1 /* 2 * Copyright (c) 2001, 2013, 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/classLoaderData.hpp" 27 #include "classfile/symbolTable.hpp" 28 #include "classfile/systemDictionary.hpp" 29 #include "code/codeCache.hpp" 30 #include "gc_implementation/concurrentMarkSweep/cmsAdaptiveSizePolicy.hpp" 31 #include "gc_implementation/concurrentMarkSweep/cmsCollectorPolicy.hpp" 32 #include "gc_implementation/concurrentMarkSweep/cmsGCAdaptivePolicyCounters.hpp" 33 #include "gc_implementation/concurrentMarkSweep/cmsOopClosures.inline.hpp" 34 #include "gc_implementation/concurrentMarkSweep/compactibleFreeListSpace.hpp" 35 #include "gc_implementation/concurrentMarkSweep/concurrentMarkSweepGeneration.inline.hpp" 36 #include "gc_implementation/concurrentMarkSweep/concurrentMarkSweepThread.hpp" 37 #include "gc_implementation/concurrentMarkSweep/vmCMSOperations.hpp" 38 #include "gc_implementation/parNew/parNewGeneration.hpp" 39 #include "gc_implementation/shared/collectorCounters.hpp" 40 #include "gc_implementation/shared/gcTimer.hpp" 41 #include "gc_implementation/shared/gcTrace.hpp" 42 #include "gc_implementation/shared/gcTraceTime.hpp" 43 #include "gc_implementation/shared/isGCActiveMark.hpp" 44 #include "gc_interface/collectedHeap.inline.hpp" 45 #include "memory/allocation.hpp" 46 #include "memory/cardTableRS.hpp" 47 #include "memory/collectorPolicy.hpp" 48 #include "memory/gcLocker.inline.hpp" 49 #include "memory/genCollectedHeap.hpp" 50 #include "memory/genMarkSweep.hpp" 51 #include "memory/genOopClosures.inline.hpp" 52 #include "memory/iterator.hpp" 53 #include "memory/padded.hpp" 54 #include "memory/referencePolicy.hpp" 55 #include "memory/resourceArea.hpp" 56 #include "memory/tenuredGeneration.hpp" 57 #include "oops/oop.inline.hpp" 58 #include "prims/jvmtiExport.hpp" 59 #include "runtime/globals_extension.hpp" 60 #include "runtime/handles.inline.hpp" 61 #include "runtime/java.hpp" 62 #include "runtime/vmThread.hpp" 63 #include "services/memoryService.hpp" 64 #include "services/runtimeService.hpp" 65 66 // statics 67 CMSCollector* ConcurrentMarkSweepGeneration::_collector = NULL; 68 bool CMSCollector::_full_gc_requested = false; 69 GCCause::Cause CMSCollector::_full_gc_cause = GCCause::_no_gc; 70 71 ////////////////////////////////////////////////////////////////// 72 // In support of CMS/VM thread synchronization 73 ////////////////////////////////////////////////////////////////// 74 // We split use of the CGC_lock into 2 "levels". 75 // The low-level locking is of the usual CGC_lock monitor. We introduce 76 // a higher level "token" (hereafter "CMS token") built on top of the 77 // low level monitor (hereafter "CGC lock"). 78 // The token-passing protocol gives priority to the VM thread. The 79 // CMS-lock doesn't provide any fairness guarantees, but clients 80 // should ensure that it is only held for very short, bounded 81 // durations. 82 // 83 // When either of the CMS thread or the VM thread is involved in 84 // collection operations during which it does not want the other 85 // thread to interfere, it obtains the CMS token. 86 // 87 // If either thread tries to get the token while the other has 88 // it, that thread waits. However, if the VM thread and CMS thread 89 // both want the token, then the VM thread gets priority while the 90 // CMS thread waits. This ensures, for instance, that the "concurrent" 91 // phases of the CMS thread's work do not block out the VM thread 92 // for long periods of time as the CMS thread continues to hog 93 // the token. (See bug 4616232). 94 // 95 // The baton-passing functions are, however, controlled by the 96 // flags _foregroundGCShouldWait and _foregroundGCIsActive, 97 // and here the low-level CMS lock, not the high level token, 98 // ensures mutual exclusion. 99 // 100 // Two important conditions that we have to satisfy: 101 // 1. if a thread does a low-level wait on the CMS lock, then it 102 // relinquishes the CMS token if it were holding that token 103 // when it acquired the low-level CMS lock. 104 // 2. any low-level notifications on the low-level lock 105 // should only be sent when a thread has relinquished the token. 106 // 107 // In the absence of either property, we'd have potential deadlock. 108 // 109 // We protect each of the CMS (concurrent and sequential) phases 110 // with the CMS _token_, not the CMS _lock_. 111 // 112 // The only code protected by CMS lock is the token acquisition code 113 // itself, see ConcurrentMarkSweepThread::[de]synchronize(), and the 114 // baton-passing code. 115 // 116 // Unfortunately, i couldn't come up with a good abstraction to factor and 117 // hide the naked CGC_lock manipulation in the baton-passing code 118 // further below. That's something we should try to do. Also, the proof 119 // of correctness of this 2-level locking scheme is far from obvious, 120 // and potentially quite slippery. We have an uneasy suspicion, for instance, 121 // that there may be a theoretical possibility of delay/starvation in the 122 // low-level lock/wait/notify scheme used for the baton-passing because of 123 // potential interference with the priority scheme embodied in the 124 // CMS-token-passing protocol. See related comments at a CGC_lock->wait() 125 // invocation further below and marked with "XXX 20011219YSR". 126 // Indeed, as we note elsewhere, this may become yet more slippery 127 // in the presence of multiple CMS and/or multiple VM threads. XXX 128 129 class CMSTokenSync: public StackObj { 130 private: 131 bool _is_cms_thread; 132 public: 133 CMSTokenSync(bool is_cms_thread): 134 _is_cms_thread(is_cms_thread) { 135 assert(is_cms_thread == Thread::current()->is_ConcurrentGC_thread(), 136 "Incorrect argument to constructor"); 137 ConcurrentMarkSweepThread::synchronize(_is_cms_thread); 138 } 139 140 ~CMSTokenSync() { 141 assert(_is_cms_thread ? 142 ConcurrentMarkSweepThread::cms_thread_has_cms_token() : 143 ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 144 "Incorrect state"); 145 ConcurrentMarkSweepThread::desynchronize(_is_cms_thread); 146 } 147 }; 148 149 // Convenience class that does a CMSTokenSync, and then acquires 150 // upto three locks. 151 class CMSTokenSyncWithLocks: public CMSTokenSync { 152 private: 153 // Note: locks are acquired in textual declaration order 154 // and released in the opposite order 155 MutexLockerEx _locker1, _locker2, _locker3; 156 public: 157 CMSTokenSyncWithLocks(bool is_cms_thread, Mutex* mutex1, 158 Mutex* mutex2 = NULL, Mutex* mutex3 = NULL): 159 CMSTokenSync(is_cms_thread), 160 _locker1(mutex1, Mutex::_no_safepoint_check_flag), 161 _locker2(mutex2, Mutex::_no_safepoint_check_flag), 162 _locker3(mutex3, Mutex::_no_safepoint_check_flag) 163 { } 164 }; 165 166 167 // Wrapper class to temporarily disable icms during a foreground cms collection. 168 class ICMSDisabler: public StackObj { 169 public: 170 // The ctor disables icms and wakes up the thread so it notices the change; 171 // the dtor re-enables icms. Note that the CMSCollector methods will check 172 // CMSIncrementalMode. 173 ICMSDisabler() { CMSCollector::disable_icms(); CMSCollector::start_icms(); } 174 ~ICMSDisabler() { CMSCollector::enable_icms(); } 175 }; 176 177 ////////////////////////////////////////////////////////////////// 178 // Concurrent Mark-Sweep Generation ///////////////////////////// 179 ////////////////////////////////////////////////////////////////// 180 181 NOT_PRODUCT(CompactibleFreeListSpace* debug_cms_space;) 182 183 // This struct contains per-thread things necessary to support parallel 184 // young-gen collection. 185 class CMSParGCThreadState: public CHeapObj<mtGC> { 186 public: 187 CFLS_LAB lab; 188 PromotionInfo promo; 189 190 // Constructor. 191 CMSParGCThreadState(CompactibleFreeListSpace* cfls) : lab(cfls) { 192 promo.setSpace(cfls); 193 } 194 }; 195 196 ConcurrentMarkSweepGeneration::ConcurrentMarkSweepGeneration( 197 ReservedSpace rs, size_t initial_byte_size, int level, 198 CardTableRS* ct, bool use_adaptive_freelists, 199 FreeBlockDictionary<FreeChunk>::DictionaryChoice dictionaryChoice) : 200 CardGeneration(rs, initial_byte_size, level, ct), 201 _dilatation_factor(((double)MinChunkSize)/((double)(CollectedHeap::min_fill_size()))), 202 _debug_collection_type(Concurrent_collection_type), 203 _did_compact(false) 204 { 205 HeapWord* bottom = (HeapWord*) _virtual_space.low(); 206 HeapWord* end = (HeapWord*) _virtual_space.high(); 207 208 _direct_allocated_words = 0; 209 NOT_PRODUCT( 210 _numObjectsPromoted = 0; 211 _numWordsPromoted = 0; 212 _numObjectsAllocated = 0; 213 _numWordsAllocated = 0; 214 ) 215 216 _cmsSpace = new CompactibleFreeListSpace(_bts, MemRegion(bottom, end), 217 use_adaptive_freelists, 218 dictionaryChoice); 219 NOT_PRODUCT(debug_cms_space = _cmsSpace;) 220 if (_cmsSpace == NULL) { 221 vm_exit_during_initialization( 222 "CompactibleFreeListSpace allocation failure"); 223 } 224 _cmsSpace->_gen = this; 225 226 _gc_stats = new CMSGCStats(); 227 228 // Verify the assumption that FreeChunk::_prev and OopDesc::_klass 229 // offsets match. The ability to tell free chunks from objects 230 // depends on this property. 231 debug_only( 232 FreeChunk* junk = NULL; 233 assert(UseCompressedClassPointers || 234 junk->prev_addr() == (void*)(oop(junk)->klass_addr()), 235 "Offset of FreeChunk::_prev within FreeChunk must match" 236 " that of OopDesc::_klass within OopDesc"); 237 ) 238 if (CollectedHeap::use_parallel_gc_threads()) { 239 typedef CMSParGCThreadState* CMSParGCThreadStatePtr; 240 _par_gc_thread_states = 241 NEW_C_HEAP_ARRAY(CMSParGCThreadStatePtr, ParallelGCThreads, mtGC); 242 if (_par_gc_thread_states == NULL) { 243 vm_exit_during_initialization("Could not allocate par gc structs"); 244 } 245 for (uint i = 0; i < ParallelGCThreads; i++) { 246 _par_gc_thread_states[i] = new CMSParGCThreadState(cmsSpace()); 247 if (_par_gc_thread_states[i] == NULL) { 248 vm_exit_during_initialization("Could not allocate par gc structs"); 249 } 250 } 251 } else { 252 _par_gc_thread_states = NULL; 253 } 254 _incremental_collection_failed = false; 255 // The "dilatation_factor" is the expansion that can occur on 256 // account of the fact that the minimum object size in the CMS 257 // generation may be larger than that in, say, a contiguous young 258 // generation. 259 // Ideally, in the calculation below, we'd compute the dilatation 260 // factor as: MinChunkSize/(promoting_gen's min object size) 261 // Since we do not have such a general query interface for the 262 // promoting generation, we'll instead just use the minimum 263 // object size (which today is a header's worth of space); 264 // note that all arithmetic is in units of HeapWords. 265 assert(MinChunkSize >= CollectedHeap::min_fill_size(), "just checking"); 266 assert(_dilatation_factor >= 1.0, "from previous assert"); 267 } 268 269 270 // The field "_initiating_occupancy" represents the occupancy percentage 271 // at which we trigger a new collection cycle. Unless explicitly specified 272 // via CMSInitiatingOccupancyFraction (argument "io" below), it 273 // is calculated by: 274 // 275 // Let "f" be MinHeapFreeRatio in 276 // 277 // _initiating_occupancy = 100-f + 278 // f * (CMSTriggerRatio/100) 279 // where CMSTriggerRatio is the argument "tr" below. 280 // 281 // That is, if we assume the heap is at its desired maximum occupancy at the 282 // end of a collection, we let CMSTriggerRatio of the (purported) free 283 // space be allocated before initiating a new collection cycle. 284 // 285 void ConcurrentMarkSweepGeneration::init_initiating_occupancy(intx io, uintx tr) { 286 assert(io <= 100 && tr <= 100, "Check the arguments"); 287 if (io >= 0) { 288 _initiating_occupancy = (double)io / 100.0; 289 } else { 290 _initiating_occupancy = ((100 - MinHeapFreeRatio) + 291 (double)(tr * MinHeapFreeRatio) / 100.0) 292 / 100.0; 293 } 294 } 295 296 void ConcurrentMarkSweepGeneration::ref_processor_init() { 297 assert(collector() != NULL, "no collector"); 298 collector()->ref_processor_init(); 299 } 300 301 void CMSCollector::ref_processor_init() { 302 if (_ref_processor == NULL) { 303 // Allocate and initialize a reference processor 304 _ref_processor = 305 new ReferenceProcessor(_span, // span 306 (ParallelGCThreads > 1) && ParallelRefProcEnabled, // mt processing 307 (int) ParallelGCThreads, // mt processing degree 308 _cmsGen->refs_discovery_is_mt(), // mt discovery 309 (int) MAX2(ConcGCThreads, ParallelGCThreads), // mt discovery degree 310 _cmsGen->refs_discovery_is_atomic(), // discovery is not atomic 311 &_is_alive_closure, // closure for liveness info 312 false); // next field updates do not need write barrier 313 // Initialize the _ref_processor field of CMSGen 314 _cmsGen->set_ref_processor(_ref_processor); 315 316 } 317 } 318 319 CMSAdaptiveSizePolicy* CMSCollector::size_policy() { 320 GenCollectedHeap* gch = GenCollectedHeap::heap(); 321 assert(gch->kind() == CollectedHeap::GenCollectedHeap, 322 "Wrong type of heap"); 323 CMSAdaptiveSizePolicy* sp = (CMSAdaptiveSizePolicy*) 324 gch->gen_policy()->size_policy(); 325 assert(sp->is_gc_cms_adaptive_size_policy(), 326 "Wrong type of size policy"); 327 return sp; 328 } 329 330 CMSGCAdaptivePolicyCounters* CMSCollector::gc_adaptive_policy_counters() { 331 CMSGCAdaptivePolicyCounters* results = 332 (CMSGCAdaptivePolicyCounters*) collector_policy()->counters(); 333 assert( 334 results->kind() == GCPolicyCounters::CMSGCAdaptivePolicyCountersKind, 335 "Wrong gc policy counter kind"); 336 return results; 337 } 338 339 340 void ConcurrentMarkSweepGeneration::initialize_performance_counters() { 341 342 const char* gen_name = "old"; 343 344 // Generation Counters - generation 1, 1 subspace 345 _gen_counters = new GenerationCounters(gen_name, 1, 1, &_virtual_space); 346 347 _space_counters = new GSpaceCounters(gen_name, 0, 348 _virtual_space.reserved_size(), 349 this, _gen_counters); 350 } 351 352 CMSStats::CMSStats(ConcurrentMarkSweepGeneration* cms_gen, unsigned int alpha): 353 _cms_gen(cms_gen) 354 { 355 assert(alpha <= 100, "bad value"); 356 _saved_alpha = alpha; 357 358 // Initialize the alphas to the bootstrap value of 100. 359 _gc0_alpha = _cms_alpha = 100; 360 361 _cms_begin_time.update(); 362 _cms_end_time.update(); 363 364 _gc0_duration = 0.0; 365 _gc0_period = 0.0; 366 _gc0_promoted = 0; 367 368 _cms_duration = 0.0; 369 _cms_period = 0.0; 370 _cms_allocated = 0; 371 372 _cms_used_at_gc0_begin = 0; 373 _cms_used_at_gc0_end = 0; 374 _allow_duty_cycle_reduction = false; 375 _valid_bits = 0; 376 _icms_duty_cycle = CMSIncrementalDutyCycle; 377 } 378 379 double CMSStats::cms_free_adjustment_factor(size_t free) const { 380 // TBD: CR 6909490 381 return 1.0; 382 } 383 384 void CMSStats::adjust_cms_free_adjustment_factor(bool fail, size_t free) { 385 } 386 387 // If promotion failure handling is on use 388 // the padded average size of the promotion for each 389 // young generation collection. 390 double CMSStats::time_until_cms_gen_full() const { 391 size_t cms_free = _cms_gen->cmsSpace()->free(); 392 GenCollectedHeap* gch = GenCollectedHeap::heap(); 393 size_t expected_promotion = MIN2(gch->get_gen(0)->capacity(), 394 (size_t) _cms_gen->gc_stats()->avg_promoted()->padded_average()); 395 if (cms_free > expected_promotion) { 396 // Start a cms collection if there isn't enough space to promote 397 // for the next minor collection. Use the padded average as 398 // a safety factor. 399 cms_free -= expected_promotion; 400 401 // Adjust by the safety factor. 402 double cms_free_dbl = (double)cms_free; 403 double cms_adjustment = (100.0 - CMSIncrementalSafetyFactor)/100.0; 404 // Apply a further correction factor which tries to adjust 405 // for recent occurance of concurrent mode failures. 406 cms_adjustment = cms_adjustment * cms_free_adjustment_factor(cms_free); 407 cms_free_dbl = cms_free_dbl * cms_adjustment; 408 409 if (PrintGCDetails && Verbose) { 410 gclog_or_tty->print_cr("CMSStats::time_until_cms_gen_full: cms_free " 411 SIZE_FORMAT " expected_promotion " SIZE_FORMAT, 412 cms_free, expected_promotion); 413 gclog_or_tty->print_cr(" cms_free_dbl %f cms_consumption_rate %f", 414 cms_free_dbl, cms_consumption_rate() + 1.0); 415 } 416 // Add 1 in case the consumption rate goes to zero. 417 return cms_free_dbl / (cms_consumption_rate() + 1.0); 418 } 419 return 0.0; 420 } 421 422 // Compare the duration of the cms collection to the 423 // time remaining before the cms generation is empty. 424 // Note that the time from the start of the cms collection 425 // to the start of the cms sweep (less than the total 426 // duration of the cms collection) can be used. This 427 // has been tried and some applications experienced 428 // promotion failures early in execution. This was 429 // possibly because the averages were not accurate 430 // enough at the beginning. 431 double CMSStats::time_until_cms_start() const { 432 // We add "gc0_period" to the "work" calculation 433 // below because this query is done (mostly) at the 434 // end of a scavenge, so we need to conservatively 435 // account for that much possible delay 436 // in the query so as to avoid concurrent mode failures 437 // due to starting the collection just a wee bit too 438 // late. 439 double work = cms_duration() + gc0_period(); 440 double deadline = time_until_cms_gen_full(); 441 // If a concurrent mode failure occurred recently, we want to be 442 // more conservative and halve our expected time_until_cms_gen_full() 443 if (work > deadline) { 444 if (Verbose && PrintGCDetails) { 445 gclog_or_tty->print( 446 " CMSCollector: collect because of anticipated promotion " 447 "before full %3.7f + %3.7f > %3.7f ", cms_duration(), 448 gc0_period(), time_until_cms_gen_full()); 449 } 450 return 0.0; 451 } 452 return work - deadline; 453 } 454 455 // Return a duty cycle based on old_duty_cycle and new_duty_cycle, limiting the 456 // amount of change to prevent wild oscillation. 457 unsigned int CMSStats::icms_damped_duty_cycle(unsigned int old_duty_cycle, 458 unsigned int new_duty_cycle) { 459 assert(old_duty_cycle <= 100, "bad input value"); 460 assert(new_duty_cycle <= 100, "bad input value"); 461 462 // Note: use subtraction with caution since it may underflow (values are 463 // unsigned). Addition is safe since we're in the range 0-100. 464 unsigned int damped_duty_cycle = new_duty_cycle; 465 if (new_duty_cycle < old_duty_cycle) { 466 const unsigned int largest_delta = MAX2(old_duty_cycle / 4, 5U); 467 if (new_duty_cycle + largest_delta < old_duty_cycle) { 468 damped_duty_cycle = old_duty_cycle - largest_delta; 469 } 470 } else if (new_duty_cycle > old_duty_cycle) { 471 const unsigned int largest_delta = MAX2(old_duty_cycle / 4, 15U); 472 if (new_duty_cycle > old_duty_cycle + largest_delta) { 473 damped_duty_cycle = MIN2(old_duty_cycle + largest_delta, 100U); 474 } 475 } 476 assert(damped_duty_cycle <= 100, "invalid duty cycle computed"); 477 478 if (CMSTraceIncrementalPacing) { 479 gclog_or_tty->print(" [icms_damped_duty_cycle(%d,%d) = %d] ", 480 old_duty_cycle, new_duty_cycle, damped_duty_cycle); 481 } 482 return damped_duty_cycle; 483 } 484 485 unsigned int CMSStats::icms_update_duty_cycle_impl() { 486 assert(CMSIncrementalPacing && valid(), 487 "should be handled in icms_update_duty_cycle()"); 488 489 double cms_time_so_far = cms_timer().seconds(); 490 double scaled_duration = cms_duration_per_mb() * _cms_used_at_gc0_end / M; 491 double scaled_duration_remaining = fabsd(scaled_duration - cms_time_so_far); 492 493 // Avoid division by 0. 494 double time_until_full = MAX2(time_until_cms_gen_full(), 0.01); 495 double duty_cycle_dbl = 100.0 * scaled_duration_remaining / time_until_full; 496 497 unsigned int new_duty_cycle = MIN2((unsigned int)duty_cycle_dbl, 100U); 498 if (new_duty_cycle > _icms_duty_cycle) { 499 // Avoid very small duty cycles (1 or 2); 0 is allowed. 500 if (new_duty_cycle > 2) { 501 _icms_duty_cycle = icms_damped_duty_cycle(_icms_duty_cycle, 502 new_duty_cycle); 503 } 504 } else if (_allow_duty_cycle_reduction) { 505 // The duty cycle is reduced only once per cms cycle (see record_cms_end()). 506 new_duty_cycle = icms_damped_duty_cycle(_icms_duty_cycle, new_duty_cycle); 507 // Respect the minimum duty cycle. 508 unsigned int min_duty_cycle = (unsigned int)CMSIncrementalDutyCycleMin; 509 _icms_duty_cycle = MAX2(new_duty_cycle, min_duty_cycle); 510 } 511 512 if (PrintGCDetails || CMSTraceIncrementalPacing) { 513 gclog_or_tty->print(" icms_dc=%d ", _icms_duty_cycle); 514 } 515 516 _allow_duty_cycle_reduction = false; 517 return _icms_duty_cycle; 518 } 519 520 #ifndef PRODUCT 521 void CMSStats::print_on(outputStream *st) const { 522 st->print(" gc0_alpha=%d,cms_alpha=%d", _gc0_alpha, _cms_alpha); 523 st->print(",gc0_dur=%g,gc0_per=%g,gc0_promo=" SIZE_FORMAT, 524 gc0_duration(), gc0_period(), gc0_promoted()); 525 st->print(",cms_dur=%g,cms_dur_per_mb=%g,cms_per=%g,cms_alloc=" SIZE_FORMAT, 526 cms_duration(), cms_duration_per_mb(), 527 cms_period(), cms_allocated()); 528 st->print(",cms_since_beg=%g,cms_since_end=%g", 529 cms_time_since_begin(), cms_time_since_end()); 530 st->print(",cms_used_beg=" SIZE_FORMAT ",cms_used_end=" SIZE_FORMAT, 531 _cms_used_at_gc0_begin, _cms_used_at_gc0_end); 532 if (CMSIncrementalMode) { 533 st->print(",dc=%d", icms_duty_cycle()); 534 } 535 536 if (valid()) { 537 st->print(",promo_rate=%g,cms_alloc_rate=%g", 538 promotion_rate(), cms_allocation_rate()); 539 st->print(",cms_consumption_rate=%g,time_until_full=%g", 540 cms_consumption_rate(), time_until_cms_gen_full()); 541 } 542 st->print(" "); 543 } 544 #endif // #ifndef PRODUCT 545 546 CMSCollector::CollectorState CMSCollector::_collectorState = 547 CMSCollector::Idling; 548 bool CMSCollector::_foregroundGCIsActive = false; 549 bool CMSCollector::_foregroundGCShouldWait = false; 550 551 CMSCollector::CMSCollector(ConcurrentMarkSweepGeneration* cmsGen, 552 CardTableRS* ct, 553 ConcurrentMarkSweepPolicy* cp): 554 _cmsGen(cmsGen), 555 _ct(ct), 556 _ref_processor(NULL), // will be set later 557 _conc_workers(NULL), // may be set later 558 _abort_preclean(false), 559 _start_sampling(false), 560 _between_prologue_and_epilogue(false), 561 _markBitMap(0, Mutex::leaf + 1, "CMS_markBitMap_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 567 _span(cmsGen->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_lock(new Mutex(Mutex::leaf + 1, "CMS_eden_chunk_lock", true)), 574 _eden_chunk_array(NULL), // may be set in ctor body 575 _eden_chunk_capacity(0), // -- ditto -- 576 _eden_chunk_index(0), // -- ditto -- 577 _survivor_plab_array(NULL), // -- ditto -- 578 _survivor_chunk_array(NULL), // -- ditto -- 579 _survivor_chunk_capacity(0), // -- ditto -- 580 _survivor_chunk_index(0), // -- ditto -- 581 _ser_pmc_preclean_ovflw(0), 582 _ser_kac_preclean_ovflw(0), 583 _ser_pmc_remark_ovflw(0), 584 _par_pmc_remark_ovflw(0), 585 _ser_kac_ovflw(0), 586 _par_kac_ovflw(0), 587 #ifndef PRODUCT 588 _num_par_pushes(0), 589 #endif 590 _collection_count_start(0), 591 _verifying(false), 592 _icms_start_limit(NULL), 593 _icms_stop_limit(NULL), 594 _verification_mark_bm(0, Mutex::leaf + 1, "CMS_verification_mark_bm_lock"), 595 _completed_initialization(false), 596 _collector_policy(cp), 597 _should_unload_classes(CMSClassUnloadingEnabled), 598 _concurrent_cycles_since_last_unload(0), 599 _roots_scanning_options(SharedHeap::SO_None), 600 _inter_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding), 601 _intra_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding), 602 _gc_tracer_cm(new (ResourceObj::C_HEAP, mtGC) CMSTracer()), 603 _gc_timer_cm(new (ResourceObj::C_HEAP, mtGC) ConcurrentGCTimer()), 604 _cms_start_registered(false) 605 { 606 if (ExplicitGCInvokesConcurrentAndUnloadsClasses) { 607 ExplicitGCInvokesConcurrent = true; 608 } 609 // Now expand the span and allocate the collection support structures 610 // (MUT, marking bit map etc.) to cover both generations subject to 611 // collection. 612 613 // For use by dirty card to oop closures. 614 _cmsGen->cmsSpace()->set_collector(this); 615 616 // Allocate MUT and marking bit map 617 { 618 MutexLockerEx x(_markBitMap.lock(), Mutex::_no_safepoint_check_flag); 619 if (!_markBitMap.allocate(_span)) { 620 warning("Failed to allocate CMS Bit Map"); 621 return; 622 } 623 assert(_markBitMap.covers(_span), "_markBitMap inconsistency?"); 624 } 625 { 626 _modUnionTable.allocate(_span); 627 assert(_modUnionTable.covers(_span), "_modUnionTable inconsistency?"); 628 } 629 630 if (!_markStack.allocate(MarkStackSize)) { 631 warning("Failed to allocate CMS Marking Stack"); 632 return; 633 } 634 635 // Support for multi-threaded concurrent phases 636 if (CMSConcurrentMTEnabled) { 637 if (FLAG_IS_DEFAULT(ConcGCThreads)) { 638 // just for now 639 FLAG_SET_DEFAULT(ConcGCThreads, (ParallelGCThreads + 3)/4); 640 } 641 if (ConcGCThreads > 1) { 642 _conc_workers = new YieldingFlexibleWorkGang("Parallel CMS Threads", 643 ConcGCThreads, true); 644 if (_conc_workers == NULL) { 645 warning("GC/CMS: _conc_workers allocation failure: " 646 "forcing -CMSConcurrentMTEnabled"); 647 CMSConcurrentMTEnabled = false; 648 } else { 649 _conc_workers->initialize_workers(); 650 } 651 } else { 652 CMSConcurrentMTEnabled = false; 653 } 654 } 655 if (!CMSConcurrentMTEnabled) { 656 ConcGCThreads = 0; 657 } else { 658 // Turn off CMSCleanOnEnter optimization temporarily for 659 // the MT case where it's not fixed yet; see 6178663. 660 CMSCleanOnEnter = false; 661 } 662 assert((_conc_workers != NULL) == (ConcGCThreads > 1), 663 "Inconsistency"); 664 665 // Parallel task queues; these are shared for the 666 // concurrent and stop-world phases of CMS, but 667 // are not shared with parallel scavenge (ParNew). 668 { 669 uint i; 670 uint num_queues = (uint) MAX2(ParallelGCThreads, ConcGCThreads); 671 672 if ((CMSParallelRemarkEnabled || CMSConcurrentMTEnabled 673 || ParallelRefProcEnabled) 674 && num_queues > 0) { 675 _task_queues = new OopTaskQueueSet(num_queues); 676 if (_task_queues == NULL) { 677 warning("task_queues allocation failure."); 678 return; 679 } 680 _hash_seed = NEW_C_HEAP_ARRAY(int, num_queues, mtGC); 681 if (_hash_seed == NULL) { 682 warning("_hash_seed array allocation failure"); 683 return; 684 } 685 686 typedef Padded<OopTaskQueue> PaddedOopTaskQueue; 687 for (i = 0; i < num_queues; i++) { 688 PaddedOopTaskQueue *q = new PaddedOopTaskQueue(); 689 if (q == NULL) { 690 warning("work_queue allocation failure."); 691 return; 692 } 693 _task_queues->register_queue(i, q); 694 } 695 for (i = 0; i < num_queues; i++) { 696 _task_queues->queue(i)->initialize(); 697 _hash_seed[i] = 17; // copied from ParNew 698 } 699 } 700 } 701 702 _cmsGen ->init_initiating_occupancy(CMSInitiatingOccupancyFraction, CMSTriggerRatio); 703 704 // Clip CMSBootstrapOccupancy between 0 and 100. 705 _bootstrap_occupancy = ((double)CMSBootstrapOccupancy)/(double)100; 706 707 _full_gcs_since_conc_gc = 0; 708 709 // Now tell CMS generations the identity of their collector 710 ConcurrentMarkSweepGeneration::set_collector(this); 711 712 // Create & start a CMS thread for this CMS collector 713 _cmsThread = ConcurrentMarkSweepThread::start(this); 714 assert(cmsThread() != NULL, "CMS Thread should have been created"); 715 assert(cmsThread()->collector() == this, 716 "CMS Thread should refer to this gen"); 717 assert(CGC_lock != NULL, "Where's the CGC_lock?"); 718 719 // Support for parallelizing young gen rescan 720 GenCollectedHeap* gch = GenCollectedHeap::heap(); 721 _young_gen = gch->prev_gen(_cmsGen); 722 if (gch->supports_inline_contig_alloc()) { 723 _top_addr = gch->top_addr(); 724 _end_addr = gch->end_addr(); 725 assert(_young_gen != NULL, "no _young_gen"); 726 _eden_chunk_index = 0; 727 _eden_chunk_capacity = (_young_gen->max_capacity()+CMSSamplingGrain)/CMSSamplingGrain; 728 _eden_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, _eden_chunk_capacity, mtGC); 729 if (_eden_chunk_array == NULL) { 730 _eden_chunk_capacity = 0; 731 warning("GC/CMS: _eden_chunk_array allocation failure"); 732 } 733 } 734 assert(_eden_chunk_array != NULL || _eden_chunk_capacity == 0, "Error"); 735 736 // Support for parallelizing survivor space rescan 737 if ((CMSParallelRemarkEnabled && CMSParallelSurvivorRemarkEnabled) || CMSParallelInitialMarkEnabled) { 738 const size_t max_plab_samples = 739 ((DefNewGeneration*)_young_gen)->max_survivor_size()/MinTLABSize; 740 741 _survivor_plab_array = NEW_C_HEAP_ARRAY(ChunkArray, ParallelGCThreads, mtGC); 742 _survivor_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, 2*max_plab_samples, mtGC); 743 _cursor = NEW_C_HEAP_ARRAY(size_t, ParallelGCThreads, mtGC); 744 if (_survivor_plab_array == NULL || _survivor_chunk_array == NULL 745 || _cursor == NULL) { 746 warning("Failed to allocate survivor plab/chunk array"); 747 if (_survivor_plab_array != NULL) { 748 FREE_C_HEAP_ARRAY(ChunkArray, _survivor_plab_array, mtGC); 749 _survivor_plab_array = NULL; 750 } 751 if (_survivor_chunk_array != NULL) { 752 FREE_C_HEAP_ARRAY(HeapWord*, _survivor_chunk_array, mtGC); 753 _survivor_chunk_array = NULL; 754 } 755 if (_cursor != NULL) { 756 FREE_C_HEAP_ARRAY(size_t, _cursor, mtGC); 757 _cursor = NULL; 758 } 759 } else { 760 _survivor_chunk_capacity = 2*max_plab_samples; 761 for (uint i = 0; i < ParallelGCThreads; i++) { 762 HeapWord** vec = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples, mtGC); 763 if (vec == NULL) { 764 warning("Failed to allocate survivor plab array"); 765 for (int j = i; j > 0; j--) { 766 FREE_C_HEAP_ARRAY(HeapWord*, _survivor_plab_array[j-1].array(), mtGC); 767 } 768 FREE_C_HEAP_ARRAY(ChunkArray, _survivor_plab_array, mtGC); 769 FREE_C_HEAP_ARRAY(HeapWord*, _survivor_chunk_array, mtGC); 770 _survivor_plab_array = NULL; 771 _survivor_chunk_array = NULL; 772 _survivor_chunk_capacity = 0; 773 break; 774 } else { 775 ChunkArray* cur = 776 ::new (&_survivor_plab_array[i]) ChunkArray(vec, 777 max_plab_samples); 778 assert(cur->end() == 0, "Should be 0"); 779 assert(cur->array() == vec, "Should be vec"); 780 assert(cur->capacity() == max_plab_samples, "Error"); 781 } 782 } 783 } 784 } 785 assert( ( _survivor_plab_array != NULL 786 && _survivor_chunk_array != NULL) 787 || ( _survivor_chunk_capacity == 0 788 && _survivor_chunk_index == 0), 789 "Error"); 790 791 NOT_PRODUCT(_overflow_counter = CMSMarkStackOverflowInterval;) 792 _gc_counters = new CollectorCounters("CMS", 1); 793 _completed_initialization = true; 794 _inter_sweep_timer.start(); // start of time 795 } 796 797 const char* ConcurrentMarkSweepGeneration::name() const { 798 return "concurrent mark-sweep generation"; 799 } 800 void ConcurrentMarkSweepGeneration::update_counters() { 801 if (UsePerfData) { 802 _space_counters->update_all(); 803 _gen_counters->update_all(); 804 } 805 } 806 807 // this is an optimized version of update_counters(). it takes the 808 // used value as a parameter rather than computing it. 809 // 810 void ConcurrentMarkSweepGeneration::update_counters(size_t used) { 811 if (UsePerfData) { 812 _space_counters->update_used(used); 813 _space_counters->update_capacity(); 814 _gen_counters->update_all(); 815 } 816 } 817 818 void ConcurrentMarkSweepGeneration::print() const { 819 Generation::print(); 820 cmsSpace()->print(); 821 } 822 823 #ifndef PRODUCT 824 void ConcurrentMarkSweepGeneration::print_statistics() { 825 cmsSpace()->printFLCensus(0); 826 } 827 #endif 828 829 void ConcurrentMarkSweepGeneration::printOccupancy(const char *s) { 830 GenCollectedHeap* gch = GenCollectedHeap::heap(); 831 if (PrintGCDetails) { 832 if (Verbose) { 833 gclog_or_tty->print("[%d %s-%s: "SIZE_FORMAT"("SIZE_FORMAT")]", 834 level(), short_name(), s, used(), capacity()); 835 } else { 836 gclog_or_tty->print("[%d %s-%s: "SIZE_FORMAT"K("SIZE_FORMAT"K)]", 837 level(), short_name(), s, used() / K, capacity() / K); 838 } 839 } 840 if (Verbose) { 841 gclog_or_tty->print(" "SIZE_FORMAT"("SIZE_FORMAT")", 842 gch->used(), gch->capacity()); 843 } else { 844 gclog_or_tty->print(" "SIZE_FORMAT"K("SIZE_FORMAT"K)", 845 gch->used() / K, gch->capacity() / K); 846 } 847 } 848 849 size_t 850 ConcurrentMarkSweepGeneration::contiguous_available() const { 851 // dld proposes an improvement in precision here. If the committed 852 // part of the space ends in a free block we should add that to 853 // uncommitted size in the calculation below. Will make this 854 // change later, staying with the approximation below for the 855 // time being. -- ysr. 856 return MAX2(_virtual_space.uncommitted_size(), unsafe_max_alloc_nogc()); 857 } 858 859 size_t 860 ConcurrentMarkSweepGeneration::unsafe_max_alloc_nogc() const { 861 return _cmsSpace->max_alloc_in_words() * HeapWordSize; 862 } 863 864 size_t ConcurrentMarkSweepGeneration::max_available() const { 865 return free() + _virtual_space.uncommitted_size(); 866 } 867 868 bool ConcurrentMarkSweepGeneration::promotion_attempt_is_safe(size_t max_promotion_in_bytes) const { 869 size_t available = max_available(); 870 size_t av_promo = (size_t)gc_stats()->avg_promoted()->padded_average(); 871 bool res = (available >= av_promo) || (available >= max_promotion_in_bytes); 872 if (Verbose && PrintGCDetails) { 873 gclog_or_tty->print_cr( 874 "CMS: promo attempt is%s safe: available("SIZE_FORMAT") %s av_promo("SIZE_FORMAT")," 875 "max_promo("SIZE_FORMAT")", 876 res? "":" not", available, res? ">=":"<", 877 av_promo, max_promotion_in_bytes); 878 } 879 return res; 880 } 881 882 // At a promotion failure dump information on block layout in heap 883 // (cms old generation). 884 void ConcurrentMarkSweepGeneration::promotion_failure_occurred() { 885 if (CMSDumpAtPromotionFailure) { 886 cmsSpace()->dump_at_safepoint_with_locks(collector(), gclog_or_tty); 887 } 888 } 889 890 CompactibleSpace* 891 ConcurrentMarkSweepGeneration::first_compaction_space() const { 892 return _cmsSpace; 893 } 894 895 void ConcurrentMarkSweepGeneration::reset_after_compaction() { 896 // Clear the promotion information. These pointers can be adjusted 897 // along with all the other pointers into the heap but 898 // compaction is expected to be a rare event with 899 // a heap using cms so don't do it without seeing the need. 900 if (CollectedHeap::use_parallel_gc_threads()) { 901 for (uint i = 0; i < ParallelGCThreads; i++) { 902 _par_gc_thread_states[i]->promo.reset(); 903 } 904 } 905 } 906 907 void ConcurrentMarkSweepGeneration::space_iterate(SpaceClosure* blk, bool usedOnly) { 908 blk->do_space(_cmsSpace); 909 } 910 911 void ConcurrentMarkSweepGeneration::compute_new_size() { 912 assert_locked_or_safepoint(Heap_lock); 913 914 // If incremental collection failed, we just want to expand 915 // to the limit. 916 if (incremental_collection_failed()) { 917 clear_incremental_collection_failed(); 918 grow_to_reserved(); 919 return; 920 } 921 922 // The heap has been compacted but not reset yet. 923 // Any metric such as free() or used() will be incorrect. 924 925 CardGeneration::compute_new_size(); 926 927 // Reset again after a possible resizing 928 if (did_compact()) { 929 cmsSpace()->reset_after_compaction(); 930 } 931 } 932 933 void ConcurrentMarkSweepGeneration::compute_new_size_free_list() { 934 assert_locked_or_safepoint(Heap_lock); 935 936 // If incremental collection failed, we just want to expand 937 // to the limit. 938 if (incremental_collection_failed()) { 939 clear_incremental_collection_failed(); 940 grow_to_reserved(); 941 return; 942 } 943 944 double free_percentage = ((double) free()) / capacity(); 945 double desired_free_percentage = (double) MinHeapFreeRatio / 100; 946 double maximum_free_percentage = (double) MaxHeapFreeRatio / 100; 947 948 // compute expansion delta needed for reaching desired free percentage 949 if (free_percentage < desired_free_percentage) { 950 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 951 assert(desired_capacity >= capacity(), "invalid expansion size"); 952 size_t expand_bytes = MAX2(desired_capacity - capacity(), MinHeapDeltaBytes); 953 if (PrintGCDetails && Verbose) { 954 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 955 gclog_or_tty->print_cr("\nFrom compute_new_size: "); 956 gclog_or_tty->print_cr(" Free fraction %f", free_percentage); 957 gclog_or_tty->print_cr(" Desired free fraction %f", 958 desired_free_percentage); 959 gclog_or_tty->print_cr(" Maximum free fraction %f", 960 maximum_free_percentage); 961 gclog_or_tty->print_cr(" Capacity "SIZE_FORMAT, capacity()/1000); 962 gclog_or_tty->print_cr(" Desired capacity "SIZE_FORMAT, 963 desired_capacity/1000); 964 int prev_level = level() - 1; 965 if (prev_level >= 0) { 966 size_t prev_size = 0; 967 GenCollectedHeap* gch = GenCollectedHeap::heap(); 968 Generation* prev_gen = gch->_gens[prev_level]; 969 prev_size = prev_gen->capacity(); 970 gclog_or_tty->print_cr(" Younger gen size "SIZE_FORMAT, 971 prev_size/1000); 972 } 973 gclog_or_tty->print_cr(" unsafe_max_alloc_nogc "SIZE_FORMAT, 974 unsafe_max_alloc_nogc()/1000); 975 gclog_or_tty->print_cr(" contiguous available "SIZE_FORMAT, 976 contiguous_available()/1000); 977 gclog_or_tty->print_cr(" Expand by "SIZE_FORMAT" (bytes)", 978 expand_bytes); 979 } 980 // safe if expansion fails 981 expand(expand_bytes, 0, CMSExpansionCause::_satisfy_free_ratio); 982 if (PrintGCDetails && Verbose) { 983 gclog_or_tty->print_cr(" Expanded free fraction %f", 984 ((double) free()) / capacity()); 985 } 986 } else { 987 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 988 assert(desired_capacity <= capacity(), "invalid expansion size"); 989 size_t shrink_bytes = capacity() - desired_capacity; 990 // Don't shrink unless the delta is greater than the minimum shrink we want 991 if (shrink_bytes >= MinHeapDeltaBytes) { 992 shrink_free_list_by(shrink_bytes); 993 } 994 } 995 } 996 997 Mutex* ConcurrentMarkSweepGeneration::freelistLock() const { 998 return cmsSpace()->freelistLock(); 999 } 1000 1001 HeapWord* ConcurrentMarkSweepGeneration::allocate(size_t size, 1002 bool tlab) { 1003 CMSSynchronousYieldRequest yr; 1004 MutexLockerEx x(freelistLock(), 1005 Mutex::_no_safepoint_check_flag); 1006 return have_lock_and_allocate(size, tlab); 1007 } 1008 1009 HeapWord* ConcurrentMarkSweepGeneration::have_lock_and_allocate(size_t size, 1010 bool tlab /* ignored */) { 1011 assert_lock_strong(freelistLock()); 1012 size_t adjustedSize = CompactibleFreeListSpace::adjustObjectSize(size); 1013 HeapWord* res = cmsSpace()->allocate(adjustedSize); 1014 // Allocate the object live (grey) if the background collector has 1015 // started marking. This is necessary because the marker may 1016 // have passed this address and consequently this object will 1017 // not otherwise be greyed and would be incorrectly swept up. 1018 // Note that if this object contains references, the writing 1019 // of those references will dirty the card containing this object 1020 // allowing the object to be blackened (and its references scanned) 1021 // either during a preclean phase or at the final checkpoint. 1022 if (res != NULL) { 1023 // We may block here with an uninitialized object with 1024 // its mark-bit or P-bits not yet set. Such objects need 1025 // to be safely navigable by block_start(). 1026 assert(oop(res)->klass_or_null() == NULL, "Object should be uninitialized here."); 1027 assert(!((FreeChunk*)res)->is_free(), "Error, block will look free but show wrong size"); 1028 collector()->direct_allocated(res, adjustedSize); 1029 _direct_allocated_words += adjustedSize; 1030 // allocation counters 1031 NOT_PRODUCT( 1032 _numObjectsAllocated++; 1033 _numWordsAllocated += (int)adjustedSize; 1034 ) 1035 } 1036 return res; 1037 } 1038 1039 // In the case of direct allocation by mutators in a generation that 1040 // is being concurrently collected, the object must be allocated 1041 // live (grey) if the background collector has started marking. 1042 // This is necessary because the marker may 1043 // have passed this address and consequently this object will 1044 // not otherwise be greyed and would be incorrectly swept up. 1045 // Note that if this object contains references, the writing 1046 // of those references will dirty the card containing this object 1047 // allowing the object to be blackened (and its references scanned) 1048 // either during a preclean phase or at the final checkpoint. 1049 void CMSCollector::direct_allocated(HeapWord* start, size_t size) { 1050 assert(_markBitMap.covers(start, size), "Out of bounds"); 1051 if (_collectorState >= Marking) { 1052 MutexLockerEx y(_markBitMap.lock(), 1053 Mutex::_no_safepoint_check_flag); 1054 // [see comments preceding SweepClosure::do_blk() below for details] 1055 // 1056 // Can the P-bits be deleted now? JJJ 1057 // 1058 // 1. need to mark the object as live so it isn't collected 1059 // 2. need to mark the 2nd bit to indicate the object may be uninitialized 1060 // 3. need to mark the end of the object so marking, precleaning or sweeping 1061 // can skip over uninitialized or unparsable objects. An allocated 1062 // object is considered uninitialized for our purposes as long as 1063 // its klass word is NULL. All old gen objects are parsable 1064 // as soon as they are initialized.) 1065 _markBitMap.mark(start); // object is live 1066 _markBitMap.mark(start + 1); // object is potentially uninitialized? 1067 _markBitMap.mark(start + size - 1); 1068 // mark end of object 1069 } 1070 // check that oop looks uninitialized 1071 assert(oop(start)->klass_or_null() == NULL, "_klass should be NULL"); 1072 } 1073 1074 void CMSCollector::promoted(bool par, HeapWord* start, 1075 bool is_obj_array, size_t obj_size) { 1076 assert(_markBitMap.covers(start), "Out of bounds"); 1077 // See comment in direct_allocated() about when objects should 1078 // be allocated live. 1079 if (_collectorState >= Marking) { 1080 // we already hold the marking bit map lock, taken in 1081 // the prologue 1082 if (par) { 1083 _markBitMap.par_mark(start); 1084 } else { 1085 _markBitMap.mark(start); 1086 } 1087 // We don't need to mark the object as uninitialized (as 1088 // in direct_allocated above) because this is being done with the 1089 // world stopped and the object will be initialized by the 1090 // time the marking, precleaning or sweeping get to look at it. 1091 // But see the code for copying objects into the CMS generation, 1092 // where we need to ensure that concurrent readers of the 1093 // block offset table are able to safely navigate a block that 1094 // is in flux from being free to being allocated (and in 1095 // transition while being copied into) and subsequently 1096 // becoming a bona-fide object when the copy/promotion is complete. 1097 assert(SafepointSynchronize::is_at_safepoint(), 1098 "expect promotion only at safepoints"); 1099 1100 if (_collectorState < Sweeping) { 1101 // Mark the appropriate cards in the modUnionTable, so that 1102 // this object gets scanned before the sweep. If this is 1103 // not done, CMS generation references in the object might 1104 // not get marked. 1105 // For the case of arrays, which are otherwise precisely 1106 // marked, we need to dirty the entire array, not just its head. 1107 if (is_obj_array) { 1108 // The [par_]mark_range() method expects mr.end() below to 1109 // be aligned to the granularity of a bit's representation 1110 // in the heap. In the case of the MUT below, that's a 1111 // card size. 1112 MemRegion mr(start, 1113 (HeapWord*)round_to((intptr_t)(start + obj_size), 1114 CardTableModRefBS::card_size /* bytes */)); 1115 if (par) { 1116 _modUnionTable.par_mark_range(mr); 1117 } else { 1118 _modUnionTable.mark_range(mr); 1119 } 1120 } else { // not an obj array; we can just mark the head 1121 if (par) { 1122 _modUnionTable.par_mark(start); 1123 } else { 1124 _modUnionTable.mark(start); 1125 } 1126 } 1127 } 1128 } 1129 } 1130 1131 static inline size_t percent_of_space(Space* space, HeapWord* addr) 1132 { 1133 size_t delta = pointer_delta(addr, space->bottom()); 1134 return (size_t)(delta * 100.0 / (space->capacity() / HeapWordSize)); 1135 } 1136 1137 void CMSCollector::icms_update_allocation_limits() 1138 { 1139 Generation* gen0 = GenCollectedHeap::heap()->get_gen(0); 1140 EdenSpace* eden = gen0->as_DefNewGeneration()->eden(); 1141 1142 const unsigned int duty_cycle = stats().icms_update_duty_cycle(); 1143 if (CMSTraceIncrementalPacing) { 1144 stats().print(); 1145 } 1146 1147 assert(duty_cycle <= 100, "invalid duty cycle"); 1148 if (duty_cycle != 0) { 1149 // The duty_cycle is a percentage between 0 and 100; convert to words and 1150 // then compute the offset from the endpoints of the space. 1151 size_t free_words = eden->free() / HeapWordSize; 1152 double free_words_dbl = (double)free_words; 1153 size_t duty_cycle_words = (size_t)(free_words_dbl * duty_cycle / 100.0); 1154 size_t offset_words = (free_words - duty_cycle_words) / 2; 1155 1156 _icms_start_limit = eden->top() + offset_words; 1157 _icms_stop_limit = eden->end() - offset_words; 1158 1159 // The limits may be adjusted (shifted to the right) by 1160 // CMSIncrementalOffset, to allow the application more mutator time after a 1161 // young gen gc (when all mutators were stopped) and before CMS starts and 1162 // takes away one or more cpus. 1163 if (CMSIncrementalOffset != 0) { 1164 double adjustment_dbl = free_words_dbl * CMSIncrementalOffset / 100.0; 1165 size_t adjustment = (size_t)adjustment_dbl; 1166 HeapWord* tmp_stop = _icms_stop_limit + adjustment; 1167 if (tmp_stop > _icms_stop_limit && tmp_stop < eden->end()) { 1168 _icms_start_limit += adjustment; 1169 _icms_stop_limit = tmp_stop; 1170 } 1171 } 1172 } 1173 if (duty_cycle == 0 || (_icms_start_limit == _icms_stop_limit)) { 1174 _icms_start_limit = _icms_stop_limit = eden->end(); 1175 } 1176 1177 // Install the new start limit. 1178 eden->set_soft_end(_icms_start_limit); 1179 1180 if (CMSTraceIncrementalMode) { 1181 gclog_or_tty->print(" icms alloc limits: " 1182 PTR_FORMAT "," PTR_FORMAT 1183 " (" SIZE_FORMAT "%%," SIZE_FORMAT "%%) ", 1184 _icms_start_limit, _icms_stop_limit, 1185 percent_of_space(eden, _icms_start_limit), 1186 percent_of_space(eden, _icms_stop_limit)); 1187 if (Verbose) { 1188 gclog_or_tty->print("eden: "); 1189 eden->print_on(gclog_or_tty); 1190 } 1191 } 1192 } 1193 1194 // Any changes here should try to maintain the invariant 1195 // that if this method is called with _icms_start_limit 1196 // and _icms_stop_limit both NULL, then it should return NULL 1197 // and not notify the icms thread. 1198 HeapWord* 1199 CMSCollector::allocation_limit_reached(Space* space, HeapWord* top, 1200 size_t word_size) 1201 { 1202 // A start_limit equal to end() means the duty cycle is 0, so treat that as a 1203 // nop. 1204 if (CMSIncrementalMode && _icms_start_limit != space->end()) { 1205 if (top <= _icms_start_limit) { 1206 if (CMSTraceIncrementalMode) { 1207 space->print_on(gclog_or_tty); 1208 gclog_or_tty->stamp(); 1209 gclog_or_tty->print_cr(" start limit top=" PTR_FORMAT 1210 ", new limit=" PTR_FORMAT 1211 " (" SIZE_FORMAT "%%)", 1212 top, _icms_stop_limit, 1213 percent_of_space(space, _icms_stop_limit)); 1214 } 1215 ConcurrentMarkSweepThread::start_icms(); 1216 assert(top < _icms_stop_limit, "Tautology"); 1217 if (word_size < pointer_delta(_icms_stop_limit, top)) { 1218 return _icms_stop_limit; 1219 } 1220 1221 // The allocation will cross both the _start and _stop limits, so do the 1222 // stop notification also and return end(). 