1 /* 2 * Copyright (c) 2001, 2015, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "classfile/classLoaderData.hpp" 27 #include "classfile/stringTable.hpp" 28 #include "classfile/systemDictionary.hpp" 29 #include "code/codeCache.hpp" 30 #include "gc/cms/cmsCollectorPolicy.hpp" 31 #include "gc/cms/cmsOopClosures.inline.hpp" 32 #include "gc/cms/compactibleFreeListSpace.hpp" 33 #include "gc/cms/concurrentMarkSweepGeneration.inline.hpp" 34 #include "gc/cms/concurrentMarkSweepThread.hpp" 35 #include "gc/cms/parNewGeneration.hpp" 36 #include "gc/cms/vmCMSOperations.hpp" 37 #include "gc/serial/genMarkSweep.hpp" 38 #include "gc/serial/tenuredGeneration.hpp" 39 #include "gc/shared/adaptiveSizePolicy.hpp" 40 #include "gc/shared/cardGeneration.inline.hpp" 41 #include "gc/shared/cardTableRS.hpp" 42 #include "gc/shared/collectedHeap.inline.hpp" 43 #include "gc/shared/collectorCounters.hpp" 44 #include "gc/shared/collectorPolicy.hpp" 45 #include "gc/shared/gcLocker.inline.hpp" 46 #include "gc/shared/gcPolicyCounters.hpp" 47 #include "gc/shared/gcTimer.hpp" 48 #include "gc/shared/gcTrace.hpp" 49 #include "gc/shared/gcTraceTime.hpp" 50 #include "gc/shared/genCollectedHeap.hpp" 51 #include "gc/shared/genOopClosures.inline.hpp" 52 #include "gc/shared/isGCActiveMark.hpp" 53 #include "gc/shared/referencePolicy.hpp" 54 #include "gc/shared/strongRootsScope.hpp" 55 #include "gc/shared/taskqueue.inline.hpp" 56 #include "memory/allocation.hpp" 57 #include "memory/iterator.inline.hpp" 58 #include "memory/padded.hpp" 59 #include "memory/resourceArea.hpp" 60 #include "oops/oop.inline.hpp" 61 #include "prims/jvmtiExport.hpp" 62 #include "runtime/atomic.inline.hpp" 63 #include "runtime/globals_extension.hpp" 64 #include "runtime/handles.inline.hpp" 65 #include "runtime/java.hpp" 66 #include "runtime/orderAccess.inline.hpp" 67 #include "runtime/vmThread.hpp" 68 #include "services/memoryService.hpp" 69 #include "services/runtimeService.hpp" 70 #include "utilities/stack.inline.hpp" 71 72 // statics 73 CMSCollector* ConcurrentMarkSweepGeneration::_collector = NULL; 74 bool CMSCollector::_full_gc_requested = false; 75 GCCause::Cause CMSCollector::_full_gc_cause = GCCause::_no_gc; 76 77 ////////////////////////////////////////////////////////////////// 78 // In support of CMS/VM thread synchronization 79 ////////////////////////////////////////////////////////////////// 80 // We split use of the CGC_lock into 2 "levels". 81 // The low-level locking is of the usual CGC_lock monitor. We introduce 82 // a higher level "token" (hereafter "CMS token") built on top of the 83 // low level monitor (hereafter "CGC lock"). 84 // The token-passing protocol gives priority to the VM thread. The 85 // CMS-lock doesn't provide any fairness guarantees, but clients 86 // should ensure that it is only held for very short, bounded 87 // durations. 88 // 89 // When either of the CMS thread or the VM thread is involved in 90 // collection operations during which it does not want the other 91 // thread to interfere, it obtains the CMS token. 92 // 93 // If either thread tries to get the token while the other has 94 // it, that thread waits. However, if the VM thread and CMS thread 95 // both want the token, then the VM thread gets priority while the 96 // CMS thread waits. This ensures, for instance, that the "concurrent" 97 // phases of the CMS thread's work do not block out the VM thread 98 // for long periods of time as the CMS thread continues to hog 99 // the token. (See bug 4616232). 100 // 101 // The baton-passing functions are, however, controlled by the 102 // flags _foregroundGCShouldWait and _foregroundGCIsActive, 103 // and here the low-level CMS lock, not the high level token, 104 // ensures mutual exclusion. 105 // 106 // Two important conditions that we have to satisfy: 107 // 1. if a thread does a low-level wait on the CMS lock, then it 108 // relinquishes the CMS token if it were holding that token 109 // when it acquired the low-level CMS lock. 110 // 2. any low-level notifications on the low-level lock 111 // should only be sent when a thread has relinquished the token. 112 // 113 // In the absence of either property, we'd have potential deadlock. 114 // 115 // We protect each of the CMS (concurrent and sequential) phases 116 // with the CMS _token_, not the CMS _lock_. 117 // 118 // The only code protected by CMS lock is the token acquisition code 119 // itself, see ConcurrentMarkSweepThread::[de]synchronize(), and the 120 // baton-passing code. 121 // 122 // Unfortunately, i couldn't come up with a good abstraction to factor and 123 // hide the naked CGC_lock manipulation in the baton-passing code 124 // further below. That's something we should try to do. Also, the proof 125 // of correctness of this 2-level locking scheme is far from obvious, 126 // and potentially quite slippery. We have an uneasy suspicion, for instance, 127 // that there may be a theoretical possibility of delay/starvation in the 128 // low-level lock/wait/notify scheme used for the baton-passing because of 129 // potential interference with the priority scheme embodied in the 130 // CMS-token-passing protocol. See related comments at a CGC_lock->wait() 131 // invocation further below and marked with "XXX 20011219YSR". 132 // Indeed, as we note elsewhere, this may become yet more slippery 133 // in the presence of multiple CMS and/or multiple VM threads. XXX 134 135 class CMSTokenSync: public StackObj { 136 private: 137 bool _is_cms_thread; 138 public: 139 CMSTokenSync(bool is_cms_thread): 140 _is_cms_thread(is_cms_thread) { 141 assert(is_cms_thread == Thread::current()->is_ConcurrentGC_thread(), 142 "Incorrect argument to constructor"); 143 ConcurrentMarkSweepThread::synchronize(_is_cms_thread); 144 } 145 146 ~CMSTokenSync() { 147 assert(_is_cms_thread ? 148 ConcurrentMarkSweepThread::cms_thread_has_cms_token() : 149 ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 150 "Incorrect state"); 151 ConcurrentMarkSweepThread::desynchronize(_is_cms_thread); 152 } 153 }; 154 155 // Convenience class that does a CMSTokenSync, and then acquires 156 // upto three locks. 157 class CMSTokenSyncWithLocks: public CMSTokenSync { 158 private: 159 // Note: locks are acquired in textual declaration order 160 // and released in the opposite order 161 MutexLockerEx _locker1, _locker2, _locker3; 162 public: 163 CMSTokenSyncWithLocks(bool is_cms_thread, Mutex* mutex1, 164 Mutex* mutex2 = NULL, Mutex* mutex3 = NULL): 165 CMSTokenSync(is_cms_thread), 166 _locker1(mutex1, Mutex::_no_safepoint_check_flag), 167 _locker2(mutex2, Mutex::_no_safepoint_check_flag), 168 _locker3(mutex3, Mutex::_no_safepoint_check_flag) 169 { } 170 }; 171 172 173 ////////////////////////////////////////////////////////////////// 174 // Concurrent Mark-Sweep Generation ///////////////////////////// 175 ////////////////////////////////////////////////////////////////// 176 177 NOT_PRODUCT(CompactibleFreeListSpace* debug_cms_space;) 178 179 // This struct contains per-thread things necessary to support parallel 180 // young-gen collection. 181 class CMSParGCThreadState: public CHeapObj<mtGC> { 182 public: 183 CFLS_LAB lab; 184 PromotionInfo promo; 185 186 // Constructor. 187 CMSParGCThreadState(CompactibleFreeListSpace* cfls) : lab(cfls) { 188 promo.setSpace(cfls); 189 } 190 }; 191 192 ConcurrentMarkSweepGeneration::ConcurrentMarkSweepGeneration( 193 ReservedSpace rs, size_t initial_byte_size, CardTableRS* ct) : 194 CardGeneration(rs, initial_byte_size, ct), 195 _dilatation_factor(((double)MinChunkSize)/((double)(CollectedHeap::min_fill_size()))), 196 _did_compact(false) 197 { 198 HeapWord* bottom = (HeapWord*) _virtual_space.low(); 199 HeapWord* end = (HeapWord*) _virtual_space.high(); 200 201 _direct_allocated_words = 0; 202 NOT_PRODUCT( 203 _numObjectsPromoted = 0; 204 _numWordsPromoted = 0; 205 _numObjectsAllocated = 0; 206 _numWordsAllocated = 0; 207 ) 208 209 _cmsSpace = new CompactibleFreeListSpace(_bts, MemRegion(bottom, end)); 210 NOT_PRODUCT(debug_cms_space = _cmsSpace;) 211 _cmsSpace->_old_gen = this; 212 213 _gc_stats = new CMSGCStats(); 214 215 // Verify the assumption that FreeChunk::_prev and OopDesc::_klass 216 // offsets match. The ability to tell free chunks from objects 217 // depends on this property. 218 debug_only( 219 FreeChunk* junk = NULL; 220 assert(UseCompressedClassPointers || 221 junk->prev_addr() == (void*)(oop(junk)->klass_addr()), 222 "Offset of FreeChunk::_prev within FreeChunk must match" 223 " that of OopDesc::_klass within OopDesc"); 224 ) 225 226 _par_gc_thread_states = NEW_C_HEAP_ARRAY(CMSParGCThreadState*, ParallelGCThreads, mtGC); 227 for (uint i = 0; i < ParallelGCThreads; i++) { 228 _par_gc_thread_states[i] = new CMSParGCThreadState(cmsSpace()); 229 } 230 231 _incremental_collection_failed = false; 232 // The "dilatation_factor" is the expansion that can occur on 233 // account of the fact that the minimum object size in the CMS 234 // generation may be larger than that in, say, a contiguous young 235 // generation. 236 // Ideally, in the calculation below, we'd compute the dilatation 237 // factor as: MinChunkSize/(promoting_gen's min object size) 238 // Since we do not have such a general query interface for the 239 // promoting generation, we'll instead just use the minimum 240 // object size (which today is a header's worth of space); 241 // note that all arithmetic is in units of HeapWords. 242 assert(MinChunkSize >= CollectedHeap::min_fill_size(), "just checking"); 243 assert(_dilatation_factor >= 1.0, "from previous assert"); 244 } 245 246 247 // The field "_initiating_occupancy" represents the occupancy percentage 248 // at which we trigger a new collection cycle. Unless explicitly specified 249 // via CMSInitiatingOccupancyFraction (argument "io" below), it 250 // is calculated by: 251 // 252 // Let "f" be MinHeapFreeRatio in 253 // 254 // _initiating_occupancy = 100-f + 255 // f * (CMSTriggerRatio/100) 256 // where CMSTriggerRatio is the argument "tr" below. 257 // 258 // That is, if we assume the heap is at its desired maximum occupancy at the 259 // end of a collection, we let CMSTriggerRatio of the (purported) free 260 // space be allocated before initiating a new collection cycle. 261 // 262 void ConcurrentMarkSweepGeneration::init_initiating_occupancy(intx io, uintx tr) { 263 assert(io <= 100 && tr <= 100, "Check the arguments"); 264 if (io >= 0) { 265 _initiating_occupancy = (double)io / 100.0; 266 } else { 267 _initiating_occupancy = ((100 - MinHeapFreeRatio) + 268 (double)(tr * MinHeapFreeRatio) / 100.0) 269 / 100.0; 270 } 271 } 272 273 void ConcurrentMarkSweepGeneration::ref_processor_init() { 274 assert(collector() != NULL, "no collector"); 275 collector()->ref_processor_init(); 276 } 277 278 void CMSCollector::ref_processor_init() { 279 if (_ref_processor == NULL) { 280 // Allocate and initialize a reference processor 281 _ref_processor = 282 new ReferenceProcessor(_span, // span 283 (ParallelGCThreads > 1) && ParallelRefProcEnabled, // mt processing 284 ParallelGCThreads, // mt processing degree 285 _cmsGen->refs_discovery_is_mt(), // mt discovery 286 MAX2(ConcGCThreads, ParallelGCThreads), // mt discovery degree 287 _cmsGen->refs_discovery_is_atomic(), // discovery is not atomic 288 &_is_alive_closure); // closure for liveness info 289 // Initialize the _ref_processor field of CMSGen 290 _cmsGen->set_ref_processor(_ref_processor); 291 292 } 293 } 294 295 AdaptiveSizePolicy* CMSCollector::size_policy() { 296 GenCollectedHeap* gch = GenCollectedHeap::heap(); 297 return gch->gen_policy()->size_policy(); 298 } 299 300 void ConcurrentMarkSweepGeneration::initialize_performance_counters() { 301 302 const char* gen_name = "old"; 303 GenCollectorPolicy* gcp = GenCollectedHeap::heap()->gen_policy(); 304 // Generation Counters - generation 1, 1 subspace 305 _gen_counters = new GenerationCounters(gen_name, 1, 1, 306 gcp->min_old_size(), gcp->max_old_size(), &_virtual_space); 307 308 _space_counters = new GSpaceCounters(gen_name, 0, 309 _virtual_space.reserved_size(), 310 this, _gen_counters); 311 } 312 313 CMSStats::CMSStats(ConcurrentMarkSweepGeneration* cms_gen, unsigned int alpha): 314 _cms_gen(cms_gen) 315 { 316 assert(alpha <= 100, "bad value"); 317 _saved_alpha = alpha; 318 319 // Initialize the alphas to the bootstrap value of 100. 320 _gc0_alpha = _cms_alpha = 100; 321 322 _cms_begin_time.update(); 323 _cms_end_time.update(); 324 325 _gc0_duration = 0.0; 326 _gc0_period = 0.0; 327 _gc0_promoted = 0; 328 329 _cms_duration = 0.0; 330 _cms_period = 0.0; 331 _cms_allocated = 0; 332 333 _cms_used_at_gc0_begin = 0; 334 _cms_used_at_gc0_end = 0; 335 _allow_duty_cycle_reduction = false; 336 _valid_bits = 0; 337 } 338 339 double CMSStats::cms_free_adjustment_factor(size_t free) const { 340 // TBD: CR 6909490 341 return 1.0; 342 } 343 344 void CMSStats::adjust_cms_free_adjustment_factor(bool fail, size_t free) { 345 } 346 347 // If promotion failure handling is on use 348 // the padded average size of the promotion for each 349 // young generation collection. 350 double CMSStats::time_until_cms_gen_full() const { 351 size_t cms_free = _cms_gen->cmsSpace()->free(); 352 GenCollectedHeap* gch = GenCollectedHeap::heap(); 353 size_t expected_promotion = MIN2(gch->young_gen()->capacity(), 354 (size_t) _cms_gen->gc_stats()->avg_promoted()->padded_average()); 355 if (cms_free > expected_promotion) { 356 // Start a cms collection if there isn't enough space to promote 357 // for the next young collection. Use the padded average as 358 // a safety factor. 359 cms_free -= expected_promotion; 360 361 // Adjust by the safety factor. 362 double cms_free_dbl = (double)cms_free; 363 double cms_adjustment = (100.0 - CMSIncrementalSafetyFactor) / 100.0; 364 // Apply a further correction factor which tries to adjust 365 // for recent occurance of concurrent mode failures. 366 cms_adjustment = cms_adjustment * cms_free_adjustment_factor(cms_free); 367 cms_free_dbl = cms_free_dbl * cms_adjustment; 368 369 if (PrintGCDetails && Verbose) { 370 gclog_or_tty->print_cr("CMSStats::time_until_cms_gen_full: cms_free " 371 SIZE_FORMAT " expected_promotion " SIZE_FORMAT, 372 cms_free, expected_promotion); 373 gclog_or_tty->print_cr(" cms_free_dbl %f cms_consumption_rate %f", 374 cms_free_dbl, cms_consumption_rate() + 1.0); 375 } 376 // Add 1 in case the consumption rate goes to zero. 377 return cms_free_dbl / (cms_consumption_rate() + 1.0); 378 } 379 return 0.0; 380 } 381 382 // Compare the duration of the cms collection to the 383 // time remaining before the cms generation is empty. 384 // Note that the time from the start of the cms collection 385 // to the start of the cms sweep (less than the total 386 // duration of the cms collection) can be used. This 387 // has been tried and some applications experienced 388 // promotion failures early in execution. This was 389 // possibly because the averages were not accurate 390 // enough at the beginning. 391 double CMSStats::time_until_cms_start() const { 392 // We add "gc0_period" to the "work" calculation 393 // below because this query is done (mostly) at the 394 // end of a scavenge, so we need to conservatively 395 // account for that much possible delay 396 // in the query so as to avoid concurrent mode failures 397 // due to starting the collection just a wee bit too 398 // late. 399 double work = cms_duration() + gc0_period(); 400 double deadline = time_until_cms_gen_full(); 401 // If a concurrent mode failure occurred recently, we want to be 402 // more conservative and halve our expected time_until_cms_gen_full() 403 if (work > deadline) { 404 if (Verbose && PrintGCDetails) { 405 gclog_or_tty->print( 406 " CMSCollector: collect because of anticipated promotion " 407 "before full %3.7f + %3.7f > %3.7f ", cms_duration(), 408 gc0_period(), time_until_cms_gen_full()); 409 } 410 return 0.0; 411 } 412 return work - deadline; 413 } 414 415 #ifndef PRODUCT 416 void CMSStats::print_on(outputStream *st) const { 417 st->print(" gc0_alpha=%d,cms_alpha=%d", _gc0_alpha, _cms_alpha); 418 st->print(",gc0_dur=%g,gc0_per=%g,gc0_promo=" SIZE_FORMAT, 419 gc0_duration(), gc0_period(), gc0_promoted()); 420 st->print(",cms_dur=%g,cms_per=%g,cms_alloc=" SIZE_FORMAT, 421 cms_duration(), cms_period(), cms_allocated()); 422 st->print(",cms_since_beg=%g,cms_since_end=%g", 423 cms_time_since_begin(), cms_time_since_end()); 424 st->print(",cms_used_beg=" SIZE_FORMAT ",cms_used_end=" SIZE_FORMAT, 425 _cms_used_at_gc0_begin, _cms_used_at_gc0_end); 426 427 if (valid()) { 428 st->print(",promo_rate=%g,cms_alloc_rate=%g", 429 promotion_rate(), cms_allocation_rate()); 430 st->print(",cms_consumption_rate=%g,time_until_full=%g", 431 cms_consumption_rate(), time_until_cms_gen_full()); 432 } 433 st->print(" "); 434 } 435 #endif // #ifndef PRODUCT 436 437 CMSCollector::CollectorState CMSCollector::_collectorState = 438 CMSCollector::Idling; 439 bool CMSCollector::_foregroundGCIsActive = false; 440 bool CMSCollector::_foregroundGCShouldWait = false; 441 442 CMSCollector::CMSCollector(ConcurrentMarkSweepGeneration* cmsGen, 443 CardTableRS* ct, 444 ConcurrentMarkSweepPolicy* cp): 445 _cmsGen(cmsGen), 446 _ct(ct), 447 _ref_processor(NULL), // will be set later 448 _conc_workers(NULL), // may be set later 449 _abort_preclean(false), 450 _start_sampling(false), 451 _between_prologue_and_epilogue(false), 452 _markBitMap(0, Mutex::leaf + 1, "CMS_markBitMap_lock"), 453 _modUnionTable((CardTableModRefBS::card_shift - LogHeapWordSize), 454 -1 /* lock-free */, "No_lock" /* dummy */), 455 _modUnionClosurePar(&_modUnionTable), 456 // Adjust my span to cover old (cms) gen 457 _span(cmsGen->reserved()), 458 // Construct the is_alive_closure with _span & markBitMap 459 _is_alive_closure(_span, &_markBitMap), 460 _restart_addr(NULL), 461 _overflow_list(NULL), 462 _stats(cmsGen), 463 _eden_chunk_lock(new Mutex(Mutex::leaf + 1, "CMS_eden_chunk_lock", true, 464 //verify that this lock should be acquired with safepoint check. 465 Monitor::_safepoint_check_sometimes)), 466 _eden_chunk_array(NULL), // may be set in ctor body 467 _eden_chunk_capacity(0), // -- ditto -- 468 _eden_chunk_index(0), // -- ditto -- 469 _survivor_plab_array(NULL), // -- ditto -- 470 _survivor_chunk_array(NULL), // -- ditto -- 471 _survivor_chunk_capacity(0), // -- ditto -- 472 _survivor_chunk_index(0), // -- ditto -- 473 _ser_pmc_preclean_ovflw(0), 474 _ser_kac_preclean_ovflw(0), 475 _ser_pmc_remark_ovflw(0), 476 _par_pmc_remark_ovflw(0), 477 _ser_kac_ovflw(0), 478 _par_kac_ovflw(0), 479 #ifndef PRODUCT 480 _num_par_pushes(0), 481 #endif 482 _collection_count_start(0), 483 _verifying(false), 484 _verification_mark_bm(0, Mutex::leaf + 1, "CMS_verification_mark_bm_lock"), 485 _completed_initialization(false), 486 _collector_policy(cp), 487 _should_unload_classes(CMSClassUnloadingEnabled), 488 _concurrent_cycles_since_last_unload(0), 489 _roots_scanning_options(GenCollectedHeap::SO_None), 490 _inter_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding), 491 _intra_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding), 492 _gc_tracer_cm(new (ResourceObj::C_HEAP, mtGC) CMSTracer()), 493 _gc_timer_cm(new (ResourceObj::C_HEAP, mtGC) ConcurrentGCTimer()), 494 _cms_start_registered(false) 495 { 496 if (ExplicitGCInvokesConcurrentAndUnloadsClasses) { 497 ExplicitGCInvokesConcurrent = true; 498 } 499 // Now expand the span and allocate the collection support structures 500 // (MUT, marking bit map etc.) to cover both generations subject to 501 // collection. 502 503 // For use by dirty card to oop closures. 504 _cmsGen->cmsSpace()->set_collector(this); 505 506 // Allocate MUT and marking bit map 507 { 508 MutexLockerEx x(_markBitMap.lock(), Mutex::_no_safepoint_check_flag); 509 if (!_markBitMap.allocate(_span)) { 510 warning("Failed to allocate CMS Bit Map"); 511 return; 512 } 513 assert(_markBitMap.covers(_span), "_markBitMap inconsistency?"); 514 } 515 { 516 _modUnionTable.allocate(_span); 517 assert(_modUnionTable.covers(_span), "_modUnionTable inconsistency?"); 518 } 519 520 if (!_markStack.allocate(MarkStackSize)) { 521 warning("Failed to allocate CMS Marking Stack"); 522 return; 523 } 524 525 // Support for multi-threaded concurrent phases 526 if (CMSConcurrentMTEnabled) { 527 if (FLAG_IS_DEFAULT(ConcGCThreads)) { 528 // just for now 529 FLAG_SET_DEFAULT(ConcGCThreads, (ParallelGCThreads + 3) / 4); 530 } 531 if (ConcGCThreads > 1) { 532 _conc_workers = new YieldingFlexibleWorkGang("CMS Thread", 533 ConcGCThreads, true); 534 if (_conc_workers == NULL) { 535 warning("GC/CMS: _conc_workers allocation failure: " 536 "forcing -CMSConcurrentMTEnabled"); 537 CMSConcurrentMTEnabled = false; 538 } else { 539 _conc_workers->initialize_workers(); 540 } 541 } else { 542 CMSConcurrentMTEnabled = false; 543 } 544 } 545 if (!CMSConcurrentMTEnabled) { 546 ConcGCThreads = 0; 547 } else { 548 // Turn off CMSCleanOnEnter optimization temporarily for 549 // the MT case where it's not fixed yet; see 6178663. 550 CMSCleanOnEnter = false; 551 } 552 assert((_conc_workers != NULL) == (ConcGCThreads > 1), 553 "Inconsistency"); 554 555 // Parallel task queues; these are shared for the 556 // concurrent and stop-world phases of CMS, but 557 // are not shared with parallel scavenge (ParNew). 558 { 559 uint i; 560 uint num_queues = MAX2(ParallelGCThreads, ConcGCThreads); 561 562 if ((CMSParallelRemarkEnabled || CMSConcurrentMTEnabled 563 || ParallelRefProcEnabled) 564 && num_queues > 0) { 565 _task_queues = new OopTaskQueueSet(num_queues); 566 if (_task_queues == NULL) { 567 warning("task_queues allocation failure."); 568 return; 569 } 570 _hash_seed = NEW_C_HEAP_ARRAY(int, num_queues, mtGC); 571 typedef Padded<OopTaskQueue> PaddedOopTaskQueue; 572 for (i = 0; i < num_queues; i++) { 573 PaddedOopTaskQueue *q = new PaddedOopTaskQueue(); 574 if (q == NULL) { 575 warning("work_queue allocation failure."); 576 return; 577 } 578 _task_queues->register_queue(i, q); 579 } 580 for (i = 0; i < num_queues; i++) { 581 _task_queues->queue(i)->initialize(); 582 _hash_seed[i] = 17; // copied from ParNew 583 } 584 } 585 } 586 587 _cmsGen ->init_initiating_occupancy(CMSInitiatingOccupancyFraction, CMSTriggerRatio); 588 589 // Clip CMSBootstrapOccupancy between 0 and 100. 590 _bootstrap_occupancy = CMSBootstrapOccupancy / 100.0; 591 592 // Now tell CMS generations the identity of their collector 593 ConcurrentMarkSweepGeneration::set_collector(this); 594 595 // Create & start a CMS thread for this CMS collector 596 _cmsThread = ConcurrentMarkSweepThread::start(this); 597 assert(cmsThread() != NULL, "CMS Thread should have been created"); 598 assert(cmsThread()->collector() == this, 599 "CMS Thread should refer to this gen"); 600 assert(CGC_lock != NULL, "Where's the CGC_lock?"); 601 602 // Support for parallelizing young gen rescan 603 GenCollectedHeap* gch = GenCollectedHeap::heap(); 604 assert(gch->young_gen()->kind() == Generation::ParNew, "CMS can only be used with ParNew"); 605 _young_gen = (ParNewGeneration*)gch->young_gen(); 606 if (gch->supports_inline_contig_alloc()) { 607 _top_addr = gch->top_addr(); 608 _end_addr = gch->end_addr(); 609 assert(_young_gen != NULL, "no _young_gen"); 610 _eden_chunk_index = 0; 611 _eden_chunk_capacity = (_young_gen->max_capacity() + CMSSamplingGrain) / CMSSamplingGrain; 612 _eden_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, _eden_chunk_capacity, mtGC); 613 } 614 615 // Support for parallelizing survivor space rescan 616 if ((CMSParallelRemarkEnabled && CMSParallelSurvivorRemarkEnabled) || CMSParallelInitialMarkEnabled) { 617 const size_t max_plab_samples = 618 _young_gen->max_survivor_size() / (PLAB::min_size() * HeapWordSize); 619 620 _survivor_plab_array = NEW_C_HEAP_ARRAY(ChunkArray, ParallelGCThreads, mtGC); 621 _survivor_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples, mtGC); 622 _cursor = NEW_C_HEAP_ARRAY(size_t, ParallelGCThreads, mtGC); 623 _survivor_chunk_capacity = max_plab_samples; 624 for (uint i = 0; i < ParallelGCThreads; i++) { 625 HeapWord** vec = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples, mtGC); 626 ChunkArray* cur = ::new (&_survivor_plab_array[i]) ChunkArray(vec, max_plab_samples); 627 assert(cur->end() == 0, "Should be 0"); 628 assert(cur->array() == vec, "Should be vec"); 629 assert(cur->capacity() == max_plab_samples, "Error"); 630 } 631 } 632 633 NOT_PRODUCT(_overflow_counter = CMSMarkStackOverflowInterval;) 634 _gc_counters = new CollectorCounters("CMS", 1); 635 _completed_initialization = true; 636 _inter_sweep_timer.start(); // start of time 637 } 638 639 const char* ConcurrentMarkSweepGeneration::name() const { 640 return "concurrent mark-sweep generation"; 641 } 642 void ConcurrentMarkSweepGeneration::update_counters() { 643 if (UsePerfData) { 644 _space_counters->update_all(); 645 _gen_counters->update_all(); 646 } 647 } 648 649 // this is an optimized version of update_counters(). it takes the 650 // used value as a parameter rather than computing it. 651 // 652 void ConcurrentMarkSweepGeneration::update_counters(size_t used) { 653 if (UsePerfData) { 654 _space_counters->update_used(used); 655 _space_counters->update_capacity(); 656 _gen_counters->update_all(); 657 } 658 } 659 660 void ConcurrentMarkSweepGeneration::print() const { 661 Generation::print(); 662 cmsSpace()->print(); 663 } 664 665 #ifndef PRODUCT 666 void ConcurrentMarkSweepGeneration::print_statistics() { 667 cmsSpace()->printFLCensus(0); 668 } 669 #endif 670 671 void ConcurrentMarkSweepGeneration::printOccupancy(const char *s) { 672 GenCollectedHeap* gch = GenCollectedHeap::heap(); 673 if (PrintGCDetails) { 674 // I didn't want to change the logging when removing the level concept, 675 // but I guess this logging could say "old" or something instead of "1". 676 assert(gch->is_old_gen(this), 677 "The CMS generation should be the old generation"); 678 uint level = 1; 679 if (Verbose) { 680 gclog_or_tty->print("[%u %s-%s: " SIZE_FORMAT "(" SIZE_FORMAT ")]", 681 level, short_name(), s, used(), capacity()); 682 } else { 683 gclog_or_tty->print("[%u %s-%s: " SIZE_FORMAT "K(" SIZE_FORMAT "K)]", 684 level, short_name(), s, used() / K, capacity() / K); 685 } 686 } 687 if (Verbose) { 688 gclog_or_tty->print(" " SIZE_FORMAT "(" SIZE_FORMAT ")", 689 gch->used(), gch->capacity()); 690 } else { 691 gclog_or_tty->print(" " SIZE_FORMAT "K(" SIZE_FORMAT "K)", 692 gch->used() / K, gch->capacity() / K); 693 } 694 } 695 696 size_t 697 ConcurrentMarkSweepGeneration::contiguous_available() const { 698 // dld proposes an improvement in precision here. If the committed 699 // part of the space ends in a free block we should add that to 700 // uncommitted size in the calculation below. Will make this 701 // change later, staying with the approximation below for the 702 // time being. -- ysr. 703 return MAX2(_virtual_space.uncommitted_size(), unsafe_max_alloc_nogc()); 704 } 705 706 size_t 707 ConcurrentMarkSweepGeneration::unsafe_max_alloc_nogc() const { 708 return _cmsSpace->max_alloc_in_words() * HeapWordSize; 709 } 710 711 size_t ConcurrentMarkSweepGeneration::max_available() const { 712 return free() + _virtual_space.uncommitted_size(); 713 } 714 715 bool ConcurrentMarkSweepGeneration::promotion_attempt_is_safe(size_t max_promotion_in_bytes) const { 716 size_t available = max_available(); 717 size_t av_promo = (size_t)gc_stats()->avg_promoted()->padded_average(); 718 bool res = (available >= av_promo) || (available >= max_promotion_in_bytes); 719 if (Verbose && PrintGCDetails) { 720 gclog_or_tty->print_cr( 721 "CMS: promo attempt is%s safe: available(" SIZE_FORMAT ") %s av_promo(" SIZE_FORMAT ")," 722 "max_promo(" SIZE_FORMAT ")", 723 res? "":" not", available, res? ">=":"<", 724 av_promo, max_promotion_in_bytes); 725 } 726 return res; 727 } 728 729 // At a promotion failure dump information on block layout in heap 730 // (cms old generation). 731 void ConcurrentMarkSweepGeneration::promotion_failure_occurred() { 732 if (CMSDumpAtPromotionFailure) { 733 cmsSpace()->dump_at_safepoint_with_locks(collector(), gclog_or_tty); 734 } 735 } 736 737 void ConcurrentMarkSweepGeneration::reset_after_compaction() { 738 // Clear the promotion information. These pointers can be adjusted 739 // along with all the other pointers into the heap but 740 // compaction is expected to be a rare event with 741 // a heap using cms so don't do it without seeing the need. 742 for (uint i = 0; i < ParallelGCThreads; i++) { 743 _par_gc_thread_states[i]->promo.reset(); 744 } 745 } 746 747 void ConcurrentMarkSweepGeneration::compute_new_size() { 748 assert_locked_or_safepoint(Heap_lock); 749 750 // If incremental collection failed, we just want to expand 751 // to the limit. 752 if (incremental_collection_failed()) { 753 clear_incremental_collection_failed(); 754 grow_to_reserved(); 755 return; 756 } 757 758 // The heap has been compacted but not reset yet. 759 // Any metric such as free() or used() will be incorrect. 760 761 CardGeneration::compute_new_size(); 762 763 // Reset again after a possible resizing 764 if (did_compact()) { 765 cmsSpace()->reset_after_compaction(); 766 } 767 } 768 769 void ConcurrentMarkSweepGeneration::compute_new_size_free_list() { 770 assert_locked_or_safepoint(Heap_lock); 771 772 // If incremental collection failed, we just want to expand 773 // to the limit. 774 if (incremental_collection_failed()) { 775 clear_incremental_collection_failed(); 776 grow_to_reserved(); 777 return; 778 } 779 780 double free_percentage = ((double) free()) / capacity(); 781 double desired_free_percentage = (double) MinHeapFreeRatio / 100; 782 double maximum_free_percentage = (double) MaxHeapFreeRatio / 100; 783 784 // compute expansion delta needed for reaching desired free percentage 785 if (free_percentage < desired_free_percentage) { 786 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 787 assert(desired_capacity >= capacity(), "invalid expansion size"); 788 size_t expand_bytes = MAX2(desired_capacity - capacity(), MinHeapDeltaBytes); 789 if (PrintGCDetails && Verbose) { 790 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 791 gclog_or_tty->print_cr("\nFrom compute_new_size: "); 792 gclog_or_tty->print_cr(" Free fraction %f", free_percentage); 793 gclog_or_tty->print_cr(" Desired free fraction %f", desired_free_percentage); 794 gclog_or_tty->print_cr(" Maximum free fraction %f", maximum_free_percentage); 795 gclog_or_tty->print_cr(" Capacity " SIZE_FORMAT, capacity() / 1000); 796 gclog_or_tty->print_cr(" Desired capacity " SIZE_FORMAT, desired_capacity / 1000); 797 GenCollectedHeap* gch = GenCollectedHeap::heap(); 798 assert(gch->is_old_gen(this), "The CMS generation should always be the old generation"); 799 size_t young_size = gch->young_gen()->capacity(); 800 gclog_or_tty->print_cr(" Young gen size " SIZE_FORMAT, young_size / 1000); 801 gclog_or_tty->print_cr(" unsafe_max_alloc_nogc " SIZE_FORMAT, unsafe_max_alloc_nogc() / 1000); 802 gclog_or_tty->print_cr(" contiguous available " SIZE_FORMAT, contiguous_available() / 1000); 803 gclog_or_tty->print_cr(" Expand by " SIZE_FORMAT " (bytes)", expand_bytes); 804 } 805 // safe if expansion fails 806 expand_for_gc_cause(expand_bytes, 0, CMSExpansionCause::_satisfy_free_ratio); 807 if (PrintGCDetails && Verbose) { 808 gclog_or_tty->print_cr(" Expanded free fraction %f", ((double) free()) / capacity()); 809 } 810 } else { 811 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 812 assert(desired_capacity <= capacity(), "invalid expansion size"); 813 size_t shrink_bytes = capacity() - desired_capacity; 814 // Don't shrink unless the delta is greater than the minimum shrink we want 815 if (shrink_bytes >= MinHeapDeltaBytes) { 816 shrink_free_list_by(shrink_bytes); 817 } 818 } 819 } 820 821 Mutex* ConcurrentMarkSweepGeneration::freelistLock() const { 822 return cmsSpace()->freelistLock(); 823 } 824 825 HeapWord* ConcurrentMarkSweepGeneration::allocate(size_t size, bool tlab) { 826 CMSSynchronousYieldRequest yr; 827 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 828 return have_lock_and_allocate(size, tlab); 829 } 830 831 HeapWord* ConcurrentMarkSweepGeneration::have_lock_and_allocate(size_t size, 832 bool tlab /* ignored */) { 833 assert_lock_strong(freelistLock()); 834 size_t adjustedSize = CompactibleFreeListSpace::adjustObjectSize(size); 835 HeapWord* res = cmsSpace()->allocate(adjustedSize); 836 // Allocate the object live (grey) if the background collector has 837 // started marking. This is necessary because the marker may 838 // have passed this address and consequently this object will 839 // not otherwise be greyed and would be incorrectly swept up. 840 // Note that if this object contains references, the writing 841 // of those references will dirty the card containing this object 842 // allowing the object to be blackened (and its references scanned) 843 // either during a preclean phase or at the final checkpoint. 844 if (res != NULL) { 845 // We may block here with an uninitialized object with 846 // its mark-bit or P-bits not yet set. Such objects need 847 // to be safely navigable by block_start(). 848 assert(oop(res)->klass_or_null() == NULL, "Object should be uninitialized here."); 849 assert(!((FreeChunk*)res)->is_free(), "Error, block will look free but show wrong size"); 850 collector()->direct_allocated(res, adjustedSize); 851 _direct_allocated_words += adjustedSize; 852 // allocation counters 853 NOT_PRODUCT( 854 _numObjectsAllocated++; 855 _numWordsAllocated += (int)adjustedSize; 856 ) 857 } 858 return res; 859 } 860 861 // In the case of direct allocation by mutators in a generation that 862 // is being concurrently collected, the object must be allocated 863 // live (grey) if the background collector has started marking. 864 // This is necessary because the marker may 865 // have passed this address and consequently this object will 866 // not otherwise be greyed and would be incorrectly swept up. 867 // Note that if this object contains references, the writing 868 // of those references will dirty the card containing this object 869 // allowing the object to be blackened (and its references scanned) 870 // either during a preclean phase or at the final checkpoint. 871 void CMSCollector::direct_allocated(HeapWord* start, size_t size) { 872 assert(_markBitMap.covers(start, size), "Out of bounds"); 873 if (_collectorState >= Marking) { 874 MutexLockerEx y(_markBitMap.lock(), 875 Mutex::_no_safepoint_check_flag); 876 // [see comments preceding SweepClosure::do_blk() below for details] 877 // 878 // Can the P-bits be deleted now? JJJ 879 // 880 // 1. need to mark the object as live so it isn't collected 881 // 2. need to mark the 2nd bit to indicate the object may be uninitialized 882 // 3. need to mark the end of the object so marking, precleaning or sweeping 883 // can skip over uninitialized or unparsable objects. An allocated 884 // object is considered uninitialized for our purposes as long as 885 // its klass word is NULL. All old gen objects are parsable 886 // as soon as they are initialized.) 887 _markBitMap.mark(start); // object is live 888 _markBitMap.mark(start + 1); // object is potentially uninitialized? 889 _markBitMap.mark(start + size - 1); 890 // mark end of object 891 } 892 // check that oop looks uninitialized 893 assert(oop(start)->klass_or_null() == NULL, "_klass should be NULL"); 894 } 895 896 void CMSCollector::promoted(bool par, HeapWord* start, 897 bool is_obj_array, size_t obj_size) { 898 assert(_markBitMap.covers(start), "Out of bounds"); 899 // See comment in direct_allocated() about when objects should 900 // be allocated live. 901 if (_collectorState >= Marking) { 902 // we already hold the marking bit map lock, taken in 903 // the prologue 904 if (par) { 905 _markBitMap.par_mark(start); 906 } else { 907 _markBitMap.mark(start); 908 } 909 // We don't need to mark the object as uninitialized (as 910 // in direct_allocated above) because this is being done with the 911 // world stopped and the object will be initialized by the 912 // time the marking, precleaning or sweeping get to look at it. 913 // But see the code for copying objects into the CMS generation, 914 // where we need to ensure that concurrent readers of the 915 // block offset table are able to safely navigate a block that 916 // is in flux from being free to being allocated (and in 917 // transition while being copied into) and subsequently 918 // becoming a bona-fide object when the copy/promotion is complete. 919 assert(SafepointSynchronize::is_at_safepoint(), 920 "expect promotion only at safepoints"); 921 922 if (_collectorState < Sweeping) { 923 // Mark the appropriate cards in the modUnionTable, so that 924 // this object gets scanned before the sweep. If this is 925 // not done, CMS generation references in the object might 926 // not get marked. 927 // For the case of arrays, which are otherwise precisely 928 // marked, we need to dirty the entire array, not just its head. 929 if (is_obj_array) { 930 // The [par_]mark_range() method expects mr.end() below to 931 // be aligned to the granularity of a bit's representation 932 // in the heap. In the case of the MUT below, that's a 933 // card size. 934 MemRegion mr(start, 935 (HeapWord*)round_to((intptr_t)(start + obj_size), 936 CardTableModRefBS::card_size /* bytes */)); 937 if (par) { 938 _modUnionTable.par_mark_range(mr); 939 } else { 940 _modUnionTable.mark_range(mr); 941 } 942 } else { // not an obj array; we can just mark the head 943 if (par) { 944 _modUnionTable.par_mark(start); 945 } else { 946 _modUnionTable.mark(start); 947 } 948 } 949 } 950 } 951 } 952 953 oop ConcurrentMarkSweepGeneration::promote(oop obj, size_t obj_size) { 954 assert(obj_size == (size_t)obj->size(), "bad obj_size passed in"); 955 // allocate, copy and if necessary update promoinfo -- 956 // delegate to underlying space. 957 assert_lock_strong(freelistLock()); 958 959 #ifndef PRODUCT 960 if (GenCollectedHeap::heap()->promotion_should_fail()) { 961 return NULL; 962 } 963 #endif // #ifndef PRODUCT 964 965 oop res = _cmsSpace->promote(obj, obj_size); 966 if (res == NULL) { 967 // expand and retry 968 size_t s = _cmsSpace->expansionSpaceRequired(obj_size); // HeapWords 969 expand_for_gc_cause(s*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_satisfy_promotion); 970 // Since this is the old generation, we don't try to promote 971 // into a more senior generation. 972 res = _cmsSpace->promote(obj, obj_size); 973 } 974 if (res != NULL) { 975 // See comment in allocate() about when objects should 976 // be allocated live. 977 assert(obj->is_oop(), "Will dereference klass pointer below"); 978 collector()->promoted(false, // Not parallel 979 (HeapWord*)res, obj->is_objArray(), obj_size); 980 // promotion counters 981 NOT_PRODUCT( 982 _numObjectsPromoted++; 983 _numWordsPromoted += 984 (int)(CompactibleFreeListSpace::adjustObjectSize(obj->size())); 985 ) 986 } 987 return res; 988 } 989 990 991 // IMPORTANT: Notes on object size recognition in CMS. 992 // --------------------------------------------------- 993 // A block of storage in the CMS generation is always in 994 // one of three states. A free block (FREE), an allocated 995 // object (OBJECT) whose size() method reports the correct size, 996 // and an intermediate state (TRANSIENT) in which its size cannot 997 // be accurately determined. 998 // STATE IDENTIFICATION: (32 bit and 64 bit w/o COOPS) 999 // ----------------------------------------------------- 1000 // FREE: klass_word & 1 == 1; mark_word holds block size 1001 // 1002 // OBJECT: klass_word installed; klass_word != 0 && klass_word & 1 == 0; 1003 // obj->size() computes correct size 1004 // 1005 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT 1006 // 1007 // STATE IDENTIFICATION: (64 bit+COOPS) 1008 // ------------------------------------ 1009 // FREE: mark_word & CMS_FREE_BIT == 1; mark_word & ~CMS_FREE_BIT gives block_size 1010 // 1011 // OBJECT: klass_word installed; klass_word != 0; 1012 // obj->size() computes correct size 1013 // 1014 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT 1015 // 1016 // 1017 // STATE TRANSITION DIAGRAM 1018 // 1019 // mut / parnew mut / parnew 1020 // FREE --------------------> TRANSIENT ---------------------> OBJECT --| 1021 // ^ | 1022 // |------------------------ DEAD <------------------------------------| 1023 // sweep mut 1024 // 1025 // While a block is in TRANSIENT state its size cannot be determined 1026 // so readers will either need to come back later or stall until 1027 // the size can be determined. Note that for the case of direct 1028 // allocation, P-bits, when available, may be used to determine the 1029 // size of an object that may not yet have been initialized. 1030 1031 // Things to support parallel young-gen collection. 1032 oop 1033 ConcurrentMarkSweepGeneration::par_promote(int thread_num, 1034 oop old, markOop m, 1035 size_t word_sz) { 1036 #ifndef PRODUCT 1037 if (GenCollectedHeap::heap()->promotion_should_fail()) { 1038 return NULL; 1039 } 1040 #endif // #ifndef PRODUCT 1041 1042 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1043 PromotionInfo* promoInfo = &ps->promo; 1044 // if we are tracking promotions, then first ensure space for 1045 // promotion (including spooling space for saving header if necessary). 1046 // then allocate and copy, then track promoted info if needed. 1047 // When tracking (see PromotionInfo::track()), the mark word may 1048 // be displaced and in this case restoration of the mark word 1049 // occurs in the (oop_since_save_marks_)iterate phase. 1050 if (promoInfo->tracking() && !promoInfo->ensure_spooling_space()) { 1051 // Out of space for allocating spooling buffers; 1052 // try expanding and allocating spooling buffers. 