994 assert(((size_t)mr.start())%CardTableModRefBS::card_size_in_words == 0, 6995 "mr should be aligned to start at a card boundary"); 6996 // We'd like to assert: 6997 // assert(mr.word_size()%CardTableModRefBS::card_size_in_words == 0, 6998 // "mr should be a range of cards"); 6999 // However, that would be too strong in one case -- the last 7000 // partition ends at _unallocated_block which, in general, can be 7001 // an arbitrary boundary, not necessarily card aligned. 7002 _num_dirty_cards += mr.word_size()/CardTableModRefBS::card_size_in_words; 7003 _space->object_iterate_mem(mr, &_scan_cl); 7004 } 7005 7006 SweepClosure::SweepClosure(CMSCollector* collector, 7007 ConcurrentMarkSweepGeneration* g, 7008 CMSBitMap* bitMap, bool should_yield) : 7009 _collector(collector), 7010 _g(g), 7011 _sp(g->cmsSpace()), 7012 _limit(_sp->sweep_limit()), 7013 _freelistLock(_sp->freelistLock()), 7014 _bitMap(bitMap), 7015 _yield(should_yield), 7016 _inFreeRange(false), // No free range at beginning of sweep 7017 _freeRangeInFreeLists(false), // No free range at beginning of sweep 7018 _lastFreeRangeCoalesced(false), 7019 _freeFinger(g->used_region().start()) 7020 { 7021 NOT_PRODUCT( 7022 _numObjectsFreed = 0; 7023 _numWordsFreed = 0; 7024 _numObjectsLive = 0; 7025 _numWordsLive = 0; 7026 _numObjectsAlreadyFree = 0; 7027 _numWordsAlreadyFree = 0; 7028 _last_fc = NULL; 7029 7030 _sp->initializeIndexedFreeListArrayReturnedBytes(); 7031 _sp->dictionary()->initialize_dict_returned_bytes(); 7032 ) 7033 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7034 "sweep _limit out of bounds"); 7035 log_develop_trace(gc, sweep)("===================="); 7036 log_develop_trace(gc, sweep)("Starting new sweep with limit " PTR_FORMAT, p2i(_limit)); 7037 } 7038 7039 void SweepClosure::print_on(outputStream* st) const { 7040 tty->print_cr("_sp = [" PTR_FORMAT "," PTR_FORMAT ")", 7041 p2i(_sp->bottom()), p2i(_sp->end())); 7042 tty->print_cr("_limit = " PTR_FORMAT, p2i(_limit)); 7043 tty->print_cr("_freeFinger = " PTR_FORMAT, p2i(_freeFinger)); 7044 NOT_PRODUCT(tty->print_cr("_last_fc = " PTR_FORMAT, p2i(_last_fc));) 7045 tty->print_cr("_inFreeRange = %d, _freeRangeInFreeLists = %d, _lastFreeRangeCoalesced = %d", 7046 _inFreeRange, _freeRangeInFreeLists, _lastFreeRangeCoalesced); 7047 } 7048 7049 #ifndef PRODUCT 7050 // Assertion checking only: no useful work in product mode -- 7051 // however, if any of the flags below become product flags, 7052 // you may need to review this code to see if it needs to be 7053 // enabled in product mode. 7054 SweepClosure::~SweepClosure() { 7055 assert_lock_strong(_freelistLock); 7056 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7057 "sweep _limit out of bounds"); 7058 if (inFreeRange()) { 7059 warning("inFreeRange() should have been reset; dumping state of SweepClosure"); 7060 print(); 7061 ShouldNotReachHere(); 7062 } 7063 7064 if (log_is_enabled(Debug, gc, sweep)) { 7065 log_debug(gc, sweep)("Collected " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes", 7066 _numObjectsFreed, _numWordsFreed*sizeof(HeapWord)); 7067 log_debug(gc, sweep)("Live " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes Already free " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes", 7068 _numObjectsLive, _numWordsLive*sizeof(HeapWord), _numObjectsAlreadyFree, _numWordsAlreadyFree*sizeof(HeapWord)); 7069 size_t totalBytes = (_numWordsFreed + _numWordsLive + _numWordsAlreadyFree) * sizeof(HeapWord); 7070 log_debug(gc, sweep)("Total sweep: " SIZE_FORMAT " bytes", totalBytes); 7071 } 7072 7073 if (log_is_enabled(Trace, gc, sweep) && CMSVerifyReturnedBytes) { 7074 size_t indexListReturnedBytes = _sp->sumIndexedFreeListArrayReturnedBytes(); 7075 size_t dict_returned_bytes = _sp->dictionary()->sum_dict_returned_bytes(); 7076 size_t returned_bytes = indexListReturnedBytes + dict_returned_bytes; 7077 log_trace(gc, sweep)("Returned " SIZE_FORMAT " bytes Indexed List Returned " SIZE_FORMAT " bytes Dictionary Returned " SIZE_FORMAT " bytes", 7078 returned_bytes, indexListReturnedBytes, dict_returned_bytes); 7079 } 7080 log_develop_trace(gc, sweep)("end of sweep with _limit = " PTR_FORMAT, p2i(_limit)); 7081 log_develop_trace(gc, sweep)("================"); 7082 } 7083 #endif // PRODUCT 7084 7085 void SweepClosure::initialize_free_range(HeapWord* freeFinger, 7086 bool freeRangeInFreeLists) { 7087 log_develop_trace(gc, sweep)("---- Start free range at " PTR_FORMAT " with free block (%d)", 7088 p2i(freeFinger), freeRangeInFreeLists); 7089 assert(!inFreeRange(), "Trampling existing free range"); 7090 set_inFreeRange(true); 7091 set_lastFreeRangeCoalesced(false); 7092 7093 set_freeFinger(freeFinger); 7094 set_freeRangeInFreeLists(freeRangeInFreeLists); 7095 if (CMSTestInFreeList) { 7096 if (freeRangeInFreeLists) { 7097 FreeChunk* fc = (FreeChunk*) freeFinger; 7098 assert(fc->is_free(), "A chunk on the free list should be free."); 7099 assert(fc->size() > 0, "Free range should have a size"); 7100 assert(_sp->verify_chunk_in_free_list(fc), "Chunk is not in free lists"); 7101 } 7102 } 7103 } 7104 7105 // Note that the sweeper runs concurrently with mutators. Thus, 7106 // it is possible for direct allocation in this generation to happen 7107 // in the middle of the sweep. Note that the sweeper also coalesces 7108 // contiguous free blocks. Thus, unless the sweeper and the allocator 7109 // synchronize appropriately freshly allocated blocks may get swept up. 7110 // This is accomplished by the sweeper locking the free lists while 7111 // it is sweeping. Thus blocks that are determined to be free are 7112 // indeed free. There is however one additional complication: 7113 // blocks that have been allocated since the final checkpoint and 7114 // mark, will not have been marked and so would be treated as 7115 // unreachable and swept up. To prevent this, the allocator marks 7116 // the bit map when allocating during the sweep phase. This leads, 7117 // however, to a further complication -- objects may have been allocated 7118 // but not yet initialized -- in the sense that the header isn't yet 7119 // installed. The sweeper can not then determine the size of the block 7120 // in order to skip over it. To deal with this case, we use a technique 7121 // (due to Printezis) to encode such uninitialized block sizes in the 7122 // bit map. Since the bit map uses a bit per every HeapWord, but the 7123 // CMS generation has a minimum object size of 3 HeapWords, it follows 7124 // that "normal marks" won't be adjacent in the bit map (there will 7125 // always be at least two 0 bits between successive 1 bits). We make use 7126 // of these "unused" bits to represent uninitialized blocks -- the bit 7127 // corresponding to the start of the uninitialized object and the next 7128 // bit are both set. Finally, a 1 bit marks the end of the object that 7129 // started with the two consecutive 1 bits to indicate its potentially 7130 // uninitialized state. 