1 /* 2 * Copyright 2001-2007 Sun Microsystems, Inc. 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 Sun Microsystems, Inc., 4150 Network Circle, Santa Clara, 20 * CA 95054 USA or visit www.sun.com if you need additional information or 21 * have any 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)(oopDesc::header_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(junk->prev_addr() == (void*)(oop(junk)->klass_addr()), 194 "Offset of FreeChunk::_prev within FreeChunk must match" 195 " that of OopDesc::_klass within OopDesc"); 196 ) 197 if (ParallelGCThreads > 0) { 198 typedef CMSParGCThreadState* CMSParGCThreadStatePtr; 199 _par_gc_thread_states = 200 NEW_C_HEAP_ARRAY(CMSParGCThreadStatePtr, ParallelGCThreads); 201 if (_par_gc_thread_states == NULL) { 202 vm_exit_during_initialization("Could not allocate par gc structs"); 203 } 204 for (uint i = 0; i < ParallelGCThreads; i++) { 205 _par_gc_thread_states[i] = new CMSParGCThreadState(cmsSpace()); 206 if (_par_gc_thread_states[i] == NULL) { 207 vm_exit_during_initialization("Could not allocate par gc structs"); 208 } 209 } 210 } else { 211 _par_gc_thread_states = NULL; 212 } 213 _incremental_collection_failed = false; 214 // The "dilatation_factor" is the expansion that can occur on 215 // account of the fact that the minimum object size in the CMS 216 // generation may be larger than that in, say, a contiguous young 217 // generation. 218 // Ideally, in the calculation below, we'd compute the dilatation 219 // factor as: MinChunkSize/(promoting_gen's min object size) 220 // Since we do not have such a general query interface for the 221 // promoting generation, we'll instead just use the mimimum 222 // object size (which today is a header's worth of space); 223 // note that all arithmetic is in units of HeapWords. 224 assert(MinChunkSize >= oopDesc::header_size(), "just checking"); 225 assert(_dilatation_factor >= 1.0, "from previous assert"); 226 } 227 228 229 // The field "_initiating_occupancy" represents the occupancy percentage 230 // at which we trigger a new collection cycle. Unless explicitly specified 231 // via CMSInitiating[Perm]OccupancyFraction (argument "io" below), it 232 // is calculated by: 233 // 234 // Let "f" be MinHeapFreeRatio in 235 // 236 // _intiating_occupancy = 100-f + 237 // f * (CMSTrigger[Perm]Ratio/100) 238 // where CMSTrigger[Perm]Ratio is the argument "tr" below. 239 // 240 // That is, if we assume the heap is at its desired maximum occupancy at the 241 // end of a collection, we let CMSTrigger[Perm]Ratio of the (purported) free 242 // space be allocated before initiating a new collection cycle. 243 // 244 void ConcurrentMarkSweepGeneration::init_initiating_occupancy(intx io, intx tr) { 245 assert(io <= 100 && tr >= 0 && tr <= 100, "Check the arguments"); 246 if (io >= 0) { 247 _initiating_occupancy = (double)io / 100.0; 248 } else { 249 _initiating_occupancy = ((100 - MinHeapFreeRatio) + 250 (double)(tr * MinHeapFreeRatio) / 100.0) 251 / 100.0; 252 } 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 // If promotion failure handling is on use 344 // the padded average size of the promotion for each 345 // young generation collection. 346 double CMSStats::time_until_cms_gen_full() const { 347 size_t cms_free = _cms_gen->cmsSpace()->free(); 348 GenCollectedHeap* gch = GenCollectedHeap::heap(); 349 size_t expected_promotion = gch->get_gen(0)->capacity(); 350 if (HandlePromotionFailure) { 351 expected_promotion = MIN2( 352 (size_t) _cms_gen->gc_stats()->avg_promoted()->padded_average(), 353 expected_promotion); 354 } 355 if (cms_free > expected_promotion) { 356 // Start a cms collection if there isn't enough space to promote 357 // for the next minor collection. Use the padded average as 358 // a safety factor. 359 cms_free -= expected_promotion; 360 361 // Adjust by the safety factor. 362 double cms_free_dbl = (double)cms_free; 363 cms_free_dbl = cms_free_dbl * (100.0 - CMSIncrementalSafetyFactor) / 100.0; 364 365 if (PrintGCDetails && Verbose) { 366 gclog_or_tty->print_cr("CMSStats::time_until_cms_gen_full: cms_free " 367 SIZE_FORMAT " expected_promotion " SIZE_FORMAT, 368 cms_free, expected_promotion); 369 gclog_or_tty->print_cr(" cms_free_dbl %f cms_consumption_rate %f", 370 cms_free_dbl, cms_consumption_rate() + 1.0); 371 } 372 // Add 1 in case the consumption rate goes to zero. 373 return cms_free_dbl / (cms_consumption_rate() + 1.0); 374 } 375 return 0.0; 376 } 377 378 // Compare the duration of the cms collection to the 379 // time remaining before the cms generation is empty. 380 // Note that the time from the start of the cms collection 381 // to the start of the cms sweep (less than the total 382 // duration of the cms collection) can be used. This 383 // has been tried and some applications experienced 384 // promotion failures early in execution. This was 385 // possibly because the averages were not accurate 386 // enough at the beginning. 387 double CMSStats::time_until_cms_start() const { 388 // We add "gc0_period" to the "work" calculation 389 // below because this query is done (mostly) at the 390 // end of a scavenge, so we need to conservatively 391 // account for that much possible delay 392 // in the query so as to avoid concurrent mode failures 393 // due to starting the collection just a wee bit too 394 // late. 395 double work = cms_duration() + gc0_period(); 396 double deadline = time_until_cms_gen_full(); 397 if (work > deadline) { 398 if (Verbose && PrintGCDetails) { 399 gclog_or_tty->print( 400 " CMSCollector: collect because of anticipated promotion " 401 "before full %3.7f + %3.7f > %3.7f ", cms_duration(), 402 gc0_period(), time_until_cms_gen_full()); 403 } 404 return 0.0; 405 } 406 return work - deadline; 407 } 408 409 // Return a duty cycle based on old_duty_cycle and new_duty_cycle, limiting the 410 // amount of change to prevent wild oscillation. 411 unsigned int CMSStats::icms_damped_duty_cycle(unsigned int old_duty_cycle, 412 unsigned int new_duty_cycle) { 413 assert(old_duty_cycle <= 100, "bad input value"); 414 assert(new_duty_cycle <= 100, "bad input value"); 415 416 // Note: use subtraction with caution since it may underflow (values are 417 // unsigned). Addition is safe since we're in the range 0-100. 418 unsigned int damped_duty_cycle = new_duty_cycle; 419 if (new_duty_cycle < old_duty_cycle) { 420 const unsigned int largest_delta = MAX2(old_duty_cycle / 4, 5U); 421 if (new_duty_cycle + largest_delta < old_duty_cycle) { 422 damped_duty_cycle = old_duty_cycle - largest_delta; 423 } 424 } else if (new_duty_cycle > old_duty_cycle) { 425 const unsigned int largest_delta = MAX2(old_duty_cycle / 4, 15U); 426 if (new_duty_cycle > old_duty_cycle + largest_delta) { 427 damped_duty_cycle = MIN2(old_duty_cycle + largest_delta, 100U); 428 } 429 } 430 assert(damped_duty_cycle <= 100, "invalid duty cycle computed"); 431 432 if (CMSTraceIncrementalPacing) { 433 gclog_or_tty->print(" [icms_damped_duty_cycle(%d,%d) = %d] ", 434 old_duty_cycle, new_duty_cycle, damped_duty_cycle); 435 } 436 return damped_duty_cycle; 437 } 438 439 unsigned int CMSStats::icms_update_duty_cycle_impl() { 440 assert(CMSIncrementalPacing && valid(), 441 "should be handled in icms_update_duty_cycle()"); 442 443 double cms_time_so_far = cms_timer().seconds(); 444 double scaled_duration = cms_duration_per_mb() * _cms_used_at_gc0_end / M; 445 double scaled_duration_remaining = fabsd(scaled_duration - cms_time_so_far); 446 447 // Avoid division by 0. 448 double time_until_full = MAX2(time_until_cms_gen_full(), 0.01); 449 double duty_cycle_dbl = 100.0 * scaled_duration_remaining / time_until_full; 450 451 unsigned int new_duty_cycle = MIN2((unsigned int)duty_cycle_dbl, 100U); 452 if (new_duty_cycle > _icms_duty_cycle) { 453 // Avoid very small duty cycles (1 or 2); 0 is allowed. 454 if (new_duty_cycle > 2) { 455 _icms_duty_cycle = icms_damped_duty_cycle(_icms_duty_cycle, 456 new_duty_cycle); 457 } 458 } else if (_allow_duty_cycle_reduction) { 459 // The duty cycle is reduced only once per cms cycle (see record_cms_end()). 460 new_duty_cycle = icms_damped_duty_cycle(_icms_duty_cycle, new_duty_cycle); 461 // Respect the minimum duty cycle. 462 unsigned int min_duty_cycle = (unsigned int)CMSIncrementalDutyCycleMin; 463 _icms_duty_cycle = MAX2(new_duty_cycle, min_duty_cycle); 464 } 465 466 if (PrintGCDetails || CMSTraceIncrementalPacing) { 467 gclog_or_tty->print(" icms_dc=%d ", _icms_duty_cycle); 468 } 469 470 _allow_duty_cycle_reduction = false; 471 return _icms_duty_cycle; 472 } 473 474 #ifndef PRODUCT 475 void CMSStats::print_on(outputStream *st) const { 476 st->print(" gc0_alpha=%d,cms_alpha=%d", _gc0_alpha, _cms_alpha); 477 st->print(",gc0_dur=%g,gc0_per=%g,gc0_promo=" SIZE_FORMAT, 478 gc0_duration(), gc0_period(), gc0_promoted()); 479 st->print(",cms_dur=%g,cms_dur_per_mb=%g,cms_per=%g,cms_alloc=" SIZE_FORMAT, 480 cms_duration(), cms_duration_per_mb(), 481 cms_period(), cms_allocated()); 482 st->print(",cms_since_beg=%g,cms_since_end=%g", 483 cms_time_since_begin(), cms_time_since_end()); 484 st->print(",cms_used_beg=" SIZE_FORMAT ",cms_used_end=" SIZE_FORMAT, 485 _cms_used_at_gc0_begin, _cms_used_at_gc0_end); 486 if (CMSIncrementalMode) { 487 st->print(",dc=%d", icms_duty_cycle()); 488 } 489 490 if (valid()) { 491 st->print(",promo_rate=%g,cms_alloc_rate=%g", 492 promotion_rate(), cms_allocation_rate()); 493 st->print(",cms_consumption_rate=%g,time_until_full=%g", 494 cms_consumption_rate(), time_until_cms_gen_full()); 495 } 496 st->print(" "); 497 } 498 #endif // #ifndef PRODUCT 499 500 CMSCollector::CollectorState CMSCollector::_collectorState = 501 CMSCollector::Idling; 502 bool CMSCollector::_foregroundGCIsActive = false; 503 bool CMSCollector::_foregroundGCShouldWait = false; 504 505 CMSCollector::CMSCollector(ConcurrentMarkSweepGeneration* cmsGen, 506 ConcurrentMarkSweepGeneration* permGen, 507 CardTableRS* ct, 508 ConcurrentMarkSweepPolicy* cp): 509 _cmsGen(cmsGen), 510 _permGen(permGen), 511 _ct(ct), 512 _ref_processor(NULL), // will be set later 513 _conc_workers(NULL), // may be set later 514 _abort_preclean(false), 515 _start_sampling(false), 516 _between_prologue_and_epilogue(false), 517 _markBitMap(0, Mutex::leaf + 1, "CMS_markBitMap_lock"), 518 _perm_gen_verify_bit_map(0, -1 /* no mutex */, "No_lock"), 519 _modUnionTable((CardTableModRefBS::card_shift - LogHeapWordSize), 520 -1 /* lock-free */, "No_lock" /* dummy */), 521 _modUnionClosure(&_modUnionTable), 522 _modUnionClosurePar(&_modUnionTable), 523 _is_alive_closure(&_markBitMap), 524 _restart_addr(NULL), 525 _overflow_list(NULL), 526 _preserved_oop_stack(NULL), 527 _preserved_mark_stack(NULL), 528 _stats(cmsGen), 529 _eden_chunk_array(NULL), // may be set in ctor body 530 _eden_chunk_capacity(0), // -- ditto -- 531 _eden_chunk_index(0), // -- ditto -- 532 _survivor_plab_array(NULL), // -- ditto -- 533 _survivor_chunk_array(NULL), // -- ditto -- 534 _survivor_chunk_capacity(0), // -- ditto -- 535 _survivor_chunk_index(0), // -- ditto -- 536 _ser_pmc_preclean_ovflw(0), 537 _ser_pmc_remark_ovflw(0), 538 _par_pmc_remark_ovflw(0), 539 _ser_kac_ovflw(0), 540 _par_kac_ovflw(0), 541 #ifndef PRODUCT 542 _num_par_pushes(0), 543 #endif 544 _collection_count_start(0), 545 _verifying(false), 546 _icms_start_limit(NULL), 547 _icms_stop_limit(NULL), 548 _verification_mark_bm(0, Mutex::leaf + 1, "CMS_verification_mark_bm_lock"), 549 _completed_initialization(false), 550 _collector_policy(cp), 551 _should_unload_classes(false), 552 _concurrent_cycles_since_last_unload(0), 553 _sweep_count(0), 554 _sweep_estimate(CMS_SweepWeight, CMS_SweepPadding) 555 { 556 if (ExplicitGCInvokesConcurrentAndUnloadsClasses) { 557 ExplicitGCInvokesConcurrent = true; 558 } 559 // Now expand the span and allocate the collection support structures 560 // (MUT, marking bit map etc.) to cover both generations subject to 561 // collection. 562 563 // First check that _permGen is adjacent to _cmsGen and above it. 564 assert( _cmsGen->reserved().word_size() > 0 565 && _permGen->reserved().word_size() > 0, 566 "generations should not be of zero size"); 567 assert(_cmsGen->reserved().intersection(_permGen->reserved()).is_empty(), 568 "_cmsGen and _permGen should not overlap"); 569 assert(_cmsGen->reserved().end() == _permGen->reserved().start(), 570 "_cmsGen->end() different from _permGen->start()"); 571 572 // For use by dirty card to oop closures. 573 _cmsGen->cmsSpace()->set_collector(this); 574 _permGen->cmsSpace()->set_collector(this); 575 576 // Adjust my span to cover old (cms) gen and perm gen 577 _span = _cmsGen->reserved()._union(_permGen->reserved()); 578 // Initialize the span of is_alive_closure 579 _is_alive_closure.set_span(_span); 580 581 // Allocate MUT and marking bit map 582 { 583 MutexLockerEx x(_markBitMap.lock(), Mutex::_no_safepoint_check_flag); 584 if (!_markBitMap.allocate(_span)) { 585 warning("Failed to allocate CMS Bit Map"); 586 return; 587 } 588 assert(_markBitMap.covers(_span), "_markBitMap inconsistency?"); 589 } 590 { 591 _modUnionTable.allocate(_span); 592 assert(_modUnionTable.covers(_span), "_modUnionTable inconsistency?"); 593 } 594 595 if (!_markStack.allocate(CMSMarkStackSize)) { 596 warning("Failed to allocate CMS Marking Stack"); 597 return; 598 } 599 if (!_revisitStack.allocate(CMSRevisitStackSize)) { 600 warning("Failed to allocate CMS Revisit Stack"); 601 return; 602 } 603 604 // Support for multi-threaded concurrent phases 605 if (ParallelGCThreads > 0 && CMSConcurrentMTEnabled) { 606 if (FLAG_IS_DEFAULT(ParallelCMSThreads)) { 607 // just for now 608 FLAG_SET_DEFAULT(ParallelCMSThreads, (ParallelGCThreads + 3)/4); 609 } 610 if (ParallelCMSThreads > 1) { 611 _conc_workers = new YieldingFlexibleWorkGang("Parallel CMS Threads", 612 ParallelCMSThreads, true); 613 if (_conc_workers == NULL) { 614 warning("GC/CMS: _conc_workers allocation failure: " 615 "forcing -CMSConcurrentMTEnabled"); 616 CMSConcurrentMTEnabled = false; 617 } 618 } else { 619 CMSConcurrentMTEnabled = false; 620 } 621 } 622 if (!CMSConcurrentMTEnabled) { 623 ParallelCMSThreads = 0; 624 } else { 625 // Turn off CMSCleanOnEnter optimization temporarily for 626 // the MT case where it's not fixed yet; see 6178663. 627 CMSCleanOnEnter = false; 628 } 629 assert((_conc_workers != NULL) == (ParallelCMSThreads > 1), 630 "Inconsistency"); 631 632 // Parallel task queues; these are shared for the 633 // concurrent and stop-world phases of CMS, but 634 // are not shared with parallel scavenge (ParNew). 635 { 636 uint i; 637 uint num_queues = (uint) MAX2(ParallelGCThreads, ParallelCMSThreads); 638 639 if ((CMSParallelRemarkEnabled || CMSConcurrentMTEnabled 640 || ParallelRefProcEnabled) 641 && num_queues > 0) { 642 _task_queues = new OopTaskQueueSet(num_queues); 643 if (_task_queues == NULL) { 644 warning("task_queues allocation failure."); 645 return; 646 } 647 _hash_seed = NEW_C_HEAP_ARRAY(int, num_queues); 648 if (_hash_seed == NULL) { 649 warning("_hash_seed array allocation failure"); 650 return; 651 } 652 653 // XXX use a global constant instead of 64! 654 typedef struct OopTaskQueuePadded { 655 OopTaskQueue work_queue; 656 char pad[64 - sizeof(OopTaskQueue)]; // prevent false sharing 657 } OopTaskQueuePadded; 658 659 for (i = 0; i < num_queues; i++) { 660 OopTaskQueuePadded *q_padded = new OopTaskQueuePadded(); 661 if (q_padded == NULL) { 662 warning("work_queue allocation failure."); 663 return; 664 } 665 _task_queues->register_queue(i, &q_padded->work_queue); 666 } 667 for (i = 0; i < num_queues; i++) { 668 _task_queues->queue(i)->initialize(); 669 _hash_seed[i] = 17; // copied from ParNew 670 } 671 } 672 } 673 674 _cmsGen ->init_initiating_occupancy(CMSInitiatingOccupancyFraction, CMSTriggerRatio); 675 _permGen->init_initiating_occupancy(CMSInitiatingPermOccupancyFraction, CMSTriggerPermRatio); 676 677 // Clip CMSBootstrapOccupancy between 0 and 100. 678 _bootstrap_occupancy = ((double)MIN2((uintx)100, MAX2((uintx)0, CMSBootstrapOccupancy))) 679 /(double)100; 680 681 _full_gcs_since_conc_gc = 0; 682 683 // Now tell CMS generations the identity of their collector 684 ConcurrentMarkSweepGeneration::set_collector(this); 685 686 // Create & start a CMS thread for this CMS collector 687 _cmsThread = ConcurrentMarkSweepThread::start(this); 688 assert(cmsThread() != NULL, "CMS Thread should have been created"); 689 assert(cmsThread()->collector() == this, 690 "CMS Thread should refer to this gen"); 691 assert(CGC_lock != NULL, "Where's the CGC_lock?"); 692 693 // Support for parallelizing young gen rescan 694 GenCollectedHeap* gch = GenCollectedHeap::heap(); 695 _young_gen = gch->prev_gen(_cmsGen); 696 if (gch->supports_inline_contig_alloc()) { 697 _top_addr = gch->top_addr(); 698 _end_addr = gch->end_addr(); 699 assert(_young_gen != NULL, "no _young_gen"); 700 _eden_chunk_index = 0; 701 _eden_chunk_capacity = (_young_gen->max_capacity()+CMSSamplingGrain)/CMSSamplingGrain; 702 _eden_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, _eden_chunk_capacity); 703 if (_eden_chunk_array == NULL) { 704 _eden_chunk_capacity = 0; 705 warning("GC/CMS: _eden_chunk_array allocation failure"); 706 } 707 } 708 assert(_eden_chunk_array != NULL || _eden_chunk_capacity == 0, "Error"); 709 710 // Support for parallelizing survivor space rescan 711 if (CMSParallelRemarkEnabled && CMSParallelSurvivorRemarkEnabled) { 712 size_t max_plab_samples = MaxNewSize/((SurvivorRatio+2)*MinTLABSize); 713 _survivor_plab_array = NEW_C_HEAP_ARRAY(ChunkArray, ParallelGCThreads); 714 _survivor_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, 2*max_plab_samples); 715 _cursor = NEW_C_HEAP_ARRAY(size_t, ParallelGCThreads); 716 if (_survivor_plab_array == NULL || _survivor_chunk_array == NULL 717 || _cursor == NULL) { 718 warning("Failed to allocate survivor plab/chunk array"); 719 if (_survivor_plab_array != NULL) { 720 FREE_C_HEAP_ARRAY(ChunkArray, _survivor_plab_array); 721 _survivor_plab_array = NULL; 722 } 723 if (_survivor_chunk_array != NULL) { 724 FREE_C_HEAP_ARRAY(HeapWord*, _survivor_chunk_array); 725 _survivor_chunk_array = NULL; 726 } 727 if (_cursor != NULL) { 728 FREE_C_HEAP_ARRAY(size_t, _cursor); 729 _cursor = NULL; 730 } 731 } else { 732 _survivor_chunk_capacity = 2*max_plab_samples; 733 for (uint i = 0; i < ParallelGCThreads; i++) { 734 HeapWord** vec = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples); 735 if (vec == NULL) { 736 warning("Failed to allocate survivor plab array"); 737 for (int j = i; j > 0; j--) { 738 FREE_C_HEAP_ARRAY(HeapWord*, _survivor_plab_array[j-1].array()); 739 } 740 FREE_C_HEAP_ARRAY(ChunkArray, _survivor_plab_array); 741 FREE_C_HEAP_ARRAY(HeapWord*, _survivor_chunk_array); 742 _survivor_plab_array = NULL; 743 _survivor_chunk_array = NULL; 744 _survivor_chunk_capacity = 0; 745 break; 746 } else { 747 ChunkArray* cur = 748 ::new (&_survivor_plab_array[i]) ChunkArray(vec, 749 max_plab_samples); 750 assert(cur->end() == 0, "Should be 0"); 751 assert(cur->array() == vec, "Should be vec"); 752 assert(cur->capacity() == max_plab_samples, "Error"); 753 } 754 } 755 } 756 } 757 assert( ( _survivor_plab_array != NULL 758 && _survivor_chunk_array != NULL) 759 || ( _survivor_chunk_capacity == 0 760 && _survivor_chunk_index == 0), 761 "Error"); 762 763 // Choose what strong roots should be scanned depending on verification options 764 // and perm gen collection mode. 765 if (!CMSClassUnloadingEnabled) { 766 // If class unloading is disabled we want to include all classes into the root set. 767 add_root_scanning_option(SharedHeap::SO_AllClasses); 768 } else { 769 add_root_scanning_option(SharedHeap::SO_SystemClasses); 770 } 771 772 NOT_PRODUCT(_overflow_counter = CMSMarkStackOverflowInterval;) 773 _gc_counters = new CollectorCounters("CMS", 1); 774 _completed_initialization = true; 775 _sweep_timer.start(); // start of time 776 } 777 778 const char* ConcurrentMarkSweepGeneration::name() const { 779 return "concurrent mark-sweep generation"; 780 } 781 void ConcurrentMarkSweepGeneration::update_counters() { 782 if (UsePerfData) { 783 _space_counters->update_all(); 784 _gen_counters->update_all(); 785 } 786 } 787 788 // this is an optimized version of update_counters(). it takes the 789 // used value as a parameter rather than computing it. 790 // 791 void ConcurrentMarkSweepGeneration::update_counters(size_t used) { 792 if (UsePerfData) { 793 _space_counters->update_used(used); 794 _space_counters->update_capacity(); 795 _gen_counters->update_all(); 796 } 797 } 798 799 void ConcurrentMarkSweepGeneration::print() const { 800 Generation::print(); 801 cmsSpace()->print(); 802 } 803 804 #ifndef PRODUCT 805 void ConcurrentMarkSweepGeneration::print_statistics() { 806 cmsSpace()->printFLCensus(0); 807 } 808 #endif 809 810 void ConcurrentMarkSweepGeneration::printOccupancy(const char *s) { 811 GenCollectedHeap* gch = GenCollectedHeap::heap(); 812 if (PrintGCDetails) { 813 if (Verbose) { 814 gclog_or_tty->print(" [%d %s-%s: "SIZE_FORMAT"("SIZE_FORMAT")]", 815 level(), short_name(), s, used(), capacity()); 816 } else { 817 gclog_or_tty->print(" [%d %s-%s: "SIZE_FORMAT"K("SIZE_FORMAT"K)]", 818 level(), short_name(), s, used() / K, capacity() / K); 819 } 820 } 821 if (Verbose) { 822 gclog_or_tty->print(" "SIZE_FORMAT"("SIZE_FORMAT")", 823 gch->used(), gch->capacity()); 824 } else { 825 gclog_or_tty->print(" "SIZE_FORMAT"K("SIZE_FORMAT"K)", 826 gch->used() / K, gch->capacity() / K); 827 } 828 } 829 830 size_t 831 ConcurrentMarkSweepGeneration::contiguous_available() const { 832 // dld proposes an improvement in precision here. If the committed 833 // part of the space ends in a free block we should add that to 834 // uncommitted size in the calculation below. Will make this 835 // change later, staying with the approximation below for the 836 // time being. -- ysr. 837 return MAX2(_virtual_space.uncommitted_size(), unsafe_max_alloc_nogc()); 838 } 839 840 size_t 841 ConcurrentMarkSweepGeneration::unsafe_max_alloc_nogc() const { 842 return _cmsSpace->max_alloc_in_words() * HeapWordSize; 843 } 844 845 size_t ConcurrentMarkSweepGeneration::max_available() const { 846 return free() + _virtual_space.uncommitted_size(); 847 } 848 849 bool ConcurrentMarkSweepGeneration::promotion_attempt_is_safe( 850 size_t max_promotion_in_bytes, 851 bool younger_handles_promotion_failure) const { 852 853 // This is the most conservative test. Full promotion is 854 // guaranteed if this is used. The multiplicative factor is to 855 // account for the worst case "dilatation". 856 double adjusted_max_promo_bytes = _dilatation_factor * max_promotion_in_bytes; 857 if (adjusted_max_promo_bytes > (double)max_uintx) { // larger than size_t 858 adjusted_max_promo_bytes = (double)max_uintx; 859 } 860 bool result = (max_contiguous_available() >= (size_t)adjusted_max_promo_bytes); 861 862 if (younger_handles_promotion_failure && !result) { 863 // Full promotion is not guaranteed because fragmentation 864 // of the cms generation can prevent the full promotion. 865 result = (max_available() >= (size_t)adjusted_max_promo_bytes); 866 867 if (!result) { 868 // With promotion failure handling the test for the ability 869 // to support the promotion does not have to be guaranteed. 870 // Use an average of the amount promoted. 871 result = max_available() >= (size_t) 872 gc_stats()->avg_promoted()->padded_average(); 873 if (PrintGC && Verbose && result) { 874 gclog_or_tty->print_cr( 875 "\nConcurrentMarkSweepGeneration::promotion_attempt_is_safe" 876 " max_available: " SIZE_FORMAT 877 " avg_promoted: " SIZE_FORMAT, 878 max_available(), (size_t) 879 gc_stats()->avg_promoted()->padded_average()); 880 } 881 } else { 882 if (PrintGC && Verbose) { 883 gclog_or_tty->print_cr( 884 "\nConcurrentMarkSweepGeneration::promotion_attempt_is_safe" 885 " max_available: " SIZE_FORMAT 886 " adj_max_promo_bytes: " SIZE_FORMAT, 887 max_available(), (size_t)adjusted_max_promo_bytes); 888 } 889 } 890 } else { 891 if (PrintGC && Verbose) { 892 gclog_or_tty->print_cr( 893 "\nConcurrentMarkSweepGeneration::promotion_attempt_is_safe" 894 " contiguous_available: " SIZE_FORMAT 895 " adj_max_promo_bytes: " SIZE_FORMAT, 896 max_contiguous_available(), (size_t)adjusted_max_promo_bytes); 897 } 898 } 899 return result; 900 } 901 902 CompactibleSpace* 903 ConcurrentMarkSweepGeneration::first_compaction_space() const { 904 return _cmsSpace; 905 } 906 907 void ConcurrentMarkSweepGeneration::reset_after_compaction() { 908 // Clear the promotion information. These pointers can be adjusted 909 // along with all the other pointers into the heap but 910 // compaction is expected to be a rare event with 911 // a heap using cms so don't do it without seeing the need. 912 if (ParallelGCThreads > 0) { 913 for (uint i = 0; i < ParallelGCThreads; i++) { 914 _par_gc_thread_states[i]->promo.reset(); 915 } 916 } 917 } 918 919 void ConcurrentMarkSweepGeneration::space_iterate(SpaceClosure* blk, bool usedOnly) { 920 blk->do_space(_cmsSpace); 921 } 922 923 void ConcurrentMarkSweepGeneration::compute_new_size() { 924 assert_locked_or_safepoint(Heap_lock); 925 926 // If incremental collection failed, we just want to expand 927 // to the limit. 928 if (incremental_collection_failed()) { 929 clear_incremental_collection_failed(); 930 grow_to_reserved(); 931 return; 932 } 933 934 size_t expand_bytes = 0; 935 double free_percentage = ((double) free()) / capacity(); 936 double desired_free_percentage = (double) MinHeapFreeRatio / 100; 937 double maximum_free_percentage = (double) MaxHeapFreeRatio / 100; 938 939 // compute expansion delta needed for reaching desired free percentage 940 if (free_percentage < desired_free_percentage) { 941 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 942 assert(desired_capacity >= capacity(), "invalid expansion size"); 943 expand_bytes = MAX2(desired_capacity - capacity(), MinHeapDeltaBytes); 944 } 945 if (expand_bytes > 0) { 946 if (PrintGCDetails && Verbose) { 947 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 948 gclog_or_tty->print_cr("\nFrom compute_new_size: "); 949 gclog_or_tty->print_cr(" Free fraction %f", free_percentage); 950 gclog_or_tty->print_cr(" Desired free fraction %f", 951 desired_free_percentage); 952 gclog_or_tty->print_cr(" Maximum free fraction %f", 953 maximum_free_percentage); 954 gclog_or_tty->print_cr(" Capactiy "SIZE_FORMAT, capacity()/1000); 955 gclog_or_tty->print_cr(" Desired capacity "SIZE_FORMAT, 956 desired_capacity/1000); 957 int prev_level = level() - 1; 958 if (prev_level >= 0) { 959 size_t prev_size = 0; 960 GenCollectedHeap* gch = GenCollectedHeap::heap(); 961 Generation* prev_gen = gch->_gens[prev_level]; 962 prev_size = prev_gen->capacity(); 963 gclog_or_tty->print_cr(" Younger gen size "SIZE_FORMAT, 964 prev_size/1000); 965 } 966 gclog_or_tty->print_cr(" unsafe_max_alloc_nogc "SIZE_FORMAT, 967 unsafe_max_alloc_nogc()/1000); 968 gclog_or_tty->print_cr(" contiguous available "SIZE_FORMAT, 969 contiguous_available()/1000); 970 gclog_or_tty->print_cr(" Expand by "SIZE_FORMAT" (bytes)", 971 expand_bytes); 972 } 973 // safe if expansion fails 974 expand(expand_bytes, 0, CMSExpansionCause::_satisfy_free_ratio); 975 if (PrintGCDetails && Verbose) { 976 gclog_or_tty->print_cr(" Expanded free fraction %f", 977 ((double) free()) / capacity()); 978 } 979 } 980 } 981 982 Mutex* ConcurrentMarkSweepGeneration::freelistLock() const { 983 return cmsSpace()->freelistLock(); 984 } 985 986 HeapWord* ConcurrentMarkSweepGeneration::allocate(size_t size, 987 bool tlab) { 988 CMSSynchronousYieldRequest yr; 989 MutexLockerEx x(freelistLock(), 990 Mutex::_no_safepoint_check_flag); 991 return have_lock_and_allocate(size, tlab); 992 } 993 994 HeapWord* ConcurrentMarkSweepGeneration::have_lock_and_allocate(size_t size, 995 bool tlab) { 996 assert_lock_strong(freelistLock()); 997 size_t adjustedSize = CompactibleFreeListSpace::adjustObjectSize(size); 998 HeapWord* res = cmsSpace()->allocate(adjustedSize); 999 // Allocate the object live (grey) if the background collector has 1000 // started marking. This is necessary because the marker may 1001 // have passed this address and consequently this object will 1002 // not otherwise be greyed and would be incorrectly swept up. 1003 // Note that if this object contains references, the writing 1004 // of those references will dirty the card containing this object 1005 // allowing the object to be blackened (and its references scanned) 1006 // either during a preclean phase or at the final checkpoint. 1007 if (res != NULL) { 1008 collector()->direct_allocated(res, adjustedSize); 1009 _direct_allocated_words += adjustedSize; 1010 // allocation counters 1011 NOT_PRODUCT( 1012 _numObjectsAllocated++; 1013 _numWordsAllocated += (int)adjustedSize; 1014 ) 1015 } 1016 return res; 1017 } 1018 1019 // In the case of direct allocation by mutators in a generation that 1020 // is being concurrently collected, the object must be allocated 1021 // live (grey) if the background collector has started marking. 1022 // This is necessary because the marker may 1023 // have passed this address and consequently this object will 1024 // not otherwise be greyed and would be incorrectly swept up. 1025 // Note that if this object contains references, the writing 1026 // of those references will dirty the card containing this object 1027 // allowing the object to be blackened (and its references scanned) 1028 // either during a preclean phase or at the final checkpoint. 1029 void CMSCollector::direct_allocated(HeapWord* start, size_t size) { 1030 assert(_markBitMap.covers(start, size), "Out of bounds"); 1031 if (_collectorState >= Marking) { 1032 MutexLockerEx y(_markBitMap.lock(), 1033 Mutex::_no_safepoint_check_flag); 1034 // [see comments preceding SweepClosure::do_blk() below for details] 1035 // 1. need to mark the object as live so it isn't collected 1036 // 2. need to mark the 2nd bit to indicate the object may be uninitialized 1037 // 3. need to mark the end of the object so sweeper can skip over it 1038 // if it's uninitialized when the sweeper reaches it. 1039 _markBitMap.mark(start); // object is live 1040 _markBitMap.mark(start + 1); // object is potentially uninitialized? 1041 _markBitMap.mark(start + size - 1); 1042 // mark end of object 1043 } 1044 // check that oop looks uninitialized 1045 assert(oop(start)->klass() == NULL, "_klass should be NULL"); 1046 } 1047 1048 void CMSCollector::promoted(bool par, HeapWord* start, 1049 bool is_obj_array, size_t obj_size) { 1050 assert(_markBitMap.covers(start), "Out of bounds"); 1051 // See comment in direct_allocated() about when objects should 1052 // be allocated live. 1053 if (_collectorState >= Marking) { 1054 // we already hold the marking bit map lock, taken in 1055 // the prologue 1056 if (par) { 1057 _markBitMap.par_mark(start); 1058 } else { 1059 _markBitMap.mark(start); 1060 } 1061 // We don't need to mark the object as uninitialized (as 1062 // in direct_allocated above) because this is being done with the 1063 // world stopped and the object will be initialized by the 1064 // time the sweeper gets to look at it. 1065 assert(SafepointSynchronize::is_at_safepoint(), 1066 "expect promotion only at safepoints"); 1067 1068 if (_collectorState < Sweeping) { 1069 // Mark the appropriate cards in the modUnionTable, so that 1070 // this object gets scanned before the sweep. If this is 1071 // not done, CMS generation references in the object might 1072 // not get marked. 1073 // For the case of arrays, which are otherwise precisely 1074 // marked, we need to dirty the entire array, not just its head. 1075 if (is_obj_array) { 1076 // The [par_]mark_range() method expects mr.end() below to 1077 // be aligned to the granularity of a bit's representation 1078 // in the heap. In the case of the MUT below, that's a 1079 // card size. 1080 MemRegion mr(start, 1081 (HeapWord*)round_to((intptr_t)(start + obj_size), 1082 CardTableModRefBS::card_size /* bytes */)); 1083 if (par) { 1084 _modUnionTable.par_mark_range(mr); 1085 } else { 1086 _modUnionTable.mark_range(mr); 1087 } 1088 } else { // not an obj array; we can just mark the head 1089 if (par) { 1090 _modUnionTable.par_mark(start); 1091 } else { 1092 _modUnionTable.mark(start); 1093 } 1094 } 1095 } 1096 } 1097 } 1098 1099 static inline size_t percent_of_space(Space* space, HeapWord* addr) 1100 { 1101 size_t delta = pointer_delta(addr, space->bottom()); 1102 return (size_t)(delta * 100.0 / (space->capacity() / HeapWordSize)); 1103 } 1104 1105 void CMSCollector::icms_update_allocation_limits() 1106 { 1107 Generation* gen0 = GenCollectedHeap::heap()->get_gen(0); 1108 EdenSpace* eden = gen0->as_DefNewGeneration()->eden(); 1109 1110 const unsigned int duty_cycle = stats().icms_update_duty_cycle(); 1111 if (CMSTraceIncrementalPacing) { 1112 stats().print(); 1113 } 1114 1115 assert(duty_cycle <= 100, "invalid duty cycle"); 1116 if (duty_cycle != 0) { 1117 // The duty_cycle is a percentage between 0 and 100; convert to words and 1118 // then compute the offset from the endpoints of the space. 1119 size_t free_words = eden->free() / HeapWordSize; 1120 double free_words_dbl = (double)free_words; 1121 size_t duty_cycle_words = (size_t)(free_words_dbl * duty_cycle / 100.0); 1122 size_t offset_words = (free_words - duty_cycle_words) / 2; 1123 1124 _icms_start_limit = eden->top() + offset_words; 1125 _icms_stop_limit = eden->end() - offset_words; 1126 1127 // The limits may be adjusted (shifted to the right) by 1128 // CMSIncrementalOffset, to allow the application more mutator time after a 1129 // young gen gc (when all mutators were stopped) and before CMS starts and 1130 // takes away one or more cpus. 1131 if (CMSIncrementalOffset != 0) { 1132 double adjustment_dbl = free_words_dbl * CMSIncrementalOffset / 100.0; 1133 size_t adjustment = (size_t)adjustment_dbl; 1134 HeapWord* tmp_stop = _icms_stop_limit + adjustment; 1135 if (tmp_stop > _icms_stop_limit && tmp_stop < eden->end()) { 1136 _icms_start_limit += adjustment; 1137 _icms_stop_limit = tmp_stop; 1138 } 1139 } 1140 } 1141 if (duty_cycle == 0 || (_icms_start_limit == _icms_stop_limit)) { 1142 _icms_start_limit = _icms_stop_limit = eden->end(); 1143 } 1144 1145 // Install the new start limit. 1146 eden->set_soft_end(_icms_start_limit); 1147 1148 if (CMSTraceIncrementalMode) { 1149 gclog_or_tty->print(" icms alloc limits: " 1150 PTR_FORMAT "," PTR_FORMAT 1151 " (" SIZE_FORMAT "%%," SIZE_FORMAT "%%) ", 1152 _icms_start_limit, _icms_stop_limit, 1153 percent_of_space(eden, _icms_start_limit), 1154 percent_of_space(eden, _icms_stop_limit)); 1155 if (Verbose) { 1156 gclog_or_tty->print("eden: "); 1157 eden->print_on(gclog_or_tty); 1158 } 1159 } 1160 } 1161 1162 // Any changes here should try to maintain the invariant 1163 // that if this method is called with _icms_start_limit 1164 // and _icms_stop_limit both NULL, then it should return NULL 1165 // and not notify the icms thread. 1166 HeapWord* 1167 CMSCollector::allocation_limit_reached(Space* space, HeapWord* top, 1168 size_t word_size) 1169 { 1170 // A start_limit equal to end() means the duty cycle is 0, so treat that as a 1171 // nop. 1172 if (CMSIncrementalMode && _icms_start_limit != space->end()) { 1173 if (top <= _icms_start_limit) { 1174 if (CMSTraceIncrementalMode) { 1175 space->print_on(gclog_or_tty); 1176 gclog_or_tty->stamp(); 1177 gclog_or_tty->print_cr(" start limit top=" PTR_FORMAT 1178 ", new limit=" PTR_FORMAT 1179 " (" SIZE_FORMAT "%%)", 1180 top, _icms_stop_limit, 1181 percent_of_space(space, _icms_stop_limit)); 1182 } 1183 ConcurrentMarkSweepThread::start_icms(); 1184 assert(top < _icms_stop_limit, "Tautology"); 1185 if (word_size < pointer_delta(_icms_stop_limit, top)) { 1186 return _icms_stop_limit; 1187 } 1188 1189 // The allocation will cross both the _start and _stop limits, so do the 1190 // stop notification also and return end(). 1191 if (CMSTraceIncrementalMode) { 1192 space->print_on(gclog_or_tty); 1193 gclog_or_tty->stamp(); 1194 gclog_or_tty->print_cr(" +stop limit top=" PTR_FORMAT 1195 ", new limit=" PTR_FORMAT 1196 " (" SIZE_FORMAT "%%)", 1197 top, space->end(), 1198 percent_of_space(space, space->end())); 1199 } 1200 ConcurrentMarkSweepThread::stop_icms(); 1201 return space->end(); 1202 } 1203 1204 if (top <= _icms_stop_limit) { 1205 if (CMSTraceIncrementalMode) { 1206 space->print_on(gclog_or_tty); 1207 gclog_or_tty->stamp(); 1208 gclog_or_tty->print_cr(" stop limit top=" PTR_FORMAT 1209 ", new limit=" PTR_FORMAT 1210 " (" SIZE_FORMAT "%%)", 1211 top, space->end(), 1212 percent_of_space(space, space->end())); 1213 } 1214 ConcurrentMarkSweepThread::stop_icms(); 1215 return space->end(); 1216 } 1217 1218 if (CMSTraceIncrementalMode) { 1219 space->print_on(gclog_or_tty); 1220 gclog_or_tty->stamp(); 1221 gclog_or_tty->print_cr(" end limit top=" PTR_FORMAT 1222 ", new limit=" PTR_FORMAT, 1223 top, NULL); 1224 } 1225 } 1226 1227 return NULL; 1228 } 1229 1230 oop ConcurrentMarkSweepGeneration::promote(oop obj, size_t obj_size, oop* ref) { 1231 assert(obj_size == (size_t)obj->size(), "bad obj_size passed in"); 1232 // allocate, copy and if necessary update promoinfo -- 1233 // delegate to underlying space. 1234 assert_lock_strong(freelistLock()); 1235 1236 #ifndef PRODUCT 1237 if (Universe::heap()->promotion_should_fail()) { 1238 return NULL; 1239 } 1240 #endif // #ifndef PRODUCT 1241 1242 oop res = _cmsSpace->promote(obj, obj_size, ref); 1243 if (res == NULL) { 1244 // expand and retry 1245 size_t s = _cmsSpace->expansionSpaceRequired(obj_size); // HeapWords 1246 expand(s*HeapWordSize, MinHeapDeltaBytes, 1247 CMSExpansionCause::_satisfy_promotion); 1248 // Since there's currently no next generation, we don't try to promote 1249 // into a more senior generation. 1250 assert(next_gen() == NULL, "assumption, based upon which no attempt " 1251 "is made to pass on a possibly failing " 1252 "promotion to next generation"); 1253 res = _cmsSpace->promote(obj, obj_size, ref); 1254 } 1255 if (res != NULL) { 1256 // See comment in allocate() about when objects should 1257 // be allocated live. 1258 assert(obj->is_oop(), "Will dereference klass pointer below"); 1259 collector()->promoted(false, // Not parallel 1260 (HeapWord*)res, obj->is_objArray(), obj_size); 1261 // promotion counters 1262 NOT_PRODUCT( 1263 _numObjectsPromoted++; 1264 _numWordsPromoted += 1265 (int)(CompactibleFreeListSpace::adjustObjectSize(obj->size())); 1266 ) 1267 } 1268 return res; 1269 } 1270 1271 1272 HeapWord* 1273 ConcurrentMarkSweepGeneration::allocation_limit_reached(Space* space, 1274 HeapWord* top, 1275 size_t word_sz) 1276 { 1277 return collector()->allocation_limit_reached(space, top, word_sz); 1278 } 1279 1280 // Things to support parallel young-gen collection. 1281 oop 1282 ConcurrentMarkSweepGeneration::par_promote(int thread_num, 1283 oop old, markOop m, 1284 size_t word_sz) { 1285 #ifndef PRODUCT 1286 if (Universe::heap()->promotion_should_fail()) { 1287 return NULL; 1288 } 1289 #endif // #ifndef PRODUCT 1290 1291 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1292 PromotionInfo* promoInfo = &ps->promo; 1293 // if we are tracking promotions, then first ensure space for 1294 // promotion (including spooling space for saving header if necessary). 1295 // then allocate and copy, then track promoted info if needed. 1296 // When tracking (see PromotionInfo::track()), the mark word may 1297 // be displaced and in this case restoration of the mark word 1298 // occurs in the (oop_since_save_marks_)iterate phase. 1299 if (promoInfo->tracking() && !promoInfo->ensure_spooling_space()) { 1300 // Out of space for allocating spooling buffers; 1301 // try expanding and allocating spooling buffers. 1302 if (!expand_and_ensure_spooling_space(promoInfo)) { 1303 return NULL; 1304 } 1305 } 1306 assert(promoInfo->has_spooling_space(), "Control point invariant"); 1307 HeapWord* obj_ptr = ps->lab.alloc(word_sz); 1308 if (obj_ptr == NULL) { 1309 obj_ptr = expand_and_par_lab_allocate(ps, word_sz); 1310 if (obj_ptr == NULL) { 1311 return NULL; 1312 } 1313 } 1314 oop obj = oop(obj_ptr); 1315 assert(obj->klass() == NULL, "Object should be uninitialized here."); 1316 // Otherwise, copy the object. Here we must be careful to insert the 1317 // klass pointer last, since this marks the block as an allocated object. 1318 HeapWord* old_ptr = (HeapWord*)old; 1319 if (word_sz > (size_t)oopDesc::header_size()) { 1320 Copy::aligned_disjoint_words(old_ptr + oopDesc::header_size(), 1321 obj_ptr + oopDesc::header_size(), 1322 word_sz - oopDesc::header_size()); 1323 } 1324 // Restore the mark word copied above. 1325 obj->set_mark(m); 1326 // Now we can track the promoted object, if necessary. We take care 1327 // To delay the transition from uninitialized to full object 1328 // (i.e., insertion of klass pointer) until after, so that it 1329 // atomically becomes a promoted object. 1330 if (promoInfo->tracking()) { 1331 promoInfo->track((PromotedObject*)obj, old->klass()); 1332 } 1333 // Finally, install the klass pointer. 1334 obj->set_klass(old->klass()); 1335 1336 assert(old->is_oop(), "Will dereference klass ptr below"); 1337 collector()->promoted(true, // parallel 1338 obj_ptr, old->is_objArray(), word_sz); 1339 1340 NOT_PRODUCT( 1341 Atomic::inc(&_numObjectsPromoted); 1342 Atomic::add((jint)CompactibleFreeListSpace::adjustObjectSize(obj->size()), 1343 &_numWordsPromoted); 1344 ) 1345 1346 return obj; 1347 } 1348 1349 void 1350 ConcurrentMarkSweepGeneration:: 1351 par_promote_alloc_undo(int thread_num, 1352 HeapWord* obj, size_t word_sz) { 1353 // CMS does not support promotion undo. 1354 ShouldNotReachHere(); 1355 } 1356 1357 void 1358 ConcurrentMarkSweepGeneration:: 1359 par_promote_alloc_done(int thread_num) { 1360 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1361 ps->lab.retire(); 1362 #if CFLS_LAB_REFILL_STATS 1363 if (thread_num == 0) { 1364 _cmsSpace->print_par_alloc_stats(); 1365 } 1366 #endif 1367 } 1368 1369 void 1370 ConcurrentMarkSweepGeneration:: 1371 par_oop_since_save_marks_iterate_done(int thread_num) { 1372 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1373 ParScanWithoutBarrierClosure* dummy_cl = NULL; 1374 ps->promo.promoted_oops_iterate_nv(dummy_cl); 1375 } 1376 1377 // XXXPERM 1378 bool ConcurrentMarkSweepGeneration::should_collect(bool full, 1379 size_t size, 1380 bool tlab) 1381 { 1382 // We allow a STW collection only if a full 1383 // collection was requested. 1384 return full || should_allocate(size, tlab); // FIX ME !!! 