1 /* 2 * Copyright (c) 2001, 2010, 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 "incls/_precompiled.incl" 26 # include "incls/_concurrentMarkSweepGeneration.cpp.incl" 27 28 // statics 29 CMSCollector* ConcurrentMarkSweepGeneration::_collector = NULL; 30 bool CMSCollector::_full_gc_requested = false; 31 32 ////////////////////////////////////////////////////////////////// 33 // In support of CMS/VM thread synchronization 34 ////////////////////////////////////////////////////////////////// 35 // We split use of the CGC_lock into 2 "levels". 36 // The low-level locking is of the usual CGC_lock monitor. We introduce 37 // a higher level "token" (hereafter "CMS token") built on top of the 38 // low level monitor (hereafter "CGC lock"). 39 // The token-passing protocol gives priority to the VM thread. The 40 // CMS-lock doesn't provide any fairness guarantees, but clients 41 // should ensure that it is only held for very short, bounded 42 // durations. 43 // 44 // When either of the CMS thread or the VM thread is involved in 45 // collection operations during which it does not want the other 46 // thread to interfere, it obtains the CMS token. 47 // 48 // If either thread tries to get the token while the other has 49 // it, that thread waits. However, if the VM thread and CMS thread 50 // both want the token, then the VM thread gets priority while the 51 // CMS thread waits. This ensures, for instance, that the "concurrent" 52 // phases of the CMS thread's work do not block out the VM thread 53 // for long periods of time as the CMS thread continues to hog 54 // the token. (See bug 4616232). 55 // 56 // The baton-passing functions are, however, controlled by the 57 // flags _foregroundGCShouldWait and _foregroundGCIsActive, 58 // and here the low-level CMS lock, not the high level token, 59 // ensures mutual exclusion. 60 // 61 // Two important conditions that we have to satisfy: 62 // 1. if a thread does a low-level wait on the CMS lock, then it 63 // relinquishes the CMS token if it were holding that token 64 // when it acquired the low-level CMS lock. 65 // 2. any low-level notifications on the low-level lock 66 // should only be sent when a thread has relinquished the token. 67 // 68 // In the absence of either property, we'd have potential deadlock. 69 // 70 // We protect each of the CMS (concurrent and sequential) phases 71 // with the CMS _token_, not the CMS _lock_. 72 // 73 // The only code protected by CMS lock is the token acquisition code 74 // itself, see ConcurrentMarkSweepThread::[de]synchronize(), and the 75 // baton-passing code. 76 // 77 // Unfortunately, i couldn't come up with a good abstraction to factor and 78 // hide the naked CGC_lock manipulation in the baton-passing code 79 // further below. That's something we should try to do. Also, the proof 80 // of correctness of this 2-level locking scheme is far from obvious, 81 // and potentially quite slippery. We have an uneasy supsicion, for instance, 82 // that there may be a theoretical possibility of delay/starvation in the 83 // low-level lock/wait/notify scheme used for the baton-passing because of 84 // potential intereference with the priority scheme embodied in the 85 // CMS-token-passing protocol. See related comments at a CGC_lock->wait() 86 // invocation further below and marked with "XXX 20011219YSR". 87 // Indeed, as we note elsewhere, this may become yet more slippery 88 // in the presence of multiple CMS and/or multiple VM threads. XXX 89 90 class CMSTokenSync: public StackObj { 91 private: 92 bool _is_cms_thread; 93 public: 94 CMSTokenSync(bool is_cms_thread): 95 _is_cms_thread(is_cms_thread) { 96 assert(is_cms_thread == Thread::current()->is_ConcurrentGC_thread(), 97 "Incorrect argument to constructor"); 98 ConcurrentMarkSweepThread::synchronize(_is_cms_thread); 99 } 100 101 ~CMSTokenSync() { 102 assert(_is_cms_thread ? 103 ConcurrentMarkSweepThread::cms_thread_has_cms_token() : 104 ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 105 "Incorrect state"); 106 ConcurrentMarkSweepThread::desynchronize(_is_cms_thread); 107 } 108 }; 109 110 // Convenience class that does a CMSTokenSync, and then acquires 111 // upto three locks. 112 class CMSTokenSyncWithLocks: public CMSTokenSync { 113 private: 114 // Note: locks are acquired in textual declaration order 115 // and released in the opposite order 116 MutexLockerEx _locker1, _locker2, _locker3; 117 public: 118 CMSTokenSyncWithLocks(bool is_cms_thread, Mutex* mutex1, 119 Mutex* mutex2 = NULL, Mutex* mutex3 = NULL): 120 CMSTokenSync(is_cms_thread), 121 _locker1(mutex1, Mutex::_no_safepoint_check_flag), 122 _locker2(mutex2, Mutex::_no_safepoint_check_flag), 123 _locker3(mutex3, Mutex::_no_safepoint_check_flag) 124 { } 125 }; 126 127 128 // Wrapper class to temporarily disable icms during a foreground cms collection. 129 class ICMSDisabler: public StackObj { 130 public: 131 // The ctor disables icms and wakes up the thread so it notices the change; 132 // the dtor re-enables icms. Note that the CMSCollector methods will check 133 // CMSIncrementalMode. 134 ICMSDisabler() { CMSCollector::disable_icms(); CMSCollector::start_icms(); } 135 ~ICMSDisabler() { CMSCollector::enable_icms(); } 136 }; 137 138 ////////////////////////////////////////////////////////////////// 139 // Concurrent Mark-Sweep Generation ///////////////////////////// 140 ////////////////////////////////////////////////////////////////// 141 142 NOT_PRODUCT(CompactibleFreeListSpace* debug_cms_space;) 143 144 // This struct contains per-thread things necessary to support parallel 145 // young-gen collection. 146 class CMSParGCThreadState: public CHeapObj { 147 public: 148 CFLS_LAB lab; 149 PromotionInfo promo; 150 151 // Constructor. 152 CMSParGCThreadState(CompactibleFreeListSpace* cfls) : lab(cfls) { 153 promo.setSpace(cfls); 154 } 155 }; 156 157 ConcurrentMarkSweepGeneration::ConcurrentMarkSweepGeneration( 158 ReservedSpace rs, size_t initial_byte_size, int level, 159 CardTableRS* ct, bool use_adaptive_freelists, 160 FreeBlockDictionary::DictionaryChoice dictionaryChoice) : 161 CardGeneration(rs, initial_byte_size, level, ct), 162 _dilatation_factor(((double)MinChunkSize)/((double)(CollectedHeap::min_fill_size()))), 163 _debug_collection_type(Concurrent_collection_type) 164 { 165 HeapWord* bottom = (HeapWord*) _virtual_space.low(); 166 HeapWord* end = (HeapWord*) _virtual_space.high(); 167 168 _direct_allocated_words = 0; 169 NOT_PRODUCT( 170 _numObjectsPromoted = 0; 171 _numWordsPromoted = 0; 172 _numObjectsAllocated = 0; 173 _numWordsAllocated = 0; 174 ) 175 176 _cmsSpace = new CompactibleFreeListSpace(_bts, MemRegion(bottom, end), 177 use_adaptive_freelists, 178 dictionaryChoice); 179 NOT_PRODUCT(debug_cms_space = _cmsSpace;) 180 if (_cmsSpace == NULL) { 181 vm_exit_during_initialization( 182 "CompactibleFreeListSpace allocation failure"); 183 } 184 _cmsSpace->_gen = this; 185 186 _gc_stats = new CMSGCStats(); 187 188 // Verify the assumption that FreeChunk::_prev and OopDesc::_klass 189 // offsets match. The ability to tell free chunks from objects 190 // depends on this property. 191 debug_only( 192 FreeChunk* junk = NULL; 193 assert(UseCompressedOops || 194 junk->prev_addr() == (void*)(oop(junk)->klass_addr()), 195 "Offset of FreeChunk::_prev within FreeChunk must match" 196 " that of OopDesc::_klass within OopDesc"); 197 ) 198 if (ParallelGCThreads > 0) { 199 typedef CMSParGCThreadState* CMSParGCThreadStatePtr; 200 _par_gc_thread_states = 201 NEW_C_HEAP_ARRAY(CMSParGCThreadStatePtr, ParallelGCThreads); 202 if (_par_gc_thread_states == NULL) { 203 vm_exit_during_initialization("Could not allocate par gc structs"); 204 } 205 for (uint i = 0; i < ParallelGCThreads; i++) { 206 _par_gc_thread_states[i] = new CMSParGCThreadState(cmsSpace()); 207 if (_par_gc_thread_states[i] == NULL) { 208 vm_exit_during_initialization("Could not allocate par gc structs"); 209 } 210 } 211 } else { 212 _par_gc_thread_states = NULL; 213 } 214 _incremental_collection_failed = false; 215 // The "dilatation_factor" is the expansion that can occur on 216 // account of the fact that the minimum object size in the CMS 217 // generation may be larger than that in, say, a contiguous young 218 // generation. 219 // Ideally, in the calculation below, we'd compute the dilatation 220 // factor as: MinChunkSize/(promoting_gen's min object size) 221 // Since we do not have such a general query interface for the 222 // promoting generation, we'll instead just use the mimimum 223 // object size (which today is a header's worth of space); 224 // note that all arithmetic is in units of HeapWords. 225 assert(MinChunkSize >= CollectedHeap::min_fill_size(), "just checking"); 226 assert(_dilatation_factor >= 1.0, "from previous assert"); 227 } 228 229 230 // The field "_initiating_occupancy" represents the occupancy percentage 231 // at which we trigger a new collection cycle. Unless explicitly specified 232 // via CMSInitiating[Perm]OccupancyFraction (argument "io" below), it 233 // is calculated by: 234 // 235 // Let "f" be MinHeapFreeRatio in 236 // 237 // _intiating_occupancy = 100-f + 238 // f * (CMSTrigger[Perm]Ratio/100) 239 // where CMSTrigger[Perm]Ratio is the argument "tr" below. 240 // 241 // That is, if we assume the heap is at its desired maximum occupancy at the 242 // end of a collection, we let CMSTrigger[Perm]Ratio of the (purported) free 243 // space be allocated before initiating a new collection cycle. 244 // 245 void ConcurrentMarkSweepGeneration::init_initiating_occupancy(intx io, intx tr) { 246 assert(io <= 100 && tr >= 0 && tr <= 100, "Check the arguments"); 247 if (io >= 0) { 248 _initiating_occupancy = (double)io / 100.0; 249 } else { 250 _initiating_occupancy = ((100 - MinHeapFreeRatio) + 251 (double)(tr * MinHeapFreeRatio) / 100.0) 252 / 100.0; 253 } 254 } 255 256 void ConcurrentMarkSweepGeneration::ref_processor_init() { 257 assert(collector() != NULL, "no collector"); 258 collector()->ref_processor_init(); 259 } 260 261 void CMSCollector::ref_processor_init() { 262 if (_ref_processor == NULL) { 263 // Allocate and initialize a reference processor 264 _ref_processor = ReferenceProcessor::create_ref_processor( 265 _span, // span 266 _cmsGen->refs_discovery_is_atomic(), // atomic_discovery 267 _cmsGen->refs_discovery_is_mt(), // mt_discovery 268 &_is_alive_closure, 269 ParallelGCThreads, 270 ParallelRefProcEnabled); 271 // Initialize the _ref_processor field of CMSGen 272 _cmsGen->set_ref_processor(_ref_processor); 273 274 // Allocate a dummy ref processor for perm gen. 275 ReferenceProcessor* rp2 = new ReferenceProcessor(); 276 if (rp2 == NULL) { 277 vm_exit_during_initialization("Could not allocate ReferenceProcessor object"); 278 } 279 _permGen->set_ref_processor(rp2); 280 } 281 } 282 283 CMSAdaptiveSizePolicy* CMSCollector::size_policy() { 284 GenCollectedHeap* gch = GenCollectedHeap::heap(); 285 assert(gch->kind() == CollectedHeap::GenCollectedHeap, 286 "Wrong type of heap"); 287 CMSAdaptiveSizePolicy* sp = (CMSAdaptiveSizePolicy*) 288 gch->gen_policy()->size_policy(); 289 assert(sp->is_gc_cms_adaptive_size_policy(), 290 "Wrong type of size policy"); 291 return sp; 292 } 293 294 CMSGCAdaptivePolicyCounters* CMSCollector::gc_adaptive_policy_counters() { 295 CMSGCAdaptivePolicyCounters* results = 296 (CMSGCAdaptivePolicyCounters*) collector_policy()->counters(); 297 assert( 298 results->kind() == GCPolicyCounters::CMSGCAdaptivePolicyCountersKind, 299 "Wrong gc policy counter kind"); 300 return results; 301 } 302 303 304 void ConcurrentMarkSweepGeneration::initialize_performance_counters() { 305 306 const char* gen_name = "old"; 307 308 // Generation Counters - generation 1, 1 subspace 309 _gen_counters = new GenerationCounters(gen_name, 1, 1, &_virtual_space); 310 311 _space_counters = new GSpaceCounters(gen_name, 0, 312 _virtual_space.reserved_size(), 313 this, _gen_counters); 314 } 315 316 CMSStats::CMSStats(ConcurrentMarkSweepGeneration* cms_gen, unsigned int alpha): 317 _cms_gen(cms_gen) 318 { 319 assert(alpha <= 100, "bad value"); 320 _saved_alpha = alpha; 321 322 // Initialize the alphas to the bootstrap value of 100. 323 _gc0_alpha = _cms_alpha = 100; 324 325 _cms_begin_time.update(); 326 _cms_end_time.update(); 327 328 _gc0_duration = 0.0; 329 _gc0_period = 0.0; 330 _gc0_promoted = 0; 331 332 _cms_duration = 0.0; 333 _cms_period = 0.0; 334 _cms_allocated = 0; 335 336 _cms_used_at_gc0_begin = 0; 337 _cms_used_at_gc0_end = 0; 338 _allow_duty_cycle_reduction = false; 339 _valid_bits = 0; 340 _icms_duty_cycle = CMSIncrementalDutyCycle; 341 } 342 343 double CMSStats::cms_free_adjustment_factor(size_t free) const { 344 // TBD: CR 6909490 345 return 1.0; 346 } 347 348 void CMSStats::adjust_cms_free_adjustment_factor(bool fail, size_t free) { 349 } 350 351 // If promotion failure handling is on use 352 // the padded average size of the promotion for each 353 // young generation collection. 354 double CMSStats::time_until_cms_gen_full() const { 355 size_t cms_free = _cms_gen->cmsSpace()->free(); 356 GenCollectedHeap* gch = GenCollectedHeap::heap(); 357 size_t expected_promotion = gch->get_gen(0)->capacity(); 358 if (HandlePromotionFailure) { 359 expected_promotion = MIN2( 360 (size_t) _cms_gen->gc_stats()->avg_promoted()->padded_average(), 361 expected_promotion); 362 } 363 if (cms_free > expected_promotion) { 364 // Start a cms collection if there isn't enough space to promote 365 // for the next minor collection. Use the padded average as 366 // a safety factor. 367 cms_free -= expected_promotion; 368 369 // Adjust by the safety factor. 370 double cms_free_dbl = (double)cms_free; 371 double cms_adjustment = (100.0 - CMSIncrementalSafetyFactor)/100.0; 372 // Apply a further correction factor which tries to adjust 373 // for recent occurance of concurrent mode failures. 374 cms_adjustment = cms_adjustment * cms_free_adjustment_factor(cms_free); 375 cms_free_dbl = cms_free_dbl * cms_adjustment; 376 377 if (PrintGCDetails && Verbose) { 378 gclog_or_tty->print_cr("CMSStats::time_until_cms_gen_full: cms_free " 379 SIZE_FORMAT " expected_promotion " SIZE_FORMAT, 380 cms_free, expected_promotion); 381 gclog_or_tty->print_cr(" cms_free_dbl %f cms_consumption_rate %f", 382 cms_free_dbl, cms_consumption_rate() + 1.0); 383 } 384 // Add 1 in case the consumption rate goes to zero. 385 return cms_free_dbl / (cms_consumption_rate() + 1.0); 386 } 387 return 0.0; 388 } 389 390 // Compare the duration of the cms collection to the 391 // time remaining before the cms generation is empty. 392 // Note that the time from the start of the cms collection 393 // to the start of the cms sweep (less than the total 394 // duration of the cms collection) can be used. This 395 // has been tried and some applications experienced 396 // promotion failures early in execution. This was 397 // possibly because the averages were not accurate 398 // enough at the beginning. 399 double CMSStats::time_until_cms_start() const { 400 // We add "gc0_period" to the "work" calculation 401 // below because this query is done (mostly) at the 402 // end of a scavenge, so we need to conservatively 403 // account for that much possible delay 404 // in the query so as to avoid concurrent mode failures 405 // due to starting the collection just a wee bit too 406 // late. 407 double work = cms_duration() + gc0_period(); 408 double deadline = time_until_cms_gen_full(); 409 // If a concurrent mode failure occurred recently, we want to be 410 // more conservative and halve our expected time_until_cms_gen_full() 411 if (work > deadline) { 412 if (Verbose && PrintGCDetails) { 413 gclog_or_tty->print( 414 " CMSCollector: collect because of anticipated promotion " 415 "before full %3.7f + %3.7f > %3.7f ", cms_duration(), 416 gc0_period(), time_until_cms_gen_full()); 417 } 418 return 0.0; 419 } 420 return work - deadline; 421 } 422 423 // Return a duty cycle based on old_duty_cycle and new_duty_cycle, limiting the 424 // amount of change to prevent wild oscillation. 425 unsigned int CMSStats::icms_damped_duty_cycle(unsigned int old_duty_cycle, 426 unsigned int new_duty_cycle) { 427 assert(old_duty_cycle <= 100, "bad input value"); 428 assert(new_duty_cycle <= 100, "bad input value"); 429 430 // Note: use subtraction with caution since it may underflow (values are 431 // unsigned). Addition is safe since we're in the range 0-100. 432 unsigned int damped_duty_cycle = new_duty_cycle; 433 if (new_duty_cycle < old_duty_cycle) { 434 const unsigned int largest_delta = MAX2(old_duty_cycle / 4, 5U); 435 if (new_duty_cycle + largest_delta < old_duty_cycle) { 436 damped_duty_cycle = old_duty_cycle - largest_delta; 437 } 438 } else if (new_duty_cycle > old_duty_cycle) { 439 const unsigned int largest_delta = MAX2(old_duty_cycle / 4, 15U); 440 if (new_duty_cycle > old_duty_cycle + largest_delta) { 441 damped_duty_cycle = MIN2(old_duty_cycle + largest_delta, 100U); 442 } 443 } 444 assert(damped_duty_cycle <= 100, "invalid duty cycle computed"); 445 446 if (CMSTraceIncrementalPacing) { 447 gclog_or_tty->print(" [icms_damped_duty_cycle(%d,%d) = %d] ", 448 old_duty_cycle, new_duty_cycle, damped_duty_cycle); 449 } 450 return damped_duty_cycle; 451 } 452 453 unsigned int CMSStats::icms_update_duty_cycle_impl() { 454 assert(CMSIncrementalPacing && valid(), 455 "should be handled in icms_update_duty_cycle()"); 456 457 double cms_time_so_far = cms_timer().seconds(); 458 double scaled_duration = cms_duration_per_mb() * _cms_used_at_gc0_end / M; 459 double scaled_duration_remaining = fabsd(scaled_duration - cms_time_so_far); 460 461 // Avoid division by 0. 462 double time_until_full = MAX2(time_until_cms_gen_full(), 0.01); 463 double duty_cycle_dbl = 100.0 * scaled_duration_remaining / time_until_full; 464 465 unsigned int new_duty_cycle = MIN2((unsigned int)duty_cycle_dbl, 100U); 466 if (new_duty_cycle > _icms_duty_cycle) { 467 // Avoid very small duty cycles (1 or 2); 0 is allowed. 468 if (new_duty_cycle > 2) { 469 _icms_duty_cycle = icms_damped_duty_cycle(_icms_duty_cycle, 470 new_duty_cycle); 471 } 472 } else if (_allow_duty_cycle_reduction) { 473 // The duty cycle is reduced only once per cms cycle (see record_cms_end()). 474 new_duty_cycle = icms_damped_duty_cycle(_icms_duty_cycle, new_duty_cycle); 475 // Respect the minimum duty cycle. 476 unsigned int min_duty_cycle = (unsigned int)CMSIncrementalDutyCycleMin; 477 _icms_duty_cycle = MAX2(new_duty_cycle, min_duty_cycle); 478 } 479 480 if (PrintGCDetails || CMSTraceIncrementalPacing) { 481 gclog_or_tty->print(" icms_dc=%d ", _icms_duty_cycle); 482 } 483 484 _allow_duty_cycle_reduction = false; 485 return _icms_duty_cycle; 486 } 487 488 #ifndef PRODUCT 489 void CMSStats::print_on(outputStream *st) const { 490 st->print(" gc0_alpha=%d,cms_alpha=%d", _gc0_alpha, _cms_alpha); 491 st->print(",gc0_dur=%g,gc0_per=%g,gc0_promo=" SIZE_FORMAT, 492 gc0_duration(), gc0_period(), gc0_promoted()); 493 st->print(",cms_dur=%g,cms_dur_per_mb=%g,cms_per=%g,cms_alloc=" SIZE_FORMAT, 494 cms_duration(), cms_duration_per_mb(), 495 cms_period(), cms_allocated()); 496 st->print(",cms_since_beg=%g,cms_since_end=%g", 497 cms_time_since_begin(), cms_time_since_end()); 498 st->print(",cms_used_beg=" SIZE_FORMAT ",cms_used_end=" SIZE_FORMAT, 499 _cms_used_at_gc0_begin, _cms_used_at_gc0_end); 500 if (CMSIncrementalMode) { 501 st->print(",dc=%d", icms_duty_cycle()); 502 } 503 504 if (valid()) { 505 st->print(",promo_rate=%g,cms_alloc_rate=%g", 506 promotion_rate(), cms_allocation_rate()); 507 st->print(",cms_consumption_rate=%g,time_until_full=%g", 508 cms_consumption_rate(), time_until_cms_gen_full()); 509 } 510 st->print(" "); 511 } 512 #endif // #ifndef PRODUCT 513 514 CMSCollector::CollectorState CMSCollector::_collectorState = 515 CMSCollector::Idling; 516 bool CMSCollector::_foregroundGCIsActive = false; 517 bool CMSCollector::_foregroundGCShouldWait = false; 518 519 CMSCollector::CMSCollector(ConcurrentMarkSweepGeneration* cmsGen, 520 ConcurrentMarkSweepGeneration* permGen, 521 CardTableRS* ct, 522 ConcurrentMarkSweepPolicy* cp): 523 _cmsGen(cmsGen), 524 _permGen(permGen), 525 _ct(ct), 526 _ref_processor(NULL), // will be set later 527 _conc_workers(NULL), // may be set later 528 _abort_preclean(false), 529 _start_sampling(false), 530 _between_prologue_and_epilogue(false), 531 _markBitMap(0, Mutex::leaf + 1, "CMS_markBitMap_lock"), 532 _perm_gen_verify_bit_map(0, -1 /* no mutex */, "No_lock"), 533 _modUnionTable((CardTableModRefBS::card_shift - LogHeapWordSize), 534 -1 /* lock-free */, "No_lock" /* dummy */), 535 _modUnionClosure(&_modUnionTable), 536 _modUnionClosurePar(&_modUnionTable), 537 // Adjust my span to cover old (cms) gen and perm gen 538 _span(cmsGen->reserved()._union(permGen->reserved())), 539 // Construct the is_alive_closure with _span & markBitMap 540 _is_alive_closure(_span, &_markBitMap), 541 _restart_addr(NULL), 542 _overflow_list(NULL), 543 _preserved_oop_stack(NULL), 544 _preserved_mark_stack(NULL), 545 _stats(cmsGen), 546 _eden_chunk_array(NULL), // may be set in ctor body 547 _eden_chunk_capacity(0), // -- ditto -- 548 _eden_chunk_index(0), // -- ditto -- 549 _survivor_plab_array(NULL), // -- ditto -- 550 _survivor_chunk_array(NULL), // -- ditto -- 551 _survivor_chunk_capacity(0), // -- ditto -- 552 _survivor_chunk_index(0), // -- ditto -- 553 _ser_pmc_preclean_ovflw(0), 554 _ser_kac_preclean_ovflw(0), 555 _ser_pmc_remark_ovflw(0), 556 _par_pmc_remark_ovflw(0), 557 _ser_kac_ovflw(0), 558 _par_kac_ovflw(0), 559 #ifndef PRODUCT 560 _num_par_pushes(0), 561 #endif 562 _collection_count_start(0), 563 _verifying(false), 564 _icms_start_limit(NULL), 565 _icms_stop_limit(NULL), 566 _verification_mark_bm(0, Mutex::leaf + 1, "CMS_verification_mark_bm_lock"), 567 _completed_initialization(false), 568 _collector_policy(cp), 569 _should_unload_classes(false), 570 _concurrent_cycles_since_last_unload(0), 571 _roots_scanning_options(0), 572 _inter_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding), 573 _intra_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding) 574 { 575 if (ExplicitGCInvokesConcurrentAndUnloadsClasses) { 576 ExplicitGCInvokesConcurrent = true; 577 } 578 // Now expand the span and allocate the collection support structures 579 // (MUT, marking bit map etc.) to cover both generations subject to 580 // collection. 581 582 // First check that _permGen is adjacent to _cmsGen and above it. 583 assert( _cmsGen->reserved().word_size() > 0 584 && _permGen->reserved().word_size() > 0, 585 "generations should not be of zero size"); 586 assert(_cmsGen->reserved().intersection(_permGen->reserved()).is_empty(), 587 "_cmsGen and _permGen should not overlap"); 588 assert(_cmsGen->reserved().end() == _permGen->reserved().start(), 589 "_cmsGen->end() different from _permGen->start()"); 590 591 // For use by dirty card to oop closures. 592 _cmsGen->cmsSpace()->set_collector(this); 593 _permGen->cmsSpace()->set_collector(this); 594 595 // Allocate MUT and marking bit map 596 { 597 MutexLockerEx x(_markBitMap.lock(), Mutex::_no_safepoint_check_flag); 598 if (!_markBitMap.allocate(_span)) { 599 warning("Failed to allocate CMS Bit Map"); 600 return; 601 } 602 assert(_markBitMap.covers(_span), "_markBitMap inconsistency?"); 603 } 604 { 605 _modUnionTable.allocate(_span); 606 assert(_modUnionTable.covers(_span), "_modUnionTable inconsistency?"); 607 } 608 609 if (!_markStack.allocate(MarkStackSize)) { 610 warning("Failed to allocate CMS Marking Stack"); 611 return; 612 } 613 if (!_revisitStack.allocate(CMSRevisitStackSize)) { 614 warning("Failed to allocate CMS Revisit Stack"); 615 return; 616 } 617 618 // Support for multi-threaded concurrent phases 619 if (ParallelGCThreads > 0 && CMSConcurrentMTEnabled) { 620 if (FLAG_IS_DEFAULT(ConcGCThreads)) { 621 // just for now 622 FLAG_SET_DEFAULT(ConcGCThreads, (ParallelGCThreads + 3)/4); 623 } 624 if (ConcGCThreads > 1) { 625 _conc_workers = new YieldingFlexibleWorkGang("Parallel CMS Threads", 626 ConcGCThreads, true); 627 if (_conc_workers == NULL) { 628 warning("GC/CMS: _conc_workers allocation failure: " 629 "forcing -CMSConcurrentMTEnabled"); 630 CMSConcurrentMTEnabled = false; 631 } 632 } else { 633 CMSConcurrentMTEnabled = false; 634 } 635 } 636 if (!CMSConcurrentMTEnabled) { 637 ConcGCThreads = 0; 638 } else { 639 // Turn off CMSCleanOnEnter optimization temporarily for 640 // the MT case where it's not fixed yet; see 6178663. 641 CMSCleanOnEnter = false; 642 } 643 assert((_conc_workers != NULL) == (ConcGCThreads > 1), 644 "Inconsistency"); 645 646 // Parallel task queues; these are shared for the 647 // concurrent and stop-world phases of CMS, but 648 // are not shared with parallel scavenge (ParNew). 649 { 650 uint i; 651 uint num_queues = (uint) MAX2(ParallelGCThreads, ConcGCThreads); 652 653 if ((CMSParallelRemarkEnabled || CMSConcurrentMTEnabled 654 || ParallelRefProcEnabled) 655 && num_queues > 0) { 656 _task_queues = new OopTaskQueueSet(num_queues); 657 if (_task_queues == NULL) { 658 warning("task_queues allocation failure."); 659 return; 660 } 661 _hash_seed = NEW_C_HEAP_ARRAY(int, num_queues); 662 if (_hash_seed == NULL) { 663 warning("_hash_seed array allocation failure"); 664 return; 665 } 666 667 // XXX use a global constant instead of 64! 668 typedef struct OopTaskQueuePadded { 669 OopTaskQueue work_queue; 670 char pad[64 - sizeof(OopTaskQueue)]; // prevent false sharing 671 } OopTaskQueuePadded; 672 673 for (i = 0; i < num_queues; i++) { 674 OopTaskQueuePadded *q_padded = new OopTaskQueuePadded(); 675 if (q_padded == NULL) { 676 warning("work_queue allocation failure."); 677 return; 678 } 679 _task_queues->register_queue(i, &q_padded->work_queue); 680 } 681 for (i = 0; i < num_queues; i++) { 682 _task_queues->queue(i)->initialize(); 683 _hash_seed[i] = 17; // copied from ParNew 684 } 685 } 686 } 687 688 _cmsGen ->init_initiating_occupancy(CMSInitiatingOccupancyFraction, CMSTriggerRatio); 689 _permGen->init_initiating_occupancy(CMSInitiatingPermOccupancyFraction, CMSTriggerPermRatio); 690 691 // Clip CMSBootstrapOccupancy between 0 and 100. 692 _bootstrap_occupancy = ((double)MIN2((uintx)100, MAX2((uintx)0, CMSBootstrapOccupancy))) 693 /(double)100; 694 695 _full_gcs_since_conc_gc = 0; 696 697 // Now tell CMS generations the identity of their collector 698 ConcurrentMarkSweepGeneration::set_collector(this); 699 700 // Create & start a CMS thread for this CMS collector 701 _cmsThread = ConcurrentMarkSweepThread::start(this); 702 assert(cmsThread() != NULL, "CMS Thread should have been created"); 703 assert(cmsThread()->collector() == this, 704 "CMS Thread should refer to this gen"); 705 assert(CGC_lock != NULL, "Where's the CGC_lock?"); 706 707 // Support for parallelizing young gen rescan 708 GenCollectedHeap* gch = GenCollectedHeap::heap(); 709 _young_gen = gch->prev_gen(_cmsGen); 710 if (gch->supports_inline_contig_alloc()) { 711 _top_addr = gch->top_addr(); 712 _end_addr = gch->end_addr(); 713 assert(_young_gen != NULL, "no _young_gen"); 714 _eden_chunk_index = 0; 715 _eden_chunk_capacity = (_young_gen->max_capacity()+CMSSamplingGrain)/CMSSamplingGrain; 716 _eden_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, _eden_chunk_capacity); 717 if (_eden_chunk_array == NULL) { 718 _eden_chunk_capacity = 0; 719 warning("GC/CMS: _eden_chunk_array allocation failure"); 720 } 721 } 722 assert(_eden_chunk_array != NULL || _eden_chunk_capacity == 0, "Error"); 723 724 // Support for parallelizing survivor space rescan 725 if (CMSParallelRemarkEnabled && CMSParallelSurvivorRemarkEnabled) { 726 const size_t max_plab_samples = 727 ((DefNewGeneration*)_young_gen)->max_survivor_size()/MinTLABSize; 728 729 _survivor_plab_array = NEW_C_HEAP_ARRAY(ChunkArray, ParallelGCThreads); 730 _survivor_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, 2*max_plab_samples); 731 _cursor = NEW_C_HEAP_ARRAY(size_t, ParallelGCThreads); 732 if (_survivor_plab_array == NULL || _survivor_chunk_array == NULL 733 || _cursor == NULL) { 734 warning("Failed to allocate survivor plab/chunk array"); 735 if (_survivor_plab_array != NULL) { 736 FREE_C_HEAP_ARRAY(ChunkArray, _survivor_plab_array); 737 _survivor_plab_array = NULL; 738 } 739 if (_survivor_chunk_array != NULL) { 740 FREE_C_HEAP_ARRAY(HeapWord*, _survivor_chunk_array); 741 _survivor_chunk_array = NULL; 742 } 743 if (_cursor != NULL) { 744 FREE_C_HEAP_ARRAY(size_t, _cursor); 745 _cursor = NULL; 746 } 747 } else { 748 _survivor_chunk_capacity = 2*max_plab_samples; 749 for (uint i = 0; i < ParallelGCThreads; i++) { 750 HeapWord** vec = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples); 751 if (vec == NULL) { 752 warning("Failed to allocate survivor plab array"); 753 for (int j = i; j > 0; j--) { 754 FREE_C_HEAP_ARRAY(HeapWord*, _survivor_plab_array[j-1].array()); 755 } 756 FREE_C_HEAP_ARRAY(ChunkArray, _survivor_plab_array); 757 FREE_C_HEAP_ARRAY(HeapWord*, _survivor_chunk_array); 758 _survivor_plab_array = NULL; 759 _survivor_chunk_array = NULL; 760 _survivor_chunk_capacity = 0; 761 break; 762 } else { 763 ChunkArray* cur = 764 ::new (&_survivor_plab_array[i]) ChunkArray(vec, 765 max_plab_samples); 766 assert(cur->end() == 0, "Should be 0"); 767 assert(cur->array() == vec, "Should be vec"); 768 assert(cur->capacity() == max_plab_samples, "Error"); 769 } 770 } 771 } 772 } 773 assert( ( _survivor_plab_array != NULL 774 && _survivor_chunk_array != NULL) 775 || ( _survivor_chunk_capacity == 0 776 && _survivor_chunk_index == 0), 777 "Error"); 778 779 // Choose what strong roots should be scanned depending on verification options 780 // and perm gen collection mode. 781 if (!CMSClassUnloadingEnabled) { 782 // If class unloading is disabled we want to include all classes into the root set. 783 add_root_scanning_option(SharedHeap::SO_AllClasses); 784 } else { 785 add_root_scanning_option(SharedHeap::SO_SystemClasses); 786 } 787 788 NOT_PRODUCT(_overflow_counter = CMSMarkStackOverflowInterval;) 789 _gc_counters = new CollectorCounters("CMS", 1); 790 _completed_initialization = true; 791 _inter_sweep_timer.start(); // start of time 792 #ifdef SPARC 793 // Issue a stern warning, but allow use for experimentation and debugging. 794 if (VM_Version::is_sun4v() && UseMemSetInBOT) { 795 assert(!FLAG_IS_DEFAULT(UseMemSetInBOT), "Error"); 796 warning("Experimental flag -XX:+UseMemSetInBOT is known to cause instability" 797 " on sun4v; please understand that you are using at your own risk!"); 798 } 799 #endif 800 } 801 802 const char* ConcurrentMarkSweepGeneration::name() const { 803 return "concurrent mark-sweep generation"; 804 } 805 void ConcurrentMarkSweepGeneration::update_counters() { 806 if (UsePerfData) { 807 _space_counters->update_all(); 808 _gen_counters->update_all(); 809 } 810 } 811 812 // this is an optimized version of update_counters(). it takes the 813 // used value as a parameter rather than computing it. 814 // 815 void ConcurrentMarkSweepGeneration::update_counters(size_t used) { 816 if (UsePerfData) { 817 _space_counters->update_used(used); 818 _space_counters->update_capacity(); 819 _gen_counters->update_all(); 820 } 821 } 822 823 void ConcurrentMarkSweepGeneration::print() const { 824 Generation::print(); 825 cmsSpace()->print(); 826 } 827 828 #ifndef PRODUCT 829 void ConcurrentMarkSweepGeneration::print_statistics() { 830 cmsSpace()->printFLCensus(0); 831 } 832 #endif 833 834 void ConcurrentMarkSweepGeneration::printOccupancy(const char *s) { 835 GenCollectedHeap* gch = GenCollectedHeap::heap(); 836 if (PrintGCDetails) { 837 if (Verbose) { 838 gclog_or_tty->print(" [%d %s-%s: "SIZE_FORMAT"("SIZE_FORMAT")]", 839 level(), short_name(), s, used(), capacity()); 840 } else { 841 gclog_or_tty->print(" [%d %s-%s: "SIZE_FORMAT"K("SIZE_FORMAT"K)]", 842 level(), short_name(), s, used() / K, capacity() / K); 843 } 844 } 845 if (Verbose) { 846 gclog_or_tty->print(" "SIZE_FORMAT"("SIZE_FORMAT")", 847 gch->used(), gch->capacity()); 848 } else { 849 gclog_or_tty->print(" "SIZE_FORMAT"K("SIZE_FORMAT"K)", 850 gch->used() / K, gch->capacity() / K); 851 } 852 } 853 854 size_t 855 ConcurrentMarkSweepGeneration::contiguous_available() const { 856 // dld proposes an improvement in precision here. If the committed 857 // part of the space ends in a free block we should add that to 858 // uncommitted size in the calculation below. Will make this 859 // change later, staying with the approximation below for the 860 // time being. -- ysr. 861 return MAX2(_virtual_space.uncommitted_size(), unsafe_max_alloc_nogc()); 862 } 863 864 size_t 865 ConcurrentMarkSweepGeneration::unsafe_max_alloc_nogc() const { 866 return _cmsSpace->max_alloc_in_words() * HeapWordSize; 867 } 868 869 size_t ConcurrentMarkSweepGeneration::max_available() const { 870 return free() + _virtual_space.uncommitted_size(); 871 } 872 873 bool ConcurrentMarkSweepGeneration::promotion_attempt_is_safe( 874 size_t max_promotion_in_bytes, 875 bool younger_handles_promotion_failure) const { 876 877 // This is the most conservative test. Full promotion is 878 // guaranteed if this is used. The multiplicative factor is to 879 // account for the worst case "dilatation". 880 double adjusted_max_promo_bytes = _dilatation_factor * max_promotion_in_bytes; 881 if (adjusted_max_promo_bytes > (double)max_uintx) { // larger than size_t 882 adjusted_max_promo_bytes = (double)max_uintx; 883 } 884 bool result = (max_contiguous_available() >= (size_t)adjusted_max_promo_bytes); 885 886 if (younger_handles_promotion_failure && !result) { 887 // Full promotion is not guaranteed because fragmentation 888 // of the cms generation can prevent the full promotion. 889 result = (max_available() >= (size_t)adjusted_max_promo_bytes); 890 891 if (!result) { 892 // With promotion failure handling the test for the ability 893 // to support the promotion does not have to be guaranteed. 894 // Use an average of the amount promoted. 895 result = max_available() >= (size_t) 896 gc_stats()->avg_promoted()->padded_average(); 897 if (PrintGC && Verbose && result) { 898 gclog_or_tty->print_cr( 899 "\nConcurrentMarkSweepGeneration::promotion_attempt_is_safe" 900 " max_available: " SIZE_FORMAT 901 " avg_promoted: " SIZE_FORMAT, 902 max_available(), (size_t) 903 gc_stats()->avg_promoted()->padded_average()); 904 } 905 } else { 906 if (PrintGC && Verbose) { 907 gclog_or_tty->print_cr( 908 "\nConcurrentMarkSweepGeneration::promotion_attempt_is_safe" 909 " max_available: " SIZE_FORMAT 910 " adj_max_promo_bytes: " SIZE_FORMAT, 911 max_available(), (size_t)adjusted_max_promo_bytes); 912 } 913 } 914 } else { 915 if (PrintGC && Verbose) { 916 gclog_or_tty->print_cr( 917 "\nConcurrentMarkSweepGeneration::promotion_attempt_is_safe" 918 " contiguous_available: " SIZE_FORMAT 919 " adj_max_promo_bytes: " SIZE_FORMAT, 920 max_contiguous_available(), (size_t)adjusted_max_promo_bytes); 921 } 922 } 923 return result; 924 } 925 926 // At a promotion failure dump information on block layout in heap 927 // (cms old generation). 928 void ConcurrentMarkSweepGeneration::promotion_failure_occurred() { 929 if (CMSDumpAtPromotionFailure) { 930 cmsSpace()->dump_at_safepoint_with_locks(collector(), gclog_or_tty); 931 } 932 } 933 934 CompactibleSpace* 935 ConcurrentMarkSweepGeneration::first_compaction_space() const { 936 return _cmsSpace; 937 } 938 939 void ConcurrentMarkSweepGeneration::reset_after_compaction() { 940 // Clear the promotion information. These pointers can be adjusted 941 // along with all the other pointers into the heap but 942 // compaction is expected to be a rare event with 943 // a heap using cms so don't do it without seeing the need. 944 if (ParallelGCThreads > 0) { 945 for (uint i = 0; i < ParallelGCThreads; i++) { 946 _par_gc_thread_states[i]->promo.reset(); 947 } 948 } 949 } 950 951 void ConcurrentMarkSweepGeneration::space_iterate(SpaceClosure* blk, bool usedOnly) { 952 blk->do_space(_cmsSpace); 953 } 954 955 void ConcurrentMarkSweepGeneration::compute_new_size() { 956 assert_locked_or_safepoint(Heap_lock); 957 958 // If incremental collection failed, we just want to expand 959 // to the limit. 960 if (incremental_collection_failed()) { 961 clear_incremental_collection_failed(); 962 grow_to_reserved(); 963 return; 964 } 965 966 size_t expand_bytes = 0; 967 double free_percentage = ((double) free()) / capacity(); 968 double desired_free_percentage = (double) MinHeapFreeRatio / 100; 969 double maximum_free_percentage = (double) MaxHeapFreeRatio / 100; 970 971 // compute expansion delta needed for reaching desired free percentage 972 if (free_percentage < desired_free_percentage) { 973 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 974 assert(desired_capacity >= capacity(), "invalid expansion size"); 975 expand_bytes = MAX2(desired_capacity - capacity(), MinHeapDeltaBytes); 976 } 977 if (expand_bytes > 0) { 978 if (PrintGCDetails && Verbose) { 979 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 980 gclog_or_tty->print_cr("\nFrom compute_new_size: "); 981 gclog_or_tty->print_cr(" Free fraction %f", free_percentage); 982 gclog_or_tty->print_cr(" Desired free fraction %f", 983 desired_free_percentage); 984 gclog_or_tty->print_cr(" Maximum free fraction %f", 985 maximum_free_percentage); 986 gclog_or_tty->print_cr(" Capactiy "SIZE_FORMAT, capacity()/1000); 987 gclog_or_tty->print_cr(" Desired capacity "SIZE_FORMAT, 988 desired_capacity/1000); 989 int prev_level = level() - 1; 990 if (prev_level >= 0) { 991 size_t prev_size = 0; 992 GenCollectedHeap* gch = GenCollectedHeap::heap(); 993 Generation* prev_gen = gch->_gens[prev_level]; 994 prev_size = prev_gen->capacity(); 995 gclog_or_tty->print_cr(" Younger gen size "SIZE_FORMAT, 996 prev_size/1000); 997 } 998 gclog_or_tty->print_cr(" unsafe_max_alloc_nogc "SIZE_FORMAT, 999 unsafe_max_alloc_nogc()/1000); 1000 gclog_or_tty->print_cr(" contiguous available "SIZE_FORMAT, 1001 contiguous_available()/1000); 1002 gclog_or_tty->print_cr(" Expand by "SIZE_FORMAT" (bytes)", 1003 expand_bytes); 1004 } 1005 // safe if expansion fails 1006 expand(expand_bytes, 0, CMSExpansionCause::_satisfy_free_ratio); 1007 if (PrintGCDetails && Verbose) { 1008 gclog_or_tty->print_cr(" Expanded free fraction %f", 1009 ((double) free()) / capacity()); 1010 } 1011 } 1012 } 1013 1014 Mutex* ConcurrentMarkSweepGeneration::freelistLock() const { 1015 return cmsSpace()->freelistLock(); 1016 } 1017 1018 HeapWord* ConcurrentMarkSweepGeneration::allocate(size_t size, 1019 bool tlab) { 1020 CMSSynchronousYieldRequest yr; 1021 MutexLockerEx x(freelistLock(), 1022 Mutex::_no_safepoint_check_flag); 1023 return have_lock_and_allocate(size, tlab); 1024 } 1025 1026 HeapWord* ConcurrentMarkSweepGeneration::have_lock_and_allocate(size_t size, 1027 bool tlab) { 1028 assert_lock_strong(freelistLock()); 1029 size_t adjustedSize = CompactibleFreeListSpace::adjustObjectSize(size); 1030 HeapWord* res = cmsSpace()->allocate(adjustedSize); 1031 // Allocate the object live (grey) if the background collector has 1032 // started marking. This is necessary because the marker may 1033 // have passed this address and consequently this object will 1034 // not otherwise be greyed and would be incorrectly swept up. 1035 // Note that if this object contains references, the writing 1036 // of those references will dirty the card containing this object 1037 // allowing the object to be blackened (and its references scanned) 1038 // either during a preclean phase or at the final checkpoint. 1039 if (res != NULL) { 1040 collector()->direct_allocated(res, adjustedSize); 1041 _direct_allocated_words += adjustedSize; 1042 // allocation counters 1043 NOT_PRODUCT( 1044 _numObjectsAllocated++; 1045 _numWordsAllocated += (int)adjustedSize; 1046 ) 1047 } 1048 return res; 1049 } 1050 1051 // In the case of direct allocation by mutators in a generation that 1052 // is being concurrently collected, the object must be allocated 1053 // live (grey) if the background collector has started marking. 1054 // This is necessary because the marker may 1055 // have passed this address and consequently this object will 1056 // not otherwise be greyed and would be incorrectly swept up. 1057 // Note that if this object contains references, the writing 1058 // of those references will dirty the card containing this object 1059 // allowing the object to be blackened (and its references scanned) 1060 // either during a preclean phase or at the final checkpoint. 1061 void CMSCollector::direct_allocated(HeapWord* start, size_t size) { 1062 assert(_markBitMap.covers(start, size), "Out of bounds"); 1063 if (_collectorState >= Marking) { 1064 MutexLockerEx y(_markBitMap.lock(), 1065 Mutex::_no_safepoint_check_flag); 1066 // [see comments preceding SweepClosure::do_blk() below for details] 1067 // 1. need to mark the object as live so it isn't collected 1068 // 2. need to mark the 2nd bit to indicate the object may be uninitialized 1069 // 3. need to mark the end of the object so sweeper can skip over it 1070 // if it's uninitialized when the sweeper reaches it. 1071 _markBitMap.mark(start); // object is live 1072 _markBitMap.mark(start + 1); // object is potentially uninitialized? 1073 _markBitMap.mark(start + size - 1); 1074 // mark end of object 1075 } 1076 // check that oop looks uninitialized 1077 assert(oop(start)->klass_or_null() == NULL, "_klass should be NULL"); 1078 } 1079 1080 void CMSCollector::promoted(bool par, HeapWord* start, 1081 bool is_obj_array, size_t obj_size) { 1082 assert(_markBitMap.covers(start), "Out of bounds"); 1083 // See comment in direct_allocated() about when objects should 1084 // be allocated live. 1085 if (_collectorState >= Marking) { 1086 // we already hold the marking bit map lock, taken in 1087 // the prologue 1088 if (par) { 1089 _markBitMap.par_mark(start); 1090 } else { 1091 _markBitMap.mark(start); 1092 } 1093 // We don't need to mark the object as uninitialized (as 1094 // in direct_allocated above) because this is being done with the 1095 // world stopped and the object will be initialized by the 1096 // time the sweeper gets to look at it. 1097 assert(SafepointSynchronize::is_at_safepoint(), 1098 "expect promotion only at safepoints"); 1099 1100 if (_collectorState < Sweeping) { 1101 // Mark the appropriate cards in the modUnionTable, so that 1102 // this object gets scanned before the sweep. If this is 1103 // not done, CMS generation references in the object might 1104 // not get marked. 1105 // For the case of arrays, which are otherwise precisely 1106 // marked, we need to dirty the entire array, not just its head. 1107 if (is_obj_array) { 1108 // The [par_]mark_range() method expects mr.end() below to 1109 // be aligned to the granularity of a bit's representation 1110 // in the heap. In the case of the MUT below, that's a 1111 // card size. 1112 MemRegion mr(start, 1113 (HeapWord*)round_to((intptr_t)(start + obj_size), 1114 CardTableModRefBS::card_size /* bytes */)); 1115 if (par) { 1116 _modUnionTable.par_mark_range(mr); 1117 } else { 1118 _modUnionTable.mark_range(mr); 1119 } 1120 } else { // not an obj array; we can just mark the head 1121 if (par) { 1122 _modUnionTable.par_mark(start); 1123 } else { 1124 _modUnionTable.mark(start); 1125 } 1126 } 1127 } 1128 } 1129 } 1130 1131 static inline size_t percent_of_space(Space* space, HeapWord* addr) 1132 { 1133 size_t delta = pointer_delta(addr, space->bottom()); 1134 return (size_t)(delta * 100.0 / (space->capacity() / HeapWordSize)); 1135 } 1136 1137 void CMSCollector::icms_update_allocation_limits() 1138 { 1139 Generation* gen0 = GenCollectedHeap::heap()->get_gen(0); 1140 EdenSpace* eden = gen0->as_DefNewGeneration()->eden(); 1141 1142 const unsigned int duty_cycle = stats().icms_update_duty_cycle(); 1143 if (CMSTraceIncrementalPacing) { 1144 stats().print(); 1145 } 1146 1147 assert(duty_cycle <= 100, "invalid duty cycle"); 1148 if (duty_cycle != 0) { 1149 // The duty_cycle is a percentage between 0 and 100; convert to words and 1150 // then compute the offset from the endpoints of the space. 1151 size_t free_words = eden->free() / HeapWordSize; 1152 double free_words_dbl = (double)free_words; 1153 size_t duty_cycle_words = (size_t)(free_words_dbl * duty_cycle / 100.0); 1154 size_t offset_words = (free_words - duty_cycle_words) / 2; 1155 1156 _icms_start_limit = eden->top() + offset_words; 1157 _icms_stop_limit = eden->end() - offset_words; 1158 1159 // The limits may be adjusted (shifted to the right) by 1160 // CMSIncrementalOffset, to allow the application more mutator time after a 1161 // young gen gc (when all mutators were stopped) and before CMS starts and 1162 // takes away one or more cpus. 1163 if (CMSIncrementalOffset != 0) { 1164 double adjustment_dbl = free_words_dbl * CMSIncrementalOffset / 100.0; 1165 size_t adjustment = (size_t)adjustment_dbl; 1166 HeapWord* tmp_stop = _icms_stop_limit + adjustment; 1167 if (tmp_stop > _icms_stop_limit && tmp_stop < eden->end()) { 1168 _icms_start_limit += adjustment; 1169 _icms_stop_limit = tmp_stop; 1170 } 1171 } 1172 } 1173 if (duty_cycle == 0 || (_icms_start_limit == _icms_stop_limit)) { 1174 _icms_start_limit = _icms_stop_limit = eden->end(); 1175 } 1176 1177 // Install the new start limit. 