1223 if (CMSTraceIncrementalMode) { 1224 space->print_on(gclog_or_tty); 1225 gclog_or_tty->stamp(); 1226 gclog_or_tty->print_cr(" +stop limit top=" PTR_FORMAT 1227 ", new limit=" PTR_FORMAT 1228 " (" SIZE_FORMAT "%%)", 1229 top, space->end(), 1230 percent_of_space(space, space->end())); 1231 } 1232 ConcurrentMarkSweepThread::stop_icms(); 1233 return space->end(); 1234 } 1235 1236 if (top <= _icms_stop_limit) { 1237 if (CMSTraceIncrementalMode) { 1238 space->print_on(gclog_or_tty); 1239 gclog_or_tty->stamp(); 1240 gclog_or_tty->print_cr(" stop limit top=" PTR_FORMAT 1241 ", new limit=" PTR_FORMAT 1242 " (" SIZE_FORMAT "%%)", 1243 top, space->end(), 1244 percent_of_space(space, space->end())); 1245 } 1246 ConcurrentMarkSweepThread::stop_icms(); 1247 return space->end(); 1248 } 1249 1250 if (CMSTraceIncrementalMode) { 1251 space->print_on(gclog_or_tty); 1252 gclog_or_tty->stamp(); 1253 gclog_or_tty->print_cr(" end limit top=" PTR_FORMAT 1254 ", new limit=" PTR_FORMAT, 1255 top, NULL); 1256 } 1257 } 1258 1259 return NULL; 1260 } 1261 1262 oop ConcurrentMarkSweepGeneration::promote(oop obj, size_t obj_size) { 1263 assert(obj_size == (size_t)obj->size(), "bad obj_size passed in"); 1264 // allocate, copy and if necessary update promoinfo -- 1265 // delegate to underlying space. 1266 assert_lock_strong(freelistLock()); 1267 1268 #ifndef PRODUCT 1269 if (Universe::heap()->promotion_should_fail()) { 1270 return NULL; 1271 } 1272 #endif // #ifndef PRODUCT 1273 1274 oop res = _cmsSpace->promote(obj, obj_size); 1275 if (res == NULL) { 1276 // expand and retry 1277 size_t s = _cmsSpace->expansionSpaceRequired(obj_size); // HeapWords 1278 expand(s*HeapWordSize, MinHeapDeltaBytes, 1279 CMSExpansionCause::_satisfy_promotion); 1280 // Since there's currently no next generation, we don't try to promote 1281 // into a more senior generation. 1282 assert(next_gen() == NULL, "assumption, based upon which no attempt " 1283 "is made to pass on a possibly failing " 1284 "promotion to next generation"); 1285 res = _cmsSpace->promote(obj, obj_size); 1286 } 1287 if (res != NULL) { 1288 // See comment in allocate() about when objects should 1289 // be allocated live. 1290 assert(obj->is_oop(), "Will dereference klass pointer below"); 1291 collector()->promoted(false, // Not parallel 1292 (HeapWord*)res, obj->is_objArray(), obj_size); 1293 // promotion counters 1294 NOT_PRODUCT( 1295 _numObjectsPromoted++; 1296 _numWordsPromoted += 1297 (int)(CompactibleFreeListSpace::adjustObjectSize(obj->size())); 1298 ) 1299 } 1300 return res; 1301 } 1302 1303 1304 HeapWord* 1305 ConcurrentMarkSweepGeneration::allocation_limit_reached(Space* space, 1306 HeapWord* top, 1307 size_t word_sz) 1308 { 1309 return collector()->allocation_limit_reached(space, top, word_sz); 1310 } 1311 1312 // IMPORTANT: Notes on object size recognition in CMS. 1313 // --------------------------------------------------- 1314 // A block of storage in the CMS generation is always in 1315 // one of three states. A free block (FREE), an allocated 1316 // object (OBJECT) whose size() method reports the correct size, 1317 // and an intermediate state (TRANSIENT) in which its size cannot 1318 // be accurately determined. 1319 // STATE IDENTIFICATION: (32 bit and 64 bit w/o COOPS) 1320 // ----------------------------------------------------- 1321 // FREE: klass_word & 1 == 1; mark_word holds block size 1322 // 1323 // OBJECT: klass_word installed; klass_word != 0 && klass_word & 1 == 0; 1324 // obj->size() computes correct size 1325 // 1326 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT 1327 // 1328 // STATE IDENTIFICATION: (64 bit+COOPS) 1329 // ------------------------------------ 1330 // FREE: mark_word & CMS_FREE_BIT == 1; mark_word & ~CMS_FREE_BIT gives block_size 1331 // 1332 // OBJECT: klass_word installed; klass_word != 0; 1333 // obj->size() computes correct size 1334 // 1335 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT 1336 // 1337 // 1338 // STATE TRANSITION DIAGRAM 1339 // 1340 // mut / parnew mut / parnew 1341 // FREE --------------------> TRANSIENT ---------------------> OBJECT --| 1342 // ^ | 1343 // |------------------------ DEAD <------------------------------------| 1344 // sweep mut 1345 // 1346 // While a block is in TRANSIENT state its size cannot be determined 1347 // so readers will either need to come back later or stall until 1348 // the size can be determined. Note that for the case of direct 1349 // allocation, P-bits, when available, may be used to determine the 1350 // size of an object that may not yet have been initialized. 1351 1352 // Things to support parallel young-gen collection. 1353 oop 1354 ConcurrentMarkSweepGeneration::par_promote(int thread_num, 1355 oop old, markOop m, 1356 size_t word_sz) { 1357 #ifndef PRODUCT 1358 if (Universe::heap()->promotion_should_fail()) { 1359 return NULL; 1360 } 1361 #endif // #ifndef PRODUCT 1362 1363 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1364 PromotionInfo* promoInfo = &ps->promo; 1365 // if we are tracking promotions, then first ensure space for 1366 // promotion (including spooling space for saving header if necessary). 1367 // then allocate and copy, then track promoted info if needed. 1368 // When tracking (see PromotionInfo::track()), the mark word may 1369 // be displaced and in this case restoration of the mark word 1370 // occurs in the (oop_since_save_marks_)iterate phase. 1371 if (promoInfo->tracking() && !promoInfo->ensure_spooling_space()) { 1372 // Out of space for allocating spooling buffers; 1373 // try expanding and allocating spooling buffers. 1374 if (!expand_and_ensure_spooling_space(promoInfo)) { 1375 return NULL; 1376 } 1377 } 1378 assert(promoInfo->has_spooling_space(), "Control point invariant"); 1379 const size_t alloc_sz = CompactibleFreeListSpace::adjustObjectSize(word_sz); 1380 HeapWord* obj_ptr = ps->lab.alloc(alloc_sz); 1381 if (obj_ptr == NULL) { 1382 obj_ptr = expand_and_par_lab_allocate(ps, alloc_sz); 1383 if (obj_ptr == NULL) { 1384 return NULL; 1385 } 1386 } 1387 oop obj = oop(obj_ptr); 1388 OrderAccess::storestore(); 1389 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1390 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1391 // IMPORTANT: See note on object initialization for CMS above. 1392 // Otherwise, copy the object. Here we must be careful to insert the 1393 // klass pointer last, since this marks the block as an allocated object. 1394 // Except with compressed oops it's the mark word. 1395 HeapWord* old_ptr = (HeapWord*)old; 1396 // Restore the mark word copied above. 1397 obj->set_mark(m); 1398 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1399 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1400 OrderAccess::storestore(); 1401 1402 if (UseCompressedClassPointers) { 1403 // Copy gap missed by (aligned) header size calculation below 1404 obj->set_klass_gap(old->klass_gap()); 1405 } 1406 if (word_sz > (size_t)oopDesc::header_size()) { 1407 Copy::aligned_disjoint_words(old_ptr + oopDesc::header_size(), 1408 obj_ptr + oopDesc::header_size(), 1409 word_sz - oopDesc::header_size()); 1410 } 1411 1412 // Now we can track the promoted object, if necessary. We take care 1413 // to delay the transition from uninitialized to full object 1414 // (i.e., insertion of klass pointer) until after, so that it 1415 // atomically becomes a promoted object. 1416 if (promoInfo->tracking()) { 1417 promoInfo->track((PromotedObject*)obj, old->klass()); 1418 } 1419 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1420 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1421 assert(old->is_oop(), "Will use and dereference old klass ptr below"); 1422 1423 // Finally, install the klass pointer (this should be volatile). 1424 OrderAccess::storestore(); 1425 obj->set_klass(old->klass()); 1426 // We should now be able to calculate the right size for this object 1427 assert(obj->is_oop() && obj->size() == (int)word_sz, "Error, incorrect size computed for promoted object"); 1428 1429 collector()->promoted(true, // parallel 1430 obj_ptr, old->is_objArray(), word_sz); 1431 1432 NOT_PRODUCT( 1433 Atomic::inc_ptr(&_numObjectsPromoted); 1434 Atomic::add_ptr(alloc_sz, &_numWordsPromoted); 1435 ) 1436 1437 return obj; 1438 } 1439 1440 void 1441 ConcurrentMarkSweepGeneration:: 1442 par_promote_alloc_undo(int thread_num, 1443 HeapWord* obj, size_t word_sz) { 1444 // CMS does not support promotion undo. 1445 ShouldNotReachHere(); 1446 } 1447 1448 void 1449 ConcurrentMarkSweepGeneration:: 1450 par_promote_alloc_done(int thread_num) { 1451 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1452 ps->lab.retire(thread_num); 1453 } 1454 1455 void 1456 ConcurrentMarkSweepGeneration:: 1457 par_oop_since_save_marks_iterate_done(int thread_num) { 1458 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1459 ParScanWithoutBarrierClosure* dummy_cl = NULL; 1460 ps->promo.promoted_oops_iterate_nv(dummy_cl); 1461 } 1462 1463 bool ConcurrentMarkSweepGeneration::should_collect(bool full, 1464 size_t size, 1465 bool tlab) 1466 { 1467 // We allow a STW collection only if a full 1468 // collection was requested. 1469 return full || should_allocate(size, tlab); // FIX ME !!! 1470 // This and promotion failure handling are connected at the 1471 // hip and should be fixed by untying them. 1472 } 1473 1474 bool CMSCollector::shouldConcurrentCollect() { 1475 if (_full_gc_requested) { 1476 if (Verbose && PrintGCDetails) { 1477 gclog_or_tty->print_cr("CMSCollector: collect because of explicit " 1478 " gc request (or gc_locker)"); 1479 } 1480 return true; 1481 } 1482 1483 // For debugging purposes, change the type of collection. 1484 // If the rotation is not on the concurrent collection 1485 // type, don't start a concurrent collection. 1486 NOT_PRODUCT( 1487 if (RotateCMSCollectionTypes && 1488 (_cmsGen->debug_collection_type() != 1489 ConcurrentMarkSweepGeneration::Concurrent_collection_type)) { 1490 assert(_cmsGen->debug_collection_type() != 1491 ConcurrentMarkSweepGeneration::Unknown_collection_type, 1492 "Bad cms collection type"); 1493 return false; 1494 } 1495 ) 1496 1497 FreelistLocker x(this); 1498 // ------------------------------------------------------------------ 1499 // Print out lots of information which affects the initiation of 1500 // a collection. 1501 if (PrintCMSInitiationStatistics && stats().valid()) { 1502 gclog_or_tty->print("CMSCollector shouldConcurrentCollect: "); 1503 gclog_or_tty->stamp(); 1504 gclog_or_tty->print_cr(""); 1505 stats().print_on(gclog_or_tty); 1506 gclog_or_tty->print_cr("time_until_cms_gen_full %3.7f", 1507 stats().time_until_cms_gen_full()); 1508 gclog_or_tty->print_cr("free="SIZE_FORMAT, _cmsGen->free()); 1509 gclog_or_tty->print_cr("contiguous_available="SIZE_FORMAT, 1510 _cmsGen->contiguous_available()); 1511 gclog_or_tty->print_cr("promotion_rate=%g", stats().promotion_rate()); 1512 gclog_or_tty->print_cr("cms_allocation_rate=%g", stats().cms_allocation_rate()); 1513 gclog_or_tty->print_cr("occupancy=%3.7f", _cmsGen->occupancy()); 1514 gclog_or_tty->print_cr("initiatingOccupancy=%3.7f", _cmsGen->initiating_occupancy()); 1515 gclog_or_tty->print_cr("metadata initialized %d", 1516 MetaspaceGC::should_concurrent_collect()); 1517 } 1518 // ------------------------------------------------------------------ 1519 1520 // If the estimated time to complete a cms collection (cms_duration()) 1521 // is less than the estimated time remaining until the cms generation 1522 // is full, start a collection. 1523 if (!UseCMSInitiatingOccupancyOnly) { 1524 if (stats().valid()) { 1525 if (stats().time_until_cms_start() == 0.0) { 1526 return true; 1527 } 1528 } else { 1529 // We want to conservatively collect somewhat early in order 1530 // to try and "bootstrap" our CMS/promotion statistics; 1531 // this branch will not fire after the first successful CMS 1532 // collection because the stats should then be valid. 1533 if (_cmsGen->occupancy() >= _bootstrap_occupancy) { 1534 if (Verbose && PrintGCDetails) { 1535 gclog_or_tty->print_cr( 1536 " CMSCollector: collect for bootstrapping statistics:" 1537 " occupancy = %f, boot occupancy = %f", _cmsGen->occupancy(), 1538 _bootstrap_occupancy); 1539 } 1540 return true; 1541 } 1542 } 1543 } 1544 1545 // Otherwise, we start a collection cycle if 1546 // old gen want a collection cycle started. Each may use 1547 // an appropriate criterion for making this decision. 1548 // XXX We need to make sure that the gen expansion 1549 // criterion dovetails well with this. XXX NEED TO FIX THIS 1550 if (_cmsGen->should_concurrent_collect()) { 1551 if (Verbose && PrintGCDetails) { 1552 gclog_or_tty->print_cr("CMS old gen initiated"); 1553 } 1554 return true; 1555 } 1556 1557 // We start a collection if we believe an incremental collection may fail; 1558 // this is not likely to be productive in practice because it's probably too 1559 // late anyway. 1560 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1561 assert(gch->collector_policy()->is_two_generation_policy(), 1562 "You may want to check the correctness of the following"); 1563 if (gch->incremental_collection_will_fail(true /* consult_young */)) { 1564 if (Verbose && PrintGCDetails) { 1565 gclog_or_tty->print("CMSCollector: collect because incremental collection will fail "); 1566 } 1567 return true; 1568 } 1569 1570 if (MetaspaceGC::should_concurrent_collect()) { 1571 if (Verbose && PrintGCDetails) { 1572 gclog_or_tty->print("CMSCollector: collect for metadata allocation "); 1573 } 1574 return true; 1575 } 1576 1577 return false; 1578 } 1579 1580 void CMSCollector::set_did_compact(bool v) { _cmsGen->set_did_compact(v); } 1581 1582 // Clear _expansion_cause fields of constituent generations 1583 void CMSCollector::clear_expansion_cause() { 1584 _cmsGen->clear_expansion_cause(); 1585 } 1586 1587 // We should be conservative in starting a collection cycle. To 1588 // start too eagerly runs the risk of collecting too often in the 1589 // extreme. To collect too rarely falls back on full collections, 1590 // which works, even if not optimum in terms of concurrent work. 1591 // As a work around for too eagerly collecting, use the flag 1592 // UseCMSInitiatingOccupancyOnly. This also has the advantage of 1593 // giving the user an easily understandable way of controlling the 1594 // collections. 1595 // We want to start a new collection cycle if any of the following 1596 // conditions hold: 1597 // . our current occupancy exceeds the configured initiating occupancy 1598 // for this generation, or 1599 // . we recently needed to expand this space and have not, since that 1600 // expansion, done a collection of this generation, or 1601 // . the underlying space believes that it may be a good idea to initiate 1602 // a concurrent collection (this may be based on criteria such as the 1603 // following: the space uses linear allocation and linear allocation is 1604 // going to fail, or there is believed to be excessive fragmentation in 1605 // the generation, etc... or ... 1606 // [.(currently done by CMSCollector::shouldConcurrentCollect() only for 1607 // the case of the old generation; see CR 6543076): 1608 // we may be approaching a point at which allocation requests may fail because 1609 // we will be out of sufficient free space given allocation rate estimates.] 1610 bool ConcurrentMarkSweepGeneration::should_concurrent_collect() const { 1611 1612 assert_lock_strong(freelistLock()); 1613 if (occupancy() > initiating_occupancy()) { 1614 if (PrintGCDetails && Verbose) { 1615 gclog_or_tty->print(" %s: collect because of occupancy %f / %f ", 1616 short_name(), occupancy(), initiating_occupancy()); 1617 } 1618 return true; 1619 } 1620 if (UseCMSInitiatingOccupancyOnly) { 1621 return false; 1622 } 1623 if (expansion_cause() == CMSExpansionCause::_satisfy_allocation) { 1624 if (PrintGCDetails && Verbose) { 1625 gclog_or_tty->print(" %s: collect because expanded for allocation ", 1626 short_name()); 1627 } 1628 return true; 1629 } 1630 if (_cmsSpace->should_concurrent_collect()) { 1631 if (PrintGCDetails && Verbose) { 1632 gclog_or_tty->print(" %s: collect because cmsSpace says so ", 1633 short_name()); 1634 } 1635 return true; 1636 } 1637 return false; 1638 } 1639 1640 void ConcurrentMarkSweepGeneration::collect(bool full, 1641 bool clear_all_soft_refs, 1642 size_t size, 1643 bool tlab) 1644 { 1645 collector()->collect(full, clear_all_soft_refs, size, tlab); 1646 } 1647 1648 void CMSCollector::collect(bool full, 1649 bool clear_all_soft_refs, 1650 size_t size, 1651 bool tlab) 1652 { 1653 if (!UseCMSCollectionPassing && _collectorState > Idling) { 1654 // For debugging purposes skip the collection if the state 1655 // is not currently idle 1656 if (TraceCMSState) { 1657 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " skipped full:%d CMS state %d", 1658 Thread::current(), full, _collectorState); 1659 } 1660 return; 1661 } 1662 1663 // The following "if" branch is present for defensive reasons. 1664 // In the current uses of this interface, it can be replaced with: 1665 // assert(!GC_locker.is_active(), "Can't be called otherwise"); 1666 // But I am not placing that assert here to allow future 1667 // generality in invoking this interface. 1668 if (GC_locker::is_active()) { 1669 // A consistency test for GC_locker 1670 assert(GC_locker::needs_gc(), "Should have been set already"); 1671 // Skip this foreground collection, instead 1672 // expanding the heap if necessary. 1673 // Need the free list locks for the call to free() in compute_new_size() 1674 compute_new_size(); 1675 return; 1676 } 1677 acquire_control_and_collect(full, clear_all_soft_refs); 1678 _full_gcs_since_conc_gc++; 1679 } 1680 1681 void CMSCollector::request_full_gc(unsigned int full_gc_count, GCCause::Cause cause) { 1682 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1683 unsigned int gc_count = gch->total_full_collections(); 1684 if (gc_count == full_gc_count) { 1685 MutexLockerEx y(CGC_lock, Mutex::_no_safepoint_check_flag); 1686 _full_gc_requested = true; 1687 _full_gc_cause = cause; 1688 CGC_lock->notify(); // nudge CMS thread 1689 } else { 1690 assert(gc_count > full_gc_count, "Error: causal loop"); 1691 } 1692 } 1693 1694 bool CMSCollector::is_external_interruption() { 1695 GCCause::Cause cause = GenCollectedHeap::heap()->gc_cause(); 1696 return GCCause::is_user_requested_gc(cause) || 1697 GCCause::is_serviceability_requested_gc(cause); 1698 } 1699 1700 void CMSCollector::report_concurrent_mode_interruption() { 1701 if (is_external_interruption()) { 1702 if (PrintGCDetails) { 1703 gclog_or_tty->print(" (concurrent mode interrupted)"); 1704 } 1705 } else { 1706 if (PrintGCDetails) { 1707 gclog_or_tty->print(" (concurrent mode failure)"); 1708 } 1709 _gc_tracer_cm->report_concurrent_mode_failure(); 1710 } 1711 } 1712 1713 1714 // The foreground and background collectors need to coordinate in order 1715 // to make sure that they do not mutually interfere with CMS collections. 1716 // When a background collection is active, 1717 // the foreground collector may need to take over (preempt) and 1718 // synchronously complete an ongoing collection. Depending on the 1719 // frequency of the background collections and the heap usage 1720 // of the application, this preemption can be seldom or frequent. 1721 // There are only certain 1722 // points in the background collection that the "collection-baton" 1723 // can be passed to the foreground collector. 1724 // 1725 // The foreground collector will wait for the baton before 1726 // starting any part of the collection. The foreground collector 1727 // will only wait at one location. 1728 // 1729 // The background collector will yield the baton before starting a new 1730 // phase of the collection (e.g., before initial marking, marking from roots, 1731 // precleaning, final re-mark, sweep etc.) This is normally done at the head 1732 // of the loop which switches the phases. The background collector does some 1733 // of the phases (initial mark, final re-mark) with the world stopped. 1734 // Because of locking involved in stopping the world, 1735 // the foreground collector should not block waiting for the background 1736 // collector when it is doing a stop-the-world phase. The background 1737 // collector will yield the baton at an additional point just before 1738 // it enters a stop-the-world phase. Once the world is stopped, the 1739 // background collector checks the phase of the collection. If the 1740 // phase has not changed, it proceeds with the collection. If the 1741 // phase has changed, it skips that phase of the collection. See 1742 // the comments on the use of the Heap_lock in collect_in_background(). 1743 // 1744 // Variable used in baton passing. 1745 // _foregroundGCIsActive - Set to true by the foreground collector when 1746 // it wants the baton. The foreground clears it when it has finished 1747 // the collection. 1748 // _foregroundGCShouldWait - Set to true by the background collector 1749 // when it is running. The foreground collector waits while 1750 // _foregroundGCShouldWait is true. 1751 // CGC_lock - monitor used to protect access to the above variables 1752 // and to notify the foreground and background collectors. 1753 // _collectorState - current state of the CMS collection. 1754 // 1755 // The foreground collector 1756 // acquires the CGC_lock 1757 // sets _foregroundGCIsActive 1758 // waits on the CGC_lock for _foregroundGCShouldWait to be false 1759 // various locks acquired in preparation for the collection 1760 // are released so as not to block the background collector 1761 // that is in the midst of a collection 1762 // proceeds with the collection 1763 // clears _foregroundGCIsActive 1764 // returns 1765 // 1766 // The background collector in a loop iterating on the phases of the 1767 // collection 1768 // acquires the CGC_lock 1769 // sets _foregroundGCShouldWait 1770 // if _foregroundGCIsActive is set 1771 // clears _foregroundGCShouldWait, notifies _CGC_lock 1772 // waits on _CGC_lock for _foregroundGCIsActive to become false 1773 // and exits the loop. 1774 // otherwise 1775 // proceed with that phase of the collection 1776 // if the phase is a stop-the-world phase, 1777 // yield the baton once more just before enqueueing 1778 // the stop-world CMS operation (executed by the VM thread). 1779 // returns after all phases of the collection are done 1780 // 1781 1782 void CMSCollector::acquire_control_and_collect(bool full, 1783 bool clear_all_soft_refs) { 1784 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint"); 1785 assert(!Thread::current()->is_ConcurrentGC_thread(), 1786 "shouldn't try to acquire control from self!"); 1787 1788 // Start the protocol for acquiring control of the 1789 // collection from the background collector (aka CMS thread). 1790 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 1791 "VM thread should have CMS token"); 1792 // Remember the possibly interrupted state of an ongoing 1793 // concurrent collection 1794 CollectorState first_state = _collectorState; 1795 1796 // Signal to a possibly ongoing concurrent collection that 1797 // we want to do a foreground collection. 1798 _foregroundGCIsActive = true; 1799 1800 // Disable incremental mode during a foreground collection. 1801 ICMSDisabler icms_disabler; 1802 1803 // release locks and wait for a notify from the background collector 1804 // releasing the locks in only necessary for phases which 1805 // do yields to improve the granularity of the collection. 1806 assert_lock_strong(bitMapLock()); 1807 // We need to lock the Free list lock for the space that we are 1808 // currently collecting. 1809 assert(haveFreelistLocks(), "Must be holding free list locks"); 1810 bitMapLock()->unlock(); 1811 releaseFreelistLocks(); 1812 { 1813 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1814 if (_foregroundGCShouldWait) { 1815 // We are going to be waiting for action for the CMS thread; 1816 // it had better not be gone (for instance at shutdown)! 1817 assert(ConcurrentMarkSweepThread::cmst() != NULL, 1818 "CMS thread must be running"); 1819 // Wait here until the background collector gives us the go-ahead 1820 ConcurrentMarkSweepThread::clear_CMS_flag( 1821 ConcurrentMarkSweepThread::CMS_vm_has_token); // release token 1822 // Get a possibly blocked CMS thread going: 1823 // Note that we set _foregroundGCIsActive true above, 1824 // without protection of the CGC_lock. 1825 CGC_lock->notify(); 1826 assert(!ConcurrentMarkSweepThread::vm_thread_wants_cms_token(), 1827 "Possible deadlock"); 1828 while (_foregroundGCShouldWait) { 1829 // wait for notification 1830 CGC_lock->wait(Mutex::_no_safepoint_check_flag); 1831 // Possibility of delay/starvation here, since CMS token does 1832 // not know to give priority to VM thread? Actually, i think 1833 // there wouldn't be any delay/starvation, but the proof of 1834 // that "fact" (?) appears non-trivial. XXX 20011219YSR 1835 } 1836 ConcurrentMarkSweepThread::set_CMS_flag( 1837 ConcurrentMarkSweepThread::CMS_vm_has_token); 1838 } 1839 } 1840 // The CMS_token is already held. Get back the other locks. 1841 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 1842 "VM thread should have CMS token"); 1843 getFreelistLocks(); 1844 bitMapLock()->lock_without_safepoint_check(); 1845 if (TraceCMSState) { 1846 gclog_or_tty->print_cr("CMS foreground collector has asked for control " 1847 INTPTR_FORMAT " with first state %d", Thread::current(), first_state); 1848 gclog_or_tty->print_cr(" gets control with state %d", _collectorState); 1849 } 1850 1851 // Check if we need to do a compaction, or if not, whether 1852 // we need to start the mark-sweep from scratch. 1853 bool should_compact = false; 1854 bool should_start_over = false; 1855 decide_foreground_collection_type(clear_all_soft_refs, 1856 &should_compact, &should_start_over); 1857 1858 NOT_PRODUCT( 1859 if (RotateCMSCollectionTypes) { 1860 if (_cmsGen->debug_collection_type() == 1861 ConcurrentMarkSweepGeneration::MSC_foreground_collection_type) { 1862 should_compact = true; 1863 } else if (_cmsGen->debug_collection_type() == 1864 ConcurrentMarkSweepGeneration::MS_foreground_collection_type) { 1865 should_compact = false; 1866 } 1867 } 1868 ) 1869 1870 if (first_state > Idling) { 1871 report_concurrent_mode_interruption(); 1872 } 1873 1874 set_did_compact(should_compact); 1875 if (should_compact) { 1876 // If the collection is being acquired from the background 1877 // collector, there may be references on the discovered 1878 // references lists that have NULL referents (being those 1879 // that were concurrently cleared by a mutator) or 1880 // that are no longer active (having been enqueued concurrently 1881 // by the mutator). 1882 // Scrub the list of those references because Mark-Sweep-Compact 1883 // code assumes referents are not NULL and that all discovered 1884 // Reference objects are active. 1885 ref_processor()->clean_up_discovered_references(); 1886 1887 if (first_state > Idling) { 1888 save_heap_summary(); 1889 } 1890 1891 do_compaction_work(clear_all_soft_refs); 1892 1893 // Has the GC time limit been exceeded? 1894 DefNewGeneration* young_gen = _young_gen->as_DefNewGeneration(); 1895 size_t max_eden_size = young_gen->max_capacity() - 1896 young_gen->to()->capacity() - 1897 young_gen->from()->capacity(); 1898 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1899 GCCause::Cause gc_cause = gch->gc_cause(); 1900 size_policy()->check_gc_overhead_limit(_young_gen->used(), 1901 young_gen->eden()->used(), 1902 _cmsGen->max_capacity(), 1903 max_eden_size, 1904 full, 1905 gc_cause, 1906 gch->collector_policy()); 1907 } else { 1908 do_mark_sweep_work(clear_all_soft_refs, first_state, 1909 should_start_over); 1910 } 1911 // Reset the expansion cause, now that we just completed 1912 // a collection cycle. 1913 clear_expansion_cause(); 1914 _foregroundGCIsActive = false; 1915 return; 1916 } 1917 1918 // Resize the tenured generation 1919 // after obtaining the free list locks for the 1920 // two generations. 1921 void CMSCollector::compute_new_size() { 1922 assert_locked_or_safepoint(Heap_lock); 1923 FreelistLocker z(this); 1924 MetaspaceGC::compute_new_size(); 1925 _cmsGen->compute_new_size_free_list(); 1926 } 1927 1928 // A work method used by foreground collection to determine 1929 // what type of collection (compacting or not, continuing or fresh) 1930 // it should do. 1931 // NOTE: the intent is to make UseCMSCompactAtFullCollection 1932 // and CMSCompactWhenClearAllSoftRefs the default in the future 1933 // and do away with the flags after a suitable period. 1934 void CMSCollector::decide_foreground_collection_type( 1935 bool clear_all_soft_refs, bool* should_compact, 1936 bool* should_start_over) { 1937 // Normally, we'll compact only if the UseCMSCompactAtFullCollection 1938 // flag is set, and we have either requested a System.gc() or 1939 // the number of full gc's since the last concurrent cycle 1940 // has exceeded the threshold set by CMSFullGCsBeforeCompaction, 1941 // or if an incremental collection has failed 1942 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1943 assert(gch->collector_policy()->is_two_generation_policy(), 1944 "You may want to check the correctness of the following"); 1945 // Inform cms gen if this was due to partial collection failing. 1946 // The CMS gen may use this fact to determine its expansion policy. 1947 if (gch->incremental_collection_will_fail(false /* don't consult_young */)) { 1948 assert(!_cmsGen->incremental_collection_failed(), 1949 "Should have been noticed, reacted to and cleared"); 1950 _cmsGen->set_incremental_collection_failed(); 1951 } 1952 *should_compact = 1953 UseCMSCompactAtFullCollection && 1954 ((_full_gcs_since_conc_gc >= CMSFullGCsBeforeCompaction) || 1955 GCCause::is_user_requested_gc(gch->gc_cause()) || 1956 gch->incremental_collection_will_fail(true /* consult_young */)); 1957 *should_start_over = false; 1958 if (clear_all_soft_refs && !*should_compact) { 1959 // We are about to do a last ditch collection attempt 1960 // so it would normally make sense to do a compaction 1961 // to reclaim as much space as possible. 1962 if (CMSCompactWhenClearAllSoftRefs) { 1963 // Default: The rationale is that in this case either 1964 // we are past the final marking phase, in which case 1965 // we'd have to start over, or so little has been done 1966 // that there's little point in saving that work. Compaction 1967 // appears to be the sensible choice in either case. 1968 *should_compact = true; 1969 } else { 1970 // We have been asked to clear all soft refs, but not to 1971 // compact. Make sure that we aren't past the final checkpoint 1972 // phase, for that is where we process soft refs. If we are already 1973 // past that phase, we'll need to redo the refs discovery phase and 1974 // if necessary clear soft refs that weren't previously 1975 // cleared. We do so by remembering the phase in which 1976 // we came in, and if we are past the refs processing 1977 // phase, we'll choose to just redo the mark-sweep 1978 // collection from scratch. 1979 if (_collectorState > FinalMarking) { 1980 // We are past the refs processing phase; 1981 // start over and do a fresh synchronous CMS cycle 1982 _collectorState = Resetting; // skip to reset to start new cycle 1983 reset(false /* == !asynch */); 1984 *should_start_over = true; 1985 } // else we can continue a possibly ongoing current cycle 1986 } 1987 } 1988 } 1989 1990 // A work method used by the foreground collector to do 1991 // a mark-sweep-compact. 1992 void CMSCollector::do_compaction_work(bool clear_all_soft_refs) { 1993 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1994 1995 STWGCTimer* gc_timer = GenMarkSweep::gc_timer(); 1996 gc_timer->register_gc_start(); 1997 1998 SerialOldTracer* gc_tracer = GenMarkSweep::gc_tracer(); 1999 gc_tracer->report_gc_start(gch->gc_cause(), gc_timer->gc_start()); 2000 2001 GCTraceTime t("CMS:MSC ", PrintGCDetails && Verbose, true, NULL); 2002 if (PrintGC && Verbose && !(GCCause::is_user_requested_gc(gch->gc_cause()))) { 2003 gclog_or_tty->print_cr("Compact ConcurrentMarkSweepGeneration after %d " 2004 "collections passed to foreground collector", _full_gcs_since_conc_gc); 2005 } 2006 2007 // Sample collection interval time and reset for collection pause. 2008 if (UseAdaptiveSizePolicy) { 2009 size_policy()->msc_collection_begin(); 2010 } 2011 2012 // Temporarily widen the span of the weak reference processing to 2013 // the entire heap. 2014 MemRegion new_span(GenCollectedHeap::heap()->reserved_region()); 2015 ReferenceProcessorSpanMutator rp_mut_span(ref_processor(), new_span); 2016 // Temporarily, clear the "is_alive_non_header" field of the 2017 // reference processor. 2018 ReferenceProcessorIsAliveMutator rp_mut_closure(ref_processor(), NULL); 2019 // Temporarily make reference _processing_ single threaded (non-MT). 2020 ReferenceProcessorMTProcMutator rp_mut_mt_processing(ref_processor(), false); 2021 // Temporarily make refs discovery atomic 2022 ReferenceProcessorAtomicMutator rp_mut_atomic(ref_processor(), true); 2023 // Temporarily make reference _discovery_ single threaded (non-MT) 2024 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false); 2025 2026 ref_processor()->set_enqueuing_is_done(false); 2027 ref_processor()->enable_discovery(false /*verify_disabled*/, false /*check_no_refs*/); 2028 ref_processor()->setup_policy(clear_all_soft_refs); 2029 // If an asynchronous collection finishes, the _modUnionTable is 2030 // all clear. If we are assuming the collection from an asynchronous 2031 // collection, clear the _modUnionTable. 2032 assert(_collectorState != Idling || _modUnionTable.isAllClear(), 2033 "_modUnionTable should be clear if the baton was not passed"); 2034 _modUnionTable.clear_all(); 2035 assert(_collectorState != Idling || _ct->klass_rem_set()->mod_union_is_clear(), 2036 "mod union for klasses should be clear if the baton was passed"); 2037 _ct->klass_rem_set()->clear_mod_union(); 2038 2039 // We must adjust the allocation statistics being maintained 2040 // in the free list space. We do so by reading and clearing 2041 // the sweep timer and updating the block flux rate estimates below. 2042 assert(!_intra_sweep_timer.is_active(), "_intra_sweep_timer should be inactive"); 2043 if (_inter_sweep_timer.is_active()) { 2044 _inter_sweep_timer.stop(); 2045 // Note that we do not use this sample to update the _inter_sweep_estimate. 2046 _cmsGen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()), 2047 _inter_sweep_estimate.padded_average(), 2048 _intra_sweep_estimate.padded_average()); 2049 } 2050 2051 GenMarkSweep::invoke_at_safepoint(_cmsGen->level(), 2052 ref_processor(), clear_all_soft_refs); 2053 #ifdef ASSERT 2054 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 2055 size_t free_size = cms_space->free(); 2056 assert(free_size == 2057 pointer_delta(cms_space->end(), cms_space->compaction_top()) 2058 * HeapWordSize, 2059 "All the free space should be compacted into one chunk at top"); 2060 assert(cms_space->dictionary()->total_chunk_size( 2061 debug_only(cms_space->freelistLock())) == 0 || 2062 cms_space->totalSizeInIndexedFreeLists() == 0, 2063 "All the free space should be in a single chunk"); 2064 size_t num = cms_space->totalCount(); 2065 assert((free_size == 0 && num == 0) || 2066 (free_size > 0 && (num == 1 || num == 2)), 2067 "There should be at most 2 free chunks after compaction"); 2068 #endif // ASSERT 2069 _collectorState = Resetting; 2070 assert(_restart_addr == NULL, 2071 "Should have been NULL'd before baton was passed"); 2072 reset(false /* == !asynch */); 2073 _cmsGen->reset_after_compaction(); 2074 _concurrent_cycles_since_last_unload = 0; 2075 2076 // Clear any data recorded in the PLAB chunk arrays. 2077 if (_survivor_plab_array != NULL) { 2078 reset_survivor_plab_arrays(); 2079 } 2080 2081 // Adjust the per-size allocation stats for the next epoch. 2082 _cmsGen->cmsSpace()->endSweepFLCensus(sweep_count() /* fake */); 2083 // Restart the "inter sweep timer" for the next epoch. 2084 _inter_sweep_timer.reset(); 2085 _inter_sweep_timer.start(); 2086 2087 // Sample collection pause time and reset for collection interval. 2088 if (UseAdaptiveSizePolicy) { 2089 size_policy()->msc_collection_end(gch->gc_cause()); 2090 } 2091 2092 gc_timer->register_gc_end(); 2093 2094 gc_tracer->report_gc_end(gc_timer->gc_end(), gc_timer->time_partitions()); 2095 2096 // For a mark-sweep-compact, compute_new_size() will be called 2097 // in the heap's do_collection() method. 2098 } 2099 2100 // A work method used by the foreground collector to do 2101 // a mark-sweep, after taking over from a possibly on-going 2102 // concurrent mark-sweep collection. 2103 void CMSCollector::do_mark_sweep_work(bool clear_all_soft_refs, 2104 CollectorState first_state, bool should_start_over) { 2105 if (PrintGC && Verbose) { 2106 gclog_or_tty->print_cr("Pass concurrent collection to foreground " 2107 "collector with count %d", 2108 _full_gcs_since_conc_gc); 2109 } 2110 switch (_collectorState) { 2111 case Idling: 2112 if (first_state == Idling || should_start_over) { 2113 // The background GC was not active, or should 2114 // restarted from scratch; start the cycle. 2115 _collectorState = InitialMarking; 2116 } 2117 // If first_state was not Idling, then a background GC 2118 // was in progress and has now finished. No need to do it 2119 // again. Leave the state as Idling. 2120 break; 2121 case Precleaning: 2122 // In the foreground case don't do the precleaning since 2123 // it is not done concurrently and there is extra work 2124 // required. 2125 _collectorState = FinalMarking; 2126 } 2127 collect_in_foreground(clear_all_soft_refs, GenCollectedHeap::heap()->gc_cause()); 2128 2129 // For a mark-sweep, compute_new_size() will be called 2130 // in the heap's do_collection() method. 2131 } 2132 2133 2134 void CMSCollector::print_eden_and_survivor_chunk_arrays() { 2135 DefNewGeneration* dng = _young_gen->as_DefNewGeneration(); 2136 EdenSpace* eden_space = dng->eden(); 2137 ContiguousSpace* from_space = dng->from(); 2138 ContiguousSpace* to_space = dng->to(); 2139 // Eden 2140 if (_eden_chunk_array != NULL) { 2141 gclog_or_tty->print_cr("eden " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")", 2142 eden_space->bottom(), eden_space->top(), 2143 eden_space->end(), eden_space->capacity()); 2144 gclog_or_tty->print_cr("_eden_chunk_index=" SIZE_FORMAT ", " 2145 "_eden_chunk_capacity=" SIZE_FORMAT, 2146 _eden_chunk_index, _eden_chunk_capacity); 2147 for (size_t i = 0; i < _eden_chunk_index; i++) { 2148 gclog_or_tty->print_cr("_eden_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT, 2149 i, _eden_chunk_array[i]); 2150 } 2151 } 2152 // Survivor 2153 if (_survivor_chunk_array != NULL) { 2154 gclog_or_tty->print_cr("survivor " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")", 2155 from_space->bottom(), from_space->top(), 2156 from_space->end(), from_space->capacity()); 2157 gclog_or_tty->print_cr("_survivor_chunk_index=" SIZE_FORMAT ", " 2158 "_survivor_chunk_capacity=" SIZE_FORMAT, 2159 _survivor_chunk_index, _survivor_chunk_capacity); 2160 for (size_t i = 0; i < _survivor_chunk_index; i++) { 2161 gclog_or_tty->print_cr("_survivor_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT, 2162 i, _survivor_chunk_array[i]); 2163 } 2164 } 2165 } 2166 2167 void CMSCollector::getFreelistLocks() const { 2168 // Get locks for all free lists in all generations that this 2169 // collector is responsible for 2170 _cmsGen->freelistLock()->lock_without_safepoint_check(); 2171 } 2172 2173 void CMSCollector::releaseFreelistLocks() const { 2174 // Release locks for all free lists in all generations that this 2175 // collector is responsible for 2176 _cmsGen->freelistLock()->unlock(); 2177 } 2178 2179 bool CMSCollector::haveFreelistLocks() const { 2180 // Check locks for all free lists in all generations that this 2181 // collector is responsible for 2182 assert_lock_strong(_cmsGen->freelistLock()); 2183 PRODUCT_ONLY(ShouldNotReachHere()); 2184 return true; 2185 } 2186 2187 // A utility class that is used by the CMS collector to 2188 // temporarily "release" the foreground collector from its 2189 // usual obligation to wait for the background collector to 2190 // complete an ongoing phase before proceeding. 2191 class ReleaseForegroundGC: public StackObj { 2192 private: 2193 CMSCollector* _c; 2194 public: 2195 ReleaseForegroundGC(CMSCollector* c) : _c(c) { 2196 assert(_c->_foregroundGCShouldWait, "Else should not need to call"); 2197 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 2198 // allow a potentially blocked foreground collector to proceed 2199 _c->_foregroundGCShouldWait = false; 2200 if (_c->_foregroundGCIsActive) { 2201 CGC_lock->notify(); 2202 } 2203 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 2204 "Possible deadlock"); 2205 } 2206 2207 ~ReleaseForegroundGC() { 2208 assert(!_c->_foregroundGCShouldWait, "Usage protocol violation?"); 2209 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 2210 _c->_foregroundGCShouldWait = true; 2211 } 2212 }; 2213 2214 // There are separate collect_in_background and collect_in_foreground because of 2215 // the different locking requirements of the background collector and the 2216 // foreground collector. There was originally an attempt to share 2217 // one "collect" method between the background collector and the foreground 2218 // collector but the if-then-else required made it cleaner to have 2219 // separate methods. 2220 void CMSCollector::collect_in_background(bool clear_all_soft_refs, GCCause::Cause cause) { 2221 assert(Thread::current()->is_ConcurrentGC_thread(), 2222 "A CMS asynchronous collection is only allowed on a CMS thread."); 2223 2224 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2225 { 2226 bool safepoint_check = Mutex::_no_safepoint_check_flag; 2227 MutexLockerEx hl(Heap_lock, safepoint_check); 2228 FreelistLocker fll(this); 2229 MutexLockerEx x(CGC_lock, safepoint_check); 2230 if (_foregroundGCIsActive || !UseAsyncConcMarkSweepGC) { 2231 // The foreground collector is active or we're 2232 // not using asynchronous collections. Skip this 2233 // background collection. 2234 assert(!_foregroundGCShouldWait, "Should be clear"); 2235 return; 2236 } else { 2237 assert(_collectorState == Idling, "Should be idling before start."); 2238 _collectorState = InitialMarking; 2239 register_gc_start(cause); 2240 // Reset the expansion cause, now that we are about to begin 2241 // a new cycle. 2242 clear_expansion_cause(); 2243 2244 // Clear the MetaspaceGC flag since a concurrent collection 2245 // is starting but also clear it after the collection. 2246 MetaspaceGC::set_should_concurrent_collect(false); 2247 } 2248 // Decide if we want to enable class unloading as part of the 2249 // ensuing concurrent GC cycle. 2250 update_should_unload_classes(); 2251 _full_gc_requested = false; // acks all outstanding full gc requests 2252 _full_gc_cause = GCCause::_no_gc; 2253 // Signal that we are about to start a collection 2254 gch->increment_total_full_collections(); // ... starting a collection cycle 2255 _collection_count_start = gch->total_full_collections(); 2256 } 2257 2258 // Used for PrintGC 2259 size_t prev_used; 2260 if (PrintGC && Verbose) { 2261 prev_used = _cmsGen->used(); // XXXPERM 2262 } 2263 2264 // The change of the collection state is normally done at this level; 2265 // the exceptions are phases that are executed while the world is 2266 // stopped. For those phases the change of state is done while the 2267 // world is stopped. For baton passing purposes this allows the 2268 // background collector to finish the phase and change state atomically. 2269 // The foreground collector cannot wait on a phase that is done 2270 // while the world is stopped because the foreground collector already 2271 // has the world stopped and would deadlock. 2272 while (_collectorState != Idling) { 2273 if (TraceCMSState) { 2274 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " in CMS state %d", 2275 Thread::current(), _collectorState); 2276 } 2277 // The foreground collector 2278 // holds the Heap_lock throughout its collection. 2279 // holds the CMS token (but not the lock) 2280 // except while it is waiting for the background collector to yield. 2281 // 2282 // The foreground collector should be blocked (not for long) 2283 // if the background collector is about to start a phase 2284 // executed with world stopped. If the background 2285 // collector has already started such a phase, the 2286 // foreground collector is blocked waiting for the 2287 // Heap_lock. The stop-world phases (InitialMarking and FinalMarking) 2288 // are executed in the VM thread. 2289 // 2290 // The locking order is 2291 // PendingListLock (PLL) -- if applicable (FinalMarking) 2292 // Heap_lock (both this & PLL locked in VM_CMS_Operation::prologue()) 2293 // CMS token (claimed in 2294 // stop_world_and_do() --> 2295 // safepoint_synchronize() --> 2296 // CMSThread::synchronize()) 2297 2298 { 2299 // Check if the FG collector wants us to yield. 2300 CMSTokenSync x(true); // is cms thread 2301 if (waitForForegroundGC()) { 2302 // We yielded to a foreground GC, nothing more to be 2303 // done this round. 2304 assert(_foregroundGCShouldWait == false, "We set it to false in " 2305 "waitForForegroundGC()"); 2306 if (TraceCMSState) { 2307 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT 2308 " exiting collection CMS state %d", 2309 Thread::current(), _collectorState); 2310 } 2311 return; 2312 } else { 2313 // The background collector can run but check to see if the 2314 // foreground collector has done a collection while the 2315 // background collector was waiting to get the CGC_lock 2316 // above. If yes, break so that _foregroundGCShouldWait 2317 // is cleared before returning. 2318 if (_collectorState == Idling) { 2319 break; 2320 } 2321 } 2322 } 2323 2324 assert(_foregroundGCShouldWait, "Foreground collector, if active, " 2325 "should be waiting"); 2326 2327 switch (_collectorState) { 2328 case InitialMarking: 2329 { 2330 ReleaseForegroundGC x(this); 2331 stats().record_cms_begin(); 2332 VM_CMS_Initial_Mark initial_mark_op(this); 2333 VMThread::execute(&initial_mark_op); 2334 } 2335 // The collector state may be any legal state at this point 2336 // since the background collector may have yielded to the 2337 // foreground collector. 2338 break; 2339 case Marking: 2340 // initial marking in checkpointRootsInitialWork has been completed 2341 if (markFromRoots(true)) { // we were successful 2342 assert(_collectorState == Precleaning, "Collector state should " 2343 "have changed"); 2344 } else { 2345 assert(_foregroundGCIsActive, "Internal state inconsistency"); 2346 } 2347 break; 2348 case Precleaning: 2349 if (UseAdaptiveSizePolicy) { 2350 size_policy()->concurrent_precleaning_begin(); 2351 } 2352 // marking from roots in markFromRoots has been completed 2353 preclean(); 2354 if (UseAdaptiveSizePolicy) { 2355 size_policy()->concurrent_precleaning_end(); 2356 } 2357 assert(_collectorState == AbortablePreclean || 2358 _collectorState == FinalMarking, 2359 "Collector state should have changed"); 2360 break; 2361 case AbortablePreclean: 2362 if (UseAdaptiveSizePolicy) { 2363 size_policy()->concurrent_phases_resume(); 2364 } 2365 abortable_preclean(); 2366 if (UseAdaptiveSizePolicy) { 2367 size_policy()->concurrent_precleaning_end(); 2368 } 2369 assert(_collectorState == FinalMarking, "Collector state should " 2370 "have changed"); 2371 break; 2372 case FinalMarking: 2373 { 2374 ReleaseForegroundGC x(this); 2375 2376 VM_CMS_Final_Remark final_remark_op(this); 2377 VMThread::execute(&final_remark_op); 2378 } 2379 assert(_foregroundGCShouldWait, "block post-condition"); 2380 break; 2381 case Sweeping: 2382 if (UseAdaptiveSizePolicy) { 2383 size_policy()->concurrent_sweeping_begin(); 2384 } 2385 // final marking in checkpointRootsFinal has been completed 2386 sweep(true); 2387 assert(_collectorState == Resizing, "Collector state change " 2388 "to Resizing must be done under the free_list_lock"); 2389 _full_gcs_since_conc_gc = 0; 2390 2391 // Stop the timers for adaptive size policy for the concurrent phases 2392 if (UseAdaptiveSizePolicy) { 2393 size_policy()->concurrent_sweeping_end(); 2394 size_policy()->concurrent_phases_end(gch->gc_cause(), 2395 gch->prev_gen(_cmsGen)->capacity(), 2396 _cmsGen->free()); 2397 } 2398 2399 case Resizing: { 2400 // Sweeping has been completed... 