7131 7132 size_t SweepClosure::do_blk_careful(HeapWord* addr) { 7133 FreeChunk* fc = (FreeChunk*)addr; 7134 size_t res; 7135 7136 // Check if we are done sweeping. Below we check "addr >= _limit" rather 7137 // than "addr == _limit" because although _limit was a block boundary when 7138 // we started the sweep, it may no longer be one because heap expansion 7139 // may have caused us to coalesce the block ending at the address _limit 7140 // with a newly expanded chunk (this happens when _limit was set to the 7141 // previous _end of the space), so we may have stepped past _limit: 7142 // see the following Zeno-like trail of CRs 6977970, 7008136, 7042740. 7143 if (addr >= _limit) { // we have swept up to or past the limit: finish up 7144 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7145 "sweep _limit out of bounds"); 7146 assert(addr < _sp->end(), "addr out of bounds"); 7147 // Flush any free range we might be holding as a single 7148 // coalesced chunk to the appropriate free list. 7149 if (inFreeRange()) { 7150 assert(freeFinger() >= _sp->bottom() && freeFinger() < _limit, 7151 "freeFinger() " PTR_FORMAT " is out-of-bounds", p2i(freeFinger())); 7152 flush_cur_free_chunk(freeFinger(), 7153 pointer_delta(addr, freeFinger())); 7154 log_develop_trace(gc, sweep)("Sweep: last chunk: put_free_blk " PTR_FORMAT " (" SIZE_FORMAT ") [coalesced:%d]", 7155 p2i(freeFinger()), pointer_delta(addr, freeFinger()), 7156 lastFreeRangeCoalesced() ? 1 : 0); 7157 } 7158 7159 // help the iterator loop finish 7160 return pointer_delta(_sp->end(), addr); 7161 } 7162 7163 assert(addr < _limit, "sweep invariant"); 7164 // check if we should yield 7165 do_yield_check(addr); 7166 if (fc->is_free()) { 7167 // Chunk that is already free 7168 res = fc->size(); 7169 do_already_free_chunk(fc); 7170 debug_only(_sp->verifyFreeLists()); 7171 // If we flush the chunk at hand in lookahead_and_flush() 7172 // and it's coalesced with a preceding chunk, then the 7173 // process of "mangling" the payload of the coalesced block 7174 // will cause erasure of the size information from the 7175 // (erstwhile) header of all the coalesced blocks but the 7176 // first, so the first disjunct in the assert will not hold 7177 // in that specific case (in which case the second disjunct 7178 // will hold). 7179 assert(res == fc->size() || ((HeapWord*)fc) + res >= _limit, 7180 "Otherwise the size info doesn't change at this step"); 7181 NOT_PRODUCT( 7182 _numObjectsAlreadyFree++; 7183 _numWordsAlreadyFree += res; 7184 ) 7185 NOT_PRODUCT(_last_fc = fc;) 7186 } else if (!_bitMap->isMarked(addr)) { 7187 // Chunk is fresh garbage 7188 res = do_garbage_chunk(fc); 7189 debug_only(_sp->verifyFreeLists()); 7190 NOT_PRODUCT( 7191 _numObjectsFreed++; 7192 _numWordsFreed += res; 7193 ) 7194 } else { 7195 // Chunk that is alive. 7196 res = do_live_chunk(fc); 7197 debug_only(_sp->verifyFreeLists()); 7198 NOT_PRODUCT( 7199 _numObjectsLive++; 7200 _numWordsLive += res; 7201 ) 7202 } 7203 return res; 7204 } 7205 7206 // For the smart allocation, record following 7207 // split deaths - a free chunk is removed from its free list because 7208 // it is being split into two or more chunks. 7209 // split birth - a free chunk is being added to its free list because 7210 // a larger free chunk has been split and resulted in this free chunk. 7211 // coal death - a free chunk is being removed from its free list because 7212 // it is being coalesced into a large free chunk. 7213 // coal birth - a free chunk is being added to its free list because 7214 // it was created when two or more free chunks where coalesced into 7215 // this free chunk. 7216 // 7217 // These statistics are used to determine the desired number of free 7218 // chunks of a given size. The desired number is chosen to be relative 7219 // to the end of a CMS sweep. The desired number at the end of a sweep 7220 // is the 7221 // count-at-end-of-previous-sweep (an amount that was enough) 7222 // - count-at-beginning-of-current-sweep (the excess) 7223 // + split-births (gains in this size during interval) 7224 // - split-deaths (demands on this size during interval) 7225 // where the interval is from the end of one sweep to the end of the 7226 // next. 7227 // 7228 // When sweeping the sweeper maintains an accumulated chunk which is 7229 // the chunk that is made up of chunks that have been coalesced. That 7230 // will be termed the left-hand chunk. A new chunk of garbage that 7231 // is being considered for coalescing will be referred to as the 7232 // right-hand chunk. 7233 // 7234 // When making a decision on whether to coalesce a right-hand chunk with 7235 // the current left-hand chunk, the current count vs. the desired count 7236 // of the left-hand chunk is considered. Also if the right-hand chunk 7237 // is near the large chunk at the end of the heap (see 7238 // ConcurrentMarkSweepGeneration::isNearLargestChunk()), then the 7239 // left-hand chunk is coalesced. 7240 // 7241 // When making a decision about whether to split a chunk, the desired count 7242 // vs. the current count of the candidate to be split is also considered. 