1178 eden->set_soft_end(_icms_start_limit); 1179 1180 if (CMSTraceIncrementalMode) { 1181 gclog_or_tty->print(" icms alloc limits: " 1182 PTR_FORMAT "," PTR_FORMAT 1183 " (" SIZE_FORMAT "%%," SIZE_FORMAT "%%) ", 1184 _icms_start_limit, _icms_stop_limit, 1185 percent_of_space(eden, _icms_start_limit), 1186 percent_of_space(eden, _icms_stop_limit)); 1187 if (Verbose) { 1188 gclog_or_tty->print("eden: "); 1189 eden->print_on(gclog_or_tty); 1190 } 1191 } 1192 } 1193 1194 // Any changes here should try to maintain the invariant 1195 // that if this method is called with _icms_start_limit 1196 // and _icms_stop_limit both NULL, then it should return NULL 1197 // and not notify the icms thread. 1198 HeapWord* 1199 CMSCollector::allocation_limit_reached(Space* space, HeapWord* top, 1200 size_t word_size) 1201 { 1202 // A start_limit equal to end() means the duty cycle is 0, so treat that as a 1203 // nop. 1204 if (CMSIncrementalMode && _icms_start_limit != space->end()) { 1205 if (top <= _icms_start_limit) { 1206 if (CMSTraceIncrementalMode) { 1207 space->print_on(gclog_or_tty); 1208 gclog_or_tty->stamp(); 1209 gclog_or_tty->print_cr(" start limit top=" PTR_FORMAT 1210 ", new limit=" PTR_FORMAT 1211 " (" SIZE_FORMAT "%%)", 1212 top, _icms_stop_limit, 1213 percent_of_space(space, _icms_stop_limit)); 1214 } 1215 ConcurrentMarkSweepThread::start_icms(); 1216 assert(top < _icms_stop_limit, "Tautology"); 1217 if (word_size < pointer_delta(_icms_stop_limit, top)) { 1218 return _icms_stop_limit; 1219 } 1220 1221 // The allocation will cross both the _start and _stop limits, so do the 1222 // stop notification also and return end(). 1223 if (CMSTraceIncrementalMode) { 1224 space->print_on(gclog_or_tty); 1225 gclog_or_tty->stamp(); 1226 gclog_or_tty->print_cr(" +stop limit top=" PTR_FORMAT 1227 ", new limit=" PTR_FORMAT 1228 " (" SIZE_FORMAT "%%)", 1229 top, space->end(), 1230 percent_of_space(space, space->end())); 1231 } 1232 ConcurrentMarkSweepThread::stop_icms(); 1233 return space->end(); 1234 } 1235 1236 if (top <= _icms_stop_limit) { 1237 if (CMSTraceIncrementalMode) { 1238 space->print_on(gclog_or_tty); 1239 gclog_or_tty->stamp(); 1240 gclog_or_tty->print_cr(" stop limit top=" PTR_FORMAT 1241 ", new limit=" PTR_FORMAT 1242 " (" SIZE_FORMAT "%%)", 1243 top, space->end(), 1244 percent_of_space(space, space->end())); 1245 } 1246 ConcurrentMarkSweepThread::stop_icms(); 1247 return space->end(); 1248 } 1249 1250 if (CMSTraceIncrementalMode) { 1251 space->print_on(gclog_or_tty); 1252 gclog_or_tty->stamp(); 1253 gclog_or_tty->print_cr(" end limit top=" PTR_FORMAT 1254 ", new limit=" PTR_FORMAT, 1255 top, NULL); 1256 } 1257 } 1258 1259 return NULL; 1260 } 1261 1262 oop ConcurrentMarkSweepGeneration::promote(oop obj, size_t obj_size) { 1263 assert(obj_size == (size_t)obj->size(), "bad obj_size passed in"); 1264 // allocate, copy and if necessary update promoinfo -- 1265 // delegate to underlying space. 1266 assert_lock_strong(freelistLock()); 1267 1268 #ifndef PRODUCT 1269 if (Universe::heap()->promotion_should_fail()) { 1270 return NULL; 1271 } 1272 #endif // #ifndef PRODUCT 1273 1274 oop res = _cmsSpace->promote(obj, obj_size); 1275 if (res == NULL) { 1276 // expand and retry 1277 size_t s = _cmsSpace->expansionSpaceRequired(obj_size); // HeapWords 1278 expand(s*HeapWordSize, MinHeapDeltaBytes, 1279 CMSExpansionCause::_satisfy_promotion); 1280 // Since there's currently no next generation, we don't try to promote 1281 // into a more senior generation. 1282 assert(next_gen() == NULL, "assumption, based upon which no attempt " 1283 "is made to pass on a possibly failing " 1284 "promotion to next generation"); 1285 res = _cmsSpace->promote(obj, obj_size); 1286 } 1287 if (res != NULL) { 1288 // See comment in allocate() about when objects should 1289 // be allocated live. 1290 assert(obj->is_oop(), "Will dereference klass pointer below"); 1291 collector()->promoted(false, // Not parallel 1292 (HeapWord*)res, obj->is_objArray(), obj_size); 1293 // promotion counters 1294 NOT_PRODUCT( 1295 _numObjectsPromoted++; 1296 _numWordsPromoted += 1297 (int)(CompactibleFreeListSpace::adjustObjectSize(obj->size())); 1298 ) 1299 } 1300 return res; 1301 } 1302 1303 1304 HeapWord* 1305 ConcurrentMarkSweepGeneration::allocation_limit_reached(Space* space, 1306 HeapWord* top, 1307 size_t word_sz) 1308 { 1309 return collector()->allocation_limit_reached(space, top, word_sz); 1310 } 1311 1312 // Things to support parallel young-gen collection. 1313 oop 1314 ConcurrentMarkSweepGeneration::par_promote(int thread_num, 1315 oop old, markOop m, 1316 size_t word_sz) { 1317 #ifndef PRODUCT 1318 if (Universe::heap()->promotion_should_fail()) { 1319 return NULL; 1320 } 1321 #endif // #ifndef PRODUCT 1322 1323 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1324 PromotionInfo* promoInfo = &ps->promo; 1325 // if we are tracking promotions, then first ensure space for 1326 // promotion (including spooling space for saving header if necessary). 1327 // then allocate and copy, then track promoted info if needed. 1328 // When tracking (see PromotionInfo::track()), the mark word may 1329 // be displaced and in this case restoration of the mark word 1330 // occurs in the (oop_since_save_marks_)iterate phase. 1331 if (promoInfo->tracking() && !promoInfo->ensure_spooling_space()) { 1332 // Out of space for allocating spooling buffers; 1333 // try expanding and allocating spooling buffers. 1334 if (!expand_and_ensure_spooling_space(promoInfo)) { 1335 return NULL; 1336 } 1337 } 1338 assert(promoInfo->has_spooling_space(), "Control point invariant"); 1339 HeapWord* obj_ptr = ps->lab.alloc(word_sz); 1340 if (obj_ptr == NULL) { 1341 obj_ptr = expand_and_par_lab_allocate(ps, word_sz); 1342 if (obj_ptr == NULL) { 1343 return NULL; 1344 } 1345 } 1346 oop obj = oop(obj_ptr); 1347 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1348 // Otherwise, copy the object. Here we must be careful to insert the 1349 // klass pointer last, since this marks the block as an allocated object. 1350 // Except with compressed oops it's the mark word. 1351 HeapWord* old_ptr = (HeapWord*)old; 1352 if (word_sz > (size_t)oopDesc::header_size()) { 1353 Copy::aligned_disjoint_words(old_ptr + oopDesc::header_size(), 1354 obj_ptr + oopDesc::header_size(), 1355 word_sz - oopDesc::header_size()); 1356 } 1357 1358 if (UseCompressedOops) { 1359 // Copy gap missed by (aligned) header size calculation above 1360 obj->set_klass_gap(old->klass_gap()); 1361 } 1362 1363 // Restore the mark word copied above. 1364 obj->set_mark(m); 1365 1366 // Now we can track the promoted object, if necessary. We take care 1367 // to delay the transition from uninitialized to full object 1368 // (i.e., insertion of klass pointer) until after, so that it 1369 // atomically becomes a promoted object. 1370 if (promoInfo->tracking()) { 1371 promoInfo->track((PromotedObject*)obj, old->klass()); 1372 } 1373 1374 // Finally, install the klass pointer (this should be volatile). 1375 obj->set_klass(old->klass()); 1376 1377 assert(old->is_oop(), "Will dereference klass ptr below"); 1378 collector()->promoted(true, // parallel 1379 obj_ptr, old->is_objArray(), word_sz); 1380 1381 NOT_PRODUCT( 1382 Atomic::inc(&_numObjectsPromoted); 1383 Atomic::add((jint)CompactibleFreeListSpace::adjustObjectSize(obj->size()), 1384 &_numWordsPromoted); 1385 ) 1386 1387 return obj; 1388 } 1389 1390 void 1391 ConcurrentMarkSweepGeneration:: 1392 par_promote_alloc_undo(int thread_num, 1393 HeapWord* obj, size_t word_sz) { 1394 // CMS does not support promotion undo. 1395 ShouldNotReachHere(); 1396 } 1397 1398 void 1399 ConcurrentMarkSweepGeneration:: 1400 par_promote_alloc_done(int thread_num) { 1401 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1402 ps->lab.retire(thread_num); 1403 } 1404 1405 void 1406 ConcurrentMarkSweepGeneration:: 1407 par_oop_since_save_marks_iterate_done(int thread_num) { 1408 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1409 ParScanWithoutBarrierClosure* dummy_cl = NULL; 1410 ps->promo.promoted_oops_iterate_nv(dummy_cl); 1411 } 1412 1413 // XXXPERM 1414 bool ConcurrentMarkSweepGeneration::should_collect(bool full, 1415 size_t size, 1416 bool tlab) 1417 { 1418 // We allow a STW collection only if a full 1419 // collection was requested. 1420 return full || should_allocate(size, tlab); // FIX ME !!! 1421 // This and promotion failure handling are connected at the 1422 // hip and should be fixed by untying them. 1423 } 1424 1425 bool CMSCollector::shouldConcurrentCollect() { 1426 if (_full_gc_requested) { 1427 if (Verbose && PrintGCDetails) { 1428 gclog_or_tty->print_cr("CMSCollector: collect because of explicit " 1429 " gc request (or gc_locker)"); 1430 } 1431 return true; 1432 } 1433 1434 // For debugging purposes, change the type of collection. 1435 // If the rotation is not on the concurrent collection 1436 // type, don't start a concurrent collection. 1437 NOT_PRODUCT( 1438 if (RotateCMSCollectionTypes && 1439 (_cmsGen->debug_collection_type() != 1440 ConcurrentMarkSweepGeneration::Concurrent_collection_type)) { 1441 assert(_cmsGen->debug_collection_type() != 1442 ConcurrentMarkSweepGeneration::Unknown_collection_type, 1443 "Bad cms collection type"); 1444 return false; 1445 } 1446 ) 1447 1448 FreelistLocker x(this); 1449 // ------------------------------------------------------------------ 1450 // Print out lots of information which affects the initiation of 1451 // a collection. 1452 if (PrintCMSInitiationStatistics && stats().valid()) { 1453 gclog_or_tty->print("CMSCollector shouldConcurrentCollect: "); 1454 gclog_or_tty->stamp(); 1455 gclog_or_tty->print_cr(""); 1456 stats().print_on(gclog_or_tty); 1457 gclog_or_tty->print_cr("time_until_cms_gen_full %3.7f", 1458 stats().time_until_cms_gen_full()); 1459 gclog_or_tty->print_cr("free="SIZE_FORMAT, _cmsGen->free()); 1460 gclog_or_tty->print_cr("contiguous_available="SIZE_FORMAT, 1461 _cmsGen->contiguous_available()); 1462 gclog_or_tty->print_cr("promotion_rate=%g", stats().promotion_rate()); 1463 gclog_or_tty->print_cr("cms_allocation_rate=%g", stats().cms_allocation_rate()); 1464 gclog_or_tty->print_cr("occupancy=%3.7f", _cmsGen->occupancy()); 1465 gclog_or_tty->print_cr("initiatingOccupancy=%3.7f", _cmsGen->initiating_occupancy()); 1466 gclog_or_tty->print_cr("initiatingPermOccupancy=%3.7f", _permGen->initiating_occupancy()); 1467 } 1468 // ------------------------------------------------------------------ 1469 1470 // If the estimated time to complete a cms collection (cms_duration()) 1471 // is less than the estimated time remaining until the cms generation 1472 // is full, start a collection. 1473 if (!UseCMSInitiatingOccupancyOnly) { 1474 if (stats().valid()) { 1475 if (stats().time_until_cms_start() == 0.0) { 1476 return true; 1477 } 1478 } else { 1479 // We want to conservatively collect somewhat early in order 1480 // to try and "bootstrap" our CMS/promotion statistics; 1481 // this branch will not fire after the first successful CMS 1482 // collection because the stats should then be valid. 1483 if (_cmsGen->occupancy() >= _bootstrap_occupancy) { 1484 if (Verbose && PrintGCDetails) { 1485 gclog_or_tty->print_cr( 1486 " CMSCollector: collect for bootstrapping statistics:" 1487 " occupancy = %f, boot occupancy = %f", _cmsGen->occupancy(), 1488 _bootstrap_occupancy); 1489 } 1490 return true; 1491 } 1492 } 1493 } 1494 1495 // Otherwise, we start a collection cycle if either the perm gen or 1496 // old gen want a collection cycle started. Each may use 1497 // an appropriate criterion for making this decision. 1498 // XXX We need to make sure that the gen expansion 1499 // criterion dovetails well with this. XXX NEED TO FIX THIS 1500 if (_cmsGen->should_concurrent_collect()) { 1501 if (Verbose && PrintGCDetails) { 1502 gclog_or_tty->print_cr("CMS old gen initiated"); 1503 } 1504 return true; 1505 } 1506 1507 // We start a collection if we believe an incremental collection may fail; 1508 // this is not likely to be productive in practice because it's probably too 1509 // late anyway. 1510 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1511 assert(gch->collector_policy()->is_two_generation_policy(), 1512 "You may want to check the correctness of the following"); 1513 if (gch->incremental_collection_will_fail()) { 1514 if (PrintGCDetails && Verbose) { 1515 gclog_or_tty->print("CMSCollector: collect because incremental collection will fail "); 1516 } 1517 return true; 1518 } 1519 1520 if (CMSClassUnloadingEnabled && _permGen->should_concurrent_collect()) { 1521 bool res = update_should_unload_classes(); 1522 if (res) { 1523 if (Verbose && PrintGCDetails) { 1524 gclog_or_tty->print_cr("CMS perm gen initiated"); 1525 } 1526 return true; 1527 } 1528 } 1529 return false; 1530 } 1531 1532 // Clear _expansion_cause fields of constituent generations 1533 void CMSCollector::clear_expansion_cause() { 1534 _cmsGen->clear_expansion_cause(); 1535 _permGen->clear_expansion_cause(); 1536 } 1537 1538 // We should be conservative in starting a collection cycle. To 1539 // start too eagerly runs the risk of collecting too often in the 1540 // extreme. To collect too rarely falls back on full collections, 1541 // which works, even if not optimum in terms of concurrent work. 1542 // As a work around for too eagerly collecting, use the flag 1543 // UseCMSInitiatingOccupancyOnly. This also has the advantage of 1544 // giving the user an easily understandable way of controlling the 1545 // collections. 1546 // We want to start a new collection cycle if any of the following 1547 // conditions hold: 1548 // . our current occupancy exceeds the configured initiating occupancy 1549 // for this generation, or 1550 // . we recently needed to expand this space and have not, since that 1551 // expansion, done a collection of this generation, or 1552 // . the underlying space believes that it may be a good idea to initiate 1553 // a concurrent collection (this may be based on criteria such as the 1554 // following: the space uses linear allocation and linear allocation is 1555 // going to fail, or there is believed to be excessive fragmentation in 1556 // the generation, etc... or ... 1557 // [.(currently done by CMSCollector::shouldConcurrentCollect() only for 1558 // the case of the old generation, not the perm generation; see CR 6543076): 1559 // we may be approaching a point at which allocation requests may fail because 1560 // we will be out of sufficient free space given allocation rate estimates.] 1561 bool ConcurrentMarkSweepGeneration::should_concurrent_collect() const { 1562 1563 assert_lock_strong(freelistLock()); 1564 if (occupancy() > initiating_occupancy()) { 1565 if (PrintGCDetails && Verbose) { 1566 gclog_or_tty->print(" %s: collect because of occupancy %f / %f ", 1567 short_name(), occupancy(), initiating_occupancy()); 1568 } 1569 return true; 1570 } 1571 if (UseCMSInitiatingOccupancyOnly) { 1572 return false; 1573 } 1574 if (expansion_cause() == CMSExpansionCause::_satisfy_allocation) { 1575 if (PrintGCDetails && Verbose) { 1576 gclog_or_tty->print(" %s: collect because expanded for allocation ", 1577 short_name()); 1578 } 1579 return true; 1580 } 1581 if (_cmsSpace->should_concurrent_collect()) { 1582 if (PrintGCDetails && Verbose) { 1583 gclog_or_tty->print(" %s: collect because cmsSpace says so ", 1584 short_name()); 1585 } 1586 return true; 1587 } 1588 return false; 1589 } 1590 1591 void ConcurrentMarkSweepGeneration::collect(bool full, 1592 bool clear_all_soft_refs, 1593 size_t size, 1594 bool tlab) 1595 { 1596 collector()->collect(full, clear_all_soft_refs, size, tlab); 1597 } 1598 1599 void CMSCollector::collect(bool full, 1600 bool clear_all_soft_refs, 1601 size_t size, 1602 bool tlab) 1603 { 1604 if (!UseCMSCollectionPassing && _collectorState > Idling) { 1605 // For debugging purposes skip the collection if the state 1606 // is not currently idle 1607 if (TraceCMSState) { 1608 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " skipped full:%d CMS state %d", 1609 Thread::current(), full, _collectorState); 1610 } 1611 return; 1612 } 1613 1614 // The following "if" branch is present for defensive reasons. 1615 // In the current uses of this interface, it can be replaced with: 1616 // assert(!GC_locker.is_active(), "Can't be called otherwise"); 1617 // But I am not placing that assert here to allow future 1618 // generality in invoking this interface. 1619 if (GC_locker::is_active()) { 1620 // A consistency test for GC_locker 1621 assert(GC_locker::needs_gc(), "Should have been set already"); 1622 // Skip this foreground collection, instead 1623 // expanding the heap if necessary. 1624 // Need the free list locks for the call to free() in compute_new_size() 1625 compute_new_size(); 1626 return; 1627 } 1628 acquire_control_and_collect(full, clear_all_soft_refs); 1629 _full_gcs_since_conc_gc++; 1630 1631 } 1632 1633 void CMSCollector::request_full_gc(unsigned int full_gc_count) { 1634 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1635 unsigned int gc_count = gch->total_full_collections(); 1636 if (gc_count == full_gc_count) { 1637 MutexLockerEx y(CGC_lock, Mutex::_no_safepoint_check_flag); 1638 _full_gc_requested = true; 1639 CGC_lock->notify(); // nudge CMS thread 1640 } 1641 } 1642 1643 1644 // The foreground and background collectors need to coordinate in order 1645 // to make sure that they do not mutually interfere with CMS collections. 1646 // When a background collection is active, 1647 // the foreground collector may need to take over (preempt) and 1648 // synchronously complete an ongoing collection. Depending on the 1649 // frequency of the background collections and the heap usage 1650 // of the application, this preemption can be seldom or frequent. 1651 // There are only certain 1652 // points in the background collection that the "collection-baton" 1653 // can be passed to the foreground collector. 1654 // 1655 // The foreground collector will wait for the baton before 1656 // starting any part of the collection. The foreground collector 1657 // will only wait at one location. 1658 // 1659 // The background collector will yield the baton before starting a new 1660 // phase of the collection (e.g., before initial marking, marking from roots, 1661 // precleaning, final re-mark, sweep etc.) This is normally done at the head 1662 // of the loop which switches the phases. The background collector does some 1663 // of the phases (initial mark, final re-mark) with the world stopped. 1664 // Because of locking involved in stopping the world, 1665 // the foreground collector should not block waiting for the background 1666 // collector when it is doing a stop-the-world phase. The background 1667 // collector will yield the baton at an additional point just before 1668 // it enters a stop-the-world phase. Once the world is stopped, the 1669 // background collector checks the phase of the collection. If the 1670 // phase has not changed, it proceeds with the collection. If the 1671 // phase has changed, it skips that phase of the collection. See 1672 // the comments on the use of the Heap_lock in collect_in_background(). 1673 // 1674 // Variable used in baton passing. 1675 // _foregroundGCIsActive - Set to true by the foreground collector when 1676 // it wants the baton. The foreground clears it when it has finished 1677 // the collection. 1678 // _foregroundGCShouldWait - Set to true by the background collector 1679 // when it is running. The foreground collector waits while 1680 // _foregroundGCShouldWait is true. 1681 // CGC_lock - monitor used to protect access to the above variables 1682 // and to notify the foreground and background collectors. 1683 // _collectorState - current state of the CMS collection. 1684 // 1685 // The foreground collector 1686 // acquires the CGC_lock 1687 // sets _foregroundGCIsActive 1688 // waits on the CGC_lock for _foregroundGCShouldWait to be false 1689 // various locks acquired in preparation for the collection 1690 // are released so as not to block the background collector 1691 // that is in the midst of a collection 1692 // proceeds with the collection 1693 // clears _foregroundGCIsActive 1694 // returns 1695 // 1696 // The background collector in a loop iterating on the phases of the 1697 // collection 1698 // acquires the CGC_lock 1699 // sets _foregroundGCShouldWait 1700 // if _foregroundGCIsActive is set 1701 // clears _foregroundGCShouldWait, notifies _CGC_lock 1702 // waits on _CGC_lock for _foregroundGCIsActive to become false 1703 // and exits the loop. 1704 // otherwise 1705 // proceed with that phase of the collection 1706 // if the phase is a stop-the-world phase, 1707 // yield the baton once more just before enqueueing 1708 // the stop-world CMS operation (executed by the VM thread). 1709 // returns after all phases of the collection are done 1710 // 1711 1712 void CMSCollector::acquire_control_and_collect(bool full, 1713 bool clear_all_soft_refs) { 1714 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint"); 1715 assert(!Thread::current()->is_ConcurrentGC_thread(), 1716 "shouldn't try to acquire control from self!"); 1717 1718 // Start the protocol for acquiring control of the 1719 // collection from the background collector (aka CMS thread). 1720 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 1721 "VM thread should have CMS token"); 1722 // Remember the possibly interrupted state of an ongoing 1723 // concurrent collection 1724 CollectorState first_state = _collectorState; 1725 1726 // Signal to a possibly ongoing concurrent collection that 1727 // we want to do a foreground collection. 1728 _foregroundGCIsActive = true; 1729 1730 // Disable incremental mode during a foreground collection. 1731 ICMSDisabler icms_disabler; 1732 1733 // release locks and wait for a notify from the background collector 1734 // releasing the locks in only necessary for phases which 1735 // do yields to improve the granularity of the collection. 1736 assert_lock_strong(bitMapLock()); 1737 // We need to lock the Free list lock for the space that we are 1738 // currently collecting. 1739 assert(haveFreelistLocks(), "Must be holding free list locks"); 1740 bitMapLock()->unlock(); 1741 releaseFreelistLocks(); 1742 { 1743 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1744 if (_foregroundGCShouldWait) { 1745 // We are going to be waiting for action for the CMS thread; 1746 // it had better not be gone (for instance at shutdown)! 1747 assert(ConcurrentMarkSweepThread::cmst() != NULL, 1748 "CMS thread must be running"); 1749 // Wait here until the background collector gives us the go-ahead 1750 ConcurrentMarkSweepThread::clear_CMS_flag( 1751 ConcurrentMarkSweepThread::CMS_vm_has_token); // release token 1752 // Get a possibly blocked CMS thread going: 1753 // Note that we set _foregroundGCIsActive true above, 1754 // without protection of the CGC_lock. 1755 CGC_lock->notify(); 1756 assert(!ConcurrentMarkSweepThread::vm_thread_wants_cms_token(), 1757 "Possible deadlock"); 1758 while (_foregroundGCShouldWait) { 1759 // wait for notification 1760 CGC_lock->wait(Mutex::_no_safepoint_check_flag); 1761 // Possibility of delay/starvation here, since CMS token does 1762 // not know to give priority to VM thread? Actually, i think 1763 // there wouldn't be any delay/starvation, but the proof of 1764 // that "fact" (?) appears non-trivial. XXX 20011219YSR 1765 } 1766 ConcurrentMarkSweepThread::set_CMS_flag( 1767 ConcurrentMarkSweepThread::CMS_vm_has_token); 1768 } 1769 } 1770 // The CMS_token is already held. Get back the other locks. 1771 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 1772 "VM thread should have CMS token"); 1773 getFreelistLocks(); 1774 bitMapLock()->lock_without_safepoint_check(); 1775 if (TraceCMSState) { 1776 gclog_or_tty->print_cr("CMS foreground collector has asked for control " 1777 INTPTR_FORMAT " with first state %d", Thread::current(), first_state); 1778 gclog_or_tty->print_cr(" gets control with state %d", _collectorState); 1779 } 1780 1781 // Check if we need to do a compaction, or if not, whether 1782 // we need to start the mark-sweep from scratch. 1783 bool should_compact = false; 1784 bool should_start_over = false; 1785 decide_foreground_collection_type(clear_all_soft_refs, 1786 &should_compact, &should_start_over); 1787 1788 NOT_PRODUCT( 1789 if (RotateCMSCollectionTypes) { 1790 if (_cmsGen->debug_collection_type() == 1791 ConcurrentMarkSweepGeneration::MSC_foreground_collection_type) { 1792 should_compact = true; 1793 } else if (_cmsGen->debug_collection_type() == 1794 ConcurrentMarkSweepGeneration::MS_foreground_collection_type) { 1795 should_compact = false; 1796 } 1797 } 1798 ) 1799 1800 if (PrintGCDetails && first_state > Idling) { 1801 GCCause::Cause cause = GenCollectedHeap::heap()->gc_cause(); 1802 if (GCCause::is_user_requested_gc(cause) || 1803 GCCause::is_serviceability_requested_gc(cause)) { 1804 gclog_or_tty->print(" (concurrent mode interrupted)"); 1805 } else { 1806 gclog_or_tty->print(" (concurrent mode failure)"); 1807 } 1808 } 1809 1810 if (should_compact) { 1811 // If the collection is being acquired from the background 1812 // collector, there may be references on the discovered 1813 // references lists that have NULL referents (being those 1814 // that were concurrently cleared by a mutator) or 1815 // that are no longer active (having been enqueued concurrently 1816 // by the mutator). 1817 // Scrub the list of those references because Mark-Sweep-Compact 1818 // code assumes referents are not NULL and that all discovered 1819 // Reference objects are active. 1820 ref_processor()->clean_up_discovered_references(); 1821 1822 do_compaction_work(clear_all_soft_refs); 1823 1824 // Has the GC time limit been exceeded? 1825 DefNewGeneration* young_gen = _young_gen->as_DefNewGeneration(); 1826 size_t max_eden_size = young_gen->max_capacity() - 1827 young_gen->to()->capacity() - 1828 young_gen->from()->capacity(); 1829 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1830 GCCause::Cause gc_cause = gch->gc_cause(); 1831 size_policy()->check_gc_overhead_limit(_young_gen->used(), 1832 young_gen->eden()->used(), 1833 _cmsGen->max_capacity(), 1834 max_eden_size, 1835 full, 1836 gc_cause, 1837 gch->collector_policy()); 1838 } else { 1839 do_mark_sweep_work(clear_all_soft_refs, first_state, 1840 should_start_over); 1841 } 1842 // Reset the expansion cause, now that we just completed 1843 // a collection cycle. 1844 clear_expansion_cause(); 1845 _foregroundGCIsActive = false; 1846 return; 1847 } 1848 1849 // Resize the perm generation and the tenured generation 1850 // after obtaining the free list locks for the 1851 // two generations. 1852 void CMSCollector::compute_new_size() { 1853 assert_locked_or_safepoint(Heap_lock); 1854 FreelistLocker z(this); 1855 _permGen->compute_new_size(); 1856 _cmsGen->compute_new_size(); 1857 } 1858 1859 // A work method used by foreground collection to determine 1860 // what type of collection (compacting or not, continuing or fresh) 1861 // it should do. 1862 // NOTE: the intent is to make UseCMSCompactAtFullCollection 1863 // and CMSCompactWhenClearAllSoftRefs the default in the future 1864 // and do away with the flags after a suitable period. 1865 void CMSCollector::decide_foreground_collection_type( 1866 bool clear_all_soft_refs, bool* should_compact, 1867 bool* should_start_over) { 1868 // Normally, we'll compact only if the UseCMSCompactAtFullCollection 1869 // flag is set, and we have either requested a System.gc() or 1870 // the number of full gc's since the last concurrent cycle 1871 // has exceeded the threshold set by CMSFullGCsBeforeCompaction, 1872 // or if an incremental collection has failed 1873 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1874 assert(gch->collector_policy()->is_two_generation_policy(), 1875 "You may want to check the correctness of the following"); 1876 // Inform cms gen if this was due to partial collection failing. 1877 // The CMS gen may use this fact to determine its expansion policy. 1878 if (gch->incremental_collection_will_fail()) { 1879 assert(!_cmsGen->incremental_collection_failed(), 1880 "Should have been noticed, reacted to and cleared"); 1881 _cmsGen->set_incremental_collection_failed(); 1882 } 1883 *should_compact = 1884 UseCMSCompactAtFullCollection && 1885 ((_full_gcs_since_conc_gc >= CMSFullGCsBeforeCompaction) || 1886 GCCause::is_user_requested_gc(gch->gc_cause()) || 1887 gch->incremental_collection_will_fail()); 1888 *should_start_over = false; 1889 if (clear_all_soft_refs && !*should_compact) { 1890 // We are about to do a last ditch collection attempt 1891 // so it would normally make sense to do a compaction 1892 // to reclaim as much space as possible. 1893 if (CMSCompactWhenClearAllSoftRefs) { 1894 // Default: The rationale is that in this case either 1895 // we are past the final marking phase, in which case 1896 // we'd have to start over, or so little has been done 1897 // that there's little point in saving that work. Compaction 1898 // appears to be the sensible choice in either case. 1899 *should_compact = true; 1900 } else { 1901 // We have been asked to clear all soft refs, but not to 1902 // compact. Make sure that we aren't past the final checkpoint 1903 // phase, for that is where we process soft refs. If we are already 1904 // past that phase, we'll need to redo the refs discovery phase and 1905 // if necessary clear soft refs that weren't previously 1906 // cleared. We do so by remembering the phase in which 1907 // we came in, and if we are past the refs processing 1908 // phase, we'll choose to just redo the mark-sweep 1909 // collection from scratch. 1910 if (_collectorState > FinalMarking) { 1911 // We are past the refs processing phase; 1912 // start over and do a fresh synchronous CMS cycle 1913 _collectorState = Resetting; // skip to reset to start new cycle 1914 reset(false /* == !asynch */); 1915 *should_start_over = true; 1916 } // else we can continue a possibly ongoing current cycle 1917 } 1918 } 1919 } 1920 1921 // A work method used by the foreground collector to do 1922 // a mark-sweep-compact. 1923 void CMSCollector::do_compaction_work(bool clear_all_soft_refs) { 1924 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1925 TraceTime t("CMS:MSC ", PrintGCDetails && Verbose, true, gclog_or_tty); 1926 if (PrintGC && Verbose && !(GCCause::is_user_requested_gc(gch->gc_cause()))) { 1927 gclog_or_tty->print_cr("Compact ConcurrentMarkSweepGeneration after %d " 1928 "collections passed to foreground collector", _full_gcs_since_conc_gc); 1929 } 1930 1931 // Sample collection interval time and reset for collection pause. 1932 if (UseAdaptiveSizePolicy) { 1933 size_policy()->msc_collection_begin(); 1934 } 1935 1936 // Temporarily widen the span of the weak reference processing to 1937 // the entire heap. 1938 MemRegion new_span(GenCollectedHeap::heap()->reserved_region()); 1939 ReferenceProcessorSpanMutator x(ref_processor(), new_span); 1940 1941 // Temporarily, clear the "is_alive_non_header" field of the 1942 // reference processor. 1943 ReferenceProcessorIsAliveMutator y(ref_processor(), NULL); 1944 1945 // Temporarily make reference _processing_ single threaded (non-MT). 1946 ReferenceProcessorMTProcMutator z(ref_processor(), false); 1947 1948 // Temporarily make refs discovery atomic 1949 ReferenceProcessorAtomicMutator w(ref_processor(), true); 1950 1951 ref_processor()->set_enqueuing_is_done(false); 1952 ref_processor()->enable_discovery(); 1953 ref_processor()->setup_policy(clear_all_soft_refs); 1954 // If an asynchronous collection finishes, the _modUnionTable is 1955 // all clear. If we are assuming the collection from an asynchronous 1956 // collection, clear the _modUnionTable. 1957 assert(_collectorState != Idling || _modUnionTable.isAllClear(), 1958 "_modUnionTable should be clear if the baton was not passed"); 1959 _modUnionTable.clear_all(); 1960 1961 // We must adjust the allocation statistics being maintained 1962 // in the free list space. We do so by reading and clearing 1963 // the sweep timer and updating the block flux rate estimates below. 1964 assert(!_intra_sweep_timer.is_active(), "_intra_sweep_timer should be inactive"); 1965 if (_inter_sweep_timer.is_active()) { 1966 _inter_sweep_timer.stop(); 1967 // Note that we do not use this sample to update the _inter_sweep_estimate. 1968 _cmsGen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()), 1969 _inter_sweep_estimate.padded_average(), 1970 _intra_sweep_estimate.padded_average()); 1971 } 1972 1973 GenMarkSweep::invoke_at_safepoint(_cmsGen->level(), 1974 ref_processor(), clear_all_soft_refs); 1975 #ifdef ASSERT 1976 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 1977 size_t free_size = cms_space->free(); 1978 assert(free_size == 1979 pointer_delta(cms_space->end(), cms_space->compaction_top()) 1980 * HeapWordSize, 1981 "All the free space should be compacted into one chunk at top"); 1982 assert(cms_space->dictionary()->totalChunkSize( 1983 debug_only(cms_space->freelistLock())) == 0 || 1984 cms_space->totalSizeInIndexedFreeLists() == 0, 1985 "All the free space should be in a single chunk"); 1986 size_t num = cms_space->totalCount(); 1987 assert((free_size == 0 && num == 0) || 1988 (free_size > 0 && (num == 1 || num == 2)), 1989 "There should be at most 2 free chunks after compaction"); 1990 #endif // ASSERT 1991 _collectorState = Resetting; 1992 assert(_restart_addr == NULL, 1993 "Should have been NULL'd before baton was passed"); 1994 reset(false /* == !asynch */); 1995 _cmsGen->reset_after_compaction(); 1996 _concurrent_cycles_since_last_unload = 0; 1997 1998 if (verifying() && !should_unload_classes()) { 1999 perm_gen_verify_bit_map()->clear_all(); 2000 } 2001 2002 // Clear any data recorded in the PLAB chunk arrays. 2003 if (_survivor_plab_array != NULL) { 2004 reset_survivor_plab_arrays(); 2005 } 2006 2007 // Adjust the per-size allocation stats for the next epoch. 2008 _cmsGen->cmsSpace()->endSweepFLCensus(sweep_count() /* fake */); 2009 // Restart the "inter sweep timer" for the next epoch. 2010 _inter_sweep_timer.reset(); 2011 _inter_sweep_timer.start(); 2012 2013 // Sample collection pause time and reset for collection interval. 2014 if (UseAdaptiveSizePolicy) { 2015 size_policy()->msc_collection_end(gch->gc_cause()); 2016 } 2017 2018 // For a mark-sweep-compact, compute_new_size() will be called 2019 // in the heap's do_collection() method. 2020 } 2021 2022 // A work method used by the foreground collector to do 2023 // a mark-sweep, after taking over from a possibly on-going 2024 // concurrent mark-sweep collection. 2025 void CMSCollector::do_mark_sweep_work(bool clear_all_soft_refs, 2026 CollectorState first_state, bool should_start_over) { 2027 if (PrintGC && Verbose) { 2028 gclog_or_tty->print_cr("Pass concurrent collection to foreground " 2029 "collector with count %d", 2030 _full_gcs_since_conc_gc); 2031 } 2032 switch (_collectorState) { 2033 case Idling: 2034 if (first_state == Idling || should_start_over) { 2035 // The background GC was not active, or should 2036 // restarted from scratch; start the cycle. 2037 _collectorState = InitialMarking; 2038 } 2039 // If first_state was not Idling, then a background GC 2040 // was in progress and has now finished. No need to do it 2041 // again. Leave the state as Idling. 2042 break; 2043 case Precleaning: 2044 // In the foreground case don't do the precleaning since 2045 // it is not done concurrently and there is extra work 2046 // required. 2047 _collectorState = FinalMarking; 2048 } 2049 if (PrintGCDetails && 2050 (_collectorState > Idling || 2051 !GCCause::is_user_requested_gc(GenCollectedHeap::heap()->gc_cause()))) { 2052 gclog_or_tty->print(" (concurrent mode failure)"); 2053 } 2054 collect_in_foreground(clear_all_soft_refs); 2055 2056 // For a mark-sweep, compute_new_size() will be called 2057 // in the heap's do_collection() method. 2058 } 2059 2060 2061 void CMSCollector::getFreelistLocks() const { 2062 // Get locks for all free lists in all generations that this 2063 // collector is responsible for 2064 _cmsGen->freelistLock()->lock_without_safepoint_check(); 2065 _permGen->freelistLock()->lock_without_safepoint_check(); 2066 } 2067 2068 void CMSCollector::releaseFreelistLocks() const { 2069 // Release locks for all free lists in all generations that this 2070 // collector is responsible for 2071 _cmsGen->freelistLock()->unlock(); 2072 _permGen->freelistLock()->unlock(); 2073 } 2074 2075 bool CMSCollector::haveFreelistLocks() const { 2076 // Check locks for all free lists in all generations that this 2077 // collector is responsible for 2078 assert_lock_strong(_cmsGen->freelistLock()); 2079 assert_lock_strong(_permGen->freelistLock()); 2080 PRODUCT_ONLY(ShouldNotReachHere()); 2081 return true; 2082 } 2083 2084 // A utility class that is used by the CMS collector to 2085 // temporarily "release" the foreground collector from its 2086 // usual obligation to wait for the background collector to 2087 // complete an ongoing phase before proceeding. 2088 class ReleaseForegroundGC: public StackObj { 2089 private: 2090 CMSCollector* _c; 2091 public: 2092 ReleaseForegroundGC(CMSCollector* c) : _c(c) { 2093 assert(_c->_foregroundGCShouldWait, "Else should not need to call"); 2094 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 2095 // allow a potentially blocked foreground collector to proceed 2096 _c->_foregroundGCShouldWait = false; 2097 if (_c->_foregroundGCIsActive) { 2098 CGC_lock->notify(); 2099 } 2100 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 2101 "Possible deadlock"); 2102 } 2103 2104 ~ReleaseForegroundGC() { 2105 assert(!_c->_foregroundGCShouldWait, "Usage protocol violation?"); 2106 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 2107 _c->_foregroundGCShouldWait = true; 2108 } 2109 }; 2110 2111 // There are separate collect_in_background and collect_in_foreground because of 2112 // the different locking requirements of the background collector and the 2113 // foreground collector. There was originally an attempt to share 2114 // one "collect" method between the background collector and the foreground 2115 // collector but the if-then-else required made it cleaner to have 2116 // separate methods. 2117 void CMSCollector::collect_in_background(bool clear_all_soft_refs) { 2118 assert(Thread::current()->is_ConcurrentGC_thread(), 2119 "A CMS asynchronous collection is only allowed on a CMS thread."); 2120 2121 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2122 { 2123 bool safepoint_check = Mutex::_no_safepoint_check_flag; 2124 MutexLockerEx hl(Heap_lock, safepoint_check); 2125 FreelistLocker fll(this); 2126 MutexLockerEx x(CGC_lock, safepoint_check); 2127 if (_foregroundGCIsActive || !UseAsyncConcMarkSweepGC) { 2128 // The foreground collector is active or we're 2129 // not using asynchronous collections. Skip this 2130 // background collection. 2131 assert(!_foregroundGCShouldWait, "Should be clear"); 2132 return; 2133 } else { 2134 assert(_collectorState == Idling, "Should be idling before start."); 2135 _collectorState = InitialMarking; 2136 // Reset the expansion cause, now that we are about to begin 2137 // a new cycle. 2138 clear_expansion_cause(); 2139 } 2140 // Decide if we want to enable class unloading as part of the 2141 // ensuing concurrent GC cycle. 2142 update_should_unload_classes(); 2143 _full_gc_requested = false; // acks all outstanding full gc requests 2144 // Signal that we are about to start a collection 2145 gch->increment_total_full_collections(); // ... starting a collection cycle 2146 _collection_count_start = gch->total_full_collections(); 2147 } 2148 2149 // Used for PrintGC 2150 size_t prev_used; 2151 if (PrintGC && Verbose) { 2152 prev_used = _cmsGen->used(); // XXXPERM 2153 } 2154 2155 // The change of the collection state is normally done at this level; 2156 // the exceptions are phases that are executed while the world is 2157 // stopped. For those phases the change of state is done while the 2158 // world is stopped. For baton passing purposes this allows the 2159 // background collector to finish the phase and change state atomically. 2160 // The foreground collector cannot wait on a phase that is done 2161 // while the world is stopped because the foreground collector already 2162 // has the world stopped and would deadlock. 2163 while (_collectorState != Idling) { 2164 if (TraceCMSState) { 2165 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " in CMS state %d", 2166 Thread::current(), _collectorState); 2167 } 2168 // The foreground collector 2169 // holds the Heap_lock throughout its collection. 2170 // holds the CMS token (but not the lock) 2171 // except while it is waiting for the background collector to yield. 2172 // 2173 // The foreground collector should be blocked (not for long) 2174 // if the background collector is about to start a phase 2175 // executed with world stopped. If the background 2176 // collector has already started such a phase, the 2177 // foreground collector is blocked waiting for the 2178 // Heap_lock. The stop-world phases (InitialMarking and FinalMarking) 2179 // are executed in the VM thread. 2180 // 2181 // The locking order is 2182 // PendingListLock (PLL) -- if applicable (FinalMarking) 2183 // Heap_lock (both this & PLL locked in VM_CMS_Operation::prologue()) 2184 // CMS token (claimed in 2185 // stop_world_and_do() --> 2186 // safepoint_synchronize() --> 2187 // CMSThread::synchronize()) 2188 2189 { 2190 // Check if the FG collector wants us to yield. 2191 CMSTokenSync x(true); // is cms thread 2192 if (waitForForegroundGC()) { 2193 // We yielded to a foreground GC, nothing more to be 2194 // done this round. 2195 assert(_foregroundGCShouldWait == false, "We set it to false in " 2196 "waitForForegroundGC()"); 2197 if (TraceCMSState) { 2198 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT 2199 " exiting collection CMS state %d", 2200 Thread::current(), _collectorState); 2201 } 2202 return; 2203 } else { 2204 // The background collector can run but check to see if the 2205 // foreground collector has done a collection while the 2206 // background collector was waiting to get the CGC_lock 2207 // above. If yes, break so that _foregroundGCShouldWait 2208 // is cleared before returning. 2209 if (_collectorState == Idling) { 2210 break; 2211 } 2212 } 2213 } 2214 2215 assert(_foregroundGCShouldWait, "Foreground collector, if active, " 2216 "should be waiting"); 2217 2218 switch (_collectorState) { 2219 case InitialMarking: 2220 { 2221 ReleaseForegroundGC x(this); 2222 stats().record_cms_begin(); 2223 2224 VM_CMS_Initial_Mark initial_mark_op(this); 2225 VMThread::execute(&initial_mark_op); 2226 } 2227 // The collector state may be any legal state at this point 2228 // since the background collector may have yielded to the 2229 // foreground collector. 2230 break; 2231 case Marking: 2232 // initial marking in checkpointRootsInitialWork has been completed 2233 if (markFromRoots(true)) { // we were successful 2234 assert(_collectorState == Precleaning, "Collector state should " 2235 "have changed"); 2236 } else { 2237 assert(_foregroundGCIsActive, "Internal state inconsistency"); 2238 } 2239 break; 2240 case Precleaning: 2241 if (UseAdaptiveSizePolicy) { 2242 size_policy()->concurrent_precleaning_begin(); 2243 } 2244 // marking from roots in markFromRoots has been completed 2245 preclean(); 2246 if (UseAdaptiveSizePolicy) { 2247 size_policy()->concurrent_precleaning_end(); 2248 } 2249 assert(_collectorState == AbortablePreclean || 2250 _collectorState == FinalMarking, 2251 "Collector state should have changed"); 2252 break; 2253 case AbortablePreclean: 2254 if (UseAdaptiveSizePolicy) { 2255 size_policy()->concurrent_phases_resume(); 2256 } 2257 abortable_preclean(); 2258 if (UseAdaptiveSizePolicy) { 2259 size_policy()->concurrent_precleaning_end(); 2260 } 2261 assert(_collectorState == FinalMarking, "Collector state should " 2262 "have changed"); 2263 break; 2264 case FinalMarking: 2265 { 2266 ReleaseForegroundGC x(this); 2267 2268 VM_CMS_Final_Remark final_remark_op(this); 2269 VMThread::execute(&final_remark_op); 2270 } 2271 assert(_foregroundGCShouldWait, "block post-condition"); 2272 break; 2273 case Sweeping: 2274 if (UseAdaptiveSizePolicy) { 2275 size_policy()->concurrent_sweeping_begin(); 2276 } 2277 // final marking in checkpointRootsFinal has been completed 2278 sweep(true); 2279 assert(_collectorState == Resizing, "Collector state change " 2280 "to Resizing must be done under the free_list_lock"); 2281 _full_gcs_since_conc_gc = 0; 2282 2283 // Stop the timers for adaptive size policy for the concurrent phases 2284 if (UseAdaptiveSizePolicy) { 2285 size_policy()->concurrent_sweeping_end(); 2286 size_policy()->concurrent_phases_end(gch->gc_cause(), 2287 gch->prev_gen(_cmsGen)->capacity(), 2288 _cmsGen->free()); 2289 } 2290 2291 case Resizing: { 2292 // Sweeping has been completed... 2293 // At this point the background collection has completed. 2294 // Don't move the call to compute_new_size() down 2295 // into code that might be executed if the background 2296 // collection was preempted. 2297 { 2298 ReleaseForegroundGC x(this); // unblock FG collection 2299 MutexLockerEx y(Heap_lock, Mutex::_no_safepoint_check_flag); 2300 CMSTokenSync z(true); // not strictly needed. 2301 if (_collectorState == Resizing) { 2302 compute_new_size(); 2303 _collectorState = Resetting; 2304 } else { 2305 assert(_collectorState == Idling, "The state should only change" 2306 " because the foreground collector has finished the collection"); 2307 } 2308 } 2309 break; 2310 } 2311 case Resetting: 2312 // CMS heap resizing has been completed 2313 reset(true); 2314 assert(_collectorState == Idling, "Collector state should " 2315 "have changed"); 2316 stats().record_cms_end(); 2317 // Don't move the concurrent_phases_end() and compute_new_size() 2318 // calls to here because a preempted background collection 2319 // has it's state set to "Resetting". 