1 /* 2 * Copyright (c) 2001, 2012, 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 "precompiled.hpp" 26 #include "classfile/symbolTable.hpp" 27 #include "classfile/systemDictionary.hpp" 28 #include "code/codeCache.hpp" 29 #include "gc_implementation/concurrentMarkSweep/cmsAdaptiveSizePolicy.hpp" 30 #include "gc_implementation/concurrentMarkSweep/cmsCollectorPolicy.hpp" 31 #include "gc_implementation/concurrentMarkSweep/cmsGCAdaptivePolicyCounters.hpp" 32 #include "gc_implementation/concurrentMarkSweep/cmsOopClosures.inline.hpp" 33 #include "gc_implementation/concurrentMarkSweep/compactibleFreeListSpace.hpp" 34 #include "gc_implementation/concurrentMarkSweep/concurrentMarkSweepGeneration.inline.hpp" 35 #include "gc_implementation/concurrentMarkSweep/concurrentMarkSweepThread.hpp" 36 #include "gc_implementation/concurrentMarkSweep/vmCMSOperations.hpp" 37 #include "gc_implementation/parNew/parNewGeneration.hpp" 38 #include "gc_implementation/shared/collectorCounters.hpp" 39 #include "gc_implementation/shared/isGCActiveMark.hpp" 40 #include "gc_interface/collectedHeap.inline.hpp" 41 #include "memory/cardTableRS.hpp" 42 #include "memory/collectorPolicy.hpp" 43 #include "memory/gcLocker.inline.hpp" 44 #include "memory/genCollectedHeap.hpp" 45 #include "memory/genMarkSweep.hpp" 46 #include "memory/genOopClosures.inline.hpp" 47 #include "memory/iterator.hpp" 48 #include "memory/referencePolicy.hpp" 49 #include "memory/resourceArea.hpp" 50 #include "oops/oop.inline.hpp" 51 #include "prims/jvmtiExport.hpp" 52 #include "runtime/globals_extension.hpp" 53 #include "runtime/handles.inline.hpp" 54 #include "runtime/java.hpp" 55 #include "runtime/vmThread.hpp" 56 #include "services/memoryService.hpp" 57 #include "services/runtimeService.hpp" 58 59 // statics 60 CMSCollector* ConcurrentMarkSweepGeneration::_collector = NULL; 61 bool CMSCollector::_full_gc_requested = false; 62 63 ////////////////////////////////////////////////////////////////// 64 // In support of CMS/VM thread synchronization 65 ////////////////////////////////////////////////////////////////// 66 // We split use of the CGC_lock into 2 "levels". 67 // The low-level locking is of the usual CGC_lock monitor. We introduce 68 // a higher level "token" (hereafter "CMS token") built on top of the 69 // low level monitor (hereafter "CGC lock"). 70 // The token-passing protocol gives priority to the VM thread. The 71 // CMS-lock doesn't provide any fairness guarantees, but clients 72 // should ensure that it is only held for very short, bounded 73 // durations. 74 // 75 // When either of the CMS thread or the VM thread is involved in 76 // collection operations during which it does not want the other 77 // thread to interfere, it obtains the CMS token. 78 // 79 // If either thread tries to get the token while the other has 80 // it, that thread waits. However, if the VM thread and CMS thread 81 // both want the token, then the VM thread gets priority while the 82 // CMS thread waits. This ensures, for instance, that the "concurrent" 83 // phases of the CMS thread's work do not block out the VM thread 84 // for long periods of time as the CMS thread continues to hog 85 // the token. (See bug 4616232). 86 // 87 // The baton-passing functions are, however, controlled by the 88 // flags _foregroundGCShouldWait and _foregroundGCIsActive, 89 // and here the low-level CMS lock, not the high level token, 90 // ensures mutual exclusion. 91 // 92 // Two important conditions that we have to satisfy: 93 // 1. if a thread does a low-level wait on the CMS lock, then it 94 // relinquishes the CMS token if it were holding that token 95 // when it acquired the low-level CMS lock. 96 // 2. any low-level notifications on the low-level lock 97 // should only be sent when a thread has relinquished the token. 98 // 99 // In the absence of either property, we'd have potential deadlock. 100 // 101 // We protect each of the CMS (concurrent and sequential) phases 102 // with the CMS _token_, not the CMS _lock_. 103 // 104 // The only code protected by CMS lock is the token acquisition code 105 // itself, see ConcurrentMarkSweepThread::[de]synchronize(), and the 106 // baton-passing code. 107 // 108 // Unfortunately, i couldn't come up with a good abstraction to factor and 109 // hide the naked CGC_lock manipulation in the baton-passing code 110 // further below. That's something we should try to do. Also, the proof 111 // of correctness of this 2-level locking scheme is far from obvious, 112 // and potentially quite slippery. We have an uneasy supsicion, for instance, 113 // that there may be a theoretical possibility of delay/starvation in the 114 // low-level lock/wait/notify scheme used for the baton-passing because of 115 // potential intereference with the priority scheme embodied in the 116 // CMS-token-passing protocol. See related comments at a CGC_lock->wait() 117 // invocation further below and marked with "XXX 20011219YSR". 118 // Indeed, as we note elsewhere, this may become yet more slippery 119 // in the presence of multiple CMS and/or multiple VM threads. XXX 120 121 class CMSTokenSync: public StackObj { 122 private: 123 bool _is_cms_thread; 124 public: 125 CMSTokenSync(bool is_cms_thread): 126 _is_cms_thread(is_cms_thread) { 127 assert(is_cms_thread == Thread::current()->is_ConcurrentGC_thread(), 128 "Incorrect argument to constructor"); 129 ConcurrentMarkSweepThread::synchronize(_is_cms_thread); 130 } 131 132 ~CMSTokenSync() { 133 assert(_is_cms_thread ? 134 ConcurrentMarkSweepThread::cms_thread_has_cms_token() : 135 ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 136 "Incorrect state"); 137 ConcurrentMarkSweepThread::desynchronize(_is_cms_thread); 138 } 139 }; 140 141 // Convenience class that does a CMSTokenSync, and then acquires 142 // upto three locks. 143 class CMSTokenSyncWithLocks: public CMSTokenSync { 144 private: 145 // Note: locks are acquired in textual declaration order 146 // and released in the opposite order 147 MutexLockerEx _locker1, _locker2, _locker3; 148 public: 149 CMSTokenSyncWithLocks(bool is_cms_thread, Mutex* mutex1, 150 Mutex* mutex2 = NULL, Mutex* mutex3 = NULL): 151 CMSTokenSync(is_cms_thread), 152 _locker1(mutex1, Mutex::_no_safepoint_check_flag), 153 _locker2(mutex2, Mutex::_no_safepoint_check_flag), 154 _locker3(mutex3, Mutex::_no_safepoint_check_flag) 155 { } 156 }; 157 158 159 // Wrapper class to temporarily disable icms during a foreground cms collection. 160 class ICMSDisabler: public StackObj { 161 public: 162 // The ctor disables icms and wakes up the thread so it notices the change; 163 // the dtor re-enables icms. Note that the CMSCollector methods will check 164 // CMSIncrementalMode. 165 ICMSDisabler() { CMSCollector::disable_icms(); CMSCollector::start_icms(); } 166 ~ICMSDisabler() { CMSCollector::enable_icms(); } 167 }; 168 169 ////////////////////////////////////////////////////////////////// 170 // Concurrent Mark-Sweep Generation ///////////////////////////// 171 ////////////////////////////////////////////////////////////////// 172 173 NOT_PRODUCT(CompactibleFreeListSpace* debug_cms_space;) 174 175 // This struct contains per-thread things necessary to support parallel 176 // young-gen collection. 177 class CMSParGCThreadState: public CHeapObj { 178 public: 179 CFLS_LAB lab; 180 PromotionInfo promo; 181 182 // Constructor. 183 CMSParGCThreadState(CompactibleFreeListSpace* cfls) : lab(cfls) { 184 promo.setSpace(cfls); 185 } 186 }; 187 188 ConcurrentMarkSweepGeneration::ConcurrentMarkSweepGeneration( 189 ReservedSpace rs, size_t initial_byte_size, int level, 190 CardTableRS* ct, bool use_adaptive_freelists, 191 FreeBlockDictionary<FreeChunk>::DictionaryChoice dictionaryChoice) : 192 CardGeneration(rs, initial_byte_size, level, ct), 193 _dilatation_factor(((double)MinChunkSize)/((double)(CollectedHeap::min_fill_size()))), 194 _debug_collection_type(Concurrent_collection_type) 195 { 196 HeapWord* bottom = (HeapWord*) _virtual_space.low(); 197 HeapWord* end = (HeapWord*) _virtual_space.high(); 198 199 _direct_allocated_words = 0; 200 NOT_PRODUCT( 201 _numObjectsPromoted = 0; 202 _numWordsPromoted = 0; 203 _numObjectsAllocated = 0; 204 _numWordsAllocated = 0; 205 ) 206 207 _cmsSpace = new CompactibleFreeListSpace(_bts, MemRegion(bottom, end), 208 use_adaptive_freelists, 209 dictionaryChoice); 210 NOT_PRODUCT(debug_cms_space = _cmsSpace;) 211 if (_cmsSpace == NULL) { 212 vm_exit_during_initialization( 213 "CompactibleFreeListSpace allocation failure"); 214 } 215 _cmsSpace->_gen = this; 216 217 _gc_stats = new CMSGCStats(); 218 219 // Verify the assumption that FreeChunk::_prev and OopDesc::_klass 220 // offsets match. The ability to tell free chunks from objects 221 // depends on this property. 222 debug_only( 223 FreeChunk* junk = NULL; 224 assert(UseCompressedOops || 225 junk->prev_addr() == (void*)(oop(junk)->klass_addr()), 226 "Offset of FreeChunk::_prev within FreeChunk must match" 227 " that of OopDesc::_klass within OopDesc"); 228 ) 229 if (CollectedHeap::use_parallel_gc_threads()) { 230 typedef CMSParGCThreadState* CMSParGCThreadStatePtr; 231 _par_gc_thread_states = 232 NEW_C_HEAP_ARRAY(CMSParGCThreadStatePtr, ParallelGCThreads); 233 if (_par_gc_thread_states == NULL) { 234 vm_exit_during_initialization("Could not allocate par gc structs"); 235 } 236 for (uint i = 0; i < ParallelGCThreads; i++) { 237 _par_gc_thread_states[i] = new CMSParGCThreadState(cmsSpace()); 238 if (_par_gc_thread_states[i] == NULL) { 239 vm_exit_during_initialization("Could not allocate par gc structs"); 240 } 241 } 242 } else { 243 _par_gc_thread_states = NULL; 244 } 245 _incremental_collection_failed = false; 246 // The "dilatation_factor" is the expansion that can occur on 247 // account of the fact that the minimum object size in the CMS 248 // generation may be larger than that in, say, a contiguous young 249 // generation. 250 // Ideally, in the calculation below, we'd compute the dilatation 251 // factor as: MinChunkSize/(promoting_gen's min object size) 252 // Since we do not have such a general query interface for the 253 // promoting generation, we'll instead just use the mimimum 254 // object size (which today is a header's worth of space); 255 // note that all arithmetic is in units of HeapWords. 256 assert(MinChunkSize >= CollectedHeap::min_fill_size(), "just checking"); 257 assert(_dilatation_factor >= 1.0, "from previous assert"); 258 } 259 260 261 // The field "_initiating_occupancy" represents the occupancy percentage 262 // at which we trigger a new collection cycle. Unless explicitly specified 263 // via CMSInitiating[Perm]OccupancyFraction (argument "io" below), it 264 // is calculated by: 265 // 266 // Let "f" be MinHeapFreeRatio in 267 // 268 // _intiating_occupancy = 100-f + 269 // f * (CMSTrigger[Perm]Ratio/100) 270 // where CMSTrigger[Perm]Ratio is the argument "tr" below. 271 // 272 // That is, if we assume the heap is at its desired maximum occupancy at the 273 // end of a collection, we let CMSTrigger[Perm]Ratio of the (purported) free 274 // space be allocated before initiating a new collection cycle. 275 // 276 void ConcurrentMarkSweepGeneration::init_initiating_occupancy(intx io, intx tr) { 277 assert(io <= 100 && tr >= 0 && tr <= 100, "Check the arguments"); 278 if (io >= 0) { 279 _initiating_occupancy = (double)io / 100.0; 280 } else { 281 _initiating_occupancy = ((100 - MinHeapFreeRatio) + 282 (double)(tr * MinHeapFreeRatio) / 100.0) 283 / 100.0; 284 } 285 } 286 287 void ConcurrentMarkSweepGeneration::ref_processor_init() { 288 assert(collector() != NULL, "no collector"); 289 collector()->ref_processor_init(); 290 } 291 292 void CMSCollector::ref_processor_init() { 293 if (_ref_processor == NULL) { 294 // Allocate and initialize a reference processor 295 _ref_processor = 296 new ReferenceProcessor(_span, // span 297 (ParallelGCThreads > 1) && ParallelRefProcEnabled, // mt processing 298 (int) ParallelGCThreads, // mt processing degree 299 _cmsGen->refs_discovery_is_mt(), // mt discovery 300 (int) MAX2(ConcGCThreads, ParallelGCThreads), // mt discovery degree 301 _cmsGen->refs_discovery_is_atomic(), // discovery is not atomic 302 &_is_alive_closure, // closure for liveness info 303 false); // next field updates do not need write barrier 304 // Initialize the _ref_processor field of CMSGen 305 _cmsGen->set_ref_processor(_ref_processor); 306 307 // Allocate a dummy ref processor for perm gen. 308 ReferenceProcessor* rp2 = new ReferenceProcessor(); 309 if (rp2 == NULL) { 310 vm_exit_during_initialization("Could not allocate ReferenceProcessor object"); 311 } 312 _permGen->set_ref_processor(rp2); 313 } 314 } 315 316 CMSAdaptiveSizePolicy* CMSCollector::size_policy() { 317 GenCollectedHeap* gch = GenCollectedHeap::heap(); 318 assert(gch->kind() == CollectedHeap::GenCollectedHeap, 319 "Wrong type of heap"); 320 CMSAdaptiveSizePolicy* sp = (CMSAdaptiveSizePolicy*) 321 gch->gen_policy()->size_policy(); 322 assert(sp->is_gc_cms_adaptive_size_policy(), 323 "Wrong type of size policy"); 324 return sp; 325 } 326 327 CMSGCAdaptivePolicyCounters* CMSCollector::gc_adaptive_policy_counters() { 328 CMSGCAdaptivePolicyCounters* results = 329 (CMSGCAdaptivePolicyCounters*) collector_policy()->counters(); 330 assert( 331 results->kind() == GCPolicyCounters::CMSGCAdaptivePolicyCountersKind, 332 "Wrong gc policy counter kind"); 333 return results; 334 } 335 336 337 void ConcurrentMarkSweepGeneration::initialize_performance_counters() { 338 339 const char* gen_name = "old"; 340 341 // Generation Counters - generation 1, 1 subspace 342 _gen_counters = new GenerationCounters(gen_name, 1, 1, &_virtual_space); 343 344 _space_counters = new GSpaceCounters(gen_name, 0, 345 _virtual_space.reserved_size(), 346 this, _gen_counters); 347 } 348 349 CMSStats::CMSStats(ConcurrentMarkSweepGeneration* cms_gen, unsigned int alpha): 350 _cms_gen(cms_gen) 351 { 352 assert(alpha <= 100, "bad value"); 353 _saved_alpha = alpha; 354 355 // Initialize the alphas to the bootstrap value of 100. 356 _gc0_alpha = _cms_alpha = 100; 357 358 _cms_begin_time.update(); 359 _cms_end_time.update(); 360 361 _gc0_duration = 0.0; 362 _gc0_period = 0.0; 363 _gc0_promoted = 0; 364 365 _cms_duration = 0.0; 366 _cms_period = 0.0; 367 _cms_allocated = 0; 368 369 _cms_used_at_gc0_begin = 0; 370 _cms_used_at_gc0_end = 0; 371 _allow_duty_cycle_reduction = false; 372 _valid_bits = 0; 373 _icms_duty_cycle = CMSIncrementalDutyCycle; 374 } 375 376 double CMSStats::cms_free_adjustment_factor(size_t free) const { 377 // TBD: CR 6909490 378 return 1.0; 379 } 380 381 void CMSStats::adjust_cms_free_adjustment_factor(bool fail, size_t free) { 382 } 383 384 // If promotion failure handling is on use 385 // the padded average size of the promotion for each 386 // young generation collection. 387 double CMSStats::time_until_cms_gen_full() const { 388 size_t cms_free = _cms_gen->cmsSpace()->free(); 389 GenCollectedHeap* gch = GenCollectedHeap::heap(); 390 size_t expected_promotion = MIN2(gch->get_gen(0)->capacity(), 391 (size_t) _cms_gen->gc_stats()->avg_promoted()->padded_average()); 392 if (cms_free > expected_promotion) { 393 // Start a cms collection if there isn't enough space to promote 394 // for the next minor collection. Use the padded average as 395 // a safety factor. 396 cms_free -= expected_promotion; 397 398 // Adjust by the safety factor. 399 double cms_free_dbl = (double)cms_free; 400 double cms_adjustment = (100.0 - CMSIncrementalSafetyFactor)/100.0; 401 // Apply a further correction factor which tries to adjust 402 // for recent occurance of concurrent mode failures. 403 cms_adjustment = cms_adjustment * cms_free_adjustment_factor(cms_free); 404 cms_free_dbl = cms_free_dbl * cms_adjustment; 405 406 if (PrintGCDetails && Verbose) { 407 gclog_or_tty->print_cr("CMSStats::time_until_cms_gen_full: cms_free " 408 SIZE_FORMAT " expected_promotion " SIZE_FORMAT, 409 cms_free, expected_promotion); 410 gclog_or_tty->print_cr(" cms_free_dbl %f cms_consumption_rate %f", 411 cms_free_dbl, cms_consumption_rate() + 1.0); 412 } 413 // Add 1 in case the consumption rate goes to zero. 414 return cms_free_dbl / (cms_consumption_rate() + 1.0); 415 } 416 return 0.0; 417 } 418 419 // Compare the duration of the cms collection to the 420 // time remaining before the cms generation is empty. 421 // Note that the time from the start of the cms collection 422 // to the start of the cms sweep (less than the total 423 // duration of the cms collection) can be used. This 424 // has been tried and some applications experienced 425 // promotion failures early in execution. This was 426 // possibly because the averages were not accurate 427 // enough at the beginning. 428 double CMSStats::time_until_cms_start() const { 429 // We add "gc0_period" to the "work" calculation 430 // below because this query is done (mostly) at the 431 // end of a scavenge, so we need to conservatively 432 // account for that much possible delay 433 // in the query so as to avoid concurrent mode failures 434 // due to starting the collection just a wee bit too 435 // late. 436 double work = cms_duration() + gc0_period(); 437 double deadline = time_until_cms_gen_full(); 438 // If a concurrent mode failure occurred recently, we want to be 439 // more conservative and halve our expected time_until_cms_gen_full() 440 if (work > deadline) { 441 if (Verbose && PrintGCDetails) { 442 gclog_or_tty->print( 443 " CMSCollector: collect because of anticipated promotion " 444 "before full %3.7f + %3.7f > %3.7f ", cms_duration(), 445 gc0_period(), time_until_cms_gen_full()); 446 } 447 return 0.0; 448 } 449 return work - deadline; 450 } 451 452 // Return a duty cycle based on old_duty_cycle and new_duty_cycle, limiting the 453 // amount of change to prevent wild oscillation. 454 unsigned int CMSStats::icms_damped_duty_cycle(unsigned int old_duty_cycle, 455 unsigned int new_duty_cycle) { 456 assert(old_duty_cycle <= 100, "bad input value"); 457 assert(new_duty_cycle <= 100, "bad input value"); 458 459 // Note: use subtraction with caution since it may underflow (values are 460 // unsigned). Addition is safe since we're in the range 0-100. 461 unsigned int damped_duty_cycle = new_duty_cycle; 462 if (new_duty_cycle < old_duty_cycle) { 463 const unsigned int largest_delta = MAX2(old_duty_cycle / 4, 5U); 464 if (new_duty_cycle + largest_delta < old_duty_cycle) { 465 damped_duty_cycle = old_duty_cycle - largest_delta; 466 } 467 } else if (new_duty_cycle > old_duty_cycle) { 468 const unsigned int largest_delta = MAX2(old_duty_cycle / 4, 15U); 469 if (new_duty_cycle > old_duty_cycle + largest_delta) { 470 damped_duty_cycle = MIN2(old_duty_cycle + largest_delta, 100U); 471 } 472 } 473 assert(damped_duty_cycle <= 100, "invalid duty cycle computed"); 474 475 if (CMSTraceIncrementalPacing) { 476 gclog_or_tty->print(" [icms_damped_duty_cycle(%d,%d) = %d] ", 477 old_duty_cycle, new_duty_cycle, damped_duty_cycle); 478 } 479 return damped_duty_cycle; 480 } 481 482 unsigned int CMSStats::icms_update_duty_cycle_impl() { 483 assert(CMSIncrementalPacing && valid(), 484 "should be handled in icms_update_duty_cycle()"); 485 486 double cms_time_so_far = cms_timer().seconds(); 487 double scaled_duration = cms_duration_per_mb() * _cms_used_at_gc0_end / M; 488 double scaled_duration_remaining = fabsd(scaled_duration - cms_time_so_far); 489 490 // Avoid division by 0. 491 double time_until_full = MAX2(time_until_cms_gen_full(), 0.01); 492 double duty_cycle_dbl = 100.0 * scaled_duration_remaining / time_until_full; 493 494 unsigned int new_duty_cycle = MIN2((unsigned int)duty_cycle_dbl, 100U); 495 if (new_duty_cycle > _icms_duty_cycle) { 496 // Avoid very small duty cycles (1 or 2); 0 is allowed. 497 if (new_duty_cycle > 2) { 498 _icms_duty_cycle = icms_damped_duty_cycle(_icms_duty_cycle, 499 new_duty_cycle); 500 } 501 } else if (_allow_duty_cycle_reduction) { 502 // The duty cycle is reduced only once per cms cycle (see record_cms_end()). 503 new_duty_cycle = icms_damped_duty_cycle(_icms_duty_cycle, new_duty_cycle); 504 // Respect the minimum duty cycle. 505 unsigned int min_duty_cycle = (unsigned int)CMSIncrementalDutyCycleMin; 506 _icms_duty_cycle = MAX2(new_duty_cycle, min_duty_cycle); 507 } 508 509 if (PrintGCDetails || CMSTraceIncrementalPacing) { 510 gclog_or_tty->print(" icms_dc=%d ", _icms_duty_cycle); 511 } 512 513 _allow_duty_cycle_reduction = false; 514 return _icms_duty_cycle; 515 } 516 517 #ifndef PRODUCT 518 void CMSStats::print_on(outputStream *st) const { 519 st->print(" gc0_alpha=%d,cms_alpha=%d", _gc0_alpha, _cms_alpha); 520 st->print(",gc0_dur=%g,gc0_per=%g,gc0_promo=" SIZE_FORMAT, 521 gc0_duration(), gc0_period(), gc0_promoted()); 522 st->print(",cms_dur=%g,cms_dur_per_mb=%g,cms_per=%g,cms_alloc=" SIZE_FORMAT, 523 cms_duration(), cms_duration_per_mb(), 524 cms_period(), cms_allocated()); 525 st->print(",cms_since_beg=%g,cms_since_end=%g", 526 cms_time_since_begin(), cms_time_since_end()); 527 st->print(",cms_used_beg=" SIZE_FORMAT ",cms_used_end=" SIZE_FORMAT, 528 _cms_used_at_gc0_begin, _cms_used_at_gc0_end); 529 if (CMSIncrementalMode) { 530 st->print(",dc=%d", icms_duty_cycle()); 531 } 532 533 if (valid()) { 534 st->print(",promo_rate=%g,cms_alloc_rate=%g", 535 promotion_rate(), cms_allocation_rate()); 536 st->print(",cms_consumption_rate=%g,time_until_full=%g", 537 cms_consumption_rate(), time_until_cms_gen_full()); 538 } 539 st->print(" "); 540 } 541 #endif // #ifndef PRODUCT 542 543 CMSCollector::CollectorState CMSCollector::_collectorState = 544 CMSCollector::Idling; 545 bool CMSCollector::_foregroundGCIsActive = false; 546 bool CMSCollector::_foregroundGCShouldWait = false; 547 548 CMSCollector::CMSCollector(ConcurrentMarkSweepGeneration* cmsGen, 549 ConcurrentMarkSweepGeneration* permGen, 550 CardTableRS* ct, 551 ConcurrentMarkSweepPolicy* cp): 552 _cmsGen(cmsGen), 553 _permGen(permGen), 554 _ct(ct), 555 _ref_processor(NULL), // will be set later 556 _conc_workers(NULL), // may be set later 557 _abort_preclean(false), 558 _start_sampling(false), 559 _between_prologue_and_epilogue(false), 560 _markBitMap(0, Mutex::leaf + 1, "CMS_markBitMap_lock"), 561 _perm_gen_verify_bit_map(0, -1 /* no mutex */, "No_lock"), 562 _modUnionTable((CardTableModRefBS::card_shift - LogHeapWordSize), 563 -1 /* lock-free */, "No_lock" /* dummy */), 564 _modUnionClosure(&_modUnionTable), 565 _modUnionClosurePar(&_modUnionTable), 566 // Adjust my span to cover old (cms) gen and perm gen 567 _span(cmsGen->reserved()._union(permGen->reserved())), 568 // Construct the is_alive_closure with _span & markBitMap 569 _is_alive_closure(_span, &_markBitMap), 570 _restart_addr(NULL), 571 _overflow_list(NULL), 572 _stats(cmsGen), 573 _eden_chunk_array(NULL), // may be set in ctor body 574 _eden_chunk_capacity(0), // -- ditto -- 575 _eden_chunk_index(0), // -- ditto -- 576 _survivor_plab_array(NULL), // -- ditto -- 577 _survivor_chunk_array(NULL), // -- ditto -- 578 _survivor_chunk_capacity(0), // -- ditto -- 579 _survivor_chunk_index(0), // -- ditto -- 580 _ser_pmc_preclean_ovflw(0), 581 _ser_kac_preclean_ovflw(0), 582 _ser_pmc_remark_ovflw(0), 583 _par_pmc_remark_ovflw(0), 584 _ser_kac_ovflw(0), 585 _par_kac_ovflw(0), 586 #ifndef PRODUCT 587 _num_par_pushes(0), 588 #endif 589 _collection_count_start(0), 590 _verifying(false), 591 _icms_start_limit(NULL), 592 _icms_stop_limit(NULL), 593 _verification_mark_bm(0, Mutex::leaf + 1, "CMS_verification_mark_bm_lock"), 594 _completed_initialization(false), 595 _collector_policy(cp), 596 _should_unload_classes(false), 597 _concurrent_cycles_since_last_unload(0), 598 _roots_scanning_options(0), 599 _inter_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding), 600 _intra_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding) 601 { 602 if (ExplicitGCInvokesConcurrentAndUnloadsClasses) { 603 ExplicitGCInvokesConcurrent = true; 604 } 605 // Now expand the span and allocate the collection support structures 606 // (MUT, marking bit map etc.) to cover both generations subject to 607 // collection. 608 609 // First check that _permGen is adjacent to _cmsGen and above it. 610 assert( _cmsGen->reserved().word_size() > 0 611 && _permGen->reserved().word_size() > 0, 612 "generations should not be of zero size"); 613 assert(_cmsGen->reserved().intersection(_permGen->reserved()).is_empty(), 614 "_cmsGen and _permGen should not overlap"); 615 assert(_cmsGen->reserved().end() == _permGen->reserved().start(), 616 "_cmsGen->end() different from _permGen->start()"); 617 618 // For use by dirty card to oop closures. 619 _cmsGen->cmsSpace()->set_collector(this); 620 _permGen->cmsSpace()->set_collector(this); 621 622 // Allocate MUT and marking bit map 623 { 624 MutexLockerEx x(_markBitMap.lock(), Mutex::_no_safepoint_check_flag); 625 if (!_markBitMap.allocate(_span)) { 626 warning("Failed to allocate CMS Bit Map"); 627 return; 628 } 629 assert(_markBitMap.covers(_span), "_markBitMap inconsistency?"); 630 } 631 { 632 _modUnionTable.allocate(_span); 633 assert(_modUnionTable.covers(_span), "_modUnionTable inconsistency?"); 634 } 635 636 if (!_markStack.allocate(MarkStackSize)) { 637 warning("Failed to allocate CMS Marking Stack"); 638 return; 639 } 640 if (!_revisitStack.allocate(CMSRevisitStackSize)) { 641 warning("Failed to allocate CMS Revisit Stack"); 642 return; 643 } 644 645 // Support for multi-threaded concurrent phases 646 if (CMSConcurrentMTEnabled) { 647 if (FLAG_IS_DEFAULT(ConcGCThreads)) { 648 // just for now 649 FLAG_SET_DEFAULT(ConcGCThreads, (ParallelGCThreads + 3)/4); 650 } 651 if (ConcGCThreads > 1) { 652 _conc_workers = new YieldingFlexibleWorkGang("Parallel CMS Threads", 653 ConcGCThreads, true); 654 if (_conc_workers == NULL) { 655 warning("GC/CMS: _conc_workers allocation failure: " 656 "forcing -CMSConcurrentMTEnabled"); 657 CMSConcurrentMTEnabled = false; 658 } else { 659 _conc_workers->initialize_workers(); 660 } 661 } else { 662 CMSConcurrentMTEnabled = false; 663 } 664 } 665 if (!CMSConcurrentMTEnabled) { 666 ConcGCThreads = 0; 667 } else { 668 // Turn off CMSCleanOnEnter optimization temporarily for 669 // the MT case where it's not fixed yet; see 6178663. 670 CMSCleanOnEnter = false; 671 } 672 assert((_conc_workers != NULL) == (ConcGCThreads > 1), 673 "Inconsistency"); 674 675 // Parallel task queues; these are shared for the 676 // concurrent and stop-world phases of CMS, but 677 // are not shared with parallel scavenge (ParNew). 678 { 679 uint i; 680 uint num_queues = (uint) MAX2(ParallelGCThreads, ConcGCThreads); 681 682 if ((CMSParallelRemarkEnabled || CMSConcurrentMTEnabled 683 || ParallelRefProcEnabled) 684 && num_queues > 0) { 685 _task_queues = new OopTaskQueueSet(num_queues); 686 if (_task_queues == NULL) { 687 warning("task_queues allocation failure."); 688 return; 689 } 690 _hash_seed = NEW_C_HEAP_ARRAY(int, num_queues); 691 if (_hash_seed == NULL) { 692 warning("_hash_seed array allocation failure"); 693 return; 694 } 695 696 typedef Padded<OopTaskQueue> PaddedOopTaskQueue; 697 for (i = 0; i < num_queues; i++) { 698 PaddedOopTaskQueue *q = new PaddedOopTaskQueue(); 699 if (q == NULL) { 700 warning("work_queue allocation failure."); 701 return; 702 } 703 _task_queues->register_queue(i, q); 704 } 705 for (i = 0; i < num_queues; i++) { 706 _task_queues->queue(i)->initialize(); 707 _hash_seed[i] = 17; // copied from ParNew 708 } 709 } 710 } 711 712 _cmsGen ->init_initiating_occupancy(CMSInitiatingOccupancyFraction, CMSTriggerRatio); 713 _permGen->init_initiating_occupancy(CMSInitiatingPermOccupancyFraction, CMSTriggerPermRatio); 714 715 // Clip CMSBootstrapOccupancy between 0 and 100. 716 _bootstrap_occupancy = ((double)MIN2((uintx)100, MAX2((uintx)0, CMSBootstrapOccupancy))) 717 /(double)100; 718 719 _full_gcs_since_conc_gc = 0; 720 721 // Now tell CMS generations the identity of their collector 722 ConcurrentMarkSweepGeneration::set_collector(this); 723 724 // Create & start a CMS thread for this CMS collector 725 _cmsThread = ConcurrentMarkSweepThread::start(this); 726 assert(cmsThread() != NULL, "CMS Thread should have been created"); 727 assert(cmsThread()->collector() == this, 728 "CMS Thread should refer to this gen"); 729 assert(CGC_lock != NULL, "Where's the CGC_lock?"); 730 731 // Support for parallelizing young gen rescan 732 GenCollectedHeap* gch = GenCollectedHeap::heap(); 733 _young_gen = gch->prev_gen(_cmsGen); 734 if (gch->supports_inline_contig_alloc()) { 735 _top_addr = gch->top_addr(); 736 _end_addr = gch->end_addr(); 737 assert(_young_gen != NULL, "no _young_gen"); 738 _eden_chunk_index = 0; 739 _eden_chunk_capacity = (_young_gen->max_capacity()+CMSSamplingGrain)/CMSSamplingGrain; 740 _eden_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, _eden_chunk_capacity); 741 if (_eden_chunk_array == NULL) { 742 _eden_chunk_capacity = 0; 743 warning("GC/CMS: _eden_chunk_array allocation failure"); 744 } 745 } 746 assert(_eden_chunk_array != NULL || _eden_chunk_capacity == 0, "Error"); 747 748 // Support for parallelizing survivor space rescan 749 if (CMSParallelRemarkEnabled && CMSParallelSurvivorRemarkEnabled) { 750 const size_t max_plab_samples = 751 ((DefNewGeneration*)_young_gen)->max_survivor_size()/MinTLABSize; 752 753 _survivor_plab_array = NEW_C_HEAP_ARRAY(ChunkArray, ParallelGCThreads); 754 _survivor_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, 2*max_plab_samples); 755 _cursor = NEW_C_HEAP_ARRAY(size_t, ParallelGCThreads); 756 if (_survivor_plab_array == NULL || _survivor_chunk_array == NULL 757 || _cursor == NULL) { 758 warning("Failed to allocate survivor plab/chunk array"); 759 if (_survivor_plab_array != NULL) { 760 FREE_C_HEAP_ARRAY(ChunkArray, _survivor_plab_array); 761 _survivor_plab_array = NULL; 762 } 763 if (_survivor_chunk_array != NULL) { 764 FREE_C_HEAP_ARRAY(HeapWord*, _survivor_chunk_array); 765 _survivor_chunk_array = NULL; 766 } 767 if (_cursor != NULL) { 768 FREE_C_HEAP_ARRAY(size_t, _cursor); 769 _cursor = NULL; 770 } 771 } else { 772 _survivor_chunk_capacity = 2*max_plab_samples; 773 for (uint i = 0; i < ParallelGCThreads; i++) { 774 HeapWord** vec = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples); 775 if (vec == NULL) { 776 warning("Failed to allocate survivor plab array"); 777 for (int j = i; j > 0; j--) { 778 FREE_C_HEAP_ARRAY(HeapWord*, _survivor_plab_array[j-1].array()); 779 } 780 FREE_C_HEAP_ARRAY(ChunkArray, _survivor_plab_array); 781 FREE_C_HEAP_ARRAY(HeapWord*, _survivor_chunk_array); 782 _survivor_plab_array = NULL; 783 _survivor_chunk_array = NULL; 784 _survivor_chunk_capacity = 0; 785 break; 786 } else { 787 ChunkArray* cur = 788 ::new (&_survivor_plab_array[i]) ChunkArray(vec, 789 max_plab_samples); 790 assert(cur->end() == 0, "Should be 0"); 791 assert(cur->array() == vec, "Should be vec"); 792 assert(cur->capacity() == max_plab_samples, "Error"); 793 } 794 } 795 } 796 } 797 assert( ( _survivor_plab_array != NULL 798 && _survivor_chunk_array != NULL) 799 || ( _survivor_chunk_capacity == 0 800 && _survivor_chunk_index == 0), 801 "Error"); 802 803 // Choose what strong roots should be scanned depending on verification options 804 // and perm gen collection mode. 805 if (!CMSClassUnloadingEnabled) { 806 // If class unloading is disabled we want to include all classes into the root set. 807 add_root_scanning_option(SharedHeap::SO_AllClasses); 808 } else { 809 add_root_scanning_option(SharedHeap::SO_SystemClasses); 810 } 811 812 NOT_PRODUCT(_overflow_counter = CMSMarkStackOverflowInterval;) 813 _gc_counters = new CollectorCounters("CMS", 1); 814 _completed_initialization = true; 815 _inter_sweep_timer.start(); // start of time 816 #ifdef SPARC 817 // Issue a stern warning, but allow use for experimentation and debugging. 818 if (VM_Version::is_sun4v() && UseMemSetInBOT) { 819 assert(!FLAG_IS_DEFAULT(UseMemSetInBOT), "Error"); 820 warning("Experimental flag -XX:+UseMemSetInBOT is known to cause instability" 821 " on sun4v; please understand that you are using at your own risk!"); 822 } 823 #endif 824 } 825 826 const char* ConcurrentMarkSweepGeneration::name() const { 827 return "concurrent mark-sweep generation"; 828 } 829 void ConcurrentMarkSweepGeneration::update_counters() { 830 if (UsePerfData) { 831 _space_counters->update_all(); 832 _gen_counters->update_all(); 833 } 834 } 835 836 // this is an optimized version of update_counters(). it takes the 837 // used value as a parameter rather than computing it. 838 // 839 void ConcurrentMarkSweepGeneration::update_counters(size_t used) { 840 if (UsePerfData) { 841 _space_counters->update_used(used); 842 _space_counters->update_capacity(); 843 _gen_counters->update_all(); 844 } 845 } 846 847 void ConcurrentMarkSweepGeneration::print() const { 848 Generation::print(); 849 cmsSpace()->print(); 850 } 851 852 #ifndef PRODUCT 853 void ConcurrentMarkSweepGeneration::print_statistics() { 854 cmsSpace()->printFLCensus(0); 855 } 856 #endif 857 858 void ConcurrentMarkSweepGeneration::printOccupancy(const char *s) { 859 GenCollectedHeap* gch = GenCollectedHeap::heap(); 860 if (PrintGCDetails) { 861 if (Verbose) { 862 gclog_or_tty->print(" [%d %s-%s: "SIZE_FORMAT"("SIZE_FORMAT")]", 863 level(), short_name(), s, used(), capacity()); 864 } else { 865 gclog_or_tty->print(" [%d %s-%s: "SIZE_FORMAT"K("SIZE_FORMAT"K)]", 866 level(), short_name(), s, used() / K, capacity() / K); 867 } 868 } 869 if (Verbose) { 870 gclog_or_tty->print(" "SIZE_FORMAT"("SIZE_FORMAT")", 871 gch->used(), gch->capacity()); 872 } else { 873 gclog_or_tty->print(" "SIZE_FORMAT"K("SIZE_FORMAT"K)", 874 gch->used() / K, gch->capacity() / K); 875 } 876 } 877 878 size_t 879 ConcurrentMarkSweepGeneration::contiguous_available() const { 880 // dld proposes an improvement in precision here. If the committed 881 // part of the space ends in a free block we should add that to 882 // uncommitted size in the calculation below. Will make this 883 // change later, staying with the approximation below for the 884 // time being. -- ysr. 885 return MAX2(_virtual_space.uncommitted_size(), unsafe_max_alloc_nogc()); 886 } 887 888 size_t 889 ConcurrentMarkSweepGeneration::unsafe_max_alloc_nogc() const { 890 return _cmsSpace->max_alloc_in_words() * HeapWordSize; 891 } 892 893 size_t ConcurrentMarkSweepGeneration::max_available() const { 894 return free() + _virtual_space.uncommitted_size(); 895 } 896 897 bool ConcurrentMarkSweepGeneration::promotion_attempt_is_safe(size_t max_promotion_in_bytes) const { 898 size_t available = max_available(); 899 size_t av_promo = (size_t)gc_stats()->avg_promoted()->padded_average(); 900 bool res = (available >= av_promo) || (available >= max_promotion_in_bytes); 901 if (Verbose && PrintGCDetails) { 902 gclog_or_tty->print_cr( 903 "CMS: promo attempt is%s safe: available("SIZE_FORMAT") %s av_promo("SIZE_FORMAT")," 904 "max_promo("SIZE_FORMAT")", 905 res? "":" not", available, res? ">=":"<", 906 av_promo, max_promotion_in_bytes); 907 } 908 return res; 909 } 910 911 // At a promotion failure dump information on block layout in heap 912 // (cms old generation). 913 void ConcurrentMarkSweepGeneration::promotion_failure_occurred() { 914 if (CMSDumpAtPromotionFailure) { 915 cmsSpace()->dump_at_safepoint_with_locks(collector(), gclog_or_tty); 916 } 917 } 918 919 CompactibleSpace* 920 ConcurrentMarkSweepGeneration::first_compaction_space() const { 921 return _cmsSpace; 922 } 923 924 void ConcurrentMarkSweepGeneration::reset_after_compaction() { 925 // Clear the promotion information. These pointers can be adjusted 926 // along with all the other pointers into the heap but 927 // compaction is expected to be a rare event with 928 // a heap using cms so don't do it without seeing the need. 929 if (CollectedHeap::use_parallel_gc_threads()) { 930 for (uint i = 0; i < ParallelGCThreads; i++) { 931 _par_gc_thread_states[i]->promo.reset(); 932 } 933 } 934 } 935 936 void ConcurrentMarkSweepGeneration::space_iterate(SpaceClosure* blk, bool usedOnly) { 937 blk->do_space(_cmsSpace); 938 } 939 940 void ConcurrentMarkSweepGeneration::compute_new_size() { 941 assert_locked_or_safepoint(Heap_lock); 942 943 // If incremental collection failed, we just want to expand 944 // to the limit. 945 if (incremental_collection_failed()) { 946 clear_incremental_collection_failed(); 947 grow_to_reserved(); 948 return; 949 } 950 951 size_t expand_bytes = 0; 952 double free_percentage = ((double) free()) / capacity(); 953 double desired_free_percentage = (double) MinHeapFreeRatio / 100; 954 double maximum_free_percentage = (double) MaxHeapFreeRatio / 100; 955 956 // compute expansion delta needed for reaching desired free percentage 957 if (free_percentage < desired_free_percentage) { 958 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 959 assert(desired_capacity >= capacity(), "invalid expansion size"); 960 expand_bytes = MAX2(desired_capacity - capacity(), MinHeapDeltaBytes); 961 } 962 if (expand_bytes > 0) { 963 if (PrintGCDetails && Verbose) { 964 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 965 gclog_or_tty->print_cr("\nFrom compute_new_size: "); 966 gclog_or_tty->print_cr(" Free fraction %f", free_percentage); 967 gclog_or_tty->print_cr(" Desired free fraction %f", 968 desired_free_percentage); 969 gclog_or_tty->print_cr(" Maximum free fraction %f", 970 maximum_free_percentage); 971 gclog_or_tty->print_cr(" Capactiy "SIZE_FORMAT, capacity()/1000); 972 gclog_or_tty->print_cr(" Desired capacity "SIZE_FORMAT, 973 desired_capacity/1000); 974 int prev_level = level() - 1; 975 if (prev_level >= 0) { 976 size_t prev_size = 0; 977 GenCollectedHeap* gch = GenCollectedHeap::heap(); 978 Generation* prev_gen = gch->_gens[prev_level]; 979 prev_size = prev_gen->capacity(); 980 gclog_or_tty->print_cr(" Younger gen size "SIZE_FORMAT, 981 prev_size/1000); 982 } 983 gclog_or_tty->print_cr(" unsafe_max_alloc_nogc "SIZE_FORMAT, 984 unsafe_max_alloc_nogc()/1000); 985 gclog_or_tty->print_cr(" contiguous available "SIZE_FORMAT, 986 contiguous_available()/1000); 987 gclog_or_tty->print_cr(" Expand by "SIZE_FORMAT" (bytes)", 988 expand_bytes); 989 } 990 // safe if expansion fails 991 expand(expand_bytes, 0, CMSExpansionCause::_satisfy_free_ratio); 992 if (PrintGCDetails && Verbose) { 993 gclog_or_tty->print_cr(" Expanded free fraction %f", 994 ((double) free()) / capacity()); 995 } 996 } 997 } 998 999 Mutex* ConcurrentMarkSweepGeneration::freelistLock() const { 1000 return cmsSpace()->freelistLock(); 1001 } 1002 1003 HeapWord* ConcurrentMarkSweepGeneration::allocate(size_t size, 1004 bool tlab) { 1005 CMSSynchronousYieldRequest yr; 1006 MutexLockerEx x(freelistLock(), 1007 Mutex::_no_safepoint_check_flag); 1008 return have_lock_and_allocate(size, tlab); 1009 } 1010 1011 HeapWord* ConcurrentMarkSweepGeneration::have_lock_and_allocate(size_t size, 1012 bool tlab /* ignored */) { 1013 assert_lock_strong(freelistLock()); 1014 size_t adjustedSize = CompactibleFreeListSpace::adjustObjectSize(size); 1015 HeapWord* res = cmsSpace()->allocate(adjustedSize); 1016 // Allocate the object live (grey) if the background collector has 1017 // started marking. This is necessary because the marker may 1018 // have passed this address and consequently this object will 1019 // not otherwise be greyed and would be incorrectly swept up. 1020 // Note that if this object contains references, the writing 1021 // of those references will dirty the card containing this object 1022 // allowing the object to be blackened (and its references scanned) 1023 // either during a preclean phase or at the final checkpoint. 1024 if (res != NULL) { 1025 // We may block here with an uninitialized object with 1026 // its mark-bit or P-bits not yet set. Such objects need 1027 // to be safely navigable by block_start(). 1028 assert(oop(res)->klass_or_null() == NULL, "Object should be uninitialized here."); 1029 assert(!((FreeChunk*)res)->is_free(), "Error, block will look free but show wrong size"); 1030 collector()->direct_allocated(res, adjustedSize); 1031 _direct_allocated_words += adjustedSize; 1032 // allocation counters 1033 NOT_PRODUCT( 1034 _numObjectsAllocated++; 1035 _numWordsAllocated += (int)adjustedSize; 1036 ) 1037 } 1038 return res; 1039 } 1040 1041 // In the case of direct allocation by mutators in a generation that 1042 // is being concurrently collected, the object must be allocated 1043 // live (grey) if the background collector has started marking. 1044 // This is necessary because the marker may 1045 // have passed this address and consequently this object will 1046 // not otherwise be greyed and would be incorrectly swept up. 1047 // Note that if this object contains references, the writing 1048 // of those references will dirty the card containing this object 1049 // allowing the object to be blackened (and its references scanned) 1050 // either during a preclean phase or at the final checkpoint. 1051 void CMSCollector::direct_allocated(HeapWord* start, size_t size) { 1052 assert(_markBitMap.covers(start, size), "Out of bounds"); 1053 if (_collectorState >= Marking) { 1054 MutexLockerEx y(_markBitMap.lock(), 1055 Mutex::_no_safepoint_check_flag); 1056 // [see comments preceding SweepClosure::do_blk() below for details] 1057 // 1. need to mark the object as live so it isn't collected 1058 // 2. need to mark the 2nd bit to indicate the object may be uninitialized 1059 // 3. need to mark the end of the object so marking, precleaning or sweeping 1060 // can skip over uninitialized or unparsable objects. An allocated 1061 // object is considered uninitialized for our purposes as long as 1062 // its klass word is NULL. (Unparsable objects are those which are 1063 // initialized in the sense just described, but whose sizes can still 1064 // not be correctly determined. Note that the class of unparsable objects 1065 // can only occur in the perm gen. All old gen objects are parsable 1066 // as soon as they are initialized.) 1067 _markBitMap.mark(start); // object is live 1068 _markBitMap.mark(start + 1); // object is potentially uninitialized? 1069 _markBitMap.mark(start + size - 1); 1070 // mark end of object 1071 } 1072 // check that oop looks uninitialized 1073 assert(oop(start)->klass_or_null() == NULL, "_klass should be NULL"); 1074 } 1075 1076 void CMSCollector::promoted(bool par, HeapWord* start, 1077 bool is_obj_array, size_t obj_size) { 1078 assert(_markBitMap.covers(start), "Out of bounds"); 1079 // See comment in direct_allocated() about when objects should 1080 // be allocated live. 1081 if (_collectorState >= Marking) { 1082 // we already hold the marking bit map lock, taken in 1083 // the prologue 1084 if (par) { 1085 _markBitMap.par_mark(start); 1086 } else { 1087 _markBitMap.mark(start); 1088 } 1089 // We don't need to mark the object as uninitialized (as 1090 // in direct_allocated above) because this is being done with the 1091 // world stopped and the object will be initialized by the 1092 // time the marking, precleaning or sweeping get to look at it. 1093 // But see the code for copying objects into the CMS generation, 1094 // where we need to ensure that concurrent readers of the 1095 // block offset table are able to safely navigate a block that 1096 // is in flux from being free to being allocated (and in 1097 // transition while being copied into) and subsequently 1098 // becoming a bona-fide object when the copy/promotion is complete. 1099 assert(SafepointSynchronize::is_at_safepoint(), 1100 "expect promotion only at safepoints"); 1101 1102 if (_collectorState < Sweeping) { 1103 // Mark the appropriate cards in the modUnionTable, so that 1104 // this object gets scanned before the sweep. If this is 1105 // not done, CMS generation references in the object might 1106 // not get marked. 1107 // For the case of arrays, which are otherwise precisely 1108 // marked, we need to dirty the entire array, not just its head. 1109 if (is_obj_array) { 1110 // The [par_]mark_range() method expects mr.end() below to 1111 // be aligned to the granularity of a bit's representation 1112 // in the heap. In the case of the MUT below, that's a 1113 // card size. 1114 MemRegion mr(start, 1115 (HeapWord*)round_to((intptr_t)(start + obj_size), 1116 CardTableModRefBS::card_size /* bytes */)); 1117 if (par) { 1118 _modUnionTable.par_mark_range(mr); 1119 } else { 1120 _modUnionTable.mark_range(mr); 1121 } 1122 } else { // not an obj array; we can just mark the head 1123 if (par) { 1124 _modUnionTable.par_mark(start); 1125 } else { 1126 _modUnionTable.mark(start); 1127 } 1128 } 1129 } 1130 } 1131 } 1132 1133 static inline size_t percent_of_space(Space* space, HeapWord* addr) 1134 { 1135 size_t delta = pointer_delta(addr, space->bottom()); 1136 return (size_t)(delta * 100.0 / (space->capacity() / HeapWordSize)); 1137 } 1138 1139 void CMSCollector::icms_update_allocation_limits() 1140 { 1141 Generation* gen0 = GenCollectedHeap::heap()->get_gen(0); 1142 EdenSpace* eden = gen0->as_DefNewGeneration()->eden(); 1143 1144 const unsigned int duty_cycle = stats().icms_update_duty_cycle(); 1145 if (CMSTraceIncrementalPacing) { 1146 stats().print(); 1147 } 1148 1149 assert(duty_cycle <= 100, "invalid duty cycle"); 1150 if (duty_cycle != 0) { 1151 // The duty_cycle is a percentage between 0 and 100; convert to words and 1152 // then compute the offset from the endpoints of the space. 1153 size_t free_words = eden->free() / HeapWordSize; 1154 double free_words_dbl = (double)free_words; 1155 size_t duty_cycle_words = (size_t)(free_words_dbl * duty_cycle / 100.0); 1156 size_t offset_words = (free_words - duty_cycle_words) / 2; 1157 1158 _icms_start_limit = eden->top() + offset_words; 1159 _icms_stop_limit = eden->end() - offset_words; 1160 1161 // The limits may be adjusted (shifted to the right) by 1162 // CMSIncrementalOffset, to allow the application more mutator time after a 1163 // young gen gc (when all mutators were stopped) and before CMS starts and 1164 // takes away one or more cpus. 1165 if (CMSIncrementalOffset != 0) { 1166 double adjustment_dbl = free_words_dbl * CMSIncrementalOffset / 100.0; 1167 size_t adjustment = (size_t)adjustment_dbl; 1168 HeapWord* tmp_stop = _icms_stop_limit + adjustment; 1169 if (tmp_stop > _icms_stop_limit && tmp_stop < eden->end()) { 1170 _icms_start_limit += adjustment; 1171 _icms_stop_limit = tmp_stop; 1172 } 1173 } 1174 } 1175 if (duty_cycle == 0 || (_icms_start_limit == _icms_stop_limit)) { 1176 _icms_start_limit = _icms_stop_limit = eden->end(); 1177 } 1178 1179 // Install the new start limit. 1180 eden->set_soft_end(_icms_start_limit); 1181 1182 if (CMSTraceIncrementalMode) { 1183 gclog_or_tty->print(" icms alloc limits: " 1184 PTR_FORMAT "," PTR_FORMAT 1185 " (" SIZE_FORMAT "%%," SIZE_FORMAT "%%) ", 1186 _icms_start_limit, _icms_stop_limit, 1187 percent_of_space(eden, _icms_start_limit), 1188 percent_of_space(eden, _icms_stop_limit)); 1189 if (Verbose) { 1190 gclog_or_tty->print("eden: "); 1191 eden->print_on(gclog_or_tty); 1192 } 1193 } 1194 } 1195 1196 // Any changes here should try to maintain the invariant 1197 // that if this method is called with _icms_start_limit 1198 // and _icms_stop_limit both NULL, then it should return NULL 1199 // and not notify the icms thread. 