1385 // This and promotion failure handling are connected at the 1386 // hip and should be fixed by untying them. 1387 } 1388 1389 bool CMSCollector::shouldConcurrentCollect() { 1390 if (_full_gc_requested) { 1391 assert(ExplicitGCInvokesConcurrent, "Unexpected state"); 1392 if (Verbose && PrintGCDetails) { 1393 gclog_or_tty->print_cr("CMSCollector: collect because of explicit " 1394 " gc request"); 1395 } 1396 return true; 1397 } 1398 1399 // For debugging purposes, change the type of collection. 1400 // If the rotation is not on the concurrent collection 1401 // type, don't start a concurrent collection. 1402 NOT_PRODUCT( 1403 if (RotateCMSCollectionTypes && 1404 (_cmsGen->debug_collection_type() != 1405 ConcurrentMarkSweepGeneration::Concurrent_collection_type)) { 1406 assert(_cmsGen->debug_collection_type() != 1407 ConcurrentMarkSweepGeneration::Unknown_collection_type, 1408 "Bad cms collection type"); 1409 return false; 1410 } 1411 ) 1412 1413 FreelistLocker x(this); 1414 // ------------------------------------------------------------------ 1415 // Print out lots of information which affects the initiation of 1416 // a collection. 1417 if (PrintCMSInitiationStatistics && stats().valid()) { 1418 gclog_or_tty->print("CMSCollector shouldConcurrentCollect: "); 1419 gclog_or_tty->stamp(); 1420 gclog_or_tty->print_cr(""); 1421 stats().print_on(gclog_or_tty); 1422 gclog_or_tty->print_cr("time_until_cms_gen_full %3.7f", 1423 stats().time_until_cms_gen_full()); 1424 gclog_or_tty->print_cr("free="SIZE_FORMAT, _cmsGen->free()); 1425 gclog_or_tty->print_cr("contiguous_available="SIZE_FORMAT, 1426 _cmsGen->contiguous_available()); 1427 gclog_or_tty->print_cr("promotion_rate=%g", stats().promotion_rate()); 1428 gclog_or_tty->print_cr("cms_allocation_rate=%g", stats().cms_allocation_rate()); 1429 gclog_or_tty->print_cr("occupancy=%3.7f", _cmsGen->occupancy()); 1430 gclog_or_tty->print_cr("initiatingOccupancy=%3.7f", _cmsGen->initiating_occupancy()); 1431 gclog_or_tty->print_cr("initiatingPermOccupancy=%3.7f", _permGen->initiating_occupancy()); 1432 } 1433 // ------------------------------------------------------------------ 1434 1435 // If the estimated time to complete a cms collection (cms_duration()) 1436 // is less than the estimated time remaining until the cms generation 1437 // is full, start a collection. 1438 if (!UseCMSInitiatingOccupancyOnly) { 1439 if (stats().valid()) { 1440 if (stats().time_until_cms_start() == 0.0) { 1441 return true; 1442 } 1443 } else { 1444 // We want to conservatively collect somewhat early in order 1445 // to try and "bootstrap" our CMS/promotion statistics; 1446 // this branch will not fire after the first successful CMS 1447 // collection because the stats should then be valid. 1448 if (_cmsGen->occupancy() >= _bootstrap_occupancy) { 1449 if (Verbose && PrintGCDetails) { 1450 gclog_or_tty->print_cr( 1451 " CMSCollector: collect for bootstrapping statistics:" 1452 " occupancy = %f, boot occupancy = %f", _cmsGen->occupancy(), 1453 _bootstrap_occupancy); 1454 } 1455 return true; 1456 } 1457 } 1458 } 1459 1460 // Otherwise, we start a collection cycle if either the perm gen or 1461 // old gen want a collection cycle started. Each may use 1462 // an appropriate criterion for making this decision. 1463 // XXX We need to make sure that the gen expansion 1464 // criterion dovetails well with this. XXX NEED TO FIX THIS 1465 if (_cmsGen->should_concurrent_collect()) { 1466 if (Verbose && PrintGCDetails) { 1467 gclog_or_tty->print_cr("CMS old gen initiated"); 1468 } 1469 return true; 1470 } 1471 1472 // We start a collection if we believe an incremental collection may fail; 1473 // this is not likely to be productive in practice because it's probably too 1474 // late anyway. 1475 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1476 assert(gch->collector_policy()->is_two_generation_policy(), 1477 "You may want to check the correctness of the following"); 1478 if (gch->incremental_collection_will_fail()) { 1479 if (PrintGCDetails && Verbose) { 1480 gclog_or_tty->print("CMSCollector: collect because incremental collection will fail "); 1481 } 1482 return true; 1483 } 1484 1485 if (CMSClassUnloadingEnabled && _permGen->should_concurrent_collect()) { 1486 bool res = update_should_unload_classes(); 1487 if (res) { 1488 if (Verbose && PrintGCDetails) { 1489 gclog_or_tty->print_cr("CMS perm gen initiated"); 1490 } 1491 return true; 1492 } 1493 } 1494 return false; 1495 } 1496 1497 // Clear _expansion_cause fields of constituent generations 1498 void CMSCollector::clear_expansion_cause() { 1499 _cmsGen->clear_expansion_cause(); 1500 _permGen->clear_expansion_cause(); 1501 } 1502 1503 // We should be conservative in starting a collection cycle. To 1504 // start too eagerly runs the risk of collecting too often in the 1505 // extreme. To collect too rarely falls back on full collections, 1506 // which works, even if not optimum in terms of concurrent work. 1507 // As a work around for too eagerly collecting, use the flag 1508 // UseCMSInitiatingOccupancyOnly. This also has the advantage of 1509 // giving the user an easily understandable way of controlling the 1510 // collections. 1511 // We want to start a new collection cycle if any of the following 1512 // conditions hold: 1513 // . our current occupancy exceeds the configured initiating occupancy 1514 // for this generation, or 1515 // . we recently needed to expand this space and have not, since that 1516 // expansion, done a collection of this generation, or 1517 // . the underlying space believes that it may be a good idea to initiate 1518 // a concurrent collection (this may be based on criteria such as the 1519 // following: the space uses linear allocation and linear allocation is 1520 // going to fail, or there is believed to be excessive fragmentation in 1521 // the generation, etc... or ... 1522 // [.(currently done by CMSCollector::shouldConcurrentCollect() only for 1523 // the case of the old generation, not the perm generation; see CR 6543076): 1524 // we may be approaching a point at which allocation requests may fail because 1525 // we will be out of sufficient free space given allocation rate estimates.] 1526 bool ConcurrentMarkSweepGeneration::should_concurrent_collect() const { 1527 1528 assert_lock_strong(freelistLock()); 1529 if (occupancy() > initiating_occupancy()) { 1530 if (PrintGCDetails && Verbose) { 1531 gclog_or_tty->print(" %s: collect because of occupancy %f / %f ", 1532 short_name(), occupancy(), initiating_occupancy()); 1533 } 1534 return true; 1535 } 1536 if (UseCMSInitiatingOccupancyOnly) { 1537 return false; 1538 } 1539 if (expansion_cause() == CMSExpansionCause::_satisfy_allocation) { 1540 if (PrintGCDetails && Verbose) { 1541 gclog_or_tty->print(" %s: collect because expanded for allocation ", 1542 short_name()); 1543 } 1544 return true; 1545 } 1546 if (_cmsSpace->should_concurrent_collect()) { 1547 if (PrintGCDetails && Verbose) { 1548 gclog_or_tty->print(" %s: collect because cmsSpace says so ", 1549 short_name()); 1550 } 1551 return true; 1552 } 1553 return false; 1554 } 1555 1556 void ConcurrentMarkSweepGeneration::collect(bool full, 1557 bool clear_all_soft_refs, 1558 size_t size, 1559 bool tlab) 1560 { 1561 collector()->collect(full, clear_all_soft_refs, size, tlab); 1562 } 1563 1564 void CMSCollector::collect(bool full, 1565 bool clear_all_soft_refs, 1566 size_t size, 1567 bool tlab) 1568 { 1569 if (!UseCMSCollectionPassing && _collectorState > Idling) { 1570 // For debugging purposes skip the collection if the state 1571 // is not currently idle 1572 if (TraceCMSState) { 1573 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " skipped full:%d CMS state %d", 1574 Thread::current(), full, _collectorState); 1575 } 1576 return; 1577 } 1578 1579 // The following "if" branch is present for defensive reasons. 1580 // In the current uses of this interface, it can be replaced with: 1581 // assert(!GC_locker.is_active(), "Can't be called otherwise"); 1582 // But I am not placing that assert here to allow future 1583 // generality in invoking this interface. 1584 if (GC_locker::is_active()) { 1585 // A consistency test for GC_locker 1586 assert(GC_locker::needs_gc(), "Should have been set already"); 1587 // Skip this foreground collection, instead 1588 // expanding the heap if necessary. 1589 // Need the free list locks for the call to free() in compute_new_size() 1590 compute_new_size(); 1591 return; 1592 } 1593 acquire_control_and_collect(full, clear_all_soft_refs); 1594 _full_gcs_since_conc_gc++; 1595 1596 } 1597 1598 void CMSCollector::request_full_gc(unsigned int full_gc_count) { 1599 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1600 unsigned int gc_count = gch->total_full_collections(); 1601 if (gc_count == full_gc_count) { 1602 MutexLockerEx y(CGC_lock, Mutex::_no_safepoint_check_flag); 1603 _full_gc_requested = true; 1604 CGC_lock->notify(); // nudge CMS thread 1605 } 1606 } 1607 1608 1609 // The foreground and background collectors need to coordinate in order 1610 // to make sure that they do not mutually interfere with CMS collections. 1611 // When a background collection is active, 1612 // the foreground collector may need to take over (preempt) and 1613 // synchronously complete an ongoing collection. Depending on the 1614 // frequency of the background collections and the heap usage 1615 // of the application, this preemption can be seldom or frequent. 1616 // There are only certain 1617 // points in the background collection that the "collection-baton" 1618 // can be passed to the foreground collector. 1619 // 1620 // The foreground collector will wait for the baton before 1621 // starting any part of the collection. The foreground collector 1622 // will only wait at one location. 1623 // 1624 // The background collector will yield the baton before starting a new 1625 // phase of the collection (e.g., before initial marking, marking from roots, 1626 // precleaning, final re-mark, sweep etc.) This is normally done at the head 1627 // of the loop which switches the phases. The background collector does some 1628 // of the phases (initial mark, final re-mark) with the world stopped. 1629 // Because of locking involved in stopping the world, 1630 // the foreground collector should not block waiting for the background 1631 // collector when it is doing a stop-the-world phase. The background 1632 // collector will yield the baton at an additional point just before 1633 // it enters a stop-the-world phase. Once the world is stopped, the 1634 // background collector checks the phase of the collection. If the 1635 // phase has not changed, it proceeds with the collection. If the 1636 // phase has changed, it skips that phase of the collection. See 1637 // the comments on the use of the Heap_lock in collect_in_background(). 1638 // 1639 // Variable used in baton passing. 1640 // _foregroundGCIsActive - Set to true by the foreground collector when 1641 // it wants the baton. The foreground clears it when it has finished 1642 // the collection. 1643 // _foregroundGCShouldWait - Set to true by the background collector 1644 // when it is running. The foreground collector waits while 1645 // _foregroundGCShouldWait is true. 1646 // CGC_lock - monitor used to protect access to the above variables 1647 // and to notify the foreground and background collectors. 1648 // _collectorState - current state of the CMS collection. 1649 // 1650 // The foreground collector 1651 // acquires the CGC_lock 1652 // sets _foregroundGCIsActive 1653 // waits on the CGC_lock for _foregroundGCShouldWait to be false 1654 // various locks acquired in preparation for the collection 1655 // are released so as not to block the background collector 1656 // that is in the midst of a collection 1657 // proceeds with the collection 1658 // clears _foregroundGCIsActive 1659 // returns 1660 // 1661 // The background collector in a loop iterating on the phases of the 1662 // collection 1663 // acquires the CGC_lock 1664 // sets _foregroundGCShouldWait 1665 // if _foregroundGCIsActive is set 1666 // clears _foregroundGCShouldWait, notifies _CGC_lock 1667 // waits on _CGC_lock for _foregroundGCIsActive to become false 1668 // and exits the loop. 1669 // otherwise 1670 // proceed with that phase of the collection 1671 // if the phase is a stop-the-world phase, 1672 // yield the baton once more just before enqueueing 1673 // the stop-world CMS operation (executed by the VM thread). 1674 // returns after all phases of the collection are done 1675 // 1676 1677 void CMSCollector::acquire_control_and_collect(bool full, 1678 bool clear_all_soft_refs) { 1679 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint"); 1680 assert(!Thread::current()->is_ConcurrentGC_thread(), 1681 "shouldn't try to acquire control from self!"); 1682 1683 // Start the protocol for acquiring control of the 1684 // collection from the background collector (aka CMS thread). 1685 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 1686 "VM thread should have CMS token"); 1687 // Remember the possibly interrupted state of an ongoing 1688 // concurrent collection 1689 CollectorState first_state = _collectorState; 1690 1691 // Signal to a possibly ongoing concurrent collection that 1692 // we want to do a foreground collection. 1693 _foregroundGCIsActive = true; 1694 1695 // Disable incremental mode during a foreground collection. 1696 ICMSDisabler icms_disabler; 1697 1698 // release locks and wait for a notify from the background collector 1699 // releasing the locks in only necessary for phases which 1700 // do yields to improve the granularity of the collection. 1701 assert_lock_strong(bitMapLock()); 1702 // We need to lock the Free list lock for the space that we are 1703 // currently collecting. 1704 assert(haveFreelistLocks(), "Must be holding free list locks"); 1705 bitMapLock()->unlock(); 1706 releaseFreelistLocks(); 1707 { 1708 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1709 if (_foregroundGCShouldWait) { 1710 // We are going to be waiting for action for the CMS thread; 1711 // it had better not be gone (for instance at shutdown)! 1712 assert(ConcurrentMarkSweepThread::cmst() != NULL, 1713 "CMS thread must be running"); 1714 // Wait here until the background collector gives us the go-ahead 1715 ConcurrentMarkSweepThread::clear_CMS_flag( 1716 ConcurrentMarkSweepThread::CMS_vm_has_token); // release token 1717 // Get a possibly blocked CMS thread going: 1718 // Note that we set _foregroundGCIsActive true above, 1719 // without protection of the CGC_lock. 1720 CGC_lock->notify(); 1721 assert(!ConcurrentMarkSweepThread::vm_thread_wants_cms_token(), 1722 "Possible deadlock"); 1723 while (_foregroundGCShouldWait) { 1724 // wait for notification 1725 CGC_lock->wait(Mutex::_no_safepoint_check_flag); 1726 // Possibility of delay/starvation here, since CMS token does 1727 // not know to give priority to VM thread? Actually, i think 1728 // there wouldn't be any delay/starvation, but the proof of 1729 // that "fact" (?) appears non-trivial. XXX 20011219YSR 1730 } 1731 ConcurrentMarkSweepThread::set_CMS_flag( 1732 ConcurrentMarkSweepThread::CMS_vm_has_token); 1733 } 1734 } 1735 // The CMS_token is already held. Get back the other locks. 1736 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 1737 "VM thread should have CMS token"); 1738 getFreelistLocks(); 1739 bitMapLock()->lock_without_safepoint_check(); 1740 if (TraceCMSState) { 1741 gclog_or_tty->print_cr("CMS foreground collector has asked for control " 1742 INTPTR_FORMAT " with first state %d", Thread::current(), first_state); 1743 gclog_or_tty->print_cr(" gets control with state %d", _collectorState); 1744 } 1745 1746 // Check if we need to do a compaction, or if not, whether 1747 // we need to start the mark-sweep from scratch. 1748 bool should_compact = false; 1749 bool should_start_over = false; 1750 decide_foreground_collection_type(clear_all_soft_refs, 1751 &should_compact, &should_start_over); 1752 1753 NOT_PRODUCT( 1754 if (RotateCMSCollectionTypes) { 1755 if (_cmsGen->debug_collection_type() == 1756 ConcurrentMarkSweepGeneration::MSC_foreground_collection_type) { 1757 should_compact = true; 1758 } else if (_cmsGen->debug_collection_type() == 1759 ConcurrentMarkSweepGeneration::MS_foreground_collection_type) { 1760 should_compact = false; 1761 } 1762 } 1763 ) 1764 1765 if (PrintGCDetails && first_state > Idling) { 1766 GCCause::Cause cause = GenCollectedHeap::heap()->gc_cause(); 1767 if (GCCause::is_user_requested_gc(cause) || 1768 GCCause::is_serviceability_requested_gc(cause)) { 1769 gclog_or_tty->print(" (concurrent mode interrupted)"); 1770 } else { 1771 gclog_or_tty->print(" (concurrent mode failure)"); 1772 } 1773 } 1774 1775 if (should_compact) { 1776 // If the collection is being acquired from the background 1777 // collector, there may be references on the discovered 1778 // references lists that have NULL referents (being those 1779 // that were concurrently cleared by a mutator) or 1780 // that are no longer active (having been enqueued concurrently 1781 // by the mutator). 1782 // Scrub the list of those references because Mark-Sweep-Compact 1783 // code assumes referents are not NULL and that all discovered 1784 // Reference objects are active. 1785 ref_processor()->clean_up_discovered_references(); 1786 1787 do_compaction_work(clear_all_soft_refs); 1788 1789 // Has the GC time limit been exceeded? 1790 check_gc_time_limit(); 1791 1792 } else { 1793 do_mark_sweep_work(clear_all_soft_refs, first_state, 1794 should_start_over); 1795 } 1796 // Reset the expansion cause, now that we just completed 1797 // a collection cycle. 1798 clear_expansion_cause(); 1799 _foregroundGCIsActive = false; 1800 return; 1801 } 1802 1803 void CMSCollector::check_gc_time_limit() { 1804 1805 // Ignore explicit GC's. Exiting here does not set the flag and 1806 // does not reset the count. Updating of the averages for system 1807 // GC's is still controlled by UseAdaptiveSizePolicyWithSystemGC. 1808 GCCause::Cause gc_cause = GenCollectedHeap::heap()->gc_cause(); 1809 if (GCCause::is_user_requested_gc(gc_cause) || 1810 GCCause::is_serviceability_requested_gc(gc_cause)) { 1811 return; 1812 } 1813 1814 // Calculate the fraction of the CMS generation was freed during 1815 // the last collection. 1816 // Only consider the STW compacting cost for now. 1817 // 1818 // Note that the gc time limit test only works for the collections 1819 // of the young gen + tenured gen and not for collections of the 1820 // permanent gen. That is because the calculation of the space 1821 // freed by the collection is the free space in the young gen + 1822 // tenured gen. 1823 1824 double fraction_free = 1825 ((double)_cmsGen->free())/((double)_cmsGen->max_capacity()); 1826 if ((100.0 * size_policy()->compacting_gc_cost()) > 1827 ((double) GCTimeLimit) && 1828 ((fraction_free * 100) < GCHeapFreeLimit)) { 1829 size_policy()->inc_gc_time_limit_count(); 1830 if (UseGCOverheadLimit && 1831 (size_policy()->gc_time_limit_count() > 1832 AdaptiveSizePolicyGCTimeLimitThreshold)) { 1833 size_policy()->set_gc_time_limit_exceeded(true); 1834 // Avoid consecutive OOM due to the gc time limit by resetting 1835 // the counter. 1836 size_policy()->reset_gc_time_limit_count(); 1837 if (PrintGCDetails) { 1838 gclog_or_tty->print_cr(" GC is exceeding overhead limit " 1839 "of %d%%", GCTimeLimit); 1840 } 1841 } else { 1842 if (PrintGCDetails) { 1843 gclog_or_tty->print_cr(" GC would exceed overhead limit " 1844 "of %d%%", GCTimeLimit); 1845 } 1846 } 1847 } else { 1848 size_policy()->reset_gc_time_limit_count(); 1849 } 1850 } 1851 1852 // Resize the perm generation and the tenured generation 1853 // after obtaining the free list locks for the 1854 // two generations. 1855 void CMSCollector::compute_new_size() { 1856 assert_locked_or_safepoint(Heap_lock); 1857 FreelistLocker z(this); 1858 _permGen->compute_new_size(); 1859 _cmsGen->compute_new_size(); 1860 } 1861 1862 // A work method used by foreground collection to determine 1863 // what type of collection (compacting or not, continuing or fresh) 1864 // it should do. 1865 // NOTE: the intent is to make UseCMSCompactAtFullCollection 1866 // and CMSCompactWhenClearAllSoftRefs the default in the future 1867 // and do away with the flags after a suitable period. 1868 void CMSCollector::decide_foreground_collection_type( 1869 bool clear_all_soft_refs, bool* should_compact, 1870 bool* should_start_over) { 1871 // Normally, we'll compact only if the UseCMSCompactAtFullCollection 1872 // flag is set, and we have either requested a System.gc() or 1873 // the number of full gc's since the last concurrent cycle 1874 // has exceeded the threshold set by CMSFullGCsBeforeCompaction, 1875 // or if an incremental collection has failed 1876 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1877 assert(gch->collector_policy()->is_two_generation_policy(), 1878 "You may want to check the correctness of the following"); 1879 // Inform cms gen if this was due to partial collection failing. 1880 // The CMS gen may use this fact to determine its expansion policy. 1881 if (gch->incremental_collection_will_fail()) { 1882 assert(!_cmsGen->incremental_collection_failed(), 1883 "Should have been noticed, reacted to and cleared"); 1884 _cmsGen->set_incremental_collection_failed(); 1885 } 1886 *should_compact = 1887 UseCMSCompactAtFullCollection && 1888 ((_full_gcs_since_conc_gc >= CMSFullGCsBeforeCompaction) || 1889 GCCause::is_user_requested_gc(gch->gc_cause()) || 1890 gch->incremental_collection_will_fail()); 1891 *should_start_over = false; 1892 if (clear_all_soft_refs && !*should_compact) { 1893 // We are about to do a last ditch collection attempt 1894 // so it would normally make sense to do a compaction 1895 // to reclaim as much space as possible. 1896 if (CMSCompactWhenClearAllSoftRefs) { 1897 // Default: The rationale is that in this case either 1898 // we are past the final marking phase, in which case 1899 // we'd have to start over, or so little has been done 1900 // that there's little point in saving that work. Compaction 1901 // appears to be the sensible choice in either case. 1902 *should_compact = true; 1903 } else { 1904 // We have been asked to clear all soft refs, but not to 1905 // compact. Make sure that we aren't past the final checkpoint 1906 // phase, for that is where we process soft refs. If we are already 1907 // past that phase, we'll need to redo the refs discovery phase and 1908 // if necessary clear soft refs that weren't previously 1909 // cleared. We do so by remembering the phase in which 1910 // we came in, and if we are past the refs processing 1911 // phase, we'll choose to just redo the mark-sweep 1912 // collection from scratch. 1913 if (_collectorState > FinalMarking) { 1914 // We are past the refs processing phase; 1915 // start over and do a fresh synchronous CMS cycle 1916 _collectorState = Resetting; // skip to reset to start new cycle 1917 reset(false /* == !asynch */); 1918 *should_start_over = true; 1919 } // else we can continue a possibly ongoing current cycle 1920 } 1921 } 1922 } 1923 1924 // A work method used by the foreground collector to do 1925 // a mark-sweep-compact. 1926 void CMSCollector::do_compaction_work(bool clear_all_soft_refs) { 1927 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1928 TraceTime t("CMS:MSC ", PrintGCDetails && Verbose, true, gclog_or_tty); 1929 if (PrintGC && Verbose && !(GCCause::is_user_requested_gc(gch->gc_cause()))) { 1930 gclog_or_tty->print_cr("Compact ConcurrentMarkSweepGeneration after %d " 1931 "collections passed to foreground collector", _full_gcs_since_conc_gc); 1932 } 1933 1934 // Sample collection interval time and reset for collection pause. 1935 if (UseAdaptiveSizePolicy) { 1936 size_policy()->msc_collection_begin(); 1937 } 1938 1939 // Temporarily widen the span of the weak reference processing to 1940 // the entire heap. 1941 MemRegion new_span(GenCollectedHeap::heap()->reserved_region()); 1942 ReferenceProcessorSpanMutator x(ref_processor(), new_span); 1943 1944 // Temporarily, clear the "is_alive_non_header" field of the 1945 // reference processor. 1946 ReferenceProcessorIsAliveMutator y(ref_processor(), NULL); 1947 1948 // Temporarily make reference _processing_ single threaded (non-MT). 1949 ReferenceProcessorMTProcMutator z(ref_processor(), false); 1950 1951 // Temporarily make refs discovery atomic 1952 ReferenceProcessorAtomicMutator w(ref_processor(), true); 1953 1954 ref_processor()->set_enqueuing_is_done(false); 1955 ref_processor()->enable_discovery(); 1956 // If an asynchronous collection finishes, the _modUnionTable is 1957 // all clear. If we are assuming the collection from an asynchronous 1958 // collection, clear the _modUnionTable. 1959 assert(_collectorState != Idling || _modUnionTable.isAllClear(), 1960 "_modUnionTable should be clear if the baton was not passed"); 1961 _modUnionTable.clear_all(); 1962 1963 // We must adjust the allocation statistics being maintained 1964 // in the free list space. We do so by reading and clearing 1965 // the sweep timer and updating the block flux rate estimates below. 1966 assert(_sweep_timer.is_active(), "We should never see the timer inactive"); 1967 _sweep_timer.stop(); 1968 // Note that we do not use this sample to update the _sweep_estimate. 1969 _cmsGen->cmsSpace()->beginSweepFLCensus((float)(_sweep_timer.seconds()), 1970 _sweep_estimate.padded_average()); 1971 1972 GenMarkSweep::invoke_at_safepoint(_cmsGen->level(), 1973 ref_processor(), clear_all_soft_refs); 1974 #ifdef ASSERT 1975 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 1976 size_t free_size = cms_space->free(); 1977 assert(free_size == 1978 pointer_delta(cms_space->end(), cms_space->compaction_top()) 1979 * HeapWordSize, 1980 "All the free space should be compacted into one chunk at top"); 1981 assert(cms_space->dictionary()->totalChunkSize( 1982 debug_only(cms_space->freelistLock())) == 0 || 1983 cms_space->totalSizeInIndexedFreeLists() == 0, 1984 "All the free space should be in a single chunk"); 1985 size_t num = cms_space->totalCount(); 1986 assert((free_size == 0 && num == 0) || 1987 (free_size > 0 && (num == 1 || num == 2)), 1988 "There should be at most 2 free chunks after compaction"); 1989 #endif // ASSERT 1990 _collectorState = Resetting; 1991 assert(_restart_addr == NULL, 1992 "Should have been NULL'd before baton was passed"); 1993 reset(false /* == !asynch */); 1994 _cmsGen->reset_after_compaction(); 1995 _concurrent_cycles_since_last_unload = 0; 1996 1997 if (verifying() && !should_unload_classes()) { 1998 perm_gen_verify_bit_map()->clear_all(); 1999 } 2000 2001 // Clear any data recorded in the PLAB chunk arrays. 2002 if (_survivor_plab_array != NULL) { 2003 reset_survivor_plab_arrays(); 2004 } 2005 2006 // Adjust the per-size allocation stats for the next epoch. 2007 _cmsGen->cmsSpace()->endSweepFLCensus(sweep_count() /* fake */); 2008 // Restart the "sweep timer" for next epoch. 2009 _sweep_timer.reset(); 2010 _sweep_timer.start(); 2011 2012 // Sample collection pause time and reset for collection interval. 2013 if (UseAdaptiveSizePolicy) { 2014 size_policy()->msc_collection_end(gch->gc_cause()); 2015 } 2016 2017 // For a mark-sweep-compact, compute_new_size() will be called 2018 // in the heap's do_collection() method. 2019 } 2020 2021 // A work method used by the foreground collector to do 2022 // a mark-sweep, after taking over from a possibly on-going 2023 // concurrent mark-sweep collection. 2024 void CMSCollector::do_mark_sweep_work(bool clear_all_soft_refs, 2025 CollectorState first_state, bool should_start_over) { 2026 if (PrintGC && Verbose) { 2027 gclog_or_tty->print_cr("Pass concurrent collection to foreground " 2028 "collector with count %d", 2029 _full_gcs_since_conc_gc); 2030 } 2031 switch (_collectorState) { 2032 case Idling: 2033 if (first_state == Idling || should_start_over) { 2034 // The background GC was not active, or should 2035 // restarted from scratch; start the cycle. 2036 _collectorState = InitialMarking; 2037 } 2038 // If first_state was not Idling, then a background GC 2039 // was in progress and has now finished. No need to do it 2040 // again. Leave the state as Idling. 2041 break; 2042 case Precleaning: 2043 // In the foreground case don't do the precleaning since 2044 // it is not done concurrently and there is extra work 2045 // required. 2046 _collectorState = FinalMarking; 2047 } 2048 if (PrintGCDetails && 2049 (_collectorState > Idling || 2050 !GCCause::is_user_requested_gc(GenCollectedHeap::heap()->gc_cause()))) { 2051 gclog_or_tty->print(" (concurrent mode failure)"); 2052 } 2053 collect_in_foreground(clear_all_soft_refs); 2054 2055 // For a mark-sweep, compute_new_size() will be called 2056 // in the heap's do_collection() method. 2057 } 2058 2059 2060 void CMSCollector::getFreelistLocks() const { 2061 // Get locks for all free lists in all generations that this 2062 // collector is responsible for 2063 _cmsGen->freelistLock()->lock_without_safepoint_check(); 2064 _permGen->freelistLock()->lock_without_safepoint_check(); 2065 } 2066 2067 void CMSCollector::releaseFreelistLocks() const { 2068 // Release locks for all free lists in all generations that this 2069 // collector is responsible for 2070 _cmsGen->freelistLock()->unlock(); 2071 _permGen->freelistLock()->unlock(); 2072 } 2073 2074 bool CMSCollector::haveFreelistLocks() const { 2075 // Check locks for all free lists in all generations that this 2076 // collector is responsible for 2077 assert_lock_strong(_cmsGen->freelistLock()); 2078 assert_lock_strong(_permGen->freelistLock()); 2079 PRODUCT_ONLY(ShouldNotReachHere()); 2080 return true; 2081 } 2082 2083 // A utility class that is used by the CMS collector to 2084 // temporarily "release" the foreground collector from its 2085 // usual obligation to wait for the background collector to 2086 // complete an ongoing phase before proceeding. 2087 class ReleaseForegroundGC: public StackObj { 2088 private: 2089 CMSCollector* _c; 2090 public: 2091 ReleaseForegroundGC(CMSCollector* c) : _c(c) { 2092 assert(_c->_foregroundGCShouldWait, "Else should not need to call"); 2093 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 2094 // allow a potentially blocked foreground collector to proceed 2095 _c->_foregroundGCShouldWait = false; 2096 if (_c->_foregroundGCIsActive) { 2097 CGC_lock->notify(); 2098 } 2099 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 2100 "Possible deadlock"); 2101 } 2102 2103 ~ReleaseForegroundGC() { 2104 assert(!_c->_foregroundGCShouldWait, "Usage protocol violation?"); 2105 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 2106 _c->_foregroundGCShouldWait = true; 2107 } 2108 }; 2109 2110 // There are separate collect_in_background and collect_in_foreground because of 2111 // the different locking requirements of the background collector and the 2112 // foreground collector. There was originally an attempt to share 2113 // one "collect" method between the background collector and the foreground 2114 // collector but the if-then-else required made it cleaner to have 2115 // separate methods. 2116 void CMSCollector::collect_in_background(bool clear_all_soft_refs) { 2117 assert(Thread::current()->is_ConcurrentGC_thread(), 2118 "A CMS asynchronous collection is only allowed on a CMS thread."); 2119 2120 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2121 { 2122 bool safepoint_check = Mutex::_no_safepoint_check_flag; 2123 MutexLockerEx hl(Heap_lock, safepoint_check); 2124 FreelistLocker fll(this); 2125 MutexLockerEx x(CGC_lock, safepoint_check); 2126 if (_foregroundGCIsActive || !UseAsyncConcMarkSweepGC) { 2127 // The foreground collector is active or we're 2128 // not using asynchronous collections. Skip this 2129 // background collection. 2130 assert(!_foregroundGCShouldWait, "Should be clear"); 2131 return; 2132 } else { 2133 assert(_collectorState == Idling, "Should be idling before start."); 2134 _collectorState = InitialMarking; 2135 // Reset the expansion cause, now that we are about to begin 2136 // a new cycle. 2137 clear_expansion_cause(); 2138 } 2139 // Decide if we want to enable class unloading as part of the 2140 // ensuing concurrent GC cycle. 2141 update_should_unload_classes(); 2142 _full_gc_requested = false; // acks all outstanding full gc requests 2143 // Signal that we are about to start a collection 2144 gch->increment_total_full_collections(); // ... starting a collection cycle 2145 _collection_count_start = gch->total_full_collections(); 2146 } 2147 2148 // Used for PrintGC 2149 size_t prev_used; 2150 if (PrintGC && Verbose) { 2151 prev_used = _cmsGen->used(); // XXXPERM 2152 } 2153 2154 // The change of the collection state is normally done at this level; 2155 // the exceptions are phases that are executed while the world is 2156 // stopped. For those phases the change of state is done while the 2157 // world is stopped. For baton passing purposes this allows the 2158 // background collector to finish the phase and change state atomically. 2159 // The foreground collector cannot wait on a phase that is done 2160 // while the world is stopped because the foreground collector already 2161 // has the world stopped and would deadlock. 2162 while (_collectorState != Idling) { 2163 if (TraceCMSState) { 2164 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " in CMS state %d", 2165 Thread::current(), _collectorState); 2166 } 2167 // The foreground collector 2168 // holds the Heap_lock throughout its collection. 2169 // holds the CMS token (but not the lock) 2170 // except while it is waiting for the background collector to yield. 2171 // 2172 // The foreground collector should be blocked (not for long) 2173 // if the background collector is about to start a phase 2174 // executed with world stopped. If the background 2175 // collector has already started such a phase, the 2176 // foreground collector is blocked waiting for the 2177 // Heap_lock. The stop-world phases (InitialMarking and FinalMarking) 2178 // are executed in the VM thread. 2179 // 2180 // The locking order is 2181 // PendingListLock (PLL) -- if applicable (FinalMarking) 2182 // Heap_lock (both this & PLL locked in VM_CMS_Operation::prologue()) 2183 // CMS token (claimed in 2184 // stop_world_and_do() --> 2185 // safepoint_synchronize() --> 2186 // CMSThread::synchronize()) 2187 2188 { 2189 // Check if the FG collector wants us to yield. 2190 CMSTokenSync x(true); // is cms thread 2191 if (waitForForegroundGC()) { 2192 // We yielded to a foreground GC, nothing more to be 2193 // done this round. 2194 assert(_foregroundGCShouldWait == false, "We set it to false in " 2195 "waitForForegroundGC()"); 2196 if (TraceCMSState) { 2197 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT 2198 " exiting collection CMS state %d", 2199 Thread::current(), _collectorState); 2200 } 2201 return; 2202 } else { 2203 // The background collector can run but check to see if the 2204 // foreground collector has done a collection while the 2205 // background collector was waiting to get the CGC_lock 2206 // above. If yes, break so that _foregroundGCShouldWait 2207 // is cleared before returning. 2208 if (_collectorState == Idling) { 2209 break; 2210 } 2211 } 2212 } 2213 2214 assert(_foregroundGCShouldWait, "Foreground collector, if active, " 2215 "should be waiting"); 2216 2217 switch (_collectorState) { 2218 case InitialMarking: 2219 { 2220 ReleaseForegroundGC x(this); 2221 stats().record_cms_begin(); 2222 2223 VM_CMS_Initial_Mark initial_mark_op(this); 2224 VMThread::execute(&initial_mark_op); 2225 } 2226 // The collector state may be any legal state at this point 2227 // since the background collector may have yielded to the 2228 // foreground collector. 2229 break; 2230 case Marking: 2231 // initial marking in checkpointRootsInitialWork has been completed 2232 if (markFromRoots(true)) { // we were successful 2233 assert(_collectorState == Precleaning, "Collector state should " 2234 "have changed"); 2235 } else { 2236 assert(_foregroundGCIsActive, "Internal state inconsistency"); 2237 } 2238 break; 2239 case Precleaning: 2240 if (UseAdaptiveSizePolicy) { 2241 size_policy()->concurrent_precleaning_begin(); 2242 } 2243 // marking from roots in markFromRoots has been completed 2244 preclean(); 2245 if (UseAdaptiveSizePolicy) { 2246 size_policy()->concurrent_precleaning_end(); 2247 } 2248 assert(_collectorState == AbortablePreclean || 2249 _collectorState == FinalMarking, 2250 "Collector state should have changed"); 2251 break; 2252 case AbortablePreclean: 2253 if (UseAdaptiveSizePolicy) { 2254 size_policy()->concurrent_phases_resume(); 2255 } 2256 abortable_preclean(); 2257 if (UseAdaptiveSizePolicy) { 2258 size_policy()->concurrent_precleaning_end(); 2259 } 2260 assert(_collectorState == FinalMarking, "Collector state should " 2261 "have changed"); 2262 break; 2263 case FinalMarking: 2264 { 2265 ReleaseForegroundGC x(this); 2266 2267 VM_CMS_Final_Remark final_remark_op(this); 2268 VMThread::execute(&final_remark_op); 2269 } 2270 assert(_foregroundGCShouldWait, "block post-condition"); 2271 break; 2272 case Sweeping: 2273 if (UseAdaptiveSizePolicy) { 2274 size_policy()->concurrent_sweeping_begin(); 2275 } 2276 // final marking in checkpointRootsFinal has been completed 2277 sweep(true); 2278 assert(_collectorState == Resizing, "Collector state change " 2279 "to Resizing must be done under the free_list_lock"); 2280 _full_gcs_since_conc_gc = 0; 2281 2282 // Stop the timers for adaptive size policy for the concurrent phases 2283 if (UseAdaptiveSizePolicy) { 2284 size_policy()->concurrent_sweeping_end(); 2285 size_policy()->concurrent_phases_end(gch->gc_cause(), 2286 gch->prev_gen(_cmsGen)->capacity(), 2287 _cmsGen->free()); 2288 } 2289 2290 case Resizing: { 2291 // Sweeping has been completed... 2292 // At this point the background collection has completed. 2293 // Don't move the call to compute_new_size() down 2294 // into code that might be executed if the background 2295 // collection was preempted. 2296 { 2297 ReleaseForegroundGC x(this); // unblock FG collection 2298 MutexLockerEx y(Heap_lock, Mutex::_no_safepoint_check_flag); 2299 CMSTokenSync z(true); // not strictly needed. 2300 if (_collectorState == Resizing) { 2301 compute_new_size(); 2302 _collectorState = Resetting; 2303 } else { 2304 assert(_collectorState == Idling, "The state should only change" 2305 " because the foreground collector has finished the collection"); 2306 } 2307 } 2308 break; 2309 } 2310 case Resetting: 2311 // CMS heap resizing has been completed 2312 reset(true); 2313 assert(_collectorState == Idling, "Collector state should " 2314 "have changed"); 2315 stats().record_cms_end(); 2316 // Don't move the concurrent_phases_end() and compute_new_size() 2317 // calls to here because a preempted background collection 2318 // has it's state set to "Resetting". 2319 break; 2320 case Idling: 2321 default: 2322 ShouldNotReachHere(); 2323 break; 2324 } 2325 if (TraceCMSState) { 2326 gclog_or_tty->print_cr(" Thread " INTPTR_FORMAT " done - next CMS state %d", 2327 Thread::current(), _collectorState); 2328 } 2329 assert(_foregroundGCShouldWait, "block post-condition"); 2330 } 2331 2332 // Should this be in gc_epilogue? 2333 collector_policy()->counters()->update_counters(); 2334 2335 { 2336 // Clear _foregroundGCShouldWait and, in the event that the 2337 // foreground collector is waiting, notify it, before 2338 // returning. 2339 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 2340 _foregroundGCShouldWait = false; 2341 if (_foregroundGCIsActive) { 2342 CGC_lock->notify(); 2343 } 2344 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 2345 "Possible deadlock"); 2346 } 2347 if (TraceCMSState) { 2348 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT 2349 " exiting collection CMS state %d", 2350 Thread::current(), _collectorState); 2351 } 2352 if (PrintGC && Verbose) { 2353 _cmsGen->print_heap_change(prev_used); 2354 } 2355 } 2356 2357 void CMSCollector::collect_in_foreground(bool clear_all_soft_refs) { 2358 assert(_foregroundGCIsActive && !_foregroundGCShouldWait, 2359 "Foreground collector should be waiting, not executing"); 2360 assert(Thread::current()->is_VM_thread(), "A foreground collection" 2361 "may only be done by the VM Thread with the world stopped"); 2362 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 2363 "VM thread should have CMS token"); 2364 2365 NOT_PRODUCT(TraceTime t("CMS:MS (foreground) ", PrintGCDetails && Verbose, 2366 true, gclog_or_tty);) 2367 if (UseAdaptiveSizePolicy) { 2368 size_policy()->ms_collection_begin(); 2369 } 2370 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact); 2371 2372 HandleMark hm; // Discard invalid handles created during verification 2373 2374 if (VerifyBeforeGC && 2375 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2376 Universe::verify(true); 2377 } 2378 2379 bool init_mark_was_synchronous = false; // until proven otherwise 2380 while (_collectorState != Idling) { 2381 if (TraceCMSState) { 2382 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " in CMS state %d", 2383 Thread::current(), _collectorState); 2384 } 2385 switch (_collectorState) { 2386 case InitialMarking: 2387 init_mark_was_synchronous = true; // fact to be exploited in re-mark 2388 checkpointRootsInitial(false); 2389 assert(_collectorState == Marking, "Collector state should have changed" 2390 " within checkpointRootsInitial()"); 2391 break; 2392 case Marking: 2393 // initial marking in checkpointRootsInitialWork has been completed 2394 if (VerifyDuringGC && 2395 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2396 gclog_or_tty->print("Verify before initial mark: "); 2397 Universe::verify(true); 2398 } 2399 { 2400 bool res = markFromRoots(false); 2401 assert(res && _collectorState == FinalMarking, "Collector state should " 2402 "have changed"); 2403 break; 2404 } 2405 case FinalMarking: 2406 if (VerifyDuringGC && 2407 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2408 gclog_or_tty->print("Verify before re-mark: "); 2409 Universe::verify(true); 2410 } 2411 checkpointRootsFinal(false, clear_all_soft_refs, 2412 init_mark_was_synchronous); 2413 assert(_collectorState == Sweeping, "Collector state should not " 2414 "have changed within checkpointRootsFinal()"); 2415 break; 2416 case Sweeping: 2417 // final marking in checkpointRootsFinal has been completed 2418 if (VerifyDuringGC && 2419 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2420 gclog_or_tty->print("Verify before sweep: "); 2421 Universe::verify(true); 2422 } 2423 sweep(false); 2424 assert(_collectorState == Resizing, "Incorrect state"); 2425 break; 2426 case Resizing: { 2427 // Sweeping has been completed; the actual resize in this case 2428 // is done separately; nothing to be done in this state. 