1 /* 2 * Copyright (c) 2001, 2013, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "classfile/classLoaderData.hpp" 27 #include "classfile/symbolTable.hpp" 28 #include "classfile/systemDictionary.hpp" 29 #include "code/codeCache.hpp" 30 #include "gc_implementation/concurrentMarkSweep/cmsAdaptiveSizePolicy.hpp" 31 #include "gc_implementation/concurrentMarkSweep/cmsCollectorPolicy.hpp" 32 #include "gc_implementation/concurrentMarkSweep/cmsGCAdaptivePolicyCounters.hpp" 33 #include "gc_implementation/concurrentMarkSweep/cmsOopClosures.inline.hpp" 34 #include "gc_implementation/concurrentMarkSweep/compactibleFreeListSpace.hpp" 35 #include "gc_implementation/concurrentMarkSweep/concurrentMarkSweepGeneration.inline.hpp" 36 #include "gc_implementation/concurrentMarkSweep/concurrentMarkSweepThread.hpp" 37 #include "gc_implementation/concurrentMarkSweep/vmCMSOperations.hpp" 38 #include "gc_implementation/parNew/parNewGeneration.hpp" 39 #include "gc_implementation/shared/collectorCounters.hpp" 40 #include "gc_implementation/shared/gcTimer.hpp" 41 #include "gc_implementation/shared/gcTrace.hpp" 42 #include "gc_implementation/shared/gcTraceTime.hpp" 43 #include "gc_implementation/shared/isGCActiveMark.hpp" 44 #include "gc_interface/collectedHeap.inline.hpp" 45 #include "memory/allocation.hpp" 46 #include "memory/cardTableRS.hpp" 47 #include "memory/collectorPolicy.hpp" 48 #include "memory/gcLocker.inline.hpp" 49 #include "memory/genCollectedHeap.hpp" 50 #include "memory/genMarkSweep.hpp" 51 #include "memory/genOopClosures.inline.hpp" 52 #include "memory/iterator.hpp" 53 #include "memory/padded.hpp" 54 #include "memory/referencePolicy.hpp" 55 #include "memory/resourceArea.hpp" 56 #include "memory/tenuredGeneration.hpp" 57 #include "oops/oop.inline.hpp" 58 #include "prims/jvmtiExport.hpp" 59 #include "runtime/globals_extension.hpp" 60 #include "runtime/handles.inline.hpp" 61 #include "runtime/java.hpp" 62 #include "runtime/vmThread.hpp" 63 #include "services/memoryService.hpp" 64 #include "services/runtimeService.hpp" 65 66 // statics 67 CMSCollector* ConcurrentMarkSweepGeneration::_collector = NULL; 68 bool CMSCollector::_full_gc_requested = false; 69 GCCause::Cause CMSCollector::_full_gc_cause = GCCause::_no_gc; 70 71 ////////////////////////////////////////////////////////////////// 72 // In support of CMS/VM thread synchronization 73 ////////////////////////////////////////////////////////////////// 74 // We split use of the CGC_lock into 2 "levels". 75 // The low-level locking is of the usual CGC_lock monitor. We introduce 76 // a higher level "token" (hereafter "CMS token") built on top of the 77 // low level monitor (hereafter "CGC lock"). 78 // The token-passing protocol gives priority to the VM thread. The 79 // CMS-lock doesn't provide any fairness guarantees, but clients 80 // should ensure that it is only held for very short, bounded 81 // durations. 82 // 83 // When either of the CMS thread or the VM thread is involved in 84 // collection operations during which it does not want the other 85 // thread to interfere, it obtains the CMS token. 86 // 87 // If either thread tries to get the token while the other has 88 // it, that thread waits. However, if the VM thread and CMS thread 89 // both want the token, then the VM thread gets priority while the 90 // CMS thread waits. This ensures, for instance, that the "concurrent" 91 // phases of the CMS thread's work do not block out the VM thread 92 // for long periods of time as the CMS thread continues to hog 93 // the token. (See bug 4616232). 94 // 95 // The baton-passing functions are, however, controlled by the 96 // flags _foregroundGCShouldWait and _foregroundGCIsActive, 97 // and here the low-level CMS lock, not the high level token, 98 // ensures mutual exclusion. 99 // 100 // Two important conditions that we have to satisfy: 101 // 1. if a thread does a low-level wait on the CMS lock, then it 102 // relinquishes the CMS token if it were holding that token 103 // when it acquired the low-level CMS lock. 104 // 2. any low-level notifications on the low-level lock 105 // should only be sent when a thread has relinquished the token. 106 // 107 // In the absence of either property, we'd have potential deadlock. 108 // 109 // We protect each of the CMS (concurrent and sequential) phases 110 // with the CMS _token_, not the CMS _lock_. 111 // 112 // The only code protected by CMS lock is the token acquisition code 113 // itself, see ConcurrentMarkSweepThread::[de]synchronize(), and the 114 // baton-passing code. 115 // 116 // Unfortunately, i couldn't come up with a good abstraction to factor and 117 // hide the naked CGC_lock manipulation in the baton-passing code 118 // further below. That's something we should try to do. Also, the proof 119 // of correctness of this 2-level locking scheme is far from obvious, 120 // and potentially quite slippery. We have an uneasy supsicion, for instance, 121 // that there may be a theoretical possibility of delay/starvation in the 122 // low-level lock/wait/notify scheme used for the baton-passing because of 123 // potential intereference with the priority scheme embodied in the 124 // CMS-token-passing protocol. See related comments at a CGC_lock->wait() 125 // invocation further below and marked with "XXX 20011219YSR". 126 // Indeed, as we note elsewhere, this may become yet more slippery 127 // in the presence of multiple CMS and/or multiple VM threads. XXX 128 129 class CMSTokenSync: public StackObj { 130 private: 131 bool _is_cms_thread; 132 public: 133 CMSTokenSync(bool is_cms_thread): 134 _is_cms_thread(is_cms_thread) { 135 assert(is_cms_thread == Thread::current()->is_ConcurrentGC_thread(), 136 "Incorrect argument to constructor"); 137 ConcurrentMarkSweepThread::synchronize(_is_cms_thread); 138 } 139 140 ~CMSTokenSync() { 141 assert(_is_cms_thread ? 142 ConcurrentMarkSweepThread::cms_thread_has_cms_token() : 143 ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 144 "Incorrect state"); 145 ConcurrentMarkSweepThread::desynchronize(_is_cms_thread); 146 } 147 }; 148 149 // Convenience class that does a CMSTokenSync, and then acquires 150 // upto three locks. 151 class CMSTokenSyncWithLocks: public CMSTokenSync { 152 private: 153 // Note: locks are acquired in textual declaration order 154 // and released in the opposite order 155 MutexLockerEx _locker1, _locker2, _locker3; 156 public: 157 CMSTokenSyncWithLocks(bool is_cms_thread, Mutex* mutex1, 158 Mutex* mutex2 = NULL, Mutex* mutex3 = NULL): 159 CMSTokenSync(is_cms_thread), 160 _locker1(mutex1, Mutex::_no_safepoint_check_flag), 161 _locker2(mutex2, Mutex::_no_safepoint_check_flag), 162 _locker3(mutex3, Mutex::_no_safepoint_check_flag) 163 { } 164 }; 165 166 167 // Wrapper class to temporarily disable icms during a foreground cms collection. 168 class ICMSDisabler: public StackObj { 169 public: 170 // The ctor disables icms and wakes up the thread so it notices the change; 171 // the dtor re-enables icms. Note that the CMSCollector methods will check 172 // CMSIncrementalMode. 