1 /* 2 * Copyright (c) 2001, 2010, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 # include "incls/_precompiled.incl" 26 # include "incls/_concurrentMarkSweepGeneration.cpp.incl" 27 28 // statics 29 CMSCollector* ConcurrentMarkSweepGeneration::_collector = NULL; 30 bool CMSCollector::_full_gc_requested = false; 31 32 ////////////////////////////////////////////////////////////////// 33 // In support of CMS/VM thread synchronization 34 ////////////////////////////////////////////////////////////////// 35 // We split use of the CGC_lock into 2 "levels". 36 // The low-level locking is of the usual CGC_lock monitor. We introduce 37 // a higher level "token" (hereafter "CMS token") built on top of the 38 // low level monitor (hereafter "CGC lock"). 39 // The token-passing protocol gives priority to the VM thread. The 40 // CMS-lock doesn't provide any fairness guarantees, but clients 41 // should ensure that it is only held for very short, bounded 42 // durations. 43 // 44 // When either of the CMS thread or the VM thread is involved in 45 // collection operations during which it does not want the other 46 // thread to interfere, it obtains the CMS token. 47 // 48 // If either thread tries to get the token while the other has 49 // it, that thread waits. However, if the VM thread and CMS thread 50 // both want the token, then the VM thread gets priority while the 51 // CMS thread waits. This ensures, for instance, that the "concurrent" 52 // phases of the CMS thread's work do not block out the VM thread 53 // for long periods of time as the CMS thread continues to hog 54 // the token. (See bug 4616232). 55 // 56 // The baton-passing functions are, however, controlled by the 57 // flags _foregroundGCShouldWait and _foregroundGCIsActive, 58 // and here the low-level CMS lock, not the high level token, 59 // ensures mutual exclusion. 60 // 61 // Two important conditions that we have to satisfy: 62 // 1. if a thread does a low-level wait on the CMS lock, then it 63 // relinquishes the CMS token if it were holding that token 64 // when it acquired the low-level CMS lock. 65 // 2. any low-level notifications on the low-level lock 66 // should only be sent when a thread has relinquished the token. 67 // 68 // In the absence of either property, we'd have potential deadlock. 69 // 70 // We protect each of the CMS (concurrent and sequential) phases 71 // with the CMS _token_, not the CMS _lock_. 72 // 73 // The only code protected by CMS lock is the token acquisition code 74 // itself, see ConcurrentMarkSweepThread::[de]synchronize(), and the 75 // baton-passing code. 76 // 77 // Unfortunately, i couldn't come up with a good abstraction to factor and 78 // hide the naked CGC_lock manipulation in the baton-passing code 79 // further below. That's something we should try to do. Also, the proof 80 // of correctness of this 2-level locking scheme is far from obvious, 81 // and potentially quite slippery. We have an uneasy supsicion, for instance, 82 // that there may be a theoretical possibility of delay/starvation in the 83 // low-level lock/wait/notify scheme used for the baton-passing because of 84 // potential intereference with the priority scheme embodied in the 85 // CMS-token-passing protocol. See related comments at a CGC_lock->wait() 86 // invocation further below and marked with "XXX 20011219YSR". 87 // Indeed, as we note elsewhere, this may become yet more slippery 88 // in the presence of multiple CMS and/or multiple VM threads. XXX 89 90 class CMSTokenSync: public StackObj { 91 private: 92 bool _is_cms_thread; 93 public: 94 CMSTokenSync(bool is_cms_thread): 95 _is_cms_thread(is_cms_thread) { 96 assert(is_cms_thread == Thread::current()->is_ConcurrentGC_thread(), 97 "Incorrect argument to constructor"); 98 ConcurrentMarkSweepThread::synchronize(_is_cms_thread); 99 } 100 101 ~CMSTokenSync() { 102 assert(_is_cms_thread ? 103 ConcurrentMarkSweepThread::cms_thread_has_cms_token() : 104 ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 105 "Incorrect state"); 106 ConcurrentMarkSweepThread::desynchronize(_is_cms_thread); 107 } 108 }; 109 110 // Convenience class that does a CMSTokenSync, and then acquires 111 // upto three locks. 112 class CMSTokenSyncWithLocks: public CMSTokenSync { 113 private: 114 // Note: locks are acquired in textual declaration order 115 // and released in the opposite order 116 MutexLockerEx _locker1, _locker2, _locker3; 117 public: 118 CMSTokenSyncWithLocks(bool is_cms_thread, Mutex* mutex1, 119 Mutex* mutex2 = NULL, Mutex* mutex3 = NULL): 120 CMSTokenSync(is_cms_thread), 121 _locker1(mutex1, Mutex::_no_safepoint_check_flag), 122 _locker2(mutex2, Mutex::_no_safepoint_check_flag), 123 _locker3(mutex3, Mutex::_no_safepoint_check_flag) 124 { } 125 }; 126 127 128 // Wrapper class to temporarily disable icms during a foreground cms collection. 129 class ICMSDisabler: public StackObj { 130 public: 131 // The ctor disables icms and wakes up the thread so it notices the change; 132 // the dtor re-enables icms. Note that the CMSCollector methods will check 133 // CMSIncrementalMode. 134 ICMSDisabler() { CMSCollector::disable_icms(); CMSCollector::start_icms(); } 135 ~ICMSDisabler() { CMSCollector::enable_icms(); } 136 }; 137 138 ////////////////////////////////////////////////////////////////// 139 // Concurrent Mark-Sweep Generation ///////////////////////////// 140 ////////////////////////////////////////////////////////////////// 141 142 NOT_PRODUCT(CompactibleFreeListSpace* debug_cms_space;) 143 144 // This struct contains per-thread things necessary to support parallel 145 // young-gen collection. 146 class CMSParGCThreadState: public CHeapObj { 147 public: 148 CFLS_LAB lab; 149 PromotionInfo promo; 150 151 // Constructor. 152 CMSParGCThreadState(CompactibleFreeListSpace* cfls) : lab(cfls) { 153 promo.setSpace(cfls); 154 } 155 }; 156 157 ConcurrentMarkSweepGeneration::ConcurrentMarkSweepGeneration( 158 ReservedSpace rs, size_t initial_byte_size, int level, 159 CardTableRS* ct, bool use_adaptive_freelists, 160 FreeBlockDictionary::DictionaryChoice dictionaryChoice) : 161 CardGeneration(rs, initial_byte_size, level, ct), 162 _dilatation_factor(((double)MinChunkSize)/((double)(CollectedHeap::min_fill_size()))), 163 _debug_collection_type(Concurrent_collection_type) 164 { 165 HeapWord* bottom = (HeapWord*) _virtual_space.low(); 166 HeapWord* end = (HeapWord*) _virtual_space.high(); 167 168 _direct_allocated_words = 0; 169 NOT_PRODUCT( 170 _numObjectsPromoted = 0; 171 _numWordsPromoted = 0; 172 _numObjectsAllocated = 0; 173 _numWordsAllocated = 0; 174 ) 175 176 _cmsSpace = new CompactibleFreeListSpace(_bts, MemRegion(bottom, end), 177 use_adaptive_freelists, 178 dictionaryChoice); 179 NOT_PRODUCT(debug_cms_space = _cmsSpace;) 180 if (_cmsSpace == NULL) { 181 vm_exit_during_initialization( 182 "CompactibleFreeListSpace allocation failure"); 183 } 184 _cmsSpace->_gen = this; 185 186 _gc_stats = new CMSGCStats(); 187 188 // Verify the assumption that FreeChunk::_prev and OopDesc::_klass 189 // offsets match. The ability to tell free chunks from objects 190 // depends on this property. 191 debug_only( 192 FreeChunk* junk = NULL; 193 assert(UseCompressedOops || 194 junk->prev_addr() == (void*)(oop(junk)->klass_addr()), 195 "Offset of FreeChunk::_prev within FreeChunk must match" 196 " that of OopDesc::_klass within OopDesc"); 197 ) 198 if (CollectedHeap::use_parallel_gc_threads()) { 199 typedef CMSParGCThreadState* CMSParGCThreadStatePtr; 200 _par_gc_thread_states = 201 NEW_C_HEAP_ARRAY(CMSParGCThreadStatePtr, ParallelGCThreads); 202 if (_par_gc_thread_states == NULL) { 203 vm_exit_during_initialization("Could not allocate par gc structs"); 204 } 205 for (uint i = 0; i < ParallelGCThreads; i++) { 206 _par_gc_thread_states[i] = new CMSParGCThreadState(cmsSpace()); 207 if (_par_gc_thread_states[i] == NULL) { 208 vm_exit_during_initialization("Could not allocate par gc structs"); 209 } 210 } 211 } else { 212 _par_gc_thread_states = NULL; 213 } 214 _incremental_collection_failed = false; 215 // The "dilatation_factor" is the expansion that can occur on 216 // account of the fact that the minimum object size in the CMS 217 // generation may be larger than that in, say, a contiguous young 218 // generation. 219 // Ideally, in the calculation below, we'd compute the dilatation 220 // factor as: MinChunkSize/(promoting_gen's min object size) 221 // Since we do not have such a general query interface for the 222 // promoting generation, we'll instead just use the mimimum 223 // object size (which today is a header's worth of space); 224 // note that all arithmetic is in units of HeapWords. 225 assert(MinChunkSize >= CollectedHeap::min_fill_size(), "just checking"); 226 assert(_dilatation_factor >= 1.0, "from previous assert"); 227 } 228 229 230 // The field "_initiating_occupancy" represents the occupancy percentage 231 // at which we trigger a new collection cycle. Unless explicitly specified 232 // via CMSInitiating[Perm]OccupancyFraction (argument "io" below), it 233 // is calculated by: 234 // 235 // Let "f" be MinHeapFreeRatio in 236 // 237 // _intiating_occupancy = 100-f + 238 // f * (CMSTrigger[Perm]Ratio/100) 239 // where CMSTrigger[Perm]Ratio is the argument "tr" below. 240 // 241 // That is, if we assume the heap is at its desired maximum occupancy at the 242 // end of a collection, we let CMSTrigger[Perm]Ratio of the (purported) free 243 // space be allocated before initiating a new collection cycle. 244 // 245 void ConcurrentMarkSweepGeneration::init_initiating_occupancy(intx io, intx tr) { 246 assert(io <= 100 && tr >= 0 && tr <= 100, "Check the arguments"); 247 if (io >= 0) { 248 _initiating_occupancy = (double)io / 100.0; 249 } else { 250 _initiating_occupancy = ((100 - MinHeapFreeRatio) + 251 (double)(tr * MinHeapFreeRatio) / 100.0) 252 / 100.0; 253 } 254 } 255 256 void ConcurrentMarkSweepGeneration::ref_processor_init() { 257 assert(collector() != NULL, "no collector"); 258 collector()->ref_processor_init(); 259 } 260 261 void CMSCollector::ref_processor_init() { 262 if (_ref_processor == NULL) { 263 // Allocate and initialize a reference processor 264 _ref_processor = ReferenceProcessor::create_ref_processor( 265 _span, // span 266 _cmsGen->refs_discovery_is_atomic(), // atomic_discovery 267 _cmsGen->refs_discovery_is_mt(), // mt_discovery 268 &_is_alive_closure, 269 ParallelGCThreads, 270 ParallelRefProcEnabled); 271 // Initialize the _ref_processor field of CMSGen 272 _cmsGen->set_ref_processor(_ref_processor); 273 274 // Allocate a dummy ref processor for perm gen. 275 ReferenceProcessor* rp2 = new ReferenceProcessor(); 276 if (rp2 == NULL) { 277 vm_exit_during_initialization("Could not allocate ReferenceProcessor object"); 278 } 279 _permGen->set_ref_processor(rp2); 280 } 281 } 282 283 CMSAdaptiveSizePolicy* CMSCollector::size_policy() { 284 GenCollectedHeap* gch = GenCollectedHeap::heap(); 285 assert(gch->kind() == CollectedHeap::GenCollectedHeap, 286 "Wrong type of heap"); 287 CMSAdaptiveSizePolicy* sp = (CMSAdaptiveSizePolicy*) 288 gch->gen_policy()->size_policy(); 289 assert(sp->is_gc_cms_adaptive_size_policy(), 290 "Wrong type of size policy"); 291 return sp; 292 } 293 294 CMSGCAdaptivePolicyCounters* CMSCollector::gc_adaptive_policy_counters() { 295 CMSGCAdaptivePolicyCounters* results = 296 (CMSGCAdaptivePolicyCounters*) collector_policy()->counters(); 297 assert( 298 results->kind() == GCPolicyCounters::CMSGCAdaptivePolicyCountersKind, 299 "Wrong gc policy counter kind"); 300 return results; 301 } 302 303 304 void ConcurrentMarkSweepGeneration::initialize_performance_counters() { 305 306 const char* gen_name = "old"; 307 308 // Generation Counters - generation 1, 1 subspace 309 _gen_counters = new GenerationCounters(gen_name, 1, 1, &_virtual_space); 310 311 _space_counters = new GSpaceCounters(gen_name, 0, 312 _virtual_space.reserved_size(), 313 this, _gen_counters); 314 } 315 316 CMSStats::CMSStats(ConcurrentMarkSweepGeneration* cms_gen, unsigned int alpha): 317 _cms_gen(cms_gen) 318 { 319 assert(alpha <= 100, "bad value"); 320 _saved_alpha = alpha; 321 322 // Initialize the alphas to the bootstrap value of 100. 323 _gc0_alpha = _cms_alpha = 100; 324 325 _cms_begin_time.update(); 326 _cms_end_time.update(); 327 328 _gc0_duration = 0.0; 329 _gc0_period = 0.0; 330 _gc0_promoted = 0; 331 332 _cms_duration = 0.0; 333 _cms_period = 0.0; 334 _cms_allocated = 0; 335 336 _cms_used_at_gc0_begin = 0; 337 _cms_used_at_gc0_end = 0; 338 _allow_duty_cycle_reduction = false; 339 _valid_bits = 0; 340 _icms_duty_cycle = CMSIncrementalDutyCycle; 341 } 342 343 double CMSStats::cms_free_adjustment_factor(size_t free) const { 344 // TBD: CR 6909490 345 return 1.0; 346 } 347 348 void CMSStats::adjust_cms_free_adjustment_factor(bool fail, size_t free) { 349 } 350 351 // If promotion failure handling is on use 352 // the padded average size of the promotion for each 353 // young generation collection. 354 double CMSStats::time_until_cms_gen_full() const { 355 size_t cms_free = _cms_gen->cmsSpace()->free(); 356 GenCollectedHeap* gch = GenCollectedHeap::heap(); 357 size_t expected_promotion = gch->get_gen(0)->capacity(); 358 if (HandlePromotionFailure) { 359 expected_promotion = MIN2( 360 (size_t) _cms_gen->gc_stats()->avg_promoted()->padded_average(), 361 expected_promotion); 362 } 363 if (cms_free > expected_promotion) { 364 // Start a cms collection if there isn't enough space to promote 365 // for the next minor collection. Use the padded average as 366 // a safety factor. 367 cms_free -= expected_promotion; 368 369 // Adjust by the safety factor. 370 double cms_free_dbl = (double)cms_free; 371 double cms_adjustment = (100.0 - CMSIncrementalSafetyFactor)/100.0; 372 // Apply a further correction factor which tries to adjust 373 // for recent occurance of concurrent mode failures. 374 cms_adjustment = cms_adjustment * cms_free_adjustment_factor(cms_free); 375 cms_free_dbl = cms_free_dbl * cms_adjustment; 376 377 if (PrintGCDetails && Verbose) { 378 gclog_or_tty->print_cr("CMSStats::time_until_cms_gen_full: cms_free " 379 SIZE_FORMAT " expected_promotion " SIZE_FORMAT, 380 cms_free, expected_promotion); 381 gclog_or_tty->print_cr(" cms_free_dbl %f cms_consumption_rate %f", 382 cms_free_dbl, cms_consumption_rate() + 1.0); 383 } 384 // Add 1 in case the consumption rate goes to zero. 385 return cms_free_dbl / (cms_consumption_rate() + 1.0); 386 } 387 return 0.0; 388 } 389 390 // Compare the duration of the cms collection to the 391 // time remaining before the cms generation is empty. 392 // Note that the time from the start of the cms collection 393 // to the start of the cms sweep (less than the total 394 // duration of the cms collection) can be used. This 395 // has been tried and some applications experienced 396 // promotion failures early in execution. This was 397 // possibly because the averages were not accurate 398 // enough at the beginning. 399 double CMSStats::time_until_cms_start() const { 400 // We add "gc0_period" to the "work" calculation 401 // below because this query is done (mostly) at the 402 // end of a scavenge, so we need to conservatively 403 // account for that much possible delay 404 // in the query so as to avoid concurrent mode failures 405 // due to starting the collection just a wee bit too 406 // late. 407 double work = cms_duration() + gc0_period(); 408 double deadline = time_until_cms_gen_full(); 409 // If a concurrent mode failure occurred recently, we want to be 410 // more conservative and halve our expected time_until_cms_gen_full() 411 if (work > deadline) { 412 if (Verbose && PrintGCDetails) { 413 gclog_or_tty->print( 414 " CMSCollector: collect because of anticipated promotion " 415 "before full %3.7f + %3.7f > %3.7f ", cms_duration(), 416 gc0_period(), time_until_cms_gen_full()); 417 } 418 return 0.0; 419 } 420 return work - deadline; 421 } 422 423 // Return a duty cycle based on old_duty_cycle and new_duty_cycle, limiting the 424 // amount of change to prevent wild oscillation. 425 unsigned int CMSStats::icms_damped_duty_cycle(unsigned int old_duty_cycle, 426 unsigned int new_duty_cycle) { 427 assert(old_duty_cycle <= 100, "bad input value"); 428 assert(new_duty_cycle <= 100, "bad input value"); 429 430 // Note: use subtraction with caution since it may underflow (values are 431 // unsigned). Addition is safe since we're in the range 0-100. 432 unsigned int damped_duty_cycle = new_duty_cycle; 433 if (new_duty_cycle < old_duty_cycle) { 434 const unsigned int largest_delta = MAX2(old_duty_cycle / 4, 5U); 435 if (new_duty_cycle + largest_delta < old_duty_cycle) { 436 damped_duty_cycle = old_duty_cycle - largest_delta; 437 } 438 } else if (new_duty_cycle > old_duty_cycle) { 439 const unsigned int largest_delta = MAX2(old_duty_cycle / 4, 15U); 440 if (new_duty_cycle > old_duty_cycle + largest_delta) { 441 damped_duty_cycle = MIN2(old_duty_cycle + largest_delta, 100U); 442 } 443 } 444 assert(damped_duty_cycle <= 100, "invalid duty cycle computed"); 445 446 if (CMSTraceIncrementalPacing) { 447 gclog_or_tty->print(" [icms_damped_duty_cycle(%d,%d) = %d] ", 448 old_duty_cycle, new_duty_cycle, damped_duty_cycle); 449 } 450 return damped_duty_cycle; 451 } 452 453 unsigned int CMSStats::icms_update_duty_cycle_impl() { 454 assert(CMSIncrementalPacing && valid(), 455 "should be handled in icms_update_duty_cycle()"); 456 457 double cms_time_so_far = cms_timer().seconds(); 458 double scaled_duration = cms_duration_per_mb() * _cms_used_at_gc0_end / M; 459 double scaled_duration_remaining = fabsd(scaled_duration - cms_time_so_far); 460 461 // Avoid division by 0. 462 double time_until_full = MAX2(time_until_cms_gen_full(), 0.01); 463 double duty_cycle_dbl = 100.0 * scaled_duration_remaining / time_until_full; 464 465 unsigned int new_duty_cycle = MIN2((unsigned int)duty_cycle_dbl, 100U); 466 if (new_duty_cycle > _icms_duty_cycle) { 467 // Avoid very small duty cycles (1 or 2); 0 is allowed. 468 if (new_duty_cycle > 2) { 469 _icms_duty_cycle = icms_damped_duty_cycle(_icms_duty_cycle, 470 new_duty_cycle); 471 } 472 } else if (_allow_duty_cycle_reduction) { 473 // The duty cycle is reduced only once per cms cycle (see record_cms_end()). 474 new_duty_cycle = icms_damped_duty_cycle(_icms_duty_cycle, new_duty_cycle); 475 // Respect the minimum duty cycle. 476 unsigned int min_duty_cycle = (unsigned int)CMSIncrementalDutyCycleMin; 477 _icms_duty_cycle = MAX2(new_duty_cycle, min_duty_cycle); 478 } 479 480 if (PrintGCDetails || CMSTraceIncrementalPacing) { 481 gclog_or_tty->print(" icms_dc=%d ", _icms_duty_cycle); 482 } 483 484 _allow_duty_cycle_reduction = false; 485 return _icms_duty_cycle; 486 } 487 488 #ifndef PRODUCT 489 void CMSStats::print_on(outputStream *st) const { 490 st->print(" gc0_alpha=%d,cms_alpha=%d", _gc0_alpha, _cms_alpha); 491 st->print(",gc0_dur=%g,gc0_per=%g,gc0_promo=" SIZE_FORMAT, 492 gc0_duration(), gc0_period(), gc0_promoted()); 493 st->print(",cms_dur=%g,cms_dur_per_mb=%g,cms_per=%g,cms_alloc=" SIZE_FORMAT, 494 cms_duration(), cms_duration_per_mb(), 495 cms_period(), cms_allocated()); 496 st->print(",cms_since_beg=%g,cms_since_end=%g", 497 cms_time_since_begin(), cms_time_since_end()); 498 st->print(",cms_used_beg=" SIZE_FORMAT ",cms_used_end=" SIZE_FORMAT, 499 _cms_used_at_gc0_begin, _cms_used_at_gc0_end); 500 if (CMSIncrementalMode) { 501 st->print(",dc=%d", icms_duty_cycle()); 502 } 503 504 if (valid()) { 505 st->print(",promo_rate=%g,cms_alloc_rate=%g", 506 promotion_rate(), cms_allocation_rate()); 507 st->print(",cms_consumption_rate=%g,time_until_full=%g", 508 cms_consumption_rate(), time_until_cms_gen_full()); 509 } 510 st->print(" "); 511 } 512 #endif // #ifndef PRODUCT 513 514 CMSCollector::CollectorState CMSCollector::_collectorState = 515 CMSCollector::Idling; 516 bool CMSCollector::_foregroundGCIsActive = false; 517 bool CMSCollector::_foregroundGCShouldWait = false; 518 519 CMSCollector::CMSCollector(ConcurrentMarkSweepGeneration* cmsGen, 520 ConcurrentMarkSweepGeneration* permGen, 521 CardTableRS* ct, 522 ConcurrentMarkSweepPolicy* cp): 523 _cmsGen(cmsGen), 524 _permGen(permGen), 525 _ct(ct), 526 _ref_processor(NULL), // will be set later 527 _conc_workers(NULL), // may be set later 528 _abort_preclean(false), 529 _start_sampling(false), 530 _between_prologue_and_epilogue(false), 531 _markBitMap(0, Mutex::leaf + 1, "CMS_markBitMap_lock"), 532 _perm_gen_verify_bit_map(0, -1 /* no mutex */, "No_lock"), 533 _modUnionTable((CardTableModRefBS::card_shift - LogHeapWordSize), 534 -1 /* lock-free */, "No_lock" /* dummy */), 535 _modUnionClosure(&_modUnionTable), 536 _modUnionClosurePar(&_modUnionTable), 537 // Adjust my span to cover old (cms) gen and perm gen 538 _span(cmsGen->reserved()._union(permGen->reserved())), 539 // Construct the is_alive_closure with _span & markBitMap 540 _is_alive_closure(_span, &_markBitMap), 541 _restart_addr(NULL), 542 _overflow_list(NULL), 543 _stats(cmsGen), 544 _eden_chunk_array(NULL), // may be set in ctor body 545 _eden_chunk_capacity(0), // -- ditto -- 546 _eden_chunk_index(0), // -- ditto -- 547 _survivor_plab_array(NULL), // -- ditto -- 548 _survivor_chunk_array(NULL), // -- ditto -- 549 _survivor_chunk_capacity(0), // -- ditto -- 550 _survivor_chunk_index(0), // -- ditto -- 551 _ser_pmc_preclean_ovflw(0), 552 _ser_kac_preclean_ovflw(0), 553 _ser_pmc_remark_ovflw(0), 554 _par_pmc_remark_ovflw(0), 555 _ser_kac_ovflw(0), 556 _par_kac_ovflw(0), 557 #ifndef PRODUCT 558 _num_par_pushes(0), 559 #endif 560 _collection_count_start(0), 561 _verifying(false), 562 _icms_start_limit(NULL), 563 _icms_stop_limit(NULL), 564 _verification_mark_bm(0, Mutex::leaf + 1, "CMS_verification_mark_bm_lock"), 565 _completed_initialization(false), 566 _collector_policy(cp), 567 _should_unload_classes(false), 568 _concurrent_cycles_since_last_unload(0), 569 _roots_scanning_options(0), 570 _inter_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding), 571 _intra_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding) 572 { 573 if (ExplicitGCInvokesConcurrentAndUnloadsClasses) { 574 ExplicitGCInvokesConcurrent = true; 575 } 576 // Now expand the span and allocate the collection support structures 577 // (MUT, marking bit map etc.) to cover both generations subject to 578 // collection. 579 580 // First check that _permGen is adjacent to _cmsGen and above it. 581 assert( _cmsGen->reserved().word_size() > 0 582 && _permGen->reserved().word_size() > 0, 583 "generations should not be of zero size"); 584 assert(_cmsGen->reserved().intersection(_permGen->reserved()).is_empty(), 585 "_cmsGen and _permGen should not overlap"); 586 assert(_cmsGen->reserved().end() == _permGen->reserved().start(), 587 "_cmsGen->end() different from _permGen->start()"); 588 589 // For use by dirty card to oop closures. 590 _cmsGen->cmsSpace()->set_collector(this); 591 _permGen->cmsSpace()->set_collector(this); 592 593 // Allocate MUT and marking bit map 594 { 595 MutexLockerEx x(_markBitMap.lock(), Mutex::_no_safepoint_check_flag); 596 if (!_markBitMap.allocate(_span)) { 597 warning("Failed to allocate CMS Bit Map"); 598 return; 599 } 600 assert(_markBitMap.covers(_span), "_markBitMap inconsistency?"); 601 } 602 { 603 _modUnionTable.allocate(_span); 604 assert(_modUnionTable.covers(_span), "_modUnionTable inconsistency?"); 605 } 606 607 if (!_markStack.allocate(MarkStackSize)) { 608 warning("Failed to allocate CMS Marking Stack"); 609 return; 610 } 611 if (!_revisitStack.allocate(CMSRevisitStackSize)) { 612 warning("Failed to allocate CMS Revisit Stack"); 613 return; 614 } 615 616 // Support for multi-threaded concurrent phases 617 if (CollectedHeap::use_parallel_gc_threads() && CMSConcurrentMTEnabled) { 618 if (FLAG_IS_DEFAULT(ConcGCThreads)) { 619 // just for now 620 FLAG_SET_DEFAULT(ConcGCThreads, (ParallelGCThreads + 3)/4); 621 } 622 if (ConcGCThreads > 1) { 623 _conc_workers = new YieldingFlexibleWorkGang("Parallel CMS Threads", 624 ConcGCThreads, true); 625 if (_conc_workers == NULL) { 626 warning("GC/CMS: _conc_workers allocation failure: " 627 "forcing -CMSConcurrentMTEnabled"); 628 CMSConcurrentMTEnabled = false; 629 } else { 630 _conc_workers->initialize_workers(); 631 } 632 } else { 633 CMSConcurrentMTEnabled = false; 634 } 635 } 636 if (!CMSConcurrentMTEnabled) { 637 ConcGCThreads = 0; 638 } else { 639 // Turn off CMSCleanOnEnter optimization temporarily for 640 // the MT case where it's not fixed yet; see 6178663. 641 CMSCleanOnEnter = false; 642 } 643 assert((_conc_workers != NULL) == (ConcGCThreads > 1), 644 "Inconsistency"); 645 646 // Parallel task queues; these are shared for the 647 // concurrent and stop-world phases of CMS, but 648 // are not shared with parallel scavenge (ParNew). 649 { 650 uint i; 651 uint num_queues = (uint) MAX2(ParallelGCThreads, ConcGCThreads); 652 653 if ((CMSParallelRemarkEnabled || CMSConcurrentMTEnabled 654 || ParallelRefProcEnabled) 655 && num_queues > 0) { 656 _task_queues = new OopTaskQueueSet(num_queues); 657 if (_task_queues == NULL) { 658 warning("task_queues allocation failure."); 659 return; 660 } 661 _hash_seed = NEW_C_HEAP_ARRAY(int, num_queues); 662 if (_hash_seed == NULL) { 663 warning("_hash_seed array allocation failure"); 664 return; 665 } 666 667 typedef Padded<OopTaskQueue> PaddedOopTaskQueue; 668 for (i = 0; i < num_queues; i++) { 669 PaddedOopTaskQueue *q = new PaddedOopTaskQueue(); 670 if (q == NULL) { 671 warning("work_queue allocation failure."); 672 return; 673 } 674 _task_queues->register_queue(i, q); 675 } 676 for (i = 0; i < num_queues; i++) { 677 _task_queues->queue(i)->initialize(); 678 _hash_seed[i] = 17; // copied from ParNew 679 } 680 } 681 } 682 683 _cmsGen ->init_initiating_occupancy(CMSInitiatingOccupancyFraction, CMSTriggerRatio); 684 _permGen->init_initiating_occupancy(CMSInitiatingPermOccupancyFraction, CMSTriggerPermRatio); 685 686 // Clip CMSBootstrapOccupancy between 0 and 100. 687 _bootstrap_occupancy = ((double)MIN2((uintx)100, MAX2((uintx)0, CMSBootstrapOccupancy))) 688 /(double)100; 689 690 _full_gcs_since_conc_gc = 0; 691 692 // Now tell CMS generations the identity of their collector 693 ConcurrentMarkSweepGeneration::set_collector(this); 694 695 // Create & start a CMS thread for this CMS collector 696 _cmsThread = ConcurrentMarkSweepThread::start(this); 697 assert(cmsThread() != NULL, "CMS Thread should have been created"); 698 assert(cmsThread()->collector() == this, 699 "CMS Thread should refer to this gen"); 700 assert(CGC_lock != NULL, "Where's the CGC_lock?"); 701 702 // Support for parallelizing young gen rescan 703 GenCollectedHeap* gch = GenCollectedHeap::heap(); 704 _young_gen = gch->prev_gen(_cmsGen); 705 if (gch->supports_inline_contig_alloc()) { 706 _top_addr = gch->top_addr(); 707 _end_addr = gch->end_addr(); 708 assert(_young_gen != NULL, "no _young_gen"); 709 _eden_chunk_index = 0; 710 _eden_chunk_capacity = (_young_gen->max_capacity()+CMSSamplingGrain)/CMSSamplingGrain; 711 _eden_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, _eden_chunk_capacity); 712 if (_eden_chunk_array == NULL) { 713 _eden_chunk_capacity = 0; 714 warning("GC/CMS: _eden_chunk_array allocation failure"); 715 } 716 } 717 assert(_eden_chunk_array != NULL || _eden_chunk_capacity == 0, "Error"); 718 719 // Support for parallelizing survivor space rescan 720 if (CMSParallelRemarkEnabled && CMSParallelSurvivorRemarkEnabled) { 721 const size_t max_plab_samples = 722 ((DefNewGeneration*)_young_gen)->max_survivor_size()/MinTLABSize; 723 724 _survivor_plab_array = NEW_C_HEAP_ARRAY(ChunkArray, ParallelGCThreads); 725 _survivor_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, 2*max_plab_samples); 726 _cursor = NEW_C_HEAP_ARRAY(size_t, ParallelGCThreads); 727 if (_survivor_plab_array == NULL || _survivor_chunk_array == NULL 728 || _cursor == NULL) { 729 warning("Failed to allocate survivor plab/chunk array"); 730 if (_survivor_plab_array != NULL) { 731 FREE_C_HEAP_ARRAY(ChunkArray, _survivor_plab_array); 732 _survivor_plab_array = NULL; 733 } 734 if (_survivor_chunk_array != NULL) { 735 FREE_C_HEAP_ARRAY(HeapWord*, _survivor_chunk_array); 736 _survivor_chunk_array = NULL; 737 } 738 if (_cursor != NULL) { 739 FREE_C_HEAP_ARRAY(size_t, _cursor); 740 _cursor = NULL; 741 } 742 } else { 743 _survivor_chunk_capacity = 2*max_plab_samples; 744 for (uint i = 0; i < ParallelGCThreads; i++) { 745 HeapWord** vec = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples); 746 if (vec == NULL) { 747 warning("Failed to allocate survivor plab array"); 748 for (int j = i; j > 0; j--) { 749 FREE_C_HEAP_ARRAY(HeapWord*, _survivor_plab_array[j-1].array()); 750 } 751 FREE_C_HEAP_ARRAY(ChunkArray, _survivor_plab_array); 752 FREE_C_HEAP_ARRAY(HeapWord*, _survivor_chunk_array); 753 _survivor_plab_array = NULL; 754 _survivor_chunk_array = NULL; 755 _survivor_chunk_capacity = 0; 756 break; 757 } else { 758 ChunkArray* cur = 759 ::new (&_survivor_plab_array[i]) ChunkArray(vec, 760 max_plab_samples); 761 assert(cur->end() == 0, "Should be 0"); 762 assert(cur->array() == vec, "Should be vec"); 763 assert(cur->capacity() == max_plab_samples, "Error"); 764 } 765 } 766 } 767 } 768 assert( ( _survivor_plab_array != NULL 769 && _survivor_chunk_array != NULL) 770 || ( _survivor_chunk_capacity == 0 771 && _survivor_chunk_index == 0), 772 "Error"); 773 774 // Choose what strong roots should be scanned depending on verification options 775 // and perm gen collection mode. 776 if (!CMSClassUnloadingEnabled) { 777 // If class unloading is disabled we want to include all classes into the root set. 778 add_root_scanning_option(SharedHeap::SO_AllClasses); 779 } else { 780 add_root_scanning_option(SharedHeap::SO_SystemClasses); 781 } 782 783 NOT_PRODUCT(_overflow_counter = CMSMarkStackOverflowInterval;) 784 _gc_counters = new CollectorCounters("CMS", 1); 785 _completed_initialization = true; 786 _inter_sweep_timer.start(); // start of time 787 #ifdef SPARC 788 // Issue a stern warning, but allow use for experimentation and debugging. 789 if (VM_Version::is_sun4v() && UseMemSetInBOT) { 790 assert(!FLAG_IS_DEFAULT(UseMemSetInBOT), "Error"); 791 warning("Experimental flag -XX:+UseMemSetInBOT is known to cause instability" 792 " on sun4v; please understand that you are using at your own risk!"); 793 } 794 #endif 795 } 796 797 const char* ConcurrentMarkSweepGeneration::name() const { 798 return "concurrent mark-sweep generation"; 799 } 800 void ConcurrentMarkSweepGeneration::update_counters() { 801 if (UsePerfData) { 802 _space_counters->update_all(); 803 _gen_counters->update_all(); 804 } 805 } 806 807 // this is an optimized version of update_counters(). it takes the 808 // used value as a parameter rather than computing it. 809 // 810 void ConcurrentMarkSweepGeneration::update_counters(size_t used) { 811 if (UsePerfData) { 812 _space_counters->update_used(used); 813 _space_counters->update_capacity(); 814 _gen_counters->update_all(); 815 } 816 } 817 818 void ConcurrentMarkSweepGeneration::print() const { 819 Generation::print(); 820 cmsSpace()->print(); 821 } 822 823 #ifndef PRODUCT 824 void ConcurrentMarkSweepGeneration::print_statistics() { 825 cmsSpace()->printFLCensus(0); 826 } 827 #endif 828 829 void ConcurrentMarkSweepGeneration::printOccupancy(const char *s) { 830 GenCollectedHeap* gch = GenCollectedHeap::heap(); 831 if (PrintGCDetails) { 832 if (Verbose) { 833 gclog_or_tty->print(" [%d %s-%s: "SIZE_FORMAT"("SIZE_FORMAT")]", 834 level(), short_name(), s, used(), capacity()); 835 } else { 836 gclog_or_tty->print(" [%d %s-%s: "SIZE_FORMAT"K("SIZE_FORMAT"K)]", 837 level(), short_name(), s, used() / K, capacity() / K); 838 } 839 } 840 if (Verbose) { 841 gclog_or_tty->print(" "SIZE_FORMAT"("SIZE_FORMAT")", 842 gch->used(), gch->capacity()); 843 } else { 844 gclog_or_tty->print(" "SIZE_FORMAT"K("SIZE_FORMAT"K)", 845 gch->used() / K, gch->capacity() / K); 846 } 847 } 848 849 size_t 850 ConcurrentMarkSweepGeneration::contiguous_available() const { 851 // dld proposes an improvement in precision here. If the committed 852 // part of the space ends in a free block we should add that to 853 // uncommitted size in the calculation below. Will make this 854 // change later, staying with the approximation below for the 855 // time being. -- ysr. 856 return MAX2(_virtual_space.uncommitted_size(), unsafe_max_alloc_nogc()); 857 } 858 859 size_t 860 ConcurrentMarkSweepGeneration::unsafe_max_alloc_nogc() const { 861 return _cmsSpace->max_alloc_in_words() * HeapWordSize; 862 } 863 864 size_t ConcurrentMarkSweepGeneration::max_available() const { 865 return free() + _virtual_space.uncommitted_size(); 866 } 867 868 bool ConcurrentMarkSweepGeneration::promotion_attempt_is_safe( 869 size_t max_promotion_in_bytes, 870 bool younger_handles_promotion_failure) const { 871 872 // This is the most conservative test. Full promotion is 873 // guaranteed if this is used. The multiplicative factor is to 874 // account for the worst case "dilatation". 875 double adjusted_max_promo_bytes = _dilatation_factor * max_promotion_in_bytes; 876 if (adjusted_max_promo_bytes > (double)max_uintx) { // larger than size_t 877 adjusted_max_promo_bytes = (double)max_uintx; 878 } 879 bool result = (max_contiguous_available() >= (size_t)adjusted_max_promo_bytes); 880 881 if (younger_handles_promotion_failure && !result) { 882 // Full promotion is not guaranteed because fragmentation 883 // of the cms generation can prevent the full promotion. 884 result = (max_available() >= (size_t)adjusted_max_promo_bytes); 885 886 if (!result) { 887 // With promotion failure handling the test for the ability 888 // to support the promotion does not have to be guaranteed. 889 // Use an average of the amount promoted. 890 result = max_available() >= (size_t) 891 gc_stats()->avg_promoted()->padded_average(); 892 if (PrintGC && Verbose && result) { 893 gclog_or_tty->print_cr( 894 "\nConcurrentMarkSweepGeneration::promotion_attempt_is_safe" 895 " max_available: " SIZE_FORMAT 896 " avg_promoted: " SIZE_FORMAT, 897 max_available(), (size_t) 898 gc_stats()->avg_promoted()->padded_average()); 899 } 900 } else { 901 if (PrintGC && Verbose) { 902 gclog_or_tty->print_cr( 903 "\nConcurrentMarkSweepGeneration::promotion_attempt_is_safe" 904 " max_available: " SIZE_FORMAT 905 " adj_max_promo_bytes: " SIZE_FORMAT, 906 max_available(), (size_t)adjusted_max_promo_bytes); 907 } 908 } 909 } else { 910 if (PrintGC && Verbose) { 911 gclog_or_tty->print_cr( 912 "\nConcurrentMarkSweepGeneration::promotion_attempt_is_safe" 913 " contiguous_available: " SIZE_FORMAT 914 " adj_max_promo_bytes: " SIZE_FORMAT, 915 max_contiguous_available(), (size_t)adjusted_max_promo_bytes); 916 } 917 } 918 return result; 919 } 920 921 // At a promotion failure dump information on block layout in heap 922 // (cms old generation). 923 void ConcurrentMarkSweepGeneration::promotion_failure_occurred() { 924 if (CMSDumpAtPromotionFailure) { 925 cmsSpace()->dump_at_safepoint_with_locks(collector(), gclog_or_tty); 926 } 927 } 928 929 CompactibleSpace* 930 ConcurrentMarkSweepGeneration::first_compaction_space() const { 931 return _cmsSpace; 932 } 933 934 void ConcurrentMarkSweepGeneration::reset_after_compaction() { 935 // Clear the promotion information. These pointers can be adjusted 936 // along with all the other pointers into the heap but 937 // compaction is expected to be a rare event with 938 // a heap using cms so don't do it without seeing the need. 939 if (CollectedHeap::use_parallel_gc_threads()) { 940 for (uint i = 0; i < ParallelGCThreads; i++) { 941 _par_gc_thread_states[i]->promo.reset(); 942 } 943 } 944 } 945 946 void ConcurrentMarkSweepGeneration::space_iterate(SpaceClosure* blk, bool usedOnly) { 947 blk->do_space(_cmsSpace); 948 } 949 950 void ConcurrentMarkSweepGeneration::compute_new_size() { 951 assert_locked_or_safepoint(Heap_lock); 952 953 // If incremental collection failed, we just want to expand 954 // to the limit. 955 if (incremental_collection_failed()) { 956 clear_incremental_collection_failed(); 957 grow_to_reserved(); 958 return; 959 } 960 961 size_t expand_bytes = 0; 962 double free_percentage = ((double) free()) / capacity(); 963 double desired_free_percentage = (double) MinHeapFreeRatio / 100; 964 double maximum_free_percentage = (double) MaxHeapFreeRatio / 100; 965 966 // compute expansion delta needed for reaching desired free percentage 967 if (free_percentage < desired_free_percentage) { 968 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 969 assert(desired_capacity >= capacity(), "invalid expansion size"); 970 expand_bytes = MAX2(desired_capacity - capacity(), MinHeapDeltaBytes); 971 } 972 if (expand_bytes > 0) { 973 if (PrintGCDetails && Verbose) { 974 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 975 gclog_or_tty->print_cr("\nFrom compute_new_size: "); 976 gclog_or_tty->print_cr(" Free fraction %f", free_percentage); 977 gclog_or_tty->print_cr(" Desired free fraction %f", 978 desired_free_percentage); 979 gclog_or_tty->print_cr(" Maximum free fraction %f", 980 maximum_free_percentage); 981 gclog_or_tty->print_cr(" Capactiy "SIZE_FORMAT, capacity()/1000); 982 gclog_or_tty->print_cr(" Desired capacity "SIZE_FORMAT, 983 desired_capacity/1000); 984 int prev_level = level() - 1; 985 if (prev_level >= 0) { 986 size_t prev_size = 0; 987 GenCollectedHeap* gch = GenCollectedHeap::heap(); 988 Generation* prev_gen = gch->_gens[prev_level]; 989 prev_size = prev_gen->capacity(); 990 gclog_or_tty->print_cr(" Younger gen size "SIZE_FORMAT, 991 prev_size/1000); 992 } 993 gclog_or_tty->print_cr(" unsafe_max_alloc_nogc "SIZE_FORMAT, 994 unsafe_max_alloc_nogc()/1000); 995 gclog_or_tty->print_cr(" contiguous available "SIZE_FORMAT, 996 contiguous_available()/1000); 997 gclog_or_tty->print_cr(" Expand by "SIZE_FORMAT" (bytes)", 998 expand_bytes); 999 } 1000 // safe if expansion fails 1001 expand(expand_bytes, 0, CMSExpansionCause::_satisfy_free_ratio); 1002 if (PrintGCDetails && Verbose) { 1003 gclog_or_tty->print_cr(" Expanded free fraction %f", 1004 ((double) free()) / capacity()); 1005 } 1006 } 1007 } 1008 1009 Mutex* ConcurrentMarkSweepGeneration::freelistLock() const { 1010 return cmsSpace()->freelistLock(); 1011 } 1012 1013 HeapWord* ConcurrentMarkSweepGeneration::allocate(size_t size, 1014 bool tlab) { 1015 CMSSynchronousYieldRequest yr; 1016 MutexLockerEx x(freelistLock(), 1017 Mutex::_no_safepoint_check_flag); 1018 return have_lock_and_allocate(size, tlab); 1019 } 1020 1021 HeapWord* ConcurrentMarkSweepGeneration::have_lock_and_allocate(size_t size, 1022 bool tlab /* ignored */) { 1023 assert_lock_strong(freelistLock()); 1024 size_t adjustedSize = CompactibleFreeListSpace::adjustObjectSize(size); 1025 HeapWord* res = cmsSpace()->allocate(adjustedSize); 1026 // Allocate the object live (grey) if the background collector has 1027 // started marking. This is necessary because the marker may 1028 // have passed this address and consequently this object will 1029 // not otherwise be greyed and would be incorrectly swept up. 1030 // Note that if this object contains references, the writing 1031 // of those references will dirty the card containing this object 1032 // allowing the object to be blackened (and its references scanned) 1033 // either during a preclean phase or at the final checkpoint. 1034 if (res != NULL) { 1035 // We may block here with an uninitialized object with 1036 // its mark-bit or P-bits not yet set. Such objects need 1037 // to be safely navigable by block_start(). 1038 assert(oop(res)->klass_or_null() == NULL, "Object should be uninitialized here."); 1039 assert(!((FreeChunk*)res)->isFree(), "Error, block will look free but show wrong size"); 1040 collector()->direct_allocated(res, adjustedSize); 1041 _direct_allocated_words += adjustedSize; 1042 // allocation counters 1043 NOT_PRODUCT( 1044 _numObjectsAllocated++; 1045 _numWordsAllocated += (int)adjustedSize; 1046 ) 1047 } 1048 return res; 1049 } 1050 1051 // In the case of direct allocation by mutators in a generation that 1052 // is being concurrently collected, the object must be allocated 1053 // live (grey) if the background collector has started marking. 1054 // This is necessary because the marker may 1055 // have passed this address and consequently this object will 1056 // not otherwise be greyed and would be incorrectly swept up. 1057 // Note that if this object contains references, the writing 1058 // of those references will dirty the card containing this object 1059 // allowing the object to be blackened (and its references scanned) 1060 // either during a preclean phase or at the final checkpoint. 1061 void CMSCollector::direct_allocated(HeapWord* start, size_t size) { 1062 assert(_markBitMap.covers(start, size), "Out of bounds"); 1063 if (_collectorState >= Marking) { 1064 MutexLockerEx y(_markBitMap.lock(), 1065 Mutex::_no_safepoint_check_flag); 1066 // [see comments preceding SweepClosure::do_blk() below for details] 1067 // 1. need to mark the object as live so it isn't collected 1068 // 2. need to mark the 2nd bit to indicate the object may be uninitialized 1069 // 3. need to mark the end of the object so marking, precleaning or sweeping 1070 // can skip over uninitialized or unparsable objects. An allocated 1071 // object is considered uninitialized for our purposes as long as 1072 // its klass word is NULL. (Unparsable objects are those which are 1073 // initialized in the sense just described, but whose sizes can still 1074 // not be correctly determined. Note that the class of unparsable objects 1075 // can only occur in the perm gen. All old gen objects are parsable 1076 // as soon as they are initialized.) 1077 _markBitMap.mark(start); // object is live 1078 _markBitMap.mark(start + 1); // object is potentially uninitialized? 1079 _markBitMap.mark(start + size - 1); 1080 // mark end of object 1081 } 1082 // check that oop looks uninitialized 1083 assert(oop(start)->klass_or_null() == NULL, "_klass should be NULL"); 1084 } 1085 1086 void CMSCollector::promoted(bool par, HeapWord* start, 1087 bool is_obj_array, size_t obj_size) { 1088 assert(_markBitMap.covers(start), "Out of bounds"); 1089 // See comment in direct_allocated() about when objects should 1090 // be allocated live. 1091 if (_collectorState >= Marking) { 1092 // we already hold the marking bit map lock, taken in 1093 // the prologue 1094 if (par) { 1095 _markBitMap.par_mark(start); 1096 } else { 1097 _markBitMap.mark(start); 1098 } 1099 // We don't need to mark the object as uninitialized (as 1100 // in direct_allocated above) because this is being done with the 1101 // world stopped and the object will be initialized by the 1102 // time the marking, precleaning or sweeping get to look at it. 1103 // But see the code for copying objects into the CMS generation, 1104 // where we need to ensure that concurrent readers of the 1105 // block offset table are able to safely navigate a block that 1106 // is in flux from being free to being allocated (and in 1107 // transition while being copied into) and subsequently 1108 // becoming a bona-fide object when the copy/promotion is complete. 1109 assert(SafepointSynchronize::is_at_safepoint(), 1110 "expect promotion only at safepoints"); 1111 1112 if (_collectorState < Sweeping) { 1113 // Mark the appropriate cards in the modUnionTable, so that 1114 // this object gets scanned before the sweep. If this is 1115 // not done, CMS generation references in the object might 1116 // not get marked. 1117 // For the case of arrays, which are otherwise precisely 1118 // marked, we need to dirty the entire array, not just its head. 1119 if (is_obj_array) { 1120 // The [par_]mark_range() method expects mr.end() below to 1121 // be aligned to the granularity of a bit's representation 1122 // in the heap. In the case of the MUT below, that's a 1123 // card size. 1124 MemRegion mr(start, 1125 (HeapWord*)round_to((intptr_t)(start + obj_size), 1126 CardTableModRefBS::card_size /* bytes */)); 1127 if (par) { 1128 _modUnionTable.par_mark_range(mr); 1129 } else { 1130 _modUnionTable.mark_range(mr); 1131 } 1132 } else { // not an obj array; we can just mark the head 1133 if (par) { 1134 _modUnionTable.par_mark(start); 1135 } else { 1136 _modUnionTable.mark(start); 1137 } 1138 } 1139 } 1140 } 1141 } 1142 1143 static inline size_t percent_of_space(Space* space, HeapWord* addr) 1144 { 1145 size_t delta = pointer_delta(addr, space->bottom()); 1146 return (size_t)(delta * 100.0 / (space->capacity() / HeapWordSize)); 1147 } 1148 1149 void CMSCollector::icms_update_allocation_limits() 1150 { 1151 Generation* gen0 = GenCollectedHeap::heap()->get_gen(0); 1152 EdenSpace* eden = gen0->as_DefNewGeneration()->eden(); 1153 1154 const unsigned int duty_cycle = stats().icms_update_duty_cycle(); 1155 if (CMSTraceIncrementalPacing) { 1156 stats().print(); 1157 } 1158 1159 assert(duty_cycle <= 100, "invalid duty cycle"); 1160 if (duty_cycle != 0) { 1161 // The duty_cycle is a percentage between 0 and 100; convert to words and 1162 // then compute the offset from the endpoints of the space. 1163 size_t free_words = eden->free() / HeapWordSize; 1164 double free_words_dbl = (double)free_words; 1165 size_t duty_cycle_words = (size_t)(free_words_dbl * duty_cycle / 100.0); 1166 size_t offset_words = (free_words - duty_cycle_words) / 2; 1167 1168 _icms_start_limit = eden->top() + offset_words; 1169 _icms_stop_limit = eden->end() - offset_words; 1170 1171 // The limits may be adjusted (shifted to the right) by 1172 // CMSIncrementalOffset, to allow the application more mutator time after a 1173 // young gen gc (when all mutators were stopped) and before CMS starts and 1174 // takes away one or more cpus. 1175 if (CMSIncrementalOffset != 0) { 1176 double adjustment_dbl = free_words_dbl * CMSIncrementalOffset / 100.0; 1177 size_t adjustment = (size_t)adjustment_dbl; 1178 HeapWord* tmp_stop = _icms_stop_limit + adjustment; 1179 if (tmp_stop > _icms_stop_limit && tmp_stop < eden->end()) { 1180 _icms_start_limit += adjustment; 1181 _icms_stop_limit = tmp_stop; 1182 } 1183 } 1184 } 1185 if (duty_cycle == 0 || (_icms_start_limit == _icms_stop_limit)) { 1186 _icms_start_limit = _icms_stop_limit = eden->end(); 1187 } 1188 1189 // Install the new start limit. 