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/referencePolicy.hpp" 54 #include "memory/resourceArea.hpp" 55 #include "memory/tenuredGeneration.hpp" 56 #include "oops/oop.inline.hpp" 57 #include "prims/jvmtiExport.hpp" 58 #include "runtime/globals_extension.hpp" 59 #include "runtime/handles.inline.hpp" 60 #include "runtime/java.hpp" 61 #include "runtime/vmThread.hpp" 62 #include "services/memoryService.hpp" 63 #include "services/runtimeService.hpp" 64 65 // statics 66 CMSCollector* ConcurrentMarkSweepGeneration::_collector = NULL; 67 bool CMSCollector::_full_gc_requested = false; 68 GCCause::Cause CMSCollector::_full_gc_cause = GCCause::_no_gc; 69 70 ////////////////////////////////////////////////////////////////// 71 // In support of CMS/VM thread synchronization 72 ////////////////////////////////////////////////////////////////// 73 // We split use of the CGC_lock into 2 "levels". 74 // The low-level locking is of the usual CGC_lock monitor. We introduce 75 // a higher level "token" (hereafter "CMS token") built on top of the 76 // low level monitor (hereafter "CGC lock"). 77 // The token-passing protocol gives priority to the VM thread. The 78 // CMS-lock doesn't provide any fairness guarantees, but clients 79 // should ensure that it is only held for very short, bounded 80 // durations. 81 // 82 // When either of the CMS thread or the VM thread is involved in 83 // collection operations during which it does not want the other 84 // thread to interfere, it obtains the CMS token. 85 // 86 // If either thread tries to get the token while the other has 87 // it, that thread waits. However, if the VM thread and CMS thread 88 // both want the token, then the VM thread gets priority while the 89 // CMS thread waits. This ensures, for instance, that the "concurrent" 90 // phases of the CMS thread's work do not block out the VM thread 91 // for long periods of time as the CMS thread continues to hog 92 // the token. (See bug 4616232). 93 // 94 // The baton-passing functions are, however, controlled by the 95 // flags _foregroundGCShouldWait and _foregroundGCIsActive, 96 // and here the low-level CMS lock, not the high level token, 97 // ensures mutual exclusion. 98 // 99 // Two important conditions that we have to satisfy: 100 // 1. if a thread does a low-level wait on the CMS lock, then it 101 // relinquishes the CMS token if it were holding that token 102 // when it acquired the low-level CMS lock. 103 // 2. any low-level notifications on the low-level lock 104 // should only be sent when a thread has relinquished the token. 105 // 106 // In the absence of either property, we'd have potential deadlock. 107 // 108 // We protect each of the CMS (concurrent and sequential) phases 109 // with the CMS _token_, not the CMS _lock_. 110 // 111 // The only code protected by CMS lock is the token acquisition code 112 // itself, see ConcurrentMarkSweepThread::[de]synchronize(), and the 113 // baton-passing code. 114 // 115 // Unfortunately, i couldn't come up with a good abstraction to factor and 116 // hide the naked CGC_lock manipulation in the baton-passing code 117 // further below. That's something we should try to do. Also, the proof 118 // of correctness of this 2-level locking scheme is far from obvious, 119 // and potentially quite slippery. We have an uneasy supsicion, for instance, 120 // that there may be a theoretical possibility of delay/starvation in the 121 // low-level lock/wait/notify scheme used for the baton-passing because of 122 // potential intereference with the priority scheme embodied in the 123 // CMS-token-passing protocol. See related comments at a CGC_lock->wait() 124 // invocation further below and marked with "XXX 20011219YSR". 125 // Indeed, as we note elsewhere, this may become yet more slippery 126 // in the presence of multiple CMS and/or multiple VM threads. XXX 127 128 class CMSTokenSync: public StackObj { 129 private: 130 bool _is_cms_thread; 131 public: 132 CMSTokenSync(bool is_cms_thread): 133 _is_cms_thread(is_cms_thread) { 134 assert(is_cms_thread == Thread::current()->is_ConcurrentGC_thread(), 135 "Incorrect argument to constructor"); 136 ConcurrentMarkSweepThread::synchronize(_is_cms_thread); 137 } 138 139 ~CMSTokenSync() { 140 assert(_is_cms_thread ? 141 ConcurrentMarkSweepThread::cms_thread_has_cms_token() : 142 ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 143 "Incorrect state"); 144 ConcurrentMarkSweepThread::desynchronize(_is_cms_thread); 145 } 146 }; 147 148 // Convenience class that does a CMSTokenSync, and then acquires 149 // upto three locks. 150 class CMSTokenSyncWithLocks: public CMSTokenSync { 151 private: 152 // Note: locks are acquired in textual declaration order 153 // and released in the opposite order 154 MutexLockerEx _locker1, _locker2, _locker3; 155 public: 156 CMSTokenSyncWithLocks(bool is_cms_thread, Mutex* mutex1, 157 Mutex* mutex2 = NULL, Mutex* mutex3 = NULL): 158 CMSTokenSync(is_cms_thread), 159 _locker1(mutex1, Mutex::_no_safepoint_check_flag), 160 _locker2(mutex2, Mutex::_no_safepoint_check_flag), 161 _locker3(mutex3, Mutex::_no_safepoint_check_flag) 162 { } 163 }; 164 165 166 // Wrapper class to temporarily disable icms during a foreground cms collection. 167 class ICMSDisabler: public StackObj { 168 public: 169 // The ctor disables icms and wakes up the thread so it notices the change; 170 // the dtor re-enables icms. Note that the CMSCollector methods will check 171 // CMSIncrementalMode. 172 ICMSDisabler() { CMSCollector::disable_icms(); CMSCollector::start_icms(); } 173 ~ICMSDisabler() { CMSCollector::enable_icms(); } 174 }; 175 176 ////////////////////////////////////////////////////////////////// 177 // Concurrent Mark-Sweep Generation ///////////////////////////// 178 ////////////////////////////////////////////////////////////////// 179 180 NOT_PRODUCT(CompactibleFreeListSpace* debug_cms_space;) 181 182 // This struct contains per-thread things necessary to support parallel 183 // young-gen collection. 184 class CMSParGCThreadState: public CHeapObj<mtGC> { 185 public: 186 CFLS_LAB lab; 187 PromotionInfo promo; 188 189 // Constructor. 190 CMSParGCThreadState(CompactibleFreeListSpace* cfls) : lab(cfls) { 191 promo.setSpace(cfls); 192 } 193 }; 194 195 ConcurrentMarkSweepGeneration::ConcurrentMarkSweepGeneration( 196 ReservedSpace rs, size_t initial_byte_size, int level, 197 CardTableRS* ct, bool use_adaptive_freelists, 198 FreeBlockDictionary<FreeChunk>::DictionaryChoice dictionaryChoice) : 199 CardGeneration(rs, initial_byte_size, level, ct), 200 _dilatation_factor(((double)MinChunkSize)/((double)(CollectedHeap::min_fill_size()))), 201 _debug_collection_type(Concurrent_collection_type), 202 _did_compact(false) 203 { 204 HeapWord* bottom = (HeapWord*) _virtual_space.low(); 205 HeapWord* end = (HeapWord*) _virtual_space.high(); 206 207 _direct_allocated_words = 0; 208 NOT_PRODUCT( 209 _numObjectsPromoted = 0; 210 _numWordsPromoted = 0; 211 _numObjectsAllocated = 0; 212 _numWordsAllocated = 0; 213 ) 214 215 _cmsSpace = new CompactibleFreeListSpace(_bts, MemRegion(bottom, end), 216 use_adaptive_freelists, 217 dictionaryChoice); 218 NOT_PRODUCT(debug_cms_space = _cmsSpace;) 219 if (_cmsSpace == NULL) { 220 vm_exit_during_initialization( 221 "CompactibleFreeListSpace allocation failure"); 222 } 223 _cmsSpace->_gen = this; 224 225 _gc_stats = new CMSGCStats(); 226 227 // Verify the assumption that FreeChunk::_prev and OopDesc::_klass 228 // offsets match. The ability to tell free chunks from objects 229 // depends on this property. 230 debug_only( 231 FreeChunk* junk = NULL; 232 assert(UseCompressedKlassPointers || 233 junk->prev_addr() == (void*)(oop(junk)->klass_addr()), 234 "Offset of FreeChunk::_prev within FreeChunk must match" 235 " that of OopDesc::_klass within OopDesc"); 236 ) 237 if (CollectedHeap::use_parallel_gc_threads()) { 238 typedef CMSParGCThreadState* CMSParGCThreadStatePtr; 239 _par_gc_thread_states = 240 NEW_C_HEAP_ARRAY(CMSParGCThreadStatePtr, ParallelGCThreads, mtGC); 241 if (_par_gc_thread_states == NULL) { 242 vm_exit_during_initialization("Could not allocate par gc structs"); 243 } 244 for (uint i = 0; i < ParallelGCThreads; i++) { 245 _par_gc_thread_states[i] = new CMSParGCThreadState(cmsSpace()); 246 if (_par_gc_thread_states[i] == NULL) { 247 vm_exit_during_initialization("Could not allocate par gc structs"); 248 } 249 } 250 } else { 251 _par_gc_thread_states = NULL; 252 } 253 _incremental_collection_failed = false; 254 // The "dilatation_factor" is the expansion that can occur on 255 // account of the fact that the minimum object size in the CMS 256 // generation may be larger than that in, say, a contiguous young 257 // generation. 258 // Ideally, in the calculation below, we'd compute the dilatation 259 // factor as: MinChunkSize/(promoting_gen's min object size) 260 // Since we do not have such a general query interface for the 261 // promoting generation, we'll instead just use the mimimum 262 // object size (which today is a header's worth of space); 263 // note that all arithmetic is in units of HeapWords. 264 assert(MinChunkSize >= CollectedHeap::min_fill_size(), "just checking"); 265 assert(_dilatation_factor >= 1.0, "from previous assert"); 266 } 267 268 269 // The field "_initiating_occupancy" represents the occupancy percentage 270 // at which we trigger a new collection cycle. Unless explicitly specified 271 // via CMSInitiatingOccupancyFraction (argument "io" below), it 272 // is calculated by: 273 // 274 // Let "f" be MinHeapFreeRatio in 275 // 276 // _intiating_occupancy = 100-f + 277 // f * (CMSTriggerRatio/100) 278 // where CMSTriggerRatio is the argument "tr" below. 279 // 280 // That is, if we assume the heap is at its desired maximum occupancy at the 281 // end of a collection, we let CMSTriggerRatio of the (purported) free 282 // space be allocated before initiating a new collection cycle. 283 // 284 void ConcurrentMarkSweepGeneration::init_initiating_occupancy(intx io, uintx tr) { 285 assert(io <= 100 && tr <= 100, "Check the arguments"); 286 if (io >= 0) { 287 _initiating_occupancy = (double)io / 100.0; 288 } else { 289 _initiating_occupancy = ((100 - MinHeapFreeRatio) + 290 (double)(tr * MinHeapFreeRatio) / 100.0) 291 / 100.0; 292 } 293 } 294 295 void ConcurrentMarkSweepGeneration::ref_processor_init() { 296 assert(collector() != NULL, "no collector"); 297 collector()->ref_processor_init(); 298 } 299 300 void CMSCollector::ref_processor_init() { 301 if (_ref_processor == NULL) { 302 // Allocate and initialize a reference processor 303 _ref_processor = 304 new ReferenceProcessor(_span, // span 305 (ParallelGCThreads > 1) && ParallelRefProcEnabled, // mt processing 306 (int) ParallelGCThreads, // mt processing degree 307 _cmsGen->refs_discovery_is_mt(), // mt discovery 308 (int) MAX2(ConcGCThreads, ParallelGCThreads), // mt discovery degree 309 _cmsGen->refs_discovery_is_atomic(), // discovery is not atomic 310 &_is_alive_closure, // closure for liveness info 311 false); // next field updates do not need write barrier 312 // Initialize the _ref_processor field of CMSGen 313 _cmsGen->set_ref_processor(_ref_processor); 314 315 } 316 } 317 318 CMSAdaptiveSizePolicy* CMSCollector::size_policy() { 319 GenCollectedHeap* gch = GenCollectedHeap::heap(); 320 assert(gch->kind() == CollectedHeap::GenCollectedHeap, 321 "Wrong type of heap"); 322 CMSAdaptiveSizePolicy* sp = (CMSAdaptiveSizePolicy*) 323 gch->gen_policy()->size_policy(); 324 assert(sp->is_gc_cms_adaptive_size_policy(), 325 "Wrong type of size policy"); 326 return sp; 327 } 328 329 CMSGCAdaptivePolicyCounters* CMSCollector::gc_adaptive_policy_counters() { 330 CMSGCAdaptivePolicyCounters* results = 331 (CMSGCAdaptivePolicyCounters*) collector_policy()->counters(); 332 assert( 333 results->kind() == GCPolicyCounters::CMSGCAdaptivePolicyCountersKind, 334 "Wrong gc policy counter kind"); 335 return results; 336 } 337 338 339 void ConcurrentMarkSweepGeneration::initialize_performance_counters() { 340 341 const char* gen_name = "old"; 342 343 // Generation Counters - generation 1, 1 subspace 344 _gen_counters = new GenerationCounters(gen_name, 1, 1, &_virtual_space); 345 346 _space_counters = new GSpaceCounters(gen_name, 0, 347 _virtual_space.reserved_size(), 348 this, _gen_counters); 349 } 350 351 CMSStats::CMSStats(ConcurrentMarkSweepGeneration* cms_gen, unsigned int alpha): 352 _cms_gen(cms_gen) 353 { 354 assert(alpha <= 100, "bad value"); 355 _saved_alpha = alpha; 356 357 // Initialize the alphas to the bootstrap value of 100. 358 _gc0_alpha = _cms_alpha = 100; 359 360 _cms_begin_time.update(); 361 _cms_end_time.update(); 362 363 _gc0_duration = 0.0; 364 _gc0_period = 0.0; 365 _gc0_promoted = 0; 366 367 _cms_duration = 0.0; 368 _cms_period = 0.0; 369 _cms_allocated = 0; 370 371 _cms_used_at_gc0_begin = 0; 372 _cms_used_at_gc0_end = 0; 373 _allow_duty_cycle_reduction = false; 374 _valid_bits = 0; 375 _icms_duty_cycle = CMSIncrementalDutyCycle; 376 } 377 378 double CMSStats::cms_free_adjustment_factor(size_t free) const { 379 // TBD: CR 6909490 380 return 1.0; 381 } 382 383 void CMSStats::adjust_cms_free_adjustment_factor(bool fail, size_t free) { 384 } 385 386 // If promotion failure handling is on use 387 // the padded average size of the promotion for each 388 // young generation collection. 389 double CMSStats::time_until_cms_gen_full() const { 390 size_t cms_free = _cms_gen->cmsSpace()->free(); 391 GenCollectedHeap* gch = GenCollectedHeap::heap(); 392 size_t expected_promotion = MIN2(gch->get_gen(0)->capacity(), 393 (size_t) _cms_gen->gc_stats()->avg_promoted()->padded_average()); 394 if (cms_free > expected_promotion) { 395 // Start a cms collection if there isn't enough space to promote 396 // for the next minor collection. Use the padded average as 397 // a safety factor. 398 cms_free -= expected_promotion; 399 400 // Adjust by the safety factor. 401 double cms_free_dbl = (double)cms_free; 402 double cms_adjustment = (100.0 - CMSIncrementalSafetyFactor)/100.0; 403 // Apply a further correction factor which tries to adjust 404 // for recent occurance of concurrent mode failures. 405 cms_adjustment = cms_adjustment * cms_free_adjustment_factor(cms_free); 406 cms_free_dbl = cms_free_dbl * cms_adjustment; 407 408 if (PrintGCDetails && Verbose) { 409 gclog_or_tty->print_cr("CMSStats::time_until_cms_gen_full: cms_free " 410 SIZE_FORMAT " expected_promotion " SIZE_FORMAT, 411 cms_free, expected_promotion); 412 gclog_or_tty->print_cr(" cms_free_dbl %f cms_consumption_rate %f", 413 cms_free_dbl, cms_consumption_rate() + 1.0); 414 } 415 // Add 1 in case the consumption rate goes to zero. 416 return cms_free_dbl / (cms_consumption_rate() + 1.0); 417 } 418 return 0.0; 419 } 420 421 // Compare the duration of the cms collection to the 422 // time remaining before the cms generation is empty. 423 // Note that the time from the start of the cms collection 424 // to the start of the cms sweep (less than the total 425 // duration of the cms collection) can be used. This 426 // has been tried and some applications experienced 427 // promotion failures early in execution. This was 428 // possibly because the averages were not accurate 429 // enough at the beginning. 430 double CMSStats::time_until_cms_start() const { 431 // We add "gc0_period" to the "work" calculation 432 // below because this query is done (mostly) at the 433 // end of a scavenge, so we need to conservatively 434 // account for that much possible delay 435 // in the query so as to avoid concurrent mode failures 436 // due to starting the collection just a wee bit too 437 // late. 438 double work = cms_duration() + gc0_period(); 439 double deadline = time_until_cms_gen_full(); 440 // If a concurrent mode failure occurred recently, we want to be 441 // more conservative and halve our expected time_until_cms_gen_full() 442 if (work > deadline) { 443 if (Verbose && PrintGCDetails) { 444 gclog_or_tty->print( 445 " CMSCollector: collect because of anticipated promotion " 446 "before full %3.7f + %3.7f > %3.7f ", cms_duration(), 447 gc0_period(), time_until_cms_gen_full()); 448 } 449 return 0.0; 450 } 451 return work - deadline; 452 } 453 454 // Return a duty cycle based on old_duty_cycle and new_duty_cycle, limiting the 455 // amount of change to prevent wild oscillation. 456 unsigned int CMSStats::icms_damped_duty_cycle(unsigned int old_duty_cycle, 457 unsigned int new_duty_cycle) { 458 assert(old_duty_cycle <= 100, "bad input value"); 459 assert(new_duty_cycle <= 100, "bad input value"); 460 461 // Note: use subtraction with caution since it may underflow (values are 462 // unsigned). Addition is safe since we're in the range 0-100. 463 unsigned int damped_duty_cycle = new_duty_cycle; 464 if (new_duty_cycle < old_duty_cycle) { 465 const unsigned int largest_delta = MAX2(old_duty_cycle / 4, 5U); 466 if (new_duty_cycle + largest_delta < old_duty_cycle) { 467 damped_duty_cycle = old_duty_cycle - largest_delta; 468 } 469 } else if (new_duty_cycle > old_duty_cycle) { 470 const unsigned int largest_delta = MAX2(old_duty_cycle / 4, 15U); 471 if (new_duty_cycle > old_duty_cycle + largest_delta) { 472 damped_duty_cycle = MIN2(old_duty_cycle + largest_delta, 100U); 473 } 474 } 475 assert(damped_duty_cycle <= 100, "invalid duty cycle computed"); 476 477 if (CMSTraceIncrementalPacing) { 478 gclog_or_tty->print(" [icms_damped_duty_cycle(%d,%d) = %d] ", 479 old_duty_cycle, new_duty_cycle, damped_duty_cycle); 480 } 481 return damped_duty_cycle; 482 } 483 484 unsigned int CMSStats::icms_update_duty_cycle_impl() { 485 assert(CMSIncrementalPacing && valid(), 486 "should be handled in icms_update_duty_cycle()"); 487 488 double cms_time_so_far = cms_timer().seconds(); 489 double scaled_duration = cms_duration_per_mb() * _cms_used_at_gc0_end / M; 490 double scaled_duration_remaining = fabsd(scaled_duration - cms_time_so_far); 491 492 // Avoid division by 0. 493 double time_until_full = MAX2(time_until_cms_gen_full(), 0.01); 494 double duty_cycle_dbl = 100.0 * scaled_duration_remaining / time_until_full; 495 496 unsigned int new_duty_cycle = MIN2((unsigned int)duty_cycle_dbl, 100U); 497 if (new_duty_cycle > _icms_duty_cycle) { 498 // Avoid very small duty cycles (1 or 2); 0 is allowed. 499 if (new_duty_cycle > 2) { 500 _icms_duty_cycle = icms_damped_duty_cycle(_icms_duty_cycle, 501 new_duty_cycle); 502 } 503 } else if (_allow_duty_cycle_reduction) { 504 // The duty cycle is reduced only once per cms cycle (see record_cms_end()). 505 new_duty_cycle = icms_damped_duty_cycle(_icms_duty_cycle, new_duty_cycle); 506 // Respect the minimum duty cycle. 507 unsigned int min_duty_cycle = (unsigned int)CMSIncrementalDutyCycleMin; 508 _icms_duty_cycle = MAX2(new_duty_cycle, min_duty_cycle); 509 } 510 511 if (PrintGCDetails || CMSTraceIncrementalPacing) { 512 gclog_or_tty->print(" icms_dc=%d ", _icms_duty_cycle); 513 } 514 515 _allow_duty_cycle_reduction = false; 516 return _icms_duty_cycle; 517 } 518 519 #ifndef PRODUCT 520 void CMSStats::print_on(outputStream *st) const { 521 st->print(" gc0_alpha=%d,cms_alpha=%d", _gc0_alpha, _cms_alpha); 522 st->print(",gc0_dur=%g,gc0_per=%g,gc0_promo=" SIZE_FORMAT, 523 gc0_duration(), gc0_period(), gc0_promoted()); 524 st->print(",cms_dur=%g,cms_dur_per_mb=%g,cms_per=%g,cms_alloc=" SIZE_FORMAT, 525 cms_duration(), cms_duration_per_mb(), 526 cms_period(), cms_allocated()); 527 st->print(",cms_since_beg=%g,cms_since_end=%g", 528 cms_time_since_begin(), cms_time_since_end()); 529 st->print(",cms_used_beg=" SIZE_FORMAT ",cms_used_end=" SIZE_FORMAT, 530 _cms_used_at_gc0_begin, _cms_used_at_gc0_end); 531 if (CMSIncrementalMode) { 532 st->print(",dc=%d", icms_duty_cycle()); 533 } 534 535 if (valid()) { 536 st->print(",promo_rate=%g,cms_alloc_rate=%g", 537 promotion_rate(), cms_allocation_rate()); 538 st->print(",cms_consumption_rate=%g,time_until_full=%g", 539 cms_consumption_rate(), time_until_cms_gen_full()); 540 } 541 st->print(" "); 542 } 543 #endif // #ifndef PRODUCT 544 545 CMSCollector::CollectorState CMSCollector::_collectorState = 546 CMSCollector::Idling; 547 bool CMSCollector::_foregroundGCIsActive = false; 548 bool CMSCollector::_foregroundGCShouldWait = false; 549 550 CMSCollector::CMSCollector(ConcurrentMarkSweepGeneration* cmsGen, 551 CardTableRS* ct, 552 ConcurrentMarkSweepPolicy* cp): 553 _cmsGen(cmsGen), 554 _ct(ct), 555 _ref_processor(NULL), // will be set later 556 _conc_workers(NULL), // may be set later 557 _abort_preclean(false), 558 _start_sampling(false), 559 _between_prologue_and_epilogue(false), 560 _markBitMap(0, Mutex::leaf + 1, "CMS_markBitMap_lock"), 561 _modUnionTable((CardTableModRefBS::card_shift - LogHeapWordSize), 562 -1 /* lock-free */, "No_lock" /* dummy */), 563 _modUnionClosure(&_modUnionTable), 564 _modUnionClosurePar(&_modUnionTable), 565 // Adjust my span to cover old (cms) gen 566 _span(cmsGen->reserved()), 567 // Construct the is_alive_closure with _span & markBitMap 568 _is_alive_closure(_span, &_markBitMap), 569 _restart_addr(NULL), 570 _overflow_list(NULL), 571 _stats(cmsGen), 572 _eden_chunk_lock(new Mutex(Mutex::leaf + 1, "CMS_eden_chunk_lock", true)), 573 _eden_chunk_array(NULL), // may be set in ctor body 574 _eden_chunk_capacity(0), // -- ditto -- 575 _eden_chunk_index(0), // -- ditto -- 576 _survivor_plab_array(NULL), // -- ditto -- 577 _survivor_chunk_array(NULL), // -- ditto -- 578 _survivor_chunk_capacity(0), // -- ditto -- 579 _survivor_chunk_index(0), // -- ditto -- 580 _ser_pmc_preclean_ovflw(0), 581 _ser_kac_preclean_ovflw(0), 582 _ser_pmc_remark_ovflw(0), 583 _par_pmc_remark_ovflw(0), 584 _ser_kac_ovflw(0), 585 _par_kac_ovflw(0), 586 #ifndef PRODUCT 587 _num_par_pushes(0), 588 #endif 589 _collection_count_start(0), 590 _verifying(false), 591 _icms_start_limit(NULL), 592 _icms_stop_limit(NULL), 593 _verification_mark_bm(0, Mutex::leaf + 1, "CMS_verification_mark_bm_lock"), 594 _completed_initialization(false), 595 _collector_policy(cp), 596 _should_unload_classes(false), 597 _concurrent_cycles_since_last_unload(0), 598 _roots_scanning_options(0), 599 _inter_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding), 600 _intra_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding), 601 _gc_tracer_cm(new (ResourceObj::C_HEAP, mtGC) CMSTracer()), 602 _gc_timer_cm(new (ResourceObj::C_HEAP, mtGC) ConcurrentGCTimer()), 603 _cms_start_registered(false) 604 { 605 if (ExplicitGCInvokesConcurrentAndUnloadsClasses) { 606 ExplicitGCInvokesConcurrent = true; 607 } 608 // Now expand the span and allocate the collection support structures 609 // (MUT, marking bit map etc.) to cover both generations subject to 610 // collection. 611 612 // For use by dirty card to oop closures. 613 _cmsGen->cmsSpace()->set_collector(this); 614 615 // Allocate MUT and marking bit map 616 { 617 MutexLockerEx x(_markBitMap.lock(), Mutex::_no_safepoint_check_flag); 618 if (!_markBitMap.allocate(_span)) { 619 warning("Failed to allocate CMS Bit Map"); 620 return; 621 } 622 assert(_markBitMap.covers(_span), "_markBitMap inconsistency?"); 623 } 624 { 625 _modUnionTable.allocate(_span); 626 assert(_modUnionTable.covers(_span), "_modUnionTable inconsistency?"); 627 } 628 629 if (!_markStack.allocate(MarkStackSize)) { 630 warning("Failed to allocate CMS Marking Stack"); 631 return; 632 } 633 634 // Support for multi-threaded concurrent phases 635 if (CMSConcurrentMTEnabled) { 636 if (FLAG_IS_DEFAULT(ConcGCThreads)) { 637 // just for now 638 FLAG_SET_DEFAULT(ConcGCThreads, (ParallelGCThreads + 3)/4); 639 } 640 if (ConcGCThreads > 1) { 641 _conc_workers = new YieldingFlexibleWorkGang("Parallel CMS Threads", 642 ConcGCThreads, true); 643 if (_conc_workers == NULL) { 644 warning("GC/CMS: _conc_workers allocation failure: " 645 "forcing -CMSConcurrentMTEnabled"); 646 CMSConcurrentMTEnabled = false; 647 } else { 648 _conc_workers->initialize_workers(); 649 } 650 } else { 651 CMSConcurrentMTEnabled = false; 652 } 653 } 654 if (!CMSConcurrentMTEnabled) { 655 ConcGCThreads = 0; 656 } else { 657 // Turn off CMSCleanOnEnter optimization temporarily for 658 // the MT case where it's not fixed yet; see 6178663. 659 CMSCleanOnEnter = false; 660 } 661 assert((_conc_workers != NULL) == (ConcGCThreads > 1), 662 "Inconsistency"); 663 664 // Parallel task queues; these are shared for the 665 // concurrent and stop-world phases of CMS, but 666 // are not shared with parallel scavenge (ParNew). 667 { 668 uint i; 669 uint num_queues = (uint) MAX2(ParallelGCThreads, ConcGCThreads); 670 671 if ((CMSParallelRemarkEnabled || CMSConcurrentMTEnabled 672 || ParallelRefProcEnabled) 673 && num_queues > 0) { 674 _task_queues = new OopTaskQueueSet(num_queues); 675 if (_task_queues == NULL) { 676 warning("task_queues allocation failure."); 677 return; 678 } 679 _hash_seed = NEW_C_HEAP_ARRAY(int, num_queues, mtGC); 680 if (_hash_seed == NULL) { 681 warning("_hash_seed array allocation failure"); 682 return; 683 } 684 685 typedef Padded<OopTaskQueue> PaddedOopTaskQueue; 686 for (i = 0; i < num_queues; i++) { 687 PaddedOopTaskQueue *q = new PaddedOopTaskQueue(); 688 if (q == NULL) { 689 warning("work_queue allocation failure."); 690 return; 691 } 692 _task_queues->register_queue(i, q); 693 } 694 for (i = 0; i < num_queues; i++) { 695 _task_queues->queue(i)->initialize(); 696 _hash_seed[i] = 17; // copied from ParNew 697 } 698 } 699 } 700 701 _cmsGen ->init_initiating_occupancy(CMSInitiatingOccupancyFraction, CMSTriggerRatio); 702 703 // Clip CMSBootstrapOccupancy between 0 and 100. 704 _bootstrap_occupancy = ((double)CMSBootstrapOccupancy)/(double)100; 705 706 _full_gcs_since_conc_gc = 0; 707 708 // Now tell CMS generations the identity of their collector 709 ConcurrentMarkSweepGeneration::set_collector(this); 710 711 // Create & start a CMS thread for this CMS collector 712 _cmsThread = ConcurrentMarkSweepThread::start(this); 713 assert(cmsThread() != NULL, "CMS Thread should have been created"); 714 assert(cmsThread()->collector() == this, 715 "CMS Thread should refer to this gen"); 716 assert(CGC_lock != NULL, "Where's the CGC_lock?"); 717 718 // Support for parallelizing young gen rescan 719 GenCollectedHeap* gch = GenCollectedHeap::heap(); 720 _young_gen = gch->prev_gen(_cmsGen); 721 if (gch->supports_inline_contig_alloc()) { 722 _top_addr = gch->top_addr(); 723 _end_addr = gch->end_addr(); 724 assert(_young_gen != NULL, "no _young_gen"); 725 _eden_chunk_index = 0; 726 _eden_chunk_capacity = (_young_gen->max_capacity()+CMSSamplingGrain)/CMSSamplingGrain; 727 _eden_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, _eden_chunk_capacity, mtGC); 728 if (_eden_chunk_array == NULL) { 729 _eden_chunk_capacity = 0; 730 warning("GC/CMS: _eden_chunk_array allocation failure"); 731 } 732 } 733 assert(_eden_chunk_array != NULL || _eden_chunk_capacity == 0, "Error"); 734 735 // Support for parallelizing survivor space rescan 736 if (CMSParallelRemarkEnabled && CMSParallelSurvivorRemarkEnabled) { 737 const size_t max_plab_samples = 738 ((DefNewGeneration*)_young_gen)->max_survivor_size()/MinTLABSize; 739 740 _survivor_plab_array = NEW_C_HEAP_ARRAY(ChunkArray, ParallelGCThreads, mtGC); 741 _survivor_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, 2*max_plab_samples, mtGC); 742 _cursor = NEW_C_HEAP_ARRAY(size_t, ParallelGCThreads, mtGC); 743 if (_survivor_plab_array == NULL || _survivor_chunk_array == NULL 744 || _cursor == NULL) { 745 warning("Failed to allocate survivor plab/chunk array"); 746 if (_survivor_plab_array != NULL) { 747 FREE_C_HEAP_ARRAY(ChunkArray, _survivor_plab_array, mtGC); 748 _survivor_plab_array = NULL; 749 } 750 if (_survivor_chunk_array != NULL) { 751 FREE_C_HEAP_ARRAY(HeapWord*, _survivor_chunk_array, mtGC); 752 _survivor_chunk_array = NULL; 753 } 754 if (_cursor != NULL) { 755 FREE_C_HEAP_ARRAY(size_t, _cursor, mtGC); 756 _cursor = NULL; 757 } 758 } else { 759 _survivor_chunk_capacity = 2*max_plab_samples; 760 for (uint i = 0; i < ParallelGCThreads; i++) { 761 HeapWord** vec = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples, mtGC); 762 if (vec == NULL) { 763 warning("Failed to allocate survivor plab array"); 764 for (int j = i; j > 0; j--) { 765 FREE_C_HEAP_ARRAY(HeapWord*, _survivor_plab_array[j-1].array(), mtGC); 766 } 767 FREE_C_HEAP_ARRAY(ChunkArray, _survivor_plab_array, mtGC); 768 FREE_C_HEAP_ARRAY(HeapWord*, _survivor_chunk_array, mtGC); 769 _survivor_plab_array = NULL; 770 _survivor_chunk_array = NULL; 771 _survivor_chunk_capacity = 0; 772 break; 773 } else { 774 ChunkArray* cur = 775 ::new (&_survivor_plab_array[i]) ChunkArray(vec, 776 max_plab_samples); 777 assert(cur->end() == 0, "Should be 0"); 778 assert(cur->array() == vec, "Should be vec"); 779 assert(cur->capacity() == max_plab_samples, "Error"); 780 } 781 } 782 } 783 } 784 assert( ( _survivor_plab_array != NULL 785 && _survivor_chunk_array != NULL) 786 || ( _survivor_chunk_capacity == 0 787 && _survivor_chunk_index == 0), 788 "Error"); 789 790 // Choose what strong roots should be scanned depending on verification options 791 if (!CMSClassUnloadingEnabled) { 792 // If class unloading is disabled we want to include all classes into the root set. 793 add_root_scanning_option(SharedHeap::SO_AllClasses); 794 } else { 795 add_root_scanning_option(SharedHeap::SO_SystemClasses); 796 } 797 798 NOT_PRODUCT(_overflow_counter = CMSMarkStackOverflowInterval;) 799 _gc_counters = new CollectorCounters("CMS", 1); 800 _completed_initialization = true; 801 _inter_sweep_timer.start(); // start of time 802 } 803 804 const char* ConcurrentMarkSweepGeneration::name() const { 805 return "concurrent mark-sweep generation"; 806 } 807 void ConcurrentMarkSweepGeneration::update_counters() { 808 if (UsePerfData) { 809 _space_counters->update_all(); 810 _gen_counters->update_all(); 811 } 812 } 813 814 // this is an optimized version of update_counters(). it takes the 815 // used value as a parameter rather than computing it. 816 // 817 void ConcurrentMarkSweepGeneration::update_counters(size_t used) { 818 if (UsePerfData) { 819 _space_counters->update_used(used); 820 _space_counters->update_capacity(); 821 _gen_counters->update_all(); 822 } 823 } 824 825 void ConcurrentMarkSweepGeneration::print() const { 826 Generation::print(); 827 cmsSpace()->print(); 828 } 829 830 #ifndef PRODUCT 831 void ConcurrentMarkSweepGeneration::print_statistics() { 832 cmsSpace()->printFLCensus(0); 833 } 834 #endif 835 836 void ConcurrentMarkSweepGeneration::printOccupancy(const char *s) { 837 GenCollectedHeap* gch = GenCollectedHeap::heap(); 838 if (PrintGCDetails) { 839 if (Verbose) { 840 gclog_or_tty->print("[%d %s-%s: "SIZE_FORMAT"("SIZE_FORMAT")]", 841 level(), short_name(), s, used(), capacity()); 842 } else { 843 gclog_or_tty->print("[%d %s-%s: "SIZE_FORMAT"K("SIZE_FORMAT"K)]", 844 level(), short_name(), s, used() / K, capacity() / K); 845 } 846 } 847 if (Verbose) { 848 gclog_or_tty->print(" "SIZE_FORMAT"("SIZE_FORMAT")", 849 gch->used(), gch->capacity()); 850 } else { 851 gclog_or_tty->print(" "SIZE_FORMAT"K("SIZE_FORMAT"K)", 852 gch->used() / K, gch->capacity() / K); 853 } 854 } 855 856 size_t 857 ConcurrentMarkSweepGeneration::contiguous_available() const { 858 // dld proposes an improvement in precision here. If the committed 859 // part of the space ends in a free block we should add that to 860 // uncommitted size in the calculation below. Will make this 861 // change later, staying with the approximation below for the 862 // time being. -- ysr. 863 return MAX2(_virtual_space.uncommitted_size(), unsafe_max_alloc_nogc()); 864 } 865 866 size_t 867 ConcurrentMarkSweepGeneration::unsafe_max_alloc_nogc() const { 868 return _cmsSpace->max_alloc_in_words() * HeapWordSize; 869 } 870 871 size_t ConcurrentMarkSweepGeneration::max_available() const { 872 return free() + _virtual_space.uncommitted_size(); 873 } 874 875 bool ConcurrentMarkSweepGeneration::promotion_attempt_is_safe(size_t max_promotion_in_bytes) const { 876 size_t available = max_available(); 877 size_t av_promo = (size_t)gc_stats()->avg_promoted()->padded_average(); 878 bool res = (available >= av_promo) || (available >= max_promotion_in_bytes); 879 if (Verbose && PrintGCDetails) { 880 gclog_or_tty->print_cr( 881 "CMS: promo attempt is%s safe: available("SIZE_FORMAT") %s av_promo("SIZE_FORMAT")," 882 "max_promo("SIZE_FORMAT")", 883 res? "":" not", available, res? ">=":"<", 884 av_promo, max_promotion_in_bytes); 885 } 886 return res; 887 } 888 889 // At a promotion failure dump information on block layout in heap 890 // (cms old generation). 891 void ConcurrentMarkSweepGeneration::promotion_failure_occurred() { 892 if (CMSDumpAtPromotionFailure) { 893 cmsSpace()->dump_at_safepoint_with_locks(collector(), gclog_or_tty); 894 } 895 } 896 897 CompactibleSpace* 898 ConcurrentMarkSweepGeneration::first_compaction_space() const { 899 return _cmsSpace; 900 } 901 902 void ConcurrentMarkSweepGeneration::reset_after_compaction() { 903 // Clear the promotion information. These pointers can be adjusted 904 // along with all the other pointers into the heap but 905 // compaction is expected to be a rare event with 906 // a heap using cms so don't do it without seeing the need. 907 if (CollectedHeap::use_parallel_gc_threads()) { 908 for (uint i = 0; i < ParallelGCThreads; i++) { 909 _par_gc_thread_states[i]->promo.reset(); 910 } 911 } 912 } 913 914 void ConcurrentMarkSweepGeneration::space_iterate(SpaceClosure* blk, bool usedOnly) { 915 blk->do_space(_cmsSpace); 916 } 917 918 void ConcurrentMarkSweepGeneration::compute_new_size() { 919 assert_locked_or_safepoint(Heap_lock); 920 921 // If incremental collection failed, we just want to expand 922 // to the limit. 923 if (incremental_collection_failed()) { 924 clear_incremental_collection_failed(); 925 grow_to_reserved(); 926 return; 927 } 928 929 // The heap has been compacted but not reset yet. 930 // Any metric such as free() or used() will be incorrect. 931 932 CardGeneration::compute_new_size(); 933 934 // Reset again after a possible resizing 935 if (did_compact()) { 936 cmsSpace()->reset_after_compaction(); 937 } 938 } 939 940 void ConcurrentMarkSweepGeneration::compute_new_size_free_list() { 941 assert_locked_or_safepoint(Heap_lock); 942 943 // If incremental collection failed, we just want to expand 944 // to the limit. 945 if (incremental_collection_failed()) { 946 clear_incremental_collection_failed(); 947 grow_to_reserved(); 948 return; 949 } 950 951 double free_percentage = ((double) free()) / capacity(); 952 double desired_free_percentage = (double) MinHeapFreeRatio / 100; 953 double maximum_free_percentage = (double) MaxHeapFreeRatio / 100; 954 955 // compute expansion delta needed for reaching desired free percentage 956 if (free_percentage < desired_free_percentage) { 957 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 958 assert(desired_capacity >= capacity(), "invalid expansion size"); 959 size_t expand_bytes = MAX2(desired_capacity - capacity(), MinHeapDeltaBytes); 960 if (PrintGCDetails && Verbose) { 961 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 962 gclog_or_tty->print_cr("\nFrom compute_new_size: "); 963 gclog_or_tty->print_cr(" Free fraction %f", free_percentage); 964 gclog_or_tty->print_cr(" Desired free fraction %f", 965 desired_free_percentage); 966 gclog_or_tty->print_cr(" Maximum free fraction %f", 967 maximum_free_percentage); 968 gclog_or_tty->print_cr(" Capactiy "SIZE_FORMAT, capacity()/1000); 969 gclog_or_tty->print_cr(" Desired capacity "SIZE_FORMAT, 970 desired_capacity/1000); 971 int prev_level = level() - 1; 972 if (prev_level >= 0) { 973 size_t prev_size = 0; 974 GenCollectedHeap* gch = GenCollectedHeap::heap(); 975 Generation* prev_gen = gch->_gens[prev_level]; 976 prev_size = prev_gen->capacity(); 977 gclog_or_tty->print_cr(" Younger gen size "SIZE_FORMAT, 978 prev_size/1000); 979 } 980 gclog_or_tty->print_cr(" unsafe_max_alloc_nogc "SIZE_FORMAT, 981 unsafe_max_alloc_nogc()/1000); 982 gclog_or_tty->print_cr(" contiguous available "SIZE_FORMAT, 983 contiguous_available()/1000); 984 gclog_or_tty->print_cr(" Expand by "SIZE_FORMAT" (bytes)", 985 expand_bytes); 986 } 987 // safe if expansion fails 988 expand(expand_bytes, 0, CMSExpansionCause::_satisfy_free_ratio); 989 if (PrintGCDetails && Verbose) { 990 gclog_or_tty->print_cr(" Expanded free fraction %f", 991 ((double) free()) / capacity()); 992 } 993 } else { 994 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 995 assert(desired_capacity <= capacity(), "invalid expansion size"); 996 size_t shrink_bytes = capacity() - desired_capacity; 997 // Don't shrink unless the delta is greater than the minimum shrink we want 998 if (shrink_bytes >= MinHeapDeltaBytes) { 999 shrink_free_list_by(shrink_bytes); 1000 } 1001 } 1002 } 1003 1004 Mutex* ConcurrentMarkSweepGeneration::freelistLock() const { 1005 return cmsSpace()->freelistLock(); 1006 } 1007 1008 HeapWord* ConcurrentMarkSweepGeneration::allocate(size_t size, 1009 bool tlab) { 1010 CMSSynchronousYieldRequest yr; 1011 MutexLockerEx x(freelistLock(), 1012 Mutex::_no_safepoint_check_flag); 1013 return have_lock_and_allocate(size, tlab); 1014 } 1015 1016 HeapWord* ConcurrentMarkSweepGeneration::have_lock_and_allocate(size_t size, 1017 bool tlab /* ignored */) { 1018 assert_lock_strong(freelistLock()); 1019 size_t adjustedSize = CompactibleFreeListSpace::adjustObjectSize(size); 1020 HeapWord* res = cmsSpace()->allocate(adjustedSize); 1021 // Allocate the object live (grey) if the background collector has 1022 // started marking. This is necessary because the marker may 1023 // have passed this address and consequently this object will 1024 // not otherwise be greyed and would be incorrectly swept up. 1025 // Note that if this object contains references, the writing 1026 // of those references will dirty the card containing this object 1027 // allowing the object to be blackened (and its references scanned) 1028 // either during a preclean phase or at the final checkpoint. 1029 if (res != NULL) { 1030 // We may block here with an uninitialized object with 1031 // its mark-bit or P-bits not yet set. Such objects need 1032 // to be safely navigable by block_start(). 1033 assert(oop(res)->klass_or_null() == NULL, "Object should be uninitialized here."); 1034 assert(!((FreeChunk*)res)->is_free(), "Error, block will look free but show wrong size"); 1035 collector()->direct_allocated(res, adjustedSize); 1036 _direct_allocated_words += adjustedSize; 1037 // allocation counters 1038 NOT_PRODUCT( 1039 _numObjectsAllocated++; 1040 _numWordsAllocated += (int)adjustedSize; 1041 ) 1042 } 1043 return res; 1044 } 1045 1046 // In the case of direct allocation by mutators in a generation that 1047 // is being concurrently collected, the object must be allocated 1048 // live (grey) if the background collector has started marking. 1049 // This is necessary because the marker may 1050 // have passed this address and consequently this object will 1051 // not otherwise be greyed and would be incorrectly swept up. 1052 // Note that if this object contains references, the writing 1053 // of those references will dirty the card containing this object 1054 // allowing the object to be blackened (and its references scanned) 1055 // either during a preclean phase or at the final checkpoint. 1056 void CMSCollector::direct_allocated(HeapWord* start, size_t size) { 1057 assert(_markBitMap.covers(start, size), "Out of bounds"); 1058 if (_collectorState >= Marking) { 1059 MutexLockerEx y(_markBitMap.lock(), 1060 Mutex::_no_safepoint_check_flag); 1061 // [see comments preceding SweepClosure::do_blk() below for details] 1062 // 1063 // Can the P-bits be deleted now? JJJ 1064 // 1065 // 1. need to mark the object as live so it isn't collected 1066 // 2. need to mark the 2nd bit to indicate the object may be uninitialized 1067 // 3. need to mark the end of the object so marking, precleaning or sweeping 1068 // can skip over uninitialized or unparsable objects. An allocated 1069 // object is considered uninitialized for our purposes as long as 1070 // its klass word is NULL. All old gen objects are parsable 1071 // as soon as they are initialized.) 1072 _markBitMap.mark(start); // object is live 1073 _markBitMap.mark(start + 1); // object is potentially uninitialized? 1074 _markBitMap.mark(start + size - 1); 1075 // mark end of object 1076 } 1077 // check that oop looks uninitialized 1078 assert(oop(start)->klass_or_null() == NULL, "_klass should be NULL"); 1079 } 1080 1081 void CMSCollector::promoted(bool par, HeapWord* start, 1082 bool is_obj_array, size_t obj_size) { 1083 assert(_markBitMap.covers(start), "Out of bounds"); 1084 // See comment in direct_allocated() about when objects should 1085 // be allocated live. 1086 if (_collectorState >= Marking) { 1087 // we already hold the marking bit map lock, taken in 1088 // the prologue 1089 if (par) { 1090 _markBitMap.par_mark(start); 1091 } else { 1092 _markBitMap.mark(start); 1093 } 1094 // We don't need to mark the object as uninitialized (as 1095 // in direct_allocated above) because this is being done with the 1096 // world stopped and the object will be initialized by the 1097 // time the marking, precleaning or sweeping get to look at it. 1098 // But see the code for copying objects into the CMS generation, 1099 // where we need to ensure that concurrent readers of the 1100 // block offset table are able to safely navigate a block that 1101 // is in flux from being free to being allocated (and in 1102 // transition while being copied into) and subsequently 1103 // becoming a bona-fide object when the copy/promotion is complete. 1104 assert(SafepointSynchronize::is_at_safepoint(), 1105 "expect promotion only at safepoints"); 1106 1107 if (_collectorState < Sweeping) { 1108 // Mark the appropriate cards in the modUnionTable, so that 1109 // this object gets scanned before the sweep. If this is 1110 // not done, CMS generation references in the object might 1111 // not get marked. 1112 // For the case of arrays, which are otherwise precisely 1113 // marked, we need to dirty the entire array, not just its head. 1114 if (is_obj_array) { 1115 // The [par_]mark_range() method expects mr.end() below to 1116 // be aligned to the granularity of a bit's representation 1117 // in the heap. In the case of the MUT below, that's a 1118 // card size. 1119 MemRegion mr(start, 1120 (HeapWord*)round_to((intptr_t)(start + obj_size), 1121 CardTableModRefBS::card_size /* bytes */)); 1122 if (par) { 1123 _modUnionTable.par_mark_range(mr); 1124 } else { 1125 _modUnionTable.mark_range(mr); 1126 } 1127 } else { // not an obj array; we can just mark the head 1128 if (par) { 1129 _modUnionTable.par_mark(start); 1130 } else { 1131 _modUnionTable.mark(start); 1132 } 1133 } 1134 } 1135 } 1136 } 1137 1138 static inline size_t percent_of_space(Space* space, HeapWord* addr) 1139 { 1140 size_t delta = pointer_delta(addr, space->bottom()); 1141 return (size_t)(delta * 100.0 / (space->capacity() / HeapWordSize)); 1142 } 1143 1144 void CMSCollector::icms_update_allocation_limits() 1145 { 1146 Generation* gen0 = GenCollectedHeap::heap()->get_gen(0); 1147 EdenSpace* eden = gen0->as_DefNewGeneration()->eden(); 1148 1149 const unsigned int duty_cycle = stats().icms_update_duty_cycle(); 1150 if (CMSTraceIncrementalPacing) { 1151 stats().print(); 1152 } 1153 1154 assert(duty_cycle <= 100, "invalid duty cycle"); 1155 if (duty_cycle != 0) { 1156 // The duty_cycle is a percentage between 0 and 100; convert to words and 1157 // then compute the offset from the endpoints of the space. 1158 size_t free_words = eden->free() / HeapWordSize; 1159 double free_words_dbl = (double)free_words; 1160 size_t duty_cycle_words = (size_t)(free_words_dbl * duty_cycle / 100.0); 1161 size_t offset_words = (free_words - duty_cycle_words) / 2; 1162 1163 _icms_start_limit = eden->top() + offset_words; 1164 _icms_stop_limit = eden->end() - offset_words; 1165 1166 // The limits may be adjusted (shifted to the right) by 1167 // CMSIncrementalOffset, to allow the application more mutator time after a 1168 // young gen gc (when all mutators were stopped) and before CMS starts and 1169 // takes away one or more cpus. 1170 if (CMSIncrementalOffset != 0) { 1171 double adjustment_dbl = free_words_dbl * CMSIncrementalOffset / 100.0; 1172 size_t adjustment = (size_t)adjustment_dbl; 1173 HeapWord* tmp_stop = _icms_stop_limit + adjustment; 1174 if (tmp_stop > _icms_stop_limit && tmp_stop < eden->end()) { 1175 _icms_start_limit += adjustment; 1176 _icms_stop_limit = tmp_stop; 1177 } 1178 } 1179 } 1180 if (duty_cycle == 0 || (_icms_start_limit == _icms_stop_limit)) { 1181 _icms_start_limit = _icms_stop_limit = eden->end(); 1182 } 1183 1184 // Install the new start limit. 1185 eden->set_soft_end(_icms_start_limit); 1186 1187 if (CMSTraceIncrementalMode) { 1188 gclog_or_tty->print(" icms alloc limits: " 1189 PTR_FORMAT "," PTR_FORMAT 1190 " (" SIZE_FORMAT "%%," SIZE_FORMAT "%%) ", 1191 _icms_start_limit, _icms_stop_limit, 1192 percent_of_space(eden, _icms_start_limit), 1193 percent_of_space(eden, _icms_stop_limit)); 1194 if (Verbose) { 1195 gclog_or_tty->print("eden: "); 1196 eden->print_on(gclog_or_tty); 1197 } 1198 } 1199 } 1200 1201 // Any changes here should try to maintain the invariant 1202 // that if this method is called with _icms_start_limit 1203 // and _icms_stop_limit both NULL, then it should return NULL 1204 // and not notify the icms thread. 1205 HeapWord* 1206 CMSCollector::allocation_limit_reached(Space* space, HeapWord* top, 1207 size_t word_size) 1208 { 1209 // A start_limit equal to end() means the duty cycle is 0, so treat that as a 1210 // nop. 1211 if (CMSIncrementalMode && _icms_start_limit != space->end()) { 1212 if (top <= _icms_start_limit) { 1213 if (CMSTraceIncrementalMode) { 1214 space->print_on(gclog_or_tty); 1215 gclog_or_tty->stamp(); 1216 gclog_or_tty->print_cr(" start limit top=" PTR_FORMAT 1217 ", new limit=" PTR_FORMAT 1218 " (" SIZE_FORMAT "%%)", 1219 top, _icms_stop_limit, 1220 percent_of_space(space, _icms_stop_limit)); 1221 } 1222 ConcurrentMarkSweepThread::start_icms(); 1223 assert(top < _icms_stop_limit, "Tautology"); 1224 if (word_size < pointer_delta(_icms_stop_limit, top)) { 1225 return _icms_stop_limit; 1226 } 1227 1228 // The allocation will cross both the _start and _stop limits, so do the 1229 // stop notification also and return end(). 1230 if (CMSTraceIncrementalMode) { 1231 space->print_on(gclog_or_tty); 1232 gclog_or_tty->stamp(); 1233 gclog_or_tty->print_cr(" +stop limit top=" PTR_FORMAT 1234 ", new limit=" PTR_FORMAT 1235 " (" SIZE_FORMAT "%%)", 1236 top, space->end(), 1237 percent_of_space(space, space->end())); 1238 } 1239 ConcurrentMarkSweepThread::stop_icms(); 1240 return space->end(); 1241 } 1242 1243 if (top <= _icms_stop_limit) { 1244 if (CMSTraceIncrementalMode) { 1245 space->print_on(gclog_or_tty); 1246 gclog_or_tty->stamp(); 1247 gclog_or_tty->print_cr(" stop limit top=" PTR_FORMAT 1248 ", new limit=" PTR_FORMAT 1249 " (" SIZE_FORMAT "%%)", 1250 top, space->end(), 1251 percent_of_space(space, space->end())); 1252 } 1253 ConcurrentMarkSweepThread::stop_icms(); 1254 return space->end(); 1255 } 1256 1257 if (CMSTraceIncrementalMode) { 1258 space->print_on(gclog_or_tty); 1259 gclog_or_tty->stamp(); 1260 gclog_or_tty->print_cr(" end limit top=" PTR_FORMAT 1261 ", new limit=" PTR_FORMAT, 1262 top, NULL); 1263 } 1264 } 1265 1266 return NULL; 1267 } 1268 1269 oop ConcurrentMarkSweepGeneration::promote(oop obj, size_t obj_size) { 1270 assert(obj_size == (size_t)obj->size(), "bad obj_size passed in"); 1271 // allocate, copy and if necessary update promoinfo -- 1272 // delegate to underlying space. 