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(" Capactiy "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, _last_metaspace_summary); 2500 } 2501 2502 void CMSCollector::collect_in_foreground(bool clear_all_soft_refs, GCCause::Cause cause) { 2503 assert(_foregroundGCIsActive && !_foregroundGCShouldWait, 2504 "Foreground collector should be waiting, not executing"); 2505 assert(Thread::current()->is_VM_thread(), "A foreground collection" 2506 "may only be done by the VM Thread with the world stopped"); 2507 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 2508 "VM thread should have CMS token"); 2509 2510 NOT_PRODUCT(GCTraceTime t("CMS:MS (foreground) ", PrintGCDetails && Verbose, 2511 true, NULL);) 2512 if (UseAdaptiveSizePolicy) { 2513 size_policy()->ms_collection_begin(); 2514 } 2515 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact); 2516 2517 HandleMark hm; // Discard invalid handles created during verification 2518 2519 if (VerifyBeforeGC && 2520 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2521 Universe::verify(); 2522 } 2523 2524 // Snapshot the soft reference policy to be used in this collection cycle. 2525 ref_processor()->setup_policy(clear_all_soft_refs); 2526 2527 // Decide if class unloading should be done 2528 update_should_unload_classes(); 2529 2530 bool init_mark_was_synchronous = false; // until proven otherwise 2531 while (_collectorState != Idling) { 2532 if (TraceCMSState) { 2533 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " in CMS state %d", 2534 Thread::current(), _collectorState); 2535 } 2536 switch (_collectorState) { 2537 case InitialMarking: 2538 register_foreground_gc_start(cause); 2539 init_mark_was_synchronous = true; // fact to be exploited in re-mark 2540 checkpointRootsInitial(false); 2541 assert(_collectorState == Marking, "Collector state should have changed" 2542 " within checkpointRootsInitial()"); 2543 break; 2544 case Marking: 2545 // initial marking in checkpointRootsInitialWork has been completed 2546 if (VerifyDuringGC && 2547 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2548 Universe::verify("Verify before initial mark: "); 2549 } 2550 { 2551 bool res = markFromRoots(false); 2552 assert(res && _collectorState == FinalMarking, "Collector state should " 2553 "have changed"); 2554 break; 2555 } 2556 case FinalMarking: 2557 if (VerifyDuringGC && 2558 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2559 Universe::verify("Verify before re-mark: "); 2560 } 2561 checkpointRootsFinal(false, clear_all_soft_refs, 2562 init_mark_was_synchronous); 2563 assert(_collectorState == Sweeping, "Collector state should not " 2564 "have changed within checkpointRootsFinal()"); 2565 break; 2566 case Sweeping: 2567 // final marking in checkpointRootsFinal has been completed 2568 if (VerifyDuringGC && 2569 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2570 Universe::verify("Verify before sweep: "); 2571 } 2572 sweep(false); 2573 assert(_collectorState == Resizing, "Incorrect state"); 2574 break; 2575 case Resizing: { 2576 // Sweeping has been completed; the actual resize in this case 2577 // is done separately; nothing to be done in this state. 2578 _collectorState = Resetting; 2579 break; 2580 } 2581 case Resetting: 2582 // The heap has been resized. 2583 if (VerifyDuringGC && 2584 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2585 Universe::verify("Verify before reset: "); 2586 } 2587 save_heap_summary(); 2588 reset(false); 2589 assert(_collectorState == Idling, "Collector state should " 2590 "have changed"); 2591 break; 2592 case Precleaning: 2593 case AbortablePreclean: 2594 // Elide the preclean phase 2595 _collectorState = FinalMarking; 2596 break; 2597 default: 2598 ShouldNotReachHere(); 2599 } 2600 if (TraceCMSState) { 2601 gclog_or_tty->print_cr(" Thread " INTPTR_FORMAT " done - next CMS state %d", 2602 Thread::current(), _collectorState); 2603 } 2604 } 2605 2606 if (UseAdaptiveSizePolicy) { 2607 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2608 size_policy()->ms_collection_end(gch->gc_cause()); 2609 } 2610 2611 if (VerifyAfterGC && 2612 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2613 Universe::verify(); 2614 } 2615 if (TraceCMSState) { 2616 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT 2617 " exiting collection CMS state %d", 2618 Thread::current(), _collectorState); 2619 } 2620 } 2621 2622 bool CMSCollector::waitForForegroundGC() { 2623 bool res = false; 2624 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 2625 "CMS thread should have CMS token"); 2626 // Block the foreground collector until the 2627 // background collectors decides whether to 2628 // yield. 2629 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 2630 _foregroundGCShouldWait = true; 2631 if (_foregroundGCIsActive) { 2632 // The background collector yields to the 2633 // foreground collector and returns a value 2634 // indicating that it has yielded. The foreground 2635 // collector can proceed. 2636 res = true; 2637 _foregroundGCShouldWait = false; 2638 ConcurrentMarkSweepThread::clear_CMS_flag( 2639 ConcurrentMarkSweepThread::CMS_cms_has_token); 2640 ConcurrentMarkSweepThread::set_CMS_flag( 2641 ConcurrentMarkSweepThread::CMS_cms_wants_token); 2642 // Get a possibly blocked foreground thread going 2643 CGC_lock->notify(); 2644 if (TraceCMSState) { 2645 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " waiting at CMS state %d", 2646 Thread::current(), _collectorState); 2647 } 2648 while (_foregroundGCIsActive) { 2649 CGC_lock->wait(Mutex::_no_safepoint_check_flag); 2650 } 2651 ConcurrentMarkSweepThread::set_CMS_flag( 2652 ConcurrentMarkSweepThread::CMS_cms_has_token); 2653 ConcurrentMarkSweepThread::clear_CMS_flag( 2654 ConcurrentMarkSweepThread::CMS_cms_wants_token); 2655 } 2656 if (TraceCMSState) { 2657 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " continuing at CMS state %d", 2658 Thread::current(), _collectorState); 2659 } 2660 return res; 2661 } 2662 2663 // Because of the need to lock the free lists and other structures in 2664 // the collector, common to all the generations that the collector is 2665 // collecting, we need the gc_prologues of individual CMS generations 2666 // delegate to their collector. It may have been simpler had the 2667 // current infrastructure allowed one to call a prologue on a 2668 // collector. In the absence of that we have the generation's 2669 // prologue delegate to the collector, which delegates back 2670 // some "local" work to a worker method in the individual generations 2671 // that it's responsible for collecting, while itself doing any 2672 // work common to all generations it's responsible for. A similar 2673 // comment applies to the gc_epilogue()'s. 2674 // The role of the variable _between_prologue_and_epilogue is to 2675 // enforce the invocation protocol. 2676 void CMSCollector::gc_prologue(bool full) { 2677 // Call gc_prologue_work() for the CMSGen 2678 // we are responsible for. 2679 2680 // The following locking discipline assumes that we are only called 2681 // when the world is stopped. 2682 assert(SafepointSynchronize::is_at_safepoint(), "world is stopped assumption"); 2683 2684 // The CMSCollector prologue must call the gc_prologues for the 2685 // "generations" that it's responsible 2686 // for. 2687 2688 assert( Thread::current()->is_VM_thread() 2689 || ( CMSScavengeBeforeRemark 2690 && Thread::current()->is_ConcurrentGC_thread()), 2691 "Incorrect thread type for prologue execution"); 2692 2693 if (_between_prologue_and_epilogue) { 2694 // We have already been invoked; this is a gc_prologue delegation 2695 // from yet another CMS generation that we are responsible for, just 2696 // ignore it since all relevant work has already been done. 2697 return; 2698 } 2699 2700 // set a bit saying prologue has been called; cleared in epilogue 2701 _between_prologue_and_epilogue = true; 2702 // Claim locks for common data structures, then call gc_prologue_work() 2703 // for each CMSGen. 2704 2705 getFreelistLocks(); // gets free list locks on constituent spaces 2706 bitMapLock()->lock_without_safepoint_check(); 2707 2708 // Should call gc_prologue_work() for all cms gens we are responsible for 2709 bool duringMarking = _collectorState >= Marking 2710 && _collectorState < Sweeping; 2711 2712 // The young collections clear the modified oops state, which tells if 2713 // there are any modified oops in the class. The remark phase also needs 2714 // that information. Tell the young collection to save the union of all 2715 // modified klasses. 2716 if (duringMarking) { 2717 _ct->klass_rem_set()->set_accumulate_modified_oops(true); 2718 } 2719 2720 bool registerClosure = duringMarking; 2721 2722 ModUnionClosure* muc = CollectedHeap::use_parallel_gc_threads() ? 2723 &_modUnionClosurePar 2724 : &_modUnionClosure; 2725 _cmsGen->gc_prologue_work(full, registerClosure, muc); 2726 2727 if (!full) { 2728 stats().record_gc0_begin(); 2729 } 2730 } 2731 2732 void ConcurrentMarkSweepGeneration::gc_prologue(bool full) { 2733 2734 _capacity_at_prologue = capacity(); 2735 _used_at_prologue = used(); 2736 2737 // Delegate to CMScollector which knows how to coordinate between 2738 // this and any other CMS generations that it is responsible for 2739 // collecting. 2740 collector()->gc_prologue(full); 2741 } 2742 2743 // This is a "private" interface for use by this generation's CMSCollector. 2744 // Not to be called directly by any other entity (for instance, 2745 // GenCollectedHeap, which calls the "public" gc_prologue method above). 2746 void ConcurrentMarkSweepGeneration::gc_prologue_work(bool full, 2747 bool registerClosure, ModUnionClosure* modUnionClosure) { 2748 assert(!incremental_collection_failed(), "Shouldn't be set yet"); 2749 assert(cmsSpace()->preconsumptionDirtyCardClosure() == NULL, 2750 "Should be NULL"); 2751 if (registerClosure) { 2752 cmsSpace()->setPreconsumptionDirtyCardClosure(modUnionClosure); 2753 } 2754 cmsSpace()->gc_prologue(); 2755 // Clear stat counters 2756 NOT_PRODUCT( 2757 assert(_numObjectsPromoted == 0, "check"); 2758 assert(_numWordsPromoted == 0, "check"); 2759 if (Verbose && PrintGC) { 2760 gclog_or_tty->print("Allocated "SIZE_FORMAT" objects, " 2761 SIZE_FORMAT" bytes concurrently", 2762 _numObjectsAllocated, _numWordsAllocated*sizeof(HeapWord)); 2763 } 2764 _numObjectsAllocated = 0; 2765 _numWordsAllocated = 0; 2766 ) 2767 } 2768 2769 void CMSCollector::gc_epilogue(bool full) { 2770 // The following locking discipline assumes that we are only called 2771 // when the world is stopped. 2772 assert(SafepointSynchronize::is_at_safepoint(), 2773 "world is stopped assumption"); 2774 2775 // Currently the CMS epilogue (see CompactibleFreeListSpace) merely checks 2776 // if linear allocation blocks need to be appropriately marked to allow the 2777 // the blocks to be parsable. We also check here whether we need to nudge the 2778 // CMS collector thread to start a new cycle (if it's not already active). 2779 assert( Thread::current()->is_VM_thread() 2780 || ( CMSScavengeBeforeRemark 2781 && Thread::current()->is_ConcurrentGC_thread()), 2782 "Incorrect thread type for epilogue execution"); 2783 2784 if (!_between_prologue_and_epilogue) { 2785 // We have already been invoked; this is a gc_epilogue delegation 2786 // from yet another CMS generation that we are responsible for, just 2787 // ignore it since all relevant work has already been done. 2788 return; 2789 } 2790 assert(haveFreelistLocks(), "must have freelist locks"); 2791 assert_lock_strong(bitMapLock()); 2792 2793 _ct->klass_rem_set()->set_accumulate_modified_oops(false); 2794 2795 _cmsGen->gc_epilogue_work(full); 2796 2797 if (_collectorState == AbortablePreclean || _collectorState == Precleaning) { 2798 // in case sampling was not already enabled, enable it 2799 _start_sampling = true; 2800 } 2801 // reset _eden_chunk_array so sampling starts afresh 2802 _eden_chunk_index = 0; 2803 2804 size_t cms_used = _cmsGen->cmsSpace()->used(); 2805 2806 // update performance counters - this uses a special version of 2807 // update_counters() that allows the utilization to be passed as a 2808 // parameter, avoiding multiple calls to used(). 2809 // 2810 _cmsGen->update_counters(cms_used); 2811 2812 if (CMSIncrementalMode) { 2813 icms_update_allocation_limits(); 2814 } 2815 2816 bitMapLock()->unlock(); 2817 releaseFreelistLocks(); 2818 2819 if (!CleanChunkPoolAsync) { 2820 Chunk::clean_chunk_pool(); 2821 } 2822 2823 set_did_compact(false); 2824 _between_prologue_and_epilogue = false; // ready for next cycle 2825 } 2826 2827 void ConcurrentMarkSweepGeneration::gc_epilogue(bool full) { 2828 collector()->gc_epilogue(full); 2829 2830 // Also reset promotion tracking in par gc thread states. 2831 if (CollectedHeap::use_parallel_gc_threads()) { 2832 for (uint i = 0; i < ParallelGCThreads; i++) { 2833 _par_gc_thread_states[i]->promo.stopTrackingPromotions(i); 2834 } 2835 } 2836 } 2837 2838 void ConcurrentMarkSweepGeneration::gc_epilogue_work(bool full) { 2839 assert(!incremental_collection_failed(), "Should have been cleared"); 2840 cmsSpace()->setPreconsumptionDirtyCardClosure(NULL); 2841 cmsSpace()->gc_epilogue(); 2842 // Print stat counters 2843 NOT_PRODUCT( 2844 assert(_numObjectsAllocated == 0, "check"); 2845 assert(_numWordsAllocated == 0, "check"); 2846 if (Verbose && PrintGC) { 2847 gclog_or_tty->print("Promoted "SIZE_FORMAT" objects, " 2848 SIZE_FORMAT" bytes", 2849 _numObjectsPromoted, _numWordsPromoted*sizeof(HeapWord)); 2850 } 2851 _numObjectsPromoted = 0; 2852 _numWordsPromoted = 0; 2853 ) 2854 2855 if (PrintGC && Verbose) { 2856 // Call down the chain in contiguous_available needs the freelistLock 2857 // so print this out before releasing the freeListLock. 2858 gclog_or_tty->print(" Contiguous available "SIZE_FORMAT" bytes ", 2859 contiguous_available()); 2860 } 2861 } 2862 2863 #ifndef PRODUCT 2864 bool CMSCollector::have_cms_token() { 2865 Thread* thr = Thread::current(); 2866 if (thr->is_VM_thread()) { 2867 return ConcurrentMarkSweepThread::vm_thread_has_cms_token(); 2868 } else if (thr->is_ConcurrentGC_thread()) { 2869 return ConcurrentMarkSweepThread::cms_thread_has_cms_token(); 2870 } else if (thr->is_GC_task_thread()) { 2871 return ConcurrentMarkSweepThread::vm_thread_has_cms_token() && 2872 ParGCRareEvent_lock->owned_by_self(); 2873 } 2874 return false; 2875 } 2876 #endif 2877 2878 // Check reachability of the given heap address in CMS generation, 2879 // treating all other generations as roots. 2880 bool CMSCollector::is_cms_reachable(HeapWord* addr) { 2881 // We could "guarantee" below, rather than assert, but I'll 2882 // leave these as "asserts" so that an adventurous debugger 2883 // could try this in the product build provided some subset of 2884 // the conditions were met, provided they were interested in the 2885 // results and knew that the computation below wouldn't interfere 2886 // with other concurrent computations mutating the structures 2887 // being read or written. 2888 assert(SafepointSynchronize::is_at_safepoint(), 2889 "Else mutations in object graph will make answer suspect"); 2890 assert(have_cms_token(), "Should hold cms token"); 2891 assert(haveFreelistLocks(), "must hold free list locks"); 2892 assert_lock_strong(bitMapLock()); 2893 2894 // Clear the marking bit map array before starting, but, just 2895 // for kicks, first report if the given address is already marked 2896 gclog_or_tty->print_cr("Start: Address 0x%x is%s marked", addr, 2897 _markBitMap.isMarked(addr) ? "" : " not"); 2898 2899 if (verify_after_remark()) { 2900 MutexLockerEx x(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag); 2901 bool result = verification_mark_bm()->isMarked(addr); 2902 gclog_or_tty->print_cr("TransitiveMark: Address 0x%x %s marked", addr, 2903 result ? "IS" : "is NOT"); 2904 return result; 2905 } else { 2906 gclog_or_tty->print_cr("Could not compute result"); 2907 return false; 2908 } 2909 } 2910 2911 2912 void 2913 CMSCollector::print_on_error(outputStream* st) { 2914 CMSCollector* collector = ConcurrentMarkSweepGeneration::_collector; 2915 if (collector != NULL) { 2916 CMSBitMap* bitmap = &collector->_markBitMap; 2917 st->print_cr("Marking Bits: (CMSBitMap*) " PTR_FORMAT, bitmap); 2918 bitmap->print_on_error(st, " Bits: "); 2919 2920 st->cr(); 2921 2922 CMSBitMap* mut_bitmap = &collector->_modUnionTable; 2923 st->print_cr("Mod Union Table: (CMSBitMap*) " PTR_FORMAT, mut_bitmap); 2924 mut_bitmap->print_on_error(st, " Bits: "); 2925 } 2926 } 2927 2928 //////////////////////////////////////////////////////// 2929 // CMS Verification Support 2930 //////////////////////////////////////////////////////// 2931 // Following the remark phase, the following invariant 2932 // should hold -- each object in the CMS heap which is 2933 // marked in markBitMap() should be marked in the verification_mark_bm(). 2934 2935 class VerifyMarkedClosure: public BitMapClosure { 2936 CMSBitMap* _marks; 2937 bool _failed; 2938 2939 public: 2940 VerifyMarkedClosure(CMSBitMap* bm): _marks(bm), _failed(false) {} 2941 2942 bool do_bit(size_t offset) { 2943 HeapWord* addr = _marks->offsetToHeapWord(offset); 2944 if (!_marks->isMarked(addr)) { 2945 oop(addr)->print_on(gclog_or_tty); 2946 gclog_or_tty->print_cr(" ("INTPTR_FORMAT" should have been marked)", addr); 2947 _failed = true; 2948 } 2949 return true; 2950 } 2951 2952 bool failed() { return _failed; } 2953 }; 2954 2955 bool CMSCollector::verify_after_remark(bool silent) { 2956 if (!silent) gclog_or_tty->print(" [Verifying CMS Marking... "); 2957 MutexLockerEx ml(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag); 2958 static bool init = false; 2959 2960 assert(SafepointSynchronize::is_at_safepoint(), 2961 "Else mutations in object graph will make answer suspect"); 2962 assert(have_cms_token(), 2963 "Else there may be mutual interference in use of " 2964 " verification data structures"); 2965 assert(_collectorState > Marking && _collectorState <= Sweeping, 2966 "Else marking info checked here may be obsolete"); 2967 assert(haveFreelistLocks(), "must hold free list locks"); 2968 assert_lock_strong(bitMapLock()); 2969 2970 2971 // Allocate marking bit map if not already allocated 2972 if (!init) { // first time 2973 if (!verification_mark_bm()->allocate(_span)) { 2974 return false; 2975 } 2976 init = true; 2977 } 2978 2979 assert(verification_mark_stack()->isEmpty(), "Should be empty"); 2980 2981 // Turn off refs discovery -- so we will be tracing through refs. 2982 // This is as intended, because by this time 2983 // GC must already have cleared any refs that need to be cleared, 2984 // and traced those that need to be marked; moreover, 2985 // the marking done here is not going to interfere in any 2986 // way with the marking information used by GC. 2987 NoRefDiscovery no_discovery(ref_processor()); 2988 2989 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;) 2990 2991 // Clear any marks from a previous round 2992 verification_mark_bm()->clear_all(); 2993 assert(verification_mark_stack()->isEmpty(), "markStack should be empty"); 2994 verify_work_stacks_empty(); 2995 2996 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2997 gch->ensure_parsability(false); // fill TLABs, but no need to retire them 2998 // Update the saved marks which may affect the root scans. 2999 gch->save_marks(); 3000 3001 if (CMSRemarkVerifyVariant == 1) { 3002 // In this first variant of verification, we complete 3003 // all marking, then check if the new marks-vector is 3004 // a subset of the CMS marks-vector. 3005 verify_after_remark_work_1(); 3006 } else if (CMSRemarkVerifyVariant == 2) { 3007 // In this second variant of verification, we flag an error 3008 // (i.e. an object reachable in the new marks-vector not reachable 3009 // in the CMS marks-vector) immediately, also indicating the 3010 // identify of an object (A) that references the unmarked object (B) -- 3011 // presumably, a mutation to A failed to be picked up by preclean/remark? 3012 verify_after_remark_work_2(); 3013 } else { 3014 warning("Unrecognized value %d for CMSRemarkVerifyVariant", 3015 CMSRemarkVerifyVariant); 3016 } 3017 if (!silent) gclog_or_tty->print(" done] "); 3018 return true; 3019 } 3020 3021 void CMSCollector::verify_after_remark_work_1() { 3022 ResourceMark rm; 3023 HandleMark hm; 3024 GenCollectedHeap* gch = GenCollectedHeap::heap(); 3025 3026 // Get a clear set of claim bits for the strong roots processing to work with. 3027 ClassLoaderDataGraph::clear_claimed_marks(); 3028 3029 // Mark from roots one level into CMS 3030 MarkRefsIntoClosure notOlder(_span, verification_mark_bm()); 3031 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 3032 3033 gch->gen_process_strong_roots(_cmsGen->level(), 3034 true, // younger gens are roots 3035 true, // activate StrongRootsScope 3036 SharedHeap::ScanningOption(roots_scanning_options()), 3037 ¬Older, 3038 true, // walk code active on stacks 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 true, // walk code active on stacks 3106 NULL, 3107 &klass_closure); 3108 3109 // Now mark from the roots 3110 MarkFromRootsVerifyClosure markFromRootsClosure(this, _span, 3111 verification_mark_bm(), markBitMap(), verification_mark_stack()); 3112 assert(_restart_addr == NULL, "Expected pre-condition"); 3113 verification_mark_bm()->iterate(&markFromRootsClosure); 3114 while (_restart_addr != NULL) { 3115 // Deal with stack overflow: by restarting at the indicated 3116 // address. 3117 HeapWord* ra = _restart_addr; 3118 markFromRootsClosure.reset(ra); 3119 _restart_addr = NULL; 3120 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end()); 3121 } 3122 assert(verification_mark_stack()->isEmpty(), "Should have been drained"); 3123 verify_work_stacks_empty(); 3124 3125 VerifyKlassOopsKlassClosure verify_klass_oops(verification_mark_bm()); 3126 ClassLoaderDataGraph::classes_do(&verify_klass_oops); 3127 3128 // Marking completed -- now verify that each bit marked in 3129 // verification_mark_bm() is also marked in markBitMap(); flag all 3130 // errors by printing corresponding objects. 3131 VerifyMarkedClosure vcl(markBitMap()); 3132 verification_mark_bm()->iterate(&vcl); 3133 assert(!vcl.failed(), "Else verification above should not have succeeded"); 3134 } 3135 3136 void ConcurrentMarkSweepGeneration::save_marks() { 3137 // delegate to CMS space 3138 cmsSpace()->save_marks(); 3139 for (uint i = 0; i < ParallelGCThreads; i++) { 3140 _par_gc_thread_states[i]->promo.startTrackingPromotions(); 3141 } 3142 } 3143 3144 bool ConcurrentMarkSweepGeneration::no_allocs_since_save_marks() { 3145 return cmsSpace()->no_allocs_since_save_marks(); 3146 } 3147 3148 #define CMS_SINCE_SAVE_MARKS_DEFN(OopClosureType, nv_suffix) \ 3149 \ 3150 void ConcurrentMarkSweepGeneration:: \ 3151 oop_since_save_marks_iterate##nv_suffix(OopClosureType* cl) { \ 3152 cl->set_generation(this); \ 3153 cmsSpace()->oop_since_save_marks_iterate##nv_suffix(cl); \ 3154 cl->reset_generation(); \ 3155 save_marks(); \ 3156 } 3157 3158 ALL_SINCE_SAVE_MARKS_CLOSURES(CMS_SINCE_SAVE_MARKS_DEFN) 3159 3160 void 3161 ConcurrentMarkSweepGeneration::younger_refs_iterate(OopsInGenClosure* cl) { 3162 cl->set_generation(this); 3163 younger_refs_in_space_iterate(_cmsSpace, cl); 3164 cl->reset_generation(); 3165 } 3166 3167 void 3168 ConcurrentMarkSweepGeneration::oop_iterate(MemRegion mr, ExtendedOopClosure* cl) { 3169 if (freelistLock()->owned_by_self()) { 3170 Generation::oop_iterate(mr, cl); 3171 } else { 3172 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 3173 Generation::oop_iterate(mr, cl); 3174 } 3175 } 3176 3177 void 3178 ConcurrentMarkSweepGeneration::oop_iterate(ExtendedOopClosure* cl) { 3179 if (freelistLock()->owned_by_self()) { 3180 Generation::oop_iterate(cl); 3181 } else { 3182 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 3183 Generation::oop_iterate(cl); 3184 } 3185 } 3186 3187 void 3188 ConcurrentMarkSweepGeneration::object_iterate(ObjectClosure* cl) { 3189 if (freelistLock()->owned_by_self()) { 3190 Generation::object_iterate(cl); 3191 } else { 3192 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 3193 Generation::object_iterate(cl); 3194 } 3195 } 3196 3197 void 3198 ConcurrentMarkSweepGeneration::safe_object_iterate(ObjectClosure* cl) { 3199 if (freelistLock()->owned_by_self()) { 3200 Generation::safe_object_iterate(cl); 3201 } else { 3202 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 3203 Generation::safe_object_iterate(cl); 3204 } 3205 } 3206 3207 void 3208 ConcurrentMarkSweepGeneration::post_compact() { 3209 } 3210 3211 void 3212 ConcurrentMarkSweepGeneration::prepare_for_verify() { 3213 // Fix the linear allocation blocks to look like free blocks. 3214 3215 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those 3216 // are not called when the heap is verified during universe initialization and 3217 // at vm shutdown. 3218 if (freelistLock()->owned_by_self()) { 3219 cmsSpace()->prepare_for_verify(); 3220 } else { 3221 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag); 3222 cmsSpace()->prepare_for_verify(); 3223 } 3224 } 3225 3226 void 3227 ConcurrentMarkSweepGeneration::verify() { 3228 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those 3229 // are not called when the heap is verified during universe initialization and 3230 // at vm shutdown. 3231 if (freelistLock()->owned_by_self()) { 3232 cmsSpace()->verify(); 3233 } else { 3234 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag); 3235 cmsSpace()->verify(); 3236 } 3237 } 3238 3239 void CMSCollector::verify() { 3240 _cmsGen->verify(); 3241 } 3242 3243 #ifndef PRODUCT 3244 bool CMSCollector::overflow_list_is_empty() const { 3245 assert(_num_par_pushes >= 0, "Inconsistency"); 3246 if (_overflow_list == NULL) { 3247 assert(_num_par_pushes == 0, "Inconsistency"); 3248 } 3249 return _overflow_list == NULL; 3250 } 3251 3252 // The methods verify_work_stacks_empty() and verify_overflow_empty() 3253 // merely consolidate assertion checks that appear to occur together frequently. 3254 void CMSCollector::verify_work_stacks_empty() const { 3255 assert(_markStack.isEmpty(), "Marking stack should be empty"); 3256 assert(overflow_list_is_empty(), "Overflow list should be empty"); 3257 } 3258 3259 void CMSCollector::verify_overflow_empty() const { 3260 assert(overflow_list_is_empty(), "Overflow list should be empty"); 3261 assert(no_preserved_marks(), "No preserved marks"); 3262 } 3263 #endif // PRODUCT 3264 3265 // Decide if we want to enable class unloading as part of the 3266 // ensuing concurrent GC cycle. We will collect and 3267 // unload classes if it's the case that: 3268 // (1) an explicit gc request has been made and the flag 3269 // ExplicitGCInvokesConcurrentAndUnloadsClasses is set, OR 3270 // (2) (a) class unloading is enabled at the command line, and 3271 // (b) old gen is getting really full 3272 // NOTE: Provided there is no change in the state of the heap between 3273 // calls to this method, it should have idempotent results. Moreover, 3274 // its results should be monotonically increasing (i.e. going from 0 to 1, 3275 // but not 1 to 0) between successive calls between which the heap was 3276 // not collected. For the implementation below, it must thus rely on 3277 // the property that concurrent_cycles_since_last_unload() 3278 // will not decrease unless a collection cycle happened and that 3279 // _cmsGen->is_too_full() are 3280 // themselves also monotonic in that sense. See check_monotonicity() 3281 // below. 3282 void CMSCollector::update_should_unload_classes() { 3283 _should_unload_classes = false; 3284 // Condition 1 above 3285 if (_full_gc_requested && ExplicitGCInvokesConcurrentAndUnloadsClasses) { 3286 _should_unload_classes = true; 3287 } else if (CMSClassUnloadingEnabled) { // Condition 2.a above 3288 // Disjuncts 2.b.(i,ii,iii) above 3289 _should_unload_classes = (concurrent_cycles_since_last_unload() >= 3290 CMSClassUnloadingMaxInterval) 3291 || _cmsGen->is_too_full(); 3292 } 3293 } 3294 3295 bool ConcurrentMarkSweepGeneration::is_too_full() const { 3296 bool res = should_concurrent_collect(); 3297 res = res && (occupancy() > (double)CMSIsTooFullPercentage/100.0); 3298 return res; 3299 } 3300 3301 void CMSCollector::setup_cms_unloading_and_verification_state() { 3302 const bool should_verify = VerifyBeforeGC || VerifyAfterGC || VerifyDuringGC 3303 || VerifyBeforeExit; 3304 const int rso = SharedHeap::SO_Strings | SharedHeap::SO_AllCodeCache; 3305 3306 // We set the proper root for this CMS cycle here. 3307 if (should_unload_classes()) { // Should unload classes this cycle 3308 remove_root_scanning_option(SharedHeap::SO_AllClasses); 3309 add_root_scanning_option(SharedHeap::SO_SystemClasses); 3310 remove_root_scanning_option(rso); // Shrink the root set appropriately 3311 set_verifying(should_verify); // Set verification state for this cycle 3312 return; // Nothing else needs to be done at this time 3313 } 3314 3315 // Not unloading classes this cycle 3316 assert(!should_unload_classes(), "Inconsitency!"); 3317 remove_root_scanning_option(SharedHeap::SO_SystemClasses); 3318 add_root_scanning_option(SharedHeap::SO_AllClasses); 3319 3320 if ((!verifying() || unloaded_classes_last_cycle()) && should_verify) { 3321 // Include symbols, strings and code cache elements to prevent their resurrection. 3322 add_root_scanning_option(rso); 3323 set_verifying(true); 3324 } else if (verifying() && !should_verify) { 3325 // We were verifying, but some verification flags got disabled. 3326 set_verifying(false); 3327 // Exclude symbols, strings and code cache elements from root scanning to 3328 // reduce IM and RM pauses. 3329 remove_root_scanning_option(rso); 3330 } 3331 } 3332 3333 3334 #ifndef PRODUCT 3335 HeapWord* CMSCollector::block_start(const void* p) const { 3336 const HeapWord* addr = (HeapWord*)p; 3337 if (_span.contains(p)) { 3338 if (_cmsGen->cmsSpace()->is_in_reserved(addr)) { 3339 return _cmsGen->cmsSpace()->block_start(p); 3340 } 3341 } 3342 return NULL; 3343 } 3344 #endif 3345 3346 HeapWord* 3347 ConcurrentMarkSweepGeneration::expand_and_allocate(size_t word_size, 3348 bool tlab, 3349 bool parallel) { 3350 CMSSynchronousYieldRequest yr; 3351 assert(!tlab, "Can't deal with TLAB allocation"); 3352 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 3353 expand(word_size*HeapWordSize, MinHeapDeltaBytes, 3354 CMSExpansionCause::_satisfy_allocation); 3355 if (GCExpandToAllocateDelayMillis > 0) { 3356 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 3357 } 3358 return have_lock_and_allocate(word_size, tlab); 3359 } 3360 3361 // YSR: All of this generation expansion/shrinking stuff is an exact copy of 3362 // OneContigSpaceCardGeneration, which makes me wonder if we should move this 3363 // to CardGeneration and share it... 3364 bool ConcurrentMarkSweepGeneration::expand(size_t bytes, size_t expand_bytes) { 3365 return CardGeneration::expand(bytes, expand_bytes); 3366 } 3367 3368 void ConcurrentMarkSweepGeneration::expand(size_t bytes, size_t expand_bytes, 3369 CMSExpansionCause::Cause cause) 3370 { 3371 3372 bool success = expand(bytes, expand_bytes); 3373 3374 // remember why we expanded; this information is used 3375 // by shouldConcurrentCollect() when making decisions on whether to start 3376 // a new CMS cycle. 3377 if (success) { 3378 set_expansion_cause(cause); 3379 if (PrintGCDetails && Verbose) { 3380 gclog_or_tty->print_cr("Expanded CMS gen for %s", 3381 CMSExpansionCause::to_string(cause)); 3382 } 3383 } 3384 } 3385 3386 HeapWord* ConcurrentMarkSweepGeneration::expand_and_par_lab_allocate(CMSParGCThreadState* ps, size_t word_sz) { 3387 HeapWord* res = NULL; 3388 MutexLocker x(ParGCRareEvent_lock); 3389 while (true) { 3390 // Expansion by some other thread might make alloc OK now: 3391 res = ps->lab.alloc(word_sz); 3392 if (res != NULL) return res; 3393 // If there's not enough expansion space available, give up. 3394 if (_virtual_space.uncommitted_size() < (word_sz * HeapWordSize)) { 3395 return NULL; 3396 } 3397 // Otherwise, we try expansion. 3398 expand(word_sz*HeapWordSize, MinHeapDeltaBytes, 3399 CMSExpansionCause::_allocate_par_lab); 3400 // Now go around the loop and try alloc again; 3401 // A competing par_promote might beat us to the expansion space, 3402 // so we may go around the loop again if promotion fails again. 3403 if (GCExpandToAllocateDelayMillis > 0) { 3404 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 3405 } 3406 } 3407 } 3408 3409 3410 bool ConcurrentMarkSweepGeneration::expand_and_ensure_spooling_space( 3411 PromotionInfo* promo) { 3412 MutexLocker x(ParGCRareEvent_lock); 3413 size_t refill_size_bytes = promo->refillSize() * HeapWordSize; 3414 while (true) { 3415 // Expansion by some other thread might make alloc OK now: 3416 if (promo->ensure_spooling_space()) { 3417 assert(promo->has_spooling_space(), 3418 "Post-condition of successful ensure_spooling_space()"); 3419 return true; 3420 } 3421 // If there's not enough expansion space available, give up. 3422 if (_virtual_space.uncommitted_size() < refill_size_bytes) { 3423 return false; 3424 } 3425 // Otherwise, we try expansion. 3426 expand(refill_size_bytes, MinHeapDeltaBytes, 3427 CMSExpansionCause::_allocate_par_spooling_space); 3428 // Now go around the loop and try alloc again; 3429 // A competing allocation might beat us to the expansion space, 3430 // so we may go around the loop again if allocation fails again. 3431 if (GCExpandToAllocateDelayMillis > 0) { 3432 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 3433 } 3434 } 3435 } 3436 3437 3438 void ConcurrentMarkSweepGeneration::shrink_by(size_t bytes) { 3439 assert_locked_or_safepoint(ExpandHeap_lock); 3440 // Shrink committed space 3441 _virtual_space.shrink_by(bytes); 3442 // Shrink space; this also shrinks the space's BOT 3443 _cmsSpace->set_end((HeapWord*) _virtual_space.high()); 3444 size_t new_word_size = heap_word_size(_cmsSpace->capacity()); 3445 // Shrink the shared block offset array 3446 _bts->resize(new_word_size); 3447 MemRegion mr(_cmsSpace->bottom(), new_word_size); 3448 // Shrink the card table 3449 Universe::heap()->barrier_set()->resize_covered_region(mr); 3450 3451 if (Verbose && PrintGC) { 3452 size_t new_mem_size = _virtual_space.committed_size(); 3453 size_t old_mem_size = new_mem_size + bytes; 3454 gclog_or_tty->print_cr("Shrinking %s from " SIZE_FORMAT "K to " SIZE_FORMAT "K", 3455 name(), old_mem_size/K, new_mem_size/K); 3456 } 3457 } 3458 3459 void ConcurrentMarkSweepGeneration::shrink(size_t bytes) { 3460 assert_locked_or_safepoint(Heap_lock); 3461 size_t size = ReservedSpace::page_align_size_down(bytes); 3462 // Only shrink if a compaction was done so that all the free space 3463 // in the generation is in a contiguous block at the end. 3464 if (size > 0 && did_compact()) { 3465 shrink_by(size); 3466 } 3467 } 3468 3469 bool ConcurrentMarkSweepGeneration::grow_by(size_t bytes) { 3470 assert_locked_or_safepoint(Heap_lock); 3471 bool result = _virtual_space.expand_by(bytes); 3472 if (result) { 3473 size_t new_word_size = 3474 heap_word_size(_virtual_space.committed_size()); 3475 MemRegion mr(_cmsSpace->bottom(), new_word_size); 3476 _bts->resize(new_word_size); // resize the block offset shared array 3477 Universe::heap()->barrier_set()->resize_covered_region(mr); 3478 // Hmmmm... why doesn't CFLS::set_end verify locking? 