2401 // At this point the background collection has completed. 2402 // Don't move the call to compute_new_size() down 2403 // into code that might be executed if the background 2404 // collection was preempted. 2405 { 2406 ReleaseForegroundGC x(this); // unblock FG collection 2407 MutexLockerEx y(Heap_lock, Mutex::_no_safepoint_check_flag); 2408 CMSTokenSync z(true); // not strictly needed. 2409 if (_collectorState == Resizing) { 2410 compute_new_size(); 2411 save_heap_summary(); 2412 _collectorState = Resetting; 2413 } else { 2414 assert(_collectorState == Idling, "The state should only change" 2415 " because the foreground collector has finished the collection"); 2416 } 2417 } 2418 break; 2419 } 2420 case Resetting: 2421 // CMS heap resizing has been completed 2422 reset(true); 2423 assert(_collectorState == Idling, "Collector state should " 2424 "have changed"); 2425 2426 MetaspaceGC::set_should_concurrent_collect(false); 2427 2428 stats().record_cms_end(); 2429 // Don't move the concurrent_phases_end() and compute_new_size() 2430 // calls to here because a preempted background collection 2431 // has it's state set to "Resetting". 2432 break; 2433 case Idling: 2434 default: 2435 ShouldNotReachHere(); 2436 break; 2437 } 2438 if (TraceCMSState) { 2439 gclog_or_tty->print_cr(" Thread " INTPTR_FORMAT " done - next CMS state %d", 2440 Thread::current(), _collectorState); 2441 } 2442 assert(_foregroundGCShouldWait, "block post-condition"); 2443 } 2444 2445 // Should this be in gc_epilogue? 2446 collector_policy()->counters()->update_counters(); 2447 2448 { 2449 // Clear _foregroundGCShouldWait and, in the event that the 2450 // foreground collector is waiting, notify it, before 2451 // returning. 2452 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 2453 _foregroundGCShouldWait = false; 2454 if (_foregroundGCIsActive) { 2455 CGC_lock->notify(); 2456 } 2457 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 2458 "Possible deadlock"); 2459 } 2460 if (TraceCMSState) { 2461 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT 2462 " exiting collection CMS state %d", 2463 Thread::current(), _collectorState); 2464 } 2465 if (PrintGC && Verbose) { 2466 _cmsGen->print_heap_change(prev_used); 2467 } 2468 } 2469 2470 void CMSCollector::register_foreground_gc_start(GCCause::Cause cause) { 2471 if (!_cms_start_registered) { 2472 register_gc_start(cause); 2473 } 2474 } 2475 2476 void CMSCollector::register_gc_start(GCCause::Cause cause) { 2477 _cms_start_registered = true; 2478 _gc_timer_cm->register_gc_start(); 2479 _gc_tracer_cm->report_gc_start(cause, _gc_timer_cm->gc_start()); 2480 } 2481 2482 void CMSCollector::register_gc_end() { 2483 if (_cms_start_registered) { 2484 report_heap_summary(GCWhen::AfterGC); 2485 2486 _gc_timer_cm->register_gc_end(); 2487 _gc_tracer_cm->report_gc_end(_gc_timer_cm->gc_end(), _gc_timer_cm->time_partitions()); 2488 _cms_start_registered = false; 2489 } 2490 } 2491 2492 void CMSCollector::save_heap_summary() { 2493 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2494 _last_heap_summary = gch->create_heap_summary(); 2495 _last_metaspace_summary = gch->create_metaspace_summary(); 2496 } 2497 2498 void CMSCollector::report_heap_summary(GCWhen::Type when) { 2499 _gc_tracer_cm->report_gc_heap_summary(when, _last_heap_summary); 2500 _gc_tracer_cm->report_metaspace_summary(when, _last_metaspace_summary); 2501 } 2502 2503 void CMSCollector::collect_in_foreground(bool clear_all_soft_refs, GCCause::Cause cause) { 2504 assert(_foregroundGCIsActive && !_foregroundGCShouldWait, 2505 "Foreground collector should be waiting, not executing"); 2506 assert(Thread::current()->is_VM_thread(), "A foreground collection" 2507 "may only be done by the VM Thread with the world stopped"); 2508 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 2509 "VM thread should have CMS token"); 2510 2511 NOT_PRODUCT(GCTraceTime t("CMS:MS (foreground) ", PrintGCDetails && Verbose, 2512 true, NULL);) 2513 if (UseAdaptiveSizePolicy) { 2514 size_policy()->ms_collection_begin(); 2515 } 2516 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact); 2517 2518 HandleMark hm; // Discard invalid handles created during verification 2519 2520 if (VerifyBeforeGC && 2521 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2522 Universe::verify(); 2523 } 2524 2525 // Snapshot the soft reference policy to be used in this collection cycle. 2526 ref_processor()->setup_policy(clear_all_soft_refs); 2527 2528 // Decide if class unloading should be done 2529 update_should_unload_classes(); 2530 2531 bool init_mark_was_synchronous = false; // until proven otherwise 2532 while (_collectorState != Idling) { 2533 if (TraceCMSState) { 2534 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " in CMS state %d", 2535 Thread::current(), _collectorState); 2536 } 2537 switch (_collectorState) { 2538 case InitialMarking: 2539 register_foreground_gc_start(cause); 2540 init_mark_was_synchronous = true; // fact to be exploited in re-mark 2541 checkpointRootsInitial(false); 2542 assert(_collectorState == Marking, "Collector state should have changed" 2543 " within checkpointRootsInitial()"); 2544 break; 2545 case Marking: 2546 // initial marking in checkpointRootsInitialWork has been completed 2547 if (VerifyDuringGC && 2548 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2549 Universe::verify("Verify before initial mark: "); 2550 } 2551 { 2552 bool res = markFromRoots(false); 2553 assert(res && _collectorState == FinalMarking, "Collector state should " 2554 "have changed"); 2555 break; 2556 } 2557 case FinalMarking: 2558 if (VerifyDuringGC && 2559 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2560 Universe::verify("Verify before re-mark: "); 2561 } 2562 checkpointRootsFinal(false, clear_all_soft_refs, 2563 init_mark_was_synchronous); 2564 assert(_collectorState == Sweeping, "Collector state should not " 2565 "have changed within checkpointRootsFinal()"); 2566 break; 2567 case Sweeping: 2568 // final marking in checkpointRootsFinal has been completed 2569 if (VerifyDuringGC && 2570 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2571 Universe::verify("Verify before sweep: "); 2572 } 2573 sweep(false); 2574 assert(_collectorState == Resizing, "Incorrect state"); 2575 break; 2576 case Resizing: { 2577 // Sweeping has been completed; the actual resize in this case 2578 // is done separately; nothing to be done in this state. 2579 _collectorState = Resetting; 2580 break; 2581 } 2582 case Resetting: 2583 // The heap has been resized. 2584 if (VerifyDuringGC && 2585 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2586 Universe::verify("Verify before reset: "); 2587 } 2588 save_heap_summary(); 2589 reset(false); 2590 assert(_collectorState == Idling, "Collector state should " 2591 "have changed"); 2592 break; 2593 case Precleaning: 2594 case AbortablePreclean: 2595 // Elide the preclean phase 2596 _collectorState = FinalMarking; 2597 break; 2598 default: 2599 ShouldNotReachHere(); 2600 } 2601 if (TraceCMSState) { 2602 gclog_or_tty->print_cr(" Thread " INTPTR_FORMAT " done - next CMS state %d", 2603 Thread::current(), _collectorState); 2604 } 2605 } 2606 2607 if (UseAdaptiveSizePolicy) { 2608 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2609 size_policy()->ms_collection_end(gch->gc_cause()); 2610 } 2611 2612 if (VerifyAfterGC && 2613 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2614 Universe::verify(); 2615 } 2616 if (TraceCMSState) { 2617 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT 2618 " exiting collection CMS state %d", 2619 Thread::current(), _collectorState); 2620 } 2621 } 2622 2623 bool CMSCollector::waitForForegroundGC() { 2624 bool res = false; 2625 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 2626 "CMS thread should have CMS token"); 2627 // Block the foreground collector until the 2628 // background collectors decides whether to 2629 // yield. 2630 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 2631 _foregroundGCShouldWait = true; 2632 if (_foregroundGCIsActive) { 2633 // The background collector yields to the 2634 // foreground collector and returns a value 2635 // indicating that it has yielded. The foreground 2636 // collector can proceed. 2637 res = true; 2638 _foregroundGCShouldWait = false; 2639 ConcurrentMarkSweepThread::clear_CMS_flag( 2640 ConcurrentMarkSweepThread::CMS_cms_has_token); 2641 ConcurrentMarkSweepThread::set_CMS_flag( 2642 ConcurrentMarkSweepThread::CMS_cms_wants_token); 2643 // Get a possibly blocked foreground thread going 2644 CGC_lock->notify(); 2645 if (TraceCMSState) { 2646 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " waiting at CMS state %d", 2647 Thread::current(), _collectorState); 2648 } 2649 while (_foregroundGCIsActive) { 2650 CGC_lock->wait(Mutex::_no_safepoint_check_flag); 2651 } 2652 ConcurrentMarkSweepThread::set_CMS_flag( 2653 ConcurrentMarkSweepThread::CMS_cms_has_token); 2654 ConcurrentMarkSweepThread::clear_CMS_flag( 2655 ConcurrentMarkSweepThread::CMS_cms_wants_token); 2656 } 2657 if (TraceCMSState) { 2658 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " continuing at CMS state %d", 2659 Thread::current(), _collectorState); 2660 } 2661 return res; 2662 } 2663 2664 // Because of the need to lock the free lists and other structures in 2665 // the collector, common to all the generations that the collector is 2666 // collecting, we need the gc_prologues of individual CMS generations 2667 // delegate to their collector. It may have been simpler had the 2668 // current infrastructure allowed one to call a prologue on a 2669 // collector. In the absence of that we have the generation's 2670 // prologue delegate to the collector, which delegates back 2671 // some "local" work to a worker method in the individual generations 2672 // that it's responsible for collecting, while itself doing any 2673 // work common to all generations it's responsible for. A similar 2674 // comment applies to the gc_epilogue()'s. 2675 // The role of the variable _between_prologue_and_epilogue is to 2676 // enforce the invocation protocol. 2677 void CMSCollector::gc_prologue(bool full) { 2678 // Call gc_prologue_work() for the CMSGen 2679 // we are responsible for. 2680 2681 // The following locking discipline assumes that we are only called 2682 // when the world is stopped. 2683 assert(SafepointSynchronize::is_at_safepoint(), "world is stopped assumption"); 2684 2685 // The CMSCollector prologue must call the gc_prologues for the 2686 // "generations" that it's responsible 2687 // for. 2688 2689 assert( Thread::current()->is_VM_thread() 2690 || ( CMSScavengeBeforeRemark 2691 && Thread::current()->is_ConcurrentGC_thread()), 2692 "Incorrect thread type for prologue execution"); 2693 2694 if (_between_prologue_and_epilogue) { 2695 // We have already been invoked; this is a gc_prologue delegation 2696 // from yet another CMS generation that we are responsible for, just 2697 // ignore it since all relevant work has already been done. 2698 return; 2699 } 2700 2701 // set a bit saying prologue has been called; cleared in epilogue 2702 _between_prologue_and_epilogue = true; 2703 // Claim locks for common data structures, then call gc_prologue_work() 2704 // for each CMSGen. 2705 2706 getFreelistLocks(); // gets free list locks on constituent spaces 2707 bitMapLock()->lock_without_safepoint_check(); 2708 2709 // Should call gc_prologue_work() for all cms gens we are responsible for 2710 bool duringMarking = _collectorState >= Marking 2711 && _collectorState < Sweeping; 2712 2713 // The young collections clear the modified oops state, which tells if 2714 // there are any modified oops in the class. The remark phase also needs 2715 // that information. Tell the young collection to save the union of all 2716 // modified klasses. 2717 if (duringMarking) { 2718 _ct->klass_rem_set()->set_accumulate_modified_oops(true); 2719 } 2720 2721 bool registerClosure = duringMarking; 2722 2723 ModUnionClosure* muc = CollectedHeap::use_parallel_gc_threads() ? 2724 &_modUnionClosurePar 2725 : &_modUnionClosure; 2726 _cmsGen->gc_prologue_work(full, registerClosure, muc); 2727 2728 if (!full) { 2729 stats().record_gc0_begin(); 2730 } 2731 } 2732 2733 void ConcurrentMarkSweepGeneration::gc_prologue(bool full) { 2734 2735 _capacity_at_prologue = capacity(); 2736 _used_at_prologue = used(); 2737 2738 // Delegate to CMScollector which knows how to coordinate between 2739 // this and any other CMS generations that it is responsible for 2740 // collecting. 2741 collector()->gc_prologue(full); 2742 } 2743 2744 // This is a "private" interface for use by this generation's CMSCollector. 2745 // Not to be called directly by any other entity (for instance, 2746 // GenCollectedHeap, which calls the "public" gc_prologue method above). 2747 void ConcurrentMarkSweepGeneration::gc_prologue_work(bool full, 2748 bool registerClosure, ModUnionClosure* modUnionClosure) { 2749 assert(!incremental_collection_failed(), "Shouldn't be set yet"); 2750 assert(cmsSpace()->preconsumptionDirtyCardClosure() == NULL, 2751 "Should be NULL"); 2752 if (registerClosure) { 2753 cmsSpace()->setPreconsumptionDirtyCardClosure(modUnionClosure); 2754 } 2755 cmsSpace()->gc_prologue(); 2756 // Clear stat counters 2757 NOT_PRODUCT( 2758 assert(_numObjectsPromoted == 0, "check"); 2759 assert(_numWordsPromoted == 0, "check"); 2760 if (Verbose && PrintGC) { 2761 gclog_or_tty->print("Allocated "SIZE_FORMAT" objects, " 2762 SIZE_FORMAT" bytes concurrently", 2763 _numObjectsAllocated, _numWordsAllocated*sizeof(HeapWord)); 2764 } 2765 _numObjectsAllocated = 0; 2766 _numWordsAllocated = 0; 2767 ) 2768 } 2769 2770 void CMSCollector::gc_epilogue(bool full) { 2771 // The following locking discipline assumes that we are only called 2772 // when the world is stopped. 2773 assert(SafepointSynchronize::is_at_safepoint(), 2774 "world is stopped assumption"); 2775 2776 // Currently the CMS epilogue (see CompactibleFreeListSpace) merely checks 2777 // if linear allocation blocks need to be appropriately marked to allow the 2778 // the blocks to be parsable. We also check here whether we need to nudge the 2779 // CMS collector thread to start a new cycle (if it's not already active). 2780 assert( Thread::current()->is_VM_thread() 2781 || ( CMSScavengeBeforeRemark 2782 && Thread::current()->is_ConcurrentGC_thread()), 2783 "Incorrect thread type for epilogue execution"); 2784 2785 if (!_between_prologue_and_epilogue) { 2786 // We have already been invoked; this is a gc_epilogue delegation 2787 // from yet another CMS generation that we are responsible for, just 2788 // ignore it since all relevant work has already been done. 2789 return; 2790 } 2791 assert(haveFreelistLocks(), "must have freelist locks"); 2792 assert_lock_strong(bitMapLock()); 2793 2794 _ct->klass_rem_set()->set_accumulate_modified_oops(false); 2795 2796 _cmsGen->gc_epilogue_work(full); 2797 2798 if (_collectorState == AbortablePreclean || _collectorState == Precleaning) { 2799 // in case sampling was not already enabled, enable it 2800 _start_sampling = true; 2801 } 2802 // reset _eden_chunk_array so sampling starts afresh 2803 _eden_chunk_index = 0; 2804 2805 size_t cms_used = _cmsGen->cmsSpace()->used(); 2806 2807 // update performance counters - this uses a special version of 2808 // update_counters() that allows the utilization to be passed as a 2809 // parameter, avoiding multiple calls to used(). 2810 // 2811 _cmsGen->update_counters(cms_used); 2812 2813 if (CMSIncrementalMode) { 2814 icms_update_allocation_limits(); 2815 } 2816 2817 bitMapLock()->unlock(); 2818 releaseFreelistLocks(); 2819 2820 if (!CleanChunkPoolAsync) { 2821 Chunk::clean_chunk_pool(); 2822 } 2823 2824 set_did_compact(false); 2825 _between_prologue_and_epilogue = false; // ready for next cycle 2826 } 2827 2828 void ConcurrentMarkSweepGeneration::gc_epilogue(bool full) { 2829 collector()->gc_epilogue(full); 2830 2831 // Also reset promotion tracking in par gc thread states. 2832 if (CollectedHeap::use_parallel_gc_threads()) { 2833 for (uint i = 0; i < ParallelGCThreads; i++) { 2834 _par_gc_thread_states[i]->promo.stopTrackingPromotions(i); 2835 } 2836 } 2837 } 2838 2839 void ConcurrentMarkSweepGeneration::gc_epilogue_work(bool full) { 2840 assert(!incremental_collection_failed(), "Should have been cleared"); 2841 cmsSpace()->setPreconsumptionDirtyCardClosure(NULL); 2842 cmsSpace()->gc_epilogue(); 2843 // Print stat counters 2844 NOT_PRODUCT( 2845 assert(_numObjectsAllocated == 0, "check"); 2846 assert(_numWordsAllocated == 0, "check"); 2847 if (Verbose && PrintGC) { 2848 gclog_or_tty->print("Promoted "SIZE_FORMAT" objects, " 2849 SIZE_FORMAT" bytes", 2850 _numObjectsPromoted, _numWordsPromoted*sizeof(HeapWord)); 2851 } 2852 _numObjectsPromoted = 0; 2853 _numWordsPromoted = 0; 2854 ) 2855 2856 if (PrintGC && Verbose) { 2857 // Call down the chain in contiguous_available needs the freelistLock 2858 // so print this out before releasing the freeListLock. 2859 gclog_or_tty->print(" Contiguous available "SIZE_FORMAT" bytes ", 2860 contiguous_available()); 2861 } 2862 } 2863 2864 #ifndef PRODUCT 2865 bool CMSCollector::have_cms_token() { 2866 Thread* thr = Thread::current(); 2867 if (thr->is_VM_thread()) { 2868 return ConcurrentMarkSweepThread::vm_thread_has_cms_token(); 2869 } else if (thr->is_ConcurrentGC_thread()) { 2870 return ConcurrentMarkSweepThread::cms_thread_has_cms_token(); 2871 } else if (thr->is_GC_task_thread()) { 2872 return ConcurrentMarkSweepThread::vm_thread_has_cms_token() && 2873 ParGCRareEvent_lock->owned_by_self(); 2874 } 2875 return false; 2876 } 2877 #endif 2878 2879 // Check reachability of the given heap address in CMS generation, 2880 // treating all other generations as roots. 2881 bool CMSCollector::is_cms_reachable(HeapWord* addr) { 2882 // We could "guarantee" below, rather than assert, but I'll 2883 // leave these as "asserts" so that an adventurous debugger 2884 // could try this in the product build provided some subset of 2885 // the conditions were met, provided they were interested in the 2886 // results and knew that the computation below wouldn't interfere 2887 // with other concurrent computations mutating the structures 2888 // being read or written. 2889 assert(SafepointSynchronize::is_at_safepoint(), 2890 "Else mutations in object graph will make answer suspect"); 2891 assert(have_cms_token(), "Should hold cms token"); 2892 assert(haveFreelistLocks(), "must hold free list locks"); 2893 assert_lock_strong(bitMapLock()); 2894 2895 // Clear the marking bit map array before starting, but, just 2896 // for kicks, first report if the given address is already marked 2897 gclog_or_tty->print_cr("Start: Address " PTR_FORMAT " is%s marked", addr, 2898 _markBitMap.isMarked(addr) ? "" : " not"); 2899 2900 if (verify_after_remark()) { 2901 MutexLockerEx x(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag); 2902 bool result = verification_mark_bm()->isMarked(addr); 2903 gclog_or_tty->print_cr("TransitiveMark: Address " PTR_FORMAT " %s marked", addr, 2904 result ? "IS" : "is NOT"); 2905 return result; 2906 } else { 2907 gclog_or_tty->print_cr("Could not compute result"); 2908 return false; 2909 } 2910 } 2911 2912 2913 void 2914 CMSCollector::print_on_error(outputStream* st) { 2915 CMSCollector* collector = ConcurrentMarkSweepGeneration::_collector; 2916 if (collector != NULL) { 2917 CMSBitMap* bitmap = &collector->_markBitMap; 2918 st->print_cr("Marking Bits: (CMSBitMap*) " PTR_FORMAT, bitmap); 2919 bitmap->print_on_error(st, " Bits: "); 2920 2921 st->cr(); 2922 2923 CMSBitMap* mut_bitmap = &collector->_modUnionTable; 2924 st->print_cr("Mod Union Table: (CMSBitMap*) " PTR_FORMAT, mut_bitmap); 2925 mut_bitmap->print_on_error(st, " Bits: "); 2926 } 2927 } 2928 2929 //////////////////////////////////////////////////////// 2930 // CMS Verification Support 2931 //////////////////////////////////////////////////////// 2932 // Following the remark phase, the following invariant 2933 // should hold -- each object in the CMS heap which is 2934 // marked in markBitMap() should be marked in the verification_mark_bm(). 2935 2936 class VerifyMarkedClosure: public BitMapClosure { 2937 CMSBitMap* _marks; 2938 bool _failed; 2939 2940 public: 2941 VerifyMarkedClosure(CMSBitMap* bm): _marks(bm), _failed(false) {} 2942 2943 bool do_bit(size_t offset) { 2944 HeapWord* addr = _marks->offsetToHeapWord(offset); 2945 if (!_marks->isMarked(addr)) { 2946 oop(addr)->print_on(gclog_or_tty); 2947 gclog_or_tty->print_cr(" ("INTPTR_FORMAT" should have been marked)", addr); 2948 _failed = true; 2949 } 2950 return true; 2951 } 2952 2953 bool failed() { return _failed; } 2954 }; 2955 2956 bool CMSCollector::verify_after_remark(bool silent) { 2957 if (!silent) gclog_or_tty->print(" [Verifying CMS Marking... "); 2958 MutexLockerEx ml(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag); 2959 static bool init = false; 2960 2961 assert(SafepointSynchronize::is_at_safepoint(), 2962 "Else mutations in object graph will make answer suspect"); 2963 assert(have_cms_token(), 2964 "Else there may be mutual interference in use of " 2965 " verification data structures"); 2966 assert(_collectorState > Marking && _collectorState <= Sweeping, 2967 "Else marking info checked here may be obsolete"); 2968 assert(haveFreelistLocks(), "must hold free list locks"); 2969 assert_lock_strong(bitMapLock()); 2970 2971 2972 // Allocate marking bit map if not already allocated 2973 if (!init) { // first time 2974 if (!verification_mark_bm()->allocate(_span)) { 2975 return false; 2976 } 2977 init = true; 2978 } 2979 2980 assert(verification_mark_stack()->isEmpty(), "Should be empty"); 2981 2982 // Turn off refs discovery -- so we will be tracing through refs. 2983 // This is as intended, because by this time 2984 // GC must already have cleared any refs that need to be cleared, 2985 // and traced those that need to be marked; moreover, 2986 // the marking done here is not going to interfere in any 2987 // way with the marking information used by GC. 2988 NoRefDiscovery no_discovery(ref_processor()); 2989 2990 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;) 2991 2992 // Clear any marks from a previous round 2993 verification_mark_bm()->clear_all(); 2994 assert(verification_mark_stack()->isEmpty(), "markStack should be empty"); 2995 verify_work_stacks_empty(); 2996 2997 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2998 gch->ensure_parsability(false); // fill TLABs, but no need to retire them 2999 // Update the saved marks which may affect the root scans. 3000 gch->save_marks(); 3001 3002 if (CMSRemarkVerifyVariant == 1) { 3003 // In this first variant of verification, we complete 3004 // all marking, then check if the new marks-vector is 3005 // a subset of the CMS marks-vector. 3006 verify_after_remark_work_1(); 3007 } else if (CMSRemarkVerifyVariant == 2) { 3008 // In this second variant of verification, we flag an error 3009 // (i.e. an object reachable in the new marks-vector not reachable 3010 // in the CMS marks-vector) immediately, also indicating the 3011 // identify of an object (A) that references the unmarked object (B) -- 3012 // presumably, a mutation to A failed to be picked up by preclean/remark? 3013 verify_after_remark_work_2(); 3014 } else { 3015 warning("Unrecognized value %d for CMSRemarkVerifyVariant", 3016 CMSRemarkVerifyVariant); 3017 } 3018 if (!silent) gclog_or_tty->print(" done] "); 3019 return true; 3020 } 3021 3022 void CMSCollector::verify_after_remark_work_1() { 3023 ResourceMark rm; 3024 HandleMark hm; 3025 GenCollectedHeap* gch = GenCollectedHeap::heap(); 3026 3027 // Get a clear set of claim bits for the strong roots processing to work with. 3028 ClassLoaderDataGraph::clear_claimed_marks(); 3029 3030 // Mark from roots one level into CMS 3031 MarkRefsIntoClosure notOlder(_span, verification_mark_bm()); 3032 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 3033 3034 gch->gen_process_strong_roots(_cmsGen->level(), 3035 true, // younger gens are roots 3036 true, // activate StrongRootsScope 3037 SharedHeap::ScanningOption(roots_scanning_options()), 3038 ¬Older, 3039 NULL, 3040 NULL); // SSS: Provide correct closure 3041 3042 // Now mark from the roots 3043 MarkFromRootsClosure markFromRootsClosure(this, _span, 3044 verification_mark_bm(), verification_mark_stack(), 3045 false /* don't yield */, true /* verifying */); 3046 assert(_restart_addr == NULL, "Expected pre-condition"); 3047 verification_mark_bm()->iterate(&markFromRootsClosure); 3048 while (_restart_addr != NULL) { 3049 // Deal with stack overflow: by restarting at the indicated 3050 // address. 3051 HeapWord* ra = _restart_addr; 3052 markFromRootsClosure.reset(ra); 3053 _restart_addr = NULL; 3054 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end()); 3055 } 3056 assert(verification_mark_stack()->isEmpty(), "Should have been drained"); 3057 verify_work_stacks_empty(); 3058 3059 // Marking completed -- now verify that each bit marked in 3060 // verification_mark_bm() is also marked in markBitMap(); flag all 3061 // errors by printing corresponding objects. 3062 VerifyMarkedClosure vcl(markBitMap()); 3063 verification_mark_bm()->iterate(&vcl); 3064 if (vcl.failed()) { 3065 gclog_or_tty->print("Verification failed"); 3066 Universe::heap()->print_on(gclog_or_tty); 3067 fatal("CMS: failed marking verification after remark"); 3068 } 3069 } 3070 3071 class VerifyKlassOopsKlassClosure : public KlassClosure { 3072 class VerifyKlassOopsClosure : public OopClosure { 3073 CMSBitMap* _bitmap; 3074 public: 3075 VerifyKlassOopsClosure(CMSBitMap* bitmap) : _bitmap(bitmap) { } 3076 void do_oop(oop* p) { guarantee(*p == NULL || _bitmap->isMarked((HeapWord*) *p), "Should be marked"); } 3077 void do_oop(narrowOop* p) { ShouldNotReachHere(); } 3078 } _oop_closure; 3079 public: 3080 VerifyKlassOopsKlassClosure(CMSBitMap* bitmap) : _oop_closure(bitmap) {} 3081 void do_klass(Klass* k) { 3082 k->oops_do(&_oop_closure); 3083 } 3084 }; 3085 3086 void CMSCollector::verify_after_remark_work_2() { 3087 ResourceMark rm; 3088 HandleMark hm; 3089 GenCollectedHeap* gch = GenCollectedHeap::heap(); 3090 3091 // Get a clear set of claim bits for the strong roots processing to work with. 3092 ClassLoaderDataGraph::clear_claimed_marks(); 3093 3094 // Mark from roots one level into CMS 3095 MarkRefsIntoVerifyClosure notOlder(_span, verification_mark_bm(), 3096 markBitMap()); 3097 CMKlassClosure klass_closure(¬Older); 3098 3099 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 3100 gch->gen_process_strong_roots(_cmsGen->level(), 3101 true, // younger gens are roots 3102 true, // activate StrongRootsScope 3103 SharedHeap::ScanningOption(roots_scanning_options()), 3104 ¬Older, 3105 NULL, 3106 &klass_closure); 3107 3108 // Now mark from the roots 3109 MarkFromRootsVerifyClosure markFromRootsClosure(this, _span, 3110 verification_mark_bm(), markBitMap(), verification_mark_stack()); 3111 assert(_restart_addr == NULL, "Expected pre-condition"); 3112 verification_mark_bm()->iterate(&markFromRootsClosure); 3113 while (_restart_addr != NULL) { 3114 // Deal with stack overflow: by restarting at the indicated 3115 // address. 3116 HeapWord* ra = _restart_addr; 3117 markFromRootsClosure.reset(ra); 3118 _restart_addr = NULL; 3119 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end()); 3120 } 3121 assert(verification_mark_stack()->isEmpty(), "Should have been drained"); 3122 verify_work_stacks_empty(); 3123 3124 VerifyKlassOopsKlassClosure verify_klass_oops(verification_mark_bm()); 3125 ClassLoaderDataGraph::classes_do(&verify_klass_oops); 3126 3127 // Marking completed -- now verify that each bit marked in 3128 // verification_mark_bm() is also marked in markBitMap(); flag all 3129 // errors by printing corresponding objects. 3130 VerifyMarkedClosure vcl(markBitMap()); 3131 verification_mark_bm()->iterate(&vcl); 3132 assert(!vcl.failed(), "Else verification above should not have succeeded"); 3133 } 3134 3135 void ConcurrentMarkSweepGeneration::save_marks() { 3136 // delegate to CMS space 3137 cmsSpace()->save_marks(); 3138 for (uint i = 0; i < ParallelGCThreads; i++) { 3139 _par_gc_thread_states[i]->promo.startTrackingPromotions(); 3140 } 3141 } 3142 3143 bool ConcurrentMarkSweepGeneration::no_allocs_since_save_marks() { 3144 return cmsSpace()->no_allocs_since_save_marks(); 3145 } 3146 3147 #define CMS_SINCE_SAVE_MARKS_DEFN(OopClosureType, nv_suffix) \ 3148 \ 3149 void ConcurrentMarkSweepGeneration:: \ 3150 oop_since_save_marks_iterate##nv_suffix(OopClosureType* cl) { \ 3151 cl->set_generation(this); \ 3152 cmsSpace()->oop_since_save_marks_iterate##nv_suffix(cl); \ 3153 cl->reset_generation(); \ 3154 save_marks(); \ 3155 } 3156 3157 ALL_SINCE_SAVE_MARKS_CLOSURES(CMS_SINCE_SAVE_MARKS_DEFN) 3158 3159 void 3160 ConcurrentMarkSweepGeneration::younger_refs_iterate(OopsInGenClosure* cl) { 3161 cl->set_generation(this); 3162 younger_refs_in_space_iterate(_cmsSpace, cl); 3163 cl->reset_generation(); 3164 } 3165 3166 void 3167 ConcurrentMarkSweepGeneration::oop_iterate(ExtendedOopClosure* cl) { 3168 if (freelistLock()->owned_by_self()) { 3169 Generation::oop_iterate(cl); 3170 } else { 3171 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 3172 Generation::oop_iterate(cl); 3173 } 3174 } 3175 3176 void 3177 ConcurrentMarkSweepGeneration::object_iterate(ObjectClosure* cl) { 3178 if (freelistLock()->owned_by_self()) { 3179 Generation::object_iterate(cl); 3180 } else { 3181 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 3182 Generation::object_iterate(cl); 3183 } 3184 } 3185 3186 void 3187 ConcurrentMarkSweepGeneration::safe_object_iterate(ObjectClosure* cl) { 3188 if (freelistLock()->owned_by_self()) { 3189 Generation::safe_object_iterate(cl); 3190 } else { 3191 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 3192 Generation::safe_object_iterate(cl); 3193 } 3194 } 3195 3196 void 3197 ConcurrentMarkSweepGeneration::post_compact() { 3198 } 3199 3200 void 3201 ConcurrentMarkSweepGeneration::prepare_for_verify() { 3202 // Fix the linear allocation blocks to look like free blocks. 3203 3204 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those 3205 // are not called when the heap is verified during universe initialization and 3206 // at vm shutdown. 3207 if (freelistLock()->owned_by_self()) { 3208 cmsSpace()->prepare_for_verify(); 3209 } else { 3210 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag); 3211 cmsSpace()->prepare_for_verify(); 3212 } 3213 } 3214 3215 void 3216 ConcurrentMarkSweepGeneration::verify() { 3217 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those 3218 // are not called when the heap is verified during universe initialization and 3219 // at vm shutdown. 3220 if (freelistLock()->owned_by_self()) { 3221 cmsSpace()->verify(); 3222 } else { 3223 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag); 3224 cmsSpace()->verify(); 3225 } 3226 } 3227 3228 void CMSCollector::verify() { 3229 _cmsGen->verify(); 3230 } 3231 3232 #ifndef PRODUCT 3233 bool CMSCollector::overflow_list_is_empty() const { 3234 assert(_num_par_pushes >= 0, "Inconsistency"); 3235 if (_overflow_list == NULL) { 3236 assert(_num_par_pushes == 0, "Inconsistency"); 3237 } 3238 return _overflow_list == NULL; 3239 } 3240 3241 // The methods verify_work_stacks_empty() and verify_overflow_empty() 3242 // merely consolidate assertion checks that appear to occur together frequently. 3243 void CMSCollector::verify_work_stacks_empty() const { 3244 assert(_markStack.isEmpty(), "Marking stack should be empty"); 3245 assert(overflow_list_is_empty(), "Overflow list should be empty"); 3246 } 3247 3248 void CMSCollector::verify_overflow_empty() const { 3249 assert(overflow_list_is_empty(), "Overflow list should be empty"); 3250 assert(no_preserved_marks(), "No preserved marks"); 3251 } 3252 #endif // PRODUCT 3253 3254 // Decide if we want to enable class unloading as part of the 3255 // ensuing concurrent GC cycle. We will collect and 3256 // unload classes if it's the case that: 3257 // (1) an explicit gc request has been made and the flag 3258 // ExplicitGCInvokesConcurrentAndUnloadsClasses is set, OR 3259 // (2) (a) class unloading is enabled at the command line, and 3260 // (b) old gen is getting really full 3261 // NOTE: Provided there is no change in the state of the heap between 3262 // calls to this method, it should have idempotent results. Moreover, 3263 // its results should be monotonically increasing (i.e. going from 0 to 1, 3264 // but not 1 to 0) between successive calls between which the heap was 3265 // not collected. For the implementation below, it must thus rely on 3266 // the property that concurrent_cycles_since_last_unload() 3267 // will not decrease unless a collection cycle happened and that 3268 // _cmsGen->is_too_full() are 3269 // themselves also monotonic in that sense. See check_monotonicity() 3270 // below. 3271 void CMSCollector::update_should_unload_classes() { 3272 _should_unload_classes = false; 3273 // Condition 1 above 3274 if (_full_gc_requested && ExplicitGCInvokesConcurrentAndUnloadsClasses) { 3275 _should_unload_classes = true; 3276 } else if (CMSClassUnloadingEnabled) { // Condition 2.a above 3277 // Disjuncts 2.b.(i,ii,iii) above 3278 _should_unload_classes = (concurrent_cycles_since_last_unload() >= 3279 CMSClassUnloadingMaxInterval) 3280 || _cmsGen->is_too_full(); 3281 } 3282 } 3283 3284 bool ConcurrentMarkSweepGeneration::is_too_full() const { 3285 bool res = should_concurrent_collect(); 3286 res = res && (occupancy() > (double)CMSIsTooFullPercentage/100.0); 3287 return res; 3288 } 3289 3290 void CMSCollector::setup_cms_unloading_and_verification_state() { 3291 const bool should_verify = VerifyBeforeGC || VerifyAfterGC || VerifyDuringGC 3292 || VerifyBeforeExit; 3293 const int rso = SharedHeap::SO_Strings | SharedHeap::SO_AllCodeCache; 3294 3295 // We set the proper root for this CMS cycle here. 3296 if (should_unload_classes()) { // Should unload classes this cycle 3297 remove_root_scanning_option(SharedHeap::SO_AllClasses); 3298 add_root_scanning_option(SharedHeap::SO_SystemClasses); 3299 remove_root_scanning_option(rso); // Shrink the root set appropriately 3300 set_verifying(should_verify); // Set verification state for this cycle 3301 return; // Nothing else needs to be done at this time 3302 } 3303 3304 // Not unloading classes this cycle 3305 assert(!should_unload_classes(), "Inconsistency!"); 3306 remove_root_scanning_option(SharedHeap::SO_SystemClasses); 3307 add_root_scanning_option(SharedHeap::SO_AllClasses); 3308 3309 if ((!verifying() || unloaded_classes_last_cycle()) && should_verify) { 3310 // Include symbols, strings and code cache elements to prevent their resurrection. 3311 add_root_scanning_option(rso); 3312 set_verifying(true); 3313 } else if (verifying() && !should_verify) { 3314 // We were verifying, but some verification flags got disabled. 3315 set_verifying(false); 3316 // Exclude symbols, strings and code cache elements from root scanning to 3317 // reduce IM and RM pauses. 3318 remove_root_scanning_option(rso); 3319 } 3320 } 3321 3322 3323 #ifndef PRODUCT 3324 HeapWord* CMSCollector::block_start(const void* p) const { 3325 const HeapWord* addr = (HeapWord*)p; 3326 if (_span.contains(p)) { 3327 if (_cmsGen->cmsSpace()->is_in_reserved(addr)) { 3328 return _cmsGen->cmsSpace()->block_start(p); 3329 } 3330 } 3331 return NULL; 3332 } 3333 #endif 3334 3335 HeapWord* 3336 ConcurrentMarkSweepGeneration::expand_and_allocate(size_t word_size, 3337 bool tlab, 3338 bool parallel) { 3339 CMSSynchronousYieldRequest yr; 3340 assert(!tlab, "Can't deal with TLAB allocation"); 3341 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 3342 expand(word_size*HeapWordSize, MinHeapDeltaBytes, 3343 CMSExpansionCause::_satisfy_allocation); 3344 if (GCExpandToAllocateDelayMillis > 0) { 3345 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 3346 } 3347 return have_lock_and_allocate(word_size, tlab); 3348 } 3349 3350 // YSR: All of this generation expansion/shrinking stuff is an exact copy of 3351 // OneContigSpaceCardGeneration, which makes me wonder if we should move this 3352 // to CardGeneration and share it... 3353 bool ConcurrentMarkSweepGeneration::expand(size_t bytes, size_t expand_bytes) { 3354 return CardGeneration::expand(bytes, expand_bytes); 3355 } 3356 3357 void ConcurrentMarkSweepGeneration::expand(size_t bytes, size_t expand_bytes, 3358 CMSExpansionCause::Cause cause) 3359 { 3360 3361 bool success = expand(bytes, expand_bytes); 3362 3363 // remember why we expanded; this information is used 3364 // by shouldConcurrentCollect() when making decisions on whether to start 3365 // a new CMS cycle. 3366 if (success) { 3367 set_expansion_cause(cause); 3368 if (PrintGCDetails && Verbose) { 3369 gclog_or_tty->print_cr("Expanded CMS gen for %s", 3370 CMSExpansionCause::to_string(cause)); 3371 } 3372 } 3373 } 3374 3375 HeapWord* ConcurrentMarkSweepGeneration::expand_and_par_lab_allocate(CMSParGCThreadState* ps, size_t word_sz) { 3376 HeapWord* res = NULL; 3377 MutexLocker x(ParGCRareEvent_lock); 3378 while (true) { 3379 // Expansion by some other thread might make alloc OK now: 3380 res = ps->lab.alloc(word_sz); 3381 if (res != NULL) return res; 3382 // If there's not enough expansion space available, give up. 3383 if (_virtual_space.uncommitted_size() < (word_sz * HeapWordSize)) { 3384 return NULL; 3385 } 3386 // Otherwise, we try expansion. 3387 expand(word_sz*HeapWordSize, MinHeapDeltaBytes, 3388 CMSExpansionCause::_allocate_par_lab); 3389 // Now go around the loop and try alloc again; 3390 // A competing par_promote might beat us to the expansion space, 3391 // so we may go around the loop again if promotion fails again. 3392 if (GCExpandToAllocateDelayMillis > 0) { 3393 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 3394 } 3395 } 3396 } 3397 3398 3399 bool ConcurrentMarkSweepGeneration::expand_and_ensure_spooling_space( 3400 PromotionInfo* promo) { 3401 MutexLocker x(ParGCRareEvent_lock); 3402 size_t refill_size_bytes = promo->refillSize() * HeapWordSize; 3403 while (true) { 3404 // Expansion by some other thread might make alloc OK now: 3405 if (promo->ensure_spooling_space()) { 3406 assert(promo->has_spooling_space(), 3407 "Post-condition of successful ensure_spooling_space()"); 3408 return true; 3409 } 3410 // If there's not enough expansion space available, give up. 3411 if (_virtual_space.uncommitted_size() < refill_size_bytes) { 3412 return false; 3413 } 3414 // Otherwise, we try expansion. 3415 expand(refill_size_bytes, MinHeapDeltaBytes, 3416 CMSExpansionCause::_allocate_par_spooling_space); 3417 // Now go around the loop and try alloc again; 3418 // A competing allocation might beat us to the expansion space, 3419 // so we may go around the loop again if allocation fails again. 3420 if (GCExpandToAllocateDelayMillis > 0) { 3421 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 3422 } 3423 } 3424 } 3425 3426 3427 void ConcurrentMarkSweepGeneration::shrink_by(size_t bytes) { 3428 assert_locked_or_safepoint(ExpandHeap_lock); 3429 // Shrink committed space 3430 _virtual_space.shrink_by(bytes); 3431 // Shrink space; this also shrinks the space's BOT 3432 _cmsSpace->set_end((HeapWord*) _virtual_space.high()); 3433 size_t new_word_size = heap_word_size(_cmsSpace->capacity()); 3434 // Shrink the shared block offset array 3435 _bts->resize(new_word_size); 3436 MemRegion mr(_cmsSpace->bottom(), new_word_size); 3437 // Shrink the card table 3438 Universe::heap()->barrier_set()->resize_covered_region(mr); 3439 3440 if (Verbose && PrintGC) { 3441 size_t new_mem_size = _virtual_space.committed_size(); 3442 size_t old_mem_size = new_mem_size + bytes; 3443 gclog_or_tty->print_cr("Shrinking %s from " SIZE_FORMAT "K to " SIZE_FORMAT "K", 3444 name(), old_mem_size/K, new_mem_size/K); 3445 } 3446 } 3447 3448 void ConcurrentMarkSweepGeneration::shrink(size_t bytes) { 3449 assert_locked_or_safepoint(Heap_lock); 3450 size_t size = ReservedSpace::page_align_size_down(bytes); 3451 // Only shrink if a compaction was done so that all the free space 3452 // in the generation is in a contiguous block at the end. 3453 if (size > 0 && did_compact()) { 3454 shrink_by(size); 3455 } 3456 } 3457 3458 bool ConcurrentMarkSweepGeneration::grow_by(size_t bytes) { 3459 assert_locked_or_safepoint(Heap_lock); 3460 bool result = _virtual_space.expand_by(bytes); 3461 if (result) { 3462 size_t new_word_size = 3463 heap_word_size(_virtual_space.committed_size()); 3464 MemRegion mr(_cmsSpace->bottom(), new_word_size); 3465 _bts->resize(new_word_size); // resize the block offset shared array 3466 Universe::heap()->barrier_set()->resize_covered_region(mr); 3467 // Hmmmm... why doesn't CFLS::set_end verify locking? 3468 // This is quite ugly; FIX ME XXX 3469 _cmsSpace->assert_locked(freelistLock()); 3470 _cmsSpace->set_end((HeapWord*)_virtual_space.high()); 3471 3472 // update the space and generation capacity counters 3473 if (UsePerfData) { 3474 _space_counters->update_capacity(); 3475 _gen_counters->update_all(); 3476 } 3477 3478 if (Verbose && PrintGC) { 3479 size_t new_mem_size = _virtual_space.committed_size(); 3480 size_t old_mem_size = new_mem_size - bytes; 3481 gclog_or_tty->print_cr("Expanding %s from " SIZE_FORMAT "K by " SIZE_FORMAT "K to " SIZE_FORMAT "K", 3482 name(), old_mem_size/K, bytes/K, new_mem_size/K); 3483 } 3484 } 3485 return result; 3486 } 3487 3488 bool ConcurrentMarkSweepGeneration::grow_to_reserved() { 3489 assert_locked_or_safepoint(Heap_lock); 3490 bool success = true; 3491 const size_t remaining_bytes = _virtual_space.uncommitted_size(); 3492 if (remaining_bytes > 0) { 3493 success = grow_by(remaining_bytes); 3494 DEBUG_ONLY(if (!success) warning("grow to reserved failed");) 3495 } 3496 return success; 3497 } 3498 3499 void ConcurrentMarkSweepGeneration::shrink_free_list_by(size_t bytes) { 3500 assert_locked_or_safepoint(Heap_lock); 3501 assert_lock_strong(freelistLock()); 3502 if (PrintGCDetails && Verbose) { 3503 warning("Shrinking of CMS not yet implemented"); 3504 } 3505 return; 3506 } 3507 3508 3509 // Simple ctor/dtor wrapper for accounting & timer chores around concurrent 3510 // phases. 3511 class CMSPhaseAccounting: public StackObj { 3512 public: 3513 CMSPhaseAccounting(CMSCollector *collector, 3514 const char *phase, 3515 bool print_cr = true); 3516 ~CMSPhaseAccounting(); 3517 3518 private: 3519 CMSCollector *_collector; 3520 const char *_phase; 3521 elapsedTimer _wallclock; 3522 bool _print_cr; 3523 3524 public: 3525 // Not MT-safe; so do not pass around these StackObj's 3526 // where they may be accessed by other threads. 3527 jlong wallclock_millis() { 3528 assert(_wallclock.is_active(), "Wall clock should not stop"); 3529 _wallclock.stop(); // to record time 3530 jlong ret = _wallclock.milliseconds(); 3531 _wallclock.start(); // restart 3532 return ret; 3533 } 3534 }; 3535 3536 CMSPhaseAccounting::CMSPhaseAccounting(CMSCollector *collector, 3537 const char *phase, 3538 bool print_cr) : 3539 _collector(collector), _phase(phase), _print_cr(print_cr) { 3540 3541 if (PrintCMSStatistics != 0) { 3542 _collector->resetYields(); 3543 } 3544 if (PrintGCDetails) { 3545 gclog_or_tty->date_stamp(PrintGCDateStamps); 3546 gclog_or_tty->stamp(PrintGCTimeStamps); 3547 gclog_or_tty->print_cr("[%s-concurrent-%s-start]", 3548 _collector->cmsGen()->short_name(), _phase); 3549 } 3550 _collector->resetTimer(); 3551 _wallclock.start(); 3552 _collector->startTimer(); 3553 } 3554 3555 CMSPhaseAccounting::~CMSPhaseAccounting() { 3556 assert(_wallclock.is_active(), "Wall clock should not have stopped"); 3557 _collector->stopTimer(); 3558 _wallclock.stop(); 3559 if (PrintGCDetails) { 3560 gclog_or_tty->date_stamp(PrintGCDateStamps); 3561 gclog_or_tty->stamp(PrintGCTimeStamps); 3562 gclog_or_tty->print("[%s-concurrent-%s: %3.3f/%3.3f secs]", 3563 _collector->cmsGen()->short_name(), 3564 _phase, _collector->timerValue(), _wallclock.seconds()); 3565 if (_print_cr) { 3566 gclog_or_tty->print_cr(""); 3567 } 3568 if (PrintCMSStatistics != 0) { 3569 gclog_or_tty->print_cr(" (CMS-concurrent-%s yielded %d times)", _phase, 3570 _collector->yields()); 3571 } 3572 } 3573 } 3574 3575 // CMS work 3576 3577 // The common parts of CMSParInitialMarkTask and CMSParRemarkTask. 3578 class CMSParMarkTask : public AbstractGangTask { 3579 protected: 3580 CMSCollector* _collector; 3581 int _n_workers; 3582 CMSParMarkTask(const char* name, CMSCollector* collector, int n_workers) : 3583 AbstractGangTask(name), 3584 _collector(collector), 3585 _n_workers(n_workers) {} 3586 // Work method in support of parallel rescan ... of young gen spaces 3587 void do_young_space_rescan(uint worker_id, OopsInGenClosure* cl, 3588 ContiguousSpace* space, 3589 HeapWord** chunk_array, size_t chunk_top); 3590 void work_on_young_gen_roots(uint worker_id, OopsInGenClosure* cl); 3591 }; 3592 3593 // Parallel initial mark task 3594 class CMSParInitialMarkTask: public CMSParMarkTask { 3595 public: 3596 CMSParInitialMarkTask(CMSCollector* collector, int n_workers) : 3597 CMSParMarkTask("Scan roots and young gen for initial mark in parallel", 3598 collector, n_workers) {} 3599 void work(uint worker_id); 3600 }; 3601 3602 // Checkpoint the roots into this generation from outside 3603 // this generation. [Note this initial checkpoint need only 3604 // be approximate -- we'll do a catch up phase subsequently.] 3605 void CMSCollector::checkpointRootsInitial(bool asynch) { 3606 assert(_collectorState == InitialMarking, "Wrong collector state"); 3607 check_correct_thread_executing(); 3608 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 3609 3610 save_heap_summary(); 3611 report_heap_summary(GCWhen::BeforeGC); 3612 3613 ReferenceProcessor* rp = ref_processor(); 3614 SpecializationStats::clear(); 3615 assert(_restart_addr == NULL, "Control point invariant"); 3616 if (asynch) { 3617 // acquire locks for subsequent manipulations 3618 MutexLockerEx x(bitMapLock(), 3619 Mutex::_no_safepoint_check_flag); 3620 checkpointRootsInitialWork(asynch); 3621 // enable ("weak") refs discovery 3622 rp->enable_discovery(true /*verify_disabled*/, true /*check_no_refs*/); 3623 _collectorState = Marking; 3624 } else { 3625 // (Weak) Refs discovery: this is controlled from genCollectedHeap::do_collection 3626 // which recognizes if we are a CMS generation, and doesn't try to turn on 3627 // discovery; verify that they aren't meddling. 