1273 assert_lock_strong(freelistLock()); 1274 1275 #ifndef PRODUCT 1276 if (Universe::heap()->promotion_should_fail()) { 1277 return NULL; 1278 } 1279 #endif // #ifndef PRODUCT 1280 1281 oop res = _cmsSpace->promote(obj, obj_size); 1282 if (res == NULL) { 1283 // expand and retry 1284 size_t s = _cmsSpace->expansionSpaceRequired(obj_size); // HeapWords 1285 expand(s*HeapWordSize, MinHeapDeltaBytes, 1286 CMSExpansionCause::_satisfy_promotion); 1287 // Since there's currently no next generation, we don't try to promote 1288 // into a more senior generation. 1289 assert(next_gen() == NULL, "assumption, based upon which no attempt " 1290 "is made to pass on a possibly failing " 1291 "promotion to next generation"); 1292 res = _cmsSpace->promote(obj, obj_size); 1293 } 1294 if (res != NULL) { 1295 // See comment in allocate() about when objects should 1296 // be allocated live. 1297 assert(obj->is_oop(), "Will dereference klass pointer below"); 1298 collector()->promoted(false, // Not parallel 1299 (HeapWord*)res, obj->is_objArray(), obj_size); 1300 // promotion counters 1301 NOT_PRODUCT( 1302 _numObjectsPromoted++; 1303 _numWordsPromoted += 1304 (int)(CompactibleFreeListSpace::adjustObjectSize(obj->size())); 1305 ) 1306 } 1307 return res; 1308 } 1309 1310 1311 HeapWord* 1312 ConcurrentMarkSweepGeneration::allocation_limit_reached(Space* space, 1313 HeapWord* top, 1314 size_t word_sz) 1315 { 1316 return collector()->allocation_limit_reached(space, top, word_sz); 1317 } 1318 1319 // IMPORTANT: Notes on object size recognition in CMS. 1320 // --------------------------------------------------- 1321 // A block of storage in the CMS generation is always in 1322 // one of three states. A free block (FREE), an allocated 1323 // object (OBJECT) whose size() method reports the correct size, 1324 // and an intermediate state (TRANSIENT) in which its size cannot 1325 // be accurately determined. 1326 // STATE IDENTIFICATION: (32 bit and 64 bit w/o COOPS) 1327 // ----------------------------------------------------- 1328 // FREE: klass_word & 1 == 1; mark_word holds block size 1329 // 1330 // OBJECT: klass_word installed; klass_word != 0 && klass_word & 1 == 0; 1331 // obj->size() computes correct size 1332 // 1333 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT 1334 // 1335 // STATE IDENTIFICATION: (64 bit+COOPS) 1336 // ------------------------------------ 1337 // FREE: mark_word & CMS_FREE_BIT == 1; mark_word & ~CMS_FREE_BIT gives block_size 1338 // 1339 // OBJECT: klass_word installed; klass_word != 0; 1340 // obj->size() computes correct size 1341 // 1342 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT 1343 // 1344 // 1345 // STATE TRANSITION DIAGRAM 1346 // 1347 // mut / parnew mut / parnew 1348 // FREE --------------------> TRANSIENT ---------------------> OBJECT --| 1349 // ^ | 1350 // |------------------------ DEAD <------------------------------------| 1351 // sweep mut 1352 // 1353 // While a block is in TRANSIENT state its size cannot be determined 1354 // so readers will either need to come back later or stall until 1355 // the size can be determined. Note that for the case of direct 1356 // allocation, P-bits, when available, may be used to determine the 1357 // size of an object that may not yet have been initialized. 1358 1359 // Things to support parallel young-gen collection. 1360 oop 1361 ConcurrentMarkSweepGeneration::par_promote(int thread_num, 1362 oop old, markOop m, 1363 size_t word_sz) { 1364 #ifndef PRODUCT 1365 if (Universe::heap()->promotion_should_fail()) { 1366 return NULL; 1367 } 1368 #endif // #ifndef PRODUCT 1369 1370 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1371 PromotionInfo* promoInfo = &ps->promo; 1372 // if we are tracking promotions, then first ensure space for 1373 // promotion (including spooling space for saving header if necessary). 1374 // then allocate and copy, then track promoted info if needed. 1375 // When tracking (see PromotionInfo::track()), the mark word may 1376 // be displaced and in this case restoration of the mark word 1377 // occurs in the (oop_since_save_marks_)iterate phase. 1378 if (promoInfo->tracking() && !promoInfo->ensure_spooling_space()) { 1379 // Out of space for allocating spooling buffers; 1380 // try expanding and allocating spooling buffers. 1381 if (!expand_and_ensure_spooling_space(promoInfo)) { 1382 return NULL; 1383 } 1384 } 1385 assert(promoInfo->has_spooling_space(), "Control point invariant"); 1386 const size_t alloc_sz = CompactibleFreeListSpace::adjustObjectSize(word_sz); 1387 HeapWord* obj_ptr = ps->lab.alloc(alloc_sz); 1388 if (obj_ptr == NULL) { 1389 obj_ptr = expand_and_par_lab_allocate(ps, alloc_sz); 1390 if (obj_ptr == NULL) { 1391 return NULL; 1392 } 1393 } 1394 oop obj = oop(obj_ptr); 1395 OrderAccess::storestore(); 1396 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1397 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1398 // IMPORTANT: See note on object initialization for CMS above. 1399 // Otherwise, copy the object. Here we must be careful to insert the 1400 // klass pointer last, since this marks the block as an allocated object. 1401 // Except with compressed oops it's the mark word. 1402 HeapWord* old_ptr = (HeapWord*)old; 1403 // Restore the mark word copied above. 1404 obj->set_mark(m); 1405 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1406 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1407 OrderAccess::storestore(); 1408 1409 if (UseCompressedKlassPointers) { 1410 // Copy gap missed by (aligned) header size calculation below 1411 obj->set_klass_gap(old->klass_gap()); 1412 } 1413 if (word_sz > (size_t)oopDesc::header_size()) { 1414 Copy::aligned_disjoint_words(old_ptr + oopDesc::header_size(), 1415 obj_ptr + oopDesc::header_size(), 1416 word_sz - oopDesc::header_size()); 1417 } 1418 1419 // Now we can track the promoted object, if necessary. We take care 1420 // to delay the transition from uninitialized to full object 1421 // (i.e., insertion of klass pointer) until after, so that it 1422 // atomically becomes a promoted object. 1423 if (promoInfo->tracking()) { 1424 promoInfo->track((PromotedObject*)obj, old->klass()); 1425 } 1426 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1427 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1428 assert(old->is_oop(), "Will use and dereference old klass ptr below"); 1429 1430 // Finally, install the klass pointer (this should be volatile). 1431 OrderAccess::storestore(); 1432 obj->set_klass(old->klass()); 1433 // We should now be able to calculate the right size for this object 1434 assert(obj->is_oop() && obj->size() == (int)word_sz, "Error, incorrect size computed for promoted object"); 1435 1436 collector()->promoted(true, // parallel 1437 obj_ptr, old->is_objArray(), word_sz); 1438 1439 NOT_PRODUCT( 1440 Atomic::inc_ptr(&_numObjectsPromoted); 1441 Atomic::add_ptr(alloc_sz, &_numWordsPromoted); 1442 ) 1443 1444 return obj; 1445 } 1446 1447 void 1448 ConcurrentMarkSweepGeneration:: 1449 par_promote_alloc_undo(int thread_num, 1450 HeapWord* obj, size_t word_sz) { 1451 // CMS does not support promotion undo. 1452 ShouldNotReachHere(); 1453 } 1454 1455 void 1456 ConcurrentMarkSweepGeneration:: 1457 par_promote_alloc_done(int thread_num) { 1458 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1459 ps->lab.retire(thread_num); 1460 } 1461 1462 void 1463 ConcurrentMarkSweepGeneration:: 1464 par_oop_since_save_marks_iterate_done(int thread_num) { 1465 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1466 ParScanWithoutBarrierClosure* dummy_cl = NULL; 1467 ps->promo.promoted_oops_iterate_nv(dummy_cl); 1468 } 1469 1470 bool ConcurrentMarkSweepGeneration::should_collect(bool full, 1471 size_t size, 1472 bool tlab) 1473 { 1474 // We allow a STW collection only if a full 1475 // collection was requested. 1476 return full || should_allocate(size, tlab); // FIX ME !!! 1477 // This and promotion failure handling are connected at the 1478 // hip and should be fixed by untying them. 1479 } 1480 1481 bool CMSCollector::shouldConcurrentCollect() { 1482 if (_full_gc_requested) { 1483 if (Verbose && PrintGCDetails) { 1484 gclog_or_tty->print_cr("CMSCollector: collect because of explicit " 1485 " gc request (or gc_locker)"); 1486 } 1487 return true; 1488 } 1489 1490 // For debugging purposes, change the type of collection. 1491 // If the rotation is not on the concurrent collection 1492 // type, don't start a concurrent collection. 1493 NOT_PRODUCT( 1494 if (RotateCMSCollectionTypes && 1495 (_cmsGen->debug_collection_type() != 1496 ConcurrentMarkSweepGeneration::Concurrent_collection_type)) { 1497 assert(_cmsGen->debug_collection_type() != 1498 ConcurrentMarkSweepGeneration::Unknown_collection_type, 1499 "Bad cms collection type"); 1500 return false; 1501 } 1502 ) 1503 1504 FreelistLocker x(this); 1505 // ------------------------------------------------------------------ 1506 // Print out lots of information which affects the initiation of 1507 // a collection. 1508 if (PrintCMSInitiationStatistics && stats().valid()) { 1509 gclog_or_tty->print("CMSCollector shouldConcurrentCollect: "); 1510 gclog_or_tty->stamp(); 1511 gclog_or_tty->print_cr(""); 1512 stats().print_on(gclog_or_tty); 1513 gclog_or_tty->print_cr("time_until_cms_gen_full %3.7f", 1514 stats().time_until_cms_gen_full()); 1515 gclog_or_tty->print_cr("free="SIZE_FORMAT, _cmsGen->free()); 1516 gclog_or_tty->print_cr("contiguous_available="SIZE_FORMAT, 1517 _cmsGen->contiguous_available()); 1518 gclog_or_tty->print_cr("promotion_rate=%g", stats().promotion_rate()); 1519 gclog_or_tty->print_cr("cms_allocation_rate=%g", stats().cms_allocation_rate()); 1520 gclog_or_tty->print_cr("occupancy=%3.7f", _cmsGen->occupancy()); 1521 gclog_or_tty->print_cr("initiatingOccupancy=%3.7f", _cmsGen->initiating_occupancy()); 1522 gclog_or_tty->print_cr("metadata initialized %d", 1523 MetaspaceGC::should_concurrent_collect()); 1524 } 1525 // ------------------------------------------------------------------ 1526 1527 // If the estimated time to complete a cms collection (cms_duration()) 1528 // is less than the estimated time remaining until the cms generation 1529 // is full, start a collection. 1530 if (!UseCMSInitiatingOccupancyOnly) { 1531 if (stats().valid()) { 1532 if (stats().time_until_cms_start() == 0.0) { 1533 return true; 1534 } 1535 } else { 1536 // We want to conservatively collect somewhat early in order 1537 // to try and "bootstrap" our CMS/promotion statistics; 1538 // this branch will not fire after the first successful CMS 1539 // collection because the stats should then be valid. 1540 if (_cmsGen->occupancy() >= _bootstrap_occupancy) { 1541 if (Verbose && PrintGCDetails) { 1542 gclog_or_tty->print_cr( 1543 " CMSCollector: collect for bootstrapping statistics:" 1544 " occupancy = %f, boot occupancy = %f", _cmsGen->occupancy(), 1545 _bootstrap_occupancy); 1546 } 1547 return true; 1548 } 1549 } 1550 } 1551 1552 // Otherwise, we start a collection cycle if 1553 // old gen want a collection cycle started. Each may use 1554 // an appropriate criterion for making this decision. 1555 // XXX We need to make sure that the gen expansion 1556 // criterion dovetails well with this. XXX NEED TO FIX THIS 1557 if (_cmsGen->should_concurrent_collect()) { 1558 if (Verbose && PrintGCDetails) { 1559 gclog_or_tty->print_cr("CMS old gen initiated"); 1560 } 1561 return true; 1562 } 1563 1564 // We start a collection if we believe an incremental collection may fail; 1565 // this is not likely to be productive in practice because it's probably too 1566 // late anyway. 1567 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1568 assert(gch->collector_policy()->is_two_generation_policy(), 1569 "You may want to check the correctness of the following"); 1570 if (gch->incremental_collection_will_fail(true /* consult_young */)) { 1571 if (Verbose && PrintGCDetails) { 1572 gclog_or_tty->print("CMSCollector: collect because incremental collection will fail "); 1573 } 1574 return true; 1575 } 1576 1577 if (MetaspaceGC::should_concurrent_collect()) { 1578 if (Verbose && PrintGCDetails) { 1579 gclog_or_tty->print("CMSCollector: collect for metadata allocation "); 1580 } 1581 return true; 1582 } 1583 1584 return false; 1585 } 1586 1587 void CMSCollector::set_did_compact(bool v) { _cmsGen->set_did_compact(v); } 1588 1589 // Clear _expansion_cause fields of constituent generations 1590 void CMSCollector::clear_expansion_cause() { 1591 _cmsGen->clear_expansion_cause(); 1592 } 1593 1594 // We should be conservative in starting a collection cycle. To 1595 // start too eagerly runs the risk of collecting too often in the 1596 // extreme. To collect too rarely falls back on full collections, 1597 // which works, even if not optimum in terms of concurrent work. 1598 // As a work around for too eagerly collecting, use the flag 1599 // UseCMSInitiatingOccupancyOnly. This also has the advantage of 1600 // giving the user an easily understandable way of controlling the 1601 // collections. 1602 // We want to start a new collection cycle if any of the following 1603 // conditions hold: 1604 // . our current occupancy exceeds the configured initiating occupancy 1605 // for this generation, or 1606 // . we recently needed to expand this space and have not, since that 1607 // expansion, done a collection of this generation, or 1608 // . the underlying space believes that it may be a good idea to initiate 1609 // a concurrent collection (this may be based on criteria such as the 1610 // following: the space uses linear allocation and linear allocation is 1611 // going to fail, or there is believed to be excessive fragmentation in 1612 // the generation, etc... or ... 1613 // [.(currently done by CMSCollector::shouldConcurrentCollect() only for 1614 // the case of the old generation; see CR 6543076): 1615 // we may be approaching a point at which allocation requests may fail because 1616 // we will be out of sufficient free space given allocation rate estimates.] 1617 bool ConcurrentMarkSweepGeneration::should_concurrent_collect() const { 1618 1619 assert_lock_strong(freelistLock()); 1620 if (occupancy() > initiating_occupancy()) { 1621 if (PrintGCDetails && Verbose) { 1622 gclog_or_tty->print(" %s: collect because of occupancy %f / %f ", 1623 short_name(), occupancy(), initiating_occupancy()); 1624 } 1625 return true; 1626 } 1627 if (UseCMSInitiatingOccupancyOnly) { 1628 return false; 1629 } 1630 if (expansion_cause() == CMSExpansionCause::_satisfy_allocation) { 1631 if (PrintGCDetails && Verbose) { 1632 gclog_or_tty->print(" %s: collect because expanded for allocation ", 1633 short_name()); 1634 } 1635 return true; 1636 } 1637 if (_cmsSpace->should_concurrent_collect()) { 1638 if (PrintGCDetails && Verbose) { 1639 gclog_or_tty->print(" %s: collect because cmsSpace says so ", 1640 short_name()); 1641 } 1642 return true; 1643 } 1644 return false; 1645 } 1646 1647 void ConcurrentMarkSweepGeneration::collect(bool full, 1648 bool clear_all_soft_refs, 1649 size_t size, 1650 bool tlab) 1651 { 1652 collector()->collect(full, clear_all_soft_refs, size, tlab); 1653 } 1654 1655 void CMSCollector::collect(bool full, 1656 bool clear_all_soft_refs, 1657 size_t size, 1658 bool tlab) 1659 { 1660 if (!UseCMSCollectionPassing && _collectorState > Idling) { 1661 // For debugging purposes skip the collection if the state 1662 // is not currently idle 1663 if (TraceCMSState) { 1664 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " skipped full:%d CMS state %d", 1665 Thread::current(), full, _collectorState); 1666 } 1667 return; 1668 } 1669 1670 // The following "if" branch is present for defensive reasons. 1671 // In the current uses of this interface, it can be replaced with: 1672 // assert(!GC_locker.is_active(), "Can't be called otherwise"); 1673 // But I am not placing that assert here to allow future 1674 // generality in invoking this interface. 1675 if (GC_locker::is_active()) { 1676 // A consistency test for GC_locker 1677 assert(GC_locker::needs_gc(), "Should have been set already"); 1678 // Skip this foreground collection, instead 1679 // expanding the heap if necessary. 1680 // Need the free list locks for the call to free() in compute_new_size() 1681 compute_new_size(); 1682 return; 1683 } 1684 acquire_control_and_collect(full, clear_all_soft_refs); 1685 _full_gcs_since_conc_gc++; 1686 } 1687 1688 void CMSCollector::request_full_gc(unsigned int full_gc_count, GCCause::Cause cause) { 1689 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1690 unsigned int gc_count = gch->total_full_collections(); 1691 if (gc_count == full_gc_count) { 1692 MutexLockerEx y(CGC_lock, Mutex::_no_safepoint_check_flag); 1693 _full_gc_requested = true; 1694 _full_gc_cause = cause; 1695 CGC_lock->notify(); // nudge CMS thread 1696 } else { 1697 assert(gc_count > full_gc_count, "Error: causal loop"); 1698 } 1699 } 1700 1701 bool CMSCollector::is_external_interruption() { 1702 GCCause::Cause cause = GenCollectedHeap::heap()->gc_cause(); 1703 return GCCause::is_user_requested_gc(cause) || 1704 GCCause::is_serviceability_requested_gc(cause); 1705 } 1706 1707 void CMSCollector::report_concurrent_mode_interruption() { 1708 if (is_external_interruption()) { 1709 if (PrintGCDetails) { 1710 gclog_or_tty->print(" (concurrent mode interrupted)"); 1711 } 1712 } else { 1713 if (PrintGCDetails) { 1714 gclog_or_tty->print(" (concurrent mode failure)"); 1715 } 1716 _gc_tracer_cm->report_concurrent_mode_failure(); 1717 } 1718 } 1719 1720 1721 // The foreground and background collectors need to coordinate in order 1722 // to make sure that they do not mutually interfere with CMS collections. 1723 // When a background collection is active, 1724 // the foreground collector may need to take over (preempt) and 1725 // synchronously complete an ongoing collection. Depending on the 1726 // frequency of the background collections and the heap usage 1727 // of the application, this preemption can be seldom or frequent. 1728 // There are only certain 1729 // points in the background collection that the "collection-baton" 1730 // can be passed to the foreground collector. 1731 // 1732 // The foreground collector will wait for the baton before 1733 // starting any part of the collection. The foreground collector 1734 // will only wait at one location. 1735 // 1736 // The background collector will yield the baton before starting a new 1737 // phase of the collection (e.g., before initial marking, marking from roots, 1738 // precleaning, final re-mark, sweep etc.) This is normally done at the head 1739 // of the loop which switches the phases. The background collector does some 1740 // of the phases (initial mark, final re-mark) with the world stopped. 1741 // Because of locking involved in stopping the world, 1742 // the foreground collector should not block waiting for the background 1743 // collector when it is doing a stop-the-world phase. The background 1744 // collector will yield the baton at an additional point just before 1745 // it enters a stop-the-world phase. Once the world is stopped, the 1746 // background collector checks the phase of the collection. If the 1747 // phase has not changed, it proceeds with the collection. If the 1748 // phase has changed, it skips that phase of the collection. See 1749 // the comments on the use of the Heap_lock in collect_in_background(). 1750 // 1751 // Variable used in baton passing. 1752 // _foregroundGCIsActive - Set to true by the foreground collector when 1753 // it wants the baton. The foreground clears it when it has finished 1754 // the collection. 1755 // _foregroundGCShouldWait - Set to true by the background collector 1756 // when it is running. The foreground collector waits while 1757 // _foregroundGCShouldWait is true. 1758 // CGC_lock - monitor used to protect access to the above variables 1759 // and to notify the foreground and background collectors. 1760 // _collectorState - current state of the CMS collection. 1761 // 1762 // The foreground collector 1763 // acquires the CGC_lock 1764 // sets _foregroundGCIsActive 1765 // waits on the CGC_lock for _foregroundGCShouldWait to be false 1766 // various locks acquired in preparation for the collection 1767 // are released so as not to block the background collector 1768 // that is in the midst of a collection 1769 // proceeds with the collection 1770 // clears _foregroundGCIsActive 1771 // returns 1772 // 1773 // The background collector in a loop iterating on the phases of the 1774 // collection 1775 // acquires the CGC_lock 1776 // sets _foregroundGCShouldWait 1777 // if _foregroundGCIsActive is set 1778 // clears _foregroundGCShouldWait, notifies _CGC_lock 1779 // waits on _CGC_lock for _foregroundGCIsActive to become false 1780 // and exits the loop. 1781 // otherwise 1782 // proceed with that phase of the collection 1783 // if the phase is a stop-the-world phase, 1784 // yield the baton once more just before enqueueing 1785 // the stop-world CMS operation (executed by the VM thread). 1786 // returns after all phases of the collection are done 1787 // 1788 1789 void CMSCollector::acquire_control_and_collect(bool full, 1790 bool clear_all_soft_refs) { 1791 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint"); 1792 assert(!Thread::current()->is_ConcurrentGC_thread(), 1793 "shouldn't try to acquire control from self!"); 1794 1795 // Start the protocol for acquiring control of the 1796 // collection from the background collector (aka CMS thread). 1797 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 1798 "VM thread should have CMS token"); 1799 // Remember the possibly interrupted state of an ongoing 1800 // concurrent collection 1801 CollectorState first_state = _collectorState; 1802 1803 // Signal to a possibly ongoing concurrent collection that 1804 // we want to do a foreground collection. 1805 _foregroundGCIsActive = true; 1806 1807 // Disable incremental mode during a foreground collection. 1808 ICMSDisabler icms_disabler; 1809 1810 // release locks and wait for a notify from the background collector 1811 // releasing the locks in only necessary for phases which 1812 // do yields to improve the granularity of the collection. 1813 assert_lock_strong(bitMapLock()); 1814 // We need to lock the Free list lock for the space that we are 1815 // currently collecting. 1816 assert(haveFreelistLocks(), "Must be holding free list locks"); 1817 bitMapLock()->unlock(); 1818 releaseFreelistLocks(); 1819 { 1820 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1821 if (_foregroundGCShouldWait) { 1822 // We are going to be waiting for action for the CMS thread; 1823 // it had better not be gone (for instance at shutdown)! 1824 assert(ConcurrentMarkSweepThread::cmst() != NULL, 1825 "CMS thread must be running"); 1826 // Wait here until the background collector gives us the go-ahead 1827 ConcurrentMarkSweepThread::clear_CMS_flag( 1828 ConcurrentMarkSweepThread::CMS_vm_has_token); // release token 1829 // Get a possibly blocked CMS thread going: 1830 // Note that we set _foregroundGCIsActive true above, 1831 // without protection of the CGC_lock. 1832 CGC_lock->notify(); 1833 assert(!ConcurrentMarkSweepThread::vm_thread_wants_cms_token(), 1834 "Possible deadlock"); 1835 while (_foregroundGCShouldWait) { 1836 // wait for notification 1837 CGC_lock->wait(Mutex::_no_safepoint_check_flag); 1838 // Possibility of delay/starvation here, since CMS token does 1839 // not know to give priority to VM thread? Actually, i think 1840 // there wouldn't be any delay/starvation, but the proof of 1841 // that "fact" (?) appears non-trivial. XXX 20011219YSR 1842 } 1843 ConcurrentMarkSweepThread::set_CMS_flag( 1844 ConcurrentMarkSweepThread::CMS_vm_has_token); 1845 } 1846 } 1847 // The CMS_token is already held. Get back the other locks. 1848 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 1849 "VM thread should have CMS token"); 1850 getFreelistLocks(); 1851 bitMapLock()->lock_without_safepoint_check(); 1852 if (TraceCMSState) { 1853 gclog_or_tty->print_cr("CMS foreground collector has asked for control " 1854 INTPTR_FORMAT " with first state %d", Thread::current(), first_state); 1855 gclog_or_tty->print_cr(" gets control with state %d", _collectorState); 1856 } 1857 1858 // Check if we need to do a compaction, or if not, whether 1859 // we need to start the mark-sweep from scratch. 1860 bool should_compact = false; 1861 bool should_start_over = false; 1862 decide_foreground_collection_type(clear_all_soft_refs, 1863 &should_compact, &should_start_over); 1864 1865 NOT_PRODUCT( 1866 if (RotateCMSCollectionTypes) { 1867 if (_cmsGen->debug_collection_type() == 1868 ConcurrentMarkSweepGeneration::MSC_foreground_collection_type) { 1869 should_compact = true; 1870 } else if (_cmsGen->debug_collection_type() == 1871 ConcurrentMarkSweepGeneration::MS_foreground_collection_type) { 1872 should_compact = false; 1873 } 1874 } 1875 ) 1876 1877 if (first_state > Idling) { 1878 report_concurrent_mode_interruption(); 1879 } 1880 1881 set_did_compact(should_compact); 1882 if (should_compact) { 1883 // If the collection is being acquired from the background 1884 // collector, there may be references on the discovered 1885 // references lists that have NULL referents (being those 1886 // that were concurrently cleared by a mutator) or 1887 // that are no longer active (having been enqueued concurrently 1888 // by the mutator). 1889 // Scrub the list of those references because Mark-Sweep-Compact 1890 // code assumes referents are not NULL and that all discovered 1891 // Reference objects are active. 1892 ref_processor()->clean_up_discovered_references(); 1893 1894 if (first_state > Idling) { 1895 save_heap_summary(); 1896 } 1897 1898 do_compaction_work(clear_all_soft_refs); 1899 1900 // Has the GC time limit been exceeded? 1901 DefNewGeneration* young_gen = _young_gen->as_DefNewGeneration(); 1902 size_t max_eden_size = young_gen->max_capacity() - 1903 young_gen->to()->capacity() - 1904 young_gen->from()->capacity(); 1905 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1906 GCCause::Cause gc_cause = gch->gc_cause(); 1907 size_policy()->check_gc_overhead_limit(_young_gen->used(), 1908 young_gen->eden()->used(), 1909 _cmsGen->max_capacity(), 1910 max_eden_size, 1911 full, 1912 gc_cause, 1913 gch->collector_policy()); 1914 } else { 1915 do_mark_sweep_work(clear_all_soft_refs, first_state, 1916 should_start_over); 1917 } 1918 // Reset the expansion cause, now that we just completed 1919 // a collection cycle. 1920 clear_expansion_cause(); 1921 _foregroundGCIsActive = false; 1922 return; 1923 } 1924 1925 // Resize the tenured generation 1926 // after obtaining the free list locks for the 1927 // two generations. 1928 void CMSCollector::compute_new_size() { 1929 assert_locked_or_safepoint(Heap_lock); 1930 FreelistLocker z(this); 1931 MetaspaceGC::compute_new_size(); 1932 _cmsGen->compute_new_size_free_list(); 1933 } 1934 1935 // A work method used by foreground collection to determine 1936 // what type of collection (compacting or not, continuing or fresh) 1937 // it should do. 1938 // NOTE: the intent is to make UseCMSCompactAtFullCollection 1939 // and CMSCompactWhenClearAllSoftRefs the default in the future 1940 // and do away with the flags after a suitable period. 1941 void CMSCollector::decide_foreground_collection_type( 1942 bool clear_all_soft_refs, bool* should_compact, 1943 bool* should_start_over) { 1944 // Normally, we'll compact only if the UseCMSCompactAtFullCollection 1945 // flag is set, and we have either requested a System.gc() or 1946 // the number of full gc's since the last concurrent cycle 1947 // has exceeded the threshold set by CMSFullGCsBeforeCompaction, 1948 // or if an incremental collection has failed 1949 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1950 assert(gch->collector_policy()->is_two_generation_policy(), 1951 "You may want to check the correctness of the following"); 1952 // Inform cms gen if this was due to partial collection failing. 1953 // The CMS gen may use this fact to determine its expansion policy. 1954 if (gch->incremental_collection_will_fail(false /* don't consult_young */)) { 1955 assert(!_cmsGen->incremental_collection_failed(), 1956 "Should have been noticed, reacted to and cleared"); 1957 _cmsGen->set_incremental_collection_failed(); 1958 } 1959 *should_compact = 1960 UseCMSCompactAtFullCollection && 1961 ((_full_gcs_since_conc_gc >= CMSFullGCsBeforeCompaction) || 1962 GCCause::is_user_requested_gc(gch->gc_cause()) || 1963 gch->incremental_collection_will_fail(true /* consult_young */)); 1964 *should_start_over = false; 1965 if (clear_all_soft_refs && !*should_compact) { 1966 // We are about to do a last ditch collection attempt 1967 // so it would normally make sense to do a compaction 1968 // to reclaim as much space as possible. 1969 if (CMSCompactWhenClearAllSoftRefs) { 1970 // Default: The rationale is that in this case either 1971 // we are past the final marking phase, in which case 1972 // we'd have to start over, or so little has been done 1973 // that there's little point in saving that work. Compaction 1974 // appears to be the sensible choice in either case. 1975 *should_compact = true; 1976 } else { 1977 // We have been asked to clear all soft refs, but not to 1978 // compact. Make sure that we aren't past the final checkpoint 1979 // phase, for that is where we process soft refs. If we are already 1980 // past that phase, we'll need to redo the refs discovery phase and 1981 // if necessary clear soft refs that weren't previously 1982 // cleared. We do so by remembering the phase in which 1983 // we came in, and if we are past the refs processing 1984 // phase, we'll choose to just redo the mark-sweep 1985 // collection from scratch. 1986 if (_collectorState > FinalMarking) { 1987 // We are past the refs processing phase; 1988 // start over and do a fresh synchronous CMS cycle 1989 _collectorState = Resetting; // skip to reset to start new cycle 1990 reset(false /* == !asynch */); 1991 *should_start_over = true; 1992 } // else we can continue a possibly ongoing current cycle 1993 } 1994 } 1995 } 1996 1997 // A work method used by the foreground collector to do 1998 // a mark-sweep-compact. 1999 void CMSCollector::do_compaction_work(bool clear_all_soft_refs) { 2000 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2001 2002 STWGCTimer* gc_timer = GenMarkSweep::gc_timer(); 2003 gc_timer->register_gc_start(os::elapsed_counter()); 2004 2005 SerialOldTracer* gc_tracer = GenMarkSweep::gc_tracer(); 2006 gc_tracer->report_gc_start(gch->gc_cause(), gc_timer->gc_start()); 2007 2008 GCTraceTime t("CMS:MSC ", PrintGCDetails && Verbose, true, NULL); 2009 if (PrintGC && Verbose && !(GCCause::is_user_requested_gc(gch->gc_cause()))) { 2010 gclog_or_tty->print_cr("Compact ConcurrentMarkSweepGeneration after %d " 2011 "collections passed to foreground collector", _full_gcs_since_conc_gc); 2012 } 2013 2014 // Sample collection interval time and reset for collection pause. 2015 if (UseAdaptiveSizePolicy) { 2016 size_policy()->msc_collection_begin(); 2017 } 2018 2019 // Temporarily widen the span of the weak reference processing to 2020 // the entire heap. 2021 MemRegion new_span(GenCollectedHeap::heap()->reserved_region()); 2022 ReferenceProcessorSpanMutator rp_mut_span(ref_processor(), new_span); 2023 // Temporarily, clear the "is_alive_non_header" field of the 2024 // reference processor. 2025 ReferenceProcessorIsAliveMutator rp_mut_closure(ref_processor(), NULL); 2026 // Temporarily make reference _processing_ single threaded (non-MT). 2027 ReferenceProcessorMTProcMutator rp_mut_mt_processing(ref_processor(), false); 2028 // Temporarily make refs discovery atomic 2029 ReferenceProcessorAtomicMutator rp_mut_atomic(ref_processor(), true); 2030 // Temporarily make reference _discovery_ single threaded (non-MT) 2031 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false); 2032 2033 ref_processor()->set_enqueuing_is_done(false); 2034 ref_processor()->enable_discovery(false /*verify_disabled*/, false /*check_no_refs*/); 2035 ref_processor()->setup_policy(clear_all_soft_refs); 2036 // If an asynchronous collection finishes, the _modUnionTable is 2037 // all clear. If we are assuming the collection from an asynchronous 2038 // collection, clear the _modUnionTable. 2039 assert(_collectorState != Idling || _modUnionTable.isAllClear(), 2040 "_modUnionTable should be clear if the baton was not passed"); 2041 _modUnionTable.clear_all(); 2042 assert(_collectorState != Idling || _ct->klass_rem_set()->mod_union_is_clear(), 2043 "mod union for klasses should be clear if the baton was passed"); 2044 _ct->klass_rem_set()->clear_mod_union(); 2045 2046 // We must adjust the allocation statistics being maintained 2047 // in the free list space. We do so by reading and clearing 2048 // the sweep timer and updating the block flux rate estimates below. 2049 assert(!_intra_sweep_timer.is_active(), "_intra_sweep_timer should be inactive"); 2050 if (_inter_sweep_timer.is_active()) { 2051 _inter_sweep_timer.stop(); 2052 // Note that we do not use this sample to update the _inter_sweep_estimate. 2053 _cmsGen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()), 2054 _inter_sweep_estimate.padded_average(), 2055 _intra_sweep_estimate.padded_average()); 2056 } 2057 2058 GenMarkSweep::invoke_at_safepoint(_cmsGen->level(), 2059 ref_processor(), clear_all_soft_refs); 2060 #ifdef ASSERT 2061 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 2062 size_t free_size = cms_space->free(); 2063 assert(free_size == 2064 pointer_delta(cms_space->end(), cms_space->compaction_top()) 2065 * HeapWordSize, 2066 "All the free space should be compacted into one chunk at top"); 2067 assert(cms_space->dictionary()->total_chunk_size( 2068 debug_only(cms_space->freelistLock())) == 0 || 2069 cms_space->totalSizeInIndexedFreeLists() == 0, 2070 "All the free space should be in a single chunk"); 2071 size_t num = cms_space->totalCount(); 2072 assert((free_size == 0 && num == 0) || 2073 (free_size > 0 && (num == 1 || num == 2)), 2074 "There should be at most 2 free chunks after compaction"); 2075 #endif // ASSERT 2076 _collectorState = Resetting; 2077 assert(_restart_addr == NULL, 2078 "Should have been NULL'd before baton was passed"); 2079 reset(false /* == !asynch */); 2080 _cmsGen->reset_after_compaction(); 2081 _concurrent_cycles_since_last_unload = 0; 2082 2083 // Clear any data recorded in the PLAB chunk arrays. 2084 if (_survivor_plab_array != NULL) { 2085 reset_survivor_plab_arrays(); 2086 } 2087 2088 // Adjust the per-size allocation stats for the next epoch. 2089 _cmsGen->cmsSpace()->endSweepFLCensus(sweep_count() /* fake */); 2090 // Restart the "inter sweep timer" for the next epoch. 2091 _inter_sweep_timer.reset(); 2092 _inter_sweep_timer.start(); 2093 2094 // Sample collection pause time and reset for collection interval. 2095 if (UseAdaptiveSizePolicy) { 2096 size_policy()->msc_collection_end(gch->gc_cause()); 2097 } 2098 2099 gc_timer->register_gc_end(os::elapsed_counter()); 2100 2101 gc_tracer->report_gc_end(gc_timer->gc_end(), gc_timer->time_partitions()); 2102 2103 // For a mark-sweep-compact, compute_new_size() will be called 2104 // in the heap's do_collection() method. 2105 } 2106 2107 // A work method used by the foreground collector to do 2108 // a mark-sweep, after taking over from a possibly on-going 2109 // concurrent mark-sweep collection. 2110 void CMSCollector::do_mark_sweep_work(bool clear_all_soft_refs, 2111 CollectorState first_state, bool should_start_over) { 2112 if (PrintGC && Verbose) { 2113 gclog_or_tty->print_cr("Pass concurrent collection to foreground " 2114 "collector with count %d", 2115 _full_gcs_since_conc_gc); 2116 } 2117 switch (_collectorState) { 2118 case Idling: 2119 if (first_state == Idling || should_start_over) { 2120 // The background GC was not active, or should 2121 // restarted from scratch; start the cycle. 2122 _collectorState = InitialMarking; 2123 } 2124 // If first_state was not Idling, then a background GC 2125 // was in progress and has now finished. No need to do it 2126 // again. Leave the state as Idling. 2127 break; 2128 case Precleaning: 2129 // In the foreground case don't do the precleaning since 2130 // it is not done concurrently and there is extra work 2131 // required. 2132 _collectorState = FinalMarking; 2133 } 2134 collect_in_foreground(clear_all_soft_refs, GenCollectedHeap::heap()->gc_cause()); 2135 2136 // For a mark-sweep, compute_new_size() will be called 2137 // in the heap's do_collection() method. 2138 } 2139 2140 2141 void CMSCollector::print_eden_and_survivor_chunk_arrays() { 2142 DefNewGeneration* dng = _young_gen->as_DefNewGeneration(); 2143 EdenSpace* eden_space = dng->eden(); 2144 ContiguousSpace* from_space = dng->from(); 2145 ContiguousSpace* to_space = dng->to(); 2146 // Eden 2147 if (_eden_chunk_array != NULL) { 2148 gclog_or_tty->print_cr("eden " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")", 2149 eden_space->bottom(), eden_space->top(), 2150 eden_space->end(), eden_space->capacity()); 2151 gclog_or_tty->print_cr("_eden_chunk_index=" SIZE_FORMAT ", " 2152 "_eden_chunk_capacity=" SIZE_FORMAT, 2153 _eden_chunk_index, _eden_chunk_capacity); 2154 for (size_t i = 0; i < _eden_chunk_index; i++) { 2155 gclog_or_tty->print_cr("_eden_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT, 2156 i, _eden_chunk_array[i]); 2157 } 2158 } 2159 // Survivor 2160 if (_survivor_chunk_array != NULL) { 2161 gclog_or_tty->print_cr("survivor " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")", 2162 from_space->bottom(), from_space->top(), 2163 from_space->end(), from_space->capacity()); 2164 gclog_or_tty->print_cr("_survivor_chunk_index=" SIZE_FORMAT ", " 2165 "_survivor_chunk_capacity=" SIZE_FORMAT, 2166 _survivor_chunk_index, _survivor_chunk_capacity); 2167 for (size_t i = 0; i < _survivor_chunk_index; i++) { 2168 gclog_or_tty->print_cr("_survivor_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT, 2169 i, _survivor_chunk_array[i]); 2170 } 2171 } 2172 } 2173 2174 void CMSCollector::getFreelistLocks() const { 2175 // Get locks for all free lists in all generations that this 2176 // collector is responsible for 2177 _cmsGen->freelistLock()->lock_without_safepoint_check(); 2178 } 2179 2180 void CMSCollector::releaseFreelistLocks() const { 2181 // Release locks for all free lists in all generations that this 2182 // collector is responsible for 2183 _cmsGen->freelistLock()->unlock(); 2184 } 2185 2186 bool CMSCollector::haveFreelistLocks() const { 2187 // Check locks for all free lists in all generations that this 2188 // collector is responsible for 2189 assert_lock_strong(_cmsGen->freelistLock()); 2190 PRODUCT_ONLY(ShouldNotReachHere()); 2191 return true; 2192 } 2193 2194 // A utility class that is used by the CMS collector to 2195 // temporarily "release" the foreground collector from its 2196 // usual obligation to wait for the background collector to 2197 // complete an ongoing phase before proceeding. 2198 class ReleaseForegroundGC: public StackObj { 2199 private: 2200 CMSCollector* _c; 2201 public: 2202 ReleaseForegroundGC(CMSCollector* c) : _c(c) { 2203 assert(_c->_foregroundGCShouldWait, "Else should not need to call"); 2204 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 2205 // allow a potentially blocked foreground collector to proceed 2206 _c->_foregroundGCShouldWait = false; 2207 if (_c->_foregroundGCIsActive) { 2208 CGC_lock->notify(); 2209 } 2210 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 2211 "Possible deadlock"); 2212 } 2213 2214 ~ReleaseForegroundGC() { 2215 assert(!_c->_foregroundGCShouldWait, "Usage protocol violation?"); 2216 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 2217 _c->_foregroundGCShouldWait = true; 2218 } 2219 }; 2220 2221 // There are separate collect_in_background and collect_in_foreground because of 2222 // the different locking requirements of the background collector and the 2223 // foreground collector. There was originally an attempt to share 2224 // one "collect" method between the background collector and the foreground 2225 // collector but the if-then-else required made it cleaner to have 2226 // separate methods. 2227 void CMSCollector::collect_in_background(bool clear_all_soft_refs, GCCause::Cause cause) { 2228 assert(Thread::current()->is_ConcurrentGC_thread(), 2229 "A CMS asynchronous collection is only allowed on a CMS thread."); 2230 2231 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2232 { 2233 bool safepoint_check = Mutex::_no_safepoint_check_flag; 2234 MutexLockerEx hl(Heap_lock, safepoint_check); 2235 FreelistLocker fll(this); 2236 MutexLockerEx x(CGC_lock, safepoint_check); 2237 if (_foregroundGCIsActive || !UseAsyncConcMarkSweepGC) { 2238 // The foreground collector is active or we're 2239 // not using asynchronous collections. Skip this 2240 // background collection. 2241 assert(!_foregroundGCShouldWait, "Should be clear"); 2242 return; 2243 } else { 2244 assert(_collectorState == Idling, "Should be idling before start."); 2245 _collectorState = InitialMarking; 2246 register_gc_start(cause); 2247 // Reset the expansion cause, now that we are about to begin 2248 // a new cycle. 2249 clear_expansion_cause(); 2250 2251 // Clear the MetaspaceGC flag since a concurrent collection 2252 // is starting but also clear it after the collection. 2253 MetaspaceGC::set_should_concurrent_collect(false); 2254 } 2255 // Decide if we want to enable class unloading as part of the 2256 // ensuing concurrent GC cycle. 2257 update_should_unload_classes(); 2258 _full_gc_requested = false; // acks all outstanding full gc requests 2259 _full_gc_cause = GCCause::_no_gc; 2260 // Signal that we are about to start a collection 2261 gch->increment_total_full_collections(); // ... starting a collection cycle 2262 _collection_count_start = gch->total_full_collections(); 2263 } 2264 2265 // Used for PrintGC 2266 size_t prev_used; 2267 if (PrintGC && Verbose) { 2268 prev_used = _cmsGen->used(); // XXXPERM 2269 } 2270 2271 // The change of the collection state is normally done at this level; 2272 // the exceptions are phases that are executed while the world is 2273 // stopped. For those phases the change of state is done while the 2274 // world is stopped. For baton passing purposes this allows the 2275 // background collector to finish the phase and change state atomically. 2276 // The foreground collector cannot wait on a phase that is done 2277 // while the world is stopped because the foreground collector already 2278 // has the world stopped and would deadlock. 2279 while (_collectorState != Idling) { 2280 if (TraceCMSState) { 2281 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " in CMS state %d", 2282 Thread::current(), _collectorState); 2283 } 2284 // The foreground collector 2285 // holds the Heap_lock throughout its collection. 2286 // holds the CMS token (but not the lock) 2287 // except while it is waiting for the background collector to yield. 2288 // 2289 // The foreground collector should be blocked (not for long) 2290 // if the background collector is about to start a phase 2291 // executed with world stopped. If the background 2292 // collector has already started such a phase, the 2293 // foreground collector is blocked waiting for the 2294 // Heap_lock. The stop-world phases (InitialMarking and FinalMarking) 2295 // are executed in the VM thread. 2296 // 2297 // The locking order is 2298 // PendingListLock (PLL) -- if applicable (FinalMarking) 2299 // Heap_lock (both this & PLL locked in VM_CMS_Operation::prologue()) 2300 // CMS token (claimed in 2301 // stop_world_and_do() --> 2302 // safepoint_synchronize() --> 2303 // CMSThread::synchronize()) 2304 2305 { 2306 // Check if the FG collector wants us to yield. 2307 CMSTokenSync x(true); // is cms thread 2308 if (waitForForegroundGC()) { 2309 // We yielded to a foreground GC, nothing more to be 2310 // done this round. 2311 assert(_foregroundGCShouldWait == false, "We set it to false in " 2312 "waitForForegroundGC()"); 2313 if (TraceCMSState) { 2314 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT 2315 " exiting collection CMS state %d", 2316 Thread::current(), _collectorState); 2317 } 2318 return; 2319 } else { 2320 // The background collector can run but check to see if the 2321 // foreground collector has done a collection while the 2322 // background collector was waiting to get the CGC_lock 2323 // above. If yes, break so that _foregroundGCShouldWait 2324 // is cleared before returning. 2325 if (_collectorState == Idling) { 2326 break; 2327 } 2328 } 2329 } 2330 2331 assert(_foregroundGCShouldWait, "Foreground collector, if active, " 2332 "should be waiting"); 2333 2334 switch (_collectorState) { 2335 case InitialMarking: 2336 { 2337 ReleaseForegroundGC x(this); 2338 stats().record_cms_begin(); 2339 VM_CMS_Initial_Mark initial_mark_op(this); 2340 VMThread::execute(&initial_mark_op); 2341 } 2342 // The collector state may be any legal state at this point 2343 // since the background collector may have yielded to the 2344 // foreground collector. 2345 break; 2346 case Marking: 2347 // initial marking in checkpointRootsInitialWork has been completed 2348 if (markFromRoots(true)) { // we were successful 2349 assert(_collectorState == Precleaning, "Collector state should " 2350 "have changed"); 2351 } else { 2352 assert(_foregroundGCIsActive, "Internal state inconsistency"); 2353 } 2354 break; 2355 case Precleaning: 2356 if (UseAdaptiveSizePolicy) { 2357 size_policy()->concurrent_precleaning_begin(); 2358 } 2359 // marking from roots in markFromRoots has been completed 2360 preclean(); 2361 if (UseAdaptiveSizePolicy) { 2362 size_policy()->concurrent_precleaning_end(); 2363 } 2364 assert(_collectorState == AbortablePreclean || 2365 _collectorState == FinalMarking, 2366 "Collector state should have changed"); 2367 break; 2368 case AbortablePreclean: 2369 if (UseAdaptiveSizePolicy) { 2370 size_policy()->concurrent_phases_resume(); 2371 } 2372 abortable_preclean(); 2373 if (UseAdaptiveSizePolicy) { 2374 size_policy()->concurrent_precleaning_end(); 2375 } 2376 assert(_collectorState == FinalMarking, "Collector state should " 2377 "have changed"); 2378 break; 2379 case FinalMarking: 2380 { 2381 ReleaseForegroundGC x(this); 2382 2383 VM_CMS_Final_Remark final_remark_op(this); 2384 VMThread::execute(&final_remark_op); 2385 } 2386 assert(_foregroundGCShouldWait, "block post-condition"); 2387 break; 2388 case Sweeping: 2389 if (UseAdaptiveSizePolicy) { 2390 size_policy()->concurrent_sweeping_begin(); 2391 } 2392 // final marking in checkpointRootsFinal has been completed 2393 sweep(true); 2394 assert(_collectorState == Resizing, "Collector state change " 2395 "to Resizing must be done under the free_list_lock"); 2396 _full_gcs_since_conc_gc = 0; 2397 2398 // Stop the timers for adaptive size policy for the concurrent phases 2399 if (UseAdaptiveSizePolicy) { 2400 size_policy()->concurrent_sweeping_end(); 2401 size_policy()->concurrent_phases_end(gch->gc_cause(), 2402 gch->prev_gen(_cmsGen)->capacity(), 2403 _cmsGen->free()); 2404 } 2405 2406 case Resizing: { 2407 // Sweeping has been completed... 