3479 // This is quite ugly; FIX ME XXX 3480 _cmsSpace->assert_locked(freelistLock()); 3481 _cmsSpace->set_end((HeapWord*)_virtual_space.high()); 3482 3483 // update the space and generation capacity counters 3484 if (UsePerfData) { 3485 _space_counters->update_capacity(); 3486 _gen_counters->update_all(); 3487 } 3488 3489 if (Verbose && PrintGC) { 3490 size_t new_mem_size = _virtual_space.committed_size(); 3491 size_t old_mem_size = new_mem_size - bytes; 3492 gclog_or_tty->print_cr("Expanding %s from " SIZE_FORMAT "K by " SIZE_FORMAT "K to " SIZE_FORMAT "K", 3493 name(), old_mem_size/K, bytes/K, new_mem_size/K); 3494 } 3495 } 3496 return result; 3497 } 3498 3499 bool ConcurrentMarkSweepGeneration::grow_to_reserved() { 3500 assert_locked_or_safepoint(Heap_lock); 3501 bool success = true; 3502 const size_t remaining_bytes = _virtual_space.uncommitted_size(); 3503 if (remaining_bytes > 0) { 3504 success = grow_by(remaining_bytes); 3505 DEBUG_ONLY(if (!success) warning("grow to reserved failed");) 3506 } 3507 return success; 3508 } 3509 3510 void ConcurrentMarkSweepGeneration::shrink_free_list_by(size_t bytes) { 3511 assert_locked_or_safepoint(Heap_lock); 3512 assert_lock_strong(freelistLock()); 3513 if (PrintGCDetails && Verbose) { 3514 warning("Shrinking of CMS not yet implemented"); 3515 } 3516 return; 3517 } 3518 3519 3520 // Simple ctor/dtor wrapper for accounting & timer chores around concurrent 3521 // phases. 3522 class CMSPhaseAccounting: public StackObj { 3523 public: 3524 CMSPhaseAccounting(CMSCollector *collector, 3525 const char *phase, 3526 bool print_cr = true); 3527 ~CMSPhaseAccounting(); 3528 3529 private: 3530 CMSCollector *_collector; 3531 const char *_phase; 3532 elapsedTimer _wallclock; 3533 bool _print_cr; 3534 3535 public: 3536 // Not MT-safe; so do not pass around these StackObj's 3537 // where they may be accessed by other threads. 3538 jlong wallclock_millis() { 3539 assert(_wallclock.is_active(), "Wall clock should not stop"); 3540 _wallclock.stop(); // to record time 3541 jlong ret = _wallclock.milliseconds(); 3542 _wallclock.start(); // restart 3543 return ret; 3544 } 3545 }; 3546 3547 CMSPhaseAccounting::CMSPhaseAccounting(CMSCollector *collector, 3548 const char *phase, 3549 bool print_cr) : 3550 _collector(collector), _phase(phase), _print_cr(print_cr) { 3551 3552 if (PrintCMSStatistics != 0) { 3553 _collector->resetYields(); 3554 } 3555 if (PrintGCDetails) { 3556 gclog_or_tty->date_stamp(PrintGCDateStamps); 3557 gclog_or_tty->stamp(PrintGCTimeStamps); 3558 gclog_or_tty->print_cr("[%s-concurrent-%s-start]", 3559 _collector->cmsGen()->short_name(), _phase); 3560 } 3561 _collector->resetTimer(); 3562 _wallclock.start(); 3563 _collector->startTimer(); 3564 } 3565 3566 CMSPhaseAccounting::~CMSPhaseAccounting() { 3567 assert(_wallclock.is_active(), "Wall clock should not have stopped"); 3568 _collector->stopTimer(); 3569 _wallclock.stop(); 3570 if (PrintGCDetails) { 3571 gclog_or_tty->date_stamp(PrintGCDateStamps); 3572 gclog_or_tty->stamp(PrintGCTimeStamps); 3573 gclog_or_tty->print("[%s-concurrent-%s: %3.3f/%3.3f secs]", 3574 _collector->cmsGen()->short_name(), 3575 _phase, _collector->timerValue(), _wallclock.seconds()); 3576 if (_print_cr) { 3577 gclog_or_tty->print_cr(""); 3578 } 3579 if (PrintCMSStatistics != 0) { 3580 gclog_or_tty->print_cr(" (CMS-concurrent-%s yielded %d times)", _phase, 3581 _collector->yields()); 3582 } 3583 } 3584 } 3585 3586 // CMS work 3587 3588 // The common parts of CMSParInitialMarkTask and CMSParRemarkTask. 3589 class CMSParMarkTask : public AbstractGangTask { 3590 protected: 3591 CMSCollector* _collector; 3592 int _n_workers; 3593 CMSParMarkTask(const char* name, CMSCollector* collector, int n_workers) : 3594 AbstractGangTask(name), 3595 _collector(collector), 3596 _n_workers(n_workers) {} 3597 // Work method in support of parallel rescan ... of young gen spaces 3598 void do_young_space_rescan(uint worker_id, OopsInGenClosure* cl, 3599 ContiguousSpace* space, 3600 HeapWord** chunk_array, size_t chunk_top); 3601 void work_on_young_gen_roots(uint worker_id, OopsInGenClosure* cl); 3602 }; 3603 3604 // Parallel initial mark task 3605 class CMSParInitialMarkTask: public CMSParMarkTask { 3606 public: 3607 CMSParInitialMarkTask(CMSCollector* collector, int n_workers) : 3608 CMSParMarkTask("Scan roots and young gen for initial mark in parallel", 3609 collector, n_workers) {} 3610 void work(uint worker_id); 3611 }; 3612 3613 // Checkpoint the roots into this generation from outside 3614 // this generation. [Note this initial checkpoint need only 3615 // be approximate -- we'll do a catch up phase subsequently.] 3616 void CMSCollector::checkpointRootsInitial(bool asynch) { 3617 assert(_collectorState == InitialMarking, "Wrong collector state"); 3618 check_correct_thread_executing(); 3619 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 3620 3621 save_heap_summary(); 3622 report_heap_summary(GCWhen::BeforeGC); 3623 3624 ReferenceProcessor* rp = ref_processor(); 3625 SpecializationStats::clear(); 3626 assert(_restart_addr == NULL, "Control point invariant"); 3627 if (asynch) { 3628 // acquire locks for subsequent manipulations 3629 MutexLockerEx x(bitMapLock(), 3630 Mutex::_no_safepoint_check_flag); 3631 checkpointRootsInitialWork(asynch); 3632 // enable ("weak") refs discovery 3633 rp->enable_discovery(true /*verify_disabled*/, true /*check_no_refs*/); 3634 _collectorState = Marking; 3635 } else { 3636 // (Weak) Refs discovery: this is controlled from genCollectedHeap::do_collection 3637 // which recognizes if we are a CMS generation, and doesn't try to turn on 3638 // discovery; verify that they aren't meddling. 3639 assert(!rp->discovery_is_atomic(), 3640 "incorrect setting of discovery predicate"); 3641 assert(!rp->discovery_enabled(), "genCollectedHeap shouldn't control " 3642 "ref discovery for this generation kind"); 3643 // already have locks 3644 checkpointRootsInitialWork(asynch); 3645 // now enable ("weak") refs discovery 3646 rp->enable_discovery(true /*verify_disabled*/, false /*verify_no_refs*/); 3647 _collectorState = Marking; 3648 } 3649 SpecializationStats::print(); 3650 } 3651 3652 void CMSCollector::checkpointRootsInitialWork(bool asynch) { 3653 assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped"); 3654 assert(_collectorState == InitialMarking, "just checking"); 3655 3656 // If there has not been a GC[n-1] since last GC[n] cycle completed, 3657 // precede our marking with a collection of all 3658 // younger generations to keep floating garbage to a minimum. 3659 // XXX: we won't do this for now -- it's an optimization to be done later. 3660 3661 // already have locks 3662 assert_lock_strong(bitMapLock()); 3663 assert(_markBitMap.isAllClear(), "was reset at end of previous cycle"); 3664 3665 // Setup the verification and class unloading state for this 3666 // CMS collection cycle. 3667 setup_cms_unloading_and_verification_state(); 3668 3669 NOT_PRODUCT(GCTraceTime t("\ncheckpointRootsInitialWork", 3670 PrintGCDetails && Verbose, true, _gc_timer_cm);) 3671 if (UseAdaptiveSizePolicy) { 3672 size_policy()->checkpoint_roots_initial_begin(); 3673 } 3674 3675 // Reset all the PLAB chunk arrays if necessary. 3676 if (_survivor_plab_array != NULL && !CMSPLABRecordAlways) { 3677 reset_survivor_plab_arrays(); 3678 } 3679 3680 ResourceMark rm; 3681 HandleMark hm; 3682 3683 FalseClosure falseClosure; 3684 // In the case of a synchronous collection, we will elide the 3685 // remark step, so it's important to catch all the nmethod oops 3686 // in this step. 3687 // The final 'true' flag to gen_process_strong_roots will ensure this. 3688 // If 'async' is true, we can relax the nmethod tracing. 3689 MarkRefsIntoClosure notOlder(_span, &_markBitMap); 3690 GenCollectedHeap* gch = GenCollectedHeap::heap(); 3691 3692 verify_work_stacks_empty(); 3693 verify_overflow_empty(); 3694 3695 gch->ensure_parsability(false); // fill TLABs, but no need to retire them 3696 // Update the saved marks which may affect the root scans. 3697 gch->save_marks(); 3698 3699 // weak reference processing has not started yet. 3700 ref_processor()->set_enqueuing_is_done(false); 3701 3702 // Need to remember all newly created CLDs, 3703 // so that we can guarantee that the remark finds them. 3704 ClassLoaderDataGraph::remember_new_clds(true); 3705 3706 // Whenever a CLD is found, it will be claimed before proceeding to mark 3707 // the klasses. The claimed marks need to be cleared before marking starts. 3708 ClassLoaderDataGraph::clear_claimed_marks(); 3709 3710 if (CMSPrintEdenSurvivorChunks) { 3711 print_eden_and_survivor_chunk_arrays(); 3712 } 3713 3714 { 3715 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;) 3716 if (CMSParallelInitialMarkEnabled && CollectedHeap::use_parallel_gc_threads()) { 3717 // The parallel version. 3718 FlexibleWorkGang* workers = gch->workers(); 3719 assert(workers != NULL, "Need parallel worker threads."); 3720 int n_workers = workers->active_workers(); 3721 CMSParInitialMarkTask tsk(this, n_workers); 3722 gch->set_par_threads(n_workers); 3723 initialize_sequential_subtasks_for_young_gen_rescan(n_workers); 3724 if (n_workers > 1) { 3725 GenCollectedHeap::StrongRootsScope srs(gch); 3726 workers->run_task(&tsk); 3727 } else { 3728 GenCollectedHeap::StrongRootsScope srs(gch); 3729 tsk.work(0); 3730 } 3731 gch->set_par_threads(0); 3732 } else { 3733 // The serial version. 3734 CMKlassClosure klass_closure(¬Older); 3735 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 3736 gch->gen_process_strong_roots(_cmsGen->level(), 3737 true, // younger gens are roots 3738 true, // activate StrongRootsScope 3739 SharedHeap::ScanningOption(roots_scanning_options()), 3740 ¬Older, 3741 true, // walk all of code cache if (so & SO_AllCodeCache) 3742 NULL, 3743 &klass_closure); 3744 } 3745 } 3746 3747 // Clear mod-union table; it will be dirtied in the prologue of 3748 // CMS generation per each younger generation collection. 3749 3750 assert(_modUnionTable.isAllClear(), 3751 "Was cleared in most recent final checkpoint phase" 3752 " or no bits are set in the gc_prologue before the start of the next " 3753 "subsequent marking phase."); 3754 3755 assert(_ct->klass_rem_set()->mod_union_is_clear(), "Must be"); 3756 3757 // Save the end of the used_region of the constituent generations 3758 // to be used to limit the extent of sweep in each generation. 3759 save_sweep_limits(); 3760 if (UseAdaptiveSizePolicy) { 3761 size_policy()->checkpoint_roots_initial_end(gch->gc_cause()); 3762 } 3763 verify_overflow_empty(); 3764 } 3765 3766 bool CMSCollector::markFromRoots(bool asynch) { 3767 // we might be tempted to assert that: 3768 // assert(asynch == !SafepointSynchronize::is_at_safepoint(), 3769 // "inconsistent argument?"); 3770 // However that wouldn't be right, because it's possible that 3771 // a safepoint is indeed in progress as a younger generation 3772 // stop-the-world GC happens even as we mark in this generation. 3773 assert(_collectorState == Marking, "inconsistent state?"); 3774 check_correct_thread_executing(); 3775 verify_overflow_empty(); 3776 3777 bool res; 3778 if (asynch) { 3779 3780 // Start the timers for adaptive size policy for the concurrent phases 3781 // Do it here so that the foreground MS can use the concurrent 3782 // timer since a foreground MS might has the sweep done concurrently 3783 // or STW. 3784 if (UseAdaptiveSizePolicy) { 3785 size_policy()->concurrent_marking_begin(); 3786 } 3787 3788 // Weak ref discovery note: We may be discovering weak 3789 // refs in this generation concurrent (but interleaved) with 3790 // weak ref discovery by a younger generation collector. 3791 3792 CMSTokenSyncWithLocks ts(true, bitMapLock()); 3793 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 3794 CMSPhaseAccounting pa(this, "mark", !PrintGCDetails); 3795 res = markFromRootsWork(asynch); 3796 if (res) { 3797 _collectorState = Precleaning; 3798 } else { // We failed and a foreground collection wants to take over 3799 assert(_foregroundGCIsActive, "internal state inconsistency"); 3800 assert(_restart_addr == NULL, "foreground will restart from scratch"); 3801 if (PrintGCDetails) { 3802 gclog_or_tty->print_cr("bailing out to foreground collection"); 3803 } 3804 } 3805 if (UseAdaptiveSizePolicy) { 3806 size_policy()->concurrent_marking_end(); 3807 } 3808 } else { 3809 assert(SafepointSynchronize::is_at_safepoint(), 3810 "inconsistent with asynch == false"); 3811 if (UseAdaptiveSizePolicy) { 3812 size_policy()->ms_collection_marking_begin(); 3813 } 3814 // already have locks 3815 res = markFromRootsWork(asynch); 3816 _collectorState = FinalMarking; 3817 if (UseAdaptiveSizePolicy) { 3818 GenCollectedHeap* gch = GenCollectedHeap::heap(); 3819 size_policy()->ms_collection_marking_end(gch->gc_cause()); 3820 } 3821 } 3822 verify_overflow_empty(); 3823 return res; 3824 } 3825 3826 bool CMSCollector::markFromRootsWork(bool asynch) { 3827 // iterate over marked bits in bit map, doing a full scan and mark 3828 // from these roots using the following algorithm: 3829 // . if oop is to the right of the current scan pointer, 3830 // mark corresponding bit (we'll process it later) 3831 // . else (oop is to left of current scan pointer) 3832 // push oop on marking stack 3833 // . drain the marking stack 3834 3835 // Note that when we do a marking step we need to hold the 3836 // bit map lock -- recall that direct allocation (by mutators) 3837 // and promotion (by younger generation collectors) is also 3838 // marking the bit map. [the so-called allocate live policy.] 3839 // Because the implementation of bit map marking is not 3840 // robust wrt simultaneous marking of bits in the same word, 3841 // we need to make sure that there is no such interference 3842 // between concurrent such updates. 3843 3844 // already have locks 3845 assert_lock_strong(bitMapLock()); 3846 3847 verify_work_stacks_empty(); 3848 verify_overflow_empty(); 3849 bool result = false; 3850 if (CMSConcurrentMTEnabled && ConcGCThreads > 0) { 3851 result = do_marking_mt(asynch); 3852 } else { 3853 result = do_marking_st(asynch); 3854 } 3855 return result; 3856 } 3857 3858 // Forward decl 3859 class CMSConcMarkingTask; 3860 3861 class CMSConcMarkingTerminator: public ParallelTaskTerminator { 3862 CMSCollector* _collector; 3863 CMSConcMarkingTask* _task; 3864 public: 3865 virtual void yield(); 3866 3867 // "n_threads" is the number of threads to be terminated. 3868 // "queue_set" is a set of work queues of other threads. 3869 // "collector" is the CMS collector associated with this task terminator. 3870 // "yield" indicates whether we need the gang as a whole to yield. 3871 CMSConcMarkingTerminator(int n_threads, TaskQueueSetSuper* queue_set, CMSCollector* collector) : 3872 ParallelTaskTerminator(n_threads, queue_set), 3873 _collector(collector) { } 3874 3875 void set_task(CMSConcMarkingTask* task) { 3876 _task = task; 3877 } 3878 }; 3879 3880 class CMSConcMarkingTerminatorTerminator: public TerminatorTerminator { 3881 CMSConcMarkingTask* _task; 3882 public: 3883 bool should_exit_termination(); 3884 void set_task(CMSConcMarkingTask* task) { 3885 _task = task; 3886 } 3887 }; 3888 3889 // MT Concurrent Marking Task 3890 class CMSConcMarkingTask: public YieldingFlexibleGangTask { 3891 CMSCollector* _collector; 3892 int _n_workers; // requested/desired # workers 3893 bool _asynch; 3894 bool _result; 3895 CompactibleFreeListSpace* _cms_space; 3896 char _pad_front[64]; // padding to ... 3897 HeapWord* _global_finger; // ... avoid sharing cache line 3898 char _pad_back[64]; 3899 HeapWord* _restart_addr; 3900 3901 // Exposed here for yielding support 3902 Mutex* const _bit_map_lock; 3903 3904 // The per thread work queues, available here for stealing 3905 OopTaskQueueSet* _task_queues; 3906 3907 // Termination (and yielding) support 3908 CMSConcMarkingTerminator _term; 3909 CMSConcMarkingTerminatorTerminator _term_term; 3910 3911 public: 3912 CMSConcMarkingTask(CMSCollector* collector, 3913 CompactibleFreeListSpace* cms_space, 3914 bool asynch, 3915 YieldingFlexibleWorkGang* workers, 3916 OopTaskQueueSet* task_queues): 3917 YieldingFlexibleGangTask("Concurrent marking done multi-threaded"), 3918 _collector(collector), 3919 _cms_space(cms_space), 3920 _asynch(asynch), _n_workers(0), _result(true), 3921 _task_queues(task_queues), 3922 _term(_n_workers, task_queues, _collector), 3923 _bit_map_lock(collector->bitMapLock()) 3924 { 3925 _requested_size = _n_workers; 3926 _term.set_task(this); 3927 _term_term.set_task(this); 3928 _restart_addr = _global_finger = _cms_space->bottom(); 3929 } 3930 3931 3932 OopTaskQueueSet* task_queues() { return _task_queues; } 3933 3934 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 3935 3936 HeapWord** global_finger_addr() { return &_global_finger; } 3937 3938 CMSConcMarkingTerminator* terminator() { return &_term; } 3939 3940 virtual void set_for_termination(int active_workers) { 3941 terminator()->reset_for_reuse(active_workers); 3942 } 3943 3944 void work(uint worker_id); 3945 bool should_yield() { 3946 return ConcurrentMarkSweepThread::should_yield() 3947 && !_collector->foregroundGCIsActive() 3948 && _asynch; 3949 } 3950 3951 virtual void coordinator_yield(); // stuff done by coordinator 3952 bool result() { return _result; } 3953 3954 void reset(HeapWord* ra) { 3955 assert(_global_finger >= _cms_space->end(), "Postcondition of ::work(i)"); 3956 _restart_addr = _global_finger = ra; 3957 _term.reset_for_reuse(); 3958 } 3959 3960 static bool get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 3961 OopTaskQueue* work_q); 3962 3963 private: 3964 void do_scan_and_mark(int i, CompactibleFreeListSpace* sp); 3965 void do_work_steal(int i); 3966 void bump_global_finger(HeapWord* f); 3967 }; 3968 3969 bool CMSConcMarkingTerminatorTerminator::should_exit_termination() { 3970 assert(_task != NULL, "Error"); 3971 return _task->yielding(); 3972 // Note that we do not need the disjunct || _task->should_yield() above 3973 // because we want terminating threads to yield only if the task 3974 // is already in the midst of yielding, which happens only after at least one 3975 // thread has yielded. 3976 } 3977 3978 void CMSConcMarkingTerminator::yield() { 3979 if (_task->should_yield()) { 3980 _task->yield(); 3981 } else { 3982 ParallelTaskTerminator::yield(); 3983 } 3984 } 3985 3986 //////////////////////////////////////////////////////////////// 3987 // Concurrent Marking Algorithm Sketch 3988 //////////////////////////////////////////////////////////////// 3989 // Until all tasks exhausted (both spaces): 3990 // -- claim next available chunk 3991 // -- bump global finger via CAS 3992 // -- find first object that starts in this chunk 3993 // and start scanning bitmap from that position 3994 // -- scan marked objects for oops 3995 // -- CAS-mark target, and if successful: 3996 // . if target oop is above global finger (volatile read) 3997 // nothing to do 3998 // . if target oop is in chunk and above local finger 3999 // then nothing to do 4000 // . else push on work-queue 4001 // -- Deal with possible overflow issues: 4002 // . local work-queue overflow causes stuff to be pushed on 4003 // global (common) overflow queue 4004 // . always first empty local work queue 4005 // . then get a batch of oops from global work queue if any 4006 // . then do work stealing 4007 // -- When all tasks claimed (both spaces) 4008 // and local work queue empty, 4009 // then in a loop do: 4010 // . check global overflow stack; steal a batch of oops and trace 4011 // . try to steal from other threads oif GOS is empty 4012 // . if neither is available, offer termination 4013 // -- Terminate and return result 4014 // 4015 void CMSConcMarkingTask::work(uint worker_id) { 4016 elapsedTimer _timer; 4017 ResourceMark rm; 4018 HandleMark hm; 4019 4020 DEBUG_ONLY(_collector->verify_overflow_empty();) 4021 4022 // Before we begin work, our work queue should be empty 4023 assert(work_queue(worker_id)->size() == 0, "Expected to be empty"); 4024 // Scan the bitmap covering _cms_space, tracing through grey objects. 4025 _timer.start(); 4026 do_scan_and_mark(worker_id, _cms_space); 4027 _timer.stop(); 4028 if (PrintCMSStatistics != 0) { 4029 gclog_or_tty->print_cr("Finished cms space scanning in %dth thread: %3.3f sec", 4030 worker_id, _timer.seconds()); 4031 // XXX: need xxx/xxx type of notation, two timers 4032 } 4033 4034 // ... do work stealing 4035 _timer.reset(); 4036 _timer.start(); 4037 do_work_steal(worker_id); 4038 _timer.stop(); 4039 if (PrintCMSStatistics != 0) { 4040 gclog_or_tty->print_cr("Finished work stealing in %dth thread: %3.3f sec", 4041 worker_id, _timer.seconds()); 4042 // XXX: need xxx/xxx type of notation, two timers 4043 } 4044 assert(_collector->_markStack.isEmpty(), "Should have been emptied"); 4045 assert(work_queue(worker_id)->size() == 0, "Should have been emptied"); 4046 // Note that under the current task protocol, the 4047 // following assertion is true even of the spaces 4048 // expanded since the completion of the concurrent 4049 // marking. XXX This will likely change under a strict 4050 // ABORT semantics. 4051 // After perm removal the comparison was changed to 4052 // greater than or equal to from strictly greater than. 4053 // Before perm removal the highest address sweep would 4054 // have been at the end of perm gen but now is at the 4055 // end of the tenured gen. 4056 assert(_global_finger >= _cms_space->end(), 4057 "All tasks have been completed"); 4058 DEBUG_ONLY(_collector->verify_overflow_empty();) 4059 } 4060 4061 void CMSConcMarkingTask::bump_global_finger(HeapWord* f) { 4062 HeapWord* read = _global_finger; 4063 HeapWord* cur = read; 4064 while (f > read) { 4065 cur = read; 4066 read = (HeapWord*) Atomic::cmpxchg_ptr(f, &_global_finger, cur); 4067 if (cur == read) { 4068 // our cas succeeded 4069 assert(_global_finger >= f, "protocol consistency"); 4070 break; 4071 } 4072 } 4073 } 4074 4075 // This is really inefficient, and should be redone by 4076 // using (not yet available) block-read and -write interfaces to the 4077 // stack and the work_queue. XXX FIX ME !!! 4078 bool CMSConcMarkingTask::get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 4079 OopTaskQueue* work_q) { 4080 // Fast lock-free check 4081 if (ovflw_stk->length() == 0) { 4082 return false; 4083 } 4084 assert(work_q->size() == 0, "Shouldn't steal"); 4085 MutexLockerEx ml(ovflw_stk->par_lock(), 4086 Mutex::_no_safepoint_check_flag); 4087 // Grab up to 1/4 the size of the work queue 4088 size_t num = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 4089 (size_t)ParGCDesiredObjsFromOverflowList); 4090 num = MIN2(num, ovflw_stk->length()); 4091 for (int i = (int) num; i > 0; i--) { 4092 oop cur = ovflw_stk->pop(); 4093 assert(cur != NULL, "Counted wrong?"); 4094 work_q->push(cur); 4095 } 4096 return num > 0; 4097 } 4098 4099 void CMSConcMarkingTask::do_scan_and_mark(int i, CompactibleFreeListSpace* sp) { 4100 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 4101 int n_tasks = pst->n_tasks(); 4102 // We allow that there may be no tasks to do here because 4103 // we are restarting after a stack overflow. 4104 assert(pst->valid() || n_tasks == 0, "Uninitialized use?"); 4105 uint nth_task = 0; 4106 4107 HeapWord* aligned_start = sp->bottom(); 4108 if (sp->used_region().contains(_restart_addr)) { 4109 // Align down to a card boundary for the start of 0th task 4110 // for this space. 4111 aligned_start = 4112 (HeapWord*)align_size_down((uintptr_t)_restart_addr, 4113 CardTableModRefBS::card_size); 4114 } 4115 4116 size_t chunk_size = sp->marking_task_size(); 4117 while (!pst->is_task_claimed(/* reference */ nth_task)) { 4118 // Having claimed the nth task in this space, 4119 // compute the chunk that it corresponds to: 4120 MemRegion span = MemRegion(aligned_start + nth_task*chunk_size, 4121 aligned_start + (nth_task+1)*chunk_size); 4122 // Try and bump the global finger via a CAS; 4123 // note that we need to do the global finger bump 4124 // _before_ taking the intersection below, because 4125 // the task corresponding to that region will be 4126 // deemed done even if the used_region() expands 4127 // because of allocation -- as it almost certainly will 4128 // during start-up while the threads yield in the 4129 // closure below. 4130 HeapWord* finger = span.end(); 4131 bump_global_finger(finger); // atomically 4132 // There are null tasks here corresponding to chunks 4133 // beyond the "top" address of the space. 4134 span = span.intersection(sp->used_region()); 4135 if (!span.is_empty()) { // Non-null task 4136 HeapWord* prev_obj; 4137 assert(!span.contains(_restart_addr) || nth_task == 0, 4138 "Inconsistency"); 4139 if (nth_task == 0) { 4140 // For the 0th task, we'll not need to compute a block_start. 4141 if (span.contains(_restart_addr)) { 4142 // In the case of a restart because of stack overflow, 4143 // we might additionally skip a chunk prefix. 4144 prev_obj = _restart_addr; 4145 } else { 4146 prev_obj = span.start(); 4147 } 4148 } else { 4149 // We want to skip the first object because 4150 // the protocol is to scan any object in its entirety 4151 // that _starts_ in this span; a fortiori, any 4152 // object starting in an earlier span is scanned 4153 // as part of an earlier claimed task. 4154 // Below we use the "careful" version of block_start 4155 // so we do not try to navigate uninitialized objects. 4156 prev_obj = sp->block_start_careful(span.start()); 4157 // Below we use a variant of block_size that uses the 4158 // Printezis bits to avoid waiting for allocated 4159 // objects to become initialized/parsable. 4160 while (prev_obj < span.start()) { 4161 size_t sz = sp->block_size_no_stall(prev_obj, _collector); 4162 if (sz > 0) { 4163 prev_obj += sz; 4164 } else { 4165 // In this case we may end up doing a bit of redundant 4166 // scanning, but that appears unavoidable, short of 4167 // locking the free list locks; see bug 6324141. 4168 break; 4169 } 4170 } 4171 } 4172 if (prev_obj < span.end()) { 4173 MemRegion my_span = MemRegion(prev_obj, span.end()); 4174 // Do the marking work within a non-empty span -- 4175 // the last argument to the constructor indicates whether the 4176 // iteration should be incremental with periodic yields. 4177 Par_MarkFromRootsClosure cl(this, _collector, my_span, 4178 &_collector->_markBitMap, 4179 work_queue(i), 4180 &_collector->_markStack, 4181 _asynch); 4182 _collector->_markBitMap.iterate(&cl, my_span.start(), my_span.end()); 4183 } // else nothing to do for this task 4184 } // else nothing to do for this task 4185 } 4186 // We'd be tempted to assert here that since there are no 4187 // more tasks left to claim in this space, the global_finger 4188 // must exceed space->top() and a fortiori space->end(). However, 4189 // that would not quite be correct because the bumping of 4190 // global_finger occurs strictly after the claiming of a task, 4191 // so by the time we reach here the global finger may not yet 4192 // have been bumped up by the thread that claimed the last 4193 // task. 4194 pst->all_tasks_completed(); 4195 } 4196 4197 class Par_ConcMarkingClosure: public CMSOopClosure { 4198 private: 4199 CMSCollector* _collector; 4200 CMSConcMarkingTask* _task; 4201 MemRegion _span; 4202 CMSBitMap* _bit_map; 4203 CMSMarkStack* _overflow_stack; 4204 OopTaskQueue* _work_queue; 4205 protected: 4206 DO_OOP_WORK_DEFN 4207 public: 4208 Par_ConcMarkingClosure(CMSCollector* collector, CMSConcMarkingTask* task, OopTaskQueue* work_queue, 4209 CMSBitMap* bit_map, CMSMarkStack* overflow_stack): 4210 CMSOopClosure(collector->ref_processor()), 4211 _collector(collector), 4212 _task(task), 4213 _span(collector->_span), 4214 _work_queue(work_queue), 4215 _bit_map(bit_map), 4216 _overflow_stack(overflow_stack) 4217 { } 4218 virtual void do_oop(oop* p); 4219 virtual void do_oop(narrowOop* p); 4220 4221 void trim_queue(size_t max); 4222 void handle_stack_overflow(HeapWord* lost); 4223 void do_yield_check() { 4224 if (_task->should_yield()) { 4225 _task->yield(); 4226 } 4227 } 4228 }; 4229 4230 // Grey object scanning during work stealing phase -- 4231 // the salient assumption here is that any references 4232 // that are in these stolen objects being scanned must 4233 // already have been initialized (else they would not have 4234 // been published), so we do not need to check for 4235 // uninitialized objects before pushing here. 4236 void Par_ConcMarkingClosure::do_oop(oop obj) { 4237 assert(obj->is_oop_or_null(true), "expected an oop or NULL"); 4238 HeapWord* addr = (HeapWord*)obj; 4239 // Check if oop points into the CMS generation 4240 // and is not marked 4241 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 4242 // a white object ... 4243 // If we manage to "claim" the object, by being the 4244 // first thread to mark it, then we push it on our 4245 // marking stack 4246 if (_bit_map->par_mark(addr)) { // ... now grey 4247 // push on work queue (grey set) 4248 bool simulate_overflow = false; 4249 NOT_PRODUCT( 4250 if (CMSMarkStackOverflowALot && 4251 _collector->simulate_overflow()) { 4252 // simulate a stack overflow 4253 simulate_overflow = true; 4254 } 4255 ) 4256 if (simulate_overflow || 4257 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) { 4258 // stack overflow 4259 if (PrintCMSStatistics != 0) { 4260 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 4261 SIZE_FORMAT, _overflow_stack->capacity()); 4262 } 4263 // We cannot assert that the overflow stack is full because 4264 // it may have been emptied since. 4265 assert(simulate_overflow || 4266 _work_queue->size() == _work_queue->max_elems(), 4267 "Else push should have succeeded"); 4268 handle_stack_overflow(addr); 4269 } 4270 } // Else, some other thread got there first 4271 do_yield_check(); 4272 } 4273 } 4274 4275 void Par_ConcMarkingClosure::do_oop(oop* p) { Par_ConcMarkingClosure::do_oop_work(p); } 4276 void Par_ConcMarkingClosure::do_oop(narrowOop* p) { Par_ConcMarkingClosure::do_oop_work(p); } 4277 4278 void Par_ConcMarkingClosure::trim_queue(size_t max) { 4279 while (_work_queue->size() > max) { 4280 oop new_oop; 4281 if (_work_queue->pop_local(new_oop)) { 4282 assert(new_oop->is_oop(), "Should be an oop"); 4283 assert(_bit_map->isMarked((HeapWord*)new_oop), "Grey object"); 4284 assert(_span.contains((HeapWord*)new_oop), "Not in span"); 4285 new_oop->oop_iterate(this); // do_oop() above 4286 do_yield_check(); 4287 } 4288 } 4289 } 4290 4291 // Upon stack overflow, we discard (part of) the stack, 4292 // remembering the least address amongst those discarded 4293 // in CMSCollector's _restart_address. 4294 void Par_ConcMarkingClosure::handle_stack_overflow(HeapWord* lost) { 4295 // We need to do this under a mutex to prevent other 4296 // workers from interfering with the work done below. 4297 MutexLockerEx ml(_overflow_stack->par_lock(), 4298 Mutex::_no_safepoint_check_flag); 4299 // Remember the least grey address discarded 4300 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 4301 _collector->lower_restart_addr(ra); 4302 _overflow_stack->reset(); // discard stack contents 4303 _overflow_stack->expand(); // expand the stack if possible 4304 } 4305 4306 4307 void CMSConcMarkingTask::do_work_steal(int i) { 4308 OopTaskQueue* work_q = work_queue(i); 4309 oop obj_to_scan; 4310 CMSBitMap* bm = &(_collector->_markBitMap); 4311 CMSMarkStack* ovflw = &(_collector->_markStack); 4312 int* seed = _collector->hash_seed(i); 4313 Par_ConcMarkingClosure cl(_collector, this, work_q, bm, ovflw); 4314 while (true) { 4315 cl.trim_queue(0); 4316 assert(work_q->size() == 0, "Should have been emptied above"); 4317 if (get_work_from_overflow_stack(ovflw, work_q)) { 4318 // Can't assert below because the work obtained from the 4319 // overflow stack may already have been stolen from us. 4320 // assert(work_q->size() > 0, "Work from overflow stack"); 4321 continue; 4322 } else if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 4323 assert(obj_to_scan->is_oop(), "Should be an oop"); 4324 assert(bm->isMarked((HeapWord*)obj_to_scan), "Grey object"); 4325 obj_to_scan->oop_iterate(&cl); 4326 } else if (terminator()->offer_termination(&_term_term)) { 4327 assert(work_q->size() == 0, "Impossible!"); 4328 break; 4329 } else if (yielding() || should_yield()) { 4330 yield(); 4331 } 4332 } 4333 } 4334 4335 // This is run by the CMS (coordinator) thread. 4336 void CMSConcMarkingTask::coordinator_yield() { 4337 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 4338 "CMS thread should hold CMS token"); 4339 // First give up the locks, then yield, then re-lock 4340 // We should probably use a constructor/destructor idiom to 4341 // do this unlock/lock or modify the MutexUnlocker class to 4342 // serve our purpose. XXX 4343 assert_lock_strong(_bit_map_lock); 4344 _bit_map_lock->unlock(); 4345 ConcurrentMarkSweepThread::desynchronize(true); 4346 ConcurrentMarkSweepThread::acknowledge_yield_request(); 4347 _collector->stopTimer(); 4348 if (PrintCMSStatistics != 0) { 4349 _collector->incrementYields(); 4350 } 4351 _collector->icms_wait(); 4352 4353 // It is possible for whichever thread initiated the yield request 4354 // not to get a chance to wake up and take the bitmap lock between 4355 // this thread releasing it and reacquiring it. So, while the 4356 // should_yield() flag is on, let's sleep for a bit to give the 4357 // other thread a chance to wake up. The limit imposed on the number 4358 // of iterations is defensive, to avoid any unforseen circumstances 4359 // putting us into an infinite loop. Since it's always been this 4360 // (coordinator_yield()) method that was observed to cause the 4361 // problem, we are using a parameter (CMSCoordinatorYieldSleepCount) 4362 // which is by default non-zero. For the other seven methods that 4363 // also perform the yield operation, as are using a different 4364 // parameter (CMSYieldSleepCount) which is by default zero. This way we 4365 // can enable the sleeping for those methods too, if necessary. 4366 // See 6442774. 4367 // 4368 // We really need to reconsider the synchronization between the GC 4369 // thread and the yield-requesting threads in the future and we 4370 // should really use wait/notify, which is the recommended 4371 // way of doing this type of interaction. Additionally, we should 4372 // consolidate the eight methods that do the yield operation and they 4373 // are almost identical into one for better maintainability and 4374 // readability. See 6445193. 