1200 HeapWord* 1201 CMSCollector::allocation_limit_reached(Space* space, HeapWord* top, 1202 size_t word_size) 1203 { 1204 // A start_limit equal to end() means the duty cycle is 0, so treat that as a 1205 // nop. 1206 if (CMSIncrementalMode && _icms_start_limit != space->end()) { 1207 if (top <= _icms_start_limit) { 1208 if (CMSTraceIncrementalMode) { 1209 space->print_on(gclog_or_tty); 1210 gclog_or_tty->stamp(); 1211 gclog_or_tty->print_cr(" start limit top=" PTR_FORMAT 1212 ", new limit=" PTR_FORMAT 1213 " (" SIZE_FORMAT "%%)", 1214 top, _icms_stop_limit, 1215 percent_of_space(space, _icms_stop_limit)); 1216 } 1217 ConcurrentMarkSweepThread::start_icms(); 1218 assert(top < _icms_stop_limit, "Tautology"); 1219 if (word_size < pointer_delta(_icms_stop_limit, top)) { 1220 return _icms_stop_limit; 1221 } 1222 1223 // The allocation will cross both the _start and _stop limits, so do the 1224 // stop notification also and return end(). 1225 if (CMSTraceIncrementalMode) { 1226 space->print_on(gclog_or_tty); 1227 gclog_or_tty->stamp(); 1228 gclog_or_tty->print_cr(" +stop limit top=" PTR_FORMAT 1229 ", new limit=" PTR_FORMAT 1230 " (" SIZE_FORMAT "%%)", 1231 top, space->end(), 1232 percent_of_space(space, space->end())); 1233 } 1234 ConcurrentMarkSweepThread::stop_icms(); 1235 return space->end(); 1236 } 1237 1238 if (top <= _icms_stop_limit) { 1239 if (CMSTraceIncrementalMode) { 1240 space->print_on(gclog_or_tty); 1241 gclog_or_tty->stamp(); 1242 gclog_or_tty->print_cr(" stop limit top=" PTR_FORMAT 1243 ", new limit=" PTR_FORMAT 1244 " (" SIZE_FORMAT "%%)", 1245 top, space->end(), 1246 percent_of_space(space, space->end())); 1247 } 1248 ConcurrentMarkSweepThread::stop_icms(); 1249 return space->end(); 1250 } 1251 1252 if (CMSTraceIncrementalMode) { 1253 space->print_on(gclog_or_tty); 1254 gclog_or_tty->stamp(); 1255 gclog_or_tty->print_cr(" end limit top=" PTR_FORMAT 1256 ", new limit=" PTR_FORMAT, 1257 top, NULL); 1258 } 1259 } 1260 1261 return NULL; 1262 } 1263 1264 oop ConcurrentMarkSweepGeneration::promote(oop obj, size_t obj_size) { 1265 assert(obj_size == (size_t)obj->size(), "bad obj_size passed in"); 1266 // allocate, copy and if necessary update promoinfo -- 1267 // delegate to underlying space. 1268 assert_lock_strong(freelistLock()); 1269 1270 #ifndef PRODUCT 1271 if (Universe::heap()->promotion_should_fail()) { 1272 return NULL; 1273 } 1274 #endif // #ifndef PRODUCT 1275 1276 oop res = _cmsSpace->promote(obj, obj_size); 1277 if (res == NULL) { 1278 // expand and retry 1279 size_t s = _cmsSpace->expansionSpaceRequired(obj_size); // HeapWords 1280 expand(s*HeapWordSize, MinHeapDeltaBytes, 1281 CMSExpansionCause::_satisfy_promotion); 1282 // Since there's currently no next generation, we don't try to promote 1283 // into a more senior generation. 1284 assert(next_gen() == NULL, "assumption, based upon which no attempt " 1285 "is made to pass on a possibly failing " 1286 "promotion to next generation"); 1287 res = _cmsSpace->promote(obj, obj_size); 1288 } 1289 if (res != NULL) { 1290 // See comment in allocate() about when objects should 1291 // be allocated live. 1292 assert(obj->is_oop(), "Will dereference klass pointer below"); 1293 collector()->promoted(false, // Not parallel 1294 (HeapWord*)res, obj->is_objArray(), obj_size); 1295 // promotion counters 1296 NOT_PRODUCT( 1297 _numObjectsPromoted++; 1298 _numWordsPromoted += 1299 (int)(CompactibleFreeListSpace::adjustObjectSize(obj->size())); 1300 ) 1301 } 1302 return res; 1303 } 1304 1305 1306 HeapWord* 1307 ConcurrentMarkSweepGeneration::allocation_limit_reached(Space* space, 1308 HeapWord* top, 1309 size_t word_sz) 1310 { 1311 return collector()->allocation_limit_reached(space, top, word_sz); 1312 } 1313 1314 // IMPORTANT: Notes on object size recognition in CMS. 1315 // --------------------------------------------------- 1316 // A block of storage in the CMS generation is always in 1317 // one of three states. A free block (FREE), an allocated 1318 // object (OBJECT) whose size() method reports the correct size, 1319 // and an intermediate state (TRANSIENT) in which its size cannot 1320 // be accurately determined. 1321 // STATE IDENTIFICATION: (32 bit and 64 bit w/o COOPS) 1322 // ----------------------------------------------------- 1323 // FREE: klass_word & 1 == 1; mark_word holds block size 1324 // 1325 // OBJECT: klass_word installed; klass_word != 0 && klass_word & 1 == 0; 1326 // obj->size() computes correct size 1327 // [Perm Gen objects needs to be "parsable" before they can be navigated] 1328 // 1329 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT 1330 // 1331 // STATE IDENTIFICATION: (64 bit+COOPS) 1332 // ------------------------------------ 1333 // FREE: mark_word & CMS_FREE_BIT == 1; mark_word & ~CMS_FREE_BIT gives block_size 1334 // 1335 // OBJECT: klass_word installed; klass_word != 0; 1336 // obj->size() computes correct size 1337 // [Perm Gen comment above continues to hold] 1338 // 1339 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT 1340 // 1341 // 1342 // STATE TRANSITION DIAGRAM 1343 // 1344 // mut / parnew mut / parnew 1345 // FREE --------------------> TRANSIENT ---------------------> OBJECT --| 1346 // ^ | 1347 // |------------------------ DEAD <------------------------------------| 1348 // sweep mut 1349 // 1350 // While a block is in TRANSIENT state its size cannot be determined 1351 // so readers will either need to come back later or stall until 1352 // the size can be determined. Note that for the case of direct 1353 // allocation, P-bits, when available, may be used to determine the 1354 // size of an object that may not yet have been initialized. 1355 1356 // Things to support parallel young-gen collection. 1357 oop 1358 ConcurrentMarkSweepGeneration::par_promote(int thread_num, 1359 oop old, markOop m, 1360 size_t word_sz) { 1361 #ifndef PRODUCT 1362 if (Universe::heap()->promotion_should_fail()) { 1363 return NULL; 1364 } 1365 #endif // #ifndef PRODUCT 1366 1367 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1368 PromotionInfo* promoInfo = &ps->promo; 1369 // if we are tracking promotions, then first ensure space for 1370 // promotion (including spooling space for saving header if necessary). 1371 // then allocate and copy, then track promoted info if needed. 1372 // When tracking (see PromotionInfo::track()), the mark word may 1373 // be displaced and in this case restoration of the mark word 1374 // occurs in the (oop_since_save_marks_)iterate phase. 1375 if (promoInfo->tracking() && !promoInfo->ensure_spooling_space()) { 1376 // Out of space for allocating spooling buffers; 1377 // try expanding and allocating spooling buffers. 1378 if (!expand_and_ensure_spooling_space(promoInfo)) { 1379 return NULL; 1380 } 1381 } 1382 assert(promoInfo->has_spooling_space(), "Control point invariant"); 1383 const size_t alloc_sz = CompactibleFreeListSpace::adjustObjectSize(word_sz); 1384 HeapWord* obj_ptr = ps->lab.alloc(alloc_sz); 1385 if (obj_ptr == NULL) { 1386 obj_ptr = expand_and_par_lab_allocate(ps, alloc_sz); 1387 if (obj_ptr == NULL) { 1388 return NULL; 1389 } 1390 } 1391 oop obj = oop(obj_ptr); 1392 OrderAccess::storestore(); 1393 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1394 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1395 // IMPORTANT: See note on object initialization for CMS above. 1396 // Otherwise, copy the object. Here we must be careful to insert the 1397 // klass pointer last, since this marks the block as an allocated object. 1398 // Except with compressed oops it's the mark word. 1399 HeapWord* old_ptr = (HeapWord*)old; 1400 // Restore the mark word copied above. 1401 obj->set_mark(m); 1402 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1403 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1404 OrderAccess::storestore(); 1405 1406 if (UseCompressedOops) { 1407 // Copy gap missed by (aligned) header size calculation below 1408 obj->set_klass_gap(old->klass_gap()); 1409 } 1410 if (word_sz > (size_t)oopDesc::header_size()) { 1411 Copy::aligned_disjoint_words(old_ptr + oopDesc::header_size(), 1412 obj_ptr + oopDesc::header_size(), 1413 word_sz - oopDesc::header_size()); 1414 } 1415 1416 // Now we can track the promoted object, if necessary. We take care 1417 // to delay the transition from uninitialized to full object 1418 // (i.e., insertion of klass pointer) until after, so that it 1419 // atomically becomes a promoted object. 1420 if (promoInfo->tracking()) { 1421 promoInfo->track((PromotedObject*)obj, old->klass()); 1422 } 1423 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1424 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1425 assert(old->is_oop(), "Will use and dereference old klass ptr below"); 1426 1427 // Finally, install the klass pointer (this should be volatile). 1428 OrderAccess::storestore(); 1429 obj->set_klass(old->klass()); 1430 // We should now be able to calculate the right size for this object 1431 assert(obj->is_oop() && obj->size() == (int)word_sz, "Error, incorrect size computed for promoted object"); 1432 1433 collector()->promoted(true, // parallel 1434 obj_ptr, old->is_objArray(), word_sz); 1435 1436 NOT_PRODUCT( 1437 Atomic::inc_ptr(&_numObjectsPromoted); 1438 Atomic::add_ptr(alloc_sz, &_numWordsPromoted); 1439 ) 1440 1441 return obj; 1442 } 1443 1444 void 1445 ConcurrentMarkSweepGeneration:: 1446 par_promote_alloc_undo(int thread_num, 1447 HeapWord* obj, size_t word_sz) { 1448 // CMS does not support promotion undo. 1449 ShouldNotReachHere(); 1450 } 1451 1452 void 1453 ConcurrentMarkSweepGeneration:: 1454 par_promote_alloc_done(int thread_num) { 1455 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1456 ps->lab.retire(thread_num); 1457 } 1458 1459 void 1460 ConcurrentMarkSweepGeneration:: 1461 par_oop_since_save_marks_iterate_done(int thread_num) { 1462 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1463 ParScanWithoutBarrierClosure* dummy_cl = NULL; 1464 ps->promo.promoted_oops_iterate_nv(dummy_cl); 1465 } 1466 1467 // XXXPERM 1468 bool ConcurrentMarkSweepGeneration::should_collect(bool full, 1469 size_t size, 1470 bool tlab) 1471 { 1472 // We allow a STW collection only if a full 1473 // collection was requested. 1474 return full || should_allocate(size, tlab); // FIX ME !!! 1475 // This and promotion failure handling are connected at the 1476 // hip and should be fixed by untying them. 1477 } 1478 1479 bool CMSCollector::shouldConcurrentCollect() { 1480 if (_full_gc_requested) { 1481 if (Verbose && PrintGCDetails) { 1482 gclog_or_tty->print_cr("CMSCollector: collect because of explicit " 1483 " gc request (or gc_locker)"); 1484 } 1485 return true; 1486 } 1487 1488 // For debugging purposes, change the type of collection. 1489 // If the rotation is not on the concurrent collection 1490 // type, don't start a concurrent collection. 1491 NOT_PRODUCT( 1492 if (RotateCMSCollectionTypes && 1493 (_cmsGen->debug_collection_type() != 1494 ConcurrentMarkSweepGeneration::Concurrent_collection_type)) { 1495 assert(_cmsGen->debug_collection_type() != 1496 ConcurrentMarkSweepGeneration::Unknown_collection_type, 1497 "Bad cms collection type"); 1498 return false; 1499 } 1500 ) 1501 1502 FreelistLocker x(this); 1503 // ------------------------------------------------------------------ 1504 // Print out lots of information which affects the initiation of 1505 // a collection. 1506 if (PrintCMSInitiationStatistics && stats().valid()) { 1507 gclog_or_tty->print("CMSCollector shouldConcurrentCollect: "); 1508 gclog_or_tty->stamp(); 1509 gclog_or_tty->print_cr(""); 1510 stats().print_on(gclog_or_tty); 1511 gclog_or_tty->print_cr("time_until_cms_gen_full %3.7f", 1512 stats().time_until_cms_gen_full()); 1513 gclog_or_tty->print_cr("free="SIZE_FORMAT, _cmsGen->free()); 1514 gclog_or_tty->print_cr("contiguous_available="SIZE_FORMAT, 1515 _cmsGen->contiguous_available()); 1516 gclog_or_tty->print_cr("promotion_rate=%g", stats().promotion_rate()); 1517 gclog_or_tty->print_cr("cms_allocation_rate=%g", stats().cms_allocation_rate()); 1518 gclog_or_tty->print_cr("occupancy=%3.7f", _cmsGen->occupancy()); 1519 gclog_or_tty->print_cr("initiatingOccupancy=%3.7f", _cmsGen->initiating_occupancy()); 1520 gclog_or_tty->print_cr("initiatingPermOccupancy=%3.7f", _permGen->initiating_occupancy()); 1521 } 1522 // ------------------------------------------------------------------ 1523 1524 // If the estimated time to complete a cms collection (cms_duration()) 1525 // is less than the estimated time remaining until the cms generation 1526 // is full, start a collection. 1527 if (!UseCMSInitiatingOccupancyOnly) { 1528 if (stats().valid()) { 1529 if (stats().time_until_cms_start() == 0.0) { 1530 return true; 1531 } 1532 } else { 1533 // We want to conservatively collect somewhat early in order 1534 // to try and "bootstrap" our CMS/promotion statistics; 1535 // this branch will not fire after the first successful CMS 1536 // collection because the stats should then be valid. 1537 if (_cmsGen->occupancy() >= _bootstrap_occupancy) { 1538 if (Verbose && PrintGCDetails) { 1539 gclog_or_tty->print_cr( 1540 " CMSCollector: collect for bootstrapping statistics:" 1541 " occupancy = %f, boot occupancy = %f", _cmsGen->occupancy(), 1542 _bootstrap_occupancy); 1543 } 1544 return true; 1545 } 1546 } 1547 } 1548 1549 // Otherwise, we start a collection cycle if either the perm gen or 1550 // old gen want a collection cycle started. Each may use 1551 // an appropriate criterion for making this decision. 1552 // XXX We need to make sure that the gen expansion 1553 // criterion dovetails well with this. XXX NEED TO FIX THIS 1554 if (_cmsGen->should_concurrent_collect()) { 1555 if (Verbose && PrintGCDetails) { 1556 gclog_or_tty->print_cr("CMS old gen initiated"); 1557 } 1558 return true; 1559 } 1560 1561 // We start a collection if we believe an incremental collection may fail; 1562 // this is not likely to be productive in practice because it's probably too 1563 // late anyway. 1564 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1565 assert(gch->collector_policy()->is_two_generation_policy(), 1566 "You may want to check the correctness of the following"); 1567 if (gch->incremental_collection_will_fail(true /* consult_young */)) { 1568 if (Verbose && PrintGCDetails) { 1569 gclog_or_tty->print("CMSCollector: collect because incremental collection will fail "); 1570 } 1571 return true; 1572 } 1573 1574 if (CMSClassUnloadingEnabled && _permGen->should_concurrent_collect()) { 1575 bool res = update_should_unload_classes(); 1576 if (res) { 1577 if (Verbose && PrintGCDetails) { 1578 gclog_or_tty->print_cr("CMS perm gen initiated"); 1579 } 1580 return true; 1581 } 1582 } 1583 return false; 1584 } 1585 1586 // Clear _expansion_cause fields of constituent generations 1587 void CMSCollector::clear_expansion_cause() { 1588 _cmsGen->clear_expansion_cause(); 1589 _permGen->clear_expansion_cause(); 1590 } 1591 1592 // We should be conservative in starting a collection cycle. To 1593 // start too eagerly runs the risk of collecting too often in the 1594 // extreme. To collect too rarely falls back on full collections, 1595 // which works, even if not optimum in terms of concurrent work. 1596 // As a work around for too eagerly collecting, use the flag 1597 // UseCMSInitiatingOccupancyOnly. This also has the advantage of 1598 // giving the user an easily understandable way of controlling the 1599 // collections. 1600 // We want to start a new collection cycle if any of the following 1601 // conditions hold: 1602 // . our current occupancy exceeds the configured initiating occupancy 1603 // for this generation, or 1604 // . we recently needed to expand this space and have not, since that 1605 // expansion, done a collection of this generation, or 1606 // . the underlying space believes that it may be a good idea to initiate 1607 // a concurrent collection (this may be based on criteria such as the 1608 // following: the space uses linear allocation and linear allocation is 1609 // going to fail, or there is believed to be excessive fragmentation in 1610 // the generation, etc... or ... 1611 // [.(currently done by CMSCollector::shouldConcurrentCollect() only for 1612 // the case of the old generation, not the perm generation; see CR 6543076): 1613 // we may be approaching a point at which allocation requests may fail because 1614 // we will be out of sufficient free space given allocation rate estimates.] 1615 bool ConcurrentMarkSweepGeneration::should_concurrent_collect() const { 1616 1617 assert_lock_strong(freelistLock()); 1618 if (occupancy() > initiating_occupancy()) { 1619 if (PrintGCDetails && Verbose) { 1620 gclog_or_tty->print(" %s: collect because of occupancy %f / %f ", 1621 short_name(), occupancy(), initiating_occupancy()); 1622 } 1623 return true; 1624 } 1625 if (UseCMSInitiatingOccupancyOnly) { 1626 return false; 1627 } 1628 if (expansion_cause() == CMSExpansionCause::_satisfy_allocation) { 1629 if (PrintGCDetails && Verbose) { 1630 gclog_or_tty->print(" %s: collect because expanded for allocation ", 1631 short_name()); 1632 } 1633 return true; 1634 } 1635 if (_cmsSpace->should_concurrent_collect()) { 1636 if (PrintGCDetails && Verbose) { 1637 gclog_or_tty->print(" %s: collect because cmsSpace says so ", 1638 short_name()); 1639 } 1640 return true; 1641 } 1642 return false; 1643 } 1644 1645 void ConcurrentMarkSweepGeneration::collect(bool full, 1646 bool clear_all_soft_refs, 1647 size_t size, 1648 bool tlab) 1649 { 1650 collector()->collect(full, clear_all_soft_refs, size, tlab); 1651 } 1652 1653 void CMSCollector::collect(bool full, 1654 bool clear_all_soft_refs, 1655 size_t size, 1656 bool tlab) 1657 { 1658 if (!UseCMSCollectionPassing && _collectorState > Idling) { 1659 // For debugging purposes skip the collection if the state 1660 // is not currently idle 1661 if (TraceCMSState) { 1662 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " skipped full:%d CMS state %d", 1663 Thread::current(), full, _collectorState); 1664 } 1665 return; 1666 } 1667 1668 // The following "if" branch is present for defensive reasons. 1669 // In the current uses of this interface, it can be replaced with: 1670 // assert(!GC_locker.is_active(), "Can't be called otherwise"); 1671 // But I am not placing that assert here to allow future 1672 // generality in invoking this interface. 1673 if (GC_locker::is_active()) { 1674 // A consistency test for GC_locker 1675 assert(GC_locker::needs_gc(), "Should have been set already"); 1676 // Skip this foreground collection, instead 1677 // expanding the heap if necessary. 1678 // Need the free list locks for the call to free() in compute_new_size() 1679 compute_new_size(); 1680 return; 1681 } 1682 acquire_control_and_collect(full, clear_all_soft_refs); 1683 _full_gcs_since_conc_gc++; 1684 1685 } 1686 1687 void CMSCollector::request_full_gc(unsigned int full_gc_count) { 1688 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1689 unsigned int gc_count = gch->total_full_collections(); 1690 if (gc_count == full_gc_count) { 1691 MutexLockerEx y(CGC_lock, Mutex::_no_safepoint_check_flag); 1692 _full_gc_requested = true; 1693 CGC_lock->notify(); // nudge CMS thread 1694 } else { 1695 assert(gc_count > full_gc_count, "Error: causal loop"); 1696 } 1697 } 1698 1699 1700 // The foreground and background collectors need to coordinate in order 1701 // to make sure that they do not mutually interfere with CMS collections. 1702 // When a background collection is active, 1703 // the foreground collector may need to take over (preempt) and 1704 // synchronously complete an ongoing collection. Depending on the 1705 // frequency of the background collections and the heap usage 1706 // of the application, this preemption can be seldom or frequent. 1707 // There are only certain 1708 // points in the background collection that the "collection-baton" 1709 // can be passed to the foreground collector. 1710 // 1711 // The foreground collector will wait for the baton before 1712 // starting any part of the collection. The foreground collector 1713 // will only wait at one location. 1714 // 1715 // The background collector will yield the baton before starting a new 1716 // phase of the collection (e.g., before initial marking, marking from roots, 1717 // precleaning, final re-mark, sweep etc.) This is normally done at the head 1718 // of the loop which switches the phases. The background collector does some 1719 // of the phases (initial mark, final re-mark) with the world stopped. 1720 // Because of locking involved in stopping the world, 1721 // the foreground collector should not block waiting for the background 1722 // collector when it is doing a stop-the-world phase. The background 1723 // collector will yield the baton at an additional point just before 1724 // it enters a stop-the-world phase. Once the world is stopped, the 1725 // background collector checks the phase of the collection. If the 1726 // phase has not changed, it proceeds with the collection. If the 1727 // phase has changed, it skips that phase of the collection. See 1728 // the comments on the use of the Heap_lock in collect_in_background(). 1729 // 1730 // Variable used in baton passing. 1731 // _foregroundGCIsActive - Set to true by the foreground collector when 1732 // it wants the baton. The foreground clears it when it has finished 1733 // the collection. 1734 // _foregroundGCShouldWait - Set to true by the background collector 1735 // when it is running. The foreground collector waits while 1736 // _foregroundGCShouldWait is true. 1737 // CGC_lock - monitor used to protect access to the above variables 1738 // and to notify the foreground and background collectors. 1739 // _collectorState - current state of the CMS collection. 1740 // 1741 // The foreground collector 1742 // acquires the CGC_lock 1743 // sets _foregroundGCIsActive 1744 // waits on the CGC_lock for _foregroundGCShouldWait to be false 1745 // various locks acquired in preparation for the collection 1746 // are released so as not to block the background collector 1747 // that is in the midst of a collection 1748 // proceeds with the collection 1749 // clears _foregroundGCIsActive 1750 // returns 1751 // 1752 // The background collector in a loop iterating on the phases of the 1753 // collection 1754 // acquires the CGC_lock 1755 // sets _foregroundGCShouldWait 1756 // if _foregroundGCIsActive is set 1757 // clears _foregroundGCShouldWait, notifies _CGC_lock 1758 // waits on _CGC_lock for _foregroundGCIsActive to become false 1759 // and exits the loop. 1760 // otherwise 1761 // proceed with that phase of the collection 1762 // if the phase is a stop-the-world phase, 1763 // yield the baton once more just before enqueueing 1764 // the stop-world CMS operation (executed by the VM thread). 1765 // returns after all phases of the collection are done 1766 // 1767 1768 void CMSCollector::acquire_control_and_collect(bool full, 1769 bool clear_all_soft_refs) { 1770 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint"); 1771 assert(!Thread::current()->is_ConcurrentGC_thread(), 1772 "shouldn't try to acquire control from self!"); 1773 1774 // Start the protocol for acquiring control of the 1775 // collection from the background collector (aka CMS thread). 1776 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 1777 "VM thread should have CMS token"); 1778 // Remember the possibly interrupted state of an ongoing 1779 // concurrent collection 1780 CollectorState first_state = _collectorState; 1781 1782 // Signal to a possibly ongoing concurrent collection that 1783 // we want to do a foreground collection. 1784 _foregroundGCIsActive = true; 1785 1786 // Disable incremental mode during a foreground collection. 1787 ICMSDisabler icms_disabler; 1788 1789 // release locks and wait for a notify from the background collector 1790 // releasing the locks in only necessary for phases which 1791 // do yields to improve the granularity of the collection. 1792 assert_lock_strong(bitMapLock()); 1793 // We need to lock the Free list lock for the space that we are 1794 // currently collecting. 1795 assert(haveFreelistLocks(), "Must be holding free list locks"); 1796 bitMapLock()->unlock(); 1797 releaseFreelistLocks(); 1798 { 1799 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1800 if (_foregroundGCShouldWait) { 1801 // We are going to be waiting for action for the CMS thread; 1802 // it had better not be gone (for instance at shutdown)! 1803 assert(ConcurrentMarkSweepThread::cmst() != NULL, 1804 "CMS thread must be running"); 1805 // Wait here until the background collector gives us the go-ahead 1806 ConcurrentMarkSweepThread::clear_CMS_flag( 1807 ConcurrentMarkSweepThread::CMS_vm_has_token); // release token 1808 // Get a possibly blocked CMS thread going: 1809 // Note that we set _foregroundGCIsActive true above, 1810 // without protection of the CGC_lock. 1811 CGC_lock->notify(); 1812 assert(!ConcurrentMarkSweepThread::vm_thread_wants_cms_token(), 1813 "Possible deadlock"); 1814 while (_foregroundGCShouldWait) { 1815 // wait for notification 1816 CGC_lock->wait(Mutex::_no_safepoint_check_flag); 1817 // Possibility of delay/starvation here, since CMS token does 1818 // not know to give priority to VM thread? Actually, i think 1819 // there wouldn't be any delay/starvation, but the proof of 1820 // that "fact" (?) appears non-trivial. XXX 20011219YSR 1821 } 1822 ConcurrentMarkSweepThread::set_CMS_flag( 1823 ConcurrentMarkSweepThread::CMS_vm_has_token); 1824 } 1825 } 1826 // The CMS_token is already held. Get back the other locks. 1827 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 1828 "VM thread should have CMS token"); 1829 getFreelistLocks(); 1830 bitMapLock()->lock_without_safepoint_check(); 1831 if (TraceCMSState) { 1832 gclog_or_tty->print_cr("CMS foreground collector has asked for control " 1833 INTPTR_FORMAT " with first state %d", Thread::current(), first_state); 1834 gclog_or_tty->print_cr(" gets control with state %d", _collectorState); 1835 } 1836 1837 // Check if we need to do a compaction, or if not, whether 1838 // we need to start the mark-sweep from scratch. 1839 bool should_compact = false; 1840 bool should_start_over = false; 1841 decide_foreground_collection_type(clear_all_soft_refs, 1842 &should_compact, &should_start_over); 1843 1844 NOT_PRODUCT( 1845 if (RotateCMSCollectionTypes) { 1846 if (_cmsGen->debug_collection_type() == 1847 ConcurrentMarkSweepGeneration::MSC_foreground_collection_type) { 1848 should_compact = true; 1849 } else if (_cmsGen->debug_collection_type() == 1850 ConcurrentMarkSweepGeneration::MS_foreground_collection_type) { 1851 should_compact = false; 1852 } 1853 } 1854 ) 1855 1856 if (PrintGCDetails && first_state > Idling) { 1857 GCCause::Cause cause = GenCollectedHeap::heap()->gc_cause(); 1858 if (GCCause::is_user_requested_gc(cause) || 1859 GCCause::is_serviceability_requested_gc(cause)) { 1860 gclog_or_tty->print(" (concurrent mode interrupted)"); 1861 } else { 1862 gclog_or_tty->print(" (concurrent mode failure)"); 1863 } 1864 } 1865 1866 if (should_compact) { 1867 // If the collection is being acquired from the background 1868 // collector, there may be references on the discovered 1869 // references lists that have NULL referents (being those 1870 // that were concurrently cleared by a mutator) or 1871 // that are no longer active (having been enqueued concurrently 1872 // by the mutator). 1873 // Scrub the list of those references because Mark-Sweep-Compact 1874 // code assumes referents are not NULL and that all discovered 1875 // Reference objects are active. 1876 ref_processor()->clean_up_discovered_references(); 1877 1878 do_compaction_work(clear_all_soft_refs); 1879 1880 // Has the GC time limit been exceeded? 1881 DefNewGeneration* young_gen = _young_gen->as_DefNewGeneration(); 1882 size_t max_eden_size = young_gen->max_capacity() - 1883 young_gen->to()->capacity() - 1884 young_gen->from()->capacity(); 1885 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1886 GCCause::Cause gc_cause = gch->gc_cause(); 1887 size_policy()->check_gc_overhead_limit(_young_gen->used(), 1888 young_gen->eden()->used(), 1889 _cmsGen->max_capacity(), 1890 max_eden_size, 1891 full, 1892 gc_cause, 1893 gch->collector_policy()); 1894 } else { 1895 do_mark_sweep_work(clear_all_soft_refs, first_state, 1896 should_start_over); 1897 } 1898 // Reset the expansion cause, now that we just completed 1899 // a collection cycle. 1900 clear_expansion_cause(); 1901 _foregroundGCIsActive = false; 1902 return; 1903 } 1904 1905 // Resize the perm generation and the tenured generation 1906 // after obtaining the free list locks for the 1907 // two generations. 1908 void CMSCollector::compute_new_size() { 1909 assert_locked_or_safepoint(Heap_lock); 1910 FreelistLocker z(this); 1911 _permGen->compute_new_size(); 1912 _cmsGen->compute_new_size(); 1913 } 1914 1915 // A work method used by foreground collection to determine 1916 // what type of collection (compacting or not, continuing or fresh) 1917 // it should do. 1918 // NOTE: the intent is to make UseCMSCompactAtFullCollection 1919 // and CMSCompactWhenClearAllSoftRefs the default in the future 1920 // and do away with the flags after a suitable period. 1921 void CMSCollector::decide_foreground_collection_type( 1922 bool clear_all_soft_refs, bool* should_compact, 1923 bool* should_start_over) { 1924 // Normally, we'll compact only if the UseCMSCompactAtFullCollection 1925 // flag is set, and we have either requested a System.gc() or 1926 // the number of full gc's since the last concurrent cycle 1927 // has exceeded the threshold set by CMSFullGCsBeforeCompaction, 1928 // or if an incremental collection has failed 1929 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1930 assert(gch->collector_policy()->is_two_generation_policy(), 1931 "You may want to check the correctness of the following"); 1932 // Inform cms gen if this was due to partial collection failing. 1933 // The CMS gen may use this fact to determine its expansion policy. 1934 if (gch->incremental_collection_will_fail(false /* don't consult_young */)) { 1935 assert(!_cmsGen->incremental_collection_failed(), 1936 "Should have been noticed, reacted to and cleared"); 1937 _cmsGen->set_incremental_collection_failed(); 1938 } 1939 *should_compact = 1940 UseCMSCompactAtFullCollection && 1941 ((_full_gcs_since_conc_gc >= CMSFullGCsBeforeCompaction) || 1942 GCCause::is_user_requested_gc(gch->gc_cause()) || 1943 gch->incremental_collection_will_fail(true /* consult_young */)); 1944 *should_start_over = false; 1945 if (clear_all_soft_refs && !*should_compact) { 1946 // We are about to do a last ditch collection attempt 1947 // so it would normally make sense to do a compaction 1948 // to reclaim as much space as possible. 1949 if (CMSCompactWhenClearAllSoftRefs) { 1950 // Default: The rationale is that in this case either 1951 // we are past the final marking phase, in which case 1952 // we'd have to start over, or so little has been done 1953 // that there's little point in saving that work. Compaction 1954 // appears to be the sensible choice in either case. 1955 *should_compact = true; 1956 } else { 1957 // We have been asked to clear all soft refs, but not to 1958 // compact. Make sure that we aren't past the final checkpoint 1959 // phase, for that is where we process soft refs. If we are already 1960 // past that phase, we'll need to redo the refs discovery phase and 1961 // if necessary clear soft refs that weren't previously 1962 // cleared. We do so by remembering the phase in which 1963 // we came in, and if we are past the refs processing 1964 // phase, we'll choose to just redo the mark-sweep 1965 // collection from scratch. 1966 if (_collectorState > FinalMarking) { 1967 // We are past the refs processing phase; 1968 // start over and do a fresh synchronous CMS cycle 1969 _collectorState = Resetting; // skip to reset to start new cycle 1970 reset(false /* == !asynch */); 1971 *should_start_over = true; 1972 } // else we can continue a possibly ongoing current cycle 1973 } 1974 } 1975 } 1976 1977 // A work method used by the foreground collector to do 1978 // a mark-sweep-compact. 1979 void CMSCollector::do_compaction_work(bool clear_all_soft_refs) { 1980 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1981 TraceTime t("CMS:MSC ", PrintGCDetails && Verbose, true, gclog_or_tty); 1982 if (PrintGC && Verbose && !(GCCause::is_user_requested_gc(gch->gc_cause()))) { 1983 gclog_or_tty->print_cr("Compact ConcurrentMarkSweepGeneration after %d " 1984 "collections passed to foreground collector", _full_gcs_since_conc_gc); 1985 } 1986 1987 // Sample collection interval time and reset for collection pause. 1988 if (UseAdaptiveSizePolicy) { 1989 size_policy()->msc_collection_begin(); 1990 } 1991 1992 // Temporarily widen the span of the weak reference processing to 1993 // the entire heap. 1994 MemRegion new_span(GenCollectedHeap::heap()->reserved_region()); 1995 ReferenceProcessorSpanMutator rp_mut_span(ref_processor(), new_span); 1996 // Temporarily, clear the "is_alive_non_header" field of the 1997 // reference processor. 1998 ReferenceProcessorIsAliveMutator rp_mut_closure(ref_processor(), NULL); 1999 // Temporarily make reference _processing_ single threaded (non-MT). 2000 ReferenceProcessorMTProcMutator rp_mut_mt_processing(ref_processor(), false); 2001 // Temporarily make refs discovery atomic 2002 ReferenceProcessorAtomicMutator rp_mut_atomic(ref_processor(), true); 2003 // Temporarily make reference _discovery_ single threaded (non-MT) 2004 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false); 2005 2006 ref_processor()->set_enqueuing_is_done(false); 2007 ref_processor()->enable_discovery(false /*verify_disabled*/, false /*check_no_refs*/); 2008 ref_processor()->setup_policy(clear_all_soft_refs); 2009 // If an asynchronous collection finishes, the _modUnionTable is 2010 // all clear. If we are assuming the collection from an asynchronous 2011 // collection, clear the _modUnionTable. 2012 assert(_collectorState != Idling || _modUnionTable.isAllClear(), 2013 "_modUnionTable should be clear if the baton was not passed"); 2014 _modUnionTable.clear_all(); 2015 2016 // We must adjust the allocation statistics being maintained 2017 // in the free list space. We do so by reading and clearing 2018 // the sweep timer and updating the block flux rate estimates below. 2019 assert(!_intra_sweep_timer.is_active(), "_intra_sweep_timer should be inactive"); 2020 if (_inter_sweep_timer.is_active()) { 2021 _inter_sweep_timer.stop(); 2022 // Note that we do not use this sample to update the _inter_sweep_estimate. 2023 _cmsGen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()), 2024 _inter_sweep_estimate.padded_average(), 2025 _intra_sweep_estimate.padded_average()); 2026 } 2027 2028 GenMarkSweep::invoke_at_safepoint(_cmsGen->level(), 2029 ref_processor(), clear_all_soft_refs); 2030 #ifdef ASSERT 2031 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 2032 size_t free_size = cms_space->free(); 2033 assert(free_size == 2034 pointer_delta(cms_space->end(), cms_space->compaction_top()) 2035 * HeapWordSize, 2036 "All the free space should be compacted into one chunk at top"); 2037 assert(cms_space->dictionary()->total_chunk_size( 2038 debug_only(cms_space->freelistLock())) == 0 || 2039 cms_space->totalSizeInIndexedFreeLists() == 0, 2040 "All the free space should be in a single chunk"); 2041 size_t num = cms_space->totalCount(); 2042 assert((free_size == 0 && num == 0) || 2043 (free_size > 0 && (num == 1 || num == 2)), 2044 "There should be at most 2 free chunks after compaction"); 2045 #endif // ASSERT 2046 _collectorState = Resetting; 2047 assert(_restart_addr == NULL, 2048 "Should have been NULL'd before baton was passed"); 2049 reset(false /* == !asynch */); 2050 _cmsGen->reset_after_compaction(); 2051 _concurrent_cycles_since_last_unload = 0; 2052 2053 if (verifying() && !should_unload_classes()) { 2054 perm_gen_verify_bit_map()->clear_all(); 2055 } 2056 2057 // Clear any data recorded in the PLAB chunk arrays. 2058 if (_survivor_plab_array != NULL) { 2059 reset_survivor_plab_arrays(); 2060 } 2061 2062 // Adjust the per-size allocation stats for the next epoch. 2063 _cmsGen->cmsSpace()->endSweepFLCensus(sweep_count() /* fake */); 2064 // Restart the "inter sweep timer" for the next epoch. 2065 _inter_sweep_timer.reset(); 2066 _inter_sweep_timer.start(); 2067 2068 // Sample collection pause time and reset for collection interval. 2069 if (UseAdaptiveSizePolicy) { 2070 size_policy()->msc_collection_end(gch->gc_cause()); 2071 } 2072 2073 // For a mark-sweep-compact, compute_new_size() will be called 2074 // in the heap's do_collection() method. 2075 } 2076 2077 // A work method used by the foreground collector to do 2078 // a mark-sweep, after taking over from a possibly on-going 2079 // concurrent mark-sweep collection. 2080 void CMSCollector::do_mark_sweep_work(bool clear_all_soft_refs, 2081 CollectorState first_state, bool should_start_over) { 2082 if (PrintGC && Verbose) { 2083 gclog_or_tty->print_cr("Pass concurrent collection to foreground " 2084 "collector with count %d", 2085 _full_gcs_since_conc_gc); 2086 } 2087 switch (_collectorState) { 2088 case Idling: 2089 if (first_state == Idling || should_start_over) { 2090 // The background GC was not active, or should 2091 // restarted from scratch; start the cycle. 2092 _collectorState = InitialMarking; 2093 } 2094 // If first_state was not Idling, then a background GC 2095 // was in progress and has now finished. No need to do it 2096 // again. Leave the state as Idling. 2097 break; 2098 case Precleaning: 2099 // In the foreground case don't do the precleaning since 2100 // it is not done concurrently and there is extra work 2101 // required. 2102 _collectorState = FinalMarking; 2103 } 2104 if (PrintGCDetails && 2105 (_collectorState > Idling || 2106 !GCCause::is_user_requested_gc(GenCollectedHeap::heap()->gc_cause()))) { 2107 gclog_or_tty->print(" (concurrent mode failure)"); 2108 } 2109 collect_in_foreground(clear_all_soft_refs); 2110 2111 // For a mark-sweep, compute_new_size() will be called 2112 // in the heap's do_collection() method. 2113 } 2114 2115 2116 void CMSCollector::getFreelistLocks() const { 2117 // Get locks for all free lists in all generations that this 2118 // collector is responsible for 2119 _cmsGen->freelistLock()->lock_without_safepoint_check(); 2120 _permGen->freelistLock()->lock_without_safepoint_check(); 2121 } 2122 2123 void CMSCollector::releaseFreelistLocks() const { 2124 // Release locks for all free lists in all generations that this 2125 // collector is responsible for 2126 _cmsGen->freelistLock()->unlock(); 2127 _permGen->freelistLock()->unlock(); 2128 } 2129 2130 bool CMSCollector::haveFreelistLocks() const { 2131 // Check locks for all free lists in all generations that this 2132 // collector is responsible for 2133 assert_lock_strong(_cmsGen->freelistLock()); 2134 assert_lock_strong(_permGen->freelistLock()); 2135 PRODUCT_ONLY(ShouldNotReachHere()); 2136 return true; 2137 } 2138 2139 // A utility class that is used by the CMS collector to 2140 // temporarily "release" the foreground collector from its 2141 // usual obligation to wait for the background collector to 2142 // complete an ongoing phase before proceeding. 2143 class ReleaseForegroundGC: public StackObj { 2144 private: 2145 CMSCollector* _c; 2146 public: 2147 ReleaseForegroundGC(CMSCollector* c) : _c(c) { 2148 assert(_c->_foregroundGCShouldWait, "Else should not need to call"); 2149 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 2150 // allow a potentially blocked foreground collector to proceed 2151 _c->_foregroundGCShouldWait = false; 2152 if (_c->_foregroundGCIsActive) { 2153 CGC_lock->notify(); 2154 } 2155 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 2156 "Possible deadlock"); 2157 } 2158 2159 ~ReleaseForegroundGC() { 2160 assert(!_c->_foregroundGCShouldWait, "Usage protocol violation?"); 2161 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 2162 _c->_foregroundGCShouldWait = true; 2163 } 2164 }; 2165 2166 // There are separate collect_in_background and collect_in_foreground because of 2167 // the different locking requirements of the background collector and the 2168 // foreground collector. There was originally an attempt to share 2169 // one "collect" method between the background collector and the foreground 2170 // collector but the if-then-else required made it cleaner to have 2171 // separate methods. 2172 void CMSCollector::collect_in_background(bool clear_all_soft_refs) { 2173 assert(Thread::current()->is_ConcurrentGC_thread(), 2174 "A CMS asynchronous collection is only allowed on a CMS thread."); 2175 2176 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2177 { 2178 bool safepoint_check = Mutex::_no_safepoint_check_flag; 2179 MutexLockerEx hl(Heap_lock, safepoint_check); 2180 FreelistLocker fll(this); 2181 MutexLockerEx x(CGC_lock, safepoint_check); 2182 if (_foregroundGCIsActive || !UseAsyncConcMarkSweepGC) { 2183 // The foreground collector is active or we're 2184 // not using asynchronous collections. Skip this 2185 // background collection. 2186 assert(!_foregroundGCShouldWait, "Should be clear"); 2187 return; 2188 } else { 2189 assert(_collectorState == Idling, "Should be idling before start."); 2190 _collectorState = InitialMarking; 2191 // Reset the expansion cause, now that we are about to begin 2192 // a new cycle. 2193 clear_expansion_cause(); 2194 } 2195 // Decide if we want to enable class unloading as part of the 2196 // ensuing concurrent GC cycle. 2197 update_should_unload_classes(); 2198 _full_gc_requested = false; // acks all outstanding full gc requests 2199 // Signal that we are about to start a collection 2200 gch->increment_total_full_collections(); // ... starting a collection cycle 2201 _collection_count_start = gch->total_full_collections(); 2202 } 2203 2204 // Used for PrintGC 2205 size_t prev_used; 2206 if (PrintGC && Verbose) { 2207 prev_used = _cmsGen->used(); // XXXPERM 2208 } 2209 2210 // The change of the collection state is normally done at this level; 2211 // the exceptions are phases that are executed while the world is 2212 // stopped. For those phases the change of state is done while the 2213 // world is stopped. For baton passing purposes this allows the 2214 // background collector to finish the phase and change state atomically. 2215 // The foreground collector cannot wait on a phase that is done 2216 // while the world is stopped because the foreground collector already 2217 // has the world stopped and would deadlock. 2218 while (_collectorState != Idling) { 2219 if (TraceCMSState) { 2220 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " in CMS state %d", 2221 Thread::current(), _collectorState); 2222 } 2223 // The foreground collector 2224 // holds the Heap_lock throughout its collection. 2225 // holds the CMS token (but not the lock) 2226 // except while it is waiting for the background collector to yield. 2227 // 2228 // The foreground collector should be blocked (not for long) 2229 // if the background collector is about to start a phase 2230 // executed with world stopped. If the background 2231 // collector has already started such a phase, the 2232 // foreground collector is blocked waiting for the 2233 // Heap_lock. The stop-world phases (InitialMarking and FinalMarking) 2234 // are executed in the VM thread. 2235 // 2236 // The locking order is 2237 // PendingListLock (PLL) -- if applicable (FinalMarking) 2238 // Heap_lock (both this & PLL locked in VM_CMS_Operation::prologue()) 2239 // CMS token (claimed in 2240 // stop_world_and_do() --> 2241 // safepoint_synchronize() --> 2242 // CMSThread::synchronize()) 2243 2244 { 2245 // Check if the FG collector wants us to yield. 2246 CMSTokenSync x(true); // is cms thread 2247 if (waitForForegroundGC()) { 2248 // We yielded to a foreground GC, nothing more to be 2249 // done this round. 2250 assert(_foregroundGCShouldWait == false, "We set it to false in " 2251 "waitForForegroundGC()"); 2252 if (TraceCMSState) { 2253 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT 2254 " exiting collection CMS state %d", 2255 Thread::current(), _collectorState); 2256 } 2257 return; 2258 } else { 2259 // The background collector can run but check to see if the 2260 // foreground collector has done a collection while the 2261 // background collector was waiting to get the CGC_lock 2262 // above. If yes, break so that _foregroundGCShouldWait 2263 // is cleared before returning. 2264 if (_collectorState == Idling) { 2265 break; 2266 } 2267 } 2268 } 2269 2270 assert(_foregroundGCShouldWait, "Foreground collector, if active, " 2271 "should be waiting"); 2272 2273 switch (_collectorState) { 2274 case InitialMarking: 2275 { 2276 ReleaseForegroundGC x(this); 2277 stats().record_cms_begin(); 2278 2279 VM_CMS_Initial_Mark initial_mark_op(this); 2280 VMThread::execute(&initial_mark_op); 2281 } 2282 // The collector state may be any legal state at this point 2283 // since the background collector may have yielded to the 2284 // foreground collector. 2285 break; 2286 case Marking: 2287 // initial marking in checkpointRootsInitialWork has been completed 2288 if (markFromRoots(true)) { // we were successful 2289 assert(_collectorState == Precleaning, "Collector state should " 2290 "have changed"); 2291 } else { 2292 assert(_foregroundGCIsActive, "Internal state inconsistency"); 2293 } 2294 break; 2295 case Precleaning: 2296 if (UseAdaptiveSizePolicy) { 2297 size_policy()->concurrent_precleaning_begin(); 2298 } 2299 // marking from roots in markFromRoots has been completed 2300 preclean(); 2301 if (UseAdaptiveSizePolicy) { 2302 size_policy()->concurrent_precleaning_end(); 2303 } 2304 assert(_collectorState == AbortablePreclean || 2305 _collectorState == FinalMarking, 2306 "Collector state should have changed"); 2307 break; 2308 case AbortablePreclean: 2309 if (UseAdaptiveSizePolicy) { 2310 size_policy()->concurrent_phases_resume(); 2311 } 2312 abortable_preclean(); 2313 if (UseAdaptiveSizePolicy) { 2314 size_policy()->concurrent_precleaning_end(); 2315 } 2316 assert(_collectorState == FinalMarking, "Collector state should " 2317 "have changed"); 2318 break; 2319 case FinalMarking: 2320 { 2321 ReleaseForegroundGC x(this); 2322 2323 VM_CMS_Final_Remark final_remark_op(this); 2324 VMThread::execute(&final_remark_op); 2325 } 2326 assert(_foregroundGCShouldWait, "block post-condition"); 2327 break; 2328 case Sweeping: 2329 if (UseAdaptiveSizePolicy) { 2330 size_policy()->concurrent_sweeping_begin(); 2331 } 2332 // final marking in checkpointRootsFinal has been completed 2333 sweep(true); 2334 assert(_collectorState == Resizing, "Collector state change " 2335 "to Resizing must be done under the free_list_lock"); 2336 _full_gcs_since_conc_gc = 0; 2337 2338 // Stop the timers for adaptive size policy for the concurrent phases 2339 if (UseAdaptiveSizePolicy) { 2340 size_policy()->concurrent_sweeping_end(); 2341 size_policy()->concurrent_phases_end(gch->gc_cause(), 2342 gch->prev_gen(_cmsGen)->capacity(), 2343 _cmsGen->free()); 2344 } 2345 2346 case Resizing: { 2347 // Sweeping has been completed... 