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 2244 // Check reachability of the given heap address in CMS generation, 2245 // treating all other generations as roots. 2246 bool CMSCollector::is_cms_reachable(HeapWord* addr) { 2247 // We could "guarantee" below, rather than assert, but I'll 2248 // leave these as "asserts" so that an adventurous debugger 2249 // could try this in the product build provided some subset of 2250 // the conditions were met, provided they were interested in the 2251 // results and knew that the computation below wouldn't interfere 2252 // with other concurrent computations mutating the structures 2253 // being read or written. 2254 assert(SafepointSynchronize::is_at_safepoint(), 2255 "Else mutations in object graph will make answer suspect"); 2256 assert(have_cms_token(), "Should hold cms token"); 2257 assert(haveFreelistLocks(), "must hold free list locks"); 2258 assert_lock_strong(bitMapLock()); 2259 2260 // Clear the marking bit map array before starting, but, just 2261 // for kicks, first report if the given address is already marked 2262 tty->print_cr("Start: Address " PTR_FORMAT " is%s marked", p2i(addr), 2263 _markBitMap.isMarked(addr) ? "" : " not"); 2264 2265 if (verify_after_remark()) { 2266 MutexLockerEx x(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag); 2267 bool result = verification_mark_bm()->isMarked(addr); 2268 tty->print_cr("TransitiveMark: Address " PTR_FORMAT " %s marked", p2i(addr), 2269 result ? "IS" : "is NOT"); 2270 return result; 2271 } else { 2272 tty->print_cr("Could not compute result"); 2273 return false; 2274 } 2275 } 2276 #endif 2277 2278 void 2279 CMSCollector::print_on_error(outputStream* st) { 2280 CMSCollector* collector = ConcurrentMarkSweepGeneration::_collector; 2281 if (collector != NULL) { 2282 CMSBitMap* bitmap = &collector->_markBitMap; 2283 st->print_cr("Marking Bits: (CMSBitMap*) " PTR_FORMAT, p2i(bitmap)); 2284 bitmap->print_on_error(st, " Bits: "); 2285 2286 st->cr(); 2287 2288 CMSBitMap* mut_bitmap = &collector->_modUnionTable; 2289 st->print_cr("Mod Union Table: (CMSBitMap*) " PTR_FORMAT, p2i(mut_bitmap)); 2290 mut_bitmap->print_on_error(st, " Bits: "); 2291 } 2292 } 2293 2294 //////////////////////////////////////////////////////// 2295 // CMS Verification Support 2296 //////////////////////////////////////////////////////// 2297 // Following the remark phase, the following invariant 2298 // should hold -- each object in the CMS heap which is 2299 // marked in markBitMap() should be marked in the verification_mark_bm(). 2300 2301 class VerifyMarkedClosure: public BitMapClosure { 2302 CMSBitMap* _marks; 2303 bool _failed; 2304 2305 public: 2306 VerifyMarkedClosure(CMSBitMap* bm): _marks(bm), _failed(false) {} 2307 2308 bool do_bit(size_t offset) { 2309 HeapWord* addr = _marks->offsetToHeapWord(offset); 2310 if (!_marks->isMarked(addr)) { 2311 oop(addr)->print_on(gclog_or_tty); 2312 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)", p2i(addr)); 2313 _failed = true; 2314 } 2315 return true; 2316 } 2317 2318 bool failed() { return _failed; } 2319 }; 2320 2321 bool CMSCollector::verify_after_remark(bool silent) { 2322 if (!silent) gclog_or_tty->print(" [Verifying CMS Marking... "); 2323 MutexLockerEx ml(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag); 2324 static bool init = false; 2325 2326 assert(SafepointSynchronize::is_at_safepoint(), 2327 "Else mutations in object graph will make answer suspect"); 2328 assert(have_cms_token(), 2329 "Else there may be mutual interference in use of " 2330 " verification data structures"); 2331 assert(_collectorState > Marking && _collectorState <= Sweeping, 2332 "Else marking info checked here may be obsolete"); 2333 assert(haveFreelistLocks(), "must hold free list locks"); 2334 assert_lock_strong(bitMapLock()); 2335 2336 2337 // Allocate marking bit map if not already allocated 2338 if (!init) { // first time 2339 if (!verification_mark_bm()->allocate(_span)) { 2340 return false; 2341 } 2342 init = true; 2343 } 2344 2345 assert(verification_mark_stack()->isEmpty(), "Should be empty"); 2346 2347 // Turn off refs discovery -- so we will be tracing through refs. 2348 // This is as intended, because by this time 2349 // GC must already have cleared any refs that need to be cleared, 2350 // and traced those that need to be marked; moreover, 2351 // the marking done here is not going to interfere in any 2352 // way with the marking information used by GC. 2353 NoRefDiscovery no_discovery(ref_processor()); 2354 2355 #if defined(COMPILER2) || INCLUDE_JVMCI 2356 DerivedPointerTableDeactivate dpt_deact; 2357 #endif 2358 2359 // Clear any marks from a previous round 2360 verification_mark_bm()->clear_all(); 2361 assert(verification_mark_stack()->isEmpty(), "markStack should be empty"); 2362 verify_work_stacks_empty(); 2363 2364 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2365 gch->ensure_parsability(false); // fill TLABs, but no need to retire them 2366 // Update the saved marks which may affect the root scans. 2367 gch->save_marks(); 2368 2369 if (CMSRemarkVerifyVariant == 1) { 2370 // In this first variant of verification, we complete 2371 // all marking, then check if the new marks-vector is 2372 // a subset of the CMS marks-vector. 2373 verify_after_remark_work_1(); 2374 } else if (CMSRemarkVerifyVariant == 2) { 2375 // In this second variant of verification, we flag an error 2376 // (i.e. an object reachable in the new marks-vector not reachable 2377 // in the CMS marks-vector) immediately, also indicating the 2378 // identify of an object (A) that references the unmarked object (B) -- 2379 // presumably, a mutation to A failed to be picked up by preclean/remark? 2380 verify_after_remark_work_2(); 2381 } else { 2382 warning("Unrecognized value " UINTX_FORMAT " for CMSRemarkVerifyVariant", 2383 CMSRemarkVerifyVariant); 2384 } 2385 if (!silent) gclog_or_tty->print(" done] "); 2386 return true; 2387 } 2388 2389 void CMSCollector::verify_after_remark_work_1() { 2390 ResourceMark rm; 2391 HandleMark hm; 2392 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2393 2394 // Get a clear set of claim bits for the roots processing to work with. 2395 ClassLoaderDataGraph::clear_claimed_marks(); 2396 2397 // Mark from roots one level into CMS 2398 MarkRefsIntoClosure notOlder(_span, verification_mark_bm()); 2399 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2400 2401 { 2402 StrongRootsScope srs(1); 2403 2404 gch->gen_process_roots(&srs, 2405 GenCollectedHeap::OldGen, 2406 true, // young gen as roots 2407 GenCollectedHeap::ScanningOption(roots_scanning_options()), 2408 should_unload_classes(), 2409 ¬Older, 2410 NULL, 2411 NULL); 2412 } 2413 2414 // Now mark from the roots 2415 MarkFromRootsClosure markFromRootsClosure(this, _span, 2416 verification_mark_bm(), verification_mark_stack(), 2417 false /* don't yield */, true /* verifying */); 2418 assert(_restart_addr == NULL, "Expected pre-condition"); 2419 verification_mark_bm()->iterate(&markFromRootsClosure); 2420 while (_restart_addr != NULL) { 2421 // Deal with stack overflow: by restarting at the indicated 2422 // address. 2423 HeapWord* ra = _restart_addr; 2424 markFromRootsClosure.reset(ra); 2425 _restart_addr = NULL; 2426 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end()); 2427 } 2428 assert(verification_mark_stack()->isEmpty(), "Should have been drained"); 2429 verify_work_stacks_empty(); 2430 2431 // Marking completed -- now verify that each bit marked in 2432 // verification_mark_bm() is also marked in markBitMap(); flag all 2433 // errors by printing corresponding objects. 2434 VerifyMarkedClosure vcl(markBitMap()); 2435 verification_mark_bm()->iterate(&vcl); 2436 if (vcl.failed()) { 2437 gclog_or_tty->print("Verification failed"); 2438 gch->print_on(gclog_or_tty); 2439 fatal("CMS: failed marking verification after remark"); 2440 } 2441 } 2442 2443 class VerifyKlassOopsKlassClosure : public KlassClosure { 2444 class VerifyKlassOopsClosure : public OopClosure { 2445 CMSBitMap* _bitmap; 2446 public: 2447 VerifyKlassOopsClosure(CMSBitMap* bitmap) : _bitmap(bitmap) { } 2448 void do_oop(oop* p) { guarantee(*p == NULL || _bitmap->isMarked((HeapWord*) *p), "Should be marked"); } 2449 void do_oop(narrowOop* p) { ShouldNotReachHere(); } 2450 } _oop_closure; 2451 public: 2452 VerifyKlassOopsKlassClosure(CMSBitMap* bitmap) : _oop_closure(bitmap) {} 2453 void do_klass(Klass* k) { 2454 k->oops_do(&_oop_closure); 2455 } 2456 }; 2457 2458 void CMSCollector::verify_after_remark_work_2() { 2459 ResourceMark rm; 2460 HandleMark hm; 2461 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2462 2463 // Get a clear set of claim bits for the roots processing to work with. 2464 ClassLoaderDataGraph::clear_claimed_marks(); 2465 2466 // Mark from roots one level into CMS 2467 MarkRefsIntoVerifyClosure notOlder(_span, verification_mark_bm(), 2468 markBitMap()); 2469 CLDToOopClosure cld_closure(¬Older, true); 2470 2471 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2472 2473 { 2474 StrongRootsScope srs(1); 2475 2476 gch->gen_process_roots(&srs, 2477 GenCollectedHeap::OldGen, 2478 true, // young gen as roots 2479 GenCollectedHeap::ScanningOption(roots_scanning_options()), 2480 should_unload_classes(), 2481 ¬Older, 2482 NULL, 2483 &cld_closure); 2484 } 2485 2486 // Now mark from the roots 2487 MarkFromRootsVerifyClosure markFromRootsClosure(this, _span, 2488 verification_mark_bm(), markBitMap(), verification_mark_stack()); 2489 assert(_restart_addr == NULL, "Expected pre-condition"); 2490 verification_mark_bm()->iterate(&markFromRootsClosure); 2491 while (_restart_addr != NULL) { 2492 // Deal with stack overflow: by restarting at the indicated 2493 // address. 2494 HeapWord* ra = _restart_addr; 2495 markFromRootsClosure.reset(ra); 2496 _restart_addr = NULL; 2497 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end()); 2498 } 2499 assert(verification_mark_stack()->isEmpty(), "Should have been drained"); 2500 verify_work_stacks_empty(); 2501 2502 VerifyKlassOopsKlassClosure verify_klass_oops(verification_mark_bm()); 2503 ClassLoaderDataGraph::classes_do(&verify_klass_oops); 2504 2505 // Marking completed -- now verify that each bit marked in 2506 // verification_mark_bm() is also marked in markBitMap(); flag all 2507 // errors by printing corresponding objects. 2508 VerifyMarkedClosure vcl(markBitMap()); 2509 verification_mark_bm()->iterate(&vcl); 2510 assert(!vcl.failed(), "Else verification above should not have succeeded"); 2511 } 2512 2513 void ConcurrentMarkSweepGeneration::save_marks() { 2514 // delegate to CMS space 2515 cmsSpace()->save_marks(); 2516 for (uint i = 0; i < ParallelGCThreads; i++) { 2517 _par_gc_thread_states[i]->promo.startTrackingPromotions(); 2518 } 2519 } 2520 2521 bool ConcurrentMarkSweepGeneration::no_allocs_since_save_marks() { 2522 return cmsSpace()->no_allocs_since_save_marks(); 2523 } 2524 2525 #define CMS_SINCE_SAVE_MARKS_DEFN(OopClosureType, nv_suffix) \ 2526 \ 2527 void ConcurrentMarkSweepGeneration:: \ 2528 oop_since_save_marks_iterate##nv_suffix(OopClosureType* cl) { \ 2529 cl->set_generation(this); \ 2530 cmsSpace()->oop_since_save_marks_iterate##nv_suffix(cl); \ 2531 cl->reset_generation(); \ 2532 save_marks(); \ 2533 } 2534 2535 ALL_SINCE_SAVE_MARKS_CLOSURES(CMS_SINCE_SAVE_MARKS_DEFN) 2536 2537 void 2538 ConcurrentMarkSweepGeneration::oop_iterate(ExtendedOopClosure* cl) { 2539 if (freelistLock()->owned_by_self()) { 2540 Generation::oop_iterate(cl); 2541 } else { 2542 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2543 Generation::oop_iterate(cl); 2544 } 2545 } 2546 2547 void 2548 ConcurrentMarkSweepGeneration::object_iterate(ObjectClosure* cl) { 2549 if (freelistLock()->owned_by_self()) { 2550 Generation::object_iterate(cl); 2551 } else { 2552 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2553 Generation::object_iterate(cl); 2554 } 2555 } 2556 2557 void 2558 ConcurrentMarkSweepGeneration::safe_object_iterate(ObjectClosure* cl) { 2559 if (freelistLock()->owned_by_self()) { 2560 Generation::safe_object_iterate(cl); 2561 } else { 2562 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2563 Generation::safe_object_iterate(cl); 2564 } 2565 } 2566 2567 void 2568 ConcurrentMarkSweepGeneration::post_compact() { 2569 } 2570 2571 void 2572 ConcurrentMarkSweepGeneration::prepare_for_verify() { 2573 // Fix the linear allocation blocks to look like free blocks. 2574 2575 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those 2576 // are not called when the heap is verified during universe initialization and 2577 // at vm shutdown. 2578 if (freelistLock()->owned_by_self()) { 2579 cmsSpace()->prepare_for_verify(); 2580 } else { 2581 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag); 2582 cmsSpace()->prepare_for_verify(); 2583 } 2584 } 2585 2586 void 2587 ConcurrentMarkSweepGeneration::verify() { 2588 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those 2589 // are not called when the heap is verified during universe initialization and 2590 // at vm shutdown. 2591 if (freelistLock()->owned_by_self()) { 2592 cmsSpace()->verify(); 2593 } else { 2594 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag); 2595 cmsSpace()->verify(); 2596 } 2597 } 2598 2599 void CMSCollector::verify() { 2600 _cmsGen->verify(); 2601 } 2602 2603 #ifndef PRODUCT 2604 bool CMSCollector::overflow_list_is_empty() const { 2605 assert(_num_par_pushes >= 0, "Inconsistency"); 2606 if (_overflow_list == NULL) { 2607 assert(_num_par_pushes == 0, "Inconsistency"); 2608 } 2609 return _overflow_list == NULL; 2610 } 2611 2612 // The methods verify_work_stacks_empty() and verify_overflow_empty() 2613 // merely consolidate assertion checks that appear to occur together frequently. 2614 void CMSCollector::verify_work_stacks_empty() const { 2615 assert(_markStack.isEmpty(), "Marking stack should be empty"); 2616 assert(overflow_list_is_empty(), "Overflow list should be empty"); 2617 } 2618 2619 void CMSCollector::verify_overflow_empty() const { 2620 assert(overflow_list_is_empty(), "Overflow list should be empty"); 2621 assert(no_preserved_marks(), "No preserved marks"); 2622 } 2623 #endif // PRODUCT 2624 2625 // Decide if we want to enable class unloading as part of the 2626 // ensuing concurrent GC cycle. We will collect and 2627 // unload classes if it's the case that: 2628 // (1) an explicit gc request has been made and the flag 2629 // ExplicitGCInvokesConcurrentAndUnloadsClasses is set, OR 2630 // (2) (a) class unloading is enabled at the command line, and 2631 // (b) old gen is getting really full 2632 // NOTE: Provided there is no change in the state of the heap between 2633 // calls to this method, it should have idempotent results. Moreover, 2634 // its results should be monotonically increasing (i.e. going from 0 to 1, 2635 // but not 1 to 0) between successive calls between which the heap was 2636 // not collected. For the implementation below, it must thus rely on 2637 // the property that concurrent_cycles_since_last_unload() 2638 // will not decrease unless a collection cycle happened and that 2639 // _cmsGen->is_too_full() are 2640 // themselves also monotonic in that sense. See check_monotonicity() 2641 // below. 2642 void CMSCollector::update_should_unload_classes() { 2643 _should_unload_classes = false; 2644 // Condition 1 above 2645 if (_full_gc_requested && ExplicitGCInvokesConcurrentAndUnloadsClasses) { 2646 _should_unload_classes = true; 2647 } else if (CMSClassUnloadingEnabled) { // Condition 2.a above 2648 // Disjuncts 2.b.(i,ii,iii) above 2649 _should_unload_classes = (concurrent_cycles_since_last_unload() >= 2650 CMSClassUnloadingMaxInterval) 2651 || _cmsGen->is_too_full(); 2652 } 2653 } 2654 2655 bool ConcurrentMarkSweepGeneration::is_too_full() const { 2656 bool res = should_concurrent_collect(); 2657 res = res && (occupancy() > (double)CMSIsTooFullPercentage/100.0); 2658 return res; 2659 } 2660 2661 void CMSCollector::setup_cms_unloading_and_verification_state() { 2662 const bool should_verify = VerifyBeforeGC || VerifyAfterGC || VerifyDuringGC 2663 || VerifyBeforeExit; 2664 const int rso = GenCollectedHeap::SO_AllCodeCache; 2665 2666 // We set the proper root for this CMS cycle here. 2667 if (should_unload_classes()) { // Should unload classes this cycle 2668 remove_root_scanning_option(rso); // Shrink the root set appropriately 2669 set_verifying(should_verify); // Set verification state for this cycle 2670 return; // Nothing else needs to be done at this time 2671 } 2672 2673 // Not unloading classes this cycle 2674 assert(!should_unload_classes(), "Inconsistency!"); 2675 2676 // If we are not unloading classes then add SO_AllCodeCache to root 2677 // scanning options. 2678 add_root_scanning_option(rso); 2679 2680 if ((!verifying() || unloaded_classes_last_cycle()) && should_verify) { 2681 set_verifying(true); 2682 } else if (verifying() && !should_verify) { 2683 // We were verifying, but some verification flags got disabled. 2684 set_verifying(false); 2685 // Exclude symbols, strings and code cache elements from root scanning to 2686 // reduce IM and RM pauses. 2687 remove_root_scanning_option(rso); 2688 } 2689 } 2690 2691 2692 #ifndef PRODUCT 2693 HeapWord* CMSCollector::block_start(const void* p) const { 2694 const HeapWord* addr = (HeapWord*)p; 2695 if (_span.contains(p)) { 2696 if (_cmsGen->cmsSpace()->is_in_reserved(addr)) { 2697 return _cmsGen->cmsSpace()->block_start(p); 2698 } 2699 } 2700 return NULL; 2701 } 2702 #endif 2703 2704 HeapWord* 2705 ConcurrentMarkSweepGeneration::expand_and_allocate(size_t word_size, 2706 bool tlab, 2707 bool parallel) { 2708 CMSSynchronousYieldRequest yr; 2709 assert(!tlab, "Can't deal with TLAB allocation"); 2710 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2711 expand_for_gc_cause(word_size*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_satisfy_allocation); 2712 if (GCExpandToAllocateDelayMillis > 0) { 2713 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 2714 } 2715 return have_lock_and_allocate(word_size, tlab); 2716 } 2717 2718 void ConcurrentMarkSweepGeneration::expand_for_gc_cause( 2719 size_t bytes, 2720 size_t expand_bytes, 2721 CMSExpansionCause::Cause cause) 2722 { 2723 2724 bool success = expand(bytes, expand_bytes); 2725 2726 // remember why we expanded; this information is used 2727 // by shouldConcurrentCollect() when making decisions on whether to start 2728 // a new CMS cycle. 2729 if (success) { 2730 set_expansion_cause(cause); 2731 if (PrintGCDetails && Verbose) { 2732 gclog_or_tty->print_cr("Expanded CMS gen for %s", 2733 CMSExpansionCause::to_string(cause)); 2734 } 2735 } 2736 } 2737 2738 HeapWord* ConcurrentMarkSweepGeneration::expand_and_par_lab_allocate(CMSParGCThreadState* ps, size_t word_sz) { 2739 HeapWord* res = NULL; 2740 MutexLocker x(ParGCRareEvent_lock); 2741 while (true) { 2742 // Expansion by some other thread might make alloc OK now: 2743 res = ps->lab.alloc(word_sz); 2744 if (res != NULL) return res; 2745 // If there's not enough expansion space available, give up. 2746 if (_virtual_space.uncommitted_size() < (word_sz * HeapWordSize)) { 2747 return NULL; 2748 } 2749 // Otherwise, we try expansion. 2750 expand_for_gc_cause(word_sz*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_allocate_par_lab); 2751 // Now go around the loop and try alloc again; 2752 // A competing par_promote might beat us to the expansion space, 2753 // so we may go around the loop again if promotion fails again. 2754 if (GCExpandToAllocateDelayMillis > 0) { 2755 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 2756 } 2757 } 2758 } 2759 2760 2761 bool ConcurrentMarkSweepGeneration::expand_and_ensure_spooling_space( 2762 PromotionInfo* promo) { 2763 MutexLocker x(ParGCRareEvent_lock); 2764 size_t refill_size_bytes = promo->refillSize() * HeapWordSize; 2765 while (true) { 2766 // Expansion by some other thread might make alloc OK now: 2767 if (promo->ensure_spooling_space()) { 2768 assert(promo->has_spooling_space(), 2769 "Post-condition of successful ensure_spooling_space()"); 2770 return true; 2771 } 2772 // If there's not enough expansion space available, give up. 2773 if (_virtual_space.uncommitted_size() < refill_size_bytes) { 2774 return false; 2775 } 2776 // Otherwise, we try expansion. 2777 expand_for_gc_cause(refill_size_bytes, MinHeapDeltaBytes, CMSExpansionCause::_allocate_par_spooling_space); 2778 // Now go around the loop and try alloc again; 2779 // A competing allocation might beat us to the expansion space, 2780 // so we may go around the loop again if allocation fails again. 2781 if (GCExpandToAllocateDelayMillis > 0) { 2782 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 2783 } 2784 } 2785 } 2786 2787 void ConcurrentMarkSweepGeneration::shrink(size_t bytes) { 2788 // Only shrink if a compaction was done so that all the free space 2789 // in the generation is in a contiguous block at the end. 2790 if (did_compact()) { 2791 CardGeneration::shrink(bytes); 2792 } 2793 } 2794 2795 void ConcurrentMarkSweepGeneration::assert_correct_size_change_locking() { 2796 assert_locked_or_safepoint(Heap_lock); 2797 } 2798 2799 void ConcurrentMarkSweepGeneration::shrink_free_list_by(size_t bytes) { 2800 assert_locked_or_safepoint(Heap_lock); 2801 assert_lock_strong(freelistLock()); 2802 if (PrintGCDetails && Verbose) { 2803 warning("Shrinking of CMS not yet implemented"); 2804 } 2805 return; 2806 } 2807 2808 2809 // Simple ctor/dtor wrapper for accounting & timer chores around concurrent 2810 // phases. 2811 class CMSPhaseAccounting: public StackObj { 2812 public: 2813 CMSPhaseAccounting(CMSCollector *collector, 2814 const char *phase, 2815 bool print_cr = true); 2816 ~CMSPhaseAccounting(); 2817 2818 private: 2819 CMSCollector *_collector; 2820 const char *_phase; 2821 elapsedTimer _wallclock; 2822 bool _print_cr; 2823 2824 public: 2825 // Not MT-safe; so do not pass around these StackObj's 2826 // where they may be accessed by other threads. 2827 jlong wallclock_millis() { 2828 assert(_wallclock.is_active(), "Wall clock should not stop"); 2829 _wallclock.stop(); // to record time 2830 jlong ret = _wallclock.milliseconds(); 2831 _wallclock.start(); // restart 2832 return ret; 2833 } 2834 }; 2835 2836 CMSPhaseAccounting::CMSPhaseAccounting(CMSCollector *collector, 2837 const char *phase, 2838 bool print_cr) : 2839 _collector(collector), _phase(phase), _print_cr(print_cr) { 2840 2841 if (PrintCMSStatistics != 0) { 2842 _collector->resetYields(); 2843 } 2844 if (PrintGCDetails) { 2845 gclog_or_tty->gclog_stamp(); 2846 gclog_or_tty->print_cr("[%s-concurrent-%s-start]", 2847 _collector->cmsGen()->short_name(), _phase); 2848 } 2849 _collector->resetTimer(); 2850 _wallclock.start(); 2851 _collector->startTimer(); 2852 } 2853 2854 CMSPhaseAccounting::~CMSPhaseAccounting() { 2855 assert(_wallclock.is_active(), "Wall clock should not have stopped"); 2856 _collector->stopTimer(); 2857 _wallclock.stop(); 2858 if (PrintGCDetails) { 2859 gclog_or_tty->gclog_stamp(); 2860 gclog_or_tty->print("[%s-concurrent-%s: %3.3f/%3.3f secs]", 2861 _collector->cmsGen()->short_name(), 2862 _phase, _collector->timerValue(), _wallclock.seconds()); 2863 if (_print_cr) { 2864 gclog_or_tty->cr(); 2865 } 2866 if (PrintCMSStatistics != 0) { 2867 gclog_or_tty->print_cr(" (CMS-concurrent-%s yielded %d times)", _phase, 2868 _collector->yields()); 2869 } 2870 } 2871 } 2872 2873 // CMS work 2874 2875 // The common parts of CMSParInitialMarkTask and CMSParRemarkTask. 2876 class CMSParMarkTask : public AbstractGangTask { 2877 protected: 2878 CMSCollector* _collector; 2879 uint _n_workers; 2880 CMSParMarkTask(const char* name, CMSCollector* collector, uint n_workers) : 2881 AbstractGangTask(name), 2882 _collector(collector), 2883 _n_workers(n_workers) {} 2884 // Work method in support of parallel rescan ... of young gen spaces 2885 void do_young_space_rescan(uint worker_id, OopsInGenClosure* cl, 2886 ContiguousSpace* space, 2887 HeapWord** chunk_array, size_t chunk_top); 2888 void work_on_young_gen_roots(uint worker_id, OopsInGenClosure* cl); 2889 }; 2890 2891 // Parallel initial mark task 2892 class CMSParInitialMarkTask: public CMSParMarkTask { 2893 StrongRootsScope* _strong_roots_scope; 2894 public: 2895 CMSParInitialMarkTask(CMSCollector* collector, StrongRootsScope* strong_roots_scope, uint n_workers) : 2896 CMSParMarkTask("Scan roots and young gen for initial mark in parallel", collector, n_workers), 2897 _strong_roots_scope(strong_roots_scope) {} 2898 void work(uint worker_id); 2899 }; 2900 2901 // Checkpoint the roots into this generation from outside 2902 // this generation. [Note this initial checkpoint need only 2903 // be approximate -- we'll do a catch up phase subsequently.] 2904 void CMSCollector::checkpointRootsInitial() { 2905 assert(_collectorState == InitialMarking, "Wrong collector state"); 2906 check_correct_thread_executing(); 2907 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 2908 2909 save_heap_summary(); 2910 report_heap_summary(GCWhen::BeforeGC); 2911 2912 ReferenceProcessor* rp = ref_processor(); 2913 assert(_restart_addr == NULL, "Control point invariant"); 2914 { 2915 // acquire locks for subsequent manipulations 2916 MutexLockerEx x(bitMapLock(), 2917 Mutex::_no_safepoint_check_flag); 2918 checkpointRootsInitialWork(); 2919 // enable ("weak") refs discovery 2920 rp->enable_discovery(); 2921 _collectorState = Marking; 2922 } 2923 } 2924 2925 void CMSCollector::checkpointRootsInitialWork() { 2926 assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped"); 2927 assert(_collectorState == InitialMarking, "just checking"); 2928 2929 // Already have locks. 2930 assert_lock_strong(bitMapLock()); 2931 assert(_markBitMap.isAllClear(), "was reset at end of previous cycle"); 2932 2933 // Setup the verification and class unloading state for this 2934 // CMS collection cycle. 2935 setup_cms_unloading_and_verification_state(); 2936 2937 NOT_PRODUCT(GCTraceTime t("\ncheckpointRootsInitialWork", 2938 PrintGCDetails && Verbose, true, _gc_timer_cm);) 2939 2940 // Reset all the PLAB chunk arrays if necessary. 2941 if (_survivor_plab_array != NULL && !CMSPLABRecordAlways) { 2942 reset_survivor_plab_arrays(); 2943 } 2944 2945 ResourceMark rm; 2946 HandleMark hm; 2947 2948 MarkRefsIntoClosure notOlder(_span, &_markBitMap); 2949 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2950 2951 verify_work_stacks_empty(); 2952 verify_overflow_empty(); 2953 2954 gch->ensure_parsability(false); // fill TLABs, but no need to retire them 2955 // Update the saved marks which may affect the root scans. 2956 gch->save_marks(); 2957 2958 // weak reference processing has not started yet. 2959 ref_processor()->set_enqueuing_is_done(false); 2960 2961 // Need to remember all newly created CLDs, 2962 // so that we can guarantee that the remark finds them. 2963 ClassLoaderDataGraph::remember_new_clds(true); 2964 2965 // Whenever a CLD is found, it will be claimed before proceeding to mark 2966 // the klasses. The claimed marks need to be cleared before marking starts. 2967 ClassLoaderDataGraph::clear_claimed_marks(); 2968 2969 if (CMSPrintEdenSurvivorChunks) { 2970 print_eden_and_survivor_chunk_arrays(); 2971 } 2972 2973 { 2974 #if defined(COMPILER2) || INCLUDE_JVMCI 2975 DerivedPointerTableDeactivate dpt_deact; 2976 #endif 2977 if (CMSParallelInitialMarkEnabled) { 2978 // The parallel version. 2979 WorkGang* workers = gch->workers(); 2980 assert(workers != NULL, "Need parallel worker threads."); 2981 uint n_workers = workers->active_workers(); 2982 2983 StrongRootsScope srs(n_workers); 2984 2985 CMSParInitialMarkTask tsk(this, &srs, n_workers); 2986 initialize_sequential_subtasks_for_young_gen_rescan(n_workers); 2987 if (n_workers > 1) { 2988 workers->run_task(&tsk); 2989 } else { 2990 tsk.work(0); 2991 } 2992 } else { 2993 // The serial version. 2994 CLDToOopClosure cld_closure(¬Older, true); 2995 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2996 2997 StrongRootsScope srs(1); 2998 2999 gch->gen_process_roots(&srs, 3000 GenCollectedHeap::OldGen, 3001 true, // young gen as roots 3002 GenCollectedHeap::ScanningOption(roots_scanning_options()), 3003 should_unload_classes(), 3004 ¬Older, 3005 NULL, 3006 &cld_closure); 3007 } 3008 } 3009 3010 // Clear mod-union table; it will be dirtied in the prologue of 3011 // CMS generation per each young generation collection. 3012 3013 assert(_modUnionTable.isAllClear(), 3014 "Was cleared in most recent final checkpoint phase" 3015 " or no bits are set in the gc_prologue before the start of the next " 3016 "subsequent marking phase."); 3017 3018 assert(_ct->klass_rem_set()->mod_union_is_clear(), "Must be"); 3019 3020 // Save the end of the used_region of the constituent generations 3021 // to be used to limit the extent of sweep in each generation. 3022 save_sweep_limits(); 3023 verify_overflow_empty(); 3024 } 3025 3026 bool CMSCollector::markFromRoots() { 3027 // we might be tempted to assert that: 3028 // assert(!SafepointSynchronize::is_at_safepoint(), 3029 // "inconsistent argument?"); 3030 // However that wouldn't be right, because it's possible that 3031 // a safepoint is indeed in progress as a young generation 3032 // stop-the-world GC happens even as we mark in this generation. 3033 assert(_collectorState == Marking, "inconsistent state?"); 3034 check_correct_thread_executing(); 3035 verify_overflow_empty(); 3036 3037 // Weak ref discovery note: We may be discovering weak 3038 // refs in this generation concurrent (but interleaved) with 3039 // weak ref discovery by the young generation collector. 3040 3041 CMSTokenSyncWithLocks ts(true, bitMapLock()); 3042 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 3043 CMSPhaseAccounting pa(this, "mark", !PrintGCDetails); 3044 bool res = markFromRootsWork(); 3045 if (res) { 3046 _collectorState = Precleaning; 3047 } else { // We failed and a foreground collection wants to take over 3048 assert(_foregroundGCIsActive, "internal state inconsistency"); 3049 assert(_restart_addr == NULL, "foreground will restart from scratch"); 3050 if (PrintGCDetails) { 3051 gclog_or_tty->print_cr("bailing out to foreground collection"); 3052 } 3053 } 3054 verify_overflow_empty(); 3055 return res; 3056 } 3057 3058 bool CMSCollector::markFromRootsWork() { 3059 // iterate over marked bits in bit map, doing a full scan and mark 3060 // from these roots using the following algorithm: 3061 // . if oop is to the right of the current scan pointer, 3062 // mark corresponding bit (we'll process it later) 3063 // . else (oop is to left of current scan pointer) 3064 // push oop on marking stack 3065 // . drain the marking stack 3066 3067 // Note that when we do a marking step we need to hold the 3068 // bit map lock -- recall that direct allocation (by mutators) 3069 // and promotion (by the young generation collector) is also 3070 // marking the bit map. [the so-called allocate live policy.] 