7243 // If the candidate is underpopulated (currently fewer chunks than desired) 7244 // a chunk of an overpopulated (currently more chunks than desired) size may 7245 // be chosen. The "hint" associated with a free list, if non-null, points 7246 // to a free list which may be overpopulated. 7247 // 7248 7249 void SweepClosure::do_already_free_chunk(FreeChunk* fc) { 7250 const size_t size = fc->size(); 7251 // Chunks that cannot be coalesced are not in the 7252 // free lists. 7253 if (CMSTestInFreeList && !fc->cantCoalesce()) { 7254 assert(_sp->verify_chunk_in_free_list(fc), 7255 "free chunk should be in free lists"); 7256 } 7257 // a chunk that is already free, should not have been 7258 // marked in the bit map 7259 HeapWord* const addr = (HeapWord*) fc; 7260 assert(!_bitMap->isMarked(addr), "free chunk should be unmarked"); 7261 // Verify that the bit map has no bits marked between 7262 // addr and purported end of this block. 7263 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 7264 7265 // Some chunks cannot be coalesced under any circumstances. 7266 // See the definition of cantCoalesce(). 7267 if (!fc->cantCoalesce()) { 7268 // This chunk can potentially be coalesced. 7269 // All the work is done in 7270 do_post_free_or_garbage_chunk(fc, size); 7271 // Note that if the chunk is not coalescable (the else arm 7272 // below), we unconditionally flush, without needing to do 7273 // a "lookahead," as we do below. 7274 if (inFreeRange()) lookahead_and_flush(fc, size); 7275 } else { 7276 // Code path common to both original and adaptive free lists. 7277 7278 // cant coalesce with previous block; this should be treated 7279 // as the end of a free run if any 7280 if (inFreeRange()) { 7281 // we kicked some butt; time to pick up the garbage 7282 assert(freeFinger() < addr, "freeFinger points too high"); 7283 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 7284 } 7285 // else, nothing to do, just continue 7286 } 7287 } 7288 7289 size_t SweepClosure::do_garbage_chunk(FreeChunk* fc) { 7290 // This is a chunk of garbage. It is not in any free list. 7291 // Add it to a free list or let it possibly be coalesced into 7292 // a larger chunk. 7293 HeapWord* const addr = (HeapWord*) fc; 7294 const size_t size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 7295 7296 // Verify that the bit map has no bits marked between 7297 // addr and purported end of just dead object. 7298 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 7299 do_post_free_or_garbage_chunk(fc, size); 7300 7301 assert(_limit >= addr + size, 7302 "A freshly garbage chunk can't possibly straddle over _limit"); 7303 if (inFreeRange()) lookahead_and_flush(fc, size); 7304 return size; 7305 } 7306 7307 size_t SweepClosure::do_live_chunk(FreeChunk* fc) { 7308 HeapWord* addr = (HeapWord*) fc; 7309 // The sweeper has just found a live object. Return any accumulated 7310 // left hand chunk to the free lists. 7311 if (inFreeRange()) { 7312 assert(freeFinger() < addr, "freeFinger points too high"); 7313 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 7314 } 7315 7316 // This object is live: we'd normally expect this to be 7317 // an oop, and like to assert the following: 7318 // assert(oop(addr)->is_oop(), "live block should be an oop"); 7319 // However, as we commented above, this may be an object whose 7320 // header hasn't yet been initialized. 7321 size_t size; 7322 assert(_bitMap->isMarked(addr), "Tautology for this control point"); 7323 if (_bitMap->isMarked(addr + 1)) { 7324 // Determine the size from the bit map, rather than trying to 7325 // compute it from the object header. 7326 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 7327 size = pointer_delta(nextOneAddr + 1, addr); 7328 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 7329 "alignment problem"); 7330 7331 #ifdef ASSERT 7332 if (oop(addr)->klass_or_null() != NULL) { 7333 // Ignore mark word because we are running concurrent with mutators 7334 assert(oop(addr)->is_oop(true), "live block should be an oop"); 7335 assert(size == 7336 CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()), 7337 "P-mark and computed size do not agree"); 7338 } 7339 #endif 7340 7341 } else { 7342 // This should be an initialized object that's alive. 7343 assert(oop(addr)->klass_or_null() != NULL, 7344 "Should be an initialized object"); 7345 // Ignore mark word because we are running concurrent with mutators 7346 assert(oop(addr)->is_oop(true), "live block should be an oop"); 7347 // Verify that the bit map has no bits marked between 7348 // addr and purported end of this block. 7349 size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 7350 assert(size >= 3, "Necessary for Printezis marks to work"); 7351 assert(!_bitMap->isMarked(addr+1), "Tautology for this control point"); 7352 DEBUG_ONLY(_bitMap->verifyNoOneBitsInRange(addr+2, addr+size);) 7353 } 7354 return size; 7355 } 7356 7357 void SweepClosure::do_post_free_or_garbage_chunk(FreeChunk* fc, 7358 size_t chunkSize) { 7359 // do_post_free_or_garbage_chunk() should only be called in the case 7360 // of the adaptive free list allocator. 7361 const bool fcInFreeLists = fc->is_free(); 7362 assert((HeapWord*)fc <= _limit, "sweep invariant"); 7363 if (CMSTestInFreeList && fcInFreeLists) { 7364 assert(_sp->verify_chunk_in_free_list(fc), "free chunk is not in free lists"); 7365 } 7366 7367 log_develop_trace(gc, sweep)(" -- pick up another chunk at " PTR_FORMAT " (" SIZE_FORMAT ")", p2i(fc), chunkSize); 7368 7369 HeapWord* const fc_addr = (HeapWord*) fc; 7370 7371 bool coalesce = false; 7372 const size_t left = pointer_delta(fc_addr, freeFinger()); 7373 const size_t right = chunkSize; 7374 switch (FLSCoalescePolicy) { 7375 // numeric value forms a coalition aggressiveness metric 7376 case 0: { // never coalesce 7377 coalesce = false; 7378 break; 7379 } 7380 case 1: { // coalesce if left & right chunks on overpopulated lists 7381 coalesce = _sp->coalOverPopulated(left) && 7382 _sp->coalOverPopulated(right); 7383 break; 7384 } 7385 case 2: { // coalesce if left chunk on overpopulated list (default) 7386 coalesce = _sp->coalOverPopulated(left); 7387 break; 7388 } 7389 case 3: { // coalesce if left OR right chunk on overpopulated list 7390 coalesce = _sp->coalOverPopulated(left) || 7391 _sp->coalOverPopulated(right); 7392 break; 7393 } 7394 case 4: { // always coalesce 7395 coalesce = true; 7396 break; 7397 } 7398 default: 7399 ShouldNotReachHere(); 7400 } 7401 7402 // Should the current free range be coalesced? 