2408 // At this point the background collection has completed. 2409 // Don't move the call to compute_new_size() down 2410 // into code that might be executed if the background 2411 // collection was preempted. 2412 { 2413 ReleaseForegroundGC x(this); // unblock FG collection 2414 MutexLockerEx y(Heap_lock, Mutex::_no_safepoint_check_flag); 2415 CMSTokenSync z(true); // not strictly needed. 2416 if (_collectorState == Resizing) { 2417 compute_new_size(); 2418 save_heap_summary(); 2419 _collectorState = Resetting; 2420 } else { 2421 assert(_collectorState == Idling, "The state should only change" 2422 " because the foreground collector has finished the collection"); 2423 } 2424 } 2425 break; 2426 } 2427 case Resetting: 2428 // CMS heap resizing has been completed 2429 reset(true); 2430 assert(_collectorState == Idling, "Collector state should " 2431 "have changed"); 2432 2433 MetaspaceGC::set_should_concurrent_collect(false); 2434 2435 stats().record_cms_end(); 2436 // Don't move the concurrent_phases_end() and compute_new_size() 2437 // calls to here because a preempted background collection 2438 // has it's state set to "Resetting". 2439 break; 2440 case Idling: 2441 default: 2442 ShouldNotReachHere(); 2443 break; 2444 } 2445 if (TraceCMSState) { 2446 gclog_or_tty->print_cr(" Thread " INTPTR_FORMAT " done - next CMS state %d", 2447 Thread::current(), _collectorState); 2448 } 2449 assert(_foregroundGCShouldWait, "block post-condition"); 2450 } 2451 2452 // Should this be in gc_epilogue? 2453 collector_policy()->counters()->update_counters(); 2454 2455 { 2456 // Clear _foregroundGCShouldWait and, in the event that the 2457 // foreground collector is waiting, notify it, before 2458 // returning. 2459 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 2460 _foregroundGCShouldWait = false; 2461 if (_foregroundGCIsActive) { 2462 CGC_lock->notify(); 2463 } 2464 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 2465 "Possible deadlock"); 2466 } 2467 if (TraceCMSState) { 2468 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT 2469 " exiting collection CMS state %d", 2470 Thread::current(), _collectorState); 2471 } 2472 if (PrintGC && Verbose) { 2473 _cmsGen->print_heap_change(prev_used); 2474 } 2475 } 2476 2477 void CMSCollector::register_foreground_gc_start(GCCause::Cause cause) { 2478 if (!_cms_start_registered) { 2479 register_gc_start(cause); 2480 } 2481 } 2482 2483 void CMSCollector::register_gc_start(GCCause::Cause cause) { 2484 _cms_start_registered = true; 2485 _gc_timer_cm->register_gc_start(os::elapsed_counter()); 2486 _gc_tracer_cm->report_gc_start(cause, _gc_timer_cm->gc_start()); 2487 } 2488 2489 void CMSCollector::register_gc_end() { 2490 if (_cms_start_registered) { 2491 report_heap_summary(GCWhen::AfterGC); 2492 2493 _gc_timer_cm->register_gc_end(os::elapsed_counter()); 2494 _gc_tracer_cm->report_gc_end(_gc_timer_cm->gc_end(), _gc_timer_cm->time_partitions()); 2495 _cms_start_registered = false; 2496 } 2497 } 2498 2499 void CMSCollector::save_heap_summary() { 2500 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2501 _last_heap_summary = gch->create_heap_summary(); 2502 _last_metaspace_summary = gch->create_metaspace_summary(); 2503 } 2504 2505 void CMSCollector::report_heap_summary(GCWhen::Type when) { 2506 _gc_tracer_cm->report_gc_heap_summary(when, _last_heap_summary, _last_metaspace_summary); 2507 } 2508 2509 void CMSCollector::collect_in_foreground(bool clear_all_soft_refs, GCCause::Cause cause) { 2510 assert(_foregroundGCIsActive && !_foregroundGCShouldWait, 2511 "Foreground collector should be waiting, not executing"); 2512 assert(Thread::current()->is_VM_thread(), "A foreground collection" 2513 "may only be done by the VM Thread with the world stopped"); 2514 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 2515 "VM thread should have CMS token"); 2516 2517 NOT_PRODUCT(GCTraceTime t("CMS:MS (foreground) ", PrintGCDetails && Verbose, 2518 true, NULL);) 2519 if (UseAdaptiveSizePolicy) { 2520 size_policy()->ms_collection_begin(); 2521 } 2522 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact); 2523 2524 HandleMark hm; // Discard invalid handles created during verification 2525 2526 if (VerifyBeforeGC && 2527 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2528 Universe::verify(); 2529 } 2530 2531 // Snapshot the soft reference policy to be used in this collection cycle. 2532 ref_processor()->setup_policy(clear_all_soft_refs); 2533 2534 bool init_mark_was_synchronous = false; // until proven otherwise 2535 while (_collectorState != Idling) { 2536 if (TraceCMSState) { 2537 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " in CMS state %d", 2538 Thread::current(), _collectorState); 2539 } 2540 switch (_collectorState) { 2541 case InitialMarking: 2542 register_foreground_gc_start(cause); 2543 init_mark_was_synchronous = true; // fact to be exploited in re-mark 2544 checkpointRootsInitial(false); 2545 assert(_collectorState == Marking, "Collector state should have changed" 2546 " within checkpointRootsInitial()"); 2547 break; 2548 case Marking: 2549 // initial marking in checkpointRootsInitialWork has been completed 2550 if (VerifyDuringGC && 2551 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2552 Universe::verify("Verify before initial mark: "); 2553 } 2554 { 2555 bool res = markFromRoots(false); 2556 assert(res && _collectorState == FinalMarking, "Collector state should " 2557 "have changed"); 2558 break; 2559 } 2560 case FinalMarking: 2561 if (VerifyDuringGC && 2562 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2563 Universe::verify("Verify before re-mark: "); 2564 } 2565 checkpointRootsFinal(false, clear_all_soft_refs, 2566 init_mark_was_synchronous); 2567 assert(_collectorState == Sweeping, "Collector state should not " 2568 "have changed within checkpointRootsFinal()"); 2569 break; 2570 case Sweeping: 2571 // final marking in checkpointRootsFinal has been completed 2572 if (VerifyDuringGC && 2573 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2574 Universe::verify("Verify before sweep: "); 2575 } 2576 sweep(false); 2577 assert(_collectorState == Resizing, "Incorrect state"); 2578 break; 2579 case Resizing: { 2580 // Sweeping has been completed; the actual resize in this case 2581 // is done separately; nothing to be done in this state. 2582 _collectorState = Resetting; 2583 break; 2584 } 2585 case Resetting: 2586 // The heap has been resized. 2587 if (VerifyDuringGC && 2588 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2589 Universe::verify("Verify before reset: "); 2590 } 2591 save_heap_summary(); 2592 reset(false); 2593 assert(_collectorState == Idling, "Collector state should " 2594 "have changed"); 2595 break; 2596 case Precleaning: 2597 case AbortablePreclean: 2598 // Elide the preclean phase 2599 _collectorState = FinalMarking; 2600 break; 2601 default: 2602 ShouldNotReachHere(); 2603 } 2604 if (TraceCMSState) { 2605 gclog_or_tty->print_cr(" Thread " INTPTR_FORMAT " done - next CMS state %d", 2606 Thread::current(), _collectorState); 2607 } 2608 } 2609 2610 if (UseAdaptiveSizePolicy) { 2611 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2612 size_policy()->ms_collection_end(gch->gc_cause()); 2613 } 2614 2615 if (VerifyAfterGC && 2616 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2617 Universe::verify(); 2618 } 2619 if (TraceCMSState) { 2620 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT 2621 " exiting collection CMS state %d", 2622 Thread::current(), _collectorState); 2623 } 2624 } 2625 2626 bool CMSCollector::waitForForegroundGC() { 2627 bool res = false; 2628 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 2629 "CMS thread should have CMS token"); 2630 // Block the foreground collector until the 2631 // background collectors decides whether to 2632 // yield. 2633 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 2634 _foregroundGCShouldWait = true; 2635 if (_foregroundGCIsActive) { 2636 // The background collector yields to the 2637 // foreground collector and returns a value 2638 // indicating that it has yielded. The foreground 2639 // collector can proceed. 2640 res = true; 2641 _foregroundGCShouldWait = false; 2642 ConcurrentMarkSweepThread::clear_CMS_flag( 2643 ConcurrentMarkSweepThread::CMS_cms_has_token); 2644 ConcurrentMarkSweepThread::set_CMS_flag( 2645 ConcurrentMarkSweepThread::CMS_cms_wants_token); 2646 // Get a possibly blocked foreground thread going 2647 CGC_lock->notify(); 2648 if (TraceCMSState) { 2649 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " waiting at CMS state %d", 2650 Thread::current(), _collectorState); 2651 } 2652 while (_foregroundGCIsActive) { 2653 CGC_lock->wait(Mutex::_no_safepoint_check_flag); 2654 } 2655 ConcurrentMarkSweepThread::set_CMS_flag( 2656 ConcurrentMarkSweepThread::CMS_cms_has_token); 2657 ConcurrentMarkSweepThread::clear_CMS_flag( 2658 ConcurrentMarkSweepThread::CMS_cms_wants_token); 2659 } 2660 if (TraceCMSState) { 2661 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " continuing at CMS state %d", 2662 Thread::current(), _collectorState); 2663 } 2664 return res; 2665 } 2666 2667 // Because of the need to lock the free lists and other structures in 2668 // the collector, common to all the generations that the collector is 2669 // collecting, we need the gc_prologues of individual CMS generations 2670 // delegate to their collector. It may have been simpler had the 2671 // current infrastructure allowed one to call a prologue on a 2672 // collector. In the absence of that we have the generation's 2673 // prologue delegate to the collector, which delegates back 2674 // some "local" work to a worker method in the individual generations 2675 // that it's responsible for collecting, while itself doing any 2676 // work common to all generations it's responsible for. A similar 2677 // comment applies to the gc_epilogue()'s. 2678 // The role of the varaible _between_prologue_and_epilogue is to 2679 // enforce the invocation protocol. 2680 void CMSCollector::gc_prologue(bool full) { 2681 // Call gc_prologue_work() for the CMSGen 2682 // we are responsible for. 2683 2684 // The following locking discipline assumes that we are only called 2685 // when the world is stopped. 2686 assert(SafepointSynchronize::is_at_safepoint(), "world is stopped assumption"); 2687 2688 // The CMSCollector prologue must call the gc_prologues for the 2689 // "generations" that it's responsible 2690 // for. 2691 2692 assert( Thread::current()->is_VM_thread() 2693 || ( CMSScavengeBeforeRemark 2694 && Thread::current()->is_ConcurrentGC_thread()), 2695 "Incorrect thread type for prologue execution"); 2696 2697 if (_between_prologue_and_epilogue) { 2698 // We have already been invoked; this is a gc_prologue delegation 2699 // from yet another CMS generation that we are responsible for, just 2700 // ignore it since all relevant work has already been done. 2701 return; 2702 } 2703 2704 // set a bit saying prologue has been called; cleared in epilogue 2705 _between_prologue_and_epilogue = true; 2706 // Claim locks for common data structures, then call gc_prologue_work() 2707 // for each CMSGen. 2708 2709 getFreelistLocks(); // gets free list locks on constituent spaces 2710 bitMapLock()->lock_without_safepoint_check(); 2711 2712 // Should call gc_prologue_work() for all cms gens we are responsible for 2713 bool duringMarking = _collectorState >= Marking 2714 && _collectorState < Sweeping; 2715 2716 // The young collections clear the modified oops state, which tells if 2717 // there are any modified oops in the class. The remark phase also needs 2718 // that information. Tell the young collection to save the union of all 2719 // modified klasses. 2720 if (duringMarking) { 2721 _ct->klass_rem_set()->set_accumulate_modified_oops(true); 2722 } 2723 2724 bool registerClosure = duringMarking; 2725 2726 ModUnionClosure* muc = CollectedHeap::use_parallel_gc_threads() ? 2727 &_modUnionClosurePar 2728 : &_modUnionClosure; 2729 _cmsGen->gc_prologue_work(full, registerClosure, muc); 2730 2731 if (!full) { 2732 stats().record_gc0_begin(); 2733 } 2734 } 2735 2736 void ConcurrentMarkSweepGeneration::gc_prologue(bool full) { 2737 2738 _capacity_at_prologue = capacity(); 2739 _used_at_prologue = used(); 2740 2741 // Delegate to CMScollector which knows how to coordinate between 2742 // this and any other CMS generations that it is responsible for 2743 // collecting. 2744 collector()->gc_prologue(full); 2745 } 2746 2747 // This is a "private" interface for use by this generation's CMSCollector. 2748 // Not to be called directly by any other entity (for instance, 2749 // GenCollectedHeap, which calls the "public" gc_prologue method above). 2750 void ConcurrentMarkSweepGeneration::gc_prologue_work(bool full, 2751 bool registerClosure, ModUnionClosure* modUnionClosure) { 2752 assert(!incremental_collection_failed(), "Shouldn't be set yet"); 2753 assert(cmsSpace()->preconsumptionDirtyCardClosure() == NULL, 2754 "Should be NULL"); 2755 if (registerClosure) { 2756 cmsSpace()->setPreconsumptionDirtyCardClosure(modUnionClosure); 2757 } 2758 cmsSpace()->gc_prologue(); 2759 // Clear stat counters 2760 NOT_PRODUCT( 2761 assert(_numObjectsPromoted == 0, "check"); 2762 assert(_numWordsPromoted == 0, "check"); 2763 if (Verbose && PrintGC) { 2764 gclog_or_tty->print("Allocated "SIZE_FORMAT" objects, " 2765 SIZE_FORMAT" bytes concurrently", 2766 _numObjectsAllocated, _numWordsAllocated*sizeof(HeapWord)); 2767 } 2768 _numObjectsAllocated = 0; 2769 _numWordsAllocated = 0; 2770 ) 2771 } 2772 2773 void CMSCollector::gc_epilogue(bool full) { 2774 // The following locking discipline assumes that we are only called 2775 // when the world is stopped. 2776 assert(SafepointSynchronize::is_at_safepoint(), 2777 "world is stopped assumption"); 2778 2779 // Currently the CMS epilogue (see CompactibleFreeListSpace) merely checks 2780 // if linear allocation blocks need to be appropriately marked to allow the 2781 // the blocks to be parsable. We also check here whether we need to nudge the 2782 // CMS collector thread to start a new cycle (if it's not already active). 2783 assert( Thread::current()->is_VM_thread() 2784 || ( CMSScavengeBeforeRemark 2785 && Thread::current()->is_ConcurrentGC_thread()), 2786 "Incorrect thread type for epilogue execution"); 2787 2788 if (!_between_prologue_and_epilogue) { 2789 // We have already been invoked; this is a gc_epilogue delegation 2790 // from yet another CMS generation that we are responsible for, just 2791 // ignore it since all relevant work has already been done. 2792 return; 2793 } 2794 assert(haveFreelistLocks(), "must have freelist locks"); 2795 assert_lock_strong(bitMapLock()); 2796 2797 _ct->klass_rem_set()->set_accumulate_modified_oops(false); 2798 2799 _cmsGen->gc_epilogue_work(full); 2800 2801 if (_collectorState == AbortablePreclean || _collectorState == Precleaning) { 2802 // in case sampling was not already enabled, enable it 2803 _start_sampling = true; 2804 } 2805 // reset _eden_chunk_array so sampling starts afresh 2806 _eden_chunk_index = 0; 2807 2808 size_t cms_used = _cmsGen->cmsSpace()->used(); 2809 2810 // update performance counters - this uses a special version of 2811 // update_counters() that allows the utilization to be passed as a 2812 // parameter, avoiding multiple calls to used(). 2813 // 2814 _cmsGen->update_counters(cms_used); 2815 2816 if (CMSIncrementalMode) { 2817 icms_update_allocation_limits(); 2818 } 2819 2820 bitMapLock()->unlock(); 2821 releaseFreelistLocks(); 2822 2823 if (!CleanChunkPoolAsync) { 2824 Chunk::clean_chunk_pool(); 2825 } 2826 2827 set_did_compact(false); 2828 _between_prologue_and_epilogue = false; // ready for next cycle 2829 } 2830 2831 void ConcurrentMarkSweepGeneration::gc_epilogue(bool full) { 2832 collector()->gc_epilogue(full); 2833 2834 // Also reset promotion tracking in par gc thread states. 2835 if (CollectedHeap::use_parallel_gc_threads()) { 2836 for (uint i = 0; i < ParallelGCThreads; i++) { 2837 _par_gc_thread_states[i]->promo.stopTrackingPromotions(i); 2838 } 2839 } 2840 } 2841 2842 void ConcurrentMarkSweepGeneration::gc_epilogue_work(bool full) { 2843 assert(!incremental_collection_failed(), "Should have been cleared"); 2844 cmsSpace()->setPreconsumptionDirtyCardClosure(NULL); 2845 cmsSpace()->gc_epilogue(); 2846 // Print stat counters 2847 NOT_PRODUCT( 2848 assert(_numObjectsAllocated == 0, "check"); 2849 assert(_numWordsAllocated == 0, "check"); 2850 if (Verbose && PrintGC) { 2851 gclog_or_tty->print("Promoted "SIZE_FORMAT" objects, " 2852 SIZE_FORMAT" bytes", 2853 _numObjectsPromoted, _numWordsPromoted*sizeof(HeapWord)); 2854 } 2855 _numObjectsPromoted = 0; 2856 _numWordsPromoted = 0; 2857 ) 2858 2859 if (PrintGC && Verbose) { 2860 // Call down the chain in contiguous_available needs the freelistLock 2861 // so print this out before releasing the freeListLock. 2862 gclog_or_tty->print(" Contiguous available "SIZE_FORMAT" bytes ", 2863 contiguous_available()); 2864 } 2865 } 2866 2867 #ifndef PRODUCT 2868 bool CMSCollector::have_cms_token() { 2869 Thread* thr = Thread::current(); 2870 if (thr->is_VM_thread()) { 2871 return ConcurrentMarkSweepThread::vm_thread_has_cms_token(); 2872 } else if (thr->is_ConcurrentGC_thread()) { 2873 return ConcurrentMarkSweepThread::cms_thread_has_cms_token(); 2874 } else if (thr->is_GC_task_thread()) { 2875 return ConcurrentMarkSweepThread::vm_thread_has_cms_token() && 2876 ParGCRareEvent_lock->owned_by_self(); 2877 } 2878 return false; 2879 } 2880 #endif 2881 2882 // Check reachability of the given heap address in CMS generation, 2883 // treating all other generations as roots. 2884 bool CMSCollector::is_cms_reachable(HeapWord* addr) { 2885 // We could "guarantee" below, rather than assert, but i'll 2886 // leave these as "asserts" so that an adventurous debugger 2887 // could try this in the product build provided some subset of 2888 // the conditions were met, provided they were intersted in the 2889 // results and knew that the computation below wouldn't interfere 2890 // with other concurrent computations mutating the structures 2891 // being read or written. 2892 assert(SafepointSynchronize::is_at_safepoint(), 2893 "Else mutations in object graph will make answer suspect"); 2894 assert(have_cms_token(), "Should hold cms token"); 2895 assert(haveFreelistLocks(), "must hold free list locks"); 2896 assert_lock_strong(bitMapLock()); 2897 2898 // Clear the marking bit map array before starting, but, just 2899 // for kicks, first report if the given address is already marked 2900 gclog_or_tty->print_cr("Start: Address 0x%x is%s marked", addr, 2901 _markBitMap.isMarked(addr) ? "" : " not"); 2902 2903 if (verify_after_remark()) { 2904 MutexLockerEx x(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag); 2905 bool result = verification_mark_bm()->isMarked(addr); 2906 gclog_or_tty->print_cr("TransitiveMark: Address 0x%x %s marked", addr, 2907 result ? "IS" : "is NOT"); 2908 return result; 2909 } else { 2910 gclog_or_tty->print_cr("Could not compute result"); 2911 return false; 2912 } 2913 } 2914 2915 2916 void 2917 CMSCollector::print_on_error(outputStream* st) { 2918 CMSCollector* collector = ConcurrentMarkSweepGeneration::_collector; 2919 if (collector != NULL) { 2920 CMSBitMap* bitmap = &collector->_markBitMap; 2921 st->print_cr("Marking Bits: (CMSBitMap*) " PTR_FORMAT, bitmap); 2922 bitmap->print_on_error(st, " Bits: "); 2923 2924 st->cr(); 2925 2926 CMSBitMap* mut_bitmap = &collector->_modUnionTable; 2927 st->print_cr("Mod Union Table: (CMSBitMap*) " PTR_FORMAT, mut_bitmap); 2928 mut_bitmap->print_on_error(st, " Bits: "); 2929 } 2930 } 2931 2932 //////////////////////////////////////////////////////// 2933 // CMS Verification Support 2934 //////////////////////////////////////////////////////// 2935 // Following the remark phase, the following invariant 2936 // should hold -- each object in the CMS heap which is 2937 // marked in markBitMap() should be marked in the verification_mark_bm(). 2938 2939 class VerifyMarkedClosure: public BitMapClosure { 2940 CMSBitMap* _marks; 2941 bool _failed; 2942 2943 public: 2944 VerifyMarkedClosure(CMSBitMap* bm): _marks(bm), _failed(false) {} 2945 2946 bool do_bit(size_t offset) { 2947 HeapWord* addr = _marks->offsetToHeapWord(offset); 2948 if (!_marks->isMarked(addr)) { 2949 oop(addr)->print_on(gclog_or_tty); 2950 gclog_or_tty->print_cr(" ("INTPTR_FORMAT" should have been marked)", addr); 2951 _failed = true; 2952 } 2953 return true; 2954 } 2955 2956 bool failed() { return _failed; } 2957 }; 2958 2959 bool CMSCollector::verify_after_remark(bool silent) { 2960 if (!silent) gclog_or_tty->print(" [Verifying CMS Marking... "); 2961 MutexLockerEx ml(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag); 2962 static bool init = false; 2963 2964 assert(SafepointSynchronize::is_at_safepoint(), 2965 "Else mutations in object graph will make answer suspect"); 2966 assert(have_cms_token(), 2967 "Else there may be mutual interference in use of " 2968 " verification data structures"); 2969 assert(_collectorState > Marking && _collectorState <= Sweeping, 2970 "Else marking info checked here may be obsolete"); 2971 assert(haveFreelistLocks(), "must hold free list locks"); 2972 assert_lock_strong(bitMapLock()); 2973 2974 2975 // Allocate marking bit map if not already allocated 2976 if (!init) { // first time 2977 if (!verification_mark_bm()->allocate(_span)) { 2978 return false; 2979 } 2980 init = true; 2981 } 2982 2983 assert(verification_mark_stack()->isEmpty(), "Should be empty"); 2984 2985 // Turn off refs discovery -- so we will be tracing through refs. 2986 // This is as intended, because by this time 2987 // GC must already have cleared any refs that need to be cleared, 2988 // and traced those that need to be marked; moreover, 2989 // the marking done here is not going to intefere in any 2990 // way with the marking information used by GC. 2991 NoRefDiscovery no_discovery(ref_processor()); 2992 2993 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;) 2994 2995 // Clear any marks from a previous round 2996 verification_mark_bm()->clear_all(); 2997 assert(verification_mark_stack()->isEmpty(), "markStack should be empty"); 2998 verify_work_stacks_empty(); 2999 3000 GenCollectedHeap* gch = GenCollectedHeap::heap(); 3001 gch->ensure_parsability(false); // fill TLABs, but no need to retire them 3002 // Update the saved marks which may affect the root scans. 3003 gch->save_marks(); 3004 3005 if (CMSRemarkVerifyVariant == 1) { 3006 // In this first variant of verification, we complete 3007 // all marking, then check if the new marks-verctor is 3008 // a subset of the CMS marks-vector. 3009 verify_after_remark_work_1(); 3010 } else if (CMSRemarkVerifyVariant == 2) { 3011 // In this second variant of verification, we flag an error 3012 // (i.e. an object reachable in the new marks-vector not reachable 3013 // in the CMS marks-vector) immediately, also indicating the 3014 // identify of an object (A) that references the unmarked object (B) -- 3015 // presumably, a mutation to A failed to be picked up by preclean/remark? 3016 verify_after_remark_work_2(); 3017 } else { 3018 warning("Unrecognized value %d for CMSRemarkVerifyVariant", 3019 CMSRemarkVerifyVariant); 3020 } 3021 if (!silent) gclog_or_tty->print(" done] "); 3022 return true; 3023 } 3024 3025 void CMSCollector::verify_after_remark_work_1() { 3026 ResourceMark rm; 3027 HandleMark hm; 3028 GenCollectedHeap* gch = GenCollectedHeap::heap(); 3029 3030 // Get a clear set of claim bits for the strong roots processing to work with. 3031 ClassLoaderDataGraph::clear_claimed_marks(); 3032 3033 // Mark from roots one level into CMS 3034 MarkRefsIntoClosure notOlder(_span, verification_mark_bm()); 3035 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 3036 3037 gch->gen_process_strong_roots(_cmsGen->level(), 3038 true, // younger gens are roots 3039 true, // activate StrongRootsScope 3040 false, // not scavenging 3041 SharedHeap::ScanningOption(roots_scanning_options()), 3042 ¬Older, 3043 true, // walk code active on stacks 3044 NULL, 3045 NULL); // SSS: Provide correct closure 3046 3047 // Now mark from the roots 3048 MarkFromRootsClosure markFromRootsClosure(this, _span, 3049 verification_mark_bm(), verification_mark_stack(), 3050 false /* don't yield */, true /* verifying */); 3051 assert(_restart_addr == NULL, "Expected pre-condition"); 3052 verification_mark_bm()->iterate(&markFromRootsClosure); 3053 while (_restart_addr != NULL) { 3054 // Deal with stack overflow: by restarting at the indicated 3055 // address. 3056 HeapWord* ra = _restart_addr; 3057 markFromRootsClosure.reset(ra); 3058 _restart_addr = NULL; 3059 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end()); 3060 } 3061 assert(verification_mark_stack()->isEmpty(), "Should have been drained"); 3062 verify_work_stacks_empty(); 3063 3064 // Marking completed -- now verify that each bit marked in 3065 // verification_mark_bm() is also marked in markBitMap(); flag all 3066 // errors by printing corresponding objects. 3067 VerifyMarkedClosure vcl(markBitMap()); 3068 verification_mark_bm()->iterate(&vcl); 3069 if (vcl.failed()) { 3070 gclog_or_tty->print("Verification failed"); 3071 Universe::heap()->print_on(gclog_or_tty); 3072 fatal("CMS: failed marking verification after remark"); 3073 } 3074 } 3075 3076 class VerifyKlassOopsKlassClosure : public KlassClosure { 3077 class VerifyKlassOopsClosure : public OopClosure { 3078 CMSBitMap* _bitmap; 3079 public: 3080 VerifyKlassOopsClosure(CMSBitMap* bitmap) : _bitmap(bitmap) { } 3081 void do_oop(oop* p) { guarantee(*p == NULL || _bitmap->isMarked((HeapWord*) *p), "Should be marked"); } 3082 void do_oop(narrowOop* p) { ShouldNotReachHere(); } 3083 } _oop_closure; 3084 public: 3085 VerifyKlassOopsKlassClosure(CMSBitMap* bitmap) : _oop_closure(bitmap) {} 3086 void do_klass(Klass* k) { 3087 k->oops_do(&_oop_closure); 3088 } 3089 }; 3090 3091 void CMSCollector::verify_after_remark_work_2() { 3092 ResourceMark rm; 3093 HandleMark hm; 3094 GenCollectedHeap* gch = GenCollectedHeap::heap(); 3095 3096 // Get a clear set of claim bits for the strong roots processing to work with. 3097 ClassLoaderDataGraph::clear_claimed_marks(); 3098 3099 // Mark from roots one level into CMS 3100 MarkRefsIntoVerifyClosure notOlder(_span, verification_mark_bm(), 3101 markBitMap()); 3102 CMKlassClosure klass_closure(¬Older); 3103 3104 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 3105 gch->gen_process_strong_roots(_cmsGen->level(), 3106 true, // younger gens are roots 3107 true, // activate StrongRootsScope 3108 false, // not scavenging 3109 SharedHeap::ScanningOption(roots_scanning_options()), 3110 ¬Older, 3111 true, // walk code active on stacks 3112 NULL, 3113 &klass_closure); 3114 3115 // Now mark from the roots 3116 MarkFromRootsVerifyClosure markFromRootsClosure(this, _span, 3117 verification_mark_bm(), markBitMap(), verification_mark_stack()); 3118 assert(_restart_addr == NULL, "Expected pre-condition"); 3119 verification_mark_bm()->iterate(&markFromRootsClosure); 3120 while (_restart_addr != NULL) { 3121 // Deal with stack overflow: by restarting at the indicated 3122 // address. 3123 HeapWord* ra = _restart_addr; 3124 markFromRootsClosure.reset(ra); 3125 _restart_addr = NULL; 3126 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end()); 3127 } 3128 assert(verification_mark_stack()->isEmpty(), "Should have been drained"); 3129 verify_work_stacks_empty(); 3130 3131 VerifyKlassOopsKlassClosure verify_klass_oops(verification_mark_bm()); 3132 ClassLoaderDataGraph::classes_do(&verify_klass_oops); 3133 3134 // Marking completed -- now verify that each bit marked in 3135 // verification_mark_bm() is also marked in markBitMap(); flag all 3136 // errors by printing corresponding objects. 3137 VerifyMarkedClosure vcl(markBitMap()); 3138 verification_mark_bm()->iterate(&vcl); 3139 assert(!vcl.failed(), "Else verification above should not have succeeded"); 3140 } 3141 3142 void ConcurrentMarkSweepGeneration::save_marks() { 3143 // delegate to CMS space 3144 cmsSpace()->save_marks(); 3145 for (uint i = 0; i < ParallelGCThreads; i++) { 3146 _par_gc_thread_states[i]->promo.startTrackingPromotions(); 3147 } 3148 } 3149 3150 bool ConcurrentMarkSweepGeneration::no_allocs_since_save_marks() { 3151 return cmsSpace()->no_allocs_since_save_marks(); 3152 } 3153 3154 #define CMS_SINCE_SAVE_MARKS_DEFN(OopClosureType, nv_suffix) \ 3155 \ 3156 void ConcurrentMarkSweepGeneration:: \ 3157 oop_since_save_marks_iterate##nv_suffix(OopClosureType* cl) { \ 3158 cl->set_generation(this); \ 3159 cmsSpace()->oop_since_save_marks_iterate##nv_suffix(cl); \ 3160 cl->reset_generation(); \ 3161 save_marks(); \ 3162 } 3163 3164 ALL_SINCE_SAVE_MARKS_CLOSURES(CMS_SINCE_SAVE_MARKS_DEFN) 3165 3166 void 3167 ConcurrentMarkSweepGeneration::object_iterate_since_last_GC(ObjectClosure* blk) 3168 { 3169 // Not currently implemented; need to do the following. -- ysr. 3170 // dld -- I think that is used for some sort of allocation profiler. So it 3171 // really means the objects allocated by the mutator since the last 3172 // GC. We could potentially implement this cheaply by recording only 3173 // the direct allocations in a side data structure. 3174 // 3175 // I think we probably ought not to be required to support these 3176 // iterations at any arbitrary point; I think there ought to be some 3177 // call to enable/disable allocation profiling in a generation/space, 3178 // and the iterator ought to return the objects allocated in the 3179 // gen/space since the enable call, or the last iterator call (which 3180 // will probably be at a GC.) That way, for gens like CM&S that would 3181 // require some extra data structure to support this, we only pay the 3182 // cost when it's in use... 3183 cmsSpace()->object_iterate_since_last_GC(blk); 3184 } 3185 3186 void 3187 ConcurrentMarkSweepGeneration::younger_refs_iterate(OopsInGenClosure* cl) { 3188 cl->set_generation(this); 3189 younger_refs_in_space_iterate(_cmsSpace, cl); 3190 cl->reset_generation(); 3191 } 3192 3193 void 3194 ConcurrentMarkSweepGeneration::oop_iterate(MemRegion mr, ExtendedOopClosure* cl) { 3195 if (freelistLock()->owned_by_self()) { 3196 Generation::oop_iterate(mr, cl); 3197 } else { 3198 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 3199 Generation::oop_iterate(mr, cl); 3200 } 3201 } 3202 3203 void 3204 ConcurrentMarkSweepGeneration::oop_iterate(ExtendedOopClosure* cl) { 3205 if (freelistLock()->owned_by_self()) { 3206 Generation::oop_iterate(cl); 3207 } else { 3208 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 3209 Generation::oop_iterate(cl); 3210 } 3211 } 3212 3213 void 3214 ConcurrentMarkSweepGeneration::object_iterate(ObjectClosure* cl) { 3215 if (freelistLock()->owned_by_self()) { 3216 Generation::object_iterate(cl); 3217 } else { 3218 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 3219 Generation::object_iterate(cl); 3220 } 3221 } 3222 3223 void 3224 ConcurrentMarkSweepGeneration::safe_object_iterate(ObjectClosure* cl) { 3225 if (freelistLock()->owned_by_self()) { 3226 Generation::safe_object_iterate(cl); 3227 } else { 3228 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 3229 Generation::safe_object_iterate(cl); 3230 } 3231 } 3232 3233 void 3234 ConcurrentMarkSweepGeneration::post_compact() { 3235 } 3236 3237 void 3238 ConcurrentMarkSweepGeneration::prepare_for_verify() { 3239 // Fix the linear allocation blocks to look like free blocks. 3240 3241 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those 3242 // are not called when the heap is verified during universe initialization and 3243 // at vm shutdown. 3244 if (freelistLock()->owned_by_self()) { 3245 cmsSpace()->prepare_for_verify(); 3246 } else { 3247 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag); 3248 cmsSpace()->prepare_for_verify(); 3249 } 3250 } 3251 3252 void 3253 ConcurrentMarkSweepGeneration::verify() { 3254 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those 3255 // are not called when the heap is verified during universe initialization and 3256 // at vm shutdown. 3257 if (freelistLock()->owned_by_self()) { 3258 cmsSpace()->verify(); 3259 } else { 3260 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag); 3261 cmsSpace()->verify(); 3262 } 3263 } 3264 3265 void CMSCollector::verify() { 3266 _cmsGen->verify(); 3267 } 3268 3269 #ifndef PRODUCT 3270 bool CMSCollector::overflow_list_is_empty() const { 3271 assert(_num_par_pushes >= 0, "Inconsistency"); 3272 if (_overflow_list == NULL) { 3273 assert(_num_par_pushes == 0, "Inconsistency"); 3274 } 3275 return _overflow_list == NULL; 3276 } 3277 3278 // The methods verify_work_stacks_empty() and verify_overflow_empty() 3279 // merely consolidate assertion checks that appear to occur together frequently. 3280 void CMSCollector::verify_work_stacks_empty() const { 3281 assert(_markStack.isEmpty(), "Marking stack should be empty"); 3282 assert(overflow_list_is_empty(), "Overflow list should be empty"); 3283 } 3284 3285 void CMSCollector::verify_overflow_empty() const { 3286 assert(overflow_list_is_empty(), "Overflow list should be empty"); 3287 assert(no_preserved_marks(), "No preserved marks"); 3288 } 3289 #endif // PRODUCT 3290 3291 // Decide if we want to enable class unloading as part of the 3292 // ensuing concurrent GC cycle. We will collect and 3293 // unload classes if it's the case that: 3294 // (1) an explicit gc request has been made and the flag 3295 // ExplicitGCInvokesConcurrentAndUnloadsClasses is set, OR 3296 // (2) (a) class unloading is enabled at the command line, and 3297 // (b) old gen is getting really full 3298 // NOTE: Provided there is no change in the state of the heap between 3299 // calls to this method, it should have idempotent results. Moreover, 3300 // its results should be monotonically increasing (i.e. going from 0 to 1, 3301 // but not 1 to 0) between successive calls between which the heap was 3302 // not collected. For the implementation below, it must thus rely on 3303 // the property that concurrent_cycles_since_last_unload() 3304 // will not decrease unless a collection cycle happened and that 3305 // _cmsGen->is_too_full() are 3306 // themselves also monotonic in that sense. See check_monotonicity() 3307 // below. 3308 void CMSCollector::update_should_unload_classes() { 3309 _should_unload_classes = false; 3310 // Condition 1 above 3311 if (_full_gc_requested && ExplicitGCInvokesConcurrentAndUnloadsClasses) { 3312 _should_unload_classes = true; 3313 } else if (CMSClassUnloadingEnabled) { // Condition 2.a above 3314 // Disjuncts 2.b.(i,ii,iii) above 3315 _should_unload_classes = (concurrent_cycles_since_last_unload() >= 3316 CMSClassUnloadingMaxInterval) 3317 || _cmsGen->is_too_full(); 3318 } 3319 } 3320 3321 bool ConcurrentMarkSweepGeneration::is_too_full() const { 3322 bool res = should_concurrent_collect(); 3323 res = res && (occupancy() > (double)CMSIsTooFullPercentage/100.0); 3324 return res; 3325 } 3326 3327 void CMSCollector::setup_cms_unloading_and_verification_state() { 3328 const bool should_verify = VerifyBeforeGC || VerifyAfterGC || VerifyDuringGC 3329 || VerifyBeforeExit; 3330 const int rso = SharedHeap::SO_Strings | SharedHeap::SO_CodeCache; 3331 3332 if (should_unload_classes()) { // Should unload classes this cycle 3333 remove_root_scanning_option(rso); // Shrink the root set appropriately 3334 set_verifying(should_verify); // Set verification state for this cycle 3335 return; // Nothing else needs to be done at this time 3336 } 3337 3338 // Not unloading classes this cycle 3339 assert(!should_unload_classes(), "Inconsitency!"); 3340 if ((!verifying() || unloaded_classes_last_cycle()) && should_verify) { 3341 // Include symbols, strings and code cache elements to prevent their resurrection. 3342 add_root_scanning_option(rso); 3343 set_verifying(true); 3344 } else if (verifying() && !should_verify) { 3345 // We were verifying, but some verification flags got disabled. 3346 set_verifying(false); 3347 // Exclude symbols, strings and code cache elements from root scanning to 3348 // reduce IM and RM pauses. 3349 remove_root_scanning_option(rso); 3350 } 3351 } 3352 3353 3354 #ifndef PRODUCT 3355 HeapWord* CMSCollector::block_start(const void* p) const { 3356 const HeapWord* addr = (HeapWord*)p; 3357 if (_span.contains(p)) { 3358 if (_cmsGen->cmsSpace()->is_in_reserved(addr)) { 3359 return _cmsGen->cmsSpace()->block_start(p); 3360 } 3361 } 3362 return NULL; 3363 } 3364 #endif 3365 3366 HeapWord* 3367 ConcurrentMarkSweepGeneration::expand_and_allocate(size_t word_size, 3368 bool tlab, 3369 bool parallel) { 3370 CMSSynchronousYieldRequest yr; 3371 assert(!tlab, "Can't deal with TLAB allocation"); 3372 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 3373 expand(word_size*HeapWordSize, MinHeapDeltaBytes, 3374 CMSExpansionCause::_satisfy_allocation); 3375 if (GCExpandToAllocateDelayMillis > 0) { 3376 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 3377 } 3378 return have_lock_and_allocate(word_size, tlab); 3379 } 3380 3381 // YSR: All of this generation expansion/shrinking stuff is an exact copy of 3382 // OneContigSpaceCardGeneration, which makes me wonder if we should move this 3383 // to CardGeneration and share it... 3384 bool ConcurrentMarkSweepGeneration::expand(size_t bytes, size_t expand_bytes) { 3385 return CardGeneration::expand(bytes, expand_bytes); 3386 } 3387 3388 void ConcurrentMarkSweepGeneration::expand(size_t bytes, size_t expand_bytes, 3389 CMSExpansionCause::Cause cause) 3390 { 3391 3392 bool success = expand(bytes, expand_bytes); 3393 3394 // remember why we expanded; this information is used 3395 // by shouldConcurrentCollect() when making decisions on whether to start 3396 // a new CMS cycle. 3397 if (success) { 3398 set_expansion_cause(cause); 3399 if (PrintGCDetails && Verbose) { 3400 gclog_or_tty->print_cr("Expanded CMS gen for %s", 3401 CMSExpansionCause::to_string(cause)); 3402 } 3403 } 3404 } 3405 3406 HeapWord* ConcurrentMarkSweepGeneration::expand_and_par_lab_allocate(CMSParGCThreadState* ps, size_t word_sz) { 3407 HeapWord* res = NULL; 3408 MutexLocker x(ParGCRareEvent_lock); 3409 while (true) { 3410 // Expansion by some other thread might make alloc OK now: 3411 res = ps->lab.alloc(word_sz); 3412 if (res != NULL) return res; 3413 // If there's not enough expansion space available, give up. 3414 if (_virtual_space.uncommitted_size() < (word_sz * HeapWordSize)) { 3415 return NULL; 3416 } 3417 // Otherwise, we try expansion. 3418 expand(word_sz*HeapWordSize, MinHeapDeltaBytes, 3419 CMSExpansionCause::_allocate_par_lab); 3420 // Now go around the loop and try alloc again; 3421 // A competing par_promote might beat us to the expansion space, 3422 // so we may go around the loop again if promotion fails agaion. 3423 if (GCExpandToAllocateDelayMillis > 0) { 3424 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 3425 } 3426 } 3427 } 3428 3429 3430 bool ConcurrentMarkSweepGeneration::expand_and_ensure_spooling_space( 3431 PromotionInfo* promo) { 3432 MutexLocker x(ParGCRareEvent_lock); 3433 size_t refill_size_bytes = promo->refillSize() * HeapWordSize; 3434 while (true) { 3435 // Expansion by some other thread might make alloc OK now: 3436 if (promo->ensure_spooling_space()) { 3437 assert(promo->has_spooling_space(), 3438 "Post-condition of successful ensure_spooling_space()"); 3439 return true; 3440 } 3441 // If there's not enough expansion space available, give up. 3442 if (_virtual_space.uncommitted_size() < refill_size_bytes) { 3443 return false; 3444 } 3445 // Otherwise, we try expansion. 3446 expand(refill_size_bytes, MinHeapDeltaBytes, 3447 CMSExpansionCause::_allocate_par_spooling_space); 3448 // Now go around the loop and try alloc again; 3449 // A competing allocation might beat us to the expansion space, 3450 // so we may go around the loop again if allocation fails again. 3451 if (GCExpandToAllocateDelayMillis > 0) { 3452 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 3453 } 3454 } 3455 } 3456 3457 3458 void ConcurrentMarkSweepGeneration::shrink_by(size_t bytes) { 3459 assert_locked_or_safepoint(ExpandHeap_lock); 3460 // Shrink committed space 3461 _virtual_space.shrink_by(bytes); 3462 // Shrink space; this also shrinks the space's BOT 3463 _cmsSpace->set_end((HeapWord*) _virtual_space.high()); 3464 size_t new_word_size = heap_word_size(_cmsSpace->capacity()); 3465 // Shrink the shared block offset array 3466 _bts->resize(new_word_size); 3467 MemRegion mr(_cmsSpace->bottom(), new_word_size); 3468 // Shrink the card table 3469 Universe::heap()->barrier_set()->resize_covered_region(mr); 3470 3471 if (Verbose && PrintGC) { 3472 size_t new_mem_size = _virtual_space.committed_size(); 3473 size_t old_mem_size = new_mem_size + bytes; 3474 gclog_or_tty->print_cr("Shrinking %s from " SIZE_FORMAT "K to " SIZE_FORMAT "K", 3475 name(), old_mem_size/K, new_mem_size/K); 3476 } 3477 } 3478 3479 void ConcurrentMarkSweepGeneration::shrink(size_t bytes) { 3480 assert_locked_or_safepoint(Heap_lock); 3481 size_t size = ReservedSpace::page_align_size_down(bytes); 3482 if (size > 0) { 3483 shrink_by(size); 3484 } 3485 } 3486 3487 bool ConcurrentMarkSweepGeneration::grow_by(size_t bytes) { 3488 assert_locked_or_safepoint(Heap_lock); 3489 bool result = _virtual_space.expand_by(bytes); 3490 if (result) { 3491 size_t new_word_size = 3492 heap_word_size(_virtual_space.committed_size()); 3493 MemRegion mr(_cmsSpace->bottom(), new_word_size); 3494 _bts->resize(new_word_size); // resize the block offset shared array 3495 Universe::heap()->barrier_set()->resize_covered_region(mr); 3496 // Hmmmm... why doesn't CFLS::set_end verify locking? 3497 // This is quite ugly; FIX ME XXX 3498 _cmsSpace->assert_locked(freelistLock()); 3499 _cmsSpace->set_end((HeapWord*)_virtual_space.high()); 3500 3501 // update the space and generation capacity counters 3502 if (UsePerfData) { 3503 _space_counters->update_capacity(); 3504 _gen_counters->update_all(); 3505 } 3506 3507 if (Verbose && PrintGC) { 3508 size_t new_mem_size = _virtual_space.committed_size(); 3509 size_t old_mem_size = new_mem_size - bytes; 3510 gclog_or_tty->print_cr("Expanding %s from " SIZE_FORMAT "K by " SIZE_FORMAT "K to " SIZE_FORMAT "K", 3511 name(), old_mem_size/K, bytes/K, new_mem_size/K); 3512 } 3513 } 3514 return result; 3515 } 3516 3517 bool ConcurrentMarkSweepGeneration::grow_to_reserved() { 3518 assert_locked_or_safepoint(Heap_lock); 3519 bool success = true; 3520 const size_t remaining_bytes = _virtual_space.uncommitted_size(); 3521 if (remaining_bytes > 0) { 3522 success = grow_by(remaining_bytes); 3523 DEBUG_ONLY(if (!success) warning("grow to reserved failed");) 3524 } 3525 return success; 3526 } 3527 3528 void ConcurrentMarkSweepGeneration::shrink_free_list_by(size_t bytes) { 3529 assert_locked_or_safepoint(Heap_lock); 3530 assert_lock_strong(freelistLock()); 3531 if (PrintGCDetails && Verbose) { 3532 warning("Shrinking of CMS not yet implemented"); 3533 } 3534 return; 3535 } 3536 3537 3538 // Simple ctor/dtor wrapper for accounting & timer chores around concurrent 3539 // phases. 3540 class CMSPhaseAccounting: public StackObj { 3541 public: 3542 CMSPhaseAccounting(CMSCollector *collector, 3543 const char *phase, 3544 bool print_cr = true); 3545 ~CMSPhaseAccounting(); 3546 3547 private: 3548 CMSCollector *_collector; 3549 const char *_phase; 3550 elapsedTimer _wallclock; 3551 bool _print_cr; 3552 3553 public: 3554 // Not MT-safe; so do not pass around these StackObj's 3555 // where they may be accessed by other threads. 3556 jlong wallclock_millis() { 3557 assert(_wallclock.is_active(), "Wall clock should not stop"); 3558 _wallclock.stop(); // to record time 3559 jlong ret = _wallclock.milliseconds(); 3560 _wallclock.start(); // restart 3561 return ret; 3562 } 3563 }; 3564 3565 CMSPhaseAccounting::CMSPhaseAccounting(CMSCollector *collector, 3566 const char *phase, 3567 bool print_cr) : 3568 _collector(collector), _phase(phase), _print_cr(print_cr) { 3569 3570 if (PrintCMSStatistics != 0) { 3571 _collector->resetYields(); 3572 } 3573 if (PrintGCDetails) { 3574 gclog_or_tty->date_stamp(PrintGCDateStamps); 3575 gclog_or_tty->stamp(PrintGCTimeStamps); 3576 gclog_or_tty->print_cr("[%s-concurrent-%s-start]", 3577 _collector->cmsGen()->short_name(), _phase); 3578 } 3579 _collector->resetTimer(); 3580 _wallclock.start(); 3581 _collector->startTimer(); 3582 } 3583 3584 CMSPhaseAccounting::~CMSPhaseAccounting() { 3585 assert(_wallclock.is_active(), "Wall clock should not have stopped"); 3586 _collector->stopTimer(); 3587 _wallclock.stop(); 3588 if (PrintGCDetails) { 3589 gclog_or_tty->date_stamp(PrintGCDateStamps); 3590 gclog_or_tty->stamp(PrintGCTimeStamps); 3591 gclog_or_tty->print("[%s-concurrent-%s: %3.3f/%3.3f secs]", 3592 _collector->cmsGen()->short_name(), 3593 _phase, _collector->timerValue(), _wallclock.seconds()); 3594 if (_print_cr) { 3595 gclog_or_tty->print_cr(""); 3596 } 3597 if (PrintCMSStatistics != 0) { 3598 gclog_or_tty->print_cr(" (CMS-concurrent-%s yielded %d times)", _phase, 3599 _collector->yields()); 3600 } 3601 } 3602 } 3603 3604 // CMS work 3605 3606 // Checkpoint the roots into this generation from outside 3607 // this generation. [Note this initial checkpoint need only 3608 // be approximate -- we'll do a catch up phase subsequently.] 3609 void CMSCollector::checkpointRootsInitial(bool asynch) { 3610 assert(_collectorState == InitialMarking, "Wrong collector state"); 3611 check_correct_thread_executing(); 3612 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 3613 3614 save_heap_summary(); 3615 report_heap_summary(GCWhen::BeforeGC); 3616 3617 ReferenceProcessor* rp = ref_processor(); 3618 SpecializationStats::clear(); 3619 assert(_restart_addr == NULL, "Control point invariant"); 3620 if (asynch) { 3621 // acquire locks for subsequent manipulations 3622 MutexLockerEx x(bitMapLock(), 3623 Mutex::_no_safepoint_check_flag); 3624 checkpointRootsInitialWork(asynch); 3625 // enable ("weak") refs discovery 3626 rp->enable_discovery(true /*verify_disabled*/, true /*check_no_refs*/); 3627 _collectorState = Marking; 3628 } else { 3629 // (Weak) Refs discovery: this is controlled from genCollectedHeap::do_collection 3630 // which recognizes if we are a CMS generation, and doesn't try to turn on 3631 // discovery; verify that they aren't meddling. 