3071 // Because the implementation of bit map marking is not 3072 // robust wrt simultaneous marking of bits in the same word, 3073 // we need to make sure that there is no such interference 3074 // between concurrent such updates. 3075 3076 // already have locks 3077 assert_lock_strong(bitMapLock()); 3078 3079 verify_work_stacks_empty(); 3080 verify_overflow_empty(); 3081 bool result = false; 3082 if (CMSConcurrentMTEnabled && ConcGCThreads > 0) { 3083 result = do_marking_mt(); 3084 } else { 3085 result = do_marking_st(); 3086 } 3087 return result; 3088 } 3089 3090 // Forward decl 3091 class CMSConcMarkingTask; 3092 3093 class CMSConcMarkingTerminator: public ParallelTaskTerminator { 3094 CMSCollector* _collector; 3095 CMSConcMarkingTask* _task; 3096 public: 3097 virtual void yield(); 3098 3099 // "n_threads" is the number of threads to be terminated. 3100 // "queue_set" is a set of work queues of other threads. 3101 // "collector" is the CMS collector associated with this task terminator. 3102 // "yield" indicates whether we need the gang as a whole to yield. 3103 CMSConcMarkingTerminator(int n_threads, TaskQueueSetSuper* queue_set, CMSCollector* collector) : 3104 ParallelTaskTerminator(n_threads, queue_set), 3105 _collector(collector) { } 3106 3107 void set_task(CMSConcMarkingTask* task) { 3108 _task = task; 3109 } 3110 }; 3111 3112 class CMSConcMarkingTerminatorTerminator: public TerminatorTerminator { 3113 CMSConcMarkingTask* _task; 3114 public: 3115 bool should_exit_termination(); 3116 void set_task(CMSConcMarkingTask* task) { 3117 _task = task; 3118 } 3119 }; 3120 3121 // MT Concurrent Marking Task 3122 class CMSConcMarkingTask: public YieldingFlexibleGangTask { 3123 CMSCollector* _collector; 3124 uint _n_workers; // requested/desired # workers 3125 bool _result; 3126 CompactibleFreeListSpace* _cms_space; 3127 char _pad_front[64]; // padding to ... 3128 HeapWord* _global_finger; // ... avoid sharing cache line 3129 char _pad_back[64]; 3130 HeapWord* _restart_addr; 3131 3132 // Exposed here for yielding support 3133 Mutex* const _bit_map_lock; 3134 3135 // The per thread work queues, available here for stealing 3136 OopTaskQueueSet* _task_queues; 3137 3138 // Termination (and yielding) support 3139 CMSConcMarkingTerminator _term; 3140 CMSConcMarkingTerminatorTerminator _term_term; 3141 3142 public: 3143 CMSConcMarkingTask(CMSCollector* collector, 3144 CompactibleFreeListSpace* cms_space, 3145 YieldingFlexibleWorkGang* workers, 3146 OopTaskQueueSet* task_queues): 3147 YieldingFlexibleGangTask("Concurrent marking done multi-threaded"), 3148 _collector(collector), 3149 _cms_space(cms_space), 3150 _n_workers(0), _result(true), 3151 _task_queues(task_queues), 3152 _term(_n_workers, task_queues, _collector), 3153 _bit_map_lock(collector->bitMapLock()) 3154 { 3155 _requested_size = _n_workers; 3156 _term.set_task(this); 3157 _term_term.set_task(this); 3158 _restart_addr = _global_finger = _cms_space->bottom(); 3159 } 3160 3161 3162 OopTaskQueueSet* task_queues() { return _task_queues; } 3163 3164 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 3165 3166 HeapWord** global_finger_addr() { return &_global_finger; } 3167 3168 CMSConcMarkingTerminator* terminator() { return &_term; } 3169 3170 virtual void set_for_termination(uint active_workers) { 3171 terminator()->reset_for_reuse(active_workers); 3172 } 3173 3174 void work(uint worker_id); 3175 bool should_yield() { 3176 return ConcurrentMarkSweepThread::should_yield() 3177 && !_collector->foregroundGCIsActive(); 3178 } 3179 3180 virtual void coordinator_yield(); // stuff done by coordinator 3181 bool result() { return _result; } 3182 3183 void reset(HeapWord* ra) { 3184 assert(_global_finger >= _cms_space->end(), "Postcondition of ::work(i)"); 3185 _restart_addr = _global_finger = ra; 3186 _term.reset_for_reuse(); 3187 } 3188 3189 static bool get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 3190 OopTaskQueue* work_q); 3191 3192 private: 3193 void do_scan_and_mark(int i, CompactibleFreeListSpace* sp); 3194 void do_work_steal(int i); 3195 void bump_global_finger(HeapWord* f); 3196 }; 3197 3198 bool CMSConcMarkingTerminatorTerminator::should_exit_termination() { 3199 assert(_task != NULL, "Error"); 3200 return _task->yielding(); 3201 // Note that we do not need the disjunct || _task->should_yield() above 3202 // because we want terminating threads to yield only if the task 3203 // is already in the midst of yielding, which happens only after at least one 3204 // thread has yielded. 3205 } 3206 3207 void CMSConcMarkingTerminator::yield() { 3208 if (_task->should_yield()) { 3209 _task->yield(); 3210 } else { 3211 ParallelTaskTerminator::yield(); 3212 } 3213 } 3214 3215 //////////////////////////////////////////////////////////////// 3216 // Concurrent Marking Algorithm Sketch 3217 //////////////////////////////////////////////////////////////// 3218 // Until all tasks exhausted (both spaces): 3219 // -- claim next available chunk 3220 // -- bump global finger via CAS 3221 // -- find first object that starts in this chunk 3222 // and start scanning bitmap from that position 3223 // -- scan marked objects for oops 3224 // -- CAS-mark target, and if successful: 3225 // . if target oop is above global finger (volatile read) 3226 // nothing to do 3227 // . if target oop is in chunk and above local finger 3228 // then nothing to do 3229 // . else push on work-queue 3230 // -- Deal with possible overflow issues: 3231 // . local work-queue overflow causes stuff to be pushed on 3232 // global (common) overflow queue 3233 // . always first empty local work queue 3234 // . then get a batch of oops from global work queue if any 3235 // . then do work stealing 3236 // -- When all tasks claimed (both spaces) 3237 // and local work queue empty, 3238 // then in a loop do: 3239 // . check global overflow stack; steal a batch of oops and trace 3240 // . try to steal from other threads oif GOS is empty 3241 // . if neither is available, offer termination 3242 // -- Terminate and return result 3243 // 3244 void CMSConcMarkingTask::work(uint worker_id) { 3245 elapsedTimer _timer; 3246 ResourceMark rm; 3247 HandleMark hm; 3248 3249 DEBUG_ONLY(_collector->verify_overflow_empty();) 3250 3251 // Before we begin work, our work queue should be empty 3252 assert(work_queue(worker_id)->size() == 0, "Expected to be empty"); 3253 // Scan the bitmap covering _cms_space, tracing through grey objects. 3254 _timer.start(); 3255 do_scan_and_mark(worker_id, _cms_space); 3256 _timer.stop(); 3257 if (PrintCMSStatistics != 0) { 3258 gclog_or_tty->print_cr("Finished cms space scanning in %dth thread: %3.3f sec", 3259 worker_id, _timer.seconds()); 3260 // XXX: need xxx/xxx type of notation, two timers 3261 } 3262 3263 // ... do work stealing 3264 _timer.reset(); 3265 _timer.start(); 3266 do_work_steal(worker_id); 3267 _timer.stop(); 3268 if (PrintCMSStatistics != 0) { 3269 gclog_or_tty->print_cr("Finished work stealing in %dth thread: %3.3f sec", 3270 worker_id, _timer.seconds()); 3271 // XXX: need xxx/xxx type of notation, two timers 3272 } 3273 assert(_collector->_markStack.isEmpty(), "Should have been emptied"); 3274 assert(work_queue(worker_id)->size() == 0, "Should have been emptied"); 3275 // Note that under the current task protocol, the 3276 // following assertion is true even of the spaces 3277 // expanded since the completion of the concurrent 3278 // marking. XXX This will likely change under a strict 3279 // ABORT semantics. 3280 // After perm removal the comparison was changed to 3281 // greater than or equal to from strictly greater than. 3282 // Before perm removal the highest address sweep would 3283 // have been at the end of perm gen but now is at the 3284 // end of the tenured gen. 3285 assert(_global_finger >= _cms_space->end(), 3286 "All tasks have been completed"); 3287 DEBUG_ONLY(_collector->verify_overflow_empty();) 3288 } 3289 3290 void CMSConcMarkingTask::bump_global_finger(HeapWord* f) { 3291 HeapWord* read = _global_finger; 3292 HeapWord* cur = read; 3293 while (f > read) { 3294 cur = read; 3295 read = (HeapWord*) Atomic::cmpxchg_ptr(f, &_global_finger, cur); 3296 if (cur == read) { 3297 // our cas succeeded 3298 assert(_global_finger >= f, "protocol consistency"); 3299 break; 3300 } 3301 } 3302 } 3303 3304 // This is really inefficient, and should be redone by 3305 // using (not yet available) block-read and -write interfaces to the 3306 // stack and the work_queue. XXX FIX ME !!! 3307 bool CMSConcMarkingTask::get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 3308 OopTaskQueue* work_q) { 3309 // Fast lock-free check 3310 if (ovflw_stk->length() == 0) { 3311 return false; 3312 } 3313 assert(work_q->size() == 0, "Shouldn't steal"); 3314 MutexLockerEx ml(ovflw_stk->par_lock(), 3315 Mutex::_no_safepoint_check_flag); 3316 // Grab up to 1/4 the size of the work queue 3317 size_t num = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 3318 (size_t)ParGCDesiredObjsFromOverflowList); 3319 num = MIN2(num, ovflw_stk->length()); 3320 for (int i = (int) num; i > 0; i--) { 3321 oop cur = ovflw_stk->pop(); 3322 assert(cur != NULL, "Counted wrong?"); 3323 work_q->push(cur); 3324 } 3325 return num > 0; 3326 } 3327 3328 void CMSConcMarkingTask::do_scan_and_mark(int i, CompactibleFreeListSpace* sp) { 3329 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 3330 int n_tasks = pst->n_tasks(); 3331 // We allow that there may be no tasks to do here because 3332 // we are restarting after a stack overflow. 3333 assert(pst->valid() || n_tasks == 0, "Uninitialized use?"); 3334 uint nth_task = 0; 3335 3336 HeapWord* aligned_start = sp->bottom(); 3337 if (sp->used_region().contains(_restart_addr)) { 3338 // Align down to a card boundary for the start of 0th task 3339 // for this space. 3340 aligned_start = 3341 (HeapWord*)align_size_down((uintptr_t)_restart_addr, 3342 CardTableModRefBS::card_size); 3343 } 3344 3345 size_t chunk_size = sp->marking_task_size(); 3346 while (!pst->is_task_claimed(/* reference */ nth_task)) { 3347 // Having claimed the nth task in this space, 3348 // compute the chunk that it corresponds to: 3349 MemRegion span = MemRegion(aligned_start + nth_task*chunk_size, 3350 aligned_start + (nth_task+1)*chunk_size); 3351 // Try and bump the global finger via a CAS; 3352 // note that we need to do the global finger bump 3353 // _before_ taking the intersection below, because 3354 // the task corresponding to that region will be 3355 // deemed done even if the used_region() expands 3356 // because of allocation -- as it almost certainly will 3357 // during start-up while the threads yield in the 3358 // closure below. 3359 HeapWord* finger = span.end(); 3360 bump_global_finger(finger); // atomically 3361 // There are null tasks here corresponding to chunks 3362 // beyond the "top" address of the space. 3363 span = span.intersection(sp->used_region()); 3364 if (!span.is_empty()) { // Non-null task 3365 HeapWord* prev_obj; 3366 assert(!span.contains(_restart_addr) || nth_task == 0, 3367 "Inconsistency"); 3368 if (nth_task == 0) { 3369 // For the 0th task, we'll not need to compute a block_start. 3370 if (span.contains(_restart_addr)) { 3371 // In the case of a restart because of stack overflow, 3372 // we might additionally skip a chunk prefix. 3373 prev_obj = _restart_addr; 3374 } else { 3375 prev_obj = span.start(); 3376 } 3377 } else { 3378 // We want to skip the first object because 3379 // the protocol is to scan any object in its entirety 3380 // that _starts_ in this span; a fortiori, any 3381 // object starting in an earlier span is scanned 3382 // as part of an earlier claimed task. 3383 // Below we use the "careful" version of block_start 3384 // so we do not try to navigate uninitialized objects. 3385 prev_obj = sp->block_start_careful(span.start()); 3386 // Below we use a variant of block_size that uses the 3387 // Printezis bits to avoid waiting for allocated 3388 // objects to become initialized/parsable. 3389 while (prev_obj < span.start()) { 3390 size_t sz = sp->block_size_no_stall(prev_obj, _collector); 3391 if (sz > 0) { 3392 prev_obj += sz; 3393 } else { 3394 // In this case we may end up doing a bit of redundant 3395 // scanning, but that appears unavoidable, short of 3396 // locking the free list locks; see bug 6324141. 3397 break; 3398 } 3399 } 3400 } 3401 if (prev_obj < span.end()) { 3402 MemRegion my_span = MemRegion(prev_obj, span.end()); 3403 // Do the marking work within a non-empty span -- 3404 // the last argument to the constructor indicates whether the 3405 // iteration should be incremental with periodic yields. 3406 Par_MarkFromRootsClosure cl(this, _collector, my_span, 3407 &_collector->_markBitMap, 3408 work_queue(i), 3409 &_collector->_markStack); 3410 _collector->_markBitMap.iterate(&cl, my_span.start(), my_span.end()); 3411 } // else nothing to do for this task 3412 } // else nothing to do for this task 3413 } 3414 // We'd be tempted to assert here that since there are no 3415 // more tasks left to claim in this space, the global_finger 3416 // must exceed space->top() and a fortiori space->end(). However, 3417 // that would not quite be correct because the bumping of 3418 // global_finger occurs strictly after the claiming of a task, 3419 // so by the time we reach here the global finger may not yet 3420 // have been bumped up by the thread that claimed the last 3421 // task. 3422 pst->all_tasks_completed(); 3423 } 3424 3425 class Par_ConcMarkingClosure: public MetadataAwareOopClosure { 3426 private: 3427 CMSCollector* _collector; 3428 CMSConcMarkingTask* _task; 3429 MemRegion _span; 3430 CMSBitMap* _bit_map; 3431 CMSMarkStack* _overflow_stack; 3432 OopTaskQueue* _work_queue; 3433 protected: 3434 DO_OOP_WORK_DEFN 3435 public: 3436 Par_ConcMarkingClosure(CMSCollector* collector, CMSConcMarkingTask* task, OopTaskQueue* work_queue, 3437 CMSBitMap* bit_map, CMSMarkStack* overflow_stack): 3438 MetadataAwareOopClosure(collector->ref_processor()), 3439 _collector(collector), 3440 _task(task), 3441 _span(collector->_span), 3442 _work_queue(work_queue), 3443 _bit_map(bit_map), 3444 _overflow_stack(overflow_stack) 3445 { } 3446 virtual void do_oop(oop* p); 3447 virtual void do_oop(narrowOop* p); 3448 3449 void trim_queue(size_t max); 3450 void handle_stack_overflow(HeapWord* lost); 3451 void do_yield_check() { 3452 if (_task->should_yield()) { 3453 _task->yield(); 3454 } 3455 } 3456 }; 3457 3458 // Grey object scanning during work stealing phase -- 3459 // the salient assumption here is that any references 3460 // that are in these stolen objects being scanned must 3461 // already have been initialized (else they would not have 3462 // been published), so we do not need to check for 3463 // uninitialized objects before pushing here. 3464 void Par_ConcMarkingClosure::do_oop(oop obj) { 3465 assert(obj->is_oop_or_null(true), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 3466 HeapWord* addr = (HeapWord*)obj; 3467 // Check if oop points into the CMS generation 3468 // and is not marked 3469 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 3470 // a white object ... 3471 // If we manage to "claim" the object, by being the 3472 // first thread to mark it, then we push it on our 3473 // marking stack 3474 if (_bit_map->par_mark(addr)) { // ... now grey 3475 // push on work queue (grey set) 3476 bool simulate_overflow = false; 3477 NOT_PRODUCT( 3478 if (CMSMarkStackOverflowALot && 3479 _collector->simulate_overflow()) { 3480 // simulate a stack overflow 3481 simulate_overflow = true; 3482 } 3483 ) 3484 if (simulate_overflow || 3485 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) { 3486 // stack overflow 3487 if (PrintCMSStatistics != 0) { 3488 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 3489 SIZE_FORMAT, _overflow_stack->capacity()); 3490 } 3491 // We cannot assert that the overflow stack is full because 3492 // it may have been emptied since. 3493 assert(simulate_overflow || 3494 _work_queue->size() == _work_queue->max_elems(), 3495 "Else push should have succeeded"); 3496 handle_stack_overflow(addr); 3497 } 3498 } // Else, some other thread got there first 3499 do_yield_check(); 3500 } 3501 } 3502 3503 void Par_ConcMarkingClosure::do_oop(oop* p) { Par_ConcMarkingClosure::do_oop_work(p); } 3504 void Par_ConcMarkingClosure::do_oop(narrowOop* p) { Par_ConcMarkingClosure::do_oop_work(p); } 3505 3506 void Par_ConcMarkingClosure::trim_queue(size_t max) { 3507 while (_work_queue->size() > max) { 3508 oop new_oop; 3509 if (_work_queue->pop_local(new_oop)) { 3510 assert(new_oop->is_oop(), "Should be an oop"); 3511 assert(_bit_map->isMarked((HeapWord*)new_oop), "Grey object"); 3512 assert(_span.contains((HeapWord*)new_oop), "Not in span"); 3513 new_oop->oop_iterate(this); // do_oop() above 3514 do_yield_check(); 3515 } 3516 } 3517 } 3518 3519 // Upon stack overflow, we discard (part of) the stack, 3520 // remembering the least address amongst those discarded 3521 // in CMSCollector's _restart_address. 3522 void Par_ConcMarkingClosure::handle_stack_overflow(HeapWord* lost) { 3523 // We need to do this under a mutex to prevent other 3524 // workers from interfering with the work done below. 3525 MutexLockerEx ml(_overflow_stack->par_lock(), 3526 Mutex::_no_safepoint_check_flag); 3527 // Remember the least grey address discarded 3528 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 3529 _collector->lower_restart_addr(ra); 3530 _overflow_stack->reset(); // discard stack contents 3531 _overflow_stack->expand(); // expand the stack if possible 3532 } 3533 3534 3535 void CMSConcMarkingTask::do_work_steal(int i) { 3536 OopTaskQueue* work_q = work_queue(i); 3537 oop obj_to_scan; 3538 CMSBitMap* bm = &(_collector->_markBitMap); 3539 CMSMarkStack* ovflw = &(_collector->_markStack); 3540 int* seed = _collector->hash_seed(i); 3541 Par_ConcMarkingClosure cl(_collector, this, work_q, bm, ovflw); 3542 while (true) { 3543 cl.trim_queue(0); 3544 assert(work_q->size() == 0, "Should have been emptied above"); 3545 if (get_work_from_overflow_stack(ovflw, work_q)) { 3546 // Can't assert below because the work obtained from the 3547 // overflow stack may already have been stolen from us. 3548 // assert(work_q->size() > 0, "Work from overflow stack"); 3549 continue; 3550 } else if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 3551 assert(obj_to_scan->is_oop(), "Should be an oop"); 3552 assert(bm->isMarked((HeapWord*)obj_to_scan), "Grey object"); 3553 obj_to_scan->oop_iterate(&cl); 3554 } else if (terminator()->offer_termination(&_term_term)) { 3555 assert(work_q->size() == 0, "Impossible!"); 3556 break; 3557 } else if (yielding() || should_yield()) { 3558 yield(); 3559 } 3560 } 3561 } 3562 3563 // This is run by the CMS (coordinator) thread. 3564 void CMSConcMarkingTask::coordinator_yield() { 3565 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 3566 "CMS thread should hold CMS token"); 3567 // First give up the locks, then yield, then re-lock 3568 // We should probably use a constructor/destructor idiom to 3569 // do this unlock/lock or modify the MutexUnlocker class to 3570 // serve our purpose. XXX 3571 assert_lock_strong(_bit_map_lock); 3572 _bit_map_lock->unlock(); 3573 ConcurrentMarkSweepThread::desynchronize(true); 3574 _collector->stopTimer(); 3575 if (PrintCMSStatistics != 0) { 3576 _collector->incrementYields(); 3577 } 3578 3579 // It is possible for whichever thread initiated the yield request 3580 // not to get a chance to wake up and take the bitmap lock between 3581 // this thread releasing it and reacquiring it. So, while the 3582 // should_yield() flag is on, let's sleep for a bit to give the 3583 // other thread a chance to wake up. The limit imposed on the number 3584 // of iterations is defensive, to avoid any unforseen circumstances 3585 // putting us into an infinite loop. Since it's always been this 3586 // (coordinator_yield()) method that was observed to cause the 3587 // problem, we are using a parameter (CMSCoordinatorYieldSleepCount) 3588 // which is by default non-zero. For the other seven methods that 3589 // also perform the yield operation, as are using a different 3590 // parameter (CMSYieldSleepCount) which is by default zero. This way we 3591 // can enable the sleeping for those methods too, if necessary. 3592 // See 6442774. 3593 // 3594 // We really need to reconsider the synchronization between the GC 3595 // thread and the yield-requesting threads in the future and we 3596 // should really use wait/notify, which is the recommended 3597 // way of doing this type of interaction. Additionally, we should 3598 // consolidate the eight methods that do the yield operation and they 3599 // are almost identical into one for better maintainability and 3600 // readability. See 6445193. 3601 // 3602 // Tony 2006.06.29 3603 for (unsigned i = 0; i < CMSCoordinatorYieldSleepCount && 3604 ConcurrentMarkSweepThread::should_yield() && 3605 !CMSCollector::foregroundGCIsActive(); ++i) { 3606 os::sleep(Thread::current(), 1, false); 3607 } 3608 3609 ConcurrentMarkSweepThread::synchronize(true); 3610 _bit_map_lock->lock_without_safepoint_check(); 3611 _collector->startTimer(); 3612 } 3613 3614 bool CMSCollector::do_marking_mt() { 3615 assert(ConcGCThreads > 0 && conc_workers() != NULL, "precondition"); 3616 uint num_workers = AdaptiveSizePolicy::calc_active_conc_workers(conc_workers()->total_workers(), 3617 conc_workers()->active_workers(), 3618 Threads::number_of_non_daemon_threads()); 3619 conc_workers()->set_active_workers(num_workers); 3620 3621 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 3622 3623 CMSConcMarkingTask tsk(this, 3624 cms_space, 3625 conc_workers(), 3626 task_queues()); 3627 3628 // Since the actual number of workers we get may be different 3629 // from the number we requested above, do we need to do anything different 3630 // below? In particular, may be we need to subclass the SequantialSubTasksDone 3631 // class?? XXX 3632 cms_space ->initialize_sequential_subtasks_for_marking(num_workers); 3633 3634 // Refs discovery is already non-atomic. 3635 assert(!ref_processor()->discovery_is_atomic(), "Should be non-atomic"); 3636 assert(ref_processor()->discovery_is_mt(), "Discovery should be MT"); 3637 conc_workers()->start_task(&tsk); 3638 while (tsk.yielded()) { 3639 tsk.coordinator_yield(); 3640 conc_workers()->continue_task(&tsk); 3641 } 3642 // If the task was aborted, _restart_addr will be non-NULL 3643 assert(tsk.completed() || _restart_addr != NULL, "Inconsistency"); 3644 while (_restart_addr != NULL) { 3645 // XXX For now we do not make use of ABORTED state and have not 3646 // yet implemented the right abort semantics (even in the original 3647 // single-threaded CMS case). That needs some more investigation 3648 // and is deferred for now; see CR# TBF. 07252005YSR. XXX 3649 assert(!CMSAbortSemantics || tsk.aborted(), "Inconsistency"); 3650 // If _restart_addr is non-NULL, a marking stack overflow 3651 // occurred; we need to do a fresh marking iteration from the 3652 // indicated restart address. 3653 if (_foregroundGCIsActive) { 3654 // We may be running into repeated stack overflows, having 3655 // reached the limit of the stack size, while making very 3656 // slow forward progress. It may be best to bail out and 3657 // let the foreground collector do its job. 3658 // Clear _restart_addr, so that foreground GC 3659 // works from scratch. This avoids the headache of 3660 // a "rescan" which would otherwise be needed because 3661 // of the dirty mod union table & card table. 3662 _restart_addr = NULL; 3663 return false; 3664 } 3665 // Adjust the task to restart from _restart_addr 3666 tsk.reset(_restart_addr); 3667 cms_space ->initialize_sequential_subtasks_for_marking(num_workers, 3668 _restart_addr); 3669 _restart_addr = NULL; 3670 // Get the workers going again 3671 conc_workers()->start_task(&tsk); 3672 while (tsk.yielded()) { 3673 tsk.coordinator_yield(); 3674 conc_workers()->continue_task(&tsk); 3675 } 3676 } 3677 assert(tsk.completed(), "Inconsistency"); 3678 assert(tsk.result() == true, "Inconsistency"); 3679 return true; 3680 } 3681 3682 bool CMSCollector::do_marking_st() { 3683 ResourceMark rm; 3684 HandleMark hm; 3685 3686 // Temporarily make refs discovery single threaded (non-MT) 3687 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false); 3688 MarkFromRootsClosure markFromRootsClosure(this, _span, &_markBitMap, 3689 &_markStack, CMSYield); 3690 // the last argument to iterate indicates whether the iteration 3691 // should be incremental with periodic yields. 3692 _markBitMap.iterate(&markFromRootsClosure); 3693 // If _restart_addr is non-NULL, a marking stack overflow 3694 // occurred; we need to do a fresh iteration from the 3695 // indicated restart address. 3696 while (_restart_addr != NULL) { 3697 if (_foregroundGCIsActive) { 3698 // We may be running into repeated stack overflows, having 3699 // reached the limit of the stack size, while making very 3700 // slow forward progress. It may be best to bail out and 3701 // let the foreground collector do its job. 3702 // Clear _restart_addr, so that foreground GC 3703 // works from scratch. This avoids the headache of 3704 // a "rescan" which would otherwise be needed because 3705 // of the dirty mod union table & card table. 3706 _restart_addr = NULL; 3707 return false; // indicating failure to complete marking 3708 } 3709 // Deal with stack overflow: 3710 // we restart marking from _restart_addr 3711 HeapWord* ra = _restart_addr; 3712 markFromRootsClosure.reset(ra); 3713 _restart_addr = NULL; 3714 _markBitMap.iterate(&markFromRootsClosure, ra, _span.end()); 3715 } 3716 return true; 3717 } 3718 3719 void CMSCollector::preclean() { 3720 check_correct_thread_executing(); 3721 assert(Thread::current()->is_ConcurrentGC_thread(), "Wrong thread"); 3722 verify_work_stacks_empty(); 3723 verify_overflow_empty(); 3724 _abort_preclean = false; 3725 if (CMSPrecleaningEnabled) { 3726 if (!CMSEdenChunksRecordAlways) { 3727 _eden_chunk_index = 0; 3728 } 3729 size_t used = get_eden_used(); 3730 size_t capacity = get_eden_capacity(); 3731 // Don't start sampling unless we will get sufficiently 3732 // many samples. 3733 if (used < (capacity/(CMSScheduleRemarkSamplingRatio * 100) 3734 * CMSScheduleRemarkEdenPenetration)) { 3735 _start_sampling = true; 3736 } else { 3737 _start_sampling = false; 3738 } 3739 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 3740 CMSPhaseAccounting pa(this, "preclean", !PrintGCDetails); 3741 preclean_work(CMSPrecleanRefLists1, CMSPrecleanSurvivors1); 3742 } 3743 CMSTokenSync x(true); // is cms thread 3744 if (CMSPrecleaningEnabled) { 3745 sample_eden(); 3746 _collectorState = AbortablePreclean; 3747 } else { 3748 _collectorState = FinalMarking; 3749 } 3750 verify_work_stacks_empty(); 3751 verify_overflow_empty(); 3752 } 3753 3754 // Try and schedule the remark such that young gen 3755 // occupancy is CMSScheduleRemarkEdenPenetration %. 3756 void CMSCollector::abortable_preclean() { 3757 check_correct_thread_executing(); 3758 assert(CMSPrecleaningEnabled, "Inconsistent control state"); 3759 assert(_collectorState == AbortablePreclean, "Inconsistent control state"); 3760 3761 // If Eden's current occupancy is below this threshold, 3762 // immediately schedule the remark; else preclean 3763 // past the next scavenge in an effort to 3764 // schedule the pause as described above. By choosing 3765 // CMSScheduleRemarkEdenSizeThreshold >= max eden size 3766 // we will never do an actual abortable preclean cycle. 3767 if (get_eden_used() > CMSScheduleRemarkEdenSizeThreshold) { 3768 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 3769 CMSPhaseAccounting pa(this, "abortable-preclean", !PrintGCDetails); 3770 // We need more smarts in the abortable preclean 3771 // loop below to deal with cases where allocation 3772 // in young gen is very very slow, and our precleaning 3773 // is running a losing race against a horde of 3774 // mutators intent on flooding us with CMS updates 3775 // (dirty cards). 3776 // One, admittedly dumb, strategy is to give up 3777 // after a certain number of abortable precleaning loops 3778 // or after a certain maximum time. We want to make 3779 // this smarter in the next iteration. 3780 // XXX FIX ME!!! YSR 3781 size_t loops = 0, workdone = 0, cumworkdone = 0, waited = 0; 3782 while (!(should_abort_preclean() || 3783 ConcurrentMarkSweepThread::should_terminate())) { 3784 workdone = preclean_work(CMSPrecleanRefLists2, CMSPrecleanSurvivors2); 3785 cumworkdone += workdone; 3786 loops++; 3787 // Voluntarily terminate abortable preclean phase if we have 3788 // been at it for too long. 3789 if ((CMSMaxAbortablePrecleanLoops != 0) && 3790 loops >= CMSMaxAbortablePrecleanLoops) { 3791 if (PrintGCDetails) { 3792 gclog_or_tty->print(" CMS: abort preclean due to loops "); 3793 } 3794 break; 3795 } 3796 if (pa.wallclock_millis() > CMSMaxAbortablePrecleanTime) { 3797 if (PrintGCDetails) { 3798 gclog_or_tty->print(" CMS: abort preclean due to time "); 3799 } 3800 break; 3801 } 3802 // If we are doing little work each iteration, we should 3803 // take a short break. 3804 if (workdone < CMSAbortablePrecleanMinWorkPerIteration) { 3805 // Sleep for some time, waiting for work to accumulate 3806 stopTimer(); 3807 cmsThread()->wait_on_cms_lock(CMSAbortablePrecleanWaitMillis); 3808 startTimer(); 3809 waited++; 3810 } 3811 } 3812 if (PrintCMSStatistics > 0) { 3813 gclog_or_tty->print(" [" SIZE_FORMAT " iterations, " SIZE_FORMAT " waits, " SIZE_FORMAT " cards)] ", 3814 loops, waited, cumworkdone); 3815 } 3816 } 3817 CMSTokenSync x(true); // is cms thread 3818 if (_collectorState != Idling) { 3819 assert(_collectorState == AbortablePreclean, 3820 "Spontaneous state transition?"); 3821 _collectorState = FinalMarking; 3822 } // Else, a foreground collection completed this CMS cycle. 3823 return; 3824 } 3825 3826 // Respond to an Eden sampling opportunity 3827 void CMSCollector::sample_eden() { 3828 // Make sure a young gc cannot sneak in between our 3829 // reading and recording of a sample. 3830 assert(Thread::current()->is_ConcurrentGC_thread(), 3831 "Only the cms thread may collect Eden samples"); 3832 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 3833 "Should collect samples while holding CMS token"); 3834 if (!_start_sampling) { 3835 return; 3836 } 3837 // When CMSEdenChunksRecordAlways is true, the eden chunk array 3838 // is populated by the young generation. 