2429 _collectorState = Resetting; 2430 break; 2431 } 2432 case Resetting: 2433 // The heap has been resized. 2434 if (VerifyDuringGC && 2435 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2436 gclog_or_tty->print("Verify before reset: "); 2437 Universe::verify(true); 2438 } 2439 reset(false); 2440 assert(_collectorState == Idling, "Collector state should " 2441 "have changed"); 2442 break; 2443 case Precleaning: 2444 case AbortablePreclean: 2445 // Elide the preclean phase 2446 _collectorState = FinalMarking; 2447 break; 2448 default: 2449 ShouldNotReachHere(); 2450 } 2451 if (TraceCMSState) { 2452 gclog_or_tty->print_cr(" Thread " INTPTR_FORMAT " done - next CMS state %d", 2453 Thread::current(), _collectorState); 2454 } 2455 } 2456 2457 if (UseAdaptiveSizePolicy) { 2458 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2459 size_policy()->ms_collection_end(gch->gc_cause()); 2460 } 2461 2462 if (VerifyAfterGC && 2463 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2464 Universe::verify(true); 2465 } 2466 if (TraceCMSState) { 2467 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT 2468 " exiting collection CMS state %d", 2469 Thread::current(), _collectorState); 2470 } 2471 } 2472 2473 bool CMSCollector::waitForForegroundGC() { 2474 bool res = false; 2475 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 2476 "CMS thread should have CMS token"); 2477 // Block the foreground collector until the 2478 // background collectors decides whether to 2479 // yield. 2480 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 2481 _foregroundGCShouldWait = true; 2482 if (_foregroundGCIsActive) { 2483 // The background collector yields to the 2484 // foreground collector and returns a value 2485 // indicating that it has yielded. The foreground 2486 // collector can proceed. 2487 res = true; 2488 _foregroundGCShouldWait = false; 2489 ConcurrentMarkSweepThread::clear_CMS_flag( 2490 ConcurrentMarkSweepThread::CMS_cms_has_token); 2491 ConcurrentMarkSweepThread::set_CMS_flag( 2492 ConcurrentMarkSweepThread::CMS_cms_wants_token); 2493 // Get a possibly blocked foreground thread going 2494 CGC_lock->notify(); 2495 if (TraceCMSState) { 2496 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " waiting at CMS state %d", 2497 Thread::current(), _collectorState); 2498 } 2499 while (_foregroundGCIsActive) { 2500 CGC_lock->wait(Mutex::_no_safepoint_check_flag); 2501 } 2502 ConcurrentMarkSweepThread::set_CMS_flag( 2503 ConcurrentMarkSweepThread::CMS_cms_has_token); 2504 ConcurrentMarkSweepThread::clear_CMS_flag( 2505 ConcurrentMarkSweepThread::CMS_cms_wants_token); 2506 } 2507 if (TraceCMSState) { 2508 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " continuing at CMS state %d", 2509 Thread::current(), _collectorState); 2510 } 2511 return res; 2512 } 2513 2514 // Because of the need to lock the free lists and other structures in 2515 // the collector, common to all the generations that the collector is 2516 // collecting, we need the gc_prologues of individual CMS generations 2517 // delegate to their collector. It may have been simpler had the 2518 // current infrastructure allowed one to call a prologue on a 2519 // collector. In the absence of that we have the generation's 2520 // prologue delegate to the collector, which delegates back 2521 // some "local" work to a worker method in the individual generations 2522 // that it's responsible for collecting, while itself doing any 2523 // work common to all generations it's responsible for. A similar 2524 // comment applies to the gc_epilogue()'s. 2525 // The role of the varaible _between_prologue_and_epilogue is to 2526 // enforce the invocation protocol. 2527 void CMSCollector::gc_prologue(bool full) { 2528 // Call gc_prologue_work() for each CMSGen and PermGen that 2529 // we are responsible for. 2530 2531 // The following locking discipline assumes that we are only called 2532 // when the world is stopped. 2533 assert(SafepointSynchronize::is_at_safepoint(), "world is stopped assumption"); 2534 2535 // The CMSCollector prologue must call the gc_prologues for the 2536 // "generations" (including PermGen if any) that it's responsible 2537 // for. 2538 2539 assert( Thread::current()->is_VM_thread() 2540 || ( CMSScavengeBeforeRemark 2541 && Thread::current()->is_ConcurrentGC_thread()), 2542 "Incorrect thread type for prologue execution"); 2543 2544 if (_between_prologue_and_epilogue) { 2545 // We have already been invoked; this is a gc_prologue delegation 2546 // from yet another CMS generation that we are responsible for, just 2547 // ignore it since all relevant work has already been done. 2548 return; 2549 } 2550 2551 // set a bit saying prologue has been called; cleared in epilogue 2552 _between_prologue_and_epilogue = true; 2553 // Claim locks for common data structures, then call gc_prologue_work() 2554 // for each CMSGen and PermGen that we are responsible for. 2555 2556 getFreelistLocks(); // gets free list locks on constituent spaces 2557 bitMapLock()->lock_without_safepoint_check(); 2558 2559 // Should call gc_prologue_work() for all cms gens we are responsible for 2560 bool registerClosure = _collectorState >= Marking 2561 && _collectorState < Sweeping; 2562 ModUnionClosure* muc = ParallelGCThreads > 0 ? &_modUnionClosurePar 2563 : &_modUnionClosure; 2564 _cmsGen->gc_prologue_work(full, registerClosure, muc); 2565 _permGen->gc_prologue_work(full, registerClosure, muc); 2566 2567 if (!full) { 2568 stats().record_gc0_begin(); 2569 } 2570 } 2571 2572 void ConcurrentMarkSweepGeneration::gc_prologue(bool full) { 2573 // Delegate to CMScollector which knows how to coordinate between 2574 // this and any other CMS generations that it is responsible for 2575 // collecting. 2576 collector()->gc_prologue(full); 2577 } 2578 2579 // This is a "private" interface for use by this generation's CMSCollector. 2580 // Not to be called directly by any other entity (for instance, 2581 // GenCollectedHeap, which calls the "public" gc_prologue method above). 2582 void ConcurrentMarkSweepGeneration::gc_prologue_work(bool full, 2583 bool registerClosure, ModUnionClosure* modUnionClosure) { 2584 assert(!incremental_collection_failed(), "Shouldn't be set yet"); 2585 assert(cmsSpace()->preconsumptionDirtyCardClosure() == NULL, 2586 "Should be NULL"); 2587 if (registerClosure) { 2588 cmsSpace()->setPreconsumptionDirtyCardClosure(modUnionClosure); 2589 } 2590 cmsSpace()->gc_prologue(); 2591 // Clear stat counters 2592 NOT_PRODUCT( 2593 assert(_numObjectsPromoted == 0, "check"); 2594 assert(_numWordsPromoted == 0, "check"); 2595 if (Verbose && PrintGC) { 2596 gclog_or_tty->print("Allocated "SIZE_FORMAT" objects, " 2597 SIZE_FORMAT" bytes concurrently", 2598 _numObjectsAllocated, _numWordsAllocated*sizeof(HeapWord)); 2599 } 2600 _numObjectsAllocated = 0; 2601 _numWordsAllocated = 0; 2602 ) 2603 } 2604 2605 void CMSCollector::gc_epilogue(bool full) { 2606 // The following locking discipline assumes that we are only called 2607 // when the world is stopped. 2608 assert(SafepointSynchronize::is_at_safepoint(), 2609 "world is stopped assumption"); 2610 2611 // Currently the CMS epilogue (see CompactibleFreeListSpace) merely checks 2612 // if linear allocation blocks need to be appropriately marked to allow the 2613 // the blocks to be parsable. We also check here whether we need to nudge the 2614 // CMS collector thread to start a new cycle (if it's not already active). 2615 assert( Thread::current()->is_VM_thread() 2616 || ( CMSScavengeBeforeRemark 2617 && Thread::current()->is_ConcurrentGC_thread()), 2618 "Incorrect thread type for epilogue execution"); 2619 2620 if (!_between_prologue_and_epilogue) { 2621 // We have already been invoked; this is a gc_epilogue delegation 2622 // from yet another CMS generation that we are responsible for, just 2623 // ignore it since all relevant work has already been done. 2624 return; 2625 } 2626 assert(haveFreelistLocks(), "must have freelist locks"); 2627 assert_lock_strong(bitMapLock()); 2628 2629 _cmsGen->gc_epilogue_work(full); 2630 _permGen->gc_epilogue_work(full); 2631 2632 if (_collectorState == AbortablePreclean || _collectorState == Precleaning) { 2633 // in case sampling was not already enabled, enable it 2634 _start_sampling = true; 2635 } 2636 // reset _eden_chunk_array so sampling starts afresh 2637 _eden_chunk_index = 0; 2638 2639 size_t cms_used = _cmsGen->cmsSpace()->used(); 2640 size_t perm_used = _permGen->cmsSpace()->used(); 2641 2642 // update performance counters - this uses a special version of 2643 // update_counters() that allows the utilization to be passed as a 2644 // parameter, avoiding multiple calls to used(). 2645 // 2646 _cmsGen->update_counters(cms_used); 2647 _permGen->update_counters(perm_used); 2648 2649 if (CMSIncrementalMode) { 2650 icms_update_allocation_limits(); 2651 } 2652 2653 bitMapLock()->unlock(); 2654 releaseFreelistLocks(); 2655 2656 _between_prologue_and_epilogue = false; // ready for next cycle 2657 } 2658 2659 void ConcurrentMarkSweepGeneration::gc_epilogue(bool full) { 2660 collector()->gc_epilogue(full); 2661 2662 // Also reset promotion tracking in par gc thread states. 2663 if (ParallelGCThreads > 0) { 2664 for (uint i = 0; i < ParallelGCThreads; i++) { 2665 _par_gc_thread_states[i]->promo.stopTrackingPromotions(); 2666 } 2667 } 2668 } 2669 2670 void ConcurrentMarkSweepGeneration::gc_epilogue_work(bool full) { 2671 assert(!incremental_collection_failed(), "Should have been cleared"); 2672 cmsSpace()->setPreconsumptionDirtyCardClosure(NULL); 2673 cmsSpace()->gc_epilogue(); 2674 // Print stat counters 2675 NOT_PRODUCT( 2676 assert(_numObjectsAllocated == 0, "check"); 2677 assert(_numWordsAllocated == 0, "check"); 2678 if (Verbose && PrintGC) { 2679 gclog_or_tty->print("Promoted "SIZE_FORMAT" objects, " 2680 SIZE_FORMAT" bytes", 2681 _numObjectsPromoted, _numWordsPromoted*sizeof(HeapWord)); 2682 } 2683 _numObjectsPromoted = 0; 2684 _numWordsPromoted = 0; 2685 ) 2686 2687 if (PrintGC && Verbose) { 2688 // Call down the chain in contiguous_available needs the freelistLock 2689 // so print this out before releasing the freeListLock. 2690 gclog_or_tty->print(" Contiguous available "SIZE_FORMAT" bytes ", 2691 contiguous_available()); 2692 } 2693 } 2694 2695 #ifndef PRODUCT 2696 bool CMSCollector::have_cms_token() { 2697 Thread* thr = Thread::current(); 2698 if (thr->is_VM_thread()) { 2699 return ConcurrentMarkSweepThread::vm_thread_has_cms_token(); 2700 } else if (thr->is_ConcurrentGC_thread()) { 2701 return ConcurrentMarkSweepThread::cms_thread_has_cms_token(); 2702 } else if (thr->is_GC_task_thread()) { 2703 return ConcurrentMarkSweepThread::vm_thread_has_cms_token() && 2704 ParGCRareEvent_lock->owned_by_self(); 2705 } 2706 return false; 2707 } 2708 #endif 2709 2710 // Check reachability of the given heap address in CMS generation, 2711 // treating all other generations as roots. 2712 bool CMSCollector::is_cms_reachable(HeapWord* addr) { 2713 // We could "guarantee" below, rather than assert, but i'll 2714 // leave these as "asserts" so that an adventurous debugger 2715 // could try this in the product build provided some subset of 2716 // the conditions were met, provided they were intersted in the 2717 // results and knew that the computation below wouldn't interfere 2718 // with other concurrent computations mutating the structures 2719 // being read or written. 2720 assert(SafepointSynchronize::is_at_safepoint(), 2721 "Else mutations in object graph will make answer suspect"); 2722 assert(have_cms_token(), "Should hold cms token"); 2723 assert(haveFreelistLocks(), "must hold free list locks"); 2724 assert_lock_strong(bitMapLock()); 2725 2726 // Clear the marking bit map array before starting, but, just 2727 // for kicks, first report if the given address is already marked 2728 gclog_or_tty->print_cr("Start: Address 0x%x is%s marked", addr, 2729 _markBitMap.isMarked(addr) ? "" : " not"); 2730 2731 if (verify_after_remark()) { 2732 MutexLockerEx x(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag); 2733 bool result = verification_mark_bm()->isMarked(addr); 2734 gclog_or_tty->print_cr("TransitiveMark: Address 0x%x %s marked", addr, 2735 result ? "IS" : "is NOT"); 2736 return result; 2737 } else { 2738 gclog_or_tty->print_cr("Could not compute result"); 2739 return false; 2740 } 2741 } 2742 2743 //////////////////////////////////////////////////////// 2744 // CMS Verification Support 2745 //////////////////////////////////////////////////////// 2746 // Following the remark phase, the following invariant 2747 // should hold -- each object in the CMS heap which is 2748 // marked in markBitMap() should be marked in the verification_mark_bm(). 2749 2750 class VerifyMarkedClosure: public BitMapClosure { 2751 CMSBitMap* _marks; 2752 bool _failed; 2753 2754 public: 2755 VerifyMarkedClosure(CMSBitMap* bm): _marks(bm), _failed(false) {} 2756 2757 void do_bit(size_t offset) { 2758 HeapWord* addr = _marks->offsetToHeapWord(offset); 2759 if (!_marks->isMarked(addr)) { 2760 oop(addr)->print(); 2761 gclog_or_tty->print_cr(" ("INTPTR_FORMAT" should have been marked)", addr); 2762 _failed = true; 2763 } 2764 } 2765 2766 bool failed() { return _failed; } 2767 }; 2768 2769 bool CMSCollector::verify_after_remark() { 2770 gclog_or_tty->print(" [Verifying CMS Marking... "); 2771 MutexLockerEx ml(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag); 2772 static bool init = false; 2773 2774 assert(SafepointSynchronize::is_at_safepoint(), 2775 "Else mutations in object graph will make answer suspect"); 2776 assert(have_cms_token(), 2777 "Else there may be mutual interference in use of " 2778 " verification data structures"); 2779 assert(_collectorState > Marking && _collectorState <= Sweeping, 2780 "Else marking info checked here may be obsolete"); 2781 assert(haveFreelistLocks(), "must hold free list locks"); 2782 assert_lock_strong(bitMapLock()); 2783 2784 2785 // Allocate marking bit map if not already allocated 2786 if (!init) { // first time 2787 if (!verification_mark_bm()->allocate(_span)) { 2788 return false; 2789 } 2790 init = true; 2791 } 2792 2793 assert(verification_mark_stack()->isEmpty(), "Should be empty"); 2794 2795 // Turn off refs discovery -- so we will be tracing through refs. 2796 // This is as intended, because by this time 2797 // GC must already have cleared any refs that need to be cleared, 2798 // and traced those that need to be marked; moreover, 2799 // the marking done here is not going to intefere in any 2800 // way with the marking information used by GC. 2801 NoRefDiscovery no_discovery(ref_processor()); 2802 2803 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;) 2804 2805 // Clear any marks from a previous round 2806 verification_mark_bm()->clear_all(); 2807 assert(verification_mark_stack()->isEmpty(), "markStack should be empty"); 2808 assert(overflow_list_is_empty(), "overflow list should be empty"); 2809 2810 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2811 gch->ensure_parsability(false); // fill TLABs, but no need to retire them 2812 // Update the saved marks which may affect the root scans. 2813 gch->save_marks(); 2814 2815 if (CMSRemarkVerifyVariant == 1) { 2816 // In this first variant of verification, we complete 2817 // all marking, then check if the new marks-verctor is 2818 // a subset of the CMS marks-vector. 2819 verify_after_remark_work_1(); 2820 } else if (CMSRemarkVerifyVariant == 2) { 2821 // In this second variant of verification, we flag an error 2822 // (i.e. an object reachable in the new marks-vector not reachable 2823 // in the CMS marks-vector) immediately, also indicating the 2824 // identify of an object (A) that references the unmarked object (B) -- 2825 // presumably, a mutation to A failed to be picked up by preclean/remark? 2826 verify_after_remark_work_2(); 2827 } else { 2828 warning("Unrecognized value %d for CMSRemarkVerifyVariant", 2829 CMSRemarkVerifyVariant); 2830 } 2831 gclog_or_tty->print(" done] "); 2832 return true; 2833 } 2834 2835 void CMSCollector::verify_after_remark_work_1() { 2836 ResourceMark rm; 2837 HandleMark hm; 2838 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2839 2840 // Mark from roots one level into CMS 2841 MarkRefsIntoClosure notOlder(_span, verification_mark_bm(), true /* nmethods */); 2842 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2843 2844 gch->gen_process_strong_roots(_cmsGen->level(), 2845 true, // younger gens are roots 2846 true, // collecting perm gen 2847 SharedHeap::ScanningOption(roots_scanning_options()), 2848 NULL, ¬Older); 2849 2850 // Now mark from the roots 2851 assert(_revisitStack.isEmpty(), "Should be empty"); 2852 MarkFromRootsClosure markFromRootsClosure(this, _span, 2853 verification_mark_bm(), verification_mark_stack(), &_revisitStack, 2854 false /* don't yield */, true /* verifying */); 2855 assert(_restart_addr == NULL, "Expected pre-condition"); 2856 verification_mark_bm()->iterate(&markFromRootsClosure); 2857 while (_restart_addr != NULL) { 2858 // Deal with stack overflow: by restarting at the indicated 2859 // address. 2860 HeapWord* ra = _restart_addr; 2861 markFromRootsClosure.reset(ra); 2862 _restart_addr = NULL; 2863 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end()); 2864 } 2865 assert(verification_mark_stack()->isEmpty(), "Should have been drained"); 2866 verify_work_stacks_empty(); 2867 // Should reset the revisit stack above, since no class tree 2868 // surgery is forthcoming. 2869 _revisitStack.reset(); // throwing away all contents 2870 2871 // Marking completed -- now verify that each bit marked in 2872 // verification_mark_bm() is also marked in markBitMap(); flag all 2873 // errors by printing corresponding objects. 2874 VerifyMarkedClosure vcl(markBitMap()); 2875 verification_mark_bm()->iterate(&vcl); 2876 if (vcl.failed()) { 2877 gclog_or_tty->print("Verification failed"); 2878 Universe::heap()->print(); 2879 fatal(" ... aborting"); 2880 } 2881 } 2882 2883 void CMSCollector::verify_after_remark_work_2() { 2884 ResourceMark rm; 2885 HandleMark hm; 2886 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2887 2888 // Mark from roots one level into CMS 2889 MarkRefsIntoVerifyClosure notOlder(_span, verification_mark_bm(), 2890 markBitMap(), true /* nmethods */); 2891 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2892 gch->gen_process_strong_roots(_cmsGen->level(), 2893 true, // younger gens are roots 2894 true, // collecting perm gen 2895 SharedHeap::ScanningOption(roots_scanning_options()), 2896 NULL, ¬Older); 2897 2898 // Now mark from the roots 2899 assert(_revisitStack.isEmpty(), "Should be empty"); 2900 MarkFromRootsVerifyClosure markFromRootsClosure(this, _span, 2901 verification_mark_bm(), markBitMap(), verification_mark_stack()); 2902 assert(_restart_addr == NULL, "Expected pre-condition"); 2903 verification_mark_bm()->iterate(&markFromRootsClosure); 2904 while (_restart_addr != NULL) { 2905 // Deal with stack overflow: by restarting at the indicated 2906 // address. 2907 HeapWord* ra = _restart_addr; 2908 markFromRootsClosure.reset(ra); 2909 _restart_addr = NULL; 2910 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end()); 2911 } 2912 assert(verification_mark_stack()->isEmpty(), "Should have been drained"); 2913 verify_work_stacks_empty(); 2914 // Should reset the revisit stack above, since no class tree 2915 // surgery is forthcoming. 2916 _revisitStack.reset(); // throwing away all contents 2917 2918 // Marking completed -- now verify that each bit marked in 2919 // verification_mark_bm() is also marked in markBitMap(); flag all 2920 // errors by printing corresponding objects. 2921 VerifyMarkedClosure vcl(markBitMap()); 2922 verification_mark_bm()->iterate(&vcl); 2923 assert(!vcl.failed(), "Else verification above should not have succeeded"); 2924 } 2925 2926 void ConcurrentMarkSweepGeneration::save_marks() { 2927 // delegate to CMS space 2928 cmsSpace()->save_marks(); 2929 for (uint i = 0; i < ParallelGCThreads; i++) { 2930 _par_gc_thread_states[i]->promo.startTrackingPromotions(); 2931 } 2932 } 2933 2934 bool ConcurrentMarkSweepGeneration::no_allocs_since_save_marks() { 2935 return cmsSpace()->no_allocs_since_save_marks(); 2936 } 2937 2938 #define CMS_SINCE_SAVE_MARKS_DEFN(OopClosureType, nv_suffix) \ 2939 \ 2940 void ConcurrentMarkSweepGeneration:: \ 2941 oop_since_save_marks_iterate##nv_suffix(OopClosureType* cl) { \ 2942 cl->set_generation(this); \ 2943 cmsSpace()->oop_since_save_marks_iterate##nv_suffix(cl); \ 2944 cl->reset_generation(); \ 2945 save_marks(); \ 2946 } 2947 2948 ALL_SINCE_SAVE_MARKS_CLOSURES(CMS_SINCE_SAVE_MARKS_DEFN) 2949 2950 void 2951 ConcurrentMarkSweepGeneration::object_iterate_since_last_GC(ObjectClosure* blk) 2952 { 2953 // Not currently implemented; need to do the following. -- ysr. 2954 // dld -- I think that is used for some sort of allocation profiler. So it 2955 // really means the objects allocated by the mutator since the last 2956 // GC. We could potentially implement this cheaply by recording only 2957 // the direct allocations in a side data structure. 2958 // 2959 // I think we probably ought not to be required to support these 2960 // iterations at any arbitrary point; I think there ought to be some 2961 // call to enable/disable allocation profiling in a generation/space, 2962 // and the iterator ought to return the objects allocated in the 2963 // gen/space since the enable call, or the last iterator call (which 2964 // will probably be at a GC.) That way, for gens like CM&S that would 2965 // require some extra data structure to support this, we only pay the 2966 // cost when it's in use... 2967 cmsSpace()->object_iterate_since_last_GC(blk); 2968 } 2969 2970 void 2971 ConcurrentMarkSweepGeneration::younger_refs_iterate(OopsInGenClosure* cl) { 2972 cl->set_generation(this); 2973 younger_refs_in_space_iterate(_cmsSpace, cl); 2974 cl->reset_generation(); 2975 } 2976 2977 void 2978 ConcurrentMarkSweepGeneration::oop_iterate(MemRegion mr, OopClosure* cl) { 2979 if (freelistLock()->owned_by_self()) { 2980 Generation::oop_iterate(mr, cl); 2981 } else { 2982 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2983 Generation::oop_iterate(mr, cl); 2984 } 2985 } 2986 2987 void 2988 ConcurrentMarkSweepGeneration::oop_iterate(OopClosure* cl) { 2989 if (freelistLock()->owned_by_self()) { 2990 Generation::oop_iterate(cl); 2991 } else { 2992 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2993 Generation::oop_iterate(cl); 2994 } 2995 } 2996 2997 void 2998 ConcurrentMarkSweepGeneration::object_iterate(ObjectClosure* cl) { 2999 if (freelistLock()->owned_by_self()) { 3000 Generation::object_iterate(cl); 3001 } else { 3002 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 3003 Generation::object_iterate(cl); 3004 } 3005 } 3006 3007 void 3008 ConcurrentMarkSweepGeneration::pre_adjust_pointers() { 3009 } 3010 3011 void 3012 ConcurrentMarkSweepGeneration::post_compact() { 3013 } 3014 3015 void 3016 ConcurrentMarkSweepGeneration::prepare_for_verify() { 3017 // Fix the linear allocation blocks to look like free blocks. 3018 3019 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those 3020 // are not called when the heap is verified during universe initialization and 3021 // at vm shutdown. 3022 if (freelistLock()->owned_by_self()) { 3023 cmsSpace()->prepare_for_verify(); 3024 } else { 3025 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag); 3026 cmsSpace()->prepare_for_verify(); 3027 } 3028 } 3029 3030 void 3031 ConcurrentMarkSweepGeneration::verify(bool allow_dirty /* ignored */) { 3032 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those 3033 // are not called when the heap is verified during universe initialization and 3034 // at vm shutdown. 3035 if (freelistLock()->owned_by_self()) { 3036 cmsSpace()->verify(false /* ignored */); 3037 } else { 3038 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag); 3039 cmsSpace()->verify(false /* ignored */); 3040 } 3041 } 3042 3043 void CMSCollector::verify(bool allow_dirty /* ignored */) { 3044 _cmsGen->verify(allow_dirty); 3045 _permGen->verify(allow_dirty); 3046 } 3047 3048 #ifndef PRODUCT 3049 bool CMSCollector::overflow_list_is_empty() const { 3050 assert(_num_par_pushes >= 0, "Inconsistency"); 3051 if (_overflow_list == NULL) { 3052 assert(_num_par_pushes == 0, "Inconsistency"); 3053 } 3054 return _overflow_list == NULL; 3055 } 3056 3057 // The methods verify_work_stacks_empty() and verify_overflow_empty() 3058 // merely consolidate assertion checks that appear to occur together frequently. 3059 void CMSCollector::verify_work_stacks_empty() const { 3060 assert(_markStack.isEmpty(), "Marking stack should be empty"); 3061 assert(overflow_list_is_empty(), "Overflow list should be empty"); 3062 } 3063 3064 void CMSCollector::verify_overflow_empty() const { 3065 assert(overflow_list_is_empty(), "Overflow list should be empty"); 3066 assert(no_preserved_marks(), "No preserved marks"); 3067 } 3068 #endif // PRODUCT 3069 3070 // Decide if we want to enable class unloading as part of the 3071 // ensuing concurrent GC cycle. We will collect the perm gen and 3072 // unload classes if it's the case that: 3073 // (1) an explicit gc request has been made and the flag 3074 // ExplicitGCInvokesConcurrentAndUnloadsClasses is set, OR 3075 // (2) (a) class unloading is enabled at the command line, and 3076 // (b) (i) perm gen threshold has been crossed, or 3077 // (ii) old gen is getting really full, or 3078 // (iii) the previous N CMS collections did not collect the 3079 // perm gen 3080 // NOTE: Provided there is no change in the state of the heap between 3081 // calls to this method, it should have idempotent results. Moreover, 3082 // its results should be monotonically increasing (i.e. going from 0 to 1, 3083 // but not 1 to 0) between successive calls between which the heap was 3084 // not collected. For the implementation below, it must thus rely on 3085 // the property that concurrent_cycles_since_last_unload() 3086 // will not decrease unless a collection cycle happened and that 3087 // _permGen->should_concurrent_collect() and _cmsGen->is_too_full() are 3088 // themselves also monotonic in that sense. See check_monotonicity() 3089 // below. 3090 bool CMSCollector::update_should_unload_classes() { 3091 // Condition 1 above 3092 if (_full_gc_requested && ExplicitGCInvokesConcurrentAndUnloadsClasses) { 3093 _should_unload_classes = true; 3094 } else if (CMSClassUnloadingEnabled) { // Condition 2.a above 3095 // Disjuncts 2.b.(i,ii,iii) above 3096 _should_unload_classes = (concurrent_cycles_since_last_unload() >= 3097 CMSClassUnloadingMaxInterval) 3098 || _permGen->should_concurrent_collect() 3099 || _cmsGen->is_too_full(); 3100 } 3101 return _should_unload_classes; 3102 } 3103 3104 bool ConcurrentMarkSweepGeneration::is_too_full() const { 3105 bool res = should_concurrent_collect(); 3106 #define CMSIsTooFullPercentage 98 3107 res = res && occupancy() > (double)CMSIsTooFullPercentage/100.0; 3108 return res; 3109 } 3110 3111 void CMSCollector::setup_cms_unloading_and_verification_state() { 3112 const bool should_verify = VerifyBeforeGC || VerifyAfterGC || VerifyDuringGC 3113 || VerifyBeforeExit; 3114 const int rso = SharedHeap::SO_Symbols | SharedHeap::SO_Strings 3115 | SharedHeap::SO_CodeCache; 3116 3117 if (should_unload_classes()) { // Should unload classes this cycle 3118 remove_root_scanning_option(rso); // Shrink the root set appropriately 3119 set_verifying(should_verify); // Set verification state for this cycle 3120 return; // Nothing else needs to be done at this time 3121 } 3122 3123 // Not unloading classes this cycle 3124 assert(!should_unload_classes(), "Inconsitency!"); 3125 if ((!verifying() || unloaded_classes_last_cycle()) && should_verify) { 3126 // We were not verifying, or we _were_ unloading classes in the last cycle, 3127 // AND some verification options are enabled this cycle; in this case, 3128 // we must make sure that the deadness map is allocated if not already so, 3129 // and cleared (if already allocated previously -- 3130 // CMSBitMap::sizeInBits() is used to determine if it's allocated). 3131 if (perm_gen_verify_bit_map()->sizeInBits() == 0) { 3132 if (!perm_gen_verify_bit_map()->allocate(_permGen->reserved())) { 3133 warning("Failed to allocate permanent generation verification CMS Bit Map;\n" 3134 "permanent generation verification disabled"); 3135 return; // Note that we leave verification disabled, so we'll retry this 3136 // allocation next cycle. We _could_ remember this failure 3137 // and skip further attempts and permanently disable verification 3138 // attempts if that is considered more desirable. 3139 } 3140 assert(perm_gen_verify_bit_map()->covers(_permGen->reserved()), 3141 "_perm_gen_ver_bit_map inconsistency?"); 3142 } else { 3143 perm_gen_verify_bit_map()->clear_all(); 3144 } 3145 // Include symbols, strings and code cache elements to prevent their resurrection. 3146 add_root_scanning_option(rso); 3147 set_verifying(true); 3148 } else if (verifying() && !should_verify) { 3149 // We were verifying, but some verification flags got disabled. 3150 set_verifying(false); 3151 // Exclude symbols, strings and code cache elements from root scanning to 3152 // reduce IM and RM pauses. 3153 remove_root_scanning_option(rso); 3154 } 3155 } 3156 3157 3158 #ifndef PRODUCT 3159 HeapWord* CMSCollector::block_start(const void* p) const { 3160 const HeapWord* addr = (HeapWord*)p; 3161 if (_span.contains(p)) { 3162 if (_cmsGen->cmsSpace()->is_in_reserved(addr)) { 3163 return _cmsGen->cmsSpace()->block_start(p); 3164 } else { 3165 assert(_permGen->cmsSpace()->is_in_reserved(addr), 3166 "Inconsistent _span?"); 3167 return _permGen->cmsSpace()->block_start(p); 3168 } 3169 } 3170 return NULL; 3171 } 3172 #endif 3173 3174 HeapWord* 3175 ConcurrentMarkSweepGeneration::expand_and_allocate(size_t word_size, 3176 bool tlab, 3177 bool parallel) { 3178 assert(!tlab, "Can't deal with TLAB allocation"); 3179 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 3180 expand(word_size*HeapWordSize, MinHeapDeltaBytes, 3181 CMSExpansionCause::_satisfy_allocation); 3182 if (GCExpandToAllocateDelayMillis > 0) { 3183 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 3184 } 3185 size_t adj_word_sz = CompactibleFreeListSpace::adjustObjectSize(word_size); 3186 if (parallel) { 3187 return cmsSpace()->par_allocate(adj_word_sz); 3188 } else { 3189 return cmsSpace()->allocate(adj_word_sz); 3190 } 3191 } 3192 3193 // YSR: All of this generation expansion/shrinking stuff is an exact copy of 3194 // OneContigSpaceCardGeneration, which makes me wonder if we should move this 3195 // to CardGeneration and share it... 3196 void ConcurrentMarkSweepGeneration::expand(size_t bytes, size_t expand_bytes, 3197 CMSExpansionCause::Cause cause) 3198 { 3199 assert_locked_or_safepoint(Heap_lock); 3200 3201 size_t aligned_bytes = ReservedSpace::page_align_size_up(bytes); 3202 size_t aligned_expand_bytes = ReservedSpace::page_align_size_up(expand_bytes); 3203 bool success = false; 3204 if (aligned_expand_bytes > aligned_bytes) { 3205 success = grow_by(aligned_expand_bytes); 3206 } 3207 if (!success) { 3208 success = grow_by(aligned_bytes); 3209 } 3210 if (!success) { 3211 size_t remaining_bytes = _virtual_space.uncommitted_size(); 3212 if (remaining_bytes > 0) { 3213 success = grow_by(remaining_bytes); 3214 } 3215 } 3216 if (GC_locker::is_active()) { 3217 if (PrintGC && Verbose) { 3218 gclog_or_tty->print_cr("Garbage collection disabled, expanded heap instead"); 3219 } 3220 } 3221 // remember why we expanded; this information is used 3222 // by shouldConcurrentCollect() when making decisions on whether to start 3223 // a new CMS cycle. 3224 if (success) { 3225 set_expansion_cause(cause); 3226 if (PrintGCDetails && Verbose) { 3227 gclog_or_tty->print_cr("Expanded CMS gen for %s", 3228 CMSExpansionCause::to_string(cause)); 3229 } 3230 } 3231 } 3232 3233 HeapWord* ConcurrentMarkSweepGeneration::expand_and_par_lab_allocate(CMSParGCThreadState* ps, size_t word_sz) { 3234 HeapWord* res = NULL; 3235 MutexLocker x(ParGCRareEvent_lock); 3236 while (true) { 3237 // Expansion by some other thread might make alloc OK now: 3238 res = ps->lab.alloc(word_sz); 3239 if (res != NULL) return res; 3240 // If there's not enough expansion space available, give up. 3241 if (_virtual_space.uncommitted_size() < (word_sz * HeapWordSize)) { 3242 return NULL; 3243 } 3244 // Otherwise, we try expansion. 3245 expand(word_sz*HeapWordSize, MinHeapDeltaBytes, 3246 CMSExpansionCause::_allocate_par_lab); 3247 // Now go around the loop and try alloc again; 3248 // A competing par_promote might beat us to the expansion space, 3249 // so we may go around the loop again if promotion fails agaion. 3250 if (GCExpandToAllocateDelayMillis > 0) { 3251 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 3252 } 3253 } 3254 } 3255 3256 3257 bool ConcurrentMarkSweepGeneration::expand_and_ensure_spooling_space( 3258 PromotionInfo* promo) { 3259 MutexLocker x(ParGCRareEvent_lock); 3260 size_t refill_size_bytes = promo->refillSize() * HeapWordSize; 3261 while (true) { 3262 // Expansion by some other thread might make alloc OK now: 3263 if (promo->ensure_spooling_space()) { 3264 assert(promo->has_spooling_space(), 3265 "Post-condition of successful ensure_spooling_space()"); 3266 return true; 3267 } 3268 // If there's not enough expansion space available, give up. 3269 if (_virtual_space.uncommitted_size() < refill_size_bytes) { 3270 return false; 3271 } 3272 // Otherwise, we try expansion. 3273 expand(refill_size_bytes, MinHeapDeltaBytes, 3274 CMSExpansionCause::_allocate_par_spooling_space); 3275 // Now go around the loop and try alloc again; 3276 // A competing allocation might beat us to the expansion space, 3277 // so we may go around the loop again if allocation fails again. 3278 if (GCExpandToAllocateDelayMillis > 0) { 3279 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 3280 } 3281 } 3282 } 3283 3284 3285 3286 void ConcurrentMarkSweepGeneration::shrink(size_t bytes) { 3287 assert_locked_or_safepoint(Heap_lock); 3288 size_t size = ReservedSpace::page_align_size_down(bytes); 3289 if (size > 0) { 3290 shrink_by(size); 3291 } 3292 } 3293 3294 bool ConcurrentMarkSweepGeneration::grow_by(size_t bytes) { 3295 assert_locked_or_safepoint(Heap_lock); 3296 bool result = _virtual_space.expand_by(bytes); 3297 if (result) { 3298 HeapWord* old_end = _cmsSpace->end(); 3299 size_t new_word_size = 3300 heap_word_size(_virtual_space.committed_size()); 3301 MemRegion mr(_cmsSpace->bottom(), new_word_size); 3302 _bts->resize(new_word_size); // resize the block offset shared array 3303 Universe::heap()->barrier_set()->resize_covered_region(mr); 3304 // Hmmmm... why doesn't CFLS::set_end verify locking? 3305 // This is quite ugly; FIX ME XXX 3306 _cmsSpace->assert_locked(); 3307 _cmsSpace->set_end((HeapWord*)_virtual_space.high()); 3308 3309 // update the space and generation capacity counters 3310 if (UsePerfData) { 3311 _space_counters->update_capacity(); 3312 _gen_counters->update_all(); 3313 } 3314 3315 if (Verbose && PrintGC) { 3316 size_t new_mem_size = _virtual_space.committed_size(); 3317 size_t old_mem_size = new_mem_size - bytes; 3318 gclog_or_tty->print_cr("Expanding %s from %ldK by %ldK to %ldK", 3319 name(), old_mem_size/K, bytes/K, new_mem_size/K); 3320 } 3321 } 3322 return result; 3323 } 3324 3325 bool ConcurrentMarkSweepGeneration::grow_to_reserved() { 3326 assert_locked_or_safepoint(Heap_lock); 3327 bool success = true; 3328 const size_t remaining_bytes = _virtual_space.uncommitted_size(); 3329 if (remaining_bytes > 0) { 3330 success = grow_by(remaining_bytes); 3331 DEBUG_ONLY(if (!success) warning("grow to reserved failed");) 3332 } 3333 return success; 3334 } 3335 3336 void ConcurrentMarkSweepGeneration::shrink_by(size_t bytes) { 3337 assert_locked_or_safepoint(Heap_lock); 3338 assert_lock_strong(freelistLock()); 3339 // XXX Fix when compaction is implemented. 3340 warning("Shrinking of CMS not yet implemented"); 3341 return; 3342 } 3343 3344 3345 // Simple ctor/dtor wrapper for accounting & timer chores around concurrent 3346 // phases. 3347 class CMSPhaseAccounting: public StackObj { 3348 public: 3349 CMSPhaseAccounting(CMSCollector *collector, 3350 const char *phase, 3351 bool print_cr = true); 3352 ~CMSPhaseAccounting(); 3353 3354 private: 3355 CMSCollector *_collector; 3356 const char *_phase; 3357 elapsedTimer _wallclock; 3358 bool _print_cr; 3359 3360 public: 3361 // Not MT-safe; so do not pass around these StackObj's 3362 // where they may be accessed by other threads. 3363 jlong wallclock_millis() { 3364 assert(_wallclock.is_active(), "Wall clock should not stop"); 3365 _wallclock.stop(); // to record time 3366 jlong ret = _wallclock.milliseconds(); 3367 _wallclock.start(); // restart 3368 return ret; 3369 } 3370 }; 3371 3372 CMSPhaseAccounting::CMSPhaseAccounting(CMSCollector *collector, 3373 const char *phase, 3374 bool print_cr) : 3375 _collector(collector), _phase(phase), _print_cr(print_cr) { 3376 3377 if (PrintCMSStatistics != 0) { 3378 _collector->resetYields(); 3379 } 3380 if (PrintGCDetails && PrintGCTimeStamps) { 3381 gclog_or_tty->date_stamp(PrintGCDateStamps); 3382 gclog_or_tty->stamp(); 3383 gclog_or_tty->print_cr(": [%s-concurrent-%s-start]", 3384 _collector->cmsGen()->short_name(), _phase); 3385 } 3386 _collector->resetTimer(); 3387 _wallclock.start(); 3388 _collector->startTimer(); 3389 } 3390 3391 CMSPhaseAccounting::~CMSPhaseAccounting() { 3392 assert(_wallclock.is_active(), "Wall clock should not have stopped"); 3393 _collector->stopTimer(); 3394 _wallclock.stop(); 3395 if (PrintGCDetails) { 3396 gclog_or_tty->date_stamp(PrintGCDateStamps); 3397 if (PrintGCTimeStamps) { 3398 gclog_or_tty->stamp(); 3399 gclog_or_tty->print(": "); 3400 } 3401 gclog_or_tty->print("[%s-concurrent-%s: %3.3f/%3.3f secs]", 3402 _collector->cmsGen()->short_name(), 3403 _phase, _collector->timerValue(), _wallclock.seconds()); 3404 if (_print_cr) { 3405 gclog_or_tty->print_cr(""); 3406 } 3407 if (PrintCMSStatistics != 0) { 3408 gclog_or_tty->print_cr(" (CMS-concurrent-%s yielded %d times)", _phase, 3409 _collector->yields()); 3410 } 3411 } 3412 } 3413 3414 // CMS work 3415 3416 // Checkpoint the roots into this generation from outside 3417 // this generation. [Note this initial checkpoint need only 3418 // be approximate -- we'll do a catch up phase subsequently.] 3419 void CMSCollector::checkpointRootsInitial(bool asynch) { 3420 assert(_collectorState == InitialMarking, "Wrong collector state"); 3421 check_correct_thread_executing(); 3422 ReferenceProcessor* rp = ref_processor(); 3423 SpecializationStats::clear(); 3424 assert(_restart_addr == NULL, "Control point invariant"); 3425 if (asynch) { 3426 // acquire locks for subsequent manipulations 3427 MutexLockerEx x(bitMapLock(), 3428 Mutex::_no_safepoint_check_flag); 3429 checkpointRootsInitialWork(asynch); 3430 rp->verify_no_references_recorded(); 3431 rp->enable_discovery(); // enable ("weak") refs discovery 3432 _collectorState = Marking; 3433 } else { 3434 // (Weak) Refs discovery: this is controlled from genCollectedHeap::do_collection 3435 // which recognizes if we are a CMS generation, and doesn't try to turn on 3436 // discovery; verify that they aren't meddling. 3437 assert(!rp->discovery_is_atomic(), 3438 "incorrect setting of discovery predicate"); 3439 assert(!rp->discovery_enabled(), "genCollectedHeap shouldn't control " 3440 "ref discovery for this generation kind"); 3441 // already have locks 3442 checkpointRootsInitialWork(asynch); 3443 rp->enable_discovery(); // now enable ("weak") refs discovery 3444 _collectorState = Marking; 3445 } 3446 SpecializationStats::print(); 3447 } 3448 3449 void CMSCollector::checkpointRootsInitialWork(bool asynch) { 3450 assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped"); 3451 assert(_collectorState == InitialMarking, "just checking"); 3452 3453 // If there has not been a GC[n-1] since last GC[n] cycle completed, 3454 // precede our marking with a collection of all 3455 // younger generations to keep floating garbage to a minimum. 3456 // XXX: we won't do this for now -- it's an optimization to be done later. 3457 3458 // already have locks 3459 assert_lock_strong(bitMapLock()); 3460 assert(_markBitMap.isAllClear(), "was reset at end of previous cycle"); 3461 3462 // Setup the verification and class unloading state for this 3463 // CMS collection cycle. 3464 setup_cms_unloading_and_verification_state(); 3465 3466 NOT_PRODUCT(TraceTime t("\ncheckpointRootsInitialWork", 3467 PrintGCDetails && Verbose, true, gclog_or_tty);) 3468 if (UseAdaptiveSizePolicy) { 3469 size_policy()->checkpoint_roots_initial_begin(); 3470 } 3471 3472 // Reset all the PLAB chunk arrays if necessary. 3473 if (_survivor_plab_array != NULL && !CMSPLABRecordAlways) { 3474 reset_survivor_plab_arrays(); 3475 } 3476 3477 ResourceMark rm; 3478 HandleMark hm; 3479 3480 FalseClosure falseClosure; 3481 // In the case of a synchronous collection, we will elide the 3482 // remark step, so it's important to catch all the nmethod oops 3483 // in this step; hence the last argument to the constrcutor below. 3484 MarkRefsIntoClosure notOlder(_span, &_markBitMap, !asynch /* nmethods */); 3485 GenCollectedHeap* gch = GenCollectedHeap::heap(); 3486 3487 verify_work_stacks_empty(); 3488 verify_overflow_empty(); 3489 3490 gch->ensure_parsability(false); // fill TLABs, but no need to retire them 3491 // Update the saved marks which may affect the root scans. 3492 gch->save_marks(); 3493 3494 // weak reference processing has not started yet. 3495 ref_processor()->set_enqueuing_is_done(false); 3496 3497 { 3498 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;) 3499 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 3500 gch->gen_process_strong_roots(_cmsGen->level(), 3501 true, // younger gens are roots 3502 true, // collecting perm gen 3503 SharedHeap::ScanningOption(roots_scanning_options()), 3504 NULL, ¬Older); 3505 } 3506 3507 // Clear mod-union table; it will be dirtied in the prologue of 3508 // CMS generation per each younger generation collection. 