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 void ConcurrentMarkSweepGeneration::ref_processor_init() { 229 assert(collector() != NULL, "no collector"); 230 collector()->ref_processor_init(); 231 } 232 233 void CMSCollector::ref_processor_init() { 234 if (_ref_processor == NULL) { 235 // Allocate and initialize a reference processor 236 _ref_processor = ReferenceProcessor::create_ref_processor( 237 _span, // span 238 _cmsGen->refs_discovery_is_atomic(), // atomic_discovery 239 _cmsGen->refs_discovery_is_mt(), // mt_discovery 240 &_is_alive_closure, 241 ParallelGCThreads, 242 ParallelRefProcEnabled); 243 // Initialize the _ref_processor field of CMSGen 244 _cmsGen->set_ref_processor(_ref_processor); 245 246 // Allocate a dummy ref processor for perm gen. 247 ReferenceProcessor* rp2 = new ReferenceProcessor(); 248 if (rp2 == NULL) { 249 vm_exit_during_initialization("Could not allocate ReferenceProcessor object"); 250 } 251 _permGen->set_ref_processor(rp2); 252 } 253 } 254 255 CMSAdaptiveSizePolicy* CMSCollector::size_policy() { 256 GenCollectedHeap* gch = GenCollectedHeap::heap(); 257 assert(gch->kind() == CollectedHeap::GenCollectedHeap, 258 "Wrong type of heap"); 259 CMSAdaptiveSizePolicy* sp = (CMSAdaptiveSizePolicy*) 260 gch->gen_policy()->size_policy(); 261 assert(sp->is_gc_cms_adaptive_size_policy(), 262 "Wrong type of size policy"); 263 return sp; 264 } 265 266 CMSGCAdaptivePolicyCounters* CMSCollector::gc_adaptive_policy_counters() { 267 CMSGCAdaptivePolicyCounters* results = 268 (CMSGCAdaptivePolicyCounters*) collector_policy()->counters(); 269 assert( 270 results->kind() == GCPolicyCounters::CMSGCAdaptivePolicyCountersKind, 271 "Wrong gc policy counter kind"); 272 return results; 273 } 274 275 276 void ConcurrentMarkSweepGeneration::initialize_performance_counters() { 277 278 const char* gen_name = "old"; 279 280 // Generation Counters - generation 1, 1 subspace 281 _gen_counters = new GenerationCounters(gen_name, 1, 1, &_virtual_space); 282 283 _space_counters = new GSpaceCounters(gen_name, 0, 284 _virtual_space.reserved_size(), 285 this, _gen_counters); 286 } 287 288 CMSStats::CMSStats(ConcurrentMarkSweepGeneration* cms_gen, unsigned int alpha): 289 _cms_gen(cms_gen) 290 { 291 assert(alpha <= 100, "bad value"); 292 _saved_alpha = alpha; 293 294 // Initialize the alphas to the bootstrap value of 100. 295 _gc0_alpha = _cms_alpha = 100; 296 297 _cms_begin_time.update(); 298 _cms_end_time.update(); 299 300 _gc0_duration = 0.0; 301 _gc0_period = 0.0; 302 _gc0_promoted = 0; 303 304 _cms_duration = 0.0; 305 _cms_period = 0.0; 306 _cms_allocated = 0; 307 308 _cms_used_at_gc0_begin = 0; 309 _cms_used_at_gc0_end = 0; 310 _allow_duty_cycle_reduction = false; 311 _valid_bits = 0; 312 _icms_duty_cycle = CMSIncrementalDutyCycle; 313 } 314 315 // If promotion failure handling is on use 316 // the padded average size of the promotion for each 317 // young generation collection. 318 double CMSStats::time_until_cms_gen_full() const { 319 size_t cms_free = _cms_gen->cmsSpace()->free(); 320 GenCollectedHeap* gch = GenCollectedHeap::heap(); 321 size_t expected_promotion = gch->get_gen(0)->capacity(); 322 if (HandlePromotionFailure) { 323 expected_promotion = MIN2( 324 (size_t) _cms_gen->gc_stats()->avg_promoted()->padded_average(), 325 expected_promotion); 326 } 327 if (cms_free > expected_promotion) { 328 // Start a cms collection if there isn't enough space to promote 329 // for the next minor collection. Use the padded average as 330 // a safety factor. 331 cms_free -= expected_promotion; 332 333 // Adjust by the safety factor. 334 double cms_free_dbl = (double)cms_free; 335 cms_free_dbl = cms_free_dbl * (100.0 - CMSIncrementalSafetyFactor) / 100.0; 336 337 if (PrintGCDetails && Verbose) { 338 gclog_or_tty->print_cr("CMSStats::time_until_cms_gen_full: cms_free " 339 SIZE_FORMAT " expected_promotion " SIZE_FORMAT, 340 cms_free, expected_promotion); 341 gclog_or_tty->print_cr(" cms_free_dbl %f cms_consumption_rate %f", 342 cms_free_dbl, cms_consumption_rate() + 1.0); 343 } 344 // Add 1 in case the consumption rate goes to zero. 345 return cms_free_dbl / (cms_consumption_rate() + 1.0); 346 } 347 return 0.0; 348 } 349 350 // Compare the duration of the cms collection to the 351 // time remaining before the cms generation is empty. 352 // Note that the time from the start of the cms collection 353 // to the start of the cms sweep (less than the total 354 // duration of the cms collection) can be used. This 355 // has been tried and some applications experienced 356 // promotion failures early in execution. This was 357 // possibly because the averages were not accurate 358 // enough at the beginning. 359 double CMSStats::time_until_cms_start() const { 360 // We add "gc0_period" to the "work" calculation 361 // below because this query is done (mostly) at the 362 // end of a scavenge, so we need to conservatively 363 // account for that much possible delay 364 // in the query so as to avoid concurrent mode failures 365 // due to starting the collection just a wee bit too 366 // late. 367 double work = cms_duration() + gc0_period(); 368 double deadline = time_until_cms_gen_full(); 369 if (work > deadline) { 370 if (Verbose && PrintGCDetails) { 371 gclog_or_tty->print( 372 " CMSCollector: collect because of anticipated promotion " 373 "before full %3.7f + %3.7f > %3.7f ", cms_duration(), 374 gc0_period(), time_until_cms_gen_full()); 375 } 376 return 0.0; 377 } 378 return work - deadline; 379 } 380 381 // Return a duty cycle based on old_duty_cycle and new_duty_cycle, limiting the 382 // amount of change to prevent wild oscillation. 383 unsigned int CMSStats::icms_damped_duty_cycle(unsigned int old_duty_cycle, 384 unsigned int new_duty_cycle) { 385 assert(old_duty_cycle <= 100, "bad input value"); 386 assert(new_duty_cycle <= 100, "bad input value"); 387 388 // Note: use subtraction with caution since it may underflow (values are 389 // unsigned). Addition is safe since we're in the range 0-100. 390 unsigned int damped_duty_cycle = new_duty_cycle; 391 if (new_duty_cycle < old_duty_cycle) { 392 const unsigned int largest_delta = MAX2(old_duty_cycle / 4, 5U); 393 if (new_duty_cycle + largest_delta < old_duty_cycle) { 394 damped_duty_cycle = old_duty_cycle - largest_delta; 395 } 396 } else if (new_duty_cycle > old_duty_cycle) { 397 const unsigned int largest_delta = MAX2(old_duty_cycle / 4, 15U); 398 if (new_duty_cycle > old_duty_cycle + largest_delta) { 399 damped_duty_cycle = MIN2(old_duty_cycle + largest_delta, 100U); 400 } 401 } 402 assert(damped_duty_cycle <= 100, "invalid duty cycle computed"); 403 404 if (CMSTraceIncrementalPacing) { 405 gclog_or_tty->print(" [icms_damped_duty_cycle(%d,%d) = %d] ", 406 old_duty_cycle, new_duty_cycle, damped_duty_cycle); 407 } 408 return damped_duty_cycle; 409 } 410 411 unsigned int CMSStats::icms_update_duty_cycle_impl() { 412 assert(CMSIncrementalPacing && valid(), 413 "should be handled in icms_update_duty_cycle()"); 414 415 double cms_time_so_far = cms_timer().seconds(); 416 double scaled_duration = cms_duration_per_mb() * _cms_used_at_gc0_end / M; 417 double scaled_duration_remaining = fabsd(scaled_duration - cms_time_so_far); 418 419 // Avoid division by 0. 420 double time_until_full = MAX2(time_until_cms_gen_full(), 0.01); 421 double duty_cycle_dbl = 100.0 * scaled_duration_remaining / time_until_full; 422 423 unsigned int new_duty_cycle = MIN2((unsigned int)duty_cycle_dbl, 100U); 424 if (new_duty_cycle > _icms_duty_cycle) { 425 // Avoid very small duty cycles (1 or 2); 0 is allowed. 426 if (new_duty_cycle > 2) { 427 _icms_duty_cycle = icms_damped_duty_cycle(_icms_duty_cycle, 428 new_duty_cycle); 429 } 430 } else if (_allow_duty_cycle_reduction) { 431 // The duty cycle is reduced only once per cms cycle (see record_cms_end()). 432 new_duty_cycle = icms_damped_duty_cycle(_icms_duty_cycle, new_duty_cycle); 433 // Respect the minimum duty cycle. 434 unsigned int min_duty_cycle = (unsigned int)CMSIncrementalDutyCycleMin; 435 _icms_duty_cycle = MAX2(new_duty_cycle, min_duty_cycle); 436 } 437 438 if (PrintGCDetails || CMSTraceIncrementalPacing) { 439 gclog_or_tty->print(" icms_dc=%d ", _icms_duty_cycle); 440 } 441 442 _allow_duty_cycle_reduction = false; 443 return _icms_duty_cycle; 444 } 445 446 #ifndef PRODUCT 447 void CMSStats::print_on(outputStream *st) const { 448 st->print(" gc0_alpha=%d,cms_alpha=%d", _gc0_alpha, _cms_alpha); 449 st->print(",gc0_dur=%g,gc0_per=%g,gc0_promo=" SIZE_FORMAT, 450 gc0_duration(), gc0_period(), gc0_promoted()); 451 st->print(",cms_dur=%g,cms_dur_per_mb=%g,cms_per=%g,cms_alloc=" SIZE_FORMAT, 452 cms_duration(), cms_duration_per_mb(), 453 cms_period(), cms_allocated()); 454 st->print(",cms_since_beg=%g,cms_since_end=%g", 455 cms_time_since_begin(), cms_time_since_end()); 456 st->print(",cms_used_beg=" SIZE_FORMAT ",cms_used_end=" SIZE_FORMAT, 457 _cms_used_at_gc0_begin, _cms_used_at_gc0_end); 458 if (CMSIncrementalMode) { 459 st->print(",dc=%d", icms_duty_cycle()); 460 } 461 462 if (valid()) { 463 st->print(",promo_rate=%g,cms_alloc_rate=%g", 464 promotion_rate(), cms_allocation_rate()); 465 st->print(",cms_consumption_rate=%g,time_until_full=%g", 466 cms_consumption_rate(), time_until_cms_gen_full()); 467 } 468 st->print(" "); 469 } 470 #endif // #ifndef PRODUCT 471 472 CMSCollector::CollectorState CMSCollector::_collectorState = 473 CMSCollector::Idling; 474 bool CMSCollector::_foregroundGCIsActive = false; 475 bool CMSCollector::_foregroundGCShouldWait = false; 476 477 CMSCollector::CMSCollector(ConcurrentMarkSweepGeneration* cmsGen, 478 ConcurrentMarkSweepGeneration* permGen, 479 CardTableRS* ct, 480 ConcurrentMarkSweepPolicy* cp): 481 _cmsGen(cmsGen), 482 _permGen(permGen), 483 _ct(ct), 484 _ref_processor(NULL), // will be set later 485 _conc_workers(NULL), // may be set later 486 _abort_preclean(false), 487 _start_sampling(false), 488 _between_prologue_and_epilogue(false), 489 _markBitMap(0, Mutex::leaf + 1, "CMS_markBitMap_lock"), 490 _perm_gen_verify_bit_map(0, -1 /* no mutex */, "No_lock"), 491 _modUnionTable((CardTableModRefBS::card_shift - LogHeapWordSize), 492 -1 /* lock-free */, "No_lock" /* dummy */), 493 _modUnionClosure(&_modUnionTable), 494 _modUnionClosurePar(&_modUnionTable), 495 _is_alive_closure(&_markBitMap), 496 _restart_addr(NULL), 497 _overflow_list(NULL), 498 _preserved_oop_stack(NULL), 499 _preserved_mark_stack(NULL), 500 _stats(cmsGen), 501 _eden_chunk_array(NULL), // may be set in ctor body 502 _eden_chunk_capacity(0), // -- ditto -- 503 _eden_chunk_index(0), // -- ditto -- 504 _survivor_plab_array(NULL), // -- ditto -- 505 _survivor_chunk_array(NULL), // -- ditto -- 506 _survivor_chunk_capacity(0), // -- ditto -- 507 _survivor_chunk_index(0), // -- ditto -- 508 _ser_pmc_preclean_ovflw(0), 509 _ser_pmc_remark_ovflw(0), 510 _par_pmc_remark_ovflw(0), 511 _ser_kac_ovflw(0), 512 _par_kac_ovflw(0), 513 #ifndef PRODUCT 514 _num_par_pushes(0), 515 #endif 516 _collection_count_start(0), 517 _verifying(false), 518 _icms_start_limit(NULL), 519 _icms_stop_limit(NULL), 520 _verification_mark_bm(0, Mutex::leaf + 1, "CMS_verification_mark_bm_lock"), 521 _completed_initialization(false), 522 _collector_policy(cp), 523 _unload_classes(false), 524 _unloaded_classes_last_cycle(false), 525 _sweep_estimate(CMS_SweepWeight, CMS_SweepPadding) 526 { 527 if (ExplicitGCInvokesConcurrentAndUnloadsClasses) { 528 ExplicitGCInvokesConcurrent = true; 529 } 530 // Now expand the span and allocate the collection support structures 531 // (MUT, marking bit map etc.) to cover both generations subject to 532 // collection. 533 534 // First check that _permGen is adjacent to _cmsGen and above it. 535 assert( _cmsGen->reserved().word_size() > 0 536 && _permGen->reserved().word_size() > 0, 537 "generations should not be of zero size"); 538 assert(_cmsGen->reserved().intersection(_permGen->reserved()).is_empty(), 539 "_cmsGen and _permGen should not overlap"); 540 assert(_cmsGen->reserved().end() == _permGen->reserved().start(), 541 "_cmsGen->end() different from _permGen->start()"); 542 543 // For use by dirty card to oop closures. 544 _cmsGen->cmsSpace()->set_collector(this); 545 _permGen->cmsSpace()->set_collector(this); 546 547 // Adjust my span to cover old (cms) gen and perm gen 548 _span = _cmsGen->reserved()._union(_permGen->reserved()); 549 // Initialize the span of is_alive_closure 550 _is_alive_closure.set_span(_span); 551 552 // Allocate MUT and marking bit map 553 { 554 MutexLockerEx x(_markBitMap.lock(), Mutex::_no_safepoint_check_flag); 555 if (!_markBitMap.allocate(_span)) { 556 warning("Failed to allocate CMS Bit Map"); 557 return; 558 } 559 assert(_markBitMap.covers(_span), "_markBitMap inconsistency?"); 560 } 561 { 562 _modUnionTable.allocate(_span); 563 assert(_modUnionTable.covers(_span), "_modUnionTable inconsistency?"); 564 } 565 566 if (!_markStack.allocate(CMSMarkStackSize)) { 567 warning("Failed to allocate CMS Marking Stack"); 568 return; 569 } 570 if (!_revisitStack.allocate(CMSRevisitStackSize)) { 571 warning("Failed to allocate CMS Revisit Stack"); 572 return; 573 } 574 575 // Support for multi-threaded concurrent phases 576 if (ParallelGCThreads > 0 && CMSConcurrentMTEnabled) { 577 if (FLAG_IS_DEFAULT(ParallelCMSThreads)) { 578 // just for now 579 FLAG_SET_DEFAULT(ParallelCMSThreads, (ParallelGCThreads + 3)/4); 580 } 581 if (ParallelCMSThreads > 1) { 582 _conc_workers = new YieldingFlexibleWorkGang("Parallel CMS Threads", 583 ParallelCMSThreads, true); 584 if (_conc_workers == NULL) { 585 warning("GC/CMS: _conc_workers allocation failure: " 586 "forcing -CMSConcurrentMTEnabled"); 587 CMSConcurrentMTEnabled = false; 588 } 589 } else { 590 CMSConcurrentMTEnabled = false; 591 } 592 } 593 if (!CMSConcurrentMTEnabled) { 594 ParallelCMSThreads = 0; 595 } else { 596 // Turn off CMSCleanOnEnter optimization temporarily for 597 // the MT case where it's not fixed yet; see 6178663. 598 CMSCleanOnEnter = false; 599 } 600 assert((_conc_workers != NULL) == (ParallelCMSThreads > 1), 601 "Inconsistency"); 602 603 // Parallel task queues; these are shared for the 604 // concurrent and stop-world phases of CMS, but 605 // are not shared with parallel scavenge (ParNew). 606 { 607 uint i; 608 uint num_queues = (uint) MAX2(ParallelGCThreads, ParallelCMSThreads); 609 610 if ((CMSParallelRemarkEnabled || CMSConcurrentMTEnabled 611 || ParallelRefProcEnabled) 612 && num_queues > 0) { 613 _task_queues = new OopTaskQueueSet(num_queues); 614 if (_task_queues == NULL) { 615 warning("task_queues allocation failure."); 616 return; 617 } 618 _hash_seed = NEW_C_HEAP_ARRAY(int, num_queues); 619 if (_hash_seed == NULL) { 620 warning("_hash_seed array allocation failure"); 621 return; 622 } 623 624 // XXX use a global constant instead of 64! 625 typedef struct OopTaskQueuePadded { 626 OopTaskQueue work_queue; 627 char pad[64 - sizeof(OopTaskQueue)]; // prevent false sharing 628 } OopTaskQueuePadded; 629 630 for (i = 0; i < num_queues; i++) { 631 OopTaskQueuePadded *q_padded = new OopTaskQueuePadded(); 632 if (q_padded == NULL) { 633 warning("work_queue allocation failure."); 634 return; 635 } 636 _task_queues->register_queue(i, &q_padded->work_queue); 637 } 638 for (i = 0; i < num_queues; i++) { 639 _task_queues->queue(i)->initialize(); 640 _hash_seed[i] = 17; // copied from ParNew 641 } 642 } 643 } 644 645 // "initiatingOccupancy" is the occupancy ratio at which we trigger 646 // a new collection cycle. Unless explicitly specified via 647 // CMSTriggerRatio, it is calculated by: 648 // Let "f" be MinHeapFreeRatio in 649 // 650 // intiatingOccupancy = 100-f + 651 // f * (CMSTriggerRatio/100) 652 // That is, if we assume the heap is at its desired maximum occupancy at the 653 // end of a collection, we let CMSTriggerRatio of the (purported) free 654 // space be allocated before initiating a new collection cycle. 655 if (CMSInitiatingOccupancyFraction > 0) { 656 _initiatingOccupancy = (double)CMSInitiatingOccupancyFraction / 100.0; 657 } else { 658 _initiatingOccupancy = ((100 - MinHeapFreeRatio) + 659 (double)(CMSTriggerRatio * 660 MinHeapFreeRatio) / 100.0) 661 / 100.0; 662 } 663 // Clip CMSBootstrapOccupancy between 0 and 100. 664 _bootstrap_occupancy = ((double)MIN2((intx)100, MAX2((intx)0, CMSBootstrapOccupancy))) 665 /(double)100; 666 667 _full_gcs_since_conc_gc = 0; 668 669 // Now tell CMS generations the identity of their collector 670 ConcurrentMarkSweepGeneration::set_collector(this); 671 672 // Create & start a CMS thread for this CMS collector 673 _cmsThread = ConcurrentMarkSweepThread::start(this); 674 assert(cmsThread() != NULL, "CMS Thread should have been created"); 675 assert(cmsThread()->collector() == this, 676 "CMS Thread should refer to this gen"); 677 assert(CGC_lock != NULL, "Where's the CGC_lock?"); 678 679 // Support for parallelizing young gen rescan 680 GenCollectedHeap* gch = GenCollectedHeap::heap(); 681 _young_gen = gch->prev_gen(_cmsGen); 682 if (gch->supports_inline_contig_alloc()) { 683 _top_addr = gch->top_addr(); 684 _end_addr = gch->end_addr(); 685 assert(_young_gen != NULL, "no _young_gen"); 686 _eden_chunk_index = 0; 687 _eden_chunk_capacity = (_young_gen->max_capacity()+CMSSamplingGrain)/CMSSamplingGrain; 688 _eden_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, _eden_chunk_capacity); 689 if (_eden_chunk_array == NULL) { 690 _eden_chunk_capacity = 0; 691 warning("GC/CMS: _eden_chunk_array allocation failure"); 692 } 693 } 694 assert(_eden_chunk_array != NULL || _eden_chunk_capacity == 0, "Error"); 695 696 // Support for parallelizing survivor space rescan 697 if (CMSParallelRemarkEnabled && CMSParallelSurvivorRemarkEnabled) { 698 size_t max_plab_samples = MaxNewSize/((SurvivorRatio+2)*MinTLABSize); 699 _survivor_plab_array = NEW_C_HEAP_ARRAY(ChunkArray, ParallelGCThreads); 700 _survivor_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, 2*max_plab_samples); 701 _cursor = NEW_C_HEAP_ARRAY(size_t, ParallelGCThreads); 702 if (_survivor_plab_array == NULL || _survivor_chunk_array == NULL 703 || _cursor == NULL) { 704 warning("Failed to allocate survivor plab/chunk array"); 705 if (_survivor_plab_array != NULL) { 706 FREE_C_HEAP_ARRAY(ChunkArray, _survivor_plab_array); 707 _survivor_plab_array = NULL; 708 } 709 if (_survivor_chunk_array != NULL) { 710 FREE_C_HEAP_ARRAY(HeapWord*, _survivor_chunk_array); 711 _survivor_chunk_array = NULL; 712 } 713 if (_cursor != NULL) { 714 FREE_C_HEAP_ARRAY(size_t, _cursor); 715 _cursor = NULL; 716 } 717 } else { 718 _survivor_chunk_capacity = 2*max_plab_samples; 719 for (uint i = 0; i < ParallelGCThreads; i++) { 720 HeapWord** vec = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples); 721 if (vec == NULL) { 722 warning("Failed to allocate survivor plab array"); 723 for (int j = i; j > 0; j--) { 724 FREE_C_HEAP_ARRAY(HeapWord*, _survivor_plab_array[j-1].array()); 725 } 726 FREE_C_HEAP_ARRAY(ChunkArray, _survivor_plab_array); 727 FREE_C_HEAP_ARRAY(HeapWord*, _survivor_chunk_array); 728 _survivor_plab_array = NULL; 729 _survivor_chunk_array = NULL; 730 _survivor_chunk_capacity = 0; 731 break; 732 } else { 733 ChunkArray* cur = 734 ::new (&_survivor_plab_array[i]) ChunkArray(vec, 735 max_plab_samples); 736 assert(cur->end() == 0, "Should be 0"); 737 assert(cur->array() == vec, "Should be vec"); 738 assert(cur->capacity() == max_plab_samples, "Error"); 739 } 740 } 741 } 742 } 743 assert( ( _survivor_plab_array != NULL 744 && _survivor_chunk_array != NULL) 745 || ( _survivor_chunk_capacity == 0 746 && _survivor_chunk_index == 0), 747 "Error"); 748 749 // Choose what strong roots should be scanned depending on verification options 750 // and perm gen collection mode. 751 if (!CMSClassUnloadingEnabled) { 752 // If class unloading is disabled we want to include all classes into the root set. 753 add_root_scanning_option(SharedHeap::SO_AllClasses); 754 } else { 755 add_root_scanning_option(SharedHeap::SO_SystemClasses); 756 } 757 758 NOT_PRODUCT(_overflow_counter = CMSMarkStackOverflowInterval;) 759 _gc_counters = new CollectorCounters("CMS", 1); 760 _completed_initialization = true; 761 _sweep_timer.start(); // start of time 762 } 763 764 const char* ConcurrentMarkSweepGeneration::name() const { 765 return "concurrent mark-sweep generation"; 766 } 767 void ConcurrentMarkSweepGeneration::update_counters() { 768 if (UsePerfData) { 769 _space_counters->update_all(); 770 _gen_counters->update_all(); 771 } 772 } 773 774 // this is an optimized version of update_counters(). it takes the 775 // used value as a parameter rather than computing it. 776 // 777 void ConcurrentMarkSweepGeneration::update_counters(size_t used) { 778 if (UsePerfData) { 779 _space_counters->update_used(used); 780 _space_counters->update_capacity(); 781 _gen_counters->update_all(); 782 } 783 } 784 785 void ConcurrentMarkSweepGeneration::print() const { 786 Generation::print(); 787 cmsSpace()->print(); 788 } 789 790 #ifndef PRODUCT 791 void ConcurrentMarkSweepGeneration::print_statistics() { 792 cmsSpace()->printFLCensus(0); 793 } 794 #endif 795 796 void ConcurrentMarkSweepGeneration::printOccupancy(const char *s) { 797 GenCollectedHeap* gch = GenCollectedHeap::heap(); 798 if (PrintGCDetails) { 799 if (Verbose) { 800 gclog_or_tty->print(" [%d %s-%s: "SIZE_FORMAT"("SIZE_FORMAT")]", 801 level(), short_name(), s, used(), capacity()); 802 } else { 803 gclog_or_tty->print(" [%d %s-%s: "SIZE_FORMAT"K("SIZE_FORMAT"K)]", 804 level(), short_name(), s, used() / K, capacity() / K); 805 } 806 } 807 if (Verbose) { 808 gclog_or_tty->print(" "SIZE_FORMAT"("SIZE_FORMAT")", 809 gch->used(), gch->capacity()); 810 } else { 811 gclog_or_tty->print(" "SIZE_FORMAT"K("SIZE_FORMAT"K)", 812 gch->used() / K, gch->capacity() / K); 813 } 814 } 815 816 size_t 817 ConcurrentMarkSweepGeneration::contiguous_available() const { 818 // dld proposes an improvement in precision here. If the committed 819 // part of the space ends in a free block we should add that to 820 // uncommitted size in the calculation below. Will make this 821 // change later, staying with the approximation below for the 822 // time being. -- ysr. 823 return MAX2(_virtual_space.uncommitted_size(), unsafe_max_alloc_nogc()); 824 } 825 826 size_t 827 ConcurrentMarkSweepGeneration::unsafe_max_alloc_nogc() const { 828 return _cmsSpace->max_alloc_in_words() * HeapWordSize; 829 } 830 831 size_t ConcurrentMarkSweepGeneration::max_available() const { 832 return free() + _virtual_space.uncommitted_size(); 833 } 834 835 bool ConcurrentMarkSweepGeneration::promotion_attempt_is_safe( 836 size_t max_promotion_in_bytes, 837 bool younger_handles_promotion_failure) const { 838 839 // This is the most conservative test. Full promotion is 840 // guaranteed if this is used. The multiplicative factor is to 841 // account for the worst case "dilatation". 842 double adjusted_max_promo_bytes = _dilatation_factor * max_promotion_in_bytes; 843 if (adjusted_max_promo_bytes > (double)max_uintx) { // larger than size_t 844 adjusted_max_promo_bytes = (double)max_uintx; 845 } 846 bool result = (max_contiguous_available() >= (size_t)adjusted_max_promo_bytes); 847 848 if (younger_handles_promotion_failure && !result) { 849 // Full promotion is not guaranteed because fragmentation 850 // of the cms generation can prevent the full promotion. 851 result = (max_available() >= (size_t)adjusted_max_promo_bytes); 852 853 if (!result) { 854 // With promotion failure handling the test for the ability 855 // to support the promotion does not have to be guaranteed. 856 // Use an average of the amount promoted. 857 result = max_available() >= (size_t) 858 gc_stats()->avg_promoted()->padded_average(); 859 if (PrintGC && Verbose && result) { 860 gclog_or_tty->print_cr( 861 "\nConcurrentMarkSweepGeneration::promotion_attempt_is_safe" 862 " max_available: " SIZE_FORMAT 863 " avg_promoted: " SIZE_FORMAT, 864 max_available(), (size_t) 865 gc_stats()->avg_promoted()->padded_average()); 866 } 867 } else { 868 if (PrintGC && Verbose) { 869 gclog_or_tty->print_cr( 870 "\nConcurrentMarkSweepGeneration::promotion_attempt_is_safe" 871 " max_available: " SIZE_FORMAT 872 " adj_max_promo_bytes: " SIZE_FORMAT, 873 max_available(), (size_t)adjusted_max_promo_bytes); 874 } 875 } 876 } else { 877 if (PrintGC && Verbose) { 878 gclog_or_tty->print_cr( 879 "\nConcurrentMarkSweepGeneration::promotion_attempt_is_safe" 880 " contiguous_available: " SIZE_FORMAT 881 " adj_max_promo_bytes: " SIZE_FORMAT, 882 max_contiguous_available(), (size_t)adjusted_max_promo_bytes); 883 } 884 } 885 return result; 886 } 887 888 CompactibleSpace* 889 ConcurrentMarkSweepGeneration::first_compaction_space() const { 890 return _cmsSpace; 891 } 892 893 void ConcurrentMarkSweepGeneration::reset_after_compaction() { 894 // Clear the promotion information. These pointers can be adjusted 895 // along with all the other pointers into the heap but 896 // compaction is expected to be a rare event with 897 // a heap using cms so don't do it without seeing the need. 898 if (ParallelGCThreads > 0) { 899 for (uint i = 0; i < ParallelGCThreads; i++) { 900 _par_gc_thread_states[i]->promo.reset(); 901 } 902 } 903 } 904 905 void ConcurrentMarkSweepGeneration::space_iterate(SpaceClosure* blk, bool usedOnly) { 906 blk->do_space(_cmsSpace); 907 } 908 909 void ConcurrentMarkSweepGeneration::compute_new_size() { 910 assert_locked_or_safepoint(Heap_lock); 911 912 // If incremental collection failed, we just want to expand 913 // to the limit. 914 if (incremental_collection_failed()) { 915 clear_incremental_collection_failed(); 916 grow_to_reserved(); 917 return; 918 } 919 920 size_t expand_bytes = 0; 921 double free_percentage = ((double) free()) / capacity(); 922 double desired_free_percentage = (double) MinHeapFreeRatio / 100; 923 double maximum_free_percentage = (double) MaxHeapFreeRatio / 100; 924 925 // compute expansion delta needed for reaching desired free percentage 926 if (free_percentage < desired_free_percentage) { 927 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 928 assert(desired_capacity >= capacity(), "invalid expansion size"); 929 expand_bytes = MAX2(desired_capacity - capacity(), MinHeapDeltaBytes); 930 } 931 if (expand_bytes > 0) { 932 if (PrintGCDetails && Verbose) { 933 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 934 gclog_or_tty->print_cr("\nFrom compute_new_size: "); 935 gclog_or_tty->print_cr(" Free fraction %f", free_percentage); 936 gclog_or_tty->print_cr(" Desired free fraction %f", 937 desired_free_percentage); 938 gclog_or_tty->print_cr(" Maximum free fraction %f", 939 maximum_free_percentage); 940 gclog_or_tty->print_cr(" Capactiy "SIZE_FORMAT, capacity()/1000); 941 gclog_or_tty->print_cr(" Desired capacity "SIZE_FORMAT, 942 desired_capacity/1000); 943 int prev_level = level() - 1; 944 if (prev_level >= 0) { 945 size_t prev_size = 0; 946 GenCollectedHeap* gch = GenCollectedHeap::heap(); 947 Generation* prev_gen = gch->_gens[prev_level]; 948 prev_size = prev_gen->capacity(); 949 gclog_or_tty->print_cr(" Younger gen size "SIZE_FORMAT, 950 prev_size/1000); 951 } 952 gclog_or_tty->print_cr(" unsafe_max_alloc_nogc "SIZE_FORMAT, 953 unsafe_max_alloc_nogc()/1000); 954 gclog_or_tty->print_cr(" contiguous available "SIZE_FORMAT, 955 contiguous_available()/1000); 956 gclog_or_tty->print_cr(" Expand by "SIZE_FORMAT" (bytes)", 957 expand_bytes); 958 } 959 // safe if expansion fails 960 expand(expand_bytes, 0, CMSExpansionCause::_satisfy_free_ratio); 961 if (PrintGCDetails && Verbose) { 962 gclog_or_tty->print_cr(" Expanded free fraction %f", 963 ((double) free()) / capacity()); 964 } 965 } 966 } 967 968 Mutex* ConcurrentMarkSweepGeneration::freelistLock() const { 969 return cmsSpace()->freelistLock(); 970 } 971 972 HeapWord* ConcurrentMarkSweepGeneration::allocate(size_t size, 973 bool tlab) { 974 CMSSynchronousYieldRequest yr; 975 MutexLockerEx x(freelistLock(), 976 Mutex::_no_safepoint_check_flag); 977 return have_lock_and_allocate(size, tlab); 978 } 979 980 HeapWord* ConcurrentMarkSweepGeneration::have_lock_and_allocate(size_t size, 981 bool tlab) { 982 assert_lock_strong(freelistLock()); 983 size_t adjustedSize = CompactibleFreeListSpace::adjustObjectSize(size); 984 HeapWord* res = cmsSpace()->allocate(adjustedSize); 985 // Allocate the object live (grey) if the background collector has 986 // started marking. This is necessary because the marker may 987 // have passed this address and consequently this object will 988 // not otherwise be greyed and would be incorrectly swept up. 989 // Note that if this object contains references, the writing 990 // of those references will dirty the card containing this object 991 // allowing the object to be blackened (and its references scanned) 992 // either during a preclean phase or at the final checkpoint. 993 if (res != NULL) { 994 collector()->direct_allocated(res, adjustedSize); 995 _direct_allocated_words += adjustedSize; 996 // allocation counters 997 NOT_PRODUCT( 998 _numObjectsAllocated++; 999 _numWordsAllocated += (int)adjustedSize; 1000 ) 1001 } 1002 return res; 1003 } 1004 1005 // In the case of direct allocation by mutators in a generation that 1006 // is being concurrently collected, the object must be allocated 1007 // live (grey) if the background collector has started marking. 1008 // This is necessary because the marker may 1009 // have passed this address and consequently this object will 1010 // not otherwise be greyed and would be incorrectly swept up. 1011 // Note that if this object contains references, the writing 1012 // of those references will dirty the card containing this object 1013 // allowing the object to be blackened (and its references scanned) 1014 // either during a preclean phase or at the final checkpoint. 1015 void CMSCollector::direct_allocated(HeapWord* start, size_t size) { 1016 assert(_markBitMap.covers(start, size), "Out of bounds"); 1017 if (_collectorState >= Marking) { 1018 MutexLockerEx y(_markBitMap.lock(), 1019 Mutex::_no_safepoint_check_flag); 1020 // [see comments preceding SweepClosure::do_blk() below for details] 1021 // 1. need to mark the object as live so it isn't collected 1022 // 2. need to mark the 2nd bit to indicate the object may be uninitialized 1023 // 3. need to mark the end of the object so sweeper can skip over it 1024 // if it's uninitialized when the sweeper reaches it. 1025 _markBitMap.mark(start); // object is live 1026 _markBitMap.mark(start + 1); // object is potentially uninitialized? 1027 _markBitMap.mark(start + size - 1); 1028 // mark end of object 1029 } 1030 // check that oop looks uninitialized 1031 assert(oop(start)->klass() == NULL, "_klass should be NULL"); 1032 } 1033 1034 void CMSCollector::promoted(bool par, HeapWord* start, 1035 bool is_obj_array, size_t obj_size) { 1036 assert(_markBitMap.covers(start), "Out of bounds"); 1037 // See comment in direct_allocated() about when objects should 1038 // be allocated live. 1039 if (_collectorState >= Marking) { 1040 // we already hold the marking bit map lock, taken in 1041 // the prologue 1042 if (par) { 1043 _markBitMap.par_mark(start); 1044 } else { 1045 _markBitMap.mark(start); 1046 } 1047 // We don't need to mark the object as uninitialized (as 1048 // in direct_allocated above) because this is being done with the 1049 // world stopped and the object will be initialized by the 1050 // time the sweeper gets to look at it. 1051 assert(SafepointSynchronize::is_at_safepoint(), 1052 "expect promotion only at safepoints"); 1053 1054 if (_collectorState < Sweeping) { 1055 // Mark the appropriate cards in the modUnionTable, so that 1056 // this object gets scanned before the sweep. If this is 1057 // not done, CMS generation references in the object might 1058 // not get marked. 1059 // For the case of arrays, which are otherwise precisely 1060 // marked, we need to dirty the entire array, not just its head. 1061 if (is_obj_array) { 1062 // The [par_]mark_range() method expects mr.end() below to 1063 // be aligned to the granularity of a bit's representation 1064 // in the heap. In the case of the MUT below, that's a 1065 // card size. 1066 MemRegion mr(start, 1067 (HeapWord*)round_to((intptr_t)(start + obj_size), 1068 CardTableModRefBS::card_size /* bytes */)); 1069 if (par) { 1070 _modUnionTable.par_mark_range(mr); 1071 } else { 1072 _modUnionTable.mark_range(mr); 1073 } 1074 } else { // not an obj array; we can just mark the head 1075 if (par) { 1076 _modUnionTable.par_mark(start); 1077 } else { 1078 _modUnionTable.mark(start); 1079 } 1080 } 1081 } 1082 } 1083 } 1084 1085 static inline size_t percent_of_space(Space* space, HeapWord* addr) 1086 { 1087 size_t delta = pointer_delta(addr, space->bottom()); 1088 return (size_t)(delta * 100.0 / (space->capacity() / HeapWordSize)); 1089 } 1090 1091 void CMSCollector::icms_update_allocation_limits() 1092 { 1093 Generation* gen0 = GenCollectedHeap::heap()->get_gen(0); 1094 EdenSpace* eden = gen0->as_DefNewGeneration()->eden(); 1095 1096 const unsigned int duty_cycle = stats().icms_update_duty_cycle(); 1097 if (CMSTraceIncrementalPacing) { 1098 stats().print(); 1099 } 1100 1101 assert(duty_cycle <= 100, "invalid duty cycle"); 1102 if (duty_cycle != 0) { 1103 // The duty_cycle is a percentage between 0 and 100; convert to words and 1104 // then compute the offset from the endpoints of the space. 1105 size_t free_words = eden->free() / HeapWordSize; 1106 double free_words_dbl = (double)free_words; 1107 size_t duty_cycle_words = (size_t)(free_words_dbl * duty_cycle / 100.0); 1108 size_t offset_words = (free_words - duty_cycle_words) / 2; 1109 1110 _icms_start_limit = eden->top() + offset_words; 1111 _icms_stop_limit = eden->end() - offset_words; 1112 1113 // The limits may be adjusted (shifted to the right) by 1114 // CMSIncrementalOffset, to allow the application more mutator time after a 1115 // young gen gc (when all mutators were stopped) and before CMS starts and 1116 // takes away one or more cpus. 1117 if (CMSIncrementalOffset != 0) { 1118 double adjustment_dbl = free_words_dbl * CMSIncrementalOffset / 100.0; 1119 size_t adjustment = (size_t)adjustment_dbl; 1120 HeapWord* tmp_stop = _icms_stop_limit + adjustment; 1121 if (tmp_stop > _icms_stop_limit && tmp_stop < eden->end()) { 1122 _icms_start_limit += adjustment; 1123 _icms_stop_limit = tmp_stop; 1124 } 1125 } 1126 } 1127 if (duty_cycle == 0 || (_icms_start_limit == _icms_stop_limit)) { 1128 _icms_start_limit = _icms_stop_limit = eden->end(); 1129 } 1130 1131 // Install the new start limit. 1132 eden->set_soft_end(_icms_start_limit); 1133 1134 if (CMSTraceIncrementalMode) { 1135 gclog_or_tty->print(" icms alloc limits: " 1136 PTR_FORMAT "," PTR_FORMAT 1137 " (" SIZE_FORMAT "%%," SIZE_FORMAT "%%) ", 1138 _icms_start_limit, _icms_stop_limit, 1139 percent_of_space(eden, _icms_start_limit), 1140 percent_of_space(eden, _icms_stop_limit)); 1141 if (Verbose) { 1142 gclog_or_tty->print("eden: "); 1143 eden->print_on(gclog_or_tty); 1144 } 1145 } 1146 } 1147 1148 // Any changes here should try to maintain the invariant 1149 // that if this method is called with _icms_start_limit 1150 // and _icms_stop_limit both NULL, then it should return NULL 1151 // and not notify the icms thread. 1152 HeapWord* 1153 CMSCollector::allocation_limit_reached(Space* space, HeapWord* top, 1154 size_t word_size) 1155 { 1156 // A start_limit equal to end() means the duty cycle is 0, so treat that as a 1157 // nop. 1158 if (CMSIncrementalMode && _icms_start_limit != space->end()) { 1159 if (top <= _icms_start_limit) { 1160 if (CMSTraceIncrementalMode) { 1161 space->print_on(gclog_or_tty); 1162 gclog_or_tty->stamp(); 1163 gclog_or_tty->print_cr(" start limit top=" PTR_FORMAT 1164 ", new limit=" PTR_FORMAT 1165 " (" SIZE_FORMAT "%%)", 1166 top, _icms_stop_limit, 1167 percent_of_space(space, _icms_stop_limit)); 1168 } 1169 ConcurrentMarkSweepThread::start_icms(); 1170 assert(top < _icms_stop_limit, "Tautology"); 1171 if (word_size < pointer_delta(_icms_stop_limit, top)) { 1172 return _icms_stop_limit; 1173 } 1174 1175 // The allocation will cross both the _start and _stop limits, so do the 1176 // stop notification also and return end(). 1177 if (CMSTraceIncrementalMode) { 1178 space->print_on(gclog_or_tty); 1179 gclog_or_tty->stamp(); 1180 gclog_or_tty->print_cr(" +stop limit top=" PTR_FORMAT 1181 ", new limit=" PTR_FORMAT 1182 " (" SIZE_FORMAT "%%)", 1183 top, space->end(), 1184 percent_of_space(space, space->end())); 1185 } 1186 ConcurrentMarkSweepThread::stop_icms(); 1187 return space->end(); 1188 } 1189 1190 if (top <= _icms_stop_limit) { 1191 if (CMSTraceIncrementalMode) { 1192 space->print_on(gclog_or_tty); 1193 gclog_or_tty->stamp(); 1194 gclog_or_tty->print_cr(" stop limit top=" PTR_FORMAT 1195 ", new limit=" PTR_FORMAT 1196 " (" SIZE_FORMAT "%%)", 1197 top, space->end(), 1198 percent_of_space(space, space->end())); 1199 } 1200 ConcurrentMarkSweepThread::stop_icms(); 1201 return space->end(); 1202 } 1203 1204 if (CMSTraceIncrementalMode) { 1205 space->print_on(gclog_or_tty); 1206 gclog_or_tty->stamp(); 1207 gclog_or_tty->print_cr(" end limit top=" PTR_FORMAT 1208 ", new limit=" PTR_FORMAT, 1209 top, NULL); 1210 } 1211 } 1212 1213 return NULL; 1214 } 1215 1216 oop ConcurrentMarkSweepGeneration::promote(oop obj, size_t obj_size, oop* ref) { 1217 assert(obj_size == (size_t)obj->size(), "bad obj_size passed in"); 1218 // allocate, copy and if necessary update promoinfo -- 1219 // delegate to underlying space. 1220 assert_lock_strong(freelistLock()); 1221 1222 #ifndef PRODUCT 1223 if (Universe::heap()->promotion_should_fail()) { 1224 return NULL; 1225 } 1226 #endif // #ifndef PRODUCT 1227 1228 oop res = _cmsSpace->promote(obj, obj_size, ref); 1229 if (res == NULL) { 1230 // expand and retry 1231 size_t s = _cmsSpace->expansionSpaceRequired(obj_size); // HeapWords 1232 expand(s*HeapWordSize, MinHeapDeltaBytes, 1233 CMSExpansionCause::_satisfy_promotion); 1234 // Since there's currently no next generation, we don't try to promote 1235 // into a more senior generation. 1236 assert(next_gen() == NULL, "assumption, based upon which no attempt " 1237 "is made to pass on a possibly failing " 1238 "promotion to next generation"); 1239 res = _cmsSpace->promote(obj, obj_size, ref); 1240 } 1241 if (res != NULL) { 1242 // See comment in allocate() about when objects should 1243 // be allocated live. 1244 assert(obj->is_oop(), "Will dereference klass pointer below"); 1245 collector()->promoted(false, // Not parallel 1246 (HeapWord*)res, obj->is_objArray(), obj_size); 1247 // promotion counters 1248 NOT_PRODUCT( 1249 _numObjectsPromoted++; 1250 _numWordsPromoted += 1251 (int)(CompactibleFreeListSpace::adjustObjectSize(obj->size())); 1252 ) 1253 } 1254 return res; 1255 } 1256 1257 1258 HeapWord* 1259 ConcurrentMarkSweepGeneration::allocation_limit_reached(Space* space, 1260 HeapWord* top, 1261 size_t word_sz) 1262 { 1263 return collector()->allocation_limit_reached(space, top, word_sz); 1264 } 1265 1266 // Things to support parallel young-gen collection. 1267 oop 1268 ConcurrentMarkSweepGeneration::par_promote(int thread_num, 1269 oop old, markOop m, 1270 size_t word_sz) { 1271 #ifndef PRODUCT 1272 if (Universe::heap()->promotion_should_fail()) { 1273 return NULL; 1274 } 1275 #endif // #ifndef PRODUCT 1276 1277 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1278 PromotionInfo* promoInfo = &ps->promo; 1279 // if we are tracking promotions, then first ensure space for 1280 // promotion (including spooling space for saving header if necessary). 1281 // then allocate and copy, then track promoted info if needed. 1282 // When tracking (see PromotionInfo::track()), the mark word may 1283 // be displaced and in this case restoration of the mark word 1284 // occurs in the (oop_since_save_marks_)iterate phase. 1285 if (promoInfo->tracking() && !promoInfo->ensure_spooling_space()) { 1286 // Out of space for allocating spooling buffers; 1287 // try expanding and allocating spooling buffers. 1288 if (!expand_and_ensure_spooling_space(promoInfo)) { 1289 return NULL; 1290 } 1291 } 1292 assert(promoInfo->has_spooling_space(), "Control point invariant"); 1293 HeapWord* obj_ptr = ps->lab.alloc(word_sz); 1294 if (obj_ptr == NULL) { 1295 obj_ptr = expand_and_par_lab_allocate(ps, word_sz); 1296 if (obj_ptr == NULL) { 1297 return NULL; 1298 } 1299 } 1300 oop obj = oop(obj_ptr); 1301 assert(obj->klass() == NULL, "Object should be uninitialized here."); 1302 // Otherwise, copy the object. Here we must be careful to insert the 1303 // klass pointer last, since this marks the block as an allocated object. 1304 HeapWord* old_ptr = (HeapWord*)old; 1305 if (word_sz > (size_t)oopDesc::header_size()) { 1306 Copy::aligned_disjoint_words(old_ptr + oopDesc::header_size(), 1307 obj_ptr + oopDesc::header_size(), 1308 word_sz - oopDesc::header_size()); 1309 } 1310 // Restore the mark word copied above. 1311 obj->set_mark(m); 1312 // Now we can track the promoted object, if necessary. We take care 1313 // To delay the transition from uninitialized to full object 1314 // (i.e., insertion of klass pointer) until after, so that it 1315 // atomically becomes a promoted object. 1316 if (promoInfo->tracking()) { 1317 promoInfo->track((PromotedObject*)obj, old->klass()); 1318 } 1319 // Finally, install the klass pointer. 1320 obj->set_klass(old->klass()); 1321 1322 assert(old->is_oop(), "Will dereference klass ptr below"); 1323 collector()->promoted(true, // parallel 1324 obj_ptr, old->is_objArray(), word_sz); 1325 1326 NOT_PRODUCT( 1327 Atomic::inc(&_numObjectsPromoted); 1328 Atomic::add((jint)CompactibleFreeListSpace::adjustObjectSize(obj->size()), 1329 &_numWordsPromoted); 1330 ) 1331 1332 return obj; 1333 } 1334 1335 void 1336 ConcurrentMarkSweepGeneration:: 1337 par_promote_alloc_undo(int thread_num, 1338 HeapWord* obj, size_t word_sz) { 1339 // CMS does not support promotion undo. 1340 ShouldNotReachHere(); 1341 } 1342 1343 void 1344 ConcurrentMarkSweepGeneration:: 1345 par_promote_alloc_done(int thread_num) { 1346 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1347 ps->lab.retire(); 1348 #if CFLS_LAB_REFILL_STATS 1349 if (thread_num == 0) { 1350 _cmsSpace->print_par_alloc_stats(); 1351 } 1352 #endif 1353 } 1354 1355 void 1356 ConcurrentMarkSweepGeneration:: 1357 par_oop_since_save_marks_iterate_done(int thread_num) { 1358 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1359 ParScanWithoutBarrierClosure* dummy_cl = NULL; 1360 ps->promo.promoted_oops_iterate_nv(dummy_cl); 1361 } 1362 1363 // XXXPERM 1364 bool ConcurrentMarkSweepGeneration::should_collect(bool full, 1365 size_t size, 1366 bool tlab) 1367 { 1368 // We allow a STW collection only if a full 1369 // collection was requested. 