3839 if (_eden_chunk_array != NULL && !CMSEdenChunksRecordAlways) { 3840 if (_eden_chunk_index < _eden_chunk_capacity) { 3841 _eden_chunk_array[_eden_chunk_index] = *_top_addr; // take sample 3842 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr, 3843 "Unexpected state of Eden"); 3844 // We'd like to check that what we just sampled is an oop-start address; 3845 // however, we cannot do that here since the object may not yet have been 3846 // initialized. So we'll instead do the check when we _use_ this sample 3847 // later. 3848 if (_eden_chunk_index == 0 || 3849 (pointer_delta(_eden_chunk_array[_eden_chunk_index], 3850 _eden_chunk_array[_eden_chunk_index-1]) 3851 >= CMSSamplingGrain)) { 3852 _eden_chunk_index++; // commit sample 3853 } 3854 } 3855 } 3856 if ((_collectorState == AbortablePreclean) && !_abort_preclean) { 3857 size_t used = get_eden_used(); 3858 size_t capacity = get_eden_capacity(); 3859 assert(used <= capacity, "Unexpected state of Eden"); 3860 if (used > (capacity/100 * CMSScheduleRemarkEdenPenetration)) { 3861 _abort_preclean = true; 3862 } 3863 } 3864 } 3865 3866 3867 size_t CMSCollector::preclean_work(bool clean_refs, bool clean_survivor) { 3868 assert(_collectorState == Precleaning || 3869 _collectorState == AbortablePreclean, "incorrect state"); 3870 ResourceMark rm; 3871 HandleMark hm; 3872 3873 // Precleaning is currently not MT but the reference processor 3874 // may be set for MT. Disable it temporarily here. 3875 ReferenceProcessor* rp = ref_processor(); 3876 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(rp, false); 3877 3878 // Do one pass of scrubbing the discovered reference lists 3879 // to remove any reference objects with strongly-reachable 3880 // referents. 3881 if (clean_refs) { 3882 CMSPrecleanRefsYieldClosure yield_cl(this); 3883 assert(rp->span().equals(_span), "Spans should be equal"); 3884 CMSKeepAliveClosure keep_alive(this, _span, &_markBitMap, 3885 &_markStack, true /* preclean */); 3886 CMSDrainMarkingStackClosure complete_trace(this, 3887 _span, &_markBitMap, &_markStack, 3888 &keep_alive, true /* preclean */); 3889 3890 // We don't want this step to interfere with a young 3891 // collection because we don't want to take CPU 3892 // or memory bandwidth away from the young GC threads 3893 // (which may be as many as there are CPUs). 3894 // Note that we don't need to protect ourselves from 3895 // interference with mutators because they can't 3896 // manipulate the discovered reference lists nor affect 3897 // the computed reachability of the referents, the 3898 // only properties manipulated by the precleaning 3899 // of these reference lists. 3900 stopTimer(); 3901 CMSTokenSyncWithLocks x(true /* is cms thread */, 3902 bitMapLock()); 3903 startTimer(); 3904 sample_eden(); 3905 3906 // The following will yield to allow foreground 3907 // collection to proceed promptly. XXX YSR: 3908 // The code in this method may need further 3909 // tweaking for better performance and some restructuring 3910 // for cleaner interfaces. 3911 GCTimer *gc_timer = NULL; // Currently not tracing concurrent phases 3912 rp->preclean_discovered_references( 3913 rp->is_alive_non_header(), &keep_alive, &complete_trace, &yield_cl, 3914 gc_timer); 3915 } 3916 3917 if (clean_survivor) { // preclean the active survivor space(s) 3918 PushAndMarkClosure pam_cl(this, _span, ref_processor(), 3919 &_markBitMap, &_modUnionTable, 3920 &_markStack, true /* precleaning phase */); 3921 stopTimer(); 3922 CMSTokenSyncWithLocks ts(true /* is cms thread */, 3923 bitMapLock()); 3924 startTimer(); 3925 unsigned int before_count = 3926 GenCollectedHeap::heap()->total_collections(); 3927 SurvivorSpacePrecleanClosure 3928 sss_cl(this, _span, &_markBitMap, &_markStack, 3929 &pam_cl, before_count, CMSYield); 3930 _young_gen->from()->object_iterate_careful(&sss_cl); 3931 _young_gen->to()->object_iterate_careful(&sss_cl); 3932 } 3933 MarkRefsIntoAndScanClosure 3934 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable, 3935 &_markStack, this, CMSYield, 3936 true /* precleaning phase */); 3937 // CAUTION: The following closure has persistent state that may need to 3938 // be reset upon a decrease in the sequence of addresses it 3939 // processes. 3940 ScanMarkedObjectsAgainCarefullyClosure 3941 smoac_cl(this, _span, 3942 &_markBitMap, &_markStack, &mrias_cl, CMSYield); 3943 3944 // Preclean dirty cards in ModUnionTable and CardTable using 3945 // appropriate convergence criterion; 3946 // repeat CMSPrecleanIter times unless we find that 3947 // we are losing. 3948 assert(CMSPrecleanIter < 10, "CMSPrecleanIter is too large"); 3949 assert(CMSPrecleanNumerator < CMSPrecleanDenominator, 3950 "Bad convergence multiplier"); 3951 assert(CMSPrecleanThreshold >= 100, 3952 "Unreasonably low CMSPrecleanThreshold"); 3953 3954 size_t numIter, cumNumCards, lastNumCards, curNumCards; 3955 for (numIter = 0, cumNumCards = lastNumCards = curNumCards = 0; 3956 numIter < CMSPrecleanIter; 3957 numIter++, lastNumCards = curNumCards, cumNumCards += curNumCards) { 3958 curNumCards = preclean_mod_union_table(_cmsGen, &smoac_cl); 3959 if (Verbose && PrintGCDetails) { 3960 gclog_or_tty->print(" (modUnionTable: " SIZE_FORMAT " cards)", curNumCards); 3961 } 3962 // Either there are very few dirty cards, so re-mark 3963 // pause will be small anyway, or our pre-cleaning isn't 3964 // that much faster than the rate at which cards are being 3965 // dirtied, so we might as well stop and re-mark since 3966 // precleaning won't improve our re-mark time by much. 3967 if (curNumCards <= CMSPrecleanThreshold || 3968 (numIter > 0 && 3969 (curNumCards * CMSPrecleanDenominator > 3970 lastNumCards * CMSPrecleanNumerator))) { 3971 numIter++; 3972 cumNumCards += curNumCards; 3973 break; 3974 } 3975 } 3976 3977 preclean_klasses(&mrias_cl, _cmsGen->freelistLock()); 3978 3979 curNumCards = preclean_card_table(_cmsGen, &smoac_cl); 3980 cumNumCards += curNumCards; 3981 if (PrintGCDetails && PrintCMSStatistics != 0) { 3982 gclog_or_tty->print_cr(" (cardTable: " SIZE_FORMAT " cards, re-scanned " SIZE_FORMAT " cards, " SIZE_FORMAT " iterations)", 3983 curNumCards, cumNumCards, numIter); 3984 } 3985 return cumNumCards; // as a measure of useful work done 3986 } 3987 3988 // PRECLEANING NOTES: 3989 // Precleaning involves: 3990 // . reading the bits of the modUnionTable and clearing the set bits. 3991 // . For the cards corresponding to the set bits, we scan the 3992 // objects on those cards. This means we need the free_list_lock 3993 // so that we can safely iterate over the CMS space when scanning 3994 // for oops. 3995 // . When we scan the objects, we'll be both reading and setting 3996 // marks in the marking bit map, so we'll need the marking bit map. 3997 // . For protecting _collector_state transitions, we take the CGC_lock. 3998 // Note that any races in the reading of of card table entries by the 3999 // CMS thread on the one hand and the clearing of those entries by the 4000 // VM thread or the setting of those entries by the mutator threads on the 4001 // other are quite benign. However, for efficiency it makes sense to keep 4002 // the VM thread from racing with the CMS thread while the latter is 4003 // dirty card info to the modUnionTable. We therefore also use the 4004 // CGC_lock to protect the reading of the card table and the mod union 4005 // table by the CM thread. 4006 // . We run concurrently with mutator updates, so scanning 4007 // needs to be done carefully -- we should not try to scan 4008 // potentially uninitialized objects. 4009 // 4010 // Locking strategy: While holding the CGC_lock, we scan over and 4011 // reset a maximal dirty range of the mod union / card tables, then lock 4012 // the free_list_lock and bitmap lock to do a full marking, then 4013 // release these locks; and repeat the cycle. This allows for a 4014 // certain amount of fairness in the sharing of these locks between 4015 // the CMS collector on the one hand, and the VM thread and the 4016 // mutators on the other. 4017 4018 // NOTE: preclean_mod_union_table() and preclean_card_table() 4019 // further below are largely identical; if you need to modify 4020 // one of these methods, please check the other method too. 4021 4022 size_t CMSCollector::preclean_mod_union_table( 4023 ConcurrentMarkSweepGeneration* old_gen, 4024 ScanMarkedObjectsAgainCarefullyClosure* cl) { 4025 verify_work_stacks_empty(); 4026 verify_overflow_empty(); 4027 4028 // strategy: starting with the first card, accumulate contiguous 4029 // ranges of dirty cards; clear these cards, then scan the region 4030 // covered by these cards. 4031 4032 // Since all of the MUT is committed ahead, we can just use 4033 // that, in case the generations expand while we are precleaning. 4034 // It might also be fine to just use the committed part of the 4035 // generation, but we might potentially miss cards when the 4036 // generation is rapidly expanding while we are in the midst 4037 // of precleaning. 4038 HeapWord* startAddr = old_gen->reserved().start(); 4039 HeapWord* endAddr = old_gen->reserved().end(); 4040 4041 cl->setFreelistLock(old_gen->freelistLock()); // needed for yielding 4042 4043 size_t numDirtyCards, cumNumDirtyCards; 4044 HeapWord *nextAddr, *lastAddr; 4045 for (cumNumDirtyCards = numDirtyCards = 0, 4046 nextAddr = lastAddr = startAddr; 4047 nextAddr < endAddr; 4048 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 4049 4050 ResourceMark rm; 4051 HandleMark hm; 4052 4053 MemRegion dirtyRegion; 4054 { 4055 stopTimer(); 4056 // Potential yield point 4057 CMSTokenSync ts(true); 4058 startTimer(); 4059 sample_eden(); 4060 // Get dirty region starting at nextOffset (inclusive), 4061 // simultaneously clearing it. 4062 dirtyRegion = 4063 _modUnionTable.getAndClearMarkedRegion(nextAddr, endAddr); 4064 assert(dirtyRegion.start() >= nextAddr, 4065 "returned region inconsistent?"); 4066 } 4067 // Remember where the next search should begin. 4068 // The returned region (if non-empty) is a right open interval, 4069 // so lastOffset is obtained from the right end of that 4070 // interval. 4071 lastAddr = dirtyRegion.end(); 4072 // Should do something more transparent and less hacky XXX 4073 numDirtyCards = 4074 _modUnionTable.heapWordDiffToOffsetDiff(dirtyRegion.word_size()); 4075 4076 // We'll scan the cards in the dirty region (with periodic 4077 // yields for foreground GC as needed). 4078 if (!dirtyRegion.is_empty()) { 4079 assert(numDirtyCards > 0, "consistency check"); 4080 HeapWord* stop_point = NULL; 4081 stopTimer(); 4082 // Potential yield point 4083 CMSTokenSyncWithLocks ts(true, old_gen->freelistLock(), 4084 bitMapLock()); 4085 startTimer(); 4086 { 4087 verify_work_stacks_empty(); 4088 verify_overflow_empty(); 4089 sample_eden(); 4090 stop_point = 4091 old_gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 4092 } 4093 if (stop_point != NULL) { 4094 // The careful iteration stopped early either because it found an 4095 // uninitialized object, or because we were in the midst of an 4096 // "abortable preclean", which should now be aborted. Redirty 4097 // the bits corresponding to the partially-scanned or unscanned 4098 // cards. We'll either restart at the next block boundary or 4099 // abort the preclean. 4100 assert((_collectorState == AbortablePreclean && should_abort_preclean()), 4101 "Should only be AbortablePreclean."); 4102 _modUnionTable.mark_range(MemRegion(stop_point, dirtyRegion.end())); 4103 if (should_abort_preclean()) { 4104 break; // out of preclean loop 4105 } else { 4106 // Compute the next address at which preclean should pick up; 4107 // might need bitMapLock in order to read P-bits. 4108 lastAddr = next_card_start_after_block(stop_point); 4109 } 4110 } 4111 } else { 4112 assert(lastAddr == endAddr, "consistency check"); 4113 assert(numDirtyCards == 0, "consistency check"); 4114 break; 4115 } 4116 } 4117 verify_work_stacks_empty(); 4118 verify_overflow_empty(); 4119 return cumNumDirtyCards; 4120 } 4121 4122 // NOTE: preclean_mod_union_table() above and preclean_card_table() 4123 // below are largely identical; if you need to modify 4124 // one of these methods, please check the other method too. 4125 4126 size_t CMSCollector::preclean_card_table(ConcurrentMarkSweepGeneration* old_gen, 4127 ScanMarkedObjectsAgainCarefullyClosure* cl) { 4128 // strategy: it's similar to precleamModUnionTable above, in that 4129 // we accumulate contiguous ranges of dirty cards, mark these cards 4130 // precleaned, then scan the region covered by these cards. 4131 HeapWord* endAddr = (HeapWord*)(old_gen->_virtual_space.high()); 4132 HeapWord* startAddr = (HeapWord*)(old_gen->_virtual_space.low()); 4133 4134 cl->setFreelistLock(old_gen->freelistLock()); // needed for yielding 4135 4136 size_t numDirtyCards, cumNumDirtyCards; 4137 HeapWord *lastAddr, *nextAddr; 4138 4139 for (cumNumDirtyCards = numDirtyCards = 0, 4140 nextAddr = lastAddr = startAddr; 4141 nextAddr < endAddr; 4142 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 4143 4144 ResourceMark rm; 4145 HandleMark hm; 4146 4147 MemRegion dirtyRegion; 4148 { 4149 // See comments in "Precleaning notes" above on why we 4150 // do this locking. XXX Could the locking overheads be 4151 // too high when dirty cards are sparse? [I don't think so.] 4152 stopTimer(); 4153 CMSTokenSync x(true); // is cms thread 4154 startTimer(); 4155 sample_eden(); 4156 // Get and clear dirty region from card table 4157 dirtyRegion = _ct->ct_bs()->dirty_card_range_after_reset( 4158 MemRegion(nextAddr, endAddr), 4159 true, 4160 CardTableModRefBS::precleaned_card_val()); 4161 4162 assert(dirtyRegion.start() >= nextAddr, 4163 "returned region inconsistent?"); 4164 } 4165 lastAddr = dirtyRegion.end(); 4166 numDirtyCards = 4167 dirtyRegion.word_size()/CardTableModRefBS::card_size_in_words; 4168 4169 if (!dirtyRegion.is_empty()) { 4170 stopTimer(); 4171 CMSTokenSyncWithLocks ts(true, old_gen->freelistLock(), bitMapLock()); 4172 startTimer(); 4173 sample_eden(); 4174 verify_work_stacks_empty(); 4175 verify_overflow_empty(); 4176 HeapWord* stop_point = 4177 old_gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 4178 if (stop_point != NULL) { 4179 assert((_collectorState == AbortablePreclean && should_abort_preclean()), 4180 "Should only be AbortablePreclean."); 4181 _ct->ct_bs()->invalidate(MemRegion(stop_point, dirtyRegion.end())); 4182 if (should_abort_preclean()) { 4183 break; // out of preclean loop 4184 } else { 4185 // Compute the next address at which preclean should pick up. 4186 lastAddr = next_card_start_after_block(stop_point); 4187 } 4188 } 4189 } else { 4190 break; 4191 } 4192 } 4193 verify_work_stacks_empty(); 4194 verify_overflow_empty(); 4195 return cumNumDirtyCards; 4196 } 4197 4198 class PrecleanKlassClosure : public KlassClosure { 4199 KlassToOopClosure _cm_klass_closure; 4200 public: 4201 PrecleanKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {} 4202 void do_klass(Klass* k) { 4203 if (k->has_accumulated_modified_oops()) { 4204 k->clear_accumulated_modified_oops(); 4205 4206 _cm_klass_closure.do_klass(k); 4207 } 4208 } 4209 }; 4210 4211 // The freelist lock is needed to prevent asserts, is it really needed? 4212 void CMSCollector::preclean_klasses(MarkRefsIntoAndScanClosure* cl, Mutex* freelistLock) { 4213 4214 cl->set_freelistLock(freelistLock); 4215 4216 CMSTokenSyncWithLocks ts(true, freelistLock, bitMapLock()); 4217 4218 // SSS: Add equivalent to ScanMarkedObjectsAgainCarefullyClosure::do_yield_check and should_abort_preclean? 4219 // SSS: We should probably check if precleaning should be aborted, at suitable intervals? 4220 PrecleanKlassClosure preclean_klass_closure(cl); 4221 ClassLoaderDataGraph::classes_do(&preclean_klass_closure); 4222 4223 verify_work_stacks_empty(); 4224 verify_overflow_empty(); 4225 } 4226 4227 void CMSCollector::checkpointRootsFinal() { 4228 assert(_collectorState == FinalMarking, "incorrect state transition?"); 4229 check_correct_thread_executing(); 4230 // world is stopped at this checkpoint 4231 assert(SafepointSynchronize::is_at_safepoint(), 4232 "world should be stopped"); 4233 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 4234 4235 verify_work_stacks_empty(); 4236 verify_overflow_empty(); 4237 4238 if (PrintGCDetails) { 4239 gclog_or_tty->print("[YG occupancy: " SIZE_FORMAT " K (" SIZE_FORMAT " K)]", 4240 _young_gen->used() / K, 4241 _young_gen->capacity() / K); 4242 } 4243 { 4244 if (CMSScavengeBeforeRemark) { 4245 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4246 // Temporarily set flag to false, GCH->do_collection will 4247 // expect it to be false and set to true 4248 FlagSetting fl(gch->_is_gc_active, false); 4249 NOT_PRODUCT(GCTraceTime t("Scavenge-Before-Remark", 4250 PrintGCDetails && Verbose, true, _gc_timer_cm);) 4251 gch->do_collection(true, // full (i.e. force, see below) 4252 false, // !clear_all_soft_refs 4253 0, // size 4254 false, // is_tlab 4255 GenCollectedHeap::YoungGen // type 4256 ); 4257 } 4258 FreelistLocker x(this); 4259 MutexLockerEx y(bitMapLock(), 4260 Mutex::_no_safepoint_check_flag); 4261 checkpointRootsFinalWork(); 4262 } 4263 verify_work_stacks_empty(); 4264 verify_overflow_empty(); 4265 } 4266 4267 void CMSCollector::checkpointRootsFinalWork() { 4268 NOT_PRODUCT(GCTraceTime tr("checkpointRootsFinalWork", PrintGCDetails, false, _gc_timer_cm);) 4269 4270 assert(haveFreelistLocks(), "must have free list locks"); 4271 assert_lock_strong(bitMapLock()); 4272 4273 ResourceMark rm; 4274 HandleMark hm; 4275 4276 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4277 4278 if (should_unload_classes()) { 4279 CodeCache::gc_prologue(); 4280 } 4281 assert(haveFreelistLocks(), "must have free list locks"); 4282 assert_lock_strong(bitMapLock()); 4283 4284 // We might assume that we need not fill TLAB's when 4285 // CMSScavengeBeforeRemark is set, because we may have just done 4286 // a scavenge which would have filled all TLAB's -- and besides 4287 // Eden would be empty. This however may not always be the case -- 4288 // for instance although we asked for a scavenge, it may not have 4289 // happened because of a JNI critical section. We probably need 4290 // a policy for deciding whether we can in that case wait until 4291 // the critical section releases and then do the remark following 4292 // the scavenge, and skip it here. In the absence of that policy, 4293 // or of an indication of whether the scavenge did indeed occur, 4294 // we cannot rely on TLAB's having been filled and must do 4295 // so here just in case a scavenge did not happen. 4296 gch->ensure_parsability(false); // fill TLAB's, but no need to retire them 4297 // Update the saved marks which may affect the root scans. 4298 gch->save_marks(); 4299 4300 if (CMSPrintEdenSurvivorChunks) { 4301 print_eden_and_survivor_chunk_arrays(); 4302 } 4303 4304 { 4305 #if defined(COMPILER2) || INCLUDE_JVMCI 4306 DerivedPointerTableDeactivate dpt_deact; 4307 #endif 4308 4309 // Note on the role of the mod union table: 4310 // Since the marker in "markFromRoots" marks concurrently with 4311 // mutators, it is possible for some reachable objects not to have been 4312 // scanned. For instance, an only reference to an object A was 4313 // placed in object B after the marker scanned B. Unless B is rescanned, 4314 // A would be collected. Such updates to references in marked objects 4315 // are detected via the mod union table which is the set of all cards 4316 // dirtied since the first checkpoint in this GC cycle and prior to 4317 // the most recent young generation GC, minus those cleaned up by the 4318 // concurrent precleaning. 4319 if (CMSParallelRemarkEnabled) { 4320 GCTraceTime t("Rescan (parallel) ", PrintGCDetails, false, _gc_timer_cm); 4321 do_remark_parallel(); 4322 } else { 4323 GCTraceTime t("Rescan (non-parallel) ", PrintGCDetails, false, _gc_timer_cm); 4324 do_remark_non_parallel(); 4325 } 4326 } 4327 verify_work_stacks_empty(); 4328 verify_overflow_empty(); 4329 4330 { 4331 NOT_PRODUCT(GCTraceTime ts("refProcessingWork", PrintGCDetails, false, _gc_timer_cm);) 4332 refProcessingWork(); 4333 } 4334 verify_work_stacks_empty(); 4335 verify_overflow_empty(); 4336 4337 if (should_unload_classes()) { 4338 CodeCache::gc_epilogue(); 4339 } 4340 JvmtiExport::gc_epilogue(); 4341 4342 // If we encountered any (marking stack / work queue) overflow 4343 // events during the current CMS cycle, take appropriate 4344 // remedial measures, where possible, so as to try and avoid 4345 // recurrence of that condition. 4346 assert(_markStack.isEmpty(), "No grey objects"); 4347 size_t ser_ovflw = _ser_pmc_remark_ovflw + _ser_pmc_preclean_ovflw + 4348 _ser_kac_ovflw + _ser_kac_preclean_ovflw; 4349 if (ser_ovflw > 0) { 4350 if (PrintCMSStatistics != 0) { 4351 gclog_or_tty->print_cr("Marking stack overflow (benign) " 4352 "(pmc_pc=" SIZE_FORMAT ", pmc_rm=" SIZE_FORMAT ", kac=" SIZE_FORMAT 4353 ", kac_preclean=" SIZE_FORMAT ")", 4354 _ser_pmc_preclean_ovflw, _ser_pmc_remark_ovflw, 4355 _ser_kac_ovflw, _ser_kac_preclean_ovflw); 4356 } 4357 _markStack.expand(); 4358 _ser_pmc_remark_ovflw = 0; 4359 _ser_pmc_preclean_ovflw = 0; 4360 _ser_kac_preclean_ovflw = 0; 4361 _ser_kac_ovflw = 0; 4362 } 4363 if (_par_pmc_remark_ovflw > 0 || _par_kac_ovflw > 0) { 4364 if (PrintCMSStatistics != 0) { 4365 gclog_or_tty->print_cr("Work queue overflow (benign) " 4366 "(pmc_rm=" SIZE_FORMAT ", kac=" SIZE_FORMAT ")", 4367 _par_pmc_remark_ovflw, _par_kac_ovflw); 4368 } 4369 _par_pmc_remark_ovflw = 0; 4370 _par_kac_ovflw = 0; 4371 } 4372 if (PrintCMSStatistics != 0) { 4373 if (_markStack._hit_limit > 0) { 4374 gclog_or_tty->print_cr(" (benign) Hit max stack size limit (" SIZE_FORMAT ")", 4375 _markStack._hit_limit); 4376 } 4377 if (_markStack._failed_double > 0) { 4378 gclog_or_tty->print_cr(" (benign) Failed stack doubling (" SIZE_FORMAT ")," 4379 " current capacity " SIZE_FORMAT, 4380 _markStack._failed_double, 4381 _markStack.capacity()); 4382 } 4383 } 4384 _markStack._hit_limit = 0; 4385 _markStack._failed_double = 0; 4386 4387 if ((VerifyAfterGC || VerifyDuringGC) && 4388 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 4389 verify_after_remark(); 4390 } 4391 4392 _gc_tracer_cm->report_object_count_after_gc(&_is_alive_closure); 4393 4394 // Change under the freelistLocks. 4395 _collectorState = Sweeping; 4396 // Call isAllClear() under bitMapLock 4397 assert(_modUnionTable.isAllClear(), 4398 "Should be clear by end of the final marking"); 4399 assert(_ct->klass_rem_set()->mod_union_is_clear(), 4400 "Should be clear by end of the final marking"); 4401 } 4402 4403 void CMSParInitialMarkTask::work(uint worker_id) { 4404 elapsedTimer _timer; 4405 ResourceMark rm; 4406 HandleMark hm; 4407 4408 // ---------- scan from roots -------------- 4409 _timer.start(); 4410 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4411 Par_MarkRefsIntoClosure par_mri_cl(_collector->_span, &(_collector->_markBitMap)); 4412 4413 // ---------- young gen roots -------------- 4414 { 4415 work_on_young_gen_roots(worker_id, &par_mri_cl); 4416 _timer.stop(); 4417 if (PrintCMSStatistics != 0) { 4418 gclog_or_tty->print_cr( 4419 "Finished young gen initial mark scan work in %dth thread: %3.3f sec", 4420 worker_id, _timer.seconds()); 4421 } 4422 } 4423 4424 // ---------- remaining roots -------------- 4425 _timer.reset(); 4426 _timer.start(); 4427 4428 CLDToOopClosure cld_closure(&par_mri_cl, true); 4429 4430 gch->gen_process_roots(_strong_roots_scope, 4431 GenCollectedHeap::OldGen, 4432 false, // yg was scanned above 4433 GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()), 4434 _collector->should_unload_classes(), 4435 &par_mri_cl, 4436 NULL, 4437 &cld_closure); 4438 assert(_collector->should_unload_classes() 4439 || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache), 4440 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 4441 _timer.stop(); 4442 if (PrintCMSStatistics != 0) { 4443 gclog_or_tty->print_cr( 4444 "Finished remaining root initial mark scan work in %dth thread: %3.3f sec", 4445 worker_id, _timer.seconds()); 4446 } 4447 } 4448 4449 // Parallel remark task 4450 class CMSParRemarkTask: public CMSParMarkTask { 4451 CompactibleFreeListSpace* _cms_space; 4452 4453 // The per-thread work queues, available here for stealing. 4454 OopTaskQueueSet* _task_queues; 4455 ParallelTaskTerminator _term; 4456 StrongRootsScope* _strong_roots_scope; 4457 4458 public: 4459 // A value of 0 passed to n_workers will cause the number of 4460 // workers to be taken from the active workers in the work gang. 4461 CMSParRemarkTask(CMSCollector* collector, 4462 CompactibleFreeListSpace* cms_space, 4463 uint n_workers, WorkGang* workers, 4464 OopTaskQueueSet* task_queues, 4465 StrongRootsScope* strong_roots_scope): 4466 CMSParMarkTask("Rescan roots and grey objects in parallel", 4467 collector, n_workers), 4468 _cms_space(cms_space), 4469 _task_queues(task_queues), 4470 _term(n_workers, task_queues), 4471 _strong_roots_scope(strong_roots_scope) { } 4472 4473 OopTaskQueueSet* task_queues() { return _task_queues; } 4474 4475 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 4476 4477 ParallelTaskTerminator* terminator() { return &_term; } 4478 uint n_workers() { return _n_workers; } 4479 4480 void work(uint worker_id); 4481 4482 private: 4483 // ... of dirty cards in old space 4484 void do_dirty_card_rescan_tasks(CompactibleFreeListSpace* sp, int i, 4485 Par_MarkRefsIntoAndScanClosure* cl); 4486 4487 // ... work stealing for the above 4488 void do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, int* seed); 4489 }; 4490 4491 class RemarkKlassClosure : public KlassClosure { 4492 KlassToOopClosure _cm_klass_closure; 4493 public: 4494 RemarkKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {} 4495 void do_klass(Klass* k) { 4496 // Check if we have modified any oops in the Klass during the concurrent marking. 4497 if (k->has_accumulated_modified_oops()) { 4498 k->clear_accumulated_modified_oops(); 4499 4500 // We could have transfered the current modified marks to the accumulated marks, 4501 // like we do with the Card Table to Mod Union Table. But it's not really necessary. 4502 } else if (k->has_modified_oops()) { 4503 // Don't clear anything, this info is needed by the next young collection. 4504 } else { 4505 // No modified oops in the Klass. 4506 return; 4507 } 4508 4509 // The klass has modified fields, need to scan the klass. 4510 _cm_klass_closure.do_klass(k); 4511 } 4512 }; 4513 4514 void CMSParMarkTask::work_on_young_gen_roots(uint worker_id, OopsInGenClosure* cl) { 4515 ParNewGeneration* young_gen = _collector->_young_gen; 4516 ContiguousSpace* eden_space = young_gen->eden(); 4517 ContiguousSpace* from_space = young_gen->from(); 4518 ContiguousSpace* to_space = young_gen->to(); 4519 4520 HeapWord** eca = _collector->_eden_chunk_array; 4521 size_t ect = _collector->_eden_chunk_index; 4522 HeapWord** sca = _collector->_survivor_chunk_array; 4523 size_t sct = _collector->_survivor_chunk_index; 4524 4525 assert(ect <= _collector->_eden_chunk_capacity, "out of bounds"); 4526 assert(sct <= _collector->_survivor_chunk_capacity, "out of bounds"); 4527 4528 do_young_space_rescan(worker_id, cl, to_space, NULL, 0); 4529 do_young_space_rescan(worker_id, cl, from_space, sca, sct); 4530 do_young_space_rescan(worker_id, cl, eden_space, eca, ect); 4531 } 4532 4533 // work_queue(i) is passed to the closure 4534 // Par_MarkRefsIntoAndScanClosure. The "i" parameter 4535 // also is passed to do_dirty_card_rescan_tasks() and to 4536 // do_work_steal() to select the i-th task_queue. 4537 4538 void CMSParRemarkTask::work(uint worker_id) { 4539 elapsedTimer _timer; 4540 ResourceMark rm; 4541 HandleMark hm; 4542 4543 // ---------- rescan from roots -------------- 4544 _timer.start(); 4545 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4546 Par_MarkRefsIntoAndScanClosure par_mrias_cl(_collector, 4547 _collector->_span, _collector->ref_processor(), 4548 &(_collector->_markBitMap), 4549 work_queue(worker_id)); 4550 4551 // Rescan young gen roots first since these are likely 4552 // coarsely partitioned and may, on that account, constitute 4553 // the critical path; thus, it's best to start off that 4554 // work first. 4555 // ---------- young gen roots -------------- 4556 { 4557 work_on_young_gen_roots(worker_id, &par_mrias_cl); 4558 _timer.stop(); 4559 if (PrintCMSStatistics != 0) { 4560 gclog_or_tty->print_cr( 4561 "Finished young gen rescan work in %dth thread: %3.3f sec", 4562 worker_id, _timer.seconds()); 4563 } 4564 } 4565 4566 // ---------- remaining roots -------------- 4567 _timer.reset(); 4568 _timer.start(); 4569 gch->gen_process_roots(_strong_roots_scope, 4570 GenCollectedHeap::OldGen, 4571 false, // yg was scanned above 4572 GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()), 4573 _collector->should_unload_classes(), 4574 &par_mrias_cl, 4575 NULL, 4576 NULL); // The dirty klasses will be handled below 4577 4578 assert(_collector->should_unload_classes() 4579 || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache), 4580 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 4581 _timer.stop(); 4582 if (PrintCMSStatistics != 0) { 4583 gclog_or_tty->print_cr( 4584 "Finished remaining root rescan work in %dth thread: %3.3f sec", 4585 worker_id, _timer.seconds()); 4586 } 4587 4588 // ---------- unhandled CLD scanning ---------- 4589 if (worker_id == 0) { // Single threaded at the moment. 4590 _timer.reset(); 4591 _timer.start(); 4592 4593 // Scan all new class loader data objects and new dependencies that were 4594 // introduced during concurrent marking. 4595 ResourceMark rm; 4596 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds(); 4597 for (int i = 0; i < array->length(); i++) { 4598 par_mrias_cl.do_cld_nv(array->at(i)); 4599 } 4600 4601 // We don't need to keep track of new CLDs anymore. 4602 ClassLoaderDataGraph::remember_new_clds(false); 4603 4604 _timer.stop(); 4605 if (PrintCMSStatistics != 0) { 4606 gclog_or_tty->print_cr( 4607 "Finished unhandled CLD scanning work in %dth thread: %3.3f sec", 4608 worker_id, _timer.seconds()); 4609 } 4610 } 4611 4612 // ---------- dirty klass scanning ---------- 4613 if (worker_id == 0) { // Single threaded at the moment. 4614 _timer.reset(); 4615 _timer.start(); 4616 4617 // Scan all classes that was dirtied during the concurrent marking phase. 4618 RemarkKlassClosure remark_klass_closure(&par_mrias_cl); 4619 ClassLoaderDataGraph::classes_do(&remark_klass_closure); 4620 4621 _timer.stop(); 4622 if (PrintCMSStatistics != 0) { 4623 gclog_or_tty->print_cr( 4624 "Finished dirty klass scanning work in %dth thread: %3.3f sec", 4625 worker_id, _timer.seconds()); 4626 } 4627 } 4628 4629 // We might have added oops to ClassLoaderData::_handles during the 4630 // concurrent marking phase. These oops point to newly allocated objects 4631 // that are guaranteed to be kept alive. Either by the direct allocation 4632 // code, or when the young collector processes the roots. Hence, 4633 // we don't have to revisit the _handles block during the remark phase. 