3509 3510 assert(_modUnionTable.isAllClear(), 3511 "Was cleared in most recent final checkpoint phase" 3512 " or no bits are set in the gc_prologue before the start of the next " 3513 "subsequent marking phase."); 3514 3515 // Temporarily disabled, since pre/post-consumption closures don't 3516 // care about precleaned cards 3517 #if 0 3518 { 3519 MemRegion mr = MemRegion((HeapWord*)_virtual_space.low(), 3520 (HeapWord*)_virtual_space.high()); 3521 _ct->ct_bs()->preclean_dirty_cards(mr); 3522 } 3523 #endif 3524 3525 // Save the end of the used_region of the constituent generations 3526 // to be used to limit the extent of sweep in each generation. 3527 save_sweep_limits(); 3528 if (UseAdaptiveSizePolicy) { 3529 size_policy()->checkpoint_roots_initial_end(gch->gc_cause()); 3530 } 3531 verify_overflow_empty(); 3532 } 3533 3534 bool CMSCollector::markFromRoots(bool asynch) { 3535 // we might be tempted to assert that: 3536 // assert(asynch == !SafepointSynchronize::is_at_safepoint(), 3537 // "inconsistent argument?"); 3538 // However that wouldn't be right, because it's possible that 3539 // a safepoint is indeed in progress as a younger generation 3540 // stop-the-world GC happens even as we mark in this generation. 3541 assert(_collectorState == Marking, "inconsistent state?"); 3542 check_correct_thread_executing(); 3543 verify_overflow_empty(); 3544 3545 bool res; 3546 if (asynch) { 3547 3548 // Start the timers for adaptive size policy for the concurrent phases 3549 // Do it here so that the foreground MS can use the concurrent 3550 // timer since a foreground MS might has the sweep done concurrently 3551 // or STW. 3552 if (UseAdaptiveSizePolicy) { 3553 size_policy()->concurrent_marking_begin(); 3554 } 3555 3556 // Weak ref discovery note: We may be discovering weak 3557 // refs in this generation concurrent (but interleaved) with 3558 // weak ref discovery by a younger generation collector. 3559 3560 CMSTokenSyncWithLocks ts(true, bitMapLock()); 3561 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 3562 CMSPhaseAccounting pa(this, "mark", !PrintGCDetails); 3563 res = markFromRootsWork(asynch); 3564 if (res) { 3565 _collectorState = Precleaning; 3566 } else { // We failed and a foreground collection wants to take over 3567 assert(_foregroundGCIsActive, "internal state inconsistency"); 3568 assert(_restart_addr == NULL, "foreground will restart from scratch"); 3569 if (PrintGCDetails) { 3570 gclog_or_tty->print_cr("bailing out to foreground collection"); 3571 } 3572 } 3573 if (UseAdaptiveSizePolicy) { 3574 size_policy()->concurrent_marking_end(); 3575 } 3576 } else { 3577 assert(SafepointSynchronize::is_at_safepoint(), 3578 "inconsistent with asynch == false"); 3579 if (UseAdaptiveSizePolicy) { 3580 size_policy()->ms_collection_marking_begin(); 3581 } 3582 // already have locks 3583 res = markFromRootsWork(asynch); 3584 _collectorState = FinalMarking; 3585 if (UseAdaptiveSizePolicy) { 3586 GenCollectedHeap* gch = GenCollectedHeap::heap(); 3587 size_policy()->ms_collection_marking_end(gch->gc_cause()); 3588 } 3589 } 3590 verify_overflow_empty(); 3591 return res; 3592 } 3593 3594 bool CMSCollector::markFromRootsWork(bool asynch) { 3595 // iterate over marked bits in bit map, doing a full scan and mark 3596 // from these roots using the following algorithm: 3597 // . if oop is to the right of the current scan pointer, 3598 // mark corresponding bit (we'll process it later) 3599 // . else (oop is to left of current scan pointer) 3600 // push oop on marking stack 3601 // . drain the marking stack 3602 3603 // Note that when we do a marking step we need to hold the 3604 // bit map lock -- recall that direct allocation (by mutators) 3605 // and promotion (by younger generation collectors) is also 3606 // marking the bit map. [the so-called allocate live policy.] 3607 // Because the implementation of bit map marking is not 3608 // robust wrt simultaneous marking of bits in the same word, 3609 // we need to make sure that there is no such interference 3610 // between concurrent such updates. 3611 3612 // already have locks 3613 assert_lock_strong(bitMapLock()); 3614 3615 // Clear the revisit stack, just in case there are any 3616 // obsolete contents from a short-circuited previous CMS cycle. 3617 _revisitStack.reset(); 3618 verify_work_stacks_empty(); 3619 verify_overflow_empty(); 3620 assert(_revisitStack.isEmpty(), "tabula rasa"); 3621 3622 bool result = false; 3623 if (CMSConcurrentMTEnabled && ParallelCMSThreads > 0) { 3624 result = do_marking_mt(asynch); 3625 } else { 3626 result = do_marking_st(asynch); 3627 } 3628 return result; 3629 } 3630 3631 // Forward decl 3632 class CMSConcMarkingTask; 3633 3634 class CMSConcMarkingTerminator: public ParallelTaskTerminator { 3635 CMSCollector* _collector; 3636 CMSConcMarkingTask* _task; 3637 bool _yield; 3638 protected: 3639 virtual void yield(); 3640 public: 3641 // "n_threads" is the number of threads to be terminated. 3642 // "queue_set" is a set of work queues of other threads. 3643 // "collector" is the CMS collector associated with this task terminator. 3644 // "yield" indicates whether we need the gang as a whole to yield. 3645 CMSConcMarkingTerminator(int n_threads, TaskQueueSetSuper* queue_set, 3646 CMSCollector* collector, bool yield) : 3647 ParallelTaskTerminator(n_threads, queue_set), 3648 _collector(collector), 3649 _yield(yield) { } 3650 3651 void set_task(CMSConcMarkingTask* task) { 3652 _task = task; 3653 } 3654 }; 3655 3656 // MT Concurrent Marking Task 3657 class CMSConcMarkingTask: public YieldingFlexibleGangTask { 3658 CMSCollector* _collector; 3659 YieldingFlexibleWorkGang* _workers; // the whole gang 3660 int _n_workers; // requested/desired # workers 3661 bool _asynch; 3662 bool _result; 3663 CompactibleFreeListSpace* _cms_space; 3664 CompactibleFreeListSpace* _perm_space; 3665 HeapWord* _global_finger; 3666 3667 // Exposed here for yielding support 3668 Mutex* const _bit_map_lock; 3669 3670 // The per thread work queues, available here for stealing 3671 OopTaskQueueSet* _task_queues; 3672 CMSConcMarkingTerminator _term; 3673 3674 public: 3675 CMSConcMarkingTask(CMSCollector* collector, 3676 CompactibleFreeListSpace* cms_space, 3677 CompactibleFreeListSpace* perm_space, 3678 bool asynch, int n_workers, 3679 YieldingFlexibleWorkGang* workers, 3680 OopTaskQueueSet* task_queues): 3681 YieldingFlexibleGangTask("Concurrent marking done multi-threaded"), 3682 _collector(collector), 3683 _cms_space(cms_space), 3684 _perm_space(perm_space), 3685 _asynch(asynch), _n_workers(n_workers), _result(true), 3686 _workers(workers), _task_queues(task_queues), 3687 _term(n_workers, task_queues, _collector, asynch), 3688 _bit_map_lock(collector->bitMapLock()) 3689 { 3690 assert(n_workers <= workers->total_workers(), 3691 "Else termination won't work correctly today"); // XXX FIX ME! 3692 _requested_size = n_workers; 3693 _term.set_task(this); 3694 assert(_cms_space->bottom() < _perm_space->bottom(), 3695 "Finger incorrectly initialized below"); 3696 _global_finger = _cms_space->bottom(); 3697 } 3698 3699 3700 OopTaskQueueSet* task_queues() { return _task_queues; } 3701 3702 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 3703 3704 HeapWord** global_finger_addr() { return &_global_finger; } 3705 3706 CMSConcMarkingTerminator* terminator() { return &_term; } 3707 3708 void work(int i); 3709 3710 virtual void coordinator_yield(); // stuff done by coordinator 3711 bool result() { return _result; } 3712 3713 void reset(HeapWord* ra) { 3714 _term.reset_for_reuse(); 3715 } 3716 3717 static bool get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 3718 OopTaskQueue* work_q); 3719 3720 private: 3721 void do_scan_and_mark(int i, CompactibleFreeListSpace* sp); 3722 void do_work_steal(int i); 3723 void bump_global_finger(HeapWord* f); 3724 }; 3725 3726 void CMSConcMarkingTerminator::yield() { 3727 if (ConcurrentMarkSweepThread::should_yield() && 3728 !_collector->foregroundGCIsActive() && 3729 _yield) { 3730 _task->yield(); 3731 } else { 3732 ParallelTaskTerminator::yield(); 3733 } 3734 } 3735 3736 //////////////////////////////////////////////////////////////// 3737 // Concurrent Marking Algorithm Sketch 3738 //////////////////////////////////////////////////////////////// 3739 // Until all tasks exhausted (both spaces): 3740 // -- claim next available chunk 3741 // -- bump global finger via CAS 3742 // -- find first object that starts in this chunk 3743 // and start scanning bitmap from that position 3744 // -- scan marked objects for oops 3745 // -- CAS-mark target, and if successful: 3746 // . if target oop is above global finger (volatile read) 3747 // nothing to do 3748 // . if target oop is in chunk and above local finger 3749 // then nothing to do 3750 // . else push on work-queue 3751 // -- Deal with possible overflow issues: 3752 // . local work-queue overflow causes stuff to be pushed on 3753 // global (common) overflow queue 3754 // . always first empty local work queue 3755 // . then get a batch of oops from global work queue if any 3756 // . then do work stealing 3757 // -- When all tasks claimed (both spaces) 3758 // and local work queue empty, 3759 // then in a loop do: 3760 // . check global overflow stack; steal a batch of oops and trace 3761 // . try to steal from other threads oif GOS is empty 3762 // . if neither is available, offer termination 3763 // -- Terminate and return result 3764 // 3765 void CMSConcMarkingTask::work(int i) { 3766 elapsedTimer _timer; 3767 ResourceMark rm; 3768 HandleMark hm; 3769 3770 DEBUG_ONLY(_collector->verify_overflow_empty();) 3771 3772 // Before we begin work, our work queue should be empty 3773 assert(work_queue(i)->size() == 0, "Expected to be empty"); 3774 // Scan the bitmap covering _cms_space, tracing through grey objects. 3775 _timer.start(); 3776 do_scan_and_mark(i, _cms_space); 3777 _timer.stop(); 3778 if (PrintCMSStatistics != 0) { 3779 gclog_or_tty->print_cr("Finished cms space scanning in %dth thread: %3.3f sec", 3780 i, _timer.seconds()); // XXX: need xxx/xxx type of notation, two timers 3781 } 3782 3783 // ... do the same for the _perm_space 3784 _timer.reset(); 3785 _timer.start(); 3786 do_scan_and_mark(i, _perm_space); 3787 _timer.stop(); 3788 if (PrintCMSStatistics != 0) { 3789 gclog_or_tty->print_cr("Finished perm space scanning in %dth thread: %3.3f sec", 3790 i, _timer.seconds()); // XXX: need xxx/xxx type of notation, two timers 3791 } 3792 3793 // ... do work stealing 3794 _timer.reset(); 3795 _timer.start(); 3796 do_work_steal(i); 3797 _timer.stop(); 3798 if (PrintCMSStatistics != 0) { 3799 gclog_or_tty->print_cr("Finished work stealing in %dth thread: %3.3f sec", 3800 i, _timer.seconds()); // XXX: need xxx/xxx type of notation, two timers 3801 } 3802 assert(_collector->_markStack.isEmpty(), "Should have been emptied"); 3803 assert(work_queue(i)->size() == 0, "Should have been emptied"); 3804 // Note that under the current task protocol, the 3805 // following assertion is true even of the spaces 3806 // expanded since the completion of the concurrent 3807 // marking. XXX This will likely change under a strict 3808 // ABORT semantics. 3809 assert(_global_finger > _cms_space->end() && 3810 _global_finger >= _perm_space->end(), 3811 "All tasks have been completed"); 3812 DEBUG_ONLY(_collector->verify_overflow_empty();) 3813 } 3814 3815 void CMSConcMarkingTask::bump_global_finger(HeapWord* f) { 3816 HeapWord* read = _global_finger; 3817 HeapWord* cur = read; 3818 while (f > read) { 3819 cur = read; 3820 read = (HeapWord*) Atomic::cmpxchg_ptr(f, &_global_finger, cur); 3821 if (cur == read) { 3822 // our cas succeeded 3823 assert(_global_finger >= f, "protocol consistency"); 3824 break; 3825 } 3826 } 3827 } 3828 3829 // This is really inefficient, and should be redone by 3830 // using (not yet available) block-read and -write interfaces to the 3831 // stack and the work_queue. XXX FIX ME !!! 3832 bool CMSConcMarkingTask::get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 3833 OopTaskQueue* work_q) { 3834 // Fast lock-free check 3835 if (ovflw_stk->length() == 0) { 3836 return false; 3837 } 3838 assert(work_q->size() == 0, "Shouldn't steal"); 3839 MutexLockerEx ml(ovflw_stk->par_lock(), 3840 Mutex::_no_safepoint_check_flag); 3841 // Grab up to 1/4 the size of the work queue 3842 size_t num = MIN2((size_t)work_q->max_elems()/4, 3843 (size_t)ParGCDesiredObjsFromOverflowList); 3844 num = MIN2(num, ovflw_stk->length()); 3845 for (int i = (int) num; i > 0; i--) { 3846 oop cur = ovflw_stk->pop(); 3847 assert(cur != NULL, "Counted wrong?"); 3848 work_q->push(cur); 3849 } 3850 return num > 0; 3851 } 3852 3853 void CMSConcMarkingTask::do_scan_and_mark(int i, CompactibleFreeListSpace* sp) { 3854 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 3855 int n_tasks = pst->n_tasks(); 3856 // We allow that there may be no tasks to do here because 3857 // we are restarting after a stack overflow. 3858 assert(pst->valid() || n_tasks == 0, "Uninitializd use?"); 3859 int nth_task = 0; 3860 3861 HeapWord* start = sp->bottom(); 3862 size_t chunk_size = sp->marking_task_size(); 3863 while (!pst->is_task_claimed(/* reference */ nth_task)) { 3864 // Having claimed the nth task in this space, 3865 // compute the chunk that it corresponds to: 3866 MemRegion span = MemRegion(start + nth_task*chunk_size, 3867 start + (nth_task+1)*chunk_size); 3868 // Try and bump the global finger via a CAS; 3869 // note that we need to do the global finger bump 3870 // _before_ taking the intersection below, because 3871 // the task corresponding to that region will be 3872 // deemed done even if the used_region() expands 3873 // because of allocation -- as it almost certainly will 3874 // during start-up while the threads yield in the 3875 // closure below. 3876 HeapWord* finger = span.end(); 3877 bump_global_finger(finger); // atomically 3878 // There are null tasks here corresponding to chunks 3879 // beyond the "top" address of the space. 3880 span = span.intersection(sp->used_region()); 3881 if (!span.is_empty()) { // Non-null task 3882 // We want to skip the first object because 3883 // the protocol is to scan any object in its entirety 3884 // that _starts_ in this span; a fortiori, any 3885 // object starting in an earlier span is scanned 3886 // as part of an earlier claimed task. 3887 // Below we use the "careful" version of block_start 3888 // so we do not try to navigate uninitialized objects. 3889 HeapWord* prev_obj = sp->block_start_careful(span.start()); 3890 // Below we use a variant of block_size that uses the 3891 // Printezis bits to avoid waiting for allocated 3892 // objects to become initialized/parsable. 3893 while (prev_obj < span.start()) { 3894 size_t sz = sp->block_size_no_stall(prev_obj, _collector); 3895 if (sz > 0) { 3896 prev_obj += sz; 3897 } else { 3898 // In this case we may end up doing a bit of redundant 3899 // scanning, but that appears unavoidable, short of 3900 // locking the free list locks; see bug 6324141. 3901 break; 3902 } 3903 } 3904 if (prev_obj < span.end()) { 3905 MemRegion my_span = MemRegion(prev_obj, span.end()); 3906 // Do the marking work within a non-empty span -- 3907 // the last argument to the constructor indicates whether the 3908 // iteration should be incremental with periodic yields. 3909 Par_MarkFromRootsClosure cl(this, _collector, my_span, 3910 &_collector->_markBitMap, 3911 work_queue(i), 3912 &_collector->_markStack, 3913 &_collector->_revisitStack, 3914 _asynch); 3915 _collector->_markBitMap.iterate(&cl, my_span.start(), my_span.end()); 3916 } // else nothing to do for this task 3917 } // else nothing to do for this task 3918 } 3919 // We'd be tempted to assert here that since there are no 3920 // more tasks left to claim in this space, the global_finger 3921 // must exceed space->top() and a fortiori space->end(). However, 3922 // that would not quite be correct because the bumping of 3923 // global_finger occurs strictly after the claiming of a task, 3924 // so by the time we reach here the global finger may not yet 3925 // have been bumped up by the thread that claimed the last 3926 // task. 3927 pst->all_tasks_completed(); 3928 } 3929 3930 class Par_ConcMarkingClosure: public OopClosure { 3931 CMSCollector* _collector; 3932 MemRegion _span; 3933 CMSBitMap* _bit_map; 3934 CMSMarkStack* _overflow_stack; 3935 CMSMarkStack* _revisit_stack; // XXXXXX Check proper use 3936 OopTaskQueue* _work_queue; 3937 3938 public: 3939 Par_ConcMarkingClosure(CMSCollector* collector, OopTaskQueue* work_queue, 3940 CMSBitMap* bit_map, CMSMarkStack* overflow_stack): 3941 _collector(collector), 3942 _span(_collector->_span), 3943 _work_queue(work_queue), 3944 _bit_map(bit_map), 3945 _overflow_stack(overflow_stack) { } // need to initialize revisit stack etc. 3946 3947 void do_oop(oop* p); 3948 void trim_queue(size_t max); 3949 void handle_stack_overflow(HeapWord* lost); 3950 }; 3951 3952 // Grey object rescan during work stealing phase -- 3953 // the salient assumption here is that stolen oops must 3954 // always be initialized, so we do not need to check for 3955 // uninitialized objects before scanning here. 3956 void Par_ConcMarkingClosure::do_oop(oop* p) { 3957 oop this_oop = *p; 3958 assert(this_oop->is_oop_or_null(), 3959 "expected an oop or NULL"); 3960 HeapWord* addr = (HeapWord*)this_oop; 3961 // Check if oop points into the CMS generation 3962 // and is not marked 3963 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 3964 // a white object ... 3965 // If we manage to "claim" the object, by being the 3966 // first thread to mark it, then we push it on our 3967 // marking stack 3968 if (_bit_map->par_mark(addr)) { // ... now grey 3969 // push on work queue (grey set) 3970 bool simulate_overflow = false; 3971 NOT_PRODUCT( 3972 if (CMSMarkStackOverflowALot && 3973 _collector->simulate_overflow()) { 3974 // simulate a stack overflow 3975 simulate_overflow = true; 3976 } 3977 ) 3978 if (simulate_overflow || 3979 !(_work_queue->push(this_oop) || _overflow_stack->par_push(this_oop))) { 3980 // stack overflow 3981 if (PrintCMSStatistics != 0) { 3982 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 3983 SIZE_FORMAT, _overflow_stack->capacity()); 3984 } 3985 // We cannot assert that the overflow stack is full because 3986 // it may have been emptied since. 3987 assert(simulate_overflow || 3988 _work_queue->size() == _work_queue->max_elems(), 3989 "Else push should have succeeded"); 3990 handle_stack_overflow(addr); 3991 } 3992 } // Else, some other thread got there first 3993 } 3994 } 3995 3996 void Par_ConcMarkingClosure::trim_queue(size_t max) { 3997 while (_work_queue->size() > max) { 3998 oop new_oop; 3999 if (_work_queue->pop_local(new_oop)) { 4000 assert(new_oop->is_oop(), "Should be an oop"); 4001 assert(_bit_map->isMarked((HeapWord*)new_oop), "Grey object"); 4002 assert(_span.contains((HeapWord*)new_oop), "Not in span"); 4003 assert(new_oop->is_parsable(), "Should be parsable"); 4004 new_oop->oop_iterate(this); // do_oop() above 4005 } 4006 } 4007 } 4008 4009 // Upon stack overflow, we discard (part of) the stack, 4010 // remembering the least address amongst those discarded 4011 // in CMSCollector's _restart_address. 4012 void Par_ConcMarkingClosure::handle_stack_overflow(HeapWord* lost) { 4013 // We need to do this under a mutex to prevent other 4014 // workers from interfering with the expansion below. 4015 MutexLockerEx ml(_overflow_stack->par_lock(), 4016 Mutex::_no_safepoint_check_flag); 4017 // Remember the least grey address discarded 4018 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 4019 _collector->lower_restart_addr(ra); 4020 _overflow_stack->reset(); // discard stack contents 4021 _overflow_stack->expand(); // expand the stack if possible 4022 } 4023 4024 4025 void CMSConcMarkingTask::do_work_steal(int i) { 4026 OopTaskQueue* work_q = work_queue(i); 4027 oop obj_to_scan; 4028 CMSBitMap* bm = &(_collector->_markBitMap); 4029 CMSMarkStack* ovflw = &(_collector->_markStack); 4030 int* seed = _collector->hash_seed(i); 4031 Par_ConcMarkingClosure cl(_collector, work_q, bm, ovflw); 4032 while (true) { 4033 cl.trim_queue(0); 4034 assert(work_q->size() == 0, "Should have been emptied above"); 4035 if (get_work_from_overflow_stack(ovflw, work_q)) { 4036 // Can't assert below because the work obtained from the 4037 // overflow stack may already have been stolen from us. 4038 // assert(work_q->size() > 0, "Work from overflow stack"); 4039 continue; 4040 } else if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 4041 assert(obj_to_scan->is_oop(), "Should be an oop"); 4042 assert(bm->isMarked((HeapWord*)obj_to_scan), "Grey object"); 4043 obj_to_scan->oop_iterate(&cl); 4044 } else if (terminator()->offer_termination()) { 4045 assert(work_q->size() == 0, "Impossible!"); 4046 break; 4047 } 4048 } 4049 } 4050 4051 // This is run by the CMS (coordinator) thread. 4052 void CMSConcMarkingTask::coordinator_yield() { 4053 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 4054 "CMS thread should hold CMS token"); 4055 4056 // First give up the locks, then yield, then re-lock 4057 // We should probably use a constructor/destructor idiom to 4058 // do this unlock/lock or modify the MutexUnlocker class to 4059 // serve our purpose. XXX 4060 assert_lock_strong(_bit_map_lock); 4061 _bit_map_lock->unlock(); 4062 ConcurrentMarkSweepThread::desynchronize(true); 4063 ConcurrentMarkSweepThread::acknowledge_yield_request(); 4064 _collector->stopTimer(); 4065 if (PrintCMSStatistics != 0) { 4066 _collector->incrementYields(); 4067 } 4068 _collector->icms_wait(); 4069 4070 // It is possible for whichever thread initiated the yield request 4071 // not to get a chance to wake up and take the bitmap lock between 4072 // this thread releasing it and reacquiring it. So, while the 4073 // should_yield() flag is on, let's sleep for a bit to give the 4074 // other thread a chance to wake up. The limit imposed on the number 4075 // of iterations is defensive, to avoid any unforseen circumstances 4076 // putting us into an infinite loop. Since it's always been this 4077 // (coordinator_yield()) method that was observed to cause the 4078 // problem, we are using a parameter (CMSCoordinatorYieldSleepCount) 4079 // which is by default non-zero. For the other seven methods that 4080 // also perform the yield operation, as are using a different 4081 // parameter (CMSYieldSleepCount) which is by default zero. This way we 4082 // can enable the sleeping for those methods too, if necessary. 4083 // See 6442774. 4084 // 4085 // We really need to reconsider the synchronization between the GC 4086 // thread and the yield-requesting threads in the future and we 4087 // should really use wait/notify, which is the recommended 4088 // way of doing this type of interaction. Additionally, we should 4089 // consolidate the eight methods that do the yield operation and they 4090 // are almost identical into one for better maintenability and 4091 // readability. See 6445193. 4092 // 4093 // Tony 2006.06.29 4094 for (unsigned i = 0; i < CMSCoordinatorYieldSleepCount && 4095 ConcurrentMarkSweepThread::should_yield() && 4096 !CMSCollector::foregroundGCIsActive(); ++i) { 4097 os::sleep(Thread::current(), 1, false); 4098 ConcurrentMarkSweepThread::acknowledge_yield_request(); 4099 } 4100 4101 ConcurrentMarkSweepThread::synchronize(true); 4102 _bit_map_lock->lock_without_safepoint_check(); 4103 _collector->startTimer(); 4104 } 4105 4106 bool CMSCollector::do_marking_mt(bool asynch) { 4107 assert(ParallelCMSThreads > 0 && conc_workers() != NULL, "precondition"); 4108 // In the future this would be determined ergonomically, based 4109 // on #cpu's, # active mutator threads (and load), and mutation rate. 4110 int num_workers = ParallelCMSThreads; 4111 4112 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 4113 CompactibleFreeListSpace* perm_space = _permGen->cmsSpace(); 4114 4115 CMSConcMarkingTask tsk(this, cms_space, perm_space, 4116 asynch, num_workers /* number requested XXX */, 4117 conc_workers(), task_queues()); 4118 4119 // Since the actual number of workers we get may be different 4120 // from the number we requested above, do we need to do anything different 4121 // below? In particular, may be we need to subclass the SequantialSubTasksDone 4122 // class?? XXX 4123 cms_space ->initialize_sequential_subtasks_for_marking(num_workers); 4124 perm_space->initialize_sequential_subtasks_for_marking(num_workers); 4125 4126 // Refs discovery is already non-atomic. 4127 assert(!ref_processor()->discovery_is_atomic(), "Should be non-atomic"); 4128 // Mutate the Refs discovery so it is MT during the 4129 // multi-threaded marking phase. 4130 ReferenceProcessorMTMutator mt(ref_processor(), num_workers > 1); 4131 4132 conc_workers()->start_task(&tsk); 4133 while (tsk.yielded()) { 4134 tsk.coordinator_yield(); 4135 conc_workers()->continue_task(&tsk); 4136 } 4137 // If the task was aborted, _restart_addr will be non-NULL 4138 assert(tsk.completed() || _restart_addr != NULL, "Inconsistency"); 4139 while (_restart_addr != NULL) { 4140 // XXX For now we do not make use of ABORTED state and have not 4141 // yet implemented the right abort semantics (even in the original 4142 // single-threaded CMS case). That needs some more investigation 4143 // and is deferred for now; see CR# TBF. 07252005YSR. XXX 4144 assert(!CMSAbortSemantics || tsk.aborted(), "Inconsistency"); 4145 // If _restart_addr is non-NULL, a marking stack overflow 4146 // occured; we need to do a fresh marking iteration from the 4147 // indicated restart address. 4148 if (_foregroundGCIsActive && asynch) { 4149 // We may be running into repeated stack overflows, having 4150 // reached the limit of the stack size, while making very 4151 // slow forward progress. It may be best to bail out and 4152 // let the foreground collector do its job. 4153 // Clear _restart_addr, so that foreground GC 4154 // works from scratch. This avoids the headache of 4155 // a "rescan" which would otherwise be needed because 4156 // of the dirty mod union table & card table. 4157 _restart_addr = NULL; 4158 return false; 4159 } 4160 // Adjust the task to restart from _restart_addr 4161 tsk.reset(_restart_addr); 4162 cms_space ->initialize_sequential_subtasks_for_marking(num_workers, 4163 _restart_addr); 4164 perm_space->initialize_sequential_subtasks_for_marking(num_workers, 4165 _restart_addr); 4166 _restart_addr = NULL; 4167 // Get the workers going again 4168 conc_workers()->start_task(&tsk); 4169 while (tsk.yielded()) { 4170 tsk.coordinator_yield(); 4171 conc_workers()->continue_task(&tsk); 4172 } 4173 } 4174 assert(tsk.completed(), "Inconsistency"); 4175 assert(tsk.result() == true, "Inconsistency"); 4176 return true; 4177 } 4178 4179 bool CMSCollector::do_marking_st(bool asynch) { 4180 ResourceMark rm; 4181 HandleMark hm; 4182 4183 MarkFromRootsClosure markFromRootsClosure(this, _span, &_markBitMap, 4184 &_markStack, &_revisitStack, CMSYield && asynch); 4185 // the last argument to iterate indicates whether the iteration 4186 // should be incremental with periodic yields. 4187 _markBitMap.iterate(&markFromRootsClosure); 4188 // If _restart_addr is non-NULL, a marking stack overflow 4189 // occured; we need to do a fresh iteration from the 4190 // indicated restart address. 4191 while (_restart_addr != NULL) { 4192 if (_foregroundGCIsActive && asynch) { 4193 // We may be running into repeated stack overflows, having 4194 // reached the limit of the stack size, while making very 4195 // slow forward progress. It may be best to bail out and 4196 // let the foreground collector do its job. 4197 // Clear _restart_addr, so that foreground GC 4198 // works from scratch. This avoids the headache of 4199 // a "rescan" which would otherwise be needed because 4200 // of the dirty mod union table & card table. 4201 _restart_addr = NULL; 4202 return false; // indicating failure to complete marking 4203 } 4204 // Deal with stack overflow: 4205 // we restart marking from _restart_addr 4206 HeapWord* ra = _restart_addr; 4207 markFromRootsClosure.reset(ra); 4208 _restart_addr = NULL; 4209 _markBitMap.iterate(&markFromRootsClosure, ra, _span.end()); 4210 } 4211 return true; 4212 } 4213 4214 void CMSCollector::preclean() { 4215 check_correct_thread_executing(); 4216 assert(Thread::current()->is_ConcurrentGC_thread(), "Wrong thread"); 4217 verify_work_stacks_empty(); 4218 verify_overflow_empty(); 4219 _abort_preclean = false; 4220 if (CMSPrecleaningEnabled) { 4221 _eden_chunk_index = 0; 4222 size_t used = get_eden_used(); 4223 size_t capacity = get_eden_capacity(); 4224 // Don't start sampling unless we will get sufficiently 4225 // many samples. 4226 if (used < (capacity/(CMSScheduleRemarkSamplingRatio * 100) 4227 * CMSScheduleRemarkEdenPenetration)) { 4228 _start_sampling = true; 4229 } else { 4230 _start_sampling = false; 4231 } 4232 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 4233 CMSPhaseAccounting pa(this, "preclean", !PrintGCDetails); 4234 preclean_work(CMSPrecleanRefLists1, CMSPrecleanSurvivors1); 4235 } 4236 CMSTokenSync x(true); // is cms thread 4237 if (CMSPrecleaningEnabled) { 4238 sample_eden(); 4239 _collectorState = AbortablePreclean; 4240 } else { 4241 _collectorState = FinalMarking; 4242 } 4243 verify_work_stacks_empty(); 4244 verify_overflow_empty(); 4245 } 4246 4247 // Try and schedule the remark such that young gen 4248 // occupancy is CMSScheduleRemarkEdenPenetration %. 4249 void CMSCollector::abortable_preclean() { 4250 check_correct_thread_executing(); 4251 assert(CMSPrecleaningEnabled, "Inconsistent control state"); 4252 assert(_collectorState == AbortablePreclean, "Inconsistent control state"); 4253 4254 // If Eden's current occupancy is below this threshold, 4255 // immediately schedule the remark; else preclean 4256 // past the next scavenge in an effort to 4257 // schedule the pause as described avove. By choosing 4258 // CMSScheduleRemarkEdenSizeThreshold >= max eden size 4259 // we will never do an actual abortable preclean cycle. 4260 if (get_eden_used() > CMSScheduleRemarkEdenSizeThreshold) { 4261 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 4262 CMSPhaseAccounting pa(this, "abortable-preclean", !PrintGCDetails); 4263 // We need more smarts in the abortable preclean 4264 // loop below to deal with cases where allocation 4265 // in young gen is very very slow, and our precleaning 4266 // is running a losing race against a horde of 4267 // mutators intent on flooding us with CMS updates 4268 // (dirty cards). 4269 // One, admittedly dumb, strategy is to give up 4270 // after a certain number of abortable precleaning loops 4271 // or after a certain maximum time. We want to make 4272 // this smarter in the next iteration. 4273 // XXX FIX ME!!! YSR 4274 size_t loops = 0, workdone = 0, cumworkdone = 0, waited = 0; 4275 while (!(should_abort_preclean() || 4276 ConcurrentMarkSweepThread::should_terminate())) { 4277 workdone = preclean_work(CMSPrecleanRefLists2, CMSPrecleanSurvivors2); 4278 cumworkdone += workdone; 4279 loops++; 4280 // Voluntarily terminate abortable preclean phase if we have 4281 // been at it for too long. 4282 if ((CMSMaxAbortablePrecleanLoops != 0) && 4283 loops >= CMSMaxAbortablePrecleanLoops) { 4284 if (PrintGCDetails) { 4285 gclog_or_tty->print(" CMS: abort preclean due to loops "); 4286 } 4287 break; 4288 } 4289 if (pa.wallclock_millis() > CMSMaxAbortablePrecleanTime) { 4290 if (PrintGCDetails) { 4291 gclog_or_tty->print(" CMS: abort preclean due to time "); 4292 } 4293 break; 4294 } 4295 // If we are doing little work each iteration, we should 4296 // take a short break. 4297 if (workdone < CMSAbortablePrecleanMinWorkPerIteration) { 4298 // Sleep for some time, waiting for work to accumulate 4299 stopTimer(); 4300 cmsThread()->wait_on_cms_lock(CMSAbortablePrecleanWaitMillis); 4301 startTimer(); 4302 waited++; 4303 } 4304 } 4305 if (PrintCMSStatistics > 0) { 4306 gclog_or_tty->print(" [%d iterations, %d waits, %d cards)] ", 4307 loops, waited, cumworkdone); 4308 } 4309 } 4310 CMSTokenSync x(true); // is cms thread 4311 if (_collectorState != Idling) { 4312 assert(_collectorState == AbortablePreclean, 4313 "Spontaneous state transition?"); 4314 _collectorState = FinalMarking; 4315 } // Else, a foreground collection completed this CMS cycle. 4316 return; 4317 } 4318 4319 // Respond to an Eden sampling opportunity 4320 void CMSCollector::sample_eden() { 4321 // Make sure a young gc cannot sneak in between our 4322 // reading and recording of a sample. 4323 assert(Thread::current()->is_ConcurrentGC_thread(), 4324 "Only the cms thread may collect Eden samples"); 4325 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 4326 "Should collect samples while holding CMS token"); 4327 if (!_start_sampling) { 4328 return; 4329 } 4330 if (_eden_chunk_array) { 4331 if (_eden_chunk_index < _eden_chunk_capacity) { 4332 _eden_chunk_array[_eden_chunk_index] = *_top_addr; // take sample 4333 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr, 4334 "Unexpected state of Eden"); 4335 // We'd like to check that what we just sampled is an oop-start address; 4336 // however, we cannot do that here since the object may not yet have been 4337 // initialized. So we'll instead do the check when we _use_ this sample 4338 // later. 4339 if (_eden_chunk_index == 0 || 4340 (pointer_delta(_eden_chunk_array[_eden_chunk_index], 4341 _eden_chunk_array[_eden_chunk_index-1]) 4342 >= CMSSamplingGrain)) { 4343 _eden_chunk_index++; // commit sample 4344 } 4345 } 4346 } 4347 if ((_collectorState == AbortablePreclean) && !_abort_preclean) { 4348 size_t used = get_eden_used(); 4349 size_t capacity = get_eden_capacity(); 4350 assert(used <= capacity, "Unexpected state of Eden"); 4351 if (used > (capacity/100 * CMSScheduleRemarkEdenPenetration)) { 4352 _abort_preclean = true; 4353 } 4354 } 4355 } 4356 4357 4358 size_t CMSCollector::preclean_work(bool clean_refs, bool clean_survivor) { 4359 assert(_collectorState == Precleaning || 4360 _collectorState == AbortablePreclean, "incorrect state"); 4361 ResourceMark rm; 4362 HandleMark hm; 4363 // Do one pass of scrubbing the discovered reference lists 4364 // to remove any reference objects with strongly-reachable 4365 // referents. 4366 if (clean_refs) { 4367 ReferenceProcessor* rp = ref_processor(); 4368 CMSPrecleanRefsYieldClosure yield_cl(this); 4369 assert(rp->span().equals(_span), "Spans should be equal"); 4370 CMSKeepAliveClosure keep_alive(this, _span, &_markBitMap, 4371 &_markStack); 4372 CMSDrainMarkingStackClosure complete_trace(this, 4373 _span, &_markBitMap, &_markStack, 4374 &keep_alive); 4375 4376 // We don't want this step to interfere with a young 4377 // collection because we don't want to take CPU 4378 // or memory bandwidth away from the young GC threads 4379 // (which may be as many as there are CPUs). 4380 // Note that we don't need to protect ourselves from 4381 // interference with mutators because they can't 4382 // manipulate the discovered reference lists nor affect 4383 // the computed reachability of the referents, the 4384 // only properties manipulated by the precleaning 4385 // of these reference lists. 4386 stopTimer(); 4387 CMSTokenSyncWithLocks x(true /* is cms thread */, 4388 bitMapLock()); 4389 startTimer(); 4390 sample_eden(); 4391 // The following will yield to allow foreground 4392 // collection to proceed promptly. XXX YSR: 4393 // The code in this method may need further 4394 // tweaking for better performance and some restructuring 4395 // for cleaner interfaces. 4396 rp->preclean_discovered_references( 4397 rp->is_alive_non_header(), &keep_alive, &complete_trace, 4398 &yield_cl); 4399 } 4400 4401 if (clean_survivor) { // preclean the active survivor space(s) 4402 assert(_young_gen->kind() == Generation::DefNew || 4403 _young_gen->kind() == Generation::ParNew || 4404 _young_gen->kind() == Generation::ASParNew, 4405 "incorrect type for cast"); 4406 DefNewGeneration* dng = (DefNewGeneration*)_young_gen; 4407 PushAndMarkClosure pam_cl(this, _span, ref_processor(), 4408 &_markBitMap, &_modUnionTable, 4409 &_markStack, &_revisitStack, 4410 true /* precleaning phase */); 4411 stopTimer(); 4412 CMSTokenSyncWithLocks ts(true /* is cms thread */, 4413 bitMapLock()); 4414 startTimer(); 4415 unsigned int before_count = 4416 GenCollectedHeap::heap()->total_collections(); 4417 SurvivorSpacePrecleanClosure 4418 sss_cl(this, _span, &_markBitMap, &_markStack, 4419 &pam_cl, before_count, CMSYield); 4420 dng->from()->object_iterate_careful(&sss_cl); 4421 dng->to()->object_iterate_careful(&sss_cl); 4422 } 4423 MarkRefsIntoAndScanClosure 4424 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable, 4425 &_markStack, &_revisitStack, this, CMSYield, 4426 true /* precleaning phase */); 4427 // CAUTION: The following closure has persistent state that may need to 4428 // be reset upon a decrease in the sequence of addresses it 4429 // processes. 4430 ScanMarkedObjectsAgainCarefullyClosure 4431 smoac_cl(this, _span, 4432 &_markBitMap, &_markStack, &_revisitStack, &mrias_cl, CMSYield); 4433 4434 // Preclean dirty cards in ModUnionTable and CardTable using 4435 // appropriate convergence criterion; 4436 // repeat CMSPrecleanIter times unless we find that 4437 // we are losing. 4438 assert(CMSPrecleanIter < 10, "CMSPrecleanIter is too large"); 4439 assert(CMSPrecleanNumerator < CMSPrecleanDenominator, 4440 "Bad convergence multiplier"); 4441 assert(CMSPrecleanThreshold >= 100, 4442 "Unreasonably low CMSPrecleanThreshold"); 4443 4444 size_t numIter, cumNumCards, lastNumCards, curNumCards; 4445 for (numIter = 0, cumNumCards = lastNumCards = curNumCards = 0; 4446 numIter < CMSPrecleanIter; 4447 numIter++, lastNumCards = curNumCards, cumNumCards += curNumCards) { 4448 curNumCards = preclean_mod_union_table(_cmsGen, &smoac_cl); 4449 if (CMSPermGenPrecleaningEnabled) { 4450 curNumCards += preclean_mod_union_table(_permGen, &smoac_cl); 4451 } 4452 if (Verbose && PrintGCDetails) { 4453 gclog_or_tty->print(" (modUnionTable: %d cards)", curNumCards); 4454 } 4455 // Either there are very few dirty cards, so re-mark 4456 // pause will be small anyway, or our pre-cleaning isn't 4457 // that much faster than the rate at which cards are being 4458 // dirtied, so we might as well stop and re-mark since 4459 // precleaning won't improve our re-mark time by much. 4460 if (curNumCards <= CMSPrecleanThreshold || 4461 (numIter > 0 && 4462 (curNumCards * CMSPrecleanDenominator > 4463 lastNumCards * CMSPrecleanNumerator))) { 4464 numIter++; 4465 cumNumCards += curNumCards; 4466 break; 4467 } 4468 } 4469 curNumCards = preclean_card_table(_cmsGen, &smoac_cl); 4470 if (CMSPermGenPrecleaningEnabled) { 4471 curNumCards += preclean_card_table(_permGen, &smoac_cl); 4472 } 4473 cumNumCards += curNumCards; 4474 if (PrintGCDetails && PrintCMSStatistics != 0) { 4475 gclog_or_tty->print_cr(" (cardTable: %d cards, re-scanned %d cards, %d iterations)", 4476 curNumCards, cumNumCards, numIter); 4477 } 4478 return cumNumCards; // as a measure of useful work done 4479 } 4480 4481 // PRECLEANING NOTES: 4482 // Precleaning involves: 4483 // . reading the bits of the modUnionTable and clearing the set bits. 4484 // . For the cards corresponding to the set bits, we scan the 4485 // objects on those cards. This means we need the free_list_lock 4486 // so that we can safely iterate over the CMS space when scanning 4487 // for oops. 4488 // . When we scan the objects, we'll be both reading and setting 4489 // marks in the marking bit map, so we'll need the marking bit map. 4490 // . For protecting _collector_state transitions, we take the CGC_lock. 4491 // Note that any races in the reading of of card table entries by the 4492 // CMS thread on the one hand and the clearing of those entries by the 4493 // VM thread or the setting of those entries by the mutator threads on the 4494 // other are quite benign. However, for efficiency it makes sense to keep 4495 // the VM thread from racing with the CMS thread while the latter is 4496 // dirty card info to the modUnionTable. We therefore also use the 4497 // CGC_lock to protect the reading of the card table and the mod union 4498 // table by the CM thread. 4499 // . We run concurrently with mutator updates, so scanning 4500 // needs to be done carefully -- we should not try to scan 4501 // potentially uninitialized objects. 4502 // 4503 // Locking strategy: While holding the CGC_lock, we scan over and 4504 // reset a maximal dirty range of the mod union / card tables, then lock 4505 // the free_list_lock and bitmap lock to do a full marking, then 4506 // release these locks; and repeat the cycle. This allows for a 4507 // certain amount of fairness in the sharing of these locks between 4508 // the CMS collector on the one hand, and the VM thread and the 4509 // mutators on the other. 4510 4511 // NOTE: preclean_mod_union_table() and preclean_card_table() 4512 // further below are largely identical; if you need to modify 4513 // one of these methods, please check the other method too. 4514 4515 size_t CMSCollector::preclean_mod_union_table( 4516 ConcurrentMarkSweepGeneration* gen, 4517 ScanMarkedObjectsAgainCarefullyClosure* cl) { 4518 verify_work_stacks_empty(); 4519 verify_overflow_empty(); 4520 4521 // strategy: starting with the first card, accumulate contiguous 4522 // ranges of dirty cards; clear these cards, then scan the region 4523 // covered by these cards. 4524 4525 // Since all of the MUT is committed ahead, we can just use 4526 // that, in case the generations expand while we are precleaning. 4527 // It might also be fine to just use the committed part of the 4528 // generation, but we might potentially miss cards when the 4529 // generation is rapidly expanding while we are in the midst 4530 // of precleaning. 4531 HeapWord* startAddr = gen->reserved().start(); 4532 HeapWord* endAddr = gen->reserved().end(); 4533 4534 cl->setFreelistLock(gen->freelistLock()); // needed for yielding 4535 4536 size_t numDirtyCards, cumNumDirtyCards; 4537 HeapWord *nextAddr, *lastAddr; 4538 for (cumNumDirtyCards = numDirtyCards = 0, 4539 nextAddr = lastAddr = startAddr; 4540 nextAddr < endAddr; 4541 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 4542 4543 ResourceMark rm; 4544 HandleMark hm; 4545 4546 MemRegion dirtyRegion; 4547 { 4548 stopTimer(); 4549 CMSTokenSync ts(true); 4550 startTimer(); 4551 sample_eden(); 4552 // Get dirty region starting at nextOffset (inclusive), 4553 // simultaneously clearing it. 4554 dirtyRegion = 4555 _modUnionTable.getAndClearMarkedRegion(nextAddr, endAddr); 4556 assert(dirtyRegion.start() >= nextAddr, 4557 "returned region inconsistent?"); 4558 } 4559 // Remember where the next search should begin. 