1370 return full || should_allocate(size, tlab); // FIX ME !!! 1371 // This and promotion failure handling are connected at the 1372 // hip and should be fixed by untying them. 1373 } 1374 1375 bool CMSCollector::shouldConcurrentCollect() { 1376 if (_full_gc_requested) { 1377 assert(ExplicitGCInvokesConcurrent, "Unexpected state"); 1378 if (Verbose && PrintGCDetails) { 1379 gclog_or_tty->print_cr("CMSCollector: collect because of explicit " 1380 " gc request"); 1381 } 1382 return true; 1383 } 1384 1385 // For debugging purposes, change the type of collection. 1386 // If the rotation is not on the concurrent collection 1387 // type, don't start a concurrent collection. 1388 NOT_PRODUCT( 1389 if (RotateCMSCollectionTypes && 1390 (_cmsGen->debug_collection_type() != 1391 ConcurrentMarkSweepGeneration::Concurrent_collection_type)) { 1392 assert(_cmsGen->debug_collection_type() != 1393 ConcurrentMarkSweepGeneration::Unknown_collection_type, 1394 "Bad cms collection type"); 1395 return false; 1396 } 1397 ) 1398 1399 FreelistLocker x(this); 1400 // ------------------------------------------------------------------ 1401 // Print out lots of information which affects the initiation of 1402 // a collection. 1403 if (PrintCMSInitiationStatistics && stats().valid()) { 1404 gclog_or_tty->print("CMSCollector shouldConcurrentCollect: "); 1405 gclog_or_tty->stamp(); 1406 gclog_or_tty->print_cr(""); 1407 stats().print_on(gclog_or_tty); 1408 gclog_or_tty->print_cr("time_until_cms_gen_full %3.7f", 1409 stats().time_until_cms_gen_full()); 1410 gclog_or_tty->print_cr("free="SIZE_FORMAT, _cmsGen->free()); 1411 gclog_or_tty->print_cr("contiguous_available="SIZE_FORMAT, 1412 _cmsGen->contiguous_available()); 1413 gclog_or_tty->print_cr("promotion_rate=%g", stats().promotion_rate()); 1414 gclog_or_tty->print_cr("cms_allocation_rate=%g", stats().cms_allocation_rate()); 1415 gclog_or_tty->print_cr("occupancy=%3.7f", _cmsGen->occupancy()); 1416 gclog_or_tty->print_cr("initiatingOccupancy=%3.7f", initiatingOccupancy()); 1417 } 1418 // ------------------------------------------------------------------ 1419 1420 // If the estimated time to complete a cms collection (cms_duration()) 1421 // is less than the estimated time remaining until the cms generation 1422 // is full, start a collection. 1423 if (!UseCMSInitiatingOccupancyOnly) { 1424 if (stats().valid()) { 1425 if (stats().time_until_cms_start() == 0.0) { 1426 return true; 1427 } 1428 } else { 1429 // We want to conservatively collect somewhat early in order 1430 // to try and "bootstrap" our CMS/promotion statistics; 1431 // this branch will not fire after the first successful CMS 1432 // collection because the stats should then be valid. 1433 if (_cmsGen->occupancy() >= _bootstrap_occupancy) { 1434 if (Verbose && PrintGCDetails) { 1435 gclog_or_tty->print_cr( 1436 " CMSCollector: collect for bootstrapping statistics:" 1437 " occupancy = %f, boot occupancy = %f", _cmsGen->occupancy(), 1438 _bootstrap_occupancy); 1439 } 1440 return true; 1441 } 1442 } 1443 } 1444 1445 // Otherwise, we start a collection cycle if either the perm gen or 1446 // old gen want a collection cycle started. Each may use 1447 // an appropriate criterion for making this decision. 1448 // XXX We need to make sure that the gen expansion 1449 // criterion dovetails well with this. 1450 if (_cmsGen->shouldConcurrentCollect(initiatingOccupancy())) { 1451 if (Verbose && PrintGCDetails) { 1452 gclog_or_tty->print_cr("CMS old gen initiated"); 1453 } 1454 return true; 1455 } 1456 1457 if (cms_should_unload_classes() && 1458 _permGen->shouldConcurrentCollect(initiatingOccupancy())) { 1459 if (Verbose && PrintGCDetails) { 1460 gclog_or_tty->print_cr("CMS perm gen initiated"); 1461 } 1462 return true; 1463 } 1464 1465 return false; 1466 } 1467 1468 // Clear _expansion_cause fields of constituent generations 1469 void CMSCollector::clear_expansion_cause() { 1470 _cmsGen->clear_expansion_cause(); 1471 _permGen->clear_expansion_cause(); 1472 } 1473 1474 bool ConcurrentMarkSweepGeneration::shouldConcurrentCollect( 1475 double initiatingOccupancy) { 1476 // We should be conservative in starting a collection cycle. To 1477 // start too eagerly runs the risk of collecting too often in the 1478 // extreme. To collect too rarely falls back on full collections, 1479 // which works, even if not optimum in terms of concurrent work. 1480 // As a work around for too eagerly collecting, use the flag 1481 // UseCMSInitiatingOccupancyOnly. This also has the advantage of 1482 // giving the user an easily understandable way of controlling the 1483 // collections. 1484 // We want to start a new collection cycle if any of the following 1485 // conditions hold: 1486 // . our current occupancy exceeds the initiating occupancy, or 1487 // . we recently needed to expand and have not since that expansion, 1488 // collected, or 1489 // . we are not using adaptive free lists and linear allocation is 1490 // going to fail, or 1491 // . (for old gen) incremental collection has already failed or 1492 // may soon fail in the near future as we may not be able to absorb 1493 // promotions. 1494 assert_lock_strong(freelistLock()); 1495 1496 if (occupancy() > initiatingOccupancy) { 1497 if (PrintGCDetails && Verbose) { 1498 gclog_or_tty->print(" %s: collect because of occupancy %f / %f ", 1499 short_name(), occupancy(), initiatingOccupancy); 1500 } 1501 return true; 1502 } 1503 if (UseCMSInitiatingOccupancyOnly) { 1504 return false; 1505 } 1506 if (expansion_cause() == CMSExpansionCause::_satisfy_allocation) { 1507 if (PrintGCDetails && Verbose) { 1508 gclog_or_tty->print(" %s: collect because expanded for allocation ", 1509 short_name()); 1510 } 1511 return true; 1512 } 1513 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1514 assert(gch->collector_policy()->is_two_generation_policy(), 1515 "You may want to check the correctness of the following"); 1516 if (gch->incremental_collection_will_fail()) { 1517 if (PrintGCDetails && Verbose) { 1518 gclog_or_tty->print(" %s: collect because incremental collection will fail ", 1519 short_name()); 1520 } 1521 return true; 1522 } 1523 if (!_cmsSpace->adaptive_freelists() && 1524 _cmsSpace->linearAllocationWouldFail()) { 1525 if (PrintGCDetails && Verbose) { 1526 gclog_or_tty->print(" %s: collect because of linAB ", 1527 short_name()); 1528 } 1529 return true; 1530 } 1531 return false; 1532 } 1533 1534 void ConcurrentMarkSweepGeneration::collect(bool full, 1535 bool clear_all_soft_refs, 1536 size_t size, 1537 bool tlab) 1538 { 1539 collector()->collect(full, clear_all_soft_refs, size, tlab); 1540 } 1541 1542 void CMSCollector::collect(bool full, 1543 bool clear_all_soft_refs, 1544 size_t size, 1545 bool tlab) 1546 { 1547 if (!UseCMSCollectionPassing && _collectorState > Idling) { 1548 // For debugging purposes skip the collection if the state 1549 // is not currently idle 1550 if (TraceCMSState) { 1551 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " skipped full:%d CMS state %d", 1552 Thread::current(), full, _collectorState); 1553 } 1554 return; 1555 } 1556 1557 // The following "if" branch is present for defensive reasons. 1558 // In the current uses of this interface, it can be replaced with: 1559 // assert(!GC_locker.is_active(), "Can't be called otherwise"); 1560 // But I am not placing that assert here to allow future 1561 // generality in invoking this interface. 1562 if (GC_locker::is_active()) { 1563 // A consistency test for GC_locker 1564 assert(GC_locker::needs_gc(), "Should have been set already"); 1565 // Skip this foreground collection, instead 1566 // expanding the heap if necessary. 1567 // Need the free list locks for the call to free() in compute_new_size() 1568 compute_new_size(); 1569 return; 1570 } 1571 acquire_control_and_collect(full, clear_all_soft_refs); 1572 _full_gcs_since_conc_gc++; 1573 1574 } 1575 1576 void CMSCollector::request_full_gc(unsigned int full_gc_count) { 1577 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1578 unsigned int gc_count = gch->total_full_collections(); 1579 if (gc_count == full_gc_count) { 1580 MutexLockerEx y(CGC_lock, Mutex::_no_safepoint_check_flag); 1581 _full_gc_requested = true; 1582 CGC_lock->notify(); // nudge CMS thread 1583 } 1584 } 1585 1586 1587 // The foreground and background collectors need to coordinate in order 1588 // to make sure that they do not mutually interfere with CMS collections. 1589 // When a background collection is active, 1590 // the foreground collector may need to take over (preempt) and 1591 // synchronously complete an ongoing collection. Depending on the 1592 // frequency of the background collections and the heap usage 1593 // of the application, this preemption can be seldom or frequent. 1594 // There are only certain 1595 // points in the background collection that the "collection-baton" 1596 // can be passed to the foreground collector. 1597 // 1598 // The foreground collector will wait for the baton before 1599 // starting any part of the collection. The foreground collector 1600 // will only wait at one location. 1601 // 1602 // The background collector will yield the baton before starting a new 1603 // phase of the collection (e.g., before initial marking, marking from roots, 1604 // precleaning, final re-mark, sweep etc.) This is normally done at the head 1605 // of the loop which switches the phases. The background collector does some 1606 // of the phases (initial mark, final re-mark) with the world stopped. 1607 // Because of locking involved in stopping the world, 1608 // the foreground collector should not block waiting for the background 1609 // collector when it is doing a stop-the-world phase. The background 1610 // collector will yield the baton at an additional point just before 1611 // it enters a stop-the-world phase. Once the world is stopped, the 1612 // background collector checks the phase of the collection. If the 1613 // phase has not changed, it proceeds with the collection. If the 1614 // phase has changed, it skips that phase of the collection. See 1615 // the comments on the use of the Heap_lock in collect_in_background(). 1616 // 1617 // Variable used in baton passing. 1618 // _foregroundGCIsActive - Set to true by the foreground collector when 1619 // it wants the baton. The foreground clears it when it has finished 1620 // the collection. 1621 // _foregroundGCShouldWait - Set to true by the background collector 1622 // when it is running. The foreground collector waits while 1623 // _foregroundGCShouldWait is true. 1624 // CGC_lock - monitor used to protect access to the above variables 1625 // and to notify the foreground and background collectors. 1626 // _collectorState - current state of the CMS collection. 1627 // 1628 // The foreground collector 1629 // acquires the CGC_lock 1630 // sets _foregroundGCIsActive 1631 // waits on the CGC_lock for _foregroundGCShouldWait to be false 1632 // various locks acquired in preparation for the collection 1633 // are released so as not to block the background collector 1634 // that is in the midst of a collection 1635 // proceeds with the collection 1636 // clears _foregroundGCIsActive 1637 // returns 1638 // 1639 // The background collector in a loop iterating on the phases of the 1640 // collection 1641 // acquires the CGC_lock 1642 // sets _foregroundGCShouldWait 1643 // if _foregroundGCIsActive is set 1644 // clears _foregroundGCShouldWait, notifies _CGC_lock 1645 // waits on _CGC_lock for _foregroundGCIsActive to become false 1646 // and exits the loop. 1647 // otherwise 1648 // proceed with that phase of the collection 1649 // if the phase is a stop-the-world phase, 1650 // yield the baton once more just before enqueueing 1651 // the stop-world CMS operation (executed by the VM thread). 1652 // returns after all phases of the collection are done 1653 // 1654 1655 void CMSCollector::acquire_control_and_collect(bool full, 1656 bool clear_all_soft_refs) { 1657 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint"); 1658 assert(!Thread::current()->is_ConcurrentGC_thread(), 1659 "shouldn't try to acquire control from self!"); 1660 1661 // Start the protocol for acquiring control of the 1662 // collection from the background collector (aka CMS thread). 1663 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 1664 "VM thread should have CMS token"); 1665 // Remember the possibly interrupted state of an ongoing 1666 // concurrent collection 1667 CollectorState first_state = _collectorState; 1668 1669 // Signal to a possibly ongoing concurrent collection that 1670 // we want to do a foreground collection. 1671 _foregroundGCIsActive = true; 1672 1673 // Disable incremental mode during a foreground collection. 1674 ICMSDisabler icms_disabler; 1675 1676 // release locks and wait for a notify from the background collector 1677 // releasing the locks in only necessary for phases which 1678 // do yields to improve the granularity of the collection. 1679 assert_lock_strong(bitMapLock()); 1680 // We need to lock the Free list lock for the space that we are 1681 // currently collecting. 1682 assert(haveFreelistLocks(), "Must be holding free list locks"); 1683 bitMapLock()->unlock(); 1684 releaseFreelistLocks(); 1685 { 1686 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1687 if (_foregroundGCShouldWait) { 1688 // We are going to be waiting for action for the CMS thread; 1689 // it had better not be gone (for instance at shutdown)! 1690 assert(ConcurrentMarkSweepThread::cmst() != NULL, 1691 "CMS thread must be running"); 1692 // Wait here until the background collector gives us the go-ahead 1693 ConcurrentMarkSweepThread::clear_CMS_flag( 1694 ConcurrentMarkSweepThread::CMS_vm_has_token); // release token 1695 // Get a possibly blocked CMS thread going: 1696 // Note that we set _foregroundGCIsActive true above, 1697 // without protection of the CGC_lock. 1698 CGC_lock->notify(); 1699 assert(!ConcurrentMarkSweepThread::vm_thread_wants_cms_token(), 1700 "Possible deadlock"); 1701 while (_foregroundGCShouldWait) { 1702 // wait for notification 1703 CGC_lock->wait(Mutex::_no_safepoint_check_flag); 1704 // Possibility of delay/starvation here, since CMS token does 1705 // not know to give priority to VM thread? Actually, i think 1706 // there wouldn't be any delay/starvation, but the proof of 1707 // that "fact" (?) appears non-trivial. XXX 20011219YSR 1708 } 1709 ConcurrentMarkSweepThread::set_CMS_flag( 1710 ConcurrentMarkSweepThread::CMS_vm_has_token); 1711 } 1712 } 1713 // The CMS_token is already held. Get back the other locks. 1714 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 1715 "VM thread should have CMS token"); 1716 getFreelistLocks(); 1717 bitMapLock()->lock_without_safepoint_check(); 1718 if (TraceCMSState) { 1719 gclog_or_tty->print_cr("CMS foreground collector has asked for control " 1720 INTPTR_FORMAT " with first state %d", Thread::current(), first_state); 1721 gclog_or_tty->print_cr(" gets control with state %d", _collectorState); 1722 } 1723 1724 // Check if we need to do a compaction, or if not, whether 1725 // we need to start the mark-sweep from scratch. 1726 bool should_compact = false; 1727 bool should_start_over = false; 1728 decide_foreground_collection_type(clear_all_soft_refs, 1729 &should_compact, &should_start_over); 1730 1731 NOT_PRODUCT( 1732 if (RotateCMSCollectionTypes) { 1733 if (_cmsGen->debug_collection_type() == 1734 ConcurrentMarkSweepGeneration::MSC_foreground_collection_type) { 1735 should_compact = true; 1736 } else if (_cmsGen->debug_collection_type() == 1737 ConcurrentMarkSweepGeneration::MS_foreground_collection_type) { 1738 should_compact = false; 1739 } 1740 } 1741 ) 1742 1743 if (PrintGCDetails && first_state > Idling) { 1744 GCCause::Cause cause = GenCollectedHeap::heap()->gc_cause(); 1745 if (GCCause::is_user_requested_gc(cause) || 1746 GCCause::is_serviceability_requested_gc(cause)) { 1747 gclog_or_tty->print(" (concurrent mode interrupted)"); 1748 } else { 1749 gclog_or_tty->print(" (concurrent mode failure)"); 1750 } 1751 } 1752 1753 if (should_compact) { 1754 // If the collection is being acquired from the background 1755 // collector, there may be references on the discovered 1756 // references lists that have NULL referents (being those 1757 // that were concurrently cleared by a mutator) or 1758 // that are no longer active (having been enqueued concurrently 1759 // by the mutator). 1760 // Scrub the list of those references because Mark-Sweep-Compact 1761 // code assumes referents are not NULL and that all discovered 1762 // Reference objects are active. 1763 ref_processor()->clean_up_discovered_references(); 1764 1765 do_compaction_work(clear_all_soft_refs); 1766 1767 // Has the GC time limit been exceeded? 1768 check_gc_time_limit(); 1769 1770 } else { 1771 do_mark_sweep_work(clear_all_soft_refs, first_state, 1772 should_start_over); 1773 } 1774 // Reset the expansion cause, now that we just completed 1775 // a collection cycle. 1776 clear_expansion_cause(); 1777 _foregroundGCIsActive = false; 1778 return; 1779 } 1780 1781 void CMSCollector::check_gc_time_limit() { 1782 1783 // Ignore explicit GC's. Exiting here does not set the flag and 1784 // does not reset the count. Updating of the averages for system 1785 // GC's is still controlled by UseAdaptiveSizePolicyWithSystemGC. 1786 GCCause::Cause gc_cause = GenCollectedHeap::heap()->gc_cause(); 1787 if (GCCause::is_user_requested_gc(gc_cause) || 1788 GCCause::is_serviceability_requested_gc(gc_cause)) { 1789 return; 1790 } 1791 1792 // Calculate the fraction of the CMS generation was freed during 1793 // the last collection. 1794 // Only consider the STW compacting cost for now. 1795 // 1796 // Note that the gc time limit test only works for the collections 1797 // of the young gen + tenured gen and not for collections of the 1798 // permanent gen. That is because the calculation of the space 1799 // freed by the collection is the free space in the young gen + 1800 // tenured gen. 1801 1802 double fraction_free = 1803 ((double)_cmsGen->free())/((double)_cmsGen->max_capacity()); 1804 if ((100.0 * size_policy()->compacting_gc_cost()) > 1805 ((double) GCTimeLimit) && 1806 ((fraction_free * 100) < GCHeapFreeLimit)) { 1807 size_policy()->inc_gc_time_limit_count(); 1808 if (UseGCOverheadLimit && 1809 (size_policy()->gc_time_limit_count() > 1810 AdaptiveSizePolicyGCTimeLimitThreshold)) { 1811 size_policy()->set_gc_time_limit_exceeded(true); 1812 // Avoid consecutive OOM due to the gc time limit by resetting 1813 // the counter. 1814 size_policy()->reset_gc_time_limit_count(); 1815 if (PrintGCDetails) { 1816 gclog_or_tty->print_cr(" GC is exceeding overhead limit " 1817 "of %d%%", GCTimeLimit); 1818 } 1819 } else { 1820 if (PrintGCDetails) { 1821 gclog_or_tty->print_cr(" GC would exceed overhead limit " 1822 "of %d%%", GCTimeLimit); 1823 } 1824 } 1825 } else { 1826 size_policy()->reset_gc_time_limit_count(); 1827 } 1828 } 1829 1830 // Resize the perm generation and the tenured generation 1831 // after obtaining the free list locks for the 1832 // two generations. 1833 void CMSCollector::compute_new_size() { 1834 assert_locked_or_safepoint(Heap_lock); 1835 FreelistLocker z(this); 1836 _permGen->compute_new_size(); 1837 _cmsGen->compute_new_size(); 1838 } 1839 1840 // A work method used by foreground collection to determine 1841 // what type of collection (compacting or not, continuing or fresh) 1842 // it should do. 1843 // NOTE: the intent is to make UseCMSCompactAtFullCollection 1844 // and CMSCompactWhenClearAllSoftRefs the default in the future 1845 // and do away with the flags after a suitable period. 1846 void CMSCollector::decide_foreground_collection_type( 1847 bool clear_all_soft_refs, bool* should_compact, 1848 bool* should_start_over) { 1849 // Normally, we'll compact only if the UseCMSCompactAtFullCollection 1850 // flag is set, and we have either requested a System.gc() or 1851 // the number of full gc's since the last concurrent cycle 1852 // has exceeded the threshold set by CMSFullGCsBeforeCompaction, 1853 // or if an incremental collection has failed 1854 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1855 assert(gch->collector_policy()->is_two_generation_policy(), 1856 "You may want to check the correctness of the following"); 1857 // Inform cms gen if this was due to partial collection failing. 1858 // The CMS gen may use this fact to determine its expansion policy. 1859 if (gch->incremental_collection_will_fail()) { 1860 assert(!_cmsGen->incremental_collection_failed(), 1861 "Should have been noticed, reacted to and cleared"); 1862 _cmsGen->set_incremental_collection_failed(); 1863 } 1864 *should_compact = 1865 UseCMSCompactAtFullCollection && 1866 ((_full_gcs_since_conc_gc >= CMSFullGCsBeforeCompaction) || 1867 GCCause::is_user_requested_gc(gch->gc_cause()) || 1868 gch->incremental_collection_will_fail()); 1869 *should_start_over = false; 1870 if (clear_all_soft_refs && !*should_compact) { 1871 // We are about to do a last ditch collection attempt 1872 // so it would normally make sense to do a compaction 1873 // to reclaim as much space as possible. 1874 if (CMSCompactWhenClearAllSoftRefs) { 1875 // Default: The rationale is that in this case either 1876 // we are past the final marking phase, in which case 1877 // we'd have to start over, or so little has been done 1878 // that there's little point in saving that work. Compaction 1879 // appears to be the sensible choice in either case. 1880 *should_compact = true; 1881 } else { 1882 // We have been asked to clear all soft refs, but not to 1883 // compact. Make sure that we aren't past the final checkpoint 1884 // phase, for that is where we process soft refs. If we are already 1885 // past that phase, we'll need to redo the refs discovery phase and 1886 // if necessary clear soft refs that weren't previously 1887 // cleared. We do so by remembering the phase in which 1888 // we came in, and if we are past the refs processing 1889 // phase, we'll choose to just redo the mark-sweep 1890 // collection from scratch. 1891 if (_collectorState > FinalMarking) { 1892 // We are past the refs processing phase; 1893 // start over and do a fresh synchronous CMS cycle 1894 _collectorState = Resetting; // skip to reset to start new cycle 1895 reset(false /* == !asynch */); 1896 *should_start_over = true; 1897 } // else we can continue a possibly ongoing current cycle 1898 } 1899 } 1900 } 1901 1902 // A work method used by the foreground collector to do 1903 // a mark-sweep-compact. 1904 void CMSCollector::do_compaction_work(bool clear_all_soft_refs) { 1905 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1906 TraceTime t("CMS:MSC ", PrintGCDetails && Verbose, true, gclog_or_tty); 1907 if (PrintGC && Verbose && !(GCCause::is_user_requested_gc(gch->gc_cause()))) { 1908 gclog_or_tty->print_cr("Compact ConcurrentMarkSweepGeneration after %d " 1909 "collections passed to foreground collector", _full_gcs_since_conc_gc); 1910 } 1911 1912 // Sample collection interval time and reset for collection pause. 1913 if (UseAdaptiveSizePolicy) { 1914 size_policy()->msc_collection_begin(); 1915 } 1916 1917 // Temporarily widen the span of the weak reference processing to 1918 // the entire heap. 1919 MemRegion new_span(GenCollectedHeap::heap()->reserved_region()); 1920 ReferenceProcessorSpanMutator x(ref_processor(), new_span); 1921 1922 // Temporarily, clear the "is_alive_non_header" field of the 1923 // reference processor. 1924 ReferenceProcessorIsAliveMutator y(ref_processor(), NULL); 1925 1926 // Temporarily make reference _processing_ single threaded (non-MT). 1927 ReferenceProcessorMTProcMutator z(ref_processor(), false); 1928 1929 // Temporarily make refs discovery atomic 1930 ReferenceProcessorAtomicMutator w(ref_processor(), true); 1931 1932 ref_processor()->set_enqueuing_is_done(false); 1933 ref_processor()->enable_discovery(); 1934 // If an asynchronous collection finishes, the _modUnionTable is 1935 // all clear. If we are assuming the collection from an asynchronous 1936 // collection, clear the _modUnionTable. 1937 assert(_collectorState != Idling || _modUnionTable.isAllClear(), 1938 "_modUnionTable should be clear if the baton was not passed"); 1939 _modUnionTable.clear_all(); 1940 1941 // We must adjust the allocation statistics being maintained 1942 // in the free list space. We do so by reading and clearing 1943 // the sweep timer and updating the block flux rate estimates below. 1944 assert(_sweep_timer.is_active(), "We should never see the timer inactive"); 1945 _sweep_timer.stop(); 1946 // Note that we do not use this sample to update the _sweep_estimate. 1947 _cmsGen->cmsSpace()->beginSweepFLCensus((float)(_sweep_timer.seconds()), 1948 _sweep_estimate.padded_average()); 1949 1950 GenMarkSweep::invoke_at_safepoint(_cmsGen->level(), 1951 ref_processor(), clear_all_soft_refs); 1952 #ifdef ASSERT 1953 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 1954 size_t free_size = cms_space->free(); 1955 assert(free_size == 1956 pointer_delta(cms_space->end(), cms_space->compaction_top()) 1957 * HeapWordSize, 1958 "All the free space should be compacted into one chunk at top"); 1959 assert(cms_space->dictionary()->totalChunkSize( 1960 debug_only(cms_space->freelistLock())) == 0 || 1961 cms_space->totalSizeInIndexedFreeLists() == 0, 1962 "All the free space should be in a single chunk"); 1963 size_t num = cms_space->totalCount(); 1964 assert((free_size == 0 && num == 0) || 1965 (free_size > 0 && (num == 1 || num == 2)), 1966 "There should be at most 2 free chunks after compaction"); 1967 #endif // ASSERT 1968 _collectorState = Resetting; 1969 assert(_restart_addr == NULL, 1970 "Should have been NULL'd before baton was passed"); 1971 reset(false /* == !asynch */); 1972 _cmsGen->reset_after_compaction(); 1973 1974 if (verifying() && !cms_should_unload_classes()) { 1975 perm_gen_verify_bit_map()->clear_all(); 1976 } 1977 1978 // Clear any data recorded in the PLAB chunk arrays. 1979 if (_survivor_plab_array != NULL) { 1980 reset_survivor_plab_arrays(); 1981 } 1982 1983 // Adjust the per-size allocation stats for the next epoch. 1984 _cmsGen->cmsSpace()->endSweepFLCensus(sweepCount() /* fake */); 1985 // Restart the "sweep timer" for next epoch. 1986 _sweep_timer.reset(); 1987 _sweep_timer.start(); 1988 1989 // Sample collection pause time and reset for collection interval. 1990 if (UseAdaptiveSizePolicy) { 1991 size_policy()->msc_collection_end(gch->gc_cause()); 1992 } 1993 1994 // For a mark-sweep-compact, compute_new_size() will be called 1995 // in the heap's do_collection() method. 1996 } 1997 1998 // A work method used by the foreground collector to do 1999 // a mark-sweep, after taking over from a possibly on-going 2000 // concurrent mark-sweep collection. 2001 void CMSCollector::do_mark_sweep_work(bool clear_all_soft_refs, 2002 CollectorState first_state, bool should_start_over) { 2003 if (PrintGC && Verbose) { 2004 gclog_or_tty->print_cr("Pass concurrent collection to foreground " 2005 "collector with count %d", 2006 _full_gcs_since_conc_gc); 2007 } 2008 switch (_collectorState) { 2009 case Idling: 2010 if (first_state == Idling || should_start_over) { 2011 // The background GC was not active, or should 2012 // restarted from scratch; start the cycle. 2013 _collectorState = InitialMarking; 2014 } 2015 // If first_state was not Idling, then a background GC 2016 // was in progress and has now finished. No need to do it 2017 // again. Leave the state as Idling. 2018 break; 2019 case Precleaning: 2020 // In the foreground case don't do the precleaning since 2021 // it is not done concurrently and there is extra work 2022 // required. 2023 _collectorState = FinalMarking; 2024 } 2025 if (PrintGCDetails && 2026 (_collectorState > Idling || 2027 !GCCause::is_user_requested_gc(GenCollectedHeap::heap()->gc_cause()))) { 2028 gclog_or_tty->print(" (concurrent mode failure)"); 2029 } 2030 collect_in_foreground(clear_all_soft_refs); 2031 2032 // For a mark-sweep, compute_new_size() will be called 2033 // in the heap's do_collection() method. 2034 } 2035 2036 2037 void CMSCollector::getFreelistLocks() const { 2038 // Get locks for all free lists in all generations that this 2039 // collector is responsible for 2040 _cmsGen->freelistLock()->lock_without_safepoint_check(); 2041 _permGen->freelistLock()->lock_without_safepoint_check(); 2042 } 2043 2044 void CMSCollector::releaseFreelistLocks() const { 2045 // Release locks for all free lists in all generations that this 2046 // collector is responsible for 2047 _cmsGen->freelistLock()->unlock(); 2048 _permGen->freelistLock()->unlock(); 2049 } 2050 2051 bool CMSCollector::haveFreelistLocks() const { 2052 // Check locks for all free lists in all generations that this 2053 // collector is responsible for 2054 assert_lock_strong(_cmsGen->freelistLock()); 2055 assert_lock_strong(_permGen->freelistLock()); 2056 PRODUCT_ONLY(ShouldNotReachHere()); 2057 return true; 2058 } 2059 2060 // A utility class that is used by the CMS collector to 2061 // temporarily "release" the foreground collector from its 2062 // usual obligation to wait for the background collector to 2063 // complete an ongoing phase before proceeding. 2064 class ReleaseForegroundGC: public StackObj { 2065 private: 2066 CMSCollector* _c; 2067 public: 2068 ReleaseForegroundGC(CMSCollector* c) : _c(c) { 2069 assert(_c->_foregroundGCShouldWait, "Else should not need to call"); 2070 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 2071 // allow a potentially blocked foreground collector to proceed 2072 _c->_foregroundGCShouldWait = false; 2073 if (_c->_foregroundGCIsActive) { 2074 CGC_lock->notify(); 2075 } 2076 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 2077 "Possible deadlock"); 2078 } 2079 2080 ~ReleaseForegroundGC() { 2081 assert(!_c->_foregroundGCShouldWait, "Usage protocol violation?"); 2082 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 2083 _c->_foregroundGCShouldWait = true; 2084 } 2085 }; 2086 2087 // There are separate collect_in_background and collect_in_foreground because of 2088 // the different locking requirements of the background collector and the 2089 // foreground collector. There was originally an attempt to share 2090 // one "collect" method between the background collector and the foreground 2091 // collector but the if-then-else required made it cleaner to have 2092 // separate methods. 2093 void CMSCollector::collect_in_background(bool clear_all_soft_refs) { 2094 assert(Thread::current()->is_ConcurrentGC_thread(), 2095 "A CMS asynchronous collection is only allowed on a CMS thread."); 2096 2097 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2098 { 2099 bool safepoint_check = Mutex::_no_safepoint_check_flag; 2100 MutexLockerEx hl(Heap_lock, safepoint_check); 2101 MutexLockerEx x(CGC_lock, safepoint_check); 2102 if (_foregroundGCIsActive || !UseAsyncConcMarkSweepGC) { 2103 // The foreground collector is active or we're 2104 // not using asynchronous collections. Skip this 2105 // background collection. 2106 assert(!_foregroundGCShouldWait, "Should be clear"); 2107 return; 2108 } else { 2109 assert(_collectorState == Idling, "Should be idling before start."); 2110 _collectorState = InitialMarking; 2111 // Reset the expansion cause, now that we are about to begin 2112 // a new cycle. 2113 clear_expansion_cause(); 2114 } 2115 _unloaded_classes_last_cycle = cms_should_unload_classes(); // ... from last cycle 2116 // This controls class unloading in response to an explicit gc request. 2117 // If ExplicitGCInvokesConcurrentAndUnloadsClasses is set, then 2118 // we will unload classes even if CMSClassUnloadingEnabled is not set. 2119 // See CR 6541037 and related CRs. 2120 _unload_classes = _full_gc_requested // ... for this cycle 2121 && ExplicitGCInvokesConcurrentAndUnloadsClasses; 2122 _full_gc_requested = false; // acks all outstanding full gc requests 2123 // Signal that we are about to start a collection 2124 gch->increment_total_full_collections(); // ... starting a collection cycle 2125 _collection_count_start = gch->total_full_collections(); 2126 } 2127 2128 // Used for PrintGC 2129 size_t prev_used; 2130 if (PrintGC && Verbose) { 2131 prev_used = _cmsGen->used(); // XXXPERM 2132 } 2133 2134 // The change of the collection state is normally done at this level; 2135 // the exceptions are phases that are executed while the world is 2136 // stopped. For those phases the change of state is done while the 2137 // world is stopped. For baton passing purposes this allows the 2138 // background collector to finish the phase and change state atomically. 2139 // The foreground collector cannot wait on a phase that is done 2140 // while the world is stopped because the foreground collector already 2141 // has the world stopped and would deadlock. 2142 while (_collectorState != Idling) { 2143 if (TraceCMSState) { 2144 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " in CMS state %d", 2145 Thread::current(), _collectorState); 2146 } 2147 // The foreground collector 2148 // holds the Heap_lock throughout its collection. 2149 // holds the CMS token (but not the lock) 2150 // except while it is waiting for the background collector to yield. 2151 // 2152 // The foreground collector should be blocked (not for long) 2153 // if the background collector is about to start a phase 2154 // executed with world stopped. If the background 2155 // collector has already started such a phase, the 2156 // foreground collector is blocked waiting for the 2157 // Heap_lock. The stop-world phases (InitialMarking and FinalMarking) 2158 // are executed in the VM thread. 2159 // 2160 // The locking order is 2161 // PendingListLock (PLL) -- if applicable (FinalMarking) 2162 // Heap_lock (both this & PLL locked in VM_CMS_Operation::prologue()) 2163 // CMS token (claimed in 2164 // stop_world_and_do() --> 2165 // safepoint_synchronize() --> 2166 // CMSThread::synchronize()) 2167 2168 { 2169 // Check if the FG collector wants us to yield. 2170 CMSTokenSync x(true); // is cms thread 2171 if (waitForForegroundGC()) { 2172 // We yielded to a foreground GC, nothing more to be 2173 // done this round. 2174 assert(_foregroundGCShouldWait == false, "We set it to false in " 2175 "waitForForegroundGC()"); 2176 if (TraceCMSState) { 2177 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT 2178 " exiting collection CMS state %d", 2179 Thread::current(), _collectorState); 2180 } 2181 return; 2182 } else { 2183 // The background collector can run but check to see if the 2184 // foreground collector has done a collection while the 2185 // background collector was waiting to get the CGC_lock 2186 // above. If yes, break so that _foregroundGCShouldWait 2187 // is cleared before returning. 2188 if (_collectorState == Idling) { 2189 break; 2190 } 2191 } 2192 } 2193 2194 assert(_foregroundGCShouldWait, "Foreground collector, if active, " 2195 "should be waiting"); 2196 2197 switch (_collectorState) { 2198 case InitialMarking: 2199 { 2200 ReleaseForegroundGC x(this); 2201 stats().record_cms_begin(); 2202 2203 VM_CMS_Initial_Mark initial_mark_op(this); 2204 VMThread::execute(&initial_mark_op); 2205 } 2206 // The collector state may be any legal state at this point 2207 // since the background collector may have yielded to the 2208 // foreground collector. 2209 break; 2210 case Marking: 2211 // initial marking in checkpointRootsInitialWork has been completed 2212 if (markFromRoots(true)) { // we were successful 2213 assert(_collectorState == Precleaning, "Collector state should " 2214 "have changed"); 2215 } else { 2216 assert(_foregroundGCIsActive, "Internal state inconsistency"); 2217 } 2218 break; 2219 case Precleaning: 2220 if (UseAdaptiveSizePolicy) { 2221 size_policy()->concurrent_precleaning_begin(); 2222 } 2223 // marking from roots in markFromRoots has been completed 2224 preclean(); 2225 if (UseAdaptiveSizePolicy) { 2226 size_policy()->concurrent_precleaning_end(); 2227 } 2228 assert(_collectorState == AbortablePreclean || 2229 _collectorState == FinalMarking, 2230 "Collector state should have changed"); 2231 break; 2232 case AbortablePreclean: 2233 if (UseAdaptiveSizePolicy) { 2234 size_policy()->concurrent_phases_resume(); 2235 } 2236 abortable_preclean(); 2237 if (UseAdaptiveSizePolicy) { 2238 size_policy()->concurrent_precleaning_end(); 2239 } 2240 assert(_collectorState == FinalMarking, "Collector state should " 2241 "have changed"); 2242 break; 2243 case FinalMarking: 2244 { 2245 ReleaseForegroundGC x(this); 2246 2247 VM_CMS_Final_Remark final_remark_op(this); 2248 VMThread::execute(&final_remark_op); 2249 } 2250 assert(_foregroundGCShouldWait, "block post-condition"); 2251 break; 2252 case Sweeping: 2253 if (UseAdaptiveSizePolicy) { 2254 size_policy()->concurrent_sweeping_begin(); 2255 } 2256 // final marking in checkpointRootsFinal has been completed 2257 sweep(true); 2258 assert(_collectorState == Resizing, "Collector state change " 2259 "to Resizing must be done under the free_list_lock"); 2260 _full_gcs_since_conc_gc = 0; 2261 2262 // Stop the timers for adaptive size policy for the concurrent phases 2263 if (UseAdaptiveSizePolicy) { 2264 size_policy()->concurrent_sweeping_end(); 2265 size_policy()->concurrent_phases_end(gch->gc_cause(), 2266 gch->prev_gen(_cmsGen)->capacity(), 2267 _cmsGen->free()); 2268 } 2269 2270 case Resizing: { 2271 // Sweeping has been completed... 2272 // At this point the background collection has completed. 2273 // Don't move the call to compute_new_size() down 2274 // into code that might be executed if the background 2275 // collection was preempted. 2276 { 2277 ReleaseForegroundGC x(this); // unblock FG collection 2278 MutexLockerEx y(Heap_lock, Mutex::_no_safepoint_check_flag); 2279 CMSTokenSync z(true); // not strictly needed. 2280 if (_collectorState == Resizing) { 2281 compute_new_size(); 2282 _collectorState = Resetting; 2283 } else { 2284 assert(_collectorState == Idling, "The state should only change" 2285 " because the foreground collector has finished the collection"); 2286 } 2287 } 2288 break; 2289 } 2290 case Resetting: 2291 // CMS heap resizing has been completed 2292 reset(true); 2293 assert(_collectorState == Idling, "Collector state should " 2294 "have changed"); 2295 stats().record_cms_end(); 2296 // Don't move the concurrent_phases_end() and compute_new_size() 2297 // calls to here because a preempted background collection 2298 // has it's state set to "Resetting". 2299 break; 2300 case Idling: 2301 default: 2302 ShouldNotReachHere(); 2303 break; 2304 } 2305 if (TraceCMSState) { 2306 gclog_or_tty->print_cr(" Thread " INTPTR_FORMAT " done - next CMS state %d", 2307 Thread::current(), _collectorState); 2308 } 2309 assert(_foregroundGCShouldWait, "block post-condition"); 2310 } 2311 2312 // Should this be in gc_epilogue? 2313 collector_policy()->counters()->update_counters(); 2314 2315 { 2316 // Clear _foregroundGCShouldWait and, in the event that the 2317 // foreground collector is waiting, notify it, before 2318 // returning. 2319 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 2320 _foregroundGCShouldWait = false; 2321 if (_foregroundGCIsActive) { 2322 CGC_lock->notify(); 2323 } 2324 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 2325 "Possible deadlock"); 2326 } 2327 if (TraceCMSState) { 2328 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT 2329 " exiting collection CMS state %d", 2330 Thread::current(), _collectorState); 2331 } 2332 if (PrintGC && Verbose) { 2333 _cmsGen->print_heap_change(prev_used); 2334 } 2335 } 2336 2337 void CMSCollector::collect_in_foreground(bool clear_all_soft_refs) { 2338 assert(_foregroundGCIsActive && !_foregroundGCShouldWait, 2339 "Foreground collector should be waiting, not executing"); 2340 assert(Thread::current()->is_VM_thread(), "A foreground collection" 2341 "may only be done by the VM Thread with the world stopped"); 2342 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 2343 "VM thread should have CMS token"); 2344 2345 NOT_PRODUCT(TraceTime t("CMS:MS (foreground) ", PrintGCDetails && Verbose, 2346 true, gclog_or_tty);) 2347 if (UseAdaptiveSizePolicy) { 2348 size_policy()->ms_collection_begin(); 2349 } 2350 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact); 2351 2352 HandleMark hm; // Discard invalid handles created during verification 2353 2354 if (VerifyBeforeGC && 2355 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2356 Universe::verify(true); 2357 } 2358 2359 bool init_mark_was_synchronous = false; // until proven otherwise 2360 while (_collectorState != Idling) { 2361 if (TraceCMSState) { 2362 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " in CMS state %d", 2363 Thread::current(), _collectorState); 2364 } 2365 switch (_collectorState) { 2366 case InitialMarking: 2367 init_mark_was_synchronous = true; // fact to be exploited in re-mark 2368 checkpointRootsInitial(false); 2369 assert(_collectorState == Marking, "Collector state should have changed" 2370 " within checkpointRootsInitial()"); 2371 break; 2372 case Marking: 2373 // initial marking in checkpointRootsInitialWork has been completed 2374 if (VerifyDuringGC && 2375 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2376 gclog_or_tty->print("Verify before initial mark: "); 2377 Universe::verify(true); 2378 } 2379 { 2380 bool res = markFromRoots(false); 2381 assert(res && _collectorState == FinalMarking, "Collector state should " 2382 "have changed"); 2383 break; 2384 } 2385 case FinalMarking: 2386 if (VerifyDuringGC && 2387 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2388 gclog_or_tty->print("Verify before re-mark: "); 2389 Universe::verify(true); 2390 } 2391 checkpointRootsFinal(false, clear_all_soft_refs, 2392 init_mark_was_synchronous); 2393 assert(_collectorState == Sweeping, "Collector state should not " 2394 "have changed within checkpointRootsFinal()"); 2395 break; 2396 case Sweeping: 2397 // final marking in checkpointRootsFinal has been completed 2398 if (VerifyDuringGC && 2399 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2400 gclog_or_tty->print("Verify before sweep: "); 2401 Universe::verify(true); 2402 } 2403 sweep(false); 2404 assert(_collectorState == Resizing, "Incorrect state"); 2405 break; 2406 case Resizing: { 2407 // Sweeping has been completed; the actual resize in this case 2408 // is done separately; nothing to be done in this state. 