2348 // At this point the background collection has completed. 2349 // Don't move the call to compute_new_size() down 2350 // into code that might be executed if the background 2351 // collection was preempted. 2352 { 2353 ReleaseForegroundGC x(this); // unblock FG collection 2354 MutexLockerEx y(Heap_lock, Mutex::_no_safepoint_check_flag); 2355 CMSTokenSync z(true); // not strictly needed. 2356 if (_collectorState == Resizing) { 2357 compute_new_size(); 2358 _collectorState = Resetting; 2359 } else { 2360 assert(_collectorState == Idling, "The state should only change" 2361 " because the foreground collector has finished the collection"); 2362 } 2363 } 2364 break; 2365 } 2366 case Resetting: 2367 // CMS heap resizing has been completed 2368 reset(true); 2369 assert(_collectorState == Idling, "Collector state should " 2370 "have changed"); 2371 stats().record_cms_end(); 2372 // Don't move the concurrent_phases_end() and compute_new_size() 2373 // calls to here because a preempted background collection 2374 // has it's state set to "Resetting". 2375 break; 2376 case Idling: 2377 default: 2378 ShouldNotReachHere(); 2379 break; 2380 } 2381 if (TraceCMSState) { 2382 gclog_or_tty->print_cr(" Thread " INTPTR_FORMAT " done - next CMS state %d", 2383 Thread::current(), _collectorState); 2384 } 2385 assert(_foregroundGCShouldWait, "block post-condition"); 2386 } 2387 2388 // Should this be in gc_epilogue? 2389 collector_policy()->counters()->update_counters(); 2390 2391 { 2392 // Clear _foregroundGCShouldWait and, in the event that the 2393 // foreground collector is waiting, notify it, before 2394 // returning. 2395 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 2396 _foregroundGCShouldWait = false; 2397 if (_foregroundGCIsActive) { 2398 CGC_lock->notify(); 2399 } 2400 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 2401 "Possible deadlock"); 2402 } 2403 if (TraceCMSState) { 2404 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT 2405 " exiting collection CMS state %d", 2406 Thread::current(), _collectorState); 2407 } 2408 if (PrintGC && Verbose) { 2409 _cmsGen->print_heap_change(prev_used); 2410 } 2411 } 2412 2413 void CMSCollector::collect_in_foreground(bool clear_all_soft_refs) { 2414 assert(_foregroundGCIsActive && !_foregroundGCShouldWait, 2415 "Foreground collector should be waiting, not executing"); 2416 assert(Thread::current()->is_VM_thread(), "A foreground collection" 2417 "may only be done by the VM Thread with the world stopped"); 2418 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 2419 "VM thread should have CMS token"); 2420 2421 NOT_PRODUCT(TraceTime t("CMS:MS (foreground) ", PrintGCDetails && Verbose, 2422 true, gclog_or_tty);) 2423 if (UseAdaptiveSizePolicy) { 2424 size_policy()->ms_collection_begin(); 2425 } 2426 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact); 2427 2428 HandleMark hm; // Discard invalid handles created during verification 2429 2430 if (VerifyBeforeGC && 2431 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2432 Universe::verify(true); 2433 } 2434 2435 // Snapshot the soft reference policy to be used in this collection cycle. 2436 ref_processor()->setup_policy(clear_all_soft_refs); 2437 2438 bool init_mark_was_synchronous = false; // until proven otherwise 2439 while (_collectorState != Idling) { 2440 if (TraceCMSState) { 2441 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " in CMS state %d", 2442 Thread::current(), _collectorState); 2443 } 2444 switch (_collectorState) { 2445 case InitialMarking: 2446 init_mark_was_synchronous = true; // fact to be exploited in re-mark 2447 checkpointRootsInitial(false); 2448 assert(_collectorState == Marking, "Collector state should have changed" 2449 " within checkpointRootsInitial()"); 2450 break; 2451 case Marking: 2452 // initial marking in checkpointRootsInitialWork has been completed 2453 if (VerifyDuringGC && 2454 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2455 gclog_or_tty->print("Verify before initial mark: "); 2456 Universe::verify(true); 2457 } 2458 { 2459 bool res = markFromRoots(false); 2460 assert(res && _collectorState == FinalMarking, "Collector state should " 2461 "have changed"); 2462 break; 2463 } 2464 case FinalMarking: 2465 if (VerifyDuringGC && 2466 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2467 gclog_or_tty->print("Verify before re-mark: "); 2468 Universe::verify(true); 2469 } 2470 checkpointRootsFinal(false, clear_all_soft_refs, 2471 init_mark_was_synchronous); 2472 assert(_collectorState == Sweeping, "Collector state should not " 2473 "have changed within checkpointRootsFinal()"); 2474 break; 2475 case Sweeping: 2476 // final marking in checkpointRootsFinal has been completed 2477 if (VerifyDuringGC && 2478 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2479 gclog_or_tty->print("Verify before sweep: "); 2480 Universe::verify(true); 2481 } 2482 sweep(false); 2483 assert(_collectorState == Resizing, "Incorrect state"); 2484 break; 2485 case Resizing: { 2486 // Sweeping has been completed; the actual resize in this case 2487 // is done separately; nothing to be done in this state. 2488 _collectorState = Resetting; 2489 break; 2490 } 2491 case Resetting: 2492 // The heap has been resized. 2493 if (VerifyDuringGC && 2494 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2495 gclog_or_tty->print("Verify before reset: "); 2496 Universe::verify(true); 2497 } 2498 reset(false); 2499 assert(_collectorState == Idling, "Collector state should " 2500 "have changed"); 2501 break; 2502 case Precleaning: 2503 case AbortablePreclean: 2504 // Elide the preclean phase 2505 _collectorState = FinalMarking; 2506 break; 2507 default: 2508 ShouldNotReachHere(); 2509 } 2510 if (TraceCMSState) { 2511 gclog_or_tty->print_cr(" Thread " INTPTR_FORMAT " done - next CMS state %d", 2512 Thread::current(), _collectorState); 2513 } 2514 } 2515 2516 if (UseAdaptiveSizePolicy) { 2517 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2518 size_policy()->ms_collection_end(gch->gc_cause()); 2519 } 2520 2521 if (VerifyAfterGC && 2522 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2523 Universe::verify(true); 2524 } 2525 if (TraceCMSState) { 2526 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT 2527 " exiting collection CMS state %d", 2528 Thread::current(), _collectorState); 2529 } 2530 } 2531 2532 bool CMSCollector::waitForForegroundGC() { 2533 bool res = false; 2534 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 2535 "CMS thread should have CMS token"); 2536 // Block the foreground collector until the 2537 // background collectors decides whether to 2538 // yield. 2539 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 2540 _foregroundGCShouldWait = true; 2541 if (_foregroundGCIsActive) { 2542 // The background collector yields to the 2543 // foreground collector and returns a value 2544 // indicating that it has yielded. The foreground 2545 // collector can proceed. 2546 res = true; 2547 _foregroundGCShouldWait = false; 2548 ConcurrentMarkSweepThread::clear_CMS_flag( 2549 ConcurrentMarkSweepThread::CMS_cms_has_token); 2550 ConcurrentMarkSweepThread::set_CMS_flag( 2551 ConcurrentMarkSweepThread::CMS_cms_wants_token); 2552 // Get a possibly blocked foreground thread going 2553 CGC_lock->notify(); 2554 if (TraceCMSState) { 2555 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " waiting at CMS state %d", 2556 Thread::current(), _collectorState); 2557 } 2558 while (_foregroundGCIsActive) { 2559 CGC_lock->wait(Mutex::_no_safepoint_check_flag); 2560 } 2561 ConcurrentMarkSweepThread::set_CMS_flag( 2562 ConcurrentMarkSweepThread::CMS_cms_has_token); 2563 ConcurrentMarkSweepThread::clear_CMS_flag( 2564 ConcurrentMarkSweepThread::CMS_cms_wants_token); 2565 } 2566 if (TraceCMSState) { 2567 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " continuing at CMS state %d", 2568 Thread::current(), _collectorState); 2569 } 2570 return res; 2571 } 2572 2573 // Because of the need to lock the free lists and other structures in 2574 // the collector, common to all the generations that the collector is 2575 // collecting, we need the gc_prologues of individual CMS generations 2576 // delegate to their collector. It may have been simpler had the 2577 // current infrastructure allowed one to call a prologue on a 2578 // collector. In the absence of that we have the generation's 2579 // prologue delegate to the collector, which delegates back 2580 // some "local" work to a worker method in the individual generations 2581 // that it's responsible for collecting, while itself doing any 2582 // work common to all generations it's responsible for. A similar 2583 // comment applies to the gc_epilogue()'s. 2584 // The role of the varaible _between_prologue_and_epilogue is to 2585 // enforce the invocation protocol. 2586 void CMSCollector::gc_prologue(bool full) { 2587 // Call gc_prologue_work() for each CMSGen and PermGen that 2588 // we are responsible for. 2589 2590 // The following locking discipline assumes that we are only called 2591 // when the world is stopped. 2592 assert(SafepointSynchronize::is_at_safepoint(), "world is stopped assumption"); 2593 2594 // The CMSCollector prologue must call the gc_prologues for the 2595 // "generations" (including PermGen if any) that it's responsible 2596 // for. 2597 2598 assert( Thread::current()->is_VM_thread() 2599 || ( CMSScavengeBeforeRemark 2600 && Thread::current()->is_ConcurrentGC_thread()), 2601 "Incorrect thread type for prologue execution"); 2602 2603 if (_between_prologue_and_epilogue) { 2604 // We have already been invoked; this is a gc_prologue delegation 2605 // from yet another CMS generation that we are responsible for, just 2606 // ignore it since all relevant work has already been done. 2607 return; 2608 } 2609 2610 // set a bit saying prologue has been called; cleared in epilogue 2611 _between_prologue_and_epilogue = true; 2612 // Claim locks for common data structures, then call gc_prologue_work() 2613 // for each CMSGen and PermGen that we are responsible for. 2614 2615 getFreelistLocks(); // gets free list locks on constituent spaces 2616 bitMapLock()->lock_without_safepoint_check(); 2617 2618 // Should call gc_prologue_work() for all cms gens we are responsible for 2619 bool registerClosure = _collectorState >= Marking 2620 && _collectorState < Sweeping; 2621 ModUnionClosure* muc = CollectedHeap::use_parallel_gc_threads() ? 2622 &_modUnionClosurePar 2623 : &_modUnionClosure; 2624 _cmsGen->gc_prologue_work(full, registerClosure, muc); 2625 _permGen->gc_prologue_work(full, registerClosure, muc); 2626 2627 if (!full) { 2628 stats().record_gc0_begin(); 2629 } 2630 } 2631 2632 void ConcurrentMarkSweepGeneration::gc_prologue(bool full) { 2633 // Delegate to CMScollector which knows how to coordinate between 2634 // this and any other CMS generations that it is responsible for 2635 // collecting. 2636 collector()->gc_prologue(full); 2637 } 2638 2639 // This is a "private" interface for use by this generation's CMSCollector. 2640 // Not to be called directly by any other entity (for instance, 2641 // GenCollectedHeap, which calls the "public" gc_prologue method above). 2642 void ConcurrentMarkSweepGeneration::gc_prologue_work(bool full, 2643 bool registerClosure, ModUnionClosure* modUnionClosure) { 2644 assert(!incremental_collection_failed(), "Shouldn't be set yet"); 2645 assert(cmsSpace()->preconsumptionDirtyCardClosure() == NULL, 2646 "Should be NULL"); 2647 if (registerClosure) { 2648 cmsSpace()->setPreconsumptionDirtyCardClosure(modUnionClosure); 2649 } 2650 cmsSpace()->gc_prologue(); 2651 // Clear stat counters 2652 NOT_PRODUCT( 2653 assert(_numObjectsPromoted == 0, "check"); 2654 assert(_numWordsPromoted == 0, "check"); 2655 if (Verbose && PrintGC) { 2656 gclog_or_tty->print("Allocated "SIZE_FORMAT" objects, " 2657 SIZE_FORMAT" bytes concurrently", 2658 _numObjectsAllocated, _numWordsAllocated*sizeof(HeapWord)); 2659 } 2660 _numObjectsAllocated = 0; 2661 _numWordsAllocated = 0; 2662 ) 2663 } 2664 2665 void CMSCollector::gc_epilogue(bool full) { 2666 // The following locking discipline assumes that we are only called 2667 // when the world is stopped. 2668 assert(SafepointSynchronize::is_at_safepoint(), 2669 "world is stopped assumption"); 2670 2671 // Currently the CMS epilogue (see CompactibleFreeListSpace) merely checks 2672 // if linear allocation blocks need to be appropriately marked to allow the 2673 // the blocks to be parsable. We also check here whether we need to nudge the 2674 // CMS collector thread to start a new cycle (if it's not already active). 2675 assert( Thread::current()->is_VM_thread() 2676 || ( CMSScavengeBeforeRemark 2677 && Thread::current()->is_ConcurrentGC_thread()), 2678 "Incorrect thread type for epilogue execution"); 2679 2680 if (!_between_prologue_and_epilogue) { 2681 // We have already been invoked; this is a gc_epilogue delegation 2682 // from yet another CMS generation that we are responsible for, just 2683 // ignore it since all relevant work has already been done. 2684 return; 2685 } 2686 assert(haveFreelistLocks(), "must have freelist locks"); 2687 assert_lock_strong(bitMapLock()); 2688 2689 _cmsGen->gc_epilogue_work(full); 2690 _permGen->gc_epilogue_work(full); 2691 2692 if (_collectorState == AbortablePreclean || _collectorState == Precleaning) { 2693 // in case sampling was not already enabled, enable it 2694 _start_sampling = true; 2695 } 2696 // reset _eden_chunk_array so sampling starts afresh 2697 _eden_chunk_index = 0; 2698 2699 size_t cms_used = _cmsGen->cmsSpace()->used(); 2700 size_t perm_used = _permGen->cmsSpace()->used(); 2701 2702 // update performance counters - this uses a special version of 2703 // update_counters() that allows the utilization to be passed as a 2704 // parameter, avoiding multiple calls to used(). 2705 // 2706 _cmsGen->update_counters(cms_used); 2707 _permGen->update_counters(perm_used); 2708 2709 if (CMSIncrementalMode) { 2710 icms_update_allocation_limits(); 2711 } 2712 2713 bitMapLock()->unlock(); 2714 releaseFreelistLocks(); 2715 2716 if (!CleanChunkPoolAsync) { 2717 Chunk::clean_chunk_pool(); 2718 } 2719 2720 _between_prologue_and_epilogue = false; // ready for next cycle 2721 } 2722 2723 void ConcurrentMarkSweepGeneration::gc_epilogue(bool full) { 2724 collector()->gc_epilogue(full); 2725 2726 // Also reset promotion tracking in par gc thread states. 2727 if (CollectedHeap::use_parallel_gc_threads()) { 2728 for (uint i = 0; i < ParallelGCThreads; i++) { 2729 _par_gc_thread_states[i]->promo.stopTrackingPromotions(i); 2730 } 2731 } 2732 } 2733 2734 void ConcurrentMarkSweepGeneration::gc_epilogue_work(bool full) { 2735 assert(!incremental_collection_failed(), "Should have been cleared"); 2736 cmsSpace()->setPreconsumptionDirtyCardClosure(NULL); 2737 cmsSpace()->gc_epilogue(); 2738 // Print stat counters 2739 NOT_PRODUCT( 2740 assert(_numObjectsAllocated == 0, "check"); 2741 assert(_numWordsAllocated == 0, "check"); 2742 if (Verbose && PrintGC) { 2743 gclog_or_tty->print("Promoted "SIZE_FORMAT" objects, " 2744 SIZE_FORMAT" bytes", 2745 _numObjectsPromoted, _numWordsPromoted*sizeof(HeapWord)); 2746 } 2747 _numObjectsPromoted = 0; 2748 _numWordsPromoted = 0; 2749 ) 2750 2751 if (PrintGC && Verbose) { 2752 // Call down the chain in contiguous_available needs the freelistLock 2753 // so print this out before releasing the freeListLock. 2754 gclog_or_tty->print(" Contiguous available "SIZE_FORMAT" bytes ", 2755 contiguous_available()); 2756 } 2757 } 2758 2759 #ifndef PRODUCT 2760 bool CMSCollector::have_cms_token() { 2761 Thread* thr = Thread::current(); 2762 if (thr->is_VM_thread()) { 2763 return ConcurrentMarkSweepThread::vm_thread_has_cms_token(); 2764 } else if (thr->is_ConcurrentGC_thread()) { 2765 return ConcurrentMarkSweepThread::cms_thread_has_cms_token(); 2766 } else if (thr->is_GC_task_thread()) { 2767 return ConcurrentMarkSweepThread::vm_thread_has_cms_token() && 2768 ParGCRareEvent_lock->owned_by_self(); 2769 } 2770 return false; 2771 } 2772 #endif 2773 2774 // Check reachability of the given heap address in CMS generation, 2775 // treating all other generations as roots. 2776 bool CMSCollector::is_cms_reachable(HeapWord* addr) { 2777 // We could "guarantee" below, rather than assert, but i'll 2778 // leave these as "asserts" so that an adventurous debugger 2779 // could try this in the product build provided some subset of 2780 // the conditions were met, provided they were intersted in the 2781 // results and knew that the computation below wouldn't interfere 2782 // with other concurrent computations mutating the structures 2783 // being read or written. 2784 assert(SafepointSynchronize::is_at_safepoint(), 2785 "Else mutations in object graph will make answer suspect"); 2786 assert(have_cms_token(), "Should hold cms token"); 2787 assert(haveFreelistLocks(), "must hold free list locks"); 2788 assert_lock_strong(bitMapLock()); 2789 2790 // Clear the marking bit map array before starting, but, just 2791 // for kicks, first report if the given address is already marked 2792 gclog_or_tty->print_cr("Start: Address 0x%x is%s marked", addr, 2793 _markBitMap.isMarked(addr) ? "" : " not"); 2794 2795 if (verify_after_remark()) { 2796 MutexLockerEx x(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag); 2797 bool result = verification_mark_bm()->isMarked(addr); 2798 gclog_or_tty->print_cr("TransitiveMark: Address 0x%x %s marked", addr, 2799 result ? "IS" : "is NOT"); 2800 return result; 2801 } else { 2802 gclog_or_tty->print_cr("Could not compute result"); 2803 return false; 2804 } 2805 } 2806 2807 //////////////////////////////////////////////////////// 2808 // CMS Verification Support 2809 //////////////////////////////////////////////////////// 2810 // Following the remark phase, the following invariant 2811 // should hold -- each object in the CMS heap which is 2812 // marked in markBitMap() should be marked in the verification_mark_bm(). 2813 2814 class VerifyMarkedClosure: public BitMapClosure { 2815 CMSBitMap* _marks; 2816 bool _failed; 2817 2818 public: 2819 VerifyMarkedClosure(CMSBitMap* bm): _marks(bm), _failed(false) {} 2820 2821 bool do_bit(size_t offset) { 2822 HeapWord* addr = _marks->offsetToHeapWord(offset); 2823 if (!_marks->isMarked(addr)) { 2824 oop(addr)->print_on(gclog_or_tty); 2825 gclog_or_tty->print_cr(" ("INTPTR_FORMAT" should have been marked)", addr); 2826 _failed = true; 2827 } 2828 return true; 2829 } 2830 2831 bool failed() { return _failed; } 2832 }; 2833 2834 bool CMSCollector::verify_after_remark() { 2835 gclog_or_tty->print(" [Verifying CMS Marking... "); 2836 MutexLockerEx ml(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag); 2837 static bool init = false; 2838 2839 assert(SafepointSynchronize::is_at_safepoint(), 2840 "Else mutations in object graph will make answer suspect"); 2841 assert(have_cms_token(), 2842 "Else there may be mutual interference in use of " 2843 " verification data structures"); 2844 assert(_collectorState > Marking && _collectorState <= Sweeping, 2845 "Else marking info checked here may be obsolete"); 2846 assert(haveFreelistLocks(), "must hold free list locks"); 2847 assert_lock_strong(bitMapLock()); 2848 2849 2850 // Allocate marking bit map if not already allocated 2851 if (!init) { // first time 2852 if (!verification_mark_bm()->allocate(_span)) { 2853 return false; 2854 } 2855 init = true; 2856 } 2857 2858 assert(verification_mark_stack()->isEmpty(), "Should be empty"); 2859 2860 // Turn off refs discovery -- so we will be tracing through refs. 2861 // This is as intended, because by this time 2862 // GC must already have cleared any refs that need to be cleared, 2863 // and traced those that need to be marked; moreover, 2864 // the marking done here is not going to intefere in any 2865 // way with the marking information used by GC. 2866 NoRefDiscovery no_discovery(ref_processor()); 2867 2868 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;) 2869 2870 // Clear any marks from a previous round 2871 verification_mark_bm()->clear_all(); 2872 assert(verification_mark_stack()->isEmpty(), "markStack should be empty"); 2873 verify_work_stacks_empty(); 2874 2875 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2876 gch->ensure_parsability(false); // fill TLABs, but no need to retire them 2877 // Update the saved marks which may affect the root scans. 2878 gch->save_marks(); 2879 2880 if (CMSRemarkVerifyVariant == 1) { 2881 // In this first variant of verification, we complete 2882 // all marking, then check if the new marks-verctor is 2883 // a subset of the CMS marks-vector. 2884 verify_after_remark_work_1(); 2885 } else if (CMSRemarkVerifyVariant == 2) { 2886 // In this second variant of verification, we flag an error 2887 // (i.e. an object reachable in the new marks-vector not reachable 2888 // in the CMS marks-vector) immediately, also indicating the 2889 // identify of an object (A) that references the unmarked object (B) -- 2890 // presumably, a mutation to A failed to be picked up by preclean/remark? 2891 verify_after_remark_work_2(); 2892 } else { 2893 warning("Unrecognized value %d for CMSRemarkVerifyVariant", 2894 CMSRemarkVerifyVariant); 2895 } 2896 gclog_or_tty->print(" done] "); 2897 return true; 2898 } 2899 2900 void CMSCollector::verify_after_remark_work_1() { 2901 ResourceMark rm; 2902 HandleMark hm; 2903 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2904 2905 // Mark from roots one level into CMS 2906 MarkRefsIntoClosure notOlder(_span, verification_mark_bm()); 2907 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2908 2909 gch->gen_process_strong_roots(_cmsGen->level(), 2910 true, // younger gens are roots 2911 true, // activate StrongRootsScope 2912 true, // collecting perm gen 2913 SharedHeap::ScanningOption(roots_scanning_options()), 2914 ¬Older, 2915 true, // walk code active on stacks 2916 NULL); 2917 2918 // Now mark from the roots 2919 assert(_revisitStack.isEmpty(), "Should be empty"); 2920 MarkFromRootsClosure markFromRootsClosure(this, _span, 2921 verification_mark_bm(), verification_mark_stack(), &_revisitStack, 2922 false /* don't yield */, true /* verifying */); 2923 assert(_restart_addr == NULL, "Expected pre-condition"); 2924 verification_mark_bm()->iterate(&markFromRootsClosure); 2925 while (_restart_addr != NULL) { 2926 // Deal with stack overflow: by restarting at the indicated 2927 // address. 2928 HeapWord* ra = _restart_addr; 2929 markFromRootsClosure.reset(ra); 2930 _restart_addr = NULL; 2931 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end()); 2932 } 2933 assert(verification_mark_stack()->isEmpty(), "Should have been drained"); 2934 verify_work_stacks_empty(); 2935 // Should reset the revisit stack above, since no class tree 2936 // surgery is forthcoming. 2937 _revisitStack.reset(); // throwing away all contents 2938 2939 // Marking completed -- now verify that each bit marked in 2940 // verification_mark_bm() is also marked in markBitMap(); flag all 2941 // errors by printing corresponding objects. 2942 VerifyMarkedClosure vcl(markBitMap()); 2943 verification_mark_bm()->iterate(&vcl); 2944 if (vcl.failed()) { 2945 gclog_or_tty->print("Verification failed"); 2946 Universe::heap()->print_on(gclog_or_tty); 2947 fatal("CMS: failed marking verification after remark"); 2948 } 2949 } 2950 2951 void CMSCollector::verify_after_remark_work_2() { 2952 ResourceMark rm; 2953 HandleMark hm; 2954 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2955 2956 // Mark from roots one level into CMS 2957 MarkRefsIntoVerifyClosure notOlder(_span, verification_mark_bm(), 2958 markBitMap()); 2959 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2960 gch->gen_process_strong_roots(_cmsGen->level(), 2961 true, // younger gens are roots 2962 true, // activate StrongRootsScope 2963 true, // collecting perm gen 2964 SharedHeap::ScanningOption(roots_scanning_options()), 2965 ¬Older, 2966 true, // walk code active on stacks 2967 NULL); 2968 2969 // Now mark from the roots 2970 assert(_revisitStack.isEmpty(), "Should be empty"); 2971 MarkFromRootsVerifyClosure markFromRootsClosure(this, _span, 2972 verification_mark_bm(), markBitMap(), verification_mark_stack()); 2973 assert(_restart_addr == NULL, "Expected pre-condition"); 2974 verification_mark_bm()->iterate(&markFromRootsClosure); 2975 while (_restart_addr != NULL) { 2976 // Deal with stack overflow: by restarting at the indicated 2977 // address. 2978 HeapWord* ra = _restart_addr; 2979 markFromRootsClosure.reset(ra); 2980 _restart_addr = NULL; 2981 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end()); 2982 } 2983 assert(verification_mark_stack()->isEmpty(), "Should have been drained"); 2984 verify_work_stacks_empty(); 2985 // Should reset the revisit stack above, since no class tree 2986 // surgery is forthcoming. 2987 _revisitStack.reset(); // throwing away all contents 2988 2989 // Marking completed -- now verify that each bit marked in 2990 // verification_mark_bm() is also marked in markBitMap(); flag all 2991 // errors by printing corresponding objects. 2992 VerifyMarkedClosure vcl(markBitMap()); 2993 verification_mark_bm()->iterate(&vcl); 2994 assert(!vcl.failed(), "Else verification above should not have succeeded"); 2995 } 2996 2997 void ConcurrentMarkSweepGeneration::save_marks() { 2998 // delegate to CMS space 2999 cmsSpace()->save_marks(); 3000 for (uint i = 0; i < ParallelGCThreads; i++) { 3001 _par_gc_thread_states[i]->promo.startTrackingPromotions(); 3002 } 3003 } 3004 3005 bool ConcurrentMarkSweepGeneration::no_allocs_since_save_marks() { 3006 return cmsSpace()->no_allocs_since_save_marks(); 3007 } 3008 3009 #define CMS_SINCE_SAVE_MARKS_DEFN(OopClosureType, nv_suffix) \ 3010 \ 3011 void ConcurrentMarkSweepGeneration:: \ 3012 oop_since_save_marks_iterate##nv_suffix(OopClosureType* cl) { \ 3013 cl->set_generation(this); \ 3014 cmsSpace()->oop_since_save_marks_iterate##nv_suffix(cl); \ 3015 cl->reset_generation(); \ 3016 save_marks(); \ 3017 } 3018 3019 ALL_SINCE_SAVE_MARKS_CLOSURES(CMS_SINCE_SAVE_MARKS_DEFN) 3020 3021 void 3022 ConcurrentMarkSweepGeneration::object_iterate_since_last_GC(ObjectClosure* blk) 3023 { 3024 // Not currently implemented; need to do the following. -- ysr. 3025 // dld -- I think that is used for some sort of allocation profiler. So it 3026 // really means the objects allocated by the mutator since the last 3027 // GC. We could potentially implement this cheaply by recording only 3028 // the direct allocations in a side data structure. 3029 // 3030 // I think we probably ought not to be required to support these 3031 // iterations at any arbitrary point; I think there ought to be some 3032 // call to enable/disable allocation profiling in a generation/space, 3033 // and the iterator ought to return the objects allocated in the 3034 // gen/space since the enable call, or the last iterator call (which 3035 // will probably be at a GC.) That way, for gens like CM&S that would 3036 // require some extra data structure to support this, we only pay the 3037 // cost when it's in use... 3038 cmsSpace()->object_iterate_since_last_GC(blk); 3039 } 3040 3041 void 3042 ConcurrentMarkSweepGeneration::younger_refs_iterate(OopsInGenClosure* cl) { 3043 cl->set_generation(this); 3044 younger_refs_in_space_iterate(_cmsSpace, cl); 3045 cl->reset_generation(); 3046 } 3047 3048 void 3049 ConcurrentMarkSweepGeneration::oop_iterate(MemRegion mr, OopClosure* cl) { 3050 if (freelistLock()->owned_by_self()) { 3051 Generation::oop_iterate(mr, cl); 3052 } else { 3053 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 3054 Generation::oop_iterate(mr, cl); 3055 } 3056 } 3057 3058 void 3059 ConcurrentMarkSweepGeneration::oop_iterate(OopClosure* cl) { 3060 if (freelistLock()->owned_by_self()) { 3061 Generation::oop_iterate(cl); 3062 } else { 3063 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 3064 Generation::oop_iterate(cl); 3065 } 3066 } 3067 3068 void 3069 ConcurrentMarkSweepGeneration::object_iterate(ObjectClosure* cl) { 3070 if (freelistLock()->owned_by_self()) { 3071 Generation::object_iterate(cl); 3072 } else { 3073 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 3074 Generation::object_iterate(cl); 3075 } 3076 } 3077 3078 void 3079 ConcurrentMarkSweepGeneration::safe_object_iterate(ObjectClosure* cl) { 3080 if (freelistLock()->owned_by_self()) { 3081 Generation::safe_object_iterate(cl); 3082 } else { 3083 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 3084 Generation::safe_object_iterate(cl); 3085 } 3086 } 3087 3088 void 3089 ConcurrentMarkSweepGeneration::pre_adjust_pointers() { 3090 } 3091 3092 void 3093 ConcurrentMarkSweepGeneration::post_compact() { 3094 } 3095 3096 void 3097 ConcurrentMarkSweepGeneration::prepare_for_verify() { 3098 // Fix the linear allocation blocks to look like free blocks. 3099 3100 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those 3101 // are not called when the heap is verified during universe initialization and 3102 // at vm shutdown. 3103 if (freelistLock()->owned_by_self()) { 3104 cmsSpace()->prepare_for_verify(); 3105 } else { 3106 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag); 3107 cmsSpace()->prepare_for_verify(); 3108 } 3109 } 3110 3111 void 3112 ConcurrentMarkSweepGeneration::verify() { 3113 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those 3114 // are not called when the heap is verified during universe initialization and 3115 // at vm shutdown. 3116 if (freelistLock()->owned_by_self()) { 3117 cmsSpace()->verify(); 3118 } else { 3119 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag); 3120 cmsSpace()->verify(); 3121 } 3122 } 3123 3124 void CMSCollector::verify() { 3125 _cmsGen->verify(); 3126 _permGen->verify(); 3127 } 3128 3129 #ifndef PRODUCT 3130 bool CMSCollector::overflow_list_is_empty() const { 3131 assert(_num_par_pushes >= 0, "Inconsistency"); 3132 if (_overflow_list == NULL) { 3133 assert(_num_par_pushes == 0, "Inconsistency"); 3134 } 3135 return _overflow_list == NULL; 3136 } 3137 3138 // The methods verify_work_stacks_empty() and verify_overflow_empty() 3139 // merely consolidate assertion checks that appear to occur together frequently. 3140 void CMSCollector::verify_work_stacks_empty() const { 3141 assert(_markStack.isEmpty(), "Marking stack should be empty"); 3142 assert(overflow_list_is_empty(), "Overflow list should be empty"); 3143 } 3144 3145 void CMSCollector::verify_overflow_empty() const { 3146 assert(overflow_list_is_empty(), "Overflow list should be empty"); 3147 assert(no_preserved_marks(), "No preserved marks"); 3148 } 3149 #endif // PRODUCT 3150 3151 // Decide if we want to enable class unloading as part of the 3152 // ensuing concurrent GC cycle. We will collect the perm gen and 3153 // unload classes if it's the case that: 3154 // (1) an explicit gc request has been made and the flag 3155 // ExplicitGCInvokesConcurrentAndUnloadsClasses is set, OR 3156 // (2) (a) class unloading is enabled at the command line, and 3157 // (b) (i) perm gen threshold has been crossed, or 3158 // (ii) old gen is getting really full, or 3159 // (iii) the previous N CMS collections did not collect the 3160 // perm gen 3161 // NOTE: Provided there is no change in the state of the heap between 3162 // calls to this method, it should have idempotent results. Moreover, 3163 // its results should be monotonically increasing (i.e. going from 0 to 1, 3164 // but not 1 to 0) between successive calls between which the heap was 3165 // not collected. For the implementation below, it must thus rely on 3166 // the property that concurrent_cycles_since_last_unload() 3167 // will not decrease unless a collection cycle happened and that 3168 // _permGen->should_concurrent_collect() and _cmsGen->is_too_full() are 3169 // themselves also monotonic in that sense. See check_monotonicity() 3170 // below. 3171 bool CMSCollector::update_should_unload_classes() { 3172 _should_unload_classes = false; 3173 // Condition 1 above 3174 if (_full_gc_requested && ExplicitGCInvokesConcurrentAndUnloadsClasses) { 3175 _should_unload_classes = true; 3176 } else if (CMSClassUnloadingEnabled) { // Condition 2.a above 3177 // Disjuncts 2.b.(i,ii,iii) above 3178 _should_unload_classes = (concurrent_cycles_since_last_unload() >= 3179 CMSClassUnloadingMaxInterval) 3180 || _permGen->should_concurrent_collect() 3181 || _cmsGen->is_too_full(); 3182 } 3183 return _should_unload_classes; 3184 } 3185 3186 bool ConcurrentMarkSweepGeneration::is_too_full() const { 3187 bool res = should_concurrent_collect(); 3188 res = res && (occupancy() > (double)CMSIsTooFullPercentage/100.0); 3189 return res; 3190 } 3191 3192 void CMSCollector::setup_cms_unloading_and_verification_state() { 3193 const bool should_verify = VerifyBeforeGC || VerifyAfterGC || VerifyDuringGC 3194 || VerifyBeforeExit; 3195 const int rso = SharedHeap::SO_Strings | SharedHeap::SO_CodeCache; 3196 3197 if (should_unload_classes()) { // Should unload classes this cycle 3198 remove_root_scanning_option(rso); // Shrink the root set appropriately 3199 set_verifying(should_verify); // Set verification state for this cycle 3200 return; // Nothing else needs to be done at this time 3201 } 3202 3203 // Not unloading classes this cycle 3204 assert(!should_unload_classes(), "Inconsitency!"); 3205 if ((!verifying() || unloaded_classes_last_cycle()) && should_verify) { 3206 // We were not verifying, or we _were_ unloading classes in the last cycle, 3207 // AND some verification options are enabled this cycle; in this case, 3208 // we must make sure that the deadness map is allocated if not already so, 3209 // and cleared (if already allocated previously -- 3210 // CMSBitMap::sizeInBits() is used to determine if it's allocated). 3211 if (perm_gen_verify_bit_map()->sizeInBits() == 0) { 3212 if (!perm_gen_verify_bit_map()->allocate(_permGen->reserved())) { 3213 warning("Failed to allocate permanent generation verification CMS Bit Map;\n" 3214 "permanent generation verification disabled"); 3215 return; // Note that we leave verification disabled, so we'll retry this 3216 // allocation next cycle. We _could_ remember this failure 3217 // and skip further attempts and permanently disable verification 3218 // attempts if that is considered more desirable. 3219 } 3220 assert(perm_gen_verify_bit_map()->covers(_permGen->reserved()), 3221 "_perm_gen_ver_bit_map inconsistency?"); 3222 } else { 3223 perm_gen_verify_bit_map()->clear_all(); 3224 } 3225 // Include symbols, strings and code cache elements to prevent their resurrection. 3226 add_root_scanning_option(rso); 3227 set_verifying(true); 3228 } else if (verifying() && !should_verify) { 3229 // We were verifying, but some verification flags got disabled. 3230 set_verifying(false); 3231 // Exclude symbols, strings and code cache elements from root scanning to 3232 // reduce IM and RM pauses. 3233 remove_root_scanning_option(rso); 3234 } 3235 } 3236 3237 3238 #ifndef PRODUCT 3239 HeapWord* CMSCollector::block_start(const void* p) const { 3240 const HeapWord* addr = (HeapWord*)p; 3241 if (_span.contains(p)) { 3242 if (_cmsGen->cmsSpace()->is_in_reserved(addr)) { 3243 return _cmsGen->cmsSpace()->block_start(p); 3244 } else { 3245 assert(_permGen->cmsSpace()->is_in_reserved(addr), 3246 "Inconsistent _span?"); 3247 return _permGen->cmsSpace()->block_start(p); 3248 } 3249 } 3250 return NULL; 3251 } 3252 #endif 3253 3254 HeapWord* 3255 ConcurrentMarkSweepGeneration::expand_and_allocate(size_t word_size, 3256 bool tlab, 3257 bool parallel) { 3258 CMSSynchronousYieldRequest yr; 3259 assert(!tlab, "Can't deal with TLAB allocation"); 3260 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 3261 expand(word_size*HeapWordSize, MinHeapDeltaBytes, 3262 CMSExpansionCause::_satisfy_allocation); 3263 if (GCExpandToAllocateDelayMillis > 0) { 3264 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 3265 } 3266 return have_lock_and_allocate(word_size, tlab); 3267 } 3268 3269 // YSR: All of this generation expansion/shrinking stuff is an exact copy of 3270 // OneContigSpaceCardGeneration, which makes me wonder if we should move this 3271 // to CardGeneration and share it... 3272 bool ConcurrentMarkSweepGeneration::expand(size_t bytes, size_t expand_bytes) { 3273 return CardGeneration::expand(bytes, expand_bytes); 3274 } 3275 3276 void ConcurrentMarkSweepGeneration::expand(size_t bytes, size_t expand_bytes, 3277 CMSExpansionCause::Cause cause) 3278 { 3279 3280 bool success = expand(bytes, expand_bytes); 3281 3282 // remember why we expanded; this information is used 3283 // by shouldConcurrentCollect() when making decisions on whether to start 3284 // a new CMS cycle. 3285 if (success) { 3286 set_expansion_cause(cause); 3287 if (PrintGCDetails && Verbose) { 3288 gclog_or_tty->print_cr("Expanded CMS gen for %s", 3289 CMSExpansionCause::to_string(cause)); 3290 } 3291 } 3292 } 3293 3294 HeapWord* ConcurrentMarkSweepGeneration::expand_and_par_lab_allocate(CMSParGCThreadState* ps, size_t word_sz) { 3295 HeapWord* res = NULL; 3296 MutexLocker x(ParGCRareEvent_lock); 3297 while (true) { 3298 // Expansion by some other thread might make alloc OK now: 3299 res = ps->lab.alloc(word_sz); 3300 if (res != NULL) return res; 3301 // If there's not enough expansion space available, give up. 3302 if (_virtual_space.uncommitted_size() < (word_sz * HeapWordSize)) { 3303 return NULL; 3304 } 3305 // Otherwise, we try expansion. 3306 expand(word_sz*HeapWordSize, MinHeapDeltaBytes, 3307 CMSExpansionCause::_allocate_par_lab); 3308 // Now go around the loop and try alloc again; 3309 // A competing par_promote might beat us to the expansion space, 3310 // so we may go around the loop again if promotion fails agaion. 3311 if (GCExpandToAllocateDelayMillis > 0) { 3312 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 3313 } 3314 } 3315 } 3316 3317 3318 bool ConcurrentMarkSweepGeneration::expand_and_ensure_spooling_space( 3319 PromotionInfo* promo) { 3320 MutexLocker x(ParGCRareEvent_lock); 3321 size_t refill_size_bytes = promo->refillSize() * HeapWordSize; 3322 while (true) { 3323 // Expansion by some other thread might make alloc OK now: 3324 if (promo->ensure_spooling_space()) { 3325 assert(promo->has_spooling_space(), 3326 "Post-condition of successful ensure_spooling_space()"); 3327 return true; 3328 } 3329 // If there's not enough expansion space available, give up. 3330 if (_virtual_space.uncommitted_size() < refill_size_bytes) { 3331 return false; 3332 } 3333 // Otherwise, we try expansion. 3334 expand(refill_size_bytes, MinHeapDeltaBytes, 3335 CMSExpansionCause::_allocate_par_spooling_space); 3336 // Now go around the loop and try alloc again; 3337 // A competing allocation might beat us to the expansion space, 3338 // so we may go around the loop again if allocation fails again. 3339 if (GCExpandToAllocateDelayMillis > 0) { 3340 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 3341 } 3342 } 3343 } 3344 3345 3346 3347 void ConcurrentMarkSweepGeneration::shrink(size_t bytes) { 3348 assert_locked_or_safepoint(Heap_lock); 3349 size_t size = ReservedSpace::page_align_size_down(bytes); 3350 if (size > 0) { 3351 shrink_by(size); 3352 } 3353 } 3354 3355 bool ConcurrentMarkSweepGeneration::grow_by(size_t bytes) { 3356 assert_locked_or_safepoint(Heap_lock); 3357 bool result = _virtual_space.expand_by(bytes); 3358 if (result) { 3359 HeapWord* old_end = _cmsSpace->end(); 3360 size_t new_word_size = 3361 heap_word_size(_virtual_space.committed_size()); 3362 MemRegion mr(_cmsSpace->bottom(), new_word_size); 3363 _bts->resize(new_word_size); // resize the block offset shared array 3364 Universe::heap()->barrier_set()->resize_covered_region(mr); 3365 // Hmmmm... why doesn't CFLS::set_end verify locking? 3366 // This is quite ugly; FIX ME XXX 3367 _cmsSpace->assert_locked(freelistLock()); 3368 _cmsSpace->set_end((HeapWord*)_virtual_space.high()); 3369 3370 // update the space and generation capacity counters 3371 if (UsePerfData) { 3372 _space_counters->update_capacity(); 3373 _gen_counters->update_all(); 3374 } 3375 3376 if (Verbose && PrintGC) { 3377 size_t new_mem_size = _virtual_space.committed_size(); 3378 size_t old_mem_size = new_mem_size - bytes; 3379 gclog_or_tty->print_cr("Expanding %s from %ldK by %ldK to %ldK", 3380 name(), old_mem_size/K, bytes/K, new_mem_size/K); 3381 } 3382 } 3383 return result; 3384 } 3385 3386 bool ConcurrentMarkSweepGeneration::grow_to_reserved() { 3387 assert_locked_or_safepoint(Heap_lock); 3388 bool success = true; 3389 const size_t remaining_bytes = _virtual_space.uncommitted_size(); 3390 if (remaining_bytes > 0) { 3391 success = grow_by(remaining_bytes); 3392 DEBUG_ONLY(if (!success) warning("grow to reserved failed");) 3393 } 3394 return success; 3395 } 3396 3397 void ConcurrentMarkSweepGeneration::shrink_by(size_t bytes) { 3398 assert_locked_or_safepoint(Heap_lock); 3399 assert_lock_strong(freelistLock()); 3400 // XXX Fix when compaction is implemented. 3401 warning("Shrinking of CMS not yet implemented"); 3402 return; 3403 } 3404 3405 3406 // Simple ctor/dtor wrapper for accounting & timer chores around concurrent 3407 // phases. 3408 class CMSPhaseAccounting: public StackObj { 3409 public: 3410 CMSPhaseAccounting(CMSCollector *collector, 3411 const char *phase, 3412 bool print_cr = true); 3413 ~CMSPhaseAccounting(); 3414 3415 private: 3416 CMSCollector *_collector; 3417 const char *_phase; 3418 elapsedTimer _wallclock; 3419 bool _print_cr; 3420 3421 public: 3422 // Not MT-safe; so do not pass around these StackObj's 3423 // where they may be accessed by other threads. 3424 jlong wallclock_millis() { 3425 assert(_wallclock.is_active(), "Wall clock should not stop"); 3426 _wallclock.stop(); // to record time 3427 jlong ret = _wallclock.milliseconds(); 3428 _wallclock.start(); // restart 3429 return ret; 3430 } 3431 }; 3432 3433 CMSPhaseAccounting::CMSPhaseAccounting(CMSCollector *collector, 3434 const char *phase, 3435 bool print_cr) : 3436 _collector(collector), _phase(phase), _print_cr(print_cr) { 3437 3438 if (PrintCMSStatistics != 0) { 3439 _collector->resetYields(); 3440 } 3441 if (PrintGCDetails && PrintGCTimeStamps) { 3442 gclog_or_tty->date_stamp(PrintGCDateStamps); 3443 gclog_or_tty->stamp(); 3444 gclog_or_tty->print_cr(": [%s-concurrent-%s-start]", 3445 _collector->cmsGen()->short_name(), _phase); 3446 } 3447 _collector->resetTimer(); 3448 _wallclock.start(); 3449 _collector->startTimer(); 3450 } 3451 3452 CMSPhaseAccounting::~CMSPhaseAccounting() { 3453 assert(_wallclock.is_active(), "Wall clock should not have stopped"); 3454 _collector->stopTimer(); 3455 _wallclock.stop(); 3456 if (PrintGCDetails) { 3457 gclog_or_tty->date_stamp(PrintGCDateStamps); 3458 if (PrintGCTimeStamps) { 3459 gclog_or_tty->stamp(); 3460 gclog_or_tty->print(": "); 3461 } 3462 gclog_or_tty->print("[%s-concurrent-%s: %3.3f/%3.3f secs]", 3463 _collector->cmsGen()->short_name(), 3464 _phase, _collector->timerValue(), _wallclock.seconds()); 3465 if (_print_cr) { 3466 gclog_or_tty->print_cr(""); 3467 } 3468 if (PrintCMSStatistics != 0) { 3469 gclog_or_tty->print_cr(" (CMS-concurrent-%s yielded %d times)", _phase, 3470 _collector->yields()); 3471 } 3472 } 3473 } 3474 3475 // CMS work 3476 3477 // Checkpoint the roots into this generation from outside 3478 // this generation. [Note this initial checkpoint need only 3479 // be approximate -- we'll do a catch up phase subsequently.] 3480 void CMSCollector::checkpointRootsInitial(bool asynch) { 3481 assert(_collectorState == InitialMarking, "Wrong collector state"); 3482 check_correct_thread_executing(); 3483 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 3484 3485 ReferenceProcessor* rp = ref_processor(); 3486 SpecializationStats::clear(); 3487 assert(_restart_addr == NULL, "Control point invariant"); 3488 if (asynch) { 3489 // acquire locks for subsequent manipulations 3490 MutexLockerEx x(bitMapLock(), 3491 Mutex::_no_safepoint_check_flag); 3492 checkpointRootsInitialWork(asynch); 3493 // enable ("weak") refs discovery 3494 rp->enable_discovery(true /*verify_disabled*/, true /*check_no_refs*/); 3495 _collectorState = Marking; 3496 } else { 3497 // (Weak) Refs discovery: this is controlled from genCollectedHeap::do_collection 3498 // which recognizes if we are a CMS generation, and doesn't try to turn on 3499 // discovery; verify that they aren't meddling. 3500 assert(!rp->discovery_is_atomic(), 3501 "incorrect setting of discovery predicate"); 3502 assert(!rp->discovery_enabled(), "genCollectedHeap shouldn't control " 3503 "ref discovery for this generation kind"); 3504 // already have locks 3505 checkpointRootsInitialWork(asynch); 3506 // now enable ("weak") refs discovery 3507 rp->enable_discovery(true /*verify_disabled*/, false /*verify_no_refs*/); 3508 _collectorState = Marking; 3509 } 3510 SpecializationStats::print(); 3511 } 3512 3513 void CMSCollector::checkpointRootsInitialWork(bool asynch) { 3514 assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped"); 3515 assert(_collectorState == InitialMarking, "just checking"); 3516 3517 // If there has not been a GC[n-1] since last GC[n] cycle completed, 3518 // precede our marking with a collection of all 3519 // younger generations to keep floating garbage to a minimum. 3520 // XXX: we won't do this for now -- it's an optimization to be done later. 3521 3522 // already have locks 3523 assert_lock_strong(bitMapLock()); 3524 assert(_markBitMap.isAllClear(), "was reset at end of previous cycle"); 3525 3526 // Setup the verification and class unloading state for this 3527 // CMS collection cycle. 3528 setup_cms_unloading_and_verification_state(); 3529 3530 NOT_PRODUCT(TraceTime t("\ncheckpointRootsInitialWork", 3531 PrintGCDetails && Verbose, true, gclog_or_tty);) 3532 if (UseAdaptiveSizePolicy) { 3533 size_policy()->checkpoint_roots_initial_begin(); 3534 } 3535 3536 // Reset all the PLAB chunk arrays if necessary. 3537 if (_survivor_plab_array != NULL && !CMSPLABRecordAlways) { 3538 reset_survivor_plab_arrays(); 3539 } 3540 3541 ResourceMark rm; 3542 HandleMark hm; 3543 3544 FalseClosure falseClosure; 3545 // In the case of a synchronous collection, we will elide the 3546 // remark step, so it's important to catch all the nmethod oops 3547 // in this step. 3548 // The final 'true' flag to gen_process_strong_roots will ensure this. 3549 // If 'async' is true, we can relax the nmethod tracing. 3550 MarkRefsIntoClosure notOlder(_span, &_markBitMap); 3551 GenCollectedHeap* gch = GenCollectedHeap::heap(); 3552 3553 verify_work_stacks_empty(); 3554 verify_overflow_empty(); 3555 3556 gch->ensure_parsability(false); // fill TLABs, but no need to retire them 3557 // Update the saved marks which may affect the root scans. 