3628 assert(!rp->discovery_is_atomic(), 3629 "incorrect setting of discovery predicate"); 3630 assert(!rp->discovery_enabled(), "genCollectedHeap shouldn't control " 3631 "ref discovery for this generation kind"); 3632 // already have locks 3633 checkpointRootsInitialWork(asynch); 3634 // now enable ("weak") refs discovery 3635 rp->enable_discovery(true /*verify_disabled*/, false /*verify_no_refs*/); 3636 _collectorState = Marking; 3637 } 3638 SpecializationStats::print(); 3639 } 3640 3641 void CMSCollector::checkpointRootsInitialWork(bool asynch) { 3642 assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped"); 3643 assert(_collectorState == InitialMarking, "just checking"); 3644 3645 // If there has not been a GC[n-1] since last GC[n] cycle completed, 3646 // precede our marking with a collection of all 3647 // younger generations to keep floating garbage to a minimum. 3648 // XXX: we won't do this for now -- it's an optimization to be done later. 3649 3650 // already have locks 3651 assert_lock_strong(bitMapLock()); 3652 assert(_markBitMap.isAllClear(), "was reset at end of previous cycle"); 3653 3654 // Setup the verification and class unloading state for this 3655 // CMS collection cycle. 3656 setup_cms_unloading_and_verification_state(); 3657 3658 NOT_PRODUCT(GCTraceTime t("\ncheckpointRootsInitialWork", 3659 PrintGCDetails && Verbose, true, _gc_timer_cm);) 3660 if (UseAdaptiveSizePolicy) { 3661 size_policy()->checkpoint_roots_initial_begin(); 3662 } 3663 3664 // Reset all the PLAB chunk arrays if necessary. 3665 if (_survivor_plab_array != NULL && !CMSPLABRecordAlways) { 3666 reset_survivor_plab_arrays(); 3667 } 3668 3669 ResourceMark rm; 3670 HandleMark hm; 3671 3672 MarkRefsIntoClosure notOlder(_span, &_markBitMap); 3673 GenCollectedHeap* gch = GenCollectedHeap::heap(); 3674 3675 verify_work_stacks_empty(); 3676 verify_overflow_empty(); 3677 3678 gch->ensure_parsability(false); // fill TLABs, but no need to retire them 3679 // Update the saved marks which may affect the root scans. 3680 gch->save_marks(); 3681 3682 // weak reference processing has not started yet. 3683 ref_processor()->set_enqueuing_is_done(false); 3684 3685 // Need to remember all newly created CLDs, 3686 // so that we can guarantee that the remark finds them. 3687 ClassLoaderDataGraph::remember_new_clds(true); 3688 3689 // Whenever a CLD is found, it will be claimed before proceeding to mark 3690 // the klasses. The claimed marks need to be cleared before marking starts. 3691 ClassLoaderDataGraph::clear_claimed_marks(); 3692 3693 if (CMSPrintEdenSurvivorChunks) { 3694 print_eden_and_survivor_chunk_arrays(); 3695 } 3696 3697 { 3698 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;) 3699 if (CMSParallelInitialMarkEnabled && CollectedHeap::use_parallel_gc_threads()) { 3700 // The parallel version. 3701 FlexibleWorkGang* workers = gch->workers(); 3702 assert(workers != NULL, "Need parallel worker threads."); 3703 int n_workers = workers->active_workers(); 3704 CMSParInitialMarkTask tsk(this, n_workers); 3705 gch->set_par_threads(n_workers); 3706 initialize_sequential_subtasks_for_young_gen_rescan(n_workers); 3707 if (n_workers > 1) { 3708 GenCollectedHeap::StrongRootsScope srs(gch); 3709 workers->run_task(&tsk); 3710 } else { 3711 GenCollectedHeap::StrongRootsScope srs(gch); 3712 tsk.work(0); 3713 } 3714 gch->set_par_threads(0); 3715 } else { 3716 // The serial version. 3717 CMKlassClosure klass_closure(¬Older); 3718 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 3719 gch->gen_process_strong_roots(_cmsGen->level(), 3720 true, // younger gens are roots 3721 true, // activate StrongRootsScope 3722 SharedHeap::ScanningOption(roots_scanning_options()), 3723 ¬Older, 3724 NULL, 3725 &klass_closure); 3726 } 3727 } 3728 3729 // Clear mod-union table; it will be dirtied in the prologue of 3730 // CMS generation per each younger generation collection. 3731 3732 assert(_modUnionTable.isAllClear(), 3733 "Was cleared in most recent final checkpoint phase" 3734 " or no bits are set in the gc_prologue before the start of the next " 3735 "subsequent marking phase."); 3736 3737 assert(_ct->klass_rem_set()->mod_union_is_clear(), "Must be"); 3738 3739 // Save the end of the used_region of the constituent generations 3740 // to be used to limit the extent of sweep in each generation. 3741 save_sweep_limits(); 3742 if (UseAdaptiveSizePolicy) { 3743 size_policy()->checkpoint_roots_initial_end(gch->gc_cause()); 3744 } 3745 verify_overflow_empty(); 3746 } 3747 3748 bool CMSCollector::markFromRoots(bool asynch) { 3749 // we might be tempted to assert that: 3750 // assert(asynch == !SafepointSynchronize::is_at_safepoint(), 3751 // "inconsistent argument?"); 3752 // However that wouldn't be right, because it's possible that 3753 // a safepoint is indeed in progress as a younger generation 3754 // stop-the-world GC happens even as we mark in this generation. 3755 assert(_collectorState == Marking, "inconsistent state?"); 3756 check_correct_thread_executing(); 3757 verify_overflow_empty(); 3758 3759 bool res; 3760 if (asynch) { 3761 3762 // Start the timers for adaptive size policy for the concurrent phases 3763 // Do it here so that the foreground MS can use the concurrent 3764 // timer since a foreground MS might has the sweep done concurrently 3765 // or STW. 3766 if (UseAdaptiveSizePolicy) { 3767 size_policy()->concurrent_marking_begin(); 3768 } 3769 3770 // Weak ref discovery note: We may be discovering weak 3771 // refs in this generation concurrent (but interleaved) with 3772 // weak ref discovery by a younger generation collector. 3773 3774 CMSTokenSyncWithLocks ts(true, bitMapLock()); 3775 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 3776 CMSPhaseAccounting pa(this, "mark", !PrintGCDetails); 3777 res = markFromRootsWork(asynch); 3778 if (res) { 3779 _collectorState = Precleaning; 3780 } else { // We failed and a foreground collection wants to take over 3781 assert(_foregroundGCIsActive, "internal state inconsistency"); 3782 assert(_restart_addr == NULL, "foreground will restart from scratch"); 3783 if (PrintGCDetails) { 3784 gclog_or_tty->print_cr("bailing out to foreground collection"); 3785 } 3786 } 3787 if (UseAdaptiveSizePolicy) { 3788 size_policy()->concurrent_marking_end(); 3789 } 3790 } else { 3791 assert(SafepointSynchronize::is_at_safepoint(), 3792 "inconsistent with asynch == false"); 3793 if (UseAdaptiveSizePolicy) { 3794 size_policy()->ms_collection_marking_begin(); 3795 } 3796 // already have locks 3797 res = markFromRootsWork(asynch); 3798 _collectorState = FinalMarking; 3799 if (UseAdaptiveSizePolicy) { 3800 GenCollectedHeap* gch = GenCollectedHeap::heap(); 3801 size_policy()->ms_collection_marking_end(gch->gc_cause()); 3802 } 3803 } 3804 verify_overflow_empty(); 3805 return res; 3806 } 3807 3808 bool CMSCollector::markFromRootsWork(bool asynch) { 3809 // iterate over marked bits in bit map, doing a full scan and mark 3810 // from these roots using the following algorithm: 3811 // . if oop is to the right of the current scan pointer, 3812 // mark corresponding bit (we'll process it later) 3813 // . else (oop is to left of current scan pointer) 3814 // push oop on marking stack 3815 // . drain the marking stack 3816 3817 // Note that when we do a marking step we need to hold the 3818 // bit map lock -- recall that direct allocation (by mutators) 3819 // and promotion (by younger generation collectors) is also 3820 // marking the bit map. [the so-called allocate live policy.] 3821 // Because the implementation of bit map marking is not 3822 // robust wrt simultaneous marking of bits in the same word, 3823 // we need to make sure that there is no such interference 3824 // between concurrent such updates. 3825 3826 // already have locks 3827 assert_lock_strong(bitMapLock()); 3828 3829 verify_work_stacks_empty(); 3830 verify_overflow_empty(); 3831 bool result = false; 3832 if (CMSConcurrentMTEnabled && ConcGCThreads > 0) { 3833 result = do_marking_mt(asynch); 3834 } else { 3835 result = do_marking_st(asynch); 3836 } 3837 return result; 3838 } 3839 3840 // Forward decl 3841 class CMSConcMarkingTask; 3842 3843 class CMSConcMarkingTerminator: public ParallelTaskTerminator { 3844 CMSCollector* _collector; 3845 CMSConcMarkingTask* _task; 3846 public: 3847 virtual void yield(); 3848 3849 // "n_threads" is the number of threads to be terminated. 3850 // "queue_set" is a set of work queues of other threads. 3851 // "collector" is the CMS collector associated with this task terminator. 3852 // "yield" indicates whether we need the gang as a whole to yield. 3853 CMSConcMarkingTerminator(int n_threads, TaskQueueSetSuper* queue_set, CMSCollector* collector) : 3854 ParallelTaskTerminator(n_threads, queue_set), 3855 _collector(collector) { } 3856 3857 void set_task(CMSConcMarkingTask* task) { 3858 _task = task; 3859 } 3860 }; 3861 3862 class CMSConcMarkingTerminatorTerminator: public TerminatorTerminator { 3863 CMSConcMarkingTask* _task; 3864 public: 3865 bool should_exit_termination(); 3866 void set_task(CMSConcMarkingTask* task) { 3867 _task = task; 3868 } 3869 }; 3870 3871 // MT Concurrent Marking Task 3872 class CMSConcMarkingTask: public YieldingFlexibleGangTask { 3873 CMSCollector* _collector; 3874 int _n_workers; // requested/desired # workers 3875 bool _asynch; 3876 bool _result; 3877 CompactibleFreeListSpace* _cms_space; 3878 char _pad_front[64]; // padding to ... 3879 HeapWord* _global_finger; // ... avoid sharing cache line 3880 char _pad_back[64]; 3881 HeapWord* _restart_addr; 3882 3883 // Exposed here for yielding support 3884 Mutex* const _bit_map_lock; 3885 3886 // The per thread work queues, available here for stealing 3887 OopTaskQueueSet* _task_queues; 3888 3889 // Termination (and yielding) support 3890 CMSConcMarkingTerminator _term; 3891 CMSConcMarkingTerminatorTerminator _term_term; 3892 3893 public: 3894 CMSConcMarkingTask(CMSCollector* collector, 3895 CompactibleFreeListSpace* cms_space, 3896 bool asynch, 3897 YieldingFlexibleWorkGang* workers, 3898 OopTaskQueueSet* task_queues): 3899 YieldingFlexibleGangTask("Concurrent marking done multi-threaded"), 3900 _collector(collector), 3901 _cms_space(cms_space), 3902 _asynch(asynch), _n_workers(0), _result(true), 3903 _task_queues(task_queues), 3904 _term(_n_workers, task_queues, _collector), 3905 _bit_map_lock(collector->bitMapLock()) 3906 { 3907 _requested_size = _n_workers; 3908 _term.set_task(this); 3909 _term_term.set_task(this); 3910 _restart_addr = _global_finger = _cms_space->bottom(); 3911 } 3912 3913 3914 OopTaskQueueSet* task_queues() { return _task_queues; } 3915 3916 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 3917 3918 HeapWord** global_finger_addr() { return &_global_finger; } 3919 3920 CMSConcMarkingTerminator* terminator() { return &_term; } 3921 3922 virtual void set_for_termination(int active_workers) { 3923 terminator()->reset_for_reuse(active_workers); 3924 } 3925 3926 void work(uint worker_id); 3927 bool should_yield() { 3928 return ConcurrentMarkSweepThread::should_yield() 3929 && !_collector->foregroundGCIsActive() 3930 && _asynch; 3931 } 3932 3933 virtual void coordinator_yield(); // stuff done by coordinator 3934 bool result() { return _result; } 3935 3936 void reset(HeapWord* ra) { 3937 assert(_global_finger >= _cms_space->end(), "Postcondition of ::work(i)"); 3938 _restart_addr = _global_finger = ra; 3939 _term.reset_for_reuse(); 3940 } 3941 3942 static bool get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 3943 OopTaskQueue* work_q); 3944 3945 private: 3946 void do_scan_and_mark(int i, CompactibleFreeListSpace* sp); 3947 void do_work_steal(int i); 3948 void bump_global_finger(HeapWord* f); 3949 }; 3950 3951 bool CMSConcMarkingTerminatorTerminator::should_exit_termination() { 3952 assert(_task != NULL, "Error"); 3953 return _task->yielding(); 3954 // Note that we do not need the disjunct || _task->should_yield() above 3955 // because we want terminating threads to yield only if the task 3956 // is already in the midst of yielding, which happens only after at least one 3957 // thread has yielded. 3958 } 3959 3960 void CMSConcMarkingTerminator::yield() { 3961 if (_task->should_yield()) { 3962 _task->yield(); 3963 } else { 3964 ParallelTaskTerminator::yield(); 3965 } 3966 } 3967 3968 //////////////////////////////////////////////////////////////// 3969 // Concurrent Marking Algorithm Sketch 3970 //////////////////////////////////////////////////////////////// 3971 // Until all tasks exhausted (both spaces): 3972 // -- claim next available chunk 3973 // -- bump global finger via CAS 3974 // -- find first object that starts in this chunk 3975 // and start scanning bitmap from that position 3976 // -- scan marked objects for oops 3977 // -- CAS-mark target, and if successful: 3978 // . if target oop is above global finger (volatile read) 3979 // nothing to do 3980 // . if target oop is in chunk and above local finger 3981 // then nothing to do 3982 // . else push on work-queue 3983 // -- Deal with possible overflow issues: 3984 // . local work-queue overflow causes stuff to be pushed on 3985 // global (common) overflow queue 3986 // . always first empty local work queue 3987 // . then get a batch of oops from global work queue if any 3988 // . then do work stealing 3989 // -- When all tasks claimed (both spaces) 3990 // and local work queue empty, 3991 // then in a loop do: 3992 // . check global overflow stack; steal a batch of oops and trace 3993 // . try to steal from other threads oif GOS is empty 3994 // . if neither is available, offer termination 3995 // -- Terminate and return result 3996 // 3997 void CMSConcMarkingTask::work(uint worker_id) { 3998 elapsedTimer _timer; 3999 ResourceMark rm; 4000 HandleMark hm; 4001 4002 DEBUG_ONLY(_collector->verify_overflow_empty();) 4003 4004 // Before we begin work, our work queue should be empty 4005 assert(work_queue(worker_id)->size() == 0, "Expected to be empty"); 4006 // Scan the bitmap covering _cms_space, tracing through grey objects. 4007 _timer.start(); 4008 do_scan_and_mark(worker_id, _cms_space); 4009 _timer.stop(); 4010 if (PrintCMSStatistics != 0) { 4011 gclog_or_tty->print_cr("Finished cms space scanning in %dth thread: %3.3f sec", 4012 worker_id, _timer.seconds()); 4013 // XXX: need xxx/xxx type of notation, two timers 4014 } 4015 4016 // ... do work stealing 4017 _timer.reset(); 4018 _timer.start(); 4019 do_work_steal(worker_id); 4020 _timer.stop(); 4021 if (PrintCMSStatistics != 0) { 4022 gclog_or_tty->print_cr("Finished work stealing in %dth thread: %3.3f sec", 4023 worker_id, _timer.seconds()); 4024 // XXX: need xxx/xxx type of notation, two timers 4025 } 4026 assert(_collector->_markStack.isEmpty(), "Should have been emptied"); 4027 assert(work_queue(worker_id)->size() == 0, "Should have been emptied"); 4028 // Note that under the current task protocol, the 4029 // following assertion is true even of the spaces 4030 // expanded since the completion of the concurrent 4031 // marking. XXX This will likely change under a strict 4032 // ABORT semantics. 4033 // After perm removal the comparison was changed to 4034 // greater than or equal to from strictly greater than. 4035 // Before perm removal the highest address sweep would 4036 // have been at the end of perm gen but now is at the 4037 // end of the tenured gen. 4038 assert(_global_finger >= _cms_space->end(), 4039 "All tasks have been completed"); 4040 DEBUG_ONLY(_collector->verify_overflow_empty();) 4041 } 4042 4043 void CMSConcMarkingTask::bump_global_finger(HeapWord* f) { 4044 HeapWord* read = _global_finger; 4045 HeapWord* cur = read; 4046 while (f > read) { 4047 cur = read; 4048 read = (HeapWord*) Atomic::cmpxchg_ptr(f, &_global_finger, cur); 4049 if (cur == read) { 4050 // our cas succeeded 4051 assert(_global_finger >= f, "protocol consistency"); 4052 break; 4053 } 4054 } 4055 } 4056 4057 // This is really inefficient, and should be redone by 4058 // using (not yet available) block-read and -write interfaces to the 4059 // stack and the work_queue. XXX FIX ME !!! 4060 bool CMSConcMarkingTask::get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 4061 OopTaskQueue* work_q) { 4062 // Fast lock-free check 4063 if (ovflw_stk->length() == 0) { 4064 return false; 4065 } 4066 assert(work_q->size() == 0, "Shouldn't steal"); 4067 MutexLockerEx ml(ovflw_stk->par_lock(), 4068 Mutex::_no_safepoint_check_flag); 4069 // Grab up to 1/4 the size of the work queue 4070 size_t num = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 4071 (size_t)ParGCDesiredObjsFromOverflowList); 4072 num = MIN2(num, ovflw_stk->length()); 4073 for (int i = (int) num; i > 0; i--) { 4074 oop cur = ovflw_stk->pop(); 4075 assert(cur != NULL, "Counted wrong?"); 4076 work_q->push(cur); 4077 } 4078 return num > 0; 4079 } 4080 4081 void CMSConcMarkingTask::do_scan_and_mark(int i, CompactibleFreeListSpace* sp) { 4082 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 4083 int n_tasks = pst->n_tasks(); 4084 // We allow that there may be no tasks to do here because 4085 // we are restarting after a stack overflow. 4086 assert(pst->valid() || n_tasks == 0, "Uninitialized use?"); 4087 uint nth_task = 0; 4088 4089 HeapWord* aligned_start = sp->bottom(); 4090 if (sp->used_region().contains(_restart_addr)) { 4091 // Align down to a card boundary for the start of 0th task 4092 // for this space. 4093 aligned_start = 4094 (HeapWord*)align_size_down((uintptr_t)_restart_addr, 4095 CardTableModRefBS::card_size); 4096 } 4097 4098 size_t chunk_size = sp->marking_task_size(); 4099 while (!pst->is_task_claimed(/* reference */ nth_task)) { 4100 // Having claimed the nth task in this space, 4101 // compute the chunk that it corresponds to: 4102 MemRegion span = MemRegion(aligned_start + nth_task*chunk_size, 4103 aligned_start + (nth_task+1)*chunk_size); 4104 // Try and bump the global finger via a CAS; 4105 // note that we need to do the global finger bump 4106 // _before_ taking the intersection below, because 4107 // the task corresponding to that region will be 4108 // deemed done even if the used_region() expands 4109 // because of allocation -- as it almost certainly will 4110 // during start-up while the threads yield in the 4111 // closure below. 4112 HeapWord* finger = span.end(); 4113 bump_global_finger(finger); // atomically 4114 // There are null tasks here corresponding to chunks 4115 // beyond the "top" address of the space. 4116 span = span.intersection(sp->used_region()); 4117 if (!span.is_empty()) { // Non-null task 4118 HeapWord* prev_obj; 4119 assert(!span.contains(_restart_addr) || nth_task == 0, 4120 "Inconsistency"); 4121 if (nth_task == 0) { 4122 // For the 0th task, we'll not need to compute a block_start. 4123 if (span.contains(_restart_addr)) { 4124 // In the case of a restart because of stack overflow, 4125 // we might additionally skip a chunk prefix. 4126 prev_obj = _restart_addr; 4127 } else { 4128 prev_obj = span.start(); 4129 } 4130 } else { 4131 // We want to skip the first object because 4132 // the protocol is to scan any object in its entirety 4133 // that _starts_ in this span; a fortiori, any 4134 // object starting in an earlier span is scanned 4135 // as part of an earlier claimed task. 4136 // Below we use the "careful" version of block_start 4137 // so we do not try to navigate uninitialized objects. 4138 prev_obj = sp->block_start_careful(span.start()); 4139 // Below we use a variant of block_size that uses the 4140 // Printezis bits to avoid waiting for allocated 4141 // objects to become initialized/parsable. 4142 while (prev_obj < span.start()) { 4143 size_t sz = sp->block_size_no_stall(prev_obj, _collector); 4144 if (sz > 0) { 4145 prev_obj += sz; 4146 } else { 4147 // In this case we may end up doing a bit of redundant 4148 // scanning, but that appears unavoidable, short of 4149 // locking the free list locks; see bug 6324141. 4150 break; 4151 } 4152 } 4153 } 4154 if (prev_obj < span.end()) { 4155 MemRegion my_span = MemRegion(prev_obj, span.end()); 4156 // Do the marking work within a non-empty span -- 4157 // the last argument to the constructor indicates whether the 4158 // iteration should be incremental with periodic yields. 4159 Par_MarkFromRootsClosure cl(this, _collector, my_span, 4160 &_collector->_markBitMap, 4161 work_queue(i), 4162 &_collector->_markStack, 4163 _asynch); 4164 _collector->_markBitMap.iterate(&cl, my_span.start(), my_span.end()); 4165 } // else nothing to do for this task 4166 } // else nothing to do for this task 4167 } 4168 // We'd be tempted to assert here that since there are no 4169 // more tasks left to claim in this space, the global_finger 4170 // must exceed space->top() and a fortiori space->end(). However, 4171 // that would not quite be correct because the bumping of 4172 // global_finger occurs strictly after the claiming of a task, 4173 // so by the time we reach here the global finger may not yet 4174 // have been bumped up by the thread that claimed the last 4175 // task. 4176 pst->all_tasks_completed(); 4177 } 4178 4179 class Par_ConcMarkingClosure: public CMSOopClosure { 4180 private: 4181 CMSCollector* _collector; 4182 CMSConcMarkingTask* _task; 4183 MemRegion _span; 4184 CMSBitMap* _bit_map; 4185 CMSMarkStack* _overflow_stack; 4186 OopTaskQueue* _work_queue; 4187 protected: 4188 DO_OOP_WORK_DEFN 4189 public: 4190 Par_ConcMarkingClosure(CMSCollector* collector, CMSConcMarkingTask* task, OopTaskQueue* work_queue, 4191 CMSBitMap* bit_map, CMSMarkStack* overflow_stack): 4192 CMSOopClosure(collector->ref_processor()), 4193 _collector(collector), 4194 _task(task), 4195 _span(collector->_span), 4196 _work_queue(work_queue), 4197 _bit_map(bit_map), 4198 _overflow_stack(overflow_stack) 4199 { } 4200 virtual void do_oop(oop* p); 4201 virtual void do_oop(narrowOop* p); 4202 4203 void trim_queue(size_t max); 4204 void handle_stack_overflow(HeapWord* lost); 4205 void do_yield_check() { 4206 if (_task->should_yield()) { 4207 _task->yield(); 4208 } 4209 } 4210 }; 4211 4212 // Grey object scanning during work stealing phase -- 4213 // the salient assumption here is that any references 4214 // that are in these stolen objects being scanned must 4215 // already have been initialized (else they would not have 4216 // been published), so we do not need to check for 4217 // uninitialized objects before pushing here. 4218 void Par_ConcMarkingClosure::do_oop(oop obj) { 4219 assert(obj->is_oop_or_null(true), "expected an oop or NULL"); 4220 HeapWord* addr = (HeapWord*)obj; 4221 // Check if oop points into the CMS generation 4222 // and is not marked 4223 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 4224 // a white object ... 4225 // If we manage to "claim" the object, by being the 4226 // first thread to mark it, then we push it on our 4227 // marking stack 4228 if (_bit_map->par_mark(addr)) { // ... now grey 4229 // push on work queue (grey set) 4230 bool simulate_overflow = false; 4231 NOT_PRODUCT( 4232 if (CMSMarkStackOverflowALot && 4233 _collector->simulate_overflow()) { 4234 // simulate a stack overflow 4235 simulate_overflow = true; 4236 } 4237 ) 4238 if (simulate_overflow || 4239 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) { 4240 // stack overflow 4241 if (PrintCMSStatistics != 0) { 4242 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 4243 SIZE_FORMAT, _overflow_stack->capacity()); 4244 } 4245 // We cannot assert that the overflow stack is full because 4246 // it may have been emptied since. 4247 assert(simulate_overflow || 4248 _work_queue->size() == _work_queue->max_elems(), 4249 "Else push should have succeeded"); 4250 handle_stack_overflow(addr); 4251 } 4252 } // Else, some other thread got there first 4253 do_yield_check(); 4254 } 4255 } 4256 4257 void Par_ConcMarkingClosure::do_oop(oop* p) { Par_ConcMarkingClosure::do_oop_work(p); } 4258 void Par_ConcMarkingClosure::do_oop(narrowOop* p) { Par_ConcMarkingClosure::do_oop_work(p); } 4259 4260 void Par_ConcMarkingClosure::trim_queue(size_t max) { 4261 while (_work_queue->size() > max) { 4262 oop new_oop; 4263 if (_work_queue->pop_local(new_oop)) { 4264 assert(new_oop->is_oop(), "Should be an oop"); 4265 assert(_bit_map->isMarked((HeapWord*)new_oop), "Grey object"); 4266 assert(_span.contains((HeapWord*)new_oop), "Not in span"); 4267 new_oop->oop_iterate(this); // do_oop() above 4268 do_yield_check(); 4269 } 4270 } 4271 } 4272 4273 // Upon stack overflow, we discard (part of) the stack, 4274 // remembering the least address amongst those discarded 4275 // in CMSCollector's _restart_address. 4276 void Par_ConcMarkingClosure::handle_stack_overflow(HeapWord* lost) { 4277 // We need to do this under a mutex to prevent other 4278 // workers from interfering with the work done below. 4279 MutexLockerEx ml(_overflow_stack->par_lock(), 4280 Mutex::_no_safepoint_check_flag); 4281 // Remember the least grey address discarded 4282 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 4283 _collector->lower_restart_addr(ra); 4284 _overflow_stack->reset(); // discard stack contents 4285 _overflow_stack->expand(); // expand the stack if possible 4286 } 4287 4288 4289 void CMSConcMarkingTask::do_work_steal(int i) { 4290 OopTaskQueue* work_q = work_queue(i); 4291 oop obj_to_scan; 4292 CMSBitMap* bm = &(_collector->_markBitMap); 4293 CMSMarkStack* ovflw = &(_collector->_markStack); 4294 int* seed = _collector->hash_seed(i); 4295 Par_ConcMarkingClosure cl(_collector, this, work_q, bm, ovflw); 4296 while (true) { 4297 cl.trim_queue(0); 4298 assert(work_q->size() == 0, "Should have been emptied above"); 4299 if (get_work_from_overflow_stack(ovflw, work_q)) { 4300 // Can't assert below because the work obtained from the 4301 // overflow stack may already have been stolen from us. 4302 // assert(work_q->size() > 0, "Work from overflow stack"); 4303 continue; 4304 } else if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 4305 assert(obj_to_scan->is_oop(), "Should be an oop"); 4306 assert(bm->isMarked((HeapWord*)obj_to_scan), "Grey object"); 4307 obj_to_scan->oop_iterate(&cl); 4308 } else if (terminator()->offer_termination(&_term_term)) { 4309 assert(work_q->size() == 0, "Impossible!"); 4310 break; 4311 } else if (yielding() || should_yield()) { 4312 yield(); 4313 } 4314 } 4315 } 4316 4317 // This is run by the CMS (coordinator) thread. 4318 void CMSConcMarkingTask::coordinator_yield() { 4319 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 4320 "CMS thread should hold CMS token"); 4321 // First give up the locks, then yield, then re-lock 4322 // We should probably use a constructor/destructor idiom to 4323 // do this unlock/lock or modify the MutexUnlocker class to 4324 // serve our purpose. XXX 4325 assert_lock_strong(_bit_map_lock); 4326 _bit_map_lock->unlock(); 4327 ConcurrentMarkSweepThread::desynchronize(true); 4328 ConcurrentMarkSweepThread::acknowledge_yield_request(); 4329 _collector->stopTimer(); 4330 if (PrintCMSStatistics != 0) { 4331 _collector->incrementYields(); 4332 } 4333 _collector->icms_wait(); 4334 4335 // It is possible for whichever thread initiated the yield request 4336 // not to get a chance to wake up and take the bitmap lock between 4337 // this thread releasing it and reacquiring it. So, while the 4338 // should_yield() flag is on, let's sleep for a bit to give the 4339 // other thread a chance to wake up. The limit imposed on the number 4340 // of iterations is defensive, to avoid any unforseen circumstances 4341 // putting us into an infinite loop. Since it's always been this 4342 // (coordinator_yield()) method that was observed to cause the 4343 // problem, we are using a parameter (CMSCoordinatorYieldSleepCount) 4344 // which is by default non-zero. For the other seven methods that 4345 // also perform the yield operation, as are using a different 4346 // parameter (CMSYieldSleepCount) which is by default zero. This way we 4347 // can enable the sleeping for those methods too, if necessary. 4348 // See 6442774. 4349 // 4350 // We really need to reconsider the synchronization between the GC 4351 // thread and the yield-requesting threads in the future and we 4352 // should really use wait/notify, which is the recommended 4353 // way of doing this type of interaction. Additionally, we should 4354 // consolidate the eight methods that do the yield operation and they 4355 // are almost identical into one for better maintainability and 4356 // readability. See 6445193. 4357 // 4358 // Tony 2006.06.29 4359 for (unsigned i = 0; i < CMSCoordinatorYieldSleepCount && 4360 ConcurrentMarkSweepThread::should_yield() && 4361 !CMSCollector::foregroundGCIsActive(); ++i) { 4362 os::sleep(Thread::current(), 1, false); 4363 ConcurrentMarkSweepThread::acknowledge_yield_request(); 4364 } 4365 4366 ConcurrentMarkSweepThread::synchronize(true); 4367 _bit_map_lock->lock_without_safepoint_check(); 4368 _collector->startTimer(); 4369 } 4370 4371 bool CMSCollector::do_marking_mt(bool asynch) { 4372 assert(ConcGCThreads > 0 && conc_workers() != NULL, "precondition"); 4373 int num_workers = AdaptiveSizePolicy::calc_active_conc_workers( 4374 conc_workers()->total_workers(), 4375 conc_workers()->active_workers(), 4376 Threads::number_of_non_daemon_threads()); 4377 conc_workers()->set_active_workers(num_workers); 4378 4379 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 4380 4381 CMSConcMarkingTask tsk(this, 4382 cms_space, 4383 asynch, 4384 conc_workers(), 4385 task_queues()); 4386 4387 // Since the actual number of workers we get may be different 4388 // from the number we requested above, do we need to do anything different 4389 // below? In particular, may be we need to subclass the SequantialSubTasksDone 4390 // class?? XXX 4391 cms_space ->initialize_sequential_subtasks_for_marking(num_workers); 4392 4393 // Refs discovery is already non-atomic. 4394 assert(!ref_processor()->discovery_is_atomic(), "Should be non-atomic"); 4395 assert(ref_processor()->discovery_is_mt(), "Discovery should be MT"); 4396 conc_workers()->start_task(&tsk); 4397 while (tsk.yielded()) { 4398 tsk.coordinator_yield(); 4399 conc_workers()->continue_task(&tsk); 4400 } 4401 // If the task was aborted, _restart_addr will be non-NULL 4402 assert(tsk.completed() || _restart_addr != NULL, "Inconsistency"); 4403 while (_restart_addr != NULL) { 4404 // XXX For now we do not make use of ABORTED state and have not 4405 // yet implemented the right abort semantics (even in the original 4406 // single-threaded CMS case). That needs some more investigation 4407 // and is deferred for now; see CR# TBF. 07252005YSR. XXX 4408 assert(!CMSAbortSemantics || tsk.aborted(), "Inconsistency"); 4409 // If _restart_addr is non-NULL, a marking stack overflow 4410 // occurred; we need to do a fresh marking iteration from the 4411 // indicated restart address. 4412 if (_foregroundGCIsActive && asynch) { 4413 // We may be running into repeated stack overflows, having 4414 // reached the limit of the stack size, while making very 4415 // slow forward progress. It may be best to bail out and 4416 // let the foreground collector do its job. 4417 // Clear _restart_addr, so that foreground GC 4418 // works from scratch. This avoids the headache of 4419 // a "rescan" which would otherwise be needed because 4420 // of the dirty mod union table & card table. 4421 _restart_addr = NULL; 4422 return false; 4423 } 4424 // Adjust the task to restart from _restart_addr 4425 tsk.reset(_restart_addr); 4426 cms_space ->initialize_sequential_subtasks_for_marking(num_workers, 4427 _restart_addr); 4428 _restart_addr = NULL; 4429 // Get the workers going again 4430 conc_workers()->start_task(&tsk); 4431 while (tsk.yielded()) { 4432 tsk.coordinator_yield(); 4433 conc_workers()->continue_task(&tsk); 4434 } 4435 } 4436 assert(tsk.completed(), "Inconsistency"); 4437 assert(tsk.result() == true, "Inconsistency"); 4438 return true; 4439 } 4440 4441 bool CMSCollector::do_marking_st(bool asynch) { 4442 ResourceMark rm; 4443 HandleMark hm; 4444 4445 // Temporarily make refs discovery single threaded (non-MT) 4446 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false); 4447 MarkFromRootsClosure markFromRootsClosure(this, _span, &_markBitMap, 4448 &_markStack, CMSYield && asynch); 4449 // the last argument to iterate indicates whether the iteration 4450 // should be incremental with periodic yields. 4451 _markBitMap.iterate(&markFromRootsClosure); 4452 // If _restart_addr is non-NULL, a marking stack overflow 4453 // occurred; we need to do a fresh iteration from the 4454 // indicated restart address. 4455 while (_restart_addr != NULL) { 4456 if (_foregroundGCIsActive && asynch) { 4457 // We may be running into repeated stack overflows, having 4458 // reached the limit of the stack size, while making very 4459 // slow forward progress. It may be best to bail out and 4460 // let the foreground collector do its job. 4461 // Clear _restart_addr, so that foreground GC 4462 // works from scratch. This avoids the headache of 4463 // a "rescan" which would otherwise be needed because 4464 // of the dirty mod union table & card table. 4465 _restart_addr = NULL; 4466 return false; // indicating failure to complete marking 4467 } 4468 // Deal with stack overflow: 4469 // we restart marking from _restart_addr 4470 HeapWord* ra = _restart_addr; 4471 markFromRootsClosure.reset(ra); 4472 _restart_addr = NULL; 4473 _markBitMap.iterate(&markFromRootsClosure, ra, _span.end()); 4474 } 4475 return true; 4476 } 4477 4478 void CMSCollector::preclean() { 4479 check_correct_thread_executing(); 4480 assert(Thread::current()->is_ConcurrentGC_thread(), "Wrong thread"); 4481 verify_work_stacks_empty(); 4482 verify_overflow_empty(); 4483 _abort_preclean = false; 4484 if (CMSPrecleaningEnabled) { 4485 if (!CMSEdenChunksRecordAlways) { 4486 _eden_chunk_index = 0; 4487 } 4488 size_t used = get_eden_used(); 4489 size_t capacity = get_eden_capacity(); 4490 // Don't start sampling unless we will get sufficiently 4491 // many samples. 4492 if (used < (capacity/(CMSScheduleRemarkSamplingRatio * 100) 4493 * CMSScheduleRemarkEdenPenetration)) { 4494 _start_sampling = true; 4495 } else { 4496 _start_sampling = false; 4497 } 4498 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 4499 CMSPhaseAccounting pa(this, "preclean", !PrintGCDetails); 4500 preclean_work(CMSPrecleanRefLists1, CMSPrecleanSurvivors1); 4501 } 4502 CMSTokenSync x(true); // is cms thread 4503 if (CMSPrecleaningEnabled) { 4504 sample_eden(); 4505 _collectorState = AbortablePreclean; 4506 } else { 4507 _collectorState = FinalMarking; 4508 } 4509 verify_work_stacks_empty(); 4510 verify_overflow_empty(); 4511 } 4512 4513 // Try and schedule the remark such that young gen 4514 // occupancy is CMSScheduleRemarkEdenPenetration %. 4515 void CMSCollector::abortable_preclean() { 4516 check_correct_thread_executing(); 4517 assert(CMSPrecleaningEnabled, "Inconsistent control state"); 4518 assert(_collectorState == AbortablePreclean, "Inconsistent control state"); 4519 4520 // If Eden's current occupancy is below this threshold, 4521 // immediately schedule the remark; else preclean 4522 // past the next scavenge in an effort to 4523 // schedule the pause as described above. By choosing 4524 // CMSScheduleRemarkEdenSizeThreshold >= max eden size 4525 // we will never do an actual abortable preclean cycle. 4526 if (get_eden_used() > CMSScheduleRemarkEdenSizeThreshold) { 4527 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 4528 CMSPhaseAccounting pa(this, "abortable-preclean", !PrintGCDetails); 4529 // We need more smarts in the abortable preclean 4530 // loop below to deal with cases where allocation 4531 // in young gen is very very slow, and our precleaning 4532 // is running a losing race against a horde of 4533 // mutators intent on flooding us with CMS updates 4534 // (dirty cards). 4535 // One, admittedly dumb, strategy is to give up 4536 // after a certain number of abortable precleaning loops 4537 // or after a certain maximum time. We want to make 4538 // this smarter in the next iteration. 4539 // XXX FIX ME!!! YSR 4540 size_t loops = 0, workdone = 0, cumworkdone = 0, waited = 0; 4541 while (!(should_abort_preclean() || 4542 ConcurrentMarkSweepThread::should_terminate())) { 4543 workdone = preclean_work(CMSPrecleanRefLists2, CMSPrecleanSurvivors2); 4544 cumworkdone += workdone; 4545 loops++; 4546 // Voluntarily terminate abortable preclean phase if we have 4547 // been at it for too long. 4548 if ((CMSMaxAbortablePrecleanLoops != 0) && 4549 loops >= CMSMaxAbortablePrecleanLoops) { 4550 if (PrintGCDetails) { 4551 gclog_or_tty->print(" CMS: abort preclean due to loops "); 4552 } 4553 break; 4554 } 4555 if (pa.wallclock_millis() > CMSMaxAbortablePrecleanTime) { 4556 if (PrintGCDetails) { 4557 gclog_or_tty->print(" CMS: abort preclean due to time "); 4558 } 4559 break; 4560 } 4561 // If we are doing little work each iteration, we should 4562 // take a short break. 4563 if (workdone < CMSAbortablePrecleanMinWorkPerIteration) { 4564 // Sleep for some time, waiting for work to accumulate 4565 stopTimer(); 4566 cmsThread()->wait_on_cms_lock(CMSAbortablePrecleanWaitMillis); 4567 startTimer(); 4568 waited++; 4569 } 4570 } 4571 if (PrintCMSStatistics > 0) { 4572 gclog_or_tty->print(" [" SIZE_FORMAT " iterations, " SIZE_FORMAT " waits, " SIZE_FORMAT " cards)] ", 4573 loops, waited, cumworkdone); 4574 } 4575 } 4576 CMSTokenSync x(true); // is cms thread 4577 if (_collectorState != Idling) { 4578 assert(_collectorState == AbortablePreclean, 4579 "Spontaneous state transition?"); 4580 _collectorState = FinalMarking; 4581 } // Else, a foreground collection completed this CMS cycle. 4582 return; 4583 } 4584 4585 // Respond to an Eden sampling opportunity 4586 void CMSCollector::sample_eden() { 4587 // Make sure a young gc cannot sneak in between our 4588 // reading and recording of a sample. 4589 assert(Thread::current()->is_ConcurrentGC_thread(), 4590 "Only the cms thread may collect Eden samples"); 4591 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 4592 "Should collect samples while holding CMS token"); 4593 if (!_start_sampling) { 4594 return; 4595 } 4596 // When CMSEdenChunksRecordAlways is true, the eden chunk array 4597 // is populated by the young generation. 4598 if (_eden_chunk_array != NULL && !CMSEdenChunksRecordAlways) { 4599 if (_eden_chunk_index < _eden_chunk_capacity) { 4600 _eden_chunk_array[_eden_chunk_index] = *_top_addr; // take sample 4601 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr, 4602 "Unexpected state of Eden"); 4603 // We'd like to check that what we just sampled is an oop-start address; 4604 // however, we cannot do that here since the object may not yet have been 4605 // initialized. So we'll instead do the check when we _use_ this sample 4606 // later. 4607 if (_eden_chunk_index == 0 || 4608 (pointer_delta(_eden_chunk_array[_eden_chunk_index], 4609 _eden_chunk_array[_eden_chunk_index-1]) 4610 >= CMSSamplingGrain)) { 4611 _eden_chunk_index++; // commit sample 4612 } 4613 } 4614 } 4615 if ((_collectorState == AbortablePreclean) && !_abort_preclean) { 4616 size_t used = get_eden_used(); 4617 size_t capacity = get_eden_capacity(); 4618 assert(used <= capacity, "Unexpected state of Eden"); 4619 if (used > (capacity/100 * CMSScheduleRemarkEdenPenetration)) { 4620 _abort_preclean = true; 4621 } 4622 } 4623 } 4624 4625 4626 size_t CMSCollector::preclean_work(bool clean_refs, bool clean_survivor) { 4627 assert(_collectorState == Precleaning || 4628 _collectorState == AbortablePreclean, "incorrect state"); 4629 ResourceMark rm; 4630 HandleMark hm; 4631 4632 // Precleaning is currently not MT but the reference processor 4633 // may be set for MT. Disable it temporarily here. 4634 ReferenceProcessor* rp = ref_processor(); 4635 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(rp, false); 4636 4637 // Do one pass of scrubbing the discovered reference lists 4638 // to remove any reference objects with strongly-reachable 4639 // referents. 4640 if (clean_refs) { 4641 CMSPrecleanRefsYieldClosure yield_cl(this); 4642 assert(rp->span().equals(_span), "Spans should be equal"); 4643 CMSKeepAliveClosure keep_alive(this, _span, &_markBitMap, 4644 &_markStack, true /* preclean */); 4645 CMSDrainMarkingStackClosure complete_trace(this, 4646 _span, &_markBitMap, &_markStack, 4647 &keep_alive, true /* preclean */); 4648 4649 // We don't want this step to interfere with a young 4650 // collection because we don't want to take CPU 4651 // or memory bandwidth away from the young GC threads 4652 // (which may be as many as there are CPUs). 4653 // Note that we don't need to protect ourselves from 4654 // interference with mutators because they can't 4655 // manipulate the discovered reference lists nor affect 4656 // the computed reachability of the referents, the 4657 // only properties manipulated by the precleaning 4658 // of these reference lists. 4659 stopTimer(); 4660 CMSTokenSyncWithLocks x(true /* is cms thread */, 4661 bitMapLock()); 4662 startTimer(); 4663 sample_eden(); 4664 4665 // The following will yield to allow foreground 4666 // collection to proceed promptly. XXX YSR: 4667 // The code in this method may need further 4668 // tweaking for better performance and some restructuring 4669 // for cleaner interfaces. 4670 GCTimer *gc_timer = NULL; // Currently not tracing concurrent phases 4671 rp->preclean_discovered_references( 4672 rp->is_alive_non_header(), &keep_alive, &complete_trace, &yield_cl, 4673 gc_timer); 4674 } 4675 4676 if (clean_survivor) { // preclean the active survivor space(s) 4677 assert(_young_gen->kind() == Generation::DefNew || 4678 _young_gen->kind() == Generation::ParNew || 4679 _young_gen->kind() == Generation::ASParNew, 4680 "incorrect type for cast"); 4681 DefNewGeneration* dng = (DefNewGeneration*)_young_gen; 4682 PushAndMarkClosure pam_cl(this, _span, ref_processor(), 4683 &_markBitMap, &_modUnionTable, 4684 &_markStack, true /* precleaning phase */); 4685 stopTimer(); 4686 CMSTokenSyncWithLocks ts(true /* is cms thread */, 4687 bitMapLock()); 4688 startTimer(); 4689 unsigned int before_count = 4690 GenCollectedHeap::heap()->total_collections(); 4691 SurvivorSpacePrecleanClosure 4692 sss_cl(this, _span, &_markBitMap, &_markStack, 4693 &pam_cl, before_count, CMSYield); 4694 dng->from()->object_iterate_careful(&sss_cl); 4695 dng->to()->object_iterate_careful(&sss_cl); 4696 } 4697 MarkRefsIntoAndScanClosure 4698 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable, 4699 &_markStack, this, CMSYield, 4700 true /* precleaning phase */); 4701 // CAUTION: The following closure has persistent state that may need to 4702 // be reset upon a decrease in the sequence of addresses it 4703 // processes. 4704 ScanMarkedObjectsAgainCarefullyClosure 4705 smoac_cl(this, _span, 4706 &_markBitMap, &_markStack, &mrias_cl, CMSYield); 4707 4708 // Preclean dirty cards in ModUnionTable and CardTable using 4709 // appropriate convergence criterion; 4710 // repeat CMSPrecleanIter times unless we find that 4711 // we are losing. 4712 assert(CMSPrecleanIter < 10, "CMSPrecleanIter is too large"); 4713 assert(CMSPrecleanNumerator < CMSPrecleanDenominator, 4714 "Bad convergence multiplier"); 4715 assert(CMSPrecleanThreshold >= 100, 4716 "Unreasonably low CMSPrecleanThreshold"); 4717 4718 size_t numIter, cumNumCards, lastNumCards, curNumCards; 4719 for (numIter = 0, cumNumCards = lastNumCards = curNumCards = 0; 4720 numIter < CMSPrecleanIter; 4721 numIter++, lastNumCards = curNumCards, cumNumCards += curNumCards) { 4722 curNumCards = preclean_mod_union_table(_cmsGen, &smoac_cl); 4723 if (Verbose && PrintGCDetails) { 4724 gclog_or_tty->print(" (modUnionTable: " SIZE_FORMAT " cards)", curNumCards); 4725 } 4726 // Either there are very few dirty cards, so re-mark 4727 // pause will be small anyway, or our pre-cleaning isn't 4728 // that much faster than the rate at which cards are being 4729 // dirtied, so we might as well stop and re-mark since 4730 // precleaning won't improve our re-mark time by much. 4731 if (curNumCards <= CMSPrecleanThreshold || 4732 (numIter > 0 && 4733 (curNumCards * CMSPrecleanDenominator > 4734 lastNumCards * CMSPrecleanNumerator))) { 4735 numIter++; 4736 cumNumCards += curNumCards; 4737 break; 4738 } 4739 } 4740 4741 preclean_klasses(&mrias_cl, _cmsGen->freelistLock()); 4742 4743 curNumCards = preclean_card_table(_cmsGen, &smoac_cl); 4744 cumNumCards += curNumCards; 4745 if (PrintGCDetails && PrintCMSStatistics != 0) { 4746 gclog_or_tty->print_cr(" (cardTable: " SIZE_FORMAT " cards, re-scanned " SIZE_FORMAT " cards, " SIZE_FORMAT " iterations)", 4747 curNumCards, cumNumCards, numIter); 4748 } 4749 return cumNumCards; // as a measure of useful work done 4750 } 4751 4752 // PRECLEANING NOTES: 4753 // Precleaning involves: 4754 // . reading the bits of the modUnionTable and clearing the set bits. 4755 // . For the cards corresponding to the set bits, we scan the 4756 // objects on those cards. This means we need the free_list_lock 4757 // so that we can safely iterate over the CMS space when scanning 4758 // for oops. 4759 // . When we scan the objects, we'll be both reading and setting 4760 // marks in the marking bit map, so we'll need the marking bit map. 4761 // . For protecting _collector_state transitions, we take the CGC_lock. 4762 // Note that any races in the reading of of card table entries by the 4763 // CMS thread on the one hand and the clearing of those entries by the 4764 // VM thread or the setting of those entries by the mutator threads on the 4765 // other are quite benign. However, for efficiency it makes sense to keep 4766 // the VM thread from racing with the CMS thread while the latter is 4767 // dirty card info to the modUnionTable. We therefore also use the 4768 // CGC_lock to protect the reading of the card table and the mod union 4769 // table by the CM thread. 4770 // . We run concurrently with mutator updates, so scanning 4771 // needs to be done carefully -- we should not try to scan 4772 // potentially uninitialized objects. 4773 // 4774 // Locking strategy: While holding the CGC_lock, we scan over and 4775 // reset a maximal dirty range of the mod union / card tables, then lock 4776 // the free_list_lock and bitmap lock to do a full marking, then 4777 // release these locks; and repeat the cycle. This allows for a 4778 // certain amount of fairness in the sharing of these locks between 4779 // the CMS collector on the one hand, and the VM thread and the 4780 // mutators on the other. 4781 4782 // NOTE: preclean_mod_union_table() and preclean_card_table() 4783 // further below are largely identical; if you need to modify 4784 // one of these methods, please check the other method too. 4785 4786 size_t CMSCollector::preclean_mod_union_table( 4787 ConcurrentMarkSweepGeneration* gen, 4788 ScanMarkedObjectsAgainCarefullyClosure* cl) { 4789 verify_work_stacks_empty(); 4790 verify_overflow_empty(); 4791 4792 // strategy: starting with the first card, accumulate contiguous 4793 // ranges of dirty cards; clear these cards, then scan the region 4794 // covered by these cards. 4795 4796 // Since all of the MUT is committed ahead, we can just use 4797 // that, in case the generations expand while we are precleaning. 4798 // It might also be fine to just use the committed part of the 4799 // generation, but we might potentially miss cards when the 4800 // generation is rapidly expanding while we are in the midst 4801 // of precleaning. 