1190 eden->set_soft_end(_icms_start_limit); 1191 1192 if (CMSTraceIncrementalMode) { 1193 gclog_or_tty->print(" icms alloc limits: " 1194 PTR_FORMAT "," PTR_FORMAT 1195 " (" SIZE_FORMAT "%%," SIZE_FORMAT "%%) ", 1196 _icms_start_limit, _icms_stop_limit, 1197 percent_of_space(eden, _icms_start_limit), 1198 percent_of_space(eden, _icms_stop_limit)); 1199 if (Verbose) { 1200 gclog_or_tty->print("eden: "); 1201 eden->print_on(gclog_or_tty); 1202 } 1203 } 1204 } 1205 1206 // Any changes here should try to maintain the invariant 1207 // that if this method is called with _icms_start_limit 1208 // and _icms_stop_limit both NULL, then it should return NULL 1209 // and not notify the icms thread. 1210 HeapWord* 1211 CMSCollector::allocation_limit_reached(Space* space, HeapWord* top, 1212 size_t word_size) 1213 { 1214 // A start_limit equal to end() means the duty cycle is 0, so treat that as a 1215 // nop. 1216 if (CMSIncrementalMode && _icms_start_limit != space->end()) { 1217 if (top <= _icms_start_limit) { 1218 if (CMSTraceIncrementalMode) { 1219 space->print_on(gclog_or_tty); 1220 gclog_or_tty->stamp(); 1221 gclog_or_tty->print_cr(" start limit top=" PTR_FORMAT 1222 ", new limit=" PTR_FORMAT 1223 " (" SIZE_FORMAT "%%)", 1224 top, _icms_stop_limit, 1225 percent_of_space(space, _icms_stop_limit)); 1226 } 1227 ConcurrentMarkSweepThread::start_icms(); 1228 assert(top < _icms_stop_limit, "Tautology"); 1229 if (word_size < pointer_delta(_icms_stop_limit, top)) { 1230 return _icms_stop_limit; 1231 } 1232 1233 // The allocation will cross both the _start and _stop limits, so do the 1234 // stop notification also and return end(). 1235 if (CMSTraceIncrementalMode) { 1236 space->print_on(gclog_or_tty); 1237 gclog_or_tty->stamp(); 1238 gclog_or_tty->print_cr(" +stop limit top=" PTR_FORMAT 1239 ", new limit=" PTR_FORMAT 1240 " (" SIZE_FORMAT "%%)", 1241 top, space->end(), 1242 percent_of_space(space, space->end())); 1243 } 1244 ConcurrentMarkSweepThread::stop_icms(); 1245 return space->end(); 1246 } 1247 1248 if (top <= _icms_stop_limit) { 1249 if (CMSTraceIncrementalMode) { 1250 space->print_on(gclog_or_tty); 1251 gclog_or_tty->stamp(); 1252 gclog_or_tty->print_cr(" stop limit top=" PTR_FORMAT 1253 ", new limit=" PTR_FORMAT 1254 " (" SIZE_FORMAT "%%)", 1255 top, space->end(), 1256 percent_of_space(space, space->end())); 1257 } 1258 ConcurrentMarkSweepThread::stop_icms(); 1259 return space->end(); 1260 } 1261 1262 if (CMSTraceIncrementalMode) { 1263 space->print_on(gclog_or_tty); 1264 gclog_or_tty->stamp(); 1265 gclog_or_tty->print_cr(" end limit top=" PTR_FORMAT 1266 ", new limit=" PTR_FORMAT, 1267 top, NULL); 1268 } 1269 } 1270 1271 return NULL; 1272 } 1273 1274 oop ConcurrentMarkSweepGeneration::promote(oop obj, size_t obj_size) { 1275 assert(obj_size == (size_t)obj->size(), "bad obj_size passed in"); 1276 // allocate, copy and if necessary update promoinfo -- 1277 // delegate to underlying space. 1278 assert_lock_strong(freelistLock()); 1279 1280 #ifndef PRODUCT 1281 if (Universe::heap()->promotion_should_fail()) { 1282 return NULL; 1283 } 1284 #endif // #ifndef PRODUCT 1285 1286 oop res = _cmsSpace->promote(obj, obj_size); 1287 if (res == NULL) { 1288 // expand and retry 1289 size_t s = _cmsSpace->expansionSpaceRequired(obj_size); // HeapWords 1290 expand(s*HeapWordSize, MinHeapDeltaBytes, 1291 CMSExpansionCause::_satisfy_promotion); 1292 // Since there's currently no next generation, we don't try to promote 1293 // into a more senior generation. 1294 assert(next_gen() == NULL, "assumption, based upon which no attempt " 1295 "is made to pass on a possibly failing " 1296 "promotion to next generation"); 1297 res = _cmsSpace->promote(obj, obj_size); 1298 } 1299 if (res != NULL) { 1300 // See comment in allocate() about when objects should 1301 // be allocated live. 1302 assert(obj->is_oop(), "Will dereference klass pointer below"); 1303 collector()->promoted(false, // Not parallel 1304 (HeapWord*)res, obj->is_objArray(), obj_size); 1305 // promotion counters 1306 NOT_PRODUCT( 1307 _numObjectsPromoted++; 1308 _numWordsPromoted += 1309 (int)(CompactibleFreeListSpace::adjustObjectSize(obj->size())); 1310 ) 1311 } 1312 return res; 1313 } 1314 1315 1316 HeapWord* 1317 ConcurrentMarkSweepGeneration::allocation_limit_reached(Space* space, 1318 HeapWord* top, 1319 size_t word_sz) 1320 { 1321 return collector()->allocation_limit_reached(space, top, word_sz); 1322 } 1323 1324 // IMPORTANT: Notes on object size recognition in CMS. 1325 // --------------------------------------------------- 1326 // A block of storage in the CMS generation is always in 1327 // one of three states. A free block (FREE), an allocated 1328 // object (OBJECT) whose size() method reports the correct size, 1329 // and an intermediate state (TRANSIENT) in which its size cannot 1330 // be accurately determined. 1331 // STATE IDENTIFICATION: (32 bit and 64 bit w/o COOPS) 1332 // ----------------------------------------------------- 1333 // FREE: klass_word & 1 == 1; mark_word holds block size 1334 // 1335 // OBJECT: klass_word installed; klass_word != 0 && klass_word & 1 == 0; 1336 // obj->size() computes correct size 1337 // [Perm Gen objects needs to be "parsable" before they can be navigated] 1338 // 1339 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT 1340 // 1341 // STATE IDENTIFICATION: (64 bit+COOPS) 1342 // ------------------------------------ 1343 // FREE: mark_word & CMS_FREE_BIT == 1; mark_word & ~CMS_FREE_BIT gives block_size 1344 // 1345 // OBJECT: klass_word installed; klass_word != 0; 1346 // obj->size() computes correct size 1347 // [Perm Gen comment above continues to hold] 1348 // 1349 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT 1350 // 1351 // 1352 // STATE TRANSITION DIAGRAM 1353 // 1354 // mut / parnew mut / parnew 1355 // FREE --------------------> TRANSIENT ---------------------> OBJECT --| 1356 // ^ | 1357 // |------------------------ DEAD <------------------------------------| 1358 // sweep mut 1359 // 1360 // While a block is in TRANSIENT state its size cannot be determined 1361 // so readers will either need to come back later or stall until 1362 // the size can be determined. Note that for the case of direct 1363 // allocation, P-bits, when available, may be used to determine the 1364 // size of an object that may not yet have been initialized. 1365 1366 // Things to support parallel young-gen collection. 1367 oop 1368 ConcurrentMarkSweepGeneration::par_promote(int thread_num, 1369 oop old, markOop m, 1370 size_t word_sz) { 1371 #ifndef PRODUCT 1372 if (Universe::heap()->promotion_should_fail()) { 1373 return NULL; 1374 } 1375 #endif // #ifndef PRODUCT 1376 1377 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1378 PromotionInfo* promoInfo = &ps->promo; 1379 // if we are tracking promotions, then first ensure space for 1380 // promotion (including spooling space for saving header if necessary). 1381 // then allocate and copy, then track promoted info if needed. 1382 // When tracking (see PromotionInfo::track()), the mark word may 1383 // be displaced and in this case restoration of the mark word 1384 // occurs in the (oop_since_save_marks_)iterate phase. 1385 if (promoInfo->tracking() && !promoInfo->ensure_spooling_space()) { 1386 // Out of space for allocating spooling buffers; 1387 // try expanding and allocating spooling buffers. 1388 if (!expand_and_ensure_spooling_space(promoInfo)) { 1389 return NULL; 1390 } 1391 } 1392 assert(promoInfo->has_spooling_space(), "Control point invariant"); 1393 const size_t alloc_sz = CompactibleFreeListSpace::adjustObjectSize(word_sz); 1394 HeapWord* obj_ptr = ps->lab.alloc(alloc_sz); 1395 if (obj_ptr == NULL) { 1396 obj_ptr = expand_and_par_lab_allocate(ps, alloc_sz); 1397 if (obj_ptr == NULL) { 1398 return NULL; 1399 } 1400 } 1401 oop obj = oop(obj_ptr); 1402 OrderAccess::storestore(); 1403 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1404 assert(!((FreeChunk*)obj_ptr)->isFree(), "Error, block will look free but show wrong size"); 1405 // IMPORTANT: See note on object initialization for CMS above. 1406 // Otherwise, copy the object. Here we must be careful to insert the 1407 // klass pointer last, since this marks the block as an allocated object. 1408 // Except with compressed oops it's the mark word. 1409 HeapWord* old_ptr = (HeapWord*)old; 1410 // Restore the mark word copied above. 1411 obj->set_mark(m); 1412 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1413 assert(!((FreeChunk*)obj_ptr)->isFree(), "Error, block will look free but show wrong size"); 1414 OrderAccess::storestore(); 1415 1416 if (UseCompressedOops) { 1417 // Copy gap missed by (aligned) header size calculation below 1418 obj->set_klass_gap(old->klass_gap()); 1419 } 1420 if (word_sz > (size_t)oopDesc::header_size()) { 1421 Copy::aligned_disjoint_words(old_ptr + oopDesc::header_size(), 1422 obj_ptr + oopDesc::header_size(), 1423 word_sz - oopDesc::header_size()); 1424 } 1425 1426 // Now we can track the promoted object, if necessary. We take care 1427 // to delay the transition from uninitialized to full object 1428 // (i.e., insertion of klass pointer) until after, so that it 1429 // atomically becomes a promoted object. 1430 if (promoInfo->tracking()) { 1431 promoInfo->track((PromotedObject*)obj, old->klass()); 1432 } 1433 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1434 assert(!((FreeChunk*)obj_ptr)->isFree(), "Error, block will look free but show wrong size"); 1435 assert(old->is_oop(), "Will use and dereference old klass ptr below"); 1436 1437 // Finally, install the klass pointer (this should be volatile). 1438 OrderAccess::storestore(); 1439 obj->set_klass(old->klass()); 1440 // We should now be able to calculate the right size for this object 1441 assert(obj->is_oop() && obj->size() == (int)word_sz, "Error, incorrect size computed for promoted object"); 1442 1443 collector()->promoted(true, // parallel 1444 obj_ptr, old->is_objArray(), word_sz); 1445 1446 NOT_PRODUCT( 1447 Atomic::inc_ptr(&_numObjectsPromoted); 1448 Atomic::add_ptr(alloc_sz, &_numWordsPromoted); 1449 ) 1450 1451 return obj; 1452 } 1453 1454 void 1455 ConcurrentMarkSweepGeneration:: 1456 par_promote_alloc_undo(int thread_num, 1457 HeapWord* obj, size_t word_sz) { 1458 // CMS does not support promotion undo. 1459 ShouldNotReachHere(); 1460 } 1461 1462 void 1463 ConcurrentMarkSweepGeneration:: 1464 par_promote_alloc_done(int thread_num) { 1465 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1466 ps->lab.retire(thread_num); 1467 } 1468 1469 void 1470 ConcurrentMarkSweepGeneration:: 1471 par_oop_since_save_marks_iterate_done(int thread_num) { 1472 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1473 ParScanWithoutBarrierClosure* dummy_cl = NULL; 1474 ps->promo.promoted_oops_iterate_nv(dummy_cl); 1475 } 1476 1477 // XXXPERM 1478 bool ConcurrentMarkSweepGeneration::should_collect(bool full, 1479 size_t size, 1480 bool tlab) 1481 { 1482 // We allow a STW collection only if a full 1483 // collection was requested. 1484 return full || should_allocate(size, tlab); // FIX ME !!! 1485 // This and promotion failure handling are connected at the 1486 // hip and should be fixed by untying them. 1487 } 1488 1489 bool CMSCollector::shouldConcurrentCollect() { 1490 if (_full_gc_requested) { 1491 if (Verbose && PrintGCDetails) { 1492 gclog_or_tty->print_cr("CMSCollector: collect because of explicit " 1493 " gc request (or gc_locker)"); 1494 } 1495 return true; 1496 } 1497 1498 // For debugging purposes, change the type of collection. 1499 // If the rotation is not on the concurrent collection 1500 // type, don't start a concurrent collection. 1501 NOT_PRODUCT( 1502 if (RotateCMSCollectionTypes && 1503 (_cmsGen->debug_collection_type() != 1504 ConcurrentMarkSweepGeneration::Concurrent_collection_type)) { 1505 assert(_cmsGen->debug_collection_type() != 1506 ConcurrentMarkSweepGeneration::Unknown_collection_type, 1507 "Bad cms collection type"); 1508 return false; 1509 } 1510 ) 1511 1512 FreelistLocker x(this); 1513 // ------------------------------------------------------------------ 1514 // Print out lots of information which affects the initiation of 1515 // a collection. 1516 if (PrintCMSInitiationStatistics && stats().valid()) { 1517 gclog_or_tty->print("CMSCollector shouldConcurrentCollect: "); 1518 gclog_or_tty->stamp(); 1519 gclog_or_tty->print_cr(""); 1520 stats().print_on(gclog_or_tty); 1521 gclog_or_tty->print_cr("time_until_cms_gen_full %3.7f", 1522 stats().time_until_cms_gen_full()); 1523 gclog_or_tty->print_cr("free="SIZE_FORMAT, _cmsGen->free()); 1524 gclog_or_tty->print_cr("contiguous_available="SIZE_FORMAT, 1525 _cmsGen->contiguous_available()); 1526 gclog_or_tty->print_cr("promotion_rate=%g", stats().promotion_rate()); 1527 gclog_or_tty->print_cr("cms_allocation_rate=%g", stats().cms_allocation_rate()); 1528 gclog_or_tty->print_cr("occupancy=%3.7f", _cmsGen->occupancy()); 1529 gclog_or_tty->print_cr("initiatingOccupancy=%3.7f", _cmsGen->initiating_occupancy()); 1530 gclog_or_tty->print_cr("initiatingPermOccupancy=%3.7f", _permGen->initiating_occupancy()); 1531 } 1532 // ------------------------------------------------------------------ 1533 1534 // If the estimated time to complete a cms collection (cms_duration()) 1535 // is less than the estimated time remaining until the cms generation 1536 // is full, start a collection. 1537 if (!UseCMSInitiatingOccupancyOnly) { 1538 if (stats().valid()) { 1539 if (stats().time_until_cms_start() == 0.0) { 1540 return true; 1541 } 1542 } else { 1543 // We want to conservatively collect somewhat early in order 1544 // to try and "bootstrap" our CMS/promotion statistics; 1545 // this branch will not fire after the first successful CMS 1546 // collection because the stats should then be valid. 1547 if (_cmsGen->occupancy() >= _bootstrap_occupancy) { 1548 if (Verbose && PrintGCDetails) { 1549 gclog_or_tty->print_cr( 1550 " CMSCollector: collect for bootstrapping statistics:" 1551 " occupancy = %f, boot occupancy = %f", _cmsGen->occupancy(), 1552 _bootstrap_occupancy); 1553 } 1554 return true; 1555 } 1556 } 1557 } 1558 1559 // Otherwise, we start a collection cycle if either the perm gen or 1560 // old gen want a collection cycle started. Each may use 1561 // an appropriate criterion for making this decision. 1562 // XXX We need to make sure that the gen expansion 1563 // criterion dovetails well with this. XXX NEED TO FIX THIS 1564 if (_cmsGen->should_concurrent_collect()) { 1565 if (Verbose && PrintGCDetails) { 1566 gclog_or_tty->print_cr("CMS old gen initiated"); 1567 } 1568 return true; 1569 } 1570 1571 // We start a collection if we believe an incremental collection may fail; 1572 // this is not likely to be productive in practice because it's probably too 1573 // late anyway. 1574 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1575 assert(gch->collector_policy()->is_two_generation_policy(), 1576 "You may want to check the correctness of the following"); 1577 if (gch->incremental_collection_will_fail()) { 1578 if (PrintGCDetails && Verbose) { 1579 gclog_or_tty->print("CMSCollector: collect because incremental collection will fail "); 1580 } 1581 return true; 1582 } 1583 1584 if (CMSClassUnloadingEnabled && _permGen->should_concurrent_collect()) { 1585 bool res = update_should_unload_classes(); 1586 if (res) { 1587 if (Verbose && PrintGCDetails) { 1588 gclog_or_tty->print_cr("CMS perm gen initiated"); 1589 } 1590 return true; 1591 } 1592 } 1593 return false; 1594 } 1595 1596 // Clear _expansion_cause fields of constituent generations 1597 void CMSCollector::clear_expansion_cause() { 1598 _cmsGen->clear_expansion_cause(); 1599 _permGen->clear_expansion_cause(); 1600 } 1601 1602 // We should be conservative in starting a collection cycle. To 1603 // start too eagerly runs the risk of collecting too often in the 1604 // extreme. To collect too rarely falls back on full collections, 1605 // which works, even if not optimum in terms of concurrent work. 1606 // As a work around for too eagerly collecting, use the flag 1607 // UseCMSInitiatingOccupancyOnly. This also has the advantage of 1608 // giving the user an easily understandable way of controlling the 1609 // collections. 1610 // We want to start a new collection cycle if any of the following 1611 // conditions hold: 1612 // . our current occupancy exceeds the configured initiating occupancy 1613 // for this generation, or 1614 // . we recently needed to expand this space and have not, since that 1615 // expansion, done a collection of this generation, or 1616 // . the underlying space believes that it may be a good idea to initiate 1617 // a concurrent collection (this may be based on criteria such as the 1618 // following: the space uses linear allocation and linear allocation is 1619 // going to fail, or there is believed to be excessive fragmentation in 1620 // the generation, etc... or ... 1621 // [.(currently done by CMSCollector::shouldConcurrentCollect() only for 1622 // the case of the old generation, not the perm generation; see CR 6543076): 1623 // we may be approaching a point at which allocation requests may fail because 1624 // we will be out of sufficient free space given allocation rate estimates.] 1625 bool ConcurrentMarkSweepGeneration::should_concurrent_collect() const { 1626 1627 assert_lock_strong(freelistLock()); 1628 if (occupancy() > initiating_occupancy()) { 1629 if (PrintGCDetails && Verbose) { 1630 gclog_or_tty->print(" %s: collect because of occupancy %f / %f ", 1631 short_name(), occupancy(), initiating_occupancy()); 1632 } 1633 return true; 1634 } 1635 if (UseCMSInitiatingOccupancyOnly) { 1636 return false; 1637 } 1638 if (expansion_cause() == CMSExpansionCause::_satisfy_allocation) { 1639 if (PrintGCDetails && Verbose) { 1640 gclog_or_tty->print(" %s: collect because expanded for allocation ", 1641 short_name()); 1642 } 1643 return true; 1644 } 1645 if (_cmsSpace->should_concurrent_collect()) { 1646 if (PrintGCDetails && Verbose) { 1647 gclog_or_tty->print(" %s: collect because cmsSpace says so ", 1648 short_name()); 1649 } 1650 return true; 1651 } 1652 return false; 1653 } 1654 1655 void ConcurrentMarkSweepGeneration::collect(bool full, 1656 bool clear_all_soft_refs, 1657 size_t size, 1658 bool tlab) 1659 { 1660 collector()->collect(full, clear_all_soft_refs, size, tlab); 1661 } 1662 1663 void CMSCollector::collect(bool full, 1664 bool clear_all_soft_refs, 1665 size_t size, 1666 bool tlab) 1667 { 1668 if (!UseCMSCollectionPassing && _collectorState > Idling) { 1669 // For debugging purposes skip the collection if the state 1670 // is not currently idle 1671 if (TraceCMSState) { 1672 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " skipped full:%d CMS state %d", 1673 Thread::current(), full, _collectorState); 1674 } 1675 return; 1676 } 1677 1678 // The following "if" branch is present for defensive reasons. 1679 // In the current uses of this interface, it can be replaced with: 1680 // assert(!GC_locker.is_active(), "Can't be called otherwise"); 1681 // But I am not placing that assert here to allow future 1682 // generality in invoking this interface. 1683 if (GC_locker::is_active()) { 1684 // A consistency test for GC_locker 1685 assert(GC_locker::needs_gc(), "Should have been set already"); 1686 // Skip this foreground collection, instead 1687 // expanding the heap if necessary. 1688 // Need the free list locks for the call to free() in compute_new_size() 1689 compute_new_size(); 1690 return; 1691 } 1692 acquire_control_and_collect(full, clear_all_soft_refs); 1693 _full_gcs_since_conc_gc++; 1694 1695 } 1696 1697 void CMSCollector::request_full_gc(unsigned int full_gc_count) { 1698 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1699 unsigned int gc_count = gch->total_full_collections(); 1700 if (gc_count == full_gc_count) { 1701 MutexLockerEx y(CGC_lock, Mutex::_no_safepoint_check_flag); 1702 _full_gc_requested = true; 1703 CGC_lock->notify(); // nudge CMS thread 1704 } 1705 } 1706 1707 1708 // The foreground and background collectors need to coordinate in order 1709 // to make sure that they do not mutually interfere with CMS collections. 1710 // When a background collection is active, 1711 // the foreground collector may need to take over (preempt) and 1712 // synchronously complete an ongoing collection. Depending on the 1713 // frequency of the background collections and the heap usage 1714 // of the application, this preemption can be seldom or frequent. 1715 // There are only certain 1716 // points in the background collection that the "collection-baton" 1717 // can be passed to the foreground collector. 1718 // 1719 // The foreground collector will wait for the baton before 1720 // starting any part of the collection. The foreground collector 1721 // will only wait at one location. 1722 // 1723 // The background collector will yield the baton before starting a new 1724 // phase of the collection (e.g., before initial marking, marking from roots, 1725 // precleaning, final re-mark, sweep etc.) This is normally done at the head 1726 // of the loop which switches the phases. The background collector does some 1727 // of the phases (initial mark, final re-mark) with the world stopped. 1728 // Because of locking involved in stopping the world, 1729 // the foreground collector should not block waiting for the background 1730 // collector when it is doing a stop-the-world phase. The background 1731 // collector will yield the baton at an additional point just before 1732 // it enters a stop-the-world phase. Once the world is stopped, the 1733 // background collector checks the phase of the collection. If the 1734 // phase has not changed, it proceeds with the collection. If the 1735 // phase has changed, it skips that phase of the collection. See 1736 // the comments on the use of the Heap_lock in collect_in_background(). 1737 // 1738 // Variable used in baton passing. 1739 // _foregroundGCIsActive - Set to true by the foreground collector when 1740 // it wants the baton. The foreground clears it when it has finished 1741 // the collection. 1742 // _foregroundGCShouldWait - Set to true by the background collector 1743 // when it is running. The foreground collector waits while 1744 // _foregroundGCShouldWait is true. 1745 // CGC_lock - monitor used to protect access to the above variables 1746 // and to notify the foreground and background collectors. 1747 // _collectorState - current state of the CMS collection. 1748 // 1749 // The foreground collector 1750 // acquires the CGC_lock 1751 // sets _foregroundGCIsActive 1752 // waits on the CGC_lock for _foregroundGCShouldWait to be false 1753 // various locks acquired in preparation for the collection 1754 // are released so as not to block the background collector 1755 // that is in the midst of a collection 1756 // proceeds with the collection 1757 // clears _foregroundGCIsActive 1758 // returns 1759 // 1760 // The background collector in a loop iterating on the phases of the 1761 // collection 1762 // acquires the CGC_lock 1763 // sets _foregroundGCShouldWait 1764 // if _foregroundGCIsActive is set 1765 // clears _foregroundGCShouldWait, notifies _CGC_lock 1766 // waits on _CGC_lock for _foregroundGCIsActive to become false 1767 // and exits the loop. 1768 // otherwise 1769 // proceed with that phase of the collection 1770 // if the phase is a stop-the-world phase, 1771 // yield the baton once more just before enqueueing 1772 // the stop-world CMS operation (executed by the VM thread). 1773 // returns after all phases of the collection are done 1774 // 1775 1776 void CMSCollector::acquire_control_and_collect(bool full, 1777 bool clear_all_soft_refs) { 1778 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint"); 1779 assert(!Thread::current()->is_ConcurrentGC_thread(), 1780 "shouldn't try to acquire control from self!"); 1781 1782 // Start the protocol for acquiring control of the 1783 // collection from the background collector (aka CMS thread). 1784 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 1785 "VM thread should have CMS token"); 1786 // Remember the possibly interrupted state of an ongoing 1787 // concurrent collection 1788 CollectorState first_state = _collectorState; 1789 1790 // Signal to a possibly ongoing concurrent collection that 1791 // we want to do a foreground collection. 1792 _foregroundGCIsActive = true; 1793 1794 // Disable incremental mode during a foreground collection. 1795 ICMSDisabler icms_disabler; 1796 1797 // release locks and wait for a notify from the background collector 1798 // releasing the locks in only necessary for phases which 1799 // do yields to improve the granularity of the collection. 1800 assert_lock_strong(bitMapLock()); 1801 // We need to lock the Free list lock for the space that we are 1802 // currently collecting. 1803 assert(haveFreelistLocks(), "Must be holding free list locks"); 1804 bitMapLock()->unlock(); 1805 releaseFreelistLocks(); 1806 { 1807 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1808 if (_foregroundGCShouldWait) { 1809 // We are going to be waiting for action for the CMS thread; 1810 // it had better not be gone (for instance at shutdown)! 1811 assert(ConcurrentMarkSweepThread::cmst() != NULL, 1812 "CMS thread must be running"); 1813 // Wait here until the background collector gives us the go-ahead 1814 ConcurrentMarkSweepThread::clear_CMS_flag( 1815 ConcurrentMarkSweepThread::CMS_vm_has_token); // release token 1816 // Get a possibly blocked CMS thread going: 1817 // Note that we set _foregroundGCIsActive true above, 1818 // without protection of the CGC_lock. 1819 CGC_lock->notify(); 1820 assert(!ConcurrentMarkSweepThread::vm_thread_wants_cms_token(), 1821 "Possible deadlock"); 1822 while (_foregroundGCShouldWait) { 1823 // wait for notification 1824 CGC_lock->wait(Mutex::_no_safepoint_check_flag); 1825 // Possibility of delay/starvation here, since CMS token does 1826 // not know to give priority to VM thread? Actually, i think 1827 // there wouldn't be any delay/starvation, but the proof of 1828 // that "fact" (?) appears non-trivial. XXX 20011219YSR 1829 } 1830 ConcurrentMarkSweepThread::set_CMS_flag( 1831 ConcurrentMarkSweepThread::CMS_vm_has_token); 1832 } 1833 } 1834 // The CMS_token is already held. Get back the other locks. 1835 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 1836 "VM thread should have CMS token"); 1837 getFreelistLocks(); 1838 bitMapLock()->lock_without_safepoint_check(); 1839 if (TraceCMSState) { 1840 gclog_or_tty->print_cr("CMS foreground collector has asked for control " 1841 INTPTR_FORMAT " with first state %d", Thread::current(), first_state); 1842 gclog_or_tty->print_cr(" gets control with state %d", _collectorState); 1843 } 1844 1845 // Check if we need to do a compaction, or if not, whether 1846 // we need to start the mark-sweep from scratch. 1847 bool should_compact = false; 1848 bool should_start_over = false; 1849 decide_foreground_collection_type(clear_all_soft_refs, 1850 &should_compact, &should_start_over); 1851 1852 NOT_PRODUCT( 1853 if (RotateCMSCollectionTypes) { 1854 if (_cmsGen->debug_collection_type() == 1855 ConcurrentMarkSweepGeneration::MSC_foreground_collection_type) { 1856 should_compact = true; 1857 } else if (_cmsGen->debug_collection_type() == 1858 ConcurrentMarkSweepGeneration::MS_foreground_collection_type) { 1859 should_compact = false; 1860 } 1861 } 1862 ) 1863 1864 if (PrintGCDetails && first_state > Idling) { 1865 GCCause::Cause cause = GenCollectedHeap::heap()->gc_cause(); 1866 if (GCCause::is_user_requested_gc(cause) || 1867 GCCause::is_serviceability_requested_gc(cause)) { 1868 gclog_or_tty->print(" (concurrent mode interrupted)"); 1869 } else { 1870 gclog_or_tty->print(" (concurrent mode failure)"); 1871 } 1872 } 1873 1874 if (should_compact) { 1875 // If the collection is being acquired from the background 1876 // collector, there may be references on the discovered 1877 // references lists that have NULL referents (being those 1878 // that were concurrently cleared by a mutator) or 1879 // that are no longer active (having been enqueued concurrently 1880 // by the mutator). 1881 // Scrub the list of those references because Mark-Sweep-Compact 1882 // code assumes referents are not NULL and that all discovered 1883 // Reference objects are active. 1884 ref_processor()->clean_up_discovered_references(); 1885 1886 do_compaction_work(clear_all_soft_refs); 1887 1888 // Has the GC time limit been exceeded? 1889 DefNewGeneration* young_gen = _young_gen->as_DefNewGeneration(); 1890 size_t max_eden_size = young_gen->max_capacity() - 1891 young_gen->to()->capacity() - 1892 young_gen->from()->capacity(); 1893 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1894 GCCause::Cause gc_cause = gch->gc_cause(); 1895 size_policy()->check_gc_overhead_limit(_young_gen->used(), 1896 young_gen->eden()->used(), 1897 _cmsGen->max_capacity(), 1898 max_eden_size, 1899 full, 1900 gc_cause, 1901 gch->collector_policy()); 1902 } else { 1903 do_mark_sweep_work(clear_all_soft_refs, first_state, 1904 should_start_over); 1905 } 1906 // Reset the expansion cause, now that we just completed 1907 // a collection cycle. 1908 clear_expansion_cause(); 1909 _foregroundGCIsActive = false; 1910 return; 1911 } 1912 1913 // Resize the perm generation and the tenured generation 1914 // after obtaining the free list locks for the 1915 // two generations. 1916 void CMSCollector::compute_new_size() { 1917 assert_locked_or_safepoint(Heap_lock); 1918 FreelistLocker z(this); 1919 _permGen->compute_new_size(); 1920 _cmsGen->compute_new_size(); 1921 } 1922 1923 // A work method used by foreground collection to determine 1924 // what type of collection (compacting or not, continuing or fresh) 1925 // it should do. 1926 // NOTE: the intent is to make UseCMSCompactAtFullCollection 1927 // and CMSCompactWhenClearAllSoftRefs the default in the future 1928 // and do away with the flags after a suitable period. 1929 void CMSCollector::decide_foreground_collection_type( 1930 bool clear_all_soft_refs, bool* should_compact, 1931 bool* should_start_over) { 1932 // Normally, we'll compact only if the UseCMSCompactAtFullCollection 1933 // flag is set, and we have either requested a System.gc() or 1934 // the number of full gc's since the last concurrent cycle 1935 // has exceeded the threshold set by CMSFullGCsBeforeCompaction, 1936 // or if an incremental collection has failed 1937 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1938 assert(gch->collector_policy()->is_two_generation_policy(), 1939 "You may want to check the correctness of the following"); 1940 // Inform cms gen if this was due to partial collection failing. 1941 // The CMS gen may use this fact to determine its expansion policy. 1942 if (gch->incremental_collection_will_fail()) { 1943 assert(!_cmsGen->incremental_collection_failed(), 1944 "Should have been noticed, reacted to and cleared"); 1945 _cmsGen->set_incremental_collection_failed(); 1946 } 1947 *should_compact = 1948 UseCMSCompactAtFullCollection && 1949 ((_full_gcs_since_conc_gc >= CMSFullGCsBeforeCompaction) || 1950 GCCause::is_user_requested_gc(gch->gc_cause()) || 1951 gch->incremental_collection_will_fail()); 1952 *should_start_over = false; 1953 if (clear_all_soft_refs && !*should_compact) { 1954 // We are about to do a last ditch collection attempt 1955 // so it would normally make sense to do a compaction 1956 // to reclaim as much space as possible. 1957 if (CMSCompactWhenClearAllSoftRefs) { 1958 // Default: The rationale is that in this case either 1959 // we are past the final marking phase, in which case 1960 // we'd have to start over, or so little has been done 1961 // that there's little point in saving that work. Compaction 1962 // appears to be the sensible choice in either case. 1963 *should_compact = true; 1964 } else { 1965 // We have been asked to clear all soft refs, but not to 1966 // compact. Make sure that we aren't past the final checkpoint 1967 // phase, for that is where we process soft refs. If we are already 1968 // past that phase, we'll need to redo the refs discovery phase and 1969 // if necessary clear soft refs that weren't previously 1970 // cleared. We do so by remembering the phase in which 1971 // we came in, and if we are past the refs processing 1972 // phase, we'll choose to just redo the mark-sweep 1973 // collection from scratch. 1974 if (_collectorState > FinalMarking) { 1975 // We are past the refs processing phase; 1976 // start over and do a fresh synchronous CMS cycle 1977 _collectorState = Resetting; // skip to reset to start new cycle 1978 reset(false /* == !asynch */); 1979 *should_start_over = true; 1980 } // else we can continue a possibly ongoing current cycle 1981 } 1982 } 1983 } 1984 1985 // A work method used by the foreground collector to do 1986 // a mark-sweep-compact. 1987 void CMSCollector::do_compaction_work(bool clear_all_soft_refs) { 1988 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1989 TraceTime t("CMS:MSC ", PrintGCDetails && Verbose, true, gclog_or_tty); 1990 if (PrintGC && Verbose && !(GCCause::is_user_requested_gc(gch->gc_cause()))) { 1991 gclog_or_tty->print_cr("Compact ConcurrentMarkSweepGeneration after %d " 1992 "collections passed to foreground collector", _full_gcs_since_conc_gc); 1993 } 1994 1995 // Sample collection interval time and reset for collection pause. 1996 if (UseAdaptiveSizePolicy) { 1997 size_policy()->msc_collection_begin(); 1998 } 1999 2000 // Temporarily widen the span of the weak reference processing to 2001 // the entire heap. 2002 MemRegion new_span(GenCollectedHeap::heap()->reserved_region()); 2003 ReferenceProcessorSpanMutator x(ref_processor(), new_span); 2004 2005 // Temporarily, clear the "is_alive_non_header" field of the 2006 // reference processor. 2007 ReferenceProcessorIsAliveMutator y(ref_processor(), NULL); 2008 2009 // Temporarily make reference _processing_ single threaded (non-MT). 2010 ReferenceProcessorMTProcMutator z(ref_processor(), false); 2011 2012 // Temporarily make refs discovery atomic 2013 ReferenceProcessorAtomicMutator w(ref_processor(), true); 2014 2015 ref_processor()->set_enqueuing_is_done(false); 2016 ref_processor()->enable_discovery(); 2017 ref_processor()->setup_policy(clear_all_soft_refs); 2018 // If an asynchronous collection finishes, the _modUnionTable is 2019 // all clear. If we are assuming the collection from an asynchronous 2020 // collection, clear the _modUnionTable. 2021 assert(_collectorState != Idling || _modUnionTable.isAllClear(), 2022 "_modUnionTable should be clear if the baton was not passed"); 2023 _modUnionTable.clear_all(); 2024 2025 // We must adjust the allocation statistics being maintained 2026 // in the free list space. We do so by reading and clearing 2027 // the sweep timer and updating the block flux rate estimates below. 2028 assert(!_intra_sweep_timer.is_active(), "_intra_sweep_timer should be inactive"); 2029 if (_inter_sweep_timer.is_active()) { 2030 _inter_sweep_timer.stop(); 2031 // Note that we do not use this sample to update the _inter_sweep_estimate. 2032 _cmsGen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()), 2033 _inter_sweep_estimate.padded_average(), 2034 _intra_sweep_estimate.padded_average()); 2035 } 2036 2037 { 2038 TraceCMSMemoryManagerStats(); 2039 } 2040 GenMarkSweep::invoke_at_safepoint(_cmsGen->level(), 2041 ref_processor(), clear_all_soft_refs); 2042 #ifdef ASSERT 2043 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 2044 size_t free_size = cms_space->free(); 2045 assert(free_size == 2046 pointer_delta(cms_space->end(), cms_space->compaction_top()) 2047 * HeapWordSize, 2048 "All the free space should be compacted into one chunk at top"); 2049 assert(cms_space->dictionary()->totalChunkSize( 2050 debug_only(cms_space->freelistLock())) == 0 || 2051 cms_space->totalSizeInIndexedFreeLists() == 0, 2052 "All the free space should be in a single chunk"); 2053 size_t num = cms_space->totalCount(); 2054 assert((free_size == 0 && num == 0) || 2055 (free_size > 0 && (num == 1 || num == 2)), 2056 "There should be at most 2 free chunks after compaction"); 2057 #endif // ASSERT 2058 _collectorState = Resetting; 2059 assert(_restart_addr == NULL, 2060 "Should have been NULL'd before baton was passed"); 2061 reset(false /* == !asynch */); 2062 _cmsGen->reset_after_compaction(); 2063 _concurrent_cycles_since_last_unload = 0; 2064 2065 if (verifying() && !should_unload_classes()) { 2066 perm_gen_verify_bit_map()->clear_all(); 2067 } 2068 2069 // Clear any data recorded in the PLAB chunk arrays. 2070 if (_survivor_plab_array != NULL) { 2071 reset_survivor_plab_arrays(); 2072 } 2073 2074 // Adjust the per-size allocation stats for the next epoch. 2075 _cmsGen->cmsSpace()->endSweepFLCensus(sweep_count() /* fake */); 2076 // Restart the "inter sweep timer" for the next epoch. 2077 _inter_sweep_timer.reset(); 2078 _inter_sweep_timer.start(); 2079 2080 // Sample collection pause time and reset for collection interval. 2081 if (UseAdaptiveSizePolicy) { 2082 size_policy()->msc_collection_end(gch->gc_cause()); 2083 } 2084 2085 // For a mark-sweep-compact, compute_new_size() will be called 2086 // in the heap's do_collection() method. 2087 } 2088 2089 // A work method used by the foreground collector to do 2090 // a mark-sweep, after taking over from a possibly on-going 2091 // concurrent mark-sweep collection. 2092 void CMSCollector::do_mark_sweep_work(bool clear_all_soft_refs, 2093 CollectorState first_state, bool should_start_over) { 2094 if (PrintGC && Verbose) { 2095 gclog_or_tty->print_cr("Pass concurrent collection to foreground " 2096 "collector with count %d", 2097 _full_gcs_since_conc_gc); 2098 } 2099 switch (_collectorState) { 2100 case Idling: 2101 if (first_state == Idling || should_start_over) { 2102 // The background GC was not active, or should 2103 // restarted from scratch; start the cycle. 2104 _collectorState = InitialMarking; 2105 } 2106 // If first_state was not Idling, then a background GC 2107 // was in progress and has now finished. No need to do it 2108 // again. Leave the state as Idling. 2109 break; 2110 case Precleaning: 2111 // In the foreground case don't do the precleaning since 2112 // it is not done concurrently and there is extra work 2113 // required. 2114 _collectorState = FinalMarking; 2115 } 2116 if (PrintGCDetails && 2117 (_collectorState > Idling || 2118 !GCCause::is_user_requested_gc(GenCollectedHeap::heap()->gc_cause()))) { 2119 gclog_or_tty->print(" (concurrent mode failure)"); 2120 } 2121 collect_in_foreground(clear_all_soft_refs); 2122 2123 // For a mark-sweep, compute_new_size() will be called 2124 // in the heap's do_collection() method. 2125 } 2126 2127 2128 void CMSCollector::getFreelistLocks() const { 2129 // Get locks for all free lists in all generations that this 2130 // collector is responsible for 2131 _cmsGen->freelistLock()->lock_without_safepoint_check(); 2132 _permGen->freelistLock()->lock_without_safepoint_check(); 2133 } 2134 2135 void CMSCollector::releaseFreelistLocks() const { 2136 // Release locks for all free lists in all generations that this 2137 // collector is responsible for 2138 _cmsGen->freelistLock()->unlock(); 2139 _permGen->freelistLock()->unlock(); 2140 } 2141 2142 bool CMSCollector::haveFreelistLocks() const { 2143 // Check locks for all free lists in all generations that this 2144 // collector is responsible for 2145 assert_lock_strong(_cmsGen->freelistLock()); 2146 assert_lock_strong(_permGen->freelistLock()); 2147 PRODUCT_ONLY(ShouldNotReachHere()); 2148 return true; 2149 } 2150 2151 // A utility class that is used by the CMS collector to 2152 // temporarily "release" the foreground collector from its 2153 // usual obligation to wait for the background collector to 2154 // complete an ongoing phase before proceeding. 2155 class ReleaseForegroundGC: public StackObj { 2156 private: 2157 CMSCollector* _c; 2158 public: 2159 ReleaseForegroundGC(CMSCollector* c) : _c(c) { 2160 assert(_c->_foregroundGCShouldWait, "Else should not need to call"); 2161 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 2162 // allow a potentially blocked foreground collector to proceed 2163 _c->_foregroundGCShouldWait = false; 2164 if (_c->_foregroundGCIsActive) { 2165 CGC_lock->notify(); 2166 } 2167 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 2168 "Possible deadlock"); 2169 } 2170 2171 ~ReleaseForegroundGC() { 2172 assert(!_c->_foregroundGCShouldWait, "Usage protocol violation?"); 2173 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 2174 _c->_foregroundGCShouldWait = true; 2175 } 2176 }; 2177 2178 // There are separate collect_in_background and collect_in_foreground because of 2179 // the different locking requirements of the background collector and the 2180 // foreground collector. There was originally an attempt to share 2181 // one "collect" method between the background collector and the foreground 2182 // collector but the if-then-else required made it cleaner to have 2183 // separate methods. 2184 void CMSCollector::collect_in_background(bool clear_all_soft_refs) { 2185 assert(Thread::current()->is_ConcurrentGC_thread(), 2186 "A CMS asynchronous collection is only allowed on a CMS thread."); 2187 2188 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2189 { 2190 bool safepoint_check = Mutex::_no_safepoint_check_flag; 2191 MutexLockerEx hl(Heap_lock, safepoint_check); 2192 FreelistLocker fll(this); 2193 MutexLockerEx x(CGC_lock, safepoint_check); 2194 if (_foregroundGCIsActive || !UseAsyncConcMarkSweepGC) { 2195 // The foreground collector is active or we're 2196 // not using asynchronous collections. Skip this 2197 // background collection. 2198 assert(!_foregroundGCShouldWait, "Should be clear"); 2199 return; 2200 } else { 2201 assert(_collectorState == Idling, "Should be idling before start."); 2202 _collectorState = InitialMarking; 2203 // Reset the expansion cause, now that we are about to begin 2204 // a new cycle. 2205 clear_expansion_cause(); 2206 } 2207 // Decide if we want to enable class unloading as part of the 2208 // ensuing concurrent GC cycle. 2209 update_should_unload_classes(); 2210 _full_gc_requested = false; // acks all outstanding full gc requests 2211 // Signal that we are about to start a collection 2212 gch->increment_total_full_collections(); // ... starting a collection cycle 2213 _collection_count_start = gch->total_full_collections(); 2214 } 2215 2216 // Used for PrintGC 2217 size_t prev_used; 2218 if (PrintGC && Verbose) { 2219 prev_used = _cmsGen->used(); // XXXPERM 2220 } 2221 2222 // The change of the collection state is normally done at this level; 2223 // the exceptions are phases that are executed while the world is 2224 // stopped. For those phases the change of state is done while the 2225 // world is stopped. For baton passing purposes this allows the 2226 // background collector to finish the phase and change state atomically. 2227 // The foreground collector cannot wait on a phase that is done 2228 // while the world is stopped because the foreground collector already 2229 // has the world stopped and would deadlock. 2230 while (_collectorState != Idling) { 2231 if (TraceCMSState) { 2232 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " in CMS state %d", 2233 Thread::current(), _collectorState); 2234 } 2235 // The foreground collector 2236 // holds the Heap_lock throughout its collection. 2237 // holds the CMS token (but not the lock) 2238 // except while it is waiting for the background collector to yield. 2239 // 2240 // The foreground collector should be blocked (not for long) 2241 // if the background collector is about to start a phase 2242 // executed with world stopped. If the background 2243 // collector has already started such a phase, the 2244 // foreground collector is blocked waiting for the 2245 // Heap_lock. The stop-world phases (InitialMarking and FinalMarking) 2246 // are executed in the VM thread. 2247 // 2248 // The locking order is 2249 // PendingListLock (PLL) -- if applicable (FinalMarking) 2250 // Heap_lock (both this & PLL locked in VM_CMS_Operation::prologue()) 2251 // CMS token (claimed in 2252 // stop_world_and_do() --> 2253 // safepoint_synchronize() --> 2254 // CMSThread::synchronize()) 2255 2256 { 2257 // Check if the FG collector wants us to yield. 2258 CMSTokenSync x(true); // is cms thread 2259 if (waitForForegroundGC()) { 2260 // We yielded to a foreground GC, nothing more to be 2261 // done this round. 2262 assert(_foregroundGCShouldWait == false, "We set it to false in " 2263 "waitForForegroundGC()"); 2264 if (TraceCMSState) { 2265 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT 2266 " exiting collection CMS state %d", 2267 Thread::current(), _collectorState); 2268 } 2269 return; 2270 } else { 2271 // The background collector can run but check to see if the 2272 // foreground collector has done a collection while the 2273 // background collector was waiting to get the CGC_lock 2274 // above. If yes, break so that _foregroundGCShouldWait 2275 // is cleared before returning. 2276 if (_collectorState == Idling) { 2277 break; 2278 } 2279 } 2280 } 2281 2282 assert(_foregroundGCShouldWait, "Foreground collector, if active, " 2283 "should be waiting"); 2284 2285 switch (_collectorState) { 2286 case InitialMarking: 2287 { 2288 ReleaseForegroundGC x(this); 2289 stats().record_cms_begin(); 2290 2291 VM_CMS_Initial_Mark initial_mark_op(this); 2292 VMThread::execute(&initial_mark_op); 2293 } 2294 // The collector state may be any legal state at this point 2295 // since the background collector may have yielded to the 2296 // foreground collector. 2297 break; 2298 case Marking: 2299 // initial marking in checkpointRootsInitialWork has been completed 2300 if (markFromRoots(true)) { // we were successful 2301 assert(_collectorState == Precleaning, "Collector state should " 2302 "have changed"); 2303 } else { 2304 assert(_foregroundGCIsActive, "Internal state inconsistency"); 2305 } 2306 break; 2307 case Precleaning: 2308 if (UseAdaptiveSizePolicy) { 2309 size_policy()->concurrent_precleaning_begin(); 2310 } 2311 // marking from roots in markFromRoots has been completed 2312 preclean(); 2313 if (UseAdaptiveSizePolicy) { 2314 size_policy()->concurrent_precleaning_end(); 2315 } 2316 assert(_collectorState == AbortablePreclean || 2317 _collectorState == FinalMarking, 2318 "Collector state should have changed"); 2319 break; 2320 case AbortablePreclean: 2321 if (UseAdaptiveSizePolicy) { 2322 size_policy()->concurrent_phases_resume(); 2323 } 2324 abortable_preclean(); 2325 if (UseAdaptiveSizePolicy) { 2326 size_policy()->concurrent_precleaning_end(); 2327 } 2328 assert(_collectorState == FinalMarking, "Collector state should " 2329 "have changed"); 2330 break; 2331 case FinalMarking: 2332 { 2333 ReleaseForegroundGC x(this); 2334 2335 VM_CMS_Final_Remark final_remark_op(this); 2336 VMThread::execute(&final_remark_op); 2337 } 2338 assert(_foregroundGCShouldWait, "block post-condition"); 2339 break; 2340 case Sweeping: 2341 if (UseAdaptiveSizePolicy) { 2342 size_policy()->concurrent_sweeping_begin(); 2343 } 2344 // final marking in checkpointRootsFinal has been completed 2345 sweep(true); 2346 assert(_collectorState == Resizing, "Collector state change " 2347 "to Resizing must be done under the free_list_lock"); 2348 _full_gcs_since_conc_gc = 0; 2349 2350 // Stop the timers for adaptive size policy for the concurrent phases 2351 if (UseAdaptiveSizePolicy) { 2352 size_policy()->concurrent_sweeping_end(); 2353 size_policy()->concurrent_phases_end(gch->gc_cause(), 2354 gch->prev_gen(_cmsGen)->capacity(), 2355 _cmsGen->free()); 2356 } 2357 2358 case Resizing: { 2359 // Sweeping has been completed... 