7403 // If the chunk is in a free range and either we decided to coalesce above 7404 // or the chunk is near the large block at the end of the heap 7405 // (isNearLargestChunk() returns true), then coalesce this chunk. 7406 const bool doCoalesce = inFreeRange() 7407 && (coalesce || _g->isNearLargestChunk(fc_addr)); 7408 if (doCoalesce) { 7409 // Coalesce the current free range on the left with the new 7410 // chunk on the right. If either is on a free list, 7411 // it must be removed from the list and stashed in the closure. 7412 if (freeRangeInFreeLists()) { 7413 FreeChunk* const ffc = (FreeChunk*)freeFinger(); 7414 assert(ffc->size() == pointer_delta(fc_addr, freeFinger()), 7415 "Size of free range is inconsistent with chunk size."); 7416 if (CMSTestInFreeList) { 7417 assert(_sp->verify_chunk_in_free_list(ffc), 7418 "Chunk is not in free lists"); 7419 } 7420 _sp->coalDeath(ffc->size()); 7421 _sp->removeFreeChunkFromFreeLists(ffc); 7422 set_freeRangeInFreeLists(false); 7423 } 7424 if (fcInFreeLists) { 7425 _sp->coalDeath(chunkSize); 7426 assert(fc->size() == chunkSize, 7427 "The chunk has the wrong size or is not in the free lists"); 7428 _sp->removeFreeChunkFromFreeLists(fc); 7429 } 7430 set_lastFreeRangeCoalesced(true); 7431 print_free_block_coalesced(fc); 7432 } else { // not in a free range and/or should not coalesce 7433 // Return the current free range and start a new one. 7434 if (inFreeRange()) { 7435 // In a free range but cannot coalesce with the right hand chunk. 7436 // Put the current free range into the free lists. 7437 flush_cur_free_chunk(freeFinger(), 7438 pointer_delta(fc_addr, freeFinger())); 7439 } 7440 // Set up for new free range. Pass along whether the right hand 7441 // chunk is in the free lists. 7442 initialize_free_range((HeapWord*)fc, fcInFreeLists); 7443 } 7444 } 7445 7446 // Lookahead flush: 7447 // If we are tracking a free range, and this is the last chunk that 7448 // we'll look at because its end crosses past _limit, we'll preemptively 7449 // flush it along with any free range we may be holding on to. Note that 7450 // this can be the case only for an already free or freshly garbage 7451 // chunk. If this block is an object, it can never straddle 7452 // over _limit. The "straddling" occurs when _limit is set at 7453 // the previous end of the space when this cycle started, and 7454 // a subsequent heap expansion caused the previously co-terminal 7455 // free block to be coalesced with the newly expanded portion, 7456 // thus rendering _limit a non-block-boundary making it dangerous 7457 // for the sweeper to step over and examine. 7458 void SweepClosure::lookahead_and_flush(FreeChunk* fc, size_t chunk_size) { 7459 assert(inFreeRange(), "Should only be called if currently in a free range."); 7460 HeapWord* const eob = ((HeapWord*)fc) + chunk_size; 7461 assert(_sp->used_region().contains(eob - 1), 7462 "eob = " PTR_FORMAT " eob-1 = " PTR_FORMAT " _limit = " PTR_FORMAT 7463 " out of bounds wrt _sp = [" PTR_FORMAT "," PTR_FORMAT ")" 7464 " when examining fc = " PTR_FORMAT "(" SIZE_FORMAT ")", 7465 p2i(eob), p2i(eob-1), p2i(_limit), p2i(_sp->bottom()), p2i(_sp->end()), p2i(fc), chunk_size); 7466 if (eob >= _limit) { 7467 assert(eob == _limit || fc->is_free(), "Only a free chunk should allow us to cross over the limit"); 7468 log_develop_trace(gc, sweep)("_limit " PTR_FORMAT " reached or crossed by block " 7469 "[" PTR_FORMAT "," PTR_FORMAT ") in space " 7470 "[" PTR_FORMAT "," PTR_FORMAT ")", 7471 p2i(_limit), p2i(fc), p2i(eob), p2i(_sp->bottom()), p2i(_sp->end())); 7472 // Return the storage we are tracking back into the free lists. 7473 log_develop_trace(gc, sweep)("Flushing ... "); 7474 assert(freeFinger() < eob, "Error"); 7475 flush_cur_free_chunk( freeFinger(), pointer_delta(eob, freeFinger())); 7476 } 7477 } 7478 7479 void SweepClosure::flush_cur_free_chunk(HeapWord* chunk, size_t size) { 7480 assert(inFreeRange(), "Should only be called if currently in a free range."); 7481 assert(size > 0, 7482 "A zero sized chunk cannot be added to the free lists."); 7483 if (!freeRangeInFreeLists()) { 7484 if (CMSTestInFreeList) { 7485 FreeChunk* fc = (FreeChunk*) chunk; 7486 fc->set_size(size); 7487 assert(!_sp->verify_chunk_in_free_list(fc), 7488 "chunk should not be in free lists yet"); 7489 } 7490 log_develop_trace(gc, sweep)(" -- add free block " PTR_FORMAT " (" SIZE_FORMAT ") to free lists", p2i(chunk), size); 7491 // A new free range is going to be starting. The current 7492 // free range has not been added to the free lists yet or 7493 // was removed so add it back. 7494 // If the current free range was coalesced, then the death 7495 // of the free range was recorded. Record a birth now. 7496 if (lastFreeRangeCoalesced()) { 7497 _sp->coalBirth(size); 7498 } 7499 _sp->addChunkAndRepairOffsetTable(chunk, size, 7500 lastFreeRangeCoalesced()); 7501 } else { 7502 log_develop_trace(gc, sweep)("Already in free list: nothing to flush"); 7503 } 7504 set_inFreeRange(false); 7505 set_freeRangeInFreeLists(false); 7506 } 7507 7508 // We take a break if we've been at this for a while, 7509 // so as to avoid monopolizing the locks involved. 7510 void SweepClosure::do_yield_work(HeapWord* addr) { 7511 // Return current free chunk being used for coalescing (if any) 7512 // to the appropriate freelist. After yielding, the next 7513 // free block encountered will start a coalescing range of 7514 // free blocks. If the next free block is adjacent to the 7515 // chunk just flushed, they will need to wait for the next 7516 // sweep to be coalesced. 