3632 assert(!rp->discovery_is_atomic(), 3633 "incorrect setting of discovery predicate"); 3634 assert(!rp->discovery_enabled(), "genCollectedHeap shouldn't control " 3635 "ref discovery for this generation kind"); 3636 // already have locks 3637 checkpointRootsInitialWork(asynch); 3638 // now enable ("weak") refs discovery 3639 rp->enable_discovery(true /*verify_disabled*/, false /*verify_no_refs*/); 3640 _collectorState = Marking; 3641 } 3642 SpecializationStats::print(); 3643 } 3644 3645 void CMSCollector::checkpointRootsInitialWork(bool asynch) { 3646 assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped"); 3647 assert(_collectorState == InitialMarking, "just checking"); 3648 3649 // If there has not been a GC[n-1] since last GC[n] cycle completed, 3650 // precede our marking with a collection of all 3651 // younger generations to keep floating garbage to a minimum. 3652 // XXX: we won't do this for now -- it's an optimization to be done later. 3653 3654 // already have locks 3655 assert_lock_strong(bitMapLock()); 3656 assert(_markBitMap.isAllClear(), "was reset at end of previous cycle"); 3657 3658 // Setup the verification and class unloading state for this 3659 // CMS collection cycle. 3660 setup_cms_unloading_and_verification_state(); 3661 3662 NOT_PRODUCT(GCTraceTime t("\ncheckpointRootsInitialWork", 3663 PrintGCDetails && Verbose, true, _gc_timer_cm);) 3664 if (UseAdaptiveSizePolicy) { 3665 size_policy()->checkpoint_roots_initial_begin(); 3666 } 3667 3668 // Reset all the PLAB chunk arrays if necessary. 3669 if (_survivor_plab_array != NULL && !CMSPLABRecordAlways) { 3670 reset_survivor_plab_arrays(); 3671 } 3672 3673 ResourceMark rm; 3674 HandleMark hm; 3675 3676 FalseClosure falseClosure; 3677 // In the case of a synchronous collection, we will elide the 3678 // remark step, so it's important to catch all the nmethod oops 3679 // in this step. 3680 // The final 'true' flag to gen_process_strong_roots will ensure this. 3681 // If 'async' is true, we can relax the nmethod tracing. 3682 MarkRefsIntoClosure notOlder(_span, &_markBitMap); 3683 GenCollectedHeap* gch = GenCollectedHeap::heap(); 3684 3685 verify_work_stacks_empty(); 3686 verify_overflow_empty(); 3687 3688 gch->ensure_parsability(false); // fill TLABs, but no need to retire them 3689 // Update the saved marks which may affect the root scans. 3690 gch->save_marks(); 3691 3692 // weak reference processing has not started yet. 3693 ref_processor()->set_enqueuing_is_done(false); 3694 3695 // Need to remember all newly created CLDs, 3696 // so that we can guarantee that the remark finds them. 3697 ClassLoaderDataGraph::remember_new_clds(true); 3698 3699 // Whenever a CLD is found, it will be claimed before proceeding to mark 3700 // the klasses. The claimed marks need to be cleared before marking starts. 3701 ClassLoaderDataGraph::clear_claimed_marks(); 3702 3703 if (CMSPrintEdenSurvivorChunks) { 3704 print_eden_and_survivor_chunk_arrays(); 3705 } 3706 3707 CMKlassClosure klass_closure(¬Older); 3708 { 3709 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;) 3710 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 3711 gch->gen_process_strong_roots(_cmsGen->level(), 3712 true, // younger gens are roots 3713 true, // activate StrongRootsScope 3714 false, // not scavenging 3715 SharedHeap::ScanningOption(roots_scanning_options()), 3716 ¬Older, 3717 true, // walk all of code cache if (so & SO_CodeCache) 3718 NULL, 3719 &klass_closure); 3720 } 3721 3722 // Clear mod-union table; it will be dirtied in the prologue of 3723 // CMS generation per each younger generation collection. 3724 3725 assert(_modUnionTable.isAllClear(), 3726 "Was cleared in most recent final checkpoint phase" 3727 " or no bits are set in the gc_prologue before the start of the next " 3728 "subsequent marking phase."); 3729 3730 assert(_ct->klass_rem_set()->mod_union_is_clear(), "Must be"); 3731 3732 // Save the end of the used_region of the constituent generations 3733 // to be used to limit the extent of sweep in each generation. 3734 save_sweep_limits(); 3735 if (UseAdaptiveSizePolicy) { 3736 size_policy()->checkpoint_roots_initial_end(gch->gc_cause()); 3737 } 3738 verify_overflow_empty(); 3739 } 3740 3741 bool CMSCollector::markFromRoots(bool asynch) { 3742 // we might be tempted to assert that: 3743 // assert(asynch == !SafepointSynchronize::is_at_safepoint(), 3744 // "inconsistent argument?"); 3745 // However that wouldn't be right, because it's possible that 3746 // a safepoint is indeed in progress as a younger generation 3747 // stop-the-world GC happens even as we mark in this generation. 3748 assert(_collectorState == Marking, "inconsistent state?"); 3749 check_correct_thread_executing(); 3750 verify_overflow_empty(); 3751 3752 bool res; 3753 if (asynch) { 3754 3755 // Start the timers for adaptive size policy for the concurrent phases 3756 // Do it here so that the foreground MS can use the concurrent 3757 // timer since a foreground MS might has the sweep done concurrently 3758 // or STW. 3759 if (UseAdaptiveSizePolicy) { 3760 size_policy()->concurrent_marking_begin(); 3761 } 3762 3763 // Weak ref discovery note: We may be discovering weak 3764 // refs in this generation concurrent (but interleaved) with 3765 // weak ref discovery by a younger generation collector. 3766 3767 CMSTokenSyncWithLocks ts(true, bitMapLock()); 3768 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 3769 CMSPhaseAccounting pa(this, "mark", !PrintGCDetails); 3770 res = markFromRootsWork(asynch); 3771 if (res) { 3772 _collectorState = Precleaning; 3773 } else { // We failed and a foreground collection wants to take over 3774 assert(_foregroundGCIsActive, "internal state inconsistency"); 3775 assert(_restart_addr == NULL, "foreground will restart from scratch"); 3776 if (PrintGCDetails) { 3777 gclog_or_tty->print_cr("bailing out to foreground collection"); 3778 } 3779 } 3780 if (UseAdaptiveSizePolicy) { 3781 size_policy()->concurrent_marking_end(); 3782 } 3783 } else { 3784 assert(SafepointSynchronize::is_at_safepoint(), 3785 "inconsistent with asynch == false"); 3786 if (UseAdaptiveSizePolicy) { 3787 size_policy()->ms_collection_marking_begin(); 3788 } 3789 // already have locks 3790 res = markFromRootsWork(asynch); 3791 _collectorState = FinalMarking; 3792 if (UseAdaptiveSizePolicy) { 3793 GenCollectedHeap* gch = GenCollectedHeap::heap(); 3794 size_policy()->ms_collection_marking_end(gch->gc_cause()); 3795 } 3796 } 3797 verify_overflow_empty(); 3798 return res; 3799 } 3800 3801 bool CMSCollector::markFromRootsWork(bool asynch) { 3802 // iterate over marked bits in bit map, doing a full scan and mark 3803 // from these roots using the following algorithm: 3804 // . if oop is to the right of the current scan pointer, 3805 // mark corresponding bit (we'll process it later) 3806 // . else (oop is to left of current scan pointer) 3807 // push oop on marking stack 3808 // . drain the marking stack 3809 3810 // Note that when we do a marking step we need to hold the 3811 // bit map lock -- recall that direct allocation (by mutators) 3812 // and promotion (by younger generation collectors) is also 3813 // marking the bit map. [the so-called allocate live policy.] 3814 // Because the implementation of bit map marking is not 3815 // robust wrt simultaneous marking of bits in the same word, 3816 // we need to make sure that there is no such interference 3817 // between concurrent such updates. 3818 3819 // already have locks 3820 assert_lock_strong(bitMapLock()); 3821 3822 verify_work_stacks_empty(); 3823 verify_overflow_empty(); 3824 bool result = false; 3825 if (CMSConcurrentMTEnabled && ConcGCThreads > 0) { 3826 result = do_marking_mt(asynch); 3827 } else { 3828 result = do_marking_st(asynch); 3829 } 3830 return result; 3831 } 3832 3833 // Forward decl 3834 class CMSConcMarkingTask; 3835 3836 class CMSConcMarkingTerminator: public ParallelTaskTerminator { 3837 CMSCollector* _collector; 3838 CMSConcMarkingTask* _task; 3839 public: 3840 virtual void yield(); 3841 3842 // "n_threads" is the number of threads to be terminated. 3843 // "queue_set" is a set of work queues of other threads. 3844 // "collector" is the CMS collector associated with this task terminator. 3845 // "yield" indicates whether we need the gang as a whole to yield. 3846 CMSConcMarkingTerminator(int n_threads, TaskQueueSetSuper* queue_set, CMSCollector* collector) : 3847 ParallelTaskTerminator(n_threads, queue_set), 3848 _collector(collector) { } 3849 3850 void set_task(CMSConcMarkingTask* task) { 3851 _task = task; 3852 } 3853 }; 3854 3855 class CMSConcMarkingTerminatorTerminator: public TerminatorTerminator { 3856 CMSConcMarkingTask* _task; 3857 public: 3858 bool should_exit_termination(); 3859 void set_task(CMSConcMarkingTask* task) { 3860 _task = task; 3861 } 3862 }; 3863 3864 // MT Concurrent Marking Task 3865 class CMSConcMarkingTask: public YieldingFlexibleGangTask { 3866 CMSCollector* _collector; 3867 int _n_workers; // requested/desired # workers 3868 bool _asynch; 3869 bool _result; 3870 CompactibleFreeListSpace* _cms_space; 3871 char _pad_front[64]; // padding to ... 3872 HeapWord* _global_finger; // ... avoid sharing cache line 3873 char _pad_back[64]; 3874 HeapWord* _restart_addr; 3875 3876 // Exposed here for yielding support 3877 Mutex* const _bit_map_lock; 3878 3879 // The per thread work queues, available here for stealing 3880 OopTaskQueueSet* _task_queues; 3881 3882 // Termination (and yielding) support 3883 CMSConcMarkingTerminator _term; 3884 CMSConcMarkingTerminatorTerminator _term_term; 3885 3886 public: 3887 CMSConcMarkingTask(CMSCollector* collector, 3888 CompactibleFreeListSpace* cms_space, 3889 bool asynch, 3890 YieldingFlexibleWorkGang* workers, 3891 OopTaskQueueSet* task_queues): 3892 YieldingFlexibleGangTask("Concurrent marking done multi-threaded"), 3893 _collector(collector), 3894 _cms_space(cms_space), 3895 _asynch(asynch), _n_workers(0), _result(true), 3896 _task_queues(task_queues), 3897 _term(_n_workers, task_queues, _collector), 3898 _bit_map_lock(collector->bitMapLock()) 3899 { 3900 _requested_size = _n_workers; 3901 _term.set_task(this); 3902 _term_term.set_task(this); 3903 _restart_addr = _global_finger = _cms_space->bottom(); 3904 } 3905 3906 3907 OopTaskQueueSet* task_queues() { return _task_queues; } 3908 3909 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 3910 3911 HeapWord** global_finger_addr() { return &_global_finger; } 3912 3913 CMSConcMarkingTerminator* terminator() { return &_term; } 3914 3915 virtual void set_for_termination(int active_workers) { 3916 terminator()->reset_for_reuse(active_workers); 3917 } 3918 3919 void work(uint worker_id); 3920 bool should_yield() { 3921 return ConcurrentMarkSweepThread::should_yield() 3922 && !_collector->foregroundGCIsActive() 3923 && _asynch; 3924 } 3925 3926 virtual void coordinator_yield(); // stuff done by coordinator 3927 bool result() { return _result; } 3928 3929 void reset(HeapWord* ra) { 3930 assert(_global_finger >= _cms_space->end(), "Postcondition of ::work(i)"); 3931 _restart_addr = _global_finger = ra; 3932 _term.reset_for_reuse(); 3933 } 3934 3935 static bool get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 3936 OopTaskQueue* work_q); 3937 3938 private: 3939 void do_scan_and_mark(int i, CompactibleFreeListSpace* sp); 3940 void do_work_steal(int i); 3941 void bump_global_finger(HeapWord* f); 3942 }; 3943 3944 bool CMSConcMarkingTerminatorTerminator::should_exit_termination() { 3945 assert(_task != NULL, "Error"); 3946 return _task->yielding(); 3947 // Note that we do not need the disjunct || _task->should_yield() above 3948 // because we want terminating threads to yield only if the task 3949 // is already in the midst of yielding, which happens only after at least one 3950 // thread has yielded. 3951 } 3952 3953 void CMSConcMarkingTerminator::yield() { 3954 if (_task->should_yield()) { 3955 _task->yield(); 3956 } else { 3957 ParallelTaskTerminator::yield(); 3958 } 3959 } 3960 3961 //////////////////////////////////////////////////////////////// 3962 // Concurrent Marking Algorithm Sketch 3963 //////////////////////////////////////////////////////////////// 3964 // Until all tasks exhausted (both spaces): 3965 // -- claim next available chunk 3966 // -- bump global finger via CAS 3967 // -- find first object that starts in this chunk 3968 // and start scanning bitmap from that position 3969 // -- scan marked objects for oops 3970 // -- CAS-mark target, and if successful: 3971 // . if target oop is above global finger (volatile read) 3972 // nothing to do 3973 // . if target oop is in chunk and above local finger 3974 // then nothing to do 3975 // . else push on work-queue 3976 // -- Deal with possible overflow issues: 3977 // . local work-queue overflow causes stuff to be pushed on 3978 // global (common) overflow queue 3979 // . always first empty local work queue 3980 // . then get a batch of oops from global work queue if any 3981 // . then do work stealing 3982 // -- When all tasks claimed (both spaces) 3983 // and local work queue empty, 3984 // then in a loop do: 3985 // . check global overflow stack; steal a batch of oops and trace 3986 // . try to steal from other threads oif GOS is empty 3987 // . if neither is available, offer termination 3988 // -- Terminate and return result 3989 // 3990 void CMSConcMarkingTask::work(uint worker_id) { 3991 elapsedTimer _timer; 3992 ResourceMark rm; 3993 HandleMark hm; 3994 3995 DEBUG_ONLY(_collector->verify_overflow_empty();) 3996 3997 // Before we begin work, our work queue should be empty 3998 assert(work_queue(worker_id)->size() == 0, "Expected to be empty"); 3999 // Scan the bitmap covering _cms_space, tracing through grey objects. 4000 _timer.start(); 4001 do_scan_and_mark(worker_id, _cms_space); 4002 _timer.stop(); 4003 if (PrintCMSStatistics != 0) { 4004 gclog_or_tty->print_cr("Finished cms space scanning in %dth thread: %3.3f sec", 4005 worker_id, _timer.seconds()); 4006 // XXX: need xxx/xxx type of notation, two timers 4007 } 4008 4009 // ... do work stealing 4010 _timer.reset(); 4011 _timer.start(); 4012 do_work_steal(worker_id); 4013 _timer.stop(); 4014 if (PrintCMSStatistics != 0) { 4015 gclog_or_tty->print_cr("Finished work stealing in %dth thread: %3.3f sec", 4016 worker_id, _timer.seconds()); 4017 // XXX: need xxx/xxx type of notation, two timers 4018 } 4019 assert(_collector->_markStack.isEmpty(), "Should have been emptied"); 4020 assert(work_queue(worker_id)->size() == 0, "Should have been emptied"); 4021 // Note that under the current task protocol, the 4022 // following assertion is true even of the spaces 4023 // expanded since the completion of the concurrent 4024 // marking. XXX This will likely change under a strict 4025 // ABORT semantics. 4026 // After perm removal the comparison was changed to 4027 // greater than or equal to from strictly greater than. 4028 // Before perm removal the highest address sweep would 4029 // have been at the end of perm gen but now is at the 4030 // end of the tenured gen. 4031 assert(_global_finger >= _cms_space->end(), 4032 "All tasks have been completed"); 4033 DEBUG_ONLY(_collector->verify_overflow_empty();) 4034 } 4035 4036 void CMSConcMarkingTask::bump_global_finger(HeapWord* f) { 4037 HeapWord* read = _global_finger; 4038 HeapWord* cur = read; 4039 while (f > read) { 4040 cur = read; 4041 read = (HeapWord*) Atomic::cmpxchg_ptr(f, &_global_finger, cur); 4042 if (cur == read) { 4043 // our cas succeeded 4044 assert(_global_finger >= f, "protocol consistency"); 4045 break; 4046 } 4047 } 4048 } 4049 4050 // This is really inefficient, and should be redone by 4051 // using (not yet available) block-read and -write interfaces to the 4052 // stack and the work_queue. XXX FIX ME !!! 4053 bool CMSConcMarkingTask::get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 4054 OopTaskQueue* work_q) { 4055 // Fast lock-free check 4056 if (ovflw_stk->length() == 0) { 4057 return false; 4058 } 4059 assert(work_q->size() == 0, "Shouldn't steal"); 4060 MutexLockerEx ml(ovflw_stk->par_lock(), 4061 Mutex::_no_safepoint_check_flag); 4062 // Grab up to 1/4 the size of the work queue 4063 size_t num = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 4064 (size_t)ParGCDesiredObjsFromOverflowList); 4065 num = MIN2(num, ovflw_stk->length()); 4066 for (int i = (int) num; i > 0; i--) { 4067 oop cur = ovflw_stk->pop(); 4068 assert(cur != NULL, "Counted wrong?"); 4069 work_q->push(cur); 4070 } 4071 return num > 0; 4072 } 4073 4074 void CMSConcMarkingTask::do_scan_and_mark(int i, CompactibleFreeListSpace* sp) { 4075 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 4076 int n_tasks = pst->n_tasks(); 4077 // We allow that there may be no tasks to do here because 4078 // we are restarting after a stack overflow. 4079 assert(pst->valid() || n_tasks == 0, "Uninitialized use?"); 4080 uint nth_task = 0; 4081 4082 HeapWord* aligned_start = sp->bottom(); 4083 if (sp->used_region().contains(_restart_addr)) { 4084 // Align down to a card boundary for the start of 0th task 4085 // for this space. 4086 aligned_start = 4087 (HeapWord*)align_size_down((uintptr_t)_restart_addr, 4088 CardTableModRefBS::card_size); 4089 } 4090 4091 size_t chunk_size = sp->marking_task_size(); 4092 while (!pst->is_task_claimed(/* reference */ nth_task)) { 4093 // Having claimed the nth task in this space, 4094 // compute the chunk that it corresponds to: 4095 MemRegion span = MemRegion(aligned_start + nth_task*chunk_size, 4096 aligned_start + (nth_task+1)*chunk_size); 4097 // Try and bump the global finger via a CAS; 4098 // note that we need to do the global finger bump 4099 // _before_ taking the intersection below, because 4100 // the task corresponding to that region will be 4101 // deemed done even if the used_region() expands 4102 // because of allocation -- as it almost certainly will 4103 // during start-up while the threads yield in the 4104 // closure below. 4105 HeapWord* finger = span.end(); 4106 bump_global_finger(finger); // atomically 4107 // There are null tasks here corresponding to chunks 4108 // beyond the "top" address of the space. 4109 span = span.intersection(sp->used_region()); 4110 if (!span.is_empty()) { // Non-null task 4111 HeapWord* prev_obj; 4112 assert(!span.contains(_restart_addr) || nth_task == 0, 4113 "Inconsistency"); 4114 if (nth_task == 0) { 4115 // For the 0th task, we'll not need to compute a block_start. 4116 if (span.contains(_restart_addr)) { 4117 // In the case of a restart because of stack overflow, 4118 // we might additionally skip a chunk prefix. 4119 prev_obj = _restart_addr; 4120 } else { 4121 prev_obj = span.start(); 4122 } 4123 } else { 4124 // We want to skip the first object because 4125 // the protocol is to scan any object in its entirety 4126 // that _starts_ in this span; a fortiori, any 4127 // object starting in an earlier span is scanned 4128 // as part of an earlier claimed task. 4129 // Below we use the "careful" version of block_start 4130 // so we do not try to navigate uninitialized objects. 4131 prev_obj = sp->block_start_careful(span.start()); 4132 // Below we use a variant of block_size that uses the 4133 // Printezis bits to avoid waiting for allocated 4134 // objects to become initialized/parsable. 4135 while (prev_obj < span.start()) { 4136 size_t sz = sp->block_size_no_stall(prev_obj, _collector); 4137 if (sz > 0) { 4138 prev_obj += sz; 4139 } else { 4140 // In this case we may end up doing a bit of redundant 4141 // scanning, but that appears unavoidable, short of 4142 // locking the free list locks; see bug 6324141. 4143 break; 4144 } 4145 } 4146 } 4147 if (prev_obj < span.end()) { 4148 MemRegion my_span = MemRegion(prev_obj, span.end()); 4149 // Do the marking work within a non-empty span -- 4150 // the last argument to the constructor indicates whether the 4151 // iteration should be incremental with periodic yields. 4152 Par_MarkFromRootsClosure cl(this, _collector, my_span, 4153 &_collector->_markBitMap, 4154 work_queue(i), 4155 &_collector->_markStack, 4156 _asynch); 4157 _collector->_markBitMap.iterate(&cl, my_span.start(), my_span.end()); 4158 } // else nothing to do for this task 4159 } // else nothing to do for this task 4160 } 4161 // We'd be tempted to assert here that since there are no 4162 // more tasks left to claim in this space, the global_finger 4163 // must exceed space->top() and a fortiori space->end(). However, 4164 // that would not quite be correct because the bumping of 4165 // global_finger occurs strictly after the claiming of a task, 4166 // so by the time we reach here the global finger may not yet 4167 // have been bumped up by the thread that claimed the last 4168 // task. 4169 pst->all_tasks_completed(); 4170 } 4171 4172 class Par_ConcMarkingClosure: public CMSOopClosure { 4173 private: 4174 CMSCollector* _collector; 4175 CMSConcMarkingTask* _task; 4176 MemRegion _span; 4177 CMSBitMap* _bit_map; 4178 CMSMarkStack* _overflow_stack; 4179 OopTaskQueue* _work_queue; 4180 protected: 4181 DO_OOP_WORK_DEFN 4182 public: 4183 Par_ConcMarkingClosure(CMSCollector* collector, CMSConcMarkingTask* task, OopTaskQueue* work_queue, 4184 CMSBitMap* bit_map, CMSMarkStack* overflow_stack): 4185 CMSOopClosure(collector->ref_processor()), 4186 _collector(collector), 4187 _task(task), 4188 _span(collector->_span), 4189 _work_queue(work_queue), 4190 _bit_map(bit_map), 4191 _overflow_stack(overflow_stack) 4192 { } 4193 virtual void do_oop(oop* p); 4194 virtual void do_oop(narrowOop* p); 4195 4196 void trim_queue(size_t max); 4197 void handle_stack_overflow(HeapWord* lost); 4198 void do_yield_check() { 4199 if (_task->should_yield()) { 4200 _task->yield(); 4201 } 4202 } 4203 }; 4204 4205 // Grey object scanning during work stealing phase -- 4206 // the salient assumption here is that any references 4207 // that are in these stolen objects being scanned must 4208 // already have been initialized (else they would not have 4209 // been published), so we do not need to check for 4210 // uninitialized objects before pushing here. 4211 void Par_ConcMarkingClosure::do_oop(oop obj) { 4212 assert(obj->is_oop_or_null(true), "expected an oop or NULL"); 4213 HeapWord* addr = (HeapWord*)obj; 4214 // Check if oop points into the CMS generation 4215 // and is not marked 4216 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 4217 // a white object ... 4218 // If we manage to "claim" the object, by being the 4219 // first thread to mark it, then we push it on our 4220 // marking stack 4221 if (_bit_map->par_mark(addr)) { // ... now grey 4222 // push on work queue (grey set) 4223 bool simulate_overflow = false; 4224 NOT_PRODUCT( 4225 if (CMSMarkStackOverflowALot && 4226 _collector->simulate_overflow()) { 4227 // simulate a stack overflow 4228 simulate_overflow = true; 4229 } 4230 ) 4231 if (simulate_overflow || 4232 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) { 4233 // stack overflow 4234 if (PrintCMSStatistics != 0) { 4235 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 4236 SIZE_FORMAT, _overflow_stack->capacity()); 4237 } 4238 // We cannot assert that the overflow stack is full because 4239 // it may have been emptied since. 4240 assert(simulate_overflow || 4241 _work_queue->size() == _work_queue->max_elems(), 4242 "Else push should have succeeded"); 4243 handle_stack_overflow(addr); 4244 } 4245 } // Else, some other thread got there first 4246 do_yield_check(); 4247 } 4248 } 4249 4250 void Par_ConcMarkingClosure::do_oop(oop* p) { Par_ConcMarkingClosure::do_oop_work(p); } 4251 void Par_ConcMarkingClosure::do_oop(narrowOop* p) { Par_ConcMarkingClosure::do_oop_work(p); } 4252 4253 void Par_ConcMarkingClosure::trim_queue(size_t max) { 4254 while (_work_queue->size() > max) { 4255 oop new_oop; 4256 if (_work_queue->pop_local(new_oop)) { 4257 assert(new_oop->is_oop(), "Should be an oop"); 4258 assert(_bit_map->isMarked((HeapWord*)new_oop), "Grey object"); 4259 assert(_span.contains((HeapWord*)new_oop), "Not in span"); 4260 new_oop->oop_iterate(this); // do_oop() above 4261 do_yield_check(); 4262 } 4263 } 4264 } 4265 4266 // Upon stack overflow, we discard (part of) the stack, 4267 // remembering the least address amongst those discarded 4268 // in CMSCollector's _restart_address. 4269 void Par_ConcMarkingClosure::handle_stack_overflow(HeapWord* lost) { 4270 // We need to do this under a mutex to prevent other 4271 // workers from interfering with the work done below. 4272 MutexLockerEx ml(_overflow_stack->par_lock(), 4273 Mutex::_no_safepoint_check_flag); 4274 // Remember the least grey address discarded 4275 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 4276 _collector->lower_restart_addr(ra); 4277 _overflow_stack->reset(); // discard stack contents 4278 _overflow_stack->expand(); // expand the stack if possible 4279 } 4280 4281 4282 void CMSConcMarkingTask::do_work_steal(int i) { 4283 OopTaskQueue* work_q = work_queue(i); 4284 oop obj_to_scan; 4285 CMSBitMap* bm = &(_collector->_markBitMap); 4286 CMSMarkStack* ovflw = &(_collector->_markStack); 4287 int* seed = _collector->hash_seed(i); 4288 Par_ConcMarkingClosure cl(_collector, this, work_q, bm, ovflw); 4289 while (true) { 4290 cl.trim_queue(0); 4291 assert(work_q->size() == 0, "Should have been emptied above"); 4292 if (get_work_from_overflow_stack(ovflw, work_q)) { 4293 // Can't assert below because the work obtained from the 4294 // overflow stack may already have been stolen from us. 4295 // assert(work_q->size() > 0, "Work from overflow stack"); 4296 continue; 4297 } else if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 4298 assert(obj_to_scan->is_oop(), "Should be an oop"); 4299 assert(bm->isMarked((HeapWord*)obj_to_scan), "Grey object"); 4300 obj_to_scan->oop_iterate(&cl); 4301 } else if (terminator()->offer_termination(&_term_term)) { 4302 assert(work_q->size() == 0, "Impossible!"); 4303 break; 4304 } else if (yielding() || should_yield()) { 4305 yield(); 4306 } 4307 } 4308 } 4309 4310 // This is run by the CMS (coordinator) thread. 4311 void CMSConcMarkingTask::coordinator_yield() { 4312 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 4313 "CMS thread should hold CMS token"); 4314 // First give up the locks, then yield, then re-lock 4315 // We should probably use a constructor/destructor idiom to 4316 // do this unlock/lock or modify the MutexUnlocker class to 4317 // serve our purpose. XXX 4318 assert_lock_strong(_bit_map_lock); 4319 _bit_map_lock->unlock(); 4320 ConcurrentMarkSweepThread::desynchronize(true); 4321 ConcurrentMarkSweepThread::acknowledge_yield_request(); 4322 _collector->stopTimer(); 4323 if (PrintCMSStatistics != 0) { 4324 _collector->incrementYields(); 4325 } 4326 _collector->icms_wait(); 4327 4328 // It is possible for whichever thread initiated the yield request 4329 // not to get a chance to wake up and take the bitmap lock between 4330 // this thread releasing it and reacquiring it. So, while the 4331 // should_yield() flag is on, let's sleep for a bit to give the 4332 // other thread a chance to wake up. The limit imposed on the number 4333 // of iterations is defensive, to avoid any unforseen circumstances 4334 // putting us into an infinite loop. Since it's always been this 4335 // (coordinator_yield()) method that was observed to cause the 4336 // problem, we are using a parameter (CMSCoordinatorYieldSleepCount) 4337 // which is by default non-zero. For the other seven methods that 4338 // also perform the yield operation, as are using a different 4339 // parameter (CMSYieldSleepCount) which is by default zero. This way we 4340 // can enable the sleeping for those methods too, if necessary. 4341 // See 6442774. 4342 // 4343 // We really need to reconsider the synchronization between the GC 4344 // thread and the yield-requesting threads in the future and we 4345 // should really use wait/notify, which is the recommended 4346 // way of doing this type of interaction. Additionally, we should 4347 // consolidate the eight methods that do the yield operation and they 4348 // are almost identical into one for better maintenability and 4349 // readability. See 6445193. 4350 // 4351 // Tony 2006.06.29 4352 for (unsigned i = 0; i < CMSCoordinatorYieldSleepCount && 4353 ConcurrentMarkSweepThread::should_yield() && 4354 !CMSCollector::foregroundGCIsActive(); ++i) { 4355 os::sleep(Thread::current(), 1, false); 4356 ConcurrentMarkSweepThread::acknowledge_yield_request(); 4357 } 4358 4359 ConcurrentMarkSweepThread::synchronize(true); 4360 _bit_map_lock->lock_without_safepoint_check(); 4361 _collector->startTimer(); 4362 } 4363 4364 bool CMSCollector::do_marking_mt(bool asynch) { 4365 assert(ConcGCThreads > 0 && conc_workers() != NULL, "precondition"); 4366 int num_workers = AdaptiveSizePolicy::calc_active_conc_workers( 4367 conc_workers()->total_workers(), 4368 conc_workers()->active_workers(), 4369 Threads::number_of_non_daemon_threads()); 4370 conc_workers()->set_active_workers(num_workers); 4371 4372 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 4373 4374 CMSConcMarkingTask tsk(this, 4375 cms_space, 4376 asynch, 4377 conc_workers(), 4378 task_queues()); 4379 4380 // Since the actual number of workers we get may be different 4381 // from the number we requested above, do we need to do anything different 4382 // below? In particular, may be we need to subclass the SequantialSubTasksDone 4383 // class?? XXX 4384 cms_space ->initialize_sequential_subtasks_for_marking(num_workers); 4385 4386 // Refs discovery is already non-atomic. 4387 assert(!ref_processor()->discovery_is_atomic(), "Should be non-atomic"); 4388 assert(ref_processor()->discovery_is_mt(), "Discovery should be MT"); 4389 conc_workers()->start_task(&tsk); 4390 while (tsk.yielded()) { 4391 tsk.coordinator_yield(); 4392 conc_workers()->continue_task(&tsk); 4393 } 4394 // If the task was aborted, _restart_addr will be non-NULL 4395 assert(tsk.completed() || _restart_addr != NULL, "Inconsistency"); 4396 while (_restart_addr != NULL) { 4397 // XXX For now we do not make use of ABORTED state and have not 4398 // yet implemented the right abort semantics (even in the original 4399 // single-threaded CMS case). That needs some more investigation 4400 // and is deferred for now; see CR# TBF. 07252005YSR. XXX 4401 assert(!CMSAbortSemantics || tsk.aborted(), "Inconsistency"); 4402 // If _restart_addr is non-NULL, a marking stack overflow 4403 // occurred; we need to do a fresh marking iteration from the 4404 // indicated restart address. 4405 if (_foregroundGCIsActive && asynch) { 4406 // We may be running into repeated stack overflows, having 4407 // reached the limit of the stack size, while making very 4408 // slow forward progress. It may be best to bail out and 4409 // let the foreground collector do its job. 4410 // Clear _restart_addr, so that foreground GC 4411 // works from scratch. This avoids the headache of 4412 // a "rescan" which would otherwise be needed because 4413 // of the dirty mod union table & card table. 4414 _restart_addr = NULL; 4415 return false; 4416 } 4417 // Adjust the task to restart from _restart_addr 4418 tsk.reset(_restart_addr); 4419 cms_space ->initialize_sequential_subtasks_for_marking(num_workers, 4420 _restart_addr); 4421 _restart_addr = NULL; 4422 // Get the workers going again 4423 conc_workers()->start_task(&tsk); 4424 while (tsk.yielded()) { 4425 tsk.coordinator_yield(); 4426 conc_workers()->continue_task(&tsk); 4427 } 4428 } 4429 assert(tsk.completed(), "Inconsistency"); 4430 assert(tsk.result() == true, "Inconsistency"); 4431 return true; 4432 } 4433 4434 bool CMSCollector::do_marking_st(bool asynch) { 4435 ResourceMark rm; 4436 HandleMark hm; 4437 4438 // Temporarily make refs discovery single threaded (non-MT) 4439 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false); 4440 MarkFromRootsClosure markFromRootsClosure(this, _span, &_markBitMap, 4441 &_markStack, CMSYield && asynch); 4442 // the last argument to iterate indicates whether the iteration 4443 // should be incremental with periodic yields. 4444 _markBitMap.iterate(&markFromRootsClosure); 4445 // If _restart_addr is non-NULL, a marking stack overflow 4446 // occurred; we need to do a fresh iteration from the 4447 // indicated restart address. 4448 while (_restart_addr != NULL) { 4449 if (_foregroundGCIsActive && asynch) { 4450 // We may be running into repeated stack overflows, having 4451 // reached the limit of the stack size, while making very 4452 // slow forward progress. It may be best to bail out and 4453 // let the foreground collector do its job. 4454 // Clear _restart_addr, so that foreground GC 4455 // works from scratch. This avoids the headache of 4456 // a "rescan" which would otherwise be needed because 4457 // of the dirty mod union table & card table. 4458 _restart_addr = NULL; 4459 return false; // indicating failure to complete marking 4460 } 4461 // Deal with stack overflow: 4462 // we restart marking from _restart_addr 4463 HeapWord* ra = _restart_addr; 4464 markFromRootsClosure.reset(ra); 4465 _restart_addr = NULL; 4466 _markBitMap.iterate(&markFromRootsClosure, ra, _span.end()); 4467 } 4468 return true; 4469 } 4470 4471 void CMSCollector::preclean() { 4472 check_correct_thread_executing(); 4473 assert(Thread::current()->is_ConcurrentGC_thread(), "Wrong thread"); 4474 verify_work_stacks_empty(); 4475 verify_overflow_empty(); 4476 _abort_preclean = false; 4477 if (CMSPrecleaningEnabled) { 4478 if (!CMSEdenChunksRecordAlways) { 4479 _eden_chunk_index = 0; 4480 } 4481 size_t used = get_eden_used(); 4482 size_t capacity = get_eden_capacity(); 4483 // Don't start sampling unless we will get sufficiently 4484 // many samples. 4485 if (used < (capacity/(CMSScheduleRemarkSamplingRatio * 100) 4486 * CMSScheduleRemarkEdenPenetration)) { 4487 _start_sampling = true; 4488 } else { 4489 _start_sampling = false; 4490 } 4491 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 4492 CMSPhaseAccounting pa(this, "preclean", !PrintGCDetails); 4493 preclean_work(CMSPrecleanRefLists1, CMSPrecleanSurvivors1); 4494 } 4495 CMSTokenSync x(true); // is cms thread 4496 if (CMSPrecleaningEnabled) { 4497 sample_eden(); 4498 _collectorState = AbortablePreclean; 4499 } else { 4500 _collectorState = FinalMarking; 4501 } 4502 verify_work_stacks_empty(); 4503 verify_overflow_empty(); 4504 } 4505 4506 // Try and schedule the remark such that young gen 4507 // occupancy is CMSScheduleRemarkEdenPenetration %. 4508 void CMSCollector::abortable_preclean() { 4509 check_correct_thread_executing(); 4510 assert(CMSPrecleaningEnabled, "Inconsistent control state"); 4511 assert(_collectorState == AbortablePreclean, "Inconsistent control state"); 4512 4513 // If Eden's current occupancy is below this threshold, 4514 // immediately schedule the remark; else preclean 4515 // past the next scavenge in an effort to 4516 // schedule the pause as described avove. By choosing 4517 // CMSScheduleRemarkEdenSizeThreshold >= max eden size 4518 // we will never do an actual abortable preclean cycle. 4519 if (get_eden_used() > CMSScheduleRemarkEdenSizeThreshold) { 4520 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 4521 CMSPhaseAccounting pa(this, "abortable-preclean", !PrintGCDetails); 4522 // We need more smarts in the abortable preclean 4523 // loop below to deal with cases where allocation 4524 // in young gen is very very slow, and our precleaning 4525 // is running a losing race against a horde of 4526 // mutators intent on flooding us with CMS updates 4527 // (dirty cards). 4528 // One, admittedly dumb, strategy is to give up 4529 // after a certain number of abortable precleaning loops 4530 // or after a certain maximum time. We want to make 4531 // this smarter in the next iteration. 4532 // XXX FIX ME!!! YSR 4533 size_t loops = 0, workdone = 0, cumworkdone = 0, waited = 0; 4534 while (!(should_abort_preclean() || 4535 ConcurrentMarkSweepThread::should_terminate())) { 4536 workdone = preclean_work(CMSPrecleanRefLists2, CMSPrecleanSurvivors2); 4537 cumworkdone += workdone; 4538 loops++; 4539 // Voluntarily terminate abortable preclean phase if we have 4540 // been at it for too long. 4541 if ((CMSMaxAbortablePrecleanLoops != 0) && 4542 loops >= CMSMaxAbortablePrecleanLoops) { 4543 if (PrintGCDetails) { 4544 gclog_or_tty->print(" CMS: abort preclean due to loops "); 4545 } 4546 break; 4547 } 4548 if (pa.wallclock_millis() > CMSMaxAbortablePrecleanTime) { 4549 if (PrintGCDetails) { 4550 gclog_or_tty->print(" CMS: abort preclean due to time "); 4551 } 4552 break; 4553 } 4554 // If we are doing little work each iteration, we should 4555 // take a short break. 4556 if (workdone < CMSAbortablePrecleanMinWorkPerIteration) { 4557 // Sleep for some time, waiting for work to accumulate 4558 stopTimer(); 4559 cmsThread()->wait_on_cms_lock(CMSAbortablePrecleanWaitMillis); 4560 startTimer(); 4561 waited++; 4562 } 4563 } 4564 if (PrintCMSStatistics > 0) { 4565 gclog_or_tty->print(" [%d iterations, %d waits, %d cards)] ", 4566 loops, waited, cumworkdone); 4567 } 4568 } 4569 CMSTokenSync x(true); // is cms thread 4570 if (_collectorState != Idling) { 4571 assert(_collectorState == AbortablePreclean, 4572 "Spontaneous state transition?"); 4573 _collectorState = FinalMarking; 4574 } // Else, a foreground collection completed this CMS cycle. 4575 return; 4576 } 4577 4578 // Respond to an Eden sampling opportunity 4579 void CMSCollector::sample_eden() { 4580 // Make sure a young gc cannot sneak in between our 4581 // reading and recording of a sample. 4582 assert(Thread::current()->is_ConcurrentGC_thread(), 4583 "Only the cms thread may collect Eden samples"); 4584 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 4585 "Should collect samples while holding CMS token"); 4586 if (!_start_sampling) { 4587 return; 4588 } 4589 // When CMSEdenChunksRecordAlways is true, the eden chunk array 4590 // is populated by the young generation. 4591 if (_eden_chunk_array != NULL && !CMSEdenChunksRecordAlways) { 4592 if (_eden_chunk_index < _eden_chunk_capacity) { 4593 _eden_chunk_array[_eden_chunk_index] = *_top_addr; // take sample 4594 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr, 4595 "Unexpected state of Eden"); 4596 // We'd like to check that what we just sampled is an oop-start address; 4597 // however, we cannot do that here since the object may not yet have been 4598 // initialized. So we'll instead do the check when we _use_ this sample 4599 // later. 4600 if (_eden_chunk_index == 0 || 4601 (pointer_delta(_eden_chunk_array[_eden_chunk_index], 4602 _eden_chunk_array[_eden_chunk_index-1]) 4603 >= CMSSamplingGrain)) { 4604 _eden_chunk_index++; // commit sample 4605 } 4606 } 4607 } 4608 if ((_collectorState == AbortablePreclean) && !_abort_preclean) { 4609 size_t used = get_eden_used(); 4610 size_t capacity = get_eden_capacity(); 4611 assert(used <= capacity, "Unexpected state of Eden"); 4612 if (used > (capacity/100 * CMSScheduleRemarkEdenPenetration)) { 4613 _abort_preclean = true; 4614 } 4615 } 4616 } 4617 4618 4619 size_t CMSCollector::preclean_work(bool clean_refs, bool clean_survivor) { 4620 assert(_collectorState == Precleaning || 4621 _collectorState == AbortablePreclean, "incorrect state"); 4622 ResourceMark rm; 4623 HandleMark hm; 4624 4625 // Precleaning is currently not MT but the reference processor 4626 // may be set for MT. Disable it temporarily here. 4627 ReferenceProcessor* rp = ref_processor(); 4628 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(rp, false); 4629 4630 // Do one pass of scrubbing the discovered reference lists 4631 // to remove any reference objects with strongly-reachable 4632 // referents. 4633 if (clean_refs) { 4634 CMSPrecleanRefsYieldClosure yield_cl(this); 4635 assert(rp->span().equals(_span), "Spans should be equal"); 4636 CMSKeepAliveClosure keep_alive(this, _span, &_markBitMap, 4637 &_markStack, true /* preclean */); 4638 CMSDrainMarkingStackClosure complete_trace(this, 4639 _span, &_markBitMap, &_markStack, 4640 &keep_alive, true /* preclean */); 4641 4642 // We don't want this step to interfere with a young 4643 // collection because we don't want to take CPU 4644 // or memory bandwidth away from the young GC threads 4645 // (which may be as many as there are CPUs). 4646 // Note that we don't need to protect ourselves from 4647 // interference with mutators because they can't 4648 // manipulate the discovered reference lists nor affect 4649 // the computed reachability of the referents, the 4650 // only properties manipulated by the precleaning 4651 // of these reference lists. 4652 stopTimer(); 4653 CMSTokenSyncWithLocks x(true /* is cms thread */, 4654 bitMapLock()); 4655 startTimer(); 4656 sample_eden(); 4657 4658 // The following will yield to allow foreground 4659 // collection to proceed promptly. XXX YSR: 4660 // The code in this method may need further 4661 // tweaking for better performance and some restructuring 4662 // for cleaner interfaces. 4663 GCTimer *gc_timer = NULL; // Currently not tracing concurrent phases 4664 rp->preclean_discovered_references( 4665 rp->is_alive_non_header(), &keep_alive, &complete_trace, &yield_cl, 4666 gc_timer); 4667 } 4668 4669 if (clean_survivor) { // preclean the active survivor space(s) 4670 assert(_young_gen->kind() == Generation::DefNew || 4671 _young_gen->kind() == Generation::ParNew || 4672 _young_gen->kind() == Generation::ASParNew, 4673 "incorrect type for cast"); 4674 DefNewGeneration* dng = (DefNewGeneration*)_young_gen; 4675 PushAndMarkClosure pam_cl(this, _span, ref_processor(), 4676 &_markBitMap, &_modUnionTable, 4677 &_markStack, true /* precleaning phase */); 4678 stopTimer(); 4679 CMSTokenSyncWithLocks ts(true /* is cms thread */, 4680 bitMapLock()); 4681 startTimer(); 4682 unsigned int before_count = 4683 GenCollectedHeap::heap()->total_collections(); 4684 SurvivorSpacePrecleanClosure 4685 sss_cl(this, _span, &_markBitMap, &_markStack, 4686 &pam_cl, before_count, CMSYield); 4687 dng->from()->object_iterate_careful(&sss_cl); 4688 dng->to()->object_iterate_careful(&sss_cl); 4689 } 4690 MarkRefsIntoAndScanClosure 4691 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable, 4692 &_markStack, this, CMSYield, 4693 true /* precleaning phase */); 4694 // CAUTION: The following closure has persistent state that may need to 4695 // be reset upon a decrease in the sequence of addresses it 4696 // processes. 4697 ScanMarkedObjectsAgainCarefullyClosure 4698 smoac_cl(this, _span, 4699 &_markBitMap, &_markStack, &mrias_cl, CMSYield); 4700 4701 // Preclean dirty cards in ModUnionTable and CardTable using 4702 // appropriate convergence criterion; 4703 // repeat CMSPrecleanIter times unless we find that 4704 // we are losing. 4705 assert(CMSPrecleanIter < 10, "CMSPrecleanIter is too large"); 4706 assert(CMSPrecleanNumerator < CMSPrecleanDenominator, 4707 "Bad convergence multiplier"); 4708 assert(CMSPrecleanThreshold >= 100, 4709 "Unreasonably low CMSPrecleanThreshold"); 4710 4711 size_t numIter, cumNumCards, lastNumCards, curNumCards; 4712 for (numIter = 0, cumNumCards = lastNumCards = curNumCards = 0; 4713 numIter < CMSPrecleanIter; 4714 numIter++, lastNumCards = curNumCards, cumNumCards += curNumCards) { 4715 curNumCards = preclean_mod_union_table(_cmsGen, &smoac_cl); 4716 if (Verbose && PrintGCDetails) { 4717 gclog_or_tty->print(" (modUnionTable: %d cards)", curNumCards); 4718 } 4719 // Either there are very few dirty cards, so re-mark 4720 // pause will be small anyway, or our pre-cleaning isn't 4721 // that much faster than the rate at which cards are being 4722 // dirtied, so we might as well stop and re-mark since 4723 // precleaning won't improve our re-mark time by much. 4724 if (curNumCards <= CMSPrecleanThreshold || 4725 (numIter > 0 && 4726 (curNumCards * CMSPrecleanDenominator > 4727 lastNumCards * CMSPrecleanNumerator))) { 4728 numIter++; 4729 cumNumCards += curNumCards; 4730 break; 4731 } 4732 } 4733 4734 preclean_klasses(&mrias_cl, _cmsGen->freelistLock()); 4735 4736 curNumCards = preclean_card_table(_cmsGen, &smoac_cl); 4737 cumNumCards += curNumCards; 4738 if (PrintGCDetails && PrintCMSStatistics != 0) { 4739 gclog_or_tty->print_cr(" (cardTable: %d cards, re-scanned %d cards, %d iterations)", 4740 curNumCards, cumNumCards, numIter); 4741 } 4742 return cumNumCards; // as a measure of useful work done 4743 } 4744 4745 // PRECLEANING NOTES: 4746 // Precleaning involves: 4747 // . reading the bits of the modUnionTable and clearing the set bits. 4748 // . For the cards corresponding to the set bits, we scan the 4749 // objects on those cards. This means we need the free_list_lock 4750 // so that we can safely iterate over the CMS space when scanning 4751 // for oops. 4752 // . When we scan the objects, we'll be both reading and setting 4753 // marks in the marking bit map, so we'll need the marking bit map. 4754 // . For protecting _collector_state transitions, we take the CGC_lock. 4755 // Note that any races in the reading of of card table entries by the 4756 // CMS thread on the one hand and the clearing of those entries by the 4757 // VM thread or the setting of those entries by the mutator threads on the 4758 // other are quite benign. However, for efficiency it makes sense to keep 4759 // the VM thread from racing with the CMS thread while the latter is 4760 // dirty card info to the modUnionTable. We therefore also use the 4761 // CGC_lock to protect the reading of the card table and the mod union 4762 // table by the CM thread. 4763 // . We run concurrently with mutator updates, so scanning 4764 // needs to be done carefully -- we should not try to scan 4765 // potentially uninitialized objects. 4766 // 4767 // Locking strategy: While holding the CGC_lock, we scan over and 4768 // reset a maximal dirty range of the mod union / card tables, then lock 4769 // the free_list_lock and bitmap lock to do a full marking, then 4770 // release these locks; and repeat the cycle. This allows for a 4771 // certain amount of fairness in the sharing of these locks between 4772 // the CMS collector on the one hand, and the VM thread and the 4773 // mutators on the other. 4774 4775 // NOTE: preclean_mod_union_table() and preclean_card_table() 4776 // further below are largely identical; if you need to modify 4777 // one of these methods, please check the other method too. 4778 4779 size_t CMSCollector::preclean_mod_union_table( 4780 ConcurrentMarkSweepGeneration* gen, 4781 ScanMarkedObjectsAgainCarefullyClosure* cl) { 4782 verify_work_stacks_empty(); 4783 verify_overflow_empty(); 4784 4785 // strategy: starting with the first card, accumulate contiguous 4786 // ranges of dirty cards; clear these cards, then scan the region 4787 // covered by these cards. 4788 4789 // Since all of the MUT is committed ahead, we can just use 4790 // that, in case the generations expand while we are precleaning. 4791 // It might also be fine to just use the committed part of the 4792 // generation, but we might potentially miss cards when the 4793 // generation is rapidly expanding while we are in the midst 4794 // of precleaning. 