1053 if (!expand_and_ensure_spooling_space(promoInfo)) { 1054 return NULL; 1055 } 1056 } 1057 assert(promoInfo->has_spooling_space(), "Control point invariant"); 1058 const size_t alloc_sz = CompactibleFreeListSpace::adjustObjectSize(word_sz); 1059 HeapWord* obj_ptr = ps->lab.alloc(alloc_sz); 1060 if (obj_ptr == NULL) { 1061 obj_ptr = expand_and_par_lab_allocate(ps, alloc_sz); 1062 if (obj_ptr == NULL) { 1063 return NULL; 1064 } 1065 } 1066 oop obj = oop(obj_ptr); 1067 OrderAccess::storestore(); 1068 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1069 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1070 // IMPORTANT: See note on object initialization for CMS above. 1071 // Otherwise, copy the object. Here we must be careful to insert the 1072 // klass pointer last, since this marks the block as an allocated object. 1073 // Except with compressed oops it's the mark word. 1074 HeapWord* old_ptr = (HeapWord*)old; 1075 // Restore the mark word copied above. 1076 obj->set_mark(m); 1077 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1078 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1079 OrderAccess::storestore(); 1080 1081 if (UseCompressedClassPointers) { 1082 // Copy gap missed by (aligned) header size calculation below 1083 obj->set_klass_gap(old->klass_gap()); 1084 } 1085 if (word_sz > (size_t)oopDesc::header_size()) { 1086 Copy::aligned_disjoint_words(old_ptr + oopDesc::header_size(), 1087 obj_ptr + oopDesc::header_size(), 1088 word_sz - oopDesc::header_size()); 1089 } 1090 1091 // Now we can track the promoted object, if necessary. We take care 1092 // to delay the transition from uninitialized to full object 1093 // (i.e., insertion of klass pointer) until after, so that it 1094 // atomically becomes a promoted object. 1095 if (promoInfo->tracking()) { 1096 promoInfo->track((PromotedObject*)obj, old->klass()); 1097 } 1098 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1099 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1100 assert(old->is_oop(), "Will use and dereference old klass ptr below"); 1101 1102 // Finally, install the klass pointer (this should be volatile). 1103 OrderAccess::storestore(); 1104 obj->set_klass(old->klass()); 1105 // We should now be able to calculate the right size for this object 1106 assert(obj->is_oop() && obj->size() == (int)word_sz, "Error, incorrect size computed for promoted object"); 1107 1108 collector()->promoted(true, // parallel 1109 obj_ptr, old->is_objArray(), word_sz); 1110 1111 NOT_PRODUCT( 1112 Atomic::inc_ptr(&_numObjectsPromoted); 1113 Atomic::add_ptr(alloc_sz, &_numWordsPromoted); 1114 ) 1115 1116 return obj; 1117 } 1118 1119 void 1120 ConcurrentMarkSweepGeneration:: 1121 par_promote_alloc_done(int thread_num) { 1122 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1123 ps->lab.retire(thread_num); 1124 } 1125 1126 void 1127 ConcurrentMarkSweepGeneration:: 1128 par_oop_since_save_marks_iterate_done(int thread_num) { 1129 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1130 ParScanWithoutBarrierClosure* dummy_cl = NULL; 1131 ps->promo.promoted_oops_iterate_nv(dummy_cl); 1132 } 1133 1134 bool ConcurrentMarkSweepGeneration::should_collect(bool full, 1135 size_t size, 1136 bool tlab) 1137 { 1138 // We allow a STW collection only if a full 1139 // collection was requested. 1140 return full || should_allocate(size, tlab); // FIX ME !!! 1141 // This and promotion failure handling are connected at the 1142 // hip and should be fixed by untying them. 1143 } 1144 1145 bool CMSCollector::shouldConcurrentCollect() { 1146 if (_full_gc_requested) { 1147 if (Verbose && PrintGCDetails) { 1148 gclog_or_tty->print_cr("CMSCollector: collect because of explicit " 1149 " gc request (or gc_locker)"); 1150 } 1151 return true; 1152 } 1153 1154 FreelistLocker x(this); 1155 // ------------------------------------------------------------------ 1156 // Print out lots of information which affects the initiation of 1157 // a collection. 1158 if (PrintCMSInitiationStatistics && stats().valid()) { 1159 gclog_or_tty->print("CMSCollector shouldConcurrentCollect: "); 1160 gclog_or_tty->stamp(); 1161 gclog_or_tty->cr(); 1162 stats().print_on(gclog_or_tty); 1163 gclog_or_tty->print_cr("time_until_cms_gen_full %3.7f", 1164 stats().time_until_cms_gen_full()); 1165 gclog_or_tty->print_cr("free=" SIZE_FORMAT, _cmsGen->free()); 1166 gclog_or_tty->print_cr("contiguous_available=" SIZE_FORMAT, 1167 _cmsGen->contiguous_available()); 1168 gclog_or_tty->print_cr("promotion_rate=%g", stats().promotion_rate()); 1169 gclog_or_tty->print_cr("cms_allocation_rate=%g", stats().cms_allocation_rate()); 1170 gclog_or_tty->print_cr("occupancy=%3.7f", _cmsGen->occupancy()); 1171 gclog_or_tty->print_cr("initiatingOccupancy=%3.7f", _cmsGen->initiating_occupancy()); 1172 gclog_or_tty->print_cr("cms_time_since_begin=%3.7f", stats().cms_time_since_begin()); 1173 gclog_or_tty->print_cr("cms_time_since_end=%3.7f", stats().cms_time_since_end()); 1174 gclog_or_tty->print_cr("metadata initialized %d", 1175 MetaspaceGC::should_concurrent_collect()); 1176 } 1177 // ------------------------------------------------------------------ 1178 1179 // If the estimated time to complete a cms collection (cms_duration()) 1180 // is less than the estimated time remaining until the cms generation 1181 // is full, start a collection. 1182 if (!UseCMSInitiatingOccupancyOnly) { 1183 if (stats().valid()) { 1184 if (stats().time_until_cms_start() == 0.0) { 1185 return true; 1186 } 1187 } else { 1188 // We want to conservatively collect somewhat early in order 1189 // to try and "bootstrap" our CMS/promotion statistics; 1190 // this branch will not fire after the first successful CMS 1191 // collection because the stats should then be valid. 1192 if (_cmsGen->occupancy() >= _bootstrap_occupancy) { 1193 if (Verbose && PrintGCDetails) { 1194 gclog_or_tty->print_cr( 1195 " CMSCollector: collect for bootstrapping statistics:" 1196 " occupancy = %f, boot occupancy = %f", _cmsGen->occupancy(), 1197 _bootstrap_occupancy); 1198 } 1199 return true; 1200 } 1201 } 1202 } 1203 1204 // Otherwise, we start a collection cycle if 1205 // old gen want a collection cycle started. Each may use 1206 // an appropriate criterion for making this decision. 1207 // XXX We need to make sure that the gen expansion 1208 // criterion dovetails well with this. XXX NEED TO FIX THIS 1209 if (_cmsGen->should_concurrent_collect()) { 1210 if (Verbose && PrintGCDetails) { 1211 gclog_or_tty->print_cr("CMS old gen initiated"); 1212 } 1213 return true; 1214 } 1215 1216 // We start a collection if we believe an incremental collection may fail; 1217 // this is not likely to be productive in practice because it's probably too 1218 // late anyway. 1219 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1220 assert(gch->collector_policy()->is_generation_policy(), 1221 "You may want to check the correctness of the following"); 1222 if (gch->incremental_collection_will_fail(true /* consult_young */)) { 1223 if (Verbose && PrintGCDetails) { 1224 gclog_or_tty->print("CMSCollector: collect because incremental collection will fail "); 1225 } 1226 return true; 1227 } 1228 1229 if (MetaspaceGC::should_concurrent_collect()) { 1230 if (Verbose && PrintGCDetails) { 1231 gclog_or_tty->print("CMSCollector: collect for metadata allocation "); 1232 } 1233 return true; 1234 } 1235 1236 // CMSTriggerInterval starts a CMS cycle if enough time has passed. 1237 if (CMSTriggerInterval >= 0) { 1238 if (CMSTriggerInterval == 0) { 1239 // Trigger always 1240 return true; 1241 } 1242 1243 // Check the CMS time since begin (we do not check the stats validity 1244 // as we want to be able to trigger the first CMS cycle as well) 1245 if (stats().cms_time_since_begin() >= (CMSTriggerInterval / ((double) MILLIUNITS))) { 1246 if (Verbose && PrintGCDetails) { 1247 if (stats().valid()) { 1248 gclog_or_tty->print_cr("CMSCollector: collect because of trigger interval (time since last begin %3.7f secs)", 1249 stats().cms_time_since_begin()); 1250 } else { 1251 gclog_or_tty->print_cr("CMSCollector: collect because of trigger interval (first collection)"); 1252 } 1253 } 1254 return true; 1255 } 1256 } 1257 1258 return false; 1259 } 1260 1261 void CMSCollector::set_did_compact(bool v) { _cmsGen->set_did_compact(v); } 1262 1263 // Clear _expansion_cause fields of constituent generations 1264 void CMSCollector::clear_expansion_cause() { 1265 _cmsGen->clear_expansion_cause(); 1266 } 1267 1268 // We should be conservative in starting a collection cycle. To 1269 // start too eagerly runs the risk of collecting too often in the 1270 // extreme. To collect too rarely falls back on full collections, 1271 // which works, even if not optimum in terms of concurrent work. 1272 // As a work around for too eagerly collecting, use the flag 1273 // UseCMSInitiatingOccupancyOnly. This also has the advantage of 1274 // giving the user an easily understandable way of controlling the 1275 // collections. 1276 // We want to start a new collection cycle if any of the following 1277 // conditions hold: 1278 // . our current occupancy exceeds the configured initiating occupancy 1279 // for this generation, or 1280 // . we recently needed to expand this space and have not, since that 1281 // expansion, done a collection of this generation, or 1282 // . the underlying space believes that it may be a good idea to initiate 1283 // a concurrent collection (this may be based on criteria such as the 1284 // following: the space uses linear allocation and linear allocation is 1285 // going to fail, or there is believed to be excessive fragmentation in 1286 // the generation, etc... or ... 1287 // [.(currently done by CMSCollector::shouldConcurrentCollect() only for 1288 // the case of the old generation; see CR 6543076): 1289 // we may be approaching a point at which allocation requests may fail because 1290 // we will be out of sufficient free space given allocation rate estimates.] 1291 bool ConcurrentMarkSweepGeneration::should_concurrent_collect() const { 1292 1293 assert_lock_strong(freelistLock()); 1294 if (occupancy() > initiating_occupancy()) { 1295 if (PrintGCDetails && Verbose) { 1296 gclog_or_tty->print(" %s: collect because of occupancy %f / %f ", 1297 short_name(), occupancy(), initiating_occupancy()); 1298 } 1299 return true; 1300 } 1301 if (UseCMSInitiatingOccupancyOnly) { 1302 return false; 1303 } 1304 if (expansion_cause() == CMSExpansionCause::_satisfy_allocation) { 1305 if (PrintGCDetails && Verbose) { 1306 gclog_or_tty->print(" %s: collect because expanded for allocation ", 1307 short_name()); 1308 } 1309 return true; 1310 } 1311 return false; 1312 } 1313 1314 void ConcurrentMarkSweepGeneration::collect(bool full, 1315 bool clear_all_soft_refs, 1316 size_t size, 1317 bool tlab) 1318 { 1319 collector()->collect(full, clear_all_soft_refs, size, tlab); 1320 } 1321 1322 void CMSCollector::collect(bool full, 1323 bool clear_all_soft_refs, 1324 size_t size, 1325 bool tlab) 1326 { 1327 // The following "if" branch is present for defensive reasons. 1328 // In the current uses of this interface, it can be replaced with: 1329 // assert(!GC_locker.is_active(), "Can't be called otherwise"); 1330 // But I am not placing that assert here to allow future 1331 // generality in invoking this interface. 1332 if (GC_locker::is_active()) { 1333 // A consistency test for GC_locker 1334 assert(GC_locker::needs_gc(), "Should have been set already"); 1335 // Skip this foreground collection, instead 1336 // expanding the heap if necessary. 1337 // Need the free list locks for the call to free() in compute_new_size() 1338 compute_new_size(); 1339 return; 1340 } 1341 acquire_control_and_collect(full, clear_all_soft_refs); 1342 } 1343 1344 void CMSCollector::request_full_gc(unsigned int full_gc_count, GCCause::Cause cause) { 1345 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1346 unsigned int gc_count = gch->total_full_collections(); 1347 if (gc_count == full_gc_count) { 1348 MutexLockerEx y(CGC_lock, Mutex::_no_safepoint_check_flag); 1349 _full_gc_requested = true; 1350 _full_gc_cause = cause; 1351 CGC_lock->notify(); // nudge CMS thread 1352 } else { 1353 assert(gc_count > full_gc_count, "Error: causal loop"); 1354 } 1355 } 1356 1357 bool CMSCollector::is_external_interruption() { 1358 GCCause::Cause cause = GenCollectedHeap::heap()->gc_cause(); 1359 return GCCause::is_user_requested_gc(cause) || 1360 GCCause::is_serviceability_requested_gc(cause); 1361 } 1362 1363 void CMSCollector::report_concurrent_mode_interruption() { 1364 if (is_external_interruption()) { 1365 if (PrintGCDetails) { 1366 gclog_or_tty->print(" (concurrent mode interrupted)"); 1367 } 1368 } else { 1369 if (PrintGCDetails) { 1370 gclog_or_tty->print(" (concurrent mode failure)"); 1371 } 1372 _gc_tracer_cm->report_concurrent_mode_failure(); 1373 } 1374 } 1375 1376 1377 // The foreground and background collectors need to coordinate in order 1378 // to make sure that they do not mutually interfere with CMS collections. 1379 // When a background collection is active, 1380 // the foreground collector may need to take over (preempt) and 1381 // synchronously complete an ongoing collection. Depending on the 1382 // frequency of the background collections and the heap usage 1383 // of the application, this preemption can be seldom or frequent. 1384 // There are only certain 1385 // points in the background collection that the "collection-baton" 1386 // can be passed to the foreground collector. 1387 // 1388 // The foreground collector will wait for the baton before 1389 // starting any part of the collection. The foreground collector 1390 // will only wait at one location. 1391 // 1392 // The background collector will yield the baton before starting a new 1393 // phase of the collection (e.g., before initial marking, marking from roots, 1394 // precleaning, final re-mark, sweep etc.) This is normally done at the head 1395 // of the loop which switches the phases. The background collector does some 1396 // of the phases (initial mark, final re-mark) with the world stopped. 1397 // Because of locking involved in stopping the world, 1398 // the foreground collector should not block waiting for the background 1399 // collector when it is doing a stop-the-world phase. The background 1400 // collector will yield the baton at an additional point just before 1401 // it enters a stop-the-world phase. Once the world is stopped, the 1402 // background collector checks the phase of the collection. If the 1403 // phase has not changed, it proceeds with the collection. If the 1404 // phase has changed, it skips that phase of the collection. See 1405 // the comments on the use of the Heap_lock in collect_in_background(). 1406 // 1407 // Variable used in baton passing. 1408 // _foregroundGCIsActive - Set to true by the foreground collector when 1409 // it wants the baton. The foreground clears it when it has finished 1410 // the collection. 1411 // _foregroundGCShouldWait - Set to true by the background collector 1412 // when it is running. The foreground collector waits while 1413 // _foregroundGCShouldWait is true. 1414 // CGC_lock - monitor used to protect access to the above variables 1415 // and to notify the foreground and background collectors. 1416 // _collectorState - current state of the CMS collection. 1417 // 1418 // The foreground collector 1419 // acquires the CGC_lock 1420 // sets _foregroundGCIsActive 1421 // waits on the CGC_lock for _foregroundGCShouldWait to be false 1422 // various locks acquired in preparation for the collection 1423 // are released so as not to block the background collector 1424 // that is in the midst of a collection 1425 // proceeds with the collection 1426 // clears _foregroundGCIsActive 1427 // returns 1428 // 1429 // The background collector in a loop iterating on the phases of the 1430 // collection 1431 // acquires the CGC_lock 1432 // sets _foregroundGCShouldWait 1433 // if _foregroundGCIsActive is set 1434 // clears _foregroundGCShouldWait, notifies _CGC_lock 1435 // waits on _CGC_lock for _foregroundGCIsActive to become false 1436 // and exits the loop. 1437 // otherwise 1438 // proceed with that phase of the collection 1439 // if the phase is a stop-the-world phase, 1440 // yield the baton once more just before enqueueing 1441 // the stop-world CMS operation (executed by the VM thread). 1442 // returns after all phases of the collection are done 1443 // 1444 1445 void CMSCollector::acquire_control_and_collect(bool full, 1446 bool clear_all_soft_refs) { 1447 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint"); 1448 assert(!Thread::current()->is_ConcurrentGC_thread(), 1449 "shouldn't try to acquire control from self!"); 1450 1451 // Start the protocol for acquiring control of the 1452 // collection from the background collector (aka CMS thread). 1453 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 1454 "VM thread should have CMS token"); 1455 // Remember the possibly interrupted state of an ongoing 1456 // concurrent collection 1457 CollectorState first_state = _collectorState; 1458 1459 // Signal to a possibly ongoing concurrent collection that 1460 // we want to do a foreground collection. 1461 _foregroundGCIsActive = true; 1462 1463 // release locks and wait for a notify from the background collector 1464 // releasing the locks in only necessary for phases which 1465 // do yields to improve the granularity of the collection. 1466 assert_lock_strong(bitMapLock()); 1467 // We need to lock the Free list lock for the space that we are 1468 // currently collecting. 1469 assert(haveFreelistLocks(), "Must be holding free list locks"); 1470 bitMapLock()->unlock(); 1471 releaseFreelistLocks(); 1472 { 1473 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1474 if (_foregroundGCShouldWait) { 1475 // We are going to be waiting for action for the CMS thread; 1476 // it had better not be gone (for instance at shutdown)! 1477 assert(ConcurrentMarkSweepThread::cmst() != NULL, 1478 "CMS thread must be running"); 1479 // Wait here until the background collector gives us the go-ahead 1480 ConcurrentMarkSweepThread::clear_CMS_flag( 1481 ConcurrentMarkSweepThread::CMS_vm_has_token); // release token 1482 // Get a possibly blocked CMS thread going: 1483 // Note that we set _foregroundGCIsActive true above, 1484 // without protection of the CGC_lock. 1485 CGC_lock->notify(); 1486 assert(!ConcurrentMarkSweepThread::vm_thread_wants_cms_token(), 1487 "Possible deadlock"); 1488 while (_foregroundGCShouldWait) { 1489 // wait for notification 1490 CGC_lock->wait(Mutex::_no_safepoint_check_flag); 1491 // Possibility of delay/starvation here, since CMS token does 1492 // not know to give priority to VM thread? Actually, i think 1493 // there wouldn't be any delay/starvation, but the proof of 1494 // that "fact" (?) appears non-trivial. XXX 20011219YSR 1495 } 1496 ConcurrentMarkSweepThread::set_CMS_flag( 1497 ConcurrentMarkSweepThread::CMS_vm_has_token); 1498 } 1499 } 1500 // The CMS_token is already held. Get back the other locks. 1501 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 1502 "VM thread should have CMS token"); 1503 getFreelistLocks(); 1504 bitMapLock()->lock_without_safepoint_check(); 1505 if (TraceCMSState) { 1506 gclog_or_tty->print_cr("CMS foreground collector has asked for control " 1507 INTPTR_FORMAT " with first state %d", p2i(Thread::current()), first_state); 1508 gclog_or_tty->print_cr(" gets control with state %d", _collectorState); 1509 } 1510 1511 // Inform cms gen if this was due to partial collection failing. 1512 // The CMS gen may use this fact to determine its expansion policy. 1513 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1514 if (gch->incremental_collection_will_fail(false /* don't consult_young */)) { 1515 assert(!_cmsGen->incremental_collection_failed(), 1516 "Should have been noticed, reacted to and cleared"); 1517 _cmsGen->set_incremental_collection_failed(); 1518 } 1519 1520 if (first_state > Idling) { 1521 report_concurrent_mode_interruption(); 1522 } 1523 1524 set_did_compact(true); 1525 1526 // If the collection is being acquired from the background 1527 // collector, there may be references on the discovered 1528 // references lists. Abandon those references, since some 1529 // of them may have become unreachable after concurrent 1530 // discovery; the STW compacting collector will redo discovery 1531 // more precisely, without being subject to floating garbage. 1532 // Leaving otherwise unreachable references in the discovered 1533 // lists would require special handling. 1534 ref_processor()->disable_discovery(); 1535 ref_processor()->abandon_partial_discovery(); 1536 ref_processor()->verify_no_references_recorded(); 1537 1538 if (first_state > Idling) { 1539 save_heap_summary(); 1540 } 1541 1542 do_compaction_work(clear_all_soft_refs); 1543 1544 // Has the GC time limit been exceeded? 1545 size_t max_eden_size = _young_gen->max_eden_size(); 1546 GCCause::Cause gc_cause = gch->gc_cause(); 1547 size_policy()->check_gc_overhead_limit(_young_gen->used(), 1548 _young_gen->eden()->used(), 1549 _cmsGen->max_capacity(), 1550 max_eden_size, 1551 full, 1552 gc_cause, 1553 gch->collector_policy()); 1554 1555 // Reset the expansion cause, now that we just completed 1556 // a collection cycle. 1557 clear_expansion_cause(); 1558 _foregroundGCIsActive = false; 1559 return; 1560 } 1561 1562 // Resize the tenured generation 1563 // after obtaining the free list locks for the 1564 // two generations. 1565 void CMSCollector::compute_new_size() { 1566 assert_locked_or_safepoint(Heap_lock); 1567 FreelistLocker z(this); 1568 MetaspaceGC::compute_new_size(); 1569 _cmsGen->compute_new_size_free_list(); 1570 } 1571 1572 // A work method used by the foreground collector to do 1573 // a mark-sweep-compact. 1574 void CMSCollector::do_compaction_work(bool clear_all_soft_refs) { 1575 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1576 1577 STWGCTimer* gc_timer = GenMarkSweep::gc_timer(); 1578 gc_timer->register_gc_start(); 1579 1580 SerialOldTracer* gc_tracer = GenMarkSweep::gc_tracer(); 1581 gc_tracer->report_gc_start(gch->gc_cause(), gc_timer->gc_start()); 1582 1583 GCTraceTime t("CMS:MSC ", PrintGCDetails && Verbose, true, NULL); 1584 1585 // Temporarily widen the span of the weak reference processing to 1586 // the entire heap. 1587 MemRegion new_span(GenCollectedHeap::heap()->reserved_region()); 1588 ReferenceProcessorSpanMutator rp_mut_span(ref_processor(), new_span); 1589 // Temporarily, clear the "is_alive_non_header" field of the 1590 // reference processor. 1591 ReferenceProcessorIsAliveMutator rp_mut_closure(ref_processor(), NULL); 1592 // Temporarily make reference _processing_ single threaded (non-MT). 1593 ReferenceProcessorMTProcMutator rp_mut_mt_processing(ref_processor(), false); 1594 // Temporarily make refs discovery atomic 1595 ReferenceProcessorAtomicMutator rp_mut_atomic(ref_processor(), true); 1596 // Temporarily make reference _discovery_ single threaded (non-MT) 1597 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false); 1598 1599 ref_processor()->set_enqueuing_is_done(false); 1600 ref_processor()->enable_discovery(); 1601 ref_processor()->setup_policy(clear_all_soft_refs); 1602 // If an asynchronous collection finishes, the _modUnionTable is 1603 // all clear. If we are assuming the collection from an asynchronous 1604 // collection, clear the _modUnionTable. 1605 assert(_collectorState != Idling || _modUnionTable.isAllClear(), 1606 "_modUnionTable should be clear if the baton was not passed"); 1607 _modUnionTable.clear_all(); 1608 assert(_collectorState != Idling || _ct->klass_rem_set()->mod_union_is_clear(), 1609 "mod union for klasses should be clear if the baton was passed"); 1610 _ct->klass_rem_set()->clear_mod_union(); 1611 1612 // We must adjust the allocation statistics being maintained 1613 // in the free list space. We do so by reading and clearing 1614 // the sweep timer and updating the block flux rate estimates below. 1615 assert(!_intra_sweep_timer.is_active(), "_intra_sweep_timer should be inactive"); 1616 if (_inter_sweep_timer.is_active()) { 1617 _inter_sweep_timer.stop(); 1618 // Note that we do not use this sample to update the _inter_sweep_estimate. 1619 _cmsGen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()), 1620 _inter_sweep_estimate.padded_average(), 1621 _intra_sweep_estimate.padded_average()); 1622 } 1623 1624 GenMarkSweep::invoke_at_safepoint(ref_processor(), clear_all_soft_refs); 1625 #ifdef ASSERT 1626 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 1627 size_t free_size = cms_space->free(); 1628 assert(free_size == 1629 pointer_delta(cms_space->end(), cms_space->compaction_top()) 1630 * HeapWordSize, 1631 "All the free space should be compacted into one chunk at top"); 1632 assert(cms_space->dictionary()->total_chunk_size( 1633 debug_only(cms_space->freelistLock())) == 0 || 1634 cms_space->totalSizeInIndexedFreeLists() == 0, 1635 "All the free space should be in a single chunk"); 1636 size_t num = cms_space->totalCount(); 1637 assert((free_size == 0 && num == 0) || 1638 (free_size > 0 && (num == 1 || num == 2)), 1639 "There should be at most 2 free chunks after compaction"); 1640 #endif // ASSERT 1641 _collectorState = Resetting; 1642 assert(_restart_addr == NULL, 1643 "Should have been NULL'd before baton was passed"); 1644 reset_stw(); 1645 _cmsGen->reset_after_compaction(); 1646 _concurrent_cycles_since_last_unload = 0; 1647 1648 // Clear any data recorded in the PLAB chunk arrays. 1649 if (_survivor_plab_array != NULL) { 1650 reset_survivor_plab_arrays(); 1651 } 1652 1653 // Adjust the per-size allocation stats for the next epoch. 1654 _cmsGen->cmsSpace()->endSweepFLCensus(sweep_count() /* fake */); 1655 // Restart the "inter sweep timer" for the next epoch. 1656 _inter_sweep_timer.reset(); 1657 _inter_sweep_timer.start(); 1658 1659 gc_timer->register_gc_end(); 1660 1661 gc_tracer->report_gc_end(gc_timer->gc_end(), gc_timer->time_partitions()); 1662 1663 // For a mark-sweep-compact, compute_new_size() will be called 1664 // in the heap's do_collection() method. 1665 } 1666 1667 void CMSCollector::print_eden_and_survivor_chunk_arrays() { 1668 ContiguousSpace* eden_space = _young_gen->eden(); 1669 ContiguousSpace* from_space = _young_gen->from(); 1670 ContiguousSpace* to_space = _young_gen->to(); 1671 // Eden 1672 if (_eden_chunk_array != NULL) { 1673 gclog_or_tty->print_cr("eden " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")", 1674 p2i(eden_space->bottom()), p2i(eden_space->top()), 1675 p2i(eden_space->end()), eden_space->capacity()); 1676 gclog_or_tty->print_cr("_eden_chunk_index=" SIZE_FORMAT ", " 1677 "_eden_chunk_capacity=" SIZE_FORMAT, 1678 _eden_chunk_index, _eden_chunk_capacity); 1679 for (size_t i = 0; i < _eden_chunk_index; i++) { 1680 gclog_or_tty->print_cr("_eden_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT, 1681 i, p2i(_eden_chunk_array[i])); 1682 } 1683 } 1684 // Survivor 1685 if (_survivor_chunk_array != NULL) { 1686 gclog_or_tty->print_cr("survivor " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")", 1687 p2i(from_space->bottom()), p2i(from_space->top()), 1688 p2i(from_space->end()), from_space->capacity()); 1689 gclog_or_tty->print_cr("_survivor_chunk_index=" SIZE_FORMAT ", " 1690 "_survivor_chunk_capacity=" SIZE_FORMAT, 1691 _survivor_chunk_index, _survivor_chunk_capacity); 1692 for (size_t i = 0; i < _survivor_chunk_index; i++) { 1693 gclog_or_tty->print_cr("_survivor_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT, 1694 i, p2i(_survivor_chunk_array[i])); 1695 } 1696 } 1697 } 1698 1699 void CMSCollector::getFreelistLocks() const { 1700 // Get locks for all free lists in all generations that this 1701 // collector is responsible for 1702 _cmsGen->freelistLock()->lock_without_safepoint_check(); 1703 } 1704 1705 void CMSCollector::releaseFreelistLocks() const { 1706 // Release locks for all free lists in all generations that this 1707 // collector is responsible for 1708 _cmsGen->freelistLock()->unlock(); 1709 } 1710 1711 bool CMSCollector::haveFreelistLocks() const { 1712 // Check locks for all free lists in all generations that this 1713 // collector is responsible for 1714 assert_lock_strong(_cmsGen->freelistLock()); 1715 PRODUCT_ONLY(ShouldNotReachHere()); 1716 return true; 1717 } 1718 1719 // A utility class that is used by the CMS collector to 1720 // temporarily "release" the foreground collector from its 1721 // usual obligation to wait for the background collector to 1722 // complete an ongoing phase before proceeding. 1723 class ReleaseForegroundGC: public StackObj { 1724 private: 1725 CMSCollector* _c; 1726 public: 1727 ReleaseForegroundGC(CMSCollector* c) : _c(c) { 1728 assert(_c->_foregroundGCShouldWait, "Else should not need to call"); 1729 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1730 // allow a potentially blocked foreground collector to proceed 1731 _c->_foregroundGCShouldWait = false; 1732 if (_c->_foregroundGCIsActive) { 1733 CGC_lock->notify(); 1734 } 1735 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 1736 "Possible deadlock"); 1737 } 1738 1739 ~ReleaseForegroundGC() { 1740 assert(!_c->_foregroundGCShouldWait, "Usage protocol violation?"); 1741 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1742 _c->_foregroundGCShouldWait = true; 1743 } 1744 }; 1745 1746 void CMSCollector::collect_in_background(GCCause::Cause cause) { 1747 assert(Thread::current()->is_ConcurrentGC_thread(), 1748 "A CMS asynchronous collection is only allowed on a CMS thread."); 1749 1750 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1751 { 1752 bool safepoint_check = Mutex::_no_safepoint_check_flag; 1753 MutexLockerEx hl(Heap_lock, safepoint_check); 1754 FreelistLocker fll(this); 1755 MutexLockerEx x(CGC_lock, safepoint_check); 1756 if (_foregroundGCIsActive) { 1757 // The foreground collector is. Skip this 1758 // background collection. 1759 assert(!_foregroundGCShouldWait, "Should be clear"); 1760 return; 1761 } else { 1762 assert(_collectorState == Idling, "Should be idling before start."); 1763 _collectorState = InitialMarking; 1764 register_gc_start(cause); 1765 // Reset the expansion cause, now that we are about to begin 1766 // a new cycle. 1767 clear_expansion_cause(); 1768 1769 // Clear the MetaspaceGC flag since a concurrent collection 1770 // is starting but also clear it after the collection. 1771 MetaspaceGC::set_should_concurrent_collect(false); 1772 } 1773 // Decide if we want to enable class unloading as part of the 1774 // ensuing concurrent GC cycle. 1775 update_should_unload_classes(); 1776 _full_gc_requested = false; // acks all outstanding full gc requests 1777 _full_gc_cause = GCCause::_no_gc; 1778 // Signal that we are about to start a collection 1779 gch->increment_total_full_collections(); // ... starting a collection cycle 1780 _collection_count_start = gch->total_full_collections(); 1781 } 1782 1783 // Used for PrintGC 1784 size_t prev_used = 0; 1785 if (PrintGC && Verbose) { 1786 prev_used = _cmsGen->used(); 1787 } 1788 1789 // The change of the collection state is normally done at this level; 1790 // the exceptions are phases that are executed while the world is 1791 // stopped. For those phases the change of state is done while the 1792 // world is stopped. For baton passing purposes this allows the 1793 // background collector to finish the phase and change state atomically. 1794 // The foreground collector cannot wait on a phase that is done 1795 // while the world is stopped because the foreground collector already 1796 // has the world stopped and would deadlock. 1797 while (_collectorState != Idling) { 1798 if (TraceCMSState) { 1799 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " in CMS state %d", 1800 p2i(Thread::current()), _collectorState); 1801 } 1802 // The foreground collector 1803 // holds the Heap_lock throughout its collection. 1804 // holds the CMS token (but not the lock) 1805 // except while it is waiting for the background collector to yield. 1806 // 1807 // The foreground collector should be blocked (not for long) 1808 // if the background collector is about to start a phase 1809 // executed with world stopped. If the background 1810 // collector has already started such a phase, the 1811 // foreground collector is blocked waiting for the 1812 // Heap_lock. The stop-world phases (InitialMarking and FinalMarking) 1813 // are executed in the VM thread. 1814 // 1815 // The locking order is 1816 // PendingListLock (PLL) -- if applicable (FinalMarking) 1817 // Heap_lock (both this & PLL locked in VM_CMS_Operation::prologue()) 1818 // CMS token (claimed in 1819 // stop_world_and_do() --> 1820 // safepoint_synchronize() --> 1821 // CMSThread::synchronize()) 1822 1823 { 1824 // Check if the FG collector wants us to yield. 1825 CMSTokenSync x(true); // is cms thread 1826 if (waitForForegroundGC()) { 1827 // We yielded to a foreground GC, nothing more to be 1828 // done this round. 1829 assert(_foregroundGCShouldWait == false, "We set it to false in " 1830 "waitForForegroundGC()"); 1831 if (TraceCMSState) { 1832 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT 1833 " exiting collection CMS state %d", 1834 p2i(Thread::current()), _collectorState); 1835 } 1836 return; 1837 } else { 1838 // The background collector can run but check to see if the 1839 // foreground collector has done a collection while the 1840 // background collector was waiting to get the CGC_lock 1841 // above. If yes, break so that _foregroundGCShouldWait 1842 // is cleared before returning. 1843 if (_collectorState == Idling) { 1844 break; 1845 } 1846 } 1847 } 1848 1849 assert(_foregroundGCShouldWait, "Foreground collector, if active, " 1850 "should be waiting"); 1851 1852 switch (_collectorState) { 1853 case InitialMarking: 1854 { 1855 ReleaseForegroundGC x(this); 1856 stats().record_cms_begin(); 1857 VM_CMS_Initial_Mark initial_mark_op(this); 1858 VMThread::execute(&initial_mark_op); 1859 } 1860 // The collector state may be any legal state at this point 1861 // since the background collector may have yielded to the 1862 // foreground collector. 1863 break; 1864 case Marking: 1865 // initial marking in checkpointRootsInitialWork has been completed 1866 if (markFromRoots()) { // we were successful 1867 assert(_collectorState == Precleaning, "Collector state should " 1868 "have changed"); 1869 } else { 1870 assert(_foregroundGCIsActive, "Internal state inconsistency"); 1871 } 1872 break; 1873 case Precleaning: 1874 // marking from roots in markFromRoots has been completed 1875 preclean(); 1876 assert(_collectorState == AbortablePreclean || 1877 _collectorState == FinalMarking, 1878 "Collector state should have changed"); 1879 break; 1880 case AbortablePreclean: 1881 abortable_preclean(); 1882 assert(_collectorState == FinalMarking, "Collector state should " 1883 "have changed"); 1884 break; 1885 case FinalMarking: 1886 { 1887 ReleaseForegroundGC x(this); 1888 1889 VM_CMS_Final_Remark final_remark_op(this); 1890 VMThread::execute(&final_remark_op); 1891 } 1892 assert(_foregroundGCShouldWait, "block post-condition"); 1893 break; 1894 case Sweeping: 1895 // final marking in checkpointRootsFinal has been completed 1896 sweep(); 1897 assert(_collectorState == Resizing, "Collector state change " 1898 "to Resizing must be done under the free_list_lock"); 1899 1900 case Resizing: { 1901 // Sweeping has been completed... 1902 // At this point the background collection has completed. 1903 // Don't move the call to compute_new_size() down 1904 // into code that might be executed if the background 1905 // collection was preempted. 1906 { 1907 ReleaseForegroundGC x(this); // unblock FG collection 1908 MutexLockerEx y(Heap_lock, Mutex::_no_safepoint_check_flag); 1909 CMSTokenSync z(true); // not strictly needed. 1910 if (_collectorState == Resizing) { 1911 compute_new_size(); 1912 save_heap_summary(); 1913 _collectorState = Resetting; 1914 } else { 1915 assert(_collectorState == Idling, "The state should only change" 1916 " because the foreground collector has finished the collection"); 1917 } 1918 } 1919 break; 1920 } 1921 case Resetting: 1922 // CMS heap resizing has been completed 1923 reset_concurrent(); 1924 assert(_collectorState == Idling, "Collector state should " 1925 "have changed"); 1926 1927 MetaspaceGC::set_should_concurrent_collect(false); 1928 1929 stats().record_cms_end(); 1930 // Don't move the concurrent_phases_end() and compute_new_size() 1931 // calls to here because a preempted background collection 1932 // has it's state set to "Resetting". 1933 break; 1934 case Idling: 1935 default: 1936 ShouldNotReachHere(); 1937 break; 1938 } 1939 if (TraceCMSState) { 1940 gclog_or_tty->print_cr(" Thread " INTPTR_FORMAT " done - next CMS state %d", 1941 p2i(Thread::current()), _collectorState); 1942 } 1943 assert(_foregroundGCShouldWait, "block post-condition"); 1944 } 1945 1946 // Should this be in gc_epilogue? 1947 collector_policy()->counters()->update_counters(); 1948 1949 { 1950 // Clear _foregroundGCShouldWait and, in the event that the 1951 // foreground collector is waiting, notify it, before 1952 // returning. 1953 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1954 _foregroundGCShouldWait = false; 1955 if (_foregroundGCIsActive) { 1956 CGC_lock->notify(); 1957 } 1958 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 1959 "Possible deadlock"); 1960 } 1961 if (TraceCMSState) { 1962 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT 1963 " exiting collection CMS state %d", 1964 p2i(Thread::current()), _collectorState); 1965 } 1966 if (PrintGC && Verbose) { 1967 _cmsGen->print_heap_change(prev_used); 1968 } 1969 } 1970 1971 void CMSCollector::register_gc_start(GCCause::Cause cause) { 1972 _cms_start_registered = true; 1973 _gc_timer_cm->register_gc_start(); 1974 _gc_tracer_cm->report_gc_start(cause, _gc_timer_cm->gc_start()); 1975 } 1976 1977 void CMSCollector::register_gc_end() { 1978 if (_cms_start_registered) { 1979 report_heap_summary(GCWhen::AfterGC); 1980 1981 _gc_timer_cm->register_gc_end(); 1982 _gc_tracer_cm->report_gc_end(_gc_timer_cm->gc_end(), _gc_timer_cm->time_partitions()); 1983 _cms_start_registered = false; 1984 } 1985 } 1986 1987 void CMSCollector::save_heap_summary() { 1988 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1989 _last_heap_summary = gch->create_heap_summary(); 1990 _last_metaspace_summary = gch->create_metaspace_summary(); 1991 } 1992 1993 void CMSCollector::report_heap_summary(GCWhen::Type when) { 1994 _gc_tracer_cm->report_gc_heap_summary(when, _last_heap_summary); 1995 _gc_tracer_cm->report_metaspace_summary(when, _last_metaspace_summary); 1996 } 1997 1998 bool CMSCollector::waitForForegroundGC() { 1999 bool res = false; 2000 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 2001 "CMS thread should have CMS token"); 2002 // Block the foreground collector until the 2003 // background collectors decides whether to 2004 // yield. 2005 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 2006 _foregroundGCShouldWait = true; 2007 if (_foregroundGCIsActive) { 2008 // The background collector yields to the 2009 // foreground collector and returns a value 2010 // indicating that it has yielded. The foreground 2011 // collector can proceed. 2012 res = true; 2013 _foregroundGCShouldWait = false; 2014 ConcurrentMarkSweepThread::clear_CMS_flag( 2015 ConcurrentMarkSweepThread::CMS_cms_has_token); 2016 ConcurrentMarkSweepThread::set_CMS_flag( 2017 ConcurrentMarkSweepThread::CMS_cms_wants_token); 2018 // Get a possibly blocked foreground thread going 2019 CGC_lock->notify(); 2020 if (TraceCMSState) { 2021 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " waiting at CMS state %d", 2022 p2i(Thread::current()), _collectorState); 2023 } 2024 while (_foregroundGCIsActive) { 2025 CGC_lock->wait(Mutex::_no_safepoint_check_flag); 2026 } 2027 ConcurrentMarkSweepThread::set_CMS_flag( 2028 ConcurrentMarkSweepThread::CMS_cms_has_token); 2029 ConcurrentMarkSweepThread::clear_CMS_flag( 2030 ConcurrentMarkSweepThread::CMS_cms_wants_token); 2031 } 2032 if (TraceCMSState) { 2033 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " continuing at CMS state %d", 2034 p2i(Thread::current()), _collectorState); 2035 } 2036 return res; 2037 } 2038 2039 // Because of the need to lock the free lists and other structures in 2040 // the collector, common to all the generations that the collector is 2041 // collecting, we need the gc_prologues of individual CMS generations 2042 // delegate to their collector. It may have been simpler had the 2043 // current infrastructure allowed one to call a prologue on a 2044 // collector. In the absence of that we have the generation's 2045 // prologue delegate to the collector, which delegates back 2046 // some "local" work to a worker method in the individual generations 2047 // that it's responsible for collecting, while itself doing any 2048 // work common to all generations it's responsible for. A similar 2049 // comment applies to the gc_epilogue()'s. 2050 // The role of the variable _between_prologue_and_epilogue is to 2051 // enforce the invocation protocol. 2052 void CMSCollector::gc_prologue(bool full) { 2053 // Call gc_prologue_work() for the CMSGen 2054 // we are responsible for. 2055 2056 // The following locking discipline assumes that we are only called 2057 // when the world is stopped. 