3558 gch->save_marks(); 3559 3560 // weak reference processing has not started yet. 3561 ref_processor()->set_enqueuing_is_done(false); 3562 3563 { 3564 // This is not needed. DEBUG_ONLY(RememberKlassesChecker imx(true);) 3565 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;) 3566 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 3567 gch->gen_process_strong_roots(_cmsGen->level(), 3568 true, // younger gens are roots 3569 true, // activate StrongRootsScope 3570 true, // collecting perm gen 3571 SharedHeap::ScanningOption(roots_scanning_options()), 3572 ¬Older, 3573 true, // walk all of code cache if (so & SO_CodeCache) 3574 NULL); 3575 } 3576 3577 // Clear mod-union table; it will be dirtied in the prologue of 3578 // CMS generation per each younger generation collection. 3579 3580 assert(_modUnionTable.isAllClear(), 3581 "Was cleared in most recent final checkpoint phase" 3582 " or no bits are set in the gc_prologue before the start of the next " 3583 "subsequent marking phase."); 3584 3585 // Save the end of the used_region of the constituent generations 3586 // to be used to limit the extent of sweep in each generation. 3587 save_sweep_limits(); 3588 if (UseAdaptiveSizePolicy) { 3589 size_policy()->checkpoint_roots_initial_end(gch->gc_cause()); 3590 } 3591 verify_overflow_empty(); 3592 } 3593 3594 bool CMSCollector::markFromRoots(bool asynch) { 3595 // we might be tempted to assert that: 3596 // assert(asynch == !SafepointSynchronize::is_at_safepoint(), 3597 // "inconsistent argument?"); 3598 // However that wouldn't be right, because it's possible that 3599 // a safepoint is indeed in progress as a younger generation 3600 // stop-the-world GC happens even as we mark in this generation. 3601 assert(_collectorState == Marking, "inconsistent state?"); 3602 check_correct_thread_executing(); 3603 verify_overflow_empty(); 3604 3605 bool res; 3606 if (asynch) { 3607 3608 // Start the timers for adaptive size policy for the concurrent phases 3609 // Do it here so that the foreground MS can use the concurrent 3610 // timer since a foreground MS might has the sweep done concurrently 3611 // or STW. 3612 if (UseAdaptiveSizePolicy) { 3613 size_policy()->concurrent_marking_begin(); 3614 } 3615 3616 // Weak ref discovery note: We may be discovering weak 3617 // refs in this generation concurrent (but interleaved) with 3618 // weak ref discovery by a younger generation collector. 3619 3620 CMSTokenSyncWithLocks ts(true, bitMapLock()); 3621 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 3622 CMSPhaseAccounting pa(this, "mark", !PrintGCDetails); 3623 res = markFromRootsWork(asynch); 3624 if (res) { 3625 _collectorState = Precleaning; 3626 } else { // We failed and a foreground collection wants to take over 3627 assert(_foregroundGCIsActive, "internal state inconsistency"); 3628 assert(_restart_addr == NULL, "foreground will restart from scratch"); 3629 if (PrintGCDetails) { 3630 gclog_or_tty->print_cr("bailing out to foreground collection"); 3631 } 3632 } 3633 if (UseAdaptiveSizePolicy) { 3634 size_policy()->concurrent_marking_end(); 3635 } 3636 } else { 3637 assert(SafepointSynchronize::is_at_safepoint(), 3638 "inconsistent with asynch == false"); 3639 if (UseAdaptiveSizePolicy) { 3640 size_policy()->ms_collection_marking_begin(); 3641 } 3642 // already have locks 3643 res = markFromRootsWork(asynch); 3644 _collectorState = FinalMarking; 3645 if (UseAdaptiveSizePolicy) { 3646 GenCollectedHeap* gch = GenCollectedHeap::heap(); 3647 size_policy()->ms_collection_marking_end(gch->gc_cause()); 3648 } 3649 } 3650 verify_overflow_empty(); 3651 return res; 3652 } 3653 3654 bool CMSCollector::markFromRootsWork(bool asynch) { 3655 // iterate over marked bits in bit map, doing a full scan and mark 3656 // from these roots using the following algorithm: 3657 // . if oop is to the right of the current scan pointer, 3658 // mark corresponding bit (we'll process it later) 3659 // . else (oop is to left of current scan pointer) 3660 // push oop on marking stack 3661 // . drain the marking stack 3662 3663 // Note that when we do a marking step we need to hold the 3664 // bit map lock -- recall that direct allocation (by mutators) 3665 // and promotion (by younger generation collectors) is also 3666 // marking the bit map. [the so-called allocate live policy.] 3667 // Because the implementation of bit map marking is not 3668 // robust wrt simultaneous marking of bits in the same word, 3669 // we need to make sure that there is no such interference 3670 // between concurrent such updates. 3671 3672 // already have locks 3673 assert_lock_strong(bitMapLock()); 3674 3675 // Clear the revisit stack, just in case there are any 3676 // obsolete contents from a short-circuited previous CMS cycle. 3677 _revisitStack.reset(); 3678 verify_work_stacks_empty(); 3679 verify_overflow_empty(); 3680 assert(_revisitStack.isEmpty(), "tabula rasa"); 3681 DEBUG_ONLY(RememberKlassesChecker cmx(should_unload_classes());) 3682 bool result = false; 3683 if (CMSConcurrentMTEnabled && ConcGCThreads > 0) { 3684 result = do_marking_mt(asynch); 3685 } else { 3686 result = do_marking_st(asynch); 3687 } 3688 return result; 3689 } 3690 3691 // Forward decl 3692 class CMSConcMarkingTask; 3693 3694 class CMSConcMarkingTerminator: public ParallelTaskTerminator { 3695 CMSCollector* _collector; 3696 CMSConcMarkingTask* _task; 3697 public: 3698 virtual void yield(); 3699 3700 // "n_threads" is the number of threads to be terminated. 3701 // "queue_set" is a set of work queues of other threads. 3702 // "collector" is the CMS collector associated with this task terminator. 3703 // "yield" indicates whether we need the gang as a whole to yield. 3704 CMSConcMarkingTerminator(int n_threads, TaskQueueSetSuper* queue_set, CMSCollector* collector) : 3705 ParallelTaskTerminator(n_threads, queue_set), 3706 _collector(collector) { } 3707 3708 void set_task(CMSConcMarkingTask* task) { 3709 _task = task; 3710 } 3711 }; 3712 3713 class CMSConcMarkingTerminatorTerminator: public TerminatorTerminator { 3714 CMSConcMarkingTask* _task; 3715 public: 3716 bool should_exit_termination(); 3717 void set_task(CMSConcMarkingTask* task) { 3718 _task = task; 3719 } 3720 }; 3721 3722 // MT Concurrent Marking Task 3723 class CMSConcMarkingTask: public YieldingFlexibleGangTask { 3724 CMSCollector* _collector; 3725 int _n_workers; // requested/desired # workers 3726 bool _asynch; 3727 bool _result; 3728 CompactibleFreeListSpace* _cms_space; 3729 CompactibleFreeListSpace* _perm_space; 3730 char _pad_front[64]; // padding to ... 3731 HeapWord* _global_finger; // ... avoid sharing cache line 3732 char _pad_back[64]; 3733 HeapWord* _restart_addr; 3734 3735 // Exposed here for yielding support 3736 Mutex* const _bit_map_lock; 3737 3738 // The per thread work queues, available here for stealing 3739 OopTaskQueueSet* _task_queues; 3740 3741 // Termination (and yielding) support 3742 CMSConcMarkingTerminator _term; 3743 CMSConcMarkingTerminatorTerminator _term_term; 3744 3745 public: 3746 CMSConcMarkingTask(CMSCollector* collector, 3747 CompactibleFreeListSpace* cms_space, 3748 CompactibleFreeListSpace* perm_space, 3749 bool asynch, 3750 YieldingFlexibleWorkGang* workers, 3751 OopTaskQueueSet* task_queues): 3752 YieldingFlexibleGangTask("Concurrent marking done multi-threaded"), 3753 _collector(collector), 3754 _cms_space(cms_space), 3755 _perm_space(perm_space), 3756 _asynch(asynch), _n_workers(0), _result(true), 3757 _task_queues(task_queues), 3758 _term(_n_workers, task_queues, _collector), 3759 _bit_map_lock(collector->bitMapLock()) 3760 { 3761 _requested_size = _n_workers; 3762 _term.set_task(this); 3763 _term_term.set_task(this); 3764 assert(_cms_space->bottom() < _perm_space->bottom(), 3765 "Finger incorrectly initialized below"); 3766 _restart_addr = _global_finger = _cms_space->bottom(); 3767 } 3768 3769 3770 OopTaskQueueSet* task_queues() { return _task_queues; } 3771 3772 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 3773 3774 HeapWord** global_finger_addr() { return &_global_finger; } 3775 3776 CMSConcMarkingTerminator* terminator() { return &_term; } 3777 3778 virtual void set_for_termination(int active_workers) { 3779 terminator()->reset_for_reuse(active_workers); 3780 } 3781 3782 void work(uint worker_id); 3783 bool should_yield() { 3784 return ConcurrentMarkSweepThread::should_yield() 3785 && !_collector->foregroundGCIsActive() 3786 && _asynch; 3787 } 3788 3789 virtual void coordinator_yield(); // stuff done by coordinator 3790 bool result() { return _result; } 3791 3792 void reset(HeapWord* ra) { 3793 assert(_global_finger >= _cms_space->end(), "Postcondition of ::work(i)"); 3794 assert(_global_finger >= _perm_space->end(), "Postcondition of ::work(i)"); 3795 assert(ra < _perm_space->end(), "ra too large"); 3796 _restart_addr = _global_finger = ra; 3797 _term.reset_for_reuse(); 3798 } 3799 3800 static bool get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 3801 OopTaskQueue* work_q); 3802 3803 private: 3804 void do_scan_and_mark(int i, CompactibleFreeListSpace* sp); 3805 void do_work_steal(int i); 3806 void bump_global_finger(HeapWord* f); 3807 }; 3808 3809 bool CMSConcMarkingTerminatorTerminator::should_exit_termination() { 3810 assert(_task != NULL, "Error"); 3811 return _task->yielding(); 3812 // Note that we do not need the disjunct || _task->should_yield() above 3813 // because we want terminating threads to yield only if the task 3814 // is already in the midst of yielding, which happens only after at least one 3815 // thread has yielded. 3816 } 3817 3818 void CMSConcMarkingTerminator::yield() { 3819 if (_task->should_yield()) { 3820 _task->yield(); 3821 } else { 3822 ParallelTaskTerminator::yield(); 3823 } 3824 } 3825 3826 //////////////////////////////////////////////////////////////// 3827 // Concurrent Marking Algorithm Sketch 3828 //////////////////////////////////////////////////////////////// 3829 // Until all tasks exhausted (both spaces): 3830 // -- claim next available chunk 3831 // -- bump global finger via CAS 3832 // -- find first object that starts in this chunk 3833 // and start scanning bitmap from that position 3834 // -- scan marked objects for oops 3835 // -- CAS-mark target, and if successful: 3836 // . if target oop is above global finger (volatile read) 3837 // nothing to do 3838 // . if target oop is in chunk and above local finger 3839 // then nothing to do 3840 // . else push on work-queue 3841 // -- Deal with possible overflow issues: 3842 // . local work-queue overflow causes stuff to be pushed on 3843 // global (common) overflow queue 3844 // . always first empty local work queue 3845 // . then get a batch of oops from global work queue if any 3846 // . then do work stealing 3847 // -- When all tasks claimed (both spaces) 3848 // and local work queue empty, 3849 // then in a loop do: 3850 // . check global overflow stack; steal a batch of oops and trace 3851 // . try to steal from other threads oif GOS is empty 3852 // . if neither is available, offer termination 3853 // -- Terminate and return result 3854 // 3855 void CMSConcMarkingTask::work(uint worker_id) { 3856 elapsedTimer _timer; 3857 ResourceMark rm; 3858 HandleMark hm; 3859 3860 DEBUG_ONLY(_collector->verify_overflow_empty();) 3861 3862 // Before we begin work, our work queue should be empty 3863 assert(work_queue(worker_id)->size() == 0, "Expected to be empty"); 3864 // Scan the bitmap covering _cms_space, tracing through grey objects. 3865 _timer.start(); 3866 do_scan_and_mark(worker_id, _cms_space); 3867 _timer.stop(); 3868 if (PrintCMSStatistics != 0) { 3869 gclog_or_tty->print_cr("Finished cms space scanning in %dth thread: %3.3f sec", 3870 worker_id, _timer.seconds()); 3871 // XXX: need xxx/xxx type of notation, two timers 3872 } 3873 3874 // ... do the same for the _perm_space 3875 _timer.reset(); 3876 _timer.start(); 3877 do_scan_and_mark(worker_id, _perm_space); 3878 _timer.stop(); 3879 if (PrintCMSStatistics != 0) { 3880 gclog_or_tty->print_cr("Finished perm space scanning in %dth thread: %3.3f sec", 3881 worker_id, _timer.seconds()); 3882 // XXX: need xxx/xxx type of notation, two timers 3883 } 3884 3885 // ... do work stealing 3886 _timer.reset(); 3887 _timer.start(); 3888 do_work_steal(worker_id); 3889 _timer.stop(); 3890 if (PrintCMSStatistics != 0) { 3891 gclog_or_tty->print_cr("Finished work stealing in %dth thread: %3.3f sec", 3892 worker_id, _timer.seconds()); 3893 // XXX: need xxx/xxx type of notation, two timers 3894 } 3895 assert(_collector->_markStack.isEmpty(), "Should have been emptied"); 3896 assert(work_queue(worker_id)->size() == 0, "Should have been emptied"); 3897 // Note that under the current task protocol, the 3898 // following assertion is true even of the spaces 3899 // expanded since the completion of the concurrent 3900 // marking. XXX This will likely change under a strict 3901 // ABORT semantics. 3902 assert(_global_finger > _cms_space->end() && 3903 _global_finger >= _perm_space->end(), 3904 "All tasks have been completed"); 3905 DEBUG_ONLY(_collector->verify_overflow_empty();) 3906 } 3907 3908 void CMSConcMarkingTask::bump_global_finger(HeapWord* f) { 3909 HeapWord* read = _global_finger; 3910 HeapWord* cur = read; 3911 while (f > read) { 3912 cur = read; 3913 read = (HeapWord*) Atomic::cmpxchg_ptr(f, &_global_finger, cur); 3914 if (cur == read) { 3915 // our cas succeeded 3916 assert(_global_finger >= f, "protocol consistency"); 3917 break; 3918 } 3919 } 3920 } 3921 3922 // This is really inefficient, and should be redone by 3923 // using (not yet available) block-read and -write interfaces to the 3924 // stack and the work_queue. XXX FIX ME !!! 3925 bool CMSConcMarkingTask::get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 3926 OopTaskQueue* work_q) { 3927 // Fast lock-free check 3928 if (ovflw_stk->length() == 0) { 3929 return false; 3930 } 3931 assert(work_q->size() == 0, "Shouldn't steal"); 3932 MutexLockerEx ml(ovflw_stk->par_lock(), 3933 Mutex::_no_safepoint_check_flag); 3934 // Grab up to 1/4 the size of the work queue 3935 size_t num = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 3936 (size_t)ParGCDesiredObjsFromOverflowList); 3937 num = MIN2(num, ovflw_stk->length()); 3938 for (int i = (int) num; i > 0; i--) { 3939 oop cur = ovflw_stk->pop(); 3940 assert(cur != NULL, "Counted wrong?"); 3941 work_q->push(cur); 3942 } 3943 return num > 0; 3944 } 3945 3946 void CMSConcMarkingTask::do_scan_and_mark(int i, CompactibleFreeListSpace* sp) { 3947 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 3948 int n_tasks = pst->n_tasks(); 3949 // We allow that there may be no tasks to do here because 3950 // we are restarting after a stack overflow. 3951 assert(pst->valid() || n_tasks == 0, "Uninitialized use?"); 3952 uint nth_task = 0; 3953 3954 HeapWord* aligned_start = sp->bottom(); 3955 if (sp->used_region().contains(_restart_addr)) { 3956 // Align down to a card boundary for the start of 0th task 3957 // for this space. 3958 aligned_start = 3959 (HeapWord*)align_size_down((uintptr_t)_restart_addr, 3960 CardTableModRefBS::card_size); 3961 } 3962 3963 size_t chunk_size = sp->marking_task_size(); 3964 while (!pst->is_task_claimed(/* reference */ nth_task)) { 3965 // Having claimed the nth task in this space, 3966 // compute the chunk that it corresponds to: 3967 MemRegion span = MemRegion(aligned_start + nth_task*chunk_size, 3968 aligned_start + (nth_task+1)*chunk_size); 3969 // Try and bump the global finger via a CAS; 3970 // note that we need to do the global finger bump 3971 // _before_ taking the intersection below, because 3972 // the task corresponding to that region will be 3973 // deemed done even if the used_region() expands 3974 // because of allocation -- as it almost certainly will 3975 // during start-up while the threads yield in the 3976 // closure below. 3977 HeapWord* finger = span.end(); 3978 bump_global_finger(finger); // atomically 3979 // There are null tasks here corresponding to chunks 3980 // beyond the "top" address of the space. 3981 span = span.intersection(sp->used_region()); 3982 if (!span.is_empty()) { // Non-null task 3983 HeapWord* prev_obj; 3984 assert(!span.contains(_restart_addr) || nth_task == 0, 3985 "Inconsistency"); 3986 if (nth_task == 0) { 3987 // For the 0th task, we'll not need to compute a block_start. 3988 if (span.contains(_restart_addr)) { 3989 // In the case of a restart because of stack overflow, 3990 // we might additionally skip a chunk prefix. 3991 prev_obj = _restart_addr; 3992 } else { 3993 prev_obj = span.start(); 3994 } 3995 } else { 3996 // We want to skip the first object because 3997 // the protocol is to scan any object in its entirety 3998 // that _starts_ in this span; a fortiori, any 3999 // object starting in an earlier span is scanned 4000 // as part of an earlier claimed task. 4001 // Below we use the "careful" version of block_start 4002 // so we do not try to navigate uninitialized objects. 4003 prev_obj = sp->block_start_careful(span.start()); 4004 // Below we use a variant of block_size that uses the 4005 // Printezis bits to avoid waiting for allocated 4006 // objects to become initialized/parsable. 4007 while (prev_obj < span.start()) { 4008 size_t sz = sp->block_size_no_stall(prev_obj, _collector); 4009 if (sz > 0) { 4010 prev_obj += sz; 4011 } else { 4012 // In this case we may end up doing a bit of redundant 4013 // scanning, but that appears unavoidable, short of 4014 // locking the free list locks; see bug 6324141. 4015 break; 4016 } 4017 } 4018 } 4019 if (prev_obj < span.end()) { 4020 MemRegion my_span = MemRegion(prev_obj, span.end()); 4021 // Do the marking work within a non-empty span -- 4022 // the last argument to the constructor indicates whether the 4023 // iteration should be incremental with periodic yields. 4024 Par_MarkFromRootsClosure cl(this, _collector, my_span, 4025 &_collector->_markBitMap, 4026 work_queue(i), 4027 &_collector->_markStack, 4028 &_collector->_revisitStack, 4029 _asynch); 4030 _collector->_markBitMap.iterate(&cl, my_span.start(), my_span.end()); 4031 } // else nothing to do for this task 4032 } // else nothing to do for this task 4033 } 4034 // We'd be tempted to assert here that since there are no 4035 // more tasks left to claim in this space, the global_finger 4036 // must exceed space->top() and a fortiori space->end(). However, 4037 // that would not quite be correct because the bumping of 4038 // global_finger occurs strictly after the claiming of a task, 4039 // so by the time we reach here the global finger may not yet 4040 // have been bumped up by the thread that claimed the last 4041 // task. 4042 pst->all_tasks_completed(); 4043 } 4044 4045 class Par_ConcMarkingClosure: public Par_KlassRememberingOopClosure { 4046 private: 4047 CMSConcMarkingTask* _task; 4048 MemRegion _span; 4049 CMSBitMap* _bit_map; 4050 CMSMarkStack* _overflow_stack; 4051 OopTaskQueue* _work_queue; 4052 protected: 4053 DO_OOP_WORK_DEFN 4054 public: 4055 Par_ConcMarkingClosure(CMSCollector* collector, CMSConcMarkingTask* task, OopTaskQueue* work_queue, 4056 CMSBitMap* bit_map, CMSMarkStack* overflow_stack, 4057 CMSMarkStack* revisit_stack): 4058 Par_KlassRememberingOopClosure(collector, collector->ref_processor(), revisit_stack), 4059 _task(task), 4060 _span(collector->_span), 4061 _work_queue(work_queue), 4062 _bit_map(bit_map), 4063 _overflow_stack(overflow_stack) 4064 { } 4065 virtual void do_oop(oop* p); 4066 virtual void do_oop(narrowOop* p); 4067 void trim_queue(size_t max); 4068 void handle_stack_overflow(HeapWord* lost); 4069 void do_yield_check() { 4070 if (_task->should_yield()) { 4071 _task->yield(); 4072 } 4073 } 4074 }; 4075 4076 // Grey object scanning during work stealing phase -- 4077 // the salient assumption here is that any references 4078 // that are in these stolen objects being scanned must 4079 // already have been initialized (else they would not have 4080 // been published), so we do not need to check for 4081 // uninitialized objects before pushing here. 4082 void Par_ConcMarkingClosure::do_oop(oop obj) { 4083 assert(obj->is_oop_or_null(true), "expected an oop or NULL"); 4084 HeapWord* addr = (HeapWord*)obj; 4085 // Check if oop points into the CMS generation 4086 // and is not marked 4087 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 4088 // a white object ... 4089 // If we manage to "claim" the object, by being the 4090 // first thread to mark it, then we push it on our 4091 // marking stack 4092 if (_bit_map->par_mark(addr)) { // ... now grey 4093 // push on work queue (grey set) 4094 bool simulate_overflow = false; 4095 NOT_PRODUCT( 4096 if (CMSMarkStackOverflowALot && 4097 _collector->simulate_overflow()) { 4098 // simulate a stack overflow 4099 simulate_overflow = true; 4100 } 4101 ) 4102 if (simulate_overflow || 4103 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) { 4104 // stack overflow 4105 if (PrintCMSStatistics != 0) { 4106 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 4107 SIZE_FORMAT, _overflow_stack->capacity()); 4108 } 4109 // We cannot assert that the overflow stack is full because 4110 // it may have been emptied since. 4111 assert(simulate_overflow || 4112 _work_queue->size() == _work_queue->max_elems(), 4113 "Else push should have succeeded"); 4114 handle_stack_overflow(addr); 4115 } 4116 } // Else, some other thread got there first 4117 do_yield_check(); 4118 } 4119 } 4120 4121 void Par_ConcMarkingClosure::do_oop(oop* p) { Par_ConcMarkingClosure::do_oop_work(p); } 4122 void Par_ConcMarkingClosure::do_oop(narrowOop* p) { Par_ConcMarkingClosure::do_oop_work(p); } 4123 4124 void Par_ConcMarkingClosure::trim_queue(size_t max) { 4125 while (_work_queue->size() > max) { 4126 oop new_oop; 4127 if (_work_queue->pop_local(new_oop)) { 4128 assert(new_oop->is_oop(), "Should be an oop"); 4129 assert(_bit_map->isMarked((HeapWord*)new_oop), "Grey object"); 4130 assert(_span.contains((HeapWord*)new_oop), "Not in span"); 4131 assert(new_oop->is_parsable(), "Should be parsable"); 4132 new_oop->oop_iterate(this); // do_oop() above 4133 do_yield_check(); 4134 } 4135 } 4136 } 4137 4138 // Upon stack overflow, we discard (part of) the stack, 4139 // remembering the least address amongst those discarded 4140 // in CMSCollector's _restart_address. 4141 void Par_ConcMarkingClosure::handle_stack_overflow(HeapWord* lost) { 4142 // We need to do this under a mutex to prevent other 4143 // workers from interfering with the work done below. 4144 MutexLockerEx ml(_overflow_stack->par_lock(), 4145 Mutex::_no_safepoint_check_flag); 4146 // Remember the least grey address discarded 4147 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 4148 _collector->lower_restart_addr(ra); 4149 _overflow_stack->reset(); // discard stack contents 4150 _overflow_stack->expand(); // expand the stack if possible 4151 } 4152 4153 4154 void CMSConcMarkingTask::do_work_steal(int i) { 4155 OopTaskQueue* work_q = work_queue(i); 4156 oop obj_to_scan; 4157 CMSBitMap* bm = &(_collector->_markBitMap); 4158 CMSMarkStack* ovflw = &(_collector->_markStack); 4159 CMSMarkStack* revisit = &(_collector->_revisitStack); 4160 int* seed = _collector->hash_seed(i); 4161 Par_ConcMarkingClosure cl(_collector, this, work_q, bm, ovflw, revisit); 4162 while (true) { 4163 cl.trim_queue(0); 4164 assert(work_q->size() == 0, "Should have been emptied above"); 4165 if (get_work_from_overflow_stack(ovflw, work_q)) { 4166 // Can't assert below because the work obtained from the 4167 // overflow stack may already have been stolen from us. 4168 // assert(work_q->size() > 0, "Work from overflow stack"); 4169 continue; 4170 } else if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 4171 assert(obj_to_scan->is_oop(), "Should be an oop"); 4172 assert(bm->isMarked((HeapWord*)obj_to_scan), "Grey object"); 4173 obj_to_scan->oop_iterate(&cl); 4174 } else if (terminator()->offer_termination(&_term_term)) { 4175 assert(work_q->size() == 0, "Impossible!"); 4176 break; 4177 } else if (yielding() || should_yield()) { 4178 yield(); 4179 } 4180 } 4181 } 4182 4183 // This is run by the CMS (coordinator) thread. 4184 void CMSConcMarkingTask::coordinator_yield() { 4185 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 4186 "CMS thread should hold CMS token"); 4187 DEBUG_ONLY(RememberKlassesChecker mux(false);) 4188 // First give up the locks, then yield, then re-lock 4189 // We should probably use a constructor/destructor idiom to 4190 // do this unlock/lock or modify the MutexUnlocker class to 4191 // serve our purpose. XXX 4192 assert_lock_strong(_bit_map_lock); 4193 _bit_map_lock->unlock(); 4194 ConcurrentMarkSweepThread::desynchronize(true); 4195 ConcurrentMarkSweepThread::acknowledge_yield_request(); 4196 _collector->stopTimer(); 4197 if (PrintCMSStatistics != 0) { 4198 _collector->incrementYields(); 4199 } 4200 _collector->icms_wait(); 4201 4202 // It is possible for whichever thread initiated the yield request 4203 // not to get a chance to wake up and take the bitmap lock between 4204 // this thread releasing it and reacquiring it. So, while the 4205 // should_yield() flag is on, let's sleep for a bit to give the 4206 // other thread a chance to wake up. The limit imposed on the number 4207 // of iterations is defensive, to avoid any unforseen circumstances 4208 // putting us into an infinite loop. Since it's always been this 4209 // (coordinator_yield()) method that was observed to cause the 4210 // problem, we are using a parameter (CMSCoordinatorYieldSleepCount) 4211 // which is by default non-zero. For the other seven methods that 4212 // also perform the yield operation, as are using a different 4213 // parameter (CMSYieldSleepCount) which is by default zero. This way we 4214 // can enable the sleeping for those methods too, if necessary. 4215 // See 6442774. 4216 // 4217 // We really need to reconsider the synchronization between the GC 4218 // thread and the yield-requesting threads in the future and we 4219 // should really use wait/notify, which is the recommended 4220 // way of doing this type of interaction. Additionally, we should 4221 // consolidate the eight methods that do the yield operation and they 4222 // are almost identical into one for better maintenability and 4223 // readability. See 6445193. 4224 // 4225 // Tony 2006.06.29 4226 for (unsigned i = 0; i < CMSCoordinatorYieldSleepCount && 4227 ConcurrentMarkSweepThread::should_yield() && 4228 !CMSCollector::foregroundGCIsActive(); ++i) { 4229 os::sleep(Thread::current(), 1, false); 4230 ConcurrentMarkSweepThread::acknowledge_yield_request(); 4231 } 4232 4233 ConcurrentMarkSweepThread::synchronize(true); 4234 _bit_map_lock->lock_without_safepoint_check(); 4235 _collector->startTimer(); 4236 } 4237 4238 bool CMSCollector::do_marking_mt(bool asynch) { 4239 assert(ConcGCThreads > 0 && conc_workers() != NULL, "precondition"); 4240 int num_workers = AdaptiveSizePolicy::calc_active_conc_workers( 4241 conc_workers()->total_workers(), 4242 conc_workers()->active_workers(), 4243 Threads::number_of_non_daemon_threads()); 4244 conc_workers()->set_active_workers(num_workers); 4245 4246 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 4247 CompactibleFreeListSpace* perm_space = _permGen->cmsSpace(); 4248 4249 CMSConcMarkingTask tsk(this, 4250 cms_space, 4251 perm_space, 4252 asynch, 4253 conc_workers(), 4254 task_queues()); 4255 4256 // Since the actual number of workers we get may be different 4257 // from the number we requested above, do we need to do anything different 4258 // below? In particular, may be we need to subclass the SequantialSubTasksDone 4259 // class?? XXX 4260 cms_space ->initialize_sequential_subtasks_for_marking(num_workers); 4261 perm_space->initialize_sequential_subtasks_for_marking(num_workers); 4262 4263 // Refs discovery is already non-atomic. 4264 assert(!ref_processor()->discovery_is_atomic(), "Should be non-atomic"); 4265 assert(ref_processor()->discovery_is_mt(), "Discovery should be MT"); 4266 DEBUG_ONLY(RememberKlassesChecker cmx(should_unload_classes());) 4267 conc_workers()->start_task(&tsk); 4268 while (tsk.yielded()) { 4269 tsk.coordinator_yield(); 4270 conc_workers()->continue_task(&tsk); 4271 } 4272 // If the task was aborted, _restart_addr will be non-NULL 4273 assert(tsk.completed() || _restart_addr != NULL, "Inconsistency"); 4274 while (_restart_addr != NULL) { 4275 // XXX For now we do not make use of ABORTED state and have not 4276 // yet implemented the right abort semantics (even in the original 4277 // single-threaded CMS case). That needs some more investigation 4278 // and is deferred for now; see CR# TBF. 07252005YSR. XXX 4279 assert(!CMSAbortSemantics || tsk.aborted(), "Inconsistency"); 4280 // If _restart_addr is non-NULL, a marking stack overflow 4281 // occurred; we need to do a fresh marking iteration from the 4282 // indicated restart address. 4283 if (_foregroundGCIsActive && asynch) { 4284 // We may be running into repeated stack overflows, having 4285 // reached the limit of the stack size, while making very 4286 // slow forward progress. It may be best to bail out and 4287 // let the foreground collector do its job. 4288 // Clear _restart_addr, so that foreground GC 4289 // works from scratch. This avoids the headache of 4290 // a "rescan" which would otherwise be needed because 4291 // of the dirty mod union table & card table. 4292 _restart_addr = NULL; 4293 return false; 4294 } 4295 // Adjust the task to restart from _restart_addr 4296 tsk.reset(_restart_addr); 4297 cms_space ->initialize_sequential_subtasks_for_marking(num_workers, 4298 _restart_addr); 4299 perm_space->initialize_sequential_subtasks_for_marking(num_workers, 4300 _restart_addr); 4301 _restart_addr = NULL; 4302 // Get the workers going again 4303 conc_workers()->start_task(&tsk); 4304 while (tsk.yielded()) { 4305 tsk.coordinator_yield(); 4306 conc_workers()->continue_task(&tsk); 4307 } 4308 } 4309 assert(tsk.completed(), "Inconsistency"); 4310 assert(tsk.result() == true, "Inconsistency"); 4311 return true; 4312 } 4313 4314 bool CMSCollector::do_marking_st(bool asynch) { 4315 ResourceMark rm; 4316 HandleMark hm; 4317 4318 // Temporarily make refs discovery single threaded (non-MT) 4319 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false); 4320 MarkFromRootsClosure markFromRootsClosure(this, _span, &_markBitMap, 4321 &_markStack, &_revisitStack, CMSYield && asynch); 4322 // the last argument to iterate indicates whether the iteration 4323 // should be incremental with periodic yields. 4324 _markBitMap.iterate(&markFromRootsClosure); 4325 // If _restart_addr is non-NULL, a marking stack overflow 4326 // occurred; we need to do a fresh iteration from the 4327 // indicated restart address. 4328 while (_restart_addr != NULL) { 4329 if (_foregroundGCIsActive && asynch) { 4330 // We may be running into repeated stack overflows, having 4331 // reached the limit of the stack size, while making very 4332 // slow forward progress. It may be best to bail out and 4333 // let the foreground collector do its job. 4334 // Clear _restart_addr, so that foreground GC 4335 // works from scratch. This avoids the headache of 4336 // a "rescan" which would otherwise be needed because 4337 // of the dirty mod union table & card table. 4338 _restart_addr = NULL; 4339 return false; // indicating failure to complete marking 4340 } 4341 // Deal with stack overflow: 4342 // we restart marking from _restart_addr 4343 HeapWord* ra = _restart_addr; 4344 markFromRootsClosure.reset(ra); 4345 _restart_addr = NULL; 4346 _markBitMap.iterate(&markFromRootsClosure, ra, _span.end()); 4347 } 4348 return true; 4349 } 4350 4351 void CMSCollector::preclean() { 4352 check_correct_thread_executing(); 4353 assert(Thread::current()->is_ConcurrentGC_thread(), "Wrong thread"); 4354 verify_work_stacks_empty(); 4355 verify_overflow_empty(); 4356 _abort_preclean = false; 4357 if (CMSPrecleaningEnabled) { 4358 _eden_chunk_index = 0; 4359 size_t used = get_eden_used(); 4360 size_t capacity = get_eden_capacity(); 4361 // Don't start sampling unless we will get sufficiently 4362 // many samples. 4363 if (used < (capacity/(CMSScheduleRemarkSamplingRatio * 100) 4364 * CMSScheduleRemarkEdenPenetration)) { 4365 _start_sampling = true; 4366 } else { 4367 _start_sampling = false; 4368 } 4369 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 4370 CMSPhaseAccounting pa(this, "preclean", !PrintGCDetails); 4371 preclean_work(CMSPrecleanRefLists1, CMSPrecleanSurvivors1); 4372 } 4373 CMSTokenSync x(true); // is cms thread 4374 if (CMSPrecleaningEnabled) { 4375 sample_eden(); 4376 _collectorState = AbortablePreclean; 4377 } else { 4378 _collectorState = FinalMarking; 4379 } 4380 verify_work_stacks_empty(); 4381 verify_overflow_empty(); 4382 } 4383 4384 // Try and schedule the remark such that young gen 4385 // occupancy is CMSScheduleRemarkEdenPenetration %. 4386 void CMSCollector::abortable_preclean() { 4387 check_correct_thread_executing(); 4388 assert(CMSPrecleaningEnabled, "Inconsistent control state"); 4389 assert(_collectorState == AbortablePreclean, "Inconsistent control state"); 4390 4391 // If Eden's current occupancy is below this threshold, 4392 // immediately schedule the remark; else preclean 4393 // past the next scavenge in an effort to 4394 // schedule the pause as described avove. By choosing 4395 // CMSScheduleRemarkEdenSizeThreshold >= max eden size 4396 // we will never do an actual abortable preclean cycle. 4397 if (get_eden_used() > CMSScheduleRemarkEdenSizeThreshold) { 4398 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 4399 CMSPhaseAccounting pa(this, "abortable-preclean", !PrintGCDetails); 4400 // We need more smarts in the abortable preclean 4401 // loop below to deal with cases where allocation 4402 // in young gen is very very slow, and our precleaning 4403 // is running a losing race against a horde of 4404 // mutators intent on flooding us with CMS updates 4405 // (dirty cards). 4406 // One, admittedly dumb, strategy is to give up 4407 // after a certain number of abortable precleaning loops 4408 // or after a certain maximum time. We want to make 4409 // this smarter in the next iteration. 4410 // XXX FIX ME!!! YSR 4411 size_t loops = 0, workdone = 0, cumworkdone = 0, waited = 0; 4412 while (!(should_abort_preclean() || 4413 ConcurrentMarkSweepThread::should_terminate())) { 4414 workdone = preclean_work(CMSPrecleanRefLists2, CMSPrecleanSurvivors2); 4415 cumworkdone += workdone; 4416 loops++; 4417 // Voluntarily terminate abortable preclean phase if we have 4418 // been at it for too long. 4419 if ((CMSMaxAbortablePrecleanLoops != 0) && 4420 loops >= CMSMaxAbortablePrecleanLoops) { 4421 if (PrintGCDetails) { 4422 gclog_or_tty->print(" CMS: abort preclean due to loops "); 4423 } 4424 break; 4425 } 4426 if (pa.wallclock_millis() > CMSMaxAbortablePrecleanTime) { 4427 if (PrintGCDetails) { 4428 gclog_or_tty->print(" CMS: abort preclean due to time "); 4429 } 4430 break; 4431 } 4432 // If we are doing little work each iteration, we should 4433 // take a short break. 4434 if (workdone < CMSAbortablePrecleanMinWorkPerIteration) { 4435 // Sleep for some time, waiting for work to accumulate 4436 stopTimer(); 4437 cmsThread()->wait_on_cms_lock(CMSAbortablePrecleanWaitMillis); 4438 startTimer(); 4439 waited++; 4440 } 4441 } 4442 if (PrintCMSStatistics > 0) { 4443 gclog_or_tty->print(" [%d iterations, %d waits, %d cards)] ", 4444 loops, waited, cumworkdone); 4445 } 4446 } 4447 CMSTokenSync x(true); // is cms thread 4448 if (_collectorState != Idling) { 4449 assert(_collectorState == AbortablePreclean, 4450 "Spontaneous state transition?"); 4451 _collectorState = FinalMarking; 4452 } // Else, a foreground collection completed this CMS cycle. 4453 return; 4454 } 4455 4456 // Respond to an Eden sampling opportunity 4457 void CMSCollector::sample_eden() { 4458 // Make sure a young gc cannot sneak in between our 4459 // reading and recording of a sample. 4460 assert(Thread::current()->is_ConcurrentGC_thread(), 4461 "Only the cms thread may collect Eden samples"); 4462 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 4463 "Should collect samples while holding CMS token"); 4464 if (!_start_sampling) { 4465 return; 4466 } 4467 if (_eden_chunk_array) { 4468 if (_eden_chunk_index < _eden_chunk_capacity) { 4469 _eden_chunk_array[_eden_chunk_index] = *_top_addr; // take sample 4470 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr, 4471 "Unexpected state of Eden"); 4472 // We'd like to check that what we just sampled is an oop-start address; 4473 // however, we cannot do that here since the object may not yet have been 4474 // initialized. So we'll instead do the check when we _use_ this sample 4475 // later. 4476 if (_eden_chunk_index == 0 || 4477 (pointer_delta(_eden_chunk_array[_eden_chunk_index], 4478 _eden_chunk_array[_eden_chunk_index-1]) 4479 >= CMSSamplingGrain)) { 4480 _eden_chunk_index++; // commit sample 4481 } 4482 } 4483 } 4484 if ((_collectorState == AbortablePreclean) && !_abort_preclean) { 4485 size_t used = get_eden_used(); 4486 size_t capacity = get_eden_capacity(); 4487 assert(used <= capacity, "Unexpected state of Eden"); 4488 if (used > (capacity/100 * CMSScheduleRemarkEdenPenetration)) { 4489 _abort_preclean = true; 4490 } 4491 } 4492 } 4493 4494 4495 size_t CMSCollector::preclean_work(bool clean_refs, bool clean_survivor) { 4496 assert(_collectorState == Precleaning || 4497 _collectorState == AbortablePreclean, "incorrect state"); 4498 ResourceMark rm; 4499 HandleMark hm; 4500 4501 // Precleaning is currently not MT but the reference processor 4502 // may be set for MT. Disable it temporarily here. 4503 ReferenceProcessor* rp = ref_processor(); 4504 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(rp, false); 4505 4506 // Do one pass of scrubbing the discovered reference lists 4507 // to remove any reference objects with strongly-reachable 4508 // referents. 4509 if (clean_refs) { 4510 CMSPrecleanRefsYieldClosure yield_cl(this); 4511 assert(rp->span().equals(_span), "Spans should be equal"); 4512 CMSKeepAliveClosure keep_alive(this, _span, &_markBitMap, 4513 &_markStack, &_revisitStack, 4514 true /* preclean */); 4515 CMSDrainMarkingStackClosure complete_trace(this, 4516 _span, &_markBitMap, &_markStack, 4517 &keep_alive, true /* preclean */); 4518 4519 // We don't want this step to interfere with a young 4520 // collection because we don't want to take CPU 4521 // or memory bandwidth away from the young GC threads 4522 // (which may be as many as there are CPUs). 4523 // Note that we don't need to protect ourselves from 4524 // interference with mutators because they can't 4525 // manipulate the discovered reference lists nor affect 4526 // the computed reachability of the referents, the 4527 // only properties manipulated by the precleaning 4528 // of these reference lists. 4529 stopTimer(); 4530 CMSTokenSyncWithLocks x(true /* is cms thread */, 4531 bitMapLock()); 4532 startTimer(); 4533 sample_eden(); 4534 4535 // The following will yield to allow foreground 4536 // collection to proceed promptly. XXX YSR: 4537 // The code in this method may need further 4538 // tweaking for better performance and some restructuring 4539 // for cleaner interfaces. 4540 rp->preclean_discovered_references( 4541 rp->is_alive_non_header(), &keep_alive, &complete_trace, 4542 &yield_cl, should_unload_classes()); 4543 } 4544 4545 if (clean_survivor) { // preclean the active survivor space(s) 4546 assert(_young_gen->kind() == Generation::DefNew || 4547 _young_gen->kind() == Generation::ParNew || 4548 _young_gen->kind() == Generation::ASParNew, 4549 "incorrect type for cast"); 4550 DefNewGeneration* dng = (DefNewGeneration*)_young_gen; 4551 PushAndMarkClosure pam_cl(this, _span, ref_processor(), 4552 &_markBitMap, &_modUnionTable, 4553 &_markStack, &_revisitStack, 4554 true /* precleaning phase */); 4555 stopTimer(); 4556 CMSTokenSyncWithLocks ts(true /* is cms thread */, 4557 bitMapLock()); 4558 startTimer(); 4559 unsigned int before_count = 4560 GenCollectedHeap::heap()->total_collections(); 4561 SurvivorSpacePrecleanClosure 4562 sss_cl(this, _span, &_markBitMap, &_markStack, 4563 &pam_cl, before_count, CMSYield); 4564 DEBUG_ONLY(RememberKlassesChecker mx(should_unload_classes());) 4565 dng->from()->object_iterate_careful(&sss_cl); 4566 dng->to()->object_iterate_careful(&sss_cl); 4567 } 4568 MarkRefsIntoAndScanClosure 4569 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable, 4570 &_markStack, &_revisitStack, this, CMSYield, 4571 true /* precleaning phase */); 4572 // CAUTION: The following closure has persistent state that may need to 4573 // be reset upon a decrease in the sequence of addresses it 4574 // processes. 4575 ScanMarkedObjectsAgainCarefullyClosure 4576 smoac_cl(this, _span, 4577 &_markBitMap, &_markStack, &_revisitStack, &mrias_cl, CMSYield); 4578 4579 // Preclean dirty cards in ModUnionTable and CardTable using 4580 // appropriate convergence criterion; 4581 // repeat CMSPrecleanIter times unless we find that 4582 // we are losing. 4583 assert(CMSPrecleanIter < 10, "CMSPrecleanIter is too large"); 4584 assert(CMSPrecleanNumerator < CMSPrecleanDenominator, 4585 "Bad convergence multiplier"); 4586 assert(CMSPrecleanThreshold >= 100, 4587 "Unreasonably low CMSPrecleanThreshold"); 4588 4589 size_t numIter, cumNumCards, lastNumCards, curNumCards; 4590 for (numIter = 0, cumNumCards = lastNumCards = curNumCards = 0; 4591 numIter < CMSPrecleanIter; 4592 numIter++, lastNumCards = curNumCards, cumNumCards += curNumCards) { 4593 curNumCards = preclean_mod_union_table(_cmsGen, &smoac_cl); 4594 if (CMSPermGenPrecleaningEnabled) { 4595 curNumCards += preclean_mod_union_table(_permGen, &smoac_cl); 4596 } 4597 if (Verbose && PrintGCDetails) { 4598 gclog_or_tty->print(" (modUnionTable: %d cards)", curNumCards); 4599 } 4600 // Either there are very few dirty cards, so re-mark 4601 // pause will be small anyway, or our pre-cleaning isn't 4602 // that much faster than the rate at which cards are being 4603 // dirtied, so we might as well stop and re-mark since 4604 // precleaning won't improve our re-mark time by much. 4605 if (curNumCards <= CMSPrecleanThreshold || 4606 (numIter > 0 && 4607 (curNumCards * CMSPrecleanDenominator > 4608 lastNumCards * CMSPrecleanNumerator))) { 4609 numIter++; 4610 cumNumCards += curNumCards; 4611 break; 4612 } 4613 } 4614 curNumCards = preclean_card_table(_cmsGen, &smoac_cl); 4615 if (CMSPermGenPrecleaningEnabled) { 4616 curNumCards += preclean_card_table(_permGen, &smoac_cl); 4617 } 4618 cumNumCards += curNumCards; 4619 if (PrintGCDetails && PrintCMSStatistics != 0) { 4620 gclog_or_tty->print_cr(" (cardTable: %d cards, re-scanned %d cards, %d iterations)", 4621 curNumCards, cumNumCards, numIter); 4622 } 4623 return cumNumCards; // as a measure of useful work done 4624 } 4625 4626 // PRECLEANING NOTES: 4627 // Precleaning involves: 4628 // . reading the bits of the modUnionTable and clearing the set bits. 4629 // . For the cards corresponding to the set bits, we scan the 4630 // objects on those cards. This means we need the free_list_lock 4631 // so that we can safely iterate over the CMS space when scanning 4632 // for oops. 4633 // . When we scan the objects, we'll be both reading and setting 4634 // marks in the marking bit map, so we'll need the marking bit map. 4635 // . For protecting _collector_state transitions, we take the CGC_lock. 4636 // Note that any races in the reading of of card table entries by the 4637 // CMS thread on the one hand and the clearing of those entries by the 4638 // VM thread or the setting of those entries by the mutator threads on the 4639 // other are quite benign. However, for efficiency it makes sense to keep 4640 // the VM thread from racing with the CMS thread while the latter is 4641 // dirty card info to the modUnionTable. We therefore also use the 4642 // CGC_lock to protect the reading of the card table and the mod union 4643 // table by the CM thread. 4644 // . We run concurrently with mutator updates, so scanning 4645 // needs to be done carefully -- we should not try to scan 4646 // potentially uninitialized objects. 4647 // 4648 // Locking strategy: While holding the CGC_lock, we scan over and 4649 // reset a maximal dirty range of the mod union / card tables, then lock 4650 // the free_list_lock and bitmap lock to do a full marking, then 4651 // release these locks; and repeat the cycle. This allows for a 4652 // certain amount of fairness in the sharing of these locks between 4653 // the CMS collector on the one hand, and the VM thread and the 4654 // mutators on the other. 4655 4656 // NOTE: preclean_mod_union_table() and preclean_card_table() 4657 // further below are largely identical; if you need to modify 4658 // one of these methods, please check the other method too. 4659 4660 size_t CMSCollector::preclean_mod_union_table( 4661 ConcurrentMarkSweepGeneration* gen, 4662 ScanMarkedObjectsAgainCarefullyClosure* cl) { 4663 verify_work_stacks_empty(); 4664 verify_overflow_empty(); 4665 4666 // Turn off checking for this method but turn it back on 4667 // selectively. There are yield points in this method 4668 // but it is difficult to turn the checking off just around 4669 // the yield points. It is simpler to selectively turn 4670 // it on. 4671 DEBUG_ONLY(RememberKlassesChecker mux(false);) 4672 4673 // strategy: starting with the first card, accumulate contiguous 4674 // ranges of dirty cards; clear these cards, then scan the region 4675 // covered by these cards. 4676 4677 // Since all of the MUT is committed ahead, we can just use 4678 // that, in case the generations expand while we are precleaning. 4679 // It might also be fine to just use the committed part of the 4680 // generation, but we might potentially miss cards when the 4681 // generation is rapidly expanding while we are in the midst 4682 // of precleaning. 4683 HeapWord* startAddr = gen->reserved().start(); 4684 HeapWord* endAddr = gen->reserved().end(); 4685 4686 cl->setFreelistLock(gen->freelistLock()); // needed for yielding 4687 4688 size_t numDirtyCards, cumNumDirtyCards; 4689 HeapWord *nextAddr, *lastAddr; 4690 for (cumNumDirtyCards = numDirtyCards = 0, 4691 nextAddr = lastAddr = startAddr; 4692 nextAddr < endAddr; 4693 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 4694 4695 ResourceMark rm; 4696 HandleMark hm; 4697 4698 MemRegion dirtyRegion; 4699 { 4700 stopTimer(); 4701 // Potential yield point 4702 CMSTokenSync ts(true); 4703 startTimer(); 4704 sample_eden(); 4705 // Get dirty region starting at nextOffset (inclusive), 4706 // simultaneously clearing it. 4707 dirtyRegion = 4708 _modUnionTable.getAndClearMarkedRegion(nextAddr, endAddr); 4709 assert(dirtyRegion.start() >= nextAddr, 4710 "returned region inconsistent?"); 4711 } 4712 // Remember where the next search should begin. 4713 // The returned region (if non-empty) is a right open interval, 4714 // so lastOffset is obtained from the right end of that 4715 // interval. 4716 lastAddr = dirtyRegion.end(); 4717 // Should do something more transparent and less hacky XXX 4718 numDirtyCards = 4719 _modUnionTable.heapWordDiffToOffsetDiff(dirtyRegion.word_size()); 4720 4721 // We'll scan the cards in the dirty region (with periodic 4722 // yields for foreground GC as needed). 