4375 // 4376 // Tony 2006.06.29 4377 for (unsigned i = 0; i < CMSCoordinatorYieldSleepCount && 4378 ConcurrentMarkSweepThread::should_yield() && 4379 !CMSCollector::foregroundGCIsActive(); ++i) { 4380 os::sleep(Thread::current(), 1, false); 4381 ConcurrentMarkSweepThread::acknowledge_yield_request(); 4382 } 4383 4384 ConcurrentMarkSweepThread::synchronize(true); 4385 _bit_map_lock->lock_without_safepoint_check(); 4386 _collector->startTimer(); 4387 } 4388 4389 bool CMSCollector::do_marking_mt(bool asynch) { 4390 assert(ConcGCThreads > 0 && conc_workers() != NULL, "precondition"); 4391 int num_workers = AdaptiveSizePolicy::calc_active_conc_workers( 4392 conc_workers()->total_workers(), 4393 conc_workers()->active_workers(), 4394 Threads::number_of_non_daemon_threads()); 4395 conc_workers()->set_active_workers(num_workers); 4396 4397 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 4398 4399 CMSConcMarkingTask tsk(this, 4400 cms_space, 4401 asynch, 4402 conc_workers(), 4403 task_queues()); 4404 4405 // Since the actual number of workers we get may be different 4406 // from the number we requested above, do we need to do anything different 4407 // below? In particular, may be we need to subclass the SequantialSubTasksDone 4408 // class?? XXX 4409 cms_space ->initialize_sequential_subtasks_for_marking(num_workers); 4410 4411 // Refs discovery is already non-atomic. 4412 assert(!ref_processor()->discovery_is_atomic(), "Should be non-atomic"); 4413 assert(ref_processor()->discovery_is_mt(), "Discovery should be MT"); 4414 conc_workers()->start_task(&tsk); 4415 while (tsk.yielded()) { 4416 tsk.coordinator_yield(); 4417 conc_workers()->continue_task(&tsk); 4418 } 4419 // If the task was aborted, _restart_addr will be non-NULL 4420 assert(tsk.completed() || _restart_addr != NULL, "Inconsistency"); 4421 while (_restart_addr != NULL) { 4422 // XXX For now we do not make use of ABORTED state and have not 4423 // yet implemented the right abort semantics (even in the original 4424 // single-threaded CMS case). That needs some more investigation 4425 // and is deferred for now; see CR# TBF. 07252005YSR. XXX 4426 assert(!CMSAbortSemantics || tsk.aborted(), "Inconsistency"); 4427 // If _restart_addr is non-NULL, a marking stack overflow 4428 // occurred; we need to do a fresh marking iteration from the 4429 // indicated restart address. 4430 if (_foregroundGCIsActive && asynch) { 4431 // We may be running into repeated stack overflows, having 4432 // reached the limit of the stack size, while making very 4433 // slow forward progress. It may be best to bail out and 4434 // let the foreground collector do its job. 4435 // Clear _restart_addr, so that foreground GC 4436 // works from scratch. This avoids the headache of 4437 // a "rescan" which would otherwise be needed because 4438 // of the dirty mod union table & card table. 4439 _restart_addr = NULL; 4440 return false; 4441 } 4442 // Adjust the task to restart from _restart_addr 4443 tsk.reset(_restart_addr); 4444 cms_space ->initialize_sequential_subtasks_for_marking(num_workers, 4445 _restart_addr); 4446 _restart_addr = NULL; 4447 // Get the workers going again 4448 conc_workers()->start_task(&tsk); 4449 while (tsk.yielded()) { 4450 tsk.coordinator_yield(); 4451 conc_workers()->continue_task(&tsk); 4452 } 4453 } 4454 assert(tsk.completed(), "Inconsistency"); 4455 assert(tsk.result() == true, "Inconsistency"); 4456 return true; 4457 } 4458 4459 bool CMSCollector::do_marking_st(bool asynch) { 4460 ResourceMark rm; 4461 HandleMark hm; 4462 4463 // Temporarily make refs discovery single threaded (non-MT) 4464 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false); 4465 MarkFromRootsClosure markFromRootsClosure(this, _span, &_markBitMap, 4466 &_markStack, CMSYield && asynch); 4467 // the last argument to iterate indicates whether the iteration 4468 // should be incremental with periodic yields. 4469 _markBitMap.iterate(&markFromRootsClosure); 4470 // If _restart_addr is non-NULL, a marking stack overflow 4471 // occurred; we need to do a fresh iteration from the 4472 // indicated restart address. 4473 while (_restart_addr != NULL) { 4474 if (_foregroundGCIsActive && asynch) { 4475 // We may be running into repeated stack overflows, having 4476 // reached the limit of the stack size, while making very 4477 // slow forward progress. It may be best to bail out and 4478 // let the foreground collector do its job. 4479 // Clear _restart_addr, so that foreground GC 4480 // works from scratch. This avoids the headache of 4481 // a "rescan" which would otherwise be needed because 4482 // of the dirty mod union table & card table. 4483 _restart_addr = NULL; 4484 return false; // indicating failure to complete marking 4485 } 4486 // Deal with stack overflow: 4487 // we restart marking from _restart_addr 4488 HeapWord* ra = _restart_addr; 4489 markFromRootsClosure.reset(ra); 4490 _restart_addr = NULL; 4491 _markBitMap.iterate(&markFromRootsClosure, ra, _span.end()); 4492 } 4493 return true; 4494 } 4495 4496 void CMSCollector::preclean() { 4497 check_correct_thread_executing(); 4498 assert(Thread::current()->is_ConcurrentGC_thread(), "Wrong thread"); 4499 verify_work_stacks_empty(); 4500 verify_overflow_empty(); 4501 _abort_preclean = false; 4502 if (CMSPrecleaningEnabled) { 4503 if (!CMSEdenChunksRecordAlways) { 4504 _eden_chunk_index = 0; 4505 } 4506 size_t used = get_eden_used(); 4507 size_t capacity = get_eden_capacity(); 4508 // Don't start sampling unless we will get sufficiently 4509 // many samples. 4510 if (used < (capacity/(CMSScheduleRemarkSamplingRatio * 100) 4511 * CMSScheduleRemarkEdenPenetration)) { 4512 _start_sampling = true; 4513 } else { 4514 _start_sampling = false; 4515 } 4516 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 4517 CMSPhaseAccounting pa(this, "preclean", !PrintGCDetails); 4518 preclean_work(CMSPrecleanRefLists1, CMSPrecleanSurvivors1); 4519 } 4520 CMSTokenSync x(true); // is cms thread 4521 if (CMSPrecleaningEnabled) { 4522 sample_eden(); 4523 _collectorState = AbortablePreclean; 4524 } else { 4525 _collectorState = FinalMarking; 4526 } 4527 verify_work_stacks_empty(); 4528 verify_overflow_empty(); 4529 } 4530 4531 // Try and schedule the remark such that young gen 4532 // occupancy is CMSScheduleRemarkEdenPenetration %. 4533 void CMSCollector::abortable_preclean() { 4534 check_correct_thread_executing(); 4535 assert(CMSPrecleaningEnabled, "Inconsistent control state"); 4536 assert(_collectorState == AbortablePreclean, "Inconsistent control state"); 4537 4538 // If Eden's current occupancy is below this threshold, 4539 // immediately schedule the remark; else preclean 4540 // past the next scavenge in an effort to 4541 // schedule the pause as described above. By choosing 4542 // CMSScheduleRemarkEdenSizeThreshold >= max eden size 4543 // we will never do an actual abortable preclean cycle. 4544 if (get_eden_used() > CMSScheduleRemarkEdenSizeThreshold) { 4545 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 4546 CMSPhaseAccounting pa(this, "abortable-preclean", !PrintGCDetails); 4547 // We need more smarts in the abortable preclean 4548 // loop below to deal with cases where allocation 4549 // in young gen is very very slow, and our precleaning 4550 // is running a losing race against a horde of 4551 // mutators intent on flooding us with CMS updates 4552 // (dirty cards). 4553 // One, admittedly dumb, strategy is to give up 4554 // after a certain number of abortable precleaning loops 4555 // or after a certain maximum time. We want to make 4556 // this smarter in the next iteration. 4557 // XXX FIX ME!!! YSR 4558 size_t loops = 0, workdone = 0, cumworkdone = 0, waited = 0; 4559 while (!(should_abort_preclean() || 4560 ConcurrentMarkSweepThread::should_terminate())) { 4561 workdone = preclean_work(CMSPrecleanRefLists2, CMSPrecleanSurvivors2); 4562 cumworkdone += workdone; 4563 loops++; 4564 // Voluntarily terminate abortable preclean phase if we have 4565 // been at it for too long. 4566 if ((CMSMaxAbortablePrecleanLoops != 0) && 4567 loops >= CMSMaxAbortablePrecleanLoops) { 4568 if (PrintGCDetails) { 4569 gclog_or_tty->print(" CMS: abort preclean due to loops "); 4570 } 4571 break; 4572 } 4573 if (pa.wallclock_millis() > CMSMaxAbortablePrecleanTime) { 4574 if (PrintGCDetails) { 4575 gclog_or_tty->print(" CMS: abort preclean due to time "); 4576 } 4577 break; 4578 } 4579 // If we are doing little work each iteration, we should 4580 // take a short break. 4581 if (workdone < CMSAbortablePrecleanMinWorkPerIteration) { 4582 // Sleep for some time, waiting for work to accumulate 4583 stopTimer(); 4584 cmsThread()->wait_on_cms_lock(CMSAbortablePrecleanWaitMillis); 4585 startTimer(); 4586 waited++; 4587 } 4588 } 4589 if (PrintCMSStatistics > 0) { 4590 gclog_or_tty->print(" [%d iterations, %d waits, %d cards)] ", 4591 loops, waited, cumworkdone); 4592 } 4593 } 4594 CMSTokenSync x(true); // is cms thread 4595 if (_collectorState != Idling) { 4596 assert(_collectorState == AbortablePreclean, 4597 "Spontaneous state transition?"); 4598 _collectorState = FinalMarking; 4599 } // Else, a foreground collection completed this CMS cycle. 4600 return; 4601 } 4602 4603 // Respond to an Eden sampling opportunity 4604 void CMSCollector::sample_eden() { 4605 // Make sure a young gc cannot sneak in between our 4606 // reading and recording of a sample. 4607 assert(Thread::current()->is_ConcurrentGC_thread(), 4608 "Only the cms thread may collect Eden samples"); 4609 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 4610 "Should collect samples while holding CMS token"); 4611 if (!_start_sampling) { 4612 return; 4613 } 4614 // When CMSEdenChunksRecordAlways is true, the eden chunk array 4615 // is populated by the young generation. 4616 if (_eden_chunk_array != NULL && !CMSEdenChunksRecordAlways) { 4617 if (_eden_chunk_index < _eden_chunk_capacity) { 4618 _eden_chunk_array[_eden_chunk_index] = *_top_addr; // take sample 4619 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr, 4620 "Unexpected state of Eden"); 4621 // We'd like to check that what we just sampled is an oop-start address; 4622 // however, we cannot do that here since the object may not yet have been 4623 // initialized. So we'll instead do the check when we _use_ this sample 4624 // later. 4625 if (_eden_chunk_index == 0 || 4626 (pointer_delta(_eden_chunk_array[_eden_chunk_index], 4627 _eden_chunk_array[_eden_chunk_index-1]) 4628 >= CMSSamplingGrain)) { 4629 _eden_chunk_index++; // commit sample 4630 } 4631 } 4632 } 4633 if ((_collectorState == AbortablePreclean) && !_abort_preclean) { 4634 size_t used = get_eden_used(); 4635 size_t capacity = get_eden_capacity(); 4636 assert(used <= capacity, "Unexpected state of Eden"); 4637 if (used > (capacity/100 * CMSScheduleRemarkEdenPenetration)) { 4638 _abort_preclean = true; 4639 } 4640 } 4641 } 4642 4643 4644 size_t CMSCollector::preclean_work(bool clean_refs, bool clean_survivor) { 4645 assert(_collectorState == Precleaning || 4646 _collectorState == AbortablePreclean, "incorrect state"); 4647 ResourceMark rm; 4648 HandleMark hm; 4649 4650 // Precleaning is currently not MT but the reference processor 4651 // may be set for MT. Disable it temporarily here. 4652 ReferenceProcessor* rp = ref_processor(); 4653 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(rp, false); 4654 4655 // Do one pass of scrubbing the discovered reference lists 4656 // to remove any reference objects with strongly-reachable 4657 // referents. 4658 if (clean_refs) { 4659 CMSPrecleanRefsYieldClosure yield_cl(this); 4660 assert(rp->span().equals(_span), "Spans should be equal"); 4661 CMSKeepAliveClosure keep_alive(this, _span, &_markBitMap, 4662 &_markStack, true /* preclean */); 4663 CMSDrainMarkingStackClosure complete_trace(this, 4664 _span, &_markBitMap, &_markStack, 4665 &keep_alive, true /* preclean */); 4666 4667 // We don't want this step to interfere with a young 4668 // collection because we don't want to take CPU 4669 // or memory bandwidth away from the young GC threads 4670 // (which may be as many as there are CPUs). 4671 // Note that we don't need to protect ourselves from 4672 // interference with mutators because they can't 4673 // manipulate the discovered reference lists nor affect 4674 // the computed reachability of the referents, the 4675 // only properties manipulated by the precleaning 4676 // of these reference lists. 4677 stopTimer(); 4678 CMSTokenSyncWithLocks x(true /* is cms thread */, 4679 bitMapLock()); 4680 startTimer(); 4681 sample_eden(); 4682 4683 // The following will yield to allow foreground 4684 // collection to proceed promptly. XXX YSR: 4685 // The code in this method may need further 4686 // tweaking for better performance and some restructuring 4687 // for cleaner interfaces. 4688 GCTimer *gc_timer = NULL; // Currently not tracing concurrent phases 4689 rp->preclean_discovered_references( 4690 rp->is_alive_non_header(), &keep_alive, &complete_trace, &yield_cl, 4691 gc_timer); 4692 } 4693 4694 if (clean_survivor) { // preclean the active survivor space(s) 4695 assert(_young_gen->kind() == Generation::DefNew || 4696 _young_gen->kind() == Generation::ParNew || 4697 _young_gen->kind() == Generation::ASParNew, 4698 "incorrect type for cast"); 4699 DefNewGeneration* dng = (DefNewGeneration*)_young_gen; 4700 PushAndMarkClosure pam_cl(this, _span, ref_processor(), 4701 &_markBitMap, &_modUnionTable, 4702 &_markStack, true /* precleaning phase */); 4703 stopTimer(); 4704 CMSTokenSyncWithLocks ts(true /* is cms thread */, 4705 bitMapLock()); 4706 startTimer(); 4707 unsigned int before_count = 4708 GenCollectedHeap::heap()->total_collections(); 4709 SurvivorSpacePrecleanClosure 4710 sss_cl(this, _span, &_markBitMap, &_markStack, 4711 &pam_cl, before_count, CMSYield); 4712 dng->from()->object_iterate_careful(&sss_cl); 4713 dng->to()->object_iterate_careful(&sss_cl); 4714 } 4715 MarkRefsIntoAndScanClosure 4716 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable, 4717 &_markStack, this, CMSYield, 4718 true /* precleaning phase */); 4719 // CAUTION: The following closure has persistent state that may need to 4720 // be reset upon a decrease in the sequence of addresses it 4721 // processes. 4722 ScanMarkedObjectsAgainCarefullyClosure 4723 smoac_cl(this, _span, 4724 &_markBitMap, &_markStack, &mrias_cl, CMSYield); 4725 4726 // Preclean dirty cards in ModUnionTable and CardTable using 4727 // appropriate convergence criterion; 4728 // repeat CMSPrecleanIter times unless we find that 4729 // we are losing. 4730 assert(CMSPrecleanIter < 10, "CMSPrecleanIter is too large"); 4731 assert(CMSPrecleanNumerator < CMSPrecleanDenominator, 4732 "Bad convergence multiplier"); 4733 assert(CMSPrecleanThreshold >= 100, 4734 "Unreasonably low CMSPrecleanThreshold"); 4735 4736 size_t numIter, cumNumCards, lastNumCards, curNumCards; 4737 for (numIter = 0, cumNumCards = lastNumCards = curNumCards = 0; 4738 numIter < CMSPrecleanIter; 4739 numIter++, lastNumCards = curNumCards, cumNumCards += curNumCards) { 4740 curNumCards = preclean_mod_union_table(_cmsGen, &smoac_cl); 4741 if (Verbose && PrintGCDetails) { 4742 gclog_or_tty->print(" (modUnionTable: %d cards)", curNumCards); 4743 } 4744 // Either there are very few dirty cards, so re-mark 4745 // pause will be small anyway, or our pre-cleaning isn't 4746 // that much faster than the rate at which cards are being 4747 // dirtied, so we might as well stop and re-mark since 4748 // precleaning won't improve our re-mark time by much. 4749 if (curNumCards <= CMSPrecleanThreshold || 4750 (numIter > 0 && 4751 (curNumCards * CMSPrecleanDenominator > 4752 lastNumCards * CMSPrecleanNumerator))) { 4753 numIter++; 4754 cumNumCards += curNumCards; 4755 break; 4756 } 4757 } 4758 4759 preclean_klasses(&mrias_cl, _cmsGen->freelistLock()); 4760 4761 curNumCards = preclean_card_table(_cmsGen, &smoac_cl); 4762 cumNumCards += curNumCards; 4763 if (PrintGCDetails && PrintCMSStatistics != 0) { 4764 gclog_or_tty->print_cr(" (cardTable: %d cards, re-scanned %d cards, %d iterations)", 4765 curNumCards, cumNumCards, numIter); 4766 } 4767 return cumNumCards; // as a measure of useful work done 4768 } 4769 4770 // PRECLEANING NOTES: 4771 // Precleaning involves: 4772 // . reading the bits of the modUnionTable and clearing the set bits. 4773 // . For the cards corresponding to the set bits, we scan the 4774 // objects on those cards. This means we need the free_list_lock 4775 // so that we can safely iterate over the CMS space when scanning 4776 // for oops. 4777 // . When we scan the objects, we'll be both reading and setting 4778 // marks in the marking bit map, so we'll need the marking bit map. 4779 // . For protecting _collector_state transitions, we take the CGC_lock. 4780 // Note that any races in the reading of of card table entries by the 4781 // CMS thread on the one hand and the clearing of those entries by the 4782 // VM thread or the setting of those entries by the mutator threads on the 4783 // other are quite benign. However, for efficiency it makes sense to keep 4784 // the VM thread from racing with the CMS thread while the latter is 4785 // dirty card info to the modUnionTable. We therefore also use the 4786 // CGC_lock to protect the reading of the card table and the mod union 4787 // table by the CM thread. 4788 // . We run concurrently with mutator updates, so scanning 4789 // needs to be done carefully -- we should not try to scan 4790 // potentially uninitialized objects. 4791 // 4792 // Locking strategy: While holding the CGC_lock, we scan over and 4793 // reset a maximal dirty range of the mod union / card tables, then lock 4794 // the free_list_lock and bitmap lock to do a full marking, then 4795 // release these locks; and repeat the cycle. This allows for a 4796 // certain amount of fairness in the sharing of these locks between 4797 // the CMS collector on the one hand, and the VM thread and the 4798 // mutators on the other. 4799 4800 // NOTE: preclean_mod_union_table() and preclean_card_table() 4801 // further below are largely identical; if you need to modify 4802 // one of these methods, please check the other method too. 4803 4804 size_t CMSCollector::preclean_mod_union_table( 4805 ConcurrentMarkSweepGeneration* gen, 4806 ScanMarkedObjectsAgainCarefullyClosure* cl) { 4807 verify_work_stacks_empty(); 4808 verify_overflow_empty(); 4809 4810 // strategy: starting with the first card, accumulate contiguous 4811 // ranges of dirty cards; clear these cards, then scan the region 4812 // covered by these cards. 4813 4814 // Since all of the MUT is committed ahead, we can just use 4815 // that, in case the generations expand while we are precleaning. 4816 // It might also be fine to just use the committed part of the 4817 // generation, but we might potentially miss cards when the 4818 // generation is rapidly expanding while we are in the midst 4819 // of precleaning. 4820 HeapWord* startAddr = gen->reserved().start(); 4821 HeapWord* endAddr = gen->reserved().end(); 4822 4823 cl->setFreelistLock(gen->freelistLock()); // needed for yielding 4824 4825 size_t numDirtyCards, cumNumDirtyCards; 4826 HeapWord *nextAddr, *lastAddr; 4827 for (cumNumDirtyCards = numDirtyCards = 0, 4828 nextAddr = lastAddr = startAddr; 4829 nextAddr < endAddr; 4830 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 4831 4832 ResourceMark rm; 4833 HandleMark hm; 4834 4835 MemRegion dirtyRegion; 4836 { 4837 stopTimer(); 4838 // Potential yield point 4839 CMSTokenSync ts(true); 4840 startTimer(); 4841 sample_eden(); 4842 // Get dirty region starting at nextOffset (inclusive), 4843 // simultaneously clearing it. 4844 dirtyRegion = 4845 _modUnionTable.getAndClearMarkedRegion(nextAddr, endAddr); 4846 assert(dirtyRegion.start() >= nextAddr, 4847 "returned region inconsistent?"); 4848 } 4849 // Remember where the next search should begin. 4850 // The returned region (if non-empty) is a right open interval, 4851 // so lastOffset is obtained from the right end of that 4852 // interval. 4853 lastAddr = dirtyRegion.end(); 4854 // Should do something more transparent and less hacky XXX 4855 numDirtyCards = 4856 _modUnionTable.heapWordDiffToOffsetDiff(dirtyRegion.word_size()); 4857 4858 // We'll scan the cards in the dirty region (with periodic 4859 // yields for foreground GC as needed). 4860 if (!dirtyRegion.is_empty()) { 4861 assert(numDirtyCards > 0, "consistency check"); 4862 HeapWord* stop_point = NULL; 4863 stopTimer(); 4864 // Potential yield point 4865 CMSTokenSyncWithLocks ts(true, gen->freelistLock(), 4866 bitMapLock()); 4867 startTimer(); 4868 { 4869 verify_work_stacks_empty(); 4870 verify_overflow_empty(); 4871 sample_eden(); 4872 stop_point = 4873 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 4874 } 4875 if (stop_point != NULL) { 4876 // The careful iteration stopped early either because it found an 4877 // uninitialized object, or because we were in the midst of an 4878 // "abortable preclean", which should now be aborted. Redirty 4879 // the bits corresponding to the partially-scanned or unscanned 4880 // cards. We'll either restart at the next block boundary or 4881 // abort the preclean. 4882 assert((_collectorState == AbortablePreclean && should_abort_preclean()), 4883 "Should only be AbortablePreclean."); 4884 _modUnionTable.mark_range(MemRegion(stop_point, dirtyRegion.end())); 4885 if (should_abort_preclean()) { 4886 break; // out of preclean loop 4887 } else { 4888 // Compute the next address at which preclean should pick up; 4889 // might need bitMapLock in order to read P-bits. 4890 lastAddr = next_card_start_after_block(stop_point); 4891 } 4892 } 4893 } else { 4894 assert(lastAddr == endAddr, "consistency check"); 4895 assert(numDirtyCards == 0, "consistency check"); 4896 break; 4897 } 4898 } 4899 verify_work_stacks_empty(); 4900 verify_overflow_empty(); 4901 return cumNumDirtyCards; 4902 } 4903 4904 // NOTE: preclean_mod_union_table() above and preclean_card_table() 4905 // below are largely identical; if you need to modify 4906 // one of these methods, please check the other method too. 4907 4908 size_t CMSCollector::preclean_card_table(ConcurrentMarkSweepGeneration* gen, 4909 ScanMarkedObjectsAgainCarefullyClosure* cl) { 4910 // strategy: it's similar to precleamModUnionTable above, in that 4911 // we accumulate contiguous ranges of dirty cards, mark these cards 4912 // precleaned, then scan the region covered by these cards. 4913 HeapWord* endAddr = (HeapWord*)(gen->_virtual_space.high()); 4914 HeapWord* startAddr = (HeapWord*)(gen->_virtual_space.low()); 4915 4916 cl->setFreelistLock(gen->freelistLock()); // needed for yielding 4917 4918 size_t numDirtyCards, cumNumDirtyCards; 4919 HeapWord *lastAddr, *nextAddr; 4920 4921 for (cumNumDirtyCards = numDirtyCards = 0, 4922 nextAddr = lastAddr = startAddr; 4923 nextAddr < endAddr; 4924 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 4925 4926 ResourceMark rm; 4927 HandleMark hm; 4928 4929 MemRegion dirtyRegion; 4930 { 4931 // See comments in "Precleaning notes" above on why we 4932 // do this locking. XXX Could the locking overheads be 4933 // too high when dirty cards are sparse? [I don't think so.] 4934 stopTimer(); 4935 CMSTokenSync x(true); // is cms thread 4936 startTimer(); 4937 sample_eden(); 4938 // Get and clear dirty region from card table 4939 dirtyRegion = _ct->ct_bs()->dirty_card_range_after_reset( 4940 MemRegion(nextAddr, endAddr), 4941 true, 4942 CardTableModRefBS::precleaned_card_val()); 4943 4944 assert(dirtyRegion.start() >= nextAddr, 4945 "returned region inconsistent?"); 4946 } 4947 lastAddr = dirtyRegion.end(); 4948 numDirtyCards = 4949 dirtyRegion.word_size()/CardTableModRefBS::card_size_in_words; 4950 4951 if (!dirtyRegion.is_empty()) { 4952 stopTimer(); 4953 CMSTokenSyncWithLocks ts(true, gen->freelistLock(), bitMapLock()); 4954 startTimer(); 4955 sample_eden(); 4956 verify_work_stacks_empty(); 4957 verify_overflow_empty(); 4958 HeapWord* stop_point = 4959 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 4960 if (stop_point != NULL) { 4961 assert((_collectorState == AbortablePreclean && should_abort_preclean()), 4962 "Should only be AbortablePreclean."); 4963 _ct->ct_bs()->invalidate(MemRegion(stop_point, dirtyRegion.end())); 4964 if (should_abort_preclean()) { 4965 break; // out of preclean loop 4966 } else { 4967 // Compute the next address at which preclean should pick up. 4968 lastAddr = next_card_start_after_block(stop_point); 4969 } 4970 } 4971 } else { 4972 break; 4973 } 4974 } 4975 verify_work_stacks_empty(); 4976 verify_overflow_empty(); 4977 return cumNumDirtyCards; 4978 } 4979 4980 class PrecleanKlassClosure : public KlassClosure { 4981 CMKlassClosure _cm_klass_closure; 4982 public: 4983 PrecleanKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {} 4984 void do_klass(Klass* k) { 4985 if (k->has_accumulated_modified_oops()) { 4986 k->clear_accumulated_modified_oops(); 4987 4988 _cm_klass_closure.do_klass(k); 4989 } 4990 } 4991 }; 4992 4993 // The freelist lock is needed to prevent asserts, is it really needed? 4994 void CMSCollector::preclean_klasses(MarkRefsIntoAndScanClosure* cl, Mutex* freelistLock) { 4995 4996 cl->set_freelistLock(freelistLock); 4997 4998 CMSTokenSyncWithLocks ts(true, freelistLock, bitMapLock()); 4999 5000 // SSS: Add equivalent to ScanMarkedObjectsAgainCarefullyClosure::do_yield_check and should_abort_preclean? 5001 // SSS: We should probably check if precleaning should be aborted, at suitable intervals? 5002 PrecleanKlassClosure preclean_klass_closure(cl); 5003 ClassLoaderDataGraph::classes_do(&preclean_klass_closure); 5004 5005 verify_work_stacks_empty(); 5006 verify_overflow_empty(); 5007 } 5008 5009 void CMSCollector::checkpointRootsFinal(bool asynch, 5010 bool clear_all_soft_refs, bool init_mark_was_synchronous) { 5011 assert(_collectorState == FinalMarking, "incorrect state transition?"); 5012 check_correct_thread_executing(); 5013 // world is stopped at this checkpoint 5014 assert(SafepointSynchronize::is_at_safepoint(), 5015 "world should be stopped"); 5016 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 5017 5018 verify_work_stacks_empty(); 5019 verify_overflow_empty(); 5020 5021 SpecializationStats::clear(); 5022 if (PrintGCDetails) { 5023 gclog_or_tty->print("[YG occupancy: "SIZE_FORMAT" K ("SIZE_FORMAT" K)]", 5024 _young_gen->used() / K, 5025 _young_gen->capacity() / K); 5026 } 5027 if (asynch) { 5028 if (CMSScavengeBeforeRemark) { 5029 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5030 // Temporarily set flag to false, GCH->do_collection will 5031 // expect it to be false and set to true 5032 FlagSetting fl(gch->_is_gc_active, false); 5033 NOT_PRODUCT(GCTraceTime t("Scavenge-Before-Remark", 5034 PrintGCDetails && Verbose, true, _gc_timer_cm);) 5035 int level = _cmsGen->level() - 1; 5036 if (level >= 0) { 5037 gch->do_collection(true, // full (i.e. force, see below) 5038 false, // !clear_all_soft_refs 5039 0, // size 5040 false, // is_tlab 5041 level // max_level 5042 ); 5043 } 5044 } 5045 FreelistLocker x(this); 5046 MutexLockerEx y(bitMapLock(), 5047 Mutex::_no_safepoint_check_flag); 5048 assert(!init_mark_was_synchronous, "but that's impossible!"); 5049 checkpointRootsFinalWork(asynch, clear_all_soft_refs, false); 5050 } else { 5051 // already have all the locks 5052 checkpointRootsFinalWork(asynch, clear_all_soft_refs, 5053 init_mark_was_synchronous); 5054 } 5055 verify_work_stacks_empty(); 5056 verify_overflow_empty(); 5057 SpecializationStats::print(); 5058 } 5059 5060 void CMSCollector::checkpointRootsFinalWork(bool asynch, 5061 bool clear_all_soft_refs, bool init_mark_was_synchronous) { 5062 5063 NOT_PRODUCT(GCTraceTime tr("checkpointRootsFinalWork", PrintGCDetails, false, _gc_timer_cm);) 5064 5065 assert(haveFreelistLocks(), "must have free list locks"); 5066 assert_lock_strong(bitMapLock()); 5067 5068 if (UseAdaptiveSizePolicy) { 5069 size_policy()->checkpoint_roots_final_begin(); 5070 } 5071 5072 ResourceMark rm; 5073 HandleMark hm; 5074 5075 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5076 5077 if (should_unload_classes()) { 5078 CodeCache::gc_prologue(); 5079 } 5080 assert(haveFreelistLocks(), "must have free list locks"); 5081 assert_lock_strong(bitMapLock()); 5082 5083 if (!init_mark_was_synchronous) { 5084 // We might assume that we need not fill TLAB's when 5085 // CMSScavengeBeforeRemark is set, because we may have just done 5086 // a scavenge which would have filled all TLAB's -- and besides 5087 // Eden would be empty. This however may not always be the case -- 5088 // for instance although we asked for a scavenge, it may not have 5089 // happened because of a JNI critical section. We probably need 5090 // a policy for deciding whether we can in that case wait until 5091 // the critical section releases and then do the remark following 5092 // the scavenge, and skip it here. In the absence of that policy, 5093 // or of an indication of whether the scavenge did indeed occur, 5094 // we cannot rely on TLAB's having been filled and must do 5095 // so here just in case a scavenge did not happen. 5096 gch->ensure_parsability(false); // fill TLAB's, but no need to retire them 5097 // Update the saved marks which may affect the root scans. 5098 gch->save_marks(); 5099 5100 if (CMSPrintEdenSurvivorChunks) { 5101 print_eden_and_survivor_chunk_arrays(); 5102 } 5103 5104 { 5105 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;) 5106 5107 // Note on the role of the mod union table: 5108 // Since the marker in "markFromRoots" marks concurrently with 5109 // mutators, it is possible for some reachable objects not to have been 5110 // scanned. For instance, an only reference to an object A was 5111 // placed in object B after the marker scanned B. Unless B is rescanned, 5112 // A would be collected. Such updates to references in marked objects 5113 // are detected via the mod union table which is the set of all cards 5114 // dirtied since the first checkpoint in this GC cycle and prior to 5115 // the most recent young generation GC, minus those cleaned up by the 5116 // concurrent precleaning. 5117 if (CMSParallelRemarkEnabled && CollectedHeap::use_parallel_gc_threads()) { 5118 GCTraceTime t("Rescan (parallel) ", PrintGCDetails, false, _gc_timer_cm); 5119 do_remark_parallel(); 5120 } else { 5121 GCTraceTime t("Rescan (non-parallel) ", PrintGCDetails, false, 5122 _gc_timer_cm); 5123 do_remark_non_parallel(); 5124 } 5125 } 5126 } else { 5127 assert(!asynch, "Can't have init_mark_was_synchronous in asynch mode"); 5128 // The initial mark was stop-world, so there's no rescanning to 5129 // do; go straight on to the next step below. 5130 } 5131 verify_work_stacks_empty(); 5132 verify_overflow_empty(); 5133 5134 { 5135 NOT_PRODUCT(GCTraceTime ts("refProcessingWork", PrintGCDetails, false, _gc_timer_cm);) 5136 refProcessingWork(asynch, clear_all_soft_refs); 5137 } 5138 verify_work_stacks_empty(); 5139 verify_overflow_empty(); 5140 5141 if (should_unload_classes()) { 5142 CodeCache::gc_epilogue(); 5143 } 5144 JvmtiExport::gc_epilogue(); 5145 5146 // If we encountered any (marking stack / work queue) overflow 5147 // events during the current CMS cycle, take appropriate 5148 // remedial measures, where possible, so as to try and avoid 5149 // recurrence of that condition. 5150 assert(_markStack.isEmpty(), "No grey objects"); 5151 size_t ser_ovflw = _ser_pmc_remark_ovflw + _ser_pmc_preclean_ovflw + 5152 _ser_kac_ovflw + _ser_kac_preclean_ovflw; 5153 if (ser_ovflw > 0) { 5154 if (PrintCMSStatistics != 0) { 5155 gclog_or_tty->print_cr("Marking stack overflow (benign) " 5156 "(pmc_pc="SIZE_FORMAT", pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT 5157 ", kac_preclean="SIZE_FORMAT")", 5158 _ser_pmc_preclean_ovflw, _ser_pmc_remark_ovflw, 5159 _ser_kac_ovflw, _ser_kac_preclean_ovflw); 5160 } 5161 _markStack.expand(); 5162 _ser_pmc_remark_ovflw = 0; 5163 _ser_pmc_preclean_ovflw = 0; 5164 _ser_kac_preclean_ovflw = 0; 5165 _ser_kac_ovflw = 0; 5166 } 5167 if (_par_pmc_remark_ovflw > 0 || _par_kac_ovflw > 0) { 5168 if (PrintCMSStatistics != 0) { 5169 gclog_or_tty->print_cr("Work queue overflow (benign) " 5170 "(pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT")", 5171 _par_pmc_remark_ovflw, _par_kac_ovflw); 5172 } 5173 _par_pmc_remark_ovflw = 0; 5174 _par_kac_ovflw = 0; 5175 } 5176 if (PrintCMSStatistics != 0) { 5177 if (_markStack._hit_limit > 0) { 5178 gclog_or_tty->print_cr(" (benign) Hit max stack size limit ("SIZE_FORMAT")", 5179 _markStack._hit_limit); 5180 } 5181 if (_markStack._failed_double > 0) { 5182 gclog_or_tty->print_cr(" (benign) Failed stack doubling ("SIZE_FORMAT")," 5183 " current capacity "SIZE_FORMAT, 5184 _markStack._failed_double, 5185 _markStack.capacity()); 5186 } 5187 } 5188 _markStack._hit_limit = 0; 5189 _markStack._failed_double = 0; 5190 5191 if ((VerifyAfterGC || VerifyDuringGC) && 5192 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 5193 verify_after_remark(); 5194 } 5195 5196 _gc_tracer_cm->report_object_count_after_gc(&_is_alive_closure); 5197 5198 // Change under the freelistLocks. 5199 _collectorState = Sweeping; 5200 // Call isAllClear() under bitMapLock 5201 assert(_modUnionTable.isAllClear(), 5202 "Should be clear by end of the final marking"); 5203 assert(_ct->klass_rem_set()->mod_union_is_clear(), 5204 "Should be clear by end of the final marking"); 5205 if (UseAdaptiveSizePolicy) { 5206 size_policy()->checkpoint_roots_final_end(gch->gc_cause()); 5207 } 5208 } 5209 5210 void CMSParInitialMarkTask::work(uint worker_id) { 5211 elapsedTimer _timer; 5212 ResourceMark rm; 5213 HandleMark hm; 5214 5215 // ---------- scan from roots -------------- 5216 _timer.start(); 5217 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5218 Par_MarkRefsIntoClosure par_mri_cl(_collector->_span, &(_collector->_markBitMap)); 5219 CMKlassClosure klass_closure(&par_mri_cl); 5220 5221 // ---------- young gen roots -------------- 5222 { 5223 work_on_young_gen_roots(worker_id, &par_mri_cl); 5224 _timer.stop(); 5225 if (PrintCMSStatistics != 0) { 5226 gclog_or_tty->print_cr( 5227 "Finished young gen initial mark scan work in %dth thread: %3.3f sec", 5228 worker_id, _timer.seconds()); 5229 } 5230 } 5231 5232 // ---------- remaining roots -------------- 5233 _timer.reset(); 5234 _timer.start(); 5235 gch->gen_process_strong_roots(_collector->_cmsGen->level(), 5236 false, // yg was scanned above 5237 false, // this is parallel code 5238 SharedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()), 5239 &par_mri_cl, 5240 true, // walk all of code cache if (so & SO_AllCodeCache) 5241 NULL, 5242 &klass_closure); 5243 assert(_collector->should_unload_classes() 5244 || (_collector->CMSCollector::roots_scanning_options() & SharedHeap::SO_AllCodeCache), 5245 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 5246 _timer.stop(); 5247 if (PrintCMSStatistics != 0) { 5248 gclog_or_tty->print_cr( 5249 "Finished remaining root initial mark scan work in %dth thread: %3.3f sec", 5250 worker_id, _timer.seconds()); 5251 } 5252 } 5253 5254 // Parallel remark task 5255 class CMSParRemarkTask: public CMSParMarkTask { 5256 CompactibleFreeListSpace* _cms_space; 5257 5258 // The per-thread work queues, available here for stealing. 