7517 if (inFreeRange()) { 7518 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 7519 } 7520 7521 // First give up the locks, then yield, then re-lock. 7522 // We should probably use a constructor/destructor idiom to 7523 // do this unlock/lock or modify the MutexUnlocker class to 7524 // serve our purpose. XXX 7525 assert_lock_strong(_bitMap->lock()); 7526 assert_lock_strong(_freelistLock); 7527 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 7528 "CMS thread should hold CMS token"); 7529 _bitMap->lock()->unlock(); 7530 _freelistLock->unlock(); 7531 ConcurrentMarkSweepThread::desynchronize(true); 7532 _collector->stopTimer(); 7533 _collector->incrementYields(); 7534 7535 // See the comment in coordinator_yield() 7536 for (unsigned i = 0; i < CMSYieldSleepCount && 7537 ConcurrentMarkSweepThread::should_yield() && 7538 !CMSCollector::foregroundGCIsActive(); ++i) { 7539 os::sleep(Thread::current(), 1, false); 7540 } 7541 7542 ConcurrentMarkSweepThread::synchronize(true); 7543 _freelistLock->lock(); 7544 _bitMap->lock()->lock_without_safepoint_check(); 7545 _collector->startTimer(); 7546 } 7547 7548 #ifndef PRODUCT 7549 // This is actually very useful in a product build if it can 7550 // be called from the debugger. Compile it into the product 7551 // as needed. 7552 bool debug_verify_chunk_in_free_list(FreeChunk* fc) { 7553 return debug_cms_space->verify_chunk_in_free_list(fc); 7554 } 7555 #endif 7556 7557 void SweepClosure::print_free_block_coalesced(FreeChunk* fc) const { 7558 log_develop_trace(gc, sweep)("Sweep:coal_free_blk " PTR_FORMAT " (" SIZE_FORMAT ")", 7559 p2i(fc), fc->size()); 7560 } 7561 7562 // CMSIsAliveClosure 7563 bool CMSIsAliveClosure::do_object_b(oop obj) { 7564 HeapWord* addr = (HeapWord*)obj; 7565 return addr != NULL && 7566 (!_span.contains(addr) || _bit_map->isMarked(addr)); 7567 } 7568 7569 7570 CMSKeepAliveClosure::CMSKeepAliveClosure( CMSCollector* collector, 7571 MemRegion span, 7572 CMSBitMap* bit_map, CMSMarkStack* mark_stack, 7573 bool cpc): 7574 _collector(collector), 7575 _span(span), 7576 _bit_map(bit_map), 7577 _mark_stack(mark_stack), 7578 _concurrent_precleaning(cpc) { 7579 assert(!_span.is_empty(), "Empty span could spell trouble"); 7580 } 7581 7582 7583 // CMSKeepAliveClosure: the serial version 7584 void CMSKeepAliveClosure::do_oop(oop obj) { 7585 HeapWord* addr = (HeapWord*)obj; 7586 if (_span.contains(addr) && 7587 !_bit_map->isMarked(addr)) { 7588 _bit_map->mark(addr); 7589 bool simulate_overflow = false; 7590 NOT_PRODUCT( 7591 if (CMSMarkStackOverflowALot && 7592 _collector->simulate_overflow()) { 7593 // simulate a stack overflow 7594 simulate_overflow = true; 7595 } 7596 ) 7597 if (simulate_overflow || !_mark_stack->push(obj)) { 7598 if (_concurrent_precleaning) { 7599 // We dirty the overflown object and let the remark 7600 // phase deal with it. 7601 assert(_collector->overflow_list_is_empty(), "Error"); 7602 // In the case of object arrays, we need to dirty all of 7603 // the cards that the object spans. No locking or atomics 7604 // are needed since no one else can be mutating the mod union 7605 // table. 7606 if (obj->is_objArray()) { 7607 size_t sz = obj->size(); 7608 HeapWord* end_card_addr = 7609 (HeapWord*)round_to((intptr_t)(addr+sz), CardTableModRefBS::card_size); 7610 MemRegion redirty_range = MemRegion(addr, end_card_addr); 7611 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 7612 _collector->_modUnionTable.mark_range(redirty_range); 7613 } else { 7614 _collector->_modUnionTable.mark(addr); 7615 } 7616 _collector->_ser_kac_preclean_ovflw++; 7617 } else { 7618 _collector->push_on_overflow_list(obj); 7619 _collector->_ser_kac_ovflw++; 7620 } 7621 } 7622 } 7623 } 7624 7625 void CMSKeepAliveClosure::do_oop(oop* p) { CMSKeepAliveClosure::do_oop_work(p); } 7626 void CMSKeepAliveClosure::do_oop(narrowOop* p) { CMSKeepAliveClosure::do_oop_work(p); } 7627 7628 // CMSParKeepAliveClosure: a parallel version of the above. 7629 // The work queues are private to each closure (thread), 7630 // but (may be) available for stealing by other threads. 7631 void CMSParKeepAliveClosure::do_oop(oop obj) { 7632 HeapWord* addr = (HeapWord*)obj; 7633 if (_span.contains(addr) && 7634 !_bit_map->isMarked(addr)) { 7635 // In general, during recursive tracing, several threads 7636 // may be concurrently getting here; the first one to 7637 // "tag" it, claims it. 7638 if (_bit_map->par_mark(addr)) { 7639 bool res = _work_queue->push(obj); 7640 assert(res, "Low water mark should be much less than capacity"); 7641 // Do a recursive trim in the hope that this will keep 7642 // stack usage lower, but leave some oops for potential stealers 7643 trim_queue(_low_water_mark); 7644 } // Else, another thread got there first 7645 } 7646 } 7647 7648 void CMSParKeepAliveClosure::do_oop(oop* p) { CMSParKeepAliveClosure::do_oop_work(p); } 7649 void CMSParKeepAliveClosure::do_oop(narrowOop* p) { CMSParKeepAliveClosure::do_oop_work(p); } 7650 7651 void CMSParKeepAliveClosure::trim_queue(uint max) { 7652 while (_work_queue->size() > max) { 7653 oop new_oop; 7654 if (_work_queue->pop_local(new_oop)) { 7655 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 7656 assert(_bit_map->isMarked((HeapWord*)new_oop), 7657 "no white objects on this stack!"); 7658 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 7659 // iterate over the oops in this oop, marking and pushing 7660 // the ones in CMS heap (i.e. in _span). 7661 new_oop->oop_iterate(&_mark_and_push); 7662 } 7663 } 7664 } 7665 7666 CMSInnerParMarkAndPushClosure::CMSInnerParMarkAndPushClosure( 7667 CMSCollector* collector, 7668 MemRegion span, CMSBitMap* bit_map, 7669 OopTaskQueue* work_queue): 7670 _collector(collector), 7671 _span(span), 7672 _bit_map(bit_map), 7673 _work_queue(work_queue) { } 7674 7675 void CMSInnerParMarkAndPushClosure::do_oop(oop obj) { 7676 HeapWord* addr = (HeapWord*)obj; 7677 if (_span.contains(addr) && 7678 !_bit_map->isMarked(addr)) { 7679 if (_bit_map->par_mark(addr)) { 7680 bool simulate_overflow = false; 7681 NOT_PRODUCT( 7682 if (CMSMarkStackOverflowALot && 7683 _collector->par_simulate_overflow()) { 7684 // simulate a stack overflow 7685 simulate_overflow = true; 7686 } 7687 ) 7688 if (simulate_overflow || !