4723 if (!dirtyRegion.is_empty()) { 4724 assert(numDirtyCards > 0, "consistency check"); 4725 HeapWord* stop_point = NULL; 4726 stopTimer(); 4727 // Potential yield point 4728 CMSTokenSyncWithLocks ts(true, gen->freelistLock(), 4729 bitMapLock()); 4730 startTimer(); 4731 { 4732 verify_work_stacks_empty(); 4733 verify_overflow_empty(); 4734 sample_eden(); 4735 DEBUG_ONLY(RememberKlassesChecker mx(should_unload_classes());) 4736 stop_point = 4737 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 4738 } 4739 if (stop_point != NULL) { 4740 // The careful iteration stopped early either because it found an 4741 // uninitialized object, or because we were in the midst of an 4742 // "abortable preclean", which should now be aborted. Redirty 4743 // the bits corresponding to the partially-scanned or unscanned 4744 // cards. We'll either restart at the next block boundary or 4745 // abort the preclean. 4746 assert((CMSPermGenPrecleaningEnabled && (gen == _permGen)) || 4747 (_collectorState == AbortablePreclean && should_abort_preclean()), 4748 "Unparsable objects should only be in perm gen."); 4749 _modUnionTable.mark_range(MemRegion(stop_point, dirtyRegion.end())); 4750 if (should_abort_preclean()) { 4751 break; // out of preclean loop 4752 } else { 4753 // Compute the next address at which preclean should pick up; 4754 // might need bitMapLock in order to read P-bits. 4755 lastAddr = next_card_start_after_block(stop_point); 4756 } 4757 } 4758 } else { 4759 assert(lastAddr == endAddr, "consistency check"); 4760 assert(numDirtyCards == 0, "consistency check"); 4761 break; 4762 } 4763 } 4764 verify_work_stacks_empty(); 4765 verify_overflow_empty(); 4766 return cumNumDirtyCards; 4767 } 4768 4769 // NOTE: preclean_mod_union_table() above and preclean_card_table() 4770 // below are largely identical; if you need to modify 4771 // one of these methods, please check the other method too. 4772 4773 size_t CMSCollector::preclean_card_table(ConcurrentMarkSweepGeneration* gen, 4774 ScanMarkedObjectsAgainCarefullyClosure* cl) { 4775 // strategy: it's similar to precleamModUnionTable above, in that 4776 // we accumulate contiguous ranges of dirty cards, mark these cards 4777 // precleaned, then scan the region covered by these cards. 4778 HeapWord* endAddr = (HeapWord*)(gen->_virtual_space.high()); 4779 HeapWord* startAddr = (HeapWord*)(gen->_virtual_space.low()); 4780 4781 cl->setFreelistLock(gen->freelistLock()); // needed for yielding 4782 4783 size_t numDirtyCards, cumNumDirtyCards; 4784 HeapWord *lastAddr, *nextAddr; 4785 4786 for (cumNumDirtyCards = numDirtyCards = 0, 4787 nextAddr = lastAddr = startAddr; 4788 nextAddr < endAddr; 4789 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 4790 4791 ResourceMark rm; 4792 HandleMark hm; 4793 4794 MemRegion dirtyRegion; 4795 { 4796 // See comments in "Precleaning notes" above on why we 4797 // do this locking. XXX Could the locking overheads be 4798 // too high when dirty cards are sparse? [I don't think so.] 4799 stopTimer(); 4800 CMSTokenSync x(true); // is cms thread 4801 startTimer(); 4802 sample_eden(); 4803 // Get and clear dirty region from card table 4804 dirtyRegion = _ct->ct_bs()->dirty_card_range_after_reset( 4805 MemRegion(nextAddr, endAddr), 4806 true, 4807 CardTableModRefBS::precleaned_card_val()); 4808 4809 assert(dirtyRegion.start() >= nextAddr, 4810 "returned region inconsistent?"); 4811 } 4812 lastAddr = dirtyRegion.end(); 4813 numDirtyCards = 4814 dirtyRegion.word_size()/CardTableModRefBS::card_size_in_words; 4815 4816 if (!dirtyRegion.is_empty()) { 4817 stopTimer(); 4818 CMSTokenSyncWithLocks ts(true, gen->freelistLock(), bitMapLock()); 4819 startTimer(); 4820 sample_eden(); 4821 verify_work_stacks_empty(); 4822 verify_overflow_empty(); 4823 DEBUG_ONLY(RememberKlassesChecker mx(should_unload_classes());) 4824 HeapWord* stop_point = 4825 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 4826 if (stop_point != NULL) { 4827 // The careful iteration stopped early because it found an 4828 // uninitialized object. Redirty the bits corresponding to the 4829 // partially-scanned or unscanned cards, and start again at the 4830 // next block boundary. 4831 assert(CMSPermGenPrecleaningEnabled || 4832 (_collectorState == AbortablePreclean && should_abort_preclean()), 4833 "Unparsable objects should only be in perm gen."); 4834 _ct->ct_bs()->invalidate(MemRegion(stop_point, dirtyRegion.end())); 4835 if (should_abort_preclean()) { 4836 break; // out of preclean loop 4837 } else { 4838 // Compute the next address at which preclean should pick up. 4839 lastAddr = next_card_start_after_block(stop_point); 4840 } 4841 } 4842 } else { 4843 break; 4844 } 4845 } 4846 verify_work_stacks_empty(); 4847 verify_overflow_empty(); 4848 return cumNumDirtyCards; 4849 } 4850 4851 void CMSCollector::checkpointRootsFinal(bool asynch, 4852 bool clear_all_soft_refs, bool init_mark_was_synchronous) { 4853 assert(_collectorState == FinalMarking, "incorrect state transition?"); 4854 check_correct_thread_executing(); 4855 // world is stopped at this checkpoint 4856 assert(SafepointSynchronize::is_at_safepoint(), 4857 "world should be stopped"); 4858 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 4859 4860 verify_work_stacks_empty(); 4861 verify_overflow_empty(); 4862 4863 SpecializationStats::clear(); 4864 if (PrintGCDetails) { 4865 gclog_or_tty->print("[YG occupancy: "SIZE_FORMAT" K ("SIZE_FORMAT" K)]", 4866 _young_gen->used() / K, 4867 _young_gen->capacity() / K); 4868 } 4869 if (asynch) { 4870 if (CMSScavengeBeforeRemark) { 4871 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4872 // Temporarily set flag to false, GCH->do_collection will 4873 // expect it to be false and set to true 4874 FlagSetting fl(gch->_is_gc_active, false); 4875 NOT_PRODUCT(TraceTime t("Scavenge-Before-Remark", 4876 PrintGCDetails && Verbose, true, gclog_or_tty);) 4877 int level = _cmsGen->level() - 1; 4878 if (level >= 0) { 4879 gch->do_collection(true, // full (i.e. force, see below) 4880 false, // !clear_all_soft_refs 4881 0, // size 4882 false, // is_tlab 4883 level // max_level 4884 ); 4885 } 4886 } 4887 FreelistLocker x(this); 4888 MutexLockerEx y(bitMapLock(), 4889 Mutex::_no_safepoint_check_flag); 4890 assert(!init_mark_was_synchronous, "but that's impossible!"); 4891 checkpointRootsFinalWork(asynch, clear_all_soft_refs, false); 4892 } else { 4893 // already have all the locks 4894 checkpointRootsFinalWork(asynch, clear_all_soft_refs, 4895 init_mark_was_synchronous); 4896 } 4897 verify_work_stacks_empty(); 4898 verify_overflow_empty(); 4899 SpecializationStats::print(); 4900 } 4901 4902 void CMSCollector::checkpointRootsFinalWork(bool asynch, 4903 bool clear_all_soft_refs, bool init_mark_was_synchronous) { 4904 4905 NOT_PRODUCT(TraceTime tr("checkpointRootsFinalWork", PrintGCDetails, false, gclog_or_tty);) 4906 4907 assert(haveFreelistLocks(), "must have free list locks"); 4908 assert_lock_strong(bitMapLock()); 4909 4910 if (UseAdaptiveSizePolicy) { 4911 size_policy()->checkpoint_roots_final_begin(); 4912 } 4913 4914 ResourceMark rm; 4915 HandleMark hm; 4916 4917 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4918 4919 if (should_unload_classes()) { 4920 CodeCache::gc_prologue(); 4921 } 4922 assert(haveFreelistLocks(), "must have free list locks"); 4923 assert_lock_strong(bitMapLock()); 4924 4925 DEBUG_ONLY(RememberKlassesChecker fmx(should_unload_classes());) 4926 if (!init_mark_was_synchronous) { 4927 // We might assume that we need not fill TLAB's when 4928 // CMSScavengeBeforeRemark is set, because we may have just done 4929 // a scavenge which would have filled all TLAB's -- and besides 4930 // Eden would be empty. This however may not always be the case -- 4931 // for instance although we asked for a scavenge, it may not have 4932 // happened because of a JNI critical section. We probably need 4933 // a policy for deciding whether we can in that case wait until 4934 // the critical section releases and then do the remark following 4935 // the scavenge, and skip it here. In the absence of that policy, 4936 // or of an indication of whether the scavenge did indeed occur, 4937 // we cannot rely on TLAB's having been filled and must do 4938 // so here just in case a scavenge did not happen. 4939 gch->ensure_parsability(false); // fill TLAB's, but no need to retire them 4940 // Update the saved marks which may affect the root scans. 4941 gch->save_marks(); 4942 4943 { 4944 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;) 4945 4946 // Note on the role of the mod union table: 4947 // Since the marker in "markFromRoots" marks concurrently with 4948 // mutators, it is possible for some reachable objects not to have been 4949 // scanned. For instance, an only reference to an object A was 4950 // placed in object B after the marker scanned B. Unless B is rescanned, 4951 // A would be collected. Such updates to references in marked objects 4952 // are detected via the mod union table which is the set of all cards 4953 // dirtied since the first checkpoint in this GC cycle and prior to 4954 // the most recent young generation GC, minus those cleaned up by the 4955 // concurrent precleaning. 4956 if (CMSParallelRemarkEnabled && CollectedHeap::use_parallel_gc_threads()) { 4957 TraceTime t("Rescan (parallel) ", PrintGCDetails, false, gclog_or_tty); 4958 do_remark_parallel(); 4959 } else { 4960 TraceTime t("Rescan (non-parallel) ", PrintGCDetails, false, 4961 gclog_or_tty); 4962 do_remark_non_parallel(); 4963 } 4964 } 4965 } else { 4966 assert(!asynch, "Can't have init_mark_was_synchronous in asynch mode"); 4967 // The initial mark was stop-world, so there's no rescanning to 4968 // do; go straight on to the next step below. 4969 } 4970 verify_work_stacks_empty(); 4971 verify_overflow_empty(); 4972 4973 { 4974 NOT_PRODUCT(TraceTime ts("refProcessingWork", PrintGCDetails, false, gclog_or_tty);) 4975 refProcessingWork(asynch, clear_all_soft_refs); 4976 } 4977 verify_work_stacks_empty(); 4978 verify_overflow_empty(); 4979 4980 if (should_unload_classes()) { 4981 CodeCache::gc_epilogue(); 4982 } 4983 JvmtiExport::gc_epilogue(); 4984 4985 // If we encountered any (marking stack / work queue) overflow 4986 // events during the current CMS cycle, take appropriate 4987 // remedial measures, where possible, so as to try and avoid 4988 // recurrence of that condition. 4989 assert(_markStack.isEmpty(), "No grey objects"); 4990 size_t ser_ovflw = _ser_pmc_remark_ovflw + _ser_pmc_preclean_ovflw + 4991 _ser_kac_ovflw + _ser_kac_preclean_ovflw; 4992 if (ser_ovflw > 0) { 4993 if (PrintCMSStatistics != 0) { 4994 gclog_or_tty->print_cr("Marking stack overflow (benign) " 4995 "(pmc_pc="SIZE_FORMAT", pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT 4996 ", kac_preclean="SIZE_FORMAT")", 4997 _ser_pmc_preclean_ovflw, _ser_pmc_remark_ovflw, 4998 _ser_kac_ovflw, _ser_kac_preclean_ovflw); 4999 } 5000 _markStack.expand(); 5001 _ser_pmc_remark_ovflw = 0; 5002 _ser_pmc_preclean_ovflw = 0; 5003 _ser_kac_preclean_ovflw = 0; 5004 _ser_kac_ovflw = 0; 5005 } 5006 if (_par_pmc_remark_ovflw > 0 || _par_kac_ovflw > 0) { 5007 if (PrintCMSStatistics != 0) { 5008 gclog_or_tty->print_cr("Work queue overflow (benign) " 5009 "(pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT")", 5010 _par_pmc_remark_ovflw, _par_kac_ovflw); 5011 } 5012 _par_pmc_remark_ovflw = 0; 5013 _par_kac_ovflw = 0; 5014 } 5015 if (PrintCMSStatistics != 0) { 5016 if (_markStack._hit_limit > 0) { 5017 gclog_or_tty->print_cr(" (benign) Hit max stack size limit ("SIZE_FORMAT")", 5018 _markStack._hit_limit); 5019 } 5020 if (_markStack._failed_double > 0) { 5021 gclog_or_tty->print_cr(" (benign) Failed stack doubling ("SIZE_FORMAT")," 5022 " current capacity "SIZE_FORMAT, 5023 _markStack._failed_double, 5024 _markStack.capacity()); 5025 } 5026 } 5027 _markStack._hit_limit = 0; 5028 _markStack._failed_double = 0; 5029 5030 // Check that all the klasses have been checked 5031 assert(_revisitStack.isEmpty(), "Not all klasses revisited"); 5032 5033 if ((VerifyAfterGC || VerifyDuringGC) && 5034 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 5035 verify_after_remark(); 5036 } 5037 5038 // Change under the freelistLocks. 5039 _collectorState = Sweeping; 5040 // Call isAllClear() under bitMapLock 5041 assert(_modUnionTable.isAllClear(), "Should be clear by end of the" 5042 " final marking"); 5043 if (UseAdaptiveSizePolicy) { 5044 size_policy()->checkpoint_roots_final_end(gch->gc_cause()); 5045 } 5046 } 5047 5048 // Parallel remark task 5049 class CMSParRemarkTask: public AbstractGangTask { 5050 CMSCollector* _collector; 5051 int _n_workers; 5052 CompactibleFreeListSpace* _cms_space; 5053 CompactibleFreeListSpace* _perm_space; 5054 5055 // The per-thread work queues, available here for stealing. 5056 OopTaskQueueSet* _task_queues; 5057 ParallelTaskTerminator _term; 5058 5059 public: 5060 // A value of 0 passed to n_workers will cause the number of 5061 // workers to be taken from the active workers in the work gang. 5062 CMSParRemarkTask(CMSCollector* collector, 5063 CompactibleFreeListSpace* cms_space, 5064 CompactibleFreeListSpace* perm_space, 5065 int n_workers, FlexibleWorkGang* workers, 5066 OopTaskQueueSet* task_queues): 5067 AbstractGangTask("Rescan roots and grey objects in parallel"), 5068 _collector(collector), 5069 _cms_space(cms_space), _perm_space(perm_space), 5070 _n_workers(n_workers), 5071 _task_queues(task_queues), 5072 _term(n_workers, task_queues) { } 5073 5074 OopTaskQueueSet* task_queues() { return _task_queues; } 5075 5076 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 5077 5078 ParallelTaskTerminator* terminator() { return &_term; } 5079 int n_workers() { return _n_workers; } 5080 5081 void work(uint worker_id); 5082 5083 private: 5084 // Work method in support of parallel rescan ... of young gen spaces 5085 void do_young_space_rescan(int i, Par_MarkRefsIntoAndScanClosure* cl, 5086 ContiguousSpace* space, 5087 HeapWord** chunk_array, size_t chunk_top); 5088 5089 // ... of dirty cards in old space 5090 void do_dirty_card_rescan_tasks(CompactibleFreeListSpace* sp, int i, 5091 Par_MarkRefsIntoAndScanClosure* cl); 5092 5093 // ... work stealing for the above 5094 void do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, int* seed); 5095 }; 5096 5097 // work_queue(i) is passed to the closure 5098 // Par_MarkRefsIntoAndScanClosure. The "i" parameter 5099 // also is passed to do_dirty_card_rescan_tasks() and to 5100 // do_work_steal() to select the i-th task_queue. 5101 5102 void CMSParRemarkTask::work(uint worker_id) { 5103 elapsedTimer _timer; 5104 ResourceMark rm; 5105 HandleMark hm; 5106 5107 // ---------- rescan from roots -------------- 5108 _timer.start(); 5109 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5110 Par_MarkRefsIntoAndScanClosure par_mrias_cl(_collector, 5111 _collector->_span, _collector->ref_processor(), 5112 &(_collector->_markBitMap), 5113 work_queue(worker_id), &(_collector->_revisitStack)); 5114 5115 // Rescan young gen roots first since these are likely 5116 // coarsely partitioned and may, on that account, constitute 5117 // the critical path; thus, it's best to start off that 5118 // work first. 5119 // ---------- young gen roots -------------- 5120 { 5121 DefNewGeneration* dng = _collector->_young_gen->as_DefNewGeneration(); 5122 EdenSpace* eden_space = dng->eden(); 5123 ContiguousSpace* from_space = dng->from(); 5124 ContiguousSpace* to_space = dng->to(); 5125 5126 HeapWord** eca = _collector->_eden_chunk_array; 5127 size_t ect = _collector->_eden_chunk_index; 5128 HeapWord** sca = _collector->_survivor_chunk_array; 5129 size_t sct = _collector->_survivor_chunk_index; 5130 5131 assert(ect <= _collector->_eden_chunk_capacity, "out of bounds"); 5132 assert(sct <= _collector->_survivor_chunk_capacity, "out of bounds"); 5133 5134 do_young_space_rescan(worker_id, &par_mrias_cl, to_space, NULL, 0); 5135 do_young_space_rescan(worker_id, &par_mrias_cl, from_space, sca, sct); 5136 do_young_space_rescan(worker_id, &par_mrias_cl, eden_space, eca, ect); 5137 5138 _timer.stop(); 5139 if (PrintCMSStatistics != 0) { 5140 gclog_or_tty->print_cr( 5141 "Finished young gen rescan work in %dth thread: %3.3f sec", 5142 worker_id, _timer.seconds()); 5143 } 5144 } 5145 5146 // ---------- remaining roots -------------- 5147 _timer.reset(); 5148 _timer.start(); 5149 gch->gen_process_strong_roots(_collector->_cmsGen->level(), 5150 false, // yg was scanned above 5151 false, // this is parallel code 5152 true, // collecting perm gen 5153 SharedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()), 5154 &par_mrias_cl, 5155 true, // walk all of code cache if (so & SO_CodeCache) 5156 NULL); 5157 assert(_collector->should_unload_classes() 5158 || (_collector->CMSCollector::roots_scanning_options() & SharedHeap::SO_CodeCache), 5159 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 5160 _timer.stop(); 5161 if (PrintCMSStatistics != 0) { 5162 gclog_or_tty->print_cr( 5163 "Finished remaining root rescan work in %dth thread: %3.3f sec", 5164 worker_id, _timer.seconds()); 5165 } 5166 5167 // ---------- rescan dirty cards ------------ 5168 _timer.reset(); 5169 _timer.start(); 5170 5171 // Do the rescan tasks for each of the two spaces 5172 // (cms_space and perm_space) in turn. 5173 // "worker_id" is passed to select the task_queue for "worker_id" 5174 do_dirty_card_rescan_tasks(_cms_space, worker_id, &par_mrias_cl); 5175 do_dirty_card_rescan_tasks(_perm_space, worker_id, &par_mrias_cl); 5176 _timer.stop(); 5177 if (PrintCMSStatistics != 0) { 5178 gclog_or_tty->print_cr( 5179 "Finished dirty card rescan work in %dth thread: %3.3f sec", 5180 worker_id, _timer.seconds()); 5181 } 5182 5183 // ---------- steal work from other threads ... 5184 // ---------- ... and drain overflow list. 5185 _timer.reset(); 5186 _timer.start(); 5187 do_work_steal(worker_id, &par_mrias_cl, _collector->hash_seed(worker_id)); 5188 _timer.stop(); 5189 if (PrintCMSStatistics != 0) { 5190 gclog_or_tty->print_cr( 5191 "Finished work stealing in %dth thread: %3.3f sec", 5192 worker_id, _timer.seconds()); 5193 } 5194 } 5195 5196 // Note that parameter "i" is not used. 5197 void 5198 CMSParRemarkTask::do_young_space_rescan(int i, 5199 Par_MarkRefsIntoAndScanClosure* cl, ContiguousSpace* space, 5200 HeapWord** chunk_array, size_t chunk_top) { 5201 // Until all tasks completed: 5202 // . claim an unclaimed task 5203 // . compute region boundaries corresponding to task claimed 5204 // using chunk_array 5205 // . par_oop_iterate(cl) over that region 5206 5207 ResourceMark rm; 5208 HandleMark hm; 5209 5210 SequentialSubTasksDone* pst = space->par_seq_tasks(); 5211 assert(pst->valid(), "Uninitialized use?"); 5212 5213 uint nth_task = 0; 5214 uint n_tasks = pst->n_tasks(); 5215 5216 HeapWord *start, *end; 5217 while (!pst->is_task_claimed(/* reference */ nth_task)) { 5218 // We claimed task # nth_task; compute its boundaries. 5219 if (chunk_top == 0) { // no samples were taken 5220 assert(nth_task == 0 && n_tasks == 1, "Can have only 1 EdenSpace task"); 5221 start = space->bottom(); 5222 end = space->top(); 5223 } else if (nth_task == 0) { 5224 start = space->bottom(); 5225 end = chunk_array[nth_task]; 5226 } else if (nth_task < (uint)chunk_top) { 5227 assert(nth_task >= 1, "Control point invariant"); 5228 start = chunk_array[nth_task - 1]; 5229 end = chunk_array[nth_task]; 5230 } else { 5231 assert(nth_task == (uint)chunk_top, "Control point invariant"); 5232 start = chunk_array[chunk_top - 1]; 5233 end = space->top(); 5234 } 5235 MemRegion mr(start, end); 5236 // Verify that mr is in space 5237 assert(mr.is_empty() || space->used_region().contains(mr), 5238 "Should be in space"); 5239 // Verify that "start" is an object boundary 5240 assert(mr.is_empty() || oop(mr.start())->is_oop(), 5241 "Should be an oop"); 5242 space->par_oop_iterate(mr, cl); 5243 } 5244 pst->all_tasks_completed(); 5245 } 5246 5247 void 5248 CMSParRemarkTask::do_dirty_card_rescan_tasks( 5249 CompactibleFreeListSpace* sp, int i, 5250 Par_MarkRefsIntoAndScanClosure* cl) { 5251 // Until all tasks completed: 5252 // . claim an unclaimed task 5253 // . compute region boundaries corresponding to task claimed 5254 // . transfer dirty bits ct->mut for that region 5255 // . apply rescanclosure to dirty mut bits for that region 5256 5257 ResourceMark rm; 5258 HandleMark hm; 5259 5260 OopTaskQueue* work_q = work_queue(i); 5261 ModUnionClosure modUnionClosure(&(_collector->_modUnionTable)); 5262 // CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! 5263 // CAUTION: This closure has state that persists across calls to 5264 // the work method dirty_range_iterate_clear() in that it has 5265 // imbedded in it a (subtype of) UpwardsObjectClosure. The 5266 // use of that state in the imbedded UpwardsObjectClosure instance 5267 // assumes that the cards are always iterated (even if in parallel 5268 // by several threads) in monotonically increasing order per each 5269 // thread. This is true of the implementation below which picks 5270 // card ranges (chunks) in monotonically increasing order globally 5271 // and, a-fortiori, in monotonically increasing order per thread 5272 // (the latter order being a subsequence of the former). 5273 // If the work code below is ever reorganized into a more chaotic 5274 // work-partitioning form than the current "sequential tasks" 5275 // paradigm, the use of that persistent state will have to be 5276 // revisited and modified appropriately. See also related 5277 // bug 4756801 work on which should examine this code to make 5278 // sure that the changes there do not run counter to the 5279 // assumptions made here and necessary for correctness and 5280 // efficiency. Note also that this code might yield inefficient 5281 // behaviour in the case of very large objects that span one or 5282 // more work chunks. Such objects would potentially be scanned 5283 // several times redundantly. Work on 4756801 should try and 5284 // address that performance anomaly if at all possible. XXX 5285 MemRegion full_span = _collector->_span; 5286 CMSBitMap* bm = &(_collector->_markBitMap); // shared 5287 CMSMarkStack* rs = &(_collector->_revisitStack); // shared 5288 MarkFromDirtyCardsClosure 5289 greyRescanClosure(_collector, full_span, // entire span of interest 5290 sp, bm, work_q, rs, cl); 5291 5292 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 5293 assert(pst->valid(), "Uninitialized use?"); 5294 uint nth_task = 0; 5295 const int alignment = CardTableModRefBS::card_size * BitsPerWord; 5296 MemRegion span = sp->used_region(); 5297 HeapWord* start_addr = span.start(); 5298 HeapWord* end_addr = (HeapWord*)round_to((intptr_t)span.end(), 5299 alignment); 5300 const size_t chunk_size = sp->rescan_task_size(); // in HeapWord units 5301 assert((HeapWord*)round_to((intptr_t)start_addr, alignment) == 5302 start_addr, "Check alignment"); 5303 assert((size_t)round_to((intptr_t)chunk_size, alignment) == 5304 chunk_size, "Check alignment"); 5305 5306 while (!pst->is_task_claimed(/* reference */ nth_task)) { 5307 // Having claimed the nth_task, compute corresponding mem-region, 5308 // which is a-fortiori aligned correctly (i.e. at a MUT bopundary). 5309 // The alignment restriction ensures that we do not need any 5310 // synchronization with other gang-workers while setting or 5311 // clearing bits in thus chunk of the MUT. 5312 MemRegion this_span = MemRegion(start_addr + nth_task*chunk_size, 5313 start_addr + (nth_task+1)*chunk_size); 5314 // The last chunk's end might be way beyond end of the 5315 // used region. In that case pull back appropriately. 5316 if (this_span.end() > end_addr) { 5317 this_span.set_end(end_addr); 5318 assert(!this_span.is_empty(), "Program logic (calculation of n_tasks)"); 5319 } 5320 // Iterate over the dirty cards covering this chunk, marking them 5321 // precleaned, and setting the corresponding bits in the mod union 5322 // table. Since we have been careful to partition at Card and MUT-word 5323 // boundaries no synchronization is needed between parallel threads. 5324 _collector->_ct->ct_bs()->dirty_card_iterate(this_span, 5325 &modUnionClosure); 5326 5327 // Having transferred these marks into the modUnionTable, 5328 // rescan the marked objects on the dirty cards in the modUnionTable. 5329 // Even if this is at a synchronous collection, the initial marking 5330 // may have been done during an asynchronous collection so there 5331 // may be dirty bits in the mod-union table. 5332 _collector->_modUnionTable.dirty_range_iterate_clear( 5333 this_span, &greyRescanClosure); 5334 _collector->_modUnionTable.verifyNoOneBitsInRange( 5335 this_span.start(), 5336 this_span.end()); 5337 } 5338 pst->all_tasks_completed(); // declare that i am done 5339 } 5340 5341 // . see if we can share work_queues with ParNew? XXX 5342 void 5343 CMSParRemarkTask::do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, 5344 int* seed) { 5345 OopTaskQueue* work_q = work_queue(i); 5346 NOT_PRODUCT(int num_steals = 0;) 5347 oop obj_to_scan; 5348 CMSBitMap* bm = &(_collector->_markBitMap); 5349 5350 while (true) { 5351 // Completely finish any left over work from (an) earlier round(s) 5352 cl->trim_queue(0); 5353 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 5354 (size_t)ParGCDesiredObjsFromOverflowList); 5355 // Now check if there's any work in the overflow list 5356 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads, 5357 // only affects the number of attempts made to get work from the 5358 // overflow list and does not affect the number of workers. Just 5359 // pass ParallelGCThreads so this behavior is unchanged. 5360 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 5361 work_q, 5362 ParallelGCThreads)) { 5363 // found something in global overflow list; 5364 // not yet ready to go stealing work from others. 5365 // We'd like to assert(work_q->size() != 0, ...) 5366 // because we just took work from the overflow list, 5367 // but of course we can't since all of that could have 5368 // been already stolen from us. 5369 // "He giveth and He taketh away." 5370 continue; 5371 } 5372 // Verify that we have no work before we resort to stealing 5373 assert(work_q->size() == 0, "Have work, shouldn't steal"); 5374 // Try to steal from other queues that have work 5375 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 5376 NOT_PRODUCT(num_steals++;) 5377 assert(obj_to_scan->is_oop(), "Oops, not an oop!"); 5378 assert(bm->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 5379 // Do scanning work 5380 obj_to_scan->oop_iterate(cl); 5381 // Loop around, finish this work, and try to steal some more 5382 } else if (terminator()->offer_termination()) { 5383 break; // nirvana from the infinite cycle 5384 } 5385 } 5386 NOT_PRODUCT( 5387 if (PrintCMSStatistics != 0) { 5388 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals); 5389 } 5390 ) 5391 assert(work_q->size() == 0 && _collector->overflow_list_is_empty(), 5392 "Else our work is not yet done"); 5393 } 5394 5395 // Return a thread-local PLAB recording array, as appropriate. 5396 void* CMSCollector::get_data_recorder(int thr_num) { 5397 if (_survivor_plab_array != NULL && 5398 (CMSPLABRecordAlways || 5399 (_collectorState > Marking && _collectorState < FinalMarking))) { 5400 assert(thr_num < (int)ParallelGCThreads, "thr_num is out of bounds"); 5401 ChunkArray* ca = &_survivor_plab_array[thr_num]; 5402 ca->reset(); // clear it so that fresh data is recorded 5403 return (void*) ca; 5404 } else { 5405 return NULL; 5406 } 5407 } 5408 5409 // Reset all the thread-local PLAB recording arrays 5410 void CMSCollector::reset_survivor_plab_arrays() { 5411 for (uint i = 0; i < ParallelGCThreads; i++) { 5412 _survivor_plab_array[i].reset(); 5413 } 5414 } 5415 5416 // Merge the per-thread plab arrays into the global survivor chunk 5417 // array which will provide the partitioning of the survivor space 5418 // for CMS rescan. 5419 void CMSCollector::merge_survivor_plab_arrays(ContiguousSpace* surv, 5420 int no_of_gc_threads) { 5421 assert(_survivor_plab_array != NULL, "Error"); 5422 assert(_survivor_chunk_array != NULL, "Error"); 5423 assert(_collectorState == FinalMarking, "Error"); 5424 for (int j = 0; j < no_of_gc_threads; j++) { 5425 _cursor[j] = 0; 5426 } 5427 HeapWord* top = surv->top(); 5428 size_t i; 5429 for (i = 0; i < _survivor_chunk_capacity; i++) { // all sca entries 5430 HeapWord* min_val = top; // Higher than any PLAB address 5431 uint min_tid = 0; // position of min_val this round 5432 for (int j = 0; j < no_of_gc_threads; j++) { 5433 ChunkArray* cur_sca = &_survivor_plab_array[j]; 5434 if (_cursor[j] == cur_sca->end()) { 5435 continue; 5436 } 5437 assert(_cursor[j] < cur_sca->end(), "ctl pt invariant"); 5438 HeapWord* cur_val = cur_sca->nth(_cursor[j]); 5439 assert(surv->used_region().contains(cur_val), "Out of bounds value"); 5440 if (cur_val < min_val) { 5441 min_tid = j; 5442 min_val = cur_val; 5443 } else { 5444 assert(cur_val < top, "All recorded addresses should be less"); 5445 } 5446 } 5447 // At this point min_val and min_tid are respectively 5448 // the least address in _survivor_plab_array[j]->nth(_cursor[j]) 5449 // and the thread (j) that witnesses that address. 5450 // We record this address in the _survivor_chunk_array[i] 5451 // and increment _cursor[min_tid] prior to the next round i. 5452 if (min_val == top) { 5453 break; 5454 } 5455 _survivor_chunk_array[i] = min_val; 5456 _cursor[min_tid]++; 5457 } 5458 // We are all done; record the size of the _survivor_chunk_array 5459 _survivor_chunk_index = i; // exclusive: [0, i) 5460 if (PrintCMSStatistics > 0) { 5461 gclog_or_tty->print(" (Survivor:" SIZE_FORMAT "chunks) ", i); 5462 } 5463 // Verify that we used up all the recorded entries 5464 #ifdef ASSERT 5465 size_t total = 0; 5466 for (int j = 0; j < no_of_gc_threads; j++) { 5467 assert(_cursor[j] == _survivor_plab_array[j].end(), "Ctl pt invariant"); 5468 total += _cursor[j]; 5469 } 5470 assert(total == _survivor_chunk_index, "Ctl Pt Invariant"); 5471 // Check that the merged array is in sorted order 5472 if (total > 0) { 5473 for (size_t i = 0; i < total - 1; i++) { 5474 if (PrintCMSStatistics > 0) { 5475 gclog_or_tty->print(" (chunk" SIZE_FORMAT ":" INTPTR_FORMAT ") ", 5476 i, _survivor_chunk_array[i]); 5477 } 5478 assert(_survivor_chunk_array[i] < _survivor_chunk_array[i+1], 5479 "Not sorted"); 5480 } 5481 } 5482 #endif // ASSERT 5483 } 5484 5485 // Set up the space's par_seq_tasks structure for work claiming 5486 // for parallel rescan of young gen. 5487 // See ParRescanTask where this is currently used. 5488 void 5489 CMSCollector:: 5490 initialize_sequential_subtasks_for_young_gen_rescan(int n_threads) { 5491 assert(n_threads > 0, "Unexpected n_threads argument"); 5492 DefNewGeneration* dng = (DefNewGeneration*)_young_gen; 5493 5494 // Eden space 5495 { 5496 SequentialSubTasksDone* pst = dng->eden()->par_seq_tasks(); 5497 assert(!pst->valid(), "Clobbering existing data?"); 5498 // Each valid entry in [0, _eden_chunk_index) represents a task. 5499 size_t n_tasks = _eden_chunk_index + 1; 5500 assert(n_tasks == 1 || _eden_chunk_array != NULL, "Error"); 5501 // Sets the condition for completion of the subtask (how many threads 5502 // need to finish in order to be done). 5503 pst->set_n_threads(n_threads); 5504 pst->set_n_tasks((int)n_tasks); 5505 } 5506 5507 // Merge the survivor plab arrays into _survivor_chunk_array 5508 if (_survivor_plab_array != NULL) { 5509 merge_survivor_plab_arrays(dng->from(), n_threads); 5510 } else { 5511 assert(_survivor_chunk_index == 0, "Error"); 5512 } 5513 5514 // To space 5515 { 5516 SequentialSubTasksDone* pst = dng->to()->par_seq_tasks(); 5517 assert(!pst->valid(), "Clobbering existing data?"); 5518 // Sets the condition for completion of the subtask (how many threads 5519 // need to finish in order to be done). 5520 pst->set_n_threads(n_threads); 5521 pst->set_n_tasks(1); 5522 assert(pst->valid(), "Error"); 5523 } 5524 5525 // From space 5526 { 5527 SequentialSubTasksDone* pst = dng->from()->par_seq_tasks(); 5528 assert(!pst->valid(), "Clobbering existing data?"); 5529 size_t n_tasks = _survivor_chunk_index + 1; 5530 assert(n_tasks == 1 || _survivor_chunk_array != NULL, "Error"); 5531 // Sets the condition for completion of the subtask (how many threads 5532 // need to finish in order to be done). 5533 pst->set_n_threads(n_threads); 5534 pst->set_n_tasks((int)n_tasks); 5535 assert(pst->valid(), "Error"); 5536 } 5537 } 5538 5539 // Parallel version of remark 5540 void CMSCollector::do_remark_parallel() { 5541 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5542 FlexibleWorkGang* workers = gch->workers(); 5543 assert(workers != NULL, "Need parallel worker threads."); 5544 // Choose to use the number of GC workers most recently set 5545 // into "active_workers". If active_workers is not set, set it 5546 // to ParallelGCThreads. 5547 int n_workers = workers->active_workers(); 5548 if (n_workers == 0) { 5549 assert(n_workers > 0, "Should have been set during scavenge"); 5550 n_workers = ParallelGCThreads; 5551 workers->set_active_workers(n_workers); 5552 } 5553 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 5554 CompactibleFreeListSpace* perm_space = _permGen->cmsSpace(); 5555 5556 CMSParRemarkTask tsk(this, 5557 cms_space, perm_space, 5558 n_workers, workers, task_queues()); 5559 5560 // Set up for parallel process_strong_roots work. 5561 gch->set_par_threads(n_workers); 5562 // We won't be iterating over the cards in the card table updating 5563 // the younger_gen cards, so we shouldn't call the following else 5564 // the verification code as well as subsequent younger_refs_iterate 5565 // code would get confused. XXX 5566 // gch->rem_set()->prepare_for_younger_refs_iterate(true); // parallel 5567 5568 // The young gen rescan work will not be done as part of 5569 // process_strong_roots (which currently doesn't knw how to 5570 // parallelize such a scan), but rather will be broken up into 5571 // a set of parallel tasks (via the sampling that the [abortable] 5572 // preclean phase did of EdenSpace, plus the [two] tasks of 5573 // scanning the [two] survivor spaces. Further fine-grain 5574 // parallelization of the scanning of the survivor spaces 5575 // themselves, and of precleaning of the younger gen itself 5576 // is deferred to the future. 5577 initialize_sequential_subtasks_for_young_gen_rescan(n_workers); 5578 5579 // The dirty card rescan work is broken up into a "sequence" 5580 // of parallel tasks (per constituent space) that are dynamically 5581 // claimed by the parallel threads. 5582 cms_space->initialize_sequential_subtasks_for_rescan(n_workers); 5583 perm_space->initialize_sequential_subtasks_for_rescan(n_workers); 5584 5585 // It turns out that even when we're using 1 thread, doing the work in a 5586 // separate thread causes wide variance in run times. We can't help this 5587 // in the multi-threaded case, but we special-case n=1 here to get 5588 // repeatable measurements of the 1-thread overhead of the parallel code. 5589 if (n_workers > 1) { 5590 // Make refs discovery MT-safe, if it isn't already: it may not 5591 // necessarily be so, since it's possible that we are doing 5592 // ST marking. 5593 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), true); 5594 GenCollectedHeap::StrongRootsScope srs(gch); 5595 workers->run_task(&tsk); 5596 } else { 5597 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false); 5598 GenCollectedHeap::StrongRootsScope srs(gch); 5599 tsk.work(0); 5600 } 5601 gch->set_par_threads(0); // 0 ==> non-parallel. 5602 // restore, single-threaded for now, any preserved marks 5603 // as a result of work_q overflow 5604 restore_preserved_marks_if_any(); 5605 } 5606 5607 // Non-parallel version of remark 5608 void CMSCollector::do_remark_non_parallel() { 5609 ResourceMark rm; 5610 HandleMark hm; 5611 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5612 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false); 5613 5614 MarkRefsIntoAndScanClosure 5615 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable, 5616 &_markStack, &_revisitStack, this, 5617 false /* should_yield */, false /* not precleaning */); 5618 MarkFromDirtyCardsClosure 5619 markFromDirtyCardsClosure(this, _span, 5620 NULL, // space is set further below 5621 &_markBitMap, &_markStack, &_revisitStack, 5622 &mrias_cl); 5623 { 5624 TraceTime t("grey object rescan", PrintGCDetails, false, gclog_or_tty); 5625 // Iterate over the dirty cards, setting the corresponding bits in the 5626 // mod union table. 5627 { 5628 ModUnionClosure modUnionClosure(&_modUnionTable); 5629 _ct->ct_bs()->dirty_card_iterate( 5630 _cmsGen->used_region(), 5631 &modUnionClosure); 5632 _ct->ct_bs()->dirty_card_iterate( 5633 _permGen->used_region(), 5634 &modUnionClosure); 5635 } 5636 // Having transferred these marks into the modUnionTable, we just need 5637 // to rescan the marked objects on the dirty cards in the modUnionTable. 5638 // The initial marking may have been done during an asynchronous 5639 // collection so there may be dirty bits in the mod-union table. 5640 const int alignment = 5641 CardTableModRefBS::card_size * BitsPerWord; 5642 { 5643 // ... First handle dirty cards in CMS gen 5644 markFromDirtyCardsClosure.set_space(_cmsGen->cmsSpace()); 5645 MemRegion ur = _cmsGen->used_region(); 5646 HeapWord* lb = ur.start(); 5647 HeapWord* ub = (HeapWord*)round_to((intptr_t)ur.end(), alignment); 5648 MemRegion cms_span(lb, ub); 5649 _modUnionTable.dirty_range_iterate_clear(cms_span, 5650 &markFromDirtyCardsClosure); 5651 verify_work_stacks_empty(); 5652 if (PrintCMSStatistics != 0) { 5653 gclog_or_tty->print(" (re-scanned "SIZE_FORMAT" dirty cards in cms gen) ", 5654 markFromDirtyCardsClosure.num_dirty_cards()); 5655 } 5656 } 5657 { 5658 // .. and then repeat for dirty cards in perm gen 5659 markFromDirtyCardsClosure.set_space(_permGen->cmsSpace()); 5660 MemRegion ur = _permGen->used_region(); 5661 HeapWord* lb = ur.start(); 5662 HeapWord* ub = (HeapWord*)round_to((intptr_t)ur.end(), alignment); 5663 MemRegion perm_span(lb, ub); 5664 _modUnionTable.dirty_range_iterate_clear(perm_span, 5665 &markFromDirtyCardsClosure); 5666 verify_work_stacks_empty(); 5667 if (PrintCMSStatistics != 0) { 5668 gclog_or_tty->print(" (re-scanned "SIZE_FORMAT" dirty cards in perm gen) ", 5669 markFromDirtyCardsClosure.num_dirty_cards()); 5670 } 5671 } 5672 } 5673 if (VerifyDuringGC && 5674 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 5675 HandleMark hm; // Discard invalid handles created during verification 5676 Universe::verify(true); 5677 } 5678 { 5679 TraceTime t("root rescan", PrintGCDetails, false, gclog_or_tty); 5680 5681 verify_work_stacks_empty(); 5682 5683 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 5684 GenCollectedHeap::StrongRootsScope srs(gch); 5685 gch->gen_process_strong_roots(_cmsGen->level(), 5686 true, // younger gens as roots 5687 false, // use the local StrongRootsScope 5688 true, // collecting perm gen 5689 SharedHeap::ScanningOption(roots_scanning_options()), 5690 &mrias_cl, 5691 true, // walk code active on stacks 5692 NULL); 5693 assert(should_unload_classes() 5694 || (roots_scanning_options() & SharedHeap::SO_CodeCache), 5695 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 5696 } 5697 verify_work_stacks_empty(); 5698 // Restore evacuated mark words, if any, used for overflow list links 5699 if (!CMSOverflowEarlyRestoration) { 5700 restore_preserved_marks_if_any(); 5701 } 5702 verify_overflow_empty(); 5703 } 5704 5705 //////////////////////////////////////////////////////// 5706 // Parallel Reference Processing Task Proxy Class 5707 //////////////////////////////////////////////////////// 5708 class CMSRefProcTaskProxy: public AbstractGangTaskWOopQueues { 5709 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask; 5710 CMSCollector* _collector; 5711 CMSBitMap* _mark_bit_map; 5712 const MemRegion _span; 5713 ProcessTask& _task; 5714 5715 public: 5716 CMSRefProcTaskProxy(ProcessTask& task, 5717 CMSCollector* collector, 5718 const MemRegion& span, 5719 CMSBitMap* mark_bit_map, 5720 AbstractWorkGang* workers, 5721 OopTaskQueueSet* task_queues): 5722 // XXX Should superclass AGTWOQ also know about AWG since it knows 5723 // about the task_queues used by the AWG? Then it could initialize 5724 // the terminator() object. See 6984287. The set_for_termination() 5725 // below is a temporary band-aid for the regression in 6984287. 5726 AbstractGangTaskWOopQueues("Process referents by policy in parallel", 5727 task_queues), 5728 _task(task), 5729 _collector(collector), _span(span), _mark_bit_map(mark_bit_map) 5730 { 5731 assert(_collector->_span.equals(_span) && !_span.is_empty(), 5732 "Inconsistency in _span"); 5733 set_for_termination(workers->active_workers()); 5734 } 5735 5736 OopTaskQueueSet* task_queues() { return queues(); } 5737 5738 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 5739 5740 void do_work_steal(int i, 5741 CMSParDrainMarkingStackClosure* drain, 5742 CMSParKeepAliveClosure* keep_alive, 5743 int* seed); 5744 5745 virtual void work(uint worker_id); 5746 }; 5747 5748 void CMSRefProcTaskProxy::work(uint worker_id) { 5749 assert(_collector->_span.equals(_span), "Inconsistency in _span"); 5750 CMSParKeepAliveClosure par_keep_alive(_collector, _span, 5751 _mark_bit_map, 5752 &_collector->_revisitStack, 5753 work_queue(worker_id)); 5754 CMSParDrainMarkingStackClosure par_drain_stack(_collector, _span, 5755 _mark_bit_map, 5756 &_collector->_revisitStack, 5757 work_queue(worker_id)); 5758 CMSIsAliveClosure is_alive_closure(_span, _mark_bit_map); 5759 _task.work(worker_id, is_alive_closure, par_keep_alive, par_drain_stack); 5760 if (_task.marks_oops_alive()) { 5761 do_work_steal(worker_id, &par_drain_stack, &par_keep_alive, 5762 _collector->hash_seed(worker_id)); 5763 } 5764 assert(work_queue(worker_id)->size() == 0, "work_queue should be empty"); 5765 assert(_collector->_overflow_list == NULL, "non-empty _overflow_list"); 5766 } 5767 5768 class CMSRefEnqueueTaskProxy: public AbstractGangTask { 5769 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask; 5770 EnqueueTask& _task; 5771 5772 public: 5773 CMSRefEnqueueTaskProxy(EnqueueTask& task) 5774 : AbstractGangTask("Enqueue reference objects in parallel"), 5775 _task(task) 5776 { } 5777 5778 virtual void work(uint worker_id) 5779 { 5780 _task.work(worker_id); 5781 } 5782 }; 5783 5784 CMSParKeepAliveClosure::CMSParKeepAliveClosure(CMSCollector* collector, 5785 MemRegion span, CMSBitMap* bit_map, CMSMarkStack* revisit_stack, 5786 OopTaskQueue* work_queue): 5787 Par_KlassRememberingOopClosure(collector, NULL, revisit_stack), 5788 _span(span), 5789 _bit_map(bit_map), 5790 _work_queue(work_queue), 5791 _mark_and_push(collector, span, bit_map, revisit_stack, work_queue), 5792 _low_water_mark(MIN2((uint)(work_queue->max_elems()/4), 5793 (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads))) 5794 { } 5795 5796 // . see if we can share work_queues with ParNew? XXX 5797 void CMSRefProcTaskProxy::do_work_steal(int i, 5798 CMSParDrainMarkingStackClosure* drain, 5799 CMSParKeepAliveClosure* keep_alive, 5800 int* seed) { 5801 OopTaskQueue* work_q = work_queue(i); 5802 NOT_PRODUCT(int num_steals = 0;) 5803 oop obj_to_scan; 5804 5805 while (true) { 5806 // Completely finish any left over work from (an) earlier round(s) 5807 drain->trim_queue(0); 5808 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 5809 (size_t)ParGCDesiredObjsFromOverflowList); 5810 // Now check if there's any work in the overflow list 5811 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads, 5812 // only affects the number of attempts made to get work from the 5813 // overflow list and does not affect the number of workers. Just 5814 // pass ParallelGCThreads so this behavior is unchanged. 5815 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 5816 work_q, 5817 ParallelGCThreads)) { 5818 // Found something in global overflow list; 5819 // not yet ready to go stealing work from others. 5820 // We'd like to assert(work_q->size() != 0, ...) 5821 // because we just took work from the overflow list, 5822 // but of course we can't, since all of that might have 5823 // been already stolen from us. 5824 continue; 5825 } 5826 // Verify that we have no work before we resort to stealing 5827 assert(work_q->size() == 0, "Have work, shouldn't steal"); 5828 // Try to steal from other queues that have work 5829 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 5830 NOT_PRODUCT(num_steals++;) 5831 assert(obj_to_scan->is_oop(), "Oops, not an oop!"); 5832 assert(_mark_bit_map->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 5833 // Do scanning work 5834 obj_to_scan->oop_iterate(keep_alive); 5835 // Loop around, finish this work, and try to steal some more 5836 } else if (terminator()->offer_termination()) { 5837 break; // nirvana from the infinite cycle 5838 } 5839 } 5840 NOT_PRODUCT( 5841 if (PrintCMSStatistics != 0) { 5842 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals); 5843 } 5844 ) 5845 } 5846 5847 void CMSRefProcTaskExecutor::execute(ProcessTask& task) 5848 { 5849 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5850 FlexibleWorkGang* workers = gch->workers(); 5851 assert(workers != NULL, "Need parallel worker threads."); 5852 CMSRefProcTaskProxy rp_task(task, &_collector, 5853 _collector.ref_processor()->span(), 5854 _collector.markBitMap(), 5855 workers, _collector.task_queues()); 5856 workers->run_task(&rp_task); 5857 } 5858 5859 void CMSRefProcTaskExecutor::execute(EnqueueTask& task) 5860 { 5861 5862 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5863 FlexibleWorkGang* workers = gch->workers(); 5864 assert(workers != NULL, "Need parallel worker threads."); 5865 CMSRefEnqueueTaskProxy enq_task(task); 5866 workers->run_task(&enq_task); 5867 } 5868 5869 void CMSCollector::refProcessingWork(bool asynch, bool clear_all_soft_refs) { 5870 5871 ResourceMark rm; 5872 HandleMark hm; 5873 5874 ReferenceProcessor* rp = ref_processor(); 5875 assert(rp->span().equals(_span), "Spans should be equal"); 5876 assert(!rp->enqueuing_is_done(), "Enqueuing should not be complete"); 5877 // Process weak references. 5878 rp->setup_policy(clear_all_soft_refs); 5879 verify_work_stacks_empty(); 5880 5881 CMSKeepAliveClosure cmsKeepAliveClosure(this, _span, &_markBitMap, 5882 &_markStack, &_revisitStack, 5883 false /* !preclean */); 5884 CMSDrainMarkingStackClosure cmsDrainMarkingStackClosure(this, 5885 _span, &_markBitMap, &_markStack, 5886 &cmsKeepAliveClosure, false /* !preclean */); 5887 { 5888 TraceTime t("weak refs processing", PrintGCDetails, false, gclog_or_tty); 5889 if (rp->processing_is_mt()) { 5890 // Set the degree of MT here. If the discovery is done MT, there 5891 // may have been a different number of threads doing the discovery 5892 // and a different number of discovered lists may have Ref objects. 5893 // That is OK as long as the Reference lists are balanced (see 5894 // balance_all_queues() and balance_queues()). 