2058 assert(SafepointSynchronize::is_at_safepoint(), "world is stopped assumption"); 2059 2060 // The CMSCollector prologue must call the gc_prologues for the 2061 // "generations" that it's responsible 2062 // for. 2063 2064 assert( Thread::current()->is_VM_thread() 2065 || ( CMSScavengeBeforeRemark 2066 && Thread::current()->is_ConcurrentGC_thread()), 2067 "Incorrect thread type for prologue execution"); 2068 2069 if (_between_prologue_and_epilogue) { 2070 // We have already been invoked; this is a gc_prologue delegation 2071 // from yet another CMS generation that we are responsible for, just 2072 // ignore it since all relevant work has already been done. 2073 return; 2074 } 2075 2076 // set a bit saying prologue has been called; cleared in epilogue 2077 _between_prologue_and_epilogue = true; 2078 // Claim locks for common data structures, then call gc_prologue_work() 2079 // for each CMSGen. 2080 2081 getFreelistLocks(); // gets free list locks on constituent spaces 2082 bitMapLock()->lock_without_safepoint_check(); 2083 2084 // Should call gc_prologue_work() for all cms gens we are responsible for 2085 bool duringMarking = _collectorState >= Marking 2086 && _collectorState < Sweeping; 2087 2088 // The young collections clear the modified oops state, which tells if 2089 // there are any modified oops in the class. The remark phase also needs 2090 // that information. Tell the young collection to save the union of all 2091 // modified klasses. 2092 if (duringMarking) { 2093 _ct->klass_rem_set()->set_accumulate_modified_oops(true); 2094 } 2095 2096 bool registerClosure = duringMarking; 2097 2098 _cmsGen->gc_prologue_work(full, registerClosure, &_modUnionClosurePar); 2099 2100 if (!full) { 2101 stats().record_gc0_begin(); 2102 } 2103 } 2104 2105 void ConcurrentMarkSweepGeneration::gc_prologue(bool full) { 2106 2107 _capacity_at_prologue = capacity(); 2108 _used_at_prologue = used(); 2109 2110 // Delegate to CMScollector which knows how to coordinate between 2111 // this and any other CMS generations that it is responsible for 2112 // collecting. 2113 collector()->gc_prologue(full); 2114 } 2115 2116 // This is a "private" interface for use by this generation's CMSCollector. 2117 // Not to be called directly by any other entity (for instance, 2118 // GenCollectedHeap, which calls the "public" gc_prologue method above). 2119 void ConcurrentMarkSweepGeneration::gc_prologue_work(bool full, 2120 bool registerClosure, ModUnionClosure* modUnionClosure) { 2121 assert(!incremental_collection_failed(), "Shouldn't be set yet"); 2122 assert(cmsSpace()->preconsumptionDirtyCardClosure() == NULL, 2123 "Should be NULL"); 2124 if (registerClosure) { 2125 cmsSpace()->setPreconsumptionDirtyCardClosure(modUnionClosure); 2126 } 2127 cmsSpace()->gc_prologue(); 2128 // Clear stat counters 2129 NOT_PRODUCT( 2130 assert(_numObjectsPromoted == 0, "check"); 2131 assert(_numWordsPromoted == 0, "check"); 2132 if (Verbose && PrintGC) { 2133 gclog_or_tty->print("Allocated " SIZE_FORMAT " objects, " 2134 SIZE_FORMAT " bytes concurrently", 2135 _numObjectsAllocated, _numWordsAllocated*sizeof(HeapWord)); 2136 } 2137 _numObjectsAllocated = 0; 2138 _numWordsAllocated = 0; 2139 ) 2140 } 2141 2142 void CMSCollector::gc_epilogue(bool full) { 2143 // The following locking discipline assumes that we are only called 2144 // when the world is stopped. 2145 assert(SafepointSynchronize::is_at_safepoint(), 2146 "world is stopped assumption"); 2147 2148 // Currently the CMS epilogue (see CompactibleFreeListSpace) merely checks 2149 // if linear allocation blocks need to be appropriately marked to allow the 2150 // the blocks to be parsable. We also check here whether we need to nudge the 2151 // CMS collector thread to start a new cycle (if it's not already active). 2152 assert( Thread::current()->is_VM_thread() 2153 || ( CMSScavengeBeforeRemark 2154 && Thread::current()->is_ConcurrentGC_thread()), 2155 "Incorrect thread type for epilogue execution"); 2156 2157 if (!_between_prologue_and_epilogue) { 2158 // We have already been invoked; this is a gc_epilogue delegation 2159 // from yet another CMS generation that we are responsible for, just 2160 // ignore it since all relevant work has already been done. 2161 return; 2162 } 2163 assert(haveFreelistLocks(), "must have freelist locks"); 2164 assert_lock_strong(bitMapLock()); 2165 2166 _ct->klass_rem_set()->set_accumulate_modified_oops(false); 2167 2168 _cmsGen->gc_epilogue_work(full); 2169 2170 if (_collectorState == AbortablePreclean || _collectorState == Precleaning) { 2171 // in case sampling was not already enabled, enable it 2172 _start_sampling = true; 2173 } 2174 // reset _eden_chunk_array so sampling starts afresh 2175 _eden_chunk_index = 0; 2176 2177 size_t cms_used = _cmsGen->cmsSpace()->used(); 2178 2179 // update performance counters - this uses a special version of 2180 // update_counters() that allows the utilization to be passed as a 2181 // parameter, avoiding multiple calls to used(). 2182 // 2183 _cmsGen->update_counters(cms_used); 2184 2185 bitMapLock()->unlock(); 2186 releaseFreelistLocks(); 2187 2188 if (!CleanChunkPoolAsync) { 2189 Chunk::clean_chunk_pool(); 2190 } 2191 2192 set_did_compact(false); 2193 _between_prologue_and_epilogue = false; // ready for next cycle 2194 } 2195 2196 void ConcurrentMarkSweepGeneration::gc_epilogue(bool full) { 2197 collector()->gc_epilogue(full); 2198 2199 // Also reset promotion tracking in par gc thread states. 2200 for (uint i = 0; i < ParallelGCThreads; i++) { 2201 _par_gc_thread_states[i]->promo.stopTrackingPromotions(i); 2202 } 2203 } 2204 2205 void ConcurrentMarkSweepGeneration::gc_epilogue_work(bool full) { 2206 assert(!incremental_collection_failed(), "Should have been cleared"); 2207 cmsSpace()->setPreconsumptionDirtyCardClosure(NULL); 2208 cmsSpace()->gc_epilogue(); 2209 // Print stat counters 2210 NOT_PRODUCT( 2211 assert(_numObjectsAllocated == 0, "check"); 2212 assert(_numWordsAllocated == 0, "check"); 2213 if (Verbose && PrintGC) { 2214 gclog_or_tty->print("Promoted " SIZE_FORMAT " objects, " 2215 SIZE_FORMAT " bytes", 2216 _numObjectsPromoted, _numWordsPromoted*sizeof(HeapWord)); 2217 } 2218 _numObjectsPromoted = 0; 2219 _numWordsPromoted = 0; 2220 ) 2221 2222 if (PrintGC && Verbose) { 2223 // Call down the chain in contiguous_available needs the freelistLock 2224 // so print this out before releasing the freeListLock. 2225 gclog_or_tty->print(" Contiguous available " SIZE_FORMAT " bytes ", 2226 contiguous_available()); 2227 } 2228 } 2229 2230 #ifndef PRODUCT 2231 bool CMSCollector::have_cms_token() { 2232 Thread* thr = Thread::current(); 2233 if (thr->is_VM_thread()) { 2234 return ConcurrentMarkSweepThread::vm_thread_has_cms_token(); 2235 } else if (thr->is_ConcurrentGC_thread()) { 2236 return ConcurrentMarkSweepThread::cms_thread_has_cms_token(); 2237 } else if (thr->is_GC_task_thread()) { 2238 return ConcurrentMarkSweepThread::vm_thread_has_cms_token() && 2239 ParGCRareEvent_lock->owned_by_self(); 2240 } 2241 return false; 2242 } 2243 #endif 2244 2245 void 2246 CMSCollector::print_on_error(outputStream* st) { 2247 CMSCollector* collector = ConcurrentMarkSweepGeneration::_collector; 2248 if (collector != NULL) { 2249 CMSBitMap* bitmap = &collector->_markBitMap; 2250 st->print_cr("Marking Bits: (CMSBitMap*) " PTR_FORMAT, p2i(bitmap)); 2251 bitmap->print_on_error(st, " Bits: "); 2252 2253 st->cr(); 2254 2255 CMSBitMap* mut_bitmap = &collector->_modUnionTable; 2256 st->print_cr("Mod Union Table: (CMSBitMap*) " PTR_FORMAT, p2i(mut_bitmap)); 2257 mut_bitmap->print_on_error(st, " Bits: "); 2258 } 2259 } 2260 2261 //////////////////////////////////////////////////////// 2262 // CMS Verification Support 2263 //////////////////////////////////////////////////////// 2264 // Following the remark phase, the following invariant 2265 // should hold -- each object in the CMS heap which is 2266 // marked in markBitMap() should be marked in the verification_mark_bm(). 2267 2268 class VerifyMarkedClosure: public BitMapClosure { 2269 CMSBitMap* _marks; 2270 bool _failed; 2271 2272 public: 2273 VerifyMarkedClosure(CMSBitMap* bm): _marks(bm), _failed(false) {} 2274 2275 bool do_bit(size_t offset) { 2276 HeapWord* addr = _marks->offsetToHeapWord(offset); 2277 if (!_marks->isMarked(addr)) { 2278 oop(addr)->print_on(gclog_or_tty); 2279 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)", p2i(addr)); 2280 _failed = true; 2281 } 2282 return true; 2283 } 2284 2285 bool failed() { return _failed; } 2286 }; 2287 2288 bool CMSCollector::verify_after_remark(bool silent) { 2289 if (!silent) gclog_or_tty->print(" [Verifying CMS Marking... "); 2290 MutexLockerEx ml(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag); 2291 static bool init = false; 2292 2293 assert(SafepointSynchronize::is_at_safepoint(), 2294 "Else mutations in object graph will make answer suspect"); 2295 assert(have_cms_token(), 2296 "Else there may be mutual interference in use of " 2297 " verification data structures"); 2298 assert(_collectorState > Marking && _collectorState <= Sweeping, 2299 "Else marking info checked here may be obsolete"); 2300 assert(haveFreelistLocks(), "must hold free list locks"); 2301 assert_lock_strong(bitMapLock()); 2302 2303 2304 // Allocate marking bit map if not already allocated 2305 if (!init) { // first time 2306 if (!verification_mark_bm()->allocate(_span)) { 2307 return false; 2308 } 2309 init = true; 2310 } 2311 2312 assert(verification_mark_stack()->isEmpty(), "Should be empty"); 2313 2314 // Turn off refs discovery -- so we will be tracing through refs. 2315 // This is as intended, because by this time 2316 // GC must already have cleared any refs that need to be cleared, 2317 // and traced those that need to be marked; moreover, 2318 // the marking done here is not going to interfere in any 2319 // way with the marking information used by GC. 2320 NoRefDiscovery no_discovery(ref_processor()); 2321 2322 #if defined(COMPILER2) || INCLUDE_JVMCI 2323 DerivedPointerTableDeactivate dpt_deact; 2324 #endif 2325 2326 // Clear any marks from a previous round 2327 verification_mark_bm()->clear_all(); 2328 assert(verification_mark_stack()->isEmpty(), "markStack should be empty"); 2329 verify_work_stacks_empty(); 2330 2331 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2332 gch->ensure_parsability(false); // fill TLABs, but no need to retire them 2333 // Update the saved marks which may affect the root scans. 2334 gch->save_marks(); 2335 2336 if (CMSRemarkVerifyVariant == 1) { 2337 // In this first variant of verification, we complete 2338 // all marking, then check if the new marks-vector is 2339 // a subset of the CMS marks-vector. 2340 verify_after_remark_work_1(); 2341 } else if (CMSRemarkVerifyVariant == 2) { 2342 // In this second variant of verification, we flag an error 2343 // (i.e. an object reachable in the new marks-vector not reachable 2344 // in the CMS marks-vector) immediately, also indicating the 2345 // identify of an object (A) that references the unmarked object (B) -- 2346 // presumably, a mutation to A failed to be picked up by preclean/remark? 2347 verify_after_remark_work_2(); 2348 } else { 2349 warning("Unrecognized value " UINTX_FORMAT " for CMSRemarkVerifyVariant", 2350 CMSRemarkVerifyVariant); 2351 } 2352 if (!silent) gclog_or_tty->print(" done] "); 2353 return true; 2354 } 2355 2356 void CMSCollector::verify_after_remark_work_1() { 2357 ResourceMark rm; 2358 HandleMark hm; 2359 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2360 2361 // Get a clear set of claim bits for the roots processing to work with. 2362 ClassLoaderDataGraph::clear_claimed_marks(); 2363 2364 // Mark from roots one level into CMS 2365 MarkRefsIntoClosure notOlder(_span, verification_mark_bm()); 2366 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2367 2368 { 2369 StrongRootsScope srs(1); 2370 2371 gch->gen_process_roots(&srs, 2372 GenCollectedHeap::OldGen, 2373 true, // young gen as roots 2374 GenCollectedHeap::ScanningOption(roots_scanning_options()), 2375 should_unload_classes(), 2376 ¬Older, 2377 NULL, 2378 NULL); 2379 } 2380 2381 // Now mark from the roots 2382 MarkFromRootsClosure markFromRootsClosure(this, _span, 2383 verification_mark_bm(), verification_mark_stack(), 2384 false /* don't yield */, true /* verifying */); 2385 assert(_restart_addr == NULL, "Expected pre-condition"); 2386 verification_mark_bm()->iterate(&markFromRootsClosure); 2387 while (_restart_addr != NULL) { 2388 // Deal with stack overflow: by restarting at the indicated 2389 // address. 2390 HeapWord* ra = _restart_addr; 2391 markFromRootsClosure.reset(ra); 2392 _restart_addr = NULL; 2393 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end()); 2394 } 2395 assert(verification_mark_stack()->isEmpty(), "Should have been drained"); 2396 verify_work_stacks_empty(); 2397 2398 // Marking completed -- now verify that each bit marked in 2399 // verification_mark_bm() is also marked in markBitMap(); flag all 2400 // errors by printing corresponding objects. 2401 VerifyMarkedClosure vcl(markBitMap()); 2402 verification_mark_bm()->iterate(&vcl); 2403 if (vcl.failed()) { 2404 gclog_or_tty->print("Verification failed"); 2405 gch->print_on(gclog_or_tty); 2406 fatal("CMS: failed marking verification after remark"); 2407 } 2408 } 2409 2410 class VerifyKlassOopsKlassClosure : public KlassClosure { 2411 class VerifyKlassOopsClosure : public OopClosure { 2412 CMSBitMap* _bitmap; 2413 public: 2414 VerifyKlassOopsClosure(CMSBitMap* bitmap) : _bitmap(bitmap) { } 2415 void do_oop(oop* p) { guarantee(*p == NULL || _bitmap->isMarked((HeapWord*) *p), "Should be marked"); } 2416 void do_oop(narrowOop* p) { ShouldNotReachHere(); } 2417 } _oop_closure; 2418 public: 2419 VerifyKlassOopsKlassClosure(CMSBitMap* bitmap) : _oop_closure(bitmap) {} 2420 void do_klass(Klass* k) { 2421 k->oops_do(&_oop_closure); 2422 } 2423 }; 2424 2425 void CMSCollector::verify_after_remark_work_2() { 2426 ResourceMark rm; 2427 HandleMark hm; 2428 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2429 2430 // Get a clear set of claim bits for the roots processing to work with. 2431 ClassLoaderDataGraph::clear_claimed_marks(); 2432 2433 // Mark from roots one level into CMS 2434 MarkRefsIntoVerifyClosure notOlder(_span, verification_mark_bm(), 2435 markBitMap()); 2436 CLDToOopClosure cld_closure(¬Older, true); 2437 2438 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2439 2440 { 2441 StrongRootsScope srs(1); 2442 2443 gch->gen_process_roots(&srs, 2444 GenCollectedHeap::OldGen, 2445 true, // young gen as roots 2446 GenCollectedHeap::ScanningOption(roots_scanning_options()), 2447 should_unload_classes(), 2448 ¬Older, 2449 NULL, 2450 &cld_closure); 2451 } 2452 2453 // Now mark from the roots 2454 MarkFromRootsVerifyClosure markFromRootsClosure(this, _span, 2455 verification_mark_bm(), markBitMap(), verification_mark_stack()); 2456 assert(_restart_addr == NULL, "Expected pre-condition"); 2457 verification_mark_bm()->iterate(&markFromRootsClosure); 2458 while (_restart_addr != NULL) { 2459 // Deal with stack overflow: by restarting at the indicated 2460 // address. 2461 HeapWord* ra = _restart_addr; 2462 markFromRootsClosure.reset(ra); 2463 _restart_addr = NULL; 2464 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end()); 2465 } 2466 assert(verification_mark_stack()->isEmpty(), "Should have been drained"); 2467 verify_work_stacks_empty(); 2468 2469 VerifyKlassOopsKlassClosure verify_klass_oops(verification_mark_bm()); 2470 ClassLoaderDataGraph::classes_do(&verify_klass_oops); 2471 2472 // Marking completed -- now verify that each bit marked in 2473 // verification_mark_bm() is also marked in markBitMap(); flag all 2474 // errors by printing corresponding objects. 2475 VerifyMarkedClosure vcl(markBitMap()); 2476 verification_mark_bm()->iterate(&vcl); 2477 assert(!vcl.failed(), "Else verification above should not have succeeded"); 2478 } 2479 2480 void ConcurrentMarkSweepGeneration::save_marks() { 2481 // delegate to CMS space 2482 cmsSpace()->save_marks(); 2483 for (uint i = 0; i < ParallelGCThreads; i++) { 2484 _par_gc_thread_states[i]->promo.startTrackingPromotions(); 2485 } 2486 } 2487 2488 bool ConcurrentMarkSweepGeneration::no_allocs_since_save_marks() { 2489 return cmsSpace()->no_allocs_since_save_marks(); 2490 } 2491 2492 #define CMS_SINCE_SAVE_MARKS_DEFN(OopClosureType, nv_suffix) \ 2493 \ 2494 void ConcurrentMarkSweepGeneration:: \ 2495 oop_since_save_marks_iterate##nv_suffix(OopClosureType* cl) { \ 2496 cl->set_generation(this); \ 2497 cmsSpace()->oop_since_save_marks_iterate##nv_suffix(cl); \ 2498 cl->reset_generation(); \ 2499 save_marks(); \ 2500 } 2501 2502 ALL_SINCE_SAVE_MARKS_CLOSURES(CMS_SINCE_SAVE_MARKS_DEFN) 2503 2504 void 2505 ConcurrentMarkSweepGeneration::oop_iterate(ExtendedOopClosure* cl) { 2506 if (freelistLock()->owned_by_self()) { 2507 Generation::oop_iterate(cl); 2508 } else { 2509 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2510 Generation::oop_iterate(cl); 2511 } 2512 } 2513 2514 void 2515 ConcurrentMarkSweepGeneration::object_iterate(ObjectClosure* cl) { 2516 if (freelistLock()->owned_by_self()) { 2517 Generation::object_iterate(cl); 2518 } else { 2519 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2520 Generation::object_iterate(cl); 2521 } 2522 } 2523 2524 void 2525 ConcurrentMarkSweepGeneration::safe_object_iterate(ObjectClosure* cl) { 2526 if (freelistLock()->owned_by_self()) { 2527 Generation::safe_object_iterate(cl); 2528 } else { 2529 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2530 Generation::safe_object_iterate(cl); 2531 } 2532 } 2533 2534 void 2535 ConcurrentMarkSweepGeneration::post_compact() { 2536 } 2537 2538 void 2539 ConcurrentMarkSweepGeneration::prepare_for_verify() { 2540 // Fix the linear allocation blocks to look like free blocks. 2541 2542 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those 2543 // are not called when the heap is verified during universe initialization and 2544 // at vm shutdown. 2545 if (freelistLock()->owned_by_self()) { 2546 cmsSpace()->prepare_for_verify(); 2547 } else { 2548 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag); 2549 cmsSpace()->prepare_for_verify(); 2550 } 2551 } 2552 2553 void 2554 ConcurrentMarkSweepGeneration::verify() { 2555 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those 2556 // are not called when the heap is verified during universe initialization and 2557 // at vm shutdown. 2558 if (freelistLock()->owned_by_self()) { 2559 cmsSpace()->verify(); 2560 } else { 2561 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag); 2562 cmsSpace()->verify(); 2563 } 2564 } 2565 2566 void CMSCollector::verify() { 2567 _cmsGen->verify(); 2568 } 2569 2570 #ifndef PRODUCT 2571 bool CMSCollector::overflow_list_is_empty() const { 2572 assert(_num_par_pushes >= 0, "Inconsistency"); 2573 if (_overflow_list == NULL) { 2574 assert(_num_par_pushes == 0, "Inconsistency"); 2575 } 2576 return _overflow_list == NULL; 2577 } 2578 2579 // The methods verify_work_stacks_empty() and verify_overflow_empty() 2580 // merely consolidate assertion checks that appear to occur together frequently. 2581 void CMSCollector::verify_work_stacks_empty() const { 2582 assert(_markStack.isEmpty(), "Marking stack should be empty"); 2583 assert(overflow_list_is_empty(), "Overflow list should be empty"); 2584 } 2585 2586 void CMSCollector::verify_overflow_empty() const { 2587 assert(overflow_list_is_empty(), "Overflow list should be empty"); 2588 assert(no_preserved_marks(), "No preserved marks"); 2589 } 2590 #endif // PRODUCT 2591 2592 // Decide if we want to enable class unloading as part of the 2593 // ensuing concurrent GC cycle. We will collect and 2594 // unload classes if it's the case that: 2595 // (1) an explicit gc request has been made and the flag 2596 // ExplicitGCInvokesConcurrentAndUnloadsClasses is set, OR 2597 // (2) (a) class unloading is enabled at the command line, and 2598 // (b) old gen is getting really full 2599 // NOTE: Provided there is no change in the state of the heap between 2600 // calls to this method, it should have idempotent results. Moreover, 2601 // its results should be monotonically increasing (i.e. going from 0 to 1, 2602 // but not 1 to 0) between successive calls between which the heap was 2603 // not collected. For the implementation below, it must thus rely on 2604 // the property that concurrent_cycles_since_last_unload() 2605 // will not decrease unless a collection cycle happened and that 2606 // _cmsGen->is_too_full() are 2607 // themselves also monotonic in that sense. See check_monotonicity() 2608 // below. 2609 void CMSCollector::update_should_unload_classes() { 2610 _should_unload_classes = false; 2611 // Condition 1 above 2612 if (_full_gc_requested && ExplicitGCInvokesConcurrentAndUnloadsClasses) { 2613 _should_unload_classes = true; 2614 } else if (CMSClassUnloadingEnabled) { // Condition 2.a above 2615 // Disjuncts 2.b.(i,ii,iii) above 2616 _should_unload_classes = (concurrent_cycles_since_last_unload() >= 2617 CMSClassUnloadingMaxInterval) 2618 || _cmsGen->is_too_full(); 2619 } 2620 } 2621 2622 bool ConcurrentMarkSweepGeneration::is_too_full() const { 2623 bool res = should_concurrent_collect(); 2624 res = res && (occupancy() > (double)CMSIsTooFullPercentage/100.0); 2625 return res; 2626 } 2627 2628 void CMSCollector::setup_cms_unloading_and_verification_state() { 2629 const bool should_verify = VerifyBeforeGC || VerifyAfterGC || VerifyDuringGC 2630 || VerifyBeforeExit; 2631 const int rso = GenCollectedHeap::SO_AllCodeCache; 2632 2633 // We set the proper root for this CMS cycle here. 2634 if (should_unload_classes()) { // Should unload classes this cycle 2635 remove_root_scanning_option(rso); // Shrink the root set appropriately 2636 set_verifying(should_verify); // Set verification state for this cycle 2637 return; // Nothing else needs to be done at this time 2638 } 2639 2640 // Not unloading classes this cycle 2641 assert(!should_unload_classes(), "Inconsistency!"); 2642 2643 // If we are not unloading classes then add SO_AllCodeCache to root 2644 // scanning options. 2645 add_root_scanning_option(rso); 2646 2647 if ((!verifying() || unloaded_classes_last_cycle()) && should_verify) { 2648 set_verifying(true); 2649 } else if (verifying() && !should_verify) { 2650 // We were verifying, but some verification flags got disabled. 2651 set_verifying(false); 2652 // Exclude symbols, strings and code cache elements from root scanning to 2653 // reduce IM and RM pauses. 2654 remove_root_scanning_option(rso); 2655 } 2656 } 2657 2658 2659 #ifndef PRODUCT 2660 HeapWord* CMSCollector::block_start(const void* p) const { 2661 const HeapWord* addr = (HeapWord*)p; 2662 if (_span.contains(p)) { 2663 if (_cmsGen->cmsSpace()->is_in_reserved(addr)) { 2664 return _cmsGen->cmsSpace()->block_start(p); 2665 } 2666 } 2667 return NULL; 2668 } 2669 #endif 2670 2671 HeapWord* 2672 ConcurrentMarkSweepGeneration::expand_and_allocate(size_t word_size, 2673 bool tlab, 2674 bool parallel) { 2675 CMSSynchronousYieldRequest yr; 2676 assert(!tlab, "Can't deal with TLAB allocation"); 2677 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2678 expand_for_gc_cause(word_size*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_satisfy_allocation); 2679 if (GCExpandToAllocateDelayMillis > 0) { 2680 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 2681 } 2682 return have_lock_and_allocate(word_size, tlab); 2683 } 2684 2685 void ConcurrentMarkSweepGeneration::expand_for_gc_cause( 2686 size_t bytes, 2687 size_t expand_bytes, 2688 CMSExpansionCause::Cause cause) 2689 { 2690 2691 bool success = expand(bytes, expand_bytes); 2692 2693 // remember why we expanded; this information is used 2694 // by shouldConcurrentCollect() when making decisions on whether to start 2695 // a new CMS cycle. 2696 if (success) { 2697 set_expansion_cause(cause); 2698 if (PrintGCDetails && Verbose) { 2699 gclog_or_tty->print_cr("Expanded CMS gen for %s", 2700 CMSExpansionCause::to_string(cause)); 2701 } 2702 } 2703 } 2704 2705 HeapWord* ConcurrentMarkSweepGeneration::expand_and_par_lab_allocate(CMSParGCThreadState* ps, size_t word_sz) { 2706 HeapWord* res = NULL; 2707 MutexLocker x(ParGCRareEvent_lock); 2708 while (true) { 2709 // Expansion by some other thread might make alloc OK now: 2710 res = ps->lab.alloc(word_sz); 2711 if (res != NULL) return res; 2712 // If there's not enough expansion space available, give up. 2713 if (_virtual_space.uncommitted_size() < (word_sz * HeapWordSize)) { 2714 return NULL; 2715 } 2716 // Otherwise, we try expansion. 2717 expand_for_gc_cause(word_sz*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_allocate_par_lab); 2718 // Now go around the loop and try alloc again; 2719 // A competing par_promote might beat us to the expansion space, 2720 // so we may go around the loop again if promotion fails again. 2721 if (GCExpandToAllocateDelayMillis > 0) { 2722 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 2723 } 2724 } 2725 } 2726 2727 2728 bool ConcurrentMarkSweepGeneration::expand_and_ensure_spooling_space( 2729 PromotionInfo* promo) { 2730 MutexLocker x(ParGCRareEvent_lock); 2731 size_t refill_size_bytes = promo->refillSize() * HeapWordSize; 2732 while (true) { 2733 // Expansion by some other thread might make alloc OK now: 2734 if (promo->ensure_spooling_space()) { 2735 assert(promo->has_spooling_space(), 2736 "Post-condition of successful ensure_spooling_space()"); 2737 return true; 2738 } 2739 // If there's not enough expansion space available, give up. 2740 if (_virtual_space.uncommitted_size() < refill_size_bytes) { 2741 return false; 2742 } 2743 // Otherwise, we try expansion. 2744 expand_for_gc_cause(refill_size_bytes, MinHeapDeltaBytes, CMSExpansionCause::_allocate_par_spooling_space); 2745 // Now go around the loop and try alloc again; 2746 // A competing allocation might beat us to the expansion space, 2747 // so we may go around the loop again if allocation fails again. 2748 if (GCExpandToAllocateDelayMillis > 0) { 2749 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 2750 } 2751 } 2752 } 2753 2754 void ConcurrentMarkSweepGeneration::shrink(size_t bytes) { 2755 // Only shrink if a compaction was done so that all the free space 2756 // in the generation is in a contiguous block at the end. 2757 if (did_compact()) { 2758 CardGeneration::shrink(bytes); 2759 } 2760 } 2761 2762 void ConcurrentMarkSweepGeneration::assert_correct_size_change_locking() { 2763 assert_locked_or_safepoint(Heap_lock); 2764 } 2765 2766 void ConcurrentMarkSweepGeneration::shrink_free_list_by(size_t bytes) { 2767 assert_locked_or_safepoint(Heap_lock); 2768 assert_lock_strong(freelistLock()); 2769 if (PrintGCDetails && Verbose) { 2770 warning("Shrinking of CMS not yet implemented"); 2771 } 2772 return; 2773 } 2774 2775 2776 // Simple ctor/dtor wrapper for accounting & timer chores around concurrent 2777 // phases. 2778 class CMSPhaseAccounting: public StackObj { 2779 public: 2780 CMSPhaseAccounting(CMSCollector *collector, 2781 const char *phase, 2782 bool print_cr = true); 2783 ~CMSPhaseAccounting(); 2784 2785 private: 2786 CMSCollector *_collector; 2787 const char *_phase; 2788 elapsedTimer _wallclock; 2789 bool _print_cr; 2790 2791 public: 2792 // Not MT-safe; so do not pass around these StackObj's 2793 // where they may be accessed by other threads. 2794 jlong wallclock_millis() { 2795 assert(_wallclock.is_active(), "Wall clock should not stop"); 2796 _wallclock.stop(); // to record time 2797 jlong ret = _wallclock.milliseconds(); 2798 _wallclock.start(); // restart 2799 return ret; 2800 } 2801 }; 2802 2803 CMSPhaseAccounting::CMSPhaseAccounting(CMSCollector *collector, 2804 const char *phase, 2805 bool print_cr) : 2806 _collector(collector), _phase(phase), _print_cr(print_cr) { 2807 2808 if (PrintCMSStatistics != 0) { 2809 _collector->resetYields(); 2810 } 2811 if (PrintGCDetails) { 2812 gclog_or_tty->gclog_stamp(); 2813 gclog_or_tty->print_cr("[%s-concurrent-%s-start]", 2814 _collector->cmsGen()->short_name(), _phase); 2815 } 2816 _collector->resetTimer(); 2817 _wallclock.start(); 2818 _collector->startTimer(); 2819 } 2820 2821 CMSPhaseAccounting::~CMSPhaseAccounting() { 2822 assert(_wallclock.is_active(), "Wall clock should not have stopped"); 2823 _collector->stopTimer(); 2824 _wallclock.stop(); 2825 if (PrintGCDetails) { 2826 gclog_or_tty->gclog_stamp(); 2827 gclog_or_tty->print("[%s-concurrent-%s: %3.3f/%3.3f secs]", 2828 _collector->cmsGen()->short_name(), 2829 _phase, _collector->timerValue(), _wallclock.seconds()); 2830 if (_print_cr) { 2831 gclog_or_tty->cr(); 2832 } 2833 if (PrintCMSStatistics != 0) { 2834 gclog_or_tty->print_cr(" (CMS-concurrent-%s yielded %d times)", _phase, 2835 _collector->yields()); 2836 } 2837 } 2838 } 2839 2840 // CMS work 2841 2842 // The common parts of CMSParInitialMarkTask and CMSParRemarkTask. 2843 class CMSParMarkTask : public AbstractGangTask { 2844 protected: 2845 CMSCollector* _collector; 2846 uint _n_workers; 2847 CMSParMarkTask(const char* name, CMSCollector* collector, uint n_workers) : 2848 AbstractGangTask(name), 2849 _collector(collector), 2850 _n_workers(n_workers) {} 2851 // Work method in support of parallel rescan ... of young gen spaces 2852 void do_young_space_rescan(uint worker_id, OopsInGenClosure* cl, 2853 ContiguousSpace* space, 2854 HeapWord** chunk_array, size_t chunk_top); 2855 void work_on_young_gen_roots(uint worker_id, OopsInGenClosure* cl); 2856 }; 2857 2858 // Parallel initial mark task 2859 class CMSParInitialMarkTask: public CMSParMarkTask { 2860 StrongRootsScope* _strong_roots_scope; 2861 public: 2862 CMSParInitialMarkTask(CMSCollector* collector, StrongRootsScope* strong_roots_scope, uint n_workers) : 2863 CMSParMarkTask("Scan roots and young gen for initial mark in parallel", collector, n_workers), 2864 _strong_roots_scope(strong_roots_scope) {} 2865 void work(uint worker_id); 2866 }; 2867 2868 // Checkpoint the roots into this generation from outside 2869 // this generation. [Note this initial checkpoint need only 2870 // be approximate -- we'll do a catch up phase subsequently.] 2871 void CMSCollector::checkpointRootsInitial() { 2872 assert(_collectorState == InitialMarking, "Wrong collector state"); 2873 check_correct_thread_executing(); 2874 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 2875 2876 save_heap_summary(); 2877 report_heap_summary(GCWhen::BeforeGC); 2878 2879 ReferenceProcessor* rp = ref_processor(); 2880 assert(_restart_addr == NULL, "Control point invariant"); 2881 { 2882 // acquire locks for subsequent manipulations 2883 MutexLockerEx x(bitMapLock(), 2884 Mutex::_no_safepoint_check_flag); 2885 checkpointRootsInitialWork(); 2886 // enable ("weak") refs discovery 2887 rp->enable_discovery(); 2888 _collectorState = Marking; 2889 } 2890 } 2891 2892 void CMSCollector::checkpointRootsInitialWork() { 2893 assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped"); 2894 assert(_collectorState == InitialMarking, "just checking"); 2895 2896 // Already have locks. 2897 assert_lock_strong(bitMapLock()); 2898 assert(_markBitMap.isAllClear(), "was reset at end of previous cycle"); 2899 2900 // Setup the verification and class unloading state for this 2901 // CMS collection cycle. 2902 setup_cms_unloading_and_verification_state(); 2903 2904 NOT_PRODUCT(GCTraceTime t("\ncheckpointRootsInitialWork", 2905 PrintGCDetails && Verbose, true, _gc_timer_cm);) 2906 2907 // Reset all the PLAB chunk arrays if necessary. 2908 if (_survivor_plab_array != NULL && !CMSPLABRecordAlways) { 2909 reset_survivor_plab_arrays(); 2910 } 2911 2912 ResourceMark rm; 2913 HandleMark hm; 2914 2915 MarkRefsIntoClosure notOlder(_span, &_markBitMap); 2916 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2917 2918 verify_work_stacks_empty(); 2919 verify_overflow_empty(); 2920 2921 gch->ensure_parsability(false); // fill TLABs, but no need to retire them 2922 // Update the saved marks which may affect the root scans. 2923 gch->save_marks(); 2924 2925 // weak reference processing has not started yet. 2926 ref_processor()->set_enqueuing_is_done(false); 2927 2928 // Need to remember all newly created CLDs, 2929 // so that we can guarantee that the remark finds them. 2930 ClassLoaderDataGraph::remember_new_clds(true); 2931 2932 // Whenever a CLD is found, it will be claimed before proceeding to mark 2933 // the klasses. The claimed marks need to be cleared before marking starts. 2934 ClassLoaderDataGraph::clear_claimed_marks(); 2935 2936 if (CMSPrintEdenSurvivorChunks) { 2937 print_eden_and_survivor_chunk_arrays(); 2938 } 2939 2940 { 2941 #if defined(COMPILER2) || INCLUDE_JVMCI 2942 DerivedPointerTableDeactivate dpt_deact; 2943 #endif 2944 if (CMSParallelInitialMarkEnabled) { 2945 // The parallel version. 2946 WorkGang* workers = gch->workers(); 2947 assert(workers != NULL, "Need parallel worker threads."); 2948 uint n_workers = workers->active_workers(); 2949 2950 StrongRootsScope srs(n_workers); 2951 2952 CMSParInitialMarkTask tsk(this, &srs, n_workers); 2953 initialize_sequential_subtasks_for_young_gen_rescan(n_workers); 2954 if (n_workers > 1) { 2955 workers->run_task(&tsk); 2956 } else { 2957 tsk.work(0); 2958 } 2959 } else { 2960 // The serial version. 2961 CLDToOopClosure cld_closure(¬Older, true); 2962 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2963 2964 StrongRootsScope srs(1); 2965 2966 gch->gen_process_roots(&srs, 2967 GenCollectedHeap::OldGen, 2968 true, // young gen as roots 2969 GenCollectedHeap::ScanningOption(roots_scanning_options()), 2970 should_unload_classes(), 2971 ¬Older, 2972 NULL, 2973 &cld_closure); 2974 } 2975 } 2976 2977 // Clear mod-union table; it will be dirtied in the prologue of 2978 // CMS generation per each young generation collection. 2979 2980 assert(_modUnionTable.isAllClear(), 2981 "Was cleared in most recent final checkpoint phase" 2982 " or no bits are set in the gc_prologue before the start of the next " 2983 "subsequent marking phase."); 2984 2985 assert(_ct->klass_rem_set()->mod_union_is_clear(), "Must be"); 2986 2987 // Save the end of the used_region of the constituent generations 2988 // to be used to limit the extent of sweep in each generation. 2989 save_sweep_limits(); 2990 verify_overflow_empty(); 2991 } 2992 2993 bool CMSCollector::markFromRoots() { 2994 // we might be tempted to assert that: 2995 // assert(!SafepointSynchronize::is_at_safepoint(), 2996 // "inconsistent argument?"); 2997 // However that wouldn't be right, because it's possible that 2998 // a safepoint is indeed in progress as a young generation 2999 // stop-the-world GC happens even as we mark in this generation. 3000 assert(_collectorState == Marking, "inconsistent state?"); 3001 check_correct_thread_executing(); 3002 verify_overflow_empty(); 3003 3004 // Weak ref discovery note: We may be discovering weak 3005 // refs in this generation concurrent (but interleaved) with 3006 // weak ref discovery by the young generation collector. 3007 3008 CMSTokenSyncWithLocks ts(true, bitMapLock()); 3009 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 3010 CMSPhaseAccounting pa(this, "mark", !PrintGCDetails); 3011 bool res = markFromRootsWork(); 3012 if (res) { 3013 _collectorState = Precleaning; 3014 } else { // We failed and a foreground collection wants to take over 3015 assert(_foregroundGCIsActive, "internal state inconsistency"); 3016 assert(_restart_addr == NULL, "foreground will restart from scratch"); 3017 if (PrintGCDetails) { 3018 gclog_or_tty->print_cr("bailing out to foreground collection"); 3019 } 3020 } 3021 verify_overflow_empty(); 3022 return res; 3023 } 3024 3025 bool CMSCollector::markFromRootsWork() { 3026 // iterate over marked bits in bit map, doing a full scan and mark 3027 // from these roots using the following algorithm: 3028 // . if oop is to the right of the current scan pointer, 3029 // mark corresponding bit (we'll process it later) 3030 // . else (oop is to left of current scan pointer) 3031 // push oop on marking stack 3032 // . drain the marking stack 3033 3034 // Note that when we do a marking step we need to hold the 3035 // bit map lock -- recall that direct allocation (by mutators) 3036 // and promotion (by the young generation collector) is also 3037 // marking the bit map. [the so-called allocate live policy.] 3038 // Because the implementation of bit map marking is not 3039 // robust wrt simultaneous marking of bits in the same word, 3040 // we need to make sure that there is no such interference 3041 // between concurrent such updates. 3042 3043 // already have locks 3044 assert_lock_strong(bitMapLock()); 3045 3046 verify_work_stacks_empty(); 3047 verify_overflow_empty(); 3048 bool result = false; 3049 if (CMSConcurrentMTEnabled && ConcGCThreads > 0) { 3050 result = do_marking_mt(); 3051 } else { 3052 result = do_marking_st(); 3053 } 3054 return result; 3055 } 3056 3057 // Forward decl 3058 class CMSConcMarkingTask; 3059 3060 class CMSConcMarkingTerminator: public ParallelTaskTerminator { 3061 CMSCollector* _collector; 3062 CMSConcMarkingTask* _task; 3063 public: 3064 virtual void yield(); 3065 3066 // "n_threads" is the number of threads to be terminated. 3067 // "queue_set" is a set of work queues of other threads. 3068 // "collector" is the CMS collector associated with this task terminator. 3069 // "yield" indicates whether we need the gang as a whole to yield. 3070 CMSConcMarkingTerminator(int n_threads, TaskQueueSetSuper* queue_set, CMSCollector* collector) : 3071 ParallelTaskTerminator(n_threads, queue_set), 3072 _collector(collector) { } 3073 3074 void set_task(CMSConcMarkingTask* task) { 3075 _task = task; 3076 } 3077 }; 3078 3079 class CMSConcMarkingTerminatorTerminator: public TerminatorTerminator { 3080 CMSConcMarkingTask* _task; 3081 public: 3082 bool should_exit_termination(); 3083 void set_task(CMSConcMarkingTask* task) { 3084 _task = task; 3085 } 3086 }; 3087 3088 // MT Concurrent Marking Task 3089 class CMSConcMarkingTask: public YieldingFlexibleGangTask { 3090 CMSCollector* _collector; 3091 uint _n_workers; // requested/desired # workers 3092 bool _result; 3093 CompactibleFreeListSpace* _cms_space; 3094 char _pad_front[64]; // padding to ... 3095 HeapWord* _global_finger; // ... avoid sharing cache line 3096 char _pad_back[64]; 3097 HeapWord* _restart_addr; 3098 3099 // Exposed here for yielding support 3100 Mutex* const _bit_map_lock; 3101 3102 // The per thread work queues, available here for stealing 3103 OopTaskQueueSet* _task_queues; 3104 3105 // Termination (and yielding) support 3106 CMSConcMarkingTerminator _term; 3107 CMSConcMarkingTerminatorTerminator _term_term; 3108 3109 public: 3110 CMSConcMarkingTask(CMSCollector* collector, 3111 CompactibleFreeListSpace* cms_space, 3112 YieldingFlexibleWorkGang* workers, 3113 OopTaskQueueSet* task_queues): 3114 YieldingFlexibleGangTask("Concurrent marking done multi-threaded"), 3115 _collector(collector), 3116 _cms_space(cms_space), 3117 _n_workers(0), _result(true), 3118 _task_queues(task_queues), 3119 _term(_n_workers, task_queues, _collector), 3120 _bit_map_lock(collector->bitMapLock()) 3121 { 3122 _requested_size = _n_workers; 3123 _term.set_task(this); 3124 _term_term.set_task(this); 3125 _restart_addr = _global_finger = _cms_space->bottom(); 3126 } 3127 3128 3129 OopTaskQueueSet* task_queues() { return _task_queues; } 3130 3131 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 3132 3133 HeapWord** global_finger_addr() { return &_global_finger; } 3134 3135 CMSConcMarkingTerminator* terminator() { return &_term; } 3136 3137 virtual void set_for_termination(uint active_workers) { 3138 terminator()->reset_for_reuse(active_workers); 3139 } 3140 3141 void work(uint worker_id); 3142 bool should_yield() { 3143 return ConcurrentMarkSweepThread::should_yield() 3144 && !_collector->foregroundGCIsActive(); 3145 } 3146 3147 virtual void coordinator_yield(); // stuff done by coordinator 3148 bool result() { return _result; } 3149 3150 void reset(HeapWord* ra) { 3151 assert(_global_finger >= _cms_space->end(), "Postcondition of ::work(i)"); 3152 _restart_addr = _global_finger = ra; 3153 _term.reset_for_reuse(); 3154 } 3155 3156 static bool get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 3157 OopTaskQueue* work_q); 3158 3159 private: 3160 void do_scan_and_mark(int i, CompactibleFreeListSpace* sp); 3161 void do_work_steal(int i); 3162 void bump_global_finger(HeapWord* f); 3163 }; 3164 3165 bool CMSConcMarkingTerminatorTerminator::should_exit_termination() { 3166 assert(_task != NULL, "Error"); 3167 return _task->yielding(); 3168 // Note that we do not need the disjunct || _task->should_yield() above 3169 // because we want terminating threads to yield only if the task 3170 // is already in the midst of yielding, which happens only after at least one 3171 // thread has yielded. 3172 } 3173 3174 void CMSConcMarkingTerminator::yield() { 3175 if (_task->should_yield()) { 3176 _task->yield(); 3177 } else { 3178 ParallelTaskTerminator::yield(); 3179 } 3180 } 3181 3182 //////////////////////////////////////////////////////////////// 3183 // Concurrent Marking Algorithm Sketch 3184 //////////////////////////////////////////////////////////////// 3185 // Until all tasks exhausted (both spaces): 3186 // -- claim next available chunk 3187 // -- bump global finger via CAS 3188 // -- find first object that starts in this chunk 3189 // and start scanning bitmap from that position 3190 // -- scan marked objects for oops 3191 // -- CAS-mark target, and if successful: 3192 // . if target oop is above global finger (volatile read) 3193 // nothing to do 3194 // . if target oop is in chunk and above local finger 3195 // then nothing to do 3196 // . else push on work-queue 3197 // -- Deal with possible overflow issues: 3198 // . local work-queue overflow causes stuff to be pushed on 3199 // global (common) overflow queue 3200 // . always first empty local work queue 3201 // . then get a batch of oops from global work queue if any 3202 // . then do work stealing 3203 // -- When all tasks claimed (both spaces) 3204 // and local work queue empty, 3205 // then in a loop do: 3206 // . check global overflow stack; steal a batch of oops and trace 3207 // . try to steal from other threads oif GOS is empty 3208 // . if neither is available, offer termination 3209 // -- Terminate and return result 3210 // 3211 void CMSConcMarkingTask::work(uint worker_id) { 3212 elapsedTimer _timer; 3213 ResourceMark rm; 3214 HandleMark hm; 3215 3216 DEBUG_ONLY(_collector->verify_overflow_empty();) 3217 3218 // Before we begin work, our work queue should be empty 3219 assert(work_queue(worker_id)->size() == 0, "Expected to be empty"); 3220 // Scan the bitmap covering _cms_space, tracing through grey objects. 