4634 4635 // ---------- rescan dirty cards ------------ 4636 _timer.reset(); 4637 _timer.start(); 4638 4639 // Do the rescan tasks for each of the two spaces 4640 // (cms_space) in turn. 4641 // "worker_id" is passed to select the task_queue for "worker_id" 4642 do_dirty_card_rescan_tasks(_cms_space, worker_id, &par_mrias_cl); 4643 _timer.stop(); 4644 if (PrintCMSStatistics != 0) { 4645 gclog_or_tty->print_cr( 4646 "Finished dirty card rescan work in %dth thread: %3.3f sec", 4647 worker_id, _timer.seconds()); 4648 } 4649 4650 // ---------- steal work from other threads ... 4651 // ---------- ... and drain overflow list. 4652 _timer.reset(); 4653 _timer.start(); 4654 do_work_steal(worker_id, &par_mrias_cl, _collector->hash_seed(worker_id)); 4655 _timer.stop(); 4656 if (PrintCMSStatistics != 0) { 4657 gclog_or_tty->print_cr( 4658 "Finished work stealing in %dth thread: %3.3f sec", 4659 worker_id, _timer.seconds()); 4660 } 4661 } 4662 4663 // Note that parameter "i" is not used. 4664 void 4665 CMSParMarkTask::do_young_space_rescan(uint worker_id, 4666 OopsInGenClosure* cl, ContiguousSpace* space, 4667 HeapWord** chunk_array, size_t chunk_top) { 4668 // Until all tasks completed: 4669 // . claim an unclaimed task 4670 // . compute region boundaries corresponding to task claimed 4671 // using chunk_array 4672 // . par_oop_iterate(cl) over that region 4673 4674 ResourceMark rm; 4675 HandleMark hm; 4676 4677 SequentialSubTasksDone* pst = space->par_seq_tasks(); 4678 4679 uint nth_task = 0; 4680 uint n_tasks = pst->n_tasks(); 4681 4682 if (n_tasks > 0) { 4683 assert(pst->valid(), "Uninitialized use?"); 4684 HeapWord *start, *end; 4685 while (!pst->is_task_claimed(/* reference */ nth_task)) { 4686 // We claimed task # nth_task; compute its boundaries. 4687 if (chunk_top == 0) { // no samples were taken 4688 assert(nth_task == 0 && n_tasks == 1, "Can have only 1 eden task"); 4689 start = space->bottom(); 4690 end = space->top(); 4691 } else if (nth_task == 0) { 4692 start = space->bottom(); 4693 end = chunk_array[nth_task]; 4694 } else if (nth_task < (uint)chunk_top) { 4695 assert(nth_task >= 1, "Control point invariant"); 4696 start = chunk_array[nth_task - 1]; 4697 end = chunk_array[nth_task]; 4698 } else { 4699 assert(nth_task == (uint)chunk_top, "Control point invariant"); 4700 start = chunk_array[chunk_top - 1]; 4701 end = space->top(); 4702 } 4703 MemRegion mr(start, end); 4704 // Verify that mr is in space 4705 assert(mr.is_empty() || space->used_region().contains(mr), 4706 "Should be in space"); 4707 // Verify that "start" is an object boundary 4708 assert(mr.is_empty() || oop(mr.start())->is_oop(), 4709 "Should be an oop"); 4710 space->par_oop_iterate(mr, cl); 4711 } 4712 pst->all_tasks_completed(); 4713 } 4714 } 4715 4716 void 4717 CMSParRemarkTask::do_dirty_card_rescan_tasks( 4718 CompactibleFreeListSpace* sp, int i, 4719 Par_MarkRefsIntoAndScanClosure* cl) { 4720 // Until all tasks completed: 4721 // . claim an unclaimed task 4722 // . compute region boundaries corresponding to task claimed 4723 // . transfer dirty bits ct->mut for that region 4724 // . apply rescanclosure to dirty mut bits for that region 4725 4726 ResourceMark rm; 4727 HandleMark hm; 4728 4729 OopTaskQueue* work_q = work_queue(i); 4730 ModUnionClosure modUnionClosure(&(_collector->_modUnionTable)); 4731 // CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! 4732 // CAUTION: This closure has state that persists across calls to 4733 // the work method dirty_range_iterate_clear() in that it has 4734 // embedded in it a (subtype of) UpwardsObjectClosure. The 4735 // use of that state in the embedded UpwardsObjectClosure instance 4736 // assumes that the cards are always iterated (even if in parallel 4737 // by several threads) in monotonically increasing order per each 4738 // thread. This is true of the implementation below which picks 4739 // card ranges (chunks) in monotonically increasing order globally 4740 // and, a-fortiori, in monotonically increasing order per thread 4741 // (the latter order being a subsequence of the former). 4742 // If the work code below is ever reorganized into a more chaotic 4743 // work-partitioning form than the current "sequential tasks" 4744 // paradigm, the use of that persistent state will have to be 4745 // revisited and modified appropriately. See also related 4746 // bug 4756801 work on which should examine this code to make 4747 // sure that the changes there do not run counter to the 4748 // assumptions made here and necessary for correctness and 4749 // efficiency. Note also that this code might yield inefficient 4750 // behavior in the case of very large objects that span one or 4751 // more work chunks. Such objects would potentially be scanned 4752 // several times redundantly. Work on 4756801 should try and 4753 // address that performance anomaly if at all possible. XXX 4754 MemRegion full_span = _collector->_span; 4755 CMSBitMap* bm = &(_collector->_markBitMap); // shared 4756 MarkFromDirtyCardsClosure 4757 greyRescanClosure(_collector, full_span, // entire span of interest 4758 sp, bm, work_q, cl); 4759 4760 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 4761 assert(pst->valid(), "Uninitialized use?"); 4762 uint nth_task = 0; 4763 const int alignment = CardTableModRefBS::card_size * BitsPerWord; 4764 MemRegion span = sp->used_region(); 4765 HeapWord* start_addr = span.start(); 4766 HeapWord* end_addr = (HeapWord*)round_to((intptr_t)span.end(), 4767 alignment); 4768 const size_t chunk_size = sp->rescan_task_size(); // in HeapWord units 4769 assert((HeapWord*)round_to((intptr_t)start_addr, alignment) == 4770 start_addr, "Check alignment"); 4771 assert((size_t)round_to((intptr_t)chunk_size, alignment) == 4772 chunk_size, "Check alignment"); 4773 4774 while (!pst->is_task_claimed(/* reference */ nth_task)) { 4775 // Having claimed the nth_task, compute corresponding mem-region, 4776 // which is a-fortiori aligned correctly (i.e. at a MUT boundary). 4777 // The alignment restriction ensures that we do not need any 4778 // synchronization with other gang-workers while setting or 4779 // clearing bits in thus chunk of the MUT. 4780 MemRegion this_span = MemRegion(start_addr + nth_task*chunk_size, 4781 start_addr + (nth_task+1)*chunk_size); 4782 // The last chunk's end might be way beyond end of the 4783 // used region. In that case pull back appropriately. 4784 if (this_span.end() > end_addr) { 4785 this_span.set_end(end_addr); 4786 assert(!this_span.is_empty(), "Program logic (calculation of n_tasks)"); 4787 } 4788 // Iterate over the dirty cards covering this chunk, marking them 4789 // precleaned, and setting the corresponding bits in the mod union 4790 // table. Since we have been careful to partition at Card and MUT-word 4791 // boundaries no synchronization is needed between parallel threads. 4792 _collector->_ct->ct_bs()->dirty_card_iterate(this_span, 4793 &modUnionClosure); 4794 4795 // Having transferred these marks into the modUnionTable, 4796 // rescan the marked objects on the dirty cards in the modUnionTable. 4797 // Even if this is at a synchronous collection, the initial marking 4798 // may have been done during an asynchronous collection so there 4799 // may be dirty bits in the mod-union table. 4800 _collector->_modUnionTable.dirty_range_iterate_clear( 4801 this_span, &greyRescanClosure); 4802 _collector->_modUnionTable.verifyNoOneBitsInRange( 4803 this_span.start(), 4804 this_span.end()); 4805 } 4806 pst->all_tasks_completed(); // declare that i am done 4807 } 4808 4809 // . see if we can share work_queues with ParNew? XXX 4810 void 4811 CMSParRemarkTask::do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, 4812 int* seed) { 4813 OopTaskQueue* work_q = work_queue(i); 4814 NOT_PRODUCT(int num_steals = 0;) 4815 oop obj_to_scan; 4816 CMSBitMap* bm = &(_collector->_markBitMap); 4817 4818 while (true) { 4819 // Completely finish any left over work from (an) earlier round(s) 4820 cl->trim_queue(0); 4821 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 4822 (size_t)ParGCDesiredObjsFromOverflowList); 4823 // Now check if there's any work in the overflow list 4824 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads, 4825 // only affects the number of attempts made to get work from the 4826 // overflow list and does not affect the number of workers. Just 4827 // pass ParallelGCThreads so this behavior is unchanged. 4828 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 4829 work_q, 4830 ParallelGCThreads)) { 4831 // found something in global overflow list; 4832 // not yet ready to go stealing work from others. 4833 // We'd like to assert(work_q->size() != 0, ...) 4834 // because we just took work from the overflow list, 4835 // but of course we can't since all of that could have 4836 // been already stolen from us. 4837 // "He giveth and He taketh away." 4838 continue; 4839 } 4840 // Verify that we have no work before we resort to stealing 4841 assert(work_q->size() == 0, "Have work, shouldn't steal"); 4842 // Try to steal from other queues that have work 4843 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 4844 NOT_PRODUCT(num_steals++;) 4845 assert(obj_to_scan->is_oop(), "Oops, not an oop!"); 4846 assert(bm->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 4847 // Do scanning work 4848 obj_to_scan->oop_iterate(cl); 4849 // Loop around, finish this work, and try to steal some more 4850 } else if (terminator()->offer_termination()) { 4851 break; // nirvana from the infinite cycle 4852 } 4853 } 4854 NOT_PRODUCT( 4855 if (PrintCMSStatistics != 0) { 4856 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals); 4857 } 4858 ) 4859 assert(work_q->size() == 0 && _collector->overflow_list_is_empty(), 4860 "Else our work is not yet done"); 4861 } 4862 4863 // Record object boundaries in _eden_chunk_array by sampling the eden 4864 // top in the slow-path eden object allocation code path and record 4865 // the boundaries, if CMSEdenChunksRecordAlways is true. If 4866 // CMSEdenChunksRecordAlways is false, we use the other asynchronous 4867 // sampling in sample_eden() that activates during the part of the 4868 // preclean phase. 4869 void CMSCollector::sample_eden_chunk() { 4870 if (CMSEdenChunksRecordAlways && _eden_chunk_array != NULL) { 4871 if (_eden_chunk_lock->try_lock()) { 4872 // Record a sample. This is the critical section. The contents 4873 // of the _eden_chunk_array have to be non-decreasing in the 4874 // address order. 4875 _eden_chunk_array[_eden_chunk_index] = *_top_addr; 4876 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr, 4877 "Unexpected state of Eden"); 4878 if (_eden_chunk_index == 0 || 4879 ((_eden_chunk_array[_eden_chunk_index] > _eden_chunk_array[_eden_chunk_index-1]) && 4880 (pointer_delta(_eden_chunk_array[_eden_chunk_index], 4881 _eden_chunk_array[_eden_chunk_index-1]) >= CMSSamplingGrain))) { 4882 _eden_chunk_index++; // commit sample 4883 } 4884 _eden_chunk_lock->unlock(); 4885 } 4886 } 4887 } 4888 4889 // Return a thread-local PLAB recording array, as appropriate. 4890 void* CMSCollector::get_data_recorder(int thr_num) { 4891 if (_survivor_plab_array != NULL && 4892 (CMSPLABRecordAlways || 4893 (_collectorState > Marking && _collectorState < FinalMarking))) { 4894 assert(thr_num < (int)ParallelGCThreads, "thr_num is out of bounds"); 4895 ChunkArray* ca = &_survivor_plab_array[thr_num]; 4896 ca->reset(); // clear it so that fresh data is recorded 4897 return (void*) ca; 4898 } else { 4899 return NULL; 4900 } 4901 } 4902 4903 // Reset all the thread-local PLAB recording arrays 4904 void CMSCollector::reset_survivor_plab_arrays() { 4905 for (uint i = 0; i < ParallelGCThreads; i++) { 4906 _survivor_plab_array[i].reset(); 4907 } 4908 } 4909 4910 // Merge the per-thread plab arrays into the global survivor chunk 4911 // array which will provide the partitioning of the survivor space 4912 // for CMS initial scan and rescan. 4913 void CMSCollector::merge_survivor_plab_arrays(ContiguousSpace* surv, 4914 int no_of_gc_threads) { 4915 assert(_survivor_plab_array != NULL, "Error"); 4916 assert(_survivor_chunk_array != NULL, "Error"); 4917 assert(_collectorState == FinalMarking || 4918 (CMSParallelInitialMarkEnabled && _collectorState == InitialMarking), "Error"); 4919 for (int j = 0; j < no_of_gc_threads; j++) { 4920 _cursor[j] = 0; 4921 } 4922 HeapWord* top = surv->top(); 4923 size_t i; 4924 for (i = 0; i < _survivor_chunk_capacity; i++) { // all sca entries 4925 HeapWord* min_val = top; // Higher than any PLAB address 4926 uint min_tid = 0; // position of min_val this round 4927 for (int j = 0; j < no_of_gc_threads; j++) { 4928 ChunkArray* cur_sca = &_survivor_plab_array[j]; 4929 if (_cursor[j] == cur_sca->end()) { 4930 continue; 4931 } 4932 assert(_cursor[j] < cur_sca->end(), "ctl pt invariant"); 4933 HeapWord* cur_val = cur_sca->nth(_cursor[j]); 4934 assert(surv->used_region().contains(cur_val), "Out of bounds value"); 4935 if (cur_val < min_val) { 4936 min_tid = j; 4937 min_val = cur_val; 4938 } else { 4939 assert(cur_val < top, "All recorded addresses should be less"); 4940 } 4941 } 4942 // At this point min_val and min_tid are respectively 4943 // the least address in _survivor_plab_array[j]->nth(_cursor[j]) 4944 // and the thread (j) that witnesses that address. 4945 // We record this address in the _survivor_chunk_array[i] 4946 // and increment _cursor[min_tid] prior to the next round i. 4947 if (min_val == top) { 4948 break; 4949 } 4950 _survivor_chunk_array[i] = min_val; 4951 _cursor[min_tid]++; 4952 } 4953 // We are all done; record the size of the _survivor_chunk_array 4954 _survivor_chunk_index = i; // exclusive: [0, i) 4955 if (PrintCMSStatistics > 0) { 4956 gclog_or_tty->print(" (Survivor:" SIZE_FORMAT "chunks) ", i); 4957 } 4958 // Verify that we used up all the recorded entries 4959 #ifdef ASSERT 4960 size_t total = 0; 4961 for (int j = 0; j < no_of_gc_threads; j++) { 4962 assert(_cursor[j] == _survivor_plab_array[j].end(), "Ctl pt invariant"); 4963 total += _cursor[j]; 4964 } 4965 assert(total == _survivor_chunk_index, "Ctl Pt Invariant"); 4966 // Check that the merged array is in sorted order 4967 if (total > 0) { 4968 for (size_t i = 0; i < total - 1; i++) { 4969 if (PrintCMSStatistics > 0) { 4970 gclog_or_tty->print(" (chunk" SIZE_FORMAT ":" INTPTR_FORMAT ") ", 4971 i, p2i(_survivor_chunk_array[i])); 4972 } 4973 assert(_survivor_chunk_array[i] < _survivor_chunk_array[i+1], 4974 "Not sorted"); 4975 } 4976 } 4977 #endif // ASSERT 4978 } 4979 4980 // Set up the space's par_seq_tasks structure for work claiming 4981 // for parallel initial scan and rescan of young gen. 4982 // See ParRescanTask where this is currently used. 4983 void 4984 CMSCollector:: 4985 initialize_sequential_subtasks_for_young_gen_rescan(int n_threads) { 4986 assert(n_threads > 0, "Unexpected n_threads argument"); 4987 4988 // Eden space 4989 if (!_young_gen->eden()->is_empty()) { 4990 SequentialSubTasksDone* pst = _young_gen->eden()->par_seq_tasks(); 4991 assert(!pst->valid(), "Clobbering existing data?"); 4992 // Each valid entry in [0, _eden_chunk_index) represents a task. 4993 size_t n_tasks = _eden_chunk_index + 1; 4994 assert(n_tasks == 1 || _eden_chunk_array != NULL, "Error"); 4995 // Sets the condition for completion of the subtask (how many threads 4996 // need to finish in order to be done). 4997 pst->set_n_threads(n_threads); 4998 pst->set_n_tasks((int)n_tasks); 4999 } 5000 5001 // Merge the survivor plab arrays into _survivor_chunk_array 5002 if (_survivor_plab_array != NULL) { 5003 merge_survivor_plab_arrays(_young_gen->from(), n_threads); 5004 } else { 5005 assert(_survivor_chunk_index == 0, "Error"); 5006 } 5007 5008 // To space 5009 { 5010 SequentialSubTasksDone* pst = _young_gen->to()->par_seq_tasks(); 5011 assert(!pst->valid(), "Clobbering existing data?"); 5012 // Sets the condition for completion of the subtask (how many threads 5013 // need to finish in order to be done). 5014 pst->set_n_threads(n_threads); 5015 pst->set_n_tasks(1); 5016 assert(pst->valid(), "Error"); 5017 } 5018 5019 // From space 5020 { 5021 SequentialSubTasksDone* pst = _young_gen->from()->par_seq_tasks(); 5022 assert(!pst->valid(), "Clobbering existing data?"); 5023 size_t n_tasks = _survivor_chunk_index + 1; 5024 assert(n_tasks == 1 || _survivor_chunk_array != NULL, "Error"); 5025 // Sets the condition for completion of the subtask (how many threads 5026 // need to finish in order to be done). 5027 pst->set_n_threads(n_threads); 5028 pst->set_n_tasks((int)n_tasks); 5029 assert(pst->valid(), "Error"); 5030 } 5031 } 5032 5033 // Parallel version of remark 5034 void CMSCollector::do_remark_parallel() { 5035 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5036 WorkGang* workers = gch->workers(); 5037 assert(workers != NULL, "Need parallel worker threads."); 5038 // Choose to use the number of GC workers most recently set 5039 // into "active_workers". 5040 uint n_workers = workers->active_workers(); 5041 5042 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 5043 5044 StrongRootsScope srs(n_workers); 5045 5046 CMSParRemarkTask tsk(this, cms_space, n_workers, workers, task_queues(), &srs); 5047 5048 // We won't be iterating over the cards in the card table updating 5049 // the younger_gen cards, so we shouldn't call the following else 5050 // the verification code as well as subsequent younger_refs_iterate 5051 // code would get confused. XXX 5052 // gch->rem_set()->prepare_for_younger_refs_iterate(true); // parallel 5053 5054 // The young gen rescan work will not be done as part of 5055 // process_roots (which currently doesn't know how to 5056 // parallelize such a scan), but rather will be broken up into 5057 // a set of parallel tasks (via the sampling that the [abortable] 5058 // preclean phase did of eden, plus the [two] tasks of 5059 // scanning the [two] survivor spaces. Further fine-grain 5060 // parallelization of the scanning of the survivor spaces 5061 // themselves, and of precleaning of the young gen itself 5062 // is deferred to the future. 5063 initialize_sequential_subtasks_for_young_gen_rescan(n_workers); 5064 5065 // The dirty card rescan work is broken up into a "sequence" 5066 // of parallel tasks (per constituent space) that are dynamically 5067 // claimed by the parallel threads. 5068 cms_space->initialize_sequential_subtasks_for_rescan(n_workers); 5069 5070 // It turns out that even when we're using 1 thread, doing the work in a 5071 // separate thread causes wide variance in run times. We can't help this 5072 // in the multi-threaded case, but we special-case n=1 here to get 5073 // repeatable measurements of the 1-thread overhead of the parallel code. 5074 if (n_workers > 1) { 5075 // Make refs discovery MT-safe, if it isn't already: it may not 5076 // necessarily be so, since it's possible that we are doing 5077 // ST marking. 5078 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), true); 5079 workers->run_task(&tsk); 5080 } else { 5081 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false); 5082 tsk.work(0); 5083 } 5084 5085 // restore, single-threaded for now, any preserved marks 5086 // as a result of work_q overflow 5087 restore_preserved_marks_if_any(); 5088 } 5089 5090 // Non-parallel version of remark 5091 void CMSCollector::do_remark_non_parallel() { 5092 ResourceMark rm; 5093 HandleMark hm; 5094 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5095 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false); 5096 5097 MarkRefsIntoAndScanClosure 5098 mrias_cl(_span, ref_processor(), &_markBitMap, NULL /* not precleaning */, 5099 &_markStack, this, 5100 false /* should_yield */, false /* not precleaning */); 5101 MarkFromDirtyCardsClosure 5102 markFromDirtyCardsClosure(this, _span, 5103 NULL, // space is set further below 5104 &_markBitMap, &_markStack, &mrias_cl); 5105 { 5106 GCTraceTime t("grey object rescan", PrintGCDetails, false, _gc_timer_cm); 5107 // Iterate over the dirty cards, setting the corresponding bits in the 5108 // mod union table. 5109 { 5110 ModUnionClosure modUnionClosure(&_modUnionTable); 5111 _ct->ct_bs()->dirty_card_iterate( 5112 _cmsGen->used_region(), 5113 &modUnionClosure); 5114 } 5115 // Having transferred these marks into the modUnionTable, we just need 5116 // to rescan the marked objects on the dirty cards in the modUnionTable. 5117 // The initial marking may have been done during an asynchronous 5118 // collection so there may be dirty bits in the mod-union table. 5119 const int alignment = 5120 CardTableModRefBS::card_size * BitsPerWord; 5121 { 5122 // ... First handle dirty cards in CMS gen 5123 markFromDirtyCardsClosure.set_space(_cmsGen->cmsSpace()); 5124 MemRegion ur = _cmsGen->used_region(); 5125 HeapWord* lb = ur.start(); 5126 HeapWord* ub = (HeapWord*)round_to((intptr_t)ur.end(), alignment); 5127 MemRegion cms_span(lb, ub); 5128 _modUnionTable.dirty_range_iterate_clear(cms_span, 5129 &markFromDirtyCardsClosure); 5130 verify_work_stacks_empty(); 5131 if (PrintCMSStatistics != 0) { 5132 gclog_or_tty->print(" (re-scanned " SIZE_FORMAT " dirty cards in cms gen) ", 5133 markFromDirtyCardsClosure.num_dirty_cards()); 5134 } 5135 } 5136 } 5137 if (VerifyDuringGC && 5138 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 5139 HandleMark hm; // Discard invalid handles created during verification 5140 Universe::verify(); 5141 } 5142 { 5143 GCTraceTime t("root rescan", PrintGCDetails, false, _gc_timer_cm); 5144 5145 verify_work_stacks_empty(); 5146 5147 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 5148 StrongRootsScope srs(1); 5149 5150 gch->gen_process_roots(&srs, 5151 GenCollectedHeap::OldGen, 5152 true, // young gen as roots 5153 GenCollectedHeap::ScanningOption(roots_scanning_options()), 5154 should_unload_classes(), 5155 &mrias_cl, 5156 NULL, 5157 NULL); // The dirty klasses will be handled below 5158 5159 assert(should_unload_classes() 5160 || (roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache), 5161 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 5162 } 5163 5164 { 5165 GCTraceTime t("visit unhandled CLDs", PrintGCDetails, false, _gc_timer_cm); 5166 5167 verify_work_stacks_empty(); 5168 5169 // Scan all class loader data objects that might have been introduced 5170 // during concurrent marking. 5171 ResourceMark rm; 5172 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds(); 5173 for (int i = 0; i < array->length(); i++) { 5174 mrias_cl.do_cld_nv(array->at(i)); 5175 } 5176 5177 // We don't need to keep track of new CLDs anymore. 5178 ClassLoaderDataGraph::remember_new_clds(false); 5179 5180 verify_work_stacks_empty(); 5181 } 5182 5183 { 5184 GCTraceTime t("dirty klass scan", PrintGCDetails, false, _gc_timer_cm); 5185 5186 verify_work_stacks_empty(); 5187 5188 RemarkKlassClosure remark_klass_closure(&mrias_cl); 5189 ClassLoaderDataGraph::classes_do(&remark_klass_closure); 5190 5191 verify_work_stacks_empty(); 5192 } 5193 5194 // We might have added oops to ClassLoaderData::_handles during the 5195 // concurrent marking phase. These oops point to newly allocated objects 5196 // that are guaranteed to be kept alive. Either by the direct allocation 5197 // code, or when the young collector processes the roots. Hence, 5198 // we don't have to revisit the _handles block during the remark phase. 5199 5200 verify_work_stacks_empty(); 5201 // Restore evacuated mark words, if any, used for overflow list links 5202 restore_preserved_marks_if_any(); 5203 5204 verify_overflow_empty(); 5205 } 5206 5207 //////////////////////////////////////////////////////// 5208 // Parallel Reference Processing Task Proxy Class 5209 //////////////////////////////////////////////////////// 5210 class AbstractGangTaskWOopQueues : public AbstractGangTask { 5211 OopTaskQueueSet* _queues; 5212 ParallelTaskTerminator _terminator; 5213 public: 5214 AbstractGangTaskWOopQueues(const char* name, OopTaskQueueSet* queues, uint n_threads) : 5215 AbstractGangTask(name), _queues(queues), _terminator(n_threads, _queues) {} 5216 ParallelTaskTerminator* terminator() { return &_terminator; } 5217 OopTaskQueueSet* queues() { return _queues; } 5218 }; 5219 5220 class CMSRefProcTaskProxy: public AbstractGangTaskWOopQueues { 5221 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask; 5222 CMSCollector* _collector; 5223 CMSBitMap* _mark_bit_map; 5224 const MemRegion _span; 5225 ProcessTask& _task; 5226 5227 public: 5228 CMSRefProcTaskProxy(ProcessTask& task, 5229 CMSCollector* collector, 5230 const MemRegion& span, 5231 CMSBitMap* mark_bit_map, 5232 AbstractWorkGang* workers, 5233 OopTaskQueueSet* task_queues): 5234 AbstractGangTaskWOopQueues("Process referents by policy in parallel", 5235 task_queues, 5236 workers->active_workers()), 5237 _task(task), 5238 _collector(collector), _span(span), _mark_bit_map(mark_bit_map) 5239 { 5240 assert(_collector->_span.equals(_span) && !_span.is_empty(), 5241 "Inconsistency in _span"); 5242 } 5243 5244 OopTaskQueueSet* task_queues() { return queues(); } 5245 5246 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 5247 5248 void do_work_steal(int i, 5249 CMSParDrainMarkingStackClosure* drain, 5250 CMSParKeepAliveClosure* keep_alive, 5251 int* seed); 5252 5253 virtual void work(uint worker_id); 5254 }; 5255 5256 void CMSRefProcTaskProxy::work(uint worker_id) { 5257 ResourceMark rm; 5258 HandleMark hm; 5259 assert(_collector->_span.equals(_span), "Inconsistency in _span"); 5260 CMSParKeepAliveClosure par_keep_alive(_collector, _span, 5261 _mark_bit_map, 5262 work_queue(worker_id)); 5263 CMSParDrainMarkingStackClosure par_drain_stack(_collector, _span, 5264 _mark_bit_map, 5265 work_queue(worker_id)); 5266 CMSIsAliveClosure is_alive_closure(_span, _mark_bit_map); 5267 _task.work(worker_id, is_alive_closure, par_keep_alive, par_drain_stack); 5268 if (_task.marks_oops_alive()) { 5269 do_work_steal(worker_id, &par_drain_stack, &par_keep_alive, 5270 _collector->hash_seed(worker_id)); 5271 } 5272 assert(work_queue(worker_id)->size() == 0, "work_queue should be empty"); 5273 assert(_collector->_overflow_list == NULL, "non-empty _overflow_list"); 5274 } 5275 5276 class CMSRefEnqueueTaskProxy: public AbstractGangTask { 5277 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask; 5278 EnqueueTask& _task; 5279 5280 public: 5281 CMSRefEnqueueTaskProxy(EnqueueTask& task) 5282 : AbstractGangTask("Enqueue reference objects in parallel"), 5283 _task(task) 5284 { } 5285 5286 virtual void work(uint worker_id) 5287 { 5288 _task.work(worker_id); 5289 } 5290 }; 5291 5292 CMSParKeepAliveClosure::CMSParKeepAliveClosure(CMSCollector* collector, 5293 MemRegion span, CMSBitMap* bit_map, OopTaskQueue* work_queue): 5294 _span(span), 5295 _bit_map(bit_map), 5296 _work_queue(work_queue), 5297 _mark_and_push(collector, span, bit_map, work_queue), 5298 _low_water_mark(MIN2((work_queue->max_elems()/4), 5299 ((uint)CMSWorkQueueDrainThreshold * ParallelGCThreads))) 5300 { } 5301 5302 // . see if we can share work_queues with ParNew? XXX 5303 void CMSRefProcTaskProxy::do_work_steal(int i, 5304 CMSParDrainMarkingStackClosure* drain, 5305 CMSParKeepAliveClosure* keep_alive, 5306 int* seed) { 5307 OopTaskQueue* work_q = work_queue(i); 5308 NOT_PRODUCT(int num_steals = 0;) 5309 oop obj_to_scan; 5310 5311 while (true) { 5312 // Completely finish any left over work from (an) earlier round(s) 5313 drain->trim_queue(0); 5314 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 5315 (size_t)ParGCDesiredObjsFromOverflowList); 5316 // Now check if there's any work in the overflow list 5317 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads, 5318 // only affects the number of attempts made to get work from the 5319 // overflow list and does not affect the number of workers. Just 5320 // pass ParallelGCThreads so this behavior is unchanged. 5321 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 5322 work_q, 5323 ParallelGCThreads)) { 5324 // Found something in global overflow list; 5325 // not yet ready to go stealing work from others. 5326 // We'd like to assert(work_q->size() != 0, ...) 5327 // because we just took work from the overflow list, 5328 // but of course we can't, since all of that might have 5329 // been already stolen from us. 5330 continue; 5331 } 5332 // Verify that we have no work before we resort to stealing 5333 assert(work_q->size() == 0, "Have work, shouldn't steal"); 5334 // Try to steal from other queues that have work 5335 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 5336 NOT_PRODUCT(num_steals++;) 5337 assert(obj_to_scan->is_oop(), "Oops, not an oop!"); 5338 assert(_mark_bit_map->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 5339 // Do scanning work 5340 obj_to_scan->oop_iterate(keep_alive); 5341 // Loop around, finish this work, and try to steal some more 5342 } else if (terminator()->offer_termination()) { 5343 break; // nirvana from the infinite cycle 5344 } 5345 } 5346 NOT_PRODUCT( 5347 if (PrintCMSStatistics != 0) { 5348 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals); 5349 } 5350 ) 5351 } 5352 5353 void CMSRefProcTaskExecutor::execute(ProcessTask& task) 5354 { 5355 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5356 WorkGang* workers = gch->workers(); 5357 assert(workers != NULL, "Need parallel worker threads."); 5358 CMSRefProcTaskProxy rp_task(task, &_collector, 5359 _collector.ref_processor()->span(), 5360 _collector.markBitMap(), 5361 workers, _collector.task_queues()); 5362 workers->run_task(&rp_task); 5363 } 5364 5365 void CMSRefProcTaskExecutor::execute(EnqueueTask& task) 5366 { 5367 5368 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5369 WorkGang* workers = gch->workers(); 5370 assert(workers != NULL, "Need parallel worker threads."); 5371 CMSRefEnqueueTaskProxy enq_task(task); 5372 workers->run_task(&enq_task); 5373 } 5374 5375 void CMSCollector::refProcessingWork() { 5376 ResourceMark rm; 5377 HandleMark hm; 5378 5379 ReferenceProcessor* rp = ref_processor(); 5380 assert(rp->span().equals(_span), "Spans should be equal"); 5381 assert(!rp->enqueuing_is_done(), "Enqueuing should not be complete"); 5382 // Process weak references. 