5259 OopTaskQueueSet* _task_queues; 5260 ParallelTaskTerminator _term; 5261 5262 public: 5263 // A value of 0 passed to n_workers will cause the number of 5264 // workers to be taken from the active workers in the work gang. 5265 CMSParRemarkTask(CMSCollector* collector, 5266 CompactibleFreeListSpace* cms_space, 5267 int n_workers, FlexibleWorkGang* workers, 5268 OopTaskQueueSet* task_queues): 5269 CMSParMarkTask("Rescan roots and grey objects in parallel", 5270 collector, n_workers), 5271 _cms_space(cms_space), 5272 _task_queues(task_queues), 5273 _term(n_workers, task_queues) { } 5274 5275 OopTaskQueueSet* task_queues() { return _task_queues; } 5276 5277 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 5278 5279 ParallelTaskTerminator* terminator() { return &_term; } 5280 int n_workers() { return _n_workers; } 5281 5282 void work(uint worker_id); 5283 5284 private: 5285 // ... of dirty cards in old space 5286 void do_dirty_card_rescan_tasks(CompactibleFreeListSpace* sp, int i, 5287 Par_MarkRefsIntoAndScanClosure* cl); 5288 5289 // ... work stealing for the above 5290 void do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, int* seed); 5291 }; 5292 5293 class RemarkKlassClosure : public KlassClosure { 5294 CMKlassClosure _cm_klass_closure; 5295 public: 5296 RemarkKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {} 5297 void do_klass(Klass* k) { 5298 // Check if we have modified any oops in the Klass during the concurrent marking. 5299 if (k->has_accumulated_modified_oops()) { 5300 k->clear_accumulated_modified_oops(); 5301 5302 // We could have transfered the current modified marks to the accumulated marks, 5303 // like we do with the Card Table to Mod Union Table. But it's not really necessary. 5304 } else if (k->has_modified_oops()) { 5305 // Don't clear anything, this info is needed by the next young collection. 5306 } else { 5307 // No modified oops in the Klass. 5308 return; 5309 } 5310 5311 // The klass has modified fields, need to scan the klass. 5312 _cm_klass_closure.do_klass(k); 5313 } 5314 }; 5315 5316 void CMSParMarkTask::work_on_young_gen_roots(uint worker_id, OopsInGenClosure* cl) { 5317 DefNewGeneration* dng = _collector->_young_gen->as_DefNewGeneration(); 5318 EdenSpace* eden_space = dng->eden(); 5319 ContiguousSpace* from_space = dng->from(); 5320 ContiguousSpace* to_space = dng->to(); 5321 5322 HeapWord** eca = _collector->_eden_chunk_array; 5323 size_t ect = _collector->_eden_chunk_index; 5324 HeapWord** sca = _collector->_survivor_chunk_array; 5325 size_t sct = _collector->_survivor_chunk_index; 5326 5327 assert(ect <= _collector->_eden_chunk_capacity, "out of bounds"); 5328 assert(sct <= _collector->_survivor_chunk_capacity, "out of bounds"); 5329 5330 do_young_space_rescan(worker_id, cl, to_space, NULL, 0); 5331 do_young_space_rescan(worker_id, cl, from_space, sca, sct); 5332 do_young_space_rescan(worker_id, cl, eden_space, eca, ect); 5333 } 5334 5335 // work_queue(i) is passed to the closure 5336 // Par_MarkRefsIntoAndScanClosure. The "i" parameter 5337 // also is passed to do_dirty_card_rescan_tasks() and to 5338 // do_work_steal() to select the i-th task_queue. 5339 5340 void CMSParRemarkTask::work(uint worker_id) { 5341 elapsedTimer _timer; 5342 ResourceMark rm; 5343 HandleMark hm; 5344 5345 // ---------- rescan from roots -------------- 5346 _timer.start(); 5347 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5348 Par_MarkRefsIntoAndScanClosure par_mrias_cl(_collector, 5349 _collector->_span, _collector->ref_processor(), 5350 &(_collector->_markBitMap), 5351 work_queue(worker_id)); 5352 5353 // Rescan young gen roots first since these are likely 5354 // coarsely partitioned and may, on that account, constitute 5355 // the critical path; thus, it's best to start off that 5356 // work first. 5357 // ---------- young gen roots -------------- 5358 { 5359 work_on_young_gen_roots(worker_id, &par_mrias_cl); 5360 _timer.stop(); 5361 if (PrintCMSStatistics != 0) { 5362 gclog_or_tty->print_cr( 5363 "Finished young gen rescan work in %dth thread: %3.3f sec", 5364 worker_id, _timer.seconds()); 5365 } 5366 } 5367 5368 // ---------- remaining roots -------------- 5369 _timer.reset(); 5370 _timer.start(); 5371 gch->gen_process_strong_roots(_collector->_cmsGen->level(), 5372 false, // yg was scanned above 5373 false, // this is parallel code 5374 SharedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()), 5375 &par_mrias_cl, 5376 true, // walk all of code cache if (so & SO_AllCodeCache) 5377 NULL, 5378 NULL); // The dirty klasses will be handled below 5379 assert(_collector->should_unload_classes() 5380 || (_collector->CMSCollector::roots_scanning_options() & SharedHeap::SO_AllCodeCache), 5381 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 5382 _timer.stop(); 5383 if (PrintCMSStatistics != 0) { 5384 gclog_or_tty->print_cr( 5385 "Finished remaining root rescan work in %dth thread: %3.3f sec", 5386 worker_id, _timer.seconds()); 5387 } 5388 5389 // ---------- unhandled CLD scanning ---------- 5390 if (worker_id == 0) { // Single threaded at the moment. 5391 _timer.reset(); 5392 _timer.start(); 5393 5394 // Scan all new class loader data objects and new dependencies that were 5395 // introduced during concurrent marking. 5396 ResourceMark rm; 5397 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds(); 5398 for (int i = 0; i < array->length(); i++) { 5399 par_mrias_cl.do_class_loader_data(array->at(i)); 5400 } 5401 5402 // We don't need to keep track of new CLDs anymore. 5403 ClassLoaderDataGraph::remember_new_clds(false); 5404 5405 _timer.stop(); 5406 if (PrintCMSStatistics != 0) { 5407 gclog_or_tty->print_cr( 5408 "Finished unhandled CLD scanning work in %dth thread: %3.3f sec", 5409 worker_id, _timer.seconds()); 5410 } 5411 } 5412 5413 // ---------- dirty klass scanning ---------- 5414 if (worker_id == 0) { // Single threaded at the moment. 5415 _timer.reset(); 5416 _timer.start(); 5417 5418 // Scan all classes that was dirtied during the concurrent marking phase. 5419 RemarkKlassClosure remark_klass_closure(&par_mrias_cl); 5420 ClassLoaderDataGraph::classes_do(&remark_klass_closure); 5421 5422 _timer.stop(); 5423 if (PrintCMSStatistics != 0) { 5424 gclog_or_tty->print_cr( 5425 "Finished dirty klass scanning work in %dth thread: %3.3f sec", 5426 worker_id, _timer.seconds()); 5427 } 5428 } 5429 5430 // We might have added oops to ClassLoaderData::_handles during the 5431 // concurrent marking phase. These oops point to newly allocated objects 5432 // that are guaranteed to be kept alive. Either by the direct allocation 5433 // code, or when the young collector processes the strong roots. Hence, 5434 // we don't have to revisit the _handles block during the remark phase. 5435 5436 // ---------- rescan dirty cards ------------ 5437 _timer.reset(); 5438 _timer.start(); 5439 5440 // Do the rescan tasks for each of the two spaces 5441 // (cms_space) in turn. 5442 // "worker_id" is passed to select the task_queue for "worker_id" 5443 do_dirty_card_rescan_tasks(_cms_space, worker_id, &par_mrias_cl); 5444 _timer.stop(); 5445 if (PrintCMSStatistics != 0) { 5446 gclog_or_tty->print_cr( 5447 "Finished dirty card rescan work in %dth thread: %3.3f sec", 5448 worker_id, _timer.seconds()); 5449 } 5450 5451 // ---------- steal work from other threads ... 5452 // ---------- ... and drain overflow list. 5453 _timer.reset(); 5454 _timer.start(); 5455 do_work_steal(worker_id, &par_mrias_cl, _collector->hash_seed(worker_id)); 5456 _timer.stop(); 5457 if (PrintCMSStatistics != 0) { 5458 gclog_or_tty->print_cr( 5459 "Finished work stealing in %dth thread: %3.3f sec", 5460 worker_id, _timer.seconds()); 5461 } 5462 } 5463 5464 // Note that parameter "i" is not used. 5465 void 5466 CMSParMarkTask::do_young_space_rescan(uint worker_id, 5467 OopsInGenClosure* cl, ContiguousSpace* space, 5468 HeapWord** chunk_array, size_t chunk_top) { 5469 // Until all tasks completed: 5470 // . claim an unclaimed task 5471 // . compute region boundaries corresponding to task claimed 5472 // using chunk_array 5473 // . par_oop_iterate(cl) over that region 5474 5475 ResourceMark rm; 5476 HandleMark hm; 5477 5478 SequentialSubTasksDone* pst = space->par_seq_tasks(); 5479 5480 uint nth_task = 0; 5481 uint n_tasks = pst->n_tasks(); 5482 5483 if (n_tasks > 0) { 5484 assert(pst->valid(), "Uninitialized use?"); 5485 HeapWord *start, *end; 5486 while (!pst->is_task_claimed(/* reference */ nth_task)) { 5487 // We claimed task # nth_task; compute its boundaries. 5488 if (chunk_top == 0) { // no samples were taken 5489 assert(nth_task == 0 && n_tasks == 1, "Can have only 1 EdenSpace task"); 5490 start = space->bottom(); 5491 end = space->top(); 5492 } else if (nth_task == 0) { 5493 start = space->bottom(); 5494 end = chunk_array[nth_task]; 5495 } else if (nth_task < (uint)chunk_top) { 5496 assert(nth_task >= 1, "Control point invariant"); 5497 start = chunk_array[nth_task - 1]; 5498 end = chunk_array[nth_task]; 5499 } else { 5500 assert(nth_task == (uint)chunk_top, "Control point invariant"); 5501 start = chunk_array[chunk_top - 1]; 5502 end = space->top(); 5503 } 5504 MemRegion mr(start, end); 5505 // Verify that mr is in space 5506 assert(mr.is_empty() || space->used_region().contains(mr), 5507 "Should be in space"); 5508 // Verify that "start" is an object boundary 5509 assert(mr.is_empty() || oop(mr.start())->is_oop(), 5510 "Should be an oop"); 5511 space->par_oop_iterate(mr, cl); 5512 } 5513 pst->all_tasks_completed(); 5514 } 5515 } 5516 5517 void 5518 CMSParRemarkTask::do_dirty_card_rescan_tasks( 5519 CompactibleFreeListSpace* sp, int i, 5520 Par_MarkRefsIntoAndScanClosure* cl) { 5521 // Until all tasks completed: 5522 // . claim an unclaimed task 5523 // . compute region boundaries corresponding to task claimed 5524 // . transfer dirty bits ct->mut for that region 5525 // . apply rescanclosure to dirty mut bits for that region 5526 5527 ResourceMark rm; 5528 HandleMark hm; 5529 5530 OopTaskQueue* work_q = work_queue(i); 5531 ModUnionClosure modUnionClosure(&(_collector->_modUnionTable)); 5532 // CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! 5533 // CAUTION: This closure has state that persists across calls to 5534 // the work method dirty_range_iterate_clear() in that it has 5535 // embedded in it a (subtype of) UpwardsObjectClosure. The 5536 // use of that state in the embedded UpwardsObjectClosure instance 5537 // assumes that the cards are always iterated (even if in parallel 5538 // by several threads) in monotonically increasing order per each 5539 // thread. This is true of the implementation below which picks 5540 // card ranges (chunks) in monotonically increasing order globally 5541 // and, a-fortiori, in monotonically increasing order per thread 5542 // (the latter order being a subsequence of the former). 5543 // If the work code below is ever reorganized into a more chaotic 5544 // work-partitioning form than the current "sequential tasks" 5545 // paradigm, the use of that persistent state will have to be 5546 // revisited and modified appropriately. See also related 5547 // bug 4756801 work on which should examine this code to make 5548 // sure that the changes there do not run counter to the 5549 // assumptions made here and necessary for correctness and 5550 // efficiency. Note also that this code might yield inefficient 5551 // behavior in the case of very large objects that span one or 5552 // more work chunks. Such objects would potentially be scanned 5553 // several times redundantly. Work on 4756801 should try and 5554 // address that performance anomaly if at all possible. XXX 5555 MemRegion full_span = _collector->_span; 5556 CMSBitMap* bm = &(_collector->_markBitMap); // shared 5557 MarkFromDirtyCardsClosure 5558 greyRescanClosure(_collector, full_span, // entire span of interest 5559 sp, bm, work_q, cl); 5560 5561 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 5562 assert(pst->valid(), "Uninitialized use?"); 5563 uint nth_task = 0; 5564 const int alignment = CardTableModRefBS::card_size * BitsPerWord; 5565 MemRegion span = sp->used_region(); 5566 HeapWord* start_addr = span.start(); 5567 HeapWord* end_addr = (HeapWord*)round_to((intptr_t)span.end(), 5568 alignment); 5569 const size_t chunk_size = sp->rescan_task_size(); // in HeapWord units 5570 assert((HeapWord*)round_to((intptr_t)start_addr, alignment) == 5571 start_addr, "Check alignment"); 5572 assert((size_t)round_to((intptr_t)chunk_size, alignment) == 5573 chunk_size, "Check alignment"); 5574 5575 while (!pst->is_task_claimed(/* reference */ nth_task)) { 5576 // Having claimed the nth_task, compute corresponding mem-region, 5577 // which is a-fortiori aligned correctly (i.e. at a MUT boundary). 5578 // The alignment restriction ensures that we do not need any 5579 // synchronization with other gang-workers while setting or 5580 // clearing bits in thus chunk of the MUT. 5581 MemRegion this_span = MemRegion(start_addr + nth_task*chunk_size, 5582 start_addr + (nth_task+1)*chunk_size); 5583 // The last chunk's end might be way beyond end of the 5584 // used region. In that case pull back appropriately. 5585 if (this_span.end() > end_addr) { 5586 this_span.set_end(end_addr); 5587 assert(!this_span.is_empty(), "Program logic (calculation of n_tasks)"); 5588 } 5589 // Iterate over the dirty cards covering this chunk, marking them 5590 // precleaned, and setting the corresponding bits in the mod union 5591 // table. Since we have been careful to partition at Card and MUT-word 5592 // boundaries no synchronization is needed between parallel threads. 5593 _collector->_ct->ct_bs()->dirty_card_iterate(this_span, 5594 &modUnionClosure); 5595 5596 // Having transferred these marks into the modUnionTable, 5597 // rescan the marked objects on the dirty cards in the modUnionTable. 5598 // Even if this is at a synchronous collection, the initial marking 5599 // may have been done during an asynchronous collection so there 5600 // may be dirty bits in the mod-union table. 5601 _collector->_modUnionTable.dirty_range_iterate_clear( 5602 this_span, &greyRescanClosure); 5603 _collector->_modUnionTable.verifyNoOneBitsInRange( 5604 this_span.start(), 5605 this_span.end()); 5606 } 5607 pst->all_tasks_completed(); // declare that i am done 5608 } 5609 5610 // . see if we can share work_queues with ParNew? XXX 5611 void 5612 CMSParRemarkTask::do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, 5613 int* seed) { 5614 OopTaskQueue* work_q = work_queue(i); 5615 NOT_PRODUCT(int num_steals = 0;) 5616 oop obj_to_scan; 5617 CMSBitMap* bm = &(_collector->_markBitMap); 5618 5619 while (true) { 5620 // Completely finish any left over work from (an) earlier round(s) 5621 cl->trim_queue(0); 5622 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 5623 (size_t)ParGCDesiredObjsFromOverflowList); 5624 // Now check if there's any work in the overflow list 5625 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads, 5626 // only affects the number of attempts made to get work from the 5627 // overflow list and does not affect the number of workers. Just 5628 // pass ParallelGCThreads so this behavior is unchanged. 5629 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 5630 work_q, 5631 ParallelGCThreads)) { 5632 // found something in global overflow list; 5633 // not yet ready to go stealing work from others. 5634 // We'd like to assert(work_q->size() != 0, ...) 5635 // because we just took work from the overflow list, 5636 // but of course we can't since all of that could have 5637 // been already stolen from us. 5638 // "He giveth and He taketh away." 5639 continue; 5640 } 5641 // Verify that we have no work before we resort to stealing 5642 assert(work_q->size() == 0, "Have work, shouldn't steal"); 5643 // Try to steal from other queues that have work 5644 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 5645 NOT_PRODUCT(num_steals++;) 5646 assert(obj_to_scan->is_oop(), "Oops, not an oop!"); 5647 assert(bm->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 5648 // Do scanning work 5649 obj_to_scan->oop_iterate(cl); 5650 // Loop around, finish this work, and try to steal some more 5651 } else if (terminator()->offer_termination()) { 5652 break; // nirvana from the infinite cycle 5653 } 5654 } 5655 NOT_PRODUCT( 5656 if (PrintCMSStatistics != 0) { 5657 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals); 5658 } 5659 ) 5660 assert(work_q->size() == 0 && _collector->overflow_list_is_empty(), 5661 "Else our work is not yet done"); 5662 } 5663 5664 // Record object boundaries in _eden_chunk_array by sampling the eden 5665 // top in the slow-path eden object allocation code path and record 5666 // the boundaries, if CMSEdenChunksRecordAlways is true. If 5667 // CMSEdenChunksRecordAlways is false, we use the other asynchronous 5668 // sampling in sample_eden() that activates during the part of the 5669 // preclean phase. 5670 void CMSCollector::sample_eden_chunk() { 5671 if (CMSEdenChunksRecordAlways && _eden_chunk_array != NULL) { 5672 if (_eden_chunk_lock->try_lock()) { 5673 // Record a sample. This is the critical section. The contents 5674 // of the _eden_chunk_array have to be non-decreasing in the 5675 // address order. 5676 _eden_chunk_array[_eden_chunk_index] = *_top_addr; 5677 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr, 5678 "Unexpected state of Eden"); 5679 if (_eden_chunk_index == 0 || 5680 ((_eden_chunk_array[_eden_chunk_index] > _eden_chunk_array[_eden_chunk_index-1]) && 5681 (pointer_delta(_eden_chunk_array[_eden_chunk_index], 5682 _eden_chunk_array[_eden_chunk_index-1]) >= CMSSamplingGrain))) { 5683 _eden_chunk_index++; // commit sample 5684 } 5685 _eden_chunk_lock->unlock(); 5686 } 5687 } 5688 } 5689 5690 // Return a thread-local PLAB recording array, as appropriate. 5691 void* CMSCollector::get_data_recorder(int thr_num) { 5692 if (_survivor_plab_array != NULL && 5693 (CMSPLABRecordAlways || 5694 (_collectorState > Marking && _collectorState < FinalMarking))) { 5695 assert(thr_num < (int)ParallelGCThreads, "thr_num is out of bounds"); 5696 ChunkArray* ca = &_survivor_plab_array[thr_num]; 5697 ca->reset(); // clear it so that fresh data is recorded 5698 return (void*) ca; 5699 } else { 5700 return NULL; 5701 } 5702 } 5703 5704 // Reset all the thread-local PLAB recording arrays 5705 void CMSCollector::reset_survivor_plab_arrays() { 5706 for (uint i = 0; i < ParallelGCThreads; i++) { 5707 _survivor_plab_array[i].reset(); 5708 } 5709 } 5710 5711 // Merge the per-thread plab arrays into the global survivor chunk 5712 // array which will provide the partitioning of the survivor space 5713 // for CMS initial scan and rescan. 5714 void CMSCollector::merge_survivor_plab_arrays(ContiguousSpace* surv, 5715 int no_of_gc_threads) { 5716 assert(_survivor_plab_array != NULL, "Error"); 5717 assert(_survivor_chunk_array != NULL, "Error"); 5718 assert(_collectorState == FinalMarking || 5719 (CMSParallelInitialMarkEnabled && _collectorState == InitialMarking), "Error"); 5720 for (int j = 0; j < no_of_gc_threads; j++) { 5721 _cursor[j] = 0; 5722 } 5723 HeapWord* top = surv->top(); 5724 size_t i; 5725 for (i = 0; i < _survivor_chunk_capacity; i++) { // all sca entries 5726 HeapWord* min_val = top; // Higher than any PLAB address 5727 uint min_tid = 0; // position of min_val this round 5728 for (int j = 0; j < no_of_gc_threads; j++) { 5729 ChunkArray* cur_sca = &_survivor_plab_array[j]; 5730 if (_cursor[j] == cur_sca->end()) { 5731 continue; 5732 } 5733 assert(_cursor[j] < cur_sca->end(), "ctl pt invariant"); 5734 HeapWord* cur_val = cur_sca->nth(_cursor[j]); 5735 assert(surv->used_region().contains(cur_val), "Out of bounds value"); 5736 if (cur_val < min_val) { 5737 min_tid = j; 5738 min_val = cur_val; 5739 } else { 5740 assert(cur_val < top, "All recorded addresses should be less"); 5741 } 5742 } 5743 // At this point min_val and min_tid are respectively 5744 // the least address in _survivor_plab_array[j]->nth(_cursor[j]) 5745 // and the thread (j) that witnesses that address. 5746 // We record this address in the _survivor_chunk_array[i] 5747 // and increment _cursor[min_tid] prior to the next round i. 5748 if (min_val == top) { 5749 break; 5750 } 5751 _survivor_chunk_array[i] = min_val; 5752 _cursor[min_tid]++; 5753 } 5754 // We are all done; record the size of the _survivor_chunk_array 5755 _survivor_chunk_index = i; // exclusive: [0, i) 5756 if (PrintCMSStatistics > 0) { 5757 gclog_or_tty->print(" (Survivor:" SIZE_FORMAT "chunks) ", i); 5758 } 5759 // Verify that we used up all the recorded entries 5760 #ifdef ASSERT 5761 size_t total = 0; 5762 for (int j = 0; j < no_of_gc_threads; j++) { 5763 assert(_cursor[j] == _survivor_plab_array[j].end(), "Ctl pt invariant"); 5764 total += _cursor[j]; 5765 } 5766 assert(total == _survivor_chunk_index, "Ctl Pt Invariant"); 5767 // Check that the merged array is in sorted order 5768 if (total > 0) { 5769 for (size_t i = 0; i < total - 1; i++) { 5770 if (PrintCMSStatistics > 0) { 5771 gclog_or_tty->print(" (chunk" SIZE_FORMAT ":" INTPTR_FORMAT ") ", 5772 i, _survivor_chunk_array[i]); 5773 } 5774 assert(_survivor_chunk_array[i] < _survivor_chunk_array[i+1], 5775 "Not sorted"); 5776 } 5777 } 5778 #endif // ASSERT 5779 } 5780 5781 // Set up the space's par_seq_tasks structure for work claiming 5782 // for parallel initial scan and rescan of young gen. 5783 // See ParRescanTask where this is currently used. 5784 void 5785 CMSCollector:: 5786 initialize_sequential_subtasks_for_young_gen_rescan(int n_threads) { 5787 assert(n_threads > 0, "Unexpected n_threads argument"); 5788 DefNewGeneration* dng = (DefNewGeneration*)_young_gen; 5789 5790 // Eden space 5791 if (!dng->eden()->is_empty()) { 5792 SequentialSubTasksDone* pst = dng->eden()->par_seq_tasks(); 5793 assert(!pst->valid(), "Clobbering existing data?"); 5794 // Each valid entry in [0, _eden_chunk_index) represents a task. 5795 size_t n_tasks = _eden_chunk_index + 1; 5796 assert(n_tasks == 1 || _eden_chunk_array != NULL, "Error"); 5797 // Sets the condition for completion of the subtask (how many threads 5798 // need to finish in order to be done). 5799 pst->set_n_threads(n_threads); 5800 pst->set_n_tasks((int)n_tasks); 5801 } 5802 5803 // Merge the survivor plab arrays into _survivor_chunk_array 5804 if (_survivor_plab_array != NULL) { 5805 merge_survivor_plab_arrays(dng->from(), n_threads); 5806 } else { 5807 assert(_survivor_chunk_index == 0, "Error"); 5808 } 5809 5810 // To space 5811 { 5812 SequentialSubTasksDone* pst = dng->to()->par_seq_tasks(); 5813 assert(!pst->valid(), "Clobbering existing data?"); 5814 // Sets the condition for completion of the subtask (how many threads 5815 // need to finish in order to be done). 5816 pst->set_n_threads(n_threads); 5817 pst->set_n_tasks(1); 5818 assert(pst->valid(), "Error"); 5819 } 5820 5821 // From space 5822 { 5823 SequentialSubTasksDone* pst = dng->from()->par_seq_tasks(); 5824 assert(!pst->valid(), "Clobbering existing data?"); 5825 size_t n_tasks = _survivor_chunk_index + 1; 5826 assert(n_tasks == 1 || _survivor_chunk_array != NULL, "Error"); 5827 // Sets the condition for completion of the subtask (how many threads 5828 // need to finish in order to be done). 5829 pst->set_n_threads(n_threads); 5830 pst->set_n_tasks((int)n_tasks); 5831 assert(pst->valid(), "Error"); 5832 } 5833 } 5834 5835 // Parallel version of remark 5836 void CMSCollector::do_remark_parallel() { 5837 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5838 FlexibleWorkGang* workers = gch->workers(); 5839 assert(workers != NULL, "Need parallel worker threads."); 5840 // Choose to use the number of GC workers most recently set 5841 // into "active_workers". If active_workers is not set, set it 5842 // to ParallelGCThreads. 5843 int n_workers = workers->active_workers(); 5844 if (n_workers == 0) { 5845 assert(n_workers > 0, "Should have been set during scavenge"); 5846 n_workers = ParallelGCThreads; 5847 workers->set_active_workers(n_workers); 5848 } 5849 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 5850 5851 CMSParRemarkTask tsk(this, 5852 cms_space, 5853 n_workers, workers, task_queues()); 5854 5855 // Set up for parallel process_strong_roots work. 5856 gch->set_par_threads(n_workers); 5857 // We won't be iterating over the cards in the card table updating 5858 // the younger_gen cards, so we shouldn't call the following else 5859 // the verification code as well as subsequent younger_refs_iterate 5860 // code would get confused. XXX 5861 // gch->rem_set()->prepare_for_younger_refs_iterate(true); // parallel 5862 5863 // The young gen rescan work will not be done as part of 5864 // process_strong_roots (which currently doesn't knw how to 5865 // parallelize such a scan), but rather will be broken up into 5866 // a set of parallel tasks (via the sampling that the [abortable] 5867 // preclean phase did of EdenSpace, plus the [two] tasks of 5868 // scanning the [two] survivor spaces. Further fine-grain 5869 // parallelization of the scanning of the survivor spaces 5870 // themselves, and of precleaning of the younger gen itself 5871 // is deferred to the future. 5872 initialize_sequential_subtasks_for_young_gen_rescan(n_workers); 5873 5874 // The dirty card rescan work is broken up into a "sequence" 5875 // of parallel tasks (per constituent space) that are dynamically 5876 // claimed by the parallel threads. 5877 cms_space->initialize_sequential_subtasks_for_rescan(n_workers); 5878 5879 // It turns out that even when we're using 1 thread, doing the work in a 5880 // separate thread causes wide variance in run times. We can't help this 5881 // in the multi-threaded case, but we special-case n=1 here to get 5882 // repeatable measurements of the 1-thread overhead of the parallel code. 5883 if (n_workers > 1) { 5884 // Make refs discovery MT-safe, if it isn't already: it may not 5885 // necessarily be so, since it's possible that we are doing 5886 // ST marking. 5887 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), true); 5888 GenCollectedHeap::StrongRootsScope srs(gch); 5889 workers->run_task(&tsk); 5890 } else { 5891 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false); 5892 GenCollectedHeap::StrongRootsScope srs(gch); 5893 tsk.work(0); 5894 } 5895 5896 gch->set_par_threads(0); // 0 ==> non-parallel. 5897 // restore, single-threaded for now, any preserved marks 5898 // as a result of work_q overflow 5899 restore_preserved_marks_if_any(); 5900 } 5901 5902 // Non-parallel version of remark 5903 void CMSCollector::do_remark_non_parallel() { 5904 ResourceMark rm; 5905 HandleMark hm; 5906 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5907 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false); 5908 5909 MarkRefsIntoAndScanClosure 5910 mrias_cl(_span, ref_processor(), &_markBitMap, NULL /* not precleaning */, 5911 &_markStack, this, 5912 false /* should_yield */, false /* not precleaning */); 5913 MarkFromDirtyCardsClosure 5914 markFromDirtyCardsClosure(this, _span, 5915 NULL, // space is set further below 5916 &_markBitMap, &_markStack, &mrias_cl); 5917 { 5918 GCTraceTime t("grey object rescan", PrintGCDetails, false, _gc_timer_cm); 5919 // Iterate over the dirty cards, setting the corresponding bits in the 5920 // mod union table. 5921 { 5922 ModUnionClosure modUnionClosure(&_modUnionTable); 5923 _ct->ct_bs()->dirty_card_iterate( 5924 _cmsGen->used_region(), 5925 &modUnionClosure); 5926 } 5927 // Having transferred these marks into the modUnionTable, we just need 5928 // to rescan the marked objects on the dirty cards in the modUnionTable. 5929 // The initial marking may have been done during an asynchronous 5930 // collection so there may be dirty bits in the mod-union table. 5931 const int alignment = 5932 CardTableModRefBS::card_size * BitsPerWord; 5933 { 5934 // ... First handle dirty cards in CMS gen 5935 markFromDirtyCardsClosure.set_space(_cmsGen->cmsSpace()); 5936 MemRegion ur = _cmsGen->used_region(); 5937 HeapWord* lb = ur.start(); 5938 HeapWord* ub = (HeapWord*)round_to((intptr_t)ur.end(), alignment); 5939 MemRegion cms_span(lb, ub); 5940 _modUnionTable.dirty_range_iterate_clear(cms_span, 5941 &markFromDirtyCardsClosure); 5942 verify_work_stacks_empty(); 5943 if (PrintCMSStatistics != 0) { 5944 gclog_or_tty->print(" (re-scanned "SIZE_FORMAT" dirty cards in cms gen) ", 5945 markFromDirtyCardsClosure.num_dirty_cards()); 5946 } 5947 } 5948 } 5949 if (VerifyDuringGC && 5950 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 5951 HandleMark hm; // Discard invalid handles created during verification 5952 Universe::verify(); 5953 } 5954 { 5955 GCTraceTime t("root rescan", PrintGCDetails, false, _gc_timer_cm); 5956 5957 verify_work_stacks_empty(); 5958 5959 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 5960 GenCollectedHeap::StrongRootsScope srs(gch); 5961 gch->gen_process_strong_roots(_cmsGen->level(), 5962 true, // younger gens as roots 5963 false, // use the local StrongRootsScope 5964 SharedHeap::ScanningOption(roots_scanning_options()), 5965 &mrias_cl, 5966 true, // walk code active on stacks 5967 NULL, 5968 NULL); // The dirty klasses will be handled below 5969 5970 assert(should_unload_classes() 5971 || (roots_scanning_options() & SharedHeap::SO_AllCodeCache), 5972 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 5973 } 5974 5975 { 5976 GCTraceTime t("visit unhandled CLDs", PrintGCDetails, false, _gc_timer_cm); 5977 5978 verify_work_stacks_empty(); 5979 5980 // Scan all class loader data objects that might have been introduced 5981 // during concurrent marking. 5982 ResourceMark rm; 5983 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds(); 5984 for (int i = 0; i < array->length(); i++) { 5985 mrias_cl.do_class_loader_data(array->at(i)); 5986 } 5987 5988 // We don't need to keep track of new CLDs anymore. 5989 ClassLoaderDataGraph::remember_new_clds(false); 5990 5991 verify_work_stacks_empty(); 5992 } 5993 5994 { 5995 GCTraceTime t("dirty klass scan", PrintGCDetails, false, _gc_timer_cm); 5996 5997 verify_work_stacks_empty(); 5998 5999 RemarkKlassClosure remark_klass_closure(&mrias_cl); 6000 ClassLoaderDataGraph::classes_do(&remark_klass_closure); 6001 6002 verify_work_stacks_empty(); 6003 } 6004 6005 // We might have added oops to ClassLoaderData::_handles during the 6006 // concurrent marking phase. These oops point to newly allocated objects 6007 // that are guaranteed to be kept alive. Either by the direct allocation 6008 // code, or when the young collector processes the strong roots. Hence, 6009 // we don't have to revisit the _handles block during the remark phase. 6010 6011 verify_work_stacks_empty(); 6012 // Restore evacuated mark words, if any, used for overflow list links 6013 if (!CMSOverflowEarlyRestoration) { 6014 restore_preserved_marks_if_any(); 6015 } 6016 verify_overflow_empty(); 6017 } 6018 6019 //////////////////////////////////////////////////////// 6020 // Parallel Reference Processing Task Proxy Class 6021 //////////////////////////////////////////////////////// 6022 class CMSRefProcTaskProxy: public AbstractGangTaskWOopQueues { 6023 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask; 6024 CMSCollector* _collector; 6025 CMSBitMap* _mark_bit_map; 6026 const MemRegion _span; 6027 ProcessTask& _task; 6028 6029 public: 6030 CMSRefProcTaskProxy(ProcessTask& task, 6031 CMSCollector* collector, 6032 const MemRegion& span, 6033 CMSBitMap* mark_bit_map, 6034 AbstractWorkGang* workers, 6035 OopTaskQueueSet* task_queues): 6036 // XXX Should superclass AGTWOQ also know about AWG since it knows 6037 // about the task_queues used by the AWG? Then it could initialize 6038 // the terminator() object. See 6984287. The set_for_termination() 6039 // below is a temporary band-aid for the regression in 6984287. 6040 AbstractGangTaskWOopQueues("Process referents by policy in parallel", 6041 task_queues), 6042 _task(task), 6043 _collector(collector), _span(span), _mark_bit_map(mark_bit_map) 6044 { 6045 assert(_collector->_span.equals(_span) && !_span.is_empty(), 6046 "Inconsistency in _span"); 6047 set_for_termination(workers->active_workers()); 6048 } 6049 6050 OopTaskQueueSet* task_queues() { return queues(); } 6051 6052 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 6053 6054 void do_work_steal(int i, 6055 CMSParDrainMarkingStackClosure* drain, 6056 CMSParKeepAliveClosure* keep_alive, 6057 int* seed); 6058 6059 virtual void work(uint worker_id); 6060 }; 6061 6062 void CMSRefProcTaskProxy::work(uint worker_id) { 6063 assert(_collector->_span.equals(_span), "Inconsistency in _span"); 6064 CMSParKeepAliveClosure par_keep_alive(_collector, _span, 6065 _mark_bit_map, 6066 work_queue(worker_id)); 6067 CMSParDrainMarkingStackClosure par_drain_stack(_collector, _span, 6068 _mark_bit_map, 6069 work_queue(worker_id)); 6070 CMSIsAliveClosure is_alive_closure(_span, _mark_bit_map); 6071 _task.work(worker_id, is_alive_closure, par_keep_alive, par_drain_stack); 6072 if (_task.marks_oops_alive()) { 6073 do_work_steal(worker_id, &par_drain_stack, &par_keep_alive, 6074 _collector->hash_seed(worker_id)); 6075 } 6076 assert(work_queue(worker_id)->size() == 0, "work_queue should be empty"); 6077 assert(_collector->_overflow_list == NULL, "non-empty _overflow_list"); 6078 } 6079 6080 class CMSRefEnqueueTaskProxy: public AbstractGangTask { 6081 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask; 6082 EnqueueTask& _task; 6083 6084 public: 6085 CMSRefEnqueueTaskProxy(EnqueueTask& task) 6086 : AbstractGangTask("Enqueue reference objects in parallel"), 6087 _task(task) 6088 { } 6089 6090 virtual void work(uint worker_id) 6091 { 6092 _task.work(worker_id); 6093 } 6094 }; 6095 6096 CMSParKeepAliveClosure::CMSParKeepAliveClosure(CMSCollector* collector, 6097 MemRegion span, CMSBitMap* bit_map, OopTaskQueue* work_queue): 6098 _span(span), 6099 _bit_map(bit_map), 6100 _work_queue(work_queue), 6101 _mark_and_push(collector, span, bit_map, work_queue), 6102 _low_water_mark(MIN2((uint)(work_queue->max_elems()/4), 6103 (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads))) 6104 { } 6105 6106 // . see if we can share work_queues with ParNew? XXX 6107 void CMSRefProcTaskProxy::do_work_steal(int i, 6108 CMSParDrainMarkingStackClosure* drain, 6109 CMSParKeepAliveClosure* keep_alive, 6110 int* seed) { 6111 OopTaskQueue* work_q = work_queue(i); 6112 NOT_PRODUCT(int num_steals = 0;) 6113 oop obj_to_scan; 6114 6115 while (true) { 6116 // Completely finish any left over work from (an) earlier round(s) 6117 drain->trim_queue(0); 6118 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 6119 (size_t)ParGCDesiredObjsFromOverflowList); 6120 // Now check if there's any work in the overflow list 6121 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads, 6122 // only affects the number of attempts made to get work from the 6123 // overflow list and does not affect the number of workers. Just 6124 // pass ParallelGCThreads so this behavior is unchanged. 