3221 _timer.start(); 3222 do_scan_and_mark(worker_id, _cms_space); 3223 _timer.stop(); 3224 if (PrintCMSStatistics != 0) { 3225 gclog_or_tty->print_cr("Finished cms space scanning in %dth thread: %3.3f sec", 3226 worker_id, _timer.seconds()); 3227 // XXX: need xxx/xxx type of notation, two timers 3228 } 3229 3230 // ... do work stealing 3231 _timer.reset(); 3232 _timer.start(); 3233 do_work_steal(worker_id); 3234 _timer.stop(); 3235 if (PrintCMSStatistics != 0) { 3236 gclog_or_tty->print_cr("Finished work stealing in %dth thread: %3.3f sec", 3237 worker_id, _timer.seconds()); 3238 // XXX: need xxx/xxx type of notation, two timers 3239 } 3240 assert(_collector->_markStack.isEmpty(), "Should have been emptied"); 3241 assert(work_queue(worker_id)->size() == 0, "Should have been emptied"); 3242 // Note that under the current task protocol, the 3243 // following assertion is true even of the spaces 3244 // expanded since the completion of the concurrent 3245 // marking. XXX This will likely change under a strict 3246 // ABORT semantics. 3247 // After perm removal the comparison was changed to 3248 // greater than or equal to from strictly greater than. 3249 // Before perm removal the highest address sweep would 3250 // have been at the end of perm gen but now is at the 3251 // end of the tenured gen. 3252 assert(_global_finger >= _cms_space->end(), 3253 "All tasks have been completed"); 3254 DEBUG_ONLY(_collector->verify_overflow_empty();) 3255 } 3256 3257 void CMSConcMarkingTask::bump_global_finger(HeapWord* f) { 3258 HeapWord* read = _global_finger; 3259 HeapWord* cur = read; 3260 while (f > read) { 3261 cur = read; 3262 read = (HeapWord*) Atomic::cmpxchg_ptr(f, &_global_finger, cur); 3263 if (cur == read) { 3264 // our cas succeeded 3265 assert(_global_finger >= f, "protocol consistency"); 3266 break; 3267 } 3268 } 3269 } 3270 3271 // This is really inefficient, and should be redone by 3272 // using (not yet available) block-read and -write interfaces to the 3273 // stack and the work_queue. XXX FIX ME !!! 3274 bool CMSConcMarkingTask::get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 3275 OopTaskQueue* work_q) { 3276 // Fast lock-free check 3277 if (ovflw_stk->length() == 0) { 3278 return false; 3279 } 3280 assert(work_q->size() == 0, "Shouldn't steal"); 3281 MutexLockerEx ml(ovflw_stk->par_lock(), 3282 Mutex::_no_safepoint_check_flag); 3283 // Grab up to 1/4 the size of the work queue 3284 size_t num = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 3285 (size_t)ParGCDesiredObjsFromOverflowList); 3286 num = MIN2(num, ovflw_stk->length()); 3287 for (int i = (int) num; i > 0; i--) { 3288 oop cur = ovflw_stk->pop(); 3289 assert(cur != NULL, "Counted wrong?"); 3290 work_q->push(cur); 3291 } 3292 return num > 0; 3293 } 3294 3295 void CMSConcMarkingTask::do_scan_and_mark(int i, CompactibleFreeListSpace* sp) { 3296 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 3297 int n_tasks = pst->n_tasks(); 3298 // We allow that there may be no tasks to do here because 3299 // we are restarting after a stack overflow. 3300 assert(pst->valid() || n_tasks == 0, "Uninitialized use?"); 3301 uint nth_task = 0; 3302 3303 HeapWord* aligned_start = sp->bottom(); 3304 if (sp->used_region().contains(_restart_addr)) { 3305 // Align down to a card boundary for the start of 0th task 3306 // for this space. 3307 aligned_start = 3308 (HeapWord*)align_size_down((uintptr_t)_restart_addr, 3309 CardTableModRefBS::card_size); 3310 } 3311 3312 size_t chunk_size = sp->marking_task_size(); 3313 while (!pst->is_task_claimed(/* reference */ nth_task)) { 3314 // Having claimed the nth task in this space, 3315 // compute the chunk that it corresponds to: 3316 MemRegion span = MemRegion(aligned_start + nth_task*chunk_size, 3317 aligned_start + (nth_task+1)*chunk_size); 3318 // Try and bump the global finger via a CAS; 3319 // note that we need to do the global finger bump 3320 // _before_ taking the intersection below, because 3321 // the task corresponding to that region will be 3322 // deemed done even if the used_region() expands 3323 // because of allocation -- as it almost certainly will 3324 // during start-up while the threads yield in the 3325 // closure below. 3326 HeapWord* finger = span.end(); 3327 bump_global_finger(finger); // atomically 3328 // There are null tasks here corresponding to chunks 3329 // beyond the "top" address of the space. 3330 span = span.intersection(sp->used_region()); 3331 if (!span.is_empty()) { // Non-null task 3332 HeapWord* prev_obj; 3333 assert(!span.contains(_restart_addr) || nth_task == 0, 3334 "Inconsistency"); 3335 if (nth_task == 0) { 3336 // For the 0th task, we'll not need to compute a block_start. 3337 if (span.contains(_restart_addr)) { 3338 // In the case of a restart because of stack overflow, 3339 // we might additionally skip a chunk prefix. 3340 prev_obj = _restart_addr; 3341 } else { 3342 prev_obj = span.start(); 3343 } 3344 } else { 3345 // We want to skip the first object because 3346 // the protocol is to scan any object in its entirety 3347 // that _starts_ in this span; a fortiori, any 3348 // object starting in an earlier span is scanned 3349 // as part of an earlier claimed task. 3350 // Below we use the "careful" version of block_start 3351 // so we do not try to navigate uninitialized objects. 3352 prev_obj = sp->block_start_careful(span.start()); 3353 // Below we use a variant of block_size that uses the 3354 // Printezis bits to avoid waiting for allocated 3355 // objects to become initialized/parsable. 3356 while (prev_obj < span.start()) { 3357 size_t sz = sp->block_size_no_stall(prev_obj, _collector); 3358 if (sz > 0) { 3359 prev_obj += sz; 3360 } else { 3361 // In this case we may end up doing a bit of redundant 3362 // scanning, but that appears unavoidable, short of 3363 // locking the free list locks; see bug 6324141. 3364 break; 3365 } 3366 } 3367 } 3368 if (prev_obj < span.end()) { 3369 MemRegion my_span = MemRegion(prev_obj, span.end()); 3370 // Do the marking work within a non-empty span -- 3371 // the last argument to the constructor indicates whether the 3372 // iteration should be incremental with periodic yields. 3373 Par_MarkFromRootsClosure cl(this, _collector, my_span, 3374 &_collector->_markBitMap, 3375 work_queue(i), 3376 &_collector->_markStack); 3377 _collector->_markBitMap.iterate(&cl, my_span.start(), my_span.end()); 3378 } // else nothing to do for this task 3379 } // else nothing to do for this task 3380 } 3381 // We'd be tempted to assert here that since there are no 3382 // more tasks left to claim in this space, the global_finger 3383 // must exceed space->top() and a fortiori space->end(). However, 3384 // that would not quite be correct because the bumping of 3385 // global_finger occurs strictly after the claiming of a task, 3386 // so by the time we reach here the global finger may not yet 3387 // have been bumped up by the thread that claimed the last 3388 // task. 3389 pst->all_tasks_completed(); 3390 } 3391 3392 class Par_ConcMarkingClosure: public MetadataAwareOopClosure { 3393 private: 3394 CMSCollector* _collector; 3395 CMSConcMarkingTask* _task; 3396 MemRegion _span; 3397 CMSBitMap* _bit_map; 3398 CMSMarkStack* _overflow_stack; 3399 OopTaskQueue* _work_queue; 3400 protected: 3401 DO_OOP_WORK_DEFN 3402 public: 3403 Par_ConcMarkingClosure(CMSCollector* collector, CMSConcMarkingTask* task, OopTaskQueue* work_queue, 3404 CMSBitMap* bit_map, CMSMarkStack* overflow_stack): 3405 MetadataAwareOopClosure(collector->ref_processor()), 3406 _collector(collector), 3407 _task(task), 3408 _span(collector->_span), 3409 _work_queue(work_queue), 3410 _bit_map(bit_map), 3411 _overflow_stack(overflow_stack) 3412 { } 3413 virtual void do_oop(oop* p); 3414 virtual void do_oop(narrowOop* p); 3415 3416 void trim_queue(size_t max); 3417 void handle_stack_overflow(HeapWord* lost); 3418 void do_yield_check() { 3419 if (_task->should_yield()) { 3420 _task->yield(); 3421 } 3422 } 3423 }; 3424 3425 // Grey object scanning during work stealing phase -- 3426 // the salient assumption here is that any references 3427 // that are in these stolen objects being scanned must 3428 // already have been initialized (else they would not have 3429 // been published), so we do not need to check for 3430 // uninitialized objects before pushing here. 3431 void Par_ConcMarkingClosure::do_oop(oop obj) { 3432 assert(obj->is_oop_or_null(true), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 3433 HeapWord* addr = (HeapWord*)obj; 3434 // Check if oop points into the CMS generation 3435 // and is not marked 3436 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 3437 // a white object ... 3438 // If we manage to "claim" the object, by being the 3439 // first thread to mark it, then we push it on our 3440 // marking stack 3441 if (_bit_map->par_mark(addr)) { // ... now grey 3442 // push on work queue (grey set) 3443 bool simulate_overflow = false; 3444 NOT_PRODUCT( 3445 if (CMSMarkStackOverflowALot && 3446 _collector->simulate_overflow()) { 3447 // simulate a stack overflow 3448 simulate_overflow = true; 3449 } 3450 ) 3451 if (simulate_overflow || 3452 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) { 3453 // stack overflow 3454 if (PrintCMSStatistics != 0) { 3455 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 3456 SIZE_FORMAT, _overflow_stack->capacity()); 3457 } 3458 // We cannot assert that the overflow stack is full because 3459 // it may have been emptied since. 3460 assert(simulate_overflow || 3461 _work_queue->size() == _work_queue->max_elems(), 3462 "Else push should have succeeded"); 3463 handle_stack_overflow(addr); 3464 } 3465 } // Else, some other thread got there first 3466 do_yield_check(); 3467 } 3468 } 3469 3470 void Par_ConcMarkingClosure::do_oop(oop* p) { Par_ConcMarkingClosure::do_oop_work(p); } 3471 void Par_ConcMarkingClosure::do_oop(narrowOop* p) { Par_ConcMarkingClosure::do_oop_work(p); } 3472 3473 void Par_ConcMarkingClosure::trim_queue(size_t max) { 3474 while (_work_queue->size() > max) { 3475 oop new_oop; 3476 if (_work_queue->pop_local(new_oop)) { 3477 assert(new_oop->is_oop(), "Should be an oop"); 3478 assert(_bit_map->isMarked((HeapWord*)new_oop), "Grey object"); 3479 assert(_span.contains((HeapWord*)new_oop), "Not in span"); 3480 new_oop->oop_iterate(this); // do_oop() above 3481 do_yield_check(); 3482 } 3483 } 3484 } 3485 3486 // Upon stack overflow, we discard (part of) the stack, 3487 // remembering the least address amongst those discarded 3488 // in CMSCollector's _restart_address. 3489 void Par_ConcMarkingClosure::handle_stack_overflow(HeapWord* lost) { 3490 // We need to do this under a mutex to prevent other 3491 // workers from interfering with the work done below. 3492 MutexLockerEx ml(_overflow_stack->par_lock(), 3493 Mutex::_no_safepoint_check_flag); 3494 // Remember the least grey address discarded 3495 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 3496 _collector->lower_restart_addr(ra); 3497 _overflow_stack->reset(); // discard stack contents 3498 _overflow_stack->expand(); // expand the stack if possible 3499 } 3500 3501 3502 void CMSConcMarkingTask::do_work_steal(int i) { 3503 OopTaskQueue* work_q = work_queue(i); 3504 oop obj_to_scan; 3505 CMSBitMap* bm = &(_collector->_markBitMap); 3506 CMSMarkStack* ovflw = &(_collector->_markStack); 3507 int* seed = _collector->hash_seed(i); 3508 Par_ConcMarkingClosure cl(_collector, this, work_q, bm, ovflw); 3509 while (true) { 3510 cl.trim_queue(0); 3511 assert(work_q->size() == 0, "Should have been emptied above"); 3512 if (get_work_from_overflow_stack(ovflw, work_q)) { 3513 // Can't assert below because the work obtained from the 3514 // overflow stack may already have been stolen from us. 3515 // assert(work_q->size() > 0, "Work from overflow stack"); 3516 continue; 3517 } else if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 3518 assert(obj_to_scan->is_oop(), "Should be an oop"); 3519 assert(bm->isMarked((HeapWord*)obj_to_scan), "Grey object"); 3520 obj_to_scan->oop_iterate(&cl); 3521 } else if (terminator()->offer_termination(&_term_term)) { 3522 assert(work_q->size() == 0, "Impossible!"); 3523 break; 3524 } else if (yielding() || should_yield()) { 3525 yield(); 3526 } 3527 } 3528 } 3529 3530 // This is run by the CMS (coordinator) thread. 3531 void CMSConcMarkingTask::coordinator_yield() { 3532 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 3533 "CMS thread should hold CMS token"); 3534 // First give up the locks, then yield, then re-lock 3535 // We should probably use a constructor/destructor idiom to 3536 // do this unlock/lock or modify the MutexUnlocker class to 3537 // serve our purpose. XXX 3538 assert_lock_strong(_bit_map_lock); 3539 _bit_map_lock->unlock(); 3540 ConcurrentMarkSweepThread::desynchronize(true); 3541 _collector->stopTimer(); 3542 if (PrintCMSStatistics != 0) { 3543 _collector->incrementYields(); 3544 } 3545 3546 // It is possible for whichever thread initiated the yield request 3547 // not to get a chance to wake up and take the bitmap lock between 3548 // this thread releasing it and reacquiring it. So, while the 3549 // should_yield() flag is on, let's sleep for a bit to give the 3550 // other thread a chance to wake up. The limit imposed on the number 3551 // of iterations is defensive, to avoid any unforseen circumstances 3552 // putting us into an infinite loop. Since it's always been this 3553 // (coordinator_yield()) method that was observed to cause the 3554 // problem, we are using a parameter (CMSCoordinatorYieldSleepCount) 3555 // which is by default non-zero. For the other seven methods that 3556 // also perform the yield operation, as are using a different 3557 // parameter (CMSYieldSleepCount) which is by default zero. This way we 3558 // can enable the sleeping for those methods too, if necessary. 3559 // See 6442774. 3560 // 3561 // We really need to reconsider the synchronization between the GC 3562 // thread and the yield-requesting threads in the future and we 3563 // should really use wait/notify, which is the recommended 3564 // way of doing this type of interaction. Additionally, we should 3565 // consolidate the eight methods that do the yield operation and they 3566 // are almost identical into one for better maintainability and 3567 // readability. See 6445193. 3568 // 3569 // Tony 2006.06.29 3570 for (unsigned i = 0; i < CMSCoordinatorYieldSleepCount && 3571 ConcurrentMarkSweepThread::should_yield() && 3572 !CMSCollector::foregroundGCIsActive(); ++i) { 3573 os::sleep(Thread::current(), 1, false); 3574 } 3575 3576 ConcurrentMarkSweepThread::synchronize(true); 3577 _bit_map_lock->lock_without_safepoint_check(); 3578 _collector->startTimer(); 3579 } 3580 3581 bool CMSCollector::do_marking_mt() { 3582 assert(ConcGCThreads > 0 && conc_workers() != NULL, "precondition"); 3583 uint num_workers = AdaptiveSizePolicy::calc_active_conc_workers(conc_workers()->total_workers(), 3584 conc_workers()->active_workers(), 3585 Threads::number_of_non_daemon_threads()); 3586 conc_workers()->set_active_workers(num_workers); 3587 3588 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 3589 3590 CMSConcMarkingTask tsk(this, 3591 cms_space, 3592 conc_workers(), 3593 task_queues()); 3594 3595 // Since the actual number of workers we get may be different 3596 // from the number we requested above, do we need to do anything different 3597 // below? In particular, may be we need to subclass the SequantialSubTasksDone 3598 // class?? XXX 3599 cms_space ->initialize_sequential_subtasks_for_marking(num_workers); 3600 3601 // Refs discovery is already non-atomic. 3602 assert(!ref_processor()->discovery_is_atomic(), "Should be non-atomic"); 3603 assert(ref_processor()->discovery_is_mt(), "Discovery should be MT"); 3604 conc_workers()->start_task(&tsk); 3605 while (tsk.yielded()) { 3606 tsk.coordinator_yield(); 3607 conc_workers()->continue_task(&tsk); 3608 } 3609 // If the task was aborted, _restart_addr will be non-NULL 3610 assert(tsk.completed() || _restart_addr != NULL, "Inconsistency"); 3611 while (_restart_addr != NULL) { 3612 // XXX For now we do not make use of ABORTED state and have not 3613 // yet implemented the right abort semantics (even in the original 3614 // single-threaded CMS case). That needs some more investigation 3615 // and is deferred for now; see CR# TBF. 07252005YSR. XXX 3616 assert(!CMSAbortSemantics || tsk.aborted(), "Inconsistency"); 3617 // If _restart_addr is non-NULL, a marking stack overflow 3618 // occurred; we need to do a fresh marking iteration from the 3619 // indicated restart address. 3620 if (_foregroundGCIsActive) { 3621 // We may be running into repeated stack overflows, having 3622 // reached the limit of the stack size, while making very 3623 // slow forward progress. It may be best to bail out and 3624 // let the foreground collector do its job. 3625 // Clear _restart_addr, so that foreground GC 3626 // works from scratch. This avoids the headache of 3627 // a "rescan" which would otherwise be needed because 3628 // of the dirty mod union table & card table. 3629 _restart_addr = NULL; 3630 return false; 3631 } 3632 // Adjust the task to restart from _restart_addr 3633 tsk.reset(_restart_addr); 3634 cms_space ->initialize_sequential_subtasks_for_marking(num_workers, 3635 _restart_addr); 3636 _restart_addr = NULL; 3637 // Get the workers going again 3638 conc_workers()->start_task(&tsk); 3639 while (tsk.yielded()) { 3640 tsk.coordinator_yield(); 3641 conc_workers()->continue_task(&tsk); 3642 } 3643 } 3644 assert(tsk.completed(), "Inconsistency"); 3645 assert(tsk.result() == true, "Inconsistency"); 3646 return true; 3647 } 3648 3649 bool CMSCollector::do_marking_st() { 3650 ResourceMark rm; 3651 HandleMark hm; 3652 3653 // Temporarily make refs discovery single threaded (non-MT) 3654 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false); 3655 MarkFromRootsClosure markFromRootsClosure(this, _span, &_markBitMap, 3656 &_markStack, CMSYield); 3657 // the last argument to iterate indicates whether the iteration 3658 // should be incremental with periodic yields. 3659 _markBitMap.iterate(&markFromRootsClosure); 3660 // If _restart_addr is non-NULL, a marking stack overflow 3661 // occurred; we need to do a fresh iteration from the 3662 // indicated restart address. 3663 while (_restart_addr != NULL) { 3664 if (_foregroundGCIsActive) { 3665 // We may be running into repeated stack overflows, having 3666 // reached the limit of the stack size, while making very 3667 // slow forward progress. It may be best to bail out and 3668 // let the foreground collector do its job. 3669 // Clear _restart_addr, so that foreground GC 3670 // works from scratch. This avoids the headache of 3671 // a "rescan" which would otherwise be needed because 3672 // of the dirty mod union table & card table. 3673 _restart_addr = NULL; 3674 return false; // indicating failure to complete marking 3675 } 3676 // Deal with stack overflow: 3677 // we restart marking from _restart_addr 3678 HeapWord* ra = _restart_addr; 3679 markFromRootsClosure.reset(ra); 3680 _restart_addr = NULL; 3681 _markBitMap.iterate(&markFromRootsClosure, ra, _span.end()); 3682 } 3683 return true; 3684 } 3685 3686 void CMSCollector::preclean() { 3687 check_correct_thread_executing(); 3688 assert(Thread::current()->is_ConcurrentGC_thread(), "Wrong thread"); 3689 verify_work_stacks_empty(); 3690 verify_overflow_empty(); 3691 _abort_preclean = false; 3692 if (CMSPrecleaningEnabled) { 3693 if (!CMSEdenChunksRecordAlways) { 3694 _eden_chunk_index = 0; 3695 } 3696 size_t used = get_eden_used(); 3697 size_t capacity = get_eden_capacity(); 3698 // Don't start sampling unless we will get sufficiently 3699 // many samples. 3700 if (used < (capacity/(CMSScheduleRemarkSamplingRatio * 100) 3701 * CMSScheduleRemarkEdenPenetration)) { 3702 _start_sampling = true; 3703 } else { 3704 _start_sampling = false; 3705 } 3706 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 3707 CMSPhaseAccounting pa(this, "preclean", !PrintGCDetails); 3708 preclean_work(CMSPrecleanRefLists1, CMSPrecleanSurvivors1); 3709 } 3710 CMSTokenSync x(true); // is cms thread 3711 if (CMSPrecleaningEnabled) { 3712 sample_eden(); 3713 _collectorState = AbortablePreclean; 3714 } else { 3715 _collectorState = FinalMarking; 3716 } 3717 verify_work_stacks_empty(); 3718 verify_overflow_empty(); 3719 } 3720 3721 // Try and schedule the remark such that young gen 3722 // occupancy is CMSScheduleRemarkEdenPenetration %. 3723 void CMSCollector::abortable_preclean() { 3724 check_correct_thread_executing(); 3725 assert(CMSPrecleaningEnabled, "Inconsistent control state"); 3726 assert(_collectorState == AbortablePreclean, "Inconsistent control state"); 3727 3728 // If Eden's current occupancy is below this threshold, 3729 // immediately schedule the remark; else preclean 3730 // past the next scavenge in an effort to 3731 // schedule the pause as described above. By choosing 3732 // CMSScheduleRemarkEdenSizeThreshold >= max eden size 3733 // we will never do an actual abortable preclean cycle. 3734 if (get_eden_used() > CMSScheduleRemarkEdenSizeThreshold) { 3735 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 3736 CMSPhaseAccounting pa(this, "abortable-preclean", !PrintGCDetails); 3737 // We need more smarts in the abortable preclean 3738 // loop below to deal with cases where allocation 3739 // in young gen is very very slow, and our precleaning 3740 // is running a losing race against a horde of 3741 // mutators intent on flooding us with CMS updates 3742 // (dirty cards). 3743 // One, admittedly dumb, strategy is to give up 3744 // after a certain number of abortable precleaning loops 3745 // or after a certain maximum time. We want to make 3746 // this smarter in the next iteration. 3747 // XXX FIX ME!!! YSR 3748 size_t loops = 0, workdone = 0, cumworkdone = 0, waited = 0; 3749 while (!(should_abort_preclean() || 3750 ConcurrentMarkSweepThread::should_terminate())) { 3751 workdone = preclean_work(CMSPrecleanRefLists2, CMSPrecleanSurvivors2); 3752 cumworkdone += workdone; 3753 loops++; 3754 // Voluntarily terminate abortable preclean phase if we have 3755 // been at it for too long. 3756 if ((CMSMaxAbortablePrecleanLoops != 0) && 3757 loops >= CMSMaxAbortablePrecleanLoops) { 3758 if (PrintGCDetails) { 3759 gclog_or_tty->print(" CMS: abort preclean due to loops "); 3760 } 3761 break; 3762 } 3763 if (pa.wallclock_millis() > CMSMaxAbortablePrecleanTime) { 3764 if (PrintGCDetails) { 3765 gclog_or_tty->print(" CMS: abort preclean due to time "); 3766 } 3767 break; 3768 } 3769 // If we are doing little work each iteration, we should 3770 // take a short break. 3771 if (workdone < CMSAbortablePrecleanMinWorkPerIteration) { 3772 // Sleep for some time, waiting for work to accumulate 3773 stopTimer(); 3774 cmsThread()->wait_on_cms_lock(CMSAbortablePrecleanWaitMillis); 3775 startTimer(); 3776 waited++; 3777 } 3778 } 3779 if (PrintCMSStatistics > 0) { 3780 gclog_or_tty->print(" [" SIZE_FORMAT " iterations, " SIZE_FORMAT " waits, " SIZE_FORMAT " cards)] ", 3781 loops, waited, cumworkdone); 3782 } 3783 } 3784 CMSTokenSync x(true); // is cms thread 3785 if (_collectorState != Idling) { 3786 assert(_collectorState == AbortablePreclean, 3787 "Spontaneous state transition?"); 3788 _collectorState = FinalMarking; 3789 } // Else, a foreground collection completed this CMS cycle. 3790 return; 3791 } 3792 3793 // Respond to an Eden sampling opportunity 3794 void CMSCollector::sample_eden() { 3795 // Make sure a young gc cannot sneak in between our 3796 // reading and recording of a sample. 3797 assert(Thread::current()->is_ConcurrentGC_thread(), 3798 "Only the cms thread may collect Eden samples"); 3799 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 3800 "Should collect samples while holding CMS token"); 3801 if (!_start_sampling) { 3802 return; 3803 } 3804 // When CMSEdenChunksRecordAlways is true, the eden chunk array 3805 // is populated by the young generation. 3806 if (_eden_chunk_array != NULL && !CMSEdenChunksRecordAlways) { 3807 if (_eden_chunk_index < _eden_chunk_capacity) { 3808 _eden_chunk_array[_eden_chunk_index] = *_top_addr; // take sample 3809 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr, 3810 "Unexpected state of Eden"); 3811 // We'd like to check that what we just sampled is an oop-start address; 3812 // however, we cannot do that here since the object may not yet have been 3813 // initialized. So we'll instead do the check when we _use_ this sample 3814 // later. 3815 if (_eden_chunk_index == 0 || 3816 (pointer_delta(_eden_chunk_array[_eden_chunk_index], 3817 _eden_chunk_array[_eden_chunk_index-1]) 3818 >= CMSSamplingGrain)) { 3819 _eden_chunk_index++; // commit sample 3820 } 3821 } 3822 } 3823 if ((_collectorState == AbortablePreclean) && !_abort_preclean) { 3824 size_t used = get_eden_used(); 3825 size_t capacity = get_eden_capacity(); 3826 assert(used <= capacity, "Unexpected state of Eden"); 3827 if (used > (capacity/100 * CMSScheduleRemarkEdenPenetration)) { 3828 _abort_preclean = true; 3829 } 3830 } 3831 } 3832 3833 3834 size_t CMSCollector::preclean_work(bool clean_refs, bool clean_survivor) { 3835 assert(_collectorState == Precleaning || 3836 _collectorState == AbortablePreclean, "incorrect state"); 3837 ResourceMark rm; 3838 HandleMark hm; 3839 3840 // Precleaning is currently not MT but the reference processor 3841 // may be set for MT. Disable it temporarily here. 3842 ReferenceProcessor* rp = ref_processor(); 3843 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(rp, false); 3844 3845 // Do one pass of scrubbing the discovered reference lists 3846 // to remove any reference objects with strongly-reachable 3847 // referents. 3848 if (clean_refs) { 3849 CMSPrecleanRefsYieldClosure yield_cl(this); 3850 assert(rp->span().equals(_span), "Spans should be equal"); 3851 CMSKeepAliveClosure keep_alive(this, _span, &_markBitMap, 3852 &_markStack, true /* preclean */); 3853 CMSDrainMarkingStackClosure complete_trace(this, 3854 _span, &_markBitMap, &_markStack, 3855 &keep_alive, true /* preclean */); 3856 3857 // We don't want this step to interfere with a young 3858 // collection because we don't want to take CPU 3859 // or memory bandwidth away from the young GC threads 3860 // (which may be as many as there are CPUs). 3861 // Note that we don't need to protect ourselves from 3862 // interference with mutators because they can't 3863 // manipulate the discovered reference lists nor affect 3864 // the computed reachability of the referents, the 3865 // only properties manipulated by the precleaning 3866 // of these reference lists. 3867 stopTimer(); 3868 CMSTokenSyncWithLocks x(true /* is cms thread */, 3869 bitMapLock()); 3870 startTimer(); 3871 sample_eden(); 3872 3873 // The following will yield to allow foreground 3874 // collection to proceed promptly. XXX YSR: 3875 // The code in this method may need further 3876 // tweaking for better performance and some restructuring 3877 // for cleaner interfaces. 3878 GCTimer *gc_timer = NULL; // Currently not tracing concurrent phases 3879 rp->preclean_discovered_references( 3880 rp->is_alive_non_header(), &keep_alive, &complete_trace, &yield_cl, 3881 gc_timer); 3882 } 3883 3884 if (clean_survivor) { // preclean the active survivor space(s) 3885 PushAndMarkClosure pam_cl(this, _span, ref_processor(), 3886 &_markBitMap, &_modUnionTable, 3887 &_markStack, true /* precleaning phase */); 3888 stopTimer(); 3889 CMSTokenSyncWithLocks ts(true /* is cms thread */, 3890 bitMapLock()); 3891 startTimer(); 3892 unsigned int before_count = 3893 GenCollectedHeap::heap()->total_collections(); 3894 SurvivorSpacePrecleanClosure 3895 sss_cl(this, _span, &_markBitMap, &_markStack, 3896 &pam_cl, before_count, CMSYield); 3897 _young_gen->from()->object_iterate_careful(&sss_cl); 3898 _young_gen->to()->object_iterate_careful(&sss_cl); 3899 } 3900 MarkRefsIntoAndScanClosure 3901 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable, 3902 &_markStack, this, CMSYield, 3903 true /* precleaning phase */); 3904 // CAUTION: The following closure has persistent state that may need to 3905 // be reset upon a decrease in the sequence of addresses it 3906 // processes. 3907 ScanMarkedObjectsAgainCarefullyClosure 3908 smoac_cl(this, _span, 3909 &_markBitMap, &_markStack, &mrias_cl, CMSYield); 3910 3911 // Preclean dirty cards in ModUnionTable and CardTable using 3912 // appropriate convergence criterion; 3913 // repeat CMSPrecleanIter times unless we find that 3914 // we are losing. 3915 assert(CMSPrecleanIter < 10, "CMSPrecleanIter is too large"); 3916 assert(CMSPrecleanNumerator < CMSPrecleanDenominator, 3917 "Bad convergence multiplier"); 3918 assert(CMSPrecleanThreshold >= 100, 3919 "Unreasonably low CMSPrecleanThreshold"); 3920 3921 size_t numIter, cumNumCards, lastNumCards, curNumCards; 3922 for (numIter = 0, cumNumCards = lastNumCards = curNumCards = 0; 3923 numIter < CMSPrecleanIter; 3924 numIter++, lastNumCards = curNumCards, cumNumCards += curNumCards) { 3925 curNumCards = preclean_mod_union_table(_cmsGen, &smoac_cl); 3926 if (Verbose && PrintGCDetails) { 3927 gclog_or_tty->print(" (modUnionTable: " SIZE_FORMAT " cards)", curNumCards); 3928 } 3929 // Either there are very few dirty cards, so re-mark 3930 // pause will be small anyway, or our pre-cleaning isn't 3931 // that much faster than the rate at which cards are being 3932 // dirtied, so we might as well stop and re-mark since 3933 // precleaning won't improve our re-mark time by much. 3934 if (curNumCards <= CMSPrecleanThreshold || 3935 (numIter > 0 && 3936 (curNumCards * CMSPrecleanDenominator > 3937 lastNumCards * CMSPrecleanNumerator))) { 3938 numIter++; 3939 cumNumCards += curNumCards; 3940 break; 3941 } 3942 } 3943 3944 preclean_klasses(&mrias_cl, _cmsGen->freelistLock()); 3945 3946 curNumCards = preclean_card_table(_cmsGen, &smoac_cl); 3947 cumNumCards += curNumCards; 3948 if (PrintGCDetails && PrintCMSStatistics != 0) { 3949 gclog_or_tty->print_cr(" (cardTable: " SIZE_FORMAT " cards, re-scanned " SIZE_FORMAT " cards, " SIZE_FORMAT " iterations)", 3950 curNumCards, cumNumCards, numIter); 3951 } 3952 return cumNumCards; // as a measure of useful work done 3953 } 3954 3955 // PRECLEANING NOTES: 3956 // Precleaning involves: 3957 // . reading the bits of the modUnionTable and clearing the set bits. 3958 // . For the cards corresponding to the set bits, we scan the 3959 // objects on those cards. This means we need the free_list_lock 3960 // so that we can safely iterate over the CMS space when scanning 3961 // for oops. 3962 // . When we scan the objects, we'll be both reading and setting 3963 // marks in the marking bit map, so we'll need the marking bit map. 3964 // . For protecting _collector_state transitions, we take the CGC_lock. 3965 // Note that any races in the reading of of card table entries by the 3966 // CMS thread on the one hand and the clearing of those entries by the 3967 // VM thread or the setting of those entries by the mutator threads on the 3968 // other are quite benign. However, for efficiency it makes sense to keep 3969 // the VM thread from racing with the CMS thread while the latter is 3970 // dirty card info to the modUnionTable. We therefore also use the 3971 // CGC_lock to protect the reading of the card table and the mod union 3972 // table by the CM thread. 3973 // . We run concurrently with mutator updates, so scanning 3974 // needs to be done carefully -- we should not try to scan 3975 // potentially uninitialized objects. 3976 // 3977 // Locking strategy: While holding the CGC_lock, we scan over and 3978 // reset a maximal dirty range of the mod union / card tables, then lock 3979 // the free_list_lock and bitmap lock to do a full marking, then 3980 // release these locks; and repeat the cycle. This allows for a 3981 // certain amount of fairness in the sharing of these locks between 3982 // the CMS collector on the one hand, and the VM thread and the 3983 // mutators on the other. 3984 3985 // NOTE: preclean_mod_union_table() and preclean_card_table() 3986 // further below are largely identical; if you need to modify 3987 // one of these methods, please check the other method too. 3988 3989 size_t CMSCollector::preclean_mod_union_table( 3990 ConcurrentMarkSweepGeneration* old_gen, 3991 ScanMarkedObjectsAgainCarefullyClosure* cl) { 3992 verify_work_stacks_empty(); 3993 verify_overflow_empty(); 3994 3995 // strategy: starting with the first card, accumulate contiguous 3996 // ranges of dirty cards; clear these cards, then scan the region 3997 // covered by these cards. 3998 3999 // Since all of the MUT is committed ahead, we can just use 4000 // that, in case the generations expand while we are precleaning. 4001 // It might also be fine to just use the committed part of the 4002 // generation, but we might potentially miss cards when the 4003 // generation is rapidly expanding while we are in the midst 4004 // of precleaning. 4005 HeapWord* startAddr = old_gen->reserved().start(); 4006 HeapWord* endAddr = old_gen->reserved().end(); 4007 4008 cl->setFreelistLock(old_gen->freelistLock()); // needed for yielding 4009 4010 size_t numDirtyCards, cumNumDirtyCards; 4011 HeapWord *nextAddr, *lastAddr; 4012 for (cumNumDirtyCards = numDirtyCards = 0, 4013 nextAddr = lastAddr = startAddr; 4014 nextAddr < endAddr; 4015 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 4016 4017 ResourceMark rm; 4018 HandleMark hm; 4019 4020 MemRegion dirtyRegion; 4021 { 4022 stopTimer(); 4023 // Potential yield point 4024 CMSTokenSync ts(true); 4025 startTimer(); 4026 sample_eden(); 4027 // Get dirty region starting at nextOffset (inclusive), 4028 // simultaneously clearing it. 4029 dirtyRegion = 4030 _modUnionTable.getAndClearMarkedRegion(nextAddr, endAddr); 4031 assert(dirtyRegion.start() >= nextAddr, 4032 "returned region inconsistent?"); 4033 } 4034 // Remember where the next search should begin. 4035 // The returned region (if non-empty) is a right open interval, 4036 // so lastOffset is obtained from the right end of that 4037 // interval. 4038 lastAddr = dirtyRegion.end(); 4039 // Should do something more transparent and less hacky XXX 4040 numDirtyCards = 4041 _modUnionTable.heapWordDiffToOffsetDiff(dirtyRegion.word_size()); 4042 4043 // We'll scan the cards in the dirty region (with periodic 4044 // yields for foreground GC as needed). 4045 if (!dirtyRegion.is_empty()) { 4046 assert(numDirtyCards > 0, "consistency check"); 4047 HeapWord* stop_point = NULL; 4048 stopTimer(); 4049 // Potential yield point 4050 CMSTokenSyncWithLocks ts(true, old_gen->freelistLock(), 4051 bitMapLock()); 4052 startTimer(); 4053 { 4054 verify_work_stacks_empty(); 4055 verify_overflow_empty(); 4056 sample_eden(); 4057 stop_point = 4058 old_gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 4059 } 4060 if (stop_point != NULL) { 4061 // The careful iteration stopped early either because it found an 4062 // uninitialized object, or because we were in the midst of an 4063 // "abortable preclean", which should now be aborted. Redirty 4064 // the bits corresponding to the partially-scanned or unscanned 4065 // cards. We'll either restart at the next block boundary or 4066 // abort the preclean. 4067 assert((_collectorState == AbortablePreclean && should_abort_preclean()), 4068 "Should only be AbortablePreclean."); 4069 _modUnionTable.mark_range(MemRegion(stop_point, dirtyRegion.end())); 4070 if (should_abort_preclean()) { 4071 break; // out of preclean loop 4072 } else { 4073 // Compute the next address at which preclean should pick up; 4074 // might need bitMapLock in order to read P-bits. 4075 lastAddr = next_card_start_after_block(stop_point); 4076 } 4077 } 4078 } else { 4079 assert(lastAddr == endAddr, "consistency check"); 4080 assert(numDirtyCards == 0, "consistency check"); 4081 break; 4082 } 4083 } 4084 verify_work_stacks_empty(); 4085 verify_overflow_empty(); 4086 return cumNumDirtyCards; 4087 } 4088 4089 // NOTE: preclean_mod_union_table() above and preclean_card_table() 4090 // below are largely identical; if you need to modify 4091 // one of these methods, please check the other method too. 4092 4093 size_t CMSCollector::preclean_card_table(ConcurrentMarkSweepGeneration* old_gen, 4094 ScanMarkedObjectsAgainCarefullyClosure* cl) { 4095 // strategy: it's similar to precleamModUnionTable above, in that 4096 // we accumulate contiguous ranges of dirty cards, mark these cards 4097 // precleaned, then scan the region covered by these cards. 4098 HeapWord* endAddr = (HeapWord*)(old_gen->_virtual_space.high()); 4099 HeapWord* startAddr = (HeapWord*)(old_gen->_virtual_space.low()); 4100 4101 cl->setFreelistLock(old_gen->freelistLock()); // needed for yielding 4102 4103 size_t numDirtyCards, cumNumDirtyCards; 4104 HeapWord *lastAddr, *nextAddr; 4105 4106 for (cumNumDirtyCards = numDirtyCards = 0, 4107 nextAddr = lastAddr = startAddr; 4108 nextAddr < endAddr; 4109 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 4110 4111 ResourceMark rm; 4112 HandleMark hm; 4113 4114 MemRegion dirtyRegion; 4115 { 4116 // See comments in "Precleaning notes" above on why we 4117 // do this locking. XXX Could the locking overheads be 4118 // too high when dirty cards are sparse? [I don't think so.] 4119 stopTimer(); 4120 CMSTokenSync x(true); // is cms thread 4121 startTimer(); 4122 sample_eden(); 4123 // Get and clear dirty region from card table 4124 dirtyRegion = _ct->ct_bs()->dirty_card_range_after_reset( 4125 MemRegion(nextAddr, endAddr), 4126 true, 4127 CardTableModRefBS::precleaned_card_val()); 4128 4129 assert(dirtyRegion.start() >= nextAddr, 4130 "returned region inconsistent?"); 4131 } 4132 lastAddr = dirtyRegion.end(); 4133 numDirtyCards = 4134 dirtyRegion.word_size()/CardTableModRefBS::card_size_in_words; 4135 4136 if (!dirtyRegion.is_empty()) { 4137 stopTimer(); 4138 CMSTokenSyncWithLocks ts(true, old_gen->freelistLock(), bitMapLock()); 4139 startTimer(); 4140 sample_eden(); 4141 verify_work_stacks_empty(); 4142 verify_overflow_empty(); 4143 HeapWord* stop_point = 4144 old_gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 4145 if (stop_point != NULL) { 4146 assert((_collectorState == AbortablePreclean && should_abort_preclean()), 4147 "Should only be AbortablePreclean."); 4148 _ct->ct_bs()->invalidate(MemRegion(stop_point, dirtyRegion.end())); 4149 if (should_abort_preclean()) { 4150 break; // out of preclean loop 4151 } else { 4152 // Compute the next address at which preclean should pick up. 4153 lastAddr = next_card_start_after_block(stop_point); 4154 } 4155 } 4156 } else { 4157 break; 4158 } 4159 } 4160 verify_work_stacks_empty(); 4161 verify_overflow_empty(); 4162 return cumNumDirtyCards; 4163 } 4164 4165 class PrecleanKlassClosure : public KlassClosure { 4166 KlassToOopClosure _cm_klass_closure; 4167 public: 4168 PrecleanKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {} 4169 void do_klass(Klass* k) { 4170 if (k->has_accumulated_modified_oops()) { 4171 k->clear_accumulated_modified_oops(); 4172 4173 _cm_klass_closure.do_klass(k); 4174 } 4175 } 4176 }; 4177 4178 // The freelist lock is needed to prevent asserts, is it really needed? 4179 void CMSCollector::preclean_klasses(MarkRefsIntoAndScanClosure* cl, Mutex* freelistLock) { 4180 4181 cl->set_freelistLock(freelistLock); 4182 4183 CMSTokenSyncWithLocks ts(true, freelistLock, bitMapLock()); 4184 4185 // SSS: Add equivalent to ScanMarkedObjectsAgainCarefullyClosure::do_yield_check and should_abort_preclean? 4186 // SSS: We should probably check if precleaning should be aborted, at suitable intervals? 4187 PrecleanKlassClosure preclean_klass_closure(cl); 4188 ClassLoaderDataGraph::classes_do(&preclean_klass_closure); 4189 4190 verify_work_stacks_empty(); 4191 verify_overflow_empty(); 4192 } 4193 4194 void CMSCollector::checkpointRootsFinal() { 4195 assert(_collectorState == FinalMarking, "incorrect state transition?"); 4196 check_correct_thread_executing(); 4197 // world is stopped at this checkpoint 4198 assert(SafepointSynchronize::is_at_safepoint(), 4199 "world should be stopped"); 4200 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 4201 4202 verify_work_stacks_empty(); 4203 verify_overflow_empty(); 4204 4205 if (PrintGCDetails) { 4206 gclog_or_tty->print("[YG occupancy: " SIZE_FORMAT " K (" SIZE_FORMAT " K)]", 4207 _young_gen->used() / K, 4208 _young_gen->capacity() / K); 4209 } 4210 { 4211 if (CMSScavengeBeforeRemark) { 4212 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4213 // Temporarily set flag to false, GCH->do_collection will 4214 // expect it to be false and set to true 4215 FlagSetting fl(gch->_is_gc_active, false); 4216 NOT_PRODUCT(GCTraceTime t("Scavenge-Before-Remark", 4217 PrintGCDetails && Verbose, true, _gc_timer_cm);) 4218 gch->do_collection(true, // full (i.e. force, see below) 4219 false, // !clear_all_soft_refs 4220 0, // size 4221 false, // is_tlab 4222 GenCollectedHeap::YoungGen // type 4223 ); 4224 } 4225 FreelistLocker x(this); 4226 MutexLockerEx y(bitMapLock(), 4227 Mutex::_no_safepoint_check_flag); 4228 checkpointRootsFinalWork(); 4229 } 4230 verify_work_stacks_empty(); 4231 verify_overflow_empty(); 4232 } 4233 4234 void CMSCollector::checkpointRootsFinalWork() { 4235 NOT_PRODUCT(GCTraceTime tr("checkpointRootsFinalWork", PrintGCDetails, false, _gc_timer_cm);) 4236 4237 assert(haveFreelistLocks(), "must have free list locks"); 4238 assert_lock_strong(bitMapLock()); 4239 4240 ResourceMark rm; 4241 HandleMark hm; 4242 4243 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4244 4245 if (should_unload_classes()) { 4246 CodeCache::gc_prologue(); 4247 } 4248 assert(haveFreelistLocks(), "must have free list locks"); 4249 assert_lock_strong(bitMapLock()); 4250 4251 // We might assume that we need not fill TLAB's when 4252 // CMSScavengeBeforeRemark is set, because we may have just done 4253 // a scavenge which would have filled all TLAB's -- and besides 4254 // Eden would be empty. This however may not always be the case -- 4255 // for instance although we asked for a scavenge, it may not have 4256 // happened because of a JNI critical section. We probably need 4257 // a policy for deciding whether we can in that case wait until 4258 // the critical section releases and then do the remark following 4259 // the scavenge, and skip it here. In the absence of that policy, 4260 // or of an indication of whether the scavenge did indeed occur, 4261 // we cannot rely on TLAB's having been filled and must do 4262 // so here just in case a scavenge did not happen. 