5895 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5896 int active_workers = ParallelGCThreads; 5897 FlexibleWorkGang* workers = gch->workers(); 5898 if (workers != NULL) { 5899 active_workers = workers->active_workers(); 5900 // The expectation is that active_workers will have already 5901 // been set to a reasonable value. If it has not been set, 5902 // investigate. 5903 assert(active_workers > 0, "Should have been set during scavenge"); 5904 } 5905 rp->set_active_mt_degree(active_workers); 5906 CMSRefProcTaskExecutor task_executor(*this); 5907 rp->process_discovered_references(&_is_alive_closure, 5908 &cmsKeepAliveClosure, 5909 &cmsDrainMarkingStackClosure, 5910 &task_executor); 5911 } else { 5912 rp->process_discovered_references(&_is_alive_closure, 5913 &cmsKeepAliveClosure, 5914 &cmsDrainMarkingStackClosure, 5915 NULL); 5916 } 5917 verify_work_stacks_empty(); 5918 } 5919 5920 if (should_unload_classes()) { 5921 { 5922 TraceTime t("class unloading", PrintGCDetails, false, gclog_or_tty); 5923 5924 // Follow SystemDictionary roots and unload classes 5925 bool purged_class = SystemDictionary::do_unloading(&_is_alive_closure); 5926 5927 // Follow CodeCache roots and unload any methods marked for unloading 5928 CodeCache::do_unloading(&_is_alive_closure, 5929 &cmsKeepAliveClosure, 5930 purged_class); 5931 5932 cmsDrainMarkingStackClosure.do_void(); 5933 verify_work_stacks_empty(); 5934 5935 // Update subklass/sibling/implementor links in KlassKlass descendants 5936 assert(!_revisitStack.isEmpty(), "revisit stack should not be empty"); 5937 oop k; 5938 while ((k = _revisitStack.pop()) != NULL) { 5939 ((Klass*)(oopDesc*)k)->follow_weak_klass_links( 5940 &_is_alive_closure, 5941 &cmsKeepAliveClosure); 5942 } 5943 assert(!ClassUnloading || 5944 (_markStack.isEmpty() && overflow_list_is_empty()), 5945 "Should not have found new reachable objects"); 5946 assert(_revisitStack.isEmpty(), "revisit stack should have been drained"); 5947 cmsDrainMarkingStackClosure.do_void(); 5948 verify_work_stacks_empty(); 5949 } 5950 5951 { 5952 TraceTime t("scrub symbol table", PrintGCDetails, false, gclog_or_tty); 5953 // Clean up unreferenced symbols in symbol table. 5954 SymbolTable::unlink(); 5955 } 5956 } 5957 5958 if (should_unload_classes() || !JavaObjectsInPerm) { 5959 TraceTime t("scrub string table", PrintGCDetails, false, gclog_or_tty); 5960 // Now clean up stale oops in StringTable 5961 StringTable::unlink(&_is_alive_closure); 5962 } 5963 5964 verify_work_stacks_empty(); 5965 // Restore any preserved marks as a result of mark stack or 5966 // work queue overflow 5967 restore_preserved_marks_if_any(); // done single-threaded for now 5968 5969 rp->set_enqueuing_is_done(true); 5970 if (rp->processing_is_mt()) { 5971 rp->balance_all_queues(); 5972 CMSRefProcTaskExecutor task_executor(*this); 5973 rp->enqueue_discovered_references(&task_executor); 5974 } else { 5975 rp->enqueue_discovered_references(NULL); 5976 } 5977 rp->verify_no_references_recorded(); 5978 assert(!rp->discovery_enabled(), "should have been disabled"); 5979 } 5980 5981 #ifndef PRODUCT 5982 void CMSCollector::check_correct_thread_executing() { 5983 Thread* t = Thread::current(); 5984 // Only the VM thread or the CMS thread should be here. 5985 assert(t->is_ConcurrentGC_thread() || t->is_VM_thread(), 5986 "Unexpected thread type"); 5987 // If this is the vm thread, the foreground process 5988 // should not be waiting. Note that _foregroundGCIsActive is 5989 // true while the foreground collector is waiting. 5990 if (_foregroundGCShouldWait) { 5991 // We cannot be the VM thread 5992 assert(t->is_ConcurrentGC_thread(), 5993 "Should be CMS thread"); 5994 } else { 5995 // We can be the CMS thread only if we are in a stop-world 5996 // phase of CMS collection. 5997 if (t->is_ConcurrentGC_thread()) { 5998 assert(_collectorState == InitialMarking || 5999 _collectorState == FinalMarking, 6000 "Should be a stop-world phase"); 6001 // The CMS thread should be holding the CMS_token. 6002 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6003 "Potential interference with concurrently " 6004 "executing VM thread"); 6005 } 6006 } 6007 } 6008 #endif 6009 6010 void CMSCollector::sweep(bool asynch) { 6011 assert(_collectorState == Sweeping, "just checking"); 6012 check_correct_thread_executing(); 6013 verify_work_stacks_empty(); 6014 verify_overflow_empty(); 6015 increment_sweep_count(); 6016 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 6017 6018 _inter_sweep_timer.stop(); 6019 _inter_sweep_estimate.sample(_inter_sweep_timer.seconds()); 6020 size_policy()->avg_cms_free_at_sweep()->sample(_cmsGen->free()); 6021 6022 // PermGen verification support: If perm gen sweeping is disabled in 6023 // this cycle, we preserve the perm gen object "deadness" information 6024 // in the perm_gen_verify_bit_map. In order to do that we traverse 6025 // all blocks in perm gen and mark all dead objects. 6026 if (verifying() && !should_unload_classes()) { 6027 assert(perm_gen_verify_bit_map()->sizeInBits() != 0, 6028 "Should have already been allocated"); 6029 MarkDeadObjectsClosure mdo(this, _permGen->cmsSpace(), 6030 markBitMap(), perm_gen_verify_bit_map()); 6031 if (asynch) { 6032 CMSTokenSyncWithLocks ts(true, _permGen->freelistLock(), 6033 bitMapLock()); 6034 _permGen->cmsSpace()->blk_iterate(&mdo); 6035 } else { 6036 // In the case of synchronous sweep, we already have 6037 // the requisite locks/tokens. 6038 _permGen->cmsSpace()->blk_iterate(&mdo); 6039 } 6040 } 6041 6042 assert(!_intra_sweep_timer.is_active(), "Should not be active"); 6043 _intra_sweep_timer.reset(); 6044 _intra_sweep_timer.start(); 6045 if (asynch) { 6046 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 6047 CMSPhaseAccounting pa(this, "sweep", !PrintGCDetails); 6048 // First sweep the old gen then the perm gen 6049 { 6050 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(), 6051 bitMapLock()); 6052 sweepWork(_cmsGen, asynch); 6053 } 6054 6055 // Now repeat for perm gen 6056 if (should_unload_classes()) { 6057 CMSTokenSyncWithLocks ts(true, _permGen->freelistLock(), 6058 bitMapLock()); 6059 sweepWork(_permGen, asynch); 6060 } 6061 6062 // Update Universe::_heap_*_at_gc figures. 6063 // We need all the free list locks to make the abstract state 6064 // transition from Sweeping to Resetting. See detailed note 6065 // further below. 6066 { 6067 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(), 6068 _permGen->freelistLock()); 6069 // Update heap occupancy information which is used as 6070 // input to soft ref clearing policy at the next gc. 6071 Universe::update_heap_info_at_gc(); 6072 _collectorState = Resizing; 6073 } 6074 } else { 6075 // already have needed locks 6076 sweepWork(_cmsGen, asynch); 6077 6078 if (should_unload_classes()) { 6079 sweepWork(_permGen, asynch); 6080 } 6081 // Update heap occupancy information which is used as 6082 // input to soft ref clearing policy at the next gc. 6083 Universe::update_heap_info_at_gc(); 6084 _collectorState = Resizing; 6085 } 6086 verify_work_stacks_empty(); 6087 verify_overflow_empty(); 6088 6089 _intra_sweep_timer.stop(); 6090 _intra_sweep_estimate.sample(_intra_sweep_timer.seconds()); 6091 6092 _inter_sweep_timer.reset(); 6093 _inter_sweep_timer.start(); 6094 6095 // We need to use a monotonically non-deccreasing time in ms 6096 // or we will see time-warp warnings and os::javaTimeMillis() 6097 // does not guarantee monotonicity. 6098 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC; 6099 update_time_of_last_gc(now); 6100 6101 // NOTE on abstract state transitions: 6102 // Mutators allocate-live and/or mark the mod-union table dirty 6103 // based on the state of the collection. The former is done in 6104 // the interval [Marking, Sweeping] and the latter in the interval 6105 // [Marking, Sweeping). Thus the transitions into the Marking state 6106 // and out of the Sweeping state must be synchronously visible 6107 // globally to the mutators. 6108 // The transition into the Marking state happens with the world 6109 // stopped so the mutators will globally see it. Sweeping is 6110 // done asynchronously by the background collector so the transition 6111 // from the Sweeping state to the Resizing state must be done 6112 // under the freelistLock (as is the check for whether to 6113 // allocate-live and whether to dirty the mod-union table). 6114 assert(_collectorState == Resizing, "Change of collector state to" 6115 " Resizing must be done under the freelistLocks (plural)"); 6116 6117 // Now that sweeping has been completed, we clear 6118 // the incremental_collection_failed flag, 6119 // thus inviting a younger gen collection to promote into 6120 // this generation. If such a promotion may still fail, 6121 // the flag will be set again when a young collection is 6122 // attempted. 6123 GenCollectedHeap* gch = GenCollectedHeap::heap(); 6124 gch->clear_incremental_collection_failed(); // Worth retrying as fresh space may have been freed up 6125 gch->update_full_collections_completed(_collection_count_start); 6126 } 6127 6128 // FIX ME!!! Looks like this belongs in CFLSpace, with 6129 // CMSGen merely delegating to it. 6130 void ConcurrentMarkSweepGeneration::setNearLargestChunk() { 6131 double nearLargestPercent = FLSLargestBlockCoalesceProximity; 6132 HeapWord* minAddr = _cmsSpace->bottom(); 6133 HeapWord* largestAddr = 6134 (HeapWord*) _cmsSpace->dictionary()->find_largest_dict(); 6135 if (largestAddr == NULL) { 6136 // The dictionary appears to be empty. In this case 6137 // try to coalesce at the end of the heap. 6138 largestAddr = _cmsSpace->end(); 6139 } 6140 size_t largestOffset = pointer_delta(largestAddr, minAddr); 6141 size_t nearLargestOffset = 6142 (size_t)((double)largestOffset * nearLargestPercent) - MinChunkSize; 6143 if (PrintFLSStatistics != 0) { 6144 gclog_or_tty->print_cr( 6145 "CMS: Large Block: " PTR_FORMAT ";" 6146 " Proximity: " PTR_FORMAT " -> " PTR_FORMAT, 6147 largestAddr, 6148 _cmsSpace->nearLargestChunk(), minAddr + nearLargestOffset); 6149 } 6150 _cmsSpace->set_nearLargestChunk(minAddr + nearLargestOffset); 6151 } 6152 6153 bool ConcurrentMarkSweepGeneration::isNearLargestChunk(HeapWord* addr) { 6154 return addr >= _cmsSpace->nearLargestChunk(); 6155 } 6156 6157 FreeChunk* ConcurrentMarkSweepGeneration::find_chunk_at_end() { 6158 return _cmsSpace->find_chunk_at_end(); 6159 } 6160 6161 void ConcurrentMarkSweepGeneration::update_gc_stats(int current_level, 6162 bool full) { 6163 // The next lower level has been collected. Gather any statistics 6164 // that are of interest at this point. 6165 if (!full && (current_level + 1) == level()) { 6166 // Gather statistics on the young generation collection. 6167 collector()->stats().record_gc0_end(used()); 6168 } 6169 } 6170 6171 CMSAdaptiveSizePolicy* ConcurrentMarkSweepGeneration::size_policy() { 6172 GenCollectedHeap* gch = GenCollectedHeap::heap(); 6173 assert(gch->kind() == CollectedHeap::GenCollectedHeap, 6174 "Wrong type of heap"); 6175 CMSAdaptiveSizePolicy* sp = (CMSAdaptiveSizePolicy*) 6176 gch->gen_policy()->size_policy(); 6177 assert(sp->is_gc_cms_adaptive_size_policy(), 6178 "Wrong type of size policy"); 6179 return sp; 6180 } 6181 6182 void ConcurrentMarkSweepGeneration::rotate_debug_collection_type() { 6183 if (PrintGCDetails && Verbose) { 6184 gclog_or_tty->print("Rotate from %d ", _debug_collection_type); 6185 } 6186 _debug_collection_type = (CollectionTypes) (_debug_collection_type + 1); 6187 _debug_collection_type = 6188 (CollectionTypes) (_debug_collection_type % Unknown_collection_type); 6189 if (PrintGCDetails && Verbose) { 6190 gclog_or_tty->print_cr("to %d ", _debug_collection_type); 6191 } 6192 } 6193 6194 void CMSCollector::sweepWork(ConcurrentMarkSweepGeneration* gen, 6195 bool asynch) { 6196 // We iterate over the space(s) underlying this generation, 6197 // checking the mark bit map to see if the bits corresponding 6198 // to specific blocks are marked or not. Blocks that are 6199 // marked are live and are not swept up. All remaining blocks 6200 // are swept up, with coalescing on-the-fly as we sweep up 6201 // contiguous free and/or garbage blocks: 6202 // We need to ensure that the sweeper synchronizes with allocators 6203 // and stop-the-world collectors. In particular, the following 6204 // locks are used: 6205 // . CMS token: if this is held, a stop the world collection cannot occur 6206 // . freelistLock: if this is held no allocation can occur from this 6207 // generation by another thread 6208 // . bitMapLock: if this is held, no other thread can access or update 6209 // 6210 6211 // Note that we need to hold the freelistLock if we use 6212 // block iterate below; else the iterator might go awry if 6213 // a mutator (or promotion) causes block contents to change 6214 // (for instance if the allocator divvies up a block). 6215 // If we hold the free list lock, for all practical purposes 6216 // young generation GC's can't occur (they'll usually need to 6217 // promote), so we might as well prevent all young generation 6218 // GC's while we do a sweeping step. For the same reason, we might 6219 // as well take the bit map lock for the entire duration 6220 6221 // check that we hold the requisite locks 6222 assert(have_cms_token(), "Should hold cms token"); 6223 assert( (asynch && ConcurrentMarkSweepThread::cms_thread_has_cms_token()) 6224 || (!asynch && ConcurrentMarkSweepThread::vm_thread_has_cms_token()), 6225 "Should possess CMS token to sweep"); 6226 assert_lock_strong(gen->freelistLock()); 6227 assert_lock_strong(bitMapLock()); 6228 6229 assert(!_inter_sweep_timer.is_active(), "Was switched off in an outer context"); 6230 assert(_intra_sweep_timer.is_active(), "Was switched on in an outer context"); 6231 gen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()), 6232 _inter_sweep_estimate.padded_average(), 6233 _intra_sweep_estimate.padded_average()); 6234 gen->setNearLargestChunk(); 6235 6236 { 6237 SweepClosure sweepClosure(this, gen, &_markBitMap, 6238 CMSYield && asynch); 6239 gen->cmsSpace()->blk_iterate_careful(&sweepClosure); 6240 // We need to free-up/coalesce garbage/blocks from a 6241 // co-terminal free run. This is done in the SweepClosure 6242 // destructor; so, do not remove this scope, else the 6243 // end-of-sweep-census below will be off by a little bit. 6244 } 6245 gen->cmsSpace()->sweep_completed(); 6246 gen->cmsSpace()->endSweepFLCensus(sweep_count()); 6247 if (should_unload_classes()) { // unloaded classes this cycle, 6248 _concurrent_cycles_since_last_unload = 0; // ... reset count 6249 } else { // did not unload classes, 6250 _concurrent_cycles_since_last_unload++; // ... increment count 6251 } 6252 } 6253 6254 // Reset CMS data structures (for now just the marking bit map) 6255 // preparatory for the next cycle. 6256 void CMSCollector::reset(bool asynch) { 6257 GenCollectedHeap* gch = GenCollectedHeap::heap(); 6258 CMSAdaptiveSizePolicy* sp = size_policy(); 6259 AdaptiveSizePolicyOutput(sp, gch->total_collections()); 6260 if (asynch) { 6261 CMSTokenSyncWithLocks ts(true, bitMapLock()); 6262 6263 // If the state is not "Resetting", the foreground thread 6264 // has done a collection and the resetting. 6265 if (_collectorState != Resetting) { 6266 assert(_collectorState == Idling, "The state should only change" 6267 " because the foreground collector has finished the collection"); 6268 return; 6269 } 6270 6271 // Clear the mark bitmap (no grey objects to start with) 6272 // for the next cycle. 6273 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 6274 CMSPhaseAccounting cmspa(this, "reset", !PrintGCDetails); 6275 6276 HeapWord* curAddr = _markBitMap.startWord(); 6277 while (curAddr < _markBitMap.endWord()) { 6278 size_t remaining = pointer_delta(_markBitMap.endWord(), curAddr); 6279 MemRegion chunk(curAddr, MIN2(CMSBitMapYieldQuantum, remaining)); 6280 _markBitMap.clear_large_range(chunk); 6281 if (ConcurrentMarkSweepThread::should_yield() && 6282 !foregroundGCIsActive() && 6283 CMSYield) { 6284 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6285 "CMS thread should hold CMS token"); 6286 assert_lock_strong(bitMapLock()); 6287 bitMapLock()->unlock(); 6288 ConcurrentMarkSweepThread::desynchronize(true); 6289 ConcurrentMarkSweepThread::acknowledge_yield_request(); 6290 stopTimer(); 6291 if (PrintCMSStatistics != 0) { 6292 incrementYields(); 6293 } 6294 icms_wait(); 6295 6296 // See the comment in coordinator_yield() 6297 for (unsigned i = 0; i < CMSYieldSleepCount && 6298 ConcurrentMarkSweepThread::should_yield() && 6299 !CMSCollector::foregroundGCIsActive(); ++i) { 6300 os::sleep(Thread::current(), 1, false); 6301 ConcurrentMarkSweepThread::acknowledge_yield_request(); 6302 } 6303 6304 ConcurrentMarkSweepThread::synchronize(true); 6305 bitMapLock()->lock_without_safepoint_check(); 6306 startTimer(); 6307 } 6308 curAddr = chunk.end(); 6309 } 6310 // A successful mostly concurrent collection has been done. 6311 // Because only the full (i.e., concurrent mode failure) collections 6312 // are being measured for gc overhead limits, clean the "near" flag 6313 // and count. 6314 sp->reset_gc_overhead_limit_count(); 6315 _collectorState = Idling; 6316 } else { 6317 // already have the lock 6318 assert(_collectorState == Resetting, "just checking"); 6319 assert_lock_strong(bitMapLock()); 6320 _markBitMap.clear_all(); 6321 _collectorState = Idling; 6322 } 6323 6324 // Stop incremental mode after a cycle completes, so that any future cycles 6325 // are triggered by allocation. 6326 stop_icms(); 6327 6328 NOT_PRODUCT( 6329 if (RotateCMSCollectionTypes) { 6330 _cmsGen->rotate_debug_collection_type(); 6331 } 6332 ) 6333 } 6334 6335 void CMSCollector::do_CMS_operation(CMS_op_type op, GCCause::Cause gc_cause) { 6336 gclog_or_tty->date_stamp(PrintGC && PrintGCDateStamps); 6337 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 6338 TraceTime t(GCCauseString("GC", gc_cause), PrintGC, !PrintGCDetails, gclog_or_tty); 6339 TraceCollectorStats tcs(counters()); 6340 6341 switch (op) { 6342 case CMS_op_checkpointRootsInitial: { 6343 SvcGCMarker sgcm(SvcGCMarker::OTHER); 6344 checkpointRootsInitial(true); // asynch 6345 if (PrintGC) { 6346 _cmsGen->printOccupancy("initial-mark"); 6347 } 6348 break; 6349 } 6350 case CMS_op_checkpointRootsFinal: { 6351 SvcGCMarker sgcm(SvcGCMarker::OTHER); 6352 checkpointRootsFinal(true, // asynch 6353 false, // !clear_all_soft_refs 6354 false); // !init_mark_was_synchronous 6355 if (PrintGC) { 6356 _cmsGen->printOccupancy("remark"); 6357 } 6358 break; 6359 } 6360 default: 6361 fatal("No such CMS_op"); 6362 } 6363 } 6364 6365 #ifndef PRODUCT 6366 size_t const CMSCollector::skip_header_HeapWords() { 6367 return FreeChunk::header_size(); 6368 } 6369 6370 // Try and collect here conditions that should hold when 6371 // CMS thread is exiting. The idea is that the foreground GC 6372 // thread should not be blocked if it wants to terminate 6373 // the CMS thread and yet continue to run the VM for a while 6374 // after that. 6375 void CMSCollector::verify_ok_to_terminate() const { 6376 assert(Thread::current()->is_ConcurrentGC_thread(), 6377 "should be called by CMS thread"); 6378 assert(!_foregroundGCShouldWait, "should be false"); 6379 // We could check here that all the various low-level locks 6380 // are not held by the CMS thread, but that is overkill; see 6381 // also CMSThread::verify_ok_to_terminate() where the CGC_lock 6382 // is checked. 6383 } 6384 #endif 6385 6386 size_t CMSCollector::block_size_using_printezis_bits(HeapWord* addr) const { 6387 assert(_markBitMap.isMarked(addr) && _markBitMap.isMarked(addr + 1), 6388 "missing Printezis mark?"); 6389 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 6390 size_t size = pointer_delta(nextOneAddr + 1, addr); 6391 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 6392 "alignment problem"); 6393 assert(size >= 3, "Necessary for Printezis marks to work"); 6394 return size; 6395 } 6396 6397 // A variant of the above (block_size_using_printezis_bits()) except 6398 // that we return 0 if the P-bits are not yet set. 6399 size_t CMSCollector::block_size_if_printezis_bits(HeapWord* addr) const { 6400 if (_markBitMap.isMarked(addr + 1)) { 6401 assert(_markBitMap.isMarked(addr), "P-bit can be set only for marked objects"); 6402 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 6403 size_t size = pointer_delta(nextOneAddr + 1, addr); 6404 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 6405 "alignment problem"); 6406 assert(size >= 3, "Necessary for Printezis marks to work"); 6407 return size; 6408 } 6409 return 0; 6410 } 6411 6412 HeapWord* CMSCollector::next_card_start_after_block(HeapWord* addr) const { 6413 size_t sz = 0; 6414 oop p = (oop)addr; 6415 if (p->klass_or_null() != NULL && p->is_parsable()) { 6416 sz = CompactibleFreeListSpace::adjustObjectSize(p->size()); 6417 } else { 6418 sz = block_size_using_printezis_bits(addr); 6419 } 6420 assert(sz > 0, "size must be nonzero"); 6421 HeapWord* next_block = addr + sz; 6422 HeapWord* next_card = (HeapWord*)round_to((uintptr_t)next_block, 6423 CardTableModRefBS::card_size); 6424 assert(round_down((uintptr_t)addr, CardTableModRefBS::card_size) < 6425 round_down((uintptr_t)next_card, CardTableModRefBS::card_size), 6426 "must be different cards"); 6427 return next_card; 6428 } 6429 6430 6431 // CMS Bit Map Wrapper ///////////////////////////////////////// 6432 6433 // Construct a CMS bit map infrastructure, but don't create the 6434 // bit vector itself. That is done by a separate call CMSBitMap::allocate() 6435 // further below. 6436 CMSBitMap::CMSBitMap(int shifter, int mutex_rank, const char* mutex_name): 6437 _bm(), 6438 _shifter(shifter), 6439 _lock(mutex_rank >= 0 ? new Mutex(mutex_rank, mutex_name, true) : NULL) 6440 { 6441 _bmStartWord = 0; 6442 _bmWordSize = 0; 6443 } 6444 6445 bool CMSBitMap::allocate(MemRegion mr) { 6446 _bmStartWord = mr.start(); 6447 _bmWordSize = mr.word_size(); 6448 ReservedSpace brs(ReservedSpace::allocation_align_size_up( 6449 (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1)); 6450 if (!brs.is_reserved()) { 6451 warning("CMS bit map allocation failure"); 6452 return false; 6453 } 6454 // For now we'll just commit all of the bit map up fromt. 6455 // Later on we'll try to be more parsimonious with swap. 6456 if (!_virtual_space.initialize(brs, brs.size())) { 6457 warning("CMS bit map backing store failure"); 6458 return false; 6459 } 6460 assert(_virtual_space.committed_size() == brs.size(), 6461 "didn't reserve backing store for all of CMS bit map?"); 6462 _bm.set_map((BitMap::bm_word_t*)_virtual_space.low()); 6463 assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >= 6464 _bmWordSize, "inconsistency in bit map sizing"); 6465 _bm.set_size(_bmWordSize >> _shifter); 6466 6467 // bm.clear(); // can we rely on getting zero'd memory? verify below 6468 assert(isAllClear(), 6469 "Expected zero'd memory from ReservedSpace constructor"); 6470 assert(_bm.size() == heapWordDiffToOffsetDiff(sizeInWords()), 6471 "consistency check"); 6472 return true; 6473 } 6474 6475 void CMSBitMap::dirty_range_iterate_clear(MemRegion mr, MemRegionClosure* cl) { 6476 HeapWord *next_addr, *end_addr, *last_addr; 6477 assert_locked(); 6478 assert(covers(mr), "out-of-range error"); 6479 // XXX assert that start and end are appropriately aligned 6480 for (next_addr = mr.start(), end_addr = mr.end(); 6481 next_addr < end_addr; next_addr = last_addr) { 6482 MemRegion dirty_region = getAndClearMarkedRegion(next_addr, end_addr); 6483 last_addr = dirty_region.end(); 6484 if (!dirty_region.is_empty()) { 6485 cl->do_MemRegion(dirty_region); 6486 } else { 6487 assert(last_addr == end_addr, "program logic"); 6488 return; 6489 } 6490 } 6491 } 6492 6493 #ifndef PRODUCT 6494 void CMSBitMap::assert_locked() const { 6495 CMSLockVerifier::assert_locked(lock()); 6496 } 6497 6498 bool CMSBitMap::covers(MemRegion mr) const { 6499 // assert(_bm.map() == _virtual_space.low(), "map inconsistency"); 6500 assert((size_t)_bm.size() == (_bmWordSize >> _shifter), 6501 "size inconsistency"); 6502 return (mr.start() >= _bmStartWord) && 6503 (mr.end() <= endWord()); 6504 } 6505 6506 bool CMSBitMap::covers(HeapWord* start, size_t size) const { 6507 return (start >= _bmStartWord && (start + size) <= endWord()); 6508 } 6509 6510 void CMSBitMap::verifyNoOneBitsInRange(HeapWord* left, HeapWord* right) { 6511 // verify that there are no 1 bits in the interval [left, right) 6512 FalseBitMapClosure falseBitMapClosure; 6513 iterate(&falseBitMapClosure, left, right); 6514 } 6515 6516 void CMSBitMap::region_invariant(MemRegion mr) 6517 { 6518 assert_locked(); 6519 // mr = mr.intersection(MemRegion(_bmStartWord, _bmWordSize)); 6520 assert(!mr.is_empty(), "unexpected empty region"); 6521 assert(covers(mr), "mr should be covered by bit map"); 6522 // convert address range into offset range 6523 size_t start_ofs = heapWordToOffset(mr.start()); 6524 // Make sure that end() is appropriately aligned 6525 assert(mr.end() == (HeapWord*)round_to((intptr_t)mr.end(), 6526 (1 << (_shifter+LogHeapWordSize))), 6527 "Misaligned mr.end()"); 6528 size_t end_ofs = heapWordToOffset(mr.end()); 6529 assert(end_ofs > start_ofs, "Should mark at least one bit"); 6530 } 6531 6532 #endif 6533 6534 bool CMSMarkStack::allocate(size_t size) { 6535 // allocate a stack of the requisite depth 6536 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 6537 size * sizeof(oop))); 6538 if (!rs.is_reserved()) { 6539 warning("CMSMarkStack allocation failure"); 6540 return false; 6541 } 6542 if (!_virtual_space.initialize(rs, rs.size())) { 6543 warning("CMSMarkStack backing store failure"); 6544 return false; 6545 } 6546 assert(_virtual_space.committed_size() == rs.size(), 6547 "didn't reserve backing store for all of CMS stack?"); 6548 _base = (oop*)(_virtual_space.low()); 6549 _index = 0; 6550 _capacity = size; 6551 NOT_PRODUCT(_max_depth = 0); 6552 return true; 6553 } 6554 6555 // XXX FIX ME !!! In the MT case we come in here holding a 6556 // leaf lock. For printing we need to take a further lock 6557 // which has lower rank. We need to recallibrate the two 6558 // lock-ranks involved in order to be able to rpint the 6559 // messages below. (Or defer the printing to the caller. 6560 // For now we take the expedient path of just disabling the 6561 // messages for the problematic case.) 6562 void CMSMarkStack::expand() { 6563 assert(_capacity <= MarkStackSizeMax, "stack bigger than permitted"); 6564 if (_capacity == MarkStackSizeMax) { 6565 if (_hit_limit++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) { 6566 // We print a warning message only once per CMS cycle. 6567 gclog_or_tty->print_cr(" (benign) Hit CMSMarkStack max size limit"); 6568 } 6569 return; 6570 } 6571 // Double capacity if possible 6572 size_t new_capacity = MIN2(_capacity*2, MarkStackSizeMax); 6573 // Do not give up existing stack until we have managed to 6574 // get the double capacity that we desired. 6575 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 6576 new_capacity * sizeof(oop))); 6577 if (rs.is_reserved()) { 6578 // Release the backing store associated with old stack 6579 _virtual_space.release(); 6580 // Reinitialize virtual space for new stack 6581 if (!_virtual_space.initialize(rs, rs.size())) { 6582 fatal("Not enough swap for expanded marking stack"); 6583 } 6584 _base = (oop*)(_virtual_space.low()); 6585 _index = 0; 6586 _capacity = new_capacity; 6587 } else if (_failed_double++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) { 6588 // Failed to double capacity, continue; 6589 // we print a detail message only once per CMS cycle. 6590 gclog_or_tty->print(" (benign) Failed to expand marking stack from "SIZE_FORMAT"K to " 6591 SIZE_FORMAT"K", 6592 _capacity / K, new_capacity / K); 6593 } 6594 } 6595 6596 6597 // Closures 6598 // XXX: there seems to be a lot of code duplication here; 6599 // should refactor and consolidate common code. 6600 6601 // This closure is used to mark refs into the CMS generation in 6602 // the CMS bit map. Called at the first checkpoint. This closure 6603 // assumes that we do not need to re-mark dirty cards; if the CMS 6604 // generation on which this is used is not an oldest (modulo perm gen) 6605 // generation then this will lose younger_gen cards! 6606 6607 MarkRefsIntoClosure::MarkRefsIntoClosure( 6608 MemRegion span, CMSBitMap* bitMap): 6609 _span(span), 6610 _bitMap(bitMap) 6611 { 6612 assert(_ref_processor == NULL, "deliberately left NULL"); 6613 assert(_bitMap->covers(_span), "_bitMap/_span mismatch"); 6614 } 6615 6616 void MarkRefsIntoClosure::do_oop(oop obj) { 6617 // if p points into _span, then mark corresponding bit in _markBitMap 6618 assert(obj->is_oop(), "expected an oop"); 6619 HeapWord* addr = (HeapWord*)obj; 6620 if (_span.contains(addr)) { 6621 // this should be made more efficient 6622 _bitMap->mark(addr); 6623 } 6624 } 6625 6626 void MarkRefsIntoClosure::do_oop(oop* p) { MarkRefsIntoClosure::do_oop_work(p); } 6627 void MarkRefsIntoClosure::do_oop(narrowOop* p) { MarkRefsIntoClosure::do_oop_work(p); } 6628 6629 // A variant of the above, used for CMS marking verification. 6630 MarkRefsIntoVerifyClosure::MarkRefsIntoVerifyClosure( 6631 MemRegion span, CMSBitMap* verification_bm, CMSBitMap* cms_bm): 6632 _span(span), 6633 _verification_bm(verification_bm), 6634 _cms_bm(cms_bm) 6635 { 6636 assert(_ref_processor == NULL, "deliberately left NULL"); 6637 assert(_verification_bm->covers(_span), "_verification_bm/_span mismatch"); 6638 } 6639 6640 void MarkRefsIntoVerifyClosure::do_oop(oop obj) { 6641 // if p points into _span, then mark corresponding bit in _markBitMap 6642 assert(obj->is_oop(), "expected an oop"); 6643 HeapWord* addr = (HeapWord*)obj; 6644 if (_span.contains(addr)) { 6645 _verification_bm->mark(addr); 6646 if (!_cms_bm->isMarked(addr)) { 6647 oop(addr)->print(); 6648 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)", addr); 6649 fatal("... aborting"); 6650 } 6651 } 6652 } 6653 6654 void MarkRefsIntoVerifyClosure::do_oop(oop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); } 6655 void MarkRefsIntoVerifyClosure::do_oop(narrowOop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); } 6656 6657 ////////////////////////////////////////////////// 6658 // MarkRefsIntoAndScanClosure 6659 ////////////////////////////////////////////////// 6660 6661 MarkRefsIntoAndScanClosure::MarkRefsIntoAndScanClosure(MemRegion span, 6662 ReferenceProcessor* rp, 6663 CMSBitMap* bit_map, 6664 CMSBitMap* mod_union_table, 6665 CMSMarkStack* mark_stack, 6666 CMSMarkStack* revisit_stack, 6667 CMSCollector* collector, 6668 bool should_yield, 6669 bool concurrent_precleaning): 6670 _collector(collector), 6671 _span(span), 6672 _bit_map(bit_map), 6673 _mark_stack(mark_stack), 6674 _pushAndMarkClosure(collector, span, rp, bit_map, mod_union_table, 6675 mark_stack, revisit_stack, concurrent_precleaning), 6676 _yield(should_yield), 6677 _concurrent_precleaning(concurrent_precleaning), 6678 _freelistLock(NULL) 6679 { 6680 _ref_processor = rp; 6681 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 6682 } 6683 6684 // This closure is used to mark refs into the CMS generation at the 6685 // second (final) checkpoint, and to scan and transitively follow 6686 // the unmarked oops. It is also used during the concurrent precleaning 6687 // phase while scanning objects on dirty cards in the CMS generation. 6688 // The marks are made in the marking bit map and the marking stack is 6689 // used for keeping the (newly) grey objects during the scan. 6690 // The parallel version (Par_...) appears further below. 6691 void MarkRefsIntoAndScanClosure::do_oop(oop obj) { 6692 if (obj != NULL) { 6693 assert(obj->is_oop(), "expected an oop"); 6694 HeapWord* addr = (HeapWord*)obj; 6695 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 6696 assert(_collector->overflow_list_is_empty(), 6697 "overflow list should be empty"); 6698 if (_span.contains(addr) && 6699 !_bit_map->isMarked(addr)) { 6700 // mark bit map (object is now grey) 6701 _bit_map->mark(addr); 6702 // push on marking stack (stack should be empty), and drain the 6703 // stack by applying this closure to the oops in the oops popped 6704 // from the stack (i.e. blacken the grey objects) 6705 bool res = _mark_stack->push(obj); 6706 assert(res, "Should have space to push on empty stack"); 6707 do { 6708 oop new_oop = _mark_stack->pop(); 6709 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 6710 assert(new_oop->is_parsable(), "Found unparsable oop"); 6711 assert(_bit_map->isMarked((HeapWord*)new_oop), 6712 "only grey objects on this stack"); 6713 // iterate over the oops in this oop, marking and pushing 6714 // the ones in CMS heap (i.e. in _span). 6715 new_oop->oop_iterate(&_pushAndMarkClosure); 6716 // check if it's time to yield 6717 do_yield_check(); 6718 } while (!_mark_stack->isEmpty() || 6719 (!_concurrent_precleaning && take_from_overflow_list())); 6720 // if marking stack is empty, and we are not doing this 6721 // during precleaning, then check the overflow list 6722 } 6723 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 6724 assert(_collector->overflow_list_is_empty(), 6725 "overflow list was drained above"); 6726 // We could restore evacuated mark words, if any, used for 6727 // overflow list links here because the overflow list is 6728 // provably empty here. That would reduce the maximum 6729 // size requirements for preserved_{oop,mark}_stack. 6730 // But we'll just postpone it until we are all done 6731 // so we can just stream through. 6732 if (!_concurrent_precleaning && CMSOverflowEarlyRestoration) { 6733 _collector->restore_preserved_marks_if_any(); 6734 assert(_collector->no_preserved_marks(), "No preserved marks"); 6735 } 6736 assert(!CMSOverflowEarlyRestoration || _collector->no_preserved_marks(), 6737 "All preserved marks should have been restored above"); 6738 } 6739 } 6740 6741 void MarkRefsIntoAndScanClosure::do_oop(oop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); } 6742 void MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); } 6743 6744 void MarkRefsIntoAndScanClosure::do_yield_work() { 6745 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6746 "CMS thread should hold CMS token"); 6747 assert_lock_strong(_freelistLock); 6748 assert_lock_strong(_bit_map->lock()); 6749 // relinquish the free_list_lock and bitMaplock() 6750 DEBUG_ONLY(RememberKlassesChecker mux(false);) 6751 _bit_map->lock()->unlock(); 6752 _freelistLock->unlock(); 6753 ConcurrentMarkSweepThread::desynchronize(true); 6754 ConcurrentMarkSweepThread::acknowledge_yield_request(); 6755 _collector->stopTimer(); 6756 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 6757 if (PrintCMSStatistics != 0) { 6758 _collector->incrementYields(); 6759 } 6760 _collector->icms_wait(); 6761 6762 // See the comment in coordinator_yield() 6763 for (unsigned i = 0; 6764 i < CMSYieldSleepCount && 6765 ConcurrentMarkSweepThread::should_yield() && 6766 !CMSCollector::foregroundGCIsActive(); 6767 ++i) { 6768 os::sleep(Thread::current(), 1, false); 6769 ConcurrentMarkSweepThread::acknowledge_yield_request(); 6770 } 6771 6772 ConcurrentMarkSweepThread::synchronize(true); 6773 _freelistLock->lock_without_safepoint_check(); 6774 _bit_map->lock()->lock_without_safepoint_check(); 6775 _collector->startTimer(); 6776 } 6777 6778 /////////////////////////////////////////////////////////// 6779 // Par_MarkRefsIntoAndScanClosure: a parallel version of 6780 // MarkRefsIntoAndScanClosure 6781 /////////////////////////////////////////////////////////// 6782 Par_MarkRefsIntoAndScanClosure::Par_MarkRefsIntoAndScanClosure( 6783 CMSCollector* collector, MemRegion span, ReferenceProcessor* rp, 6784 CMSBitMap* bit_map, OopTaskQueue* work_queue, CMSMarkStack* revisit_stack): 6785 _span(span), 6786 _bit_map(bit_map), 6787 _work_queue(work_queue), 6788 _low_water_mark(MIN2((uint)(work_queue->max_elems()/4), 6789 (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads))), 6790 _par_pushAndMarkClosure(collector, span, rp, bit_map, work_queue, 6791 revisit_stack) 6792 { 6793 _ref_processor = rp; 6794 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 6795 } 6796 6797 // This closure is used to mark refs into the CMS generation at the 6798 // second (final) checkpoint, and to scan and transitively follow 6799 // the unmarked oops. The marks are made in the marking bit map and 6800 // the work_queue is used for keeping the (newly) grey objects during 6801 // the scan phase whence they are also available for stealing by parallel 6802 // threads. Since the marking bit map is shared, updates are 6803 // synchronized (via CAS). 6804 void Par_MarkRefsIntoAndScanClosure::do_oop(oop obj) { 6805 if (obj != NULL) { 6806 // Ignore mark word because this could be an already marked oop 6807 // that may be chained at the end of the overflow list. 6808 assert(obj->is_oop(true), "expected an oop"); 6809 HeapWord* addr = (HeapWord*)obj; 6810 if (_span.contains(addr) && 6811 !_bit_map->isMarked(addr)) { 6812 // mark bit map (object will become grey): 6813 // It is possible for several threads to be 6814 // trying to "claim" this object concurrently; 6815 // the unique thread that succeeds in marking the 6816 // object first will do the subsequent push on 6817 // to the work queue (or overflow list). 6818 if (_bit_map->par_mark(addr)) { 6819 // push on work_queue (which may not be empty), and trim the 6820 // queue to an appropriate length by applying this closure to 6821 // the oops in the oops popped from the stack (i.e. blacken the 6822 // grey objects) 6823 bool res = _work_queue->push(obj); 6824 assert(res, "Low water mark should be less than capacity?"); 6825 trim_queue(_low_water_mark); 6826 } // Else, another thread claimed the object 6827 } 6828 } 6829 } 6830 6831 void Par_MarkRefsIntoAndScanClosure::do_oop(oop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); } 6832 void Par_MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); } 6833 6834 // This closure is used to rescan the marked objects on the dirty cards 6835 // in the mod union table and the card table proper. 6836 size_t ScanMarkedObjectsAgainCarefullyClosure::do_object_careful_m( 6837 oop p, MemRegion mr) { 6838 6839 size_t size = 0; 6840 HeapWord* addr = (HeapWord*)p; 6841 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6842 assert(_span.contains(addr), "we are scanning the CMS generation"); 6843 // check if it's time to yield 6844 if (do_yield_check()) { 6845 // We yielded for some foreground stop-world work, 6846 // and we have been asked to abort this ongoing preclean cycle. 6847 return 0; 6848 } 6849 if (_bitMap->isMarked(addr)) { 6850 // it's marked; is it potentially uninitialized? 6851 if (p->klass_or_null() != NULL) { 6852 // If is_conc_safe is false, the object may be undergoing 6853 // change by the VM outside a safepoint. Don't try to 6854 // scan it, but rather leave it for the remark phase. 6855 if (CMSPermGenPrecleaningEnabled && 6856 (!p->is_conc_safe() || !p->is_parsable())) { 6857 // Signal precleaning to redirty the card since 6858 // the klass pointer is already installed. 6859 assert(size == 0, "Initial value"); 6860 } else { 6861 assert(p->is_parsable(), "must be parsable."); 6862 // an initialized object; ignore mark word in verification below 6863 // since we are running concurrent with mutators 6864 assert(p->is_oop(true), "should be an oop"); 6865 if (p->is_objArray()) { 6866 // objArrays are precisely marked; restrict scanning 6867 // to dirty cards only. 6868 size = CompactibleFreeListSpace::adjustObjectSize( 6869 p->oop_iterate(_scanningClosure, mr)); 6870 } else { 6871 // A non-array may have been imprecisely marked; we need 6872 // to scan object in its entirety. 6873 size = CompactibleFreeListSpace::adjustObjectSize( 6874 p->oop_iterate(_scanningClosure)); 6875 } 6876 #ifdef DEBUG 6877 size_t direct_size = 6878 CompactibleFreeListSpace::adjustObjectSize(p->size()); 6879 assert(size == direct_size, "Inconsistency in size"); 6880 assert(size >= 3, "Necessary for Printezis marks to work"); 6881 if (!_bitMap->isMarked(addr+1)) { 6882 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size); 6883 } else { 6884 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size-1); 6885 assert(_bitMap->isMarked(addr+size-1), 6886 "inconsistent Printezis mark"); 6887 } 6888 #endif // DEBUG 6889 } 6890 } else { 6891 // an unitialized object 6892 assert(_bitMap->isMarked(addr+1), "missing Printezis mark?"); 6893 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 6894 size = pointer_delta(nextOneAddr + 1, addr); 6895 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 6896 "alignment problem"); 6897 // Note that pre-cleaning needn't redirty the card. OopDesc::set_klass() 6898 // will dirty the card when the klass pointer is installed in the 6899 // object (signalling the completion of initialization). 6900 } 6901 } else { 6902 // Either a not yet marked object or an uninitialized object 6903 if (p->klass_or_null() == NULL || !p->is_parsable()) { 6904 // An uninitialized object, skip to the next card, since 6905 // we may not be able to read its P-bits yet. 6906 assert(size == 0, "Initial value"); 6907 } else { 6908 // An object not (yet) reached by marking: we merely need to 6909 // compute its size so as to go look at the next block. 6910 assert(p->is_oop(true), "should be an oop"); 6911 size = CompactibleFreeListSpace::adjustObjectSize(p->size()); 6912 } 6913 } 6914 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6915 return size; 6916 } 6917 6918 void ScanMarkedObjectsAgainCarefullyClosure::do_yield_work() { 6919 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6920 "CMS thread should hold CMS token"); 6921 assert_lock_strong(_freelistLock); 6922 assert_lock_strong(_bitMap->lock()); 6923 DEBUG_ONLY(RememberKlassesChecker mux(false);) 6924 // relinquish the free_list_lock and bitMaplock() 6925 _bitMap->lock()->unlock(); 6926 _freelistLock->unlock(); 6927 ConcurrentMarkSweepThread::desynchronize(true); 6928 ConcurrentMarkSweepThread::acknowledge_yield_request(); 6929 _collector->stopTimer(); 6930 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 6931 if (PrintCMSStatistics != 0) { 6932 _collector->incrementYields(); 6933 } 6934 _collector->icms_wait(); 6935 6936 // See the comment in coordinator_yield() 6937 for (unsigned i = 0; i < CMSYieldSleepCount && 6938 ConcurrentMarkSweepThread::should_yield() && 6939 !CMSCollector::foregroundGCIsActive(); ++i) { 6940 os::sleep(Thread::current(), 1, false); 6941 ConcurrentMarkSweepThread::acknowledge_yield_request(); 6942 } 6943 6944 ConcurrentMarkSweepThread::synchronize(true); 6945 _freelistLock->lock_without_safepoint_check(); 6946 _bitMap->lock()->lock_without_safepoint_check(); 6947 _collector->startTimer(); 6948 } 6949 6950 6951 ////////////////////////////////////////////////////////////////// 6952 // SurvivorSpacePrecleanClosure 6953 ////////////////////////////////////////////////////////////////// 6954 // This (single-threaded) closure is used to preclean the oops in 6955 // the survivor spaces. 6956 size_t SurvivorSpacePrecleanClosure::do_object_careful(oop p) { 6957 6958 HeapWord* addr = (HeapWord*)p; 6959 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6960 assert(!_span.contains(addr), "we are scanning the survivor spaces"); 6961 assert(p->klass_or_null() != NULL, "object should be initializd"); 6962 assert(p->is_parsable(), "must be parsable."); 6963 // an initialized object; ignore mark word in verification below 6964 // since we are running concurrent with mutators 6965 assert(p->is_oop(true), "should be an oop"); 6966 // Note that we do not yield while we iterate over 6967 // the interior oops of p, pushing the relevant ones 6968 // on our marking stack. 6969 size_t size = p->oop_iterate(_scanning_closure); 6970 do_yield_check(); 6971 // Observe that below, we do not abandon the preclean 6972 // phase as soon as we should; rather we empty the 6973 // marking stack before returning. This is to satisfy 6974 // some existing assertions. In general, it may be a 6975 // good idea to abort immediately and complete the marking 6976 // from the grey objects at a later time. 6977 while (!_mark_stack->isEmpty()) { 6978 oop new_oop = _mark_stack->pop(); 6979 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 6980 assert(new_oop->is_parsable(), "Found unparsable oop"); 6981 assert(_bit_map->isMarked((HeapWord*)new_oop), 6982 "only grey objects on this stack"); 6983 // iterate over the oops in this oop, marking and pushing 6984 // the ones in CMS heap (i.e. in _span). 6985 new_oop->oop_iterate(_scanning_closure); 6986 // check if it's time to yield 6987 do_yield_check(); 6988 } 6989 unsigned int after_count = 6990 GenCollectedHeap::heap()->total_collections(); 6991 bool abort = (_before_count != after_count) || 6992 _collector->should_abort_preclean(); 6993 return abort ? 0 : size; 6994 } 6995 6996 void SurvivorSpacePrecleanClosure::do_yield_work() { 6997 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6998 "CMS thread should hold CMS token"); 6999 assert_lock_strong(_bit_map->lock()); 7000 DEBUG_ONLY(RememberKlassesChecker smx(false);) 7001 // Relinquish the bit map lock 7002 _bit_map->lock()->unlock(); 7003 ConcurrentMarkSweepThread::desynchronize(true); 7004 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7005 _collector->stopTimer(); 7006 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 7007 if (PrintCMSStatistics != 0) { 7008 _collector->incrementYields(); 7009 } 7010 _collector->icms_wait(); 7011 7012 // See the comment in coordinator_yield() 7013 for (unsigned i = 0; i < CMSYieldSleepCount && 7014 ConcurrentMarkSweepThread::should_yield() && 7015 !CMSCollector::foregroundGCIsActive(); ++i) { 7016 os::sleep(Thread::current(), 1, false); 7017 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7018 } 7019 7020 ConcurrentMarkSweepThread::synchronize(true); 7021 _bit_map->lock()->lock_without_safepoint_check(); 7022 _collector->startTimer(); 7023 } 7024 7025 // This closure is used to rescan the marked objects on the dirty cards 7026 // in the mod union table and the card table proper. In the parallel 7027 // case, although the bitMap is shared, we do a single read so the 7028 // isMarked() query is "safe". 7029 bool ScanMarkedObjectsAgainClosure::do_object_bm(oop p, MemRegion mr) { 7030 // Ignore mark word because we are running concurrent with mutators 7031 assert(p->is_oop_or_null(true), "expected an oop or null"); 7032 HeapWord* addr = (HeapWord*)p; 7033 assert(_span.contains(addr), "we are scanning the CMS generation"); 7034 bool is_obj_array = false; 7035 #ifdef DEBUG 7036 if (!_parallel) { 7037 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 7038 assert(_collector->overflow_list_is_empty(), 7039 "overflow list should be empty"); 7040 7041 } 7042 #endif // DEBUG 7043 if (_bit_map->isMarked(addr)) { 7044 // Obj arrays are precisely marked, non-arrays are not; 7045 // so we scan objArrays precisely and non-arrays in their 7046 // entirety. 