4560 // The returned region (if non-empty) is a right open interval, 4561 // so lastOffset is obtained from the right end of that 4562 // interval. 4563 lastAddr = dirtyRegion.end(); 4564 // Should do something more transparent and less hacky XXX 4565 numDirtyCards = 4566 _modUnionTable.heapWordDiffToOffsetDiff(dirtyRegion.word_size()); 4567 4568 // We'll scan the cards in the dirty region (with periodic 4569 // yields for foreground GC as needed). 4570 if (!dirtyRegion.is_empty()) { 4571 assert(numDirtyCards > 0, "consistency check"); 4572 HeapWord* stop_point = NULL; 4573 { 4574 stopTimer(); 4575 CMSTokenSyncWithLocks ts(true, gen->freelistLock(), 4576 bitMapLock()); 4577 startTimer(); 4578 verify_work_stacks_empty(); 4579 verify_overflow_empty(); 4580 sample_eden(); 4581 stop_point = 4582 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 4583 } 4584 if (stop_point != NULL) { 4585 // The careful iteration stopped early either because it found an 4586 // uninitialized object, or because we were in the midst of an 4587 // "abortable preclean", which should now be aborted. Redirty 4588 // the bits corresponding to the partially-scanned or unscanned 4589 // cards. We'll either restart at the next block boundary or 4590 // abort the preclean. 4591 assert((CMSPermGenPrecleaningEnabled && (gen == _permGen)) || 4592 (_collectorState == AbortablePreclean && should_abort_preclean()), 4593 "Unparsable objects should only be in perm gen."); 4594 4595 stopTimer(); 4596 CMSTokenSyncWithLocks ts(true, bitMapLock()); 4597 startTimer(); 4598 _modUnionTable.mark_range(MemRegion(stop_point, dirtyRegion.end())); 4599 if (should_abort_preclean()) { 4600 break; // out of preclean loop 4601 } else { 4602 // Compute the next address at which preclean should pick up; 4603 // might need bitMapLock in order to read P-bits. 4604 lastAddr = next_card_start_after_block(stop_point); 4605 } 4606 } 4607 } else { 4608 assert(lastAddr == endAddr, "consistency check"); 4609 assert(numDirtyCards == 0, "consistency check"); 4610 break; 4611 } 4612 } 4613 verify_work_stacks_empty(); 4614 verify_overflow_empty(); 4615 return cumNumDirtyCards; 4616 } 4617 4618 // NOTE: preclean_mod_union_table() above and preclean_card_table() 4619 // below are largely identical; if you need to modify 4620 // one of these methods, please check the other method too. 4621 4622 size_t CMSCollector::preclean_card_table(ConcurrentMarkSweepGeneration* gen, 4623 ScanMarkedObjectsAgainCarefullyClosure* cl) { 4624 // strategy: it's similar to precleamModUnionTable above, in that 4625 // we accumulate contiguous ranges of dirty cards, mark these cards 4626 // precleaned, then scan the region covered by these cards. 4627 HeapWord* endAddr = (HeapWord*)(gen->_virtual_space.high()); 4628 HeapWord* startAddr = (HeapWord*)(gen->_virtual_space.low()); 4629 4630 cl->setFreelistLock(gen->freelistLock()); // needed for yielding 4631 4632 size_t numDirtyCards, cumNumDirtyCards; 4633 HeapWord *lastAddr, *nextAddr; 4634 4635 for (cumNumDirtyCards = numDirtyCards = 0, 4636 nextAddr = lastAddr = startAddr; 4637 nextAddr < endAddr; 4638 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 4639 4640 ResourceMark rm; 4641 HandleMark hm; 4642 4643 MemRegion dirtyRegion; 4644 { 4645 // See comments in "Precleaning notes" above on why we 4646 // do this locking. XXX Could the locking overheads be 4647 // too high when dirty cards are sparse? [I don't think so.] 4648 stopTimer(); 4649 CMSTokenSync x(true); // is cms thread 4650 startTimer(); 4651 sample_eden(); 4652 // Get and clear dirty region from card table 4653 dirtyRegion = _ct->ct_bs()->dirty_card_range_after_preclean( 4654 MemRegion(nextAddr, endAddr)); 4655 assert(dirtyRegion.start() >= nextAddr, 4656 "returned region inconsistent?"); 4657 } 4658 lastAddr = dirtyRegion.end(); 4659 numDirtyCards = 4660 dirtyRegion.word_size()/CardTableModRefBS::card_size_in_words; 4661 4662 if (!dirtyRegion.is_empty()) { 4663 stopTimer(); 4664 CMSTokenSyncWithLocks ts(true, gen->freelistLock(), bitMapLock()); 4665 startTimer(); 4666 sample_eden(); 4667 verify_work_stacks_empty(); 4668 verify_overflow_empty(); 4669 HeapWord* stop_point = 4670 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 4671 if (stop_point != NULL) { 4672 // The careful iteration stopped early because it found an 4673 // uninitialized object. Redirty the bits corresponding to the 4674 // partially-scanned or unscanned cards, and start again at the 4675 // next block boundary. 4676 assert(CMSPermGenPrecleaningEnabled || 4677 (_collectorState == AbortablePreclean && should_abort_preclean()), 4678 "Unparsable objects should only be in perm gen."); 4679 _ct->ct_bs()->invalidate(MemRegion(stop_point, dirtyRegion.end())); 4680 if (should_abort_preclean()) { 4681 break; // out of preclean loop 4682 } else { 4683 // Compute the next address at which preclean should pick up. 4684 lastAddr = next_card_start_after_block(stop_point); 4685 } 4686 } 4687 } else { 4688 break; 4689 } 4690 } 4691 verify_work_stacks_empty(); 4692 verify_overflow_empty(); 4693 return cumNumDirtyCards; 4694 } 4695 4696 void CMSCollector::checkpointRootsFinal(bool asynch, 4697 bool clear_all_soft_refs, bool init_mark_was_synchronous) { 4698 assert(_collectorState == FinalMarking, "incorrect state transition?"); 4699 check_correct_thread_executing(); 4700 // world is stopped at this checkpoint 4701 assert(SafepointSynchronize::is_at_safepoint(), 4702 "world should be stopped"); 4703 verify_work_stacks_empty(); 4704 verify_overflow_empty(); 4705 4706 SpecializationStats::clear(); 4707 if (PrintGCDetails) { 4708 gclog_or_tty->print("[YG occupancy: "SIZE_FORMAT" K ("SIZE_FORMAT" K)]", 4709 _young_gen->used() / K, 4710 _young_gen->capacity() / K); 4711 } 4712 if (asynch) { 4713 if (CMSScavengeBeforeRemark) { 4714 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4715 // Temporarily set flag to false, GCH->do_collection will 4716 // expect it to be false and set to true 4717 FlagSetting fl(gch->_is_gc_active, false); 4718 NOT_PRODUCT(TraceTime t("Scavenge-Before-Remark", 4719 PrintGCDetails && Verbose, true, gclog_or_tty);) 4720 int level = _cmsGen->level() - 1; 4721 if (level >= 0) { 4722 gch->do_collection(true, // full (i.e. force, see below) 4723 false, // !clear_all_soft_refs 4724 0, // size 4725 false, // is_tlab 4726 level // max_level 4727 ); 4728 } 4729 } 4730 FreelistLocker x(this); 4731 MutexLockerEx y(bitMapLock(), 4732 Mutex::_no_safepoint_check_flag); 4733 assert(!init_mark_was_synchronous, "but that's impossible!"); 4734 checkpointRootsFinalWork(asynch, clear_all_soft_refs, false); 4735 } else { 4736 // already have all the locks 4737 checkpointRootsFinalWork(asynch, clear_all_soft_refs, 4738 init_mark_was_synchronous); 4739 } 4740 verify_work_stacks_empty(); 4741 verify_overflow_empty(); 4742 SpecializationStats::print(); 4743 } 4744 4745 void CMSCollector::checkpointRootsFinalWork(bool asynch, 4746 bool clear_all_soft_refs, bool init_mark_was_synchronous) { 4747 4748 NOT_PRODUCT(TraceTime tr("checkpointRootsFinalWork", PrintGCDetails, false, gclog_or_tty);) 4749 4750 assert(haveFreelistLocks(), "must have free list locks"); 4751 assert_lock_strong(bitMapLock()); 4752 4753 if (UseAdaptiveSizePolicy) { 4754 size_policy()->checkpoint_roots_final_begin(); 4755 } 4756 4757 ResourceMark rm; 4758 HandleMark hm; 4759 4760 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4761 4762 if (should_unload_classes()) { 4763 CodeCache::gc_prologue(); 4764 } 4765 assert(haveFreelistLocks(), "must have free list locks"); 4766 assert_lock_strong(bitMapLock()); 4767 4768 if (!init_mark_was_synchronous) { 4769 // We might assume that we need not fill TLAB's when 4770 // CMSScavengeBeforeRemark is set, because we may have just done 4771 // a scavenge which would have filled all TLAB's -- and besides 4772 // Eden would be empty. This however may not always be the case -- 4773 // for instance although we asked for a scavenge, it may not have 4774 // happened because of a JNI critical section. We probably need 4775 // a policy for deciding whether we can in that case wait until 4776 // the critical section releases and then do the remark following 4777 // the scavenge, and skip it here. In the absence of that policy, 4778 // or of an indication of whether the scavenge did indeed occur, 4779 // we cannot rely on TLAB's having been filled and must do 4780 // so here just in case a scavenge did not happen. 4781 gch->ensure_parsability(false); // fill TLAB's, but no need to retire them 4782 // Update the saved marks which may affect the root scans. 4783 gch->save_marks(); 4784 4785 { 4786 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;) 4787 4788 // Note on the role of the mod union table: 4789 // Since the marker in "markFromRoots" marks concurrently with 4790 // mutators, it is possible for some reachable objects not to have been 4791 // scanned. For instance, an only reference to an object A was 4792 // placed in object B after the marker scanned B. Unless B is rescanned, 4793 // A would be collected. Such updates to references in marked objects 4794 // are detected via the mod union table which is the set of all cards 4795 // dirtied since the first checkpoint in this GC cycle and prior to 4796 // the most recent young generation GC, minus those cleaned up by the 4797 // concurrent precleaning. 4798 if (CMSParallelRemarkEnabled && ParallelGCThreads > 0) { 4799 TraceTime t("Rescan (parallel) ", PrintGCDetails, false, gclog_or_tty); 4800 do_remark_parallel(); 4801 } else { 4802 TraceTime t("Rescan (non-parallel) ", PrintGCDetails, false, 4803 gclog_or_tty); 4804 do_remark_non_parallel(); 4805 } 4806 } 4807 } else { 4808 assert(!asynch, "Can't have init_mark_was_synchronous in asynch mode"); 4809 // The initial mark was stop-world, so there's no rescanning to 4810 // do; go straight on to the next step below. 4811 } 4812 verify_work_stacks_empty(); 4813 verify_overflow_empty(); 4814 4815 { 4816 NOT_PRODUCT(TraceTime ts("refProcessingWork", PrintGCDetails, false, gclog_or_tty);) 4817 refProcessingWork(asynch, clear_all_soft_refs); 4818 } 4819 verify_work_stacks_empty(); 4820 verify_overflow_empty(); 4821 4822 if (should_unload_classes()) { 4823 CodeCache::gc_epilogue(); 4824 } 4825 4826 // If we encountered any (marking stack / work queue) overflow 4827 // events during the current CMS cycle, take appropriate 4828 // remedial measures, where possible, so as to try and avoid 4829 // recurrence of that condition. 4830 assert(_markStack.isEmpty(), "No grey objects"); 4831 size_t ser_ovflw = _ser_pmc_remark_ovflw + _ser_pmc_preclean_ovflw + 4832 _ser_kac_ovflw; 4833 if (ser_ovflw > 0) { 4834 if (PrintCMSStatistics != 0) { 4835 gclog_or_tty->print_cr("Marking stack overflow (benign) " 4836 "(pmc_pc="SIZE_FORMAT", pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT")", 4837 _ser_pmc_preclean_ovflw, _ser_pmc_remark_ovflw, 4838 _ser_kac_ovflw); 4839 } 4840 _markStack.expand(); 4841 _ser_pmc_remark_ovflw = 0; 4842 _ser_pmc_preclean_ovflw = 0; 4843 _ser_kac_ovflw = 0; 4844 } 4845 if (_par_pmc_remark_ovflw > 0 || _par_kac_ovflw > 0) { 4846 if (PrintCMSStatistics != 0) { 4847 gclog_or_tty->print_cr("Work queue overflow (benign) " 4848 "(pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT")", 4849 _par_pmc_remark_ovflw, _par_kac_ovflw); 4850 } 4851 _par_pmc_remark_ovflw = 0; 4852 _par_kac_ovflw = 0; 4853 } 4854 if (PrintCMSStatistics != 0) { 4855 if (_markStack._hit_limit > 0) { 4856 gclog_or_tty->print_cr(" (benign) Hit max stack size limit ("SIZE_FORMAT")", 4857 _markStack._hit_limit); 4858 } 4859 if (_markStack._failed_double > 0) { 4860 gclog_or_tty->print_cr(" (benign) Failed stack doubling ("SIZE_FORMAT")," 4861 " current capacity "SIZE_FORMAT, 4862 _markStack._failed_double, 4863 _markStack.capacity()); 4864 } 4865 } 4866 _markStack._hit_limit = 0; 4867 _markStack._failed_double = 0; 4868 4869 if ((VerifyAfterGC || VerifyDuringGC) && 4870 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 4871 verify_after_remark(); 4872 } 4873 4874 // Change under the freelistLocks. 4875 _collectorState = Sweeping; 4876 // Call isAllClear() under bitMapLock 4877 assert(_modUnionTable.isAllClear(), "Should be clear by end of the" 4878 " final marking"); 4879 if (UseAdaptiveSizePolicy) { 4880 size_policy()->checkpoint_roots_final_end(gch->gc_cause()); 4881 } 4882 } 4883 4884 // Parallel remark task 4885 class CMSParRemarkTask: public AbstractGangTask { 4886 CMSCollector* _collector; 4887 WorkGang* _workers; 4888 int _n_workers; 4889 CompactibleFreeListSpace* _cms_space; 4890 CompactibleFreeListSpace* _perm_space; 4891 4892 // The per-thread work queues, available here for stealing. 4893 OopTaskQueueSet* _task_queues; 4894 ParallelTaskTerminator _term; 4895 4896 public: 4897 CMSParRemarkTask(CMSCollector* collector, 4898 CompactibleFreeListSpace* cms_space, 4899 CompactibleFreeListSpace* perm_space, 4900 int n_workers, WorkGang* workers, 4901 OopTaskQueueSet* task_queues): 4902 AbstractGangTask("Rescan roots and grey objects in parallel"), 4903 _collector(collector), 4904 _cms_space(cms_space), _perm_space(perm_space), 4905 _n_workers(n_workers), 4906 _workers(workers), 4907 _task_queues(task_queues), 4908 _term(workers->total_workers(), task_queues) { } 4909 4910 OopTaskQueueSet* task_queues() { return _task_queues; } 4911 4912 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 4913 4914 ParallelTaskTerminator* terminator() { return &_term; } 4915 4916 void work(int i); 4917 4918 private: 4919 // Work method in support of parallel rescan ... of young gen spaces 4920 void do_young_space_rescan(int i, Par_MarkRefsIntoAndScanClosure* cl, 4921 ContiguousSpace* space, 4922 HeapWord** chunk_array, size_t chunk_top); 4923 4924 // ... of dirty cards in old space 4925 void do_dirty_card_rescan_tasks(CompactibleFreeListSpace* sp, int i, 4926 Par_MarkRefsIntoAndScanClosure* cl); 4927 4928 // ... work stealing for the above 4929 void do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, int* seed); 4930 }; 4931 4932 void CMSParRemarkTask::work(int i) { 4933 elapsedTimer _timer; 4934 ResourceMark rm; 4935 HandleMark hm; 4936 4937 // ---------- rescan from roots -------------- 4938 _timer.start(); 4939 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4940 Par_MarkRefsIntoAndScanClosure par_mrias_cl(_collector, 4941 _collector->_span, _collector->ref_processor(), 4942 &(_collector->_markBitMap), 4943 work_queue(i), &(_collector->_revisitStack)); 4944 4945 // Rescan young gen roots first since these are likely 4946 // coarsely partitioned and may, on that account, constitute 4947 // the critical path; thus, it's best to start off that 4948 // work first. 4949 // ---------- young gen roots -------------- 4950 { 4951 DefNewGeneration* dng = _collector->_young_gen->as_DefNewGeneration(); 4952 EdenSpace* eden_space = dng->eden(); 4953 ContiguousSpace* from_space = dng->from(); 4954 ContiguousSpace* to_space = dng->to(); 4955 4956 HeapWord** eca = _collector->_eden_chunk_array; 4957 size_t ect = _collector->_eden_chunk_index; 4958 HeapWord** sca = _collector->_survivor_chunk_array; 4959 size_t sct = _collector->_survivor_chunk_index; 4960 4961 assert(ect <= _collector->_eden_chunk_capacity, "out of bounds"); 4962 assert(sct <= _collector->_survivor_chunk_capacity, "out of bounds"); 4963 4964 do_young_space_rescan(i, &par_mrias_cl, to_space, NULL, 0); 4965 do_young_space_rescan(i, &par_mrias_cl, from_space, sca, sct); 4966 do_young_space_rescan(i, &par_mrias_cl, eden_space, eca, ect); 4967 4968 _timer.stop(); 4969 if (PrintCMSStatistics != 0) { 4970 gclog_or_tty->print_cr( 4971 "Finished young gen rescan work in %dth thread: %3.3f sec", 4972 i, _timer.seconds()); 4973 } 4974 } 4975 4976 // ---------- remaining roots -------------- 4977 _timer.reset(); 4978 _timer.start(); 4979 gch->gen_process_strong_roots(_collector->_cmsGen->level(), 4980 false, // yg was scanned above 4981 true, // collecting perm gen 4982 SharedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()), 4983 NULL, &par_mrias_cl); 4984 _timer.stop(); 4985 if (PrintCMSStatistics != 0) { 4986 gclog_or_tty->print_cr( 4987 "Finished remaining root rescan work in %dth thread: %3.3f sec", 4988 i, _timer.seconds()); 4989 } 4990 4991 // ---------- rescan dirty cards ------------ 4992 _timer.reset(); 4993 _timer.start(); 4994 4995 // Do the rescan tasks for each of the two spaces 4996 // (cms_space and perm_space) in turn. 4997 do_dirty_card_rescan_tasks(_cms_space, i, &par_mrias_cl); 4998 do_dirty_card_rescan_tasks(_perm_space, i, &par_mrias_cl); 4999 _timer.stop(); 5000 if (PrintCMSStatistics != 0) { 5001 gclog_or_tty->print_cr( 5002 "Finished dirty card rescan work in %dth thread: %3.3f sec", 5003 i, _timer.seconds()); 5004 } 5005 5006 // ---------- steal work from other threads ... 5007 // ---------- ... and drain overflow list. 5008 _timer.reset(); 5009 _timer.start(); 5010 do_work_steal(i, &par_mrias_cl, _collector->hash_seed(i)); 5011 _timer.stop(); 5012 if (PrintCMSStatistics != 0) { 5013 gclog_or_tty->print_cr( 5014 "Finished work stealing in %dth thread: %3.3f sec", 5015 i, _timer.seconds()); 5016 } 5017 } 5018 5019 void 5020 CMSParRemarkTask::do_young_space_rescan(int i, 5021 Par_MarkRefsIntoAndScanClosure* cl, ContiguousSpace* space, 5022 HeapWord** chunk_array, size_t chunk_top) { 5023 // Until all tasks completed: 5024 // . claim an unclaimed task 5025 // . compute region boundaries corresponding to task claimed 5026 // using chunk_array 5027 // . par_oop_iterate(cl) over that region 5028 5029 ResourceMark rm; 5030 HandleMark hm; 5031 5032 SequentialSubTasksDone* pst = space->par_seq_tasks(); 5033 assert(pst->valid(), "Uninitialized use?"); 5034 5035 int nth_task = 0; 5036 int n_tasks = pst->n_tasks(); 5037 5038 HeapWord *start, *end; 5039 while (!pst->is_task_claimed(/* reference */ nth_task)) { 5040 // We claimed task # nth_task; compute its boundaries. 5041 if (chunk_top == 0) { // no samples were taken 5042 assert(nth_task == 0 && n_tasks == 1, "Can have only 1 EdenSpace task"); 5043 start = space->bottom(); 5044 end = space->top(); 5045 } else if (nth_task == 0) { 5046 start = space->bottom(); 5047 end = chunk_array[nth_task]; 5048 } else if (nth_task < (jint)chunk_top) { 5049 assert(nth_task >= 1, "Control point invariant"); 5050 start = chunk_array[nth_task - 1]; 5051 end = chunk_array[nth_task]; 5052 } else { 5053 assert(nth_task == (jint)chunk_top, "Control point invariant"); 5054 start = chunk_array[chunk_top - 1]; 5055 end = space->top(); 5056 } 5057 MemRegion mr(start, end); 5058 // Verify that mr is in space 5059 assert(mr.is_empty() || space->used_region().contains(mr), 5060 "Should be in space"); 5061 // Verify that "start" is an object boundary 5062 assert(mr.is_empty() || oop(mr.start())->is_oop(), 5063 "Should be an oop"); 5064 space->par_oop_iterate(mr, cl); 5065 } 5066 pst->all_tasks_completed(); 5067 } 5068 5069 void 5070 CMSParRemarkTask::do_dirty_card_rescan_tasks( 5071 CompactibleFreeListSpace* sp, int i, 5072 Par_MarkRefsIntoAndScanClosure* cl) { 5073 // Until all tasks completed: 5074 // . claim an unclaimed task 5075 // . compute region boundaries corresponding to task claimed 5076 // . transfer dirty bits ct->mut for that region 5077 // . apply rescanclosure to dirty mut bits for that region 5078 5079 ResourceMark rm; 5080 HandleMark hm; 5081 5082 OopTaskQueue* work_q = work_queue(i); 5083 ModUnionClosure modUnionClosure(&(_collector->_modUnionTable)); 5084 // CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! 5085 // CAUTION: This closure has state that persists across calls to 5086 // the work method dirty_range_iterate_clear() in that it has 5087 // imbedded in it a (subtype of) UpwardsObjectClosure. The 5088 // use of that state in the imbedded UpwardsObjectClosure instance 5089 // assumes that the cards are always iterated (even if in parallel 5090 // by several threads) in monotonically increasing order per each 5091 // thread. This is true of the implementation below which picks 5092 // card ranges (chunks) in monotonically increasing order globally 5093 // and, a-fortiori, in monotonically increasing order per thread 5094 // (the latter order being a subsequence of the former). 5095 // If the work code below is ever reorganized into a more chaotic 5096 // work-partitioning form than the current "sequential tasks" 5097 // paradigm, the use of that persistent state will have to be 5098 // revisited and modified appropriately. See also related 5099 // bug 4756801 work on which should examine this code to make 5100 // sure that the changes there do not run counter to the 5101 // assumptions made here and necessary for correctness and 5102 // efficiency. Note also that this code might yield inefficient 5103 // behaviour in the case of very large objects that span one or 5104 // more work chunks. Such objects would potentially be scanned 5105 // several times redundantly. Work on 4756801 should try and 5106 // address that performance anomaly if at all possible. XXX 5107 MemRegion full_span = _collector->_span; 5108 CMSBitMap* bm = &(_collector->_markBitMap); // shared 5109 CMSMarkStack* rs = &(_collector->_revisitStack); // shared 5110 MarkFromDirtyCardsClosure 5111 greyRescanClosure(_collector, full_span, // entire span of interest 5112 sp, bm, work_q, rs, cl); 5113 5114 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 5115 assert(pst->valid(), "Uninitialized use?"); 5116 int nth_task = 0; 5117 const int alignment = CardTableModRefBS::card_size * BitsPerWord; 5118 MemRegion span = sp->used_region(); 5119 HeapWord* start_addr = span.start(); 5120 HeapWord* end_addr = (HeapWord*)round_to((intptr_t)span.end(), 5121 alignment); 5122 const size_t chunk_size = sp->rescan_task_size(); // in HeapWord units 5123 assert((HeapWord*)round_to((intptr_t)start_addr, alignment) == 5124 start_addr, "Check alignment"); 5125 assert((size_t)round_to((intptr_t)chunk_size, alignment) == 5126 chunk_size, "Check alignment"); 5127 5128 while (!pst->is_task_claimed(/* reference */ nth_task)) { 5129 // Having claimed the nth_task, compute corresponding mem-region, 5130 // which is a-fortiori aligned correctly (i.e. at a MUT bopundary). 5131 // The alignment restriction ensures that we do not need any 5132 // synchronization with other gang-workers while setting or 5133 // clearing bits in thus chunk of the MUT. 5134 MemRegion this_span = MemRegion(start_addr + nth_task*chunk_size, 5135 start_addr + (nth_task+1)*chunk_size); 5136 // The last chunk's end might be way beyond end of the 5137 // used region. In that case pull back appropriately. 5138 if (this_span.end() > end_addr) { 5139 this_span.set_end(end_addr); 5140 assert(!this_span.is_empty(), "Program logic (calculation of n_tasks)"); 5141 } 5142 // Iterate over the dirty cards covering this chunk, marking them 5143 // precleaned, and setting the corresponding bits in the mod union 5144 // table. Since we have been careful to partition at Card and MUT-word 5145 // boundaries no synchronization is needed between parallel threads. 5146 _collector->_ct->ct_bs()->dirty_card_iterate(this_span, 5147 &modUnionClosure); 5148 5149 // Having transferred these marks into the modUnionTable, 5150 // rescan the marked objects on the dirty cards in the modUnionTable. 5151 // Even if this is at a synchronous collection, the initial marking 5152 // may have been done during an asynchronous collection so there 5153 // may be dirty bits in the mod-union table. 5154 _collector->_modUnionTable.dirty_range_iterate_clear( 5155 this_span, &greyRescanClosure); 5156 _collector->_modUnionTable.verifyNoOneBitsInRange( 5157 this_span.start(), 5158 this_span.end()); 5159 } 5160 pst->all_tasks_completed(); // declare that i am done 5161 } 5162 5163 // . see if we can share work_queues with ParNew? XXX 5164 void 5165 CMSParRemarkTask::do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, 5166 int* seed) { 5167 OopTaskQueue* work_q = work_queue(i); 5168 NOT_PRODUCT(int num_steals = 0;) 5169 oop obj_to_scan; 5170 CMSBitMap* bm = &(_collector->_markBitMap); 5171 size_t num_from_overflow_list = 5172 MIN2((size_t)work_q->max_elems()/4, 5173 (size_t)ParGCDesiredObjsFromOverflowList); 5174 5175 while (true) { 5176 // Completely finish any left over work from (an) earlier round(s) 5177 cl->trim_queue(0); 5178 // Now check if there's any work in the overflow list 5179 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 5180 work_q)) { 5181 // found something in global overflow list; 5182 // not yet ready to go stealing work from others. 5183 // We'd like to assert(work_q->size() != 0, ...) 5184 // because we just took work from the overflow list, 5185 // but of course we can't since all of that could have 5186 // been already stolen from us. 5187 // "He giveth and He taketh away." 5188 continue; 5189 } 5190 // Verify that we have no work before we resort to stealing 5191 assert(work_q->size() == 0, "Have work, shouldn't steal"); 5192 // Try to steal from other queues that have work 5193 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 5194 NOT_PRODUCT(num_steals++;) 5195 assert(obj_to_scan->is_oop(), "Oops, not an oop!"); 5196 assert(bm->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 5197 // Do scanning work 5198 obj_to_scan->oop_iterate(cl); 5199 // Loop around, finish this work, and try to steal some more 5200 } else if (terminator()->offer_termination()) { 5201 break; // nirvana from the infinite cycle 5202 } 5203 } 5204 NOT_PRODUCT( 5205 if (PrintCMSStatistics != 0) { 5206 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals); 5207 } 5208 ) 5209 assert(work_q->size() == 0 && _collector->overflow_list_is_empty(), 5210 "Else our work is not yet done"); 5211 } 5212 5213 // Return a thread-local PLAB recording array, as appropriate. 5214 void* CMSCollector::get_data_recorder(int thr_num) { 5215 if (_survivor_plab_array != NULL && 5216 (CMSPLABRecordAlways || 5217 (_collectorState > Marking && _collectorState < FinalMarking))) { 5218 assert(thr_num < (int)ParallelGCThreads, "thr_num is out of bounds"); 5219 ChunkArray* ca = &_survivor_plab_array[thr_num]; 5220 ca->reset(); // clear it so that fresh data is recorded 5221 return (void*) ca; 5222 } else { 5223 return NULL; 5224 } 5225 } 5226 5227 // Reset all the thread-local PLAB recording arrays 5228 void CMSCollector::reset_survivor_plab_arrays() { 5229 for (uint i = 0; i < ParallelGCThreads; i++) { 5230 _survivor_plab_array[i].reset(); 5231 } 5232 } 5233 5234 // Merge the per-thread plab arrays into the global survivor chunk 5235 // array which will provide the partitioning of the survivor space 5236 // for CMS rescan. 5237 void CMSCollector::merge_survivor_plab_arrays(ContiguousSpace* surv) { 5238 assert(_survivor_plab_array != NULL, "Error"); 5239 assert(_survivor_chunk_array != NULL, "Error"); 5240 assert(_collectorState == FinalMarking, "Error"); 5241 for (uint j = 0; j < ParallelGCThreads; j++) { 5242 _cursor[j] = 0; 5243 } 5244 HeapWord* top = surv->top(); 5245 size_t i; 5246 for (i = 0; i < _survivor_chunk_capacity; i++) { // all sca entries 5247 HeapWord* min_val = top; // Higher than any PLAB address 5248 uint min_tid = 0; // position of min_val this round 5249 for (uint j = 0; j < ParallelGCThreads; j++) { 5250 ChunkArray* cur_sca = &_survivor_plab_array[j]; 5251 if (_cursor[j] == cur_sca->end()) { 5252 continue; 5253 } 5254 assert(_cursor[j] < cur_sca->end(), "ctl pt invariant"); 5255 HeapWord* cur_val = cur_sca->nth(_cursor[j]); 5256 assert(surv->used_region().contains(cur_val), "Out of bounds value"); 5257 if (cur_val < min_val) { 5258 min_tid = j; 5259 min_val = cur_val; 5260 } else { 5261 assert(cur_val < top, "All recorded addresses should be less"); 5262 } 5263 } 5264 // At this point min_val and min_tid are respectively 5265 // the least address in _survivor_plab_array[j]->nth(_cursor[j]) 5266 // and the thread (j) that witnesses that address. 5267 // We record this address in the _survivor_chunk_array[i] 5268 // and increment _cursor[min_tid] prior to the next round i. 5269 if (min_val == top) { 5270 break; 5271 } 5272 _survivor_chunk_array[i] = min_val; 5273 _cursor[min_tid]++; 5274 } 5275 // We are all done; record the size of the _survivor_chunk_array 5276 _survivor_chunk_index = i; // exclusive: [0, i) 5277 if (PrintCMSStatistics > 0) { 5278 gclog_or_tty->print(" (Survivor:" SIZE_FORMAT "chunks) ", i); 5279 } 5280 // Verify that we used up all the recorded entries 5281 #ifdef ASSERT 5282 size_t total = 0; 5283 for (uint j = 0; j < ParallelGCThreads; j++) { 5284 assert(_cursor[j] == _survivor_plab_array[j].end(), "Ctl pt invariant"); 5285 total += _cursor[j]; 5286 } 5287 assert(total == _survivor_chunk_index, "Ctl Pt Invariant"); 5288 // Check that the merged array is in sorted order 5289 if (total > 0) { 5290 for (size_t i = 0; i < total - 1; i++) { 5291 if (PrintCMSStatistics > 0) { 5292 gclog_or_tty->print(" (chunk" SIZE_FORMAT ":" INTPTR_FORMAT ") ", 5293 i, _survivor_chunk_array[i]); 5294 } 5295 assert(_survivor_chunk_array[i] < _survivor_chunk_array[i+1], 5296 "Not sorted"); 5297 } 5298 } 5299 #endif // ASSERT 5300 } 5301 5302 // Set up the space's par_seq_tasks structure for work claiming 5303 // for parallel rescan of young gen. 5304 // See ParRescanTask where this is currently used. 5305 void 5306 CMSCollector:: 5307 initialize_sequential_subtasks_for_young_gen_rescan(int n_threads) { 5308 assert(n_threads > 0, "Unexpected n_threads argument"); 5309 DefNewGeneration* dng = (DefNewGeneration*)_young_gen; 5310 5311 // Eden space 5312 { 5313 SequentialSubTasksDone* pst = dng->eden()->par_seq_tasks(); 5314 assert(!pst->valid(), "Clobbering existing data?"); 5315 // Each valid entry in [0, _eden_chunk_index) represents a task. 5316 size_t n_tasks = _eden_chunk_index + 1; 5317 assert(n_tasks == 1 || _eden_chunk_array != NULL, "Error"); 5318 pst->set_par_threads(n_threads); 5319 pst->set_n_tasks((int)n_tasks); 5320 } 5321 5322 // Merge the survivor plab arrays into _survivor_chunk_array 5323 if (_survivor_plab_array != NULL) { 5324 merge_survivor_plab_arrays(dng->from()); 5325 } else { 5326 assert(_survivor_chunk_index == 0, "Error"); 5327 } 5328 5329 // To space 5330 { 5331 SequentialSubTasksDone* pst = dng->to()->par_seq_tasks(); 5332 assert(!pst->valid(), "Clobbering existing data?"); 5333 pst->set_par_threads(n_threads); 5334 pst->set_n_tasks(1); 5335 assert(pst->valid(), "Error"); 5336 } 5337 5338 // From space 5339 { 5340 SequentialSubTasksDone* pst = dng->from()->par_seq_tasks(); 5341 assert(!pst->valid(), "Clobbering existing data?"); 5342 size_t n_tasks = _survivor_chunk_index + 1; 5343 assert(n_tasks == 1 || _survivor_chunk_array != NULL, "Error"); 5344 pst->set_par_threads(n_threads); 5345 pst->set_n_tasks((int)n_tasks); 5346 assert(pst->valid(), "Error"); 5347 } 5348 } 5349 5350 // Parallel version of remark 5351 void CMSCollector::do_remark_parallel() { 5352 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5353 WorkGang* workers = gch->workers(); 5354 assert(workers != NULL, "Need parallel worker threads."); 5355 int n_workers = workers->total_workers(); 5356 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 5357 CompactibleFreeListSpace* perm_space = _permGen->cmsSpace(); 5358 5359 CMSParRemarkTask tsk(this, 5360 cms_space, perm_space, 5361 n_workers, workers, task_queues()); 5362 5363 // Set up for parallel process_strong_roots work. 5364 gch->set_par_threads(n_workers); 5365 gch->change_strong_roots_parity(); 5366 // We won't be iterating over the cards in the card table updating 5367 // the younger_gen cards, so we shouldn't call the following else 5368 // the verification code as well as subsequent younger_refs_iterate 5369 // code would get confused. XXX 5370 // gch->rem_set()->prepare_for_younger_refs_iterate(true); // parallel 5371 5372 // The young gen rescan work will not be done as part of 5373 // process_strong_roots (which currently doesn't knw how to 5374 // parallelize such a scan), but rather will be broken up into 5375 // a set of parallel tasks (via the sampling that the [abortable] 5376 // preclean phase did of EdenSpace, plus the [two] tasks of 5377 // scanning the [two] survivor spaces. Further fine-grain 5378 // parallelization of the scanning of the survivor spaces 5379 // themselves, and of precleaning of the younger gen itself 5380 // is deferred to the future. 5381 initialize_sequential_subtasks_for_young_gen_rescan(n_workers); 5382 5383 // The dirty card rescan work is broken up into a "sequence" 5384 // of parallel tasks (per constituent space) that are dynamically 5385 // claimed by the parallel threads. 5386 cms_space->initialize_sequential_subtasks_for_rescan(n_workers); 5387 perm_space->initialize_sequential_subtasks_for_rescan(n_workers); 5388 5389 // It turns out that even when we're using 1 thread, doing the work in a 5390 // separate thread causes wide variance in run times. We can't help this 5391 // in the multi-threaded case, but we special-case n=1 here to get 5392 // repeatable measurements of the 1-thread overhead of the parallel code. 5393 if (n_workers > 1) { 5394 // Make refs discovery MT-safe 5395 ReferenceProcessorMTMutator mt(ref_processor(), true); 5396 workers->run_task(&tsk); 5397 } else { 5398 tsk.work(0); 5399 } 5400 gch->set_par_threads(0); // 0 ==> non-parallel. 5401 // restore, single-threaded for now, any preserved marks 5402 // as a result of work_q overflow 5403 restore_preserved_marks_if_any(); 5404 } 5405 5406 // Non-parallel version of remark 5407 void CMSCollector::do_remark_non_parallel() { 5408 ResourceMark rm; 5409 HandleMark hm; 5410 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5411 MarkRefsIntoAndScanClosure 5412 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable, 5413 &_markStack, &_revisitStack, this, 5414 false /* should_yield */, false /* not precleaning */); 5415 MarkFromDirtyCardsClosure 5416 markFromDirtyCardsClosure(this, _span, 5417 NULL, // space is set further below 5418 &_markBitMap, &_markStack, &_revisitStack, 5419 &mrias_cl); 5420 { 5421 TraceTime t("grey object rescan", PrintGCDetails, false, gclog_or_tty); 5422 // Iterate over the dirty cards, marking them precleaned, and 5423 // setting the corresponding bits in the mod union table. 5424 { 5425 ModUnionClosure modUnionClosure(&_modUnionTable); 5426 _ct->ct_bs()->dirty_card_iterate( 5427 _cmsGen->used_region(), 5428 &modUnionClosure); 5429 _ct->ct_bs()->dirty_card_iterate( 5430 _permGen->used_region(), 5431 &modUnionClosure); 5432 } 5433 // Having transferred these marks into the modUnionTable, we just need 5434 // to rescan the marked objects on the dirty cards in the modUnionTable. 5435 // The initial marking may have been done during an asynchronous 5436 // collection so there may be dirty bits in the mod-union table. 5437 const int alignment = 5438 CardTableModRefBS::card_size * BitsPerWord; 5439 { 5440 // ... First handle dirty cards in CMS gen 5441 markFromDirtyCardsClosure.set_space(_cmsGen->cmsSpace()); 5442 MemRegion ur = _cmsGen->used_region(); 5443 HeapWord* lb = ur.start(); 5444 HeapWord* ub = (HeapWord*)round_to((intptr_t)ur.end(), alignment); 5445 MemRegion cms_span(lb, ub); 5446 _modUnionTable.dirty_range_iterate_clear(cms_span, 5447 &markFromDirtyCardsClosure); 5448 verify_work_stacks_empty(); 5449 if (PrintCMSStatistics != 0) { 5450 gclog_or_tty->print(" (re-scanned "SIZE_FORMAT" dirty cards in cms gen) ", 5451 markFromDirtyCardsClosure.num_dirty_cards()); 5452 } 5453 } 5454 { 5455 // .. and then repeat for dirty cards in perm gen 5456 markFromDirtyCardsClosure.set_space(_permGen->cmsSpace()); 5457 MemRegion ur = _permGen->used_region(); 5458 HeapWord* lb = ur.start(); 5459 HeapWord* ub = (HeapWord*)round_to((intptr_t)ur.end(), alignment); 5460 MemRegion perm_span(lb, ub); 5461 _modUnionTable.dirty_range_iterate_clear(perm_span, 5462 &markFromDirtyCardsClosure); 5463 verify_work_stacks_empty(); 5464 if (PrintCMSStatistics != 0) { 5465 gclog_or_tty->print(" (re-scanned "SIZE_FORMAT" dirty cards in perm gen) ", 5466 markFromDirtyCardsClosure.num_dirty_cards()); 5467 } 5468 } 5469 } 5470 if (VerifyDuringGC && 5471 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 5472 HandleMark hm; // Discard invalid handles created during verification 5473 Universe::verify(true); 5474 } 5475 { 5476 TraceTime t("root rescan", PrintGCDetails, false, gclog_or_tty); 5477 5478 verify_work_stacks_empty(); 5479 5480 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 5481 gch->gen_process_strong_roots(_cmsGen->level(), 5482 true, // younger gens as roots 5483 true, // collecting perm gen 5484 SharedHeap::ScanningOption(roots_scanning_options()), 5485 NULL, &mrias_cl); 5486 } 5487 verify_work_stacks_empty(); 5488 // Restore evacuated mark words, if any, used for overflow list links 5489 if (!CMSOverflowEarlyRestoration) { 5490 restore_preserved_marks_if_any(); 5491 } 5492 verify_overflow_empty(); 5493 } 5494 5495 //////////////////////////////////////////////////////// 5496 // Parallel Reference Processing Task Proxy Class 5497 //////////////////////////////////////////////////////// 5498 class CMSRefProcTaskProxy: public AbstractGangTask { 5499 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask; 5500 CMSCollector* _collector; 5501 CMSBitMap* _mark_bit_map; 5502 MemRegion _span; 5503 OopTaskQueueSet* _task_queues; 5504 ParallelTaskTerminator _term; 5505 ProcessTask& _task; 5506 5507 public: 5508 CMSRefProcTaskProxy(ProcessTask& task, 5509 CMSCollector* collector, 5510 const MemRegion& span, 5511 CMSBitMap* mark_bit_map, 5512 int total_workers, 5513 OopTaskQueueSet* task_queues): 5514 AbstractGangTask("Process referents by policy in parallel"), 5515 _task(task), 5516 _collector(collector), _span(span), _mark_bit_map(mark_bit_map), 5517 _task_queues(task_queues), 5518 _term(total_workers, task_queues) 5519 { } 5520 5521 OopTaskQueueSet* task_queues() { return _task_queues; } 5522 5523 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 5524 5525 ParallelTaskTerminator* terminator() { return &_term; } 5526 5527 void do_work_steal(int i, 5528 CMSParDrainMarkingStackClosure* drain, 5529 CMSParKeepAliveClosure* keep_alive, 5530 int* seed); 5531 5532 virtual void work(int i); 5533 }; 5534 5535 void CMSRefProcTaskProxy::work(int i) { 5536 CMSParKeepAliveClosure par_keep_alive(_collector, _span, 5537 _mark_bit_map, work_queue(i)); 5538 CMSParDrainMarkingStackClosure par_drain_stack(_collector, _span, 5539 _mark_bit_map, work_queue(i)); 5540 CMSIsAliveClosure is_alive_closure(_mark_bit_map); 5541 _task.work(i, is_alive_closure, par_keep_alive, par_drain_stack); 5542 if (_task.marks_oops_alive()) { 5543 do_work_steal(i, &par_drain_stack, &par_keep_alive, 5544 _collector->hash_seed(i)); 5545 } 5546 assert(work_queue(i)->size() == 0, "work_queue should be empty"); 5547 assert(_collector->_overflow_list == NULL, "non-empty _overflow_list"); 5548 } 5549 5550 class CMSRefEnqueueTaskProxy: public AbstractGangTask { 5551 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask; 5552 EnqueueTask& _task; 5553 5554 public: 5555 CMSRefEnqueueTaskProxy(EnqueueTask& task) 5556 : AbstractGangTask("Enqueue reference objects in parallel"), 5557 _task(task) 5558 { } 5559 5560 virtual void work(int i) 5561 { 5562 _task.work(i); 5563 } 5564 }; 5565 5566 CMSParKeepAliveClosure::CMSParKeepAliveClosure(CMSCollector* collector, 5567 MemRegion span, CMSBitMap* bit_map, OopTaskQueue* work_queue): 5568 _collector(collector), 5569 _span(span), 5570 _bit_map(bit_map), 5571 _work_queue(work_queue), 5572 _mark_and_push(collector, span, bit_map, work_queue), 5573 _low_water_mark(MIN2((uint)(work_queue->max_elems()/4), 5574 (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads))) 5575 { } 5576 5577 // . see if we can share work_queues with ParNew? XXX 5578 void CMSRefProcTaskProxy::do_work_steal(int i, 5579 CMSParDrainMarkingStackClosure* drain, 5580 CMSParKeepAliveClosure* keep_alive, 5581 int* seed) { 5582 OopTaskQueue* work_q = work_queue(i); 5583 NOT_PRODUCT(int num_steals = 0;) 5584 oop obj_to_scan; 5585 size_t num_from_overflow_list = 5586 MIN2((size_t)work_q->max_elems()/4, 5587 (size_t)ParGCDesiredObjsFromOverflowList); 5588 5589 while (true) { 5590 // Completely finish any left over work from (an) earlier round(s) 5591 drain->trim_queue(0); 5592 // Now check if there's any work in the overflow list 5593 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 5594 work_q)) { 5595 // Found something in global overflow list; 5596 // not yet ready to go stealing work from others. 5597 // We'd like to assert(work_q->size() != 0, ...) 5598 // because we just took work from the overflow list, 5599 // but of course we can't, since all of that might have 5600 // been already stolen from us. 5601 continue; 5602 } 5603 // Verify that we have no work before we resort to stealing 5604 assert(work_q->size() == 0, "Have work, shouldn't steal"); 5605 // Try to steal from other queues that have work 5606 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 5607 NOT_PRODUCT(num_steals++;) 5608 assert(obj_to_scan->is_oop(), "Oops, not an oop!"); 5609 assert(_mark_bit_map->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 5610 // Do scanning work 5611 obj_to_scan->oop_iterate(keep_alive); 5612 // Loop around, finish this work, and try to steal some more 5613 } else if (terminator()->offer_termination()) { 5614 break; // nirvana from the infinite cycle 5615 } 5616 } 5617 NOT_PRODUCT( 5618 if (PrintCMSStatistics != 0) { 5619 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals); 5620 } 5621 ) 5622 } 5623 5624 void CMSRefProcTaskExecutor::execute(ProcessTask& task) 5625 { 5626 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5627 WorkGang* workers = gch->workers(); 5628 assert(workers != NULL, "Need parallel worker threads."); 5629 int n_workers = workers->total_workers(); 5630 CMSRefProcTaskProxy rp_task(task, &_collector, 5631 _collector.ref_processor()->span(), 5632 _collector.markBitMap(), 5633 n_workers, _collector.task_queues()); 5634 workers->run_task(&rp_task); 5635 } 5636 5637 void CMSRefProcTaskExecutor::execute(EnqueueTask& task) 5638 { 5639 5640 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5641 WorkGang* workers = gch->workers(); 5642 assert(workers != NULL, "Need parallel worker threads."); 5643 CMSRefEnqueueTaskProxy enq_task(task); 5644 workers->run_task(&enq_task); 5645 } 5646 5647 void CMSCollector::refProcessingWork(bool asynch, bool clear_all_soft_refs) { 5648 5649 ResourceMark rm; 5650 HandleMark hm; 5651 ReferencePolicy* soft_ref_policy; 5652 5653 assert(!ref_processor()->enqueuing_is_done(), "Enqueuing should not be complete"); 5654 // Process weak references. 