4795 HeapWord* startAddr = gen->reserved().start(); 4796 HeapWord* endAddr = gen->reserved().end(); 4797 4798 cl->setFreelistLock(gen->freelistLock()); // needed for yielding 4799 4800 size_t numDirtyCards, cumNumDirtyCards; 4801 HeapWord *nextAddr, *lastAddr; 4802 for (cumNumDirtyCards = numDirtyCards = 0, 4803 nextAddr = lastAddr = startAddr; 4804 nextAddr < endAddr; 4805 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 4806 4807 ResourceMark rm; 4808 HandleMark hm; 4809 4810 MemRegion dirtyRegion; 4811 { 4812 stopTimer(); 4813 // Potential yield point 4814 CMSTokenSync ts(true); 4815 startTimer(); 4816 sample_eden(); 4817 // Get dirty region starting at nextOffset (inclusive), 4818 // simultaneously clearing it. 4819 dirtyRegion = 4820 _modUnionTable.getAndClearMarkedRegion(nextAddr, endAddr); 4821 assert(dirtyRegion.start() >= nextAddr, 4822 "returned region inconsistent?"); 4823 } 4824 // Remember where the next search should begin. 4825 // The returned region (if non-empty) is a right open interval, 4826 // so lastOffset is obtained from the right end of that 4827 // interval. 4828 lastAddr = dirtyRegion.end(); 4829 // Should do something more transparent and less hacky XXX 4830 numDirtyCards = 4831 _modUnionTable.heapWordDiffToOffsetDiff(dirtyRegion.word_size()); 4832 4833 // We'll scan the cards in the dirty region (with periodic 4834 // yields for foreground GC as needed). 4835 if (!dirtyRegion.is_empty()) { 4836 assert(numDirtyCards > 0, "consistency check"); 4837 HeapWord* stop_point = NULL; 4838 stopTimer(); 4839 // Potential yield point 4840 CMSTokenSyncWithLocks ts(true, gen->freelistLock(), 4841 bitMapLock()); 4842 startTimer(); 4843 { 4844 verify_work_stacks_empty(); 4845 verify_overflow_empty(); 4846 sample_eden(); 4847 stop_point = 4848 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 4849 } 4850 if (stop_point != NULL) { 4851 // The careful iteration stopped early either because it found an 4852 // uninitialized object, or because we were in the midst of an 4853 // "abortable preclean", which should now be aborted. Redirty 4854 // the bits corresponding to the partially-scanned or unscanned 4855 // cards. We'll either restart at the next block boundary or 4856 // abort the preclean. 4857 assert((_collectorState == AbortablePreclean && should_abort_preclean()), 4858 "Should only be AbortablePreclean."); 4859 _modUnionTable.mark_range(MemRegion(stop_point, dirtyRegion.end())); 4860 if (should_abort_preclean()) { 4861 break; // out of preclean loop 4862 } else { 4863 // Compute the next address at which preclean should pick up; 4864 // might need bitMapLock in order to read P-bits. 4865 lastAddr = next_card_start_after_block(stop_point); 4866 } 4867 } 4868 } else { 4869 assert(lastAddr == endAddr, "consistency check"); 4870 assert(numDirtyCards == 0, "consistency check"); 4871 break; 4872 } 4873 } 4874 verify_work_stacks_empty(); 4875 verify_overflow_empty(); 4876 return cumNumDirtyCards; 4877 } 4878 4879 // NOTE: preclean_mod_union_table() above and preclean_card_table() 4880 // below are largely identical; if you need to modify 4881 // one of these methods, please check the other method too. 4882 4883 size_t CMSCollector::preclean_card_table(ConcurrentMarkSweepGeneration* gen, 4884 ScanMarkedObjectsAgainCarefullyClosure* cl) { 4885 // strategy: it's similar to precleamModUnionTable above, in that 4886 // we accumulate contiguous ranges of dirty cards, mark these cards 4887 // precleaned, then scan the region covered by these cards. 4888 HeapWord* endAddr = (HeapWord*)(gen->_virtual_space.high()); 4889 HeapWord* startAddr = (HeapWord*)(gen->_virtual_space.low()); 4890 4891 cl->setFreelistLock(gen->freelistLock()); // needed for yielding 4892 4893 size_t numDirtyCards, cumNumDirtyCards; 4894 HeapWord *lastAddr, *nextAddr; 4895 4896 for (cumNumDirtyCards = numDirtyCards = 0, 4897 nextAddr = lastAddr = startAddr; 4898 nextAddr < endAddr; 4899 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 4900 4901 ResourceMark rm; 4902 HandleMark hm; 4903 4904 MemRegion dirtyRegion; 4905 { 4906 // See comments in "Precleaning notes" above on why we 4907 // do this locking. XXX Could the locking overheads be 4908 // too high when dirty cards are sparse? [I don't think so.] 4909 stopTimer(); 4910 CMSTokenSync x(true); // is cms thread 4911 startTimer(); 4912 sample_eden(); 4913 // Get and clear dirty region from card table 4914 dirtyRegion = _ct->ct_bs()->dirty_card_range_after_reset( 4915 MemRegion(nextAddr, endAddr), 4916 true, 4917 CardTableModRefBS::precleaned_card_val()); 4918 4919 assert(dirtyRegion.start() >= nextAddr, 4920 "returned region inconsistent?"); 4921 } 4922 lastAddr = dirtyRegion.end(); 4923 numDirtyCards = 4924 dirtyRegion.word_size()/CardTableModRefBS::card_size_in_words; 4925 4926 if (!dirtyRegion.is_empty()) { 4927 stopTimer(); 4928 CMSTokenSyncWithLocks ts(true, gen->freelistLock(), bitMapLock()); 4929 startTimer(); 4930 sample_eden(); 4931 verify_work_stacks_empty(); 4932 verify_overflow_empty(); 4933 HeapWord* stop_point = 4934 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 4935 if (stop_point != NULL) { 4936 assert((_collectorState == AbortablePreclean && should_abort_preclean()), 4937 "Should only be AbortablePreclean."); 4938 _ct->ct_bs()->invalidate(MemRegion(stop_point, dirtyRegion.end())); 4939 if (should_abort_preclean()) { 4940 break; // out of preclean loop 4941 } else { 4942 // Compute the next address at which preclean should pick up. 4943 lastAddr = next_card_start_after_block(stop_point); 4944 } 4945 } 4946 } else { 4947 break; 4948 } 4949 } 4950 verify_work_stacks_empty(); 4951 verify_overflow_empty(); 4952 return cumNumDirtyCards; 4953 } 4954 4955 class PrecleanKlassClosure : public KlassClosure { 4956 CMKlassClosure _cm_klass_closure; 4957 public: 4958 PrecleanKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {} 4959 void do_klass(Klass* k) { 4960 if (k->has_accumulated_modified_oops()) { 4961 k->clear_accumulated_modified_oops(); 4962 4963 _cm_klass_closure.do_klass(k); 4964 } 4965 } 4966 }; 4967 4968 // The freelist lock is needed to prevent asserts, is it really needed? 4969 void CMSCollector::preclean_klasses(MarkRefsIntoAndScanClosure* cl, Mutex* freelistLock) { 4970 4971 cl->set_freelistLock(freelistLock); 4972 4973 CMSTokenSyncWithLocks ts(true, freelistLock, bitMapLock()); 4974 4975 // SSS: Add equivalent to ScanMarkedObjectsAgainCarefullyClosure::do_yield_check and should_abort_preclean? 4976 // SSS: We should probably check if precleaning should be aborted, at suitable intervals? 4977 PrecleanKlassClosure preclean_klass_closure(cl); 4978 ClassLoaderDataGraph::classes_do(&preclean_klass_closure); 4979 4980 verify_work_stacks_empty(); 4981 verify_overflow_empty(); 4982 } 4983 4984 void CMSCollector::checkpointRootsFinal(bool asynch, 4985 bool clear_all_soft_refs, bool init_mark_was_synchronous) { 4986 assert(_collectorState == FinalMarking, "incorrect state transition?"); 4987 check_correct_thread_executing(); 4988 // world is stopped at this checkpoint 4989 assert(SafepointSynchronize::is_at_safepoint(), 4990 "world should be stopped"); 4991 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 4992 4993 verify_work_stacks_empty(); 4994 verify_overflow_empty(); 4995 4996 SpecializationStats::clear(); 4997 if (PrintGCDetails) { 4998 gclog_or_tty->print("[YG occupancy: "SIZE_FORMAT" K ("SIZE_FORMAT" K)]", 4999 _young_gen->used() / K, 5000 _young_gen->capacity() / K); 5001 } 5002 if (asynch) { 5003 if (CMSScavengeBeforeRemark) { 5004 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5005 // Temporarily set flag to false, GCH->do_collection will 5006 // expect it to be false and set to true 5007 FlagSetting fl(gch->_is_gc_active, false); 5008 NOT_PRODUCT(GCTraceTime t("Scavenge-Before-Remark", 5009 PrintGCDetails && Verbose, true, _gc_timer_cm);) 5010 int level = _cmsGen->level() - 1; 5011 if (level >= 0) { 5012 gch->do_collection(true, // full (i.e. force, see below) 5013 false, // !clear_all_soft_refs 5014 0, // size 5015 false, // is_tlab 5016 level // max_level 5017 ); 5018 } 5019 } 5020 FreelistLocker x(this); 5021 MutexLockerEx y(bitMapLock(), 5022 Mutex::_no_safepoint_check_flag); 5023 assert(!init_mark_was_synchronous, "but that's impossible!"); 5024 checkpointRootsFinalWork(asynch, clear_all_soft_refs, false); 5025 } else { 5026 // already have all the locks 5027 checkpointRootsFinalWork(asynch, clear_all_soft_refs, 5028 init_mark_was_synchronous); 5029 } 5030 verify_work_stacks_empty(); 5031 verify_overflow_empty(); 5032 SpecializationStats::print(); 5033 } 5034 5035 void CMSCollector::checkpointRootsFinalWork(bool asynch, 5036 bool clear_all_soft_refs, bool init_mark_was_synchronous) { 5037 5038 NOT_PRODUCT(GCTraceTime tr("checkpointRootsFinalWork", PrintGCDetails, false, _gc_timer_cm);) 5039 5040 assert(haveFreelistLocks(), "must have free list locks"); 5041 assert_lock_strong(bitMapLock()); 5042 5043 if (UseAdaptiveSizePolicy) { 5044 size_policy()->checkpoint_roots_final_begin(); 5045 } 5046 5047 ResourceMark rm; 5048 HandleMark hm; 5049 5050 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5051 5052 if (should_unload_classes()) { 5053 CodeCache::gc_prologue(); 5054 } 5055 assert(haveFreelistLocks(), "must have free list locks"); 5056 assert_lock_strong(bitMapLock()); 5057 5058 if (!init_mark_was_synchronous) { 5059 // We might assume that we need not fill TLAB's when 5060 // CMSScavengeBeforeRemark is set, because we may have just done 5061 // a scavenge which would have filled all TLAB's -- and besides 5062 // Eden would be empty. This however may not always be the case -- 5063 // for instance although we asked for a scavenge, it may not have 5064 // happened because of a JNI critical section. We probably need 5065 // a policy for deciding whether we can in that case wait until 5066 // the critical section releases and then do the remark following 5067 // the scavenge, and skip it here. In the absence of that policy, 5068 // or of an indication of whether the scavenge did indeed occur, 5069 // we cannot rely on TLAB's having been filled and must do 5070 // so here just in case a scavenge did not happen. 5071 gch->ensure_parsability(false); // fill TLAB's, but no need to retire them 5072 // Update the saved marks which may affect the root scans. 5073 gch->save_marks(); 5074 5075 if (CMSPrintEdenSurvivorChunks) { 5076 print_eden_and_survivor_chunk_arrays(); 5077 } 5078 5079 { 5080 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;) 5081 5082 // Note on the role of the mod union table: 5083 // Since the marker in "markFromRoots" marks concurrently with 5084 // mutators, it is possible for some reachable objects not to have been 5085 // scanned. For instance, an only reference to an object A was 5086 // placed in object B after the marker scanned B. Unless B is rescanned, 5087 // A would be collected. Such updates to references in marked objects 5088 // are detected via the mod union table which is the set of all cards 5089 // dirtied since the first checkpoint in this GC cycle and prior to 5090 // the most recent young generation GC, minus those cleaned up by the 5091 // concurrent precleaning. 5092 if (CMSParallelRemarkEnabled && CollectedHeap::use_parallel_gc_threads()) { 5093 GCTraceTime t("Rescan (parallel) ", PrintGCDetails, false, _gc_timer_cm); 5094 do_remark_parallel(); 5095 } else { 5096 GCTraceTime t("Rescan (non-parallel) ", PrintGCDetails, false, 5097 _gc_timer_cm); 5098 do_remark_non_parallel(); 5099 } 5100 } 5101 } else { 5102 assert(!asynch, "Can't have init_mark_was_synchronous in asynch mode"); 5103 // The initial mark was stop-world, so there's no rescanning to 5104 // do; go straight on to the next step below. 5105 } 5106 verify_work_stacks_empty(); 5107 verify_overflow_empty(); 5108 5109 { 5110 NOT_PRODUCT(GCTraceTime ts("refProcessingWork", PrintGCDetails, false, _gc_timer_cm);) 5111 refProcessingWork(asynch, clear_all_soft_refs); 5112 } 5113 verify_work_stacks_empty(); 5114 verify_overflow_empty(); 5115 5116 if (should_unload_classes()) { 5117 CodeCache::gc_epilogue(); 5118 } 5119 JvmtiExport::gc_epilogue(); 5120 5121 // If we encountered any (marking stack / work queue) overflow 5122 // events during the current CMS cycle, take appropriate 5123 // remedial measures, where possible, so as to try and avoid 5124 // recurrence of that condition. 5125 assert(_markStack.isEmpty(), "No grey objects"); 5126 size_t ser_ovflw = _ser_pmc_remark_ovflw + _ser_pmc_preclean_ovflw + 5127 _ser_kac_ovflw + _ser_kac_preclean_ovflw; 5128 if (ser_ovflw > 0) { 5129 if (PrintCMSStatistics != 0) { 5130 gclog_or_tty->print_cr("Marking stack overflow (benign) " 5131 "(pmc_pc="SIZE_FORMAT", pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT 5132 ", kac_preclean="SIZE_FORMAT")", 5133 _ser_pmc_preclean_ovflw, _ser_pmc_remark_ovflw, 5134 _ser_kac_ovflw, _ser_kac_preclean_ovflw); 5135 } 5136 _markStack.expand(); 5137 _ser_pmc_remark_ovflw = 0; 5138 _ser_pmc_preclean_ovflw = 0; 5139 _ser_kac_preclean_ovflw = 0; 5140 _ser_kac_ovflw = 0; 5141 } 5142 if (_par_pmc_remark_ovflw > 0 || _par_kac_ovflw > 0) { 5143 if (PrintCMSStatistics != 0) { 5144 gclog_or_tty->print_cr("Work queue overflow (benign) " 5145 "(pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT")", 5146 _par_pmc_remark_ovflw, _par_kac_ovflw); 5147 } 5148 _par_pmc_remark_ovflw = 0; 5149 _par_kac_ovflw = 0; 5150 } 5151 if (PrintCMSStatistics != 0) { 5152 if (_markStack._hit_limit > 0) { 5153 gclog_or_tty->print_cr(" (benign) Hit max stack size limit ("SIZE_FORMAT")", 5154 _markStack._hit_limit); 5155 } 5156 if (_markStack._failed_double > 0) { 5157 gclog_or_tty->print_cr(" (benign) Failed stack doubling ("SIZE_FORMAT")," 5158 " current capacity "SIZE_FORMAT, 5159 _markStack._failed_double, 5160 _markStack.capacity()); 5161 } 5162 } 5163 _markStack._hit_limit = 0; 5164 _markStack._failed_double = 0; 5165 5166 if ((VerifyAfterGC || VerifyDuringGC) && 5167 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 5168 verify_after_remark(); 5169 } 5170 5171 _gc_tracer_cm->report_object_count_after_gc(&_is_alive_closure); 5172 5173 // Change under the freelistLocks. 5174 _collectorState = Sweeping; 5175 // Call isAllClear() under bitMapLock 5176 assert(_modUnionTable.isAllClear(), 5177 "Should be clear by end of the final marking"); 5178 assert(_ct->klass_rem_set()->mod_union_is_clear(), 5179 "Should be clear by end of the final marking"); 5180 if (UseAdaptiveSizePolicy) { 5181 size_policy()->checkpoint_roots_final_end(gch->gc_cause()); 5182 } 5183 } 5184 5185 // Parallel remark task 5186 class CMSParRemarkTask: public AbstractGangTask { 5187 CMSCollector* _collector; 5188 int _n_workers; 5189 CompactibleFreeListSpace* _cms_space; 5190 5191 // The per-thread work queues, available here for stealing. 5192 OopTaskQueueSet* _task_queues; 5193 ParallelTaskTerminator _term; 5194 5195 public: 5196 // A value of 0 passed to n_workers will cause the number of 5197 // workers to be taken from the active workers in the work gang. 5198 CMSParRemarkTask(CMSCollector* collector, 5199 CompactibleFreeListSpace* cms_space, 5200 int n_workers, FlexibleWorkGang* workers, 5201 OopTaskQueueSet* task_queues): 5202 AbstractGangTask("Rescan roots and grey objects in parallel"), 5203 _collector(collector), 5204 _cms_space(cms_space), 5205 _n_workers(n_workers), 5206 _task_queues(task_queues), 5207 _term(n_workers, task_queues) { } 5208 5209 OopTaskQueueSet* task_queues() { return _task_queues; } 5210 5211 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 5212 5213 ParallelTaskTerminator* terminator() { return &_term; } 5214 int n_workers() { return _n_workers; } 5215 5216 void work(uint worker_id); 5217 5218 private: 5219 // Work method in support of parallel rescan ... of young gen spaces 5220 void do_young_space_rescan(int i, Par_MarkRefsIntoAndScanClosure* cl, 5221 ContiguousSpace* space, 5222 HeapWord** chunk_array, size_t chunk_top); 5223 5224 // ... of dirty cards in old space 5225 void do_dirty_card_rescan_tasks(CompactibleFreeListSpace* sp, int i, 5226 Par_MarkRefsIntoAndScanClosure* cl); 5227 5228 // ... work stealing for the above 5229 void do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, int* seed); 5230 }; 5231 5232 class RemarkKlassClosure : public KlassClosure { 5233 CMKlassClosure _cm_klass_closure; 5234 public: 5235 RemarkKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {} 5236 void do_klass(Klass* k) { 5237 // Check if we have modified any oops in the Klass during the concurrent marking. 5238 if (k->has_accumulated_modified_oops()) { 5239 k->clear_accumulated_modified_oops(); 5240 5241 // We could have transfered the current modified marks to the accumulated marks, 5242 // like we do with the Card Table to Mod Union Table. But it's not really necessary. 5243 } else if (k->has_modified_oops()) { 5244 // Don't clear anything, this info is needed by the next young collection. 5245 } else { 5246 // No modified oops in the Klass. 5247 return; 5248 } 5249 5250 // The klass has modified fields, need to scan the klass. 5251 _cm_klass_closure.do_klass(k); 5252 } 5253 }; 5254 5255 // work_queue(i) is passed to the closure 5256 // Par_MarkRefsIntoAndScanClosure. The "i" parameter 5257 // also is passed to do_dirty_card_rescan_tasks() and to 5258 // do_work_steal() to select the i-th task_queue. 5259 5260 void CMSParRemarkTask::work(uint worker_id) { 5261 elapsedTimer _timer; 5262 ResourceMark rm; 5263 HandleMark hm; 5264 5265 // ---------- rescan from roots -------------- 5266 _timer.start(); 5267 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5268 Par_MarkRefsIntoAndScanClosure par_mrias_cl(_collector, 5269 _collector->_span, _collector->ref_processor(), 5270 &(_collector->_markBitMap), 5271 work_queue(worker_id)); 5272 5273 // Rescan young gen roots first since these are likely 5274 // coarsely partitioned and may, on that account, constitute 5275 // the critical path; thus, it's best to start off that 5276 // work first. 5277 // ---------- young gen roots -------------- 5278 { 5279 DefNewGeneration* dng = _collector->_young_gen->as_DefNewGeneration(); 5280 EdenSpace* eden_space = dng->eden(); 5281 ContiguousSpace* from_space = dng->from(); 5282 ContiguousSpace* to_space = dng->to(); 5283 5284 HeapWord** eca = _collector->_eden_chunk_array; 5285 size_t ect = _collector->_eden_chunk_index; 5286 HeapWord** sca = _collector->_survivor_chunk_array; 5287 size_t sct = _collector->_survivor_chunk_index; 5288 5289 assert(ect <= _collector->_eden_chunk_capacity, "out of bounds"); 5290 assert(sct <= _collector->_survivor_chunk_capacity, "out of bounds"); 5291 5292 do_young_space_rescan(worker_id, &par_mrias_cl, to_space, NULL, 0); 5293 do_young_space_rescan(worker_id, &par_mrias_cl, from_space, sca, sct); 5294 do_young_space_rescan(worker_id, &par_mrias_cl, eden_space, eca, ect); 5295 5296 _timer.stop(); 5297 if (PrintCMSStatistics != 0) { 5298 gclog_or_tty->print_cr( 5299 "Finished young gen rescan work in %dth thread: %3.3f sec", 5300 worker_id, _timer.seconds()); 5301 } 5302 } 5303 5304 // ---------- remaining roots -------------- 5305 _timer.reset(); 5306 _timer.start(); 5307 gch->gen_process_strong_roots(_collector->_cmsGen->level(), 5308 false, // yg was scanned above 5309 false, // this is parallel code 5310 false, // not scavenging 5311 SharedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()), 5312 &par_mrias_cl, 5313 true, // walk all of code cache if (so & SO_CodeCache) 5314 NULL, 5315 NULL); // The dirty klasses will be handled below 5316 assert(_collector->should_unload_classes() 5317 || (_collector->CMSCollector::roots_scanning_options() & SharedHeap::SO_CodeCache), 5318 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 5319 _timer.stop(); 5320 if (PrintCMSStatistics != 0) { 5321 gclog_or_tty->print_cr( 5322 "Finished remaining root rescan work in %dth thread: %3.3f sec", 5323 worker_id, _timer.seconds()); 5324 } 5325 5326 // ---------- unhandled CLD scanning ---------- 5327 if (worker_id == 0) { // Single threaded at the moment. 5328 _timer.reset(); 5329 _timer.start(); 5330 5331 // Scan all new class loader data objects and new dependencies that were 5332 // introduced during concurrent marking. 5333 ResourceMark rm; 5334 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds(); 5335 for (int i = 0; i < array->length(); i++) { 5336 par_mrias_cl.do_class_loader_data(array->at(i)); 5337 } 5338 5339 // We don't need to keep track of new CLDs anymore. 5340 ClassLoaderDataGraph::remember_new_clds(false); 5341 5342 _timer.stop(); 5343 if (PrintCMSStatistics != 0) { 5344 gclog_or_tty->print_cr( 5345 "Finished unhandled CLD scanning work in %dth thread: %3.3f sec", 5346 worker_id, _timer.seconds()); 5347 } 5348 } 5349 5350 // ---------- dirty klass scanning ---------- 5351 if (worker_id == 0) { // Single threaded at the moment. 5352 _timer.reset(); 5353 _timer.start(); 5354 5355 // Scan all classes that was dirtied during the concurrent marking phase. 5356 RemarkKlassClosure remark_klass_closure(&par_mrias_cl); 5357 ClassLoaderDataGraph::classes_do(&remark_klass_closure); 5358 5359 _timer.stop(); 5360 if (PrintCMSStatistics != 0) { 5361 gclog_or_tty->print_cr( 5362 "Finished dirty klass scanning work in %dth thread: %3.3f sec", 5363 worker_id, _timer.seconds()); 5364 } 5365 } 5366 5367 // We might have added oops to ClassLoaderData::_handles during the 5368 // concurrent marking phase. These oops point to newly allocated objects 5369 // that are guaranteed to be kept alive. Either by the direct allocation 5370 // code, or when the young collector processes the strong roots. Hence, 5371 // we don't have to revisit the _handles block during the remark phase. 5372 5373 // ---------- rescan dirty cards ------------ 5374 _timer.reset(); 5375 _timer.start(); 5376 5377 // Do the rescan tasks for each of the two spaces 5378 // (cms_space) in turn. 5379 // "worker_id" is passed to select the task_queue for "worker_id" 5380 do_dirty_card_rescan_tasks(_cms_space, worker_id, &par_mrias_cl); 5381 _timer.stop(); 5382 if (PrintCMSStatistics != 0) { 5383 gclog_or_tty->print_cr( 5384 "Finished dirty card rescan work in %dth thread: %3.3f sec", 5385 worker_id, _timer.seconds()); 5386 } 5387 5388 // ---------- steal work from other threads ... 5389 // ---------- ... and drain overflow list. 5390 _timer.reset(); 5391 _timer.start(); 5392 do_work_steal(worker_id, &par_mrias_cl, _collector->hash_seed(worker_id)); 5393 _timer.stop(); 5394 if (PrintCMSStatistics != 0) { 5395 gclog_or_tty->print_cr( 5396 "Finished work stealing in %dth thread: %3.3f sec", 5397 worker_id, _timer.seconds()); 5398 } 5399 } 5400 5401 // Note that parameter "i" is not used. 5402 void 5403 CMSParRemarkTask::do_young_space_rescan(int i, 5404 Par_MarkRefsIntoAndScanClosure* cl, ContiguousSpace* space, 5405 HeapWord** chunk_array, size_t chunk_top) { 5406 // Until all tasks completed: 5407 // . claim an unclaimed task 5408 // . compute region boundaries corresponding to task claimed 5409 // using chunk_array 5410 // . par_oop_iterate(cl) over that region 5411 5412 ResourceMark rm; 5413 HandleMark hm; 5414 5415 SequentialSubTasksDone* pst = space->par_seq_tasks(); 5416 assert(pst->valid(), "Uninitialized use?"); 5417 5418 uint nth_task = 0; 5419 uint n_tasks = pst->n_tasks(); 5420 5421 HeapWord *start, *end; 5422 while (!pst->is_task_claimed(/* reference */ nth_task)) { 5423 // We claimed task # nth_task; compute its boundaries. 5424 if (chunk_top == 0) { // no samples were taken 5425 assert(nth_task == 0 && n_tasks == 1, "Can have only 1 EdenSpace task"); 5426 start = space->bottom(); 5427 end = space->top(); 5428 } else if (nth_task == 0) { 5429 start = space->bottom(); 5430 end = chunk_array[nth_task]; 5431 } else if (nth_task < (uint)chunk_top) { 5432 assert(nth_task >= 1, "Control point invariant"); 5433 start = chunk_array[nth_task - 1]; 5434 end = chunk_array[nth_task]; 5435 } else { 5436 assert(nth_task == (uint)chunk_top, "Control point invariant"); 5437 start = chunk_array[chunk_top - 1]; 5438 end = space->top(); 5439 } 5440 MemRegion mr(start, end); 5441 // Verify that mr is in space 5442 assert(mr.is_empty() || space->used_region().contains(mr), 5443 "Should be in space"); 5444 // Verify that "start" is an object boundary 5445 assert(mr.is_empty() || oop(mr.start())->is_oop(), 5446 "Should be an oop"); 5447 space->par_oop_iterate(mr, cl); 5448 } 5449 pst->all_tasks_completed(); 5450 } 5451 5452 void 5453 CMSParRemarkTask::do_dirty_card_rescan_tasks( 5454 CompactibleFreeListSpace* sp, int i, 5455 Par_MarkRefsIntoAndScanClosure* cl) { 5456 // Until all tasks completed: 5457 // . claim an unclaimed task 5458 // . compute region boundaries corresponding to task claimed 5459 // . transfer dirty bits ct->mut for that region 5460 // . apply rescanclosure to dirty mut bits for that region 5461 5462 ResourceMark rm; 5463 HandleMark hm; 5464 5465 OopTaskQueue* work_q = work_queue(i); 5466 ModUnionClosure modUnionClosure(&(_collector->_modUnionTable)); 5467 // CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! 5468 // CAUTION: This closure has state that persists across calls to 5469 // the work method dirty_range_iterate_clear() in that it has 5470 // imbedded in it a (subtype of) UpwardsObjectClosure. The 5471 // use of that state in the imbedded UpwardsObjectClosure instance 5472 // assumes that the cards are always iterated (even if in parallel 5473 // by several threads) in monotonically increasing order per each 5474 // thread. This is true of the implementation below which picks 5475 // card ranges (chunks) in monotonically increasing order globally 5476 // and, a-fortiori, in monotonically increasing order per thread 5477 // (the latter order being a subsequence of the former). 5478 // If the work code below is ever reorganized into a more chaotic 5479 // work-partitioning form than the current "sequential tasks" 5480 // paradigm, the use of that persistent state will have to be 5481 // revisited and modified appropriately. See also related 5482 // bug 4756801 work on which should examine this code to make 5483 // sure that the changes there do not run counter to the 5484 // assumptions made here and necessary for correctness and 5485 // efficiency. Note also that this code might yield inefficient 5486 // behaviour in the case of very large objects that span one or 5487 // more work chunks. Such objects would potentially be scanned 5488 // several times redundantly. Work on 4756801 should try and 5489 // address that performance anomaly if at all possible. XXX 5490 MemRegion full_span = _collector->_span; 5491 CMSBitMap* bm = &(_collector->_markBitMap); // shared 5492 MarkFromDirtyCardsClosure 5493 greyRescanClosure(_collector, full_span, // entire span of interest 5494 sp, bm, work_q, cl); 5495 5496 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 5497 assert(pst->valid(), "Uninitialized use?"); 5498 uint nth_task = 0; 5499 const int alignment = CardTableModRefBS::card_size * BitsPerWord; 5500 MemRegion span = sp->used_region(); 5501 HeapWord* start_addr = span.start(); 5502 HeapWord* end_addr = (HeapWord*)round_to((intptr_t)span.end(), 5503 alignment); 5504 const size_t chunk_size = sp->rescan_task_size(); // in HeapWord units 5505 assert((HeapWord*)round_to((intptr_t)start_addr, alignment) == 5506 start_addr, "Check alignment"); 5507 assert((size_t)round_to((intptr_t)chunk_size, alignment) == 5508 chunk_size, "Check alignment"); 5509 5510 while (!pst->is_task_claimed(/* reference */ nth_task)) { 5511 // Having claimed the nth_task, compute corresponding mem-region, 5512 // which is a-fortiori aligned correctly (i.e. at a MUT bopundary). 5513 // The alignment restriction ensures that we do not need any 5514 // synchronization with other gang-workers while setting or 5515 // clearing bits in thus chunk of the MUT. 5516 MemRegion this_span = MemRegion(start_addr + nth_task*chunk_size, 5517 start_addr + (nth_task+1)*chunk_size); 5518 // The last chunk's end might be way beyond end of the 5519 // used region. In that case pull back appropriately. 5520 if (this_span.end() > end_addr) { 5521 this_span.set_end(end_addr); 5522 assert(!this_span.is_empty(), "Program logic (calculation of n_tasks)"); 5523 } 5524 // Iterate over the dirty cards covering this chunk, marking them 5525 // precleaned, and setting the corresponding bits in the mod union 5526 // table. Since we have been careful to partition at Card and MUT-word 5527 // boundaries no synchronization is needed between parallel threads. 5528 _collector->_ct->ct_bs()->dirty_card_iterate(this_span, 5529 &modUnionClosure); 5530 5531 // Having transferred these marks into the modUnionTable, 5532 // rescan the marked objects on the dirty cards in the modUnionTable. 5533 // Even if this is at a synchronous collection, the initial marking 5534 // may have been done during an asynchronous collection so there 5535 // may be dirty bits in the mod-union table. 5536 _collector->_modUnionTable.dirty_range_iterate_clear( 5537 this_span, &greyRescanClosure); 5538 _collector->_modUnionTable.verifyNoOneBitsInRange( 5539 this_span.start(), 5540 this_span.end()); 5541 } 5542 pst->all_tasks_completed(); // declare that i am done 5543 } 5544 5545 // . see if we can share work_queues with ParNew? XXX 5546 void 5547 CMSParRemarkTask::do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, 5548 int* seed) { 5549 OopTaskQueue* work_q = work_queue(i); 5550 NOT_PRODUCT(int num_steals = 0;) 5551 oop obj_to_scan; 5552 CMSBitMap* bm = &(_collector->_markBitMap); 5553 5554 while (true) { 5555 // Completely finish any left over work from (an) earlier round(s) 5556 cl->trim_queue(0); 5557 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 5558 (size_t)ParGCDesiredObjsFromOverflowList); 5559 // Now check if there's any work in the overflow list 5560 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads, 5561 // only affects the number of attempts made to get work from the 5562 // overflow list and does not affect the number of workers. Just 5563 // pass ParallelGCThreads so this behavior is unchanged. 5564 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 5565 work_q, 5566 ParallelGCThreads)) { 5567 // found something in global overflow list; 5568 // not yet ready to go stealing work from others. 5569 // We'd like to assert(work_q->size() != 0, ...) 5570 // because we just took work from the overflow list, 5571 // but of course we can't since all of that could have 5572 // been already stolen from us. 5573 // "He giveth and He taketh away." 5574 continue; 5575 } 5576 // Verify that we have no work before we resort to stealing 5577 assert(work_q->size() == 0, "Have work, shouldn't steal"); 5578 // Try to steal from other queues that have work 5579 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 5580 NOT_PRODUCT(num_steals++;) 5581 assert(obj_to_scan->is_oop(), "Oops, not an oop!"); 5582 assert(bm->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 5583 // Do scanning work 5584 obj_to_scan->oop_iterate(cl); 5585 // Loop around, finish this work, and try to steal some more 5586 } else if (terminator()->offer_termination()) { 5587 break; // nirvana from the infinite cycle 5588 } 5589 } 5590 NOT_PRODUCT( 5591 if (PrintCMSStatistics != 0) { 5592 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals); 5593 } 5594 ) 5595 assert(work_q->size() == 0 && _collector->overflow_list_is_empty(), 5596 "Else our work is not yet done"); 5597 } 5598 5599 // Record object boundaries in _eden_chunk_array by sampling the eden 5600 // top in the slow-path eden object allocation code path and record 5601 // the boundaries, if CMSEdenChunksRecordAlways is true. If 5602 // CMSEdenChunksRecordAlways is false, we use the other asynchronous 5603 // sampling in sample_eden() that activates during the part of the 5604 // preclean phase. 5605 void CMSCollector::sample_eden_chunk() { 5606 if (CMSEdenChunksRecordAlways && _eden_chunk_array != NULL) { 5607 if (_eden_chunk_lock->try_lock()) { 5608 // Record a sample. This is the critical section. The contents 5609 // of the _eden_chunk_array have to be non-decreasing in the 5610 // address order. 5611 _eden_chunk_array[_eden_chunk_index] = *_top_addr; 5612 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr, 5613 "Unexpected state of Eden"); 5614 if (_eden_chunk_index == 0 || 5615 ((_eden_chunk_array[_eden_chunk_index] > _eden_chunk_array[_eden_chunk_index-1]) && 5616 (pointer_delta(_eden_chunk_array[_eden_chunk_index], 5617 _eden_chunk_array[_eden_chunk_index-1]) >= CMSSamplingGrain))) { 5618 _eden_chunk_index++; // commit sample 5619 } 5620 _eden_chunk_lock->unlock(); 5621 } 5622 } 5623 } 5624 5625 // Return a thread-local PLAB recording array, as appropriate. 5626 void* CMSCollector::get_data_recorder(int thr_num) { 5627 if (_survivor_plab_array != NULL && 5628 (CMSPLABRecordAlways || 5629 (_collectorState > Marking && _collectorState < FinalMarking))) { 5630 assert(thr_num < (int)ParallelGCThreads, "thr_num is out of bounds"); 5631 ChunkArray* ca = &_survivor_plab_array[thr_num]; 5632 ca->reset(); // clear it so that fresh data is recorded 5633 return (void*) ca; 5634 } else { 5635 return NULL; 5636 } 5637 } 5638 5639 // Reset all the thread-local PLAB recording arrays 5640 void CMSCollector::reset_survivor_plab_arrays() { 5641 for (uint i = 0; i < ParallelGCThreads; i++) { 5642 _survivor_plab_array[i].reset(); 5643 } 5644 } 5645 5646 // Merge the per-thread plab arrays into the global survivor chunk 5647 // array which will provide the partitioning of the survivor space 5648 // for CMS rescan. 5649 void CMSCollector::merge_survivor_plab_arrays(ContiguousSpace* surv, 5650 int no_of_gc_threads) { 5651 assert(_survivor_plab_array != NULL, "Error"); 5652 assert(_survivor_chunk_array != NULL, "Error"); 5653 assert(_collectorState == FinalMarking, "Error"); 5654 for (int j = 0; j < no_of_gc_threads; j++) { 5655 _cursor[j] = 0; 5656 } 5657 HeapWord* top = surv->top(); 5658 size_t i; 5659 for (i = 0; i < _survivor_chunk_capacity; i++) { // all sca entries 5660 HeapWord* min_val = top; // Higher than any PLAB address 5661 uint min_tid = 0; // position of min_val this round 5662 for (int j = 0; j < no_of_gc_threads; j++) { 5663 ChunkArray* cur_sca = &_survivor_plab_array[j]; 5664 if (_cursor[j] == cur_sca->end()) { 5665 continue; 5666 } 5667 assert(_cursor[j] < cur_sca->end(), "ctl pt invariant"); 5668 HeapWord* cur_val = cur_sca->nth(_cursor[j]); 5669 assert(surv->used_region().contains(cur_val), "Out of bounds value"); 5670 if (cur_val < min_val) { 5671 min_tid = j; 5672 min_val = cur_val; 5673 } else { 5674 assert(cur_val < top, "All recorded addresses should be less"); 5675 } 5676 } 5677 // At this point min_val and min_tid are respectively 5678 // the least address in _survivor_plab_array[j]->nth(_cursor[j]) 5679 // and the thread (j) that witnesses that address. 5680 // We record this address in the _survivor_chunk_array[i] 5681 // and increment _cursor[min_tid] prior to the next round i. 5682 if (min_val == top) { 5683 break; 5684 } 5685 _survivor_chunk_array[i] = min_val; 5686 _cursor[min_tid]++; 5687 } 5688 // We are all done; record the size of the _survivor_chunk_array 5689 _survivor_chunk_index = i; // exclusive: [0, i) 5690 if (PrintCMSStatistics > 0) { 5691 gclog_or_tty->print(" (Survivor:" SIZE_FORMAT "chunks) ", i); 5692 } 5693 // Verify that we used up all the recorded entries 5694 #ifdef ASSERT 5695 size_t total = 0; 5696 for (int j = 0; j < no_of_gc_threads; j++) { 5697 assert(_cursor[j] == _survivor_plab_array[j].end(), "Ctl pt invariant"); 5698 total += _cursor[j]; 5699 } 5700 assert(total == _survivor_chunk_index, "Ctl Pt Invariant"); 5701 // Check that the merged array is in sorted order 5702 if (total > 0) { 5703 for (size_t i = 0; i < total - 1; i++) { 5704 if (PrintCMSStatistics > 0) { 5705 gclog_or_tty->print(" (chunk" SIZE_FORMAT ":" INTPTR_FORMAT ") ", 5706 i, _survivor_chunk_array[i]); 5707 } 5708 assert(_survivor_chunk_array[i] < _survivor_chunk_array[i+1], 5709 "Not sorted"); 5710 } 5711 } 5712 #endif // ASSERT 5713 } 5714 5715 // Set up the space's par_seq_tasks structure for work claiming 5716 // for parallel rescan of young gen. 5717 // See ParRescanTask where this is currently used. 5718 void 5719 CMSCollector:: 5720 initialize_sequential_subtasks_for_young_gen_rescan(int n_threads) { 5721 assert(n_threads > 0, "Unexpected n_threads argument"); 5722 DefNewGeneration* dng = (DefNewGeneration*)_young_gen; 5723 5724 // Eden space 5725 { 5726 SequentialSubTasksDone* pst = dng->eden()->par_seq_tasks(); 5727 assert(!pst->valid(), "Clobbering existing data?"); 5728 // Each valid entry in [0, _eden_chunk_index) represents a task. 5729 size_t n_tasks = _eden_chunk_index + 1; 5730 assert(n_tasks == 1 || _eden_chunk_array != NULL, "Error"); 5731 // Sets the condition for completion of the subtask (how many threads 5732 // need to finish in order to be done). 5733 pst->set_n_threads(n_threads); 5734 pst->set_n_tasks((int)n_tasks); 5735 } 5736 5737 // Merge the survivor plab arrays into _survivor_chunk_array 5738 if (_survivor_plab_array != NULL) { 5739 merge_survivor_plab_arrays(dng->from(), n_threads); 5740 } else { 5741 assert(_survivor_chunk_index == 0, "Error"); 5742 } 5743 5744 // To space 5745 { 5746 SequentialSubTasksDone* pst = dng->to()->par_seq_tasks(); 5747 assert(!pst->valid(), "Clobbering existing data?"); 5748 // Sets the condition for completion of the subtask (how many threads 5749 // need to finish in order to be done). 5750 pst->set_n_threads(n_threads); 5751 pst->set_n_tasks(1); 5752 assert(pst->valid(), "Error"); 5753 } 5754 5755 // From space 5756 { 5757 SequentialSubTasksDone* pst = dng->from()->par_seq_tasks(); 5758 assert(!pst->valid(), "Clobbering existing data?"); 5759 size_t n_tasks = _survivor_chunk_index + 1; 5760 assert(n_tasks == 1 || _survivor_chunk_array != NULL, "Error"); 5761 // Sets the condition for completion of the subtask (how many threads 5762 // need to finish in order to be done). 5763 pst->set_n_threads(n_threads); 5764 pst->set_n_tasks((int)n_tasks); 5765 assert(pst->valid(), "Error"); 5766 } 5767 } 5768 5769 // Parallel version of remark 5770 void CMSCollector::do_remark_parallel() { 5771 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5772 FlexibleWorkGang* workers = gch->workers(); 5773 assert(workers != NULL, "Need parallel worker threads."); 5774 // Choose to use the number of GC workers most recently set 5775 // into "active_workers". If active_workers is not set, set it 5776 // to ParallelGCThreads. 5777 int n_workers = workers->active_workers(); 5778 if (n_workers == 0) { 5779 assert(n_workers > 0, "Should have been set during scavenge"); 5780 n_workers = ParallelGCThreads; 5781 workers->set_active_workers(n_workers); 5782 } 5783 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 5784 5785 CMSParRemarkTask tsk(this, 5786 cms_space, 5787 n_workers, workers, task_queues()); 5788 5789 // Set up for parallel process_strong_roots work. 5790 gch->set_par_threads(n_workers); 5791 // We won't be iterating over the cards in the card table updating 5792 // the younger_gen cards, so we shouldn't call the following else 5793 // the verification code as well as subsequent younger_refs_iterate 5794 // code would get confused. XXX 5795 // gch->rem_set()->prepare_for_younger_refs_iterate(true); // parallel 5796 5797 // The young gen rescan work will not be done as part of 5798 // process_strong_roots (which currently doesn't knw how to 5799 // parallelize such a scan), but rather will be broken up into 5800 // a set of parallel tasks (via the sampling that the [abortable] 5801 // preclean phase did of EdenSpace, plus the [two] tasks of 5802 // scanning the [two] survivor spaces. Further fine-grain 5803 // parallelization of the scanning of the survivor spaces 5804 // themselves, and of precleaning of the younger gen itself 5805 // is deferred to the future. 5806 initialize_sequential_subtasks_for_young_gen_rescan(n_workers); 5807 5808 // The dirty card rescan work is broken up into a "sequence" 5809 // of parallel tasks (per constituent space) that are dynamically 5810 // claimed by the parallel threads. 5811 cms_space->initialize_sequential_subtasks_for_rescan(n_workers); 5812 5813 // It turns out that even when we're using 1 thread, doing the work in a 5814 // separate thread causes wide variance in run times. We can't help this 5815 // in the multi-threaded case, but we special-case n=1 here to get 5816 // repeatable measurements of the 1-thread overhead of the parallel code. 5817 if (n_workers > 1) { 5818 // Make refs discovery MT-safe, if it isn't already: it may not 5819 // necessarily be so, since it's possible that we are doing 5820 // ST marking. 5821 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), true); 5822 GenCollectedHeap::StrongRootsScope srs(gch); 5823 workers->run_task(&tsk); 5824 } else { 5825 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false); 5826 GenCollectedHeap::StrongRootsScope srs(gch); 5827 tsk.work(0); 5828 } 5829 5830 gch->set_par_threads(0); // 0 ==> non-parallel. 5831 // restore, single-threaded for now, any preserved marks 5832 // as a result of work_q overflow 5833 restore_preserved_marks_if_any(); 5834 } 5835 5836 // Non-parallel version of remark 5837 void CMSCollector::do_remark_non_parallel() { 5838 ResourceMark rm; 5839 HandleMark hm; 5840 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5841 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false); 5842 5843 MarkRefsIntoAndScanClosure 5844 mrias_cl(_span, ref_processor(), &_markBitMap, NULL /* not precleaning */, 5845 &_markStack, this, 5846 false /* should_yield */, false /* not precleaning */); 5847 MarkFromDirtyCardsClosure 5848 markFromDirtyCardsClosure(this, _span, 5849 NULL, // space is set further below 5850 &_markBitMap, &_markStack, &mrias_cl); 5851 { 5852 GCTraceTime t("grey object rescan", PrintGCDetails, false, _gc_timer_cm); 5853 // Iterate over the dirty cards, setting the corresponding bits in the 5854 // mod union table. 5855 { 5856 ModUnionClosure modUnionClosure(&_modUnionTable); 5857 _ct->ct_bs()->dirty_card_iterate( 5858 _cmsGen->used_region(), 5859 &modUnionClosure); 5860 } 5861 // Having transferred these marks into the modUnionTable, we just need 5862 // to rescan the marked objects on the dirty cards in the modUnionTable. 5863 // The initial marking may have been done during an asynchronous 5864 // collection so there may be dirty bits in the mod-union table. 5865 const int alignment = 5866 CardTableModRefBS::card_size * BitsPerWord; 5867 { 5868 // ... First handle dirty cards in CMS gen 5869 markFromDirtyCardsClosure.set_space(_cmsGen->cmsSpace()); 5870 MemRegion ur = _cmsGen->used_region(); 5871 HeapWord* lb = ur.start(); 5872 HeapWord* ub = (HeapWord*)round_to((intptr_t)ur.end(), alignment); 5873 MemRegion cms_span(lb, ub); 5874 _modUnionTable.dirty_range_iterate_clear(cms_span, 5875 &markFromDirtyCardsClosure); 5876 verify_work_stacks_empty(); 5877 if (PrintCMSStatistics != 0) { 5878 gclog_or_tty->print(" (re-scanned "SIZE_FORMAT" dirty cards in cms gen) ", 5879 markFromDirtyCardsClosure.num_dirty_cards()); 5880 } 5881 } 5882 } 5883 if (VerifyDuringGC && 5884 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 5885 HandleMark hm; // Discard invalid handles created during verification 5886 Universe::verify(); 5887 } 5888 { 5889 GCTraceTime t("root rescan", PrintGCDetails, false, _gc_timer_cm); 5890 5891 verify_work_stacks_empty(); 5892 5893 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 5894 GenCollectedHeap::StrongRootsScope srs(gch); 5895 gch->gen_process_strong_roots(_cmsGen->level(), 5896 true, // younger gens as roots 5897 false, // use the local StrongRootsScope 5898 false, // not scavenging 5899 SharedHeap::ScanningOption(roots_scanning_options()), 5900 &mrias_cl, 5901 true, // walk code active on stacks 5902 NULL, 5903 NULL); // The dirty klasses will be handled below 5904 5905 assert(should_unload_classes() 5906 || (roots_scanning_options() & SharedHeap::SO_CodeCache), 5907 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 5908 } 5909 5910 { 5911 GCTraceTime t("visit unhandled CLDs", PrintGCDetails, false, _gc_timer_cm); 5912 5913 verify_work_stacks_empty(); 5914 5915 // Scan all class loader data objects that might have been introduced 5916 // during concurrent marking. 5917 ResourceMark rm; 5918 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds(); 5919 for (int i = 0; i < array->length(); i++) { 5920 mrias_cl.do_class_loader_data(array->at(i)); 5921 } 5922 5923 // We don't need to keep track of new CLDs anymore. 5924 ClassLoaderDataGraph::remember_new_clds(false); 5925 5926 verify_work_stacks_empty(); 5927 } 5928 5929 { 5930 GCTraceTime t("dirty klass scan", PrintGCDetails, false, _gc_timer_cm); 5931 5932 verify_work_stacks_empty(); 5933 5934 RemarkKlassClosure remark_klass_closure(&mrias_cl); 5935 ClassLoaderDataGraph::classes_do(&remark_klass_closure); 5936 5937 verify_work_stacks_empty(); 5938 } 5939 5940 // We might have added oops to ClassLoaderData::_handles during the 5941 // concurrent marking phase. These oops point to newly allocated objects 5942 // that are guaranteed to be kept alive. Either by the direct allocation 5943 // code, or when the young collector processes the strong roots. Hence, 5944 // we don't have to revisit the _handles block during the remark phase. 5945 5946 verify_work_stacks_empty(); 5947 // Restore evacuated mark words, if any, used for overflow list links 5948 if (!CMSOverflowEarlyRestoration) { 5949 restore_preserved_marks_if_any(); 5950 } 5951 verify_overflow_empty(); 5952 } 5953 5954 //////////////////////////////////////////////////////// 5955 // Parallel Reference Processing Task Proxy Class 5956 //////////////////////////////////////////////////////// 5957 class CMSRefProcTaskProxy: public AbstractGangTaskWOopQueues { 5958 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask; 5959 CMSCollector* _collector; 5960 CMSBitMap* _mark_bit_map; 5961 const MemRegion _span; 5962 ProcessTask& _task; 5963 5964 public: 5965 CMSRefProcTaskProxy(ProcessTask& task, 5966 CMSCollector* collector, 5967 const MemRegion& span, 5968 CMSBitMap* mark_bit_map, 5969 AbstractWorkGang* workers, 5970 OopTaskQueueSet* task_queues): 5971 // XXX Should superclass AGTWOQ also know about AWG since it knows 5972 // about the task_queues used by the AWG? Then it could initialize 5973 // the terminator() object. See 6984287. The set_for_termination() 5974 // below is a temporary band-aid for the regression in 6984287. 5975 AbstractGangTaskWOopQueues("Process referents by policy in parallel", 5976 task_queues), 5977 _task(task), 5978 _collector(collector), _span(span), _mark_bit_map(mark_bit_map) 5979 { 5980 assert(_collector->_span.equals(_span) && !