7047 if (p->is_objArray()) { 7048 is_obj_array = true; 7049 if (_parallel) { 7050 p->oop_iterate(_par_scan_closure, mr); 7051 } else { 7052 p->oop_iterate(_scan_closure, mr); 7053 } 7054 } else { 7055 if (_parallel) { 7056 p->oop_iterate(_par_scan_closure); 7057 } else { 7058 p->oop_iterate(_scan_closure); 7059 } 7060 } 7061 } 7062 #ifdef DEBUG 7063 if (!_parallel) { 7064 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 7065 assert(_collector->overflow_list_is_empty(), 7066 "overflow list should be empty"); 7067 7068 } 7069 #endif // DEBUG 7070 return is_obj_array; 7071 } 7072 7073 MarkFromRootsClosure::MarkFromRootsClosure(CMSCollector* collector, 7074 MemRegion span, 7075 CMSBitMap* bitMap, CMSMarkStack* markStack, 7076 CMSMarkStack* revisitStack, 7077 bool should_yield, bool verifying): 7078 _collector(collector), 7079 _span(span), 7080 _bitMap(bitMap), 7081 _mut(&collector->_modUnionTable), 7082 _markStack(markStack), 7083 _revisitStack(revisitStack), 7084 _yield(should_yield), 7085 _skipBits(0) 7086 { 7087 assert(_markStack->isEmpty(), "stack should be empty"); 7088 _finger = _bitMap->startWord(); 7089 _threshold = _finger; 7090 assert(_collector->_restart_addr == NULL, "Sanity check"); 7091 assert(_span.contains(_finger), "Out of bounds _finger?"); 7092 DEBUG_ONLY(_verifying = verifying;) 7093 } 7094 7095 void MarkFromRootsClosure::reset(HeapWord* addr) { 7096 assert(_markStack->isEmpty(), "would cause duplicates on stack"); 7097 assert(_span.contains(addr), "Out of bounds _finger?"); 7098 _finger = addr; 7099 _threshold = (HeapWord*)round_to( 7100 (intptr_t)_finger, CardTableModRefBS::card_size); 7101 } 7102 7103 // Should revisit to see if this should be restructured for 7104 // greater efficiency. 7105 bool MarkFromRootsClosure::do_bit(size_t offset) { 7106 if (_skipBits > 0) { 7107 _skipBits--; 7108 return true; 7109 } 7110 // convert offset into a HeapWord* 7111 HeapWord* addr = _bitMap->startWord() + offset; 7112 assert(_bitMap->endWord() && addr < _bitMap->endWord(), 7113 "address out of range"); 7114 assert(_bitMap->isMarked(addr), "tautology"); 7115 if (_bitMap->isMarked(addr+1)) { 7116 // this is an allocated but not yet initialized object 7117 assert(_skipBits == 0, "tautology"); 7118 _skipBits = 2; // skip next two marked bits ("Printezis-marks") 7119 oop p = oop(addr); 7120 if (p->klass_or_null() == NULL || !p->is_parsable()) { 7121 DEBUG_ONLY(if (!_verifying) {) 7122 // We re-dirty the cards on which this object lies and increase 7123 // the _threshold so that we'll come back to scan this object 7124 // during the preclean or remark phase. (CMSCleanOnEnter) 7125 if (CMSCleanOnEnter) { 7126 size_t sz = _collector->block_size_using_printezis_bits(addr); 7127 HeapWord* end_card_addr = (HeapWord*)round_to( 7128 (intptr_t)(addr+sz), CardTableModRefBS::card_size); 7129 MemRegion redirty_range = MemRegion(addr, end_card_addr); 7130 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 7131 // Bump _threshold to end_card_addr; note that 7132 // _threshold cannot possibly exceed end_card_addr, anyhow. 7133 // This prevents future clearing of the card as the scan proceeds 7134 // to the right. 7135 assert(_threshold <= end_card_addr, 7136 "Because we are just scanning into this object"); 7137 if (_threshold < end_card_addr) { 7138 _threshold = end_card_addr; 7139 } 7140 if (p->klass_or_null() != NULL) { 7141 // Redirty the range of cards... 7142 _mut->mark_range(redirty_range); 7143 } // ...else the setting of klass will dirty the card anyway. 7144 } 7145 DEBUG_ONLY(}) 7146 return true; 7147 } 7148 } 7149 scanOopsInOop(addr); 7150 return true; 7151 } 7152 7153 // We take a break if we've been at this for a while, 7154 // so as to avoid monopolizing the locks involved. 7155 void MarkFromRootsClosure::do_yield_work() { 7156 // First give up the locks, then yield, then re-lock 7157 // We should probably use a constructor/destructor idiom to 7158 // do this unlock/lock or modify the MutexUnlocker class to 7159 // serve our purpose. XXX 7160 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 7161 "CMS thread should hold CMS token"); 7162 assert_lock_strong(_bitMap->lock()); 7163 DEBUG_ONLY(RememberKlassesChecker mux(false);) 7164 _bitMap->lock()->unlock(); 7165 ConcurrentMarkSweepThread::desynchronize(true); 7166 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7167 _collector->stopTimer(); 7168 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 7169 if (PrintCMSStatistics != 0) { 7170 _collector->incrementYields(); 7171 } 7172 _collector->icms_wait(); 7173 7174 // See the comment in coordinator_yield() 7175 for (unsigned i = 0; i < CMSYieldSleepCount && 7176 ConcurrentMarkSweepThread::should_yield() && 7177 !CMSCollector::foregroundGCIsActive(); ++i) { 7178 os::sleep(Thread::current(), 1, false); 7179 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7180 } 7181 7182 ConcurrentMarkSweepThread::synchronize(true); 7183 _bitMap->lock()->lock_without_safepoint_check(); 7184 _collector->startTimer(); 7185 } 7186 7187 void MarkFromRootsClosure::scanOopsInOop(HeapWord* ptr) { 7188 assert(_bitMap->isMarked(ptr), "expected bit to be set"); 7189 assert(_markStack->isEmpty(), 7190 "should drain stack to limit stack usage"); 7191 // convert ptr to an oop preparatory to scanning 7192 oop obj = oop(ptr); 7193 // Ignore mark word in verification below, since we 7194 // may be running concurrent with mutators. 7195 assert(obj->is_oop(true), "should be an oop"); 7196 assert(_finger <= ptr, "_finger runneth ahead"); 7197 // advance the finger to right end of this object 7198 _finger = ptr + obj->size(); 7199 assert(_finger > ptr, "we just incremented it above"); 7200 // On large heaps, it may take us some time to get through 7201 // the marking phase (especially if running iCMS). During 7202 // this time it's possible that a lot of mutations have 7203 // accumulated in the card table and the mod union table -- 7204 // these mutation records are redundant until we have 7205 // actually traced into the corresponding card. 7206 // Here, we check whether advancing the finger would make 7207 // us cross into a new card, and if so clear corresponding 7208 // cards in the MUT (preclean them in the card-table in the 7209 // future). 7210 7211 DEBUG_ONLY(if (!_verifying) {) 7212 // The clean-on-enter optimization is disabled by default, 7213 // until we fix 6178663. 7214 if (CMSCleanOnEnter && (_finger > _threshold)) { 7215 // [_threshold, _finger) represents the interval 7216 // of cards to be cleared in MUT (or precleaned in card table). 7217 // The set of cards to be cleared is all those that overlap 7218 // with the interval [_threshold, _finger); note that 7219 // _threshold is always kept card-aligned but _finger isn't 7220 // always card-aligned. 7221 HeapWord* old_threshold = _threshold; 7222 assert(old_threshold == (HeapWord*)round_to( 7223 (intptr_t)old_threshold, CardTableModRefBS::card_size), 7224 "_threshold should always be card-aligned"); 7225 _threshold = (HeapWord*)round_to( 7226 (intptr_t)_finger, CardTableModRefBS::card_size); 7227 MemRegion mr(old_threshold, _threshold); 7228 assert(!mr.is_empty(), "Control point invariant"); 7229 assert(_span.contains(mr), "Should clear within span"); 7230 // XXX When _finger crosses from old gen into perm gen 7231 // we may be doing unnecessary cleaning; do better in the 7232 // future by detecting that condition and clearing fewer 7233 // MUT/CT entries. 7234 _mut->clear_range(mr); 7235 } 7236 DEBUG_ONLY(}) 7237 // Note: the finger doesn't advance while we drain 7238 // the stack below. 7239 PushOrMarkClosure pushOrMarkClosure(_collector, 7240 _span, _bitMap, _markStack, 7241 _revisitStack, 7242 _finger, this); 7243 bool res = _markStack->push(obj); 7244 assert(res, "Empty non-zero size stack should have space for single push"); 7245 while (!_markStack->isEmpty()) { 7246 oop new_oop = _markStack->pop(); 7247 // Skip verifying header mark word below because we are 7248 // running concurrent with mutators. 7249 assert(new_oop->is_oop(true), "Oops! expected to pop an oop"); 7250 // now scan this oop's oops 7251 new_oop->oop_iterate(&pushOrMarkClosure); 7252 do_yield_check(); 7253 } 7254 assert(_markStack->isEmpty(), "tautology, emphasizing post-condition"); 7255 } 7256 7257 Par_MarkFromRootsClosure::Par_MarkFromRootsClosure(CMSConcMarkingTask* task, 7258 CMSCollector* collector, MemRegion span, 7259 CMSBitMap* bit_map, 7260 OopTaskQueue* work_queue, 7261 CMSMarkStack* overflow_stack, 7262 CMSMarkStack* revisit_stack, 7263 bool should_yield): 7264 _collector(collector), 7265 _whole_span(collector->_span), 7266 _span(span), 7267 _bit_map(bit_map), 7268 _mut(&collector->_modUnionTable), 7269 _work_queue(work_queue), 7270 _overflow_stack(overflow_stack), 7271 _revisit_stack(revisit_stack), 7272 _yield(should_yield), 7273 _skip_bits(0), 7274 _task(task) 7275 { 7276 assert(_work_queue->size() == 0, "work_queue should be empty"); 7277 _finger = span.start(); 7278 _threshold = _finger; // XXX Defer clear-on-enter optimization for now 7279 assert(_span.contains(_finger), "Out of bounds _finger?"); 7280 } 7281 7282 // Should revisit to see if this should be restructured for 7283 // greater efficiency. 7284 bool Par_MarkFromRootsClosure::do_bit(size_t offset) { 7285 if (_skip_bits > 0) { 7286 _skip_bits--; 7287 return true; 7288 } 7289 // convert offset into a HeapWord* 7290 HeapWord* addr = _bit_map->startWord() + offset; 7291 assert(_bit_map->endWord() && addr < _bit_map->endWord(), 7292 "address out of range"); 7293 assert(_bit_map->isMarked(addr), "tautology"); 7294 if (_bit_map->isMarked(addr+1)) { 7295 // this is an allocated object that might not yet be initialized 7296 assert(_skip_bits == 0, "tautology"); 7297 _skip_bits = 2; // skip next two marked bits ("Printezis-marks") 7298 oop p = oop(addr); 7299 if (p->klass_or_null() == NULL || !p->is_parsable()) { 7300 // in the case of Clean-on-Enter optimization, redirty card 7301 // and avoid clearing card by increasing the threshold. 7302 return true; 7303 } 7304 } 7305 scan_oops_in_oop(addr); 7306 return true; 7307 } 7308 7309 void Par_MarkFromRootsClosure::scan_oops_in_oop(HeapWord* ptr) { 7310 assert(_bit_map->isMarked(ptr), "expected bit to be set"); 7311 // Should we assert that our work queue is empty or 7312 // below some drain limit? 7313 assert(_work_queue->size() == 0, 7314 "should drain stack to limit stack usage"); 7315 // convert ptr to an oop preparatory to scanning 7316 oop obj = oop(ptr); 7317 // Ignore mark word in verification below, since we 7318 // may be running concurrent with mutators. 7319 assert(obj->is_oop(true), "should be an oop"); 7320 assert(_finger <= ptr, "_finger runneth ahead"); 7321 // advance the finger to right end of this object 7322 _finger = ptr + obj->size(); 7323 assert(_finger > ptr, "we just incremented it above"); 7324 // On large heaps, it may take us some time to get through 7325 // the marking phase (especially if running iCMS). During 7326 // this time it's possible that a lot of mutations have 7327 // accumulated in the card table and the mod union table -- 7328 // these mutation records are redundant until we have 7329 // actually traced into the corresponding card. 7330 // Here, we check whether advancing the finger would make 7331 // us cross into a new card, and if so clear corresponding 7332 // cards in the MUT (preclean them in the card-table in the 7333 // future). 7334 7335 // The clean-on-enter optimization is disabled by default, 7336 // until we fix 6178663. 7337 if (CMSCleanOnEnter && (_finger > _threshold)) { 7338 // [_threshold, _finger) represents the interval 7339 // of cards to be cleared in MUT (or precleaned in card table). 7340 // The set of cards to be cleared is all those that overlap 7341 // with the interval [_threshold, _finger); note that 7342 // _threshold is always kept card-aligned but _finger isn't 7343 // always card-aligned. 7344 HeapWord* old_threshold = _threshold; 7345 assert(old_threshold == (HeapWord*)round_to( 7346 (intptr_t)old_threshold, CardTableModRefBS::card_size), 7347 "_threshold should always be card-aligned"); 7348 _threshold = (HeapWord*)round_to( 7349 (intptr_t)_finger, CardTableModRefBS::card_size); 7350 MemRegion mr(old_threshold, _threshold); 7351 assert(!mr.is_empty(), "Control point invariant"); 7352 assert(_span.contains(mr), "Should clear within span"); // _whole_span ?? 7353 // XXX When _finger crosses from old gen into perm gen 7354 // we may be doing unnecessary cleaning; do better in the 7355 // future by detecting that condition and clearing fewer 7356 // MUT/CT entries. 7357 _mut->clear_range(mr); 7358 } 7359 7360 // Note: the local finger doesn't advance while we drain 7361 // the stack below, but the global finger sure can and will. 7362 HeapWord** gfa = _task->global_finger_addr(); 7363 Par_PushOrMarkClosure pushOrMarkClosure(_collector, 7364 _span, _bit_map, 7365 _work_queue, 7366 _overflow_stack, 7367 _revisit_stack, 7368 _finger, 7369 gfa, this); 7370 bool res = _work_queue->push(obj); // overflow could occur here 7371 assert(res, "Will hold once we use workqueues"); 7372 while (true) { 7373 oop new_oop; 7374 if (!_work_queue->pop_local(new_oop)) { 7375 // We emptied our work_queue; check if there's stuff that can 7376 // be gotten from the overflow stack. 7377 if (CMSConcMarkingTask::get_work_from_overflow_stack( 7378 _overflow_stack, _work_queue)) { 7379 do_yield_check(); 7380 continue; 7381 } else { // done 7382 break; 7383 } 7384 } 7385 // Skip verifying header mark word below because we are 7386 // running concurrent with mutators. 7387 assert(new_oop->is_oop(true), "Oops! expected to pop an oop"); 7388 // now scan this oop's oops 7389 new_oop->oop_iterate(&pushOrMarkClosure); 7390 do_yield_check(); 7391 } 7392 assert(_work_queue->size() == 0, "tautology, emphasizing post-condition"); 7393 } 7394 7395 // Yield in response to a request from VM Thread or 7396 // from mutators. 7397 void Par_MarkFromRootsClosure::do_yield_work() { 7398 assert(_task != NULL, "sanity"); 7399 _task->yield(); 7400 } 7401 7402 // A variant of the above used for verifying CMS marking work. 7403 MarkFromRootsVerifyClosure::MarkFromRootsVerifyClosure(CMSCollector* collector, 7404 MemRegion span, 7405 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 7406 CMSMarkStack* mark_stack): 7407 _collector(collector), 7408 _span(span), 7409 _verification_bm(verification_bm), 7410 _cms_bm(cms_bm), 7411 _mark_stack(mark_stack), 7412 _pam_verify_closure(collector, span, verification_bm, cms_bm, 7413 mark_stack) 7414 { 7415 assert(_mark_stack->isEmpty(), "stack should be empty"); 7416 _finger = _verification_bm->startWord(); 7417 assert(_collector->_restart_addr == NULL, "Sanity check"); 7418 assert(_span.contains(_finger), "Out of bounds _finger?"); 7419 } 7420 7421 void MarkFromRootsVerifyClosure::reset(HeapWord* addr) { 7422 assert(_mark_stack->isEmpty(), "would cause duplicates on stack"); 7423 assert(_span.contains(addr), "Out of bounds _finger?"); 7424 _finger = addr; 7425 } 7426 7427 // Should revisit to see if this should be restructured for 7428 // greater efficiency. 7429 bool MarkFromRootsVerifyClosure::do_bit(size_t offset) { 7430 // convert offset into a HeapWord* 7431 HeapWord* addr = _verification_bm->startWord() + offset; 7432 assert(_verification_bm->endWord() && addr < _verification_bm->endWord(), 7433 "address out of range"); 7434 assert(_verification_bm->isMarked(addr), "tautology"); 7435 assert(_cms_bm->isMarked(addr), "tautology"); 7436 7437 assert(_mark_stack->isEmpty(), 7438 "should drain stack to limit stack usage"); 7439 // convert addr to an oop preparatory to scanning 7440 oop obj = oop(addr); 7441 assert(obj->is_oop(), "should be an oop"); 7442 assert(_finger <= addr, "_finger runneth ahead"); 7443 // advance the finger to right end of this object 7444 _finger = addr + obj->size(); 7445 assert(_finger > addr, "we just incremented it above"); 7446 // Note: the finger doesn't advance while we drain 7447 // the stack below. 7448 bool res = _mark_stack->push(obj); 7449 assert(res, "Empty non-zero size stack should have space for single push"); 7450 while (!_mark_stack->isEmpty()) { 7451 oop new_oop = _mark_stack->pop(); 7452 assert(new_oop->is_oop(), "Oops! expected to pop an oop"); 7453 // now scan this oop's oops 7454 new_oop->oop_iterate(&_pam_verify_closure); 7455 } 7456 assert(_mark_stack->isEmpty(), "tautology, emphasizing post-condition"); 7457 return true; 7458 } 7459 7460 PushAndMarkVerifyClosure::PushAndMarkVerifyClosure( 7461 CMSCollector* collector, MemRegion span, 7462 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 7463 CMSMarkStack* mark_stack): 7464 OopClosure(collector->ref_processor()), 7465 _collector(collector), 7466 _span(span), 7467 _verification_bm(verification_bm), 7468 _cms_bm(cms_bm), 7469 _mark_stack(mark_stack) 7470 { } 7471 7472 void PushAndMarkVerifyClosure::do_oop(oop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 7473 void PushAndMarkVerifyClosure::do_oop(narrowOop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 7474 7475 // Upon stack overflow, we discard (part of) the stack, 7476 // remembering the least address amongst those discarded 7477 // in CMSCollector's _restart_address. 7478 void PushAndMarkVerifyClosure::handle_stack_overflow(HeapWord* lost) { 7479 // Remember the least grey address discarded 7480 HeapWord* ra = (HeapWord*)_mark_stack->least_value(lost); 7481 _collector->lower_restart_addr(ra); 7482 _mark_stack->reset(); // discard stack contents 7483 _mark_stack->expand(); // expand the stack if possible 7484 } 7485 7486 void PushAndMarkVerifyClosure::do_oop(oop obj) { 7487 assert(obj->is_oop_or_null(), "expected an oop or NULL"); 7488 HeapWord* addr = (HeapWord*)obj; 7489 if (_span.contains(addr) && !_verification_bm->isMarked(addr)) { 7490 // Oop lies in _span and isn't yet grey or black 7491 _verification_bm->mark(addr); // now grey 7492 if (!_cms_bm->isMarked(addr)) { 7493 oop(addr)->print(); 7494 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)", 7495 addr); 7496 fatal("... aborting"); 7497 } 7498 7499 if (!_mark_stack->push(obj)) { // stack overflow 7500 if (PrintCMSStatistics != 0) { 7501 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 7502 SIZE_FORMAT, _mark_stack->capacity()); 7503 } 7504 assert(_mark_stack->isFull(), "Else push should have succeeded"); 7505 handle_stack_overflow(addr); 7506 } 7507 // anything including and to the right of _finger 7508 // will be scanned as we iterate over the remainder of the 7509 // bit map 7510 } 7511 } 7512 7513 PushOrMarkClosure::PushOrMarkClosure(CMSCollector* collector, 7514 MemRegion span, 7515 CMSBitMap* bitMap, CMSMarkStack* markStack, 7516 CMSMarkStack* revisitStack, 7517 HeapWord* finger, MarkFromRootsClosure* parent) : 7518 KlassRememberingOopClosure(collector, collector->ref_processor(), revisitStack), 7519 _span(span), 7520 _bitMap(bitMap), 7521 _markStack(markStack), 7522 _finger(finger), 7523 _parent(parent) 7524 { } 7525 7526 Par_PushOrMarkClosure::Par_PushOrMarkClosure(CMSCollector* collector, 7527 MemRegion span, 7528 CMSBitMap* bit_map, 7529 OopTaskQueue* work_queue, 7530 CMSMarkStack* overflow_stack, 7531 CMSMarkStack* revisit_stack, 7532 HeapWord* finger, 7533 HeapWord** global_finger_addr, 7534 Par_MarkFromRootsClosure* parent) : 7535 Par_KlassRememberingOopClosure(collector, 7536 collector->ref_processor(), 7537 revisit_stack), 7538 _whole_span(collector->_span), 7539 _span(span), 7540 _bit_map(bit_map), 7541 _work_queue(work_queue), 7542 _overflow_stack(overflow_stack), 7543 _finger(finger), 7544 _global_finger_addr(global_finger_addr), 7545 _parent(parent) 7546 { } 7547 7548 // Assumes thread-safe access by callers, who are 7549 // responsible for mutual exclusion. 7550 void CMSCollector::lower_restart_addr(HeapWord* low) { 7551 assert(_span.contains(low), "Out of bounds addr"); 7552 if (_restart_addr == NULL) { 7553 _restart_addr = low; 7554 } else { 7555 _restart_addr = MIN2(_restart_addr, low); 7556 } 7557 } 7558 7559 // Upon stack overflow, we discard (part of) the stack, 7560 // remembering the least address amongst those discarded 7561 // in CMSCollector's _restart_address. 7562 void PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 7563 // Remember the least grey address discarded 7564 HeapWord* ra = (HeapWord*)_markStack->least_value(lost); 7565 _collector->lower_restart_addr(ra); 7566 _markStack->reset(); // discard stack contents 7567 _markStack->expand(); // expand the stack if possible 7568 } 7569 7570 // Upon stack overflow, we discard (part of) the stack, 7571 // remembering the least address amongst those discarded 7572 // in CMSCollector's _restart_address. 7573 void Par_PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 7574 // We need to do this under a mutex to prevent other 7575 // workers from interfering with the work done below. 7576 MutexLockerEx ml(_overflow_stack->par_lock(), 7577 Mutex::_no_safepoint_check_flag); 7578 // Remember the least grey address discarded 7579 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 7580 _collector->lower_restart_addr(ra); 7581 _overflow_stack->reset(); // discard stack contents 7582 _overflow_stack->expand(); // expand the stack if possible 7583 } 7584 7585 void PushOrMarkClosure::do_oop(oop obj) { 7586 // Ignore mark word because we are running concurrent with mutators. 7587 assert(obj->is_oop_or_null(true), "expected an oop or NULL"); 7588 HeapWord* addr = (HeapWord*)obj; 7589 if (_span.contains(addr) && !_bitMap->isMarked(addr)) { 7590 // Oop lies in _span and isn't yet grey or black 7591 _bitMap->mark(addr); // now grey 7592 if (addr < _finger) { 7593 // the bit map iteration has already either passed, or 7594 // sampled, this bit in the bit map; we'll need to 7595 // use the marking stack to scan this oop's oops. 7596 bool simulate_overflow = false; 7597 NOT_PRODUCT( 7598 if (CMSMarkStackOverflowALot && 7599 _collector->simulate_overflow()) { 7600 // simulate a stack overflow 7601 simulate_overflow = true; 7602 } 7603 ) 7604 if (simulate_overflow || !_markStack->push(obj)) { // stack overflow 7605 if (PrintCMSStatistics != 0) { 7606 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 7607 SIZE_FORMAT, _markStack->capacity()); 7608 } 7609 assert(simulate_overflow || _markStack->isFull(), "Else push should have succeeded"); 7610 handle_stack_overflow(addr); 7611 } 7612 } 7613 // anything including and to the right of _finger 7614 // will be scanned as we iterate over the remainder of the 7615 // bit map 7616 do_yield_check(); 7617 } 7618 } 7619 7620 void PushOrMarkClosure::do_oop(oop* p) { PushOrMarkClosure::do_oop_work(p); } 7621 void PushOrMarkClosure::do_oop(narrowOop* p) { PushOrMarkClosure::do_oop_work(p); } 7622 7623 void Par_PushOrMarkClosure::do_oop(oop obj) { 7624 // Ignore mark word because we are running concurrent with mutators. 7625 assert(obj->is_oop_or_null(true), "expected an oop or NULL"); 7626 HeapWord* addr = (HeapWord*)obj; 7627 if (_whole_span.contains(addr) && !_bit_map->isMarked(addr)) { 7628 // Oop lies in _span and isn't yet grey or black 7629 // We read the global_finger (volatile read) strictly after marking oop 7630 bool res = _bit_map->par_mark(addr); // now grey 7631 volatile HeapWord** gfa = (volatile HeapWord**)_global_finger_addr; 7632 // Should we push this marked oop on our stack? 7633 // -- if someone else marked it, nothing to do 7634 // -- if target oop is above global finger nothing to do 7635 // -- if target oop is in chunk and above local finger 7636 // then nothing to do 7637 // -- else push on work queue 7638 if ( !res // someone else marked it, they will deal with it 7639 || (addr >= *gfa) // will be scanned in a later task 7640 || (_span.contains(addr) && addr >= _finger)) { // later in this chunk 7641 return; 7642 } 7643 // the bit map iteration has already either passed, or 7644 // sampled, this bit in the bit map; we'll need to 7645 // use the marking stack to scan this oop's oops. 7646 bool simulate_overflow = false; 7647 NOT_PRODUCT( 7648 if (CMSMarkStackOverflowALot && 7649 _collector->simulate_overflow()) { 7650 // simulate a stack overflow 7651 simulate_overflow = true; 7652 } 7653 ) 7654 if (simulate_overflow || 7655 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) { 7656 // stack overflow 7657 if (PrintCMSStatistics != 0) { 7658 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 7659 SIZE_FORMAT, _overflow_stack->capacity()); 7660 } 7661 // We cannot assert that the overflow stack is full because 7662 // it may have been emptied since. 7663 assert(simulate_overflow || 7664 _work_queue->size() == _work_queue->max_elems(), 7665 "Else push should have succeeded"); 7666 handle_stack_overflow(addr); 7667 } 7668 do_yield_check(); 7669 } 7670 } 7671 7672 void Par_PushOrMarkClosure::do_oop(oop* p) { Par_PushOrMarkClosure::do_oop_work(p); } 7673 void Par_PushOrMarkClosure::do_oop(narrowOop* p) { Par_PushOrMarkClosure::do_oop_work(p); } 7674 7675 KlassRememberingOopClosure::KlassRememberingOopClosure(CMSCollector* collector, 7676 ReferenceProcessor* rp, 7677 CMSMarkStack* revisit_stack) : 7678 OopClosure(rp), 7679 _collector(collector), 7680 _revisit_stack(revisit_stack), 7681 _should_remember_klasses(collector->should_unload_classes()) {} 7682 7683 PushAndMarkClosure::PushAndMarkClosure(CMSCollector* collector, 7684 MemRegion span, 7685 ReferenceProcessor* rp, 7686 CMSBitMap* bit_map, 7687 CMSBitMap* mod_union_table, 7688 CMSMarkStack* mark_stack, 7689 CMSMarkStack* revisit_stack, 7690 bool concurrent_precleaning): 7691 KlassRememberingOopClosure(collector, rp, revisit_stack), 7692 _span(span), 7693 _bit_map(bit_map), 7694 _mod_union_table(mod_union_table), 7695 _mark_stack(mark_stack), 7696 _concurrent_precleaning(concurrent_precleaning) 7697 { 7698 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 7699 } 7700 7701 // Grey object rescan during pre-cleaning and second checkpoint phases -- 7702 // the non-parallel version (the parallel version appears further below.) 7703 void PushAndMarkClosure::do_oop(oop obj) { 7704 // Ignore mark word verification. If during concurrent precleaning, 7705 // the object monitor may be locked. If during the checkpoint 7706 // phases, the object may already have been reached by a different 7707 // path and may be at the end of the global overflow list (so 7708 // the mark word may be NULL). 7709 assert(obj->is_oop_or_null(true /* ignore mark word */), 7710 "expected an oop or NULL"); 7711 HeapWord* addr = (HeapWord*)obj; 7712 // Check if oop points into the CMS generation 7713 // and is not marked 7714 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 7715 // a white object ... 7716 _bit_map->mark(addr); // ... now grey 7717 // push on the marking stack (grey set) 7718 bool simulate_overflow = false; 7719 NOT_PRODUCT( 7720 if (CMSMarkStackOverflowALot && 7721 _collector->simulate_overflow()) { 7722 // simulate a stack overflow 7723 simulate_overflow = true; 7724 } 7725 ) 7726 if (simulate_overflow || !_mark_stack->push(obj)) { 7727 if (_concurrent_precleaning) { 7728 // During precleaning we can just dirty the appropriate card(s) 7729 // in the mod union table, thus ensuring that the object remains 7730 // in the grey set and continue. In the case of object arrays 7731 // we need to dirty all of the cards that the object spans, 7732 // since the rescan of object arrays will be limited to the 7733 // dirty cards. 7734 // Note that no one can be intefering with us in this action 7735 // of dirtying the mod union table, so no locking or atomics 7736 // are required. 7737 if (obj->is_objArray()) { 7738 size_t sz = obj->size(); 7739 HeapWord* end_card_addr = (HeapWord*)round_to( 7740 (intptr_t)(addr+sz), CardTableModRefBS::card_size); 7741 MemRegion redirty_range = MemRegion(addr, end_card_addr); 7742 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 7743 _mod_union_table->mark_range(redirty_range); 7744 } else { 7745 _mod_union_table->mark(addr); 7746 } 7747 _collector->_ser_pmc_preclean_ovflw++; 7748 } else { 7749 // During the remark phase, we need to remember this oop 7750 // in the overflow list. 7751 _collector->push_on_overflow_list(obj); 7752 _collector->_ser_pmc_remark_ovflw++; 7753 } 7754 } 7755 } 7756 } 7757 7758 Par_PushAndMarkClosure::Par_PushAndMarkClosure(CMSCollector* collector, 7759 MemRegion span, 7760 ReferenceProcessor* rp, 7761 CMSBitMap* bit_map, 7762 OopTaskQueue* work_queue, 7763 CMSMarkStack* revisit_stack): 7764 Par_KlassRememberingOopClosure(collector, rp, revisit_stack), 7765 _span(span), 7766 _bit_map(bit_map), 7767 _work_queue(work_queue) 7768 { 7769 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 7770 } 7771 7772 void PushAndMarkClosure::do_oop(oop* p) { PushAndMarkClosure::do_oop_work(p); } 7773 void PushAndMarkClosure::do_oop(narrowOop* p) { PushAndMarkClosure::do_oop_work(p); } 7774 7775 // Grey object rescan during second checkpoint phase -- 7776 // the parallel version. 7777 void Par_PushAndMarkClosure::do_oop(oop obj) { 7778 // In the assert below, we ignore the mark word because 7779 // this oop may point to an already visited object that is 7780 // on the overflow stack (in which case the mark word has 7781 // been hijacked for chaining into the overflow stack -- 7782 // if this is the last object in the overflow stack then 7783 // its mark word will be NULL). Because this object may 7784 // have been subsequently popped off the global overflow 7785 // stack, and the mark word possibly restored to the prototypical 7786 // value, by the time we get to examined this failing assert in 7787 // the debugger, is_oop_or_null(false) may subsequently start 7788 // to hold. 7789 assert(obj->is_oop_or_null(true), 7790 "expected an oop or NULL"); 7791 HeapWord* addr = (HeapWord*)obj; 7792 // Check if oop points into the CMS generation 7793 // and is not marked 7794 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 7795 // a white object ... 7796 // If we manage to "claim" the object, by being the 7797 // first thread to mark it, then we push it on our 7798 // marking stack 7799 if (_bit_map->par_mark(addr)) { // ... now grey 7800 // push on work queue (grey set) 7801 bool simulate_overflow = false; 7802 NOT_PRODUCT( 7803 if (CMSMarkStackOverflowALot && 7804 _collector->par_simulate_overflow()) { 7805 // simulate a stack overflow 7806 simulate_overflow = true; 7807 } 7808 ) 7809 if (simulate_overflow || !_work_queue->push(obj)) { 7810 _collector->par_push_on_overflow_list(obj); 7811 _collector->_par_pmc_remark_ovflw++; // imprecise OK: no need to CAS 7812 } 7813 } // Else, some other thread got there first 7814 } 7815 } 7816 7817 void Par_PushAndMarkClosure::do_oop(oop* p) { Par_PushAndMarkClosure::do_oop_work(p); } 7818 void Par_PushAndMarkClosure::do_oop(narrowOop* p) { Par_PushAndMarkClosure::do_oop_work(p); } 7819 7820 void PushAndMarkClosure::remember_mdo(DataLayout* v) { 7821 // TBD 7822 } 7823 7824 void Par_PushAndMarkClosure::remember_mdo(DataLayout* v) { 7825 // TBD 7826 } 7827 7828 void CMSPrecleanRefsYieldClosure::do_yield_work() { 7829 DEBUG_ONLY(RememberKlassesChecker mux(false);) 7830 Mutex* bml = _collector->bitMapLock(); 7831 assert_lock_strong(bml); 7832 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 7833 "CMS thread should hold CMS token"); 7834 7835 bml->unlock(); 7836 ConcurrentMarkSweepThread::desynchronize(true); 7837 7838 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7839 7840 _collector->stopTimer(); 7841 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 7842 if (PrintCMSStatistics != 0) { 7843 _collector->incrementYields(); 7844 } 7845 _collector->icms_wait(); 7846 7847 // See the comment in coordinator_yield() 7848 for (unsigned i = 0; i < CMSYieldSleepCount && 7849 ConcurrentMarkSweepThread::should_yield() && 7850 !CMSCollector::foregroundGCIsActive(); ++i) { 7851 os::sleep(Thread::current(), 1, false); 7852 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7853 } 7854 7855 ConcurrentMarkSweepThread::synchronize(true); 7856 bml->lock(); 7857 7858 _collector->startTimer(); 7859 } 7860 7861 bool CMSPrecleanRefsYieldClosure::should_return() { 7862 if (ConcurrentMarkSweepThread::should_yield()) { 7863 do_yield_work(); 7864 } 7865 return _collector->foregroundGCIsActive(); 7866 } 7867 7868 void MarkFromDirtyCardsClosure::do_MemRegion(MemRegion mr) { 7869 assert(((size_t)mr.start())%CardTableModRefBS::card_size_in_words == 0, 7870 "mr should be aligned to start at a card boundary"); 7871 // We'd like to assert: 7872 // assert(mr.word_size()%CardTableModRefBS::card_size_in_words == 0, 7873 // "mr should be a range of cards"); 7874 // However, that would be too strong in one case -- the last 7875 // partition ends at _unallocated_block which, in general, can be 7876 // an arbitrary boundary, not necessarily card aligned. 7877 if (PrintCMSStatistics != 0) { 7878 _num_dirty_cards += 7879 mr.word_size()/CardTableModRefBS::card_size_in_words; 7880 } 7881 _space->object_iterate_mem(mr, &_scan_cl); 7882 } 7883 7884 SweepClosure::SweepClosure(CMSCollector* collector, 7885 ConcurrentMarkSweepGeneration* g, 7886 CMSBitMap* bitMap, bool should_yield) : 7887 _collector(collector), 7888 _g(g), 7889 _sp(g->cmsSpace()), 7890 _limit(_sp->sweep_limit()), 7891 _freelistLock(_sp->freelistLock()), 7892 _bitMap(bitMap), 7893 _yield(should_yield), 7894 _inFreeRange(false), // No free range at beginning of sweep 7895 _freeRangeInFreeLists(false), // No free range at beginning of sweep 7896 _lastFreeRangeCoalesced(false), 7897 _freeFinger(g->used_region().start()) 7898 { 7899 NOT_PRODUCT( 7900 _numObjectsFreed = 0; 7901 _numWordsFreed = 0; 7902 _numObjectsLive = 0; 7903 _numWordsLive = 0; 7904 _numObjectsAlreadyFree = 0; 7905 _numWordsAlreadyFree = 0; 7906 _last_fc = NULL; 7907 7908 _sp->initializeIndexedFreeListArrayReturnedBytes(); 7909 _sp->dictionary()->initialize_dict_returned_bytes(); 7910 ) 7911 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7912 "sweep _limit out of bounds"); 7913 if (CMSTraceSweeper) { 7914 gclog_or_tty->print_cr("\n====================\nStarting new sweep with limit " PTR_FORMAT, 7915 _limit); 7916 } 7917 } 7918 7919 void SweepClosure::print_on(outputStream* st) const { 7920 tty->print_cr("_sp = [" PTR_FORMAT "," PTR_FORMAT ")", 7921 _sp->bottom(), _sp->end()); 7922 tty->print_cr("_limit = " PTR_FORMAT, _limit); 7923 tty->print_cr("_freeFinger = " PTR_FORMAT, _freeFinger); 7924 NOT_PRODUCT(tty->print_cr("_last_fc = " PTR_FORMAT, _last_fc);) 7925 tty->print_cr("_inFreeRange = %d, _freeRangeInFreeLists = %d, _lastFreeRangeCoalesced = %d", 7926 _inFreeRange, _freeRangeInFreeLists, _lastFreeRangeCoalesced); 7927 } 7928 7929 #ifndef PRODUCT 7930 // Assertion checking only: no useful work in product mode -- 7931 // however, if any of the flags below become product flags, 7932 // you may need to review this code to see if it needs to be 7933 // enabled in product mode. 7934 SweepClosure::~SweepClosure() { 7935 assert_lock_strong(_freelistLock); 7936 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7937 "sweep _limit out of bounds"); 7938 if (inFreeRange()) { 7939 warning("inFreeRange() should have been reset; dumping state of SweepClosure"); 7940 print(); 7941 ShouldNotReachHere(); 7942 } 7943 if (Verbose && PrintGC) { 7944 gclog_or_tty->print("Collected "SIZE_FORMAT" objects, " SIZE_FORMAT " bytes", 7945 _numObjectsFreed, _numWordsFreed*sizeof(HeapWord)); 7946 gclog_or_tty->print_cr("\nLive "SIZE_FORMAT" objects, " 7947 SIZE_FORMAT" bytes " 7948 "Already free "SIZE_FORMAT" objects, "SIZE_FORMAT" bytes", 7949 _numObjectsLive, _numWordsLive*sizeof(HeapWord), 7950 _numObjectsAlreadyFree, _numWordsAlreadyFree*sizeof(HeapWord)); 7951 size_t totalBytes = (_numWordsFreed + _numWordsLive + _numWordsAlreadyFree) 7952 * sizeof(HeapWord); 7953 gclog_or_tty->print_cr("Total sweep: "SIZE_FORMAT" bytes", totalBytes); 7954 7955 if (PrintCMSStatistics && CMSVerifyReturnedBytes) { 7956 size_t indexListReturnedBytes = _sp->sumIndexedFreeListArrayReturnedBytes(); 7957 size_t dict_returned_bytes = _sp->dictionary()->sum_dict_returned_bytes(); 7958 size_t returned_bytes = indexListReturnedBytes + dict_returned_bytes; 7959 gclog_or_tty->print("Returned "SIZE_FORMAT" bytes", returned_bytes); 7960 gclog_or_tty->print(" Indexed List Returned "SIZE_FORMAT" bytes", 7961 indexListReturnedBytes); 7962 gclog_or_tty->print_cr(" Dictionary Returned "SIZE_FORMAT" bytes", 7963 dict_returned_bytes); 7964 } 7965 } 7966 if (CMSTraceSweeper) { 7967 gclog_or_tty->print_cr("end of sweep with _limit = " PTR_FORMAT "\n================", 7968 _limit); 7969 } 7970 } 7971 #endif // PRODUCT 7972 7973 void SweepClosure::initialize_free_range(HeapWord* freeFinger, 7974 bool freeRangeInFreeLists) { 7975 if (CMSTraceSweeper) { 7976 gclog_or_tty->print("---- Start free range at 0x%x with free block (%d)\n", 7977 freeFinger, freeRangeInFreeLists); 7978 } 7979 assert(!inFreeRange(), "Trampling existing free range"); 7980 set_inFreeRange(true); 7981 set_lastFreeRangeCoalesced(false); 7982 7983 set_freeFinger(freeFinger); 7984 set_freeRangeInFreeLists(freeRangeInFreeLists); 7985 if (CMSTestInFreeList) { 7986 if (freeRangeInFreeLists) { 7987 FreeChunk* fc = (FreeChunk*) freeFinger; 7988 assert(fc->is_free(), "A chunk on the free list should be free."); 7989 assert(fc->size() > 0, "Free range should have a size"); 7990 assert(_sp->verify_chunk_in_free_list(fc), "Chunk is not in free lists"); 7991 } 7992 } 7993 } 7994 7995 // Note that the sweeper runs concurrently with mutators. Thus, 7996 // it is possible for direct allocation in this generation to happen 7997 // in the middle of the sweep. Note that the sweeper also coalesces 7998 // contiguous free blocks. Thus, unless the sweeper and the allocator 7999 // synchronize appropriately freshly allocated blocks may get swept up. 8000 // This is accomplished by the sweeper locking the free lists while 8001 // it is sweeping. Thus blocks that are determined to be free are 8002 // indeed free. There is however one additional complication: 8003 // blocks that have been allocated since the final checkpoint and 8004 // mark, will not have been marked and so would be treated as 8005 // unreachable and swept up. To prevent this, the allocator marks 8006 // the bit map when allocating during the sweep phase. This leads, 8007 // however, to a further complication -- objects may have been allocated 8008 // but not yet initialized -- in the sense that the header isn't yet 8009 // installed. The sweeper can not then determine the size of the block 8010 // in order to skip over it. To deal with this case, we use a technique 8011 // (due to Printezis) to encode such uninitialized block sizes in the 8012 // bit map. Since the bit map uses a bit per every HeapWord, but the 8013 // CMS generation has a minimum object size of 3 HeapWords, it follows 8014 // that "normal marks" won't be adjacent in the bit map (there will 8015 // always be at least two 0 bits between successive 1 bits). We make use 8016 // of these "unused" bits to represent uninitialized blocks -- the bit 8017 // corresponding to the start of the uninitialized object and the next 8018 // bit are both set. Finally, a 1 bit marks the end of the object that 8019 // started with the two consecutive 1 bits to indicate its potentially 8020 // uninitialized state. 8021 8022 size_t SweepClosure::do_blk_careful(HeapWord* addr) { 8023 FreeChunk* fc = (FreeChunk*)addr; 8024 size_t res; 8025 8026 // Check if we are done sweeping. Below we check "addr >= _limit" rather 8027 // than "addr == _limit" because although _limit was a block boundary when 8028 // we started the sweep, it may no longer be one because heap expansion 8029 // may have caused us to coalesce the block ending at the address _limit 8030 // with a newly expanded chunk (this happens when _limit was set to the 8031 // previous _end of the space), so we may have stepped past _limit: 8032 // see the following Zeno-like trail of CRs 6977970, 7008136, 7042740. 8033 if (addr >= _limit) { // we have swept up to or past the limit: finish up 8034 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 8035 "sweep _limit out of bounds"); 8036 assert(addr < _sp->end(), "addr out of bounds"); 8037 // Flush any free range we might be holding as a single 8038 // coalesced chunk to the appropriate free list. 8039 if (inFreeRange()) { 8040 assert(freeFinger() >= _sp->bottom() && freeFinger() < _limit, 8041 err_msg("freeFinger() " PTR_FORMAT" is out-of-bounds", freeFinger())); 8042 flush_cur_free_chunk(freeFinger(), 8043 pointer_delta(addr, freeFinger())); 8044 if (CMSTraceSweeper) { 8045 gclog_or_tty->print("Sweep: last chunk: "); 8046 gclog_or_tty->print("put_free_blk 0x%x ("SIZE_FORMAT") " 8047 "[coalesced:"SIZE_FORMAT"]\n", 8048 freeFinger(), pointer_delta(addr, freeFinger()), 8049 lastFreeRangeCoalesced()); 8050 } 8051 } 8052 8053 // help the iterator loop finish 8054 return pointer_delta(_sp->end(), addr); 8055 } 8056 8057 assert(addr < _limit, "sweep invariant"); 8058 // check if we should yield 8059 do_yield_check(addr); 8060 if (fc->is_free()) { 8061 // Chunk that is already free 8062 res = fc->size(); 8063 do_already_free_chunk(fc); 8064 debug_only(_sp->verifyFreeLists()); 8065 // If we flush the chunk at hand in lookahead_and_flush() 8066 // and it's coalesced with a preceding chunk, then the 8067 // process of "mangling" the payload of the coalesced block 8068 // will cause erasure of the size information from the 8069 // (erstwhile) header of all the coalesced blocks but the 8070 // first, so the first disjunct in the assert will not hold 8071 // in that specific case (in which case the second disjunct 8072 // will hold). 8073 assert(res == fc->size() || ((HeapWord*)fc) + res >= _limit, 8074 "Otherwise the size info doesn't change at this step"); 8075 NOT_PRODUCT( 8076 _numObjectsAlreadyFree++; 8077 _numWordsAlreadyFree += res; 8078 ) 8079 NOT_PRODUCT(_last_fc = fc;) 8080 } else if (!_bitMap->isMarked(addr)) { 8081 // Chunk is fresh garbage 8082 res = do_garbage_chunk(fc); 8083 debug_only(_sp->verifyFreeLists()); 8084 NOT_PRODUCT( 8085 _numObjectsFreed++; 8086 _numWordsFreed += res; 8087 ) 8088 } else { 8089 // Chunk that is alive. 8090 res = do_live_chunk(fc); 8091 debug_only(_sp->verifyFreeLists()); 8092 NOT_PRODUCT( 8093 _numObjectsLive++; 8094 _numWordsLive += res; 8095 ) 8096 } 8097 return res; 8098 } 8099 8100 // For the smart allocation, record following 8101 // split deaths - a free chunk is removed from its free list because 8102 // it is being split into two or more chunks. 8103 // split birth - a free chunk is being added to its free list because 8104 // a larger free chunk has been split and resulted in this free chunk. 8105 // coal death - a free chunk is being removed from its free list because 8106 // it is being coalesced into a large free chunk. 8107 // coal birth - a free chunk is being added to its free list because 8108 // it was created when two or more free chunks where coalesced into 8109 // this free chunk. 8110 // 8111 // These statistics are used to determine the desired number of free 8112 // chunks of a given size. The desired number is chosen to be relative 8113 // to the end of a CMS sweep. The desired number at the end of a sweep 8114 // is the 8115 // count-at-end-of-previous-sweep (an amount that was enough) 8116 // - count-at-beginning-of-current-sweep (the excess) 8117 // + split-births (gains in this size during interval) 8118 // - split-deaths (demands on this size during interval) 8119 // where the interval is from the end of one sweep to the end of the 8120 // next. 8121 // 8122 // When sweeping the sweeper maintains an accumulated chunk which is 8123 // the chunk that is made up of chunks that have been coalesced. That 8124 // will be termed the left-hand chunk. A new chunk of garbage that 8125 // is being considered for coalescing will be referred to as the 8126 // right-hand chunk. 8127 // 8128 // When making a decision on whether to coalesce a right-hand chunk with 8129 // the current left-hand chunk, the current count vs. the desired count 8130 // of the left-hand chunk is considered. Also if the right-hand chunk 8131 // is near the large chunk at the end of the heap (see 8132 // ConcurrentMarkSweepGeneration::isNearLargestChunk()), then the 8133 // left-hand chunk is coalesced. 8134 // 8135 // When making a decision about whether to split a chunk, the desired count 8136 // vs. the current count of the candidate to be split is also considered. 8137 // If the candidate is underpopulated (currently fewer chunks than desired) 8138 // a chunk of an overpopulated (currently more chunks than desired) size may 8139 // be chosen. The "hint" associated with a free list, if non-null, points 8140 // to a free list which may be overpopulated. 8141 // 8142 8143 void SweepClosure::do_already_free_chunk(FreeChunk* fc) { 8144 const size_t size = fc->size(); 8145 // Chunks that cannot be coalesced are not in the 8146 // free lists. 8147 if (CMSTestInFreeList && !fc->cantCoalesce()) { 8148 assert(_sp->verify_chunk_in_free_list(fc), 8149 "free chunk should be in free lists"); 8150 } 8151 // a chunk that is already free, should not have been 8152 // marked in the bit map 8153 HeapWord* const addr = (HeapWord*) fc; 8154 assert(!_bitMap->isMarked(addr), "free chunk should be unmarked"); 8155 // Verify that the bit map has no bits marked between 8156 // addr and purported end of this block. 8157 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 8158 8159 // Some chunks cannot be coalesced under any circumstances. 8160 // See the definition of cantCoalesce(). 8161 if (!fc->cantCoalesce()) { 8162 // This chunk can potentially be coalesced. 8163 if (_sp->adaptive_freelists()) { 8164 // All the work is done in 8165 do_post_free_or_garbage_chunk(fc, size); 8166 } else { // Not adaptive free lists 8167 // this is a free chunk that can potentially be coalesced by the sweeper; 8168 if (!inFreeRange()) { 8169 // if the next chunk is a free block that can't be coalesced 8170 // it doesn't make sense to remove this chunk from the free lists 8171 FreeChunk* nextChunk = (FreeChunk*)(addr + size); 8172 assert((HeapWord*)nextChunk <= _sp->end(), "Chunk size out of bounds?"); 8173 if ((HeapWord*)nextChunk < _sp->end() && // There is another free chunk to the right ... 8174 nextChunk->is_free() && // ... which is free... 8175 nextChunk->cantCoalesce()) { // ... but can't be coalesced 8176 // nothing to do 8177 } else { 8178 // Potentially the start of a new free range: 8179 // Don't eagerly remove it from the free lists. 8180 // No need to remove it if it will just be put 8181 // back again. (Also from a pragmatic point of view 8182 // if it is a free block in a region that is beyond 8183 // any allocated blocks, an assertion will fail) 8184 // Remember the start of a free run. 8185 initialize_free_range(addr, true); 8186 // end - can coalesce with next chunk 8187 } 8188 } else { 8189 // the midst of a free range, we are coalescing 8190 print_free_block_coalesced(fc); 8191 if (CMSTraceSweeper) { 8192 gclog_or_tty->print(" -- pick up free block 0x%x (%d)\n", fc, size); 8193 } 8194 // remove it from the free lists 8195 _sp->removeFreeChunkFromFreeLists(fc); 8196 set_lastFreeRangeCoalesced(true); 8197 // If the chunk is being coalesced and the current free range is 8198 // in the free lists, remove the current free range so that it 8199 // will be returned to the free lists in its entirety - all 8200 // the coalesced pieces included. 