4802 HeapWord* startAddr = gen->reserved().start(); 4803 HeapWord* endAddr = gen->reserved().end(); 4804 4805 cl->setFreelistLock(gen->freelistLock()); // needed for yielding 4806 4807 size_t numDirtyCards, cumNumDirtyCards; 4808 HeapWord *nextAddr, *lastAddr; 4809 for (cumNumDirtyCards = numDirtyCards = 0, 4810 nextAddr = lastAddr = startAddr; 4811 nextAddr < endAddr; 4812 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 4813 4814 ResourceMark rm; 4815 HandleMark hm; 4816 4817 MemRegion dirtyRegion; 4818 { 4819 stopTimer(); 4820 // Potential yield point 4821 CMSTokenSync ts(true); 4822 startTimer(); 4823 sample_eden(); 4824 // Get dirty region starting at nextOffset (inclusive), 4825 // simultaneously clearing it. 4826 dirtyRegion = 4827 _modUnionTable.getAndClearMarkedRegion(nextAddr, endAddr); 4828 assert(dirtyRegion.start() >= nextAddr, 4829 "returned region inconsistent?"); 4830 } 4831 // Remember where the next search should begin. 4832 // The returned region (if non-empty) is a right open interval, 4833 // so lastOffset is obtained from the right end of that 4834 // interval. 4835 lastAddr = dirtyRegion.end(); 4836 // Should do something more transparent and less hacky XXX 4837 numDirtyCards = 4838 _modUnionTable.heapWordDiffToOffsetDiff(dirtyRegion.word_size()); 4839 4840 // We'll scan the cards in the dirty region (with periodic 4841 // yields for foreground GC as needed). 4842 if (!dirtyRegion.is_empty()) { 4843 assert(numDirtyCards > 0, "consistency check"); 4844 HeapWord* stop_point = NULL; 4845 stopTimer(); 4846 // Potential yield point 4847 CMSTokenSyncWithLocks ts(true, gen->freelistLock(), 4848 bitMapLock()); 4849 startTimer(); 4850 { 4851 verify_work_stacks_empty(); 4852 verify_overflow_empty(); 4853 sample_eden(); 4854 stop_point = 4855 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 4856 } 4857 if (stop_point != NULL) { 4858 // The careful iteration stopped early either because it found an 4859 // uninitialized object, or because we were in the midst of an 4860 // "abortable preclean", which should now be aborted. Redirty 4861 // the bits corresponding to the partially-scanned or unscanned 4862 // cards. We'll either restart at the next block boundary or 4863 // abort the preclean. 4864 assert((_collectorState == AbortablePreclean && should_abort_preclean()), 4865 "Should only be AbortablePreclean."); 4866 _modUnionTable.mark_range(MemRegion(stop_point, dirtyRegion.end())); 4867 if (should_abort_preclean()) { 4868 break; // out of preclean loop 4869 } else { 4870 // Compute the next address at which preclean should pick up; 4871 // might need bitMapLock in order to read P-bits. 4872 lastAddr = next_card_start_after_block(stop_point); 4873 } 4874 } 4875 } else { 4876 assert(lastAddr == endAddr, "consistency check"); 4877 assert(numDirtyCards == 0, "consistency check"); 4878 break; 4879 } 4880 } 4881 verify_work_stacks_empty(); 4882 verify_overflow_empty(); 4883 return cumNumDirtyCards; 4884 } 4885 4886 // NOTE: preclean_mod_union_table() above and preclean_card_table() 4887 // below are largely identical; if you need to modify 4888 // one of these methods, please check the other method too. 4889 4890 size_t CMSCollector::preclean_card_table(ConcurrentMarkSweepGeneration* gen, 4891 ScanMarkedObjectsAgainCarefullyClosure* cl) { 4892 // strategy: it's similar to precleamModUnionTable above, in that 4893 // we accumulate contiguous ranges of dirty cards, mark these cards 4894 // precleaned, then scan the region covered by these cards. 4895 HeapWord* endAddr = (HeapWord*)(gen->_virtual_space.high()); 4896 HeapWord* startAddr = (HeapWord*)(gen->_virtual_space.low()); 4897 4898 cl->setFreelistLock(gen->freelistLock()); // needed for yielding 4899 4900 size_t numDirtyCards, cumNumDirtyCards; 4901 HeapWord *lastAddr, *nextAddr; 4902 4903 for (cumNumDirtyCards = numDirtyCards = 0, 4904 nextAddr = lastAddr = startAddr; 4905 nextAddr < endAddr; 4906 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 4907 4908 ResourceMark rm; 4909 HandleMark hm; 4910 4911 MemRegion dirtyRegion; 4912 { 4913 // See comments in "Precleaning notes" above on why we 4914 // do this locking. XXX Could the locking overheads be 4915 // too high when dirty cards are sparse? [I don't think so.] 4916 stopTimer(); 4917 CMSTokenSync x(true); // is cms thread 4918 startTimer(); 4919 sample_eden(); 4920 // Get and clear dirty region from card table 4921 dirtyRegion = _ct->ct_bs()->dirty_card_range_after_reset( 4922 MemRegion(nextAddr, endAddr), 4923 true, 4924 CardTableModRefBS::precleaned_card_val()); 4925 4926 assert(dirtyRegion.start() >= nextAddr, 4927 "returned region inconsistent?"); 4928 } 4929 lastAddr = dirtyRegion.end(); 4930 numDirtyCards = 4931 dirtyRegion.word_size()/CardTableModRefBS::card_size_in_words; 4932 4933 if (!dirtyRegion.is_empty()) { 4934 stopTimer(); 4935 CMSTokenSyncWithLocks ts(true, gen->freelistLock(), bitMapLock()); 4936 startTimer(); 4937 sample_eden(); 4938 verify_work_stacks_empty(); 4939 verify_overflow_empty(); 4940 HeapWord* stop_point = 4941 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 4942 if (stop_point != NULL) { 4943 assert((_collectorState == AbortablePreclean && should_abort_preclean()), 4944 "Should only be AbortablePreclean."); 4945 _ct->ct_bs()->invalidate(MemRegion(stop_point, dirtyRegion.end())); 4946 if (should_abort_preclean()) { 4947 break; // out of preclean loop 4948 } else { 4949 // Compute the next address at which preclean should pick up. 4950 lastAddr = next_card_start_after_block(stop_point); 4951 } 4952 } 4953 } else { 4954 break; 4955 } 4956 } 4957 verify_work_stacks_empty(); 4958 verify_overflow_empty(); 4959 return cumNumDirtyCards; 4960 } 4961 4962 class PrecleanKlassClosure : public KlassClosure { 4963 CMKlassClosure _cm_klass_closure; 4964 public: 4965 PrecleanKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {} 4966 void do_klass(Klass* k) { 4967 if (k->has_accumulated_modified_oops()) { 4968 k->clear_accumulated_modified_oops(); 4969 4970 _cm_klass_closure.do_klass(k); 4971 } 4972 } 4973 }; 4974 4975 // The freelist lock is needed to prevent asserts, is it really needed? 4976 void CMSCollector::preclean_klasses(MarkRefsIntoAndScanClosure* cl, Mutex* freelistLock) { 4977 4978 cl->set_freelistLock(freelistLock); 4979 4980 CMSTokenSyncWithLocks ts(true, freelistLock, bitMapLock()); 4981 4982 // SSS: Add equivalent to ScanMarkedObjectsAgainCarefullyClosure::do_yield_check and should_abort_preclean? 4983 // SSS: We should probably check if precleaning should be aborted, at suitable intervals? 4984 PrecleanKlassClosure preclean_klass_closure(cl); 4985 ClassLoaderDataGraph::classes_do(&preclean_klass_closure); 4986 4987 verify_work_stacks_empty(); 4988 verify_overflow_empty(); 4989 } 4990 4991 void CMSCollector::checkpointRootsFinal(bool asynch, 4992 bool clear_all_soft_refs, bool init_mark_was_synchronous) { 4993 assert(_collectorState == FinalMarking, "incorrect state transition?"); 4994 check_correct_thread_executing(); 4995 // world is stopped at this checkpoint 4996 assert(SafepointSynchronize::is_at_safepoint(), 4997 "world should be stopped"); 4998 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 4999 5000 verify_work_stacks_empty(); 5001 verify_overflow_empty(); 5002 5003 SpecializationStats::clear(); 5004 if (PrintGCDetails) { 5005 gclog_or_tty->print("[YG occupancy: "SIZE_FORMAT" K ("SIZE_FORMAT" K)]", 5006 _young_gen->used() / K, 5007 _young_gen->capacity() / K); 5008 } 5009 if (asynch) { 5010 if (CMSScavengeBeforeRemark) { 5011 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5012 // Temporarily set flag to false, GCH->do_collection will 5013 // expect it to be false and set to true 5014 FlagSetting fl(gch->_is_gc_active, false); 5015 NOT_PRODUCT(GCTraceTime t("Scavenge-Before-Remark", 5016 PrintGCDetails && Verbose, true, _gc_timer_cm);) 5017 int level = _cmsGen->level() - 1; 5018 if (level >= 0) { 5019 gch->do_collection(true, // full (i.e. force, see below) 5020 false, // !clear_all_soft_refs 5021 0, // size 5022 false, // is_tlab 5023 level // max_level 5024 ); 5025 } 5026 } 5027 FreelistLocker x(this); 5028 MutexLockerEx y(bitMapLock(), 5029 Mutex::_no_safepoint_check_flag); 5030 assert(!init_mark_was_synchronous, "but that's impossible!"); 5031 checkpointRootsFinalWork(asynch, clear_all_soft_refs, false); 5032 } else { 5033 // already have all the locks 5034 checkpointRootsFinalWork(asynch, clear_all_soft_refs, 5035 init_mark_was_synchronous); 5036 } 5037 verify_work_stacks_empty(); 5038 verify_overflow_empty(); 5039 SpecializationStats::print(); 5040 } 5041 5042 void CMSCollector::checkpointRootsFinalWork(bool asynch, 5043 bool clear_all_soft_refs, bool init_mark_was_synchronous) { 5044 5045 NOT_PRODUCT(GCTraceTime tr("checkpointRootsFinalWork", PrintGCDetails, false, _gc_timer_cm);) 5046 5047 assert(haveFreelistLocks(), "must have free list locks"); 5048 assert_lock_strong(bitMapLock()); 5049 5050 if (UseAdaptiveSizePolicy) { 5051 size_policy()->checkpoint_roots_final_begin(); 5052 } 5053 5054 ResourceMark rm; 5055 HandleMark hm; 5056 5057 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5058 5059 if (should_unload_classes()) { 5060 CodeCache::gc_prologue(); 5061 } 5062 assert(haveFreelistLocks(), "must have free list locks"); 5063 assert_lock_strong(bitMapLock()); 5064 5065 if (!init_mark_was_synchronous) { 5066 // We might assume that we need not fill TLAB's when 5067 // CMSScavengeBeforeRemark is set, because we may have just done 5068 // a scavenge which would have filled all TLAB's -- and besides 5069 // Eden would be empty. This however may not always be the case -- 5070 // for instance although we asked for a scavenge, it may not have 5071 // happened because of a JNI critical section. We probably need 5072 // a policy for deciding whether we can in that case wait until 5073 // the critical section releases and then do the remark following 5074 // the scavenge, and skip it here. In the absence of that policy, 5075 // or of an indication of whether the scavenge did indeed occur, 5076 // we cannot rely on TLAB's having been filled and must do 5077 // so here just in case a scavenge did not happen. 5078 gch->ensure_parsability(false); // fill TLAB's, but no need to retire them 5079 // Update the saved marks which may affect the root scans. 5080 gch->save_marks(); 5081 5082 if (CMSPrintEdenSurvivorChunks) { 5083 print_eden_and_survivor_chunk_arrays(); 5084 } 5085 5086 { 5087 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;) 5088 5089 // Note on the role of the mod union table: 5090 // Since the marker in "markFromRoots" marks concurrently with 5091 // mutators, it is possible for some reachable objects not to have been 5092 // scanned. For instance, an only reference to an object A was 5093 // placed in object B after the marker scanned B. Unless B is rescanned, 5094 // A would be collected. Such updates to references in marked objects 5095 // are detected via the mod union table which is the set of all cards 5096 // dirtied since the first checkpoint in this GC cycle and prior to 5097 // the most recent young generation GC, minus those cleaned up by the 5098 // concurrent precleaning. 5099 if (CMSParallelRemarkEnabled && CollectedHeap::use_parallel_gc_threads()) { 5100 GCTraceTime t("Rescan (parallel) ", PrintGCDetails, false, _gc_timer_cm); 5101 do_remark_parallel(); 5102 } else { 5103 GCTraceTime t("Rescan (non-parallel) ", PrintGCDetails, false, 5104 _gc_timer_cm); 5105 do_remark_non_parallel(); 5106 } 5107 } 5108 } else { 5109 assert(!asynch, "Can't have init_mark_was_synchronous in asynch mode"); 5110 // The initial mark was stop-world, so there's no rescanning to 5111 // do; go straight on to the next step below. 5112 } 5113 verify_work_stacks_empty(); 5114 verify_overflow_empty(); 5115 5116 { 5117 NOT_PRODUCT(GCTraceTime ts("refProcessingWork", PrintGCDetails, false, _gc_timer_cm);) 5118 refProcessingWork(asynch, clear_all_soft_refs); 5119 } 5120 verify_work_stacks_empty(); 5121 verify_overflow_empty(); 5122 5123 if (should_unload_classes()) { 5124 CodeCache::gc_epilogue(); 5125 } 5126 JvmtiExport::gc_epilogue(); 5127 5128 // If we encountered any (marking stack / work queue) overflow 5129 // events during the current CMS cycle, take appropriate 5130 // remedial measures, where possible, so as to try and avoid 5131 // recurrence of that condition. 5132 assert(_markStack.isEmpty(), "No grey objects"); 5133 size_t ser_ovflw = _ser_pmc_remark_ovflw + _ser_pmc_preclean_ovflw + 5134 _ser_kac_ovflw + _ser_kac_preclean_ovflw; 5135 if (ser_ovflw > 0) { 5136 if (PrintCMSStatistics != 0) { 5137 gclog_or_tty->print_cr("Marking stack overflow (benign) " 5138 "(pmc_pc="SIZE_FORMAT", pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT 5139 ", kac_preclean="SIZE_FORMAT")", 5140 _ser_pmc_preclean_ovflw, _ser_pmc_remark_ovflw, 5141 _ser_kac_ovflw, _ser_kac_preclean_ovflw); 5142 } 5143 _markStack.expand(); 5144 _ser_pmc_remark_ovflw = 0; 5145 _ser_pmc_preclean_ovflw = 0; 5146 _ser_kac_preclean_ovflw = 0; 5147 _ser_kac_ovflw = 0; 5148 } 5149 if (_par_pmc_remark_ovflw > 0 || _par_kac_ovflw > 0) { 5150 if (PrintCMSStatistics != 0) { 5151 gclog_or_tty->print_cr("Work queue overflow (benign) " 5152 "(pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT")", 5153 _par_pmc_remark_ovflw, _par_kac_ovflw); 5154 } 5155 _par_pmc_remark_ovflw = 0; 5156 _par_kac_ovflw = 0; 5157 } 5158 if (PrintCMSStatistics != 0) { 5159 if (_markStack._hit_limit > 0) { 5160 gclog_or_tty->print_cr(" (benign) Hit max stack size limit ("SIZE_FORMAT")", 5161 _markStack._hit_limit); 5162 } 5163 if (_markStack._failed_double > 0) { 5164 gclog_or_tty->print_cr(" (benign) Failed stack doubling ("SIZE_FORMAT")," 5165 " current capacity "SIZE_FORMAT, 5166 _markStack._failed_double, 5167 _markStack.capacity()); 5168 } 5169 } 5170 _markStack._hit_limit = 0; 5171 _markStack._failed_double = 0; 5172 5173 if ((VerifyAfterGC || VerifyDuringGC) && 5174 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 5175 verify_after_remark(); 5176 } 5177 5178 _gc_tracer_cm->report_object_count_after_gc(&_is_alive_closure); 5179 5180 // Change under the freelistLocks. 5181 _collectorState = Sweeping; 5182 // Call isAllClear() under bitMapLock 5183 assert(_modUnionTable.isAllClear(), 5184 "Should be clear by end of the final marking"); 5185 assert(_ct->klass_rem_set()->mod_union_is_clear(), 5186 "Should be clear by end of the final marking"); 5187 if (UseAdaptiveSizePolicy) { 5188 size_policy()->checkpoint_roots_final_end(gch->gc_cause()); 5189 } 5190 } 5191 5192 void CMSParInitialMarkTask::work(uint worker_id) { 5193 elapsedTimer _timer; 5194 ResourceMark rm; 5195 HandleMark hm; 5196 5197 // ---------- scan from roots -------------- 5198 _timer.start(); 5199 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5200 Par_MarkRefsIntoClosure par_mri_cl(_collector->_span, &(_collector->_markBitMap)); 5201 CMKlassClosure klass_closure(&par_mri_cl); 5202 5203 // ---------- young gen roots -------------- 5204 { 5205 work_on_young_gen_roots(worker_id, &par_mri_cl); 5206 _timer.stop(); 5207 if (PrintCMSStatistics != 0) { 5208 gclog_or_tty->print_cr( 5209 "Finished young gen initial mark scan work in %dth thread: %3.3f sec", 5210 worker_id, _timer.seconds()); 5211 } 5212 } 5213 5214 // ---------- remaining roots -------------- 5215 _timer.reset(); 5216 _timer.start(); 5217 gch->gen_process_strong_roots(_collector->_cmsGen->level(), 5218 false, // yg was scanned above 5219 false, // this is parallel code 5220 SharedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()), 5221 &par_mri_cl, 5222 NULL, 5223 &klass_closure); 5224 assert(_collector->should_unload_classes() 5225 || (_collector->CMSCollector::roots_scanning_options() & SharedHeap::SO_AllCodeCache), 5226 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 5227 _timer.stop(); 5228 if (PrintCMSStatistics != 0) { 5229 gclog_or_tty->print_cr( 5230 "Finished remaining root initial mark scan work in %dth thread: %3.3f sec", 5231 worker_id, _timer.seconds()); 5232 } 5233 } 5234 5235 // Parallel remark task 5236 class CMSParRemarkTask: public CMSParMarkTask { 5237 CompactibleFreeListSpace* _cms_space; 5238 5239 // The per-thread work queues, available here for stealing. 5240 OopTaskQueueSet* _task_queues; 5241 ParallelTaskTerminator _term; 5242 5243 public: 5244 // A value of 0 passed to n_workers will cause the number of 5245 // workers to be taken from the active workers in the work gang. 5246 CMSParRemarkTask(CMSCollector* collector, 5247 CompactibleFreeListSpace* cms_space, 5248 int n_workers, FlexibleWorkGang* workers, 5249 OopTaskQueueSet* task_queues): 5250 CMSParMarkTask("Rescan roots and grey objects in parallel", 5251 collector, n_workers), 5252 _cms_space(cms_space), 5253 _task_queues(task_queues), 5254 _term(n_workers, task_queues) { } 5255 5256 OopTaskQueueSet* task_queues() { return _task_queues; } 5257 5258 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 5259 5260 ParallelTaskTerminator* terminator() { return &_term; } 5261 int n_workers() { return _n_workers; } 5262 5263 void work(uint worker_id); 5264 5265 private: 5266 // ... of dirty cards in old space 5267 void do_dirty_card_rescan_tasks(CompactibleFreeListSpace* sp, int i, 5268 Par_MarkRefsIntoAndScanClosure* cl); 5269 5270 // ... work stealing for the above 5271 void do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, int* seed); 5272 }; 5273 5274 class RemarkKlassClosure : public KlassClosure { 5275 CMKlassClosure _cm_klass_closure; 5276 public: 5277 RemarkKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {} 5278 void do_klass(Klass* k) { 5279 // Check if we have modified any oops in the Klass during the concurrent marking. 5280 if (k->has_accumulated_modified_oops()) { 5281 k->clear_accumulated_modified_oops(); 5282 5283 // We could have transfered the current modified marks to the accumulated marks, 5284 // like we do with the Card Table to Mod Union Table. But it's not really necessary. 5285 } else if (k->has_modified_oops()) { 5286 // Don't clear anything, this info is needed by the next young collection. 5287 } else { 5288 // No modified oops in the Klass. 5289 return; 5290 } 5291 5292 // The klass has modified fields, need to scan the klass. 5293 _cm_klass_closure.do_klass(k); 5294 } 5295 }; 5296 5297 void CMSParMarkTask::work_on_young_gen_roots(uint worker_id, OopsInGenClosure* cl) { 5298 DefNewGeneration* dng = _collector->_young_gen->as_DefNewGeneration(); 5299 EdenSpace* eden_space = dng->eden(); 5300 ContiguousSpace* from_space = dng->from(); 5301 ContiguousSpace* to_space = dng->to(); 5302 5303 HeapWord** eca = _collector->_eden_chunk_array; 5304 size_t ect = _collector->_eden_chunk_index; 5305 HeapWord** sca = _collector->_survivor_chunk_array; 5306 size_t sct = _collector->_survivor_chunk_index; 5307 5308 assert(ect <= _collector->_eden_chunk_capacity, "out of bounds"); 5309 assert(sct <= _collector->_survivor_chunk_capacity, "out of bounds"); 5310 5311 do_young_space_rescan(worker_id, cl, to_space, NULL, 0); 5312 do_young_space_rescan(worker_id, cl, from_space, sca, sct); 5313 do_young_space_rescan(worker_id, cl, eden_space, eca, ect); 5314 } 5315 5316 // work_queue(i) is passed to the closure 5317 // Par_MarkRefsIntoAndScanClosure. The "i" parameter 5318 // also is passed to do_dirty_card_rescan_tasks() and to 5319 // do_work_steal() to select the i-th task_queue. 5320 5321 void CMSParRemarkTask::work(uint worker_id) { 5322 elapsedTimer _timer; 5323 ResourceMark rm; 5324 HandleMark hm; 5325 5326 // ---------- rescan from roots -------------- 5327 _timer.start(); 5328 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5329 Par_MarkRefsIntoAndScanClosure par_mrias_cl(_collector, 5330 _collector->_span, _collector->ref_processor(), 5331 &(_collector->_markBitMap), 5332 work_queue(worker_id)); 5333 5334 // Rescan young gen roots first since these are likely 5335 // coarsely partitioned and may, on that account, constitute 5336 // the critical path; thus, it's best to start off that 5337 // work first. 5338 // ---------- young gen roots -------------- 5339 { 5340 work_on_young_gen_roots(worker_id, &par_mrias_cl); 5341 _timer.stop(); 5342 if (PrintCMSStatistics != 0) { 5343 gclog_or_tty->print_cr( 5344 "Finished young gen rescan work in %dth thread: %3.3f sec", 5345 worker_id, _timer.seconds()); 5346 } 5347 } 5348 5349 // ---------- remaining roots -------------- 5350 _timer.reset(); 5351 _timer.start(); 5352 gch->gen_process_strong_roots(_collector->_cmsGen->level(), 5353 false, // yg was scanned above 5354 false, // this is parallel code 5355 SharedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()), 5356 &par_mrias_cl, 5357 NULL, 5358 NULL); // The dirty klasses will be handled below 5359 assert(_collector->should_unload_classes() 5360 || (_collector->CMSCollector::roots_scanning_options() & SharedHeap::SO_AllCodeCache), 5361 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 5362 _timer.stop(); 5363 if (PrintCMSStatistics != 0) { 5364 gclog_or_tty->print_cr( 5365 "Finished remaining root rescan work in %dth thread: %3.3f sec", 5366 worker_id, _timer.seconds()); 5367 } 5368 5369 // ---------- unhandled CLD scanning ---------- 5370 if (worker_id == 0) { // Single threaded at the moment. 5371 _timer.reset(); 5372 _timer.start(); 5373 5374 // Scan all new class loader data objects and new dependencies that were 5375 // introduced during concurrent marking. 5376 ResourceMark rm; 5377 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds(); 5378 for (int i = 0; i < array->length(); i++) { 5379 par_mrias_cl.do_class_loader_data(array->at(i)); 5380 } 5381 5382 // We don't need to keep track of new CLDs anymore. 5383 ClassLoaderDataGraph::remember_new_clds(false); 5384 5385 _timer.stop(); 5386 if (PrintCMSStatistics != 0) { 5387 gclog_or_tty->print_cr( 5388 "Finished unhandled CLD scanning work in %dth thread: %3.3f sec", 5389 worker_id, _timer.seconds()); 5390 } 5391 } 5392 5393 // ---------- dirty klass scanning ---------- 5394 if (worker_id == 0) { // Single threaded at the moment. 5395 _timer.reset(); 5396 _timer.start(); 5397 5398 // Scan all classes that was dirtied during the concurrent marking phase. 5399 RemarkKlassClosure remark_klass_closure(&par_mrias_cl); 5400 ClassLoaderDataGraph::classes_do(&remark_klass_closure); 5401 5402 _timer.stop(); 5403 if (PrintCMSStatistics != 0) { 5404 gclog_or_tty->print_cr( 5405 "Finished dirty klass scanning work in %dth thread: %3.3f sec", 5406 worker_id, _timer.seconds()); 5407 } 5408 } 5409 5410 // We might have added oops to ClassLoaderData::_handles during the 5411 // concurrent marking phase. These oops point to newly allocated objects 5412 // that are guaranteed to be kept alive. Either by the direct allocation 5413 // code, or when the young collector processes the strong roots. Hence, 5414 // we don't have to revisit the _handles block during the remark phase. 5415 5416 // ---------- rescan dirty cards ------------ 5417 _timer.reset(); 5418 _timer.start(); 5419 5420 // Do the rescan tasks for each of the two spaces 5421 // (cms_space) in turn. 5422 // "worker_id" is passed to select the task_queue for "worker_id" 5423 do_dirty_card_rescan_tasks(_cms_space, worker_id, &par_mrias_cl); 5424 _timer.stop(); 5425 if (PrintCMSStatistics != 0) { 5426 gclog_or_tty->print_cr( 5427 "Finished dirty card rescan work in %dth thread: %3.3f sec", 5428 worker_id, _timer.seconds()); 5429 } 5430 5431 // ---------- steal work from other threads ... 5432 // ---------- ... and drain overflow list. 5433 _timer.reset(); 5434 _timer.start(); 5435 do_work_steal(worker_id, &par_mrias_cl, _collector->hash_seed(worker_id)); 5436 _timer.stop(); 5437 if (PrintCMSStatistics != 0) { 5438 gclog_or_tty->print_cr( 5439 "Finished work stealing in %dth thread: %3.3f sec", 5440 worker_id, _timer.seconds()); 5441 } 5442 } 5443 5444 // Note that parameter "i" is not used. 5445 void 5446 CMSParMarkTask::do_young_space_rescan(uint worker_id, 5447 OopsInGenClosure* cl, ContiguousSpace* space, 5448 HeapWord** chunk_array, size_t chunk_top) { 5449 // Until all tasks completed: 5450 // . claim an unclaimed task 5451 // . compute region boundaries corresponding to task claimed 5452 // using chunk_array 5453 // . par_oop_iterate(cl) over that region 5454 5455 ResourceMark rm; 5456 HandleMark hm; 5457 5458 SequentialSubTasksDone* pst = space->par_seq_tasks(); 5459 5460 uint nth_task = 0; 5461 uint n_tasks = pst->n_tasks(); 5462 5463 if (n_tasks > 0) { 5464 assert(pst->valid(), "Uninitialized use?"); 5465 HeapWord *start, *end; 5466 while (!pst->is_task_claimed(/* reference */ nth_task)) { 5467 // We claimed task # nth_task; compute its boundaries. 5468 if (chunk_top == 0) { // no samples were taken 5469 assert(nth_task == 0 && n_tasks == 1, "Can have only 1 EdenSpace task"); 5470 start = space->bottom(); 5471 end = space->top(); 5472 } else if (nth_task == 0) { 5473 start = space->bottom(); 5474 end = chunk_array[nth_task]; 5475 } else if (nth_task < (uint)chunk_top) { 5476 assert(nth_task >= 1, "Control point invariant"); 5477 start = chunk_array[nth_task - 1]; 5478 end = chunk_array[nth_task]; 5479 } else { 5480 assert(nth_task == (uint)chunk_top, "Control point invariant"); 5481 start = chunk_array[chunk_top - 1]; 5482 end = space->top(); 5483 } 5484 MemRegion mr(start, end); 5485 // Verify that mr is in space 5486 assert(mr.is_empty() || space->used_region().contains(mr), 5487 "Should be in space"); 5488 // Verify that "start" is an object boundary 5489 assert(mr.is_empty() || oop(mr.start())->is_oop(), 5490 "Should be an oop"); 5491 space->par_oop_iterate(mr, cl); 5492 } 5493 pst->all_tasks_completed(); 5494 } 5495 } 5496 5497 void 5498 CMSParRemarkTask::do_dirty_card_rescan_tasks( 5499 CompactibleFreeListSpace* sp, int i, 5500 Par_MarkRefsIntoAndScanClosure* cl) { 5501 // Until all tasks completed: 5502 // . claim an unclaimed task 5503 // . compute region boundaries corresponding to task claimed 5504 // . transfer dirty bits ct->mut for that region 5505 // . apply rescanclosure to dirty mut bits for that region 5506 5507 ResourceMark rm; 5508 HandleMark hm; 5509 5510 OopTaskQueue* work_q = work_queue(i); 5511 ModUnionClosure modUnionClosure(&(_collector->_modUnionTable)); 5512 // CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! 5513 // CAUTION: This closure has state that persists across calls to 5514 // the work method dirty_range_iterate_clear() in that it has 5515 // embedded in it a (subtype of) UpwardsObjectClosure. The 5516 // use of that state in the embedded UpwardsObjectClosure instance 5517 // assumes that the cards are always iterated (even if in parallel 5518 // by several threads) in monotonically increasing order per each 5519 // thread. This is true of the implementation below which picks 5520 // card ranges (chunks) in monotonically increasing order globally 5521 // and, a-fortiori, in monotonically increasing order per thread 5522 // (the latter order being a subsequence of the former). 5523 // If the work code below is ever reorganized into a more chaotic 5524 // work-partitioning form than the current "sequential tasks" 5525 // paradigm, the use of that persistent state will have to be 5526 // revisited and modified appropriately. See also related 5527 // bug 4756801 work on which should examine this code to make 5528 // sure that the changes there do not run counter to the 5529 // assumptions made here and necessary for correctness and 5530 // efficiency. Note also that this code might yield inefficient 5531 // behavior in the case of very large objects that span one or 5532 // more work chunks. Such objects would potentially be scanned 5533 // several times redundantly. Work on 4756801 should try and 5534 // address that performance anomaly if at all possible. XXX 5535 MemRegion full_span = _collector->_span; 5536 CMSBitMap* bm = &(_collector->_markBitMap); // shared 5537 MarkFromDirtyCardsClosure 5538 greyRescanClosure(_collector, full_span, // entire span of interest 5539 sp, bm, work_q, cl); 5540 5541 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 5542 assert(pst->valid(), "Uninitialized use?"); 5543 uint nth_task = 0; 5544 const int alignment = CardTableModRefBS::card_size * BitsPerWord; 5545 MemRegion span = sp->used_region(); 5546 HeapWord* start_addr = span.start(); 5547 HeapWord* end_addr = (HeapWord*)round_to((intptr_t)span.end(), 5548 alignment); 5549 const size_t chunk_size = sp->rescan_task_size(); // in HeapWord units 5550 assert((HeapWord*)round_to((intptr_t)start_addr, alignment) == 5551 start_addr, "Check alignment"); 5552 assert((size_t)round_to((intptr_t)chunk_size, alignment) == 5553 chunk_size, "Check alignment"); 5554 5555 while (!pst->is_task_claimed(/* reference */ nth_task)) { 5556 // Having claimed the nth_task, compute corresponding mem-region, 5557 // which is a-fortiori aligned correctly (i.e. at a MUT boundary). 5558 // The alignment restriction ensures that we do not need any 5559 // synchronization with other gang-workers while setting or 5560 // clearing bits in thus chunk of the MUT. 5561 MemRegion this_span = MemRegion(start_addr + nth_task*chunk_size, 5562 start_addr + (nth_task+1)*chunk_size); 5563 // The last chunk's end might be way beyond end of the 5564 // used region. In that case pull back appropriately. 5565 if (this_span.end() > end_addr) { 5566 this_span.set_end(end_addr); 5567 assert(!this_span.is_empty(), "Program logic (calculation of n_tasks)"); 5568 } 5569 // Iterate over the dirty cards covering this chunk, marking them 5570 // precleaned, and setting the corresponding bits in the mod union 5571 // table. Since we have been careful to partition at Card and MUT-word 5572 // boundaries no synchronization is needed between parallel threads. 5573 _collector->_ct->ct_bs()->dirty_card_iterate(this_span, 5574 &modUnionClosure); 5575 5576 // Having transferred these marks into the modUnionTable, 5577 // rescan the marked objects on the dirty cards in the modUnionTable. 5578 // Even if this is at a synchronous collection, the initial marking 5579 // may have been done during an asynchronous collection so there 5580 // may be dirty bits in the mod-union table. 5581 _collector->_modUnionTable.dirty_range_iterate_clear( 5582 this_span, &greyRescanClosure); 5583 _collector->_modUnionTable.verifyNoOneBitsInRange( 5584 this_span.start(), 5585 this_span.end()); 5586 } 5587 pst->all_tasks_completed(); // declare that i am done 5588 } 5589 5590 // . see if we can share work_queues with ParNew? XXX 5591 void 5592 CMSParRemarkTask::do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, 5593 int* seed) { 5594 OopTaskQueue* work_q = work_queue(i); 5595 NOT_PRODUCT(int num_steals = 0;) 5596 oop obj_to_scan; 5597 CMSBitMap* bm = &(_collector->_markBitMap); 5598 5599 while (true) { 5600 // Completely finish any left over work from (an) earlier round(s) 5601 cl->trim_queue(0); 5602 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 5603 (size_t)ParGCDesiredObjsFromOverflowList); 5604 // Now check if there's any work in the overflow list 5605 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads, 5606 // only affects the number of attempts made to get work from the 5607 // overflow list and does not affect the number of workers. Just 5608 // pass ParallelGCThreads so this behavior is unchanged. 5609 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 5610 work_q, 5611 ParallelGCThreads)) { 5612 // found something in global overflow list; 5613 // not yet ready to go stealing work from others. 5614 // We'd like to assert(work_q->size() != 0, ...) 5615 // because we just took work from the overflow list, 5616 // but of course we can't since all of that could have 5617 // been already stolen from us. 5618 // "He giveth and He taketh away." 5619 continue; 5620 } 5621 // Verify that we have no work before we resort to stealing 5622 assert(work_q->size() == 0, "Have work, shouldn't steal"); 5623 // Try to steal from other queues that have work 5624 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 5625 NOT_PRODUCT(num_steals++;) 5626 assert(obj_to_scan->is_oop(), "Oops, not an oop!"); 5627 assert(bm->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 5628 // Do scanning work 5629 obj_to_scan->oop_iterate(cl); 5630 // Loop around, finish this work, and try to steal some more 5631 } else if (terminator()->offer_termination()) { 5632 break; // nirvana from the infinite cycle 5633 } 5634 } 5635 NOT_PRODUCT( 5636 if (PrintCMSStatistics != 0) { 5637 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals); 5638 } 5639 ) 5640 assert(work_q->size() == 0 && _collector->overflow_list_is_empty(), 5641 "Else our work is not yet done"); 5642 } 5643 5644 // Record object boundaries in _eden_chunk_array by sampling the eden 5645 // top in the slow-path eden object allocation code path and record 5646 // the boundaries, if CMSEdenChunksRecordAlways is true. If 5647 // CMSEdenChunksRecordAlways is false, we use the other asynchronous 5648 // sampling in sample_eden() that activates during the part of the 5649 // preclean phase. 5650 void CMSCollector::sample_eden_chunk() { 5651 if (CMSEdenChunksRecordAlways && _eden_chunk_array != NULL) { 5652 if (_eden_chunk_lock->try_lock()) { 5653 // Record a sample. This is the critical section. The contents 5654 // of the _eden_chunk_array have to be non-decreasing in the 5655 // address order. 5656 _eden_chunk_array[_eden_chunk_index] = *_top_addr; 5657 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr, 5658 "Unexpected state of Eden"); 5659 if (_eden_chunk_index == 0 || 5660 ((_eden_chunk_array[_eden_chunk_index] > _eden_chunk_array[_eden_chunk_index-1]) && 5661 (pointer_delta(_eden_chunk_array[_eden_chunk_index], 5662 _eden_chunk_array[_eden_chunk_index-1]) >= CMSSamplingGrain))) { 5663 _eden_chunk_index++; // commit sample 5664 } 5665 _eden_chunk_lock->unlock(); 5666 } 5667 } 5668 } 5669 5670 // Return a thread-local PLAB recording array, as appropriate. 5671 void* CMSCollector::get_data_recorder(int thr_num) { 5672 if (_survivor_plab_array != NULL && 5673 (CMSPLABRecordAlways || 5674 (_collectorState > Marking && _collectorState < FinalMarking))) { 5675 assert(thr_num < (int)ParallelGCThreads, "thr_num is out of bounds"); 5676 ChunkArray* ca = &_survivor_plab_array[thr_num]; 5677 ca->reset(); // clear it so that fresh data is recorded 5678 return (void*) ca; 5679 } else { 5680 return NULL; 5681 } 5682 } 5683 5684 // Reset all the thread-local PLAB recording arrays 5685 void CMSCollector::reset_survivor_plab_arrays() { 5686 for (uint i = 0; i < ParallelGCThreads; i++) { 5687 _survivor_plab_array[i].reset(); 5688 } 5689 } 5690 5691 // Merge the per-thread plab arrays into the global survivor chunk 5692 // array which will provide the partitioning of the survivor space 5693 // for CMS initial scan and rescan. 5694 void CMSCollector::merge_survivor_plab_arrays(ContiguousSpace* surv, 5695 int no_of_gc_threads) { 5696 assert(_survivor_plab_array != NULL, "Error"); 5697 assert(_survivor_chunk_array != NULL, "Error"); 5698 assert(_collectorState == FinalMarking || 5699 (CMSParallelInitialMarkEnabled && _collectorState == InitialMarking), "Error"); 5700 for (int j = 0; j < no_of_gc_threads; j++) { 5701 _cursor[j] = 0; 5702 } 5703 HeapWord* top = surv->top(); 5704 size_t i; 5705 for (i = 0; i < _survivor_chunk_capacity; i++) { // all sca entries 5706 HeapWord* min_val = top; // Higher than any PLAB address 5707 uint min_tid = 0; // position of min_val this round 5708 for (int j = 0; j < no_of_gc_threads; j++) { 5709 ChunkArray* cur_sca = &_survivor_plab_array[j]; 5710 if (_cursor[j] == cur_sca->end()) { 5711 continue; 5712 } 5713 assert(_cursor[j] < cur_sca->end(), "ctl pt invariant"); 5714 HeapWord* cur_val = cur_sca->nth(_cursor[j]); 5715 assert(surv->used_region().contains(cur_val), "Out of bounds value"); 5716 if (cur_val < min_val) { 5717 min_tid = j; 5718 min_val = cur_val; 5719 } else { 5720 assert(cur_val < top, "All recorded addresses should be less"); 5721 } 5722 } 5723 // At this point min_val and min_tid are respectively 5724 // the least address in _survivor_plab_array[j]->nth(_cursor[j]) 5725 // and the thread (j) that witnesses that address. 5726 // We record this address in the _survivor_chunk_array[i] 5727 // and increment _cursor[min_tid] prior to the next round i. 5728 if (min_val == top) { 5729 break; 5730 } 5731 _survivor_chunk_array[i] = min_val; 5732 _cursor[min_tid]++; 5733 } 5734 // We are all done; record the size of the _survivor_chunk_array 5735 _survivor_chunk_index = i; // exclusive: [0, i) 5736 if (PrintCMSStatistics > 0) { 5737 gclog_or_tty->print(" (Survivor:" SIZE_FORMAT "chunks) ", i); 5738 } 5739 // Verify that we used up all the recorded entries 5740 #ifdef ASSERT 5741 size_t total = 0; 5742 for (int j = 0; j < no_of_gc_threads; j++) { 5743 assert(_cursor[j] == _survivor_plab_array[j].end(), "Ctl pt invariant"); 5744 total += _cursor[j]; 5745 } 5746 assert(total == _survivor_chunk_index, "Ctl Pt Invariant"); 5747 // Check that the merged array is in sorted order 5748 if (total > 0) { 5749 for (size_t i = 0; i < total - 1; i++) { 5750 if (PrintCMSStatistics > 0) { 5751 gclog_or_tty->print(" (chunk" SIZE_FORMAT ":" INTPTR_FORMAT ") ", 5752 i, _survivor_chunk_array[i]); 5753 } 5754 assert(_survivor_chunk_array[i] < _survivor_chunk_array[i+1], 5755 "Not sorted"); 5756 } 5757 } 5758 #endif // ASSERT 5759 } 5760 5761 // Set up the space's par_seq_tasks structure for work claiming 5762 // for parallel initial scan and rescan of young gen. 5763 // See ParRescanTask where this is currently used. 5764 void 5765 CMSCollector:: 5766 initialize_sequential_subtasks_for_young_gen_rescan(int n_threads) { 5767 assert(n_threads > 0, "Unexpected n_threads argument"); 5768 DefNewGeneration* dng = (DefNewGeneration*)_young_gen; 5769 5770 // Eden space 5771 if (!dng->eden()->is_empty()) { 5772 SequentialSubTasksDone* pst = dng->eden()->par_seq_tasks(); 5773 assert(!pst->valid(), "Clobbering existing data?"); 5774 // Each valid entry in [0, _eden_chunk_index) represents a task. 5775 size_t n_tasks = _eden_chunk_index + 1; 5776 assert(n_tasks == 1 || _eden_chunk_array != NULL, "Error"); 5777 // Sets the condition for completion of the subtask (how many threads 5778 // need to finish in order to be done). 5779 pst->set_n_threads(n_threads); 5780 pst->set_n_tasks((int)n_tasks); 5781 } 5782 5783 // Merge the survivor plab arrays into _survivor_chunk_array 5784 if (_survivor_plab_array != NULL) { 5785 merge_survivor_plab_arrays(dng->from(), n_threads); 5786 } else { 5787 assert(_survivor_chunk_index == 0, "Error"); 5788 } 5789 5790 // To space 5791 { 5792 SequentialSubTasksDone* pst = dng->to()->par_seq_tasks(); 5793 assert(!pst->valid(), "Clobbering existing data?"); 5794 // Sets the condition for completion of the subtask (how many threads 5795 // need to finish in order to be done). 5796 pst->set_n_threads(n_threads); 5797 pst->set_n_tasks(1); 5798 assert(pst->valid(), "Error"); 5799 } 5800 5801 // From space 5802 { 5803 SequentialSubTasksDone* pst = dng->from()->par_seq_tasks(); 5804 assert(!pst->valid(), "Clobbering existing data?"); 5805 size_t n_tasks = _survivor_chunk_index + 1; 5806 assert(n_tasks == 1 || _survivor_chunk_array != NULL, "Error"); 5807 // Sets the condition for completion of the subtask (how many threads 5808 // need to finish in order to be done). 5809 pst->set_n_threads(n_threads); 5810 pst->set_n_tasks((int)n_tasks); 5811 assert(pst->valid(), "Error"); 5812 } 5813 } 5814 5815 // Parallel version of remark 5816 void CMSCollector::do_remark_parallel() { 5817 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5818 FlexibleWorkGang* workers = gch->workers(); 5819 assert(workers != NULL, "Need parallel worker threads."); 5820 // Choose to use the number of GC workers most recently set 5821 // into "active_workers". If active_workers is not set, set it 5822 // to ParallelGCThreads. 5823 int n_workers = workers->active_workers(); 5824 if (n_workers == 0) { 5825 assert(n_workers > 0, "Should have been set during scavenge"); 5826 n_workers = ParallelGCThreads; 5827 workers->set_active_workers(n_workers); 5828 } 5829 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 5830 5831 CMSParRemarkTask tsk(this, 5832 cms_space, 5833 n_workers, workers, task_queues()); 5834 5835 // Set up for parallel process_strong_roots work. 5836 gch->set_par_threads(n_workers); 5837 // We won't be iterating over the cards in the card table updating 5838 // the younger_gen cards, so we shouldn't call the following else 5839 // the verification code as well as subsequent younger_refs_iterate 5840 // code would get confused. XXX 5841 // gch->rem_set()->prepare_for_younger_refs_iterate(true); // parallel 5842 5843 // The young gen rescan work will not be done as part of 5844 // process_strong_roots (which currently doesn't knw how to 5845 // parallelize such a scan), but rather will be broken up into 5846 // a set of parallel tasks (via the sampling that the [abortable] 5847 // preclean phase did of EdenSpace, plus the [two] tasks of 5848 // scanning the [two] survivor spaces. Further fine-grain 5849 // parallelization of the scanning of the survivor spaces 5850 // themselves, and of precleaning of the younger gen itself 5851 // is deferred to the future. 5852 initialize_sequential_subtasks_for_young_gen_rescan(n_workers); 5853 5854 // The dirty card rescan work is broken up into a "sequence" 5855 // of parallel tasks (per constituent space) that are dynamically 5856 // claimed by the parallel threads. 5857 cms_space->initialize_sequential_subtasks_for_rescan(n_workers); 5858 5859 // It turns out that even when we're using 1 thread, doing the work in a 5860 // separate thread causes wide variance in run times. We can't help this 5861 // in the multi-threaded case, but we special-case n=1 here to get 5862 // repeatable measurements of the 1-thread overhead of the parallel code. 5863 if (n_workers > 1) { 5864 // Make refs discovery MT-safe, if it isn't already: it may not 5865 // necessarily be so, since it's possible that we are doing 5866 // ST marking. 5867 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), true); 5868 GenCollectedHeap::StrongRootsScope srs(gch); 5869 workers->run_task(&tsk); 5870 } else { 5871 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false); 5872 GenCollectedHeap::StrongRootsScope srs(gch); 5873 tsk.work(0); 5874 } 5875 5876 gch->set_par_threads(0); // 0 ==> non-parallel. 5877 // restore, single-threaded for now, any preserved marks 5878 // as a result of work_q overflow 5879 restore_preserved_marks_if_any(); 5880 } 5881 5882 // Non-parallel version of remark 5883 void CMSCollector::do_remark_non_parallel() { 5884 ResourceMark rm; 5885 HandleMark hm; 5886 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5887 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false); 5888 5889 MarkRefsIntoAndScanClosure 5890 mrias_cl(_span, ref_processor(), &_markBitMap, NULL /* not precleaning */, 5891 &_markStack, this, 5892 false /* should_yield */, false /* not precleaning */); 5893 MarkFromDirtyCardsClosure 5894 markFromDirtyCardsClosure(this, _span, 5895 NULL, // space is set further below 5896 &_markBitMap, &_markStack, &mrias_cl); 5897 { 5898 GCTraceTime t("grey object rescan", PrintGCDetails, false, _gc_timer_cm); 5899 // Iterate over the dirty cards, setting the corresponding bits in the 5900 // mod union table. 5901 { 5902 ModUnionClosure modUnionClosure(&_modUnionTable); 5903 _ct->ct_bs()->dirty_card_iterate( 5904 _cmsGen->used_region(), 5905 &modUnionClosure); 5906 } 5907 // Having transferred these marks into the modUnionTable, we just need 5908 // to rescan the marked objects on the dirty cards in the modUnionTable. 5909 // The initial marking may have been done during an asynchronous 5910 // collection so there may be dirty bits in the mod-union table. 5911 const int alignment = 5912 CardTableModRefBS::card_size * BitsPerWord; 5913 { 5914 // ... First handle dirty cards in CMS gen 5915 markFromDirtyCardsClosure.set_space(_cmsGen->cmsSpace()); 5916 MemRegion ur = _cmsGen->used_region(); 5917 HeapWord* lb = ur.start(); 5918 HeapWord* ub = (HeapWord*)round_to((intptr_t)ur.end(), alignment); 5919 MemRegion cms_span(lb, ub); 5920 _modUnionTable.dirty_range_iterate_clear(cms_span, 5921 &markFromDirtyCardsClosure); 5922 verify_work_stacks_empty(); 5923 if (PrintCMSStatistics != 0) { 5924 gclog_or_tty->print(" (re-scanned "SIZE_FORMAT" dirty cards in cms gen) ", 5925 markFromDirtyCardsClosure.num_dirty_cards()); 5926 } 5927 } 5928 } 5929 if (VerifyDuringGC && 5930 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 5931 HandleMark hm; // Discard invalid handles created during verification 5932 Universe::verify(); 5933 } 5934 { 5935 GCTraceTime t("root rescan", PrintGCDetails, false, _gc_timer_cm); 5936 5937 verify_work_stacks_empty(); 5938 5939 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 5940 GenCollectedHeap::StrongRootsScope srs(gch); 5941 gch->gen_process_strong_roots(_cmsGen->level(), 5942 true, // younger gens as roots 5943 false, // use the local StrongRootsScope 5944 SharedHeap::ScanningOption(roots_scanning_options()), 5945 &mrias_cl, 5946 NULL, 5947 NULL); // The dirty klasses will be handled below 5948 5949 assert(should_unload_classes() 5950 || (roots_scanning_options() & SharedHeap::SO_AllCodeCache), 5951 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 5952 } 5953 5954 { 5955 GCTraceTime t("visit unhandled CLDs", PrintGCDetails, false, _gc_timer_cm); 5956 5957 verify_work_stacks_empty(); 5958 5959 // Scan all class loader data objects that might have been introduced 5960 // during concurrent marking. 5961 ResourceMark rm; 5962 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds(); 5963 for (int i = 0; i < array->length(); i++) { 5964 mrias_cl.do_class_loader_data(array->at(i)); 5965 } 5966 5967 // We don't need to keep track of new CLDs anymore. 5968 ClassLoaderDataGraph::remember_new_clds(false); 5969 5970 verify_work_stacks_empty(); 5971 } 5972 5973 { 5974 GCTraceTime t("dirty klass scan", PrintGCDetails, false, _gc_timer_cm); 5975 5976 verify_work_stacks_empty(); 5977 5978 RemarkKlassClosure remark_klass_closure(&mrias_cl); 5979 ClassLoaderDataGraph::classes_do(&remark_klass_closure); 5980 5981 verify_work_stacks_empty(); 5982 } 5983 5984 // We might have added oops to ClassLoaderData::_handles during the 5985 // concurrent marking phase. These oops point to newly allocated objects 5986 // that are guaranteed to be kept alive. Either by the direct allocation 5987 // code, or when the young collector processes the strong roots. Hence, 5988 // we don't have to revisit the _handles block during the remark phase. 5989 5990 verify_work_stacks_empty(); 5991 // Restore evacuated mark words, if any, used for overflow list links 5992 if (!CMSOverflowEarlyRestoration) { 5993 restore_preserved_marks_if_any(); 5994 } 5995 verify_overflow_empty(); 5996 } 5997 5998 //////////////////////////////////////////////////////// 5999 // Parallel Reference Processing Task Proxy Class 6000 //////////////////////////////////////////////////////// 6001 class CMSRefProcTaskProxy: public AbstractGangTaskWOopQueues { 6002 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask; 6003 CMSCollector* _collector; 6004 CMSBitMap* _mark_bit_map; 6005 const MemRegion _span; 6006 ProcessTask& _task; 6007 6008 public: 6009 CMSRefProcTaskProxy(ProcessTask& task, 6010 CMSCollector* collector, 6011 const MemRegion& span, 6012 CMSBitMap* mark_bit_map, 6013 AbstractWorkGang* workers, 6014 OopTaskQueueSet* task_queues): 6015 // XXX Should superclass AGTWOQ also know about AWG since it knows 6016 // about the task_queues used by the AWG? Then it could initialize 6017 // the terminator() object. See 6984287. The set_for_termination() 6018 // below is a temporary band-aid for the regression in 6984287. 