2360 // At this point the background collection has completed. 2361 // Don't move the call to compute_new_size() down 2362 // into code that might be executed if the background 2363 // collection was preempted. 2364 { 2365 ReleaseForegroundGC x(this); // unblock FG collection 2366 MutexLockerEx y(Heap_lock, Mutex::_no_safepoint_check_flag); 2367 CMSTokenSync z(true); // not strictly needed. 2368 if (_collectorState == Resizing) { 2369 compute_new_size(); 2370 _collectorState = Resetting; 2371 } else { 2372 assert(_collectorState == Idling, "The state should only change" 2373 " because the foreground collector has finished the collection"); 2374 } 2375 } 2376 break; 2377 } 2378 case Resetting: 2379 // CMS heap resizing has been completed 2380 reset(true); 2381 assert(_collectorState == Idling, "Collector state should " 2382 "have changed"); 2383 stats().record_cms_end(); 2384 // Don't move the concurrent_phases_end() and compute_new_size() 2385 // calls to here because a preempted background collection 2386 // has it's state set to "Resetting". 2387 break; 2388 case Idling: 2389 default: 2390 ShouldNotReachHere(); 2391 break; 2392 } 2393 if (TraceCMSState) { 2394 gclog_or_tty->print_cr(" Thread " INTPTR_FORMAT " done - next CMS state %d", 2395 Thread::current(), _collectorState); 2396 } 2397 assert(_foregroundGCShouldWait, "block post-condition"); 2398 } 2399 2400 // Should this be in gc_epilogue? 2401 collector_policy()->counters()->update_counters(); 2402 2403 { 2404 // Clear _foregroundGCShouldWait and, in the event that the 2405 // foreground collector is waiting, notify it, before 2406 // returning. 2407 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 2408 _foregroundGCShouldWait = false; 2409 if (_foregroundGCIsActive) { 2410 CGC_lock->notify(); 2411 } 2412 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 2413 "Possible deadlock"); 2414 } 2415 if (TraceCMSState) { 2416 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT 2417 " exiting collection CMS state %d", 2418 Thread::current(), _collectorState); 2419 } 2420 if (PrintGC && Verbose) { 2421 _cmsGen->print_heap_change(prev_used); 2422 } 2423 } 2424 2425 void CMSCollector::collect_in_foreground(bool clear_all_soft_refs) { 2426 assert(_foregroundGCIsActive && !_foregroundGCShouldWait, 2427 "Foreground collector should be waiting, not executing"); 2428 assert(Thread::current()->is_VM_thread(), "A foreground collection" 2429 "may only be done by the VM Thread with the world stopped"); 2430 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 2431 "VM thread should have CMS token"); 2432 2433 NOT_PRODUCT(TraceTime t("CMS:MS (foreground) ", PrintGCDetails && Verbose, 2434 true, gclog_or_tty);) 2435 if (UseAdaptiveSizePolicy) { 2436 size_policy()->ms_collection_begin(); 2437 } 2438 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact); 2439 2440 HandleMark hm; // Discard invalid handles created during verification 2441 2442 if (VerifyBeforeGC && 2443 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2444 Universe::verify(true); 2445 } 2446 2447 // Snapshot the soft reference policy to be used in this collection cycle. 2448 ref_processor()->setup_policy(clear_all_soft_refs); 2449 2450 bool init_mark_was_synchronous = false; // until proven otherwise 2451 while (_collectorState != Idling) { 2452 if (TraceCMSState) { 2453 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " in CMS state %d", 2454 Thread::current(), _collectorState); 2455 } 2456 switch (_collectorState) { 2457 case InitialMarking: 2458 init_mark_was_synchronous = true; // fact to be exploited in re-mark 2459 checkpointRootsInitial(false); 2460 assert(_collectorState == Marking, "Collector state should have changed" 2461 " within checkpointRootsInitial()"); 2462 break; 2463 case Marking: 2464 // initial marking in checkpointRootsInitialWork has been completed 2465 if (VerifyDuringGC && 2466 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2467 gclog_or_tty->print("Verify before initial mark: "); 2468 Universe::verify(true); 2469 } 2470 { 2471 bool res = markFromRoots(false); 2472 assert(res && _collectorState == FinalMarking, "Collector state should " 2473 "have changed"); 2474 break; 2475 } 2476 case FinalMarking: 2477 if (VerifyDuringGC && 2478 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2479 gclog_or_tty->print("Verify before re-mark: "); 2480 Universe::verify(true); 2481 } 2482 checkpointRootsFinal(false, clear_all_soft_refs, 2483 init_mark_was_synchronous); 2484 assert(_collectorState == Sweeping, "Collector state should not " 2485 "have changed within checkpointRootsFinal()"); 2486 break; 2487 case Sweeping: 2488 // final marking in checkpointRootsFinal has been completed 2489 if (VerifyDuringGC && 2490 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2491 gclog_or_tty->print("Verify before sweep: "); 2492 Universe::verify(true); 2493 } 2494 sweep(false); 2495 assert(_collectorState == Resizing, "Incorrect state"); 2496 break; 2497 case Resizing: { 2498 // Sweeping has been completed; the actual resize in this case 2499 // is done separately; nothing to be done in this state. 2500 _collectorState = Resetting; 2501 break; 2502 } 2503 case Resetting: 2504 // The heap has been resized. 2505 if (VerifyDuringGC && 2506 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2507 gclog_or_tty->print("Verify before reset: "); 2508 Universe::verify(true); 2509 } 2510 reset(false); 2511 assert(_collectorState == Idling, "Collector state should " 2512 "have changed"); 2513 break; 2514 case Precleaning: 2515 case AbortablePreclean: 2516 // Elide the preclean phase 2517 _collectorState = FinalMarking; 2518 break; 2519 default: 2520 ShouldNotReachHere(); 2521 } 2522 if (TraceCMSState) { 2523 gclog_or_tty->print_cr(" Thread " INTPTR_FORMAT " done - next CMS state %d", 2524 Thread::current(), _collectorState); 2525 } 2526 } 2527 2528 if (UseAdaptiveSizePolicy) { 2529 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2530 size_policy()->ms_collection_end(gch->gc_cause()); 2531 } 2532 2533 if (VerifyAfterGC && 2534 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2535 Universe::verify(true); 2536 } 2537 if (TraceCMSState) { 2538 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT 2539 " exiting collection CMS state %d", 2540 Thread::current(), _collectorState); 2541 } 2542 } 2543 2544 bool CMSCollector::waitForForegroundGC() { 2545 bool res = false; 2546 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 2547 "CMS thread should have CMS token"); 2548 // Block the foreground collector until the 2549 // background collectors decides whether to 2550 // yield. 2551 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 2552 _foregroundGCShouldWait = true; 2553 if (_foregroundGCIsActive) { 2554 // The background collector yields to the 2555 // foreground collector and returns a value 2556 // indicating that it has yielded. The foreground 2557 // collector can proceed. 2558 res = true; 2559 _foregroundGCShouldWait = false; 2560 ConcurrentMarkSweepThread::clear_CMS_flag( 2561 ConcurrentMarkSweepThread::CMS_cms_has_token); 2562 ConcurrentMarkSweepThread::set_CMS_flag( 2563 ConcurrentMarkSweepThread::CMS_cms_wants_token); 2564 // Get a possibly blocked foreground thread going 2565 CGC_lock->notify(); 2566 if (TraceCMSState) { 2567 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " waiting at CMS state %d", 2568 Thread::current(), _collectorState); 2569 } 2570 while (_foregroundGCIsActive) { 2571 CGC_lock->wait(Mutex::_no_safepoint_check_flag); 2572 } 2573 ConcurrentMarkSweepThread::set_CMS_flag( 2574 ConcurrentMarkSweepThread::CMS_cms_has_token); 2575 ConcurrentMarkSweepThread::clear_CMS_flag( 2576 ConcurrentMarkSweepThread::CMS_cms_wants_token); 2577 } 2578 if (TraceCMSState) { 2579 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " continuing at CMS state %d", 2580 Thread::current(), _collectorState); 2581 } 2582 return res; 2583 } 2584 2585 // Because of the need to lock the free lists and other structures in 2586 // the collector, common to all the generations that the collector is 2587 // collecting, we need the gc_prologues of individual CMS generations 2588 // delegate to their collector. It may have been simpler had the 2589 // current infrastructure allowed one to call a prologue on a 2590 // collector. In the absence of that we have the generation's 2591 // prologue delegate to the collector, which delegates back 2592 // some "local" work to a worker method in the individual generations 2593 // that it's responsible for collecting, while itself doing any 2594 // work common to all generations it's responsible for. A similar 2595 // comment applies to the gc_epilogue()'s. 2596 // The role of the varaible _between_prologue_and_epilogue is to 2597 // enforce the invocation protocol. 2598 void CMSCollector::gc_prologue(bool full) { 2599 // Call gc_prologue_work() for each CMSGen and PermGen that 2600 // we are responsible for. 2601 2602 // The following locking discipline assumes that we are only called 2603 // when the world is stopped. 2604 assert(SafepointSynchronize::is_at_safepoint(), "world is stopped assumption"); 2605 2606 // The CMSCollector prologue must call the gc_prologues for the 2607 // "generations" (including PermGen if any) that it's responsible 2608 // for. 2609 2610 assert( Thread::current()->is_VM_thread() 2611 || ( CMSScavengeBeforeRemark 2612 && Thread::current()->is_ConcurrentGC_thread()), 2613 "Incorrect thread type for prologue execution"); 2614 2615 if (_between_prologue_and_epilogue) { 2616 // We have already been invoked; this is a gc_prologue delegation 2617 // from yet another CMS generation that we are responsible for, just 2618 // ignore it since all relevant work has already been done. 2619 return; 2620 } 2621 2622 // set a bit saying prologue has been called; cleared in epilogue 2623 _between_prologue_and_epilogue = true; 2624 // Claim locks for common data structures, then call gc_prologue_work() 2625 // for each CMSGen and PermGen that we are responsible for. 2626 2627 getFreelistLocks(); // gets free list locks on constituent spaces 2628 bitMapLock()->lock_without_safepoint_check(); 2629 2630 // Should call gc_prologue_work() for all cms gens we are responsible for 2631 bool registerClosure = _collectorState >= Marking 2632 && _collectorState < Sweeping; 2633 ModUnionClosure* muc = CollectedHeap::use_parallel_gc_threads() ? 2634 &_modUnionClosurePar 2635 : &_modUnionClosure; 2636 _cmsGen->gc_prologue_work(full, registerClosure, muc); 2637 _permGen->gc_prologue_work(full, registerClosure, muc); 2638 2639 if (!full) { 2640 stats().record_gc0_begin(); 2641 } 2642 } 2643 2644 void ConcurrentMarkSweepGeneration::gc_prologue(bool full) { 2645 // Delegate to CMScollector which knows how to coordinate between 2646 // this and any other CMS generations that it is responsible for 2647 // collecting. 2648 collector()->gc_prologue(full); 2649 } 2650 2651 // This is a "private" interface for use by this generation's CMSCollector. 2652 // Not to be called directly by any other entity (for instance, 2653 // GenCollectedHeap, which calls the "public" gc_prologue method above). 2654 void ConcurrentMarkSweepGeneration::gc_prologue_work(bool full, 2655 bool registerClosure, ModUnionClosure* modUnionClosure) { 2656 assert(!incremental_collection_failed(), "Shouldn't be set yet"); 2657 assert(cmsSpace()->preconsumptionDirtyCardClosure() == NULL, 2658 "Should be NULL"); 2659 if (registerClosure) { 2660 cmsSpace()->setPreconsumptionDirtyCardClosure(modUnionClosure); 2661 } 2662 cmsSpace()->gc_prologue(); 2663 // Clear stat counters 2664 NOT_PRODUCT( 2665 assert(_numObjectsPromoted == 0, "check"); 2666 assert(_numWordsPromoted == 0, "check"); 2667 if (Verbose && PrintGC) { 2668 gclog_or_tty->print("Allocated "SIZE_FORMAT" objects, " 2669 SIZE_FORMAT" bytes concurrently", 2670 _numObjectsAllocated, _numWordsAllocated*sizeof(HeapWord)); 2671 } 2672 _numObjectsAllocated = 0; 2673 _numWordsAllocated = 0; 2674 ) 2675 } 2676 2677 void CMSCollector::gc_epilogue(bool full) { 2678 // The following locking discipline assumes that we are only called 2679 // when the world is stopped. 2680 assert(SafepointSynchronize::is_at_safepoint(), 2681 "world is stopped assumption"); 2682 2683 // Currently the CMS epilogue (see CompactibleFreeListSpace) merely checks 2684 // if linear allocation blocks need to be appropriately marked to allow the 2685 // the blocks to be parsable. We also check here whether we need to nudge the 2686 // CMS collector thread to start a new cycle (if it's not already active). 2687 assert( Thread::current()->is_VM_thread() 2688 || ( CMSScavengeBeforeRemark 2689 && Thread::current()->is_ConcurrentGC_thread()), 2690 "Incorrect thread type for epilogue execution"); 2691 2692 if (!_between_prologue_and_epilogue) { 2693 // We have already been invoked; this is a gc_epilogue delegation 2694 // from yet another CMS generation that we are responsible for, just 2695 // ignore it since all relevant work has already been done. 2696 return; 2697 } 2698 assert(haveFreelistLocks(), "must have freelist locks"); 2699 assert_lock_strong(bitMapLock()); 2700 2701 _cmsGen->gc_epilogue_work(full); 2702 _permGen->gc_epilogue_work(full); 2703 2704 if (_collectorState == AbortablePreclean || _collectorState == Precleaning) { 2705 // in case sampling was not already enabled, enable it 2706 _start_sampling = true; 2707 } 2708 // reset _eden_chunk_array so sampling starts afresh 2709 _eden_chunk_index = 0; 2710 2711 size_t cms_used = _cmsGen->cmsSpace()->used(); 2712 size_t perm_used = _permGen->cmsSpace()->used(); 2713 2714 // update performance counters - this uses a special version of 2715 // update_counters() that allows the utilization to be passed as a 2716 // parameter, avoiding multiple calls to used(). 2717 // 2718 _cmsGen->update_counters(cms_used); 2719 _permGen->update_counters(perm_used); 2720 2721 if (CMSIncrementalMode) { 2722 icms_update_allocation_limits(); 2723 } 2724 2725 bitMapLock()->unlock(); 2726 releaseFreelistLocks(); 2727 2728 _between_prologue_and_epilogue = false; // ready for next cycle 2729 } 2730 2731 void ConcurrentMarkSweepGeneration::gc_epilogue(bool full) { 2732 collector()->gc_epilogue(full); 2733 2734 // Also reset promotion tracking in par gc thread states. 2735 if (CollectedHeap::use_parallel_gc_threads()) { 2736 for (uint i = 0; i < ParallelGCThreads; i++) { 2737 _par_gc_thread_states[i]->promo.stopTrackingPromotions(i); 2738 } 2739 } 2740 } 2741 2742 void ConcurrentMarkSweepGeneration::gc_epilogue_work(bool full) { 2743 assert(!incremental_collection_failed(), "Should have been cleared"); 2744 cmsSpace()->setPreconsumptionDirtyCardClosure(NULL); 2745 cmsSpace()->gc_epilogue(); 2746 // Print stat counters 2747 NOT_PRODUCT( 2748 assert(_numObjectsAllocated == 0, "check"); 2749 assert(_numWordsAllocated == 0, "check"); 2750 if (Verbose && PrintGC) { 2751 gclog_or_tty->print("Promoted "SIZE_FORMAT" objects, " 2752 SIZE_FORMAT" bytes", 2753 _numObjectsPromoted, _numWordsPromoted*sizeof(HeapWord)); 2754 } 2755 _numObjectsPromoted = 0; 2756 _numWordsPromoted = 0; 2757 ) 2758 2759 if (PrintGC && Verbose) { 2760 // Call down the chain in contiguous_available needs the freelistLock 2761 // so print this out before releasing the freeListLock. 2762 gclog_or_tty->print(" Contiguous available "SIZE_FORMAT" bytes ", 2763 contiguous_available()); 2764 } 2765 } 2766 2767 #ifndef PRODUCT 2768 bool CMSCollector::have_cms_token() { 2769 Thread* thr = Thread::current(); 2770 if (thr->is_VM_thread()) { 2771 return ConcurrentMarkSweepThread::vm_thread_has_cms_token(); 2772 } else if (thr->is_ConcurrentGC_thread()) { 2773 return ConcurrentMarkSweepThread::cms_thread_has_cms_token(); 2774 } else if (thr->is_GC_task_thread()) { 2775 return ConcurrentMarkSweepThread::vm_thread_has_cms_token() && 2776 ParGCRareEvent_lock->owned_by_self(); 2777 } 2778 return false; 2779 } 2780 #endif 2781 2782 // Check reachability of the given heap address in CMS generation, 2783 // treating all other generations as roots. 2784 bool CMSCollector::is_cms_reachable(HeapWord* addr) { 2785 // We could "guarantee" below, rather than assert, but i'll 2786 // leave these as "asserts" so that an adventurous debugger 2787 // could try this in the product build provided some subset of 2788 // the conditions were met, provided they were intersted in the 2789 // results and knew that the computation below wouldn't interfere 2790 // with other concurrent computations mutating the structures 2791 // being read or written. 2792 assert(SafepointSynchronize::is_at_safepoint(), 2793 "Else mutations in object graph will make answer suspect"); 2794 assert(have_cms_token(), "Should hold cms token"); 2795 assert(haveFreelistLocks(), "must hold free list locks"); 2796 assert_lock_strong(bitMapLock()); 2797 2798 // Clear the marking bit map array before starting, but, just 2799 // for kicks, first report if the given address is already marked 2800 gclog_or_tty->print_cr("Start: Address 0x%x is%s marked", addr, 2801 _markBitMap.isMarked(addr) ? "" : " not"); 2802 2803 if (verify_after_remark()) { 2804 MutexLockerEx x(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag); 2805 bool result = verification_mark_bm()->isMarked(addr); 2806 gclog_or_tty->print_cr("TransitiveMark: Address 0x%x %s marked", addr, 2807 result ? "IS" : "is NOT"); 2808 return result; 2809 } else { 2810 gclog_or_tty->print_cr("Could not compute result"); 2811 return false; 2812 } 2813 } 2814 2815 //////////////////////////////////////////////////////// 2816 // CMS Verification Support 2817 //////////////////////////////////////////////////////// 2818 // Following the remark phase, the following invariant 2819 // should hold -- each object in the CMS heap which is 2820 // marked in markBitMap() should be marked in the verification_mark_bm(). 2821 2822 class VerifyMarkedClosure: public BitMapClosure { 2823 CMSBitMap* _marks; 2824 bool _failed; 2825 2826 public: 2827 VerifyMarkedClosure(CMSBitMap* bm): _marks(bm), _failed(false) {} 2828 2829 bool do_bit(size_t offset) { 2830 HeapWord* addr = _marks->offsetToHeapWord(offset); 2831 if (!_marks->isMarked(addr)) { 2832 oop(addr)->print_on(gclog_or_tty); 2833 gclog_or_tty->print_cr(" ("INTPTR_FORMAT" should have been marked)", addr); 2834 _failed = true; 2835 } 2836 return true; 2837 } 2838 2839 bool failed() { return _failed; } 2840 }; 2841 2842 bool CMSCollector::verify_after_remark() { 2843 gclog_or_tty->print(" [Verifying CMS Marking... "); 2844 MutexLockerEx ml(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag); 2845 static bool init = false; 2846 2847 assert(SafepointSynchronize::is_at_safepoint(), 2848 "Else mutations in object graph will make answer suspect"); 2849 assert(have_cms_token(), 2850 "Else there may be mutual interference in use of " 2851 " verification data structures"); 2852 assert(_collectorState > Marking && _collectorState <= Sweeping, 2853 "Else marking info checked here may be obsolete"); 2854 assert(haveFreelistLocks(), "must hold free list locks"); 2855 assert_lock_strong(bitMapLock()); 2856 2857 2858 // Allocate marking bit map if not already allocated 2859 if (!init) { // first time 2860 if (!verification_mark_bm()->allocate(_span)) { 2861 return false; 2862 } 2863 init = true; 2864 } 2865 2866 assert(verification_mark_stack()->isEmpty(), "Should be empty"); 2867 2868 // Turn off refs discovery -- so we will be tracing through refs. 2869 // This is as intended, because by this time 2870 // GC must already have cleared any refs that need to be cleared, 2871 // and traced those that need to be marked; moreover, 2872 // the marking done here is not going to intefere in any 2873 // way with the marking information used by GC. 2874 NoRefDiscovery no_discovery(ref_processor()); 2875 2876 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;) 2877 2878 // Clear any marks from a previous round 2879 verification_mark_bm()->clear_all(); 2880 assert(verification_mark_stack()->isEmpty(), "markStack should be empty"); 2881 verify_work_stacks_empty(); 2882 2883 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2884 gch->ensure_parsability(false); // fill TLABs, but no need to retire them 2885 // Update the saved marks which may affect the root scans. 2886 gch->save_marks(); 2887 2888 if (CMSRemarkVerifyVariant == 1) { 2889 // In this first variant of verification, we complete 2890 // all marking, then check if the new marks-verctor is 2891 // a subset of the CMS marks-vector. 2892 verify_after_remark_work_1(); 2893 } else if (CMSRemarkVerifyVariant == 2) { 2894 // In this second variant of verification, we flag an error 2895 // (i.e. an object reachable in the new marks-vector not reachable 2896 // in the CMS marks-vector) immediately, also indicating the 2897 // identify of an object (A) that references the unmarked object (B) -- 2898 // presumably, a mutation to A failed to be picked up by preclean/remark? 2899 verify_after_remark_work_2(); 2900 } else { 2901 warning("Unrecognized value %d for CMSRemarkVerifyVariant", 2902 CMSRemarkVerifyVariant); 2903 } 2904 gclog_or_tty->print(" done] "); 2905 return true; 2906 } 2907 2908 void CMSCollector::verify_after_remark_work_1() { 2909 ResourceMark rm; 2910 HandleMark hm; 2911 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2912 2913 // Mark from roots one level into CMS 2914 MarkRefsIntoClosure notOlder(_span, verification_mark_bm()); 2915 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2916 2917 gch->gen_process_strong_roots(_cmsGen->level(), 2918 true, // younger gens are roots 2919 true, // activate StrongRootsScope 2920 true, // collecting perm gen 2921 SharedHeap::ScanningOption(roots_scanning_options()), 2922 ¬Older, 2923 true, // walk code active on stacks 2924 NULL); 2925 2926 // Now mark from the roots 2927 assert(_revisitStack.isEmpty(), "Should be empty"); 2928 MarkFromRootsClosure markFromRootsClosure(this, _span, 2929 verification_mark_bm(), verification_mark_stack(), &_revisitStack, 2930 false /* don't yield */, true /* verifying */); 2931 assert(_restart_addr == NULL, "Expected pre-condition"); 2932 verification_mark_bm()->iterate(&markFromRootsClosure); 2933 while (_restart_addr != NULL) { 2934 // Deal with stack overflow: by restarting at the indicated 2935 // address. 2936 HeapWord* ra = _restart_addr; 2937 markFromRootsClosure.reset(ra); 2938 _restart_addr = NULL; 2939 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end()); 2940 } 2941 assert(verification_mark_stack()->isEmpty(), "Should have been drained"); 2942 verify_work_stacks_empty(); 2943 // Should reset the revisit stack above, since no class tree 2944 // surgery is forthcoming. 2945 _revisitStack.reset(); // throwing away all contents 2946 2947 // Marking completed -- now verify that each bit marked in 2948 // verification_mark_bm() is also marked in markBitMap(); flag all 2949 // errors by printing corresponding objects. 2950 VerifyMarkedClosure vcl(markBitMap()); 2951 verification_mark_bm()->iterate(&vcl); 2952 if (vcl.failed()) { 2953 gclog_or_tty->print("Verification failed"); 2954 Universe::heap()->print_on(gclog_or_tty); 2955 fatal("CMS: failed marking verification after remark"); 2956 } 2957 } 2958 2959 void CMSCollector::verify_after_remark_work_2() { 2960 ResourceMark rm; 2961 HandleMark hm; 2962 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2963 2964 // Mark from roots one level into CMS 2965 MarkRefsIntoVerifyClosure notOlder(_span, verification_mark_bm(), 2966 markBitMap()); 2967 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2968 gch->gen_process_strong_roots(_cmsGen->level(), 2969 true, // younger gens are roots 2970 true, // activate StrongRootsScope 2971 true, // collecting perm gen 2972 SharedHeap::ScanningOption(roots_scanning_options()), 2973 ¬Older, 2974 true, // walk code active on stacks 2975 NULL); 2976 2977 // Now mark from the roots 2978 assert(_revisitStack.isEmpty(), "Should be empty"); 2979 MarkFromRootsVerifyClosure markFromRootsClosure(this, _span, 2980 verification_mark_bm(), markBitMap(), verification_mark_stack()); 2981 assert(_restart_addr == NULL, "Expected pre-condition"); 2982 verification_mark_bm()->iterate(&markFromRootsClosure); 2983 while (_restart_addr != NULL) { 2984 // Deal with stack overflow: by restarting at the indicated 2985 // address. 2986 HeapWord* ra = _restart_addr; 2987 markFromRootsClosure.reset(ra); 2988 _restart_addr = NULL; 2989 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end()); 2990 } 2991 assert(verification_mark_stack()->isEmpty(), "Should have been drained"); 2992 verify_work_stacks_empty(); 2993 // Should reset the revisit stack above, since no class tree 2994 // surgery is forthcoming. 2995 _revisitStack.reset(); // throwing away all contents 2996 2997 // Marking completed -- now verify that each bit marked in 2998 // verification_mark_bm() is also marked in markBitMap(); flag all 2999 // errors by printing corresponding objects. 3000 VerifyMarkedClosure vcl(markBitMap()); 3001 verification_mark_bm()->iterate(&vcl); 3002 assert(!vcl.failed(), "Else verification above should not have succeeded"); 3003 } 3004 3005 void ConcurrentMarkSweepGeneration::save_marks() { 3006 // delegate to CMS space 3007 cmsSpace()->save_marks(); 3008 for (uint i = 0; i < ParallelGCThreads; i++) { 3009 _par_gc_thread_states[i]->promo.startTrackingPromotions(); 3010 } 3011 } 3012 3013 bool ConcurrentMarkSweepGeneration::no_allocs_since_save_marks() { 3014 return cmsSpace()->no_allocs_since_save_marks(); 3015 } 3016 3017 #define CMS_SINCE_SAVE_MARKS_DEFN(OopClosureType, nv_suffix) \ 3018 \ 3019 void ConcurrentMarkSweepGeneration:: \ 3020 oop_since_save_marks_iterate##nv_suffix(OopClosureType* cl) { \ 3021 cl->set_generation(this); \ 3022 cmsSpace()->oop_since_save_marks_iterate##nv_suffix(cl); \ 3023 cl->reset_generation(); \ 3024 save_marks(); \ 3025 } 3026 3027 ALL_SINCE_SAVE_MARKS_CLOSURES(CMS_SINCE_SAVE_MARKS_DEFN) 3028 3029 void 3030 ConcurrentMarkSweepGeneration::object_iterate_since_last_GC(ObjectClosure* blk) 3031 { 3032 // Not currently implemented; need to do the following. -- ysr. 3033 // dld -- I think that is used for some sort of allocation profiler. So it 3034 // really means the objects allocated by the mutator since the last 3035 // GC. We could potentially implement this cheaply by recording only 3036 // the direct allocations in a side data structure. 3037 // 3038 // I think we probably ought not to be required to support these 3039 // iterations at any arbitrary point; I think there ought to be some 3040 // call to enable/disable allocation profiling in a generation/space, 3041 // and the iterator ought to return the objects allocated in the 3042 // gen/space since the enable call, or the last iterator call (which 3043 // will probably be at a GC.) That way, for gens like CM&S that would 3044 // require some extra data structure to support this, we only pay the 3045 // cost when it's in use... 3046 cmsSpace()->object_iterate_since_last_GC(blk); 3047 } 3048 3049 void 3050 ConcurrentMarkSweepGeneration::younger_refs_iterate(OopsInGenClosure* cl) { 3051 cl->set_generation(this); 3052 younger_refs_in_space_iterate(_cmsSpace, cl); 3053 cl->reset_generation(); 3054 } 3055 3056 void 3057 ConcurrentMarkSweepGeneration::oop_iterate(MemRegion mr, OopClosure* cl) { 3058 if (freelistLock()->owned_by_self()) { 3059 Generation::oop_iterate(mr, cl); 3060 } else { 3061 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 3062 Generation::oop_iterate(mr, cl); 3063 } 3064 } 3065 3066 void 3067 ConcurrentMarkSweepGeneration::oop_iterate(OopClosure* cl) { 3068 if (freelistLock()->owned_by_self()) { 3069 Generation::oop_iterate(cl); 3070 } else { 3071 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 3072 Generation::oop_iterate(cl); 3073 } 3074 } 3075 3076 void 3077 ConcurrentMarkSweepGeneration::object_iterate(ObjectClosure* cl) { 3078 if (freelistLock()->owned_by_self()) { 3079 Generation::object_iterate(cl); 3080 } else { 3081 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 3082 Generation::object_iterate(cl); 3083 } 3084 } 3085 3086 void 3087 ConcurrentMarkSweepGeneration::safe_object_iterate(ObjectClosure* cl) { 3088 if (freelistLock()->owned_by_self()) { 3089 Generation::safe_object_iterate(cl); 3090 } else { 3091 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 3092 Generation::safe_object_iterate(cl); 3093 } 3094 } 3095 3096 void 3097 ConcurrentMarkSweepGeneration::pre_adjust_pointers() { 3098 } 3099 3100 void 3101 ConcurrentMarkSweepGeneration::post_compact() { 3102 } 3103 3104 void 3105 ConcurrentMarkSweepGeneration::prepare_for_verify() { 3106 // Fix the linear allocation blocks to look like free blocks. 3107 3108 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those 3109 // are not called when the heap is verified during universe initialization and 3110 // at vm shutdown. 3111 if (freelistLock()->owned_by_self()) { 3112 cmsSpace()->prepare_for_verify(); 3113 } else { 3114 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag); 3115 cmsSpace()->prepare_for_verify(); 3116 } 3117 } 3118 3119 void 3120 ConcurrentMarkSweepGeneration::verify(bool allow_dirty /* ignored */) { 3121 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those 3122 // are not called when the heap is verified during universe initialization and 3123 // at vm shutdown. 3124 if (freelistLock()->owned_by_self()) { 3125 cmsSpace()->verify(false /* ignored */); 3126 } else { 3127 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag); 3128 cmsSpace()->verify(false /* ignored */); 3129 } 3130 } 3131 3132 void CMSCollector::verify(bool allow_dirty /* ignored */) { 3133 _cmsGen->verify(allow_dirty); 3134 _permGen->verify(allow_dirty); 3135 } 3136 3137 #ifndef PRODUCT 3138 bool CMSCollector::overflow_list_is_empty() const { 3139 assert(_num_par_pushes >= 0, "Inconsistency"); 3140 if (_overflow_list == NULL) { 3141 assert(_num_par_pushes == 0, "Inconsistency"); 3142 } 3143 return _overflow_list == NULL; 3144 } 3145 3146 // The methods verify_work_stacks_empty() and verify_overflow_empty() 3147 // merely consolidate assertion checks that appear to occur together frequently. 3148 void CMSCollector::verify_work_stacks_empty() const { 3149 assert(_markStack.isEmpty(), "Marking stack should be empty"); 3150 assert(overflow_list_is_empty(), "Overflow list should be empty"); 3151 } 3152 3153 void CMSCollector::verify_overflow_empty() const { 3154 assert(overflow_list_is_empty(), "Overflow list should be empty"); 3155 assert(no_preserved_marks(), "No preserved marks"); 3156 } 3157 #endif // PRODUCT 3158 3159 // Decide if we want to enable class unloading as part of the 3160 // ensuing concurrent GC cycle. We will collect the perm gen and 3161 // unload classes if it's the case that: 3162 // (1) an explicit gc request has been made and the flag 3163 // ExplicitGCInvokesConcurrentAndUnloadsClasses is set, OR 3164 // (2) (a) class unloading is enabled at the command line, and 3165 // (b) (i) perm gen threshold has been crossed, or 3166 // (ii) old gen is getting really full, or 3167 // (iii) the previous N CMS collections did not collect the 3168 // perm gen 3169 // NOTE: Provided there is no change in the state of the heap between 3170 // calls to this method, it should have idempotent results. Moreover, 3171 // its results should be monotonically increasing (i.e. going from 0 to 1, 3172 // but not 1 to 0) between successive calls between which the heap was 3173 // not collected. For the implementation below, it must thus rely on 3174 // the property that concurrent_cycles_since_last_unload() 3175 // will not decrease unless a collection cycle happened and that 3176 // _permGen->should_concurrent_collect() and _cmsGen->is_too_full() are 3177 // themselves also monotonic in that sense. See check_monotonicity() 3178 // below. 3179 bool CMSCollector::update_should_unload_classes() { 3180 _should_unload_classes = false; 3181 // Condition 1 above 3182 if (_full_gc_requested && ExplicitGCInvokesConcurrentAndUnloadsClasses) { 3183 _should_unload_classes = true; 3184 } else if (CMSClassUnloadingEnabled) { // Condition 2.a above 3185 // Disjuncts 2.b.(i,ii,iii) above 3186 _should_unload_classes = (concurrent_cycles_since_last_unload() >= 3187 CMSClassUnloadingMaxInterval) 3188 || _permGen->should_concurrent_collect() 3189 || _cmsGen->is_too_full(); 3190 } 3191 return _should_unload_classes; 3192 } 3193 3194 bool ConcurrentMarkSweepGeneration::is_too_full() const { 3195 bool res = should_concurrent_collect(); 3196 res = res && (occupancy() > (double)CMSIsTooFullPercentage/100.0); 3197 return res; 3198 } 3199 3200 void CMSCollector::setup_cms_unloading_and_verification_state() { 3201 const bool should_verify = VerifyBeforeGC || VerifyAfterGC || VerifyDuringGC 3202 || VerifyBeforeExit; 3203 const int rso = SharedHeap::SO_Symbols | SharedHeap::SO_Strings 3204 | SharedHeap::SO_CodeCache; 3205 3206 if (should_unload_classes()) { // Should unload classes this cycle 3207 remove_root_scanning_option(rso); // Shrink the root set appropriately 3208 set_verifying(should_verify); // Set verification state for this cycle 3209 return; // Nothing else needs to be done at this time 3210 } 3211 3212 // Not unloading classes this cycle 3213 assert(!should_unload_classes(), "Inconsitency!"); 3214 if ((!verifying() || unloaded_classes_last_cycle()) && should_verify) { 3215 // We were not verifying, or we _were_ unloading classes in the last cycle, 3216 // AND some verification options are enabled this cycle; in this case, 3217 // we must make sure that the deadness map is allocated if not already so, 3218 // and cleared (if already allocated previously -- 3219 // CMSBitMap::sizeInBits() is used to determine if it's allocated). 3220 if (perm_gen_verify_bit_map()->sizeInBits() == 0) { 3221 if (!perm_gen_verify_bit_map()->allocate(_permGen->reserved())) { 3222 warning("Failed to allocate permanent generation verification CMS Bit Map;\n" 3223 "permanent generation verification disabled"); 3224 return; // Note that we leave verification disabled, so we'll retry this 3225 // allocation next cycle. We _could_ remember this failure 3226 // and skip further attempts and permanently disable verification 3227 // attempts if that is considered more desirable. 3228 } 3229 assert(perm_gen_verify_bit_map()->covers(_permGen->reserved()), 3230 "_perm_gen_ver_bit_map inconsistency?"); 3231 } else { 3232 perm_gen_verify_bit_map()->clear_all(); 3233 } 3234 // Include symbols, strings and code cache elements to prevent their resurrection. 3235 add_root_scanning_option(rso); 3236 set_verifying(true); 3237 } else if (verifying() && !should_verify) { 3238 // We were verifying, but some verification flags got disabled. 3239 set_verifying(false); 3240 // Exclude symbols, strings and code cache elements from root scanning to 3241 // reduce IM and RM pauses. 3242 remove_root_scanning_option(rso); 3243 } 3244 } 3245 3246 3247 #ifndef PRODUCT 3248 HeapWord* CMSCollector::block_start(const void* p) const { 3249 const HeapWord* addr = (HeapWord*)p; 3250 if (_span.contains(p)) { 3251 if (_cmsGen->cmsSpace()->is_in_reserved(addr)) { 3252 return _cmsGen->cmsSpace()->block_start(p); 3253 } else { 3254 assert(_permGen->cmsSpace()->is_in_reserved(addr), 3255 "Inconsistent _span?"); 3256 return _permGen->cmsSpace()->block_start(p); 3257 } 3258 } 3259 return NULL; 3260 } 3261 #endif 3262 3263 HeapWord* 3264 ConcurrentMarkSweepGeneration::expand_and_allocate(size_t word_size, 3265 bool tlab, 3266 bool parallel) { 3267 CMSSynchronousYieldRequest yr; 3268 assert(!tlab, "Can't deal with TLAB allocation"); 3269 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 3270 expand(word_size*HeapWordSize, MinHeapDeltaBytes, 3271 CMSExpansionCause::_satisfy_allocation); 3272 if (GCExpandToAllocateDelayMillis > 0) { 3273 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 3274 } 3275 return have_lock_and_allocate(word_size, tlab); 3276 } 3277 3278 // YSR: All of this generation expansion/shrinking stuff is an exact copy of 3279 // OneContigSpaceCardGeneration, which makes me wonder if we should move this 3280 // to CardGeneration and share it... 3281 bool ConcurrentMarkSweepGeneration::expand(size_t bytes, size_t expand_bytes) { 3282 return CardGeneration::expand(bytes, expand_bytes); 3283 } 3284 3285 void ConcurrentMarkSweepGeneration::expand(size_t bytes, size_t expand_bytes, 3286 CMSExpansionCause::Cause cause) 3287 { 3288 3289 bool success = expand(bytes, expand_bytes); 3290 3291 // remember why we expanded; this information is used 3292 // by shouldConcurrentCollect() when making decisions on whether to start 3293 // a new CMS cycle. 3294 if (success) { 3295 set_expansion_cause(cause); 3296 if (PrintGCDetails && Verbose) { 3297 gclog_or_tty->print_cr("Expanded CMS gen for %s", 3298 CMSExpansionCause::to_string(cause)); 3299 } 3300 } 3301 } 3302 3303 HeapWord* ConcurrentMarkSweepGeneration::expand_and_par_lab_allocate(CMSParGCThreadState* ps, size_t word_sz) { 3304 HeapWord* res = NULL; 3305 MutexLocker x(ParGCRareEvent_lock); 3306 while (true) { 3307 // Expansion by some other thread might make alloc OK now: 3308 res = ps->lab.alloc(word_sz); 3309 if (res != NULL) return res; 3310 // If there's not enough expansion space available, give up. 3311 if (_virtual_space.uncommitted_size() < (word_sz * HeapWordSize)) { 3312 return NULL; 3313 } 3314 // Otherwise, we try expansion. 3315 expand(word_sz*HeapWordSize, MinHeapDeltaBytes, 3316 CMSExpansionCause::_allocate_par_lab); 3317 // Now go around the loop and try alloc again; 3318 // A competing par_promote might beat us to the expansion space, 3319 // so we may go around the loop again if promotion fails agaion. 3320 if (GCExpandToAllocateDelayMillis > 0) { 3321 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 3322 } 3323 } 3324 } 3325 3326 3327 bool ConcurrentMarkSweepGeneration::expand_and_ensure_spooling_space( 3328 PromotionInfo* promo) { 3329 MutexLocker x(ParGCRareEvent_lock); 3330 size_t refill_size_bytes = promo->refillSize() * HeapWordSize; 3331 while (true) { 3332 // Expansion by some other thread might make alloc OK now: 3333 if (promo->ensure_spooling_space()) { 3334 assert(promo->has_spooling_space(), 3335 "Post-condition of successful ensure_spooling_space()"); 3336 return true; 3337 } 3338 // If there's not enough expansion space available, give up. 3339 if (_virtual_space.uncommitted_size() < refill_size_bytes) { 3340 return false; 3341 } 3342 // Otherwise, we try expansion. 3343 expand(refill_size_bytes, MinHeapDeltaBytes, 3344 CMSExpansionCause::_allocate_par_spooling_space); 3345 // Now go around the loop and try alloc again; 3346 // A competing allocation might beat us to the expansion space, 3347 // so we may go around the loop again if allocation fails again. 3348 if (GCExpandToAllocateDelayMillis > 0) { 3349 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 3350 } 3351 } 3352 } 3353 3354 3355 3356 void ConcurrentMarkSweepGeneration::shrink(size_t bytes) { 3357 assert_locked_or_safepoint(Heap_lock); 3358 size_t size = ReservedSpace::page_align_size_down(bytes); 3359 if (size > 0) { 3360 shrink_by(size); 3361 } 3362 } 3363 3364 bool ConcurrentMarkSweepGeneration::grow_by(size_t bytes) { 3365 assert_locked_or_safepoint(Heap_lock); 3366 bool result = _virtual_space.expand_by(bytes); 3367 if (result) { 3368 HeapWord* old_end = _cmsSpace->end(); 3369 size_t new_word_size = 3370 heap_word_size(_virtual_space.committed_size()); 3371 MemRegion mr(_cmsSpace->bottom(), new_word_size); 3372 _bts->resize(new_word_size); // resize the block offset shared array 3373 Universe::heap()->barrier_set()->resize_covered_region(mr); 3374 // Hmmmm... why doesn't CFLS::set_end verify locking? 3375 // This is quite ugly; FIX ME XXX 3376 _cmsSpace->assert_locked(freelistLock()); 3377 _cmsSpace->set_end((HeapWord*)_virtual_space.high()); 3378 3379 // update the space and generation capacity counters 3380 if (UsePerfData) { 3381 _space_counters->update_capacity(); 3382 _gen_counters->update_all(); 3383 } 3384 3385 if (Verbose && PrintGC) { 3386 size_t new_mem_size = _virtual_space.committed_size(); 3387 size_t old_mem_size = new_mem_size - bytes; 3388 gclog_or_tty->print_cr("Expanding %s from %ldK by %ldK to %ldK", 3389 name(), old_mem_size/K, bytes/K, new_mem_size/K); 3390 } 3391 } 3392 return result; 3393 } 3394 3395 bool ConcurrentMarkSweepGeneration::grow_to_reserved() { 3396 assert_locked_or_safepoint(Heap_lock); 3397 bool success = true; 3398 const size_t remaining_bytes = _virtual_space.uncommitted_size(); 3399 if (remaining_bytes > 0) { 3400 success = grow_by(remaining_bytes); 3401 DEBUG_ONLY(if (!success) warning("grow to reserved failed");) 3402 } 3403 return success; 3404 } 3405 3406 void ConcurrentMarkSweepGeneration::shrink_by(size_t bytes) { 3407 assert_locked_or_safepoint(Heap_lock); 3408 assert_lock_strong(freelistLock()); 3409 // XXX Fix when compaction is implemented. 3410 warning("Shrinking of CMS not yet implemented"); 3411 return; 3412 } 3413 3414 3415 // Simple ctor/dtor wrapper for accounting & timer chores around concurrent 3416 // phases. 3417 class CMSPhaseAccounting: public StackObj { 3418 public: 3419 CMSPhaseAccounting(CMSCollector *collector, 3420 const char *phase, 3421 bool print_cr = true); 3422 ~CMSPhaseAccounting(); 3423 3424 private: 3425 CMSCollector *_collector; 3426 const char *_phase; 3427 elapsedTimer _wallclock; 3428 bool _print_cr; 3429 3430 public: 3431 // Not MT-safe; so do not pass around these StackObj's 3432 // where they may be accessed by other threads. 3433 jlong wallclock_millis() { 3434 assert(_wallclock.is_active(), "Wall clock should not stop"); 3435 _wallclock.stop(); // to record time 3436 jlong ret = _wallclock.milliseconds(); 3437 _wallclock.start(); // restart 3438 return ret; 3439 } 3440 }; 3441 3442 CMSPhaseAccounting::CMSPhaseAccounting(CMSCollector *collector, 3443 const char *phase, 3444 bool print_cr) : 3445 _collector(collector), _phase(phase), _print_cr(print_cr) { 3446 3447 if (PrintCMSStatistics != 0) { 3448 _collector->resetYields(); 3449 } 3450 if (PrintGCDetails && PrintGCTimeStamps) { 3451 gclog_or_tty->date_stamp(PrintGCDateStamps); 3452 gclog_or_tty->stamp(); 3453 gclog_or_tty->print_cr(": [%s-concurrent-%s-start]", 3454 _collector->cmsGen()->short_name(), _phase); 3455 } 3456 _collector->resetTimer(); 3457 _wallclock.start(); 3458 _collector->startTimer(); 3459 } 3460 3461 CMSPhaseAccounting::~CMSPhaseAccounting() { 3462 assert(_wallclock.is_active(), "Wall clock should not have stopped"); 3463 _collector->stopTimer(); 3464 _wallclock.stop(); 3465 if (PrintGCDetails) { 3466 gclog_or_tty->date_stamp(PrintGCDateStamps); 3467 if (PrintGCTimeStamps) { 3468 gclog_or_tty->stamp(); 3469 gclog_or_tty->print(": "); 3470 } 3471 gclog_or_tty->print("[%s-concurrent-%s: %3.3f/%3.3f secs]", 3472 _collector->cmsGen()->short_name(), 3473 _phase, _collector->timerValue(), _wallclock.seconds()); 3474 if (_print_cr) { 3475 gclog_or_tty->print_cr(""); 3476 } 3477 if (PrintCMSStatistics != 0) { 3478 gclog_or_tty->print_cr(" (CMS-concurrent-%s yielded %d times)", _phase, 3479 _collector->yields()); 3480 } 3481 } 3482 } 3483 3484 // CMS work 3485 3486 // Checkpoint the roots into this generation from outside 3487 // this generation. [Note this initial checkpoint need only 3488 // be approximate -- we'll do a catch up phase subsequently.] 3489 void CMSCollector::checkpointRootsInitial(bool asynch) { 3490 assert(_collectorState == InitialMarking, "Wrong collector state"); 3491 check_correct_thread_executing(); 3492 TraceCMSMemoryManagerStats tms(_collectorState); 3493 ReferenceProcessor* rp = ref_processor(); 3494 SpecializationStats::clear(); 3495 assert(_restart_addr == NULL, "Control point invariant"); 3496 if (asynch) { 3497 // acquire locks for subsequent manipulations 3498 MutexLockerEx x(bitMapLock(), 3499 Mutex::_no_safepoint_check_flag); 3500 checkpointRootsInitialWork(asynch); 3501 rp->verify_no_references_recorded(); 3502 rp->enable_discovery(); // enable ("weak") refs discovery 3503 _collectorState = Marking; 3504 } else { 3505 // (Weak) Refs discovery: this is controlled from genCollectedHeap::do_collection 3506 // which recognizes if we are a CMS generation, and doesn't try to turn on 3507 // discovery; verify that they aren't meddling. 3508 assert(!rp->discovery_is_atomic(), 3509 "incorrect setting of discovery predicate"); 3510 assert(!rp->discovery_enabled(), "genCollectedHeap shouldn't control " 3511 "ref discovery for this generation kind"); 3512 // already have locks 3513 checkpointRootsInitialWork(asynch); 3514 rp->enable_discovery(); // now enable ("weak") refs discovery 3515 _collectorState = Marking; 3516 } 3517 SpecializationStats::print(); 3518 } 3519 3520 void CMSCollector::checkpointRootsInitialWork(bool asynch) { 3521 assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped"); 3522 assert(_collectorState == InitialMarking, "just checking"); 3523 3524 // If there has not been a GC[n-1] since last GC[n] cycle completed, 3525 // precede our marking with a collection of all 3526 // younger generations to keep floating garbage to a minimum. 3527 // XXX: we won't do this for now -- it's an optimization to be done later. 3528 3529 // already have locks 3530 assert_lock_strong(bitMapLock()); 3531 assert(_markBitMap.isAllClear(), "was reset at end of previous cycle"); 3532 3533 // Setup the verification and class unloading state for this 3534 // CMS collection cycle. 