173 ICMSDisabler() { CMSCollector::disable_icms(); CMSCollector::start_icms(); } 174 ~ICMSDisabler() { CMSCollector::enable_icms(); } 175 }; 176 177 ////////////////////////////////////////////////////////////////// 178 // Concurrent Mark-Sweep Generation ///////////////////////////// 179 ////////////////////////////////////////////////////////////////// 180 181 NOT_PRODUCT(CompactibleFreeListSpace* debug_cms_space;) 182 183 // This struct contains per-thread things necessary to support parallel 184 // young-gen collection. 185 class CMSParGCThreadState: public CHeapObj<mtGC> { 186 public: 187 CFLS_LAB lab; 188 PromotionInfo promo; 189 190 // Constructor. 191 CMSParGCThreadState(CompactibleFreeListSpace* cfls) : lab(cfls) { 192 promo.setSpace(cfls); 193 } 194 }; 195 196 ConcurrentMarkSweepGeneration::ConcurrentMarkSweepGeneration( 197 ReservedSpace rs, size_t initial_byte_size, int level, 198 CardTableRS* ct, bool use_adaptive_freelists, 199 FreeBlockDictionary<FreeChunk>::DictionaryChoice dictionaryChoice) : 200 CardGeneration(rs, initial_byte_size, level, ct), 201 _dilatation_factor(((double)MinChunkSize)/((double)(CollectedHeap::min_fill_size()))), 202 _debug_collection_type(Concurrent_collection_type), 203 _did_compact(false) 204 { 205 HeapWord* bottom = (HeapWord*) _virtual_space.low(); 206 HeapWord* end = (HeapWord*) _virtual_space.high(); 207 208 _direct_allocated_words = 0; 209 NOT_PRODUCT( 210 _numObjectsPromoted = 0; 211 _numWordsPromoted = 0; 212 _numObjectsAllocated = 0; 213 _numWordsAllocated = 0; 214 ) 215 216 _cmsSpace = new CompactibleFreeListSpace(_bts, MemRegion(bottom, end), 217 use_adaptive_freelists, 218 dictionaryChoice); 219 NOT_PRODUCT(debug_cms_space = _cmsSpace;) 220 if (_cmsSpace == NULL) { 221 vm_exit_during_initialization( 222 "CompactibleFreeListSpace allocation failure"); 223 } 224 _cmsSpace->_gen = this; 225 226 _gc_stats = new CMSGCStats(); 227 228 // Verify the assumption that FreeChunk::_prev and OopDesc::_klass 229 // offsets match. The ability to tell free chunks from objects 230 // depends on this property. 231 debug_only( 232 FreeChunk* junk = NULL; 233 assert(UseCompressedKlassPointers || 234 junk->prev_addr() == (void*)(oop(junk)->klass_addr()), 235 "Offset of FreeChunk::_prev within FreeChunk must match" 236 " that of OopDesc::_klass within OopDesc"); 237 ) 238 if (CollectedHeap::use_parallel_gc_threads()) { 239 typedef CMSParGCThreadState* CMSParGCThreadStatePtr; 240 _par_gc_thread_states = 241 NEW_C_HEAP_ARRAY(CMSParGCThreadStatePtr, ParallelGCThreads, mtGC); 242 if (_par_gc_thread_states == NULL) { 243 vm_exit_during_initialization("Could not allocate par gc structs"); 244 } 245 for (uint i = 0; i < ParallelGCThreads; i++) { 246 _par_gc_thread_states[i] = new CMSParGCThreadState(cmsSpace()); 247 if (_par_gc_thread_states[i] == NULL) { 248 vm_exit_during_initialization("Could not allocate par gc structs"); 249 } 250 } 251 } else { 252 _par_gc_thread_states = NULL; 253 } 254 _incremental_collection_failed = false; 255 // The "dilatation_factor" is the expansion that can occur on 256 // account of the fact that the minimum object size in the CMS 257 // generation may be larger than that in, say, a contiguous young 258 // generation. 259 // Ideally, in the calculation below, we'd compute the dilatation 260 // factor as: MinChunkSize/(promoting_gen's min object size) 261 // Since we do not have such a general query interface for the 262 // promoting generation, we'll instead just use the mimimum 263 // object size (which today is a header's worth of space); 264 // note that all arithmetic is in units of HeapWords. 265 assert(MinChunkSize >= CollectedHeap::min_fill_size(), "just checking"); 266 assert(_dilatation_factor >= 1.0, "from previous assert"); 267 } 268 269 270 // The field "_initiating_occupancy" represents the occupancy percentage 271 // at which we trigger a new collection cycle. Unless explicitly specified 272 // via CMSInitiatingOccupancyFraction (argument "io" below), it 273 // is calculated by: 274 // 275 // Let "f" be MinHeapFreeRatio in 276 // 277 // _intiating_occupancy = 100-f + 278 // f * (CMSTriggerRatio/100) 279 // where CMSTriggerRatio is the argument "tr" below. 280 // 281 // That is, if we assume the heap is at its desired maximum occupancy at the 282 // end of a collection, we let CMSTriggerRatio of the (purported) free 283 // space be allocated before initiating a new collection cycle. 284 // 285 void ConcurrentMarkSweepGeneration::init_initiating_occupancy(intx io, uintx tr) { 286 assert(io <= 100 && tr <= 100, "Check the arguments"); 287 if (io >= 0) { 288 _initiating_occupancy = (double)io / 100.0; 289 } else { 290 _initiating_occupancy = ((100 - MinHeapFreeRatio) + 291 (double)(tr * MinHeapFreeRatio) / 100.0) 292 / 100.0; 293 } 294 } 295 296 void ConcurrentMarkSweepGeneration::ref_processor_init() { 297 assert(collector() != NULL, "no collector"); 298 collector()->ref_processor_init(); 299 } 300 301 void CMSCollector::ref_processor_init() { 302 if (_ref_processor == NULL) { 303 // Allocate and initialize a reference processor 304 _ref_processor = 305 new ReferenceProcessor(_span, // span 306 (ParallelGCThreads > 1) && ParallelRefProcEnabled, // mt processing 307 (int) ParallelGCThreads, // mt processing degree 308 _cmsGen->refs_discovery_is_mt(), // mt discovery 309 (int) MAX2(ConcGCThreads, ParallelGCThreads), // mt discovery degree 310 _cmsGen->refs_discovery_is_atomic(), // discovery is not atomic 311 &_is_alive_closure, // closure for liveness info 312 false); // next field updates do not need write barrier 313 // Initialize the _ref_processor field of CMSGen 314 _cmsGen->set_ref_processor(_ref_processor); 315 316 } 317 } 318 319 CMSAdaptiveSizePolicy* CMSCollector::size_policy() { 320 GenCollectedHeap* gch = GenCollectedHeap::heap(); 321 assert(gch->kind() == CollectedHeap::GenCollectedHeap, 322 "Wrong type of heap"); 323 CMSAdaptiveSizePolicy* sp = (CMSAdaptiveSizePolicy*) 324 gch->gen_policy()->size_policy(); 325 assert(sp->is_gc_cms_adaptive_size_policy(), 326 "Wrong type of size policy"); 327 return sp; 328 } 329 330 CMSGCAdaptivePolicyCounters* CMSCollector::gc_adaptive_policy_counters() { 331 CMSGCAdaptivePolicyCounters* results = 332 (CMSGCAdaptivePolicyCounters*) collector_policy()->counters(); 333 assert( 334 results->kind() == GCPolicyCounters::CMSGCAdaptivePolicyCountersKind, 335 "Wrong gc policy counter kind"); 336 return results; 337 } 338 339 340 void ConcurrentMarkSweepGeneration::initialize_performance_counters() { 341 342 const char* gen_name = "old"; 343 344 // Generation Counters - generation 1, 1 subspace 345 _gen_counters = new GenerationCounters(gen_name, 1, 1, &_virtual_space); 346 347 _space_counters = new GSpaceCounters(gen_name, 0, 348 _virtual_space.reserved_size(), 349 this, _gen_counters); 350 } 351 352 CMSStats::CMSStats(ConcurrentMarkSweepGeneration* cms_gen, unsigned int alpha): 353 _cms_gen(cms_gen) 354 { 355 assert(alpha <= 100, "bad value"); 356 _saved_alpha = alpha; 357 358 // Initialize the alphas to the bootstrap value of 100. 