_span.is_empty(), 5981 "Inconsistency in _span"); 5982 set_for_termination(workers->active_workers()); 5983 } 5984 5985 OopTaskQueueSet* task_queues() { return queues(); } 5986 5987 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 5988 5989 void do_work_steal(int i, 5990 CMSParDrainMarkingStackClosure* drain, 5991 CMSParKeepAliveClosure* keep_alive, 5992 int* seed); 5993 5994 virtual void work(uint worker_id); 5995 }; 5996 5997 void CMSRefProcTaskProxy::work(uint worker_id) { 5998 assert(_collector->_span.equals(_span), "Inconsistency in _span"); 5999 CMSParKeepAliveClosure par_keep_alive(_collector, _span, 6000 _mark_bit_map, 6001 work_queue(worker_id)); 6002 CMSParDrainMarkingStackClosure par_drain_stack(_collector, _span, 6003 _mark_bit_map, 6004 work_queue(worker_id)); 6005 CMSIsAliveClosure is_alive_closure(_span, _mark_bit_map); 6006 _task.work(worker_id, is_alive_closure, par_keep_alive, par_drain_stack); 6007 if (_task.marks_oops_alive()) { 6008 do_work_steal(worker_id, &par_drain_stack, &par_keep_alive, 6009 _collector->hash_seed(worker_id)); 6010 } 6011 assert(work_queue(worker_id)->size() == 0, "work_queue should be empty"); 6012 assert(_collector->_overflow_list == NULL, "non-empty _overflow_list"); 6013 } 6014 6015 class CMSRefEnqueueTaskProxy: public AbstractGangTask { 6016 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask; 6017 EnqueueTask& _task; 6018 6019 public: 6020 CMSRefEnqueueTaskProxy(EnqueueTask& task) 6021 : AbstractGangTask("Enqueue reference objects in parallel"), 6022 _task(task) 6023 { } 6024 6025 virtual void work(uint worker_id) 6026 { 6027 _task.work(worker_id); 6028 } 6029 }; 6030 6031 CMSParKeepAliveClosure::CMSParKeepAliveClosure(CMSCollector* collector, 6032 MemRegion span, CMSBitMap* bit_map, OopTaskQueue* work_queue): 6033 _span(span), 6034 _bit_map(bit_map), 6035 _work_queue(work_queue), 6036 _mark_and_push(collector, span, bit_map, work_queue), 6037 _low_water_mark(MIN2((uint)(work_queue->max_elems()/4), 6038 (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads))) 6039 { } 6040 6041 // . see if we can share work_queues with ParNew? XXX 6042 void CMSRefProcTaskProxy::do_work_steal(int i, 6043 CMSParDrainMarkingStackClosure* drain, 6044 CMSParKeepAliveClosure* keep_alive, 6045 int* seed) { 6046 OopTaskQueue* work_q = work_queue(i); 6047 NOT_PRODUCT(int num_steals = 0;) 6048 oop obj_to_scan; 6049 6050 while (true) { 6051 // Completely finish any left over work from (an) earlier round(s) 6052 drain->trim_queue(0); 6053 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 6054 (size_t)ParGCDesiredObjsFromOverflowList); 6055 // Now check if there's any work in the overflow list 6056 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads, 6057 // only affects the number of attempts made to get work from the 6058 // overflow list and does not affect the number of workers. Just 6059 // pass ParallelGCThreads so this behavior is unchanged. 6060 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 6061 work_q, 6062 ParallelGCThreads)) { 6063 // Found something in global overflow list; 6064 // not yet ready to go stealing work from others. 6065 // We'd like to assert(work_q->size() != 0, ...) 6066 // because we just took work from the overflow list, 6067 // but of course we can't, since all of that might have 6068 // been already stolen from us. 6069 continue; 6070 } 6071 // Verify that we have no work before we resort to stealing 6072 assert(work_q->size() == 0, "Have work, shouldn't steal"); 6073 // Try to steal from other queues that have work 6074 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 6075 NOT_PRODUCT(num_steals++;) 6076 assert(obj_to_scan->is_oop(), "Oops, not an oop!"); 6077 assert(_mark_bit_map->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 6078 // Do scanning work 6079 obj_to_scan->oop_iterate(keep_alive); 6080 // Loop around, finish this work, and try to steal some more 6081 } else if (terminator()->offer_termination()) { 6082 break; // nirvana from the infinite cycle 6083 } 6084 } 6085 NOT_PRODUCT( 6086 if (PrintCMSStatistics != 0) { 6087 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals); 6088 } 6089 ) 6090 } 6091 6092 void CMSRefProcTaskExecutor::execute(ProcessTask& task) 6093 { 6094 GenCollectedHeap* gch = GenCollectedHeap::heap(); 6095 FlexibleWorkGang* workers = gch->workers(); 6096 assert(workers != NULL, "Need parallel worker threads."); 6097 CMSRefProcTaskProxy rp_task(task, &_collector, 6098 _collector.ref_processor()->span(), 6099 _collector.markBitMap(), 6100 workers, _collector.task_queues()); 6101 workers->run_task(&rp_task); 6102 } 6103 6104 void CMSRefProcTaskExecutor::execute(EnqueueTask& task) 6105 { 6106 6107 GenCollectedHeap* gch = GenCollectedHeap::heap(); 6108 FlexibleWorkGang* workers = gch->workers(); 6109 assert(workers != NULL, "Need parallel worker threads."); 6110 CMSRefEnqueueTaskProxy enq_task(task); 6111 workers->run_task(&enq_task); 6112 } 6113 6114 void CMSCollector::refProcessingWork(bool asynch, bool clear_all_soft_refs) { 6115 6116 ResourceMark rm; 6117 HandleMark hm; 6118 6119 ReferenceProcessor* rp = ref_processor(); 6120 assert(rp->span().equals(_span), "Spans should be equal"); 6121 assert(!rp->enqueuing_is_done(), "Enqueuing should not be complete"); 6122 // Process weak references. 6123 rp->setup_policy(clear_all_soft_refs); 6124 verify_work_stacks_empty(); 6125 6126 CMSKeepAliveClosure cmsKeepAliveClosure(this, _span, &_markBitMap, 6127 &_markStack, false /* !preclean */); 6128 CMSDrainMarkingStackClosure cmsDrainMarkingStackClosure(this, 6129 _span, &_markBitMap, &_markStack, 6130 &cmsKeepAliveClosure, false /* !preclean */); 6131 { 6132 GCTraceTime t("weak refs processing", PrintGCDetails, false, _gc_timer_cm); 6133 6134 ReferenceProcessorStats stats; 6135 if (rp->processing_is_mt()) { 6136 // Set the degree of MT here. If the discovery is done MT, there 6137 // may have been a different number of threads doing the discovery 6138 // and a different number of discovered lists may have Ref objects. 6139 // That is OK as long as the Reference lists are balanced (see 6140 // balance_all_queues() and balance_queues()). 6141 GenCollectedHeap* gch = GenCollectedHeap::heap(); 6142 int active_workers = ParallelGCThreads; 6143 FlexibleWorkGang* workers = gch->workers(); 6144 if (workers != NULL) { 6145 active_workers = workers->active_workers(); 6146 // The expectation is that active_workers will have already 6147 // been set to a reasonable value. If it has not been set, 6148 // investigate. 6149 assert(active_workers > 0, "Should have been set during scavenge"); 6150 } 6151 rp->set_active_mt_degree(active_workers); 6152 CMSRefProcTaskExecutor task_executor(*this); 6153 stats = rp->process_discovered_references(&_is_alive_closure, 6154 &cmsKeepAliveClosure, 6155 &cmsDrainMarkingStackClosure, 6156 &task_executor, 6157 _gc_timer_cm); 6158 } else { 6159 stats = rp->process_discovered_references(&_is_alive_closure, 6160 &cmsKeepAliveClosure, 6161 &cmsDrainMarkingStackClosure, 6162 NULL, 6163 _gc_timer_cm); 6164 } 6165 _gc_tracer_cm->report_gc_reference_stats(stats); 6166 6167 } 6168 6169 // This is the point where the entire marking should have completed. 6170 verify_work_stacks_empty(); 6171 6172 if (should_unload_classes()) { 6173 { 6174 GCTraceTime t("class unloading", PrintGCDetails, false, _gc_timer_cm); 6175 6176 // Unload classes and purge the SystemDictionary. 6177 bool purged_class = SystemDictionary::do_unloading(&_is_alive_closure); 6178 6179 // Unload nmethods. 6180 CodeCache::do_unloading(&_is_alive_closure, purged_class); 6181 6182 // Prune dead klasses from subklass/sibling/implementor lists. 6183 Klass::clean_weak_klass_links(&_is_alive_closure); 6184 } 6185 6186 { 6187 GCTraceTime t("scrub symbol table", PrintGCDetails, false, _gc_timer_cm); 6188 // Clean up unreferenced symbols in symbol table. 6189 SymbolTable::unlink(); 6190 } 6191 } 6192 6193 // CMS doesn't use the StringTable as hard roots when class unloading is turned off. 6194 // Need to check if we really scanned the StringTable. 6195 if ((roots_scanning_options() & SharedHeap::SO_Strings) == 0) { 6196 GCTraceTime t("scrub string table", PrintGCDetails, false, _gc_timer_cm); 6197 // Delete entries for dead interned strings. 6198 StringTable::unlink(&_is_alive_closure); 6199 } 6200 6201 // Restore any preserved marks as a result of mark stack or 6202 // work queue overflow 6203 restore_preserved_marks_if_any(); // done single-threaded for now 6204 6205 rp->set_enqueuing_is_done(true); 6206 if (rp->processing_is_mt()) { 6207 rp->balance_all_queues(); 6208 CMSRefProcTaskExecutor task_executor(*this); 6209 rp->enqueue_discovered_references(&task_executor); 6210 } else { 6211 rp->enqueue_discovered_references(NULL); 6212 } 6213 rp->verify_no_references_recorded(); 6214 assert(!rp->discovery_enabled(), "should have been disabled"); 6215 } 6216 6217 #ifndef PRODUCT 6218 void CMSCollector::check_correct_thread_executing() { 6219 Thread* t = Thread::current(); 6220 // Only the VM thread or the CMS thread should be here. 6221 assert(t->is_ConcurrentGC_thread() || t->is_VM_thread(), 6222 "Unexpected thread type"); 6223 // If this is the vm thread, the foreground process 6224 // should not be waiting. Note that _foregroundGCIsActive is 6225 // true while the foreground collector is waiting. 6226 if (_foregroundGCShouldWait) { 6227 // We cannot be the VM thread 6228 assert(t->is_ConcurrentGC_thread(), 6229 "Should be CMS thread"); 6230 } else { 6231 // We can be the CMS thread only if we are in a stop-world 6232 // phase of CMS collection. 6233 if (t->is_ConcurrentGC_thread()) { 6234 assert(_collectorState == InitialMarking || 6235 _collectorState == FinalMarking, 6236 "Should be a stop-world phase"); 6237 // The CMS thread should be holding the CMS_token. 6238 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6239 "Potential interference with concurrently " 6240 "executing VM thread"); 6241 } 6242 } 6243 } 6244 #endif 6245 6246 void CMSCollector::sweep(bool asynch) { 6247 assert(_collectorState == Sweeping, "just checking"); 6248 check_correct_thread_executing(); 6249 verify_work_stacks_empty(); 6250 verify_overflow_empty(); 6251 increment_sweep_count(); 6252 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 6253 6254 _inter_sweep_timer.stop(); 6255 _inter_sweep_estimate.sample(_inter_sweep_timer.seconds()); 6256 size_policy()->avg_cms_free_at_sweep()->sample(_cmsGen->free()); 6257 6258 assert(!_intra_sweep_timer.is_active(), "Should not be active"); 6259 _intra_sweep_timer.reset(); 6260 _intra_sweep_timer.start(); 6261 if (asynch) { 6262 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 6263 CMSPhaseAccounting pa(this, "sweep", !PrintGCDetails); 6264 // First sweep the old gen 6265 { 6266 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(), 6267 bitMapLock()); 6268 sweepWork(_cmsGen, asynch); 6269 } 6270 6271 // Update Universe::_heap_*_at_gc figures. 6272 // We need all the free list locks to make the abstract state 6273 // transition from Sweeping to Resetting. See detailed note 6274 // further below. 6275 { 6276 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock()); 6277 // Update heap occupancy information which is used as 6278 // input to soft ref clearing policy at the next gc. 6279 Universe::update_heap_info_at_gc(); 6280 _collectorState = Resizing; 6281 } 6282 } else { 6283 // already have needed locks 6284 sweepWork(_cmsGen, asynch); 6285 // Update heap occupancy information which is used as 6286 // input to soft ref clearing policy at the next gc. 6287 Universe::update_heap_info_at_gc(); 6288 _collectorState = Resizing; 6289 } 6290 verify_work_stacks_empty(); 6291 verify_overflow_empty(); 6292 6293 if (should_unload_classes()) { 6294 ClassLoaderDataGraph::purge(); 6295 } 6296 6297 _intra_sweep_timer.stop(); 6298 _intra_sweep_estimate.sample(_intra_sweep_timer.seconds()); 6299 6300 _inter_sweep_timer.reset(); 6301 _inter_sweep_timer.start(); 6302 6303 // We need to use a monotonically non-deccreasing time in ms 6304 // or we will see time-warp warnings and os::javaTimeMillis() 6305 // does not guarantee monotonicity. 6306 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC; 6307 update_time_of_last_gc(now); 6308 6309 // NOTE on abstract state transitions: 6310 // Mutators allocate-live and/or mark the mod-union table dirty 6311 // based on the state of the collection. The former is done in 6312 // the interval [Marking, Sweeping] and the latter in the interval 6313 // [Marking, Sweeping). Thus the transitions into the Marking state 6314 // and out of the Sweeping state must be synchronously visible 6315 // globally to the mutators. 6316 // The transition into the Marking state happens with the world 6317 // stopped so the mutators will globally see it. Sweeping is 6318 // done asynchronously by the background collector so the transition 6319 // from the Sweeping state to the Resizing state must be done 6320 // under the freelistLock (as is the check for whether to 6321 // allocate-live and whether to dirty the mod-union table). 6322 assert(_collectorState == Resizing, "Change of collector state to" 6323 " Resizing must be done under the freelistLocks (plural)"); 6324 6325 // Now that sweeping has been completed, we clear 6326 // the incremental_collection_failed flag, 6327 // thus inviting a younger gen collection to promote into 6328 // this generation. If such a promotion may still fail, 6329 // the flag will be set again when a young collection is 6330 // attempted. 6331 GenCollectedHeap* gch = GenCollectedHeap::heap(); 6332 gch->clear_incremental_collection_failed(); // Worth retrying as fresh space may have been freed up 6333 gch->update_full_collections_completed(_collection_count_start); 6334 } 6335 6336 // FIX ME!!! Looks like this belongs in CFLSpace, with 6337 // CMSGen merely delegating to it. 6338 void ConcurrentMarkSweepGeneration::setNearLargestChunk() { 6339 double nearLargestPercent = FLSLargestBlockCoalesceProximity; 6340 HeapWord* minAddr = _cmsSpace->bottom(); 6341 HeapWord* largestAddr = 6342 (HeapWord*) _cmsSpace->dictionary()->find_largest_dict(); 6343 if (largestAddr == NULL) { 6344 // The dictionary appears to be empty. In this case 6345 // try to coalesce at the end of the heap. 6346 largestAddr = _cmsSpace->end(); 6347 } 6348 size_t largestOffset = pointer_delta(largestAddr, minAddr); 6349 size_t nearLargestOffset = 6350 (size_t)((double)largestOffset * nearLargestPercent) - MinChunkSize; 6351 if (PrintFLSStatistics != 0) { 6352 gclog_or_tty->print_cr( 6353 "CMS: Large Block: " PTR_FORMAT ";" 6354 " Proximity: " PTR_FORMAT " -> " PTR_FORMAT, 6355 largestAddr, 6356 _cmsSpace->nearLargestChunk(), minAddr + nearLargestOffset); 6357 } 6358 _cmsSpace->set_nearLargestChunk(minAddr + nearLargestOffset); 6359 } 6360 6361 bool ConcurrentMarkSweepGeneration::isNearLargestChunk(HeapWord* addr) { 6362 return addr >= _cmsSpace->nearLargestChunk(); 6363 } 6364 6365 FreeChunk* ConcurrentMarkSweepGeneration::find_chunk_at_end() { 6366 return _cmsSpace->find_chunk_at_end(); 6367 } 6368 6369 void ConcurrentMarkSweepGeneration::update_gc_stats(int current_level, 6370 bool full) { 6371 // The next lower level has been collected. Gather any statistics 6372 // that are of interest at this point. 6373 if (!full && (current_level + 1) == level()) { 6374 // Gather statistics on the young generation collection. 6375 collector()->stats().record_gc0_end(used()); 6376 } 6377 } 6378 6379 CMSAdaptiveSizePolicy* ConcurrentMarkSweepGeneration::size_policy() { 6380 GenCollectedHeap* gch = GenCollectedHeap::heap(); 6381 assert(gch->kind() == CollectedHeap::GenCollectedHeap, 6382 "Wrong type of heap"); 6383 CMSAdaptiveSizePolicy* sp = (CMSAdaptiveSizePolicy*) 6384 gch->gen_policy()->size_policy(); 6385 assert(sp->is_gc_cms_adaptive_size_policy(), 6386 "Wrong type of size policy"); 6387 return sp; 6388 } 6389 6390 void ConcurrentMarkSweepGeneration::rotate_debug_collection_type() { 6391 if (PrintGCDetails && Verbose) { 6392 gclog_or_tty->print("Rotate from %d ", _debug_collection_type); 6393 } 6394 _debug_collection_type = (CollectionTypes) (_debug_collection_type + 1); 6395 _debug_collection_type = 6396 (CollectionTypes) (_debug_collection_type % Unknown_collection_type); 6397 if (PrintGCDetails && Verbose) { 6398 gclog_or_tty->print_cr("to %d ", _debug_collection_type); 6399 } 6400 } 6401 6402 void CMSCollector::sweepWork(ConcurrentMarkSweepGeneration* gen, 6403 bool asynch) { 6404 // We iterate over the space(s) underlying this generation, 6405 // checking the mark bit map to see if the bits corresponding 6406 // to specific blocks are marked or not. Blocks that are 6407 // marked are live and are not swept up. All remaining blocks 6408 // are swept up, with coalescing on-the-fly as we sweep up 6409 // contiguous free and/or garbage blocks: 6410 // We need to ensure that the sweeper synchronizes with allocators 6411 // and stop-the-world collectors. In particular, the following 6412 // locks are used: 6413 // . CMS token: if this is held, a stop the world collection cannot occur 6414 // . freelistLock: if this is held no allocation can occur from this 6415 // generation by another thread 6416 // . bitMapLock: if this is held, no other thread can access or update 6417 // 6418 6419 // Note that we need to hold the freelistLock if we use 6420 // block iterate below; else the iterator might go awry if 6421 // a mutator (or promotion) causes block contents to change 6422 // (for instance if the allocator divvies up a block). 6423 // If we hold the free list lock, for all practical purposes 6424 // young generation GC's can't occur (they'll usually need to 6425 // promote), so we might as well prevent all young generation 6426 // GC's while we do a sweeping step. For the same reason, we might 6427 // as well take the bit map lock for the entire duration 6428 6429 // check that we hold the requisite locks 6430 assert(have_cms_token(), "Should hold cms token"); 6431 assert( (asynch && ConcurrentMarkSweepThread::cms_thread_has_cms_token()) 6432 || (!asynch && ConcurrentMarkSweepThread::vm_thread_has_cms_token()), 6433 "Should possess CMS token to sweep"); 6434 assert_lock_strong(gen->freelistLock()); 6435 assert_lock_strong(bitMapLock()); 6436 6437 assert(!_inter_sweep_timer.is_active(), "Was switched off in an outer context"); 6438 assert(_intra_sweep_timer.is_active(), "Was switched on in an outer context"); 6439 gen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()), 6440 _inter_sweep_estimate.padded_average(), 6441 _intra_sweep_estimate.padded_average()); 6442 gen->setNearLargestChunk(); 6443 6444 { 6445 SweepClosure sweepClosure(this, gen, &_markBitMap, 6446 CMSYield && asynch); 6447 gen->cmsSpace()->blk_iterate_careful(&sweepClosure); 6448 // We need to free-up/coalesce garbage/blocks from a 6449 // co-terminal free run. This is done in the SweepClosure 6450 // destructor; so, do not remove this scope, else the 6451 // end-of-sweep-census below will be off by a little bit. 6452 } 6453 gen->cmsSpace()->sweep_completed(); 6454 gen->cmsSpace()->endSweepFLCensus(sweep_count()); 6455 if (should_unload_classes()) { // unloaded classes this cycle, 6456 _concurrent_cycles_since_last_unload = 0; // ... reset count 6457 } else { // did not unload classes, 6458 _concurrent_cycles_since_last_unload++; // ... increment count 6459 } 6460 } 6461 6462 // Reset CMS data structures (for now just the marking bit map) 6463 // preparatory for the next cycle. 6464 void CMSCollector::reset(bool asynch) { 6465 GenCollectedHeap* gch = GenCollectedHeap::heap(); 6466 CMSAdaptiveSizePolicy* sp = size_policy(); 6467 AdaptiveSizePolicyOutput(sp, gch->total_collections()); 6468 if (asynch) { 6469 CMSTokenSyncWithLocks ts(true, bitMapLock()); 6470 6471 // If the state is not "Resetting", the foreground thread 6472 // has done a collection and the resetting. 6473 if (_collectorState != Resetting) { 6474 assert(_collectorState == Idling, "The state should only change" 6475 " because the foreground collector has finished the collection"); 6476 return; 6477 } 6478 6479 // Clear the mark bitmap (no grey objects to start with) 6480 // for the next cycle. 6481 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 6482 CMSPhaseAccounting cmspa(this, "reset", !PrintGCDetails); 6483 6484 HeapWord* curAddr = _markBitMap.startWord(); 6485 while (curAddr < _markBitMap.endWord()) { 6486 size_t remaining = pointer_delta(_markBitMap.endWord(), curAddr); 6487 MemRegion chunk(curAddr, MIN2(CMSBitMapYieldQuantum, remaining)); 6488 _markBitMap.clear_large_range(chunk); 6489 if (ConcurrentMarkSweepThread::should_yield() && 6490 !foregroundGCIsActive() && 6491 CMSYield) { 6492 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6493 "CMS thread should hold CMS token"); 6494 assert_lock_strong(bitMapLock()); 6495 bitMapLock()->unlock(); 6496 ConcurrentMarkSweepThread::desynchronize(true); 6497 ConcurrentMarkSweepThread::acknowledge_yield_request(); 6498 stopTimer(); 6499 if (PrintCMSStatistics != 0) { 6500 incrementYields(); 6501 } 6502 icms_wait(); 6503 6504 // See the comment in coordinator_yield() 6505 for (unsigned i = 0; i < CMSYieldSleepCount && 6506 ConcurrentMarkSweepThread::should_yield() && 6507 !CMSCollector::foregroundGCIsActive(); ++i) { 6508 os::sleep(Thread::current(), 1, false); 6509 ConcurrentMarkSweepThread::acknowledge_yield_request(); 6510 } 6511 6512 ConcurrentMarkSweepThread::synchronize(true); 6513 bitMapLock()->lock_without_safepoint_check(); 6514 startTimer(); 6515 } 6516 curAddr = chunk.end(); 6517 } 6518 // A successful mostly concurrent collection has been done. 6519 // Because only the full (i.e., concurrent mode failure) collections 6520 // are being measured for gc overhead limits, clean the "near" flag 6521 // and count. 6522 sp->reset_gc_overhead_limit_count(); 6523 _collectorState = Idling; 6524 } else { 6525 // already have the lock 6526 assert(_collectorState == Resetting, "just checking"); 6527 assert_lock_strong(bitMapLock()); 6528 _markBitMap.clear_all(); 6529 _collectorState = Idling; 6530 } 6531 6532 // Stop incremental mode after a cycle completes, so that any future cycles 6533 // are triggered by allocation. 6534 stop_icms(); 6535 6536 NOT_PRODUCT( 6537 if (RotateCMSCollectionTypes) { 6538 _cmsGen->rotate_debug_collection_type(); 6539 } 6540 ) 6541 6542 register_gc_end(); 6543 } 6544 6545 void CMSCollector::do_CMS_operation(CMS_op_type op, GCCause::Cause gc_cause) { 6546 gclog_or_tty->date_stamp(PrintGC && PrintGCDateStamps); 6547 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 6548 GCTraceTime t(GCCauseString("GC", gc_cause), PrintGC, !PrintGCDetails, NULL); 6549 TraceCollectorStats tcs(counters()); 6550 6551 switch (op) { 6552 case CMS_op_checkpointRootsInitial: { 6553 SvcGCMarker sgcm(SvcGCMarker::OTHER); 6554 checkpointRootsInitial(true); // asynch 6555 if (PrintGC) { 6556 _cmsGen->printOccupancy("initial-mark"); 6557 } 6558 break; 6559 } 6560 case CMS_op_checkpointRootsFinal: { 6561 SvcGCMarker sgcm(SvcGCMarker::OTHER); 6562 checkpointRootsFinal(true, // asynch 6563 false, // !clear_all_soft_refs 6564 false); // !init_mark_was_synchronous 6565 if (PrintGC) { 6566 _cmsGen->printOccupancy("remark"); 6567 } 6568 break; 6569 } 6570 default: 6571 fatal("No such CMS_op"); 6572 } 6573 } 6574 6575 #ifndef PRODUCT 6576 size_t const CMSCollector::skip_header_HeapWords() { 6577 return FreeChunk::header_size(); 6578 } 6579 6580 // Try and collect here conditions that should hold when 6581 // CMS thread is exiting. The idea is that the foreground GC 6582 // thread should not be blocked if it wants to terminate 6583 // the CMS thread and yet continue to run the VM for a while 6584 // after that. 6585 void CMSCollector::verify_ok_to_terminate() const { 6586 assert(Thread::current()->is_ConcurrentGC_thread(), 6587 "should be called by CMS thread"); 6588 assert(!_foregroundGCShouldWait, "should be false"); 6589 // We could check here that all the various low-level locks 6590 // are not held by the CMS thread, but that is overkill; see 6591 // also CMSThread::verify_ok_to_terminate() where the CGC_lock 6592 // is checked. 6593 } 6594 #endif 6595 6596 size_t CMSCollector::block_size_using_printezis_bits(HeapWord* addr) const { 6597 assert(_markBitMap.isMarked(addr) && _markBitMap.isMarked(addr + 1), 6598 "missing Printezis mark?"); 6599 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 6600 size_t size = pointer_delta(nextOneAddr + 1, addr); 6601 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 6602 "alignment problem"); 6603 assert(size >= 3, "Necessary for Printezis marks to work"); 6604 return size; 6605 } 6606 6607 // A variant of the above (block_size_using_printezis_bits()) except 6608 // that we return 0 if the P-bits are not yet set. 6609 size_t CMSCollector::block_size_if_printezis_bits(HeapWord* addr) const { 6610 if (_markBitMap.isMarked(addr + 1)) { 6611 assert(_markBitMap.isMarked(addr), "P-bit can be set only for marked objects"); 6612 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 6613 size_t size = pointer_delta(nextOneAddr + 1, addr); 6614 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 6615 "alignment problem"); 6616 assert(size >= 3, "Necessary for Printezis marks to work"); 6617 return size; 6618 } 6619 return 0; 6620 } 6621 6622 HeapWord* CMSCollector::next_card_start_after_block(HeapWord* addr) const { 6623 size_t sz = 0; 6624 oop p = (oop)addr; 6625 if (p->klass_or_null() != NULL) { 6626 sz = CompactibleFreeListSpace::adjustObjectSize(p->size()); 6627 } else { 6628 sz = block_size_using_printezis_bits(addr); 6629 } 6630 assert(sz > 0, "size must be nonzero"); 6631 HeapWord* next_block = addr + sz; 6632 HeapWord* next_card = (HeapWord*)round_to((uintptr_t)next_block, 6633 CardTableModRefBS::card_size); 6634 assert(round_down((uintptr_t)addr, CardTableModRefBS::card_size) < 6635 round_down((uintptr_t)next_card, CardTableModRefBS::card_size), 6636 "must be different cards"); 6637 return next_card; 6638 } 6639 6640 6641 // CMS Bit Map Wrapper ///////////////////////////////////////// 6642 6643 // Construct a CMS bit map infrastructure, but don't create the 6644 // bit vector itself. That is done by a separate call CMSBitMap::allocate() 6645 // further below. 6646 CMSBitMap::CMSBitMap(int shifter, int mutex_rank, const char* mutex_name): 6647 _bm(), 6648 _shifter(shifter), 6649 _lock(mutex_rank >= 0 ? new Mutex(mutex_rank, mutex_name, true) : NULL) 6650 { 6651 _bmStartWord = 0; 6652 _bmWordSize = 0; 6653 } 6654 6655 bool CMSBitMap::allocate(MemRegion mr) { 6656 _bmStartWord = mr.start(); 6657 _bmWordSize = mr.word_size(); 6658 ReservedSpace brs(ReservedSpace::allocation_align_size_up( 6659 (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1)); 6660 if (!brs.is_reserved()) { 6661 warning("CMS bit map allocation failure"); 6662 return false; 6663 } 6664 // For now we'll just commit all of the bit map up fromt. 6665 // Later on we'll try to be more parsimonious with swap. 6666 if (!_virtual_space.initialize(brs, brs.size())) { 6667 warning("CMS bit map backing store failure"); 6668 return false; 6669 } 6670 assert(_virtual_space.committed_size() == brs.size(), 6671 "didn't reserve backing store for all of CMS bit map?"); 6672 _bm.set_map((BitMap::bm_word_t*)_virtual_space.low()); 6673 assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >= 6674 _bmWordSize, "inconsistency in bit map sizing"); 6675 _bm.set_size(_bmWordSize >> _shifter); 6676 6677 // bm.clear(); // can we rely on getting zero'd memory? verify below 6678 assert(isAllClear(), 6679 "Expected zero'd memory from ReservedSpace constructor"); 6680 assert(_bm.size() == heapWordDiffToOffsetDiff(sizeInWords()), 6681 "consistency check"); 6682 return true; 6683 } 6684 6685 void CMSBitMap::dirty_range_iterate_clear(MemRegion mr, MemRegionClosure* cl) { 6686 HeapWord *next_addr, *end_addr, *last_addr; 6687 assert_locked(); 6688 assert(covers(mr), "out-of-range error"); 6689 // XXX assert that start and end are appropriately aligned 6690 for (next_addr = mr.start(), end_addr = mr.end(); 6691 next_addr < end_addr; next_addr = last_addr) { 6692 MemRegion dirty_region = getAndClearMarkedRegion(next_addr, end_addr); 6693 last_addr = dirty_region.end(); 6694 if (!dirty_region.is_empty()) { 6695 cl->do_MemRegion(dirty_region); 6696 } else { 6697 assert(last_addr == end_addr, "program logic"); 6698 return; 6699 } 6700 } 6701 } 6702 6703 void CMSBitMap::print_on_error(outputStream* st, const char* prefix) const { 6704 _bm.print_on_error(st, prefix); 6705 } 6706 6707 #ifndef PRODUCT 6708 void CMSBitMap::assert_locked() const { 6709 CMSLockVerifier::assert_locked(lock()); 6710 } 6711 6712 bool CMSBitMap::covers(MemRegion mr) const { 6713 // assert(_bm.map() == _virtual_space.low(), "map inconsistency"); 6714 assert((size_t)_bm.size() == (_bmWordSize >> _shifter), 6715 "size inconsistency"); 6716 return (mr.start() >= _bmStartWord) && 6717 (mr.end() <= endWord()); 6718 } 6719 6720 bool CMSBitMap::covers(HeapWord* start, size_t size) const { 6721 return (start >= _bmStartWord && (start + size) <= endWord()); 6722 } 6723 6724 void CMSBitMap::verifyNoOneBitsInRange(HeapWord* left, HeapWord* right) { 6725 // verify that there are no 1 bits in the interval [left, right) 6726 FalseBitMapClosure falseBitMapClosure; 6727 iterate(&falseBitMapClosure, left, right); 6728 } 6729 6730 void CMSBitMap::region_invariant(MemRegion mr) 6731 { 6732 assert_locked(); 6733 // mr = mr.intersection(MemRegion(_bmStartWord, _bmWordSize)); 6734 assert(!mr.is_empty(), "unexpected empty region"); 6735 assert(covers(mr), "mr should be covered by bit map"); 6736 // convert address range into offset range 6737 size_t start_ofs = heapWordToOffset(mr.start()); 6738 // Make sure that end() is appropriately aligned 6739 assert(mr.end() == (HeapWord*)round_to((intptr_t)mr.end(), 6740 (1 << (_shifter+LogHeapWordSize))), 6741 "Misaligned mr.end()"); 6742 size_t end_ofs = heapWordToOffset(mr.end()); 6743 assert(end_ofs > start_ofs, "Should mark at least one bit"); 6744 } 6745 6746 #endif 6747 6748 bool CMSMarkStack::allocate(size_t size) { 6749 // allocate a stack of the requisite depth 6750 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 6751 size * sizeof(oop))); 6752 if (!rs.is_reserved()) { 6753 warning("CMSMarkStack allocation failure"); 6754 return false; 6755 } 6756 if (!_virtual_space.initialize(rs, rs.size())) { 6757 warning("CMSMarkStack backing store failure"); 6758 return false; 6759 } 6760 assert(_virtual_space.committed_size() == rs.size(), 6761 "didn't reserve backing store for all of CMS stack?"); 6762 _base = (oop*)(_virtual_space.low()); 6763 _index = 0; 6764 _capacity = size; 6765 NOT_PRODUCT(_max_depth = 0); 6766 return true; 6767 } 6768 6769 // XXX FIX ME !!! In the MT case we come in here holding a 6770 // leaf lock. For printing we need to take a further lock 6771 // which has lower rank. We need to recallibrate the two 6772 // lock-ranks involved in order to be able to rpint the 6773 // messages below. (Or defer the printing to the caller. 6774 // For now we take the expedient path of just disabling the 6775 // messages for the problematic case.) 6776 void CMSMarkStack::expand() { 6777 assert(_capacity <= MarkStackSizeMax, "stack bigger than permitted"); 6778 if (_capacity == MarkStackSizeMax) { 6779 if (_hit_limit++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) { 6780 // We print a warning message only once per CMS cycle. 6781 gclog_or_tty->print_cr(" (benign) Hit CMSMarkStack max size limit"); 6782 } 6783 return; 6784 } 6785 // Double capacity if possible 6786 size_t new_capacity = MIN2(_capacity*2, MarkStackSizeMax); 6787 // Do not give up existing stack until we have managed to 6788 // get the double capacity that we desired. 6789 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 6790 new_capacity * sizeof(oop))); 6791 if (rs.is_reserved()) { 6792 // Release the backing store associated with old stack 6793 _virtual_space.release(); 6794 // Reinitialize virtual space for new stack 6795 if (!_virtual_space.initialize(rs, rs.size())) { 6796 fatal("Not enough swap for expanded marking stack"); 6797 } 6798 _base = (oop*)(_virtual_space.low()); 6799 _index = 0; 6800 _capacity = new_capacity; 6801 } else if (_failed_double++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) { 6802 // Failed to double capacity, continue; 6803 // we print a detail message only once per CMS cycle. 6804 gclog_or_tty->print(" (benign) Failed to expand marking stack from "SIZE_FORMAT"K to " 6805 SIZE_FORMAT"K", 6806 _capacity / K, new_capacity / K); 6807 } 6808 } 6809 6810 6811 // Closures 6812 // XXX: there seems to be a lot of code duplication here; 6813 // should refactor and consolidate common code. 6814 6815 // This closure is used to mark refs into the CMS generation in 6816 // the CMS bit map. Called at the first checkpoint. This closure 6817 // assumes that we do not need to re-mark dirty cards; if the CMS 6818 // generation on which this is used is not an oldest 6819 // generation then this will lose younger_gen cards! 6820 6821 MarkRefsIntoClosure::MarkRefsIntoClosure( 6822 MemRegion span, CMSBitMap* bitMap): 6823 _span(span), 6824 _bitMap(bitMap) 6825 { 6826 assert(_ref_processor == NULL, "deliberately left NULL"); 6827 assert(_bitMap->covers(_span), "_bitMap/_span mismatch"); 6828 } 6829 6830 void MarkRefsIntoClosure::do_oop(oop obj) { 6831 // if p points into _span, then mark corresponding bit in _markBitMap 6832 assert(obj->is_oop(), "expected an oop"); 6833 HeapWord* addr = (HeapWord*)obj; 6834 if (_span.contains(addr)) { 6835 // this should be made more efficient 6836 _bitMap->mark(addr); 6837 } 6838 } 6839 6840 void MarkRefsIntoClosure::do_oop(oop* p) { MarkRefsIntoClosure::do_oop_work(p); } 6841 void MarkRefsIntoClosure::do_oop(narrowOop* p) { MarkRefsIntoClosure::do_oop_work(p); } 6842 6843 // A variant of the above, used for CMS marking verification. 6844 MarkRefsIntoVerifyClosure::MarkRefsIntoVerifyClosure( 6845 MemRegion span, CMSBitMap* verification_bm, CMSBitMap* cms_bm): 6846 _span(span), 6847 _verification_bm(verification_bm), 6848 _cms_bm(cms_bm) 6849 { 6850 assert(_ref_processor == NULL, "deliberately left NULL"); 6851 assert(_verification_bm->covers(_span), "_verification_bm/_span mismatch"); 6852 } 6853 6854 void MarkRefsIntoVerifyClosure::do_oop(oop obj) { 6855 // if p points into _span, then mark corresponding bit in _markBitMap 6856 assert(obj->is_oop(), "expected an oop"); 6857 HeapWord* addr = (HeapWord*)obj; 6858 if (_span.contains(addr)) { 6859 _verification_bm->mark(addr); 6860 if (!_cms_bm->isMarked(addr)) { 6861 oop(addr)->print(); 6862 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)", addr); 6863 fatal("... aborting"); 6864 } 6865 } 6866 } 6867 6868 void MarkRefsIntoVerifyClosure::do_oop(oop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); } 6869 void MarkRefsIntoVerifyClosure::do_oop(narrowOop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); } 6870 6871 ////////////////////////////////////////////////// 6872 // MarkRefsIntoAndScanClosure 6873 ////////////////////////////////////////////////// 6874 6875 MarkRefsIntoAndScanClosure::MarkRefsIntoAndScanClosure(MemRegion span, 6876 ReferenceProcessor* rp, 6877 CMSBitMap* bit_map, 6878 CMSBitMap* mod_union_table, 6879 CMSMarkStack* mark_stack, 6880 CMSCollector* collector, 6881 bool should_yield, 6882 bool concurrent_precleaning): 6883 _collector(collector), 6884 _span(span), 6885 _bit_map(bit_map), 6886 _mark_stack(mark_stack), 6887 _pushAndMarkClosure(collector, span, rp, bit_map, mod_union_table, 6888 mark_stack, concurrent_precleaning), 6889 _yield(should_yield), 6890 _concurrent_precleaning(concurrent_precleaning), 6891 _freelistLock(NULL) 6892 { 6893 _ref_processor = rp; 6894 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 6895 } 6896 6897 // This closure is used to mark refs into the CMS generation at the 6898 // second (final) checkpoint, and to scan and transitively follow 6899 // the unmarked oops. It is also used during the concurrent precleaning 6900 // phase while scanning objects on dirty cards in the CMS generation. 6901 // The marks are made in the marking bit map and the marking stack is 6902 // used for keeping the (newly) grey objects during the scan. 6903 // The parallel version (Par_...) appears further below. 6904 void MarkRefsIntoAndScanClosure::do_oop(oop obj) { 6905 if (obj != NULL) { 6906 assert(obj->is_oop(), "expected an oop"); 6907 HeapWord* addr = (HeapWord*)obj; 6908 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 6909 assert(_collector->overflow_list_is_empty(), 6910 "overflow list should be empty"); 6911 if (_span.contains(addr) && 6912 !_bit_map->isMarked(addr)) { 6913 // mark bit map (object is now grey) 6914 _bit_map->mark(addr); 6915 // push on marking stack (stack should be empty), and drain the 6916 // stack by applying this closure to the oops in the oops popped 6917 // from the stack (i.e. blacken the grey objects) 6918 bool res = _mark_stack->push(obj); 6919 assert(res, "Should have space to push on empty stack"); 6920 do { 6921 oop new_oop = _mark_stack->pop(); 6922 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 6923 assert(_bit_map->isMarked((HeapWord*)new_oop), 6924 "only grey objects on this stack"); 6925 // iterate over the oops in this oop, marking and pushing 6926 // the ones in CMS heap (i.e. in _span). 6927 new_oop->oop_iterate(&_pushAndMarkClosure); 6928 // check if it's time to yield 6929 do_yield_check(); 6930 } while (!_mark_stack->isEmpty() || 6931 (!_concurrent_precleaning && take_from_overflow_list())); 6932 // if marking stack is empty, and we are not doing this 6933 // during precleaning, then check the overflow list 6934 } 6935 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 6936 assert(_collector->overflow_list_is_empty(), 6937 "overflow list was drained above"); 6938 // We could restore evacuated mark words, if any, used for 6939 // overflow list links here because the overflow list is 6940 // provably empty here. That would reduce the maximum 6941 // size requirements for preserved_{oop,mark}_stack. 6942 // But we'll just postpone it until we are all done 6943 // so we can just stream through. 6944 if (!_concurrent_precleaning && CMSOverflowEarlyRestoration) { 6945 _collector->restore_preserved_marks_if_any(); 6946 assert(_collector->no_preserved_marks(), "No preserved marks"); 6947 } 6948 assert(!CMSOverflowEarlyRestoration || _collector->no_preserved_marks(), 6949 "All preserved marks should have been restored above"); 6950 } 6951 } 6952 6953 void MarkRefsIntoAndScanClosure::do_oop(oop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); } 6954 void MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); } 6955 6956 void MarkRefsIntoAndScanClosure::do_yield_work() { 6957 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6958 "CMS thread should hold CMS token"); 6959 assert_lock_strong(_freelistLock); 6960 assert_lock_strong(_bit_map->lock()); 6961 // relinquish the free_list_lock and bitMaplock() 6962 _bit_map->lock()->unlock(); 6963 _freelistLock->unlock(); 6964 ConcurrentMarkSweepThread::desynchronize(true); 6965 ConcurrentMarkSweepThread::acknowledge_yield_request(); 6966 _collector->stopTimer(); 6967 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 6968 if (PrintCMSStatistics != 0) { 6969 _collector->incrementYields(); 6970 } 6971 _collector->icms_wait(); 6972 6973 // See the comment in coordinator_yield() 6974 for (unsigned i = 0; 6975 i < CMSYieldSleepCount && 6976 ConcurrentMarkSweepThread::should_yield() && 6977 !CMSCollector::foregroundGCIsActive(); 6978 ++i) { 6979 os::sleep(Thread::current(), 1, false); 6980 ConcurrentMarkSweepThread::acknowledge_yield_request(); 6981 } 6982 6983 ConcurrentMarkSweepThread::synchronize(true); 6984 _freelistLock->lock_without_safepoint_check(); 6985 _bit_map->lock()->lock_without_safepoint_check(); 6986 _collector->startTimer(); 6987 } 6988 6989 /////////////////////////////////////////////////////////// 6990 // Par_MarkRefsIntoAndScanClosure: a parallel version of 6991 // MarkRefsIntoAndScanClosure 6992 /////////////////////////////////////////////////////////// 6993 Par_MarkRefsIntoAndScanClosure::Par_MarkRefsIntoAndScanClosure( 6994 CMSCollector* collector, MemRegion span, ReferenceProcessor* rp, 6995 CMSBitMap* bit_map, OopTaskQueue* work_queue): 6996 _span(span), 6997 _bit_map(bit_map), 6998 _work_queue(work_queue), 6999 _low_water_mark(MIN2((uint)(work_queue->max_elems()/4), 7000 (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads))), 7001 _par_pushAndMarkClosure(collector, span, rp, bit_map, work_queue) 7002 { 7003 _ref_processor = rp; 7004 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 7005 } 7006 7007 // This closure is used to mark refs into the CMS generation at the 7008 // second (final) checkpoint, and to scan and transitively follow 7009 // the unmarked oops. The marks are made in the marking bit map and 7010 // the work_queue is used for keeping the (newly) grey objects during 7011 // the scan phase whence they are also available for stealing by parallel 7012 // threads. Since the marking bit map is shared, updates are 7013 // synchronized (via CAS). 7014 void Par_MarkRefsIntoAndScanClosure::do_oop(oop obj) { 7015 if (obj != NULL) { 7016 // Ignore mark word because this could be an already marked oop 7017 // that may be chained at the end of the overflow list. 7018 assert(obj->is_oop(true), "expected an oop"); 7019 HeapWord* addr = (HeapWord*)obj; 7020 if (_span.contains(addr) && 7021 !_bit_map->isMarked(addr)) { 7022 // mark bit map (object will become grey): 7023 // It is possible for several threads to be 7024 // trying to "claim" this object concurrently; 7025 // the unique thread that succeeds in marking the 7026 // object first will do the subsequent push on 7027 // to the work queue (or overflow list). 7028 if (_bit_map->par_mark(addr)) { 7029 // push on work_queue (which may not be empty), and trim the 7030 // queue to an appropriate length by applying this closure to 7031 // the oops in the oops popped from the stack (i.e. blacken the 7032 // grey objects) 7033 bool res = _work_queue->push(obj); 7034 assert(res, "Low water mark should be less than capacity?"); 7035 trim_queue(_low_water_mark); 7036 } // Else, another thread claimed the object 7037 } 7038 } 7039 } 7040 7041 void Par_MarkRefsIntoAndScanClosure::do_oop(oop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); } 7042 void Par_MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); } 7043 7044 // This closure is used to rescan the marked objects on the dirty cards 7045 // in the mod union table and the card table proper. 7046 size_t ScanMarkedObjectsAgainCarefullyClosure::do_object_careful_m( 7047 oop p, MemRegion mr) { 7048 7049 size_t size = 0; 7050 HeapWord* addr = (HeapWord*)p; 7051 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 7052 assert(_span.contains(addr), "we are scanning the CMS generation"); 7053 // check if it's time to yield 7054 if (do_yield_check()) { 7055 // We yielded for some foreground stop-world work, 7056 // and we have been asked to abort this ongoing preclean cycle. 7057 return 0; 7058 } 7059 if (_bitMap->isMarked(addr)) { 7060 // it's marked; is it potentially uninitialized? 7061 if (p->klass_or_null() != NULL) { 7062 // an initialized object; ignore mark word in verification below 7063 // since we are running concurrent with mutators 7064 assert(p->is_oop(true), "should be an oop"); 7065 if (p->is_objArray()) { 7066 // objArrays are precisely marked; restrict scanning 7067 // to dirty cards only. 7068 size = CompactibleFreeListSpace::adjustObjectSize( 7069 p->oop_iterate(_scanningClosure, mr)); 7070 } else { 7071 // A non-array may have been imprecisely marked; we need 7072 // to scan object in its entirety. 7073 size = CompactibleFreeListSpace::adjustObjectSize( 7074 p->oop_iterate(_scanningClosure)); 7075 } 7076 #ifdef ASSERT 7077 size_t direct_size = 7078 CompactibleFreeListSpace::adjustObjectSize(p->size()); 7079 assert(size == direct_size, "Inconsistency in size"); 7080 assert(size >= 3, "Necessary for Printezis marks to work"); 7081 if (!_bitMap->isMarked(addr+1)) { 7082 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size); 7083 } else { 7084 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size-1); 7085 assert(_bitMap->isMarked(addr+size-1), 7086 "inconsistent Printezis mark"); 7087 } 7088 #endif // ASSERT 7089 } else { 7090 // an unitialized object 7091 assert(_bitMap->isMarked(addr+1), "missing Printezis mark?"); 7092 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 7093 size = pointer_delta(nextOneAddr + 1, addr); 7094 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 7095 "alignment problem"); 7096 // Note that pre-cleaning needn't redirty the card. OopDesc::set_klass() 7097 // will dirty the card when the klass pointer is installed in the 7098 // object (signalling the completion of initialization). 7099 } 7100 } else { 7101 // Either a not yet marked object or an uninitialized object 7102 if (p->klass_or_null() == NULL) { 7103 // An uninitialized object, skip to the next card, since 7104 // we may not be able to read its P-bits yet. 7105 assert(size == 0, "Initial value"); 7106 } else { 7107 // An object not (yet) reached by marking: we merely need to 7108 // compute its size so as to go look at the next block. 7109 assert(p->is_oop(true), "should be an oop"); 7110 size = CompactibleFreeListSpace::adjustObjectSize(p->size()); 7111 } 7112 } 7113 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 7114 return size; 7115 } 7116 7117 void ScanMarkedObjectsAgainCarefullyClosure::do_yield_work() { 7118 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 7119 "CMS thread should hold CMS token"); 7120 assert_lock_strong(_freelistLock); 7121 assert_lock_strong(_bitMap->lock()); 7122 // relinquish the free_list_lock and bitMaplock() 7123 _bitMap->lock()->unlock(); 7124 _freelistLock->unlock(); 7125 ConcurrentMarkSweepThread::desynchronize(true); 7126 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7127 _collector->stopTimer(); 7128 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 7129 if (PrintCMSStatistics != 0) { 7130 _collector->incrementYields(); 7131 } 7132 _collector->icms_wait(); 7133 7134 // See the comment in coordinator_yield() 7135 for (unsigned i = 0; i < CMSYieldSleepCount && 7136 ConcurrentMarkSweepThread::should_yield() && 7137 !CMSCollector::foregroundGCIsActive(); ++i) { 7138 os::sleep(Thread::current(), 1, false); 7139 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7140 } 7141 7142 ConcurrentMarkSweepThread::synchronize(true); 7143 _freelistLock->lock_without_safepoint_check(); 7144 _bitMap->lock()->lock_without_safepoint_check(); 7145 _collector->startTimer(); 7146 } 7147 7148 7149 ////////////////////////////////////////////////////////////////// 7150 // SurvivorSpacePrecleanClosure 7151 ////////////////////////////////////////////////////////////////// 7152 // This (single-threaded) closure is used to preclean the oops in 7153 // the survivor spaces. 