6019 AbstractGangTaskWOopQueues("Process referents by policy in parallel", 6020 task_queues), 6021 _task(task), 6022 _collector(collector), _span(span), _mark_bit_map(mark_bit_map) 6023 { 6024 assert(_collector->_span.equals(_span) && !_span.is_empty(), 6025 "Inconsistency in _span"); 6026 set_for_termination(workers->active_workers()); 6027 } 6028 6029 OopTaskQueueSet* task_queues() { return queues(); } 6030 6031 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 6032 6033 void do_work_steal(int i, 6034 CMSParDrainMarkingStackClosure* drain, 6035 CMSParKeepAliveClosure* keep_alive, 6036 int* seed); 6037 6038 virtual void work(uint worker_id); 6039 }; 6040 6041 void CMSRefProcTaskProxy::work(uint worker_id) { 6042 assert(_collector->_span.equals(_span), "Inconsistency in _span"); 6043 CMSParKeepAliveClosure par_keep_alive(_collector, _span, 6044 _mark_bit_map, 6045 work_queue(worker_id)); 6046 CMSParDrainMarkingStackClosure par_drain_stack(_collector, _span, 6047 _mark_bit_map, 6048 work_queue(worker_id)); 6049 CMSIsAliveClosure is_alive_closure(_span, _mark_bit_map); 6050 _task.work(worker_id, is_alive_closure, par_keep_alive, par_drain_stack); 6051 if (_task.marks_oops_alive()) { 6052 do_work_steal(worker_id, &par_drain_stack, &par_keep_alive, 6053 _collector->hash_seed(worker_id)); 6054 } 6055 assert(work_queue(worker_id)->size() == 0, "work_queue should be empty"); 6056 assert(_collector->_overflow_list == NULL, "non-empty _overflow_list"); 6057 } 6058 6059 class CMSRefEnqueueTaskProxy: public AbstractGangTask { 6060 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask; 6061 EnqueueTask& _task; 6062 6063 public: 6064 CMSRefEnqueueTaskProxy(EnqueueTask& task) 6065 : AbstractGangTask("Enqueue reference objects in parallel"), 6066 _task(task) 6067 { } 6068 6069 virtual void work(uint worker_id) 6070 { 6071 _task.work(worker_id); 6072 } 6073 }; 6074 6075 CMSParKeepAliveClosure::CMSParKeepAliveClosure(CMSCollector* collector, 6076 MemRegion span, CMSBitMap* bit_map, OopTaskQueue* work_queue): 6077 _span(span), 6078 _bit_map(bit_map), 6079 _work_queue(work_queue), 6080 _mark_and_push(collector, span, bit_map, work_queue), 6081 _low_water_mark(MIN2((uint)(work_queue->max_elems()/4), 6082 (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads))) 6083 { } 6084 6085 // . see if we can share work_queues with ParNew? XXX 6086 void CMSRefProcTaskProxy::do_work_steal(int i, 6087 CMSParDrainMarkingStackClosure* drain, 6088 CMSParKeepAliveClosure* keep_alive, 6089 int* seed) { 6090 OopTaskQueue* work_q = work_queue(i); 6091 NOT_PRODUCT(int num_steals = 0;) 6092 oop obj_to_scan; 6093 6094 while (true) { 6095 // Completely finish any left over work from (an) earlier round(s) 6096 drain->trim_queue(0); 6097 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 6098 (size_t)ParGCDesiredObjsFromOverflowList); 6099 // Now check if there's any work in the overflow list 6100 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads, 6101 // only affects the number of attempts made to get work from the 6102 // overflow list and does not affect the number of workers. Just 6103 // pass ParallelGCThreads so this behavior is unchanged. 6104 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 6105 work_q, 6106 ParallelGCThreads)) { 6107 // Found something in global overflow list; 6108 // not yet ready to go stealing work from others. 6109 // We'd like to assert(work_q->size() != 0, ...) 6110 // because we just took work from the overflow list, 6111 // but of course we can't, since all of that might have 6112 // been already stolen from us. 6113 continue; 6114 } 6115 // Verify that we have no work before we resort to stealing 6116 assert(work_q->size() == 0, "Have work, shouldn't steal"); 6117 // Try to steal from other queues that have work 6118 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 6119 NOT_PRODUCT(num_steals++;) 6120 assert(obj_to_scan->is_oop(), "Oops, not an oop!"); 6121 assert(_mark_bit_map->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 6122 // Do scanning work 6123 obj_to_scan->oop_iterate(keep_alive); 6124 // Loop around, finish this work, and try to steal some more 6125 } else if (terminator()->offer_termination()) { 6126 break; // nirvana from the infinite cycle 6127 } 6128 } 6129 NOT_PRODUCT( 6130 if (PrintCMSStatistics != 0) { 6131 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals); 6132 } 6133 ) 6134 } 6135 6136 void CMSRefProcTaskExecutor::execute(ProcessTask& task) 6137 { 6138 GenCollectedHeap* gch = GenCollectedHeap::heap(); 6139 FlexibleWorkGang* workers = gch->workers(); 6140 assert(workers != NULL, "Need parallel worker threads."); 6141 CMSRefProcTaskProxy rp_task(task, &_collector, 6142 _collector.ref_processor()->span(), 6143 _collector.markBitMap(), 6144 workers, _collector.task_queues()); 6145 workers->run_task(&rp_task); 6146 } 6147 6148 void CMSRefProcTaskExecutor::execute(EnqueueTask& task) 6149 { 6150 6151 GenCollectedHeap* gch = GenCollectedHeap::heap(); 6152 FlexibleWorkGang* workers = gch->workers(); 6153 assert(workers != NULL, "Need parallel worker threads."); 6154 CMSRefEnqueueTaskProxy enq_task(task); 6155 workers->run_task(&enq_task); 6156 } 6157 6158 void CMSCollector::refProcessingWork(bool asynch, bool clear_all_soft_refs) { 6159 6160 ResourceMark rm; 6161 HandleMark hm; 6162 6163 ReferenceProcessor* rp = ref_processor(); 6164 assert(rp->span().equals(_span), "Spans should be equal"); 6165 assert(!rp->enqueuing_is_done(), "Enqueuing should not be complete"); 6166 // Process weak references. 6167 rp->setup_policy(clear_all_soft_refs); 6168 verify_work_stacks_empty(); 6169 6170 CMSKeepAliveClosure cmsKeepAliveClosure(this, _span, &_markBitMap, 6171 &_markStack, false /* !preclean */); 6172 CMSDrainMarkingStackClosure cmsDrainMarkingStackClosure(this, 6173 _span, &_markBitMap, &_markStack, 6174 &cmsKeepAliveClosure, false /* !preclean */); 6175 { 6176 GCTraceTime t("weak refs processing", PrintGCDetails, false, _gc_timer_cm); 6177 6178 ReferenceProcessorStats stats; 6179 if (rp->processing_is_mt()) { 6180 // Set the degree of MT here. If the discovery is done MT, there 6181 // may have been a different number of threads doing the discovery 6182 // and a different number of discovered lists may have Ref objects. 6183 // That is OK as long as the Reference lists are balanced (see 6184 // balance_all_queues() and balance_queues()). 6185 GenCollectedHeap* gch = GenCollectedHeap::heap(); 6186 int active_workers = ParallelGCThreads; 6187 FlexibleWorkGang* workers = gch->workers(); 6188 if (workers != NULL) { 6189 active_workers = workers->active_workers(); 6190 // The expectation is that active_workers will have already 6191 // been set to a reasonable value. If it has not been set, 6192 // investigate. 6193 assert(active_workers > 0, "Should have been set during scavenge"); 6194 } 6195 rp->set_active_mt_degree(active_workers); 6196 CMSRefProcTaskExecutor task_executor(*this); 6197 stats = rp->process_discovered_references(&_is_alive_closure, 6198 &cmsKeepAliveClosure, 6199 &cmsDrainMarkingStackClosure, 6200 &task_executor, 6201 _gc_timer_cm); 6202 } else { 6203 stats = rp->process_discovered_references(&_is_alive_closure, 6204 &cmsKeepAliveClosure, 6205 &cmsDrainMarkingStackClosure, 6206 NULL, 6207 _gc_timer_cm); 6208 } 6209 _gc_tracer_cm->report_gc_reference_stats(stats); 6210 6211 } 6212 6213 // This is the point where the entire marking should have completed. 6214 verify_work_stacks_empty(); 6215 6216 if (should_unload_classes()) { 6217 { 6218 GCTraceTime t("class unloading", PrintGCDetails, false, _gc_timer_cm); 6219 6220 // Unload classes and purge the SystemDictionary. 6221 bool purged_class = SystemDictionary::do_unloading(&_is_alive_closure); 6222 6223 // Unload nmethods. 6224 CodeCache::do_unloading(&_is_alive_closure, purged_class); 6225 6226 // Prune dead klasses from subklass/sibling/implementor lists. 6227 Klass::clean_weak_klass_links(&_is_alive_closure); 6228 } 6229 6230 { 6231 GCTraceTime t("scrub symbol table", PrintGCDetails, false, _gc_timer_cm); 6232 // Clean up unreferenced symbols in symbol table. 6233 SymbolTable::unlink(); 6234 } 6235 } 6236 6237 // CMS doesn't use the StringTable as hard roots when class unloading is turned off. 6238 // Need to check if we really scanned the StringTable. 6239 if ((roots_scanning_options() & SharedHeap::SO_Strings) == 0) { 6240 GCTraceTime t("scrub string table", PrintGCDetails, false, _gc_timer_cm); 6241 // Delete entries for dead interned strings. 6242 StringTable::unlink(&_is_alive_closure); 6243 } 6244 6245 // Restore any preserved marks as a result of mark stack or 6246 // work queue overflow 6247 restore_preserved_marks_if_any(); // done single-threaded for now 6248 6249 rp->set_enqueuing_is_done(true); 6250 if (rp->processing_is_mt()) { 6251 rp->balance_all_queues(); 6252 CMSRefProcTaskExecutor task_executor(*this); 6253 rp->enqueue_discovered_references(&task_executor); 6254 } else { 6255 rp->enqueue_discovered_references(NULL); 6256 } 6257 rp->verify_no_references_recorded(); 6258 assert(!rp->discovery_enabled(), "should have been disabled"); 6259 } 6260 6261 #ifndef PRODUCT 6262 void CMSCollector::check_correct_thread_executing() { 6263 Thread* t = Thread::current(); 6264 // Only the VM thread or the CMS thread should be here. 6265 assert(t->is_ConcurrentGC_thread() || t->is_VM_thread(), 6266 "Unexpected thread type"); 6267 // If this is the vm thread, the foreground process 6268 // should not be waiting. Note that _foregroundGCIsActive is 6269 // true while the foreground collector is waiting. 6270 if (_foregroundGCShouldWait) { 6271 // We cannot be the VM thread 6272 assert(t->is_ConcurrentGC_thread(), 6273 "Should be CMS thread"); 6274 } else { 6275 // We can be the CMS thread only if we are in a stop-world 6276 // phase of CMS collection. 6277 if (t->is_ConcurrentGC_thread()) { 6278 assert(_collectorState == InitialMarking || 6279 _collectorState == FinalMarking, 6280 "Should be a stop-world phase"); 6281 // The CMS thread should be holding the CMS_token. 6282 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6283 "Potential interference with concurrently " 6284 "executing VM thread"); 6285 } 6286 } 6287 } 6288 #endif 6289 6290 void CMSCollector::sweep(bool asynch) { 6291 assert(_collectorState == Sweeping, "just checking"); 6292 check_correct_thread_executing(); 6293 verify_work_stacks_empty(); 6294 verify_overflow_empty(); 6295 increment_sweep_count(); 6296 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 6297 6298 _inter_sweep_timer.stop(); 6299 _inter_sweep_estimate.sample(_inter_sweep_timer.seconds()); 6300 size_policy()->avg_cms_free_at_sweep()->sample(_cmsGen->free()); 6301 6302 assert(!_intra_sweep_timer.is_active(), "Should not be active"); 6303 _intra_sweep_timer.reset(); 6304 _intra_sweep_timer.start(); 6305 if (asynch) { 6306 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 6307 CMSPhaseAccounting pa(this, "sweep", !PrintGCDetails); 6308 // First sweep the old gen 6309 { 6310 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(), 6311 bitMapLock()); 6312 sweepWork(_cmsGen, asynch); 6313 } 6314 6315 // Update Universe::_heap_*_at_gc figures. 6316 // We need all the free list locks to make the abstract state 6317 // transition from Sweeping to Resetting. See detailed note 6318 // further below. 6319 { 6320 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock()); 6321 // Update heap occupancy information which is used as 6322 // input to soft ref clearing policy at the next gc. 6323 Universe::update_heap_info_at_gc(); 6324 _collectorState = Resizing; 6325 } 6326 } else { 6327 // already have needed locks 6328 sweepWork(_cmsGen, asynch); 6329 // Update heap occupancy information which is used as 6330 // input to soft ref clearing policy at the next gc. 6331 Universe::update_heap_info_at_gc(); 6332 _collectorState = Resizing; 6333 } 6334 verify_work_stacks_empty(); 6335 verify_overflow_empty(); 6336 6337 if (should_unload_classes()) { 6338 ClassLoaderDataGraph::purge(); 6339 } 6340 6341 _intra_sweep_timer.stop(); 6342 _intra_sweep_estimate.sample(_intra_sweep_timer.seconds()); 6343 6344 _inter_sweep_timer.reset(); 6345 _inter_sweep_timer.start(); 6346 6347 // We need to use a monotonically non-decreasing time in ms 6348 // or we will see time-warp warnings and os::javaTimeMillis() 6349 // does not guarantee monotonicity. 6350 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC; 6351 update_time_of_last_gc(now); 6352 6353 // NOTE on abstract state transitions: 6354 // Mutators allocate-live and/or mark the mod-union table dirty 6355 // based on the state of the collection. The former is done in 6356 // the interval [Marking, Sweeping] and the latter in the interval 6357 // [Marking, Sweeping). Thus the transitions into the Marking state 6358 // and out of the Sweeping state must be synchronously visible 6359 // globally to the mutators. 6360 // The transition into the Marking state happens with the world 6361 // stopped so the mutators will globally see it. Sweeping is 6362 // done asynchronously by the background collector so the transition 6363 // from the Sweeping state to the Resizing state must be done 6364 // under the freelistLock (as is the check for whether to 6365 // allocate-live and whether to dirty the mod-union table). 6366 assert(_collectorState == Resizing, "Change of collector state to" 6367 " Resizing must be done under the freelistLocks (plural)"); 6368 6369 // Now that sweeping has been completed, we clear 6370 // the incremental_collection_failed flag, 6371 // thus inviting a younger gen collection to promote into 6372 // this generation. If such a promotion may still fail, 6373 // the flag will be set again when a young collection is 6374 // attempted. 6375 GenCollectedHeap* gch = GenCollectedHeap::heap(); 6376 gch->clear_incremental_collection_failed(); // Worth retrying as fresh space may have been freed up 6377 gch->update_full_collections_completed(_collection_count_start); 6378 } 6379 6380 // FIX ME!!! Looks like this belongs in CFLSpace, with 6381 // CMSGen merely delegating to it. 6382 void ConcurrentMarkSweepGeneration::setNearLargestChunk() { 6383 double nearLargestPercent = FLSLargestBlockCoalesceProximity; 6384 HeapWord* minAddr = _cmsSpace->bottom(); 6385 HeapWord* largestAddr = 6386 (HeapWord*) _cmsSpace->dictionary()->find_largest_dict(); 6387 if (largestAddr == NULL) { 6388 // The dictionary appears to be empty. In this case 6389 // try to coalesce at the end of the heap. 6390 largestAddr = _cmsSpace->end(); 6391 } 6392 size_t largestOffset = pointer_delta(largestAddr, minAddr); 6393 size_t nearLargestOffset = 6394 (size_t)((double)largestOffset * nearLargestPercent) - MinChunkSize; 6395 if (PrintFLSStatistics != 0) { 6396 gclog_or_tty->print_cr( 6397 "CMS: Large Block: " PTR_FORMAT ";" 6398 " Proximity: " PTR_FORMAT " -> " PTR_FORMAT, 6399 largestAddr, 6400 _cmsSpace->nearLargestChunk(), minAddr + nearLargestOffset); 6401 } 6402 _cmsSpace->set_nearLargestChunk(minAddr + nearLargestOffset); 6403 } 6404 6405 bool ConcurrentMarkSweepGeneration::isNearLargestChunk(HeapWord* addr) { 6406 return addr >= _cmsSpace->nearLargestChunk(); 6407 } 6408 6409 FreeChunk* ConcurrentMarkSweepGeneration::find_chunk_at_end() { 6410 return _cmsSpace->find_chunk_at_end(); 6411 } 6412 6413 void ConcurrentMarkSweepGeneration::update_gc_stats(int current_level, 6414 bool full) { 6415 // The next lower level has been collected. Gather any statistics 6416 // that are of interest at this point. 6417 if (!full && (current_level + 1) == level()) { 6418 // Gather statistics on the young generation collection. 6419 collector()->stats().record_gc0_end(used()); 6420 } 6421 } 6422 6423 CMSAdaptiveSizePolicy* ConcurrentMarkSweepGeneration::size_policy() { 6424 GenCollectedHeap* gch = GenCollectedHeap::heap(); 6425 assert(gch->kind() == CollectedHeap::GenCollectedHeap, 6426 "Wrong type of heap"); 6427 CMSAdaptiveSizePolicy* sp = (CMSAdaptiveSizePolicy*) 6428 gch->gen_policy()->size_policy(); 6429 assert(sp->is_gc_cms_adaptive_size_policy(), 6430 "Wrong type of size policy"); 6431 return sp; 6432 } 6433 6434 void ConcurrentMarkSweepGeneration::rotate_debug_collection_type() { 6435 if (PrintGCDetails && Verbose) { 6436 gclog_or_tty->print("Rotate from %d ", _debug_collection_type); 6437 } 6438 _debug_collection_type = (CollectionTypes) (_debug_collection_type + 1); 6439 _debug_collection_type = 6440 (CollectionTypes) (_debug_collection_type % Unknown_collection_type); 6441 if (PrintGCDetails && Verbose) { 6442 gclog_or_tty->print_cr("to %d ", _debug_collection_type); 6443 } 6444 } 6445 6446 void CMSCollector::sweepWork(ConcurrentMarkSweepGeneration* gen, 6447 bool asynch) { 6448 // We iterate over the space(s) underlying this generation, 6449 // checking the mark bit map to see if the bits corresponding 6450 // to specific blocks are marked or not. Blocks that are 6451 // marked are live and are not swept up. All remaining blocks 6452 // are swept up, with coalescing on-the-fly as we sweep up 6453 // contiguous free and/or garbage blocks: 6454 // We need to ensure that the sweeper synchronizes with allocators 6455 // and stop-the-world collectors. In particular, the following 6456 // locks are used: 6457 // . CMS token: if this is held, a stop the world collection cannot occur 6458 // . freelistLock: if this is held no allocation can occur from this 6459 // generation by another thread 6460 // . bitMapLock: if this is held, no other thread can access or update 6461 // 6462 6463 // Note that we need to hold the freelistLock if we use 6464 // block iterate below; else the iterator might go awry if 6465 // a mutator (or promotion) causes block contents to change 6466 // (for instance if the allocator divvies up a block). 6467 // If we hold the free list lock, for all practical purposes 6468 // young generation GC's can't occur (they'll usually need to 6469 // promote), so we might as well prevent all young generation 6470 // GC's while we do a sweeping step. For the same reason, we might 6471 // as well take the bit map lock for the entire duration 6472 6473 // check that we hold the requisite locks 6474 assert(have_cms_token(), "Should hold cms token"); 6475 assert( (asynch && ConcurrentMarkSweepThread::cms_thread_has_cms_token()) 6476 || (!asynch && ConcurrentMarkSweepThread::vm_thread_has_cms_token()), 6477 "Should possess CMS token to sweep"); 6478 assert_lock_strong(gen->freelistLock()); 6479 assert_lock_strong(bitMapLock()); 6480 6481 assert(!_inter_sweep_timer.is_active(), "Was switched off in an outer context"); 6482 assert(_intra_sweep_timer.is_active(), "Was switched on in an outer context"); 6483 gen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()), 6484 _inter_sweep_estimate.padded_average(), 6485 _intra_sweep_estimate.padded_average()); 6486 gen->setNearLargestChunk(); 6487 6488 { 6489 SweepClosure sweepClosure(this, gen, &_markBitMap, 6490 CMSYield && asynch); 6491 gen->cmsSpace()->blk_iterate_careful(&sweepClosure); 6492 // We need to free-up/coalesce garbage/blocks from a 6493 // co-terminal free run. This is done in the SweepClosure 6494 // destructor; so, do not remove this scope, else the 6495 // end-of-sweep-census below will be off by a little bit. 6496 } 6497 gen->cmsSpace()->sweep_completed(); 6498 gen->cmsSpace()->endSweepFLCensus(sweep_count()); 6499 if (should_unload_classes()) { // unloaded classes this cycle, 6500 _concurrent_cycles_since_last_unload = 0; // ... reset count 6501 } else { // did not unload classes, 6502 _concurrent_cycles_since_last_unload++; // ... increment count 6503 } 6504 } 6505 6506 // Reset CMS data structures (for now just the marking bit map) 6507 // preparatory for the next cycle. 6508 void CMSCollector::reset(bool asynch) { 6509 GenCollectedHeap* gch = GenCollectedHeap::heap(); 6510 CMSAdaptiveSizePolicy* sp = size_policy(); 6511 AdaptiveSizePolicyOutput(sp, gch->total_collections()); 6512 if (asynch) { 6513 CMSTokenSyncWithLocks ts(true, bitMapLock()); 6514 6515 // If the state is not "Resetting", the foreground thread 6516 // has done a collection and the resetting. 6517 if (_collectorState != Resetting) { 6518 assert(_collectorState == Idling, "The state should only change" 6519 " because the foreground collector has finished the collection"); 6520 return; 6521 } 6522 6523 // Clear the mark bitmap (no grey objects to start with) 6524 // for the next cycle. 6525 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 6526 CMSPhaseAccounting cmspa(this, "reset", !PrintGCDetails); 6527 6528 HeapWord* curAddr = _markBitMap.startWord(); 6529 while (curAddr < _markBitMap.endWord()) { 6530 size_t remaining = pointer_delta(_markBitMap.endWord(), curAddr); 6531 MemRegion chunk(curAddr, MIN2(CMSBitMapYieldQuantum, remaining)); 6532 _markBitMap.clear_large_range(chunk); 6533 if (ConcurrentMarkSweepThread::should_yield() && 6534 !foregroundGCIsActive() && 6535 CMSYield) { 6536 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6537 "CMS thread should hold CMS token"); 6538 assert_lock_strong(bitMapLock()); 6539 bitMapLock()->unlock(); 6540 ConcurrentMarkSweepThread::desynchronize(true); 6541 ConcurrentMarkSweepThread::acknowledge_yield_request(); 6542 stopTimer(); 6543 if (PrintCMSStatistics != 0) { 6544 incrementYields(); 6545 } 6546 icms_wait(); 6547 6548 // See the comment in coordinator_yield() 6549 for (unsigned i = 0; i < CMSYieldSleepCount && 6550 ConcurrentMarkSweepThread::should_yield() && 6551 !CMSCollector::foregroundGCIsActive(); ++i) { 6552 os::sleep(Thread::current(), 1, false); 6553 ConcurrentMarkSweepThread::acknowledge_yield_request(); 6554 } 6555 6556 ConcurrentMarkSweepThread::synchronize(true); 6557 bitMapLock()->lock_without_safepoint_check(); 6558 startTimer(); 6559 } 6560 curAddr = chunk.end(); 6561 } 6562 // A successful mostly concurrent collection has been done. 6563 // Because only the full (i.e., concurrent mode failure) collections 6564 // are being measured for gc overhead limits, clean the "near" flag 6565 // and count. 6566 sp->reset_gc_overhead_limit_count(); 6567 _collectorState = Idling; 6568 } else { 6569 // already have the lock 6570 assert(_collectorState == Resetting, "just checking"); 6571 assert_lock_strong(bitMapLock()); 6572 _markBitMap.clear_all(); 6573 _collectorState = Idling; 6574 } 6575 6576 // Stop incremental mode after a cycle completes, so that any future cycles 6577 // are triggered by allocation. 6578 stop_icms(); 6579 6580 NOT_PRODUCT( 6581 if (RotateCMSCollectionTypes) { 6582 _cmsGen->rotate_debug_collection_type(); 6583 } 6584 ) 6585 6586 register_gc_end(); 6587 } 6588 6589 void CMSCollector::do_CMS_operation(CMS_op_type op, GCCause::Cause gc_cause) { 6590 gclog_or_tty->date_stamp(PrintGC && PrintGCDateStamps); 6591 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 6592 GCTraceTime t(GCCauseString("GC", gc_cause), PrintGC, !PrintGCDetails, NULL); 6593 TraceCollectorStats tcs(counters()); 6594 6595 switch (op) { 6596 case CMS_op_checkpointRootsInitial: { 6597 SvcGCMarker sgcm(SvcGCMarker::OTHER); 6598 checkpointRootsInitial(true); // asynch 6599 if (PrintGC) { 6600 _cmsGen->printOccupancy("initial-mark"); 6601 } 6602 break; 6603 } 6604 case CMS_op_checkpointRootsFinal: { 6605 SvcGCMarker sgcm(SvcGCMarker::OTHER); 6606 checkpointRootsFinal(true, // asynch 6607 false, // !clear_all_soft_refs 6608 false); // !init_mark_was_synchronous 6609 if (PrintGC) { 6610 _cmsGen->printOccupancy("remark"); 6611 } 6612 break; 6613 } 6614 default: 6615 fatal("No such CMS_op"); 6616 } 6617 } 6618 6619 #ifndef PRODUCT 6620 size_t const CMSCollector::skip_header_HeapWords() { 6621 return FreeChunk::header_size(); 6622 } 6623 6624 // Try and collect here conditions that should hold when 6625 // CMS thread is exiting. The idea is that the foreground GC 6626 // thread should not be blocked if it wants to terminate 6627 // the CMS thread and yet continue to run the VM for a while 6628 // after that. 6629 void CMSCollector::verify_ok_to_terminate() const { 6630 assert(Thread::current()->is_ConcurrentGC_thread(), 6631 "should be called by CMS thread"); 6632 assert(!_foregroundGCShouldWait, "should be false"); 6633 // We could check here that all the various low-level locks 6634 // are not held by the CMS thread, but that is overkill; see 6635 // also CMSThread::verify_ok_to_terminate() where the CGC_lock 6636 // is checked. 6637 } 6638 #endif 6639 6640 size_t CMSCollector::block_size_using_printezis_bits(HeapWord* addr) const { 6641 assert(_markBitMap.isMarked(addr) && _markBitMap.isMarked(addr + 1), 6642 "missing Printezis mark?"); 6643 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 6644 size_t size = pointer_delta(nextOneAddr + 1, addr); 6645 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 6646 "alignment problem"); 6647 assert(size >= 3, "Necessary for Printezis marks to work"); 6648 return size; 6649 } 6650 6651 // A variant of the above (block_size_using_printezis_bits()) except 6652 // that we return 0 if the P-bits are not yet set. 6653 size_t CMSCollector::block_size_if_printezis_bits(HeapWord* addr) const { 6654 if (_markBitMap.isMarked(addr + 1)) { 6655 assert(_markBitMap.isMarked(addr), "P-bit can be set only for marked objects"); 6656 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 6657 size_t size = pointer_delta(nextOneAddr + 1, addr); 6658 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 6659 "alignment problem"); 6660 assert(size >= 3, "Necessary for Printezis marks to work"); 6661 return size; 6662 } 6663 return 0; 6664 } 6665 6666 HeapWord* CMSCollector::next_card_start_after_block(HeapWord* addr) const { 6667 size_t sz = 0; 6668 oop p = (oop)addr; 6669 if (p->klass_or_null() != NULL) { 6670 sz = CompactibleFreeListSpace::adjustObjectSize(p->size()); 6671 } else { 6672 sz = block_size_using_printezis_bits(addr); 6673 } 6674 assert(sz > 0, "size must be nonzero"); 6675 HeapWord* next_block = addr + sz; 6676 HeapWord* next_card = (HeapWord*)round_to((uintptr_t)next_block, 6677 CardTableModRefBS::card_size); 6678 assert(round_down((uintptr_t)addr, CardTableModRefBS::card_size) < 6679 round_down((uintptr_t)next_card, CardTableModRefBS::card_size), 6680 "must be different cards"); 6681 return next_card; 6682 } 6683 6684 6685 // CMS Bit Map Wrapper ///////////////////////////////////////// 6686 6687 // Construct a CMS bit map infrastructure, but don't create the 6688 // bit vector itself. That is done by a separate call CMSBitMap::allocate() 6689 // further below. 6690 CMSBitMap::CMSBitMap(int shifter, int mutex_rank, const char* mutex_name): 6691 _bm(), 6692 _shifter(shifter), 6693 _lock(mutex_rank >= 0 ? new Mutex(mutex_rank, mutex_name, true) : NULL) 6694 { 6695 _bmStartWord = 0; 6696 _bmWordSize = 0; 6697 } 6698 6699 bool CMSBitMap::allocate(MemRegion mr) { 6700 _bmStartWord = mr.start(); 6701 _bmWordSize = mr.word_size(); 6702 ReservedSpace brs(ReservedSpace::allocation_align_size_up( 6703 (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1)); 6704 if (!brs.is_reserved()) { 6705 warning("CMS bit map allocation failure"); 6706 return false; 6707 } 6708 // For now we'll just commit all of the bit map up front. 6709 // Later on we'll try to be more parsimonious with swap. 6710 if (!_virtual_space.initialize(brs, brs.size())) { 6711 warning("CMS bit map backing store failure"); 6712 return false; 6713 } 6714 assert(_virtual_space.committed_size() == brs.size(), 6715 "didn't reserve backing store for all of CMS bit map?"); 6716 _bm.set_map((BitMap::bm_word_t*)_virtual_space.low()); 6717 assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >= 6718 _bmWordSize, "inconsistency in bit map sizing"); 6719 _bm.set_size(_bmWordSize >> _shifter); 6720 6721 // bm.clear(); // can we rely on getting zero'd memory? verify below 6722 assert(isAllClear(), 6723 "Expected zero'd memory from ReservedSpace constructor"); 6724 assert(_bm.size() == heapWordDiffToOffsetDiff(sizeInWords()), 6725 "consistency check"); 6726 return true; 6727 } 6728 6729 void CMSBitMap::dirty_range_iterate_clear(MemRegion mr, MemRegionClosure* cl) { 6730 HeapWord *next_addr, *end_addr, *last_addr; 6731 assert_locked(); 6732 assert(covers(mr), "out-of-range error"); 6733 // XXX assert that start and end are appropriately aligned 6734 for (next_addr = mr.start(), end_addr = mr.end(); 6735 next_addr < end_addr; next_addr = last_addr) { 6736 MemRegion dirty_region = getAndClearMarkedRegion(next_addr, end_addr); 6737 last_addr = dirty_region.end(); 6738 if (!dirty_region.is_empty()) { 6739 cl->do_MemRegion(dirty_region); 6740 } else { 6741 assert(last_addr == end_addr, "program logic"); 6742 return; 6743 } 6744 } 6745 } 6746 6747 void CMSBitMap::print_on_error(outputStream* st, const char* prefix) const { 6748 _bm.print_on_error(st, prefix); 6749 } 6750 6751 #ifndef PRODUCT 6752 void CMSBitMap::assert_locked() const { 6753 CMSLockVerifier::assert_locked(lock()); 6754 } 6755 6756 bool CMSBitMap::covers(MemRegion mr) const { 6757 // assert(_bm.map() == _virtual_space.low(), "map inconsistency"); 6758 assert((size_t)_bm.size() == (_bmWordSize >> _shifter), 6759 "size inconsistency"); 6760 return (mr.start() >= _bmStartWord) && 6761 (mr.end() <= endWord()); 6762 } 6763 6764 bool CMSBitMap::covers(HeapWord* start, size_t size) const { 6765 return (start >= _bmStartWord && (start + size) <= endWord()); 6766 } 6767 6768 void CMSBitMap::verifyNoOneBitsInRange(HeapWord* left, HeapWord* right) { 6769 // verify that there are no 1 bits in the interval [left, right) 6770 FalseBitMapClosure falseBitMapClosure; 6771 iterate(&falseBitMapClosure, left, right); 6772 } 6773 6774 void CMSBitMap::region_invariant(MemRegion mr) 6775 { 6776 assert_locked(); 6777 // mr = mr.intersection(MemRegion(_bmStartWord, _bmWordSize)); 6778 assert(!mr.is_empty(), "unexpected empty region"); 6779 assert(covers(mr), "mr should be covered by bit map"); 6780 // convert address range into offset range 6781 size_t start_ofs = heapWordToOffset(mr.start()); 6782 // Make sure that end() is appropriately aligned 6783 assert(mr.end() == (HeapWord*)round_to((intptr_t)mr.end(), 6784 (1 << (_shifter+LogHeapWordSize))), 6785 "Misaligned mr.end()"); 6786 size_t end_ofs = heapWordToOffset(mr.end()); 6787 assert(end_ofs > start_ofs, "Should mark at least one bit"); 6788 } 6789 6790 #endif 6791 6792 bool CMSMarkStack::allocate(size_t size) { 6793 // allocate a stack of the requisite depth 6794 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 6795 size * sizeof(oop))); 6796 if (!rs.is_reserved()) { 6797 warning("CMSMarkStack allocation failure"); 6798 return false; 6799 } 6800 if (!_virtual_space.initialize(rs, rs.size())) { 6801 warning("CMSMarkStack backing store failure"); 6802 return false; 6803 } 6804 assert(_virtual_space.committed_size() == rs.size(), 6805 "didn't reserve backing store for all of CMS stack?"); 6806 _base = (oop*)(_virtual_space.low()); 6807 _index = 0; 6808 _capacity = size; 6809 NOT_PRODUCT(_max_depth = 0); 6810 return true; 6811 } 6812 6813 // XXX FIX ME !!! In the MT case we come in here holding a 6814 // leaf lock. For printing we need to take a further lock 6815 // which has lower rank. We need to recalibrate the two 6816 // lock-ranks involved in order to be able to print the 6817 // messages below. (Or defer the printing to the caller. 6818 // For now we take the expedient path of just disabling the 6819 // messages for the problematic case.) 6820 void CMSMarkStack::expand() { 6821 assert(_capacity <= MarkStackSizeMax, "stack bigger than permitted"); 6822 if (_capacity == MarkStackSizeMax) { 6823 if (_hit_limit++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) { 6824 // We print a warning message only once per CMS cycle. 6825 gclog_or_tty->print_cr(" (benign) Hit CMSMarkStack max size limit"); 6826 } 6827 return; 6828 } 6829 // Double capacity if possible 6830 size_t new_capacity = MIN2(_capacity*2, MarkStackSizeMax); 6831 // Do not give up existing stack until we have managed to 6832 // get the double capacity that we desired. 6833 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 6834 new_capacity * sizeof(oop))); 6835 if (rs.is_reserved()) { 6836 // Release the backing store associated with old stack 6837 _virtual_space.release(); 6838 // Reinitialize virtual space for new stack 6839 if (!_virtual_space.initialize(rs, rs.size())) { 6840 fatal("Not enough swap for expanded marking stack"); 6841 } 6842 _base = (oop*)(_virtual_space.low()); 6843 _index = 0; 6844 _capacity = new_capacity; 6845 } else if (_failed_double++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) { 6846 // Failed to double capacity, continue; 6847 // we print a detail message only once per CMS cycle. 6848 gclog_or_tty->print(" (benign) Failed to expand marking stack from "SIZE_FORMAT"K to " 6849 SIZE_FORMAT"K", 6850 _capacity / K, new_capacity / K); 6851 } 6852 } 6853 6854 6855 // Closures 6856 // XXX: there seems to be a lot of code duplication here; 6857 // should refactor and consolidate common code. 6858 6859 // This closure is used to mark refs into the CMS generation in 6860 // the CMS bit map. Called at the first checkpoint. This closure 6861 // assumes that we do not need to re-mark dirty cards; if the CMS 6862 // generation on which this is used is not an oldest 6863 // generation then this will lose younger_gen cards! 6864 6865 MarkRefsIntoClosure::MarkRefsIntoClosure( 6866 MemRegion span, CMSBitMap* bitMap): 6867 _span(span), 6868 _bitMap(bitMap) 6869 { 6870 assert(_ref_processor == NULL, "deliberately left NULL"); 6871 assert(_bitMap->covers(_span), "_bitMap/_span mismatch"); 6872 } 6873 6874 void MarkRefsIntoClosure::do_oop(oop obj) { 6875 // if p points into _span, then mark corresponding bit in _markBitMap 6876 assert(obj->is_oop(), "expected an oop"); 6877 HeapWord* addr = (HeapWord*)obj; 6878 if (_span.contains(addr)) { 6879 // this should be made more efficient 6880 _bitMap->mark(addr); 6881 } 6882 } 6883 6884 void MarkRefsIntoClosure::do_oop(oop* p) { MarkRefsIntoClosure::do_oop_work(p); } 6885 void MarkRefsIntoClosure::do_oop(narrowOop* p) { MarkRefsIntoClosure::do_oop_work(p); } 6886 6887 Par_MarkRefsIntoClosure::Par_MarkRefsIntoClosure( 6888 MemRegion span, CMSBitMap* bitMap): 6889 _span(span), 6890 _bitMap(bitMap) 6891 { 6892 assert(_ref_processor == NULL, "deliberately left NULL"); 6893 assert(_bitMap->covers(_span), "_bitMap/_span mismatch"); 6894 } 6895 6896 void Par_MarkRefsIntoClosure::do_oop(oop obj) { 6897 // if p points into _span, then mark corresponding bit in _markBitMap 6898 assert(obj->is_oop(), "expected an oop"); 6899 HeapWord* addr = (HeapWord*)obj; 6900 if (_span.contains(addr)) { 6901 // this should be made more efficient 6902 _bitMap->par_mark(addr); 6903 } 6904 } 6905 6906 void Par_MarkRefsIntoClosure::do_oop(oop* p) { Par_MarkRefsIntoClosure::do_oop_work(p); } 6907 void Par_MarkRefsIntoClosure::do_oop(narrowOop* p) { Par_MarkRefsIntoClosure::do_oop_work(p); } 6908 6909 // A variant of the above, used for CMS marking verification. 6910 MarkRefsIntoVerifyClosure::MarkRefsIntoVerifyClosure( 6911 MemRegion span, CMSBitMap* verification_bm, CMSBitMap* cms_bm): 6912 _span(span), 6913 _verification_bm(verification_bm), 6914 _cms_bm(cms_bm) 6915 { 6916 assert(_ref_processor == NULL, "deliberately left NULL"); 6917 assert(_verification_bm->covers(_span), "_verification_bm/_span mismatch"); 6918 } 6919 6920 void MarkRefsIntoVerifyClosure::do_oop(oop obj) { 6921 // if p points into _span, then mark corresponding bit in _markBitMap 6922 assert(obj->is_oop(), "expected an oop"); 6923 HeapWord* addr = (HeapWord*)obj; 6924 if (_span.contains(addr)) { 6925 _verification_bm->mark(addr); 6926 if (!_cms_bm->isMarked(addr)) { 6927 oop(addr)->print(); 6928 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)", addr); 6929 fatal("... aborting"); 6930 } 6931 } 6932 } 6933 6934 void MarkRefsIntoVerifyClosure::do_oop(oop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); } 6935 void MarkRefsIntoVerifyClosure::do_oop(narrowOop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); } 6936 6937 ////////////////////////////////////////////////// 6938 // MarkRefsIntoAndScanClosure 6939 ////////////////////////////////////////////////// 6940 6941 MarkRefsIntoAndScanClosure::MarkRefsIntoAndScanClosure(MemRegion span, 6942 ReferenceProcessor* rp, 6943 CMSBitMap* bit_map, 6944 CMSBitMap* mod_union_table, 6945 CMSMarkStack* mark_stack, 6946 CMSCollector* collector, 6947 bool should_yield, 6948 bool concurrent_precleaning): 6949 _collector(collector), 6950 _span(span), 6951 _bit_map(bit_map), 6952 _mark_stack(mark_stack), 6953 _pushAndMarkClosure(collector, span, rp, bit_map, mod_union_table, 6954 mark_stack, concurrent_precleaning), 6955 _yield(should_yield), 6956 _concurrent_precleaning(concurrent_precleaning), 6957 _freelistLock(NULL) 6958 { 6959 _ref_processor = rp; 6960 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 6961 } 6962 6963 // This closure is used to mark refs into the CMS generation at the 6964 // second (final) checkpoint, and to scan and transitively follow 6965 // the unmarked oops. It is also used during the concurrent precleaning 6966 // phase while scanning objects on dirty cards in the CMS generation. 6967 // The marks are made in the marking bit map and the marking stack is 6968 // used for keeping the (newly) grey objects during the scan. 6969 // The parallel version (Par_...) appears further below. 6970 void MarkRefsIntoAndScanClosure::do_oop(oop obj) { 6971 if (obj != NULL) { 6972 assert(obj->is_oop(), "expected an oop"); 6973 HeapWord* addr = (HeapWord*)obj; 6974 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 6975 assert(_collector->overflow_list_is_empty(), 6976 "overflow list should be empty"); 6977 if (_span.contains(addr) && 6978 !_bit_map->isMarked(addr)) { 6979 // mark bit map (object is now grey) 6980 _bit_map->mark(addr); 6981 // push on marking stack (stack should be empty), and drain the 6982 // stack by applying this closure to the oops in the oops popped 6983 // from the stack (i.e. blacken the grey objects) 6984 bool res = _mark_stack->push(obj); 6985 assert(res, "Should have space to push on empty stack"); 6986 do { 6987 oop new_oop = _mark_stack->pop(); 6988 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 6989 assert(_bit_map->isMarked((HeapWord*)new_oop), 6990 "only grey objects on this stack"); 6991 // iterate over the oops in this oop, marking and pushing 6992 // the ones in CMS heap (i.e. in _span). 6993 new_oop->oop_iterate(&_pushAndMarkClosure); 6994 // check if it's time to yield 6995 do_yield_check(); 6996 } while (!_mark_stack->isEmpty() || 6997 (!_concurrent_precleaning && take_from_overflow_list())); 6998 // if marking stack is empty, and we are not doing this 6999 // during precleaning, then check the overflow list 7000 } 7001 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 7002 assert(_collector->overflow_list_is_empty(), 7003 "overflow list was drained above"); 7004 // We could restore evacuated mark words, if any, used for 7005 // overflow list links here because the overflow list is 7006 // provably empty here. That would reduce the maximum 7007 // size requirements for preserved_{oop,mark}_stack. 7008 // But we'll just postpone it until we are all done 7009 // so we can just stream through. 7010 if (!_concurrent_precleaning && CMSOverflowEarlyRestoration) { 7011 _collector->restore_preserved_marks_if_any(); 7012 assert(_collector->no_preserved_marks(), "No preserved marks"); 7013 } 7014 assert(!CMSOverflowEarlyRestoration || _collector->no_preserved_marks(), 7015 "All preserved marks should have been restored above"); 7016 } 7017 } 7018 7019 void MarkRefsIntoAndScanClosure::do_oop(oop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); } 7020 void MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); } 7021 7022 void MarkRefsIntoAndScanClosure::do_yield_work() { 7023 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 7024 "CMS thread should hold CMS token"); 7025 assert_lock_strong(_freelistLock); 7026 assert_lock_strong(_bit_map->lock()); 7027 // relinquish the free_list_lock and bitMaplock() 7028 _bit_map->lock()->unlock(); 7029 _freelistLock->unlock(); 7030 ConcurrentMarkSweepThread::desynchronize(true); 7031 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7032 _collector->stopTimer(); 7033 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 7034 if (PrintCMSStatistics != 0) { 7035 _collector->incrementYields(); 7036 } 7037 _collector->icms_wait(); 7038 7039 // See the comment in coordinator_yield() 7040 for (unsigned i = 0; 7041 i < CMSYieldSleepCount && 7042 ConcurrentMarkSweepThread::should_yield() && 7043 !CMSCollector::foregroundGCIsActive(); 7044 ++i) { 7045 os::sleep(Thread::current(), 1, false); 7046 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7047 } 7048 7049 ConcurrentMarkSweepThread::synchronize(true); 7050 _freelistLock->lock_without_safepoint_check(); 7051 _bit_map->lock()->lock_without_safepoint_check(); 7052 _collector->startTimer(); 7053 } 7054 7055 /////////////////////////////////////////////////////////// 7056 // Par_MarkRefsIntoAndScanClosure: a parallel version of 7057 // MarkRefsIntoAndScanClosure 7058 /////////////////////////////////////////////////////////// 7059 Par_MarkRefsIntoAndScanClosure::Par_MarkRefsIntoAndScanClosure( 7060 CMSCollector* collector, MemRegion span, ReferenceProcessor* rp, 7061 CMSBitMap* bit_map, OopTaskQueue* work_queue): 7062 _span(span), 7063 _bit_map(bit_map), 7064 _work_queue(work_queue), 7065 _low_water_mark(MIN2((uint)(work_queue->max_elems()/4), 7066 (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads))), 7067 _par_pushAndMarkClosure(collector, span, rp, bit_map, work_queue) 7068 { 7069 _ref_processor = rp; 7070 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 7071 } 7072 7073 // This closure is used to mark refs into the CMS generation at the 7074 // second (final) checkpoint, and to scan and transitively follow 7075 // the unmarked oops. The marks are made in the marking bit map and 7076 // the work_queue is used for keeping the (newly) grey objects during 7077 // the scan phase whence they are also available for stealing by parallel 7078 // threads. Since the marking bit map is shared, updates are 7079 // synchronized (via CAS). 7080 void Par_MarkRefsIntoAndScanClosure::do_oop(oop obj) { 7081 if (obj != NULL) { 7082 // Ignore mark word because this could be an already marked oop 7083 // that may be chained at the end of the overflow list. 7084 assert(obj->is_oop(true), "expected an oop"); 7085 HeapWord* addr = (HeapWord*)obj; 7086 if (_span.contains(addr) && 7087 !_bit_map->isMarked(addr)) { 7088 // mark bit map (object will become grey): 7089 // It is possible for several threads to be 7090 // trying to "claim" this object concurrently; 7091 // the unique thread that succeeds in marking the 7092 // object first will do the subsequent push on 7093 // to the work queue (or overflow list). 7094 if (_bit_map->par_mark(addr)) { 7095 // push on work_queue (which may not be empty), and trim the 7096 // queue to an appropriate length by applying this closure to 7097 // the oops in the oops popped from the stack (i.e. blacken the 7098 // grey objects) 7099 bool res = _work_queue->push(obj); 7100 assert(res, "Low water mark should be less than capacity?"); 7101 trim_queue(_low_water_mark); 7102 } // Else, another thread claimed the object 7103 } 7104 } 7105 } 7106 7107 void Par_MarkRefsIntoAndScanClosure::do_oop(oop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); } 7108 void Par_MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); } 7109 7110 // This closure is used to rescan the marked objects on the dirty cards 7111 // in the mod union table and the card table proper. 7112 size_t ScanMarkedObjectsAgainCarefullyClosure::do_object_careful_m( 7113 oop p, MemRegion mr) { 7114 7115 size_t size = 0; 7116 HeapWord* addr = (HeapWord*)p; 7117 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 7118 assert(_span.contains(addr), "we are scanning the CMS generation"); 7119 // check if it's time to yield 7120 if (do_yield_check()) { 7121 // We yielded for some foreground stop-world work, 7122 // and we have been asked to abort this ongoing preclean cycle. 7123 return 0; 7124 } 7125 if (_bitMap->isMarked(addr)) { 7126 // it's marked; is it potentially uninitialized? 7127 if (p->klass_or_null() != NULL) { 7128 // an initialized object; ignore mark word in verification below 7129 // since we are running concurrent with mutators 7130 assert(p->is_oop(true), "should be an oop"); 7131 if (p->is_objArray()) { 7132 // objArrays are precisely marked; restrict scanning 7133 // to dirty cards only. 7134 size = CompactibleFreeListSpace::adjustObjectSize( 7135 p->oop_iterate(_scanningClosure, mr)); 7136 } else { 7137 // A non-array may have been imprecisely marked; we need 7138 // to scan object in its entirety. 7139 size = CompactibleFreeListSpace::adjustObjectSize( 7140 p->oop_iterate(_scanningClosure)); 7141 } 7142 #ifdef ASSERT 7143 size_t direct_size = 7144 CompactibleFreeListSpace::adjustObjectSize(p->size()); 7145 assert(size == direct_size, "Inconsistency in size"); 7146 assert(size >= 3, "Necessary for Printezis marks to work"); 7147 if (!_bitMap->isMarked(addr+1)) { 7148 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size); 7149 } else { 7150 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size-1); 7151 assert(_bitMap->isMarked(addr+size-1), 7152 "inconsistent Printezis mark"); 7153 } 7154 #endif // ASSERT 7155 } else { 7156 // An uninitialized object. 7157 assert(_bitMap->isMarked(addr+1), "missing Printezis mark?"); 7158 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 7159 size = pointer_delta(nextOneAddr + 1, addr); 7160 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 7161 "alignment problem"); 7162 // Note that pre-cleaning needn't redirty the card. OopDesc::set_klass() 7163 // will dirty the card when the klass pointer is installed in the 7164 // object (signaling the completion of initialization). 7165 } 7166 } else { 7167 // Either a not yet marked object or an uninitialized object 7168 if (p->klass_or_null() == NULL) { 7169 // An uninitialized object, skip to the next card, since 7170 // we may not be able to read its P-bits yet. 7171 assert(size == 0, "Initial value"); 7172 } else { 7173 // An object not (yet) reached by marking: we merely need to 7174 // compute its size so as to go look at the next block. 7175 assert(p->is_oop(true), "should be an oop"); 7176 size = CompactibleFreeListSpace::adjustObjectSize(p->size()); 7177 } 7178 } 7179 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 7180 return size; 7181 } 7182 7183 void ScanMarkedObjectsAgainCarefullyClosure::do_yield_work() { 7184 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 7185 "CMS thread should hold CMS token"); 7186 assert_lock_strong(_freelistLock); 7187 assert_lock_strong(_bitMap->lock()); 7188 // relinquish the free_list_lock and bitMaplock() 7189 _bitMap->lock()->unlock(); 7190 _freelistLock->unlock(); 7191 ConcurrentMarkSweepThread::desynchronize(true); 7192 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7193 _collector->stopTimer(); 7194 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 7195 if (PrintCMSStatistics != 0) { 7196 _collector->incrementYields(); 7197 } 7198 _collector->icms_wait(); 7199 7200 // See the comment in coordinator_yield() 7201 for (unsigned i = 0; i < CMSYieldSleepCount && 7202 ConcurrentMarkSweepThread::should_yield() && 7203 !CMSCollector::foregroundGCIsActive(); ++i) { 7204 os::sleep(Thread::current(), 1, false); 7205 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7206 } 7207 7208 ConcurrentMarkSweepThread::synchronize(true); 7209 _freelistLock->lock_without_safepoint_check(); 7210 _bitMap->lock()->lock_without_safepoint_check(); 7211 _collector->startTimer(); 7212 } 7213 7214 7215 ////////////////////////////////////////////////////////////////// 7216 // SurvivorSpacePrecleanClosure 7217 ////////////////////////////////////////////////////////////////// 7218 // This (single-threaded) closure is used to preclean the oops in 7219 // the survivor spaces. 