5655 if (clear_all_soft_refs) { 5656 soft_ref_policy = new AlwaysClearPolicy(); 5657 } else { 5658 #ifdef COMPILER2 5659 soft_ref_policy = new LRUMaxHeapPolicy(); 5660 #else 5661 soft_ref_policy = new LRUCurrentHeapPolicy(); 5662 #endif // COMPILER2 5663 } 5664 verify_work_stacks_empty(); 5665 5666 ReferenceProcessor* rp = ref_processor(); 5667 assert(rp->span().equals(_span), "Spans should be equal"); 5668 CMSKeepAliveClosure cmsKeepAliveClosure(this, _span, &_markBitMap, 5669 &_markStack); 5670 CMSDrainMarkingStackClosure cmsDrainMarkingStackClosure(this, 5671 _span, &_markBitMap, &_markStack, 5672 &cmsKeepAliveClosure); 5673 { 5674 TraceTime t("weak refs processing", PrintGCDetails, false, gclog_or_tty); 5675 if (rp->processing_is_mt()) { 5676 CMSRefProcTaskExecutor task_executor(*this); 5677 rp->process_discovered_references(soft_ref_policy, 5678 &_is_alive_closure, 5679 &cmsKeepAliveClosure, 5680 &cmsDrainMarkingStackClosure, 5681 &task_executor); 5682 } else { 5683 rp->process_discovered_references(soft_ref_policy, 5684 &_is_alive_closure, 5685 &cmsKeepAliveClosure, 5686 &cmsDrainMarkingStackClosure, 5687 NULL); 5688 } 5689 verify_work_stacks_empty(); 5690 } 5691 5692 if (should_unload_classes()) { 5693 { 5694 TraceTime t("class unloading", PrintGCDetails, false, gclog_or_tty); 5695 5696 // Follow SystemDictionary roots and unload classes 5697 bool purged_class = SystemDictionary::do_unloading(&_is_alive_closure); 5698 5699 // Follow CodeCache roots and unload any methods marked for unloading 5700 CodeCache::do_unloading(&_is_alive_closure, 5701 &cmsKeepAliveClosure, 5702 purged_class); 5703 5704 cmsDrainMarkingStackClosure.do_void(); 5705 verify_work_stacks_empty(); 5706 5707 // Update subklass/sibling/implementor links in KlassKlass descendants 5708 assert(!_revisitStack.isEmpty(), "revisit stack should not be empty"); 5709 oop k; 5710 while ((k = _revisitStack.pop()) != NULL) { 5711 ((Klass*)(oopDesc*)k)->follow_weak_klass_links( 5712 &_is_alive_closure, 5713 &cmsKeepAliveClosure); 5714 } 5715 assert(!ClassUnloading || 5716 (_markStack.isEmpty() && overflow_list_is_empty()), 5717 "Should not have found new reachable objects"); 5718 assert(_revisitStack.isEmpty(), "revisit stack should have been drained"); 5719 cmsDrainMarkingStackClosure.do_void(); 5720 verify_work_stacks_empty(); 5721 } 5722 5723 { 5724 TraceTime t("scrub symbol & string tables", PrintGCDetails, false, gclog_or_tty); 5725 // Now clean up stale oops in SymbolTable and StringTable 5726 SymbolTable::unlink(&_is_alive_closure); 5727 StringTable::unlink(&_is_alive_closure); 5728 } 5729 } 5730 5731 verify_work_stacks_empty(); 5732 // Restore any preserved marks as a result of mark stack or 5733 // work queue overflow 5734 restore_preserved_marks_if_any(); // done single-threaded for now 5735 5736 rp->set_enqueuing_is_done(true); 5737 if (rp->processing_is_mt()) { 5738 CMSRefProcTaskExecutor task_executor(*this); 5739 rp->enqueue_discovered_references(&task_executor); 5740 } else { 5741 rp->enqueue_discovered_references(NULL); 5742 } 5743 rp->verify_no_references_recorded(); 5744 assert(!rp->discovery_enabled(), "should have been disabled"); 5745 5746 // JVMTI object tagging is based on JNI weak refs. If any of these 5747 // refs were cleared then JVMTI needs to update its maps and 5748 // maybe post ObjectFrees to agents. 5749 JvmtiExport::cms_ref_processing_epilogue(); 5750 } 5751 5752 #ifndef PRODUCT 5753 void CMSCollector::check_correct_thread_executing() { 5754 Thread* t = Thread::current(); 5755 // Only the VM thread or the CMS thread should be here. 5756 assert(t->is_ConcurrentGC_thread() || t->is_VM_thread(), 5757 "Unexpected thread type"); 5758 // If this is the vm thread, the foreground process 5759 // should not be waiting. Note that _foregroundGCIsActive is 5760 // true while the foreground collector is waiting. 5761 if (_foregroundGCShouldWait) { 5762 // We cannot be the VM thread 5763 assert(t->is_ConcurrentGC_thread(), 5764 "Should be CMS thread"); 5765 } else { 5766 // We can be the CMS thread only if we are in a stop-world 5767 // phase of CMS collection. 5768 if (t->is_ConcurrentGC_thread()) { 5769 assert(_collectorState == InitialMarking || 5770 _collectorState == FinalMarking, 5771 "Should be a stop-world phase"); 5772 // The CMS thread should be holding the CMS_token. 5773 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 5774 "Potential interference with concurrently " 5775 "executing VM thread"); 5776 } 5777 } 5778 } 5779 #endif 5780 5781 void CMSCollector::sweep(bool asynch) { 5782 assert(_collectorState == Sweeping, "just checking"); 5783 check_correct_thread_executing(); 5784 verify_work_stacks_empty(); 5785 verify_overflow_empty(); 5786 increment_sweep_count(); 5787 _sweep_timer.stop(); 5788 _sweep_estimate.sample(_sweep_timer.seconds()); 5789 size_policy()->avg_cms_free_at_sweep()->sample(_cmsGen->free()); 5790 5791 // PermGen verification support: If perm gen sweeping is disabled in 5792 // this cycle, we preserve the perm gen object "deadness" information 5793 // in the perm_gen_verify_bit_map. In order to do that we traverse 5794 // all blocks in perm gen and mark all dead objects. 5795 if (verifying() && !should_unload_classes()) { 5796 assert(perm_gen_verify_bit_map()->sizeInBits() != 0, 5797 "Should have already been allocated"); 5798 MarkDeadObjectsClosure mdo(this, _permGen->cmsSpace(), 5799 markBitMap(), perm_gen_verify_bit_map()); 5800 if (asynch) { 5801 CMSTokenSyncWithLocks ts(true, _permGen->freelistLock(), 5802 bitMapLock()); 5803 _permGen->cmsSpace()->blk_iterate(&mdo); 5804 } else { 5805 // In the case of synchronous sweep, we already have 5806 // the requisite locks/tokens. 5807 _permGen->cmsSpace()->blk_iterate(&mdo); 5808 } 5809 } 5810 5811 if (asynch) { 5812 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 5813 CMSPhaseAccounting pa(this, "sweep", !PrintGCDetails); 5814 // First sweep the old gen then the perm gen 5815 { 5816 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(), 5817 bitMapLock()); 5818 sweepWork(_cmsGen, asynch); 5819 } 5820 5821 // Now repeat for perm gen 5822 if (should_unload_classes()) { 5823 CMSTokenSyncWithLocks ts(true, _permGen->freelistLock(), 5824 bitMapLock()); 5825 sweepWork(_permGen, asynch); 5826 } 5827 5828 // Update Universe::_heap_*_at_gc figures. 5829 // We need all the free list locks to make the abstract state 5830 // transition from Sweeping to Resetting. See detailed note 5831 // further below. 5832 { 5833 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(), 5834 _permGen->freelistLock()); 5835 // Update heap occupancy information which is used as 5836 // input to soft ref clearing policy at the next gc. 5837 Universe::update_heap_info_at_gc(); 5838 _collectorState = Resizing; 5839 } 5840 } else { 5841 // already have needed locks 5842 sweepWork(_cmsGen, asynch); 5843 5844 if (should_unload_classes()) { 5845 sweepWork(_permGen, asynch); 5846 } 5847 // Update heap occupancy information which is used as 5848 // input to soft ref clearing policy at the next gc. 5849 Universe::update_heap_info_at_gc(); 5850 _collectorState = Resizing; 5851 } 5852 verify_work_stacks_empty(); 5853 verify_overflow_empty(); 5854 5855 _sweep_timer.reset(); 5856 _sweep_timer.start(); 5857 5858 update_time_of_last_gc(os::javaTimeMillis()); 5859 5860 // NOTE on abstract state transitions: 5861 // Mutators allocate-live and/or mark the mod-union table dirty 5862 // based on the state of the collection. The former is done in 5863 // the interval [Marking, Sweeping] and the latter in the interval 5864 // [Marking, Sweeping). Thus the transitions into the Marking state 5865 // and out of the Sweeping state must be synchronously visible 5866 // globally to the mutators. 5867 // The transition into the Marking state happens with the world 5868 // stopped so the mutators will globally see it. Sweeping is 5869 // done asynchronously by the background collector so the transition 5870 // from the Sweeping state to the Resizing state must be done 5871 // under the freelistLock (as is the check for whether to 5872 // allocate-live and whether to dirty the mod-union table). 5873 assert(_collectorState == Resizing, "Change of collector state to" 5874 " Resizing must be done under the freelistLocks (plural)"); 5875 5876 // Now that sweeping has been completed, if the GCH's 5877 // incremental_collection_will_fail flag is set, clear it, 5878 // thus inviting a younger gen collection to promote into 5879 // this generation. If such a promotion may still fail, 5880 // the flag will be set again when a young collection is 5881 // attempted. 5882 // I think the incremental_collection_will_fail flag's use 5883 // is specific to a 2 generation collection policy, so i'll 5884 // assert that that's the configuration we are operating within. 5885 // The use of the flag can and should be generalized appropriately 5886 // in the future to deal with a general n-generation system. 5887 5888 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5889 assert(gch->collector_policy()->is_two_generation_policy(), 5890 "Resetting of incremental_collection_will_fail flag" 5891 " may be incorrect otherwise"); 5892 gch->clear_incremental_collection_will_fail(); 5893 gch->update_full_collections_completed(_collection_count_start); 5894 } 5895 5896 // FIX ME!!! Looks like this belongs in CFLSpace, with 5897 // CMSGen merely delegating to it. 5898 void ConcurrentMarkSweepGeneration::setNearLargestChunk() { 5899 double nearLargestPercent = 0.999; 5900 HeapWord* minAddr = _cmsSpace->bottom(); 5901 HeapWord* largestAddr = 5902 (HeapWord*) _cmsSpace->dictionary()->findLargestDict(); 5903 if (largestAddr == 0) { 5904 // The dictionary appears to be empty. In this case 5905 // try to coalesce at the end of the heap. 5906 largestAddr = _cmsSpace->end(); 5907 } 5908 size_t largestOffset = pointer_delta(largestAddr, minAddr); 5909 size_t nearLargestOffset = 5910 (size_t)((double)largestOffset * nearLargestPercent) - MinChunkSize; 5911 _cmsSpace->set_nearLargestChunk(minAddr + nearLargestOffset); 5912 } 5913 5914 bool ConcurrentMarkSweepGeneration::isNearLargestChunk(HeapWord* addr) { 5915 return addr >= _cmsSpace->nearLargestChunk(); 5916 } 5917 5918 FreeChunk* ConcurrentMarkSweepGeneration::find_chunk_at_end() { 5919 return _cmsSpace->find_chunk_at_end(); 5920 } 5921 5922 void ConcurrentMarkSweepGeneration::update_gc_stats(int current_level, 5923 bool full) { 5924 // The next lower level has been collected. Gather any statistics 5925 // that are of interest at this point. 5926 if (!full && (current_level + 1) == level()) { 5927 // Gather statistics on the young generation collection. 5928 collector()->stats().record_gc0_end(used()); 5929 } 5930 } 5931 5932 CMSAdaptiveSizePolicy* ConcurrentMarkSweepGeneration::size_policy() { 5933 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5934 assert(gch->kind() == CollectedHeap::GenCollectedHeap, 5935 "Wrong type of heap"); 5936 CMSAdaptiveSizePolicy* sp = (CMSAdaptiveSizePolicy*) 5937 gch->gen_policy()->size_policy(); 5938 assert(sp->is_gc_cms_adaptive_size_policy(), 5939 "Wrong type of size policy"); 5940 return sp; 5941 } 5942 5943 void ConcurrentMarkSweepGeneration::rotate_debug_collection_type() { 5944 if (PrintGCDetails && Verbose) { 5945 gclog_or_tty->print("Rotate from %d ", _debug_collection_type); 5946 } 5947 _debug_collection_type = (CollectionTypes) (_debug_collection_type + 1); 5948 _debug_collection_type = 5949 (CollectionTypes) (_debug_collection_type % Unknown_collection_type); 5950 if (PrintGCDetails && Verbose) { 5951 gclog_or_tty->print_cr("to %d ", _debug_collection_type); 5952 } 5953 } 5954 5955 void CMSCollector::sweepWork(ConcurrentMarkSweepGeneration* gen, 5956 bool asynch) { 5957 // We iterate over the space(s) underlying this generation, 5958 // checking the mark bit map to see if the bits corresponding 5959 // to specific blocks are marked or not. Blocks that are 5960 // marked are live and are not swept up. All remaining blocks 5961 // are swept up, with coalescing on-the-fly as we sweep up 5962 // contiguous free and/or garbage blocks: 5963 // We need to ensure that the sweeper synchronizes with allocators 5964 // and stop-the-world collectors. In particular, the following 5965 // locks are used: 5966 // . CMS token: if this is held, a stop the world collection cannot occur 5967 // . freelistLock: if this is held no allocation can occur from this 5968 // generation by another thread 5969 // . bitMapLock: if this is held, no other thread can access or update 5970 // 5971 5972 // Note that we need to hold the freelistLock if we use 5973 // block iterate below; else the iterator might go awry if 5974 // a mutator (or promotion) causes block contents to change 5975 // (for instance if the allocator divvies up a block). 5976 // If we hold the free list lock, for all practical purposes 5977 // young generation GC's can't occur (they'll usually need to 5978 // promote), so we might as well prevent all young generation 5979 // GC's while we do a sweeping step. For the same reason, we might 5980 // as well take the bit map lock for the entire duration 5981 5982 // check that we hold the requisite locks 5983 assert(have_cms_token(), "Should hold cms token"); 5984 assert( (asynch && ConcurrentMarkSweepThread::cms_thread_has_cms_token()) 5985 || (!asynch && ConcurrentMarkSweepThread::vm_thread_has_cms_token()), 5986 "Should possess CMS token to sweep"); 5987 assert_lock_strong(gen->freelistLock()); 5988 assert_lock_strong(bitMapLock()); 5989 5990 assert(!_sweep_timer.is_active(), "Was switched off in an outer context"); 5991 gen->cmsSpace()->beginSweepFLCensus((float)(_sweep_timer.seconds()), 5992 _sweep_estimate.padded_average()); 5993 gen->setNearLargestChunk(); 5994 5995 { 5996 SweepClosure sweepClosure(this, gen, &_markBitMap, 5997 CMSYield && asynch); 5998 gen->cmsSpace()->blk_iterate_careful(&sweepClosure); 5999 // We need to free-up/coalesce garbage/blocks from a 6000 // co-terminal free run. This is done in the SweepClosure 6001 // destructor; so, do not remove this scope, else the 6002 // end-of-sweep-census below will be off by a little bit. 6003 } 6004 gen->cmsSpace()->sweep_completed(); 6005 gen->cmsSpace()->endSweepFLCensus(sweep_count()); 6006 if (should_unload_classes()) { // unloaded classes this cycle, 6007 _concurrent_cycles_since_last_unload = 0; // ... reset count 6008 } else { // did not unload classes, 6009 _concurrent_cycles_since_last_unload++; // ... increment count 6010 } 6011 } 6012 6013 // Reset CMS data structures (for now just the marking bit map) 6014 // preparatory for the next cycle. 6015 void CMSCollector::reset(bool asynch) { 6016 GenCollectedHeap* gch = GenCollectedHeap::heap(); 6017 CMSAdaptiveSizePolicy* sp = size_policy(); 6018 AdaptiveSizePolicyOutput(sp, gch->total_collections()); 6019 if (asynch) { 6020 CMSTokenSyncWithLocks ts(true, bitMapLock()); 6021 6022 // If the state is not "Resetting", the foreground thread 6023 // has done a collection and the resetting. 6024 if (_collectorState != Resetting) { 6025 assert(_collectorState == Idling, "The state should only change" 6026 " because the foreground collector has finished the collection"); 6027 return; 6028 } 6029 6030 // Clear the mark bitmap (no grey objects to start with) 6031 // for the next cycle. 6032 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 6033 CMSPhaseAccounting cmspa(this, "reset", !PrintGCDetails); 6034 6035 HeapWord* curAddr = _markBitMap.startWord(); 6036 while (curAddr < _markBitMap.endWord()) { 6037 size_t remaining = pointer_delta(_markBitMap.endWord(), curAddr); 6038 MemRegion chunk(curAddr, MIN2(CMSBitMapYieldQuantum, remaining)); 6039 _markBitMap.clear_large_range(chunk); 6040 if (ConcurrentMarkSweepThread::should_yield() && 6041 !foregroundGCIsActive() && 6042 CMSYield) { 6043 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6044 "CMS thread should hold CMS token"); 6045 assert_lock_strong(bitMapLock()); 6046 bitMapLock()->unlock(); 6047 ConcurrentMarkSweepThread::desynchronize(true); 6048 ConcurrentMarkSweepThread::acknowledge_yield_request(); 6049 stopTimer(); 6050 if (PrintCMSStatistics != 0) { 6051 incrementYields(); 6052 } 6053 icms_wait(); 6054 6055 // See the comment in coordinator_yield() 6056 for (unsigned i = 0; i < CMSYieldSleepCount && 6057 ConcurrentMarkSweepThread::should_yield() && 6058 !CMSCollector::foregroundGCIsActive(); ++i) { 6059 os::sleep(Thread::current(), 1, false); 6060 ConcurrentMarkSweepThread::acknowledge_yield_request(); 6061 } 6062 6063 ConcurrentMarkSweepThread::synchronize(true); 6064 bitMapLock()->lock_without_safepoint_check(); 6065 startTimer(); 6066 } 6067 curAddr = chunk.end(); 6068 } 6069 _collectorState = Idling; 6070 } else { 6071 // already have the lock 6072 assert(_collectorState == Resetting, "just checking"); 6073 assert_lock_strong(bitMapLock()); 6074 _markBitMap.clear_all(); 6075 _collectorState = Idling; 6076 } 6077 6078 // Stop incremental mode after a cycle completes, so that any future cycles 6079 // are triggered by allocation. 6080 stop_icms(); 6081 6082 NOT_PRODUCT( 6083 if (RotateCMSCollectionTypes) { 6084 _cmsGen->rotate_debug_collection_type(); 6085 } 6086 ) 6087 } 6088 6089 void CMSCollector::do_CMS_operation(CMS_op_type op) { 6090 gclog_or_tty->date_stamp(PrintGC && PrintGCDateStamps); 6091 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 6092 TraceTime t("GC", PrintGC, !PrintGCDetails, gclog_or_tty); 6093 TraceCollectorStats tcs(counters()); 6094 6095 switch (op) { 6096 case CMS_op_checkpointRootsInitial: { 6097 checkpointRootsInitial(true); // asynch 6098 if (PrintGC) { 6099 _cmsGen->printOccupancy("initial-mark"); 6100 } 6101 break; 6102 } 6103 case CMS_op_checkpointRootsFinal: { 6104 checkpointRootsFinal(true, // asynch 6105 false, // !clear_all_soft_refs 6106 false); // !init_mark_was_synchronous 6107 if (PrintGC) { 6108 _cmsGen->printOccupancy("remark"); 6109 } 6110 break; 6111 } 6112 default: 6113 fatal("No such CMS_op"); 6114 } 6115 } 6116 6117 #ifndef PRODUCT 6118 size_t const CMSCollector::skip_header_HeapWords() { 6119 return FreeChunk::header_size(); 6120 } 6121 6122 // Try and collect here conditions that should hold when 6123 // CMS thread is exiting. The idea is that the foreground GC 6124 // thread should not be blocked if it wants to terminate 6125 // the CMS thread and yet continue to run the VM for a while 6126 // after that. 6127 void CMSCollector::verify_ok_to_terminate() const { 6128 assert(Thread::current()->is_ConcurrentGC_thread(), 6129 "should be called by CMS thread"); 6130 assert(!_foregroundGCShouldWait, "should be false"); 6131 // We could check here that all the various low-level locks 6132 // are not held by the CMS thread, but that is overkill; see 6133 // also CMSThread::verify_ok_to_terminate() where the CGC_lock 6134 // is checked. 6135 } 6136 #endif 6137 6138 size_t CMSCollector::block_size_using_printezis_bits(HeapWord* addr) const { 6139 assert(_markBitMap.isMarked(addr) && _markBitMap.isMarked(addr + 1), 6140 "missing Printezis mark?"); 6141 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 6142 size_t size = pointer_delta(nextOneAddr + 1, addr); 6143 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 6144 "alignment problem"); 6145 assert(size >= 3, "Necessary for Printezis marks to work"); 6146 return size; 6147 } 6148 6149 // A variant of the above (block_size_using_printezis_bits()) except 6150 // that we return 0 if the P-bits are not yet set. 6151 size_t CMSCollector::block_size_if_printezis_bits(HeapWord* addr) const { 6152 if (_markBitMap.isMarked(addr)) { 6153 assert(_markBitMap.isMarked(addr + 1), "Missing Printezis bit?"); 6154 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 6155 size_t size = pointer_delta(nextOneAddr + 1, addr); 6156 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 6157 "alignment problem"); 6158 assert(size >= 3, "Necessary for Printezis marks to work"); 6159 return size; 6160 } else { 6161 assert(!_markBitMap.isMarked(addr + 1), "Bit map inconsistency?"); 6162 return 0; 6163 } 6164 } 6165 6166 HeapWord* CMSCollector::next_card_start_after_block(HeapWord* addr) const { 6167 size_t sz = 0; 6168 oop p = (oop)addr; 6169 if (p->klass() != NULL && p->is_parsable()) { 6170 sz = CompactibleFreeListSpace::adjustObjectSize(p->size()); 6171 } else { 6172 sz = block_size_using_printezis_bits(addr); 6173 } 6174 assert(sz > 0, "size must be nonzero"); 6175 HeapWord* next_block = addr + sz; 6176 HeapWord* next_card = (HeapWord*)round_to((uintptr_t)next_block, 6177 CardTableModRefBS::card_size); 6178 assert(round_down((uintptr_t)addr, CardTableModRefBS::card_size) < 6179 round_down((uintptr_t)next_card, CardTableModRefBS::card_size), 6180 "must be different cards"); 6181 return next_card; 6182 } 6183 6184 6185 // CMS Bit Map Wrapper ///////////////////////////////////////// 6186 6187 // Construct a CMS bit map infrastructure, but don't create the 6188 // bit vector itself. That is done by a separate call CMSBitMap::allocate() 6189 // further below. 6190 CMSBitMap::CMSBitMap(int shifter, int mutex_rank, const char* mutex_name): 6191 _bm(NULL,0), 6192 _shifter(shifter), 6193 _lock(mutex_rank >= 0 ? new Mutex(mutex_rank, mutex_name, true) : NULL) 6194 { 6195 _bmStartWord = 0; 6196 _bmWordSize = 0; 6197 } 6198 6199 bool CMSBitMap::allocate(MemRegion mr) { 6200 _bmStartWord = mr.start(); 6201 _bmWordSize = mr.word_size(); 6202 ReservedSpace brs(ReservedSpace::allocation_align_size_up( 6203 (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1)); 6204 if (!brs.is_reserved()) { 6205 warning("CMS bit map allocation failure"); 6206 return false; 6207 } 6208 // For now we'll just commit all of the bit map up fromt. 6209 // Later on we'll try to be more parsimonious with swap. 6210 if (!_virtual_space.initialize(brs, brs.size())) { 6211 warning("CMS bit map backing store failure"); 6212 return false; 6213 } 6214 assert(_virtual_space.committed_size() == brs.size(), 6215 "didn't reserve backing store for all of CMS bit map?"); 6216 _bm.set_map((uintptr_t*)_virtual_space.low()); 6217 assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >= 6218 _bmWordSize, "inconsistency in bit map sizing"); 6219 _bm.set_size(_bmWordSize >> _shifter); 6220 6221 // bm.clear(); // can we rely on getting zero'd memory? verify below 6222 assert(isAllClear(), 6223 "Expected zero'd memory from ReservedSpace constructor"); 6224 assert(_bm.size() == heapWordDiffToOffsetDiff(sizeInWords()), 6225 "consistency check"); 6226 return true; 6227 } 6228 6229 void CMSBitMap::dirty_range_iterate_clear(MemRegion mr, MemRegionClosure* cl) { 6230 HeapWord *next_addr, *end_addr, *last_addr; 6231 assert_locked(); 6232 assert(covers(mr), "out-of-range error"); 6233 // XXX assert that start and end are appropriately aligned 6234 for (next_addr = mr.start(), end_addr = mr.end(); 6235 next_addr < end_addr; next_addr = last_addr) { 6236 MemRegion dirty_region = getAndClearMarkedRegion(next_addr, end_addr); 6237 last_addr = dirty_region.end(); 6238 if (!dirty_region.is_empty()) { 6239 cl->do_MemRegion(dirty_region); 6240 } else { 6241 assert(last_addr == end_addr, "program logic"); 6242 return; 6243 } 6244 } 6245 } 6246 6247 #ifndef PRODUCT 6248 void CMSBitMap::assert_locked() const { 6249 CMSLockVerifier::assert_locked(lock()); 6250 } 6251 6252 bool CMSBitMap::covers(MemRegion mr) const { 6253 // assert(_bm.map() == _virtual_space.low(), "map inconsistency"); 6254 assert((size_t)_bm.size() == (_bmWordSize >> _shifter), 6255 "size inconsistency"); 6256 return (mr.start() >= _bmStartWord) && 6257 (mr.end() <= endWord()); 6258 } 6259 6260 bool CMSBitMap::covers(HeapWord* start, size_t size) const { 6261 return (start >= _bmStartWord && (start + size) <= endWord()); 6262 } 6263 6264 void CMSBitMap::verifyNoOneBitsInRange(HeapWord* left, HeapWord* right) { 6265 // verify that there are no 1 bits in the interval [left, right) 6266 FalseBitMapClosure falseBitMapClosure; 6267 iterate(&falseBitMapClosure, left, right); 6268 } 6269 6270 void CMSBitMap::region_invariant(MemRegion mr) 6271 { 6272 assert_locked(); 6273 // mr = mr.intersection(MemRegion(_bmStartWord, _bmWordSize)); 6274 assert(!mr.is_empty(), "unexpected empty region"); 6275 assert(covers(mr), "mr should be covered by bit map"); 6276 // convert address range into offset range 6277 size_t start_ofs = heapWordToOffset(mr.start()); 6278 // Make sure that end() is appropriately aligned 6279 assert(mr.end() == (HeapWord*)round_to((intptr_t)mr.end(), 6280 (1 << (_shifter+LogHeapWordSize))), 6281 "Misaligned mr.end()"); 6282 size_t end_ofs = heapWordToOffset(mr.end()); 6283 assert(end_ofs > start_ofs, "Should mark at least one bit"); 6284 } 6285 6286 #endif 6287 6288 bool CMSMarkStack::allocate(size_t size) { 6289 // allocate a stack of the requisite depth 6290 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 6291 size * sizeof(oop))); 6292 if (!rs.is_reserved()) { 6293 warning("CMSMarkStack allocation failure"); 6294 return false; 6295 } 6296 if (!_virtual_space.initialize(rs, rs.size())) { 6297 warning("CMSMarkStack backing store failure"); 6298 return false; 6299 } 6300 assert(_virtual_space.committed_size() == rs.size(), 6301 "didn't reserve backing store for all of CMS stack?"); 6302 _base = (oop*)(_virtual_space.low()); 6303 _index = 0; 6304 _capacity = size; 6305 NOT_PRODUCT(_max_depth = 0); 6306 return true; 6307 } 6308 6309 // XXX FIX ME !!! In the MT case we come in here holding a 6310 // leaf lock. For printing we need to take a further lock 6311 // which has lower rank. We need to recallibrate the two 6312 // lock-ranks involved in order to be able to rpint the 6313 // messages below. (Or defer the printing to the caller. 6314 // For now we take the expedient path of just disabling the 6315 // messages for the problematic case.) 6316 void CMSMarkStack::expand() { 6317 assert(_capacity <= CMSMarkStackSizeMax, "stack bigger than permitted"); 6318 if (_capacity == CMSMarkStackSizeMax) { 6319 if (_hit_limit++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) { 6320 // We print a warning message only once per CMS cycle. 6321 gclog_or_tty->print_cr(" (benign) Hit CMSMarkStack max size limit"); 6322 } 6323 return; 6324 } 6325 // Double capacity if possible 6326 size_t new_capacity = MIN2(_capacity*2, CMSMarkStackSizeMax); 6327 // Do not give up existing stack until we have managed to 6328 // get the double capacity that we desired. 6329 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 6330 new_capacity * sizeof(oop))); 6331 if (rs.is_reserved()) { 6332 // Release the backing store associated with old stack 6333 _virtual_space.release(); 6334 // Reinitialize virtual space for new stack 6335 if (!_virtual_space.initialize(rs, rs.size())) { 6336 fatal("Not enough swap for expanded marking stack"); 6337 } 6338 _base = (oop*)(_virtual_space.low()); 6339 _index = 0; 6340 _capacity = new_capacity; 6341 } else if (_failed_double++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) { 6342 // Failed to double capacity, continue; 6343 // we print a detail message only once per CMS cycle. 6344 gclog_or_tty->print(" (benign) Failed to expand marking stack from "SIZE_FORMAT"K to " 6345 SIZE_FORMAT"K", 6346 _capacity / K, new_capacity / K); 6347 } 6348 } 6349 6350 6351 // Closures 6352 // XXX: there seems to be a lot of code duplication here; 6353 // should refactor and consolidate common code. 6354 6355 // This closure is used to mark refs into the CMS generation in 6356 // the CMS bit map. Called at the first checkpoint. This closure 6357 // assumes that we do not need to re-mark dirty cards; if the CMS 6358 // generation on which this is used is not an oldest (modulo perm gen) 6359 // generation then this will lose younger_gen cards! 6360 6361 MarkRefsIntoClosure::MarkRefsIntoClosure( 6362 MemRegion span, CMSBitMap* bitMap, bool should_do_nmethods): 6363 _span(span), 6364 _bitMap(bitMap), 6365 _should_do_nmethods(should_do_nmethods) 6366 { 6367 assert(_ref_processor == NULL, "deliberately left NULL"); 6368 assert(_bitMap->covers(_span), "_bitMap/_span mismatch"); 6369 } 6370 6371 void MarkRefsIntoClosure::do_oop(oop* p) { 6372 // if p points into _span, then mark corresponding bit in _markBitMap 6373 oop thisOop = *p; 6374 if (thisOop != NULL) { 6375 assert(thisOop->is_oop(), "expected an oop"); 6376 HeapWord* addr = (HeapWord*)thisOop; 6377 if (_span.contains(addr)) { 6378 // this should be made more efficient 6379 _bitMap->mark(addr); 6380 } 6381 } 6382 } 6383 6384 // A variant of the above, used for CMS marking verification. 6385 MarkRefsIntoVerifyClosure::MarkRefsIntoVerifyClosure( 6386 MemRegion span, CMSBitMap* verification_bm, CMSBitMap* cms_bm, 6387 bool should_do_nmethods): 6388 _span(span), 6389 _verification_bm(verification_bm), 6390 _cms_bm(cms_bm), 6391 _should_do_nmethods(should_do_nmethods) { 6392 assert(_ref_processor == NULL, "deliberately left NULL"); 6393 assert(_verification_bm->covers(_span), "_verification_bm/_span mismatch"); 6394 } 6395 6396 void MarkRefsIntoVerifyClosure::do_oop(oop* p) { 6397 // if p points into _span, then mark corresponding bit in _markBitMap 6398 oop this_oop = *p; 6399 if (this_oop != NULL) { 6400 assert(this_oop->is_oop(), "expected an oop"); 6401 HeapWord* addr = (HeapWord*)this_oop; 6402 if (_span.contains(addr)) { 6403 _verification_bm->mark(addr); 6404 if (!_cms_bm->isMarked(addr)) { 6405 oop(addr)->print(); 6406 gclog_or_tty->print_cr(" ("INTPTR_FORMAT" should have been marked)", addr); 6407 fatal("... aborting"); 6408 } 6409 } 6410 } 6411 } 6412 6413 ////////////////////////////////////////////////// 6414 // MarkRefsIntoAndScanClosure 6415 ////////////////////////////////////////////////// 6416 6417 MarkRefsIntoAndScanClosure::MarkRefsIntoAndScanClosure(MemRegion span, 6418 ReferenceProcessor* rp, 6419 CMSBitMap* bit_map, 6420 CMSBitMap* mod_union_table, 6421 CMSMarkStack* mark_stack, 6422 CMSMarkStack* revisit_stack, 6423 CMSCollector* collector, 6424 bool should_yield, 6425 bool concurrent_precleaning): 6426 _collector(collector), 6427 _span(span), 6428 _bit_map(bit_map), 6429 _mark_stack(mark_stack), 6430 _pushAndMarkClosure(collector, span, rp, bit_map, mod_union_table, 6431 mark_stack, revisit_stack, concurrent_precleaning), 6432 _yield(should_yield), 6433 _concurrent_precleaning(concurrent_precleaning), 6434 _freelistLock(NULL) 6435 { 6436 _ref_processor = rp; 6437 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 6438 } 6439 6440 // This closure is used to mark refs into the CMS generation at the 6441 // second (final) checkpoint, and to scan and transitively follow 6442 // the unmarked oops. It is also used during the concurrent precleaning 6443 // phase while scanning objects on dirty cards in the CMS generation. 6444 // The marks are made in the marking bit map and the marking stack is 6445 // used for keeping the (newly) grey objects during the scan. 6446 // The parallel version (Par_...) appears further below. 6447 void MarkRefsIntoAndScanClosure::do_oop(oop* p) { 6448 oop this_oop = *p; 6449 if (this_oop != NULL) { 6450 assert(this_oop->is_oop(), "expected an oop"); 6451 HeapWord* addr = (HeapWord*)this_oop; 6452 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 6453 assert(_collector->overflow_list_is_empty(), "should be empty"); 6454 if (_span.contains(addr) && 6455 !_bit_map->isMarked(addr)) { 6456 // mark bit map (object is now grey) 6457 _bit_map->mark(addr); 6458 // push on marking stack (stack should be empty), and drain the 6459 // stack by applying this closure to the oops in the oops popped 6460 // from the stack (i.e. blacken the grey objects) 6461 bool res = _mark_stack->push(this_oop); 6462 assert(res, "Should have space to push on empty stack"); 6463 do { 6464 oop new_oop = _mark_stack->pop(); 6465 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 6466 assert(new_oop->is_parsable(), "Found unparsable oop"); 6467 assert(_bit_map->isMarked((HeapWord*)new_oop), 6468 "only grey objects on this stack"); 6469 // iterate over the oops in this oop, marking and pushing 6470 // the ones in CMS heap (i.e. in _span). 6471 new_oop->oop_iterate(&_pushAndMarkClosure); 6472 // check if it's time to yield 6473 do_yield_check(); 6474 } while (!_mark_stack->isEmpty() || 6475 (!_concurrent_precleaning && take_from_overflow_list())); 6476 // if marking stack is empty, and we are not doing this 6477 // during precleaning, then check the overflow list 6478 } 6479 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 6480 assert(_collector->overflow_list_is_empty(), 6481 "overflow list was drained above"); 6482 // We could restore evacuated mark words, if any, used for 6483 // overflow list links here because the overflow list is 6484 // provably empty here. That would reduce the maximum 6485 // size requirements for preserved_{oop,mark}_stack. 6486 // But we'll just postpone it until we are all done 6487 // so we can just stream through. 6488 if (!_concurrent_precleaning && CMSOverflowEarlyRestoration) { 6489 _collector->restore_preserved_marks_if_any(); 6490 assert(_collector->no_preserved_marks(), "No preserved marks"); 6491 } 6492 assert(!CMSOverflowEarlyRestoration || _collector->no_preserved_marks(), 6493 "All preserved marks should have been restored above"); 6494 } 6495 } 6496 6497 void MarkRefsIntoAndScanClosure::do_yield_work() { 6498 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6499 "CMS thread should hold CMS token"); 6500 assert_lock_strong(_freelistLock); 6501 assert_lock_strong(_bit_map->lock()); 6502 // relinquish the free_list_lock and bitMaplock() 6503 _bit_map->lock()->unlock(); 6504 _freelistLock->unlock(); 6505 ConcurrentMarkSweepThread::desynchronize(true); 6506 ConcurrentMarkSweepThread::acknowledge_yield_request(); 6507 _collector->stopTimer(); 6508 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 6509 if (PrintCMSStatistics != 0) { 6510 _collector->incrementYields(); 6511 } 6512 _collector->icms_wait(); 6513 6514 // See the comment in coordinator_yield() 6515 for (unsigned i = 0; i < CMSYieldSleepCount && 6516 ConcurrentMarkSweepThread::should_yield() && 6517 !CMSCollector::foregroundGCIsActive(); ++i) { 6518 os::sleep(Thread::current(), 1, false); 6519 ConcurrentMarkSweepThread::acknowledge_yield_request(); 6520 } 6521 6522 ConcurrentMarkSweepThread::synchronize(true); 6523 _freelistLock->lock_without_safepoint_check(); 6524 _bit_map->lock()->lock_without_safepoint_check(); 6525 _collector->startTimer(); 6526 } 6527 6528 /////////////////////////////////////////////////////////// 6529 // Par_MarkRefsIntoAndScanClosure: a parallel version of 6530 // MarkRefsIntoAndScanClosure 6531 /////////////////////////////////////////////////////////// 6532 Par_MarkRefsIntoAndScanClosure::Par_MarkRefsIntoAndScanClosure( 6533 CMSCollector* collector, MemRegion span, ReferenceProcessor* rp, 6534 CMSBitMap* bit_map, OopTaskQueue* work_queue, CMSMarkStack* revisit_stack): 6535 _span(span), 6536 _bit_map(bit_map), 6537 _work_queue(work_queue), 6538 _low_water_mark(MIN2((uint)(work_queue->max_elems()/4), 6539 (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads))), 6540 _par_pushAndMarkClosure(collector, span, rp, bit_map, work_queue, 6541 revisit_stack) 6542 { 6543 _ref_processor = rp; 6544 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 6545 } 6546 6547 // This closure is used to mark refs into the CMS generation at the 6548 // second (final) checkpoint, and to scan and transitively follow 6549 // the unmarked oops. The marks are made in the marking bit map and 6550 // the work_queue is used for keeping the (newly) grey objects during 6551 // the scan phase whence they are also available for stealing by parallel 6552 // threads. Since the marking bit map is shared, updates are 6553 // synchronized (via CAS). 6554 void Par_MarkRefsIntoAndScanClosure::do_oop(oop* p) { 6555 oop this_oop = *p; 6556 if (this_oop != NULL) { 6557 // Ignore mark word because this could be an already marked oop 6558 // that may be chained at the end of the overflow list. 6559 assert(this_oop->is_oop(true /* ignore mark word */), "expected an oop"); 6560 HeapWord* addr = (HeapWord*)this_oop; 6561 if (_span.contains(addr) && 6562 !_bit_map->isMarked(addr)) { 6563 // mark bit map (object will become grey): 6564 // It is possible for several threads to be 6565 // trying to "claim" this object concurrently; 6566 // the unique thread that succeeds in marking the 6567 // object first will do the subsequent push on 6568 // to the work queue (or overflow list). 6569 if (_bit_map->par_mark(addr)) { 6570 // push on work_queue (which may not be empty), and trim the 6571 // queue to an appropriate length by applying this closure to 6572 // the oops in the oops popped from the stack (i.e. blacken the 6573 // grey objects) 6574 bool res = _work_queue->push(this_oop); 6575 assert(res, "Low water mark should be less than capacity?"); 6576 trim_queue(_low_water_mark); 6577 } // Else, another thread claimed the object 6578 } 6579 } 6580 } 6581 6582 // This closure is used to rescan the marked objects on the dirty cards 6583 // in the mod union table and the card table proper. 6584 size_t ScanMarkedObjectsAgainCarefullyClosure::do_object_careful_m( 6585 oop p, MemRegion mr) { 6586 6587 size_t size = 0; 6588 HeapWord* addr = (HeapWord*)p; 6589 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6590 assert(_span.contains(addr), "we are scanning the CMS generation"); 6591 // check if it's time to yield 6592 if (do_yield_check()) { 6593 // We yielded for some foreground stop-world work, 6594 // and we have been asked to abort this ongoing preclean cycle. 6595 return 0; 6596 } 6597 if (_bitMap->isMarked(addr)) { 6598 // it's marked; is it potentially uninitialized? 6599 if (p->klass() != NULL) { 6600 if (CMSPermGenPrecleaningEnabled && !p->is_parsable()) { 6601 // Signal precleaning to redirty the card since 6602 // the klass pointer is already installed. 6603 assert(size == 0, "Initial value"); 6604 } else { 6605 assert(p->is_parsable(), "must be parsable."); 6606 // an initialized object; ignore mark word in verification below 6607 // since we are running concurrent with mutators 6608 assert(p->is_oop(true), "should be an oop"); 6609 if (p->is_objArray()) { 6610 // objArrays are precisely marked; restrict scanning 6611 // to dirty cards only. 6612 size = p->oop_iterate(_scanningClosure, mr); 6613 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 6614 "adjustObjectSize should be the identity for array sizes, " 6615 "which are necessarily larger than minimum object size of " 6616 "two heap words"); 6617 } else { 6618 // A non-array may have been imprecisely marked; we need 6619 // to scan object in its entirety. 6620 size = CompactibleFreeListSpace::adjustObjectSize( 6621 p->oop_iterate(_scanningClosure)); 6622 } 6623 #ifdef DEBUG 6624 size_t direct_size = 6625 CompactibleFreeListSpace::adjustObjectSize(p->size()); 6626 assert(size == direct_size, "Inconsistency in size"); 6627 assert(size >= 3, "Necessary for Printezis marks to work"); 6628 if (!_bitMap->isMarked(addr+1)) { 6629 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size); 6630 } else { 6631 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size-1); 6632 assert(_bitMap->isMarked(addr+size-1), 6633 "inconsistent Printezis mark"); 6634 } 6635 #endif // DEBUG 6636 } 6637 } else { 6638 // an unitialized object 6639 assert(_bitMap->isMarked(addr+1), "missing Printezis mark?"); 6640 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 6641 size = pointer_delta(nextOneAddr + 1, addr); 6642 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 6643 "alignment problem"); 6644 // Note that pre-cleaning needn't redirty the card. OopDesc::set_klass() 6645 // will dirty the card when the klass pointer is installed in the 6646 // object (signalling the completion of initialization). 6647 } 6648 } else { 6649 // Either a not yet marked object or an uninitialized object 6650 if (p->klass() == NULL || !p->is_parsable()) { 6651 // An uninitialized object, skip to the next card, since 6652 // we may not be able to read its P-bits yet. 6653 assert(size == 0, "Initial value"); 6654 } else { 6655 // An object not (yet) reached by marking: we merely need to 6656 // compute its size so as to go look at the next block. 6657 assert(p->is_oop(true), "should be an oop"); 6658 size = CompactibleFreeListSpace::adjustObjectSize(p->size()); 6659 } 6660 } 6661 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6662 return size; 6663 } 6664 6665 void ScanMarkedObjectsAgainCarefullyClosure::do_yield_work() { 6666 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6667 "CMS thread should hold CMS token"); 6668 assert_lock_strong(_freelistLock); 6669 assert_lock_strong(_bitMap->lock()); 6670 // relinquish the free_list_lock and bitMaplock() 6671 _bitMap->lock()->unlock(); 6672 _freelistLock->unlock(); 6673 ConcurrentMarkSweepThread::desynchronize(true); 6674 ConcurrentMarkSweepThread::acknowledge_yield_request(); 6675 _collector->stopTimer(); 6676 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 6677 if (PrintCMSStatistics != 0) { 6678 _collector->incrementYields(); 6679 } 6680 _collector->icms_wait(); 6681 6682 // See the comment in coordinator_yield() 6683 for (unsigned i = 0; i < CMSYieldSleepCount && 6684 ConcurrentMarkSweepThread::should_yield() && 6685 !CMSCollector::foregroundGCIsActive(); ++i) { 6686 os::sleep(Thread::current(), 1, false); 6687 ConcurrentMarkSweepThread::acknowledge_yield_request(); 6688 } 6689 6690 ConcurrentMarkSweepThread::synchronize(true); 6691 _freelistLock->lock_without_safepoint_check(); 6692 _bitMap->lock()->lock_without_safepoint_check(); 6693 _collector->startTimer(); 6694 } 6695 6696 6697 ////////////////////////////////////////////////////////////////// 6698 // SurvivorSpacePrecleanClosure 6699 ////////////////////////////////////////////////////////////////// 6700 // This (single-threaded) closure is used to preclean the oops in 6701 // the survivor spaces. 6702 size_t SurvivorSpacePrecleanClosure::do_object_careful(oop p) { 6703 6704 HeapWord* addr = (HeapWord*)p; 6705 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6706 assert(!