3535 setup_cms_unloading_and_verification_state(); 3536 3537 NOT_PRODUCT(TraceTime t("\ncheckpointRootsInitialWork", 3538 PrintGCDetails && Verbose, true, gclog_or_tty);) 3539 if (UseAdaptiveSizePolicy) { 3540 size_policy()->checkpoint_roots_initial_begin(); 3541 } 3542 3543 // Reset all the PLAB chunk arrays if necessary. 3544 if (_survivor_plab_array != NULL && !CMSPLABRecordAlways) { 3545 reset_survivor_plab_arrays(); 3546 } 3547 3548 ResourceMark rm; 3549 HandleMark hm; 3550 3551 FalseClosure falseClosure; 3552 // In the case of a synchronous collection, we will elide the 3553 // remark step, so it's important to catch all the nmethod oops 3554 // in this step. 3555 // The final 'true' flag to gen_process_strong_roots will ensure this. 3556 // If 'async' is true, we can relax the nmethod tracing. 3557 MarkRefsIntoClosure notOlder(_span, &_markBitMap); 3558 GenCollectedHeap* gch = GenCollectedHeap::heap(); 3559 3560 verify_work_stacks_empty(); 3561 verify_overflow_empty(); 3562 3563 gch->ensure_parsability(false); // fill TLABs, but no need to retire them 3564 // Update the saved marks which may affect the root scans. 3565 gch->save_marks(); 3566 3567 // weak reference processing has not started yet. 3568 ref_processor()->set_enqueuing_is_done(false); 3569 3570 { 3571 // This is not needed. DEBUG_ONLY(RememberKlassesChecker imx(true);) 3572 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;) 3573 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 3574 gch->gen_process_strong_roots(_cmsGen->level(), 3575 true, // younger gens are roots 3576 true, // activate StrongRootsScope 3577 true, // collecting perm gen 3578 SharedHeap::ScanningOption(roots_scanning_options()), 3579 ¬Older, 3580 true, // walk all of code cache if (so & SO_CodeCache) 3581 NULL); 3582 } 3583 3584 // Clear mod-union table; it will be dirtied in the prologue of 3585 // CMS generation per each younger generation collection. 3586 3587 assert(_modUnionTable.isAllClear(), 3588 "Was cleared in most recent final checkpoint phase" 3589 " or no bits are set in the gc_prologue before the start of the next " 3590 "subsequent marking phase."); 3591 3592 // Temporarily disabled, since pre/post-consumption closures don't 3593 // care about precleaned cards 3594 #if 0 3595 { 3596 MemRegion mr = MemRegion((HeapWord*)_virtual_space.low(), 3597 (HeapWord*)_virtual_space.high()); 3598 _ct->ct_bs()->preclean_dirty_cards(mr); 3599 } 3600 #endif 3601 3602 // Save the end of the used_region of the constituent generations 3603 // to be used to limit the extent of sweep in each generation. 3604 save_sweep_limits(); 3605 if (UseAdaptiveSizePolicy) { 3606 size_policy()->checkpoint_roots_initial_end(gch->gc_cause()); 3607 } 3608 verify_overflow_empty(); 3609 } 3610 3611 bool CMSCollector::markFromRoots(bool asynch) { 3612 // we might be tempted to assert that: 3613 // assert(asynch == !SafepointSynchronize::is_at_safepoint(), 3614 // "inconsistent argument?"); 3615 // However that wouldn't be right, because it's possible that 3616 // a safepoint is indeed in progress as a younger generation 3617 // stop-the-world GC happens even as we mark in this generation. 3618 assert(_collectorState == Marking, "inconsistent state?"); 3619 check_correct_thread_executing(); 3620 verify_overflow_empty(); 3621 3622 bool res; 3623 if (asynch) { 3624 3625 // Start the timers for adaptive size policy for the concurrent phases 3626 // Do it here so that the foreground MS can use the concurrent 3627 // timer since a foreground MS might has the sweep done concurrently 3628 // or STW. 3629 if (UseAdaptiveSizePolicy) { 3630 size_policy()->concurrent_marking_begin(); 3631 } 3632 3633 // Weak ref discovery note: We may be discovering weak 3634 // refs in this generation concurrent (but interleaved) with 3635 // weak ref discovery by a younger generation collector. 3636 3637 CMSTokenSyncWithLocks ts(true, bitMapLock()); 3638 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 3639 CMSPhaseAccounting pa(this, "mark", !PrintGCDetails); 3640 res = markFromRootsWork(asynch); 3641 if (res) { 3642 _collectorState = Precleaning; 3643 } else { // We failed and a foreground collection wants to take over 3644 assert(_foregroundGCIsActive, "internal state inconsistency"); 3645 assert(_restart_addr == NULL, "foreground will restart from scratch"); 3646 if (PrintGCDetails) { 3647 gclog_or_tty->print_cr("bailing out to foreground collection"); 3648 } 3649 } 3650 if (UseAdaptiveSizePolicy) { 3651 size_policy()->concurrent_marking_end(); 3652 } 3653 } else { 3654 assert(SafepointSynchronize::is_at_safepoint(), 3655 "inconsistent with asynch == false"); 3656 if (UseAdaptiveSizePolicy) { 3657 size_policy()->ms_collection_marking_begin(); 3658 } 3659 // already have locks 3660 res = markFromRootsWork(asynch); 3661 _collectorState = FinalMarking; 3662 if (UseAdaptiveSizePolicy) { 3663 GenCollectedHeap* gch = GenCollectedHeap::heap(); 3664 size_policy()->ms_collection_marking_end(gch->gc_cause()); 3665 } 3666 } 3667 verify_overflow_empty(); 3668 return res; 3669 } 3670 3671 bool CMSCollector::markFromRootsWork(bool asynch) { 3672 // iterate over marked bits in bit map, doing a full scan and mark 3673 // from these roots using the following algorithm: 3674 // . if oop is to the right of the current scan pointer, 3675 // mark corresponding bit (we'll process it later) 3676 // . else (oop is to left of current scan pointer) 3677 // push oop on marking stack 3678 // . drain the marking stack 3679 3680 // Note that when we do a marking step we need to hold the 3681 // bit map lock -- recall that direct allocation (by mutators) 3682 // and promotion (by younger generation collectors) is also 3683 // marking the bit map. [the so-called allocate live policy.] 3684 // Because the implementation of bit map marking is not 3685 // robust wrt simultaneous marking of bits in the same word, 3686 // we need to make sure that there is no such interference 3687 // between concurrent such updates. 3688 3689 // already have locks 3690 assert_lock_strong(bitMapLock()); 3691 3692 // Clear the revisit stack, just in case there are any 3693 // obsolete contents from a short-circuited previous CMS cycle. 3694 _revisitStack.reset(); 3695 verify_work_stacks_empty(); 3696 verify_overflow_empty(); 3697 assert(_revisitStack.isEmpty(), "tabula rasa"); 3698 DEBUG_ONLY(RememberKlassesChecker cmx(should_unload_classes());) 3699 bool result = false; 3700 if (CMSConcurrentMTEnabled && ConcGCThreads > 0) { 3701 result = do_marking_mt(asynch); 3702 } else { 3703 result = do_marking_st(asynch); 3704 } 3705 return result; 3706 } 3707 3708 // Forward decl 3709 class CMSConcMarkingTask; 3710 3711 class CMSConcMarkingTerminator: public ParallelTaskTerminator { 3712 CMSCollector* _collector; 3713 CMSConcMarkingTask* _task; 3714 public: 3715 virtual void yield(); 3716 3717 // "n_threads" is the number of threads to be terminated. 3718 // "queue_set" is a set of work queues of other threads. 3719 // "collector" is the CMS collector associated with this task terminator. 3720 // "yield" indicates whether we need the gang as a whole to yield. 3721 CMSConcMarkingTerminator(int n_threads, TaskQueueSetSuper* queue_set, CMSCollector* collector) : 3722 ParallelTaskTerminator(n_threads, queue_set), 3723 _collector(collector) { } 3724 3725 void set_task(CMSConcMarkingTask* task) { 3726 _task = task; 3727 } 3728 }; 3729 3730 class CMSConcMarkingTerminatorTerminator: public TerminatorTerminator { 3731 CMSConcMarkingTask* _task; 3732 public: 3733 bool should_exit_termination(); 3734 void set_task(CMSConcMarkingTask* task) { 3735 _task = task; 3736 } 3737 }; 3738 3739 // MT Concurrent Marking Task 3740 class CMSConcMarkingTask: public YieldingFlexibleGangTask { 3741 CMSCollector* _collector; 3742 int _n_workers; // requested/desired # workers 3743 bool _asynch; 3744 bool _result; 3745 CompactibleFreeListSpace* _cms_space; 3746 CompactibleFreeListSpace* _perm_space; 3747 char _pad_front[64]; // padding to ... 3748 HeapWord* _global_finger; // ... avoid sharing cache line 3749 char _pad_back[64]; 3750 HeapWord* _restart_addr; 3751 3752 // Exposed here for yielding support 3753 Mutex* const _bit_map_lock; 3754 3755 // The per thread work queues, available here for stealing 3756 OopTaskQueueSet* _task_queues; 3757 3758 // Termination (and yielding) support 3759 CMSConcMarkingTerminator _term; 3760 CMSConcMarkingTerminatorTerminator _term_term; 3761 3762 public: 3763 CMSConcMarkingTask(CMSCollector* collector, 3764 CompactibleFreeListSpace* cms_space, 3765 CompactibleFreeListSpace* perm_space, 3766 bool asynch, 3767 YieldingFlexibleWorkGang* workers, 3768 OopTaskQueueSet* task_queues): 3769 YieldingFlexibleGangTask("Concurrent marking done multi-threaded"), 3770 _collector(collector), 3771 _cms_space(cms_space), 3772 _perm_space(perm_space), 3773 _asynch(asynch), _n_workers(0), _result(true), 3774 _task_queues(task_queues), 3775 _term(_n_workers, task_queues, _collector), 3776 _bit_map_lock(collector->bitMapLock()) 3777 { 3778 _requested_size = _n_workers; 3779 _term.set_task(this); 3780 _term_term.set_task(this); 3781 assert(_cms_space->bottom() < _perm_space->bottom(), 3782 "Finger incorrectly initialized below"); 3783 _restart_addr = _global_finger = _cms_space->bottom(); 3784 } 3785 3786 3787 OopTaskQueueSet* task_queues() { return _task_queues; } 3788 3789 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 3790 3791 HeapWord** global_finger_addr() { return &_global_finger; } 3792 3793 CMSConcMarkingTerminator* terminator() { return &_term; } 3794 3795 virtual void set_for_termination(int active_workers) { 3796 terminator()->reset_for_reuse(active_workers); 3797 } 3798 3799 void work(int i); 3800 bool should_yield() { 3801 return ConcurrentMarkSweepThread::should_yield() 3802 && !_collector->foregroundGCIsActive() 3803 && _asynch; 3804 } 3805 3806 virtual void coordinator_yield(); // stuff done by coordinator 3807 bool result() { return _result; } 3808 3809 void reset(HeapWord* ra) { 3810 assert(_global_finger >= _cms_space->end(), "Postcondition of ::work(i)"); 3811 assert(_global_finger >= _perm_space->end(), "Postcondition of ::work(i)"); 3812 assert(ra < _perm_space->end(), "ra too large"); 3813 _restart_addr = _global_finger = ra; 3814 _term.reset_for_reuse(); 3815 } 3816 3817 static bool get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 3818 OopTaskQueue* work_q); 3819 3820 private: 3821 void do_scan_and_mark(int i, CompactibleFreeListSpace* sp); 3822 void do_work_steal(int i); 3823 void bump_global_finger(HeapWord* f); 3824 }; 3825 3826 bool CMSConcMarkingTerminatorTerminator::should_exit_termination() { 3827 assert(_task != NULL, "Error"); 3828 return _task->yielding(); 3829 // Note that we do not need the disjunct || _task->should_yield() above 3830 // because we want terminating threads to yield only if the task 3831 // is already in the midst of yielding, which happens only after at least one 3832 // thread has yielded. 3833 } 3834 3835 void CMSConcMarkingTerminator::yield() { 3836 if (_task->should_yield()) { 3837 _task->yield(); 3838 } else { 3839 ParallelTaskTerminator::yield(); 3840 } 3841 } 3842 3843 //////////////////////////////////////////////////////////////// 3844 // Concurrent Marking Algorithm Sketch 3845 //////////////////////////////////////////////////////////////// 3846 // Until all tasks exhausted (both spaces): 3847 // -- claim next available chunk 3848 // -- bump global finger via CAS 3849 // -- find first object that starts in this chunk 3850 // and start scanning bitmap from that position 3851 // -- scan marked objects for oops 3852 // -- CAS-mark target, and if successful: 3853 // . if target oop is above global finger (volatile read) 3854 // nothing to do 3855 // . if target oop is in chunk and above local finger 3856 // then nothing to do 3857 // . else push on work-queue 3858 // -- Deal with possible overflow issues: 3859 // . local work-queue overflow causes stuff to be pushed on 3860 // global (common) overflow queue 3861 // . always first empty local work queue 3862 // . then get a batch of oops from global work queue if any 3863 // . then do work stealing 3864 // -- When all tasks claimed (both spaces) 3865 // and local work queue empty, 3866 // then in a loop do: 3867 // . check global overflow stack; steal a batch of oops and trace 3868 // . try to steal from other threads oif GOS is empty 3869 // . if neither is available, offer termination 3870 // -- Terminate and return result 3871 // 3872 void CMSConcMarkingTask::work(int i) { 3873 elapsedTimer _timer; 3874 ResourceMark rm; 3875 HandleMark hm; 3876 3877 DEBUG_ONLY(_collector->verify_overflow_empty();) 3878 3879 // Before we begin work, our work queue should be empty 3880 assert(work_queue(i)->size() == 0, "Expected to be empty"); 3881 // Scan the bitmap covering _cms_space, tracing through grey objects. 3882 _timer.start(); 3883 do_scan_and_mark(i, _cms_space); 3884 _timer.stop(); 3885 if (PrintCMSStatistics != 0) { 3886 gclog_or_tty->print_cr("Finished cms space scanning in %dth thread: %3.3f sec", 3887 i, _timer.seconds()); // XXX: need xxx/xxx type of notation, two timers 3888 } 3889 3890 // ... do the same for the _perm_space 3891 _timer.reset(); 3892 _timer.start(); 3893 do_scan_and_mark(i, _perm_space); 3894 _timer.stop(); 3895 if (PrintCMSStatistics != 0) { 3896 gclog_or_tty->print_cr("Finished perm space scanning in %dth thread: %3.3f sec", 3897 i, _timer.seconds()); // XXX: need xxx/xxx type of notation, two timers 3898 } 3899 3900 // ... do work stealing 3901 _timer.reset(); 3902 _timer.start(); 3903 do_work_steal(i); 3904 _timer.stop(); 3905 if (PrintCMSStatistics != 0) { 3906 gclog_or_tty->print_cr("Finished work stealing in %dth thread: %3.3f sec", 3907 i, _timer.seconds()); // XXX: need xxx/xxx type of notation, two timers 3908 } 3909 assert(_collector->_markStack.isEmpty(), "Should have been emptied"); 3910 assert(work_queue(i)->size() == 0, "Should have been emptied"); 3911 // Note that under the current task protocol, the 3912 // following assertion is true even of the spaces 3913 // expanded since the completion of the concurrent 3914 // marking. XXX This will likely change under a strict 3915 // ABORT semantics. 3916 assert(_global_finger > _cms_space->end() && 3917 _global_finger >= _perm_space->end(), 3918 "All tasks have been completed"); 3919 DEBUG_ONLY(_collector->verify_overflow_empty();) 3920 } 3921 3922 void CMSConcMarkingTask::bump_global_finger(HeapWord* f) { 3923 HeapWord* read = _global_finger; 3924 HeapWord* cur = read; 3925 while (f > read) { 3926 cur = read; 3927 read = (HeapWord*) Atomic::cmpxchg_ptr(f, &_global_finger, cur); 3928 if (cur == read) { 3929 // our cas succeeded 3930 assert(_global_finger >= f, "protocol consistency"); 3931 break; 3932 } 3933 } 3934 } 3935 3936 // This is really inefficient, and should be redone by 3937 // using (not yet available) block-read and -write interfaces to the 3938 // stack and the work_queue. XXX FIX ME !!! 3939 bool CMSConcMarkingTask::get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 3940 OopTaskQueue* work_q) { 3941 // Fast lock-free check 3942 if (ovflw_stk->length() == 0) { 3943 return false; 3944 } 3945 assert(work_q->size() == 0, "Shouldn't steal"); 3946 MutexLockerEx ml(ovflw_stk->par_lock(), 3947 Mutex::_no_safepoint_check_flag); 3948 // Grab up to 1/4 the size of the work queue 3949 size_t num = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 3950 (size_t)ParGCDesiredObjsFromOverflowList); 3951 num = MIN2(num, ovflw_stk->length()); 3952 for (int i = (int) num; i > 0; i--) { 3953 oop cur = ovflw_stk->pop(); 3954 assert(cur != NULL, "Counted wrong?"); 3955 work_q->push(cur); 3956 } 3957 return num > 0; 3958 } 3959 3960 void CMSConcMarkingTask::do_scan_and_mark(int i, CompactibleFreeListSpace* sp) { 3961 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 3962 int n_tasks = pst->n_tasks(); 3963 // We allow that there may be no tasks to do here because 3964 // we are restarting after a stack overflow. 3965 assert(pst->valid() || n_tasks == 0, "Uninitialized use?"); 3966 int nth_task = 0; 3967 3968 HeapWord* aligned_start = sp->bottom(); 3969 if (sp->used_region().contains(_restart_addr)) { 3970 // Align down to a card boundary for the start of 0th task 3971 // for this space. 3972 aligned_start = 3973 (HeapWord*)align_size_down((uintptr_t)_restart_addr, 3974 CardTableModRefBS::card_size); 3975 } 3976 3977 size_t chunk_size = sp->marking_task_size(); 3978 while (!pst->is_task_claimed(/* reference */ nth_task)) { 3979 // Having claimed the nth task in this space, 3980 // compute the chunk that it corresponds to: 3981 MemRegion span = MemRegion(aligned_start + nth_task*chunk_size, 3982 aligned_start + (nth_task+1)*chunk_size); 3983 // Try and bump the global finger via a CAS; 3984 // note that we need to do the global finger bump 3985 // _before_ taking the intersection below, because 3986 // the task corresponding to that region will be 3987 // deemed done even if the used_region() expands 3988 // because of allocation -- as it almost certainly will 3989 // during start-up while the threads yield in the 3990 // closure below. 3991 HeapWord* finger = span.end(); 3992 bump_global_finger(finger); // atomically 3993 // There are null tasks here corresponding to chunks 3994 // beyond the "top" address of the space. 3995 span = span.intersection(sp->used_region()); 3996 if (!span.is_empty()) { // Non-null task 3997 HeapWord* prev_obj; 3998 assert(!span.contains(_restart_addr) || nth_task == 0, 3999 "Inconsistency"); 4000 if (nth_task == 0) { 4001 // For the 0th task, we'll not need to compute a block_start. 4002 if (span.contains(_restart_addr)) { 4003 // In the case of a restart because of stack overflow, 4004 // we might additionally skip a chunk prefix. 4005 prev_obj = _restart_addr; 4006 } else { 4007 prev_obj = span.start(); 4008 } 4009 } else { 4010 // We want to skip the first object because 4011 // the protocol is to scan any object in its entirety 4012 // that _starts_ in this span; a fortiori, any 4013 // object starting in an earlier span is scanned 4014 // as part of an earlier claimed task. 4015 // Below we use the "careful" version of block_start 4016 // so we do not try to navigate uninitialized objects. 4017 prev_obj = sp->block_start_careful(span.start()); 4018 // Below we use a variant of block_size that uses the 4019 // Printezis bits to avoid waiting for allocated 4020 // objects to become initialized/parsable. 4021 while (prev_obj < span.start()) { 4022 size_t sz = sp->block_size_no_stall(prev_obj, _collector); 4023 if (sz > 0) { 4024 prev_obj += sz; 4025 } else { 4026 // In this case we may end up doing a bit of redundant 4027 // scanning, but that appears unavoidable, short of 4028 // locking the free list locks; see bug 6324141. 4029 break; 4030 } 4031 } 4032 } 4033 if (prev_obj < span.end()) { 4034 MemRegion my_span = MemRegion(prev_obj, span.end()); 4035 // Do the marking work within a non-empty span -- 4036 // the last argument to the constructor indicates whether the 4037 // iteration should be incremental with periodic yields. 4038 Par_MarkFromRootsClosure cl(this, _collector, my_span, 4039 &_collector->_markBitMap, 4040 work_queue(i), 4041 &_collector->_markStack, 4042 &_collector->_revisitStack, 4043 _asynch); 4044 _collector->_markBitMap.iterate(&cl, my_span.start(), my_span.end()); 4045 } // else nothing to do for this task 4046 } // else nothing to do for this task 4047 } 4048 // We'd be tempted to assert here that since there are no 4049 // more tasks left to claim in this space, the global_finger 4050 // must exceed space->top() and a fortiori space->end(). However, 4051 // that would not quite be correct because the bumping of 4052 // global_finger occurs strictly after the claiming of a task, 4053 // so by the time we reach here the global finger may not yet 4054 // have been bumped up by the thread that claimed the last 4055 // task. 4056 pst->all_tasks_completed(); 4057 } 4058 4059 class Par_ConcMarkingClosure: public Par_KlassRememberingOopClosure { 4060 private: 4061 CMSConcMarkingTask* _task; 4062 MemRegion _span; 4063 CMSBitMap* _bit_map; 4064 CMSMarkStack* _overflow_stack; 4065 OopTaskQueue* _work_queue; 4066 protected: 4067 DO_OOP_WORK_DEFN 4068 public: 4069 Par_ConcMarkingClosure(CMSCollector* collector, CMSConcMarkingTask* task, OopTaskQueue* work_queue, 4070 CMSBitMap* bit_map, CMSMarkStack* overflow_stack, 4071 CMSMarkStack* revisit_stack): 4072 Par_KlassRememberingOopClosure(collector, NULL, revisit_stack), 4073 _task(task), 4074 _span(collector->_span), 4075 _work_queue(work_queue), 4076 _bit_map(bit_map), 4077 _overflow_stack(overflow_stack) 4078 { } 4079 virtual void do_oop(oop* p); 4080 virtual void do_oop(narrowOop* p); 4081 void trim_queue(size_t max); 4082 void handle_stack_overflow(HeapWord* lost); 4083 void do_yield_check() { 4084 if (_task->should_yield()) { 4085 _task->yield(); 4086 } 4087 } 4088 }; 4089 4090 // Grey object scanning during work stealing phase -- 4091 // the salient assumption here is that any references 4092 // that are in these stolen objects being scanned must 4093 // already have been initialized (else they would not have 4094 // been published), so we do not need to check for 4095 // uninitialized objects before pushing here. 4096 void Par_ConcMarkingClosure::do_oop(oop obj) { 4097 assert(obj->is_oop_or_null(true), "expected an oop or NULL"); 4098 HeapWord* addr = (HeapWord*)obj; 4099 // Check if oop points into the CMS generation 4100 // and is not marked 4101 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 4102 // a white object ... 4103 // If we manage to "claim" the object, by being the 4104 // first thread to mark it, then we push it on our 4105 // marking stack 4106 if (_bit_map->par_mark(addr)) { // ... now grey 4107 // push on work queue (grey set) 4108 bool simulate_overflow = false; 4109 NOT_PRODUCT( 4110 if (CMSMarkStackOverflowALot && 4111 _collector->simulate_overflow()) { 4112 // simulate a stack overflow 4113 simulate_overflow = true; 4114 } 4115 ) 4116 if (simulate_overflow || 4117 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) { 4118 // stack overflow 4119 if (PrintCMSStatistics != 0) { 4120 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 4121 SIZE_FORMAT, _overflow_stack->capacity()); 4122 } 4123 // We cannot assert that the overflow stack is full because 4124 // it may have been emptied since. 4125 assert(simulate_overflow || 4126 _work_queue->size() == _work_queue->max_elems(), 4127 "Else push should have succeeded"); 4128 handle_stack_overflow(addr); 4129 } 4130 } // Else, some other thread got there first 4131 do_yield_check(); 4132 } 4133 } 4134 4135 void Par_ConcMarkingClosure::do_oop(oop* p) { Par_ConcMarkingClosure::do_oop_work(p); } 4136 void Par_ConcMarkingClosure::do_oop(narrowOop* p) { Par_ConcMarkingClosure::do_oop_work(p); } 4137 4138 void Par_ConcMarkingClosure::trim_queue(size_t max) { 4139 while (_work_queue->size() > max) { 4140 oop new_oop; 4141 if (_work_queue->pop_local(new_oop)) { 4142 assert(new_oop->is_oop(), "Should be an oop"); 4143 assert(_bit_map->isMarked((HeapWord*)new_oop), "Grey object"); 4144 assert(_span.contains((HeapWord*)new_oop), "Not in span"); 4145 assert(new_oop->is_parsable(), "Should be parsable"); 4146 new_oop->oop_iterate(this); // do_oop() above 4147 do_yield_check(); 4148 } 4149 } 4150 } 4151 4152 // Upon stack overflow, we discard (part of) the stack, 4153 // remembering the least address amongst those discarded 4154 // in CMSCollector's _restart_address. 4155 void Par_ConcMarkingClosure::handle_stack_overflow(HeapWord* lost) { 4156 // We need to do this under a mutex to prevent other 4157 // workers from interfering with the work done below. 4158 MutexLockerEx ml(_overflow_stack->par_lock(), 4159 Mutex::_no_safepoint_check_flag); 4160 // Remember the least grey address discarded 4161 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 4162 _collector->lower_restart_addr(ra); 4163 _overflow_stack->reset(); // discard stack contents 4164 _overflow_stack->expand(); // expand the stack if possible 4165 } 4166 4167 4168 void CMSConcMarkingTask::do_work_steal(int i) { 4169 OopTaskQueue* work_q = work_queue(i); 4170 oop obj_to_scan; 4171 CMSBitMap* bm = &(_collector->_markBitMap); 4172 CMSMarkStack* ovflw = &(_collector->_markStack); 4173 CMSMarkStack* revisit = &(_collector->_revisitStack); 4174 int* seed = _collector->hash_seed(i); 4175 Par_ConcMarkingClosure cl(_collector, this, work_q, bm, ovflw, revisit); 4176 while (true) { 4177 cl.trim_queue(0); 4178 assert(work_q->size() == 0, "Should have been emptied above"); 4179 if (get_work_from_overflow_stack(ovflw, work_q)) { 4180 // Can't assert below because the work obtained from the 4181 // overflow stack may already have been stolen from us. 4182 // assert(work_q->size() > 0, "Work from overflow stack"); 4183 continue; 4184 } else if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 4185 assert(obj_to_scan->is_oop(), "Should be an oop"); 4186 assert(bm->isMarked((HeapWord*)obj_to_scan), "Grey object"); 4187 obj_to_scan->oop_iterate(&cl); 4188 } else if (terminator()->offer_termination(&_term_term)) { 4189 assert(work_q->size() == 0, "Impossible!"); 4190 break; 4191 } else if (yielding() || should_yield()) { 4192 yield(); 4193 } 4194 } 4195 } 4196 4197 // This is run by the CMS (coordinator) thread. 4198 void CMSConcMarkingTask::coordinator_yield() { 4199 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 4200 "CMS thread should hold CMS token"); 4201 DEBUG_ONLY(RememberKlassesChecker mux(false);) 4202 // First give up the locks, then yield, then re-lock 4203 // We should probably use a constructor/destructor idiom to 4204 // do this unlock/lock or modify the MutexUnlocker class to 4205 // serve our purpose. XXX 4206 assert_lock_strong(_bit_map_lock); 4207 _bit_map_lock->unlock(); 4208 ConcurrentMarkSweepThread::desynchronize(true); 4209 ConcurrentMarkSweepThread::acknowledge_yield_request(); 4210 _collector->stopTimer(); 4211 if (PrintCMSStatistics != 0) { 4212 _collector->incrementYields(); 4213 } 4214 _collector->icms_wait(); 4215 4216 // It is possible for whichever thread initiated the yield request 4217 // not to get a chance to wake up and take the bitmap lock between 4218 // this thread releasing it and reacquiring it. So, while the 4219 // should_yield() flag is on, let's sleep for a bit to give the 4220 // other thread a chance to wake up. The limit imposed on the number 4221 // of iterations is defensive, to avoid any unforseen circumstances 4222 // putting us into an infinite loop. Since it's always been this 4223 // (coordinator_yield()) method that was observed to cause the 4224 // problem, we are using a parameter (CMSCoordinatorYieldSleepCount) 4225 // which is by default non-zero. For the other seven methods that 4226 // also perform the yield operation, as are using a different 4227 // parameter (CMSYieldSleepCount) which is by default zero. This way we 4228 // can enable the sleeping for those methods too, if necessary. 4229 // See 6442774. 4230 // 4231 // We really need to reconsider the synchronization between the GC 4232 // thread and the yield-requesting threads in the future and we 4233 // should really use wait/notify, which is the recommended 4234 // way of doing this type of interaction. Additionally, we should 4235 // consolidate the eight methods that do the yield operation and they 4236 // are almost identical into one for better maintenability and 4237 // readability. See 6445193. 4238 // 4239 // Tony 2006.06.29 4240 for (unsigned i = 0; i < CMSCoordinatorYieldSleepCount && 4241 ConcurrentMarkSweepThread::should_yield() && 4242 !CMSCollector::foregroundGCIsActive(); ++i) { 4243 os::sleep(Thread::current(), 1, false); 4244 ConcurrentMarkSweepThread::acknowledge_yield_request(); 4245 } 4246 4247 ConcurrentMarkSweepThread::synchronize(true); 4248 _bit_map_lock->lock_without_safepoint_check(); 4249 _collector->startTimer(); 4250 } 4251 4252 bool CMSCollector::do_marking_mt(bool asynch) { 4253 assert(ConcGCThreads > 0 && conc_workers() != NULL, "precondition"); 4254 // In the future this would be determined ergonomically, based 4255 // on #cpu's, # active mutator threads (and load), and mutation rate. 4256 int num_workers = ConcGCThreads; 4257 4258 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 4259 CompactibleFreeListSpace* perm_space = _permGen->cmsSpace(); 4260 4261 CMSConcMarkingTask tsk(this, 4262 cms_space, 4263 perm_space, 4264 asynch, 4265 conc_workers(), 4266 task_queues()); 4267 4268 // Since the actual number of workers we get may be different 4269 // from the number we requested above, do we need to do anything different 4270 // below? In particular, may be we need to subclass the SequantialSubTasksDone 4271 // class?? XXX 4272 cms_space ->initialize_sequential_subtasks_for_marking(num_workers); 4273 perm_space->initialize_sequential_subtasks_for_marking(num_workers); 4274 4275 // Refs discovery is already non-atomic. 4276 assert(!ref_processor()->discovery_is_atomic(), "Should be non-atomic"); 4277 // Mutate the Refs discovery so it is MT during the 4278 // multi-threaded marking phase. 4279 ReferenceProcessorMTMutator mt(ref_processor(), num_workers > 1); 4280 DEBUG_ONLY(RememberKlassesChecker cmx(should_unload_classes());) 4281 conc_workers()->start_task(&tsk); 4282 while (tsk.yielded()) { 4283 tsk.coordinator_yield(); 4284 conc_workers()->continue_task(&tsk); 4285 } 4286 // If the task was aborted, _restart_addr will be non-NULL 4287 assert(tsk.completed() || _restart_addr != NULL, "Inconsistency"); 4288 while (_restart_addr != NULL) { 4289 // XXX For now we do not make use of ABORTED state and have not 4290 // yet implemented the right abort semantics (even in the original 4291 // single-threaded CMS case). That needs some more investigation 4292 // and is deferred for now; see CR# TBF. 07252005YSR. XXX 4293 assert(!CMSAbortSemantics || tsk.aborted(), "Inconsistency"); 4294 // If _restart_addr is non-NULL, a marking stack overflow 4295 // occurred; we need to do a fresh marking iteration from the 4296 // indicated restart address. 4297 if (_foregroundGCIsActive && asynch) { 4298 // We may be running into repeated stack overflows, having 4299 // reached the limit of the stack size, while making very 4300 // slow forward progress. It may be best to bail out and 4301 // let the foreground collector do its job. 4302 // Clear _restart_addr, so that foreground GC 4303 // works from scratch. This avoids the headache of 4304 // a "rescan" which would otherwise be needed because 4305 // of the dirty mod union table & card table. 4306 _restart_addr = NULL; 4307 return false; 4308 } 4309 // Adjust the task to restart from _restart_addr 4310 tsk.reset(_restart_addr); 4311 cms_space ->initialize_sequential_subtasks_for_marking(num_workers, 4312 _restart_addr); 4313 perm_space->initialize_sequential_subtasks_for_marking(num_workers, 4314 _restart_addr); 4315 _restart_addr = NULL; 4316 // Get the workers going again 4317 conc_workers()->start_task(&tsk); 4318 while (tsk.yielded()) { 4319 tsk.coordinator_yield(); 4320 conc_workers()->continue_task(&tsk); 4321 } 4322 } 4323 assert(tsk.completed(), "Inconsistency"); 4324 assert(tsk.result() == true, "Inconsistency"); 4325 return true; 4326 } 4327 4328 bool CMSCollector::do_marking_st(bool asynch) { 4329 ResourceMark rm; 4330 HandleMark hm; 4331 4332 MarkFromRootsClosure markFromRootsClosure(this, _span, &_markBitMap, 4333 &_markStack, &_revisitStack, CMSYield && asynch); 4334 // the last argument to iterate indicates whether the iteration 4335 // should be incremental with periodic yields. 4336 _markBitMap.iterate(&markFromRootsClosure); 4337 // If _restart_addr is non-NULL, a marking stack overflow 4338 // occurred; we need to do a fresh iteration from the 4339 // indicated restart address. 4340 while (_restart_addr != NULL) { 4341 if (_foregroundGCIsActive && asynch) { 4342 // We may be running into repeated stack overflows, having 4343 // reached the limit of the stack size, while making very 4344 // slow forward progress. It may be best to bail out and 4345 // let the foreground collector do its job. 4346 // Clear _restart_addr, so that foreground GC 4347 // works from scratch. This avoids the headache of 4348 // a "rescan" which would otherwise be needed because 4349 // of the dirty mod union table & card table. 4350 _restart_addr = NULL; 4351 return false; // indicating failure to complete marking 4352 } 4353 // Deal with stack overflow: 4354 // we restart marking from _restart_addr 4355 HeapWord* ra = _restart_addr; 4356 markFromRootsClosure.reset(ra); 4357 _restart_addr = NULL; 4358 _markBitMap.iterate(&markFromRootsClosure, ra, _span.end()); 4359 } 4360 return true; 4361 } 4362 4363 void CMSCollector::preclean() { 4364 check_correct_thread_executing(); 4365 assert(Thread::current()->is_ConcurrentGC_thread(), "Wrong thread"); 4366 verify_work_stacks_empty(); 4367 verify_overflow_empty(); 4368 _abort_preclean = false; 4369 if (CMSPrecleaningEnabled) { 4370 // Precleaning is currently not MT but the reference processor 4371 // may be set for MT. Disable it temporarily here. 4372 ReferenceProcessor* rp = ref_processor(); 4373 ReferenceProcessorMTProcMutator z(rp, false); 4374 _eden_chunk_index = 0; 4375 size_t used = get_eden_used(); 4376 size_t capacity = get_eden_capacity(); 4377 // Don't start sampling unless we will get sufficiently 4378 // many samples. 4379 if (used < (capacity/(CMSScheduleRemarkSamplingRatio * 100) 4380 * CMSScheduleRemarkEdenPenetration)) { 4381 _start_sampling = true; 4382 } else { 4383 _start_sampling = false; 4384 } 4385 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 4386 CMSPhaseAccounting pa(this, "preclean", !PrintGCDetails); 4387 preclean_work(CMSPrecleanRefLists1, CMSPrecleanSurvivors1); 4388 } 4389 CMSTokenSync x(true); // is cms thread 4390 if (CMSPrecleaningEnabled) { 4391 sample_eden(); 4392 _collectorState = AbortablePreclean; 4393 } else { 4394 _collectorState = FinalMarking; 4395 } 4396 verify_work_stacks_empty(); 4397 verify_overflow_empty(); 4398 } 4399 4400 // Try and schedule the remark such that young gen 4401 // occupancy is CMSScheduleRemarkEdenPenetration %. 4402 void CMSCollector::abortable_preclean() { 4403 check_correct_thread_executing(); 4404 assert(CMSPrecleaningEnabled, "Inconsistent control state"); 4405 assert(_collectorState == AbortablePreclean, "Inconsistent control state"); 4406 4407 // If Eden's current occupancy is below this threshold, 4408 // immediately schedule the remark; else preclean 4409 // past the next scavenge in an effort to 4410 // schedule the pause as described avove. By choosing 4411 // CMSScheduleRemarkEdenSizeThreshold >= max eden size 4412 // we will never do an actual abortable preclean cycle. 4413 if (get_eden_used() > CMSScheduleRemarkEdenSizeThreshold) { 4414 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 4415 CMSPhaseAccounting pa(this, "abortable-preclean", !PrintGCDetails); 4416 // We need more smarts in the abortable preclean 4417 // loop below to deal with cases where allocation 4418 // in young gen is very very slow, and our precleaning 4419 // is running a losing race against a horde of 4420 // mutators intent on flooding us with CMS updates 4421 // (dirty cards). 4422 // One, admittedly dumb, strategy is to give up 4423 // after a certain number of abortable precleaning loops 4424 // or after a certain maximum time. We want to make 4425 // this smarter in the next iteration. 4426 // XXX FIX ME!!! YSR 4427 size_t loops = 0, workdone = 0, cumworkdone = 0, waited = 0; 4428 while (!(should_abort_preclean() || 4429 ConcurrentMarkSweepThread::should_terminate())) { 4430 workdone = preclean_work(CMSPrecleanRefLists2, CMSPrecleanSurvivors2); 4431 cumworkdone += workdone; 4432 loops++; 4433 // Voluntarily terminate abortable preclean phase if we have 4434 // been at it for too long. 4435 if ((CMSMaxAbortablePrecleanLoops != 0) && 4436 loops >= CMSMaxAbortablePrecleanLoops) { 4437 if (PrintGCDetails) { 4438 gclog_or_tty->print(" CMS: abort preclean due to loops "); 4439 } 4440 break; 4441 } 4442 if (pa.wallclock_millis() > CMSMaxAbortablePrecleanTime) { 4443 if (PrintGCDetails) { 4444 gclog_or_tty->print(" CMS: abort preclean due to time "); 4445 } 4446 break; 4447 } 4448 // If we are doing little work each iteration, we should 4449 // take a short break. 4450 if (workdone < CMSAbortablePrecleanMinWorkPerIteration) { 4451 // Sleep for some time, waiting for work to accumulate 4452 stopTimer(); 4453 cmsThread()->wait_on_cms_lock(CMSAbortablePrecleanWaitMillis); 4454 startTimer(); 4455 waited++; 4456 } 4457 } 4458 if (PrintCMSStatistics > 0) { 4459 gclog_or_tty->print(" [%d iterations, %d waits, %d cards)] ", 4460 loops, waited, cumworkdone); 4461 } 4462 } 4463 CMSTokenSync x(true); // is cms thread 4464 if (_collectorState != Idling) { 4465 assert(_collectorState == AbortablePreclean, 4466 "Spontaneous state transition?"); 4467 _collectorState = FinalMarking; 4468 } // Else, a foreground collection completed this CMS cycle. 4469 return; 4470 } 4471 4472 // Respond to an Eden sampling opportunity 4473 void CMSCollector::sample_eden() { 4474 // Make sure a young gc cannot sneak in between our 4475 // reading and recording of a sample. 4476 assert(Thread::current()->is_ConcurrentGC_thread(), 4477 "Only the cms thread may collect Eden samples"); 4478 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 4479 "Should collect samples while holding CMS token"); 4480 if (!_start_sampling) { 4481 return; 4482 } 4483 if (_eden_chunk_array) { 4484 if (_eden_chunk_index < _eden_chunk_capacity) { 4485 _eden_chunk_array[_eden_chunk_index] = *_top_addr; // take sample 4486 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr, 4487 "Unexpected state of Eden"); 4488 // We'd like to check that what we just sampled is an oop-start address; 4489 // however, we cannot do that here since the object may not yet have been 4490 // initialized. So we'll instead do the check when we _use_ this sample 4491 // later. 4492 if (_eden_chunk_index == 0 || 4493 (pointer_delta(_eden_chunk_array[_eden_chunk_index], 4494 _eden_chunk_array[_eden_chunk_index-1]) 4495 >= CMSSamplingGrain)) { 4496 _eden_chunk_index++; // commit sample 4497 } 4498 } 4499 } 4500 if ((_collectorState == AbortablePreclean) && !_abort_preclean) { 4501 size_t used = get_eden_used(); 4502 size_t capacity = get_eden_capacity(); 4503 assert(used <= capacity, "Unexpected state of Eden"); 4504 if (used > (capacity/100 * CMSScheduleRemarkEdenPenetration)) { 4505 _abort_preclean = true; 4506 } 4507 } 4508 } 4509 4510 4511 size_t CMSCollector::preclean_work(bool clean_refs, bool clean_survivor) { 4512 assert(_collectorState == Precleaning || 4513 _collectorState == AbortablePreclean, "incorrect state"); 4514 ResourceMark rm; 4515 HandleMark hm; 4516 // Do one pass of scrubbing the discovered reference lists 4517 // to remove any reference objects with strongly-reachable 4518 // referents. 4519 if (clean_refs) { 4520 ReferenceProcessor* rp = ref_processor(); 4521 CMSPrecleanRefsYieldClosure yield_cl(this); 4522 assert(rp->span().equals(_span), "Spans should be equal"); 4523 CMSKeepAliveClosure keep_alive(this, _span, &_markBitMap, 4524 &_markStack, &_revisitStack, 4525 true /* preclean */); 4526 CMSDrainMarkingStackClosure complete_trace(this, 4527 _span, &_markBitMap, &_markStack, 4528 &keep_alive, true /* preclean */); 4529 4530 // We don't want this step to interfere with a young 4531 // collection because we don't want to take CPU 4532 // or memory bandwidth away from the young GC threads 4533 // (which may be as many as there are CPUs). 4534 // Note that we don't need to protect ourselves from 4535 // interference with mutators because they can't 4536 // manipulate the discovered reference lists nor affect 4537 // the computed reachability of the referents, the 4538 // only properties manipulated by the precleaning 4539 // of these reference lists. 4540 stopTimer(); 4541 CMSTokenSyncWithLocks x(true /* is cms thread */, 4542 bitMapLock()); 4543 startTimer(); 4544 sample_eden(); 4545 4546 // The following will yield to allow foreground 4547 // collection to proceed promptly. XXX YSR: 4548 // The code in this method may need further 4549 // tweaking for better performance and some restructuring 4550 // for cleaner interfaces. 4551 rp->preclean_discovered_references( 4552 rp->is_alive_non_header(), &keep_alive, &complete_trace, 4553 &yield_cl, should_unload_classes()); 4554 } 4555 4556 if (clean_survivor) { // preclean the active survivor space(s) 4557 assert(_young_gen->kind() == Generation::DefNew || 4558 _young_gen->kind() == Generation::ParNew || 4559 _young_gen->kind() == Generation::ASParNew, 4560 "incorrect type for cast"); 4561 DefNewGeneration* dng = (DefNewGeneration*)_young_gen; 4562 PushAndMarkClosure pam_cl(this, _span, ref_processor(), 4563 &_markBitMap, &_modUnionTable, 4564 &_markStack, &_revisitStack, 4565 true /* precleaning phase */); 4566 stopTimer(); 4567 CMSTokenSyncWithLocks ts(true /* is cms thread */, 4568 bitMapLock()); 4569 startTimer(); 4570 unsigned int before_count = 4571 GenCollectedHeap::heap()->total_collections(); 4572 SurvivorSpacePrecleanClosure 4573 sss_cl(this, _span, &_markBitMap, &_markStack, 4574 &pam_cl, before_count, CMSYield); 4575 DEBUG_ONLY(RememberKlassesChecker mx(should_unload_classes());) 4576 dng->from()->object_iterate_careful(&sss_cl); 4577 dng->to()->object_iterate_careful(&sss_cl); 4578 } 4579 MarkRefsIntoAndScanClosure 4580 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable, 4581 &_markStack, &_revisitStack, this, CMSYield, 4582 true /* precleaning phase */); 4583 // CAUTION: The following closure has persistent state that may need to 4584 // be reset upon a decrease in the sequence of addresses it 4585 // processes. 4586 ScanMarkedObjectsAgainCarefullyClosure 4587 smoac_cl(this, _span, 4588 &_markBitMap, &_markStack, &_revisitStack, &mrias_cl, CMSYield); 4589 4590 // Preclean dirty cards in ModUnionTable and CardTable using 4591 // appropriate convergence criterion; 4592 // repeat CMSPrecleanIter times unless we find that 4593 // we are losing. 4594 assert(CMSPrecleanIter < 10, "CMSPrecleanIter is too large"); 4595 assert(CMSPrecleanNumerator < CMSPrecleanDenominator, 4596 "Bad convergence multiplier"); 4597 assert(CMSPrecleanThreshold >= 100, 4598 "Unreasonably low CMSPrecleanThreshold"); 4599 4600 size_t numIter, cumNumCards, lastNumCards, curNumCards; 4601 for (numIter = 0, cumNumCards = lastNumCards = curNumCards = 0; 4602 numIter < CMSPrecleanIter; 4603 numIter++, lastNumCards = curNumCards, cumNumCards += curNumCards) { 4604 curNumCards = preclean_mod_union_table(_cmsGen, &smoac_cl); 4605 if (CMSPermGenPrecleaningEnabled) { 4606 curNumCards += preclean_mod_union_table(_permGen, &smoac_cl); 4607 } 4608 if (Verbose && PrintGCDetails) { 4609 gclog_or_tty->print(" (modUnionTable: %d cards)", curNumCards); 4610 } 4611 // Either there are very few dirty cards, so re-mark 4612 // pause will be small anyway, or our pre-cleaning isn't 4613 // that much faster than the rate at which cards are being 4614 // dirtied, so we might as well stop and re-mark since 4615 // precleaning won't improve our re-mark time by much. 4616 if (curNumCards <= CMSPrecleanThreshold || 4617 (numIter > 0 && 4618 (curNumCards * CMSPrecleanDenominator > 4619 lastNumCards * CMSPrecleanNumerator))) { 4620 numIter++; 4621 cumNumCards += curNumCards; 4622 break; 4623 } 4624 } 4625 curNumCards = preclean_card_table(_cmsGen, &smoac_cl); 4626 if (CMSPermGenPrecleaningEnabled) { 4627 curNumCards += preclean_card_table(_permGen, &smoac_cl); 4628 } 4629 cumNumCards += curNumCards; 4630 if (PrintGCDetails && PrintCMSStatistics != 0) { 4631 gclog_or_tty->print_cr(" (cardTable: %d cards, re-scanned %d cards, %d iterations)", 4632 curNumCards, cumNumCards, numIter); 4633 } 4634 return cumNumCards; // as a measure of useful work done 4635 } 4636 4637 // PRECLEANING NOTES: 4638 // Precleaning involves: 4639 // . reading the bits of the modUnionTable and clearing the set bits. 4640 // . For the cards corresponding to the set bits, we scan the 4641 // objects on those cards. This means we need the free_list_lock 4642 // so that we can safely iterate over the CMS space when scanning 4643 // for oops. 4644 // . When we scan the objects, we'll be both reading and setting 4645 // marks in the marking bit map, so we'll need the marking bit map. 4646 // . For protecting _collector_state transitions, we take the CGC_lock. 4647 // Note that any races in the reading of of card table entries by the 4648 // CMS thread on the one hand and the clearing of those entries by the 4649 // VM thread or the setting of those entries by the mutator threads on the 4650 // other are quite benign. However, for efficiency it makes sense to keep 4651 // the VM thread from racing with the CMS thread while the latter is 4652 // dirty card info to the modUnionTable. We therefore also use the 4653 // CGC_lock to protect the reading of the card table and the mod union 4654 // table by the CM thread. 4655 // . We run concurrently with mutator updates, so scanning 4656 // needs to be done carefully -- we should not try to scan 4657 // potentially uninitialized objects. 4658 // 4659 // Locking strategy: While holding the CGC_lock, we scan over and 4660 // reset a maximal dirty range of the mod union / card tables, then lock 4661 // the free_list_lock and bitmap lock to do a full marking, then 4662 // release these locks; and repeat the cycle. This allows for a 4663 // certain amount of fairness in the sharing of these locks between 4664 // the CMS collector on the one hand, and the VM thread and the 4665 // mutators on the other. 4666 4667 // NOTE: preclean_mod_union_table() and preclean_card_table() 4668 // further below are largely identical; if you need to modify 4669 // one of these methods, please check the other method too. 4670 4671 size_t CMSCollector::preclean_mod_union_table( 4672 ConcurrentMarkSweepGeneration* gen, 4673 ScanMarkedObjectsAgainCarefullyClosure* cl) { 4674 verify_work_stacks_empty(); 4675 verify_overflow_empty(); 4676 4677 // Turn off checking for this method but turn it back on 4678 // selectively. There are yield points in this method 4679 // but it is difficult to turn the checking off just around 4680 // the yield points. It is simpler to selectively turn 4681 // it on. 4682 DEBUG_ONLY(RememberKlassesChecker mux(false);) 4683 4684 // strategy: starting with the first card, accumulate contiguous 4685 // ranges of dirty cards; clear these cards, then scan the region 4686 // covered by these cards. 