7220 size_t SurvivorSpacePrecleanClosure::do_object_careful(oop p) { 7221 7222 HeapWord* addr = (HeapWord*)p; 7223 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 7224 assert(!_span.contains(addr), "we are scanning the survivor spaces"); 7225 assert(p->klass_or_null() != NULL, "object should be initialized"); 7226 // an initialized object; ignore mark word in verification below 7227 // since we are running concurrent with mutators 7228 assert(p->is_oop(true), "should be an oop"); 7229 // Note that we do not yield while we iterate over 7230 // the interior oops of p, pushing the relevant ones 7231 // on our marking stack. 7232 size_t size = p->oop_iterate(_scanning_closure); 7233 do_yield_check(); 7234 // Observe that below, we do not abandon the preclean 7235 // phase as soon as we should; rather we empty the 7236 // marking stack before returning. This is to satisfy 7237 // some existing assertions. In general, it may be a 7238 // good idea to abort immediately and complete the marking 7239 // from the grey objects at a later time. 7240 while (!_mark_stack->isEmpty()) { 7241 oop new_oop = _mark_stack->pop(); 7242 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 7243 assert(_bit_map->isMarked((HeapWord*)new_oop), 7244 "only grey objects on this stack"); 7245 // iterate over the oops in this oop, marking and pushing 7246 // the ones in CMS heap (i.e. in _span). 7247 new_oop->oop_iterate(_scanning_closure); 7248 // check if it's time to yield 7249 do_yield_check(); 7250 } 7251 unsigned int after_count = 7252 GenCollectedHeap::heap()->total_collections(); 7253 bool abort = (_before_count != after_count) || 7254 _collector->should_abort_preclean(); 7255 return abort ? 0 : size; 7256 } 7257 7258 void SurvivorSpacePrecleanClosure::do_yield_work() { 7259 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 7260 "CMS thread should hold CMS token"); 7261 assert_lock_strong(_bit_map->lock()); 7262 // Relinquish the bit map lock 7263 _bit_map->lock()->unlock(); 7264 ConcurrentMarkSweepThread::desynchronize(true); 7265 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7266 _collector->stopTimer(); 7267 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 7268 if (PrintCMSStatistics != 0) { 7269 _collector->incrementYields(); 7270 } 7271 _collector->icms_wait(); 7272 7273 // See the comment in coordinator_yield() 7274 for (unsigned i = 0; i < CMSYieldSleepCount && 7275 ConcurrentMarkSweepThread::should_yield() && 7276 !CMSCollector::foregroundGCIsActive(); ++i) { 7277 os::sleep(Thread::current(), 1, false); 7278 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7279 } 7280 7281 ConcurrentMarkSweepThread::synchronize(true); 7282 _bit_map->lock()->lock_without_safepoint_check(); 7283 _collector->startTimer(); 7284 } 7285 7286 // This closure is used to rescan the marked objects on the dirty cards 7287 // in the mod union table and the card table proper. In the parallel 7288 // case, although the bitMap is shared, we do a single read so the 7289 // isMarked() query is "safe". 7290 bool ScanMarkedObjectsAgainClosure::do_object_bm(oop p, MemRegion mr) { 7291 // Ignore mark word because we are running concurrent with mutators 7292 assert(p->is_oop_or_null(true), "expected an oop or null"); 7293 HeapWord* addr = (HeapWord*)p; 7294 assert(_span.contains(addr), "we are scanning the CMS generation"); 7295 bool is_obj_array = false; 7296 #ifdef ASSERT 7297 if (!_parallel) { 7298 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 7299 assert(_collector->overflow_list_is_empty(), 7300 "overflow list should be empty"); 7301 7302 } 7303 #endif // ASSERT 7304 if (_bit_map->isMarked(addr)) { 7305 // Obj arrays are precisely marked, non-arrays are not; 7306 // so we scan objArrays precisely and non-arrays in their 7307 // entirety. 7308 if (p->is_objArray()) { 7309 is_obj_array = true; 7310 if (_parallel) { 7311 p->oop_iterate(_par_scan_closure, mr); 7312 } else { 7313 p->oop_iterate(_scan_closure, mr); 7314 } 7315 } else { 7316 if (_parallel) { 7317 p->oop_iterate(_par_scan_closure); 7318 } else { 7319 p->oop_iterate(_scan_closure); 7320 } 7321 } 7322 } 7323 #ifdef ASSERT 7324 if (!_parallel) { 7325 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 7326 assert(_collector->overflow_list_is_empty(), 7327 "overflow list should be empty"); 7328 7329 } 7330 #endif // ASSERT 7331 return is_obj_array; 7332 } 7333 7334 MarkFromRootsClosure::MarkFromRootsClosure(CMSCollector* collector, 7335 MemRegion span, 7336 CMSBitMap* bitMap, CMSMarkStack* markStack, 7337 bool should_yield, bool verifying): 7338 _collector(collector), 7339 _span(span), 7340 _bitMap(bitMap), 7341 _mut(&collector->_modUnionTable), 7342 _markStack(markStack), 7343 _yield(should_yield), 7344 _skipBits(0) 7345 { 7346 assert(_markStack->isEmpty(), "stack should be empty"); 7347 _finger = _bitMap->startWord(); 7348 _threshold = _finger; 7349 assert(_collector->_restart_addr == NULL, "Sanity check"); 7350 assert(_span.contains(_finger), "Out of bounds _finger?"); 7351 DEBUG_ONLY(_verifying = verifying;) 7352 } 7353 7354 void MarkFromRootsClosure::reset(HeapWord* addr) { 7355 assert(_markStack->isEmpty(), "would cause duplicates on stack"); 7356 assert(_span.contains(addr), "Out of bounds _finger?"); 7357 _finger = addr; 7358 _threshold = (HeapWord*)round_to( 7359 (intptr_t)_finger, CardTableModRefBS::card_size); 7360 } 7361 7362 // Should revisit to see if this should be restructured for 7363 // greater efficiency. 7364 bool MarkFromRootsClosure::do_bit(size_t offset) { 7365 if (_skipBits > 0) { 7366 _skipBits--; 7367 return true; 7368 } 7369 // convert offset into a HeapWord* 7370 HeapWord* addr = _bitMap->startWord() + offset; 7371 assert(_bitMap->endWord() && addr < _bitMap->endWord(), 7372 "address out of range"); 7373 assert(_bitMap->isMarked(addr), "tautology"); 7374 if (_bitMap->isMarked(addr+1)) { 7375 // this is an allocated but not yet initialized object 7376 assert(_skipBits == 0, "tautology"); 7377 _skipBits = 2; // skip next two marked bits ("Printezis-marks") 7378 oop p = oop(addr); 7379 if (p->klass_or_null() == NULL) { 7380 DEBUG_ONLY(if (!_verifying) {) 7381 // We re-dirty the cards on which this object lies and increase 7382 // the _threshold so that we'll come back to scan this object 7383 // during the preclean or remark phase. (CMSCleanOnEnter) 7384 if (CMSCleanOnEnter) { 7385 size_t sz = _collector->block_size_using_printezis_bits(addr); 7386 HeapWord* end_card_addr = (HeapWord*)round_to( 7387 (intptr_t)(addr+sz), CardTableModRefBS::card_size); 7388 MemRegion redirty_range = MemRegion(addr, end_card_addr); 7389 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 7390 // Bump _threshold to end_card_addr; note that 7391 // _threshold cannot possibly exceed end_card_addr, anyhow. 7392 // This prevents future clearing of the card as the scan proceeds 7393 // to the right. 7394 assert(_threshold <= end_card_addr, 7395 "Because we are just scanning into this object"); 7396 if (_threshold < end_card_addr) { 7397 _threshold = end_card_addr; 7398 } 7399 if (p->klass_or_null() != NULL) { 7400 // Redirty the range of cards... 7401 _mut->mark_range(redirty_range); 7402 } // ...else the setting of klass will dirty the card anyway. 7403 } 7404 DEBUG_ONLY(}) 7405 return true; 7406 } 7407 } 7408 scanOopsInOop(addr); 7409 return true; 7410 } 7411 7412 // We take a break if we've been at this for a while, 7413 // so as to avoid monopolizing the locks involved. 7414 void MarkFromRootsClosure::do_yield_work() { 7415 // First give up the locks, then yield, then re-lock 7416 // We should probably use a constructor/destructor idiom to 7417 // do this unlock/lock or modify the MutexUnlocker class to 7418 // serve our purpose. XXX 7419 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 7420 "CMS thread should hold CMS token"); 7421 assert_lock_strong(_bitMap->lock()); 7422 _bitMap->lock()->unlock(); 7423 ConcurrentMarkSweepThread::desynchronize(true); 7424 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7425 _collector->stopTimer(); 7426 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 7427 if (PrintCMSStatistics != 0) { 7428 _collector->incrementYields(); 7429 } 7430 _collector->icms_wait(); 7431 7432 // See the comment in coordinator_yield() 7433 for (unsigned i = 0; i < CMSYieldSleepCount && 7434 ConcurrentMarkSweepThread::should_yield() && 7435 !CMSCollector::foregroundGCIsActive(); ++i) { 7436 os::sleep(Thread::current(), 1, false); 7437 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7438 } 7439 7440 ConcurrentMarkSweepThread::synchronize(true); 7441 _bitMap->lock()->lock_without_safepoint_check(); 7442 _collector->startTimer(); 7443 } 7444 7445 void MarkFromRootsClosure::scanOopsInOop(HeapWord* ptr) { 7446 assert(_bitMap->isMarked(ptr), "expected bit to be set"); 7447 assert(_markStack->isEmpty(), 7448 "should drain stack to limit stack usage"); 7449 // convert ptr to an oop preparatory to scanning 7450 oop obj = oop(ptr); 7451 // Ignore mark word in verification below, since we 7452 // may be running concurrent with mutators. 7453 assert(obj->is_oop(true), "should be an oop"); 7454 assert(_finger <= ptr, "_finger runneth ahead"); 7455 // advance the finger to right end of this object 7456 _finger = ptr + obj->size(); 7457 assert(_finger > ptr, "we just incremented it above"); 7458 // On large heaps, it may take us some time to get through 7459 // the marking phase (especially if running iCMS). During 7460 // this time it's possible that a lot of mutations have 7461 // accumulated in the card table and the mod union table -- 7462 // these mutation records are redundant until we have 7463 // actually traced into the corresponding card. 7464 // Here, we check whether advancing the finger would make 7465 // us cross into a new card, and if so clear corresponding 7466 // cards in the MUT (preclean them in the card-table in the 7467 // future). 7468 7469 DEBUG_ONLY(if (!_verifying) {) 7470 // The clean-on-enter optimization is disabled by default, 7471 // until we fix 6178663. 7472 if (CMSCleanOnEnter && (_finger > _threshold)) { 7473 // [_threshold, _finger) represents the interval 7474 // of cards to be cleared in MUT (or precleaned in card table). 7475 // The set of cards to be cleared is all those that overlap 7476 // with the interval [_threshold, _finger); note that 7477 // _threshold is always kept card-aligned but _finger isn't 7478 // always card-aligned. 7479 HeapWord* old_threshold = _threshold; 7480 assert(old_threshold == (HeapWord*)round_to( 7481 (intptr_t)old_threshold, CardTableModRefBS::card_size), 7482 "_threshold should always be card-aligned"); 7483 _threshold = (HeapWord*)round_to( 7484 (intptr_t)_finger, CardTableModRefBS::card_size); 7485 MemRegion mr(old_threshold, _threshold); 7486 assert(!mr.is_empty(), "Control point invariant"); 7487 assert(_span.contains(mr), "Should clear within span"); 7488 _mut->clear_range(mr); 7489 } 7490 DEBUG_ONLY(}) 7491 // Note: the finger doesn't advance while we drain 7492 // the stack below. 7493 PushOrMarkClosure pushOrMarkClosure(_collector, 7494 _span, _bitMap, _markStack, 7495 _finger, this); 7496 bool res = _markStack->push(obj); 7497 assert(res, "Empty non-zero size stack should have space for single push"); 7498 while (!_markStack->isEmpty()) { 7499 oop new_oop = _markStack->pop(); 7500 // Skip verifying header mark word below because we are 7501 // running concurrent with mutators. 7502 assert(new_oop->is_oop(true), "Oops! expected to pop an oop"); 7503 // now scan this oop's oops 7504 new_oop->oop_iterate(&pushOrMarkClosure); 7505 do_yield_check(); 7506 } 7507 assert(_markStack->isEmpty(), "tautology, emphasizing post-condition"); 7508 } 7509 7510 Par_MarkFromRootsClosure::Par_MarkFromRootsClosure(CMSConcMarkingTask* task, 7511 CMSCollector* collector, MemRegion span, 7512 CMSBitMap* bit_map, 7513 OopTaskQueue* work_queue, 7514 CMSMarkStack* overflow_stack, 7515 bool should_yield): 7516 _collector(collector), 7517 _whole_span(collector->_span), 7518 _span(span), 7519 _bit_map(bit_map), 7520 _mut(&collector->_modUnionTable), 7521 _work_queue(work_queue), 7522 _overflow_stack(overflow_stack), 7523 _yield(should_yield), 7524 _skip_bits(0), 7525 _task(task) 7526 { 7527 assert(_work_queue->size() == 0, "work_queue should be empty"); 7528 _finger = span.start(); 7529 _threshold = _finger; // XXX Defer clear-on-enter optimization for now 7530 assert(_span.contains(_finger), "Out of bounds _finger?"); 7531 } 7532 7533 // Should revisit to see if this should be restructured for 7534 // greater efficiency. 7535 bool Par_MarkFromRootsClosure::do_bit(size_t offset) { 7536 if (_skip_bits > 0) { 7537 _skip_bits--; 7538 return true; 7539 } 7540 // convert offset into a HeapWord* 7541 HeapWord* addr = _bit_map->startWord() + offset; 7542 assert(_bit_map->endWord() && addr < _bit_map->endWord(), 7543 "address out of range"); 7544 assert(_bit_map->isMarked(addr), "tautology"); 7545 if (_bit_map->isMarked(addr+1)) { 7546 // this is an allocated object that might not yet be initialized 7547 assert(_skip_bits == 0, "tautology"); 7548 _skip_bits = 2; // skip next two marked bits ("Printezis-marks") 7549 oop p = oop(addr); 7550 if (p->klass_or_null() == NULL) { 7551 // in the case of Clean-on-Enter optimization, redirty card 7552 // and avoid clearing card by increasing the threshold. 7553 return true; 7554 } 7555 } 7556 scan_oops_in_oop(addr); 7557 return true; 7558 } 7559 7560 void Par_MarkFromRootsClosure::scan_oops_in_oop(HeapWord* ptr) { 7561 assert(_bit_map->isMarked(ptr), "expected bit to be set"); 7562 // Should we assert that our work queue is empty or 7563 // below some drain limit? 7564 assert(_work_queue->size() == 0, 7565 "should drain stack to limit stack usage"); 7566 // convert ptr to an oop preparatory to scanning 7567 oop obj = oop(ptr); 7568 // Ignore mark word in verification below, since we 7569 // may be running concurrent with mutators. 7570 assert(obj->is_oop(true), "should be an oop"); 7571 assert(_finger <= ptr, "_finger runneth ahead"); 7572 // advance the finger to right end of this object 7573 _finger = ptr + obj->size(); 7574 assert(_finger > ptr, "we just incremented it above"); 7575 // On large heaps, it may take us some time to get through 7576 // the marking phase (especially if running iCMS). During 7577 // this time it's possible that a lot of mutations have 7578 // accumulated in the card table and the mod union table -- 7579 // these mutation records are redundant until we have 7580 // actually traced into the corresponding card. 7581 // Here, we check whether advancing the finger would make 7582 // us cross into a new card, and if so clear corresponding 7583 // cards in the MUT (preclean them in the card-table in the 7584 // future). 7585 7586 // The clean-on-enter optimization is disabled by default, 7587 // until we fix 6178663. 7588 if (CMSCleanOnEnter && (_finger > _threshold)) { 7589 // [_threshold, _finger) represents the interval 7590 // of cards to be cleared in MUT (or precleaned in card table). 7591 // The set of cards to be cleared is all those that overlap 7592 // with the interval [_threshold, _finger); note that 7593 // _threshold is always kept card-aligned but _finger isn't 7594 // always card-aligned. 7595 HeapWord* old_threshold = _threshold; 7596 assert(old_threshold == (HeapWord*)round_to( 7597 (intptr_t)old_threshold, CardTableModRefBS::card_size), 7598 "_threshold should always be card-aligned"); 7599 _threshold = (HeapWord*)round_to( 7600 (intptr_t)_finger, CardTableModRefBS::card_size); 7601 MemRegion mr(old_threshold, _threshold); 7602 assert(!mr.is_empty(), "Control point invariant"); 7603 assert(_span.contains(mr), "Should clear within span"); // _whole_span ?? 7604 _mut->clear_range(mr); 7605 } 7606 7607 // Note: the local finger doesn't advance while we drain 7608 // the stack below, but the global finger sure can and will. 7609 HeapWord** gfa = _task->global_finger_addr(); 7610 Par_PushOrMarkClosure pushOrMarkClosure(_collector, 7611 _span, _bit_map, 7612 _work_queue, 7613 _overflow_stack, 7614 _finger, 7615 gfa, this); 7616 bool res = _work_queue->push(obj); // overflow could occur here 7617 assert(res, "Will hold once we use workqueues"); 7618 while (true) { 7619 oop new_oop; 7620 if (!_work_queue->pop_local(new_oop)) { 7621 // We emptied our work_queue; check if there's stuff that can 7622 // be gotten from the overflow stack. 7623 if (CMSConcMarkingTask::get_work_from_overflow_stack( 7624 _overflow_stack, _work_queue)) { 7625 do_yield_check(); 7626 continue; 7627 } else { // done 7628 break; 7629 } 7630 } 7631 // Skip verifying header mark word below because we are 7632 // running concurrent with mutators. 7633 assert(new_oop->is_oop(true), "Oops! expected to pop an oop"); 7634 // now scan this oop's oops 7635 new_oop->oop_iterate(&pushOrMarkClosure); 7636 do_yield_check(); 7637 } 7638 assert(_work_queue->size() == 0, "tautology, emphasizing post-condition"); 7639 } 7640 7641 // Yield in response to a request from VM Thread or 7642 // from mutators. 7643 void Par_MarkFromRootsClosure::do_yield_work() { 7644 assert(_task != NULL, "sanity"); 7645 _task->yield(); 7646 } 7647 7648 // A variant of the above used for verifying CMS marking work. 7649 MarkFromRootsVerifyClosure::MarkFromRootsVerifyClosure(CMSCollector* collector, 7650 MemRegion span, 7651 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 7652 CMSMarkStack* mark_stack): 7653 _collector(collector), 7654 _span(span), 7655 _verification_bm(verification_bm), 7656 _cms_bm(cms_bm), 7657 _mark_stack(mark_stack), 7658 _pam_verify_closure(collector, span, verification_bm, cms_bm, 7659 mark_stack) 7660 { 7661 assert(_mark_stack->isEmpty(), "stack should be empty"); 7662 _finger = _verification_bm->startWord(); 7663 assert(_collector->_restart_addr == NULL, "Sanity check"); 7664 assert(_span.contains(_finger), "Out of bounds _finger?"); 7665 } 7666 7667 void MarkFromRootsVerifyClosure::reset(HeapWord* addr) { 7668 assert(_mark_stack->isEmpty(), "would cause duplicates on stack"); 7669 assert(_span.contains(addr), "Out of bounds _finger?"); 7670 _finger = addr; 7671 } 7672 7673 // Should revisit to see if this should be restructured for 7674 // greater efficiency. 7675 bool MarkFromRootsVerifyClosure::do_bit(size_t offset) { 7676 // convert offset into a HeapWord* 7677 HeapWord* addr = _verification_bm->startWord() + offset; 7678 assert(_verification_bm->endWord() && addr < _verification_bm->endWord(), 7679 "address out of range"); 7680 assert(_verification_bm->isMarked(addr), "tautology"); 7681 assert(_cms_bm->isMarked(addr), "tautology"); 7682 7683 assert(_mark_stack->isEmpty(), 7684 "should drain stack to limit stack usage"); 7685 // convert addr to an oop preparatory to scanning 7686 oop obj = oop(addr); 7687 assert(obj->is_oop(), "should be an oop"); 7688 assert(_finger <= addr, "_finger runneth ahead"); 7689 // advance the finger to right end of this object 7690 _finger = addr + obj->size(); 7691 assert(_finger > addr, "we just incremented it above"); 7692 // Note: the finger doesn't advance while we drain 7693 // the stack below. 7694 bool res = _mark_stack->push(obj); 7695 assert(res, "Empty non-zero size stack should have space for single push"); 7696 while (!_mark_stack->isEmpty()) { 7697 oop new_oop = _mark_stack->pop(); 7698 assert(new_oop->is_oop(), "Oops! expected to pop an oop"); 7699 // now scan this oop's oops 7700 new_oop->oop_iterate(&_pam_verify_closure); 7701 } 7702 assert(_mark_stack->isEmpty(), "tautology, emphasizing post-condition"); 7703 return true; 7704 } 7705 7706 PushAndMarkVerifyClosure::PushAndMarkVerifyClosure( 7707 CMSCollector* collector, MemRegion span, 7708 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 7709 CMSMarkStack* mark_stack): 7710 CMSOopClosure(collector->ref_processor()), 7711 _collector(collector), 7712 _span(span), 7713 _verification_bm(verification_bm), 7714 _cms_bm(cms_bm), 7715 _mark_stack(mark_stack) 7716 { } 7717 7718 void PushAndMarkVerifyClosure::do_oop(oop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 7719 void PushAndMarkVerifyClosure::do_oop(narrowOop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 7720 7721 // Upon stack overflow, we discard (part of) the stack, 7722 // remembering the least address amongst those discarded 7723 // in CMSCollector's _restart_address. 7724 void PushAndMarkVerifyClosure::handle_stack_overflow(HeapWord* lost) { 7725 // Remember the least grey address discarded 7726 HeapWord* ra = (HeapWord*)_mark_stack->least_value(lost); 7727 _collector->lower_restart_addr(ra); 7728 _mark_stack->reset(); // discard stack contents 7729 _mark_stack->expand(); // expand the stack if possible 7730 } 7731 7732 void PushAndMarkVerifyClosure::do_oop(oop obj) { 7733 assert(obj->is_oop_or_null(), "expected an oop or NULL"); 7734 HeapWord* addr = (HeapWord*)obj; 7735 if (_span.contains(addr) && !_verification_bm->isMarked(addr)) { 7736 // Oop lies in _span and isn't yet grey or black 7737 _verification_bm->mark(addr); // now grey 7738 if (!_cms_bm->isMarked(addr)) { 7739 oop(addr)->print(); 7740 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)", 7741 addr); 7742 fatal("... aborting"); 7743 } 7744 7745 if (!_mark_stack->push(obj)) { // stack overflow 7746 if (PrintCMSStatistics != 0) { 7747 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 7748 SIZE_FORMAT, _mark_stack->capacity()); 7749 } 7750 assert(_mark_stack->isFull(), "Else push should have succeeded"); 7751 handle_stack_overflow(addr); 7752 } 7753 // anything including and to the right of _finger 7754 // will be scanned as we iterate over the remainder of the 7755 // bit map 7756 } 7757 } 7758 7759 PushOrMarkClosure::PushOrMarkClosure(CMSCollector* collector, 7760 MemRegion span, 7761 CMSBitMap* bitMap, CMSMarkStack* markStack, 7762 HeapWord* finger, MarkFromRootsClosure* parent) : 7763 CMSOopClosure(collector->ref_processor()), 7764 _collector(collector), 7765 _span(span), 7766 _bitMap(bitMap), 7767 _markStack(markStack), 7768 _finger(finger), 7769 _parent(parent) 7770 { } 7771 7772 Par_PushOrMarkClosure::Par_PushOrMarkClosure(CMSCollector* collector, 7773 MemRegion span, 7774 CMSBitMap* bit_map, 7775 OopTaskQueue* work_queue, 7776 CMSMarkStack* overflow_stack, 7777 HeapWord* finger, 7778 HeapWord** global_finger_addr, 7779 Par_MarkFromRootsClosure* parent) : 7780 CMSOopClosure(collector->ref_processor()), 7781 _collector(collector), 7782 _whole_span(collector->_span), 7783 _span(span), 7784 _bit_map(bit_map), 7785 _work_queue(work_queue), 7786 _overflow_stack(overflow_stack), 7787 _finger(finger), 7788 _global_finger_addr(global_finger_addr), 7789 _parent(parent) 7790 { } 7791 7792 // Assumes thread-safe access by callers, who are 7793 // responsible for mutual exclusion. 7794 void CMSCollector::lower_restart_addr(HeapWord* low) { 7795 assert(_span.contains(low), "Out of bounds addr"); 7796 if (_restart_addr == NULL) { 7797 _restart_addr = low; 7798 } else { 7799 _restart_addr = MIN2(_restart_addr, low); 7800 } 7801 } 7802 7803 // Upon stack overflow, we discard (part of) the stack, 7804 // remembering the least address amongst those discarded 7805 // in CMSCollector's _restart_address. 7806 void PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 7807 // Remember the least grey address discarded 7808 HeapWord* ra = (HeapWord*)_markStack->least_value(lost); 7809 _collector->lower_restart_addr(ra); 7810 _markStack->reset(); // discard stack contents 7811 _markStack->expand(); // expand the stack if possible 7812 } 7813 7814 // Upon stack overflow, we discard (part of) the stack, 7815 // remembering the least address amongst those discarded 7816 // in CMSCollector's _restart_address. 7817 void Par_PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 7818 // We need to do this under a mutex to prevent other 7819 // workers from interfering with the work done below. 7820 MutexLockerEx ml(_overflow_stack->par_lock(), 7821 Mutex::_no_safepoint_check_flag); 7822 // Remember the least grey address discarded 7823 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 7824 _collector->lower_restart_addr(ra); 7825 _overflow_stack->reset(); // discard stack contents 7826 _overflow_stack->expand(); // expand the stack if possible 7827 } 7828 7829 void CMKlassClosure::do_klass(Klass* k) { 7830 assert(_oop_closure != NULL, "Not initialized?"); 7831 k->oops_do(_oop_closure); 7832 } 7833 7834 void PushOrMarkClosure::do_oop(oop obj) { 7835 // Ignore mark word because we are running concurrent with mutators. 7836 assert(obj->is_oop_or_null(true), "expected an oop or NULL"); 7837 HeapWord* addr = (HeapWord*)obj; 7838 if (_span.contains(addr) && !_bitMap->isMarked(addr)) { 7839 // Oop lies in _span and isn't yet grey or black 7840 _bitMap->mark(addr); // now grey 7841 if (addr < _finger) { 7842 // the bit map iteration has already either passed, or 7843 // sampled, this bit in the bit map; we'll need to 7844 // use the marking stack to scan this oop's oops. 7845 bool simulate_overflow = false; 7846 NOT_PRODUCT( 7847 if (CMSMarkStackOverflowALot && 7848 _collector->simulate_overflow()) { 7849 // simulate a stack overflow 7850 simulate_overflow = true; 7851 } 7852 ) 7853 if (simulate_overflow || !_markStack->push(obj)) { // stack overflow 7854 if (PrintCMSStatistics != 0) { 7855 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 7856 SIZE_FORMAT, _markStack->capacity()); 7857 } 7858 assert(simulate_overflow || _markStack->isFull(), "Else push should have succeeded"); 7859 handle_stack_overflow(addr); 7860 } 7861 } 7862 // anything including and to the right of _finger 7863 // will be scanned as we iterate over the remainder of the 7864 // bit map 7865 do_yield_check(); 7866 } 7867 } 7868 7869 void PushOrMarkClosure::do_oop(oop* p) { PushOrMarkClosure::do_oop_work(p); } 7870 void PushOrMarkClosure::do_oop(narrowOop* p) { PushOrMarkClosure::do_oop_work(p); } 7871 7872 void Par_PushOrMarkClosure::do_oop(oop obj) { 7873 // Ignore mark word because we are running concurrent with mutators. 7874 assert(obj->is_oop_or_null(true), "expected an oop or NULL"); 7875 HeapWord* addr = (HeapWord*)obj; 7876 if (_whole_span.contains(addr) && !_bit_map->isMarked(addr)) { 7877 // Oop lies in _span and isn't yet grey or black 7878 // We read the global_finger (volatile read) strictly after marking oop 7879 bool res = _bit_map->par_mark(addr); // now grey 7880 volatile HeapWord** gfa = (volatile HeapWord**)_global_finger_addr; 7881 // Should we push this marked oop on our stack? 7882 // -- if someone else marked it, nothing to do 7883 // -- if target oop is above global finger nothing to do 7884 // -- if target oop is in chunk and above local finger 7885 // then nothing to do 7886 // -- else push on work queue 7887 if ( !res // someone else marked it, they will deal with it 7888 || (addr >= *gfa) // will be scanned in a later task 7889 || (_span.contains(addr) && addr >= _finger)) { // later in this chunk 7890 return; 7891 } 7892 // the bit map iteration has already either passed, or 7893 // sampled, this bit in the bit map; we'll need to 7894 // use the marking stack to scan this oop's oops. 7895 bool simulate_overflow = false; 7896 NOT_PRODUCT( 7897 if (CMSMarkStackOverflowALot && 7898 _collector->simulate_overflow()) { 7899 // simulate a stack overflow 7900 simulate_overflow = true; 7901 } 7902 ) 7903 if (simulate_overflow || 7904 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) { 7905 // stack overflow 7906 if (PrintCMSStatistics != 0) { 7907 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 7908 SIZE_FORMAT, _overflow_stack->capacity()); 7909 } 7910 // We cannot assert that the overflow stack is full because 7911 // it may have been emptied since. 7912 assert(simulate_overflow || 7913 _work_queue->size() == _work_queue->max_elems(), 7914 "Else push should have succeeded"); 7915 handle_stack_overflow(addr); 7916 } 7917 do_yield_check(); 7918 } 7919 } 7920 7921 void Par_PushOrMarkClosure::do_oop(oop* p) { Par_PushOrMarkClosure::do_oop_work(p); } 7922 void Par_PushOrMarkClosure::do_oop(narrowOop* p) { Par_PushOrMarkClosure::do_oop_work(p); } 7923 7924 PushAndMarkClosure::PushAndMarkClosure(CMSCollector* collector, 7925 MemRegion span, 7926 ReferenceProcessor* rp, 7927 CMSBitMap* bit_map, 7928 CMSBitMap* mod_union_table, 7929 CMSMarkStack* mark_stack, 7930 bool concurrent_precleaning): 7931 CMSOopClosure(rp), 7932 _collector(collector), 7933 _span(span), 7934 _bit_map(bit_map), 7935 _mod_union_table(mod_union_table), 7936 _mark_stack(mark_stack), 7937 _concurrent_precleaning(concurrent_precleaning) 7938 { 7939 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 7940 } 7941 7942 // Grey object rescan during pre-cleaning and second checkpoint phases -- 7943 // the non-parallel version (the parallel version appears further below.) 7944 void PushAndMarkClosure::do_oop(oop obj) { 7945 // Ignore mark word verification. If during concurrent precleaning, 7946 // the object monitor may be locked. If during the checkpoint 7947 // phases, the object may already have been reached by a different 7948 // path and may be at the end of the global overflow list (so 7949 // the mark word may be NULL). 7950 assert(obj->is_oop_or_null(true /* ignore mark word */), 7951 "expected an oop or NULL"); 7952 HeapWord* addr = (HeapWord*)obj; 7953 // Check if oop points into the CMS generation 7954 // and is not marked 7955 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 7956 // a white object ... 7957 _bit_map->mark(addr); // ... now grey 7958 // push on the marking stack (grey set) 7959 bool simulate_overflow = false; 7960 NOT_PRODUCT( 7961 if (CMSMarkStackOverflowALot && 7962 _collector->simulate_overflow()) { 7963 // simulate a stack overflow 7964 simulate_overflow = true; 7965 } 7966 ) 7967 if (simulate_overflow || !_mark_stack->push(obj)) { 7968 if (_concurrent_precleaning) { 7969 // During precleaning we can just dirty the appropriate card(s) 7970 // in the mod union table, thus ensuring that the object remains 7971 // in the grey set and continue. In the case of object arrays 7972 // we need to dirty all of the cards that the object spans, 7973 // since the rescan of object arrays will be limited to the 7974 // dirty cards. 7975 // Note that no one can be interfering with us in this action 7976 // of dirtying the mod union table, so no locking or atomics 7977 // are required. 7978 if (obj->is_objArray()) { 7979 size_t sz = obj->size(); 7980 HeapWord* end_card_addr = (HeapWord*)round_to( 7981 (intptr_t)(addr+sz), CardTableModRefBS::card_size); 7982 MemRegion redirty_range = MemRegion(addr, end_card_addr); 7983 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 7984 _mod_union_table->mark_range(redirty_range); 7985 } else { 7986 _mod_union_table->mark(addr); 7987 } 7988 _collector->_ser_pmc_preclean_ovflw++; 7989 } else { 7990 // During the remark phase, we need to remember this oop 7991 // in the overflow list. 7992 _collector->push_on_overflow_list(obj); 7993 _collector->_ser_pmc_remark_ovflw++; 7994 } 7995 } 7996 } 7997 } 7998 7999 Par_PushAndMarkClosure::Par_PushAndMarkClosure(CMSCollector* collector, 8000 MemRegion span, 8001 ReferenceProcessor* rp, 8002 CMSBitMap* bit_map, 8003 OopTaskQueue* work_queue): 8004 CMSOopClosure(rp), 8005 _collector(collector), 8006 _span(span), 8007 _bit_map(bit_map), 8008 _work_queue(work_queue) 8009 { 8010 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 8011 } 8012 8013 void PushAndMarkClosure::do_oop(oop* p) { PushAndMarkClosure::do_oop_work(p); } 8014 void PushAndMarkClosure::do_oop(narrowOop* p) { PushAndMarkClosure::do_oop_work(p); } 8015 8016 // Grey object rescan during second checkpoint phase -- 8017 // the parallel version. 8018 void Par_PushAndMarkClosure::do_oop(oop obj) { 8019 // In the assert below, we ignore the mark word because 8020 // this oop may point to an already visited object that is 8021 // on the overflow stack (in which case the mark word has 8022 // been hijacked for chaining into the overflow stack -- 8023 // if this is the last object in the overflow stack then 8024 // its mark word will be NULL). Because this object may 8025 // have been subsequently popped off the global overflow 8026 // stack, and the mark word possibly restored to the prototypical 8027 // value, by the time we get to examined this failing assert in 8028 // the debugger, is_oop_or_null(false) may subsequently start 8029 // to hold. 8030 assert(obj->is_oop_or_null(true), 8031 "expected an oop or NULL"); 8032 HeapWord* addr = (HeapWord*)obj; 8033 // Check if oop points into the CMS generation 8034 // and is not marked 8035 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 8036 // a white object ... 8037 // If we manage to "claim" the object, by being the 8038 // first thread to mark it, then we push it on our 8039 // marking stack 8040 if (_bit_map->par_mark(addr)) { // ... now grey 8041 // push on work queue (grey set) 8042 bool simulate_overflow = false; 8043 NOT_PRODUCT( 8044 if (CMSMarkStackOverflowALot && 8045 _collector->par_simulate_overflow()) { 8046 // simulate a stack overflow 8047 simulate_overflow = true; 8048 } 8049 ) 8050 if (simulate_overflow || !_work_queue->push(obj)) { 8051 _collector->par_push_on_overflow_list(obj); 8052 _collector->_par_pmc_remark_ovflw++; // imprecise OK: no need to CAS 8053 } 8054 } // Else, some other thread got there first 8055 } 8056 } 8057 8058 void Par_PushAndMarkClosure::do_oop(oop* p) { Par_PushAndMarkClosure::do_oop_work(p); } 8059 void Par_PushAndMarkClosure::do_oop(narrowOop* p) { Par_PushAndMarkClosure::do_oop_work(p); } 8060 8061 void CMSPrecleanRefsYieldClosure::do_yield_work() { 8062 Mutex* bml = _collector->bitMapLock(); 8063 assert_lock_strong(bml); 8064 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 8065 "CMS thread should hold CMS token"); 8066 8067 bml->unlock(); 8068 ConcurrentMarkSweepThread::desynchronize(true); 8069 8070 ConcurrentMarkSweepThread::acknowledge_yield_request(); 8071 8072 _collector->stopTimer(); 8073 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 8074 if (PrintCMSStatistics != 0) { 8075 _collector->incrementYields(); 8076 } 8077 _collector->icms_wait(); 8078 8079 // See the comment in coordinator_yield() 8080 for (unsigned i = 0; i < CMSYieldSleepCount && 8081 ConcurrentMarkSweepThread::should_yield() && 8082 !CMSCollector::foregroundGCIsActive(); ++i) { 8083 os::sleep(Thread::current(), 1, false); 8084 ConcurrentMarkSweepThread::acknowledge_yield_request(); 8085 } 8086 8087 ConcurrentMarkSweepThread::synchronize(true); 8088 bml->lock(); 8089 8090 _collector->startTimer(); 8091 } 8092 8093 bool CMSPrecleanRefsYieldClosure::should_return() { 8094 if (ConcurrentMarkSweepThread::should_yield()) { 8095 do_yield_work(); 8096 } 8097 return _collector->foregroundGCIsActive(); 8098 } 8099 8100 void MarkFromDirtyCardsClosure::do_MemRegion(MemRegion mr) { 8101 assert(((size_t)mr.start())%CardTableModRefBS::card_size_in_words == 0, 8102 "mr should be aligned to start at a card boundary"); 8103 // We'd like to assert: 8104 // assert(mr.word_size()%CardTableModRefBS::card_size_in_words == 0, 8105 // "mr should be a range of cards"); 8106 // However, that would be too strong in one case -- the last 8107 // partition ends at _unallocated_block which, in general, can be 8108 // an arbitrary boundary, not necessarily card aligned. 8109 if (PrintCMSStatistics != 0) { 8110 _num_dirty_cards += 8111 mr.word_size()/CardTableModRefBS::card_size_in_words; 8112 } 8113 _space->object_iterate_mem(mr, &_scan_cl); 8114 } 8115 8116 SweepClosure::SweepClosure(CMSCollector* collector, 8117 ConcurrentMarkSweepGeneration* g, 8118 CMSBitMap* bitMap, bool should_yield) : 8119 _collector(collector), 8120 _g(g), 8121 _sp(g->cmsSpace()), 8122 _limit(_sp->sweep_limit()), 8123 _freelistLock(_sp->freelistLock()), 8124 _bitMap(bitMap), 8125 _yield(should_yield), 8126 _inFreeRange(false), // No free range at beginning of sweep 8127 _freeRangeInFreeLists(false), // No free range at beginning of sweep 8128 _lastFreeRangeCoalesced(false), 8129 _freeFinger(g->used_region().start()) 8130 { 8131 NOT_PRODUCT( 8132 _numObjectsFreed = 0; 8133 _numWordsFreed = 0; 8134 _numObjectsLive = 0; 8135 _numWordsLive = 0; 8136 _numObjectsAlreadyFree = 0; 8137 _numWordsAlreadyFree = 0; 8138 _last_fc = NULL; 8139 8140 _sp->initializeIndexedFreeListArrayReturnedBytes(); 8141 _sp->dictionary()->initialize_dict_returned_bytes(); 8142 ) 8143 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 8144 "sweep _limit out of bounds"); 8145 if (CMSTraceSweeper) { 8146 gclog_or_tty->print_cr("\n====================\nStarting new sweep with limit " PTR_FORMAT, 8147 _limit); 8148 } 8149 } 8150 8151 void SweepClosure::print_on(outputStream* st) const { 8152 tty->print_cr("_sp = [" PTR_FORMAT "," PTR_FORMAT ")", 8153 _sp->bottom(), _sp->end()); 8154 tty->print_cr("_limit = " PTR_FORMAT, _limit); 8155 tty->print_cr("_freeFinger = " PTR_FORMAT, _freeFinger); 8156 NOT_PRODUCT(tty->print_cr("_last_fc = " PTR_FORMAT, _last_fc);) 8157 tty->print_cr("_inFreeRange = %d, _freeRangeInFreeLists = %d, _lastFreeRangeCoalesced = %d", 8158 _inFreeRange, _freeRangeInFreeLists, _lastFreeRangeCoalesced); 8159 } 8160 8161 #ifndef PRODUCT 8162 // Assertion checking only: no useful work in product mode -- 8163 // however, if any of the flags below become product flags, 8164 // you may need to review this code to see if it needs to be 8165 // enabled in product mode. 8166 SweepClosure::~SweepClosure() { 8167 assert_lock_strong(_freelistLock); 8168 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 8169 "sweep _limit out of bounds"); 8170 if (inFreeRange()) { 8171 warning("inFreeRange() should have been reset; dumping state of SweepClosure"); 8172 print(); 8173 ShouldNotReachHere(); 8174 } 8175 if (Verbose && PrintGC) { 8176 gclog_or_tty->print("Collected "SIZE_FORMAT" objects, " SIZE_FORMAT " bytes", 8177 _numObjectsFreed, _numWordsFreed*sizeof(HeapWord)); 8178 gclog_or_tty->print_cr("\nLive "SIZE_FORMAT" objects, " 8179 SIZE_FORMAT" bytes " 8180 "Already free "SIZE_FORMAT" objects, "SIZE_FORMAT" bytes", 8181 _numObjectsLive, _numWordsLive*sizeof(HeapWord), 8182 _numObjectsAlreadyFree, _numWordsAlreadyFree*sizeof(HeapWord)); 8183 size_t totalBytes = (_numWordsFreed + _numWordsLive + _numWordsAlreadyFree) 8184 * sizeof(HeapWord); 8185 gclog_or_tty->print_cr("Total sweep: "SIZE_FORMAT" bytes", totalBytes); 8186 8187 if (PrintCMSStatistics && CMSVerifyReturnedBytes) { 8188 size_t indexListReturnedBytes = _sp->sumIndexedFreeListArrayReturnedBytes(); 8189 size_t dict_returned_bytes = _sp->dictionary()->sum_dict_returned_bytes(); 8190 size_t returned_bytes = indexListReturnedBytes + dict_returned_bytes; 8191 gclog_or_tty->print("Returned "SIZE_FORMAT" bytes", returned_bytes); 8192 gclog_or_tty->print(" Indexed List Returned "SIZE_FORMAT" bytes", 8193 indexListReturnedBytes); 8194 gclog_or_tty->print_cr(" Dictionary Returned "SIZE_FORMAT" bytes", 8195 dict_returned_bytes); 8196 } 8197 } 8198 if (CMSTraceSweeper) { 8199 gclog_or_tty->print_cr("end of sweep with _limit = " PTR_FORMAT "\n================", 8200 _limit); 8201 } 8202 } 8203 #endif // PRODUCT 8204 8205 void SweepClosure::initialize_free_range(HeapWord* freeFinger, 8206 bool freeRangeInFreeLists) { 8207 if (CMSTraceSweeper) { 8208 gclog_or_tty->print("---- Start free range at " PTR_FORMAT " with free block (%d)\n", 8209 freeFinger, freeRangeInFreeLists); 8210 } 8211 assert(!inFreeRange(), "Trampling existing free range"); 8212 set_inFreeRange(true); 8213 set_lastFreeRangeCoalesced(false); 8214 8215 set_freeFinger(freeFinger); 8216 set_freeRangeInFreeLists(freeRangeInFreeLists); 8217 if (CMSTestInFreeList) { 8218 if (freeRangeInFreeLists) { 8219 FreeChunk* fc = (FreeChunk*) freeFinger; 8220 assert(fc->is_free(), "A chunk on the free list should be free."); 8221 assert(fc->size() > 0, "Free range should have a size"); 8222 assert(_sp->verify_chunk_in_free_list(fc), "Chunk is not in free lists"); 8223 } 8224 } 8225 } 8226 8227 // Note that the sweeper runs concurrently with mutators. Thus, 8228 // it is possible for direct allocation in this generation to happen 8229 // in the middle of the sweep. Note that the sweeper also coalesces 8230 // contiguous free blocks. Thus, unless the sweeper and the allocator 8231 // synchronize appropriately freshly allocated blocks may get swept up. 8232 // This is accomplished by the sweeper locking the free lists while 8233 // it is sweeping. Thus blocks that are determined to be free are 8234 // indeed free. There is however one additional complication: 8235 // blocks that have been allocated since the final checkpoint and 8236 // mark, will not have been marked and so would be treated as 8237 // unreachable and swept up. To prevent this, the allocator marks 8238 // the bit map when allocating during the sweep phase. This leads, 8239 // however, to a further complication -- objects may have been allocated 8240 // but not yet initialized -- in the sense that the header isn't yet 8241 // installed. The sweeper can not then determine the size of the block 8242 // in order to skip over it. To deal with this case, we use a technique 8243 // (due to Printezis) to encode such uninitialized block sizes in the 8244 // bit map. Since the bit map uses a bit per every HeapWord, but the 8245 // CMS generation has a minimum object size of 3 HeapWords, it follows 8246 // that "normal marks" won't be adjacent in the bit map (there will 8247 // always be at least two 0 bits between successive 1 bits). We make use 8248 // of these "unused" bits to represent uninitialized blocks -- the bit 8249 // corresponding to the start of the uninitialized object and the next 8250 // bit are both set. Finally, a 1 bit marks the end of the object that 8251 // started with the two consecutive 1 bits to indicate its potentially 8252 // uninitialized state. 8253 8254 size_t SweepClosure::do_blk_careful(HeapWord* addr) { 8255 FreeChunk* fc = (FreeChunk*)addr; 8256 size_t res; 8257 8258 // Check if we are done sweeping. Below we check "addr >= _limit" rather 8259 // than "addr == _limit" because although _limit was a block boundary when 8260 // we started the sweep, it may no longer be one because heap expansion 8261 // may have caused us to coalesce the block ending at the address _limit 8262 // with a newly expanded chunk (this happens when _limit was set to the 8263 // previous _end of the space), so we may have stepped past _limit: 8264 // see the following Zeno-like trail of CRs 6977970, 7008136, 7042740. 8265 if (addr >= _limit) { // we have swept up to or past the limit: finish up 8266 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 8267 "sweep _limit out of bounds"); 8268 assert(addr < _sp->end(), "addr out of bounds"); 8269 // Flush any free range we might be holding as a single 8270 // coalesced chunk to the appropriate free list. 8271 if (inFreeRange()) { 8272 assert(freeFinger() >= _sp->bottom() && freeFinger() < _limit, 8273 err_msg("freeFinger() " PTR_FORMAT" is out-of-bounds", freeFinger())); 8274 flush_cur_free_chunk(freeFinger(), 8275 pointer_delta(addr, freeFinger())); 8276 if (CMSTraceSweeper) { 8277 gclog_or_tty->print("Sweep: last chunk: "); 8278 gclog_or_tty->print("put_free_blk " PTR_FORMAT " ("SIZE_FORMAT") " 8279 "[coalesced:%d]\n", 8280 freeFinger(), pointer_delta(addr, freeFinger()), 8281 lastFreeRangeCoalesced() ? 1 : 0); 8282 } 8283 } 8284 8285 // help the iterator loop finish 8286 return pointer_delta(_sp->end(), addr); 8287 } 8288 8289 assert(addr < _limit, "sweep invariant"); 8290 // check if we should yield 8291 do_yield_check(addr); 8292 if (fc->is_free()) { 8293 // Chunk that is already free 8294 res = fc->size(); 8295 do_already_free_chunk(fc); 8296 debug_only(_sp->verifyFreeLists()); 8297 // If we flush the chunk at hand in lookahead_and_flush() 8298 // and it's coalesced with a preceding chunk, then the 8299 // process of "mangling" the payload of the coalesced block 8300 // will cause erasure of the size information from the 8301 // (erstwhile) header of all the coalesced blocks but the 8302 // first, so the first disjunct in the assert will not hold 8303 // in that specific case (in which case the second disjunct 8304 // will hold). 8305 assert(res == fc->size() || ((HeapWord*)fc) + res >= _limit, 8306 "Otherwise the size info doesn't change at this step"); 8307 NOT_PRODUCT( 8308 _numObjectsAlreadyFree++; 8309 _numWordsAlreadyFree += res; 8310 ) 8311 NOT_PRODUCT(_last_fc = fc;) 8312 } else if (!_bitMap->isMarked(addr)) { 8313 // Chunk is fresh garbage 8314 res = do_garbage_chunk(fc); 8315 debug_only(_sp->verifyFreeLists()); 8316 NOT_PRODUCT( 8317 _numObjectsFreed++; 8318 _numWordsFreed += res; 8319 ) 8320 } else { 8321 // Chunk that is alive. 8322 res = do_live_chunk(fc); 8323 debug_only(_sp->verifyFreeLists()); 8324 NOT_PRODUCT( 8325 _numObjectsLive++; 8326 _numWordsLive += res; 8327 ) 8328 } 8329 return res; 8330 } 8331 8332 // For the smart allocation, record following 8333 // split deaths - a free chunk is removed from its free list because 8334 // it is being split into two or more chunks. 8335 // split birth - a free chunk is being added to its free list because 8336 // a larger free chunk has been split and resulted in this free chunk. 8337 // coal death - a free chunk is being removed from its free list because 8338 // it is being coalesced into a large free chunk. 8339 // coal birth - a free chunk is being added to its free list because 8340 // it was created when two or more free chunks where coalesced into 8341 // this free chunk. 8342 // 8343 // These statistics are used to determine the desired number of free 8344 // chunks of a given size. The desired number is chosen to be relative 8345 // to the end of a CMS sweep. The desired number at the end of a sweep 8346 // is the 8347 // count-at-end-of-previous-sweep (an amount that was enough) 8348 // - count-at-beginning-of-current-sweep (the excess) 8349 // + split-births (gains in this size during interval) 8350 // - split-deaths (demands on this size during interval) 8351 // where the interval is from the end of one sweep to the end of the 8352 // next. 8353 // 8354 // When sweeping the sweeper maintains an accumulated chunk which is 8355 // the chunk that is made up of chunks that have been coalesced. That 8356 // will be termed the left-hand chunk. A new chunk of garbage that 8357 // is being considered for coalescing will be referred to as the 8358 // right-hand chunk. 8359 // 8360 // When making a decision on whether to coalesce a right-hand chunk with 8361 // the current left-hand chunk, the current count vs. the desired count 8362 // of the left-hand chunk is considered. Also if the right-hand chunk 8363 // is near the large chunk at the end of the heap (see 8364 // ConcurrentMarkSweepGeneration::isNearLargestChunk()), then the 8365 // left-hand chunk is coalesced. 