359 _gc0_alpha = _cms_alpha = 100; 360 361 _cms_begin_time.update(); 362 _cms_end_time.update(); 363 364 _gc0_duration = 0.0; 365 _gc0_period = 0.0; 366 _gc0_promoted = 0; 367 368 _cms_duration = 0.0; 369 _cms_period = 0.0; 370 _cms_allocated = 0; 371 372 _cms_used_at_gc0_begin = 0; 373 _cms_used_at_gc0_end = 0; 374 _allow_duty_cycle_reduction = false; 375 _valid_bits = 0; 376 _icms_duty_cycle = CMSIncrementalDutyCycle; 377 } 378 379 double CMSStats::cms_free_adjustment_factor(size_t free) const { 380 // TBD: CR 6909490 381 return 1.0; 382 } 383 384 void CMSStats::adjust_cms_free_adjustment_factor(bool fail, size_t free) { 385 } 386 387 // If promotion failure handling is on use 388 // the padded average size of the promotion for each 389 // young generation collection. 390 double CMSStats::time_until_cms_gen_full() const { 391 size_t cms_free = _cms_gen->cmsSpace()->free(); 392 GenCollectedHeap* gch = GenCollectedHeap::heap(); 393 size_t expected_promotion = MIN2(gch->get_gen(0)->capacity(), 394 (size_t) _cms_gen->gc_stats()->avg_promoted()->padded_average()); 395 if (cms_free > expected_promotion) { 396 // Start a cms collection if there isn't enough space to promote 397 // for the next minor collection. Use the padded average as 398 // a safety factor. 399 cms_free -= expected_promotion; 400 401 // Adjust by the safety factor. 402 double cms_free_dbl = (double)cms_free; 403 double cms_adjustment = (100.0 - CMSIncrementalSafetyFactor)/100.0; 404 // Apply a further correction factor which tries to adjust 405 // for recent occurance of concurrent mode failures. 406 cms_adjustment = cms_adjustment * cms_free_adjustment_factor(cms_free); 407 cms_free_dbl = cms_free_dbl * cms_adjustment; 408 409 if (PrintGCDetails && Verbose) { 410 gclog_or_tty->print_cr("CMSStats::time_until_cms_gen_full: cms_free " 411 SIZE_FORMAT " expected_promotion " SIZE_FORMAT, 412 cms_free, expected_promotion); 413 gclog_or_tty->print_cr(" cms_free_dbl %f cms_consumption_rate %f", 414 cms_free_dbl, cms_consumption_rate() + 1.0); 415 } 416 // Add 1 in case the consumption rate goes to zero. 417 return cms_free_dbl / (cms_consumption_rate() + 1.0); 418 } 419 return 0.0; 420 } 421 422 // Compare the duration of the cms collection to the 423 // time remaining before the cms generation is empty. 424 // Note that the time from the start of the cms collection 425 // to the start of the cms sweep (less than the total 426 // duration of the cms collection) can be used. This 427 // has been tried and some applications experienced 428 // promotion failures early in execution. This was 429 // possibly because the averages were not accurate 430 // enough at the beginning. 431 double CMSStats::time_until_cms_start() const { 432 // We add "gc0_period" to the "work" calculation 433 // below because this query is done (mostly) at the 434 // end of a scavenge, so we need to conservatively 435 // account for that much possible delay 436 // in the query so as to avoid concurrent mode failures 437 // due to starting the collection just a wee bit too 438 // late. 439 double work = cms_duration() + gc0_period(); 440 double deadline = time_until_cms_gen_full(); 441 // If a concurrent mode failure occurred recently, we want to be 442 // more conservative and halve our expected time_until_cms_gen_full() 443 if (work > deadline) { 444 if (Verbose && PrintGCDetails) { 445 gclog_or_tty->print( 446 " CMSCollector: collect because of anticipated promotion " 447 "before full %3.7f + %3.7f > %3.7f ", cms_duration(), 448 gc0_period(), time_until_cms_gen_full()); 449 } 450 return 0.0; 451 } 452 return work - deadline; 453 } 454 455 // Return a duty cycle based on old_duty_cycle and new_duty_cycle, limiting the 456 // amount of change to prevent wild oscillation. 457 unsigned int CMSStats::icms_damped_duty_cycle(unsigned int old_duty_cycle, 458 unsigned int new_duty_cycle) { 459 assert(old_duty_cycle <= 100, "bad input value"); 460 assert(new_duty_cycle <= 100, "bad input value"); 461 462 // Note: use subtraction with caution since it may underflow (values are 463 // unsigned). Addition is safe since we're in the range 0-100. 464 unsigned int damped_duty_cycle = new_duty_cycle; 465 if (new_duty_cycle < old_duty_cycle) { 466 const unsigned int largest_delta = MAX2(old_duty_cycle / 4, 5U); 467 if (new_duty_cycle + largest_delta < old_duty_cycle) { 468 damped_duty_cycle = old_duty_cycle - largest_delta; 469 } 470 } else if (new_duty_cycle > old_duty_cycle) { 471 const unsigned int largest_delta = MAX2(old_duty_cycle / 4, 15U); 472 if (new_duty_cycle > old_duty_cycle + largest_delta) { 473 damped_duty_cycle = MIN2(old_duty_cycle + largest_delta, 100U); 474 } 475 } 476 assert(damped_duty_cycle <= 100, "invalid duty cycle computed"); 477 478 if (CMSTraceIncrementalPacing) { 479 gclog_or_tty->print(" [icms_damped_duty_cycle(%d,%d) = %d] ", 480 old_duty_cycle, new_duty_cycle, damped_duty_cycle); 481 } 482 return damped_duty_cycle; 483 } 484 485 unsigned int CMSStats::icms_update_duty_cycle_impl() { 486 assert(CMSIncrementalPacing && valid(), 487 "should be handled in icms_update_duty_cycle()"); 488 489 double cms_time_so_far = cms_timer().seconds(); 490 double scaled_duration = cms_duration_per_mb() * _cms_used_at_gc0_end / M; 491 double scaled_duration_remaining = fabsd(scaled_duration - cms_time_so_far); 492 493 // Avoid division by 0. 494 double time_until_full = MAX2(time_until_cms_gen_full(), 0.01); 495 double duty_cycle_dbl = 100.0 * scaled_duration_remaining / time_until_full; 496 497 unsigned int new_duty_cycle = MIN2((unsigned int)duty_cycle_dbl, 100U); 498 if (new_duty_cycle > _icms_duty_cycle) { 499 // Avoid very small duty cycles (1 or 2); 0 is allowed. 500 if (new_duty_cycle > 2) { 501 _icms_duty_cycle = icms_damped_duty_cycle(_icms_duty_cycle, 502 new_duty_cycle); 503 } 504 } else if (_allow_duty_cycle_reduction) { 505 // The duty cycle is reduced only once per cms cycle (see record_cms_end()). 506 new_duty_cycle = icms_damped_duty_cycle(_icms_duty_cycle, new_duty_cycle); 507 // Respect the minimum duty cycle. 508 unsigned int min_duty_cycle = (unsigned int)CMSIncrementalDutyCycleMin; 509 _icms_duty_cycle = MAX2(new_duty_cycle, min_duty_cycle); 510 } 511 512 if (PrintGCDetails || CMSTraceIncrementalPacing) { 513 gclog_or_tty->print(" icms_dc=%d ", _icms_duty_cycle); 514 } 515 516 _allow_duty_cycle_reduction = false; 517 return _icms_duty_cycle; 518 } 519 520 #ifndef PRODUCT 521 void CMSStats::print_on(outputStream *st) const { 522 st->print(" gc0_alpha=%d,cms_alpha=%d", _gc0_alpha, _cms_alpha); 523 st->print(",gc0_dur=%g,gc0_per=%g,gc0_promo=" SIZE_FORMAT, 524 gc0_duration(), gc0_period(), gc0_promoted()); 525 st->print(",cms_dur=%g,cms_dur_per_mb=%g,cms_per=%g,cms_alloc=" SIZE_FORMAT, 526 cms_duration(), cms_duration_per_mb(), 527 cms_period(), cms_allocated()); 528 st->print(",cms_since_beg=%g,cms_since_end=%g", 529 cms_time_since_begin(), cms_time_since_end()); 530 st->print(",cms_used_beg=" SIZE_FORMAT ",cms_used_end=" SIZE_FORMAT, 531 _cms_used_at_gc0_begin, _cms_used_at_gc0_end); 532 if (CMSIncrementalMode) { 533 st->print(",dc=%d", icms_duty_cycle()); 534 } 535 536 if (valid()) { 537 st->print(",promo_rate=%g,cms_alloc_rate=%g", 538 promotion_rate(), cms_allocation_rate()); 539 st->print(",cms_consumption_rate=%g,time_until_full=%g", 540 cms_consumption_rate(), time_until_cms_gen_full()); 541 } 542 st->print(" "); 543 } 544 #endif // #ifndef PRODUCT 545 546 CMSCollector::CollectorState CMSCollector::_collectorState = 547 CMSCollector::Idling; 548 bool CMSCollector::_foregroundGCIsActive = false; 549 bool CMSCollector::_foregroundGCShouldWait = false; 550 551 CMSCollector::CMSCollector(ConcurrentMarkSweepGeneration* cmsGen, 552 CardTableRS* ct, 553 ConcurrentMarkSweepPolicy* cp): 554 _cmsGen(cmsGen), 555 _ct(ct), 556 _ref_processor(NULL), // will be set later 557 _conc_workers(NULL), // may be set later 558 _abort_preclean(false), 559 _start_sampling(false), 560 _between_prologue_and_epilogue(false), 561 _markBitMap(0, Mutex::leaf + 1, "CMS_markBitMap_lock"), 562 _modUnionTable((CardTableModRefBS::card_shift - LogHeapWordSize), 563 -1 /* lock-free */, "No_lock" /* dummy */), 564 _modUnionClosure(&_modUnionTable), 565 _modUnionClosurePar(&_modUnionTable), 566 // Adjust my span to cover old (cms) gen 567 _span(cmsGen->reserved()), 568 // Construct the is_alive_closure with _span & markBitMap 569 _is_alive_closure(_span, &_markBitMap), 570 _restart_addr(NULL), 571 _overflow_list(NULL), 572 _stats(cmsGen), 573 _eden_chunk_lock(new Mutex(Mutex::leaf + 1, "CMS_eden_chunk_lock", true)), 574 _eden_chunk_array(NULL), // may be set in ctor body 575 _eden_chunk_capacity(0), // -- ditto -- 576 _eden_chunk_index(0), // -- ditto -- 577 _survivor_plab_array(NULL), // -- ditto -- 578 _survivor_chunk_array(NULL), // -- ditto -- 579 _survivor_chunk_capacity(0), // -- ditto -- 580 _survivor_chunk_index(0), // -- ditto -- 581 _ser_pmc_preclean_ovflw(0), 582 _ser_kac_preclean_ovflw(0), 583 _ser_pmc_remark_ovflw(0), 584 _par_pmc_remark_ovflw(0), 585 _ser_kac_ovflw(0), 586 _par_kac_ovflw(0), 587 #ifndef PRODUCT 588 _num_par_pushes(0), 589 #endif 590 _collection_count_start(0), 591 _verifying(false), 592 _icms_start_limit(NULL), 593 _icms_stop_limit(NULL), 594 _verification_mark_bm(0, Mutex::leaf + 1, "CMS_verification_mark_bm_lock"), 595 _completed_initialization(false), 596 _collector_policy(cp), 597 _should_unload_classes(false), 598 _concurrent_cycles_since_last_unload(0), 599 _roots_scanning_options(0), 600 _inter_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding), 601 _intra_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding), 602 _gc_tracer_cm(new (ResourceObj::C_HEAP, mtGC) CMSTracer()), 603 _gc_timer_cm(new (ResourceObj::C_HEAP, mtGC) ConcurrentGCTimer()), 604 _cms_start_registered(false) 605 { 606 if (ExplicitGCInvokesConcurrentAndUnloadsClasses) { 607 ExplicitGCInvokesConcurrent = true; 608 } 609 // Now expand the span and allocate the collection support structures 610 // (MUT, marking bit map etc.) to cover both generations subject to 611 // collection. 612 613 // For use by dirty card to oop closures. 614 _cmsGen->cmsSpace()->set_collector(this); 615 616 // Allocate MUT and marking bit map 617 { 618 MutexLockerEx x(_markBitMap.lock(), Mutex::_no_safepoint_check_flag); 619 if (!_markBitMap.allocate(_span)) { 620 warning("Failed to allocate CMS Bit Map"); 621 return; 622 } 623 assert(_markBitMap.covers(_span), "_markBitMap inconsistency?"); 624 } 625 { 626 _modUnionTable.allocate(_span); 627 assert(_modUnionTable.covers(_span), "_modUnionTable inconsistency?"); 628 } 629 630 if (!_markStack.allocate(MarkStackSize)) { 631 warning("Failed to allocate CMS Marking Stack"); 632 return; 633 } 634 635 // Support for multi-threaded concurrent phases 636 if (CMSConcurrentMTEnabled) { 637 if (FLAG_IS_DEFAULT(ConcGCThreads)) { 638 // just for now 639 FLAG_SET_DEFAULT(ConcGCThreads, (ParallelGCThreads + 3)/4); 640 } 641 if (ConcGCThreads > 1) { 642 _conc_workers = new YieldingFlexibleWorkGang("Parallel CMS Threads", 643 ConcGCThreads, true); 644 if (_conc_workers == NULL) { 645 warning("GC/CMS: _conc_workers allocation failure: " 646 "forcing -CMSConcurrentMTEnabled"); 647 CMSConcurrentMTEnabled = false; 648 } else { 649 _conc_workers->initialize_workers(); 650 } 651 } else { 652 CMSConcurrentMTEnabled = false; 653 } 654 } 655 if (!CMSConcurrentMTEnabled) { 656 ConcGCThreads = 0; 657 } else { 658 // Turn off CMSCleanOnEnter optimization temporarily for 659 // the MT case where it's not fixed yet; see 6178663. 660 CMSCleanOnEnter = false; 661 } 662 assert((_conc_workers != NULL) == (ConcGCThreads > 1), 663 "Inconsistency"); 664 665 // Parallel task queues; these are shared for the 666 // concurrent and stop-world phases of CMS, but 667 // are not shared with parallel scavenge (ParNew). 668 { 669 uint i; 670 uint num_queues = (uint) MAX2(ParallelGCThreads, ConcGCThreads); 671 672 if ((CMSParallelRemarkEnabled || CMSConcurrentMTEnabled 673 || ParallelRefProcEnabled) 674 && num_queues > 0) { 675 _task_queues = new OopTaskQueueSet(num_queues); 676 if (_task_queues == NULL) { 677 warning("task_queues allocation failure."); 678 return; 679 } 680 _hash_seed = NEW_C_HEAP_ARRAY(int, num_queues, mtGC); 681 if (_hash_seed == NULL) { 682 warning("_hash_seed array allocation failure"); 683 return; 684 } 685 686 typedef Padded<OopTaskQueue> PaddedOopTaskQueue; 687 for (i = 0; i < num_queues; i++) { 688 PaddedOopTaskQueue *q = new PaddedOopTaskQueue(); 689 if (q == NULL) { 690 warning("work_queue allocation failure."); 691 return; 692 } 693 _task_queues->register_queue(i, q); 694 } 695 for (i = 0; i < num_queues; i++) { 696 _task_queues->queue(i)->initialize(); 697 _hash_seed[i] = 17; // copied from ParNew 698 } 699 } 700 } 701 702 _cmsGen ->init_initiating_occupancy(CMSInitiatingOccupancyFraction, CMSTriggerRatio); 703 704 // Clip CMSBootstrapOccupancy between 0 and 100. 705 _bootstrap_occupancy = ((double)CMSBootstrapOccupancy)/(double)100; 706 707 _full_gcs_since_conc_gc = 0; 708 709 // Now tell CMS generations the identity of their collector 710 ConcurrentMarkSweepGeneration::set_collector(this); 711 712 // Create & start a CMS thread for this CMS collector 713 _cmsThread = ConcurrentMarkSweepThread::start(this); 714 assert(cmsThread() != NULL, "CMS Thread should have been created"); 715 assert(cmsThread()->collector() == this, 716 "CMS Thread should refer to this gen"); 717 assert(CGC_lock != NULL, "Where's the CGC_lock?"); 718 719 // Support for parallelizing young gen rescan 720 GenCollectedHeap* gch = GenCollectedHeap::heap(); 721 _young_gen = gch->prev_gen(_cmsGen); 722 if (gch->supports_inline_contig_alloc()) { 723 _top_addr = gch->top_addr(); 724 _end_addr = gch->end_addr(); 725 assert(_young_gen != NULL, "no _young_gen"); 726 _eden_chunk_index = 0; 727 _eden_chunk_capacity = (_young_gen->max_capacity()+CMSSamplingGrain)/CMSSamplingGrain; 728 _eden_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, _eden_chunk_capacity, mtGC); 729 if (_eden_chunk_array == NULL) { 730 _eden_chunk_capacity = 0; 731 warning("GC/CMS: _eden_chunk_array allocation failure"); 732 } 733 } 734 assert(_eden_chunk_array != NULL || _eden_chunk_capacity == 0, "Error"); 735 736 // Support for parallelizing survivor space rescan 737 if ((CMSParallelRemarkEnabled && CMSParallelSurvivorRemarkEnabled) || CMSParallelInitialMarkEnabled) { 738 const size_t max_plab_samples = 739 ((DefNewGeneration*)_young_gen)->max_survivor_size()/MinTLABSize; 740 741 _survivor_plab_array = NEW_C_HEAP_ARRAY(ChunkArray, ParallelGCThreads, mtGC); 742 _survivor_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, 2*max_plab_samples, mtGC); 743 _cursor = NEW_C_HEAP_ARRAY(size_t, ParallelGCThreads, mtGC); 744 if (_survivor_plab_array == NULL || _survivor_chunk_array == NULL 745 || _cursor == NULL) { 746 