5383 rp->setup_policy(false); 5384 verify_work_stacks_empty(); 5385 5386 CMSKeepAliveClosure cmsKeepAliveClosure(this, _span, &_markBitMap, 5387 &_markStack, false /* !preclean */); 5388 CMSDrainMarkingStackClosure cmsDrainMarkingStackClosure(this, 5389 _span, &_markBitMap, &_markStack, 5390 &cmsKeepAliveClosure, false /* !preclean */); 5391 { 5392 GCTraceTime t("weak refs processing", PrintGCDetails, false, _gc_timer_cm); 5393 5394 ReferenceProcessorStats stats; 5395 if (rp->processing_is_mt()) { 5396 // Set the degree of MT here. If the discovery is done MT, there 5397 // may have been a different number of threads doing the discovery 5398 // and a different number of discovered lists may have Ref objects. 5399 // That is OK as long as the Reference lists are balanced (see 5400 // balance_all_queues() and balance_queues()). 5401 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5402 uint active_workers = ParallelGCThreads; 5403 WorkGang* workers = gch->workers(); 5404 if (workers != NULL) { 5405 active_workers = workers->active_workers(); 5406 // The expectation is that active_workers will have already 5407 // been set to a reasonable value. If it has not been set, 5408 // investigate. 5409 assert(active_workers > 0, "Should have been set during scavenge"); 5410 } 5411 rp->set_active_mt_degree(active_workers); 5412 CMSRefProcTaskExecutor task_executor(*this); 5413 stats = rp->process_discovered_references(&_is_alive_closure, 5414 &cmsKeepAliveClosure, 5415 &cmsDrainMarkingStackClosure, 5416 &task_executor, 5417 _gc_timer_cm); 5418 } else { 5419 stats = rp->process_discovered_references(&_is_alive_closure, 5420 &cmsKeepAliveClosure, 5421 &cmsDrainMarkingStackClosure, 5422 NULL, 5423 _gc_timer_cm); 5424 } 5425 _gc_tracer_cm->report_gc_reference_stats(stats); 5426 5427 } 5428 5429 // This is the point where the entire marking should have completed. 5430 verify_work_stacks_empty(); 5431 5432 if (should_unload_classes()) { 5433 { 5434 GCTraceTime t("class unloading", PrintGCDetails, false, _gc_timer_cm); 5435 5436 // Unload classes and purge the SystemDictionary. 5437 bool purged_class = SystemDictionary::do_unloading(&_is_alive_closure); 5438 5439 // Unload nmethods. 5440 CodeCache::do_unloading(&_is_alive_closure, purged_class); 5441 5442 // Prune dead klasses from subklass/sibling/implementor lists. 5443 Klass::clean_weak_klass_links(&_is_alive_closure); 5444 } 5445 5446 { 5447 GCTraceTime t("scrub symbol table", PrintGCDetails, false, _gc_timer_cm); 5448 // Clean up unreferenced symbols in symbol table. 5449 SymbolTable::unlink(); 5450 } 5451 5452 { 5453 GCTraceTime t("scrub string table", PrintGCDetails, false, _gc_timer_cm); 5454 // Delete entries for dead interned strings. 5455 StringTable::unlink(&_is_alive_closure); 5456 } 5457 } 5458 5459 5460 // Restore any preserved marks as a result of mark stack or 5461 // work queue overflow 5462 restore_preserved_marks_if_any(); // done single-threaded for now 5463 5464 rp->set_enqueuing_is_done(true); 5465 if (rp->processing_is_mt()) { 5466 rp->balance_all_queues(); 5467 CMSRefProcTaskExecutor task_executor(*this); 5468 rp->enqueue_discovered_references(&task_executor); 5469 } else { 5470 rp->enqueue_discovered_references(NULL); 5471 } 5472 rp->verify_no_references_recorded(); 5473 assert(!rp->discovery_enabled(), "should have been disabled"); 5474 } 5475 5476 #ifndef PRODUCT 5477 void CMSCollector::check_correct_thread_executing() { 5478 Thread* t = Thread::current(); 5479 // Only the VM thread or the CMS thread should be here. 5480 assert(t->is_ConcurrentGC_thread() || t->is_VM_thread(), 5481 "Unexpected thread type"); 5482 // If this is the vm thread, the foreground process 5483 // should not be waiting. Note that _foregroundGCIsActive is 5484 // true while the foreground collector is waiting. 5485 if (_foregroundGCShouldWait) { 5486 // We cannot be the VM thread 5487 assert(t->is_ConcurrentGC_thread(), 5488 "Should be CMS thread"); 5489 } else { 5490 // We can be the CMS thread only if we are in a stop-world 5491 // phase of CMS collection. 5492 if (t->is_ConcurrentGC_thread()) { 5493 assert(_collectorState == InitialMarking || 5494 _collectorState == FinalMarking, 5495 "Should be a stop-world phase"); 5496 // The CMS thread should be holding the CMS_token. 5497 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 5498 "Potential interference with concurrently " 5499 "executing VM thread"); 5500 } 5501 } 5502 } 5503 #endif 5504 5505 void CMSCollector::sweep() { 5506 assert(_collectorState == Sweeping, "just checking"); 5507 check_correct_thread_executing(); 5508 verify_work_stacks_empty(); 5509 verify_overflow_empty(); 5510 increment_sweep_count(); 5511 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 5512 5513 _inter_sweep_timer.stop(); 5514 _inter_sweep_estimate.sample(_inter_sweep_timer.seconds()); 5515 5516 assert(!_intra_sweep_timer.is_active(), "Should not be active"); 5517 _intra_sweep_timer.reset(); 5518 _intra_sweep_timer.start(); 5519 { 5520 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 5521 CMSPhaseAccounting pa(this, "sweep", !PrintGCDetails); 5522 // First sweep the old gen 5523 { 5524 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(), 5525 bitMapLock()); 5526 sweepWork(_cmsGen); 5527 } 5528 5529 // Update Universe::_heap_*_at_gc figures. 5530 // We need all the free list locks to make the abstract state 5531 // transition from Sweeping to Resetting. See detailed note 5532 // further below. 5533 { 5534 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock()); 5535 // Update heap occupancy information which is used as 5536 // input to soft ref clearing policy at the next gc. 5537 Universe::update_heap_info_at_gc(); 5538 _collectorState = Resizing; 5539 } 5540 } 5541 verify_work_stacks_empty(); 5542 verify_overflow_empty(); 5543 5544 if (should_unload_classes()) { 5545 // Delay purge to the beginning of the next safepoint. Metaspace::contains 5546 // requires that the virtual spaces are stable and not deleted. 5547 ClassLoaderDataGraph::set_should_purge(true); 5548 } 5549 5550 _intra_sweep_timer.stop(); 5551 _intra_sweep_estimate.sample(_intra_sweep_timer.seconds()); 5552 5553 _inter_sweep_timer.reset(); 5554 _inter_sweep_timer.start(); 5555 5556 // We need to use a monotonically non-decreasing time in ms 5557 // or we will see time-warp warnings and os::javaTimeMillis() 5558 // does not guarantee monotonicity. 5559 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC; 5560 update_time_of_last_gc(now); 5561 5562 // NOTE on abstract state transitions: 5563 // Mutators allocate-live and/or mark the mod-union table dirty 5564 // based on the state of the collection. The former is done in 5565 // the interval [Marking, Sweeping] and the latter in the interval 5566 // [Marking, Sweeping). Thus the transitions into the Marking state 5567 // and out of the Sweeping state must be synchronously visible 5568 // globally to the mutators. 5569 // The transition into the Marking state happens with the world 5570 // stopped so the mutators will globally see it. Sweeping is 5571 // done asynchronously by the background collector so the transition 5572 // from the Sweeping state to the Resizing state must be done 5573 // under the freelistLock (as is the check for whether to 5574 // allocate-live and whether to dirty the mod-union table). 5575 assert(_collectorState == Resizing, "Change of collector state to" 5576 " Resizing must be done under the freelistLocks (plural)"); 5577 5578 // Now that sweeping has been completed, we clear 5579 // the incremental_collection_failed flag, 5580 // thus inviting a younger gen collection to promote into 5581 // this generation. If such a promotion may still fail, 5582 // the flag will be set again when a young collection is 5583 // attempted. 5584 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5585 gch->clear_incremental_collection_failed(); // Worth retrying as fresh space may have been freed up 5586 gch->update_full_collections_completed(_collection_count_start); 5587 } 5588 5589 // FIX ME!!! Looks like this belongs in CFLSpace, with 5590 // CMSGen merely delegating to it. 5591 void ConcurrentMarkSweepGeneration::setNearLargestChunk() { 5592 double nearLargestPercent = FLSLargestBlockCoalesceProximity; 5593 HeapWord* minAddr = _cmsSpace->bottom(); 5594 HeapWord* largestAddr = 5595 (HeapWord*) _cmsSpace->dictionary()->find_largest_dict(); 5596 if (largestAddr == NULL) { 5597 // The dictionary appears to be empty. In this case 5598 // try to coalesce at the end of the heap. 5599 largestAddr = _cmsSpace->end(); 5600 } 5601 size_t largestOffset = pointer_delta(largestAddr, minAddr); 5602 size_t nearLargestOffset = 5603 (size_t)((double)largestOffset * nearLargestPercent) - MinChunkSize; 5604 if (PrintFLSStatistics != 0) { 5605 gclog_or_tty->print_cr( 5606 "CMS: Large Block: " PTR_FORMAT ";" 5607 " Proximity: " PTR_FORMAT " -> " PTR_FORMAT, 5608 p2i(largestAddr), 5609 p2i(_cmsSpace->nearLargestChunk()), p2i(minAddr + nearLargestOffset)); 5610 } 5611 _cmsSpace->set_nearLargestChunk(minAddr + nearLargestOffset); 5612 } 5613 5614 bool ConcurrentMarkSweepGeneration::isNearLargestChunk(HeapWord* addr) { 5615 return addr >= _cmsSpace->nearLargestChunk(); 5616 } 5617 5618 FreeChunk* ConcurrentMarkSweepGeneration::find_chunk_at_end() { 5619 return _cmsSpace->find_chunk_at_end(); 5620 } 5621 5622 void ConcurrentMarkSweepGeneration::update_gc_stats(Generation* current_generation, 5623 bool full) { 5624 // If the young generation has been collected, gather any statistics 5625 // that are of interest at this point. 5626 bool current_is_young = GenCollectedHeap::heap()->is_young_gen(current_generation); 5627 if (!full && current_is_young) { 5628 // Gather statistics on the young generation collection. 5629 collector()->stats().record_gc0_end(used()); 5630 } 5631 } 5632 5633 void CMSCollector::sweepWork(ConcurrentMarkSweepGeneration* old_gen) { 5634 // We iterate over the space(s) underlying this generation, 5635 // checking the mark bit map to see if the bits corresponding 5636 // to specific blocks are marked or not. Blocks that are 5637 // marked are live and are not swept up. All remaining blocks 5638 // are swept up, with coalescing on-the-fly as we sweep up 5639 // contiguous free and/or garbage blocks: 5640 // We need to ensure that the sweeper synchronizes with allocators 5641 // and stop-the-world collectors. In particular, the following 5642 // locks are used: 5643 // . CMS token: if this is held, a stop the world collection cannot occur 5644 // . freelistLock: if this is held no allocation can occur from this 5645 // generation by another thread 5646 // . bitMapLock: if this is held, no other thread can access or update 5647 // 5648 5649 // Note that we need to hold the freelistLock if we use 5650 // block iterate below; else the iterator might go awry if 5651 // a mutator (or promotion) causes block contents to change 5652 // (for instance if the allocator divvies up a block). 5653 // If we hold the free list lock, for all practical purposes 5654 // young generation GC's can't occur (they'll usually need to 5655 // promote), so we might as well prevent all young generation 5656 // GC's while we do a sweeping step. For the same reason, we might 5657 // as well take the bit map lock for the entire duration 5658 5659 // check that we hold the requisite locks 5660 assert(have_cms_token(), "Should hold cms token"); 5661 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), "Should possess CMS token to sweep"); 5662 assert_lock_strong(old_gen->freelistLock()); 5663 assert_lock_strong(bitMapLock()); 5664 5665 assert(!_inter_sweep_timer.is_active(), "Was switched off in an outer context"); 5666 assert(_intra_sweep_timer.is_active(), "Was switched on in an outer context"); 5667 old_gen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()), 5668 _inter_sweep_estimate.padded_average(), 5669 _intra_sweep_estimate.padded_average()); 5670 old_gen->setNearLargestChunk(); 5671 5672 { 5673 SweepClosure sweepClosure(this, old_gen, &_markBitMap, CMSYield); 5674 old_gen->cmsSpace()->blk_iterate_careful(&sweepClosure); 5675 // We need to free-up/coalesce garbage/blocks from a 5676 // co-terminal free run. This is done in the SweepClosure 5677 // destructor; so, do not remove this scope, else the 5678 // end-of-sweep-census below will be off by a little bit. 5679 } 5680 old_gen->cmsSpace()->sweep_completed(); 5681 old_gen->cmsSpace()->endSweepFLCensus(sweep_count()); 5682 if (should_unload_classes()) { // unloaded classes this cycle, 5683 _concurrent_cycles_since_last_unload = 0; // ... reset count 5684 } else { // did not unload classes, 5685 _concurrent_cycles_since_last_unload++; // ... increment count 5686 } 5687 } 5688 5689 // Reset CMS data structures (for now just the marking bit map) 5690 // preparatory for the next cycle. 5691 void CMSCollector::reset_concurrent() { 5692 CMSTokenSyncWithLocks ts(true, bitMapLock()); 5693 5694 // If the state is not "Resetting", the foreground thread 5695 // has done a collection and the resetting. 5696 if (_collectorState != Resetting) { 5697 assert(_collectorState == Idling, "The state should only change" 5698 " because the foreground collector has finished the collection"); 5699 return; 5700 } 5701 5702 // Clear the mark bitmap (no grey objects to start with) 5703 // for the next cycle. 5704 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 5705 CMSPhaseAccounting cmspa(this, "reset", !PrintGCDetails); 5706 5707 HeapWord* curAddr = _markBitMap.startWord(); 5708 while (curAddr < _markBitMap.endWord()) { 5709 size_t remaining = pointer_delta(_markBitMap.endWord(), curAddr); 5710 MemRegion chunk(curAddr, MIN2(CMSBitMapYieldQuantum, remaining)); 5711 _markBitMap.clear_large_range(chunk); 5712 if (ConcurrentMarkSweepThread::should_yield() && 5713 !foregroundGCIsActive() && 5714 CMSYield) { 5715 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 5716 "CMS thread should hold CMS token"); 5717 assert_lock_strong(bitMapLock()); 5718 bitMapLock()->unlock(); 5719 ConcurrentMarkSweepThread::desynchronize(true); 5720 stopTimer(); 5721 if (PrintCMSStatistics != 0) { 5722 incrementYields(); 5723 } 5724 5725 // See the comment in coordinator_yield() 5726 for (unsigned i = 0; i < CMSYieldSleepCount && 5727 ConcurrentMarkSweepThread::should_yield() && 5728 !CMSCollector::foregroundGCIsActive(); ++i) { 5729 os::sleep(Thread::current(), 1, false); 5730 } 5731 5732 ConcurrentMarkSweepThread::synchronize(true); 5733 bitMapLock()->lock_without_safepoint_check(); 5734 startTimer(); 5735 } 5736 curAddr = chunk.end(); 5737 } 5738 // A successful mostly concurrent collection has been done. 5739 // Because only the full (i.e., concurrent mode failure) collections 5740 // are being measured for gc overhead limits, clean the "near" flag 5741 // and count. 5742 size_policy()->reset_gc_overhead_limit_count(); 5743 _collectorState = Idling; 5744 5745 register_gc_end(); 5746 } 5747 5748 // Same as above but for STW paths 5749 void CMSCollector::reset_stw() { 5750 // already have the lock 5751 assert(_collectorState == Resetting, "just checking"); 5752 assert_lock_strong(bitMapLock()); 5753 GCIdMarkAndRestore gc_id_mark(_cmsThread->gc_id()); 5754 _markBitMap.clear_all(); 5755 _collectorState = Idling; 5756 register_gc_end(); 5757 } 5758 5759 void CMSCollector::do_CMS_operation(CMS_op_type op, GCCause::Cause gc_cause) { 5760 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 5761 GCTraceTime t(GCCauseString("GC", gc_cause), PrintGC, !PrintGCDetails, NULL); 5762 TraceCollectorStats tcs(counters()); 5763 5764 switch (op) { 5765 case CMS_op_checkpointRootsInitial: { 5766 SvcGCMarker sgcm(SvcGCMarker::OTHER); 5767 checkpointRootsInitial(); 5768 if (PrintGC) { 5769 _cmsGen->printOccupancy("initial-mark"); 5770 } 5771 break; 5772 } 5773 case CMS_op_checkpointRootsFinal: { 5774 SvcGCMarker sgcm(SvcGCMarker::OTHER); 5775 checkpointRootsFinal(); 5776 if (PrintGC) { 5777 _cmsGen->printOccupancy("remark"); 5778 } 5779 break; 5780 } 5781 default: 5782 fatal("No such CMS_op"); 5783 } 5784 } 5785 5786 #ifndef PRODUCT 5787 size_t const CMSCollector::skip_header_HeapWords() { 5788 return FreeChunk::header_size(); 5789 } 5790 5791 // Try and collect here conditions that should hold when 5792 // CMS thread is exiting. The idea is that the foreground GC 5793 // thread should not be blocked if it wants to terminate 5794 // the CMS thread and yet continue to run the VM for a while 5795 // after that. 5796 void CMSCollector::verify_ok_to_terminate() const { 5797 assert(Thread::current()->is_ConcurrentGC_thread(), 5798 "should be called by CMS thread"); 5799 assert(!_foregroundGCShouldWait, "should be false"); 5800 // We could check here that all the various low-level locks 5801 // are not held by the CMS thread, but that is overkill; see 5802 // also CMSThread::verify_ok_to_terminate() where the CGC_lock 5803 // is checked. 5804 } 5805 #endif 5806 5807 size_t CMSCollector::block_size_using_printezis_bits(HeapWord* addr) const { 5808 assert(_markBitMap.isMarked(addr) && _markBitMap.isMarked(addr + 1), 5809 "missing Printezis mark?"); 5810 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 5811 size_t size = pointer_delta(nextOneAddr + 1, addr); 5812 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 5813 "alignment problem"); 5814 assert(size >= 3, "Necessary for Printezis marks to work"); 5815 return size; 5816 } 5817 5818 // A variant of the above (block_size_using_printezis_bits()) except 5819 // that we return 0 if the P-bits are not yet set. 5820 size_t CMSCollector::block_size_if_printezis_bits(HeapWord* addr) const { 5821 if (_markBitMap.isMarked(addr + 1)) { 5822 assert(_markBitMap.isMarked(addr), "P-bit can be set only for marked objects"); 5823 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 5824 size_t size = pointer_delta(nextOneAddr + 1, addr); 5825 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 5826 "alignment problem"); 5827 assert(size >= 3, "Necessary for Printezis marks to work"); 5828 return size; 5829 } 5830 return 0; 5831 } 5832 5833 HeapWord* CMSCollector::next_card_start_after_block(HeapWord* addr) const { 5834 size_t sz = 0; 5835 oop p = (oop)addr; 5836 if (p->klass_or_null() != NULL) { 5837 sz = CompactibleFreeListSpace::adjustObjectSize(p->size()); 5838 } else { 5839 sz = block_size_using_printezis_bits(addr); 5840 } 5841 assert(sz > 0, "size must be nonzero"); 5842 HeapWord* next_block = addr + sz; 5843 HeapWord* next_card = (HeapWord*)round_to((uintptr_t)next_block, 5844 CardTableModRefBS::card_size); 5845 assert(round_down((uintptr_t)addr, CardTableModRefBS::card_size) < 5846 round_down((uintptr_t)next_card, CardTableModRefBS::card_size), 5847 "must be different cards"); 5848 return next_card; 5849 } 5850 5851 5852 // CMS Bit Map Wrapper ///////////////////////////////////////// 5853 5854 // Construct a CMS bit map infrastructure, but don't create the 5855 // bit vector itself. That is done by a separate call CMSBitMap::allocate() 5856 // further below. 5857 CMSBitMap::CMSBitMap(int shifter, int mutex_rank, const char* mutex_name): 5858 _bm(), 5859 _shifter(shifter), 5860 _lock(mutex_rank >= 0 ? new Mutex(mutex_rank, mutex_name, true, 5861 Monitor::_safepoint_check_sometimes) : NULL) 5862 { 5863 _bmStartWord = 0; 5864 _bmWordSize = 0; 5865 } 5866 5867 bool CMSBitMap::allocate(MemRegion mr) { 5868 _bmStartWord = mr.start(); 5869 _bmWordSize = mr.word_size(); 5870 ReservedSpace brs(ReservedSpace::allocation_align_size_up( 5871 (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1)); 5872 if (!brs.is_reserved()) { 5873 warning("CMS bit map allocation failure"); 5874 return false; 5875 } 5876 // For now we'll just commit all of the bit map up front. 5877 // Later on we'll try to be more parsimonious with swap. 5878 if (!_virtual_space.initialize(brs, brs.size())) { 5879 warning("CMS bit map backing store failure"); 5880 return false; 5881 } 5882 assert(_virtual_space.committed_size() == brs.size(), 5883 "didn't reserve backing store for all of CMS bit map?"); 5884 _bm.set_map((BitMap::bm_word_t*)_virtual_space.low()); 5885 assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >= 5886 _bmWordSize, "inconsistency in bit map sizing"); 5887 _bm.set_size(_bmWordSize >> _shifter); 5888 5889 // bm.clear(); // can we rely on getting zero'd memory? verify below 5890 assert(isAllClear(), 5891 "Expected zero'd memory from ReservedSpace constructor"); 5892 assert(_bm.size() == heapWordDiffToOffsetDiff(sizeInWords()), 5893 "consistency check"); 5894 return true; 5895 } 5896 5897 void CMSBitMap::dirty_range_iterate_clear(MemRegion mr, MemRegionClosure* cl) { 5898 HeapWord *next_addr, *end_addr, *last_addr; 5899 assert_locked(); 5900 assert(covers(mr), "out-of-range error"); 5901 // XXX assert that start and end are appropriately aligned 5902 for (next_addr = mr.start(), end_addr = mr.end(); 5903 next_addr < end_addr; next_addr = last_addr) { 5904 MemRegion dirty_region = getAndClearMarkedRegion(next_addr, end_addr); 5905 last_addr = dirty_region.end(); 5906 if (!dirty_region.is_empty()) { 5907 cl->do_MemRegion(dirty_region); 5908 } else { 5909 assert(last_addr == end_addr, "program logic"); 5910 return; 5911 } 5912 } 5913 } 5914 5915 void CMSBitMap::print_on_error(outputStream* st, const char* prefix) const { 5916 _bm.print_on_error(st, prefix); 5917 } 5918 5919 #ifndef PRODUCT 5920 void CMSBitMap::assert_locked() const { 5921 CMSLockVerifier::assert_locked(lock()); 5922 } 5923 5924 bool CMSBitMap::covers(MemRegion mr) const { 5925 // assert(_bm.map() == _virtual_space.low(), "map inconsistency"); 5926 assert((size_t)_bm.size() == (_bmWordSize >> _shifter), 5927 "size inconsistency"); 5928 return (mr.start() >= _bmStartWord) && 5929 (mr.end() <= endWord()); 5930 } 5931 5932 bool CMSBitMap::covers(HeapWord* start, size_t size) const { 5933 return (start >= _bmStartWord && (start + size) <= endWord()); 5934 } 5935 5936 void CMSBitMap::verifyNoOneBitsInRange(HeapWord* left, HeapWord* right) { 5937 // verify that there are no 1 bits in the interval [left, right) 5938 FalseBitMapClosure falseBitMapClosure; 5939 iterate(&falseBitMapClosure, left, right); 5940 } 5941 5942 void CMSBitMap::region_invariant(MemRegion mr) 5943 { 5944 assert_locked(); 5945 // mr = mr.intersection(MemRegion(_bmStartWord, _bmWordSize)); 5946 assert(!mr.is_empty(), "unexpected empty region"); 5947 assert(covers(mr), "mr should be covered by bit map"); 5948 // convert address range into offset range 5949 size_t start_ofs = heapWordToOffset(mr.start()); 5950 // Make sure that end() is appropriately aligned 5951 assert(mr.end() == (HeapWord*)round_to((intptr_t)mr.end(), 5952 (1 << (_shifter+LogHeapWordSize))), 5953 "Misaligned mr.end()"); 5954 size_t end_ofs = heapWordToOffset(mr.end()); 5955 assert(end_ofs > start_ofs, "Should mark at least one bit"); 5956 } 5957 5958 #endif 5959 5960 bool CMSMarkStack::allocate(size_t size) { 5961 // allocate a stack of the requisite depth 5962 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 5963 size * sizeof(oop))); 5964 if (!rs.is_reserved()) { 5965 warning("CMSMarkStack allocation failure"); 5966 return false; 5967 } 5968 if (!_virtual_space.initialize(rs, rs.size())) { 5969 warning("CMSMarkStack backing store failure"); 5970 return false; 5971 } 5972 assert(_virtual_space.committed_size() == rs.size(), 5973 "didn't reserve backing store for all of CMS stack?"); 5974 _base = (oop*)(_virtual_space.low()); 5975 _index = 0; 5976 _capacity = size; 5977 NOT_PRODUCT(_max_depth = 0); 5978 return true; 5979 } 5980 5981 // XXX FIX ME !!! In the MT case we come in here holding a 5982 // leaf lock. For printing we need to take a further lock 5983 // which has lower rank. We need to recalibrate the two 5984 // lock-ranks involved in order to be able to print the 5985 // messages below. (Or defer the printing to the caller. 5986 // For now we take the expedient path of just disabling the 5987 // messages for the problematic case.) 5988 void CMSMarkStack::expand() { 5989 assert(_capacity <= MarkStackSizeMax, "stack bigger than permitted"); 5990 if (_capacity == MarkStackSizeMax) { 5991 if (_hit_limit++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) { 5992 // We print a warning message only once per CMS cycle. 5993 gclog_or_tty->print_cr(" (benign) Hit CMSMarkStack max size limit"); 5994 } 5995 return; 5996 } 5997 // Double capacity if possible 5998 size_t new_capacity = MIN2(_capacity*2, MarkStackSizeMax); 5999 // Do not give up existing stack until we have managed to 6000 // get the double capacity that we desired. 6001 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 6002 new_capacity * sizeof(oop))); 6003 if (rs.is_reserved()) { 6004 // Release the backing store associated with old stack 6005 _virtual_space.release(); 6006 // Reinitialize virtual space for new stack 6007 if (!_virtual_space.initialize(rs, rs.size())) { 6008 fatal("Not enough swap for expanded marking stack"); 6009 } 6010 _base = (oop*)(_virtual_space.low()); 6011 _index = 0; 6012 _capacity = new_capacity; 6013 } else if (_failed_double++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) { 6014 // Failed to double capacity, continue; 6015 // we print a detail message only once per CMS cycle. 6016 gclog_or_tty->print(" (benign) Failed to expand marking stack from " SIZE_FORMAT "K to " 6017 SIZE_FORMAT "K", 6018 _capacity / K, new_capacity / K); 6019 } 6020 } 6021 6022 6023 // Closures 6024 // XXX: there seems to be a lot of code duplication here; 6025 // should refactor and consolidate common code. 6026 6027 // This closure is used to mark refs into the CMS generation in 6028 // the CMS bit map. Called at the first checkpoint. This closure 6029 // assumes that we do not need to re-mark dirty cards; if the CMS 6030 // generation on which this is used is not an oldest 6031 // generation then this will lose younger_gen cards! 6032 6033 MarkRefsIntoClosure::MarkRefsIntoClosure( 6034 MemRegion span, CMSBitMap* bitMap): 6035 _span(span), 6036 _bitMap(bitMap) 6037 { 6038 assert(ref_processor() == NULL, "deliberately left NULL"); 6039 assert(_bitMap->covers(_span), "_bitMap/_span mismatch"); 6040 } 6041 6042 void MarkRefsIntoClosure::do_oop(oop obj) { 6043 // if p points into _span, then mark corresponding bit in _markBitMap 6044 assert(obj->is_oop(), "expected an oop"); 6045 HeapWord* addr = (HeapWord*)obj; 6046 if (_span.contains(addr)) { 6047 // this should be made more efficient 6048 _bitMap->mark(addr); 6049 } 6050 } 6051 6052 void MarkRefsIntoClosure::do_oop(oop* p) { MarkRefsIntoClosure::do_oop_work(p); } 6053 void MarkRefsIntoClosure::do_oop(narrowOop* p) { MarkRefsIntoClosure::do_oop_work(p); } 6054 6055 Par_MarkRefsIntoClosure::Par_MarkRefsIntoClosure( 6056 MemRegion span, CMSBitMap* bitMap): 6057 _span(span), 6058 _bitMap(bitMap) 6059 { 6060 assert(ref_processor() == NULL, "deliberately left NULL"); 6061 assert(_bitMap->covers(_span), "_bitMap/_span mismatch"); 6062 } 6063 6064 void Par_MarkRefsIntoClosure::do_oop(oop obj) { 6065 // if p points into _span, then mark corresponding bit in _markBitMap 6066 assert(obj->is_oop(), "expected an oop"); 6067 HeapWord* addr = (HeapWord*)obj; 6068 if (_span.contains(addr)) { 6069 // this should be made more efficient 6070 _bitMap->par_mark(addr); 6071 } 6072 } 6073 6074 void Par_MarkRefsIntoClosure::do_oop(oop* p) { Par_MarkRefsIntoClosure::do_oop_work(p); } 6075 void Par_MarkRefsIntoClosure::do_oop(narrowOop* p) { Par_MarkRefsIntoClosure::do_oop_work(p); } 6076 6077 // A variant of the above, used for CMS marking verification. 6078 MarkRefsIntoVerifyClosure::MarkRefsIntoVerifyClosure( 6079 MemRegion span, CMSBitMap* verification_bm, CMSBitMap* cms_bm): 6080 _span(span), 6081 _verification_bm(verification_bm), 6082 _cms_bm(cms_bm) 6083 { 6084 assert(ref_processor() == NULL, "deliberately left NULL"); 6085 assert(_verification_bm->covers(_span), "_verification_bm/_span mismatch"); 6086 } 6087 6088 void MarkRefsIntoVerifyClosure::do_oop(oop obj) { 6089 // if p points into _span, then mark corresponding bit in _markBitMap 6090 assert(obj->is_oop(), "expected an oop"); 6091 HeapWord* addr = (HeapWord*)obj; 6092 if (_span.contains(addr)) { 6093 _verification_bm->mark(addr); 6094 if (!_cms_bm->isMarked(addr)) { 6095 oop(addr)->print(); 6096 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)", p2i(addr)); 6097 fatal("... aborting"); 6098 } 6099 } 6100 } 6101 6102 void MarkRefsIntoVerifyClosure::do_oop(oop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); } 6103 void MarkRefsIntoVerifyClosure::do_oop(narrowOop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); } 6104 6105 ////////////////////////////////////////////////// 6106 // MarkRefsIntoAndScanClosure 6107 ////////////////////////////////////////////////// 6108 6109 MarkRefsIntoAndScanClosure::MarkRefsIntoAndScanClosure(MemRegion span, 6110 ReferenceProcessor* rp, 6111 CMSBitMap* bit_map, 6112 CMSBitMap* mod_union_table, 6113 CMSMarkStack* mark_stack, 6114 CMSCollector* collector, 6115 bool should_yield, 6116 bool concurrent_precleaning): 6117 _collector(collector), 6118 _span(span), 6119 _bit_map(bit_map), 6120 _mark_stack(mark_stack), 6121 _pushAndMarkClosure(collector, span, rp, bit_map, mod_union_table, 6122 mark_stack, concurrent_precleaning), 6123 _yield(should_yield), 6124 _concurrent_precleaning(concurrent_precleaning), 6125 _freelistLock(NULL) 6126 { 6127 // FIXME: Should initialize in base class constructor. 6128 assert(rp != NULL, "ref_processor shouldn't be NULL"); 6129 set_ref_processor_internal(rp); 6130 } 6131 6132 // This closure is used to mark refs into the CMS generation at the 6133 // second (final) checkpoint, and to scan and transitively follow 6134 // the unmarked oops. It is also used during the concurrent precleaning 6135 // phase while scanning objects on dirty cards in the CMS generation. 6136 // The marks are made in the marking bit map and the marking stack is 6137 // used for keeping the (newly) grey objects during the scan. 6138 // The parallel version (Par_...) appears further below. 6139 void MarkRefsIntoAndScanClosure::do_oop(oop obj) { 6140 if (obj != NULL) { 6141 assert(obj->is_oop(), "expected an oop"); 6142 HeapWord* addr = (HeapWord*)obj; 6143 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 6144 assert(_collector->overflow_list_is_empty(), 6145 "overflow list should be empty"); 6146 if (_span.contains(addr) && 6147 !_bit_map->isMarked(addr)) { 6148 // mark bit map (object is now grey) 6149 _bit_map->mark(addr); 6150 // push on marking stack (stack should be empty), and drain the 6151 // stack by applying this closure to the oops in the oops popped 6152 // from the stack (i.e. blacken the grey objects) 6153 bool res = _mark_stack->push(obj); 6154 assert(res, "Should have space to push on empty stack"); 6155 do { 6156 oop new_oop = _mark_stack->pop(); 6157 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 6158 assert(_bit_map->isMarked((HeapWord*)new_oop), 6159 "only grey objects on this stack"); 6160 // iterate over the oops in this oop, marking and pushing 6161 // the ones in CMS heap (i.e. in _span). 