_work_queue->push(obj)) { 7689 _collector->par_push_on_overflow_list(obj); 7690 _collector->_par_kac_ovflw++; 7691 } 7692 } // Else another thread got there already 7693 } 7694 } 7695 7696 void CMSInnerParMarkAndPushClosure::do_oop(oop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); } 7697 void CMSInnerParMarkAndPushClosure::do_oop(narrowOop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); } 7698 7699 ////////////////////////////////////////////////////////////////// 7700 // CMSExpansionCause ///////////////////////////// 7701 ////////////////////////////////////////////////////////////////// 7702 const char* CMSExpansionCause::to_string(CMSExpansionCause::Cause cause) { 7703 switch (cause) { 7704 case _no_expansion: 7705 return "No expansion"; 7706 case _satisfy_free_ratio: 7707 return "Free ratio"; 7708 case _satisfy_promotion: 7709 return "Satisfy promotion"; 7710 case _satisfy_allocation: 7711 return "allocation"; 7712 case _allocate_par_lab: 7713 return "Par LAB"; 7714 case _allocate_par_spooling_space: 7715 return "Par Spooling Space"; 7716 case _adaptive_size_policy: 7717 return "Ergonomics"; 7718 default: 7719 return "unknown"; 7720 } 7721 } 7722 7723 void CMSDrainMarkingStackClosure::do_void() { 7724 // the max number to take from overflow list at a time 7725 const size_t num = _mark_stack->capacity()/4; 7726 assert(!_concurrent_precleaning || _collector->overflow_list_is_empty(), 7727 "Overflow list should be NULL during concurrent phases"); 7728 while (!_mark_stack->isEmpty() || 7729 // if stack is empty, check the overflow list 7730 _collector->take_from_overflow_list(num, _mark_stack)) { 7731 oop obj = _mark_stack->pop(); 7732 HeapWord* addr = (HeapWord*)obj; 7733 assert(_span.contains(addr), "Should be within span"); 7734 assert(_bit_map->isMarked(addr), "Should be marked"); 7735 assert(obj->is_oop(), "Should be an oop"); 7736 obj->oop_iterate(_keep_alive); 7737 } 7738 } 7739 7740 void CMSParDrainMarkingStackClosure::do_void() { 7741 // drain queue 7742 trim_queue(0); 7743 } 7744 7745 // Trim our work_queue so its length is below max at return 7746 void CMSParDrainMarkingStackClosure::trim_queue(uint max) { 7747 while (_work_queue->size() > max) { 7748 oop new_oop; 7749 if (_work_queue->pop_local(new_oop)) { 7750 assert(new_oop->is_oop(), "Expected an oop"); 7751 assert(_bit_map->isMarked((HeapWord*)new_oop), 7752 "no white objects on this stack!"); 7753 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 7754 // iterate over the oops in this oop, marking and pushing 7755 // the ones in CMS heap (i.e. in _span). 7756 new_oop->oop_iterate(&_mark_and_push); 7757 } 7758 } 7759 } 7760 7761 //////////////////////////////////////////////////////////////////// 7762 // Support for Marking Stack Overflow list handling and related code 7763 //////////////////////////////////////////////////////////////////// 7764 // Much of the following code is similar in shape and spirit to the 7765 // code used in ParNewGC. We should try and share that code 7766 // as much as possible in the future. 7767 7768 #ifndef PRODUCT 7769 // Debugging support for CMSStackOverflowALot 7770 7771 // It's OK to call this multi-threaded; the worst thing 7772 // that can happen is that we'll get a bunch of closely 7773 // spaced simulated overflows, but that's OK, in fact 7774 // probably good as it would exercise the overflow code 7775 // under contention. 7776 bool CMSCollector::simulate_overflow() { 7777 if (_overflow_counter-- <= 0) { // just being defensive 7778 _overflow_counter = CMSMarkStackOverflowInterval; 7779 return true; 7780 } else { 7781 return false; 7782 } 7783 } 7784 7785 bool CMSCollector::par_simulate_overflow() { 7786 return simulate_overflow(); 7787 } 7788 #endif 7789 7790 // Single-threaded 7791 bool CMSCollector::take_from_overflow_list(size_t num, CMSMarkStack* stack) { 7792 assert(stack->isEmpty(), "Expected precondition"); 7793 assert(stack->capacity() > num, "Shouldn't bite more than can chew"); 7794 size_t i = num; 7795 oop cur = _overflow_list; 7796 const markOop proto = markOopDesc::prototype(); 7797 NOT_PRODUCT(ssize_t n = 0;) 7798 for (oop next; i > 0 && cur != NULL; cur = next, i--) { 7799 next = oop(cur->mark()); 7800 cur->set_mark(proto); // until proven otherwise 7801 assert(cur->is_oop(), "Should be an oop"); 7802 bool res = stack->push(cur); 7803 assert(res, "Bit off more than can chew?"); 7804 NOT_PRODUCT(n++;) 7805 } 7806 _overflow_list = cur; 7807 #ifndef PRODUCT 7808 assert(_num_par_pushes >= n, "Too many pops?"); 7809 _num_par_pushes -=n; 7810 #endif 7811 return !stack->isEmpty(); 7812 } 7813 7814 #define BUSY (cast_to_oop<intptr_t>(0x1aff1aff)) 7815 // (MT-safe) Get a prefix of at most "num" from the list. 7816 // The overflow list is chained through the mark word of 7817 // each object in the list. We fetch the entire list, 7818 // break off a prefix of the right size and return the 7819 // remainder. If other threads try to take objects from 7820 // the overflow list at that time, they will wait for 7821 // some time to see if data becomes available. If (and 7822 // only if) another thread places one or more object(s) 7823 // on the global list before we have returned the suffix 7824 // to the global list, we will walk down our local list 7825 // to find its end and append the global list to 7826 // our suffix before returning it. This suffix walk can 7827 // prove to be expensive (quadratic in the amount of traffic) 7828 // when there are many objects in the overflow list and 7829 // there is much producer-consumer contention on the list. 7830 // *NOTE*: The overflow list manipulation code here and 7831 // in ParNewGeneration:: are very similar in shape, 7832 // except that in the ParNew case we use the old (from/eden) 7833 // copy of the object to thread the list via its klass word. 7834 // Because of the common code, if you make any changes in 7835 // the code below, please check the ParNew version to see if 7836 // similar changes might be needed. 