2409 _collectorState = Resetting; 2410 break; 2411 } 2412 case Resetting: 2413 // The heap has been resized. 2414 if (VerifyDuringGC && 2415 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2416 gclog_or_tty->print("Verify before reset: "); 2417 Universe::verify(true); 2418 } 2419 reset(false); 2420 assert(_collectorState == Idling, "Collector state should " 2421 "have changed"); 2422 break; 2423 case Precleaning: 2424 case AbortablePreclean: 2425 // Elide the preclean phase 2426 _collectorState = FinalMarking; 2427 break; 2428 default: 2429 ShouldNotReachHere(); 2430 } 2431 if (TraceCMSState) { 2432 gclog_or_tty->print_cr(" Thread " INTPTR_FORMAT " done - next CMS state %d", 2433 Thread::current(), _collectorState); 2434 } 2435 } 2436 2437 if (UseAdaptiveSizePolicy) { 2438 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2439 size_policy()->ms_collection_end(gch->gc_cause()); 2440 } 2441 2442 if (VerifyAfterGC && 2443 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2444 Universe::verify(true); 2445 } 2446 if (TraceCMSState) { 2447 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT 2448 " exiting collection CMS state %d", 2449 Thread::current(), _collectorState); 2450 } 2451 } 2452 2453 bool CMSCollector::waitForForegroundGC() { 2454 bool res = false; 2455 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 2456 "CMS thread should have CMS token"); 2457 // Block the foreground collector until the 2458 // background collectors decides whether to 2459 // yield. 2460 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 2461 _foregroundGCShouldWait = true; 2462 if (_foregroundGCIsActive) { 2463 // The background collector yields to the 2464 // foreground collector and returns a value 2465 // indicating that it has yielded. The foreground 2466 // collector can proceed. 2467 res = true; 2468 _foregroundGCShouldWait = false; 2469 ConcurrentMarkSweepThread::clear_CMS_flag( 2470 ConcurrentMarkSweepThread::CMS_cms_has_token); 2471 ConcurrentMarkSweepThread::set_CMS_flag( 2472 ConcurrentMarkSweepThread::CMS_cms_wants_token); 2473 // Get a possibly blocked foreground thread going 2474 CGC_lock->notify(); 2475 if (TraceCMSState) { 2476 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " waiting at CMS state %d", 2477 Thread::current(), _collectorState); 2478 } 2479 while (_foregroundGCIsActive) { 2480 CGC_lock->wait(Mutex::_no_safepoint_check_flag); 2481 } 2482 ConcurrentMarkSweepThread::set_CMS_flag( 2483 ConcurrentMarkSweepThread::CMS_cms_has_token); 2484 ConcurrentMarkSweepThread::clear_CMS_flag( 2485 ConcurrentMarkSweepThread::CMS_cms_wants_token); 2486 } 2487 if (TraceCMSState) { 2488 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " continuing at CMS state %d", 2489 Thread::current(), _collectorState); 2490 } 2491 return res; 2492 } 2493 2494 // Because of the need to lock the free lists and other structures in 2495 // the collector, common to all the generations that the collector is 2496 // collecting, we need the gc_prologues of individual CMS generations 2497 // delegate to their collector. It may have been simpler had the 2498 // current infrastructure allowed one to call a prologue on a 2499 // collector. In the absence of that we have the generation's 2500 // prologue delegate to the collector, which delegates back 2501 // some "local" work to a worker method in the individual generations 2502 // that it's responsible for collecting, while itself doing any 2503 // work common to all generations it's responsible for. A similar 2504 // comment applies to the gc_epilogue()'s. 2505 // The role of the varaible _between_prologue_and_epilogue is to 2506 // enforce the invocation protocol. 2507 void CMSCollector::gc_prologue(bool full) { 2508 // Call gc_prologue_work() for each CMSGen and PermGen that 2509 // we are responsible for. 2510 2511 // The following locking discipline assumes that we are only called 2512 // when the world is stopped. 2513 assert(SafepointSynchronize::is_at_safepoint(), "world is stopped assumption"); 2514 2515 // The CMSCollector prologue must call the gc_prologues for the 2516 // "generations" (including PermGen if any) that it's responsible 2517 // for. 2518 2519 assert( Thread::current()->is_VM_thread() 2520 || ( CMSScavengeBeforeRemark 2521 && Thread::current()->is_ConcurrentGC_thread()), 2522 "Incorrect thread type for prologue execution"); 2523 2524 if (_between_prologue_and_epilogue) { 2525 // We have already been invoked; this is a gc_prologue delegation 2526 // from yet another CMS generation that we are responsible for, just 2527 // ignore it since all relevant work has already been done. 2528 return; 2529 } 2530 2531 // set a bit saying prologue has been called; cleared in epilogue 2532 _between_prologue_and_epilogue = true; 2533 // Claim locks for common data structures, then call gc_prologue_work() 2534 // for each CMSGen and PermGen that we are responsible for. 2535 2536 getFreelistLocks(); // gets free list locks on constituent spaces 2537 bitMapLock()->lock_without_safepoint_check(); 2538 2539 // Should call gc_prologue_work() for all cms gens we are responsible for 2540 bool registerClosure = _collectorState >= Marking 2541 && _collectorState < Sweeping; 2542 ModUnionClosure* muc = ParallelGCThreads > 0 ? &_modUnionClosurePar 2543 : &_modUnionClosure; 2544 _cmsGen->gc_prologue_work(full, registerClosure, muc); 2545 _permGen->gc_prologue_work(full, registerClosure, muc); 2546 2547 if (!full) { 2548 stats().record_gc0_begin(); 2549 } 2550 } 2551 2552 void ConcurrentMarkSweepGeneration::gc_prologue(bool full) { 2553 // Delegate to CMScollector which knows how to coordinate between 2554 // this and any other CMS generations that it is responsible for 2555 // collecting. 2556 collector()->gc_prologue(full); 2557 } 2558 2559 // This is a "private" interface for use by this generation's CMSCollector. 2560 // Not to be called directly by any other entity (for instance, 2561 // GenCollectedHeap, which calls the "public" gc_prologue method above). 2562 void ConcurrentMarkSweepGeneration::gc_prologue_work(bool full, 2563 bool registerClosure, ModUnionClosure* modUnionClosure) { 2564 assert(!incremental_collection_failed(), "Shouldn't be set yet"); 2565 assert(cmsSpace()->preconsumptionDirtyCardClosure() == NULL, 2566 "Should be NULL"); 2567 if (registerClosure) { 2568 cmsSpace()->setPreconsumptionDirtyCardClosure(modUnionClosure); 2569 } 2570 cmsSpace()->gc_prologue(); 2571 // Clear stat counters 2572 NOT_PRODUCT( 2573 assert(_numObjectsPromoted == 0, "check"); 2574 assert(_numWordsPromoted == 0, "check"); 2575 if (Verbose && PrintGC) { 2576 gclog_or_tty->print("Allocated "SIZE_FORMAT" objects, " 2577 SIZE_FORMAT" bytes concurrently", 2578 _numObjectsAllocated, _numWordsAllocated*sizeof(HeapWord)); 2579 } 2580 _numObjectsAllocated = 0; 2581 _numWordsAllocated = 0; 2582 ) 2583 } 2584 2585 void CMSCollector::gc_epilogue(bool full) { 2586 // The following locking discipline assumes that we are only called 2587 // when the world is stopped. 2588 assert(SafepointSynchronize::is_at_safepoint(), 2589 "world is stopped assumption"); 2590 2591 // Currently the CMS epilogue (see CompactibleFreeListSpace) merely checks 2592 // if linear allocation blocks need to be appropriately marked to allow the 2593 // the blocks to be parsable. We also check here whether we need to nudge the 2594 // CMS collector thread to start a new cycle (if it's not already active). 2595 assert( Thread::current()->is_VM_thread() 2596 || ( CMSScavengeBeforeRemark 2597 && Thread::current()->is_ConcurrentGC_thread()), 2598 "Incorrect thread type for epilogue execution"); 2599 2600 if (!_between_prologue_and_epilogue) { 2601 // We have already been invoked; this is a gc_epilogue delegation 2602 // from yet another CMS generation that we are responsible for, just 2603 // ignore it since all relevant work has already been done. 2604 return; 2605 } 2606 assert(haveFreelistLocks(), "must have freelist locks"); 2607 assert_lock_strong(bitMapLock()); 2608 2609 _cmsGen->gc_epilogue_work(full); 2610 _permGen->gc_epilogue_work(full); 2611 2612 if (_collectorState == AbortablePreclean || _collectorState == Precleaning) { 2613 // in case sampling was not already enabled, enable it 2614 _start_sampling = true; 2615 } 2616 // reset _eden_chunk_array so sampling starts afresh 2617 _eden_chunk_index = 0; 2618 2619 size_t cms_used = _cmsGen->cmsSpace()->used(); 2620 size_t perm_used = _permGen->cmsSpace()->used(); 2621 2622 // update performance counters - this uses a special version of 2623 // update_counters() that allows the utilization to be passed as a 2624 // parameter, avoiding multiple calls to used(). 2625 // 2626 _cmsGen->update_counters(cms_used); 2627 _permGen->update_counters(perm_used); 2628 2629 if (CMSIncrementalMode) { 2630 icms_update_allocation_limits(); 2631 } 2632 2633 bitMapLock()->unlock(); 2634 releaseFreelistLocks(); 2635 2636 _between_prologue_and_epilogue = false; // ready for next cycle 2637 } 2638 2639 void ConcurrentMarkSweepGeneration::gc_epilogue(bool full) { 2640 collector()->gc_epilogue(full); 2641 2642 // Also reset promotion tracking in par gc thread states. 2643 if (ParallelGCThreads > 0) { 2644 for (uint i = 0; i < ParallelGCThreads; i++) { 2645 _par_gc_thread_states[i]->promo.stopTrackingPromotions(); 2646 } 2647 } 2648 } 2649 2650 void ConcurrentMarkSweepGeneration::gc_epilogue_work(bool full) { 2651 assert(!incremental_collection_failed(), "Should have been cleared"); 2652 cmsSpace()->setPreconsumptionDirtyCardClosure(NULL); 2653 cmsSpace()->gc_epilogue(); 2654 // Print stat counters 2655 NOT_PRODUCT( 2656 assert(_numObjectsAllocated == 0, "check"); 2657 assert(_numWordsAllocated == 0, "check"); 2658 if (Verbose && PrintGC) { 2659 gclog_or_tty->print("Promoted "SIZE_FORMAT" objects, " 2660 SIZE_FORMAT" bytes", 2661 _numObjectsPromoted, _numWordsPromoted*sizeof(HeapWord)); 2662 } 2663 _numObjectsPromoted = 0; 2664 _numWordsPromoted = 0; 2665 ) 2666 2667 if (PrintGC && Verbose) { 2668 // Call down the chain in contiguous_available needs the freelistLock 2669 // so print this out before releasing the freeListLock. 2670 gclog_or_tty->print(" Contiguous available "SIZE_FORMAT" bytes ", 2671 contiguous_available()); 2672 } 2673 } 2674 2675 #ifndef PRODUCT 2676 bool CMSCollector::have_cms_token() { 2677 Thread* thr = Thread::current(); 2678 if (thr->is_VM_thread()) { 2679 return ConcurrentMarkSweepThread::vm_thread_has_cms_token(); 2680 } else if (thr->is_ConcurrentGC_thread()) { 2681 return ConcurrentMarkSweepThread::cms_thread_has_cms_token(); 2682 } else if (thr->is_GC_task_thread()) { 2683 return ConcurrentMarkSweepThread::vm_thread_has_cms_token() && 2684 ParGCRareEvent_lock->owned_by_self(); 2685 } 2686 return false; 2687 } 2688 #endif 2689 2690 // Check reachability of the given heap address in CMS generation, 2691 // treating all other generations as roots. 2692 bool CMSCollector::is_cms_reachable(HeapWord* addr) { 2693 // We could "guarantee" below, rather than assert, but i'll 2694 // leave these as "asserts" so that an adventurous debugger 2695 // could try this in the product build provided some subset of 2696 // the conditions were met, provided they were intersted in the 2697 // results and knew that the computation below wouldn't interfere 2698 // with other concurrent computations mutating the structures 2699 // being read or written. 2700 assert(SafepointSynchronize::is_at_safepoint(), 2701 "Else mutations in object graph will make answer suspect"); 2702 assert(have_cms_token(), "Should hold cms token"); 2703 assert(haveFreelistLocks(), "must hold free list locks"); 2704 assert_lock_strong(bitMapLock()); 2705 2706 // Clear the marking bit map array before starting, but, just 2707 // for kicks, first report if the given address is already marked 2708 gclog_or_tty->print_cr("Start: Address 0x%x is%s marked", addr, 2709 _markBitMap.isMarked(addr) ? "" : " not"); 2710 2711 if (verify_after_remark()) { 2712 MutexLockerEx x(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag); 2713 bool result = verification_mark_bm()->isMarked(addr); 2714 gclog_or_tty->print_cr("TransitiveMark: Address 0x%x %s marked", addr, 2715 result ? "IS" : "is NOT"); 2716 return result; 2717 } else { 2718 gclog_or_tty->print_cr("Could not compute result"); 2719 return false; 2720 } 2721 } 2722 2723 //////////////////////////////////////////////////////// 2724 // CMS Verification Support 2725 //////////////////////////////////////////////////////// 2726 // Following the remark phase, the following invariant 2727 // should hold -- each object in the CMS heap which is 2728 // marked in markBitMap() should be marked in the verification_mark_bm(). 2729 2730 class VerifyMarkedClosure: public BitMapClosure { 2731 CMSBitMap* _marks; 2732 bool _failed; 2733 2734 public: 2735 VerifyMarkedClosure(CMSBitMap* bm): _marks(bm), _failed(false) {} 2736 2737 void do_bit(size_t offset) { 2738 HeapWord* addr = _marks->offsetToHeapWord(offset); 2739 if (!_marks->isMarked(addr)) { 2740 oop(addr)->print(); 2741 gclog_or_tty->print_cr(" ("INTPTR_FORMAT" should have been marked)", addr); 2742 _failed = true; 2743 } 2744 } 2745 2746 bool failed() { return _failed; } 2747 }; 2748 2749 bool CMSCollector::verify_after_remark() { 2750 gclog_or_tty->print(" [Verifying CMS Marking... "); 2751 MutexLockerEx ml(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag); 2752 static bool init = false; 2753 2754 assert(SafepointSynchronize::is_at_safepoint(), 2755 "Else mutations in object graph will make answer suspect"); 2756 assert(have_cms_token(), 2757 "Else there may be mutual interference in use of " 2758 " verification data structures"); 2759 assert(_collectorState > Marking && _collectorState <= Sweeping, 2760 "Else marking info checked here may be obsolete"); 2761 assert(haveFreelistLocks(), "must hold free list locks"); 2762 assert_lock_strong(bitMapLock()); 2763 2764 2765 // Allocate marking bit map if not already allocated 2766 if (!init) { // first time 2767 if (!verification_mark_bm()->allocate(_span)) { 2768 return false; 2769 } 2770 init = true; 2771 } 2772 2773 assert(verification_mark_stack()->isEmpty(), "Should be empty"); 2774 2775 // Turn off refs discovery -- so we will be tracing through refs. 2776 // This is as intended, because by this time 2777 // GC must already have cleared any refs that need to be cleared, 2778 // and traced those that need to be marked; moreover, 2779 // the marking done here is not going to intefere in any 2780 // way with the marking information used by GC. 2781 NoRefDiscovery no_discovery(ref_processor()); 2782 2783 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;) 2784 2785 // Clear any marks from a previous round 2786 verification_mark_bm()->clear_all(); 2787 assert(verification_mark_stack()->isEmpty(), "markStack should be empty"); 2788 assert(overflow_list_is_empty(), "overflow list should be empty"); 2789 2790 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2791 gch->ensure_parsability(false); // fill TLABs, but no need to retire them 2792 // Update the saved marks which may affect the root scans. 2793 gch->save_marks(); 2794 2795 if (CMSRemarkVerifyVariant == 1) { 2796 // In this first variant of verification, we complete 2797 // all marking, then check if the new marks-verctor is 2798 // a subset of the CMS marks-vector. 2799 verify_after_remark_work_1(); 2800 } else if (CMSRemarkVerifyVariant == 2) { 2801 // In this second variant of verification, we flag an error 2802 // (i.e. an object reachable in the new marks-vector not reachable 2803 // in the CMS marks-vector) immediately, also indicating the 2804 // identify of an object (A) that references the unmarked object (B) -- 2805 // presumably, a mutation to A failed to be picked up by preclean/remark? 2806 verify_after_remark_work_2(); 2807 } else { 2808 warning("Unrecognized value %d for CMSRemarkVerifyVariant", 2809 CMSRemarkVerifyVariant); 2810 } 2811 gclog_or_tty->print(" done] "); 2812 return true; 2813 } 2814 2815 void CMSCollector::verify_after_remark_work_1() { 2816 ResourceMark rm; 2817 HandleMark hm; 2818 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2819 2820 // Mark from roots one level into CMS 2821 MarkRefsIntoClosure notOlder(_span, verification_mark_bm(), true /* nmethods */); 2822 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2823 2824 gch->gen_process_strong_roots(_cmsGen->level(), 2825 true, // younger gens are roots 2826 true, // collecting perm gen 2827 SharedHeap::ScanningOption(roots_scanning_options()), 2828 NULL, ¬Older); 2829 2830 // Now mark from the roots 2831 assert(_revisitStack.isEmpty(), "Should be empty"); 2832 MarkFromRootsClosure markFromRootsClosure(this, _span, 2833 verification_mark_bm(), verification_mark_stack(), &_revisitStack, 2834 false /* don't yield */, true /* verifying */); 2835 assert(_restart_addr == NULL, "Expected pre-condition"); 2836 verification_mark_bm()->iterate(&markFromRootsClosure); 2837 while (_restart_addr != NULL) { 2838 // Deal with stack overflow: by restarting at the indicated 2839 // address. 2840 HeapWord* ra = _restart_addr; 2841 markFromRootsClosure.reset(ra); 2842 _restart_addr = NULL; 2843 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end()); 2844 } 2845 assert(verification_mark_stack()->isEmpty(), "Should have been drained"); 2846 verify_work_stacks_empty(); 2847 // Should reset the revisit stack above, since no class tree 2848 // surgery is forthcoming. 2849 _revisitStack.reset(); // throwing away all contents 2850 2851 // Marking completed -- now verify that each bit marked in 2852 // verification_mark_bm() is also marked in markBitMap(); flag all 2853 // errors by printing corresponding objects. 2854 VerifyMarkedClosure vcl(markBitMap()); 2855 verification_mark_bm()->iterate(&vcl); 2856 if (vcl.failed()) { 2857 gclog_or_tty->print("Verification failed"); 2858 Universe::heap()->print(); 2859 fatal(" ... aborting"); 2860 } 2861 } 2862 2863 void CMSCollector::verify_after_remark_work_2() { 2864 ResourceMark rm; 2865 HandleMark hm; 2866 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2867 2868 // Mark from roots one level into CMS 2869 MarkRefsIntoVerifyClosure notOlder(_span, verification_mark_bm(), 2870 markBitMap(), true /* nmethods */); 2871 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2872 gch->gen_process_strong_roots(_cmsGen->level(), 2873 true, // younger gens are roots 2874 true, // collecting perm gen 2875 SharedHeap::ScanningOption(roots_scanning_options()), 2876 NULL, ¬Older); 2877 2878 // Now mark from the roots 2879 assert(_revisitStack.isEmpty(), "Should be empty"); 2880 MarkFromRootsVerifyClosure markFromRootsClosure(this, _span, 2881 verification_mark_bm(), markBitMap(), verification_mark_stack()); 2882 assert(_restart_addr == NULL, "Expected pre-condition"); 2883 verification_mark_bm()->iterate(&markFromRootsClosure); 2884 while (_restart_addr != NULL) { 2885 // Deal with stack overflow: by restarting at the indicated 2886 // address. 2887 HeapWord* ra = _restart_addr; 2888 markFromRootsClosure.reset(ra); 2889 _restart_addr = NULL; 2890 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end()); 2891 } 2892 assert(verification_mark_stack()->isEmpty(), "Should have been drained"); 2893 verify_work_stacks_empty(); 2894 // Should reset the revisit stack above, since no class tree 2895 // surgery is forthcoming. 2896 _revisitStack.reset(); // throwing away all contents 2897 2898 // Marking completed -- now verify that each bit marked in 2899 // verification_mark_bm() is also marked in markBitMap(); flag all 2900 // errors by printing corresponding objects. 2901 VerifyMarkedClosure vcl(markBitMap()); 2902 verification_mark_bm()->iterate(&vcl); 2903 assert(!vcl.failed(), "Else verification above should not have succeeded"); 2904 } 2905 2906 void ConcurrentMarkSweepGeneration::save_marks() { 2907 // delegate to CMS space 2908 cmsSpace()->save_marks(); 2909 for (uint i = 0; i < ParallelGCThreads; i++) { 2910 _par_gc_thread_states[i]->promo.startTrackingPromotions(); 2911 } 2912 } 2913 2914 bool ConcurrentMarkSweepGeneration::no_allocs_since_save_marks() { 2915 return cmsSpace()->no_allocs_since_save_marks(); 2916 } 2917 2918 #define CMS_SINCE_SAVE_MARKS_DEFN(OopClosureType, nv_suffix) \ 2919 \ 2920 void ConcurrentMarkSweepGeneration:: \ 2921 oop_since_save_marks_iterate##nv_suffix(OopClosureType* cl) { \ 2922 cl->set_generation(this); \ 2923 cmsSpace()->oop_since_save_marks_iterate##nv_suffix(cl); \ 2924 cl->reset_generation(); \ 2925 save_marks(); \ 2926 } 2927 2928 ALL_SINCE_SAVE_MARKS_CLOSURES(CMS_SINCE_SAVE_MARKS_DEFN) 2929 2930 void 2931 ConcurrentMarkSweepGeneration::object_iterate_since_last_GC(ObjectClosure* blk) 2932 { 2933 // Not currently implemented; need to do the following. -- ysr. 2934 // dld -- I think that is used for some sort of allocation profiler. So it 2935 // really means the objects allocated by the mutator since the last 2936 // GC. We could potentially implement this cheaply by recording only 2937 // the direct allocations in a side data structure. 2938 // 2939 // I think we probably ought not to be required to support these 2940 // iterations at any arbitrary point; I think there ought to be some 2941 // call to enable/disable allocation profiling in a generation/space, 2942 // and the iterator ought to return the objects allocated in the 2943 // gen/space since the enable call, or the last iterator call (which 2944 // will probably be at a GC.) That way, for gens like CM&S that would 2945 // require some extra data structure to support this, we only pay the 2946 // cost when it's in use... 2947 cmsSpace()->object_iterate_since_last_GC(blk); 2948 } 2949 2950 void 2951 ConcurrentMarkSweepGeneration::younger_refs_iterate(OopsInGenClosure* cl) { 2952 cl->set_generation(this); 2953 younger_refs_in_space_iterate(_cmsSpace, cl); 2954 cl->reset_generation(); 2955 } 2956 2957 void 2958 ConcurrentMarkSweepGeneration::oop_iterate(MemRegion mr, OopClosure* cl) { 2959 if (freelistLock()->owned_by_self()) { 2960 Generation::oop_iterate(mr, cl); 2961 } else { 2962 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2963 Generation::oop_iterate(mr, cl); 2964 } 2965 } 2966 2967 void 2968 ConcurrentMarkSweepGeneration::oop_iterate(OopClosure* cl) { 2969 if (freelistLock()->owned_by_self()) { 2970 Generation::oop_iterate(cl); 2971 } else { 2972 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2973 Generation::oop_iterate(cl); 2974 } 2975 } 2976 2977 void 2978 ConcurrentMarkSweepGeneration::object_iterate(ObjectClosure* cl) { 2979 if (freelistLock()->owned_by_self()) { 2980 Generation::object_iterate(cl); 2981 } else { 2982 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2983 Generation::object_iterate(cl); 2984 } 2985 } 2986 2987 void 2988 ConcurrentMarkSweepGeneration::pre_adjust_pointers() { 2989 } 2990 2991 void 2992 ConcurrentMarkSweepGeneration::post_compact() { 2993 } 2994 2995 void 2996 ConcurrentMarkSweepGeneration::prepare_for_verify() { 2997 // Fix the linear allocation blocks to look like free blocks. 2998 2999 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those 3000 // are not called when the heap is verified during universe initialization and 3001 // at vm shutdown. 3002 if (freelistLock()->owned_by_self()) { 3003 cmsSpace()->prepare_for_verify(); 3004 } else { 3005 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag); 3006 cmsSpace()->prepare_for_verify(); 3007 } 3008 } 3009 3010 void 3011 ConcurrentMarkSweepGeneration::verify(bool allow_dirty /* ignored */) { 3012 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those 3013 // are not called when the heap is verified during universe initialization and 3014 // at vm shutdown. 3015 if (freelistLock()->owned_by_self()) { 3016 cmsSpace()->verify(false /* ignored */); 3017 } else { 3018 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag); 3019 cmsSpace()->verify(false /* ignored */); 3020 } 3021 } 3022 3023 void CMSCollector::verify(bool allow_dirty /* ignored */) { 3024 _cmsGen->verify(allow_dirty); 3025 _permGen->verify(allow_dirty); 3026 } 3027 3028 #ifndef PRODUCT 3029 bool CMSCollector::overflow_list_is_empty() const { 3030 assert(_num_par_pushes >= 0, "Inconsistency"); 3031 if (_overflow_list == NULL) { 3032 assert(_num_par_pushes == 0, "Inconsistency"); 3033 } 3034 return _overflow_list == NULL; 3035 } 3036 3037 // The methods verify_work_stacks_empty() and verify_overflow_empty() 3038 // merely consolidate assertion checks that appear to occur together frequently. 3039 void CMSCollector::verify_work_stacks_empty() const { 3040 assert(_markStack.isEmpty(), "Marking stack should be empty"); 3041 assert(overflow_list_is_empty(), "Overflow list should be empty"); 3042 } 3043 3044 void CMSCollector::verify_overflow_empty() const { 3045 assert(overflow_list_is_empty(), "Overflow list should be empty"); 3046 assert(no_preserved_marks(), "No preserved marks"); 3047 } 3048 #endif // PRODUCT 3049 3050 void CMSCollector::setup_cms_unloading_and_verification_state() { 3051 const bool should_verify = VerifyBeforeGC || VerifyAfterGC || VerifyDuringGC 3052 || VerifyBeforeExit; 3053 const int rso = SharedHeap::SO_Symbols | SharedHeap::SO_Strings 3054 | SharedHeap::SO_CodeCache; 3055 3056 if (cms_should_unload_classes()) { // Should unload classes this cycle 3057 remove_root_scanning_option(rso); // Shrink the root set appropriately 3058 set_verifying(should_verify); // Set verification state for this cycle 3059 return; // Nothing else needs to be done at this time 3060 } 3061 3062 // Not unloading classes this cycle 3063 assert(!cms_should_unload_classes(), "Inconsitency!"); 3064 if ((!verifying() || cms_unloaded_classes_last_cycle()) && should_verify) { 3065 // We were not verifying, or we _were_ unloading classes in the last cycle, 3066 // AND some verification options are enabled this cycle; in this case, 3067 // we must make sure that the deadness map is allocated if not already so, 3068 // and cleared (if already allocated previously -- 3069 // CMSBitMap::sizeInBits() is used to determine if it's allocated). 3070 if (perm_gen_verify_bit_map()->sizeInBits() == 0) { 3071 if (!perm_gen_verify_bit_map()->allocate(_permGen->reserved())) { 3072 warning("Failed to allocate permanent generation verification CMS Bit Map;\n" 3073 "permanent generation verification disabled"); 3074 return; // Note that we leave verification disabled, so we'll retry this 3075 // allocation next cycle. We _could_ remember this failure 3076 // and skip further attempts and permanently disable verification 3077 // attempts if that is considered more desirable. 3078 } 3079 assert(perm_gen_verify_bit_map()->covers(_permGen->reserved()), 3080 "_perm_gen_ver_bit_map inconsistency?"); 3081 } else { 3082 perm_gen_verify_bit_map()->clear_all(); 3083 } 3084 // Include symbols, strings and code cache elements to prevent their resurrection. 3085 add_root_scanning_option(rso); 3086 set_verifying(true); 3087 } else if (verifying() && !should_verify) { 3088 // We were verifying, but some verification flags got disabled. 3089 set_verifying(false); 3090 // Exclude symbols, strings and code cache elements from root scanning to 3091 // reduce IM and RM pauses. 3092 remove_root_scanning_option(rso); 3093 } 3094 } 3095 3096 3097 #ifndef PRODUCT 3098 HeapWord* CMSCollector::block_start(const void* p) const { 3099 const HeapWord* addr = (HeapWord*)p; 3100 if (_span.contains(p)) { 3101 if (_cmsGen->cmsSpace()->is_in_reserved(addr)) { 3102 return _cmsGen->cmsSpace()->block_start(p); 3103 } else { 3104 assert(_permGen->cmsSpace()->is_in_reserved(addr), 3105 "Inconsistent _span?"); 3106 return _permGen->cmsSpace()->block_start(p); 3107 } 3108 } 3109 return NULL; 3110 } 3111 #endif 3112 3113 HeapWord* 3114 ConcurrentMarkSweepGeneration::expand_and_allocate(size_t word_size, 3115 bool tlab, 3116 bool parallel) { 3117 assert(!tlab, "Can't deal with TLAB allocation"); 3118 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 3119 expand(word_size*HeapWordSize, MinHeapDeltaBytes, 3120 CMSExpansionCause::_satisfy_allocation); 3121 if (GCExpandToAllocateDelayMillis > 0) { 3122 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 3123 } 3124 size_t adj_word_sz = CompactibleFreeListSpace::adjustObjectSize(word_size); 3125 if (parallel) { 3126 return cmsSpace()->par_allocate(adj_word_sz); 3127 } else { 3128 return cmsSpace()->allocate(adj_word_sz); 3129 } 3130 } 3131 3132 // YSR: All of this generation expansion/shrinking stuff is an exact copy of 3133 // OneContigSpaceCardGeneration, which makes me wonder if we should move this 3134 // to CardGeneration and share it... 3135 void ConcurrentMarkSweepGeneration::expand(size_t bytes, size_t expand_bytes, 3136 CMSExpansionCause::Cause cause) 3137 { 3138 assert_locked_or_safepoint(Heap_lock); 3139 3140 size_t aligned_bytes = ReservedSpace::page_align_size_up(bytes); 3141 size_t aligned_expand_bytes = ReservedSpace::page_align_size_up(expand_bytes); 3142 bool success = false; 3143 if (aligned_expand_bytes > aligned_bytes) { 3144 success = grow_by(aligned_expand_bytes); 3145 } 3146 if (!success) { 3147 success = grow_by(aligned_bytes); 3148 } 3149 if (!success) { 3150 size_t remaining_bytes = _virtual_space.uncommitted_size(); 3151 if (remaining_bytes > 0) { 3152 success = grow_by(remaining_bytes); 3153 } 3154 } 3155 if (GC_locker::is_active()) { 3156 if (PrintGC && Verbose) { 3157 gclog_or_tty->print_cr("Garbage collection disabled, expanded heap instead"); 3158 } 3159 } 3160 // remember why we expanded; this information is used 3161 // by shouldConcurrentCollect() when making decisions on whether to start 3162 // a new CMS cycle. 3163 if (success) { 3164 set_expansion_cause(cause); 3165 if (PrintGCDetails && Verbose) { 3166 gclog_or_tty->print_cr("Expanded CMS gen for %s", 3167 CMSExpansionCause::to_string(cause)); 3168 } 3169 } 3170 } 3171 3172 HeapWord* ConcurrentMarkSweepGeneration::expand_and_par_lab_allocate(CMSParGCThreadState* ps, size_t word_sz) { 3173 HeapWord* res = NULL; 3174 MutexLocker x(ParGCRareEvent_lock); 3175 while (true) { 3176 // Expansion by some other thread might make alloc OK now: 3177 res = ps->lab.alloc(word_sz); 3178 if (res != NULL) return res; 3179 // If there's not enough expansion space available, give up. 3180 if (_virtual_space.uncommitted_size() < (word_sz * HeapWordSize)) { 3181 return NULL; 3182 } 3183 // Otherwise, we try expansion. 3184 expand(word_sz*HeapWordSize, MinHeapDeltaBytes, 3185 CMSExpansionCause::_allocate_par_lab); 3186 // Now go around the loop and try alloc again; 3187 // A competing par_promote might beat us to the expansion space, 3188 // so we may go around the loop again if promotion fails agaion. 3189 if (GCExpandToAllocateDelayMillis > 0) { 3190 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 3191 } 3192 } 3193 } 3194 3195 3196 bool ConcurrentMarkSweepGeneration::expand_and_ensure_spooling_space( 3197 PromotionInfo* promo) { 3198 MutexLocker x(ParGCRareEvent_lock); 3199 size_t refill_size_bytes = promo->refillSize() * HeapWordSize; 3200 while (true) { 3201 // Expansion by some other thread might make alloc OK now: 3202 if (promo->ensure_spooling_space()) { 3203 assert(promo->has_spooling_space(), 3204 "Post-condition of successful ensure_spooling_space()"); 3205 return true; 3206 } 3207 // If there's not enough expansion space available, give up. 3208 if (_virtual_space.uncommitted_size() < refill_size_bytes) { 3209 return false; 3210 } 3211 // Otherwise, we try expansion. 3212 expand(refill_size_bytes, MinHeapDeltaBytes, 3213 CMSExpansionCause::_allocate_par_spooling_space); 3214 // Now go around the loop and try alloc again; 3215 // A competing allocation might beat us to the expansion space, 3216 // so we may go around the loop again if allocation fails again. 3217 if (GCExpandToAllocateDelayMillis > 0) { 3218 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 3219 } 3220 } 3221 } 3222 3223 3224 3225 void ConcurrentMarkSweepGeneration::shrink(size_t bytes) { 3226 assert_locked_or_safepoint(Heap_lock); 3227 size_t size = ReservedSpace::page_align_size_down(bytes); 3228 if (size > 0) { 3229 shrink_by(size); 3230 } 3231 } 3232 3233 bool ConcurrentMarkSweepGeneration::grow_by(size_t bytes) { 3234 assert_locked_or_safepoint(Heap_lock); 3235 bool result = _virtual_space.expand_by(bytes); 3236 if (result) { 3237 HeapWord* old_end = _cmsSpace->end(); 3238 size_t new_word_size = 3239 heap_word_size(_virtual_space.committed_size()); 3240 MemRegion mr(_cmsSpace->bottom(), new_word_size); 3241 _bts->resize(new_word_size); // resize the block offset shared array 3242 Universe::heap()->barrier_set()->resize_covered_region(mr); 3243 // Hmmmm... why doesn't CFLS::set_end verify locking? 3244 // This is quite ugly; FIX ME XXX 3245 _cmsSpace->assert_locked(); 3246 _cmsSpace->set_end((HeapWord*)_virtual_space.high()); 3247 3248 // update the space and generation capacity counters 3249 if (UsePerfData) { 3250 _space_counters->update_capacity(); 3251 _gen_counters->update_all(); 3252 } 3253 3254 if (Verbose && PrintGC) { 3255 size_t new_mem_size = _virtual_space.committed_size(); 3256 size_t old_mem_size = new_mem_size - bytes; 3257 gclog_or_tty->print_cr("Expanding %s from %ldK by %ldK to %ldK", 3258 name(), old_mem_size/K, bytes/K, new_mem_size/K); 3259 } 3260 } 3261 return result; 3262 } 3263 3264 bool ConcurrentMarkSweepGeneration::grow_to_reserved() { 3265 assert_locked_or_safepoint(Heap_lock); 3266 bool success = true; 3267 const size_t remaining_bytes = _virtual_space.uncommitted_size(); 3268 if (remaining_bytes > 0) { 3269 success = grow_by(remaining_bytes); 3270 DEBUG_ONLY(if (!success) warning("grow to reserved failed");) 3271 } 3272 return success; 3273 } 3274 3275 void ConcurrentMarkSweepGeneration::shrink_by(size_t bytes) { 3276 assert_locked_or_safepoint(Heap_lock); 3277 assert_lock_strong(freelistLock()); 3278 // XXX Fix when compaction is implemented. 3279 warning("Shrinking of CMS not yet implemented"); 3280 return; 3281 } 3282 3283 3284 // Simple ctor/dtor wrapper for accounting & timer chores around concurrent 3285 // phases. 3286 class CMSPhaseAccounting: public StackObj { 3287 public: 3288 CMSPhaseAccounting(CMSCollector *collector, 3289 const char *phase, 3290 bool print_cr = true); 3291 ~CMSPhaseAccounting(); 3292 3293 private: 3294 CMSCollector *_collector; 3295 const char *_phase; 3296 elapsedTimer _wallclock; 3297 bool _print_cr; 3298 3299 public: 3300 // Not MT-safe; so do not pass around these StackObj's 3301 // where they may be accessed by other threads. 3302 jlong wallclock_millis() { 3303 assert(_wallclock.is_active(), "Wall clock should not stop"); 3304 _wallclock.stop(); // to record time 3305 jlong ret = _wallclock.milliseconds(); 3306 _wallclock.start(); // restart 3307 return ret; 3308 } 3309 }; 3310 3311 CMSPhaseAccounting::CMSPhaseAccounting(CMSCollector *collector, 3312 const char *phase, 3313 bool print_cr) : 3314 _collector(collector), _phase(phase), _print_cr(print_cr) { 3315 3316 if (PrintCMSStatistics != 0) { 3317 _collector->resetYields(); 3318 } 3319 if (PrintGCDetails && PrintGCTimeStamps) { 3320 gclog_or_tty->date_stamp(PrintGCDateStamps); 3321 gclog_or_tty->stamp(); 3322 gclog_or_tty->print_cr(": [%s-concurrent-%s-start]", 3323 _collector->cmsGen()->short_name(), _phase); 3324 } 3325 _collector->resetTimer(); 3326 _wallclock.start(); 3327 _collector->startTimer(); 3328 } 3329 3330 CMSPhaseAccounting::~CMSPhaseAccounting() { 3331 assert(_wallclock.is_active(), "Wall clock should not have stopped"); 3332 _collector->stopTimer(); 3333 _wallclock.stop(); 3334 if (PrintGCDetails) { 3335 gclog_or_tty->date_stamp(PrintGCDateStamps); 3336 if (PrintGCTimeStamps) { 3337 gclog_or_tty->stamp(); 3338 gclog_or_tty->print(": "); 3339 } 3340 gclog_or_tty->print("[%s-concurrent-%s: %3.3f/%3.3f secs]", 3341 _collector->cmsGen()->short_name(), 3342 _phase, _collector->timerValue(), _wallclock.seconds()); 3343 if (_print_cr) { 3344 gclog_or_tty->print_cr(""); 3345 } 3346 if (PrintCMSStatistics != 0) { 3347 gclog_or_tty->print_cr(" (CMS-concurrent-%s yielded %d times)", _phase, 3348 _collector->yields()); 3349 } 3350 } 3351 } 3352 3353 // CMS work 3354 3355 // Checkpoint the roots into this generation from outside 3356 // this generation. [Note this initial checkpoint need only 3357 // be approximate -- we'll do a catch up phase subsequently.] 3358 void CMSCollector::checkpointRootsInitial(bool asynch) { 3359 assert(_collectorState == InitialMarking, "Wrong collector state"); 3360 check_correct_thread_executing(); 3361 ReferenceProcessor* rp = ref_processor(); 3362 SpecializationStats::clear(); 3363 assert(_restart_addr == NULL, "Control point invariant"); 3364 if (asynch) { 3365 // acquire locks for subsequent manipulations 3366 MutexLockerEx x(bitMapLock(), 3367 Mutex::_no_safepoint_check_flag); 3368 checkpointRootsInitialWork(asynch); 3369 rp->verify_no_references_recorded(); 3370 rp->enable_discovery(); // enable ("weak") refs discovery 3371 _collectorState = Marking; 3372 } else { 3373 // (Weak) Refs discovery: this is controlled from genCollectedHeap::do_collection 3374 // which recognizes if we are a CMS generation, and doesn't try to turn on 3375 // discovery; verify that they aren't meddling. 3376 assert(!rp->discovery_is_atomic(), 3377 "incorrect setting of discovery predicate"); 3378 assert(!rp->discovery_enabled(), "genCollectedHeap shouldn't control " 3379 "ref discovery for this generation kind"); 3380 // already have locks 3381 checkpointRootsInitialWork(asynch); 3382 rp->enable_discovery(); // now enable ("weak") refs discovery 3383 _collectorState = Marking; 3384 } 3385 SpecializationStats::print(); 3386 } 3387 3388 void CMSCollector::checkpointRootsInitialWork(bool asynch) { 3389 assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped"); 3390 assert(_collectorState == InitialMarking, "just checking"); 3391 3392 // If there has not been a GC[n-1] since last GC[n] cycle completed, 3393 // precede our marking with a collection of all 3394 // younger generations to keep floating garbage to a minimum. 3395 // XXX: we won't do this for now -- it's an optimization to be done later. 3396 3397 // already have locks 3398 assert_lock_strong(bitMapLock()); 3399 assert(_markBitMap.isAllClear(), "was reset at end of previous cycle"); 3400 3401 // Setup the verification and class unloading state for this 3402 // CMS collection cycle. 3403 setup_cms_unloading_and_verification_state(); 3404 3405 NOT_PRODUCT(TraceTime t("\ncheckpointRootsInitialWork", 3406 PrintGCDetails && Verbose, true, gclog_or_tty);) 3407 if (UseAdaptiveSizePolicy) { 3408 size_policy()->checkpoint_roots_initial_begin(); 3409 } 3410 3411 // Reset all the PLAB chunk arrays if necessary. 3412 if (_survivor_plab_array != NULL && !CMSPLABRecordAlways) { 3413 reset_survivor_plab_arrays(); 3414 } 3415 3416 ResourceMark rm; 3417 HandleMark hm; 3418 3419 FalseClosure falseClosure; 3420 // In the case of a synchronous collection, we will elide the 3421 // remark step, so it's important to catch all the nmethod oops 3422 // in this step; hence the last argument to the constrcutor below. 3423 MarkRefsIntoClosure notOlder(_span, &_markBitMap, !asynch /* nmethods */); 3424 GenCollectedHeap* gch = GenCollectedHeap::heap(); 3425 3426 verify_work_stacks_empty(); 3427 verify_overflow_empty(); 3428 3429 gch->ensure_parsability(false); // fill TLABs, but no need to retire them 3430 // Update the saved marks which may affect the root scans. 3431 gch->save_marks(); 3432 3433 // weak reference processing has not started yet. 3434 ref_processor()->set_enqueuing_is_done(false); 3435 3436 { 3437 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;) 3438 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 3439 gch->gen_process_strong_roots(_cmsGen->level(), 3440 true, // younger gens are roots 3441 true, // collecting perm gen 3442 SharedHeap::ScanningOption(roots_scanning_options()), 3443 NULL, ¬Older); 3444 } 3445 3446 // Clear mod-union table; it will be dirtied in the prologue of 3447 // CMS generation per each younger generation collection. 3448 3449 assert(_modUnionTable.isAllClear(), 3450 "Was cleared in most recent final checkpoint phase" 3451 " or no bits are set in the gc_prologue before the start of the next " 3452 "subsequent marking phase."); 3453 3454 // Temporarily disabled, since pre/post-consumption closures don't 3455 // care about precleaned cards 3456 #if 0 3457 { 3458 MemRegion mr = MemRegion((HeapWord*)_virtual_space.low(), 3459 (HeapWord*)_virtual_space.high()); 3460 _ct->ct_bs()->preclean_dirty_cards(mr); 3461 } 3462 #endif 3463 3464 // Save the end of the used_region of the constituent generations 3465 // to be used to limit the extent of sweep in each generation. 3466 save_sweep_limits(); 3467 if (UseAdaptiveSizePolicy) { 3468 size_policy()->checkpoint_roots_initial_end(gch->gc_cause()); 3469 } 3470 verify_overflow_empty(); 3471 } 3472 3473 bool CMSCollector::markFromRoots(bool asynch) { 3474 // we might be tempted to assert that: 3475 // assert(asynch == !SafepointSynchronize::is_at_safepoint(), 3476 // "inconsistent argument?"); 3477 // However that wouldn't be right, because it's possible that 3478 // a safepoint is indeed in progress as a younger generation 3479 // stop-the-world GC happens even as we mark in this generation. 