6162 new_oop->oop_iterate(&_pushAndMarkClosure); 6163 // check if it's time to yield 6164 do_yield_check(); 6165 } while (!_mark_stack->isEmpty() || 6166 (!_concurrent_precleaning && take_from_overflow_list())); 6167 // if marking stack is empty, and we are not doing this 6168 // during precleaning, then check the overflow list 6169 } 6170 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 6171 assert(_collector->overflow_list_is_empty(), 6172 "overflow list was drained above"); 6173 6174 assert(_collector->no_preserved_marks(), 6175 "All preserved marks should have been restored above"); 6176 } 6177 } 6178 6179 void MarkRefsIntoAndScanClosure::do_oop(oop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); } 6180 void MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); } 6181 6182 void MarkRefsIntoAndScanClosure::do_yield_work() { 6183 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6184 "CMS thread should hold CMS token"); 6185 assert_lock_strong(_freelistLock); 6186 assert_lock_strong(_bit_map->lock()); 6187 // relinquish the free_list_lock and bitMaplock() 6188 _bit_map->lock()->unlock(); 6189 _freelistLock->unlock(); 6190 ConcurrentMarkSweepThread::desynchronize(true); 6191 _collector->stopTimer(); 6192 if (PrintCMSStatistics != 0) { 6193 _collector->incrementYields(); 6194 } 6195 6196 // See the comment in coordinator_yield() 6197 for (unsigned i = 0; 6198 i < CMSYieldSleepCount && 6199 ConcurrentMarkSweepThread::should_yield() && 6200 !CMSCollector::foregroundGCIsActive(); 6201 ++i) { 6202 os::sleep(Thread::current(), 1, false); 6203 } 6204 6205 ConcurrentMarkSweepThread::synchronize(true); 6206 _freelistLock->lock_without_safepoint_check(); 6207 _bit_map->lock()->lock_without_safepoint_check(); 6208 _collector->startTimer(); 6209 } 6210 6211 /////////////////////////////////////////////////////////// 6212 // Par_MarkRefsIntoAndScanClosure: a parallel version of 6213 // MarkRefsIntoAndScanClosure 6214 /////////////////////////////////////////////////////////// 6215 Par_MarkRefsIntoAndScanClosure::Par_MarkRefsIntoAndScanClosure( 6216 CMSCollector* collector, MemRegion span, ReferenceProcessor* rp, 6217 CMSBitMap* bit_map, OopTaskQueue* work_queue): 6218 _span(span), 6219 _bit_map(bit_map), 6220 _work_queue(work_queue), 6221 _low_water_mark(MIN2((work_queue->max_elems()/4), 6222 ((uint)CMSWorkQueueDrainThreshold * ParallelGCThreads))), 6223 _par_pushAndMarkClosure(collector, span, rp, bit_map, work_queue) 6224 { 6225 // FIXME: Should initialize in base class constructor. 6226 assert(rp != NULL, "ref_processor shouldn't be NULL"); 6227 set_ref_processor_internal(rp); 6228 } 6229 6230 // This closure is used to mark refs into the CMS generation at the 6231 // second (final) checkpoint, and to scan and transitively follow 6232 // the unmarked oops. The marks are made in the marking bit map and 6233 // the work_queue is used for keeping the (newly) grey objects during 6234 // the scan phase whence they are also available for stealing by parallel 6235 // threads. Since the marking bit map is shared, updates are 6236 // synchronized (via CAS). 6237 void Par_MarkRefsIntoAndScanClosure::do_oop(oop obj) { 6238 if (obj != NULL) { 6239 // Ignore mark word because this could be an already marked oop 6240 // that may be chained at the end of the overflow list. 6241 assert(obj->is_oop(true), "expected an oop"); 6242 HeapWord* addr = (HeapWord*)obj; 6243 if (_span.contains(addr) && 6244 !_bit_map->isMarked(addr)) { 6245 // mark bit map (object will become grey): 6246 // It is possible for several threads to be 6247 // trying to "claim" this object concurrently; 6248 // the unique thread that succeeds in marking the 6249 // object first will do the subsequent push on 6250 // to the work queue (or overflow list). 6251 if (_bit_map->par_mark(addr)) { 6252 // push on work_queue (which may not be empty), and trim the 6253 // queue to an appropriate length by applying this closure to 6254 // the oops in the oops popped from the stack (i.e. blacken the 6255 // grey objects) 6256 bool res = _work_queue->push(obj); 6257 assert(res, "Low water mark should be less than capacity?"); 6258 trim_queue(_low_water_mark); 6259 } // Else, another thread claimed the object 6260 } 6261 } 6262 } 6263 6264 void Par_MarkRefsIntoAndScanClosure::do_oop(oop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); } 6265 void Par_MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); } 6266 6267 // This closure is used to rescan the marked objects on the dirty cards 6268 // in the mod union table and the card table proper. 6269 size_t ScanMarkedObjectsAgainCarefullyClosure::do_object_careful_m( 6270 oop p, MemRegion mr) { 6271 6272 size_t size = 0; 6273 HeapWord* addr = (HeapWord*)p; 6274 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6275 assert(_span.contains(addr), "we are scanning the CMS generation"); 6276 // check if it's time to yield 6277 if (do_yield_check()) { 6278 // We yielded for some foreground stop-world work, 6279 // and we have been asked to abort this ongoing preclean cycle. 6280 return 0; 6281 } 6282 if (_bitMap->isMarked(addr)) { 6283 // it's marked; is it potentially uninitialized? 6284 if (p->klass_or_null() != NULL) { 6285 // an initialized object; ignore mark word in verification below 6286 // since we are running concurrent with mutators 6287 assert(p->is_oop(true), "should be an oop"); 6288 if (p->is_objArray()) { 6289 // objArrays are precisely marked; restrict scanning 6290 // to dirty cards only. 6291 size = CompactibleFreeListSpace::adjustObjectSize( 6292 p->oop_iterate_size(_scanningClosure, mr)); 6293 } else { 6294 // A non-array may have been imprecisely marked; we need 6295 // to scan object in its entirety. 6296 size = CompactibleFreeListSpace::adjustObjectSize( 6297 p->oop_iterate_size(_scanningClosure)); 6298 } 6299 #ifdef ASSERT 6300 size_t direct_size = 6301 CompactibleFreeListSpace::adjustObjectSize(p->size()); 6302 assert(size == direct_size, "Inconsistency in size"); 6303 assert(size >= 3, "Necessary for Printezis marks to work"); 6304 if (!_bitMap->isMarked(addr+1)) { 6305 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size); 6306 } else { 6307 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size-1); 6308 assert(_bitMap->isMarked(addr+size-1), 6309 "inconsistent Printezis mark"); 6310 } 6311 #endif // ASSERT 6312 } else { 6313 // An uninitialized object. 6314 assert(_bitMap->isMarked(addr+1), "missing Printezis mark?"); 6315 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 6316 size = pointer_delta(nextOneAddr + 1, addr); 6317 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 6318 "alignment problem"); 6319 // Note that pre-cleaning needn't redirty the card. OopDesc::set_klass() 6320 // will dirty the card when the klass pointer is installed in the 6321 // object (signaling the completion of initialization). 6322 } 6323 } else { 6324 // Either a not yet marked object or an uninitialized object 6325 if (p->klass_or_null() == NULL) { 6326 // An uninitialized object, skip to the next card, since 6327 // we may not be able to read its P-bits yet. 6328 assert(size == 0, "Initial value"); 6329 } else { 6330 // An object not (yet) reached by marking: we merely need to 6331 // compute its size so as to go look at the next block. 6332 assert(p->is_oop(true), "should be an oop"); 6333 size = CompactibleFreeListSpace::adjustObjectSize(p->size()); 6334 } 6335 } 6336 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6337 return size; 6338 } 6339 6340 void ScanMarkedObjectsAgainCarefullyClosure::do_yield_work() { 6341 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6342 "CMS thread should hold CMS token"); 6343 assert_lock_strong(_freelistLock); 6344 assert_lock_strong(_bitMap->lock()); 6345 // relinquish the free_list_lock and bitMaplock() 6346 _bitMap->lock()->unlock(); 6347 _freelistLock->unlock(); 6348 ConcurrentMarkSweepThread::desynchronize(true); 6349 _collector->stopTimer(); 6350 if (PrintCMSStatistics != 0) { 6351 _collector->incrementYields(); 6352 } 6353 6354 // See the comment in coordinator_yield() 6355 for (unsigned i = 0; i < CMSYieldSleepCount && 6356 ConcurrentMarkSweepThread::should_yield() && 6357 !CMSCollector::foregroundGCIsActive(); ++i) { 6358 os::sleep(Thread::current(), 1, false); 6359 } 6360 6361 ConcurrentMarkSweepThread::synchronize(true); 6362 _freelistLock->lock_without_safepoint_check(); 6363 _bitMap->lock()->lock_without_safepoint_check(); 6364 _collector->startTimer(); 6365 } 6366 6367 6368 ////////////////////////////////////////////////////////////////// 6369 // SurvivorSpacePrecleanClosure 6370 ////////////////////////////////////////////////////////////////// 6371 // This (single-threaded) closure is used to preclean the oops in 6372 // the survivor spaces. 6373 size_t SurvivorSpacePrecleanClosure::do_object_careful(oop p) { 6374 6375 HeapWord* addr = (HeapWord*)p; 6376 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6377 assert(!_span.contains(addr), "we are scanning the survivor spaces"); 6378 assert(p->klass_or_null() != NULL, "object should be initialized"); 6379 // an initialized object; ignore mark word in verification below 6380 // since we are running concurrent with mutators 6381 assert(p->is_oop(true), "should be an oop"); 6382 // Note that we do not yield while we iterate over 6383 // the interior oops of p, pushing the relevant ones 6384 // on our marking stack. 6385 size_t size = p->oop_iterate_size(_scanning_closure); 6386 do_yield_check(); 6387 // Observe that below, we do not abandon the preclean 6388 // phase as soon as we should; rather we empty the 6389 // marking stack before returning. This is to satisfy 6390 // some existing assertions. In general, it may be a 6391 // good idea to abort immediately and complete the marking 6392 // from the grey objects at a later time. 6393 while (!_mark_stack->isEmpty()) { 6394 oop new_oop = _mark_stack->pop(); 6395 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 6396 assert(_bit_map->isMarked((HeapWord*)new_oop), 6397 "only grey objects on this stack"); 6398 // iterate over the oops in this oop, marking and pushing 6399 // the ones in CMS heap (i.e. in _span). 6400 new_oop->oop_iterate(_scanning_closure); 6401 // check if it's time to yield 6402 do_yield_check(); 6403 } 6404 unsigned int after_count = 6405 GenCollectedHeap::heap()->total_collections(); 6406 bool abort = (_before_count != after_count) || 6407 _collector->should_abort_preclean(); 6408 return abort ? 0 : size; 6409 } 6410 6411 void SurvivorSpacePrecleanClosure::do_yield_work() { 6412 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6413 "CMS thread should hold CMS token"); 6414 assert_lock_strong(_bit_map->lock()); 6415 // Relinquish the bit map lock 6416 _bit_map->lock()->unlock(); 6417 ConcurrentMarkSweepThread::desynchronize(true); 6418 _collector->stopTimer(); 6419 if (PrintCMSStatistics != 0) { 6420 _collector->incrementYields(); 6421 } 6422 6423 // See the comment in coordinator_yield() 6424 for (unsigned i = 0; i < CMSYieldSleepCount && 6425 ConcurrentMarkSweepThread::should_yield() && 6426 !CMSCollector::foregroundGCIsActive(); ++i) { 6427 os::sleep(Thread::current(), 1, false); 6428 } 6429 6430 ConcurrentMarkSweepThread::synchronize(true); 6431 _bit_map->lock()->lock_without_safepoint_check(); 6432 _collector->startTimer(); 6433 } 6434 6435 // This closure is used to rescan the marked objects on the dirty cards 6436 // in the mod union table and the card table proper. In the parallel 6437 // case, although the bitMap is shared, we do a single read so the 6438 // isMarked() query is "safe". 6439 bool ScanMarkedObjectsAgainClosure::do_object_bm(oop p, MemRegion mr) { 6440 // Ignore mark word because we are running concurrent with mutators 6441 assert(p->is_oop_or_null(true), "Expected an oop or NULL at " PTR_FORMAT, p2i(p)); 6442 HeapWord* addr = (HeapWord*)p; 6443 assert(_span.contains(addr), "we are scanning the CMS generation"); 6444 bool is_obj_array = false; 6445 #ifdef ASSERT 6446 if (!_parallel) { 6447 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 6448 assert(_collector->overflow_list_is_empty(), 6449 "overflow list should be empty"); 6450 6451 } 6452 #endif // ASSERT 6453 if (_bit_map->isMarked(addr)) { 6454 // Obj arrays are precisely marked, non-arrays are not; 6455 // so we scan objArrays precisely and non-arrays in their 6456 // entirety. 6457 if (p->is_objArray()) { 6458 is_obj_array = true; 6459 if (_parallel) { 6460 p->oop_iterate(_par_scan_closure, mr); 6461 } else { 6462 p->oop_iterate(_scan_closure, mr); 6463 } 6464 } else { 6465 if (_parallel) { 6466 p->oop_iterate(_par_scan_closure); 6467 } else { 6468 p->oop_iterate(_scan_closure); 6469 } 6470 } 6471 } 6472 #ifdef ASSERT 6473 if (!_parallel) { 6474 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 6475 assert(_collector->overflow_list_is_empty(), 6476 "overflow list should be empty"); 6477 6478 } 6479 #endif // ASSERT 6480 return is_obj_array; 6481 } 6482 6483 MarkFromRootsClosure::MarkFromRootsClosure(CMSCollector* collector, 6484 MemRegion span, 6485 CMSBitMap* bitMap, CMSMarkStack* markStack, 6486 bool should_yield, bool verifying): 6487 _collector(collector), 6488 _span(span), 6489 _bitMap(bitMap), 6490 _mut(&collector->_modUnionTable), 6491 _markStack(markStack), 6492 _yield(should_yield), 6493 _skipBits(0) 6494 { 6495 assert(_markStack->isEmpty(), "stack should be empty"); 6496 _finger = _bitMap->startWord(); 6497 _threshold = _finger; 6498 assert(_collector->_restart_addr == NULL, "Sanity check"); 6499 assert(_span.contains(_finger), "Out of bounds _finger?"); 6500 DEBUG_ONLY(_verifying = verifying;) 6501 } 6502 6503 void MarkFromRootsClosure::reset(HeapWord* addr) { 6504 assert(_markStack->isEmpty(), "would cause duplicates on stack"); 6505 assert(_span.contains(addr), "Out of bounds _finger?"); 6506 _finger = addr; 6507 _threshold = (HeapWord*)round_to( 6508 (intptr_t)_finger, CardTableModRefBS::card_size); 6509 } 6510 6511 // Should revisit to see if this should be restructured for 6512 // greater efficiency. 6513 bool MarkFromRootsClosure::do_bit(size_t offset) { 6514 if (_skipBits > 0) { 6515 _skipBits--; 6516 return true; 6517 } 6518 // convert offset into a HeapWord* 6519 HeapWord* addr = _bitMap->startWord() + offset; 6520 assert(_bitMap->endWord() && addr < _bitMap->endWord(), 6521 "address out of range"); 6522 assert(_bitMap->isMarked(addr), "tautology"); 6523 if (_bitMap->isMarked(addr+1)) { 6524 // this is an allocated but not yet initialized object 6525 assert(_skipBits == 0, "tautology"); 6526 _skipBits = 2; // skip next two marked bits ("Printezis-marks") 6527 oop p = oop(addr); 6528 if (p->klass_or_null() == NULL) { 6529 DEBUG_ONLY(if (!_verifying) {) 6530 // We re-dirty the cards on which this object lies and increase 6531 // the _threshold so that we'll come back to scan this object 6532 // during the preclean or remark phase. (CMSCleanOnEnter) 6533 if (CMSCleanOnEnter) { 6534 size_t sz = _collector->block_size_using_printezis_bits(addr); 6535 HeapWord* end_card_addr = (HeapWord*)round_to( 6536 (intptr_t)(addr+sz), CardTableModRefBS::card_size); 6537 MemRegion redirty_range = MemRegion(addr, end_card_addr); 6538 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 6539 // Bump _threshold to end_card_addr; note that 6540 // _threshold cannot possibly exceed end_card_addr, anyhow. 6541 // This prevents future clearing of the card as the scan proceeds 6542 // to the right. 6543 assert(_threshold <= end_card_addr, 6544 "Because we are just scanning into this object"); 6545 if (_threshold < end_card_addr) { 6546 _threshold = end_card_addr; 6547 } 6548 if (p->klass_or_null() != NULL) { 6549 // Redirty the range of cards... 6550 _mut->mark_range(redirty_range); 6551 } // ...else the setting of klass will dirty the card anyway. 6552 } 6553 DEBUG_ONLY(}) 6554 return true; 6555 } 6556 } 6557 scanOopsInOop(addr); 6558 return true; 6559 } 6560 6561 // We take a break if we've been at this for a while, 6562 // so as to avoid monopolizing the locks involved. 6563 void MarkFromRootsClosure::do_yield_work() { 6564 // First give up the locks, then yield, then re-lock 6565 // We should probably use a constructor/destructor idiom to 6566 // do this unlock/lock or modify the MutexUnlocker class to 6567 // serve our purpose. XXX 6568 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6569 "CMS thread should hold CMS token"); 6570 assert_lock_strong(_bitMap->lock()); 6571 _bitMap->lock()->unlock(); 6572 ConcurrentMarkSweepThread::desynchronize(true); 6573 _collector->stopTimer(); 6574 if (PrintCMSStatistics != 0) { 6575 _collector->incrementYields(); 6576 } 6577 6578 // See the comment in coordinator_yield() 6579 for (unsigned i = 0; i < CMSYieldSleepCount && 6580 ConcurrentMarkSweepThread::should_yield() && 6581 !CMSCollector::foregroundGCIsActive(); ++i) { 6582 os::sleep(Thread::current(), 1, false); 6583 } 6584 6585 ConcurrentMarkSweepThread::synchronize(true); 6586 _bitMap->lock()->lock_without_safepoint_check(); 6587 _collector->startTimer(); 6588 } 6589 6590 void MarkFromRootsClosure::scanOopsInOop(HeapWord* ptr) { 6591 assert(_bitMap->isMarked(ptr), "expected bit to be set"); 6592 assert(_markStack->isEmpty(), 6593 "should drain stack to limit stack usage"); 6594 // convert ptr to an oop preparatory to scanning 6595 oop obj = oop(ptr); 6596 // Ignore mark word in verification below, since we 6597 // may be running concurrent with mutators. 6598 assert(obj->is_oop(true), "should be an oop"); 6599 assert(_finger <= ptr, "_finger runneth ahead"); 6600 // advance the finger to right end of this object 6601 _finger = ptr + obj->size(); 6602 assert(_finger > ptr, "we just incremented it above"); 6603 // On large heaps, it may take us some time to get through 6604 // the marking phase. During 6605 // this time it's possible that a lot of mutations have 6606 // accumulated in the card table and the mod union table -- 6607 // these mutation records are redundant until we have 6608 // actually traced into the corresponding card. 6609 // Here, we check whether advancing the finger would make 6610 // us cross into a new card, and if so clear corresponding 6611 // cards in the MUT (preclean them in the card-table in the 6612 // future). 6613 6614 DEBUG_ONLY(if (!_verifying) {) 6615 // The clean-on-enter optimization is disabled by default, 6616 // until we fix 6178663. 6617 if (CMSCleanOnEnter && (_finger > _threshold)) { 6618 // [_threshold, _finger) represents the interval 6619 // of cards to be cleared in MUT (or precleaned in card table). 6620 // The set of cards to be cleared is all those that overlap 6621 // with the interval [_threshold, _finger); note that 6622 // _threshold is always kept card-aligned but _finger isn't 6623 // always card-aligned. 6624 HeapWord* old_threshold = _threshold; 6625 assert(old_threshold == (HeapWord*)round_to( 6626 (intptr_t)old_threshold, CardTableModRefBS::card_size), 6627 "_threshold should always be card-aligned"); 6628 _threshold = (HeapWord*)round_to( 6629 (intptr_t)_finger, CardTableModRefBS::card_size); 6630 MemRegion mr(old_threshold, _threshold); 6631 assert(!mr.is_empty(), "Control point invariant"); 6632 assert(_span.contains(mr), "Should clear within span"); 6633 _mut->clear_range(mr); 6634 } 6635 DEBUG_ONLY(}) 6636 // Note: the finger doesn't advance while we drain 6637 // the stack below. 6638 PushOrMarkClosure pushOrMarkClosure(_collector, 6639 _span, _bitMap, _markStack, 6640 _finger, this); 6641 bool res = _markStack->push(obj); 6642 assert(res, "Empty non-zero size stack should have space for single push"); 6643 while (!_markStack->isEmpty()) { 6644 oop new_oop = _markStack->pop(); 6645 // Skip verifying header mark word below because we are 6646 // running concurrent with mutators. 6647 assert(new_oop->is_oop(true), "Oops! expected to pop an oop"); 6648 // now scan this oop's oops 6649 new_oop->oop_iterate(&pushOrMarkClosure); 6650 do_yield_check(); 6651 } 6652 assert(_markStack->isEmpty(), "tautology, emphasizing post-condition"); 6653 } 6654 6655 Par_MarkFromRootsClosure::Par_MarkFromRootsClosure(CMSConcMarkingTask* task, 6656 CMSCollector* collector, MemRegion span, 6657 CMSBitMap* bit_map, 6658 OopTaskQueue* work_queue, 6659 CMSMarkStack* overflow_stack): 6660 _collector(collector), 6661 _whole_span(collector->_span), 6662 _span(span), 6663 _bit_map(bit_map), 6664 _mut(&collector->_modUnionTable), 6665 _work_queue(work_queue), 6666 _overflow_stack(overflow_stack), 6667 _skip_bits(0), 6668 _task(task) 6669 { 6670 assert(_work_queue->size() == 0, "work_queue should be empty"); 6671 _finger = span.start(); 6672 _threshold = _finger; // XXX Defer clear-on-enter optimization for now 6673 assert(_span.contains(_finger), "Out of bounds _finger?"); 6674 } 6675 6676 // Should revisit to see if this should be restructured for 6677 // greater efficiency. 6678 bool Par_MarkFromRootsClosure::do_bit(size_t offset) { 6679 if (_skip_bits > 0) { 6680 _skip_bits--; 6681 return true; 6682 } 6683 // convert offset into a HeapWord* 6684 HeapWord* addr = _bit_map->startWord() + offset; 6685 assert(_bit_map->endWord() && addr < _bit_map->endWord(), 6686 "address out of range"); 6687 assert(_bit_map->isMarked(addr), "tautology"); 6688 if (_bit_map->isMarked(addr+1)) { 6689 // this is an allocated object that might not yet be initialized 6690 assert(_skip_bits == 0, "tautology"); 6691 _skip_bits = 2; // skip next two marked bits ("Printezis-marks") 6692 oop p = oop(addr); 6693 if (p->klass_or_null() == NULL) { 6694 // in the case of Clean-on-Enter optimization, redirty card 6695 // and avoid clearing card by increasing the threshold. 6696 return true; 6697 } 6698 } 6699 scan_oops_in_oop(addr); 6700 return true; 6701 } 6702 6703 void Par_MarkFromRootsClosure::scan_oops_in_oop(HeapWord* ptr) { 6704 assert(_bit_map->isMarked(ptr), "expected bit to be set"); 6705 // Should we assert that our work queue is empty or 6706 // below some drain limit? 6707 assert(_work_queue->size() == 0, 6708 "should drain stack to limit stack usage"); 6709 // convert ptr to an oop preparatory to scanning 6710 oop obj = oop(ptr); 6711 // Ignore mark word in verification below, since we 6712 // may be running concurrent with mutators. 6713 assert(obj->is_oop(true), "should be an oop"); 6714 assert(_finger <= ptr, "_finger runneth ahead"); 6715 // advance the finger to right end of this object 6716 _finger = ptr + obj->size(); 6717 assert(_finger > ptr, "we just incremented it above"); 6718 // On large heaps, it may take us some time to get through 6719 // the marking phase. During 6720 // this time it's possible that a lot of mutations have 6721 // accumulated in the card table and the mod union table -- 6722 // these mutation records are redundant until we have 6723 // actually traced into the corresponding card. 6724 // Here, we check whether advancing the finger would make 6725 // us cross into a new card, and if so clear corresponding 6726 // cards in the MUT (preclean them in the card-table in the 6727 // future). 6728 6729 // The clean-on-enter optimization is disabled by default, 6730 // until we fix 6178663. 6731 if (CMSCleanOnEnter && (_finger > _threshold)) { 6732 // [_threshold, _finger) represents the interval 6733 // of cards to be cleared in MUT (or precleaned in card table). 6734 // The set of cards to be cleared is all those that overlap 6735 // with the interval [_threshold, _finger); note that 6736 // _threshold is always kept card-aligned but _finger isn't 6737 // always card-aligned. 6738 HeapWord* old_threshold = _threshold; 6739 assert(old_threshold == (HeapWord*)round_to( 6740 (intptr_t)old_threshold, CardTableModRefBS::card_size), 6741 "_threshold should always be card-aligned"); 6742 _threshold = (HeapWord*)round_to( 6743 (intptr_t)_finger, CardTableModRefBS::card_size); 6744 MemRegion mr(old_threshold, _threshold); 6745 assert(!mr.is_empty(), "Control point invariant"); 6746 assert(_span.contains(mr), "Should clear within span"); // _whole_span ?? 6747 _mut->clear_range(mr); 6748 } 6749 6750 // Note: the local finger doesn't advance while we drain 6751 // the stack below, but the global finger sure can and will. 6752 HeapWord** gfa = _task->global_finger_addr(); 6753 Par_PushOrMarkClosure pushOrMarkClosure(_collector, 6754 _span, _bit_map, 6755 _work_queue, 6756 _overflow_stack, 6757 _finger, 6758 gfa, this); 6759 bool res = _work_queue->push(obj); // overflow could occur here 6760 assert(res, "Will hold once we use workqueues"); 6761 while (true) { 6762 oop new_oop; 6763 if (!_work_queue->pop_local(new_oop)) { 6764 // We emptied our work_queue; check if there's stuff that can 6765 // be gotten from the overflow stack. 6766 if (CMSConcMarkingTask::get_work_from_overflow_stack( 6767 _overflow_stack, _work_queue)) { 6768 do_yield_check(); 6769 continue; 6770 } else { // done 6771 break; 6772 } 6773 } 6774 // Skip verifying header mark word below because we are 6775 // running concurrent with mutators. 6776 assert(new_oop->is_oop(true), "Oops! expected to pop an oop"); 6777 // now scan this oop's oops 6778 new_oop->oop_iterate(&pushOrMarkClosure); 6779 do_yield_check(); 6780 } 6781 assert(_work_queue->size() == 0, "tautology, emphasizing post-condition"); 6782 } 6783 6784 // Yield in response to a request from VM Thread or 6785 // from mutators. 6786 void Par_MarkFromRootsClosure::do_yield_work() { 6787 assert(_task != NULL, "sanity"); 6788 _task->yield(); 6789 } 6790 6791 // A variant of the above used for verifying CMS marking work. 6792 MarkFromRootsVerifyClosure::MarkFromRootsVerifyClosure(CMSCollector* collector, 6793 MemRegion span, 6794 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 6795 CMSMarkStack* mark_stack): 6796 _collector(collector), 6797 _span(span), 6798 _verification_bm(verification_bm), 6799 _cms_bm(cms_bm), 6800 _mark_stack(mark_stack), 6801 _pam_verify_closure(collector, span, verification_bm, cms_bm, 6802 mark_stack) 6803 { 6804 assert(_mark_stack->isEmpty(), "stack should be empty"); 6805 _finger = _verification_bm->startWord(); 6806 assert(_collector->_restart_addr == NULL, "Sanity check"); 6807 assert(_span.contains(_finger), "Out of bounds _finger?"); 6808 } 6809 6810 void MarkFromRootsVerifyClosure::reset(HeapWord* addr) { 6811 assert(_mark_stack->isEmpty(), "would cause duplicates on stack"); 6812 assert(_span.contains(addr), "Out of bounds _finger?"); 6813 _finger = addr; 6814 } 6815 6816 // Should revisit to see if this should be restructured for 6817 // greater efficiency. 6818 bool MarkFromRootsVerifyClosure::do_bit(size_t offset) { 6819 // convert offset into a HeapWord* 6820 HeapWord* addr = _verification_bm->startWord() + offset; 6821 assert(_verification_bm->endWord() && addr < _verification_bm->endWord(), 6822 "address out of range"); 6823 assert(_verification_bm->isMarked(addr), "tautology"); 6824 assert(_cms_bm->isMarked(addr), "tautology"); 6825 6826 assert(_mark_stack->isEmpty(), 6827 "should drain stack to limit stack usage"); 6828 // convert addr to an oop preparatory to scanning 6829 oop obj = oop(addr); 6830 assert(obj->is_oop(), "should be an oop"); 6831 assert(_finger <= addr, "_finger runneth ahead"); 6832 // advance the finger to right end of this object 6833 _finger = addr + obj->size(); 6834 assert(_finger > addr, "we just incremented it above"); 6835 // Note: the finger doesn't advance while we drain 6836 // the stack below. 6837 bool res = _mark_stack->push(obj); 6838 assert(res, "Empty non-zero size stack should have space for single push"); 6839 while (!_mark_stack->isEmpty()) { 6840 oop new_oop = _mark_stack->pop(); 6841 assert(new_oop->is_oop(), "Oops! expected to pop an oop"); 6842 // now scan this oop's oops 6843 new_oop->oop_iterate(&_pam_verify_closure); 6844 } 6845 assert(_mark_stack->isEmpty(), "tautology, emphasizing post-condition"); 6846 return true; 6847 } 6848 6849 PushAndMarkVerifyClosure::PushAndMarkVerifyClosure( 6850 CMSCollector* collector, MemRegion span, 6851 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 6852 CMSMarkStack* mark_stack): 6853 MetadataAwareOopClosure(collector->ref_processor()), 6854 _collector(collector), 6855 _span(span), 6856 _verification_bm(verification_bm), 6857 _cms_bm(cms_bm), 6858 _mark_stack(mark_stack) 6859 { } 6860 6861 void PushAndMarkVerifyClosure::do_oop(oop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 6862 void PushAndMarkVerifyClosure::do_oop(narrowOop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 6863 6864 // Upon stack overflow, we discard (part of) the stack, 6865 // remembering the least address amongst those discarded 6866 // in CMSCollector's _restart_address. 6867 void PushAndMarkVerifyClosure::handle_stack_overflow(HeapWord* lost) { 6868 // Remember the least grey address discarded 6869 HeapWord* ra = (HeapWord*)_mark_stack->least_value(lost); 6870 _collector->lower_restart_addr(ra); 6871 _mark_stack->reset(); // discard stack contents 6872 _mark_stack->expand(); // expand the stack if possible 6873 } 6874 6875 void PushAndMarkVerifyClosure::do_oop(oop obj) { 6876 assert(obj->is_oop_or_null(), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6877 HeapWord* addr = (HeapWord*)obj; 6878 if (_span.contains(addr) && !_verification_bm->isMarked(addr)) { 6879 // Oop lies in _span and isn't yet grey or black 6880 _verification_bm->mark(addr); // now grey 6881 if (!_cms_bm->isMarked(addr)) { 6882 oop(addr)->print(); 6883 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)", 6884 p2i(addr)); 6885 fatal("... aborting"); 6886 } 6887 6888 if (!_mark_stack->push(obj)) { // stack overflow 6889 if (PrintCMSStatistics != 0) { 6890 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 6891 SIZE_FORMAT, _mark_stack->capacity()); 6892 } 6893 assert(_mark_stack->isFull(), "Else push should have succeeded"); 6894 handle_stack_overflow(addr); 6895 } 6896 // anything including and to the right of _finger 6897 // will be scanned as we iterate over the remainder of the 6898 // bit map 6899 } 6900 } 6901 6902 PushOrMarkClosure::PushOrMarkClosure(CMSCollector* collector, 6903 MemRegion span, 6904 CMSBitMap* bitMap, CMSMarkStack* markStack, 6905 HeapWord* finger, MarkFromRootsClosure* parent) : 6906 MetadataAwareOopClosure(collector->ref_processor()), 6907 _collector(collector), 6908 _span(span), 6909 _bitMap(bitMap), 6910 _markStack(markStack), 6911 _finger(finger), 6912 _parent(parent) 6913 { } 6914 6915 Par_PushOrMarkClosure::Par_PushOrMarkClosure(CMSCollector* collector, 6916 MemRegion span, 6917 CMSBitMap* bit_map, 6918 OopTaskQueue* work_queue, 6919 CMSMarkStack* overflow_stack, 6920 HeapWord* finger, 6921 HeapWord** global_finger_addr, 6922 Par_MarkFromRootsClosure* parent) : 6923 MetadataAwareOopClosure(collector->ref_processor()), 6924 _collector(collector), 6925 _whole_span(collector->_span), 6926 _span(span), 6927 _bit_map(bit_map), 6928 _work_queue(work_queue), 6929 _overflow_stack(overflow_stack), 6930 _finger(finger), 6931 _global_finger_addr(global_finger_addr), 6932 _parent(parent) 6933 { } 6934 6935 // Assumes thread-safe access by callers, who are 6936 // responsible for mutual exclusion. 