4687 4688 // Since all of the MUT is committed ahead, we can just use 4689 // that, in case the generations expand while we are precleaning. 4690 // It might also be fine to just use the committed part of the 4691 // generation, but we might potentially miss cards when the 4692 // generation is rapidly expanding while we are in the midst 4693 // of precleaning. 4694 HeapWord* startAddr = gen->reserved().start(); 4695 HeapWord* endAddr = gen->reserved().end(); 4696 4697 cl->setFreelistLock(gen->freelistLock()); // needed for yielding 4698 4699 size_t numDirtyCards, cumNumDirtyCards; 4700 HeapWord *nextAddr, *lastAddr; 4701 for (cumNumDirtyCards = numDirtyCards = 0, 4702 nextAddr = lastAddr = startAddr; 4703 nextAddr < endAddr; 4704 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 4705 4706 ResourceMark rm; 4707 HandleMark hm; 4708 4709 MemRegion dirtyRegion; 4710 { 4711 stopTimer(); 4712 // Potential yield point 4713 CMSTokenSync ts(true); 4714 startTimer(); 4715 sample_eden(); 4716 // Get dirty region starting at nextOffset (inclusive), 4717 // simultaneously clearing it. 4718 dirtyRegion = 4719 _modUnionTable.getAndClearMarkedRegion(nextAddr, endAddr); 4720 assert(dirtyRegion.start() >= nextAddr, 4721 "returned region inconsistent?"); 4722 } 4723 // Remember where the next search should begin. 4724 // The returned region (if non-empty) is a right open interval, 4725 // so lastOffset is obtained from the right end of that 4726 // interval. 4727 lastAddr = dirtyRegion.end(); 4728 // Should do something more transparent and less hacky XXX 4729 numDirtyCards = 4730 _modUnionTable.heapWordDiffToOffsetDiff(dirtyRegion.word_size()); 4731 4732 // We'll scan the cards in the dirty region (with periodic 4733 // yields for foreground GC as needed). 4734 if (!dirtyRegion.is_empty()) { 4735 assert(numDirtyCards > 0, "consistency check"); 4736 HeapWord* stop_point = NULL; 4737 stopTimer(); 4738 // Potential yield point 4739 CMSTokenSyncWithLocks ts(true, gen->freelistLock(), 4740 bitMapLock()); 4741 startTimer(); 4742 { 4743 verify_work_stacks_empty(); 4744 verify_overflow_empty(); 4745 sample_eden(); 4746 DEBUG_ONLY(RememberKlassesChecker mx(should_unload_classes());) 4747 stop_point = 4748 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 4749 } 4750 if (stop_point != NULL) { 4751 // The careful iteration stopped early either because it found an 4752 // uninitialized object, or because we were in the midst of an 4753 // "abortable preclean", which should now be aborted. Redirty 4754 // the bits corresponding to the partially-scanned or unscanned 4755 // cards. We'll either restart at the next block boundary or 4756 // abort the preclean. 4757 assert((CMSPermGenPrecleaningEnabled && (gen == _permGen)) || 4758 (_collectorState == AbortablePreclean && should_abort_preclean()), 4759 "Unparsable objects should only be in perm gen."); 4760 _modUnionTable.mark_range(MemRegion(stop_point, dirtyRegion.end())); 4761 if (should_abort_preclean()) { 4762 break; // out of preclean loop 4763 } else { 4764 // Compute the next address at which preclean should pick up; 4765 // might need bitMapLock in order to read P-bits. 4766 lastAddr = next_card_start_after_block(stop_point); 4767 } 4768 } 4769 } else { 4770 assert(lastAddr == endAddr, "consistency check"); 4771 assert(numDirtyCards == 0, "consistency check"); 4772 break; 4773 } 4774 } 4775 verify_work_stacks_empty(); 4776 verify_overflow_empty(); 4777 return cumNumDirtyCards; 4778 } 4779 4780 // NOTE: preclean_mod_union_table() above and preclean_card_table() 4781 // below are largely identical; if you need to modify 4782 // one of these methods, please check the other method too. 4783 4784 size_t CMSCollector::preclean_card_table(ConcurrentMarkSweepGeneration* gen, 4785 ScanMarkedObjectsAgainCarefullyClosure* cl) { 4786 // strategy: it's similar to precleamModUnionTable above, in that 4787 // we accumulate contiguous ranges of dirty cards, mark these cards 4788 // precleaned, then scan the region covered by these cards. 4789 HeapWord* endAddr = (HeapWord*)(gen->_virtual_space.high()); 4790 HeapWord* startAddr = (HeapWord*)(gen->_virtual_space.low()); 4791 4792 cl->setFreelistLock(gen->freelistLock()); // needed for yielding 4793 4794 size_t numDirtyCards, cumNumDirtyCards; 4795 HeapWord *lastAddr, *nextAddr; 4796 4797 for (cumNumDirtyCards = numDirtyCards = 0, 4798 nextAddr = lastAddr = startAddr; 4799 nextAddr < endAddr; 4800 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 4801 4802 ResourceMark rm; 4803 HandleMark hm; 4804 4805 MemRegion dirtyRegion; 4806 { 4807 // See comments in "Precleaning notes" above on why we 4808 // do this locking. XXX Could the locking overheads be 4809 // too high when dirty cards are sparse? [I don't think so.] 4810 stopTimer(); 4811 CMSTokenSync x(true); // is cms thread 4812 startTimer(); 4813 sample_eden(); 4814 // Get and clear dirty region from card table 4815 dirtyRegion = _ct->ct_bs()->dirty_card_range_after_reset( 4816 MemRegion(nextAddr, endAddr), 4817 true, 4818 CardTableModRefBS::precleaned_card_val()); 4819 4820 assert(dirtyRegion.start() >= nextAddr, 4821 "returned region inconsistent?"); 4822 } 4823 lastAddr = dirtyRegion.end(); 4824 numDirtyCards = 4825 dirtyRegion.word_size()/CardTableModRefBS::card_size_in_words; 4826 4827 if (!dirtyRegion.is_empty()) { 4828 stopTimer(); 4829 CMSTokenSyncWithLocks ts(true, gen->freelistLock(), bitMapLock()); 4830 startTimer(); 4831 sample_eden(); 4832 verify_work_stacks_empty(); 4833 verify_overflow_empty(); 4834 DEBUG_ONLY(RememberKlassesChecker mx(should_unload_classes());) 4835 HeapWord* stop_point = 4836 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 4837 if (stop_point != NULL) { 4838 // The careful iteration stopped early because it found an 4839 // uninitialized object. Redirty the bits corresponding to the 4840 // partially-scanned or unscanned cards, and start again at the 4841 // next block boundary. 4842 assert(CMSPermGenPrecleaningEnabled || 4843 (_collectorState == AbortablePreclean && should_abort_preclean()), 4844 "Unparsable objects should only be in perm gen."); 4845 _ct->ct_bs()->invalidate(MemRegion(stop_point, dirtyRegion.end())); 4846 if (should_abort_preclean()) { 4847 break; // out of preclean loop 4848 } else { 4849 // Compute the next address at which preclean should pick up. 4850 lastAddr = next_card_start_after_block(stop_point); 4851 } 4852 } 4853 } else { 4854 break; 4855 } 4856 } 4857 verify_work_stacks_empty(); 4858 verify_overflow_empty(); 4859 return cumNumDirtyCards; 4860 } 4861 4862 void CMSCollector::checkpointRootsFinal(bool asynch, 4863 bool clear_all_soft_refs, bool init_mark_was_synchronous) { 4864 assert(_collectorState == FinalMarking, "incorrect state transition?"); 4865 check_correct_thread_executing(); 4866 // world is stopped at this checkpoint 4867 assert(SafepointSynchronize::is_at_safepoint(), 4868 "world should be stopped"); 4869 TraceCMSMemoryManagerStats tms(_collectorState); 4870 verify_work_stacks_empty(); 4871 verify_overflow_empty(); 4872 4873 SpecializationStats::clear(); 4874 if (PrintGCDetails) { 4875 gclog_or_tty->print("[YG occupancy: "SIZE_FORMAT" K ("SIZE_FORMAT" K)]", 4876 _young_gen->used() / K, 4877 _young_gen->capacity() / K); 4878 } 4879 if (asynch) { 4880 if (CMSScavengeBeforeRemark) { 4881 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4882 // Temporarily set flag to false, GCH->do_collection will 4883 // expect it to be false and set to true 4884 FlagSetting fl(gch->_is_gc_active, false); 4885 NOT_PRODUCT(TraceTime t("Scavenge-Before-Remark", 4886 PrintGCDetails && Verbose, true, gclog_or_tty);) 4887 int level = _cmsGen->level() - 1; 4888 if (level >= 0) { 4889 gch->do_collection(true, // full (i.e. force, see below) 4890 false, // !clear_all_soft_refs 4891 0, // size 4892 false, // is_tlab 4893 level // max_level 4894 ); 4895 } 4896 } 4897 FreelistLocker x(this); 4898 MutexLockerEx y(bitMapLock(), 4899 Mutex::_no_safepoint_check_flag); 4900 assert(!init_mark_was_synchronous, "but that's impossible!"); 4901 checkpointRootsFinalWork(asynch, clear_all_soft_refs, false); 4902 } else { 4903 // already have all the locks 4904 checkpointRootsFinalWork(asynch, clear_all_soft_refs, 4905 init_mark_was_synchronous); 4906 } 4907 verify_work_stacks_empty(); 4908 verify_overflow_empty(); 4909 SpecializationStats::print(); 4910 } 4911 4912 void CMSCollector::checkpointRootsFinalWork(bool asynch, 4913 bool clear_all_soft_refs, bool init_mark_was_synchronous) { 4914 4915 NOT_PRODUCT(TraceTime tr("checkpointRootsFinalWork", PrintGCDetails, false, gclog_or_tty);) 4916 4917 assert(haveFreelistLocks(), "must have free list locks"); 4918 assert_lock_strong(bitMapLock()); 4919 4920 if (UseAdaptiveSizePolicy) { 4921 size_policy()->checkpoint_roots_final_begin(); 4922 } 4923 4924 ResourceMark rm; 4925 HandleMark hm; 4926 4927 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4928 4929 if (should_unload_classes()) { 4930 CodeCache::gc_prologue(); 4931 } 4932 assert(haveFreelistLocks(), "must have free list locks"); 4933 assert_lock_strong(bitMapLock()); 4934 4935 DEBUG_ONLY(RememberKlassesChecker fmx(should_unload_classes());) 4936 if (!init_mark_was_synchronous) { 4937 // We might assume that we need not fill TLAB's when 4938 // CMSScavengeBeforeRemark is set, because we may have just done 4939 // a scavenge which would have filled all TLAB's -- and besides 4940 // Eden would be empty. This however may not always be the case -- 4941 // for instance although we asked for a scavenge, it may not have 4942 // happened because of a JNI critical section. We probably need 4943 // a policy for deciding whether we can in that case wait until 4944 // the critical section releases and then do the remark following 4945 // the scavenge, and skip it here. In the absence of that policy, 4946 // or of an indication of whether the scavenge did indeed occur, 4947 // we cannot rely on TLAB's having been filled and must do 4948 // so here just in case a scavenge did not happen. 4949 gch->ensure_parsability(false); // fill TLAB's, but no need to retire them 4950 // Update the saved marks which may affect the root scans. 4951 gch->save_marks(); 4952 4953 { 4954 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;) 4955 4956 // Note on the role of the mod union table: 4957 // Since the marker in "markFromRoots" marks concurrently with 4958 // mutators, it is possible for some reachable objects not to have been 4959 // scanned. For instance, an only reference to an object A was 4960 // placed in object B after the marker scanned B. Unless B is rescanned, 4961 // A would be collected. Such updates to references in marked objects 4962 // are detected via the mod union table which is the set of all cards 4963 // dirtied since the first checkpoint in this GC cycle and prior to 4964 // the most recent young generation GC, minus those cleaned up by the 4965 // concurrent precleaning. 4966 if (CMSParallelRemarkEnabled && CollectedHeap::use_parallel_gc_threads()) { 4967 TraceTime t("Rescan (parallel) ", PrintGCDetails, false, gclog_or_tty); 4968 do_remark_parallel(); 4969 } else { 4970 TraceTime t("Rescan (non-parallel) ", PrintGCDetails, false, 4971 gclog_or_tty); 4972 do_remark_non_parallel(); 4973 } 4974 } 4975 } else { 4976 assert(!asynch, "Can't have init_mark_was_synchronous in asynch mode"); 4977 // The initial mark was stop-world, so there's no rescanning to 4978 // do; go straight on to the next step below. 4979 } 4980 verify_work_stacks_empty(); 4981 verify_overflow_empty(); 4982 4983 { 4984 NOT_PRODUCT(TraceTime ts("refProcessingWork", PrintGCDetails, false, gclog_or_tty);) 4985 refProcessingWork(asynch, clear_all_soft_refs); 4986 } 4987 verify_work_stacks_empty(); 4988 verify_overflow_empty(); 4989 4990 if (should_unload_classes()) { 4991 CodeCache::gc_epilogue(); 4992 } 4993 4994 // If we encountered any (marking stack / work queue) overflow 4995 // events during the current CMS cycle, take appropriate 4996 // remedial measures, where possible, so as to try and avoid 4997 // recurrence of that condition. 4998 assert(_markStack.isEmpty(), "No grey objects"); 4999 size_t ser_ovflw = _ser_pmc_remark_ovflw + _ser_pmc_preclean_ovflw + 5000 _ser_kac_ovflw + _ser_kac_preclean_ovflw; 5001 if (ser_ovflw > 0) { 5002 if (PrintCMSStatistics != 0) { 5003 gclog_or_tty->print_cr("Marking stack overflow (benign) " 5004 "(pmc_pc="SIZE_FORMAT", pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT 5005 ", kac_preclean="SIZE_FORMAT")", 5006 _ser_pmc_preclean_ovflw, _ser_pmc_remark_ovflw, 5007 _ser_kac_ovflw, _ser_kac_preclean_ovflw); 5008 } 5009 _markStack.expand(); 5010 _ser_pmc_remark_ovflw = 0; 5011 _ser_pmc_preclean_ovflw = 0; 5012 _ser_kac_preclean_ovflw = 0; 5013 _ser_kac_ovflw = 0; 5014 } 5015 if (_par_pmc_remark_ovflw > 0 || _par_kac_ovflw > 0) { 5016 if (PrintCMSStatistics != 0) { 5017 gclog_or_tty->print_cr("Work queue overflow (benign) " 5018 "(pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT")", 5019 _par_pmc_remark_ovflw, _par_kac_ovflw); 5020 } 5021 _par_pmc_remark_ovflw = 0; 5022 _par_kac_ovflw = 0; 5023 } 5024 if (PrintCMSStatistics != 0) { 5025 if (_markStack._hit_limit > 0) { 5026 gclog_or_tty->print_cr(" (benign) Hit max stack size limit ("SIZE_FORMAT")", 5027 _markStack._hit_limit); 5028 } 5029 if (_markStack._failed_double > 0) { 5030 gclog_or_tty->print_cr(" (benign) Failed stack doubling ("SIZE_FORMAT")," 5031 " current capacity "SIZE_FORMAT, 5032 _markStack._failed_double, 5033 _markStack.capacity()); 5034 } 5035 } 5036 _markStack._hit_limit = 0; 5037 _markStack._failed_double = 0; 5038 5039 // Check that all the klasses have been checked 5040 assert(_revisitStack.isEmpty(), "Not all klasses revisited"); 5041 5042 if ((VerifyAfterGC || VerifyDuringGC) && 5043 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 5044 verify_after_remark(); 5045 } 5046 5047 // Change under the freelistLocks. 5048 _collectorState = Sweeping; 5049 // Call isAllClear() under bitMapLock 5050 assert(_modUnionTable.isAllClear(), "Should be clear by end of the" 5051 " final marking"); 5052 if (UseAdaptiveSizePolicy) { 5053 size_policy()->checkpoint_roots_final_end(gch->gc_cause()); 5054 } 5055 } 5056 5057 // Parallel remark task 5058 class CMSParRemarkTask: public AbstractGangTask { 5059 CMSCollector* _collector; 5060 int _n_workers; 5061 CompactibleFreeListSpace* _cms_space; 5062 CompactibleFreeListSpace* _perm_space; 5063 5064 // The per-thread work queues, available here for stealing. 5065 OopTaskQueueSet* _task_queues; 5066 ParallelTaskTerminator _term; 5067 5068 public: 5069 CMSParRemarkTask(CMSCollector* collector, 5070 CompactibleFreeListSpace* cms_space, 5071 CompactibleFreeListSpace* perm_space, 5072 int n_workers, FlexibleWorkGang* workers, 5073 OopTaskQueueSet* task_queues): 5074 AbstractGangTask("Rescan roots and grey objects in parallel"), 5075 _collector(collector), 5076 _cms_space(cms_space), _perm_space(perm_space), 5077 _n_workers(n_workers), 5078 _task_queues(task_queues), 5079 _term(n_workers, task_queues) { } 5080 5081 OopTaskQueueSet* task_queues() { return _task_queues; } 5082 5083 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 5084 5085 ParallelTaskTerminator* terminator() { return &_term; } 5086 int n_workers() { return _n_workers; } 5087 5088 void work(int i); 5089 5090 private: 5091 // Work method in support of parallel rescan ... of young gen spaces 5092 void do_young_space_rescan(int i, Par_MarkRefsIntoAndScanClosure* cl, 5093 ContiguousSpace* space, 5094 HeapWord** chunk_array, size_t chunk_top); 5095 5096 // ... of dirty cards in old space 5097 void do_dirty_card_rescan_tasks(CompactibleFreeListSpace* sp, int i, 5098 Par_MarkRefsIntoAndScanClosure* cl); 5099 5100 // ... work stealing for the above 5101 void do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, int* seed); 5102 }; 5103 5104 // work_queue(i) is passed to the closure 5105 // Par_MarkRefsIntoAndScanClosure. The "i" parameter 5106 // also is passed to do_dirty_card_rescan_tasks() and to 5107 // do_work_steal() to select the i-th task_queue. 5108 5109 void CMSParRemarkTask::work(int i) { 5110 elapsedTimer _timer; 5111 ResourceMark rm; 5112 HandleMark hm; 5113 5114 // ---------- rescan from roots -------------- 5115 _timer.start(); 5116 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5117 Par_MarkRefsIntoAndScanClosure par_mrias_cl(_collector, 5118 _collector->_span, _collector->ref_processor(), 5119 &(_collector->_markBitMap), 5120 work_queue(i), &(_collector->_revisitStack)); 5121 5122 // Rescan young gen roots first since these are likely 5123 // coarsely partitioned and may, on that account, constitute 5124 // the critical path; thus, it's best to start off that 5125 // work first. 5126 // ---------- young gen roots -------------- 5127 { 5128 DefNewGeneration* dng = _collector->_young_gen->as_DefNewGeneration(); 5129 EdenSpace* eden_space = dng->eden(); 5130 ContiguousSpace* from_space = dng->from(); 5131 ContiguousSpace* to_space = dng->to(); 5132 5133 HeapWord** eca = _collector->_eden_chunk_array; 5134 size_t ect = _collector->_eden_chunk_index; 5135 HeapWord** sca = _collector->_survivor_chunk_array; 5136 size_t sct = _collector->_survivor_chunk_index; 5137 5138 assert(ect <= _collector->_eden_chunk_capacity, "out of bounds"); 5139 assert(sct <= _collector->_survivor_chunk_capacity, "out of bounds"); 5140 5141 do_young_space_rescan(i, &par_mrias_cl, to_space, NULL, 0); 5142 do_young_space_rescan(i, &par_mrias_cl, from_space, sca, sct); 5143 do_young_space_rescan(i, &par_mrias_cl, eden_space, eca, ect); 5144 5145 _timer.stop(); 5146 if (PrintCMSStatistics != 0) { 5147 gclog_or_tty->print_cr( 5148 "Finished young gen rescan work in %dth thread: %3.3f sec", 5149 i, _timer.seconds()); 5150 } 5151 } 5152 5153 // ---------- remaining roots -------------- 5154 _timer.reset(); 5155 _timer.start(); 5156 gch->gen_process_strong_roots(_collector->_cmsGen->level(), 5157 false, // yg was scanned above 5158 false, // this is parallel code 5159 true, // collecting perm gen 5160 SharedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()), 5161 &par_mrias_cl, 5162 true, // walk all of code cache if (so & SO_CodeCache) 5163 NULL); 5164 assert(_collector->should_unload_classes() 5165 || (_collector->CMSCollector::roots_scanning_options() & SharedHeap::SO_CodeCache), 5166 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 5167 _timer.stop(); 5168 if (PrintCMSStatistics != 0) { 5169 gclog_or_tty->print_cr( 5170 "Finished remaining root rescan work in %dth thread: %3.3f sec", 5171 i, _timer.seconds()); 5172 } 5173 5174 // ---------- rescan dirty cards ------------ 5175 _timer.reset(); 5176 _timer.start(); 5177 5178 // Do the rescan tasks for each of the two spaces 5179 // (cms_space and perm_space) in turn. 5180 // "i" is passed to select the "i-th" task_queue 5181 do_dirty_card_rescan_tasks(_cms_space, i, &par_mrias_cl); 5182 do_dirty_card_rescan_tasks(_perm_space, i, &par_mrias_cl); 5183 _timer.stop(); 5184 if (PrintCMSStatistics != 0) { 5185 gclog_or_tty->print_cr( 5186 "Finished dirty card rescan work in %dth thread: %3.3f sec", 5187 i, _timer.seconds()); 5188 } 5189 5190 // ---------- steal work from other threads ... 5191 // ---------- ... and drain overflow list. 5192 _timer.reset(); 5193 _timer.start(); 5194 do_work_steal(i, &par_mrias_cl, _collector->hash_seed(i)); 5195 _timer.stop(); 5196 if (PrintCMSStatistics != 0) { 5197 gclog_or_tty->print_cr( 5198 "Finished work stealing in %dth thread: %3.3f sec", 5199 i, _timer.seconds()); 5200 } 5201 } 5202 5203 // Note that parameter "i" is not used. 5204 void 5205 CMSParRemarkTask::do_young_space_rescan(int i, 5206 Par_MarkRefsIntoAndScanClosure* cl, ContiguousSpace* space, 5207 HeapWord** chunk_array, size_t chunk_top) { 5208 // Until all tasks completed: 5209 // . claim an unclaimed task 5210 // . compute region boundaries corresponding to task claimed 5211 // using chunk_array 5212 // . par_oop_iterate(cl) over that region 5213 5214 ResourceMark rm; 5215 HandleMark hm; 5216 5217 SequentialSubTasksDone* pst = space->par_seq_tasks(); 5218 assert(pst->valid(), "Uninitialized use?"); 5219 5220 int nth_task = 0; 5221 int n_tasks = pst->n_tasks(); 5222 5223 HeapWord *start, *end; 5224 while (!pst->is_task_claimed(/* reference */ nth_task)) { 5225 // We claimed task # nth_task; compute its boundaries. 5226 if (chunk_top == 0) { // no samples were taken 5227 assert(nth_task == 0 && n_tasks == 1, "Can have only 1 EdenSpace task"); 5228 start = space->bottom(); 5229 end = space->top(); 5230 } else if (nth_task == 0) { 5231 start = space->bottom(); 5232 end = chunk_array[nth_task]; 5233 } else if (nth_task < (jint)chunk_top) { 5234 assert(nth_task >= 1, "Control point invariant"); 5235 start = chunk_array[nth_task - 1]; 5236 end = chunk_array[nth_task]; 5237 } else { 5238 assert(nth_task == (jint)chunk_top, "Control point invariant"); 5239 start = chunk_array[chunk_top - 1]; 5240 end = space->top(); 5241 } 5242 MemRegion mr(start, end); 5243 // Verify that mr is in space 5244 assert(mr.is_empty() || space->used_region().contains(mr), 5245 "Should be in space"); 5246 // Verify that "start" is an object boundary 5247 assert(mr.is_empty() || oop(mr.start())->is_oop(), 5248 "Should be an oop"); 5249 space->par_oop_iterate(mr, cl); 5250 } 5251 pst->all_tasks_completed(); 5252 } 5253 5254 void 5255 CMSParRemarkTask::do_dirty_card_rescan_tasks( 5256 CompactibleFreeListSpace* sp, int i, 5257 Par_MarkRefsIntoAndScanClosure* cl) { 5258 // Until all tasks completed: 5259 // . claim an unclaimed task 5260 // . compute region boundaries corresponding to task claimed 5261 // . transfer dirty bits ct->mut for that region 5262 // . apply rescanclosure to dirty mut bits for that region 5263 5264 ResourceMark rm; 5265 HandleMark hm; 5266 5267 OopTaskQueue* work_q = work_queue(i); 5268 ModUnionClosure modUnionClosure(&(_collector->_modUnionTable)); 5269 // CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! 5270 // CAUTION: This closure has state that persists across calls to 5271 // the work method dirty_range_iterate_clear() in that it has 5272 // imbedded in it a (subtype of) UpwardsObjectClosure. The 5273 // use of that state in the imbedded UpwardsObjectClosure instance 5274 // assumes that the cards are always iterated (even if in parallel 5275 // by several threads) in monotonically increasing order per each 5276 // thread. This is true of the implementation below which picks 5277 // card ranges (chunks) in monotonically increasing order globally 5278 // and, a-fortiori, in monotonically increasing order per thread 5279 // (the latter order being a subsequence of the former). 5280 // If the work code below is ever reorganized into a more chaotic 5281 // work-partitioning form than the current "sequential tasks" 5282 // paradigm, the use of that persistent state will have to be 5283 // revisited and modified appropriately. See also related 5284 // bug 4756801 work on which should examine this code to make 5285 // sure that the changes there do not run counter to the 5286 // assumptions made here and necessary for correctness and 5287 // efficiency. Note also that this code might yield inefficient 5288 // behaviour in the case of very large objects that span one or 5289 // more work chunks. Such objects would potentially be scanned 5290 // several times redundantly. Work on 4756801 should try and 5291 // address that performance anomaly if at all possible. XXX 5292 MemRegion full_span = _collector->_span; 5293 CMSBitMap* bm = &(_collector->_markBitMap); // shared 5294 CMSMarkStack* rs = &(_collector->_revisitStack); // shared 5295 MarkFromDirtyCardsClosure 5296 greyRescanClosure(_collector, full_span, // entire span of interest 5297 sp, bm, work_q, rs, cl); 5298 5299 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 5300 assert(pst->valid(), "Uninitialized use?"); 5301 int nth_task = 0; 5302 const int alignment = CardTableModRefBS::card_size * BitsPerWord; 5303 MemRegion span = sp->used_region(); 5304 HeapWord* start_addr = span.start(); 5305 HeapWord* end_addr = (HeapWord*)round_to((intptr_t)span.end(), 5306 alignment); 5307 const size_t chunk_size = sp->rescan_task_size(); // in HeapWord units 5308 assert((HeapWord*)round_to((intptr_t)start_addr, alignment) == 5309 start_addr, "Check alignment"); 5310 assert((size_t)round_to((intptr_t)chunk_size, alignment) == 5311 chunk_size, "Check alignment"); 5312 5313 while (!pst->is_task_claimed(/* reference */ nth_task)) { 5314 // Having claimed the nth_task, compute corresponding mem-region, 5315 // which is a-fortiori aligned correctly (i.e. at a MUT bopundary). 5316 // The alignment restriction ensures that we do not need any 5317 // synchronization with other gang-workers while setting or 5318 // clearing bits in thus chunk of the MUT. 5319 MemRegion this_span = MemRegion(start_addr + nth_task*chunk_size, 5320 start_addr + (nth_task+1)*chunk_size); 5321 // The last chunk's end might be way beyond end of the 5322 // used region. In that case pull back appropriately. 5323 if (this_span.end() > end_addr) { 5324 this_span.set_end(end_addr); 5325 assert(!this_span.is_empty(), "Program logic (calculation of n_tasks)"); 5326 } 5327 // Iterate over the dirty cards covering this chunk, marking them 5328 // precleaned, and setting the corresponding bits in the mod union 5329 // table. Since we have been careful to partition at Card and MUT-word 5330 // boundaries no synchronization is needed between parallel threads. 5331 _collector->_ct->ct_bs()->dirty_card_iterate(this_span, 5332 &modUnionClosure); 5333 5334 // Having transferred these marks into the modUnionTable, 5335 // rescan the marked objects on the dirty cards in the modUnionTable. 5336 // Even if this is at a synchronous collection, the initial marking 5337 // may have been done during an asynchronous collection so there 5338 // may be dirty bits in the mod-union table. 5339 _collector->_modUnionTable.dirty_range_iterate_clear( 5340 this_span, &greyRescanClosure); 5341 _collector->_modUnionTable.verifyNoOneBitsInRange( 5342 this_span.start(), 5343 this_span.end()); 5344 } 5345 pst->all_tasks_completed(); // declare that i am done 5346 } 5347 5348 // . see if we can share work_queues with ParNew? XXX 5349 void 5350 CMSParRemarkTask::do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, 5351 int* seed) { 5352 OopTaskQueue* work_q = work_queue(i); 5353 NOT_PRODUCT(int num_steals = 0;) 5354 oop obj_to_scan; 5355 CMSBitMap* bm = &(_collector->_markBitMap); 5356 5357 while (true) { 5358 // Completely finish any left over work from (an) earlier round(s) 5359 cl->trim_queue(0); 5360 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 5361 (size_t)ParGCDesiredObjsFromOverflowList); 5362 // Now check if there's any work in the overflow list 5363 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads, 5364 // only affects the number of attempts made to get work from the 5365 // overflow list and does not affect the number of workers. Just 5366 // pass ParallelGCThreads so this behavior is unchanged. 5367 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 5368 work_q, 5369 ParallelGCThreads)) { 5370 // found something in global overflow list; 5371 // not yet ready to go stealing work from others. 5372 // We'd like to assert(work_q->size() != 0, ...) 5373 // because we just took work from the overflow list, 5374 // but of course we can't since all of that could have 5375 // been already stolen from us. 5376 // "He giveth and He taketh away." 5377 continue; 5378 } 5379 // Verify that we have no work before we resort to stealing 5380 assert(work_q->size() == 0, "Have work, shouldn't steal"); 5381 // Try to steal from other queues that have work 5382 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 5383 NOT_PRODUCT(num_steals++;) 5384 assert(obj_to_scan->is_oop(), "Oops, not an oop!"); 5385 assert(bm->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 5386 // Do scanning work 5387 obj_to_scan->oop_iterate(cl); 5388 // Loop around, finish this work, and try to steal some more 5389 } else if (terminator()->offer_termination()) { 5390 break; // nirvana from the infinite cycle 5391 } 5392 } 5393 NOT_PRODUCT( 5394 if (PrintCMSStatistics != 0) { 5395 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals); 5396 } 5397 ) 5398 assert(work_q->size() == 0 && _collector->overflow_list_is_empty(), 5399 "Else our work is not yet done"); 5400 } 5401 5402 // Return a thread-local PLAB recording array, as appropriate. 5403 void* CMSCollector::get_data_recorder(int thr_num) { 5404 if (_survivor_plab_array != NULL && 5405 (CMSPLABRecordAlways || 5406 (_collectorState > Marking && _collectorState < FinalMarking))) { 5407 assert(thr_num < (int)ParallelGCThreads, "thr_num is out of bounds"); 5408 ChunkArray* ca = &_survivor_plab_array[thr_num]; 5409 ca->reset(); // clear it so that fresh data is recorded 5410 return (void*) ca; 5411 } else { 5412 return NULL; 5413 } 5414 } 5415 5416 // Reset all the thread-local PLAB recording arrays 5417 void CMSCollector::reset_survivor_plab_arrays() { 5418 for (uint i = 0; i < ParallelGCThreads; i++) { 5419 _survivor_plab_array[i].reset(); 5420 } 5421 } 5422 5423 // Merge the per-thread plab arrays into the global survivor chunk 5424 // array which will provide the partitioning of the survivor space 5425 // for CMS rescan. 5426 void CMSCollector::merge_survivor_plab_arrays(ContiguousSpace* surv, 5427 int no_of_gc_threads) { 5428 assert(_survivor_plab_array != NULL, "Error"); 5429 assert(_survivor_chunk_array != NULL, "Error"); 5430 assert(_collectorState == FinalMarking, "Error"); 5431 for (int j = 0; j < no_of_gc_threads; j++) { 5432 _cursor[j] = 0; 5433 } 5434 HeapWord* top = surv->top(); 5435 size_t i; 5436 for (i = 0; i < _survivor_chunk_capacity; i++) { // all sca entries 5437 HeapWord* min_val = top; // Higher than any PLAB address 5438 uint min_tid = 0; // position of min_val this round 5439 for (int j = 0; j < no_of_gc_threads; j++) { 5440 ChunkArray* cur_sca = &_survivor_plab_array[j]; 5441 if (_cursor[j] == cur_sca->end()) { 5442 continue; 5443 } 5444 assert(_cursor[j] < cur_sca->end(), "ctl pt invariant"); 5445 HeapWord* cur_val = cur_sca->nth(_cursor[j]); 5446 assert(surv->used_region().contains(cur_val), "Out of bounds value"); 5447 if (cur_val < min_val) { 5448 min_tid = j; 5449 min_val = cur_val; 5450 } else { 5451 assert(cur_val < top, "All recorded addresses should be less"); 5452 } 5453 } 5454 // At this point min_val and min_tid are respectively 5455 // the least address in _survivor_plab_array[j]->nth(_cursor[j]) 5456 // and the thread (j) that witnesses that address. 5457 // We record this address in the _survivor_chunk_array[i] 5458 // and increment _cursor[min_tid] prior to the next round i. 5459 if (min_val == top) { 5460 break; 5461 } 5462 _survivor_chunk_array[i] = min_val; 5463 _cursor[min_tid]++; 5464 } 5465 // We are all done; record the size of the _survivor_chunk_array 5466 _survivor_chunk_index = i; // exclusive: [0, i) 5467 if (PrintCMSStatistics > 0) { 5468 gclog_or_tty->print(" (Survivor:" SIZE_FORMAT "chunks) ", i); 5469 } 5470 // Verify that we used up all the recorded entries 5471 #ifdef ASSERT 5472 size_t total = 0; 5473 for (int j = 0; j < no_of_gc_threads; j++) { 5474 assert(_cursor[j] == _survivor_plab_array[j].end(), "Ctl pt invariant"); 5475 total += _cursor[j]; 5476 } 5477 assert(total == _survivor_chunk_index, "Ctl Pt Invariant"); 5478 // Check that the merged array is in sorted order 5479 if (total > 0) { 5480 for (size_t i = 0; i < total - 1; i++) { 5481 if (PrintCMSStatistics > 0) { 5482 gclog_or_tty->print(" (chunk" SIZE_FORMAT ":" INTPTR_FORMAT ") ", 5483 i, _survivor_chunk_array[i]); 5484 } 5485 assert(_survivor_chunk_array[i] < _survivor_chunk_array[i+1], 5486 "Not sorted"); 5487 } 5488 } 5489 #endif // ASSERT 5490 } 5491 5492 // Set up the space's par_seq_tasks structure for work claiming 5493 // for parallel rescan of young gen. 5494 // See ParRescanTask where this is currently used. 5495 void 5496 CMSCollector:: 5497 initialize_sequential_subtasks_for_young_gen_rescan(int n_threads) { 5498 assert(n_threads > 0, "Unexpected n_threads argument"); 5499 DefNewGeneration* dng = (DefNewGeneration*)_young_gen; 5500 5501 // Eden space 5502 { 5503 SequentialSubTasksDone* pst = dng->eden()->par_seq_tasks(); 5504 assert(!pst->valid(), "Clobbering existing data?"); 5505 // Each valid entry in [0, _eden_chunk_index) represents a task. 5506 size_t n_tasks = _eden_chunk_index + 1; 5507 assert(n_tasks == 1 || _eden_chunk_array != NULL, "Error"); 5508 // Sets the condition for completion of the subtask (how many threads 5509 // need to finish in order to be done). 5510 pst->set_n_threads(n_threads); 5511 pst->set_n_tasks((int)n_tasks); 5512 } 5513 5514 // Merge the survivor plab arrays into _survivor_chunk_array 5515 if (_survivor_plab_array != NULL) { 5516 merge_survivor_plab_arrays(dng->from(), n_threads); 5517 } else { 5518 assert(_survivor_chunk_index == 0, "Error"); 5519 } 5520 5521 // To space 5522 { 5523 SequentialSubTasksDone* pst = dng->to()->par_seq_tasks(); 5524 assert(!pst->valid(), "Clobbering existing data?"); 5525 // Sets the condition for completion of the subtask (how many threads 5526 // need to finish in order to be done). 5527 pst->set_n_threads(n_threads); 5528 pst->set_n_tasks(1); 5529 assert(pst->valid(), "Error"); 5530 } 5531 5532 // From space 5533 { 5534 SequentialSubTasksDone* pst = dng->from()->par_seq_tasks(); 5535 assert(!pst->valid(), "Clobbering existing data?"); 5536 size_t n_tasks = _survivor_chunk_index + 1; 5537 assert(n_tasks == 1 || _survivor_chunk_array != NULL, "Error"); 5538 // Sets the condition for completion of the subtask (how many threads 5539 // need to finish in order to be done). 5540 pst->set_n_threads(n_threads); 5541 pst->set_n_tasks((int)n_tasks); 5542 assert(pst->valid(), "Error"); 5543 } 5544 } 5545 5546 // Parallel version of remark 5547 void CMSCollector::do_remark_parallel() { 5548 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5549 FlexibleWorkGang* workers = gch->workers(); 5550 assert(workers != NULL, "Need parallel worker threads."); 5551 int n_workers = workers->total_workers(); 5552 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 5553 CompactibleFreeListSpace* perm_space = _permGen->cmsSpace(); 5554 5555 CMSParRemarkTask tsk(this, 5556 cms_space, perm_space, 5557 n_workers, workers, task_queues()); 5558 5559 // Set up for parallel process_strong_roots work. 5560 gch->set_par_threads(n_workers); 5561 // We won't be iterating over the cards in the card table updating 5562 // the younger_gen cards, so we shouldn't call the following else 5563 // the verification code as well as subsequent younger_refs_iterate 5564 // code would get confused. XXX 5565 // gch->rem_set()->prepare_for_younger_refs_iterate(true); // parallel 5566 5567 // The young gen rescan work will not be done as part of 5568 // process_strong_roots (which currently doesn't knw how to 5569 // parallelize such a scan), but rather will be broken up into 5570 // a set of parallel tasks (via the sampling that the [abortable] 5571 // preclean phase did of EdenSpace, plus the [two] tasks of 5572 // scanning the [two] survivor spaces. Further fine-grain 5573 // parallelization of the scanning of the survivor spaces 5574 // themselves, and of precleaning of the younger gen itself 5575 // is deferred to the future. 5576 initialize_sequential_subtasks_for_young_gen_rescan(n_workers); 5577 5578 // The dirty card rescan work is broken up into a "sequence" 5579 // of parallel tasks (per constituent space) that are dynamically 5580 // claimed by the parallel threads. 5581 cms_space->initialize_sequential_subtasks_for_rescan(n_workers); 5582 perm_space->initialize_sequential_subtasks_for_rescan(n_workers); 5583 5584 // It turns out that even when we're using 1 thread, doing the work in a 5585 // separate thread causes wide variance in run times. We can't help this 5586 // in the multi-threaded case, but we special-case n=1 here to get 5587 // repeatable measurements of the 1-thread overhead of the parallel code. 5588 if (n_workers > 1) { 5589 // Make refs discovery MT-safe 5590 ReferenceProcessorMTMutator mt(ref_processor(), true); 5591 GenCollectedHeap::StrongRootsScope srs(gch); 5592 workers->run_task(&tsk); 5593 } else { 5594 GenCollectedHeap::StrongRootsScope srs(gch); 5595 tsk.work(0); 5596 } 5597 gch->set_par_threads(0); // 0 ==> non-parallel. 5598 // restore, single-threaded for now, any preserved marks 5599 // as a result of work_q overflow 5600 restore_preserved_marks_if_any(); 5601 } 5602 5603 // Non-parallel version of remark 5604 void CMSCollector::do_remark_non_parallel() { 5605 ResourceMark rm; 5606 HandleMark hm; 5607 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5608 MarkRefsIntoAndScanClosure 5609 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable, 5610 &_markStack, &_revisitStack, this, 5611 false /* should_yield */, false /* not precleaning */); 5612 MarkFromDirtyCardsClosure 5613 markFromDirtyCardsClosure(this, _span, 5614 NULL, // space is set further below 5615 &_markBitMap, &_markStack, &_revisitStack, 5616 &mrias_cl); 5617 { 5618 TraceTime t("grey object rescan", PrintGCDetails, false, gclog_or_tty); 5619 // Iterate over the dirty cards, setting the corresponding bits in the 5620 // mod union table. 5621 { 5622 ModUnionClosure modUnionClosure(&_modUnionTable); 5623 _ct->ct_bs()->dirty_card_iterate( 5624 _cmsGen->used_region(), 5625 &modUnionClosure); 5626 _ct->ct_bs()->dirty_card_iterate( 5627 _permGen->used_region(), 5628 &modUnionClosure); 5629 } 5630 // Having transferred these marks into the modUnionTable, we just need 5631 // to rescan the marked objects on the dirty cards in the modUnionTable. 5632 // The initial marking may have been done during an asynchronous 5633 // collection so there may be dirty bits in the mod-union table. 5634 const int alignment = 5635 CardTableModRefBS::card_size * BitsPerWord; 5636 { 5637 // ... First handle dirty cards in CMS gen 5638 markFromDirtyCardsClosure.set_space(_cmsGen->cmsSpace()); 5639 MemRegion ur = _cmsGen->used_region(); 5640 HeapWord* lb = ur.start(); 5641 HeapWord* ub = (HeapWord*)round_to((intptr_t)ur.end(), alignment); 5642 MemRegion cms_span(lb, ub); 5643 _modUnionTable.dirty_range_iterate_clear(cms_span, 5644 &markFromDirtyCardsClosure); 5645 verify_work_stacks_empty(); 5646 if (PrintCMSStatistics != 0) { 5647 gclog_or_tty->print(" (re-scanned "SIZE_FORMAT" dirty cards in cms gen) ", 5648 markFromDirtyCardsClosure.num_dirty_cards()); 5649 } 5650 } 5651 { 5652 // .. and then repeat for dirty cards in perm gen 5653 markFromDirtyCardsClosure.set_space(_permGen->cmsSpace()); 5654 MemRegion ur = _permGen->used_region(); 5655 HeapWord* lb = ur.start(); 5656 HeapWord* ub = (HeapWord*)round_to((intptr_t)ur.end(), alignment); 5657 MemRegion perm_span(lb, ub); 5658 _modUnionTable.dirty_range_iterate_clear(perm_span, 5659 &markFromDirtyCardsClosure); 5660 verify_work_stacks_empty(); 5661 if (PrintCMSStatistics != 0) { 5662 gclog_or_tty->print(" (re-scanned "SIZE_FORMAT" dirty cards in perm gen) ", 5663 markFromDirtyCardsClosure.num_dirty_cards()); 5664 } 5665 } 5666 } 5667 if (VerifyDuringGC && 5668 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 5669 HandleMark hm; // Discard invalid handles created during verification 5670 Universe::verify(true); 5671 } 5672 { 5673 TraceTime t("root rescan", PrintGCDetails, false, gclog_or_tty); 5674 5675 verify_work_stacks_empty(); 5676 5677 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 5678 GenCollectedHeap::StrongRootsScope srs(gch); 5679 gch->gen_process_strong_roots(_cmsGen->level(), 5680 true, // younger gens as roots 5681 false, // use the local StrongRootsScope 5682 true, // collecting perm gen 5683 SharedHeap::ScanningOption(roots_scanning_options()), 5684 &mrias_cl, 5685 true, // walk code active on stacks 5686 NULL); 5687 assert(should_unload_classes() 5688 || (roots_scanning_options() & SharedHeap::SO_CodeCache), 5689 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 5690 } 5691 verify_work_stacks_empty(); 5692 // Restore evacuated mark words, if any, used for overflow list links 5693 if (!CMSOverflowEarlyRestoration) { 5694 restore_preserved_marks_if_any(); 5695 } 5696 verify_overflow_empty(); 5697 } 5698 5699 //////////////////////////////////////////////////////// 5700 // Parallel Reference Processing Task Proxy Class 5701 //////////////////////////////////////////////////////// 5702 class CMSRefProcTaskProxy: public AbstractGangTaskWOopQueues { 5703 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask; 5704 CMSCollector* _collector; 5705 CMSBitMap* _mark_bit_map; 5706 const MemRegion _span; 5707 ProcessTask& _task; 5708 5709 public: 5710 CMSRefProcTaskProxy(ProcessTask& task, 5711 CMSCollector* collector, 5712 const MemRegion& span, 5713 CMSBitMap* mark_bit_map, 5714 AbstractWorkGang* workers, 5715 OopTaskQueueSet* task_queues): 5716 AbstractGangTaskWOopQueues("Process referents by policy in parallel", 5717 task_queues), 5718 _task(task), 5719 _collector(collector), _span(span), _mark_bit_map(mark_bit_map) 5720 { 5721 assert(_collector->_span.equals(_span) && !_span.is_empty(), 5722 "Inconsistency in _span"); 5723 } 5724 5725 OopTaskQueueSet* task_queues() { return queues(); } 5726 5727 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 5728 5729 void do_work_steal(int i, 5730 CMSParDrainMarkingStackClosure* drain, 5731 CMSParKeepAliveClosure* keep_alive, 5732 int* seed); 5733 5734 virtual void work(int i); 5735 }; 5736 5737 void CMSRefProcTaskProxy::work(int i) { 5738 assert(_collector->_span.equals(_span), "Inconsistency in _span"); 5739 CMSParKeepAliveClosure par_keep_alive(_collector, _span, 5740 _mark_bit_map, 5741 &_collector->_revisitStack, 5742 work_queue(i)); 5743 CMSParDrainMarkingStackClosure par_drain_stack(_collector, _span, 5744 _mark_bit_map, 5745 &_collector->_revisitStack, 5746 work_queue(i)); 5747 CMSIsAliveClosure is_alive_closure(_span, _mark_bit_map); 5748 _task.work(i, is_alive_closure, par_keep_alive, par_drain_stack); 5749 if (_task.marks_oops_alive()) { 5750 do_work_steal(i, &par_drain_stack, &par_keep_alive, 5751 _collector->hash_seed(i)); 5752 } 5753 assert(work_queue(i)->size() == 0, "work_queue should be empty"); 5754 assert(_collector->_overflow_list == NULL, "non-empty _overflow_list"); 5755 } 5756 5757 class CMSRefEnqueueTaskProxy: public AbstractGangTask { 5758 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask; 5759 EnqueueTask& _task; 5760 5761 public: 5762 CMSRefEnqueueTaskProxy(EnqueueTask& task) 5763 : AbstractGangTask("Enqueue reference objects in parallel"), 5764 _task(task) 5765 { } 5766 5767 virtual void work(int i) 5768 { 5769 _task.work(i); 5770 } 5771 }; 5772 5773 CMSParKeepAliveClosure::CMSParKeepAliveClosure(CMSCollector* collector, 5774 MemRegion span, CMSBitMap* bit_map, CMSMarkStack* revisit_stack, 5775 OopTaskQueue* work_queue): 5776 Par_KlassRememberingOopClosure(collector, NULL, revisit_stack), 5777 _span(span), 5778 _bit_map(bit_map), 5779 _work_queue(work_queue), 5780 _mark_and_push(collector, span, bit_map, revisit_stack, work_queue), 5781 _low_water_mark(MIN2((uint)(work_queue->max_elems()/4), 5782 (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads))) 5783 { } 5784 5785 // . see if we can share work_queues with ParNew? XXX 5786 void CMSRefProcTaskProxy::do_work_steal(int i, 5787 CMSParDrainMarkingStackClosure* drain, 5788 CMSParKeepAliveClosure* keep_alive, 5789 int* seed) { 5790 OopTaskQueue* work_q = work_queue(i); 5791 NOT_PRODUCT(int num_steals = 0;) 5792 oop obj_to_scan; 5793 5794 while (true) { 5795 // Completely finish any left over work from (an) earlier round(s) 5796 drain->trim_queue(0); 5797 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 5798 (size_t)ParGCDesiredObjsFromOverflowList); 5799 // Now check if there's any work in the overflow list 5800 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads, 5801 // only affects the number of attempts made to get work from the 5802 // overflow list and does not affect the number of workers. Just 5803 // pass ParallelGCThreads so this behavior is unchanged. 5804 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 5805 work_q, 5806 ParallelGCThreads)) { 5807 // Found something in global overflow list; 5808 // not yet ready to go stealing work from others. 5809 // We'd like to assert(work_q->size() != 0, ...) 5810 // because we just took work from the overflow list, 5811 // but of course we can't, since all of that might have 5812 // been already stolen from us. 5813 continue; 5814 } 5815 // Verify that we have no work before we resort to stealing 5816 assert(work_q->size() == 0, "Have work, shouldn't steal"); 5817 // Try to steal from other queues that have work 5818 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 5819 NOT_PRODUCT(num_steals++;) 5820 assert(obj_to_scan->is_oop(), "Oops, not an oop!"); 5821 assert(_mark_bit_map->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 5822 // Do scanning work 5823 obj_to_scan->oop_iterate(keep_alive); 5824 // Loop around, finish this work, and try to steal some more 5825 } else if (terminator()->offer_termination()) { 5826 break; // nirvana from the infinite cycle 5827 } 5828 } 5829 NOT_PRODUCT( 5830 if (PrintCMSStatistics != 0) { 5831 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals); 5832 } 5833 ) 5834 } 5835 5836 void CMSRefProcTaskExecutor::execute(ProcessTask& task) 5837 { 5838 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5839 FlexibleWorkGang* workers = gch->workers(); 5840 assert(workers != NULL, "Need parallel worker threads."); 5841 CMSRefProcTaskProxy rp_task(task, &_collector, 5842 _collector.ref_processor()->span(), 5843 _collector.markBitMap(), 5844 workers, _collector.task_queues()); 5845 workers->run_task(&rp_task); 5846 } 5847 5848 void CMSRefProcTaskExecutor::execute(EnqueueTask& task) 5849 { 5850 5851 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5852 FlexibleWorkGang* workers = gch->workers(); 5853 assert(workers != NULL, "Need parallel worker threads."); 5854 CMSRefEnqueueTaskProxy enq_task(task); 5855 workers->run_task(&enq_task); 5856 } 5857 5858 void CMSCollector::refProcessingWork(bool asynch, bool clear_all_soft_refs) { 5859 5860 ResourceMark rm; 5861 HandleMark hm; 5862 5863 ReferenceProcessor* rp = ref_processor(); 5864 assert(rp->span().equals(_span), "Spans should be equal"); 5865 assert(!rp->enqueuing_is_done(), "Enqueuing should not be complete"); 5866 // Process weak references. 