3480 assert(_collectorState == Marking, "inconsistent state?"); 3481 check_correct_thread_executing(); 3482 verify_overflow_empty(); 3483 3484 bool res; 3485 if (asynch) { 3486 3487 // Start the timers for adaptive size policy for the concurrent phases 3488 // Do it here so that the foreground MS can use the concurrent 3489 // timer since a foreground MS might has the sweep done concurrently 3490 // or STW. 3491 if (UseAdaptiveSizePolicy) { 3492 size_policy()->concurrent_marking_begin(); 3493 } 3494 3495 // Weak ref discovery note: We may be discovering weak 3496 // refs in this generation concurrent (but interleaved) with 3497 // weak ref discovery by a younger generation collector. 3498 3499 CMSTokenSyncWithLocks ts(true, bitMapLock()); 3500 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 3501 CMSPhaseAccounting pa(this, "mark", !PrintGCDetails); 3502 res = markFromRootsWork(asynch); 3503 if (res) { 3504 _collectorState = Precleaning; 3505 } else { // We failed and a foreground collection wants to take over 3506 assert(_foregroundGCIsActive, "internal state inconsistency"); 3507 assert(_restart_addr == NULL, "foreground will restart from scratch"); 3508 if (PrintGCDetails) { 3509 gclog_or_tty->print_cr("bailing out to foreground collection"); 3510 } 3511 } 3512 if (UseAdaptiveSizePolicy) { 3513 size_policy()->concurrent_marking_end(); 3514 } 3515 } else { 3516 assert(SafepointSynchronize::is_at_safepoint(), 3517 "inconsistent with asynch == false"); 3518 if (UseAdaptiveSizePolicy) { 3519 size_policy()->ms_collection_marking_begin(); 3520 } 3521 // already have locks 3522 res = markFromRootsWork(asynch); 3523 _collectorState = FinalMarking; 3524 if (UseAdaptiveSizePolicy) { 3525 GenCollectedHeap* gch = GenCollectedHeap::heap(); 3526 size_policy()->ms_collection_marking_end(gch->gc_cause()); 3527 } 3528 } 3529 verify_overflow_empty(); 3530 return res; 3531 } 3532 3533 bool CMSCollector::markFromRootsWork(bool asynch) { 3534 // iterate over marked bits in bit map, doing a full scan and mark 3535 // from these roots using the following algorithm: 3536 // . if oop is to the right of the current scan pointer, 3537 // mark corresponding bit (we'll process it later) 3538 // . else (oop is to left of current scan pointer) 3539 // push oop on marking stack 3540 // . drain the marking stack 3541 3542 // Note that when we do a marking step we need to hold the 3543 // bit map lock -- recall that direct allocation (by mutators) 3544 // and promotion (by younger generation collectors) is also 3545 // marking the bit map. [the so-called allocate live policy.] 3546 // Because the implementation of bit map marking is not 3547 // robust wrt simultaneous marking of bits in the same word, 3548 // we need to make sure that there is no such interference 3549 // between concurrent such updates. 3550 3551 // already have locks 3552 assert_lock_strong(bitMapLock()); 3553 3554 // Clear the revisit stack, just in case there are any 3555 // obsolete contents from a short-circuited previous CMS cycle. 3556 _revisitStack.reset(); 3557 verify_work_stacks_empty(); 3558 verify_overflow_empty(); 3559 assert(_revisitStack.isEmpty(), "tabula rasa"); 3560 3561 bool result = false; 3562 if (CMSConcurrentMTEnabled && ParallelCMSThreads > 0) { 3563 result = do_marking_mt(asynch); 3564 } else { 3565 result = do_marking_st(asynch); 3566 } 3567 return result; 3568 } 3569 3570 // Forward decl 3571 class CMSConcMarkingTask; 3572 3573 class CMSConcMarkingTerminator: public ParallelTaskTerminator { 3574 CMSCollector* _collector; 3575 CMSConcMarkingTask* _task; 3576 bool _yield; 3577 protected: 3578 virtual void yield(); 3579 public: 3580 // "n_threads" is the number of threads to be terminated. 3581 // "queue_set" is a set of work queues of other threads. 3582 // "collector" is the CMS collector associated with this task terminator. 3583 // "yield" indicates whether we need the gang as a whole to yield. 3584 CMSConcMarkingTerminator(int n_threads, TaskQueueSetSuper* queue_set, 3585 CMSCollector* collector, bool yield) : 3586 ParallelTaskTerminator(n_threads, queue_set), 3587 _collector(collector), 3588 _yield(yield) { } 3589 3590 void set_task(CMSConcMarkingTask* task) { 3591 _task = task; 3592 } 3593 }; 3594 3595 // MT Concurrent Marking Task 3596 class CMSConcMarkingTask: public YieldingFlexibleGangTask { 3597 CMSCollector* _collector; 3598 YieldingFlexibleWorkGang* _workers; // the whole gang 3599 int _n_workers; // requested/desired # workers 3600 bool _asynch; 3601 bool _result; 3602 CompactibleFreeListSpace* _cms_space; 3603 CompactibleFreeListSpace* _perm_space; 3604 HeapWord* _global_finger; 3605 3606 // Exposed here for yielding support 3607 Mutex* const _bit_map_lock; 3608 3609 // The per thread work queues, available here for stealing 3610 OopTaskQueueSet* _task_queues; 3611 CMSConcMarkingTerminator _term; 3612 3613 public: 3614 CMSConcMarkingTask(CMSCollector* collector, 3615 CompactibleFreeListSpace* cms_space, 3616 CompactibleFreeListSpace* perm_space, 3617 bool asynch, int n_workers, 3618 YieldingFlexibleWorkGang* workers, 3619 OopTaskQueueSet* task_queues): 3620 YieldingFlexibleGangTask("Concurrent marking done multi-threaded"), 3621 _collector(collector), 3622 _cms_space(cms_space), 3623 _perm_space(perm_space), 3624 _asynch(asynch), _n_workers(n_workers), _result(true), 3625 _workers(workers), _task_queues(task_queues), 3626 _term(n_workers, task_queues, _collector, asynch), 3627 _bit_map_lock(collector->bitMapLock()) 3628 { 3629 assert(n_workers <= workers->total_workers(), 3630 "Else termination won't work correctly today"); // XXX FIX ME! 3631 _requested_size = n_workers; 3632 _term.set_task(this); 3633 assert(_cms_space->bottom() < _perm_space->bottom(), 3634 "Finger incorrectly initialized below"); 3635 _global_finger = _cms_space->bottom(); 3636 } 3637 3638 3639 OopTaskQueueSet* task_queues() { return _task_queues; } 3640 3641 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 3642 3643 HeapWord** global_finger_addr() { return &_global_finger; } 3644 3645 CMSConcMarkingTerminator* terminator() { return &_term; } 3646 3647 void work(int i); 3648 3649 virtual void coordinator_yield(); // stuff done by coordinator 3650 bool result() { return _result; } 3651 3652 void reset(HeapWord* ra) { 3653 _term.reset_for_reuse(); 3654 } 3655 3656 static bool get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 3657 OopTaskQueue* work_q); 3658 3659 private: 3660 void do_scan_and_mark(int i, CompactibleFreeListSpace* sp); 3661 void do_work_steal(int i); 3662 void bump_global_finger(HeapWord* f); 3663 }; 3664 3665 void CMSConcMarkingTerminator::yield() { 3666 if (ConcurrentMarkSweepThread::should_yield() && 3667 !_collector->foregroundGCIsActive() && 3668 _yield) { 3669 _task->yield(); 3670 } else { 3671 ParallelTaskTerminator::yield(); 3672 } 3673 } 3674 3675 //////////////////////////////////////////////////////////////// 3676 // Concurrent Marking Algorithm Sketch 3677 //////////////////////////////////////////////////////////////// 3678 // Until all tasks exhausted (both spaces): 3679 // -- claim next available chunk 3680 // -- bump global finger via CAS 3681 // -- find first object that starts in this chunk 3682 // and start scanning bitmap from that position 3683 // -- scan marked objects for oops 3684 // -- CAS-mark target, and if successful: 3685 // . if target oop is above global finger (volatile read) 3686 // nothing to do 3687 // . if target oop is in chunk and above local finger 3688 // then nothing to do 3689 // . else push on work-queue 3690 // -- Deal with possible overflow issues: 3691 // . local work-queue overflow causes stuff to be pushed on 3692 // global (common) overflow queue 3693 // . always first empty local work queue 3694 // . then get a batch of oops from global work queue if any 3695 // . then do work stealing 3696 // -- When all tasks claimed (both spaces) 3697 // and local work queue empty, 3698 // then in a loop do: 3699 // . check global overflow stack; steal a batch of oops and trace 3700 // . try to steal from other threads oif GOS is empty 3701 // . if neither is available, offer termination 3702 // -- Terminate and return result 3703 // 3704 void CMSConcMarkingTask::work(int i) { 3705 elapsedTimer _timer; 3706 ResourceMark rm; 3707 HandleMark hm; 3708 3709 DEBUG_ONLY(_collector->verify_overflow_empty();) 3710 3711 // Before we begin work, our work queue should be empty 3712 assert(work_queue(i)->size() == 0, "Expected to be empty"); 3713 // Scan the bitmap covering _cms_space, tracing through grey objects. 3714 _timer.start(); 3715 do_scan_and_mark(i, _cms_space); 3716 _timer.stop(); 3717 if (PrintCMSStatistics != 0) { 3718 gclog_or_tty->print_cr("Finished cms space scanning in %dth thread: %3.3f sec", 3719 i, _timer.seconds()); // XXX: need xxx/xxx type of notation, two timers 3720 } 3721 3722 // ... do the same for the _perm_space 3723 _timer.reset(); 3724 _timer.start(); 3725 do_scan_and_mark(i, _perm_space); 3726 _timer.stop(); 3727 if (PrintCMSStatistics != 0) { 3728 gclog_or_tty->print_cr("Finished perm space scanning in %dth thread: %3.3f sec", 3729 i, _timer.seconds()); // XXX: need xxx/xxx type of notation, two timers 3730 } 3731 3732 // ... do work stealing 3733 _timer.reset(); 3734 _timer.start(); 3735 do_work_steal(i); 3736 _timer.stop(); 3737 if (PrintCMSStatistics != 0) { 3738 gclog_or_tty->print_cr("Finished work stealing in %dth thread: %3.3f sec", 3739 i, _timer.seconds()); // XXX: need xxx/xxx type of notation, two timers 3740 } 3741 assert(_collector->_markStack.isEmpty(), "Should have been emptied"); 3742 assert(work_queue(i)->size() == 0, "Should have been emptied"); 3743 // Note that under the current task protocol, the 3744 // following assertion is true even of the spaces 3745 // expanded since the completion of the concurrent 3746 // marking. XXX This will likely change under a strict 3747 // ABORT semantics. 3748 assert(_global_finger > _cms_space->end() && 3749 _global_finger >= _perm_space->end(), 3750 "All tasks have been completed"); 3751 DEBUG_ONLY(_collector->verify_overflow_empty();) 3752 } 3753 3754 void CMSConcMarkingTask::bump_global_finger(HeapWord* f) { 3755 HeapWord* read = _global_finger; 3756 HeapWord* cur = read; 3757 while (f > read) { 3758 cur = read; 3759 read = (HeapWord*) Atomic::cmpxchg_ptr(f, &_global_finger, cur); 3760 if (cur == read) { 3761 // our cas succeeded 3762 assert(_global_finger >= f, "protocol consistency"); 3763 break; 3764 } 3765 } 3766 } 3767 3768 // This is really inefficient, and should be redone by 3769 // using (not yet available) block-read and -write interfaces to the 3770 // stack and the work_queue. XXX FIX ME !!! 3771 bool CMSConcMarkingTask::get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 3772 OopTaskQueue* work_q) { 3773 // Fast lock-free check 3774 if (ovflw_stk->length() == 0) { 3775 return false; 3776 } 3777 assert(work_q->size() == 0, "Shouldn't steal"); 3778 MutexLockerEx ml(ovflw_stk->par_lock(), 3779 Mutex::_no_safepoint_check_flag); 3780 // Grab up to 1/4 the size of the work queue 3781 size_t num = MIN2((size_t)work_q->max_elems()/4, 3782 (size_t)ParGCDesiredObjsFromOverflowList); 3783 num = MIN2(num, ovflw_stk->length()); 3784 for (int i = (int) num; i > 0; i--) { 3785 oop cur = ovflw_stk->pop(); 3786 assert(cur != NULL, "Counted wrong?"); 3787 work_q->push(cur); 3788 } 3789 return num > 0; 3790 } 3791 3792 void CMSConcMarkingTask::do_scan_and_mark(int i, CompactibleFreeListSpace* sp) { 3793 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 3794 int n_tasks = pst->n_tasks(); 3795 // We allow that there may be no tasks to do here because 3796 // we are restarting after a stack overflow. 3797 assert(pst->valid() || n_tasks == 0, "Uninitializd use?"); 3798 int nth_task = 0; 3799 3800 HeapWord* start = sp->bottom(); 3801 size_t chunk_size = sp->marking_task_size(); 3802 while (!pst->is_task_claimed(/* reference */ nth_task)) { 3803 // Having claimed the nth task in this space, 3804 // compute the chunk that it corresponds to: 3805 MemRegion span = MemRegion(start + nth_task*chunk_size, 3806 start + (nth_task+1)*chunk_size); 3807 // Try and bump the global finger via a CAS; 3808 // note that we need to do the global finger bump 3809 // _before_ taking the intersection below, because 3810 // the task corresponding to that region will be 3811 // deemed done even if the used_region() expands 3812 // because of allocation -- as it almost certainly will 3813 // during start-up while the threads yield in the 3814 // closure below. 3815 HeapWord* finger = span.end(); 3816 bump_global_finger(finger); // atomically 3817 // There are null tasks here corresponding to chunks 3818 // beyond the "top" address of the space. 3819 span = span.intersection(sp->used_region()); 3820 if (!span.is_empty()) { // Non-null task 3821 // We want to skip the first object because 3822 // the protocol is to scan any object in its entirety 3823 // that _starts_ in this span; a fortiori, any 3824 // object starting in an earlier span is scanned 3825 // as part of an earlier claimed task. 3826 // Below we use the "careful" version of block_start 3827 // so we do not try to navigate uninitialized objects. 3828 HeapWord* prev_obj = sp->block_start_careful(span.start()); 3829 // Below we use a variant of block_size that uses the 3830 // Printezis bits to avoid waiting for allocated 3831 // objects to become initialized/parsable. 3832 while (prev_obj < span.start()) { 3833 size_t sz = sp->block_size_no_stall(prev_obj, _collector); 3834 if (sz > 0) { 3835 prev_obj += sz; 3836 } else { 3837 // In this case we may end up doing a bit of redundant 3838 // scanning, but that appears unavoidable, short of 3839 // locking the free list locks; see bug 6324141. 3840 break; 3841 } 3842 } 3843 if (prev_obj < span.end()) { 3844 MemRegion my_span = MemRegion(prev_obj, span.end()); 3845 // Do the marking work within a non-empty span -- 3846 // the last argument to the constructor indicates whether the 3847 // iteration should be incremental with periodic yields. 3848 Par_MarkFromRootsClosure cl(this, _collector, my_span, 3849 &_collector->_markBitMap, 3850 work_queue(i), 3851 &_collector->_markStack, 3852 &_collector->_revisitStack, 3853 _asynch); 3854 _collector->_markBitMap.iterate(&cl, my_span.start(), my_span.end()); 3855 } // else nothing to do for this task 3856 } // else nothing to do for this task 3857 } 3858 // We'd be tempted to assert here that since there are no 3859 // more tasks left to claim in this space, the global_finger 3860 // must exceed space->top() and a fortiori space->end(). However, 3861 // that would not quite be correct because the bumping of 3862 // global_finger occurs strictly after the claiming of a task, 3863 // so by the time we reach here the global finger may not yet 3864 // have been bumped up by the thread that claimed the last 3865 // task. 3866 pst->all_tasks_completed(); 3867 } 3868 3869 class Par_ConcMarkingClosure: public OopClosure { 3870 CMSCollector* _collector; 3871 MemRegion _span; 3872 CMSBitMap* _bit_map; 3873 CMSMarkStack* _overflow_stack; 3874 CMSMarkStack* _revisit_stack; // XXXXXX Check proper use 3875 OopTaskQueue* _work_queue; 3876 3877 public: 3878 Par_ConcMarkingClosure(CMSCollector* collector, OopTaskQueue* work_queue, 3879 CMSBitMap* bit_map, CMSMarkStack* overflow_stack): 3880 _collector(collector), 3881 _span(_collector->_span), 3882 _work_queue(work_queue), 3883 _bit_map(bit_map), 3884 _overflow_stack(overflow_stack) { } // need to initialize revisit stack etc. 3885 3886 void do_oop(oop* p); 3887 void trim_queue(size_t max); 3888 void handle_stack_overflow(HeapWord* lost); 3889 }; 3890 3891 // Grey object rescan during work stealing phase -- 3892 // the salient assumption here is that stolen oops must 3893 // always be initialized, so we do not need to check for 3894 // uninitialized objects before scanning here. 3895 void Par_ConcMarkingClosure::do_oop(oop* p) { 3896 oop this_oop = *p; 3897 assert(this_oop->is_oop_or_null(), 3898 "expected an oop or NULL"); 3899 HeapWord* addr = (HeapWord*)this_oop; 3900 // Check if oop points into the CMS generation 3901 // and is not marked 3902 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 3903 // a white object ... 3904 // If we manage to "claim" the object, by being the 3905 // first thread to mark it, then we push it on our 3906 // marking stack 3907 if (_bit_map->par_mark(addr)) { // ... now grey 3908 // push on work queue (grey set) 3909 bool simulate_overflow = false; 3910 NOT_PRODUCT( 3911 if (CMSMarkStackOverflowALot && 3912 _collector->simulate_overflow()) { 3913 // simulate a stack overflow 3914 simulate_overflow = true; 3915 } 3916 ) 3917 if (simulate_overflow || 3918 !(_work_queue->push(this_oop) || _overflow_stack->par_push(this_oop))) { 3919 // stack overflow 3920 if (PrintCMSStatistics != 0) { 3921 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 3922 SIZE_FORMAT, _overflow_stack->capacity()); 3923 } 3924 // We cannot assert that the overflow stack is full because 3925 // it may have been emptied since. 3926 assert(simulate_overflow || 3927 _work_queue->size() == _work_queue->max_elems(), 3928 "Else push should have succeeded"); 3929 handle_stack_overflow(addr); 3930 } 3931 } // Else, some other thread got there first 3932 } 3933 } 3934 3935 void Par_ConcMarkingClosure::trim_queue(size_t max) { 3936 while (_work_queue->size() > max) { 3937 oop new_oop; 3938 if (_work_queue->pop_local(new_oop)) { 3939 assert(new_oop->is_oop(), "Should be an oop"); 3940 assert(_bit_map->isMarked((HeapWord*)new_oop), "Grey object"); 3941 assert(_span.contains((HeapWord*)new_oop), "Not in span"); 3942 assert(new_oop->is_parsable(), "Should be parsable"); 3943 new_oop->oop_iterate(this); // do_oop() above 3944 } 3945 } 3946 } 3947 3948 // Upon stack overflow, we discard (part of) the stack, 3949 // remembering the least address amongst those discarded 3950 // in CMSCollector's _restart_address. 3951 void Par_ConcMarkingClosure::handle_stack_overflow(HeapWord* lost) { 3952 // We need to do this under a mutex to prevent other 3953 // workers from interfering with the expansion below. 3954 MutexLockerEx ml(_overflow_stack->par_lock(), 3955 Mutex::_no_safepoint_check_flag); 3956 // Remember the least grey address discarded 3957 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 3958 _collector->lower_restart_addr(ra); 3959 _overflow_stack->reset(); // discard stack contents 3960 _overflow_stack->expand(); // expand the stack if possible 3961 } 3962 3963 3964 void CMSConcMarkingTask::do_work_steal(int i) { 3965 OopTaskQueue* work_q = work_queue(i); 3966 oop obj_to_scan; 3967 CMSBitMap* bm = &(_collector->_markBitMap); 3968 CMSMarkStack* ovflw = &(_collector->_markStack); 3969 int* seed = _collector->hash_seed(i); 3970 Par_ConcMarkingClosure cl(_collector, work_q, bm, ovflw); 3971 while (true) { 3972 cl.trim_queue(0); 3973 assert(work_q->size() == 0, "Should have been emptied above"); 3974 if (get_work_from_overflow_stack(ovflw, work_q)) { 3975 // Can't assert below because the work obtained from the 3976 // overflow stack may already have been stolen from us. 3977 // assert(work_q->size() > 0, "Work from overflow stack"); 3978 continue; 3979 } else if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 3980 assert(obj_to_scan->is_oop(), "Should be an oop"); 3981 assert(bm->isMarked((HeapWord*)obj_to_scan), "Grey object"); 3982 obj_to_scan->oop_iterate(&cl); 3983 } else if (terminator()->offer_termination()) { 3984 assert(work_q->size() == 0, "Impossible!"); 3985 break; 3986 } 3987 } 3988 } 3989 3990 // This is run by the CMS (coordinator) thread. 3991 void CMSConcMarkingTask::coordinator_yield() { 3992 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 3993 "CMS thread should hold CMS token"); 3994 3995 // First give up the locks, then yield, then re-lock 3996 // We should probably use a constructor/destructor idiom to 3997 // do this unlock/lock or modify the MutexUnlocker class to 3998 // serve our purpose. XXX 3999 assert_lock_strong(_bit_map_lock); 4000 _bit_map_lock->unlock(); 4001 ConcurrentMarkSweepThread::desynchronize(true); 4002 ConcurrentMarkSweepThread::acknowledge_yield_request(); 4003 _collector->stopTimer(); 4004 if (PrintCMSStatistics != 0) { 4005 _collector->incrementYields(); 4006 } 4007 _collector->icms_wait(); 4008 4009 // It is possible for whichever thread initiated the yield request 4010 // not to get a chance to wake up and take the bitmap lock between 4011 // this thread releasing it and reacquiring it. So, while the 4012 // should_yield() flag is on, let's sleep for a bit to give the 4013 // other thread a chance to wake up. The limit imposed on the number 4014 // of iterations is defensive, to avoid any unforseen circumstances 4015 // putting us into an infinite loop. Since it's always been this 4016 // (coordinator_yield()) method that was observed to cause the 4017 // problem, we are using a parameter (CMSCoordinatorYieldSleepCount) 4018 // which is by default non-zero. For the other seven methods that 4019 // also perform the yield operation, as are using a different 4020 // parameter (CMSYieldSleepCount) which is by default zero. This way we 4021 // can enable the sleeping for those methods too, if necessary. 4022 // See 6442774. 4023 // 4024 // We really need to reconsider the synchronization between the GC 4025 // thread and the yield-requesting threads in the future and we 4026 // should really use wait/notify, which is the recommended 4027 // way of doing this type of interaction. Additionally, we should 4028 // consolidate the eight methods that do the yield operation and they 4029 // are almost identical into one for better maintenability and 4030 // readability. See 6445193. 4031 // 4032 // Tony 2006.06.29 4033 for (unsigned i = 0; i < CMSCoordinatorYieldSleepCount && 4034 ConcurrentMarkSweepThread::should_yield() && 4035 !CMSCollector::foregroundGCIsActive(); ++i) { 4036 os::sleep(Thread::current(), 1, false); 4037 ConcurrentMarkSweepThread::acknowledge_yield_request(); 4038 } 4039 4040 ConcurrentMarkSweepThread::synchronize(true); 4041 _bit_map_lock->lock_without_safepoint_check(); 4042 _collector->startTimer(); 4043 } 4044 4045 bool CMSCollector::do_marking_mt(bool asynch) { 4046 assert(ParallelCMSThreads > 0 && conc_workers() != NULL, "precondition"); 4047 // In the future this would be determined ergonomically, based 4048 // on #cpu's, # active mutator threads (and load), and mutation rate. 4049 int num_workers = ParallelCMSThreads; 4050 4051 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 4052 CompactibleFreeListSpace* perm_space = _permGen->cmsSpace(); 4053 4054 CMSConcMarkingTask tsk(this, cms_space, perm_space, 4055 asynch, num_workers /* number requested XXX */, 4056 conc_workers(), task_queues()); 4057 4058 // Since the actual number of workers we get may be different 4059 // from the number we requested above, do we need to do anything different 4060 // below? In particular, may be we need to subclass the SequantialSubTasksDone 4061 // class?? XXX 4062 cms_space ->initialize_sequential_subtasks_for_marking(num_workers); 4063 perm_space->initialize_sequential_subtasks_for_marking(num_workers); 4064 4065 // Refs discovery is already non-atomic. 4066 assert(!ref_processor()->discovery_is_atomic(), "Should be non-atomic"); 4067 // Mutate the Refs discovery so it is MT during the 4068 // multi-threaded marking phase. 4069 ReferenceProcessorMTMutator mt(ref_processor(), num_workers > 1); 4070 4071 conc_workers()->start_task(&tsk); 4072 while (tsk.yielded()) { 4073 tsk.coordinator_yield(); 4074 conc_workers()->continue_task(&tsk); 4075 } 4076 // If the task was aborted, _restart_addr will be non-NULL 4077 assert(tsk.completed() || _restart_addr != NULL, "Inconsistency"); 4078 while (_restart_addr != NULL) { 4079 // XXX For now we do not make use of ABORTED state and have not 4080 // yet implemented the right abort semantics (even in the original 4081 // single-threaded CMS case). That needs some more investigation 4082 // and is deferred for now; see CR# TBF. 07252005YSR. XXX 4083 assert(!CMSAbortSemantics || tsk.aborted(), "Inconsistency"); 4084 // If _restart_addr is non-NULL, a marking stack overflow 4085 // occured; we need to do a fresh marking iteration from the 4086 // indicated restart address. 4087 if (_foregroundGCIsActive && asynch) { 4088 // We may be running into repeated stack overflows, having 4089 // reached the limit of the stack size, while making very 4090 // slow forward progress. It may be best to bail out and 4091 // let the foreground collector do its job. 4092 // Clear _restart_addr, so that foreground GC 4093 // works from scratch. This avoids the headache of 4094 // a "rescan" which would otherwise be needed because 4095 // of the dirty mod union table & card table. 4096 _restart_addr = NULL; 4097 return false; 4098 } 4099 // Adjust the task to restart from _restart_addr 4100 tsk.reset(_restart_addr); 4101 cms_space ->initialize_sequential_subtasks_for_marking(num_workers, 4102 _restart_addr); 4103 perm_space->initialize_sequential_subtasks_for_marking(num_workers, 4104 _restart_addr); 4105 _restart_addr = NULL; 4106 // Get the workers going again 4107 conc_workers()->start_task(&tsk); 4108 while (tsk.yielded()) { 4109 tsk.coordinator_yield(); 4110 conc_workers()->continue_task(&tsk); 4111 } 4112 } 4113 assert(tsk.completed(), "Inconsistency"); 4114 assert(tsk.result() == true, "Inconsistency"); 4115 return true; 4116 } 4117 4118 bool CMSCollector::do_marking_st(bool asynch) { 4119 ResourceMark rm; 4120 HandleMark hm; 4121 4122 MarkFromRootsClosure markFromRootsClosure(this, _span, &_markBitMap, 4123 &_markStack, &_revisitStack, CMSYield && asynch); 4124 // the last argument to iterate indicates whether the iteration 4125 // should be incremental with periodic yields. 4126 _markBitMap.iterate(&markFromRootsClosure); 4127 // If _restart_addr is non-NULL, a marking stack overflow 4128 // occured; we need to do a fresh iteration from the 4129 // indicated restart address. 4130 while (_restart_addr != NULL) { 4131 if (_foregroundGCIsActive && asynch) { 4132 // We may be running into repeated stack overflows, having 4133 // reached the limit of the stack size, while making very 4134 // slow forward progress. It may be best to bail out and 4135 // let the foreground collector do its job. 4136 // Clear _restart_addr, so that foreground GC 4137 // works from scratch. This avoids the headache of 4138 // a "rescan" which would otherwise be needed because 4139 // of the dirty mod union table & card table. 4140 _restart_addr = NULL; 4141 return false; // indicating failure to complete marking 4142 } 4143 // Deal with stack overflow: 4144 // we restart marking from _restart_addr 4145 HeapWord* ra = _restart_addr; 4146 markFromRootsClosure.reset(ra); 4147 _restart_addr = NULL; 4148 _markBitMap.iterate(&markFromRootsClosure, ra, _span.end()); 4149 } 4150 return true; 4151 } 4152 4153 void CMSCollector::preclean() { 4154 check_correct_thread_executing(); 4155 assert(Thread::current()->is_ConcurrentGC_thread(), "Wrong thread"); 4156 verify_work_stacks_empty(); 4157 verify_overflow_empty(); 4158 _abort_preclean = false; 4159 if (CMSPrecleaningEnabled) { 4160 _eden_chunk_index = 0; 4161 size_t used = get_eden_used(); 4162 size_t capacity = get_eden_capacity(); 4163 // Don't start sampling unless we will get sufficiently 4164 // many samples. 4165 if (used < (capacity/(CMSScheduleRemarkSamplingRatio * 100) 4166 * CMSScheduleRemarkEdenPenetration)) { 4167 _start_sampling = true; 4168 } else { 4169 _start_sampling = false; 4170 } 4171 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 4172 CMSPhaseAccounting pa(this, "preclean", !PrintGCDetails); 4173 preclean_work(CMSPrecleanRefLists1, CMSPrecleanSurvivors1); 4174 } 4175 CMSTokenSync x(true); // is cms thread 4176 if (CMSPrecleaningEnabled) { 4177 sample_eden(); 4178 _collectorState = AbortablePreclean; 4179 } else { 4180 _collectorState = FinalMarking; 4181 } 4182 verify_work_stacks_empty(); 4183 verify_overflow_empty(); 4184 } 4185 4186 // Try and schedule the remark such that young gen 4187 // occupancy is CMSScheduleRemarkEdenPenetration %. 4188 void CMSCollector::abortable_preclean() { 4189 check_correct_thread_executing(); 4190 assert(CMSPrecleaningEnabled, "Inconsistent control state"); 4191 assert(_collectorState == AbortablePreclean, "Inconsistent control state"); 4192 4193 // If Eden's current occupancy is below this threshold, 4194 // immediately schedule the remark; else preclean 4195 // past the next scavenge in an effort to 4196 // schedule the pause as described avove. By choosing 4197 // CMSScheduleRemarkEdenSizeThreshold >= max eden size 4198 // we will never do an actual abortable preclean cycle. 4199 if (get_eden_used() > CMSScheduleRemarkEdenSizeThreshold) { 4200 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 4201 CMSPhaseAccounting pa(this, "abortable-preclean", !PrintGCDetails); 4202 // We need more smarts in the abortable preclean 4203 // loop below to deal with cases where allocation 4204 // in young gen is very very slow, and our precleaning 4205 // is running a losing race against a horde of 4206 // mutators intent on flooding us with CMS updates 4207 // (dirty cards). 4208 // One, admittedly dumb, strategy is to give up 4209 // after a certain number of abortable precleaning loops 4210 // or after a certain maximum time. We want to make 4211 // this smarter in the next iteration. 4212 // XXX FIX ME!!! YSR 4213 size_t loops = 0, workdone = 0, cumworkdone = 0, waited = 0; 4214 while (!(should_abort_preclean() || 4215 ConcurrentMarkSweepThread::should_terminate())) { 4216 workdone = preclean_work(CMSPrecleanRefLists2, CMSPrecleanSurvivors2); 4217 cumworkdone += workdone; 4218 loops++; 4219 // Voluntarily terminate abortable preclean phase if we have 4220 // been at it for too long. 4221 if ((CMSMaxAbortablePrecleanLoops != 0) && 4222 loops >= CMSMaxAbortablePrecleanLoops) { 4223 if (PrintGCDetails) { 4224 gclog_or_tty->print(" CMS: abort preclean due to loops "); 4225 } 4226 break; 4227 } 4228 if (pa.wallclock_millis() > CMSMaxAbortablePrecleanTime) { 4229 if (PrintGCDetails) { 4230 gclog_or_tty->print(" CMS: abort preclean due to time "); 4231 } 4232 break; 4233 } 4234 // If we are doing little work each iteration, we should 4235 // take a short break. 4236 if (workdone < CMSAbortablePrecleanMinWorkPerIteration) { 4237 // Sleep for some time, waiting for work to accumulate 4238 stopTimer(); 4239 cmsThread()->wait_on_cms_lock(CMSAbortablePrecleanWaitMillis); 4240 startTimer(); 4241 waited++; 4242 } 4243 } 4244 if (PrintCMSStatistics > 0) { 4245 gclog_or_tty->print(" [%d iterations, %d waits, %d cards)] ", 4246 loops, waited, cumworkdone); 4247 } 4248 } 4249 CMSTokenSync x(true); // is cms thread 4250 if (_collectorState != Idling) { 4251 assert(_collectorState == AbortablePreclean, 4252 "Spontaneous state transition?"); 4253 _collectorState = FinalMarking; 4254 } // Else, a foreground collection completed this CMS cycle. 4255 return; 4256 } 4257 4258 // Respond to an Eden sampling opportunity 4259 void CMSCollector::sample_eden() { 4260 // Make sure a young gc cannot sneak in between our 4261 // reading and recording of a sample. 4262 assert(Thread::current()->is_ConcurrentGC_thread(), 4263 "Only the cms thread may collect Eden samples"); 4264 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 4265 "Should collect samples while holding CMS token"); 4266 if (!_start_sampling) { 4267 return; 4268 } 4269 if (_eden_chunk_array) { 4270 if (_eden_chunk_index < _eden_chunk_capacity) { 4271 _eden_chunk_array[_eden_chunk_index] = *_top_addr; // take sample 4272 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr, 4273 "Unexpected state of Eden"); 4274 // We'd like to check that what we just sampled is an oop-start address; 4275 // however, we cannot do that here since the object may not yet have been 4276 // initialized. So we'll instead do the check when we _use_ this sample 4277 // later. 4278 if (_eden_chunk_index == 0 || 4279 (pointer_delta(_eden_chunk_array[_eden_chunk_index], 4280 _eden_chunk_array[_eden_chunk_index-1]) 4281 >= CMSSamplingGrain)) { 4282 _eden_chunk_index++; // commit sample 4283 } 4284 } 4285 } 4286 if ((_collectorState == AbortablePreclean) && !_abort_preclean) { 4287 size_t used = get_eden_used(); 4288 size_t capacity = get_eden_capacity(); 4289 assert(used <= capacity, "Unexpected state of Eden"); 4290 if (used > (capacity/100 * CMSScheduleRemarkEdenPenetration)) { 4291 _abort_preclean = true; 4292 } 4293 } 4294 } 4295 4296 4297 size_t CMSCollector::preclean_work(bool clean_refs, bool clean_survivor) { 4298 assert(_collectorState == Precleaning || 4299 _collectorState == AbortablePreclean, "incorrect state"); 4300 ResourceMark rm; 4301 HandleMark hm; 4302 // Do one pass of scrubbing the discovered reference lists 4303 // to remove any reference objects with strongly-reachable 4304 // referents. 4305 if (clean_refs) { 4306 ReferenceProcessor* rp = ref_processor(); 4307 CMSPrecleanRefsYieldClosure yield_cl(this); 4308 assert(rp->span().equals(_span), "Spans should be equal"); 4309 CMSKeepAliveClosure keep_alive(this, _span, &_markBitMap, 4310 &_markStack); 4311 CMSDrainMarkingStackClosure complete_trace(this, 4312 _span, &_markBitMap, &_markStack, 4313 &keep_alive); 4314 4315 // We don't want this step to interfere with a young 4316 // collection because we don't want to take CPU 4317 // or memory bandwidth away from the young GC threads 4318 // (which may be as many as there are CPUs). 4319 // Note that we don't need to protect ourselves from 4320 // interference with mutators because they can't 4321 // manipulate the discovered reference lists nor affect 4322 // the computed reachability of the referents, the 4323 // only properties manipulated by the precleaning 4324 // of these reference lists. 4325 stopTimer(); 4326 CMSTokenSyncWithLocks x(true /* is cms thread */, 4327 bitMapLock()); 4328 startTimer(); 4329 sample_eden(); 4330 // The following will yield to allow foreground 4331 // collection to proceed promptly. XXX YSR: 4332 // The code in this method may need further 4333 // tweaking for better performance and some restructuring 4334 // for cleaner interfaces. 4335 rp->preclean_discovered_references( 4336 rp->is_alive_non_header(), &keep_alive, &complete_trace, 4337 &yield_cl); 4338 } 4339 4340 if (clean_survivor) { // preclean the active survivor space(s) 4341 assert(_young_gen->kind() == Generation::DefNew || 4342 _young_gen->kind() == Generation::ParNew || 4343 _young_gen->kind() == Generation::ASParNew, 4344 "incorrect type for cast"); 4345 DefNewGeneration* dng = (DefNewGeneration*)_young_gen; 4346 PushAndMarkClosure pam_cl(this, _span, ref_processor(), 4347 &_markBitMap, &_modUnionTable, 4348 &_markStack, &_revisitStack, 4349 true /* precleaning phase */); 4350 stopTimer(); 4351 CMSTokenSyncWithLocks ts(true /* is cms thread */, 4352 bitMapLock()); 4353 startTimer(); 4354 unsigned int before_count = 4355 GenCollectedHeap::heap()->total_collections(); 4356 SurvivorSpacePrecleanClosure 4357 sss_cl(this, _span, &_markBitMap, &_markStack, 4358 &pam_cl, before_count, CMSYield); 4359 dng->from()->object_iterate_careful(&sss_cl); 4360 dng->to()->object_iterate_careful(&sss_cl); 4361 } 4362 MarkRefsIntoAndScanClosure 4363 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable, 4364 &_markStack, &_revisitStack, this, CMSYield, 4365 true /* precleaning phase */); 4366 // CAUTION: The following closure has persistent state that may need to 4367 // be reset upon a decrease in the sequence of addresses it 4368 // processes. 4369 ScanMarkedObjectsAgainCarefullyClosure 4370 smoac_cl(this, _span, 4371 &_markBitMap, &_markStack, &_revisitStack, &mrias_cl, CMSYield); 4372 4373 // Preclean dirty cards in ModUnionTable and CardTable using 4374 // appropriate convergence criterion; 4375 // repeat CMSPrecleanIter times unless we find that 4376 // we are losing. 4377 assert(CMSPrecleanIter < 10, "CMSPrecleanIter is too large"); 4378 assert(CMSPrecleanNumerator < CMSPrecleanDenominator, 4379 "Bad convergence multiplier"); 4380 assert(CMSPrecleanThreshold >= 100, 4381 "Unreasonably low CMSPrecleanThreshold"); 4382 4383 size_t numIter, cumNumCards, lastNumCards, curNumCards; 4384 for (numIter = 0, cumNumCards = lastNumCards = curNumCards = 0; 4385 numIter < CMSPrecleanIter; 4386 numIter++, lastNumCards = curNumCards, cumNumCards += curNumCards) { 4387 curNumCards = preclean_mod_union_table(_cmsGen, &smoac_cl); 4388 if (CMSPermGenPrecleaningEnabled) { 4389 curNumCards += preclean_mod_union_table(_permGen, &smoac_cl); 4390 } 4391 if (Verbose && PrintGCDetails) { 4392 gclog_or_tty->print(" (modUnionTable: %d cards)", curNumCards); 4393 } 4394 // Either there are very few dirty cards, so re-mark 4395 // pause will be small anyway, or our pre-cleaning isn't 4396 // that much faster than the rate at which cards are being 4397 // dirtied, so we might as well stop and re-mark since 4398 // precleaning won't improve our re-mark time by much. 4399 if (curNumCards <= CMSPrecleanThreshold || 4400 (numIter > 0 && 4401 (curNumCards * CMSPrecleanDenominator > 4402 lastNumCards * CMSPrecleanNumerator))) { 4403 numIter++; 4404 cumNumCards += curNumCards; 4405 break; 4406 } 4407 } 4408 curNumCards = preclean_card_table(_cmsGen, &smoac_cl); 4409 if (CMSPermGenPrecleaningEnabled) { 4410 curNumCards += preclean_card_table(_permGen, &smoac_cl); 4411 } 4412 cumNumCards += curNumCards; 4413 if (PrintGCDetails && PrintCMSStatistics != 0) { 4414 gclog_or_tty->print_cr(" (cardTable: %d cards, re-scanned %d cards, %d iterations)", 4415 curNumCards, cumNumCards, numIter); 4416 } 4417 return cumNumCards; // as a measure of useful work done 4418 } 4419 4420 // PRECLEANING NOTES: 4421 // Precleaning involves: 4422 // . reading the bits of the modUnionTable and clearing the set bits. 4423 // . For the cards corresponding to the set bits, we scan the 4424 // objects on those cards. This means we need the free_list_lock 4425 // so that we can safely iterate over the CMS space when scanning 4426 // for oops. 4427 // . When we scan the objects, we'll be both reading and setting 4428 // marks in the marking bit map, so we'll need the marking bit map. 4429 // . For protecting _collector_state transitions, we take the CGC_lock. 4430 // Note that any races in the reading of of card table entries by the 4431 // CMS thread on the one hand and the clearing of those entries by the 4432 // VM thread or the setting of those entries by the mutator threads on the 4433 // other are quite benign. However, for efficiency it makes sense to keep 4434 // the VM thread from racing with the CMS thread while the latter is 4435 // dirty card info to the modUnionTable. We therefore also use the 4436 // CGC_lock to protect the reading of the card table and the mod union 4437 // table by the CM thread. 4438 // . We run concurrently with mutator updates, so scanning 4439 // needs to be done carefully -- we should not try to scan 4440 // potentially uninitialized objects. 4441 // 4442 // Locking strategy: While holding the CGC_lock, we scan over and 4443 // reset a maximal dirty range of the mod union / card tables, then lock 4444 // the free_list_lock and bitmap lock to do a full marking, then 4445 // release these locks; and repeat the cycle. This allows for a 4446 // certain amount of fairness in the sharing of these locks between 4447 // the CMS collector on the one hand, and the VM thread and the 4448 // mutators on the other. 4449 4450 // NOTE: preclean_mod_union_table() and preclean_card_table() 4451 // further below are largely identical; if you need to modify 4452 // one of these methods, please check the other method too. 4453 4454 size_t CMSCollector::preclean_mod_union_table( 4455 ConcurrentMarkSweepGeneration* gen, 4456 ScanMarkedObjectsAgainCarefullyClosure* cl) { 4457 verify_work_stacks_empty(); 4458 verify_overflow_empty(); 4459 4460 // strategy: starting with the first card, accumulate contiguous 4461 // ranges of dirty cards; clear these cards, then scan the region 4462 // covered by these cards. 4463 4464 // Since all of the MUT is committed ahead, we can just use 4465 // that, in case the generations expand while we are precleaning. 4466 // It might also be fine to just use the committed part of the 4467 // generation, but we might potentially miss cards when the 4468 // generation is rapidly expanding while we are in the midst 4469 // of precleaning. 4470 HeapWord* startAddr = gen->reserved().start(); 4471 HeapWord* endAddr = gen->reserved().end(); 4472 4473 cl->setFreelistLock(gen->freelistLock()); // needed for yielding 4474 4475 size_t numDirtyCards, cumNumDirtyCards; 4476 HeapWord *nextAddr, *lastAddr; 4477 for (cumNumDirtyCards = numDirtyCards = 0, 4478 nextAddr = lastAddr = startAddr; 4479 nextAddr < endAddr; 4480 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 4481 4482 ResourceMark rm; 4483 HandleMark hm; 4484 4485 MemRegion dirtyRegion; 4486 { 4487 stopTimer(); 4488 CMSTokenSync ts(true); 4489 startTimer(); 4490 sample_eden(); 4491 // Get dirty region starting at nextOffset (inclusive), 4492 // simultaneously clearing it. 4493 dirtyRegion = 4494 _modUnionTable.getAndClearMarkedRegion(nextAddr, endAddr); 4495 assert(dirtyRegion.start() >= nextAddr, 4496 "returned region inconsistent?"); 4497 } 4498 // Remember where the next search should begin. 4499 // The returned region (if non-empty) is a right open interval, 4500 // so lastOffset is obtained from the right end of that 4501 // interval. 4502 lastAddr = dirtyRegion.end(); 4503 // Should do something more transparent and less hacky XXX 4504 numDirtyCards = 4505 _modUnionTable.heapWordDiffToOffsetDiff(dirtyRegion.word_size()); 4506 4507 // We'll scan the cards in the dirty region (with periodic 4508 // yields for foreground GC as needed). 4509 if (!dirtyRegion.is_empty()) { 4510 assert(numDirtyCards > 0, "consistency check"); 4511 HeapWord* stop_point = NULL; 4512 { 4513 stopTimer(); 4514 CMSTokenSyncWithLocks ts(true, gen->freelistLock(), 4515 bitMapLock()); 4516 startTimer(); 4517 verify_work_stacks_empty(); 4518 verify_overflow_empty(); 4519 sample_eden(); 4520 stop_point = 4521 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 4522 } 4523 if (stop_point != NULL) { 4524 // The careful iteration stopped early either because it found an 4525 // uninitialized object, or because we were in the midst of an 4526 // "abortable preclean", which should now be aborted. Redirty 4527 // the bits corresponding to the partially-scanned or unscanned 4528 // cards. We'll either restart at the next block boundary or 4529 // abort the preclean. 