7837 // CR 6797058 has been filed to consolidate the common code. 7838 bool CMSCollector::par_take_from_overflow_list(size_t num, 7839 OopTaskQueue* work_q, 7840 int no_of_gc_threads) { 7841 assert(work_q->size() == 0, "First empty local work queue"); 7842 assert(num < work_q->max_elems(), "Can't bite more than we can chew"); 7843 if (_overflow_list == NULL) { 7844 return false; 7845 } 7846 // Grab the entire list; we'll put back a suffix 7847 oop prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list)); 7848 Thread* tid = Thread::current(); 7849 // Before "no_of_gc_threads" was introduced CMSOverflowSpinCount was 7850 // set to ParallelGCThreads. 7851 size_t CMSOverflowSpinCount = (size_t) no_of_gc_threads; // was ParallelGCThreads; 7852 size_t sleep_time_millis = MAX2((size_t)1, num/100); 7853 // If the list is busy, we spin for a short while, 7854 // sleeping between attempts to get the list. 7855 for (size_t spin = 0; prefix == BUSY && spin < CMSOverflowSpinCount; spin++) { 7856 os::sleep(tid, sleep_time_millis, false); 7857 if (_overflow_list == NULL) { 7858 // Nothing left to take 7859 return false; 7860 } else if (_overflow_list != BUSY) { 7861 // Try and grab the prefix 7862 prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list)); 7863 } 7864 } 7865 // If the list was found to be empty, or we spun long 7866 // enough, we give up and return empty-handed. If we leave 7867 // the list in the BUSY state below, it must be the case that 7868 // some other thread holds the overflow list and will set it 7869 // to a non-BUSY state in the future. 7870 if (prefix == NULL || prefix == BUSY) { 7871 // Nothing to take or waited long enough 7872 if (prefix == NULL) { 7873 // Write back the NULL in case we overwrote it with BUSY above 7874 // and it is still the same value. 7875 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY); 7876 } 7877 return false; 7878 } 7879 assert(prefix != NULL && prefix != BUSY, "Error"); 7880 size_t i = num; 7881 oop cur = prefix; 7882 // Walk down the first "num" objects, unless we reach the end. 7883 for (; i > 1 && cur->mark() != NULL; cur = oop(cur->mark()), i--); 7884 if (cur->mark() == NULL) { 7885 // We have "num" or fewer elements in the list, so there 7886 // is nothing to return to the global list. 7887 // Write back the NULL in lieu of the BUSY we wrote 7888 // above, if it is still the same value. 7889 if (_overflow_list == BUSY) { 7890 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY); 7891 } 7892 } else { 7893 // Chop off the suffix and return it to the global list. 7894 assert(cur->mark() != BUSY, "Error"); 7895 oop suffix_head = cur->mark(); // suffix will be put back on global list 7896 cur->set_mark(NULL); // break off suffix 7897 // It's possible that the list is still in the empty(busy) state 7898 // we left it in a short while ago; in that case we may be 7899 // able to place back the suffix without incurring the cost 7900 // of a walk down the list. 7901 oop observed_overflow_list = _overflow_list; 7902 oop cur_overflow_list = observed_overflow_list; 7903 bool attached = false; 7904 while (observed_overflow_list == BUSY || observed_overflow_list == NULL) { 7905 observed_overflow_list = 7906 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list); 7907 if (cur_overflow_list == observed_overflow_list) { 7908 attached = true; 7909 break; 7910 } else cur_overflow_list = observed_overflow_list; 7911 } 7912 if (!attached) { 7913 // Too bad, someone else sneaked in (at least) an element; we'll need 7914 // to do a splice. Find tail of suffix so we can prepend suffix to global 7915 // list. 7916 for (cur = suffix_head; cur->mark() != NULL; cur = (oop)(cur->mark())); 7917 oop suffix_tail = cur; 7918 assert(suffix_tail != NULL && suffix_tail->mark() == NULL, 7919 "Tautology"); 7920 observed_overflow_list = _overflow_list; 7921 do { 7922 cur_overflow_list = observed_overflow_list; 7923 if (cur_overflow_list != BUSY) { 7924 // Do the splice ... 7925 suffix_tail->set_mark(markOop(cur_overflow_list)); 7926 } else { // cur_overflow_list == BUSY 7927 suffix_tail->set_mark(NULL); 7928 } 7929 // ... and try to place spliced list back on overflow_list ... 7930 observed_overflow_list = 7931 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list); 7932 } while (cur_overflow_list != observed_overflow_list); 7933 // ... until we have succeeded in doing so. 7934 } 7935 } 7936 7937 // Push the prefix elements on work_q 7938 assert(prefix != NULL, "control point invariant"); 7939 const markOop proto = markOopDesc::prototype(); 7940 oop next; 7941 NOT_PRODUCT(ssize_t n = 0;) 7942 for (cur = prefix; cur != NULL; cur = next) { 7943 next = oop(cur->mark()); 7944 cur->set_mark(proto); // until proven otherwise 7945 assert(cur->is_oop(), "Should be an oop"); 7946 bool res = work_q->push(cur); 7947 assert(res, "Bit off more than we can chew?"); 7948 NOT_PRODUCT(n++;) 7949 } 7950 #ifndef PRODUCT 7951 assert(_num_par_pushes >= n, "Too many pops?"); 7952 Atomic::add_ptr(-(intptr_t)n, &_num_par_pushes); 7953 #endif 7954 return true; 7955 } 7956 7957 // Single-threaded 7958 void CMSCollector::push_on_overflow_list(oop p) { 7959 NOT_PRODUCT(_num_par_pushes++;) 7960 assert(p->is_oop(), "Not an oop"); 7961 preserve_mark_if_necessary(p); 7962 p->set_mark((markOop)_overflow_list); 7963 _overflow_list = p; 7964 } 7965 7966 // Multi-threaded; use CAS to prepend to overflow list 7967 void CMSCollector::par_push_on_overflow_list(oop p) { 7968 NOT_PRODUCT(Atomic::inc_ptr(&_num_par_pushes);) 7969 assert(p->is_oop(), "Not an oop"); 7970 par_preserve_mark_if_necessary(p); 7971 oop observed_overflow_list = _overflow_list; 7972 oop cur_overflow_list; 7973 do { 7974 cur_overflow_list = observed_overflow_list; 7975 if (cur_overflow_list != BUSY) { 7976 p->set_mark(markOop(cur_overflow_list)); 7977 } else { 7978 p->set_mark(NULL); 7979 } 7980 observed_overflow_list = 7981 (oop) Atomic::cmpxchg_ptr(p, &_overflow_list, cur_overflow_list); 7982 } while (cur_overflow_list != observed_overflow_list); 7983 } 7984 #undef BUSY 7985 7986 // Single threaded 7987 // General Note on GrowableArray: pushes may silently fail 7988 // because we are (temporarily) out of C-heap for expanding 7989 // the stack. The problem is quite ubiquitous and affects 7990 // a lot of code in the JVM. The prudent thing for GrowableArray 7991 // to do (for now) is to exit with an error. However, that may 7992 // be too draconian in some cases because the caller may be 7993 // able to recover without much harm. For such cases, we 7994 // should probably introduce a "soft_push" method which returns 7995 // an indication of success or failure with the assumption that 7996 // the caller may be able to recover from a failure; code in 7997 // the VM can then be changed, incrementally, to deal with such 7998 // failures where possible, thus, incrementally hardening the VM 7999 // in such low resource situations. 