2320 break; 2321 case Idling: 2322 default: 2323 ShouldNotReachHere(); 2324 break; 2325 } 2326 if (TraceCMSState) { 2327 gclog_or_tty->print_cr(" Thread " INTPTR_FORMAT " done - next CMS state %d", 2328 Thread::current(), _collectorState); 2329 } 2330 assert(_foregroundGCShouldWait, "block post-condition"); 2331 } 2332 2333 // Should this be in gc_epilogue? 2334 collector_policy()->counters()->update_counters(); 2335 2336 { 2337 // Clear _foregroundGCShouldWait and, in the event that the 2338 // foreground collector is waiting, notify it, before 2339 // returning. 2340 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 2341 _foregroundGCShouldWait = false; 2342 if (_foregroundGCIsActive) { 2343 CGC_lock->notify(); 2344 } 2345 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 2346 "Possible deadlock"); 2347 } 2348 if (TraceCMSState) { 2349 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT 2350 " exiting collection CMS state %d", 2351 Thread::current(), _collectorState); 2352 } 2353 if (PrintGC && Verbose) { 2354 _cmsGen->print_heap_change(prev_used); 2355 } 2356 } 2357 2358 void CMSCollector::collect_in_foreground(bool clear_all_soft_refs) { 2359 assert(_foregroundGCIsActive && !_foregroundGCShouldWait, 2360 "Foreground collector should be waiting, not executing"); 2361 assert(Thread::current()->is_VM_thread(), "A foreground collection" 2362 "may only be done by the VM Thread with the world stopped"); 2363 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 2364 "VM thread should have CMS token"); 2365 2366 NOT_PRODUCT(TraceTime t("CMS:MS (foreground) ", PrintGCDetails && Verbose, 2367 true, gclog_or_tty);) 2368 if (UseAdaptiveSizePolicy) { 2369 size_policy()->ms_collection_begin(); 2370 } 2371 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact); 2372 2373 HandleMark hm; // Discard invalid handles created during verification 2374 2375 if (VerifyBeforeGC && 2376 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2377 Universe::verify(true); 2378 } 2379 2380 // Snapshot the soft reference policy to be used in this collection cycle. 2381 ref_processor()->setup_policy(clear_all_soft_refs); 2382 2383 bool init_mark_was_synchronous = false; // until proven otherwise 2384 while (_collectorState != Idling) { 2385 if (TraceCMSState) { 2386 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " in CMS state %d", 2387 Thread::current(), _collectorState); 2388 } 2389 switch (_collectorState) { 2390 case InitialMarking: 2391 init_mark_was_synchronous = true; // fact to be exploited in re-mark 2392 checkpointRootsInitial(false); 2393 assert(_collectorState == Marking, "Collector state should have changed" 2394 " within checkpointRootsInitial()"); 2395 break; 2396 case Marking: 2397 // initial marking in checkpointRootsInitialWork has been completed 2398 if (VerifyDuringGC && 2399 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2400 gclog_or_tty->print("Verify before initial mark: "); 2401 Universe::verify(true); 2402 } 2403 { 2404 bool res = markFromRoots(false); 2405 assert(res && _collectorState == FinalMarking, "Collector state should " 2406 "have changed"); 2407 break; 2408 } 2409 case FinalMarking: 2410 if (VerifyDuringGC && 2411 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2412 gclog_or_tty->print("Verify before re-mark: "); 2413 Universe::verify(true); 2414 } 2415 checkpointRootsFinal(false, clear_all_soft_refs, 2416 init_mark_was_synchronous); 2417 assert(_collectorState == Sweeping, "Collector state should not " 2418 "have changed within checkpointRootsFinal()"); 2419 break; 2420 case Sweeping: 2421 // final marking in checkpointRootsFinal has been completed 2422 if (VerifyDuringGC && 2423 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2424 gclog_or_tty->print("Verify before sweep: "); 2425 Universe::verify(true); 2426 } 2427 sweep(false); 2428 assert(_collectorState == Resizing, "Incorrect state"); 2429 break; 2430 case Resizing: { 2431 // Sweeping has been completed; the actual resize in this case 2432 // is done separately; nothing to be done in this state. 2433 _collectorState = Resetting; 2434 break; 2435 } 2436 case Resetting: 2437 // The heap has been resized. 2438 if (VerifyDuringGC && 2439 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2440 gclog_or_tty->print("Verify before reset: "); 2441 Universe::verify(true); 2442 } 2443 reset(false); 2444 assert(_collectorState == Idling, "Collector state should " 2445 "have changed"); 2446 break; 2447 case Precleaning: 2448 case AbortablePreclean: 2449 // Elide the preclean phase 2450 _collectorState = FinalMarking; 2451 break; 2452 default: 2453 ShouldNotReachHere(); 2454 } 2455 if (TraceCMSState) { 2456 gclog_or_tty->print_cr(" Thread " INTPTR_FORMAT " done - next CMS state %d", 2457 Thread::current(), _collectorState); 2458 } 2459 } 2460 2461 if (UseAdaptiveSizePolicy) { 2462 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2463 size_policy()->ms_collection_end(gch->gc_cause()); 2464 } 2465 2466 if (VerifyAfterGC && 2467 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2468 Universe::verify(true); 2469 } 2470 if (TraceCMSState) { 2471 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT 2472 " exiting collection CMS state %d", 2473 Thread::current(), _collectorState); 2474 } 2475 } 2476 2477 bool CMSCollector::waitForForegroundGC() { 2478 bool res = false; 2479 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 2480 "CMS thread should have CMS token"); 2481 // Block the foreground collector until the 2482 // background collectors decides whether to 2483 // yield. 2484 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 2485 _foregroundGCShouldWait = true; 2486 if (_foregroundGCIsActive) { 2487 // The background collector yields to the 2488 // foreground collector and returns a value 2489 // indicating that it has yielded. The foreground 2490 // collector can proceed. 2491 res = true; 2492 _foregroundGCShouldWait = false; 2493 ConcurrentMarkSweepThread::clear_CMS_flag( 2494 ConcurrentMarkSweepThread::CMS_cms_has_token); 2495 ConcurrentMarkSweepThread::set_CMS_flag( 2496 ConcurrentMarkSweepThread::CMS_cms_wants_token); 2497 // Get a possibly blocked foreground thread going 2498 CGC_lock->notify(); 2499 if (TraceCMSState) { 2500 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " waiting at CMS state %d", 2501 Thread::current(), _collectorState); 2502 } 2503 while (_foregroundGCIsActive) { 2504 CGC_lock->wait(Mutex::_no_safepoint_check_flag); 2505 } 2506 ConcurrentMarkSweepThread::set_CMS_flag( 2507 ConcurrentMarkSweepThread::CMS_cms_has_token); 2508 ConcurrentMarkSweepThread::clear_CMS_flag( 2509 ConcurrentMarkSweepThread::CMS_cms_wants_token); 2510 } 2511 if (TraceCMSState) { 2512 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " continuing at CMS state %d", 2513 Thread::current(), _collectorState); 2514 } 2515 return res; 2516 } 2517 2518 // Because of the need to lock the free lists and other structures in 2519 // the collector, common to all the generations that the collector is 2520 // collecting, we need the gc_prologues of individual CMS generations 2521 // delegate to their collector. It may have been simpler had the 2522 // current infrastructure allowed one to call a prologue on a 2523 // collector. In the absence of that we have the generation's 2524 // prologue delegate to the collector, which delegates back 2525 // some "local" work to a worker method in the individual generations 2526 // that it's responsible for collecting, while itself doing any 2527 // work common to all generations it's responsible for. A similar 2528 // comment applies to the gc_epilogue()'s. 2529 // The role of the varaible _between_prologue_and_epilogue is to 2530 // enforce the invocation protocol. 2531 void CMSCollector::gc_prologue(bool full) { 2532 // Call gc_prologue_work() for each CMSGen and PermGen that 2533 // we are responsible for. 2534 2535 // The following locking discipline assumes that we are only called 2536 // when the world is stopped. 2537 assert(SafepointSynchronize::is_at_safepoint(), "world is stopped assumption"); 2538 2539 // The CMSCollector prologue must call the gc_prologues for the 2540 // "generations" (including PermGen if any) that it's responsible 2541 // for. 2542 2543 assert( Thread::current()->is_VM_thread() 2544 || ( CMSScavengeBeforeRemark 2545 && Thread::current()->is_ConcurrentGC_thread()), 2546 "Incorrect thread type for prologue execution"); 2547 2548 if (_between_prologue_and_epilogue) { 2549 // We have already been invoked; this is a gc_prologue delegation 2550 // from yet another CMS generation that we are responsible for, just 2551 // ignore it since all relevant work has already been done. 2552 return; 2553 } 2554 2555 // set a bit saying prologue has been called; cleared in epilogue 2556 _between_prologue_and_epilogue = true; 2557 // Claim locks for common data structures, then call gc_prologue_work() 2558 // for each CMSGen and PermGen that we are responsible for. 2559 2560 getFreelistLocks(); // gets free list locks on constituent spaces 2561 bitMapLock()->lock_without_safepoint_check(); 2562 2563 // Should call gc_prologue_work() for all cms gens we are responsible for 2564 bool registerClosure = _collectorState >= Marking 2565 && _collectorState < Sweeping; 2566 ModUnionClosure* muc = ParallelGCThreads > 0 ? &_modUnionClosurePar 2567 : &_modUnionClosure; 2568 _cmsGen->gc_prologue_work(full, registerClosure, muc); 2569 _permGen->gc_prologue_work(full, registerClosure, muc); 2570 2571 if (!full) { 2572 stats().record_gc0_begin(); 2573 } 2574 } 2575 2576 void ConcurrentMarkSweepGeneration::gc_prologue(bool full) { 2577 // Delegate to CMScollector which knows how to coordinate between 2578 // this and any other CMS generations that it is responsible for 2579 // collecting. 2580 collector()->gc_prologue(full); 2581 } 2582 2583 // This is a "private" interface for use by this generation's CMSCollector. 2584 // Not to be called directly by any other entity (for instance, 2585 // GenCollectedHeap, which calls the "public" gc_prologue method above). 2586 void ConcurrentMarkSweepGeneration::gc_prologue_work(bool full, 2587 bool registerClosure, ModUnionClosure* modUnionClosure) { 2588 assert(!incremental_collection_failed(), "Shouldn't be set yet"); 2589 assert(cmsSpace()->preconsumptionDirtyCardClosure() == NULL, 2590 "Should be NULL"); 2591 if (registerClosure) { 2592 cmsSpace()->setPreconsumptionDirtyCardClosure(modUnionClosure); 2593 } 2594 cmsSpace()->gc_prologue(); 2595 // Clear stat counters 2596 NOT_PRODUCT( 2597 assert(_numObjectsPromoted == 0, "check"); 2598 assert(_numWordsPromoted == 0, "check"); 2599 if (Verbose && PrintGC) { 2600 gclog_or_tty->print("Allocated "SIZE_FORMAT" objects, " 2601 SIZE_FORMAT" bytes concurrently", 2602 _numObjectsAllocated, _numWordsAllocated*sizeof(HeapWord)); 2603 } 2604 _numObjectsAllocated = 0; 2605 _numWordsAllocated = 0; 2606 ) 2607 } 2608 2609 void CMSCollector::gc_epilogue(bool full) { 2610 // The following locking discipline assumes that we are only called 2611 // when the world is stopped. 2612 assert(SafepointSynchronize::is_at_safepoint(), 2613 "world is stopped assumption"); 2614 2615 // Currently the CMS epilogue (see CompactibleFreeListSpace) merely checks 2616 // if linear allocation blocks need to be appropriately marked to allow the 2617 // the blocks to be parsable. We also check here whether we need to nudge the 2618 // CMS collector thread to start a new cycle (if it's not already active). 2619 assert( Thread::current()->is_VM_thread() 2620 || ( CMSScavengeBeforeRemark 2621 && Thread::current()->is_ConcurrentGC_thread()), 2622 "Incorrect thread type for epilogue execution"); 2623 2624 if (!_between_prologue_and_epilogue) { 2625 // We have already been invoked; this is a gc_epilogue delegation 2626 // from yet another CMS generation that we are responsible for, just 2627 // ignore it since all relevant work has already been done. 2628 return; 2629 } 2630 assert(haveFreelistLocks(), "must have freelist locks"); 2631 assert_lock_strong(bitMapLock()); 2632 2633 _cmsGen->gc_epilogue_work(full); 2634 _permGen->gc_epilogue_work(full); 2635 2636 if (_collectorState == AbortablePreclean || _collectorState == Precleaning) { 2637 // in case sampling was not already enabled, enable it 2638 _start_sampling = true; 2639 } 2640 // reset _eden_chunk_array so sampling starts afresh 2641 _eden_chunk_index = 0; 2642 2643 size_t cms_used = _cmsGen->cmsSpace()->used(); 2644 size_t perm_used = _permGen->cmsSpace()->used(); 2645 2646 // update performance counters - this uses a special version of 2647 // update_counters() that allows the utilization to be passed as a 2648 // parameter, avoiding multiple calls to used(). 2649 // 2650 _cmsGen->update_counters(cms_used); 2651 _permGen->update_counters(perm_used); 2652 2653 if (CMSIncrementalMode) { 2654 icms_update_allocation_limits(); 2655 } 2656 2657 bitMapLock()->unlock(); 2658 releaseFreelistLocks(); 2659 2660 _between_prologue_and_epilogue = false; // ready for next cycle 2661 } 2662 2663 void ConcurrentMarkSweepGeneration::gc_epilogue(bool full) { 2664 collector()->gc_epilogue(full); 2665 2666 // Also reset promotion tracking in par gc thread states. 2667 if (ParallelGCThreads > 0) { 2668 for (uint i = 0; i < ParallelGCThreads; i++) { 2669 _par_gc_thread_states[i]->promo.stopTrackingPromotions(i); 2670 } 2671 } 2672 } 2673 2674 void ConcurrentMarkSweepGeneration::gc_epilogue_work(bool full) { 2675 assert(!incremental_collection_failed(), "Should have been cleared"); 2676 cmsSpace()->setPreconsumptionDirtyCardClosure(NULL); 2677 cmsSpace()->gc_epilogue(); 2678 // Print stat counters 2679 NOT_PRODUCT( 2680 assert(_numObjectsAllocated == 0, "check"); 2681 assert(_numWordsAllocated == 0, "check"); 2682 if (Verbose && PrintGC) { 2683 gclog_or_tty->print("Promoted "SIZE_FORMAT" objects, " 2684 SIZE_FORMAT" bytes", 2685 _numObjectsPromoted, _numWordsPromoted*sizeof(HeapWord)); 2686 } 2687 _numObjectsPromoted = 0; 2688 _numWordsPromoted = 0; 2689 ) 2690 2691 if (PrintGC && Verbose) { 2692 // Call down the chain in contiguous_available needs the freelistLock 2693 // so print this out before releasing the freeListLock. 2694 gclog_or_tty->print(" Contiguous available "SIZE_FORMAT" bytes ", 2695 contiguous_available()); 2696 } 2697 } 2698 2699 #ifndef PRODUCT 2700 bool CMSCollector::have_cms_token() { 2701 Thread* thr = Thread::current(); 2702 if (thr->is_VM_thread()) { 2703 return ConcurrentMarkSweepThread::vm_thread_has_cms_token(); 2704 } else if (thr->is_ConcurrentGC_thread()) { 2705 return ConcurrentMarkSweepThread::cms_thread_has_cms_token(); 2706 } else if (thr->is_GC_task_thread()) { 2707 return ConcurrentMarkSweepThread::vm_thread_has_cms_token() && 2708 ParGCRareEvent_lock->owned_by_self(); 2709 } 2710 return false; 2711 } 2712 #endif 2713 2714 // Check reachability of the given heap address in CMS generation, 2715 // treating all other generations as roots. 2716 bool CMSCollector::is_cms_reachable(HeapWord* addr) { 2717 // We could "guarantee" below, rather than assert, but i'll 2718 // leave these as "asserts" so that an adventurous debugger 2719 // could try this in the product build provided some subset of 2720 // the conditions were met, provided they were intersted in the 2721 // results and knew that the computation below wouldn't interfere 2722 // with other concurrent computations mutating the structures 2723 // being read or written. 2724 assert(SafepointSynchronize::is_at_safepoint(), 2725 "Else mutations in object graph will make answer suspect"); 2726 assert(have_cms_token(), "Should hold cms token"); 2727 assert(haveFreelistLocks(), "must hold free list locks"); 2728 assert_lock_strong(bitMapLock()); 2729 2730 // Clear the marking bit map array before starting, but, just 2731 // for kicks, first report if the given address is already marked 2732 gclog_or_tty->print_cr("Start: Address 0x%x is%s marked", addr, 2733 _markBitMap.isMarked(addr) ? "" : " not"); 2734 2735 if (verify_after_remark()) { 2736 MutexLockerEx x(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag); 2737 bool result = verification_mark_bm()->isMarked(addr); 2738 gclog_or_tty->print_cr("TransitiveMark: Address 0x%x %s marked", addr, 2739 result ? "IS" : "is NOT"); 2740 return result; 2741 } else { 2742 gclog_or_tty->print_cr("Could not compute result"); 2743 return false; 2744 } 2745 } 2746 2747 //////////////////////////////////////////////////////// 2748 // CMS Verification Support 2749 //////////////////////////////////////////////////////// 2750 // Following the remark phase, the following invariant 2751 // should hold -- each object in the CMS heap which is 2752 // marked in markBitMap() should be marked in the verification_mark_bm(). 2753 2754 class VerifyMarkedClosure: public BitMapClosure { 2755 CMSBitMap* _marks; 2756 bool _failed; 2757 2758 public: 2759 VerifyMarkedClosure(CMSBitMap* bm): _marks(bm), _failed(false) {} 2760 2761 bool do_bit(size_t offset) { 2762 HeapWord* addr = _marks->offsetToHeapWord(offset); 2763 if (!_marks->isMarked(addr)) { 2764 oop(addr)->print_on(gclog_or_tty); 2765 gclog_or_tty->print_cr(" ("INTPTR_FORMAT" should have been marked)", addr); 2766 _failed = true; 2767 } 2768 return true; 2769 } 2770 2771 bool failed() { return _failed; } 2772 }; 2773 2774 bool CMSCollector::verify_after_remark() { 2775 gclog_or_tty->print(" [Verifying CMS Marking... "); 2776 MutexLockerEx ml(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag); 2777 static bool init = false; 2778 2779 assert(SafepointSynchronize::is_at_safepoint(), 2780 "Else mutations in object graph will make answer suspect"); 2781 assert(have_cms_token(), 2782 "Else there may be mutual interference in use of " 2783 " verification data structures"); 2784 assert(_collectorState > Marking && _collectorState <= Sweeping, 2785 "Else marking info checked here may be obsolete"); 2786 assert(haveFreelistLocks(), "must hold free list locks"); 2787 assert_lock_strong(bitMapLock()); 2788 2789 2790 // Allocate marking bit map if not already allocated 2791 if (!init) { // first time 2792 if (!verification_mark_bm()->allocate(_span)) { 2793 return false; 2794 } 2795 init = true; 2796 } 2797 2798 assert(verification_mark_stack()->isEmpty(), "Should be empty"); 2799 2800 // Turn off refs discovery -- so we will be tracing through refs. 2801 // This is as intended, because by this time 2802 // GC must already have cleared any refs that need to be cleared, 2803 // and traced those that need to be marked; moreover, 2804 // the marking done here is not going to intefere in any 2805 // way with the marking information used by GC. 2806 NoRefDiscovery no_discovery(ref_processor()); 2807 2808 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;) 2809 2810 // Clear any marks from a previous round 2811 verification_mark_bm()->clear_all(); 2812 assert(verification_mark_stack()->isEmpty(), "markStack should be empty"); 2813 verify_work_stacks_empty(); 2814 2815 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2816 gch->ensure_parsability(false); // fill TLABs, but no need to retire them 2817 // Update the saved marks which may affect the root scans. 2818 gch->save_marks(); 2819 2820 if (CMSRemarkVerifyVariant == 1) { 2821 // In this first variant of verification, we complete 2822 // all marking, then check if the new marks-verctor is 2823 // a subset of the CMS marks-vector. 2824 verify_after_remark_work_1(); 2825 } else if (CMSRemarkVerifyVariant == 2) { 2826 // In this second variant of verification, we flag an error 2827 // (i.e. an object reachable in the new marks-vector not reachable 2828 // in the CMS marks-vector) immediately, also indicating the 2829 // identify of an object (A) that references the unmarked object (B) -- 2830 // presumably, a mutation to A failed to be picked up by preclean/remark? 2831 verify_after_remark_work_2(); 2832 } else { 2833 warning("Unrecognized value %d for CMSRemarkVerifyVariant", 2834 CMSRemarkVerifyVariant); 2835 } 2836 gclog_or_tty->print(" done] "); 2837 return true; 2838 } 2839 2840 void CMSCollector::verify_after_remark_work_1() { 2841 ResourceMark rm; 2842 HandleMark hm; 2843 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2844 2845 // Mark from roots one level into CMS 2846 MarkRefsIntoClosure notOlder(_span, verification_mark_bm()); 2847 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2848 2849 gch->gen_process_strong_roots(_cmsGen->level(), 2850 true, // younger gens are roots 2851 true, // activate StrongRootsScope 2852 true, // collecting perm gen 2853 SharedHeap::ScanningOption(roots_scanning_options()), 2854 ¬Older, 2855 true, // walk code active on stacks 2856 NULL); 2857 2858 // Now mark from the roots 2859 assert(_revisitStack.isEmpty(), "Should be empty"); 2860 MarkFromRootsClosure markFromRootsClosure(this, _span, 2861 verification_mark_bm(), verification_mark_stack(), &_revisitStack, 2862 false /* don't yield */, true /* verifying */); 2863 assert(_restart_addr == NULL, "Expected pre-condition"); 2864 verification_mark_bm()->iterate(&markFromRootsClosure); 2865 while (_restart_addr != NULL) { 2866 // Deal with stack overflow: by restarting at the indicated 2867 // address. 2868 HeapWord* ra = _restart_addr; 2869 markFromRootsClosure.reset(ra); 2870 _restart_addr = NULL; 2871 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end()); 2872 } 2873 assert(verification_mark_stack()->isEmpty(), "Should have been drained"); 2874 verify_work_stacks_empty(); 2875 // Should reset the revisit stack above, since no class tree 2876 // surgery is forthcoming. 2877 _revisitStack.reset(); // throwing away all contents 2878 2879 // Marking completed -- now verify that each bit marked in 2880 // verification_mark_bm() is also marked in markBitMap(); flag all 2881 // errors by printing corresponding objects. 2882 VerifyMarkedClosure vcl(markBitMap()); 2883 verification_mark_bm()->iterate(&vcl); 2884 if (vcl.failed()) { 2885 gclog_or_tty->print("Verification failed"); 2886 Universe::heap()->print_on(gclog_or_tty); 2887 fatal("CMS: failed marking verification after remark"); 2888 } 2889 } 2890 2891 void CMSCollector::verify_after_remark_work_2() { 2892 ResourceMark rm; 2893 HandleMark hm; 2894 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2895 2896 // Mark from roots one level into CMS 2897 MarkRefsIntoVerifyClosure notOlder(_span, verification_mark_bm(), 2898 markBitMap()); 2899 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2900 gch->gen_process_strong_roots(_cmsGen->level(), 2901 true, // younger gens are roots 2902 true, // activate StrongRootsScope 2903 true, // collecting perm gen 2904 SharedHeap::ScanningOption(roots_scanning_options()), 2905 ¬Older, 2906 true, // walk code active on stacks 2907 NULL); 2908 2909 // Now mark from the roots 2910 assert(_revisitStack.isEmpty(), "Should be empty"); 2911 MarkFromRootsVerifyClosure markFromRootsClosure(this, _span, 2912 verification_mark_bm(), markBitMap(), verification_mark_stack()); 2913 assert(_restart_addr == NULL, "Expected pre-condition"); 2914 verification_mark_bm()->iterate(&markFromRootsClosure); 2915 while (_restart_addr != NULL) { 2916 // Deal with stack overflow: by restarting at the indicated 2917 // address. 2918 HeapWord* ra = _restart_addr; 2919 markFromRootsClosure.reset(ra); 2920 _restart_addr = NULL; 2921 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end()); 2922 } 2923 assert(verification_mark_stack()->isEmpty(), "Should have been drained"); 2924 verify_work_stacks_empty(); 2925 // Should reset the revisit stack above, since no class tree 2926 // surgery is forthcoming. 2927 _revisitStack.reset(); // throwing away all contents 2928 2929 // Marking completed -- now verify that each bit marked in 2930 // verification_mark_bm() is also marked in markBitMap(); flag all 2931 // errors by printing corresponding objects. 2932 VerifyMarkedClosure vcl(markBitMap()); 2933 verification_mark_bm()->iterate(&vcl); 2934 assert(!vcl.failed(), "Else verification above should not have succeeded"); 2935 } 2936 2937 void ConcurrentMarkSweepGeneration::save_marks() { 2938 // delegate to CMS space 2939 cmsSpace()->save_marks(); 2940 for (uint i = 0; i < ParallelGCThreads; i++) { 2941 _par_gc_thread_states[i]->promo.startTrackingPromotions(); 2942 } 2943 } 2944 2945 bool ConcurrentMarkSweepGeneration::no_allocs_since_save_marks() { 2946 return cmsSpace()->no_allocs_since_save_marks(); 2947 } 2948 2949 #define CMS_SINCE_SAVE_MARKS_DEFN(OopClosureType, nv_suffix) \ 2950 \ 2951 void ConcurrentMarkSweepGeneration:: \ 2952 oop_since_save_marks_iterate##nv_suffix(OopClosureType* cl) { \ 2953 cl->set_generation(this); \ 2954 cmsSpace()->oop_since_save_marks_iterate##nv_suffix(cl); \ 2955 cl->reset_generation(); \ 2956 save_marks(); \ 2957 } 2958 2959 ALL_SINCE_SAVE_MARKS_CLOSURES(CMS_SINCE_SAVE_MARKS_DEFN) 2960 2961 void 2962 ConcurrentMarkSweepGeneration::object_iterate_since_last_GC(ObjectClosure* blk) 2963 { 2964 // Not currently implemented; need to do the following. -- ysr. 2965 // dld -- I think that is used for some sort of allocation profiler. So it 2966 // really means the objects allocated by the mutator since the last 2967 // GC. We could potentially implement this cheaply by recording only 2968 // the direct allocations in a side data structure. 2969 // 2970 // I think we probably ought not to be required to support these 2971 // iterations at any arbitrary point; I think there ought to be some 2972 // call to enable/disable allocation profiling in a generation/space, 2973 // and the iterator ought to return the objects allocated in the 2974 // gen/space since the enable call, or the last iterator call (which 2975 // will probably be at a GC.) That way, for gens like CM&S that would 2976 // require some extra data structure to support this, we only pay the 2977 // cost when it's in use... 2978 cmsSpace()->object_iterate_since_last_GC(blk); 2979 } 2980 2981 void 2982 ConcurrentMarkSweepGeneration::younger_refs_iterate(OopsInGenClosure* cl) { 2983 cl->set_generation(this); 2984 younger_refs_in_space_iterate(_cmsSpace, cl); 2985 cl->reset_generation(); 2986 } 2987 2988 void 2989 ConcurrentMarkSweepGeneration::oop_iterate(MemRegion mr, OopClosure* cl) { 2990 if (freelistLock()->owned_by_self()) { 2991 Generation::oop_iterate(mr, cl); 2992 } else { 2993 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2994 Generation::oop_iterate(mr, cl); 2995 } 2996 } 2997 2998 void 2999 ConcurrentMarkSweepGeneration::oop_iterate(OopClosure* cl) { 3000 if (freelistLock()->owned_by_self()) { 3001 Generation::oop_iterate(cl); 3002 } else { 3003 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 3004 Generation::oop_iterate(cl); 3005 } 3006 } 3007 3008 void 3009 ConcurrentMarkSweepGeneration::object_iterate(ObjectClosure* cl) { 3010 if (freelistLock()->owned_by_self()) { 3011 Generation::object_iterate(cl); 3012 } else { 3013 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 3014 Generation::object_iterate(cl); 3015 } 3016 } 3017 3018 void 3019 ConcurrentMarkSweepGeneration::safe_object_iterate(ObjectClosure* cl) { 3020 if (freelistLock()->owned_by_self()) { 3021 Generation::safe_object_iterate(cl); 3022 } else { 3023 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 3024 Generation::safe_object_iterate(cl); 3025 } 3026 } 3027 3028 void 3029 ConcurrentMarkSweepGeneration::pre_adjust_pointers() { 3030 } 3031 3032 void 3033 ConcurrentMarkSweepGeneration::post_compact() { 3034 } 3035 3036 void 3037 ConcurrentMarkSweepGeneration::prepare_for_verify() { 3038 // Fix the linear allocation blocks to look like free blocks. 3039 3040 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those 3041 // are not called when the heap is verified during universe initialization and 3042 // at vm shutdown. 3043 if (freelistLock()->owned_by_self()) { 3044 cmsSpace()->prepare_for_verify(); 3045 } else { 3046 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag); 3047 cmsSpace()->prepare_for_verify(); 3048 } 3049 } 3050 3051 void 3052 ConcurrentMarkSweepGeneration::verify(bool allow_dirty /* ignored */) { 3053 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those 3054 // are not called when the heap is verified during universe initialization and 3055 // at vm shutdown. 3056 if (freelistLock()->owned_by_self()) { 3057 cmsSpace()->verify(false /* ignored */); 3058 } else { 3059 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag); 3060 cmsSpace()->verify(false /* ignored */); 3061 } 3062 } 3063 3064 void CMSCollector::verify(bool allow_dirty /* ignored */) { 3065 _cmsGen->verify(allow_dirty); 3066 _permGen->verify(allow_dirty); 3067 } 3068 3069 #ifndef PRODUCT 3070 bool CMSCollector::overflow_list_is_empty() const { 3071 assert(_num_par_pushes >= 0, "Inconsistency"); 3072 if (_overflow_list == NULL) { 3073 assert(_num_par_pushes == 0, "Inconsistency"); 3074 } 3075 return _overflow_list == NULL; 3076 } 3077 3078 // The methods verify_work_stacks_empty() and verify_overflow_empty() 3079 // merely consolidate assertion checks that appear to occur together frequently. 3080 void CMSCollector::verify_work_stacks_empty() const { 3081 assert(_markStack.isEmpty(), "Marking stack should be empty"); 3082 assert(overflow_list_is_empty(), "Overflow list should be empty"); 3083 } 3084 3085 void CMSCollector::verify_overflow_empty() const { 3086 assert(overflow_list_is_empty(), "Overflow list should be empty"); 3087 assert(no_preserved_marks(), "No preserved marks"); 3088 } 3089 #endif // PRODUCT 3090 3091 // Decide if we want to enable class unloading as part of the 3092 // ensuing concurrent GC cycle. We will collect the perm gen and 3093 // unload classes if it's the case that: 3094 // (1) an explicit gc request has been made and the flag 3095 // ExplicitGCInvokesConcurrentAndUnloadsClasses is set, OR 3096 // (2) (a) class unloading is enabled at the command line, and 3097 // (b) (i) perm gen threshold has been crossed, or 3098 // (ii) old gen is getting really full, or 3099 // (iii) the previous N CMS collections did not collect the 3100 // perm gen 3101 // NOTE: Provided there is no change in the state of the heap between 3102 // calls to this method, it should have idempotent results. Moreover, 3103 // its results should be monotonically increasing (i.e. going from 0 to 1, 3104 // but not 1 to 0) between successive calls between which the heap was 3105 // not collected. For the implementation below, it must thus rely on 3106 // the property that concurrent_cycles_since_last_unload() 3107 // will not decrease unless a collection cycle happened and that 3108 // _permGen->should_concurrent_collect() and _cmsGen->is_too_full() are 3109 // themselves also monotonic in that sense. See check_monotonicity() 3110 // below. 3111 bool CMSCollector::update_should_unload_classes() { 3112 _should_unload_classes = false; 3113 // Condition 1 above 3114 if (_full_gc_requested && ExplicitGCInvokesConcurrentAndUnloadsClasses) { 3115 _should_unload_classes = true; 3116 } else if (CMSClassUnloadingEnabled) { // Condition 2.a above 3117 // Disjuncts 2.b.(i,ii,iii) above 3118 _should_unload_classes = (concurrent_cycles_since_last_unload() >= 3119 CMSClassUnloadingMaxInterval) 3120 || _permGen->should_concurrent_collect() 3121 || _cmsGen->is_too_full(); 3122 } 3123 return _should_unload_classes; 3124 } 3125 3126 bool ConcurrentMarkSweepGeneration::is_too_full() const { 3127 bool res = should_concurrent_collect(); 3128 res = res && (occupancy() > (double)CMSIsTooFullPercentage/100.0); 3129 return res; 3130 } 3131 3132 void CMSCollector::setup_cms_unloading_and_verification_state() { 3133 const bool should_verify = VerifyBeforeGC || VerifyAfterGC || VerifyDuringGC 3134 || VerifyBeforeExit; 3135 const int rso = SharedHeap::SO_Symbols | SharedHeap::SO_Strings 3136 | SharedHeap::SO_CodeCache; 3137 3138 if (should_unload_classes()) { // Should unload classes this cycle 3139 remove_root_scanning_option(rso); // Shrink the root set appropriately 3140 set_verifying(should_verify); // Set verification state for this cycle 3141 return; // Nothing else needs to be done at this time 3142 } 3143 3144 // Not unloading classes this cycle 3145 assert(!should_unload_classes(), "Inconsitency!"); 3146 if ((!verifying() || unloaded_classes_last_cycle()) && should_verify) { 3147 // We were not verifying, or we _were_ unloading classes in the last cycle, 3148 // AND some verification options are enabled this cycle; in this case, 3149 // we must make sure that the deadness map is allocated if not already so, 3150 // and cleared (if already allocated previously -- 3151 // CMSBitMap::sizeInBits() is used to determine if it's allocated). 3152 if (perm_gen_verify_bit_map()->sizeInBits() == 0) { 3153 if (!perm_gen_verify_bit_map()->allocate(_permGen->reserved())) { 3154 warning("Failed to allocate permanent generation verification CMS Bit Map;\n" 3155 "permanent generation verification disabled"); 3156 return; // Note that we leave verification disabled, so we'll retry this 3157 // allocation next cycle. We _could_ remember this failure 3158 // and skip further attempts and permanently disable verification 3159 // attempts if that is considered more desirable. 3160 } 3161 assert(perm_gen_verify_bit_map()->covers(_permGen->reserved()), 3162 "_perm_gen_ver_bit_map inconsistency?"); 3163 } else { 3164 perm_gen_verify_bit_map()->clear_all(); 3165 } 3166 // Include symbols, strings and code cache elements to prevent their resurrection. 3167 add_root_scanning_option(rso); 3168 set_verifying(true); 3169 } else if (verifying() && !should_verify) { 3170 // We were verifying, but some verification flags got disabled. 3171 set_verifying(false); 3172 // Exclude symbols, strings and code cache elements from root scanning to 3173 // reduce IM and RM pauses. 3174 remove_root_scanning_option(rso); 3175 } 3176 } 3177 3178 3179 #ifndef PRODUCT 3180 HeapWord* CMSCollector::block_start(const void* p) const { 3181 const HeapWord* addr = (HeapWord*)p; 3182 if (_span.contains(p)) { 3183 if (_cmsGen->cmsSpace()->is_in_reserved(addr)) { 3184 return _cmsGen->cmsSpace()->block_start(p); 3185 } else { 3186 assert(_permGen->cmsSpace()->is_in_reserved(addr), 3187 "Inconsistent _span?"); 3188 return _permGen->cmsSpace()->block_start(p); 3189 } 3190 } 3191 return NULL; 3192 } 3193 #endif 3194 3195 HeapWord* 3196 ConcurrentMarkSweepGeneration::expand_and_allocate(size_t word_size, 3197 bool tlab, 3198 bool parallel) { 3199 assert(!tlab, "Can't deal with TLAB allocation"); 3200 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 3201 expand(word_size*HeapWordSize, MinHeapDeltaBytes, 3202 CMSExpansionCause::_satisfy_allocation); 3203 if (GCExpandToAllocateDelayMillis > 0) { 3204 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 3205 } 3206 return have_lock_and_allocate(word_size, tlab); 3207 } 3208 3209 // YSR: All of this generation expansion/shrinking stuff is an exact copy of 3210 // OneContigSpaceCardGeneration, which makes me wonder if we should move this 3211 // to CardGeneration and share it... 3212 bool ConcurrentMarkSweepGeneration::expand(size_t bytes, size_t expand_bytes) { 3213 return CardGeneration::expand(bytes, expand_bytes); 3214 } 3215 3216 void ConcurrentMarkSweepGeneration::expand(size_t bytes, size_t expand_bytes, 3217 CMSExpansionCause::Cause cause) 3218 { 3219 3220 bool success = expand(bytes, expand_bytes); 3221 3222 // remember why we expanded; this information is used 3223 // by shouldConcurrentCollect() when making decisions on whether to start 3224 // a new CMS cycle. 3225 if (success) { 3226 set_expansion_cause(cause); 3227 if (PrintGCDetails && Verbose) { 3228 gclog_or_tty->print_cr("Expanded CMS gen for %s", 3229 CMSExpansionCause::to_string(cause)); 3230 } 3231 } 3232 } 3233 3234 HeapWord* ConcurrentMarkSweepGeneration::expand_and_par_lab_allocate(CMSParGCThreadState* ps, size_t word_sz) { 3235 HeapWord* res = NULL; 3236 MutexLocker x(ParGCRareEvent_lock); 3237 while (true) { 3238 // Expansion by some other thread might make alloc OK now: 3239 res = ps->lab.alloc(word_sz); 3240 if (res != NULL) return res; 3241 // If there's not enough expansion space available, give up. 3242 if (_virtual_space.uncommitted_size() < (word_sz * HeapWordSize)) { 3243 return NULL; 3244 } 3245 // Otherwise, we try expansion. 3246 expand(word_sz*HeapWordSize, MinHeapDeltaBytes, 3247 CMSExpansionCause::_allocate_par_lab); 3248 // Now go around the loop and try alloc again; 3249 // A competing par_promote might beat us to the expansion space, 3250 // so we may go around the loop again if promotion fails agaion. 3251 if (GCExpandToAllocateDelayMillis > 0) { 3252 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 3253 } 3254 } 3255 } 3256 3257 3258 bool ConcurrentMarkSweepGeneration::expand_and_ensure_spooling_space( 3259 PromotionInfo* promo) { 3260 MutexLocker x(ParGCRareEvent_lock); 3261 size_t refill_size_bytes = promo->refillSize() * HeapWordSize; 3262 while (true) { 3263 // Expansion by some other thread might make alloc OK now: 3264 if (promo->ensure_spooling_space()) { 3265 assert(promo->has_spooling_space(), 3266 "Post-condition of successful ensure_spooling_space()"); 3267 return true; 3268 } 3269 // If there's not enough expansion space available, give up. 3270 if (_virtual_space.uncommitted_size() < refill_size_bytes) { 3271 return false; 3272 } 3273 // Otherwise, we try expansion. 3274 expand(refill_size_bytes, MinHeapDeltaBytes, 3275 CMSExpansionCause::_allocate_par_spooling_space); 3276 // Now go around the loop and try alloc again; 3277 // A competing allocation might beat us to the expansion space, 3278 // so we may go around the loop again if allocation fails again. 3279 if (GCExpandToAllocateDelayMillis > 0) { 3280 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 3281 } 3282 } 3283 } 3284 3285 3286 3287 void ConcurrentMarkSweepGeneration::shrink(size_t bytes) { 3288 assert_locked_or_safepoint(Heap_lock); 3289 size_t size = ReservedSpace::page_align_size_down(bytes); 3290 if (size > 0) { 3291 shrink_by(size); 3292 } 3293 } 3294 3295 bool ConcurrentMarkSweepGeneration::grow_by(size_t bytes) { 3296 assert_locked_or_safepoint(Heap_lock); 3297 bool result = _virtual_space.expand_by(bytes); 3298 if (result) { 3299 HeapWord* old_end = _cmsSpace->end(); 3300 size_t new_word_size = 3301 heap_word_size(_virtual_space.committed_size()); 3302 MemRegion mr(_cmsSpace->bottom(), new_word_size); 3303 _bts->resize(new_word_size); // resize the block offset shared array 3304 Universe::heap()->barrier_set()->resize_covered_region(mr); 3305 // Hmmmm... why doesn't CFLS::set_end verify locking? 3306 // This is quite ugly; FIX ME XXX 3307 _cmsSpace->assert_locked(freelistLock()); 3308 _cmsSpace->set_end((HeapWord*)_virtual_space.high()); 3309 3310 // update the space and generation capacity counters 3311 if (UsePerfData) { 3312 _space_counters->update_capacity(); 3313 _gen_counters->update_all(); 3314 } 3315 3316 if (Verbose && PrintGC) { 3317 size_t new_mem_size = _virtual_space.committed_size(); 3318 size_t old_mem_size = new_mem_size - bytes; 3319 gclog_or_tty->print_cr("Expanding %s from %ldK by %ldK to %ldK", 3320 name(), old_mem_size/K, bytes/K, new_mem_size/K); 3321 } 3322 } 3323 return result; 3324 } 3325 3326 bool ConcurrentMarkSweepGeneration::grow_to_reserved() { 3327 assert_locked_or_safepoint(Heap_lock); 3328 bool success = true; 3329 const size_t remaining_bytes = _virtual_space.uncommitted_size(); 3330 if (remaining_bytes > 0) { 3331 success = grow_by(remaining_bytes); 3332 DEBUG_ONLY(if (!success) warning("grow to reserved failed");) 3333 } 3334 return success; 3335 } 3336 3337 void ConcurrentMarkSweepGeneration::shrink_by(size_t bytes) { 3338 assert_locked_or_safepoint(Heap_lock); 3339 assert_lock_strong(freelistLock()); 3340 // XXX Fix when compaction is implemented. 3341 warning("Shrinking of CMS not yet implemented"); 3342 return; 3343 } 3344 3345 3346 // Simple ctor/dtor wrapper for accounting & timer chores around concurrent 3347 // phases. 3348 class CMSPhaseAccounting: public StackObj { 3349 public: 3350 CMSPhaseAccounting(CMSCollector *collector, 3351 const char *phase, 3352 bool print_cr = true); 3353 ~CMSPhaseAccounting(); 3354 3355 private: 3356 CMSCollector *_collector; 3357 const char *_phase; 3358 elapsedTimer _wallclock; 3359 bool _print_cr; 3360 3361 public: 3362 // Not MT-safe; so do not pass around these StackObj's 3363 // where they may be accessed by other threads. 3364 jlong wallclock_millis() { 3365 assert(_wallclock.is_active(), "Wall clock should not stop"); 3366 _wallclock.stop(); // to record time 3367 jlong ret = _wallclock.milliseconds(); 3368 _wallclock.start(); // restart 3369 return ret; 3370 } 3371 }; 3372 3373 CMSPhaseAccounting::CMSPhaseAccounting(CMSCollector *collector, 3374 const char *phase, 3375 bool print_cr) : 3376 _collector(collector), _phase(phase), _print_cr(print_cr) { 3377 3378 if (PrintCMSStatistics != 0) { 3379 _collector->resetYields(); 3380 } 3381 if (PrintGCDetails && PrintGCTimeStamps) { 3382 gclog_or_tty->date_stamp(PrintGCDateStamps); 3383 gclog_or_tty->stamp(); 3384 gclog_or_tty->print_cr(": [%s-concurrent-%s-start]", 3385 _collector->cmsGen()->short_name(), _phase); 3386 } 3387 _collector->resetTimer(); 3388 _wallclock.start(); 3389 _collector->startTimer(); 3390 } 3391 3392 CMSPhaseAccounting::~CMSPhaseAccounting() { 3393 assert(_wallclock.is_active(), "Wall clock should not have stopped"); 3394 _collector->stopTimer(); 3395 _wallclock.stop(); 3396 if (PrintGCDetails) { 3397 gclog_or_tty->date_stamp(PrintGCDateStamps); 3398 if (PrintGCTimeStamps) { 3399 gclog_or_tty->stamp(); 3400 gclog_or_tty->print(": "); 3401 } 3402 gclog_or_tty->print("[%s-concurrent-%s: %3.3f/%3.3f secs]", 3403 _collector->cmsGen()->short_name(), 3404 _phase, _collector->timerValue(), _wallclock.seconds()); 3405 if (_print_cr) { 3406 gclog_or_tty->print_cr(""); 3407 } 3408 if (PrintCMSStatistics != 0) { 3409 gclog_or_tty->print_cr(" (CMS-concurrent-%s yielded %d times)", _phase, 3410 _collector->yields()); 3411 } 3412 } 3413 } 3414 3415 // CMS work 3416 3417 // Checkpoint the roots into this generation from outside 3418 // this generation. [Note this initial checkpoint need only 3419 // be approximate -- we'll do a catch up phase subsequently.] 3420 void CMSCollector::checkpointRootsInitial(bool asynch) { 3421 assert(_collectorState == InitialMarking, "Wrong collector state"); 3422 check_correct_thread_executing(); 3423 ReferenceProcessor* rp = ref_processor(); 3424 SpecializationStats::clear(); 3425 assert(_restart_addr == NULL, "Control point invariant"); 3426 if (asynch) { 3427 // acquire locks for subsequent manipulations 3428 MutexLockerEx x(bitMapLock(), 3429 Mutex::_no_safepoint_check_flag); 3430 checkpointRootsInitialWork(asynch); 3431 rp->verify_no_references_recorded(); 3432 rp->enable_discovery(); // enable ("weak") refs discovery 3433 _collectorState = Marking; 3434 } else { 3435 // (Weak) Refs discovery: this is controlled from genCollectedHeap::do_collection 3436 // which recognizes if we are a CMS generation, and doesn't try to turn on 3437 // discovery; verify that they aren't meddling. 3438 assert(!rp->discovery_is_atomic(), 3439 "incorrect setting of discovery predicate"); 3440 assert(!rp->discovery_enabled(), "genCollectedHeap shouldn't control " 3441 "ref discovery for this generation kind"); 3442 // already have locks 3443 checkpointRootsInitialWork(asynch); 3444 rp->enable_discovery(); // now enable ("weak") refs discovery 3445 _collectorState = Marking; 3446 } 3447 SpecializationStats::print(); 3448 } 3449 3450 void CMSCollector::checkpointRootsInitialWork(bool asynch) { 3451 assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped"); 3452 assert(_collectorState == InitialMarking, "just checking"); 3453 3454 // If there has not been a GC[n-1] since last GC[n] cycle completed, 3455 // precede our marking with a collection of all 3456 // younger generations to keep floating garbage to a minimum. 3457 // XXX: we won't do this for now -- it's an optimization to be done later. 3458 3459 // already have locks 3460 assert_lock_strong(bitMapLock()); 3461 assert(_markBitMap.isAllClear(), "was reset at end of previous cycle"); 3462 3463 // Setup the verification and class unloading state for this 3464 // CMS collection cycle. 3465 setup_cms_unloading_and_verification_state(); 3466 3467 NOT_PRODUCT(TraceTime t("\ncheckpointRootsInitialWork", 3468 PrintGCDetails && Verbose, true, gclog_or_tty);) 3469 if (UseAdaptiveSizePolicy) { 3470 size_policy()->checkpoint_roots_initial_begin(); 3471 } 3472 3473 // Reset all the PLAB chunk arrays if necessary. 3474 if (_survivor_plab_array != NULL && !CMSPLABRecordAlways) { 3475 reset_survivor_plab_arrays(); 3476 } 3477 3478 ResourceMark rm; 3479 HandleMark hm; 3480 3481 FalseClosure falseClosure; 3482 // In the case of a synchronous collection, we will elide the 3483 // remark step, so it's important to catch all the nmethod oops 3484 // in this step. 3485 // The final 'true' flag to gen_process_strong_roots will ensure this. 3486 // If 'async' is true, we can relax the nmethod tracing. 