8201 if (freeRangeInFreeLists()) { 8202 FreeChunk* ffc = (FreeChunk*) freeFinger(); 8203 assert(ffc->size() == pointer_delta(addr, freeFinger()), 8204 "Size of free range is inconsistent with chunk size."); 8205 if (CMSTestInFreeList) { 8206 assert(_sp->verify_chunk_in_free_list(ffc), 8207 "free range is not in free lists"); 8208 } 8209 _sp->removeFreeChunkFromFreeLists(ffc); 8210 set_freeRangeInFreeLists(false); 8211 } 8212 } 8213 } 8214 // Note that if the chunk is not coalescable (the else arm 8215 // below), we unconditionally flush, without needing to do 8216 // a "lookahead," as we do below. 8217 if (inFreeRange()) lookahead_and_flush(fc, size); 8218 } else { 8219 // Code path common to both original and adaptive free lists. 8220 8221 // cant coalesce with previous block; this should be treated 8222 // as the end of a free run if any 8223 if (inFreeRange()) { 8224 // we kicked some butt; time to pick up the garbage 8225 assert(freeFinger() < addr, "freeFinger points too high"); 8226 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 8227 } 8228 // else, nothing to do, just continue 8229 } 8230 } 8231 8232 size_t SweepClosure::do_garbage_chunk(FreeChunk* fc) { 8233 // This is a chunk of garbage. It is not in any free list. 8234 // Add it to a free list or let it possibly be coalesced into 8235 // a larger chunk. 8236 HeapWord* const addr = (HeapWord*) fc; 8237 const size_t size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 8238 8239 if (_sp->adaptive_freelists()) { 8240 // Verify that the bit map has no bits marked between 8241 // addr and purported end of just dead object. 8242 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 8243 8244 do_post_free_or_garbage_chunk(fc, size); 8245 } else { 8246 if (!inFreeRange()) { 8247 // start of a new free range 8248 assert(size > 0, "A free range should have a size"); 8249 initialize_free_range(addr, false); 8250 } else { 8251 // this will be swept up when we hit the end of the 8252 // free range 8253 if (CMSTraceSweeper) { 8254 gclog_or_tty->print(" -- pick up garbage 0x%x (%d) \n", fc, size); 8255 } 8256 // If the chunk is being coalesced and the current free range is 8257 // in the free lists, remove the current free range so that it 8258 // will be returned to the free lists in its entirety - all 8259 // the coalesced pieces included. 8260 if (freeRangeInFreeLists()) { 8261 FreeChunk* ffc = (FreeChunk*)freeFinger(); 8262 assert(ffc->size() == pointer_delta(addr, freeFinger()), 8263 "Size of free range is inconsistent with chunk size."); 8264 if (CMSTestInFreeList) { 8265 assert(_sp->verify_chunk_in_free_list(ffc), 8266 "free range is not in free lists"); 8267 } 8268 _sp->removeFreeChunkFromFreeLists(ffc); 8269 set_freeRangeInFreeLists(false); 8270 } 8271 set_lastFreeRangeCoalesced(true); 8272 } 8273 // this will be swept up when we hit the end of the free range 8274 8275 // Verify that the bit map has no bits marked between 8276 // addr and purported end of just dead object. 8277 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 8278 } 8279 assert(_limit >= addr + size, 8280 "A freshly garbage chunk can't possibly straddle over _limit"); 8281 if (inFreeRange()) lookahead_and_flush(fc, size); 8282 return size; 8283 } 8284 8285 size_t SweepClosure::do_live_chunk(FreeChunk* fc) { 8286 HeapWord* addr = (HeapWord*) fc; 8287 // The sweeper has just found a live object. Return any accumulated 8288 // left hand chunk to the free lists. 8289 if (inFreeRange()) { 8290 assert(freeFinger() < addr, "freeFinger points too high"); 8291 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 8292 } 8293 8294 // This object is live: we'd normally expect this to be 8295 // an oop, and like to assert the following: 8296 // assert(oop(addr)->is_oop(), "live block should be an oop"); 8297 // However, as we commented above, this may be an object whose 8298 // header hasn't yet been initialized. 8299 size_t size; 8300 assert(_bitMap->isMarked(addr), "Tautology for this control point"); 8301 if (_bitMap->isMarked(addr + 1)) { 8302 // Determine the size from the bit map, rather than trying to 8303 // compute it from the object header. 8304 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 8305 size = pointer_delta(nextOneAddr + 1, addr); 8306 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 8307 "alignment problem"); 8308 8309 #ifdef DEBUG 8310 if (oop(addr)->klass_or_null() != NULL && 8311 ( !_collector->should_unload_classes() 8312 || (oop(addr)->is_parsable()) && 8313 oop(addr)->is_conc_safe())) { 8314 // Ignore mark word because we are running concurrent with mutators 8315 assert(oop(addr)->is_oop(true), "live block should be an oop"); 8316 // is_conc_safe is checked before performing this assertion 8317 // because an object that is not is_conc_safe may yet have 8318 // the return from size() correct. 8319 assert(size == 8320 CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()), 8321 "P-mark and computed size do not agree"); 8322 } 8323 #endif 8324 8325 } else { 8326 // This should be an initialized object that's alive. 8327 assert(oop(addr)->klass_or_null() != NULL && 8328 (!_collector->should_unload_classes() 8329 || oop(addr)->is_parsable()), 8330 "Should be an initialized object"); 8331 // Note that there are objects used during class redefinition, 8332 // e.g. merge_cp in VM_RedefineClasses::merge_cp_and_rewrite(), 8333 // which are discarded with their is_conc_safe state still 8334 // false. These object may be floating garbage so may be 8335 // seen here. If they are floating garbage their size 8336 // should be attainable from their klass. Do not that 8337 // is_conc_safe() is true for oop(addr). 8338 // Ignore mark word because we are running concurrent with mutators 8339 assert(oop(addr)->is_oop(true), "live block should be an oop"); 8340 // Verify that the bit map has no bits marked between 8341 // addr and purported end of this block. 8342 size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 8343 assert(size >= 3, "Necessary for Printezis marks to work"); 8344 assert(!_bitMap->isMarked(addr+1), "Tautology for this control point"); 8345 DEBUG_ONLY(_bitMap->verifyNoOneBitsInRange(addr+2, addr+size);) 8346 } 8347 return size; 8348 } 8349 8350 void SweepClosure::do_post_free_or_garbage_chunk(FreeChunk* fc, 8351 size_t chunkSize) { 8352 // do_post_free_or_garbage_chunk() should only be called in the case 8353 // of the adaptive free list allocator. 8354 const bool fcInFreeLists = fc->is_free(); 8355 assert(_sp->adaptive_freelists(), "Should only be used in this case."); 8356 assert((HeapWord*)fc <= _limit, "sweep invariant"); 8357 if (CMSTestInFreeList && fcInFreeLists) { 8358 assert(_sp->verify_chunk_in_free_list(fc), "free chunk is not in free lists"); 8359 } 8360 8361 if (CMSTraceSweeper) { 8362 gclog_or_tty->print_cr(" -- pick up another chunk at 0x%x (%d)", fc, chunkSize); 8363 } 8364 8365 HeapWord* const fc_addr = (HeapWord*) fc; 8366 8367 bool coalesce; 8368 const size_t left = pointer_delta(fc_addr, freeFinger()); 8369 const size_t right = chunkSize; 8370 switch (FLSCoalescePolicy) { 8371 // numeric value forms a coalition aggressiveness metric 8372 case 0: { // never coalesce 8373 coalesce = false; 8374 break; 8375 } 8376 case 1: { // coalesce if left & right chunks on overpopulated lists 8377 coalesce = _sp->coalOverPopulated(left) && 8378 _sp->coalOverPopulated(right); 8379 break; 8380 } 8381 case 2: { // coalesce if left chunk on overpopulated list (default) 8382 coalesce = _sp->coalOverPopulated(left); 8383 break; 8384 } 8385 case 3: { // coalesce if left OR right chunk on overpopulated list 8386 coalesce = _sp->coalOverPopulated(left) || 8387 _sp->coalOverPopulated(right); 8388 break; 8389 } 8390 case 4: { // always coalesce 8391 coalesce = true; 8392 break; 8393 } 8394 default: 8395 ShouldNotReachHere(); 8396 } 8397 8398 // Should the current free range be coalesced? 8399 // If the chunk is in a free range and either we decided to coalesce above 8400 // or the chunk is near the large block at the end of the heap 8401 // (isNearLargestChunk() returns true), then coalesce this chunk. 8402 const bool doCoalesce = inFreeRange() 8403 && (coalesce || _g->isNearLargestChunk(fc_addr)); 8404 if (doCoalesce) { 8405 // Coalesce the current free range on the left with the new 8406 // chunk on the right. If either is on a free list, 8407 // it must be removed from the list and stashed in the closure. 8408 if (freeRangeInFreeLists()) { 8409 FreeChunk* const ffc = (FreeChunk*)freeFinger(); 8410 assert(ffc->size() == pointer_delta(fc_addr, freeFinger()), 8411 "Size of free range is inconsistent with chunk size."); 8412 if (CMSTestInFreeList) { 8413 assert(_sp->verify_chunk_in_free_list(ffc), 8414 "Chunk is not in free lists"); 8415 } 8416 _sp->coalDeath(ffc->size()); 8417 _sp->removeFreeChunkFromFreeLists(ffc); 8418 set_freeRangeInFreeLists(false); 8419 } 8420 if (fcInFreeLists) { 8421 _sp->coalDeath(chunkSize); 8422 assert(fc->size() == chunkSize, 8423 "The chunk has the wrong size or is not in the free lists"); 8424 _sp->removeFreeChunkFromFreeLists(fc); 8425 } 8426 set_lastFreeRangeCoalesced(true); 8427 print_free_block_coalesced(fc); 8428 } else { // not in a free range and/or should not coalesce 8429 // Return the current free range and start a new one. 8430 if (inFreeRange()) { 8431 // In a free range but cannot coalesce with the right hand chunk. 8432 // Put the current free range into the free lists. 8433 flush_cur_free_chunk(freeFinger(), 8434 pointer_delta(fc_addr, freeFinger())); 8435 } 8436 // Set up for new free range. Pass along whether the right hand 8437 // chunk is in the free lists. 8438 initialize_free_range((HeapWord*)fc, fcInFreeLists); 8439 } 8440 } 8441 8442 // Lookahead flush: 8443 // If we are tracking a free range, and this is the last chunk that 8444 // we'll look at because its end crosses past _limit, we'll preemptively 8445 // flush it along with any free range we may be holding on to. Note that 8446 // this can be the case only for an already free or freshly garbage 8447 // chunk. If this block is an object, it can never straddle 8448 // over _limit. The "straddling" occurs when _limit is set at 8449 // the previous end of the space when this cycle started, and 8450 // a subsequent heap expansion caused the previously co-terminal 8451 // free block to be coalesced with the newly expanded portion, 8452 // thus rendering _limit a non-block-boundary making it dangerous 8453 // for the sweeper to step over and examine. 8454 void SweepClosure::lookahead_and_flush(FreeChunk* fc, size_t chunk_size) { 8455 assert(inFreeRange(), "Should only be called if currently in a free range."); 8456 HeapWord* const eob = ((HeapWord*)fc) + chunk_size; 8457 assert(_sp->used_region().contains(eob - 1), 8458 err_msg("eob = " PTR_FORMAT " out of bounds wrt _sp = [" PTR_FORMAT "," PTR_FORMAT ")" 8459 " when examining fc = " PTR_FORMAT "(" SIZE_FORMAT ")", 8460 _limit, _sp->bottom(), _sp->end(), fc, chunk_size)); 8461 if (eob >= _limit) { 8462 assert(eob == _limit || fc->is_free(), "Only a free chunk should allow us to cross over the limit"); 8463 if (CMSTraceSweeper) { 8464 gclog_or_tty->print_cr("_limit " PTR_FORMAT " reached or crossed by block " 8465 "[" PTR_FORMAT "," PTR_FORMAT ") in space " 8466 "[" PTR_FORMAT "," PTR_FORMAT ")", 8467 _limit, fc, eob, _sp->bottom(), _sp->end()); 8468 } 8469 // Return the storage we are tracking back into the free lists. 8470 if (CMSTraceSweeper) { 8471 gclog_or_tty->print_cr("Flushing ... "); 8472 } 8473 assert(freeFinger() < eob, "Error"); 8474 flush_cur_free_chunk( freeFinger(), pointer_delta(eob, freeFinger())); 8475 } 8476 } 8477 8478 void SweepClosure::flush_cur_free_chunk(HeapWord* chunk, size_t size) { 8479 assert(inFreeRange(), "Should only be called if currently in a free range."); 8480 assert(size > 0, 8481 "A zero sized chunk cannot be added to the free lists."); 8482 if (!freeRangeInFreeLists()) { 8483 if (CMSTestInFreeList) { 8484 FreeChunk* fc = (FreeChunk*) chunk; 8485 fc->set_size(size); 8486 assert(!_sp->verify_chunk_in_free_list(fc), 8487 "chunk should not be in free lists yet"); 8488 } 8489 if (CMSTraceSweeper) { 8490 gclog_or_tty->print_cr(" -- add free block 0x%x (%d) to free lists", 8491 chunk, size); 8492 } 8493 // A new free range is going to be starting. The current 8494 // free range has not been added to the free lists yet or 8495 // was removed so add it back. 8496 // If the current free range was coalesced, then the death 8497 // of the free range was recorded. Record a birth now. 8498 if (lastFreeRangeCoalesced()) { 8499 _sp->coalBirth(size); 8500 } 8501 _sp->addChunkAndRepairOffsetTable(chunk, size, 8502 lastFreeRangeCoalesced()); 8503 } else if (CMSTraceSweeper) { 8504 gclog_or_tty->print_cr("Already in free list: nothing to flush"); 8505 } 8506 set_inFreeRange(false); 8507 set_freeRangeInFreeLists(false); 8508 } 8509 8510 // We take a break if we've been at this for a while, 8511 // so as to avoid monopolizing the locks involved. 8512 void SweepClosure::do_yield_work(HeapWord* addr) { 8513 // Return current free chunk being used for coalescing (if any) 8514 // to the appropriate freelist. After yielding, the next 8515 // free block encountered will start a coalescing range of 8516 // free blocks. If the next free block is adjacent to the 8517 // chunk just flushed, they will need to wait for the next 8518 // sweep to be coalesced. 8519 if (inFreeRange()) { 8520 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 8521 } 8522 8523 // First give up the locks, then yield, then re-lock. 8524 // We should probably use a constructor/destructor idiom to 8525 // do this unlock/lock or modify the MutexUnlocker class to 8526 // serve our purpose. XXX 8527 assert_lock_strong(_bitMap->lock()); 8528 assert_lock_strong(_freelistLock); 8529 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 8530 "CMS thread should hold CMS token"); 8531 _bitMap->lock()->unlock(); 8532 _freelistLock->unlock(); 8533 ConcurrentMarkSweepThread::desynchronize(true); 8534 ConcurrentMarkSweepThread::acknowledge_yield_request(); 8535 _collector->stopTimer(); 8536 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 8537 if (PrintCMSStatistics != 0) { 8538 _collector->incrementYields(); 8539 } 8540 _collector->icms_wait(); 8541 8542 // See the comment in coordinator_yield() 8543 for (unsigned i = 0; i < CMSYieldSleepCount && 8544 ConcurrentMarkSweepThread::should_yield() && 8545 !CMSCollector::foregroundGCIsActive(); ++i) { 8546 os::sleep(Thread::current(), 1, false); 8547 ConcurrentMarkSweepThread::acknowledge_yield_request(); 8548 } 8549 8550 ConcurrentMarkSweepThread::synchronize(true); 8551 _freelistLock->lock(); 8552 _bitMap->lock()->lock_without_safepoint_check(); 8553 _collector->startTimer(); 8554 } 8555 8556 #ifndef PRODUCT 8557 // This is actually very useful in a product build if it can 8558 // be called from the debugger. Compile it into the product 8559 // as needed. 8560 bool debug_verify_chunk_in_free_list(FreeChunk* fc) { 8561 return debug_cms_space->verify_chunk_in_free_list(fc); 8562 } 8563 #endif 8564 8565 void SweepClosure::print_free_block_coalesced(FreeChunk* fc) const { 8566 if (CMSTraceSweeper) { 8567 gclog_or_tty->print_cr("Sweep:coal_free_blk " PTR_FORMAT " (" SIZE_FORMAT ")", 8568 fc, fc->size()); 8569 } 8570 } 8571 8572 // CMSIsAliveClosure 8573 bool CMSIsAliveClosure::do_object_b(oop obj) { 8574 HeapWord* addr = (HeapWord*)obj; 8575 return addr != NULL && 8576 (!_span.contains(addr) || _bit_map->isMarked(addr)); 8577 } 8578 8579 CMSKeepAliveClosure::CMSKeepAliveClosure( CMSCollector* collector, 8580 MemRegion span, 8581 CMSBitMap* bit_map, CMSMarkStack* mark_stack, 8582 CMSMarkStack* revisit_stack, bool cpc): 8583 KlassRememberingOopClosure(collector, NULL, revisit_stack), 8584 _span(span), 8585 _bit_map(bit_map), 8586 _mark_stack(mark_stack), 8587 _concurrent_precleaning(cpc) { 8588 assert(!_span.is_empty(), "Empty span could spell trouble"); 8589 } 8590 8591 8592 // CMSKeepAliveClosure: the serial version 8593 void CMSKeepAliveClosure::do_oop(oop obj) { 8594 HeapWord* addr = (HeapWord*)obj; 8595 if (_span.contains(addr) && 8596 !_bit_map->isMarked(addr)) { 8597 _bit_map->mark(addr); 8598 bool simulate_overflow = false; 8599 NOT_PRODUCT( 8600 if (CMSMarkStackOverflowALot && 8601 _collector->simulate_overflow()) { 8602 // simulate a stack overflow 8603 simulate_overflow = true; 8604 } 8605 ) 8606 if (simulate_overflow || !_mark_stack->push(obj)) { 8607 if (_concurrent_precleaning) { 8608 // We dirty the overflown object and let the remark 8609 // phase deal with it. 8610 assert(_collector->overflow_list_is_empty(), "Error"); 8611 // In the case of object arrays, we need to dirty all of 8612 // the cards that the object spans. No locking or atomics 8613 // are needed since no one else can be mutating the mod union 8614 // table. 8615 if (obj->is_objArray()) { 8616 size_t sz = obj->size(); 8617 HeapWord* end_card_addr = 8618 (HeapWord*)round_to((intptr_t)(addr+sz), CardTableModRefBS::card_size); 8619 MemRegion redirty_range = MemRegion(addr, end_card_addr); 8620 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 8621 _collector->_modUnionTable.mark_range(redirty_range); 8622 } else { 8623 _collector->_modUnionTable.mark(addr); 8624 } 8625 _collector->_ser_kac_preclean_ovflw++; 8626 } else { 8627 _collector->push_on_overflow_list(obj); 8628 _collector->_ser_kac_ovflw++; 8629 } 8630 } 8631 } 8632 } 8633 8634 void CMSKeepAliveClosure::do_oop(oop* p) { CMSKeepAliveClosure::do_oop_work(p); } 8635 void CMSKeepAliveClosure::do_oop(narrowOop* p) { CMSKeepAliveClosure::do_oop_work(p); } 8636 8637 // CMSParKeepAliveClosure: a parallel version of the above. 8638 // The work queues are private to each closure (thread), 8639 // but (may be) available for stealing by other threads. 8640 void CMSParKeepAliveClosure::do_oop(oop obj) { 8641 HeapWord* addr = (HeapWord*)obj; 8642 if (_span.contains(addr) && 8643 !_bit_map->isMarked(addr)) { 8644 // In general, during recursive tracing, several threads 8645 // may be concurrently getting here; the first one to 8646 // "tag" it, claims it. 8647 if (_bit_map->par_mark(addr)) { 8648 bool res = _work_queue->push(obj); 8649 assert(res, "Low water mark should be much less than capacity"); 8650 // Do a recursive trim in the hope that this will keep 8651 // stack usage lower, but leave some oops for potential stealers 8652 trim_queue(_low_water_mark); 8653 } // Else, another thread got there first 8654 } 8655 } 8656 8657 void CMSParKeepAliveClosure::do_oop(oop* p) { CMSParKeepAliveClosure::do_oop_work(p); } 8658 void CMSParKeepAliveClosure::do_oop(narrowOop* p) { CMSParKeepAliveClosure::do_oop_work(p); } 8659 8660 void CMSParKeepAliveClosure::trim_queue(uint max) { 8661 while (_work_queue->size() > max) { 8662 oop new_oop; 8663 if (_work_queue->pop_local(new_oop)) { 8664 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 8665 assert(_bit_map->isMarked((HeapWord*)new_oop), 8666 "no white objects on this stack!"); 8667 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 8668 // iterate over the oops in this oop, marking and pushing 8669 // the ones in CMS heap (i.e. in _span). 8670 new_oop->oop_iterate(&_mark_and_push); 8671 } 8672 } 8673 } 8674 8675 CMSInnerParMarkAndPushClosure::CMSInnerParMarkAndPushClosure( 8676 CMSCollector* collector, 8677 MemRegion span, CMSBitMap* bit_map, 8678 CMSMarkStack* revisit_stack, 8679 OopTaskQueue* work_queue): 8680 Par_KlassRememberingOopClosure(collector, NULL, revisit_stack), 8681 _span(span), 8682 _bit_map(bit_map), 8683 _work_queue(work_queue) { } 8684 8685 void CMSInnerParMarkAndPushClosure::do_oop(oop obj) { 8686 HeapWord* addr = (HeapWord*)obj; 8687 if (_span.contains(addr) && 8688 !_bit_map->isMarked(addr)) { 8689 if (_bit_map->par_mark(addr)) { 8690 bool simulate_overflow = false; 8691 NOT_PRODUCT( 8692 if (CMSMarkStackOverflowALot && 8693 _collector->par_simulate_overflow()) { 8694 // simulate a stack overflow 8695 simulate_overflow = true; 8696 } 8697 ) 8698 if (simulate_overflow || !_work_queue->push(obj)) { 8699 _collector->par_push_on_overflow_list(obj); 8700 _collector->_par_kac_ovflw++; 8701 } 8702 } // Else another thread got there already 8703 } 8704 } 8705 8706 void CMSInnerParMarkAndPushClosure::do_oop(oop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); } 8707 void CMSInnerParMarkAndPushClosure::do_oop(narrowOop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); } 8708 8709 ////////////////////////////////////////////////////////////////// 8710 // CMSExpansionCause ///////////////////////////// 8711 ////////////////////////////////////////////////////////////////// 8712 const char* CMSExpansionCause::to_string(CMSExpansionCause::Cause cause) { 8713 switch (cause) { 8714 case _no_expansion: 8715 return "No expansion"; 8716 case _satisfy_free_ratio: 8717 return "Free ratio"; 8718 case _satisfy_promotion: 8719 return "Satisfy promotion"; 8720 case _satisfy_allocation: 8721 return "allocation"; 8722 case _allocate_par_lab: 8723 return "Par LAB"; 8724 case _allocate_par_spooling_space: 8725 return "Par Spooling Space"; 8726 case _adaptive_size_policy: 8727 return "Ergonomics"; 8728 default: 8729 return "unknown"; 8730 } 8731 } 8732 8733 void CMSDrainMarkingStackClosure::do_void() { 8734 // the max number to take from overflow list at a time 8735 const size_t num = _mark_stack->capacity()/4; 8736 assert(!_concurrent_precleaning || _collector->overflow_list_is_empty(), 8737 "Overflow list should be NULL during concurrent phases"); 8738 while (!_mark_stack->isEmpty() || 8739 // if stack is empty, check the overflow list 8740 _collector->take_from_overflow_list(num, _mark_stack)) { 8741 oop obj = _mark_stack->pop(); 8742 HeapWord* addr = (HeapWord*)obj; 8743 assert(_span.contains(addr), "Should be within span"); 8744 assert(_bit_map->isMarked(addr), "Should be marked"); 8745 assert(obj->is_oop(), "Should be an oop"); 8746 obj->oop_iterate(_keep_alive); 8747 } 8748 } 8749 8750 void CMSParDrainMarkingStackClosure::do_void() { 8751 // drain queue 8752 trim_queue(0); 8753 } 8754 8755 // Trim our work_queue so its length is below max at return 8756 void CMSParDrainMarkingStackClosure::trim_queue(uint max) { 8757 while (_work_queue->size() > max) { 8758 oop new_oop; 8759 if (_work_queue->pop_local(new_oop)) { 8760 assert(new_oop->is_oop(), "Expected an oop"); 8761 assert(_bit_map->isMarked((HeapWord*)new_oop), 8762 "no white objects on this stack!"); 8763 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 8764 // iterate over the oops in this oop, marking and pushing 8765 // the ones in CMS heap (i.e. in _span). 8766 new_oop->oop_iterate(&_mark_and_push); 8767 } 8768 } 8769 } 8770 8771 //////////////////////////////////////////////////////////////////// 8772 // Support for Marking Stack Overflow list handling and related code 8773 //////////////////////////////////////////////////////////////////// 8774 // Much of the following code is similar in shape and spirit to the 8775 // code used in ParNewGC. We should try and share that code 8776 // as much as possible in the future. 8777 8778 #ifndef PRODUCT 8779 // Debugging support for CMSStackOverflowALot 8780 8781 // It's OK to call this multi-threaded; the worst thing 8782 // that can happen is that we'll get a bunch of closely 8783 // spaced simulated oveflows, but that's OK, in fact 8784 // probably good as it would exercise the overflow code 8785 // under contention. 8786 bool CMSCollector::simulate_overflow() { 8787 if (_overflow_counter-- <= 0) { // just being defensive 8788 _overflow_counter = CMSMarkStackOverflowInterval; 8789 return true; 8790 } else { 8791 return false; 8792 } 8793 } 8794 8795 bool CMSCollector::par_simulate_overflow() { 8796 return simulate_overflow(); 8797 } 8798 #endif 8799 8800 // Single-threaded 8801 bool CMSCollector::take_from_overflow_list(size_t num, CMSMarkStack* stack) { 8802 assert(stack->isEmpty(), "Expected precondition"); 8803 assert(stack->capacity() > num, "Shouldn't bite more than can chew"); 8804 size_t i = num; 8805 oop cur = _overflow_list; 8806 const markOop proto = markOopDesc::prototype(); 8807 NOT_PRODUCT(ssize_t n = 0;) 8808 for (oop next; i > 0 && cur != NULL; cur = next, i--) { 8809 next = oop(cur->mark()); 8810 cur->set_mark(proto); // until proven otherwise 8811 assert(cur->is_oop(), "Should be an oop"); 8812 bool res = stack->push(cur); 8813 assert(res, "Bit off more than can chew?"); 8814 NOT_PRODUCT(n++;) 8815 } 8816 _overflow_list = cur; 8817 #ifndef PRODUCT 8818 assert(_num_par_pushes >= n, "Too many pops?"); 8819 _num_par_pushes -=n; 8820 #endif 8821 return !stack->isEmpty(); 8822 } 8823 8824 #define BUSY (oop(0x1aff1aff)) 8825 // (MT-safe) Get a prefix of at most "num" from the list. 8826 // The overflow list is chained through the mark word of 8827 // each object in the list. We fetch the entire list, 8828 // break off a prefix of the right size and return the 8829 // remainder. If other threads try to take objects from 8830 // the overflow list at that time, they will wait for 8831 // some time to see if data becomes available. If (and 8832 // only if) another thread places one or more object(s) 8833 // on the global list before we have returned the suffix 8834 // to the global list, we will walk down our local list 8835 // to find its end and append the global list to 8836 // our suffix before returning it. This suffix walk can 8837 // prove to be expensive (quadratic in the amount of traffic) 8838 // when there are many objects in the overflow list and 8839 // there is much producer-consumer contention on the list. 8840 // *NOTE*: The overflow list manipulation code here and 8841 // in ParNewGeneration:: are very similar in shape, 8842 // except that in the ParNew case we use the old (from/eden) 8843 // copy of the object to thread the list via its klass word. 8844 // Because of the common code, if you make any changes in 8845 // the code below, please check the ParNew version to see if 8846 // similar changes might be needed. 8847 // CR 6797058 has been filed to consolidate the common code. 8848 bool CMSCollector::par_take_from_overflow_list(size_t num, 8849 OopTaskQueue* work_q, 8850 int no_of_gc_threads) { 8851 assert(work_q->size() == 0, "First empty local work queue"); 8852 assert(num < work_q->max_elems(), "Can't bite more than we can chew"); 8853 if (_overflow_list == NULL) { 8854 return false; 8855 } 8856 // Grab the entire list; we'll put back a suffix 8857 oop prefix = (oop)Atomic::xchg_ptr(BUSY, &_overflow_list); 8858 Thread* tid = Thread::current(); 8859 // Before "no_of_gc_threads" was introduced CMSOverflowSpinCount was 8860 // set to ParallelGCThreads. 8861 size_t CMSOverflowSpinCount = (size_t) no_of_gc_threads; // was ParallelGCThreads; 8862 size_t sleep_time_millis = MAX2((size_t)1, num/100); 8863 // If the list is busy, we spin for a short while, 8864 // sleeping between attempts to get the list. 8865 for (size_t spin = 0; prefix == BUSY && spin < CMSOverflowSpinCount; spin++) { 8866 os::sleep(tid, sleep_time_millis, false); 8867 if (_overflow_list == NULL) { 8868 // Nothing left to take 8869 return false; 8870 } else if (_overflow_list != BUSY) { 8871 // Try and grab the prefix 8872 prefix = (oop)Atomic::xchg_ptr(BUSY, &_overflow_list); 8873 } 8874 } 8875 // If the list was found to be empty, or we spun long 8876 // enough, we give up and return empty-handed. If we leave 8877 // the list in the BUSY state below, it must be the case that 8878 // some other thread holds the overflow list and will set it 8879 // to a non-BUSY state in the future. 8880 if (prefix == NULL || prefix == BUSY) { 8881 // Nothing to take or waited long enough 8882 if (prefix == NULL) { 8883 // Write back the NULL in case we overwrote it with BUSY above 8884 // and it is still the same value. 8885 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY); 8886 } 8887 return false; 8888 } 8889 assert(prefix != NULL && prefix != BUSY, "Error"); 8890 size_t i = num; 8891 oop cur = prefix; 8892 // Walk down the first "num" objects, unless we reach the end. 8893 for (; i > 1 && cur->mark() != NULL; cur = oop(cur->mark()), i--); 8894 if (cur->mark() == NULL) { 8895 // We have "num" or fewer elements in the list, so there 8896 // is nothing to return to the global list. 8897 // Write back the NULL in lieu of the BUSY we wrote 8898 // above, if it is still the same value. 8899 if (_overflow_list == BUSY) { 8900 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY); 8901 } 8902 } else { 8903 // Chop off the suffix and rerturn it to the global list. 8904 assert(cur->mark() != BUSY, "Error"); 8905 oop suffix_head = cur->mark(); // suffix will be put back on global list 8906 cur->set_mark(NULL); // break off suffix 8907 // It's possible that the list is still in the empty(busy) state 8908 // we left it in a short while ago; in that case we may be 8909 // able to place back the suffix without incurring the cost 8910 // of a walk down the list. 8911 oop observed_overflow_list = _overflow_list; 8912 oop cur_overflow_list = observed_overflow_list; 8913 bool attached = false; 8914 while (observed_overflow_list == BUSY || observed_overflow_list == NULL) { 8915 observed_overflow_list = 8916 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list); 8917 if (cur_overflow_list == observed_overflow_list) { 8918 attached = true; 8919 break; 8920 } else cur_overflow_list = observed_overflow_list; 8921 } 8922 if (!attached) { 8923 // Too bad, someone else sneaked in (at least) an element; we'll need 8924 // to do a splice. Find tail of suffix so we can prepend suffix to global 8925 // list. 8926 for (cur = suffix_head; cur->mark() != NULL; cur = (oop)(cur->mark())); 8927 oop suffix_tail = cur; 8928 assert(suffix_tail != NULL && suffix_tail->mark() == NULL, 8929 "Tautology"); 8930 observed_overflow_list = _overflow_list; 8931 do { 8932 cur_overflow_list = observed_overflow_list; 8933 if (cur_overflow_list != BUSY) { 8934 // Do the splice ... 8935 suffix_tail->set_mark(markOop(cur_overflow_list)); 8936 } else { // cur_overflow_list == BUSY 8937 suffix_tail->set_mark(NULL); 8938 } 8939 // ... and try to place spliced list back on overflow_list ... 8940 observed_overflow_list = 8941 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list); 8942 } while (cur_overflow_list != observed_overflow_list); 8943 // ... until we have succeeded in doing so. 8944 } 8945 } 8946 8947 // Push the prefix elements on work_q 8948 assert(prefix != NULL, "control point invariant"); 8949 const markOop proto = markOopDesc::prototype(); 8950 oop next; 8951 NOT_PRODUCT(ssize_t n = 0;) 8952 for (cur = prefix; cur != NULL; cur = next) { 8953 next = oop(cur->mark()); 8954 cur->set_mark(proto); // until proven otherwise 8955 assert(cur->is_oop(), "Should be an oop"); 8956 bool res = work_q->push(cur); 8957 assert(res, "Bit off more than we can chew?"); 8958 NOT_PRODUCT(n++;) 8959 } 8960 #ifndef PRODUCT 8961 assert(_num_par_pushes >= n, "Too many pops?"); 8962 Atomic::add_ptr(-(intptr_t)n, &_num_par_pushes); 8963 #endif 8964 return true; 8965 } 8966 8967 // Single-threaded 8968 void CMSCollector::push_on_overflow_list(oop p) { 8969 NOT_PRODUCT(_num_par_pushes++;) 8970 assert(p->is_oop(), "Not an oop"); 8971 preserve_mark_if_necessary(p); 8972 p->set_mark((markOop)_overflow_list); 8973 _overflow_list = p; 8974 } 8975 8976 // Multi-threaded; use CAS to prepend to overflow list 8977 void CMSCollector::par_push_on_overflow_list(oop p) { 8978 NOT_PRODUCT(Atomic::inc_ptr(&_num_par_pushes);) 8979 assert(p->is_oop(), "Not an oop"); 8980 par_preserve_mark_if_necessary(p); 8981 oop observed_overflow_list = _overflow_list; 8982 oop cur_overflow_list; 8983 do { 8984 cur_overflow_list = observed_overflow_list; 8985 if (cur_overflow_list != BUSY) { 8986 p->set_mark(markOop(cur_overflow_list)); 8987 } else { 8988 p->set_mark(NULL); 8989 } 8990 observed_overflow_list = 8991 (oop) Atomic::cmpxchg_ptr(p, &_overflow_list, cur_overflow_list); 8992 } while (cur_overflow_list != observed_overflow_list); 8993 } 8994 #undef BUSY 8995 8996 // Single threaded 8997 // General Note on GrowableArray: pushes may silently fail 8998 // because we are (temporarily) out of C-heap for expanding 8999 // the stack. The problem is quite ubiquitous and affects 9000 // a lot of code in the JVM. The prudent thing for GrowableArray 9001 // to do (for now) is to exit with an error. However, that may 9002 // be too draconian in some cases because the caller may be 9003 // able to recover without much harm. For such cases, we 9004 // should probably introduce a "soft_push" method which returns 9005 // an indication of success or failure with the assumption that 9006 // the caller may be able to recover from a failure; code in 9007 // the VM can then be changed, incrementally, to deal with such 9008 // failures where possible, thus, incrementally hardening the VM 9009 // in such low resource situations. 9010 void CMSCollector::preserve_mark_work(oop p, markOop m) { 9011 _preserved_oop_stack.push(p); 9012 _preserved_mark_stack.push(m); 9013 assert(m == p->mark(), "Mark word changed"); 9014 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), 9015 "bijection"); 9016 } 9017 9018 // Single threaded 9019 void CMSCollector::preserve_mark_if_necessary(oop p) { 9020 markOop m = p->mark(); 9021 if (m->must_be_preserved(p)) { 9022 preserve_mark_work(p, m); 9023 } 9024 } 9025 9026 void CMSCollector::par_preserve_mark_if_necessary(oop p) { 9027 markOop m = p->mark(); 9028 if (m->must_be_preserved(p)) { 9029 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag); 9030 // Even though we read the mark word without holding 9031 // the lock, we are assured that it will not change 9032 // because we "own" this oop, so no other thread can 9033 // be trying to push it on the overflow list; see 9034 // the assertion in preserve_mark_work() that checks 9035 // that m == p->mark(). 9036 preserve_mark_work(p, m); 9037 } 9038 } 9039 9040 // We should be able to do this multi-threaded, 9041 // a chunk of stack being a task (this is 9042 // correct because each oop only ever appears 9043 // once in the overflow list. However, it's 9044 // not very easy to completely overlap this with 9045 // other operations, so will generally not be done 9046 // until all work's been completed. Because we 9047 // expect the preserved oop stack (set) to be small, 9048 // it's probably fine to do this single-threaded. 9049 // We can explore cleverer concurrent/overlapped/parallel 9050 // processing of preserved marks if we feel the 9051 // need for this in the future. Stack overflow should 9052 // be so rare in practice and, when it happens, its 9053 // effect on performance so great that this will 9054 // likely just be in the noise anyway. 9055 void CMSCollector::restore_preserved_marks_if_any() { 9056 assert(SafepointSynchronize::is_at_safepoint(), 9057 "world should be stopped"); 9058 assert(Thread::current()->is_ConcurrentGC_thread() || 9059 Thread::current()->is_VM_thread(), 9060 "should be single-threaded"); 9061 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), 9062 "bijection"); 9063 9064 while (!_preserved_oop_stack.is_empty()) { 9065 oop p = _preserved_oop_stack.pop(); 9066 assert(p->is_oop(), "Should be an oop"); 9067 assert(_span.contains(p), "oop should be in _span"); 9068 assert(p->mark() == markOopDesc::prototype(), 9069 "Set when taken from overflow list"); 9070 markOop m = _preserved_mark_stack.pop(); 9071 p->set_mark(m); 9072 } 9073 assert(_preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(), 9074 "stacks were cleared above"); 9075 } 9076 9077 #ifndef PRODUCT 9078 bool CMSCollector::no_preserved_marks() const { 9079 return _preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(); 9080 } 9081 #endif 9082 9083 CMSAdaptiveSizePolicy* ASConcurrentMarkSweepGeneration::cms_size_policy() const 9084 { 9085 GenCollectedHeap* gch = (GenCollectedHeap*) GenCollectedHeap::heap(); 9086 CMSAdaptiveSizePolicy* size_policy = 9087 (CMSAdaptiveSizePolicy*) gch->gen_policy()->size_policy(); 9088 assert(size_policy->is_gc_cms_adaptive_size_policy(), 9089 "Wrong type for size policy"); 9090 return size_policy; 9091 } 9092 9093 void ASConcurrentMarkSweepGeneration::resize(size_t cur_promo_size, 9094 size_t desired_promo_size) { 9095 if (cur_promo_size < desired_promo_size) { 9096 size_t expand_bytes = desired_promo_size - cur_promo_size; 9097 if (PrintAdaptiveSizePolicy && Verbose) { 9098 gclog_or_tty->print_cr(" ASConcurrentMarkSweepGeneration::resize " 9099 "Expanding tenured generation by " SIZE_FORMAT " (bytes)", 9100 expand_bytes); 9101 } 9102 expand(expand_bytes, 9103 MinHeapDeltaBytes, 9104 CMSExpansionCause::_adaptive_size_policy); 9105 } else if (desired_promo_size < cur_promo_size) { 9106 size_t shrink_bytes = cur_promo_size - desired_promo_size; 9107 if (PrintAdaptiveSizePolicy && Verbose) { 9108 gclog_or_tty->print_cr(" ASConcurrentMarkSweepGeneration::resize " 9109 "Shrinking tenured generation by " SIZE_FORMAT " (bytes)", 9110 shrink_bytes); 9111 } 9112 shrink(shrink_bytes); 9113 } 9114 } 9115 9116 CMSGCAdaptivePolicyCounters* ASConcurrentMarkSweepGeneration::gc_adaptive_policy_counters() { 9117 GenCollectedHeap* gch = GenCollectedHeap::heap(); 9118 CMSGCAdaptivePolicyCounters* counters = 9119 (CMSGCAdaptivePolicyCounters*) gch->collector_policy()->counters(); 9120 assert(counters->kind() == GCPolicyCounters::CMSGCAdaptivePolicyCountersKind, 9121 "Wrong kind of counters"); 9122 return counters; 9123 } 9124 9125 9126 void ASConcurrentMarkSweepGeneration::update_counters() { 9127 if (UsePerfData) { 9128 _space_counters->update_all(); 9129 _gen_counters->update_all(); 9130 CMSGCAdaptivePolicyCounters* counters = gc_adaptive_policy_counters(); 9131 GenCollectedHeap* gch = GenCollectedHeap::heap(); 9132 CMSGCStats* gc_stats_l = (CMSGCStats*) gc_stats(); 9133 assert(gc_stats_l->kind() == GCStats::CMSGCStatsKind, 9134 "Wrong gc statistics type"); 9135 counters->update_counters(gc_stats_l); 9136 } 9137 } 9138 9139 void ASConcurrentMarkSweepGeneration::update_counters(size_t used) { 9140 if (UsePerfData) { 9141 _space_counters->update_used(used); 9142 _space_counters->update_capacity(); 9143 _gen_counters->update_all(); 9144 9145 CMSGCAdaptivePolicyCounters* counters = gc_adaptive_policy_counters(); 9146 GenCollectedHeap* gch = GenCollectedHeap::heap(); 9147 CMSGCStats* gc_stats_l = (CMSGCStats*) gc_stats(); 9148 assert(gc_stats_l->kind() == GCStats::CMSGCStatsKind, 9149 "Wrong gc statistics type"); 9150 counters->update_counters(gc_stats_l); 9151 } 9152 } 9153 9154 // The desired expansion delta is computed so that: 9155 // . desired free percentage or greater is used 9156 void ASConcurrentMarkSweepGeneration::compute_new_size() { 9157 assert_locked_or_safepoint(Heap_lock); 9158 9159 GenCollectedHeap* gch = (GenCollectedHeap*) GenCollectedHeap::heap(); 9160 9161 // If incremental collection failed, we just want to expand 9162 // to the limit. 9163 if (incremental_collection_failed()) { 9164 clear_incremental_collection_failed(); 9165 grow_to_reserved(); 9166 return; 9167 } 9168 9169 assert(UseAdaptiveSizePolicy, "Should be using adaptive sizing"); 9170 9171 assert(gch->kind() == CollectedHeap::GenCollectedHeap, 9172 "Wrong type of heap"); 9173 int prev_level = level() - 1; 9174 assert(prev_level >= 0, "The cms generation is the lowest generation"); 9175 Generation* prev_gen = gch->get_gen(prev_level); 9176 assert(prev_gen->kind() == Generation::ASParNew, 9177 "Wrong type of young generation"); 9178 ParNewGeneration* younger_gen = (ParNewGeneration*) prev_gen; 9179 size_t cur_eden = younger_gen->eden()->capacity(); 9180 CMSAdaptiveSizePolicy* size_policy = cms_size_policy(); 9181 size_t cur_promo = free(); 9182 size_policy->compute_tenured_generation_free_space(cur_promo, 9183 max_available(), 9184 cur_eden); 9185 resize(cur_promo, size_policy->promo_size()); 9186 9187 // Record the new size of the space in the cms generation 9188 // that is available for promotions. This is temporary. 9189 // It should be the desired promo size. 9190 size_policy->avg_cms_promo()->sample(free()); 9191 size_policy->avg_old_live()->sample(used()); 9192 9193 if (UsePerfData) { 9194 CMSGCAdaptivePolicyCounters* counters = gc_adaptive_policy_counters(); 9195 counters->update_cms_capacity_counter(capacity()); 9196 } 9197 } 9198 9199 void ASConcurrentMarkSweepGeneration::shrink_by(size_t desired_bytes) { 9200 assert_locked_or_safepoint(Heap_lock); 9201 assert_lock_strong(freelistLock()); 9202 HeapWord* old_end = _cmsSpace->end(); 9203 HeapWord* unallocated_start = _cmsSpace->unallocated_block(); 9204 assert(old_end >= unallocated_start, "Miscalculation of unallocated_start"); 9205 FreeChunk* chunk_at_end = find_chunk_at_end(); 9206 if (chunk_at_end == NULL) { 9207 // No room to shrink 9208 if (PrintGCDetails && Verbose) { 9209 gclog_or_tty->print_cr("No room to shrink: old_end " 9210 PTR_FORMAT " unallocated_start " PTR_FORMAT 9211 " chunk_at_end " PTR_FORMAT, 9212 old_end, unallocated_start, chunk_at_end); 9213 } 9214 return; 9215 } else { 9216 9217 // Find the chunk at the end of the space and determine 9218 // how much it can be shrunk. 9219 size_t shrinkable_size_in_bytes = chunk_at_end->size(); 9220 size_t aligned_shrinkable_size_in_bytes = 9221 align_size_down(shrinkable_size_in_bytes, os::vm_page_size()); 9222 assert(unallocated_start <= chunk_at_end->end(), 9223 "Inconsistent chunk at end of space"); 9224 size_t bytes = MIN2(desired_bytes, aligned_shrinkable_size_in_bytes); 9225 size_t word_size_before = heap_word_size(_virtual_space.committed_size()); 9226 9227 // Shrink the underlying space 9228 _virtual_space.shrink_by(bytes); 9229 if (PrintGCDetails && Verbose) { 9230 gclog_or_tty->print_cr("ConcurrentMarkSweepGeneration::shrink_by:" 9231 " desired_bytes " SIZE_FORMAT 9232 " shrinkable_size_in_bytes " SIZE_FORMAT 9233 " aligned_shrinkable_size_in_bytes " SIZE_FORMAT 9234 " bytes " SIZE_FORMAT, 9235 desired_bytes, shrinkable_size_in_bytes, 9236 aligned_shrinkable_size_in_bytes, bytes); 9237 gclog_or_tty->print_cr(" old_end " SIZE_FORMAT 9238 " unallocated_start " SIZE_FORMAT, 9239 old_end, unallocated_start); 9240 } 9241 9242 // If the space did shrink (shrinking is not guaranteed), 9243 // shrink the chunk at the end by the appropriate amount. 9244 if (((HeapWord*)_virtual_space.high()) < old_end) { 9245 size_t new_word_size = 9246 heap_word_size(_virtual_space.committed_size()); 9247 9248 // Have to remove the chunk from the dictionary because it is changing 9249 // size and might be someplace elsewhere in the dictionary. 9250 9251 // Get the chunk at end, shrink it, and put it 9252 // back. 9253 _cmsSpace->removeChunkFromDictionary(chunk_at_end); 9254 size_t word_size_change = word_size_before - new_word_size; 9255 size_t chunk_at_end_old_size = chunk_at_end->size(); 9256 assert(chunk_at_end_old_size >= word_size_change, 9257 "Shrink is too large"); 9258 chunk_at_end->set_size(chunk_at_end_old_size - 9259 word_size_change); 9260 _cmsSpace->freed((HeapWord*) chunk_at_end->end(), 9261 word_size_change); 9262 9263 _cmsSpace->returnChunkToDictionary(chunk_at_end); 9264 9265 MemRegion mr(_cmsSpace->bottom(), new_word_size); 9266 _bts->resize(new_word_size); // resize the block offset shared array 9267 Universe::heap()->barrier_set()->resize_covered_region(mr); 9268 _cmsSpace->assert_locked(); 9269 _cmsSpace->set_end((HeapWord*)_virtual_space.high()); 9270 9271 NOT_PRODUCT(_cmsSpace->dictionary()->verify()); 9272 9273 // update the space and generation capacity counters 9274 if (UsePerfData) { 9275 _space_counters->update_capacity(); 9276 _gen_counters->update_all(); 9277 } 9278 9279 if (Verbose && PrintGCDetails) { 9280 size_t new_mem_size = _virtual_space.committed_size(); 9281 size_t old_mem_size = new_mem_size + bytes; 9282 gclog_or_tty->print_cr("Shrinking %s from %ldK by %ldK to %ldK", 9283 name(), old_mem_size/K, bytes/K, new_mem_size/K); 9284 } 9285 } 9286 9287 assert(_cmsSpace->unallocated_block() <= _cmsSpace->end(), 9288 "Inconsistency at end of space"); 9289 assert(chunk_at_end->end() == _cmsSpace->end(), 9290 "Shrinking is inconsistent"); 9291 return; 9292 } 9293 } 9294 9295 // Transfer some number of overflown objects to usual marking 9296 // stack. Return true if some objects were transferred. 9297 bool MarkRefsIntoAndScanClosure::take_from_overflow_list() { 9298 size_t num = MIN2((size_t)(_mark_stack->capacity() - _mark_stack->length())/4, 9299 (size_t)ParGCDesiredObjsFromOverflowList); 9300 9301 bool res = _collector->take_from_overflow_list(num, _mark_stack); 9302 assert(_collector->overflow_list_is_empty() || res, 9303 "If list is not empty, we should have taken something"); 9304 assert(!res || !_mark_stack->isEmpty(), 9305 "If we took something, it should now be on our stack"); 9306 return res; 9307 } 9308 9309 size_t MarkDeadObjectsClosure::do_blk(HeapWord* addr) { 9310 size_t res = _sp->block_size_no_stall(addr, _collector); 9311 if (_sp->block_is_obj(addr)) { 9312 if (_live_bit_map->isMarked(addr)) { 9313 // It can't have been dead in a previous cycle 9314 guarantee(!_dead_bit_map->isMarked(addr), "No resurrection!"); 9315 } else { 9316 _dead_bit_map->mark(addr); // mark the dead object 9317 } 9318 } 9319 // Could be 0, if the block size could not be computed without stalling. 9320 return res; 9321 } 9322 9323 TraceCMSMemoryManagerStats::TraceCMSMemoryManagerStats(CMSCollector::CollectorState phase, GCCause::Cause cause): TraceMemoryManagerStats() { 9324 9325 switch (phase) { 9326 case CMSCollector::InitialMarking: 9327 initialize(true /* fullGC */ , 9328 cause /* cause of the GC */, 9329 true /* recordGCBeginTime */, 9330 true /* recordPreGCUsage */, 9331 false /* recordPeakUsage */, 9332 false /* recordPostGCusage */, 9333 true /* recordAccumulatedGCTime */, 9334 false /* recordGCEndTime */, 9335 false /* countCollection */ ); 9336 break; 9337 9338 case CMSCollector::FinalMarking: 9339 initialize(true /* fullGC */ , 9340 cause /* cause of the GC */, 9341 false /* recordGCBeginTime */, 9342 false /* recordPreGCUsage */, 9343 false /* recordPeakUsage */, 9344 false /* recordPostGCusage */, 9345 true /* recordAccumulatedGCTime */, 9346 false /* recordGCEndTime */, 9347 false /* countCollection */ ); 9348 break; 9349 9350 case CMSCollector::Sweeping: 9351 initialize(true /* fullGC */ , 9352 cause /* cause of the GC */, 9353 false /* recordGCBeginTime */, 9354 false /* recordPreGCUsage */, 9355 true /* recordPeakUsage */, 9356 true /* recordPostGCusage */, 9357 false /* recordAccumulatedGCTime */, 9358 true /* recordGCEndTime */, 9359 true /* countCollection */ ); 9360 break; 9361 9362 default: 9363 ShouldNotReachHere(); 9364 } 9365 } 9366