_span.contains(addr), "we are scanning the survivor spaces"); 6707 assert(p->klass() != NULL, "object should be initializd"); 6708 assert(p->is_parsable(), "must be parsable."); 6709 // an initialized object; ignore mark word in verification below 6710 // since we are running concurrent with mutators 6711 assert(p->is_oop(true), "should be an oop"); 6712 // Note that we do not yield while we iterate over 6713 // the interior oops of p, pushing the relevant ones 6714 // on our marking stack. 6715 size_t size = p->oop_iterate(_scanning_closure); 6716 do_yield_check(); 6717 // Observe that below, we do not abandon the preclean 6718 // phase as soon as we should; rather we empty the 6719 // marking stack before returning. This is to satisfy 6720 // some existing assertions. In general, it may be a 6721 // good idea to abort immediately and complete the marking 6722 // from the grey objects at a later time. 6723 while (!_mark_stack->isEmpty()) { 6724 oop new_oop = _mark_stack->pop(); 6725 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 6726 assert(new_oop->is_parsable(), "Found unparsable oop"); 6727 assert(_bit_map->isMarked((HeapWord*)new_oop), 6728 "only grey objects on this stack"); 6729 // iterate over the oops in this oop, marking and pushing 6730 // the ones in CMS heap (i.e. in _span). 6731 new_oop->oop_iterate(_scanning_closure); 6732 // check if it's time to yield 6733 do_yield_check(); 6734 } 6735 unsigned int after_count = 6736 GenCollectedHeap::heap()->total_collections(); 6737 bool abort = (_before_count != after_count) || 6738 _collector->should_abort_preclean(); 6739 return abort ? 0 : size; 6740 } 6741 6742 void SurvivorSpacePrecleanClosure::do_yield_work() { 6743 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6744 "CMS thread should hold CMS token"); 6745 assert_lock_strong(_bit_map->lock()); 6746 // Relinquish the bit map lock 6747 _bit_map->lock()->unlock(); 6748 ConcurrentMarkSweepThread::desynchronize(true); 6749 ConcurrentMarkSweepThread::acknowledge_yield_request(); 6750 _collector->stopTimer(); 6751 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 6752 if (PrintCMSStatistics != 0) { 6753 _collector->incrementYields(); 6754 } 6755 _collector->icms_wait(); 6756 6757 // See the comment in coordinator_yield() 6758 for (unsigned i = 0; i < CMSYieldSleepCount && 6759 ConcurrentMarkSweepThread::should_yield() && 6760 !CMSCollector::foregroundGCIsActive(); ++i) { 6761 os::sleep(Thread::current(), 1, false); 6762 ConcurrentMarkSweepThread::acknowledge_yield_request(); 6763 } 6764 6765 ConcurrentMarkSweepThread::synchronize(true); 6766 _bit_map->lock()->lock_without_safepoint_check(); 6767 _collector->startTimer(); 6768 } 6769 6770 // This closure is used to rescan the marked objects on the dirty cards 6771 // in the mod union table and the card table proper. In the parallel 6772 // case, although the bitMap is shared, we do a single read so the 6773 // isMarked() query is "safe". 6774 bool ScanMarkedObjectsAgainClosure::do_object_bm(oop p, MemRegion mr) { 6775 // Ignore mark word because we are running concurrent with mutators 6776 assert(p->is_oop_or_null(true), "expected an oop or null"); 6777 HeapWord* addr = (HeapWord*)p; 6778 assert(_span.contains(addr), "we are scanning the CMS generation"); 6779 bool is_obj_array = false; 6780 #ifdef DEBUG 6781 if (!_parallel) { 6782 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 6783 assert(_collector->overflow_list_is_empty(), 6784 "overflow list should be empty"); 6785 6786 } 6787 #endif // DEBUG 6788 if (_bit_map->isMarked(addr)) { 6789 // Obj arrays are precisely marked, non-arrays are not; 6790 // so we scan objArrays precisely and non-arrays in their 6791 // entirety. 6792 if (p->is_objArray()) { 6793 is_obj_array = true; 6794 if (_parallel) { 6795 p->oop_iterate(_par_scan_closure, mr); 6796 } else { 6797 p->oop_iterate(_scan_closure, mr); 6798 } 6799 } else { 6800 if (_parallel) { 6801 p->oop_iterate(_par_scan_closure); 6802 } else { 6803 p->oop_iterate(_scan_closure); 6804 } 6805 } 6806 } 6807 #ifdef DEBUG 6808 if (!_parallel) { 6809 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 6810 assert(_collector->overflow_list_is_empty(), 6811 "overflow list should be empty"); 6812 6813 } 6814 #endif // DEBUG 6815 return is_obj_array; 6816 } 6817 6818 MarkFromRootsClosure::MarkFromRootsClosure(CMSCollector* collector, 6819 MemRegion span, 6820 CMSBitMap* bitMap, CMSMarkStack* markStack, 6821 CMSMarkStack* revisitStack, 6822 bool should_yield, bool verifying): 6823 _collector(collector), 6824 _span(span), 6825 _bitMap(bitMap), 6826 _mut(&collector->_modUnionTable), 6827 _markStack(markStack), 6828 _revisitStack(revisitStack), 6829 _yield(should_yield), 6830 _skipBits(0) 6831 { 6832 assert(_markStack->isEmpty(), "stack should be empty"); 6833 _finger = _bitMap->startWord(); 6834 _threshold = _finger; 6835 assert(_collector->_restart_addr == NULL, "Sanity check"); 6836 assert(_span.contains(_finger), "Out of bounds _finger?"); 6837 DEBUG_ONLY(_verifying = verifying;) 6838 } 6839 6840 void MarkFromRootsClosure::reset(HeapWord* addr) { 6841 assert(_markStack->isEmpty(), "would cause duplicates on stack"); 6842 assert(_span.contains(addr), "Out of bounds _finger?"); 6843 _finger = addr; 6844 _threshold = (HeapWord*)round_to( 6845 (intptr_t)_finger, CardTableModRefBS::card_size); 6846 } 6847 6848 // Should revisit to see if this should be restructured for 6849 // greater efficiency. 6850 void MarkFromRootsClosure::do_bit(size_t offset) { 6851 if (_skipBits > 0) { 6852 _skipBits--; 6853 return; 6854 } 6855 // convert offset into a HeapWord* 6856 HeapWord* addr = _bitMap->startWord() + offset; 6857 assert(_bitMap->endWord() && addr < _bitMap->endWord(), 6858 "address out of range"); 6859 assert(_bitMap->isMarked(addr), "tautology"); 6860 if (_bitMap->isMarked(addr+1)) { 6861 // this is an allocated but not yet initialized object 6862 assert(_skipBits == 0, "tautology"); 6863 _skipBits = 2; // skip next two marked bits ("Printezis-marks") 6864 oop p = oop(addr); 6865 if (p->klass() == NULL || !p->is_parsable()) { 6866 DEBUG_ONLY(if (!_verifying) {) 6867 // We re-dirty the cards on which this object lies and increase 6868 // the _threshold so that we'll come back to scan this object 6869 // during the preclean or remark phase. (CMSCleanOnEnter) 6870 if (CMSCleanOnEnter) { 6871 size_t sz = _collector->block_size_using_printezis_bits(addr); 6872 HeapWord* start_card_addr = (HeapWord*)round_down( 6873 (intptr_t)addr, CardTableModRefBS::card_size); 6874 HeapWord* end_card_addr = (HeapWord*)round_to( 6875 (intptr_t)(addr+sz), CardTableModRefBS::card_size); 6876 MemRegion redirty_range = MemRegion(start_card_addr, end_card_addr); 6877 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 6878 // Bump _threshold to end_card_addr; note that 6879 // _threshold cannot possibly exceed end_card_addr, anyhow. 6880 // This prevents future clearing of the card as the scan proceeds 6881 // to the right. 6882 assert(_threshold <= end_card_addr, 6883 "Because we are just scanning into this object"); 6884 if (_threshold < end_card_addr) { 6885 _threshold = end_card_addr; 6886 } 6887 if (p->klass() != NULL) { 6888 // Redirty the range of cards... 6889 _mut->mark_range(redirty_range); 6890 } // ...else the setting of klass will dirty the card anyway. 6891 } 6892 DEBUG_ONLY(}) 6893 return; 6894 } 6895 } 6896 scanOopsInOop(addr); 6897 } 6898 6899 // We take a break if we've been at this for a while, 6900 // so as to avoid monopolizing the locks involved. 6901 void MarkFromRootsClosure::do_yield_work() { 6902 // First give up the locks, then yield, then re-lock 6903 // We should probably use a constructor/destructor idiom to 6904 // do this unlock/lock or modify the MutexUnlocker class to 6905 // serve our purpose. XXX 6906 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6907 "CMS thread should hold CMS token"); 6908 assert_lock_strong(_bitMap->lock()); 6909 _bitMap->lock()->unlock(); 6910 ConcurrentMarkSweepThread::desynchronize(true); 6911 ConcurrentMarkSweepThread::acknowledge_yield_request(); 6912 _collector->stopTimer(); 6913 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 6914 if (PrintCMSStatistics != 0) { 6915 _collector->incrementYields(); 6916 } 6917 _collector->icms_wait(); 6918 6919 // See the comment in coordinator_yield() 6920 for (unsigned i = 0; i < CMSYieldSleepCount && 6921 ConcurrentMarkSweepThread::should_yield() && 6922 !CMSCollector::foregroundGCIsActive(); ++i) { 6923 os::sleep(Thread::current(), 1, false); 6924 ConcurrentMarkSweepThread::acknowledge_yield_request(); 6925 } 6926 6927 ConcurrentMarkSweepThread::synchronize(true); 6928 _bitMap->lock()->lock_without_safepoint_check(); 6929 _collector->startTimer(); 6930 } 6931 6932 void MarkFromRootsClosure::scanOopsInOop(HeapWord* ptr) { 6933 assert(_bitMap->isMarked(ptr), "expected bit to be set"); 6934 assert(_markStack->isEmpty(), 6935 "should drain stack to limit stack usage"); 6936 // convert ptr to an oop preparatory to scanning 6937 oop this_oop = oop(ptr); 6938 // Ignore mark word in verification below, since we 6939 // may be running concurrent with mutators. 6940 assert(this_oop->is_oop(true), "should be an oop"); 6941 assert(_finger <= ptr, "_finger runneth ahead"); 6942 // advance the finger to right end of this object 6943 _finger = ptr + this_oop->size(); 6944 assert(_finger > ptr, "we just incremented it above"); 6945 // On large heaps, it may take us some time to get through 6946 // the marking phase (especially if running iCMS). During 6947 // this time it's possible that a lot of mutations have 6948 // accumulated in the card table and the mod union table -- 6949 // these mutation records are redundant until we have 6950 // actually traced into the corresponding card. 6951 // Here, we check whether advancing the finger would make 6952 // us cross into a new card, and if so clear corresponding 6953 // cards in the MUT (preclean them in the card-table in the 6954 // future). 6955 6956 DEBUG_ONLY(if (!_verifying) {) 6957 // The clean-on-enter optimization is disabled by default, 6958 // until we fix 6178663. 6959 if (CMSCleanOnEnter && (_finger > _threshold)) { 6960 // [_threshold, _finger) represents the interval 6961 // of cards to be cleared in MUT (or precleaned in card table). 6962 // The set of cards to be cleared is all those that overlap 6963 // with the interval [_threshold, _finger); note that 6964 // _threshold is always kept card-aligned but _finger isn't 6965 // always card-aligned. 6966 HeapWord* old_threshold = _threshold; 6967 assert(old_threshold == (HeapWord*)round_to( 6968 (intptr_t)old_threshold, CardTableModRefBS::card_size), 6969 "_threshold should always be card-aligned"); 6970 _threshold = (HeapWord*)round_to( 6971 (intptr_t)_finger, CardTableModRefBS::card_size); 6972 MemRegion mr(old_threshold, _threshold); 6973 assert(!mr.is_empty(), "Control point invariant"); 6974 assert(_span.contains(mr), "Should clear within span"); 6975 // XXX When _finger crosses from old gen into perm gen 6976 // we may be doing unnecessary cleaning; do better in the 6977 // future by detecting that condition and clearing fewer 6978 // MUT/CT entries. 6979 _mut->clear_range(mr); 6980 } 6981 DEBUG_ONLY(}) 6982 6983 // Note: the finger doesn't advance while we drain 6984 // the stack below. 6985 PushOrMarkClosure pushOrMarkClosure(_collector, 6986 _span, _bitMap, _markStack, 6987 _revisitStack, 6988 _finger, this); 6989 bool res = _markStack->push(this_oop); 6990 assert(res, "Empty non-zero size stack should have space for single push"); 6991 while (!_markStack->isEmpty()) { 6992 oop new_oop = _markStack->pop(); 6993 // Skip verifying header mark word below because we are 6994 // running concurrent with mutators. 6995 assert(new_oop->is_oop(true), "Oops! expected to pop an oop"); 6996 // now scan this oop's oops 6997 new_oop->oop_iterate(&pushOrMarkClosure); 6998 do_yield_check(); 6999 } 7000 assert(_markStack->isEmpty(), "tautology, emphasizing post-condition"); 7001 } 7002 7003 Par_MarkFromRootsClosure::Par_MarkFromRootsClosure(CMSConcMarkingTask* task, 7004 CMSCollector* collector, MemRegion span, 7005 CMSBitMap* bit_map, 7006 OopTaskQueue* work_queue, 7007 CMSMarkStack* overflow_stack, 7008 CMSMarkStack* revisit_stack, 7009 bool should_yield): 7010 _collector(collector), 7011 _whole_span(collector->_span), 7012 _span(span), 7013 _bit_map(bit_map), 7014 _mut(&collector->_modUnionTable), 7015 _work_queue(work_queue), 7016 _overflow_stack(overflow_stack), 7017 _revisit_stack(revisit_stack), 7018 _yield(should_yield), 7019 _skip_bits(0), 7020 _task(task) 7021 { 7022 assert(_work_queue->size() == 0, "work_queue should be empty"); 7023 _finger = span.start(); 7024 _threshold = _finger; // XXX Defer clear-on-enter optimization for now 7025 assert(_span.contains(_finger), "Out of bounds _finger?"); 7026 } 7027 7028 // Should revisit to see if this should be restructured for 7029 // greater efficiency. 7030 void Par_MarkFromRootsClosure::do_bit(size_t offset) { 7031 if (_skip_bits > 0) { 7032 _skip_bits--; 7033 return; 7034 } 7035 // convert offset into a HeapWord* 7036 HeapWord* addr = _bit_map->startWord() + offset; 7037 assert(_bit_map->endWord() && addr < _bit_map->endWord(), 7038 "address out of range"); 7039 assert(_bit_map->isMarked(addr), "tautology"); 7040 if (_bit_map->isMarked(addr+1)) { 7041 // this is an allocated object that might not yet be initialized 7042 assert(_skip_bits == 0, "tautology"); 7043 _skip_bits = 2; // skip next two marked bits ("Printezis-marks") 7044 oop p = oop(addr); 7045 if (p->klass() == NULL || !p->is_parsable()) { 7046 // in the case of Clean-on-Enter optimization, redirty card 7047 // and avoid clearing card by increasing the threshold. 7048 return; 7049 } 7050 } 7051 scan_oops_in_oop(addr); 7052 } 7053 7054 void Par_MarkFromRootsClosure::scan_oops_in_oop(HeapWord* ptr) { 7055 assert(_bit_map->isMarked(ptr), "expected bit to be set"); 7056 // Should we assert that our work queue is empty or 7057 // below some drain limit? 7058 assert(_work_queue->size() == 0, 7059 "should drain stack to limit stack usage"); 7060 // convert ptr to an oop preparatory to scanning 7061 oop this_oop = oop(ptr); 7062 // Ignore mark word in verification below, since we 7063 // may be running concurrent with mutators. 7064 assert(this_oop->is_oop(true), "should be an oop"); 7065 assert(_finger <= ptr, "_finger runneth ahead"); 7066 // advance the finger to right end of this object 7067 _finger = ptr + this_oop->size(); 7068 assert(_finger > ptr, "we just incremented it above"); 7069 // On large heaps, it may take us some time to get through 7070 // the marking phase (especially if running iCMS). During 7071 // this time it's possible that a lot of mutations have 7072 // accumulated in the card table and the mod union table -- 7073 // these mutation records are redundant until we have 7074 // actually traced into the corresponding card. 7075 // Here, we check whether advancing the finger would make 7076 // us cross into a new card, and if so clear corresponding 7077 // cards in the MUT (preclean them in the card-table in the 7078 // future). 7079 7080 // The clean-on-enter optimization is disabled by default, 7081 // until we fix 6178663. 7082 if (CMSCleanOnEnter && (_finger > _threshold)) { 7083 // [_threshold, _finger) represents the interval 7084 // of cards to be cleared in MUT (or precleaned in card table). 7085 // The set of cards to be cleared is all those that overlap 7086 // with the interval [_threshold, _finger); note that 7087 // _threshold is always kept card-aligned but _finger isn't 7088 // always card-aligned. 7089 HeapWord* old_threshold = _threshold; 7090 assert(old_threshold == (HeapWord*)round_to( 7091 (intptr_t)old_threshold, CardTableModRefBS::card_size), 7092 "_threshold should always be card-aligned"); 7093 _threshold = (HeapWord*)round_to( 7094 (intptr_t)_finger, CardTableModRefBS::card_size); 7095 MemRegion mr(old_threshold, _threshold); 7096 assert(!mr.is_empty(), "Control point invariant"); 7097 assert(_span.contains(mr), "Should clear within span"); // _whole_span ?? 7098 // XXX When _finger crosses from old gen into perm gen 7099 // we may be doing unnecessary cleaning; do better in the 7100 // future by detecting that condition and clearing fewer 7101 // MUT/CT entries. 7102 _mut->clear_range(mr); 7103 } 7104 7105 // Note: the local finger doesn't advance while we drain 7106 // the stack below, but the global finger sure can and will. 7107 HeapWord** gfa = _task->global_finger_addr(); 7108 Par_PushOrMarkClosure pushOrMarkClosure(_collector, 7109 _span, _bit_map, 7110 _work_queue, 7111 _overflow_stack, 7112 _revisit_stack, 7113 _finger, 7114 gfa, this); 7115 bool res = _work_queue->push(this_oop); // overflow could occur here 7116 assert(res, "Will hold once we use workqueues"); 7117 while (true) { 7118 oop new_oop; 7119 if (!_work_queue->pop_local(new_oop)) { 7120 // We emptied our work_queue; check if there's stuff that can 7121 // be gotten from the overflow stack. 7122 if (CMSConcMarkingTask::get_work_from_overflow_stack( 7123 _overflow_stack, _work_queue)) { 7124 do_yield_check(); 7125 continue; 7126 } else { // done 7127 break; 7128 } 7129 } 7130 // Skip verifying header mark word below because we are 7131 // running concurrent with mutators. 7132 assert(new_oop->is_oop(true), "Oops! expected to pop an oop"); 7133 // now scan this oop's oops 7134 new_oop->oop_iterate(&pushOrMarkClosure); 7135 do_yield_check(); 7136 } 7137 assert(_work_queue->size() == 0, "tautology, emphasizing post-condition"); 7138 } 7139 7140 // Yield in response to a request from VM Thread or 7141 // from mutators. 7142 void Par_MarkFromRootsClosure::do_yield_work() { 7143 assert(_task != NULL, "sanity"); 7144 _task->yield(); 7145 } 7146 7147 // A variant of the above used for verifying CMS marking work. 7148 MarkFromRootsVerifyClosure::MarkFromRootsVerifyClosure(CMSCollector* collector, 7149 MemRegion span, 7150 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 7151 CMSMarkStack* mark_stack): 7152 _collector(collector), 7153 _span(span), 7154 _verification_bm(verification_bm), 7155 _cms_bm(cms_bm), 7156 _mark_stack(mark_stack), 7157 _pam_verify_closure(collector, span, verification_bm, cms_bm, 7158 mark_stack) 7159 { 7160 assert(_mark_stack->isEmpty(), "stack should be empty"); 7161 _finger = _verification_bm->startWord(); 7162 assert(_collector->_restart_addr == NULL, "Sanity check"); 7163 assert(_span.contains(_finger), "Out of bounds _finger?"); 7164 } 7165 7166 void MarkFromRootsVerifyClosure::reset(HeapWord* addr) { 7167 assert(_mark_stack->isEmpty(), "would cause duplicates on stack"); 7168 assert(_span.contains(addr), "Out of bounds _finger?"); 7169 _finger = addr; 7170 } 7171 7172 // Should revisit to see if this should be restructured for 7173 // greater efficiency. 7174 void MarkFromRootsVerifyClosure::do_bit(size_t offset) { 7175 // convert offset into a HeapWord* 7176 HeapWord* addr = _verification_bm->startWord() + offset; 7177 assert(_verification_bm->endWord() && addr < _verification_bm->endWord(), 7178 "address out of range"); 7179 assert(_verification_bm->isMarked(addr), "tautology"); 7180 assert(_cms_bm->isMarked(addr), "tautology"); 7181 7182 assert(_mark_stack->isEmpty(), 7183 "should drain stack to limit stack usage"); 7184 // convert addr to an oop preparatory to scanning 7185 oop this_oop = oop(addr); 7186 assert(this_oop->is_oop(), "should be an oop"); 7187 assert(_finger <= addr, "_finger runneth ahead"); 7188 // advance the finger to right end of this object 7189 _finger = addr + this_oop->size(); 7190 assert(_finger > addr, "we just incremented it above"); 7191 // Note: the finger doesn't advance while we drain 7192 // the stack below. 7193 bool res = _mark_stack->push(this_oop); 7194 assert(res, "Empty non-zero size stack should have space for single push"); 7195 while (!_mark_stack->isEmpty()) { 7196 oop new_oop = _mark_stack->pop(); 7197 assert(new_oop->is_oop(), "Oops! expected to pop an oop"); 7198 // now scan this oop's oops 7199 new_oop->oop_iterate(&_pam_verify_closure); 7200 } 7201 assert(_mark_stack->isEmpty(), "tautology, emphasizing post-condition"); 7202 } 7203 7204 PushAndMarkVerifyClosure::PushAndMarkVerifyClosure( 7205 CMSCollector* collector, MemRegion span, 7206 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 7207 CMSMarkStack* mark_stack): 7208 OopClosure(collector->ref_processor()), 7209 _collector(collector), 7210 _span(span), 7211 _verification_bm(verification_bm), 7212 _cms_bm(cms_bm), 7213 _mark_stack(mark_stack) 7214 { } 7215 7216 7217 // Upon stack overflow, we discard (part of) the stack, 7218 // remembering the least address amongst those discarded 7219 // in CMSCollector's _restart_address. 7220 void PushAndMarkVerifyClosure::handle_stack_overflow(HeapWord* lost) { 7221 // Remember the least grey address discarded 7222 HeapWord* ra = (HeapWord*)_mark_stack->least_value(lost); 7223 _collector->lower_restart_addr(ra); 7224 _mark_stack->reset(); // discard stack contents 7225 _mark_stack->expand(); // expand the stack if possible 7226 } 7227 7228 void PushAndMarkVerifyClosure::do_oop(oop* p) { 7229 oop this_oop = *p; 7230 assert(this_oop->is_oop_or_null(), "expected an oop or NULL"); 7231 HeapWord* addr = (HeapWord*)this_oop; 7232 if (_span.contains(addr) && !_verification_bm->isMarked(addr)) { 7233 // Oop lies in _span and isn't yet grey or black 7234 _verification_bm->mark(addr); // now grey 7235 if (!_cms_bm->isMarked(addr)) { 7236 oop(addr)->print(); 7237 gclog_or_tty->print_cr(" ("INTPTR_FORMAT" should have been marked)", addr); 7238 fatal("... aborting"); 7239 } 7240 7241 if (!_mark_stack->push(this_oop)) { // stack overflow 7242 if (PrintCMSStatistics != 0) { 7243 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 7244 SIZE_FORMAT, _mark_stack->capacity()); 7245 } 7246 assert(_mark_stack->isFull(), "Else push should have succeeded"); 7247 handle_stack_overflow(addr); 7248 } 7249 // anything including and to the right of _finger 7250 // will be scanned as we iterate over the remainder of the 7251 // bit map 7252 } 7253 } 7254 7255 PushOrMarkClosure::PushOrMarkClosure(CMSCollector* collector, 7256 MemRegion span, 7257 CMSBitMap* bitMap, CMSMarkStack* markStack, 7258 CMSMarkStack* revisitStack, 7259 HeapWord* finger, MarkFromRootsClosure* parent) : 7260 OopClosure(collector->ref_processor()), 7261 _collector(collector), 7262 _span(span), 7263 _bitMap(bitMap), 7264 _markStack(markStack), 7265 _revisitStack(revisitStack), 7266 _finger(finger), 7267 _parent(parent), 7268 _should_remember_klasses(collector->should_unload_classes()) 7269 { } 7270 7271 Par_PushOrMarkClosure::Par_PushOrMarkClosure(CMSCollector* collector, 7272 MemRegion span, 7273 CMSBitMap* bit_map, 7274 OopTaskQueue* work_queue, 7275 CMSMarkStack* overflow_stack, 7276 CMSMarkStack* revisit_stack, 7277 HeapWord* finger, 7278 HeapWord** global_finger_addr, 7279 Par_MarkFromRootsClosure* parent) : 7280 OopClosure(collector->ref_processor()), 7281 _collector(collector), 7282 _whole_span(collector->_span), 7283 _span(span), 7284 _bit_map(bit_map), 7285 _work_queue(work_queue), 7286 _overflow_stack(overflow_stack), 7287 _revisit_stack(revisit_stack), 7288 _finger(finger), 7289 _global_finger_addr(global_finger_addr), 7290 _parent(parent), 7291 _should_remember_klasses(collector->should_unload_classes()) 7292 { } 7293 7294 7295 void CMSCollector::lower_restart_addr(HeapWord* low) { 7296 assert(_span.contains(low), "Out of bounds addr"); 7297 if (_restart_addr == NULL) { 7298 _restart_addr = low; 7299 } else { 7300 _restart_addr = MIN2(_restart_addr, low); 7301 } 7302 } 7303 7304 // Upon stack overflow, we discard (part of) the stack, 7305 // remembering the least address amongst those discarded 7306 // in CMSCollector's _restart_address. 7307 void PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 7308 // Remember the least grey address discarded 7309 HeapWord* ra = (HeapWord*)_markStack->least_value(lost); 7310 _collector->lower_restart_addr(ra); 7311 _markStack->reset(); // discard stack contents 7312 _markStack->expand(); // expand the stack if possible 7313 } 7314 7315 // Upon stack overflow, we discard (part of) the stack, 7316 // remembering the least address amongst those discarded 7317 // in CMSCollector's _restart_address. 7318 void Par_PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 7319 // We need to do this under a mutex to prevent other 7320 // workers from interfering with the expansion below. 7321 MutexLockerEx ml(_overflow_stack->par_lock(), 7322 Mutex::_no_safepoint_check_flag); 7323 // Remember the least grey address discarded 7324 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 7325 _collector->lower_restart_addr(ra); 7326 _overflow_stack->reset(); // discard stack contents 7327 _overflow_stack->expand(); // expand the stack if possible 7328 } 7329 7330 7331 void PushOrMarkClosure::do_oop(oop* p) { 7332 oop thisOop = *p; 7333 // Ignore mark word because we are running concurrent with mutators. 7334 assert(thisOop->is_oop_or_null(true), "expected an oop or NULL"); 7335 HeapWord* addr = (HeapWord*)thisOop; 7336 if (_span.contains(addr) && !_bitMap->isMarked(addr)) { 7337 // Oop lies in _span and isn't yet grey or black 7338 _bitMap->mark(addr); // now grey 7339 if (addr < _finger) { 7340 // the bit map iteration has already either passed, or 7341 // sampled, this bit in the bit map; we'll need to 7342 // use the marking stack to scan this oop's oops. 7343 bool simulate_overflow = false; 7344 NOT_PRODUCT( 7345 if (CMSMarkStackOverflowALot && 7346 _collector->simulate_overflow()) { 7347 // simulate a stack overflow 7348 simulate_overflow = true; 7349 } 7350 ) 7351 if (simulate_overflow || !_markStack->push(thisOop)) { // stack overflow 7352 if (PrintCMSStatistics != 0) { 7353 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 7354 SIZE_FORMAT, _markStack->capacity()); 7355 } 7356 assert(simulate_overflow || _markStack->isFull(), "Else push should have succeeded"); 7357 handle_stack_overflow(addr); 7358 } 7359 } 7360 // anything including and to the right of _finger 7361 // will be scanned as we iterate over the remainder of the 7362 // bit map 7363 do_yield_check(); 7364 } 7365 } 7366 7367 void Par_PushOrMarkClosure::do_oop(oop* p) { 7368 oop this_oop = *p; 7369 // Ignore mark word because we are running concurrent with mutators. 7370 assert(this_oop->is_oop_or_null(true), "expected an oop or NULL"); 7371 HeapWord* addr = (HeapWord*)this_oop; 7372 if (_whole_span.contains(addr) && !_bit_map->isMarked(addr)) { 7373 // Oop lies in _span and isn't yet grey or black 7374 // We read the global_finger (volatile read) strictly after marking oop 7375 bool res = _bit_map->par_mark(addr); // now grey 7376 volatile HeapWord** gfa = (volatile HeapWord**)_global_finger_addr; 7377 // Should we push this marked oop on our stack? 7378 // -- if someone else marked it, nothing to do 7379 // -- if target oop is above global finger nothing to do 7380 // -- if target oop is in chunk and above local finger 7381 // then nothing to do 7382 // -- else push on work queue 7383 if ( !res // someone else marked it, they will deal with it 7384 || (addr >= *gfa) // will be scanned in a later task 7385 || (_span.contains(addr) && addr >= _finger)) { // later in this chunk 7386 return; 7387 } 7388 // the bit map iteration has already either passed, or 7389 // sampled, this bit in the bit map; we'll need to 7390 // use the marking stack to scan this oop's oops. 7391 bool simulate_overflow = false; 7392 NOT_PRODUCT( 7393 if (CMSMarkStackOverflowALot && 7394 _collector->simulate_overflow()) { 7395 // simulate a stack overflow 7396 simulate_overflow = true; 7397 } 7398 ) 7399 if (simulate_overflow || 7400 !(_work_queue->push(this_oop) || _overflow_stack->par_push(this_oop))) { 7401 // stack overflow 7402 if (PrintCMSStatistics != 0) { 7403 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 7404 SIZE_FORMAT, _overflow_stack->capacity()); 7405 } 7406 // We cannot assert that the overflow stack is full because 7407 // it may have been emptied since. 7408 assert(simulate_overflow || 7409 _work_queue->size() == _work_queue->max_elems(), 7410 "Else push should have succeeded"); 7411 handle_stack_overflow(addr); 7412 } 7413 do_yield_check(); 7414 } 7415 } 7416 7417 7418 PushAndMarkClosure::PushAndMarkClosure(CMSCollector* collector, 7419 MemRegion span, 7420 ReferenceProcessor* rp, 7421 CMSBitMap* bit_map, 7422 CMSBitMap* mod_union_table, 7423 CMSMarkStack* mark_stack, 7424 CMSMarkStack* revisit_stack, 7425 bool concurrent_precleaning): 7426 OopClosure(rp), 7427 _collector(collector), 7428 _span(span), 7429 _bit_map(bit_map), 7430 _mod_union_table(mod_union_table), 7431 _mark_stack(mark_stack), 7432 _revisit_stack(revisit_stack), 7433 _concurrent_precleaning(concurrent_precleaning), 7434 _should_remember_klasses(collector->should_unload_classes()) 7435 { 7436 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 7437 } 7438 7439 // Grey object rescan during pre-cleaning and second checkpoint phases -- 7440 // the non-parallel version (the parallel version appears further below.) 7441 void PushAndMarkClosure::do_oop(oop* p) { 7442 oop this_oop = *p; 7443 // Ignore mark word verification. If during concurrent precleaning 7444 // the object monitor may be locked. If during the checkpoint 7445 // phases, the object may already have been reached by a different 7446 // path and may be at the end of the global overflow list (so 7447 // the mark word may be NULL). 7448 assert(this_oop->is_oop_or_null(true/* ignore mark word */), 7449 "expected an oop or NULL"); 7450 HeapWord* addr = (HeapWord*)this_oop; 7451 // Check if oop points into the CMS generation 7452 // and is not marked 7453 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 7454 // a white object ... 7455 _bit_map->mark(addr); // ... now grey 7456 // push on the marking stack (grey set) 7457 bool simulate_overflow = false; 7458 NOT_PRODUCT( 7459 if (CMSMarkStackOverflowALot && 7460 _collector->simulate_overflow()) { 7461 // simulate a stack overflow 7462 simulate_overflow = true; 7463 } 7464 ) 7465 if (simulate_overflow || !_mark_stack->push(this_oop)) { 7466 if (_concurrent_precleaning) { 7467 // During precleaning we can just dirty the appropriate card 7468 // in the mod union table, thus ensuring that the object remains 7469 // in the grey set and continue. Note that no one can be intefering 7470 // with us in this action of dirtying the mod union table, so 7471 // no locking is required. 7472 _mod_union_table->mark(addr); 7473 _collector->_ser_pmc_preclean_ovflw++; 7474 } else { 7475 // During the remark phase, we need to remember this oop 7476 // in the overflow list. 7477 _collector->push_on_overflow_list(this_oop); 7478 _collector->_ser_pmc_remark_ovflw++; 7479 } 7480 } 7481 } 7482 } 7483 7484 Par_PushAndMarkClosure::Par_PushAndMarkClosure(CMSCollector* collector, 7485 MemRegion span, 7486 ReferenceProcessor* rp, 7487 CMSBitMap* bit_map, 7488 OopTaskQueue* work_queue, 7489 CMSMarkStack* revisit_stack): 7490 OopClosure(rp), 7491 _collector(collector), 7492 _span(span), 7493 _bit_map(bit_map), 7494 _work_queue(work_queue), 7495 _revisit_stack(revisit_stack), 7496 _should_remember_klasses(collector->should_unload_classes()) 7497 { 7498 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 7499 } 7500 7501 // Grey object rescan during second checkpoint phase -- 7502 // the parallel version. 7503 void Par_PushAndMarkClosure::do_oop(oop* p) { 7504 oop this_oop = *p; 7505 // In the assert below, we ignore the mark word because 7506 // this oop may point to an already visited object that is 7507 // on the overflow stack (in which case the mark word has 7508 // been hijacked for chaining into the overflow stack -- 7509 // if this is the last object in the overflow stack then 7510 // its mark word will be NULL). Because this object may 7511 // have been subsequently popped off the global overflow 7512 // stack, and the mark word possibly restored to the prototypical 7513 // value, by the time we get to examined this failing assert in 7514 // the debugger, is_oop_or_null(false) may subsequently start 7515 // to hold. 7516 assert(this_oop->is_oop_or_null(true), 7517 "expected an oop or NULL"); 7518 HeapWord* addr = (HeapWord*)this_oop; 7519 // Check if oop points into the CMS generation 7520 // and is not marked 7521 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 7522 // a white object ... 7523 // If we manage to "claim" the object, by being the 7524 // first thread to mark it, then we push it on our 7525 // marking stack 7526 if (_bit_map->par_mark(addr)) { // ... now grey 7527 // push on work queue (grey set) 7528 bool simulate_overflow = false; 7529 NOT_PRODUCT( 7530 if (CMSMarkStackOverflowALot && 7531 _collector->par_simulate_overflow()) { 7532 // simulate a stack overflow 7533 simulate_overflow = true; 7534 } 7535 ) 7536 if (simulate_overflow || !_work_queue->push(this_oop)) { 7537 _collector->par_push_on_overflow_list(this_oop); 7538 _collector->_par_pmc_remark_ovflw++; // imprecise OK: no need to CAS 7539 } 7540 } // Else, some other thread got there first 7541 } 7542 } 7543 7544 void PushAndMarkClosure::remember_klass(Klass* k) { 7545 if (!_revisit_stack->push(oop(k))) { 7546 fatal("Revisit stack overflowed in PushAndMarkClosure"); 7547 } 7548 } 7549 7550 void Par_PushAndMarkClosure::remember_klass(Klass* k) { 7551 if (!_revisit_stack->par_push(oop(k))) { 7552 fatal("Revist stack overflowed in Par_PushAndMarkClosure"); 7553 } 7554 } 7555 7556 void CMSPrecleanRefsYieldClosure::do_yield_work() { 7557 Mutex* bml = _collector->bitMapLock(); 7558 assert_lock_strong(bml); 7559 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 7560 "CMS thread should hold CMS token"); 7561 7562 bml->unlock(); 7563 ConcurrentMarkSweepThread::desynchronize(true); 7564 7565 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7566 7567 _collector->stopTimer(); 7568 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 7569 if (PrintCMSStatistics != 0) { 7570 _collector->incrementYields(); 7571 } 7572 _collector->icms_wait(); 7573 7574 // See the comment in coordinator_yield() 7575 for (unsigned i = 0; i < CMSYieldSleepCount && 7576 ConcurrentMarkSweepThread::should_yield() && 7577 !CMSCollector::foregroundGCIsActive(); ++i) { 7578 os::sleep(Thread::current(), 1, false); 7579 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7580 } 7581 7582 ConcurrentMarkSweepThread::synchronize(true); 7583 bml->lock(); 7584 7585 _collector->startTimer(); 7586 } 7587 7588 bool CMSPrecleanRefsYieldClosure::should_return() { 7589 if (ConcurrentMarkSweepThread::should_yield()) { 7590 do_yield_work(); 7591 } 7592 return _collector->foregroundGCIsActive(); 7593 } 7594 7595 void MarkFromDirtyCardsClosure::do_MemRegion(MemRegion mr) { 7596 assert(((size_t)mr.start())%CardTableModRefBS::card_size_in_words == 0, 7597 "mr should be aligned to start at a card boundary"); 7598 // We'd like to assert: 7599 // assert(mr.word_size()%CardTableModRefBS::card_size_in_words == 0, 7600 // "mr should be a range of cards"); 7601 // However, that would be too strong in one case -- the last 7602 // partition ends at _unallocated_block which, in general, can be 7603 // an arbitrary boundary, not necessarily card aligned. 7604 if (PrintCMSStatistics != 0) { 7605 _num_dirty_cards += 7606 mr.word_size()/CardTableModRefBS::card_size_in_words; 7607 } 7608 _space->object_iterate_mem(mr, &_scan_cl); 7609 } 7610 7611 SweepClosure::SweepClosure(CMSCollector* collector, 7612 ConcurrentMarkSweepGeneration* g, 7613 CMSBitMap* bitMap, bool should_yield) : 7614 _collector(collector), 7615 _g(g), 7616 _sp(g->cmsSpace()), 7617 _limit(_sp->sweep_limit()), 7618 _freelistLock(_sp->freelistLock()), 7619 _bitMap(bitMap), 7620 _yield(should_yield), 7621 _inFreeRange(false), // No free range at beginning of sweep 7622 _freeRangeInFreeLists(false), // No free range at beginning of sweep 7623 _lastFreeRangeCoalesced(false), 7624 _freeFinger(g->used_region().start()) 7625 { 7626 NOT_PRODUCT( 7627 _numObjectsFreed = 0; 7628 _numWordsFreed = 0; 7629 _numObjectsLive = 0; 7630 _numWordsLive = 0; 7631 _numObjectsAlreadyFree = 0; 7632 _numWordsAlreadyFree = 0; 7633 _last_fc = NULL; 7634 7635 _sp->initializeIndexedFreeListArrayReturnedBytes(); 7636 _sp->dictionary()->initializeDictReturnedBytes(); 7637 ) 7638 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7639 "sweep _limit out of bounds"); 7640 if (CMSTraceSweeper) { 7641 gclog_or_tty->print("\n====================\nStarting new sweep\n"); 7642 } 7643 } 7644 7645 // We need this destructor to reclaim any space at the end 7646 // of the space, which do_blk below may not have added back to 7647 // the free lists. [basically dealing with the "fringe effect"] 7648 SweepClosure::~SweepClosure() { 7649 assert_lock_strong(_freelistLock); 7650 // this should be treated as the end of a free run if any 7651 // The current free range should be returned to the free lists 7652 // as one coalesced chunk. 7653 if (inFreeRange()) { 7654 flushCurFreeChunk(freeFinger(), 7655 pointer_delta(_limit, freeFinger())); 7656 assert(freeFinger() < _limit, "the finger pointeth off base"); 7657 if (CMSTraceSweeper) { 7658 gclog_or_tty->print("destructor:"); 7659 gclog_or_tty->print("Sweep:put_free_blk 0x%x ("SIZE_FORMAT") " 7660 "[coalesced:"SIZE_FORMAT"]\n", 7661 freeFinger(), pointer_delta(_limit, freeFinger()), 7662 lastFreeRangeCoalesced()); 7663 } 7664 } 7665 NOT_PRODUCT( 7666 if (Verbose && PrintGC) { 7667 gclog_or_tty->print("Collected "SIZE_FORMAT" objects, " 7668 SIZE_FORMAT " bytes", 7669 _numObjectsFreed, _numWordsFreed*sizeof(HeapWord)); 7670 gclog_or_tty->print_cr("\nLive "SIZE_FORMAT" objects, " 7671 SIZE_FORMAT" bytes " 7672 "Already free "SIZE_FORMAT" objects, "SIZE_FORMAT" bytes", 7673 _numObjectsLive, _numWordsLive*sizeof(HeapWord), 7674 _numObjectsAlreadyFree, _numWordsAlreadyFree*sizeof(HeapWord)); 7675 size_t totalBytes = (_numWordsFreed + _numWordsLive + _numWordsAlreadyFree) * 7676 sizeof(HeapWord); 7677 gclog_or_tty->print_cr("Total sweep: "SIZE_FORMAT" bytes", totalBytes); 7678 7679 if (PrintCMSStatistics && CMSVerifyReturnedBytes) { 7680 size_t indexListReturnedBytes = _sp->sumIndexedFreeListArrayReturnedBytes(); 7681 size_t dictReturnedBytes = _sp->dictionary()->sumDictReturnedBytes(); 7682 size_t returnedBytes = indexListReturnedBytes + dictReturnedBytes; 7683 gclog_or_tty->print("Returned "SIZE_FORMAT" bytes", returnedBytes); 7684 gclog_or_tty->print(" Indexed List Returned "SIZE_FORMAT" bytes", 7685 indexListReturnedBytes); 7686 gclog_or_tty->print_cr(" Dictionary Returned "SIZE_FORMAT" bytes", 7687 dictReturnedBytes); 7688 } 7689 } 7690 ) 7691 // Now, in debug mode, just null out the sweep_limit 7692 NOT_PRODUCT(_sp->clear_sweep_limit();) 7693 if (CMSTraceSweeper) { 7694 gclog_or_tty->print("end of sweep\n================\n"); 7695 } 7696 } 7697 7698 void SweepClosure::initialize_free_range(HeapWord* freeFinger, 7699 bool freeRangeInFreeLists) { 7700 if (CMSTraceSweeper) { 7701 gclog_or_tty->print("---- Start free range 0x%x with free block [%d] (%d)\n", 7702 freeFinger, _sp->block_size(freeFinger), 7703 freeRangeInFreeLists); 7704 } 7705 assert(!inFreeRange(), "Trampling existing free range"); 7706 set_inFreeRange(true); 7707 set_lastFreeRangeCoalesced(false); 7708 7709 set_freeFinger(freeFinger); 7710 set_freeRangeInFreeLists(freeRangeInFreeLists); 7711 if (CMSTestInFreeList) { 7712 if (freeRangeInFreeLists) { 7713 FreeChunk* fc = (FreeChunk*) freeFinger; 7714 assert(fc->isFree(), "A chunk on the free list should be free."); 7715 assert(fc->size() > 0, "Free range should have a size"); 7716 assert(_sp->verifyChunkInFreeLists(fc), "Chunk is not in free lists"); 7717 } 7718 } 7719 } 7720 7721 // Note that the sweeper runs concurrently with mutators. Thus, 7722 // it is possible for direct allocation in this generation to happen 7723 // in the middle of the sweep. Note that the sweeper also coalesces 7724 // contiguous free blocks. Thus, unless the sweeper and the allocator 7725 // synchronize appropriately freshly allocated blocks may get swept up. 7726 // This is accomplished by the sweeper locking the free lists while 7727 // it is sweeping. Thus blocks that are determined to be free are 7728 // indeed free. There is however one additional complication: 7729 // blocks that have been allocated since the final checkpoint and 7730 // mark, will not have been marked and so would be treated as 7731 // unreachable and swept up. To prevent this, the allocator marks 7732 // the bit map when allocating during the sweep phase. This leads, 7733 // however, to a further complication -- objects may have been allocated 7734 // but not yet initialized -- in the sense that the header isn't yet 7735 // installed. The sweeper can not then determine the size of the block 7736 // in order to skip over it. To deal with this case, we use a technique 7737 // (due to Printezis) to encode such uninitialized block sizes in the 7738 // bit map. Since the bit map uses a bit per every HeapWord, but the 7739 // CMS generation has a minimum object size of 3 HeapWords, it follows 7740 // that "normal marks" won't be adjacent in the bit map (there will 7741 // always be at least two 0 bits between successive 1 bits). We make use 7742 // of these "unused" bits to represent uninitialized blocks -- the bit 7743 // corresponding to the start of the uninitialized object and the next 7744 // bit are both set. Finally, a 1 bit marks the end of the object that 7745 // started with the two consecutive 1 bits to indicate its potentially 7746 // uninitialized state. 7747 7748 size_t SweepClosure::do_blk_careful(HeapWord* addr) { 7749 FreeChunk* fc = (FreeChunk*)addr; 7750 size_t res; 7751 7752 // check if we are done sweepinrg 7753 if (addr == _limit) { // we have swept up to the limit, do nothing more 7754 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7755 "sweep _limit out of bounds"); 7756 // help the closure application finish 7757 return pointer_delta(_sp->end(), _limit); 7758 } 7759 assert(addr <= _limit, "sweep invariant"); 7760 7761 // check if we should yield 7762 do_yield_check(addr); 7763 if (fc->isFree()) { 7764 // Chunk that is already free 7765 res = fc->size(); 7766 doAlreadyFreeChunk(fc); 7767 debug_only(_sp->verifyFreeLists()); 7768 assert(res == fc->size(), "Don't expect the size to change"); 7769 NOT_PRODUCT( 7770 _numObjectsAlreadyFree++; 7771 _numWordsAlreadyFree += res; 7772 ) 7773 NOT_PRODUCT(_last_fc = fc;) 7774 } else if (!