6125 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 6126 work_q, 6127 ParallelGCThreads)) { 6128 // Found something in global overflow list; 6129 // not yet ready to go stealing work from others. 6130 // We'd like to assert(work_q->size() != 0, ...) 6131 // because we just took work from the overflow list, 6132 // but of course we can't, since all of that might have 6133 // been already stolen from us. 6134 continue; 6135 } 6136 // Verify that we have no work before we resort to stealing 6137 assert(work_q->size() == 0, "Have work, shouldn't steal"); 6138 // Try to steal from other queues that have work 6139 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 6140 NOT_PRODUCT(num_steals++;) 6141 assert(obj_to_scan->is_oop(), "Oops, not an oop!"); 6142 assert(_mark_bit_map->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 6143 // Do scanning work 6144 obj_to_scan->oop_iterate(keep_alive); 6145 // Loop around, finish this work, and try to steal some more 6146 } else if (terminator()->offer_termination()) { 6147 break; // nirvana from the infinite cycle 6148 } 6149 } 6150 NOT_PRODUCT( 6151 if (PrintCMSStatistics != 0) { 6152 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals); 6153 } 6154 ) 6155 } 6156 6157 void CMSRefProcTaskExecutor::execute(ProcessTask& task) 6158 { 6159 GenCollectedHeap* gch = GenCollectedHeap::heap(); 6160 FlexibleWorkGang* workers = gch->workers(); 6161 assert(workers != NULL, "Need parallel worker threads."); 6162 CMSRefProcTaskProxy rp_task(task, &_collector, 6163 _collector.ref_processor()->span(), 6164 _collector.markBitMap(), 6165 workers, _collector.task_queues()); 6166 workers->run_task(&rp_task); 6167 } 6168 6169 void CMSRefProcTaskExecutor::execute(EnqueueTask& task) 6170 { 6171 6172 GenCollectedHeap* gch = GenCollectedHeap::heap(); 6173 FlexibleWorkGang* workers = gch->workers(); 6174 assert(workers != NULL, "Need parallel worker threads."); 6175 CMSRefEnqueueTaskProxy enq_task(task); 6176 workers->run_task(&enq_task); 6177 } 6178 6179 void CMSCollector::refProcessingWork(bool asynch, bool clear_all_soft_refs) { 6180 6181 ResourceMark rm; 6182 HandleMark hm; 6183 6184 ReferenceProcessor* rp = ref_processor(); 6185 assert(rp->span().equals(_span), "Spans should be equal"); 6186 assert(!rp->enqueuing_is_done(), "Enqueuing should not be complete"); 6187 // Process weak references. 6188 rp->setup_policy(clear_all_soft_refs); 6189 verify_work_stacks_empty(); 6190 6191 CMSKeepAliveClosure cmsKeepAliveClosure(this, _span, &_markBitMap, 6192 &_markStack, false /* !preclean */); 6193 CMSDrainMarkingStackClosure cmsDrainMarkingStackClosure(this, 6194 _span, &_markBitMap, &_markStack, 6195 &cmsKeepAliveClosure, false /* !preclean */); 6196 { 6197 GCTraceTime t("weak refs processing", PrintGCDetails, false, _gc_timer_cm); 6198 6199 ReferenceProcessorStats stats; 6200 if (rp->processing_is_mt()) { 6201 // Set the degree of MT here. If the discovery is done MT, there 6202 // may have been a different number of threads doing the discovery 6203 // and a different number of discovered lists may have Ref objects. 6204 // That is OK as long as the Reference lists are balanced (see 6205 // balance_all_queues() and balance_queues()). 6206 GenCollectedHeap* gch = GenCollectedHeap::heap(); 6207 int active_workers = ParallelGCThreads; 6208 FlexibleWorkGang* workers = gch->workers(); 6209 if (workers != NULL) { 6210 active_workers = workers->active_workers(); 6211 // The expectation is that active_workers will have already 6212 // been set to a reasonable value. If it has not been set, 6213 // investigate. 6214 assert(active_workers > 0, "Should have been set during scavenge"); 6215 } 6216 rp->set_active_mt_degree(active_workers); 6217 CMSRefProcTaskExecutor task_executor(*this); 6218 stats = rp->process_discovered_references(&_is_alive_closure, 6219 &cmsKeepAliveClosure, 6220 &cmsDrainMarkingStackClosure, 6221 &task_executor, 6222 _gc_timer_cm); 6223 } else { 6224 stats = rp->process_discovered_references(&_is_alive_closure, 6225 &cmsKeepAliveClosure, 6226 &cmsDrainMarkingStackClosure, 6227 NULL, 6228 _gc_timer_cm); 6229 } 6230 _gc_tracer_cm->report_gc_reference_stats(stats); 6231 6232 } 6233 6234 // This is the point where the entire marking should have completed. 6235 verify_work_stacks_empty(); 6236 6237 if (should_unload_classes()) { 6238 { 6239 GCTraceTime t("class unloading", PrintGCDetails, false, _gc_timer_cm); 6240 6241 // Unload classes and purge the SystemDictionary. 6242 bool purged_class = SystemDictionary::do_unloading(&_is_alive_closure); 6243 6244 // Unload nmethods. 6245 CodeCache::do_unloading(&_is_alive_closure, purged_class); 6246 6247 // Prune dead klasses from subklass/sibling/implementor lists. 6248 Klass::clean_weak_klass_links(&_is_alive_closure); 6249 } 6250 6251 { 6252 GCTraceTime t("scrub symbol table", PrintGCDetails, false, _gc_timer_cm); 6253 // Clean up unreferenced symbols in symbol table. 6254 SymbolTable::unlink(); 6255 } 6256 } 6257 6258 // CMS doesn't use the StringTable as hard roots when class unloading is turned off. 6259 // Need to check if we really scanned the StringTable. 6260 if ((roots_scanning_options() & SharedHeap::SO_Strings) == 0) { 6261 GCTraceTime t("scrub string table", PrintGCDetails, false, _gc_timer_cm); 6262 // Delete entries for dead interned strings. 6263 StringTable::unlink(&_is_alive_closure); 6264 } 6265 6266 // Restore any preserved marks as a result of mark stack or 6267 // work queue overflow 6268 restore_preserved_marks_if_any(); // done single-threaded for now 6269 6270 rp->set_enqueuing_is_done(true); 6271 if (rp->processing_is_mt()) { 6272 rp->balance_all_queues(); 6273 CMSRefProcTaskExecutor task_executor(*this); 6274 rp->enqueue_discovered_references(&task_executor); 6275 } else { 6276 rp->enqueue_discovered_references(NULL); 6277 } 6278 rp->verify_no_references_recorded(); 6279 assert(!rp->discovery_enabled(), "should have been disabled"); 6280 } 6281 6282 #ifndef PRODUCT 6283 void CMSCollector::check_correct_thread_executing() { 6284 Thread* t = Thread::current(); 6285 // Only the VM thread or the CMS thread should be here. 6286 assert(t->is_ConcurrentGC_thread() || t->is_VM_thread(), 6287 "Unexpected thread type"); 6288 // If this is the vm thread, the foreground process 6289 // should not be waiting. Note that _foregroundGCIsActive is 6290 // true while the foreground collector is waiting. 6291 if (_foregroundGCShouldWait) { 6292 // We cannot be the VM thread 6293 assert(t->is_ConcurrentGC_thread(), 6294 "Should be CMS thread"); 6295 } else { 6296 // We can be the CMS thread only if we are in a stop-world 6297 // phase of CMS collection. 6298 if (t->is_ConcurrentGC_thread()) { 6299 assert(_collectorState == InitialMarking || 6300 _collectorState == FinalMarking, 6301 "Should be a stop-world phase"); 6302 // The CMS thread should be holding the CMS_token. 6303 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6304 "Potential interference with concurrently " 6305 "executing VM thread"); 6306 } 6307 } 6308 } 6309 #endif 6310 6311 void CMSCollector::sweep(bool asynch) { 6312 assert(_collectorState == Sweeping, "just checking"); 6313 check_correct_thread_executing(); 6314 verify_work_stacks_empty(); 6315 verify_overflow_empty(); 6316 increment_sweep_count(); 6317 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 6318 6319 _inter_sweep_timer.stop(); 6320 _inter_sweep_estimate.sample(_inter_sweep_timer.seconds()); 6321 size_policy()->avg_cms_free_at_sweep()->sample(_cmsGen->free()); 6322 6323 assert(!_intra_sweep_timer.is_active(), "Should not be active"); 6324 _intra_sweep_timer.reset(); 6325 _intra_sweep_timer.start(); 6326 if (asynch) { 6327 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 6328 CMSPhaseAccounting pa(this, "sweep", !PrintGCDetails); 6329 // First sweep the old gen 6330 { 6331 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(), 6332 bitMapLock()); 6333 sweepWork(_cmsGen, asynch); 6334 } 6335 6336 // Update Universe::_heap_*_at_gc figures. 6337 // We need all the free list locks to make the abstract state 6338 // transition from Sweeping to Resetting. See detailed note 6339 // further below. 6340 { 6341 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock()); 6342 // Update heap occupancy information which is used as 6343 // input to soft ref clearing policy at the next gc. 6344 Universe::update_heap_info_at_gc(); 6345 _collectorState = Resizing; 6346 } 6347 } else { 6348 // already have needed locks 6349 sweepWork(_cmsGen, asynch); 6350 // Update heap occupancy information which is used as 6351 // input to soft ref clearing policy at the next gc. 6352 Universe::update_heap_info_at_gc(); 6353 _collectorState = Resizing; 6354 } 6355 verify_work_stacks_empty(); 6356 verify_overflow_empty(); 6357 6358 if (should_unload_classes()) { 6359 ClassLoaderDataGraph::purge(); 6360 } 6361 6362 _intra_sweep_timer.stop(); 6363 _intra_sweep_estimate.sample(_intra_sweep_timer.seconds()); 6364 6365 _inter_sweep_timer.reset(); 6366 _inter_sweep_timer.start(); 6367 6368 // We need to use a monotonically non-decreasing time in ms 6369 // or we will see time-warp warnings and os::javaTimeMillis() 6370 // does not guarantee monotonicity. 6371 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC; 6372 update_time_of_last_gc(now); 6373 6374 // NOTE on abstract state transitions: 6375 // Mutators allocate-live and/or mark the mod-union table dirty 6376 // based on the state of the collection. The former is done in 6377 // the interval [Marking, Sweeping] and the latter in the interval 6378 // [Marking, Sweeping). Thus the transitions into the Marking state 6379 // and out of the Sweeping state must be synchronously visible 6380 // globally to the mutators. 6381 // The transition into the Marking state happens with the world 6382 // stopped so the mutators will globally see it. Sweeping is 6383 // done asynchronously by the background collector so the transition 6384 // from the Sweeping state to the Resizing state must be done 6385 // under the freelistLock (as is the check for whether to 6386 // allocate-live and whether to dirty the mod-union table). 6387 assert(_collectorState == Resizing, "Change of collector state to" 6388 " Resizing must be done under the freelistLocks (plural)"); 6389 6390 // Now that sweeping has been completed, we clear 6391 // the incremental_collection_failed flag, 6392 // thus inviting a younger gen collection to promote into 6393 // this generation. If such a promotion may still fail, 6394 // the flag will be set again when a young collection is 6395 // attempted. 6396 GenCollectedHeap* gch = GenCollectedHeap::heap(); 6397 gch->clear_incremental_collection_failed(); // Worth retrying as fresh space may have been freed up 6398 gch->update_full_collections_completed(_collection_count_start); 6399 } 6400 6401 // FIX ME!!! Looks like this belongs in CFLSpace, with 6402 // CMSGen merely delegating to it. 6403 void ConcurrentMarkSweepGeneration::setNearLargestChunk() { 6404 double nearLargestPercent = FLSLargestBlockCoalesceProximity; 6405 HeapWord* minAddr = _cmsSpace->bottom(); 6406 HeapWord* largestAddr = 6407 (HeapWord*) _cmsSpace->dictionary()->find_largest_dict(); 6408 if (largestAddr == NULL) { 6409 // The dictionary appears to be empty. In this case 6410 // try to coalesce at the end of the heap. 6411 largestAddr = _cmsSpace->end(); 6412 } 6413 size_t largestOffset = pointer_delta(largestAddr, minAddr); 6414 size_t nearLargestOffset = 6415 (size_t)((double)largestOffset * nearLargestPercent) - MinChunkSize; 6416 if (PrintFLSStatistics != 0) { 6417 gclog_or_tty->print_cr( 6418 "CMS: Large Block: " PTR_FORMAT ";" 6419 " Proximity: " PTR_FORMAT " -> " PTR_FORMAT, 6420 largestAddr, 6421 _cmsSpace->nearLargestChunk(), minAddr + nearLargestOffset); 6422 } 6423 _cmsSpace->set_nearLargestChunk(minAddr + nearLargestOffset); 6424 } 6425 6426 bool ConcurrentMarkSweepGeneration::isNearLargestChunk(HeapWord* addr) { 6427 return addr >= _cmsSpace->nearLargestChunk(); 6428 } 6429 6430 FreeChunk* ConcurrentMarkSweepGeneration::find_chunk_at_end() { 6431 return _cmsSpace->find_chunk_at_end(); 6432 } 6433 6434 void ConcurrentMarkSweepGeneration::update_gc_stats(int current_level, 6435 bool full) { 6436 // The next lower level has been collected. Gather any statistics 6437 // that are of interest at this point. 6438 if (!full && (current_level + 1) == level()) { 6439 // Gather statistics on the young generation collection. 6440 collector()->stats().record_gc0_end(used()); 6441 } 6442 } 6443 6444 CMSAdaptiveSizePolicy* ConcurrentMarkSweepGeneration::size_policy() { 6445 GenCollectedHeap* gch = GenCollectedHeap::heap(); 6446 assert(gch->kind() == CollectedHeap::GenCollectedHeap, 6447 "Wrong type of heap"); 6448 CMSAdaptiveSizePolicy* sp = (CMSAdaptiveSizePolicy*) 6449 gch->gen_policy()->size_policy(); 6450 assert(sp->is_gc_cms_adaptive_size_policy(), 6451 "Wrong type of size policy"); 6452 return sp; 6453 } 6454 6455 void ConcurrentMarkSweepGeneration::rotate_debug_collection_type() { 6456 if (PrintGCDetails && Verbose) { 6457 gclog_or_tty->print("Rotate from %d ", _debug_collection_type); 6458 } 6459 _debug_collection_type = (CollectionTypes) (_debug_collection_type + 1); 6460 _debug_collection_type = 6461 (CollectionTypes) (_debug_collection_type % Unknown_collection_type); 6462 if (PrintGCDetails && Verbose) { 6463 gclog_or_tty->print_cr("to %d ", _debug_collection_type); 6464 } 6465 } 6466 6467 void CMSCollector::sweepWork(ConcurrentMarkSweepGeneration* gen, 6468 bool asynch) { 6469 // We iterate over the space(s) underlying this generation, 6470 // checking the mark bit map to see if the bits corresponding 6471 // to specific blocks are marked or not. Blocks that are 6472 // marked are live and are not swept up. All remaining blocks 6473 // are swept up, with coalescing on-the-fly as we sweep up 6474 // contiguous free and/or garbage blocks: 6475 // We need to ensure that the sweeper synchronizes with allocators 6476 // and stop-the-world collectors. In particular, the following 6477 // locks are used: 6478 // . CMS token: if this is held, a stop the world collection cannot occur 6479 // . freelistLock: if this is held no allocation can occur from this 6480 // generation by another thread 6481 // . bitMapLock: if this is held, no other thread can access or update 6482 // 6483 6484 // Note that we need to hold the freelistLock if we use 6485 // block iterate below; else the iterator might go awry if 6486 // a mutator (or promotion) causes block contents to change 6487 // (for instance if the allocator divvies up a block). 6488 // If we hold the free list lock, for all practical purposes 6489 // young generation GC's can't occur (they'll usually need to 6490 // promote), so we might as well prevent all young generation 6491 // GC's while we do a sweeping step. For the same reason, we might 6492 // as well take the bit map lock for the entire duration 6493 6494 // check that we hold the requisite locks 6495 assert(have_cms_token(), "Should hold cms token"); 6496 assert( (asynch && ConcurrentMarkSweepThread::cms_thread_has_cms_token()) 6497 || (!asynch && ConcurrentMarkSweepThread::vm_thread_has_cms_token()), 6498 "Should possess CMS token to sweep"); 6499 assert_lock_strong(gen->freelistLock()); 6500 assert_lock_strong(bitMapLock()); 6501 6502 assert(!_inter_sweep_timer.is_active(), "Was switched off in an outer context"); 6503 assert(_intra_sweep_timer.is_active(), "Was switched on in an outer context"); 6504 gen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()), 6505 _inter_sweep_estimate.padded_average(), 6506 _intra_sweep_estimate.padded_average()); 6507 gen->setNearLargestChunk(); 6508 6509 { 6510 SweepClosure sweepClosure(this, gen, &_markBitMap, 6511 CMSYield && asynch); 6512 gen->cmsSpace()->blk_iterate_careful(&sweepClosure); 6513 // We need to free-up/coalesce garbage/blocks from a 6514 // co-terminal free run. This is done in the SweepClosure 6515 // destructor; so, do not remove this scope, else the 6516 // end-of-sweep-census below will be off by a little bit. 6517 } 6518 gen->cmsSpace()->sweep_completed(); 6519 gen->cmsSpace()->endSweepFLCensus(sweep_count()); 6520 if (should_unload_classes()) { // unloaded classes this cycle, 6521 _concurrent_cycles_since_last_unload = 0; // ... reset count 6522 } else { // did not unload classes, 6523 _concurrent_cycles_since_last_unload++; // ... increment count 6524 } 6525 } 6526 6527 // Reset CMS data structures (for now just the marking bit map) 6528 // preparatory for the next cycle. 6529 void CMSCollector::reset(bool asynch) { 6530 GenCollectedHeap* gch = GenCollectedHeap::heap(); 6531 CMSAdaptiveSizePolicy* sp = size_policy(); 6532 AdaptiveSizePolicyOutput(sp, gch->total_collections()); 6533 if (asynch) { 6534 CMSTokenSyncWithLocks ts(true, bitMapLock()); 6535 6536 // If the state is not "Resetting", the foreground thread 6537 // has done a collection and the resetting. 6538 if (_collectorState != Resetting) { 6539 assert(_collectorState == Idling, "The state should only change" 6540 " because the foreground collector has finished the collection"); 6541 return; 6542 } 6543 6544 // Clear the mark bitmap (no grey objects to start with) 6545 // for the next cycle. 6546 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 6547 CMSPhaseAccounting cmspa(this, "reset", !PrintGCDetails); 6548 6549 HeapWord* curAddr = _markBitMap.startWord(); 6550 while (curAddr < _markBitMap.endWord()) { 6551 size_t remaining = pointer_delta(_markBitMap.endWord(), curAddr); 6552 MemRegion chunk(curAddr, MIN2(CMSBitMapYieldQuantum, remaining)); 6553 _markBitMap.clear_large_range(chunk); 6554 if (ConcurrentMarkSweepThread::should_yield() && 6555 !foregroundGCIsActive() && 6556 CMSYield) { 6557 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6558 "CMS thread should hold CMS token"); 6559 assert_lock_strong(bitMapLock()); 6560 bitMapLock()->unlock(); 6561 ConcurrentMarkSweepThread::desynchronize(true); 6562 ConcurrentMarkSweepThread::acknowledge_yield_request(); 6563 stopTimer(); 6564 if (PrintCMSStatistics != 0) { 6565 incrementYields(); 6566 } 6567 icms_wait(); 6568 6569 // See the comment in coordinator_yield() 6570 for (unsigned i = 0; i < CMSYieldSleepCount && 6571 ConcurrentMarkSweepThread::should_yield() && 6572 !CMSCollector::foregroundGCIsActive(); ++i) { 6573 os::sleep(Thread::current(), 1, false); 6574 ConcurrentMarkSweepThread::acknowledge_yield_request(); 6575 } 6576 6577 ConcurrentMarkSweepThread::synchronize(true); 6578 bitMapLock()->lock_without_safepoint_check(); 6579 startTimer(); 6580 } 6581 curAddr = chunk.end(); 6582 } 6583 // A successful mostly concurrent collection has been done. 6584 // Because only the full (i.e., concurrent mode failure) collections 6585 // are being measured for gc overhead limits, clean the "near" flag 6586 // and count. 6587 sp->reset_gc_overhead_limit_count(); 6588 _collectorState = Idling; 6589 } else { 6590 // already have the lock 6591 assert(_collectorState == Resetting, "just checking"); 6592 assert_lock_strong(bitMapLock()); 6593 _markBitMap.clear_all(); 6594 _collectorState = Idling; 6595 } 6596 6597 // Stop incremental mode after a cycle completes, so that any future cycles 6598 // are triggered by allocation. 6599 stop_icms(); 6600 6601 NOT_PRODUCT( 6602 if (RotateCMSCollectionTypes) { 6603 _cmsGen->rotate_debug_collection_type(); 6604 } 6605 ) 6606 6607 register_gc_end(); 6608 } 6609 6610 void CMSCollector::do_CMS_operation(CMS_op_type op, GCCause::Cause gc_cause) { 6611 gclog_or_tty->date_stamp(PrintGC && PrintGCDateStamps); 6612 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 6613 GCTraceTime t(GCCauseString("GC", gc_cause), PrintGC, !PrintGCDetails, NULL); 6614 TraceCollectorStats tcs(counters()); 6615 6616 switch (op) { 6617 case CMS_op_checkpointRootsInitial: { 6618 SvcGCMarker sgcm(SvcGCMarker::OTHER); 6619 checkpointRootsInitial(true); // asynch 6620 if (PrintGC) { 6621 _cmsGen->printOccupancy("initial-mark"); 6622 } 6623 break; 6624 } 6625 case CMS_op_checkpointRootsFinal: { 6626 SvcGCMarker sgcm(SvcGCMarker::OTHER); 6627 checkpointRootsFinal(true, // asynch 6628 false, // !clear_all_soft_refs 6629 false); // !init_mark_was_synchronous 6630 if (PrintGC) { 6631 _cmsGen->printOccupancy("remark"); 6632 } 6633 break; 6634 } 6635 default: 6636 fatal("No such CMS_op"); 6637 } 6638 } 6639 6640 #ifndef PRODUCT 6641 size_t const CMSCollector::skip_header_HeapWords() { 6642 return FreeChunk::header_size(); 6643 } 6644 6645 // Try and collect here conditions that should hold when 6646 // CMS thread is exiting. The idea is that the foreground GC 6647 // thread should not be blocked if it wants to terminate 6648 // the CMS thread and yet continue to run the VM for a while 6649 // after that. 6650 void CMSCollector::verify_ok_to_terminate() const { 6651 assert(Thread::current()->is_ConcurrentGC_thread(), 6652 "should be called by CMS thread"); 6653 assert(!_foregroundGCShouldWait, "should be false"); 6654 // We could check here that all the various low-level locks 6655 // are not held by the CMS thread, but that is overkill; see 6656 // also CMSThread::verify_ok_to_terminate() where the CGC_lock 6657 // is checked. 6658 } 6659 #endif 6660 6661 size_t CMSCollector::block_size_using_printezis_bits(HeapWord* addr) const { 6662 assert(_markBitMap.isMarked(addr) && _markBitMap.isMarked(addr + 1), 6663 "missing Printezis mark?"); 6664 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 6665 size_t size = pointer_delta(nextOneAddr + 1, addr); 6666 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 6667 "alignment problem"); 6668 assert(size >= 3, "Necessary for Printezis marks to work"); 6669 return size; 6670 } 6671 6672 // A variant of the above (block_size_using_printezis_bits()) except 6673 // that we return 0 if the P-bits are not yet set. 6674 size_t CMSCollector::block_size_if_printezis_bits(HeapWord* addr) const { 6675 if (_markBitMap.isMarked(addr + 1)) { 6676 assert(_markBitMap.isMarked(addr), "P-bit can be set only for marked objects"); 6677 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 6678 size_t size = pointer_delta(nextOneAddr + 1, addr); 6679 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 6680 "alignment problem"); 6681 assert(size >= 3, "Necessary for Printezis marks to work"); 6682 return size; 6683 } 6684 return 0; 6685 } 6686 6687 HeapWord* CMSCollector::next_card_start_after_block(HeapWord* addr) const { 6688 size_t sz = 0; 6689 oop p = (oop)addr; 6690 if (p->klass_or_null() != NULL) { 6691 sz = CompactibleFreeListSpace::adjustObjectSize(p->size()); 6692 } else { 6693 sz = block_size_using_printezis_bits(addr); 6694 } 6695 assert(sz > 0, "size must be nonzero"); 6696 HeapWord* next_block = addr + sz; 6697 HeapWord* next_card = (HeapWord*)round_to((uintptr_t)next_block, 6698 CardTableModRefBS::card_size); 6699 assert(round_down((uintptr_t)addr, CardTableModRefBS::card_size) < 6700 round_down((uintptr_t)next_card, CardTableModRefBS::card_size), 6701 "must be different cards"); 6702 return next_card; 6703 } 6704 6705 6706 // CMS Bit Map Wrapper ///////////////////////////////////////// 6707 6708 // Construct a CMS bit map infrastructure, but don't create the 6709 // bit vector itself. That is done by a separate call CMSBitMap::allocate() 6710 // further below. 6711 CMSBitMap::CMSBitMap(int shifter, int mutex_rank, const char* mutex_name): 6712 _bm(), 6713 _shifter(shifter), 6714 _lock(mutex_rank >= 0 ? new Mutex(mutex_rank, mutex_name, true) : NULL) 6715 { 6716 _bmStartWord = 0; 6717 _bmWordSize = 0; 6718 } 6719 6720 bool CMSBitMap::allocate(MemRegion mr) { 6721 _bmStartWord = mr.start(); 6722 _bmWordSize = mr.word_size(); 6723 ReservedSpace brs(ReservedSpace::allocation_align_size_up( 6724 (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1)); 6725 if (!brs.is_reserved()) { 6726 warning("CMS bit map allocation failure"); 6727 return false; 6728 } 6729 // For now we'll just commit all of the bit map up front. 6730 // Later on we'll try to be more parsimonious with swap. 6731 if (!_virtual_space.initialize(brs, brs.size())) { 6732 warning("CMS bit map backing store failure"); 6733 return false; 6734 } 6735 assert(_virtual_space.committed_size() == brs.size(), 6736 "didn't reserve backing store for all of CMS bit map?"); 6737 _bm.set_map((BitMap::bm_word_t*)_virtual_space.low()); 6738 assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >= 6739 _bmWordSize, "inconsistency in bit map sizing"); 6740 _bm.set_size(_bmWordSize >> _shifter); 6741 6742 // bm.clear(); // can we rely on getting zero'd memory? verify below 6743 assert(isAllClear(), 6744 "Expected zero'd memory from ReservedSpace constructor"); 6745 assert(_bm.size() == heapWordDiffToOffsetDiff(sizeInWords()), 6746 "consistency check"); 6747 return true; 6748 } 6749 6750 void CMSBitMap::dirty_range_iterate_clear(MemRegion mr, MemRegionClosure* cl) { 6751 HeapWord *next_addr, *end_addr, *last_addr; 6752 assert_locked(); 6753 assert(covers(mr), "out-of-range error"); 6754 // XXX assert that start and end are appropriately aligned 6755 for (next_addr = mr.start(), end_addr = mr.end(); 6756 next_addr < end_addr; next_addr = last_addr) { 6757 MemRegion dirty_region = getAndClearMarkedRegion(next_addr, end_addr); 6758 last_addr = dirty_region.end(); 6759 if (!dirty_region.is_empty()) { 6760 cl->do_MemRegion(dirty_region); 6761 } else { 6762 assert(last_addr == end_addr, "program logic"); 6763 return; 6764 } 6765 } 6766 } 6767 6768 void CMSBitMap::print_on_error(outputStream* st, const char* prefix) const { 6769 _bm.print_on_error(st, prefix); 6770 } 6771 6772 #ifndef PRODUCT 6773 void CMSBitMap::assert_locked() const { 6774 CMSLockVerifier::assert_locked(lock()); 6775 } 6776 6777 bool CMSBitMap::covers(MemRegion mr) const { 6778 // assert(_bm.map() == _virtual_space.low(), "map inconsistency"); 6779 assert((size_t)_bm.size() == (_bmWordSize >> _shifter), 6780 "size inconsistency"); 6781 return (mr.start() >= _bmStartWord) && 6782 (mr.end() <= endWord()); 6783 } 6784 6785 bool CMSBitMap::covers(HeapWord* start, size_t size) const { 6786 return (start >= _bmStartWord && (start + size) <= endWord()); 6787 } 6788 6789 void CMSBitMap::verifyNoOneBitsInRange(HeapWord* left, HeapWord* right) { 6790 // verify that there are no 1 bits in the interval [left, right) 6791 FalseBitMapClosure falseBitMapClosure; 6792 iterate(&falseBitMapClosure, left, right); 6793 } 6794 6795 void CMSBitMap::region_invariant(MemRegion mr) 6796 { 6797 assert_locked(); 6798 // mr = mr.intersection(MemRegion(_bmStartWord, _bmWordSize)); 6799 assert(!mr.is_empty(), "unexpected empty region"); 6800 assert(covers(mr), "mr should be covered by bit map"); 6801 // convert address range into offset range 6802 size_t start_ofs = heapWordToOffset(mr.start()); 6803 // Make sure that end() is appropriately aligned 6804 assert(mr.end() == (HeapWord*)round_to((intptr_t)mr.end(), 6805 (1 << (_shifter+LogHeapWordSize))), 6806 "Misaligned mr.end()"); 6807 size_t end_ofs = heapWordToOffset(mr.end()); 6808 assert(end_ofs > start_ofs, "Should mark at least one bit"); 6809 } 6810 6811 #endif 6812 6813 bool CMSMarkStack::allocate(size_t size) { 6814 // allocate a stack of the requisite depth 6815 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 6816 size * sizeof(oop))); 6817 if (!rs.is_reserved()) { 6818 warning("CMSMarkStack allocation failure"); 6819 return false; 6820 } 6821 if (!_virtual_space.initialize(rs, rs.size())) { 6822 warning("CMSMarkStack backing store failure"); 6823 return false; 6824 } 6825 assert(_virtual_space.committed_size() == rs.size(), 6826 "didn't reserve backing store for all of CMS stack?"); 6827 _base = (oop*)(_virtual_space.low()); 6828 _index = 0; 6829 _capacity = size; 6830 NOT_PRODUCT(_max_depth = 0); 6831 return true; 6832 } 6833 6834 // XXX FIX ME !!! In the MT case we come in here holding a 6835 // leaf lock. For printing we need to take a further lock 6836 // which has lower rank. We need to recalibrate the two 6837 // lock-ranks involved in order to be able to print the 6838 // messages below. (Or defer the printing to the caller. 6839 // For now we take the expedient path of just disabling the 6840 // messages for the problematic case.) 6841 void CMSMarkStack::expand() { 6842 assert(_capacity <= MarkStackSizeMax, "stack bigger than permitted"); 6843 if (_capacity == MarkStackSizeMax) { 6844 if (_hit_limit++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) { 6845 // We print a warning message only once per CMS cycle. 6846 gclog_or_tty->print_cr(" (benign) Hit CMSMarkStack max size limit"); 6847 } 6848 return; 6849 } 6850 // Double capacity if possible 6851 size_t new_capacity = MIN2(_capacity*2, MarkStackSizeMax); 6852 // Do not give up existing stack until we have managed to 6853 // get the double capacity that we desired. 6854 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 6855 new_capacity * sizeof(oop))); 6856 if (rs.is_reserved()) { 6857 // Release the backing store associated with old stack 6858 _virtual_space.release(); 6859 // Reinitialize virtual space for new stack 6860 if (!_virtual_space.initialize(rs, rs.size())) { 6861 fatal("Not enough swap for expanded marking stack"); 6862 } 6863 _base = (oop*)(_virtual_space.low()); 6864 _index = 0; 6865 _capacity = new_capacity; 6866 } else if (_failed_double++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) { 6867 // Failed to double capacity, continue; 6868 // we print a detail message only once per CMS cycle. 6869 gclog_or_tty->print(" (benign) Failed to expand marking stack from "SIZE_FORMAT"K to " 6870 SIZE_FORMAT"K", 6871 _capacity / K, new_capacity / K); 6872 } 6873 } 6874 6875 6876 // Closures 6877 // XXX: there seems to be a lot of code duplication here; 6878 // should refactor and consolidate common code. 6879 6880 // This closure is used to mark refs into the CMS generation in 6881 // the CMS bit map. Called at the first checkpoint. This closure 6882 // assumes that we do not need to re-mark dirty cards; if the CMS 6883 // generation on which this is used is not an oldest 6884 // generation then this will lose younger_gen cards! 6885 6886 MarkRefsIntoClosure::MarkRefsIntoClosure( 6887 MemRegion span, CMSBitMap* bitMap): 6888 _span(span), 6889 _bitMap(bitMap) 6890 { 6891 assert(_ref_processor == NULL, "deliberately left NULL"); 6892 assert(_bitMap->covers(_span), "_bitMap/_span mismatch"); 6893 } 6894 6895 void MarkRefsIntoClosure::do_oop(oop obj) { 6896 // if p points into _span, then mark corresponding bit in _markBitMap 6897 assert(obj->is_oop(), "expected an oop"); 6898 HeapWord* addr = (HeapWord*)obj; 6899 if (_span.contains(addr)) { 6900 // this should be made more efficient 6901 _bitMap->mark(addr); 6902 } 6903 } 6904 6905 void MarkRefsIntoClosure::do_oop(oop* p) { MarkRefsIntoClosure::do_oop_work(p); } 6906 void MarkRefsIntoClosure::do_oop(narrowOop* p) { MarkRefsIntoClosure::do_oop_work(p); } 6907 6908 Par_MarkRefsIntoClosure::Par_MarkRefsIntoClosure( 6909 MemRegion span, CMSBitMap* bitMap): 6910 _span(span), 6911 _bitMap(bitMap) 6912 { 6913 assert(_ref_processor == NULL, "deliberately left NULL"); 6914 assert(_bitMap->covers(_span), "_bitMap/_span mismatch"); 6915 } 6916 6917 void Par_MarkRefsIntoClosure::do_oop(oop obj) { 6918 // if p points into _span, then mark corresponding bit in _markBitMap 6919 assert(obj->is_oop(), "expected an oop"); 6920 HeapWord* addr = (HeapWord*)obj; 6921 if (_span.contains(addr)) { 6922 // this should be made more efficient 6923 _bitMap->par_mark(addr); 6924 } 6925 } 6926 6927 void Par_MarkRefsIntoClosure::do_oop(oop* p) { Par_MarkRefsIntoClosure::do_oop_work(p); } 6928 void Par_MarkRefsIntoClosure::do_oop(narrowOop* p) { Par_MarkRefsIntoClosure::do_oop_work(p); } 6929 6930 // A variant of the above, used for CMS marking verification. 6931 MarkRefsIntoVerifyClosure::MarkRefsIntoVerifyClosure( 6932 MemRegion span, CMSBitMap* verification_bm, CMSBitMap* cms_bm): 6933 _span(span), 6934 _verification_bm(verification_bm), 6935 _cms_bm(cms_bm) 6936 { 6937 assert(_ref_processor == NULL, "deliberately left NULL"); 6938 assert(_verification_bm->covers(_span), "_verification_bm/_span mismatch"); 6939 } 6940 6941 void MarkRefsIntoVerifyClosure::do_oop(oop obj) { 6942 // if p points into _span, then mark corresponding bit in _markBitMap 6943 assert(obj->is_oop(), "expected an oop"); 6944 HeapWord* addr = (HeapWord*)obj; 6945 if (_span.contains(addr)) { 6946 _verification_bm->mark(addr); 6947 if (!_cms_bm->isMarked(addr)) { 6948 oop(addr)->print(); 6949 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)", addr); 6950 fatal("... aborting"); 6951 } 6952 } 6953 } 6954 6955 void MarkRefsIntoVerifyClosure::do_oop(oop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); } 6956 void MarkRefsIntoVerifyClosure::do_oop(narrowOop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); } 6957 6958 ////////////////////////////////////////////////// 6959 // MarkRefsIntoAndScanClosure 6960 ////////////////////////////////////////////////// 6961 6962 MarkRefsIntoAndScanClosure::MarkRefsIntoAndScanClosure(MemRegion span, 6963 ReferenceProcessor* rp, 6964 CMSBitMap* bit_map, 6965 CMSBitMap* mod_union_table, 6966 CMSMarkStack* mark_stack, 6967 CMSCollector* collector, 6968 bool should_yield, 6969 bool concurrent_precleaning): 6970 _collector(collector), 6971 _span(span), 6972 _bit_map(bit_map), 6973 _mark_stack(mark_stack), 6974 _pushAndMarkClosure(collector, span, rp, bit_map, mod_union_table, 6975 mark_stack, concurrent_precleaning), 6976 _yield(should_yield), 6977 _concurrent_precleaning(concurrent_precleaning), 6978 _freelistLock(NULL) 6979 { 6980 _ref_processor = rp; 6981 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 6982 } 6983 6984 // This closure is used to mark refs into the CMS generation at the 6985 // second (final) checkpoint, and to scan and transitively follow 6986 // the unmarked oops. It is also used during the concurrent precleaning 6987 // phase while scanning objects on dirty cards in the CMS generation. 6988 // The marks are made in the marking bit map and the marking stack is 6989 // used for keeping the (newly) grey objects during the scan. 6990 // The parallel version (Par_...) appears further below. 6991 void MarkRefsIntoAndScanClosure::do_oop(oop obj) { 6992 if (obj != NULL) { 6993 assert(obj->is_oop(), "expected an oop"); 6994 HeapWord* addr = (HeapWord*)obj; 6995 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 6996 assert(_collector->overflow_list_is_empty(), 6997 "overflow list should be empty"); 6998 if (_span.contains(addr) && 6999 !