warning("Failed to allocate survivor plab/chunk array"); 747 if (_survivor_plab_array != NULL) { 748 FREE_C_HEAP_ARRAY(ChunkArray, _survivor_plab_array, mtGC); 749 _survivor_plab_array = NULL; 750 } 751 if (_survivor_chunk_array != NULL) { 752 FREE_C_HEAP_ARRAY(HeapWord*, _survivor_chunk_array, mtGC); 753 _survivor_chunk_array = NULL; 754 } 755 if (_cursor != NULL) { 756 FREE_C_HEAP_ARRAY(size_t, _cursor, mtGC); 757 _cursor = NULL; 758 } 759 } else { 760 _survivor_chunk_capacity = 2*max_plab_samples; 761 for (uint i = 0; i < ParallelGCThreads; i++) { 762 HeapWord** vec = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples, mtGC); 763 if (vec == NULL) { 764 warning("Failed to allocate survivor plab array"); 765 for (int j = i; j > 0; j--) { 766 FREE_C_HEAP_ARRAY(HeapWord*, _survivor_plab_array[j-1].array(), mtGC); 767 } 768 FREE_C_HEAP_ARRAY(ChunkArray, _survivor_plab_array, mtGC); 769 FREE_C_HEAP_ARRAY(HeapWord*, _survivor_chunk_array, mtGC); 770 _survivor_plab_array = NULL; 771 _survivor_chunk_array = NULL; 772 _survivor_chunk_capacity = 0; 773 break; 774 } else { 775 ChunkArray* cur = 776 ::new (&_survivor_plab_array[i]) ChunkArray(vec, 777 max_plab_samples); 778 assert(cur->end() == 0, "Should be 0"); 779 assert(cur->array() == vec, "Should be vec"); 780 assert(cur->capacity() == max_plab_samples, "Error"); 781 } 782 } 783 } 784 } 785 assert( ( _survivor_plab_array != NULL 786 && _survivor_chunk_array != NULL) 787 || ( _survivor_chunk_capacity == 0 788 && _survivor_chunk_index == 0), 789 "Error"); 790 791 // Choose what strong roots should be scanned depending on verification options. 792 // If CMSClassUnloadingEnabled is on, root scanning option is set during 793 // setup_cms_unloading_and_verification_state() on each CMS cycle. 794 if (!CMSClassUnloadingEnabled) { 795 // If class unloading is disabled we want to include all classes into the root set. 796 add_root_scanning_option(SharedHeap::SO_AllClasses); 797 } 798 799 NOT_PRODUCT(_overflow_counter = CMSMarkStackOverflowInterval;) 800 _gc_counters = new CollectorCounters("CMS", 1); 801 _completed_initialization = true; 802 _inter_sweep_timer.start(); // start of time 803 } 804 805 const char* ConcurrentMarkSweepGeneration::name() const { 806 return "concurrent mark-sweep generation"; 807 } 808 void ConcurrentMarkSweepGeneration::update_counters() { 809 if (UsePerfData) { 810 _space_counters->update_all(); 811 _gen_counters->update_all(); 812 } 813 } 814 815 // this is an optimized version of update_counters(). it takes the 816 // used value as a parameter rather than computing it. 817 // 818 void ConcurrentMarkSweepGeneration::update_counters(size_t used) { 819 if (UsePerfData) { 820 _space_counters->update_used(used); 821 _space_counters->update_capacity(); 822 _gen_counters->update_all(); 823 } 824 } 825 826 void ConcurrentMarkSweepGeneration::print() const { 827 Generation::print(); 828 cmsSpace()->print(); 829 } 830 831 #ifndef PRODUCT 832 void ConcurrentMarkSweepGeneration::print_statistics() { 833 cmsSpace()->printFLCensus(0); 834 } 835 #endif 836 837 void ConcurrentMarkSweepGeneration::printOccupancy(const char *s) { 838 GenCollectedHeap* gch = GenCollectedHeap::heap(); 839 if (PrintGCDetails) { 840 if (Verbose) { 841 gclog_or_tty->print("[%d %s-%s: "SIZE_FORMAT"("SIZE_FORMAT")]", 842 level(), short_name(), s, used(), capacity()); 843 } else { 844 gclog_or_tty->print("[%d %s-%s: "SIZE_FORMAT"K("SIZE_FORMAT"K)]", 845 level(), short_name(), s, used() / K, capacity() / K); 846 } 847 } 848 if (Verbose) { 849 gclog_or_tty->print(" "SIZE_FORMAT"("SIZE_FORMAT")", 850 gch->used(), gch->capacity()); 851 } else { 852 gclog_or_tty->print(" "SIZE_FORMAT"K("SIZE_FORMAT"K)", 853 gch->used() / K, gch->capacity() / K); 854 } 855 } 856 857 size_t 858 ConcurrentMarkSweepGeneration::contiguous_available() const { 859 // dld proposes an improvement in precision here. If the committed 860 // part of the space ends in a free block we should add that to 861 // uncommitted size in the calculation below. Will make this 862 // change later, staying with the approximation below for the 863 // time being. -- ysr. 864 return MAX2(_virtual_space.uncommitted_size(), unsafe_max_alloc_nogc()); 865 } 866 867 size_t 868 ConcurrentMarkSweepGeneration::unsafe_max_alloc_nogc() const { 869 return _cmsSpace->max_alloc_in_words() * HeapWordSize; 870 } 871 872 size_t ConcurrentMarkSweepGeneration::max_available() const { 873 return free() + _virtual_space.uncommitted_size(); 874 } 875 876 bool ConcurrentMarkSweepGeneration::promotion_attempt_is_safe(size_t max_promotion_in_bytes) const { 877 size_t available = max_available(); 878 size_t av_promo = (size_t)gc_stats()->avg_promoted()->padded_average(); 879 bool res = (available >= av_promo) || (available >= max_promotion_in_bytes); 880 if (Verbose && PrintGCDetails) { 881 gclog_or_tty->print_cr( 882 "CMS: promo attempt is%s safe: available("SIZE_FORMAT") %s av_promo("SIZE_FORMAT")," 883 "max_promo("SIZE_FORMAT")", 884 res? "":" not", available, res? ">=":"<", 885 av_promo, max_promotion_in_bytes); 886 } 887 return res; 888 } 889 890 // At a promotion failure dump information on block layout in heap 891 // (cms old generation). 892 void ConcurrentMarkSweepGeneration::promotion_failure_occurred() { 893 if (CMSDumpAtPromotionFailure) { 894 cmsSpace()->dump_at_safepoint_with_locks(collector(), gclog_or_tty); 895 } 896 } 897 898 CompactibleSpace* 899 ConcurrentMarkSweepGeneration::first_compaction_space() const { 900 return _cmsSpace; 901 } 902 903 void ConcurrentMarkSweepGeneration::reset_after_compaction() { 904 // Clear the promotion information. These pointers can be adjusted 905 // along with all the other pointers into the heap but 906 // compaction is expected to be a rare event with 907 // a heap using cms so don't do it without seeing the need. 908 if (CollectedHeap::use_parallel_gc_threads()) { 909 for (uint i = 0; i < ParallelGCThreads; i++) { 910 _par_gc_thread_states[i]->promo.reset(); 911 } 912 } 913 } 914 915 void ConcurrentMarkSweepGeneration::space_iterate(SpaceClosure* blk, bool usedOnly) { 916 blk->do_space(_cmsSpace); 917 } 918 919 void ConcurrentMarkSweepGeneration::compute_new_size() { 920 assert_locked_or_safepoint(Heap_lock); 921 922 // If incremental collection failed, we just want to expand 923 // to the limit. 924 if (incremental_collection_failed()) { 925 clear_incremental_collection_failed(); 926 grow_to_reserved(); 927 return; 928 } 929 930 // The heap has been compacted but not reset yet. 931 // Any metric such as free() or used() will be incorrect. 932 933 CardGeneration::compute_new_size(); 934 935 // Reset again after a possible resizing 936 if (did_compact()) { 937 cmsSpace()->reset_after_compaction(); 938 } 939 } 940 941 void ConcurrentMarkSweepGeneration::compute_new_size_free_list() { 942 assert_locked_or_safepoint(Heap_lock); 943 944 // If incremental collection failed, we just want to expand 945 // to the limit. 