8366 // 8367 // When making a decision about whether to split a chunk, the desired count 8368 // vs. the current count of the candidate to be split is also considered. 8369 // If the candidate is underpopulated (currently fewer chunks than desired) 8370 // a chunk of an overpopulated (currently more chunks than desired) size may 8371 // be chosen. The "hint" associated with a free list, if non-null, points 8372 // to a free list which may be overpopulated. 8373 // 8374 8375 void SweepClosure::do_already_free_chunk(FreeChunk* fc) { 8376 const size_t size = fc->size(); 8377 // Chunks that cannot be coalesced are not in the 8378 // free lists. 8379 if (CMSTestInFreeList && !fc->cantCoalesce()) { 8380 assert(_sp->verify_chunk_in_free_list(fc), 8381 "free chunk should be in free lists"); 8382 } 8383 // a chunk that is already free, should not have been 8384 // marked in the bit map 8385 HeapWord* const addr = (HeapWord*) fc; 8386 assert(!_bitMap->isMarked(addr), "free chunk should be unmarked"); 8387 // Verify that the bit map has no bits marked between 8388 // addr and purported end of this block. 8389 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 8390 8391 // Some chunks cannot be coalesced under any circumstances. 8392 // See the definition of cantCoalesce(). 8393 if (!fc->cantCoalesce()) { 8394 // This chunk can potentially be coalesced. 8395 if (_sp->adaptive_freelists()) { 8396 // All the work is done in 8397 do_post_free_or_garbage_chunk(fc, size); 8398 } else { // Not adaptive free lists 8399 // this is a free chunk that can potentially be coalesced by the sweeper; 8400 if (!inFreeRange()) { 8401 // if the next chunk is a free block that can't be coalesced 8402 // it doesn't make sense to remove this chunk from the free lists 8403 FreeChunk* nextChunk = (FreeChunk*)(addr + size); 8404 assert((HeapWord*)nextChunk <= _sp->end(), "Chunk size out of bounds?"); 8405 if ((HeapWord*)nextChunk < _sp->end() && // There is another free chunk to the right ... 8406 nextChunk->is_free() && // ... which is free... 8407 nextChunk->cantCoalesce()) { // ... but can't be coalesced 8408 // nothing to do 8409 } else { 8410 // Potentially the start of a new free range: 8411 // Don't eagerly remove it from the free lists. 8412 // No need to remove it if it will just be put 8413 // back again. (Also from a pragmatic point of view 8414 // if it is a free block in a region that is beyond 8415 // any allocated blocks, an assertion will fail) 8416 // Remember the start of a free run. 8417 initialize_free_range(addr, true); 8418 // end - can coalesce with next chunk 8419 } 8420 } else { 8421 // the midst of a free range, we are coalescing 8422 print_free_block_coalesced(fc); 8423 if (CMSTraceSweeper) { 8424 gclog_or_tty->print(" -- pick up free block " PTR_FORMAT " (" SIZE_FORMAT ")\n", fc, size); 8425 } 8426 // remove it from the free lists 8427 _sp->removeFreeChunkFromFreeLists(fc); 8428 set_lastFreeRangeCoalesced(true); 8429 // If the chunk is being coalesced and the current free range is 8430 // in the free lists, remove the current free range so that it 8431 // will be returned to the free lists in its entirety - all 8432 // the coalesced pieces included. 8433 if (freeRangeInFreeLists()) { 8434 FreeChunk* ffc = (FreeChunk*) freeFinger(); 8435 assert(ffc->size() == pointer_delta(addr, freeFinger()), 8436 "Size of free range is inconsistent with chunk size."); 8437 if (CMSTestInFreeList) { 8438 assert(_sp->verify_chunk_in_free_list(ffc), 8439 "free range is not in free lists"); 8440 } 8441 _sp->removeFreeChunkFromFreeLists(ffc); 8442 set_freeRangeInFreeLists(false); 8443 } 8444 } 8445 } 8446 // Note that if the chunk is not coalescable (the else arm 8447 // below), we unconditionally flush, without needing to do 8448 // a "lookahead," as we do below. 8449 if (inFreeRange()) lookahead_and_flush(fc, size); 8450 } else { 8451 // Code path common to both original and adaptive free lists. 8452 8453 // cant coalesce with previous block; this should be treated 8454 // as the end of a free run if any 8455 if (inFreeRange()) { 8456 // we kicked some butt; time to pick up the garbage 8457 assert(freeFinger() < addr, "freeFinger points too high"); 8458 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 8459 } 8460 // else, nothing to do, just continue 8461 } 8462 } 8463 8464 size_t SweepClosure::do_garbage_chunk(FreeChunk* fc) { 8465 // This is a chunk of garbage. It is not in any free list. 8466 // Add it to a free list or let it possibly be coalesced into 8467 // a larger chunk. 8468 HeapWord* const addr = (HeapWord*) fc; 8469 const size_t size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 8470 8471 if (_sp->adaptive_freelists()) { 8472 // Verify that the bit map has no bits marked between 8473 // addr and purported end of just dead object. 8474 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 8475 8476 do_post_free_or_garbage_chunk(fc, size); 8477 } else { 8478 if (!inFreeRange()) { 8479 // start of a new free range 8480 assert(size > 0, "A free range should have a size"); 8481 initialize_free_range(addr, false); 8482 } else { 8483 // this will be swept up when we hit the end of the 8484 // free range 8485 if (CMSTraceSweeper) { 8486 gclog_or_tty->print(" -- pick up garbage " PTR_FORMAT " (" SIZE_FORMAT ")\n", fc, size); 8487 } 8488 // If the chunk is being coalesced and the current free range is 8489 // in the free lists, remove the current free range so that it 8490 // will be returned to the free lists in its entirety - all 8491 // the coalesced pieces included. 8492 if (freeRangeInFreeLists()) { 8493 FreeChunk* ffc = (FreeChunk*)freeFinger(); 8494 assert(ffc->size() == pointer_delta(addr, freeFinger()), 8495 "Size of free range is inconsistent with chunk size."); 8496 if (CMSTestInFreeList) { 8497 assert(_sp->verify_chunk_in_free_list(ffc), 8498 "free range is not in free lists"); 8499 } 8500 _sp->removeFreeChunkFromFreeLists(ffc); 8501 set_freeRangeInFreeLists(false); 8502 } 8503 set_lastFreeRangeCoalesced(true); 8504 } 8505 // this will be swept up when we hit the end of the free range 8506 8507 // Verify that the bit map has no bits marked between 8508 // addr and purported end of just dead object. 8509 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 8510 } 8511 assert(_limit >= addr + size, 8512 "A freshly garbage chunk can't possibly straddle over _limit"); 8513 if (inFreeRange()) lookahead_and_flush(fc, size); 8514 return size; 8515 } 8516 8517 size_t SweepClosure::do_live_chunk(FreeChunk* fc) { 8518 HeapWord* addr = (HeapWord*) fc; 8519 // The sweeper has just found a live object. Return any accumulated 8520 // left hand chunk to the free lists. 8521 if (inFreeRange()) { 8522 assert(freeFinger() < addr, "freeFinger points too high"); 8523 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 8524 } 8525 8526 // This object is live: we'd normally expect this to be 8527 // an oop, and like to assert the following: 8528 // assert(oop(addr)->is_oop(), "live block should be an oop"); 8529 // However, as we commented above, this may be an object whose 8530 // header hasn't yet been initialized. 8531 size_t size; 8532 assert(_bitMap->isMarked(addr), "Tautology for this control point"); 8533 if (_bitMap->isMarked(addr + 1)) { 8534 // Determine the size from the bit map, rather than trying to 8535 // compute it from the object header. 8536 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 8537 size = pointer_delta(nextOneAddr + 1, addr); 8538 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 8539 "alignment problem"); 8540 8541 #ifdef ASSERT 8542 if (oop(addr)->klass_or_null() != NULL) { 8543 // Ignore mark word because we are running concurrent with mutators 8544 assert(oop(addr)->is_oop(true), "live block should be an oop"); 8545 assert(size == 8546 CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()), 8547 "P-mark and computed size do not agree"); 8548 } 8549 #endif 8550 8551 } else { 8552 // This should be an initialized object that's alive. 8553 assert(oop(addr)->klass_or_null() != NULL, 8554 "Should be an initialized object"); 8555 // Ignore mark word because we are running concurrent with mutators 8556 assert(oop(addr)->is_oop(true), "live block should be an oop"); 8557 // Verify that the bit map has no bits marked between 8558 // addr and purported end of this block. 8559 size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 8560 assert(size >= 3, "Necessary for Printezis marks to work"); 8561 assert(!_bitMap->isMarked(addr+1), "Tautology for this control point"); 8562 DEBUG_ONLY(_bitMap->verifyNoOneBitsInRange(addr+2, addr+size);) 8563 } 8564 return size; 8565 } 8566 8567 void SweepClosure::do_post_free_or_garbage_chunk(FreeChunk* fc, 8568 size_t chunkSize) { 8569 // do_post_free_or_garbage_chunk() should only be called in the case 8570 // of the adaptive free list allocator. 8571 const bool fcInFreeLists = fc->is_free(); 8572 assert(_sp->adaptive_freelists(), "Should only be used in this case."); 8573 assert((HeapWord*)fc <= _limit, "sweep invariant"); 8574 if (CMSTestInFreeList && fcInFreeLists) { 8575 assert(_sp->verify_chunk_in_free_list(fc), "free chunk is not in free lists"); 8576 } 8577 8578 if (CMSTraceSweeper) { 8579 gclog_or_tty->print_cr(" -- pick up another chunk at " PTR_FORMAT " (" SIZE_FORMAT ")", fc, chunkSize); 8580 } 8581 8582 HeapWord* const fc_addr = (HeapWord*) fc; 8583 8584 bool coalesce; 8585 const size_t left = pointer_delta(fc_addr, freeFinger()); 8586 const size_t right = chunkSize; 8587 switch (FLSCoalescePolicy) { 8588 // numeric value forms a coalition aggressiveness metric 8589 case 0: { // never coalesce 8590 coalesce = false; 8591 break; 8592 } 8593 case 1: { // coalesce if left & right chunks on overpopulated lists 8594 coalesce = _sp->coalOverPopulated(left) && 8595 _sp->coalOverPopulated(right); 8596 break; 8597 } 8598 case 2: { // coalesce if left chunk on overpopulated list (default) 8599 coalesce = _sp->coalOverPopulated(left); 8600 break; 8601 } 8602 case 3: { // coalesce if left OR right chunk on overpopulated list 8603 coalesce = _sp->coalOverPopulated(left) || 8604 _sp->coalOverPopulated(right); 8605 break; 8606 } 8607 case 4: { // always coalesce 8608 coalesce = true; 8609 break; 8610 } 8611 default: 8612 ShouldNotReachHere(); 8613 } 8614 8615 // Should the current free range be coalesced? 8616 // If the chunk is in a free range and either we decided to coalesce above 8617 // or the chunk is near the large block at the end of the heap 8618 // (isNearLargestChunk() returns true), then coalesce this chunk. 8619 const bool doCoalesce = inFreeRange() 8620 && (coalesce || _g->isNearLargestChunk(fc_addr)); 8621 if (doCoalesce) { 8622 // Coalesce the current free range on the left with the new 8623 // chunk on the right. If either is on a free list, 8624 // it must be removed from the list and stashed in the closure. 8625 if (freeRangeInFreeLists()) { 8626 FreeChunk* const ffc = (FreeChunk*)freeFinger(); 8627 assert(ffc->size() == pointer_delta(fc_addr, freeFinger()), 8628 "Size of free range is inconsistent with chunk size."); 8629 if (CMSTestInFreeList) { 8630 assert(_sp->verify_chunk_in_free_list(ffc), 8631 "Chunk is not in free lists"); 8632 } 8633 _sp->coalDeath(ffc->size()); 8634 _sp->removeFreeChunkFromFreeLists(ffc); 8635 set_freeRangeInFreeLists(false); 8636 } 8637 if (fcInFreeLists) { 8638 _sp->coalDeath(chunkSize); 8639 assert(fc->size() == chunkSize, 8640 "The chunk has the wrong size or is not in the free lists"); 8641 _sp->removeFreeChunkFromFreeLists(fc); 8642 } 8643 set_lastFreeRangeCoalesced(true); 8644 print_free_block_coalesced(fc); 8645 } else { // not in a free range and/or should not coalesce 8646 // Return the current free range and start a new one. 8647 if (inFreeRange()) { 8648 // In a free range but cannot coalesce with the right hand chunk. 8649 // Put the current free range into the free lists. 8650 flush_cur_free_chunk(freeFinger(), 8651 pointer_delta(fc_addr, freeFinger())); 8652 } 8653 // Set up for new free range. Pass along whether the right hand 8654 // chunk is in the free lists. 8655 initialize_free_range((HeapWord*)fc, fcInFreeLists); 8656 } 8657 } 8658 8659 // Lookahead flush: 8660 // If we are tracking a free range, and this is the last chunk that 8661 // we'll look at because its end crosses past _limit, we'll preemptively 8662 // flush it along with any free range we may be holding on to. Note that 8663 // this can be the case only for an already free or freshly garbage 8664 // chunk. If this block is an object, it can never straddle 8665 // over _limit. The "straddling" occurs when _limit is set at 8666 // the previous end of the space when this cycle started, and 8667 // a subsequent heap expansion caused the previously co-terminal 8668 // free block to be coalesced with the newly expanded portion, 8669 // thus rendering _limit a non-block-boundary making it dangerous 8670 // for the sweeper to step over and examine. 8671 void SweepClosure::lookahead_and_flush(FreeChunk* fc, size_t chunk_size) { 8672 assert(inFreeRange(), "Should only be called if currently in a free range."); 8673 HeapWord* const eob = ((HeapWord*)fc) + chunk_size; 8674 assert(_sp->used_region().contains(eob - 1), 8675 err_msg("eob = " PTR_FORMAT " eob-1 = " PTR_FORMAT " _limit = " PTR_FORMAT 8676 " out of bounds wrt _sp = [" PTR_FORMAT "," PTR_FORMAT ")" 8677 " when examining fc = " PTR_FORMAT "(" SIZE_FORMAT ")", 8678 eob, eob-1, _limit, _sp->bottom(), _sp->end(), fc, chunk_size)); 8679 if (eob >= _limit) { 8680 assert(eob == _limit || fc->is_free(), "Only a free chunk should allow us to cross over the limit"); 8681 if (CMSTraceSweeper) { 8682 gclog_or_tty->print_cr("_limit " PTR_FORMAT " reached or crossed by block " 8683 "[" PTR_FORMAT "," PTR_FORMAT ") in space " 8684 "[" PTR_FORMAT "," PTR_FORMAT ")", 8685 _limit, fc, eob, _sp->bottom(), _sp->end()); 8686 } 8687 // Return the storage we are tracking back into the free lists. 8688 if (CMSTraceSweeper) { 8689 gclog_or_tty->print_cr("Flushing ... "); 8690 } 8691 assert(freeFinger() < eob, "Error"); 8692 flush_cur_free_chunk( freeFinger(), pointer_delta(eob, freeFinger())); 8693 } 8694 } 8695 8696 void SweepClosure::flush_cur_free_chunk(HeapWord* chunk, size_t size) { 8697 assert(inFreeRange(), "Should only be called if currently in a free range."); 8698 assert(size > 0, 8699 "A zero sized chunk cannot be added to the free lists."); 8700 if (!freeRangeInFreeLists()) { 8701 if (CMSTestInFreeList) { 8702 FreeChunk* fc = (FreeChunk*) chunk; 8703 fc->set_size(size); 8704 assert(!_sp->verify_chunk_in_free_list(fc), 8705 "chunk should not be in free lists yet"); 8706 } 8707 if (CMSTraceSweeper) { 8708 gclog_or_tty->print_cr(" -- add free block " PTR_FORMAT " (" SIZE_FORMAT ") to free lists", 8709 chunk, size); 8710 } 8711 // A new free range is going to be starting. The current 8712 // free range has not been added to the free lists yet or 8713 // was removed so add it back. 8714 // If the current free range was coalesced, then the death 8715 // of the free range was recorded. Record a birth now. 8716 if (lastFreeRangeCoalesced()) { 8717 _sp->coalBirth(size); 8718 } 8719 _sp->addChunkAndRepairOffsetTable(chunk, size, 8720 lastFreeRangeCoalesced()); 8721 } else if (CMSTraceSweeper) { 8722 gclog_or_tty->print_cr("Already in free list: nothing to flush"); 8723 } 8724 set_inFreeRange(false); 8725 set_freeRangeInFreeLists(false); 8726 } 8727 8728 // We take a break if we've been at this for a while, 8729 // so as to avoid monopolizing the locks involved. 8730 void SweepClosure::do_yield_work(HeapWord* addr) { 8731 // Return current free chunk being used for coalescing (if any) 8732 // to the appropriate freelist. After yielding, the next 8733 // free block encountered will start a coalescing range of 8734 // free blocks. If the next free block is adjacent to the 8735 // chunk just flushed, they will need to wait for the next 8736 // sweep to be coalesced. 8737 if (inFreeRange()) { 8738 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 8739 } 8740 8741 // First give up the locks, then yield, then re-lock. 8742 // We should probably use a constructor/destructor idiom to 8743 // do this unlock/lock or modify the MutexUnlocker class to 8744 // serve our purpose. XXX 8745 assert_lock_strong(_bitMap->lock()); 8746 assert_lock_strong(_freelistLock); 8747 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 8748 "CMS thread should hold CMS token"); 8749 _bitMap->lock()->unlock(); 8750 _freelistLock->unlock(); 8751 ConcurrentMarkSweepThread::desynchronize(true); 8752 ConcurrentMarkSweepThread::acknowledge_yield_request(); 8753 _collector->stopTimer(); 8754 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 8755 if (PrintCMSStatistics != 0) { 8756 _collector->incrementYields(); 8757 } 8758 _collector->icms_wait(); 8759 8760 // See the comment in coordinator_yield() 8761 for (unsigned i = 0; i < CMSYieldSleepCount && 8762 ConcurrentMarkSweepThread::should_yield() && 8763 !CMSCollector::foregroundGCIsActive(); ++i) { 8764 os::sleep(Thread::current(), 1, false); 8765 ConcurrentMarkSweepThread::acknowledge_yield_request(); 8766 } 8767 8768 ConcurrentMarkSweepThread::synchronize(true); 8769 _freelistLock->lock(); 8770 _bitMap->lock()->lock_without_safepoint_check(); 8771 _collector->startTimer(); 8772 } 8773 8774 #ifndef PRODUCT 8775 // This is actually very useful in a product build if it can 8776 // be called from the debugger. Compile it into the product 8777 // as needed. 8778 bool debug_verify_chunk_in_free_list(FreeChunk* fc) { 8779 return debug_cms_space->verify_chunk_in_free_list(fc); 8780 } 8781 #endif 8782 8783 void SweepClosure::print_free_block_coalesced(FreeChunk* fc) const { 8784 if (CMSTraceSweeper) { 8785 gclog_or_tty->print_cr("Sweep:coal_free_blk " PTR_FORMAT " (" SIZE_FORMAT ")", 8786 fc, fc->size()); 8787 } 8788 } 8789 8790 // CMSIsAliveClosure 8791 bool CMSIsAliveClosure::do_object_b(oop obj) { 8792 HeapWord* addr = (HeapWord*)obj; 8793 return addr != NULL && 8794 (!_span.contains(addr) || _bit_map->isMarked(addr)); 8795 } 8796 8797 8798 CMSKeepAliveClosure::CMSKeepAliveClosure( CMSCollector* collector, 8799 MemRegion span, 8800 CMSBitMap* bit_map, CMSMarkStack* mark_stack, 8801 bool cpc): 8802 _collector(collector), 8803 _span(span), 8804 _bit_map(bit_map), 8805 _mark_stack(mark_stack), 8806 _concurrent_precleaning(cpc) { 8807 assert(!_span.is_empty(), "Empty span could spell trouble"); 8808 } 8809 8810 8811 // CMSKeepAliveClosure: the serial version 8812 void CMSKeepAliveClosure::do_oop(oop obj) { 8813 HeapWord* addr = (HeapWord*)obj; 8814 if (_span.contains(addr) && 8815 !_bit_map->isMarked(addr)) { 8816 _bit_map->mark(addr); 8817 bool simulate_overflow = false; 8818 NOT_PRODUCT( 8819 if (CMSMarkStackOverflowALot && 8820 _collector->simulate_overflow()) { 8821 // simulate a stack overflow 8822 simulate_overflow = true; 8823 } 8824 ) 8825 if (simulate_overflow || !_mark_stack->push(obj)) { 8826 if (_concurrent_precleaning) { 8827 // We dirty the overflown object and let the remark 8828 // phase deal with it. 8829 assert(_collector->overflow_list_is_empty(), "Error"); 8830 // In the case of object arrays, we need to dirty all of 8831 // the cards that the object spans. No locking or atomics 8832 // are needed since no one else can be mutating the mod union 8833 // table. 8834 if (obj->is_objArray()) { 8835 size_t sz = obj->size(); 8836 HeapWord* end_card_addr = 8837 (HeapWord*)round_to((intptr_t)(addr+sz), CardTableModRefBS::card_size); 8838 MemRegion redirty_range = MemRegion(addr, end_card_addr); 8839 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 8840 _collector->_modUnionTable.mark_range(redirty_range); 8841 } else { 8842 _collector->_modUnionTable.mark(addr); 8843 } 8844 _collector->_ser_kac_preclean_ovflw++; 8845 } else { 8846 _collector->push_on_overflow_list(obj); 8847 _collector->_ser_kac_ovflw++; 8848 } 8849 } 8850 } 8851 } 8852 8853 void CMSKeepAliveClosure::do_oop(oop* p) { CMSKeepAliveClosure::do_oop_work(p); } 8854 void CMSKeepAliveClosure::do_oop(narrowOop* p) { CMSKeepAliveClosure::do_oop_work(p); } 8855 8856 // CMSParKeepAliveClosure: a parallel version of the above. 8857 // The work queues are private to each closure (thread), 8858 // but (may be) available for stealing by other threads. 8859 void CMSParKeepAliveClosure::do_oop(oop obj) { 8860 HeapWord* addr = (HeapWord*)obj; 8861 if (_span.contains(addr) && 8862 !_bit_map->isMarked(addr)) { 8863 // In general, during recursive tracing, several threads 8864 // may be concurrently getting here; the first one to 8865 // "tag" it, claims it. 8866 if (_bit_map->par_mark(addr)) { 8867 bool res = _work_queue->push(obj); 8868 assert(res, "Low water mark should be much less than capacity"); 8869 // Do a recursive trim in the hope that this will keep 8870 // stack usage lower, but leave some oops for potential stealers 8871 trim_queue(_low_water_mark); 8872 } // Else, another thread got there first 8873 } 8874 } 8875 8876 void CMSParKeepAliveClosure::do_oop(oop* p) { CMSParKeepAliveClosure::do_oop_work(p); } 8877 void CMSParKeepAliveClosure::do_oop(narrowOop* p) { CMSParKeepAliveClosure::do_oop_work(p); } 8878 8879 void CMSParKeepAliveClosure::trim_queue(uint max) { 8880 while (_work_queue->size() > max) { 8881 oop new_oop; 8882 if (_work_queue->pop_local(new_oop)) { 8883 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 8884 assert(_bit_map->isMarked((HeapWord*)new_oop), 8885 "no white objects on this stack!"); 8886 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 8887 // iterate over the oops in this oop, marking and pushing 8888 // the ones in CMS heap (i.e. in _span). 8889 new_oop->oop_iterate(&_mark_and_push); 8890 } 8891 } 8892 } 8893 8894 CMSInnerParMarkAndPushClosure::CMSInnerParMarkAndPushClosure( 8895 CMSCollector* collector, 8896 MemRegion span, CMSBitMap* bit_map, 8897 OopTaskQueue* work_queue): 8898 _collector(collector), 8899 _span(span), 8900 _bit_map(bit_map), 8901 _work_queue(work_queue) { } 8902 8903 void CMSInnerParMarkAndPushClosure::do_oop(oop obj) { 8904 HeapWord* addr = (HeapWord*)obj; 8905 if (_span.contains(addr) && 8906 !_bit_map->isMarked(addr)) { 8907 if (_bit_map->par_mark(addr)) { 8908 bool simulate_overflow = false; 8909 NOT_PRODUCT( 8910 if (CMSMarkStackOverflowALot && 8911 _collector->par_simulate_overflow()) { 8912 // simulate a stack overflow 8913 simulate_overflow = true; 8914 } 8915 ) 8916 if (simulate_overflow || !_work_queue->push(obj)) { 8917 _collector->par_push_on_overflow_list(obj); 8918 _collector->_par_kac_ovflw++; 8919 } 8920 } // Else another thread got there already 8921 } 8922 } 8923 8924 void CMSInnerParMarkAndPushClosure::do_oop(oop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); } 8925 void CMSInnerParMarkAndPushClosure::do_oop(narrowOop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); } 8926 8927 ////////////////////////////////////////////////////////////////// 8928 // CMSExpansionCause ///////////////////////////// 8929 ////////////////////////////////////////////////////////////////// 8930 const char* CMSExpansionCause::to_string(CMSExpansionCause::Cause cause) { 8931 switch (cause) { 8932 case _no_expansion: 8933 return "No expansion"; 8934 case _satisfy_free_ratio: 8935 return "Free ratio"; 8936 case _satisfy_promotion: 8937 return "Satisfy promotion"; 8938 case _satisfy_allocation: 8939 return "allocation"; 8940 case _allocate_par_lab: 8941 return "Par LAB"; 8942 case _allocate_par_spooling_space: 8943 return "Par Spooling Space"; 8944 case _adaptive_size_policy: 8945 return "Ergonomics"; 8946 default: 8947 return "unknown"; 8948 } 8949 } 8950 8951 void CMSDrainMarkingStackClosure::do_void() { 8952 // the max number to take from overflow list at a time 8953 const size_t num = _mark_stack->capacity()/4; 8954 assert(!_concurrent_precleaning || _collector->overflow_list_is_empty(), 8955 "Overflow list should be NULL during concurrent phases"); 8956 while (!_mark_stack->isEmpty() || 8957 // if stack is empty, check the overflow list 8958 _collector->take_from_overflow_list(num, _mark_stack)) { 8959 oop obj = _mark_stack->pop(); 8960 HeapWord* addr = (HeapWord*)obj; 8961 assert(_span.contains(addr), "Should be within span"); 8962 assert(_bit_map->isMarked(addr), "Should be marked"); 8963 assert(obj->is_oop(), "Should be an oop"); 8964 obj->oop_iterate(_keep_alive); 8965 } 8966 } 8967 8968 void CMSParDrainMarkingStackClosure::do_void() { 8969 // drain queue 8970 trim_queue(0); 8971 } 8972 8973 // Trim our work_queue so its length is below max at return 8974 void CMSParDrainMarkingStackClosure::trim_queue(uint max) { 8975 while (_work_queue->size() > max) { 8976 oop new_oop; 8977 if (_work_queue->pop_local(new_oop)) { 8978 assert(new_oop->is_oop(), "Expected an oop"); 8979 assert(_bit_map->isMarked((HeapWord*)new_oop), 8980 "no white objects on this stack!"); 8981 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 8982 // iterate over the oops in this oop, marking and pushing 8983 // the ones in CMS heap (i.e. in _span). 8984 new_oop->oop_iterate(&_mark_and_push); 8985 } 8986 } 8987 } 8988 8989 //////////////////////////////////////////////////////////////////// 8990 // Support for Marking Stack Overflow list handling and related code 8991 //////////////////////////////////////////////////////////////////// 8992 // Much of the following code is similar in shape and spirit to the 8993 // code used in ParNewGC. We should try and share that code 8994 // as much as possible in the future. 8995 8996 #ifndef PRODUCT 8997 // Debugging support for CMSStackOverflowALot 8998 8999 // It's OK to call this multi-threaded; the worst thing 9000 // that can happen is that we'll get a bunch of closely 9001 // spaced simulated overflows, but that's OK, in fact 9002 // probably good as it would exercise the overflow code 9003 // under contention. 9004 bool CMSCollector::simulate_overflow() { 9005 if (_overflow_counter-- <= 0) { // just being defensive 9006 _overflow_counter = CMSMarkStackOverflowInterval; 9007 return true; 9008 } else { 9009 return false; 9010 } 9011 } 9012 9013 bool CMSCollector::par_simulate_overflow() { 9014 return simulate_overflow(); 9015 } 9016 #endif 9017 9018 // Single-threaded 9019 bool CMSCollector::take_from_overflow_list(size_t num, CMSMarkStack* stack) { 9020 assert(stack->isEmpty(), "Expected precondition"); 9021 assert(stack->capacity() > num, "Shouldn't bite more than can chew"); 9022 size_t i = num; 9023 oop cur = _overflow_list; 9024 const markOop proto = markOopDesc::prototype(); 9025 NOT_PRODUCT(ssize_t n = 0;) 9026 for (oop next; i > 0 && cur != NULL; cur = next, i--) { 9027 next = oop(cur->mark()); 9028 cur->set_mark(proto); // until proven otherwise 9029 assert(cur->is_oop(), "Should be an oop"); 9030 bool res = stack->push(cur); 9031 assert(res, "Bit off more than can chew?"); 9032 NOT_PRODUCT(n++;) 9033 } 9034 _overflow_list = cur; 9035 #ifndef PRODUCT 9036 assert(_num_par_pushes >= n, "Too many pops?"); 9037 _num_par_pushes -=n; 9038 #endif 9039 return !stack->isEmpty(); 9040 } 9041 9042 #define BUSY (cast_to_oop<intptr_t>(0x1aff1aff)) 9043 // (MT-safe) Get a prefix of at most "num" from the list. 9044 // The overflow list is chained through the mark word of 9045 // each object in the list. We fetch the entire list, 9046 // break off a prefix of the right size and return the 9047 // remainder. If other threads try to take objects from 9048 // the overflow list at that time, they will wait for 9049 // some time to see if data becomes available. If (and 9050 // only if) another thread places one or more object(s) 9051 // on the global list before we have returned the suffix 9052 // to the global list, we will walk down our local list 9053 // to find its end and append the global list to 9054 // our suffix before returning it. This suffix walk can 9055 // prove to be expensive (quadratic in the amount of traffic) 9056 // when there are many objects in the overflow list and 9057 // there is much producer-consumer contention on the list. 9058 // *NOTE*: The overflow list manipulation code here and 9059 // in ParNewGeneration:: are very similar in shape, 9060 // except that in the ParNew case we use the old (from/eden) 9061 // copy of the object to thread the list via its klass word. 9062 // Because of the common code, if you make any changes in 9063 // the code below, please check the ParNew version to see if 9064 // similar changes might be needed. 9065 // CR 6797058 has been filed to consolidate the common code. 9066 bool CMSCollector::par_take_from_overflow_list(size_t num, 9067 OopTaskQueue* work_q, 9068 int no_of_gc_threads) { 9069 assert(work_q->size() == 0, "First empty local work queue"); 9070 assert(num < work_q->max_elems(), "Can't bite more than we can chew"); 9071 if (_overflow_list == NULL) { 9072 return false; 9073 } 9074 // Grab the entire list; we'll put back a suffix 9075 oop prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list)); 9076 Thread* tid = Thread::current(); 9077 // Before "no_of_gc_threads" was introduced CMSOverflowSpinCount was 9078 // set to ParallelGCThreads. 9079 size_t CMSOverflowSpinCount = (size_t) no_of_gc_threads; // was ParallelGCThreads; 9080 size_t sleep_time_millis = MAX2((size_t)1, num/100); 9081 // If the list is busy, we spin for a short while, 9082 // sleeping between attempts to get the list. 9083 for (size_t spin = 0; prefix == BUSY && spin < CMSOverflowSpinCount; spin++) { 9084 os::sleep(tid, sleep_time_millis, false); 9085 if (_overflow_list == NULL) { 9086 // Nothing left to take 9087 return false; 9088 } else if (_overflow_list != BUSY) { 9089 // Try and grab the prefix 9090 prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list)); 9091 } 9092 } 9093 // If the list was found to be empty, or we spun long 9094 // enough, we give up and return empty-handed. If we leave 9095 // the list in the BUSY state below, it must be the case that 9096 // some other thread holds the overflow list and will set it 9097 // to a non-BUSY state in the future. 9098 if (prefix == NULL || prefix == BUSY) { 9099 // Nothing to take or waited long enough 9100 if (prefix == NULL) { 9101 // Write back the NULL in case we overwrote it with BUSY above 9102 // and it is still the same value. 9103 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY); 9104 } 9105 return false; 9106 } 9107 assert(prefix != NULL && prefix != BUSY, "Error"); 9108 size_t i = num; 9109 oop cur = prefix; 9110 // Walk down the first "num" objects, unless we reach the end. 9111 for (; i > 1 && cur->mark() != NULL; cur = oop(cur->mark()), i--); 9112 if (cur->mark() == NULL) { 9113 // We have "num" or fewer elements in the list, so there 9114 // is nothing to return to the global list. 9115 // Write back the NULL in lieu of the BUSY we wrote 9116 // above, if it is still the same value. 9117 if (_overflow_list == BUSY) { 9118 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY); 9119 } 9120 } else { 9121 // Chop off the suffix and return it to the global list. 9122 assert(cur->mark() != BUSY, "Error"); 9123 oop suffix_head = cur->mark(); // suffix will be put back on global list 9124 cur->set_mark(NULL); // break off suffix 9125 // It's possible that the list is still in the empty(busy) state 9126 // we left it in a short while ago; in that case we may be 9127 // able to place back the suffix without incurring the cost 9128 // of a walk down the list. 9129 oop observed_overflow_list = _overflow_list; 9130 oop cur_overflow_list = observed_overflow_list; 9131 bool attached = false; 9132 while (observed_overflow_list == BUSY || observed_overflow_list == NULL) { 9133 observed_overflow_list = 9134 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list); 9135 if (cur_overflow_list == observed_overflow_list) { 9136 attached = true; 9137 break; 9138 } else cur_overflow_list = observed_overflow_list; 9139 } 9140 if (!attached) { 9141 // Too bad, someone else sneaked in (at least) an element; we'll need 9142 // to do a splice. Find tail of suffix so we can prepend suffix to global 9143 // list. 9144 for (cur = suffix_head; cur->mark() != NULL; cur = (oop)(cur->mark())); 9145 oop suffix_tail = cur; 9146 assert(suffix_tail != NULL && suffix_tail->mark() == NULL, 9147 "Tautology"); 9148 observed_overflow_list = _overflow_list; 9149 do { 9150 cur_overflow_list = observed_overflow_list; 9151 if (cur_overflow_list != BUSY) { 9152 // Do the splice ... 9153 suffix_tail->set_mark(markOop(cur_overflow_list)); 9154 } else { // cur_overflow_list == BUSY 9155 suffix_tail->set_mark(NULL); 9156 } 9157 // ... and try to place spliced list back on overflow_list ... 9158 observed_overflow_list = 9159 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list); 9160 } while (cur_overflow_list != observed_overflow_list); 9161 // ... until we have succeeded in doing so. 9162 } 9163 } 9164 9165 // Push the prefix elements on work_q 9166 assert(prefix != NULL, "control point invariant"); 9167 const markOop proto = markOopDesc::prototype(); 9168 oop next; 9169 NOT_PRODUCT(ssize_t n = 0;) 9170 for (cur = prefix; cur != NULL; cur = next) { 9171 next = oop(cur->mark()); 9172 cur->set_mark(proto); // until proven otherwise 9173 assert(cur->is_oop(), "Should be an oop"); 9174 bool res = work_q->push(cur); 9175 assert(res, "Bit off more than we can chew?"); 9176 NOT_PRODUCT(n++;) 9177 } 9178 #ifndef PRODUCT 9179 assert(_num_par_pushes >= n, "Too many pops?"); 9180 Atomic::add_ptr(-(intptr_t)n, &_num_par_pushes); 9181 #endif 9182 return true; 9183 } 9184 9185 // Single-threaded 9186 void CMSCollector::push_on_overflow_list(oop p) { 9187 NOT_PRODUCT(_num_par_pushes++;) 9188 assert(p->is_oop(), "Not an oop"); 9189 preserve_mark_if_necessary(p); 9190 p->set_mark((markOop)_overflow_list); 9191 _overflow_list = p; 9192 } 9193 9194 // Multi-threaded; use CAS to prepend to overflow list 9195 void CMSCollector::par_push_on_overflow_list(oop p) { 9196 NOT_PRODUCT(Atomic::inc_ptr(&_num_par_pushes);) 9197 assert(p->is_oop(), "Not an oop"); 9198 par_preserve_mark_if_necessary(p); 9199 oop observed_overflow_list = _overflow_list; 9200 oop cur_overflow_list; 9201 do { 9202 cur_overflow_list = observed_overflow_list; 9203 if (cur_overflow_list != BUSY) { 9204 p->set_mark(markOop(cur_overflow_list)); 9205 } else { 9206 p->set_mark(NULL); 9207 } 9208 observed_overflow_list = 9209 (oop) Atomic::cmpxchg_ptr(p, &_overflow_list, cur_overflow_list); 9210 } while (cur_overflow_list != observed_overflow_list); 9211 } 9212 #undef BUSY 9213 9214 // Single threaded 9215 // General Note on GrowableArray: pushes may silently fail 9216 // because we are (temporarily) out of C-heap for expanding 9217 // the stack. The problem is quite ubiquitous and affects 9218 // a lot of code in the JVM. The prudent thing for GrowableArray 9219 // to do (for now) is to exit with an error. However, that may 9220 // be too draconian in some cases because the caller may be 9221 // able to recover without much harm. For such cases, we 9222 // should probably introduce a "soft_push" method which returns 9223 // an indication of success or failure with the assumption that 9224 // the caller may be able to recover from a failure; code in 9225 // the VM can then be changed, incrementally, to deal with such 9226 // failures where possible, thus, incrementally hardening the VM 9227 // in such low resource situations. 9228 void CMSCollector::preserve_mark_work(oop p, markOop m) { 9229 _preserved_oop_stack.push(p); 9230 _preserved_mark_stack.push(m); 9231 assert(m == p->mark(), "Mark word changed"); 9232 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), 9233 "bijection"); 9234 } 9235 9236 // Single threaded 9237 void CMSCollector::preserve_mark_if_necessary(oop p) { 9238 markOop m = p->mark(); 9239 if (m->must_be_preserved(p)) { 9240 preserve_mark_work(p, m); 9241 } 9242 } 9243 9244 void CMSCollector::par_preserve_mark_if_necessary(oop p) { 9245 markOop m = p->mark(); 9246 if (m->must_be_preserved(p)) { 9247 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag); 9248 // Even though we read the mark word without holding 9249 // the lock, we are assured that it will not change 9250 // because we "own" this oop, so no other thread can 9251 // be trying to push it on the overflow list; see 9252 // the assertion in preserve_mark_work() that checks 9253 // that m == p->mark(). 9254 preserve_mark_work(p, m); 9255 } 9256 } 9257 9258 // We should be able to do this multi-threaded, 9259 // a chunk of stack being a task (this is 9260 // correct because each oop only ever appears 9261 // once in the overflow list. However, it's 9262 // not very easy to completely overlap this with 9263 // other operations, so will generally not be done 9264 // until all work's been completed. Because we 9265 // expect the preserved oop stack (set) to be small, 9266 // it's probably fine to do this single-threaded. 9267 // We can explore cleverer concurrent/overlapped/parallel 9268 // processing of preserved marks if we feel the 9269 // need for this in the future. Stack overflow should 9270 // be so rare in practice and, when it happens, its 9271 // effect on performance so great that this will 9272 // likely just be in the noise anyway. 9273 void CMSCollector::restore_preserved_marks_if_any() { 9274 assert(SafepointSynchronize::is_at_safepoint(), 9275 "world should be stopped"); 9276 assert(Thread::current()->is_ConcurrentGC_thread() || 9277 Thread::current()->is_VM_thread(), 9278 "should be single-threaded"); 9279 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), 9280 "bijection"); 9281 9282 while (!_preserved_oop_stack.is_empty()) { 9283 oop p = _preserved_oop_stack.pop(); 9284 assert(p->is_oop(), "Should be an oop"); 9285 assert(_span.contains(p), "oop should be in _span"); 9286 assert(p->mark() == markOopDesc::prototype(), 9287 "Set when taken from overflow list"); 9288 markOop m = _preserved_mark_stack.pop(); 9289 p->set_mark(m); 9290 } 9291 assert(_preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(), 9292 "stacks were cleared above"); 9293 } 9294 9295 #ifndef PRODUCT 9296 bool CMSCollector::no_preserved_marks() const { 9297 return _preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(); 9298 } 9299 #endif 9300 9301 CMSAdaptiveSizePolicy* ASConcurrentMarkSweepGeneration::cms_size_policy() const 9302 { 9303 GenCollectedHeap* gch = (GenCollectedHeap*) GenCollectedHeap::heap(); 9304 CMSAdaptiveSizePolicy* size_policy = 9305 (CMSAdaptiveSizePolicy*) gch->gen_policy()->size_policy(); 9306 assert(size_policy->is_gc_cms_adaptive_size_policy(), 9307 "Wrong type for size policy"); 9308 return size_policy; 9309 } 9310 9311 void ASConcurrentMarkSweepGeneration::resize(size_t cur_promo_size, 9312 size_t desired_promo_size) { 9313 if (cur_promo_size < desired_promo_size) { 9314 size_t expand_bytes = desired_promo_size - cur_promo_size; 9315 if (PrintAdaptiveSizePolicy && Verbose) { 9316 gclog_or_tty->print_cr(" ASConcurrentMarkSweepGeneration::resize " 9317 "Expanding tenured generation by " SIZE_FORMAT " (bytes)", 9318 expand_bytes); 9319 } 9320 expand(expand_bytes, 9321 MinHeapDeltaBytes, 9322 CMSExpansionCause::_adaptive_size_policy); 9323 } else if (desired_promo_size < cur_promo_size) { 9324 size_t shrink_bytes = cur_promo_size - desired_promo_size; 9325 if (PrintAdaptiveSizePolicy && Verbose) { 9326 gclog_or_tty->print_cr(" ASConcurrentMarkSweepGeneration::resize " 9327 "Shrinking tenured generation by " SIZE_FORMAT " (bytes)", 9328 shrink_bytes); 9329 } 9330 shrink(shrink_bytes); 9331 } 9332 } 9333 9334 CMSGCAdaptivePolicyCounters* ASConcurrentMarkSweepGeneration::gc_adaptive_policy_counters() { 9335 GenCollectedHeap* gch = GenCollectedHeap::heap(); 9336 CMSGCAdaptivePolicyCounters* counters = 9337 (CMSGCAdaptivePolicyCounters*) gch->collector_policy()->counters(); 9338 assert(counters->kind() == GCPolicyCounters::CMSGCAdaptivePolicyCountersKind, 9339 "Wrong kind of counters"); 9340 return counters; 9341 } 9342 9343 9344 void ASConcurrentMarkSweepGeneration::update_counters() { 9345 if (UsePerfData) { 9346 _space_counters->update_all(); 9347 _gen_counters->update_all(); 9348 CMSGCAdaptivePolicyCounters* counters = gc_adaptive_policy_counters(); 9349 GenCollectedHeap* gch = GenCollectedHeap::heap(); 9350 CMSGCStats* gc_stats_l = (CMSGCStats*) gc_stats(); 9351 assert(gc_stats_l->kind() == GCStats::CMSGCStatsKind, 9352 "Wrong gc statistics type"); 9353 counters->update_counters(gc_stats_l); 9354 } 9355 } 9356 9357 void ASConcurrentMarkSweepGeneration::update_counters(size_t used) { 9358 if (UsePerfData) { 9359 _space_counters->update_used(used); 9360 _space_counters->update_capacity(); 9361 _gen_counters->update_all(); 9362 9363 CMSGCAdaptivePolicyCounters* counters = gc_adaptive_policy_counters(); 9364 GenCollectedHeap* gch = GenCollectedHeap::heap(); 9365 CMSGCStats* gc_stats_l = (CMSGCStats*) gc_stats(); 9366 assert(gc_stats_l->kind() == GCStats::CMSGCStatsKind, 9367 "Wrong gc statistics type"); 9368 counters->update_counters(gc_stats_l); 9369 } 9370 } 9371 9372 void ASConcurrentMarkSweepGeneration::shrink_by(size_t desired_bytes) { 9373 assert_locked_or_safepoint(Heap_lock); 9374 assert_lock_strong(freelistLock()); 9375 HeapWord* old_end = _cmsSpace->end(); 9376 HeapWord* unallocated_start = _cmsSpace->unallocated_block(); 9377 assert(old_end >= unallocated_start, "Miscalculation of unallocated_start"); 9378 FreeChunk* chunk_at_end = find_chunk_at_end(); 9379 if (chunk_at_end == NULL) { 9380 // No room to shrink 9381 if (PrintGCDetails && Verbose) { 9382 gclog_or_tty->print_cr("No room to shrink: old_end " 9383 PTR_FORMAT " unallocated_start " PTR_FORMAT 9384 " chunk_at_end " PTR_FORMAT, 9385 old_end, unallocated_start, chunk_at_end); 9386 } 9387 return; 9388 } else { 9389 9390 // Find the chunk at the end of the space and determine 9391 // how much it can be shrunk. 9392 size_t shrinkable_size_in_bytes = chunk_at_end->size(); 9393 size_t aligned_shrinkable_size_in_bytes = 9394 align_size_down(shrinkable_size_in_bytes, os::vm_page_size()); 9395 assert(unallocated_start <= (HeapWord*) chunk_at_end->end(), 9396 "Inconsistent chunk at end of space"); 9397 size_t bytes = MIN2(desired_bytes, aligned_shrinkable_size_in_bytes); 9398 size_t word_size_before = heap_word_size(_virtual_space.committed_size()); 9399 9400 // Shrink the underlying space 9401 _virtual_space.shrink_by(bytes); 9402 if (PrintGCDetails && Verbose) { 9403 gclog_or_tty->print_cr("ConcurrentMarkSweepGeneration::shrink_by:" 9404 " desired_bytes " SIZE_FORMAT 9405 " shrinkable_size_in_bytes " SIZE_FORMAT 9406 " aligned_shrinkable_size_in_bytes " SIZE_FORMAT 9407 " bytes " SIZE_FORMAT, 9408 desired_bytes, shrinkable_size_in_bytes, 9409 aligned_shrinkable_size_in_bytes, bytes); 9410 gclog_or_tty->print_cr(" old_end " SIZE_FORMAT 9411 " unallocated_start " SIZE_FORMAT, 9412 old_end, unallocated_start); 9413 } 9414 9415 // If the space did shrink (shrinking is not guaranteed), 9416 // shrink the chunk at the end by the appropriate amount. 9417 if (((HeapWord*)_virtual_space.high()) < old_end) { 9418 size_t new_word_size = 9419 heap_word_size(_virtual_space.committed_size()); 9420 9421 // Have to remove the chunk from the dictionary because it is changing 9422 // size and might be someplace elsewhere in the dictionary. 9423 9424 // Get the chunk at end, shrink it, and put it 9425 // back. 9426 _cmsSpace->removeChunkFromDictionary(chunk_at_end); 9427 size_t word_size_change = word_size_before - new_word_size; 9428 size_t chunk_at_end_old_size = chunk_at_end->size(); 9429 assert(chunk_at_end_old_size >= word_size_change, 9430 "Shrink is too large"); 9431 chunk_at_end->set_size(chunk_at_end_old_size - 9432 word_size_change); 9433 _cmsSpace->freed((HeapWord*) chunk_at_end->end(), 9434 word_size_change); 9435 9436 _cmsSpace->returnChunkToDictionary(chunk_at_end); 9437 9438 MemRegion mr(_cmsSpace->bottom(), new_word_size); 9439 _bts->resize(new_word_size); // resize the block offset shared array 9440 Universe::heap()->barrier_set()->resize_covered_region(mr); 9441 _cmsSpace->assert_locked(); 9442 _cmsSpace->set_end((HeapWord*)_virtual_space.high()); 9443 9444 NOT_PRODUCT(_cmsSpace->dictionary()->verify()); 9445 9446 // update the space and generation capacity counters 9447 if (UsePerfData) { 9448 _space_counters->update_capacity(); 9449 _gen_counters->update_all(); 9450 } 9451 9452 if (Verbose && PrintGCDetails) { 9453 size_t new_mem_size = _virtual_space.committed_size(); 9454 size_t old_mem_size = new_mem_size + bytes; 9455 gclog_or_tty->print_cr("Shrinking %s from " SIZE_FORMAT "K by " SIZE_FORMAT "K to " SIZE_FORMAT "K", 9456 name(), old_mem_size/K, bytes/K, new_mem_size/K); 9457 } 9458 } 9459 9460 assert(_cmsSpace->unallocated_block() <= _cmsSpace->end(), 9461 "Inconsistency at end of space"); 9462 assert(chunk_at_end->end() == (uintptr_t*) _cmsSpace->end(), 9463 "Shrinking is inconsistent"); 9464 return; 9465 } 9466 } 9467 // Transfer some number of overflown objects to usual marking 9468 // stack. Return true if some objects were transferred. 9469 bool MarkRefsIntoAndScanClosure::take_from_overflow_list() { 9470 size_t num = MIN2((size_t)(_mark_stack->capacity() - _mark_stack->length())/4, 9471 (size_t)ParGCDesiredObjsFromOverflowList); 9472 9473 bool res = _collector->take_from_overflow_list(num, _mark_stack); 9474 assert(_collector->overflow_list_is_empty() || res, 9475 "If list is not empty, we should have taken something"); 9476 assert(!res || !_mark_stack->isEmpty(), 9477 "If we took something, it should now be on our stack"); 9478 return res; 9479 } 9480 9481 size_t MarkDeadObjectsClosure::do_blk(HeapWord* addr) { 9482 size_t res = _sp->block_size_no_stall(addr, _collector); 9483 if (_sp->block_is_obj(addr)) { 9484 if (_live_bit_map->isMarked(addr)) { 9485 // It can't have been dead in a previous cycle 9486 guarantee(!_dead_bit_map->isMarked(addr), "No resurrection!"); 9487 } else { 9488 _dead_bit_map->mark(addr); // mark the dead object 9489 } 9490 } 9491 // Could be 0, if the block size could not be computed without stalling. 9492 return res; 9493 } 9494 9495 TraceCMSMemoryManagerStats::TraceCMSMemoryManagerStats(CMSCollector::CollectorState phase, GCCause::Cause cause): TraceMemoryManagerStats() { 9496 9497 switch (phase) { 9498 case CMSCollector::InitialMarking: 9499 initialize(true /* fullGC */ , 9500 cause /* cause of the GC */, 9501 true /* recordGCBeginTime */, 9502 true /* recordPreGCUsage */, 9503 false /* recordPeakUsage */, 9504 false /* recordPostGCusage */, 9505 true /* recordAccumulatedGCTime */, 9506 false /* recordGCEndTime */, 9507 false /* countCollection */ ); 9508 break; 9509 9510 case CMSCollector::FinalMarking: 9511 initialize(true /* fullGC */ , 9512 cause /* cause of the GC */, 9513 false /* recordGCBeginTime */, 9514 false /* recordPreGCUsage */, 9515 false /* recordPeakUsage */, 9516 false /* recordPostGCusage */, 9517 true /* recordAccumulatedGCTime */, 9518 false /* recordGCEndTime */, 9519 false /* countCollection */ ); 9520 break; 9521 9522 case CMSCollector::Sweeping: 9523 initialize(true /* fullGC */ , 9524 cause /* cause of the GC */, 9525 false /* recordGCBeginTime */, 9526 false /* recordPreGCUsage */, 9527 true /* recordPeakUsage */, 9528 true /* recordPostGCusage */, 9529 false /* recordAccumulatedGCTime */, 9530 true /* recordGCEndTime */, 9531 true /* countCollection */ ); 9532 break; 9533 9534 default: 9535 ShouldNotReachHere(); 9536 } 9537 }