4263 gch->ensure_parsability(false); // fill TLAB's, but no need to retire them 4264 // Update the saved marks which may affect the root scans. 4265 gch->save_marks(); 4266 4267 if (CMSPrintEdenSurvivorChunks) { 4268 print_eden_and_survivor_chunk_arrays(); 4269 } 4270 4271 { 4272 #if defined(COMPILER2) || INCLUDE_JVMCI 4273 DerivedPointerTableDeactivate dpt_deact; 4274 #endif 4275 4276 // Note on the role of the mod union table: 4277 // Since the marker in "markFromRoots" marks concurrently with 4278 // mutators, it is possible for some reachable objects not to have been 4279 // scanned. For instance, an only reference to an object A was 4280 // placed in object B after the marker scanned B. Unless B is rescanned, 4281 // A would be collected. Such updates to references in marked objects 4282 // are detected via the mod union table which is the set of all cards 4283 // dirtied since the first checkpoint in this GC cycle and prior to 4284 // the most recent young generation GC, minus those cleaned up by the 4285 // concurrent precleaning. 4286 if (CMSParallelRemarkEnabled) { 4287 GCTraceTime t("Rescan (parallel) ", PrintGCDetails, false, _gc_timer_cm); 4288 do_remark_parallel(); 4289 } else { 4290 GCTraceTime t("Rescan (non-parallel) ", PrintGCDetails, false, _gc_timer_cm); 4291 do_remark_non_parallel(); 4292 } 4293 } 4294 verify_work_stacks_empty(); 4295 verify_overflow_empty(); 4296 4297 { 4298 NOT_PRODUCT(GCTraceTime ts("refProcessingWork", PrintGCDetails, false, _gc_timer_cm);) 4299 refProcessingWork(); 4300 } 4301 verify_work_stacks_empty(); 4302 verify_overflow_empty(); 4303 4304 if (should_unload_classes()) { 4305 CodeCache::gc_epilogue(); 4306 } 4307 JvmtiExport::gc_epilogue(); 4308 4309 // If we encountered any (marking stack / work queue) overflow 4310 // events during the current CMS cycle, take appropriate 4311 // remedial measures, where possible, so as to try and avoid 4312 // recurrence of that condition. 4313 assert(_markStack.isEmpty(), "No grey objects"); 4314 size_t ser_ovflw = _ser_pmc_remark_ovflw + _ser_pmc_preclean_ovflw + 4315 _ser_kac_ovflw + _ser_kac_preclean_ovflw; 4316 if (ser_ovflw > 0) { 4317 if (PrintCMSStatistics != 0) { 4318 gclog_or_tty->print_cr("Marking stack overflow (benign) " 4319 "(pmc_pc=" SIZE_FORMAT ", pmc_rm=" SIZE_FORMAT ", kac=" SIZE_FORMAT 4320 ", kac_preclean=" SIZE_FORMAT ")", 4321 _ser_pmc_preclean_ovflw, _ser_pmc_remark_ovflw, 4322 _ser_kac_ovflw, _ser_kac_preclean_ovflw); 4323 } 4324 _markStack.expand(); 4325 _ser_pmc_remark_ovflw = 0; 4326 _ser_pmc_preclean_ovflw = 0; 4327 _ser_kac_preclean_ovflw = 0; 4328 _ser_kac_ovflw = 0; 4329 } 4330 if (_par_pmc_remark_ovflw > 0 || _par_kac_ovflw > 0) { 4331 if (PrintCMSStatistics != 0) { 4332 gclog_or_tty->print_cr("Work queue overflow (benign) " 4333 "(pmc_rm=" SIZE_FORMAT ", kac=" SIZE_FORMAT ")", 4334 _par_pmc_remark_ovflw, _par_kac_ovflw); 4335 } 4336 _par_pmc_remark_ovflw = 0; 4337 _par_kac_ovflw = 0; 4338 } 4339 if (PrintCMSStatistics != 0) { 4340 if (_markStack._hit_limit > 0) { 4341 gclog_or_tty->print_cr(" (benign) Hit max stack size limit (" SIZE_FORMAT ")", 4342 _markStack._hit_limit); 4343 } 4344 if (_markStack._failed_double > 0) { 4345 gclog_or_tty->print_cr(" (benign) Failed stack doubling (" SIZE_FORMAT ")," 4346 " current capacity " SIZE_FORMAT, 4347 _markStack._failed_double, 4348 _markStack.capacity()); 4349 } 4350 } 4351 _markStack._hit_limit = 0; 4352 _markStack._failed_double = 0; 4353 4354 if ((VerifyAfterGC || VerifyDuringGC) && 4355 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 4356 verify_after_remark(); 4357 } 4358 4359 _gc_tracer_cm->report_object_count_after_gc(&_is_alive_closure); 4360 4361 // Change under the freelistLocks. 4362 _collectorState = Sweeping; 4363 // Call isAllClear() under bitMapLock 4364 assert(_modUnionTable.isAllClear(), 4365 "Should be clear by end of the final marking"); 4366 assert(_ct->klass_rem_set()->mod_union_is_clear(), 4367 "Should be clear by end of the final marking"); 4368 } 4369 4370 void CMSParInitialMarkTask::work(uint worker_id) { 4371 elapsedTimer _timer; 4372 ResourceMark rm; 4373 HandleMark hm; 4374 4375 // ---------- scan from roots -------------- 4376 _timer.start(); 4377 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4378 Par_MarkRefsIntoClosure par_mri_cl(_collector->_span, &(_collector->_markBitMap)); 4379 4380 // ---------- young gen roots -------------- 4381 { 4382 work_on_young_gen_roots(worker_id, &par_mri_cl); 4383 _timer.stop(); 4384 if (PrintCMSStatistics != 0) { 4385 gclog_or_tty->print_cr( 4386 "Finished young gen initial mark scan work in %dth thread: %3.3f sec", 4387 worker_id, _timer.seconds()); 4388 } 4389 } 4390 4391 // ---------- remaining roots -------------- 4392 _timer.reset(); 4393 _timer.start(); 4394 4395 CLDToOopClosure cld_closure(&par_mri_cl, true); 4396 4397 gch->gen_process_roots(_strong_roots_scope, 4398 GenCollectedHeap::OldGen, 4399 false, // yg was scanned above 4400 GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()), 4401 _collector->should_unload_classes(), 4402 &par_mri_cl, 4403 NULL, 4404 &cld_closure); 4405 assert(_collector->should_unload_classes() 4406 || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache), 4407 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 4408 _timer.stop(); 4409 if (PrintCMSStatistics != 0) { 4410 gclog_or_tty->print_cr( 4411 "Finished remaining root initial mark scan work in %dth thread: %3.3f sec", 4412 worker_id, _timer.seconds()); 4413 } 4414 } 4415 4416 // Parallel remark task 4417 class CMSParRemarkTask: public CMSParMarkTask { 4418 CompactibleFreeListSpace* _cms_space; 4419 4420 // The per-thread work queues, available here for stealing. 4421 OopTaskQueueSet* _task_queues; 4422 ParallelTaskTerminator _term; 4423 StrongRootsScope* _strong_roots_scope; 4424 4425 public: 4426 // A value of 0 passed to n_workers will cause the number of 4427 // workers to be taken from the active workers in the work gang. 4428 CMSParRemarkTask(CMSCollector* collector, 4429 CompactibleFreeListSpace* cms_space, 4430 uint n_workers, WorkGang* workers, 4431 OopTaskQueueSet* task_queues, 4432 StrongRootsScope* strong_roots_scope): 4433 CMSParMarkTask("Rescan roots and grey objects in parallel", 4434 collector, n_workers), 4435 _cms_space(cms_space), 4436 _task_queues(task_queues), 4437 _term(n_workers, task_queues), 4438 _strong_roots_scope(strong_roots_scope) { } 4439 4440 OopTaskQueueSet* task_queues() { return _task_queues; } 4441 4442 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 4443 4444 ParallelTaskTerminator* terminator() { return &_term; } 4445 uint n_workers() { return _n_workers; } 4446 4447 void work(uint worker_id); 4448 4449 private: 4450 // ... of dirty cards in old space 4451 void do_dirty_card_rescan_tasks(CompactibleFreeListSpace* sp, int i, 4452 Par_MarkRefsIntoAndScanClosure* cl); 4453 4454 // ... work stealing for the above 4455 void do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, int* seed); 4456 }; 4457 4458 class RemarkKlassClosure : public KlassClosure { 4459 KlassToOopClosure _cm_klass_closure; 4460 public: 4461 RemarkKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {} 4462 void do_klass(Klass* k) { 4463 // Check if we have modified any oops in the Klass during the concurrent marking. 4464 if (k->has_accumulated_modified_oops()) { 4465 k->clear_accumulated_modified_oops(); 4466 4467 // We could have transfered the current modified marks to the accumulated marks, 4468 // like we do with the Card Table to Mod Union Table. But it's not really necessary. 4469 } else if (k->has_modified_oops()) { 4470 // Don't clear anything, this info is needed by the next young collection. 4471 } else { 4472 // No modified oops in the Klass. 4473 return; 4474 } 4475 4476 // The klass has modified fields, need to scan the klass. 4477 _cm_klass_closure.do_klass(k); 4478 } 4479 }; 4480 4481 void CMSParMarkTask::work_on_young_gen_roots(uint worker_id, OopsInGenClosure* cl) { 4482 ParNewGeneration* young_gen = _collector->_young_gen; 4483 ContiguousSpace* eden_space = young_gen->eden(); 4484 ContiguousSpace* from_space = young_gen->from(); 4485 ContiguousSpace* to_space = young_gen->to(); 4486 4487 HeapWord** eca = _collector->_eden_chunk_array; 4488 size_t ect = _collector->_eden_chunk_index; 4489 HeapWord** sca = _collector->_survivor_chunk_array; 4490 size_t sct = _collector->_survivor_chunk_index; 4491 4492 assert(ect <= _collector->_eden_chunk_capacity, "out of bounds"); 4493 assert(sct <= _collector->_survivor_chunk_capacity, "out of bounds"); 4494 4495 do_young_space_rescan(worker_id, cl, to_space, NULL, 0); 4496 do_young_space_rescan(worker_id, cl, from_space, sca, sct); 4497 do_young_space_rescan(worker_id, cl, eden_space, eca, ect); 4498 } 4499 4500 // work_queue(i) is passed to the closure 4501 // Par_MarkRefsIntoAndScanClosure. The "i" parameter 4502 // also is passed to do_dirty_card_rescan_tasks() and to 4503 // do_work_steal() to select the i-th task_queue. 4504 4505 void CMSParRemarkTask::work(uint worker_id) { 4506 elapsedTimer _timer; 4507 ResourceMark rm; 4508 HandleMark hm; 4509 4510 // ---------- rescan from roots -------------- 4511 _timer.start(); 4512 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4513 Par_MarkRefsIntoAndScanClosure par_mrias_cl(_collector, 4514 _collector->_span, _collector->ref_processor(), 4515 &(_collector->_markBitMap), 4516 work_queue(worker_id)); 4517 4518 // Rescan young gen roots first since these are likely 4519 // coarsely partitioned and may, on that account, constitute 4520 // the critical path; thus, it's best to start off that 4521 // work first. 4522 // ---------- young gen roots -------------- 4523 { 4524 work_on_young_gen_roots(worker_id, &par_mrias_cl); 4525 _timer.stop(); 4526 if (PrintCMSStatistics != 0) { 4527 gclog_or_tty->print_cr( 4528 "Finished young gen rescan work in %dth thread: %3.3f sec", 4529 worker_id, _timer.seconds()); 4530 } 4531 } 4532 4533 // ---------- remaining roots -------------- 4534 _timer.reset(); 4535 _timer.start(); 4536 gch->gen_process_roots(_strong_roots_scope, 4537 GenCollectedHeap::OldGen, 4538 false, // yg was scanned above 4539 GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()), 4540 _collector->should_unload_classes(), 4541 &par_mrias_cl, 4542 NULL, 4543 NULL); // The dirty klasses will be handled below 4544 4545 assert(_collector->should_unload_classes() 4546 || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache), 4547 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 4548 _timer.stop(); 4549 if (PrintCMSStatistics != 0) { 4550 gclog_or_tty->print_cr( 4551 "Finished remaining root rescan work in %dth thread: %3.3f sec", 4552 worker_id, _timer.seconds()); 4553 } 4554 4555 // ---------- unhandled CLD scanning ---------- 4556 if (worker_id == 0) { // Single threaded at the moment. 4557 _timer.reset(); 4558 _timer.start(); 4559 4560 // Scan all new class loader data objects and new dependencies that were 4561 // introduced during concurrent marking. 4562 ResourceMark rm; 4563 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds(); 4564 for (int i = 0; i < array->length(); i++) { 4565 par_mrias_cl.do_cld_nv(array->at(i)); 4566 } 4567 4568 // We don't need to keep track of new CLDs anymore. 4569 ClassLoaderDataGraph::remember_new_clds(false); 4570 4571 _timer.stop(); 4572 if (PrintCMSStatistics != 0) { 4573 gclog_or_tty->print_cr( 4574 "Finished unhandled CLD scanning work in %dth thread: %3.3f sec", 4575 worker_id, _timer.seconds()); 4576 } 4577 } 4578 4579 // ---------- dirty klass scanning ---------- 4580 if (worker_id == 0) { // Single threaded at the moment. 4581 _timer.reset(); 4582 _timer.start(); 4583 4584 // Scan all classes that was dirtied during the concurrent marking phase. 4585 RemarkKlassClosure remark_klass_closure(&par_mrias_cl); 4586 ClassLoaderDataGraph::classes_do(&remark_klass_closure); 4587 4588 _timer.stop(); 4589 if (PrintCMSStatistics != 0) { 4590 gclog_or_tty->print_cr( 4591 "Finished dirty klass scanning work in %dth thread: %3.3f sec", 4592 worker_id, _timer.seconds()); 4593 } 4594 } 4595 4596 // We might have added oops to ClassLoaderData::_handles during the 4597 // concurrent marking phase. These oops point to newly allocated objects 4598 // that are guaranteed to be kept alive. Either by the direct allocation 4599 // code, or when the young collector processes the roots. Hence, 4600 // we don't have to revisit the _handles block during the remark phase. 4601 4602 // ---------- rescan dirty cards ------------ 4603 _timer.reset(); 4604 _timer.start(); 4605 4606 // Do the rescan tasks for each of the two spaces 4607 // (cms_space) in turn. 4608 // "worker_id" is passed to select the task_queue for "worker_id" 4609 do_dirty_card_rescan_tasks(_cms_space, worker_id, &par_mrias_cl); 4610 _timer.stop(); 4611 if (PrintCMSStatistics != 0) { 4612 gclog_or_tty->print_cr( 4613 "Finished dirty card rescan work in %dth thread: %3.3f sec", 4614 worker_id, _timer.seconds()); 4615 } 4616 4617 // ---------- steal work from other threads ... 4618 // ---------- ... and drain overflow list. 4619 _timer.reset(); 4620 _timer.start(); 4621 do_work_steal(worker_id, &par_mrias_cl, _collector->hash_seed(worker_id)); 4622 _timer.stop(); 4623 if (PrintCMSStatistics != 0) { 4624 gclog_or_tty->print_cr( 4625 "Finished work stealing in %dth thread: %3.3f sec", 4626 worker_id, _timer.seconds()); 4627 } 4628 } 4629 4630 // Note that parameter "i" is not used. 4631 void 4632 CMSParMarkTask::do_young_space_rescan(uint worker_id, 4633 OopsInGenClosure* cl, ContiguousSpace* space, 4634 HeapWord** chunk_array, size_t chunk_top) { 4635 // Until all tasks completed: 4636 // . claim an unclaimed task 4637 // . compute region boundaries corresponding to task claimed 4638 // using chunk_array 4639 // . par_oop_iterate(cl) over that region 4640 4641 ResourceMark rm; 4642 HandleMark hm; 4643 4644 SequentialSubTasksDone* pst = space->par_seq_tasks(); 4645 4646 uint nth_task = 0; 4647 uint n_tasks = pst->n_tasks(); 4648 4649 if (n_tasks > 0) { 4650 assert(pst->valid(), "Uninitialized use?"); 4651 HeapWord *start, *end; 4652 while (!pst->is_task_claimed(/* reference */ nth_task)) { 4653 // We claimed task # nth_task; compute its boundaries. 4654 if (chunk_top == 0) { // no samples were taken 4655 assert(nth_task == 0 && n_tasks == 1, "Can have only 1 eden task"); 4656 start = space->bottom(); 4657 end = space->top(); 4658 } else if (nth_task == 0) { 4659 start = space->bottom(); 4660 end = chunk_array[nth_task]; 4661 } else if (nth_task < (uint)chunk_top) { 4662 assert(nth_task >= 1, "Control point invariant"); 4663 start = chunk_array[nth_task - 1]; 4664 end = chunk_array[nth_task]; 4665 } else { 4666 assert(nth_task == (uint)chunk_top, "Control point invariant"); 4667 start = chunk_array[chunk_top - 1]; 4668 end = space->top(); 4669 } 4670 MemRegion mr(start, end); 4671 // Verify that mr is in space 4672 assert(mr.is_empty() || space->used_region().contains(mr), 4673 "Should be in space"); 4674 // Verify that "start" is an object boundary 4675 assert(mr.is_empty() || oop(mr.start())->is_oop(), 4676 "Should be an oop"); 4677 space->par_oop_iterate(mr, cl); 4678 } 4679 pst->all_tasks_completed(); 4680 } 4681 } 4682 4683 void 4684 CMSParRemarkTask::do_dirty_card_rescan_tasks( 4685 CompactibleFreeListSpace* sp, int i, 4686 Par_MarkRefsIntoAndScanClosure* cl) { 4687 // Until all tasks completed: 4688 // . claim an unclaimed task 4689 // . compute region boundaries corresponding to task claimed 4690 // . transfer dirty bits ct->mut for that region 4691 // . apply rescanclosure to dirty mut bits for that region 4692 4693 ResourceMark rm; 4694 HandleMark hm; 4695 4696 OopTaskQueue* work_q = work_queue(i); 4697 ModUnionClosure modUnionClosure(&(_collector->_modUnionTable)); 4698 // CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! 4699 // CAUTION: This closure has state that persists across calls to 4700 // the work method dirty_range_iterate_clear() in that it has 4701 // embedded in it a (subtype of) UpwardsObjectClosure. The 4702 // use of that state in the embedded UpwardsObjectClosure instance 4703 // assumes that the cards are always iterated (even if in parallel 4704 // by several threads) in monotonically increasing order per each 4705 // thread. This is true of the implementation below which picks 4706 // card ranges (chunks) in monotonically increasing order globally 4707 // and, a-fortiori, in monotonically increasing order per thread 4708 // (the latter order being a subsequence of the former). 4709 // If the work code below is ever reorganized into a more chaotic 4710 // work-partitioning form than the current "sequential tasks" 4711 // paradigm, the use of that persistent state will have to be 4712 // revisited and modified appropriately. See also related 4713 // bug 4756801 work on which should examine this code to make 4714 // sure that the changes there do not run counter to the 4715 // assumptions made here and necessary for correctness and 4716 // efficiency. Note also that this code might yield inefficient 4717 // behavior in the case of very large objects that span one or 4718 // more work chunks. Such objects would potentially be scanned 4719 // several times redundantly. Work on 4756801 should try and 4720 // address that performance anomaly if at all possible. XXX 4721 MemRegion full_span = _collector->_span; 4722 CMSBitMap* bm = &(_collector->_markBitMap); // shared 4723 MarkFromDirtyCardsClosure 4724 greyRescanClosure(_collector, full_span, // entire span of interest 4725 sp, bm, work_q, cl); 4726 4727 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 4728 assert(pst->valid(), "Uninitialized use?"); 4729 uint nth_task = 0; 4730 const int alignment = CardTableModRefBS::card_size * BitsPerWord; 4731 MemRegion span = sp->used_region(); 4732 HeapWord* start_addr = span.start(); 4733 HeapWord* end_addr = (HeapWord*)round_to((intptr_t)span.end(), 4734 alignment); 4735 const size_t chunk_size = sp->rescan_task_size(); // in HeapWord units 4736 assert((HeapWord*)round_to((intptr_t)start_addr, alignment) == 4737 start_addr, "Check alignment"); 4738 assert((size_t)round_to((intptr_t)chunk_size, alignment) == 4739 chunk_size, "Check alignment"); 4740 4741 while (!pst->is_task_claimed(/* reference */ nth_task)) { 4742 // Having claimed the nth_task, compute corresponding mem-region, 4743 // which is a-fortiori aligned correctly (i.e. at a MUT boundary). 4744 // The alignment restriction ensures that we do not need any 4745 // synchronization with other gang-workers while setting or 4746 // clearing bits in thus chunk of the MUT. 4747 MemRegion this_span = MemRegion(start_addr + nth_task*chunk_size, 4748 start_addr + (nth_task+1)*chunk_size); 4749 // The last chunk's end might be way beyond end of the 4750 // used region. In that case pull back appropriately. 4751 if (this_span.end() > end_addr) { 4752 this_span.set_end(end_addr); 4753 assert(!this_span.is_empty(), "Program logic (calculation of n_tasks)"); 4754 } 4755 // Iterate over the dirty cards covering this chunk, marking them 4756 // precleaned, and setting the corresponding bits in the mod union 4757 // table. Since we have been careful to partition at Card and MUT-word 4758 // boundaries no synchronization is needed between parallel threads. 4759 _collector->_ct->ct_bs()->dirty_card_iterate(this_span, 4760 &modUnionClosure); 4761 4762 // Having transferred these marks into the modUnionTable, 4763 // rescan the marked objects on the dirty cards in the modUnionTable. 4764 // Even if this is at a synchronous collection, the initial marking 4765 // may have been done during an asynchronous collection so there 4766 // may be dirty bits in the mod-union table. 4767 _collector->_modUnionTable.dirty_range_iterate_clear( 4768 this_span, &greyRescanClosure); 4769 _collector->_modUnionTable.verifyNoOneBitsInRange( 4770 this_span.start(), 4771 this_span.end()); 4772 } 4773 pst->all_tasks_completed(); // declare that i am done 4774 } 4775 4776 // . see if we can share work_queues with ParNew? XXX 4777 void 4778 CMSParRemarkTask::do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, 4779 int* seed) { 4780 OopTaskQueue* work_q = work_queue(i); 4781 NOT_PRODUCT(int num_steals = 0;) 4782 oop obj_to_scan; 4783 CMSBitMap* bm = &(_collector->_markBitMap); 4784 4785 while (true) { 4786 // Completely finish any left over work from (an) earlier round(s) 4787 cl->trim_queue(0); 4788 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 4789 (size_t)ParGCDesiredObjsFromOverflowList); 4790 // Now check if there's any work in the overflow list 4791 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads, 4792 // only affects the number of attempts made to get work from the 4793 // overflow list and does not affect the number of workers. Just 4794 // pass ParallelGCThreads so this behavior is unchanged. 4795 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 4796 work_q, 4797 ParallelGCThreads)) { 4798 // found something in global overflow list; 4799 // not yet ready to go stealing work from others. 4800 // We'd like to assert(work_q->size() != 0, ...) 4801 // because we just took work from the overflow list, 4802 // but of course we can't since all of that could have 4803 // been already stolen from us. 4804 // "He giveth and He taketh away." 4805 continue; 4806 } 4807 // Verify that we have no work before we resort to stealing 4808 assert(work_q->size() == 0, "Have work, shouldn't steal"); 4809 // Try to steal from other queues that have work 4810 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 4811 NOT_PRODUCT(num_steals++;) 4812 assert(obj_to_scan->is_oop(), "Oops, not an oop!"); 4813 assert(bm->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 4814 // Do scanning work 4815 obj_to_scan->oop_iterate(cl); 4816 // Loop around, finish this work, and try to steal some more 4817 } else if (terminator()->offer_termination()) { 4818 break; // nirvana from the infinite cycle 4819 } 4820 } 4821 NOT_PRODUCT( 4822 if (PrintCMSStatistics != 0) { 4823 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals); 4824 } 4825 ) 4826 assert(work_q->size() == 0 && _collector->overflow_list_is_empty(), 4827 "Else our work is not yet done"); 4828 } 4829 4830 // Record object boundaries in _eden_chunk_array by sampling the eden 4831 // top in the slow-path eden object allocation code path and record 4832 // the boundaries, if CMSEdenChunksRecordAlways is true. If 4833 // CMSEdenChunksRecordAlways is false, we use the other asynchronous 4834 // sampling in sample_eden() that activates during the part of the 4835 // preclean phase. 4836 void CMSCollector::sample_eden_chunk() { 4837 if (CMSEdenChunksRecordAlways && _eden_chunk_array != NULL) { 4838 if (_eden_chunk_lock->try_lock()) { 4839 // Record a sample. This is the critical section. The contents 4840 // of the _eden_chunk_array have to be non-decreasing in the 4841 // address order. 4842 _eden_chunk_array[_eden_chunk_index] = *_top_addr; 4843 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr, 4844 "Unexpected state of Eden"); 4845 if (_eden_chunk_index == 0 || 4846 ((_eden_chunk_array[_eden_chunk_index] > _eden_chunk_array[_eden_chunk_index-1]) && 4847 (pointer_delta(_eden_chunk_array[_eden_chunk_index], 4848 _eden_chunk_array[_eden_chunk_index-1]) >= CMSSamplingGrain))) { 4849 _eden_chunk_index++; // commit sample 4850 } 4851 _eden_chunk_lock->unlock(); 4852 } 4853 } 4854 } 4855 4856 // Return a thread-local PLAB recording array, as appropriate. 4857 void* CMSCollector::get_data_recorder(int thr_num) { 4858 if (_survivor_plab_array != NULL && 4859 (CMSPLABRecordAlways || 4860 (_collectorState > Marking && _collectorState < FinalMarking))) { 4861 assert(thr_num < (int)ParallelGCThreads, "thr_num is out of bounds"); 4862 ChunkArray* ca = &_survivor_plab_array[thr_num]; 4863 ca->reset(); // clear it so that fresh data is recorded 4864 return (void*) ca; 4865 } else { 4866 return NULL; 4867 } 4868 } 4869 4870 // Reset all the thread-local PLAB recording arrays 4871 void CMSCollector::reset_survivor_plab_arrays() { 4872 for (uint i = 0; i < ParallelGCThreads; i++) { 4873 _survivor_plab_array[i].reset(); 4874 } 4875 } 4876 4877 // Merge the per-thread plab arrays into the global survivor chunk 4878 // array which will provide the partitioning of the survivor space 4879 // for CMS initial scan and rescan. 4880 void CMSCollector::merge_survivor_plab_arrays(ContiguousSpace* surv, 4881 int no_of_gc_threads) { 4882 assert(_survivor_plab_array != NULL, "Error"); 4883 assert(_survivor_chunk_array != NULL, "Error"); 4884 assert(_collectorState == FinalMarking || 4885 (CMSParallelInitialMarkEnabled && _collectorState == InitialMarking), "Error"); 4886 for (int j = 0; j < no_of_gc_threads; j++) { 4887 _cursor[j] = 0; 4888 } 4889 HeapWord* top = surv->top(); 4890 size_t i; 4891 for (i = 0; i < _survivor_chunk_capacity; i++) { // all sca entries 4892 HeapWord* min_val = top; // Higher than any PLAB address 4893 uint min_tid = 0; // position of min_val this round 4894 for (int j = 0; j < no_of_gc_threads; j++) { 4895 ChunkArray* cur_sca = &_survivor_plab_array[j]; 4896 if (_cursor[j] == cur_sca->end()) { 4897 continue; 4898 } 4899 assert(_cursor[j] < cur_sca->end(), "ctl pt invariant"); 4900 HeapWord* cur_val = cur_sca->nth(_cursor[j]); 4901 assert(surv->used_region().contains(cur_val), "Out of bounds value"); 4902 if (cur_val < min_val) { 4903 min_tid = j; 4904 min_val = cur_val; 4905 } else { 4906 assert(cur_val < top, "All recorded addresses should be less"); 4907 } 4908 } 4909 // At this point min_val and min_tid are respectively 4910 // the least address in _survivor_plab_array[j]->nth(_cursor[j]) 4911 // and the thread (j) that witnesses that address. 4912 // We record this address in the _survivor_chunk_array[i] 4913 // and increment _cursor[min_tid] prior to the next round i. 4914 if (min_val == top) { 4915 break; 4916 } 4917 _survivor_chunk_array[i] = min_val; 4918 _cursor[min_tid]++; 4919 } 4920 // We are all done; record the size of the _survivor_chunk_array 4921 _survivor_chunk_index = i; // exclusive: [0, i) 4922 if (PrintCMSStatistics > 0) { 4923 gclog_or_tty->print(" (Survivor:" SIZE_FORMAT "chunks) ", i); 4924 } 4925 // Verify that we used up all the recorded entries 4926 #ifdef ASSERT 4927 size_t total = 0; 4928 for (int j = 0; j < no_of_gc_threads; j++) { 4929 assert(_cursor[j] == _survivor_plab_array[j].end(), "Ctl pt invariant"); 4930 total += _cursor[j]; 4931 } 4932 assert(total == _survivor_chunk_index, "Ctl Pt Invariant"); 4933 // Check that the merged array is in sorted order 4934 if (total > 0) { 4935 for (size_t i = 0; i < total - 1; i++) { 4936 if (PrintCMSStatistics > 0) { 4937 gclog_or_tty->print(" (chunk" SIZE_FORMAT ":" INTPTR_FORMAT ") ", 4938 i, p2i(_survivor_chunk_array[i])); 4939 } 4940 assert(_survivor_chunk_array[i] < _survivor_chunk_array[i+1], 4941 "Not sorted"); 4942 } 4943 } 4944 #endif // ASSERT 4945 } 4946 4947 // Set up the space's par_seq_tasks structure for work claiming 4948 // for parallel initial scan and rescan of young gen. 4949 // See ParRescanTask where this is currently used. 4950 void 4951 CMSCollector:: 4952 initialize_sequential_subtasks_for_young_gen_rescan(int n_threads) { 4953 assert(n_threads > 0, "Unexpected n_threads argument"); 4954 4955 // Eden space 4956 if (!_young_gen->eden()->is_empty()) { 4957 SequentialSubTasksDone* pst = _young_gen->eden()->par_seq_tasks(); 4958 assert(!pst->valid(), "Clobbering existing data?"); 4959 // Each valid entry in [0, _eden_chunk_index) represents a task. 4960 size_t n_tasks = _eden_chunk_index + 1; 4961 assert(n_tasks == 1 || _eden_chunk_array != NULL, "Error"); 4962 // Sets the condition for completion of the subtask (how many threads 4963 // need to finish in order to be done). 4964 pst->set_n_threads(n_threads); 4965 pst->set_n_tasks((int)n_tasks); 4966 } 4967 4968 // Merge the survivor plab arrays into _survivor_chunk_array 4969 if (_survivor_plab_array != NULL) { 4970 merge_survivor_plab_arrays(_young_gen->from(), n_threads); 4971 } else { 4972 assert(_survivor_chunk_index == 0, "Error"); 4973 } 4974 4975 // To space 4976 { 4977 SequentialSubTasksDone* pst = _young_gen->to()->par_seq_tasks(); 4978 assert(!pst->valid(), "Clobbering existing data?"); 4979 // Sets the condition for completion of the subtask (how many threads 4980 // need to finish in order to be done). 4981 pst->set_n_threads(n_threads); 4982 pst->set_n_tasks(1); 4983 assert(pst->valid(), "Error"); 4984 } 4985 4986 // From space 4987 { 4988 SequentialSubTasksDone* pst = _young_gen->from()->par_seq_tasks(); 4989 assert(!pst->valid(), "Clobbering existing data?"); 4990 size_t n_tasks = _survivor_chunk_index + 1; 4991 assert(n_tasks == 1 || _survivor_chunk_array != NULL, "Error"); 4992 // Sets the condition for completion of the subtask (how many threads 4993 // need to finish in order to be done). 4994 pst->set_n_threads(n_threads); 4995 pst->set_n_tasks((int)n_tasks); 4996 assert(pst->valid(), "Error"); 4997 } 4998 } 4999 5000 // Parallel version of remark 5001 void CMSCollector::do_remark_parallel() { 5002 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5003 WorkGang* workers = gch->workers(); 5004 assert(workers != NULL, "Need parallel worker threads."); 5005 // Choose to use the number of GC workers most recently set 5006 // into "active_workers". 5007 uint n_workers = workers->active_workers(); 5008 5009 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 5010 5011 StrongRootsScope srs(n_workers); 5012 5013 CMSParRemarkTask tsk(this, cms_space, n_workers, workers, task_queues(), &srs); 5014 5015 // We won't be iterating over the cards in the card table updating 5016 // the younger_gen cards, so we shouldn't call the following else 5017 // the verification code as well as subsequent younger_refs_iterate 5018 // code would get confused. XXX 5019 // gch->rem_set()->prepare_for_younger_refs_iterate(true); // parallel 5020 5021 // The young gen rescan work will not be done as part of 5022 // process_roots (which currently doesn't know how to 5023 // parallelize such a scan), but rather will be broken up into 5024 // a set of parallel tasks (via the sampling that the [abortable] 5025 // preclean phase did of eden, plus the [two] tasks of 5026 // scanning the [two] survivor spaces. Further fine-grain 5027 // parallelization of the scanning of the survivor spaces 5028 // themselves, and of precleaning of the young gen itself 5029 // is deferred to the future. 5030 initialize_sequential_subtasks_for_young_gen_rescan(n_workers); 5031 5032 // The dirty card rescan work is broken up into a "sequence" 5033 // of parallel tasks (per constituent space) that are dynamically 5034 // claimed by the parallel threads. 5035 cms_space->initialize_sequential_subtasks_for_rescan(n_workers); 5036 5037 // It turns out that even when we're using 1 thread, doing the work in a 5038 // separate thread causes wide variance in run times. We can't help this 5039 // in the multi-threaded case, but we special-case n=1 here to get 5040 // repeatable measurements of the 1-thread overhead of the parallel code. 5041 if (n_workers > 1) { 5042 // Make refs discovery MT-safe, if it isn't already: it may not 5043 // necessarily be so, since it's possible that we are doing 5044 // ST marking. 5045 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), true); 5046 workers->run_task(&tsk); 5047 } else { 5048 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false); 5049 tsk.work(0); 5050 } 5051 5052 // restore, single-threaded for now, any preserved marks 5053 // as a result of work_q overflow 5054 restore_preserved_marks_if_any(); 5055 } 5056 5057 // Non-parallel version of remark 5058 void CMSCollector::do_remark_non_parallel() { 5059 ResourceMark rm; 5060 HandleMark hm; 5061 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5062 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false); 5063 5064 MarkRefsIntoAndScanClosure 5065 mrias_cl(_span, ref_processor(), &_markBitMap, NULL /* not precleaning */, 5066 &_markStack, this, 5067 false /* should_yield */, false /* not precleaning */); 5068 MarkFromDirtyCardsClosure 5069 markFromDirtyCardsClosure(this, _span, 5070 NULL, // space is set further below 5071 &_markBitMap, &_markStack, &mrias_cl); 5072 { 5073 GCTraceTime t("grey object rescan", PrintGCDetails, false, _gc_timer_cm); 5074 // Iterate over the dirty cards, setting the corresponding bits in the 5075 // mod union table. 5076 { 5077 ModUnionClosure modUnionClosure(&_modUnionTable); 5078 _ct->ct_bs()->dirty_card_iterate( 5079 _cmsGen->used_region(), 5080 &modUnionClosure); 5081 } 5082 // Having transferred these marks into the modUnionTable, we just need 5083 // to rescan the marked objects on the dirty cards in the modUnionTable. 5084 // The initial marking may have been done during an asynchronous 5085 // collection so there may be dirty bits in the mod-union table. 5086 const int alignment = 5087 CardTableModRefBS::card_size * BitsPerWord; 5088 { 5089 // ... First handle dirty cards in CMS gen 5090 markFromDirtyCardsClosure.set_space(_cmsGen->cmsSpace()); 5091 MemRegion ur = _cmsGen->used_region(); 5092 HeapWord* lb = ur.start(); 5093 HeapWord* ub = (HeapWord*)round_to((intptr_t)ur.end(), alignment); 5094 MemRegion cms_span(lb, ub); 5095 _modUnionTable.dirty_range_iterate_clear(cms_span, 5096 &markFromDirtyCardsClosure); 5097 verify_work_stacks_empty(); 5098 if (PrintCMSStatistics != 0) { 5099 gclog_or_tty->print(" (re-scanned " SIZE_FORMAT " dirty cards in cms gen) ", 5100 markFromDirtyCardsClosure.num_dirty_cards()); 5101 } 5102 } 5103 } 5104 if (VerifyDuringGC && 5105 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 5106 HandleMark hm; // Discard invalid handles created during verification 5107 Universe::verify(); 5108 } 5109 { 5110 GCTraceTime t("root rescan", PrintGCDetails, false, _gc_timer_cm); 5111 5112 verify_work_stacks_empty(); 5113 5114 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 5115 StrongRootsScope srs(1); 5116 5117 gch->gen_process_roots(&srs, 5118 GenCollectedHeap::OldGen, 5119 true, // young gen as roots 5120 GenCollectedHeap::ScanningOption(roots_scanning_options()), 5121 should_unload_classes(), 5122 &mrias_cl, 5123 NULL, 5124 NULL); // The dirty klasses will be handled below 5125 5126 assert(should_unload_classes() 5127 || (roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache), 5128 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 5129 } 5130 5131 { 5132 GCTraceTime t("visit unhandled CLDs", PrintGCDetails, false, _gc_timer_cm); 5133 5134 verify_work_stacks_empty(); 5135 5136 // Scan all class loader data objects that might have been introduced 5137 // during concurrent marking. 5138 ResourceMark rm; 5139 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds(); 5140 for (int i = 0; i < array->length(); i++) { 5141 mrias_cl.do_cld_nv(array->at(i)); 5142 } 5143 5144 // We don't need to keep track of new CLDs anymore. 5145 ClassLoaderDataGraph::remember_new_clds(false); 5146 5147 verify_work_stacks_empty(); 5148 } 5149 5150 { 5151 GCTraceTime t("dirty klass scan", PrintGCDetails, false, _gc_timer_cm); 5152 5153 verify_work_stacks_empty(); 5154 5155 RemarkKlassClosure remark_klass_closure(&mrias_cl); 5156 ClassLoaderDataGraph::classes_do(&remark_klass_closure); 5157 5158 verify_work_stacks_empty(); 5159 } 5160 5161 // We might have added oops to ClassLoaderData::_handles during the 5162 // concurrent marking phase. These oops point to newly allocated objects 5163 // that are guaranteed to be kept alive. Either by the direct allocation 5164 // code, or when the young collector processes the roots. Hence, 5165 // we don't have to revisit the _handles block during the remark phase. 5166 5167 verify_work_stacks_empty(); 5168 // Restore evacuated mark words, if any, used for overflow list links 5169 restore_preserved_marks_if_any(); 5170 5171 verify_overflow_empty(); 5172 } 5173 5174 //////////////////////////////////////////////////////// 5175 // Parallel Reference Processing Task Proxy Class 5176 //////////////////////////////////////////////////////// 5177 class AbstractGangTaskWOopQueues : public AbstractGangTask { 5178 OopTaskQueueSet* _queues; 5179 ParallelTaskTerminator _terminator; 5180 public: 5181 AbstractGangTaskWOopQueues(const char* name, OopTaskQueueSet* queues, uint n_threads) : 5182 AbstractGangTask(name), _queues(queues), _terminator(n_threads, _queues) {} 5183 ParallelTaskTerminator* terminator() { return &_terminator; } 5184 OopTaskQueueSet* queues() { return _queues; } 5185 }; 5186 5187 class CMSRefProcTaskProxy: public AbstractGangTaskWOopQueues { 5188 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask; 5189 CMSCollector* _collector; 5190 CMSBitMap* _mark_bit_map; 5191 const MemRegion _span; 5192 ProcessTask& _task; 5193 5194 public: 5195 CMSRefProcTaskProxy(ProcessTask& task, 5196 CMSCollector* collector, 5197 const MemRegion& span, 5198 CMSBitMap* mark_bit_map, 5199 AbstractWorkGang* workers, 5200 OopTaskQueueSet* task_queues): 5201 AbstractGangTaskWOopQueues("Process referents by policy in parallel", 5202 task_queues, 5203 workers->active_workers()), 5204 _task(task), 5205 _collector(collector), _span(span), _mark_bit_map(mark_bit_map) 5206 { 5207 assert(_collector->_span.equals(_span) && !_span.is_empty(), 5208 "Inconsistency in _span"); 5209 } 5210 5211 OopTaskQueueSet* task_queues() { return queues(); } 5212 5213 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 5214 5215 void do_work_steal(int i, 5216 CMSParDrainMarkingStackClosure* drain, 5217 CMSParKeepAliveClosure* keep_alive, 5218 int* seed); 5219 5220 virtual void work(uint worker_id); 5221 }; 5222 5223 void CMSRefProcTaskProxy::work(uint worker_id) { 5224 ResourceMark rm; 5225 HandleMark hm; 5226 assert(_collector->_span.equals(_span), "Inconsistency in _span"); 5227 CMSParKeepAliveClosure par_keep_alive(_collector, _span, 5228 _mark_bit_map, 5229 work_queue(worker_id)); 5230 CMSParDrainMarkingStackClosure par_drain_stack(_collector, _span, 5231 _mark_bit_map, 5232 work_queue(worker_id)); 5233 CMSIsAliveClosure is_alive_closure(_span, _mark_bit_map); 5234 _task.work(worker_id, is_alive_closure, par_keep_alive, par_drain_stack); 5235 if (_task.marks_oops_alive()) { 5236 do_work_steal(worker_id, &par_drain_stack, &par_keep_alive, 5237 _collector->hash_seed(worker_id)); 5238 } 5239 assert(work_queue(worker_id)->size() == 0, "work_queue should be empty"); 5240 assert(_collector->_overflow_list == NULL, "non-empty _overflow_list"); 5241 } 5242 5243 class CMSRefEnqueueTaskProxy: public AbstractGangTask { 5244 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask; 5245 EnqueueTask& _task; 5246 5247 public: 5248 CMSRefEnqueueTaskProxy(EnqueueTask& task) 5249 : AbstractGangTask("Enqueue reference objects in parallel"), 5250 _task(task) 5251 { } 5252 5253 virtual void work(uint worker_id) 5254 { 5255 _task.work(worker_id); 5256 } 5257 }; 5258 5259 CMSParKeepAliveClosure::CMSParKeepAliveClosure(CMSCollector* collector, 5260 MemRegion span, CMSBitMap* bit_map, OopTaskQueue* work_queue): 5261 _span(span), 5262 _bit_map(bit_map), 5263 _work_queue(work_queue), 5264 _mark_and_push(collector, span, bit_map, work_queue), 5265 _low_water_mark(MIN2((work_queue->max_elems()/4), 5266 ((uint)CMSWorkQueueDrainThreshold * ParallelGCThreads))) 5267 { } 5268 5269 // . see if we can share work_queues with ParNew? XXX 5270 void CMSRefProcTaskProxy::do_work_steal(int i, 5271 CMSParDrainMarkingStackClosure* drain, 5272 CMSParKeepAliveClosure* keep_alive, 5273 int* seed) { 5274 OopTaskQueue* work_q = work_queue(i); 5275 NOT_PRODUCT(int num_steals = 0;) 5276 oop obj_to_scan; 5277 5278 while (true) { 5279 // Completely finish any left over work from (an) earlier round(s) 5280 drain->trim_queue(0); 5281 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 5282 (size_t)ParGCDesiredObjsFromOverflowList); 5283 // Now check if there's any work in the overflow list 5284 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads, 5285 // only affects the number of attempts made to get work from the 5286 // overflow list and does not affect the number of workers. Just 5287 // pass ParallelGCThreads so this behavior is unchanged. 5288 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 5289 work_q, 5290 ParallelGCThreads)) { 5291 // Found something in global overflow list; 5292 // not yet ready to go stealing work from others. 