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