_bit_map->isMarked(addr)) { 7000 // mark bit map (object is now grey) 7001 _bit_map->mark(addr); 7002 // push on marking stack (stack should be empty), and drain the 7003 // stack by applying this closure to the oops in the oops popped 7004 // from the stack (i.e. blacken the grey objects) 7005 bool res = _mark_stack->push(obj); 7006 assert(res, "Should have space to push on empty stack"); 7007 do { 7008 oop new_oop = _mark_stack->pop(); 7009 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 7010 assert(_bit_map->isMarked((HeapWord*)new_oop), 7011 "only grey objects on this stack"); 7012 // iterate over the oops in this oop, marking and pushing 7013 // the ones in CMS heap (i.e. in _span). 7014 new_oop->oop_iterate(&_pushAndMarkClosure); 7015 // check if it's time to yield 7016 do_yield_check(); 7017 } while (!_mark_stack->isEmpty() || 7018 (!_concurrent_precleaning && take_from_overflow_list())); 7019 // if marking stack is empty, and we are not doing this 7020 // during precleaning, then check the overflow list 7021 } 7022 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 7023 assert(_collector->overflow_list_is_empty(), 7024 "overflow list was drained above"); 7025 // We could restore evacuated mark words, if any, used for 7026 // overflow list links here because the overflow list is 7027 // provably empty here. That would reduce the maximum 7028 // size requirements for preserved_{oop,mark}_stack. 7029 // But we'll just postpone it until we are all done 7030 // so we can just stream through. 7031 if (!_concurrent_precleaning && CMSOverflowEarlyRestoration) { 7032 _collector->restore_preserved_marks_if_any(); 7033 assert(_collector->no_preserved_marks(), "No preserved marks"); 7034 } 7035 assert(!CMSOverflowEarlyRestoration || _collector->no_preserved_marks(), 7036 "All preserved marks should have been restored above"); 7037 } 7038 } 7039 7040 void MarkRefsIntoAndScanClosure::do_oop(oop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); } 7041 void MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); } 7042 7043 void MarkRefsIntoAndScanClosure::do_yield_work() { 7044 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 7045 "CMS thread should hold CMS token"); 7046 assert_lock_strong(_freelistLock); 7047 assert_lock_strong(_bit_map->lock()); 7048 // relinquish the free_list_lock and bitMaplock() 7049 _bit_map->lock()->unlock(); 7050 _freelistLock->unlock(); 7051 ConcurrentMarkSweepThread::desynchronize(true); 7052 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7053 _collector->stopTimer(); 7054 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 7055 if (PrintCMSStatistics != 0) { 7056 _collector->incrementYields(); 7057 } 7058 _collector->icms_wait(); 7059 7060 // See the comment in coordinator_yield() 7061 for (unsigned i = 0; 7062 i < CMSYieldSleepCount && 7063 ConcurrentMarkSweepThread::should_yield() && 7064 !CMSCollector::foregroundGCIsActive(); 7065 ++i) { 7066 os::sleep(Thread::current(), 1, false); 7067 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7068 } 7069 7070 ConcurrentMarkSweepThread::synchronize(true); 7071 _freelistLock->lock_without_safepoint_check(); 7072 _bit_map->lock()->lock_without_safepoint_check(); 7073 _collector->startTimer(); 7074 } 7075 7076 /////////////////////////////////////////////////////////// 7077 // Par_MarkRefsIntoAndScanClosure: a parallel version of 7078 // MarkRefsIntoAndScanClosure 7079 /////////////////////////////////////////////////////////// 7080 Par_MarkRefsIntoAndScanClosure::Par_MarkRefsIntoAndScanClosure( 7081 CMSCollector* collector, MemRegion span, ReferenceProcessor* rp, 7082 CMSBitMap* bit_map, OopTaskQueue* work_queue): 7083 _span(span), 7084 _bit_map(bit_map), 7085 _work_queue(work_queue), 7086 _low_water_mark(MIN2((uint)(work_queue->max_elems()/4), 7087 (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads))), 7088 _par_pushAndMarkClosure(collector, span, rp, bit_map, work_queue) 7089 { 7090 _ref_processor = rp; 7091 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 7092 } 7093 7094 // This closure is used to mark refs into the CMS generation at the 7095 // second (final) checkpoint, and to scan and transitively follow 7096 // the unmarked oops. The marks are made in the marking bit map and 7097 // the work_queue is used for keeping the (newly) grey objects during 7098 // the scan phase whence they are also available for stealing by parallel 7099 // threads. Since the marking bit map is shared, updates are 7100 // synchronized (via CAS). 7101 void Par_MarkRefsIntoAndScanClosure::do_oop(oop obj) { 7102 if (obj != NULL) { 7103 // Ignore mark word because this could be an already marked oop 7104 // that may be chained at the end of the overflow list. 7105 assert(obj->is_oop(true), "expected an oop"); 7106 HeapWord* addr = (HeapWord*)obj; 7107 if (_span.contains(addr) && 7108 !_bit_map->isMarked(addr)) { 7109 // mark bit map (object will become grey): 7110 // It is possible for several threads to be 7111 // trying to "claim" this object concurrently; 7112 // the unique thread that succeeds in marking the 7113 // object first will do the subsequent push on 7114 // to the work queue (or overflow list). 7115 if (_bit_map->par_mark(addr)) { 7116 // push on work_queue (which may not be empty), and trim the 7117 // queue to an appropriate length by applying this closure to 7118 // the oops in the oops popped from the stack (i.e. blacken the 7119 // grey objects) 7120 bool res = _work_queue->push(obj); 7121 assert(res, "Low water mark should be less than capacity?"); 7122 trim_queue(_low_water_mark); 7123 } // Else, another thread claimed the object 7124 } 7125 } 7126 } 7127 7128 void Par_MarkRefsIntoAndScanClosure::do_oop(oop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); } 7129 void Par_MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); } 7130 7131 // This closure is used to rescan the marked objects on the dirty cards 7132 // in the mod union table and the card table proper. 7133 size_t ScanMarkedObjectsAgainCarefullyClosure::do_object_careful_m( 7134 oop p, MemRegion mr) { 7135 7136 size_t size = 0; 7137 HeapWord* addr = (HeapWord*)p; 7138 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 7139 assert(_span.contains(addr), "we are scanning the CMS generation"); 7140 // check if it's time to yield 7141 if (do_yield_check()) { 7142 // We yielded for some foreground stop-world work, 7143 // and we have been asked to abort this ongoing preclean cycle. 7144 return 0; 7145 } 7146 if (_bitMap->isMarked(addr)) { 7147 // it's marked; is it potentially uninitialized? 7148 if (p->klass_or_null() != NULL) { 7149 // an initialized object; ignore mark word in verification below 7150 // since we are running concurrent with mutators 7151 assert(p->is_oop(true), "should be an oop"); 7152 if (p->is_objArray()) { 7153 // objArrays are precisely marked; restrict scanning 7154 // to dirty cards only. 7155 size = CompactibleFreeListSpace::adjustObjectSize( 7156 p->oop_iterate(_scanningClosure, mr)); 7157 } else { 7158 // A non-array may have been imprecisely marked; we need 7159 // to scan object in its entirety. 7160 size = CompactibleFreeListSpace::adjustObjectSize( 7161 p->oop_iterate(_scanningClosure)); 7162 } 7163 #ifdef ASSERT 7164 size_t direct_size = 7165 CompactibleFreeListSpace::adjustObjectSize(p->size()); 7166 assert(size == direct_size, "Inconsistency in size"); 7167 assert(size >= 3, "Necessary for Printezis marks to work"); 7168 if (!_bitMap->isMarked(addr+1)) { 7169 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size); 7170 } else { 7171 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size-1); 7172 assert(_bitMap->isMarked(addr+size-1), 7173 "inconsistent Printezis mark"); 7174 } 7175 #endif // ASSERT 7176 } else { 7177 // An uninitialized object. 7178 assert(_bitMap->isMarked(addr+1), "missing Printezis mark?"); 7179 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 7180 size = pointer_delta(nextOneAddr + 1, addr); 7181 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 7182 "alignment problem"); 7183 // Note that pre-cleaning needn't redirty the card. OopDesc::set_klass() 7184 // will dirty the card when the klass pointer is installed in the 7185 // object (signaling the completion of initialization). 7186 } 7187 } else { 7188 // Either a not yet marked object or an uninitialized object 7189 if (p->klass_or_null() == NULL) { 7190 // An uninitialized object, skip to the next card, since 7191 // we may not be able to read its P-bits yet. 7192 assert(size == 0, "Initial value"); 7193 } else { 7194 // An object not (yet) reached by marking: we merely need to 7195 // compute its size so as to go look at the next block. 7196 assert(p->is_oop(true), "should be an oop"); 7197 size = CompactibleFreeListSpace::adjustObjectSize(p->size()); 7198 } 7199 } 7200 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 7201 return size; 7202 } 7203 7204 void ScanMarkedObjectsAgainCarefullyClosure::do_yield_work() { 7205 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 7206 "CMS thread should hold CMS token"); 7207 assert_lock_strong(_freelistLock); 7208 assert_lock_strong(_bitMap->lock()); 7209 // relinquish the free_list_lock and bitMaplock() 7210 _bitMap->lock()->unlock(); 7211 _freelistLock->unlock(); 7212 ConcurrentMarkSweepThread::desynchronize(true); 7213 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7214 _collector->stopTimer(); 7215 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 7216 if (PrintCMSStatistics != 0) { 7217 _collector->incrementYields(); 7218 } 7219 _collector->icms_wait(); 7220 7221 // See the comment in coordinator_yield() 7222 for (unsigned i = 0; i < CMSYieldSleepCount && 7223 ConcurrentMarkSweepThread::should_yield() && 7224 !CMSCollector::foregroundGCIsActive(); ++i) { 7225 os::sleep(Thread::current(), 1, false); 7226 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7227 } 7228 7229 ConcurrentMarkSweepThread::synchronize(true); 7230 _freelistLock->lock_without_safepoint_check(); 7231 _bitMap->lock()->lock_without_safepoint_check(); 7232 _collector->startTimer(); 7233 } 7234 7235 7236 ////////////////////////////////////////////////////////////////// 7237 // SurvivorSpacePrecleanClosure 7238 ////////////////////////////////////////////////////////////////// 7239 // This (single-threaded) closure is used to preclean the oops in 7240 // the survivor spaces. 7241 size_t SurvivorSpacePrecleanClosure::do_object_careful(oop p) { 7242 7243 HeapWord* addr = (HeapWord*)p; 7244 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 7245 assert(!_span.contains(addr), "we are scanning the survivor spaces"); 7246 assert(p->klass_or_null() != NULL, "object should be initializd"); 7247 // an initialized object; ignore mark word in verification below 7248 // since we are running concurrent with mutators 7249 assert(p->is_oop(true), "should be an oop"); 7250 // Note that we do not yield while we iterate over 7251 // the interior oops of p, pushing the relevant ones 7252 // on our marking stack. 7253 size_t size = p->oop_iterate(_scanning_closure); 7254 do_yield_check(); 7255 // Observe that below, we do not abandon the preclean 7256 // phase as soon as we should; rather we empty the 7257 // marking stack before returning. This is to satisfy 7258 // some existing assertions. In general, it may be a 7259 // good idea to abort immediately and complete the marking 7260 // from the grey objects at a later time. 7261 while (!_mark_stack->isEmpty()) { 7262 oop new_oop = _mark_stack->pop(); 7263 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 7264 assert(_bit_map->isMarked((HeapWord*)new_oop), 7265 "only grey objects on this stack"); 7266 // iterate over the oops in this oop, marking and pushing 7267 // the ones in CMS heap (i.e. in _span). 7268 new_oop->oop_iterate(_scanning_closure); 7269 // check if it's time to yield 7270 do_yield_check(); 7271 } 7272 unsigned int after_count = 7273 GenCollectedHeap::heap()->total_collections(); 7274 bool abort = (_before_count != after_count) || 7275 _collector->should_abort_preclean(); 7276 return abort ? 0 : size; 7277 } 7278 7279 void SurvivorSpacePrecleanClosure::do_yield_work() { 7280 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 7281 "CMS thread should hold CMS token"); 7282 assert_lock_strong(_bit_map->lock()); 7283 // Relinquish the bit map lock 7284 _bit_map->lock()->unlock(); 7285 ConcurrentMarkSweepThread::desynchronize(true); 7286 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7287 _collector->stopTimer(); 7288 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 7289 if (PrintCMSStatistics != 0) { 7290 _collector->incrementYields(); 7291 } 7292 _collector->icms_wait(); 7293 7294 // See the comment in coordinator_yield() 7295 for (unsigned i = 0; i < CMSYieldSleepCount && 7296 ConcurrentMarkSweepThread::should_yield() && 7297 !CMSCollector::foregroundGCIsActive(); ++i) { 7298 os::sleep(Thread::current(), 1, false); 7299 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7300 } 7301 7302 ConcurrentMarkSweepThread::synchronize(true); 7303 _bit_map->lock()->lock_without_safepoint_check(); 7304 _collector->startTimer(); 7305 } 7306 7307 // This closure is used to rescan the marked objects on the dirty cards 7308 // in the mod union table and the card table proper. In the parallel 7309 // case, although the bitMap is shared, we do a single read so the 7310 // isMarked() query is "safe". 7311 bool ScanMarkedObjectsAgainClosure::do_object_bm(oop p, MemRegion mr) { 7312 // Ignore mark word because we are running concurrent with mutators 7313 assert(p->is_oop_or_null(true), "expected an oop or null"); 7314 HeapWord* addr = (HeapWord*)p; 7315 assert(_span.contains(addr), "we are scanning the CMS generation"); 7316 bool is_obj_array = false; 7317 #ifdef ASSERT 7318 if (!_parallel) { 7319 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 7320 assert(_collector->overflow_list_is_empty(), 7321 "overflow list should be empty"); 7322 7323 } 7324 #endif // ASSERT 7325 if (_bit_map->isMarked(addr)) { 7326 // Obj arrays are precisely marked, non-arrays are not; 7327 // so we scan objArrays precisely and non-arrays in their 7328 // entirety. 7329 if (p->is_objArray()) { 7330 is_obj_array = true; 7331 if (_parallel) { 7332 p->oop_iterate(_par_scan_closure, mr); 7333 } else { 7334 p->oop_iterate(_scan_closure, mr); 7335 } 7336 } else { 7337 if (_parallel) { 7338 p->oop_iterate(_par_scan_closure); 7339 } else { 7340 p->oop_iterate(_scan_closure); 7341 } 7342 } 7343 } 7344 #ifdef ASSERT 7345 if (!_parallel) { 7346 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 7347 assert(_collector->overflow_list_is_empty(), 7348 "overflow list should be empty"); 7349 7350 } 7351 #endif // ASSERT 7352 return is_obj_array; 7353 } 7354 7355 MarkFromRootsClosure::MarkFromRootsClosure(CMSCollector* collector, 7356 MemRegion span, 7357 CMSBitMap* bitMap, CMSMarkStack* markStack, 7358 bool should_yield, bool verifying): 7359 _collector(collector), 7360 _span(span), 7361 _bitMap(bitMap), 7362 _mut(&collector->_modUnionTable), 7363 _markStack(markStack), 7364 _yield(should_yield), 7365 _skipBits(0) 7366 { 7367 assert(_markStack->isEmpty(), "stack should be empty"); 7368 _finger = _bitMap->startWord(); 7369 _threshold = _finger; 7370 assert(_collector->_restart_addr == NULL, "Sanity check"); 7371 assert(_span.contains(_finger), "Out of bounds _finger?"); 7372 DEBUG_ONLY(_verifying = verifying;) 7373 } 7374 7375 void MarkFromRootsClosure::reset(HeapWord* addr) { 7376 assert(_markStack->isEmpty(), "would cause duplicates on stack"); 7377 assert(_span.contains(addr), "Out of bounds _finger?"); 7378 _finger = addr; 7379 _threshold = (HeapWord*)round_to( 7380 (intptr_t)_finger, CardTableModRefBS::card_size); 7381 } 7382 7383 // Should revisit to see if this should be restructured for 7384 // greater efficiency. 7385 bool MarkFromRootsClosure::do_bit(size_t offset) { 7386 if (_skipBits > 0) { 7387 _skipBits--; 7388 return true; 7389 } 7390 // convert offset into a HeapWord* 7391 HeapWord* addr = _bitMap->startWord() + offset; 7392 assert(_bitMap->endWord() && addr < _bitMap->endWord(), 7393 "address out of range"); 7394 assert(_bitMap->isMarked(addr), "tautology"); 7395 if (_bitMap->isMarked(addr+1)) { 7396 // this is an allocated but not yet initialized object 7397 assert(_skipBits == 0, "tautology"); 7398 _skipBits = 2; // skip next two marked bits ("Printezis-marks") 7399 oop p = oop(addr); 7400 if (p->klass_or_null() == NULL) { 7401 DEBUG_ONLY(if (!_verifying) {) 7402 // We re-dirty the cards on which this object lies and increase 7403 // the _threshold so that we'll come back to scan this object 7404 // during the preclean or remark phase. (CMSCleanOnEnter) 7405 if (CMSCleanOnEnter) { 7406 size_t sz = _collector->block_size_using_printezis_bits(addr); 7407 HeapWord* end_card_addr = (HeapWord*)round_to( 7408 (intptr_t)(addr+sz), CardTableModRefBS::card_size); 7409 MemRegion redirty_range = MemRegion(addr, end_card_addr); 7410 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 7411 // Bump _threshold to end_card_addr; note that 7412 // _threshold cannot possibly exceed end_card_addr, anyhow. 7413 // This prevents future clearing of the card as the scan proceeds 7414 // to the right. 7415 assert(_threshold <= end_card_addr, 7416 "Because we are just scanning into this object"); 7417 if (_threshold < end_card_addr) { 7418 _threshold = end_card_addr; 7419 } 7420 if (p->klass_or_null() != NULL) { 7421 // Redirty the range of cards... 7422 _mut->mark_range(redirty_range); 7423 } // ...else the setting of klass will dirty the card anyway. 7424 } 7425 DEBUG_ONLY(}) 7426 return true; 7427 } 7428 } 7429 scanOopsInOop(addr); 7430 return true; 7431 } 7432 7433 // We take a break if we've been at this for a while, 7434 // so as to avoid monopolizing the locks involved. 7435 void MarkFromRootsClosure::do_yield_work() { 7436 // First give up the locks, then yield, then re-lock 7437 // We should probably use a constructor/destructor idiom to 7438 // do this unlock/lock or modify the MutexUnlocker class to 7439 // serve our purpose. XXX 7440 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 7441 "CMS thread should hold CMS token"); 7442 assert_lock_strong(_bitMap->lock()); 7443 _bitMap->lock()->unlock(); 7444 ConcurrentMarkSweepThread::desynchronize(true); 7445 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7446 _collector->stopTimer(); 7447 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 7448 if (PrintCMSStatistics != 0) { 7449 _collector->incrementYields(); 7450 } 7451 _collector->icms_wait(); 7452 7453 // See the comment in coordinator_yield() 7454 for (unsigned i = 0; i < CMSYieldSleepCount && 7455 ConcurrentMarkSweepThread::should_yield() && 7456 !CMSCollector::foregroundGCIsActive(); ++i) { 7457 os::sleep(Thread::current(), 1, false); 7458 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7459 } 7460 7461 ConcurrentMarkSweepThread::synchronize(true); 7462 _bitMap->lock()->lock_without_safepoint_check(); 7463 _collector->startTimer(); 7464 } 7465 7466 void MarkFromRootsClosure::scanOopsInOop(HeapWord* ptr) { 7467 assert(_bitMap->isMarked(ptr), "expected bit to be set"); 7468 assert(_markStack->isEmpty(), 7469 "should drain stack to limit stack usage"); 7470 // convert ptr to an oop preparatory to scanning 7471 oop obj = oop(ptr); 7472 // Ignore mark word in verification below, since we 7473 // may be running concurrent with mutators. 7474 assert(obj->is_oop(true), "should be an oop"); 7475 assert(_finger <= ptr, "_finger runneth ahead"); 7476 // advance the finger to right end of this object 7477 _finger = ptr + obj->size(); 7478 assert(_finger > ptr, "we just incremented it above"); 7479 // On large heaps, it may take us some time to get through 7480 // the marking phase (especially if running iCMS). During 7481 // this time it's possible that a lot of mutations have 7482 // accumulated in the card table and the mod union table -- 7483 // these mutation records are redundant until we have 7484 // actually traced into the corresponding card. 7485 // Here, we check whether advancing the finger would make 7486 // us cross into a new card, and if so clear corresponding 7487 // cards in the MUT (preclean them in the card-table in the 7488 // future). 7489 7490 DEBUG_ONLY(if (!_verifying) {) 7491 // The clean-on-enter optimization is disabled by default, 7492 // until we fix 6178663. 7493 if (CMSCleanOnEnter && (_finger > _threshold)) { 7494 // [_threshold, _finger) represents the interval 7495 // of cards to be cleared in MUT (or precleaned in card table). 7496 // The set of cards to be cleared is all those that overlap 7497 // with the interval [_threshold, _finger); note that 7498 // _threshold is always kept card-aligned but _finger isn't 7499 // always card-aligned. 7500 HeapWord* old_threshold = _threshold; 7501 assert(old_threshold == (HeapWord*)round_to( 7502 (intptr_t)old_threshold, CardTableModRefBS::card_size), 7503 "_threshold should always be card-aligned"); 7504 _threshold = (HeapWord*)round_to( 7505 (intptr_t)_finger, CardTableModRefBS::card_size); 7506 MemRegion mr(old_threshold, _threshold); 7507 assert(!mr.is_empty(), "Control point invariant"); 7508 assert(_span.contains(mr), "Should clear within span"); 7509 _mut->clear_range(mr); 7510 } 7511 DEBUG_ONLY(}) 7512 // Note: the finger doesn't advance while we drain 7513 // the stack below. 7514 PushOrMarkClosure pushOrMarkClosure(_collector, 7515 _span, _bitMap, _markStack, 7516 _finger, this); 7517 bool res = _markStack->push(obj); 7518 assert(res, "Empty non-zero size stack should have space for single push"); 7519 while (!_markStack->isEmpty()) { 7520 oop new_oop = _markStack->pop(); 7521 // Skip verifying header mark word below because we are 7522 // running concurrent with mutators. 7523 assert(new_oop->is_oop(true), "Oops! expected to pop an oop"); 7524 // now scan this oop's oops 7525 new_oop->oop_iterate(&pushOrMarkClosure); 7526 do_yield_check(); 7527 } 7528 assert(_markStack->isEmpty(), "tautology, emphasizing post-condition"); 7529 } 7530 7531 Par_MarkFromRootsClosure::Par_MarkFromRootsClosure(CMSConcMarkingTask* task, 7532 CMSCollector* collector, MemRegion span, 7533 CMSBitMap* bit_map, 7534 OopTaskQueue* work_queue, 7535 CMSMarkStack* overflow_stack, 7536 bool should_yield): 7537 _collector(collector), 7538 _whole_span(collector->_span), 7539 _span(span), 7540 _bit_map(bit_map), 7541 _mut(&collector->_modUnionTable), 7542 _work_queue(work_queue), 7543 _overflow_stack(overflow_stack), 7544 _yield(should_yield), 7545 _skip_bits(0), 7546 _task(task) 7547 { 7548 assert(_work_queue->size() == 0, "work_queue should be empty"); 7549 _finger = span.start(); 7550 _threshold = _finger; // XXX Defer clear-on-enter optimization for now 7551 assert(_span.contains(_finger), "Out of bounds _finger?"); 7552 } 7553 7554 // Should revisit to see if this should be restructured for 7555 // greater efficiency. 7556 bool Par_MarkFromRootsClosure::do_bit(size_t offset) { 7557 if (_skip_bits > 0) { 7558 _skip_bits--; 7559 return true; 7560 } 7561 // convert offset into a HeapWord* 7562 HeapWord* addr = _bit_map->startWord() + offset; 7563 assert(_bit_map->endWord() && addr < _bit_map->endWord(), 7564 "address out of range"); 7565 assert(_bit_map->isMarked(addr), "tautology"); 7566 if (_bit_map->isMarked(addr+1)) { 7567 // this is an allocated object that might not yet be initialized 7568 assert(_skip_bits == 0, "tautology"); 7569 _skip_bits = 2; // skip next two marked bits ("Printezis-marks") 7570 oop p = oop(addr); 7571 if (p->klass_or_null() == NULL) { 7572 // in the case of Clean-on-Enter optimization, redirty card 7573 // and avoid clearing card by increasing the threshold. 7574 return true; 7575 } 7576 } 7577 scan_oops_in_oop(addr); 7578 return true; 7579 } 7580 7581 void Par_MarkFromRootsClosure::scan_oops_in_oop(HeapWord* ptr) { 7582 assert(_bit_map->isMarked(ptr), "expected bit to be set"); 7583 // Should we assert that our work queue is empty or 7584 // below some drain limit? 7585 assert(_work_queue->size() == 0, 7586 "should drain stack to limit stack usage"); 7587 // convert ptr to an oop preparatory to scanning 7588 oop obj = oop(ptr); 7589 // Ignore mark word in verification below, since we 7590 // may be running concurrent with mutators. 7591 assert(obj->is_oop(true), "should be an oop"); 7592 assert(_finger <= ptr, "_finger runneth ahead"); 7593 // advance the finger to right end of this object 7594 _finger = ptr + obj->size(); 7595 assert(_finger > ptr, "we just incremented it above"); 7596 // On large heaps, it may take us some time to get through 7597 // the marking phase (especially if running iCMS). During 7598 // this time it's possible that a lot of mutations have 7599 // accumulated in the card table and the mod union table -- 7600 // these mutation records are redundant until we have 7601 // actually traced into the corresponding card. 7602 // Here, we check whether advancing the finger would make 7603 // us cross into a new card, and if so clear corresponding 7604 // cards in the MUT (preclean them in the card-table in the 7605 // future). 7606 7607 // The clean-on-enter optimization is disabled by default, 7608 // until we fix 6178663. 7609 if (CMSCleanOnEnter && (_finger > _threshold)) { 7610 // [_threshold, _finger) represents the interval 7611 // of cards to be cleared in MUT (or precleaned in card table). 7612 // The set of cards to be cleared is all those that overlap 7613 // with the interval [_threshold, _finger); note that 7614 // _threshold is always kept card-aligned but _finger isn't 7615 // always card-aligned. 7616 HeapWord* old_threshold = _threshold; 7617 assert(old_threshold == (HeapWord*)round_to( 7618 (intptr_t)old_threshold, CardTableModRefBS::card_size), 7619 "_threshold should always be card-aligned"); 7620 _threshold = (HeapWord*)round_to( 7621 (intptr_t)_finger, CardTableModRefBS::card_size); 7622 MemRegion mr(old_threshold, _threshold); 7623 assert(!mr.is_empty(), "Control point invariant"); 7624 assert(_span.contains(mr), "Should clear within span"); // _whole_span ?? 7625 _mut->clear_range(mr); 7626 } 7627 7628 // Note: the local finger doesn't advance while we drain 7629 // the stack below, but the global finger sure can and will. 7630 HeapWord** gfa = _task->global_finger_addr(); 7631 Par_PushOrMarkClosure pushOrMarkClosure(_collector, 7632 _span, _bit_map, 7633 _work_queue, 7634 _overflow_stack, 7635 _finger, 7636 gfa, this); 7637 bool res = _work_queue->push(obj); // overflow could occur here 7638 assert(res, "Will hold once we use workqueues"); 7639 while (true) { 7640 oop new_oop; 7641 if (!_work_queue->pop_local(new_oop)) { 7642 // We emptied our work_queue; check if there's stuff that can 7643 // be gotten from the overflow stack. 7644 if (CMSConcMarkingTask::get_work_from_overflow_stack( 7645 _overflow_stack, _work_queue)) { 7646 do_yield_check(); 7647 continue; 7648 } else { // done 7649 break; 7650 } 7651 } 7652 // Skip verifying header mark word below because we are 7653 // running concurrent with mutators. 7654 assert(new_oop->is_oop(true), "Oops! expected to pop an oop"); 7655 // now scan this oop's oops 7656 new_oop->oop_iterate(&pushOrMarkClosure); 7657 do_yield_check(); 7658 } 7659 assert(_work_queue->size() == 0, "tautology, emphasizing post-condition"); 7660 } 7661 7662 // Yield in response to a request from VM Thread or 7663 // from mutators. 7664 void Par_MarkFromRootsClosure::do_yield_work() { 7665 assert(_task != NULL, "sanity"); 7666 _task->yield(); 7667 } 7668 7669 // A variant of the above used for verifying CMS marking work. 7670 MarkFromRootsVerifyClosure::MarkFromRootsVerifyClosure(CMSCollector* collector, 7671 MemRegion span, 7672 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 7673 CMSMarkStack* mark_stack): 7674 _collector(collector), 7675 _span(span), 7676 _verification_bm(verification_bm), 7677 _cms_bm(cms_bm), 7678 _mark_stack(mark_stack), 7679 _pam_verify_closure(collector, span, verification_bm, cms_bm, 7680 mark_stack) 7681 { 7682 assert(_mark_stack->isEmpty(), "stack should be empty"); 7683 _finger = _verification_bm->startWord(); 7684 assert(_collector->_restart_addr == NULL, "Sanity check"); 7685 assert(_span.contains(_finger), "Out of bounds _finger?"); 7686 } 7687 7688 void MarkFromRootsVerifyClosure::reset(HeapWord* addr) { 7689 assert(_mark_stack->isEmpty(), "would cause duplicates on stack"); 7690 assert(_span.contains(addr), "Out of bounds _finger?"); 7691 _finger = addr; 7692 } 7693 7694 // Should revisit to see if this should be restructured for 7695 // greater efficiency. 7696 bool MarkFromRootsVerifyClosure::do_bit(size_t offset) { 7697 // convert offset into a HeapWord* 7698 HeapWord* addr = _verification_bm->startWord() + offset; 7699 assert(_verification_bm->endWord() && addr < _verification_bm->endWord(), 7700 "address out of range"); 7701 assert(_verification_bm->isMarked(addr), "tautology"); 7702 assert(_cms_bm->isMarked(addr), "tautology"); 7703 7704 assert(_mark_stack->isEmpty(), 7705 "should drain stack to limit stack usage"); 7706 // convert addr to an oop preparatory to scanning 7707 oop obj = oop(addr); 7708 assert(obj->is_oop(), "should be an oop"); 7709 assert(_finger <= addr, "_finger runneth ahead"); 7710 // advance the finger to right end of this object 7711 _finger = addr + obj->size(); 7712 assert(_finger > addr, "we just incremented it above"); 7713 // Note: the finger doesn't advance while we drain 7714 // the stack below. 7715 bool res = _mark_stack->push(obj); 7716 assert(res, "Empty non-zero size stack should have space for single push"); 7717 while (!_mark_stack->isEmpty()) { 7718 oop new_oop = _mark_stack->pop(); 7719 assert(new_oop->is_oop(), "Oops! expected to pop an oop"); 7720 // now scan this oop's oops 7721 new_oop->oop_iterate(&_pam_verify_closure); 7722 } 7723 assert(_mark_stack->isEmpty(), "tautology, emphasizing post-condition"); 7724 return true; 7725 } 7726 7727 PushAndMarkVerifyClosure::PushAndMarkVerifyClosure( 7728 CMSCollector* collector, MemRegion span, 7729 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 7730 CMSMarkStack* mark_stack): 7731 CMSOopClosure(collector->ref_processor()), 7732 _collector(collector), 7733 _span(span), 7734 _verification_bm(verification_bm), 7735 _cms_bm(cms_bm), 7736 _mark_stack(mark_stack) 7737 { } 7738 7739 void PushAndMarkVerifyClosure::do_oop(oop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 7740 void PushAndMarkVerifyClosure::do_oop(narrowOop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 7741 7742 // Upon stack overflow, we discard (part of) the stack, 7743 // remembering the least address amongst those discarded 7744 // in CMSCollector's _restart_address. 7745 void PushAndMarkVerifyClosure::handle_stack_overflow(HeapWord* lost) { 7746 // Remember the least grey address discarded 7747 HeapWord* ra = (HeapWord*)_mark_stack->least_value(lost); 7748 _collector->lower_restart_addr(ra); 7749 _mark_stack->reset(); // discard stack contents 7750 _mark_stack->expand(); // expand the stack if possible 7751 } 7752 7753 void PushAndMarkVerifyClosure::do_oop(oop obj) { 7754 assert(obj->is_oop_or_null(), "expected an oop or NULL"); 7755 HeapWord* addr = (HeapWord*)obj; 7756 if (_span.contains(addr) && !_verification_bm->isMarked(addr)) { 7757 // Oop lies in _span and isn't yet grey or black 7758 _verification_bm->mark(addr); // now grey 7759 if (!_cms_bm->isMarked(addr)) { 7760 oop(addr)->print(); 7761 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)", 7762 addr); 7763 fatal("... aborting"); 7764 } 7765 7766 if (!_mark_stack->push(obj)) { // stack overflow 7767 if (PrintCMSStatistics != 0) { 7768 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 7769 SIZE_FORMAT, _mark_stack->capacity()); 7770 } 7771 assert(_mark_stack->isFull(), "Else push should have succeeded"); 7772 handle_stack_overflow(addr); 7773 } 7774 // anything including and to the right of _finger 7775 // will be scanned as we iterate over the remainder of the 7776 // bit map 7777 } 7778 } 7779 7780 PushOrMarkClosure::PushOrMarkClosure(CMSCollector* collector, 7781 MemRegion span, 7782 CMSBitMap* bitMap, CMSMarkStack* markStack, 7783 HeapWord* finger, MarkFromRootsClosure* parent) : 7784 CMSOopClosure(collector->ref_processor()), 7785 _collector(collector), 7786 _span(span), 7787 _bitMap(bitMap), 7788 _markStack(markStack), 7789 _finger(finger), 7790 _parent(parent) 7791 { } 7792 7793 Par_PushOrMarkClosure::Par_PushOrMarkClosure(CMSCollector* collector, 7794 MemRegion span, 7795 CMSBitMap* bit_map, 7796 OopTaskQueue* work_queue, 7797 CMSMarkStack* overflow_stack, 7798 HeapWord* finger, 7799 HeapWord** global_finger_addr, 7800 Par_MarkFromRootsClosure* parent) : 7801 CMSOopClosure(collector->ref_processor()), 7802 _collector(collector), 7803 _whole_span(collector->_span), 7804 _span(span), 7805 _bit_map(bit_map), 7806 _work_queue(work_queue), 7807 _overflow_stack(overflow_stack), 7808 _finger(finger), 7809 _global_finger_addr(global_finger_addr), 7810 _parent(parent) 7811 { } 7812 7813 // Assumes thread-safe access by callers, who are 7814 // responsible for mutual exclusion. 7815 void CMSCollector::lower_restart_addr(HeapWord* low) { 7816 assert(_span.contains(low), "Out of bounds addr"); 7817 if (_restart_addr == NULL) { 7818 _restart_addr = low; 7819 } else { 7820 _restart_addr = MIN2(_restart_addr, low); 7821 } 7822 } 7823 7824 // Upon stack overflow, we discard (part of) the stack, 7825 // remembering the least address amongst those discarded 7826 // in CMSCollector's _restart_address. 7827 void PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 7828 // Remember the least grey address discarded 7829 HeapWord* ra = (HeapWord*)_markStack->least_value(lost); 7830 _collector->lower_restart_addr(ra); 7831 _markStack->reset(); // discard stack contents 7832 _markStack->expand(); // expand the stack if possible 7833 } 7834 7835 // Upon stack overflow, we discard (part of) the stack, 7836 // remembering the least address amongst those discarded 7837 // in CMSCollector's _restart_address. 7838 void Par_PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 7839 // We need to do this under a mutex to prevent other 7840 // workers from interfering with the work done below. 7841 MutexLockerEx ml(_overflow_stack->par_lock(), 7842 Mutex::_no_safepoint_check_flag); 7843 // Remember the least grey address discarded 7844 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 7845 _collector->lower_restart_addr(ra); 7846 _overflow_stack->reset(); // discard stack contents 7847 _overflow_stack->expand(); // expand the stack if possible 7848 } 7849 7850 void CMKlassClosure::do_klass(Klass* k) { 7851 assert(_oop_closure != NULL, "Not initialized?"); 7852 k->oops_do(_oop_closure); 7853 } 7854 7855 void PushOrMarkClosure::do_oop(oop obj) { 7856 // Ignore mark word because we are running concurrent with mutators. 