5293 // We'd like to assert(work_q->size() != 0, ...) 5294 // because we just took work from the overflow list, 5295 // but of course we can't, since all of that might have 5296 // been already stolen from us. 5297 continue; 5298 } 5299 // Verify that we have no work before we resort to stealing 5300 assert(work_q->size() == 0, "Have work, shouldn't steal"); 5301 // Try to steal from other queues that have work 5302 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 5303 NOT_PRODUCT(num_steals++;) 5304 assert(obj_to_scan->is_oop(), "Oops, not an oop!"); 5305 assert(_mark_bit_map->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 5306 // Do scanning work 5307 obj_to_scan->oop_iterate(keep_alive); 5308 // Loop around, finish this work, and try to steal some more 5309 } else if (terminator()->offer_termination()) { 5310 break; // nirvana from the infinite cycle 5311 } 5312 } 5313 NOT_PRODUCT( 5314 if (PrintCMSStatistics != 0) { 5315 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals); 5316 } 5317 ) 5318 } 5319 5320 void CMSRefProcTaskExecutor::execute(ProcessTask& task) 5321 { 5322 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5323 WorkGang* workers = gch->workers(); 5324 assert(workers != NULL, "Need parallel worker threads."); 5325 CMSRefProcTaskProxy rp_task(task, &_collector, 5326 _collector.ref_processor()->span(), 5327 _collector.markBitMap(), 5328 workers, _collector.task_queues()); 5329 workers->run_task(&rp_task); 5330 } 5331 5332 void CMSRefProcTaskExecutor::execute(EnqueueTask& task) 5333 { 5334 5335 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5336 WorkGang* workers = gch->workers(); 5337 assert(workers != NULL, "Need parallel worker threads."); 5338 CMSRefEnqueueTaskProxy enq_task(task); 5339 workers->run_task(&enq_task); 5340 } 5341 5342 void CMSCollector::refProcessingWork() { 5343 ResourceMark rm; 5344 HandleMark hm; 5345 5346 ReferenceProcessor* rp = ref_processor(); 5347 assert(rp->span().equals(_span), "Spans should be equal"); 5348 assert(!rp->enqueuing_is_done(), "Enqueuing should not be complete"); 5349 // Process weak references. 5350 rp->setup_policy(false); 5351 verify_work_stacks_empty(); 5352 5353 CMSKeepAliveClosure cmsKeepAliveClosure(this, _span, &_markBitMap, 5354 &_markStack, false /* !preclean */); 5355 CMSDrainMarkingStackClosure cmsDrainMarkingStackClosure(this, 5356 _span, &_markBitMap, &_markStack, 5357 &cmsKeepAliveClosure, false /* !preclean */); 5358 { 5359 GCTraceTime t("weak refs processing", PrintGCDetails, false, _gc_timer_cm); 5360 5361 ReferenceProcessorStats stats; 5362 if (rp->processing_is_mt()) { 5363 // Set the degree of MT here. If the discovery is done MT, there 5364 // may have been a different number of threads doing the discovery 5365 // and a different number of discovered lists may have Ref objects. 5366 // That is OK as long as the Reference lists are balanced (see 5367 // balance_all_queues() and balance_queues()). 5368 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5369 uint active_workers = ParallelGCThreads; 5370 WorkGang* workers = gch->workers(); 5371 if (workers != NULL) { 5372 active_workers = workers->active_workers(); 5373 // The expectation is that active_workers will have already 5374 // been set to a reasonable value. If it has not been set, 5375 // investigate. 5376 assert(active_workers > 0, "Should have been set during scavenge"); 5377 } 5378 rp->set_active_mt_degree(active_workers); 5379 CMSRefProcTaskExecutor task_executor(*this); 5380 stats = rp->process_discovered_references(&_is_alive_closure, 5381 &cmsKeepAliveClosure, 5382 &cmsDrainMarkingStackClosure, 5383 &task_executor, 5384 _gc_timer_cm); 5385 } else { 5386 stats = rp->process_discovered_references(&_is_alive_closure, 5387 &cmsKeepAliveClosure, 5388 &cmsDrainMarkingStackClosure, 5389 NULL, 5390 _gc_timer_cm); 5391 } 5392 _gc_tracer_cm->report_gc_reference_stats(stats); 5393 5394 } 5395 5396 // This is the point where the entire marking should have completed. 5397 verify_work_stacks_empty(); 5398 5399 if (should_unload_classes()) { 5400 { 5401 GCTraceTime t("class unloading", PrintGCDetails, false, _gc_timer_cm); 5402 5403 // Unload classes and purge the SystemDictionary. 5404 bool purged_class = SystemDictionary::do_unloading(&_is_alive_closure); 5405 5406 // Unload nmethods. 5407 CodeCache::do_unloading(&_is_alive_closure, purged_class); 5408 5409 // Prune dead klasses from subklass/sibling/implementor lists. 5410 Klass::clean_weak_klass_links(&_is_alive_closure); 5411 } 5412 5413 { 5414 GCTraceTime t("scrub symbol table", PrintGCDetails, false, _gc_timer_cm); 5415 // Clean up unreferenced symbols in symbol table. 5416 SymbolTable::unlink(); 5417 } 5418 5419 { 5420 GCTraceTime t("scrub string table", PrintGCDetails, false, _gc_timer_cm); 5421 // Delete entries for dead interned strings. 5422 StringTable::unlink(&_is_alive_closure); 5423 } 5424 } 5425 5426 5427 // Restore any preserved marks as a result of mark stack or 5428 // work queue overflow 5429 restore_preserved_marks_if_any(); // done single-threaded for now 5430 5431 rp->set_enqueuing_is_done(true); 5432 if (rp->processing_is_mt()) { 5433 rp->balance_all_queues(); 5434 CMSRefProcTaskExecutor task_executor(*this); 5435 rp->enqueue_discovered_references(&task_executor); 5436 } else { 5437 rp->enqueue_discovered_references(NULL); 5438 } 5439 rp->verify_no_references_recorded(); 5440 assert(!rp->discovery_enabled(), "should have been disabled"); 5441 } 5442 5443 #ifndef PRODUCT 5444 void CMSCollector::check_correct_thread_executing() { 5445 Thread* t = Thread::current(); 5446 // Only the VM thread or the CMS thread should be here. 5447 assert(t->is_ConcurrentGC_thread() || t->is_VM_thread(), 5448 "Unexpected thread type"); 5449 // If this is the vm thread, the foreground process 5450 // should not be waiting. Note that _foregroundGCIsActive is 5451 // true while the foreground collector is waiting. 5452 if (_foregroundGCShouldWait) { 5453 // We cannot be the VM thread 5454 assert(t->is_ConcurrentGC_thread(), 5455 "Should be CMS thread"); 5456 } else { 5457 // We can be the CMS thread only if we are in a stop-world 5458 // phase of CMS collection. 5459 if (t->is_ConcurrentGC_thread()) { 5460 assert(_collectorState == InitialMarking || 5461 _collectorState == FinalMarking, 5462 "Should be a stop-world phase"); 5463 // The CMS thread should be holding the CMS_token. 5464 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 5465 "Potential interference with concurrently " 5466 "executing VM thread"); 5467 } 5468 } 5469 } 5470 #endif 5471 5472 void CMSCollector::sweep() { 5473 assert(_collectorState == Sweeping, "just checking"); 5474 check_correct_thread_executing(); 5475 verify_work_stacks_empty(); 5476 verify_overflow_empty(); 5477 increment_sweep_count(); 5478 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 5479 5480 _inter_sweep_timer.stop(); 5481 _inter_sweep_estimate.sample(_inter_sweep_timer.seconds()); 5482 5483 assert(!_intra_sweep_timer.is_active(), "Should not be active"); 5484 _intra_sweep_timer.reset(); 5485 _intra_sweep_timer.start(); 5486 { 5487 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 5488 CMSPhaseAccounting pa(this, "sweep", !PrintGCDetails); 5489 // First sweep the old gen 5490 { 5491 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(), 5492 bitMapLock()); 5493 sweepWork(_cmsGen); 5494 } 5495 5496 // Update Universe::_heap_*_at_gc figures. 5497 // We need all the free list locks to make the abstract state 5498 // transition from Sweeping to Resetting. See detailed note 5499 // further below. 5500 { 5501 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock()); 5502 // Update heap occupancy information which is used as 5503 // input to soft ref clearing policy at the next gc. 5504 Universe::update_heap_info_at_gc(); 5505 _collectorState = Resizing; 5506 } 5507 } 5508 verify_work_stacks_empty(); 5509 verify_overflow_empty(); 5510 5511 if (should_unload_classes()) { 5512 // Delay purge to the beginning of the next safepoint. Metaspace::contains 5513 // requires that the virtual spaces are stable and not deleted. 5514 ClassLoaderDataGraph::set_should_purge(true); 5515 } 5516 5517 _intra_sweep_timer.stop(); 5518 _intra_sweep_estimate.sample(_intra_sweep_timer.seconds()); 5519 5520 _inter_sweep_timer.reset(); 5521 _inter_sweep_timer.start(); 5522 5523 // We need to use a monotonically non-decreasing time in ms 5524 // or we will see time-warp warnings and os::javaTimeMillis() 5525 // does not guarantee monotonicity. 5526 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC; 5527 update_time_of_last_gc(now); 5528 5529 // NOTE on abstract state transitions: 5530 // Mutators allocate-live and/or mark the mod-union table dirty 5531 // based on the state of the collection. The former is done in 5532 // the interval [Marking, Sweeping] and the latter in the interval 5533 // [Marking, Sweeping). Thus the transitions into the Marking state 5534 // and out of the Sweeping state must be synchronously visible 5535 // globally to the mutators. 5536 // The transition into the Marking state happens with the world 5537 // stopped so the mutators will globally see it. Sweeping is 5538 // done asynchronously by the background collector so the transition 5539 // from the Sweeping state to the Resizing state must be done 5540 // under the freelistLock (as is the check for whether to 5541 // allocate-live and whether to dirty the mod-union table). 5542 assert(_collectorState == Resizing, "Change of collector state to" 5543 " Resizing must be done under the freelistLocks (plural)"); 5544 5545 // Now that sweeping has been completed, we clear 5546 // the incremental_collection_failed flag, 5547 // thus inviting a younger gen collection to promote into 5548 // this generation. If such a promotion may still fail, 5549 // the flag will be set again when a young collection is 5550 // attempted. 5551 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5552 gch->clear_incremental_collection_failed(); // Worth retrying as fresh space may have been freed up 5553 gch->update_full_collections_completed(_collection_count_start); 5554 } 5555 5556 // FIX ME!!! Looks like this belongs in CFLSpace, with 5557 // CMSGen merely delegating to it. 5558 void ConcurrentMarkSweepGeneration::setNearLargestChunk() { 5559 double nearLargestPercent = FLSLargestBlockCoalesceProximity; 5560 HeapWord* minAddr = _cmsSpace->bottom(); 5561 HeapWord* largestAddr = 5562 (HeapWord*) _cmsSpace->dictionary()->find_largest_dict(); 5563 if (largestAddr == NULL) { 5564 // The dictionary appears to be empty. In this case 5565 // try to coalesce at the end of the heap. 5566 largestAddr = _cmsSpace->end(); 5567 } 5568 size_t largestOffset = pointer_delta(largestAddr, minAddr); 5569 size_t nearLargestOffset = 5570 (size_t)((double)largestOffset * nearLargestPercent) - MinChunkSize; 5571 if (PrintFLSStatistics != 0) { 5572 gclog_or_tty->print_cr( 5573 "CMS: Large Block: " PTR_FORMAT ";" 5574 " Proximity: " PTR_FORMAT " -> " PTR_FORMAT, 5575 p2i(largestAddr), 5576 p2i(_cmsSpace->nearLargestChunk()), p2i(minAddr + nearLargestOffset)); 5577 } 5578 _cmsSpace->set_nearLargestChunk(minAddr + nearLargestOffset); 5579 } 5580 5581 bool ConcurrentMarkSweepGeneration::isNearLargestChunk(HeapWord* addr) { 5582 return addr >= _cmsSpace->nearLargestChunk(); 5583 } 5584 5585 FreeChunk* ConcurrentMarkSweepGeneration::find_chunk_at_end() { 5586 return _cmsSpace->find_chunk_at_end(); 5587 } 5588 5589 void ConcurrentMarkSweepGeneration::update_gc_stats(Generation* current_generation, 5590 bool full) { 5591 // If the young generation has been collected, gather any statistics 5592 // that are of interest at this point. 5593 bool current_is_young = GenCollectedHeap::heap()->is_young_gen(current_generation); 5594 if (!full && current_is_young) { 5595 // Gather statistics on the young generation collection. 5596 collector()->stats().record_gc0_end(used()); 5597 } 5598 } 5599 5600 void CMSCollector::sweepWork(ConcurrentMarkSweepGeneration* old_gen) { 5601 // We iterate over the space(s) underlying this generation, 5602 // checking the mark bit map to see if the bits corresponding 5603 // to specific blocks are marked or not. Blocks that are 5604 // marked are live and are not swept up. All remaining blocks 5605 // are swept up, with coalescing on-the-fly as we sweep up 5606 // contiguous free and/or garbage blocks: 5607 // We need to ensure that the sweeper synchronizes with allocators 5608 // and stop-the-world collectors. In particular, the following 5609 // locks are used: 5610 // . CMS token: if this is held, a stop the world collection cannot occur 5611 // . freelistLock: if this is held no allocation can occur from this 5612 // generation by another thread 5613 // . bitMapLock: if this is held, no other thread can access or update 5614 // 5615 5616 // Note that we need to hold the freelistLock if we use 5617 // block iterate below; else the iterator might go awry if 5618 // a mutator (or promotion) causes block contents to change 5619 // (for instance if the allocator divvies up a block). 5620 // If we hold the free list lock, for all practical purposes 5621 // young generation GC's can't occur (they'll usually need to 5622 // promote), so we might as well prevent all young generation 5623 // GC's while we do a sweeping step. For the same reason, we might 5624 // as well take the bit map lock for the entire duration 5625 5626 // check that we hold the requisite locks 5627 assert(have_cms_token(), "Should hold cms token"); 5628 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), "Should possess CMS token to sweep"); 5629 assert_lock_strong(old_gen->freelistLock()); 5630 assert_lock_strong(bitMapLock()); 5631 5632 assert(!_inter_sweep_timer.is_active(), "Was switched off in an outer context"); 5633 assert(_intra_sweep_timer.is_active(), "Was switched on in an outer context"); 5634 old_gen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()), 5635 _inter_sweep_estimate.padded_average(), 5636 _intra_sweep_estimate.padded_average()); 5637 old_gen->setNearLargestChunk(); 5638 5639 { 5640 SweepClosure sweepClosure(this, old_gen, &_markBitMap, CMSYield); 5641 old_gen->cmsSpace()->blk_iterate_careful(&sweepClosure); 5642 // We need to free-up/coalesce garbage/blocks from a 5643 // co-terminal free run. This is done in the SweepClosure 5644 // destructor; so, do not remove this scope, else the 5645 // end-of-sweep-census below will be off by a little bit. 5646 } 5647 old_gen->cmsSpace()->sweep_completed(); 5648 old_gen->cmsSpace()->endSweepFLCensus(sweep_count()); 5649 if (should_unload_classes()) { // unloaded classes this cycle, 5650 _concurrent_cycles_since_last_unload = 0; // ... reset count 5651 } else { // did not unload classes, 5652 _concurrent_cycles_since_last_unload++; // ... increment count 5653 } 5654 } 5655 5656 // Reset CMS data structures (for now just the marking bit map) 5657 // preparatory for the next cycle. 5658 void CMSCollector::reset_concurrent() { 5659 CMSTokenSyncWithLocks ts(true, bitMapLock()); 5660 5661 // If the state is not "Resetting", the foreground thread 5662 // has done a collection and the resetting. 5663 if (_collectorState != Resetting) { 5664 assert(_collectorState == Idling, "The state should only change" 5665 " because the foreground collector has finished the collection"); 5666 return; 5667 } 5668 5669 // Clear the mark bitmap (no grey objects to start with) 5670 // for the next cycle. 5671 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 5672 CMSPhaseAccounting cmspa(this, "reset", !PrintGCDetails); 5673 5674 HeapWord* curAddr = _markBitMap.startWord(); 5675 while (curAddr < _markBitMap.endWord()) { 5676 size_t remaining = pointer_delta(_markBitMap.endWord(), curAddr); 5677 MemRegion chunk(curAddr, MIN2(CMSBitMapYieldQuantum, remaining)); 5678 _markBitMap.clear_large_range(chunk); 5679 if (ConcurrentMarkSweepThread::should_yield() && 5680 !foregroundGCIsActive() && 5681 CMSYield) { 5682 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 5683 "CMS thread should hold CMS token"); 5684 assert_lock_strong(bitMapLock()); 5685 bitMapLock()->unlock(); 5686 ConcurrentMarkSweepThread::desynchronize(true); 5687 stopTimer(); 5688 if (PrintCMSStatistics != 0) { 5689 incrementYields(); 5690 } 5691 5692 // See the comment in coordinator_yield() 5693 for (unsigned i = 0; i < CMSYieldSleepCount && 5694 ConcurrentMarkSweepThread::should_yield() && 5695 !CMSCollector::foregroundGCIsActive(); ++i) { 5696 os::sleep(Thread::current(), 1, false); 5697 } 5698 5699 ConcurrentMarkSweepThread::synchronize(true); 5700 bitMapLock()->lock_without_safepoint_check(); 5701 startTimer(); 5702 } 5703 curAddr = chunk.end(); 5704 } 5705 // A successful mostly concurrent collection has been done. 5706 // Because only the full (i.e., concurrent mode failure) collections 5707 // are being measured for gc overhead limits, clean the "near" flag 5708 // and count. 5709 size_policy()->reset_gc_overhead_limit_count(); 5710 _collectorState = Idling; 5711 5712 register_gc_end(); 5713 } 5714 5715 // Same as above but for STW paths 5716 void CMSCollector::reset_stw() { 5717 // already have the lock 5718 assert(_collectorState == Resetting, "just checking"); 5719 assert_lock_strong(bitMapLock()); 5720 GCIdMarkAndRestore gc_id_mark(_cmsThread->gc_id()); 5721 _markBitMap.clear_all(); 5722 _collectorState = Idling; 5723 register_gc_end(); 5724 } 5725 5726 void CMSCollector::do_CMS_operation(CMS_op_type op, GCCause::Cause gc_cause) { 5727 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 5728 GCTraceTime t(GCCauseString("GC", gc_cause), PrintGC, !PrintGCDetails, NULL); 5729 TraceCollectorStats tcs(counters()); 5730 5731 switch (op) { 5732 case CMS_op_checkpointRootsInitial: { 5733 SvcGCMarker sgcm(SvcGCMarker::OTHER); 5734 checkpointRootsInitial(); 5735 if (PrintGC) { 5736 _cmsGen->printOccupancy("initial-mark"); 5737 } 5738 break; 5739 } 5740 case CMS_op_checkpointRootsFinal: { 5741 SvcGCMarker sgcm(SvcGCMarker::OTHER); 5742 checkpointRootsFinal(); 5743 if (PrintGC) { 5744 _cmsGen->printOccupancy("remark"); 5745 } 5746 break; 5747 } 5748 default: 5749 fatal("No such CMS_op"); 5750 } 5751 } 5752 5753 #ifndef PRODUCT 5754 size_t const CMSCollector::skip_header_HeapWords() { 5755 return FreeChunk::header_size(); 5756 } 5757 5758 // Try and collect here conditions that should hold when 5759 // CMS thread is exiting. The idea is that the foreground GC 5760 // thread should not be blocked if it wants to terminate 5761 // the CMS thread and yet continue to run the VM for a while 5762 // after that. 5763 void CMSCollector::verify_ok_to_terminate() const { 5764 assert(Thread::current()->is_ConcurrentGC_thread(), 5765 "should be called by CMS thread"); 5766 assert(!_foregroundGCShouldWait, "should be false"); 5767 // We could check here that all the various low-level locks 5768 // are not held by the CMS thread, but that is overkill; see 5769 // also CMSThread::verify_ok_to_terminate() where the CGC_lock 5770 // is checked. 5771 } 5772 #endif 5773 5774 size_t CMSCollector::block_size_using_printezis_bits(HeapWord* addr) const { 5775 assert(_markBitMap.isMarked(addr) && _markBitMap.isMarked(addr + 1), 5776 "missing Printezis mark?"); 5777 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 5778 size_t size = pointer_delta(nextOneAddr + 1, addr); 5779 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 5780 "alignment problem"); 5781 assert(size >= 3, "Necessary for Printezis marks to work"); 5782 return size; 5783 } 5784 5785 // A variant of the above (block_size_using_printezis_bits()) except 5786 // that we return 0 if the P-bits are not yet set. 5787 size_t CMSCollector::block_size_if_printezis_bits(HeapWord* addr) const { 5788 if (_markBitMap.isMarked(addr + 1)) { 5789 assert(_markBitMap.isMarked(addr), "P-bit can be set only for marked objects"); 5790 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 5791 size_t size = pointer_delta(nextOneAddr + 1, addr); 5792 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 5793 "alignment problem"); 5794 assert(size >= 3, "Necessary for Printezis marks to work"); 5795 return size; 5796 } 5797 return 0; 5798 } 5799 5800 HeapWord* CMSCollector::next_card_start_after_block(HeapWord* addr) const { 5801 size_t sz = 0; 5802 oop p = (oop)addr; 5803 if (p->klass_or_null() != NULL) { 5804 sz = CompactibleFreeListSpace::adjustObjectSize(p->size()); 5805 } else { 5806 sz = block_size_using_printezis_bits(addr); 5807 } 5808 assert(sz > 0, "size must be nonzero"); 5809 HeapWord* next_block = addr + sz; 5810 HeapWord* next_card = (HeapWord*)round_to((uintptr_t)next_block, 5811 CardTableModRefBS::card_size); 5812 assert(round_down((uintptr_t)addr, CardTableModRefBS::card_size) < 5813 round_down((uintptr_t)next_card, CardTableModRefBS::card_size), 5814 "must be different cards"); 5815 return next_card; 5816 } 5817 5818 5819 // CMS Bit Map Wrapper ///////////////////////////////////////// 5820 5821 // Construct a CMS bit map infrastructure, but don't create the 5822 // bit vector itself. That is done by a separate call CMSBitMap::allocate() 5823 // further below. 5824 CMSBitMap::CMSBitMap(int shifter, int mutex_rank, const char* mutex_name): 5825 _bm(), 5826 _shifter(shifter), 5827 _lock(mutex_rank >= 0 ? new Mutex(mutex_rank, mutex_name, true, 5828 Monitor::_safepoint_check_sometimes) : NULL) 5829 { 5830 _bmStartWord = 0; 5831 _bmWordSize = 0; 5832 } 5833 5834 bool CMSBitMap::allocate(MemRegion mr) { 5835 _bmStartWord = mr.start(); 5836 _bmWordSize = mr.word_size(); 5837 ReservedSpace brs(ReservedSpace::allocation_align_size_up( 5838 (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1)); 5839 if (!brs.is_reserved()) { 5840 warning("CMS bit map allocation failure"); 5841 return false; 5842 } 5843 // For now we'll just commit all of the bit map up front. 5844 // Later on we'll try to be more parsimonious with swap. 5845 if (!_virtual_space.initialize(brs, brs.size())) { 5846 warning("CMS bit map backing store failure"); 5847 return false; 5848 } 5849 assert(_virtual_space.committed_size() == brs.size(), 5850 "didn't reserve backing store for all of CMS bit map?"); 5851 _bm.set_map((BitMap::bm_word_t*)_virtual_space.low()); 5852 assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >= 5853 _bmWordSize, "inconsistency in bit map sizing"); 5854 _bm.set_size(_bmWordSize >> _shifter); 5855 5856 // bm.clear(); // can we rely on getting zero'd memory? verify below 5857 assert(isAllClear(), 5858 "Expected zero'd memory from ReservedSpace constructor"); 5859 assert(_bm.size() == heapWordDiffToOffsetDiff(sizeInWords()), 5860 "consistency check"); 5861 return true; 5862 } 5863 5864 void CMSBitMap::dirty_range_iterate_clear(MemRegion mr, MemRegionClosure* cl) { 5865 HeapWord *next_addr, *end_addr, *last_addr; 5866 assert_locked(); 5867 assert(covers(mr), "out-of-range error"); 5868 // XXX assert that start and end are appropriately aligned 5869 for (next_addr = mr.start(), end_addr = mr.end(); 5870 next_addr < end_addr; next_addr = last_addr) { 5871 MemRegion dirty_region = getAndClearMarkedRegion(next_addr, end_addr); 5872 last_addr = dirty_region.end(); 5873 if (!dirty_region.is_empty()) { 5874 cl->do_MemRegion(dirty_region); 5875 } else { 5876 assert(last_addr == end_addr, "program logic"); 5877 return; 5878 } 5879 } 5880 } 5881 5882 void CMSBitMap::print_on_error(outputStream* st, const char* prefix) const { 5883 _bm.print_on_error(st, prefix); 5884 } 5885 5886 #ifndef PRODUCT 5887 void CMSBitMap::assert_locked() const { 5888 CMSLockVerifier::assert_locked(lock()); 5889 } 5890 5891 bool CMSBitMap::covers(MemRegion mr) const { 5892 // assert(_bm.map() == _virtual_space.low(), "map inconsistency"); 5893 assert((size_t)_bm.size() == (_bmWordSize >> _shifter), 5894 "size inconsistency"); 5895 return (mr.start() >= _bmStartWord) && 5896 (mr.end() <= endWord()); 5897 } 5898 5899 bool CMSBitMap::covers(HeapWord* start, size_t size) const { 5900 return (start >= _bmStartWord && (start + size) <= endWord()); 5901 } 5902 5903 void CMSBitMap::verifyNoOneBitsInRange(HeapWord* left, HeapWord* right) { 5904 // verify that there are no 1 bits in the interval [left, right) 5905 FalseBitMapClosure falseBitMapClosure; 5906 iterate(&falseBitMapClosure, left, right); 5907 } 5908 5909 void CMSBitMap::region_invariant(MemRegion mr) 5910 { 5911 assert_locked(); 5912 // mr = mr.intersection(MemRegion(_bmStartWord, _bmWordSize)); 5913 assert(!mr.is_empty(), "unexpected empty region"); 5914 assert(covers(mr), "mr should be covered by bit map"); 5915 // convert address range into offset range 5916 size_t start_ofs = heapWordToOffset(mr.start()); 5917 // Make sure that end() is appropriately aligned 5918 assert(mr.end() == (HeapWord*)round_to((intptr_t)mr.end(), 5919 (1 << (_shifter+LogHeapWordSize))), 5920 "Misaligned mr.end()"); 5921 size_t end_ofs = heapWordToOffset(mr.end()); 5922 assert(end_ofs > start_ofs, "Should mark at least one bit"); 5923 } 5924 5925 #endif 5926 5927 bool CMSMarkStack::allocate(size_t size) { 5928 // allocate a stack of the requisite depth 5929 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 5930 size * sizeof(oop))); 5931 if (!rs.is_reserved()) { 5932 warning("CMSMarkStack allocation failure"); 5933 return false; 5934 } 5935 if (!_virtual_space.initialize(rs, rs.size())) { 5936 warning("CMSMarkStack backing store failure"); 5937 return false; 5938 } 5939 assert(_virtual_space.committed_size() == rs.size(), 5940 "didn't reserve backing store for all of CMS stack?"); 5941 _base = (oop*)(_virtual_space.low()); 5942 _index = 0; 5943 _capacity = size; 5944 NOT_PRODUCT(_max_depth = 0); 5945 return true; 5946 } 5947 5948 // XXX FIX ME !!! In the MT case we come in here holding a 5949 // leaf lock. For printing we need to take a further lock 5950 // which has lower rank. We need to recalibrate the two 5951 // lock-ranks involved in order to be able to print the 5952 // messages below. (Or defer the printing to the caller. 5953 // For now we take the expedient path of just disabling the 5954 // messages for the problematic case.) 5955 void CMSMarkStack::expand() { 5956 assert(_capacity <= MarkStackSizeMax, "stack bigger than permitted"); 5957 if (_capacity == MarkStackSizeMax) { 5958 if (_hit_limit++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) { 5959 // We print a warning message only once per CMS cycle. 5960 gclog_or_tty->print_cr(" (benign) Hit CMSMarkStack max size limit"); 5961 } 5962 return; 5963 } 5964 // Double capacity if possible 5965 size_t new_capacity = MIN2(_capacity*2, MarkStackSizeMax); 5966 // Do not give up existing stack until we have managed to 5967 // get the double capacity that we desired. 5968 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 5969 new_capacity * sizeof(oop))); 5970 if (rs.is_reserved()) { 5971 // Release the backing store associated with old stack 5972 _virtual_space.release(); 5973 // Reinitialize virtual space for new stack 5974 if (!_virtual_space.initialize(rs, rs.size())) { 5975 fatal("Not enough swap for expanded marking stack"); 5976 } 5977 _base = (oop*)(_virtual_space.low()); 5978 _index = 0; 5979 _capacity = new_capacity; 5980 } else if (_failed_double++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) { 5981 // Failed to double capacity, continue; 5982 // we print a detail message only once per CMS cycle. 5983 gclog_or_tty->print(" (benign) Failed to expand marking stack from " SIZE_FORMAT "K to " 5984 SIZE_FORMAT "K", 5985 _capacity / K, new_capacity / K); 5986 } 5987 } 5988 5989 5990 // Closures 5991 // XXX: there seems to be a lot of code duplication here; 5992 // should refactor and consolidate common code. 5993 5994 // This closure is used to mark refs into the CMS generation in 5995 // the CMS bit map. Called at the first checkpoint. This closure 5996 // assumes that we do not need to re-mark dirty cards; if the CMS 5997 // generation on which this is used is not an oldest 5998 // generation then this will lose younger_gen cards! 5999 6000 MarkRefsIntoClosure::MarkRefsIntoClosure( 6001 MemRegion span, CMSBitMap* bitMap): 6002 _span(span), 6003 _bitMap(bitMap) 6004 { 6005 assert(ref_processor() == NULL, "deliberately left NULL"); 6006 assert(_bitMap->covers(_span), "_bitMap/_span mismatch"); 6007 } 6008 6009 void MarkRefsIntoClosure::do_oop(oop obj) { 6010 // if p points into _span, then mark corresponding bit in _markBitMap 6011 assert(obj->is_oop(), "expected an oop"); 6012 HeapWord* addr = (HeapWord*)obj; 6013 if (_span.contains(addr)) { 6014 // this should be made more efficient 6015 _bitMap->mark(addr); 6016 } 6017 } 6018 6019 void MarkRefsIntoClosure::do_oop(oop* p) { MarkRefsIntoClosure::do_oop_work(p); } 6020 void MarkRefsIntoClosure::do_oop(narrowOop* p) { MarkRefsIntoClosure::do_oop_work(p); } 6021 6022 Par_MarkRefsIntoClosure::Par_MarkRefsIntoClosure( 6023 MemRegion span, CMSBitMap* bitMap): 6024 _span(span), 6025 _bitMap(bitMap) 6026 { 6027 assert(ref_processor() == NULL, "deliberately left NULL"); 6028 assert(_bitMap->covers(_span), "_bitMap/_span mismatch"); 6029 } 6030 6031 void Par_MarkRefsIntoClosure::do_oop(oop obj) { 6032 // if p points into _span, then mark corresponding bit in _markBitMap 6033 assert(obj->is_oop(), "expected an oop"); 6034 HeapWord* addr = (HeapWord*)obj; 6035 if (_span.contains(addr)) { 6036 // this should be made more efficient 6037 _bitMap->par_mark(addr); 6038 } 6039 } 6040 6041 void Par_MarkRefsIntoClosure::do_oop(oop* p) { Par_MarkRefsIntoClosure::do_oop_work(p); } 6042 void Par_MarkRefsIntoClosure::do_oop(narrowOop* p) { Par_MarkRefsIntoClosure::do_oop_work(p); } 6043 6044 // A variant of the above, used for CMS marking verification. 6045 MarkRefsIntoVerifyClosure::MarkRefsIntoVerifyClosure( 6046 MemRegion span, CMSBitMap* verification_bm, CMSBitMap* cms_bm): 6047 _span(span), 6048 _verification_bm(verification_bm), 6049 _cms_bm(cms_bm) 6050 { 6051 assert(ref_processor() == NULL, "deliberately left NULL"); 6052 assert(_verification_bm->covers(_span), "_verification_bm/_span mismatch"); 6053 } 6054 6055 void MarkRefsIntoVerifyClosure::do_oop(oop obj) { 6056 // if p points into _span, then mark corresponding bit in _markBitMap 6057 assert(obj->is_oop(), "expected an oop"); 6058 HeapWord* addr = (HeapWord*)obj; 6059 if (_span.contains(addr)) { 6060 _verification_bm->mark(addr); 6061 if (!_cms_bm->isMarked(addr)) { 6062 oop(addr)->print(); 6063 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)", p2i(addr)); 6064 fatal("... aborting"); 6065 } 6066 } 6067 } 6068 6069 void MarkRefsIntoVerifyClosure::do_oop(oop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); } 6070 void MarkRefsIntoVerifyClosure::do_oop(narrowOop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); } 6071 6072 ////////////////////////////////////////////////// 6073 // MarkRefsIntoAndScanClosure 6074 ////////////////////////////////////////////////// 6075 6076 MarkRefsIntoAndScanClosure::MarkRefsIntoAndScanClosure(MemRegion span, 6077 ReferenceProcessor* rp, 6078 CMSBitMap* bit_map, 6079 CMSBitMap* mod_union_table, 6080 CMSMarkStack* mark_stack, 6081 CMSCollector* collector, 6082 bool should_yield, 6083 bool concurrent_precleaning): 6084 _collector(collector), 6085 _span(span), 6086 _bit_map(bit_map), 6087 _mark_stack(mark_stack), 6088 _pushAndMarkClosure(collector, span, rp, bit_map, mod_union_table, 6089 mark_stack, concurrent_precleaning), 6090 _yield(should_yield), 6091 _concurrent_precleaning(concurrent_precleaning), 6092 _freelistLock(NULL) 6093 { 6094 // FIXME: Should initialize in base class constructor. 6095 assert(rp != NULL, "ref_processor shouldn't be NULL"); 6096 set_ref_processor_internal(rp); 6097 } 6098 6099 // This closure is used to mark refs into the CMS generation at the 6100 // second (final) checkpoint, and to scan and transitively follow 6101 // the unmarked oops. It is also used during the concurrent precleaning 6102 // phase while scanning objects on dirty cards in the CMS generation. 6103 // The marks are made in the marking bit map and the marking stack is 6104 // used for keeping the (newly) grey objects during the scan. 6105 // The parallel version (Par_...) appears further below. 6106 void MarkRefsIntoAndScanClosure::do_oop(oop obj) { 6107 if (obj != NULL) { 6108 assert(obj->is_oop(), "expected an oop"); 6109 HeapWord* addr = (HeapWord*)obj; 6110 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 6111 assert(_collector->overflow_list_is_empty(), 6112 "overflow list should be empty"); 6113 if (_span.contains(addr) && 6114 !_bit_map->isMarked(addr)) { 6115 // mark bit map (object is now grey) 6116 _bit_map->mark(addr); 6117 // push on marking stack (stack should be empty), and drain the 6118 // stack by applying this closure to the oops in the oops popped 6119 // from the stack (i.e. blacken the grey objects) 6120 bool res = _mark_stack->push(obj); 6121 assert(res, "Should have space to push on empty stack"); 6122 do { 6123 oop new_oop = _mark_stack->pop(); 6124 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 6125 assert(_bit_map->isMarked((HeapWord*)new_oop), 6126 "only grey objects on this stack"); 6127 // iterate over the oops in this oop, marking and pushing 6128 // the ones in CMS heap (i.e. in _span). 6129 new_oop->oop_iterate(&_pushAndMarkClosure); 6130 // check if it's time to yield 6131 do_yield_check(); 6132 } while (!_mark_stack->isEmpty() || 6133 (!_concurrent_precleaning && take_from_overflow_list())); 6134 // if marking stack is empty, and we are not doing this 6135 // during precleaning, then check the overflow list 6136 } 6137 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 6138 assert(_collector->overflow_list_is_empty(), 6139 "overflow list was drained above"); 6140 6141 assert(_collector->no_preserved_marks(), 6142 "All preserved marks should have been restored above"); 6143 } 6144 } 6145 6146 void MarkRefsIntoAndScanClosure::do_oop(oop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); } 6147 void MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); } 6148 6149 void MarkRefsIntoAndScanClosure::do_yield_work() { 6150 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6151 "CMS thread should hold CMS token"); 6152 assert_lock_strong(_freelistLock); 6153 assert_lock_strong(_bit_map->lock()); 6154 // relinquish the free_list_lock and bitMaplock() 6155 _bit_map->lock()->unlock(); 6156 _freelistLock->unlock(); 6157 ConcurrentMarkSweepThread::desynchronize(true); 6158 _collector->stopTimer(); 6159 if (PrintCMSStatistics != 0) { 6160 _collector->incrementYields(); 6161 } 6162 6163 // See the comment in coordinator_yield() 6164 for (unsigned i = 0; 6165 i < CMSYieldSleepCount && 6166 ConcurrentMarkSweepThread::should_yield() && 6167 !CMSCollector::foregroundGCIsActive(); 6168 ++i) { 6169 os::sleep(Thread::current(), 1, false); 6170 } 6171 6172 ConcurrentMarkSweepThread::synchronize(true); 6173 _freelistLock->lock_without_safepoint_check(); 6174 _bit_map->lock()->lock_without_safepoint_check(); 6175 _collector->startTimer(); 6176 } 6177 6178 /////////////////////////////////////////////////////////// 6179 // Par_MarkRefsIntoAndScanClosure: a parallel version of 6180 // MarkRefsIntoAndScanClosure 6181 /////////////////////////////////////////////////////////// 6182 Par_MarkRefsIntoAndScanClosure::Par_MarkRefsIntoAndScanClosure( 6183 CMSCollector* collector, MemRegion span, ReferenceProcessor* rp, 6184 CMSBitMap* bit_map, OopTaskQueue* work_queue): 6185 _span(span), 6186 _bit_map(bit_map), 6187 _work_queue(work_queue), 6188 _low_water_mark(MIN2((work_queue->max_elems()/4), 6189 ((uint)CMSWorkQueueDrainThreshold * ParallelGCThreads))), 6190 _par_pushAndMarkClosure(collector, span, rp, bit_map, work_queue) 6191 { 6192 // FIXME: Should initialize in base class constructor. 6193 assert(rp != NULL, "ref_processor shouldn't be NULL"); 6194 set_ref_processor_internal(rp); 6195 } 6196 6197 // This closure is used to mark refs into the CMS generation at the 6198 // second (final) checkpoint, and to scan and transitively follow 6199 // the unmarked oops. The marks are made in the marking bit map and 6200 // the work_queue is used for keeping the (newly) grey objects during 6201 // the scan phase whence they are also available for stealing by parallel 6202 // threads. Since the marking bit map is shared, updates are 6203 // synchronized (via CAS). 6204 void Par_MarkRefsIntoAndScanClosure::do_oop(oop obj) { 6205 if (obj != NULL) { 6206 // Ignore mark word because this could be an already marked oop 6207 // that may be chained at the end of the overflow list. 6208 assert(obj->is_oop(true), "expected an oop"); 6209 HeapWord* addr = (HeapWord*)obj; 6210 if (_span.contains(addr) && 6211 !_bit_map->isMarked(addr)) { 6212 // mark bit map (object will become grey): 6213 // It is possible for several threads to be 6214 // trying to "claim" this object concurrently; 6215 // the unique thread that succeeds in marking the 6216 // object first will do the subsequent push on 6217 // to the work queue (or overflow list). 6218 if (_bit_map->par_mark(addr)) { 6219 // push on work_queue (which may not be empty), and trim the 6220 // queue to an appropriate length by applying this closure to 6221 // the oops in the oops popped from the stack (i.e. blacken the 6222 // grey objects) 6223 bool res = _work_queue->push(obj); 6224 assert(res, "Low water mark should be less than capacity?"); 6225 trim_queue(_low_water_mark); 6226 } // Else, another thread claimed the object 6227 } 6228 } 6229 } 6230 6231 void Par_MarkRefsIntoAndScanClosure::do_oop(oop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); } 6232 void Par_MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); } 6233 6234 // This closure is used to rescan the marked objects on the dirty cards 6235 // in the mod union table and the card table proper. 6236 size_t ScanMarkedObjectsAgainCarefullyClosure::do_object_careful_m( 6237 oop p, MemRegion mr) { 6238 6239 size_t size = 0; 6240 HeapWord* addr = (HeapWord*)p; 6241 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6242 assert(_span.contains(addr), "we are scanning the CMS generation"); 6243 // check if it's time to yield 6244 if (do_yield_check()) { 6245 // We yielded for some foreground stop-world work, 6246 // and we have been asked to abort this ongoing preclean cycle. 6247 return 0; 6248 } 6249 if (_bitMap->isMarked(addr)) { 6250 // it's marked; is it potentially uninitialized? 6251 if (p->klass_or_null() != NULL) { 6252 // an initialized object; ignore mark word in verification below 6253 // since we are running concurrent with mutators 6254 assert(p->is_oop(true), "should be an oop"); 6255 if (p->is_objArray()) { 6256 // objArrays are precisely marked; restrict scanning 6257 // to dirty cards only. 6258 size = CompactibleFreeListSpace::adjustObjectSize( 6259 p->oop_iterate_size(_scanningClosure, mr)); 6260 } else { 6261 // A non-array may have been imprecisely marked; we need 6262 // to scan object in its entirety. 6263 size = CompactibleFreeListSpace::adjustObjectSize( 6264 p->oop_iterate_size(_scanningClosure)); 6265 } 6266 #ifdef ASSERT 6267 size_t direct_size = 6268 CompactibleFreeListSpace::adjustObjectSize(p->size()); 6269 assert(size == direct_size, "Inconsistency in size"); 6270 assert(size >= 3, "Necessary for Printezis marks to work"); 6271 if (!_bitMap->isMarked(addr+1)) { 6272 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size); 6273 } else { 6274 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size-1); 6275 assert(_bitMap->isMarked(addr+size-1), 6276 "inconsistent Printezis mark"); 6277 } 6278 #endif // ASSERT 6279 } else { 6280 // An uninitialized object. 6281 assert(_bitMap->isMarked(addr+1), "missing Printezis mark?"); 6282 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 6283 size = pointer_delta(nextOneAddr + 1, addr); 6284 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 6285 "alignment problem"); 6286 // Note that pre-cleaning needn't redirty the card. OopDesc::set_klass() 6287 // will dirty the card when the klass pointer is installed in the 6288 // object (signaling the completion of initialization). 6289 } 6290 } else { 6291 // Either a not yet marked object or an uninitialized object 6292 if (p->klass_or_null() == NULL) { 6293 // An uninitialized object, skip to the next card, since 6294 // we may not be able to read its P-bits yet. 6295 assert(size == 0, "Initial value"); 6296 } else { 6297 // An object not (yet) reached by marking: we merely need to 6298 // compute its size so as to go look at the next block. 6299 assert(p->is_oop(true), "should be an oop"); 6300 size = CompactibleFreeListSpace::adjustObjectSize(p->size()); 6301 } 6302 } 6303 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6304 return size; 6305 } 6306 6307 void ScanMarkedObjectsAgainCarefullyClosure::do_yield_work() { 6308 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6309 "CMS thread should hold CMS token"); 6310 assert_lock_strong(_freelistLock); 6311 assert_lock_strong(_bitMap->lock()); 6312 // relinquish the free_list_lock and bitMaplock() 6313 _bitMap->lock()->unlock(); 6314 _freelistLock->unlock(); 6315 ConcurrentMarkSweepThread::desynchronize(true); 6316 _collector->stopTimer(); 6317 if (PrintCMSStatistics != 0) { 6318 _collector->incrementYields(); 6319 } 6320 6321 // See the comment in coordinator_yield() 6322 for (unsigned i = 0; i < CMSYieldSleepCount && 6323 ConcurrentMarkSweepThread::should_yield() && 6324 !CMSCollector::foregroundGCIsActive(); ++i) { 6325 os::sleep(Thread::current(), 1, false); 6326 } 6327 6328 ConcurrentMarkSweepThread::synchronize(true); 6329 _freelistLock->lock_without_safepoint_check(); 6330 _bitMap->lock()->lock_without_safepoint_check(); 6331 _collector->startTimer(); 6332 } 6333 6334 6335 ////////////////////////////////////////////////////////////////// 6336 // SurvivorSpacePrecleanClosure 6337 ////////////////////////////////////////////////////////////////// 6338 // This (single-threaded) closure is used to preclean the oops in 6339 // the survivor spaces. 