5867 rp->setup_policy(clear_all_soft_refs); 5868 verify_work_stacks_empty(); 5869 5870 CMSKeepAliveClosure cmsKeepAliveClosure(this, _span, &_markBitMap, 5871 &_markStack, &_revisitStack, 5872 false /* !preclean */); 5873 CMSDrainMarkingStackClosure cmsDrainMarkingStackClosure(this, 5874 _span, &_markBitMap, &_markStack, 5875 &cmsKeepAliveClosure, false /* !preclean */); 5876 { 5877 TraceTime t("weak refs processing", PrintGCDetails, false, gclog_or_tty); 5878 if (rp->processing_is_mt()) { 5879 // Set the degree of MT here. If the discovery is done MT, there 5880 // may have been a different number of threads doing the discovery 5881 // and a different number of discovered lists may have Ref objects. 5882 // That is OK as long as the Reference lists are balanced (see 5883 // balance_all_queues() and balance_queues()). 5884 5885 5886 rp->set_mt_degree(ParallelGCThreads); 5887 CMSRefProcTaskExecutor task_executor(*this); 5888 rp->process_discovered_references(&_is_alive_closure, 5889 &cmsKeepAliveClosure, 5890 &cmsDrainMarkingStackClosure, 5891 &task_executor); 5892 } else { 5893 rp->process_discovered_references(&_is_alive_closure, 5894 &cmsKeepAliveClosure, 5895 &cmsDrainMarkingStackClosure, 5896 NULL); 5897 } 5898 verify_work_stacks_empty(); 5899 } 5900 5901 if (should_unload_classes()) { 5902 { 5903 TraceTime t("class unloading", PrintGCDetails, false, gclog_or_tty); 5904 5905 // Follow SystemDictionary roots and unload classes 5906 bool purged_class = SystemDictionary::do_unloading(&_is_alive_closure); 5907 5908 // Follow CodeCache roots and unload any methods marked for unloading 5909 CodeCache::do_unloading(&_is_alive_closure, 5910 &cmsKeepAliveClosure, 5911 purged_class); 5912 5913 cmsDrainMarkingStackClosure.do_void(); 5914 verify_work_stacks_empty(); 5915 5916 // Update subklass/sibling/implementor links in KlassKlass descendants 5917 assert(!_revisitStack.isEmpty(), "revisit stack should not be empty"); 5918 oop k; 5919 while ((k = _revisitStack.pop()) != NULL) { 5920 ((Klass*)(oopDesc*)k)->follow_weak_klass_links( 5921 &_is_alive_closure, 5922 &cmsKeepAliveClosure); 5923 } 5924 assert(!ClassUnloading || 5925 (_markStack.isEmpty() && overflow_list_is_empty()), 5926 "Should not have found new reachable objects"); 5927 assert(_revisitStack.isEmpty(), "revisit stack should have been drained"); 5928 cmsDrainMarkingStackClosure.do_void(); 5929 verify_work_stacks_empty(); 5930 } 5931 5932 { 5933 TraceTime t("scrub symbol & string tables", PrintGCDetails, false, gclog_or_tty); 5934 // Now clean up stale oops in SymbolTable and StringTable 5935 SymbolTable::unlink(&_is_alive_closure); 5936 StringTable::unlink(&_is_alive_closure); 5937 } 5938 } 5939 5940 verify_work_stacks_empty(); 5941 // Restore any preserved marks as a result of mark stack or 5942 // work queue overflow 5943 restore_preserved_marks_if_any(); // done single-threaded for now 5944 5945 rp->set_enqueuing_is_done(true); 5946 if (rp->processing_is_mt()) { 5947 rp->balance_all_queues(); 5948 CMSRefProcTaskExecutor task_executor(*this); 5949 rp->enqueue_discovered_references(&task_executor); 5950 } else { 5951 rp->enqueue_discovered_references(NULL); 5952 } 5953 rp->verify_no_references_recorded(); 5954 assert(!rp->discovery_enabled(), "should have been disabled"); 5955 5956 // JVMTI object tagging is based on JNI weak refs. If any of these 5957 // refs were cleared then JVMTI needs to update its maps and 5958 // maybe post ObjectFrees to agents. 5959 JvmtiExport::cms_ref_processing_epilogue(); 5960 } 5961 5962 #ifndef PRODUCT 5963 void CMSCollector::check_correct_thread_executing() { 5964 Thread* t = Thread::current(); 5965 // Only the VM thread or the CMS thread should be here. 5966 assert(t->is_ConcurrentGC_thread() || t->is_VM_thread(), 5967 "Unexpected thread type"); 5968 // If this is the vm thread, the foreground process 5969 // should not be waiting. Note that _foregroundGCIsActive is 5970 // true while the foreground collector is waiting. 5971 if (_foregroundGCShouldWait) { 5972 // We cannot be the VM thread 5973 assert(t->is_ConcurrentGC_thread(), 5974 "Should be CMS thread"); 5975 } else { 5976 // We can be the CMS thread only if we are in a stop-world 5977 // phase of CMS collection. 5978 if (t->is_ConcurrentGC_thread()) { 5979 assert(_collectorState == InitialMarking || 5980 _collectorState == FinalMarking, 5981 "Should be a stop-world phase"); 5982 // The CMS thread should be holding the CMS_token. 5983 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 5984 "Potential interference with concurrently " 5985 "executing VM thread"); 5986 } 5987 } 5988 } 5989 #endif 5990 5991 void CMSCollector::sweep(bool asynch) { 5992 assert(_collectorState == Sweeping, "just checking"); 5993 check_correct_thread_executing(); 5994 verify_work_stacks_empty(); 5995 verify_overflow_empty(); 5996 increment_sweep_count(); 5997 TraceCMSMemoryManagerStats tms(_collectorState); 5998 5999 _inter_sweep_timer.stop(); 6000 _inter_sweep_estimate.sample(_inter_sweep_timer.seconds()); 6001 size_policy()->avg_cms_free_at_sweep()->sample(_cmsGen->free()); 6002 6003 // PermGen verification support: If perm gen sweeping is disabled in 6004 // this cycle, we preserve the perm gen object "deadness" information 6005 // in the perm_gen_verify_bit_map. In order to do that we traverse 6006 // all blocks in perm gen and mark all dead objects. 6007 if (verifying() && !should_unload_classes()) { 6008 assert(perm_gen_verify_bit_map()->sizeInBits() != 0, 6009 "Should have already been allocated"); 6010 MarkDeadObjectsClosure mdo(this, _permGen->cmsSpace(), 6011 markBitMap(), perm_gen_verify_bit_map()); 6012 if (asynch) { 6013 CMSTokenSyncWithLocks ts(true, _permGen->freelistLock(), 6014 bitMapLock()); 6015 _permGen->cmsSpace()->blk_iterate(&mdo); 6016 } else { 6017 // In the case of synchronous sweep, we already have 6018 // the requisite locks/tokens. 6019 _permGen->cmsSpace()->blk_iterate(&mdo); 6020 } 6021 } 6022 6023 assert(!_intra_sweep_timer.is_active(), "Should not be active"); 6024 _intra_sweep_timer.reset(); 6025 _intra_sweep_timer.start(); 6026 if (asynch) { 6027 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 6028 CMSPhaseAccounting pa(this, "sweep", !PrintGCDetails); 6029 // First sweep the old gen then the perm gen 6030 { 6031 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(), 6032 bitMapLock()); 6033 sweepWork(_cmsGen, asynch); 6034 } 6035 6036 // Now repeat for perm gen 6037 if (should_unload_classes()) { 6038 CMSTokenSyncWithLocks ts(true, _permGen->freelistLock(), 6039 bitMapLock()); 6040 sweepWork(_permGen, asynch); 6041 } 6042 6043 // Update Universe::_heap_*_at_gc figures. 6044 // We need all the free list locks to make the abstract state 6045 // transition from Sweeping to Resetting. See detailed note 6046 // further below. 6047 { 6048 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(), 6049 _permGen->freelistLock()); 6050 // Update heap occupancy information which is used as 6051 // input to soft ref clearing policy at the next gc. 6052 Universe::update_heap_info_at_gc(); 6053 _collectorState = Resizing; 6054 } 6055 } else { 6056 // already have needed locks 6057 sweepWork(_cmsGen, asynch); 6058 6059 if (should_unload_classes()) { 6060 sweepWork(_permGen, asynch); 6061 } 6062 // Update heap occupancy information which is used as 6063 // input to soft ref clearing policy at the next gc. 6064 Universe::update_heap_info_at_gc(); 6065 _collectorState = Resizing; 6066 } 6067 verify_work_stacks_empty(); 6068 verify_overflow_empty(); 6069 6070 _intra_sweep_timer.stop(); 6071 _intra_sweep_estimate.sample(_intra_sweep_timer.seconds()); 6072 6073 _inter_sweep_timer.reset(); 6074 _inter_sweep_timer.start(); 6075 6076 update_time_of_last_gc(os::javaTimeMillis()); 6077 6078 // NOTE on abstract state transitions: 6079 // Mutators allocate-live and/or mark the mod-union table dirty 6080 // based on the state of the collection. The former is done in 6081 // the interval [Marking, Sweeping] and the latter in the interval 6082 // [Marking, Sweeping). Thus the transitions into the Marking state 6083 // and out of the Sweeping state must be synchronously visible 6084 // globally to the mutators. 6085 // The transition into the Marking state happens with the world 6086 // stopped so the mutators will globally see it. Sweeping is 6087 // done asynchronously by the background collector so the transition 6088 // from the Sweeping state to the Resizing state must be done 6089 // under the freelistLock (as is the check for whether to 6090 // allocate-live and whether to dirty the mod-union table). 6091 assert(_collectorState == Resizing, "Change of collector state to" 6092 " Resizing must be done under the freelistLocks (plural)"); 6093 6094 // Now that sweeping has been completed, if the GCH's 6095 // incremental_collection_will_fail flag is set, clear it, 6096 // thus inviting a younger gen collection to promote into 6097 // this generation. If such a promotion may still fail, 6098 // the flag will be set again when a young collection is 6099 // attempted. 6100 // I think the incremental_collection_will_fail flag's use 6101 // is specific to a 2 generation collection policy, so i'll 6102 // assert that that's the configuration we are operating within. 6103 // The use of the flag can and should be generalized appropriately 6104 // in the future to deal with a general n-generation system. 6105 6106 GenCollectedHeap* gch = GenCollectedHeap::heap(); 6107 assert(gch->collector_policy()->is_two_generation_policy(), 6108 "Resetting of incremental_collection_will_fail flag" 6109 " may be incorrect otherwise"); 6110 gch->clear_incremental_collection_will_fail(); 6111 gch->update_full_collections_completed(_collection_count_start); 6112 } 6113 6114 // FIX ME!!! Looks like this belongs in CFLSpace, with 6115 // CMSGen merely delegating to it. 6116 void ConcurrentMarkSweepGeneration::setNearLargestChunk() { 6117 double nearLargestPercent = FLSLargestBlockCoalesceProximity; 6118 HeapWord* minAddr = _cmsSpace->bottom(); 6119 HeapWord* largestAddr = 6120 (HeapWord*) _cmsSpace->dictionary()->findLargestDict(); 6121 if (largestAddr == NULL) { 6122 // The dictionary appears to be empty. In this case 6123 // try to coalesce at the end of the heap. 6124 largestAddr = _cmsSpace->end(); 6125 } 6126 size_t largestOffset = pointer_delta(largestAddr, minAddr); 6127 size_t nearLargestOffset = 6128 (size_t)((double)largestOffset * nearLargestPercent) - MinChunkSize; 6129 if (PrintFLSStatistics != 0) { 6130 gclog_or_tty->print_cr( 6131 "CMS: Large Block: " PTR_FORMAT ";" 6132 " Proximity: " PTR_FORMAT " -> " PTR_FORMAT, 6133 largestAddr, 6134 _cmsSpace->nearLargestChunk(), minAddr + nearLargestOffset); 6135 } 6136 _cmsSpace->set_nearLargestChunk(minAddr + nearLargestOffset); 6137 } 6138 6139 bool ConcurrentMarkSweepGeneration::isNearLargestChunk(HeapWord* addr) { 6140 return addr >= _cmsSpace->nearLargestChunk(); 6141 } 6142 6143 FreeChunk* ConcurrentMarkSweepGeneration::find_chunk_at_end() { 6144 return _cmsSpace->find_chunk_at_end(); 6145 } 6146 6147 void ConcurrentMarkSweepGeneration::update_gc_stats(int current_level, 6148 bool full) { 6149 // The next lower level has been collected. Gather any statistics 6150 // that are of interest at this point. 6151 if (!full && (current_level + 1) == level()) { 6152 // Gather statistics on the young generation collection. 6153 collector()->stats().record_gc0_end(used()); 6154 } 6155 } 6156 6157 CMSAdaptiveSizePolicy* ConcurrentMarkSweepGeneration::size_policy() { 6158 GenCollectedHeap* gch = GenCollectedHeap::heap(); 6159 assert(gch->kind() == CollectedHeap::GenCollectedHeap, 6160 "Wrong type of heap"); 6161 CMSAdaptiveSizePolicy* sp = (CMSAdaptiveSizePolicy*) 6162 gch->gen_policy()->size_policy(); 6163 assert(sp->is_gc_cms_adaptive_size_policy(), 6164 "Wrong type of size policy"); 6165 return sp; 6166 } 6167 6168 void ConcurrentMarkSweepGeneration::rotate_debug_collection_type() { 6169 if (PrintGCDetails && Verbose) { 6170 gclog_or_tty->print("Rotate from %d ", _debug_collection_type); 6171 } 6172 _debug_collection_type = (CollectionTypes) (_debug_collection_type + 1); 6173 _debug_collection_type = 6174 (CollectionTypes) (_debug_collection_type % Unknown_collection_type); 6175 if (PrintGCDetails && Verbose) { 6176 gclog_or_tty->print_cr("to %d ", _debug_collection_type); 6177 } 6178 } 6179 6180 void CMSCollector::sweepWork(ConcurrentMarkSweepGeneration* gen, 6181 bool asynch) { 6182 // We iterate over the space(s) underlying this generation, 6183 // checking the mark bit map to see if the bits corresponding 6184 // to specific blocks are marked or not. Blocks that are 6185 // marked are live and are not swept up. All remaining blocks 6186 // are swept up, with coalescing on-the-fly as we sweep up 6187 // contiguous free and/or garbage blocks: 6188 // We need to ensure that the sweeper synchronizes with allocators 6189 // and stop-the-world collectors. In particular, the following 6190 // locks are used: 6191 // . CMS token: if this is held, a stop the world collection cannot occur 6192 // . freelistLock: if this is held no allocation can occur from this 6193 // generation by another thread 6194 // . bitMapLock: if this is held, no other thread can access or update 6195 // 6196 6197 // Note that we need to hold the freelistLock if we use 6198 // block iterate below; else the iterator might go awry if 6199 // a mutator (or promotion) causes block contents to change 6200 // (for instance if the allocator divvies up a block). 6201 // If we hold the free list lock, for all practical purposes 6202 // young generation GC's can't occur (they'll usually need to 6203 // promote), so we might as well prevent all young generation 6204 // GC's while we do a sweeping step. For the same reason, we might 6205 // as well take the bit map lock for the entire duration 6206 6207 // check that we hold the requisite locks 6208 assert(have_cms_token(), "Should hold cms token"); 6209 assert( (asynch && ConcurrentMarkSweepThread::cms_thread_has_cms_token()) 6210 || (!asynch && ConcurrentMarkSweepThread::vm_thread_has_cms_token()), 6211 "Should possess CMS token to sweep"); 6212 assert_lock_strong(gen->freelistLock()); 6213 assert_lock_strong(bitMapLock()); 6214 6215 assert(!_inter_sweep_timer.is_active(), "Was switched off in an outer context"); 6216 assert(_intra_sweep_timer.is_active(), "Was switched on in an outer context"); 6217 gen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()), 6218 _inter_sweep_estimate.padded_average(), 6219 _intra_sweep_estimate.padded_average()); 6220 gen->setNearLargestChunk(); 6221 6222 { 6223 SweepClosure sweepClosure(this, gen, &_markBitMap, 6224 CMSYield && asynch); 6225 gen->cmsSpace()->blk_iterate_careful(&sweepClosure); 6226 // We need to free-up/coalesce garbage/blocks from a 6227 // co-terminal free run. This is done in the SweepClosure 6228 // destructor; so, do not remove this scope, else the 6229 // end-of-sweep-census below will be off by a little bit. 6230 } 6231 gen->cmsSpace()->sweep_completed(); 6232 gen->cmsSpace()->endSweepFLCensus(sweep_count()); 6233 if (should_unload_classes()) { // unloaded classes this cycle, 6234 _concurrent_cycles_since_last_unload = 0; // ... reset count 6235 } else { // did not unload classes, 6236 _concurrent_cycles_since_last_unload++; // ... increment count 6237 } 6238 } 6239 6240 // Reset CMS data structures (for now just the marking bit map) 6241 // preparatory for the next cycle. 6242 void CMSCollector::reset(bool asynch) { 6243 GenCollectedHeap* gch = GenCollectedHeap::heap(); 6244 CMSAdaptiveSizePolicy* sp = size_policy(); 6245 AdaptiveSizePolicyOutput(sp, gch->total_collections()); 6246 if (asynch) { 6247 CMSTokenSyncWithLocks ts(true, bitMapLock()); 6248 6249 // If the state is not "Resetting", the foreground thread 6250 // has done a collection and the resetting. 6251 if (_collectorState != Resetting) { 6252 assert(_collectorState == Idling, "The state should only change" 6253 " because the foreground collector has finished the collection"); 6254 return; 6255 } 6256 6257 // Clear the mark bitmap (no grey objects to start with) 6258 // for the next cycle. 6259 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 6260 CMSPhaseAccounting cmspa(this, "reset", !PrintGCDetails); 6261 6262 HeapWord* curAddr = _markBitMap.startWord(); 6263 while (curAddr < _markBitMap.endWord()) { 6264 size_t remaining = pointer_delta(_markBitMap.endWord(), curAddr); 6265 MemRegion chunk(curAddr, MIN2(CMSBitMapYieldQuantum, remaining)); 6266 _markBitMap.clear_large_range(chunk); 6267 if (ConcurrentMarkSweepThread::should_yield() && 6268 !foregroundGCIsActive() && 6269 CMSYield) { 6270 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6271 "CMS thread should hold CMS token"); 6272 assert_lock_strong(bitMapLock()); 6273 bitMapLock()->unlock(); 6274 ConcurrentMarkSweepThread::desynchronize(true); 6275 ConcurrentMarkSweepThread::acknowledge_yield_request(); 6276 stopTimer(); 6277 if (PrintCMSStatistics != 0) { 6278 incrementYields(); 6279 } 6280 icms_wait(); 6281 6282 // See the comment in coordinator_yield() 6283 for (unsigned i = 0; i < CMSYieldSleepCount && 6284 ConcurrentMarkSweepThread::should_yield() && 6285 !CMSCollector::foregroundGCIsActive(); ++i) { 6286 os::sleep(Thread::current(), 1, false); 6287 ConcurrentMarkSweepThread::acknowledge_yield_request(); 6288 } 6289 6290 ConcurrentMarkSweepThread::synchronize(true); 6291 bitMapLock()->lock_without_safepoint_check(); 6292 startTimer(); 6293 } 6294 curAddr = chunk.end(); 6295 } 6296 // A successful mostly concurrent collection has been done. 6297 // Because only the full (i.e., concurrent mode failure) collections 6298 // are being measured for gc overhead limits, clean the "near" flag 6299 // and count. 6300 sp->reset_gc_overhead_limit_count(); 6301 _collectorState = Idling; 6302 } else { 6303 // already have the lock 6304 assert(_collectorState == Resetting, "just checking"); 6305 assert_lock_strong(bitMapLock()); 6306 _markBitMap.clear_all(); 6307 _collectorState = Idling; 6308 } 6309 6310 // Stop incremental mode after a cycle completes, so that any future cycles 6311 // are triggered by allocation. 6312 stop_icms(); 6313 6314 NOT_PRODUCT( 6315 if (RotateCMSCollectionTypes) { 6316 _cmsGen->rotate_debug_collection_type(); 6317 } 6318 ) 6319 } 6320 6321 void CMSCollector::do_CMS_operation(CMS_op_type op) { 6322 gclog_or_tty->date_stamp(PrintGC && PrintGCDateStamps); 6323 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 6324 TraceTime t("GC", PrintGC, !PrintGCDetails, gclog_or_tty); 6325 TraceCollectorStats tcs(counters()); 6326 6327 switch (op) { 6328 case CMS_op_checkpointRootsInitial: { 6329 checkpointRootsInitial(true); // asynch 6330 if (PrintGC) { 6331 _cmsGen->printOccupancy("initial-mark"); 6332 } 6333 break; 6334 } 6335 case CMS_op_checkpointRootsFinal: { 6336 checkpointRootsFinal(true, // asynch 6337 false, // !clear_all_soft_refs 6338 false); // !init_mark_was_synchronous 6339 if (PrintGC) { 6340 _cmsGen->printOccupancy("remark"); 6341 } 6342 break; 6343 } 6344 default: 6345 fatal("No such CMS_op"); 6346 } 6347 } 6348 6349 #ifndef PRODUCT 6350 size_t const CMSCollector::skip_header_HeapWords() { 6351 return FreeChunk::header_size(); 6352 } 6353 6354 // Try and collect here conditions that should hold when 6355 // CMS thread is exiting. The idea is that the foreground GC 6356 // thread should not be blocked if it wants to terminate 6357 // the CMS thread and yet continue to run the VM for a while 6358 // after that. 6359 void CMSCollector::verify_ok_to_terminate() const { 6360 assert(Thread::current()->is_ConcurrentGC_thread(), 6361 "should be called by CMS thread"); 6362 assert(!_foregroundGCShouldWait, "should be false"); 6363 // We could check here that all the various low-level locks 6364 // are not held by the CMS thread, but that is overkill; see 6365 // also CMSThread::verify_ok_to_terminate() where the CGC_lock 6366 // is checked. 6367 } 6368 #endif 6369 6370 size_t CMSCollector::block_size_using_printezis_bits(HeapWord* addr) const { 6371 assert(_markBitMap.isMarked(addr) && _markBitMap.isMarked(addr + 1), 6372 "missing Printezis mark?"); 6373 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 6374 size_t size = pointer_delta(nextOneAddr + 1, addr); 6375 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 6376 "alignment problem"); 6377 assert(size >= 3, "Necessary for Printezis marks to work"); 6378 return size; 6379 } 6380 6381 // A variant of the above (block_size_using_printezis_bits()) except 6382 // that we return 0 if the P-bits are not yet set. 6383 size_t CMSCollector::block_size_if_printezis_bits(HeapWord* addr) const { 6384 if (_markBitMap.isMarked(addr)) { 6385 assert(_markBitMap.isMarked(addr + 1), "Missing Printezis bit?"); 6386 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 6387 size_t size = pointer_delta(nextOneAddr + 1, addr); 6388 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 6389 "alignment problem"); 6390 assert(size >= 3, "Necessary for Printezis marks to work"); 6391 return size; 6392 } else { 6393 assert(!_markBitMap.isMarked(addr + 1), "Bit map inconsistency?"); 6394 return 0; 6395 } 6396 } 6397 6398 HeapWord* CMSCollector::next_card_start_after_block(HeapWord* addr) const { 6399 size_t sz = 0; 6400 oop p = (oop)addr; 6401 if (p->klass_or_null() != NULL && p->is_parsable()) { 6402 sz = CompactibleFreeListSpace::adjustObjectSize(p->size()); 6403 } else { 6404 sz = block_size_using_printezis_bits(addr); 6405 } 6406 assert(sz > 0, "size must be nonzero"); 6407 HeapWord* next_block = addr + sz; 6408 HeapWord* next_card = (HeapWord*)round_to((uintptr_t)next_block, 6409 CardTableModRefBS::card_size); 6410 assert(round_down((uintptr_t)addr, CardTableModRefBS::card_size) < 6411 round_down((uintptr_t)next_card, CardTableModRefBS::card_size), 6412 "must be different cards"); 6413 return next_card; 6414 } 6415 6416 6417 // CMS Bit Map Wrapper ///////////////////////////////////////// 6418 6419 // Construct a CMS bit map infrastructure, but don't create the 6420 // bit vector itself. That is done by a separate call CMSBitMap::allocate() 6421 // further below. 6422 CMSBitMap::CMSBitMap(int shifter, int mutex_rank, const char* mutex_name): 6423 _bm(), 6424 _shifter(shifter), 6425 _lock(mutex_rank >= 0 ? new Mutex(mutex_rank, mutex_name, true) : NULL) 6426 { 6427 _bmStartWord = 0; 6428 _bmWordSize = 0; 6429 } 6430 6431 bool CMSBitMap::allocate(MemRegion mr) { 6432 _bmStartWord = mr.start(); 6433 _bmWordSize = mr.word_size(); 6434 ReservedSpace brs(ReservedSpace::allocation_align_size_up( 6435 (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1)); 6436 if (!brs.is_reserved()) { 6437 warning("CMS bit map allocation failure"); 6438 return false; 6439 } 6440 // For now we'll just commit all of the bit map up fromt. 6441 // Later on we'll try to be more parsimonious with swap. 6442 if (!_virtual_space.initialize(brs, brs.size())) { 6443 warning("CMS bit map backing store failure"); 6444 return false; 6445 } 6446 assert(_virtual_space.committed_size() == brs.size(), 6447 "didn't reserve backing store for all of CMS bit map?"); 6448 _bm.set_map((BitMap::bm_word_t*)_virtual_space.low()); 6449 assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >= 6450 _bmWordSize, "inconsistency in bit map sizing"); 6451 _bm.set_size(_bmWordSize >> _shifter); 6452 6453 // bm.clear(); // can we rely on getting zero'd memory? verify below 6454 assert(isAllClear(), 6455 "Expected zero'd memory from ReservedSpace constructor"); 6456 assert(_bm.size() == heapWordDiffToOffsetDiff(sizeInWords()), 6457 "consistency check"); 6458 return true; 6459 } 6460 6461 void CMSBitMap::dirty_range_iterate_clear(MemRegion mr, MemRegionClosure* cl) { 6462 HeapWord *next_addr, *end_addr, *last_addr; 6463 assert_locked(); 6464 assert(covers(mr), "out-of-range error"); 6465 // XXX assert that start and end are appropriately aligned 6466 for (next_addr = mr.start(), end_addr = mr.end(); 6467 next_addr < end_addr; next_addr = last_addr) { 6468 MemRegion dirty_region = getAndClearMarkedRegion(next_addr, end_addr); 6469 last_addr = dirty_region.end(); 6470 if (!dirty_region.is_empty()) { 6471 cl->do_MemRegion(dirty_region); 6472 } else { 6473 assert(last_addr == end_addr, "program logic"); 6474 return; 6475 } 6476 } 6477 } 6478 6479 #ifndef PRODUCT 6480 void CMSBitMap::assert_locked() const { 6481 CMSLockVerifier::assert_locked(lock()); 6482 } 6483 6484 bool CMSBitMap::covers(MemRegion mr) const { 6485 // assert(_bm.map() == _virtual_space.low(), "map inconsistency"); 6486 assert((size_t)_bm.size() == (_bmWordSize >> _shifter), 6487 "size inconsistency"); 6488 return (mr.start() >= _bmStartWord) && 6489 (mr.end() <= endWord()); 6490 } 6491 6492 bool CMSBitMap::covers(HeapWord* start, size_t size) const { 6493 return (start >= _bmStartWord && (start + size) <= endWord()); 6494 } 6495 6496 void CMSBitMap::verifyNoOneBitsInRange(HeapWord* left, HeapWord* right) { 6497 // verify that there are no 1 bits in the interval [left, right) 6498 FalseBitMapClosure falseBitMapClosure; 6499 iterate(&falseBitMapClosure, left, right); 6500 } 6501 6502 void CMSBitMap::region_invariant(MemRegion mr) 6503 { 6504 assert_locked(); 6505 // mr = mr.intersection(MemRegion(_bmStartWord, _bmWordSize)); 6506 assert(!mr.is_empty(), "unexpected empty region"); 6507 assert(covers(mr), "mr should be covered by bit map"); 6508 // convert address range into offset range 6509 size_t start_ofs = heapWordToOffset(mr.start()); 6510 // Make sure that end() is appropriately aligned 6511 assert(mr.end() == (HeapWord*)round_to((intptr_t)mr.end(), 6512 (1 << (_shifter+LogHeapWordSize))), 6513 "Misaligned mr.end()"); 6514 size_t end_ofs = heapWordToOffset(mr.end()); 6515 assert(end_ofs > start_ofs, "Should mark at least one bit"); 6516 } 6517 6518 #endif 6519 6520 bool CMSMarkStack::allocate(size_t size) { 6521 // allocate a stack of the requisite depth 6522 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 6523 size * sizeof(oop))); 6524 if (!rs.is_reserved()) { 6525 warning("CMSMarkStack allocation failure"); 6526 return false; 6527 } 6528 if (!_virtual_space.initialize(rs, rs.size())) { 6529 warning("CMSMarkStack backing store failure"); 6530 return false; 6531 } 6532 assert(_virtual_space.committed_size() == rs.size(), 6533 "didn't reserve backing store for all of CMS stack?"); 6534 _base = (oop*)(_virtual_space.low()); 6535 _index = 0; 6536 _capacity = size; 6537 NOT_PRODUCT(_max_depth = 0); 6538 return true; 6539 } 6540 6541 // XXX FIX ME !!! In the MT case we come in here holding a 6542 // leaf lock. For printing we need to take a further lock 6543 // which has lower rank. We need to recallibrate the two 6544 // lock-ranks involved in order to be able to rpint the 6545 // messages below. (Or defer the printing to the caller. 6546 // For now we take the expedient path of just disabling the 6547 // messages for the problematic case.) 6548 void CMSMarkStack::expand() { 6549 assert(_capacity <= MarkStackSizeMax, "stack bigger than permitted"); 6550 if (_capacity == MarkStackSizeMax) { 6551 if (_hit_limit++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) { 6552 // We print a warning message only once per CMS cycle. 6553 gclog_or_tty->print_cr(" (benign) Hit CMSMarkStack max size limit"); 6554 } 6555 return; 6556 } 6557 // Double capacity if possible 6558 size_t new_capacity = MIN2(_capacity*2, MarkStackSizeMax); 6559 // Do not give up existing stack until we have managed to 6560 // get the double capacity that we desired. 6561 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 6562 new_capacity * sizeof(oop))); 6563 if (rs.is_reserved()) { 6564 // Release the backing store associated with old stack 6565 _virtual_space.release(); 6566 // Reinitialize virtual space for new stack 6567 if (!_virtual_space.initialize(rs, rs.size())) { 6568 fatal("Not enough swap for expanded marking stack"); 6569 } 6570 _base = (oop*)(_virtual_space.low()); 6571 _index = 0; 6572 _capacity = new_capacity; 6573 } else if (_failed_double++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) { 6574 // Failed to double capacity, continue; 6575 // we print a detail message only once per CMS cycle. 6576 gclog_or_tty->print(" (benign) Failed to expand marking stack from "SIZE_FORMAT"K to " 6577 SIZE_FORMAT"K", 6578 _capacity / K, new_capacity / K); 6579 } 6580 } 6581 6582 6583 // Closures 6584 // XXX: there seems to be a lot of code duplication here; 6585 // should refactor and consolidate common code. 6586 6587 // This closure is used to mark refs into the CMS generation in 6588 // the CMS bit map. Called at the first checkpoint. This closure 6589 // assumes that we do not need to re-mark dirty cards; if the CMS 6590 // generation on which this is used is not an oldest (modulo perm gen) 6591 // generation then this will lose younger_gen cards! 6592 6593 MarkRefsIntoClosure::MarkRefsIntoClosure( 6594 MemRegion span, CMSBitMap* bitMap): 6595 _span(span), 6596 _bitMap(bitMap) 6597 { 6598 assert(_ref_processor == NULL, "deliberately left NULL"); 6599 assert(_bitMap->covers(_span), "_bitMap/_span mismatch"); 6600 } 6601 6602 void MarkRefsIntoClosure::do_oop(oop obj) { 6603 // if p points into _span, then mark corresponding bit in _markBitMap 6604 assert(obj->is_oop(), "expected an oop"); 6605 HeapWord* addr = (HeapWord*)obj; 6606 if (_span.contains(addr)) { 6607 // this should be made more efficient 6608 _bitMap->mark(addr); 6609 } 6610 } 6611 6612 void MarkRefsIntoClosure::do_oop(oop* p) { MarkRefsIntoClosure::do_oop_work(p); } 6613 void MarkRefsIntoClosure::do_oop(narrowOop* p) { MarkRefsIntoClosure::do_oop_work(p); } 6614 6615 // A variant of the above, used for CMS marking verification. 6616 MarkRefsIntoVerifyClosure::MarkRefsIntoVerifyClosure( 6617 MemRegion span, CMSBitMap* verification_bm, CMSBitMap* cms_bm): 6618 _span(span), 6619 _verification_bm(verification_bm), 6620 _cms_bm(cms_bm) 6621 { 6622 assert(_ref_processor == NULL, "deliberately left NULL"); 6623 assert(_verification_bm->covers(_span), "_verification_bm/_span mismatch"); 6624 } 6625 6626 void MarkRefsIntoVerifyClosure::do_oop(oop obj) { 6627 // if p points into _span, then mark corresponding bit in _markBitMap 6628 assert(obj->is_oop(), "expected an oop"); 6629 HeapWord* addr = (HeapWord*)obj; 6630 if (_span.contains(addr)) { 6631 _verification_bm->mark(addr); 6632 if (!_cms_bm->isMarked(addr)) { 6633 oop(addr)->print(); 6634 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)", addr); 6635 fatal("... aborting"); 6636 } 6637 } 6638 } 6639 6640 void MarkRefsIntoVerifyClosure::do_oop(oop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); } 6641 void MarkRefsIntoVerifyClosure::do_oop(narrowOop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); } 6642 6643 ////////////////////////////////////////////////// 6644 // MarkRefsIntoAndScanClosure 6645 ////////////////////////////////////////////////// 6646 6647 MarkRefsIntoAndScanClosure::MarkRefsIntoAndScanClosure(MemRegion span, 6648 ReferenceProcessor* rp, 6649 CMSBitMap* bit_map, 6650 CMSBitMap* mod_union_table, 6651 CMSMarkStack* mark_stack, 6652 CMSMarkStack* revisit_stack, 6653 CMSCollector* collector, 6654 bool should_yield, 6655 bool concurrent_precleaning): 6656 _collector(collector), 6657 _span(span), 6658 _bit_map(bit_map), 6659 _mark_stack(mark_stack), 6660 _pushAndMarkClosure(collector, span, rp, bit_map, mod_union_table, 6661 mark_stack, revisit_stack, concurrent_precleaning), 6662 _yield(should_yield), 6663 _concurrent_precleaning(concurrent_precleaning), 6664 _freelistLock(NULL) 6665 { 6666 _ref_processor = rp; 6667 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 6668 } 6669 6670 // This closure is used to mark refs into the CMS generation at the 6671 // second (final) checkpoint, and to scan and transitively follow 6672 // the unmarked oops. It is also used during the concurrent precleaning 6673 // phase while scanning objects on dirty cards in the CMS generation. 6674 // The marks are made in the marking bit map and the marking stack is 6675 // used for keeping the (newly) grey objects during the scan. 6676 // The parallel version (Par_...) appears further below. 6677 void MarkRefsIntoAndScanClosure::do_oop(oop obj) { 6678 if (obj != NULL) { 6679 assert(obj->is_oop(), "expected an oop"); 6680 HeapWord* addr = (HeapWord*)obj; 6681 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 6682 assert(_collector->overflow_list_is_empty(), 6683 "overflow list should be empty"); 6684 if (_span.contains(addr) && 6685 !_bit_map->isMarked(addr)) { 6686 // mark bit map (object is now grey) 6687 _bit_map->mark(addr); 6688 // push on marking stack (stack should be empty), and drain the 6689 // stack by applying this closure to the oops in the oops popped 6690 // from the stack (i.e. blacken the grey objects) 6691 bool res = _mark_stack->push(obj); 6692 assert(res, "Should have space to push on empty stack"); 6693 do { 6694 oop new_oop = _mark_stack->pop(); 6695 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 6696 assert(new_oop->is_parsable(), "Found unparsable oop"); 6697 assert(_bit_map->isMarked((HeapWord*)new_oop), 6698 "only grey objects on this stack"); 6699 // iterate over the oops in this oop, marking and pushing 6700 // the ones in CMS heap (i.e. in _span). 6701 new_oop->oop_iterate(&_pushAndMarkClosure); 6702 // check if it's time to yield 6703 do_yield_check(); 6704 } while (!_mark_stack->isEmpty() || 6705 (!_concurrent_precleaning && take_from_overflow_list())); 6706 // if marking stack is empty, and we are not doing this 6707 // during precleaning, then check the overflow list 6708 } 6709 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 6710 assert(_collector->overflow_list_is_empty(), 6711 "overflow list was drained above"); 6712 // We could restore evacuated mark words, if any, used for 6713 // overflow list links here because the overflow list is 6714 // provably empty here. That would reduce the maximum 6715 // size requirements for preserved_{oop,mark}_stack. 6716 // But we'll just postpone it until we are all done 6717 // so we can just stream through. 6718 if (!_concurrent_precleaning && CMSOverflowEarlyRestoration) { 6719 _collector->restore_preserved_marks_if_any(); 6720 assert(_collector->no_preserved_marks(), "No preserved marks"); 6721 } 6722 assert(!CMSOverflowEarlyRestoration || _collector->no_preserved_marks(), 6723 "All preserved marks should have been restored above"); 6724 } 6725 } 6726 6727 void MarkRefsIntoAndScanClosure::do_oop(oop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); } 6728 void MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); } 6729 6730 void MarkRefsIntoAndScanClosure::do_yield_work() { 6731 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6732 "CMS thread should hold CMS token"); 6733 assert_lock_strong(_freelistLock); 6734 assert_lock_strong(_bit_map->lock()); 6735 // relinquish the free_list_lock and bitMaplock() 6736 DEBUG_ONLY(RememberKlassesChecker mux(false);) 6737 _bit_map->lock()->unlock(); 6738 _freelistLock->unlock(); 6739 ConcurrentMarkSweepThread::desynchronize(true); 6740 ConcurrentMarkSweepThread::acknowledge_yield_request(); 6741 _collector->stopTimer(); 6742 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 6743 if (PrintCMSStatistics != 0) { 6744 _collector->incrementYields(); 6745 } 6746 _collector->icms_wait(); 6747 6748 // See the comment in coordinator_yield() 6749 for (unsigned i = 0; 6750 i < CMSYieldSleepCount && 6751 ConcurrentMarkSweepThread::should_yield() && 6752 !CMSCollector::foregroundGCIsActive(); 6753 ++i) { 6754 os::sleep(Thread::current(), 1, false); 6755 ConcurrentMarkSweepThread::acknowledge_yield_request(); 6756 } 6757 6758 ConcurrentMarkSweepThread::synchronize(true); 6759 _freelistLock->lock_without_safepoint_check(); 6760 _bit_map->lock()->lock_without_safepoint_check(); 6761 _collector->startTimer(); 6762 } 6763 6764 /////////////////////////////////////////////////////////// 6765 // Par_MarkRefsIntoAndScanClosure: a parallel version of 6766 // MarkRefsIntoAndScanClosure 6767 /////////////////////////////////////////////////////////// 6768 Par_MarkRefsIntoAndScanClosure::Par_MarkRefsIntoAndScanClosure( 6769 CMSCollector* collector, MemRegion span, ReferenceProcessor* rp, 6770 CMSBitMap* bit_map, OopTaskQueue* work_queue, CMSMarkStack* revisit_stack): 6771 _span(span), 6772 _bit_map(bit_map), 6773 _work_queue(work_queue), 6774 _low_water_mark(MIN2((uint)(work_queue->max_elems()/4), 6775 (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads))), 6776 _par_pushAndMarkClosure(collector, span, rp, bit_map, work_queue, 6777 revisit_stack) 6778 { 6779 _ref_processor = rp; 6780 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 6781 } 6782 6783 // This closure is used to mark refs into the CMS generation at the 6784 // second (final) checkpoint, and to scan and transitively follow 6785 // the unmarked oops. The marks are made in the marking bit map and 6786 // the work_queue is used for keeping the (newly) grey objects during 6787 // the scan phase whence they are also available for stealing by parallel 6788 // threads. Since the marking bit map is shared, updates are 6789 // synchronized (via CAS). 6790 void Par_MarkRefsIntoAndScanClosure::do_oop(oop obj) { 6791 if (obj != NULL) { 6792 // Ignore mark word because this could be an already marked oop 6793 // that may be chained at the end of the overflow list. 6794 assert(obj->is_oop(true), "expected an oop"); 6795 HeapWord* addr = (HeapWord*)obj; 6796 if (_span.contains(addr) && 6797 !_bit_map->isMarked(addr)) { 6798 // mark bit map (object will become grey): 6799 // It is possible for several threads to be 6800 // trying to "claim" this object concurrently; 6801 // the unique thread that succeeds in marking the 6802 // object first will do the subsequent push on 6803 // to the work queue (or overflow list). 6804 if (_bit_map->par_mark(addr)) { 6805 // push on work_queue (which may not be empty), and trim the 6806 // queue to an appropriate length by applying this closure to 6807 // the oops in the oops popped from the stack (i.e. blacken the 6808 // grey objects) 6809 bool res = _work_queue->push(obj); 6810 assert(res, "Low water mark should be less than capacity?"); 6811 trim_queue(_low_water_mark); 6812 } // Else, another thread claimed the object 6813 } 6814 } 6815 } 6816 6817 void Par_MarkRefsIntoAndScanClosure::do_oop(oop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); } 6818 void Par_MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); } 6819 6820 // This closure is used to rescan the marked objects on the dirty cards 6821 // in the mod union table and the card table proper. 6822 size_t ScanMarkedObjectsAgainCarefullyClosure::do_object_careful_m( 6823 oop p, MemRegion mr) { 6824 6825 size_t size = 0; 6826 HeapWord* addr = (HeapWord*)p; 6827 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6828 assert(_span.contains(addr), "we are scanning the CMS generation"); 6829 // check if it's time to yield 6830 if (do_yield_check()) { 6831 // We yielded for some foreground stop-world work, 6832 // and we have been asked to abort this ongoing preclean cycle. 6833 return 0; 6834 } 6835 if (_bitMap->isMarked(addr)) { 6836 // it's marked; is it potentially uninitialized? 6837 if (p->klass_or_null() != NULL) { 6838 // If is_conc_safe is false, the object may be undergoing 6839 // change by the VM outside a safepoint. Don't try to 6840 // scan it, but rather leave it for the remark phase. 6841 if (CMSPermGenPrecleaningEnabled && 6842 (!p->is_conc_safe() || !p->is_parsable())) { 6843 // Signal precleaning to redirty the card since 6844 // the klass pointer is already installed. 6845 assert(size == 0, "Initial value"); 6846 } else { 6847 assert(p->is_parsable(), "must be parsable."); 6848 // an initialized object; ignore mark word in verification below 6849 // since we are running concurrent with mutators 6850 assert(p->is_oop(true), "should be an oop"); 6851 if (p->is_objArray()) { 6852 // objArrays are precisely marked; restrict scanning 6853 // to dirty cards only. 6854 size = CompactibleFreeListSpace::adjustObjectSize( 6855 p->oop_iterate(_scanningClosure, mr)); 6856 } else { 6857 // A non-array may have been imprecisely marked; we need 6858 // to scan object in its entirety. 6859 size = CompactibleFreeListSpace::adjustObjectSize( 6860 p->oop_iterate(_scanningClosure)); 6861 } 6862 #ifdef DEBUG 6863 size_t direct_size = 6864 CompactibleFreeListSpace::adjustObjectSize(p->size()); 6865 assert(size == direct_size, "Inconsistency in size"); 6866 assert(size >= 3, "Necessary for Printezis marks to work"); 6867 if (!_bitMap->isMarked(addr+1)) { 6868 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size); 6869 } else { 6870 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size-1); 6871 assert(_bitMap->isMarked(addr+size-1), 6872 "inconsistent Printezis mark"); 6873 } 6874 #endif // DEBUG 6875 } 6876 } else { 6877 // an unitialized object 6878 assert(_bitMap->isMarked(addr+1), "missing Printezis mark?"); 6879 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 6880 size = pointer_delta(nextOneAddr + 1, addr); 6881 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 6882 "alignment problem"); 6883 // Note that pre-cleaning needn't redirty the card. OopDesc::set_klass() 6884 // will dirty the card when the klass pointer is installed in the 6885 // object (signalling the completion of initialization). 6886 } 6887 } else { 6888 // Either a not yet marked object or an uninitialized object 6889 if (p->klass_or_null() == NULL || !p->is_parsable()) { 6890 // An uninitialized object, skip to the next card, since 6891 // we may not be able to read its P-bits yet. 6892 assert(size == 0, "Initial value"); 6893 } else { 6894 // An object not (yet) reached by marking: we merely need to 6895 // compute its size so as to go look at the next block. 6896 assert(p->is_oop(true), "should be an oop"); 6897 size = CompactibleFreeListSpace::adjustObjectSize(p->size()); 6898 } 6899 } 6900 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6901 return size; 6902 } 6903 6904 void ScanMarkedObjectsAgainCarefullyClosure::do_yield_work() { 6905 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6906 "CMS thread should hold CMS token"); 6907 assert_lock_strong(_freelistLock); 6908 assert_lock_strong(_bitMap->lock()); 6909 DEBUG_ONLY(RememberKlassesChecker mux(false);) 6910 // relinquish the free_list_lock and bitMaplock() 6911 _bitMap->lock()->unlock(); 6912 _freelistLock->unlock(); 6913 ConcurrentMarkSweepThread::desynchronize(true); 6914 ConcurrentMarkSweepThread::acknowledge_yield_request(); 6915 _collector->stopTimer(); 6916 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 6917 if (PrintCMSStatistics != 0) { 6918 _collector->incrementYields(); 6919 } 6920 _collector->icms_wait(); 6921 6922 // See the comment in coordinator_yield() 6923 for (unsigned i = 0; i < CMSYieldSleepCount && 6924 ConcurrentMarkSweepThread::should_yield() && 6925 !CMSCollector::foregroundGCIsActive(); ++i) { 6926 os::sleep(Thread::current(), 1, false); 6927 ConcurrentMarkSweepThread::acknowledge_yield_request(); 6928 } 6929 6930 ConcurrentMarkSweepThread::synchronize(true); 6931 _freelistLock->lock_without_safepoint_check(); 6932 _bitMap->lock()->lock_without_safepoint_check(); 6933 _collector->startTimer(); 6934 } 6935 6936 6937 ////////////////////////////////////////////////////////////////// 6938 // SurvivorSpacePrecleanClosure 6939 ////////////////////////////////////////////////////////////////// 6940 // This (single-threaded) closure is used to preclean the oops in 6941 // the survivor spaces. 6942 size_t SurvivorSpacePrecleanClosure::do_object_careful(oop p) { 6943 6944 HeapWord* addr = (HeapWord*)p; 6945 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6946 assert(!_span.contains(addr), "we are scanning the survivor spaces"); 6947 assert(p->klass_or_null() != NULL, "object should be initializd"); 6948 assert(p->is_parsable(), "must be parsable."); 6949 // an initialized object; ignore mark word in verification below 6950 // since we are running concurrent with mutators 6951 assert(p->is_oop(true), "should be an oop"); 6952 // Note that we do not yield while we iterate over 6953 // the interior oops of p, pushing the relevant ones 6954 // on our marking stack. 6955 size_t size = p->oop_iterate(_scanning_closure); 6956 do_yield_check(); 6957 // Observe that below, we do not abandon the preclean 6958 // phase as soon as we should; rather we empty the 6959 // marking stack before returning. This is to satisfy 6960 // some existing assertions. In general, it may be a 6961 // good idea to abort immediately and complete the marking 6962 // from the grey objects at a later time. 6963 while (!_mark_stack->isEmpty()) { 6964 oop new_oop = _mark_stack->pop(); 6965 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 6966 assert(new_oop->is_parsable(), "Found unparsable oop"); 6967 assert(_bit_map->isMarked((HeapWord*)new_oop), 6968 "only grey objects on this stack"); 6969 // iterate over the oops in this oop, marking and pushing 6970 // the ones in CMS heap (i.e. in _span). 6971 new_oop->oop_iterate(_scanning_closure); 6972 // check if it's time to yield 6973 do_yield_check(); 6974 } 6975 unsigned int after_count = 6976 GenCollectedHeap::heap()->total_collections(); 6977 bool abort = (_before_count != after_count) || 6978 _collector->should_abort_preclean(); 6979 return abort ? 