3487 MarkRefsIntoClosure notOlder(_span, &_markBitMap); 3488 GenCollectedHeap* gch = GenCollectedHeap::heap(); 3489 3490 verify_work_stacks_empty(); 3491 verify_overflow_empty(); 3492 3493 gch->ensure_parsability(false); // fill TLABs, but no need to retire them 3494 // Update the saved marks which may affect the root scans. 3495 gch->save_marks(); 3496 3497 // weak reference processing has not started yet. 3498 ref_processor()->set_enqueuing_is_done(false); 3499 3500 { 3501 // This is not needed. DEBUG_ONLY(RememberKlassesChecker imx(true);) 3502 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;) 3503 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 3504 gch->gen_process_strong_roots(_cmsGen->level(), 3505 true, // younger gens are roots 3506 true, // activate StrongRootsScope 3507 true, // collecting perm gen 3508 SharedHeap::ScanningOption(roots_scanning_options()), 3509 ¬Older, 3510 true, // walk all of code cache if (so & SO_CodeCache) 3511 NULL); 3512 } 3513 3514 // Clear mod-union table; it will be dirtied in the prologue of 3515 // CMS generation per each younger generation collection. 3516 3517 assert(_modUnionTable.isAllClear(), 3518 "Was cleared in most recent final checkpoint phase" 3519 " or no bits are set in the gc_prologue before the start of the next " 3520 "subsequent marking phase."); 3521 3522 // Temporarily disabled, since pre/post-consumption closures don't 3523 // care about precleaned cards 3524 #if 0 3525 { 3526 MemRegion mr = MemRegion((HeapWord*)_virtual_space.low(), 3527 (HeapWord*)_virtual_space.high()); 3528 _ct->ct_bs()->preclean_dirty_cards(mr); 3529 } 3530 #endif 3531 3532 // Save the end of the used_region of the constituent generations 3533 // to be used to limit the extent of sweep in each generation. 3534 save_sweep_limits(); 3535 if (UseAdaptiveSizePolicy) { 3536 size_policy()->checkpoint_roots_initial_end(gch->gc_cause()); 3537 } 3538 verify_overflow_empty(); 3539 } 3540 3541 bool CMSCollector::markFromRoots(bool asynch) { 3542 // we might be tempted to assert that: 3543 // assert(asynch == !SafepointSynchronize::is_at_safepoint(), 3544 // "inconsistent argument?"); 3545 // However that wouldn't be right, because it's possible that 3546 // a safepoint is indeed in progress as a younger generation 3547 // stop-the-world GC happens even as we mark in this generation. 3548 assert(_collectorState == Marking, "inconsistent state?"); 3549 check_correct_thread_executing(); 3550 verify_overflow_empty(); 3551 3552 bool res; 3553 if (asynch) { 3554 3555 // Start the timers for adaptive size policy for the concurrent phases 3556 // Do it here so that the foreground MS can use the concurrent 3557 // timer since a foreground MS might has the sweep done concurrently 3558 // or STW. 3559 if (UseAdaptiveSizePolicy) { 3560 size_policy()->concurrent_marking_begin(); 3561 } 3562 3563 // Weak ref discovery note: We may be discovering weak 3564 // refs in this generation concurrent (but interleaved) with 3565 // weak ref discovery by a younger generation collector. 3566 3567 CMSTokenSyncWithLocks ts(true, bitMapLock()); 3568 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 3569 CMSPhaseAccounting pa(this, "mark", !PrintGCDetails); 3570 res = markFromRootsWork(asynch); 3571 if (res) { 3572 _collectorState = Precleaning; 3573 } else { // We failed and a foreground collection wants to take over 3574 assert(_foregroundGCIsActive, "internal state inconsistency"); 3575 assert(_restart_addr == NULL, "foreground will restart from scratch"); 3576 if (PrintGCDetails) { 3577 gclog_or_tty->print_cr("bailing out to foreground collection"); 3578 } 3579 } 3580 if (UseAdaptiveSizePolicy) { 3581 size_policy()->concurrent_marking_end(); 3582 } 3583 } else { 3584 assert(SafepointSynchronize::is_at_safepoint(), 3585 "inconsistent with asynch == false"); 3586 if (UseAdaptiveSizePolicy) { 3587 size_policy()->ms_collection_marking_begin(); 3588 } 3589 // already have locks 3590 res = markFromRootsWork(asynch); 3591 _collectorState = FinalMarking; 3592 if (UseAdaptiveSizePolicy) { 3593 GenCollectedHeap* gch = GenCollectedHeap::heap(); 3594 size_policy()->ms_collection_marking_end(gch->gc_cause()); 3595 } 3596 } 3597 verify_overflow_empty(); 3598 return res; 3599 } 3600 3601 bool CMSCollector::markFromRootsWork(bool asynch) { 3602 // iterate over marked bits in bit map, doing a full scan and mark 3603 // from these roots using the following algorithm: 3604 // . if oop is to the right of the current scan pointer, 3605 // mark corresponding bit (we'll process it later) 3606 // . else (oop is to left of current scan pointer) 3607 // push oop on marking stack 3608 // . drain the marking stack 3609 3610 // Note that when we do a marking step we need to hold the 3611 // bit map lock -- recall that direct allocation (by mutators) 3612 // and promotion (by younger generation collectors) is also 3613 // marking the bit map. [the so-called allocate live policy.] 3614 // Because the implementation of bit map marking is not 3615 // robust wrt simultaneous marking of bits in the same word, 3616 // we need to make sure that there is no such interference 3617 // between concurrent such updates. 3618 3619 // already have locks 3620 assert_lock_strong(bitMapLock()); 3621 3622 // Clear the revisit stack, just in case there are any 3623 // obsolete contents from a short-circuited previous CMS cycle. 3624 _revisitStack.reset(); 3625 verify_work_stacks_empty(); 3626 verify_overflow_empty(); 3627 assert(_revisitStack.isEmpty(), "tabula rasa"); 3628 DEBUG_ONLY(RememberKlassesChecker cmx(should_unload_classes());) 3629 bool result = false; 3630 if (CMSConcurrentMTEnabled && ConcGCThreads > 0) { 3631 result = do_marking_mt(asynch); 3632 } else { 3633 result = do_marking_st(asynch); 3634 } 3635 return result; 3636 } 3637 3638 // Forward decl 3639 class CMSConcMarkingTask; 3640 3641 class CMSConcMarkingTerminator: public ParallelTaskTerminator { 3642 CMSCollector* _collector; 3643 CMSConcMarkingTask* _task; 3644 bool _yield; 3645 protected: 3646 virtual void yield(); 3647 public: 3648 // "n_threads" is the number of threads to be terminated. 3649 // "queue_set" is a set of work queues of other threads. 3650 // "collector" is the CMS collector associated with this task terminator. 3651 // "yield" indicates whether we need the gang as a whole to yield. 3652 CMSConcMarkingTerminator(int n_threads, TaskQueueSetSuper* queue_set, 3653 CMSCollector* collector, bool yield) : 3654 ParallelTaskTerminator(n_threads, queue_set), 3655 _collector(collector), 3656 _yield(yield) { } 3657 3658 void set_task(CMSConcMarkingTask* task) { 3659 _task = task; 3660 } 3661 }; 3662 3663 // MT Concurrent Marking Task 3664 class CMSConcMarkingTask: public YieldingFlexibleGangTask { 3665 CMSCollector* _collector; 3666 YieldingFlexibleWorkGang* _workers; // the whole gang 3667 int _n_workers; // requested/desired # workers 3668 bool _asynch; 3669 bool _result; 3670 CompactibleFreeListSpace* _cms_space; 3671 CompactibleFreeListSpace* _perm_space; 3672 HeapWord* _global_finger; 3673 HeapWord* _restart_addr; 3674 3675 // Exposed here for yielding support 3676 Mutex* const _bit_map_lock; 3677 3678 // The per thread work queues, available here for stealing 3679 OopTaskQueueSet* _task_queues; 3680 CMSConcMarkingTerminator _term; 3681 3682 public: 3683 CMSConcMarkingTask(CMSCollector* collector, 3684 CompactibleFreeListSpace* cms_space, 3685 CompactibleFreeListSpace* perm_space, 3686 bool asynch, int n_workers, 3687 YieldingFlexibleWorkGang* workers, 3688 OopTaskQueueSet* task_queues): 3689 YieldingFlexibleGangTask("Concurrent marking done multi-threaded"), 3690 _collector(collector), 3691 _cms_space(cms_space), 3692 _perm_space(perm_space), 3693 _asynch(asynch), _n_workers(n_workers), _result(true), 3694 _workers(workers), _task_queues(task_queues), 3695 _term(n_workers, task_queues, _collector, asynch), 3696 _bit_map_lock(collector->bitMapLock()) 3697 { 3698 assert(n_workers <= workers->total_workers(), 3699 "Else termination won't work correctly today"); // XXX FIX ME! 3700 _requested_size = n_workers; 3701 _term.set_task(this); 3702 assert(_cms_space->bottom() < _perm_space->bottom(), 3703 "Finger incorrectly initialized below"); 3704 _restart_addr = _global_finger = _cms_space->bottom(); 3705 } 3706 3707 3708 OopTaskQueueSet* task_queues() { return _task_queues; } 3709 3710 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 3711 3712 HeapWord** global_finger_addr() { return &_global_finger; } 3713 3714 CMSConcMarkingTerminator* terminator() { return &_term; } 3715 3716 void work(int i); 3717 3718 virtual void coordinator_yield(); // stuff done by coordinator 3719 bool result() { return _result; } 3720 3721 void reset(HeapWord* ra) { 3722 assert(_global_finger >= _cms_space->end(), "Postcondition of ::work(i)"); 3723 assert(_global_finger >= _perm_space->end(), "Postcondition of ::work(i)"); 3724 assert(ra < _perm_space->end(), "ra too large"); 3725 _restart_addr = _global_finger = ra; 3726 _term.reset_for_reuse(); 3727 } 3728 3729 static bool get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 3730 OopTaskQueue* work_q); 3731 3732 private: 3733 void do_scan_and_mark(int i, CompactibleFreeListSpace* sp); 3734 void do_work_steal(int i); 3735 void bump_global_finger(HeapWord* f); 3736 }; 3737 3738 void CMSConcMarkingTerminator::yield() { 3739 if (ConcurrentMarkSweepThread::should_yield() && 3740 !_collector->foregroundGCIsActive() && 3741 _yield) { 3742 _task->yield(); 3743 } else { 3744 ParallelTaskTerminator::yield(); 3745 } 3746 } 3747 3748 //////////////////////////////////////////////////////////////// 3749 // Concurrent Marking Algorithm Sketch 3750 //////////////////////////////////////////////////////////////// 3751 // Until all tasks exhausted (both spaces): 3752 // -- claim next available chunk 3753 // -- bump global finger via CAS 3754 // -- find first object that starts in this chunk 3755 // and start scanning bitmap from that position 3756 // -- scan marked objects for oops 3757 // -- CAS-mark target, and if successful: 3758 // . if target oop is above global finger (volatile read) 3759 // nothing to do 3760 // . if target oop is in chunk and above local finger 3761 // then nothing to do 3762 // . else push on work-queue 3763 // -- Deal with possible overflow issues: 3764 // . local work-queue overflow causes stuff to be pushed on 3765 // global (common) overflow queue 3766 // . always first empty local work queue 3767 // . then get a batch of oops from global work queue if any 3768 // . then do work stealing 3769 // -- When all tasks claimed (both spaces) 3770 // and local work queue empty, 3771 // then in a loop do: 3772 // . check global overflow stack; steal a batch of oops and trace 3773 // . try to steal from other threads oif GOS is empty 3774 // . if neither is available, offer termination 3775 // -- Terminate and return result 3776 // 3777 void CMSConcMarkingTask::work(int i) { 3778 elapsedTimer _timer; 3779 ResourceMark rm; 3780 HandleMark hm; 3781 3782 DEBUG_ONLY(_collector->verify_overflow_empty();) 3783 3784 // Before we begin work, our work queue should be empty 3785 assert(work_queue(i)->size() == 0, "Expected to be empty"); 3786 // Scan the bitmap covering _cms_space, tracing through grey objects. 3787 _timer.start(); 3788 do_scan_and_mark(i, _cms_space); 3789 _timer.stop(); 3790 if (PrintCMSStatistics != 0) { 3791 gclog_or_tty->print_cr("Finished cms space scanning in %dth thread: %3.3f sec", 3792 i, _timer.seconds()); // XXX: need xxx/xxx type of notation, two timers 3793 } 3794 3795 // ... do the same for the _perm_space 3796 _timer.reset(); 3797 _timer.start(); 3798 do_scan_and_mark(i, _perm_space); 3799 _timer.stop(); 3800 if (PrintCMSStatistics != 0) { 3801 gclog_or_tty->print_cr("Finished perm space scanning in %dth thread: %3.3f sec", 3802 i, _timer.seconds()); // XXX: need xxx/xxx type of notation, two timers 3803 } 3804 3805 // ... do work stealing 3806 _timer.reset(); 3807 _timer.start(); 3808 do_work_steal(i); 3809 _timer.stop(); 3810 if (PrintCMSStatistics != 0) { 3811 gclog_or_tty->print_cr("Finished work stealing in %dth thread: %3.3f sec", 3812 i, _timer.seconds()); // XXX: need xxx/xxx type of notation, two timers 3813 } 3814 assert(_collector->_markStack.isEmpty(), "Should have been emptied"); 3815 assert(work_queue(i)->size() == 0, "Should have been emptied"); 3816 // Note that under the current task protocol, the 3817 // following assertion is true even of the spaces 3818 // expanded since the completion of the concurrent 3819 // marking. XXX This will likely change under a strict 3820 // ABORT semantics. 3821 assert(_global_finger > _cms_space->end() && 3822 _global_finger >= _perm_space->end(), 3823 "All tasks have been completed"); 3824 DEBUG_ONLY(_collector->verify_overflow_empty();) 3825 } 3826 3827 void CMSConcMarkingTask::bump_global_finger(HeapWord* f) { 3828 HeapWord* read = _global_finger; 3829 HeapWord* cur = read; 3830 while (f > read) { 3831 cur = read; 3832 read = (HeapWord*) Atomic::cmpxchg_ptr(f, &_global_finger, cur); 3833 if (cur == read) { 3834 // our cas succeeded 3835 assert(_global_finger >= f, "protocol consistency"); 3836 break; 3837 } 3838 } 3839 } 3840 3841 // This is really inefficient, and should be redone by 3842 // using (not yet available) block-read and -write interfaces to the 3843 // stack and the work_queue. XXX FIX ME !!! 3844 bool CMSConcMarkingTask::get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 3845 OopTaskQueue* work_q) { 3846 // Fast lock-free check 3847 if (ovflw_stk->length() == 0) { 3848 return false; 3849 } 3850 assert(work_q->size() == 0, "Shouldn't steal"); 3851 MutexLockerEx ml(ovflw_stk->par_lock(), 3852 Mutex::_no_safepoint_check_flag); 3853 // Grab up to 1/4 the size of the work queue 3854 size_t num = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 3855 (size_t)ParGCDesiredObjsFromOverflowList); 3856 num = MIN2(num, ovflw_stk->length()); 3857 for (int i = (int) num; i > 0; i--) { 3858 oop cur = ovflw_stk->pop(); 3859 assert(cur != NULL, "Counted wrong?"); 3860 work_q->push(cur); 3861 } 3862 return num > 0; 3863 } 3864 3865 void CMSConcMarkingTask::do_scan_and_mark(int i, CompactibleFreeListSpace* sp) { 3866 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 3867 int n_tasks = pst->n_tasks(); 3868 // We allow that there may be no tasks to do here because 3869 // we are restarting after a stack overflow. 3870 assert(pst->valid() || n_tasks == 0, "Uninitialized use?"); 3871 int nth_task = 0; 3872 3873 HeapWord* aligned_start = sp->bottom(); 3874 if (sp->used_region().contains(_restart_addr)) { 3875 // Align down to a card boundary for the start of 0th task 3876 // for this space. 3877 aligned_start = 3878 (HeapWord*)align_size_down((uintptr_t)_restart_addr, 3879 CardTableModRefBS::card_size); 3880 } 3881 3882 size_t chunk_size = sp->marking_task_size(); 3883 while (!pst->is_task_claimed(/* reference */ nth_task)) { 3884 // Having claimed the nth task in this space, 3885 // compute the chunk that it corresponds to: 3886 MemRegion span = MemRegion(aligned_start + nth_task*chunk_size, 3887 aligned_start + (nth_task+1)*chunk_size); 3888 // Try and bump the global finger via a CAS; 3889 // note that we need to do the global finger bump 3890 // _before_ taking the intersection below, because 3891 // the task corresponding to that region will be 3892 // deemed done even if the used_region() expands 3893 // because of allocation -- as it almost certainly will 3894 // during start-up while the threads yield in the 3895 // closure below. 3896 HeapWord* finger = span.end(); 3897 bump_global_finger(finger); // atomically 3898 // There are null tasks here corresponding to chunks 3899 // beyond the "top" address of the space. 3900 span = span.intersection(sp->used_region()); 3901 if (!span.is_empty()) { // Non-null task 3902 HeapWord* prev_obj; 3903 assert(!span.contains(_restart_addr) || nth_task == 0, 3904 "Inconsistency"); 3905 if (nth_task == 0) { 3906 // For the 0th task, we'll not need to compute a block_start. 3907 if (span.contains(_restart_addr)) { 3908 // In the case of a restart because of stack overflow, 3909 // we might additionally skip a chunk prefix. 3910 prev_obj = _restart_addr; 3911 } else { 3912 prev_obj = span.start(); 3913 } 3914 } else { 3915 // We want to skip the first object because 3916 // the protocol is to scan any object in its entirety 3917 // that _starts_ in this span; a fortiori, any 3918 // object starting in an earlier span is scanned 3919 // as part of an earlier claimed task. 3920 // Below we use the "careful" version of block_start 3921 // so we do not try to navigate uninitialized objects. 3922 prev_obj = sp->block_start_careful(span.start()); 3923 // Below we use a variant of block_size that uses the 3924 // Printezis bits to avoid waiting for allocated 3925 // objects to become initialized/parsable. 3926 while (prev_obj < span.start()) { 3927 size_t sz = sp->block_size_no_stall(prev_obj, _collector); 3928 if (sz > 0) { 3929 prev_obj += sz; 3930 } else { 3931 // In this case we may end up doing a bit of redundant 3932 // scanning, but that appears unavoidable, short of 3933 // locking the free list locks; see bug 6324141. 3934 break; 3935 } 3936 } 3937 } 3938 if (prev_obj < span.end()) { 3939 MemRegion my_span = MemRegion(prev_obj, span.end()); 3940 // Do the marking work within a non-empty span -- 3941 // the last argument to the constructor indicates whether the 3942 // iteration should be incremental with periodic yields. 3943 Par_MarkFromRootsClosure cl(this, _collector, my_span, 3944 &_collector->_markBitMap, 3945 work_queue(i), 3946 &_collector->_markStack, 3947 &_collector->_revisitStack, 3948 _asynch); 3949 _collector->_markBitMap.iterate(&cl, my_span.start(), my_span.end()); 3950 } // else nothing to do for this task 3951 } // else nothing to do for this task 3952 } 3953 // We'd be tempted to assert here that since there are no 3954 // more tasks left to claim in this space, the global_finger 3955 // must exceed space->top() and a fortiori space->end(). However, 3956 // that would not quite be correct because the bumping of 3957 // global_finger occurs strictly after the claiming of a task, 3958 // so by the time we reach here the global finger may not yet 3959 // have been bumped up by the thread that claimed the last 3960 // task. 3961 pst->all_tasks_completed(); 3962 } 3963 3964 class Par_ConcMarkingClosure: public Par_KlassRememberingOopClosure { 3965 private: 3966 MemRegion _span; 3967 CMSBitMap* _bit_map; 3968 CMSMarkStack* _overflow_stack; 3969 OopTaskQueue* _work_queue; 3970 protected: 3971 DO_OOP_WORK_DEFN 3972 public: 3973 Par_ConcMarkingClosure(CMSCollector* collector, OopTaskQueue* work_queue, 3974 CMSBitMap* bit_map, CMSMarkStack* overflow_stack, 3975 CMSMarkStack* revisit_stack): 3976 Par_KlassRememberingOopClosure(collector, NULL, revisit_stack), 3977 _span(_collector->_span), 3978 _work_queue(work_queue), 3979 _bit_map(bit_map), 3980 _overflow_stack(overflow_stack) 3981 { } 3982 virtual void do_oop(oop* p); 3983 virtual void do_oop(narrowOop* p); 3984 void trim_queue(size_t max); 3985 void handle_stack_overflow(HeapWord* lost); 3986 }; 3987 3988 // Grey object scanning during work stealing phase -- 3989 // the salient assumption here is that any references 3990 // that are in these stolen objects being scanned must 3991 // already have been initialized (else they would not have 3992 // been published), so we do not need to check for 3993 // uninitialized objects before pushing here. 3994 void Par_ConcMarkingClosure::do_oop(oop obj) { 3995 assert(obj->is_oop_or_null(true), "expected an oop or NULL"); 3996 HeapWord* addr = (HeapWord*)obj; 3997 // Check if oop points into the CMS generation 3998 // and is not marked 3999 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 4000 // a white object ... 4001 // If we manage to "claim" the object, by being the 4002 // first thread to mark it, then we push it on our 4003 // marking stack 4004 if (_bit_map->par_mark(addr)) { // ... now grey 4005 // push on work queue (grey set) 4006 bool simulate_overflow = false; 4007 NOT_PRODUCT( 4008 if (CMSMarkStackOverflowALot && 4009 _collector->simulate_overflow()) { 4010 // simulate a stack overflow 4011 simulate_overflow = true; 4012 } 4013 ) 4014 if (simulate_overflow || 4015 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) { 4016 // stack overflow 4017 if (PrintCMSStatistics != 0) { 4018 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 4019 SIZE_FORMAT, _overflow_stack->capacity()); 4020 } 4021 // We cannot assert that the overflow stack is full because 4022 // it may have been emptied since. 4023 assert(simulate_overflow || 4024 _work_queue->size() == _work_queue->max_elems(), 4025 "Else push should have succeeded"); 4026 handle_stack_overflow(addr); 4027 } 4028 } // Else, some other thread got there first 4029 } 4030 } 4031 4032 void Par_ConcMarkingClosure::do_oop(oop* p) { Par_ConcMarkingClosure::do_oop_work(p); } 4033 void Par_ConcMarkingClosure::do_oop(narrowOop* p) { Par_ConcMarkingClosure::do_oop_work(p); } 4034 4035 void Par_ConcMarkingClosure::trim_queue(size_t max) { 4036 while (_work_queue->size() > max) { 4037 oop new_oop; 4038 if (_work_queue->pop_local(new_oop)) { 4039 assert(new_oop->is_oop(), "Should be an oop"); 4040 assert(_bit_map->isMarked((HeapWord*)new_oop), "Grey object"); 4041 assert(_span.contains((HeapWord*)new_oop), "Not in span"); 4042 assert(new_oop->is_parsable(), "Should be parsable"); 4043 new_oop->oop_iterate(this); // do_oop() above 4044 } 4045 } 4046 } 4047 4048 // Upon stack overflow, we discard (part of) the stack, 4049 // remembering the least address amongst those discarded 4050 // in CMSCollector's _restart_address. 4051 void Par_ConcMarkingClosure::handle_stack_overflow(HeapWord* lost) { 4052 // We need to do this under a mutex to prevent other 4053 // workers from interfering with the work done below. 4054 MutexLockerEx ml(_overflow_stack->par_lock(), 4055 Mutex::_no_safepoint_check_flag); 4056 // Remember the least grey address discarded 4057 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 4058 _collector->lower_restart_addr(ra); 4059 _overflow_stack->reset(); // discard stack contents 4060 _overflow_stack->expand(); // expand the stack if possible 4061 } 4062 4063 4064 void CMSConcMarkingTask::do_work_steal(int i) { 4065 OopTaskQueue* work_q = work_queue(i); 4066 oop obj_to_scan; 4067 CMSBitMap* bm = &(_collector->_markBitMap); 4068 CMSMarkStack* ovflw = &(_collector->_markStack); 4069 CMSMarkStack* revisit = &(_collector->_revisitStack); 4070 int* seed = _collector->hash_seed(i); 4071 Par_ConcMarkingClosure cl(_collector, work_q, bm, ovflw, revisit); 4072 while (true) { 4073 cl.trim_queue(0); 4074 assert(work_q->size() == 0, "Should have been emptied above"); 4075 if (get_work_from_overflow_stack(ovflw, work_q)) { 4076 // Can't assert below because the work obtained from the 4077 // overflow stack may already have been stolen from us. 4078 // assert(work_q->size() > 0, "Work from overflow stack"); 4079 continue; 4080 } else if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 4081 assert(obj_to_scan->is_oop(), "Should be an oop"); 4082 assert(bm->isMarked((HeapWord*)obj_to_scan), "Grey object"); 4083 obj_to_scan->oop_iterate(&cl); 4084 } else if (terminator()->offer_termination()) { 4085 assert(work_q->size() == 0, "Impossible!"); 4086 break; 4087 } 4088 } 4089 } 4090 4091 // This is run by the CMS (coordinator) thread. 4092 void CMSConcMarkingTask::coordinator_yield() { 4093 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 4094 "CMS thread should hold CMS token"); 4095 DEBUG_ONLY(RememberKlassesChecker mux(false);) 4096 // First give up the locks, then yield, then re-lock 4097 // We should probably use a constructor/destructor idiom to 4098 // do this unlock/lock or modify the MutexUnlocker class to 4099 // serve our purpose. XXX 4100 assert_lock_strong(_bit_map_lock); 4101 _bit_map_lock->unlock(); 4102 ConcurrentMarkSweepThread::desynchronize(true); 4103 ConcurrentMarkSweepThread::acknowledge_yield_request(); 4104 _collector->stopTimer(); 4105 if (PrintCMSStatistics != 0) { 4106 _collector->incrementYields(); 4107 } 4108 _collector->icms_wait(); 4109 4110 // It is possible for whichever thread initiated the yield request 4111 // not to get a chance to wake up and take the bitmap lock between 4112 // this thread releasing it and reacquiring it. So, while the 4113 // should_yield() flag is on, let's sleep for a bit to give the 4114 // other thread a chance to wake up. The limit imposed on the number 4115 // of iterations is defensive, to avoid any unforseen circumstances 4116 // putting us into an infinite loop. Since it's always been this 4117 // (coordinator_yield()) method that was observed to cause the 4118 // problem, we are using a parameter (CMSCoordinatorYieldSleepCount) 4119 // which is by default non-zero. For the other seven methods that 4120 // also perform the yield operation, as are using a different 4121 // parameter (CMSYieldSleepCount) which is by default zero. This way we 4122 // can enable the sleeping for those methods too, if necessary. 4123 // See 6442774. 4124 // 4125 // We really need to reconsider the synchronization between the GC 4126 // thread and the yield-requesting threads in the future and we 4127 // should really use wait/notify, which is the recommended 4128 // way of doing this type of interaction. Additionally, we should 4129 // consolidate the eight methods that do the yield operation and they 4130 // are almost identical into one for better maintenability and 4131 // readability. See 6445193. 4132 // 4133 // Tony 2006.06.29 4134 for (unsigned i = 0; i < CMSCoordinatorYieldSleepCount && 4135 ConcurrentMarkSweepThread::should_yield() && 4136 !CMSCollector::foregroundGCIsActive(); ++i) { 4137 os::sleep(Thread::current(), 1, false); 4138 ConcurrentMarkSweepThread::acknowledge_yield_request(); 4139 } 4140 4141 ConcurrentMarkSweepThread::synchronize(true); 4142 _bit_map_lock->lock_without_safepoint_check(); 4143 _collector->startTimer(); 4144 } 4145 4146 bool CMSCollector::do_marking_mt(bool asynch) { 4147 assert(ConcGCThreads > 0 && conc_workers() != NULL, "precondition"); 4148 // In the future this would be determined ergonomically, based 4149 // on #cpu's, # active mutator threads (and load), and mutation rate. 4150 int num_workers = ConcGCThreads; 4151 4152 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 4153 CompactibleFreeListSpace* perm_space = _permGen->cmsSpace(); 4154 4155 CMSConcMarkingTask tsk(this, cms_space, perm_space, 4156 asynch, num_workers /* number requested XXX */, 4157 conc_workers(), task_queues()); 4158 4159 // Since the actual number of workers we get may be different 4160 // from the number we requested above, do we need to do anything different 4161 // below? In particular, may be we need to subclass the SequantialSubTasksDone 4162 // class?? XXX 4163 cms_space ->initialize_sequential_subtasks_for_marking(num_workers); 4164 perm_space->initialize_sequential_subtasks_for_marking(num_workers); 4165 4166 // Refs discovery is already non-atomic. 4167 assert(!ref_processor()->discovery_is_atomic(), "Should be non-atomic"); 4168 // Mutate the Refs discovery so it is MT during the 4169 // multi-threaded marking phase. 4170 ReferenceProcessorMTMutator mt(ref_processor(), num_workers > 1); 4171 DEBUG_ONLY(RememberKlassesChecker cmx(should_unload_classes());) 4172 conc_workers()->start_task(&tsk); 4173 while (tsk.yielded()) { 4174 tsk.coordinator_yield(); 4175 conc_workers()->continue_task(&tsk); 4176 } 4177 // If the task was aborted, _restart_addr will be non-NULL 4178 assert(tsk.completed() || _restart_addr != NULL, "Inconsistency"); 4179 while (_restart_addr != NULL) { 4180 // XXX For now we do not make use of ABORTED state and have not 4181 // yet implemented the right abort semantics (even in the original 4182 // single-threaded CMS case). That needs some more investigation 4183 // and is deferred for now; see CR# TBF. 07252005YSR. XXX 4184 assert(!CMSAbortSemantics || tsk.aborted(), "Inconsistency"); 4185 // If _restart_addr is non-NULL, a marking stack overflow 4186 // occurred; we need to do a fresh marking iteration from the 4187 // indicated restart address. 4188 if (_foregroundGCIsActive && asynch) { 4189 // We may be running into repeated stack overflows, having 4190 // reached the limit of the stack size, while making very 4191 // slow forward progress. It may be best to bail out and 4192 // let the foreground collector do its job. 4193 // Clear _restart_addr, so that foreground GC 4194 // works from scratch. This avoids the headache of 4195 // a "rescan" which would otherwise be needed because 4196 // of the dirty mod union table & card table. 4197 _restart_addr = NULL; 4198 return false; 4199 } 4200 // Adjust the task to restart from _restart_addr 4201 tsk.reset(_restart_addr); 4202 cms_space ->initialize_sequential_subtasks_for_marking(num_workers, 4203 _restart_addr); 4204 perm_space->initialize_sequential_subtasks_for_marking(num_workers, 4205 _restart_addr); 4206 _restart_addr = NULL; 4207 // Get the workers going again 4208 conc_workers()->start_task(&tsk); 4209 while (tsk.yielded()) { 4210 tsk.coordinator_yield(); 4211 conc_workers()->continue_task(&tsk); 4212 } 4213 } 4214 assert(tsk.completed(), "Inconsistency"); 4215 assert(tsk.result() == true, "Inconsistency"); 4216 return true; 4217 } 4218 4219 bool CMSCollector::do_marking_st(bool asynch) { 4220 ResourceMark rm; 4221 HandleMark hm; 4222 4223 MarkFromRootsClosure markFromRootsClosure(this, _span, &_markBitMap, 4224 &_markStack, &_revisitStack, CMSYield && asynch); 4225 // the last argument to iterate indicates whether the iteration 4226 // should be incremental with periodic yields. 4227 _markBitMap.iterate(&markFromRootsClosure); 4228 // If _restart_addr is non-NULL, a marking stack overflow 4229 // occurred; we need to do a fresh iteration from the 4230 // indicated restart address. 4231 while (_restart_addr != NULL) { 4232 if (_foregroundGCIsActive && asynch) { 4233 // We may be running into repeated stack overflows, having 4234 // reached the limit of the stack size, while making very 4235 // slow forward progress. It may be best to bail out and 4236 // let the foreground collector do its job. 4237 // Clear _restart_addr, so that foreground GC 4238 // works from scratch. This avoids the headache of 4239 // a "rescan" which would otherwise be needed because 4240 // of the dirty mod union table & card table. 4241 _restart_addr = NULL; 4242 return false; // indicating failure to complete marking 4243 } 4244 // Deal with stack overflow: 4245 // we restart marking from _restart_addr 4246 HeapWord* ra = _restart_addr; 4247 markFromRootsClosure.reset(ra); 4248 _restart_addr = NULL; 4249 _markBitMap.iterate(&markFromRootsClosure, ra, _span.end()); 4250 } 4251 return true; 4252 } 4253 4254 void CMSCollector::preclean() { 4255 check_correct_thread_executing(); 4256 assert(Thread::current()->is_ConcurrentGC_thread(), "Wrong thread"); 4257 verify_work_stacks_empty(); 4258 verify_overflow_empty(); 4259 _abort_preclean = false; 4260 if (CMSPrecleaningEnabled) { 4261 _eden_chunk_index = 0; 4262 size_t used = get_eden_used(); 4263 size_t capacity = get_eden_capacity(); 4264 // Don't start sampling unless we will get sufficiently 4265 // many samples. 4266 if (used < (capacity/(CMSScheduleRemarkSamplingRatio * 100) 4267 * CMSScheduleRemarkEdenPenetration)) { 4268 _start_sampling = true; 4269 } else { 4270 _start_sampling = false; 4271 } 4272 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 4273 CMSPhaseAccounting pa(this, "preclean", !PrintGCDetails); 4274 preclean_work(CMSPrecleanRefLists1, CMSPrecleanSurvivors1); 4275 } 4276 CMSTokenSync x(true); // is cms thread 4277 if (CMSPrecleaningEnabled) { 4278 sample_eden(); 4279 _collectorState = AbortablePreclean; 4280 } else { 4281 _collectorState = FinalMarking; 4282 } 4283 verify_work_stacks_empty(); 4284 verify_overflow_empty(); 4285 } 4286 4287 // Try and schedule the remark such that young gen 4288 // occupancy is CMSScheduleRemarkEdenPenetration %. 4289 void CMSCollector::abortable_preclean() { 4290 check_correct_thread_executing(); 4291 assert(CMSPrecleaningEnabled, "Inconsistent control state"); 4292 assert(_collectorState == AbortablePreclean, "Inconsistent control state"); 4293 4294 // If Eden's current occupancy is below this threshold, 4295 // immediately schedule the remark; else preclean 4296 // past the next scavenge in an effort to 4297 // schedule the pause as described avove. By choosing 4298 // CMSScheduleRemarkEdenSizeThreshold >= max eden size 4299 // we will never do an actual abortable preclean cycle. 4300 if (get_eden_used() > CMSScheduleRemarkEdenSizeThreshold) { 4301 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 4302 CMSPhaseAccounting pa(this, "abortable-preclean", !PrintGCDetails); 4303 // We need more smarts in the abortable preclean 4304 // loop below to deal with cases where allocation 4305 // in young gen is very very slow, and our precleaning 4306 // is running a losing race against a horde of 4307 // mutators intent on flooding us with CMS updates 4308 // (dirty cards). 4309 // One, admittedly dumb, strategy is to give up 4310 // after a certain number of abortable precleaning loops 4311 // or after a certain maximum time. We want to make 4312 // this smarter in the next iteration. 4313 // XXX FIX ME!!! YSR 4314 size_t loops = 0, workdone = 0, cumworkdone = 0, waited = 0; 4315 while (!(should_abort_preclean() || 4316 ConcurrentMarkSweepThread::should_terminate())) { 4317 workdone = preclean_work(CMSPrecleanRefLists2, CMSPrecleanSurvivors2); 4318 cumworkdone += workdone; 4319 loops++; 4320 // Voluntarily terminate abortable preclean phase if we have 4321 // been at it for too long. 4322 if ((CMSMaxAbortablePrecleanLoops != 0) && 4323 loops >= CMSMaxAbortablePrecleanLoops) { 4324 if (PrintGCDetails) { 4325 gclog_or_tty->print(" CMS: abort preclean due to loops "); 4326 } 4327 break; 4328 } 4329 if (pa.wallclock_millis() > CMSMaxAbortablePrecleanTime) { 4330 if (PrintGCDetails) { 4331 gclog_or_tty->print(" CMS: abort preclean due to time "); 4332 } 4333 break; 4334 } 4335 // If we are doing little work each iteration, we should 4336 // take a short break. 4337 if (workdone < CMSAbortablePrecleanMinWorkPerIteration) { 4338 // Sleep for some time, waiting for work to accumulate 4339 stopTimer(); 4340 cmsThread()->wait_on_cms_lock(CMSAbortablePrecleanWaitMillis); 4341 startTimer(); 4342 waited++; 4343 } 4344 } 4345 if (PrintCMSStatistics > 0) { 4346 gclog_or_tty->print(" [%d iterations, %d waits, %d cards)] ", 4347 loops, waited, cumworkdone); 4348 } 4349 } 4350 CMSTokenSync x(true); // is cms thread 4351 if (_collectorState != Idling) { 4352 assert(_collectorState == AbortablePreclean, 4353 "Spontaneous state transition?"); 4354 _collectorState = FinalMarking; 4355 } // Else, a foreground collection completed this CMS cycle. 4356 return; 4357 } 4358 4359 // Respond to an Eden sampling opportunity 4360 void CMSCollector::sample_eden() { 4361 // Make sure a young gc cannot sneak in between our 4362 // reading and recording of a sample. 4363 assert(Thread::current()->is_ConcurrentGC_thread(), 4364 "Only the cms thread may collect Eden samples"); 4365 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 4366 "Should collect samples while holding CMS token"); 4367 if (!_start_sampling) { 4368 return; 4369 } 4370 if (_eden_chunk_array) { 4371 if (_eden_chunk_index < _eden_chunk_capacity) { 4372 _eden_chunk_array[_eden_chunk_index] = *_top_addr; // take sample 4373 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr, 4374 "Unexpected state of Eden"); 4375 // We'd like to check that what we just sampled is an oop-start address; 4376 // however, we cannot do that here since the object may not yet have been 4377 // initialized. So we'll instead do the check when we _use_ this sample 4378 // later. 4379 if (_eden_chunk_index == 0 || 4380 (pointer_delta(_eden_chunk_array[_eden_chunk_index], 4381 _eden_chunk_array[_eden_chunk_index-1]) 4382 >= CMSSamplingGrain)) { 4383 _eden_chunk_index++; // commit sample 4384 } 4385 } 4386 } 4387 if ((_collectorState == AbortablePreclean) && !_abort_preclean) { 4388 size_t used = get_eden_used(); 4389 size_t capacity = get_eden_capacity(); 4390 assert(used <= capacity, "Unexpected state of Eden"); 4391 if (used > (capacity/100 * CMSScheduleRemarkEdenPenetration)) { 4392 _abort_preclean = true; 4393 } 4394 } 4395 } 4396 4397 4398 size_t CMSCollector::preclean_work(bool clean_refs, bool clean_survivor) { 4399 assert(_collectorState == Precleaning || 4400 _collectorState == AbortablePreclean, "incorrect state"); 4401 ResourceMark rm; 4402 HandleMark hm; 4403 // Do one pass of scrubbing the discovered reference lists 4404 // to remove any reference objects with strongly-reachable 4405 // referents. 4406 if (clean_refs) { 4407 ReferenceProcessor* rp = ref_processor(); 4408 CMSPrecleanRefsYieldClosure yield_cl(this); 4409 assert(rp->span().equals(_span), "Spans should be equal"); 4410 CMSKeepAliveClosure keep_alive(this, _span, &_markBitMap, 4411 &_markStack, &_revisitStack, 4412 true /* preclean */); 4413 CMSDrainMarkingStackClosure complete_trace(this, 4414 _span, &_markBitMap, &_markStack, 4415 &keep_alive, true /* preclean */); 4416 4417 // We don't want this step to interfere with a young 4418 // collection because we don't want to take CPU 4419 // or memory bandwidth away from the young GC threads 4420 // (which may be as many as there are CPUs). 4421 // Note that we don't need to protect ourselves from 4422 // interference with mutators because they can't 4423 // manipulate the discovered reference lists nor affect 4424 // the computed reachability of the referents, the 4425 // only properties manipulated by the precleaning 4426 // of these reference lists. 4427 stopTimer(); 4428 CMSTokenSyncWithLocks x(true /* is cms thread */, 4429 bitMapLock()); 4430 startTimer(); 4431 sample_eden(); 4432 4433 // The following will yield to allow foreground 4434 // collection to proceed promptly. XXX YSR: 4435 // The code in this method may need further 4436 // tweaking for better performance and some restructuring 4437 // for cleaner interfaces. 4438 rp->preclean_discovered_references( 4439 rp->is_alive_non_header(), &keep_alive, &complete_trace, 4440 &yield_cl, should_unload_classes()); 4441 } 4442 4443 if (clean_survivor) { // preclean the active survivor space(s) 4444 assert(_young_gen->kind() == Generation::DefNew || 4445 _young_gen->kind() == Generation::ParNew || 4446 _young_gen->kind() == Generation::ASParNew, 4447 "incorrect type for cast"); 4448 DefNewGeneration* dng = (DefNewGeneration*)_young_gen; 4449 PushAndMarkClosure pam_cl(this, _span, ref_processor(), 4450 &_markBitMap, &_modUnionTable, 4451 &_markStack, &_revisitStack, 4452 true /* precleaning phase */); 4453 stopTimer(); 4454 CMSTokenSyncWithLocks ts(true /* is cms thread */, 4455 bitMapLock()); 4456 startTimer(); 4457 unsigned int before_count = 4458 GenCollectedHeap::heap()->total_collections(); 4459 SurvivorSpacePrecleanClosure 4460 sss_cl(this, _span, &_markBitMap, &_markStack, 4461 &pam_cl, before_count, CMSYield); 4462 DEBUG_ONLY(RememberKlassesChecker mx(should_unload_classes());) 4463 dng->from()->object_iterate_careful(&sss_cl); 4464 dng->to()->object_iterate_careful(&sss_cl); 4465 } 4466 MarkRefsIntoAndScanClosure 4467 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable, 4468 &_markStack, &_revisitStack, this, CMSYield, 4469 true /* precleaning phase */); 4470 // CAUTION: The following closure has persistent state that may need to 4471 // be reset upon a decrease in the sequence of addresses it 4472 // processes. 4473 ScanMarkedObjectsAgainCarefullyClosure 4474 smoac_cl(this, _span, 4475 &_markBitMap, &_markStack, &_revisitStack, &mrias_cl, CMSYield); 4476 4477 // Preclean dirty cards in ModUnionTable and CardTable using 4478 // appropriate convergence criterion; 4479 // repeat CMSPrecleanIter times unless we find that 4480 // we are losing. 4481 assert(CMSPrecleanIter < 10, "CMSPrecleanIter is too large"); 4482 assert(CMSPrecleanNumerator < CMSPrecleanDenominator, 4483 "Bad convergence multiplier"); 4484 assert(CMSPrecleanThreshold >= 100, 4485 "Unreasonably low CMSPrecleanThreshold"); 4486 4487 size_t numIter, cumNumCards, lastNumCards, curNumCards; 4488 for (numIter = 0, cumNumCards = lastNumCards = curNumCards = 0; 4489 numIter < CMSPrecleanIter; 4490 numIter++, lastNumCards = curNumCards, cumNumCards += curNumCards) { 4491 curNumCards = preclean_mod_union_table(_cmsGen, &smoac_cl); 4492 if (CMSPermGenPrecleaningEnabled) { 4493 curNumCards += preclean_mod_union_table(_permGen, &smoac_cl); 4494 } 4495 if (Verbose && PrintGCDetails) { 4496 gclog_or_tty->print(" (modUnionTable: %d cards)", curNumCards); 4497 } 4498 // Either there are very few dirty cards, so re-mark 4499 // pause will be small anyway, or our pre-cleaning isn't 4500 // that much faster than the rate at which cards are being 4501 // dirtied, so we might as well stop and re-mark since 4502 // precleaning won't improve our re-mark time by much. 4503 if (curNumCards <= CMSPrecleanThreshold || 4504 (numIter > 0 && 4505 (curNumCards * CMSPrecleanDenominator > 4506 lastNumCards * CMSPrecleanNumerator))) { 4507 numIter++; 4508 cumNumCards += curNumCards; 4509 break; 4510 } 4511 } 4512 curNumCards = preclean_card_table(_cmsGen, &smoac_cl); 4513 if (CMSPermGenPrecleaningEnabled) { 4514 curNumCards += preclean_card_table(_permGen, &smoac_cl); 4515 } 4516 cumNumCards += curNumCards; 4517 if (PrintGCDetails && PrintCMSStatistics != 0) { 4518 gclog_or_tty->print_cr(" (cardTable: %d cards, re-scanned %d cards, %d iterations)", 4519 curNumCards, cumNumCards, numIter); 4520 } 4521 return cumNumCards; // as a measure of useful work done 4522 } 4523 4524 // PRECLEANING NOTES: 4525 // Precleaning involves: 4526 // . reading the bits of the modUnionTable and clearing the set bits. 4527 // . For the cards corresponding to the set bits, we scan the 4528 // objects on those cards. This means we need the free_list_lock 4529 // so that we can safely iterate over the CMS space when scanning 4530 // for oops. 4531 // . When we scan the objects, we'll be both reading and setting 4532 // marks in the marking bit map, so we'll need the marking bit map. 4533 // . For protecting _collector_state transitions, we take the CGC_lock. 4534 // Note that any races in the reading of of card table entries by the 4535 // CMS thread on the one hand and the clearing of those entries by the 4536 // VM thread or the setting of those entries by the mutator threads on the 4537 // other are quite benign. However, for efficiency it makes sense to keep 4538 // the VM thread from racing with the CMS thread while the latter is 4539 // dirty card info to the modUnionTable. We therefore also use the 4540 // CGC_lock to protect the reading of the card table and the mod union 4541 // table by the CM thread. 4542 // . We run concurrently with mutator updates, so scanning 4543 // needs to be done carefully -- we should not try to scan 4544 // potentially uninitialized objects. 4545 // 4546 // Locking strategy: While holding the CGC_lock, we scan over and 4547 // reset a maximal dirty range of the mod union / card tables, then lock 4548 // the free_list_lock and bitmap lock to do a full marking, then 4549 // release these locks; and repeat the cycle. This allows for a 4550 // certain amount of fairness in the sharing of these locks between 4551 // the CMS collector on the one hand, and the VM thread and the 4552 // mutators on the other. 4553 4554 // NOTE: preclean_mod_union_table() and preclean_card_table() 4555 // further below are largely identical; if you need to modify 4556 // one of these methods, please check the other method too. 4557 4558 size_t CMSCollector::preclean_mod_union_table( 4559 ConcurrentMarkSweepGeneration* gen, 4560 ScanMarkedObjectsAgainCarefullyClosure* cl) { 4561 verify_work_stacks_empty(); 4562 verify_overflow_empty(); 4563 4564 // Turn off checking for this method but turn it back on 4565 // selectively. There are yield points in this method 4566 // but it is difficult to turn the checking off just around 4567 // the yield points. It is simpler to selectively turn 4568 // it on. 4569 DEBUG_ONLY(RememberKlassesChecker mux(false);) 4570 4571 // strategy: starting with the first card, accumulate contiguous 4572 // ranges of dirty cards; clear these cards, then scan the region 4573 // covered by these cards. 4574 4575 // Since all of the MUT is committed ahead, we can just use 4576 // that, in case the generations expand while we are precleaning. 4577 // It might also be fine to just use the committed part of the 4578 // generation, but we might potentially miss cards when the 4579 // generation is rapidly expanding while we are in the midst 4580 // of precleaning. 