6340 size_t SurvivorSpacePrecleanClosure::do_object_careful(oop p) { 6341 6342 HeapWord* addr = (HeapWord*)p; 6343 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6344 assert(!_span.contains(addr), "we are scanning the survivor spaces"); 6345 assert(p->klass_or_null() != NULL, "object should be initialized"); 6346 // an initialized object; ignore mark word in verification below 6347 // since we are running concurrent with mutators 6348 assert(p->is_oop(true), "should be an oop"); 6349 // Note that we do not yield while we iterate over 6350 // the interior oops of p, pushing the relevant ones 6351 // on our marking stack. 6352 size_t size = p->oop_iterate_size(_scanning_closure); 6353 do_yield_check(); 6354 // Observe that below, we do not abandon the preclean 6355 // phase as soon as we should; rather we empty the 6356 // marking stack before returning. This is to satisfy 6357 // some existing assertions. In general, it may be a 6358 // good idea to abort immediately and complete the marking 6359 // from the grey objects at a later time. 6360 while (!_mark_stack->isEmpty()) { 6361 oop new_oop = _mark_stack->pop(); 6362 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 6363 assert(_bit_map->isMarked((HeapWord*)new_oop), 6364 "only grey objects on this stack"); 6365 // iterate over the oops in this oop, marking and pushing 6366 // the ones in CMS heap (i.e. in _span). 6367 new_oop->oop_iterate(_scanning_closure); 6368 // check if it's time to yield 6369 do_yield_check(); 6370 } 6371 unsigned int after_count = 6372 GenCollectedHeap::heap()->total_collections(); 6373 bool abort = (_before_count != after_count) || 6374 _collector->should_abort_preclean(); 6375 return abort ? 0 : size; 6376 } 6377 6378 void SurvivorSpacePrecleanClosure::do_yield_work() { 6379 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6380 "CMS thread should hold CMS token"); 6381 assert_lock_strong(_bit_map->lock()); 6382 // Relinquish the bit map lock 6383 _bit_map->lock()->unlock(); 6384 ConcurrentMarkSweepThread::desynchronize(true); 6385 _collector->stopTimer(); 6386 if (PrintCMSStatistics != 0) { 6387 _collector->incrementYields(); 6388 } 6389 6390 // See the comment in coordinator_yield() 6391 for (unsigned i = 0; i < CMSYieldSleepCount && 6392 ConcurrentMarkSweepThread::should_yield() && 6393 !CMSCollector::foregroundGCIsActive(); ++i) { 6394 os::sleep(Thread::current(), 1, false); 6395 } 6396 6397 ConcurrentMarkSweepThread::synchronize(true); 6398 _bit_map->lock()->lock_without_safepoint_check(); 6399 _collector->startTimer(); 6400 } 6401 6402 // This closure is used to rescan the marked objects on the dirty cards 6403 // in the mod union table and the card table proper. In the parallel 6404 // case, although the bitMap is shared, we do a single read so the 6405 // isMarked() query is "safe". 6406 bool ScanMarkedObjectsAgainClosure::do_object_bm(oop p, MemRegion mr) { 6407 // Ignore mark word because we are running concurrent with mutators 6408 assert(p->is_oop_or_null(true), "Expected an oop or NULL at " PTR_FORMAT, p2i(p)); 6409 HeapWord* addr = (HeapWord*)p; 6410 assert(_span.contains(addr), "we are scanning the CMS generation"); 6411 bool is_obj_array = false; 6412 #ifdef ASSERT 6413 if (!_parallel) { 6414 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 6415 assert(_collector->overflow_list_is_empty(), 6416 "overflow list should be empty"); 6417 6418 } 6419 #endif // ASSERT 6420 if (_bit_map->isMarked(addr)) { 6421 // Obj arrays are precisely marked, non-arrays are not; 6422 // so we scan objArrays precisely and non-arrays in their 6423 // entirety. 6424 if (p->is_objArray()) { 6425 is_obj_array = true; 6426 if (_parallel) { 6427 p->oop_iterate(_par_scan_closure, mr); 6428 } else { 6429 p->oop_iterate(_scan_closure, mr); 6430 } 6431 } else { 6432 if (_parallel) { 6433 p->oop_iterate(_par_scan_closure); 6434 } else { 6435 p->oop_iterate(_scan_closure); 6436 } 6437 } 6438 } 6439 #ifdef ASSERT 6440 if (!_parallel) { 6441 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 6442 assert(_collector->overflow_list_is_empty(), 6443 "overflow list should be empty"); 6444 6445 } 6446 #endif // ASSERT 6447 return is_obj_array; 6448 } 6449 6450 MarkFromRootsClosure::MarkFromRootsClosure(CMSCollector* collector, 6451 MemRegion span, 6452 CMSBitMap* bitMap, CMSMarkStack* markStack, 6453 bool should_yield, bool verifying): 6454 _collector(collector), 6455 _span(span), 6456 _bitMap(bitMap), 6457 _mut(&collector->_modUnionTable), 6458 _markStack(markStack), 6459 _yield(should_yield), 6460 _skipBits(0) 6461 { 6462 assert(_markStack->isEmpty(), "stack should be empty"); 6463 _finger = _bitMap->startWord(); 6464 _threshold = _finger; 6465 assert(_collector->_restart_addr == NULL, "Sanity check"); 6466 assert(_span.contains(_finger), "Out of bounds _finger?"); 6467 DEBUG_ONLY(_verifying = verifying;) 6468 } 6469 6470 void MarkFromRootsClosure::reset(HeapWord* addr) { 6471 assert(_markStack->isEmpty(), "would cause duplicates on stack"); 6472 assert(_span.contains(addr), "Out of bounds _finger?"); 6473 _finger = addr; 6474 _threshold = (HeapWord*)round_to( 6475 (intptr_t)_finger, CardTableModRefBS::card_size); 6476 } 6477 6478 // Should revisit to see if this should be restructured for 6479 // greater efficiency. 6480 bool MarkFromRootsClosure::do_bit(size_t offset) { 6481 if (_skipBits > 0) { 6482 _skipBits--; 6483 return true; 6484 } 6485 // convert offset into a HeapWord* 6486 HeapWord* addr = _bitMap->startWord() + offset; 6487 assert(_bitMap->endWord() && addr < _bitMap->endWord(), 6488 "address out of range"); 6489 assert(_bitMap->isMarked(addr), "tautology"); 6490 if (_bitMap->isMarked(addr+1)) { 6491 // this is an allocated but not yet initialized object 6492 assert(_skipBits == 0, "tautology"); 6493 _skipBits = 2; // skip next two marked bits ("Printezis-marks") 6494 oop p = oop(addr); 6495 if (p->klass_or_null() == NULL) { 6496 DEBUG_ONLY(if (!_verifying) {) 6497 // We re-dirty the cards on which this object lies and increase 6498 // the _threshold so that we'll come back to scan this object 6499 // during the preclean or remark phase. (CMSCleanOnEnter) 6500 if (CMSCleanOnEnter) { 6501 size_t sz = _collector->block_size_using_printezis_bits(addr); 6502 HeapWord* end_card_addr = (HeapWord*)round_to( 6503 (intptr_t)(addr+sz), CardTableModRefBS::card_size); 6504 MemRegion redirty_range = MemRegion(addr, end_card_addr); 6505 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 6506 // Bump _threshold to end_card_addr; note that 6507 // _threshold cannot possibly exceed end_card_addr, anyhow. 6508 // This prevents future clearing of the card as the scan proceeds 6509 // to the right. 6510 assert(_threshold <= end_card_addr, 6511 "Because we are just scanning into this object"); 6512 if (_threshold < end_card_addr) { 6513 _threshold = end_card_addr; 6514 } 6515 if (p->klass_or_null() != NULL) { 6516 // Redirty the range of cards... 6517 _mut->mark_range(redirty_range); 6518 } // ...else the setting of klass will dirty the card anyway. 6519 } 6520 DEBUG_ONLY(}) 6521 return true; 6522 } 6523 } 6524 scanOopsInOop(addr); 6525 return true; 6526 } 6527 6528 // We take a break if we've been at this for a while, 6529 // so as to avoid monopolizing the locks involved. 6530 void MarkFromRootsClosure::do_yield_work() { 6531 // First give up the locks, then yield, then re-lock 6532 // We should probably use a constructor/destructor idiom to 6533 // do this unlock/lock or modify the MutexUnlocker class to 6534 // serve our purpose. XXX 6535 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6536 "CMS thread should hold CMS token"); 6537 assert_lock_strong(_bitMap->lock()); 6538 _bitMap->lock()->unlock(); 6539 ConcurrentMarkSweepThread::desynchronize(true); 6540 _collector->stopTimer(); 6541 if (PrintCMSStatistics != 0) { 6542 _collector->incrementYields(); 6543 } 6544 6545 // See the comment in coordinator_yield() 6546 for (unsigned i = 0; i < CMSYieldSleepCount && 6547 ConcurrentMarkSweepThread::should_yield() && 6548 !CMSCollector::foregroundGCIsActive(); ++i) { 6549 os::sleep(Thread::current(), 1, false); 6550 } 6551 6552 ConcurrentMarkSweepThread::synchronize(true); 6553 _bitMap->lock()->lock_without_safepoint_check(); 6554 _collector->startTimer(); 6555 } 6556 6557 void MarkFromRootsClosure::scanOopsInOop(HeapWord* ptr) { 6558 assert(_bitMap->isMarked(ptr), "expected bit to be set"); 6559 assert(_markStack->isEmpty(), 6560 "should drain stack to limit stack usage"); 6561 // convert ptr to an oop preparatory to scanning 6562 oop obj = oop(ptr); 6563 // Ignore mark word in verification below, since we 6564 // may be running concurrent with mutators. 6565 assert(obj->is_oop(true), "should be an oop"); 6566 assert(_finger <= ptr, "_finger runneth ahead"); 6567 // advance the finger to right end of this object 6568 _finger = ptr + obj->size(); 6569 assert(_finger > ptr, "we just incremented it above"); 6570 // On large heaps, it may take us some time to get through 6571 // the marking phase. During 6572 // this time it's possible that a lot of mutations have 6573 // accumulated in the card table and the mod union table -- 6574 // these mutation records are redundant until we have 6575 // actually traced into the corresponding card. 6576 // Here, we check whether advancing the finger would make 6577 // us cross into a new card, and if so clear corresponding 6578 // cards in the MUT (preclean them in the card-table in the 6579 // future). 6580 6581 DEBUG_ONLY(if (!_verifying) {) 6582 // The clean-on-enter optimization is disabled by default, 6583 // until we fix 6178663. 6584 if (CMSCleanOnEnter && (_finger > _threshold)) { 6585 // [_threshold, _finger) represents the interval 6586 // of cards to be cleared in MUT (or precleaned in card table). 6587 // The set of cards to be cleared is all those that overlap 6588 // with the interval [_threshold, _finger); note that 6589 // _threshold is always kept card-aligned but _finger isn't 6590 // always card-aligned. 6591 HeapWord* old_threshold = _threshold; 6592 assert(old_threshold == (HeapWord*)round_to( 6593 (intptr_t)old_threshold, CardTableModRefBS::card_size), 6594 "_threshold should always be card-aligned"); 6595 _threshold = (HeapWord*)round_to( 6596 (intptr_t)_finger, CardTableModRefBS::card_size); 6597 MemRegion mr(old_threshold, _threshold); 6598 assert(!mr.is_empty(), "Control point invariant"); 6599 assert(_span.contains(mr), "Should clear within span"); 6600 _mut->clear_range(mr); 6601 } 6602 DEBUG_ONLY(}) 6603 // Note: the finger doesn't advance while we drain 6604 // the stack below. 6605 PushOrMarkClosure pushOrMarkClosure(_collector, 6606 _span, _bitMap, _markStack, 6607 _finger, this); 6608 bool res = _markStack->push(obj); 6609 assert(res, "Empty non-zero size stack should have space for single push"); 6610 while (!_markStack->isEmpty()) { 6611 oop new_oop = _markStack->pop(); 6612 // Skip verifying header mark word below because we are 6613 // running concurrent with mutators. 6614 assert(new_oop->is_oop(true), "Oops! expected to pop an oop"); 6615 // now scan this oop's oops 6616 new_oop->oop_iterate(&pushOrMarkClosure); 6617 do_yield_check(); 6618 } 6619 assert(_markStack->isEmpty(), "tautology, emphasizing post-condition"); 6620 } 6621 6622 Par_MarkFromRootsClosure::Par_MarkFromRootsClosure(CMSConcMarkingTask* task, 6623 CMSCollector* collector, MemRegion span, 6624 CMSBitMap* bit_map, 6625 OopTaskQueue* work_queue, 6626 CMSMarkStack* overflow_stack): 6627 _collector(collector), 6628 _whole_span(collector->_span), 6629 _span(span), 6630 _bit_map(bit_map), 6631 _mut(&collector->_modUnionTable), 6632 _work_queue(work_queue), 6633 _overflow_stack(overflow_stack), 6634 _skip_bits(0), 6635 _task(task) 6636 { 6637 assert(_work_queue->size() == 0, "work_queue should be empty"); 6638 _finger = span.start(); 6639 _threshold = _finger; // XXX Defer clear-on-enter optimization for now 6640 assert(_span.contains(_finger), "Out of bounds _finger?"); 6641 } 6642 6643 // Should revisit to see if this should be restructured for 6644 // greater efficiency. 6645 bool Par_MarkFromRootsClosure::do_bit(size_t offset) { 6646 if (_skip_bits > 0) { 6647 _skip_bits--; 6648 return true; 6649 } 6650 // convert offset into a HeapWord* 6651 HeapWord* addr = _bit_map->startWord() + offset; 6652 assert(_bit_map->endWord() && addr < _bit_map->endWord(), 6653 "address out of range"); 6654 assert(_bit_map->isMarked(addr), "tautology"); 6655 if (_bit_map->isMarked(addr+1)) { 6656 // this is an allocated object that might not yet be initialized 6657 assert(_skip_bits == 0, "tautology"); 6658 _skip_bits = 2; // skip next two marked bits ("Printezis-marks") 6659 oop p = oop(addr); 6660 if (p->klass_or_null() == NULL) { 6661 // in the case of Clean-on-Enter optimization, redirty card 6662 // and avoid clearing card by increasing the threshold. 6663 return true; 6664 } 6665 } 6666 scan_oops_in_oop(addr); 6667 return true; 6668 } 6669 6670 void Par_MarkFromRootsClosure::scan_oops_in_oop(HeapWord* ptr) { 6671 assert(_bit_map->isMarked(ptr), "expected bit to be set"); 6672 // Should we assert that our work queue is empty or 6673 // below some drain limit? 6674 assert(_work_queue->size() == 0, 6675 "should drain stack to limit stack usage"); 6676 // convert ptr to an oop preparatory to scanning 6677 oop obj = oop(ptr); 6678 // Ignore mark word in verification below, since we 6679 // may be running concurrent with mutators. 6680 assert(obj->is_oop(true), "should be an oop"); 6681 assert(_finger <= ptr, "_finger runneth ahead"); 6682 // advance the finger to right end of this object 6683 _finger = ptr + obj->size(); 6684 assert(_finger > ptr, "we just incremented it above"); 6685 // On large heaps, it may take us some time to get through 6686 // the marking phase. During 6687 // this time it's possible that a lot of mutations have 6688 // accumulated in the card table and the mod union table -- 6689 // these mutation records are redundant until we have 6690 // actually traced into the corresponding card. 6691 // Here, we check whether advancing the finger would make 6692 // us cross into a new card, and if so clear corresponding 6693 // cards in the MUT (preclean them in the card-table in the 6694 // future). 6695 6696 // The clean-on-enter optimization is disabled by default, 6697 // until we fix 6178663. 6698 if (CMSCleanOnEnter && (_finger > _threshold)) { 6699 // [_threshold, _finger) represents the interval 6700 // of cards to be cleared in MUT (or precleaned in card table). 6701 // The set of cards to be cleared is all those that overlap 6702 // with the interval [_threshold, _finger); note that 6703 // _threshold is always kept card-aligned but _finger isn't 6704 // always card-aligned. 6705 HeapWord* old_threshold = _threshold; 6706 assert(old_threshold == (HeapWord*)round_to( 6707 (intptr_t)old_threshold, CardTableModRefBS::card_size), 6708 "_threshold should always be card-aligned"); 6709 _threshold = (HeapWord*)round_to( 6710 (intptr_t)_finger, CardTableModRefBS::card_size); 6711 MemRegion mr(old_threshold, _threshold); 6712 assert(!mr.is_empty(), "Control point invariant"); 6713 assert(_span.contains(mr), "Should clear within span"); // _whole_span ?? 6714 _mut->clear_range(mr); 6715 } 6716 6717 // Note: the local finger doesn't advance while we drain 6718 // the stack below, but the global finger sure can and will. 6719 HeapWord** gfa = _task->global_finger_addr(); 6720 Par_PushOrMarkClosure pushOrMarkClosure(_collector, 6721 _span, _bit_map, 6722 _work_queue, 6723 _overflow_stack, 6724 _finger, 6725 gfa, this); 6726 bool res = _work_queue->push(obj); // overflow could occur here 6727 assert(res, "Will hold once we use workqueues"); 6728 while (true) { 6729 oop new_oop; 6730 if (!_work_queue->pop_local(new_oop)) { 6731 // We emptied our work_queue; check if there's stuff that can 6732 // be gotten from the overflow stack. 6733 if (CMSConcMarkingTask::get_work_from_overflow_stack( 6734 _overflow_stack, _work_queue)) { 6735 do_yield_check(); 6736 continue; 6737 } else { // done 6738 break; 6739 } 6740 } 6741 // Skip verifying header mark word below because we are 6742 // running concurrent with mutators. 6743 assert(new_oop->is_oop(true), "Oops! expected to pop an oop"); 6744 // now scan this oop's oops 6745 new_oop->oop_iterate(&pushOrMarkClosure); 6746 do_yield_check(); 6747 } 6748 assert(_work_queue->size() == 0, "tautology, emphasizing post-condition"); 6749 } 6750 6751 // Yield in response to a request from VM Thread or 6752 // from mutators. 6753 void Par_MarkFromRootsClosure::do_yield_work() { 6754 assert(_task != NULL, "sanity"); 6755 _task->yield(); 6756 } 6757 6758 // A variant of the above used for verifying CMS marking work. 6759 MarkFromRootsVerifyClosure::MarkFromRootsVerifyClosure(CMSCollector* collector, 6760 MemRegion span, 6761 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 6762 CMSMarkStack* mark_stack): 6763 _collector(collector), 6764 _span(span), 6765 _verification_bm(verification_bm), 6766 _cms_bm(cms_bm), 6767 _mark_stack(mark_stack), 6768 _pam_verify_closure(collector, span, verification_bm, cms_bm, 6769 mark_stack) 6770 { 6771 assert(_mark_stack->isEmpty(), "stack should be empty"); 6772 _finger = _verification_bm->startWord(); 6773 assert(_collector->_restart_addr == NULL, "Sanity check"); 6774 assert(_span.contains(_finger), "Out of bounds _finger?"); 6775 } 6776 6777 void MarkFromRootsVerifyClosure::reset(HeapWord* addr) { 6778 assert(_mark_stack->isEmpty(), "would cause duplicates on stack"); 6779 assert(_span.contains(addr), "Out of bounds _finger?"); 6780 _finger = addr; 6781 } 6782 6783 // Should revisit to see if this should be restructured for 6784 // greater efficiency. 6785 bool MarkFromRootsVerifyClosure::do_bit(size_t offset) { 6786 // convert offset into a HeapWord* 6787 HeapWord* addr = _verification_bm->startWord() + offset; 6788 assert(_verification_bm->endWord() && addr < _verification_bm->endWord(), 6789 "address out of range"); 6790 assert(_verification_bm->isMarked(addr), "tautology"); 6791 assert(_cms_bm->isMarked(addr), "tautology"); 6792 6793 assert(_mark_stack->isEmpty(), 6794 "should drain stack to limit stack usage"); 6795 // convert addr to an oop preparatory to scanning 6796 oop obj = oop(addr); 6797 assert(obj->is_oop(), "should be an oop"); 6798 assert(_finger <= addr, "_finger runneth ahead"); 6799 // advance the finger to right end of this object 6800 _finger = addr + obj->size(); 6801 assert(_finger > addr, "we just incremented it above"); 6802 // Note: the finger doesn't advance while we drain 6803 // the stack below. 6804 bool res = _mark_stack->push(obj); 6805 assert(res, "Empty non-zero size stack should have space for single push"); 6806 while (!_mark_stack->isEmpty()) { 6807 oop new_oop = _mark_stack->pop(); 6808 assert(new_oop->is_oop(), "Oops! expected to pop an oop"); 6809 // now scan this oop's oops 6810 new_oop->oop_iterate(&_pam_verify_closure); 6811 } 6812 assert(_mark_stack->isEmpty(), "tautology, emphasizing post-condition"); 6813 return true; 6814 } 6815 6816 PushAndMarkVerifyClosure::PushAndMarkVerifyClosure( 6817 CMSCollector* collector, MemRegion span, 6818 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 6819 CMSMarkStack* mark_stack): 6820 MetadataAwareOopClosure(collector->ref_processor()), 6821 _collector(collector), 6822 _span(span), 6823 _verification_bm(verification_bm), 6824 _cms_bm(cms_bm), 6825 _mark_stack(mark_stack) 6826 { } 6827 6828 void PushAndMarkVerifyClosure::do_oop(oop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 6829 void PushAndMarkVerifyClosure::do_oop(narrowOop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 6830 6831 // Upon stack overflow, we discard (part of) the stack, 6832 // remembering the least address amongst those discarded 6833 // in CMSCollector's _restart_address. 6834 void PushAndMarkVerifyClosure::handle_stack_overflow(HeapWord* lost) { 6835 // Remember the least grey address discarded 6836 HeapWord* ra = (HeapWord*)_mark_stack->least_value(lost); 6837 _collector->lower_restart_addr(ra); 6838 _mark_stack->reset(); // discard stack contents 6839 _mark_stack->expand(); // expand the stack if possible 6840 } 6841 6842 void PushAndMarkVerifyClosure::do_oop(oop obj) { 6843 assert(obj->is_oop_or_null(), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6844 HeapWord* addr = (HeapWord*)obj; 6845 if (_span.contains(addr) && !_verification_bm->isMarked(addr)) { 6846 // Oop lies in _span and isn't yet grey or black 6847 _verification_bm->mark(addr); // now grey 6848 if (!_cms_bm->isMarked(addr)) { 6849 oop(addr)->print(); 6850 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)", 6851 p2i(addr)); 6852 fatal("... aborting"); 6853 } 6854 6855 if (!_mark_stack->push(obj)) { // stack overflow 6856 if (PrintCMSStatistics != 0) { 6857 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 6858 SIZE_FORMAT, _mark_stack->capacity()); 6859 } 6860 assert(_mark_stack->isFull(), "Else push should have succeeded"); 6861 handle_stack_overflow(addr); 6862 } 6863 // anything including and to the right of _finger 6864 // will be scanned as we iterate over the remainder of the 6865 // bit map 6866 } 6867 } 6868 6869 PushOrMarkClosure::PushOrMarkClosure(CMSCollector* collector, 6870 MemRegion span, 6871 CMSBitMap* bitMap, CMSMarkStack* markStack, 6872 HeapWord* finger, MarkFromRootsClosure* parent) : 6873 MetadataAwareOopClosure(collector->ref_processor()), 6874 _collector(collector), 6875 _span(span), 6876 _bitMap(bitMap), 6877 _markStack(markStack), 6878 _finger(finger), 6879 _parent(parent) 6880 { } 6881 6882 Par_PushOrMarkClosure::Par_PushOrMarkClosure(CMSCollector* collector, 6883 MemRegion span, 6884 CMSBitMap* bit_map, 6885 OopTaskQueue* work_queue, 6886 CMSMarkStack* overflow_stack, 6887 HeapWord* finger, 6888 HeapWord** global_finger_addr, 6889 Par_MarkFromRootsClosure* parent) : 6890 MetadataAwareOopClosure(collector->ref_processor()), 6891 _collector(collector), 6892 _whole_span(collector->_span), 6893 _span(span), 6894 _bit_map(bit_map), 6895 _work_queue(work_queue), 6896 _overflow_stack(overflow_stack), 6897 _finger(finger), 6898 _global_finger_addr(global_finger_addr), 6899 _parent(parent) 6900 { } 6901 6902 // Assumes thread-safe access by callers, who are 6903 // responsible for mutual exclusion. 6904 void CMSCollector::lower_restart_addr(HeapWord* low) { 6905 assert(_span.contains(low), "Out of bounds addr"); 6906 if (_restart_addr == NULL) { 6907 _restart_addr = low; 6908 } else { 6909 _restart_addr = MIN2(_restart_addr, low); 6910 } 6911 } 6912 6913 // Upon stack overflow, we discard (part of) the stack, 6914 // remembering the least address amongst those discarded 6915 // in CMSCollector's _restart_address. 6916 void PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 6917 // Remember the least grey address discarded 6918 HeapWord* ra = (HeapWord*)_markStack->least_value(lost); 6919 _collector->lower_restart_addr(ra); 6920 _markStack->reset(); // discard stack contents 6921 _markStack->expand(); // expand the stack if possible 6922 } 6923 6924 // Upon stack overflow, we discard (part of) the stack, 6925 // remembering the least address amongst those discarded 6926 // in CMSCollector's _restart_address. 6927 void Par_PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 6928 // We need to do this under a mutex to prevent other 6929 // workers from interfering with the work done below. 6930 MutexLockerEx ml(_overflow_stack->par_lock(), 6931 Mutex::_no_safepoint_check_flag); 6932 // Remember the least grey address discarded 6933 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 6934 _collector->lower_restart_addr(ra); 6935 _overflow_stack->reset(); // discard stack contents 6936 _overflow_stack->expand(); // expand the stack if possible 6937 } 6938 6939 void PushOrMarkClosure::do_oop(oop obj) { 6940 // Ignore mark word because we are running concurrent with mutators. 6941 assert(obj->is_oop_or_null(true), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6942 HeapWord* addr = (HeapWord*)obj; 6943 if (_span.contains(addr) && !_bitMap->isMarked(addr)) { 6944 // Oop lies in _span and isn't yet grey or black 6945 _bitMap->mark(addr); // now grey 6946 if (addr < _finger) { 6947 // the bit map iteration has already either passed, or 6948 // sampled, this bit in the bit map; we'll need to 6949 // use the marking stack to scan this oop's oops. 6950 bool simulate_overflow = false; 6951 NOT_PRODUCT( 6952 if (CMSMarkStackOverflowALot && 6953 _collector->simulate_overflow()) { 6954 // simulate a stack overflow 6955 simulate_overflow = true; 6956 } 6957 ) 6958 if (simulate_overflow || !_markStack->push(obj)) { // stack overflow 6959 if (PrintCMSStatistics != 0) { 6960 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 6961 SIZE_FORMAT, _markStack->capacity()); 6962 } 6963 assert(simulate_overflow || _markStack->isFull(), "Else push should have succeeded"); 6964 handle_stack_overflow(addr); 6965 } 6966 } 6967 // anything including and to the right of _finger 6968 // will be scanned as we iterate over the remainder of the 6969 // bit map 6970 do_yield_check(); 6971 } 6972 } 6973 6974 void PushOrMarkClosure::do_oop(oop* p) { PushOrMarkClosure::do_oop_work(p); } 6975 void PushOrMarkClosure::do_oop(narrowOop* p) { PushOrMarkClosure::do_oop_work(p); } 6976 6977 void Par_PushOrMarkClosure::do_oop(oop obj) { 6978 // Ignore mark word because we are running concurrent with mutators. 6979 assert(obj->is_oop_or_null(true), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6980 HeapWord* addr = (HeapWord*)obj; 6981 if (_whole_span.contains(addr) && !_bit_map->isMarked(addr)) { 6982 // Oop lies in _span and isn't yet grey or black 6983 // We read the global_finger (volatile read) strictly after marking oop 6984 bool res = _bit_map->par_mark(addr); // now grey 6985 volatile HeapWord** gfa = (volatile HeapWord**)_global_finger_addr; 6986 // Should we push this marked oop on our stack? 6987 // -- if someone else marked it, nothing to do 6988 // -- if target oop is above global finger nothing to do 6989 // -- if target oop is in chunk and above local finger 6990 // then nothing to do 6991 // -- else push on work queue 6992 if ( !res // someone else marked it, they will deal with it 6993 || (addr >= *gfa) // will be scanned in a later task 6994 || (_span.contains(addr) && addr >= _finger)) { // later in this chunk 6995 return; 6996 } 6997 // the bit map iteration has already either passed, or 6998 // sampled, this bit in the bit map; we'll need to 6999 // use the marking stack to scan this oop's oops. 7000 bool simulate_overflow = false; 7001 NOT_PRODUCT( 7002 if (CMSMarkStackOverflowALot && 7003 _collector->simulate_overflow()) { 7004 // simulate a stack overflow 7005 simulate_overflow = true; 7006 } 7007 ) 7008 if (simulate_overflow || 7009 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) { 7010 // stack overflow 7011 if (PrintCMSStatistics != 0) { 7012 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 7013 SIZE_FORMAT, _overflow_stack->capacity()); 7014 } 7015 // We cannot assert that the overflow stack is full because 7016 // it may have been emptied since. 7017 assert(simulate_overflow || 7018 _work_queue->size() == _work_queue->max_elems(), 7019 "Else push should have succeeded"); 7020 handle_stack_overflow(addr); 7021 } 7022 do_yield_check(); 7023 } 7024 } 7025 7026 void Par_PushOrMarkClosure::do_oop(oop* p) { Par_PushOrMarkClosure::do_oop_work(p); } 7027 void Par_PushOrMarkClosure::do_oop(narrowOop* p) { Par_PushOrMarkClosure::do_oop_work(p); } 7028 7029 PushAndMarkClosure::PushAndMarkClosure(CMSCollector* collector, 7030 MemRegion span, 7031 ReferenceProcessor* rp, 7032 CMSBitMap* bit_map, 7033 CMSBitMap* mod_union_table, 7034 CMSMarkStack* mark_stack, 7035 bool concurrent_precleaning): 7036 MetadataAwareOopClosure(rp), 7037 _collector(collector), 7038 _span(span), 7039 _bit_map(bit_map), 7040 _mod_union_table(mod_union_table), 7041 _mark_stack(mark_stack), 7042 _concurrent_precleaning(concurrent_precleaning) 7043 { 7044 assert(ref_processor() != NULL, "ref_processor shouldn't be NULL"); 7045 } 7046 7047 // Grey object rescan during pre-cleaning and second checkpoint phases -- 7048 // the non-parallel version (the parallel version appears further below.) 7049 void PushAndMarkClosure::do_oop(oop obj) { 7050 // Ignore mark word verification. If during concurrent precleaning, 7051 // the object monitor may be locked. If during the checkpoint 7052 // phases, the object may already have been reached by a different 7053 // path and may be at the end of the global overflow list (so 7054 // the mark word may be NULL). 7055 assert(obj->is_oop_or_null(true /* ignore mark word */), 7056 "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 7057 HeapWord* addr = (HeapWord*)obj; 7058 // Check if oop points into the CMS generation 7059 // and is not marked 7060 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 7061 // a white object ... 7062 _bit_map->mark(addr); // ... now grey 7063 // push on the marking stack (grey set) 7064 bool simulate_overflow = false; 7065 NOT_PRODUCT( 7066 if (CMSMarkStackOverflowALot && 7067 _collector->simulate_overflow()) { 7068 // simulate a stack overflow 7069 simulate_overflow = true; 7070 } 7071 ) 7072 if (simulate_overflow || !_mark_stack->push(obj)) { 7073 if (_concurrent_precleaning) { 7074 // During precleaning we can just dirty the appropriate card(s) 7075 // in the mod union table, thus ensuring that the object remains 7076 // in the grey set and continue. In the case of object arrays 7077 // we need to dirty all of the cards that the object spans, 7078 // since the rescan of object arrays will be limited to the 7079 // dirty cards. 7080 // Note that no one can be interfering with us in this action 7081 // of dirtying the mod union table, so no locking or atomics 7082 // are required. 7083 if (obj->is_objArray()) { 7084 size_t sz = obj->size(); 7085 HeapWord* end_card_addr = (HeapWord*)round_to( 7086 (intptr_t)(addr+sz), CardTableModRefBS::card_size); 7087 MemRegion redirty_range = MemRegion(addr, end_card_addr); 7088 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 7089 _mod_union_table->mark_range(redirty_range); 7090 } else { 7091 _mod_union_table->mark(addr); 7092 } 7093 _collector->_ser_pmc_preclean_ovflw++; 7094 } else { 7095 // During the remark phase, we need to remember this oop 7096 // in the overflow list. 7097 _collector->push_on_overflow_list(obj); 7098 _collector->_ser_pmc_remark_ovflw++; 7099 } 7100 } 7101 } 7102 } 7103 7104 Par_PushAndMarkClosure::Par_PushAndMarkClosure(CMSCollector* collector, 7105 MemRegion span, 7106 ReferenceProcessor* rp, 7107 CMSBitMap* bit_map, 7108 OopTaskQueue* work_queue): 7109 MetadataAwareOopClosure(rp), 7110 _collector(collector), 7111 _span(span), 7112 _bit_map(bit_map), 7113 _work_queue(work_queue) 7114 { 7115 assert(ref_processor() != NULL, "ref_processor shouldn't be NULL"); 7116 } 7117 7118 void PushAndMarkClosure::do_oop(oop* p) { PushAndMarkClosure::do_oop_work(p); } 7119 void PushAndMarkClosure::do_oop(narrowOop* p) { PushAndMarkClosure::do_oop_work(p); } 7120 7121 // Grey object rescan during second checkpoint phase -- 7122 // the parallel version. 7123 void Par_PushAndMarkClosure::do_oop(oop obj) { 7124 // In the assert below, we ignore the mark word because 7125 // this oop may point to an already visited object that is 7126 // on the overflow stack (in which case the mark word has 7127 // been hijacked for chaining into the overflow stack -- 7128 // if this is the last object in the overflow stack then 7129 // its mark word will be NULL). Because this object may 7130 // have been subsequently popped off the global overflow 7131 // stack, and the mark word possibly restored to the prototypical 7132 // value, by the time we get to examined this failing assert in 7133 // the debugger, is_oop_or_null(false) may subsequently start 7134 // to hold. 7135 assert(obj->is_oop_or_null(true), 7136 "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 7137 HeapWord* addr = (HeapWord*)obj; 7138 // Check if oop points into the CMS generation 7139 // and is not marked 7140 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 7141 // a white object ... 7142 // If we manage to "claim" the object, by being the 7143 // first thread to mark it, then we push it on our 7144 // marking stack 7145 if (_bit_map->par_mark(addr)) { // ... now grey 7146 // push on work queue (grey set) 7147 bool simulate_overflow = false; 7148 NOT_PRODUCT( 7149 if (CMSMarkStackOverflowALot && 7150 _collector->par_simulate_overflow()) { 7151 // simulate a stack overflow 7152 simulate_overflow = true; 7153 } 7154 ) 7155 if (simulate_overflow || !_work_queue->push(obj)) { 7156 _collector->par_push_on_overflow_list(obj); 7157 _collector->_par_pmc_remark_ovflw++; // imprecise OK: no need to CAS 7158 } 7159 } // Else, some other thread got there first 7160 } 7161 } 7162 7163 void Par_PushAndMarkClosure::do_oop(oop* p) { Par_PushAndMarkClosure::do_oop_work(p); } 7164 void Par_PushAndMarkClosure::do_oop(narrowOop* p) { Par_PushAndMarkClosure::do_oop_work(p); } 7165 7166 void CMSPrecleanRefsYieldClosure::do_yield_work() { 7167 Mutex* bml = _collector->bitMapLock(); 7168 assert_lock_strong(bml); 7169 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 7170 "CMS thread should hold CMS token"); 7171 7172 bml->unlock(); 7173 ConcurrentMarkSweepThread::desynchronize(true); 7174 7175 _collector->stopTimer(); 7176 if (PrintCMSStatistics != 0) { 7177 _collector->incrementYields(); 7178 } 7179 7180 // See the comment in coordinator_yield() 7181 for (unsigned i = 0; i < CMSYieldSleepCount && 7182 ConcurrentMarkSweepThread::should_yield() && 7183 !CMSCollector::foregroundGCIsActive(); ++i) { 7184 os::sleep(Thread::current(), 1, false); 7185 } 7186 7187 ConcurrentMarkSweepThread::synchronize(true); 7188 bml->lock(); 7189 7190 _collector->startTimer(); 7191 } 7192 7193 bool CMSPrecleanRefsYieldClosure::should_return() { 7194 if (ConcurrentMarkSweepThread::should_yield()) { 7195 do_yield_work(); 7196 } 7197 return _collector->foregroundGCIsActive(); 7198 } 7199 7200 void MarkFromDirtyCardsClosure::do_MemRegion(MemRegion mr) { 7201 assert(((size_t)mr.start())%CardTableModRefBS::card_size_in_words == 0, 7202 "mr should be aligned to start at a card boundary"); 7203 // We'd like to assert: 7204 // assert(mr.word_size()%CardTableModRefBS::card_size_in_words == 0, 7205 // "mr should be a range of cards"); 7206 // However, that would be too strong in one case -- the last 7207 // partition ends at _unallocated_block which, in general, can be 7208 // an arbitrary boundary, not necessarily card aligned. 7209 if (PrintCMSStatistics != 0) { 7210 _num_dirty_cards += 7211 mr.word_size()/CardTableModRefBS::card_size_in_words; 7212 } 7213 _space->object_iterate_mem(mr, &_scan_cl); 7214 } 7215 7216 SweepClosure::SweepClosure(CMSCollector* collector, 7217 ConcurrentMarkSweepGeneration* g, 7218 CMSBitMap* bitMap, bool should_yield) : 7219 _collector(collector), 7220 _g(g), 7221 _sp(g->cmsSpace()), 7222 _limit(_sp->sweep_limit()), 7223 _freelistLock(_sp->freelistLock()), 7224 _bitMap(bitMap), 7225 _yield(should_yield), 7226 _inFreeRange(false), // No free range at beginning of sweep 7227 _freeRangeInFreeLists(false), // No free range at beginning of sweep 7228 _lastFreeRangeCoalesced(false), 7229 _freeFinger(g->used_region().start()) 7230 { 7231 NOT_PRODUCT( 7232 _numObjectsFreed = 0; 7233 _numWordsFreed = 0; 7234 _numObjectsLive = 0; 7235 _numWordsLive = 0; 7236 _numObjectsAlreadyFree = 0; 7237 _numWordsAlreadyFree = 0; 7238 _last_fc = NULL; 7239 7240 _sp->initializeIndexedFreeListArrayReturnedBytes(); 7241 _sp->dictionary()->initialize_dict_returned_bytes(); 7242 ) 7243 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7244 "sweep _limit out of bounds"); 7245 if (CMSTraceSweeper) { 7246 gclog_or_tty->print_cr("\n====================\nStarting new sweep with limit " PTR_FORMAT, 7247 p2i(_limit)); 7248 } 7249 } 7250 7251 void SweepClosure::print_on(outputStream* st) const { 7252 tty->print_cr("_sp = [" PTR_FORMAT "," PTR_FORMAT ")", 7253 p2i(_sp->bottom()), p2i(_sp->end())); 7254 tty->print_cr("_limit = " PTR_FORMAT, p2i(_limit)); 7255 tty->print_cr("_freeFinger = " PTR_FORMAT, p2i(_freeFinger)); 7256 NOT_PRODUCT(tty->print_cr("_last_fc = " PTR_FORMAT, p2i(_last_fc));) 7257 tty->print_cr("_inFreeRange = %d, _freeRangeInFreeLists = %d, _lastFreeRangeCoalesced = %d", 7258 _inFreeRange, _freeRangeInFreeLists, _lastFreeRangeCoalesced); 7259 } 7260 7261 #ifndef PRODUCT 7262 // Assertion checking only: no useful work in product mode -- 7263 // however, if any of the flags below become product flags, 7264 // you may need to review this code to see if it needs to be 7265 // enabled in product mode. 7266 SweepClosure::~SweepClosure() { 7267 assert_lock_strong(_freelistLock); 7268 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7269 "sweep _limit out of bounds"); 7270 if (inFreeRange()) { 7271 warning("inFreeRange() should have been reset; dumping state of SweepClosure"); 7272 print(); 7273 ShouldNotReachHere(); 7274 } 7275 if (Verbose && PrintGC) { 7276 gclog_or_tty->print("Collected " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes", 7277 _numObjectsFreed, _numWordsFreed*sizeof(HeapWord)); 7278 gclog_or_tty->print_cr("\nLive " SIZE_FORMAT " objects, " 7279 SIZE_FORMAT " bytes " 7280 "Already free " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes", 7281 _numObjectsLive, _numWordsLive*sizeof(HeapWord), 7282 _numObjectsAlreadyFree, _numWordsAlreadyFree*sizeof(HeapWord)); 7283 size_t totalBytes = (_numWordsFreed + _numWordsLive + _numWordsAlreadyFree) 7284 * sizeof(HeapWord); 7285 gclog_or_tty->print_cr("Total sweep: " SIZE_FORMAT " bytes", totalBytes); 7286 7287 if (PrintCMSStatistics && CMSVerifyReturnedBytes) { 7288 size_t indexListReturnedBytes = _sp->sumIndexedFreeListArrayReturnedBytes(); 7289 size_t dict_returned_bytes = _sp->dictionary()->sum_dict_returned_bytes(); 7290 size_t returned_bytes = indexListReturnedBytes + dict_returned_bytes; 7291 gclog_or_tty->print("Returned " SIZE_FORMAT " bytes", returned_bytes); 7292 gclog_or_tty->print(" Indexed List Returned " SIZE_FORMAT " bytes", 7293 indexListReturnedBytes); 7294 gclog_or_tty->print_cr(" Dictionary Returned " SIZE_FORMAT " bytes", 7295 dict_returned_bytes); 7296 } 7297 } 7298 if (CMSTraceSweeper) { 7299 gclog_or_tty->print_cr("end of sweep with _limit = " PTR_FORMAT "\n================", 7300 p2i(_limit)); 7301 } 7302 } 7303 #endif // PRODUCT 7304 7305 void SweepClosure::initialize_free_range(HeapWord* freeFinger, 7306 bool freeRangeInFreeLists) { 7307 if (CMSTraceSweeper) { 7308 gclog_or_tty->print("---- Start free range at " PTR_FORMAT " with free block (%d)\n", 7309 p2i(freeFinger), freeRangeInFreeLists); 7310 } 7311 assert(!inFreeRange(), "Trampling existing free range"); 7312 set_inFreeRange(true); 7313 set_lastFreeRangeCoalesced(false); 7314 7315 set_freeFinger(freeFinger); 7316 set_freeRangeInFreeLists(freeRangeInFreeLists); 7317 if (CMSTestInFreeList) { 7318 if (freeRangeInFreeLists) { 7319 FreeChunk* fc = (FreeChunk*) freeFinger; 7320 assert(fc->is_free(), "A chunk on the free list should be free."); 7321 assert(fc->size() > 0, "Free range should have a size"); 7322 assert(_sp->verify_chunk_in_free_list(fc), "Chunk is not in free lists"); 7323 } 7324 } 7325 } 7326 7327 // Note that the sweeper runs concurrently with mutators. Thus, 7328 // it is possible for direct allocation in this generation to happen 7329 // in the middle of the sweep. Note that the sweeper also coalesces 7330 // contiguous free blocks. Thus, unless the sweeper and the allocator 7331 // synchronize appropriately freshly allocated blocks may get swept up. 7332 // This is accomplished by the sweeper locking the free lists while 7333 // it is sweeping. Thus blocks that are determined to be free are 7334 // indeed free. There is however one additional complication: 7335 // blocks that have been allocated since the final checkpoint and 7336 // mark, will not have been marked and so would be treated as 7337 // unreachable and swept up. To prevent this, the allocator marks 7338 // the bit map when allocating during the sweep phase. This leads, 7339 // however, to a further complication -- objects may have been allocated 7340 // but not yet initialized -- in the sense that the header isn't yet 7341 // installed. The sweeper can not then determine the size of the block 7342 // in order to skip over it. To deal with this case, we use a technique 7343 // (due to Printezis) to encode such uninitialized block sizes in the 7344 // bit map. Since the bit map uses a bit per every HeapWord, but the 7345 // CMS generation has a minimum object size of 3 HeapWords, it follows 7346 // that "normal marks" won't be adjacent in the bit map (there will 7347 // always be at least two 0 bits between successive 1 bits). We make use 7348 // of these "unused" bits to represent uninitialized blocks -- the bit 7349 // corresponding to the start of the uninitialized object and the next 7350 // bit are both set. Finally, a 1 bit marks the end of the object that 7351 // started with the two consecutive 1 bits to indicate its potentially 7352 // uninitialized state. 