4530 assert((CMSPermGenPrecleaningEnabled && (gen == _permGen)) || 4531 (_collectorState == AbortablePreclean && should_abort_preclean()), 4532 "Unparsable objects should only be in perm gen."); 4533 4534 stopTimer(); 4535 CMSTokenSyncWithLocks ts(true, bitMapLock()); 4536 startTimer(); 4537 _modUnionTable.mark_range(MemRegion(stop_point, dirtyRegion.end())); 4538 if (should_abort_preclean()) { 4539 break; // out of preclean loop 4540 } else { 4541 // Compute the next address at which preclean should pick up; 4542 // might need bitMapLock in order to read P-bits. 4543 lastAddr = next_card_start_after_block(stop_point); 4544 } 4545 } 4546 } else { 4547 assert(lastAddr == endAddr, "consistency check"); 4548 assert(numDirtyCards == 0, "consistency check"); 4549 break; 4550 } 4551 } 4552 verify_work_stacks_empty(); 4553 verify_overflow_empty(); 4554 return cumNumDirtyCards; 4555 } 4556 4557 // NOTE: preclean_mod_union_table() above and preclean_card_table() 4558 // below are largely identical; if you need to modify 4559 // one of these methods, please check the other method too. 4560 4561 size_t CMSCollector::preclean_card_table(ConcurrentMarkSweepGeneration* gen, 4562 ScanMarkedObjectsAgainCarefullyClosure* cl) { 4563 // strategy: it's similar to precleamModUnionTable above, in that 4564 // we accumulate contiguous ranges of dirty cards, mark these cards 4565 // precleaned, then scan the region covered by these cards. 4566 HeapWord* endAddr = (HeapWord*)(gen->_virtual_space.high()); 4567 HeapWord* startAddr = (HeapWord*)(gen->_virtual_space.low()); 4568 4569 cl->setFreelistLock(gen->freelistLock()); // needed for yielding 4570 4571 size_t numDirtyCards, cumNumDirtyCards; 4572 HeapWord *lastAddr, *nextAddr; 4573 4574 for (cumNumDirtyCards = numDirtyCards = 0, 4575 nextAddr = lastAddr = startAddr; 4576 nextAddr < endAddr; 4577 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 4578 4579 ResourceMark rm; 4580 HandleMark hm; 4581 4582 MemRegion dirtyRegion; 4583 { 4584 // See comments in "Precleaning notes" above on why we 4585 // do this locking. XXX Could the locking overheads be 4586 // too high when dirty cards are sparse? [I don't think so.] 4587 stopTimer(); 4588 CMSTokenSync x(true); // is cms thread 4589 startTimer(); 4590 sample_eden(); 4591 // Get and clear dirty region from card table 4592 dirtyRegion = _ct->ct_bs()->dirty_card_range_after_preclean( 4593 MemRegion(nextAddr, endAddr)); 4594 assert(dirtyRegion.start() >= nextAddr, 4595 "returned region inconsistent?"); 4596 } 4597 lastAddr = dirtyRegion.end(); 4598 numDirtyCards = 4599 dirtyRegion.word_size()/CardTableModRefBS::card_size_in_words; 4600 4601 if (!dirtyRegion.is_empty()) { 4602 stopTimer(); 4603 CMSTokenSyncWithLocks ts(true, gen->freelistLock(), bitMapLock()); 4604 startTimer(); 4605 sample_eden(); 4606 verify_work_stacks_empty(); 4607 verify_overflow_empty(); 4608 HeapWord* stop_point = 4609 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 4610 if (stop_point != NULL) { 4611 // The careful iteration stopped early because it found an 4612 // uninitialized object. Redirty the bits corresponding to the 4613 // partially-scanned or unscanned cards, and start again at the 4614 // next block boundary. 4615 assert(CMSPermGenPrecleaningEnabled || 4616 (_collectorState == AbortablePreclean && should_abort_preclean()), 4617 "Unparsable objects should only be in perm gen."); 4618 _ct->ct_bs()->invalidate(MemRegion(stop_point, dirtyRegion.end())); 4619 if (should_abort_preclean()) { 4620 break; // out of preclean loop 4621 } else { 4622 // Compute the next address at which preclean should pick up. 4623 lastAddr = next_card_start_after_block(stop_point); 4624 } 4625 } 4626 } else { 4627 break; 4628 } 4629 } 4630 verify_work_stacks_empty(); 4631 verify_overflow_empty(); 4632 return cumNumDirtyCards; 4633 } 4634 4635 void CMSCollector::checkpointRootsFinal(bool asynch, 4636 bool clear_all_soft_refs, bool init_mark_was_synchronous) { 4637 assert(_collectorState == FinalMarking, "incorrect state transition?"); 4638 check_correct_thread_executing(); 4639 // world is stopped at this checkpoint 4640 assert(SafepointSynchronize::is_at_safepoint(), 4641 "world should be stopped"); 4642 verify_work_stacks_empty(); 4643 verify_overflow_empty(); 4644 4645 SpecializationStats::clear(); 4646 if (PrintGCDetails) { 4647 gclog_or_tty->print("[YG occupancy: "SIZE_FORMAT" K ("SIZE_FORMAT" K)]", 4648 _young_gen->used() / K, 4649 _young_gen->capacity() / K); 4650 } 4651 if (asynch) { 4652 if (CMSScavengeBeforeRemark) { 4653 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4654 // Temporarily set flag to false, GCH->do_collection will 4655 // expect it to be false and set to true 4656 FlagSetting fl(gch->_is_gc_active, false); 4657 NOT_PRODUCT(TraceTime t("Scavenge-Before-Remark", 4658 PrintGCDetails && Verbose, true, gclog_or_tty);) 4659 int level = _cmsGen->level() - 1; 4660 if (level >= 0) { 4661 gch->do_collection(true, // full (i.e. force, see below) 4662 false, // !clear_all_soft_refs 4663 0, // size 4664 false, // is_tlab 4665 level // max_level 4666 ); 4667 } 4668 } 4669 FreelistLocker x(this); 4670 MutexLockerEx y(bitMapLock(), 4671 Mutex::_no_safepoint_check_flag); 4672 assert(!init_mark_was_synchronous, "but that's impossible!"); 4673 checkpointRootsFinalWork(asynch, clear_all_soft_refs, false); 4674 } else { 4675 // already have all the locks 4676 checkpointRootsFinalWork(asynch, clear_all_soft_refs, 4677 init_mark_was_synchronous); 4678 } 4679 verify_work_stacks_empty(); 4680 verify_overflow_empty(); 4681 SpecializationStats::print(); 4682 } 4683 4684 void CMSCollector::checkpointRootsFinalWork(bool asynch, 4685 bool clear_all_soft_refs, bool init_mark_was_synchronous) { 4686 4687 NOT_PRODUCT(TraceTime tr("checkpointRootsFinalWork", PrintGCDetails, false, gclog_or_tty);) 4688 4689 assert(haveFreelistLocks(), "must have free list locks"); 4690 assert_lock_strong(bitMapLock()); 4691 4692 if (UseAdaptiveSizePolicy) { 4693 size_policy()->checkpoint_roots_final_begin(); 4694 } 4695 4696 ResourceMark rm; 4697 HandleMark hm; 4698 4699 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4700 4701 if (cms_should_unload_classes()) { 4702 CodeCache::gc_prologue(); 4703 } 4704 assert(haveFreelistLocks(), "must have free list locks"); 4705 assert_lock_strong(bitMapLock()); 4706 4707 if (!init_mark_was_synchronous) { 4708 // We might assume that we need not fill TLAB's when 4709 // CMSScavengeBeforeRemark is set, because we may have just done 4710 // a scavenge which would have filled all TLAB's -- and besides 4711 // Eden would be empty. This however may not always be the case -- 4712 // for instance although we asked for a scavenge, it may not have 4713 // happened because of a JNI critical section. We probably need 4714 // a policy for deciding whether we can in that case wait until 4715 // the critical section releases and then do the remark following 4716 // the scavenge, and skip it here. In the absence of that policy, 4717 // or of an indication of whether the scavenge did indeed occur, 4718 // we cannot rely on TLAB's having been filled and must do 4719 // so here just in case a scavenge did not happen. 4720 gch->ensure_parsability(false); // fill TLAB's, but no need to retire them 4721 // Update the saved marks which may affect the root scans. 4722 gch->save_marks(); 4723 4724 { 4725 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;) 4726 4727 // Note on the role of the mod union table: 4728 // Since the marker in "markFromRoots" marks concurrently with 4729 // mutators, it is possible for some reachable objects not to have been 4730 // scanned. For instance, an only reference to an object A was 4731 // placed in object B after the marker scanned B. Unless B is rescanned, 4732 // A would be collected. Such updates to references in marked objects 4733 // are detected via the mod union table which is the set of all cards 4734 // dirtied since the first checkpoint in this GC cycle and prior to 4735 // the most recent young generation GC, minus those cleaned up by the 4736 // concurrent precleaning. 4737 if (CMSParallelRemarkEnabled && ParallelGCThreads > 0) { 4738 TraceTime t("Rescan (parallel) ", PrintGCDetails, false, gclog_or_tty); 4739 do_remark_parallel(); 4740 } else { 4741 TraceTime t("Rescan (non-parallel) ", PrintGCDetails, false, 4742 gclog_or_tty); 4743 do_remark_non_parallel(); 4744 } 4745 } 4746 } else { 4747 assert(!asynch, "Can't have init_mark_was_synchronous in asynch mode"); 4748 // The initial mark was stop-world, so there's no rescanning to 4749 // do; go straight on to the next step below. 4750 } 4751 verify_work_stacks_empty(); 4752 verify_overflow_empty(); 4753 4754 { 4755 NOT_PRODUCT(TraceTime ts("refProcessingWork", PrintGCDetails, false, gclog_or_tty);) 4756 refProcessingWork(asynch, clear_all_soft_refs); 4757 } 4758 verify_work_stacks_empty(); 4759 verify_overflow_empty(); 4760 4761 if (cms_should_unload_classes()) { 4762 CodeCache::gc_epilogue(); 4763 } 4764 4765 // If we encountered any (marking stack / work queue) overflow 4766 // events during the current CMS cycle, take appropriate 4767 // remedial measures, where possible, so as to try and avoid 4768 // recurrence of that condition. 4769 assert(_markStack.isEmpty(), "No grey objects"); 4770 size_t ser_ovflw = _ser_pmc_remark_ovflw + _ser_pmc_preclean_ovflw + 4771 _ser_kac_ovflw; 4772 if (ser_ovflw > 0) { 4773 if (PrintCMSStatistics != 0) { 4774 gclog_or_tty->print_cr("Marking stack overflow (benign) " 4775 "(pmc_pc="SIZE_FORMAT", pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT")", 4776 _ser_pmc_preclean_ovflw, _ser_pmc_remark_ovflw, 4777 _ser_kac_ovflw); 4778 } 4779 _markStack.expand(); 4780 _ser_pmc_remark_ovflw = 0; 4781 _ser_pmc_preclean_ovflw = 0; 4782 _ser_kac_ovflw = 0; 4783 } 4784 if (_par_pmc_remark_ovflw > 0 || _par_kac_ovflw > 0) { 4785 if (PrintCMSStatistics != 0) { 4786 gclog_or_tty->print_cr("Work queue overflow (benign) " 4787 "(pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT")", 4788 _par_pmc_remark_ovflw, _par_kac_ovflw); 4789 } 4790 _par_pmc_remark_ovflw = 0; 4791 _par_kac_ovflw = 0; 4792 } 4793 if (PrintCMSStatistics != 0) { 4794 if (_markStack._hit_limit > 0) { 4795 gclog_or_tty->print_cr(" (benign) Hit max stack size limit ("SIZE_FORMAT")", 4796 _markStack._hit_limit); 4797 } 4798 if (_markStack._failed_double > 0) { 4799 gclog_or_tty->print_cr(" (benign) Failed stack doubling ("SIZE_FORMAT")," 4800 " current capacity "SIZE_FORMAT, 4801 _markStack._failed_double, 4802 _markStack.capacity()); 4803 } 4804 } 4805 _markStack._hit_limit = 0; 4806 _markStack._failed_double = 0; 4807 4808 if ((VerifyAfterGC || VerifyDuringGC) && 4809 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 4810 verify_after_remark(); 4811 } 4812 4813 // Change under the freelistLocks. 4814 _collectorState = Sweeping; 4815 // Call isAllClear() under bitMapLock 4816 assert(_modUnionTable.isAllClear(), "Should be clear by end of the" 4817 " final marking"); 4818 if (UseAdaptiveSizePolicy) { 4819 size_policy()->checkpoint_roots_final_end(gch->gc_cause()); 4820 } 4821 } 4822 4823 // Parallel remark task 4824 class CMSParRemarkTask: public AbstractGangTask { 4825 CMSCollector* _collector; 4826 WorkGang* _workers; 4827 int _n_workers; 4828 CompactibleFreeListSpace* _cms_space; 4829 CompactibleFreeListSpace* _perm_space; 4830 4831 // The per-thread work queues, available here for stealing. 4832 OopTaskQueueSet* _task_queues; 4833 ParallelTaskTerminator _term; 4834 4835 public: 4836 CMSParRemarkTask(CMSCollector* collector, 4837 CompactibleFreeListSpace* cms_space, 4838 CompactibleFreeListSpace* perm_space, 4839 int n_workers, WorkGang* workers, 4840 OopTaskQueueSet* task_queues): 4841 AbstractGangTask("Rescan roots and grey objects in parallel"), 4842 _collector(collector), 4843 _cms_space(cms_space), _perm_space(perm_space), 4844 _n_workers(n_workers), 4845 _workers(workers), 4846 _task_queues(task_queues), 4847 _term(workers->total_workers(), task_queues) { } 4848 4849 OopTaskQueueSet* task_queues() { return _task_queues; } 4850 4851 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 4852 4853 ParallelTaskTerminator* terminator() { return &_term; } 4854 4855 void work(int i); 4856 4857 private: 4858 // Work method in support of parallel rescan ... of young gen spaces 4859 void do_young_space_rescan(int i, Par_MarkRefsIntoAndScanClosure* cl, 4860 ContiguousSpace* space, 4861 HeapWord** chunk_array, size_t chunk_top); 4862 4863 // ... of dirty cards in old space 4864 void do_dirty_card_rescan_tasks(CompactibleFreeListSpace* sp, int i, 4865 Par_MarkRefsIntoAndScanClosure* cl); 4866 4867 // ... work stealing for the above 4868 void do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, int* seed); 4869 }; 4870 4871 void CMSParRemarkTask::work(int i) { 4872 elapsedTimer _timer; 4873 ResourceMark rm; 4874 HandleMark hm; 4875 4876 // ---------- rescan from roots -------------- 4877 _timer.start(); 4878 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4879 Par_MarkRefsIntoAndScanClosure par_mrias_cl(_collector, 4880 _collector->_span, _collector->ref_processor(), 4881 &(_collector->_markBitMap), 4882 work_queue(i), &(_collector->_revisitStack)); 4883 4884 // Rescan young gen roots first since these are likely 4885 // coarsely partitioned and may, on that account, constitute 4886 // the critical path; thus, it's best to start off that 4887 // work first. 4888 // ---------- young gen roots -------------- 4889 { 4890 DefNewGeneration* dng = _collector->_young_gen->as_DefNewGeneration(); 4891 EdenSpace* eden_space = dng->eden(); 4892 ContiguousSpace* from_space = dng->from(); 4893 ContiguousSpace* to_space = dng->to(); 4894 4895 HeapWord** eca = _collector->_eden_chunk_array; 4896 size_t ect = _collector->_eden_chunk_index; 4897 HeapWord** sca = _collector->_survivor_chunk_array; 4898 size_t sct = _collector->_survivor_chunk_index; 4899 4900 assert(ect <= _collector->_eden_chunk_capacity, "out of bounds"); 4901 assert(sct <= _collector->_survivor_chunk_capacity, "out of bounds"); 4902 4903 do_young_space_rescan(i, &par_mrias_cl, to_space, NULL, 0); 4904 do_young_space_rescan(i, &par_mrias_cl, from_space, sca, sct); 4905 do_young_space_rescan(i, &par_mrias_cl, eden_space, eca, ect); 4906 4907 _timer.stop(); 4908 if (PrintCMSStatistics != 0) { 4909 gclog_or_tty->print_cr( 4910 "Finished young gen rescan work in %dth thread: %3.3f sec", 4911 i, _timer.seconds()); 4912 } 4913 } 4914 4915 // ---------- remaining roots -------------- 4916 _timer.reset(); 4917 _timer.start(); 4918 gch->gen_process_strong_roots(_collector->_cmsGen->level(), 4919 false, // yg was scanned above 4920 true, // collecting perm gen 4921 SharedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()), 4922 NULL, &par_mrias_cl); 4923 _timer.stop(); 4924 if (PrintCMSStatistics != 0) { 4925 gclog_or_tty->print_cr( 4926 "Finished remaining root rescan work in %dth thread: %3.3f sec", 4927 i, _timer.seconds()); 4928 } 4929 4930 // ---------- rescan dirty cards ------------ 4931 _timer.reset(); 4932 _timer.start(); 4933 4934 // Do the rescan tasks for each of the two spaces 4935 // (cms_space and perm_space) in turn. 4936 do_dirty_card_rescan_tasks(_cms_space, i, &par_mrias_cl); 4937 do_dirty_card_rescan_tasks(_perm_space, i, &par_mrias_cl); 4938 _timer.stop(); 4939 if (PrintCMSStatistics != 0) { 4940 gclog_or_tty->print_cr( 4941 "Finished dirty card rescan work in %dth thread: %3.3f sec", 4942 i, _timer.seconds()); 4943 } 4944 4945 // ---------- steal work from other threads ... 4946 // ---------- ... and drain overflow list. 4947 _timer.reset(); 4948 _timer.start(); 4949 do_work_steal(i, &par_mrias_cl, _collector->hash_seed(i)); 4950 _timer.stop(); 4951 if (PrintCMSStatistics != 0) { 4952 gclog_or_tty->print_cr( 4953 "Finished work stealing in %dth thread: %3.3f sec", 4954 i, _timer.seconds()); 4955 } 4956 } 4957 4958 void 4959 CMSParRemarkTask::do_young_space_rescan(int i, 4960 Par_MarkRefsIntoAndScanClosure* cl, ContiguousSpace* space, 4961 HeapWord** chunk_array, size_t chunk_top) { 4962 // Until all tasks completed: 4963 // . claim an unclaimed task 4964 // . compute region boundaries corresponding to task claimed 4965 // using chunk_array 4966 // . par_oop_iterate(cl) over that region 4967 4968 ResourceMark rm; 4969 HandleMark hm; 4970 4971 SequentialSubTasksDone* pst = space->par_seq_tasks(); 4972 assert(pst->valid(), "Uninitialized use?"); 4973 4974 int nth_task = 0; 4975 int n_tasks = pst->n_tasks(); 4976 4977 HeapWord *start, *end; 4978 while (!pst->is_task_claimed(/* reference */ nth_task)) { 4979 // We claimed task # nth_task; compute its boundaries. 4980 if (chunk_top == 0) { // no samples were taken 4981 assert(nth_task == 0 && n_tasks == 1, "Can have only 1 EdenSpace task"); 4982 start = space->bottom(); 4983 end = space->top(); 4984 } else if (nth_task == 0) { 4985 start = space->bottom(); 4986 end = chunk_array[nth_task]; 4987 } else if (nth_task < (jint)chunk_top) { 4988 assert(nth_task >= 1, "Control point invariant"); 4989 start = chunk_array[nth_task - 1]; 4990 end = chunk_array[nth_task]; 4991 } else { 4992 assert(nth_task == (jint)chunk_top, "Control point invariant"); 4993 start = chunk_array[chunk_top - 1]; 4994 end = space->top(); 4995 } 4996 MemRegion mr(start, end); 4997 // Verify that mr is in space 4998 assert(mr.is_empty() || space->used_region().contains(mr), 4999 "Should be in space"); 5000 // Verify that "start" is an object boundary 5001 assert(mr.is_empty() || oop(mr.start())->is_oop(), 5002 "Should be an oop"); 5003 space->par_oop_iterate(mr, cl); 5004 } 5005 pst->all_tasks_completed(); 5006 } 5007 5008 void 5009 CMSParRemarkTask::do_dirty_card_rescan_tasks( 5010 CompactibleFreeListSpace* sp, int i, 5011 Par_MarkRefsIntoAndScanClosure* cl) { 5012 // Until all tasks completed: 5013 // . claim an unclaimed task 5014 // . compute region boundaries corresponding to task claimed 5015 // . transfer dirty bits ct->mut for that region 5016 // . apply rescanclosure to dirty mut bits for that region 5017 5018 ResourceMark rm; 5019 HandleMark hm; 5020 5021 OopTaskQueue* work_q = work_queue(i); 5022 ModUnionClosure modUnionClosure(&(_collector->_modUnionTable)); 5023 // CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! 5024 // CAUTION: This closure has state that persists across calls to 5025 // the work method dirty_range_iterate_clear() in that it has 5026 // imbedded in it a (subtype of) UpwardsObjectClosure. The 5027 // use of that state in the imbedded UpwardsObjectClosure instance 5028 // assumes that the cards are always iterated (even if in parallel 5029 // by several threads) in monotonically increasing order per each 5030 // thread. This is true of the implementation below which picks 5031 // card ranges (chunks) in monotonically increasing order globally 5032 // and, a-fortiori, in monotonically increasing order per thread 5033 // (the latter order being a subsequence of the former). 5034 // If the work code below is ever reorganized into a more chaotic 5035 // work-partitioning form than the current "sequential tasks" 5036 // paradigm, the use of that persistent state will have to be 5037 // revisited and modified appropriately. See also related 5038 // bug 4756801 work on which should examine this code to make 5039 // sure that the changes there do not run counter to the 5040 // assumptions made here and necessary for correctness and 5041 // efficiency. Note also that this code might yield inefficient 5042 // behaviour in the case of very large objects that span one or 5043 // more work chunks. Such objects would potentially be scanned 5044 // several times redundantly. Work on 4756801 should try and 5045 // address that performance anomaly if at all possible. XXX 5046 MemRegion full_span = _collector->_span; 5047 CMSBitMap* bm = &(_collector->_markBitMap); // shared 5048 CMSMarkStack* rs = &(_collector->_revisitStack); // shared 5049 MarkFromDirtyCardsClosure 5050 greyRescanClosure(_collector, full_span, // entire span of interest 5051 sp, bm, work_q, rs, cl); 5052 5053 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 5054 assert(pst->valid(), "Uninitialized use?"); 5055 int nth_task = 0; 5056 const int alignment = CardTableModRefBS::card_size * BitsPerWord; 5057 MemRegion span = sp->used_region(); 5058 HeapWord* start_addr = span.start(); 5059 HeapWord* end_addr = (HeapWord*)round_to((intptr_t)span.end(), 5060 alignment); 5061 const size_t chunk_size = sp->rescan_task_size(); // in HeapWord units 5062 assert((HeapWord*)round_to((intptr_t)start_addr, alignment) == 5063 start_addr, "Check alignment"); 5064 assert((size_t)round_to((intptr_t)chunk_size, alignment) == 5065 chunk_size, "Check alignment"); 5066 5067 while (!pst->is_task_claimed(/* reference */ nth_task)) { 5068 // Having claimed the nth_task, compute corresponding mem-region, 5069 // which is a-fortiori aligned correctly (i.e. at a MUT bopundary). 5070 // The alignment restriction ensures that we do not need any 5071 // synchronization with other gang-workers while setting or 5072 // clearing bits in thus chunk of the MUT. 5073 MemRegion this_span = MemRegion(start_addr + nth_task*chunk_size, 5074 start_addr + (nth_task+1)*chunk_size); 5075 // The last chunk's end might be way beyond end of the 5076 // used region. In that case pull back appropriately. 5077 if (this_span.end() > end_addr) { 5078 this_span.set_end(end_addr); 5079 assert(!this_span.is_empty(), "Program logic (calculation of n_tasks)"); 5080 } 5081 // Iterate over the dirty cards covering this chunk, marking them 5082 // precleaned, and setting the corresponding bits in the mod union 5083 // table. Since we have been careful to partition at Card and MUT-word 5084 // boundaries no synchronization is needed between parallel threads. 5085 _collector->_ct->ct_bs()->dirty_card_iterate(this_span, 5086 &modUnionClosure); 5087 5088 // Having transferred these marks into the modUnionTable, 5089 // rescan the marked objects on the dirty cards in the modUnionTable. 5090 // Even if this is at a synchronous collection, the initial marking 5091 // may have been done during an asynchronous collection so there 5092 // may be dirty bits in the mod-union table. 5093 _collector->_modUnionTable.dirty_range_iterate_clear( 5094 this_span, &greyRescanClosure); 5095 _collector->_modUnionTable.verifyNoOneBitsInRange( 5096 this_span.start(), 5097 this_span.end()); 5098 } 5099 pst->all_tasks_completed(); // declare that i am done 5100 } 5101 5102 // . see if we can share work_queues with ParNew? XXX 5103 void 5104 CMSParRemarkTask::do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, 5105 int* seed) { 5106 OopTaskQueue* work_q = work_queue(i); 5107 NOT_PRODUCT(int num_steals = 0;) 5108 oop obj_to_scan; 5109 CMSBitMap* bm = &(_collector->_markBitMap); 5110 size_t num_from_overflow_list = 5111 MIN2((size_t)work_q->max_elems()/4, 5112 (size_t)ParGCDesiredObjsFromOverflowList); 5113 5114 while (true) { 5115 // Completely finish any left over work from (an) earlier round(s) 5116 cl->trim_queue(0); 5117 // Now check if there's any work in the overflow list 5118 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 5119 work_q)) { 5120 // found something in global overflow list; 5121 // not yet ready to go stealing work from others. 5122 // We'd like to assert(work_q->size() != 0, ...) 5123 // because we just took work from the overflow list, 5124 // but of course we can't since all of that could have 5125 // been already stolen from us. 5126 // "He giveth and He taketh away." 5127 continue; 5128 } 5129 // Verify that we have no work before we resort to stealing 5130 assert(work_q->size() == 0, "Have work, shouldn't steal"); 5131 // Try to steal from other queues that have work 5132 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 5133 NOT_PRODUCT(num_steals++;) 5134 assert(obj_to_scan->is_oop(), "Oops, not an oop!"); 5135 assert(bm->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 5136 // Do scanning work 5137 obj_to_scan->oop_iterate(cl); 5138 // Loop around, finish this work, and try to steal some more 5139 } else if (terminator()->offer_termination()) { 5140 break; // nirvana from the infinite cycle 5141 } 5142 } 5143 NOT_PRODUCT( 5144 if (PrintCMSStatistics != 0) { 5145 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals); 5146 } 5147 ) 5148 assert(work_q->size() == 0 && _collector->overflow_list_is_empty(), 5149 "Else our work is not yet done"); 5150 } 5151 5152 // Return a thread-local PLAB recording array, as appropriate. 5153 void* CMSCollector::get_data_recorder(int thr_num) { 5154 if (_survivor_plab_array != NULL && 5155 (CMSPLABRecordAlways || 5156 (_collectorState > Marking && _collectorState < FinalMarking))) { 5157 assert(thr_num < (int)ParallelGCThreads, "thr_num is out of bounds"); 5158 ChunkArray* ca = &_survivor_plab_array[thr_num]; 5159 ca->reset(); // clear it so that fresh data is recorded 5160 return (void*) ca; 5161 } else { 5162 return NULL; 5163 } 5164 } 5165 5166 // Reset all the thread-local PLAB recording arrays 5167 void CMSCollector::reset_survivor_plab_arrays() { 5168 for (uint i = 0; i < ParallelGCThreads; i++) { 5169 _survivor_plab_array[i].reset(); 5170 } 5171 } 5172 5173 // Merge the per-thread plab arrays into the global survivor chunk 5174 // array which will provide the partitioning of the survivor space 5175 // for CMS rescan. 5176 void CMSCollector::merge_survivor_plab_arrays(ContiguousSpace* surv) { 5177 assert(_survivor_plab_array != NULL, "Error"); 5178 assert(_survivor_chunk_array != NULL, "Error"); 5179 assert(_collectorState == FinalMarking, "Error"); 5180 for (uint j = 0; j < ParallelGCThreads; j++) { 5181 _cursor[j] = 0; 5182 } 5183 HeapWord* top = surv->top(); 5184 size_t i; 5185 for (i = 0; i < _survivor_chunk_capacity; i++) { // all sca entries 5186 HeapWord* min_val = top; // Higher than any PLAB address 5187 uint min_tid = 0; // position of min_val this round 5188 for (uint j = 0; j < ParallelGCThreads; j++) { 5189 ChunkArray* cur_sca = &_survivor_plab_array[j]; 5190 if (_cursor[j] == cur_sca->end()) { 5191 continue; 5192 } 5193 assert(_cursor[j] < cur_sca->end(), "ctl pt invariant"); 5194 HeapWord* cur_val = cur_sca->nth(_cursor[j]); 5195 assert(surv->used_region().contains(cur_val), "Out of bounds value"); 5196 if (cur_val < min_val) { 5197 min_tid = j; 5198 min_val = cur_val; 5199 } else { 5200 assert(cur_val < top, "All recorded addresses should be less"); 5201 } 5202 } 5203 // At this point min_val and min_tid are respectively 5204 // the least address in _survivor_plab_array[j]->nth(_cursor[j]) 5205 // and the thread (j) that witnesses that address. 5206 // We record this address in the _survivor_chunk_array[i] 5207 // and increment _cursor[min_tid] prior to the next round i. 5208 if (min_val == top) { 5209 break; 5210 } 5211 _survivor_chunk_array[i] = min_val; 5212 _cursor[min_tid]++; 5213 } 5214 // We are all done; record the size of the _survivor_chunk_array 5215 _survivor_chunk_index = i; // exclusive: [0, i) 5216 if (PrintCMSStatistics > 0) { 5217 gclog_or_tty->print(" (Survivor:" SIZE_FORMAT "chunks) ", i); 5218 } 5219 // Verify that we used up all the recorded entries 5220 #ifdef ASSERT 5221 size_t total = 0; 5222 for (uint j = 0; j < ParallelGCThreads; j++) { 5223 assert(_cursor[j] == _survivor_plab_array[j].end(), "Ctl pt invariant"); 5224 total += _cursor[j]; 5225 } 5226 assert(total == _survivor_chunk_index, "Ctl Pt Invariant"); 5227 // Check that the merged array is in sorted order 5228 if (total > 0) { 5229 for (size_t i = 0; i < total - 1; i++) { 5230 if (PrintCMSStatistics > 0) { 5231 gclog_or_tty->print(" (chunk" SIZE_FORMAT ":" INTPTR_FORMAT ") ", 5232 i, _survivor_chunk_array[i]); 5233 } 5234 assert(_survivor_chunk_array[i] < _survivor_chunk_array[i+1], 5235 "Not sorted"); 5236 } 5237 } 5238 #endif // ASSERT 5239 } 5240 5241 // Set up the space's par_seq_tasks structure for work claiming 5242 // for parallel rescan of young gen. 5243 // See ParRescanTask where this is currently used. 5244 void 5245 CMSCollector:: 5246 initialize_sequential_subtasks_for_young_gen_rescan(int n_threads) { 5247 assert(n_threads > 0, "Unexpected n_threads argument"); 5248 DefNewGeneration* dng = (DefNewGeneration*)_young_gen; 5249 5250 // Eden space 5251 { 5252 SequentialSubTasksDone* pst = dng->eden()->par_seq_tasks(); 5253 assert(!pst->valid(), "Clobbering existing data?"); 5254 // Each valid entry in [0, _eden_chunk_index) represents a task. 5255 size_t n_tasks = _eden_chunk_index + 1; 5256 assert(n_tasks == 1 || _eden_chunk_array != NULL, "Error"); 5257 pst->set_par_threads(n_threads); 5258 pst->set_n_tasks((int)n_tasks); 5259 } 5260 5261 // Merge the survivor plab arrays into _survivor_chunk_array 5262 if (_survivor_plab_array != NULL) { 5263 merge_survivor_plab_arrays(dng->from()); 5264 } else { 5265 assert(_survivor_chunk_index == 0, "Error"); 5266 } 5267 5268 // To space 5269 { 5270 SequentialSubTasksDone* pst = dng->to()->par_seq_tasks(); 5271 assert(!pst->valid(), "Clobbering existing data?"); 5272 pst->set_par_threads(n_threads); 5273 pst->set_n_tasks(1); 5274 assert(pst->valid(), "Error"); 5275 } 5276 5277 // From space 5278 { 5279 SequentialSubTasksDone* pst = dng->from()->par_seq_tasks(); 5280 assert(!pst->valid(), "Clobbering existing data?"); 5281 size_t n_tasks = _survivor_chunk_index + 1; 5282 assert(n_tasks == 1 || _survivor_chunk_array != NULL, "Error"); 5283 pst->set_par_threads(n_threads); 5284 pst->set_n_tasks((int)n_tasks); 5285 assert(pst->valid(), "Error"); 5286 } 5287 } 5288 5289 // Parallel version of remark 5290 void CMSCollector::do_remark_parallel() { 5291 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5292 WorkGang* workers = gch->workers(); 5293 assert(workers != NULL, "Need parallel worker threads."); 5294 int n_workers = workers->total_workers(); 5295 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 5296 CompactibleFreeListSpace* perm_space = _permGen->cmsSpace(); 5297 5298 CMSParRemarkTask tsk(this, 5299 cms_space, perm_space, 5300 n_workers, workers, task_queues()); 5301 5302 // Set up for parallel process_strong_roots work. 5303 gch->set_par_threads(n_workers); 5304 gch->change_strong_roots_parity(); 5305 // We won't be iterating over the cards in the card table updating 5306 // the younger_gen cards, so we shouldn't call the following else 5307 // the verification code as well as subsequent younger_refs_iterate 5308 // code would get confused. XXX 5309 // gch->rem_set()->prepare_for_younger_refs_iterate(true); // parallel 5310 5311 // The young gen rescan work will not be done as part of 5312 // process_strong_roots (which currently doesn't knw how to 5313 // parallelize such a scan), but rather will be broken up into 5314 // a set of parallel tasks (via the sampling that the [abortable] 5315 // preclean phase did of EdenSpace, plus the [two] tasks of 5316 // scanning the [two] survivor spaces. Further fine-grain 5317 // parallelization of the scanning of the survivor spaces 5318 // themselves, and of precleaning of the younger gen itself 5319 // is deferred to the future. 5320 initialize_sequential_subtasks_for_young_gen_rescan(n_workers); 5321 5322 // The dirty card rescan work is broken up into a "sequence" 5323 // of parallel tasks (per constituent space) that are dynamically 5324 // claimed by the parallel threads. 5325 cms_space->initialize_sequential_subtasks_for_rescan(n_workers); 5326 perm_space->initialize_sequential_subtasks_for_rescan(n_workers); 5327 5328 // It turns out that even when we're using 1 thread, doing the work in a 5329 // separate thread causes wide variance in run times. We can't help this 5330 // in the multi-threaded case, but we special-case n=1 here to get 5331 // repeatable measurements of the 1-thread overhead of the parallel code. 5332 if (n_workers > 1) { 5333 // Make refs discovery MT-safe 5334 ReferenceProcessorMTMutator mt(ref_processor(), true); 5335 workers->run_task(&tsk); 5336 } else { 5337 tsk.work(0); 5338 } 5339 gch->set_par_threads(0); // 0 ==> non-parallel. 5340 // restore, single-threaded for now, any preserved marks 5341 // as a result of work_q overflow 5342 restore_preserved_marks_if_any(); 5343 } 5344 5345 // Non-parallel version of remark 5346 void CMSCollector::do_remark_non_parallel() { 5347 ResourceMark rm; 5348 HandleMark hm; 5349 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5350 MarkRefsIntoAndScanClosure 5351 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable, 5352 &_markStack, &_revisitStack, this, 5353 false /* should_yield */, false /* not precleaning */); 5354 MarkFromDirtyCardsClosure 5355 markFromDirtyCardsClosure(this, _span, 5356 NULL, // space is set further below 5357 &_markBitMap, &_markStack, &_revisitStack, 5358 &mrias_cl); 5359 { 5360 TraceTime t("grey object rescan", PrintGCDetails, false, gclog_or_tty); 5361 // Iterate over the dirty cards, marking them precleaned, and 5362 // setting the corresponding bits in the mod union table. 5363 { 5364 ModUnionClosure modUnionClosure(&_modUnionTable); 5365 _ct->ct_bs()->dirty_card_iterate( 5366 _cmsGen->used_region(), 5367 &modUnionClosure); 5368 _ct->ct_bs()->dirty_card_iterate( 5369 _permGen->used_region(), 5370 &modUnionClosure); 5371 } 5372 // Having transferred these marks into the modUnionTable, we just need 5373 // to rescan the marked objects on the dirty cards in the modUnionTable. 5374 // The initial marking may have been done during an asynchronous 5375 // collection so there may be dirty bits in the mod-union table. 5376 const int alignment = 5377 CardTableModRefBS::card_size * BitsPerWord; 5378 { 5379 // ... First handle dirty cards in CMS gen 5380 markFromDirtyCardsClosure.set_space(_cmsGen->cmsSpace()); 5381 MemRegion ur = _cmsGen->used_region(); 5382 HeapWord* lb = ur.start(); 5383 HeapWord* ub = (HeapWord*)round_to((intptr_t)ur.end(), alignment); 5384 MemRegion cms_span(lb, ub); 5385 _modUnionTable.dirty_range_iterate_clear(cms_span, 5386 &markFromDirtyCardsClosure); 5387 verify_work_stacks_empty(); 5388 if (PrintCMSStatistics != 0) { 5389 gclog_or_tty->print(" (re-scanned "SIZE_FORMAT" dirty cards in cms gen) ", 5390 markFromDirtyCardsClosure.num_dirty_cards()); 5391 } 5392 } 5393 { 5394 // .. and then repeat for dirty cards in perm gen 5395 markFromDirtyCardsClosure.set_space(_permGen->cmsSpace()); 5396 MemRegion ur = _permGen->used_region(); 5397 HeapWord* lb = ur.start(); 5398 HeapWord* ub = (HeapWord*)round_to((intptr_t)ur.end(), alignment); 5399 MemRegion perm_span(lb, ub); 5400 _modUnionTable.dirty_range_iterate_clear(perm_span, 5401 &markFromDirtyCardsClosure); 5402 verify_work_stacks_empty(); 5403 if (PrintCMSStatistics != 0) { 5404 gclog_or_tty->print(" (re-scanned "SIZE_FORMAT" dirty cards in perm gen) ", 5405 markFromDirtyCardsClosure.num_dirty_cards()); 5406 } 5407 } 5408 } 5409 if (VerifyDuringGC && 5410 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 5411 HandleMark hm; // Discard invalid handles created during verification 5412 Universe::verify(true); 5413 } 5414 { 5415 TraceTime t("root rescan", PrintGCDetails, false, gclog_or_tty); 5416 5417 verify_work_stacks_empty(); 5418 5419 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 5420 gch->gen_process_strong_roots(_cmsGen->level(), 5421 true, // younger gens as roots 5422 true, // collecting perm gen 5423 SharedHeap::ScanningOption(roots_scanning_options()), 5424 NULL, &mrias_cl); 5425 } 5426 verify_work_stacks_empty(); 5427 // Restore evacuated mark words, if any, used for overflow list links 5428 if (!CMSOverflowEarlyRestoration) { 5429 restore_preserved_marks_if_any(); 5430 } 5431 verify_overflow_empty(); 5432 } 5433 5434 //////////////////////////////////////////////////////// 5435 // Parallel Reference Processing Task Proxy Class 5436 //////////////////////////////////////////////////////// 5437 class CMSRefProcTaskProxy: public AbstractGangTask { 5438 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask; 5439 CMSCollector* _collector; 5440 CMSBitMap* _mark_bit_map; 5441 MemRegion _span; 5442 OopTaskQueueSet* _task_queues; 5443 ParallelTaskTerminator _term; 5444 ProcessTask& _task; 5445 5446 public: 5447 CMSRefProcTaskProxy(ProcessTask& task, 5448 CMSCollector* collector, 5449 const MemRegion& span, 5450 CMSBitMap* mark_bit_map, 5451 int total_workers, 5452 OopTaskQueueSet* task_queues): 5453 AbstractGangTask("Process referents by policy in parallel"), 5454 _task(task), 5455 _collector(collector), _span(span), _mark_bit_map(mark_bit_map), 5456 _task_queues(task_queues), 5457 _term(total_workers, task_queues) 5458 { } 5459 5460 OopTaskQueueSet* task_queues() { return _task_queues; } 5461 5462 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 5463 5464 ParallelTaskTerminator* terminator() { return &_term; } 5465 5466 void do_work_steal(int i, 5467 CMSParDrainMarkingStackClosure* drain, 5468 CMSParKeepAliveClosure* keep_alive, 5469 int* seed); 5470 5471 virtual void work(int i); 5472 }; 5473 5474 void CMSRefProcTaskProxy::work(int i) { 5475 CMSParKeepAliveClosure par_keep_alive(_collector, _span, 5476 _mark_bit_map, work_queue(i)); 5477 CMSParDrainMarkingStackClosure par_drain_stack(_collector, _span, 5478 _mark_bit_map, work_queue(i)); 5479 CMSIsAliveClosure is_alive_closure(_mark_bit_map); 5480 _task.work(i, is_alive_closure, par_keep_alive, par_drain_stack); 5481 if (_task.marks_oops_alive()) { 5482 do_work_steal(i, &par_drain_stack, &par_keep_alive, 5483 _collector->hash_seed(i)); 5484 } 5485 assert(work_queue(i)->size() == 0, "work_queue should be empty"); 5486 assert(_collector->_overflow_list == NULL, "non-empty _overflow_list"); 5487 } 5488 5489 class CMSRefEnqueueTaskProxy: public AbstractGangTask { 5490 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask; 5491 EnqueueTask& _task; 5492 5493 public: 5494 CMSRefEnqueueTaskProxy(EnqueueTask& task) 5495 : AbstractGangTask("Enqueue reference objects in parallel"), 5496 _task(task) 5497 { } 5498 5499 virtual void work(int i) 5500 { 5501 _task.work(i); 5502 } 5503 }; 5504 5505 CMSParKeepAliveClosure::CMSParKeepAliveClosure(CMSCollector* collector, 5506 MemRegion span, CMSBitMap* bit_map, OopTaskQueue* work_queue): 5507 _collector(collector), 5508 _span(span), 5509 _bit_map(bit_map), 5510 _work_queue(work_queue), 5511 _mark_and_push(collector, span, bit_map, work_queue), 5512 _low_water_mark(MIN2((uint)(work_queue->max_elems()/4), 5513 (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads))) 5514 { } 5515 5516 // . see if we can share work_queues with ParNew? XXX 5517 void CMSRefProcTaskProxy::do_work_steal(int i, 5518 CMSParDrainMarkingStackClosure* drain, 5519 CMSParKeepAliveClosure* keep_alive, 5520 int* seed) { 5521 OopTaskQueue* work_q = work_queue(i); 5522 NOT_PRODUCT(int num_steals = 0;) 5523 oop obj_to_scan; 5524 size_t num_from_overflow_list = 5525 MIN2((size_t)work_q->max_elems()/4, 5526 (size_t)ParGCDesiredObjsFromOverflowList); 5527 5528 while (true) { 5529 // Completely finish any left over work from (an) earlier round(s) 5530 drain->trim_queue(0); 5531 // Now check if there's any work in the overflow list 5532 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 5533 work_q)) { 5534 // Found something in global overflow list; 5535 // not yet ready to go stealing work from others. 5536 // We'd like to assert(work_q->size() != 0, ...) 5537 // because we just took work from the overflow list, 5538 // but of course we can't, since all of that might have 5539 // been already stolen from us. 5540 continue; 5541 } 5542 // Verify that we have no work before we resort to stealing 5543 assert(work_q->size() == 0, "Have work, shouldn't steal"); 5544 // Try to steal from other queues that have work 5545 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 5546 NOT_PRODUCT(num_steals++;) 5547 assert(obj_to_scan->is_oop(), "Oops, not an oop!"); 5548 assert(_mark_bit_map->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 5549 // Do scanning work 5550 obj_to_scan->oop_iterate(keep_alive); 5551 // Loop around, finish this work, and try to steal some more 5552 } else if (terminator()->offer_termination()) { 5553 break; // nirvana from the infinite cycle 5554 } 5555 } 5556 NOT_PRODUCT( 5557 if (PrintCMSStatistics != 0) { 5558 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals); 5559 } 5560 ) 5561 } 5562 5563 void CMSRefProcTaskExecutor::execute(ProcessTask& task) 5564 { 5565 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5566 WorkGang* workers = gch->workers(); 5567 assert(workers != NULL, "Need parallel worker threads."); 5568 int n_workers = workers->total_workers(); 5569 CMSRefProcTaskProxy rp_task(task, &_collector, 5570 _collector.ref_processor()->span(), 5571 _collector.markBitMap(), 5572 n_workers, _collector.task_queues()); 5573 workers->run_task(&rp_task); 5574 } 5575 5576 void CMSRefProcTaskExecutor::execute(EnqueueTask& task) 5577 { 5578 5579 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5580 WorkGang* workers = gch->workers(); 5581 assert(workers != NULL, "Need parallel worker threads."); 5582 CMSRefEnqueueTaskProxy enq_task(task); 5583 workers->run_task(&enq_task); 5584 } 5585 5586 void CMSCollector::refProcessingWork(bool asynch, bool clear_all_soft_refs) { 5587 5588 ResourceMark rm; 5589 HandleMark hm; 5590 ReferencePolicy* soft_ref_policy; 5591 5592 assert(!ref_processor()->enqueuing_is_done(), "Enqueuing should not be complete"); 5593 // Process weak references. 