4581 HeapWord* startAddr = gen->reserved().start(); 4582 HeapWord* endAddr = gen->reserved().end(); 4583 4584 cl->setFreelistLock(gen->freelistLock()); // needed for yielding 4585 4586 size_t numDirtyCards, cumNumDirtyCards; 4587 HeapWord *nextAddr, *lastAddr; 4588 for (cumNumDirtyCards = numDirtyCards = 0, 4589 nextAddr = lastAddr = startAddr; 4590 nextAddr < endAddr; 4591 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 4592 4593 ResourceMark rm; 4594 HandleMark hm; 4595 4596 MemRegion dirtyRegion; 4597 { 4598 stopTimer(); 4599 // Potential yield point 4600 CMSTokenSync ts(true); 4601 startTimer(); 4602 sample_eden(); 4603 // Get dirty region starting at nextOffset (inclusive), 4604 // simultaneously clearing it. 4605 dirtyRegion = 4606 _modUnionTable.getAndClearMarkedRegion(nextAddr, endAddr); 4607 assert(dirtyRegion.start() >= nextAddr, 4608 "returned region inconsistent?"); 4609 } 4610 // Remember where the next search should begin. 4611 // The returned region (if non-empty) is a right open interval, 4612 // so lastOffset is obtained from the right end of that 4613 // interval. 4614 lastAddr = dirtyRegion.end(); 4615 // Should do something more transparent and less hacky XXX 4616 numDirtyCards = 4617 _modUnionTable.heapWordDiffToOffsetDiff(dirtyRegion.word_size()); 4618 4619 // We'll scan the cards in the dirty region (with periodic 4620 // yields for foreground GC as needed). 4621 if (!dirtyRegion.is_empty()) { 4622 assert(numDirtyCards > 0, "consistency check"); 4623 HeapWord* stop_point = NULL; 4624 stopTimer(); 4625 // Potential yield point 4626 CMSTokenSyncWithLocks ts(true, gen->freelistLock(), 4627 bitMapLock()); 4628 startTimer(); 4629 { 4630 verify_work_stacks_empty(); 4631 verify_overflow_empty(); 4632 sample_eden(); 4633 DEBUG_ONLY(RememberKlassesChecker mx(should_unload_classes());) 4634 stop_point = 4635 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 4636 } 4637 if (stop_point != NULL) { 4638 // The careful iteration stopped early either because it found an 4639 // uninitialized object, or because we were in the midst of an 4640 // "abortable preclean", which should now be aborted. Redirty 4641 // the bits corresponding to the partially-scanned or unscanned 4642 // cards. We'll either restart at the next block boundary or 4643 // abort the preclean. 4644 assert((CMSPermGenPrecleaningEnabled && (gen == _permGen)) || 4645 (_collectorState == AbortablePreclean && should_abort_preclean()), 4646 "Unparsable objects should only be in perm gen."); 4647 _modUnionTable.mark_range(MemRegion(stop_point, dirtyRegion.end())); 4648 if (should_abort_preclean()) { 4649 break; // out of preclean loop 4650 } else { 4651 // Compute the next address at which preclean should pick up; 4652 // might need bitMapLock in order to read P-bits. 4653 lastAddr = next_card_start_after_block(stop_point); 4654 } 4655 } 4656 } else { 4657 assert(lastAddr == endAddr, "consistency check"); 4658 assert(numDirtyCards == 0, "consistency check"); 4659 break; 4660 } 4661 } 4662 verify_work_stacks_empty(); 4663 verify_overflow_empty(); 4664 return cumNumDirtyCards; 4665 } 4666 4667 // NOTE: preclean_mod_union_table() above and preclean_card_table() 4668 // below are largely identical; if you need to modify 4669 // one of these methods, please check the other method too. 4670 4671 size_t CMSCollector::preclean_card_table(ConcurrentMarkSweepGeneration* gen, 4672 ScanMarkedObjectsAgainCarefullyClosure* cl) { 4673 // strategy: it's similar to precleamModUnionTable above, in that 4674 // we accumulate contiguous ranges of dirty cards, mark these cards 4675 // precleaned, then scan the region covered by these cards. 4676 HeapWord* endAddr = (HeapWord*)(gen->_virtual_space.high()); 4677 HeapWord* startAddr = (HeapWord*)(gen->_virtual_space.low()); 4678 4679 cl->setFreelistLock(gen->freelistLock()); // needed for yielding 4680 4681 size_t numDirtyCards, cumNumDirtyCards; 4682 HeapWord *lastAddr, *nextAddr; 4683 4684 for (cumNumDirtyCards = numDirtyCards = 0, 4685 nextAddr = lastAddr = startAddr; 4686 nextAddr < endAddr; 4687 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 4688 4689 ResourceMark rm; 4690 HandleMark hm; 4691 4692 MemRegion dirtyRegion; 4693 { 4694 // See comments in "Precleaning notes" above on why we 4695 // do this locking. XXX Could the locking overheads be 4696 // too high when dirty cards are sparse? [I don't think so.] 4697 stopTimer(); 4698 CMSTokenSync x(true); // is cms thread 4699 startTimer(); 4700 sample_eden(); 4701 // Get and clear dirty region from card table 4702 dirtyRegion = _ct->ct_bs()->dirty_card_range_after_reset( 4703 MemRegion(nextAddr, endAddr), 4704 true, 4705 CardTableModRefBS::precleaned_card_val()); 4706 4707 assert(dirtyRegion.start() >= nextAddr, 4708 "returned region inconsistent?"); 4709 } 4710 lastAddr = dirtyRegion.end(); 4711 numDirtyCards = 4712 dirtyRegion.word_size()/CardTableModRefBS::card_size_in_words; 4713 4714 if (!dirtyRegion.is_empty()) { 4715 stopTimer(); 4716 CMSTokenSyncWithLocks ts(true, gen->freelistLock(), bitMapLock()); 4717 startTimer(); 4718 sample_eden(); 4719 verify_work_stacks_empty(); 4720 verify_overflow_empty(); 4721 DEBUG_ONLY(RememberKlassesChecker mx(should_unload_classes());) 4722 HeapWord* stop_point = 4723 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 4724 if (stop_point != NULL) { 4725 // The careful iteration stopped early because it found an 4726 // uninitialized object. Redirty the bits corresponding to the 4727 // partially-scanned or unscanned cards, and start again at the 4728 // next block boundary. 4729 assert(CMSPermGenPrecleaningEnabled || 4730 (_collectorState == AbortablePreclean && should_abort_preclean()), 4731 "Unparsable objects should only be in perm gen."); 4732 _ct->ct_bs()->invalidate(MemRegion(stop_point, dirtyRegion.end())); 4733 if (should_abort_preclean()) { 4734 break; // out of preclean loop 4735 } else { 4736 // Compute the next address at which preclean should pick up. 4737 lastAddr = next_card_start_after_block(stop_point); 4738 } 4739 } 4740 } else { 4741 break; 4742 } 4743 } 4744 verify_work_stacks_empty(); 4745 verify_overflow_empty(); 4746 return cumNumDirtyCards; 4747 } 4748 4749 void CMSCollector::checkpointRootsFinal(bool asynch, 4750 bool clear_all_soft_refs, bool init_mark_was_synchronous) { 4751 assert(_collectorState == FinalMarking, "incorrect state transition?"); 4752 check_correct_thread_executing(); 4753 // world is stopped at this checkpoint 4754 assert(SafepointSynchronize::is_at_safepoint(), 4755 "world should be stopped"); 4756 verify_work_stacks_empty(); 4757 verify_overflow_empty(); 4758 4759 SpecializationStats::clear(); 4760 if (PrintGCDetails) { 4761 gclog_or_tty->print("[YG occupancy: "SIZE_FORMAT" K ("SIZE_FORMAT" K)]", 4762 _young_gen->used() / K, 4763 _young_gen->capacity() / K); 4764 } 4765 if (asynch) { 4766 if (CMSScavengeBeforeRemark) { 4767 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4768 // Temporarily set flag to false, GCH->do_collection will 4769 // expect it to be false and set to true 4770 FlagSetting fl(gch->_is_gc_active, false); 4771 NOT_PRODUCT(TraceTime t("Scavenge-Before-Remark", 4772 PrintGCDetails && Verbose, true, gclog_or_tty);) 4773 int level = _cmsGen->level() - 1; 4774 if (level >= 0) { 4775 gch->do_collection(true, // full (i.e. force, see below) 4776 false, // !clear_all_soft_refs 4777 0, // size 4778 false, // is_tlab 4779 level // max_level 4780 ); 4781 } 4782 } 4783 FreelistLocker x(this); 4784 MutexLockerEx y(bitMapLock(), 4785 Mutex::_no_safepoint_check_flag); 4786 assert(!init_mark_was_synchronous, "but that's impossible!"); 4787 checkpointRootsFinalWork(asynch, clear_all_soft_refs, false); 4788 } else { 4789 // already have all the locks 4790 checkpointRootsFinalWork(asynch, clear_all_soft_refs, 4791 init_mark_was_synchronous); 4792 } 4793 verify_work_stacks_empty(); 4794 verify_overflow_empty(); 4795 SpecializationStats::print(); 4796 } 4797 4798 void CMSCollector::checkpointRootsFinalWork(bool asynch, 4799 bool clear_all_soft_refs, bool init_mark_was_synchronous) { 4800 4801 NOT_PRODUCT(TraceTime tr("checkpointRootsFinalWork", PrintGCDetails, false, gclog_or_tty);) 4802 4803 assert(haveFreelistLocks(), "must have free list locks"); 4804 assert_lock_strong(bitMapLock()); 4805 4806 if (UseAdaptiveSizePolicy) { 4807 size_policy()->checkpoint_roots_final_begin(); 4808 } 4809 4810 ResourceMark rm; 4811 HandleMark hm; 4812 4813 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4814 4815 if (should_unload_classes()) { 4816 CodeCache::gc_prologue(); 4817 } 4818 assert(haveFreelistLocks(), "must have free list locks"); 4819 assert_lock_strong(bitMapLock()); 4820 4821 DEBUG_ONLY(RememberKlassesChecker fmx(should_unload_classes());) 4822 if (!init_mark_was_synchronous) { 4823 // We might assume that we need not fill TLAB's when 4824 // CMSScavengeBeforeRemark is set, because we may have just done 4825 // a scavenge which would have filled all TLAB's -- and besides 4826 // Eden would be empty. This however may not always be the case -- 4827 // for instance although we asked for a scavenge, it may not have 4828 // happened because of a JNI critical section. We probably need 4829 // a policy for deciding whether we can in that case wait until 4830 // the critical section releases and then do the remark following 4831 // the scavenge, and skip it here. In the absence of that policy, 4832 // or of an indication of whether the scavenge did indeed occur, 4833 // we cannot rely on TLAB's having been filled and must do 4834 // so here just in case a scavenge did not happen. 4835 gch->ensure_parsability(false); // fill TLAB's, but no need to retire them 4836 // Update the saved marks which may affect the root scans. 4837 gch->save_marks(); 4838 4839 { 4840 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;) 4841 4842 // Note on the role of the mod union table: 4843 // Since the marker in "markFromRoots" marks concurrently with 4844 // mutators, it is possible for some reachable objects not to have been 4845 // scanned. For instance, an only reference to an object A was 4846 // placed in object B after the marker scanned B. Unless B is rescanned, 4847 // A would be collected. Such updates to references in marked objects 4848 // are detected via the mod union table which is the set of all cards 4849 // dirtied since the first checkpoint in this GC cycle and prior to 4850 // the most recent young generation GC, minus those cleaned up by the 4851 // concurrent precleaning. 4852 if (CMSParallelRemarkEnabled && ParallelGCThreads > 0) { 4853 TraceTime t("Rescan (parallel) ", PrintGCDetails, false, gclog_or_tty); 4854 do_remark_parallel(); 4855 } else { 4856 TraceTime t("Rescan (non-parallel) ", PrintGCDetails, false, 4857 gclog_or_tty); 4858 do_remark_non_parallel(); 4859 } 4860 } 4861 } else { 4862 assert(!asynch, "Can't have init_mark_was_synchronous in asynch mode"); 4863 // The initial mark was stop-world, so there's no rescanning to 4864 // do; go straight on to the next step below. 4865 } 4866 verify_work_stacks_empty(); 4867 verify_overflow_empty(); 4868 4869 { 4870 NOT_PRODUCT(TraceTime ts("refProcessingWork", PrintGCDetails, false, gclog_or_tty);) 4871 refProcessingWork(asynch, clear_all_soft_refs); 4872 } 4873 verify_work_stacks_empty(); 4874 verify_overflow_empty(); 4875 4876 if (should_unload_classes()) { 4877 CodeCache::gc_epilogue(); 4878 } 4879 4880 // If we encountered any (marking stack / work queue) overflow 4881 // events during the current CMS cycle, take appropriate 4882 // remedial measures, where possible, so as to try and avoid 4883 // recurrence of that condition. 4884 assert(_markStack.isEmpty(), "No grey objects"); 4885 size_t ser_ovflw = _ser_pmc_remark_ovflw + _ser_pmc_preclean_ovflw + 4886 _ser_kac_ovflw + _ser_kac_preclean_ovflw; 4887 if (ser_ovflw > 0) { 4888 if (PrintCMSStatistics != 0) { 4889 gclog_or_tty->print_cr("Marking stack overflow (benign) " 4890 "(pmc_pc="SIZE_FORMAT", pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT 4891 ", kac_preclean="SIZE_FORMAT")", 4892 _ser_pmc_preclean_ovflw, _ser_pmc_remark_ovflw, 4893 _ser_kac_ovflw, _ser_kac_preclean_ovflw); 4894 } 4895 _markStack.expand(); 4896 _ser_pmc_remark_ovflw = 0; 4897 _ser_pmc_preclean_ovflw = 0; 4898 _ser_kac_preclean_ovflw = 0; 4899 _ser_kac_ovflw = 0; 4900 } 4901 if (_par_pmc_remark_ovflw > 0 || _par_kac_ovflw > 0) { 4902 if (PrintCMSStatistics != 0) { 4903 gclog_or_tty->print_cr("Work queue overflow (benign) " 4904 "(pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT")", 4905 _par_pmc_remark_ovflw, _par_kac_ovflw); 4906 } 4907 _par_pmc_remark_ovflw = 0; 4908 _par_kac_ovflw = 0; 4909 } 4910 if (PrintCMSStatistics != 0) { 4911 if (_markStack._hit_limit > 0) { 4912 gclog_or_tty->print_cr(" (benign) Hit max stack size limit ("SIZE_FORMAT")", 4913 _markStack._hit_limit); 4914 } 4915 if (_markStack._failed_double > 0) { 4916 gclog_or_tty->print_cr(" (benign) Failed stack doubling ("SIZE_FORMAT")," 4917 " current capacity "SIZE_FORMAT, 4918 _markStack._failed_double, 4919 _markStack.capacity()); 4920 } 4921 } 4922 _markStack._hit_limit = 0; 4923 _markStack._failed_double = 0; 4924 4925 // Check that all the klasses have been checked 4926 assert(_revisitStack.isEmpty(), "Not all klasses revisited"); 4927 4928 if ((VerifyAfterGC || VerifyDuringGC) && 4929 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 4930 verify_after_remark(); 4931 } 4932 4933 // Change under the freelistLocks. 4934 _collectorState = Sweeping; 4935 // Call isAllClear() under bitMapLock 4936 assert(_modUnionTable.isAllClear(), "Should be clear by end of the" 4937 " final marking"); 4938 if (UseAdaptiveSizePolicy) { 4939 size_policy()->checkpoint_roots_final_end(gch->gc_cause()); 4940 } 4941 } 4942 4943 // Parallel remark task 4944 class CMSParRemarkTask: public AbstractGangTask { 4945 CMSCollector* _collector; 4946 WorkGang* _workers; 4947 int _n_workers; 4948 CompactibleFreeListSpace* _cms_space; 4949 CompactibleFreeListSpace* _perm_space; 4950 4951 // The per-thread work queues, available here for stealing. 4952 OopTaskQueueSet* _task_queues; 4953 ParallelTaskTerminator _term; 4954 4955 public: 4956 CMSParRemarkTask(CMSCollector* collector, 4957 CompactibleFreeListSpace* cms_space, 4958 CompactibleFreeListSpace* perm_space, 4959 int n_workers, WorkGang* workers, 4960 OopTaskQueueSet* task_queues): 4961 AbstractGangTask("Rescan roots and grey objects in parallel"), 4962 _collector(collector), 4963 _cms_space(cms_space), _perm_space(perm_space), 4964 _n_workers(n_workers), 4965 _workers(workers), 4966 _task_queues(task_queues), 4967 _term(workers->total_workers(), task_queues) { } 4968 4969 OopTaskQueueSet* task_queues() { return _task_queues; } 4970 4971 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 4972 4973 ParallelTaskTerminator* terminator() { return &_term; } 4974 4975 void work(int i); 4976 4977 private: 4978 // Work method in support of parallel rescan ... of young gen spaces 4979 void do_young_space_rescan(int i, Par_MarkRefsIntoAndScanClosure* cl, 4980 ContiguousSpace* space, 4981 HeapWord** chunk_array, size_t chunk_top); 4982 4983 // ... of dirty cards in old space 4984 void do_dirty_card_rescan_tasks(CompactibleFreeListSpace* sp, int i, 4985 Par_MarkRefsIntoAndScanClosure* cl); 4986 4987 // ... work stealing for the above 4988 void do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, int* seed); 4989 }; 4990 4991 void CMSParRemarkTask::work(int i) { 4992 elapsedTimer _timer; 4993 ResourceMark rm; 4994 HandleMark hm; 4995 4996 // ---------- rescan from roots -------------- 4997 _timer.start(); 4998 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4999 Par_MarkRefsIntoAndScanClosure par_mrias_cl(_collector, 5000 _collector->_span, _collector->ref_processor(), 5001 &(_collector->_markBitMap), 5002 work_queue(i), &(_collector->_revisitStack)); 5003 5004 // Rescan young gen roots first since these are likely 5005 // coarsely partitioned and may, on that account, constitute 5006 // the critical path; thus, it's best to start off that 5007 // work first. 5008 // ---------- young gen roots -------------- 5009 { 5010 DefNewGeneration* dng = _collector->_young_gen->as_DefNewGeneration(); 5011 EdenSpace* eden_space = dng->eden(); 5012 ContiguousSpace* from_space = dng->from(); 5013 ContiguousSpace* to_space = dng->to(); 5014 5015 HeapWord** eca = _collector->_eden_chunk_array; 5016 size_t ect = _collector->_eden_chunk_index; 5017 HeapWord** sca = _collector->_survivor_chunk_array; 5018 size_t sct = _collector->_survivor_chunk_index; 5019 5020 assert(ect <= _collector->_eden_chunk_capacity, "out of bounds"); 5021 assert(sct <= _collector->_survivor_chunk_capacity, "out of bounds"); 5022 5023 do_young_space_rescan(i, &par_mrias_cl, to_space, NULL, 0); 5024 do_young_space_rescan(i, &par_mrias_cl, from_space, sca, sct); 5025 do_young_space_rescan(i, &par_mrias_cl, eden_space, eca, ect); 5026 5027 _timer.stop(); 5028 if (PrintCMSStatistics != 0) { 5029 gclog_or_tty->print_cr( 5030 "Finished young gen rescan work in %dth thread: %3.3f sec", 5031 i, _timer.seconds()); 5032 } 5033 } 5034 5035 // ---------- remaining roots -------------- 5036 _timer.reset(); 5037 _timer.start(); 5038 gch->gen_process_strong_roots(_collector->_cmsGen->level(), 5039 false, // yg was scanned above 5040 false, // this is parallel code 5041 true, // collecting perm gen 5042 SharedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()), 5043 &par_mrias_cl, 5044 true, // walk all of code cache if (so & SO_CodeCache) 5045 NULL); 5046 assert(_collector->should_unload_classes() 5047 || (_collector->CMSCollector::roots_scanning_options() & SharedHeap::SO_CodeCache), 5048 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 5049 _timer.stop(); 5050 if (PrintCMSStatistics != 0) { 5051 gclog_or_tty->print_cr( 5052 "Finished remaining root rescan work in %dth thread: %3.3f sec", 5053 i, _timer.seconds()); 5054 } 5055 5056 // ---------- rescan dirty cards ------------ 5057 _timer.reset(); 5058 _timer.start(); 5059 5060 // Do the rescan tasks for each of the two spaces 5061 // (cms_space and perm_space) in turn. 5062 do_dirty_card_rescan_tasks(_cms_space, i, &par_mrias_cl); 5063 do_dirty_card_rescan_tasks(_perm_space, i, &par_mrias_cl); 5064 _timer.stop(); 5065 if (PrintCMSStatistics != 0) { 5066 gclog_or_tty->print_cr( 5067 "Finished dirty card rescan work in %dth thread: %3.3f sec", 5068 i, _timer.seconds()); 5069 } 5070 5071 // ---------- steal work from other threads ... 5072 // ---------- ... and drain overflow list. 5073 _timer.reset(); 5074 _timer.start(); 5075 do_work_steal(i, &par_mrias_cl, _collector->hash_seed(i)); 5076 _timer.stop(); 5077 if (PrintCMSStatistics != 0) { 5078 gclog_or_tty->print_cr( 5079 "Finished work stealing in %dth thread: %3.3f sec", 5080 i, _timer.seconds()); 5081 } 5082 } 5083 5084 void 5085 CMSParRemarkTask::do_young_space_rescan(int i, 5086 Par_MarkRefsIntoAndScanClosure* cl, ContiguousSpace* space, 5087 HeapWord** chunk_array, size_t chunk_top) { 5088 // Until all tasks completed: 5089 // . claim an unclaimed task 5090 // . compute region boundaries corresponding to task claimed 5091 // using chunk_array 5092 // . par_oop_iterate(cl) over that region 5093 5094 ResourceMark rm; 5095 HandleMark hm; 5096 5097 SequentialSubTasksDone* pst = space->par_seq_tasks(); 5098 assert(pst->valid(), "Uninitialized use?"); 5099 5100 int nth_task = 0; 5101 int n_tasks = pst->n_tasks(); 5102 5103 HeapWord *start, *end; 5104 while (!pst->is_task_claimed(/* reference */ nth_task)) { 5105 // We claimed task # nth_task; compute its boundaries. 5106 if (chunk_top == 0) { // no samples were taken 5107 assert(nth_task == 0 && n_tasks == 1, "Can have only 1 EdenSpace task"); 5108 start = space->bottom(); 5109 end = space->top(); 5110 } else if (nth_task == 0) { 5111 start = space->bottom(); 5112 end = chunk_array[nth_task]; 5113 } else if (nth_task < (jint)chunk_top) { 5114 assert(nth_task >= 1, "Control point invariant"); 5115 start = chunk_array[nth_task - 1]; 5116 end = chunk_array[nth_task]; 5117 } else { 5118 assert(nth_task == (jint)chunk_top, "Control point invariant"); 5119 start = chunk_array[chunk_top - 1]; 5120 end = space->top(); 5121 } 5122 MemRegion mr(start, end); 5123 // Verify that mr is in space 5124 assert(mr.is_empty() || space->used_region().contains(mr), 5125 "Should be in space"); 5126 // Verify that "start" is an object boundary 5127 assert(mr.is_empty() || oop(mr.start())->is_oop(), 5128 "Should be an oop"); 5129 space->par_oop_iterate(mr, cl); 5130 } 5131 pst->all_tasks_completed(); 5132 } 5133 5134 void 5135 CMSParRemarkTask::do_dirty_card_rescan_tasks( 5136 CompactibleFreeListSpace* sp, int i, 5137 Par_MarkRefsIntoAndScanClosure* cl) { 5138 // Until all tasks completed: 5139 // . claim an unclaimed task 5140 // . compute region boundaries corresponding to task claimed 5141 // . transfer dirty bits ct->mut for that region 5142 // . apply rescanclosure to dirty mut bits for that region 5143 5144 ResourceMark rm; 5145 HandleMark hm; 5146 5147 OopTaskQueue* work_q = work_queue(i); 5148 ModUnionClosure modUnionClosure(&(_collector->_modUnionTable)); 5149 // CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! 5150 // CAUTION: This closure has state that persists across calls to 5151 // the work method dirty_range_iterate_clear() in that it has 5152 // imbedded in it a (subtype of) UpwardsObjectClosure. The 5153 // use of that state in the imbedded UpwardsObjectClosure instance 5154 // assumes that the cards are always iterated (even if in parallel 5155 // by several threads) in monotonically increasing order per each 5156 // thread. This is true of the implementation below which picks 5157 // card ranges (chunks) in monotonically increasing order globally 5158 // and, a-fortiori, in monotonically increasing order per thread 5159 // (the latter order being a subsequence of the former). 5160 // If the work code below is ever reorganized into a more chaotic 5161 // work-partitioning form than the current "sequential tasks" 5162 // paradigm, the use of that persistent state will have to be 5163 // revisited and modified appropriately. See also related 5164 // bug 4756801 work on which should examine this code to make 5165 // sure that the changes there do not run counter to the 5166 // assumptions made here and necessary for correctness and 5167 // efficiency. Note also that this code might yield inefficient 5168 // behaviour in the case of very large objects that span one or 5169 // more work chunks. Such objects would potentially be scanned 5170 // several times redundantly. Work on 4756801 should try and 5171 // address that performance anomaly if at all possible. XXX 5172 MemRegion full_span = _collector->_span; 5173 CMSBitMap* bm = &(_collector->_markBitMap); // shared 5174 CMSMarkStack* rs = &(_collector->_revisitStack); // shared 5175 MarkFromDirtyCardsClosure 5176 greyRescanClosure(_collector, full_span, // entire span of interest 5177 sp, bm, work_q, rs, cl); 5178 5179 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 5180 assert(pst->valid(), "Uninitialized use?"); 5181 int nth_task = 0; 5182 const int alignment = CardTableModRefBS::card_size * BitsPerWord; 5183 MemRegion span = sp->used_region(); 5184 HeapWord* start_addr = span.start(); 5185 HeapWord* end_addr = (HeapWord*)round_to((intptr_t)span.end(), 5186 alignment); 5187 const size_t chunk_size = sp->rescan_task_size(); // in HeapWord units 5188 assert((HeapWord*)round_to((intptr_t)start_addr, alignment) == 5189 start_addr, "Check alignment"); 5190 assert((size_t)round_to((intptr_t)chunk_size, alignment) == 5191 chunk_size, "Check alignment"); 5192 5193 while (!pst->is_task_claimed(/* reference */ nth_task)) { 5194 // Having claimed the nth_task, compute corresponding mem-region, 5195 // which is a-fortiori aligned correctly (i.e. at a MUT bopundary). 5196 // The alignment restriction ensures that we do not need any 5197 // synchronization with other gang-workers while setting or 5198 // clearing bits in thus chunk of the MUT. 5199 MemRegion this_span = MemRegion(start_addr + nth_task*chunk_size, 5200 start_addr + (nth_task+1)*chunk_size); 5201 // The last chunk's end might be way beyond end of the 5202 // used region. In that case pull back appropriately. 5203 if (this_span.end() > end_addr) { 5204 this_span.set_end(end_addr); 5205 assert(!this_span.is_empty(), "Program logic (calculation of n_tasks)"); 5206 } 5207 // Iterate over the dirty cards covering this chunk, marking them 5208 // precleaned, and setting the corresponding bits in the mod union 5209 // table. Since we have been careful to partition at Card and MUT-word 5210 // boundaries no synchronization is needed between parallel threads. 5211 _collector->_ct->ct_bs()->dirty_card_iterate(this_span, 5212 &modUnionClosure); 5213 5214 // Having transferred these marks into the modUnionTable, 5215 // rescan the marked objects on the dirty cards in the modUnionTable. 5216 // Even if this is at a synchronous collection, the initial marking 5217 // may have been done during an asynchronous collection so there 5218 // may be dirty bits in the mod-union table. 5219 _collector->_modUnionTable.dirty_range_iterate_clear( 5220 this_span, &greyRescanClosure); 5221 _collector->_modUnionTable.verifyNoOneBitsInRange( 5222 this_span.start(), 5223 this_span.end()); 5224 } 5225 pst->all_tasks_completed(); // declare that i am done 5226 } 5227 5228 // . see if we can share work_queues with ParNew? XXX 5229 void 5230 CMSParRemarkTask::do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, 5231 int* seed) { 5232 OopTaskQueue* work_q = work_queue(i); 5233 NOT_PRODUCT(int num_steals = 0;) 5234 oop obj_to_scan; 5235 CMSBitMap* bm = &(_collector->_markBitMap); 5236 5237 while (true) { 5238 // Completely finish any left over work from (an) earlier round(s) 5239 cl->trim_queue(0); 5240 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 5241 (size_t)ParGCDesiredObjsFromOverflowList); 5242 // Now check if there's any work in the overflow list 5243 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 5244 work_q)) { 5245 // found something in global overflow list; 5246 // not yet ready to go stealing work from others. 5247 // We'd like to assert(work_q->size() != 0, ...) 5248 // because we just took work from the overflow list, 5249 // but of course we can't since all of that could have 5250 // been already stolen from us. 5251 // "He giveth and He taketh away." 5252 continue; 5253 } 5254 // Verify that we have no work before we resort to stealing 5255 assert(work_q->size() == 0, "Have work, shouldn't steal"); 5256 // Try to steal from other queues that have work 5257 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 5258 NOT_PRODUCT(num_steals++;) 5259 assert(obj_to_scan->is_oop(), "Oops, not an oop!"); 5260 assert(bm->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 5261 // Do scanning work 5262 obj_to_scan->oop_iterate(cl); 5263 // Loop around, finish this work, and try to steal some more 5264 } else if (terminator()->offer_termination()) { 5265 break; // nirvana from the infinite cycle 5266 } 5267 } 5268 NOT_PRODUCT( 5269 if (PrintCMSStatistics != 0) { 5270 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals); 5271 } 5272 ) 5273 assert(work_q->size() == 0 && _collector->overflow_list_is_empty(), 5274 "Else our work is not yet done"); 5275 } 5276 5277 // Return a thread-local PLAB recording array, as appropriate. 5278 void* CMSCollector::get_data_recorder(int thr_num) { 5279 if (_survivor_plab_array != NULL && 5280 (CMSPLABRecordAlways || 5281 (_collectorState > Marking && _collectorState < FinalMarking))) { 5282 assert(thr_num < (int)ParallelGCThreads, "thr_num is out of bounds"); 5283 ChunkArray* ca = &_survivor_plab_array[thr_num]; 5284 ca->reset(); // clear it so that fresh data is recorded 5285 return (void*) ca; 5286 } else { 5287 return NULL; 5288 } 5289 } 5290 5291 // Reset all the thread-local PLAB recording arrays 5292 void CMSCollector::reset_survivor_plab_arrays() { 5293 for (uint i = 0; i < ParallelGCThreads; i++) { 5294 _survivor_plab_array[i].reset(); 5295 } 5296 } 5297 5298 // Merge the per-thread plab arrays into the global survivor chunk 5299 // array which will provide the partitioning of the survivor space 5300 // for CMS rescan. 5301 void CMSCollector::merge_survivor_plab_arrays(ContiguousSpace* surv) { 5302 assert(_survivor_plab_array != NULL, "Error"); 5303 assert(_survivor_chunk_array != NULL, "Error"); 5304 assert(_collectorState == FinalMarking, "Error"); 5305 for (uint j = 0; j < ParallelGCThreads; j++) { 5306 _cursor[j] = 0; 5307 } 5308 HeapWord* top = surv->top(); 5309 size_t i; 5310 for (i = 0; i < _survivor_chunk_capacity; i++) { // all sca entries 5311 HeapWord* min_val = top; // Higher than any PLAB address 5312 uint min_tid = 0; // position of min_val this round 5313 for (uint j = 0; j < ParallelGCThreads; j++) { 5314 ChunkArray* cur_sca = &_survivor_plab_array[j]; 5315 if (_cursor[j] == cur_sca->end()) { 5316 continue; 5317 } 5318 assert(_cursor[j] < cur_sca->end(), "ctl pt invariant"); 5319 HeapWord* cur_val = cur_sca->nth(_cursor[j]); 5320 assert(surv->used_region().contains(cur_val), "Out of bounds value"); 5321 if (cur_val < min_val) { 5322 min_tid = j; 5323 min_val = cur_val; 5324 } else { 5325 assert(cur_val < top, "All recorded addresses should be less"); 5326 } 5327 } 5328 // At this point min_val and min_tid are respectively 5329 // the least address in _survivor_plab_array[j]->nth(_cursor[j]) 5330 // and the thread (j) that witnesses that address. 5331 // We record this address in the _survivor_chunk_array[i] 5332 // and increment _cursor[min_tid] prior to the next round i. 5333 if (min_val == top) { 5334 break; 5335 } 5336 _survivor_chunk_array[i] = min_val; 5337 _cursor[min_tid]++; 5338 } 5339 // We are all done; record the size of the _survivor_chunk_array 5340 _survivor_chunk_index = i; // exclusive: [0, i) 5341 if (PrintCMSStatistics > 0) { 5342 gclog_or_tty->print(" (Survivor:" SIZE_FORMAT "chunks) ", i); 5343 } 5344 // Verify that we used up all the recorded entries 5345 #ifdef ASSERT 5346 size_t total = 0; 5347 for (uint j = 0; j < ParallelGCThreads; j++) { 5348 assert(_cursor[j] == _survivor_plab_array[j].end(), "Ctl pt invariant"); 5349 total += _cursor[j]; 5350 } 5351 assert(total == _survivor_chunk_index, "Ctl Pt Invariant"); 5352 // Check that the merged array is in sorted order 5353 if (total > 0) { 5354 for (size_t i = 0; i < total - 1; i++) { 5355 if (PrintCMSStatistics > 0) { 5356 gclog_or_tty->print(" (chunk" SIZE_FORMAT ":" INTPTR_FORMAT ") ", 5357 i, _survivor_chunk_array[i]); 5358 } 5359 assert(_survivor_chunk_array[i] < _survivor_chunk_array[i+1], 5360 "Not sorted"); 5361 } 5362 } 5363 #endif // ASSERT 5364 } 5365 5366 // Set up the space's par_seq_tasks structure for work claiming 5367 // for parallel rescan of young gen. 5368 // See ParRescanTask where this is currently used. 5369 void 5370 CMSCollector:: 5371 initialize_sequential_subtasks_for_young_gen_rescan(int n_threads) { 5372 assert(n_threads > 0, "Unexpected n_threads argument"); 5373 DefNewGeneration* dng = (DefNewGeneration*)_young_gen; 5374 5375 // Eden space 5376 { 5377 SequentialSubTasksDone* pst = dng->eden()->par_seq_tasks(); 5378 assert(!pst->valid(), "Clobbering existing data?"); 5379 // Each valid entry in [0, _eden_chunk_index) represents a task. 5380 size_t n_tasks = _eden_chunk_index + 1; 5381 assert(n_tasks == 1 || _eden_chunk_array != NULL, "Error"); 5382 pst->set_par_threads(n_threads); 5383 pst->set_n_tasks((int)n_tasks); 5384 } 5385 5386 // Merge the survivor plab arrays into _survivor_chunk_array 5387 if (_survivor_plab_array != NULL) { 5388 merge_survivor_plab_arrays(dng->from()); 5389 } else { 5390 assert(_survivor_chunk_index == 0, "Error"); 5391 } 5392 5393 // To space 5394 { 5395 SequentialSubTasksDone* pst = dng->to()->par_seq_tasks(); 5396 assert(!pst->valid(), "Clobbering existing data?"); 5397 pst->set_par_threads(n_threads); 5398 pst->set_n_tasks(1); 5399 assert(pst->valid(), "Error"); 5400 } 5401 5402 // From space 5403 { 5404 SequentialSubTasksDone* pst = dng->from()->par_seq_tasks(); 5405 assert(!pst->valid(), "Clobbering existing data?"); 5406 size_t n_tasks = _survivor_chunk_index + 1; 5407 assert(n_tasks == 1 || _survivor_chunk_array != NULL, "Error"); 5408 pst->set_par_threads(n_threads); 5409 pst->set_n_tasks((int)n_tasks); 5410 assert(pst->valid(), "Error"); 5411 } 5412 } 5413 5414 // Parallel version of remark 5415 void CMSCollector::do_remark_parallel() { 5416 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5417 WorkGang* workers = gch->workers(); 5418 assert(workers != NULL, "Need parallel worker threads."); 5419 int n_workers = workers->total_workers(); 5420 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 5421 CompactibleFreeListSpace* perm_space = _permGen->cmsSpace(); 5422 5423 CMSParRemarkTask tsk(this, 5424 cms_space, perm_space, 5425 n_workers, workers, task_queues()); 5426 5427 // Set up for parallel process_strong_roots work. 5428 gch->set_par_threads(n_workers); 5429 // We won't be iterating over the cards in the card table updating 5430 // the younger_gen cards, so we shouldn't call the following else 5431 // the verification code as well as subsequent younger_refs_iterate 5432 // code would get confused. XXX 5433 // gch->rem_set()->prepare_for_younger_refs_iterate(true); // parallel 5434 5435 // The young gen rescan work will not be done as part of 5436 // process_strong_roots (which currently doesn't knw how to 5437 // parallelize such a scan), but rather will be broken up into 5438 // a set of parallel tasks (via the sampling that the [abortable] 5439 // preclean phase did of EdenSpace, plus the [two] tasks of 5440 // scanning the [two] survivor spaces. Further fine-grain 5441 // parallelization of the scanning of the survivor spaces 5442 // themselves, and of precleaning of the younger gen itself 5443 // is deferred to the future. 5444 initialize_sequential_subtasks_for_young_gen_rescan(n_workers); 5445 5446 // The dirty card rescan work is broken up into a "sequence" 5447 // of parallel tasks (per constituent space) that are dynamically 5448 // claimed by the parallel threads. 5449 cms_space->initialize_sequential_subtasks_for_rescan(n_workers); 5450 perm_space->initialize_sequential_subtasks_for_rescan(n_workers); 5451 5452 // It turns out that even when we're using 1 thread, doing the work in a 5453 // separate thread causes wide variance in run times. We can't help this 5454 // in the multi-threaded case, but we special-case n=1 here to get 5455 // repeatable measurements of the 1-thread overhead of the parallel code. 5456 if (n_workers > 1) { 5457 // Make refs discovery MT-safe 5458 ReferenceProcessorMTMutator mt(ref_processor(), true); 5459 GenCollectedHeap::StrongRootsScope srs(gch); 5460 workers->run_task(&tsk); 5461 } else { 5462 GenCollectedHeap::StrongRootsScope srs(gch); 5463 tsk.work(0); 5464 } 5465 gch->set_par_threads(0); // 0 ==> non-parallel. 5466 // restore, single-threaded for now, any preserved marks 5467 // as a result of work_q overflow 5468 restore_preserved_marks_if_any(); 5469 } 5470 5471 // Non-parallel version of remark 5472 void CMSCollector::do_remark_non_parallel() { 5473 ResourceMark rm; 5474 HandleMark hm; 5475 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5476 MarkRefsIntoAndScanClosure 5477 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable, 5478 &_markStack, &_revisitStack, this, 5479 false /* should_yield */, false /* not precleaning */); 5480 MarkFromDirtyCardsClosure 5481 markFromDirtyCardsClosure(this, _span, 5482 NULL, // space is set further below 5483 &_markBitMap, &_markStack, &_revisitStack, 5484 &mrias_cl); 5485 { 5486 TraceTime t("grey object rescan", PrintGCDetails, false, gclog_or_tty); 5487 // Iterate over the dirty cards, setting the corresponding bits in the 5488 // mod union table. 5489 { 5490 ModUnionClosure modUnionClosure(&_modUnionTable); 5491 _ct->ct_bs()->dirty_card_iterate( 5492 _cmsGen->used_region(), 5493 &modUnionClosure); 5494 _ct->ct_bs()->dirty_card_iterate( 5495 _permGen->used_region(), 5496 &modUnionClosure); 5497 } 5498 // Having transferred these marks into the modUnionTable, we just need 5499 // to rescan the marked objects on the dirty cards in the modUnionTable. 5500 // The initial marking may have been done during an asynchronous 5501 // collection so there may be dirty bits in the mod-union table. 5502 const int alignment = 5503 CardTableModRefBS::card_size * BitsPerWord; 5504 { 5505 // ... First handle dirty cards in CMS gen 5506 markFromDirtyCardsClosure.set_space(_cmsGen->cmsSpace()); 5507 MemRegion ur = _cmsGen->used_region(); 5508 HeapWord* lb = ur.start(); 5509 HeapWord* ub = (HeapWord*)round_to((intptr_t)ur.end(), alignment); 5510 MemRegion cms_span(lb, ub); 5511 _modUnionTable.dirty_range_iterate_clear(cms_span, 5512 &markFromDirtyCardsClosure); 5513 verify_work_stacks_empty(); 5514 if (PrintCMSStatistics != 0) { 5515 gclog_or_tty->print(" (re-scanned "SIZE_FORMAT" dirty cards in cms gen) ", 5516 markFromDirtyCardsClosure.num_dirty_cards()); 5517 } 5518 } 5519 { 5520 // .. and then repeat for dirty cards in perm gen 5521 markFromDirtyCardsClosure.set_space(_permGen->cmsSpace()); 5522 MemRegion ur = _permGen->used_region(); 5523 HeapWord* lb = ur.start(); 5524 HeapWord* ub = (HeapWord*)round_to((intptr_t)ur.end(), alignment); 5525 MemRegion perm_span(lb, ub); 5526 _modUnionTable.dirty_range_iterate_clear(perm_span, 5527 &markFromDirtyCardsClosure); 5528 verify_work_stacks_empty(); 5529 if (PrintCMSStatistics != 0) { 5530 gclog_or_tty->print(" (re-scanned "SIZE_FORMAT" dirty cards in perm gen) ", 5531 markFromDirtyCardsClosure.num_dirty_cards()); 5532 } 5533 } 5534 } 5535 if (VerifyDuringGC && 5536 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 5537 HandleMark hm; // Discard invalid handles created during verification 5538 Universe::verify(true); 5539 } 5540 { 5541 TraceTime t("root rescan", PrintGCDetails, false, gclog_or_tty); 5542 5543 verify_work_stacks_empty(); 5544 5545 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 5546 GenCollectedHeap::StrongRootsScope srs(gch); 5547 gch->gen_process_strong_roots(_cmsGen->level(), 5548 true, // younger gens as roots 5549 false, // use the local StrongRootsScope 5550 true, // collecting perm gen 5551 SharedHeap::ScanningOption(roots_scanning_options()), 5552 &mrias_cl, 5553 true, // walk code active on stacks 5554 NULL); 5555 assert(should_unload_classes() 5556 || (roots_scanning_options() & SharedHeap::SO_CodeCache), 5557 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 5558 } 5559 verify_work_stacks_empty(); 5560 // Restore evacuated mark words, if any, used for overflow list links 5561 if (!CMSOverflowEarlyRestoration) { 5562 restore_preserved_marks_if_any(); 5563 } 5564 verify_overflow_empty(); 5565 } 5566 5567 //////////////////////////////////////////////////////// 5568 // Parallel Reference Processing Task Proxy Class 5569 //////////////////////////////////////////////////////// 5570 class CMSRefProcTaskProxy: public AbstractGangTask { 5571 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask; 5572 CMSCollector* _collector; 5573 CMSBitMap* _mark_bit_map; 5574 const MemRegion _span; 5575 OopTaskQueueSet* _task_queues; 5576 ParallelTaskTerminator _term; 5577 ProcessTask& _task; 5578 5579 public: 5580 CMSRefProcTaskProxy(ProcessTask& task, 5581 CMSCollector* collector, 5582 const MemRegion& span, 5583 CMSBitMap* mark_bit_map, 5584 int total_workers, 5585 OopTaskQueueSet* task_queues): 5586 AbstractGangTask("Process referents by policy in parallel"), 5587 _task(task), 5588 _collector(collector), _span(span), _mark_bit_map(mark_bit_map), 5589 _task_queues(task_queues), 5590 _term(total_workers, task_queues) 5591 { 5592 assert(_collector->_span.equals(_span) && !_span.is_empty(), 5593 "Inconsistency in _span"); 5594 } 5595 5596 OopTaskQueueSet* task_queues() { return _task_queues; } 5597 5598 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 5599 5600 ParallelTaskTerminator* terminator() { return &_term; } 5601 5602 void do_work_steal(int i, 5603 CMSParDrainMarkingStackClosure* drain, 5604 CMSParKeepAliveClosure* keep_alive, 5605 int* seed); 5606 5607 virtual void work(int i); 5608 }; 5609 5610 void CMSRefProcTaskProxy::work(int i) { 5611 assert(_collector->_span.equals(_span), "Inconsistency in _span"); 5612 CMSParKeepAliveClosure par_keep_alive(_collector, _span, 5613 _mark_bit_map, 5614 &_collector->_revisitStack, 5615 work_queue(i)); 5616 CMSParDrainMarkingStackClosure par_drain_stack(_collector, _span, 5617 _mark_bit_map, 5618 &_collector->_revisitStack, 5619 work_queue(i)); 5620 CMSIsAliveClosure is_alive_closure(_span, _mark_bit_map); 5621 _task.work(i, is_alive_closure, par_keep_alive, par_drain_stack); 5622 if (_task.marks_oops_alive()) { 5623 do_work_steal(i, &par_drain_stack, &par_keep_alive, 5624 _collector->hash_seed(i)); 5625 } 5626 assert(work_queue(i)->size() == 0, "work_queue should be empty"); 5627 assert(_collector->_overflow_list == NULL, "non-empty _overflow_list"); 5628 } 5629 5630 class CMSRefEnqueueTaskProxy: public AbstractGangTask { 5631 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask; 5632 EnqueueTask& _task; 5633 5634 public: 5635 CMSRefEnqueueTaskProxy(EnqueueTask& task) 5636 : AbstractGangTask("Enqueue reference objects in parallel"), 5637 _task(task) 5638 { } 5639 5640 virtual void work(int i) 5641 { 5642 _task.work(i); 5643 } 5644 }; 5645 5646 CMSParKeepAliveClosure::CMSParKeepAliveClosure(CMSCollector* collector, 5647 MemRegion span, CMSBitMap* bit_map, CMSMarkStack* revisit_stack, 5648 OopTaskQueue* work_queue): 5649 Par_KlassRememberingOopClosure(collector, NULL, revisit_stack), 5650 _span(span), 5651 _bit_map(bit_map), 5652 _work_queue(work_queue), 5653 _mark_and_push(collector, span, bit_map, revisit_stack, work_queue), 5654 _low_water_mark(MIN2((uint)(work_queue->max_elems()/4), 5655 (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads))) 5656 { } 5657 5658 // . see if we can share work_queues with ParNew? XXX 5659 void CMSRefProcTaskProxy::do_work_steal(int i, 5660 CMSParDrainMarkingStackClosure* drain, 5661 CMSParKeepAliveClosure* keep_alive, 5662 int* seed) { 5663 OopTaskQueue* work_q = work_queue(i); 5664 NOT_PRODUCT(int num_steals = 0;) 5665 oop obj_to_scan; 5666 5667 while (true) { 5668 // Completely finish any left over work from (an) earlier round(s) 5669 drain->trim_queue(0); 5670 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 5671 (size_t)ParGCDesiredObjsFromOverflowList); 5672 // Now check if there's any work in the overflow list 5673 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 5674 work_q)) { 5675 // Found something in global overflow list; 5676 // not yet ready to go stealing work from others. 5677 // We'd like to assert(work_q->size() != 0, ...) 5678 // because we just took work from the overflow list, 5679 // but of course we can't, since all of that might have 5680 // been already stolen from us. 5681 continue; 5682 } 5683 // Verify that we have no work before we resort to stealing 5684 assert(work_q->size() == 0, "Have work, shouldn't steal"); 5685 // Try to steal from other queues that have work 5686 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 5687 NOT_PRODUCT(num_steals++;) 5688 assert(obj_to_scan->is_oop(), "Oops, not an oop!"); 5689 assert(_mark_bit_map->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 5690 // Do scanning work 5691 obj_to_scan->oop_iterate(keep_alive); 5692 // Loop around, finish this work, and try to steal some more 5693 } else if (terminator()->offer_termination()) { 5694 break; // nirvana from the infinite cycle 5695 } 5696 } 5697 NOT_PRODUCT( 5698 if (PrintCMSStatistics != 0) { 5699 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals); 5700 } 5701 ) 5702 } 5703 5704 void CMSRefProcTaskExecutor::execute(ProcessTask& task) 5705 { 5706 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5707 WorkGang* workers = gch->workers(); 5708 assert(workers != NULL, "Need parallel worker threads."); 5709 int n_workers = workers->total_workers(); 5710 CMSRefProcTaskProxy rp_task(task, &_collector, 5711 _collector.ref_processor()->span(), 5712 _collector.markBitMap(), 5713 n_workers, _collector.task_queues()); 5714 workers->run_task(&rp_task); 5715 } 5716 5717 void CMSRefProcTaskExecutor::execute(EnqueueTask& task) 5718 { 5719 5720 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5721 WorkGang* workers = gch->workers(); 5722 assert(workers != NULL, "Need parallel worker threads."); 5723 CMSRefEnqueueTaskProxy enq_task(task); 5724 workers->run_task(&enq_task); 5725 } 5726 5727 void CMSCollector::refProcessingWork(bool asynch, bool clear_all_soft_refs) { 5728 5729 ResourceMark rm; 5730 HandleMark hm; 5731 5732 ReferenceProcessor* rp = ref_processor(); 5733 assert(rp->span().equals(_span), "Spans should be equal"); 5734 assert(!rp->enqueuing_is_done(), "Enqueuing should not be complete"); 5735 // Process weak references. 5736 rp->setup_policy(clear_all_soft_refs); 5737 verify_work_stacks_empty(); 5738 5739 CMSKeepAliveClosure cmsKeepAliveClosure(this, _span, &_markBitMap, 5740 &_markStack, &_revisitStack, 5741 false /* !preclean */); 5742 CMSDrainMarkingStackClosure cmsDrainMarkingStackClosure(this, 5743 _span, &_markBitMap, &_markStack, 5744 &cmsKeepAliveClosure, false /* !preclean */); 5745 { 5746 TraceTime t("weak refs processing", PrintGCDetails, false, gclog_or_tty); 5747 if (rp->processing_is_mt()) { 5748 CMSRefProcTaskExecutor task_executor(*this); 5749 rp->process_discovered_references(&_is_alive_closure, 5750 &cmsKeepAliveClosure, 5751 &cmsDrainMarkingStackClosure, 5752 &task_executor); 5753 } else { 5754 rp->process_discovered_references(&_is_alive_closure, 5755 &cmsKeepAliveClosure, 5756 &cmsDrainMarkingStackClosure, 5757 NULL); 5758 } 5759 verify_work_stacks_empty(); 5760 } 5761 5762 if (should_unload_classes()) { 5763 { 5764 TraceTime t("class unloading", PrintGCDetails, false, gclog_or_tty); 5765 5766 // Follow SystemDictionary roots and unload classes 5767 bool purged_class = SystemDictionary::do_unloading(&_is_alive_closure); 5768 5769 // Follow CodeCache roots and unload any methods marked for unloading 5770 CodeCache::do_unloading(&_is_alive_closure, 5771 &cmsKeepAliveClosure, 5772 purged_class); 5773 5774 cmsDrainMarkingStackClosure.do_void(); 5775 verify_work_stacks_empty(); 5776 5777 // Update subklass/sibling/implementor links in KlassKlass descendants 5778 assert(!