7353 7354 size_t SweepClosure::do_blk_careful(HeapWord* addr) { 7355 FreeChunk* fc = (FreeChunk*)addr; 7356 size_t res; 7357 7358 // Check if we are done sweeping. Below we check "addr >= _limit" rather 7359 // than "addr == _limit" because although _limit was a block boundary when 7360 // we started the sweep, it may no longer be one because heap expansion 7361 // may have caused us to coalesce the block ending at the address _limit 7362 // with a newly expanded chunk (this happens when _limit was set to the 7363 // previous _end of the space), so we may have stepped past _limit: 7364 // see the following Zeno-like trail of CRs 6977970, 7008136, 7042740. 7365 if (addr >= _limit) { // we have swept up to or past the limit: finish up 7366 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7367 "sweep _limit out of bounds"); 7368 assert(addr < _sp->end(), "addr out of bounds"); 7369 // Flush any free range we might be holding as a single 7370 // coalesced chunk to the appropriate free list. 7371 if (inFreeRange()) { 7372 assert(freeFinger() >= _sp->bottom() && freeFinger() < _limit, 7373 "freeFinger() " PTR_FORMAT " is out-of-bounds", p2i(freeFinger())); 7374 flush_cur_free_chunk(freeFinger(), 7375 pointer_delta(addr, freeFinger())); 7376 if (CMSTraceSweeper) { 7377 gclog_or_tty->print("Sweep: last chunk: "); 7378 gclog_or_tty->print("put_free_blk " PTR_FORMAT " (" SIZE_FORMAT ") " 7379 "[coalesced:%d]\n", 7380 p2i(freeFinger()), pointer_delta(addr, freeFinger()), 7381 lastFreeRangeCoalesced() ? 1 : 0); 7382 } 7383 } 7384 7385 // help the iterator loop finish 7386 return pointer_delta(_sp->end(), addr); 7387 } 7388 7389 assert(addr < _limit, "sweep invariant"); 7390 // check if we should yield 7391 do_yield_check(addr); 7392 if (fc->is_free()) { 7393 // Chunk that is already free 7394 res = fc->size(); 7395 do_already_free_chunk(fc); 7396 debug_only(_sp->verifyFreeLists()); 7397 // If we flush the chunk at hand in lookahead_and_flush() 7398 // and it's coalesced with a preceding chunk, then the 7399 // process of "mangling" the payload of the coalesced block 7400 // will cause erasure of the size information from the 7401 // (erstwhile) header of all the coalesced blocks but the 7402 // first, so the first disjunct in the assert will not hold 7403 // in that specific case (in which case the second disjunct 7404 // will hold). 7405 assert(res == fc->size() || ((HeapWord*)fc) + res >= _limit, 7406 "Otherwise the size info doesn't change at this step"); 7407 NOT_PRODUCT( 7408 _numObjectsAlreadyFree++; 7409 _numWordsAlreadyFree += res; 7410 ) 7411 NOT_PRODUCT(_last_fc = fc;) 7412 } else if (!_bitMap->isMarked(addr)) { 7413 // Chunk is fresh garbage 7414 res = do_garbage_chunk(fc); 7415 debug_only(_sp->verifyFreeLists()); 7416 NOT_PRODUCT( 7417 _numObjectsFreed++; 7418 _numWordsFreed += res; 7419 ) 7420 } else { 7421 // Chunk that is alive. 7422 res = do_live_chunk(fc); 7423 debug_only(_sp->verifyFreeLists()); 7424 NOT_PRODUCT( 7425 _numObjectsLive++; 7426 _numWordsLive += res; 7427 ) 7428 } 7429 return res; 7430 } 7431 7432 // For the smart allocation, record following 7433 // split deaths - a free chunk is removed from its free list because 7434 // it is being split into two or more chunks. 7435 // split birth - a free chunk is being added to its free list because 7436 // a larger free chunk has been split and resulted in this free chunk. 7437 // coal death - a free chunk is being removed from its free list because 7438 // it is being coalesced into a large free chunk. 7439 // coal birth - a free chunk is being added to its free list because 7440 // it was created when two or more free chunks where coalesced into 7441 // this free chunk. 7442 // 7443 // These statistics are used to determine the desired number of free 7444 // chunks of a given size. The desired number is chosen to be relative 7445 // to the end of a CMS sweep. The desired number at the end of a sweep 7446 // is the 7447 // count-at-end-of-previous-sweep (an amount that was enough) 7448 // - count-at-beginning-of-current-sweep (the excess) 7449 // + split-births (gains in this size during interval) 7450 // - split-deaths (demands on this size during interval) 7451 // where the interval is from the end of one sweep to the end of the 7452 // next. 7453 // 7454 // When sweeping the sweeper maintains an accumulated chunk which is 7455 // the chunk that is made up of chunks that have been coalesced. That 7456 // will be termed the left-hand chunk. A new chunk of garbage that 7457 // is being considered for coalescing will be referred to as the 7458 // right-hand chunk. 7459 // 7460 // When making a decision on whether to coalesce a right-hand chunk with 7461 // the current left-hand chunk, the current count vs. the desired count 7462 // of the left-hand chunk is considered. Also if the right-hand chunk 7463 // is near the large chunk at the end of the heap (see 7464 // ConcurrentMarkSweepGeneration::isNearLargestChunk()), then the 7465 // left-hand chunk is coalesced. 7466 // 7467 // When making a decision about whether to split a chunk, the desired count 7468 // vs. the current count of the candidate to be split is also considered. 7469 // If the candidate is underpopulated (currently fewer chunks than desired) 7470 // a chunk of an overpopulated (currently more chunks than desired) size may 7471 // be chosen. The "hint" associated with a free list, if non-null, points 7472 // to a free list which may be overpopulated. 7473 // 7474 7475 void SweepClosure::do_already_free_chunk(FreeChunk* fc) { 7476 const size_t size = fc->size(); 7477 // Chunks that cannot be coalesced are not in the 7478 // free lists. 7479 if (CMSTestInFreeList && !fc->cantCoalesce()) { 7480 assert(_sp->verify_chunk_in_free_list(fc), 7481 "free chunk should be in free lists"); 7482 } 7483 // a chunk that is already free, should not have been 7484 // marked in the bit map 7485 HeapWord* const addr = (HeapWord*) fc; 7486 assert(!_bitMap->isMarked(addr), "free chunk should be unmarked"); 7487 // Verify that the bit map has no bits marked between 7488 // addr and purported end of this block. 7489 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 7490 7491 // Some chunks cannot be coalesced under any circumstances. 7492 // See the definition of cantCoalesce(). 7493 if (!fc->cantCoalesce()) { 7494 // This chunk can potentially be coalesced. 7495 // All the work is done in 7496 do_post_free_or_garbage_chunk(fc, size); 7497 // Note that if the chunk is not coalescable (the else arm 7498 // below), we unconditionally flush, without needing to do 7499 // a "lookahead," as we do below. 7500 if (inFreeRange()) lookahead_and_flush(fc, size); 7501 } else { 7502 // Code path common to both original and adaptive free lists. 7503 7504 // cant coalesce with previous block; this should be treated 7505 // as the end of a free run if any 7506 if (inFreeRange()) { 7507 // we kicked some butt; time to pick up the garbage 7508 assert(freeFinger() < addr, "freeFinger points too high"); 7509 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 7510 } 7511 // else, nothing to do, just continue 7512 } 7513 } 7514 7515 size_t SweepClosure::do_garbage_chunk(FreeChunk* fc) { 7516 // This is a chunk of garbage. It is not in any free list. 7517 // Add it to a free list or let it possibly be coalesced into 7518 // a larger chunk. 7519 HeapWord* const addr = (HeapWord*) fc; 7520 const size_t size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 7521 7522 // Verify that the bit map has no bits marked between 7523 // addr and purported end of just dead object. 7524 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 7525 do_post_free_or_garbage_chunk(fc, size); 7526 7527 assert(_limit >= addr + size, 7528 "A freshly garbage chunk can't possibly straddle over _limit"); 7529 if (inFreeRange()) lookahead_and_flush(fc, size); 7530 return size; 7531 } 7532 7533 size_t SweepClosure::do_live_chunk(FreeChunk* fc) { 7534 HeapWord* addr = (HeapWord*) fc; 7535 // The sweeper has just found a live object. Return any accumulated 7536 // left hand chunk to the free lists. 7537 if (inFreeRange()) { 7538 assert(freeFinger() < addr, "freeFinger points too high"); 7539 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 7540 } 7541 7542 // This object is live: we'd normally expect this to be 7543 // an oop, and like to assert the following: 7544 // assert(oop(addr)->is_oop(), "live block should be an oop"); 7545 // However, as we commented above, this may be an object whose 7546 // header hasn't yet been initialized. 7547 size_t size; 7548 assert(_bitMap->isMarked(addr), "Tautology for this control point"); 7549 if (_bitMap->isMarked(addr + 1)) { 7550 // Determine the size from the bit map, rather than trying to 7551 // compute it from the object header. 7552 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 7553 size = pointer_delta(nextOneAddr + 1, addr); 7554 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 7555 "alignment problem"); 7556 7557 #ifdef ASSERT 7558 if (oop(addr)->klass_or_null() != NULL) { 7559 // Ignore mark word because we are running concurrent with mutators 7560 assert(oop(addr)->is_oop(true), "live block should be an oop"); 7561 assert(size == 7562 CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()), 7563 "P-mark and computed size do not agree"); 7564 } 7565 #endif 7566 7567 } else { 7568 // This should be an initialized object that's alive. 7569 assert(oop(addr)->klass_or_null() != NULL, 7570 "Should be an initialized object"); 7571 // Ignore mark word because we are running concurrent with mutators 7572 assert(oop(addr)->is_oop(true), "live block should be an oop"); 7573 // Verify that the bit map has no bits marked between 7574 // addr and purported end of this block. 7575 size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 7576 assert(size >= 3, "Necessary for Printezis marks to work"); 7577 assert(!_bitMap->isMarked(addr+1), "Tautology for this control point"); 7578 DEBUG_ONLY(_bitMap->verifyNoOneBitsInRange(addr+2, addr+size);) 7579 } 7580 return size; 7581 } 7582 7583 void SweepClosure::do_post_free_or_garbage_chunk(FreeChunk* fc, 7584 size_t chunkSize) { 7585 // do_post_free_or_garbage_chunk() should only be called in the case 7586 // of the adaptive free list allocator. 7587 const bool fcInFreeLists = fc->is_free(); 7588 assert((HeapWord*)fc <= _limit, "sweep invariant"); 7589 if (CMSTestInFreeList && fcInFreeLists) { 7590 assert(_sp->verify_chunk_in_free_list(fc), "free chunk is not in free lists"); 7591 } 7592 7593 if (CMSTraceSweeper) { 7594 gclog_or_tty->print_cr(" -- pick up another chunk at " PTR_FORMAT " (" SIZE_FORMAT ")", p2i(fc), chunkSize); 7595 } 7596 7597 HeapWord* const fc_addr = (HeapWord*) fc; 7598 7599 bool coalesce = false; 7600 const size_t left = pointer_delta(fc_addr, freeFinger()); 7601 const size_t right = chunkSize; 7602 switch (FLSCoalescePolicy) { 7603 // numeric value forms a coalition aggressiveness metric 7604 case 0: { // never coalesce 7605 coalesce = false; 7606 break; 7607 } 7608 case 1: { // coalesce if left & right chunks on overpopulated lists 7609 coalesce = _sp->coalOverPopulated(left) && 7610 _sp->coalOverPopulated(right); 7611 break; 7612 } 7613 case 2: { // coalesce if left chunk on overpopulated list (default) 7614 coalesce = _sp->coalOverPopulated(left); 7615 break; 7616 } 7617 case 3: { // coalesce if left OR right chunk on overpopulated list 7618 coalesce = _sp->coalOverPopulated(left) || 7619 _sp->coalOverPopulated(right); 7620 break; 7621 } 7622 case 4: { // always coalesce 7623 coalesce = true; 7624 break; 7625 } 7626 default: 7627 ShouldNotReachHere(); 7628 } 7629 7630 // Should the current free range be coalesced? 7631 // If the chunk is in a free range and either we decided to coalesce above 7632 // or the chunk is near the large block at the end of the heap 7633 // (isNearLargestChunk() returns true), then coalesce this chunk. 7634 const bool doCoalesce = inFreeRange() 7635 && (coalesce || _g->isNearLargestChunk(fc_addr)); 7636 if (doCoalesce) { 7637 // Coalesce the current free range on the left with the new 7638 // chunk on the right. If either is on a free list, 7639 // it must be removed from the list and stashed in the closure. 7640 if (freeRangeInFreeLists()) { 7641 FreeChunk* const ffc = (FreeChunk*)freeFinger(); 7642 assert(ffc->size() == pointer_delta(fc_addr, freeFinger()), 7643 "Size of free range is inconsistent with chunk size."); 7644 if (CMSTestInFreeList) { 7645 assert(_sp->verify_chunk_in_free_list(ffc), 7646 "Chunk is not in free lists"); 7647 } 7648 _sp->coalDeath(ffc->size()); 7649 _sp->removeFreeChunkFromFreeLists(ffc); 7650 set_freeRangeInFreeLists(false); 7651 } 7652 if (fcInFreeLists) { 7653 _sp->coalDeath(chunkSize); 7654 assert(fc->size() == chunkSize, 7655 "The chunk has the wrong size or is not in the free lists"); 7656 _sp->removeFreeChunkFromFreeLists(fc); 7657 } 7658 set_lastFreeRangeCoalesced(true); 7659 print_free_block_coalesced(fc); 7660 } else { // not in a free range and/or should not coalesce 7661 // Return the current free range and start a new one. 7662 if (inFreeRange()) { 7663 // In a free range but cannot coalesce with the right hand chunk. 7664 // Put the current free range into the free lists. 7665 flush_cur_free_chunk(freeFinger(), 7666 pointer_delta(fc_addr, freeFinger())); 7667 } 7668 // Set up for new free range. Pass along whether the right hand 7669 // chunk is in the free lists. 7670 initialize_free_range((HeapWord*)fc, fcInFreeLists); 7671 } 7672 } 7673 7674 // Lookahead flush: 7675 // If we are tracking a free range, and this is the last chunk that 7676 // we'll look at because its end crosses past _limit, we'll preemptively 7677 // flush it along with any free range we may be holding on to. Note that 7678 // this can be the case only for an already free or freshly garbage 7679 // chunk. If this block is an object, it can never straddle 7680 // over _limit. The "straddling" occurs when _limit is set at 7681 // the previous end of the space when this cycle started, and 7682 // a subsequent heap expansion caused the previously co-terminal 7683 // free block to be coalesced with the newly expanded portion, 7684 // thus rendering _limit a non-block-boundary making it dangerous 7685 // for the sweeper to step over and examine. 7686 void SweepClosure::lookahead_and_flush(FreeChunk* fc, size_t chunk_size) { 7687 assert(inFreeRange(), "Should only be called if currently in a free range."); 7688 HeapWord* const eob = ((HeapWord*)fc) + chunk_size; 7689 assert(_sp->used_region().contains(eob - 1), 7690 "eob = " PTR_FORMAT " eob-1 = " PTR_FORMAT " _limit = " PTR_FORMAT 7691 " out of bounds wrt _sp = [" PTR_FORMAT "," PTR_FORMAT ")" 7692 " when examining fc = " PTR_FORMAT "(" SIZE_FORMAT ")", 7693 p2i(eob), p2i(eob-1), p2i(_limit), p2i(_sp->bottom()), p2i(_sp->end()), p2i(fc), chunk_size); 7694 if (eob >= _limit) { 7695 assert(eob == _limit || fc->is_free(), "Only a free chunk should allow us to cross over the limit"); 7696 if (CMSTraceSweeper) { 7697 gclog_or_tty->print_cr("_limit " PTR_FORMAT " reached or crossed by block " 7698 "[" PTR_FORMAT "," PTR_FORMAT ") in space " 7699 "[" PTR_FORMAT "," PTR_FORMAT ")", 7700 p2i(_limit), p2i(fc), p2i(eob), p2i(_sp->bottom()), p2i(_sp->end())); 7701 } 7702 // Return the storage we are tracking back into the free lists. 7703 if (CMSTraceSweeper) { 7704 gclog_or_tty->print_cr("Flushing ... "); 7705 } 7706 assert(freeFinger() < eob, "Error"); 7707 flush_cur_free_chunk( freeFinger(), pointer_delta(eob, freeFinger())); 7708 } 7709 } 7710 7711 void SweepClosure::flush_cur_free_chunk(HeapWord* chunk, size_t size) { 7712 assert(inFreeRange(), "Should only be called if currently in a free range."); 7713 assert(size > 0, 7714 "A zero sized chunk cannot be added to the free lists."); 7715 if (!freeRangeInFreeLists()) { 7716 if (CMSTestInFreeList) { 7717 FreeChunk* fc = (FreeChunk*) chunk; 7718 fc->set_size(size); 7719 assert(!_sp->verify_chunk_in_free_list(fc), 7720 "chunk should not be in free lists yet"); 7721 } 7722 if (CMSTraceSweeper) { 7723 gclog_or_tty->print_cr(" -- add free block " PTR_FORMAT " (" SIZE_FORMAT ") to free lists", 7724 p2i(chunk), size); 7725 } 7726 // A new free range is going to be starting. The current 7727 // free range has not been added to the free lists yet or 7728 // was removed so add it back. 7729 // If the current free range was coalesced, then the death 7730 // of the free range was recorded. Record a birth now. 7731 if (lastFreeRangeCoalesced()) { 7732 _sp->coalBirth(size); 7733 } 7734 _sp->addChunkAndRepairOffsetTable(chunk, size, 7735 lastFreeRangeCoalesced()); 7736 } else if (CMSTraceSweeper) { 7737 gclog_or_tty->print_cr("Already in free list: nothing to flush"); 7738 } 7739 set_inFreeRange(false); 7740 set_freeRangeInFreeLists(false); 7741 } 7742 7743 // We take a break if we've been at this for a while, 7744 // so as to avoid monopolizing the locks involved. 7745 void SweepClosure::do_yield_work(HeapWord* addr) { 7746 // Return current free chunk being used for coalescing (if any) 7747 // to the appropriate freelist. After yielding, the next 7748 // free block encountered will start a coalescing range of 7749 // free blocks. If the next free block is adjacent to the 7750 // chunk just flushed, they will need to wait for the next 7751 // sweep to be coalesced. 7752 if (inFreeRange()) { 7753 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 7754 } 7755 7756 // First give up the locks, then yield, then re-lock. 7757 // We should probably use a constructor/destructor idiom to 7758 // do this unlock/lock or modify the MutexUnlocker class to 7759 // serve our purpose. XXX 7760 assert_lock_strong(_bitMap->lock()); 7761 assert_lock_strong(_freelistLock); 7762 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 7763 "CMS thread should hold CMS token"); 7764 _bitMap->lock()->unlock(); 7765 _freelistLock->unlock(); 7766 ConcurrentMarkSweepThread::desynchronize(true); 7767 _collector->stopTimer(); 7768 if (PrintCMSStatistics != 0) { 7769 _collector->incrementYields(); 7770 } 7771 7772 // See the comment in coordinator_yield() 7773 for (unsigned i = 0; i < CMSYieldSleepCount && 7774 ConcurrentMarkSweepThread::should_yield() && 7775 !CMSCollector::foregroundGCIsActive(); ++i) { 7776 os::sleep(Thread::current(), 1, false); 7777 } 7778 7779 ConcurrentMarkSweepThread::synchronize(true); 7780 _freelistLock->lock(); 7781 _bitMap->lock()->lock_without_safepoint_check(); 7782 _collector->startTimer(); 7783 } 7784 7785 #ifndef PRODUCT 7786 // This is actually very useful in a product build if it can 7787 // be called from the debugger. Compile it into the product 7788 // as needed. 7789 bool debug_verify_chunk_in_free_list(FreeChunk* fc) { 7790 return debug_cms_space->verify_chunk_in_free_list(fc); 7791 } 7792 #endif 7793 7794 void SweepClosure::print_free_block_coalesced(FreeChunk* fc) const { 7795 if (CMSTraceSweeper) { 7796 gclog_or_tty->print_cr("Sweep:coal_free_blk " PTR_FORMAT " (" SIZE_FORMAT ")", 7797 p2i(fc), fc->size()); 7798 } 7799 } 7800 7801 // CMSIsAliveClosure 7802 bool CMSIsAliveClosure::do_object_b(oop obj) { 7803 HeapWord* addr = (HeapWord*)obj; 7804 return addr != NULL && 7805 (!_span.contains(addr) || _bit_map->isMarked(addr)); 7806 } 7807 7808 7809 CMSKeepAliveClosure::CMSKeepAliveClosure( CMSCollector* collector, 7810 MemRegion span, 7811 CMSBitMap* bit_map, CMSMarkStack* mark_stack, 7812 bool cpc): 7813 _collector(collector), 7814 _span(span), 7815 _bit_map(bit_map), 7816 _mark_stack(mark_stack), 7817 _concurrent_precleaning(cpc) { 7818 assert(!_span.is_empty(), "Empty span could spell trouble"); 7819 } 7820 7821 7822 // CMSKeepAliveClosure: the serial version 7823 void CMSKeepAliveClosure::do_oop(oop obj) { 7824 HeapWord* addr = (HeapWord*)obj; 7825 if (_span.contains(addr) && 7826 !_bit_map->isMarked(addr)) { 7827 _bit_map->mark(addr); 7828 bool simulate_overflow = false; 7829 NOT_PRODUCT( 7830 if (CMSMarkStackOverflowALot && 7831 _collector->simulate_overflow()) { 7832 // simulate a stack overflow 7833 simulate_overflow = true; 7834 } 7835 ) 7836 if (simulate_overflow || !_mark_stack->push(obj)) { 7837 if (_concurrent_precleaning) { 7838 // We dirty the overflown object and let the remark 7839 // phase deal with it. 7840 assert(_collector->overflow_list_is_empty(), "Error"); 7841 // In the case of object arrays, we need to dirty all of 7842 // the cards that the object spans. No locking or atomics 7843 // are needed since no one else can be mutating the mod union 7844 // table. 7845 if (obj->is_objArray()) { 7846 size_t sz = obj->size(); 7847 HeapWord* end_card_addr = 7848 (HeapWord*)round_to((intptr_t)(addr+sz), CardTableModRefBS::card_size); 7849 MemRegion redirty_range = MemRegion(addr, end_card_addr); 7850 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 7851 _collector->_modUnionTable.mark_range(redirty_range); 7852 } else { 7853 _collector->_modUnionTable.mark(addr); 7854 } 7855 _collector->_ser_kac_preclean_ovflw++; 7856 } else { 7857 _collector->push_on_overflow_list(obj); 7858 _collector->_ser_kac_ovflw++; 7859 } 7860 } 7861 } 7862 } 7863 7864 void CMSKeepAliveClosure::do_oop(oop* p) { CMSKeepAliveClosure::do_oop_work(p); } 7865 void CMSKeepAliveClosure::do_oop(narrowOop* p) { CMSKeepAliveClosure::do_oop_work(p); } 7866 7867 // CMSParKeepAliveClosure: a parallel version of the above. 7868 // The work queues are private to each closure (thread), 7869 // but (may be) available for stealing by other threads. 7870 void CMSParKeepAliveClosure::do_oop(oop obj) { 7871 HeapWord* addr = (HeapWord*)obj; 7872 if (_span.contains(addr) && 7873 !_bit_map->isMarked(addr)) { 7874 // In general, during recursive tracing, several threads 7875 // may be concurrently getting here; the first one to 7876 // "tag" it, claims it. 7877 if (_bit_map->par_mark(addr)) { 7878 bool res = _work_queue->push(obj); 7879 assert(res, "Low water mark should be much less than capacity"); 7880 // Do a recursive trim in the hope that this will keep 7881 // stack usage lower, but leave some oops for potential stealers 7882 trim_queue(_low_water_mark); 7883 } // Else, another thread got there first 7884 } 7885 } 7886 7887 void CMSParKeepAliveClosure::do_oop(oop* p) { CMSParKeepAliveClosure::do_oop_work(p); } 7888 void CMSParKeepAliveClosure::do_oop(narrowOop* p) { CMSParKeepAliveClosure::do_oop_work(p); } 7889 7890 void CMSParKeepAliveClosure::trim_queue(uint max) { 7891 while (_work_queue->size() > max) { 7892 oop new_oop; 7893 if (_work_queue->pop_local(new_oop)) { 7894 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 7895 assert(_bit_map->isMarked((HeapWord*)new_oop), 7896 "no white objects on this stack!"); 7897 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 7898 // iterate over the oops in this oop, marking and pushing 7899 // the ones in CMS heap (i.e. in _span). 7900 new_oop->oop_iterate(&_mark_and_push); 7901 } 7902 } 7903 } 7904 7905 CMSInnerParMarkAndPushClosure::CMSInnerParMarkAndPushClosure( 7906 CMSCollector* collector, 7907 MemRegion span, CMSBitMap* bit_map, 7908 OopTaskQueue* work_queue): 7909 _collector(collector), 7910 _span(span), 7911 _bit_map(bit_map), 7912 _work_queue(work_queue) { } 7913 7914 void CMSInnerParMarkAndPushClosure::do_oop(oop obj) { 7915 HeapWord* addr = (HeapWord*)obj; 7916 if (_span.contains(addr) && 7917 !_bit_map->isMarked(addr)) { 7918 if (_bit_map->par_mark(addr)) { 7919 bool simulate_overflow = false; 7920 NOT_PRODUCT( 7921 if (CMSMarkStackOverflowALot && 7922 _collector->par_simulate_overflow()) { 7923 // simulate a stack overflow 7924 simulate_overflow = true; 7925 } 7926 ) 7927 if (simulate_overflow || !_work_queue->push(obj)) { 7928 _collector->par_push_on_overflow_list(obj); 7929 _collector->_par_kac_ovflw++; 7930 } 7931 } // Else another thread got there already 7932 } 7933 } 7934 7935 void CMSInnerParMarkAndPushClosure::do_oop(oop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); } 7936 void CMSInnerParMarkAndPushClosure::do_oop(narrowOop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); } 7937 7938 ////////////////////////////////////////////////////////////////// 7939 // CMSExpansionCause ///////////////////////////// 7940 ////////////////////////////////////////////////////////////////// 7941 const char* CMSExpansionCause::to_string(CMSExpansionCause::Cause cause) { 7942 switch (cause) { 7943 case _no_expansion: 7944 return "No expansion"; 7945 case _satisfy_free_ratio: 7946 return "Free ratio"; 7947 case _satisfy_promotion: 7948 return "Satisfy promotion"; 7949 case _satisfy_allocation: 7950 return "allocation"; 7951 case _allocate_par_lab: 7952 return "Par LAB"; 7953 case _allocate_par_spooling_space: 7954 return "Par Spooling Space"; 7955 case _adaptive_size_policy: 7956 return "Ergonomics"; 7957 default: 7958 return "unknown"; 7959 } 7960 } 7961 7962 void CMSDrainMarkingStackClosure::do_void() { 7963 // the max number to take from overflow list at a time 7964 const size_t num = _mark_stack->capacity()/4; 7965 assert(!_concurrent_precleaning || _collector->overflow_list_is_empty(), 7966 "Overflow list should be NULL during concurrent phases"); 7967 while (!_mark_stack->isEmpty() || 7968 // if stack is empty, check the overflow list 7969 _collector->take_from_overflow_list(num, _mark_stack)) { 7970 oop obj = _mark_stack->pop(); 7971 HeapWord* addr = (HeapWord*)obj; 7972 assert(_span.contains(addr), "Should be within span"); 7973 assert(_bit_map->isMarked(addr), "Should be marked"); 7974 assert(obj->is_oop(), "Should be an oop"); 7975 obj->oop_iterate(_keep_alive); 7976 } 7977 } 7978 7979 void CMSParDrainMarkingStackClosure::do_void() { 7980 // drain queue 7981 trim_queue(0); 7982 } 7983 7984 // Trim our work_queue so its length is below max at return 7985 void CMSParDrainMarkingStackClosure::trim_queue(uint max) { 7986 while (_work_queue->size() > max) { 7987 oop new_oop; 7988 if (_work_queue->pop_local(new_oop)) { 7989 assert(new_oop->is_oop(), "Expected an oop"); 7990 assert(_bit_map->isMarked((HeapWord*)new_oop), 7991 "no white objects on this stack!"); 7992 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 7993 // iterate over the oops in this oop, marking and pushing 7994 // the ones in CMS heap (i.e. in _span). 7995 new_oop->oop_iterate(&_mark_and_push); 7996 } 7997 } 7998 } 7999 8000 //////////////////////////////////////////////////////////////////// 8001 // Support for Marking Stack Overflow list handling and related code 8002 //////////////////////////////////////////////////////////////////// 8003 // Much of the following code is similar in shape and spirit to the 8004 // code used in ParNewGC. We should try and share that code 8005 // as much as possible in the future. 8006 8007 #ifndef PRODUCT 8008 // Debugging support for CMSStackOverflowALot 8009 8010 // It's OK to call this multi-threaded; the worst thing 8011 // that can happen is that we'll get a bunch of closely 8012 // spaced simulated overflows, but that's OK, in fact 8013 // probably good as it would exercise the overflow code 8014 // under contention. 8015 bool CMSCollector::simulate_overflow() { 8016 if (_overflow_counter-- <= 0) { // just being defensive 8017 _overflow_counter = CMSMarkStackOverflowInterval; 8018 return true; 8019 } else { 8020 return false; 8021 } 8022 } 8023 8024 bool CMSCollector::par_simulate_overflow() { 8025 return simulate_overflow(); 8026 } 8027 #endif 8028 8029 // Single-threaded 8030 bool CMSCollector::take_from_overflow_list(size_t num, CMSMarkStack* stack) { 8031 assert(stack->isEmpty(), "Expected precondition"); 8032 assert(stack->capacity() > num, "Shouldn't bite more than can chew"); 8033 size_t i = num; 8034 oop cur = _overflow_list; 8035 const markOop proto = markOopDesc::prototype(); 8036 NOT_PRODUCT(ssize_t n = 0;) 8037 for (oop next; i > 0 && cur != NULL; cur = next, i--) { 8038 next = oop(cur->mark()); 8039 cur->set_mark(proto); // until proven otherwise 8040 assert(cur->is_oop(), "Should be an oop"); 8041 bool res = stack->push(cur); 8042 assert(res, "Bit off more than can chew?"); 8043 NOT_PRODUCT(n++;) 8044 } 8045 _overflow_list = cur; 8046 #ifndef PRODUCT 8047 assert(_num_par_pushes >= n, "Too many pops?"); 8048 _num_par_pushes -=n; 8049 #endif 8050 return !stack->isEmpty(); 8051 } 8052 8053 #define BUSY (cast_to_oop<intptr_t>(0x1aff1aff)) 8054 // (MT-safe) Get a prefix of at most "num" from the list. 8055 // The overflow list is chained through the mark word of 8056 // each object in the list. We fetch the entire list, 8057 // break off a prefix of the right size and return the 8058 // remainder. If other threads try to take objects from 8059 // the overflow list at that time, they will wait for 8060 // some time to see if data becomes available. If (and 8061 // only if) another thread places one or more object(s) 8062 // on the global list before we have returned the suffix 8063 // to the global list, we will walk down our local list 8064 // to find its end and append the global list to 8065 // our suffix before returning it. This suffix walk can 8066 // prove to be expensive (quadratic in the amount of traffic) 8067 // when there are many objects in the overflow list and 8068 // there is much producer-consumer contention on the list. 8069 // *NOTE*: The overflow list manipulation code here and 8070 // in ParNewGeneration:: are very similar in shape, 8071 // except that in the ParNew case we use the old (from/eden) 8072 // copy of the object to thread the list via its klass word. 8073 // Because of the common code, if you make any changes in 8074 // the code below, please check the ParNew version to see if 8075 // similar changes might be needed. 8076 // CR 6797058 has been filed to consolidate the common code. 8077 bool CMSCollector::par_take_from_overflow_list(size_t num, 8078 OopTaskQueue* work_q, 8079 int no_of_gc_threads) { 8080 assert(work_q->size() == 0, "First empty local work queue"); 8081 assert(num < work_q->max_elems(), "Can't bite more than we can chew"); 8082 if (_overflow_list == NULL) { 8083 return false; 8084 } 8085 // Grab the entire list; we'll put back a suffix 8086 oop prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list)); 8087 Thread* tid = Thread::current(); 8088 // Before "no_of_gc_threads" was introduced CMSOverflowSpinCount was 8089 // set to ParallelGCThreads. 8090 size_t CMSOverflowSpinCount = (size_t) no_of_gc_threads; // was ParallelGCThreads; 8091 size_t sleep_time_millis = MAX2((size_t)1, num/100); 8092 // If the list is busy, we spin for a short while, 8093 // sleeping between attempts to get the list. 8094 for (size_t spin = 0; prefix == BUSY && spin < CMSOverflowSpinCount; spin++) { 8095 os::sleep(tid, sleep_time_millis, false); 8096 if (_overflow_list == NULL) { 8097 // Nothing left to take 8098 return false; 8099 } else if (_overflow_list != BUSY) { 8100 // Try and grab the prefix 8101 prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list)); 8102 } 8103 } 8104 // If the list was found to be empty, or we spun long 8105 // enough, we give up and return empty-handed. If we leave 8106 // the list in the BUSY state below, it must be the case that 8107 // some other thread holds the overflow list and will set it 8108 // to a non-BUSY state in the future. 8109 if (prefix == NULL || prefix == BUSY) { 8110 // Nothing to take or waited long enough 8111 if (prefix == NULL) { 8112 // Write back the NULL in case we overwrote it with BUSY above 8113 // and it is still the same value. 8114 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY); 8115 } 8116 return false; 8117 } 8118 assert(prefix != NULL && prefix != BUSY, "Error"); 8119 size_t i = num; 8120 oop cur = prefix; 8121 // Walk down the first "num" objects, unless we reach the end. 8122 for (; i > 1 && cur->mark() != NULL; cur = oop(cur->mark()), i--); 8123 if (cur->mark() == NULL) { 8124 // We have "num" or fewer elements in the list, so there 8125 // is nothing to return to the global list. 8126 // Write back the NULL in lieu of the BUSY we wrote 8127 // above, if it is still the same value. 8128 if (_overflow_list == BUSY) { 8129 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY); 8130 } 8131 } else { 8132 // Chop off the suffix and return it to the global list. 8133 assert(cur->mark() != BUSY, "Error"); 8134 oop suffix_head = cur->mark(); // suffix will be put back on global list 8135 cur->set_mark(NULL); // break off suffix 8136 // It's possible that the list is still in the empty(busy) state 8137 // we left it in a short while ago; in that case we may be 8138 // able to place back the suffix without incurring the cost 8139 // of a walk down the list. 8140 oop observed_overflow_list = _overflow_list; 8141 oop cur_overflow_list = observed_overflow_list; 8142 bool attached = false; 8143 while (observed_overflow_list == BUSY || observed_overflow_list == NULL) { 8144 observed_overflow_list = 8145 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list); 8146 if (cur_overflow_list == observed_overflow_list) { 8147 attached = true; 8148 break; 8149 } else cur_overflow_list = observed_overflow_list; 8150 } 8151 if (!attached) { 8152 // Too bad, someone else sneaked in (at least) an element; we'll need 8153 // to do a splice. Find tail of suffix so we can prepend suffix to global 8154 // list. 8155 for (cur = suffix_head; cur->mark() != NULL; cur = (oop)(cur->mark())); 8156 oop suffix_tail = cur; 8157 assert(suffix_tail != NULL && suffix_tail->mark() == NULL, 8158 "Tautology"); 8159 observed_overflow_list = _overflow_list; 8160 do { 8161 cur_overflow_list = observed_overflow_list; 8162 if (cur_overflow_list != BUSY) { 8163 // Do the splice ... 8164 suffix_tail->set_mark(markOop(cur_overflow_list)); 8165 } else { // cur_overflow_list == BUSY 8166 suffix_tail->set_mark(NULL); 8167 } 8168 // ... and try to place spliced list back on overflow_list ... 8169 observed_overflow_list = 8170 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list); 8171 } while (cur_overflow_list != observed_overflow_list); 8172 // ... until we have succeeded in doing so. 8173 } 8174 } 8175 8176 // Push the prefix elements on work_q 8177 assert(prefix != NULL, "control point invariant"); 8178 const markOop proto = markOopDesc::prototype(); 8179 oop next; 8180 NOT_PRODUCT(ssize_t n = 0;) 8181 for (cur = prefix; cur != NULL; cur = next) { 8182 next = oop(cur->mark()); 8183 cur->set_mark(proto); // until proven otherwise 8184 assert(cur->is_oop(), "Should be an oop"); 8185 bool res = work_q->push(cur); 8186 assert(res, "Bit off more than we can chew?"); 8187 NOT_PRODUCT(n++;) 8188 } 8189 #ifndef PRODUCT 8190 assert(_num_par_pushes >= n, "Too many pops?"); 8191 Atomic::add_ptr(-(intptr_t)n, &_num_par_pushes); 8192 #endif 8193 return true; 8194 } 8195 8196 // Single-threaded 8197 void CMSCollector::push_on_overflow_list(oop p) { 8198 NOT_PRODUCT(_num_par_pushes++;) 8199 assert(p->is_oop(), "Not an oop"); 8200 preserve_mark_if_necessary(p); 8201 p->set_mark((markOop)_overflow_list); 8202 _overflow_list = p; 8203 } 8204 8205 // Multi-threaded; use CAS to prepend to overflow list 8206 void CMSCollector::par_push_on_overflow_list(oop p) { 8207 NOT_PRODUCT(Atomic::inc_ptr(&_num_par_pushes);) 8208 assert(p->is_oop(), "Not an oop"); 8209 par_preserve_mark_if_necessary(p); 8210 oop observed_overflow_list = _overflow_list; 8211 oop cur_overflow_list; 8212 do { 8213 cur_overflow_list = observed_overflow_list; 8214 if (cur_overflow_list != BUSY) { 8215 p->set_mark(markOop(cur_overflow_list)); 8216 } else { 8217 p->set_mark(NULL); 8218 } 8219 observed_overflow_list = 8220 (oop) Atomic::cmpxchg_ptr(p, &_overflow_list, cur_overflow_list); 8221 } while (cur_overflow_list != observed_overflow_list); 8222 } 8223 #undef BUSY 8224 8225 // Single threaded 8226 // General Note on GrowableArray: pushes may silently fail 8227 // because we are (temporarily) out of C-heap for expanding 8228 // the stack. The problem is quite ubiquitous and affects 8229 // a lot of code in the JVM. The prudent thing for GrowableArray 8230 // to do (for now) is to exit with an error. However, that may 8231 // be too draconian in some cases because the caller may be 8232 // able to recover without much harm. For such cases, we 8233 // should probably introduce a "soft_push" method which returns 8234 // an indication of success or failure with the assumption that 8235 // the caller may be able to recover from a failure; code in 8236 // the VM can then be changed, incrementally, to deal with such 8237 // failures where possible, thus, incrementally hardening the VM 8238 // in such low resource situations. 8239 void CMSCollector::preserve_mark_work(oop p, markOop m) { 8240 _preserved_oop_stack.push(p); 8241 _preserved_mark_stack.push(m); 8242 assert(m == p->mark(), "Mark word changed"); 8243 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), 8244 "bijection"); 8245 } 8246 8247 // Single threaded 8248 void CMSCollector::preserve_mark_if_necessary(oop p) { 8249 markOop m = p->mark(); 8250 if (m->must_be_preserved(p)) { 8251 preserve_mark_work(p, m); 8252 } 8253 } 8254 8255 void CMSCollector::par_preserve_mark_if_necessary(oop p) { 8256 markOop m = p->mark(); 8257 if (m->must_be_preserved(p)) { 8258 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag); 8259 // Even though we read the mark word without holding 8260 // the lock, we are assured that it will not change 8261 // because we "own" this oop, so no other thread can 8262 // be trying to push it on the overflow list; see 8263 // the assertion in preserve_mark_work() that checks 8264 // that m == p->mark(). 8265 preserve_mark_work(p, m); 8266 } 8267 } 8268 8269 // We should be able to do this multi-threaded, 8270 // a chunk of stack being a task (this is 8271 // correct because each oop only ever appears 8272 // once in the overflow list. However, it's 8273 // not very easy to completely overlap this with 8274 // other operations, so will generally not be done 8275 // until all work's been completed. Because we 8276 // expect the preserved oop stack (set) to be small, 8277 // it's probably fine to do this single-threaded. 8278 // We can explore cleverer concurrent/overlapped/parallel 8279 // processing of preserved marks if we feel the 8280 // need for this in the future. Stack overflow should 8281 // be so rare in practice and, when it happens, its 8282 // effect on performance so great that this will 8283 // likely just be in the noise anyway. 8284 void CMSCollector::restore_preserved_marks_if_any() { 8285 assert(SafepointSynchronize::is_at_safepoint(), 8286 "world should be stopped"); 8287 assert(Thread::current()->is_ConcurrentGC_thread() || 8288 Thread::current()->is_VM_thread(), 8289 "should be single-threaded"); 8290 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), 8291 "bijection"); 8292 8293 while (!_preserved_oop_stack.is_empty()) { 8294 oop p = _preserved_oop_stack.pop(); 8295 assert(p->is_oop(), "Should be an oop"); 8296 assert(_span.contains(p), "oop should be in _span"); 8297 assert(p->mark() == markOopDesc::prototype(), 8298 "Set when taken from overflow list"); 8299 markOop m = _preserved_mark_stack.pop(); 8300 p->set_mark(m); 8301 } 8302 assert(_preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(), 8303 "stacks were cleared above"); 8304 } 8305 8306 #ifndef PRODUCT 8307 bool CMSCollector::no_preserved_marks() const { 8308 return _preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(); 8309 } 8310 #endif 8311 8312 // Transfer some number of overflown objects to usual marking 8313 // stack. Return true if some objects were transferred. 8314 bool MarkRefsIntoAndScanClosure::take_from_overflow_list() { 8315 size_t num = MIN2((size_t)(_mark_stack->capacity() - _mark_stack->length())/4, 8316 (size_t)ParGCDesiredObjsFromOverflowList); 8317 8318 bool res = _collector->take_from_overflow_list(num, _mark_stack); 8319 assert(_collector->overflow_list_is_empty() || res, 8320 "If list is not empty, we should have taken something"); 8321 assert(!res || !_mark_stack->isEmpty(), 8322 "If we took something, it should now be on our stack"); 8323 return res; 8324 } 8325 8326 size_t MarkDeadObjectsClosure::do_blk(HeapWord* addr) { 8327 size_t res = _sp->block_size_no_stall(addr, _collector); 8328 if (_sp->block_is_obj(addr)) { 8329 if (_live_bit_map->isMarked(addr)) { 8330 // It can't have been dead in a previous cycle 8331 guarantee(!_dead_bit_map->isMarked(addr), "No resurrection!"); 8332 } else { 8333 _dead_bit_map->mark(addr); // mark the dead object 8334 } 8335 } 8336 // Could be 0, if the block size could not be computed without stalling. 8337 return res; 8338 } 8339 8340 TraceCMSMemoryManagerStats::TraceCMSMemoryManagerStats(CMSCollector::CollectorState phase, GCCause::Cause cause): TraceMemoryManagerStats() { 8341 8342 switch (phase) { 8343 case CMSCollector::InitialMarking: 8344 initialize(true /* fullGC */ , 8345 cause /* cause of the GC */, 8346 true /* recordGCBeginTime */, 8347 true /* recordPreGCUsage */, 8348 false /* recordPeakUsage */, 8349 false /* recordPostGCusage */, 8350 true /* recordAccumulatedGCTime */, 8351 false /* recordGCEndTime */, 8352 false /* countCollection */ ); 8353 break; 8354 8355 case CMSCollector::FinalMarking: 8356 initialize(true /* fullGC */ , 8357 cause /* cause of the GC */, 8358 false /* recordGCBeginTime */, 8359 false /* recordPreGCUsage */, 8360 false /* recordPeakUsage */, 8361 false /* recordPostGCusage */, 8362 true /* recordAccumulatedGCTime */, 8363 false /* recordGCEndTime */, 8364 false /* countCollection */ ); 8365 break; 8366 8367 case CMSCollector::Sweeping: 8368 initialize(true /* fullGC */ , 8369 cause /* cause of the GC */, 8370 false /* recordGCBeginTime */, 8371 false /* recordPreGCUsage */, 8372 true /* recordPeakUsage */, 8373 true /* recordPostGCusage */, 8374 false /* recordAccumulatedGCTime */, 8375 true /* recordGCEndTime */, 8376 true /* countCollection */ ); 8377 break; 8378 8379 default: 8380 ShouldNotReachHere(); 8381 } 8382 }