7857 assert(obj->is_oop_or_null(true), "expected an oop or NULL"); 7858 HeapWord* addr = (HeapWord*)obj; 7859 if (_span.contains(addr) && !_bitMap->isMarked(addr)) { 7860 // Oop lies in _span and isn't yet grey or black 7861 _bitMap->mark(addr); // now grey 7862 if (addr < _finger) { 7863 // the bit map iteration has already either passed, or 7864 // sampled, this bit in the bit map; we'll need to 7865 // use the marking stack to scan this oop's oops. 7866 bool simulate_overflow = false; 7867 NOT_PRODUCT( 7868 if (CMSMarkStackOverflowALot && 7869 _collector->simulate_overflow()) { 7870 // simulate a stack overflow 7871 simulate_overflow = true; 7872 } 7873 ) 7874 if (simulate_overflow || !_markStack->push(obj)) { // stack overflow 7875 if (PrintCMSStatistics != 0) { 7876 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 7877 SIZE_FORMAT, _markStack->capacity()); 7878 } 7879 assert(simulate_overflow || _markStack->isFull(), "Else push should have succeeded"); 7880 handle_stack_overflow(addr); 7881 } 7882 } 7883 // anything including and to the right of _finger 7884 // will be scanned as we iterate over the remainder of the 7885 // bit map 7886 do_yield_check(); 7887 } 7888 } 7889 7890 void PushOrMarkClosure::do_oop(oop* p) { PushOrMarkClosure::do_oop_work(p); } 7891 void PushOrMarkClosure::do_oop(narrowOop* p) { PushOrMarkClosure::do_oop_work(p); } 7892 7893 void Par_PushOrMarkClosure::do_oop(oop obj) { 7894 // Ignore mark word because we are running concurrent with mutators. 7895 assert(obj->is_oop_or_null(true), "expected an oop or NULL"); 7896 HeapWord* addr = (HeapWord*)obj; 7897 if (_whole_span.contains(addr) && !_bit_map->isMarked(addr)) { 7898 // Oop lies in _span and isn't yet grey or black 7899 // We read the global_finger (volatile read) strictly after marking oop 7900 bool res = _bit_map->par_mark(addr); // now grey 7901 volatile HeapWord** gfa = (volatile HeapWord**)_global_finger_addr; 7902 // Should we push this marked oop on our stack? 7903 // -- if someone else marked it, nothing to do 7904 // -- if target oop is above global finger nothing to do 7905 // -- if target oop is in chunk and above local finger 7906 // then nothing to do 7907 // -- else push on work queue 7908 if ( !res // someone else marked it, they will deal with it 7909 || (addr >= *gfa) // will be scanned in a later task 7910 || (_span.contains(addr) && addr >= _finger)) { // later in this chunk 7911 return; 7912 } 7913 // the bit map iteration has already either passed, or 7914 // sampled, this bit in the bit map; we'll need to 7915 // use the marking stack to scan this oop's oops. 7916 bool simulate_overflow = false; 7917 NOT_PRODUCT( 7918 if (CMSMarkStackOverflowALot && 7919 _collector->simulate_overflow()) { 7920 // simulate a stack overflow 7921 simulate_overflow = true; 7922 } 7923 ) 7924 if (simulate_overflow || 7925 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) { 7926 // stack overflow 7927 if (PrintCMSStatistics != 0) { 7928 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 7929 SIZE_FORMAT, _overflow_stack->capacity()); 7930 } 7931 // We cannot assert that the overflow stack is full because 7932 // it may have been emptied since. 7933 assert(simulate_overflow || 7934 _work_queue->size() == _work_queue->max_elems(), 7935 "Else push should have succeeded"); 7936 handle_stack_overflow(addr); 7937 } 7938 do_yield_check(); 7939 } 7940 } 7941 7942 void Par_PushOrMarkClosure::do_oop(oop* p) { Par_PushOrMarkClosure::do_oop_work(p); } 7943 void Par_PushOrMarkClosure::do_oop(narrowOop* p) { Par_PushOrMarkClosure::do_oop_work(p); } 7944 7945 PushAndMarkClosure::PushAndMarkClosure(CMSCollector* collector, 7946 MemRegion span, 7947 ReferenceProcessor* rp, 7948 CMSBitMap* bit_map, 7949 CMSBitMap* mod_union_table, 7950 CMSMarkStack* mark_stack, 7951 bool concurrent_precleaning): 7952 CMSOopClosure(rp), 7953 _collector(collector), 7954 _span(span), 7955 _bit_map(bit_map), 7956 _mod_union_table(mod_union_table), 7957 _mark_stack(mark_stack), 7958 _concurrent_precleaning(concurrent_precleaning) 7959 { 7960 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 7961 } 7962 7963 // Grey object rescan during pre-cleaning and second checkpoint phases -- 7964 // the non-parallel version (the parallel version appears further below.) 7965 void PushAndMarkClosure::do_oop(oop obj) { 7966 // Ignore mark word verification. If during concurrent precleaning, 7967 // the object monitor may be locked. If during the checkpoint 7968 // phases, the object may already have been reached by a different 7969 // path and may be at the end of the global overflow list (so 7970 // the mark word may be NULL). 7971 assert(obj->is_oop_or_null(true /* ignore mark word */), 7972 "expected an oop or NULL"); 7973 HeapWord* addr = (HeapWord*)obj; 7974 // Check if oop points into the CMS generation 7975 // and is not marked 7976 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 7977 // a white object ... 7978 _bit_map->mark(addr); // ... now grey 7979 // push on the marking stack (grey set) 7980 bool simulate_overflow = false; 7981 NOT_PRODUCT( 7982 if (CMSMarkStackOverflowALot && 7983 _collector->simulate_overflow()) { 7984 // simulate a stack overflow 7985 simulate_overflow = true; 7986 } 7987 ) 7988 if (simulate_overflow || !_mark_stack->push(obj)) { 7989 if (_concurrent_precleaning) { 7990 // During precleaning we can just dirty the appropriate card(s) 7991 // in the mod union table, thus ensuring that the object remains 7992 // in the grey set and continue. In the case of object arrays 7993 // we need to dirty all of the cards that the object spans, 7994 // since the rescan of object arrays will be limited to the 7995 // dirty cards. 7996 // Note that no one can be interfering with us in this action 7997 // of dirtying the mod union table, so no locking or atomics 7998 // are required. 7999 if (obj->is_objArray()) { 8000 size_t sz = obj->size(); 8001 HeapWord* end_card_addr = (HeapWord*)round_to( 8002 (intptr_t)(addr+sz), CardTableModRefBS::card_size); 8003 MemRegion redirty_range = MemRegion(addr, end_card_addr); 8004 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 8005 _mod_union_table->mark_range(redirty_range); 8006 } else { 8007 _mod_union_table->mark(addr); 8008 } 8009 _collector->_ser_pmc_preclean_ovflw++; 8010 } else { 8011 // During the remark phase, we need to remember this oop 8012 // in the overflow list. 8013 _collector->push_on_overflow_list(obj); 8014 _collector->_ser_pmc_remark_ovflw++; 8015 } 8016 } 8017 } 8018 } 8019 8020 Par_PushAndMarkClosure::Par_PushAndMarkClosure(CMSCollector* collector, 8021 MemRegion span, 8022 ReferenceProcessor* rp, 8023 CMSBitMap* bit_map, 8024 OopTaskQueue* work_queue): 8025 CMSOopClosure(rp), 8026 _collector(collector), 8027 _span(span), 8028 _bit_map(bit_map), 8029 _work_queue(work_queue) 8030 { 8031 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 8032 } 8033 8034 void PushAndMarkClosure::do_oop(oop* p) { PushAndMarkClosure::do_oop_work(p); } 8035 void PushAndMarkClosure::do_oop(narrowOop* p) { PushAndMarkClosure::do_oop_work(p); } 8036 8037 // Grey object rescan during second checkpoint phase -- 8038 // the parallel version. 8039 void Par_PushAndMarkClosure::do_oop(oop obj) { 8040 // In the assert below, we ignore the mark word because 8041 // this oop may point to an already visited object that is 8042 // on the overflow stack (in which case the mark word has 8043 // been hijacked for chaining into the overflow stack -- 8044 // if this is the last object in the overflow stack then 8045 // its mark word will be NULL). Because this object may 8046 // have been subsequently popped off the global overflow 8047 // stack, and the mark word possibly restored to the prototypical 8048 // value, by the time we get to examined this failing assert in 8049 // the debugger, is_oop_or_null(false) may subsequently start 8050 // to hold. 8051 assert(obj->is_oop_or_null(true), 8052 "expected an oop or NULL"); 8053 HeapWord* addr = (HeapWord*)obj; 8054 // Check if oop points into the CMS generation 8055 // and is not marked 8056 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 8057 // a white object ... 8058 // If we manage to "claim" the object, by being the 8059 // first thread to mark it, then we push it on our 8060 // marking stack 8061 if (_bit_map->par_mark(addr)) { // ... now grey 8062 // push on work queue (grey set) 8063 bool simulate_overflow = false; 8064 NOT_PRODUCT( 8065 if (CMSMarkStackOverflowALot && 8066 _collector->par_simulate_overflow()) { 8067 // simulate a stack overflow 8068 simulate_overflow = true; 8069 } 8070 ) 8071 if (simulate_overflow || !_work_queue->push(obj)) { 8072 _collector->par_push_on_overflow_list(obj); 8073 _collector->_par_pmc_remark_ovflw++; // imprecise OK: no need to CAS 8074 } 8075 } // Else, some other thread got there first 8076 } 8077 } 8078 8079 void Par_PushAndMarkClosure::do_oop(oop* p) { Par_PushAndMarkClosure::do_oop_work(p); } 8080 void Par_PushAndMarkClosure::do_oop(narrowOop* p) { Par_PushAndMarkClosure::do_oop_work(p); } 8081 8082 void CMSPrecleanRefsYieldClosure::do_yield_work() { 8083 Mutex* bml = _collector->bitMapLock(); 8084 assert_lock_strong(bml); 8085 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 8086 "CMS thread should hold CMS token"); 8087 8088 bml->unlock(); 8089 ConcurrentMarkSweepThread::desynchronize(true); 8090 8091 ConcurrentMarkSweepThread::acknowledge_yield_request(); 8092 8093 _collector->stopTimer(); 8094 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 8095 if (PrintCMSStatistics != 0) { 8096 _collector->incrementYields(); 8097 } 8098 _collector->icms_wait(); 8099 8100 // See the comment in coordinator_yield() 8101 for (unsigned i = 0; i < CMSYieldSleepCount && 8102 ConcurrentMarkSweepThread::should_yield() && 8103 !CMSCollector::foregroundGCIsActive(); ++i) { 8104 os::sleep(Thread::current(), 1, false); 8105 ConcurrentMarkSweepThread::acknowledge_yield_request(); 8106 } 8107 8108 ConcurrentMarkSweepThread::synchronize(true); 8109 bml->lock(); 8110 8111 _collector->startTimer(); 8112 } 8113 8114 bool CMSPrecleanRefsYieldClosure::should_return() { 8115 if (ConcurrentMarkSweepThread::should_yield()) { 8116 do_yield_work(); 8117 } 8118 return _collector->foregroundGCIsActive(); 8119 } 8120 8121 void MarkFromDirtyCardsClosure::do_MemRegion(MemRegion mr) { 8122 assert(((size_t)mr.start())%CardTableModRefBS::card_size_in_words == 0, 8123 "mr should be aligned to start at a card boundary"); 8124 // We'd like to assert: 8125 // assert(mr.word_size()%CardTableModRefBS::card_size_in_words == 0, 8126 // "mr should be a range of cards"); 8127 // However, that would be too strong in one case -- the last 8128 // partition ends at _unallocated_block which, in general, can be 8129 // an arbitrary boundary, not necessarily card aligned. 8130 if (PrintCMSStatistics != 0) { 8131 _num_dirty_cards += 8132 mr.word_size()/CardTableModRefBS::card_size_in_words; 8133 } 8134 _space->object_iterate_mem(mr, &_scan_cl); 8135 } 8136 8137 SweepClosure::SweepClosure(CMSCollector* collector, 8138 ConcurrentMarkSweepGeneration* g, 8139 CMSBitMap* bitMap, bool should_yield) : 8140 _collector(collector), 8141 _g(g), 8142 _sp(g->cmsSpace()), 8143 _limit(_sp->sweep_limit()), 8144 _freelistLock(_sp->freelistLock()), 8145 _bitMap(bitMap), 8146 _yield(should_yield), 8147 _inFreeRange(false), // No free range at beginning of sweep 8148 _freeRangeInFreeLists(false), // No free range at beginning of sweep 8149 _lastFreeRangeCoalesced(false), 8150 _freeFinger(g->used_region().start()) 8151 { 8152 NOT_PRODUCT( 8153 _numObjectsFreed = 0; 8154 _numWordsFreed = 0; 8155 _numObjectsLive = 0; 8156 _numWordsLive = 0; 8157 _numObjectsAlreadyFree = 0; 8158 _numWordsAlreadyFree = 0; 8159 _last_fc = NULL; 8160 8161 _sp->initializeIndexedFreeListArrayReturnedBytes(); 8162 _sp->dictionary()->initialize_dict_returned_bytes(); 8163 ) 8164 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 8165 "sweep _limit out of bounds"); 8166 if (CMSTraceSweeper) { 8167 gclog_or_tty->print_cr("\n====================\nStarting new sweep with limit " PTR_FORMAT, 8168 _limit); 8169 } 8170 } 8171 8172 void SweepClosure::print_on(outputStream* st) const { 8173 tty->print_cr("_sp = [" PTR_FORMAT "," PTR_FORMAT ")", 8174 _sp->bottom(), _sp->end()); 8175 tty->print_cr("_limit = " PTR_FORMAT, _limit); 8176 tty->print_cr("_freeFinger = " PTR_FORMAT, _freeFinger); 8177 NOT_PRODUCT(tty->print_cr("_last_fc = " PTR_FORMAT, _last_fc);) 8178 tty->print_cr("_inFreeRange = %d, _freeRangeInFreeLists = %d, _lastFreeRangeCoalesced = %d", 8179 _inFreeRange, _freeRangeInFreeLists, _lastFreeRangeCoalesced); 8180 } 8181 8182 #ifndef PRODUCT 8183 // Assertion checking only: no useful work in product mode -- 8184 // however, if any of the flags below become product flags, 8185 // you may need to review this code to see if it needs to be 8186 // enabled in product mode. 8187 SweepClosure::~SweepClosure() { 8188 assert_lock_strong(_freelistLock); 8189 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 8190 "sweep _limit out of bounds"); 8191 if (inFreeRange()) { 8192 warning("inFreeRange() should have been reset; dumping state of SweepClosure"); 8193 print(); 8194 ShouldNotReachHere(); 8195 } 8196 if (Verbose && PrintGC) { 8197 gclog_or_tty->print("Collected "SIZE_FORMAT" objects, " SIZE_FORMAT " bytes", 8198 _numObjectsFreed, _numWordsFreed*sizeof(HeapWord)); 8199 gclog_or_tty->print_cr("\nLive "SIZE_FORMAT" objects, " 8200 SIZE_FORMAT" bytes " 8201 "Already free "SIZE_FORMAT" objects, "SIZE_FORMAT" bytes", 8202 _numObjectsLive, _numWordsLive*sizeof(HeapWord), 8203 _numObjectsAlreadyFree, _numWordsAlreadyFree*sizeof(HeapWord)); 8204 size_t totalBytes = (_numWordsFreed + _numWordsLive + _numWordsAlreadyFree) 8205 * sizeof(HeapWord); 8206 gclog_or_tty->print_cr("Total sweep: "SIZE_FORMAT" bytes", totalBytes); 8207 8208 if (PrintCMSStatistics && CMSVerifyReturnedBytes) { 8209 size_t indexListReturnedBytes = _sp->sumIndexedFreeListArrayReturnedBytes(); 8210 size_t dict_returned_bytes = _sp->dictionary()->sum_dict_returned_bytes(); 8211 size_t returned_bytes = indexListReturnedBytes + dict_returned_bytes; 8212 gclog_or_tty->print("Returned "SIZE_FORMAT" bytes", returned_bytes); 8213 gclog_or_tty->print(" Indexed List Returned "SIZE_FORMAT" bytes", 8214 indexListReturnedBytes); 8215 gclog_or_tty->print_cr(" Dictionary Returned "SIZE_FORMAT" bytes", 8216 dict_returned_bytes); 8217 } 8218 } 8219 if (CMSTraceSweeper) { 8220 gclog_or_tty->print_cr("end of sweep with _limit = " PTR_FORMAT "\n================", 8221 _limit); 8222 } 8223 } 8224 #endif // PRODUCT 8225 8226 void SweepClosure::initialize_free_range(HeapWord* freeFinger, 8227 bool freeRangeInFreeLists) { 8228 if (CMSTraceSweeper) { 8229 gclog_or_tty->print("---- Start free range at 0x%x with free block (%d)\n", 8230 freeFinger, freeRangeInFreeLists); 8231 } 8232 assert(!inFreeRange(), "Trampling existing free range"); 8233 set_inFreeRange(true); 8234 set_lastFreeRangeCoalesced(false); 8235 8236 set_freeFinger(freeFinger); 8237 set_freeRangeInFreeLists(freeRangeInFreeLists); 8238 if (CMSTestInFreeList) { 8239 if (freeRangeInFreeLists) { 8240 FreeChunk* fc = (FreeChunk*) freeFinger; 8241 assert(fc->is_free(), "A chunk on the free list should be free."); 8242 assert(fc->size() > 0, "Free range should have a size"); 8243 assert(_sp->verify_chunk_in_free_list(fc), "Chunk is not in free lists"); 8244 } 8245 } 8246 } 8247 8248 // Note that the sweeper runs concurrently with mutators. Thus, 8249 // it is possible for direct allocation in this generation to happen 8250 // in the middle of the sweep. Note that the sweeper also coalesces 8251 // contiguous free blocks. Thus, unless the sweeper and the allocator 8252 // synchronize appropriately freshly allocated blocks may get swept up. 8253 // This is accomplished by the sweeper locking the free lists while 8254 // it is sweeping. Thus blocks that are determined to be free are 8255 // indeed free. There is however one additional complication: 8256 // blocks that have been allocated since the final checkpoint and 8257 // mark, will not have been marked and so would be treated as 8258 // unreachable and swept up. To prevent this, the allocator marks 8259 // the bit map when allocating during the sweep phase. This leads, 8260 // however, to a further complication -- objects may have been allocated 8261 // but not yet initialized -- in the sense that the header isn't yet 8262 // installed. The sweeper can not then determine the size of the block 8263 // in order to skip over it. To deal with this case, we use a technique 8264 // (due to Printezis) to encode such uninitialized block sizes in the 8265 // bit map. Since the bit map uses a bit per every HeapWord, but the 8266 // CMS generation has a minimum object size of 3 HeapWords, it follows 8267 // that "normal marks" won't be adjacent in the bit map (there will 8268 // always be at least two 0 bits between successive 1 bits). We make use 8269 // of these "unused" bits to represent uninitialized blocks -- the bit 8270 // corresponding to the start of the uninitialized object and the next 8271 // bit are both set. Finally, a 1 bit marks the end of the object that 8272 // started with the two consecutive 1 bits to indicate its potentially 8273 // uninitialized state. 8274 8275 size_t SweepClosure::do_blk_careful(HeapWord* addr) { 8276 FreeChunk* fc = (FreeChunk*)addr; 8277 size_t res; 8278 8279 // Check if we are done sweeping. Below we check "addr >= _limit" rather 8280 // than "addr == _limit" because although _limit was a block boundary when 8281 // we started the sweep, it may no longer be one because heap expansion 8282 // may have caused us to coalesce the block ending at the address _limit 8283 // with a newly expanded chunk (this happens when _limit was set to the 8284 // previous _end of the space), so we may have stepped past _limit: 8285 // see the following Zeno-like trail of CRs 6977970, 7008136, 7042740. 8286 if (addr >= _limit) { // we have swept up to or past the limit: finish up 8287 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 8288 "sweep _limit out of bounds"); 8289 assert(addr < _sp->end(), "addr out of bounds"); 8290 // Flush any free range we might be holding as a single 8291 // coalesced chunk to the appropriate free list. 8292 if (inFreeRange()) { 8293 assert(freeFinger() >= _sp->bottom() && freeFinger() < _limit, 8294 err_msg("freeFinger() " PTR_FORMAT" is out-of-bounds", freeFinger())); 8295 flush_cur_free_chunk(freeFinger(), 8296 pointer_delta(addr, freeFinger())); 8297 if (CMSTraceSweeper) { 8298 gclog_or_tty->print("Sweep: last chunk: "); 8299 gclog_or_tty->print("put_free_blk 0x%x ("SIZE_FORMAT") " 8300 "[coalesced:"SIZE_FORMAT"]\n", 8301 freeFinger(), pointer_delta(addr, freeFinger()), 8302 lastFreeRangeCoalesced()); 8303 } 8304 } 8305 8306 // help the iterator loop finish 8307 return pointer_delta(_sp->end(), addr); 8308 } 8309 8310 assert(addr < _limit, "sweep invariant"); 8311 // check if we should yield 8312 do_yield_check(addr); 8313 if (fc->is_free()) { 8314 // Chunk that is already free 8315 res = fc->size(); 8316 do_already_free_chunk(fc); 8317 debug_only(_sp->verifyFreeLists()); 8318 // If we flush the chunk at hand in lookahead_and_flush() 8319 // and it's coalesced with a preceding chunk, then the 8320 // process of "mangling" the payload of the coalesced block 8321 // will cause erasure of the size information from the 8322 // (erstwhile) header of all the coalesced blocks but the 8323 // first, so the first disjunct in the assert will not hold 8324 // in that specific case (in which case the second disjunct 8325 // will hold). 8326 assert(res == fc->size() || ((HeapWord*)fc) + res >= _limit, 8327 "Otherwise the size info doesn't change at this step"); 8328 NOT_PRODUCT( 8329 _numObjectsAlreadyFree++; 8330 _numWordsAlreadyFree += res; 8331 ) 8332 NOT_PRODUCT(_last_fc = fc;) 8333 } else if (!_bitMap->isMarked(addr)) { 8334 // Chunk is fresh garbage 8335 res = do_garbage_chunk(fc); 8336 debug_only(_sp->verifyFreeLists()); 8337 NOT_PRODUCT( 8338 _numObjectsFreed++; 8339 _numWordsFreed += res; 8340 ) 8341 } else { 8342 // Chunk that is alive. 8343 res = do_live_chunk(fc); 8344 debug_only(_sp->verifyFreeLists()); 8345 NOT_PRODUCT( 8346 _numObjectsLive++; 8347 _numWordsLive += res; 8348 ) 8349 } 8350 return res; 8351 } 8352 8353 // For the smart allocation, record following 8354 // split deaths - a free chunk is removed from its free list because 8355 // it is being split into two or more chunks. 8356 // split birth - a free chunk is being added to its free list because 8357 // a larger free chunk has been split and resulted in this free chunk. 8358 // coal death - a free chunk is being removed from its free list because 8359 // it is being coalesced into a large free chunk. 8360 // coal birth - a free chunk is being added to its free list because 8361 // it was created when two or more free chunks where coalesced into 8362 // this free chunk. 8363 // 8364 // These statistics are used to determine the desired number of free 8365 // chunks of a given size. The desired number is chosen to be relative 8366 // to the end of a CMS sweep. The desired number at the end of a sweep 8367 // is the 8368 // count-at-end-of-previous-sweep (an amount that was enough) 8369 // - count-at-beginning-of-current-sweep (the excess) 8370 // + split-births (gains in this size during interval) 8371 // - split-deaths (demands on this size during interval) 8372 // where the interval is from the end of one sweep to the end of the 8373 // next. 8374 // 8375 // When sweeping the sweeper maintains an accumulated chunk which is 8376 // the chunk that is made up of chunks that have been coalesced. That 8377 // will be termed the left-hand chunk. A new chunk of garbage that 8378 // is being considered for coalescing will be referred to as the 8379 // right-hand chunk. 8380 // 8381 // When making a decision on whether to coalesce a right-hand chunk with 8382 // the current left-hand chunk, the current count vs. the desired count 8383 // of the left-hand chunk is considered. Also if the right-hand chunk 8384 // is near the large chunk at the end of the heap (see 8385 // ConcurrentMarkSweepGeneration::isNearLargestChunk()), then the 8386 // left-hand chunk is coalesced. 8387 // 8388 // When making a decision about whether to split a chunk, the desired count 8389 // vs. the current count of the candidate to be split is also considered. 8390 // If the candidate is underpopulated (currently fewer chunks than desired) 8391 // a chunk of an overpopulated (currently more chunks than desired) size may 8392 // be chosen. The "hint" associated with a free list, if non-null, points 8393 // to a free list which may be overpopulated. 8394 // 8395 8396 void SweepClosure::do_already_free_chunk(FreeChunk* fc) { 8397 const size_t size = fc->size(); 8398 // Chunks that cannot be coalesced are not in the 8399 // free lists. 8400 if (CMSTestInFreeList && !fc->cantCoalesce()) { 8401 assert(_sp->verify_chunk_in_free_list(fc), 8402 "free chunk should be in free lists"); 8403 } 8404 // a chunk that is already free, should not have been 8405 // marked in the bit map 8406 HeapWord* const addr = (HeapWord*) fc; 8407 assert(!_bitMap->isMarked(addr), "free chunk should be unmarked"); 8408 // Verify that the bit map has no bits marked between 8409 // addr and purported end of this block. 8410 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 8411 8412 // Some chunks cannot be coalesced under any circumstances. 8413 // See the definition of cantCoalesce(). 8414 if (!fc->cantCoalesce()) { 8415 // This chunk can potentially be coalesced. 8416 if (_sp->adaptive_freelists()) { 8417 // All the work is done in 8418 do_post_free_or_garbage_chunk(fc, size); 8419 } else { // Not adaptive free lists 8420 // this is a free chunk that can potentially be coalesced by the sweeper; 8421 if (!inFreeRange()) { 8422 // if the next chunk is a free block that can't be coalesced 8423 // it doesn't make sense to remove this chunk from the free lists 8424 FreeChunk* nextChunk = (FreeChunk*)(addr + size); 8425 assert((HeapWord*)nextChunk <= _sp->end(), "Chunk size out of bounds?"); 8426 if ((HeapWord*)nextChunk < _sp->end() && // There is another free chunk to the right ... 8427 nextChunk->is_free() && // ... which is free... 8428 nextChunk->cantCoalesce()) { // ... but can't be coalesced 8429 // nothing to do 8430 } else { 8431 // Potentially the start of a new free range: 8432 // Don't eagerly remove it from the free lists. 8433 // No need to remove it if it will just be put 8434 // back again. (Also from a pragmatic point of view 8435 // if it is a free block in a region that is beyond 8436 // any allocated blocks, an assertion will fail) 8437 // Remember the start of a free run. 8438 initialize_free_range(addr, true); 8439 // end - can coalesce with next chunk 8440 } 8441 } else { 8442 // the midst of a free range, we are coalescing 8443 print_free_block_coalesced(fc); 8444 if (CMSTraceSweeper) { 8445 gclog_or_tty->print(" -- pick up free block 0x%x (%d)\n", fc, size); 8446 } 8447 // remove it from the free lists 8448 _sp->removeFreeChunkFromFreeLists(fc); 8449 set_lastFreeRangeCoalesced(true); 8450 // If the chunk is being coalesced and the current free range is 8451 // in the free lists, remove the current free range so that it 8452 // will be returned to the free lists in its entirety - all 8453 // the coalesced pieces included. 8454 if (freeRangeInFreeLists()) { 8455 FreeChunk* ffc = (FreeChunk*) freeFinger(); 8456 assert(ffc->size() == pointer_delta(addr, freeFinger()), 8457 "Size of free range is inconsistent with chunk size."); 8458 if (CMSTestInFreeList) { 8459 assert(_sp->verify_chunk_in_free_list(ffc), 8460 "free range is not in free lists"); 8461 } 8462 _sp->removeFreeChunkFromFreeLists(ffc); 8463 set_freeRangeInFreeLists(false); 8464 } 8465 } 8466 } 8467 // Note that if the chunk is not coalescable (the else arm 8468 // below), we unconditionally flush, without needing to do 8469 // a "lookahead," as we do below. 8470 if (inFreeRange()) lookahead_and_flush(fc, size); 8471 } else { 8472 // Code path common to both original and adaptive free lists. 8473 8474 // cant coalesce with previous block; this should be treated 8475 // as the end of a free run if any 8476 if (inFreeRange()) { 8477 // we kicked some butt; time to pick up the garbage 8478 assert(freeFinger() < addr, "freeFinger points too high"); 8479 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 8480 } 8481 // else, nothing to do, just continue 8482 } 8483 } 8484 8485 size_t SweepClosure::do_garbage_chunk(FreeChunk* fc) { 8486 // This is a chunk of garbage. It is not in any free list. 8487 // Add it to a free list or let it possibly be coalesced into 8488 // a larger chunk. 8489 HeapWord* const addr = (HeapWord*) fc; 8490 const size_t size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 8491 8492 if (_sp->adaptive_freelists()) { 8493 // Verify that the bit map has no bits marked between 8494 // addr and purported end of just dead object. 8495 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 8496 8497 do_post_free_or_garbage_chunk(fc, size); 8498 } else { 8499 if (!inFreeRange()) { 8500 // start of a new free range 8501 assert(size > 0, "A free range should have a size"); 8502 initialize_free_range(addr, false); 8503 } else { 8504 // this will be swept up when we hit the end of the 8505 // free range 8506 if (CMSTraceSweeper) { 8507 gclog_or_tty->print(" -- pick up garbage 0x%x (%d) \n", fc, size); 8508 } 8509 // If the chunk is being coalesced and the current free range is 8510 // in the free lists, remove the current free range so that it 8511 // will be returned to the free lists in its entirety - all 8512 // the coalesced pieces included. 8513 if (freeRangeInFreeLists()) { 8514 FreeChunk* ffc = (FreeChunk*)freeFinger(); 8515 assert(ffc->size() == pointer_delta(addr, freeFinger()), 8516 "Size of free range is inconsistent with chunk size."); 8517 if (CMSTestInFreeList) { 8518 assert(_sp->verify_chunk_in_free_list(ffc), 8519 "free range is not in free lists"); 8520 } 8521 _sp->removeFreeChunkFromFreeLists(ffc); 8522 set_freeRangeInFreeLists(false); 8523 } 8524 set_lastFreeRangeCoalesced(true); 8525 } 8526 // this will be swept up when we hit the end of the free range 8527 8528 // Verify that the bit map has no bits marked between 8529 // addr and purported end of just dead object. 8530 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 8531 } 8532 assert(_limit >= addr + size, 8533 "A freshly garbage chunk can't possibly straddle over _limit"); 8534 if (inFreeRange()) lookahead_and_flush(fc, size); 8535 return size; 8536 } 8537 8538 size_t SweepClosure::do_live_chunk(FreeChunk* fc) { 8539 HeapWord* addr = (HeapWord*) fc; 8540 // The sweeper has just found a live object. Return any accumulated 8541 // left hand chunk to the free lists. 8542 if (inFreeRange()) { 8543 assert(freeFinger() < addr, "freeFinger points too high"); 8544 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 8545 } 8546 8547 // This object is live: we'd normally expect this to be 8548 // an oop, and like to assert the following: 8549 // assert(oop(addr)->is_oop(), "live block should be an oop"); 8550 // However, as we commented above, this may be an object whose 8551 // header hasn't yet been initialized. 8552 size_t size; 8553 assert(_bitMap->isMarked(addr), "Tautology for this control point"); 8554 if (_bitMap->isMarked(addr + 1)) { 8555 // Determine the size from the bit map, rather than trying to 8556 // compute it from the object header. 8557 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 8558 size = pointer_delta(nextOneAddr + 1, addr); 8559 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 8560 "alignment problem"); 8561 8562 #ifdef ASSERT 8563 if (oop(addr)->klass_or_null() != NULL) { 8564 // Ignore mark word because we are running concurrent with mutators 8565 assert(oop(addr)->is_oop(true), "live block should be an oop"); 8566 assert(size == 8567 CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()), 8568 "P-mark and computed size do not agree"); 8569 } 8570 #endif 8571 8572 } else { 8573 // This should be an initialized object that's alive. 8574 assert(oop(addr)->klass_or_null() != NULL, 8575 "Should be an initialized object"); 8576 // Ignore mark word because we are running concurrent with mutators 8577 assert(oop(addr)->is_oop(true), "live block should be an oop"); 8578 // Verify that the bit map has no bits marked between 8579 // addr and purported end of this block. 8580 size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 8581 assert(size >= 3, "Necessary for Printezis marks to work"); 8582 assert(!_bitMap->isMarked(addr+1), "Tautology for this control point"); 8583 DEBUG_ONLY(_bitMap->verifyNoOneBitsInRange(addr+2, addr+size);) 8584 } 8585 return size; 8586 } 8587 8588 void SweepClosure::do_post_free_or_garbage_chunk(FreeChunk* fc, 8589 size_t chunkSize) { 8590 // do_post_free_or_garbage_chunk() should only be called in the case 8591 // of the adaptive free list allocator. 8592 const bool fcInFreeLists = fc->is_free(); 8593 assert(_sp->adaptive_freelists(), "Should only be used in this case."); 8594 assert((HeapWord*)fc <= _limit, "sweep invariant"); 8595 if (CMSTestInFreeList && fcInFreeLists) { 8596 assert(_sp->verify_chunk_in_free_list(fc), "free chunk is not in free lists"); 8597 } 8598 8599 if (CMSTraceSweeper) { 8600 gclog_or_tty->print_cr(" -- pick up another chunk at 0x%x (%d)", fc, chunkSize); 8601 } 8602 8603 HeapWord* const fc_addr = (HeapWord*) fc; 8604 8605 bool coalesce; 8606 const size_t left = pointer_delta(fc_addr, freeFinger()); 8607 const size_t right = chunkSize; 8608 switch (FLSCoalescePolicy) { 8609 // numeric value forms a coalition aggressiveness metric 8610 case 0: { // never coalesce 8611 coalesce = false; 8612 break; 8613 } 8614 case 1: { // coalesce if left & right chunks on overpopulated lists 8615 coalesce = _sp->coalOverPopulated(left) && 8616 _sp->coalOverPopulated(right); 8617 break; 8618 } 8619 case 2: { // coalesce if left chunk on overpopulated list (default) 8620 coalesce = _sp->coalOverPopulated(left); 8621 break; 8622 } 8623 case 3: { // coalesce if left OR right chunk on overpopulated list 8624 coalesce = _sp->coalOverPopulated(left) || 8625 _sp->coalOverPopulated(right); 8626 break; 8627 } 8628 case 4: { // always coalesce 8629 coalesce = true; 8630 break; 8631 } 8632 default: 8633 ShouldNotReachHere(); 8634 } 8635 8636 // Should the current free range be coalesced? 8637 // If the chunk is in a free range and either we decided to coalesce above 8638 // or the chunk is near the large block at the end of the heap 8639 // (isNearLargestChunk() returns true), then coalesce this chunk. 8640 const bool doCoalesce = inFreeRange() 8641 && (coalesce || _g->isNearLargestChunk(fc_addr)); 8642 if (doCoalesce) { 8643 // Coalesce the current free range on the left with the new 8644 // chunk on the right. If either is on a free list, 8645 // it must be removed from the list and stashed in the closure. 8646 if (freeRangeInFreeLists()) { 8647 FreeChunk* const ffc = (FreeChunk*)freeFinger(); 8648 assert(ffc->size() == pointer_delta(fc_addr, freeFinger()), 8649 "Size of free range is inconsistent with chunk size."); 8650 if (CMSTestInFreeList) { 8651 assert(_sp->verify_chunk_in_free_list(ffc), 8652 "Chunk is not in free lists"); 8653 } 8654 _sp->coalDeath(ffc->size()); 8655 _sp->removeFreeChunkFromFreeLists(ffc); 8656 set_freeRangeInFreeLists(false); 8657 } 8658 if (fcInFreeLists) { 8659 _sp->coalDeath(chunkSize); 8660 assert(fc->size() == chunkSize, 8661 "The chunk has the wrong size or is not in the free lists"); 8662 _sp->removeFreeChunkFromFreeLists(fc); 8663 } 8664 set_lastFreeRangeCoalesced(true); 8665 print_free_block_coalesced(fc); 8666 } else { // not in a free range and/or should not coalesce 8667 // Return the current free range and start a new one. 8668 if (inFreeRange()) { 8669 // In a free range but cannot coalesce with the right hand chunk. 8670 // Put the current free range into the free lists. 8671 flush_cur_free_chunk(freeFinger(), 8672 pointer_delta(fc_addr, freeFinger())); 8673 } 8674 // Set up for new free range. Pass along whether the right hand 8675 // chunk is in the free lists. 