5594 if (clear_all_soft_refs) { 5595 soft_ref_policy = new AlwaysClearPolicy(); 5596 } else { 5597 #ifdef COMPILER2 5598 soft_ref_policy = new LRUMaxHeapPolicy(); 5599 #else 5600 soft_ref_policy = new LRUCurrentHeapPolicy(); 5601 #endif // COMPILER2 5602 } 5603 verify_work_stacks_empty(); 5604 5605 ReferenceProcessor* rp = ref_processor(); 5606 assert(rp->span().equals(_span), "Spans should be equal"); 5607 CMSKeepAliveClosure cmsKeepAliveClosure(this, _span, &_markBitMap, 5608 &_markStack); 5609 CMSDrainMarkingStackClosure cmsDrainMarkingStackClosure(this, 5610 _span, &_markBitMap, &_markStack, 5611 &cmsKeepAliveClosure); 5612 { 5613 TraceTime t("weak refs processing", PrintGCDetails, false, gclog_or_tty); 5614 if (rp->processing_is_mt()) { 5615 CMSRefProcTaskExecutor task_executor(*this); 5616 rp->process_discovered_references(soft_ref_policy, 5617 &_is_alive_closure, 5618 &cmsKeepAliveClosure, 5619 &cmsDrainMarkingStackClosure, 5620 &task_executor); 5621 } else { 5622 rp->process_discovered_references(soft_ref_policy, 5623 &_is_alive_closure, 5624 &cmsKeepAliveClosure, 5625 &cmsDrainMarkingStackClosure, 5626 NULL); 5627 } 5628 verify_work_stacks_empty(); 5629 } 5630 5631 if (cms_should_unload_classes()) { 5632 { 5633 TraceTime t("class unloading", PrintGCDetails, false, gclog_or_tty); 5634 5635 // Follow SystemDictionary roots and unload classes 5636 bool purged_class = SystemDictionary::do_unloading(&_is_alive_closure); 5637 5638 // Follow CodeCache roots and unload any methods marked for unloading 5639 CodeCache::do_unloading(&_is_alive_closure, 5640 &cmsKeepAliveClosure, 5641 purged_class); 5642 5643 cmsDrainMarkingStackClosure.do_void(); 5644 verify_work_stacks_empty(); 5645 5646 // Update subklass/sibling/implementor links in KlassKlass descendants 5647 assert(!_revisitStack.isEmpty(), "revisit stack should not be empty"); 5648 oop k; 5649 while ((k = _revisitStack.pop()) != NULL) { 5650 ((Klass*)(oopDesc*)k)->follow_weak_klass_links( 5651 &_is_alive_closure, 5652 &cmsKeepAliveClosure); 5653 } 5654 assert(!ClassUnloading || 5655 (_markStack.isEmpty() && overflow_list_is_empty()), 5656 "Should not have found new reachable objects"); 5657 assert(_revisitStack.isEmpty(), "revisit stack should have been drained"); 5658 cmsDrainMarkingStackClosure.do_void(); 5659 verify_work_stacks_empty(); 5660 } 5661 5662 { 5663 TraceTime t("scrub symbol & string tables", PrintGCDetails, false, gclog_or_tty); 5664 // Now clean up stale oops in SymbolTable and StringTable 5665 SymbolTable::unlink(&_is_alive_closure); 5666 StringTable::unlink(&_is_alive_closure); 5667 } 5668 } 5669 5670 verify_work_stacks_empty(); 5671 // Restore any preserved marks as a result of mark stack or 5672 // work queue overflow 5673 restore_preserved_marks_if_any(); // done single-threaded for now 5674 5675 rp->set_enqueuing_is_done(true); 5676 if (rp->processing_is_mt()) { 5677 CMSRefProcTaskExecutor task_executor(*this); 5678 rp->enqueue_discovered_references(&task_executor); 5679 } else { 5680 rp->enqueue_discovered_references(NULL); 5681 } 5682 rp->verify_no_references_recorded(); 5683 assert(!rp->discovery_enabled(), "should have been disabled"); 5684 5685 // JVMTI object tagging is based on JNI weak refs. If any of these 5686 // refs were cleared then JVMTI needs to update its maps and 5687 // maybe post ObjectFrees to agents. 5688 JvmtiExport::cms_ref_processing_epilogue(); 5689 } 5690 5691 #ifndef PRODUCT 5692 void CMSCollector::check_correct_thread_executing() { 5693 Thread* t = Thread::current(); 5694 // Only the VM thread or the CMS thread should be here. 5695 assert(t->is_ConcurrentGC_thread() || t->is_VM_thread(), 5696 "Unexpected thread type"); 5697 // If this is the vm thread, the foreground process 5698 // should not be waiting. Note that _foregroundGCIsActive is 5699 // true while the foreground collector is waiting. 5700 if (_foregroundGCShouldWait) { 5701 // We cannot be the VM thread 5702 assert(t->is_ConcurrentGC_thread(), 5703 "Should be CMS thread"); 5704 } else { 5705 // We can be the CMS thread only if we are in a stop-world 5706 // phase of CMS collection. 5707 if (t->is_ConcurrentGC_thread()) { 5708 assert(_collectorState == InitialMarking || 5709 _collectorState == FinalMarking, 5710 "Should be a stop-world phase"); 5711 // The CMS thread should be holding the CMS_token. 5712 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 5713 "Potential interference with concurrently " 5714 "executing VM thread"); 5715 } 5716 } 5717 } 5718 #endif 5719 5720 void CMSCollector::sweep(bool asynch) { 5721 assert(_collectorState == Sweeping, "just checking"); 5722 check_correct_thread_executing(); 5723 verify_work_stacks_empty(); 5724 verify_overflow_empty(); 5725 incrementSweepCount(); 5726 _sweep_timer.stop(); 5727 _sweep_estimate.sample(_sweep_timer.seconds()); 5728 size_policy()->avg_cms_free_at_sweep()->sample(_cmsGen->free()); 5729 5730 // PermGen verification support: If perm gen sweeping is disabled in 5731 // this cycle, we preserve the perm gen object "deadness" information 5732 // in the perm_gen_verify_bit_map. In order to do that we traverse 5733 // all blocks in perm gen and mark all dead objects. 5734 if (verifying() && !cms_should_unload_classes()) { 5735 CMSTokenSyncWithLocks ts(true, _permGen->freelistLock(), 5736 bitMapLock()); 5737 assert(perm_gen_verify_bit_map()->sizeInBits() != 0, 5738 "Should have already been allocated"); 5739 MarkDeadObjectsClosure mdo(this, _permGen->cmsSpace(), 5740 markBitMap(), perm_gen_verify_bit_map()); 5741 _permGen->cmsSpace()->blk_iterate(&mdo); 5742 } 5743 5744 if (asynch) { 5745 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 5746 CMSPhaseAccounting pa(this, "sweep", !PrintGCDetails); 5747 // First sweep the old gen then the perm gen 5748 { 5749 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(), 5750 bitMapLock()); 5751 sweepWork(_cmsGen, asynch); 5752 } 5753 5754 // Now repeat for perm gen 5755 if (cms_should_unload_classes()) { 5756 CMSTokenSyncWithLocks ts(true, _permGen->freelistLock(), 5757 bitMapLock()); 5758 sweepWork(_permGen, asynch); 5759 } 5760 5761 // Update Universe::_heap_*_at_gc figures. 5762 // We need all the free list locks to make the abstract state 5763 // transition from Sweeping to Resetting. See detailed note 5764 // further below. 5765 { 5766 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(), 5767 _permGen->freelistLock()); 5768 // Update heap occupancy information which is used as 5769 // input to soft ref clearing policy at the next gc. 5770 Universe::update_heap_info_at_gc(); 5771 _collectorState = Resizing; 5772 } 5773 } else { 5774 // already have needed locks 5775 sweepWork(_cmsGen, asynch); 5776 5777 if (cms_should_unload_classes()) { 5778 sweepWork(_permGen, asynch); 5779 } 5780 // Update heap occupancy information which is used as 5781 // input to soft ref clearing policy at the next gc. 5782 Universe::update_heap_info_at_gc(); 5783 _collectorState = Resizing; 5784 } 5785 verify_work_stacks_empty(); 5786 verify_overflow_empty(); 5787 5788 _sweep_timer.reset(); 5789 _sweep_timer.start(); 5790 5791 update_time_of_last_gc(os::javaTimeMillis()); 5792 5793 // NOTE on abstract state transitions: 5794 // Mutators allocate-live and/or mark the mod-union table dirty 5795 // based on the state of the collection. The former is done in 5796 // the interval [Marking, Sweeping] and the latter in the interval 5797 // [Marking, Sweeping). Thus the transitions into the Marking state 5798 // and out of the Sweeping state must be synchronously visible 5799 // globally to the mutators. 5800 // The transition into the Marking state happens with the world 5801 // stopped so the mutators will globally see it. Sweeping is 5802 // done asynchronously by the background collector so the transition 5803 // from the Sweeping state to the Resizing state must be done 5804 // under the freelistLock (as is the check for whether to 5805 // allocate-live and whether to dirty the mod-union table). 5806 assert(_collectorState == Resizing, "Change of collector state to" 5807 " Resizing must be done under the freelistLocks (plural)"); 5808 5809 // Now that sweeping has been completed, if the GCH's 5810 // incremental_collection_will_fail flag is set, clear it, 5811 // thus inviting a younger gen collection to promote into 5812 // this generation. If such a promotion may still fail, 5813 // the flag will be set again when a young collection is 5814 // attempted. 5815 // I think the incremental_collection_will_fail flag's use 5816 // is specific to a 2 generation collection policy, so i'll 5817 // assert that that's the configuration we are operating within. 5818 // The use of the flag can and should be generalized appropriately 5819 // in the future to deal with a general n-generation system. 5820 5821 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5822 assert(gch->collector_policy()->is_two_generation_policy(), 5823 "Resetting of incremental_collection_will_fail flag" 5824 " may be incorrect otherwise"); 5825 gch->clear_incremental_collection_will_fail(); 5826 gch->update_full_collections_completed(_collection_count_start); 5827 } 5828 5829 // FIX ME!!! Looks like this belongs in CFLSpace, with 5830 // CMSGen merely delegating to it. 5831 void ConcurrentMarkSweepGeneration::setNearLargestChunk() { 5832 double nearLargestPercent = 0.999; 5833 HeapWord* minAddr = _cmsSpace->bottom(); 5834 HeapWord* largestAddr = 5835 (HeapWord*) _cmsSpace->dictionary()->findLargestDict(); 5836 if (largestAddr == 0) { 5837 // The dictionary appears to be empty. In this case 5838 // try to coalesce at the end of the heap. 5839 largestAddr = _cmsSpace->end(); 5840 } 5841 size_t largestOffset = pointer_delta(largestAddr, minAddr); 5842 size_t nearLargestOffset = 5843 (size_t)((double)largestOffset * nearLargestPercent) - MinChunkSize; 5844 _cmsSpace->set_nearLargestChunk(minAddr + nearLargestOffset); 5845 } 5846 5847 bool ConcurrentMarkSweepGeneration::isNearLargestChunk(HeapWord* addr) { 5848 return addr >= _cmsSpace->nearLargestChunk(); 5849 } 5850 5851 FreeChunk* ConcurrentMarkSweepGeneration::find_chunk_at_end() { 5852 return _cmsSpace->find_chunk_at_end(); 5853 } 5854 5855 void ConcurrentMarkSweepGeneration::update_gc_stats(int current_level, 5856 bool full) { 5857 // The next lower level has been collected. Gather any statistics 5858 // that are of interest at this point. 5859 if (!full && (current_level + 1) == level()) { 5860 // Gather statistics on the young generation collection. 5861 collector()->stats().record_gc0_end(used()); 5862 } 5863 } 5864 5865 CMSAdaptiveSizePolicy* ConcurrentMarkSweepGeneration::size_policy() { 5866 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5867 assert(gch->kind() == CollectedHeap::GenCollectedHeap, 5868 "Wrong type of heap"); 5869 CMSAdaptiveSizePolicy* sp = (CMSAdaptiveSizePolicy*) 5870 gch->gen_policy()->size_policy(); 5871 assert(sp->is_gc_cms_adaptive_size_policy(), 5872 "Wrong type of size policy"); 5873 return sp; 5874 } 5875 5876 void ConcurrentMarkSweepGeneration::rotate_debug_collection_type() { 5877 if (PrintGCDetails && Verbose) { 5878 gclog_or_tty->print("Rotate from %d ", _debug_collection_type); 5879 } 5880 _debug_collection_type = (CollectionTypes) (_debug_collection_type + 1); 5881 _debug_collection_type = 5882 (CollectionTypes) (_debug_collection_type % Unknown_collection_type); 5883 if (PrintGCDetails && Verbose) { 5884 gclog_or_tty->print_cr("to %d ", _debug_collection_type); 5885 } 5886 } 5887 5888 void CMSCollector::sweepWork(ConcurrentMarkSweepGeneration* gen, 5889 bool asynch) { 5890 // We iterate over the space(s) underlying this generation, 5891 // checking the mark bit map to see if the bits corresponding 5892 // to specific blocks are marked or not. Blocks that are 5893 // marked are live and are not swept up. All remaining blocks 5894 // are swept up, with coalescing on-the-fly as we sweep up 5895 // contiguous free and/or garbage blocks: 5896 // We need to ensure that the sweeper synchronizes with allocators 5897 // and stop-the-world collectors. In particular, the following 5898 // locks are used: 5899 // . CMS token: if this is held, a stop the world collection cannot occur 5900 // . freelistLock: if this is held no allocation can occur from this 5901 // generation by another thread 5902 // . bitMapLock: if this is held, no other thread can access or update 5903 // 5904 5905 // Note that we need to hold the freelistLock if we use 5906 // block iterate below; else the iterator might go awry if 5907 // a mutator (or promotion) causes block contents to change 5908 // (for instance if the allocator divvies up a block). 5909 // If we hold the free list lock, for all practical purposes 5910 // young generation GC's can't occur (they'll usually need to 5911 // promote), so we might as well prevent all young generation 5912 // GC's while we do a sweeping step. For the same reason, we might 5913 // as well take the bit map lock for the entire duration 5914 5915 // check that we hold the requisite locks 5916 assert(have_cms_token(), "Should hold cms token"); 5917 assert( (asynch && ConcurrentMarkSweepThread::cms_thread_has_cms_token()) 5918 || (!asynch && ConcurrentMarkSweepThread::vm_thread_has_cms_token()), 5919 "Should possess CMS token to sweep"); 5920 assert_lock_strong(gen->freelistLock()); 5921 assert_lock_strong(bitMapLock()); 5922 5923 assert(!_sweep_timer.is_active(), "Was switched off in an outer context"); 5924 gen->cmsSpace()->beginSweepFLCensus((float)(_sweep_timer.seconds()), 5925 _sweep_estimate.padded_average()); 5926 gen->setNearLargestChunk(); 5927 5928 { 5929 SweepClosure sweepClosure(this, gen, &_markBitMap, 5930 CMSYield && asynch); 5931 gen->cmsSpace()->blk_iterate_careful(&sweepClosure); 5932 // We need to free-up/coalesce garbage/blocks from a 5933 // co-terminal free run. This is done in the SweepClosure 5934 // destructor; so, do not remove this scope, else the 5935 // end-of-sweep-census below will be off by a little bit. 5936 } 5937 gen->cmsSpace()->sweep_completed(); 5938 gen->cmsSpace()->endSweepFLCensus(sweepCount()); 5939 } 5940 5941 // Reset CMS data structures (for now just the marking bit map) 5942 // preparatory for the next cycle. 5943 void CMSCollector::reset(bool asynch) { 5944 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5945 CMSAdaptiveSizePolicy* sp = size_policy(); 5946 AdaptiveSizePolicyOutput(sp, gch->total_collections()); 5947 if (asynch) { 5948 CMSTokenSyncWithLocks ts(true, bitMapLock()); 5949 5950 // If the state is not "Resetting", the foreground thread 5951 // has done a collection and the resetting. 5952 if (_collectorState != Resetting) { 5953 assert(_collectorState == Idling, "The state should only change" 5954 " because the foreground collector has finished the collection"); 5955 return; 5956 } 5957 5958 // Clear the mark bitmap (no grey objects to start with) 5959 // for the next cycle. 5960 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 5961 CMSPhaseAccounting cmspa(this, "reset", !PrintGCDetails); 5962 5963 HeapWord* curAddr = _markBitMap.startWord(); 5964 while (curAddr < _markBitMap.endWord()) { 5965 size_t remaining = pointer_delta(_markBitMap.endWord(), curAddr); 5966 MemRegion chunk(curAddr, MIN2(CMSBitMapYieldQuantum, remaining)); 5967 _markBitMap.clear_large_range(chunk); 5968 if (ConcurrentMarkSweepThread::should_yield() && 5969 !foregroundGCIsActive() && 5970 CMSYield) { 5971 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 5972 "CMS thread should hold CMS token"); 5973 assert_lock_strong(bitMapLock()); 5974 bitMapLock()->unlock(); 5975 ConcurrentMarkSweepThread::desynchronize(true); 5976 ConcurrentMarkSweepThread::acknowledge_yield_request(); 5977 stopTimer(); 5978 if (PrintCMSStatistics != 0) { 5979 incrementYields(); 5980 } 5981 icms_wait(); 5982 5983 // See the comment in coordinator_yield() 5984 for (unsigned i = 0; i < CMSYieldSleepCount && 5985 ConcurrentMarkSweepThread::should_yield() && 5986 !CMSCollector::foregroundGCIsActive(); ++i) { 5987 os::sleep(Thread::current(), 1, false); 5988 ConcurrentMarkSweepThread::acknowledge_yield_request(); 5989 } 5990 5991 ConcurrentMarkSweepThread::synchronize(true); 5992 bitMapLock()->lock_without_safepoint_check(); 5993 startTimer(); 5994 } 5995 curAddr = chunk.end(); 5996 } 5997 _collectorState = Idling; 5998 } else { 5999 // already have the lock 6000 assert(_collectorState == Resetting, "just checking"); 6001 assert_lock_strong(bitMapLock()); 6002 _markBitMap.clear_all(); 6003 _collectorState = Idling; 6004 } 6005 6006 // Stop incremental mode after a cycle completes, so that any future cycles 6007 // are triggered by allocation. 6008 stop_icms(); 6009 6010 NOT_PRODUCT( 6011 if (RotateCMSCollectionTypes) { 6012 _cmsGen->rotate_debug_collection_type(); 6013 } 6014 ) 6015 } 6016 6017 void CMSCollector::do_CMS_operation(CMS_op_type op) { 6018 gclog_or_tty->date_stamp(PrintGC && PrintGCDateStamps); 6019 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 6020 TraceTime t("GC", PrintGC, !PrintGCDetails, gclog_or_tty); 6021 TraceCollectorStats tcs(counters()); 6022 6023 switch (op) { 6024 case CMS_op_checkpointRootsInitial: { 6025 checkpointRootsInitial(true); // asynch 6026 if (PrintGC) { 6027 _cmsGen->printOccupancy("initial-mark"); 6028 } 6029 break; 6030 } 6031 case CMS_op_checkpointRootsFinal: { 6032 checkpointRootsFinal(true, // asynch 6033 false, // !clear_all_soft_refs 6034 false); // !init_mark_was_synchronous 6035 if (PrintGC) { 6036 _cmsGen->printOccupancy("remark"); 6037 } 6038 break; 6039 } 6040 default: 6041 fatal("No such CMS_op"); 6042 } 6043 } 6044 6045 #ifndef PRODUCT 6046 size_t const CMSCollector::skip_header_HeapWords() { 6047 return FreeChunk::header_size(); 6048 } 6049 6050 // Try and collect here conditions that should hold when 6051 // CMS thread is exiting. The idea is that the foreground GC 6052 // thread should not be blocked if it wants to terminate 6053 // the CMS thread and yet continue to run the VM for a while 6054 // after that. 6055 void CMSCollector::verify_ok_to_terminate() const { 6056 assert(Thread::current()->is_ConcurrentGC_thread(), 6057 "should be called by CMS thread"); 6058 assert(!_foregroundGCShouldWait, "should be false"); 6059 // We could check here that all the various low-level locks 6060 // are not held by the CMS thread, but that is overkill; see 6061 // also CMSThread::verify_ok_to_terminate() where the CGC_lock 6062 // is checked. 6063 } 6064 #endif 6065 6066 size_t CMSCollector::block_size_using_printezis_bits(HeapWord* addr) const { 6067 assert(_markBitMap.isMarked(addr) && _markBitMap.isMarked(addr + 1), 6068 "missing Printezis mark?"); 6069 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 6070 size_t size = pointer_delta(nextOneAddr + 1, addr); 6071 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 6072 "alignment problem"); 6073 assert(size >= 3, "Necessary for Printezis marks to work"); 6074 return size; 6075 } 6076 6077 // A variant of the above (block_size_using_printezis_bits()) except 6078 // that we return 0 if the P-bits are not yet set. 6079 size_t CMSCollector::block_size_if_printezis_bits(HeapWord* addr) const { 6080 if (_markBitMap.isMarked(addr)) { 6081 assert(_markBitMap.isMarked(addr + 1), "Missing Printezis bit?"); 6082 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 6083 size_t size = pointer_delta(nextOneAddr + 1, addr); 6084 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 6085 "alignment problem"); 6086 assert(size >= 3, "Necessary for Printezis marks to work"); 6087 return size; 6088 } else { 6089 assert(!_markBitMap.isMarked(addr + 1), "Bit map inconsistency?"); 6090 return 0; 6091 } 6092 } 6093 6094 HeapWord* CMSCollector::next_card_start_after_block(HeapWord* addr) const { 6095 size_t sz = 0; 6096 oop p = (oop)addr; 6097 if (p->klass() != NULL && p->is_parsable()) { 6098 sz = CompactibleFreeListSpace::adjustObjectSize(p->size()); 6099 } else { 6100 sz = block_size_using_printezis_bits(addr); 6101 } 6102 assert(sz > 0, "size must be nonzero"); 6103 HeapWord* next_block = addr + sz; 6104 HeapWord* next_card = (HeapWord*)round_to((uintptr_t)next_block, 6105 CardTableModRefBS::card_size); 6106 assert(round_down((uintptr_t)addr, CardTableModRefBS::card_size) < 6107 round_down((uintptr_t)next_card, CardTableModRefBS::card_size), 6108 "must be different cards"); 6109 return next_card; 6110 } 6111 6112 6113 // CMS Bit Map Wrapper ///////////////////////////////////////// 6114 6115 // Construct a CMS bit map infrastructure, but don't create the 6116 // bit vector itself. That is done by a separate call CMSBitMap::allocate() 6117 // further below. 6118 CMSBitMap::CMSBitMap(int shifter, int mutex_rank, const char* mutex_name): 6119 _bm(NULL,0), 6120 _shifter(shifter), 6121 _lock(mutex_rank >= 0 ? new Mutex(mutex_rank, mutex_name, true) : NULL) 6122 { 6123 _bmStartWord = 0; 6124 _bmWordSize = 0; 6125 } 6126 6127 bool CMSBitMap::allocate(MemRegion mr) { 6128 _bmStartWord = mr.start(); 6129 _bmWordSize = mr.word_size(); 6130 ReservedSpace brs(ReservedSpace::allocation_align_size_up( 6131 (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1)); 6132 if (!brs.is_reserved()) { 6133 warning("CMS bit map allocation failure"); 6134 return false; 6135 } 6136 // For now we'll just commit all of the bit map up fromt. 6137 // Later on we'll try to be more parsimonious with swap. 6138 if (!_virtual_space.initialize(brs, brs.size())) { 6139 warning("CMS bit map backing store failure"); 6140 return false; 6141 } 6142 assert(_virtual_space.committed_size() == brs.size(), 6143 "didn't reserve backing store for all of CMS bit map?"); 6144 _bm.set_map((uintptr_t*)_virtual_space.low()); 6145 assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >= 6146 _bmWordSize, "inconsistency in bit map sizing"); 6147 _bm.set_size(_bmWordSize >> _shifter); 6148 6149 // bm.clear(); // can we rely on getting zero'd memory? verify below 6150 assert(isAllClear(), 6151 "Expected zero'd memory from ReservedSpace constructor"); 6152 assert(_bm.size() == heapWordDiffToOffsetDiff(sizeInWords()), 6153 "consistency check"); 6154 return true; 6155 } 6156 6157 void CMSBitMap::dirty_range_iterate_clear(MemRegion mr, MemRegionClosure* cl) { 6158 HeapWord *next_addr, *end_addr, *last_addr; 6159 assert_locked(); 6160 assert(covers(mr), "out-of-range error"); 6161 // XXX assert that start and end are appropriately aligned 6162 for (next_addr = mr.start(), end_addr = mr.end(); 6163 next_addr < end_addr; next_addr = last_addr) { 6164 MemRegion dirty_region = getAndClearMarkedRegion(next_addr, end_addr); 6165 last_addr = dirty_region.end(); 6166 if (!dirty_region.is_empty()) { 6167 cl->do_MemRegion(dirty_region); 6168 } else { 6169 assert(last_addr == end_addr, "program logic"); 6170 return; 6171 } 6172 } 6173 } 6174 6175 #ifndef PRODUCT 6176 void CMSBitMap::assert_locked() const { 6177 CMSLockVerifier::assert_locked(lock()); 6178 } 6179 6180 bool CMSBitMap::covers(MemRegion mr) const { 6181 // assert(_bm.map() == _virtual_space.low(), "map inconsistency"); 6182 assert((size_t)_bm.size() == (_bmWordSize >> _shifter), 6183 "size inconsistency"); 6184 return (mr.start() >= _bmStartWord) && 6185 (mr.end() <= endWord()); 6186 } 6187 6188 bool CMSBitMap::covers(HeapWord* start, size_t size) const { 6189 return (start >= _bmStartWord && (start + size) <= endWord()); 6190 } 6191 6192 void CMSBitMap::verifyNoOneBitsInRange(HeapWord* left, HeapWord* right) { 6193 // verify that there are no 1 bits in the interval [left, right) 6194 FalseBitMapClosure falseBitMapClosure; 6195 iterate(&falseBitMapClosure, left, right); 6196 } 6197 6198 void CMSBitMap::region_invariant(MemRegion mr) 6199 { 6200 assert_locked(); 6201 // mr = mr.intersection(MemRegion(_bmStartWord, _bmWordSize)); 6202 assert(!mr.is_empty(), "unexpected empty region"); 6203 assert(covers(mr), "mr should be covered by bit map"); 6204 // convert address range into offset range 6205 size_t start_ofs = heapWordToOffset(mr.start()); 6206 // Make sure that end() is appropriately aligned 6207 assert(mr.end() == (HeapWord*)round_to((intptr_t)mr.end(), 6208 (1 << (_shifter+LogHeapWordSize))), 6209 "Misaligned mr.end()"); 6210 size_t end_ofs = heapWordToOffset(mr.end()); 6211 assert(end_ofs > start_ofs, "Should mark at least one bit"); 6212 } 6213 6214 #endif 6215 6216 bool CMSMarkStack::allocate(size_t size) { 6217 // allocate a stack of the requisite depth 6218 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 6219 size * sizeof(oop))); 6220 if (!rs.is_reserved()) { 6221 warning("CMSMarkStack allocation failure"); 6222 return false; 6223 } 6224 if (!_virtual_space.initialize(rs, rs.size())) { 6225 warning("CMSMarkStack backing store failure"); 6226 return false; 6227 } 6228 assert(_virtual_space.committed_size() == rs.size(), 6229 "didn't reserve backing store for all of CMS stack?"); 6230 _base = (oop*)(_virtual_space.low()); 6231 _index = 0; 6232 _capacity = size; 6233 NOT_PRODUCT(_max_depth = 0); 6234 return true; 6235 } 6236 6237 // XXX FIX ME !!! In the MT case we come in here holding a 6238 // leaf lock. For printing we need to take a further lock 6239 // which has lower rank. We need to recallibrate the two 6240 // lock-ranks involved in order to be able to rpint the 6241 // messages below. (Or defer the printing to the caller. 6242 // For now we take the expedient path of just disabling the 6243 // messages for the problematic case.) 6244 void CMSMarkStack::expand() { 6245 assert(_capacity <= CMSMarkStackSizeMax, "stack bigger than permitted"); 6246 if (_capacity == CMSMarkStackSizeMax) { 6247 if (_hit_limit++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) { 6248 // We print a warning message only once per CMS cycle. 6249 gclog_or_tty->print_cr(" (benign) Hit CMSMarkStack max size limit"); 6250 } 6251 return; 6252 } 6253 // Double capacity if possible 6254 size_t new_capacity = MIN2(_capacity*2, CMSMarkStackSizeMax); 6255 // Do not give up existing stack until we have managed to 6256 // get the double capacity that we desired. 6257 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 6258 new_capacity * sizeof(oop))); 6259 if (rs.is_reserved()) { 6260 // Release the backing store associated with old stack 6261 _virtual_space.release(); 6262 // Reinitialize virtual space for new stack 6263 if (!_virtual_space.initialize(rs, rs.size())) { 6264 fatal("Not enough swap for expanded marking stack"); 6265 } 6266 _base = (oop*)(_virtual_space.low()); 6267 _index = 0; 6268 _capacity = new_capacity; 6269 } else if (_failed_double++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) { 6270 // Failed to double capacity, continue; 6271 // we print a detail message only once per CMS cycle. 6272 gclog_or_tty->print(" (benign) Failed to expand marking stack from "SIZE_FORMAT"K to " 6273 SIZE_FORMAT"K", 6274 _capacity / K, new_capacity / K); 6275 } 6276 } 6277 6278 6279 // Closures 6280 // XXX: there seems to be a lot of code duplication here; 6281 // should refactor and consolidate common code. 6282 6283 // This closure is used to mark refs into the CMS generation in 6284 // the CMS bit map. Called at the first checkpoint. This closure 6285 // assumes that we do not need to re-mark dirty cards; if the CMS 6286 // generation on which this is used is not an oldest (modulo perm gen) 6287 // generation then this will lose younger_gen cards! 6288 6289 MarkRefsIntoClosure::MarkRefsIntoClosure( 6290 MemRegion span, CMSBitMap* bitMap, bool should_do_nmethods): 6291 _span(span), 6292 _bitMap(bitMap), 6293 _should_do_nmethods(should_do_nmethods) 6294 { 6295 assert(_ref_processor == NULL, "deliberately left NULL"); 6296 assert(_bitMap->covers(_span), "_bitMap/_span mismatch"); 6297 } 6298 6299 void MarkRefsIntoClosure::do_oop(oop* p) { 6300 // if p points into _span, then mark corresponding bit in _markBitMap 6301 oop thisOop = *p; 6302 if (thisOop != NULL) { 6303 assert(thisOop->is_oop(), "expected an oop"); 6304 HeapWord* addr = (HeapWord*)thisOop; 6305 if (_span.contains(addr)) { 6306 // this should be made more efficient 6307 _bitMap->mark(addr); 6308 } 6309 } 6310 } 6311 6312 // A variant of the above, used for CMS marking verification. 6313 MarkRefsIntoVerifyClosure::MarkRefsIntoVerifyClosure( 6314 MemRegion span, CMSBitMap* verification_bm, CMSBitMap* cms_bm, 6315 bool should_do_nmethods): 6316 _span(span), 6317 _verification_bm(verification_bm), 6318 _cms_bm(cms_bm), 6319 _should_do_nmethods(should_do_nmethods) { 6320 assert(_ref_processor == NULL, "deliberately left NULL"); 6321 assert(_verification_bm->covers(_span), "_verification_bm/_span mismatch"); 6322 } 6323 6324 void MarkRefsIntoVerifyClosure::do_oop(oop* p) { 6325 // if p points into _span, then mark corresponding bit in _markBitMap 6326 oop this_oop = *p; 6327 if (this_oop != NULL) { 6328 assert(this_oop->is_oop(), "expected an oop"); 6329 HeapWord* addr = (HeapWord*)this_oop; 6330 if (_span.contains(addr)) { 6331 _verification_bm->mark(addr); 6332 if (!_cms_bm->isMarked(addr)) { 6333 oop(addr)->print(); 6334 gclog_or_tty->print_cr(" ("INTPTR_FORMAT" should have been marked)", addr); 6335 fatal("... aborting"); 6336 } 6337 } 6338 } 6339 } 6340 6341 ////////////////////////////////////////////////// 6342 // MarkRefsIntoAndScanClosure 6343 ////////////////////////////////////////////////// 6344 6345 MarkRefsIntoAndScanClosure::MarkRefsIntoAndScanClosure(MemRegion span, 6346 ReferenceProcessor* rp, 6347 CMSBitMap* bit_map, 6348 CMSBitMap* mod_union_table, 6349 CMSMarkStack* mark_stack, 6350 CMSMarkStack* revisit_stack, 6351 CMSCollector* collector, 6352 bool should_yield, 6353 bool concurrent_precleaning): 6354 _collector(collector), 6355 _span(span), 6356 _bit_map(bit_map), 6357 _mark_stack(mark_stack), 6358 _pushAndMarkClosure(collector, span, rp, bit_map, mod_union_table, 6359 mark_stack, revisit_stack, concurrent_precleaning), 6360 _yield(should_yield), 6361 _concurrent_precleaning(concurrent_precleaning), 6362 _freelistLock(NULL) 6363 { 6364 _ref_processor = rp; 6365 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 6366 } 6367 6368 // This closure is used to mark refs into the CMS generation at the 6369 // second (final) checkpoint, and to scan and transitively follow 6370 // the unmarked oops. It is also used during the concurrent precleaning 6371 // phase while scanning objects on dirty cards in the CMS generation. 6372 // The marks are made in the marking bit map and the marking stack is 6373 // used for keeping the (newly) grey objects during the scan. 6374 // The parallel version (Par_...) appears further below. 6375 void MarkRefsIntoAndScanClosure::do_oop(oop* p) { 6376 oop this_oop = *p; 6377 if (this_oop != NULL) { 6378 assert(this_oop->is_oop(), "expected an oop"); 6379 HeapWord* addr = (HeapWord*)this_oop; 6380 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 6381 assert(_collector->overflow_list_is_empty(), "should be empty"); 6382 if (_span.contains(addr) && 6383 !_bit_map->isMarked(addr)) { 6384 // mark bit map (object is now grey) 6385 _bit_map->mark(addr); 6386 // push on marking stack (stack should be empty), and drain the 6387 // stack by applying this closure to the oops in the oops popped 6388 // from the stack (i.e. blacken the grey objects) 6389 bool res = _mark_stack->push(this_oop); 6390 assert(res, "Should have space to push on empty stack"); 6391 do { 6392 oop new_oop = _mark_stack->pop(); 6393 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 6394 assert(new_oop->is_parsable(), "Found unparsable oop"); 6395 assert(_bit_map->isMarked((HeapWord*)new_oop), 6396 "only grey objects on this stack"); 6397 // iterate over the oops in this oop, marking and pushing 6398 // the ones in CMS heap (i.e. in _span). 6399 new_oop->oop_iterate(&_pushAndMarkClosure); 6400 // check if it's time to yield 6401 do_yield_check(); 6402 } while (!_mark_stack->isEmpty() || 6403 (!_concurrent_precleaning && take_from_overflow_list())); 6404 // if marking stack is empty, and we are not doing this 6405 // during precleaning, then check the overflow list 6406 } 6407 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 6408 assert(_collector->overflow_list_is_empty(), 6409 "overflow list was drained above"); 6410 // We could restore evacuated mark words, if any, used for 6411 // overflow list links here because the overflow list is 6412 // provably empty here. That would reduce the maximum 6413 // size requirements for preserved_{oop,mark}_stack. 6414 // But we'll just postpone it until we are all done 6415 // so we can just stream through. 6416 if (!_concurrent_precleaning && CMSOverflowEarlyRestoration) { 6417 _collector->restore_preserved_marks_if_any(); 6418 assert(_collector->no_preserved_marks(), "No preserved marks"); 6419 } 6420 assert(!CMSOverflowEarlyRestoration || _collector->no_preserved_marks(), 6421 "All preserved marks should have been restored above"); 6422 } 6423 } 6424 6425 void MarkRefsIntoAndScanClosure::do_yield_work() { 6426 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6427 "CMS thread should hold CMS token"); 6428 assert_lock_strong(_freelistLock); 6429 assert_lock_strong(_bit_map->lock()); 6430 // relinquish the free_list_lock and bitMaplock() 6431 _bit_map->lock()->unlock(); 6432 _freelistLock->unlock(); 6433 ConcurrentMarkSweepThread::desynchronize(true); 6434 ConcurrentMarkSweepThread::acknowledge_yield_request(); 6435 _collector->stopTimer(); 6436 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 6437 if (PrintCMSStatistics != 0) { 6438 _collector->incrementYields(); 6439 } 6440 _collector->icms_wait(); 6441 6442 // See the comment in coordinator_yield() 6443 for (unsigned i = 0; i < CMSYieldSleepCount && 6444 ConcurrentMarkSweepThread::should_yield() && 6445 !CMSCollector::foregroundGCIsActive(); ++i) { 6446 os::sleep(Thread::current(), 1, false); 6447 ConcurrentMarkSweepThread::acknowledge_yield_request(); 6448 } 6449 6450 ConcurrentMarkSweepThread::synchronize(true); 6451 _freelistLock->lock_without_safepoint_check(); 6452 _bit_map->lock()->lock_without_safepoint_check(); 6453 _collector->startTimer(); 6454 } 6455 6456 /////////////////////////////////////////////////////////// 6457 // Par_MarkRefsIntoAndScanClosure: a parallel version of 6458 // MarkRefsIntoAndScanClosure 6459 /////////////////////////////////////////////////////////// 6460 Par_MarkRefsIntoAndScanClosure::Par_MarkRefsIntoAndScanClosure( 6461 CMSCollector* collector, MemRegion span, ReferenceProcessor* rp, 6462 CMSBitMap* bit_map, OopTaskQueue* work_queue, CMSMarkStack* revisit_stack): 6463 _span(span), 6464 _bit_map(bit_map), 6465 _work_queue(work_queue), 6466 _low_water_mark(MIN2((uint)(work_queue->max_elems()/4), 6467 (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads))), 6468 _par_pushAndMarkClosure(collector, span, rp, bit_map, work_queue, 6469 revisit_stack) 6470 { 6471 _ref_processor = rp; 6472 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 6473 } 6474 6475 // This closure is used to mark refs into the CMS generation at the 6476 // second (final) checkpoint, and to scan and transitively follow 6477 // the unmarked oops. The marks are made in the marking bit map and 6478 // the work_queue is used for keeping the (newly) grey objects during 6479 // the scan phase whence they are also available for stealing by parallel 6480 // threads. Since the marking bit map is shared, updates are 6481 // synchronized (via CAS). 6482 void Par_MarkRefsIntoAndScanClosure::do_oop(oop* p) { 6483 oop this_oop = *p; 6484 if (this_oop != NULL) { 6485 // Ignore mark word because this could be an already marked oop 6486 // that may be chained at the end of the overflow list. 6487 assert(this_oop->is_oop(true /* ignore mark word */), "expected an oop"); 6488 HeapWord* addr = (HeapWord*)this_oop; 6489 if (_span.contains(addr) && 6490 !_bit_map->isMarked(addr)) { 6491 // mark bit map (object will become grey): 6492 // It is possible for several threads to be 6493 // trying to "claim" this object concurrently; 6494 // the unique thread that succeeds in marking the 6495 // object first will do the subsequent push on 6496 // to the work queue (or overflow list). 6497 if (_bit_map->par_mark(addr)) { 6498 // push on work_queue (which may not be empty), and trim the 6499 // queue to an appropriate length by applying this closure to 6500 // the oops in the oops popped from the stack (i.e. blacken the 6501 // grey objects) 6502 bool res = _work_queue->push(this_oop); 6503 assert(res, "Low water mark should be less than capacity?"); 6504 trim_queue(_low_water_mark); 6505 } // Else, another thread claimed the object 6506 } 6507 } 6508 } 6509 6510 // This closure is used to rescan the marked objects on the dirty cards 6511 // in the mod union table and the card table proper. 6512 size_t ScanMarkedObjectsAgainCarefullyClosure::do_object_careful_m( 6513 oop p, MemRegion mr) { 6514 6515 size_t size = 0; 6516 HeapWord* addr = (HeapWord*)p; 6517 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6518 assert(_span.contains(addr), "we are scanning the CMS generation"); 6519 // check if it's time to yield 6520 if (do_yield_check()) { 6521 // We yielded for some foreground stop-world work, 6522 // and we have been asked to abort this ongoing preclean cycle. 6523 return 0; 6524 } 6525 if (_bitMap->isMarked(addr)) { 6526 // it's marked; is it potentially uninitialized? 6527 if (p->klass() != NULL) { 6528 if (CMSPermGenPrecleaningEnabled && !p->is_parsable()) { 6529 // Signal precleaning to redirty the card since 6530 // the klass pointer is already installed. 6531 assert(size == 0, "Initial value"); 6532 } else { 6533 assert(p->is_parsable(), "must be parsable."); 6534 // an initialized object; ignore mark word in verification below 6535 // since we are running concurrent with mutators 6536 assert(p->is_oop(true), "should be an oop"); 6537 if (p->is_objArray()) { 6538 // objArrays are precisely marked; restrict scanning 6539 // to dirty cards only. 6540 size = p->oop_iterate(_scanningClosure, mr); 6541 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 6542 "adjustObjectSize should be the identity for array sizes, " 6543 "which are necessarily larger than minimum object size of " 6544 "two heap words"); 6545 } else { 6546 // A non-array may have been imprecisely marked; we need 6547 // to scan object in its entirety. 6548 size = CompactibleFreeListSpace::adjustObjectSize( 6549 p->oop_iterate(_scanningClosure)); 6550 } 6551 #ifdef DEBUG 6552 size_t direct_size = 6553 CompactibleFreeListSpace::adjustObjectSize(p->size()); 6554 assert(size == direct_size, "Inconsistency in size"); 6555 assert(size >= 3, "Necessary for Printezis marks to work"); 6556 if (!_bitMap->isMarked(addr+1)) { 6557 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size); 6558 } else { 6559 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size-1); 6560 assert(_bitMap->isMarked(addr+size-1), 6561 "inconsistent Printezis mark"); 6562 } 6563 #endif // DEBUG 6564 } 6565 } else { 6566 // an unitialized object 6567 assert(_bitMap->isMarked(addr+1), "missing Printezis mark?"); 6568 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 6569 size = pointer_delta(nextOneAddr + 1, addr); 6570 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 6571 "alignment problem"); 6572 // Note that pre-cleaning needn't redirty the card. OopDesc::set_klass() 6573 // will dirty the card when the klass pointer is installed in the 6574 // object (signalling the completion of initialization). 6575 } 6576 } else { 6577 // Either a not yet marked object or an uninitialized object 6578 if (p->klass() == NULL || !p->is_parsable()) { 6579 // An uninitialized object, skip to the next card, since 6580 // we may not be able to read its P-bits yet. 6581 assert(size == 0, "Initial value"); 6582 } else { 6583 // An object not (yet) reached by marking: we merely need to 6584 // compute its size so as to go look at the next block. 6585 assert(p->is_oop(true), "should be an oop"); 6586 size = CompactibleFreeListSpace::adjustObjectSize(p->size()); 6587 } 6588 } 6589 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6590 return size; 6591 } 6592 6593 void ScanMarkedObjectsAgainCarefullyClosure::do_yield_work() { 6594 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6595 "CMS thread should hold CMS token"); 6596 assert_lock_strong(_freelistLock); 6597 assert_lock_strong(_bitMap->lock()); 6598 // relinquish the free_list_lock and bitMaplock() 6599 _bitMap->lock()->unlock(); 6600 _freelistLock->unlock(); 6601 ConcurrentMarkSweepThread::desynchronize(true); 6602 ConcurrentMarkSweepThread::acknowledge_yield_request(); 6603 _collector->stopTimer(); 6604 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 6605 if (PrintCMSStatistics != 0) { 6606 _collector->incrementYields(); 6607 } 6608 _collector->icms_wait(); 6609 6610 // See the comment in coordinator_yield() 6611 for (unsigned i = 0; i < CMSYieldSleepCount && 6612 ConcurrentMarkSweepThread::should_yield() && 6613 !CMSCollector::foregroundGCIsActive(); ++i) { 6614 os::sleep(Thread::current(), 1, false); 6615 ConcurrentMarkSweepThread::acknowledge_yield_request(); 6616 } 6617 6618 ConcurrentMarkSweepThread::synchronize(true); 6619 _freelistLock->lock_without_safepoint_check(); 6620 _bitMap->lock()->lock_without_safepoint_check(); 6621 _collector->startTimer(); 6622 } 6623 6624 6625 ////////////////////////////////////////////////////////////////// 6626 // SurvivorSpacePrecleanClosure 6627 ////////////////////////////////////////////////////////////////// 6628 // This (single-threaded) closure is used to preclean the oops in 6629 // the survivor spaces. 6630 size_t SurvivorSpacePrecleanClosure::do_object_careful(oop p) { 6631 6632 HeapWord* addr = (HeapWord*)p; 6633 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6634 assert(!_span.contains(addr), "we are scanning the survivor spaces"); 6635 assert(p->klass() != NULL, "object should be initializd"); 6636 assert(p->is_parsable(), "must be parsable."); 6637 // an initialized object; ignore mark word in verification below 6638 // since we are running concurrent with mutators 6639 assert(p->is_oop(true), "should be an oop"); 6640 // Note that we do not yield while we iterate over 6641 // the interior oops of p, pushing the relevant ones 6642 // on our marking stack. 6643 size_t size = p->oop_iterate(_scanning_closure); 6644 do_yield_check(); 6645 // Observe that below, we do not abandon the preclean 6646 // phase as soon as we should; rather we empty the 6647 // marking stack before returning. This is to satisfy 6648 // some existing assertions. In general, it may be a 6649 // good idea to abort immediately and complete the marking 6650 // from the grey objects at a later time. 6651 while (!