0 : size; 6980 } 6981 6982 void SurvivorSpacePrecleanClosure::do_yield_work() { 6983 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6984 "CMS thread should hold CMS token"); 6985 assert_lock_strong(_bit_map->lock()); 6986 DEBUG_ONLY(RememberKlassesChecker smx(false);) 6987 // Relinquish the bit map lock 6988 _bit_map->lock()->unlock(); 6989 ConcurrentMarkSweepThread::desynchronize(true); 6990 ConcurrentMarkSweepThread::acknowledge_yield_request(); 6991 _collector->stopTimer(); 6992 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 6993 if (PrintCMSStatistics != 0) { 6994 _collector->incrementYields(); 6995 } 6996 _collector->icms_wait(); 6997 6998 // See the comment in coordinator_yield() 6999 for (unsigned i = 0; i < CMSYieldSleepCount && 7000 ConcurrentMarkSweepThread::should_yield() && 7001 !CMSCollector::foregroundGCIsActive(); ++i) { 7002 os::sleep(Thread::current(), 1, false); 7003 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7004 } 7005 7006 ConcurrentMarkSweepThread::synchronize(true); 7007 _bit_map->lock()->lock_without_safepoint_check(); 7008 _collector->startTimer(); 7009 } 7010 7011 // This closure is used to rescan the marked objects on the dirty cards 7012 // in the mod union table and the card table proper. In the parallel 7013 // case, although the bitMap is shared, we do a single read so the 7014 // isMarked() query is "safe". 7015 bool ScanMarkedObjectsAgainClosure::do_object_bm(oop p, MemRegion mr) { 7016 // Ignore mark word because we are running concurrent with mutators 7017 assert(p->is_oop_or_null(true), "expected an oop or null"); 7018 HeapWord* addr = (HeapWord*)p; 7019 assert(_span.contains(addr), "we are scanning the CMS generation"); 7020 bool is_obj_array = false; 7021 #ifdef DEBUG 7022 if (!_parallel) { 7023 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 7024 assert(_collector->overflow_list_is_empty(), 7025 "overflow list should be empty"); 7026 7027 } 7028 #endif // DEBUG 7029 if (_bit_map->isMarked(addr)) { 7030 // Obj arrays are precisely marked, non-arrays are not; 7031 // so we scan objArrays precisely and non-arrays in their 7032 // entirety. 7033 if (p->is_objArray()) { 7034 is_obj_array = true; 7035 if (_parallel) { 7036 p->oop_iterate(_par_scan_closure, mr); 7037 } else { 7038 p->oop_iterate(_scan_closure, mr); 7039 } 7040 } else { 7041 if (_parallel) { 7042 p->oop_iterate(_par_scan_closure); 7043 } else { 7044 p->oop_iterate(_scan_closure); 7045 } 7046 } 7047 } 7048 #ifdef DEBUG 7049 if (!_parallel) { 7050 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 7051 assert(_collector->overflow_list_is_empty(), 7052 "overflow list should be empty"); 7053 7054 } 7055 #endif // DEBUG 7056 return is_obj_array; 7057 } 7058 7059 MarkFromRootsClosure::MarkFromRootsClosure(CMSCollector* collector, 7060 MemRegion span, 7061 CMSBitMap* bitMap, CMSMarkStack* markStack, 7062 CMSMarkStack* revisitStack, 7063 bool should_yield, bool verifying): 7064 _collector(collector), 7065 _span(span), 7066 _bitMap(bitMap), 7067 _mut(&collector->_modUnionTable), 7068 _markStack(markStack), 7069 _revisitStack(revisitStack), 7070 _yield(should_yield), 7071 _skipBits(0) 7072 { 7073 assert(_markStack->isEmpty(), "stack should be empty"); 7074 _finger = _bitMap->startWord(); 7075 _threshold = _finger; 7076 assert(_collector->_restart_addr == NULL, "Sanity check"); 7077 assert(_span.contains(_finger), "Out of bounds _finger?"); 7078 DEBUG_ONLY(_verifying = verifying;) 7079 } 7080 7081 void MarkFromRootsClosure::reset(HeapWord* addr) { 7082 assert(_markStack->isEmpty(), "would cause duplicates on stack"); 7083 assert(_span.contains(addr), "Out of bounds _finger?"); 7084 _finger = addr; 7085 _threshold = (HeapWord*)round_to( 7086 (intptr_t)_finger, CardTableModRefBS::card_size); 7087 } 7088 7089 // Should revisit to see if this should be restructured for 7090 // greater efficiency. 7091 bool MarkFromRootsClosure::do_bit(size_t offset) { 7092 if (_skipBits > 0) { 7093 _skipBits--; 7094 return true; 7095 } 7096 // convert offset into a HeapWord* 7097 HeapWord* addr = _bitMap->startWord() + offset; 7098 assert(_bitMap->endWord() && addr < _bitMap->endWord(), 7099 "address out of range"); 7100 assert(_bitMap->isMarked(addr), "tautology"); 7101 if (_bitMap->isMarked(addr+1)) { 7102 // this is an allocated but not yet initialized object 7103 assert(_skipBits == 0, "tautology"); 7104 _skipBits = 2; // skip next two marked bits ("Printezis-marks") 7105 oop p = oop(addr); 7106 if (p->klass_or_null() == NULL || !p->is_parsable()) { 7107 DEBUG_ONLY(if (!_verifying) {) 7108 // We re-dirty the cards on which this object lies and increase 7109 // the _threshold so that we'll come back to scan this object 7110 // during the preclean or remark phase. (CMSCleanOnEnter) 7111 if (CMSCleanOnEnter) { 7112 size_t sz = _collector->block_size_using_printezis_bits(addr); 7113 HeapWord* end_card_addr = (HeapWord*)round_to( 7114 (intptr_t)(addr+sz), CardTableModRefBS::card_size); 7115 MemRegion redirty_range = MemRegion(addr, end_card_addr); 7116 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 7117 // Bump _threshold to end_card_addr; note that 7118 // _threshold cannot possibly exceed end_card_addr, anyhow. 7119 // This prevents future clearing of the card as the scan proceeds 7120 // to the right. 7121 assert(_threshold <= end_card_addr, 7122 "Because we are just scanning into this object"); 7123 if (_threshold < end_card_addr) { 7124 _threshold = end_card_addr; 7125 } 7126 if (p->klass_or_null() != NULL) { 7127 // Redirty the range of cards... 7128 _mut->mark_range(redirty_range); 7129 } // ...else the setting of klass will dirty the card anyway. 7130 } 7131 DEBUG_ONLY(}) 7132 return true; 7133 } 7134 } 7135 scanOopsInOop(addr); 7136 return true; 7137 } 7138 7139 // We take a break if we've been at this for a while, 7140 // so as to avoid monopolizing the locks involved. 7141 void MarkFromRootsClosure::do_yield_work() { 7142 // First give up the locks, then yield, then re-lock 7143 // We should probably use a constructor/destructor idiom to 7144 // do this unlock/lock or modify the MutexUnlocker class to 7145 // serve our purpose. XXX 7146 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 7147 "CMS thread should hold CMS token"); 7148 assert_lock_strong(_bitMap->lock()); 7149 DEBUG_ONLY(RememberKlassesChecker mux(false);) 7150 _bitMap->lock()->unlock(); 7151 ConcurrentMarkSweepThread::desynchronize(true); 7152 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7153 _collector->stopTimer(); 7154 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 7155 if (PrintCMSStatistics != 0) { 7156 _collector->incrementYields(); 7157 } 7158 _collector->icms_wait(); 7159 7160 // See the comment in coordinator_yield() 7161 for (unsigned i = 0; i < CMSYieldSleepCount && 7162 ConcurrentMarkSweepThread::should_yield() && 7163 !CMSCollector::foregroundGCIsActive(); ++i) { 7164 os::sleep(Thread::current(), 1, false); 7165 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7166 } 7167 7168 ConcurrentMarkSweepThread::synchronize(true); 7169 _bitMap->lock()->lock_without_safepoint_check(); 7170 _collector->startTimer(); 7171 } 7172 7173 void MarkFromRootsClosure::scanOopsInOop(HeapWord* ptr) { 7174 assert(_bitMap->isMarked(ptr), "expected bit to be set"); 7175 assert(_markStack->isEmpty(), 7176 "should drain stack to limit stack usage"); 7177 // convert ptr to an oop preparatory to scanning 7178 oop obj = oop(ptr); 7179 // Ignore mark word in verification below, since we 7180 // may be running concurrent with mutators. 7181 assert(obj->is_oop(true), "should be an oop"); 7182 assert(_finger <= ptr, "_finger runneth ahead"); 7183 // advance the finger to right end of this object 7184 _finger = ptr + obj->size(); 7185 assert(_finger > ptr, "we just incremented it above"); 7186 // On large heaps, it may take us some time to get through 7187 // the marking phase (especially if running iCMS). During 7188 // this time it's possible that a lot of mutations have 7189 // accumulated in the card table and the mod union table -- 7190 // these mutation records are redundant until we have 7191 // actually traced into the corresponding card. 7192 // Here, we check whether advancing the finger would make 7193 // us cross into a new card, and if so clear corresponding 7194 // cards in the MUT (preclean them in the card-table in the 7195 // future). 7196 7197 DEBUG_ONLY(if (!_verifying) {) 7198 // The clean-on-enter optimization is disabled by default, 7199 // until we fix 6178663. 7200 if (CMSCleanOnEnter && (_finger > _threshold)) { 7201 // [_threshold, _finger) represents the interval 7202 // of cards to be cleared in MUT (or precleaned in card table). 7203 // The set of cards to be cleared is all those that overlap 7204 // with the interval [_threshold, _finger); note that 7205 // _threshold is always kept card-aligned but _finger isn't 7206 // always card-aligned. 7207 HeapWord* old_threshold = _threshold; 7208 assert(old_threshold == (HeapWord*)round_to( 7209 (intptr_t)old_threshold, CardTableModRefBS::card_size), 7210 "_threshold should always be card-aligned"); 7211 _threshold = (HeapWord*)round_to( 7212 (intptr_t)_finger, CardTableModRefBS::card_size); 7213 MemRegion mr(old_threshold, _threshold); 7214 assert(!mr.is_empty(), "Control point invariant"); 7215 assert(_span.contains(mr), "Should clear within span"); 7216 // XXX When _finger crosses from old gen into perm gen 7217 // we may be doing unnecessary cleaning; do better in the 7218 // future by detecting that condition and clearing fewer 7219 // MUT/CT entries. 7220 _mut->clear_range(mr); 7221 } 7222 DEBUG_ONLY(}) 7223 // Note: the finger doesn't advance while we drain 7224 // the stack below. 7225 PushOrMarkClosure pushOrMarkClosure(_collector, 7226 _span, _bitMap, _markStack, 7227 _revisitStack, 7228 _finger, this); 7229 bool res = _markStack->push(obj); 7230 assert(res, "Empty non-zero size stack should have space for single push"); 7231 while (!_markStack->isEmpty()) { 7232 oop new_oop = _markStack->pop(); 7233 // Skip verifying header mark word below because we are 7234 // running concurrent with mutators. 7235 assert(new_oop->is_oop(true), "Oops! expected to pop an oop"); 7236 // now scan this oop's oops 7237 new_oop->oop_iterate(&pushOrMarkClosure); 7238 do_yield_check(); 7239 } 7240 assert(_markStack->isEmpty(), "tautology, emphasizing post-condition"); 7241 } 7242 7243 Par_MarkFromRootsClosure::Par_MarkFromRootsClosure(CMSConcMarkingTask* task, 7244 CMSCollector* collector, MemRegion span, 7245 CMSBitMap* bit_map, 7246 OopTaskQueue* work_queue, 7247 CMSMarkStack* overflow_stack, 7248 CMSMarkStack* revisit_stack, 7249 bool should_yield): 7250 _collector(collector), 7251 _whole_span(collector->_span), 7252 _span(span), 7253 _bit_map(bit_map), 7254 _mut(&collector->_modUnionTable), 7255 _work_queue(work_queue), 7256 _overflow_stack(overflow_stack), 7257 _revisit_stack(revisit_stack), 7258 _yield(should_yield), 7259 _skip_bits(0), 7260 _task(task) 7261 { 7262 assert(_work_queue->size() == 0, "work_queue should be empty"); 7263 _finger = span.start(); 7264 _threshold = _finger; // XXX Defer clear-on-enter optimization for now 7265 assert(_span.contains(_finger), "Out of bounds _finger?"); 7266 } 7267 7268 // Should revisit to see if this should be restructured for 7269 // greater efficiency. 7270 bool Par_MarkFromRootsClosure::do_bit(size_t offset) { 7271 if (_skip_bits > 0) { 7272 _skip_bits--; 7273 return true; 7274 } 7275 // convert offset into a HeapWord* 7276 HeapWord* addr = _bit_map->startWord() + offset; 7277 assert(_bit_map->endWord() && addr < _bit_map->endWord(), 7278 "address out of range"); 7279 assert(_bit_map->isMarked(addr), "tautology"); 7280 if (_bit_map->isMarked(addr+1)) { 7281 // this is an allocated object that might not yet be initialized 7282 assert(_skip_bits == 0, "tautology"); 7283 _skip_bits = 2; // skip next two marked bits ("Printezis-marks") 7284 oop p = oop(addr); 7285 if (p->klass_or_null() == NULL || !p->is_parsable()) { 7286 // in the case of Clean-on-Enter optimization, redirty card 7287 // and avoid clearing card by increasing the threshold. 7288 return true; 7289 } 7290 } 7291 scan_oops_in_oop(addr); 7292 return true; 7293 } 7294 7295 void Par_MarkFromRootsClosure::scan_oops_in_oop(HeapWord* ptr) { 7296 assert(_bit_map->isMarked(ptr), "expected bit to be set"); 7297 // Should we assert that our work queue is empty or 7298 // below some drain limit? 7299 assert(_work_queue->size() == 0, 7300 "should drain stack to limit stack usage"); 7301 // convert ptr to an oop preparatory to scanning 7302 oop obj = oop(ptr); 7303 // Ignore mark word in verification below, since we 7304 // may be running concurrent with mutators. 7305 assert(obj->is_oop(true), "should be an oop"); 7306 assert(_finger <= ptr, "_finger runneth ahead"); 7307 // advance the finger to right end of this object 7308 _finger = ptr + obj->size(); 7309 assert(_finger > ptr, "we just incremented it above"); 7310 // On large heaps, it may take us some time to get through 7311 // the marking phase (especially if running iCMS). During 7312 // this time it's possible that a lot of mutations have 7313 // accumulated in the card table and the mod union table -- 7314 // these mutation records are redundant until we have 7315 // actually traced into the corresponding card. 7316 // Here, we check whether advancing the finger would make 7317 // us cross into a new card, and if so clear corresponding 7318 // cards in the MUT (preclean them in the card-table in the 7319 // future). 7320 7321 // The clean-on-enter optimization is disabled by default, 7322 // until we fix 6178663. 7323 if (CMSCleanOnEnter && (_finger > _threshold)) { 7324 // [_threshold, _finger) represents the interval 7325 // of cards to be cleared in MUT (or precleaned in card table). 7326 // The set of cards to be cleared is all those that overlap 7327 // with the interval [_threshold, _finger); note that 7328 // _threshold is always kept card-aligned but _finger isn't 7329 // always card-aligned. 7330 HeapWord* old_threshold = _threshold; 7331 assert(old_threshold == (HeapWord*)round_to( 7332 (intptr_t)old_threshold, CardTableModRefBS::card_size), 7333 "_threshold should always be card-aligned"); 7334 _threshold = (HeapWord*)round_to( 7335 (intptr_t)_finger, CardTableModRefBS::card_size); 7336 MemRegion mr(old_threshold, _threshold); 7337 assert(!mr.is_empty(), "Control point invariant"); 7338 assert(_span.contains(mr), "Should clear within span"); // _whole_span ?? 7339 // XXX When _finger crosses from old gen into perm gen 7340 // we may be doing unnecessary cleaning; do better in the 7341 // future by detecting that condition and clearing fewer 7342 // MUT/CT entries. 7343 _mut->clear_range(mr); 7344 } 7345 7346 // Note: the local finger doesn't advance while we drain 7347 // the stack below, but the global finger sure can and will. 7348 HeapWord** gfa = _task->global_finger_addr(); 7349 Par_PushOrMarkClosure pushOrMarkClosure(_collector, 7350 _span, _bit_map, 7351 _work_queue, 7352 _overflow_stack, 7353 _revisit_stack, 7354 _finger, 7355 gfa, this); 7356 bool res = _work_queue->push(obj); // overflow could occur here 7357 assert(res, "Will hold once we use workqueues"); 7358 while (true) { 7359 oop new_oop; 7360 if (!_work_queue->pop_local(new_oop)) { 7361 // We emptied our work_queue; check if there's stuff that can 7362 // be gotten from the overflow stack. 7363 if (CMSConcMarkingTask::get_work_from_overflow_stack( 7364 _overflow_stack, _work_queue)) { 7365 do_yield_check(); 7366 continue; 7367 } else { // done 7368 break; 7369 } 7370 } 7371 // Skip verifying header mark word below because we are 7372 // running concurrent with mutators. 7373 assert(new_oop->is_oop(true), "Oops! expected to pop an oop"); 7374 // now scan this oop's oops 7375 new_oop->oop_iterate(&pushOrMarkClosure); 7376 do_yield_check(); 7377 } 7378 assert(_work_queue->size() == 0, "tautology, emphasizing post-condition"); 7379 } 7380 7381 // Yield in response to a request from VM Thread or 7382 // from mutators. 7383 void Par_MarkFromRootsClosure::do_yield_work() { 7384 assert(_task != NULL, "sanity"); 7385 _task->yield(); 7386 } 7387 7388 // A variant of the above used for verifying CMS marking work. 7389 MarkFromRootsVerifyClosure::MarkFromRootsVerifyClosure(CMSCollector* collector, 7390 MemRegion span, 7391 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 7392 CMSMarkStack* mark_stack): 7393 _collector(collector), 7394 _span(span), 7395 _verification_bm(verification_bm), 7396 _cms_bm(cms_bm), 7397 _mark_stack(mark_stack), 7398 _pam_verify_closure(collector, span, verification_bm, cms_bm, 7399 mark_stack) 7400 { 7401 assert(_mark_stack->isEmpty(), "stack should be empty"); 7402 _finger = _verification_bm->startWord(); 7403 assert(_collector->_restart_addr == NULL, "Sanity check"); 7404 assert(_span.contains(_finger), "Out of bounds _finger?"); 7405 } 7406 7407 void MarkFromRootsVerifyClosure::reset(HeapWord* addr) { 7408 assert(_mark_stack->isEmpty(), "would cause duplicates on stack"); 7409 assert(_span.contains(addr), "Out of bounds _finger?"); 7410 _finger = addr; 7411 } 7412 7413 // Should revisit to see if this should be restructured for 7414 // greater efficiency. 7415 bool MarkFromRootsVerifyClosure::do_bit(size_t offset) { 7416 // convert offset into a HeapWord* 7417 HeapWord* addr = _verification_bm->startWord() + offset; 7418 assert(_verification_bm->endWord() && addr < _verification_bm->endWord(), 7419 "address out of range"); 7420 assert(_verification_bm->isMarked(addr), "tautology"); 7421 assert(_cms_bm->isMarked(addr), "tautology"); 7422 7423 assert(_mark_stack->isEmpty(), 7424 "should drain stack to limit stack usage"); 7425 // convert addr to an oop preparatory to scanning 7426 oop obj = oop(addr); 7427 assert(obj->is_oop(), "should be an oop"); 7428 assert(_finger <= addr, "_finger runneth ahead"); 7429 // advance the finger to right end of this object 7430 _finger = addr + obj->size(); 7431 assert(_finger > addr, "we just incremented it above"); 7432 // Note: the finger doesn't advance while we drain 7433 // the stack below. 7434 bool res = _mark_stack->push(obj); 7435 assert(res, "Empty non-zero size stack should have space for single push"); 7436 while (!_mark_stack->isEmpty()) { 7437 oop new_oop = _mark_stack->pop(); 7438 assert(new_oop->is_oop(), "Oops! expected to pop an oop"); 7439 // now scan this oop's oops 7440 new_oop->oop_iterate(&_pam_verify_closure); 7441 } 7442 assert(_mark_stack->isEmpty(), "tautology, emphasizing post-condition"); 7443 return true; 7444 } 7445 7446 PushAndMarkVerifyClosure::PushAndMarkVerifyClosure( 7447 CMSCollector* collector, MemRegion span, 7448 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 7449 CMSMarkStack* mark_stack): 7450 OopClosure(collector->ref_processor()), 7451 _collector(collector), 7452 _span(span), 7453 _verification_bm(verification_bm), 7454 _cms_bm(cms_bm), 7455 _mark_stack(mark_stack) 7456 { } 7457 7458 void PushAndMarkVerifyClosure::do_oop(oop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 7459 void PushAndMarkVerifyClosure::do_oop(narrowOop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 7460 7461 // Upon stack overflow, we discard (part of) the stack, 7462 // remembering the least address amongst those discarded 7463 // in CMSCollector's _restart_address. 7464 void PushAndMarkVerifyClosure::handle_stack_overflow(HeapWord* lost) { 7465 // Remember the least grey address discarded 7466 HeapWord* ra = (HeapWord*)_mark_stack->least_value(lost); 7467 _collector->lower_restart_addr(ra); 7468 _mark_stack->reset(); // discard stack contents 7469 _mark_stack->expand(); // expand the stack if possible 7470 } 7471 7472 void PushAndMarkVerifyClosure::do_oop(oop obj) { 7473 assert(obj->is_oop_or_null(), "expected an oop or NULL"); 7474 HeapWord* addr = (HeapWord*)obj; 7475 if (_span.contains(addr) && !_verification_bm->isMarked(addr)) { 7476 // Oop lies in _span and isn't yet grey or black 7477 _verification_bm->mark(addr); // now grey 7478 if (!_cms_bm->isMarked(addr)) { 7479 oop(addr)->print(); 7480 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)", 7481 addr); 7482 fatal("... aborting"); 7483 } 7484 7485 if (!_mark_stack->push(obj)) { // stack overflow 7486 if (PrintCMSStatistics != 0) { 7487 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 7488 SIZE_FORMAT, _mark_stack->capacity()); 7489 } 7490 assert(_mark_stack->isFull(), "Else push should have succeeded"); 7491 handle_stack_overflow(addr); 7492 } 7493 // anything including and to the right of _finger 7494 // will be scanned as we iterate over the remainder of the 7495 // bit map 7496 } 7497 } 7498 7499 PushOrMarkClosure::PushOrMarkClosure(CMSCollector* collector, 7500 MemRegion span, 7501 CMSBitMap* bitMap, CMSMarkStack* markStack, 7502 CMSMarkStack* revisitStack, 7503 HeapWord* finger, MarkFromRootsClosure* parent) : 7504 KlassRememberingOopClosure(collector, collector->ref_processor(), revisitStack), 7505 _span(span), 7506 _bitMap(bitMap), 7507 _markStack(markStack), 7508 _finger(finger), 7509 _parent(parent) 7510 { } 7511 7512 Par_PushOrMarkClosure::Par_PushOrMarkClosure(CMSCollector* collector, 7513 MemRegion span, 7514 CMSBitMap* bit_map, 7515 OopTaskQueue* work_queue, 7516 CMSMarkStack* overflow_stack, 7517 CMSMarkStack* revisit_stack, 7518 HeapWord* finger, 7519 HeapWord** global_finger_addr, 7520 Par_MarkFromRootsClosure* parent) : 7521 Par_KlassRememberingOopClosure(collector, 7522 collector->ref_processor(), 7523 revisit_stack), 7524 _whole_span(collector->_span), 7525 _span(span), 7526 _bit_map(bit_map), 7527 _work_queue(work_queue), 7528 _overflow_stack(overflow_stack), 7529 _finger(finger), 7530 _global_finger_addr(global_finger_addr), 7531 _parent(parent) 7532 { } 7533 7534 // Assumes thread-safe access by callers, who are 7535 // responsible for mutual exclusion. 7536 void CMSCollector::lower_restart_addr(HeapWord* low) { 7537 assert(_span.contains(low), "Out of bounds addr"); 7538 if (_restart_addr == NULL) { 7539 _restart_addr = low; 7540 } else { 7541 _restart_addr = MIN2(_restart_addr, low); 7542 } 7543 } 7544 7545 // Upon stack overflow, we discard (part of) the stack, 7546 // remembering the least address amongst those discarded 7547 // in CMSCollector's _restart_address. 7548 void PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 7549 // Remember the least grey address discarded 7550 HeapWord* ra = (HeapWord*)_markStack->least_value(lost); 7551 _collector->lower_restart_addr(ra); 7552 _markStack->reset(); // discard stack contents 7553 _markStack->expand(); // expand the stack if possible 7554 } 7555 7556 // Upon stack overflow, we discard (part of) the stack, 7557 // remembering the least address amongst those discarded 7558 // in CMSCollector's _restart_address. 7559 void Par_PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 7560 // We need to do this under a mutex to prevent other 7561 // workers from interfering with the work done below. 7562 MutexLockerEx ml(_overflow_stack->par_lock(), 7563 Mutex::_no_safepoint_check_flag); 7564 // Remember the least grey address discarded 7565 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 7566 _collector->lower_restart_addr(ra); 7567 _overflow_stack->reset(); // discard stack contents 7568 _overflow_stack->expand(); // expand the stack if possible 7569 } 7570 7571 void PushOrMarkClosure::do_oop(oop obj) { 7572 // Ignore mark word because we are running concurrent with mutators. 7573 assert(obj->is_oop_or_null(true), "expected an oop or NULL"); 7574 HeapWord* addr = (HeapWord*)obj; 7575 if (_span.contains(addr) && !_bitMap->isMarked(addr)) { 7576 // Oop lies in _span and isn't yet grey or black 7577 _bitMap->mark(addr); // now grey 7578 if (addr < _finger) { 7579 // the bit map iteration has already either passed, or 7580 // sampled, this bit in the bit map; we'll need to 7581 // use the marking stack to scan this oop's oops. 7582 bool simulate_overflow = false; 7583 NOT_PRODUCT( 7584 if (CMSMarkStackOverflowALot && 7585 _collector->simulate_overflow()) { 7586 // simulate a stack overflow 7587 simulate_overflow = true; 7588 } 7589 ) 7590 if (simulate_overflow || !_markStack->push(obj)) { // stack overflow 7591 if (PrintCMSStatistics != 0) { 7592 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 7593 SIZE_FORMAT, _markStack->capacity()); 7594 } 7595 assert(simulate_overflow || _markStack->isFull(), "Else push should have succeeded"); 7596 handle_stack_overflow(addr); 7597 } 7598 } 7599 // anything including and to the right of _finger 7600 // will be scanned as we iterate over the remainder of the 7601 // bit map 7602 do_yield_check(); 7603 } 7604 } 7605 7606 void PushOrMarkClosure::do_oop(oop* p) { PushOrMarkClosure::do_oop_work(p); } 7607 void PushOrMarkClosure::do_oop(narrowOop* p) { PushOrMarkClosure::do_oop_work(p); } 7608 7609 void Par_PushOrMarkClosure::do_oop(oop obj) { 7610 // Ignore mark word because we are running concurrent with mutators. 7611 assert(obj->is_oop_or_null(true), "expected an oop or NULL"); 7612 HeapWord* addr = (HeapWord*)obj; 7613 if (_whole_span.contains(addr) && !_bit_map->isMarked(addr)) { 7614 // Oop lies in _span and isn't yet grey or black 7615 // We read the global_finger (volatile read) strictly after marking oop 7616 bool res = _bit_map->par_mark(addr); // now grey 7617 volatile HeapWord** gfa = (volatile HeapWord**)_global_finger_addr; 7618 // Should we push this marked oop on our stack? 7619 // -- if someone else marked it, nothing to do 7620 // -- if target oop is above global finger nothing to do 7621 // -- if target oop is in chunk and above local finger 7622 // then nothing to do 7623 // -- else push on work queue 7624 if ( !res // someone else marked it, they will deal with it 7625 || (addr >= *gfa) // will be scanned in a later task 7626 || (_span.contains(addr) && addr >= _finger)) { // later in this chunk 7627 return; 7628 } 7629 // the bit map iteration has already either passed, or 7630 // sampled, this bit in the bit map; we'll need to 7631 // use the marking stack to scan this oop's oops. 7632 bool simulate_overflow = false; 7633 NOT_PRODUCT( 7634 if (CMSMarkStackOverflowALot && 7635 _collector->simulate_overflow()) { 7636 // simulate a stack overflow 7637 simulate_overflow = true; 7638 } 7639 ) 7640 if (simulate_overflow || 7641 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) { 7642 // stack overflow 7643 if (PrintCMSStatistics != 0) { 7644 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 7645 SIZE_FORMAT, _overflow_stack->capacity()); 7646 } 7647 // We cannot assert that the overflow stack is full because 7648 // it may have been emptied since. 7649 assert(simulate_overflow || 7650 _work_queue->size() == _work_queue->max_elems(), 7651 "Else push should have succeeded"); 7652 handle_stack_overflow(addr); 7653 } 7654 do_yield_check(); 7655 } 7656 } 7657 7658 void Par_PushOrMarkClosure::do_oop(oop* p) { Par_PushOrMarkClosure::do_oop_work(p); } 7659 void Par_PushOrMarkClosure::do_oop(narrowOop* p) { Par_PushOrMarkClosure::do_oop_work(p); } 7660 7661 KlassRememberingOopClosure::KlassRememberingOopClosure(CMSCollector* collector, 7662 ReferenceProcessor* rp, 7663 CMSMarkStack* revisit_stack) : 7664 OopClosure(rp), 7665 _collector(collector), 7666 _revisit_stack(revisit_stack), 7667 _should_remember_klasses(collector->should_unload_classes()) {} 7668 7669 PushAndMarkClosure::PushAndMarkClosure(CMSCollector* collector, 7670 MemRegion span, 7671 ReferenceProcessor* rp, 7672 CMSBitMap* bit_map, 7673 CMSBitMap* mod_union_table, 7674 CMSMarkStack* mark_stack, 7675 CMSMarkStack* revisit_stack, 7676 bool concurrent_precleaning): 7677 KlassRememberingOopClosure(collector, rp, revisit_stack), 7678 _span(span), 7679 _bit_map(bit_map), 7680 _mod_union_table(mod_union_table), 7681 _mark_stack(mark_stack), 7682 _concurrent_precleaning(concurrent_precleaning) 7683 { 7684 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 7685 } 7686 7687 // Grey object rescan during pre-cleaning and second checkpoint phases -- 7688 // the non-parallel version (the parallel version appears further below.) 7689 void PushAndMarkClosure::do_oop(oop obj) { 7690 // Ignore mark word verification. If during concurrent precleaning, 7691 // the object monitor may be locked. If during the checkpoint 7692 // phases, the object may already have been reached by a different 7693 // path and may be at the end of the global overflow list (so 7694 // the mark word may be NULL). 7695 assert(obj->is_oop_or_null(true /* ignore mark word */), 7696 "expected an oop or NULL"); 7697 HeapWord* addr = (HeapWord*)obj; 7698 // Check if oop points into the CMS generation 7699 // and is not marked 7700 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 7701 // a white object ... 7702 _bit_map->mark(addr); // ... now grey 7703 // push on the marking stack (grey set) 7704 bool simulate_overflow = false; 7705 NOT_PRODUCT( 7706 if (CMSMarkStackOverflowALot && 7707 _collector->simulate_overflow()) { 7708 // simulate a stack overflow 7709 simulate_overflow = true; 7710 } 7711 ) 7712 if (simulate_overflow || !_mark_stack->push(obj)) { 7713 if (_concurrent_precleaning) { 7714 // During precleaning we can just dirty the appropriate card(s) 7715 // in the mod union table, thus ensuring that the object remains 7716 // in the grey set and continue. In the case of object arrays 7717 // we need to dirty all of the cards that the object spans, 7718 // since the rescan of object arrays will be limited to the 7719 // dirty cards. 7720 // Note that no one can be intefering with us in this action 7721 // of dirtying the mod union table, so no locking or atomics 7722 // are required. 7723 if (obj->is_objArray()) { 7724 size_t sz = obj->size(); 7725 HeapWord* end_card_addr = (HeapWord*)round_to( 7726 (intptr_t)(addr+sz), CardTableModRefBS::card_size); 7727 MemRegion redirty_range = MemRegion(addr, end_card_addr); 7728 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 7729 _mod_union_table->mark_range(redirty_range); 7730 } else { 7731 _mod_union_table->mark(addr); 7732 } 7733 _collector->_ser_pmc_preclean_ovflw++; 7734 } else { 7735 // During the remark phase, we need to remember this oop 7736 // in the overflow list. 7737 _collector->push_on_overflow_list(obj); 7738 _collector->_ser_pmc_remark_ovflw++; 7739 } 7740 } 7741 } 7742 } 7743 7744 Par_PushAndMarkClosure::Par_PushAndMarkClosure(CMSCollector* collector, 7745 MemRegion span, 7746 ReferenceProcessor* rp, 7747 CMSBitMap* bit_map, 7748 OopTaskQueue* work_queue, 7749 CMSMarkStack* revisit_stack): 7750 Par_KlassRememberingOopClosure(collector, rp, revisit_stack), 7751 _span(span), 7752 _bit_map(bit_map), 7753 _work_queue(work_queue) 7754 { 7755 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 7756 } 7757 7758 void PushAndMarkClosure::do_oop(oop* p) { PushAndMarkClosure::do_oop_work(p); } 7759 void PushAndMarkClosure::do_oop(narrowOop* p) { PushAndMarkClosure::do_oop_work(p); } 7760 7761 // Grey object rescan during second checkpoint phase -- 7762 // the parallel version. 7763 void Par_PushAndMarkClosure::do_oop(oop obj) { 7764 // In the assert below, we ignore the mark word because 7765 // this oop may point to an already visited object that is 7766 // on the overflow stack (in which case the mark word has 7767 // been hijacked for chaining into the overflow stack -- 7768 // if this is the last object in the overflow stack then 7769 // its mark word will be NULL). Because this object may 7770 // have been subsequently popped off the global overflow 7771 // stack, and the mark word possibly restored to the prototypical 7772 // value, by the time we get to examined this failing assert in 7773 // the debugger, is_oop_or_null(false) may subsequently start 7774 // to hold. 7775 assert(obj->is_oop_or_null(true), 7776 "expected an oop or NULL"); 7777 HeapWord* addr = (HeapWord*)obj; 7778 // Check if oop points into the CMS generation 7779 // and is not marked 7780 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 7781 // a white object ... 7782 // If we manage to "claim" the object, by being the 7783 // first thread to mark it, then we push it on our 7784 // marking stack 7785 if (_bit_map->par_mark(addr)) { // ... now grey 7786 // push on work queue (grey set) 7787 bool simulate_overflow = false; 7788 NOT_PRODUCT( 7789 if (CMSMarkStackOverflowALot && 7790 _collector->par_simulate_overflow()) { 7791 // simulate a stack overflow 7792 simulate_overflow = true; 7793 } 7794 ) 7795 if (simulate_overflow || !_work_queue->push(obj)) { 7796 _collector->par_push_on_overflow_list(obj); 7797 _collector->_par_pmc_remark_ovflw++; // imprecise OK: no need to CAS 7798 } 7799 } // Else, some other thread got there first 7800 } 7801 } 7802 7803 void Par_PushAndMarkClosure::do_oop(oop* p) { Par_PushAndMarkClosure::do_oop_work(p); } 7804 void Par_PushAndMarkClosure::do_oop(narrowOop* p) { Par_PushAndMarkClosure::do_oop_work(p); } 7805 7806 void PushAndMarkClosure::remember_mdo(DataLayout* v) { 7807 // TBD 7808 } 7809 7810 void Par_PushAndMarkClosure::remember_mdo(DataLayout* v) { 7811 // TBD 7812 } 7813 7814 void CMSPrecleanRefsYieldClosure::do_yield_work() { 7815 DEBUG_ONLY(RememberKlassesChecker mux(false);) 7816 Mutex* bml = _collector->bitMapLock(); 7817 assert_lock_strong(bml); 7818 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 7819 "CMS thread should hold CMS token"); 7820 7821 bml->unlock(); 7822 ConcurrentMarkSweepThread::desynchronize(true); 7823 7824 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7825 7826 _collector->stopTimer(); 7827 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 7828 if (PrintCMSStatistics != 0) { 7829 _collector->incrementYields(); 7830 } 7831 _collector->icms_wait(); 7832 7833 // See the comment in coordinator_yield() 7834 for (unsigned i = 0; i < CMSYieldSleepCount && 7835 ConcurrentMarkSweepThread::should_yield() && 7836 !CMSCollector::foregroundGCIsActive(); ++i) { 7837 os::sleep(Thread::current(), 1, false); 7838 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7839 } 7840 7841 ConcurrentMarkSweepThread::synchronize(true); 7842 bml->lock(); 7843 7844 _collector->startTimer(); 7845 } 7846 7847 bool CMSPrecleanRefsYieldClosure::should_return() { 7848 if (ConcurrentMarkSweepThread::should_yield()) { 7849 do_yield_work(); 7850 } 7851 return _collector->foregroundGCIsActive(); 7852 } 7853 7854 void MarkFromDirtyCardsClosure::do_MemRegion(MemRegion mr) { 7855 assert(((size_t)mr.start())%CardTableModRefBS::card_size_in_words == 0, 7856 "mr should be aligned to start at a card boundary"); 7857 // We'd like to assert: 7858 // assert(mr.word_size()%CardTableModRefBS::card_size_in_words == 0, 7859 // "mr should be a range of cards"); 7860 // However, that would be too strong in one case -- the last 7861 // partition ends at _unallocated_block which, in general, can be 7862 // an arbitrary boundary, not necessarily card aligned. 7863 if (PrintCMSStatistics != 0) { 7864 _num_dirty_cards += 7865 mr.word_size()/CardTableModRefBS::card_size_in_words; 7866 } 7867 _space->object_iterate_mem(mr, &_scan_cl); 7868 } 7869 7870 SweepClosure::SweepClosure(CMSCollector* collector, 7871 ConcurrentMarkSweepGeneration* g, 7872 CMSBitMap* bitMap, bool should_yield) : 7873 _collector(collector), 7874 _g(g), 7875 _sp(g->cmsSpace()), 7876 _limit(_sp->sweep_limit()), 7877 _freelistLock(_sp->freelistLock()), 7878 _bitMap(bitMap), 7879 _yield(should_yield), 7880 _inFreeRange(false), // No free range at beginning of sweep 7881 _freeRangeInFreeLists(false), // No free range at beginning of sweep 7882 _lastFreeRangeCoalesced(false), 7883 _freeFinger(g->used_region().start()) 7884 { 7885 NOT_PRODUCT( 7886 _numObjectsFreed = 0; 7887 _numWordsFreed = 0; 7888 _numObjectsLive = 0; 7889 _numWordsLive = 0; 7890 _numObjectsAlreadyFree = 0; 7891 _numWordsAlreadyFree = 0; 7892 _last_fc = NULL; 7893 7894 _sp->initializeIndexedFreeListArrayReturnedBytes(); 7895 _sp->dictionary()->initializeDictReturnedBytes(); 7896 ) 7897 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7898 "sweep _limit out of bounds"); 7899 if (CMSTraceSweeper) { 7900 gclog_or_tty->print("\n====================\nStarting new sweep\n"); 7901 } 7902 } 7903 7904 // We need this destructor to reclaim any space at the end 7905 // of the space, which do_blk below may not have added back to 7906 // the free lists. [basically dealing with the "fringe effect"] 7907 SweepClosure::~SweepClosure() { 7908 assert_lock_strong(_freelistLock); 7909 // this should be treated as the end of a free run if any 7910 // The current free range should be returned to the free lists 7911 // as one coalesced chunk. 7912 if (inFreeRange()) { 7913 flushCurFreeChunk(freeFinger(), 7914 pointer_delta(_limit, freeFinger())); 7915 assert(freeFinger() < _limit, "the finger pointeth off base"); 7916 if (CMSTraceSweeper) { 7917 gclog_or_tty->print("destructor:"); 7918 gclog_or_tty->print("Sweep:put_free_blk 0x%x ("SIZE_FORMAT") " 7919 "[coalesced:"SIZE_FORMAT"]\n", 7920 freeFinger(), pointer_delta(_limit, freeFinger()), 7921 lastFreeRangeCoalesced()); 7922 } 7923 } 7924 NOT_PRODUCT( 7925 if (Verbose && PrintGC) { 7926 gclog_or_tty->print("Collected "SIZE_FORMAT" objects, " 7927 SIZE_FORMAT " bytes", 7928 _numObjectsFreed, _numWordsFreed*sizeof(HeapWord)); 7929 gclog_or_tty->print_cr("\nLive "SIZE_FORMAT" objects, " 7930 SIZE_FORMAT" bytes " 7931 "Already free "SIZE_FORMAT" objects, "SIZE_FORMAT" bytes", 7932 _numObjectsLive, _numWordsLive*sizeof(HeapWord), 7933 _numObjectsAlreadyFree, _numWordsAlreadyFree*sizeof(HeapWord)); 7934 size_t totalBytes = (_numWordsFreed + _numWordsLive + _numWordsAlreadyFree) * 7935 sizeof(HeapWord); 7936 gclog_or_tty->print_cr("Total sweep: "SIZE_FORMAT" bytes", totalBytes); 7937 7938 if (PrintCMSStatistics && CMSVerifyReturnedBytes) { 7939 size_t indexListReturnedBytes = _sp->sumIndexedFreeListArrayReturnedBytes(); 7940 size_t dictReturnedBytes = _sp->dictionary()->sumDictReturnedBytes(); 7941 size_t returnedBytes = indexListReturnedBytes + dictReturnedBytes; 7942 gclog_or_tty->print("Returned "SIZE_FORMAT" bytes", returnedBytes); 7943 gclog_or_tty->print(" Indexed List Returned "SIZE_FORMAT" bytes", 7944 indexListReturnedBytes); 7945 gclog_or_tty->print_cr(" Dictionary Returned "SIZE_FORMAT" bytes", 7946 dictReturnedBytes); 7947 } 7948 } 7949 ) 7950 // Now, in debug mode, just null out the sweep_limit 7951 NOT_PRODUCT(_sp->clear_sweep_limit();) 7952 if (CMSTraceSweeper) { 7953 gclog_or_tty->print("end of sweep\n================\n"); 7954 } 7955 } 7956 7957 void SweepClosure::initialize_free_range(HeapWord* freeFinger, 7958 bool freeRangeInFreeLists) { 7959 if (CMSTraceSweeper) { 7960 gclog_or_tty->print("---- Start free range 0x%x with free block [%d] (%d)\n", 7961 freeFinger, _sp->block_size(freeFinger), 7962 freeRangeInFreeLists); 7963 } 7964 assert(!inFreeRange(), "Trampling existing free range"); 7965 set_inFreeRange(true); 7966 set_lastFreeRangeCoalesced(false); 7967 7968 set_freeFinger(freeFinger); 7969 set_freeRangeInFreeLists(freeRangeInFreeLists); 7970 if (CMSTestInFreeList) { 7971 if (freeRangeInFreeLists) { 7972 FreeChunk* fc = (FreeChunk*) freeFinger; 7973 assert(fc->isFree(), "A chunk on the free list should be free."); 7974 assert(fc->size() > 0, "Free range should have a size"); 7975 assert(_sp->verifyChunkInFreeLists(fc), "Chunk is not in free lists"); 7976 } 7977 } 7978 } 7979 7980 // Note that the sweeper runs concurrently with mutators. Thus, 7981 // it is possible for direct allocation in this generation to happen 7982 // in the middle of the sweep. Note that the sweeper also coalesces 7983 // contiguous free blocks. Thus, unless the sweeper and the allocator 7984 // synchronize appropriately freshly allocated blocks may get swept up. 7985 // This is accomplished by the sweeper locking the free lists while 7986 // it is sweeping. Thus blocks that are determined to be free are 7987 // indeed free. There is however one additional complication: 7988 // blocks that have been allocated since the final checkpoint and 7989 // mark, will not have been marked and so would be treated as 7990 // unreachable and swept up. To prevent this, the allocator marks 7991 // the bit map when allocating during the sweep phase. This leads, 7992 // however, to a further complication -- objects may have been allocated 7993 // but not yet initialized -- in the sense that the header isn't yet 7994 // installed. The sweeper can not then determine the size of the block 7995 // in order to skip over it. To deal with this case, we use a technique 7996 // (due to Printezis) to encode such uninitialized block sizes in the 7997 // bit map. Since the bit map uses a bit per every HeapWord, but the 7998 // CMS generation has a minimum object size of 3 HeapWords, it follows 7999 // that "normal marks" won't be adjacent in the bit map (there will 8000 // always be at least two 0 bits between successive 1 bits). We make use 8001 // of these "unused" bits to represent uninitialized blocks -- the bit 8002 // corresponding to the start of the uninitialized object and the next 8003 // bit are both set. Finally, a 1 bit marks the end of the object that 8004 // started with the two consecutive 1 bits to indicate its potentially 8005 // uninitialized state. 8006 8007 size_t SweepClosure::do_blk_careful(HeapWord* addr) { 8008 FreeChunk* fc = (FreeChunk*)addr; 8009 size_t res; 8010 8011 // Check if we are done sweeping. Below we check "addr >= _limit" rather 8012 // than "addr == _limit" because although _limit was a block boundary when 8013 // we started the sweep, it may no longer be one because heap expansion 8014 // may have caused us to coalesce the block ending at the address _limit 8015 // with a newly expanded chunk (this happens when _limit was set to the 8016 // previous _end of the space), so we may have stepped past _limit; see CR 6977970. 8017 if (addr >= _limit) { // we have swept up to or past the limit, do nothing more 8018 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 8019 "sweep _limit out of bounds"); 8020 assert(addr < _sp->end(), "addr out of bounds"); 8021 // help the closure application finish 8022 return pointer_delta(_sp->end(), addr); 8023 } 8024 assert(addr < _limit, "sweep invariant"); 8025 8026 // check if we should yield 8027 do_yield_check(addr); 8028 if (fc->isFree()) { 8029 // Chunk that is already free 8030 res = fc->size(); 8031 doAlreadyFreeChunk(fc); 8032 debug_only(_sp->verifyFreeLists()); 8033 assert(res == fc->size(), "Don't expect the size to change"); 8034 NOT_PRODUCT( 8035 _numObjectsAlreadyFree++; 8036 _numWordsAlreadyFree += res; 8037 ) 8038 NOT_PRODUCT(_last_fc = fc;) 8039 } else if (!_bitMap->isMarked(addr)) { 8040 // Chunk is fresh garbage 8041 res = doGarbageChunk(fc); 8042 debug_only(_sp->verifyFreeLists()); 8043 NOT_PRODUCT( 8044 _numObjectsFreed++; 8045 _numWordsFreed += res; 8046 ) 8047 } else { 8048 // Chunk that is alive. 8049 res = doLiveChunk(fc); 8050 debug_only(_sp->verifyFreeLists()); 8051 NOT_PRODUCT( 8052 _numObjectsLive++; 8053 _numWordsLive += res; 8054 ) 8055 } 8056 return res; 8057 } 8058 8059 // For the smart allocation, record following 8060 // split deaths - a free chunk is removed from its free list because 8061 // it is being split into two or more chunks. 8062 // split birth - a free chunk is being added to its free list because 8063 // a larger free chunk has been split and resulted in this free chunk. 8064 // coal death - a free chunk is being removed from its free list because 8065 // it is being coalesced into a large free chunk. 8066 // coal birth - a free chunk is being added to its free list because 8067 // it was created when two or more free chunks where coalesced into 8068 // this free chunk. 8069 // 8070 // These statistics are used to determine the desired number of free 8071 // chunks of a given size. The desired number is chosen to be relative 8072 // to the end of a CMS sweep. The desired number at the end of a sweep 8073 // is the 8074 // count-at-end-of-previous-sweep (an amount that was enough) 8075 // - count-at-beginning-of-current-sweep (the excess) 8076 // + split-births (gains in this size during interval) 8077 // - split-deaths (demands on this size during interval) 8078 // where the interval is from the end of one sweep to the end of the 8079 // next. 8080 // 8081 // When sweeping the sweeper maintains an accumulated chunk which is 8082 // the chunk that is made up of chunks that have been coalesced. That 8083 // will be termed the left-hand chunk. A new chunk of garbage that 8084 // is being considered for coalescing will be referred to as the 8085 // right-hand chunk. 8086 // 8087 // When making a decision on whether to coalesce a right-hand chunk with 8088 // the current left-hand chunk, the current count vs. the desired count 8089 // of the left-hand chunk is considered. Also if the right-hand chunk 8090 // is near the large chunk at the end of the heap (see 8091 // ConcurrentMarkSweepGeneration::isNearLargestChunk()), then the 8092 // left-hand chunk is coalesced. 8093 // 8094 // When making a decision about whether to split a chunk, the desired count 8095 // vs. the current count of the candidate to be split is also considered. 8096 // If the candidate is underpopulated (currently fewer chunks than desired) 8097 // a chunk of an overpopulated (currently more chunks than desired) size may 8098 // be chosen. The "hint" associated with a free list, if non-null, points 8099 // to a free list which may be overpopulated. 8100 // 8101 8102 void SweepClosure::doAlreadyFreeChunk(FreeChunk* fc) { 8103 size_t size = fc->size(); 8104 // Chunks that cannot be coalesced are not in the 8105 // free lists. 8106 if (CMSTestInFreeList && !fc->cantCoalesce()) { 8107 assert(_sp->verifyChunkInFreeLists(fc), 8108 "free chunk should be in free lists"); 8109 } 8110 // a chunk that is already free, should not have been 8111 // marked in the bit map 8112 HeapWord* addr = (HeapWord*) fc; 8113 assert(!_bitMap->isMarked(addr), "free chunk should be unmarked"); 8114 // Verify that the bit map has no bits marked between 8115 // addr and purported end of this block. 8116 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 8117 8118 // Some chunks cannot be coalesced in under any circumstances. 8119 // See the definition of cantCoalesce(). 8120 if (!fc->cantCoalesce()) { 8121 // This chunk can potentially be coalesced. 8122 if (_sp->adaptive_freelists()) { 8123 // All the work is done in 8124 doPostIsFreeOrGarbageChunk(fc, size); 8125 } else { // Not adaptive free lists 8126 // this is a free chunk that can potentially be coalesced by the sweeper; 8127 if (!inFreeRange()) { 8128 // if the next chunk is a free block that can't be coalesced 8129 // it doesn't make sense to remove this chunk from the free lists 8130 FreeChunk* nextChunk = (FreeChunk*)(addr + size); 8131 assert((HeapWord*)nextChunk <= _limit, "sweep invariant"); 8132 if ((HeapWord*)nextChunk < _limit && // there's a next chunk... 