8000 void CMSCollector::preserve_mark_work(oop p, markOop m) { 8001 _preserved_oop_stack.push(p); 8002 _preserved_mark_stack.push(m); 8003 assert(m == p->mark(), "Mark word changed"); 8004 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), 8005 "bijection"); 8006 } 8007 8008 // Single threaded 8009 void CMSCollector::preserve_mark_if_necessary(oop p) { 8010 markOop m = p->mark(); 8011 if (m->must_be_preserved(p)) { 8012 preserve_mark_work(p, m); 8013 } 8014 } 8015 8016 void CMSCollector::par_preserve_mark_if_necessary(oop p) { 8017 markOop m = p->mark(); 8018 if (m->must_be_preserved(p)) { 8019 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag); 8020 // Even though we read the mark word without holding 8021 // the lock, we are assured that it will not change 8022 // because we "own" this oop, so no other thread can 8023 // be trying to push it on the overflow list; see 8024 // the assertion in preserve_mark_work() that checks 8025 // that m == p->mark(). 8026 preserve_mark_work(p, m); 8027 } 8028 } 8029 8030 // We should be able to do this multi-threaded, 8031 // a chunk of stack being a task (this is 8032 // correct because each oop only ever appears 8033 // once in the overflow list. However, it's 8034 // not very easy to completely overlap this with 8035 // other operations, so will generally not be done 8036 // until all work's been completed. Because we 8037 // expect the preserved oop stack (set) to be small, 8038 // it's probably fine to do this single-threaded. 8039 // We can explore cleverer concurrent/overlapped/parallel 8040 // processing of preserved marks if we feel the 8041 // need for this in the future. Stack overflow should 8042 // be so rare in practice and, when it happens, its 8043 // effect on performance so great that this will 8044 // likely just be in the noise anyway. 8045 void CMSCollector::restore_preserved_marks_if_any() { 8046 assert(SafepointSynchronize::is_at_safepoint(), 8047 "world should be stopped"); 8048 assert(Thread::current()->is_ConcurrentGC_thread() || 8049 Thread::current()->is_VM_thread(), 8050 "should be single-threaded"); 8051 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), 8052 "bijection"); 8053 8054 while (!_preserved_oop_stack.is_empty()) { 8055 oop p = _preserved_oop_stack.pop(); 8056 assert(p->is_oop(), "Should be an oop"); 8057 assert(_span.contains(p), "oop should be in _span"); 8058 assert(p->mark() == markOopDesc::prototype(), 8059 "Set when taken from overflow list"); 8060 markOop m = _preserved_mark_stack.pop(); 8061 p->set_mark(m); 8062 } 8063 assert(_preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(), 8064 "stacks were cleared above"); 8065 } 8066 8067 #ifndef PRODUCT 8068 bool CMSCollector::no_preserved_marks() const { 8069 return _preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(); 8070 } 8071 #endif 8072 8073 // Transfer some number of overflown objects to usual marking 8074 // stack. Return true if some objects were transferred. 8075 bool MarkRefsIntoAndScanClosure::take_from_overflow_list() { 8076 size_t num = MIN2((size_t)(_mark_stack->capacity() - _mark_stack->length())/4, 8077 (size_t)ParGCDesiredObjsFromOverflowList); 8078 8079 bool res = _collector->take_from_overflow_list(num, _mark_stack); 8080 assert(_collector->overflow_list_is_empty() || res, 8081 "If list is not empty, we should have taken something"); 8082 assert(!res || !_mark_stack->isEmpty(), 8083 "If we took something, it should now be on our stack"); 8084 return res; 8085 } 8086 8087 size_t MarkDeadObjectsClosure::do_blk(HeapWord* addr) { 8088 size_t res = _sp->block_size_no_stall(addr, _collector); 8089 if (_sp->block_is_obj(addr)) { 8090 if (_live_bit_map->isMarked(addr)) { 8091 // It can't have been dead in a previous cycle 8092 guarantee(!_dead_bit_map->isMarked(addr), "No resurrection!"); 8093 } else { 8094 _dead_bit_map->mark(addr); // mark the dead object 8095 } 8096 } 8097 // Could be 0, if the block size could not be computed without stalling. 8098 return res; 8099 } 8100 8101 TraceCMSMemoryManagerStats::TraceCMSMemoryManagerStats(CMSCollector::CollectorState phase, GCCause::Cause cause): TraceMemoryManagerStats() { 8102 8103 switch (phase) { 8104 case CMSCollector::InitialMarking: 8105 initialize(true /* fullGC */ , 8106 cause /* cause of the GC */, 8107 true /* recordGCBeginTime */, 8108 true /* recordPreGCUsage */, 8109 false /* recordPeakUsage */, 8110 false /* recordPostGCusage */, 8111 true /* recordAccumulatedGCTime */, 8112 false /* recordGCEndTime */, 8113 false /* countCollection */ ); 8114 break; 8115 8116 case CMSCollector::FinalMarking: 8117 initialize(true /* fullGC */ , 8118 cause /* cause of the GC */, 8119 false /* recordGCBeginTime */, 8120 false /* recordPreGCUsage */, 8121 false /* recordPeakUsage */, 8122 false /* recordPostGCusage */, 8123 true /* recordAccumulatedGCTime */, 8124 false /* recordGCEndTime */, 8125 false /* countCollection */ ); 8126 break; 8127 8128 case CMSCollector::Sweeping: 8129 initialize(true /* fullGC */ , 8130 cause /* cause of the GC */, 8131 false /* recordGCBeginTime */, 8132 false /* recordPreGCUsage */, 8133 true /* recordPeakUsage */, 8134 true /* recordPostGCusage */, 8135 false /* recordAccumulatedGCTime */, 8136 true /* recordGCEndTime */, 8137 true /* countCollection */ ); 8138 break; 8139 8140 default: 8141 ShouldNotReachHere(); 8142 } 8143 }