_mark_stack->isEmpty()) { 6652 oop new_oop = _mark_stack->pop(); 6653 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 6654 assert(new_oop->is_parsable(), "Found unparsable oop"); 6655 assert(_bit_map->isMarked((HeapWord*)new_oop), 6656 "only grey objects on this stack"); 6657 // iterate over the oops in this oop, marking and pushing 6658 // the ones in CMS heap (i.e. in _span). 6659 new_oop->oop_iterate(_scanning_closure); 6660 // check if it's time to yield 6661 do_yield_check(); 6662 } 6663 unsigned int after_count = 6664 GenCollectedHeap::heap()->total_collections(); 6665 bool abort = (_before_count != after_count) || 6666 _collector->should_abort_preclean(); 6667 return abort ? 0 : size; 6668 } 6669 6670 void SurvivorSpacePrecleanClosure::do_yield_work() { 6671 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6672 "CMS thread should hold CMS token"); 6673 assert_lock_strong(_bit_map->lock()); 6674 // Relinquish the bit map lock 6675 _bit_map->lock()->unlock(); 6676 ConcurrentMarkSweepThread::desynchronize(true); 6677 ConcurrentMarkSweepThread::acknowledge_yield_request(); 6678 _collector->stopTimer(); 6679 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 6680 if (PrintCMSStatistics != 0) { 6681 _collector->incrementYields(); 6682 } 6683 _collector->icms_wait(); 6684 6685 // See the comment in coordinator_yield() 6686 for (unsigned i = 0; i < CMSYieldSleepCount && 6687 ConcurrentMarkSweepThread::should_yield() && 6688 !CMSCollector::foregroundGCIsActive(); ++i) { 6689 os::sleep(Thread::current(), 1, false); 6690 ConcurrentMarkSweepThread::acknowledge_yield_request(); 6691 } 6692 6693 ConcurrentMarkSweepThread::synchronize(true); 6694 _bit_map->lock()->lock_without_safepoint_check(); 6695 _collector->startTimer(); 6696 } 6697 6698 // This closure is used to rescan the marked objects on the dirty cards 6699 // in the mod union table and the card table proper. In the parallel 6700 // case, although the bitMap is shared, we do a single read so the 6701 // isMarked() query is "safe". 6702 bool ScanMarkedObjectsAgainClosure::do_object_bm(oop p, MemRegion mr) { 6703 // Ignore mark word because we are running concurrent with mutators 6704 assert(p->is_oop_or_null(true), "expected an oop or null"); 6705 HeapWord* addr = (HeapWord*)p; 6706 assert(_span.contains(addr), "we are scanning the CMS generation"); 6707 bool is_obj_array = false; 6708 #ifdef DEBUG 6709 if (!_parallel) { 6710 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 6711 assert(_collector->overflow_list_is_empty(), 6712 "overflow list should be empty"); 6713 6714 } 6715 #endif // DEBUG 6716 if (_bit_map->isMarked(addr)) { 6717 // Obj arrays are precisely marked, non-arrays are not; 6718 // so we scan objArrays precisely and non-arrays in their 6719 // entirety. 6720 if (p->is_objArray()) { 6721 is_obj_array = true; 6722 if (_parallel) { 6723 p->oop_iterate(_par_scan_closure, mr); 6724 } else { 6725 p->oop_iterate(_scan_closure, mr); 6726 } 6727 } else { 6728 if (_parallel) { 6729 p->oop_iterate(_par_scan_closure); 6730 } else { 6731 p->oop_iterate(_scan_closure); 6732 } 6733 } 6734 } 6735 #ifdef DEBUG 6736 if (!_parallel) { 6737 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 6738 assert(_collector->overflow_list_is_empty(), 6739 "overflow list should be empty"); 6740 6741 } 6742 #endif // DEBUG 6743 return is_obj_array; 6744 } 6745 6746 MarkFromRootsClosure::MarkFromRootsClosure(CMSCollector* collector, 6747 MemRegion span, 6748 CMSBitMap* bitMap, CMSMarkStack* markStack, 6749 CMSMarkStack* revisitStack, 6750 bool should_yield, bool verifying): 6751 _collector(collector), 6752 _span(span), 6753 _bitMap(bitMap), 6754 _mut(&collector->_modUnionTable), 6755 _markStack(markStack), 6756 _revisitStack(revisitStack), 6757 _yield(should_yield), 6758 _skipBits(0) 6759 { 6760 assert(_markStack->isEmpty(), "stack should be empty"); 6761 _finger = _bitMap->startWord(); 6762 _threshold = _finger; 6763 assert(_collector->_restart_addr == NULL, "Sanity check"); 6764 assert(_span.contains(_finger), "Out of bounds _finger?"); 6765 DEBUG_ONLY(_verifying = verifying;) 6766 } 6767 6768 void MarkFromRootsClosure::reset(HeapWord* addr) { 6769 assert(_markStack->isEmpty(), "would cause duplicates on stack"); 6770 assert(_span.contains(addr), "Out of bounds _finger?"); 6771 _finger = addr; 6772 _threshold = (HeapWord*)round_to( 6773 (intptr_t)_finger, CardTableModRefBS::card_size); 6774 } 6775 6776 // Should revisit to see if this should be restructured for 6777 // greater efficiency. 6778 void MarkFromRootsClosure::do_bit(size_t offset) { 6779 if (_skipBits > 0) { 6780 _skipBits--; 6781 return; 6782 } 6783 // convert offset into a HeapWord* 6784 HeapWord* addr = _bitMap->startWord() + offset; 6785 assert(_bitMap->endWord() && addr < _bitMap->endWord(), 6786 "address out of range"); 6787 assert(_bitMap->isMarked(addr), "tautology"); 6788 if (_bitMap->isMarked(addr+1)) { 6789 // this is an allocated but not yet initialized object 6790 assert(_skipBits == 0, "tautology"); 6791 _skipBits = 2; // skip next two marked bits ("Printezis-marks") 6792 oop p = oop(addr); 6793 if (p->klass() == NULL || !p->is_parsable()) { 6794 DEBUG_ONLY(if (!_verifying) {) 6795 // We re-dirty the cards on which this object lies and increase 6796 // the _threshold so that we'll come back to scan this object 6797 // during the preclean or remark phase. (CMSCleanOnEnter) 6798 if (CMSCleanOnEnter) { 6799 size_t sz = _collector->block_size_using_printezis_bits(addr); 6800 HeapWord* start_card_addr = (HeapWord*)round_down( 6801 (intptr_t)addr, CardTableModRefBS::card_size); 6802 HeapWord* end_card_addr = (HeapWord*)round_to( 6803 (intptr_t)(addr+sz), CardTableModRefBS::card_size); 6804 MemRegion redirty_range = MemRegion(start_card_addr, end_card_addr); 6805 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 6806 // Bump _threshold to end_card_addr; note that 6807 // _threshold cannot possibly exceed end_card_addr, anyhow. 6808 // This prevents future clearing of the card as the scan proceeds 6809 // to the right. 6810 assert(_threshold <= end_card_addr, 6811 "Because we are just scanning into this object"); 6812 if (_threshold < end_card_addr) { 6813 _threshold = end_card_addr; 6814 } 6815 if (p->klass() != NULL) { 6816 // Redirty the range of cards... 6817 _mut->mark_range(redirty_range); 6818 } // ...else the setting of klass will dirty the card anyway. 6819 } 6820 DEBUG_ONLY(}) 6821 return; 6822 } 6823 } 6824 scanOopsInOop(addr); 6825 } 6826 6827 // We take a break if we've been at this for a while, 6828 // so as to avoid monopolizing the locks involved. 6829 void MarkFromRootsClosure::do_yield_work() { 6830 // First give up the locks, then yield, then re-lock 6831 // We should probably use a constructor/destructor idiom to 6832 // do this unlock/lock or modify the MutexUnlocker class to 6833 // serve our purpose. XXX 6834 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6835 "CMS thread should hold CMS token"); 6836 assert_lock_strong(_bitMap->lock()); 6837 _bitMap->lock()->unlock(); 6838 ConcurrentMarkSweepThread::desynchronize(true); 6839 ConcurrentMarkSweepThread::acknowledge_yield_request(); 6840 _collector->stopTimer(); 6841 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 6842 if (PrintCMSStatistics != 0) { 6843 _collector->incrementYields(); 6844 } 6845 _collector->icms_wait(); 6846 6847 // See the comment in coordinator_yield() 6848 for (unsigned i = 0; i < CMSYieldSleepCount && 6849 ConcurrentMarkSweepThread::should_yield() && 6850 !CMSCollector::foregroundGCIsActive(); ++i) { 6851 os::sleep(Thread::current(), 1, false); 6852 ConcurrentMarkSweepThread::acknowledge_yield_request(); 6853 } 6854 6855 ConcurrentMarkSweepThread::synchronize(true); 6856 _bitMap->lock()->lock_without_safepoint_check(); 6857 _collector->startTimer(); 6858 } 6859 6860 void MarkFromRootsClosure::scanOopsInOop(HeapWord* ptr) { 6861 assert(_bitMap->isMarked(ptr), "expected bit to be set"); 6862 assert(_markStack->isEmpty(), 6863 "should drain stack to limit stack usage"); 6864 // convert ptr to an oop preparatory to scanning 6865 oop this_oop = oop(ptr); 6866 // Ignore mark word in verification below, since we 6867 // may be running concurrent with mutators. 6868 assert(this_oop->is_oop(true), "should be an oop"); 6869 assert(_finger <= ptr, "_finger runneth ahead"); 6870 // advance the finger to right end of this object 6871 _finger = ptr + this_oop->size(); 6872 assert(_finger > ptr, "we just incremented it above"); 6873 // On large heaps, it may take us some time to get through 6874 // the marking phase (especially if running iCMS). During 6875 // this time it's possible that a lot of mutations have 6876 // accumulated in the card table and the mod union table -- 6877 // these mutation records are redundant until we have 6878 // actually traced into the corresponding card. 6879 // Here, we check whether advancing the finger would make 6880 // us cross into a new card, and if so clear corresponding 6881 // cards in the MUT (preclean them in the card-table in the 6882 // future). 6883 6884 DEBUG_ONLY(if (!_verifying) {) 6885 // The clean-on-enter optimization is disabled by default, 6886 // until we fix 6178663. 6887 if (CMSCleanOnEnter && (_finger > _threshold)) { 6888 // [_threshold, _finger) represents the interval 6889 // of cards to be cleared in MUT (or precleaned in card table). 6890 // The set of cards to be cleared is all those that overlap 6891 // with the interval [_threshold, _finger); note that 6892 // _threshold is always kept card-aligned but _finger isn't 6893 // always card-aligned. 6894 HeapWord* old_threshold = _threshold; 6895 assert(old_threshold == (HeapWord*)round_to( 6896 (intptr_t)old_threshold, CardTableModRefBS::card_size), 6897 "_threshold should always be card-aligned"); 6898 _threshold = (HeapWord*)round_to( 6899 (intptr_t)_finger, CardTableModRefBS::card_size); 6900 MemRegion mr(old_threshold, _threshold); 6901 assert(!mr.is_empty(), "Control point invariant"); 6902 assert(_span.contains(mr), "Should clear within span"); 6903 // XXX When _finger crosses from old gen into perm gen 6904 // we may be doing unnecessary cleaning; do better in the 6905 // future by detecting that condition and clearing fewer 6906 // MUT/CT entries. 6907 _mut->clear_range(mr); 6908 } 6909 DEBUG_ONLY(}) 6910 6911 // Note: the finger doesn't advance while we drain 6912 // the stack below. 6913 PushOrMarkClosure pushOrMarkClosure(_collector, 6914 _span, _bitMap, _markStack, 6915 _revisitStack, 6916 _finger, this); 6917 bool res = _markStack->push(this_oop); 6918 assert(res, "Empty non-zero size stack should have space for single push"); 6919 while (!_markStack->isEmpty()) { 6920 oop new_oop = _markStack->pop(); 6921 // Skip verifying header mark word below because we are 6922 // running concurrent with mutators. 6923 assert(new_oop->is_oop(true), "Oops! expected to pop an oop"); 6924 // now scan this oop's oops 6925 new_oop->oop_iterate(&pushOrMarkClosure); 6926 do_yield_check(); 6927 } 6928 assert(_markStack->isEmpty(), "tautology, emphasizing post-condition"); 6929 } 6930 6931 Par_MarkFromRootsClosure::Par_MarkFromRootsClosure(CMSConcMarkingTask* task, 6932 CMSCollector* collector, MemRegion span, 6933 CMSBitMap* bit_map, 6934 OopTaskQueue* work_queue, 6935 CMSMarkStack* overflow_stack, 6936 CMSMarkStack* revisit_stack, 6937 bool should_yield): 6938 _collector(collector), 6939 _whole_span(collector->_span), 6940 _span(span), 6941 _bit_map(bit_map), 6942 _mut(&collector->_modUnionTable), 6943 _work_queue(work_queue), 6944 _overflow_stack(overflow_stack), 6945 _revisit_stack(revisit_stack), 6946 _yield(should_yield), 6947 _skip_bits(0), 6948 _task(task) 6949 { 6950 assert(_work_queue->size() == 0, "work_queue should be empty"); 6951 _finger = span.start(); 6952 _threshold = _finger; // XXX Defer clear-on-enter optimization for now 6953 assert(_span.contains(_finger), "Out of bounds _finger?"); 6954 } 6955 6956 // Should revisit to see if this should be restructured for 6957 // greater efficiency. 6958 void Par_MarkFromRootsClosure::do_bit(size_t offset) { 6959 if (_skip_bits > 0) { 6960 _skip_bits--; 6961 return; 6962 } 6963 // convert offset into a HeapWord* 6964 HeapWord* addr = _bit_map->startWord() + offset; 6965 assert(_bit_map->endWord() && addr < _bit_map->endWord(), 6966 "address out of range"); 6967 assert(_bit_map->isMarked(addr), "tautology"); 6968 if (_bit_map->isMarked(addr+1)) { 6969 // this is an allocated object that might not yet be initialized 6970 assert(_skip_bits == 0, "tautology"); 6971 _skip_bits = 2; // skip next two marked bits ("Printezis-marks") 6972 oop p = oop(addr); 6973 if (p->klass() == NULL || !p->is_parsable()) { 6974 // in the case of Clean-on-Enter optimization, redirty card 6975 // and avoid clearing card by increasing the threshold. 6976 return; 6977 } 6978 } 6979 scan_oops_in_oop(addr); 6980 } 6981 6982 void Par_MarkFromRootsClosure::scan_oops_in_oop(HeapWord* ptr) { 6983 assert(_bit_map->isMarked(ptr), "expected bit to be set"); 6984 // Should we assert that our work queue is empty or 6985 // below some drain limit? 6986 assert(_work_queue->size() == 0, 6987 "should drain stack to limit stack usage"); 6988 // convert ptr to an oop preparatory to scanning 6989 oop this_oop = oop(ptr); 6990 // Ignore mark word in verification below, since we 6991 // may be running concurrent with mutators. 6992 assert(this_oop->is_oop(true), "should be an oop"); 6993 assert(_finger <= ptr, "_finger runneth ahead"); 6994 // advance the finger to right end of this object 6995 _finger = ptr + this_oop->size(); 6996 assert(_finger > ptr, "we just incremented it above"); 6997 // On large heaps, it may take us some time to get through 6998 // the marking phase (especially if running iCMS). During 6999 // this time it's possible that a lot of mutations have 7000 // accumulated in the card table and the mod union table -- 7001 // these mutation records are redundant until we have 7002 // actually traced into the corresponding card. 7003 // Here, we check whether advancing the finger would make 7004 // us cross into a new card, and if so clear corresponding 7005 // cards in the MUT (preclean them in the card-table in the 7006 // future). 7007 7008 // The clean-on-enter optimization is disabled by default, 7009 // until we fix 6178663. 7010 if (CMSCleanOnEnter && (_finger > _threshold)) { 7011 // [_threshold, _finger) represents the interval 7012 // of cards to be cleared in MUT (or precleaned in card table). 7013 // The set of cards to be cleared is all those that overlap 7014 // with the interval [_threshold, _finger); note that 7015 // _threshold is always kept card-aligned but _finger isn't 7016 // always card-aligned. 7017 HeapWord* old_threshold = _threshold; 7018 assert(old_threshold == (HeapWord*)round_to( 7019 (intptr_t)old_threshold, CardTableModRefBS::card_size), 7020 "_threshold should always be card-aligned"); 7021 _threshold = (HeapWord*)round_to( 7022 (intptr_t)_finger, CardTableModRefBS::card_size); 7023 MemRegion mr(old_threshold, _threshold); 7024 assert(!mr.is_empty(), "Control point invariant"); 7025 assert(_span.contains(mr), "Should clear within span"); // _whole_span ?? 7026 // XXX When _finger crosses from old gen into perm gen 7027 // we may be doing unnecessary cleaning; do better in the 7028 // future by detecting that condition and clearing fewer 7029 // MUT/CT entries. 7030 _mut->clear_range(mr); 7031 } 7032 7033 // Note: the local finger doesn't advance while we drain 7034 // the stack below, but the global finger sure can and will. 7035 HeapWord** gfa = _task->global_finger_addr(); 7036 Par_PushOrMarkClosure pushOrMarkClosure(_collector, 7037 _span, _bit_map, 7038 _work_queue, 7039 _overflow_stack, 7040 _revisit_stack, 7041 _finger, 7042 gfa, this); 7043 bool res = _work_queue->push(this_oop); // overflow could occur here 7044 assert(res, "Will hold once we use workqueues"); 7045 while (true) { 7046 oop new_oop; 7047 if (!_work_queue->pop_local(new_oop)) { 7048 // We emptied our work_queue; check if there's stuff that can 7049 // be gotten from the overflow stack. 7050 if (CMSConcMarkingTask::get_work_from_overflow_stack( 7051 _overflow_stack, _work_queue)) { 7052 do_yield_check(); 7053 continue; 7054 } else { // done 7055 break; 7056 } 7057 } 7058 // Skip verifying header mark word below because we are 7059 // running concurrent with mutators. 7060 assert(new_oop->is_oop(true), "Oops! expected to pop an oop"); 7061 // now scan this oop's oops 7062 new_oop->oop_iterate(&pushOrMarkClosure); 7063 do_yield_check(); 7064 } 7065 assert(_work_queue->size() == 0, "tautology, emphasizing post-condition"); 7066 } 7067 7068 // Yield in response to a request from VM Thread or 7069 // from mutators. 7070 void Par_MarkFromRootsClosure::do_yield_work() { 7071 assert(_task != NULL, "sanity"); 7072 _task->yield(); 7073 } 7074 7075 // A variant of the above used for verifying CMS marking work. 7076 MarkFromRootsVerifyClosure::MarkFromRootsVerifyClosure(CMSCollector* collector, 7077 MemRegion span, 7078 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 7079 CMSMarkStack* mark_stack): 7080 _collector(collector), 7081 _span(span), 7082 _verification_bm(verification_bm), 7083 _cms_bm(cms_bm), 7084 _mark_stack(mark_stack), 7085 _pam_verify_closure(collector, span, verification_bm, cms_bm, 7086 mark_stack) 7087 { 7088 assert(_mark_stack->isEmpty(), "stack should be empty"); 7089 _finger = _verification_bm->startWord(); 7090 assert(_collector->_restart_addr == NULL, "Sanity check"); 7091 assert(_span.contains(_finger), "Out of bounds _finger?"); 7092 } 7093 7094 void MarkFromRootsVerifyClosure::reset(HeapWord* addr) { 7095 assert(_mark_stack->isEmpty(), "would cause duplicates on stack"); 7096 assert(_span.contains(addr), "Out of bounds _finger?"); 7097 _finger = addr; 7098 } 7099 7100 // Should revisit to see if this should be restructured for 7101 // greater efficiency. 7102 void MarkFromRootsVerifyClosure::do_bit(size_t offset) { 7103 // convert offset into a HeapWord* 7104 HeapWord* addr = _verification_bm->startWord() + offset; 7105 assert(_verification_bm->endWord() && addr < _verification_bm->endWord(), 7106 "address out of range"); 7107 assert(_verification_bm->isMarked(addr), "tautology"); 7108 assert(_cms_bm->isMarked(addr), "tautology"); 7109 7110 assert(_mark_stack->isEmpty(), 7111 "should drain stack to limit stack usage"); 7112 // convert addr to an oop preparatory to scanning 7113 oop this_oop = oop(addr); 7114 assert(this_oop->is_oop(), "should be an oop"); 7115 assert(_finger <= addr, "_finger runneth ahead"); 7116 // advance the finger to right end of this object 7117 _finger = addr + this_oop->size(); 7118 assert(_finger > addr, "we just incremented it above"); 7119 // Note: the finger doesn't advance while we drain 7120 // the stack below. 7121 bool res = _mark_stack->push(this_oop); 7122 assert(res, "Empty non-zero size stack should have space for single push"); 7123 while (!_mark_stack->isEmpty()) { 7124 oop new_oop = _mark_stack->pop(); 7125 assert(new_oop->is_oop(), "Oops! expected to pop an oop"); 7126 // now scan this oop's oops 7127 new_oop->oop_iterate(&_pam_verify_closure); 7128 } 7129 assert(_mark_stack->isEmpty(), "tautology, emphasizing post-condition"); 7130 } 7131 7132 PushAndMarkVerifyClosure::PushAndMarkVerifyClosure( 7133 CMSCollector* collector, MemRegion span, 7134 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 7135 CMSMarkStack* mark_stack): 7136 OopClosure(collector->ref_processor()), 7137 _collector(collector), 7138 _span(span), 7139 _verification_bm(verification_bm), 7140 _cms_bm(cms_bm), 7141 _mark_stack(mark_stack) 7142 { } 7143 7144 7145 // Upon stack overflow, we discard (part of) the stack, 7146 // remembering the least address amongst those discarded 7147 // in CMSCollector's _restart_address. 7148 void PushAndMarkVerifyClosure::handle_stack_overflow(HeapWord* lost) { 7149 // Remember the least grey address discarded 7150 HeapWord* ra = (HeapWord*)_mark_stack->least_value(lost); 7151 _collector->lower_restart_addr(ra); 7152 _mark_stack->reset(); // discard stack contents 7153 _mark_stack->expand(); // expand the stack if possible 7154 } 7155 7156 void PushAndMarkVerifyClosure::do_oop(oop* p) { 7157 oop this_oop = *p; 7158 assert(this_oop->is_oop_or_null(), "expected an oop or NULL"); 7159 HeapWord* addr = (HeapWord*)this_oop; 7160 if (_span.contains(addr) && !_verification_bm->isMarked(addr)) { 7161 // Oop lies in _span and isn't yet grey or black 7162 _verification_bm->mark(addr); // now grey 7163 if (!_cms_bm->isMarked(addr)) { 7164 oop(addr)->print(); 7165 gclog_or_tty->print_cr(" ("INTPTR_FORMAT" should have been marked)", addr); 7166 fatal("... aborting"); 7167 } 7168 7169 if (!_mark_stack->push(this_oop)) { // stack overflow 7170 if (PrintCMSStatistics != 0) { 7171 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 7172 SIZE_FORMAT, _mark_stack->capacity()); 7173 } 7174 assert(_mark_stack->isFull(), "Else push should have succeeded"); 7175 handle_stack_overflow(addr); 7176 } 7177 // anything including and to the right of _finger 7178 // will be scanned as we iterate over the remainder of the 7179 // bit map 7180 } 7181 } 7182 7183 PushOrMarkClosure::PushOrMarkClosure(CMSCollector* collector, 7184 MemRegion span, 7185 CMSBitMap* bitMap, CMSMarkStack* markStack, 7186 CMSMarkStack* revisitStack, 7187 HeapWord* finger, MarkFromRootsClosure* parent) : 7188 OopClosure(collector->ref_processor()), 7189 _collector(collector), 7190 _span(span), 7191 _bitMap(bitMap), 7192 _markStack(markStack), 7193 _revisitStack(revisitStack), 7194 _finger(finger), 7195 _parent(parent), 7196 _should_remember_klasses(collector->cms_should_unload_classes()) 7197 { } 7198 7199 Par_PushOrMarkClosure::Par_PushOrMarkClosure(CMSCollector* collector, 7200 MemRegion span, 7201 CMSBitMap* bit_map, 7202 OopTaskQueue* work_queue, 7203 CMSMarkStack* overflow_stack, 7204 CMSMarkStack* revisit_stack, 7205 HeapWord* finger, 7206 HeapWord** global_finger_addr, 7207 Par_MarkFromRootsClosure* parent) : 7208 OopClosure(collector->ref_processor()), 7209 _collector(collector), 7210 _whole_span(collector->_span), 7211 _span(span), 7212 _bit_map(bit_map), 7213 _work_queue(work_queue), 7214 _overflow_stack(overflow_stack), 7215 _revisit_stack(revisit_stack), 7216 _finger(finger), 7217 _global_finger_addr(global_finger_addr), 7218 _parent(parent), 7219 _should_remember_klasses(collector->cms_should_unload_classes()) 7220 { } 7221 7222 7223 void CMSCollector::lower_restart_addr(HeapWord* low) { 7224 assert(_span.contains(low), "Out of bounds addr"); 7225 if (_restart_addr == NULL) { 7226 _restart_addr = low; 7227 } else { 7228 _restart_addr = MIN2(_restart_addr, low); 7229 } 7230 } 7231 7232 // Upon stack overflow, we discard (part of) the stack, 7233 // remembering the least address amongst those discarded 7234 // in CMSCollector's _restart_address. 7235 void PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 7236 // Remember the least grey address discarded 7237 HeapWord* ra = (HeapWord*)_markStack->least_value(lost); 7238 _collector->lower_restart_addr(ra); 7239 _markStack->reset(); // discard stack contents 7240 _markStack->expand(); // expand the stack if possible 7241 } 7242 7243 // Upon stack overflow, we discard (part of) the stack, 7244 // remembering the least address amongst those discarded 7245 // in CMSCollector's _restart_address. 7246 void Par_PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 7247 // We need to do this under a mutex to prevent other 7248 // workers from interfering with the expansion below. 7249 MutexLockerEx ml(_overflow_stack->par_lock(), 7250 Mutex::_no_safepoint_check_flag); 7251 // Remember the least grey address discarded 7252 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 7253 _collector->lower_restart_addr(ra); 7254 _overflow_stack->reset(); // discard stack contents 7255 _overflow_stack->expand(); // expand the stack if possible 7256 } 7257 7258 7259 void PushOrMarkClosure::do_oop(oop* p) { 7260 oop thisOop = *p; 7261 // Ignore mark word because we are running concurrent with mutators. 7262 assert(thisOop->is_oop_or_null(true), "expected an oop or NULL"); 7263 HeapWord* addr = (HeapWord*)thisOop; 7264 if (_span.contains(addr) && !_bitMap->isMarked(addr)) { 7265 // Oop lies in _span and isn't yet grey or black 7266 _bitMap->mark(addr); // now grey 7267 if (addr < _finger) { 7268 // the bit map iteration has already either passed, or 7269 // sampled, this bit in the bit map; we'll need to 7270 // use the marking stack to scan this oop's oops. 7271 bool simulate_overflow = false; 7272 NOT_PRODUCT( 7273 if (CMSMarkStackOverflowALot && 7274 _collector->simulate_overflow()) { 7275 // simulate a stack overflow 7276 simulate_overflow = true; 7277 } 7278 ) 7279 if (simulate_overflow || !_markStack->push(thisOop)) { // stack overflow 7280 if (PrintCMSStatistics != 0) { 7281 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 7282 SIZE_FORMAT, _markStack->capacity()); 7283 } 7284 assert(simulate_overflow || _markStack->isFull(), "Else push should have succeeded"); 7285 handle_stack_overflow(addr); 7286 } 7287 } 7288 // anything including and to the right of _finger 7289 // will be scanned as we iterate over the remainder of the 7290 // bit map 7291 do_yield_check(); 7292 } 7293 } 7294 7295 void Par_PushOrMarkClosure::do_oop(oop* p) { 7296 oop this_oop = *p; 7297 // Ignore mark word because we are running concurrent with mutators. 7298 assert(this_oop->is_oop_or_null(true), "expected an oop or NULL"); 7299 HeapWord* addr = (HeapWord*)this_oop; 7300 if (_whole_span.contains(addr) && !_bit_map->isMarked(addr)) { 7301 // Oop lies in _span and isn't yet grey or black 7302 // We read the global_finger (volatile read) strictly after marking oop 7303 bool res = _bit_map->par_mark(addr); // now grey 7304 volatile HeapWord** gfa = (volatile HeapWord**)_global_finger_addr; 7305 // Should we push this marked oop on our stack? 7306 // -- if someone else marked it, nothing to do 7307 // -- if target oop is above global finger nothing to do 7308 // -- if target oop is in chunk and above local finger 7309 // then nothing to do 7310 // -- else push on work queue 7311 if ( !res // someone else marked it, they will deal with it 7312 || (addr >= *gfa) // will be scanned in a later task 7313 || (_span.contains(addr) && addr >= _finger)) { // later in this chunk 7314 return; 7315 } 7316 // the bit map iteration has already either passed, or 7317 // sampled, this bit in the bit map; we'll need to 7318 // use the marking stack to scan this oop's oops. 7319 bool simulate_overflow = false; 7320 NOT_PRODUCT( 7321 if (CMSMarkStackOverflowALot && 7322 _collector->simulate_overflow()) { 7323 // simulate a stack overflow 7324 simulate_overflow = true; 7325 } 7326 ) 7327 if (simulate_overflow || 7328 !(_work_queue->push(this_oop) || _overflow_stack->par_push(this_oop))) { 7329 // stack overflow 7330 if (PrintCMSStatistics != 0) { 7331 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 7332 SIZE_FORMAT, _overflow_stack->capacity()); 7333 } 7334 // We cannot assert that the overflow stack is full because 7335 // it may have been emptied since. 7336 assert(simulate_overflow || 7337 _work_queue->size() == _work_queue->max_elems(), 7338 "Else push should have succeeded"); 7339 handle_stack_overflow(addr); 7340 } 7341 do_yield_check(); 7342 } 7343 } 7344 7345 7346 PushAndMarkClosure::PushAndMarkClosure(CMSCollector* collector, 7347 MemRegion span, 7348 ReferenceProcessor* rp, 7349 CMSBitMap* bit_map, 7350 CMSBitMap* mod_union_table, 7351 CMSMarkStack* mark_stack, 7352 CMSMarkStack* revisit_stack, 7353 bool concurrent_precleaning): 7354 OopClosure(rp), 7355 _collector(collector), 7356 _span(span), 7357 _bit_map(bit_map), 7358 _mod_union_table(mod_union_table), 7359 _mark_stack(mark_stack), 7360 _revisit_stack(revisit_stack), 7361 _concurrent_precleaning(concurrent_precleaning), 7362 _should_remember_klasses(collector->cms_should_unload_classes()) 7363 { 7364 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 7365 } 7366 7367 // Grey object rescan during pre-cleaning and second checkpoint phases -- 7368 // the non-parallel version (the parallel version appears further below.) 7369 void PushAndMarkClosure::do_oop(oop* p) { 7370 oop this_oop = *p; 7371 // Ignore mark word verification. If during concurrent precleaning 7372 // the object monitor may be locked. If during the checkpoint 7373 // phases, the object may already have been reached by a different 7374 // path and may be at the end of the global overflow list (so 7375 // the mark word may be NULL). 7376 assert(this_oop->is_oop_or_null(true/* ignore mark word */), 7377 "expected an oop or NULL"); 7378 HeapWord* addr = (HeapWord*)this_oop; 7379 // Check if oop points into the CMS generation 7380 // and is not marked 7381 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 7382 // a white object ... 7383 _bit_map->mark(addr); // ... now grey 7384 // push on the marking stack (grey set) 7385 bool simulate_overflow = false; 7386 NOT_PRODUCT( 7387 if (CMSMarkStackOverflowALot && 7388 _collector->simulate_overflow()) { 7389 // simulate a stack overflow 7390 simulate_overflow = true; 7391 } 7392 ) 7393 if (simulate_overflow || !_mark_stack->push(this_oop)) { 7394 if (_concurrent_precleaning) { 7395 // During precleaning we can just dirty the appropriate card 7396 // in the mod union table, thus ensuring that the object remains 7397 // in the grey set and continue. Note that no one can be intefering 7398 // with us in this action of dirtying the mod union table, so 7399 // no locking is required. 7400 _mod_union_table->mark(addr); 7401 _collector->_ser_pmc_preclean_ovflw++; 7402 } else { 7403 // During the remark phase, we need to remember this oop 7404 // in the overflow list. 7405 _collector->push_on_overflow_list(this_oop); 7406 _collector->_ser_pmc_remark_ovflw++; 7407 } 7408 } 7409 } 7410 } 7411 7412 Par_PushAndMarkClosure::Par_PushAndMarkClosure(CMSCollector* collector, 7413 MemRegion span, 7414 ReferenceProcessor* rp, 7415 CMSBitMap* bit_map, 7416 OopTaskQueue* work_queue, 7417 CMSMarkStack* revisit_stack): 7418 OopClosure(rp), 7419 _collector(collector), 7420 _span(span), 7421 _bit_map(bit_map), 7422 _work_queue(work_queue), 7423 _revisit_stack(revisit_stack), 7424 _should_remember_klasses(collector->cms_should_unload_classes()) 7425 { 7426 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 7427 } 7428 7429 // Grey object rescan during second checkpoint phase -- 7430 // the parallel version. 7431 void Par_PushAndMarkClosure::do_oop(oop* p) { 7432 oop this_oop = *p; 7433 // In the assert below, we ignore the mark word because 7434 // this oop may point to an already visited object that is 7435 // on the overflow stack (in which case the mark word has 7436 // been hijacked for chaining into the overflow stack -- 7437 // if this is the last object in the overflow stack then 7438 // its mark word will be NULL). Because this object may 7439 // have been subsequently popped off the global overflow 7440 // stack, and the mark word possibly restored to the prototypical 7441 // value, by the time we get to examined this failing assert in 7442 // the debugger, is_oop_or_null(false) may subsequently start 7443 // to hold. 7444 assert(this_oop->is_oop_or_null(true), 7445 "expected an oop or NULL"); 7446 HeapWord* addr = (HeapWord*)this_oop; 7447 // Check if oop points into the CMS generation 7448 // and is not marked 7449 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 7450 // a white object ... 7451 // If we manage to "claim" the object, by being the 7452 // first thread to mark it, then we push it on our 7453 // marking stack 7454 if (_bit_map->par_mark(addr)) { // ... now grey 7455 // push on work queue (grey set) 7456 bool simulate_overflow = false; 7457 NOT_PRODUCT( 7458 if (CMSMarkStackOverflowALot && 7459 _collector->par_simulate_overflow()) { 7460 // simulate a stack overflow 7461 simulate_overflow = true; 7462 } 7463 ) 7464 if (simulate_overflow || !_work_queue->push(this_oop)) { 7465 _collector->par_push_on_overflow_list(this_oop); 7466 _collector->_par_pmc_remark_ovflw++; // imprecise OK: no need to CAS 7467 } 7468 } // Else, some other thread got there first 7469 } 7470 } 7471 7472 void PushAndMarkClosure::remember_klass(Klass* k) { 7473 if (!_revisit_stack->push(oop(k))) { 7474 fatal("Revisit stack overflowed in PushAndMarkClosure"); 7475 } 7476 } 7477 7478 void Par_PushAndMarkClosure::remember_klass(Klass* k) { 7479 if (!_revisit_stack->par_push(oop(k))) { 7480 fatal("Revist stack overflowed in Par_PushAndMarkClosure"); 7481 } 7482 } 7483 7484 void CMSPrecleanRefsYieldClosure::do_yield_work() { 7485 Mutex* bml = _collector->bitMapLock(); 7486 assert_lock_strong(bml); 7487 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 7488 "CMS thread should hold CMS token"); 7489 7490 bml->unlock(); 7491 ConcurrentMarkSweepThread::desynchronize(true); 7492 7493 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7494 7495 _collector->stopTimer(); 7496 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 7497 if (PrintCMSStatistics != 0) { 7498 _collector->incrementYields(); 7499 } 7500 _collector->icms_wait(); 7501 7502 // See the comment in coordinator_yield() 7503 for (unsigned i = 0; i < CMSYieldSleepCount && 7504 ConcurrentMarkSweepThread::should_yield() && 7505 !CMSCollector::foregroundGCIsActive(); ++i) { 7506 os::sleep(Thread::current(), 1, false); 7507 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7508 } 7509 7510 ConcurrentMarkSweepThread::synchronize(true); 7511 bml->lock(); 7512 7513 _collector->startTimer(); 7514 } 7515 7516 bool CMSPrecleanRefsYieldClosure::should_return() { 7517 if (ConcurrentMarkSweepThread::should_yield()) { 7518 do_yield_work(); 7519 } 7520 return _collector->foregroundGCIsActive(); 7521 } 7522 7523 void MarkFromDirtyCardsClosure::do_MemRegion(MemRegion mr) { 7524 assert(((size_t)mr.start())%CardTableModRefBS::card_size_in_words == 0, 7525 "mr should be aligned to start at a card boundary"); 7526 // We'd like to assert: 7527 // assert(mr.word_size()%CardTableModRefBS::card_size_in_words == 0, 7528 // "mr should be a range of cards"); 7529 // However, that would be too strong in one case -- the last 7530 // partition ends at _unallocated_block which, in general, can be 7531 // an arbitrary boundary, not necessarily card aligned. 7532 if (PrintCMSStatistics != 0) { 7533 _num_dirty_cards += 7534 mr.word_size()/CardTableModRefBS::card_size_in_words; 7535 } 7536 _space->object_iterate_mem(mr, &_scan_cl); 7537 } 7538 7539 SweepClosure::SweepClosure(CMSCollector* collector, 7540 ConcurrentMarkSweepGeneration* g, 7541 CMSBitMap* bitMap, bool should_yield) : 7542 _collector(collector), 7543 _g(g), 7544 _sp(g->cmsSpace()), 7545 _limit(_sp->sweep_limit()), 7546 _freelistLock(_sp->freelistLock()), 7547 _bitMap(bitMap), 7548 _yield(should_yield), 7549 _inFreeRange(false), // No free range at beginning of sweep 7550 _freeRangeInFreeLists(false), // No free range at beginning of sweep 7551 _lastFreeRangeCoalesced(false), 7552 _freeFinger(g->used_region().start()) 7553 { 7554 NOT_PRODUCT( 7555 _numObjectsFreed = 0; 7556 _numWordsFreed = 0; 7557 _numObjectsLive = 0; 7558 _numWordsLive = 0; 7559 _numObjectsAlreadyFree = 0; 7560 _numWordsAlreadyFree = 0; 7561 _last_fc = NULL; 7562 7563 _sp->initializeIndexedFreeListArrayReturnedBytes(); 7564 _sp->dictionary()->initializeDictReturnedBytes(); 7565 ) 7566 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7567 "sweep _limit out of bounds"); 7568 if (CMSTraceSweeper) { 7569 gclog_or_tty->print("\n====================\nStarting new sweep\n"); 7570 } 7571 } 7572 7573 // We need this destructor to reclaim any space at the end 7574 // of the space, which do_blk below may not have added back to 7575 // the free lists. [basically dealing with the "fringe effect"] 7576 SweepClosure::~SweepClosure() { 7577 assert_lock_strong(_freelistLock); 7578 // this should be treated as the end of a free run if any 7579 // The current free range should be returned to the free lists 7580 // as one coalesced chunk. 7581 if (inFreeRange()) { 7582 flushCurFreeChunk(freeFinger(), 7583 pointer_delta(_limit, freeFinger())); 7584 assert(freeFinger() < _limit, "the finger pointeth off base"); 7585 if (CMSTraceSweeper) { 7586 gclog_or_tty->print("destructor:"); 7587 gclog_or_tty->print("Sweep:put_free_blk 0x%x ("SIZE_FORMAT") " 7588 "[coalesced:"SIZE_FORMAT"]\n", 7589 freeFinger(), pointer_delta(_limit, freeFinger()), 7590 lastFreeRangeCoalesced()); 7591 } 7592 } 7593 NOT_PRODUCT( 7594 if (Verbose && PrintGC) { 7595 gclog_or_tty->print("Collected "SIZE_FORMAT" objects, " 7596 SIZE_FORMAT " bytes", 7597 _numObjectsFreed, _numWordsFreed*sizeof(HeapWord)); 7598 gclog_or_tty->print_cr("\nLive "SIZE_FORMAT" objects, " 7599 SIZE_FORMAT" bytes " 7600 "Already free "SIZE_FORMAT" objects, "SIZE_FORMAT" bytes", 7601 _numObjectsLive, _numWordsLive*sizeof(HeapWord), 7602 _numObjectsAlreadyFree, _numWordsAlreadyFree*sizeof(HeapWord)); 7603 size_t totalBytes = (_numWordsFreed + _numWordsLive + _numWordsAlreadyFree) * 7604 sizeof(HeapWord); 7605 gclog_or_tty->print_cr("Total sweep: "SIZE_FORMAT" bytes", totalBytes); 7606 7607 if (PrintCMSStatistics && CMSVerifyReturnedBytes) { 7608 size_t indexListReturnedBytes = _sp->sumIndexedFreeListArrayReturnedBytes(); 7609 size_t dictReturnedBytes = _sp->dictionary()->sumDictReturnedBytes(); 7610 size_t returnedBytes = indexListReturnedBytes + dictReturnedBytes; 7611 gclog_or_tty->print("Returned "SIZE_FORMAT" bytes", returnedBytes); 7612 gclog_or_tty->print(" Indexed List Returned "SIZE_FORMAT" bytes", 7613 indexListReturnedBytes); 7614 gclog_or_tty->print_cr(" Dictionary Returned "SIZE_FORMAT" bytes", 7615 dictReturnedBytes); 7616 } 7617 } 7618 ) 7619 // Now, in debug mode, just null out the sweep_limit 7620 NOT_PRODUCT(_sp->clear_sweep_limit();) 7621 if (CMSTraceSweeper) { 7622 gclog_or_tty->print("end of sweep\n================\n"); 7623 } 7624 } 7625 7626 void SweepClosure::initialize_free_range(HeapWord* freeFinger, 7627 bool freeRangeInFreeLists) { 7628 if (CMSTraceSweeper) { 7629 gclog_or_tty->print("---- Start free range 0x%x with free block [%d] (%d)\n", 7630 freeFinger, _sp->block_size(freeFinger), 7631 freeRangeInFreeLists); 7632 } 7633 assert(!inFreeRange(), "Trampling existing free range"); 7634 set_inFreeRange(true); 7635 set_lastFreeRangeCoalesced(false); 7636 7637 set_freeFinger(freeFinger); 7638 set_freeRangeInFreeLists(freeRangeInFreeLists); 7639 if (CMSTestInFreeList) { 7640 if (freeRangeInFreeLists) { 7641 FreeChunk* fc = (FreeChunk*) freeFinger; 7642 assert(fc->isFree(), "A chunk on the free list should be free."); 7643 assert(fc->size() > 0, "Free range should have a size"); 7644 assert(_sp->verifyChunkInFreeLists(fc), "Chunk is not in free lists"); 7645 } 7646 } 7647 } 7648 7649 // Note that the sweeper runs concurrently with mutators. Thus, 7650 // it is possible for direct allocation in this generation to happen 7651 // in the middle of the sweep. Note that the sweeper also coalesces 7652 // contiguous free blocks. Thus, unless the sweeper and the allocator 7653 // synchronize appropriately freshly allocated blocks may get swept up. 7654 // This is accomplished by the sweeper locking the free lists while 7655 // it is sweeping. Thus blocks that are determined to be free are 7656 // indeed free. There is however one additional complication: 7657 // blocks that have been allocated since the final checkpoint and 7658 // mark, will not have been marked and so would be treated as 7659 // unreachable and swept up. To prevent this, the allocator marks 7660 // the bit map when allocating during the sweep phase. This leads, 7661 // however, to a further complication -- objects may have been allocated 7662 // but not yet initialized -- in the sense that the header isn't yet 7663 // installed. The sweeper can not then determine the size of the block 7664 // in order to skip over it. To deal with this case, we use a technique 7665 // (due to Printezis) to encode such uninitialized block sizes in the 7666 // bit map. Since the bit map uses a bit per every HeapWord, but the 7667 // CMS generation has a minimum object size of 3 HeapWords, it follows 7668 // that "normal marks" won't be adjacent in the bit map (there will 7669 // always be at least two 0 bits between successive 1 bits). We make use 7670 // of these "unused" bits to represent uninitialized blocks -- the bit 7671 // corresponding to the start of the uninitialized object and the next 7672 // bit are both set. Finally, a 1 bit marks the end of the object that 7673 // started with the two consecutive 1 bits to indicate its potentially 7674 // uninitialized state. 7675 7676 size_t SweepClosure::do_blk_careful(HeapWord* addr) { 7677 FreeChunk* fc = (FreeChunk*)addr; 7678 size_t res; 7679 7680 // check if we are done sweepinrg 7681 if (addr == _limit) { // we have swept up to the limit, do nothing more 7682 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7683 "sweep _limit out of bounds"); 7684 // help the closure application finish 7685 return pointer_delta(_sp->end(), _limit); 7686 } 7687 assert(addr <= _limit, "sweep invariant"); 7688 7689 // check if we should yield 7690 do_yield_check(addr); 7691 if (fc->isFree()) { 7692 // Chunk that is already free 7693 res = fc->size(); 7694 doAlreadyFreeChunk(fc); 7695 debug_only(_sp->verifyFreeLists()); 7696 assert(res == fc->size(), "Don't expect the size to change"); 7697 NOT_PRODUCT( 7698 _numObjectsAlreadyFree++; 7699 _numWordsAlreadyFree += res; 7700 ) 7701 NOT_PRODUCT(_last_fc = fc;) 7702 } else if (!_bitMap->isMarked(addr)) { 7703 // Chunk is fresh garbage 7704 res = doGarbageChunk(fc); 7705 debug_only(_sp->verifyFreeLists()); 7706 NOT_PRODUCT( 7707 _numObjectsFreed++; 7708 _numWordsFreed += res; 7709 ) 7710 } else { 7711 // Chunk that is alive. 