946 if (incremental_collection_failed()) { 947 clear_incremental_collection_failed(); 948 grow_to_reserved(); 949 return; 950 } 951 952 double free_percentage = ((double) free()) / capacity(); 953 double desired_free_percentage = (double) MinHeapFreeRatio / 100; 954 double maximum_free_percentage = (double) MaxHeapFreeRatio / 100; 955 956 // compute expansion delta needed for reaching desired free percentage 957 if (free_percentage < desired_free_percentage) { 958 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 959 assert(desired_capacity >= capacity(), "invalid expansion size"); 960 size_t expand_bytes = MAX2(desired_capacity - capacity(), MinHeapDeltaBytes); 961 if (PrintGCDetails && Verbose) { 962 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 963 gclog_or_tty->print_cr("\nFrom compute_new_size: "); 964 gclog_or_tty->print_cr(" Free fraction %f", free_percentage); 965 gclog_or_tty->print_cr(" Desired free fraction %f", 966 desired_free_percentage); 967 gclog_or_tty->print_cr(" Maximum free fraction %f", 968 maximum_free_percentage); 969 gclog_or_tty->print_cr(" Capactiy "SIZE_FORMAT, capacity()/1000); 970 gclog_or_tty->print_cr(" Desired capacity "SIZE_FORMAT, 971 desired_capacity/1000); 972 int prev_level = level() - 1; 973 if (prev_level >= 0) { 974 size_t prev_size = 0; 975 GenCollectedHeap* gch = GenCollectedHeap::heap(); 976 Generation* prev_gen = gch->_gens[prev_level]; 977 prev_size = prev_gen->capacity(); 978 gclog_or_tty->print_cr(" Younger gen size "SIZE_FORMAT, 979 prev_size/1000); 980 } 981 gclog_or_tty->print_cr(" unsafe_max_alloc_nogc "SIZE_FORMAT, 982 unsafe_max_alloc_nogc()/1000); 983 gclog_or_tty->print_cr(" contiguous available "SIZE_FORMAT, 984 contiguous_available()/1000); 985 gclog_or_tty->print_cr(" Expand by "SIZE_FORMAT" (bytes)", 986 expand_bytes); 987 } 988 // safe if expansion fails 989 expand(expand_bytes, 0, CMSExpansionCause::_satisfy_free_ratio); 990 if (PrintGCDetails && Verbose) { 991 gclog_or_tty->print_cr(" Expanded free fraction %f", 992 ((double) free()) / capacity()); 993 } 994 } else { 995 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 996 assert(desired_capacity <= capacity(), "invalid expansion size"); 997 size_t shrink_bytes = capacity() - desired_capacity; 998 // Don't shrink unless the delta is greater than the minimum shrink we want 999 if (shrink_bytes >= MinHeapDeltaBytes) { 1000 shrink_free_list_by(shrink_bytes); 1001 } 1002 } 1003 } 1004 1005 Mutex* ConcurrentMarkSweepGeneration::freelistLock() const { 1006 return cmsSpace()->freelistLock(); 1007 } 1008 1009 HeapWord* ConcurrentMarkSweepGeneration::allocate(size_t size, 1010 bool tlab) { 1011 CMSSynchronousYieldRequest yr; 1012 MutexLockerEx x(freelistLock(), 1013 Mutex::_no_safepoint_check_flag); 1014 return have_lock_and_allocate(size, tlab); 1015 } 1016 1017 HeapWord* ConcurrentMarkSweepGeneration::have_lock_and_allocate(size_t size, 1018 bool tlab /* ignored */) { 1019 assert_lock_strong(freelistLock()); 1020 size_t adjustedSize = CompactibleFreeListSpace::adjustObjectSize(size); 1021 HeapWord* res = cmsSpace()->allocate(adjustedSize); 1022 // Allocate the object live (grey) if the background collector has 1023 // started marking. This is necessary because the marker may 1024 // have passed this address and consequently this object will 1025 // not otherwise be greyed and would be incorrectly swept up. 1026 // Note that if this object contains references, the writing 1027 // of those references will dirty the card containing this object 1028 // allowing the object to be blackened (and its references scanned) 1029 // either during a preclean phase or at the final checkpoint. 1030 if (res != NULL) { 1031 // We may block here with an uninitialized object with 1032 // its mark-bit or P-bits not yet set. Such objects need 1033 // to be safely navigable by block_start(). 1034 assert(oop(res)->klass_or_null() == NULL, "Object should be uninitialized here."); 1035 assert(!((FreeChunk*)res)->is_free(), "Error, block will look free but show wrong size"); 1036 collector()->direct_allocated(res, adjustedSize); 1037 _direct_allocated_words += adjustedSize; 1038 // allocation counters 1039 NOT_PRODUCT( 1040 _numObjectsAllocated++; 1041 _numWordsAllocated += (int)adjustedSize; 1042 ) 1043 } 1044 return res; 1045 } 1046 1047 // In the case of direct allocation by mutators in a generation that 1048 // is being concurrently collected, the object must be allocated 1049 // live (grey) if the background collector has started marking. 1050 // This is necessary because the marker may 1051 // have passed this address and consequently this object will 1052 // not otherwise be greyed and would be incorrectly swept up. 1053 // Note that if this object contains references, the writing 1054 // of those references will dirty the card containing this object 1055 // allowing the object to be blackened (and its references scanned) 1056 // either during a preclean phase or at the final checkpoint. 1057 void CMSCollector::direct_allocated(HeapWord* start, size_t size) { 1058 assert(_markBitMap.covers(start, size), "Out of bounds"); 1059 if (_collectorState >= Marking) { 1060 MutexLockerEx y(_markBitMap.lock(), 1061 Mutex::_no_safepoint_check_flag); 1062 // [see comments preceding SweepClosure::do_blk() below for details] 1063 // 1064 // Can the P-bits be deleted now? JJJ 1065 // 1066 // 1. need to mark the object as live so it isn't collected 1067 // 2. need to mark the 2nd bit to indicate the object may be uninitialized 1068 // 3. need to mark the end of the object so marking, precleaning or sweeping 1069 // can skip over uninitialized or unparsable objects. An allocated 1070 // object is considered uninitialized for our purposes as long as 1071 // its klass word is NULL. All old gen objects are parsable 1072 // as soon as they are initialized.) 1073 _markBitMap.mark(start); // object is live 1074 _markBitMap.mark(start + 1); // object is potentially uninitialized? 1075 _markBitMap.mark(start + size - 1); 1076 // mark end of object 1077 } 1078 // check that oop looks uninitialized 1079 assert(oop(start)->klass_or_null() == NULL, "_klass should be NULL"); 1080 } 1081 1082 void CMSCollector::promoted(bool par, HeapWord* start, 1083 bool is_obj_array, size_t obj_size) { 1084 assert(_markBitMap.covers(start), "Out of bounds"); 1085 // See comment in direct_allocated() about when objects should 1086 // be allocated live. 1087 if (_collectorState >= Marking) { 1088 // we already hold the marking bit map lock, taken in 1089 // the prologue 1090 if (par) { 1091 _markBitMap.par_mark(start); 1092 } else { 1093 _markBitMap.mark(start); 1094 } 1095 // We don't need to mark the object as uninitialized (as 1096 // in direct_allocated above) because this is being done with the 1097 // world stopped and the object will be initialized by the 1098 // time the marking, precleaning or sweeping get to look at it. 1099 // But see the code for copying objects into the CMS generation, 1100 // where we need to ensure that concurrent readers of the 1101 // block offset table are able to safely navigate a block that 1102 // is in flux from being free to being allocated (and in 1103 // transition while being copied into) and subsequently 1104 // becoming a bona-fide object when the copy/promotion is complete. 1105 assert(SafepointSynchronize::is_at_safepoint(), 1106 "expect promotion only at safepoints"); 1107 1108 if (_collectorState < Sweeping) { 1109 // Mark the appropriate cards in the modUnionTable, so that 1110 // this object gets scanned before the sweep. If this is 1111 // not done, CMS generation references in the object might 1112 // not get marked. 