_revisitStack.isEmpty(), "revisit stack should not be empty"); 5779 oop k; 5780 while ((k = _revisitStack.pop()) != NULL) { 5781 ((Klass*)(oopDesc*)k)->follow_weak_klass_links( 5782 &_is_alive_closure, 5783 &cmsKeepAliveClosure); 5784 } 5785 assert(!ClassUnloading || 5786 (_markStack.isEmpty() && overflow_list_is_empty()), 5787 "Should not have found new reachable objects"); 5788 assert(_revisitStack.isEmpty(), "revisit stack should have been drained"); 5789 cmsDrainMarkingStackClosure.do_void(); 5790 verify_work_stacks_empty(); 5791 } 5792 5793 { 5794 TraceTime t("scrub symbol & string tables", PrintGCDetails, false, gclog_or_tty); 5795 // Now clean up stale oops in SymbolTable and StringTable 5796 SymbolTable::unlink(&_is_alive_closure); 5797 StringTable::unlink(&_is_alive_closure); 5798 } 5799 } 5800 5801 verify_work_stacks_empty(); 5802 // Restore any preserved marks as a result of mark stack or 5803 // work queue overflow 5804 restore_preserved_marks_if_any(); // done single-threaded for now 5805 5806 rp->set_enqueuing_is_done(true); 5807 if (rp->processing_is_mt()) { 5808 CMSRefProcTaskExecutor task_executor(*this); 5809 rp->enqueue_discovered_references(&task_executor); 5810 } else { 5811 rp->enqueue_discovered_references(NULL); 5812 } 5813 rp->verify_no_references_recorded(); 5814 assert(!rp->discovery_enabled(), "should have been disabled"); 5815 5816 // JVMTI object tagging is based on JNI weak refs. If any of these 5817 // refs were cleared then JVMTI needs to update its maps and 5818 // maybe post ObjectFrees to agents. 5819 JvmtiExport::cms_ref_processing_epilogue(); 5820 } 5821 5822 #ifndef PRODUCT 5823 void CMSCollector::check_correct_thread_executing() { 5824 Thread* t = Thread::current(); 5825 // Only the VM thread or the CMS thread should be here. 5826 assert(t->is_ConcurrentGC_thread() || t->is_VM_thread(), 5827 "Unexpected thread type"); 5828 // If this is the vm thread, the foreground process 5829 // should not be waiting. Note that _foregroundGCIsActive is 5830 // true while the foreground collector is waiting. 5831 if (_foregroundGCShouldWait) { 5832 // We cannot be the VM thread 5833 assert(t->is_ConcurrentGC_thread(), 5834 "Should be CMS thread"); 5835 } else { 5836 // We can be the CMS thread only if we are in a stop-world 5837 // phase of CMS collection. 5838 if (t->is_ConcurrentGC_thread()) { 5839 assert(_collectorState == InitialMarking || 5840 _collectorState == FinalMarking, 5841 "Should be a stop-world phase"); 5842 // The CMS thread should be holding the CMS_token. 5843 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 5844 "Potential interference with concurrently " 5845 "executing VM thread"); 5846 } 5847 } 5848 } 5849 #endif 5850 5851 void CMSCollector::sweep(bool asynch) { 5852 assert(_collectorState == Sweeping, "just checking"); 5853 check_correct_thread_executing(); 5854 verify_work_stacks_empty(); 5855 verify_overflow_empty(); 5856 increment_sweep_count(); 5857 _inter_sweep_timer.stop(); 5858 _inter_sweep_estimate.sample(_inter_sweep_timer.seconds()); 5859 size_policy()->avg_cms_free_at_sweep()->sample(_cmsGen->free()); 5860 5861 // PermGen verification support: If perm gen sweeping is disabled in 5862 // this cycle, we preserve the perm gen object "deadness" information 5863 // in the perm_gen_verify_bit_map. In order to do that we traverse 5864 // all blocks in perm gen and mark all dead objects. 5865 if (verifying() && !should_unload_classes()) { 5866 assert(perm_gen_verify_bit_map()->sizeInBits() != 0, 5867 "Should have already been allocated"); 5868 MarkDeadObjectsClosure mdo(this, _permGen->cmsSpace(), 5869 markBitMap(), perm_gen_verify_bit_map()); 5870 if (asynch) { 5871 CMSTokenSyncWithLocks ts(true, _permGen->freelistLock(), 5872 bitMapLock()); 5873 _permGen->cmsSpace()->blk_iterate(&mdo); 5874 } else { 5875 // In the case of synchronous sweep, we already have 5876 // the requisite locks/tokens. 5877 _permGen->cmsSpace()->blk_iterate(&mdo); 5878 } 5879 } 5880 5881 assert(!_intra_sweep_timer.is_active(), "Should not be active"); 5882 _intra_sweep_timer.reset(); 5883 _intra_sweep_timer.start(); 5884 if (asynch) { 5885 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 5886 CMSPhaseAccounting pa(this, "sweep", !PrintGCDetails); 5887 // First sweep the old gen then the perm gen 5888 { 5889 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(), 5890 bitMapLock()); 5891 sweepWork(_cmsGen, asynch); 5892 } 5893 5894 // Now repeat for perm gen 5895 if (should_unload_classes()) { 5896 CMSTokenSyncWithLocks ts(true, _permGen->freelistLock(), 5897 bitMapLock()); 5898 sweepWork(_permGen, asynch); 5899 } 5900 5901 // Update Universe::_heap_*_at_gc figures. 5902 // We need all the free list locks to make the abstract state 5903 // transition from Sweeping to Resetting. See detailed note 5904 // further below. 5905 { 5906 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(), 5907 _permGen->freelistLock()); 5908 // Update heap occupancy information which is used as 5909 // input to soft ref clearing policy at the next gc. 5910 Universe::update_heap_info_at_gc(); 5911 _collectorState = Resizing; 5912 } 5913 } else { 5914 // already have needed locks 5915 sweepWork(_cmsGen, asynch); 5916 5917 if (should_unload_classes()) { 5918 sweepWork(_permGen, asynch); 5919 } 5920 // Update heap occupancy information which is used as 5921 // input to soft ref clearing policy at the next gc. 5922 Universe::update_heap_info_at_gc(); 5923 _collectorState = Resizing; 5924 } 5925 verify_work_stacks_empty(); 5926 verify_overflow_empty(); 5927 5928 _intra_sweep_timer.stop(); 5929 _intra_sweep_estimate.sample(_intra_sweep_timer.seconds()); 5930 5931 _inter_sweep_timer.reset(); 5932 _inter_sweep_timer.start(); 5933 5934 update_time_of_last_gc(os::javaTimeMillis()); 5935 5936 // NOTE on abstract state transitions: 5937 // Mutators allocate-live and/or mark the mod-union table dirty 5938 // based on the state of the collection. The former is done in 5939 // the interval [Marking, Sweeping] and the latter in the interval 5940 // [Marking, Sweeping). Thus the transitions into the Marking state 5941 // and out of the Sweeping state must be synchronously visible 5942 // globally to the mutators. 5943 // The transition into the Marking state happens with the world 5944 // stopped so the mutators will globally see it. Sweeping is 5945 // done asynchronously by the background collector so the transition 5946 // from the Sweeping state to the Resizing state must be done 5947 // under the freelistLock (as is the check for whether to 5948 // allocate-live and whether to dirty the mod-union table). 5949 assert(_collectorState == Resizing, "Change of collector state to" 5950 " Resizing must be done under the freelistLocks (plural)"); 5951 5952 // Now that sweeping has been completed, if the GCH's 5953 // incremental_collection_will_fail flag is set, clear it, 5954 // thus inviting a younger gen collection to promote into 5955 // this generation. If such a promotion may still fail, 5956 // the flag will be set again when a young collection is 5957 // attempted. 5958 // I think the incremental_collection_will_fail flag's use 5959 // is specific to a 2 generation collection policy, so i'll 5960 // assert that that's the configuration we are operating within. 5961 // The use of the flag can and should be generalized appropriately 5962 // in the future to deal with a general n-generation system. 5963 5964 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5965 assert(gch->collector_policy()->is_two_generation_policy(), 5966 "Resetting of incremental_collection_will_fail flag" 5967 " may be incorrect otherwise"); 5968 gch->clear_incremental_collection_will_fail(); 5969 gch->update_full_collections_completed(_collection_count_start); 5970 } 5971 5972 // FIX ME!!! Looks like this belongs in CFLSpace, with 5973 // CMSGen merely delegating to it. 5974 void ConcurrentMarkSweepGeneration::setNearLargestChunk() { 5975 double nearLargestPercent = FLSLargestBlockCoalesceProximity; 5976 HeapWord* minAddr = _cmsSpace->bottom(); 5977 HeapWord* largestAddr = 5978 (HeapWord*) _cmsSpace->dictionary()->findLargestDict(); 5979 if (largestAddr == NULL) { 5980 // The dictionary appears to be empty. In this case 5981 // try to coalesce at the end of the heap. 5982 largestAddr = _cmsSpace->end(); 5983 } 5984 size_t largestOffset = pointer_delta(largestAddr, minAddr); 5985 size_t nearLargestOffset = 5986 (size_t)((double)largestOffset * nearLargestPercent) - MinChunkSize; 5987 if (PrintFLSStatistics != 0) { 5988 gclog_or_tty->print_cr( 5989 "CMS: Large Block: " PTR_FORMAT ";" 5990 " Proximity: " PTR_FORMAT " -> " PTR_FORMAT, 5991 largestAddr, 5992 _cmsSpace->nearLargestChunk(), minAddr + nearLargestOffset); 5993 } 5994 _cmsSpace->set_nearLargestChunk(minAddr + nearLargestOffset); 5995 } 5996 5997 bool ConcurrentMarkSweepGeneration::isNearLargestChunk(HeapWord* addr) { 5998 return addr >= _cmsSpace->nearLargestChunk(); 5999 } 6000 6001 FreeChunk* ConcurrentMarkSweepGeneration::find_chunk_at_end() { 6002 return _cmsSpace->find_chunk_at_end(); 6003 } 6004 6005 void ConcurrentMarkSweepGeneration::update_gc_stats(int current_level, 6006 bool full) { 6007 // The next lower level has been collected. Gather any statistics 6008 // that are of interest at this point. 6009 if (!full && (current_level + 1) == level()) { 6010 // Gather statistics on the young generation collection. 6011 collector()->stats().record_gc0_end(used()); 6012 } 6013 } 6014 6015 CMSAdaptiveSizePolicy* ConcurrentMarkSweepGeneration::size_policy() { 6016 GenCollectedHeap* gch = GenCollectedHeap::heap(); 6017 assert(gch->kind() == CollectedHeap::GenCollectedHeap, 6018 "Wrong type of heap"); 6019 CMSAdaptiveSizePolicy* sp = (CMSAdaptiveSizePolicy*) 6020 gch->gen_policy()->size_policy(); 6021 assert(sp->is_gc_cms_adaptive_size_policy(), 6022 "Wrong type of size policy"); 6023 return sp; 6024 } 6025 6026 void ConcurrentMarkSweepGeneration::rotate_debug_collection_type() { 6027 if (PrintGCDetails && Verbose) { 6028 gclog_or_tty->print("Rotate from %d ", _debug_collection_type); 6029 } 6030 _debug_collection_type = (CollectionTypes) (_debug_collection_type + 1); 6031 _debug_collection_type = 6032 (CollectionTypes) (_debug_collection_type % Unknown_collection_type); 6033 if (PrintGCDetails && Verbose) { 6034 gclog_or_tty->print_cr("to %d ", _debug_collection_type); 6035 } 6036 } 6037 6038 void CMSCollector::sweepWork(ConcurrentMarkSweepGeneration* gen, 6039 bool asynch) { 6040 // We iterate over the space(s) underlying this generation, 6041 // checking the mark bit map to see if the bits corresponding 6042 // to specific blocks are marked or not. Blocks that are 6043 // marked are live and are not swept up. All remaining blocks 6044 // are swept up, with coalescing on-the-fly as we sweep up 6045 // contiguous free and/or garbage blocks: 6046 // We need to ensure that the sweeper synchronizes with allocators 6047 // and stop-the-world collectors. In particular, the following 6048 // locks are used: 6049 // . CMS token: if this is held, a stop the world collection cannot occur 6050 // . freelistLock: if this is held no allocation can occur from this 6051 // generation by another thread 6052 // . bitMapLock: if this is held, no other thread can access or update 6053 // 6054 6055 // Note that we need to hold the freelistLock if we use 6056 // block iterate below; else the iterator might go awry if 6057 // a mutator (or promotion) causes block contents to change 6058 // (for instance if the allocator divvies up a block). 6059 // If we hold the free list lock, for all practical purposes 6060 // young generation GC's can't occur (they'll usually need to 6061 // promote), so we might as well prevent all young generation 6062 // GC's while we do a sweeping step. For the same reason, we might 6063 // as well take the bit map lock for the entire duration 6064 6065 // check that we hold the requisite locks 6066 assert(have_cms_token(), "Should hold cms token"); 6067 assert( (asynch && ConcurrentMarkSweepThread::cms_thread_has_cms_token()) 6068 || (!asynch && ConcurrentMarkSweepThread::vm_thread_has_cms_token()), 6069 "Should possess CMS token to sweep"); 6070 assert_lock_strong(gen->freelistLock()); 6071 assert_lock_strong(bitMapLock()); 6072 6073 assert(!_inter_sweep_timer.is_active(), "Was switched off in an outer context"); 6074 assert(_intra_sweep_timer.is_active(), "Was switched on in an outer context"); 6075 gen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()), 6076 _inter_sweep_estimate.padded_average(), 6077 _intra_sweep_estimate.padded_average()); 6078 gen->setNearLargestChunk(); 6079 6080 { 6081 SweepClosure sweepClosure(this, gen, &_markBitMap, 6082 CMSYield && asynch); 6083 gen->cmsSpace()->blk_iterate_careful(&sweepClosure); 6084 // We need to free-up/coalesce garbage/blocks from a 6085 // co-terminal free run. This is done in the SweepClosure 6086 // destructor; so, do not remove this scope, else the 6087 // end-of-sweep-census below will be off by a little bit. 6088 } 6089 gen->cmsSpace()->sweep_completed(); 6090 gen->cmsSpace()->endSweepFLCensus(sweep_count()); 6091 if (should_unload_classes()) { // unloaded classes this cycle, 6092 _concurrent_cycles_since_last_unload = 0; // ... reset count 6093 } else { // did not unload classes, 6094 _concurrent_cycles_since_last_unload++; // ... increment count 6095 } 6096 } 6097 6098 // Reset CMS data structures (for now just the marking bit map) 6099 // preparatory for the next cycle. 6100 void CMSCollector::reset(bool asynch) { 6101 GenCollectedHeap* gch = GenCollectedHeap::heap(); 6102 CMSAdaptiveSizePolicy* sp = size_policy(); 6103 AdaptiveSizePolicyOutput(sp, gch->total_collections()); 6104 if (asynch) { 6105 CMSTokenSyncWithLocks ts(true, bitMapLock()); 6106 6107 // If the state is not "Resetting", the foreground thread 6108 // has done a collection and the resetting. 6109 if (_collectorState != Resetting) { 6110 assert(_collectorState == Idling, "The state should only change" 6111 " because the foreground collector has finished the collection"); 6112 return; 6113 } 6114 6115 // Clear the mark bitmap (no grey objects to start with) 6116 // for the next cycle. 6117 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 6118 CMSPhaseAccounting cmspa(this, "reset", !PrintGCDetails); 6119 6120 HeapWord* curAddr = _markBitMap.startWord(); 6121 while (curAddr < _markBitMap.endWord()) { 6122 size_t remaining = pointer_delta(_markBitMap.endWord(), curAddr); 6123 MemRegion chunk(curAddr, MIN2(CMSBitMapYieldQuantum, remaining)); 6124 _markBitMap.clear_large_range(chunk); 6125 if (ConcurrentMarkSweepThread::should_yield() && 6126 !foregroundGCIsActive() && 6127 CMSYield) { 6128 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6129 "CMS thread should hold CMS token"); 6130 assert_lock_strong(bitMapLock()); 6131 bitMapLock()->unlock(); 6132 ConcurrentMarkSweepThread::desynchronize(true); 6133 ConcurrentMarkSweepThread::acknowledge_yield_request(); 6134 stopTimer(); 6135 if (PrintCMSStatistics != 0) { 6136 incrementYields(); 6137 } 6138 icms_wait(); 6139 6140 // See the comment in coordinator_yield() 6141 for (unsigned i = 0; i < CMSYieldSleepCount && 6142 ConcurrentMarkSweepThread::should_yield() && 6143 !CMSCollector::foregroundGCIsActive(); ++i) { 6144 os::sleep(Thread::current(), 1, false); 6145 ConcurrentMarkSweepThread::acknowledge_yield_request(); 6146 } 6147 6148 ConcurrentMarkSweepThread::synchronize(true); 6149 bitMapLock()->lock_without_safepoint_check(); 6150 startTimer(); 6151 } 6152 curAddr = chunk.end(); 6153 } 6154 // A successful mostly concurrent collection has been done. 6155 // Because only the full (i.e., concurrent mode failure) collections 6156 // are being measured for gc overhead limits, clean the "near" flag 6157 // and count. 6158 sp->reset_gc_overhead_limit_count(); 6159 _collectorState = Idling; 6160 } else { 6161 // already have the lock 6162 assert(_collectorState == Resetting, "just checking"); 6163 assert_lock_strong(bitMapLock()); 6164 _markBitMap.clear_all(); 6165 _collectorState = Idling; 6166 } 6167 6168 // Stop incremental mode after a cycle completes, so that any future cycles 6169 // are triggered by allocation. 6170 stop_icms(); 6171 6172 NOT_PRODUCT( 6173 if (RotateCMSCollectionTypes) { 6174 _cmsGen->rotate_debug_collection_type(); 6175 } 6176 ) 6177 } 6178 6179 void CMSCollector::do_CMS_operation(CMS_op_type op) { 6180 gclog_or_tty->date_stamp(PrintGC && PrintGCDateStamps); 6181 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 6182 TraceTime t("GC", PrintGC, !PrintGCDetails, gclog_or_tty); 6183 TraceCollectorStats tcs(counters()); 6184 6185 switch (op) { 6186 case CMS_op_checkpointRootsInitial: { 6187 checkpointRootsInitial(true); // asynch 6188 if (PrintGC) { 6189 _cmsGen->printOccupancy("initial-mark"); 6190 } 6191 break; 6192 } 6193 case CMS_op_checkpointRootsFinal: { 6194 checkpointRootsFinal(true, // asynch 6195 false, // !clear_all_soft_refs 6196 false); // !init_mark_was_synchronous 6197 if (PrintGC) { 6198 _cmsGen->printOccupancy("remark"); 6199 } 6200 break; 6201 } 6202 default: 6203 fatal("No such CMS_op"); 6204 } 6205 } 6206 6207 #ifndef PRODUCT 6208 size_t const CMSCollector::skip_header_HeapWords() { 6209 return FreeChunk::header_size(); 6210 } 6211 6212 // Try and collect here conditions that should hold when 6213 // CMS thread is exiting. The idea is that the foreground GC 6214 // thread should not be blocked if it wants to terminate 6215 // the CMS thread and yet continue to run the VM for a while 6216 // after that. 6217 void CMSCollector::verify_ok_to_terminate() const { 6218 assert(Thread::current()->is_ConcurrentGC_thread(), 6219 "should be called by CMS thread"); 6220 assert(!_foregroundGCShouldWait, "should be false"); 6221 // We could check here that all the various low-level locks 6222 // are not held by the CMS thread, but that is overkill; see 6223 // also CMSThread::verify_ok_to_terminate() where the CGC_lock 6224 // is checked. 6225 } 6226 #endif 6227 6228 size_t CMSCollector::block_size_using_printezis_bits(HeapWord* addr) const { 6229 assert(_markBitMap.isMarked(addr) && _markBitMap.isMarked(addr + 1), 6230 "missing Printezis mark?"); 6231 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 6232 size_t size = pointer_delta(nextOneAddr + 1, addr); 6233 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 6234 "alignment problem"); 6235 assert(size >= 3, "Necessary for Printezis marks to work"); 6236 return size; 6237 } 6238 6239 // A variant of the above (block_size_using_printezis_bits()) except 6240 // that we return 0 if the P-bits are not yet set. 6241 size_t CMSCollector::block_size_if_printezis_bits(HeapWord* addr) const { 6242 if (_markBitMap.isMarked(addr)) { 6243 assert(_markBitMap.isMarked(addr + 1), "Missing Printezis bit?"); 6244 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 6245 size_t size = pointer_delta(nextOneAddr + 1, addr); 6246 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 6247 "alignment problem"); 6248 assert(size >= 3, "Necessary for Printezis marks to work"); 6249 return size; 6250 } else { 6251 assert(!_markBitMap.isMarked(addr + 1), "Bit map inconsistency?"); 6252 return 0; 6253 } 6254 } 6255 6256 HeapWord* CMSCollector::next_card_start_after_block(HeapWord* addr) const { 6257 size_t sz = 0; 6258 oop p = (oop)addr; 6259 if (p->klass_or_null() != NULL && p->is_parsable()) { 6260 sz = CompactibleFreeListSpace::adjustObjectSize(p->size()); 6261 } else { 6262 sz = block_size_using_printezis_bits(addr); 6263 } 6264 assert(sz > 0, "size must be nonzero"); 6265 HeapWord* next_block = addr + sz; 6266 HeapWord* next_card = (HeapWord*)round_to((uintptr_t)next_block, 6267 CardTableModRefBS::card_size); 6268 assert(round_down((uintptr_t)addr, CardTableModRefBS::card_size) < 6269 round_down((uintptr_t)next_card, CardTableModRefBS::card_size), 6270 "must be different cards"); 6271 return next_card; 6272 } 6273 6274 6275 // CMS Bit Map Wrapper ///////////////////////////////////////// 6276 6277 // Construct a CMS bit map infrastructure, but don't create the 6278 // bit vector itself. That is done by a separate call CMSBitMap::allocate() 6279 // further below. 6280 CMSBitMap::CMSBitMap(int shifter, int mutex_rank, const char* mutex_name): 6281 _bm(), 6282 _shifter(shifter), 6283 _lock(mutex_rank >= 0 ? new Mutex(mutex_rank, mutex_name, true) : NULL) 6284 { 6285 _bmStartWord = 0; 6286 _bmWordSize = 0; 6287 } 6288 6289 bool CMSBitMap::allocate(MemRegion mr) { 6290 _bmStartWord = mr.start(); 6291 _bmWordSize = mr.word_size(); 6292 ReservedSpace brs(ReservedSpace::allocation_align_size_up( 6293 (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1)); 6294 if (!brs.is_reserved()) { 6295 warning("CMS bit map allocation failure"); 6296 return false; 6297 } 6298 // For now we'll just commit all of the bit map up fromt. 6299 // Later on we'll try to be more parsimonious with swap. 6300 if (!_virtual_space.initialize(brs, brs.size())) { 6301 warning("CMS bit map backing store failure"); 6302 return false; 6303 } 6304 assert(_virtual_space.committed_size() == brs.size(), 6305 "didn't reserve backing store for all of CMS bit map?"); 6306 _bm.set_map((BitMap::bm_word_t*)_virtual_space.low()); 6307 assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >= 6308 _bmWordSize, "inconsistency in bit map sizing"); 6309 _bm.set_size(_bmWordSize >> _shifter); 6310 6311 // bm.clear(); // can we rely on getting zero'd memory? verify below 6312 assert(isAllClear(), 6313 "Expected zero'd memory from ReservedSpace constructor"); 6314 assert(_bm.size() == heapWordDiffToOffsetDiff(sizeInWords()), 6315 "consistency check"); 6316 return true; 6317 } 6318 6319 void CMSBitMap::dirty_range_iterate_clear(MemRegion mr, MemRegionClosure* cl) { 6320 HeapWord *next_addr, *end_addr, *last_addr; 6321 assert_locked(); 6322 assert(covers(mr), "out-of-range error"); 6323 // XXX assert that start and end are appropriately aligned 6324 for (next_addr = mr.start(), end_addr = mr.end(); 6325 next_addr < end_addr; next_addr = last_addr) { 6326 MemRegion dirty_region = getAndClearMarkedRegion(next_addr, end_addr); 6327 last_addr = dirty_region.end(); 6328 if (!dirty_region.is_empty()) { 6329 cl->do_MemRegion(dirty_region); 6330 } else { 6331 assert(last_addr == end_addr, "program logic"); 6332 return; 6333 } 6334 } 6335 } 6336 6337 #ifndef PRODUCT 6338 void CMSBitMap::assert_locked() const { 6339 CMSLockVerifier::assert_locked(lock()); 6340 } 6341 6342 bool CMSBitMap::covers(MemRegion mr) const { 6343 // assert(_bm.map() == _virtual_space.low(), "map inconsistency"); 6344 assert((size_t)_bm.size() == (_bmWordSize >> _shifter), 6345 "size inconsistency"); 6346 return (mr.start() >= _bmStartWord) && 6347 (mr.end() <= endWord()); 6348 } 6349 6350 bool CMSBitMap::covers(HeapWord* start, size_t size) const { 6351 return (start >= _bmStartWord && (start + size) <= endWord()); 6352 } 6353 6354 void CMSBitMap::verifyNoOneBitsInRange(HeapWord* left, HeapWord* right) { 6355 // verify that there are no 1 bits in the interval [left, right) 6356 FalseBitMapClosure falseBitMapClosure; 6357 iterate(&falseBitMapClosure, left, right); 6358 } 6359 6360 void CMSBitMap::region_invariant(MemRegion mr) 6361 { 6362 assert_locked(); 6363 // mr = mr.intersection(MemRegion(_bmStartWord, _bmWordSize)); 6364 assert(!mr.is_empty(), "unexpected empty region"); 6365 assert(covers(mr), "mr should be covered by bit map"); 6366 // convert address range into offset range 6367 size_t start_ofs = heapWordToOffset(mr.start()); 6368 // Make sure that end() is appropriately aligned 6369 assert(mr.end() == (HeapWord*)round_to((intptr_t)mr.end(), 6370 (1 << (_shifter+LogHeapWordSize))), 6371 "Misaligned mr.end()"); 6372 size_t end_ofs = heapWordToOffset(mr.end()); 6373 assert(end_ofs > start_ofs, "Should mark at least one bit"); 6374 } 6375 6376 #endif 6377 6378 bool CMSMarkStack::allocate(size_t size) { 6379 // allocate a stack of the requisite depth 6380 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 6381 size * sizeof(oop))); 6382 if (!rs.is_reserved()) { 6383 warning("CMSMarkStack allocation failure"); 6384 return false; 6385 } 6386 if (!_virtual_space.initialize(rs, rs.size())) { 6387 warning("CMSMarkStack backing store failure"); 6388 return false; 6389 } 6390 assert(_virtual_space.committed_size() == rs.size(), 6391 "didn't reserve backing store for all of CMS stack?"); 6392 _base = (oop*)(_virtual_space.low()); 6393 _index = 0; 6394 _capacity = size; 6395 NOT_PRODUCT(_max_depth = 0); 6396 return true; 6397 } 6398 6399 // XXX FIX ME !!! In the MT case we come in here holding a 6400 // leaf lock. For printing we need to take a further lock 6401 // which has lower rank. We need to recallibrate the two 6402 // lock-ranks involved in order to be able to rpint the 6403 // messages below. (Or defer the printing to the caller. 6404 // For now we take the expedient path of just disabling the 6405 // messages for the problematic case.) 6406 void CMSMarkStack::expand() { 6407 assert(_capacity <= MarkStackSizeMax, "stack bigger than permitted"); 6408 if (_capacity == MarkStackSizeMax) { 6409 if (_hit_limit++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) { 6410 // We print a warning message only once per CMS cycle. 6411 gclog_or_tty->print_cr(" (benign) Hit CMSMarkStack max size limit"); 6412 } 6413 return; 6414 } 6415 // Double capacity if possible 6416 size_t new_capacity = MIN2(_capacity*2, MarkStackSizeMax); 6417 // Do not give up existing stack until we have managed to 6418 // get the double capacity that we desired. 6419 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 6420 new_capacity * sizeof(oop))); 6421 if (rs.is_reserved()) { 6422 // Release the backing store associated with old stack 6423 _virtual_space.release(); 6424 // Reinitialize virtual space for new stack 6425 if (!_virtual_space.initialize(rs, rs.size())) { 6426 fatal("Not enough swap for expanded marking stack"); 6427 } 6428 _base = (oop*)(_virtual_space.low()); 6429 _index = 0; 6430 _capacity = new_capacity; 6431 } else if (_failed_double++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) { 6432 // Failed to double capacity, continue; 6433 // we print a detail message only once per CMS cycle. 6434 gclog_or_tty->print(" (benign) Failed to expand marking stack from "SIZE_FORMAT"K to " 6435 SIZE_FORMAT"K", 6436 _capacity / K, new_capacity / K); 6437 } 6438 } 6439 6440 6441 // Closures 6442 // XXX: there seems to be a lot of code duplication here; 6443 // should refactor and consolidate common code. 6444 6445 // This closure is used to mark refs into the CMS generation in 6446 // the CMS bit map. Called at the first checkpoint. This closure 6447 // assumes that we do not need to re-mark dirty cards; if the CMS 6448 // generation on which this is used is not an oldest (modulo perm gen) 6449 // generation then this will lose younger_gen cards! 6450 6451 MarkRefsIntoClosure::MarkRefsIntoClosure( 6452 MemRegion span, CMSBitMap* bitMap): 6453 _span(span), 6454 _bitMap(bitMap) 6455 { 6456 assert(_ref_processor == NULL, "deliberately left NULL"); 6457 assert(_bitMap->covers(_span), "_bitMap/_span mismatch"); 6458 } 6459 6460 void MarkRefsIntoClosure::do_oop(oop obj) { 6461 // if p points into _span, then mark corresponding bit in _markBitMap 6462 assert(obj->is_oop(), "expected an oop"); 6463 HeapWord* addr = (HeapWord*)obj; 6464 if (_span.contains(addr)) { 6465 // this should be made more efficient 6466 _bitMap->mark(addr); 6467 } 6468 } 6469 6470 void MarkRefsIntoClosure::do_oop(oop* p) { MarkRefsIntoClosure::do_oop_work(p); } 6471 void MarkRefsIntoClosure::do_oop(narrowOop* p) { MarkRefsIntoClosure::do_oop_work(p); } 6472 6473 // A variant of the above, used for CMS marking verification. 6474 MarkRefsIntoVerifyClosure::MarkRefsIntoVerifyClosure( 6475 MemRegion span, CMSBitMap* verification_bm, CMSBitMap* cms_bm): 6476 _span(span), 6477 _verification_bm(verification_bm), 6478 _cms_bm(cms_bm) 6479 { 6480 assert(_ref_processor == NULL, "deliberately left NULL"); 6481 assert(_verification_bm->covers(_span), "_verification_bm/_span mismatch"); 6482 } 6483 6484 void MarkRefsIntoVerifyClosure::do_oop(oop obj) { 6485 // if p points into _span, then mark corresponding bit in _markBitMap 6486 assert(obj->is_oop(), "expected an oop"); 6487 HeapWord* addr = (HeapWord*)obj; 6488 if (_span.contains(addr)) { 6489 _verification_bm->mark(addr); 6490 if (!_cms_bm->isMarked(addr)) { 6491 oop(addr)->print(); 6492 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)", addr); 6493 fatal("... aborting"); 6494 } 6495 } 6496 } 6497 6498 void MarkRefsIntoVerifyClosure::do_oop(oop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); } 6499 void MarkRefsIntoVerifyClosure::do_oop(narrowOop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); } 6500 6501 ////////////////////////////////////////////////// 6502 // MarkRefsIntoAndScanClosure 6503 ////////////////////////////////////////////////// 6504 6505 MarkRefsIntoAndScanClosure::MarkRefsIntoAndScanClosure(MemRegion span, 6506 ReferenceProcessor* rp, 6507 CMSBitMap* bit_map, 6508 CMSBitMap* mod_union_table, 6509 CMSMarkStack* mark_stack, 6510 CMSMarkStack* revisit_stack, 6511 CMSCollector* collector, 6512 bool should_yield, 6513 bool concurrent_precleaning): 6514 _collector(collector), 6515 _span(span), 6516 _bit_map(bit_map), 6517 _mark_stack(mark_stack), 6518 _pushAndMarkClosure(collector, span, rp, bit_map, mod_union_table, 6519 mark_stack, revisit_stack, concurrent_precleaning), 6520 _yield(should_yield), 6521 _concurrent_precleaning(concurrent_precleaning), 6522 _freelistLock(NULL) 6523 { 6524 _ref_processor = rp; 6525 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 6526 } 6527 6528 // This closure is used to mark refs into the CMS generation at the 6529 // second (final) checkpoint, and to scan and transitively follow 6530 // the unmarked oops. It is also used during the concurrent precleaning 6531 // phase while scanning objects on dirty cards in the CMS generation. 6532 // The marks are made in the marking bit map and the marking stack is 6533 // used for keeping the (newly) grey objects during the scan. 6534 // The parallel version (Par_...) appears further below. 6535 void MarkRefsIntoAndScanClosure::do_oop(oop obj) { 6536 if (obj != NULL) { 6537 assert(obj->is_oop(), "expected an oop"); 6538 HeapWord* addr = (HeapWord*)obj; 6539 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 6540 assert(_collector->overflow_list_is_empty(), 6541 "overflow list should be empty"); 6542 if (_span.contains(addr) && 6543 !_bit_map->isMarked(addr)) { 6544 // mark bit map (object is now grey) 6545 _bit_map->mark(addr); 6546 // push on marking stack (stack should be empty), and drain the 6547 // stack by applying this closure to the oops in the oops popped 6548 // from the stack (i.e. blacken the grey objects) 6549 bool res = _mark_stack->push(obj); 6550 assert(res, "Should have space to push on empty stack"); 6551 do { 6552 oop new_oop = _mark_stack->pop(); 6553 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 6554 assert(new_oop->is_parsable(), "Found unparsable oop"); 6555 assert(_bit_map->isMarked((HeapWord*)new_oop), 6556 "only grey objects on this stack"); 6557 // iterate over the oops in this oop, marking and pushing 6558 // the ones in CMS heap (i.e. in _span). 6559 new_oop->oop_iterate(&_pushAndMarkClosure); 6560 // check if it's time to yield 6561 do_yield_check(); 6562 } while (!_mark_stack->isEmpty() || 6563 (!_concurrent_precleaning && take_from_overflow_list())); 6564 // if marking stack is empty, and we are not doing this 6565 // during precleaning, then check the overflow list 6566 } 6567 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 6568 assert(_collector->overflow_list_is_empty(), 6569 "overflow list was drained above"); 6570 // We could restore evacuated mark words, if any, used for 6571 // overflow list links here because the overflow list is 6572 // provably empty here. That would reduce the maximum 6573 // size requirements for preserved_{oop,mark}_stack. 6574 // But we'll just postpone it until we are all done 6575 // so we can just stream through. 6576 if (!_concurrent_precleaning && CMSOverflowEarlyRestoration) { 6577 _collector->restore_preserved_marks_if_any(); 6578 assert(_collector->no_preserved_marks(), "No preserved marks"); 6579 } 6580 assert(!CMSOverflowEarlyRestoration || _collector->no_preserved_marks(), 6581 "All preserved marks should have been restored above"); 6582 } 6583 } 6584 6585 void MarkRefsIntoAndScanClosure::do_oop(oop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); } 6586 void MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); } 6587 6588 void MarkRefsIntoAndScanClosure::do_yield_work() { 6589 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6590 "CMS thread should hold CMS token"); 6591 assert_lock_strong(_freelistLock); 6592 assert_lock_strong(_bit_map->lock()); 6593 // relinquish the free_list_lock and bitMaplock() 6594 DEBUG_ONLY(RememberKlassesChecker mux(false);) 6595 _bit_map->lock()->unlock(); 6596 _freelistLock->unlock(); 6597 ConcurrentMarkSweepThread::desynchronize(true); 6598 ConcurrentMarkSweepThread::acknowledge_yield_request(); 6599 _collector->stopTimer(); 6600 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 6601 if (PrintCMSStatistics != 0) { 6602 _collector->incrementYields(); 6603 } 6604 _collector->icms_wait(); 6605 6606 // See the comment in coordinator_yield() 6607 for (unsigned i = 0; 6608 i < CMSYieldSleepCount && 6609 ConcurrentMarkSweepThread::should_yield() && 6610 !CMSCollector::foregroundGCIsActive(); 6611 ++i) { 6612 os::sleep(Thread::current(), 1, false); 6613 ConcurrentMarkSweepThread::acknowledge_yield_request(); 6614 } 6615 6616 ConcurrentMarkSweepThread::synchronize(true); 6617 _freelistLock->lock_without_safepoint_check(); 6618 _bit_map->lock()->lock_without_safepoint_check(); 6619 _collector->startTimer(); 6620 } 6621 6622 /////////////////////////////////////////////////////////// 6623 // Par_MarkRefsIntoAndScanClosure: a parallel version of 6624 // MarkRefsIntoAndScanClosure 6625 /////////////////////////////////////////////////////////// 6626 Par_MarkRefsIntoAndScanClosure::Par_MarkRefsIntoAndScanClosure( 6627 CMSCollector* collector, MemRegion span, ReferenceProcessor* rp, 6628 CMSBitMap* bit_map, OopTaskQueue* work_queue, CMSMarkStack* revisit_stack): 6629 _span(span), 6630 _bit_map(bit_map), 6631 _work_queue(work_queue), 6632 _low_water_mark(MIN2((uint)(work_queue->max_elems()/4), 6633 (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads))), 6634 _par_pushAndMarkClosure(collector, span, rp, bit_map, work_queue, 6635 revisit_stack) 6636 { 6637 _ref_processor = rp; 6638 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 6639 } 6640 6641 // This closure is used to mark refs into the CMS generation at the 6642 // second (final) checkpoint, and to scan and transitively follow 6643 // the unmarked oops. The marks are made in the marking bit map and 6644 // the work_queue is used for keeping the (newly) grey objects during 6645 // the scan phase whence they are also available for stealing by parallel 6646 // threads. Since the marking bit map is shared, updates are 6647 // synchronized (via CAS). 6648 void Par_MarkRefsIntoAndScanClosure::do_oop(oop obj) { 6649 if (obj != NULL) { 6650 // Ignore mark word because this could be an already marked oop 6651 // that may be chained at the end of the overflow list. 6652 assert(obj->is_oop(true), "expected an oop"); 6653 HeapWord* addr = (HeapWord*)obj; 6654 if (_span.contains(addr) && 6655 !_bit_map->isMarked(addr)) { 6656 // mark bit map (object will become grey): 6657 // It is possible for several threads to be 6658 // trying to "claim" this object concurrently; 6659 // the unique thread that succeeds in marking the 6660 // object first will do the subsequent push on 6661 // to the work queue (or overflow list). 6662 if (_bit_map->par_mark(addr)) { 6663 // push on work_queue (which may not be empty), and trim the 6664 // queue to an appropriate length by applying this closure to 6665 // the oops in the oops popped from the stack (i.e. blacken the 6666 // grey objects) 6667 bool res = _work_queue->push(obj); 6668 assert(res, "Low water mark should be less than capacity?"); 6669 trim_queue(_low_water_mark); 6670 } // Else, another thread claimed the object 6671 } 6672 } 6673 } 6674 6675 void Par_MarkRefsIntoAndScanClosure::do_oop(oop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); } 6676 void Par_MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); } 6677 6678 // This closure is used to rescan the marked objects on the dirty cards 6679 // in the mod union table and the card table proper. 6680 size_t ScanMarkedObjectsAgainCarefullyClosure::do_object_careful_m( 6681 oop p, MemRegion mr) { 6682 6683 size_t size = 0; 6684 HeapWord* addr = (HeapWord*)p; 6685 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6686 assert(_span.contains(addr), "we are scanning the CMS generation"); 6687 // check if it's time to yield 6688 if (do_yield_check()) { 6689 // We yielded for some foreground stop-world work, 6690 // and we have been asked to abort this ongoing preclean cycle. 6691 return 0; 6692 } 6693 if (_bitMap->isMarked(addr)) { 6694 // it's marked; is it potentially uninitialized? 6695 if (p->klass_or_null() != NULL) { 6696 // If is_conc_safe is false, the object may be undergoing 6697 // change by the VM outside a safepoint. Don't try to 6698 // scan it, but rather leave it for the remark phase. 6699 if (CMSPermGenPrecleaningEnabled && 6700 (!p->is_conc_safe() || !p->is_parsable())) { 6701 // Signal precleaning to redirty the card since 6702 // the klass pointer is already installed. 6703 assert(size == 0, "Initial value"); 6704 } else { 6705 assert(p->is_parsable(), "must be parsable."); 6706 // an initialized object; ignore mark word in verification below 6707 // since we are running concurrent with mutators 6708 assert(p->is_oop(true), "should be an oop"); 6709 if (p->is_objArray()) { 6710 // objArrays are precisely marked; restrict scanning 6711 // to dirty cards only. 6712 size = CompactibleFreeListSpace::adjustObjectSize( 6713 p->oop_iterate(_scanningClosure, mr)); 6714 } else { 6715 // A non-array may have been imprecisely marked; we need 6716 // to scan object in its entirety. 6717 size = CompactibleFreeListSpace::adjustObjectSize( 6718 p->oop_iterate(_scanningClosure)); 6719 } 6720 #ifdef DEBUG 6721 size_t direct_size = 6722 CompactibleFreeListSpace::adjustObjectSize(p->size()); 6723 assert(size == direct_size, "Inconsistency in size"); 6724 assert(size >= 3, "Necessary for Printezis marks to work"); 6725 if (!_bitMap->isMarked(addr+1)) { 6726 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size); 6727 } else { 6728 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size-1); 6729 assert(_bitMap->isMarked(addr+size-1), 6730 "inconsistent Printezis mark"); 6731 } 6732 #endif // DEBUG 6733 } 6734 } else { 6735 // an unitialized object 6736 assert(_bitMap->isMarked(addr+1), "missing Printezis mark?"); 6737 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 6738 size = pointer_delta(nextOneAddr + 1, addr); 6739 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 6740 "alignment problem"); 6741 // Note that pre-cleaning needn't redirty the card. OopDesc::set_klass() 6742 // will dirty the card when the klass pointer is installed in the 6743 // object (signalling the completion of initialization). 6744 } 6745 } else { 6746 // Either a not yet marked object or an uninitialized object 6747 if (p->klass_or_null() == NULL || !p->is_parsable()) { 6748 // An uninitialized object, skip to the next card, since 6749 // we may not be able to read its P-bits yet. 6750 assert(size == 0, "Initial value"); 6751 } else { 6752 // An object not (yet) reached by marking: we merely need to 6753 // compute its size so as to go look at the next block. 6754 assert(p->is_oop(true), "should be an oop"); 6755 size = CompactibleFreeListSpace::adjustObjectSize(p->size()); 6756 } 6757 } 6758 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6759 return size; 6760 } 6761 6762 void ScanMarkedObjectsAgainCarefullyClosure::do_yield_work() { 6763 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6764 "CMS thread should hold CMS token"); 6765 assert_lock_strong(_freelistLock); 6766 assert_lock_strong(_bitMap->lock()); 6767 DEBUG_ONLY(RememberKlassesChecker mux(false);) 6768 // relinquish the free_list_lock and bitMaplock() 6769 _bitMap->lock()->unlock(); 6770 _freelistLock->unlock(); 6771 ConcurrentMarkSweepThread::desynchronize(true); 6772 ConcurrentMarkSweepThread::acknowledge_yield_request(); 6773 _collector->stopTimer(); 6774 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 6775 if (PrintCMSStatistics != 0) { 6776 _collector->incrementYields(); 6777 } 6778 _collector->icms_wait(); 6779 6780 // See the comment in coordinator_yield() 6781 for (unsigned i = 0; i < CMSYieldSleepCount && 6782 ConcurrentMarkSweepThread::should_yield() && 6783 !CMSCollector::foregroundGCIsActive(); ++i) { 6784 os::sleep(Thread::current(), 1, false); 6785 ConcurrentMarkSweepThread::acknowledge_yield_request(); 6786 } 6787 6788 ConcurrentMarkSweepThread::synchronize(true); 6789 _freelistLock->lock_without_safepoint_check(); 6790 _bitMap->lock()->lock_without_safepoint_check(); 6791 _collector->startTimer(); 6792 } 6793 6794 6795 ////////////////////////////////////////////////////////////////// 6796 // SurvivorSpacePrecleanClosure 6797 ////////////////////////////////////////////////////////////////// 6798 // This (single-threaded) closure is used to preclean the oops in 6799 // the survivor spaces. 6800 size_t SurvivorSpacePrecleanClosure::do_object_careful(oop p) { 6801 6802 HeapWord* addr = (HeapWord*)p; 6803 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6804 assert(!_span.contains(addr), "we are scanning the survivor spaces"); 6805 assert(p->klass_or_null() != NULL, "object should be initializd"); 6806 assert(p->is_parsable(), "must be parsable."); 6807 // an initialized object; ignore mark word in verification below 6808 // since we are running concurrent with mutators 6809 assert(p->is_oop(true), "should be an oop"); 6810 // Note that we do not yield while we iterate over 6811 // the interior oops of p, pushing the relevant ones 6812 // on our marking stack. 6813 size_t size = p->oop_iterate(_scanning_closure); 6814 do_yield_check(); 6815 // Observe that below, we do not abandon the preclean 6816 // phase as soon as we should; rather we empty the 6817 // marking stack before returning. This is to satisfy 6818 // some existing assertions. In general, it may be a 6819 // good idea to abort immediately and complete the marking 6820 // from the grey objects at a later time. 6821 while (!_mark_stack->isEmpty()) { 6822 oop new_oop = _mark_stack->pop(); 6823 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 6824 assert(new_oop->is_parsable(), "Found unparsable oop"); 6825 assert(_bit_map->isMarked((HeapWord*)new_oop), 6826 "only grey objects on this stack"); 6827 // iterate over the oops in this oop, marking and pushing 6828 // the ones in CMS heap (i.e. in _span). 6829 new_oop->oop_iterate(_scanning_closure); 6830 // check if it's time to yield 6831 do_yield_check(); 6832 } 6833 unsigned int after_count = 6834 GenCollectedHeap::heap()->total_collections(); 6835 bool abort = (_before_count != after_count) || 6836 _collector->should_abort_preclean(); 6837 return abort ? 0 : size; 6838 } 6839 6840 void SurvivorSpacePrecleanClosure::do_yield_work() { 6841 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6842 "CMS thread should hold CMS token"); 6843 assert_lock_strong(_bit_map->lock()); 6844 DEBUG_ONLY(RememberKlassesChecker smx(false);) 6845 // Relinquish the bit map lock 6846 _bit_map->lock()->unlock(); 6847 ConcurrentMarkSweepThread::desynchronize(true); 6848 ConcurrentMarkSweepThread::acknowledge_yield_request(); 6849 _collector->stopTimer(); 6850 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 6851 if (PrintCMSStatistics != 0) { 6852 _collector->incrementYields(); 6853 } 6854 _collector->icms_wait(); 6855 6856 // See the comment in coordinator_yield() 6857 for (unsigned i = 0; i < CMSYieldSleepCount && 6858 ConcurrentMarkSweepThread::should_yield() && 6859 !CMSCollector::foregroundGCIsActive(); ++i) { 6860 os::sleep(Thread::current(), 1, false); 6861 ConcurrentMarkSweepThread::acknowledge_yield_request(); 6862 } 6863 6864 ConcurrentMarkSweepThread::synchronize(true); 6865 _bit_map->lock()->lock_without_safepoint_check(); 6866 _collector->startTimer(); 6867 } 6868 6869 // This closure is used to rescan the marked objects on the dirty cards 6870 // in the mod union table and the card table proper. In the parallel 6871 // case, although the bitMap is shared, we do a single read so the 6872 // isMarked() query is "safe". 