7712 res = doLiveChunk(fc); 7713 debug_only(_sp->verifyFreeLists()); 7714 NOT_PRODUCT( 7715 _numObjectsLive++; 7716 _numWordsLive += res; 7717 ) 7718 } 7719 return res; 7720 } 7721 7722 // For the smart allocation, record following 7723 // split deaths - a free chunk is removed from its free list because 7724 // it is being split into two or more chunks. 7725 // split birth - a free chunk is being added to its free list because 7726 // a larger free chunk has been split and resulted in this free chunk. 7727 // coal death - a free chunk is being removed from its free list because 7728 // it is being coalesced into a large free chunk. 7729 // coal birth - a free chunk is being added to its free list because 7730 // it was created when two or more free chunks where coalesced into 7731 // this free chunk. 7732 // 7733 // These statistics are used to determine the desired number of free 7734 // chunks of a given size. The desired number is chosen to be relative 7735 // to the end of a CMS sweep. The desired number at the end of a sweep 7736 // is the 7737 // count-at-end-of-previous-sweep (an amount that was enough) 7738 // - count-at-beginning-of-current-sweep (the excess) 7739 // + split-births (gains in this size during interval) 7740 // - split-deaths (demands on this size during interval) 7741 // where the interval is from the end of one sweep to the end of the 7742 // next. 7743 // 7744 // When sweeping the sweeper maintains an accumulated chunk which is 7745 // the chunk that is made up of chunks that have been coalesced. That 7746 // will be termed the left-hand chunk. A new chunk of garbage that 7747 // is being considered for coalescing will be referred to as the 7748 // right-hand chunk. 7749 // 7750 // When making a decision on whether to coalesce a right-hand chunk with 7751 // the current left-hand chunk, the current count vs. the desired count 7752 // of the left-hand chunk is considered. Also if the right-hand chunk 7753 // is near the large chunk at the end of the heap (see 7754 // ConcurrentMarkSweepGeneration::isNearLargestChunk()), then the 7755 // left-hand chunk is coalesced. 7756 // 7757 // When making a decision about whether to split a chunk, the desired count 7758 // vs. the current count of the candidate to be split is also considered. 7759 // If the candidate is underpopulated (currently fewer chunks than desired) 7760 // a chunk of an overpopulated (currently more chunks than desired) size may 7761 // be chosen. The "hint" associated with a free list, if non-null, points 7762 // to a free list which may be overpopulated. 7763 // 7764 7765 void SweepClosure::doAlreadyFreeChunk(FreeChunk* fc) { 7766 size_t size = fc->size(); 7767 // Chunks that cannot be coalesced are not in the 7768 // free lists. 7769 if (CMSTestInFreeList && !fc->cantCoalesce()) { 7770 assert(_sp->verifyChunkInFreeLists(fc), 7771 "free chunk should be in free lists"); 7772 } 7773 // a chunk that is already free, should not have been 7774 // marked in the bit map 7775 HeapWord* addr = (HeapWord*) fc; 7776 assert(!_bitMap->isMarked(addr), "free chunk should be unmarked"); 7777 // Verify that the bit map has no bits marked between 7778 // addr and purported end of this block. 7779 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 7780 7781 // Some chunks cannot be coalesced in under any circumstances. 7782 // See the definition of cantCoalesce(). 7783 if (!fc->cantCoalesce()) { 7784 // This chunk can potentially be coalesced. 7785 if (_sp->adaptive_freelists()) { 7786 // All the work is done in 7787 doPostIsFreeOrGarbageChunk(fc, size); 7788 } else { // Not adaptive free lists 7789 // this is a free chunk that can potentially be coalesced by the sweeper; 7790 if (!inFreeRange()) { 7791 // if the next chunk is a free block that can't be coalesced 7792 // it doesn't make sense to remove this chunk from the free lists 7793 FreeChunk* nextChunk = (FreeChunk*)(addr + size); 7794 assert((HeapWord*)nextChunk <= _limit, "sweep invariant"); 7795 if ((HeapWord*)nextChunk < _limit && // there's a next chunk... 7796 nextChunk->isFree() && // which is free... 7797 nextChunk->cantCoalesce()) { // ... but cant be coalesced 7798 // nothing to do 7799 } else { 7800 // Potentially the start of a new free range: 7801 // Don't eagerly remove it from the free lists. 7802 // No need to remove it if it will just be put 7803 // back again. (Also from a pragmatic point of view 7804 // if it is a free block in a region that is beyond 7805 // any allocated blocks, an assertion will fail) 7806 // Remember the start of a free run. 7807 initialize_free_range(addr, true); 7808 // end - can coalesce with next chunk 7809 } 7810 } else { 7811 // the midst of a free range, we are coalescing 7812 debug_only(record_free_block_coalesced(fc);) 7813 if (CMSTraceSweeper) { 7814 gclog_or_tty->print(" -- pick up free block 0x%x (%d)\n", fc, size); 7815 } 7816 // remove it from the free lists 7817 _sp->removeFreeChunkFromFreeLists(fc); 7818 set_lastFreeRangeCoalesced(true); 7819 // If the chunk is being coalesced and the current free range is 7820 // in the free lists, remove the current free range so that it 7821 // will be returned to the free lists in its entirety - all 7822 // the coalesced pieces included. 7823 if (freeRangeInFreeLists()) { 7824 FreeChunk* ffc = (FreeChunk*) freeFinger(); 7825 assert(ffc->size() == pointer_delta(addr, freeFinger()), 7826 "Size of free range is inconsistent with chunk size."); 7827 if (CMSTestInFreeList) { 7828 assert(_sp->verifyChunkInFreeLists(ffc), 7829 "free range is not in free lists"); 7830 } 7831 _sp->removeFreeChunkFromFreeLists(ffc); 7832 set_freeRangeInFreeLists(false); 7833 } 7834 } 7835 } 7836 } else { 7837 // Code path common to both original and adaptive free lists. 7838 7839 // cant coalesce with previous block; this should be treated 7840 // as the end of a free run if any 7841 if (inFreeRange()) { 7842 // we kicked some butt; time to pick up the garbage 7843 assert(freeFinger() < addr, "the finger pointeth off base"); 7844 flushCurFreeChunk(freeFinger(), pointer_delta(addr, freeFinger())); 7845 } 7846 // else, nothing to do, just continue 7847 } 7848 } 7849 7850 size_t SweepClosure::doGarbageChunk(FreeChunk* fc) { 7851 // This is a chunk of garbage. It is not in any free list. 7852 // Add it to a free list or let it possibly be coalesced into 7853 // a larger chunk. 7854 HeapWord* addr = (HeapWord*) fc; 7855 size_t size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 7856 7857 if (_sp->adaptive_freelists()) { 7858 // Verify that the bit map has no bits marked between 7859 // addr and purported end of just dead object. 7860 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 7861 7862 doPostIsFreeOrGarbageChunk(fc, size); 7863 } else { 7864 if (!inFreeRange()) { 7865 // start of a new free range 7866 assert(size > 0, "A free range should have a size"); 7867 initialize_free_range(addr, false); 7868 7869 } else { 7870 // this will be swept up when we hit the end of the 7871 // free range 7872 if (CMSTraceSweeper) { 7873 gclog_or_tty->print(" -- pick up garbage 0x%x (%d) \n", fc, size); 7874 } 7875 // If the chunk is being coalesced and the current free range is 7876 // in the free lists, remove the current free range so that it 7877 // will be returned to the free lists in its entirety - all 7878 // the coalesced pieces included. 7879 if (freeRangeInFreeLists()) { 7880 FreeChunk* ffc = (FreeChunk*)freeFinger(); 7881 assert(ffc->size() == pointer_delta(addr, freeFinger()), 7882 "Size of free range is inconsistent with chunk size."); 7883 if (CMSTestInFreeList) { 7884 assert(_sp->verifyChunkInFreeLists(ffc), 7885 "free range is not in free lists"); 7886 } 7887 _sp->removeFreeChunkFromFreeLists(ffc); 7888 set_freeRangeInFreeLists(false); 7889 } 7890 set_lastFreeRangeCoalesced(true); 7891 } 7892 // this will be swept up when we hit the end of the free range 7893 7894 // Verify that the bit map has no bits marked between 7895 // addr and purported end of just dead object. 7896 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 7897 } 7898 return size; 7899 } 7900 7901 size_t SweepClosure::doLiveChunk(FreeChunk* fc) { 7902 HeapWord* addr = (HeapWord*) fc; 7903 // The sweeper has just found a live object. Return any accumulated 7904 // left hand chunk to the free lists. 7905 if (inFreeRange()) { 7906 if (_sp->adaptive_freelists()) { 7907 flushCurFreeChunk(freeFinger(), 7908 pointer_delta(addr, freeFinger())); 7909 } else { // not adaptive freelists 7910 set_inFreeRange(false); 7911 // Add the free range back to the free list if it is not already 7912 // there. 7913 if (!freeRangeInFreeLists()) { 7914 assert(freeFinger() < addr, "the finger pointeth off base"); 7915 if (CMSTraceSweeper) { 7916 gclog_or_tty->print("Sweep:put_free_blk 0x%x (%d) " 7917 "[coalesced:%d]\n", 7918 freeFinger(), pointer_delta(addr, freeFinger()), 7919 lastFreeRangeCoalesced()); 7920 } 7921 _sp->addChunkAndRepairOffsetTable(freeFinger(), 7922 pointer_delta(addr, freeFinger()), lastFreeRangeCoalesced()); 7923 } 7924 } 7925 } 7926 7927 // Common code path for original and adaptive free lists. 7928 7929 // this object is live: we'd normally expect this to be 7930 // an oop, and like to assert the following: 7931 // assert(oop(addr)->is_oop(), "live block should be an oop"); 7932 // However, as we commented above, this may be an object whose 7933 // header hasn't yet been initialized. 7934 size_t size; 7935 assert(_bitMap->isMarked(addr), "Tautology for this control point"); 7936 if (_bitMap->isMarked(addr + 1)) { 7937 // Determine the size from the bit map, rather than trying to 7938 // compute it from the object header. 7939 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 7940 size = pointer_delta(nextOneAddr + 1, addr); 7941 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 7942 "alignment problem"); 7943 7944 #ifdef DEBUG 7945 if (oop(addr)->klass() != NULL && 7946 ( !_collector->cms_should_unload_classes() 7947 || oop(addr)->is_parsable())) { 7948 // Ignore mark word because we are running concurrent with mutators 7949 assert(oop(addr)->is_oop(true), "live block should be an oop"); 7950 assert(size == 7951 CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()), 7952 "P-mark and computed size do not agree"); 7953 } 7954 #endif 7955 7956 } else { 7957 // This should be an initialized object that's alive. 7958 assert(oop(addr)->klass() != NULL && 7959 (!_collector->cms_should_unload_classes() 7960 || oop(addr)->is_parsable()), 7961 "Should be an initialized object"); 7962 // Ignore mark word because we are running concurrent with mutators 7963 assert(oop(addr)->is_oop(true), "live block should be an oop"); 7964 // Verify that the bit map has no bits marked between 7965 // addr and purported end of this block. 7966 size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 7967 assert(size >= 3, "Necessary for Printezis marks to work"); 7968 assert(!_bitMap->isMarked(addr+1), "Tautology for this control point"); 7969 DEBUG_ONLY(_bitMap->verifyNoOneBitsInRange(addr+2, addr+size);) 7970 } 7971 return size; 7972 } 7973 7974 void SweepClosure::doPostIsFreeOrGarbageChunk(FreeChunk* fc, 7975 size_t chunkSize) { 7976 // doPostIsFreeOrGarbageChunk() should only be called in the smart allocation 7977 // scheme. 7978 bool fcInFreeLists = fc->isFree(); 7979 assert(_sp->adaptive_freelists(), "Should only be used in this case."); 7980 assert((HeapWord*)fc <= _limit, "sweep invariant"); 7981 if (CMSTestInFreeList && fcInFreeLists) { 7982 assert(_sp->verifyChunkInFreeLists(fc), 7983 "free chunk is not in free lists"); 7984 } 7985 7986 7987 if (CMSTraceSweeper) { 7988 gclog_or_tty->print_cr(" -- pick up another chunk at 0x%x (%d)", fc, chunkSize); 7989 } 7990 7991 HeapWord* addr = (HeapWord*) fc; 7992 7993 bool coalesce; 7994 size_t left = pointer_delta(addr, freeFinger()); 7995 size_t right = chunkSize; 7996 switch (FLSCoalescePolicy) { 7997 // numeric value forms a coalition aggressiveness metric 7998 case 0: { // never coalesce 7999 coalesce = false; 8000 break; 8001 } 8002 case 1: { // coalesce if left & right chunks on overpopulated lists 8003 coalesce = _sp->coalOverPopulated(left) && 8004 _sp->coalOverPopulated(right); 8005 break; 8006 } 8007 case 2: { // coalesce if left chunk on overpopulated list (default) 8008 coalesce = _sp->coalOverPopulated(left); 8009 break; 8010 } 8011 case 3: { // coalesce if left OR right chunk on overpopulated list 8012 coalesce = _sp->coalOverPopulated(left) || 8013 _sp->coalOverPopulated(right); 8014 break; 8015 } 8016 case 4: { // always coalesce 8017 coalesce = true; 8018 break; 8019 } 8020 default: 8021 ShouldNotReachHere(); 8022 } 8023 8024 // Should the current free range be coalesced? 8025 // If the chunk is in a free range and either we decided to coalesce above 8026 // or the chunk is near the large block at the end of the heap 8027 // (isNearLargestChunk() returns true), then coalesce this chunk. 8028 bool doCoalesce = inFreeRange() && 8029 (coalesce || _g->isNearLargestChunk((HeapWord*)fc)); 8030 if (doCoalesce) { 8031 // Coalesce the current free range on the left with the new 8032 // chunk on the right. If either is on a free list, 8033 // it must be removed from the list and stashed in the closure. 8034 if (freeRangeInFreeLists()) { 8035 FreeChunk* ffc = (FreeChunk*)freeFinger(); 8036 assert(ffc->size() == pointer_delta(addr, freeFinger()), 8037 "Size of free range is inconsistent with chunk size."); 8038 if (CMSTestInFreeList) { 8039 assert(_sp->verifyChunkInFreeLists(ffc), 8040 "Chunk is not in free lists"); 8041 } 8042 _sp->coalDeath(ffc->size()); 8043 _sp->removeFreeChunkFromFreeLists(ffc); 8044 set_freeRangeInFreeLists(false); 8045 } 8046 if (fcInFreeLists) { 8047 _sp->coalDeath(chunkSize); 8048 assert(fc->size() == chunkSize, 8049 "The chunk has the wrong size or is not in the free lists"); 8050 _sp->removeFreeChunkFromFreeLists(fc); 8051 } 8052 set_lastFreeRangeCoalesced(true); 8053 } else { // not in a free range and/or should not coalesce 8054 // Return the current free range and start a new one. 8055 if (inFreeRange()) { 8056 // In a free range but cannot coalesce with the right hand chunk. 8057 // Put the current free range into the free lists. 8058 flushCurFreeChunk(freeFinger(), 8059 pointer_delta(addr, freeFinger())); 8060 } 8061 // Set up for new free range. Pass along whether the right hand 8062 // chunk is in the free lists. 8063 initialize_free_range((HeapWord*)fc, fcInFreeLists); 8064 } 8065 } 8066 void SweepClosure::flushCurFreeChunk(HeapWord* chunk, size_t size) { 8067 assert(inFreeRange(), "Should only be called if currently in a free range."); 8068 assert(size > 0, 8069 "A zero sized chunk cannot be added to the free lists."); 8070 if (!freeRangeInFreeLists()) { 8071 if(CMSTestInFreeList) { 8072 FreeChunk* fc = (FreeChunk*) chunk; 8073 fc->setSize(size); 8074 assert(!_sp->verifyChunkInFreeLists(fc), 8075 "chunk should not be in free lists yet"); 8076 } 8077 if (CMSTraceSweeper) { 8078 gclog_or_tty->print_cr(" -- add free block 0x%x (%d) to free lists", 8079 chunk, size); 8080 } 8081 // A new free range is going to be starting. The current 8082 // free range has not been added to the free lists yet or 8083 // was removed so add it back. 8084 // If the current free range was coalesced, then the death 8085 // of the free range was recorded. Record a birth now. 8086 if (lastFreeRangeCoalesced()) { 8087 _sp->coalBirth(size); 8088 } 8089 _sp->addChunkAndRepairOffsetTable(chunk, size, 8090 lastFreeRangeCoalesced()); 8091 } 8092 set_inFreeRange(false); 8093 set_freeRangeInFreeLists(false); 8094 } 8095 8096 // We take a break if we've been at this for a while, 8097 // so as to avoid monopolizing the locks involved. 8098 void SweepClosure::do_yield_work(HeapWord* addr) { 8099 // Return current free chunk being used for coalescing (if any) 8100 // to the appropriate freelist. After yielding, the next 8101 // free block encountered will start a coalescing range of 8102 // free blocks. If the next free block is adjacent to the 8103 // chunk just flushed, they will need to wait for the next 8104 // sweep to be coalesced. 8105 if (inFreeRange()) { 8106 flushCurFreeChunk(freeFinger(), pointer_delta(addr, freeFinger())); 8107 } 8108 8109 // First give up the locks, then yield, then re-lock. 8110 // We should probably use a constructor/destructor idiom to 8111 // do this unlock/lock or modify the MutexUnlocker class to 8112 // serve our purpose. XXX 8113 assert_lock_strong(_bitMap->lock()); 8114 assert_lock_strong(_freelistLock); 8115 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 8116 "CMS thread should hold CMS token"); 8117 _bitMap->lock()->unlock(); 8118 _freelistLock->unlock(); 8119 ConcurrentMarkSweepThread::desynchronize(true); 8120 ConcurrentMarkSweepThread::acknowledge_yield_request(); 8121 _collector->stopTimer(); 8122 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 8123 if (PrintCMSStatistics != 0) { 8124 _collector->incrementYields(); 8125 } 8126 _collector->icms_wait(); 8127 8128 // See the comment in coordinator_yield() 8129 for (unsigned i = 0; i < CMSYieldSleepCount && 8130 ConcurrentMarkSweepThread::should_yield() && 8131 !CMSCollector::foregroundGCIsActive(); ++i) { 8132 os::sleep(Thread::current(), 1, false); 8133 ConcurrentMarkSweepThread::acknowledge_yield_request(); 8134 } 8135 8136 ConcurrentMarkSweepThread::synchronize(true); 8137 _freelistLock->lock(); 8138 _bitMap->lock()->lock_without_safepoint_check(); 8139 _collector->startTimer(); 8140 } 8141 8142 #ifndef PRODUCT 8143 // This is actually very useful in a product build if it can 8144 // be called from the debugger. Compile it into the product 8145 // as needed. 8146 bool debug_verifyChunkInFreeLists(FreeChunk* fc) { 8147 return debug_cms_space->verifyChunkInFreeLists(fc); 8148 } 8149 8150 void SweepClosure::record_free_block_coalesced(FreeChunk* fc) const { 8151 if (CMSTraceSweeper) { 8152 gclog_or_tty->print("Sweep:coal_free_blk 0x%x (%d)\n", fc, fc->size()); 8153 } 8154 } 8155 #endif 8156 8157 // CMSIsAliveClosure 8158 bool CMSIsAliveClosure::do_object_b(oop obj) { 8159 HeapWord* addr = (HeapWord*)obj; 8160 return addr != NULL && 8161 (!_span.contains(addr) || _bit_map->isMarked(addr)); 8162 } 8163 8164 // CMSKeepAliveClosure: the serial version 8165 void CMSKeepAliveClosure::do_oop(oop* p) { 8166 oop this_oop = *p; 8167 HeapWord* addr = (HeapWord*)this_oop; 8168 if (_span.contains(addr) && 8169 !_bit_map->isMarked(addr)) { 8170 _bit_map->mark(addr); 8171 bool simulate_overflow = false; 8172 NOT_PRODUCT( 8173 if (CMSMarkStackOverflowALot && 8174 _collector->simulate_overflow()) { 8175 // simulate a stack overflow 8176 simulate_overflow = true; 8177 } 8178 ) 8179 if (simulate_overflow || !_mark_stack->push(this_oop)) { 8180 _collector->push_on_overflow_list(this_oop); 8181 _collector->_ser_kac_ovflw++; 8182 } 8183 } 8184 } 8185 8186 // CMSParKeepAliveClosure: a parallel version of the above. 8187 // The work queues are private to each closure (thread), 8188 // but (may be) available for stealing by other threads. 8189 void CMSParKeepAliveClosure::do_oop(oop* p) { 8190 oop this_oop = *p; 8191 HeapWord* addr = (HeapWord*)this_oop; 8192 if (_span.contains(addr) && 8193 !_bit_map->isMarked(addr)) { 8194 // In general, during recursive tracing, several threads 8195 // may be concurrently getting here; the first one to 8196 // "tag" it, claims it. 8197 if (_bit_map->par_mark(addr)) { 8198 bool res = _work_queue->push(this_oop); 8199 assert(res, "Low water mark should be much less than capacity"); 8200 // Do a recursive trim in the hope that this will keep 8201 // stack usage lower, but leave some oops for potential stealers 8202 trim_queue(_low_water_mark); 8203 } // Else, another thread got there first 8204 } 8205 } 8206 8207 void CMSParKeepAliveClosure::trim_queue(uint max) { 8208 while (_work_queue->size() > max) { 8209 oop new_oop; 8210 if (_work_queue->pop_local(new_oop)) { 8211 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 8212 assert(_bit_map->isMarked((HeapWord*)new_oop), 8213 "no white objects on this stack!"); 8214 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 8215 // iterate over the oops in this oop, marking and pushing 8216 // the ones in CMS heap (i.e. in _span). 8217 new_oop->oop_iterate(&_mark_and_push); 8218 } 8219 } 8220 } 8221 8222 void CMSInnerParMarkAndPushClosure::do_oop(oop* p) { 8223 oop this_oop = *p; 8224 HeapWord* addr = (HeapWord*)this_oop; 8225 if (_span.contains(addr) && 8226 !_bit_map->isMarked(addr)) { 8227 if (_bit_map->par_mark(addr)) { 8228 bool simulate_overflow = false; 8229 NOT_PRODUCT( 8230 if (CMSMarkStackOverflowALot && 8231 _collector->par_simulate_overflow()) { 8232 // simulate a stack overflow 8233 simulate_overflow = true; 8234 } 8235 ) 8236 if (simulate_overflow || !_work_queue->push(this_oop)) { 8237 _collector->par_push_on_overflow_list(this_oop); 8238 _collector->_par_kac_ovflw++; 8239 } 8240 } // Else another thread got there already 8241 } 8242 } 8243 8244 ////////////////////////////////////////////////////////////////// 8245 // CMSExpansionCause ///////////////////////////// 8246 ////////////////////////////////////////////////////////////////// 8247 const char* CMSExpansionCause::to_string(CMSExpansionCause::Cause cause) { 8248 switch (cause) { 8249 case _no_expansion: 8250 return "No expansion"; 8251 case _satisfy_free_ratio: 8252 return "Free ratio"; 8253 case _satisfy_promotion: 8254 return "Satisfy promotion"; 8255 case _satisfy_allocation: 8256 return "allocation"; 8257 case _allocate_par_lab: 8258 return "Par LAB"; 8259 case _allocate_par_spooling_space: 8260 return "Par Spooling Space"; 8261 case _adaptive_size_policy: 8262 return "Ergonomics"; 8263 default: 8264 return "unknown"; 8265 } 8266 } 8267 8268 void CMSDrainMarkingStackClosure::do_void() { 8269 // the max number to take from overflow list at a time 8270 const size_t num = _mark_stack->capacity()/4; 8271 while (!_mark_stack->isEmpty() || 8272 // if stack is empty, check the overflow list 8273 _collector->take_from_overflow_list(num, _mark_stack)) { 8274 oop this_oop = _mark_stack->pop(); 8275 HeapWord* addr = (HeapWord*)this_oop; 8276 assert(_span.contains(addr), "Should be within span"); 8277 assert(_bit_map->isMarked(addr), "Should be marked"); 8278 assert(this_oop->is_oop(), "Should be an oop"); 8279 this_oop->oop_iterate(_keep_alive); 8280 } 8281 } 8282 8283 void CMSParDrainMarkingStackClosure::do_void() { 8284 // drain queue 8285 trim_queue(0); 8286 } 8287 8288 // Trim our work_queue so its length is below max at return 8289 void CMSParDrainMarkingStackClosure::trim_queue(uint max) { 8290 while (_work_queue->size() > max) { 8291 oop new_oop; 8292 if (_work_queue->pop_local(new_oop)) { 8293 assert(new_oop->is_oop(), "Expected an oop"); 8294 assert(_bit_map->isMarked((HeapWord*)new_oop), 8295 "no white objects on this stack!"); 8296 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 8297 // iterate over the oops in this oop, marking and pushing 8298 // the ones in CMS heap (i.e. in _span). 8299 new_oop->oop_iterate(&_mark_and_push); 8300 } 8301 } 8302 } 8303 8304 //////////////////////////////////////////////////////////////////// 8305 // Support for Marking Stack Overflow list handling and related code 8306 //////////////////////////////////////////////////////////////////// 8307 // Much of the following code is similar in shape and spirit to the 8308 // code used in ParNewGC. We should try and share that code 8309 // as much as possible in the future. 8310 8311 #ifndef PRODUCT 8312 // Debugging support for CMSStackOverflowALot 8313 8314 // It's OK to call this multi-threaded; the worst thing 8315 // that can happen is that we'll get a bunch of closely 8316 // spaced simulated oveflows, but that's OK, in fact 8317 // probably good as it would exercise the overflow code 8318 // under contention. 8319 bool CMSCollector::simulate_overflow() { 8320 if (_overflow_counter-- <= 0) { // just being defensive 8321 _overflow_counter = CMSMarkStackOverflowInterval; 8322 return true; 8323 } else { 8324 return false; 8325 } 8326 } 8327 8328 bool CMSCollector::par_simulate_overflow() { 8329 return simulate_overflow(); 8330 } 8331 #endif 8332 8333 // Single-threaded 8334 bool CMSCollector::take_from_overflow_list(size_t num, CMSMarkStack* stack) { 8335 assert(stack->isEmpty(), "Expected precondition"); 8336 assert(stack->capacity() > num, "Shouldn't bite more than can chew"); 8337 size_t i = num; 8338 oop cur = _overflow_list; 8339 const markOop proto = markOopDesc::prototype(); 8340 NOT_PRODUCT(size_t n = 0;) 8341 for (oop next; i > 0 && cur != NULL; cur = next, i--) { 8342 next = oop(cur->mark()); 8343 cur->set_mark(proto); // until proven otherwise 8344 assert(cur->is_oop(), "Should be an oop"); 8345 bool res = stack->push(cur); 8346 assert(res, "Bit off more than can chew?"); 8347 NOT_PRODUCT(n++;) 8348 } 8349 _overflow_list = cur; 8350 #ifndef PRODUCT 8351 assert(_num_par_pushes >= n, "Too many pops?"); 8352 _num_par_pushes -=n; 8353 #endif 8354 return !stack->isEmpty(); 8355 } 8356 8357 // Multi-threaded; use CAS to break off a prefix 8358 bool CMSCollector::par_take_from_overflow_list(size_t num, 8359 OopTaskQueue* work_q) { 8360 assert(work_q->size() == 0, "That's the current policy"); 8361 assert(num < work_q->max_elems(), "Can't bite more than we can chew"); 8362 if (_overflow_list == NULL) { 8363 return false; 8364 } 8365 // Grab the entire list; we'll put back a suffix 8366 oop prefix = (oop)Atomic::xchg_ptr(NULL, &_overflow_list); 8367 if (prefix == NULL) { // someone grabbed it before we did ... 8368 // ... we could spin for a short while, but for now we don't 8369 return false; 8370 } 8371 size_t i = num; 8372 oop cur = prefix; 8373 for (; i > 1 && cur->mark() != NULL; cur = oop(cur->mark()), i--); 8374 if (cur->mark() != NULL) { 8375 oop suffix_head = cur->mark(); // suffix will be put back on global list 8376 cur->set_mark(NULL); // break off suffix 8377 // Find tail of suffix so we can prepend suffix to global list 8378 for (cur = suffix_head; cur->mark() != NULL; cur = (oop)(cur->mark())); 8379 oop suffix_tail = cur; 8380 assert(suffix_tail != NULL && suffix_tail->mark() == NULL, 8381 "Tautology"); 8382 oop observed_overflow_list = _overflow_list; 8383 do { 8384 cur = observed_overflow_list; 8385 suffix_tail->set_mark(markOop(cur)); 8386 observed_overflow_list = 8387 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur); 8388 } while (cur != observed_overflow_list); 8389 } 8390 8391 // Push the prefix elements on work_q 8392 assert(prefix != NULL, "control point invariant"); 8393 const markOop proto = markOopDesc::prototype(); 8394 oop next; 8395 NOT_PRODUCT(size_t n = 0;) 8396 for (cur = prefix; cur != NULL; cur = next) { 8397 next = oop(cur->mark()); 8398 cur->set_mark(proto); // until proven otherwise 8399 assert(cur->is_oop(), "Should be an oop"); 8400 bool res = work_q->push(cur); 8401 assert(res, "Bit off more than we can chew?"); 8402 NOT_PRODUCT(n++;) 8403 } 8404 #ifndef PRODUCT 8405 assert(_num_par_pushes >= n, "Too many pops?"); 8406 Atomic::add_ptr(-(intptr_t)n, &_num_par_pushes); 8407 #endif 8408 return true; 8409 } 8410 8411 // Single-threaded 8412 void CMSCollector::push_on_overflow_list(oop p) { 8413 NOT_PRODUCT(_num_par_pushes++;) 8414 assert(p->is_oop(), "Not an oop"); 8415 preserve_mark_if_necessary(p); 8416 p->set_mark((markOop)_overflow_list); 8417 _overflow_list = p; 8418 } 8419 8420 // Multi-threaded; use CAS to prepend to overflow list 8421 void CMSCollector::par_push_on_overflow_list(oop p) { 8422 NOT_PRODUCT(Atomic::inc_ptr(&_num_par_pushes);) 8423 assert(p->is_oop(), "Not an oop"); 8424 par_preserve_mark_if_necessary(p); 8425 oop observed_overflow_list = _overflow_list; 8426 oop cur_overflow_list; 8427 do { 8428 cur_overflow_list = observed_overflow_list; 8429 p->set_mark(markOop(cur_overflow_list)); 8430 observed_overflow_list = 8431 (oop) Atomic::cmpxchg_ptr(p, &_overflow_list, cur_overflow_list); 8432 } while (cur_overflow_list != observed_overflow_list); 8433 } 8434 8435 // Single threaded 8436 // General Note on GrowableArray: pushes may silently fail 8437 // because we are (temporarily) out of C-heap for expanding 8438 // the stack. The problem is quite ubiquitous and affects 8439 // a lot of code in the JVM. The prudent thing for GrowableArray 8440 // to do (for now) is to exit with an error. However, that may 8441 // be too draconian in some cases because the caller may be 8442 // able to recover without much harm. For suych cases, we 8443 // should probably introduce a "soft_push" method which returns 8444 // an indication of success or failure with the assumption that 8445 // the caller may be able to recover from a failure; code in 8446 // the VM can then be changed, incrementally, to deal with such 8447 // failures where possible, thus, incrementally hardening the VM 8448 // in such low resource situations. 8449 void CMSCollector::preserve_mark_work(oop p, markOop m) { 8450 int PreserveMarkStackSize = 128; 8451 8452 if (_preserved_oop_stack == NULL) { 8453 assert(_preserved_mark_stack == NULL, 8454 "bijection with preserved_oop_stack"); 8455 // Allocate the stacks 8456 _preserved_oop_stack = new (ResourceObj::C_HEAP) 8457 GrowableArray<oop>(PreserveMarkStackSize, true); 8458 _preserved_mark_stack = new (ResourceObj::C_HEAP) 8459 GrowableArray<markOop>(PreserveMarkStackSize, true); 8460 if (_preserved_oop_stack == NULL || _preserved_mark_stack == NULL) { 8461 vm_exit_out_of_memory(2* PreserveMarkStackSize * sizeof(oop) /* punt */, 8462 "Preserved Mark/Oop Stack for CMS (C-heap)"); 8463 } 8464 } 8465 _preserved_oop_stack->push(p); 8466 _preserved_mark_stack->push(m); 8467 assert(m == p->mark(), "Mark word changed"); 8468 assert(_preserved_oop_stack->length() == _preserved_mark_stack->length(), 8469 "bijection"); 8470 } 8471 8472 // Single threaded 8473 void CMSCollector::preserve_mark_if_necessary(oop p) { 8474 markOop m = p->mark(); 8475 if (m->must_be_preserved(p)) { 8476 preserve_mark_work(p, m); 8477 } 8478 } 8479 8480 void CMSCollector::par_preserve_mark_if_necessary(oop p) { 8481 markOop m = p->mark(); 8482 if (m->must_be_preserved(p)) { 8483 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag); 8484 // Even though we read the mark word without holding 8485 // the lock, we are assured that it will not change 8486 // because we "own" this oop, so no other thread can 8487 // be trying to push it on the overflow list; see 8488 // the assertion in preserve_mark_work() that checks 8489 // that m == p->mark(). 8490 preserve_mark_work(p, m); 8491 } 8492 } 8493 8494 // We should be able to do this multi-threaded, 8495 // a chunk of stack being a task (this is 8496 // correct because each oop only ever appears 8497 // once in the overflow list. However, it's 8498 // not very easy to completely overlap this with 8499 // other operations, so will generally not be done 8500 // until all work's been completed. Because we 8501 // expect the preserved oop stack (set) to be small, 8502 // it's probably fine to do this single-threaded. 8503 // We can explore cleverer concurrent/overlapped/parallel 8504 // processing of preserved marks if we feel the 8505 // need for this in the future. Stack overflow should 8506 // be so rare in practice and, when it happens, its 8507 // effect on performance so great that this will 8508 // likely just be in the noise anyway. 8509 void CMSCollector::restore_preserved_marks_if_any() { 8510 if (_preserved_oop_stack == NULL) { 8511 assert(_preserved_mark_stack == NULL, 8512 "bijection with preserved_oop_stack"); 8513 return; 8514 } 8515 8516 assert(SafepointSynchronize::is_at_safepoint(), 8517 "world should be stopped"); 8518 assert(Thread::current()->is_ConcurrentGC_thread() || 8519 Thread::current()->is_VM_thread(), 8520 "should be single-threaded"); 8521 8522 int length = _preserved_oop_stack->length(); 8523 assert(_preserved_mark_stack->length() == length, "bijection"); 8524 for (int i = 0; i < length; i++) { 8525 oop p = _preserved_oop_stack->at(i); 8526 assert(p->is_oop(), "Should be an oop"); 8527 assert(_span.contains(p), "oop should be in _span"); 8528 assert(p->mark() == markOopDesc::prototype(), 8529 "Set when taken from overflow list"); 8530 markOop m = _preserved_mark_stack->at(i); 8531 p->set_mark(m); 8532 } 8533 _preserved_mark_stack->clear(); 8534 _preserved_oop_stack->clear(); 8535 assert(_preserved_mark_stack->is_empty() && 8536 _preserved_oop_stack->is_empty(), 8537 "stacks were cleared above"); 8538 } 8539 8540 #ifndef PRODUCT 8541 bool CMSCollector::no_preserved_marks() const { 8542 return ( ( _preserved_mark_stack == NULL 8543 && _preserved_oop_stack == NULL) 8544 || ( _preserved_mark_stack->is_empty() 8545 && _preserved_oop_stack->is_empty())); 8546 } 8547 #endif 8548 8549 CMSAdaptiveSizePolicy* ASConcurrentMarkSweepGeneration::cms_size_policy() const 8550 { 8551 GenCollectedHeap* gch = (GenCollectedHeap*) GenCollectedHeap::heap(); 8552 CMSAdaptiveSizePolicy* size_policy = 8553 (CMSAdaptiveSizePolicy*) gch->gen_policy()->size_policy(); 8554 assert(size_policy->is_gc_cms_adaptive_size_policy(), 8555 "Wrong type for size policy"); 8556 return size_policy; 8557 } 8558 8559 void ASConcurrentMarkSweepGeneration::resize(size_t cur_promo_size, 8560 size_t desired_promo_size) { 8561 if (cur_promo_size < desired_promo_size) { 8562 size_t expand_bytes = desired_promo_size - cur_promo_size; 8563 if (PrintAdaptiveSizePolicy && Verbose) { 8564 gclog_or_tty->print_cr(" ASConcurrentMarkSweepGeneration::resize " 8565 "Expanding tenured generation by " SIZE_FORMAT " (bytes)", 8566 expand_bytes); 8567 } 8568 expand(expand_bytes, 8569 MinHeapDeltaBytes, 8570 CMSExpansionCause::_adaptive_size_policy); 8571 } else if (desired_promo_size < cur_promo_size) { 8572 size_t shrink_bytes = cur_promo_size - desired_promo_size; 8573 if (PrintAdaptiveSizePolicy && Verbose) { 8574 gclog_or_tty->print_cr(" ASConcurrentMarkSweepGeneration::resize " 8575 "Shrinking tenured generation by " SIZE_FORMAT " (bytes)", 8576 shrink_bytes); 8577 } 8578 shrink(shrink_bytes); 8579 } 8580 } 8581 8582 CMSGCAdaptivePolicyCounters* ASConcurrentMarkSweepGeneration::gc_adaptive_policy_counters() { 8583 GenCollectedHeap* gch = GenCollectedHeap::heap(); 8584 CMSGCAdaptivePolicyCounters* counters = 8585 (CMSGCAdaptivePolicyCounters*) gch->collector_policy()->counters(); 8586 assert(counters->kind() == GCPolicyCounters::CMSGCAdaptivePolicyCountersKind, 8587 "Wrong kind of counters"); 8588 return counters; 8589 } 8590 8591 8592 void ASConcurrentMarkSweepGeneration::update_counters() { 8593 if (UsePerfData) { 8594 _space_counters->update_all(); 8595 _gen_counters->update_all(); 8596 CMSGCAdaptivePolicyCounters* counters = gc_adaptive_policy_counters(); 8597 GenCollectedHeap* gch = GenCollectedHeap::heap(); 8598 CMSGCStats* gc_stats_l = (CMSGCStats*) gc_stats(); 8599 assert(gc_stats_l->kind() == GCStats::CMSGCStatsKind, 8600 "Wrong gc statistics type"); 8601 counters->update_counters(gc_stats_l); 8602 } 8603 } 8604 8605 void ASConcurrentMarkSweepGeneration::update_counters(size_t used) { 8606 if (UsePerfData) { 8607 _space_counters->update_used(used); 8608 _space_counters->update_capacity(); 8609 _gen_counters->update_all(); 8610 8611 CMSGCAdaptivePolicyCounters* counters = gc_adaptive_policy_counters(); 8612 GenCollectedHeap* gch = GenCollectedHeap::heap(); 8613 CMSGCStats* gc_stats_l = (CMSGCStats*) gc_stats(); 8614 assert(gc_stats_l->kind() == GCStats::CMSGCStatsKind, 8615 "Wrong gc statistics type"); 8616 counters->update_counters(gc_stats_l); 8617 } 8618 } 8619 8620 // The desired expansion delta is computed so that: 8621 // . desired free percentage or greater is used 8622 void ASConcurrentMarkSweepGeneration::compute_new_size() { 8623 assert_locked_or_safepoint(Heap_lock); 8624 8625 GenCollectedHeap* gch = (GenCollectedHeap*) GenCollectedHeap::heap(); 8626 8627 // If incremental collection failed, we just want to expand 8628 // to the limit. 8629 if (incremental_collection_failed()) { 8630 clear_incremental_collection_failed(); 8631 grow_to_reserved(); 8632 return; 8633 } 8634 8635 assert(UseAdaptiveSizePolicy, "Should be using adaptive sizing"); 8636 8637 assert(gch->kind() == CollectedHeap::GenCollectedHeap, 8638 "Wrong type of heap"); 8639 int prev_level = level() - 1; 8640 assert(prev_level >= 0, "The cms generation is the lowest generation"); 8641 Generation* prev_gen = gch->get_gen(prev_level); 8642 assert(prev_gen->kind() == Generation::ASParNew, 8643 "Wrong type of young generation"); 8644 ParNewGeneration* younger_gen = (ParNewGeneration*) prev_gen; 8645 size_t cur_eden = younger_gen->eden()->capacity(); 8646 CMSAdaptiveSizePolicy* size_policy = cms_size_policy(); 8647 size_t cur_promo = free(); 8648 size_policy->compute_tenured_generation_free_space(cur_promo, 8649 max_available(), 8650 cur_eden); 8651 resize(cur_promo, size_policy->promo_size()); 8652 8653 // Record the new size of the space in the cms generation 8654 // that is available for promotions. This is temporary. 8655 // It should be the desired promo size. 8656 size_policy->avg_cms_promo()->sample(free()); 8657 size_policy->avg_old_live()->sample(used()); 8658 8659 if (UsePerfData) { 8660 CMSGCAdaptivePolicyCounters* counters = gc_adaptive_policy_counters(); 8661 counters->update_cms_capacity_counter(capacity()); 8662 } 8663 } 8664 8665 void ASConcurrentMarkSweepGeneration::shrink_by(size_t desired_bytes) { 8666 assert_locked_or_safepoint(Heap_lock); 8667 assert_lock_strong(freelistLock()); 8668 HeapWord* old_end = _cmsSpace->end(); 8669 HeapWord* unallocated_start = _cmsSpace->unallocated_block(); 8670 assert(old_end >= unallocated_start, "Miscalculation of unallocated_start"); 8671 FreeChunk* chunk_at_end = find_chunk_at_end(); 8672 if (chunk_at_end == NULL) { 8673 // No room to shrink 8674 if (PrintGCDetails && Verbose) { 8675 gclog_or_tty->print_cr("No room to shrink: old_end " 8676 PTR_FORMAT " unallocated_start " PTR_FORMAT 8677 " chunk_at_end " PTR_FORMAT, 8678 old_end, unallocated_start, chunk_at_end); 8679 } 8680 return; 8681 } else { 8682 8683 // Find the chunk at the end of the space and determine 8684 // how much it can be shrunk. 8685 size_t shrinkable_size_in_bytes = chunk_at_end->size(); 8686 size_t aligned_shrinkable_size_in_bytes = 8687 align_size_down(shrinkable_size_in_bytes, os::vm_page_size()); 8688 assert(unallocated_start <= chunk_at_end->end(), 8689 "Inconsistent chunk at end of space"); 8690 size_t bytes = MIN2(desired_bytes, aligned_shrinkable_size_in_bytes); 8691 size_t word_size_before = heap_word_size(_virtual_space.committed_size()); 8692 8693 // Shrink the underlying space 8694 _virtual_space.shrink_by(bytes); 8695 if (PrintGCDetails && Verbose) { 8696 gclog_or_tty->print_cr("ConcurrentMarkSweepGeneration::shrink_by:" 8697 " desired_bytes " SIZE_FORMAT 8698 " shrinkable_size_in_bytes " SIZE_FORMAT 8699 " aligned_shrinkable_size_in_bytes " SIZE_FORMAT 8700 " bytes " SIZE_FORMAT, 8701 desired_bytes, shrinkable_size_in_bytes, 8702 aligned_shrinkable_size_in_bytes, bytes); 8703 gclog_or_tty->print_cr(" old_end " SIZE_FORMAT 8704 " unallocated_start " SIZE_FORMAT, 8705 old_end, unallocated_start); 8706 } 8707 8708 // If the space did shrink (shrinking is not guaranteed), 8709 // shrink the chunk at the end by the appropriate amount. 8710 if (((HeapWord*)_virtual_space.high()) < old_end) { 8711 size_t new_word_size = 8712 heap_word_size(_virtual_space.committed_size()); 8713 8714 // Have to remove the chunk from the dictionary because it is changing 8715 // size and might be someplace elsewhere in the dictionary. 8716 8717 // Get the chunk at end, shrink it, and put it 8718 // back. 8719 _cmsSpace->removeChunkFromDictionary(chunk_at_end); 8720 size_t word_size_change = word_size_before - new_word_size; 8721 size_t chunk_at_end_old_size = chunk_at_end->size(); 8722 assert(chunk_at_end_old_size >= word_size_change, 8723 "Shrink is too large"); 8724 chunk_at_end->setSize(chunk_at_end_old_size - 8725 word_size_change); 8726 _cmsSpace->freed((HeapWord*) chunk_at_end->end(), 8727 word_size_change); 8728 8729 _cmsSpace->returnChunkToDictionary(chunk_at_end); 8730 8731 MemRegion mr(_cmsSpace->bottom(), new_word_size); 8732 _bts->resize(new_word_size); // resize the block offset shared array 8733 Universe::heap()->barrier_set()->resize_covered_region(mr); 8734 _cmsSpace->assert_locked(); 8735 _cmsSpace->set_end((HeapWord*)_virtual_space.high()); 8736 8737 NOT_PRODUCT(_cmsSpace->dictionary()->verify()); 8738 8739 // update the space and generation capacity counters 8740 if (UsePerfData) { 8741 _space_counters->update_capacity(); 8742 _gen_counters->update_all(); 8743 } 8744 8745 if (Verbose && PrintGCDetails) { 8746 size_t new_mem_size = _virtual_space.committed_size(); 8747 size_t old_mem_size = new_mem_size + bytes; 8748 gclog_or_tty->print_cr("Shrinking %s from %ldK by %ldK to %ldK", 8749 name(), old_mem_size/K, bytes/K, new_mem_size/K); 8750 } 8751 } 8752 8753 assert(_cmsSpace->unallocated_block() <= _cmsSpace->end(), 8754 "Inconsistency at end of space"); 8755 assert(chunk_at_end->end() == _cmsSpace->end(), 8756 "Shrinking is inconsistent"); 8757 return; 8758 } 8759 } 8760 8761 // Transfer some number of overflown objects to usual marking 8762 // stack. Return true if some objects were transferred. 8763 bool MarkRefsIntoAndScanClosure::take_from_overflow_list() { 8764 size_t num = MIN2((size_t)_mark_stack->capacity()/4, 8765 (size_t)ParGCDesiredObjsFromOverflowList); 8766 8767 bool res = _collector->take_from_overflow_list(num, _mark_stack); 8768 assert(_collector->overflow_list_is_empty() || res, 8769 "If list is not empty, we should have taken something"); 8770 assert(!res || !_mark_stack->isEmpty(), 8771 "If we took something, it should now be on our stack"); 8772 return res; 8773 } 8774 8775 size_t MarkDeadObjectsClosure::do_blk(HeapWord* addr) { 8776 size_t res = _sp->block_size_no_stall(addr, _collector); 8777 assert(res != 0, "Should always be able to compute a size"); 8778 if (_sp->block_is_obj(addr)) { 8779 if (_live_bit_map->isMarked(addr)) { 8780 // It can't have been dead in a previous cycle 8781 guarantee(!_dead_bit_map->isMarked(addr), "No resurrection!"); 8782 } else { 8783 _dead_bit_map->mark(addr); // mark the dead object 8784 } 8785 } 8786 return res; 8787 }