7154 size_t SurvivorSpacePrecleanClosure::do_object_careful(oop p) { 7155 7156 HeapWord* addr = (HeapWord*)p; 7157 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 7158 assert(!_span.contains(addr), "we are scanning the survivor spaces"); 7159 assert(p->klass_or_null() != NULL, "object should be initializd"); 7160 // an initialized object; ignore mark word in verification below 7161 // since we are running concurrent with mutators 7162 assert(p->is_oop(true), "should be an oop"); 7163 // Note that we do not yield while we iterate over 7164 // the interior oops of p, pushing the relevant ones 7165 // on our marking stack. 7166 size_t size = p->oop_iterate(_scanning_closure); 7167 do_yield_check(); 7168 // Observe that below, we do not abandon the preclean 7169 // phase as soon as we should; rather we empty the 7170 // marking stack before returning. This is to satisfy 7171 // some existing assertions. In general, it may be a 7172 // good idea to abort immediately and complete the marking 7173 // from the grey objects at a later time. 7174 while (!_mark_stack->isEmpty()) { 7175 oop new_oop = _mark_stack->pop(); 7176 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 7177 assert(_bit_map->isMarked((HeapWord*)new_oop), 7178 "only grey objects on this stack"); 7179 // iterate over the oops in this oop, marking and pushing 7180 // the ones in CMS heap (i.e. in _span). 7181 new_oop->oop_iterate(_scanning_closure); 7182 // check if it's time to yield 7183 do_yield_check(); 7184 } 7185 unsigned int after_count = 7186 GenCollectedHeap::heap()->total_collections(); 7187 bool abort = (_before_count != after_count) || 7188 _collector->should_abort_preclean(); 7189 return abort ? 0 : size; 7190 } 7191 7192 void SurvivorSpacePrecleanClosure::do_yield_work() { 7193 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 7194 "CMS thread should hold CMS token"); 7195 assert_lock_strong(_bit_map->lock()); 7196 // Relinquish the bit map lock 7197 _bit_map->lock()->unlock(); 7198 ConcurrentMarkSweepThread::desynchronize(true); 7199 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7200 _collector->stopTimer(); 7201 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 7202 if (PrintCMSStatistics != 0) { 7203 _collector->incrementYields(); 7204 } 7205 _collector->icms_wait(); 7206 7207 // See the comment in coordinator_yield() 7208 for (unsigned i = 0; i < CMSYieldSleepCount && 7209 ConcurrentMarkSweepThread::should_yield() && 7210 !CMSCollector::foregroundGCIsActive(); ++i) { 7211 os::sleep(Thread::current(), 1, false); 7212 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7213 } 7214 7215 ConcurrentMarkSweepThread::synchronize(true); 7216 _bit_map->lock()->lock_without_safepoint_check(); 7217 _collector->startTimer(); 7218 } 7219 7220 // This closure is used to rescan the marked objects on the dirty cards 7221 // in the mod union table and the card table proper. In the parallel 7222 // case, although the bitMap is shared, we do a single read so the 7223 // isMarked() query is "safe". 7224 bool ScanMarkedObjectsAgainClosure::do_object_bm(oop p, MemRegion mr) { 7225 // Ignore mark word because we are running concurrent with mutators 7226 assert(p->is_oop_or_null(true), "expected an oop or null"); 7227 HeapWord* addr = (HeapWord*)p; 7228 assert(_span.contains(addr), "we are scanning the CMS generation"); 7229 bool is_obj_array = false; 7230 #ifdef ASSERT 7231 if (!_parallel) { 7232 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 7233 assert(_collector->overflow_list_is_empty(), 7234 "overflow list should be empty"); 7235 7236 } 7237 #endif // ASSERT 7238 if (_bit_map->isMarked(addr)) { 7239 // Obj arrays are precisely marked, non-arrays are not; 7240 // so we scan objArrays precisely and non-arrays in their 7241 // entirety. 7242 if (p->is_objArray()) { 7243 is_obj_array = true; 7244 if (_parallel) { 7245 p->oop_iterate(_par_scan_closure, mr); 7246 } else { 7247 p->oop_iterate(_scan_closure, mr); 7248 } 7249 } else { 7250 if (_parallel) { 7251 p->oop_iterate(_par_scan_closure); 7252 } else { 7253 p->oop_iterate(_scan_closure); 7254 } 7255 } 7256 } 7257 #ifdef ASSERT 7258 if (!_parallel) { 7259 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 7260 assert(_collector->overflow_list_is_empty(), 7261 "overflow list should be empty"); 7262 7263 } 7264 #endif // ASSERT 7265 return is_obj_array; 7266 } 7267 7268 MarkFromRootsClosure::MarkFromRootsClosure(CMSCollector* collector, 7269 MemRegion span, 7270 CMSBitMap* bitMap, CMSMarkStack* markStack, 7271 bool should_yield, bool verifying): 7272 _collector(collector), 7273 _span(span), 7274 _bitMap(bitMap), 7275 _mut(&collector->_modUnionTable), 7276 _markStack(markStack), 7277 _yield(should_yield), 7278 _skipBits(0) 7279 { 7280 assert(_markStack->isEmpty(), "stack should be empty"); 7281 _finger = _bitMap->startWord(); 7282 _threshold = _finger; 7283 assert(_collector->_restart_addr == NULL, "Sanity check"); 7284 assert(_span.contains(_finger), "Out of bounds _finger?"); 7285 DEBUG_ONLY(_verifying = verifying;) 7286 } 7287 7288 void MarkFromRootsClosure::reset(HeapWord* addr) { 7289 assert(_markStack->isEmpty(), "would cause duplicates on stack"); 7290 assert(_span.contains(addr), "Out of bounds _finger?"); 7291 _finger = addr; 7292 _threshold = (HeapWord*)round_to( 7293 (intptr_t)_finger, CardTableModRefBS::card_size); 7294 } 7295 7296 // Should revisit to see if this should be restructured for 7297 // greater efficiency. 7298 bool MarkFromRootsClosure::do_bit(size_t offset) { 7299 if (_skipBits > 0) { 7300 _skipBits--; 7301 return true; 7302 } 7303 // convert offset into a HeapWord* 7304 HeapWord* addr = _bitMap->startWord() + offset; 7305 assert(_bitMap->endWord() && addr < _bitMap->endWord(), 7306 "address out of range"); 7307 assert(_bitMap->isMarked(addr), "tautology"); 7308 if (_bitMap->isMarked(addr+1)) { 7309 // this is an allocated but not yet initialized object 7310 assert(_skipBits == 0, "tautology"); 7311 _skipBits = 2; // skip next two marked bits ("Printezis-marks") 7312 oop p = oop(addr); 7313 if (p->klass_or_null() == NULL) { 7314 DEBUG_ONLY(if (!_verifying) {) 7315 // We re-dirty the cards on which this object lies and increase 7316 // the _threshold so that we'll come back to scan this object 7317 // during the preclean or remark phase. (CMSCleanOnEnter) 7318 if (CMSCleanOnEnter) { 7319 size_t sz = _collector->block_size_using_printezis_bits(addr); 7320 HeapWord* end_card_addr = (HeapWord*)round_to( 7321 (intptr_t)(addr+sz), CardTableModRefBS::card_size); 7322 MemRegion redirty_range = MemRegion(addr, end_card_addr); 7323 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 7324 // Bump _threshold to end_card_addr; note that 7325 // _threshold cannot possibly exceed end_card_addr, anyhow. 7326 // This prevents future clearing of the card as the scan proceeds 7327 // to the right. 7328 assert(_threshold <= end_card_addr, 7329 "Because we are just scanning into this object"); 7330 if (_threshold < end_card_addr) { 7331 _threshold = end_card_addr; 7332 } 7333 if (p->klass_or_null() != NULL) { 7334 // Redirty the range of cards... 7335 _mut->mark_range(redirty_range); 7336 } // ...else the setting of klass will dirty the card anyway. 7337 } 7338 DEBUG_ONLY(}) 7339 return true; 7340 } 7341 } 7342 scanOopsInOop(addr); 7343 return true; 7344 } 7345 7346 // We take a break if we've been at this for a while, 7347 // so as to avoid monopolizing the locks involved. 7348 void MarkFromRootsClosure::do_yield_work() { 7349 // First give up the locks, then yield, then re-lock 7350 // We should probably use a constructor/destructor idiom to 7351 // do this unlock/lock or modify the MutexUnlocker class to 7352 // serve our purpose. XXX 7353 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 7354 "CMS thread should hold CMS token"); 7355 assert_lock_strong(_bitMap->lock()); 7356 _bitMap->lock()->unlock(); 7357 ConcurrentMarkSweepThread::desynchronize(true); 7358 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7359 _collector->stopTimer(); 7360 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 7361 if (PrintCMSStatistics != 0) { 7362 _collector->incrementYields(); 7363 } 7364 _collector->icms_wait(); 7365 7366 // See the comment in coordinator_yield() 7367 for (unsigned i = 0; i < CMSYieldSleepCount && 7368 ConcurrentMarkSweepThread::should_yield() && 7369 !CMSCollector::foregroundGCIsActive(); ++i) { 7370 os::sleep(Thread::current(), 1, false); 7371 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7372 } 7373 7374 ConcurrentMarkSweepThread::synchronize(true); 7375 _bitMap->lock()->lock_without_safepoint_check(); 7376 _collector->startTimer(); 7377 } 7378 7379 void MarkFromRootsClosure::scanOopsInOop(HeapWord* ptr) { 7380 assert(_bitMap->isMarked(ptr), "expected bit to be set"); 7381 assert(_markStack->isEmpty(), 7382 "should drain stack to limit stack usage"); 7383 // convert ptr to an oop preparatory to scanning 7384 oop obj = oop(ptr); 7385 // Ignore mark word in verification below, since we 7386 // may be running concurrent with mutators. 7387 assert(obj->is_oop(true), "should be an oop"); 7388 assert(_finger <= ptr, "_finger runneth ahead"); 7389 // advance the finger to right end of this object 7390 _finger = ptr + obj->size(); 7391 assert(_finger > ptr, "we just incremented it above"); 7392 // On large heaps, it may take us some time to get through 7393 // the marking phase (especially if running iCMS). During 7394 // this time it's possible that a lot of mutations have 7395 // accumulated in the card table and the mod union table -- 7396 // these mutation records are redundant until we have 7397 // actually traced into the corresponding card. 7398 // Here, we check whether advancing the finger would make 7399 // us cross into a new card, and if so clear corresponding 7400 // cards in the MUT (preclean them in the card-table in the 7401 // future). 7402 7403 DEBUG_ONLY(if (!_verifying) {) 7404 // The clean-on-enter optimization is disabled by default, 7405 // until we fix 6178663. 7406 if (CMSCleanOnEnter && (_finger > _threshold)) { 7407 // [_threshold, _finger) represents the interval 7408 // of cards to be cleared in MUT (or precleaned in card table). 7409 // The set of cards to be cleared is all those that overlap 7410 // with the interval [_threshold, _finger); note that 7411 // _threshold is always kept card-aligned but _finger isn't 7412 // always card-aligned. 7413 HeapWord* old_threshold = _threshold; 7414 assert(old_threshold == (HeapWord*)round_to( 7415 (intptr_t)old_threshold, CardTableModRefBS::card_size), 7416 "_threshold should always be card-aligned"); 7417 _threshold = (HeapWord*)round_to( 7418 (intptr_t)_finger, CardTableModRefBS::card_size); 7419 MemRegion mr(old_threshold, _threshold); 7420 assert(!mr.is_empty(), "Control point invariant"); 7421 assert(_span.contains(mr), "Should clear within span"); 7422 _mut->clear_range(mr); 7423 } 7424 DEBUG_ONLY(}) 7425 // Note: the finger doesn't advance while we drain 7426 // the stack below. 7427 PushOrMarkClosure pushOrMarkClosure(_collector, 7428 _span, _bitMap, _markStack, 7429 _finger, this); 7430 bool res = _markStack->push(obj); 7431 assert(res, "Empty non-zero size stack should have space for single push"); 7432 while (!_markStack->isEmpty()) { 7433 oop new_oop = _markStack->pop(); 7434 // Skip verifying header mark word below because we are 7435 // running concurrent with mutators. 7436 assert(new_oop->is_oop(true), "Oops! expected to pop an oop"); 7437 // now scan this oop's oops 7438 new_oop->oop_iterate(&pushOrMarkClosure); 7439 do_yield_check(); 7440 } 7441 assert(_markStack->isEmpty(), "tautology, emphasizing post-condition"); 7442 } 7443 7444 Par_MarkFromRootsClosure::Par_MarkFromRootsClosure(CMSConcMarkingTask* task, 7445 CMSCollector* collector, MemRegion span, 7446 CMSBitMap* bit_map, 7447 OopTaskQueue* work_queue, 7448 CMSMarkStack* overflow_stack, 7449 bool should_yield): 7450 _collector(collector), 7451 _whole_span(collector->_span), 7452 _span(span), 7453 _bit_map(bit_map), 7454 _mut(&collector->_modUnionTable), 7455 _work_queue(work_queue), 7456 _overflow_stack(overflow_stack), 7457 _yield(should_yield), 7458 _skip_bits(0), 7459 _task(task) 7460 { 7461 assert(_work_queue->size() == 0, "work_queue should be empty"); 7462 _finger = span.start(); 7463 _threshold = _finger; // XXX Defer clear-on-enter optimization for now 7464 assert(_span.contains(_finger), "Out of bounds _finger?"); 7465 } 7466 7467 // Should revisit to see if this should be restructured for 7468 // greater efficiency. 7469 bool Par_MarkFromRootsClosure::do_bit(size_t offset) { 7470 if (_skip_bits > 0) { 7471 _skip_bits--; 7472 return true; 7473 } 7474 // convert offset into a HeapWord* 7475 HeapWord* addr = _bit_map->startWord() + offset; 7476 assert(_bit_map->endWord() && addr < _bit_map->endWord(), 7477 "address out of range"); 7478 assert(_bit_map->isMarked(addr), "tautology"); 7479 if (_bit_map->isMarked(addr+1)) { 7480 // this is an allocated object that might not yet be initialized 7481 assert(_skip_bits == 0, "tautology"); 7482 _skip_bits = 2; // skip next two marked bits ("Printezis-marks") 7483 oop p = oop(addr); 7484 if (p->klass_or_null() == NULL) { 7485 // in the case of Clean-on-Enter optimization, redirty card 7486 // and avoid clearing card by increasing the threshold. 7487 return true; 7488 } 7489 } 7490 scan_oops_in_oop(addr); 7491 return true; 7492 } 7493 7494 void Par_MarkFromRootsClosure::scan_oops_in_oop(HeapWord* ptr) { 7495 assert(_bit_map->isMarked(ptr), "expected bit to be set"); 7496 // Should we assert that our work queue is empty or 7497 // below some drain limit? 7498 assert(_work_queue->size() == 0, 7499 "should drain stack to limit stack usage"); 7500 // convert ptr to an oop preparatory to scanning 7501 oop obj = oop(ptr); 7502 // Ignore mark word in verification below, since we 7503 // may be running concurrent with mutators. 7504 assert(obj->is_oop(true), "should be an oop"); 7505 assert(_finger <= ptr, "_finger runneth ahead"); 7506 // advance the finger to right end of this object 7507 _finger = ptr + obj->size(); 7508 assert(_finger > ptr, "we just incremented it above"); 7509 // On large heaps, it may take us some time to get through 7510 // the marking phase (especially if running iCMS). During 7511 // this time it's possible that a lot of mutations have 7512 // accumulated in the card table and the mod union table -- 7513 // these mutation records are redundant until we have 7514 // actually traced into the corresponding card. 7515 // Here, we check whether advancing the finger would make 7516 // us cross into a new card, and if so clear corresponding 7517 // cards in the MUT (preclean them in the card-table in the 7518 // future). 7519 7520 // The clean-on-enter optimization is disabled by default, 7521 // until we fix 6178663. 7522 if (CMSCleanOnEnter && (_finger > _threshold)) { 7523 // [_threshold, _finger) represents the interval 7524 // of cards to be cleared in MUT (or precleaned in card table). 7525 // The set of cards to be cleared is all those that overlap 7526 // with the interval [_threshold, _finger); note that 7527 // _threshold is always kept card-aligned but _finger isn't 7528 // always card-aligned. 7529 HeapWord* old_threshold = _threshold; 7530 assert(old_threshold == (HeapWord*)round_to( 7531 (intptr_t)old_threshold, CardTableModRefBS::card_size), 7532 "_threshold should always be card-aligned"); 7533 _threshold = (HeapWord*)round_to( 7534 (intptr_t)_finger, CardTableModRefBS::card_size); 7535 MemRegion mr(old_threshold, _threshold); 7536 assert(!mr.is_empty(), "Control point invariant"); 7537 assert(_span.contains(mr), "Should clear within span"); // _whole_span ?? 7538 _mut->clear_range(mr); 7539 } 7540 7541 // Note: the local finger doesn't advance while we drain 7542 // the stack below, but the global finger sure can and will. 7543 HeapWord** gfa = _task->global_finger_addr(); 7544 Par_PushOrMarkClosure pushOrMarkClosure(_collector, 7545 _span, _bit_map, 7546 _work_queue, 7547 _overflow_stack, 7548 _finger, 7549 gfa, this); 7550 bool res = _work_queue->push(obj); // overflow could occur here 7551 assert(res, "Will hold once we use workqueues"); 7552 while (true) { 7553 oop new_oop; 7554 if (!_work_queue->pop_local(new_oop)) { 7555 // We emptied our work_queue; check if there's stuff that can 7556 // be gotten from the overflow stack. 7557 if (CMSConcMarkingTask::get_work_from_overflow_stack( 7558 _overflow_stack, _work_queue)) { 7559 do_yield_check(); 7560 continue; 7561 } else { // done 7562 break; 7563 } 7564 } 7565 // Skip verifying header mark word below because we are 7566 // running concurrent with mutators. 7567 assert(new_oop->is_oop(true), "Oops! expected to pop an oop"); 7568 // now scan this oop's oops 7569 new_oop->oop_iterate(&pushOrMarkClosure); 7570 do_yield_check(); 7571 } 7572 assert(_work_queue->size() == 0, "tautology, emphasizing post-condition"); 7573 } 7574 7575 // Yield in response to a request from VM Thread or 7576 // from mutators. 7577 void Par_MarkFromRootsClosure::do_yield_work() { 7578 assert(_task != NULL, "sanity"); 7579 _task->yield(); 7580 } 7581 7582 // A variant of the above used for verifying CMS marking work. 7583 MarkFromRootsVerifyClosure::MarkFromRootsVerifyClosure(CMSCollector* collector, 7584 MemRegion span, 7585 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 7586 CMSMarkStack* mark_stack): 7587 _collector(collector), 7588 _span(span), 7589 _verification_bm(verification_bm), 7590 _cms_bm(cms_bm), 7591 _mark_stack(mark_stack), 7592 _pam_verify_closure(collector, span, verification_bm, cms_bm, 7593 mark_stack) 7594 { 7595 assert(_mark_stack->isEmpty(), "stack should be empty"); 7596 _finger = _verification_bm->startWord(); 7597 assert(_collector->_restart_addr == NULL, "Sanity check"); 7598 assert(_span.contains(_finger), "Out of bounds _finger?"); 7599 } 7600 7601 void MarkFromRootsVerifyClosure::reset(HeapWord* addr) { 7602 assert(_mark_stack->isEmpty(), "would cause duplicates on stack"); 7603 assert(_span.contains(addr), "Out of bounds _finger?"); 7604 _finger = addr; 7605 } 7606 7607 // Should revisit to see if this should be restructured for 7608 // greater efficiency. 7609 bool MarkFromRootsVerifyClosure::do_bit(size_t offset) { 7610 // convert offset into a HeapWord* 7611 HeapWord* addr = _verification_bm->startWord() + offset; 7612 assert(_verification_bm->endWord() && addr < _verification_bm->endWord(), 7613 "address out of range"); 7614 assert(_verification_bm->isMarked(addr), "tautology"); 7615 assert(_cms_bm->isMarked(addr), "tautology"); 7616 7617 assert(_mark_stack->isEmpty(), 7618 "should drain stack to limit stack usage"); 7619 // convert addr to an oop preparatory to scanning 7620 oop obj = oop(addr); 7621 assert(obj->is_oop(), "should be an oop"); 7622 assert(_finger <= addr, "_finger runneth ahead"); 7623 // advance the finger to right end of this object 7624 _finger = addr + obj->size(); 7625 assert(_finger > addr, "we just incremented it above"); 7626 // Note: the finger doesn't advance while we drain 7627 // the stack below. 7628 bool res = _mark_stack->push(obj); 7629 assert(res, "Empty non-zero size stack should have space for single push"); 7630 while (!_mark_stack->isEmpty()) { 7631 oop new_oop = _mark_stack->pop(); 7632 assert(new_oop->is_oop(), "Oops! expected to pop an oop"); 7633 // now scan this oop's oops 7634 new_oop->oop_iterate(&_pam_verify_closure); 7635 } 7636 assert(_mark_stack->isEmpty(), "tautology, emphasizing post-condition"); 7637 return true; 7638 } 7639 7640 PushAndMarkVerifyClosure::PushAndMarkVerifyClosure( 7641 CMSCollector* collector, MemRegion span, 7642 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 7643 CMSMarkStack* mark_stack): 7644 CMSOopClosure(collector->ref_processor()), 7645 _collector(collector), 7646 _span(span), 7647 _verification_bm(verification_bm), 7648 _cms_bm(cms_bm), 7649 _mark_stack(mark_stack) 7650 { } 7651 7652 void PushAndMarkVerifyClosure::do_oop(oop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 7653 void PushAndMarkVerifyClosure::do_oop(narrowOop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 7654 7655 // Upon stack overflow, we discard (part of) the stack, 7656 // remembering the least address amongst those discarded 7657 // in CMSCollector's _restart_address. 7658 void PushAndMarkVerifyClosure::handle_stack_overflow(HeapWord* lost) { 7659 // Remember the least grey address discarded 7660 HeapWord* ra = (HeapWord*)_mark_stack->least_value(lost); 7661 _collector->lower_restart_addr(ra); 7662 _mark_stack->reset(); // discard stack contents 7663 _mark_stack->expand(); // expand the stack if possible 7664 } 7665 7666 void PushAndMarkVerifyClosure::do_oop(oop obj) { 7667 assert(obj->is_oop_or_null(), "expected an oop or NULL"); 7668 HeapWord* addr = (HeapWord*)obj; 7669 if (_span.contains(addr) && !_verification_bm->isMarked(addr)) { 7670 // Oop lies in _span and isn't yet grey or black 7671 _verification_bm->mark(addr); // now grey 7672 if (!_cms_bm->isMarked(addr)) { 7673 oop(addr)->print(); 7674 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)", 7675 addr); 7676 fatal("... aborting"); 7677 } 7678 7679 if (!_mark_stack->push(obj)) { // stack overflow 7680 if (PrintCMSStatistics != 0) { 7681 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 7682 SIZE_FORMAT, _mark_stack->capacity()); 7683 } 7684 assert(_mark_stack->isFull(), "Else push should have succeeded"); 7685 handle_stack_overflow(addr); 7686 } 7687 // anything including and to the right of _finger 7688 // will be scanned as we iterate over the remainder of the 7689 // bit map 7690 } 7691 } 7692 7693 PushOrMarkClosure::PushOrMarkClosure(CMSCollector* collector, 7694 MemRegion span, 7695 CMSBitMap* bitMap, CMSMarkStack* markStack, 7696 HeapWord* finger, MarkFromRootsClosure* parent) : 7697 CMSOopClosure(collector->ref_processor()), 7698 _collector(collector), 7699 _span(span), 7700 _bitMap(bitMap), 7701 _markStack(markStack), 7702 _finger(finger), 7703 _parent(parent) 7704 { } 7705 7706 Par_PushOrMarkClosure::Par_PushOrMarkClosure(CMSCollector* collector, 7707 MemRegion span, 7708 CMSBitMap* bit_map, 7709 OopTaskQueue* work_queue, 7710 CMSMarkStack* overflow_stack, 7711 HeapWord* finger, 7712 HeapWord** global_finger_addr, 7713 Par_MarkFromRootsClosure* parent) : 7714 CMSOopClosure(collector->ref_processor()), 7715 _collector(collector), 7716 _whole_span(collector->_span), 7717 _span(span), 7718 _bit_map(bit_map), 7719 _work_queue(work_queue), 7720 _overflow_stack(overflow_stack), 7721 _finger(finger), 7722 _global_finger_addr(global_finger_addr), 7723 _parent(parent) 7724 { } 7725 7726 // Assumes thread-safe access by callers, who are 7727 // responsible for mutual exclusion. 7728 void CMSCollector::lower_restart_addr(HeapWord* low) { 7729 assert(_span.contains(low), "Out of bounds addr"); 7730 if (_restart_addr == NULL) { 7731 _restart_addr = low; 7732 } else { 7733 _restart_addr = MIN2(_restart_addr, low); 7734 } 7735 } 7736 7737 // Upon stack overflow, we discard (part of) the stack, 7738 // remembering the least address amongst those discarded 7739 // in CMSCollector's _restart_address. 7740 void PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 7741 // Remember the least grey address discarded 7742 HeapWord* ra = (HeapWord*)_markStack->least_value(lost); 7743 _collector->lower_restart_addr(ra); 7744 _markStack->reset(); // discard stack contents 7745 _markStack->expand(); // expand the stack if possible 7746 } 7747 7748 // Upon stack overflow, we discard (part of) the stack, 7749 // remembering the least address amongst those discarded 7750 // in CMSCollector's _restart_address. 7751 void Par_PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 7752 // We need to do this under a mutex to prevent other 7753 // workers from interfering with the work done below. 7754 MutexLockerEx ml(_overflow_stack->par_lock(), 7755 Mutex::_no_safepoint_check_flag); 7756 // Remember the least grey address discarded 7757 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 7758 _collector->lower_restart_addr(ra); 7759 _overflow_stack->reset(); // discard stack contents 7760 _overflow_stack->expand(); // expand the stack if possible 7761 } 7762 7763 void CMKlassClosure::do_klass(Klass* k) { 7764 assert(_oop_closure != NULL, "Not initialized?"); 7765 k->oops_do(_oop_closure); 7766 } 7767 7768 void PushOrMarkClosure::do_oop(oop obj) { 7769 // Ignore mark word because we are running concurrent with mutators. 7770 assert(obj->is_oop_or_null(true), "expected an oop or NULL"); 7771 HeapWord* addr = (HeapWord*)obj; 7772 if (_span.contains(addr) && !_bitMap->isMarked(addr)) { 7773 // Oop lies in _span and isn't yet grey or black 7774 _bitMap->mark(addr); // now grey 7775 if (addr < _finger) { 7776 // the bit map iteration has already either passed, or 7777 // sampled, this bit in the bit map; we'll need to 7778 // use the marking stack to scan this oop's oops. 7779 bool simulate_overflow = false; 7780 NOT_PRODUCT( 7781 if (CMSMarkStackOverflowALot && 7782 _collector->simulate_overflow()) { 7783 // simulate a stack overflow 7784 simulate_overflow = true; 7785 } 7786 ) 7787 if (simulate_overflow || !_markStack->push(obj)) { // stack overflow 7788 if (PrintCMSStatistics != 0) { 7789 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 7790 SIZE_FORMAT, _markStack->capacity()); 7791 } 7792 assert(simulate_overflow || _markStack->isFull(), "Else push should have succeeded"); 7793 handle_stack_overflow(addr); 7794 } 7795 } 7796 // anything including and to the right of _finger 7797 // will be scanned as we iterate over the remainder of the 7798 // bit map 7799 do_yield_check(); 7800 } 7801 } 7802 7803 void PushOrMarkClosure::do_oop(oop* p) { PushOrMarkClosure::do_oop_work(p); } 7804 void PushOrMarkClosure::do_oop(narrowOop* p) { PushOrMarkClosure::do_oop_work(p); } 7805 7806 void Par_PushOrMarkClosure::do_oop(oop obj) { 7807 // Ignore mark word because we are running concurrent with mutators. 7808 assert(obj->is_oop_or_null(true), "expected an oop or NULL"); 7809 HeapWord* addr = (HeapWord*)obj; 7810 if (_whole_span.contains(addr) && !_bit_map->isMarked(addr)) { 7811 // Oop lies in _span and isn't yet grey or black 7812 // We read the global_finger (volatile read) strictly after marking oop 7813 bool res = _bit_map->par_mark(addr); // now grey 7814 volatile HeapWord** gfa = (volatile HeapWord**)_global_finger_addr; 7815 // Should we push this marked oop on our stack? 7816 // -- if someone else marked it, nothing to do 7817 // -- if target oop is above global finger nothing to do 7818 // -- if target oop is in chunk and above local finger 7819 // then nothing to do 7820 // -- else push on work queue 7821 if ( !res // someone else marked it, they will deal with it 7822 || (addr >= *gfa) // will be scanned in a later task 7823 || (_span.contains(addr) && addr >= _finger)) { // later in this chunk 7824 return; 7825 } 7826 // the bit map iteration has already either passed, or 7827 // sampled, this bit in the bit map; we'll need to 7828 // use the marking stack to scan this oop's oops. 7829 bool simulate_overflow = false; 7830 NOT_PRODUCT( 7831 if (CMSMarkStackOverflowALot && 7832 _collector->simulate_overflow()) { 7833 // simulate a stack overflow 7834 simulate_overflow = true; 7835 } 7836 ) 7837 if (simulate_overflow || 7838 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) { 7839 // stack overflow 7840 if (PrintCMSStatistics != 0) { 7841 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 7842 SIZE_FORMAT, _overflow_stack->capacity()); 7843 } 7844 // We cannot assert that the overflow stack is full because 7845 // it may have been emptied since. 7846 assert(simulate_overflow || 7847 _work_queue->size() == _work_queue->max_elems(), 7848 "Else push should have succeeded"); 7849 handle_stack_overflow(addr); 7850 } 7851 do_yield_check(); 7852 } 7853 } 7854 7855 void Par_PushOrMarkClosure::do_oop(oop* p) { Par_PushOrMarkClosure::do_oop_work(p); } 7856 void Par_PushOrMarkClosure::do_oop(narrowOop* p) { Par_PushOrMarkClosure::do_oop_work(p); } 7857 7858 PushAndMarkClosure::PushAndMarkClosure(CMSCollector* collector, 7859 MemRegion span, 7860 ReferenceProcessor* rp, 7861 CMSBitMap* bit_map, 7862 CMSBitMap* mod_union_table, 7863 CMSMarkStack* mark_stack, 7864 bool concurrent_precleaning): 7865 CMSOopClosure(rp), 7866 _collector(collector), 7867 _span(span), 7868 _bit_map(bit_map), 7869 _mod_union_table(mod_union_table), 7870 _mark_stack(mark_stack), 7871 _concurrent_precleaning(concurrent_precleaning) 7872 { 7873 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 7874 } 7875 7876 // Grey object rescan during pre-cleaning and second checkpoint phases -- 7877 // the non-parallel version (the parallel version appears further below.) 7878 void PushAndMarkClosure::do_oop(oop obj) { 7879 // Ignore mark word verification. If during concurrent precleaning, 7880 // the object monitor may be locked. If during the checkpoint 7881 // phases, the object may already have been reached by a different 7882 // path and may be at the end of the global overflow list (so 7883 // the mark word may be NULL). 7884 assert(obj->is_oop_or_null(true /* ignore mark word */), 7885 "expected an oop or NULL"); 7886 HeapWord* addr = (HeapWord*)obj; 7887 // Check if oop points into the CMS generation 7888 // and is not marked 7889 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 7890 // a white object ... 7891 _bit_map->mark(addr); // ... now grey 7892 // push on the marking stack (grey set) 7893 bool simulate_overflow = false; 7894 NOT_PRODUCT( 7895 if (CMSMarkStackOverflowALot && 7896 _collector->simulate_overflow()) { 7897 // simulate a stack overflow 7898 simulate_overflow = true; 7899 } 7900 ) 7901 if (simulate_overflow || !_mark_stack->push(obj)) { 7902 if (_concurrent_precleaning) { 7903 // During precleaning we can just dirty the appropriate card(s) 7904 // in the mod union table, thus ensuring that the object remains 7905 // in the grey set and continue. In the case of object arrays 7906 // we need to dirty all of the cards that the object spans, 7907 // since the rescan of object arrays will be limited to the 7908 // dirty cards. 7909 // Note that no one can be intefering with us in this action 7910 // of dirtying the mod union table, so no locking or atomics 7911 // are required. 7912 if (obj->is_objArray()) { 7913 size_t sz = obj->size(); 7914 HeapWord* end_card_addr = (HeapWord*)round_to( 7915 (intptr_t)(addr+sz), CardTableModRefBS::card_size); 7916 MemRegion redirty_range = MemRegion(addr, end_card_addr); 7917 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 7918 _mod_union_table->mark_range(redirty_range); 7919 } else { 7920 _mod_union_table->mark(addr); 7921 } 7922 _collector->_ser_pmc_preclean_ovflw++; 7923 } else { 7924 // During the remark phase, we need to remember this oop 7925 // in the overflow list. 7926 _collector->push_on_overflow_list(obj); 7927 _collector->_ser_pmc_remark_ovflw++; 7928 } 7929 } 7930 } 7931 } 7932 7933 Par_PushAndMarkClosure::Par_PushAndMarkClosure(CMSCollector* collector, 7934 MemRegion span, 7935 ReferenceProcessor* rp, 7936 CMSBitMap* bit_map, 7937 OopTaskQueue* work_queue): 7938 CMSOopClosure(rp), 7939 _collector(collector), 7940 _span(span), 7941 _bit_map(bit_map), 7942 _work_queue(work_queue) 7943 { 7944 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 7945 } 7946 7947 void PushAndMarkClosure::do_oop(oop* p) { PushAndMarkClosure::do_oop_work(p); } 7948 void PushAndMarkClosure::do_oop(narrowOop* p) { PushAndMarkClosure::do_oop_work(p); } 7949 7950 // Grey object rescan during second checkpoint phase -- 7951 // the parallel version. 7952 void Par_PushAndMarkClosure::do_oop(oop obj) { 7953 // In the assert below, we ignore the mark word because 7954 // this oop may point to an already visited object that is 7955 // on the overflow stack (in which case the mark word has 7956 // been hijacked for chaining into the overflow stack -- 7957 // if this is the last object in the overflow stack then 7958 // its mark word will be NULL). Because this object may 7959 // have been subsequently popped off the global overflow 7960 // stack, and the mark word possibly restored to the prototypical 7961 // value, by the time we get to examined this failing assert in 7962 // the debugger, is_oop_or_null(false) may subsequently start 7963 // to hold. 7964 assert(obj->is_oop_or_null(true), 7965 "expected an oop or NULL"); 7966 HeapWord* addr = (HeapWord*)obj; 7967 // Check if oop points into the CMS generation 7968 // and is not marked 7969 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 7970 // a white object ... 7971 // If we manage to "claim" the object, by being the 7972 // first thread to mark it, then we push it on our 7973 // marking stack 7974 if (_bit_map->par_mark(addr)) { // ... now grey 7975 // push on work queue (grey set) 7976 bool simulate_overflow = false; 7977 NOT_PRODUCT( 7978 if (CMSMarkStackOverflowALot && 7979 _collector->par_simulate_overflow()) { 7980 // simulate a stack overflow 7981 simulate_overflow = true; 7982 } 7983 ) 7984 if (simulate_overflow || !_work_queue->push(obj)) { 7985 _collector->par_push_on_overflow_list(obj); 7986 _collector->_par_pmc_remark_ovflw++; // imprecise OK: no need to CAS 7987 } 7988 } // Else, some other thread got there first 7989 } 7990 } 7991 7992 void Par_PushAndMarkClosure::do_oop(oop* p) { Par_PushAndMarkClosure::do_oop_work(p); } 7993 void Par_PushAndMarkClosure::do_oop(narrowOop* p) { Par_PushAndMarkClosure::do_oop_work(p); } 7994 7995 void CMSPrecleanRefsYieldClosure::do_yield_work() { 7996 Mutex* bml = _collector->bitMapLock(); 7997 assert_lock_strong(bml); 7998 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 7999 "CMS thread should hold CMS token"); 8000 8001 bml->unlock(); 8002 ConcurrentMarkSweepThread::desynchronize(true); 8003 8004 ConcurrentMarkSweepThread::acknowledge_yield_request(); 8005 8006 _collector->stopTimer(); 8007 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 8008 if (PrintCMSStatistics != 0) { 8009 _collector->incrementYields(); 8010 } 8011 _collector->icms_wait(); 8012 8013 // See the comment in coordinator_yield() 8014 for (unsigned i = 0; i < CMSYieldSleepCount && 8015 ConcurrentMarkSweepThread::should_yield() && 8016 !CMSCollector::foregroundGCIsActive(); ++i) { 8017 os::sleep(Thread::current(), 1, false); 8018 ConcurrentMarkSweepThread::acknowledge_yield_request(); 8019 } 8020 8021 ConcurrentMarkSweepThread::synchronize(true); 8022 bml->lock(); 8023 8024 _collector->startTimer(); 8025 } 8026 8027 bool CMSPrecleanRefsYieldClosure::should_return() { 8028 if (ConcurrentMarkSweepThread::should_yield()) { 8029 do_yield_work(); 8030 } 8031 return _collector->foregroundGCIsActive(); 8032 } 8033 8034 void MarkFromDirtyCardsClosure::do_MemRegion(MemRegion mr) { 8035 assert(((size_t)mr.start())%CardTableModRefBS::card_size_in_words == 0, 8036 "mr should be aligned to start at a card boundary"); 8037 // We'd like to assert: 8038 // assert(mr.word_size()%CardTableModRefBS::card_size_in_words == 0, 8039 // "mr should be a range of cards"); 8040 // However, that would be too strong in one case -- the last 8041 // partition ends at _unallocated_block which, in general, can be 8042 // an arbitrary boundary, not necessarily card aligned. 8043 if (PrintCMSStatistics != 0) { 8044 _num_dirty_cards += 8045 mr.word_size()/CardTableModRefBS::card_size_in_words; 8046 } 8047 _space->object_iterate_mem(mr, &_scan_cl); 8048 } 8049 8050 SweepClosure::SweepClosure(CMSCollector* collector, 8051 ConcurrentMarkSweepGeneration* g, 8052 CMSBitMap* bitMap, bool should_yield) : 8053 _collector(collector), 8054 _g(g), 8055 _sp(g->cmsSpace()), 8056 _limit(_sp->sweep_limit()), 8057 _freelistLock(_sp->freelistLock()), 8058 _bitMap(bitMap), 8059 _yield(should_yield), 8060 _inFreeRange(false), // No free range at beginning of sweep 8061 _freeRangeInFreeLists(false), // No free range at beginning of sweep 8062 _lastFreeRangeCoalesced(false), 8063 _freeFinger(g->used_region().start()) 8064 { 8065 NOT_PRODUCT( 8066 _numObjectsFreed = 0; 8067 _numWordsFreed = 0; 8068 _numObjectsLive = 0; 8069 _numWordsLive = 0; 8070 _numObjectsAlreadyFree = 0; 8071 _numWordsAlreadyFree = 0; 8072 _last_fc = NULL; 8073 8074 _sp->initializeIndexedFreeListArrayReturnedBytes(); 8075 _sp->dictionary()->initialize_dict_returned_bytes(); 8076 ) 8077 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 8078 "sweep _limit out of bounds"); 8079 if (CMSTraceSweeper) { 8080 gclog_or_tty->print_cr("\n====================\nStarting new sweep with limit " PTR_FORMAT, 8081 _limit); 8082 } 8083 } 8084 8085 void SweepClosure::print_on(outputStream* st) const { 8086 tty->print_cr("_sp = [" PTR_FORMAT "," PTR_FORMAT ")", 8087 _sp->bottom(), _sp->end()); 8088 tty->print_cr("_limit = " PTR_FORMAT, _limit); 8089 tty->print_cr("_freeFinger = " PTR_FORMAT, _freeFinger); 8090 NOT_PRODUCT(tty->print_cr("_last_fc = " PTR_FORMAT, _last_fc);) 8091 tty->print_cr("_inFreeRange = %d, _freeRangeInFreeLists = %d, _lastFreeRangeCoalesced = %d", 8092 _inFreeRange, _freeRangeInFreeLists, _lastFreeRangeCoalesced); 8093 } 8094 8095 #ifndef PRODUCT 8096 // Assertion checking only: no useful work in product mode -- 8097 // however, if any of the flags below become product flags, 8098 // you may need to review this code to see if it needs to be 8099 // enabled in product mode. 8100 SweepClosure::~SweepClosure() { 8101 assert_lock_strong(_freelistLock); 8102 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 8103 "sweep _limit out of bounds"); 8104 if (inFreeRange()) { 8105 warning("inFreeRange() should have been reset; dumping state of SweepClosure"); 8106 print(); 8107 ShouldNotReachHere(); 8108 } 8109 if (Verbose && PrintGC) { 8110 gclog_or_tty->print("Collected "SIZE_FORMAT" objects, " SIZE_FORMAT " bytes", 8111 _numObjectsFreed, _numWordsFreed*sizeof(HeapWord)); 8112 gclog_or_tty->print_cr("\nLive "SIZE_FORMAT" objects, " 8113 SIZE_FORMAT" bytes " 8114 "Already free "SIZE_FORMAT" objects, "SIZE_FORMAT" bytes", 8115 _numObjectsLive, _numWordsLive*sizeof(HeapWord), 8116 _numObjectsAlreadyFree, _numWordsAlreadyFree*sizeof(HeapWord)); 8117 size_t totalBytes = (_numWordsFreed + _numWordsLive + _numWordsAlreadyFree) 8118 * sizeof(HeapWord); 8119 gclog_or_tty->print_cr("Total sweep: "SIZE_FORMAT" bytes", totalBytes); 8120 8121 if (PrintCMSStatistics && CMSVerifyReturnedBytes) { 8122 size_t indexListReturnedBytes = _sp->sumIndexedFreeListArrayReturnedBytes(); 8123 size_t dict_returned_bytes = _sp->dictionary()->sum_dict_returned_bytes(); 8124 size_t returned_bytes = indexListReturnedBytes + dict_returned_bytes; 8125 gclog_or_tty->print("Returned "SIZE_FORMAT" bytes", returned_bytes); 8126 gclog_or_tty->print(" Indexed List Returned "SIZE_FORMAT" bytes", 8127 indexListReturnedBytes); 8128 gclog_or_tty->print_cr(" Dictionary Returned "SIZE_FORMAT" bytes", 8129 dict_returned_bytes); 8130 } 8131 } 8132 if (CMSTraceSweeper) { 8133 gclog_or_tty->print_cr("end of sweep with _limit = " PTR_FORMAT "\n================", 8134 _limit); 8135 } 8136 } 8137 #endif // PRODUCT 8138 8139 void SweepClosure::initialize_free_range(HeapWord* freeFinger, 8140 bool freeRangeInFreeLists) { 8141 if (CMSTraceSweeper) { 8142 gclog_or_tty->print("---- Start free range at 0x%x with free block (%d)\n", 8143 freeFinger, freeRangeInFreeLists); 8144 } 8145 assert(!inFreeRange(), "Trampling existing free range"); 8146 set_inFreeRange(true); 8147 set_lastFreeRangeCoalesced(false); 8148 8149 set_freeFinger(freeFinger); 8150 set_freeRangeInFreeLists(freeRangeInFreeLists); 8151 if (CMSTestInFreeList) { 8152 if (freeRangeInFreeLists) { 8153 FreeChunk* fc = (FreeChunk*) freeFinger; 8154 assert(fc->is_free(), "A chunk on the free list should be free."); 8155 assert(fc->size() > 0, "Free range should have a size"); 8156 assert(_sp->verify_chunk_in_free_list(fc), "Chunk is not in free lists"); 8157 } 8158 } 8159 } 8160 8161 // Note that the sweeper runs concurrently with mutators. Thus, 8162 // it is possible for direct allocation in this generation to happen 8163 // in the middle of the sweep. Note that the sweeper also coalesces 8164 // contiguous free blocks. Thus, unless the sweeper and the allocator 8165 // synchronize appropriately freshly allocated blocks may get swept up. 8166 // This is accomplished by the sweeper locking the free lists while 8167 // it is sweeping. Thus blocks that are determined to be free are 8168 // indeed free. There is however one additional complication: 8169 // blocks that have been allocated since the final checkpoint and 8170 // mark, will not have been marked and so would be treated as 8171 // unreachable and swept up. To prevent this, the allocator marks 8172 // the bit map when allocating during the sweep phase. This leads, 8173 // however, to a further complication -- objects may have been allocated 8174 // but not yet initialized -- in the sense that the header isn't yet 8175 // installed. The sweeper can not then determine the size of the block 8176 // in order to skip over it. To deal with this case, we use a technique 8177 // (due to Printezis) to encode such uninitialized block sizes in the 8178 // bit map. Since the bit map uses a bit per every HeapWord, but the 8179 // CMS generation has a minimum object size of 3 HeapWords, it follows 8180 // that "normal marks" won't be adjacent in the bit map (there will 8181 // always be at least two 0 bits between successive 1 bits). We make use 8182 // of these "unused" bits to represent uninitialized blocks -- the bit 8183 // corresponding to the start of the uninitialized object and the next 8184 // bit are both set. Finally, a 1 bit marks the end of the object that 8185 // started with the two consecutive 1 bits to indicate its potentially 8186 // uninitialized state. 8187 8188 size_t SweepClosure::do_blk_careful(HeapWord* addr) { 8189 FreeChunk* fc = (FreeChunk*)addr; 8190 size_t res; 8191 8192 // Check if we are done sweeping. Below we check "addr >= _limit" rather 8193 // than "addr == _limit" because although _limit was a block boundary when 8194 // we started the sweep, it may no longer be one because heap expansion 8195 // may have caused us to coalesce the block ending at the address _limit 8196 // with a newly expanded chunk (this happens when _limit was set to the 8197 // previous _end of the space), so we may have stepped past _limit: 8198 // see the following Zeno-like trail of CRs 6977970, 7008136, 7042740. 8199 if (addr >= _limit) { // we have swept up to or past the limit: finish up 8200 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 8201 "sweep _limit out of bounds"); 8202 assert(addr < _sp->end(), "addr out of bounds"); 8203 // Flush any free range we might be holding as a single 8204 // coalesced chunk to the appropriate free list. 8205 if (inFreeRange()) { 8206 assert(freeFinger() >= _sp->bottom() && freeFinger() < _limit, 8207 err_msg("freeFinger() " PTR_FORMAT" is out-of-bounds", freeFinger())); 8208 flush_cur_free_chunk(freeFinger(), 8209 pointer_delta(addr, freeFinger())); 8210 if (CMSTraceSweeper) { 8211 gclog_or_tty->print("Sweep: last chunk: "); 8212 gclog_or_tty->print("put_free_blk 0x%x ("SIZE_FORMAT") " 8213 "[coalesced:"SIZE_FORMAT"]\n", 8214 freeFinger(), pointer_delta(addr, freeFinger()), 8215 lastFreeRangeCoalesced()); 8216 } 8217 } 8218 8219 // help the iterator loop finish 8220 return pointer_delta(_sp->end(), addr); 8221 } 8222 8223 assert(addr < _limit, "sweep invariant"); 8224 // check if we should yield 8225 do_yield_check(addr); 8226 if (fc->is_free()) { 8227 // Chunk that is already free 8228 res = fc->size(); 8229 do_already_free_chunk(fc); 8230 debug_only(_sp->verifyFreeLists()); 8231 // If we flush the chunk at hand in lookahead_and_flush() 8232 // and it's coalesced with a preceding chunk, then the 8233 // process of "mangling" the payload of the coalesced block 8234 // will cause erasure of the size information from the 8235 // (erstwhile) header of all the coalesced blocks but the 8236 // first, so the first disjunct in the assert will not hold 8237 // in that specific case (in which case the second disjunct 8238 // will hold). 8239 assert(res == fc->size() || ((HeapWord*)fc) + res >= _limit, 8240 "Otherwise the size info doesn't change at this step"); 8241 NOT_PRODUCT( 8242 _numObjectsAlreadyFree++; 8243 _numWordsAlreadyFree += res; 8244 ) 8245 NOT_PRODUCT(_last_fc = fc;) 8246 } else if (!_bitMap->isMarked(addr)) { 8247 // Chunk is fresh garbage 8248 res = do_garbage_chunk(fc); 8249 debug_only(_sp->verifyFreeLists()); 8250 NOT_PRODUCT( 8251 _numObjectsFreed++; 8252 _numWordsFreed += res; 8253 ) 8254 } else { 8255 // Chunk that is alive. 8256 res = do_live_chunk(fc); 8257 debug_only(_sp->verifyFreeLists()); 8258 NOT_PRODUCT( 8259 _numObjectsLive++; 8260 _numWordsLive += res; 8261 ) 8262 } 8263 return res; 8264 } 8265 8266 // For the smart allocation, record following 8267 // split deaths - a free chunk is removed from its free list because 8268 // it is being split into two or more chunks. 8269 // split birth - a free chunk is being added to its free list because 8270 // a larger free chunk has been split and resulted in this free chunk. 8271 // coal death - a free chunk is being removed from its free list because 8272 // it is being coalesced into a large free chunk. 8273 // coal birth - a free chunk is being added to its free list because 8274 // it was created when two or more free chunks where coalesced into 8275 // this free chunk. 8276 // 8277 // These statistics are used to determine the desired number of free 8278 // chunks of a given size. The desired number is chosen to be relative 8279 // to the end of a CMS sweep. The desired number at the end of a sweep 8280 // is the 8281 // count-at-end-of-previous-sweep (an amount that was enough) 8282 // - count-at-beginning-of-current-sweep (the excess) 8283 // + split-births (gains in this size during interval) 8284 // - split-deaths (demands on this size during interval) 8285 // where the interval is from the end of one sweep to the end of the 8286 // next. 8287 // 8288 // When sweeping the sweeper maintains an accumulated chunk which is 8289 // the chunk that is made up of chunks that have been coalesced. That 8290 // will be termed the left-hand chunk. A new chunk of garbage that 8291 // is being considered for coalescing will be referred to as the 8292 // right-hand chunk. 8293 // 8294 // When making a decision on whether to coalesce a right-hand chunk with 8295 // the current left-hand chunk, the current count vs. the desired count 8296 // of the left-hand chunk is considered. Also if the right-hand chunk 8297 // is near the large chunk at the end of the heap (see 8298 // ConcurrentMarkSweepGeneration::isNearLargestChunk()), then the 8299 // left-hand chunk is coalesced. 8300 // 8301 // When making a decision about whether to split a chunk, the desired count 8302 // vs. the current count of the candidate to be split is also considered. 