1113 // For the case of arrays, which are otherwise precisely 1114 // marked, we need to dirty the entire array, not just its head. 1115 if (is_obj_array) { 1116 // The [par_]mark_range() method expects mr.end() below to 1117 // be aligned to the granularity of a bit's representation 1118 // in the heap. In the case of the MUT below, that's a 1119 // card size. 1120 MemRegion mr(start, 1121 (HeapWord*)round_to((intptr_t)(start + obj_size), 1122 CardTableModRefBS::card_size /* bytes */)); 1123 if (par) { 1124 _modUnionTable.par_mark_range(mr); 1125 } else { 1126 _modUnionTable.mark_range(mr); 1127 } 1128 } else { // not an obj array; we can just mark the head 1129 if (par) { 1130 _modUnionTable.par_mark(start); 1131 } else { 1132 _modUnionTable.mark(start); 1133 } 1134 } 1135 } 1136 } 1137 } 1138 1139 static inline size_t percent_of_space(Space* space, HeapWord* addr) 1140 { 1141 size_t delta = pointer_delta(addr, space->bottom()); 1142 return (size_t)(delta * 100.0 / (space->capacity() / HeapWordSize)); 1143 } 1144 1145 void CMSCollector::icms_update_allocation_limits() 1146 { 1147 Generation* gen0 = GenCollectedHeap::heap()->get_gen(0); 1148 EdenSpace* eden = gen0->as_DefNewGeneration()->eden(); 1149 1150 const unsigned int duty_cycle = stats().icms_update_duty_cycle(); 1151 if (CMSTraceIncrementalPacing) { 1152 stats().print(); 1153 } 1154 1155 assert(duty_cycle <= 100, "invalid duty cycle"); 1156 if (duty_cycle != 0) { 1157 // The duty_cycle is a percentage between 0 and 100; convert to words and 1158 // then compute the offset from the endpoints of the space. 1159 size_t free_words = eden->free() / HeapWordSize; 1160 double free_words_dbl = (double)free_words; 1161 size_t duty_cycle_words = (size_t)(free_words_dbl * duty_cycle / 100.0); 1162 size_t offset_words = (free_words - duty_cycle_words) / 2; 1163 1164 _icms_start_limit = eden->top() + offset_words; 1165 _icms_stop_limit = eden->end() - offset_words; 1166 1167 // The limits may be adjusted (shifted to the right) by 1168 // CMSIncrementalOffset, to allow the application more mutator time after a 1169 // young gen gc (when all mutators were stopped) and before CMS starts and 1170 // takes away one or more cpus. 1171 if (CMSIncrementalOffset != 0) { 1172 double adjustment_dbl = free_words_dbl * CMSIncrementalOffset / 100.0; 1173 size_t adjustment = (size_t)adjustment_dbl; 1174 HeapWord* tmp_stop = _icms_stop_limit + adjustment; 1175 if (tmp_stop > _icms_stop_limit && tmp_stop < eden->end()) { 1176 _icms_start_limit += adjustment; 1177 _icms_stop_limit = tmp_stop; 1178 } 1179 } 1180 } 1181 if (duty_cycle == 0 || (_icms_start_limit == _icms_stop_limit)) { 1182 _icms_start_limit = _icms_stop_limit = eden->end(); 1183 } 1184 1185 // Install the new start limit. 1186 eden->set_soft_end(_icms_start_limit); 1187 1188 if (CMSTraceIncrementalMode) { 1189 gclog_or_tty->print(" icms alloc limits: " 1190 PTR_FORMAT "," PTR_FORMAT 1191 " (" SIZE_FORMAT "%%," SIZE_FORMAT "%%) ", 1192 _icms_start_limit, _icms_stop_limit, 1193 percent_of_space(eden, _icms_start_limit), 1194 percent_of_space(eden, _icms_stop_limit)); 1195 if (Verbose) { 1196 gclog_or_tty->print("eden: "); 1197 eden->print_on(gclog_or_tty); 1198 } 1199 } 1200 } 1201 1202 // Any changes here should try to maintain the invariant 1203 // that if this method is called with _icms_start_limit 1204 // and _icms_stop_limit both NULL, then it should return NULL 1205 // and not notify the icms thread. 1206 HeapWord* 1207 CMSCollector::allocation_limit_reached(Space* space, HeapWord* top, 1208 size_t word_size) 1209 { 1210 // A start_limit equal to end() means the duty cycle is 0, so treat that as a 1211 // nop. 1212 if (CMSIncrementalMode && _icms_start_limit != space->end()) { 1213 if (top <= _icms_start_limit) { 1214 if (CMSTraceIncrementalMode) { 1215 space->print_on(gclog_or_tty); 1216 gclog_or_tty->stamp(); 1217 gclog_or_tty->print_cr(" start limit top=" PTR_FORMAT 1218 ", new limit=" PTR_FORMAT 1219 " (" SIZE_FORMAT "%%)", 1220 top, _icms_stop_limit, 1221 percent_of_space(space, _icms_stop_limit)); 1222 } 1223 ConcurrentMarkSweepThread::start_icms(); 1224 assert(top < _icms_stop_limit, "Tautology"); 1225 if (word_size < pointer_delta(_icms_stop_limit, top)) { 1226 return _icms_stop_limit; 1227 } 1228 1229 // The allocation will cross both the _start and _stop limits, so do the 1230 // stop notification also and return end(). 1231 if (CMSTraceIncrementalMode) { 1232 space->print_on(gclog_or_tty); 1233 gclog_or_tty->stamp(); 1234 gclog_or_tty->print_cr(" +stop limit top=" PTR_FORMAT 1235 ", new limit=" PTR_FORMAT 1236 " (" SIZE_FORMAT "%%)", 1237 top, space->end(), 1238 percent_of_space(space, space->end())); 1239 } 1240 ConcurrentMarkSweepThread::stop_icms(); 1241 return space->end(); 1242 } 1243 1244 if (top <= _icms_stop_limit) { 1245 if (CMSTraceIncrementalMode) { 1246 space->print_on(gclog_or_tty); 1247 gclog_or_tty->stamp(); 1248 gclog_or_tty->print_cr(" stop limit top=" PTR_FORMAT 1249 ", new limit=" PTR_FORMAT 1250 " (" SIZE_FORMAT "%%)", 1251 top, space->end(), 1252 percent_of_space(space, space->end())); 1253 } 1254 ConcurrentMarkSweepThread::stop_icms(); 1255 return space->end(); 1256 } 1257 1258 if (CMSTraceIncrementalMode) { 1259 space->print_on(gclog_or_tty); 1260 gclog_or_tty->stamp(); 1261 gclog_or_tty->print_cr(" end limit top=" PTR_FORMAT 1262 ", new limit=" PTR_FORMAT, 1263 top, NULL); 1264 } 1265 } 1266 1267 return NULL; 1268 } 1269 1270 oop ConcurrentMarkSweepGeneration::promote(oop obj, size_t obj_size) { 1271 assert(obj_size == (size_t)obj->size(), "bad obj_size passed in"); 1272 // allocate, copy and if necessary update promoinfo -- 1273 // delegate to underlying space. 1274 assert_lock_strong(freelistLock()); 1275 1276 #ifndef PRODUCT 1277 if (Universe::heap()->promotion_should_fail()) { 1278 return NULL; 1279 } 1280 #endif // #ifndef PRODUCT 1281 1282 oop res = _cmsSpace->promote(obj, obj_size); 1283 if (res == NULL) { 1284 // expand and retry 1285 size_t s = _cmsSpace->expansionSpaceRequired(obj_size); // HeapWords 1286 expand(s*HeapWordSize, MinHeapDeltaBytes, 1287 CMSExpansionCause::_satisfy_promotion); 1288 // Since there's currently no next generation, we don't try to promote 1289 // into a more senior generation. 1290 assert(next_gen() == NULL, "assumption, based upon which no attempt " 1291 "is made to pass on a possibly failing " 1292 "promotion to next generation"); 1293 res = _cmsSpace->promote(obj, obj_size); 1294 } 1295 if (res != NULL) { 1296 // See comment in allocate() about when objects should 1297 // be allocated live. 1298 assert(obj->is_oop(), "Will dereference klass pointer below"); 1299 collector()->promoted(false, // Not parallel 1300 (HeapWord*)res, obj->is_objArray(), obj_size); 1301 // promotion counters 1302 NOT_PRODUCT( 1303 _numObjectsPromoted++; 1304 _numWordsPromoted += 1305 (int)(CompactibleFreeListSpace::adjustObjectSize(obj->size())); 1306 ) 1307 } 1308 return res; 1309 } 1310 1311 1312 HeapWord* 1313 ConcurrentMarkSweepGeneration::allocation_limit_reached(Space* space, 1314 HeapWord* top, 1315 size_t word_sz) 1316 { 1317 return collector()->allocation_limit_reached(space, top, word_sz); 1318 } 1319 1320 // IMPORTANT: Notes on object size recognition in CMS. 1321 // --------------------------------------------------- 1322 // A block of storage in the CMS generation is always in 1323 // one of three states. A free block (FREE), an allocated 1324 // object (OBJECT) whose size() method reports the correct size, 1325 // and an intermediate state (TRANSIENT) in which its size cannot 1326 // be accurately determined. 