8133 nextChunk->isFree() && // which is free... 8134 nextChunk->cantCoalesce()) { // ... but cant be coalesced 8135 // nothing to do 8136 } else { 8137 // Potentially the start of a new free range: 8138 // Don't eagerly remove it from the free lists. 8139 // No need to remove it if it will just be put 8140 // back again. (Also from a pragmatic point of view 8141 // if it is a free block in a region that is beyond 8142 // any allocated blocks, an assertion will fail) 8143 // Remember the start of a free run. 8144 initialize_free_range(addr, true); 8145 // end - can coalesce with next chunk 8146 } 8147 } else { 8148 // the midst of a free range, we are coalescing 8149 debug_only(record_free_block_coalesced(fc);) 8150 if (CMSTraceSweeper) { 8151 gclog_or_tty->print(" -- pick up free block 0x%x (%d)\n", fc, size); 8152 } 8153 // remove it from the free lists 8154 _sp->removeFreeChunkFromFreeLists(fc); 8155 set_lastFreeRangeCoalesced(true); 8156 // If the chunk is being coalesced and the current free range is 8157 // in the free lists, remove the current free range so that it 8158 // will be returned to the free lists in its entirety - all 8159 // the coalesced pieces included. 8160 if (freeRangeInFreeLists()) { 8161 FreeChunk* ffc = (FreeChunk*) freeFinger(); 8162 assert(ffc->size() == pointer_delta(addr, freeFinger()), 8163 "Size of free range is inconsistent with chunk size."); 8164 if (CMSTestInFreeList) { 8165 assert(_sp->verifyChunkInFreeLists(ffc), 8166 "free range is not in free lists"); 8167 } 8168 _sp->removeFreeChunkFromFreeLists(ffc); 8169 set_freeRangeInFreeLists(false); 8170 } 8171 } 8172 } 8173 } else { 8174 // Code path common to both original and adaptive free lists. 8175 8176 // cant coalesce with previous block; this should be treated 8177 // as the end of a free run if any 8178 if (inFreeRange()) { 8179 // we kicked some butt; time to pick up the garbage 8180 assert(freeFinger() < addr, "the finger pointeth off base"); 8181 flushCurFreeChunk(freeFinger(), pointer_delta(addr, freeFinger())); 8182 } 8183 // else, nothing to do, just continue 8184 } 8185 } 8186 8187 size_t SweepClosure::doGarbageChunk(FreeChunk* fc) { 8188 // This is a chunk of garbage. It is not in any free list. 8189 // Add it to a free list or let it possibly be coalesced into 8190 // a larger chunk. 8191 HeapWord* addr = (HeapWord*) fc; 8192 size_t size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 8193 8194 if (_sp->adaptive_freelists()) { 8195 // Verify that the bit map has no bits marked between 8196 // addr and purported end of just dead object. 8197 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 8198 8199 doPostIsFreeOrGarbageChunk(fc, size); 8200 } else { 8201 if (!inFreeRange()) { 8202 // start of a new free range 8203 assert(size > 0, "A free range should have a size"); 8204 initialize_free_range(addr, false); 8205 8206 } else { 8207 // this will be swept up when we hit the end of the 8208 // free range 8209 if (CMSTraceSweeper) { 8210 gclog_or_tty->print(" -- pick up garbage 0x%x (%d) \n", fc, size); 8211 } 8212 // If the chunk is being coalesced and the current free range is 8213 // in the free lists, remove the current free range so that it 8214 // will be returned to the free lists in its entirety - all 8215 // the coalesced pieces included. 8216 if (freeRangeInFreeLists()) { 8217 FreeChunk* ffc = (FreeChunk*)freeFinger(); 8218 assert(ffc->size() == pointer_delta(addr, freeFinger()), 8219 "Size of free range is inconsistent with chunk size."); 8220 if (CMSTestInFreeList) { 8221 assert(_sp->verifyChunkInFreeLists(ffc), 8222 "free range is not in free lists"); 8223 } 8224 _sp->removeFreeChunkFromFreeLists(ffc); 8225 set_freeRangeInFreeLists(false); 8226 } 8227 set_lastFreeRangeCoalesced(true); 8228 } 8229 // this will be swept up when we hit the end of the free range 8230 8231 // Verify that the bit map has no bits marked between 8232 // addr and purported end of just dead object. 8233 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 8234 } 8235 return size; 8236 } 8237 8238 size_t SweepClosure::doLiveChunk(FreeChunk* fc) { 8239 HeapWord* addr = (HeapWord*) fc; 8240 // The sweeper has just found a live object. Return any accumulated 8241 // left hand chunk to the free lists. 8242 if (inFreeRange()) { 8243 if (_sp->adaptive_freelists()) { 8244 flushCurFreeChunk(freeFinger(), 8245 pointer_delta(addr, freeFinger())); 8246 } else { // not adaptive freelists 8247 set_inFreeRange(false); 8248 // Add the free range back to the free list if it is not already 8249 // there. 8250 if (!freeRangeInFreeLists()) { 8251 assert(freeFinger() < addr, "the finger pointeth off base"); 8252 if (CMSTraceSweeper) { 8253 gclog_or_tty->print("Sweep:put_free_blk 0x%x (%d) " 8254 "[coalesced:%d]\n", 8255 freeFinger(), pointer_delta(addr, freeFinger()), 8256 lastFreeRangeCoalesced()); 8257 } 8258 _sp->addChunkAndRepairOffsetTable(freeFinger(), 8259 pointer_delta(addr, freeFinger()), lastFreeRangeCoalesced()); 8260 } 8261 } 8262 } 8263 8264 // Common code path for original and adaptive free lists. 8265 8266 // this object is live: we'd normally expect this to be 8267 // an oop, and like to assert the following: 8268 // assert(oop(addr)->is_oop(), "live block should be an oop"); 8269 // However, as we commented above, this may be an object whose 8270 // header hasn't yet been initialized. 8271 size_t size; 8272 assert(_bitMap->isMarked(addr), "Tautology for this control point"); 8273 if (_bitMap->isMarked(addr + 1)) { 8274 // Determine the size from the bit map, rather than trying to 8275 // compute it from the object header. 8276 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 8277 size = pointer_delta(nextOneAddr + 1, addr); 8278 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 8279 "alignment problem"); 8280 8281 #ifdef DEBUG 8282 if (oop(addr)->klass_or_null() != NULL && 8283 ( !_collector->should_unload_classes() 8284 || (oop(addr)->is_parsable()) && 8285 oop(addr)->is_conc_safe())) { 8286 // Ignore mark word because we are running concurrent with mutators 8287 assert(oop(addr)->is_oop(true), "live block should be an oop"); 8288 // is_conc_safe is checked before performing this assertion 8289 // because an object that is not is_conc_safe may yet have 8290 // the return from size() correct. 8291 assert(size == 8292 CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()), 8293 "P-mark and computed size do not agree"); 8294 } 8295 #endif 8296 8297 } else { 8298 // This should be an initialized object that's alive. 8299 assert(oop(addr)->klass_or_null() != NULL && 8300 (!_collector->should_unload_classes() 8301 || oop(addr)->is_parsable()), 8302 "Should be an initialized object"); 8303 // Note that there are objects used during class redefinition 8304 // (e.g., merge_cp in VM_RedefineClasses::merge_cp_and_rewrite() 8305 // which are discarded with their is_conc_safe state still 8306 // false. These object may be floating garbage so may be 8307 // seen here. If they are floating garbage their size 8308 // should be attainable from their klass. Do not that 8309 // is_conc_safe() is true for oop(addr). 8310 // Ignore mark word because we are running concurrent with mutators 8311 assert(oop(addr)->is_oop(true), "live block should be an oop"); 8312 // Verify that the bit map has no bits marked between 8313 // addr and purported end of this block. 8314 size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 8315 assert(size >= 3, "Necessary for Printezis marks to work"); 8316 assert(!_bitMap->isMarked(addr+1), "Tautology for this control point"); 8317 DEBUG_ONLY(_bitMap->verifyNoOneBitsInRange(addr+2, addr+size);) 8318 } 8319 return size; 8320 } 8321 8322 void SweepClosure::doPostIsFreeOrGarbageChunk(FreeChunk* fc, 8323 size_t chunkSize) { 8324 // doPostIsFreeOrGarbageChunk() should only be called in the smart allocation 8325 // scheme. 8326 bool fcInFreeLists = fc->isFree(); 8327 assert(_sp->adaptive_freelists(), "Should only be used in this case."); 8328 assert((HeapWord*)fc <= _limit, "sweep invariant"); 8329 if (CMSTestInFreeList && fcInFreeLists) { 8330 assert(_sp->verifyChunkInFreeLists(fc), 8331 "free chunk is not in free lists"); 8332 } 8333 8334 8335 if (CMSTraceSweeper) { 8336 gclog_or_tty->print_cr(" -- pick up another chunk at 0x%x (%d)", fc, chunkSize); 8337 } 8338 8339 HeapWord* addr = (HeapWord*) fc; 8340 8341 bool coalesce; 8342 size_t left = pointer_delta(addr, freeFinger()); 8343 size_t right = chunkSize; 8344 switch (FLSCoalescePolicy) { 8345 // numeric value forms a coalition aggressiveness metric 8346 case 0: { // never coalesce 8347 coalesce = false; 8348 break; 8349 } 8350 case 1: { // coalesce if left & right chunks on overpopulated lists 8351 coalesce = _sp->coalOverPopulated(left) && 8352 _sp->coalOverPopulated(right); 8353 break; 8354 } 8355 case 2: { // coalesce if left chunk on overpopulated list (default) 8356 coalesce = _sp->coalOverPopulated(left); 8357 break; 8358 } 8359 case 3: { // coalesce if left OR right chunk on overpopulated list 8360 coalesce = _sp->coalOverPopulated(left) || 8361 _sp->coalOverPopulated(right); 8362 break; 8363 } 8364 case 4: { // always coalesce 8365 coalesce = true; 8366 break; 8367 } 8368 default: 8369 ShouldNotReachHere(); 8370 } 8371 8372 // Should the current free range be coalesced? 8373 // If the chunk is in a free range and either we decided to coalesce above 8374 // or the chunk is near the large block at the end of the heap 8375 // (isNearLargestChunk() returns true), then coalesce this chunk. 8376 bool doCoalesce = inFreeRange() && 8377 (coalesce || _g->isNearLargestChunk((HeapWord*)fc)); 8378 if (doCoalesce) { 8379 // Coalesce the current free range on the left with the new 8380 // chunk on the right. If either is on a free list, 8381 // it must be removed from the list and stashed in the closure. 8382 if (freeRangeInFreeLists()) { 8383 FreeChunk* ffc = (FreeChunk*)freeFinger(); 8384 assert(ffc->size() == pointer_delta(addr, freeFinger()), 8385 "Size of free range is inconsistent with chunk size."); 8386 if (CMSTestInFreeList) { 8387 assert(_sp->verifyChunkInFreeLists(ffc), 8388 "Chunk is not in free lists"); 8389 } 8390 _sp->coalDeath(ffc->size()); 8391 _sp->removeFreeChunkFromFreeLists(ffc); 8392 set_freeRangeInFreeLists(false); 8393 } 8394 if (fcInFreeLists) { 8395 _sp->coalDeath(chunkSize); 8396 assert(fc->size() == chunkSize, 8397 "The chunk has the wrong size or is not in the free lists"); 8398 _sp->removeFreeChunkFromFreeLists(fc); 8399 } 8400 set_lastFreeRangeCoalesced(true); 8401 } else { // not in a free range and/or should not coalesce 8402 // Return the current free range and start a new one. 8403 if (inFreeRange()) { 8404 // In a free range but cannot coalesce with the right hand chunk. 8405 // Put the current free range into the free lists. 8406 flushCurFreeChunk(freeFinger(), 8407 pointer_delta(addr, freeFinger())); 8408 } 8409 // Set up for new free range. Pass along whether the right hand 8410 // chunk is in the free lists. 8411 initialize_free_range((HeapWord*)fc, fcInFreeLists); 8412 } 8413 } 8414 void SweepClosure::flushCurFreeChunk(HeapWord* chunk, size_t size) { 8415 assert(inFreeRange(), "Should only be called if currently in a free range."); 8416 assert(size > 0, 8417 "A zero sized chunk cannot be added to the free lists."); 8418 if (!freeRangeInFreeLists()) { 8419 if(CMSTestInFreeList) { 8420 FreeChunk* fc = (FreeChunk*) chunk; 8421 fc->setSize(size); 8422 assert(!_sp->verifyChunkInFreeLists(fc), 8423 "chunk should not be in free lists yet"); 8424 } 8425 if (CMSTraceSweeper) { 8426 gclog_or_tty->print_cr(" -- add free block 0x%x (%d) to free lists", 8427 chunk, size); 8428 } 8429 // A new free range is going to be starting. The current 8430 // free range has not been added to the free lists yet or 8431 // was removed so add it back. 8432 // If the current free range was coalesced, then the death 8433 // of the free range was recorded. Record a birth now. 8434 if (lastFreeRangeCoalesced()) { 8435 _sp->coalBirth(size); 8436 } 8437 _sp->addChunkAndRepairOffsetTable(chunk, size, 8438 lastFreeRangeCoalesced()); 8439 } 8440 set_inFreeRange(false); 8441 set_freeRangeInFreeLists(false); 8442 } 8443 8444 // We take a break if we've been at this for a while, 8445 // so as to avoid monopolizing the locks involved. 8446 void SweepClosure::do_yield_work(HeapWord* addr) { 8447 // Return current free chunk being used for coalescing (if any) 8448 // to the appropriate freelist. After yielding, the next 8449 // free block encountered will start a coalescing range of 8450 // free blocks. If the next free block is adjacent to the 8451 // chunk just flushed, they will need to wait for the next 8452 // sweep to be coalesced. 8453 if (inFreeRange()) { 8454 flushCurFreeChunk(freeFinger(), pointer_delta(addr, freeFinger())); 8455 } 8456 8457 // First give up the locks, then yield, then re-lock. 8458 // We should probably use a constructor/destructor idiom to 8459 // do this unlock/lock or modify the MutexUnlocker class to 8460 // serve our purpose. XXX 8461 assert_lock_strong(_bitMap->lock()); 8462 assert_lock_strong(_freelistLock); 8463 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 8464 "CMS thread should hold CMS token"); 8465 _bitMap->lock()->unlock(); 8466 _freelistLock->unlock(); 8467 ConcurrentMarkSweepThread::desynchronize(true); 8468 ConcurrentMarkSweepThread::acknowledge_yield_request(); 8469 _collector->stopTimer(); 8470 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 8471 if (PrintCMSStatistics != 0) { 8472 _collector->incrementYields(); 8473 } 8474 _collector->icms_wait(); 8475 8476 // See the comment in coordinator_yield() 8477 for (unsigned i = 0; i < CMSYieldSleepCount && 8478 ConcurrentMarkSweepThread::should_yield() && 8479 !CMSCollector::foregroundGCIsActive(); ++i) { 8480 os::sleep(Thread::current(), 1, false); 8481 ConcurrentMarkSweepThread::acknowledge_yield_request(); 8482 } 8483 8484 ConcurrentMarkSweepThread::synchronize(true); 8485 _freelistLock->lock(); 8486 _bitMap->lock()->lock_without_safepoint_check(); 8487 _collector->startTimer(); 8488 } 8489 8490 #ifndef PRODUCT 8491 // This is actually very useful in a product build if it can 8492 // be called from the debugger. Compile it into the product 8493 // as needed. 8494 bool debug_verifyChunkInFreeLists(FreeChunk* fc) { 8495 return debug_cms_space->verifyChunkInFreeLists(fc); 8496 } 8497 8498 void SweepClosure::record_free_block_coalesced(FreeChunk* fc) const { 8499 if (CMSTraceSweeper) { 8500 gclog_or_tty->print("Sweep:coal_free_blk 0x%x (%d)\n", fc, fc->size()); 8501 } 8502 } 8503 #endif 8504 8505 // CMSIsAliveClosure 8506 bool CMSIsAliveClosure::do_object_b(oop obj) { 8507 HeapWord* addr = (HeapWord*)obj; 8508 return addr != NULL && 8509 (!_span.contains(addr) || _bit_map->isMarked(addr)); 8510 } 8511 8512 CMSKeepAliveClosure::CMSKeepAliveClosure( CMSCollector* collector, 8513 MemRegion span, 8514 CMSBitMap* bit_map, CMSMarkStack* mark_stack, 8515 CMSMarkStack* revisit_stack, bool cpc): 8516 KlassRememberingOopClosure(collector, NULL, revisit_stack), 8517 _span(span), 8518 _bit_map(bit_map), 8519 _mark_stack(mark_stack), 8520 _concurrent_precleaning(cpc) { 8521 assert(!_span.is_empty(), "Empty span could spell trouble"); 8522 } 8523 8524 8525 // CMSKeepAliveClosure: the serial version 8526 void CMSKeepAliveClosure::do_oop(oop obj) { 8527 HeapWord* addr = (HeapWord*)obj; 8528 if (_span.contains(addr) && 8529 !_bit_map->isMarked(addr)) { 8530 _bit_map->mark(addr); 8531 bool simulate_overflow = false; 8532 NOT_PRODUCT( 8533 if (CMSMarkStackOverflowALot && 8534 _collector->simulate_overflow()) { 8535 // simulate a stack overflow 8536 simulate_overflow = true; 8537 } 8538 ) 8539 if (simulate_overflow || !_mark_stack->push(obj)) { 8540 if (_concurrent_precleaning) { 8541 // We dirty the overflown object and let the remark 8542 // phase deal with it. 8543 assert(_collector->overflow_list_is_empty(), "Error"); 8544 // In the case of object arrays, we need to dirty all of 8545 // the cards that the object spans. No locking or atomics 8546 // are needed since no one else can be mutating the mod union 8547 // table. 8548 if (obj->is_objArray()) { 8549 size_t sz = obj->size(); 8550 HeapWord* end_card_addr = 8551 (HeapWord*)round_to((intptr_t)(addr+sz), CardTableModRefBS::card_size); 8552 MemRegion redirty_range = MemRegion(addr, end_card_addr); 8553 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 8554 _collector->_modUnionTable.mark_range(redirty_range); 8555 } else { 8556 _collector->_modUnionTable.mark(addr); 8557 } 8558 _collector->_ser_kac_preclean_ovflw++; 8559 } else { 8560 _collector->push_on_overflow_list(obj); 8561 _collector->_ser_kac_ovflw++; 8562 } 8563 } 8564 } 8565 } 8566 8567 void CMSKeepAliveClosure::do_oop(oop* p) { CMSKeepAliveClosure::do_oop_work(p); } 8568 void CMSKeepAliveClosure::do_oop(narrowOop* p) { CMSKeepAliveClosure::do_oop_work(p); } 8569 8570 // CMSParKeepAliveClosure: a parallel version of the above. 8571 // The work queues are private to each closure (thread), 8572 // but (may be) available for stealing by other threads. 8573 void CMSParKeepAliveClosure::do_oop(oop obj) { 8574 HeapWord* addr = (HeapWord*)obj; 8575 if (_span.contains(addr) && 8576 !_bit_map->isMarked(addr)) { 8577 // In general, during recursive tracing, several threads 8578 // may be concurrently getting here; the first one to 8579 // "tag" it, claims it. 8580 if (_bit_map->par_mark(addr)) { 8581 bool res = _work_queue->push(obj); 8582 assert(res, "Low water mark should be much less than capacity"); 8583 // Do a recursive trim in the hope that this will keep 8584 // stack usage lower, but leave some oops for potential stealers 8585 trim_queue(_low_water_mark); 8586 } // Else, another thread got there first 8587 } 8588 } 8589 8590 void CMSParKeepAliveClosure::do_oop(oop* p) { CMSParKeepAliveClosure::do_oop_work(p); } 8591 void CMSParKeepAliveClosure::do_oop(narrowOop* p) { CMSParKeepAliveClosure::do_oop_work(p); } 8592 8593 void CMSParKeepAliveClosure::trim_queue(uint max) { 8594 while (_work_queue->size() > max) { 8595 oop new_oop; 8596 if (_work_queue->pop_local(new_oop)) { 8597 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 8598 assert(_bit_map->isMarked((HeapWord*)new_oop), 8599 "no white objects on this stack!"); 8600 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 8601 // iterate over the oops in this oop, marking and pushing 8602 // the ones in CMS heap (i.e. in _span). 8603 new_oop->oop_iterate(&_mark_and_push); 8604 } 8605 } 8606 } 8607 8608 CMSInnerParMarkAndPushClosure::CMSInnerParMarkAndPushClosure( 8609 CMSCollector* collector, 8610 MemRegion span, CMSBitMap* bit_map, 8611 CMSMarkStack* revisit_stack, 8612 OopTaskQueue* work_queue): 8613 Par_KlassRememberingOopClosure(collector, NULL, revisit_stack), 8614 _span(span), 8615 _bit_map(bit_map), 8616 _work_queue(work_queue) { } 8617 8618 void CMSInnerParMarkAndPushClosure::do_oop(oop obj) { 8619 HeapWord* addr = (HeapWord*)obj; 8620 if (_span.contains(addr) && 8621 !_bit_map->isMarked(addr)) { 8622 if (_bit_map->par_mark(addr)) { 8623 bool simulate_overflow = false; 8624 NOT_PRODUCT( 8625 if (CMSMarkStackOverflowALot && 8626 _collector->par_simulate_overflow()) { 8627 // simulate a stack overflow 8628 simulate_overflow = true; 8629 } 8630 ) 8631 if (simulate_overflow || !_work_queue->push(obj)) { 8632 _collector->par_push_on_overflow_list(obj); 8633 _collector->_par_kac_ovflw++; 8634 } 8635 } // Else another thread got there already 8636 } 8637 } 8638 8639 void CMSInnerParMarkAndPushClosure::do_oop(oop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); } 8640 void CMSInnerParMarkAndPushClosure::do_oop(narrowOop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); } 8641 8642 ////////////////////////////////////////////////////////////////// 8643 // CMSExpansionCause ///////////////////////////// 8644 ////////////////////////////////////////////////////////////////// 8645 const char* CMSExpansionCause::to_string(CMSExpansionCause::Cause cause) { 8646 switch (cause) { 8647 case _no_expansion: 8648 return "No expansion"; 8649 case _satisfy_free_ratio: 8650 return "Free ratio"; 8651 case _satisfy_promotion: 8652 return "Satisfy promotion"; 8653 case _satisfy_allocation: 8654 return "allocation"; 8655 case _allocate_par_lab: 8656 return "Par LAB"; 8657 case _allocate_par_spooling_space: 8658 return "Par Spooling Space"; 8659 case _adaptive_size_policy: 8660 return "Ergonomics"; 8661 default: 8662 return "unknown"; 8663 } 8664 } 8665 8666 void CMSDrainMarkingStackClosure::do_void() { 8667 // the max number to take from overflow list at a time 8668 const size_t num = _mark_stack->capacity()/4; 8669 assert(!_concurrent_precleaning || _collector->overflow_list_is_empty(), 8670 "Overflow list should be NULL during concurrent phases"); 8671 while (!_mark_stack->isEmpty() || 8672 // if stack is empty, check the overflow list 8673 _collector->take_from_overflow_list(num, _mark_stack)) { 8674 oop obj = _mark_stack->pop(); 8675 HeapWord* addr = (HeapWord*)obj; 8676 assert(_span.contains(addr), "Should be within span"); 8677 assert(_bit_map->isMarked(addr), "Should be marked"); 8678 assert(obj->is_oop(), "Should be an oop"); 8679 obj->oop_iterate(_keep_alive); 8680 } 8681 } 8682 8683 void CMSParDrainMarkingStackClosure::do_void() { 8684 // drain queue 8685 trim_queue(0); 8686 } 8687 8688 // Trim our work_queue so its length is below max at return 8689 void CMSParDrainMarkingStackClosure::trim_queue(uint max) { 8690 while (_work_queue->size() > max) { 8691 oop new_oop; 8692 if (_work_queue->pop_local(new_oop)) { 8693 assert(new_oop->is_oop(), "Expected an oop"); 8694 assert(_bit_map->isMarked((HeapWord*)new_oop), 8695 "no white objects on this stack!"); 8696 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 8697 // iterate over the oops in this oop, marking and pushing 8698 // the ones in CMS heap (i.e. in _span). 8699 new_oop->oop_iterate(&_mark_and_push); 8700 } 8701 } 8702 } 8703 8704 //////////////////////////////////////////////////////////////////// 8705 // Support for Marking Stack Overflow list handling and related code 8706 //////////////////////////////////////////////////////////////////// 8707 // Much of the following code is similar in shape and spirit to the 8708 // code used in ParNewGC. We should try and share that code 8709 // as much as possible in the future. 8710 8711 #ifndef PRODUCT 8712 // Debugging support for CMSStackOverflowALot 8713 8714 // It's OK to call this multi-threaded; the worst thing 8715 // that can happen is that we'll get a bunch of closely 8716 // spaced simulated oveflows, but that's OK, in fact 8717 // probably good as it would exercise the overflow code 8718 // under contention. 8719 bool CMSCollector::simulate_overflow() { 8720 if (_overflow_counter-- <= 0) { // just being defensive 8721 _overflow_counter = CMSMarkStackOverflowInterval; 8722 return true; 8723 } else { 8724 return false; 8725 } 8726 } 8727 8728 bool CMSCollector::par_simulate_overflow() { 8729 return simulate_overflow(); 8730 } 8731 #endif 8732 8733 // Single-threaded 8734 bool CMSCollector::take_from_overflow_list(size_t num, CMSMarkStack* stack) { 8735 assert(stack->isEmpty(), "Expected precondition"); 8736 assert(stack->capacity() > num, "Shouldn't bite more than can chew"); 8737 size_t i = num; 8738 oop cur = _overflow_list; 8739 const markOop proto = markOopDesc::prototype(); 8740 NOT_PRODUCT(ssize_t n = 0;) 8741 for (oop next; i > 0 && cur != NULL; cur = next, i--) { 8742 next = oop(cur->mark()); 8743 cur->set_mark(proto); // until proven otherwise 8744 assert(cur->is_oop(), "Should be an oop"); 8745 bool res = stack->push(cur); 8746 assert(res, "Bit off more than can chew?"); 8747 NOT_PRODUCT(n++;) 8748 } 8749 _overflow_list = cur; 8750 #ifndef PRODUCT 8751 assert(_num_par_pushes >= n, "Too many pops?"); 8752 _num_par_pushes -=n; 8753 #endif 8754 return !stack->isEmpty(); 8755 } 8756 8757 #define BUSY (oop(0x1aff1aff)) 8758 // (MT-safe) Get a prefix of at most "num" from the list. 8759 // The overflow list is chained through the mark word of 8760 // each object in the list. We fetch the entire list, 8761 // break off a prefix of the right size and return the 8762 // remainder. If other threads try to take objects from 8763 // the overflow list at that time, they will wait for 8764 // some time to see if data becomes available. If (and 8765 // only if) another thread places one or more object(s) 8766 // on the global list before we have returned the suffix 8767 // to the global list, we will walk down our local list 8768 // to find its end and append the global list to 8769 // our suffix before returning it. This suffix walk can 8770 // prove to be expensive (quadratic in the amount of traffic) 8771 // when there are many objects in the overflow list and 8772 // there is much producer-consumer contention on the list. 8773 // *NOTE*: The overflow list manipulation code here and 8774 // in ParNewGeneration:: are very similar in shape, 8775 // except that in the ParNew case we use the old (from/eden) 8776 // copy of the object to thread the list via its klass word. 8777 // Because of the common code, if you make any changes in 8778 // the code below, please check the ParNew version to see if 8779 // similar changes might be needed. 8780 // CR 6797058 has been filed to consolidate the common code. 8781 bool CMSCollector::par_take_from_overflow_list(size_t num, 8782 OopTaskQueue* work_q, 8783 int no_of_gc_threads) { 8784 assert(work_q->size() == 0, "First empty local work queue"); 8785 assert(num < work_q->max_elems(), "Can't bite more than we can chew"); 8786 if (_overflow_list == NULL) { 8787 return false; 8788 } 8789 // Grab the entire list; we'll put back a suffix 8790 oop prefix = (oop)Atomic::xchg_ptr(BUSY, &_overflow_list); 8791 Thread* tid = Thread::current(); 8792 // Before "no_of_gc_threads" was introduced CMSOverflowSpinCount was 8793 // set to ParallelGCThreads. 8794 size_t CMSOverflowSpinCount = (size_t) no_of_gc_threads; // was ParallelGCThreads; 8795 size_t sleep_time_millis = MAX2((size_t)1, num/100); 8796 // If the list is busy, we spin for a short while, 8797 // sleeping between attempts to get the list. 8798 for (size_t spin = 0; prefix == BUSY && spin < CMSOverflowSpinCount; spin++) { 8799 os::sleep(tid, sleep_time_millis, false); 8800 if (_overflow_list == NULL) { 8801 // Nothing left to take 8802 return false; 8803 } else if (_overflow_list != BUSY) { 8804 // Try and grab the prefix 8805 prefix = (oop)Atomic::xchg_ptr(BUSY, &_overflow_list); 8806 } 8807 } 8808 // If the list was found to be empty, or we spun long 8809 // enough, we give up and return empty-handed. If we leave 8810 // the list in the BUSY state below, it must be the case that 8811 // some other thread holds the overflow list and will set it 8812 // to a non-BUSY state in the future. 8813 if (prefix == NULL || prefix == BUSY) { 8814 // Nothing to take or waited long enough 8815 if (prefix == NULL) { 8816 // Write back the NULL in case we overwrote it with BUSY above 8817 // and it is still the same value. 8818 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY); 8819 } 8820 return false; 8821 } 8822 assert(prefix != NULL && prefix != BUSY, "Error"); 8823 size_t i = num; 8824 oop cur = prefix; 8825 // Walk down the first "num" objects, unless we reach the end. 8826 for (; i > 1 && cur->mark() != NULL; cur = oop(cur->mark()), i--); 8827 if (cur->mark() == NULL) { 8828 // We have "num" or fewer elements in the list, so there 8829 // is nothing to return to the global list. 8830 // Write back the NULL in lieu of the BUSY we wrote 8831 // above, if it is still the same value. 8832 if (_overflow_list == BUSY) { 8833 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY); 8834 } 8835 } else { 8836 // Chop off the suffix and rerturn it to the global list. 8837 assert(cur->mark() != BUSY, "Error"); 8838 oop suffix_head = cur->mark(); // suffix will be put back on global list 8839 cur->set_mark(NULL); // break off suffix 8840 // It's possible that the list is still in the empty(busy) state 8841 // we left it in a short while ago; in that case we may be 8842 // able to place back the suffix without incurring the cost 8843 // of a walk down the list. 8844 oop observed_overflow_list = _overflow_list; 8845 oop cur_overflow_list = observed_overflow_list; 8846 bool attached = false; 8847 while (observed_overflow_list == BUSY || observed_overflow_list == NULL) { 8848 observed_overflow_list = 8849 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list); 8850 if (cur_overflow_list == observed_overflow_list) { 8851 attached = true; 8852 break; 8853 } else cur_overflow_list = observed_overflow_list; 8854 } 8855 if (!attached) { 8856 // Too bad, someone else sneaked in (at least) an element; we'll need 8857 // to do a splice. Find tail of suffix so we can prepend suffix to global 8858 // list. 8859 for (cur = suffix_head; cur->mark() != NULL; cur = (oop)(cur->mark())); 8860 oop suffix_tail = cur; 8861 assert(suffix_tail != NULL && suffix_tail->mark() == NULL, 8862 "Tautology"); 8863 observed_overflow_list = _overflow_list; 8864 do { 8865 cur_overflow_list = observed_overflow_list; 8866 if (cur_overflow_list != BUSY) { 8867 // Do the splice ... 8868 suffix_tail->set_mark(markOop(cur_overflow_list)); 8869 } else { // cur_overflow_list == BUSY 8870 suffix_tail->set_mark(NULL); 8871 } 8872 // ... and try to place spliced list back on overflow_list ... 8873 observed_overflow_list = 8874 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list); 8875 } while (cur_overflow_list != observed_overflow_list); 8876 // ... until we have succeeded in doing so. 8877 } 8878 } 8879 8880 // Push the prefix elements on work_q 8881 assert(prefix != NULL, "control point invariant"); 8882 const markOop proto = markOopDesc::prototype(); 8883 oop next; 8884 NOT_PRODUCT(ssize_t n = 0;) 8885 for (cur = prefix; cur != NULL; cur = next) { 8886 next = oop(cur->mark()); 8887 cur->set_mark(proto); // until proven otherwise 8888 assert(cur->is_oop(), "Should be an oop"); 8889 bool res = work_q->push(cur); 8890 assert(res, "Bit off more than we can chew?"); 8891 NOT_PRODUCT(n++;) 8892 } 8893 #ifndef PRODUCT 8894 assert(_num_par_pushes >= n, "Too many pops?"); 8895 Atomic::add_ptr(-(intptr_t)n, &_num_par_pushes); 8896 #endif 8897 return true; 8898 } 8899 8900 // Single-threaded 8901 void CMSCollector::push_on_overflow_list(oop p) { 8902 NOT_PRODUCT(_num_par_pushes++;) 8903 assert(p->is_oop(), "Not an oop"); 8904 preserve_mark_if_necessary(p); 8905 p->set_mark((markOop)_overflow_list); 8906 _overflow_list = p; 8907 } 8908 8909 // Multi-threaded; use CAS to prepend to overflow list 8910 void CMSCollector::par_push_on_overflow_list(oop p) { 8911 NOT_PRODUCT(Atomic::inc_ptr(&_num_par_pushes);) 8912 assert(p->is_oop(), "Not an oop"); 8913 par_preserve_mark_if_necessary(p); 8914 oop observed_overflow_list = _overflow_list; 8915 oop cur_overflow_list; 8916 do { 8917 cur_overflow_list = observed_overflow_list; 8918 if (cur_overflow_list != BUSY) { 8919 p->set_mark(markOop(cur_overflow_list)); 8920 } else { 8921 p->set_mark(NULL); 8922 } 8923 observed_overflow_list = 8924 (oop) Atomic::cmpxchg_ptr(p, &_overflow_list, cur_overflow_list); 8925 } while (cur_overflow_list != observed_overflow_list); 8926 } 8927 #undef BUSY 8928 8929 // Single threaded 8930 // General Note on GrowableArray: pushes may silently fail 8931 // because we are (temporarily) out of C-heap for expanding 8932 // the stack. The problem is quite ubiquitous and affects 8933 // a lot of code in the JVM. The prudent thing for GrowableArray 8934 // to do (for now) is to exit with an error. However, that may 8935 // be too draconian in some cases because the caller may be 8936 // able to recover without much harm. For such cases, we 8937 // should probably introduce a "soft_push" method which returns 8938 // an indication of success or failure with the assumption that 8939 // the caller may be able to recover from a failure; code in 8940 // the VM can then be changed, incrementally, to deal with such 8941 // failures where possible, thus, incrementally hardening the VM 8942 // in such low resource situations. 8943 void CMSCollector::preserve_mark_work(oop p, markOop m) { 8944 _preserved_oop_stack.push(p); 8945 _preserved_mark_stack.push(m); 8946 assert(m == p->mark(), "Mark word changed"); 8947 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), 8948 "bijection"); 8949 } 8950 8951 // Single threaded 8952 void CMSCollector::preserve_mark_if_necessary(oop p) { 8953 markOop m = p->mark(); 8954 if (m->must_be_preserved(p)) { 8955 preserve_mark_work(p, m); 8956 } 8957 } 8958 8959 void CMSCollector::par_preserve_mark_if_necessary(oop p) { 8960 markOop m = p->mark(); 8961 if (m->must_be_preserved(p)) { 8962 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag); 8963 // Even though we read the mark word without holding 8964 // the lock, we are assured that it will not change 8965 // because we "own" this oop, so no other thread can 8966 // be trying to push it on the overflow list; see 8967 // the assertion in preserve_mark_work() that checks 8968 // that m == p->mark(). 8969 preserve_mark_work(p, m); 8970 } 8971 } 8972 8973 // We should be able to do this multi-threaded, 8974 // a chunk of stack being a task (this is 8975 // correct because each oop only ever appears 8976 // once in the overflow list. However, it's 8977 // not very easy to completely overlap this with 8978 // other operations, so will generally not be done 8979 // until all work's been completed. Because we 8980 // expect the preserved oop stack (set) to be small, 8981 // it's probably fine to do this single-threaded. 8982 // We can explore cleverer concurrent/overlapped/parallel 8983 // processing of preserved marks if we feel the 8984 // need for this in the future. Stack overflow should 8985 // be so rare in practice and, when it happens, its 8986 // effect on performance so great that this will 8987 // likely just be in the noise anyway. 8988 void CMSCollector::restore_preserved_marks_if_any() { 8989 assert(SafepointSynchronize::is_at_safepoint(), 8990 "world should be stopped"); 8991 assert(Thread::current()->is_ConcurrentGC_thread() || 8992 Thread::current()->is_VM_thread(), 8993 "should be single-threaded"); 8994 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), 8995 "bijection"); 8996 8997 while (!_preserved_oop_stack.is_empty()) { 8998 oop p = _preserved_oop_stack.pop(); 8999 assert(p->is_oop(), "Should be an oop"); 9000 assert(_span.contains(p), "oop should be in _span"); 9001 assert(p->mark() == markOopDesc::prototype(), 9002 "Set when taken from overflow list"); 9003 markOop m = _preserved_mark_stack.pop(); 9004 p->set_mark(m); 9005 } 9006 assert(_preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(), 9007 "stacks were cleared above"); 9008 } 9009 9010 #ifndef PRODUCT 9011 bool CMSCollector::no_preserved_marks() const { 9012 return _preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(); 9013 } 9014 #endif 9015 9016 CMSAdaptiveSizePolicy* ASConcurrentMarkSweepGeneration::cms_size_policy() const 9017 { 9018 GenCollectedHeap* gch = (GenCollectedHeap*) GenCollectedHeap::heap(); 9019 CMSAdaptiveSizePolicy* size_policy = 9020 (CMSAdaptiveSizePolicy*) gch->gen_policy()->size_policy(); 9021 assert(size_policy->is_gc_cms_adaptive_size_policy(), 9022 "Wrong type for size policy"); 9023 return size_policy; 9024 } 9025 9026 void ASConcurrentMarkSweepGeneration::resize(size_t cur_promo_size, 9027 size_t desired_promo_size) { 9028 if (cur_promo_size < desired_promo_size) { 9029 size_t expand_bytes = desired_promo_size - cur_promo_size; 9030 if (PrintAdaptiveSizePolicy && Verbose) { 9031 gclog_or_tty->print_cr(" ASConcurrentMarkSweepGeneration::resize " 9032 "Expanding tenured generation by " SIZE_FORMAT " (bytes)", 9033 expand_bytes); 9034 } 9035 expand(expand_bytes, 9036 MinHeapDeltaBytes, 9037 CMSExpansionCause::_adaptive_size_policy); 9038 } else if (desired_promo_size < cur_promo_size) { 9039 size_t shrink_bytes = cur_promo_size - desired_promo_size; 9040 if (PrintAdaptiveSizePolicy && Verbose) { 9041 gclog_or_tty->print_cr(" ASConcurrentMarkSweepGeneration::resize " 9042 "Shrinking tenured generation by " SIZE_FORMAT " (bytes)", 9043 shrink_bytes); 9044 } 9045 shrink(shrink_bytes); 9046 } 9047 } 9048 9049 CMSGCAdaptivePolicyCounters* ASConcurrentMarkSweepGeneration::gc_adaptive_policy_counters() { 9050 GenCollectedHeap* gch = GenCollectedHeap::heap(); 9051 CMSGCAdaptivePolicyCounters* counters = 9052 (CMSGCAdaptivePolicyCounters*) gch->collector_policy()->counters(); 9053 assert(counters->kind() == GCPolicyCounters::CMSGCAdaptivePolicyCountersKind, 9054 "Wrong kind of counters"); 9055 return counters; 9056 } 9057 9058 9059 void ASConcurrentMarkSweepGeneration::update_counters() { 9060 if (UsePerfData) { 9061 _space_counters->update_all(); 9062 _gen_counters->update_all(); 9063 CMSGCAdaptivePolicyCounters* counters = gc_adaptive_policy_counters(); 9064 GenCollectedHeap* gch = GenCollectedHeap::heap(); 9065 CMSGCStats* gc_stats_l = (CMSGCStats*) gc_stats(); 9066 assert(gc_stats_l->kind() == GCStats::CMSGCStatsKind, 9067 "Wrong gc statistics type"); 9068 counters->update_counters(gc_stats_l); 9069 } 9070 } 9071 9072 void ASConcurrentMarkSweepGeneration::update_counters(size_t used) { 9073 if (UsePerfData) { 9074 _space_counters->update_used(used); 9075 _space_counters->update_capacity(); 9076 _gen_counters->update_all(); 9077 9078 CMSGCAdaptivePolicyCounters* counters = gc_adaptive_policy_counters(); 9079 GenCollectedHeap* gch = GenCollectedHeap::heap(); 9080 CMSGCStats* gc_stats_l = (CMSGCStats*) gc_stats(); 9081 assert(gc_stats_l->kind() == GCStats::CMSGCStatsKind, 9082 "Wrong gc statistics type"); 9083 counters->update_counters(gc_stats_l); 9084 } 9085 } 9086 9087 // The desired expansion delta is computed so that: 9088 // . desired free percentage or greater is used 9089 void ASConcurrentMarkSweepGeneration::compute_new_size() { 9090 assert_locked_or_safepoint(Heap_lock); 9091 9092 GenCollectedHeap* gch = (GenCollectedHeap*) GenCollectedHeap::heap(); 9093 9094 // If incremental collection failed, we just want to expand 9095 // to the limit. 9096 if (incremental_collection_failed()) { 9097 clear_incremental_collection_failed(); 9098 grow_to_reserved(); 9099 return; 9100 } 9101 9102 assert(UseAdaptiveSizePolicy, "Should be using adaptive sizing"); 9103 9104 assert(gch->kind() == CollectedHeap::GenCollectedHeap, 9105 "Wrong type of heap"); 9106 int prev_level = level() - 1; 9107 assert(prev_level >= 0, "The cms generation is the lowest generation"); 9108 Generation* prev_gen = gch->get_gen(prev_level); 9109 assert(prev_gen->kind() == Generation::ASParNew, 9110 "Wrong type of young generation"); 9111 ParNewGeneration* younger_gen = (ParNewGeneration*) prev_gen; 9112 size_t cur_eden = younger_gen->eden()->capacity(); 9113 CMSAdaptiveSizePolicy* size_policy = cms_size_policy(); 9114 size_t cur_promo = free(); 9115 size_policy->compute_tenured_generation_free_space(cur_promo, 9116 max_available(), 9117 cur_eden); 9118 resize(cur_promo, size_policy->promo_size()); 9119 9120 // Record the new size of the space in the cms generation 9121 // that is available for promotions. This is temporary. 9122 // It should be the desired promo size. 9123 size_policy->avg_cms_promo()->sample(free()); 9124 size_policy->avg_old_live()->sample(used()); 9125 9126 if (UsePerfData) { 9127 CMSGCAdaptivePolicyCounters* counters = gc_adaptive_policy_counters(); 9128 counters->update_cms_capacity_counter(capacity()); 9129 } 9130 } 9131 9132 void ASConcurrentMarkSweepGeneration::shrink_by(size_t desired_bytes) { 9133 assert_locked_or_safepoint(Heap_lock); 9134 assert_lock_strong(freelistLock()); 9135 HeapWord* old_end = _cmsSpace->end(); 9136 HeapWord* unallocated_start = _cmsSpace->unallocated_block(); 9137 assert(old_end >= unallocated_start, "Miscalculation of unallocated_start"); 9138 FreeChunk* chunk_at_end = find_chunk_at_end(); 9139 if (chunk_at_end == NULL) { 9140 // No room to shrink 9141 if (PrintGCDetails && Verbose) { 9142 gclog_or_tty->print_cr("No room to shrink: old_end " 9143 PTR_FORMAT " unallocated_start " PTR_FORMAT 9144 " chunk_at_end " PTR_FORMAT, 9145 old_end, unallocated_start, chunk_at_end); 9146 } 9147 return; 9148 } else { 9149 9150 // Find the chunk at the end of the space and determine 9151 // how much it can be shrunk. 9152 size_t shrinkable_size_in_bytes = chunk_at_end->size(); 9153 size_t aligned_shrinkable_size_in_bytes = 9154 align_size_down(shrinkable_size_in_bytes, os::vm_page_size()); 9155 assert(unallocated_start <= chunk_at_end->end(), 9156 "Inconsistent chunk at end of space"); 9157 size_t bytes = MIN2(desired_bytes, aligned_shrinkable_size_in_bytes); 9158 size_t word_size_before = heap_word_size(_virtual_space.committed_size()); 9159 9160 // Shrink the underlying space 9161 _virtual_space.shrink_by(bytes); 9162 if (PrintGCDetails && Verbose) { 9163 gclog_or_tty->print_cr("ConcurrentMarkSweepGeneration::shrink_by:" 9164 " desired_bytes " SIZE_FORMAT 9165 " shrinkable_size_in_bytes " SIZE_FORMAT 9166 " aligned_shrinkable_size_in_bytes " SIZE_FORMAT 9167 " bytes " SIZE_FORMAT, 9168 desired_bytes, shrinkable_size_in_bytes, 9169 aligned_shrinkable_size_in_bytes, bytes); 9170 gclog_or_tty->print_cr(" old_end " SIZE_FORMAT 9171 " unallocated_start " SIZE_FORMAT, 9172 old_end, unallocated_start); 9173 } 9174 9175 // If the space did shrink (shrinking is not guaranteed), 9176 // shrink the chunk at the end by the appropriate amount. 9177 if (((HeapWord*)_virtual_space.high()) < old_end) { 9178 size_t new_word_size = 9179 heap_word_size(_virtual_space.committed_size()); 9180 9181 // Have to remove the chunk from the dictionary because it is changing 9182 // size and might be someplace elsewhere in the dictionary. 9183 9184 // Get the chunk at end, shrink it, and put it 9185 // back. 9186 _cmsSpace->removeChunkFromDictionary(chunk_at_end); 9187 size_t word_size_change = word_size_before - new_word_size; 9188 size_t chunk_at_end_old_size = chunk_at_end->size(); 9189 assert(chunk_at_end_old_size >= word_size_change, 9190 "Shrink is too large"); 9191 chunk_at_end->setSize(chunk_at_end_old_size - 9192 word_size_change); 9193 _cmsSpace->freed((HeapWord*) chunk_at_end->end(), 9194 word_size_change); 9195 9196 _cmsSpace->returnChunkToDictionary(chunk_at_end); 9197 9198 MemRegion mr(_cmsSpace->bottom(), new_word_size); 9199 _bts->resize(new_word_size); // resize the block offset shared array 9200 Universe::heap()->barrier_set()->resize_covered_region(mr); 9201 _cmsSpace->assert_locked(); 9202 _cmsSpace->set_end((HeapWord*)_virtual_space.high()); 9203 9204 NOT_PRODUCT(_cmsSpace->dictionary()->verify()); 9205 9206 // update the space and generation capacity counters 9207 if (UsePerfData) { 9208 _space_counters->update_capacity(); 9209 _gen_counters->update_all(); 9210 } 9211 9212 if (Verbose && PrintGCDetails) { 9213 size_t new_mem_size = _virtual_space.committed_size(); 9214 size_t old_mem_size = new_mem_size + bytes; 9215 gclog_or_tty->print_cr("Shrinking %s from %ldK by %ldK to %ldK", 9216 name(), old_mem_size/K, bytes/K, new_mem_size/K); 9217 } 9218 } 9219 9220 assert(_cmsSpace->unallocated_block() <= _cmsSpace->end(), 9221 "Inconsistency at end of space"); 9222 assert(chunk_at_end->end() == _cmsSpace->end(), 9223 "Shrinking is inconsistent"); 9224 return; 9225 } 9226 } 9227 9228 // Transfer some number of overflown objects to usual marking 9229 // stack. Return true if some objects were transferred. 9230 bool MarkRefsIntoAndScanClosure::take_from_overflow_list() { 9231 size_t num = MIN2((size_t)(_mark_stack->capacity() - _mark_stack->length())/4, 9232 (size_t)ParGCDesiredObjsFromOverflowList); 9233 9234 bool res = _collector->take_from_overflow_list(num, _mark_stack); 9235 assert(_collector->overflow_list_is_empty() || res, 9236 "If list is not empty, we should have taken something"); 9237 assert(!res || !_mark_stack->isEmpty(), 9238 "If we took something, it should now be on our stack"); 9239 return res; 9240 } 9241 9242 size_t MarkDeadObjectsClosure::do_blk(HeapWord* addr) { 9243 size_t res = _sp->block_size_no_stall(addr, _collector); 9244 assert(res != 0, "Should always be able to compute a size"); 9245 if (_sp->block_is_obj(addr)) { 9246 if (_live_bit_map->isMarked(addr)) { 9247 // It can't have been dead in a previous cycle 9248 guarantee(!_dead_bit_map->isMarked(addr), "No resurrection!"); 9249 } else { 9250 _dead_bit_map->mark(addr); // mark the dead object 9251 } 9252 } 9253 return res; 9254 } 9255 9256 TraceCMSMemoryManagerStats::TraceCMSMemoryManagerStats(CMSCollector::CollectorState phase): TraceMemoryManagerStats() { 9257 9258 switch (phase) { 9259 case CMSCollector::InitialMarking: 9260 initialize(true /* fullGC */ , 9261 true /* recordGCBeginTime */, 9262 true /* recordPreGCUsage */, 9263 false /* recordPeakUsage */, 9264 false /* recordPostGCusage */, 9265 true /* recordAccumulatedGCTime */, 9266 false /* recordGCEndTime */, 9267 false /* countCollection */ ); 9268 break; 9269 9270 case CMSCollector::FinalMarking: 9271 initialize(true /* fullGC */ , 9272 false /* recordGCBeginTime */, 9273 false /* recordPreGCUsage */, 9274 false /* recordPeakUsage */, 9275 false /* recordPostGCusage */, 9276 true /* recordAccumulatedGCTime */, 9277 false /* recordGCEndTime */, 9278 false /* countCollection */ ); 9279 break; 9280 9281 case CMSCollector::Sweeping: 9282 initialize(true /* fullGC */ , 9283 false /* recordGCBeginTime */, 9284 false /* recordPreGCUsage */, 9285 true /* recordPeakUsage */, 9286 true /* recordPostGCusage */, 9287 false /* recordAccumulatedGCTime */, 9288 true /* recordGCEndTime */, 9289 true /* countCollection */ ); 9290 break; 9291 9292 default: 9293 ShouldNotReachHere(); 9294 } 9295 } 9296 9297 // when bailing out of cms in concurrent mode failure 9298 TraceCMSMemoryManagerStats::TraceCMSMemoryManagerStats(): TraceMemoryManagerStats() { 9299 initialize(true /* fullGC */ , 9300 true /* recordGCBeginTime */, 9301 true /* recordPreGCUsage */, 9302 true /* recordPeakUsage */, 9303 true /* recordPostGCusage */, 9304 true /* recordAccumulatedGCTime */, 9305 true /* recordGCEndTime */, 9306 true /* countCollection */ ); 9307 }