6873 bool ScanMarkedObjectsAgainClosure::do_object_bm(oop p, MemRegion mr) { 6874 // Ignore mark word because we are running concurrent with mutators 6875 assert(p->is_oop_or_null(true), "expected an oop or null"); 6876 HeapWord* addr = (HeapWord*)p; 6877 assert(_span.contains(addr), "we are scanning the CMS generation"); 6878 bool is_obj_array = false; 6879 #ifdef DEBUG 6880 if (!_parallel) { 6881 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 6882 assert(_collector->overflow_list_is_empty(), 6883 "overflow list should be empty"); 6884 6885 } 6886 #endif // DEBUG 6887 if (_bit_map->isMarked(addr)) { 6888 // Obj arrays are precisely marked, non-arrays are not; 6889 // so we scan objArrays precisely and non-arrays in their 6890 // entirety. 6891 if (p->is_objArray()) { 6892 is_obj_array = true; 6893 if (_parallel) { 6894 p->oop_iterate(_par_scan_closure, mr); 6895 } else { 6896 p->oop_iterate(_scan_closure, mr); 6897 } 6898 } else { 6899 if (_parallel) { 6900 p->oop_iterate(_par_scan_closure); 6901 } else { 6902 p->oop_iterate(_scan_closure); 6903 } 6904 } 6905 } 6906 #ifdef DEBUG 6907 if (!_parallel) { 6908 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 6909 assert(_collector->overflow_list_is_empty(), 6910 "overflow list should be empty"); 6911 6912 } 6913 #endif // DEBUG 6914 return is_obj_array; 6915 } 6916 6917 MarkFromRootsClosure::MarkFromRootsClosure(CMSCollector* collector, 6918 MemRegion span, 6919 CMSBitMap* bitMap, CMSMarkStack* markStack, 6920 CMSMarkStack* revisitStack, 6921 bool should_yield, bool verifying): 6922 _collector(collector), 6923 _span(span), 6924 _bitMap(bitMap), 6925 _mut(&collector->_modUnionTable), 6926 _markStack(markStack), 6927 _revisitStack(revisitStack), 6928 _yield(should_yield), 6929 _skipBits(0) 6930 { 6931 assert(_markStack->isEmpty(), "stack should be empty"); 6932 _finger = _bitMap->startWord(); 6933 _threshold = _finger; 6934 assert(_collector->_restart_addr == NULL, "Sanity check"); 6935 assert(_span.contains(_finger), "Out of bounds _finger?"); 6936 DEBUG_ONLY(_verifying = verifying;) 6937 } 6938 6939 void MarkFromRootsClosure::reset(HeapWord* addr) { 6940 assert(_markStack->isEmpty(), "would cause duplicates on stack"); 6941 assert(_span.contains(addr), "Out of bounds _finger?"); 6942 _finger = addr; 6943 _threshold = (HeapWord*)round_to( 6944 (intptr_t)_finger, CardTableModRefBS::card_size); 6945 } 6946 6947 // Should revisit to see if this should be restructured for 6948 // greater efficiency. 6949 bool MarkFromRootsClosure::do_bit(size_t offset) { 6950 if (_skipBits > 0) { 6951 _skipBits--; 6952 return true; 6953 } 6954 // convert offset into a HeapWord* 6955 HeapWord* addr = _bitMap->startWord() + offset; 6956 assert(_bitMap->endWord() && addr < _bitMap->endWord(), 6957 "address out of range"); 6958 assert(_bitMap->isMarked(addr), "tautology"); 6959 if (_bitMap->isMarked(addr+1)) { 6960 // this is an allocated but not yet initialized object 6961 assert(_skipBits == 0, "tautology"); 6962 _skipBits = 2; // skip next two marked bits ("Printezis-marks") 6963 oop p = oop(addr); 6964 if (p->klass_or_null() == NULL || !p->is_parsable()) { 6965 DEBUG_ONLY(if (!_verifying) {) 6966 // We re-dirty the cards on which this object lies and increase 6967 // the _threshold so that we'll come back to scan this object 6968 // during the preclean or remark phase. (CMSCleanOnEnter) 6969 if (CMSCleanOnEnter) { 6970 size_t sz = _collector->block_size_using_printezis_bits(addr); 6971 HeapWord* end_card_addr = (HeapWord*)round_to( 6972 (intptr_t)(addr+sz), CardTableModRefBS::card_size); 6973 MemRegion redirty_range = MemRegion(addr, end_card_addr); 6974 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 6975 // Bump _threshold to end_card_addr; note that 6976 // _threshold cannot possibly exceed end_card_addr, anyhow. 6977 // This prevents future clearing of the card as the scan proceeds 6978 // to the right. 6979 assert(_threshold <= end_card_addr, 6980 "Because we are just scanning into this object"); 6981 if (_threshold < end_card_addr) { 6982 _threshold = end_card_addr; 6983 } 6984 if (p->klass_or_null() != NULL) { 6985 // Redirty the range of cards... 6986 _mut->mark_range(redirty_range); 6987 } // ...else the setting of klass will dirty the card anyway. 6988 } 6989 DEBUG_ONLY(}) 6990 return true; 6991 } 6992 } 6993 scanOopsInOop(addr); 6994 return true; 6995 } 6996 6997 // We take a break if we've been at this for a while, 6998 // so as to avoid monopolizing the locks involved. 6999 void MarkFromRootsClosure::do_yield_work() { 7000 // First give up the locks, then yield, then re-lock 7001 // We should probably use a constructor/destructor idiom to 7002 // do this unlock/lock or modify the MutexUnlocker class to 7003 // serve our purpose. XXX 7004 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 7005 "CMS thread should hold CMS token"); 7006 assert_lock_strong(_bitMap->lock()); 7007 DEBUG_ONLY(RememberKlassesChecker mux(false);) 7008 _bitMap->lock()->unlock(); 7009 ConcurrentMarkSweepThread::desynchronize(true); 7010 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7011 _collector->stopTimer(); 7012 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 7013 if (PrintCMSStatistics != 0) { 7014 _collector->incrementYields(); 7015 } 7016 _collector->icms_wait(); 7017 7018 // See the comment in coordinator_yield() 7019 for (unsigned i = 0; i < CMSYieldSleepCount && 7020 ConcurrentMarkSweepThread::should_yield() && 7021 !CMSCollector::foregroundGCIsActive(); ++i) { 7022 os::sleep(Thread::current(), 1, false); 7023 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7024 } 7025 7026 ConcurrentMarkSweepThread::synchronize(true); 7027 _bitMap->lock()->lock_without_safepoint_check(); 7028 _collector->startTimer(); 7029 } 7030 7031 void MarkFromRootsClosure::scanOopsInOop(HeapWord* ptr) { 7032 assert(_bitMap->isMarked(ptr), "expected bit to be set"); 7033 assert(_markStack->isEmpty(), 7034 "should drain stack to limit stack usage"); 7035 // convert ptr to an oop preparatory to scanning 7036 oop obj = oop(ptr); 7037 // Ignore mark word in verification below, since we 7038 // may be running concurrent with mutators. 7039 assert(obj->is_oop(true), "should be an oop"); 7040 assert(_finger <= ptr, "_finger runneth ahead"); 7041 // advance the finger to right end of this object 7042 _finger = ptr + obj->size(); 7043 assert(_finger > ptr, "we just incremented it above"); 7044 // On large heaps, it may take us some time to get through 7045 // the marking phase (especially if running iCMS). During 7046 // this time it's possible that a lot of mutations have 7047 // accumulated in the card table and the mod union table -- 7048 // these mutation records are redundant until we have 7049 // actually traced into the corresponding card. 7050 // Here, we check whether advancing the finger would make 7051 // us cross into a new card, and if so clear corresponding 7052 // cards in the MUT (preclean them in the card-table in the 7053 // future). 7054 7055 DEBUG_ONLY(if (!_verifying) {) 7056 // The clean-on-enter optimization is disabled by default, 7057 // until we fix 6178663. 7058 if (CMSCleanOnEnter && (_finger > _threshold)) { 7059 // [_threshold, _finger) represents the interval 7060 // of cards to be cleared in MUT (or precleaned in card table). 7061 // The set of cards to be cleared is all those that overlap 7062 // with the interval [_threshold, _finger); note that 7063 // _threshold is always kept card-aligned but _finger isn't 7064 // always card-aligned. 7065 HeapWord* old_threshold = _threshold; 7066 assert(old_threshold == (HeapWord*)round_to( 7067 (intptr_t)old_threshold, CardTableModRefBS::card_size), 7068 "_threshold should always be card-aligned"); 7069 _threshold = (HeapWord*)round_to( 7070 (intptr_t)_finger, CardTableModRefBS::card_size); 7071 MemRegion mr(old_threshold, _threshold); 7072 assert(!mr.is_empty(), "Control point invariant"); 7073 assert(_span.contains(mr), "Should clear within span"); 7074 // XXX When _finger crosses from old gen into perm gen 7075 // we may be doing unnecessary cleaning; do better in the 7076 // future by detecting that condition and clearing fewer 7077 // MUT/CT entries. 7078 _mut->clear_range(mr); 7079 } 7080 DEBUG_ONLY(}) 7081 // Note: the finger doesn't advance while we drain 7082 // the stack below. 7083 PushOrMarkClosure pushOrMarkClosure(_collector, 7084 _span, _bitMap, _markStack, 7085 _revisitStack, 7086 _finger, this); 7087 bool res = _markStack->push(obj); 7088 assert(res, "Empty non-zero size stack should have space for single push"); 7089 while (!_markStack->isEmpty()) { 7090 oop new_oop = _markStack->pop(); 7091 // Skip verifying header mark word below because we are 7092 // running concurrent with mutators. 7093 assert(new_oop->is_oop(true), "Oops! expected to pop an oop"); 7094 // now scan this oop's oops 7095 new_oop->oop_iterate(&pushOrMarkClosure); 7096 do_yield_check(); 7097 } 7098 assert(_markStack->isEmpty(), "tautology, emphasizing post-condition"); 7099 } 7100 7101 Par_MarkFromRootsClosure::Par_MarkFromRootsClosure(CMSConcMarkingTask* task, 7102 CMSCollector* collector, MemRegion span, 7103 CMSBitMap* bit_map, 7104 OopTaskQueue* work_queue, 7105 CMSMarkStack* overflow_stack, 7106 CMSMarkStack* revisit_stack, 7107 bool should_yield): 7108 _collector(collector), 7109 _whole_span(collector->_span), 7110 _span(span), 7111 _bit_map(bit_map), 7112 _mut(&collector->_modUnionTable), 7113 _work_queue(work_queue), 7114 _overflow_stack(overflow_stack), 7115 _revisit_stack(revisit_stack), 7116 _yield(should_yield), 7117 _skip_bits(0), 7118 _task(task) 7119 { 7120 assert(_work_queue->size() == 0, "work_queue should be empty"); 7121 _finger = span.start(); 7122 _threshold = _finger; // XXX Defer clear-on-enter optimization for now 7123 assert(_span.contains(_finger), "Out of bounds _finger?"); 7124 } 7125 7126 // Should revisit to see if this should be restructured for 7127 // greater efficiency. 7128 bool Par_MarkFromRootsClosure::do_bit(size_t offset) { 7129 if (_skip_bits > 0) { 7130 _skip_bits--; 7131 return true; 7132 } 7133 // convert offset into a HeapWord* 7134 HeapWord* addr = _bit_map->startWord() + offset; 7135 assert(_bit_map->endWord() && addr < _bit_map->endWord(), 7136 "address out of range"); 7137 assert(_bit_map->isMarked(addr), "tautology"); 7138 if (_bit_map->isMarked(addr+1)) { 7139 // this is an allocated object that might not yet be initialized 7140 assert(_skip_bits == 0, "tautology"); 7141 _skip_bits = 2; // skip next two marked bits ("Printezis-marks") 7142 oop p = oop(addr); 7143 if (p->klass_or_null() == NULL || !p->is_parsable()) { 7144 // in the case of Clean-on-Enter optimization, redirty card 7145 // and avoid clearing card by increasing the threshold. 7146 return true; 7147 } 7148 } 7149 scan_oops_in_oop(addr); 7150 return true; 7151 } 7152 7153 void Par_MarkFromRootsClosure::scan_oops_in_oop(HeapWord* ptr) { 7154 assert(_bit_map->isMarked(ptr), "expected bit to be set"); 7155 // Should we assert that our work queue is empty or 7156 // below some drain limit? 7157 assert(_work_queue->size() == 0, 7158 "should drain stack to limit stack usage"); 7159 // convert ptr to an oop preparatory to scanning 7160 oop obj = oop(ptr); 7161 // Ignore mark word in verification below, since we 7162 // may be running concurrent with mutators. 7163 assert(obj->is_oop(true), "should be an oop"); 7164 assert(_finger <= ptr, "_finger runneth ahead"); 7165 // advance the finger to right end of this object 7166 _finger = ptr + obj->size(); 7167 assert(_finger > ptr, "we just incremented it above"); 7168 // On large heaps, it may take us some time to get through 7169 // the marking phase (especially if running iCMS). During 7170 // this time it's possible that a lot of mutations have 7171 // accumulated in the card table and the mod union table -- 7172 // these mutation records are redundant until we have 7173 // actually traced into the corresponding card. 7174 // Here, we check whether advancing the finger would make 7175 // us cross into a new card, and if so clear corresponding 7176 // cards in the MUT (preclean them in the card-table in the 7177 // future). 7178 7179 // The clean-on-enter optimization is disabled by default, 7180 // until we fix 6178663. 7181 if (CMSCleanOnEnter && (_finger > _threshold)) { 7182 // [_threshold, _finger) represents the interval 7183 // of cards to be cleared in MUT (or precleaned in card table). 7184 // The set of cards to be cleared is all those that overlap 7185 // with the interval [_threshold, _finger); note that 7186 // _threshold is always kept card-aligned but _finger isn't 7187 // always card-aligned. 7188 HeapWord* old_threshold = _threshold; 7189 assert(old_threshold == (HeapWord*)round_to( 7190 (intptr_t)old_threshold, CardTableModRefBS::card_size), 7191 "_threshold should always be card-aligned"); 7192 _threshold = (HeapWord*)round_to( 7193 (intptr_t)_finger, CardTableModRefBS::card_size); 7194 MemRegion mr(old_threshold, _threshold); 7195 assert(!mr.is_empty(), "Control point invariant"); 7196 assert(_span.contains(mr), "Should clear within span"); // _whole_span ?? 7197 // XXX When _finger crosses from old gen into perm gen 7198 // we may be doing unnecessary cleaning; do better in the 7199 // future by detecting that condition and clearing fewer 7200 // MUT/CT entries. 7201 _mut->clear_range(mr); 7202 } 7203 7204 // Note: the local finger doesn't advance while we drain 7205 // the stack below, but the global finger sure can and will. 7206 HeapWord** gfa = _task->global_finger_addr(); 7207 Par_PushOrMarkClosure pushOrMarkClosure(_collector, 7208 _span, _bit_map, 7209 _work_queue, 7210 _overflow_stack, 7211 _revisit_stack, 7212 _finger, 7213 gfa, this); 7214 bool res = _work_queue->push(obj); // overflow could occur here 7215 assert(res, "Will hold once we use workqueues"); 7216 while (true) { 7217 oop new_oop; 7218 if (!_work_queue->pop_local(new_oop)) { 7219 // We emptied our work_queue; check if there's stuff that can 7220 // be gotten from the overflow stack. 7221 if (CMSConcMarkingTask::get_work_from_overflow_stack( 7222 _overflow_stack, _work_queue)) { 7223 do_yield_check(); 7224 continue; 7225 } else { // done 7226 break; 7227 } 7228 } 7229 // Skip verifying header mark word below because we are 7230 // running concurrent with mutators. 7231 assert(new_oop->is_oop(true), "Oops! expected to pop an oop"); 7232 // now scan this oop's oops 7233 new_oop->oop_iterate(&pushOrMarkClosure); 7234 do_yield_check(); 7235 } 7236 assert(_work_queue->size() == 0, "tautology, emphasizing post-condition"); 7237 } 7238 7239 // Yield in response to a request from VM Thread or 7240 // from mutators. 7241 void Par_MarkFromRootsClosure::do_yield_work() { 7242 assert(_task != NULL, "sanity"); 7243 _task->yield(); 7244 } 7245 7246 // A variant of the above used for verifying CMS marking work. 7247 MarkFromRootsVerifyClosure::MarkFromRootsVerifyClosure(CMSCollector* collector, 7248 MemRegion span, 7249 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 7250 CMSMarkStack* mark_stack): 7251 _collector(collector), 7252 _span(span), 7253 _verification_bm(verification_bm), 7254 _cms_bm(cms_bm), 7255 _mark_stack(mark_stack), 7256 _pam_verify_closure(collector, span, verification_bm, cms_bm, 7257 mark_stack) 7258 { 7259 assert(_mark_stack->isEmpty(), "stack should be empty"); 7260 _finger = _verification_bm->startWord(); 7261 assert(_collector->_restart_addr == NULL, "Sanity check"); 7262 assert(_span.contains(_finger), "Out of bounds _finger?"); 7263 } 7264 7265 void MarkFromRootsVerifyClosure::reset(HeapWord* addr) { 7266 assert(_mark_stack->isEmpty(), "would cause duplicates on stack"); 7267 assert(_span.contains(addr), "Out of bounds _finger?"); 7268 _finger = addr; 7269 } 7270 7271 // Should revisit to see if this should be restructured for 7272 // greater efficiency. 7273 bool MarkFromRootsVerifyClosure::do_bit(size_t offset) { 7274 // convert offset into a HeapWord* 7275 HeapWord* addr = _verification_bm->startWord() + offset; 7276 assert(_verification_bm->endWord() && addr < _verification_bm->endWord(), 7277 "address out of range"); 7278 assert(_verification_bm->isMarked(addr), "tautology"); 7279 assert(_cms_bm->isMarked(addr), "tautology"); 7280 7281 assert(_mark_stack->isEmpty(), 7282 "should drain stack to limit stack usage"); 7283 // convert addr to an oop preparatory to scanning 7284 oop obj = oop(addr); 7285 assert(obj->is_oop(), "should be an oop"); 7286 assert(_finger <= addr, "_finger runneth ahead"); 7287 // advance the finger to right end of this object 7288 _finger = addr + obj->size(); 7289 assert(_finger > addr, "we just incremented it above"); 7290 // Note: the finger doesn't advance while we drain 7291 // the stack below. 7292 bool res = _mark_stack->push(obj); 7293 assert(res, "Empty non-zero size stack should have space for single push"); 7294 while (!_mark_stack->isEmpty()) { 7295 oop new_oop = _mark_stack->pop(); 7296 assert(new_oop->is_oop(), "Oops! expected to pop an oop"); 7297 // now scan this oop's oops 7298 new_oop->oop_iterate(&_pam_verify_closure); 7299 } 7300 assert(_mark_stack->isEmpty(), "tautology, emphasizing post-condition"); 7301 return true; 7302 } 7303 7304 PushAndMarkVerifyClosure::PushAndMarkVerifyClosure( 7305 CMSCollector* collector, MemRegion span, 7306 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 7307 CMSMarkStack* mark_stack): 7308 OopClosure(collector->ref_processor()), 7309 _collector(collector), 7310 _span(span), 7311 _verification_bm(verification_bm), 7312 _cms_bm(cms_bm), 7313 _mark_stack(mark_stack) 7314 { } 7315 7316 void PushAndMarkVerifyClosure::do_oop(oop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 7317 void PushAndMarkVerifyClosure::do_oop(narrowOop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 7318 7319 // Upon stack overflow, we discard (part of) the stack, 7320 // remembering the least address amongst those discarded 7321 // in CMSCollector's _restart_address. 7322 void PushAndMarkVerifyClosure::handle_stack_overflow(HeapWord* lost) { 7323 // Remember the least grey address discarded 7324 HeapWord* ra = (HeapWord*)_mark_stack->least_value(lost); 7325 _collector->lower_restart_addr(ra); 7326 _mark_stack->reset(); // discard stack contents 7327 _mark_stack->expand(); // expand the stack if possible 7328 } 7329 7330 void PushAndMarkVerifyClosure::do_oop(oop obj) { 7331 assert(obj->is_oop_or_null(), "expected an oop or NULL"); 7332 HeapWord* addr = (HeapWord*)obj; 7333 if (_span.contains(addr) && !_verification_bm->isMarked(addr)) { 7334 // Oop lies in _span and isn't yet grey or black 7335 _verification_bm->mark(addr); // now grey 7336 if (!_cms_bm->isMarked(addr)) { 7337 oop(addr)->print(); 7338 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)", 7339 addr); 7340 fatal("... aborting"); 7341 } 7342 7343 if (!_mark_stack->push(obj)) { // stack overflow 7344 if (PrintCMSStatistics != 0) { 7345 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 7346 SIZE_FORMAT, _mark_stack->capacity()); 7347 } 7348 assert(_mark_stack->isFull(), "Else push should have succeeded"); 7349 handle_stack_overflow(addr); 7350 } 7351 // anything including and to the right of _finger 7352 // will be scanned as we iterate over the remainder of the 7353 // bit map 7354 } 7355 } 7356 7357 PushOrMarkClosure::PushOrMarkClosure(CMSCollector* collector, 7358 MemRegion span, 7359 CMSBitMap* bitMap, CMSMarkStack* markStack, 7360 CMSMarkStack* revisitStack, 7361 HeapWord* finger, MarkFromRootsClosure* parent) : 7362 KlassRememberingOopClosure(collector, collector->ref_processor(), revisitStack), 7363 _span(span), 7364 _bitMap(bitMap), 7365 _markStack(markStack), 7366 _finger(finger), 7367 _parent(parent) 7368 { } 7369 7370 Par_PushOrMarkClosure::Par_PushOrMarkClosure(CMSCollector* collector, 7371 MemRegion span, 7372 CMSBitMap* bit_map, 7373 OopTaskQueue* work_queue, 7374 CMSMarkStack* overflow_stack, 7375 CMSMarkStack* revisit_stack, 7376 HeapWord* finger, 7377 HeapWord** global_finger_addr, 7378 Par_MarkFromRootsClosure* parent) : 7379 Par_KlassRememberingOopClosure(collector, 7380 collector->ref_processor(), 7381 revisit_stack), 7382 _whole_span(collector->_span), 7383 _span(span), 7384 _bit_map(bit_map), 7385 _work_queue(work_queue), 7386 _overflow_stack(overflow_stack), 7387 _finger(finger), 7388 _global_finger_addr(global_finger_addr), 7389 _parent(parent) 7390 { } 7391 7392 // Assumes thread-safe access by callers, who are 7393 // responsible for mutual exclusion. 7394 void CMSCollector::lower_restart_addr(HeapWord* low) { 7395 assert(_span.contains(low), "Out of bounds addr"); 7396 if (_restart_addr == NULL) { 7397 _restart_addr = low; 7398 } else { 7399 _restart_addr = MIN2(_restart_addr, low); 7400 } 7401 } 7402 7403 // Upon stack overflow, we discard (part of) the stack, 7404 // remembering the least address amongst those discarded 7405 // in CMSCollector's _restart_address. 7406 void PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 7407 // Remember the least grey address discarded 7408 HeapWord* ra = (HeapWord*)_markStack->least_value(lost); 7409 _collector->lower_restart_addr(ra); 7410 _markStack->reset(); // discard stack contents 7411 _markStack->expand(); // expand the stack if possible 7412 } 7413 7414 // Upon stack overflow, we discard (part of) the stack, 7415 // remembering the least address amongst those discarded 7416 // in CMSCollector's _restart_address. 7417 void Par_PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 7418 // We need to do this under a mutex to prevent other 7419 // workers from interfering with the work done below. 7420 MutexLockerEx ml(_overflow_stack->par_lock(), 7421 Mutex::_no_safepoint_check_flag); 7422 // Remember the least grey address discarded 7423 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 7424 _collector->lower_restart_addr(ra); 7425 _overflow_stack->reset(); // discard stack contents 7426 _overflow_stack->expand(); // expand the stack if possible 7427 } 7428 7429 void PushOrMarkClosure::do_oop(oop obj) { 7430 // Ignore mark word because we are running concurrent with mutators. 7431 assert(obj->is_oop_or_null(true), "expected an oop or NULL"); 7432 HeapWord* addr = (HeapWord*)obj; 7433 if (_span.contains(addr) && !_bitMap->isMarked(addr)) { 7434 // Oop lies in _span and isn't yet grey or black 7435 _bitMap->mark(addr); // now grey 7436 if (addr < _finger) { 7437 // the bit map iteration has already either passed, or 7438 // sampled, this bit in the bit map; we'll need to 7439 // use the marking stack to scan this oop's oops. 7440 bool simulate_overflow = false; 7441 NOT_PRODUCT( 7442 if (CMSMarkStackOverflowALot && 7443 _collector->simulate_overflow()) { 7444 // simulate a stack overflow 7445 simulate_overflow = true; 7446 } 7447 ) 7448 if (simulate_overflow || !_markStack->push(obj)) { // stack overflow 7449 if (PrintCMSStatistics != 0) { 7450 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 7451 SIZE_FORMAT, _markStack->capacity()); 7452 } 7453 assert(simulate_overflow || _markStack->isFull(), "Else push should have succeeded"); 7454 handle_stack_overflow(addr); 7455 } 7456 } 7457 // anything including and to the right of _finger 7458 // will be scanned as we iterate over the remainder of the 7459 // bit map 7460 do_yield_check(); 7461 } 7462 } 7463 7464 void PushOrMarkClosure::do_oop(oop* p) { PushOrMarkClosure::do_oop_work(p); } 7465 void PushOrMarkClosure::do_oop(narrowOop* p) { PushOrMarkClosure::do_oop_work(p); } 7466 7467 void Par_PushOrMarkClosure::do_oop(oop obj) { 7468 // Ignore mark word because we are running concurrent with mutators. 7469 assert(obj->is_oop_or_null(true), "expected an oop or NULL"); 7470 HeapWord* addr = (HeapWord*)obj; 7471 if (_whole_span.contains(addr) && !_bit_map->isMarked(addr)) { 7472 // Oop lies in _span and isn't yet grey or black 7473 // We read the global_finger (volatile read) strictly after marking oop 7474 bool res = _bit_map->par_mark(addr); // now grey 7475 volatile HeapWord** gfa = (volatile HeapWord**)_global_finger_addr; 7476 // Should we push this marked oop on our stack? 7477 // -- if someone else marked it, nothing to do 7478 // -- if target oop is above global finger nothing to do 7479 // -- if target oop is in chunk and above local finger 7480 // then nothing to do 7481 // -- else push on work queue 7482 if ( !res // someone else marked it, they will deal with it 7483 || (addr >= *gfa) // will be scanned in a later task 7484 || (_span.contains(addr) && addr >= _finger)) { // later in this chunk 7485 return; 7486 } 7487 // the bit map iteration has already either passed, or 7488 // sampled, this bit in the bit map; we'll need to 7489 // use the marking stack to scan this oop's oops. 7490 bool simulate_overflow = false; 7491 NOT_PRODUCT( 7492 if (CMSMarkStackOverflowALot && 7493 _collector->simulate_overflow()) { 7494 // simulate a stack overflow 7495 simulate_overflow = true; 7496 } 7497 ) 7498 if (simulate_overflow || 7499 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) { 7500 // stack overflow 7501 if (PrintCMSStatistics != 0) { 7502 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 7503 SIZE_FORMAT, _overflow_stack->capacity()); 7504 } 7505 // We cannot assert that the overflow stack is full because 7506 // it may have been emptied since. 7507 assert(simulate_overflow || 7508 _work_queue->size() == _work_queue->max_elems(), 7509 "Else push should have succeeded"); 7510 handle_stack_overflow(addr); 7511 } 7512 do_yield_check(); 7513 } 7514 } 7515 7516 void Par_PushOrMarkClosure::do_oop(oop* p) { Par_PushOrMarkClosure::do_oop_work(p); } 7517 void Par_PushOrMarkClosure::do_oop(narrowOop* p) { Par_PushOrMarkClosure::do_oop_work(p); } 7518 7519 KlassRememberingOopClosure::KlassRememberingOopClosure(CMSCollector* collector, 7520 ReferenceProcessor* rp, 7521 CMSMarkStack* revisit_stack) : 7522 OopClosure(rp), 7523 _collector(collector), 7524 _revisit_stack(revisit_stack), 7525 _should_remember_klasses(collector->should_unload_classes()) {} 7526 7527 PushAndMarkClosure::PushAndMarkClosure(CMSCollector* collector, 7528 MemRegion span, 7529 ReferenceProcessor* rp, 7530 CMSBitMap* bit_map, 7531 CMSBitMap* mod_union_table, 7532 CMSMarkStack* mark_stack, 7533 CMSMarkStack* revisit_stack, 7534 bool concurrent_precleaning): 7535 KlassRememberingOopClosure(collector, rp, revisit_stack), 7536 _span(span), 7537 _bit_map(bit_map), 7538 _mod_union_table(mod_union_table), 7539 _mark_stack(mark_stack), 7540 _concurrent_precleaning(concurrent_precleaning) 7541 { 7542 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 7543 } 7544 7545 // Grey object rescan during pre-cleaning and second checkpoint phases -- 7546 // the non-parallel version (the parallel version appears further below.) 7547 void PushAndMarkClosure::do_oop(oop obj) { 7548 // Ignore mark word verification. If during concurrent precleaning, 7549 // the object monitor may be locked. If during the checkpoint 7550 // phases, the object may already have been reached by a different 7551 // path and may be at the end of the global overflow list (so 7552 // the mark word may be NULL). 7553 assert(obj->is_oop_or_null(true /* ignore mark word */), 7554 "expected an oop or NULL"); 7555 HeapWord* addr = (HeapWord*)obj; 7556 // Check if oop points into the CMS generation 7557 // and is not marked 7558 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 7559 // a white object ... 7560 _bit_map->mark(addr); // ... now grey 7561 // push on the marking stack (grey set) 7562 bool simulate_overflow = false; 7563 NOT_PRODUCT( 7564 if (CMSMarkStackOverflowALot && 7565 _collector->simulate_overflow()) { 7566 // simulate a stack overflow 7567 simulate_overflow = true; 7568 } 7569 ) 7570 if (simulate_overflow || !_mark_stack->push(obj)) { 7571 if (_concurrent_precleaning) { 7572 // During precleaning we can just dirty the appropriate card(s) 7573 // in the mod union table, thus ensuring that the object remains 7574 // in the grey set and continue. In the case of object arrays 7575 // we need to dirty all of the cards that the object spans, 7576 // since the rescan of object arrays will be limited to the 7577 // dirty cards. 7578 // Note that no one can be intefering with us in this action 7579 // of dirtying the mod union table, so no locking or atomics 7580 // are required. 7581 if (obj->is_objArray()) { 7582 size_t sz = obj->size(); 7583 HeapWord* end_card_addr = (HeapWord*)round_to( 7584 (intptr_t)(addr+sz), CardTableModRefBS::card_size); 7585 MemRegion redirty_range = MemRegion(addr, end_card_addr); 7586 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 7587 _mod_union_table->mark_range(redirty_range); 7588 } else { 7589 _mod_union_table->mark(addr); 7590 } 7591 _collector->_ser_pmc_preclean_ovflw++; 7592 } else { 7593 // During the remark phase, we need to remember this oop 7594 // in the overflow list. 7595 _collector->push_on_overflow_list(obj); 7596 _collector->_ser_pmc_remark_ovflw++; 7597 } 7598 } 7599 } 7600 } 7601 7602 Par_PushAndMarkClosure::Par_PushAndMarkClosure(CMSCollector* collector, 7603 MemRegion span, 7604 ReferenceProcessor* rp, 7605 CMSBitMap* bit_map, 7606 OopTaskQueue* work_queue, 7607 CMSMarkStack* revisit_stack): 7608 Par_KlassRememberingOopClosure(collector, rp, revisit_stack), 7609 _span(span), 7610 _bit_map(bit_map), 7611 _work_queue(work_queue) 7612 { 7613 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 7614 } 7615 7616 void PushAndMarkClosure::do_oop(oop* p) { PushAndMarkClosure::do_oop_work(p); } 7617 void PushAndMarkClosure::do_oop(narrowOop* p) { PushAndMarkClosure::do_oop_work(p); } 7618 7619 // Grey object rescan during second checkpoint phase -- 7620 // the parallel version. 7621 void Par_PushAndMarkClosure::do_oop(oop obj) { 7622 // In the assert below, we ignore the mark word because 7623 // this oop may point to an already visited object that is 7624 // on the overflow stack (in which case the mark word has 7625 // been hijacked for chaining into the overflow stack -- 7626 // if this is the last object in the overflow stack then 7627 // its mark word will be NULL). Because this object may 7628 // have been subsequently popped off the global overflow 7629 // stack, and the mark word possibly restored to the prototypical 7630 // value, by the time we get to examined this failing assert in 7631 // the debugger, is_oop_or_null(false) may subsequently start 7632 // to hold. 7633 assert(obj->is_oop_or_null(true), 7634 "expected an oop or NULL"); 7635 HeapWord* addr = (HeapWord*)obj; 7636 // Check if oop points into the CMS generation 7637 // and is not marked 7638 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 7639 // a white object ... 7640 // If we manage to "claim" the object, by being the 7641 // first thread to mark it, then we push it on our 7642 // marking stack 7643 if (_bit_map->par_mark(addr)) { // ... now grey 7644 // push on work queue (grey set) 7645 bool simulate_overflow = false; 7646 NOT_PRODUCT( 7647 if (CMSMarkStackOverflowALot && 7648 _collector->par_simulate_overflow()) { 7649 // simulate a stack overflow 7650 simulate_overflow = true; 7651 } 7652 ) 7653 if (simulate_overflow || !_work_queue->push(obj)) { 7654 _collector->par_push_on_overflow_list(obj); 7655 _collector->_par_pmc_remark_ovflw++; // imprecise OK: no need to CAS 7656 } 7657 } // Else, some other thread got there first 7658 } 7659 } 7660 7661 void Par_PushAndMarkClosure::do_oop(oop* p) { Par_PushAndMarkClosure::do_oop_work(p); } 7662 void Par_PushAndMarkClosure::do_oop(narrowOop* p) { Par_PushAndMarkClosure::do_oop_work(p); } 7663 7664 void PushAndMarkClosure::remember_mdo(DataLayout* v) { 7665 // TBD 7666 } 7667 7668 void Par_PushAndMarkClosure::remember_mdo(DataLayout* v) { 7669 // TBD 7670 } 7671 7672 void CMSPrecleanRefsYieldClosure::do_yield_work() { 7673 DEBUG_ONLY(RememberKlassesChecker mux(false);) 7674 Mutex* bml = _collector->bitMapLock(); 7675 assert_lock_strong(bml); 7676 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 7677 "CMS thread should hold CMS token"); 7678 7679 bml->unlock(); 7680 ConcurrentMarkSweepThread::desynchronize(true); 7681 7682 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7683 7684 _collector->stopTimer(); 7685 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 7686 if (PrintCMSStatistics != 0) { 7687 _collector->incrementYields(); 7688 } 7689 _collector->icms_wait(); 7690 7691 // See the comment in coordinator_yield() 7692 for (unsigned i = 0; i < CMSYieldSleepCount && 7693 ConcurrentMarkSweepThread::should_yield() && 7694 !CMSCollector::foregroundGCIsActive(); ++i) { 7695 os::sleep(Thread::current(), 1, false); 7696 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7697 } 7698 7699 ConcurrentMarkSweepThread::synchronize(true); 7700 bml->lock(); 7701 7702 _collector->startTimer(); 7703 } 7704 7705 bool CMSPrecleanRefsYieldClosure::should_return() { 7706 if (ConcurrentMarkSweepThread::should_yield()) { 7707 do_yield_work(); 7708 } 7709 return _collector->foregroundGCIsActive(); 7710 } 7711 7712 void MarkFromDirtyCardsClosure::do_MemRegion(MemRegion mr) { 7713 assert(((size_t)mr.start())%CardTableModRefBS::card_size_in_words == 0, 7714 "mr should be aligned to start at a card boundary"); 7715 // We'd like to assert: 7716 // assert(mr.word_size()%CardTableModRefBS::card_size_in_words == 0, 7717 // "mr should be a range of cards"); 7718 // However, that would be too strong in one case -- the last 7719 // partition ends at _unallocated_block which, in general, can be 7720 // an arbitrary boundary, not necessarily card aligned. 7721 if (PrintCMSStatistics != 0) { 7722 _num_dirty_cards += 7723 mr.word_size()/CardTableModRefBS::card_size_in_words; 7724 } 7725 _space->object_iterate_mem(mr, &_scan_cl); 7726 } 7727 7728 SweepClosure::SweepClosure(CMSCollector* collector, 7729 ConcurrentMarkSweepGeneration* g, 7730 CMSBitMap* bitMap, bool should_yield) : 7731 _collector(collector), 7732 _g(g), 7733 _sp(g->cmsSpace()), 7734 _limit(_sp->sweep_limit()), 7735 _freelistLock(_sp->freelistLock()), 7736 _bitMap(bitMap), 7737 _yield(should_yield), 7738 _inFreeRange(false), // No free range at beginning of sweep 7739 _freeRangeInFreeLists(false), // No free range at beginning of sweep 7740 _lastFreeRangeCoalesced(false), 7741 _freeFinger(g->used_region().start()) 7742 { 7743 NOT_PRODUCT( 7744 _numObjectsFreed = 0; 7745 _numWordsFreed = 0; 7746 _numObjectsLive = 0; 7747 _numWordsLive = 0; 7748 _numObjectsAlreadyFree = 0; 7749 _numWordsAlreadyFree = 0; 7750 _last_fc = NULL; 7751 7752 _sp->initializeIndexedFreeListArrayReturnedBytes(); 7753 _sp->dictionary()->initializeDictReturnedBytes(); 7754 ) 7755 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7756 "sweep _limit out of bounds"); 7757 if (CMSTraceSweeper) { 7758 gclog_or_tty->print("\n====================\nStarting new sweep\n"); 7759 } 7760 } 7761 7762 // We need this destructor to reclaim any space at the end 7763 // of the space, which do_blk below may not have added back to 7764 // the free lists. [basically dealing with the "fringe effect"] 7765 SweepClosure::~SweepClosure() { 7766 assert_lock_strong(_freelistLock); 7767 // this should be treated as the end of a free run if any 7768 // The current free range should be returned to the free lists 7769 // as one coalesced chunk. 7770 if (inFreeRange()) { 7771 flushCurFreeChunk(freeFinger(), 7772 pointer_delta(_limit, freeFinger())); 7773 assert(freeFinger() < _limit, "the finger pointeth off base"); 7774 if (CMSTraceSweeper) { 7775 gclog_or_tty->print("destructor:"); 7776 gclog_or_tty->print("Sweep:put_free_blk 0x%x ("SIZE_FORMAT") " 7777 "[coalesced:"SIZE_FORMAT"]\n", 7778 freeFinger(), pointer_delta(_limit, freeFinger()), 7779 lastFreeRangeCoalesced()); 7780 } 7781 } 7782 NOT_PRODUCT( 7783 if (Verbose && PrintGC) { 7784 gclog_or_tty->print("Collected "SIZE_FORMAT" objects, " 7785 SIZE_FORMAT " bytes", 7786 _numObjectsFreed, _numWordsFreed*sizeof(HeapWord)); 7787 gclog_or_tty->print_cr("\nLive "SIZE_FORMAT" objects, " 7788 SIZE_FORMAT" bytes " 7789 "Already free "SIZE_FORMAT" objects, "SIZE_FORMAT" bytes", 7790 _numObjectsLive, _numWordsLive*sizeof(HeapWord), 7791 _numObjectsAlreadyFree, _numWordsAlreadyFree*sizeof(HeapWord)); 7792 size_t totalBytes = (_numWordsFreed + _numWordsLive + _numWordsAlreadyFree) * 7793 sizeof(HeapWord); 7794 gclog_or_tty->print_cr("Total sweep: "SIZE_FORMAT" bytes", totalBytes); 7795 7796 if (PrintCMSStatistics && CMSVerifyReturnedBytes) { 7797 size_t indexListReturnedBytes = _sp->sumIndexedFreeListArrayReturnedBytes(); 7798 size_t dictReturnedBytes = _sp->dictionary()->sumDictReturnedBytes(); 7799 size_t returnedBytes = indexListReturnedBytes + dictReturnedBytes; 7800 gclog_or_tty->print("Returned "SIZE_FORMAT" bytes", returnedBytes); 7801 gclog_or_tty->print(" Indexed List Returned "SIZE_FORMAT" bytes", 7802 indexListReturnedBytes); 7803 gclog_or_tty->print_cr(" Dictionary Returned "SIZE_FORMAT" bytes", 7804 dictReturnedBytes); 7805 } 7806 } 7807 ) 7808 // Now, in debug mode, just null out the sweep_limit 7809 NOT_PRODUCT(_sp->clear_sweep_limit();) 7810 if (CMSTraceSweeper) { 7811 gclog_or_tty->print("end of sweep\n================\n"); 7812 } 7813 } 7814 7815 void SweepClosure::initialize_free_range(HeapWord* freeFinger, 7816 bool freeRangeInFreeLists) { 7817 if (CMSTraceSweeper) { 7818 gclog_or_tty->print("---- Start free range 0x%x with free block [%d] (%d)\n", 7819 freeFinger, _sp->block_size(freeFinger), 7820 freeRangeInFreeLists); 7821 } 7822 assert(!inFreeRange(), "Trampling existing free range"); 7823 set_inFreeRange(true); 7824 set_lastFreeRangeCoalesced(false); 7825 7826 set_freeFinger(freeFinger); 7827 set_freeRangeInFreeLists(freeRangeInFreeLists); 7828 if (CMSTestInFreeList) { 7829 if (freeRangeInFreeLists) { 7830 FreeChunk* fc = (FreeChunk*) freeFinger; 7831 assert(fc->isFree(), "A chunk on the free list should be free."); 7832 assert(fc->size() > 0, "Free range should have a size"); 7833 assert(_sp->verifyChunkInFreeLists(fc), "Chunk is not in free lists"); 7834 } 7835 } 7836 } 7837 7838 // Note that the sweeper runs concurrently with mutators. Thus, 7839 // it is possible for direct allocation in this generation to happen 7840 // in the middle of the sweep. Note that the sweeper also coalesces 7841 // contiguous free blocks. Thus, unless the sweeper and the allocator 7842 // synchronize appropriately freshly allocated blocks may get swept up. 7843 // This is accomplished by the sweeper locking the free lists while 7844 // it is sweeping. Thus blocks that are determined to be free are 7845 // indeed free. There is however one additional complication: 7846 // blocks that have been allocated since the final checkpoint and 7847 // mark, will not have been marked and so would be treated as 7848 // unreachable and swept up. To prevent this, the allocator marks 7849 // the bit map when allocating during the sweep phase. This leads, 7850 // however, to a further complication -- objects may have been allocated 7851 // but not yet initialized -- in the sense that the header isn't yet 7852 // installed. The sweeper can not then determine the size of the block 7853 // in order to skip over it. To deal with this case, we use a technique 7854 // (due to Printezis) to encode such uninitialized block sizes in the 7855 // bit map. Since the bit map uses a bit per every HeapWord, but the 7856 // CMS generation has a minimum object size of 3 HeapWords, it follows 7857 // that "normal marks" won't be adjacent in the bit map (there will 7858 // always be at least two 0 bits between successive 1 bits). We make use 7859 // of these "unused" bits to represent uninitialized blocks -- the bit 7860 // corresponding to the start of the uninitialized object and the next 7861 // bit are both set. Finally, a 1 bit marks the end of the object that 7862 // started with the two consecutive 1 bits to indicate its potentially 7863 // uninitialized state. 7864 7865 size_t SweepClosure::do_blk_careful(HeapWord* addr) { 7866 FreeChunk* fc = (FreeChunk*)addr; 7867 size_t res; 7868 7869 // check if we are done sweepinrg 7870 if (addr == _limit) { // we have swept up to the limit, do nothing more 7871 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7872 "sweep _limit out of bounds"); 7873 // help the closure application finish 7874 return pointer_delta(_sp->end(), _limit); 7875 } 7876 assert(addr <= _limit, "sweep invariant"); 7877 7878 // check if we should yield 7879 do_yield_check(addr); 7880 if (fc->isFree()) { 7881 // Chunk that is already free 7882 res = fc->size(); 7883 doAlreadyFreeChunk(fc); 7884 debug_only(_sp->verifyFreeLists()); 7885 assert(res == fc->size(), "Don't expect the size to change"); 7886 NOT_PRODUCT( 7887 _numObjectsAlreadyFree++; 7888 _numWordsAlreadyFree += res; 7889 ) 7890 NOT_PRODUCT(_last_fc = fc;) 7891 } else if (!_bitMap->isMarked(addr)) { 7892 // Chunk is fresh garbage 7893 res = doGarbageChunk(fc); 7894 debug_only(_sp->verifyFreeLists()); 7895 NOT_PRODUCT( 7896 _numObjectsFreed++; 7897 _numWordsFreed += res; 7898 ) 7899 } else { 7900 // Chunk that is alive. 7901 res = doLiveChunk(fc); 7902 debug_only(_sp->verifyFreeLists()); 7903 NOT_PRODUCT( 7904 _numObjectsLive++; 7905 _numWordsLive += res; 7906 ) 7907 } 7908 return res; 7909 } 7910 7911 // For the smart allocation, record following 7912 // split deaths - a free chunk is removed from its free list because 7913 // it is being split into two or more chunks. 7914 // split birth - a free chunk is being added to its free list because 7915 // a larger free chunk has been split and resulted in this free chunk. 7916 // coal death - a free chunk is being removed from its free list because 7917 // it is being coalesced into a large free chunk. 7918 // coal birth - a free chunk is being added to its free list because 7919 // it was created when two or more free chunks where coalesced into 7920 // this free chunk. 7921 // 7922 // These statistics are used to determine the desired number of free 7923 // chunks of a given size. The desired number is chosen to be relative 7924 // to the end of a CMS sweep. The desired number at the end of a sweep 7925 // is the 7926 // count-at-end-of-previous-sweep (an amount that was enough) 7927 // - count-at-beginning-of-current-sweep (the excess) 7928 // + split-births (gains in this size during interval) 7929 // - split-deaths (demands on this size during interval) 7930 // where the interval is from the end of one sweep to the end of the 7931 // next. 7932 // 7933 // When sweeping the sweeper maintains an accumulated chunk which is 7934 // the chunk that is made up of chunks that have been coalesced. That 7935 // will be termed the left-hand chunk. A new chunk of garbage that 7936 // is being considered for coalescing will be referred to as the 7937 // right-hand chunk. 7938 // 7939 // When making a decision on whether to coalesce a right-hand chunk with 7940 // the current left-hand chunk, the current count vs. the desired count 7941 // of the left-hand chunk is considered. Also if the right-hand chunk 7942 // is near the large chunk at the end of the heap (see 7943 // ConcurrentMarkSweepGeneration::isNearLargestChunk()), then the 7944 // left-hand chunk is coalesced. 7945 // 7946 // When making a decision about whether to split a chunk, the desired count 7947 // vs. the current count of the candidate to be split is also considered. 7948 // If the candidate is underpopulated (currently fewer chunks than desired) 7949 // a chunk of an overpopulated (currently more chunks than desired) size may 7950 // be chosen. The "hint" associated with a free list, if non-null, points 7951 // to a free list which may be overpopulated. 7952 // 7953 7954 void SweepClosure::doAlreadyFreeChunk(FreeChunk* fc) { 7955 size_t size = fc->size(); 7956 // Chunks that cannot be coalesced are not in the 7957 // free lists. 7958 if (CMSTestInFreeList && !fc->cantCoalesce()) { 7959 assert(_sp->verifyChunkInFreeLists(fc), 7960 "free chunk should be in free lists"); 7961 } 7962 // a chunk that is already free, should not have been 7963 // marked in the bit map 7964 HeapWord* addr = (HeapWord*) fc; 7965 assert(!_bitMap->isMarked(addr), "free chunk should be unmarked"); 7966 // Verify that the bit map has no bits marked between 7967 // addr and purported end of this block. 7968 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 7969 7970 // Some chunks cannot be coalesced in under any circumstances. 7971 // See the definition of cantCoalesce(). 7972 if (!fc->cantCoalesce()) { 7973 // This chunk can potentially be coalesced. 7974 if (_sp->adaptive_freelists()) { 7975 // All the work is done in 7976 doPostIsFreeOrGarbageChunk(fc, size); 7977 } else { // Not adaptive free lists 7978 // this is a free chunk that can potentially be coalesced by the sweeper; 7979 if (!inFreeRange()) { 7980 // if the next chunk is a free block that can't be coalesced 7981 // it doesn't make sense to remove this chunk from the free lists 7982 FreeChunk* nextChunk = (FreeChunk*)(addr + size); 7983 assert((HeapWord*)nextChunk <= _limit, "sweep invariant"); 7984 if ((HeapWord*)nextChunk < _limit && // there's a next chunk... 