8676 initialize_free_range((HeapWord*)fc, fcInFreeLists); 8677 } 8678 } 8679 8680 // Lookahead flush: 8681 // If we are tracking a free range, and this is the last chunk that 8682 // we'll look at because its end crosses past _limit, we'll preemptively 8683 // flush it along with any free range we may be holding on to. Note that 8684 // this can be the case only for an already free or freshly garbage 8685 // chunk. If this block is an object, it can never straddle 8686 // over _limit. The "straddling" occurs when _limit is set at 8687 // the previous end of the space when this cycle started, and 8688 // a subsequent heap expansion caused the previously co-terminal 8689 // free block to be coalesced with the newly expanded portion, 8690 // thus rendering _limit a non-block-boundary making it dangerous 8691 // for the sweeper to step over and examine. 8692 void SweepClosure::lookahead_and_flush(FreeChunk* fc, size_t chunk_size) { 8693 assert(inFreeRange(), "Should only be called if currently in a free range."); 8694 HeapWord* const eob = ((HeapWord*)fc) + chunk_size; 8695 assert(_sp->used_region().contains(eob - 1), 8696 err_msg("eob = " PTR_FORMAT " eob-1 = " PTR_FORMAT " _limit = " PTR_FORMAT 8697 " out of bounds wrt _sp = [" PTR_FORMAT "," PTR_FORMAT ")" 8698 " when examining fc = " PTR_FORMAT "(" SIZE_FORMAT ")", 8699 eob, eob-1, _limit, _sp->bottom(), _sp->end(), fc, chunk_size)); 8700 if (eob >= _limit) { 8701 assert(eob == _limit || fc->is_free(), "Only a free chunk should allow us to cross over the limit"); 8702 if (CMSTraceSweeper) { 8703 gclog_or_tty->print_cr("_limit " PTR_FORMAT " reached or crossed by block " 8704 "[" PTR_FORMAT "," PTR_FORMAT ") in space " 8705 "[" PTR_FORMAT "," PTR_FORMAT ")", 8706 _limit, fc, eob, _sp->bottom(), _sp->end()); 8707 } 8708 // Return the storage we are tracking back into the free lists. 8709 if (CMSTraceSweeper) { 8710 gclog_or_tty->print_cr("Flushing ... "); 8711 } 8712 assert(freeFinger() < eob, "Error"); 8713 flush_cur_free_chunk( freeFinger(), pointer_delta(eob, freeFinger())); 8714 } 8715 } 8716 8717 void SweepClosure::flush_cur_free_chunk(HeapWord* chunk, size_t size) { 8718 assert(inFreeRange(), "Should only be called if currently in a free range."); 8719 assert(size > 0, 8720 "A zero sized chunk cannot be added to the free lists."); 8721 if (!freeRangeInFreeLists()) { 8722 if (CMSTestInFreeList) { 8723 FreeChunk* fc = (FreeChunk*) chunk; 8724 fc->set_size(size); 8725 assert(!_sp->verify_chunk_in_free_list(fc), 8726 "chunk should not be in free lists yet"); 8727 } 8728 if (CMSTraceSweeper) { 8729 gclog_or_tty->print_cr(" -- add free block 0x%x (%d) to free lists", 8730 chunk, size); 8731 } 8732 // A new free range is going to be starting. The current 8733 // free range has not been added to the free lists yet or 8734 // was removed so add it back. 8735 // If the current free range was coalesced, then the death 8736 // of the free range was recorded. Record a birth now. 8737 if (lastFreeRangeCoalesced()) { 8738 _sp->coalBirth(size); 8739 } 8740 _sp->addChunkAndRepairOffsetTable(chunk, size, 8741 lastFreeRangeCoalesced()); 8742 } else if (CMSTraceSweeper) { 8743 gclog_or_tty->print_cr("Already in free list: nothing to flush"); 8744 } 8745 set_inFreeRange(false); 8746 set_freeRangeInFreeLists(false); 8747 } 8748 8749 // We take a break if we've been at this for a while, 8750 // so as to avoid monopolizing the locks involved. 8751 void SweepClosure::do_yield_work(HeapWord* addr) { 8752 // Return current free chunk being used for coalescing (if any) 8753 // to the appropriate freelist. After yielding, the next 8754 // free block encountered will start a coalescing range of 8755 // free blocks. If the next free block is adjacent to the 8756 // chunk just flushed, they will need to wait for the next 8757 // sweep to be coalesced. 8758 if (inFreeRange()) { 8759 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 8760 } 8761 8762 // First give up the locks, then yield, then re-lock. 8763 // We should probably use a constructor/destructor idiom to 8764 // do this unlock/lock or modify the MutexUnlocker class to 8765 // serve our purpose. XXX 8766 assert_lock_strong(_bitMap->lock()); 8767 assert_lock_strong(_freelistLock); 8768 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 8769 "CMS thread should hold CMS token"); 8770 _bitMap->lock()->unlock(); 8771 _freelistLock->unlock(); 8772 ConcurrentMarkSweepThread::desynchronize(true); 8773 ConcurrentMarkSweepThread::acknowledge_yield_request(); 8774 _collector->stopTimer(); 8775 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 8776 if (PrintCMSStatistics != 0) { 8777 _collector->incrementYields(); 8778 } 8779 _collector->icms_wait(); 8780 8781 // See the comment in coordinator_yield() 8782 for (unsigned i = 0; i < CMSYieldSleepCount && 8783 ConcurrentMarkSweepThread::should_yield() && 8784 !CMSCollector::foregroundGCIsActive(); ++i) { 8785 os::sleep(Thread::current(), 1, false); 8786 ConcurrentMarkSweepThread::acknowledge_yield_request(); 8787 } 8788 8789 ConcurrentMarkSweepThread::synchronize(true); 8790 _freelistLock->lock(); 8791 _bitMap->lock()->lock_without_safepoint_check(); 8792 _collector->startTimer(); 8793 } 8794 8795 #ifndef PRODUCT 8796 // This is actually very useful in a product build if it can 8797 // be called from the debugger. Compile it into the product 8798 // as needed. 8799 bool debug_verify_chunk_in_free_list(FreeChunk* fc) { 8800 return debug_cms_space->verify_chunk_in_free_list(fc); 8801 } 8802 #endif 8803 8804 void SweepClosure::print_free_block_coalesced(FreeChunk* fc) const { 8805 if (CMSTraceSweeper) { 8806 gclog_or_tty->print_cr("Sweep:coal_free_blk " PTR_FORMAT " (" SIZE_FORMAT ")", 8807 fc, fc->size()); 8808 } 8809 } 8810 8811 // CMSIsAliveClosure 8812 bool CMSIsAliveClosure::do_object_b(oop obj) { 8813 HeapWord* addr = (HeapWord*)obj; 8814 return addr != NULL && 8815 (!_span.contains(addr) || _bit_map->isMarked(addr)); 8816 } 8817 8818 8819 CMSKeepAliveClosure::CMSKeepAliveClosure( CMSCollector* collector, 8820 MemRegion span, 8821 CMSBitMap* bit_map, CMSMarkStack* mark_stack, 8822 bool cpc): 8823 _collector(collector), 8824 _span(span), 8825 _bit_map(bit_map), 8826 _mark_stack(mark_stack), 8827 _concurrent_precleaning(cpc) { 8828 assert(!_span.is_empty(), "Empty span could spell trouble"); 8829 } 8830 8831 8832 // CMSKeepAliveClosure: the serial version 8833 void CMSKeepAliveClosure::do_oop(oop obj) { 8834 HeapWord* addr = (HeapWord*)obj; 8835 if (_span.contains(addr) && 8836 !_bit_map->isMarked(addr)) { 8837 _bit_map->mark(addr); 8838 bool simulate_overflow = false; 8839 NOT_PRODUCT( 8840 if (CMSMarkStackOverflowALot && 8841 _collector->simulate_overflow()) { 8842 // simulate a stack overflow 8843 simulate_overflow = true; 8844 } 8845 ) 8846 if (simulate_overflow || !_mark_stack->push(obj)) { 8847 if (_concurrent_precleaning) { 8848 // We dirty the overflown object and let the remark 8849 // phase deal with it. 8850 assert(_collector->overflow_list_is_empty(), "Error"); 8851 // In the case of object arrays, we need to dirty all of 8852 // the cards that the object spans. No locking or atomics 8853 // are needed since no one else can be mutating the mod union 8854 // table. 8855 if (obj->is_objArray()) { 8856 size_t sz = obj->size(); 8857 HeapWord* end_card_addr = 8858 (HeapWord*)round_to((intptr_t)(addr+sz), CardTableModRefBS::card_size); 8859 MemRegion redirty_range = MemRegion(addr, end_card_addr); 8860 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 8861 _collector->_modUnionTable.mark_range(redirty_range); 8862 } else { 8863 _collector->_modUnionTable.mark(addr); 8864 } 8865 _collector->_ser_kac_preclean_ovflw++; 8866 } else { 8867 _collector->push_on_overflow_list(obj); 8868 _collector->_ser_kac_ovflw++; 8869 } 8870 } 8871 } 8872 } 8873 8874 void CMSKeepAliveClosure::do_oop(oop* p) { CMSKeepAliveClosure::do_oop_work(p); } 8875 void CMSKeepAliveClosure::do_oop(narrowOop* p) { CMSKeepAliveClosure::do_oop_work(p); } 8876 8877 // CMSParKeepAliveClosure: a parallel version of the above. 8878 // The work queues are private to each closure (thread), 8879 // but (may be) available for stealing by other threads. 8880 void CMSParKeepAliveClosure::do_oop(oop obj) { 8881 HeapWord* addr = (HeapWord*)obj; 8882 if (_span.contains(addr) && 8883 !_bit_map->isMarked(addr)) { 8884 // In general, during recursive tracing, several threads 8885 // may be concurrently getting here; the first one to 8886 // "tag" it, claims it. 8887 if (_bit_map->par_mark(addr)) { 8888 bool res = _work_queue->push(obj); 8889 assert(res, "Low water mark should be much less than capacity"); 8890 // Do a recursive trim in the hope that this will keep 8891 // stack usage lower, but leave some oops for potential stealers 8892 trim_queue(_low_water_mark); 8893 } // Else, another thread got there first 8894 } 8895 } 8896 8897 void CMSParKeepAliveClosure::do_oop(oop* p) { CMSParKeepAliveClosure::do_oop_work(p); } 8898 void CMSParKeepAliveClosure::do_oop(narrowOop* p) { CMSParKeepAliveClosure::do_oop_work(p); } 8899 8900 void CMSParKeepAliveClosure::trim_queue(uint max) { 8901 while (_work_queue->size() > max) { 8902 oop new_oop; 8903 if (_work_queue->pop_local(new_oop)) { 8904 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 8905 assert(_bit_map->isMarked((HeapWord*)new_oop), 8906 "no white objects on this stack!"); 8907 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 8908 // iterate over the oops in this oop, marking and pushing 8909 // the ones in CMS heap (i.e. in _span). 8910 new_oop->oop_iterate(&_mark_and_push); 8911 } 8912 } 8913 } 8914 8915 CMSInnerParMarkAndPushClosure::CMSInnerParMarkAndPushClosure( 8916 CMSCollector* collector, 8917 MemRegion span, CMSBitMap* bit_map, 8918 OopTaskQueue* work_queue): 8919 _collector(collector), 8920 _span(span), 8921 _bit_map(bit_map), 8922 _work_queue(work_queue) { } 8923 8924 void CMSInnerParMarkAndPushClosure::do_oop(oop obj) { 8925 HeapWord* addr = (HeapWord*)obj; 8926 if (_span.contains(addr) && 8927 !_bit_map->isMarked(addr)) { 8928 if (_bit_map->par_mark(addr)) { 8929 bool simulate_overflow = false; 8930 NOT_PRODUCT( 8931 if (CMSMarkStackOverflowALot && 8932 _collector->par_simulate_overflow()) { 8933 // simulate a stack overflow 8934 simulate_overflow = true; 8935 } 8936 ) 8937 if (simulate_overflow || !_work_queue->push(obj)) { 8938 _collector->par_push_on_overflow_list(obj); 8939 _collector->_par_kac_ovflw++; 8940 } 8941 } // Else another thread got there already 8942 } 8943 } 8944 8945 void CMSInnerParMarkAndPushClosure::do_oop(oop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); } 8946 void CMSInnerParMarkAndPushClosure::do_oop(narrowOop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); } 8947 8948 ////////////////////////////////////////////////////////////////// 8949 // CMSExpansionCause ///////////////////////////// 8950 ////////////////////////////////////////////////////////////////// 8951 const char* CMSExpansionCause::to_string(CMSExpansionCause::Cause cause) { 8952 switch (cause) { 8953 case _no_expansion: 8954 return "No expansion"; 8955 case _satisfy_free_ratio: 8956 return "Free ratio"; 8957 case _satisfy_promotion: 8958 return "Satisfy promotion"; 8959 case _satisfy_allocation: 8960 return "allocation"; 8961 case _allocate_par_lab: 8962 return "Par LAB"; 8963 case _allocate_par_spooling_space: 8964 return "Par Spooling Space"; 8965 case _adaptive_size_policy: 8966 return "Ergonomics"; 8967 default: 8968 return "unknown"; 8969 } 8970 } 8971 8972 void CMSDrainMarkingStackClosure::do_void() { 8973 // the max number to take from overflow list at a time 8974 const size_t num = _mark_stack->capacity()/4; 8975 assert(!_concurrent_precleaning || _collector->overflow_list_is_empty(), 8976 "Overflow list should be NULL during concurrent phases"); 8977 while (!_mark_stack->isEmpty() || 8978 // if stack is empty, check the overflow list 8979 _collector->take_from_overflow_list(num, _mark_stack)) { 8980 oop obj = _mark_stack->pop(); 8981 HeapWord* addr = (HeapWord*)obj; 8982 assert(_span.contains(addr), "Should be within span"); 8983 assert(_bit_map->isMarked(addr), "Should be marked"); 8984 assert(obj->is_oop(), "Should be an oop"); 8985 obj->oop_iterate(_keep_alive); 8986 } 8987 } 8988 8989 void CMSParDrainMarkingStackClosure::do_void() { 8990 // drain queue 8991 trim_queue(0); 8992 } 8993 8994 // Trim our work_queue so its length is below max at return 8995 void CMSParDrainMarkingStackClosure::trim_queue(uint max) { 8996 while (_work_queue->size() > max) { 8997 oop new_oop; 8998 if (_work_queue->pop_local(new_oop)) { 8999 assert(new_oop->is_oop(), "Expected an oop"); 9000 assert(_bit_map->isMarked((HeapWord*)new_oop), 9001 "no white objects on this stack!"); 9002 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 9003 // iterate over the oops in this oop, marking and pushing 9004 // the ones in CMS heap (i.e. in _span). 9005 new_oop->oop_iterate(&_mark_and_push); 9006 } 9007 } 9008 } 9009 9010 //////////////////////////////////////////////////////////////////// 9011 // Support for Marking Stack Overflow list handling and related code 9012 //////////////////////////////////////////////////////////////////// 9013 // Much of the following code is similar in shape and spirit to the 9014 // code used in ParNewGC. We should try and share that code 9015 // as much as possible in the future. 9016 9017 #ifndef PRODUCT 9018 // Debugging support for CMSStackOverflowALot 9019 9020 // It's OK to call this multi-threaded; the worst thing 9021 // that can happen is that we'll get a bunch of closely 9022 // spaced simulated overflows, but that's OK, in fact 9023 // probably good as it would exercise the overflow code 9024 // under contention. 9025 bool CMSCollector::simulate_overflow() { 9026 if (_overflow_counter-- <= 0) { // just being defensive 9027 _overflow_counter = CMSMarkStackOverflowInterval; 9028 return true; 9029 } else { 9030 return false; 9031 } 9032 } 9033 9034 bool CMSCollector::par_simulate_overflow() { 9035 return simulate_overflow(); 9036 } 9037 #endif 9038 9039 // Single-threaded 9040 bool CMSCollector::take_from_overflow_list(size_t num, CMSMarkStack* stack) { 9041 assert(stack->isEmpty(), "Expected precondition"); 9042 assert(stack->capacity() > num, "Shouldn't bite more than can chew"); 9043 size_t i = num; 9044 oop cur = _overflow_list; 9045 const markOop proto = markOopDesc::prototype(); 9046 NOT_PRODUCT(ssize_t n = 0;) 9047 for (oop next; i > 0 && cur != NULL; cur = next, i--) { 9048 next = oop(cur->mark()); 9049 cur->set_mark(proto); // until proven otherwise 9050 assert(cur->is_oop(), "Should be an oop"); 9051 bool res = stack->push(cur); 9052 assert(res, "Bit off more than can chew?"); 9053 NOT_PRODUCT(n++;) 9054 } 9055 _overflow_list = cur; 9056 #ifndef PRODUCT 9057 assert(_num_par_pushes >= n, "Too many pops?"); 9058 _num_par_pushes -=n; 9059 #endif 9060 return !stack->isEmpty(); 9061 } 9062 9063 #define BUSY (cast_to_oop<intptr_t>(0x1aff1aff)) 9064 // (MT-safe) Get a prefix of at most "num" from the list. 9065 // The overflow list is chained through the mark word of 9066 // each object in the list. We fetch the entire list, 9067 // break off a prefix of the right size and return the 9068 // remainder. If other threads try to take objects from 9069 // the overflow list at that time, they will wait for 9070 // some time to see if data becomes available. If (and 9071 // only if) another thread places one or more object(s) 9072 // on the global list before we have returned the suffix 9073 // to the global list, we will walk down our local list 9074 // to find its end and append the global list to 9075 // our suffix before returning it. This suffix walk can 9076 // prove to be expensive (quadratic in the amount of traffic) 9077 // when there are many objects in the overflow list and 9078 // there is much producer-consumer contention on the list. 9079 // *NOTE*: The overflow list manipulation code here and 9080 // in ParNewGeneration:: are very similar in shape, 9081 // except that in the ParNew case we use the old (from/eden) 9082 // copy of the object to thread the list via its klass word. 9083 // Because of the common code, if you make any changes in 9084 // the code below, please check the ParNew version to see if 9085 // similar changes might be needed. 9086 // CR 6797058 has been filed to consolidate the common code. 9087 bool CMSCollector::par_take_from_overflow_list(size_t num, 9088 OopTaskQueue* work_q, 9089 int no_of_gc_threads) { 9090 assert(work_q->size() == 0, "First empty local work queue"); 9091 assert(num < work_q->max_elems(), "Can't bite more than we can chew"); 9092 if (_overflow_list == NULL) { 9093 return false; 9094 } 9095 // Grab the entire list; we'll put back a suffix 9096 oop prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list)); 9097 Thread* tid = Thread::current(); 9098 // Before "no_of_gc_threads" was introduced CMSOverflowSpinCount was 9099 // set to ParallelGCThreads. 9100 size_t CMSOverflowSpinCount = (size_t) no_of_gc_threads; // was ParallelGCThreads; 9101 size_t sleep_time_millis = MAX2((size_t)1, num/100); 9102 // If the list is busy, we spin for a short while, 9103 // sleeping between attempts to get the list. 9104 for (size_t spin = 0; prefix == BUSY && spin < CMSOverflowSpinCount; spin++) { 9105 os::sleep(tid, sleep_time_millis, false); 9106 if (_overflow_list == NULL) { 9107 // Nothing left to take 9108 return false; 9109 } else if (_overflow_list != BUSY) { 9110 // Try and grab the prefix 9111 prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list)); 9112 } 9113 } 9114 // If the list was found to be empty, or we spun long 9115 // enough, we give up and return empty-handed. If we leave 9116 // the list in the BUSY state below, it must be the case that 9117 // some other thread holds the overflow list and will set it 9118 // to a non-BUSY state in the future. 9119 if (prefix == NULL || prefix == BUSY) { 9120 // Nothing to take or waited long enough 9121 if (prefix == NULL) { 9122 // Write back the NULL in case we overwrote it with BUSY above 9123 // and it is still the same value. 9124 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY); 9125 } 9126 return false; 9127 } 9128 assert(prefix != NULL && prefix != BUSY, "Error"); 9129 size_t i = num; 9130 oop cur = prefix; 9131 // Walk down the first "num" objects, unless we reach the end. 9132 for (; i > 1 && cur->mark() != NULL; cur = oop(cur->mark()), i--); 9133 if (cur->mark() == NULL) { 9134 // We have "num" or fewer elements in the list, so there 9135 // is nothing to return to the global list. 9136 // Write back the NULL in lieu of the BUSY we wrote 9137 // above, if it is still the same value. 9138 if (_overflow_list == BUSY) { 9139 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY); 9140 } 9141 } else { 9142 // Chop off the suffix and return it to the global list. 9143 assert(cur->mark() != BUSY, "Error"); 9144 oop suffix_head = cur->mark(); // suffix will be put back on global list 9145 cur->set_mark(NULL); // break off suffix 9146 // It's possible that the list is still in the empty(busy) state 9147 // we left it in a short while ago; in that case we may be 9148 // able to place back the suffix without incurring the cost 9149 // of a walk down the list. 9150 oop observed_overflow_list = _overflow_list; 9151 oop cur_overflow_list = observed_overflow_list; 9152 bool attached = false; 9153 while (observed_overflow_list == BUSY || observed_overflow_list == NULL) { 9154 observed_overflow_list = 9155 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list); 9156 if (cur_overflow_list == observed_overflow_list) { 9157 attached = true; 9158 break; 9159 } else cur_overflow_list = observed_overflow_list; 9160 } 9161 if (!attached) { 9162 // Too bad, someone else sneaked in (at least) an element; we'll need 9163 // to do a splice. Find tail of suffix so we can prepend suffix to global 9164 // list. 9165 for (cur = suffix_head; cur->mark() != NULL; cur = (oop)(cur->mark())); 9166 oop suffix_tail = cur; 9167 assert(suffix_tail != NULL && suffix_tail->mark() == NULL, 9168 "Tautology"); 9169 observed_overflow_list = _overflow_list; 9170 do { 9171 cur_overflow_list = observed_overflow_list; 9172 if (cur_overflow_list != BUSY) { 9173 // Do the splice ... 9174 suffix_tail->set_mark(markOop(cur_overflow_list)); 9175 } else { // cur_overflow_list == BUSY 9176 suffix_tail->set_mark(NULL); 9177 } 9178 // ... and try to place spliced list back on overflow_list ... 9179 observed_overflow_list = 9180 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list); 9181 } while (cur_overflow_list != observed_overflow_list); 9182 // ... until we have succeeded in doing so. 9183 } 9184 } 9185 9186 // Push the prefix elements on work_q 9187 assert(prefix != NULL, "control point invariant"); 9188 const markOop proto = markOopDesc::prototype(); 9189 oop next; 9190 NOT_PRODUCT(ssize_t n = 0;) 9191 for (cur = prefix; cur != NULL; cur = next) { 9192 next = oop(cur->mark()); 9193 cur->set_mark(proto); // until proven otherwise 9194 assert(cur->is_oop(), "Should be an oop"); 9195 bool res = work_q->push(cur); 9196 assert(res, "Bit off more than we can chew?"); 9197 NOT_PRODUCT(n++;) 9198 } 9199 #ifndef PRODUCT 9200 assert(_num_par_pushes >= n, "Too many pops?"); 9201 Atomic::add_ptr(-(intptr_t)n, &_num_par_pushes); 9202 #endif 9203 return true; 9204 } 9205 9206 // Single-threaded 9207 void CMSCollector::push_on_overflow_list(oop p) { 9208 NOT_PRODUCT(_num_par_pushes++;) 9209 assert(p->is_oop(), "Not an oop"); 9210 preserve_mark_if_necessary(p); 9211 p->set_mark((markOop)_overflow_list); 9212 _overflow_list = p; 9213 } 9214 9215 // Multi-threaded; use CAS to prepend to overflow list 9216 void CMSCollector::par_push_on_overflow_list(oop p) { 9217 NOT_PRODUCT(Atomic::inc_ptr(&_num_par_pushes);) 9218 assert(p->is_oop(), "Not an oop"); 9219 par_preserve_mark_if_necessary(p); 9220 oop observed_overflow_list = _overflow_list; 9221 oop cur_overflow_list; 9222 do { 9223 cur_overflow_list = observed_overflow_list; 9224 if (cur_overflow_list != BUSY) { 9225 p->set_mark(markOop(cur_overflow_list)); 9226 } else { 9227 p->set_mark(NULL); 9228 } 9229 observed_overflow_list = 9230 (oop) Atomic::cmpxchg_ptr(p, &_overflow_list, cur_overflow_list); 9231 } while (cur_overflow_list != observed_overflow_list); 9232 } 9233 #undef BUSY 9234 9235 // Single threaded 9236 // General Note on GrowableArray: pushes may silently fail 9237 // because we are (temporarily) out of C-heap for expanding 9238 // the stack. The problem is quite ubiquitous and affects 9239 // a lot of code in the JVM. The prudent thing for GrowableArray 9240 // to do (for now) is to exit with an error. However, that may 9241 // be too draconian in some cases because the caller may be 9242 // able to recover without much harm. For such cases, we 9243 // should probably introduce a "soft_push" method which returns 9244 // an indication of success or failure with the assumption that 9245 // the caller may be able to recover from a failure; code in 9246 // the VM can then be changed, incrementally, to deal with such 9247 // failures where possible, thus, incrementally hardening the VM 9248 // in such low resource situations. 9249 void CMSCollector::preserve_mark_work(oop p, markOop m) { 9250 _preserved_oop_stack.push(p); 9251 _preserved_mark_stack.push(m); 9252 assert(m == p->mark(), "Mark word changed"); 9253 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), 9254 "bijection"); 9255 } 9256 9257 // Single threaded 9258 void CMSCollector::preserve_mark_if_necessary(oop p) { 9259 markOop m = p->mark(); 9260 if (m->must_be_preserved(p)) { 9261 preserve_mark_work(p, m); 9262 } 9263 } 9264 9265 void CMSCollector::par_preserve_mark_if_necessary(oop p) { 9266 markOop m = p->mark(); 9267 if (m->must_be_preserved(p)) { 9268 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag); 9269 // Even though we read the mark word without holding 9270 // the lock, we are assured that it will not change 9271 // because we "own" this oop, so no other thread can 9272 // be trying to push it on the overflow list; see 9273 // the assertion in preserve_mark_work() that checks 9274 // that m == p->mark(). 9275 preserve_mark_work(p, m); 9276 } 9277 } 9278 9279 // We should be able to do this multi-threaded, 9280 // a chunk of stack being a task (this is 9281 // correct because each oop only ever appears 9282 // once in the overflow list. However, it's 9283 // not very easy to completely overlap this with 9284 // other operations, so will generally not be done 9285 // until all work's been completed. Because we 9286 // expect the preserved oop stack (set) to be small, 9287 // it's probably fine to do this single-threaded. 9288 // We can explore cleverer concurrent/overlapped/parallel 9289 // processing of preserved marks if we feel the 9290 // need for this in the future. Stack overflow should 9291 // be so rare in practice and, when it happens, its 9292 // effect on performance so great that this will 9293 // likely just be in the noise anyway. 9294 void CMSCollector::restore_preserved_marks_if_any() { 9295 assert(SafepointSynchronize::is_at_safepoint(), 9296 "world should be stopped"); 9297 assert(Thread::current()->is_ConcurrentGC_thread() || 9298 Thread::current()->is_VM_thread(), 9299 "should be single-threaded"); 9300 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), 9301 "bijection"); 9302 9303 while (!_preserved_oop_stack.is_empty()) { 9304 oop p = _preserved_oop_stack.pop(); 9305 assert(p->is_oop(), "Should be an oop"); 9306 assert(_span.contains(p), "oop should be in _span"); 9307 assert(p->mark() == markOopDesc::prototype(), 9308 "Set when taken from overflow list"); 9309 markOop m = _preserved_mark_stack.pop(); 9310 p->set_mark(m); 9311 } 9312 assert(_preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(), 9313 "stacks were cleared above"); 9314 } 9315 9316 #ifndef PRODUCT 9317 bool CMSCollector::no_preserved_marks() const { 9318 return _preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(); 9319 } 9320 #endif 9321 9322 CMSAdaptiveSizePolicy* ASConcurrentMarkSweepGeneration::cms_size_policy() const 9323 { 9324 GenCollectedHeap* gch = (GenCollectedHeap*) GenCollectedHeap::heap(); 9325 CMSAdaptiveSizePolicy* size_policy = 9326 (CMSAdaptiveSizePolicy*) gch->gen_policy()->size_policy(); 9327 assert(size_policy->is_gc_cms_adaptive_size_policy(), 9328 "Wrong type for size policy"); 9329 return size_policy; 9330 } 9331 9332 void ASConcurrentMarkSweepGeneration::resize(size_t cur_promo_size, 9333 size_t desired_promo_size) { 9334 if (cur_promo_size < desired_promo_size) { 9335 size_t expand_bytes = desired_promo_size - cur_promo_size; 9336 if (PrintAdaptiveSizePolicy && Verbose) { 9337 gclog_or_tty->print_cr(" ASConcurrentMarkSweepGeneration::resize " 9338 "Expanding tenured generation by " SIZE_FORMAT " (bytes)", 9339 expand_bytes); 9340 } 9341 expand(expand_bytes, 9342 MinHeapDeltaBytes, 9343 CMSExpansionCause::_adaptive_size_policy); 9344 } else if (desired_promo_size < cur_promo_size) { 9345 size_t shrink_bytes = cur_promo_size - desired_promo_size; 9346 if (PrintAdaptiveSizePolicy && Verbose) { 9347 gclog_or_tty->print_cr(" ASConcurrentMarkSweepGeneration::resize " 9348 "Shrinking tenured generation by " SIZE_FORMAT " (bytes)", 9349 shrink_bytes); 9350 } 9351 shrink(shrink_bytes); 9352 } 9353 } 9354 9355 CMSGCAdaptivePolicyCounters* ASConcurrentMarkSweepGeneration::gc_adaptive_policy_counters() { 9356 GenCollectedHeap* gch = GenCollectedHeap::heap(); 9357 CMSGCAdaptivePolicyCounters* counters = 9358 (CMSGCAdaptivePolicyCounters*) gch->collector_policy()->counters(); 9359 assert(counters->kind() == GCPolicyCounters::CMSGCAdaptivePolicyCountersKind, 9360 "Wrong kind of counters"); 9361 return counters; 9362 } 9363 9364 9365 void ASConcurrentMarkSweepGeneration::update_counters() { 9366 if (UsePerfData) { 9367 _space_counters->update_all(); 9368 _gen_counters->update_all(); 9369 CMSGCAdaptivePolicyCounters* counters = gc_adaptive_policy_counters(); 9370 GenCollectedHeap* gch = GenCollectedHeap::heap(); 9371 CMSGCStats* gc_stats_l = (CMSGCStats*) gc_stats(); 9372 assert(gc_stats_l->kind() == GCStats::CMSGCStatsKind, 9373 "Wrong gc statistics type"); 9374 counters->update_counters(gc_stats_l); 9375 } 9376 } 9377 9378 void ASConcurrentMarkSweepGeneration::update_counters(size_t used) { 9379 if (UsePerfData) { 9380 _space_counters->update_used(used); 9381 _space_counters->update_capacity(); 9382 _gen_counters->update_all(); 9383 9384 CMSGCAdaptivePolicyCounters* counters = gc_adaptive_policy_counters(); 9385 GenCollectedHeap* gch = GenCollectedHeap::heap(); 9386 CMSGCStats* gc_stats_l = (CMSGCStats*) gc_stats(); 9387 assert(gc_stats_l->kind() == GCStats::CMSGCStatsKind, 9388 "Wrong gc statistics type"); 9389 counters->update_counters(gc_stats_l); 9390 } 9391 } 9392 9393 void ASConcurrentMarkSweepGeneration::shrink_by(size_t desired_bytes) { 9394 assert_locked_or_safepoint(Heap_lock); 9395 assert_lock_strong(freelistLock()); 9396 HeapWord* old_end = _cmsSpace->end(); 9397 HeapWord* unallocated_start = _cmsSpace->unallocated_block(); 9398 assert(old_end >= unallocated_start, "Miscalculation of unallocated_start"); 9399 FreeChunk* chunk_at_end = find_chunk_at_end(); 9400 if (chunk_at_end == NULL) { 9401 // No room to shrink 9402 if (PrintGCDetails && Verbose) { 9403 gclog_or_tty->print_cr("No room to shrink: old_end " 9404 PTR_FORMAT " unallocated_start " PTR_FORMAT 9405 " chunk_at_end " PTR_FORMAT, 9406 old_end, unallocated_start, chunk_at_end); 9407 } 9408 return; 9409 } else { 9410 9411 // Find the chunk at the end of the space and determine 9412 // how much it can be shrunk. 9413 size_t shrinkable_size_in_bytes = chunk_at_end->size(); 9414 size_t aligned_shrinkable_size_in_bytes = 9415 align_size_down(shrinkable_size_in_bytes, os::vm_page_size()); 9416 assert(unallocated_start <= (HeapWord*) chunk_at_end->end(), 9417 "Inconsistent chunk at end of space"); 9418 size_t bytes = MIN2(desired_bytes, aligned_shrinkable_size_in_bytes); 9419 size_t word_size_before = heap_word_size(_virtual_space.committed_size()); 9420 9421 // Shrink the underlying space 9422 _virtual_space.shrink_by(bytes); 9423 if (PrintGCDetails && Verbose) { 9424 gclog_or_tty->print_cr("ConcurrentMarkSweepGeneration::shrink_by:" 9425 " desired_bytes " SIZE_FORMAT 9426 " shrinkable_size_in_bytes " SIZE_FORMAT 9427 " aligned_shrinkable_size_in_bytes " SIZE_FORMAT 9428 " bytes " SIZE_FORMAT, 9429 desired_bytes, shrinkable_size_in_bytes, 9430 aligned_shrinkable_size_in_bytes, bytes); 9431 gclog_or_tty->print_cr(" old_end " SIZE_FORMAT 9432 " unallocated_start " SIZE_FORMAT, 9433 old_end, unallocated_start); 9434 } 9435 9436 // If the space did shrink (shrinking is not guaranteed), 9437 // shrink the chunk at the end by the appropriate amount. 9438 if (((HeapWord*)_virtual_space.high()) < old_end) { 9439 size_t new_word_size = 9440 heap_word_size(_virtual_space.committed_size()); 9441 9442 // Have to remove the chunk from the dictionary because it is changing 9443 // size and might be someplace elsewhere in the dictionary. 9444 9445 // Get the chunk at end, shrink it, and put it 9446 // back. 9447 _cmsSpace->removeChunkFromDictionary(chunk_at_end); 9448 size_t word_size_change = word_size_before - new_word_size; 9449 size_t chunk_at_end_old_size = chunk_at_end->size(); 9450 assert(chunk_at_end_old_size >= word_size_change, 9451 "Shrink is too large"); 9452 chunk_at_end->set_size(chunk_at_end_old_size - 9453 word_size_change); 9454 _cmsSpace->freed((HeapWord*) chunk_at_end->end(), 9455 word_size_change); 9456 9457 _cmsSpace->returnChunkToDictionary(chunk_at_end); 9458 9459 MemRegion mr(_cmsSpace->bottom(), new_word_size); 9460 _bts->resize(new_word_size); // resize the block offset shared array 9461 Universe::heap()->barrier_set()->resize_covered_region(mr); 9462 _cmsSpace->assert_locked(); 9463 _cmsSpace->set_end((HeapWord*)_virtual_space.high()); 9464 9465 NOT_PRODUCT(_cmsSpace->dictionary()->verify()); 9466 9467 // update the space and generation capacity counters 9468 if (UsePerfData) { 9469 _space_counters->update_capacity(); 9470 _gen_counters->update_all(); 9471 } 9472 9473 if (Verbose && PrintGCDetails) { 9474 size_t new_mem_size = _virtual_space.committed_size(); 9475 size_t old_mem_size = new_mem_size + bytes; 9476 gclog_or_tty->print_cr("Shrinking %s from " SIZE_FORMAT "K by " SIZE_FORMAT "K to " SIZE_FORMAT "K", 9477 name(), old_mem_size/K, bytes/K, new_mem_size/K); 9478 } 9479 } 9480 9481 assert(_cmsSpace->unallocated_block() <= _cmsSpace->end(), 9482 "Inconsistency at end of space"); 9483 assert(chunk_at_end->end() == (uintptr_t*) _cmsSpace->end(), 9484 "Shrinking is inconsistent"); 9485 return; 9486 } 9487 } 9488 // Transfer some number of overflown objects to usual marking 9489 // stack. Return true if some objects were transferred. 9490 bool MarkRefsIntoAndScanClosure::take_from_overflow_list() { 9491 size_t num = MIN2((size_t)(_mark_stack->capacity() - _mark_stack->length())/4, 9492 (size_t)ParGCDesiredObjsFromOverflowList); 9493 9494 bool res = _collector->take_from_overflow_list(num, _mark_stack); 9495 assert(_collector->overflow_list_is_empty() || res, 9496 "If list is not empty, we should have taken something"); 9497 assert(!res || !_mark_stack->isEmpty(), 9498 "If we took something, it should now be on our stack"); 9499 return res; 9500 } 9501 9502 size_t MarkDeadObjectsClosure::do_blk(HeapWord* addr) { 9503 size_t res = _sp->block_size_no_stall(addr, _collector); 9504 if (_sp->block_is_obj(addr)) { 9505 if (_live_bit_map->isMarked(addr)) { 9506 // It can't have been dead in a previous cycle 9507 guarantee(!_dead_bit_map->isMarked(addr), "No resurrection!"); 9508 } else { 9509 _dead_bit_map->mark(addr); // mark the dead object 9510 } 9511 } 9512 // Could be 0, if the block size could not be computed without stalling. 9513 return res; 9514 } 9515 9516 TraceCMSMemoryManagerStats::TraceCMSMemoryManagerStats(CMSCollector::CollectorState phase, GCCause::Cause cause): TraceMemoryManagerStats() { 9517 9518 switch (phase) { 9519 case CMSCollector::InitialMarking: 9520 initialize(true /* fullGC */ , 9521 cause /* cause of the GC */, 9522 true /* recordGCBeginTime */, 9523 true /* recordPreGCUsage */, 9524 false /* recordPeakUsage */, 9525 false /* recordPostGCusage */, 9526 true /* recordAccumulatedGCTime */, 9527 false /* recordGCEndTime */, 9528 false /* countCollection */ ); 9529 break; 9530 9531 case CMSCollector::FinalMarking: 9532 initialize(true /* fullGC */ , 9533 cause /* cause of the GC */, 9534 false /* recordGCBeginTime */, 9535 false /* recordPreGCUsage */, 9536 false /* recordPeakUsage */, 9537 false /* recordPostGCusage */, 9538 true /* recordAccumulatedGCTime */, 9539 false /* recordGCEndTime */, 9540 false /* countCollection */ ); 9541 break; 9542 9543 case CMSCollector::Sweeping: 9544 initialize(true /* fullGC */ , 9545 cause /* cause of the GC */, 9546 false /* recordGCBeginTime */, 9547 false /* recordPreGCUsage */, 9548 true /* recordPeakUsage */, 9549 true /* recordPostGCusage */, 9550 false /* recordAccumulatedGCTime */, 9551 true /* recordGCEndTime */, 9552 true /* countCollection */ ); 9553 break; 9554 9555 default: 9556 ShouldNotReachHere(); 9557 } 9558 }