6937 void CMSCollector::lower_restart_addr(HeapWord* low) { 6938 assert(_span.contains(low), "Out of bounds addr"); 6939 if (_restart_addr == NULL) { 6940 _restart_addr = low; 6941 } else { 6942 _restart_addr = MIN2(_restart_addr, low); 6943 } 6944 } 6945 6946 // Upon stack overflow, we discard (part of) the stack, 6947 // remembering the least address amongst those discarded 6948 // in CMSCollector's _restart_address. 6949 void PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 6950 // Remember the least grey address discarded 6951 HeapWord* ra = (HeapWord*)_markStack->least_value(lost); 6952 _collector->lower_restart_addr(ra); 6953 _markStack->reset(); // discard stack contents 6954 _markStack->expand(); // expand the stack if possible 6955 } 6956 6957 // Upon stack overflow, we discard (part of) the stack, 6958 // remembering the least address amongst those discarded 6959 // in CMSCollector's _restart_address. 6960 void Par_PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 6961 // We need to do this under a mutex to prevent other 6962 // workers from interfering with the work done below. 6963 MutexLockerEx ml(_overflow_stack->par_lock(), 6964 Mutex::_no_safepoint_check_flag); 6965 // Remember the least grey address discarded 6966 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 6967 _collector->lower_restart_addr(ra); 6968 _overflow_stack->reset(); // discard stack contents 6969 _overflow_stack->expand(); // expand the stack if possible 6970 } 6971 6972 void PushOrMarkClosure::do_oop(oop obj) { 6973 // Ignore mark word because we are running concurrent with mutators. 6974 assert(obj->is_oop_or_null(true), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6975 HeapWord* addr = (HeapWord*)obj; 6976 if (_span.contains(addr) && !_bitMap->isMarked(addr)) { 6977 // Oop lies in _span and isn't yet grey or black 6978 _bitMap->mark(addr); // now grey 6979 if (addr < _finger) { 6980 // the bit map iteration has already either passed, or 6981 // sampled, this bit in the bit map; we'll need to 6982 // use the marking stack to scan this oop's oops. 6983 bool simulate_overflow = false; 6984 NOT_PRODUCT( 6985 if (CMSMarkStackOverflowALot && 6986 _collector->simulate_overflow()) { 6987 // simulate a stack overflow 6988 simulate_overflow = true; 6989 } 6990 ) 6991 if (simulate_overflow || !_markStack->push(obj)) { // stack overflow 6992 if (PrintCMSStatistics != 0) { 6993 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 6994 SIZE_FORMAT, _markStack->capacity()); 6995 } 6996 assert(simulate_overflow || _markStack->isFull(), "Else push should have succeeded"); 6997 handle_stack_overflow(addr); 6998 } 6999 } 7000 // anything including and to the right of _finger 7001 // will be scanned as we iterate over the remainder of the 7002 // bit map 7003 do_yield_check(); 7004 } 7005 } 7006 7007 void PushOrMarkClosure::do_oop(oop* p) { PushOrMarkClosure::do_oop_work(p); } 7008 void PushOrMarkClosure::do_oop(narrowOop* p) { PushOrMarkClosure::do_oop_work(p); } 7009 7010 void Par_PushOrMarkClosure::do_oop(oop obj) { 7011 // Ignore mark word because we are running concurrent with mutators. 7012 assert(obj->is_oop_or_null(true), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 7013 HeapWord* addr = (HeapWord*)obj; 7014 if (_whole_span.contains(addr) && !_bit_map->isMarked(addr)) { 7015 // Oop lies in _span and isn't yet grey or black 7016 // We read the global_finger (volatile read) strictly after marking oop 7017 bool res = _bit_map->par_mark(addr); // now grey 7018 volatile HeapWord** gfa = (volatile HeapWord**)_global_finger_addr; 7019 // Should we push this marked oop on our stack? 7020 // -- if someone else marked it, nothing to do 7021 // -- if target oop is above global finger nothing to do 7022 // -- if target oop is in chunk and above local finger 7023 // then nothing to do 7024 // -- else push on work queue 7025 if ( !res // someone else marked it, they will deal with it 7026 || (addr >= *gfa) // will be scanned in a later task 7027 || (_span.contains(addr) && addr >= _finger)) { // later in this chunk 7028 return; 7029 } 7030 // the bit map iteration has already either passed, or 7031 // sampled, this bit in the bit map; we'll need to 7032 // use the marking stack to scan this oop's oops. 7033 bool simulate_overflow = false; 7034 NOT_PRODUCT( 7035 if (CMSMarkStackOverflowALot && 7036 _collector->simulate_overflow()) { 7037 // simulate a stack overflow 7038 simulate_overflow = true; 7039 } 7040 ) 7041 if (simulate_overflow || 7042 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) { 7043 // stack overflow 7044 if (PrintCMSStatistics != 0) { 7045 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 7046 SIZE_FORMAT, _overflow_stack->capacity()); 7047 } 7048 // We cannot assert that the overflow stack is full because 7049 // it may have been emptied since. 7050 assert(simulate_overflow || 7051 _work_queue->size() == _work_queue->max_elems(), 7052 "Else push should have succeeded"); 7053 handle_stack_overflow(addr); 7054 } 7055 do_yield_check(); 7056 } 7057 } 7058 7059 void Par_PushOrMarkClosure::do_oop(oop* p) { Par_PushOrMarkClosure::do_oop_work(p); } 7060 void Par_PushOrMarkClosure::do_oop(narrowOop* p) { Par_PushOrMarkClosure::do_oop_work(p); } 7061 7062 PushAndMarkClosure::PushAndMarkClosure(CMSCollector* collector, 7063 MemRegion span, 7064 ReferenceProcessor* rp, 7065 CMSBitMap* bit_map, 7066 CMSBitMap* mod_union_table, 7067 CMSMarkStack* mark_stack, 7068 bool concurrent_precleaning): 7069 MetadataAwareOopClosure(rp), 7070 _collector(collector), 7071 _span(span), 7072 _bit_map(bit_map), 7073 _mod_union_table(mod_union_table), 7074 _mark_stack(mark_stack), 7075 _concurrent_precleaning(concurrent_precleaning) 7076 { 7077 assert(ref_processor() != NULL, "ref_processor shouldn't be NULL"); 7078 } 7079 7080 // Grey object rescan during pre-cleaning and second checkpoint phases -- 7081 // the non-parallel version (the parallel version appears further below.) 7082 void PushAndMarkClosure::do_oop(oop obj) { 7083 // Ignore mark word verification. If during concurrent precleaning, 7084 // the object monitor may be locked. If during the checkpoint 7085 // phases, the object may already have been reached by a different 7086 // path and may be at the end of the global overflow list (so 7087 // the mark word may be NULL). 7088 assert(obj->is_oop_or_null(true /* ignore mark word */), 7089 "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 7090 HeapWord* addr = (HeapWord*)obj; 7091 // Check if oop points into the CMS generation 7092 // and is not marked 7093 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 7094 // a white object ... 7095 _bit_map->mark(addr); // ... now grey 7096 // push on the marking stack (grey set) 7097 bool simulate_overflow = false; 7098 NOT_PRODUCT( 7099 if (CMSMarkStackOverflowALot && 7100 _collector->simulate_overflow()) { 7101 // simulate a stack overflow 7102 simulate_overflow = true; 7103 } 7104 ) 7105 if (simulate_overflow || !_mark_stack->push(obj)) { 7106 if (_concurrent_precleaning) { 7107 // During precleaning we can just dirty the appropriate card(s) 7108 // in the mod union table, thus ensuring that the object remains 7109 // in the grey set and continue. In the case of object arrays 7110 // we need to dirty all of the cards that the object spans, 7111 // since the rescan of object arrays will be limited to the 7112 // dirty cards. 7113 // Note that no one can be interfering with us in this action 7114 // of dirtying the mod union table, so no locking or atomics 7115 // are required. 7116 if (obj->is_objArray()) { 7117 size_t sz = obj->size(); 7118 HeapWord* end_card_addr = (HeapWord*)round_to( 7119 (intptr_t)(addr+sz), CardTableModRefBS::card_size); 7120 MemRegion redirty_range = MemRegion(addr, end_card_addr); 7121 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 7122 _mod_union_table->mark_range(redirty_range); 7123 } else { 7124 _mod_union_table->mark(addr); 7125 } 7126 _collector->_ser_pmc_preclean_ovflw++; 7127 } else { 7128 // During the remark phase, we need to remember this oop 7129 // in the overflow list. 7130 _collector->push_on_overflow_list(obj); 7131 _collector->_ser_pmc_remark_ovflw++; 7132 } 7133 } 7134 } 7135 } 7136 7137 Par_PushAndMarkClosure::Par_PushAndMarkClosure(CMSCollector* collector, 7138 MemRegion span, 7139 ReferenceProcessor* rp, 7140 CMSBitMap* bit_map, 7141 OopTaskQueue* work_queue): 7142 MetadataAwareOopClosure(rp), 7143 _collector(collector), 7144 _span(span), 7145 _bit_map(bit_map), 7146 _work_queue(work_queue) 7147 { 7148 assert(ref_processor() != NULL, "ref_processor shouldn't be NULL"); 7149 } 7150 7151 void PushAndMarkClosure::do_oop(oop* p) { PushAndMarkClosure::do_oop_work(p); } 7152 void PushAndMarkClosure::do_oop(narrowOop* p) { PushAndMarkClosure::do_oop_work(p); } 7153 7154 // Grey object rescan during second checkpoint phase -- 7155 // the parallel version. 7156 void Par_PushAndMarkClosure::do_oop(oop obj) { 7157 // In the assert below, we ignore the mark word because 7158 // this oop may point to an already visited object that is 7159 // on the overflow stack (in which case the mark word has 7160 // been hijacked for chaining into the overflow stack -- 7161 // if this is the last object in the overflow stack then 7162 // its mark word will be NULL). Because this object may 7163 // have been subsequently popped off the global overflow 7164 // stack, and the mark word possibly restored to the prototypical 7165 // value, by the time we get to examined this failing assert in 7166 // the debugger, is_oop_or_null(false) may subsequently start 7167 // to hold. 7168 assert(obj->is_oop_or_null(true), 7169 "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 7170 HeapWord* addr = (HeapWord*)obj; 7171 // Check if oop points into the CMS generation 7172 // and is not marked 7173 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 7174 // a white object ... 7175 // If we manage to "claim" the object, by being the 7176 // first thread to mark it, then we push it on our 7177 // marking stack 7178 if (_bit_map->par_mark(addr)) { // ... now grey 7179 // push on work queue (grey set) 7180 bool simulate_overflow = false; 7181 NOT_PRODUCT( 7182 if (CMSMarkStackOverflowALot && 7183 _collector->par_simulate_overflow()) { 7184 // simulate a stack overflow 7185 simulate_overflow = true; 7186 } 7187 ) 7188 if (simulate_overflow || !_work_queue->push(obj)) { 7189 _collector->par_push_on_overflow_list(obj); 7190 _collector->_par_pmc_remark_ovflw++; // imprecise OK: no need to CAS 7191 } 7192 } // Else, some other thread got there first 7193 } 7194 } 7195 7196 void Par_PushAndMarkClosure::do_oop(oop* p) { Par_PushAndMarkClosure::do_oop_work(p); } 7197 void Par_PushAndMarkClosure::do_oop(narrowOop* p) { Par_PushAndMarkClosure::do_oop_work(p); } 7198 7199 void CMSPrecleanRefsYieldClosure::do_yield_work() { 7200 Mutex* bml = _collector->bitMapLock(); 7201 assert_lock_strong(bml); 7202 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 7203 "CMS thread should hold CMS token"); 7204 7205 bml->unlock(); 7206 ConcurrentMarkSweepThread::desynchronize(true); 7207 7208 _collector->stopTimer(); 7209 if (PrintCMSStatistics != 0) { 7210 _collector->incrementYields(); 7211 } 7212 7213 // See the comment in coordinator_yield() 7214 for (unsigned i = 0; i < CMSYieldSleepCount && 7215 ConcurrentMarkSweepThread::should_yield() && 7216 !CMSCollector::foregroundGCIsActive(); ++i) { 7217 os::sleep(Thread::current(), 1, false); 7218 } 7219 7220 ConcurrentMarkSweepThread::synchronize(true); 7221 bml->lock(); 7222 7223 _collector->startTimer(); 7224 } 7225 7226 bool CMSPrecleanRefsYieldClosure::should_return() { 7227 if (ConcurrentMarkSweepThread::should_yield()) { 7228 do_yield_work(); 7229 } 7230 return _collector->foregroundGCIsActive(); 7231 } 7232 7233 void MarkFromDirtyCardsClosure::do_MemRegion(MemRegion mr) { 7234 assert(((size_t)mr.start())%CardTableModRefBS::card_size_in_words == 0, 7235 "mr should be aligned to start at a card boundary"); 7236 // We'd like to assert: 7237 // assert(mr.word_size()%CardTableModRefBS::card_size_in_words == 0, 7238 // "mr should be a range of cards"); 7239 // However, that would be too strong in one case -- the last 7240 // partition ends at _unallocated_block which, in general, can be 7241 // an arbitrary boundary, not necessarily card aligned. 7242 if (PrintCMSStatistics != 0) { 7243 _num_dirty_cards += 7244 mr.word_size()/CardTableModRefBS::card_size_in_words; 7245 } 7246 _space->object_iterate_mem(mr, &_scan_cl); 7247 } 7248 7249 SweepClosure::SweepClosure(CMSCollector* collector, 7250 ConcurrentMarkSweepGeneration* g, 7251 CMSBitMap* bitMap, bool should_yield) : 7252 _collector(collector), 7253 _g(g), 7254 _sp(g->cmsSpace()), 7255 _limit(_sp->sweep_limit()), 7256 _freelistLock(_sp->freelistLock()), 7257 _bitMap(bitMap), 7258 _yield(should_yield), 7259 _inFreeRange(false), // No free range at beginning of sweep 7260 _freeRangeInFreeLists(false), // No free range at beginning of sweep 7261 _lastFreeRangeCoalesced(false), 7262 _freeFinger(g->used_region().start()) 7263 { 7264 NOT_PRODUCT( 7265 _numObjectsFreed = 0; 7266 _numWordsFreed = 0; 7267 _numObjectsLive = 0; 7268 _numWordsLive = 0; 7269 _numObjectsAlreadyFree = 0; 7270 _numWordsAlreadyFree = 0; 7271 _last_fc = NULL; 7272 7273 _sp->initializeIndexedFreeListArrayReturnedBytes(); 7274 _sp->dictionary()->initialize_dict_returned_bytes(); 7275 ) 7276 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7277 "sweep _limit out of bounds"); 7278 if (CMSTraceSweeper) { 7279 gclog_or_tty->print_cr("\n====================\nStarting new sweep with limit " PTR_FORMAT, 7280 p2i(_limit)); 7281 } 7282 } 7283 7284 void SweepClosure::print_on(outputStream* st) const { 7285 tty->print_cr("_sp = [" PTR_FORMAT "," PTR_FORMAT ")", 7286 p2i(_sp->bottom()), p2i(_sp->end())); 7287 tty->print_cr("_limit = " PTR_FORMAT, p2i(_limit)); 7288 tty->print_cr("_freeFinger = " PTR_FORMAT, p2i(_freeFinger)); 7289 NOT_PRODUCT(tty->print_cr("_last_fc = " PTR_FORMAT, p2i(_last_fc));) 7290 tty->print_cr("_inFreeRange = %d, _freeRangeInFreeLists = %d, _lastFreeRangeCoalesced = %d", 7291 _inFreeRange, _freeRangeInFreeLists, _lastFreeRangeCoalesced); 7292 } 7293 7294 #ifndef PRODUCT 7295 // Assertion checking only: no useful work in product mode -- 7296 // however, if any of the flags below become product flags, 7297 // you may need to review this code to see if it needs to be 7298 // enabled in product mode. 7299 SweepClosure::~SweepClosure() { 7300 assert_lock_strong(_freelistLock); 7301 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7302 "sweep _limit out of bounds"); 7303 if (inFreeRange()) { 7304 warning("inFreeRange() should have been reset; dumping state of SweepClosure"); 7305 print(); 7306 ShouldNotReachHere(); 7307 } 7308 if (Verbose && PrintGC) { 7309 gclog_or_tty->print("Collected " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes", 7310 _numObjectsFreed, _numWordsFreed*sizeof(HeapWord)); 7311 gclog_or_tty->print_cr("\nLive " SIZE_FORMAT " objects, " 7312 SIZE_FORMAT " bytes " 7313 "Already free " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes", 7314 _numObjectsLive, _numWordsLive*sizeof(HeapWord), 7315 _numObjectsAlreadyFree, _numWordsAlreadyFree*sizeof(HeapWord)); 7316 size_t totalBytes = (_numWordsFreed + _numWordsLive + _numWordsAlreadyFree) 7317 * sizeof(HeapWord); 7318 gclog_or_tty->print_cr("Total sweep: " SIZE_FORMAT " bytes", totalBytes); 7319 7320 if (PrintCMSStatistics && CMSVerifyReturnedBytes) { 7321 size_t indexListReturnedBytes = _sp->sumIndexedFreeListArrayReturnedBytes(); 7322 size_t dict_returned_bytes = _sp->dictionary()->sum_dict_returned_bytes(); 7323 size_t returned_bytes = indexListReturnedBytes + dict_returned_bytes; 7324 gclog_or_tty->print("Returned " SIZE_FORMAT " bytes", returned_bytes); 7325 gclog_or_tty->print(" Indexed List Returned " SIZE_FORMAT " bytes", 7326 indexListReturnedBytes); 7327 gclog_or_tty->print_cr(" Dictionary Returned " SIZE_FORMAT " bytes", 7328 dict_returned_bytes); 7329 } 7330 } 7331 if (CMSTraceSweeper) { 7332 gclog_or_tty->print_cr("end of sweep with _limit = " PTR_FORMAT "\n================", 7333 p2i(_limit)); 7334 } 7335 } 7336 #endif // PRODUCT 7337 7338 void SweepClosure::initialize_free_range(HeapWord* freeFinger, 7339 bool freeRangeInFreeLists) { 7340 if (CMSTraceSweeper) { 7341 gclog_or_tty->print("---- Start free range at " PTR_FORMAT " with free block (%d)\n", 7342 p2i(freeFinger), freeRangeInFreeLists); 7343 } 7344 assert(!inFreeRange(), "Trampling existing free range"); 7345 set_inFreeRange(true); 7346 set_lastFreeRangeCoalesced(false); 7347 7348 set_freeFinger(freeFinger); 7349 set_freeRangeInFreeLists(freeRangeInFreeLists); 7350 if (CMSTestInFreeList) { 7351 if (freeRangeInFreeLists) { 7352 FreeChunk* fc = (FreeChunk*) freeFinger; 7353 assert(fc->is_free(), "A chunk on the free list should be free."); 7354 assert(fc->size() > 0, "Free range should have a size"); 7355 assert(_sp->verify_chunk_in_free_list(fc), "Chunk is not in free lists"); 7356 } 7357 } 7358 } 7359 7360 // Note that the sweeper runs concurrently with mutators. Thus, 7361 // it is possible for direct allocation in this generation to happen 7362 // in the middle of the sweep. Note that the sweeper also coalesces 7363 // contiguous free blocks. Thus, unless the sweeper and the allocator 7364 // synchronize appropriately freshly allocated blocks may get swept up. 7365 // This is accomplished by the sweeper locking the free lists while 7366 // it is sweeping. Thus blocks that are determined to be free are 7367 // indeed free. There is however one additional complication: 7368 // blocks that have been allocated since the final checkpoint and 7369 // mark, will not have been marked and so would be treated as 7370 // unreachable and swept up. To prevent this, the allocator marks 7371 // the bit map when allocating during the sweep phase. This leads, 7372 // however, to a further complication -- objects may have been allocated 7373 // but not yet initialized -- in the sense that the header isn't yet 7374 // installed. The sweeper can not then determine the size of the block 7375 // in order to skip over it. To deal with this case, we use a technique 7376 // (due to Printezis) to encode such uninitialized block sizes in the 7377 // bit map. Since the bit map uses a bit per every HeapWord, but the 7378 // CMS generation has a minimum object size of 3 HeapWords, it follows 7379 // that "normal marks" won't be adjacent in the bit map (there will 7380 // always be at least two 0 bits between successive 1 bits). We make use 7381 // of these "unused" bits to represent uninitialized blocks -- the bit 7382 // corresponding to the start of the uninitialized object and the next 7383 // bit are both set. Finally, a 1 bit marks the end of the object that 7384 // started with the two consecutive 1 bits to indicate its potentially 7385 // uninitialized state. 7386 7387 size_t SweepClosure::do_blk_careful(HeapWord* addr) { 7388 FreeChunk* fc = (FreeChunk*)addr; 7389 size_t res; 7390 7391 // Check if we are done sweeping. Below we check "addr >= _limit" rather 7392 // than "addr == _limit" because although _limit was a block boundary when 7393 // we started the sweep, it may no longer be one because heap expansion 7394 // may have caused us to coalesce the block ending at the address _limit 7395 // with a newly expanded chunk (this happens when _limit was set to the 7396 // previous _end of the space), so we may have stepped past _limit: 7397 // see the following Zeno-like trail of CRs 6977970, 7008136, 7042740. 7398 if (addr >= _limit) { // we have swept up to or past the limit: finish up 7399 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7400 "sweep _limit out of bounds"); 7401 assert(addr < _sp->end(), "addr out of bounds"); 7402 // Flush any free range we might be holding as a single 7403 // coalesced chunk to the appropriate free list. 7404 if (inFreeRange()) { 7405 assert(freeFinger() >= _sp->bottom() && freeFinger() < _limit, 7406 "freeFinger() " PTR_FORMAT " is out-of-bounds", p2i(freeFinger())); 7407 flush_cur_free_chunk(freeFinger(), 7408 pointer_delta(addr, freeFinger())); 7409 if (CMSTraceSweeper) { 7410 gclog_or_tty->print("Sweep: last chunk: "); 7411 gclog_or_tty->print("put_free_blk " PTR_FORMAT " (" SIZE_FORMAT ") " 7412 "[coalesced:%d]\n", 7413 p2i(freeFinger()), pointer_delta(addr, freeFinger()), 7414 lastFreeRangeCoalesced() ? 1 : 0); 7415 } 7416 } 7417 7418 // help the iterator loop finish 7419 return pointer_delta(_sp->end(), addr); 7420 } 7421 7422 assert(addr < _limit, "sweep invariant"); 7423 // check if we should yield 7424 do_yield_check(addr); 7425 if (fc->is_free()) { 7426 // Chunk that is already free 7427 res = fc->size(); 7428 do_already_free_chunk(fc); 7429 debug_only(_sp->verifyFreeLists()); 7430 // If we flush the chunk at hand in lookahead_and_flush() 7431 // and it's coalesced with a preceding chunk, then the 7432 // process of "mangling" the payload of the coalesced block 7433 // will cause erasure of the size information from the 7434 // (erstwhile) header of all the coalesced blocks but the 7435 // first, so the first disjunct in the assert will not hold 7436 // in that specific case (in which case the second disjunct 7437 // will hold). 7438 assert(res == fc->size() || ((HeapWord*)fc) + res >= _limit, 7439 "Otherwise the size info doesn't change at this step"); 7440 NOT_PRODUCT( 7441 _numObjectsAlreadyFree++; 7442 _numWordsAlreadyFree += res; 7443 ) 7444 NOT_PRODUCT(_last_fc = fc;) 7445 } else if (!_bitMap->isMarked(addr)) { 7446 // Chunk is fresh garbage 7447 res = do_garbage_chunk(fc); 7448 debug_only(_sp->verifyFreeLists()); 7449 NOT_PRODUCT( 7450 _numObjectsFreed++; 7451 _numWordsFreed += res; 7452 ) 7453 } else { 7454 // Chunk that is alive. 7455 res = do_live_chunk(fc); 7456 debug_only(_sp->verifyFreeLists()); 7457 NOT_PRODUCT( 7458 _numObjectsLive++; 7459 _numWordsLive += res; 7460 ) 7461 } 7462 return res; 7463 } 7464 7465 // For the smart allocation, record following 7466 // split deaths - a free chunk is removed from its free list because 7467 // it is being split into two or more chunks. 7468 // split birth - a free chunk is being added to its free list because 7469 // a larger free chunk has been split and resulted in this free chunk. 7470 // coal death - a free chunk is being removed from its free list because 7471 // it is being coalesced into a large free chunk. 7472 // coal birth - a free chunk is being added to its free list because 7473 // it was created when two or more free chunks where coalesced into 7474 // this free chunk. 7475 // 7476 // These statistics are used to determine the desired number of free 7477 // chunks of a given size. The desired number is chosen to be relative 7478 // to the end of a CMS sweep. The desired number at the end of a sweep 7479 // is the 7480 // count-at-end-of-previous-sweep (an amount that was enough) 7481 // - count-at-beginning-of-current-sweep (the excess) 7482 // + split-births (gains in this size during interval) 7483 // - split-deaths (demands on this size during interval) 7484 // where the interval is from the end of one sweep to the end of the 7485 // next. 7486 // 7487 // When sweeping the sweeper maintains an accumulated chunk which is 7488 // the chunk that is made up of chunks that have been coalesced. That 7489 // will be termed the left-hand chunk. A new chunk of garbage that 7490 // is being considered for coalescing will be referred to as the 7491 // right-hand chunk. 7492 // 7493 // When making a decision on whether to coalesce a right-hand chunk with 7494 // the current left-hand chunk, the current count vs. the desired count 7495 // of the left-hand chunk is considered. Also if the right-hand chunk 7496 // is near the large chunk at the end of the heap (see 7497 // ConcurrentMarkSweepGeneration::isNearLargestChunk()), then the 7498 // left-hand chunk is coalesced. 7499 // 7500 // When making a decision about whether to split a chunk, the desired count 7501 // vs. the current count of the candidate to be split is also considered. 7502 // If the candidate is underpopulated (currently fewer chunks than desired) 7503 // a chunk of an overpopulated (currently more chunks than desired) size may 7504 // be chosen. The "hint" associated with a free list, if non-null, points 7505 // to a free list which may be overpopulated. 7506 // 7507 7508 void SweepClosure::do_already_free_chunk(FreeChunk* fc) { 7509 const size_t size = fc->size(); 7510 // Chunks that cannot be coalesced are not in the 7511 // free lists. 7512 if (CMSTestInFreeList && !fc->cantCoalesce()) { 7513 assert(_sp->verify_chunk_in_free_list(fc), 7514 "free chunk should be in free lists"); 7515 } 7516 // a chunk that is already free, should not have been 7517 // marked in the bit map 7518 HeapWord* const addr = (HeapWord*) fc; 7519 assert(!_bitMap->isMarked(addr), "free chunk should be unmarked"); 7520 // Verify that the bit map has no bits marked between 7521 // addr and purported end of this block. 7522 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 7523 7524 // Some chunks cannot be coalesced under any circumstances. 7525 // See the definition of cantCoalesce(). 7526 if (!fc->cantCoalesce()) { 7527 // This chunk can potentially be coalesced. 7528 // All the work is done in 7529 do_post_free_or_garbage_chunk(fc, size); 7530 // Note that if the chunk is not coalescable (the else arm 7531 // below), we unconditionally flush, without needing to do 7532 // a "lookahead," as we do below. 7533 if (inFreeRange()) lookahead_and_flush(fc, size); 7534 } else { 7535 // Code path common to both original and adaptive free lists. 7536 7537 // cant coalesce with previous block; this should be treated 7538 // as the end of a free run if any 7539 if (inFreeRange()) { 7540 // we kicked some butt; time to pick up the garbage 7541 assert(freeFinger() < addr, "freeFinger points too high"); 7542 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 7543 } 7544 // else, nothing to do, just continue 7545 } 7546 } 7547 7548 size_t SweepClosure::do_garbage_chunk(FreeChunk* fc) { 7549 // This is a chunk of garbage. It is not in any free list. 7550 // Add it to a free list or let it possibly be coalesced into 7551 // a larger chunk. 7552 HeapWord* const addr = (HeapWord*) fc; 7553 const size_t size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 7554 7555 // Verify that the bit map has no bits marked between 7556 // addr and purported end of just dead object. 7557 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 7558 do_post_free_or_garbage_chunk(fc, size); 7559 7560 assert(_limit >= addr + size, 7561 "A freshly garbage chunk can't possibly straddle over _limit"); 7562 if (inFreeRange()) lookahead_and_flush(fc, size); 7563 return size; 7564 } 7565 7566 size_t SweepClosure::do_live_chunk(FreeChunk* fc) { 7567 HeapWord* addr = (HeapWord*) fc; 7568 // The sweeper has just found a live object. Return any accumulated 7569 // left hand chunk to the free lists. 7570 if (inFreeRange()) { 7571 assert(freeFinger() < addr, "freeFinger points too high"); 7572 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 7573 } 7574 7575 // This object is live: we'd normally expect this to be 7576 // an oop, and like to assert the following: 7577 // assert(oop(addr)->is_oop(), "live block should be an oop"); 7578 // However, as we commented above, this may be an object whose 7579 // header hasn't yet been initialized. 7580 size_t size; 7581 assert(_bitMap->isMarked(addr), "Tautology for this control point"); 7582 if (_bitMap->isMarked(addr + 1)) { 7583 // Determine the size from the bit map, rather than trying to 7584 // compute it from the object header. 7585 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 7586 size = pointer_delta(nextOneAddr + 1, addr); 7587 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 7588 "alignment problem"); 7589 7590 #ifdef ASSERT 7591 if (oop(addr)->klass_or_null() != NULL) { 7592 // Ignore mark word because we are running concurrent with mutators 7593 assert(oop(addr)->is_oop(true), "live block should be an oop"); 7594 assert(size == 7595 CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()), 7596 "P-mark and computed size do not agree"); 7597 } 7598 #endif 7599 7600 } else { 7601 // This should be an initialized object that's alive. 7602 assert(oop(addr)->klass_or_null() != NULL, 7603 "Should be an initialized object"); 7604 // Ignore mark word because we are running concurrent with mutators 7605 assert(oop(addr)->is_oop(true), "live block should be an oop"); 7606 // Verify that the bit map has no bits marked between 7607 // addr and purported end of this block. 7608 size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 7609 assert(size >= 3, "Necessary for Printezis marks to work"); 7610 assert(!_bitMap->isMarked(addr+1), "Tautology for this control point"); 7611 DEBUG_ONLY(_bitMap->verifyNoOneBitsInRange(addr+2, addr+size);) 7612 } 7613 return size; 7614 } 7615 7616 void SweepClosure::do_post_free_or_garbage_chunk(FreeChunk* fc, 7617 size_t chunkSize) { 7618 // do_post_free_or_garbage_chunk() should only be called in the case 7619 // of the adaptive free list allocator. 7620 const bool fcInFreeLists = fc->is_free(); 7621 assert((HeapWord*)fc <= _limit, "sweep invariant"); 7622 if (CMSTestInFreeList && fcInFreeLists) { 7623 assert(_sp->verify_chunk_in_free_list(fc), "free chunk is not in free lists"); 7624 } 7625 7626 if (CMSTraceSweeper) { 7627 gclog_or_tty->print_cr(" -- pick up another chunk at " PTR_FORMAT " (" SIZE_FORMAT ")", p2i(fc), chunkSize); 7628 } 7629 7630 HeapWord* const fc_addr = (HeapWord*) fc; 7631 7632 bool coalesce = false; 7633 const size_t left = pointer_delta(fc_addr, freeFinger()); 7634 const size_t right = chunkSize; 7635 switch (FLSCoalescePolicy) { 7636 // numeric value forms a coalition aggressiveness metric 7637 case 0: { // never coalesce 7638 coalesce = false; 7639 break; 7640 } 7641 case 1: { // coalesce if left & right chunks on overpopulated lists 7642 coalesce = _sp->coalOverPopulated(left) && 7643 _sp->coalOverPopulated(right); 7644 break; 7645 } 7646 case 2: { // coalesce if left chunk on overpopulated list (default) 7647 coalesce = _sp->coalOverPopulated(left); 7648 break; 7649 } 7650 case 3: { // coalesce if left OR right chunk on overpopulated list 7651 coalesce = _sp->coalOverPopulated(left) || 7652 _sp->coalOverPopulated(right); 7653 break; 7654 } 7655 case 4: { // always coalesce 7656 coalesce = true; 7657 break; 7658 } 7659 default: 7660 ShouldNotReachHere(); 7661 } 7662 7663 // Should the current free range be coalesced? 