8303 // If the candidate is underpopulated (currently fewer chunks than desired) 8304 // a chunk of an overpopulated (currently more chunks than desired) size may 8305 // be chosen. The "hint" associated with a free list, if non-null, points 8306 // to a free list which may be overpopulated. 8307 // 8308 8309 void SweepClosure::do_already_free_chunk(FreeChunk* fc) { 8310 const size_t size = fc->size(); 8311 // Chunks that cannot be coalesced are not in the 8312 // free lists. 8313 if (CMSTestInFreeList && !fc->cantCoalesce()) { 8314 assert(_sp->verify_chunk_in_free_list(fc), 8315 "free chunk should be in free lists"); 8316 } 8317 // a chunk that is already free, should not have been 8318 // marked in the bit map 8319 HeapWord* const addr = (HeapWord*) fc; 8320 assert(!_bitMap->isMarked(addr), "free chunk should be unmarked"); 8321 // Verify that the bit map has no bits marked between 8322 // addr and purported end of this block. 8323 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 8324 8325 // Some chunks cannot be coalesced under any circumstances. 8326 // See the definition of cantCoalesce(). 8327 if (!fc->cantCoalesce()) { 8328 // This chunk can potentially be coalesced. 8329 if (_sp->adaptive_freelists()) { 8330 // All the work is done in 8331 do_post_free_or_garbage_chunk(fc, size); 8332 } else { // Not adaptive free lists 8333 // this is a free chunk that can potentially be coalesced by the sweeper; 8334 if (!inFreeRange()) { 8335 // if the next chunk is a free block that can't be coalesced 8336 // it doesn't make sense to remove this chunk from the free lists 8337 FreeChunk* nextChunk = (FreeChunk*)(addr + size); 8338 assert((HeapWord*)nextChunk <= _sp->end(), "Chunk size out of bounds?"); 8339 if ((HeapWord*)nextChunk < _sp->end() && // There is another free chunk to the right ... 8340 nextChunk->is_free() && // ... which is free... 8341 nextChunk->cantCoalesce()) { // ... but can't be coalesced 8342 // nothing to do 8343 } else { 8344 // Potentially the start of a new free range: 8345 // Don't eagerly remove it from the free lists. 8346 // No need to remove it if it will just be put 8347 // back again. (Also from a pragmatic point of view 8348 // if it is a free block in a region that is beyond 8349 // any allocated blocks, an assertion will fail) 8350 // Remember the start of a free run. 8351 initialize_free_range(addr, true); 8352 // end - can coalesce with next chunk 8353 } 8354 } else { 8355 // the midst of a free range, we are coalescing 8356 print_free_block_coalesced(fc); 8357 if (CMSTraceSweeper) { 8358 gclog_or_tty->print(" -- pick up free block 0x%x (%d)\n", fc, size); 8359 } 8360 // remove it from the free lists 8361 _sp->removeFreeChunkFromFreeLists(fc); 8362 set_lastFreeRangeCoalesced(true); 8363 // If the chunk is being coalesced and the current free range is 8364 // in the free lists, remove the current free range so that it 8365 // will be returned to the free lists in its entirety - all 8366 // the coalesced pieces included. 8367 if (freeRangeInFreeLists()) { 8368 FreeChunk* ffc = (FreeChunk*) freeFinger(); 8369 assert(ffc->size() == pointer_delta(addr, freeFinger()), 8370 "Size of free range is inconsistent with chunk size."); 8371 if (CMSTestInFreeList) { 8372 assert(_sp->verify_chunk_in_free_list(ffc), 8373 "free range is not in free lists"); 8374 } 8375 _sp->removeFreeChunkFromFreeLists(ffc); 8376 set_freeRangeInFreeLists(false); 8377 } 8378 } 8379 } 8380 // Note that if the chunk is not coalescable (the else arm 8381 // below), we unconditionally flush, without needing to do 8382 // a "lookahead," as we do below. 8383 if (inFreeRange()) lookahead_and_flush(fc, size); 8384 } else { 8385 // Code path common to both original and adaptive free lists. 8386 8387 // cant coalesce with previous block; this should be treated 8388 // as the end of a free run if any 8389 if (inFreeRange()) { 8390 // we kicked some butt; time to pick up the garbage 8391 assert(freeFinger() < addr, "freeFinger points too high"); 8392 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 8393 } 8394 // else, nothing to do, just continue 8395 } 8396 } 8397 8398 size_t SweepClosure::do_garbage_chunk(FreeChunk* fc) { 8399 // This is a chunk of garbage. It is not in any free list. 8400 // Add it to a free list or let it possibly be coalesced into 8401 // a larger chunk. 8402 HeapWord* const addr = (HeapWord*) fc; 8403 const size_t size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 8404 8405 if (_sp->adaptive_freelists()) { 8406 // Verify that the bit map has no bits marked between 8407 // addr and purported end of just dead object. 8408 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 8409 8410 do_post_free_or_garbage_chunk(fc, size); 8411 } else { 8412 if (!inFreeRange()) { 8413 // start of a new free range 8414 assert(size > 0, "A free range should have a size"); 8415 initialize_free_range(addr, false); 8416 } else { 8417 // this will be swept up when we hit the end of the 8418 // free range 8419 if (CMSTraceSweeper) { 8420 gclog_or_tty->print(" -- pick up garbage 0x%x (%d) \n", fc, size); 8421 } 8422 // If the chunk is being coalesced and the current free range is 8423 // in the free lists, remove the current free range so that it 8424 // will be returned to the free lists in its entirety - all 8425 // the coalesced pieces included. 8426 if (freeRangeInFreeLists()) { 8427 FreeChunk* ffc = (FreeChunk*)freeFinger(); 8428 assert(ffc->size() == pointer_delta(addr, freeFinger()), 8429 "Size of free range is inconsistent with chunk size."); 8430 if (CMSTestInFreeList) { 8431 assert(_sp->verify_chunk_in_free_list(ffc), 8432 "free range is not in free lists"); 8433 } 8434 _sp->removeFreeChunkFromFreeLists(ffc); 8435 set_freeRangeInFreeLists(false); 8436 } 8437 set_lastFreeRangeCoalesced(true); 8438 } 8439 // this will be swept up when we hit the end of the free range 8440 8441 // Verify that the bit map has no bits marked between 8442 // addr and purported end of just dead object. 8443 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 8444 } 8445 assert(_limit >= addr + size, 8446 "A freshly garbage chunk can't possibly straddle over _limit"); 8447 if (inFreeRange()) lookahead_and_flush(fc, size); 8448 return size; 8449 } 8450 8451 size_t SweepClosure::do_live_chunk(FreeChunk* fc) { 8452 HeapWord* addr = (HeapWord*) fc; 8453 // The sweeper has just found a live object. Return any accumulated 8454 // left hand chunk to the free lists. 8455 if (inFreeRange()) { 8456 assert(freeFinger() < addr, "freeFinger points too high"); 8457 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 8458 } 8459 8460 // This object is live: we'd normally expect this to be 8461 // an oop, and like to assert the following: 8462 // assert(oop(addr)->is_oop(), "live block should be an oop"); 8463 // However, as we commented above, this may be an object whose 8464 // header hasn't yet been initialized. 8465 size_t size; 8466 assert(_bitMap->isMarked(addr), "Tautology for this control point"); 8467 if (_bitMap->isMarked(addr + 1)) { 8468 // Determine the size from the bit map, rather than trying to 8469 // compute it from the object header. 8470 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 8471 size = pointer_delta(nextOneAddr + 1, addr); 8472 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 8473 "alignment problem"); 8474 8475 #ifdef ASSERT 8476 if (oop(addr)->klass_or_null() != NULL) { 8477 // Ignore mark word because we are running concurrent with mutators 8478 assert(oop(addr)->is_oop(true), "live block should be an oop"); 8479 assert(size == 8480 CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()), 8481 "P-mark and computed size do not agree"); 8482 } 8483 #endif 8484 8485 } else { 8486 // This should be an initialized object that's alive. 8487 assert(oop(addr)->klass_or_null() != NULL, 8488 "Should be an initialized object"); 8489 // Ignore mark word because we are running concurrent with mutators 8490 assert(oop(addr)->is_oop(true), "live block should be an oop"); 8491 // Verify that the bit map has no bits marked between 8492 // addr and purported end of this block. 8493 size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 8494 assert(size >= 3, "Necessary for Printezis marks to work"); 8495 assert(!_bitMap->isMarked(addr+1), "Tautology for this control point"); 8496 DEBUG_ONLY(_bitMap->verifyNoOneBitsInRange(addr+2, addr+size);) 8497 } 8498 return size; 8499 } 8500 8501 void SweepClosure::do_post_free_or_garbage_chunk(FreeChunk* fc, 8502 size_t chunkSize) { 8503 // do_post_free_or_garbage_chunk() should only be called in the case 8504 // of the adaptive free list allocator. 8505 const bool fcInFreeLists = fc->is_free(); 8506 assert(_sp->adaptive_freelists(), "Should only be used in this case."); 8507 assert((HeapWord*)fc <= _limit, "sweep invariant"); 8508 if (CMSTestInFreeList && fcInFreeLists) { 8509 assert(_sp->verify_chunk_in_free_list(fc), "free chunk is not in free lists"); 8510 } 8511 8512 if (CMSTraceSweeper) { 8513 gclog_or_tty->print_cr(" -- pick up another chunk at 0x%x (%d)", fc, chunkSize); 8514 } 8515 8516 HeapWord* const fc_addr = (HeapWord*) fc; 8517 8518 bool coalesce; 8519 const size_t left = pointer_delta(fc_addr, freeFinger()); 8520 const size_t right = chunkSize; 8521 switch (FLSCoalescePolicy) { 8522 // numeric value forms a coalition aggressiveness metric 8523 case 0: { // never coalesce 8524 coalesce = false; 8525 break; 8526 } 8527 case 1: { // coalesce if left & right chunks on overpopulated lists 8528 coalesce = _sp->coalOverPopulated(left) && 8529 _sp->coalOverPopulated(right); 8530 break; 8531 } 8532 case 2: { // coalesce if left chunk on overpopulated list (default) 8533 coalesce = _sp->coalOverPopulated(left); 8534 break; 8535 } 8536 case 3: { // coalesce if left OR right chunk on overpopulated list 8537 coalesce = _sp->coalOverPopulated(left) || 8538 _sp->coalOverPopulated(right); 8539 break; 8540 } 8541 case 4: { // always coalesce 8542 coalesce = true; 8543 break; 8544 } 8545 default: 8546 ShouldNotReachHere(); 8547 } 8548 8549 // Should the current free range be coalesced? 8550 // If the chunk is in a free range and either we decided to coalesce above 8551 // or the chunk is near the large block at the end of the heap 8552 // (isNearLargestChunk() returns true), then coalesce this chunk. 8553 const bool doCoalesce = inFreeRange() 8554 && (coalesce || _g->isNearLargestChunk(fc_addr)); 8555 if (doCoalesce) { 8556 // Coalesce the current free range on the left with the new 8557 // chunk on the right. If either is on a free list, 8558 // it must be removed from the list and stashed in the closure. 8559 if (freeRangeInFreeLists()) { 8560 FreeChunk* const ffc = (FreeChunk*)freeFinger(); 8561 assert(ffc->size() == pointer_delta(fc_addr, freeFinger()), 8562 "Size of free range is inconsistent with chunk size."); 8563 if (CMSTestInFreeList) { 8564 assert(_sp->verify_chunk_in_free_list(ffc), 8565 "Chunk is not in free lists"); 8566 } 8567 _sp->coalDeath(ffc->size()); 8568 _sp->removeFreeChunkFromFreeLists(ffc); 8569 set_freeRangeInFreeLists(false); 8570 } 8571 if (fcInFreeLists) { 8572 _sp->coalDeath(chunkSize); 8573 assert(fc->size() == chunkSize, 8574 "The chunk has the wrong size or is not in the free lists"); 8575 _sp->removeFreeChunkFromFreeLists(fc); 8576 } 8577 set_lastFreeRangeCoalesced(true); 8578 print_free_block_coalesced(fc); 8579 } else { // not in a free range and/or should not coalesce 8580 // Return the current free range and start a new one. 8581 if (inFreeRange()) { 8582 // In a free range but cannot coalesce with the right hand chunk. 8583 // Put the current free range into the free lists. 8584 flush_cur_free_chunk(freeFinger(), 8585 pointer_delta(fc_addr, freeFinger())); 8586 } 8587 // Set up for new free range. Pass along whether the right hand 8588 // chunk is in the free lists. 8589 initialize_free_range((HeapWord*)fc, fcInFreeLists); 8590 } 8591 } 8592 8593 // Lookahead flush: 8594 // If we are tracking a free range, and this is the last chunk that 8595 // we'll look at because its end crosses past _limit, we'll preemptively 8596 // flush it along with any free range we may be holding on to. Note that 8597 // this can be the case only for an already free or freshly garbage 8598 // chunk. If this block is an object, it can never straddle 8599 // over _limit. The "straddling" occurs when _limit is set at 8600 // the previous end of the space when this cycle started, and 8601 // a subsequent heap expansion caused the previously co-terminal 8602 // free block to be coalesced with the newly expanded portion, 8603 // thus rendering _limit a non-block-boundary making it dangerous 8604 // for the sweeper to step over and examine. 8605 void SweepClosure::lookahead_and_flush(FreeChunk* fc, size_t chunk_size) { 8606 assert(inFreeRange(), "Should only be called if currently in a free range."); 8607 HeapWord* const eob = ((HeapWord*)fc) + chunk_size; 8608 assert(_sp->used_region().contains(eob - 1), 8609 err_msg("eob = " PTR_FORMAT " out of bounds wrt _sp = [" PTR_FORMAT "," PTR_FORMAT ")" 8610 " when examining fc = " PTR_FORMAT "(" SIZE_FORMAT ")", 8611 _limit, _sp->bottom(), _sp->end(), fc, chunk_size)); 8612 if (eob >= _limit) { 8613 assert(eob == _limit || fc->is_free(), "Only a free chunk should allow us to cross over the limit"); 8614 if (CMSTraceSweeper) { 8615 gclog_or_tty->print_cr("_limit " PTR_FORMAT " reached or crossed by block " 8616 "[" PTR_FORMAT "," PTR_FORMAT ") in space " 8617 "[" PTR_FORMAT "," PTR_FORMAT ")", 8618 _limit, fc, eob, _sp->bottom(), _sp->end()); 8619 } 8620 // Return the storage we are tracking back into the free lists. 8621 if (CMSTraceSweeper) { 8622 gclog_or_tty->print_cr("Flushing ... "); 8623 } 8624 assert(freeFinger() < eob, "Error"); 8625 flush_cur_free_chunk( freeFinger(), pointer_delta(eob, freeFinger())); 8626 } 8627 } 8628 8629 void SweepClosure::flush_cur_free_chunk(HeapWord* chunk, size_t size) { 8630 assert(inFreeRange(), "Should only be called if currently in a free range."); 8631 assert(size > 0, 8632 "A zero sized chunk cannot be added to the free lists."); 8633 if (!freeRangeInFreeLists()) { 8634 if (CMSTestInFreeList) { 8635 FreeChunk* fc = (FreeChunk*) chunk; 8636 fc->set_size(size); 8637 assert(!_sp->verify_chunk_in_free_list(fc), 8638 "chunk should not be in free lists yet"); 8639 } 8640 if (CMSTraceSweeper) { 8641 gclog_or_tty->print_cr(" -- add free block 0x%x (%d) to free lists", 8642 chunk, size); 8643 } 8644 // A new free range is going to be starting. The current 8645 // free range has not been added to the free lists yet or 8646 // was removed so add it back. 8647 // If the current free range was coalesced, then the death 8648 // of the free range was recorded. Record a birth now. 8649 if (lastFreeRangeCoalesced()) { 8650 _sp->coalBirth(size); 8651 } 8652 _sp->addChunkAndRepairOffsetTable(chunk, size, 8653 lastFreeRangeCoalesced()); 8654 } else if (CMSTraceSweeper) { 8655 gclog_or_tty->print_cr("Already in free list: nothing to flush"); 8656 } 8657 set_inFreeRange(false); 8658 set_freeRangeInFreeLists(false); 8659 } 8660 8661 // We take a break if we've been at this for a while, 8662 // so as to avoid monopolizing the locks involved. 8663 void SweepClosure::do_yield_work(HeapWord* addr) { 8664 // Return current free chunk being used for coalescing (if any) 8665 // to the appropriate freelist. After yielding, the next 8666 // free block encountered will start a coalescing range of 8667 // free blocks. If the next free block is adjacent to the 8668 // chunk just flushed, they will need to wait for the next 8669 // sweep to be coalesced. 8670 if (inFreeRange()) { 8671 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 8672 } 8673 8674 // First give up the locks, then yield, then re-lock. 8675 // We should probably use a constructor/destructor idiom to 8676 // do this unlock/lock or modify the MutexUnlocker class to 8677 // serve our purpose. XXX 8678 assert_lock_strong(_bitMap->lock()); 8679 assert_lock_strong(_freelistLock); 8680 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 8681 "CMS thread should hold CMS token"); 8682 _bitMap->lock()->unlock(); 8683 _freelistLock->unlock(); 8684 ConcurrentMarkSweepThread::desynchronize(true); 8685 ConcurrentMarkSweepThread::acknowledge_yield_request(); 8686 _collector->stopTimer(); 8687 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 8688 if (PrintCMSStatistics != 0) { 8689 _collector->incrementYields(); 8690 } 8691 _collector->icms_wait(); 8692 8693 // See the comment in coordinator_yield() 8694 for (unsigned i = 0; i < CMSYieldSleepCount && 8695 ConcurrentMarkSweepThread::should_yield() && 8696 !CMSCollector::foregroundGCIsActive(); ++i) { 8697 os::sleep(Thread::current(), 1, false); 8698 ConcurrentMarkSweepThread::acknowledge_yield_request(); 8699 } 8700 8701 ConcurrentMarkSweepThread::synchronize(true); 8702 _freelistLock->lock(); 8703 _bitMap->lock()->lock_without_safepoint_check(); 8704 _collector->startTimer(); 8705 } 8706 8707 #ifndef PRODUCT 8708 // This is actually very useful in a product build if it can 8709 // be called from the debugger. Compile it into the product 8710 // as needed. 8711 bool debug_verify_chunk_in_free_list(FreeChunk* fc) { 8712 return debug_cms_space->verify_chunk_in_free_list(fc); 8713 } 8714 #endif 8715 8716 void SweepClosure::print_free_block_coalesced(FreeChunk* fc) const { 8717 if (CMSTraceSweeper) { 8718 gclog_or_tty->print_cr("Sweep:coal_free_blk " PTR_FORMAT " (" SIZE_FORMAT ")", 8719 fc, fc->size()); 8720 } 8721 } 8722 8723 // CMSIsAliveClosure 8724 bool CMSIsAliveClosure::do_object_b(oop obj) { 8725 HeapWord* addr = (HeapWord*)obj; 8726 return addr != NULL && 8727 (!_span.contains(addr) || _bit_map->isMarked(addr)); 8728 } 8729 8730 8731 CMSKeepAliveClosure::CMSKeepAliveClosure( CMSCollector* collector, 8732 MemRegion span, 8733 CMSBitMap* bit_map, CMSMarkStack* mark_stack, 8734 bool cpc): 8735 _collector(collector), 8736 _span(span), 8737 _bit_map(bit_map), 8738 _mark_stack(mark_stack), 8739 _concurrent_precleaning(cpc) { 8740 assert(!_span.is_empty(), "Empty span could spell trouble"); 8741 } 8742 8743 8744 // CMSKeepAliveClosure: the serial version 8745 void CMSKeepAliveClosure::do_oop(oop obj) { 8746 HeapWord* addr = (HeapWord*)obj; 8747 if (_span.contains(addr) && 8748 !_bit_map->isMarked(addr)) { 8749 _bit_map->mark(addr); 8750 bool simulate_overflow = false; 8751 NOT_PRODUCT( 8752 if (CMSMarkStackOverflowALot && 8753 _collector->simulate_overflow()) { 8754 // simulate a stack overflow 8755 simulate_overflow = true; 8756 } 8757 ) 8758 if (simulate_overflow || !_mark_stack->push(obj)) { 8759 if (_concurrent_precleaning) { 8760 // We dirty the overflown object and let the remark 8761 // phase deal with it. 8762 assert(_collector->overflow_list_is_empty(), "Error"); 8763 // In the case of object arrays, we need to dirty all of 8764 // the cards that the object spans. No locking or atomics 8765 // are needed since no one else can be mutating the mod union 8766 // table. 8767 if (obj->is_objArray()) { 8768 size_t sz = obj->size(); 8769 HeapWord* end_card_addr = 8770 (HeapWord*)round_to((intptr_t)(addr+sz), CardTableModRefBS::card_size); 8771 MemRegion redirty_range = MemRegion(addr, end_card_addr); 8772 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 8773 _collector->_modUnionTable.mark_range(redirty_range); 8774 } else { 8775 _collector->_modUnionTable.mark(addr); 8776 } 8777 _collector->_ser_kac_preclean_ovflw++; 8778 } else { 8779 _collector->push_on_overflow_list(obj); 8780 _collector->_ser_kac_ovflw++; 8781 } 8782 } 8783 } 8784 } 8785 8786 void CMSKeepAliveClosure::do_oop(oop* p) { CMSKeepAliveClosure::do_oop_work(p); } 8787 void CMSKeepAliveClosure::do_oop(narrowOop* p) { CMSKeepAliveClosure::do_oop_work(p); } 8788 8789 // CMSParKeepAliveClosure: a parallel version of the above. 8790 // The work queues are private to each closure (thread), 8791 // but (may be) available for stealing by other threads. 8792 void CMSParKeepAliveClosure::do_oop(oop obj) { 8793 HeapWord* addr = (HeapWord*)obj; 8794 if (_span.contains(addr) && 8795 !_bit_map->isMarked(addr)) { 8796 // In general, during recursive tracing, several threads 8797 // may be concurrently getting here; the first one to 8798 // "tag" it, claims it. 8799 if (_bit_map->par_mark(addr)) { 8800 bool res = _work_queue->push(obj); 8801 assert(res, "Low water mark should be much less than capacity"); 8802 // Do a recursive trim in the hope that this will keep 8803 // stack usage lower, but leave some oops for potential stealers 8804 trim_queue(_low_water_mark); 8805 } // Else, another thread got there first 8806 } 8807 } 8808 8809 void CMSParKeepAliveClosure::do_oop(oop* p) { CMSParKeepAliveClosure::do_oop_work(p); } 8810 void CMSParKeepAliveClosure::do_oop(narrowOop* p) { CMSParKeepAliveClosure::do_oop_work(p); } 8811 8812 void CMSParKeepAliveClosure::trim_queue(uint max) { 8813 while (_work_queue->size() > max) { 8814 oop new_oop; 8815 if (_work_queue->pop_local(new_oop)) { 8816 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 8817 assert(_bit_map->isMarked((HeapWord*)new_oop), 8818 "no white objects on this stack!"); 8819 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 8820 // iterate over the oops in this oop, marking and pushing 8821 // the ones in CMS heap (i.e. in _span). 8822 new_oop->oop_iterate(&_mark_and_push); 8823 } 8824 } 8825 } 8826 8827 CMSInnerParMarkAndPushClosure::CMSInnerParMarkAndPushClosure( 8828 CMSCollector* collector, 8829 MemRegion span, CMSBitMap* bit_map, 8830 OopTaskQueue* work_queue): 8831 _collector(collector), 8832 _span(span), 8833 _bit_map(bit_map), 8834 _work_queue(work_queue) { } 8835 8836 void CMSInnerParMarkAndPushClosure::do_oop(oop obj) { 8837 HeapWord* addr = (HeapWord*)obj; 8838 if (_span.contains(addr) && 8839 !_bit_map->isMarked(addr)) { 8840 if (_bit_map->par_mark(addr)) { 8841 bool simulate_overflow = false; 8842 NOT_PRODUCT( 8843 if (CMSMarkStackOverflowALot && 8844 _collector->par_simulate_overflow()) { 8845 // simulate a stack overflow 8846 simulate_overflow = true; 8847 } 8848 ) 8849 if (simulate_overflow || !_work_queue->push(obj)) { 8850 _collector->par_push_on_overflow_list(obj); 8851 _collector->_par_kac_ovflw++; 8852 } 8853 } // Else another thread got there already 8854 } 8855 } 8856 8857 void CMSInnerParMarkAndPushClosure::do_oop(oop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); } 8858 void CMSInnerParMarkAndPushClosure::do_oop(narrowOop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); } 8859 8860 ////////////////////////////////////////////////////////////////// 8861 // CMSExpansionCause ///////////////////////////// 8862 ////////////////////////////////////////////////////////////////// 8863 const char* CMSExpansionCause::to_string(CMSExpansionCause::Cause cause) { 8864 switch (cause) { 8865 case _no_expansion: 8866 return "No expansion"; 8867 case _satisfy_free_ratio: 8868 return "Free ratio"; 8869 case _satisfy_promotion: 8870 return "Satisfy promotion"; 8871 case _satisfy_allocation: 8872 return "allocation"; 8873 case _allocate_par_lab: 8874 return "Par LAB"; 8875 case _allocate_par_spooling_space: 8876 return "Par Spooling Space"; 8877 case _adaptive_size_policy: 8878 return "Ergonomics"; 8879 default: 8880 return "unknown"; 8881 } 8882 } 8883 8884 void CMSDrainMarkingStackClosure::do_void() { 8885 // the max number to take from overflow list at a time 8886 const size_t num = _mark_stack->capacity()/4; 8887 assert(!_concurrent_precleaning || _collector->overflow_list_is_empty(), 8888 "Overflow list should be NULL during concurrent phases"); 8889 while (!_mark_stack->isEmpty() || 8890 // if stack is empty, check the overflow list 8891 _collector->take_from_overflow_list(num, _mark_stack)) { 8892 oop obj = _mark_stack->pop(); 8893 HeapWord* addr = (HeapWord*)obj; 8894 assert(_span.contains(addr), "Should be within span"); 8895 assert(_bit_map->isMarked(addr), "Should be marked"); 8896 assert(obj->is_oop(), "Should be an oop"); 8897 obj->oop_iterate(_keep_alive); 8898 } 8899 } 8900 8901 void CMSParDrainMarkingStackClosure::do_void() { 8902 // drain queue 8903 trim_queue(0); 8904 } 8905 8906 // Trim our work_queue so its length is below max at return 8907 void CMSParDrainMarkingStackClosure::trim_queue(uint max) { 8908 while (_work_queue->size() > max) { 8909 oop new_oop; 8910 if (_work_queue->pop_local(new_oop)) { 8911 assert(new_oop->is_oop(), "Expected an oop"); 8912 assert(_bit_map->isMarked((HeapWord*)new_oop), 8913 "no white objects on this stack!"); 8914 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 8915 // iterate over the oops in this oop, marking and pushing 8916 // the ones in CMS heap (i.e. in _span). 8917 new_oop->oop_iterate(&_mark_and_push); 8918 } 8919 } 8920 } 8921 8922 //////////////////////////////////////////////////////////////////// 8923 // Support for Marking Stack Overflow list handling and related code 8924 //////////////////////////////////////////////////////////////////// 8925 // Much of the following code is similar in shape and spirit to the 8926 // code used in ParNewGC. We should try and share that code 8927 // as much as possible in the future. 8928 8929 #ifndef PRODUCT 8930 // Debugging support for CMSStackOverflowALot 8931 8932 // It's OK to call this multi-threaded; the worst thing 8933 // that can happen is that we'll get a bunch of closely 8934 // spaced simulated oveflows, but that's OK, in fact 8935 // probably good as it would exercise the overflow code 8936 // under contention. 8937 bool CMSCollector::simulate_overflow() { 8938 if (_overflow_counter-- <= 0) { // just being defensive 8939 _overflow_counter = CMSMarkStackOverflowInterval; 8940 return true; 8941 } else { 8942 return false; 8943 } 8944 } 8945 8946 bool CMSCollector::par_simulate_overflow() { 8947 return simulate_overflow(); 8948 } 8949 #endif 8950 8951 // Single-threaded 8952 bool CMSCollector::take_from_overflow_list(size_t num, CMSMarkStack* stack) { 8953 assert(stack->isEmpty(), "Expected precondition"); 8954 assert(stack->capacity() > num, "Shouldn't bite more than can chew"); 8955 size_t i = num; 8956 oop cur = _overflow_list; 8957 const markOop proto = markOopDesc::prototype(); 8958 NOT_PRODUCT(ssize_t n = 0;) 8959 for (oop next; i > 0 && cur != NULL; cur = next, i--) { 8960 next = oop(cur->mark()); 8961 cur->set_mark(proto); // until proven otherwise 8962 assert(cur->is_oop(), "Should be an oop"); 8963 bool res = stack->push(cur); 8964 assert(res, "Bit off more than can chew?"); 8965 NOT_PRODUCT(n++;) 8966 } 8967 _overflow_list = cur; 8968 #ifndef PRODUCT 8969 assert(_num_par_pushes >= n, "Too many pops?"); 8970 _num_par_pushes -=n; 8971 #endif 8972 return !stack->isEmpty(); 8973 } 8974 8975 #define BUSY (oop(0x1aff1aff)) 8976 // (MT-safe) Get a prefix of at most "num" from the list. 8977 // The overflow list is chained through the mark word of 8978 // each object in the list. We fetch the entire list, 8979 // break off a prefix of the right size and return the 8980 // remainder. If other threads try to take objects from 8981 // the overflow list at that time, they will wait for 8982 // some time to see if data becomes available. If (and 8983 // only if) another thread places one or more object(s) 8984 // on the global list before we have returned the suffix 8985 // to the global list, we will walk down our local list 8986 // to find its end and append the global list to 8987 // our suffix before returning it. This suffix walk can 8988 // prove to be expensive (quadratic in the amount of traffic) 8989 // when there are many objects in the overflow list and 8990 // there is much producer-consumer contention on the list. 8991 // *NOTE*: The overflow list manipulation code here and 8992 // in ParNewGeneration:: are very similar in shape, 8993 // except that in the ParNew case we use the old (from/eden) 8994 // copy of the object to thread the list via its klass word. 8995 // Because of the common code, if you make any changes in 8996 // the code below, please check the ParNew version to see if 8997 // similar changes might be needed. 8998 // CR 6797058 has been filed to consolidate the common code. 8999 bool CMSCollector::par_take_from_overflow_list(size_t num, 9000 OopTaskQueue* work_q, 9001 int no_of_gc_threads) { 9002 assert(work_q->size() == 0, "First empty local work queue"); 9003 assert(num < work_q->max_elems(), "Can't bite more than we can chew"); 9004 if (_overflow_list == NULL) { 9005 return false; 9006 } 9007 // Grab the entire list; we'll put back a suffix 9008 oop prefix = (oop)Atomic::xchg_ptr(BUSY, &_overflow_list); 9009 Thread* tid = Thread::current(); 9010 // Before "no_of_gc_threads" was introduced CMSOverflowSpinCount was 9011 // set to ParallelGCThreads. 9012 size_t CMSOverflowSpinCount = (size_t) no_of_gc_threads; // was ParallelGCThreads; 9013 size_t sleep_time_millis = MAX2((size_t)1, num/100); 9014 // If the list is busy, we spin for a short while, 9015 // sleeping between attempts to get the list. 9016 for (size_t spin = 0; prefix == BUSY && spin < CMSOverflowSpinCount; spin++) { 9017 os::sleep(tid, sleep_time_millis, false); 9018 if (_overflow_list == NULL) { 9019 // Nothing left to take 9020 return false; 9021 } else if (_overflow_list != BUSY) { 9022 // Try and grab the prefix 9023 prefix = (oop)Atomic::xchg_ptr(BUSY, &_overflow_list); 9024 } 9025 } 9026 // If the list was found to be empty, or we spun long 9027 // enough, we give up and return empty-handed. If we leave 9028 // the list in the BUSY state below, it must be the case that 9029 // some other thread holds the overflow list and will set it 9030 // to a non-BUSY state in the future. 9031 if (prefix == NULL || prefix == BUSY) { 9032 // Nothing to take or waited long enough 9033 if (prefix == NULL) { 9034 // Write back the NULL in case we overwrote it with BUSY above 9035 // and it is still the same value. 9036 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY); 9037 } 9038 return false; 9039 } 9040 assert(prefix != NULL && prefix != BUSY, "Error"); 9041 size_t i = num; 9042 oop cur = prefix; 9043 // Walk down the first "num" objects, unless we reach the end. 9044 for (; i > 1 && cur->mark() != NULL; cur = oop(cur->mark()), i--); 9045 if (cur->mark() == NULL) { 9046 // We have "num" or fewer elements in the list, so there 9047 // is nothing to return to the global list. 9048 // Write back the NULL in lieu of the BUSY we wrote 9049 // above, if it is still the same value. 9050 if (_overflow_list == BUSY) { 9051 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY); 9052 } 9053 } else { 9054 // Chop off the suffix and rerturn it to the global list. 9055 assert(cur->mark() != BUSY, "Error"); 9056 oop suffix_head = cur->mark(); // suffix will be put back on global list 9057 cur->set_mark(NULL); // break off suffix 9058 // It's possible that the list is still in the empty(busy) state 9059 // we left it in a short while ago; in that case we may be 9060 // able to place back the suffix without incurring the cost 9061 // of a walk down the list. 9062 oop observed_overflow_list = _overflow_list; 9063 oop cur_overflow_list = observed_overflow_list; 9064 bool attached = false; 9065 while (observed_overflow_list == BUSY || observed_overflow_list == NULL) { 9066 observed_overflow_list = 9067 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list); 9068 if (cur_overflow_list == observed_overflow_list) { 9069 attached = true; 9070 break; 9071 } else cur_overflow_list = observed_overflow_list; 9072 } 9073 if (!attached) { 9074 // Too bad, someone else sneaked in (at least) an element; we'll need 9075 // to do a splice. Find tail of suffix so we can prepend suffix to global 9076 // list. 9077 for (cur = suffix_head; cur->mark() != NULL; cur = (oop)(cur->mark())); 9078 oop suffix_tail = cur; 9079 assert(suffix_tail != NULL && suffix_tail->mark() == NULL, 9080 "Tautology"); 9081 observed_overflow_list = _overflow_list; 9082 do { 9083 cur_overflow_list = observed_overflow_list; 9084 if (cur_overflow_list != BUSY) { 9085 // Do the splice ... 9086 suffix_tail->set_mark(markOop(cur_overflow_list)); 9087 } else { // cur_overflow_list == BUSY 9088 suffix_tail->set_mark(NULL); 9089 } 9090 // ... and try to place spliced list back on overflow_list ... 9091 observed_overflow_list = 9092 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list); 9093 } while (cur_overflow_list != observed_overflow_list); 9094 // ... until we have succeeded in doing so. 9095 } 9096 } 9097 9098 // Push the prefix elements on work_q 9099 assert(prefix != NULL, "control point invariant"); 9100 const markOop proto = markOopDesc::prototype(); 9101 oop next; 9102 NOT_PRODUCT(ssize_t n = 0;) 9103 for (cur = prefix; cur != NULL; cur = next) { 9104 next = oop(cur->mark()); 9105 cur->set_mark(proto); // until proven otherwise 9106 assert(cur->is_oop(), "Should be an oop"); 9107 bool res = work_q->push(cur); 9108 assert(res, "Bit off more than we can chew?"); 9109 NOT_PRODUCT(n++;) 9110 } 9111 #ifndef PRODUCT 9112 assert(_num_par_pushes >= n, "Too many pops?"); 9113 Atomic::add_ptr(-(intptr_t)n, &_num_par_pushes); 9114 #endif 9115 return true; 9116 } 9117 9118 // Single-threaded 9119 void CMSCollector::push_on_overflow_list(oop p) { 9120 NOT_PRODUCT(_num_par_pushes++;) 9121 assert(p->is_oop(), "Not an oop"); 9122 preserve_mark_if_necessary(p); 9123 p->set_mark((markOop)_overflow_list); 9124 _overflow_list = p; 9125 } 9126 9127 // Multi-threaded; use CAS to prepend to overflow list 9128 void CMSCollector::par_push_on_overflow_list(oop p) { 9129 NOT_PRODUCT(Atomic::inc_ptr(&_num_par_pushes);) 9130 assert(p->is_oop(), "Not an oop"); 9131 par_preserve_mark_if_necessary(p); 9132 oop observed_overflow_list = _overflow_list; 9133 oop cur_overflow_list; 9134 do { 9135 cur_overflow_list = observed_overflow_list; 9136 if (cur_overflow_list != BUSY) { 9137 p->set_mark(markOop(cur_overflow_list)); 9138 } else { 9139 p->set_mark(NULL); 9140 } 9141 observed_overflow_list = 9142 (oop) Atomic::cmpxchg_ptr(p, &_overflow_list, cur_overflow_list); 9143 } while (cur_overflow_list != observed_overflow_list); 9144 } 9145 #undef BUSY 9146 9147 // Single threaded 9148 // General Note on GrowableArray: pushes may silently fail 9149 // because we are (temporarily) out of C-heap for expanding 9150 // the stack. The problem is quite ubiquitous and affects 9151 // a lot of code in the JVM. The prudent thing for GrowableArray 9152 // to do (for now) is to exit with an error. However, that may 9153 // be too draconian in some cases because the caller may be 9154 // able to recover without much harm. For such cases, we 9155 // should probably introduce a "soft_push" method which returns 9156 // an indication of success or failure with the assumption that 9157 // the caller may be able to recover from a failure; code in 9158 // the VM can then be changed, incrementally, to deal with such 9159 // failures where possible, thus, incrementally hardening the VM 9160 // in such low resource situations. 9161 void CMSCollector::preserve_mark_work(oop p, markOop m) { 9162 _preserved_oop_stack.push(p); 9163 _preserved_mark_stack.push(m); 9164 assert(m == p->mark(), "Mark word changed"); 9165 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), 9166 "bijection"); 9167 } 9168 9169 // Single threaded 9170 void CMSCollector::preserve_mark_if_necessary(oop p) { 9171 markOop m = p->mark(); 9172 if (m->must_be_preserved(p)) { 9173 preserve_mark_work(p, m); 9174 } 9175 } 9176 9177 void CMSCollector::par_preserve_mark_if_necessary(oop p) { 9178 markOop m = p->mark(); 9179 if (m->must_be_preserved(p)) { 9180 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag); 9181 // Even though we read the mark word without holding 9182 // the lock, we are assured that it will not change 9183 // because we "own" this oop, so no other thread can 9184 // be trying to push it on the overflow list; see 9185 // the assertion in preserve_mark_work() that checks 9186 // that m == p->mark(). 9187 preserve_mark_work(p, m); 9188 } 9189 } 9190 9191 // We should be able to do this multi-threaded, 9192 // a chunk of stack being a task (this is 9193 // correct because each oop only ever appears 9194 // once in the overflow list. However, it's 9195 // not very easy to completely overlap this with 9196 // other operations, so will generally not be done 9197 // until all work's been completed. Because we 9198 // expect the preserved oop stack (set) to be small, 9199 // it's probably fine to do this single-threaded. 9200 // We can explore cleverer concurrent/overlapped/parallel 9201 // processing of preserved marks if we feel the 9202 // need for this in the future. Stack overflow should 9203 // be so rare in practice and, when it happens, its 9204 // effect on performance so great that this will 9205 // likely just be in the noise anyway. 9206 void CMSCollector::restore_preserved_marks_if_any() { 9207 assert(SafepointSynchronize::is_at_safepoint(), 9208 "world should be stopped"); 9209 assert(Thread::current()->is_ConcurrentGC_thread() || 9210 Thread::current()->is_VM_thread(), 9211 "should be single-threaded"); 9212 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), 9213 "bijection"); 9214 9215 while (!_preserved_oop_stack.is_empty()) { 9216 oop p = _preserved_oop_stack.pop(); 9217 assert(p->is_oop(), "Should be an oop"); 9218 assert(_span.contains(p), "oop should be in _span"); 9219 assert(p->mark() == markOopDesc::prototype(), 9220 "Set when taken from overflow list"); 9221 markOop m = _preserved_mark_stack.pop(); 9222 p->set_mark(m); 9223 } 9224 assert(_preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(), 9225 "stacks were cleared above"); 9226 } 9227 9228 #ifndef PRODUCT 9229 bool CMSCollector::no_preserved_marks() const { 9230 return _preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(); 9231 } 9232 #endif 9233 9234 CMSAdaptiveSizePolicy* ASConcurrentMarkSweepGeneration::cms_size_policy() const 9235 { 9236 GenCollectedHeap* gch = (GenCollectedHeap*) GenCollectedHeap::heap(); 9237 CMSAdaptiveSizePolicy* size_policy = 9238 (CMSAdaptiveSizePolicy*) gch->gen_policy()->size_policy(); 9239 assert(size_policy->is_gc_cms_adaptive_size_policy(), 9240 "Wrong type for size policy"); 9241 return size_policy; 9242 } 9243 9244 void ASConcurrentMarkSweepGeneration::resize(size_t cur_promo_size, 9245 size_t desired_promo_size) { 9246 if (cur_promo_size < desired_promo_size) { 9247 size_t expand_bytes = desired_promo_size - cur_promo_size; 9248 if (PrintAdaptiveSizePolicy && Verbose) { 9249 gclog_or_tty->print_cr(" ASConcurrentMarkSweepGeneration::resize " 9250 "Expanding tenured generation by " SIZE_FORMAT " (bytes)", 9251 expand_bytes); 9252 } 9253 expand(expand_bytes, 9254 MinHeapDeltaBytes, 9255 CMSExpansionCause::_adaptive_size_policy); 9256 } else if (desired_promo_size < cur_promo_size) { 9257 size_t shrink_bytes = cur_promo_size - desired_promo_size; 9258 if (PrintAdaptiveSizePolicy && Verbose) { 9259 gclog_or_tty->print_cr(" ASConcurrentMarkSweepGeneration::resize " 9260 "Shrinking tenured generation by " SIZE_FORMAT " (bytes)", 9261 shrink_bytes); 9262 } 9263 shrink(shrink_bytes); 9264 } 9265 } 9266 9267 CMSGCAdaptivePolicyCounters* ASConcurrentMarkSweepGeneration::gc_adaptive_policy_counters() { 9268 GenCollectedHeap* gch = GenCollectedHeap::heap(); 9269 CMSGCAdaptivePolicyCounters* counters = 9270 (CMSGCAdaptivePolicyCounters*) gch->collector_policy()->counters(); 9271 assert(counters->kind() == GCPolicyCounters::CMSGCAdaptivePolicyCountersKind, 9272 "Wrong kind of counters"); 9273 return counters; 9274 } 9275 9276 9277 void ASConcurrentMarkSweepGeneration::update_counters() { 9278 if (UsePerfData) { 9279 _space_counters->update_all(); 9280 _gen_counters->update_all(); 9281 CMSGCAdaptivePolicyCounters* counters = gc_adaptive_policy_counters(); 9282 GenCollectedHeap* gch = GenCollectedHeap::heap(); 9283 CMSGCStats* gc_stats_l = (CMSGCStats*) gc_stats(); 9284 assert(gc_stats_l->kind() == GCStats::CMSGCStatsKind, 9285 "Wrong gc statistics type"); 9286 counters->update_counters(gc_stats_l); 9287 } 9288 } 9289 9290 void ASConcurrentMarkSweepGeneration::update_counters(size_t used) { 9291 if (UsePerfData) { 9292 _space_counters->update_used(used); 9293 _space_counters->update_capacity(); 9294 _gen_counters->update_all(); 9295 9296 CMSGCAdaptivePolicyCounters* counters = gc_adaptive_policy_counters(); 9297 GenCollectedHeap* gch = GenCollectedHeap::heap(); 9298 CMSGCStats* gc_stats_l = (CMSGCStats*) gc_stats(); 9299 assert(gc_stats_l->kind() == GCStats::CMSGCStatsKind, 9300 "Wrong gc statistics type"); 9301 counters->update_counters(gc_stats_l); 9302 } 9303 } 9304 9305 void ASConcurrentMarkSweepGeneration::shrink_by(size_t desired_bytes) { 9306 assert_locked_or_safepoint(Heap_lock); 9307 assert_lock_strong(freelistLock()); 9308 HeapWord* old_end = _cmsSpace->end(); 9309 HeapWord* unallocated_start = _cmsSpace->unallocated_block(); 9310 assert(old_end >= unallocated_start, "Miscalculation of unallocated_start"); 9311 FreeChunk* chunk_at_end = find_chunk_at_end(); 9312 if (chunk_at_end == NULL) { 9313 // No room to shrink 9314 if (PrintGCDetails && Verbose) { 9315 gclog_or_tty->print_cr("No room to shrink: old_end " 9316 PTR_FORMAT " unallocated_start " PTR_FORMAT 9317 " chunk_at_end " PTR_FORMAT, 9318 old_end, unallocated_start, chunk_at_end); 9319 } 9320 return; 9321 } else { 9322 9323 // Find the chunk at the end of the space and determine 9324 // how much it can be shrunk. 9325 size_t shrinkable_size_in_bytes = chunk_at_end->size(); 9326 size_t aligned_shrinkable_size_in_bytes = 9327 align_size_down(shrinkable_size_in_bytes, os::vm_page_size()); 9328 assert(unallocated_start <= (HeapWord*) chunk_at_end->end(), 9329 "Inconsistent chunk at end of space"); 9330 size_t bytes = MIN2(desired_bytes, aligned_shrinkable_size_in_bytes); 9331 size_t word_size_before = heap_word_size(_virtual_space.committed_size()); 9332 9333 // Shrink the underlying space 9334 _virtual_space.shrink_by(bytes); 9335 if (PrintGCDetails && Verbose) { 9336 gclog_or_tty->print_cr("ConcurrentMarkSweepGeneration::shrink_by:" 9337 " desired_bytes " SIZE_FORMAT 9338 " shrinkable_size_in_bytes " SIZE_FORMAT 9339 " aligned_shrinkable_size_in_bytes " SIZE_FORMAT 9340 " bytes " SIZE_FORMAT, 9341 desired_bytes, shrinkable_size_in_bytes, 9342 aligned_shrinkable_size_in_bytes, bytes); 9343 gclog_or_tty->print_cr(" old_end " SIZE_FORMAT 9344 " unallocated_start " SIZE_FORMAT, 9345 old_end, unallocated_start); 9346 } 9347 9348 // If the space did shrink (shrinking is not guaranteed), 9349 // shrink the chunk at the end by the appropriate amount. 9350 if (((HeapWord*)_virtual_space.high()) < old_end) { 9351 size_t new_word_size = 9352 heap_word_size(_virtual_space.committed_size()); 9353 9354 // Have to remove the chunk from the dictionary because it is changing 9355 // size and might be someplace elsewhere in the dictionary. 9356 9357 // Get the chunk at end, shrink it, and put it 9358 // back. 9359 _cmsSpace->removeChunkFromDictionary(chunk_at_end); 9360 size_t word_size_change = word_size_before - new_word_size; 9361 size_t chunk_at_end_old_size = chunk_at_end->size(); 9362 assert(chunk_at_end_old_size >= word_size_change, 9363 "Shrink is too large"); 9364 chunk_at_end->set_size(chunk_at_end_old_size - 9365 word_size_change); 9366 _cmsSpace->freed((HeapWord*) chunk_at_end->end(), 9367 word_size_change); 9368 9369 _cmsSpace->returnChunkToDictionary(chunk_at_end); 9370 9371 MemRegion mr(_cmsSpace->bottom(), new_word_size); 9372 _bts->resize(new_word_size); // resize the block offset shared array 9373 Universe::heap()->barrier_set()->resize_covered_region(mr); 9374 _cmsSpace->assert_locked(); 9375 _cmsSpace->set_end((HeapWord*)_virtual_space.high()); 9376 9377 NOT_PRODUCT(_cmsSpace->dictionary()->verify()); 9378 9379 // update the space and generation capacity counters 9380 if (UsePerfData) { 9381 _space_counters->update_capacity(); 9382 _gen_counters->update_all(); 9383 } 9384 9385 if (Verbose && PrintGCDetails) { 9386 size_t new_mem_size = _virtual_space.committed_size(); 9387 size_t old_mem_size = new_mem_size + bytes; 9388 gclog_or_tty->print_cr("Shrinking %s from " SIZE_FORMAT "K by " SIZE_FORMAT "K to " SIZE_FORMAT "K", 9389 name(), old_mem_size/K, bytes/K, new_mem_size/K); 9390 } 9391 } 9392 9393 assert(_cmsSpace->unallocated_block() <= _cmsSpace->end(), 9394 "Inconsistency at end of space"); 9395 assert(chunk_at_end->end() == (uintptr_t*) _cmsSpace->end(), 9396 "Shrinking is inconsistent"); 9397 return; 9398 } 9399 } 9400 9401 // Transfer some number of overflown objects to usual marking 9402 // stack. Return true if some objects were transferred. 9403 bool MarkRefsIntoAndScanClosure::take_from_overflow_list() { 9404 size_t num = MIN2((size_t)(_mark_stack->capacity() - _mark_stack->length())/4, 9405 (size_t)ParGCDesiredObjsFromOverflowList); 9406 9407 bool res = _collector->take_from_overflow_list(num, _mark_stack); 9408 assert(_collector->overflow_list_is_empty() || res, 9409 "If list is not empty, we should have taken something"); 9410 assert(!res || !_mark_stack->isEmpty(), 9411 "If we took something, it should now be on our stack"); 9412 return res; 9413 } 9414 9415 size_t MarkDeadObjectsClosure::do_blk(HeapWord* addr) { 9416 size_t res = _sp->block_size_no_stall(addr, _collector); 9417 if (_sp->block_is_obj(addr)) { 9418 if (_live_bit_map->isMarked(addr)) { 9419 // It can't have been dead in a previous cycle 9420 guarantee(!_dead_bit_map->isMarked(addr), "No resurrection!"); 9421 } else { 9422 _dead_bit_map->mark(addr); // mark the dead object 9423 } 9424 } 9425 // Could be 0, if the block size could not be computed without stalling. 9426 return res; 9427 } 9428 9429 TraceCMSMemoryManagerStats::TraceCMSMemoryManagerStats(CMSCollector::CollectorState phase, GCCause::Cause cause): TraceMemoryManagerStats() { 9430 9431 switch (phase) { 9432 case CMSCollector::InitialMarking: 9433 initialize(true /* fullGC */ , 9434 cause /* cause of the GC */, 9435 true /* recordGCBeginTime */, 9436 true /* recordPreGCUsage */, 9437 false /* recordPeakUsage */, 9438 false /* recordPostGCusage */, 9439 true /* recordAccumulatedGCTime */, 9440 false /* recordGCEndTime */, 9441 false /* countCollection */ ); 9442 break; 9443 9444 case CMSCollector::FinalMarking: 9445 initialize(true /* fullGC */ , 9446 cause /* cause of the GC */, 9447 false /* recordGCBeginTime */, 9448 false /* recordPreGCUsage */, 9449 false /* recordPeakUsage */, 9450 false /* recordPostGCusage */, 9451 true /* recordAccumulatedGCTime */, 9452 false /* recordGCEndTime */, 9453 false /* countCollection */ ); 9454 break; 9455 9456 case CMSCollector::Sweeping: 9457 initialize(true /* fullGC */ , 9458 cause /* cause of the GC */, 9459 false /* recordGCBeginTime */, 9460 false /* recordPreGCUsage */, 9461 true /* recordPeakUsage */, 9462 true /* recordPostGCusage */, 9463 false /* recordAccumulatedGCTime */, 9464 true /* recordGCEndTime */, 9465 true /* countCollection */ ); 9466 break; 9467 9468 default: 9469 ShouldNotReachHere(); 9470 } 9471 }