7985 nextChunk->isFree() && // which is free... 7986 nextChunk->cantCoalesce()) { // ... but cant be coalesced 7987 // nothing to do 7988 } else { 7989 // Potentially the start of a new free range: 7990 // Don't eagerly remove it from the free lists. 7991 // No need to remove it if it will just be put 7992 // back again. (Also from a pragmatic point of view 7993 // if it is a free block in a region that is beyond 7994 // any allocated blocks, an assertion will fail) 7995 // Remember the start of a free run. 7996 initialize_free_range(addr, true); 7997 // end - can coalesce with next chunk 7998 } 7999 } else { 8000 // the midst of a free range, we are coalescing 8001 debug_only(record_free_block_coalesced(fc);) 8002 if (CMSTraceSweeper) { 8003 gclog_or_tty->print(" -- pick up free block 0x%x (%d)\n", fc, size); 8004 } 8005 // remove it from the free lists 8006 _sp->removeFreeChunkFromFreeLists(fc); 8007 set_lastFreeRangeCoalesced(true); 8008 // If the chunk is being coalesced and the current free range is 8009 // in the free lists, remove the current free range so that it 8010 // will be returned to the free lists in its entirety - all 8011 // the coalesced pieces included. 8012 if (freeRangeInFreeLists()) { 8013 FreeChunk* ffc = (FreeChunk*) freeFinger(); 8014 assert(ffc->size() == pointer_delta(addr, freeFinger()), 8015 "Size of free range is inconsistent with chunk size."); 8016 if (CMSTestInFreeList) { 8017 assert(_sp->verifyChunkInFreeLists(ffc), 8018 "free range is not in free lists"); 8019 } 8020 _sp->removeFreeChunkFromFreeLists(ffc); 8021 set_freeRangeInFreeLists(false); 8022 } 8023 } 8024 } 8025 } else { 8026 // Code path common to both original and adaptive free lists. 8027 8028 // cant coalesce with previous block; this should be treated 8029 // as the end of a free run if any 8030 if (inFreeRange()) { 8031 // we kicked some butt; time to pick up the garbage 8032 assert(freeFinger() < addr, "the finger pointeth off base"); 8033 flushCurFreeChunk(freeFinger(), pointer_delta(addr, freeFinger())); 8034 } 8035 // else, nothing to do, just continue 8036 } 8037 } 8038 8039 size_t SweepClosure::doGarbageChunk(FreeChunk* fc) { 8040 // This is a chunk of garbage. It is not in any free list. 8041 // Add it to a free list or let it possibly be coalesced into 8042 // a larger chunk. 8043 HeapWord* addr = (HeapWord*) fc; 8044 size_t size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 8045 8046 if (_sp->adaptive_freelists()) { 8047 // Verify that the bit map has no bits marked between 8048 // addr and purported end of just dead object. 8049 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 8050 8051 doPostIsFreeOrGarbageChunk(fc, size); 8052 } else { 8053 if (!inFreeRange()) { 8054 // start of a new free range 8055 assert(size > 0, "A free range should have a size"); 8056 initialize_free_range(addr, false); 8057 8058 } else { 8059 // this will be swept up when we hit the end of the 8060 // free range 8061 if (CMSTraceSweeper) { 8062 gclog_or_tty->print(" -- pick up garbage 0x%x (%d) \n", fc, size); 8063 } 8064 // If the chunk is being coalesced and the current free range is 8065 // in the free lists, remove the current free range so that it 8066 // will be returned to the free lists in its entirety - all 8067 // the coalesced pieces included. 8068 if (freeRangeInFreeLists()) { 8069 FreeChunk* ffc = (FreeChunk*)freeFinger(); 8070 assert(ffc->size() == pointer_delta(addr, freeFinger()), 8071 "Size of free range is inconsistent with chunk size."); 8072 if (CMSTestInFreeList) { 8073 assert(_sp->verifyChunkInFreeLists(ffc), 8074 "free range is not in free lists"); 8075 } 8076 _sp->removeFreeChunkFromFreeLists(ffc); 8077 set_freeRangeInFreeLists(false); 8078 } 8079 set_lastFreeRangeCoalesced(true); 8080 } 8081 // this will be swept up when we hit the end of the free range 8082 8083 // Verify that the bit map has no bits marked between 8084 // addr and purported end of just dead object. 8085 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 8086 } 8087 return size; 8088 } 8089 8090 size_t SweepClosure::doLiveChunk(FreeChunk* fc) { 8091 HeapWord* addr = (HeapWord*) fc; 8092 // The sweeper has just found a live object. Return any accumulated 8093 // left hand chunk to the free lists. 8094 if (inFreeRange()) { 8095 if (_sp->adaptive_freelists()) { 8096 flushCurFreeChunk(freeFinger(), 8097 pointer_delta(addr, freeFinger())); 8098 } else { // not adaptive freelists 8099 set_inFreeRange(false); 8100 // Add the free range back to the free list if it is not already 8101 // there. 8102 if (!freeRangeInFreeLists()) { 8103 assert(freeFinger() < addr, "the finger pointeth off base"); 8104 if (CMSTraceSweeper) { 8105 gclog_or_tty->print("Sweep:put_free_blk 0x%x (%d) " 8106 "[coalesced:%d]\n", 8107 freeFinger(), pointer_delta(addr, freeFinger()), 8108 lastFreeRangeCoalesced()); 8109 } 8110 _sp->addChunkAndRepairOffsetTable(freeFinger(), 8111 pointer_delta(addr, freeFinger()), lastFreeRangeCoalesced()); 8112 } 8113 } 8114 } 8115 8116 // Common code path for original and adaptive free lists. 8117 8118 // this object is live: we'd normally expect this to be 8119 // an oop, and like to assert the following: 8120 // assert(oop(addr)->is_oop(), "live block should be an oop"); 8121 // However, as we commented above, this may be an object whose 8122 // header hasn't yet been initialized. 8123 size_t size; 8124 assert(_bitMap->isMarked(addr), "Tautology for this control point"); 8125 if (_bitMap->isMarked(addr + 1)) { 8126 // Determine the size from the bit map, rather than trying to 8127 // compute it from the object header. 8128 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 8129 size = pointer_delta(nextOneAddr + 1, addr); 8130 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 8131 "alignment problem"); 8132 8133 #ifdef DEBUG 8134 if (oop(addr)->klass_or_null() != NULL && 8135 ( !_collector->should_unload_classes() 8136 || (oop(addr)->is_parsable()) && 8137 oop(addr)->is_conc_safe())) { 8138 // Ignore mark word because we are running concurrent with mutators 8139 assert(oop(addr)->is_oop(true), "live block should be an oop"); 8140 // is_conc_safe is checked before performing this assertion 8141 // because an object that is not is_conc_safe may yet have 8142 // the return from size() correct. 8143 assert(size == 8144 CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()), 8145 "P-mark and computed size do not agree"); 8146 } 8147 #endif 8148 8149 } else { 8150 // This should be an initialized object that's alive. 8151 assert(oop(addr)->klass_or_null() != NULL && 8152 (!_collector->should_unload_classes() 8153 || oop(addr)->is_parsable()), 8154 "Should be an initialized object"); 8155 // Note that there are objects used during class redefinition 8156 // (e.g., merge_cp in VM_RedefineClasses::merge_cp_and_rewrite() 8157 // which are discarded with their is_conc_safe state still 8158 // false. These object may be floating garbage so may be 8159 // seen here. If they are floating garbage their size 8160 // should be attainable from their klass. Do not that 8161 // is_conc_safe() is true for oop(addr). 8162 // Ignore mark word because we are running concurrent with mutators 8163 assert(oop(addr)->is_oop(true), "live block should be an oop"); 8164 // Verify that the bit map has no bits marked between 8165 // addr and purported end of this block. 8166 size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 8167 assert(size >= 3, "Necessary for Printezis marks to work"); 8168 assert(!_bitMap->isMarked(addr+1), "Tautology for this control point"); 8169 DEBUG_ONLY(_bitMap->verifyNoOneBitsInRange(addr+2, addr+size);) 8170 } 8171 return size; 8172 } 8173 8174 void SweepClosure::doPostIsFreeOrGarbageChunk(FreeChunk* fc, 8175 size_t chunkSize) { 8176 // doPostIsFreeOrGarbageChunk() should only be called in the smart allocation 8177 // scheme. 8178 bool fcInFreeLists = fc->isFree(); 8179 assert(_sp->adaptive_freelists(), "Should only be used in this case."); 8180 assert((HeapWord*)fc <= _limit, "sweep invariant"); 8181 if (CMSTestInFreeList && fcInFreeLists) { 8182 assert(_sp->verifyChunkInFreeLists(fc), 8183 "free chunk is not in free lists"); 8184 } 8185 8186 8187 if (CMSTraceSweeper) { 8188 gclog_or_tty->print_cr(" -- pick up another chunk at 0x%x (%d)", fc, chunkSize); 8189 } 8190 8191 HeapWord* addr = (HeapWord*) fc; 8192 8193 bool coalesce; 8194 size_t left = pointer_delta(addr, freeFinger()); 8195 size_t right = chunkSize; 8196 switch (FLSCoalescePolicy) { 8197 // numeric value forms a coalition aggressiveness metric 8198 case 0: { // never coalesce 8199 coalesce = false; 8200 break; 8201 } 8202 case 1: { // coalesce if left & right chunks on overpopulated lists 8203 coalesce = _sp->coalOverPopulated(left) && 8204 _sp->coalOverPopulated(right); 8205 break; 8206 } 8207 case 2: { // coalesce if left chunk on overpopulated list (default) 8208 coalesce = _sp->coalOverPopulated(left); 8209 break; 8210 } 8211 case 3: { // coalesce if left OR right chunk on overpopulated list 8212 coalesce = _sp->coalOverPopulated(left) || 8213 _sp->coalOverPopulated(right); 8214 break; 8215 } 8216 case 4: { // always coalesce 8217 coalesce = true; 8218 break; 8219 } 8220 default: 8221 ShouldNotReachHere(); 8222 } 8223 8224 // Should the current free range be coalesced? 8225 // If the chunk is in a free range and either we decided to coalesce above 8226 // or the chunk is near the large block at the end of the heap 8227 // (isNearLargestChunk() returns true), then coalesce this chunk. 8228 bool doCoalesce = inFreeRange() && 8229 (coalesce || _g->isNearLargestChunk((HeapWord*)fc)); 8230 if (doCoalesce) { 8231 // Coalesce the current free range on the left with the new 8232 // chunk on the right. If either is on a free list, 8233 // it must be removed from the list and stashed in the closure. 8234 if (freeRangeInFreeLists()) { 8235 FreeChunk* ffc = (FreeChunk*)freeFinger(); 8236 assert(ffc->size() == pointer_delta(addr, freeFinger()), 8237 "Size of free range is inconsistent with chunk size."); 8238 if (CMSTestInFreeList) { 8239 assert(_sp->verifyChunkInFreeLists(ffc), 8240 "Chunk is not in free lists"); 8241 } 8242 _sp->coalDeath(ffc->size()); 8243 _sp->removeFreeChunkFromFreeLists(ffc); 8244 set_freeRangeInFreeLists(false); 8245 } 8246 if (fcInFreeLists) { 8247 _sp->coalDeath(chunkSize); 8248 assert(fc->size() == chunkSize, 8249 "The chunk has the wrong size or is not in the free lists"); 8250 _sp->removeFreeChunkFromFreeLists(fc); 8251 } 8252 set_lastFreeRangeCoalesced(true); 8253 } else { // not in a free range and/or should not coalesce 8254 // Return the current free range and start a new one. 8255 if (inFreeRange()) { 8256 // In a free range but cannot coalesce with the right hand chunk. 8257 // Put the current free range into the free lists. 8258 flushCurFreeChunk(freeFinger(), 8259 pointer_delta(addr, freeFinger())); 8260 } 8261 // Set up for new free range. Pass along whether the right hand 8262 // chunk is in the free lists. 8263 initialize_free_range((HeapWord*)fc, fcInFreeLists); 8264 } 8265 } 8266 void SweepClosure::flushCurFreeChunk(HeapWord* chunk, size_t size) { 8267 assert(inFreeRange(), "Should only be called if currently in a free range."); 8268 assert(size > 0, 8269 "A zero sized chunk cannot be added to the free lists."); 8270 if (!freeRangeInFreeLists()) { 8271 if(CMSTestInFreeList) { 8272 FreeChunk* fc = (FreeChunk*) chunk; 8273 fc->setSize(size); 8274 assert(!_sp->verifyChunkInFreeLists(fc), 8275 "chunk should not be in free lists yet"); 8276 } 8277 if (CMSTraceSweeper) { 8278 gclog_or_tty->print_cr(" -- add free block 0x%x (%d) to free lists", 8279 chunk, size); 8280 } 8281 // A new free range is going to be starting. The current 8282 // free range has not been added to the free lists yet or 8283 // was removed so add it back. 8284 // If the current free range was coalesced, then the death 8285 // of the free range was recorded. Record a birth now. 8286 if (lastFreeRangeCoalesced()) { 8287 _sp->coalBirth(size); 8288 } 8289 _sp->addChunkAndRepairOffsetTable(chunk, size, 8290 lastFreeRangeCoalesced()); 8291 } 8292 set_inFreeRange(false); 8293 set_freeRangeInFreeLists(false); 8294 } 8295 8296 // We take a break if we've been at this for a while, 8297 // so as to avoid monopolizing the locks involved. 8298 void SweepClosure::do_yield_work(HeapWord* addr) { 8299 // Return current free chunk being used for coalescing (if any) 8300 // to the appropriate freelist. After yielding, the next 8301 // free block encountered will start a coalescing range of 8302 // free blocks. If the next free block is adjacent to the 8303 // chunk just flushed, they will need to wait for the next 8304 // sweep to be coalesced. 8305 if (inFreeRange()) { 8306 flushCurFreeChunk(freeFinger(), pointer_delta(addr, freeFinger())); 8307 } 8308 8309 // First give up the locks, then yield, then re-lock. 8310 // We should probably use a constructor/destructor idiom to 8311 // do this unlock/lock or modify the MutexUnlocker class to 8312 // serve our purpose. XXX 8313 assert_lock_strong(_bitMap->lock()); 8314 assert_lock_strong(_freelistLock); 8315 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 8316 "CMS thread should hold CMS token"); 8317 _bitMap->lock()->unlock(); 8318 _freelistLock->unlock(); 8319 ConcurrentMarkSweepThread::desynchronize(true); 8320 ConcurrentMarkSweepThread::acknowledge_yield_request(); 8321 _collector->stopTimer(); 8322 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 8323 if (PrintCMSStatistics != 0) { 8324 _collector->incrementYields(); 8325 } 8326 _collector->icms_wait(); 8327 8328 // See the comment in coordinator_yield() 8329 for (unsigned i = 0; i < CMSYieldSleepCount && 8330 ConcurrentMarkSweepThread::should_yield() && 8331 !CMSCollector::foregroundGCIsActive(); ++i) { 8332 os::sleep(Thread::current(), 1, false); 8333 ConcurrentMarkSweepThread::acknowledge_yield_request(); 8334 } 8335 8336 ConcurrentMarkSweepThread::synchronize(true); 8337 _freelistLock->lock(); 8338 _bitMap->lock()->lock_without_safepoint_check(); 8339 _collector->startTimer(); 8340 } 8341 8342 #ifndef PRODUCT 8343 // This is actually very useful in a product build if it can 8344 // be called from the debugger. Compile it into the product 8345 // as needed. 8346 bool debug_verifyChunkInFreeLists(FreeChunk* fc) { 8347 return debug_cms_space->verifyChunkInFreeLists(fc); 8348 } 8349 8350 void SweepClosure::record_free_block_coalesced(FreeChunk* fc) const { 8351 if (CMSTraceSweeper) { 8352 gclog_or_tty->print("Sweep:coal_free_blk 0x%x (%d)\n", fc, fc->size()); 8353 } 8354 } 8355 #endif 8356 8357 // CMSIsAliveClosure 8358 bool CMSIsAliveClosure::do_object_b(oop obj) { 8359 HeapWord* addr = (HeapWord*)obj; 8360 return addr != NULL && 8361 (!_span.contains(addr) || _bit_map->isMarked(addr)); 8362 } 8363 8364 CMSKeepAliveClosure::CMSKeepAliveClosure( CMSCollector* collector, 8365 MemRegion span, 8366 CMSBitMap* bit_map, CMSMarkStack* mark_stack, 8367 CMSMarkStack* revisit_stack, bool cpc): 8368 KlassRememberingOopClosure(collector, NULL, revisit_stack), 8369 _span(span), 8370 _bit_map(bit_map), 8371 _mark_stack(mark_stack), 8372 _concurrent_precleaning(cpc) { 8373 assert(!_span.is_empty(), "Empty span could spell trouble"); 8374 } 8375 8376 8377 // CMSKeepAliveClosure: the serial version 8378 void CMSKeepAliveClosure::do_oop(oop obj) { 8379 HeapWord* addr = (HeapWord*)obj; 8380 if (_span.contains(addr) && 8381 !_bit_map->isMarked(addr)) { 8382 _bit_map->mark(addr); 8383 bool simulate_overflow = false; 8384 NOT_PRODUCT( 8385 if (CMSMarkStackOverflowALot && 8386 _collector->simulate_overflow()) { 8387 // simulate a stack overflow 8388 simulate_overflow = true; 8389 } 8390 ) 8391 if (simulate_overflow || !_mark_stack->push(obj)) { 8392 if (_concurrent_precleaning) { 8393 // We dirty the overflown object and let the remark 8394 // phase deal with it. 8395 assert(_collector->overflow_list_is_empty(), "Error"); 8396 // In the case of object arrays, we need to dirty all of 8397 // the cards that the object spans. No locking or atomics 8398 // are needed since no one else can be mutating the mod union 8399 // table. 8400 if (obj->is_objArray()) { 8401 size_t sz = obj->size(); 8402 HeapWord* end_card_addr = 8403 (HeapWord*)round_to((intptr_t)(addr+sz), CardTableModRefBS::card_size); 8404 MemRegion redirty_range = MemRegion(addr, end_card_addr); 8405 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 8406 _collector->_modUnionTable.mark_range(redirty_range); 8407 } else { 8408 _collector->_modUnionTable.mark(addr); 8409 } 8410 _collector->_ser_kac_preclean_ovflw++; 8411 } else { 8412 _collector->push_on_overflow_list(obj); 8413 _collector->_ser_kac_ovflw++; 8414 } 8415 } 8416 } 8417 } 8418 8419 void CMSKeepAliveClosure::do_oop(oop* p) { CMSKeepAliveClosure::do_oop_work(p); } 8420 void CMSKeepAliveClosure::do_oop(narrowOop* p) { CMSKeepAliveClosure::do_oop_work(p); } 8421 8422 // CMSParKeepAliveClosure: a parallel version of the above. 8423 // The work queues are private to each closure (thread), 8424 // but (may be) available for stealing by other threads. 8425 void CMSParKeepAliveClosure::do_oop(oop obj) { 8426 HeapWord* addr = (HeapWord*)obj; 8427 if (_span.contains(addr) && 8428 !_bit_map->isMarked(addr)) { 8429 // In general, during recursive tracing, several threads 8430 // may be concurrently getting here; the first one to 8431 // "tag" it, claims it. 8432 if (_bit_map->par_mark(addr)) { 8433 bool res = _work_queue->push(obj); 8434 assert(res, "Low water mark should be much less than capacity"); 8435 // Do a recursive trim in the hope that this will keep 8436 // stack usage lower, but leave some oops for potential stealers 8437 trim_queue(_low_water_mark); 8438 } // Else, another thread got there first 8439 } 8440 } 8441 8442 void CMSParKeepAliveClosure::do_oop(oop* p) { CMSParKeepAliveClosure::do_oop_work(p); } 8443 void CMSParKeepAliveClosure::do_oop(narrowOop* p) { CMSParKeepAliveClosure::do_oop_work(p); } 8444 8445 void CMSParKeepAliveClosure::trim_queue(uint max) { 8446 while (_work_queue->size() > max) { 8447 oop new_oop; 8448 if (_work_queue->pop_local(new_oop)) { 8449 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 8450 assert(_bit_map->isMarked((HeapWord*)new_oop), 8451 "no white objects on this stack!"); 8452 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 8453 // iterate over the oops in this oop, marking and pushing 8454 // the ones in CMS heap (i.e. in _span). 8455 new_oop->oop_iterate(&_mark_and_push); 8456 } 8457 } 8458 } 8459 8460 CMSInnerParMarkAndPushClosure::CMSInnerParMarkAndPushClosure( 8461 CMSCollector* collector, 8462 MemRegion span, CMSBitMap* bit_map, 8463 CMSMarkStack* revisit_stack, 8464 OopTaskQueue* work_queue): 8465 Par_KlassRememberingOopClosure(collector, NULL, revisit_stack), 8466 _span(span), 8467 _bit_map(bit_map), 8468 _work_queue(work_queue) { } 8469 8470 void CMSInnerParMarkAndPushClosure::do_oop(oop obj) { 8471 HeapWord* addr = (HeapWord*)obj; 8472 if (_span.contains(addr) && 8473 !_bit_map->isMarked(addr)) { 8474 if (_bit_map->par_mark(addr)) { 8475 bool simulate_overflow = false; 8476 NOT_PRODUCT( 8477 if (CMSMarkStackOverflowALot && 8478 _collector->par_simulate_overflow()) { 8479 // simulate a stack overflow 8480 simulate_overflow = true; 8481 } 8482 ) 8483 if (simulate_overflow || !_work_queue->push(obj)) { 8484 _collector->par_push_on_overflow_list(obj); 8485 _collector->_par_kac_ovflw++; 8486 } 8487 } // Else another thread got there already 8488 } 8489 } 8490 8491 void CMSInnerParMarkAndPushClosure::do_oop(oop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); } 8492 void CMSInnerParMarkAndPushClosure::do_oop(narrowOop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); } 8493 8494 ////////////////////////////////////////////////////////////////// 8495 // CMSExpansionCause ///////////////////////////// 8496 ////////////////////////////////////////////////////////////////// 8497 const char* CMSExpansionCause::to_string(CMSExpansionCause::Cause cause) { 8498 switch (cause) { 8499 case _no_expansion: 8500 return "No expansion"; 8501 case _satisfy_free_ratio: 8502 return "Free ratio"; 8503 case _satisfy_promotion: 8504 return "Satisfy promotion"; 8505 case _satisfy_allocation: 8506 return "allocation"; 8507 case _allocate_par_lab: 8508 return "Par LAB"; 8509 case _allocate_par_spooling_space: 8510 return "Par Spooling Space"; 8511 case _adaptive_size_policy: 8512 return "Ergonomics"; 8513 default: 8514 return "unknown"; 8515 } 8516 } 8517 8518 void CMSDrainMarkingStackClosure::do_void() { 8519 // the max number to take from overflow list at a time 8520 const size_t num = _mark_stack->capacity()/4; 8521 assert(!_concurrent_precleaning || _collector->overflow_list_is_empty(), 8522 "Overflow list should be NULL during concurrent phases"); 8523 while (!_mark_stack->isEmpty() || 8524 // if stack is empty, check the overflow list 8525 _collector->take_from_overflow_list(num, _mark_stack)) { 8526 oop obj = _mark_stack->pop(); 8527 HeapWord* addr = (HeapWord*)obj; 8528 assert(_span.contains(addr), "Should be within span"); 8529 assert(_bit_map->isMarked(addr), "Should be marked"); 8530 assert(obj->is_oop(), "Should be an oop"); 8531 obj->oop_iterate(_keep_alive); 8532 } 8533 } 8534 8535 void CMSParDrainMarkingStackClosure::do_void() { 8536 // drain queue 8537 trim_queue(0); 8538 } 8539 8540 // Trim our work_queue so its length is below max at return 8541 void CMSParDrainMarkingStackClosure::trim_queue(uint max) { 8542 while (_work_queue->size() > max) { 8543 oop new_oop; 8544 if (_work_queue->pop_local(new_oop)) { 8545 assert(new_oop->is_oop(), "Expected an oop"); 8546 assert(_bit_map->isMarked((HeapWord*)new_oop), 8547 "no white objects on this stack!"); 8548 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 8549 // iterate over the oops in this oop, marking and pushing 8550 // the ones in CMS heap (i.e. in _span). 8551 new_oop->oop_iterate(&_mark_and_push); 8552 } 8553 } 8554 } 8555 8556 //////////////////////////////////////////////////////////////////// 8557 // Support for Marking Stack Overflow list handling and related code 8558 //////////////////////////////////////////////////////////////////// 8559 // Much of the following code is similar in shape and spirit to the 8560 // code used in ParNewGC. We should try and share that code 8561 // as much as possible in the future. 8562 8563 #ifndef PRODUCT 8564 // Debugging support for CMSStackOverflowALot 8565 8566 // It's OK to call this multi-threaded; the worst thing 8567 // that can happen is that we'll get a bunch of closely 8568 // spaced simulated oveflows, but that's OK, in fact 8569 // probably good as it would exercise the overflow code 8570 // under contention. 8571 bool CMSCollector::simulate_overflow() { 8572 if (_overflow_counter-- <= 0) { // just being defensive 8573 _overflow_counter = CMSMarkStackOverflowInterval; 8574 return true; 8575 } else { 8576 return false; 8577 } 8578 } 8579 8580 bool CMSCollector::par_simulate_overflow() { 8581 return simulate_overflow(); 8582 } 8583 #endif 8584 8585 // Single-threaded 8586 bool CMSCollector::take_from_overflow_list(size_t num, CMSMarkStack* stack) { 8587 assert(stack->isEmpty(), "Expected precondition"); 8588 assert(stack->capacity() > num, "Shouldn't bite more than can chew"); 8589 size_t i = num; 8590 oop cur = _overflow_list; 8591 const markOop proto = markOopDesc::prototype(); 8592 NOT_PRODUCT(ssize_t n = 0;) 8593 for (oop next; i > 0 && cur != NULL; cur = next, i--) { 8594 next = oop(cur->mark()); 8595 cur->set_mark(proto); // until proven otherwise 8596 assert(cur->is_oop(), "Should be an oop"); 8597 bool res = stack->push(cur); 8598 assert(res, "Bit off more than can chew?"); 8599 NOT_PRODUCT(n++;) 8600 } 8601 _overflow_list = cur; 8602 #ifndef PRODUCT 8603 assert(_num_par_pushes >= n, "Too many pops?"); 8604 _num_par_pushes -=n; 8605 #endif 8606 return !stack->isEmpty(); 8607 } 8608 8609 #define BUSY (oop(0x1aff1aff)) 8610 // (MT-safe) Get a prefix of at most "num" from the list. 8611 // The overflow list is chained through the mark word of 8612 // each object in the list. We fetch the entire list, 8613 // break off a prefix of the right size and return the 8614 // remainder. If other threads try to take objects from 8615 // the overflow list at that time, they will wait for 8616 // some time to see if data becomes available. If (and 8617 // only if) another thread places one or more object(s) 8618 // on the global list before we have returned the suffix 8619 // to the global list, we will walk down our local list 8620 // to find its end and append the global list to 8621 // our suffix before returning it. This suffix walk can 8622 // prove to be expensive (quadratic in the amount of traffic) 8623 // when there are many objects in the overflow list and 8624 // there is much producer-consumer contention on the list. 8625 // *NOTE*: The overflow list manipulation code here and 8626 // in ParNewGeneration:: are very similar in shape, 8627 // except that in the ParNew case we use the old (from/eden) 8628 // copy of the object to thread the list via its klass word. 8629 // Because of the common code, if you make any changes in 8630 // the code below, please check the ParNew version to see if 8631 // similar changes might be needed. 8632 // CR 6797058 has been filed to consolidate the common code. 8633 bool CMSCollector::par_take_from_overflow_list(size_t num, 8634 OopTaskQueue* work_q) { 8635 assert(work_q->size() == 0, "First empty local work queue"); 8636 assert(num < work_q->max_elems(), "Can't bite more than we can chew"); 8637 if (_overflow_list == NULL) { 8638 return false; 8639 } 8640 // Grab the entire list; we'll put back a suffix 8641 oop prefix = (oop)Atomic::xchg_ptr(BUSY, &_overflow_list); 8642 Thread* tid = Thread::current(); 8643 size_t CMSOverflowSpinCount = (size_t)ParallelGCThreads; 8644 size_t sleep_time_millis = MAX2((size_t)1, num/100); 8645 // If the list is busy, we spin for a short while, 8646 // sleeping between attempts to get the list. 8647 for (size_t spin = 0; prefix == BUSY && spin < CMSOverflowSpinCount; spin++) { 8648 os::sleep(tid, sleep_time_millis, false); 8649 if (_overflow_list == NULL) { 8650 // Nothing left to take 8651 return false; 8652 } else if (_overflow_list != BUSY) { 8653 // Try and grab the prefix 8654 prefix = (oop)Atomic::xchg_ptr(BUSY, &_overflow_list); 8655 } 8656 } 8657 // If the list was found to be empty, or we spun long 8658 // enough, we give up and return empty-handed. If we leave 8659 // the list in the BUSY state below, it must be the case that 8660 // some other thread holds the overflow list and will set it 8661 // to a non-BUSY state in the future. 8662 if (prefix == NULL || prefix == BUSY) { 8663 // Nothing to take or waited long enough 8664 if (prefix == NULL) { 8665 // Write back the NULL in case we overwrote it with BUSY above 8666 // and it is still the same value. 8667 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY); 8668 } 8669 return false; 8670 } 8671 assert(prefix != NULL && prefix != BUSY, "Error"); 8672 size_t i = num; 8673 oop cur = prefix; 8674 // Walk down the first "num" objects, unless we reach the end. 8675 for (; i > 1 && cur->mark() != NULL; cur = oop(cur->mark()), i--); 8676 if (cur->mark() == NULL) { 8677 // We have "num" or fewer elements in the list, so there 8678 // is nothing to return to the global list. 8679 // Write back the NULL in lieu of the BUSY we wrote 8680 // above, if it is still the same value. 8681 if (_overflow_list == BUSY) { 8682 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY); 8683 } 8684 } else { 8685 // Chop off the suffix and rerturn it to the global list. 8686 assert(cur->mark() != BUSY, "Error"); 8687 oop suffix_head = cur->mark(); // suffix will be put back on global list 8688 cur->set_mark(NULL); // break off suffix 8689 // It's possible that the list is still in the empty(busy) state 8690 // we left it in a short while ago; in that case we may be 8691 // able to place back the suffix without incurring the cost 8692 // of a walk down the list. 8693 oop observed_overflow_list = _overflow_list; 8694 oop cur_overflow_list = observed_overflow_list; 8695 bool attached = false; 8696 while (observed_overflow_list == BUSY || observed_overflow_list == NULL) { 8697 observed_overflow_list = 8698 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list); 8699 if (cur_overflow_list == observed_overflow_list) { 8700 attached = true; 8701 break; 8702 } else cur_overflow_list = observed_overflow_list; 8703 } 8704 if (!attached) { 8705 // Too bad, someone else sneaked in (at least) an element; we'll need 8706 // to do a splice. Find tail of suffix so we can prepend suffix to global 8707 // list. 8708 for (cur = suffix_head; cur->mark() != NULL; cur = (oop)(cur->mark())); 8709 oop suffix_tail = cur; 8710 assert(suffix_tail != NULL && suffix_tail->mark() == NULL, 8711 "Tautology"); 8712 observed_overflow_list = _overflow_list; 8713 do { 8714 cur_overflow_list = observed_overflow_list; 8715 if (cur_overflow_list != BUSY) { 8716 // Do the splice ... 8717 suffix_tail->set_mark(markOop(cur_overflow_list)); 8718 } else { // cur_overflow_list == BUSY 8719 suffix_tail->set_mark(NULL); 8720 } 8721 // ... and try to place spliced list back on overflow_list ... 8722 observed_overflow_list = 8723 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list); 8724 } while (cur_overflow_list != observed_overflow_list); 8725 // ... until we have succeeded in doing so. 8726 } 8727 } 8728 8729 // Push the prefix elements on work_q 8730 assert(prefix != NULL, "control point invariant"); 8731 const markOop proto = markOopDesc::prototype(); 8732 oop next; 8733 NOT_PRODUCT(ssize_t n = 0;) 8734 for (cur = prefix; cur != NULL; cur = next) { 8735 next = oop(cur->mark()); 8736 cur->set_mark(proto); // until proven otherwise 8737 assert(cur->is_oop(), "Should be an oop"); 8738 bool res = work_q->push(cur); 8739 assert(res, "Bit off more than we can chew?"); 8740 NOT_PRODUCT(n++;) 8741 } 8742 #ifndef PRODUCT 8743 assert(_num_par_pushes >= n, "Too many pops?"); 8744 Atomic::add_ptr(-(intptr_t)n, &_num_par_pushes); 8745 #endif 8746 return true; 8747 } 8748 8749 // Single-threaded 8750 void CMSCollector::push_on_overflow_list(oop p) { 8751 NOT_PRODUCT(_num_par_pushes++;) 8752 assert(p->is_oop(), "Not an oop"); 8753 preserve_mark_if_necessary(p); 8754 p->set_mark((markOop)_overflow_list); 8755 _overflow_list = p; 8756 } 8757 8758 // Multi-threaded; use CAS to prepend to overflow list 8759 void CMSCollector::par_push_on_overflow_list(oop p) { 8760 NOT_PRODUCT(Atomic::inc_ptr(&_num_par_pushes);) 8761 assert(p->is_oop(), "Not an oop"); 8762 par_preserve_mark_if_necessary(p); 8763 oop observed_overflow_list = _overflow_list; 8764 oop cur_overflow_list; 8765 do { 8766 cur_overflow_list = observed_overflow_list; 8767 if (cur_overflow_list != BUSY) { 8768 p->set_mark(markOop(cur_overflow_list)); 8769 } else { 8770 p->set_mark(NULL); 8771 } 8772 observed_overflow_list = 8773 (oop) Atomic::cmpxchg_ptr(p, &_overflow_list, cur_overflow_list); 8774 } while (cur_overflow_list != observed_overflow_list); 8775 } 8776 #undef BUSY 8777 8778 // Single threaded 8779 // General Note on GrowableArray: pushes may silently fail 8780 // because we are (temporarily) out of C-heap for expanding 8781 // the stack. The problem is quite ubiquitous and affects 8782 // a lot of code in the JVM. The prudent thing for GrowableArray 8783 // to do (for now) is to exit with an error. However, that may 8784 // be too draconian in some cases because the caller may be 8785 // able to recover without much harm. For such cases, we 8786 // should probably introduce a "soft_push" method which returns 8787 // an indication of success or failure with the assumption that 8788 // the caller may be able to recover from a failure; code in 8789 // the VM can then be changed, incrementally, to deal with such 8790 // failures where possible, thus, incrementally hardening the VM 8791 // in such low resource situations. 8792 void CMSCollector::preserve_mark_work(oop p, markOop m) { 8793 if (_preserved_oop_stack == NULL) { 8794 assert(_preserved_mark_stack == NULL, 8795 "bijection with preserved_oop_stack"); 8796 // Allocate the stacks 8797 _preserved_oop_stack = new (ResourceObj::C_HEAP) 8798 GrowableArray<oop>(PreserveMarkStackSize, true); 8799 _preserved_mark_stack = new (ResourceObj::C_HEAP) 8800 GrowableArray<markOop>(PreserveMarkStackSize, true); 8801 if (_preserved_oop_stack == NULL || _preserved_mark_stack == NULL) { 8802 vm_exit_out_of_memory(2* PreserveMarkStackSize * sizeof(oop) /* punt */, 8803 "Preserved Mark/Oop Stack for CMS (C-heap)"); 8804 } 8805 } 8806 _preserved_oop_stack->push(p); 8807 _preserved_mark_stack->push(m); 8808 assert(m == p->mark(), "Mark word changed"); 8809 assert(_preserved_oop_stack->length() == _preserved_mark_stack->length(), 8810 "bijection"); 8811 } 8812 8813 // Single threaded 8814 void CMSCollector::preserve_mark_if_necessary(oop p) { 8815 markOop m = p->mark(); 8816 if (m->must_be_preserved(p)) { 8817 preserve_mark_work(p, m); 8818 } 8819 } 8820 8821 void CMSCollector::par_preserve_mark_if_necessary(oop p) { 8822 markOop m = p->mark(); 8823 if (m->must_be_preserved(p)) { 8824 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag); 8825 // Even though we read the mark word without holding 8826 // the lock, we are assured that it will not change 8827 // because we "own" this oop, so no other thread can 8828 // be trying to push it on the overflow list; see 8829 // the assertion in preserve_mark_work() that checks 8830 // that m == p->mark(). 8831 preserve_mark_work(p, m); 8832 } 8833 } 8834 8835 // We should be able to do this multi-threaded, 8836 // a chunk of stack being a task (this is 8837 // correct because each oop only ever appears 8838 // once in the overflow list. However, it's 8839 // not very easy to completely overlap this with 8840 // other operations, so will generally not be done 8841 // until all work's been completed. Because we 8842 // expect the preserved oop stack (set) to be small, 8843 // it's probably fine to do this single-threaded. 8844 // We can explore cleverer concurrent/overlapped/parallel 8845 // processing of preserved marks if we feel the 8846 // need for this in the future. Stack overflow should 8847 // be so rare in practice and, when it happens, its 8848 // effect on performance so great that this will 8849 // likely just be in the noise anyway. 8850 void CMSCollector::restore_preserved_marks_if_any() { 8851 if (_preserved_oop_stack == NULL) { 8852 assert(_preserved_mark_stack == NULL, 8853 "bijection with preserved_oop_stack"); 8854 return; 8855 } 8856 8857 assert(SafepointSynchronize::is_at_safepoint(), 8858 "world should be stopped"); 8859 assert(Thread::current()->is_ConcurrentGC_thread() || 8860 Thread::current()->is_VM_thread(), 8861 "should be single-threaded"); 8862 8863 int length = _preserved_oop_stack->length(); 8864 assert(_preserved_mark_stack->length() == length, "bijection"); 8865 for (int i = 0; i < length; i++) { 8866 oop p = _preserved_oop_stack->at(i); 8867 assert(p->is_oop(), "Should be an oop"); 8868 assert(_span.contains(p), "oop should be in _span"); 8869 assert(p->mark() == markOopDesc::prototype(), 8870 "Set when taken from overflow list"); 8871 markOop m = _preserved_mark_stack->at(i); 8872 p->set_mark(m); 8873 } 8874 _preserved_mark_stack->clear(); 8875 _preserved_oop_stack->clear(); 8876 assert(_preserved_mark_stack->is_empty() && 8877 _preserved_oop_stack->is_empty(), 8878 "stacks were cleared above"); 8879 } 8880 8881 #ifndef PRODUCT 8882 bool CMSCollector::no_preserved_marks() const { 8883 return ( ( _preserved_mark_stack == NULL 8884 && _preserved_oop_stack == NULL) 8885 || ( _preserved_mark_stack->is_empty() 8886 && _preserved_oop_stack->is_empty())); 8887 } 8888 #endif 8889 8890 CMSAdaptiveSizePolicy* ASConcurrentMarkSweepGeneration::cms_size_policy() const 8891 { 8892 GenCollectedHeap* gch = (GenCollectedHeap*) GenCollectedHeap::heap(); 8893 CMSAdaptiveSizePolicy* size_policy = 8894 (CMSAdaptiveSizePolicy*) gch->gen_policy()->size_policy(); 8895 assert(size_policy->is_gc_cms_adaptive_size_policy(), 8896 "Wrong type for size policy"); 8897 return size_policy; 8898 } 8899 8900 void ASConcurrentMarkSweepGeneration::resize(size_t cur_promo_size, 8901 size_t desired_promo_size) { 8902 if (cur_promo_size < desired_promo_size) { 8903 size_t expand_bytes = desired_promo_size - cur_promo_size; 8904 if (PrintAdaptiveSizePolicy && Verbose) { 8905 gclog_or_tty->print_cr(" ASConcurrentMarkSweepGeneration::resize " 8906 "Expanding tenured generation by " SIZE_FORMAT " (bytes)", 8907 expand_bytes); 8908 } 8909 expand(expand_bytes, 8910 MinHeapDeltaBytes, 8911 CMSExpansionCause::_adaptive_size_policy); 8912 } else if (desired_promo_size < cur_promo_size) { 8913 size_t shrink_bytes = cur_promo_size - desired_promo_size; 8914 if (PrintAdaptiveSizePolicy && Verbose) { 8915 gclog_or_tty->print_cr(" ASConcurrentMarkSweepGeneration::resize " 8916 "Shrinking tenured generation by " SIZE_FORMAT " (bytes)", 8917 shrink_bytes); 8918 } 8919 shrink(shrink_bytes); 8920 } 8921 } 8922 8923 CMSGCAdaptivePolicyCounters* ASConcurrentMarkSweepGeneration::gc_adaptive_policy_counters() { 8924 GenCollectedHeap* gch = GenCollectedHeap::heap(); 8925 CMSGCAdaptivePolicyCounters* counters = 8926 (CMSGCAdaptivePolicyCounters*) gch->collector_policy()->counters(); 8927 assert(counters->kind() == GCPolicyCounters::CMSGCAdaptivePolicyCountersKind, 8928 "Wrong kind of counters"); 8929 return counters; 8930 } 8931 8932 8933 void ASConcurrentMarkSweepGeneration::update_counters() { 8934 if (UsePerfData) { 8935 _space_counters->update_all(); 8936 _gen_counters->update_all(); 8937 CMSGCAdaptivePolicyCounters* counters = gc_adaptive_policy_counters(); 8938 GenCollectedHeap* gch = GenCollectedHeap::heap(); 8939 CMSGCStats* gc_stats_l = (CMSGCStats*) gc_stats(); 8940 assert(gc_stats_l->kind() == GCStats::CMSGCStatsKind, 8941 "Wrong gc statistics type"); 8942 counters->update_counters(gc_stats_l); 8943 } 8944 } 8945 8946 void ASConcurrentMarkSweepGeneration::update_counters(size_t used) { 8947 if (UsePerfData) { 8948 _space_counters->update_used(used); 8949 _space_counters->update_capacity(); 8950 _gen_counters->update_all(); 8951 8952 CMSGCAdaptivePolicyCounters* counters = gc_adaptive_policy_counters(); 8953 GenCollectedHeap* gch = GenCollectedHeap::heap(); 8954 CMSGCStats* gc_stats_l = (CMSGCStats*) gc_stats(); 8955 assert(gc_stats_l->kind() == GCStats::CMSGCStatsKind, 8956 "Wrong gc statistics type"); 8957 counters->update_counters(gc_stats_l); 8958 } 8959 } 8960 8961 // The desired expansion delta is computed so that: 8962 // . desired free percentage or greater is used 8963 void ASConcurrentMarkSweepGeneration::compute_new_size() { 8964 assert_locked_or_safepoint(Heap_lock); 8965 8966 GenCollectedHeap* gch = (GenCollectedHeap*) GenCollectedHeap::heap(); 8967 8968 // If incremental collection failed, we just want to expand 8969 // to the limit. 8970 if (incremental_collection_failed()) { 8971 clear_incremental_collection_failed(); 8972 grow_to_reserved(); 8973 return; 8974 } 8975 8976 assert(UseAdaptiveSizePolicy, "Should be using adaptive sizing"); 8977 8978 assert(gch->kind() == CollectedHeap::GenCollectedHeap, 8979 "Wrong type of heap"); 8980 int prev_level = level() - 1; 8981 assert(prev_level >= 0, "The cms generation is the lowest generation"); 8982 Generation* prev_gen = gch->get_gen(prev_level); 8983 assert(prev_gen->kind() == Generation::ASParNew, 8984 "Wrong type of young generation"); 8985 ParNewGeneration* younger_gen = (ParNewGeneration*) prev_gen; 8986 size_t cur_eden = younger_gen->eden()->capacity(); 8987 CMSAdaptiveSizePolicy* size_policy = cms_size_policy(); 8988 size_t cur_promo = free(); 8989 size_policy->compute_tenured_generation_free_space(cur_promo, 8990 max_available(), 8991 cur_eden); 8992 resize(cur_promo, size_policy->promo_size()); 8993 8994 // Record the new size of the space in the cms generation 8995 // that is available for promotions. This is temporary. 8996 // It should be the desired promo size. 8997 size_policy->avg_cms_promo()->sample(free()); 8998 size_policy->avg_old_live()->sample(used()); 8999 9000 if (UsePerfData) { 9001 CMSGCAdaptivePolicyCounters* counters = gc_adaptive_policy_counters(); 9002 counters->update_cms_capacity_counter(capacity()); 9003 } 9004 } 9005 9006 void ASConcurrentMarkSweepGeneration::shrink_by(size_t desired_bytes) { 9007 assert_locked_or_safepoint(Heap_lock); 9008 assert_lock_strong(freelistLock()); 9009 HeapWord* old_end = _cmsSpace->end(); 9010 HeapWord* unallocated_start = _cmsSpace->unallocated_block(); 9011 assert(old_end >= unallocated_start, "Miscalculation of unallocated_start"); 9012 FreeChunk* chunk_at_end = find_chunk_at_end(); 9013 if (chunk_at_end == NULL) { 9014 // No room to shrink 9015 if (PrintGCDetails && Verbose) { 9016 gclog_or_tty->print_cr("No room to shrink: old_end " 9017 PTR_FORMAT " unallocated_start " PTR_FORMAT 9018 " chunk_at_end " PTR_FORMAT, 9019 old_end, unallocated_start, chunk_at_end); 9020 } 9021 return; 9022 } else { 9023 9024 // Find the chunk at the end of the space and determine 9025 // how much it can be shrunk. 9026 size_t shrinkable_size_in_bytes = chunk_at_end->size(); 9027 size_t aligned_shrinkable_size_in_bytes = 9028 align_size_down(shrinkable_size_in_bytes, os::vm_page_size()); 9029 assert(unallocated_start <= chunk_at_end->end(), 9030 "Inconsistent chunk at end of space"); 9031 size_t bytes = MIN2(desired_bytes, aligned_shrinkable_size_in_bytes); 9032 size_t word_size_before = heap_word_size(_virtual_space.committed_size()); 9033 9034 // Shrink the underlying space 9035 _virtual_space.shrink_by(bytes); 9036 if (PrintGCDetails && Verbose) { 9037 gclog_or_tty->print_cr("ConcurrentMarkSweepGeneration::shrink_by:" 9038 " desired_bytes " SIZE_FORMAT 9039 " shrinkable_size_in_bytes " SIZE_FORMAT 9040 " aligned_shrinkable_size_in_bytes " SIZE_FORMAT 9041 " bytes " SIZE_FORMAT, 9042 desired_bytes, shrinkable_size_in_bytes, 9043 aligned_shrinkable_size_in_bytes, bytes); 9044 gclog_or_tty->print_cr(" old_end " SIZE_FORMAT 9045 " unallocated_start " SIZE_FORMAT, 9046 old_end, unallocated_start); 9047 } 9048 9049 // If the space did shrink (shrinking is not guaranteed), 9050 // shrink the chunk at the end by the appropriate amount. 9051 if (((HeapWord*)_virtual_space.high()) < old_end) { 9052 size_t new_word_size = 9053 heap_word_size(_virtual_space.committed_size()); 9054 9055 // Have to remove the chunk from the dictionary because it is changing 9056 // size and might be someplace elsewhere in the dictionary. 9057 9058 // Get the chunk at end, shrink it, and put it 9059 // back. 9060 _cmsSpace->removeChunkFromDictionary(chunk_at_end); 9061 size_t word_size_change = word_size_before - new_word_size; 9062 size_t chunk_at_end_old_size = chunk_at_end->size(); 9063 assert(chunk_at_end_old_size >= word_size_change, 9064 "Shrink is too large"); 9065 chunk_at_end->setSize(chunk_at_end_old_size - 9066 word_size_change); 9067 _cmsSpace->freed((HeapWord*) chunk_at_end->end(), 9068 word_size_change); 9069 9070 _cmsSpace->returnChunkToDictionary(chunk_at_end); 9071 9072 MemRegion mr(_cmsSpace->bottom(), new_word_size); 9073 _bts->resize(new_word_size); // resize the block offset shared array 9074 Universe::heap()->barrier_set()->resize_covered_region(mr); 9075 _cmsSpace->assert_locked(); 9076 _cmsSpace->set_end((HeapWord*)_virtual_space.high()); 9077 9078 NOT_PRODUCT(_cmsSpace->dictionary()->verify()); 9079 9080 // update the space and generation capacity counters 9081 if (UsePerfData) { 9082 _space_counters->update_capacity(); 9083 _gen_counters->update_all(); 9084 } 9085 9086 if (Verbose && PrintGCDetails) { 9087 size_t new_mem_size = _virtual_space.committed_size(); 9088 size_t old_mem_size = new_mem_size + bytes; 9089 gclog_or_tty->print_cr("Shrinking %s from %ldK by %ldK to %ldK", 9090 name(), old_mem_size/K, bytes/K, new_mem_size/K); 9091 } 9092 } 9093 9094 assert(_cmsSpace->unallocated_block() <= _cmsSpace->end(), 9095 "Inconsistency at end of space"); 9096 assert(chunk_at_end->end() == _cmsSpace->end(), 9097 "Shrinking is inconsistent"); 9098 return; 9099 } 9100 } 9101 9102 // Transfer some number of overflown objects to usual marking 9103 // stack. Return true if some objects were transferred. 9104 bool MarkRefsIntoAndScanClosure::take_from_overflow_list() { 9105 size_t num = MIN2((size_t)(_mark_stack->capacity() - _mark_stack->length())/4, 9106 (size_t)ParGCDesiredObjsFromOverflowList); 9107 9108 bool res = _collector->take_from_overflow_list(num, _mark_stack); 9109 assert(_collector->overflow_list_is_empty() || res, 9110 "If list is not empty, we should have taken something"); 9111 assert(!res || !_mark_stack->isEmpty(), 9112 "If we took something, it should now be on our stack"); 9113 return res; 9114 } 9115 9116 size_t MarkDeadObjectsClosure::do_blk(HeapWord* addr) { 9117 size_t res = _sp->block_size_no_stall(addr, _collector); 9118 assert(res != 0, "Should always be able to compute a size"); 9119 if (_sp->block_is_obj(addr)) { 9120 if (_live_bit_map->isMarked(addr)) { 9121 // It can't have been dead in a previous cycle 9122 guarantee(!_dead_bit_map->isMarked(addr), "No resurrection!"); 9123 } else { 9124 _dead_bit_map->mark(addr); // mark the dead object 9125 } 9126 } 9127 return res; 9128 }