_bitMap->isMarked(addr)) { 7775 // Chunk is fresh garbage 7776 res = doGarbageChunk(fc); 7777 debug_only(_sp->verifyFreeLists()); 7778 NOT_PRODUCT( 7779 _numObjectsFreed++; 7780 _numWordsFreed += res; 7781 ) 7782 } else { 7783 // Chunk that is alive. 7784 res = doLiveChunk(fc); 7785 debug_only(_sp->verifyFreeLists()); 7786 NOT_PRODUCT( 7787 _numObjectsLive++; 7788 _numWordsLive += res; 7789 ) 7790 } 7791 return res; 7792 } 7793 7794 // For the smart allocation, record following 7795 // split deaths - a free chunk is removed from its free list because 7796 // it is being split into two or more chunks. 7797 // split birth - a free chunk is being added to its free list because 7798 // a larger free chunk has been split and resulted in this free chunk. 7799 // coal death - a free chunk is being removed from its free list because 7800 // it is being coalesced into a large free chunk. 7801 // coal birth - a free chunk is being added to its free list because 7802 // it was created when two or more free chunks where coalesced into 7803 // this free chunk. 7804 // 7805 // These statistics are used to determine the desired number of free 7806 // chunks of a given size. The desired number is chosen to be relative 7807 // to the end of a CMS sweep. The desired number at the end of a sweep 7808 // is the 7809 // count-at-end-of-previous-sweep (an amount that was enough) 7810 // - count-at-beginning-of-current-sweep (the excess) 7811 // + split-births (gains in this size during interval) 7812 // - split-deaths (demands on this size during interval) 7813 // where the interval is from the end of one sweep to the end of the 7814 // next. 7815 // 7816 // When sweeping the sweeper maintains an accumulated chunk which is 7817 // the chunk that is made up of chunks that have been coalesced. That 7818 // will be termed the left-hand chunk. A new chunk of garbage that 7819 // is being considered for coalescing will be referred to as the 7820 // right-hand chunk. 7821 // 7822 // When making a decision on whether to coalesce a right-hand chunk with 7823 // the current left-hand chunk, the current count vs. the desired count 7824 // of the left-hand chunk is considered. Also if the right-hand chunk 7825 // is near the large chunk at the end of the heap (see 7826 // ConcurrentMarkSweepGeneration::isNearLargestChunk()), then the 7827 // left-hand chunk is coalesced. 7828 // 7829 // When making a decision about whether to split a chunk, the desired count 7830 // vs. the current count of the candidate to be split is also considered. 7831 // If the candidate is underpopulated (currently fewer chunks than desired) 7832 // a chunk of an overpopulated (currently more chunks than desired) size may 7833 // be chosen. The "hint" associated with a free list, if non-null, points 7834 // to a free list which may be overpopulated. 7835 // 7836 7837 void SweepClosure::doAlreadyFreeChunk(FreeChunk* fc) { 7838 size_t size = fc->size(); 7839 // Chunks that cannot be coalesced are not in the 7840 // free lists. 7841 if (CMSTestInFreeList && !fc->cantCoalesce()) { 7842 assert(_sp->verifyChunkInFreeLists(fc), 7843 "free chunk should be in free lists"); 7844 } 7845 // a chunk that is already free, should not have been 7846 // marked in the bit map 7847 HeapWord* addr = (HeapWord*) fc; 7848 assert(!_bitMap->isMarked(addr), "free chunk should be unmarked"); 7849 // Verify that the bit map has no bits marked between 7850 // addr and purported end of this block. 7851 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 7852 7853 // Some chunks cannot be coalesced in under any circumstances. 7854 // See the definition of cantCoalesce(). 7855 if (!fc->cantCoalesce()) { 7856 // This chunk can potentially be coalesced. 7857 if (_sp->adaptive_freelists()) { 7858 // All the work is done in 7859 doPostIsFreeOrGarbageChunk(fc, size); 7860 } else { // Not adaptive free lists 7861 // this is a free chunk that can potentially be coalesced by the sweeper; 7862 if (!inFreeRange()) { 7863 // if the next chunk is a free block that can't be coalesced 7864 // it doesn't make sense to remove this chunk from the free lists 7865 FreeChunk* nextChunk = (FreeChunk*)(addr + size); 7866 assert((HeapWord*)nextChunk <= _limit, "sweep invariant"); 7867 if ((HeapWord*)nextChunk < _limit && // there's a next chunk... 7868 nextChunk->isFree() && // which is free... 7869 nextChunk->cantCoalesce()) { // ... but cant be coalesced 7870 // nothing to do 7871 } else { 7872 // Potentially the start of a new free range: 7873 // Don't eagerly remove it from the free lists. 7874 // No need to remove it if it will just be put 7875 // back again. (Also from a pragmatic point of view 7876 // if it is a free block in a region that is beyond 7877 // any allocated blocks, an assertion will fail) 7878 // Remember the start of a free run. 7879 initialize_free_range(addr, true); 7880 // end - can coalesce with next chunk 7881 } 7882 } else { 7883 // the midst of a free range, we are coalescing 7884 debug_only(record_free_block_coalesced(fc);) 7885 if (CMSTraceSweeper) { 7886 gclog_or_tty->print(" -- pick up free block 0x%x (%d)\n", fc, size); 7887 } 7888 // remove it from the free lists 7889 _sp->removeFreeChunkFromFreeLists(fc); 7890 set_lastFreeRangeCoalesced(true); 7891 // If the chunk is being coalesced and the current free range is 7892 // in the free lists, remove the current free range so that it 7893 // will be returned to the free lists in its entirety - all 7894 // the coalesced pieces included. 7895 if (freeRangeInFreeLists()) { 7896 FreeChunk* ffc = (FreeChunk*) freeFinger(); 7897 assert(ffc->size() == pointer_delta(addr, freeFinger()), 7898 "Size of free range is inconsistent with chunk size."); 7899 if (CMSTestInFreeList) { 7900 assert(_sp->verifyChunkInFreeLists(ffc), 7901 "free range is not in free lists"); 7902 } 7903 _sp->removeFreeChunkFromFreeLists(ffc); 7904 set_freeRangeInFreeLists(false); 7905 } 7906 } 7907 } 7908 } else { 7909 // Code path common to both original and adaptive free lists. 7910 7911 // cant coalesce with previous block; this should be treated 7912 // as the end of a free run if any 7913 if (inFreeRange()) { 7914 // we kicked some butt; time to pick up the garbage 7915 assert(freeFinger() < addr, "the finger pointeth off base"); 7916 flushCurFreeChunk(freeFinger(), pointer_delta(addr, freeFinger())); 7917 } 7918 // else, nothing to do, just continue 7919 } 7920 } 7921 7922 size_t SweepClosure::doGarbageChunk(FreeChunk* fc) { 7923 // This is a chunk of garbage. It is not in any free list. 7924 // Add it to a free list or let it possibly be coalesced into 7925 // a larger chunk. 7926 HeapWord* addr = (HeapWord*) fc; 7927 size_t size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 7928 7929 if (_sp->adaptive_freelists()) { 7930 // Verify that the bit map has no bits marked between 7931 // addr and purported end of just dead object. 7932 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 7933 7934 doPostIsFreeOrGarbageChunk(fc, size); 7935 } else { 7936 if (!inFreeRange()) { 7937 // start of a new free range 7938 assert(size > 0, "A free range should have a size"); 7939 initialize_free_range(addr, false); 7940 7941 } else { 7942 // this will be swept up when we hit the end of the 7943 // free range 7944 if (CMSTraceSweeper) { 7945 gclog_or_tty->print(" -- pick up garbage 0x%x (%d) \n", fc, size); 7946 } 7947 // If the chunk is being coalesced and the current free range is 7948 // in the free lists, remove the current free range so that it 7949 // will be returned to the free lists in its entirety - all 7950 // the coalesced pieces included. 7951 if (freeRangeInFreeLists()) { 7952 FreeChunk* ffc = (FreeChunk*)freeFinger(); 7953 assert(ffc->size() == pointer_delta(addr, freeFinger()), 7954 "Size of free range is inconsistent with chunk size."); 7955 if (CMSTestInFreeList) { 7956 assert(_sp->verifyChunkInFreeLists(ffc), 7957 "free range is not in free lists"); 7958 } 7959 _sp->removeFreeChunkFromFreeLists(ffc); 7960 set_freeRangeInFreeLists(false); 7961 } 7962 set_lastFreeRangeCoalesced(true); 7963 } 7964 // this will be swept up when we hit the end of the free range 7965 7966 // Verify that the bit map has no bits marked between 7967 // addr and purported end of just dead object. 7968 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 7969 } 7970 return size; 7971 } 7972 7973 size_t SweepClosure::doLiveChunk(FreeChunk* fc) { 7974 HeapWord* addr = (HeapWord*) fc; 7975 // The sweeper has just found a live object. Return any accumulated 7976 // left hand chunk to the free lists. 7977 if (inFreeRange()) { 7978 if (_sp->adaptive_freelists()) { 7979 flushCurFreeChunk(freeFinger(), 7980 pointer_delta(addr, freeFinger())); 7981 } else { // not adaptive freelists 7982 set_inFreeRange(false); 7983 // Add the free range back to the free list if it is not already 7984 // there. 7985 if (!freeRangeInFreeLists()) { 7986 assert(freeFinger() < addr, "the finger pointeth off base"); 7987 if (CMSTraceSweeper) { 7988 gclog_or_tty->print("Sweep:put_free_blk 0x%x (%d) " 7989 "[coalesced:%d]\n", 7990 freeFinger(), pointer_delta(addr, freeFinger()), 7991 lastFreeRangeCoalesced()); 7992 } 7993 _sp->addChunkAndRepairOffsetTable(freeFinger(), 7994 pointer_delta(addr, freeFinger()), lastFreeRangeCoalesced()); 7995 } 7996 } 7997 } 7998 7999 // Common code path for original and adaptive free lists. 8000 8001 // this object is live: we'd normally expect this to be 8002 // an oop, and like to assert the following: 8003 // assert(oop(addr)->is_oop(), "live block should be an oop"); 8004 // However, as we commented above, this may be an object whose 8005 // header hasn't yet been initialized. 8006 size_t size; 8007 assert(_bitMap->isMarked(addr), "Tautology for this control point"); 8008 if (_bitMap->isMarked(addr + 1)) { 8009 // Determine the size from the bit map, rather than trying to 8010 // compute it from the object header. 8011 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 8012 size = pointer_delta(nextOneAddr + 1, addr); 8013 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 8014 "alignment problem"); 8015 8016 #ifdef DEBUG 8017 if (oop(addr)->klass() != NULL && 8018 ( !_collector->should_unload_classes() 8019 || oop(addr)->is_parsable())) { 8020 // Ignore mark word because we are running concurrent with mutators 8021 assert(oop(addr)->is_oop(true), "live block should be an oop"); 8022 assert(size == 8023 CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()), 8024 "P-mark and computed size do not agree"); 8025 } 8026 #endif 8027 8028 } else { 8029 // This should be an initialized object that's alive. 8030 assert(oop(addr)->klass() != NULL && 8031 (!_collector->should_unload_classes() 8032 || oop(addr)->is_parsable()), 8033 "Should be an initialized object"); 8034 // Ignore mark word because we are running concurrent with mutators 8035 assert(oop(addr)->is_oop(true), "live block should be an oop"); 8036 // Verify that the bit map has no bits marked between 8037 // addr and purported end of this block. 8038 size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 8039 assert(size >= 3, "Necessary for Printezis marks to work"); 8040 assert(!_bitMap->isMarked(addr+1), "Tautology for this control point"); 8041 DEBUG_ONLY(_bitMap->verifyNoOneBitsInRange(addr+2, addr+size);) 8042 } 8043 return size; 8044 } 8045 8046 void SweepClosure::doPostIsFreeOrGarbageChunk(FreeChunk* fc, 8047 size_t chunkSize) { 8048 // doPostIsFreeOrGarbageChunk() should only be called in the smart allocation 8049 // scheme. 8050 bool fcInFreeLists = fc->isFree(); 8051 assert(_sp->adaptive_freelists(), "Should only be used in this case."); 8052 assert((HeapWord*)fc <= _limit, "sweep invariant"); 8053 if (CMSTestInFreeList && fcInFreeLists) { 8054 assert(_sp->verifyChunkInFreeLists(fc), 8055 "free chunk is not in free lists"); 8056 } 8057 8058 8059 if (CMSTraceSweeper) { 8060 gclog_or_tty->print_cr(" -- pick up another chunk at 0x%x (%d)", fc, chunkSize); 8061 } 8062 8063 HeapWord* addr = (HeapWord*) fc; 8064 8065 bool coalesce; 8066 size_t left = pointer_delta(addr, freeFinger()); 8067 size_t right = chunkSize; 8068 switch (FLSCoalescePolicy) { 8069 // numeric value forms a coalition aggressiveness metric 8070 case 0: { // never coalesce 8071 coalesce = false; 8072 break; 8073 } 8074 case 1: { // coalesce if left & right chunks on overpopulated lists 8075 coalesce = _sp->coalOverPopulated(left) && 8076 _sp->coalOverPopulated(right); 8077 break; 8078 } 8079 case 2: { // coalesce if left chunk on overpopulated list (default) 8080 coalesce = _sp->coalOverPopulated(left); 8081 break; 8082 } 8083 case 3: { // coalesce if left OR right chunk on overpopulated list 8084 coalesce = _sp->coalOverPopulated(left) || 8085 _sp->coalOverPopulated(right); 8086 break; 8087 } 8088 case 4: { // always coalesce 8089 coalesce = true; 8090 break; 8091 } 8092 default: 8093 ShouldNotReachHere(); 8094 } 8095 8096 // Should the current free range be coalesced? 8097 // If the chunk is in a free range and either we decided to coalesce above 8098 // or the chunk is near the large block at the end of the heap 8099 // (isNearLargestChunk() returns true), then coalesce this chunk. 8100 bool doCoalesce = inFreeRange() && 8101 (coalesce || _g->isNearLargestChunk((HeapWord*)fc)); 8102 if (doCoalesce) { 8103 // Coalesce the current free range on the left with the new 8104 // chunk on the right. If either is on a free list, 8105 // it must be removed from the list and stashed in the closure. 8106 if (freeRangeInFreeLists()) { 8107 FreeChunk* ffc = (FreeChunk*)freeFinger(); 8108 assert(ffc->size() == pointer_delta(addr, freeFinger()), 8109 "Size of free range is inconsistent with chunk size."); 8110 if (CMSTestInFreeList) { 8111 assert(_sp->verifyChunkInFreeLists(ffc), 8112 "Chunk is not in free lists"); 8113 } 8114 _sp->coalDeath(ffc->size()); 8115 _sp->removeFreeChunkFromFreeLists(ffc); 8116 set_freeRangeInFreeLists(false); 8117 } 8118 if (fcInFreeLists) { 8119 _sp->coalDeath(chunkSize); 8120 assert(fc->size() == chunkSize, 8121 "The chunk has the wrong size or is not in the free lists"); 8122 _sp->removeFreeChunkFromFreeLists(fc); 8123 } 8124 set_lastFreeRangeCoalesced(true); 8125 } else { // not in a free range and/or should not coalesce 8126 // Return the current free range and start a new one. 8127 if (inFreeRange()) { 8128 // In a free range but cannot coalesce with the right hand chunk. 8129 // Put the current free range into the free lists. 8130 flushCurFreeChunk(freeFinger(), 8131 pointer_delta(addr, freeFinger())); 8132 } 8133 // Set up for new free range. Pass along whether the right hand 8134 // chunk is in the free lists. 8135 initialize_free_range((HeapWord*)fc, fcInFreeLists); 8136 } 8137 } 8138 void SweepClosure::flushCurFreeChunk(HeapWord* chunk, size_t size) { 8139 assert(inFreeRange(), "Should only be called if currently in a free range."); 8140 assert(size > 0, 8141 "A zero sized chunk cannot be added to the free lists."); 8142 if (!freeRangeInFreeLists()) { 8143 if(CMSTestInFreeList) { 8144 FreeChunk* fc = (FreeChunk*) chunk; 8145 fc->setSize(size); 8146 assert(!_sp->verifyChunkInFreeLists(fc), 8147 "chunk should not be in free lists yet"); 8148 } 8149 if (CMSTraceSweeper) { 8150 gclog_or_tty->print_cr(" -- add free block 0x%x (%d) to free lists", 8151 chunk, size); 8152 } 8153 // A new free range is going to be starting. The current 8154 // free range has not been added to the free lists yet or 8155 // was removed so add it back. 8156 // If the current free range was coalesced, then the death 8157 // of the free range was recorded. Record a birth now. 8158 if (lastFreeRangeCoalesced()) { 8159 _sp->coalBirth(size); 8160 } 8161 _sp->addChunkAndRepairOffsetTable(chunk, size, 8162 lastFreeRangeCoalesced()); 8163 } 8164 set_inFreeRange(false); 8165 set_freeRangeInFreeLists(false); 8166 } 8167 8168 // We take a break if we've been at this for a while, 8169 // so as to avoid monopolizing the locks involved. 8170 void SweepClosure::do_yield_work(HeapWord* addr) { 8171 // Return current free chunk being used for coalescing (if any) 8172 // to the appropriate freelist. After yielding, the next 8173 // free block encountered will start a coalescing range of 8174 // free blocks. If the next free block is adjacent to the 8175 // chunk just flushed, they will need to wait for the next 8176 // sweep to be coalesced. 8177 if (inFreeRange()) { 8178 flushCurFreeChunk(freeFinger(), pointer_delta(addr, freeFinger())); 8179 } 8180 8181 // First give up the locks, then yield, then re-lock. 8182 // We should probably use a constructor/destructor idiom to 8183 // do this unlock/lock or modify the MutexUnlocker class to 8184 // serve our purpose. XXX 8185 assert_lock_strong(_bitMap->lock()); 8186 assert_lock_strong(_freelistLock); 8187 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 8188 "CMS thread should hold CMS token"); 8189 _bitMap->lock()->unlock(); 8190 _freelistLock->unlock(); 8191 ConcurrentMarkSweepThread::desynchronize(true); 8192 ConcurrentMarkSweepThread::acknowledge_yield_request(); 8193 _collector->stopTimer(); 8194 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 8195 if (PrintCMSStatistics != 0) { 8196 _collector->incrementYields(); 8197 } 8198 _collector->icms_wait(); 8199 8200 // See the comment in coordinator_yield() 8201 for (unsigned i = 0; i < CMSYieldSleepCount && 8202 ConcurrentMarkSweepThread::should_yield() && 8203 !CMSCollector::foregroundGCIsActive(); ++i) { 8204 os::sleep(Thread::current(), 1, false); 8205 ConcurrentMarkSweepThread::acknowledge_yield_request(); 8206 } 8207 8208 ConcurrentMarkSweepThread::synchronize(true); 8209 _freelistLock->lock(); 8210 _bitMap->lock()->lock_without_safepoint_check(); 8211 _collector->startTimer(); 8212 } 8213 8214 #ifndef PRODUCT 8215 // This is actually very useful in a product build if it can 8216 // be called from the debugger. Compile it into the product 8217 // as needed. 8218 bool debug_verifyChunkInFreeLists(FreeChunk* fc) { 8219 return debug_cms_space->verifyChunkInFreeLists(fc); 8220 } 8221 8222 void SweepClosure::record_free_block_coalesced(FreeChunk* fc) const { 8223 if (CMSTraceSweeper) { 8224 gclog_or_tty->print("Sweep:coal_free_blk 0x%x (%d)\n", fc, fc->size()); 8225 } 8226 } 8227 #endif 8228 8229 // CMSIsAliveClosure 8230 bool CMSIsAliveClosure::do_object_b(oop obj) { 8231 HeapWord* addr = (HeapWord*)obj; 8232 return addr != NULL && 8233 (!_span.contains(addr) || _bit_map->isMarked(addr)); 8234 } 8235 8236 // CMSKeepAliveClosure: the serial version 8237 void CMSKeepAliveClosure::do_oop(oop* p) { 8238 oop this_oop = *p; 8239 HeapWord* addr = (HeapWord*)this_oop; 8240 if (_span.contains(addr) && 8241 !_bit_map->isMarked(addr)) { 8242 _bit_map->mark(addr); 8243 bool simulate_overflow = false; 8244 NOT_PRODUCT( 8245 if (CMSMarkStackOverflowALot && 8246 _collector->simulate_overflow()) { 8247 // simulate a stack overflow 8248 simulate_overflow = true; 8249 } 8250 ) 8251 if (simulate_overflow || !_mark_stack->push(this_oop)) { 8252 _collector->push_on_overflow_list(this_oop); 8253 _collector->_ser_kac_ovflw++; 8254 } 8255 } 8256 } 8257 8258 // CMSParKeepAliveClosure: a parallel version of the above. 8259 // The work queues are private to each closure (thread), 8260 // but (may be) available for stealing by other threads. 8261 void CMSParKeepAliveClosure::do_oop(oop* p) { 8262 oop this_oop = *p; 8263 HeapWord* addr = (HeapWord*)this_oop; 8264 if (_span.contains(addr) && 8265 !_bit_map->isMarked(addr)) { 8266 // In general, during recursive tracing, several threads 8267 // may be concurrently getting here; the first one to 8268 // "tag" it, claims it. 8269 if (_bit_map->par_mark(addr)) { 8270 bool res = _work_queue->push(this_oop); 8271 assert(res, "Low water mark should be much less than capacity"); 8272 // Do a recursive trim in the hope that this will keep 8273 // stack usage lower, but leave some oops for potential stealers 8274 trim_queue(_low_water_mark); 8275 } // Else, another thread got there first 8276 } 8277 } 8278 8279 void CMSParKeepAliveClosure::trim_queue(uint max) { 8280 while (_work_queue->size() > max) { 8281 oop new_oop; 8282 if (_work_queue->pop_local(new_oop)) { 8283 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 8284 assert(_bit_map->isMarked((HeapWord*)new_oop), 8285 "no white objects on this stack!"); 8286 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 8287 // iterate over the oops in this oop, marking and pushing 8288 // the ones in CMS heap (i.e. in _span). 8289 new_oop->oop_iterate(&_mark_and_push); 8290 } 8291 } 8292 } 8293 8294 void CMSInnerParMarkAndPushClosure::do_oop(oop* p) { 8295 oop this_oop = *p; 8296 HeapWord* addr = (HeapWord*)this_oop; 8297 if (_span.contains(addr) && 8298 !_bit_map->isMarked(addr)) { 8299 if (_bit_map->par_mark(addr)) { 8300 bool simulate_overflow = false; 8301 NOT_PRODUCT( 8302 if (CMSMarkStackOverflowALot && 8303 _collector->par_simulate_overflow()) { 8304 // simulate a stack overflow 8305 simulate_overflow = true; 8306 } 8307 ) 8308 if (simulate_overflow || !_work_queue->push(this_oop)) { 8309 _collector->par_push_on_overflow_list(this_oop); 8310 _collector->_par_kac_ovflw++; 8311 } 8312 } // Else another thread got there already 8313 } 8314 } 8315 8316 ////////////////////////////////////////////////////////////////// 8317 // CMSExpansionCause ///////////////////////////// 8318 ////////////////////////////////////////////////////////////////// 8319 const char* CMSExpansionCause::to_string(CMSExpansionCause::Cause cause) { 8320 switch (cause) { 8321 case _no_expansion: 8322 return "No expansion"; 8323 case _satisfy_free_ratio: 8324 return "Free ratio"; 8325 case _satisfy_promotion: 8326 return "Satisfy promotion"; 8327 case _satisfy_allocation: 8328 return "allocation"; 8329 case _allocate_par_lab: 8330 return "Par LAB"; 8331 case _allocate_par_spooling_space: 8332 return "Par Spooling Space"; 8333 case _adaptive_size_policy: 8334 return "Ergonomics"; 8335 default: 8336 return "unknown"; 8337 } 8338 } 8339 8340 void CMSDrainMarkingStackClosure::do_void() { 8341 // the max number to take from overflow list at a time 8342 const size_t num = _mark_stack->capacity()/4; 8343 while (!_mark_stack->isEmpty() || 8344 // if stack is empty, check the overflow list 8345 _collector->take_from_overflow_list(num, _mark_stack)) { 8346 oop this_oop = _mark_stack->pop(); 8347 HeapWord* addr = (HeapWord*)this_oop; 8348 assert(_span.contains(addr), "Should be within span"); 8349 assert(_bit_map->isMarked(addr), "Should be marked"); 8350 assert(this_oop->is_oop(), "Should be an oop"); 8351 this_oop->oop_iterate(_keep_alive); 8352 } 8353 } 8354 8355 void CMSParDrainMarkingStackClosure::do_void() { 8356 // drain queue 8357 trim_queue(0); 8358 } 8359 8360 // Trim our work_queue so its length is below max at return 8361 void CMSParDrainMarkingStackClosure::trim_queue(uint max) { 8362 while (_work_queue->size() > max) { 8363 oop new_oop; 8364 if (_work_queue->pop_local(new_oop)) { 8365 assert(new_oop->is_oop(), "Expected an oop"); 8366 assert(_bit_map->isMarked((HeapWord*)new_oop), 8367 "no white objects on this stack!"); 8368 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 8369 // iterate over the oops in this oop, marking and pushing 8370 // the ones in CMS heap (i.e. in _span). 8371 new_oop->oop_iterate(&_mark_and_push); 8372 } 8373 } 8374 } 8375 8376 //////////////////////////////////////////////////////////////////// 8377 // Support for Marking Stack Overflow list handling and related code 8378 //////////////////////////////////////////////////////////////////// 8379 // Much of the following code is similar in shape and spirit to the 8380 // code used in ParNewGC. We should try and share that code 8381 // as much as possible in the future. 8382 8383 #ifndef PRODUCT 8384 // Debugging support for CMSStackOverflowALot 8385 8386 // It's OK to call this multi-threaded; the worst thing 8387 // that can happen is that we'll get a bunch of closely 8388 // spaced simulated oveflows, but that's OK, in fact 8389 // probably good as it would exercise the overflow code 8390 // under contention. 8391 bool CMSCollector::simulate_overflow() { 8392 if (_overflow_counter-- <= 0) { // just being defensive 8393 _overflow_counter = CMSMarkStackOverflowInterval; 8394 return true; 8395 } else { 8396 return false; 8397 } 8398 } 8399 8400 bool CMSCollector::par_simulate_overflow() { 8401 return simulate_overflow(); 8402 } 8403 #endif 8404 8405 // Single-threaded 8406 bool CMSCollector::take_from_overflow_list(size_t num, CMSMarkStack* stack) { 8407 assert(stack->isEmpty(), "Expected precondition"); 8408 assert(stack->capacity() > num, "Shouldn't bite more than can chew"); 8409 size_t i = num; 8410 oop cur = _overflow_list; 8411 const markOop proto = markOopDesc::prototype(); 8412 NOT_PRODUCT(size_t n = 0;) 8413 for (oop next; i > 0 && cur != NULL; cur = next, i--) { 8414 next = oop(cur->mark()); 8415 cur->set_mark(proto); // until proven otherwise 8416 assert(cur->is_oop(), "Should be an oop"); 8417 bool res = stack->push(cur); 8418 assert(res, "Bit off more than can chew?"); 8419 NOT_PRODUCT(n++;) 8420 } 8421 _overflow_list = cur; 8422 #ifndef PRODUCT 8423 assert(_num_par_pushes >= n, "Too many pops?"); 8424 _num_par_pushes -=n; 8425 #endif 8426 return !stack->isEmpty(); 8427 } 8428 8429 // Multi-threaded; use CAS to break off a prefix 8430 bool CMSCollector::par_take_from_overflow_list(size_t num, 8431 OopTaskQueue* work_q) { 8432 assert(work_q->size() == 0, "That's the current policy"); 8433 assert(num < work_q->max_elems(), "Can't bite more than we can chew"); 8434 if (_overflow_list == NULL) { 8435 return false; 8436 } 8437 // Grab the entire list; we'll put back a suffix 8438 oop prefix = (oop)Atomic::xchg_ptr(NULL, &_overflow_list); 8439 if (prefix == NULL) { // someone grabbed it before we did ... 8440 // ... we could spin for a short while, but for now we don't 8441 return false; 8442 } 8443 size_t i = num; 8444 oop cur = prefix; 8445 for (; i > 1 && cur->mark() != NULL; cur = oop(cur->mark()), i--); 8446 if (cur->mark() != NULL) { 8447 oop suffix_head = cur->mark(); // suffix will be put back on global list 8448 cur->set_mark(NULL); // break off suffix 8449 // Find tail of suffix so we can prepend suffix to global list 8450 for (cur = suffix_head; cur->mark() != NULL; cur = (oop)(cur->mark())); 8451 oop suffix_tail = cur; 8452 assert(suffix_tail != NULL && suffix_tail->mark() == NULL, 8453 "Tautology"); 8454 oop observed_overflow_list = _overflow_list; 8455 do { 8456 cur = observed_overflow_list; 8457 suffix_tail->set_mark(markOop(cur)); 8458 observed_overflow_list = 8459 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur); 8460 } while (cur != observed_overflow_list); 8461 } 8462 8463 // Push the prefix elements on work_q 8464 assert(prefix != NULL, "control point invariant"); 8465 const markOop proto = markOopDesc::prototype(); 8466 oop next; 8467 NOT_PRODUCT(size_t n = 0;) 8468 for (cur = prefix; cur != NULL; cur = next) { 8469 next = oop(cur->mark()); 8470 cur->set_mark(proto); // until proven otherwise 8471 assert(cur->is_oop(), "Should be an oop"); 8472 bool res = work_q->push(cur); 8473 assert(res, "Bit off more than we can chew?"); 8474 NOT_PRODUCT(n++;) 8475 } 8476 #ifndef PRODUCT 8477 assert(_num_par_pushes >= n, "Too many pops?"); 8478 Atomic::add_ptr(-(intptr_t)n, &_num_par_pushes); 8479 #endif 8480 return true; 8481 } 8482 8483 // Single-threaded 8484 void CMSCollector::push_on_overflow_list(oop p) { 8485 NOT_PRODUCT(_num_par_pushes++;) 8486 assert(p->is_oop(), "Not an oop"); 8487 preserve_mark_if_necessary(p); 8488 p->set_mark((markOop)_overflow_list); 8489 _overflow_list = p; 8490 } 8491 8492 // Multi-threaded; use CAS to prepend to overflow list 8493 void CMSCollector::par_push_on_overflow_list(oop p) { 8494 NOT_PRODUCT(Atomic::inc_ptr(&_num_par_pushes);) 8495 assert(p->is_oop(), "Not an oop"); 8496 par_preserve_mark_if_necessary(p); 8497 oop observed_overflow_list = _overflow_list; 8498 oop cur_overflow_list; 8499 do { 8500 cur_overflow_list = observed_overflow_list; 8501 p->set_mark(markOop(cur_overflow_list)); 8502 observed_overflow_list = 8503 (oop) Atomic::cmpxchg_ptr(p, &_overflow_list, cur_overflow_list); 8504 } while (cur_overflow_list != observed_overflow_list); 8505 } 8506 8507 // Single threaded 8508 // General Note on GrowableArray: pushes may silently fail 8509 // because we are (temporarily) out of C-heap for expanding 8510 // the stack. The problem is quite ubiquitous and affects 8511 // a lot of code in the JVM. The prudent thing for GrowableArray 8512 // to do (for now) is to exit with an error. However, that may 8513 // be too draconian in some cases because the caller may be 8514 // able to recover without much harm. For suych cases, we 8515 // should probably introduce a "soft_push" method which returns 8516 // an indication of success or failure with the assumption that 8517 // the caller may be able to recover from a failure; code in 8518 // the VM can then be changed, incrementally, to deal with such 8519 // failures where possible, thus, incrementally hardening the VM 8520 // in such low resource situations. 8521 void CMSCollector::preserve_mark_work(oop p, markOop m) { 8522 int PreserveMarkStackSize = 128; 8523 8524 if (_preserved_oop_stack == NULL) { 8525 assert(_preserved_mark_stack == NULL, 8526 "bijection with preserved_oop_stack"); 8527 // Allocate the stacks 8528 _preserved_oop_stack = new (ResourceObj::C_HEAP) 8529 GrowableArray<oop>(PreserveMarkStackSize, true); 8530 _preserved_mark_stack = new (ResourceObj::C_HEAP) 8531 GrowableArray<markOop>(PreserveMarkStackSize, true); 8532 if (_preserved_oop_stack == NULL || _preserved_mark_stack == NULL) { 8533 vm_exit_out_of_memory(2* PreserveMarkStackSize * sizeof(oop) /* punt */, 8534 "Preserved Mark/Oop Stack for CMS (C-heap)"); 8535 } 8536 } 8537 _preserved_oop_stack->push(p); 8538 _preserved_mark_stack->push(m); 8539 assert(m == p->mark(), "Mark word changed"); 8540 assert(_preserved_oop_stack->length() == _preserved_mark_stack->length(), 8541 "bijection"); 8542 } 8543 8544 // Single threaded 8545 void CMSCollector::preserve_mark_if_necessary(oop p) { 8546 markOop m = p->mark(); 8547 if (m->must_be_preserved(p)) { 8548 preserve_mark_work(p, m); 8549 } 8550 } 8551 8552 void CMSCollector::par_preserve_mark_if_necessary(oop p) { 8553 markOop m = p->mark(); 8554 if (m->must_be_preserved(p)) { 8555 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag); 8556 // Even though we read the mark word without holding 8557 // the lock, we are assured that it will not change 8558 // because we "own" this oop, so no other thread can 8559 // be trying to push it on the overflow list; see 8560 // the assertion in preserve_mark_work() that checks 8561 // that m == p->mark(). 8562 preserve_mark_work(p, m); 8563 } 8564 } 8565 8566 // We should be able to do this multi-threaded, 8567 // a chunk of stack being a task (this is 8568 // correct because each oop only ever appears 8569 // once in the overflow list. However, it's 8570 // not very easy to completely overlap this with 8571 // other operations, so will generally not be done 8572 // until all work's been completed. Because we 8573 // expect the preserved oop stack (set) to be small, 8574 // it's probably fine to do this single-threaded. 8575 // We can explore cleverer concurrent/overlapped/parallel 8576 // processing of preserved marks if we feel the 8577 // need for this in the future. Stack overflow should 8578 // be so rare in practice and, when it happens, its 8579 // effect on performance so great that this will 8580 // likely just be in the noise anyway. 8581 void CMSCollector::restore_preserved_marks_if_any() { 8582 if (_preserved_oop_stack == NULL) { 8583 assert(_preserved_mark_stack == NULL, 8584 "bijection with preserved_oop_stack"); 8585 return; 8586 } 8587 8588 assert(SafepointSynchronize::is_at_safepoint(), 8589 "world should be stopped"); 8590 assert(Thread::current()->is_ConcurrentGC_thread() || 8591 Thread::current()->is_VM_thread(), 8592 "should be single-threaded"); 8593 8594 int length = _preserved_oop_stack->length(); 8595 assert(_preserved_mark_stack->length() == length, "bijection"); 8596 for (int i = 0; i < length; i++) { 8597 oop p = _preserved_oop_stack->at(i); 8598 assert(p->is_oop(), "Should be an oop"); 8599 assert(_span.contains(p), "oop should be in _span"); 8600 assert(p->mark() == markOopDesc::prototype(), 8601 "Set when taken from overflow list"); 8602 markOop m = _preserved_mark_stack->at(i); 8603 p->set_mark(m); 8604 } 8605 _preserved_mark_stack->clear(); 8606 _preserved_oop_stack->clear(); 8607 assert(_preserved_mark_stack->is_empty() && 8608 _preserved_oop_stack->is_empty(), 8609 "stacks were cleared above"); 8610 } 8611 8612 #ifndef PRODUCT 8613 bool CMSCollector::no_preserved_marks() const { 8614 return ( ( _preserved_mark_stack == NULL 8615 && _preserved_oop_stack == NULL) 8616 || ( _preserved_mark_stack->is_empty() 8617 && _preserved_oop_stack->is_empty())); 8618 } 8619 #endif 8620 8621 CMSAdaptiveSizePolicy* ASConcurrentMarkSweepGeneration::cms_size_policy() const 8622 { 8623 GenCollectedHeap* gch = (GenCollectedHeap*) GenCollectedHeap::heap(); 8624 CMSAdaptiveSizePolicy* size_policy = 8625 (CMSAdaptiveSizePolicy*) gch->gen_policy()->size_policy(); 8626 assert(size_policy->is_gc_cms_adaptive_size_policy(), 8627 "Wrong type for size policy"); 8628 return size_policy; 8629 } 8630 8631 void ASConcurrentMarkSweepGeneration::resize(size_t cur_promo_size, 8632 size_t desired_promo_size) { 8633 if (cur_promo_size < desired_promo_size) { 8634 size_t expand_bytes = desired_promo_size - cur_promo_size; 8635 if (PrintAdaptiveSizePolicy && Verbose) { 8636 gclog_or_tty->print_cr(" ASConcurrentMarkSweepGeneration::resize " 8637 "Expanding tenured generation by " SIZE_FORMAT " (bytes)", 8638 expand_bytes); 8639 } 8640 expand(expand_bytes, 8641 MinHeapDeltaBytes, 8642 CMSExpansionCause::_adaptive_size_policy); 8643 } else if (desired_promo_size < cur_promo_size) { 8644 size_t shrink_bytes = cur_promo_size - desired_promo_size; 8645 if (PrintAdaptiveSizePolicy && Verbose) { 8646 gclog_or_tty->print_cr(" ASConcurrentMarkSweepGeneration::resize " 8647 "Shrinking tenured generation by " SIZE_FORMAT " (bytes)", 8648 shrink_bytes); 8649 } 8650 shrink(shrink_bytes); 8651 } 8652 } 8653 8654 CMSGCAdaptivePolicyCounters* ASConcurrentMarkSweepGeneration::gc_adaptive_policy_counters() { 8655 GenCollectedHeap* gch = GenCollectedHeap::heap(); 8656 CMSGCAdaptivePolicyCounters* counters = 8657 (CMSGCAdaptivePolicyCounters*) gch->collector_policy()->counters(); 8658 assert(counters->kind() == GCPolicyCounters::CMSGCAdaptivePolicyCountersKind, 8659 "Wrong kind of counters"); 8660 return counters; 8661 } 8662 8663 8664 void ASConcurrentMarkSweepGeneration::update_counters() { 8665 if (UsePerfData) { 8666 _space_counters->update_all(); 8667 _gen_counters->update_all(); 8668 CMSGCAdaptivePolicyCounters* counters = gc_adaptive_policy_counters(); 8669 GenCollectedHeap* gch = GenCollectedHeap::heap(); 8670 CMSGCStats* gc_stats_l = (CMSGCStats*) gc_stats(); 8671 assert(gc_stats_l->kind() == GCStats::CMSGCStatsKind, 8672 "Wrong gc statistics type"); 8673 counters->update_counters(gc_stats_l); 8674 } 8675 } 8676 8677 void ASConcurrentMarkSweepGeneration::update_counters(size_t used) { 8678 if (UsePerfData) { 8679 _space_counters->update_used(used); 8680 _space_counters->update_capacity(); 8681 _gen_counters->update_all(); 8682 8683 CMSGCAdaptivePolicyCounters* counters = gc_adaptive_policy_counters(); 8684 GenCollectedHeap* gch = GenCollectedHeap::heap(); 8685 CMSGCStats* gc_stats_l = (CMSGCStats*) gc_stats(); 8686 assert(gc_stats_l->kind() == GCStats::CMSGCStatsKind, 8687 "Wrong gc statistics type"); 8688 counters->update_counters(gc_stats_l); 8689 } 8690 } 8691 8692 // The desired expansion delta is computed so that: 8693 // . desired free percentage or greater is used 8694 void ASConcurrentMarkSweepGeneration::compute_new_size() { 8695 assert_locked_or_safepoint(Heap_lock); 8696 8697 GenCollectedHeap* gch = (GenCollectedHeap*) GenCollectedHeap::heap(); 8698 8699 // If incremental collection failed, we just want to expand 8700 // to the limit. 8701 if (incremental_collection_failed()) { 8702 clear_incremental_collection_failed(); 8703 grow_to_reserved(); 8704 return; 8705 } 8706 8707 assert(UseAdaptiveSizePolicy, "Should be using adaptive sizing"); 8708 8709 assert(gch->kind() == CollectedHeap::GenCollectedHeap, 8710 "Wrong type of heap"); 8711 int prev_level = level() - 1; 8712 assert(prev_level >= 0, "The cms generation is the lowest generation"); 8713 Generation* prev_gen = gch->get_gen(prev_level); 8714 assert(prev_gen->kind() == Generation::ASParNew, 8715 "Wrong type of young generation"); 8716 ParNewGeneration* younger_gen = (ParNewGeneration*) prev_gen; 8717 size_t cur_eden = younger_gen->eden()->capacity(); 8718 CMSAdaptiveSizePolicy* size_policy = cms_size_policy(); 8719 size_t cur_promo = free(); 8720 size_policy->compute_tenured_generation_free_space(cur_promo, 8721 max_available(), 8722 cur_eden); 8723 resize(cur_promo, size_policy->promo_size()); 8724 8725 // Record the new size of the space in the cms generation 8726 // that is available for promotions. This is temporary. 8727 // It should be the desired promo size. 8728 size_policy->avg_cms_promo()->sample(free()); 8729 size_policy->avg_old_live()->sample(used()); 8730 8731 if (UsePerfData) { 8732 CMSGCAdaptivePolicyCounters* counters = gc_adaptive_policy_counters(); 8733 counters->update_cms_capacity_counter(capacity()); 8734 } 8735 } 8736 8737 void ASConcurrentMarkSweepGeneration::shrink_by(size_t desired_bytes) { 8738 assert_locked_or_safepoint(Heap_lock); 8739 assert_lock_strong(freelistLock()); 8740 HeapWord* old_end = _cmsSpace->end(); 8741 HeapWord* unallocated_start = _cmsSpace->unallocated_block(); 8742 assert(old_end >= unallocated_start, "Miscalculation of unallocated_start"); 8743 FreeChunk* chunk_at_end = find_chunk_at_end(); 8744 if (chunk_at_end == NULL) { 8745 // No room to shrink 8746 if (PrintGCDetails && Verbose) { 8747 gclog_or_tty->print_cr("No room to shrink: old_end " 8748 PTR_FORMAT " unallocated_start " PTR_FORMAT 8749 " chunk_at_end " PTR_FORMAT, 8750 old_end, unallocated_start, chunk_at_end); 8751 } 8752 return; 8753 } else { 8754 8755 // Find the chunk at the end of the space and determine 8756 // how much it can be shrunk. 8757 size_t shrinkable_size_in_bytes = chunk_at_end->size(); 8758 size_t aligned_shrinkable_size_in_bytes = 8759 align_size_down(shrinkable_size_in_bytes, os::vm_page_size()); 8760 assert(unallocated_start <= chunk_at_end->end(), 8761 "Inconsistent chunk at end of space"); 8762 size_t bytes = MIN2(desired_bytes, aligned_shrinkable_size_in_bytes); 8763 size_t word_size_before = heap_word_size(_virtual_space.committed_size()); 8764 8765 // Shrink the underlying space 8766 _virtual_space.shrink_by(bytes); 8767 if (PrintGCDetails && Verbose) { 8768 gclog_or_tty->print_cr("ConcurrentMarkSweepGeneration::shrink_by:" 8769 " desired_bytes " SIZE_FORMAT 8770 " shrinkable_size_in_bytes " SIZE_FORMAT 8771 " aligned_shrinkable_size_in_bytes " SIZE_FORMAT 8772 " bytes " SIZE_FORMAT, 8773 desired_bytes, shrinkable_size_in_bytes, 8774 aligned_shrinkable_size_in_bytes, bytes); 8775 gclog_or_tty->print_cr(" old_end " SIZE_FORMAT 8776 " unallocated_start " SIZE_FORMAT, 8777 old_end, unallocated_start); 8778 } 8779 8780 // If the space did shrink (shrinking is not guaranteed), 8781 // shrink the chunk at the end by the appropriate amount. 8782 if (((HeapWord*)_virtual_space.high()) < old_end) { 8783 size_t new_word_size = 8784 heap_word_size(_virtual_space.committed_size()); 8785 8786 // Have to remove the chunk from the dictionary because it is changing 8787 // size and might be someplace elsewhere in the dictionary. 8788 8789 // Get the chunk at end, shrink it, and put it 8790 // back. 8791 _cmsSpace->removeChunkFromDictionary(chunk_at_end); 8792 size_t word_size_change = word_size_before - new_word_size; 8793 size_t chunk_at_end_old_size = chunk_at_end->size(); 8794 assert(chunk_at_end_old_size >= word_size_change, 8795 "Shrink is too large"); 8796 chunk_at_end->setSize(chunk_at_end_old_size - 8797 word_size_change); 8798 _cmsSpace->freed((HeapWord*) chunk_at_end->end(), 8799 word_size_change); 8800 8801 _cmsSpace->returnChunkToDictionary(chunk_at_end); 8802 8803 MemRegion mr(_cmsSpace->bottom(), new_word_size); 8804 _bts->resize(new_word_size); // resize the block offset shared array 8805 Universe::heap()->barrier_set()->resize_covered_region(mr); 8806 _cmsSpace->assert_locked(); 8807 _cmsSpace->set_end((HeapWord*)_virtual_space.high()); 8808 8809 NOT_PRODUCT(_cmsSpace->dictionary()->verify()); 8810 8811 // update the space and generation capacity counters 8812 if (UsePerfData) { 8813 _space_counters->update_capacity(); 8814 _gen_counters->update_all(); 8815 } 8816 8817 if (Verbose && PrintGCDetails) { 8818 size_t new_mem_size = _virtual_space.committed_size(); 8819 size_t old_mem_size = new_mem_size + bytes; 8820 gclog_or_tty->print_cr("Shrinking %s from %ldK by %ldK to %ldK", 8821 name(), old_mem_size/K, bytes/K, new_mem_size/K); 8822 } 8823 } 8824 8825 assert(_cmsSpace->unallocated_block() <= _cmsSpace->end(), 8826 "Inconsistency at end of space"); 8827 assert(chunk_at_end->end() == _cmsSpace->end(), 8828 "Shrinking is inconsistent"); 8829 return; 8830 } 8831 } 8832 8833 // Transfer some number of overflown objects to usual marking 8834 // stack. Return true if some objects were transferred. 8835 bool MarkRefsIntoAndScanClosure::take_from_overflow_list() { 8836 size_t num = MIN2((size_t)_mark_stack->capacity()/4, 8837 (size_t)ParGCDesiredObjsFromOverflowList); 8838 8839 bool res = _collector->take_from_overflow_list(num, _mark_stack); 8840 assert(_collector->overflow_list_is_empty() || res, 8841 "If list is not empty, we should have taken something"); 8842 assert(!res || !_mark_stack->isEmpty(), 8843 "If we took something, it should now be on our stack"); 8844 return res; 8845 } 8846 8847 size_t MarkDeadObjectsClosure::do_blk(HeapWord* addr) { 8848 size_t res = _sp->block_size_no_stall(addr, _collector); 8849 assert(res != 0, "Should always be able to compute a size"); 8850 if (_sp->block_is_obj(addr)) { 8851 if (_live_bit_map->isMarked(addr)) { 8852 // It can't have been dead in a previous cycle 8853 guarantee(!_dead_bit_map->isMarked(addr), "No resurrection!"); 8854 } else { 8855 _dead_bit_map->mark(addr); // mark the dead object 8856 } 8857 } 8858 return res; 8859 }