7664 // If the chunk is in a free range and either we decided to coalesce above 7665 // or the chunk is near the large block at the end of the heap 7666 // (isNearLargestChunk() returns true), then coalesce this chunk. 7667 const bool doCoalesce = inFreeRange() 7668 && (coalesce || _g->isNearLargestChunk(fc_addr)); 7669 if (doCoalesce) { 7670 // Coalesce the current free range on the left with the new 7671 // chunk on the right. If either is on a free list, 7672 // it must be removed from the list and stashed in the closure. 7673 if (freeRangeInFreeLists()) { 7674 FreeChunk* const ffc = (FreeChunk*)freeFinger(); 7675 assert(ffc->size() == pointer_delta(fc_addr, freeFinger()), 7676 "Size of free range is inconsistent with chunk size."); 7677 if (CMSTestInFreeList) { 7678 assert(_sp->verify_chunk_in_free_list(ffc), 7679 "Chunk is not in free lists"); 7680 } 7681 _sp->coalDeath(ffc->size()); 7682 _sp->removeFreeChunkFromFreeLists(ffc); 7683 set_freeRangeInFreeLists(false); 7684 } 7685 if (fcInFreeLists) { 7686 _sp->coalDeath(chunkSize); 7687 assert(fc->size() == chunkSize, 7688 "The chunk has the wrong size or is not in the free lists"); 7689 _sp->removeFreeChunkFromFreeLists(fc); 7690 } 7691 set_lastFreeRangeCoalesced(true); 7692 print_free_block_coalesced(fc); 7693 } else { // not in a free range and/or should not coalesce 7694 // Return the current free range and start a new one. 7695 if (inFreeRange()) { 7696 // In a free range but cannot coalesce with the right hand chunk. 7697 // Put the current free range into the free lists. 7698 flush_cur_free_chunk(freeFinger(), 7699 pointer_delta(fc_addr, freeFinger())); 7700 } 7701 // Set up for new free range. Pass along whether the right hand 7702 // chunk is in the free lists. 7703 initialize_free_range((HeapWord*)fc, fcInFreeLists); 7704 } 7705 } 7706 7707 // Lookahead flush: 7708 // If we are tracking a free range, and this is the last chunk that 7709 // we'll look at because its end crosses past _limit, we'll preemptively 7710 // flush it along with any free range we may be holding on to. Note that 7711 // this can be the case only for an already free or freshly garbage 7712 // chunk. If this block is an object, it can never straddle 7713 // over _limit. The "straddling" occurs when _limit is set at 7714 // the previous end of the space when this cycle started, and 7715 // a subsequent heap expansion caused the previously co-terminal 7716 // free block to be coalesced with the newly expanded portion, 7717 // thus rendering _limit a non-block-boundary making it dangerous 7718 // for the sweeper to step over and examine. 7719 void SweepClosure::lookahead_and_flush(FreeChunk* fc, size_t chunk_size) { 7720 assert(inFreeRange(), "Should only be called if currently in a free range."); 7721 HeapWord* const eob = ((HeapWord*)fc) + chunk_size; 7722 assert(_sp->used_region().contains(eob - 1), 7723 "eob = " PTR_FORMAT " eob-1 = " PTR_FORMAT " _limit = " PTR_FORMAT 7724 " out of bounds wrt _sp = [" PTR_FORMAT "," PTR_FORMAT ")" 7725 " when examining fc = " PTR_FORMAT "(" SIZE_FORMAT ")", 7726 p2i(eob), p2i(eob-1), p2i(_limit), p2i(_sp->bottom()), p2i(_sp->end()), p2i(fc), chunk_size); 7727 if (eob >= _limit) { 7728 assert(eob == _limit || fc->is_free(), "Only a free chunk should allow us to cross over the limit"); 7729 if (CMSTraceSweeper) { 7730 gclog_or_tty->print_cr("_limit " PTR_FORMAT " reached or crossed by block " 7731 "[" PTR_FORMAT "," PTR_FORMAT ") in space " 7732 "[" PTR_FORMAT "," PTR_FORMAT ")", 7733 p2i(_limit), p2i(fc), p2i(eob), p2i(_sp->bottom()), p2i(_sp->end())); 7734 } 7735 // Return the storage we are tracking back into the free lists. 7736 if (CMSTraceSweeper) { 7737 gclog_or_tty->print_cr("Flushing ... "); 7738 } 7739 assert(freeFinger() < eob, "Error"); 7740 flush_cur_free_chunk( freeFinger(), pointer_delta(eob, freeFinger())); 7741 } 7742 } 7743 7744 void SweepClosure::flush_cur_free_chunk(HeapWord* chunk, size_t size) { 7745 assert(inFreeRange(), "Should only be called if currently in a free range."); 7746 assert(size > 0, 7747 "A zero sized chunk cannot be added to the free lists."); 7748 if (!freeRangeInFreeLists()) { 7749 if (CMSTestInFreeList) { 7750 FreeChunk* fc = (FreeChunk*) chunk; 7751 fc->set_size(size); 7752 assert(!_sp->verify_chunk_in_free_list(fc), 7753 "chunk should not be in free lists yet"); 7754 } 7755 if (CMSTraceSweeper) { 7756 gclog_or_tty->print_cr(" -- add free block " PTR_FORMAT " (" SIZE_FORMAT ") to free lists", 7757 p2i(chunk), size); 7758 } 7759 // A new free range is going to be starting. The current 7760 // free range has not been added to the free lists yet or 7761 // was removed so add it back. 7762 // If the current free range was coalesced, then the death 7763 // of the free range was recorded. Record a birth now. 7764 if (lastFreeRangeCoalesced()) { 7765 _sp->coalBirth(size); 7766 } 7767 _sp->addChunkAndRepairOffsetTable(chunk, size, 7768 lastFreeRangeCoalesced()); 7769 } else if (CMSTraceSweeper) { 7770 gclog_or_tty->print_cr("Already in free list: nothing to flush"); 7771 } 7772 set_inFreeRange(false); 7773 set_freeRangeInFreeLists(false); 7774 } 7775 7776 // We take a break if we've been at this for a while, 7777 // so as to avoid monopolizing the locks involved. 7778 void SweepClosure::do_yield_work(HeapWord* addr) { 7779 // Return current free chunk being used for coalescing (if any) 7780 // to the appropriate freelist. After yielding, the next 7781 // free block encountered will start a coalescing range of 7782 // free blocks. If the next free block is adjacent to the 7783 // chunk just flushed, they will need to wait for the next 7784 // sweep to be coalesced. 7785 if (inFreeRange()) { 7786 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 7787 } 7788 7789 // First give up the locks, then yield, then re-lock. 7790 // We should probably use a constructor/destructor idiom to 7791 // do this unlock/lock or modify the MutexUnlocker class to 7792 // serve our purpose. XXX 7793 assert_lock_strong(_bitMap->lock()); 7794 assert_lock_strong(_freelistLock); 7795 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 7796 "CMS thread should hold CMS token"); 7797 _bitMap->lock()->unlock(); 7798 _freelistLock->unlock(); 7799 ConcurrentMarkSweepThread::desynchronize(true); 7800 _collector->stopTimer(); 7801 if (PrintCMSStatistics != 0) { 7802 _collector->incrementYields(); 7803 } 7804 7805 // See the comment in coordinator_yield() 7806 for (unsigned i = 0; i < CMSYieldSleepCount && 7807 ConcurrentMarkSweepThread::should_yield() && 7808 !CMSCollector::foregroundGCIsActive(); ++i) { 7809 os::sleep(Thread::current(), 1, false); 7810 } 7811 7812 ConcurrentMarkSweepThread::synchronize(true); 7813 _freelistLock->lock(); 7814 _bitMap->lock()->lock_without_safepoint_check(); 7815 _collector->startTimer(); 7816 } 7817 7818 #ifndef PRODUCT 7819 // This is actually very useful in a product build if it can 7820 // be called from the debugger. Compile it into the product 7821 // as needed. 7822 bool debug_verify_chunk_in_free_list(FreeChunk* fc) { 7823 return debug_cms_space->verify_chunk_in_free_list(fc); 7824 } 7825 #endif 7826 7827 void SweepClosure::print_free_block_coalesced(FreeChunk* fc) const { 7828 if (CMSTraceSweeper) { 7829 gclog_or_tty->print_cr("Sweep:coal_free_blk " PTR_FORMAT " (" SIZE_FORMAT ")", 7830 p2i(fc), fc->size()); 7831 } 7832 } 7833 7834 // CMSIsAliveClosure 7835 bool CMSIsAliveClosure::do_object_b(oop obj) { 7836 HeapWord* addr = (HeapWord*)obj; 7837 return addr != NULL && 7838 (!_span.contains(addr) || _bit_map->isMarked(addr)); 7839 } 7840 7841 7842 CMSKeepAliveClosure::CMSKeepAliveClosure( CMSCollector* collector, 7843 MemRegion span, 7844 CMSBitMap* bit_map, CMSMarkStack* mark_stack, 7845 bool cpc): 7846 _collector(collector), 7847 _span(span), 7848 _bit_map(bit_map), 7849 _mark_stack(mark_stack), 7850 _concurrent_precleaning(cpc) { 7851 assert(!_span.is_empty(), "Empty span could spell trouble"); 7852 } 7853 7854 7855 // CMSKeepAliveClosure: the serial version 7856 void CMSKeepAliveClosure::do_oop(oop obj) { 7857 HeapWord* addr = (HeapWord*)obj; 7858 if (_span.contains(addr) && 7859 !_bit_map->isMarked(addr)) { 7860 _bit_map->mark(addr); 7861 bool simulate_overflow = false; 7862 NOT_PRODUCT( 7863 if (CMSMarkStackOverflowALot && 7864 _collector->simulate_overflow()) { 7865 // simulate a stack overflow 7866 simulate_overflow = true; 7867 } 7868 ) 7869 if (simulate_overflow || !_mark_stack->push(obj)) { 7870 if (_concurrent_precleaning) { 7871 // We dirty the overflown object and let the remark 7872 // phase deal with it. 7873 assert(_collector->overflow_list_is_empty(), "Error"); 7874 // In the case of object arrays, we need to dirty all of 7875 // the cards that the object spans. No locking or atomics 7876 // are needed since no one else can be mutating the mod union 7877 // table. 7878 if (obj->is_objArray()) { 7879 size_t sz = obj->size(); 7880 HeapWord* end_card_addr = 7881 (HeapWord*)round_to((intptr_t)(addr+sz), CardTableModRefBS::card_size); 7882 MemRegion redirty_range = MemRegion(addr, end_card_addr); 7883 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 7884 _collector->_modUnionTable.mark_range(redirty_range); 7885 } else { 7886 _collector->_modUnionTable.mark(addr); 7887 } 7888 _collector->_ser_kac_preclean_ovflw++; 7889 } else { 7890 _collector->push_on_overflow_list(obj); 7891 _collector->_ser_kac_ovflw++; 7892 } 7893 } 7894 } 7895 } 7896 7897 void CMSKeepAliveClosure::do_oop(oop* p) { CMSKeepAliveClosure::do_oop_work(p); } 7898 void CMSKeepAliveClosure::do_oop(narrowOop* p) { CMSKeepAliveClosure::do_oop_work(p); } 7899 7900 // CMSParKeepAliveClosure: a parallel version of the above. 7901 // The work queues are private to each closure (thread), 7902 // but (may be) available for stealing by other threads. 7903 void CMSParKeepAliveClosure::do_oop(oop obj) { 7904 HeapWord* addr = (HeapWord*)obj; 7905 if (_span.contains(addr) && 7906 !_bit_map->isMarked(addr)) { 7907 // In general, during recursive tracing, several threads 7908 // may be concurrently getting here; the first one to 7909 // "tag" it, claims it. 7910 if (_bit_map->par_mark(addr)) { 7911 bool res = _work_queue->push(obj); 7912 assert(res, "Low water mark should be much less than capacity"); 7913 // Do a recursive trim in the hope that this will keep 7914 // stack usage lower, but leave some oops for potential stealers 7915 trim_queue(_low_water_mark); 7916 } // Else, another thread got there first 7917 } 7918 } 7919 7920 void CMSParKeepAliveClosure::do_oop(oop* p) { CMSParKeepAliveClosure::do_oop_work(p); } 7921 void CMSParKeepAliveClosure::do_oop(narrowOop* p) { CMSParKeepAliveClosure::do_oop_work(p); } 7922 7923 void CMSParKeepAliveClosure::trim_queue(uint max) { 7924 while (_work_queue->size() > max) { 7925 oop new_oop; 7926 if (_work_queue->pop_local(new_oop)) { 7927 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 7928 assert(_bit_map->isMarked((HeapWord*)new_oop), 7929 "no white objects on this stack!"); 7930 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 7931 // iterate over the oops in this oop, marking and pushing 7932 // the ones in CMS heap (i.e. in _span). 7933 new_oop->oop_iterate(&_mark_and_push); 7934 } 7935 } 7936 } 7937 7938 CMSInnerParMarkAndPushClosure::CMSInnerParMarkAndPushClosure( 7939 CMSCollector* collector, 7940 MemRegion span, CMSBitMap* bit_map, 7941 OopTaskQueue* work_queue): 7942 _collector(collector), 7943 _span(span), 7944 _bit_map(bit_map), 7945 _work_queue(work_queue) { } 7946 7947 void CMSInnerParMarkAndPushClosure::do_oop(oop obj) { 7948 HeapWord* addr = (HeapWord*)obj; 7949 if (_span.contains(addr) && 7950 !_bit_map->isMarked(addr)) { 7951 if (_bit_map->par_mark(addr)) { 7952 bool simulate_overflow = false; 7953 NOT_PRODUCT( 7954 if (CMSMarkStackOverflowALot && 7955 _collector->par_simulate_overflow()) { 7956 // simulate a stack overflow 7957 simulate_overflow = true; 7958 } 7959 ) 7960 if (simulate_overflow || !_work_queue->push(obj)) { 7961 _collector->par_push_on_overflow_list(obj); 7962 _collector->_par_kac_ovflw++; 7963 } 7964 } // Else another thread got there already 7965 } 7966 } 7967 7968 void CMSInnerParMarkAndPushClosure::do_oop(oop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); } 7969 void CMSInnerParMarkAndPushClosure::do_oop(narrowOop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); } 7970 7971 ////////////////////////////////////////////////////////////////// 7972 // CMSExpansionCause ///////////////////////////// 7973 ////////////////////////////////////////////////////////////////// 7974 const char* CMSExpansionCause::to_string(CMSExpansionCause::Cause cause) { 7975 switch (cause) { 7976 case _no_expansion: 7977 return "No expansion"; 7978 case _satisfy_free_ratio: 7979 return "Free ratio"; 7980 case _satisfy_promotion: 7981 return "Satisfy promotion"; 7982 case _satisfy_allocation: 7983 return "allocation"; 7984 case _allocate_par_lab: 7985 return "Par LAB"; 7986 case _allocate_par_spooling_space: 7987 return "Par Spooling Space"; 7988 case _adaptive_size_policy: 7989 return "Ergonomics"; 7990 default: 7991 return "unknown"; 7992 } 7993 } 7994 7995 void CMSDrainMarkingStackClosure::do_void() { 7996 // the max number to take from overflow list at a time 7997 const size_t num = _mark_stack->capacity()/4; 7998 assert(!_concurrent_precleaning || _collector->overflow_list_is_empty(), 7999 "Overflow list should be NULL during concurrent phases"); 8000 while (!_mark_stack->isEmpty() || 8001 // if stack is empty, check the overflow list 8002 _collector->take_from_overflow_list(num, _mark_stack)) { 8003 oop obj = _mark_stack->pop(); 8004 HeapWord* addr = (HeapWord*)obj; 8005 assert(_span.contains(addr), "Should be within span"); 8006 assert(_bit_map->isMarked(addr), "Should be marked"); 8007 assert(obj->is_oop(), "Should be an oop"); 8008 obj->oop_iterate(_keep_alive); 8009 } 8010 } 8011 8012 void CMSParDrainMarkingStackClosure::do_void() { 8013 // drain queue 8014 trim_queue(0); 8015 } 8016 8017 // Trim our work_queue so its length is below max at return 8018 void CMSParDrainMarkingStackClosure::trim_queue(uint max) { 8019 while (_work_queue->size() > max) { 8020 oop new_oop; 8021 if (_work_queue->pop_local(new_oop)) { 8022 assert(new_oop->is_oop(), "Expected an oop"); 8023 assert(_bit_map->isMarked((HeapWord*)new_oop), 8024 "no white objects on this stack!"); 8025 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 8026 // iterate over the oops in this oop, marking and pushing 8027 // the ones in CMS heap (i.e. in _span). 8028 new_oop->oop_iterate(&_mark_and_push); 8029 } 8030 } 8031 } 8032 8033 //////////////////////////////////////////////////////////////////// 8034 // Support for Marking Stack Overflow list handling and related code 8035 //////////////////////////////////////////////////////////////////// 8036 // Much of the following code is similar in shape and spirit to the 8037 // code used in ParNewGC. We should try and share that code 8038 // as much as possible in the future. 8039 8040 #ifndef PRODUCT 8041 // Debugging support for CMSStackOverflowALot 8042 8043 // It's OK to call this multi-threaded; the worst thing 8044 // that can happen is that we'll get a bunch of closely 8045 // spaced simulated overflows, but that's OK, in fact 8046 // probably good as it would exercise the overflow code 8047 // under contention. 8048 bool CMSCollector::simulate_overflow() { 8049 if (_overflow_counter-- <= 0) { // just being defensive 8050 _overflow_counter = CMSMarkStackOverflowInterval; 8051 return true; 8052 } else { 8053 return false; 8054 } 8055 } 8056 8057 bool CMSCollector::par_simulate_overflow() { 8058 return simulate_overflow(); 8059 } 8060 #endif 8061 8062 // Single-threaded 8063 bool CMSCollector::take_from_overflow_list(size_t num, CMSMarkStack* stack) { 8064 assert(stack->isEmpty(), "Expected precondition"); 8065 assert(stack->capacity() > num, "Shouldn't bite more than can chew"); 8066 size_t i = num; 8067 oop cur = _overflow_list; 8068 const markOop proto = markOopDesc::prototype(); 8069 NOT_PRODUCT(ssize_t n = 0;) 8070 for (oop next; i > 0 && cur != NULL; cur = next, i--) { 8071 next = oop(cur->mark()); 8072 cur->set_mark(proto); // until proven otherwise 8073 assert(cur->is_oop(), "Should be an oop"); 8074 bool res = stack->push(cur); 8075 assert(res, "Bit off more than can chew?"); 8076 NOT_PRODUCT(n++;) 8077 } 8078 _overflow_list = cur; 8079 #ifndef PRODUCT 8080 assert(_num_par_pushes >= n, "Too many pops?"); 8081 _num_par_pushes -=n; 8082 #endif 8083 return !stack->isEmpty(); 8084 } 8085 8086 #define BUSY (cast_to_oop<intptr_t>(0x1aff1aff)) 8087 // (MT-safe) Get a prefix of at most "num" from the list. 8088 // The overflow list is chained through the mark word of 8089 // each object in the list. We fetch the entire list, 8090 // break off a prefix of the right size and return the 8091 // remainder. If other threads try to take objects from 8092 // the overflow list at that time, they will wait for 8093 // some time to see if data becomes available. If (and 8094 // only if) another thread places one or more object(s) 8095 // on the global list before we have returned the suffix 8096 // to the global list, we will walk down our local list 8097 // to find its end and append the global list to 8098 // our suffix before returning it. This suffix walk can 8099 // prove to be expensive (quadratic in the amount of traffic) 8100 // when there are many objects in the overflow list and 8101 // there is much producer-consumer contention on the list. 8102 // *NOTE*: The overflow list manipulation code here and 8103 // in ParNewGeneration:: are very similar in shape, 8104 // except that in the ParNew case we use the old (from/eden) 8105 // copy of the object to thread the list via its klass word. 8106 // Because of the common code, if you make any changes in 8107 // the code below, please check the ParNew version to see if 8108 // similar changes might be needed. 8109 // CR 6797058 has been filed to consolidate the common code. 8110 bool CMSCollector::par_take_from_overflow_list(size_t num, 8111 OopTaskQueue* work_q, 8112 int no_of_gc_threads) { 8113 assert(work_q->size() == 0, "First empty local work queue"); 8114 assert(num < work_q->max_elems(), "Can't bite more than we can chew"); 8115 if (_overflow_list == NULL) { 8116 return false; 8117 } 8118 // Grab the entire list; we'll put back a suffix 8119 oop prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list)); 8120 Thread* tid = Thread::current(); 8121 // Before "no_of_gc_threads" was introduced CMSOverflowSpinCount was 8122 // set to ParallelGCThreads. 8123 size_t CMSOverflowSpinCount = (size_t) no_of_gc_threads; // was ParallelGCThreads; 8124 size_t sleep_time_millis = MAX2((size_t)1, num/100); 8125 // If the list is busy, we spin for a short while, 8126 // sleeping between attempts to get the list. 8127 for (size_t spin = 0; prefix == BUSY && spin < CMSOverflowSpinCount; spin++) { 8128 os::sleep(tid, sleep_time_millis, false); 8129 if (_overflow_list == NULL) { 8130 // Nothing left to take 8131 return false; 8132 } else if (_overflow_list != BUSY) { 8133 // Try and grab the prefix 8134 prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list)); 8135 } 8136 } 8137 // If the list was found to be empty, or we spun long 8138 // enough, we give up and return empty-handed. If we leave 8139 // the list in the BUSY state below, it must be the case that 8140 // some other thread holds the overflow list and will set it 8141 // to a non-BUSY state in the future. 8142 if (prefix == NULL || prefix == BUSY) { 8143 // Nothing to take or waited long enough 8144 if (prefix == NULL) { 8145 // Write back the NULL in case we overwrote it with BUSY above 8146 // and it is still the same value. 8147 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY); 8148 } 8149 return false; 8150 } 8151 assert(prefix != NULL && prefix != BUSY, "Error"); 8152 size_t i = num; 8153 oop cur = prefix; 8154 // Walk down the first "num" objects, unless we reach the end. 8155 for (; i > 1 && cur->mark() != NULL; cur = oop(cur->mark()), i--); 8156 if (cur->mark() == NULL) { 8157 // We have "num" or fewer elements in the list, so there 8158 // is nothing to return to the global list. 8159 // Write back the NULL in lieu of the BUSY we wrote 8160 // above, if it is still the same value. 8161 if (_overflow_list == BUSY) { 8162 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY); 8163 } 8164 } else { 8165 // Chop off the suffix and return it to the global list. 8166 assert(cur->mark() != BUSY, "Error"); 8167 oop suffix_head = cur->mark(); // suffix will be put back on global list 8168 cur->set_mark(NULL); // break off suffix 8169 // It's possible that the list is still in the empty(busy) state 8170 // we left it in a short while ago; in that case we may be 8171 // able to place back the suffix without incurring the cost 8172 // of a walk down the list. 8173 oop observed_overflow_list = _overflow_list; 8174 oop cur_overflow_list = observed_overflow_list; 8175 bool attached = false; 8176 while (observed_overflow_list == BUSY || observed_overflow_list == NULL) { 8177 observed_overflow_list = 8178 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list); 8179 if (cur_overflow_list == observed_overflow_list) { 8180 attached = true; 8181 break; 8182 } else cur_overflow_list = observed_overflow_list; 8183 } 8184 if (!attached) { 8185 // Too bad, someone else sneaked in (at least) an element; we'll need 8186 // to do a splice. Find tail of suffix so we can prepend suffix to global 8187 // list. 8188 for (cur = suffix_head; cur->mark() != NULL; cur = (oop)(cur->mark())); 8189 oop suffix_tail = cur; 8190 assert(suffix_tail != NULL && suffix_tail->mark() == NULL, 8191 "Tautology"); 8192 observed_overflow_list = _overflow_list; 8193 do { 8194 cur_overflow_list = observed_overflow_list; 8195 if (cur_overflow_list != BUSY) { 8196 // Do the splice ... 8197 suffix_tail->set_mark(markOop(cur_overflow_list)); 8198 } else { // cur_overflow_list == BUSY 8199 suffix_tail->set_mark(NULL); 8200 } 8201 // ... and try to place spliced list back on overflow_list ... 8202 observed_overflow_list = 8203 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list); 8204 } while (cur_overflow_list != observed_overflow_list); 8205 // ... until we have succeeded in doing so. 8206 } 8207 } 8208 8209 // Push the prefix elements on work_q 8210 assert(prefix != NULL, "control point invariant"); 8211 const markOop proto = markOopDesc::prototype(); 8212 oop next; 8213 NOT_PRODUCT(ssize_t n = 0;) 8214 for (cur = prefix; cur != NULL; cur = next) { 8215 next = oop(cur->mark()); 8216 cur->set_mark(proto); // until proven otherwise 8217 assert(cur->is_oop(), "Should be an oop"); 8218 bool res = work_q->push(cur); 8219 assert(res, "Bit off more than we can chew?"); 8220 NOT_PRODUCT(n++;) 8221 } 8222 #ifndef PRODUCT 8223 assert(_num_par_pushes >= n, "Too many pops?"); 8224 Atomic::add_ptr(-(intptr_t)n, &_num_par_pushes); 8225 #endif 8226 return true; 8227 } 8228 8229 // Single-threaded 8230 void CMSCollector::push_on_overflow_list(oop p) { 8231 NOT_PRODUCT(_num_par_pushes++;) 8232 assert(p->is_oop(), "Not an oop"); 8233 preserve_mark_if_necessary(p); 8234 p->set_mark((markOop)_overflow_list); 8235 _overflow_list = p; 8236 } 8237 8238 // Multi-threaded; use CAS to prepend to overflow list 8239 void CMSCollector::par_push_on_overflow_list(oop p) { 8240 NOT_PRODUCT(Atomic::inc_ptr(&_num_par_pushes);) 8241 assert(p->is_oop(), "Not an oop"); 8242 par_preserve_mark_if_necessary(p); 8243 oop observed_overflow_list = _overflow_list; 8244 oop cur_overflow_list; 8245 do { 8246 cur_overflow_list = observed_overflow_list; 8247 if (cur_overflow_list != BUSY) { 8248 p->set_mark(markOop(cur_overflow_list)); 8249 } else { 8250 p->set_mark(NULL); 8251 } 8252 observed_overflow_list = 8253 (oop) Atomic::cmpxchg_ptr(p, &_overflow_list, cur_overflow_list); 8254 } while (cur_overflow_list != observed_overflow_list); 8255 } 8256 #undef BUSY 8257 8258 // Single threaded 8259 // General Note on GrowableArray: pushes may silently fail 8260 // because we are (temporarily) out of C-heap for expanding 8261 // the stack. The problem is quite ubiquitous and affects 8262 // a lot of code in the JVM. The prudent thing for GrowableArray 8263 // to do (for now) is to exit with an error. However, that may 8264 // be too draconian in some cases because the caller may be 8265 // able to recover without much harm. For such cases, we 8266 // should probably introduce a "soft_push" method which returns 8267 // an indication of success or failure with the assumption that 8268 // the caller may be able to recover from a failure; code in 8269 // the VM can then be changed, incrementally, to deal with such 8270 // failures where possible, thus, incrementally hardening the VM 8271 // in such low resource situations. 8272 void CMSCollector::preserve_mark_work(oop p, markOop m) { 8273 _preserved_oop_stack.push(p); 8274 _preserved_mark_stack.push(m); 8275 assert(m == p->mark(), "Mark word changed"); 8276 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), 8277 "bijection"); 8278 } 8279 8280 // Single threaded 8281 void CMSCollector::preserve_mark_if_necessary(oop p) { 8282 markOop m = p->mark(); 8283 if (m->must_be_preserved(p)) { 8284 preserve_mark_work(p, m); 8285 } 8286 } 8287 8288 void CMSCollector::par_preserve_mark_if_necessary(oop p) { 8289 markOop m = p->mark(); 8290 if (m->must_be_preserved(p)) { 8291 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag); 8292 // Even though we read the mark word without holding 8293 // the lock, we are assured that it will not change 8294 // because we "own" this oop, so no other thread can 8295 // be trying to push it on the overflow list; see 8296 // the assertion in preserve_mark_work() that checks 8297 // that m == p->mark(). 8298 preserve_mark_work(p, m); 8299 } 8300 } 8301 8302 // We should be able to do this multi-threaded, 8303 // a chunk of stack being a task (this is 8304 // correct because each oop only ever appears 8305 // once in the overflow list. However, it's 8306 // not very easy to completely overlap this with 8307 // other operations, so will generally not be done 8308 // until all work's been completed. Because we 8309 // expect the preserved oop stack (set) to be small, 8310 // it's probably fine to do this single-threaded. 8311 // We can explore cleverer concurrent/overlapped/parallel 8312 // processing of preserved marks if we feel the 8313 // need for this in the future. Stack overflow should 8314 // be so rare in practice and, when it happens, its 8315 // effect on performance so great that this will 8316 // likely just be in the noise anyway. 8317 void CMSCollector::restore_preserved_marks_if_any() { 8318 assert(SafepointSynchronize::is_at_safepoint(), 8319 "world should be stopped"); 8320 assert(Thread::current()->is_ConcurrentGC_thread() || 8321 Thread::current()->is_VM_thread(), 8322 "should be single-threaded"); 8323 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), 8324 "bijection"); 8325 8326 while (!_preserved_oop_stack.is_empty()) { 8327 oop p = _preserved_oop_stack.pop(); 8328 assert(p->is_oop(), "Should be an oop"); 8329 assert(_span.contains(p), "oop should be in _span"); 8330 assert(p->mark() == markOopDesc::prototype(), 8331 "Set when taken from overflow list"); 8332 markOop m = _preserved_mark_stack.pop(); 8333 p->set_mark(m); 8334 } 8335 assert(_preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(), 8336 "stacks were cleared above"); 8337 } 8338 8339 #ifndef PRODUCT 8340 bool CMSCollector::no_preserved_marks() const { 8341 return _preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(); 8342 } 8343 #endif 8344 8345 // Transfer some number of overflown objects to usual marking 8346 // stack. Return true if some objects were transferred. 8347 bool MarkRefsIntoAndScanClosure::take_from_overflow_list() { 8348 size_t num = MIN2((size_t)(_mark_stack->capacity() - _mark_stack->length())/4, 8349 (size_t)ParGCDesiredObjsFromOverflowList); 8350 8351 bool res = _collector->take_from_overflow_list(num, _mark_stack); 8352 assert(_collector->overflow_list_is_empty() || res, 8353 "If list is not empty, we should have taken something"); 8354 assert(!res || !_mark_stack->isEmpty(), 8355 "If we took something, it should now be on our stack"); 8356 return res; 8357 } 8358 8359 size_t MarkDeadObjectsClosure::do_blk(HeapWord* addr) { 8360 size_t res = _sp->block_size_no_stall(addr, _collector); 8361 if (_sp->block_is_obj(addr)) { 8362 if (_live_bit_map->isMarked(addr)) { 8363 // It can't have been dead in a previous cycle 8364 guarantee(!_dead_bit_map->isMarked(addr), "No resurrection!"); 8365 } else { 8366 _dead_bit_map->mark(addr); // mark the dead object 8367 } 8368 } 8369 // Could be 0, if the block size could not be computed without stalling. 8370 return res; 8371 } 8372 8373 TraceCMSMemoryManagerStats::TraceCMSMemoryManagerStats(CMSCollector::CollectorState phase, GCCause::Cause cause): TraceMemoryManagerStats() { 8374 8375 switch (phase) { 8376 case CMSCollector::InitialMarking: 8377 initialize(true /* fullGC */ , 8378 cause /* cause of the GC */, 8379 true /* recordGCBeginTime */, 8380 true /* recordPreGCUsage */, 8381 false /* recordPeakUsage */, 8382 false /* recordPostGCusage */, 8383 true /* recordAccumulatedGCTime */, 8384 false /* recordGCEndTime */, 8385 false /* countCollection */ ); 8386 break; 8387 8388 case CMSCollector::FinalMarking: 8389 initialize(true /* fullGC */ , 8390 cause /* cause of the GC */, 8391 false /* recordGCBeginTime */, 8392 false /* recordPreGCUsage */, 8393 false /* recordPeakUsage */, 8394 false /* recordPostGCusage */, 8395 true /* recordAccumulatedGCTime */, 8396 false /* recordGCEndTime */, 8397 false /* countCollection */ ); 8398 break; 8399 8400 case CMSCollector::Sweeping: 8401 initialize(true /* fullGC */ , 8402 cause /* cause of the GC */, 8403 false /* recordGCBeginTime */, 8404 false /* recordPreGCUsage */, 8405 true /* recordPeakUsage */, 8406 true /* recordPostGCusage */, 8407 false /* recordAccumulatedGCTime */, 8408 true /* recordGCEndTime */, 8409 true /* countCollection */ ); 8410 break; 8411 8412 default: 8413 ShouldNotReachHere(); 8414 } 8415 }