1327 // STATE IDENTIFICATION: (32 bit and 64 bit w/o COOPS) 1328 // ----------------------------------------------------- 1329 // FREE: klass_word & 1 == 1; mark_word holds block size 1330 // 1331 // OBJECT: klass_word installed; klass_word != 0 && klass_word & 1 == 0; 1332 // obj->size() computes correct size 1333 // 1334 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT 1335 // 1336 // STATE IDENTIFICATION: (64 bit+COOPS) 1337 // ------------------------------------ 1338 // FREE: mark_word & CMS_FREE_BIT == 1; mark_word & ~CMS_FREE_BIT gives block_size 1339 // 1340 // OBJECT: klass_word installed; klass_word != 0; 1341 // obj->size() computes correct size 1342 // 1343 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT 1344 // 1345 // 1346 // STATE TRANSITION DIAGRAM 1347 // 1348 // mut / parnew mut / parnew 1349 // FREE --------------------> TRANSIENT ---------------------> OBJECT --| 1350 // ^ | 1351 // |------------------------ DEAD <------------------------------------| 1352 // sweep mut 1353 // 1354 // While a block is in TRANSIENT state its size cannot be determined 1355 // so readers will either need to come back later or stall until 1356 // the size can be determined. Note that for the case of direct 1357 // allocation, P-bits, when available, may be used to determine the 1358 // size of an object that may not yet have been initialized. 1359 1360 // Things to support parallel young-gen collection. 1361 oop 1362 ConcurrentMarkSweepGeneration::par_promote(int thread_num, 1363 oop old, markOop m, 1364 size_t word_sz) { 1365 #ifndef PRODUCT 1366 if (Universe::heap()->promotion_should_fail()) { 1367 return NULL; 1368 } 1369 #endif // #ifndef PRODUCT 1370 1371 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1372 PromotionInfo* promoInfo = &ps->promo; 1373 // if we are tracking promotions, then first ensure space for 1374 // promotion (including spooling space for saving header if necessary). 1375 // then allocate and copy, then track promoted info if needed. 1376 // When tracking (see PromotionInfo::track()), the mark word may 1377 // be displaced and in this case restoration of the mark word 1378 // occurs in the (oop_since_save_marks_)iterate phase. 1379 if (promoInfo->tracking() && !promoInfo->ensure_spooling_space()) { 1380 // Out of space for allocating spooling buffers; 1381 // try expanding and allocating spooling buffers. 1382 if (!expand_and_ensure_spooling_space(promoInfo)) { 1383 return NULL; 1384 } 1385 } 1386 assert(promoInfo->has_spooling_space(), "Control point invariant"); 1387 const size_t alloc_sz = CompactibleFreeListSpace::adjustObjectSize(word_sz); 1388 HeapWord* obj_ptr = ps->lab.alloc(alloc_sz); 1389 if (obj_ptr == NULL) { 1390 obj_ptr = expand_and_par_lab_allocate(ps, alloc_sz); 1391 if (obj_ptr == NULL) { 1392 return NULL; 1393 } 1394 } 1395 oop obj = oop(obj_ptr); 1396 OrderAccess::storestore(); 1397 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1398 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1399 // IMPORTANT: See note on object initialization for CMS above. 1400 // Otherwise, copy the object. Here we must be careful to insert the 1401 // klass pointer last, since this marks the block as an allocated object. 1402 // Except with compressed oops it's the mark word. 1403 HeapWord* old_ptr = (HeapWord*)old; 1404 // Restore the mark word copied above. 1405 obj->set_mark(m); 1406 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1407 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1408 OrderAccess::storestore(); 1409 1410 if (UseCompressedKlassPointers) { 1411 // Copy gap missed by (aligned) header size calculation below 1412 obj->set_klass_gap(old->klass_gap()); 1413 } 1414 if (word_sz > (size_t)oopDesc::header_size()) { 1415 Copy::aligned_disjoint_words(old_ptr + oopDesc::header_size(), 1416 obj_ptr + oopDesc::header_size(), 1417 word_sz - oopDesc::header_size()); 1418 } 1419 1420 // Now we can track the promoted object, if necessary. We take care 1421 // to delay the transition from uninitialized to full object 1422 // (i.e., insertion of klass pointer) until after, so that it 1423 // atomically becomes a promoted object. 1424 if (promoInfo->tracking()) { 1425 promoInfo->track((PromotedObject*)obj, old->klass()); 1426 } 1427 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1428 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1429 assert(old->is_oop(), "Will use and dereference old klass ptr below"); 1430 1431 // Finally, install the klass pointer (this should be volatile). 1432 OrderAccess::storestore(); 1433 obj->set_klass(old->klass()); 1434 // We should now be able to calculate the right size for this object 1435 assert(obj->is_oop() && obj->size() == (int)word_sz, "Error, incorrect size computed for promoted object"); 1436 1437 collector()->promoted(true, // parallel 1438 obj_ptr, old->is_objArray(), word_sz); 1439 1440 NOT_PRODUCT( 1441 Atomic::inc_ptr(&_numObjectsPromoted); 1442 Atomic::add_ptr(alloc_sz, &_numWordsPromoted); 1443 ) 1444 1445 return obj; 1446 } 1447 1448 void 1449 ConcurrentMarkSweepGeneration:: 1450 par_promote_alloc_undo(int thread_num, 1451 HeapWord* obj, size_t word_sz) { 1452 // CMS does not support promotion undo. 1453 ShouldNotReachHere(); 1454 } 1455 1456 void 1457 ConcurrentMarkSweepGeneration:: 1458 par_promote_alloc_done(int thread_num) { 1459 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1460 ps->lab.retire(thread_num); 1461 } 1462 1463 void 1464 ConcurrentMarkSweepGeneration:: 1465 par_oop_since_save_marks_iterate_done(int thread_num) { 1466 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1467 ParScanWithoutBarrierClosure* dummy_cl = NULL; 1468 ps->promo.promoted_oops_iterate_nv(dummy_cl); 1469 } 1470 1471 bool ConcurrentMarkSweepGeneration::should_collect(bool full, 1472 size_t size, 1473 bool tlab) 1474 { 1475 // We allow a STW collection only if a full 1476 // collection was requested. 1477 return full || should_allocate(size, tlab); // FIX ME !!! 1478 // This and promotion failure handling are connected at the 1479 // hip and should be fixed by untying them. 1480 } 1481 1482 bool CMSCollector::shouldConcurrentCollect() { 1483 if (_full_gc_requested) { 1484 if (Verbose && PrintGCDetails) { 1485 gclog_or_tty->print_cr("CMSCollector: collect because of explicit " 1486 " gc request (or gc_locker)"); 1487 } 1488 return true; 1489 } 1490 1491 // For debugging purposes, change the type of collection. 1492 // If the rotation is not on the concurrent collection 1493 // type, don't start a concurrent collection. 1494 NOT_PRODUCT( 1495 if (RotateCMSCollectionTypes && 1496 (_cmsGen->debug_collection_type() != 1497 ConcurrentMarkSweepGeneration::Concurrent_collection_type)) { 1498 assert(_cmsGen->debug_collection_type() != 1499 ConcurrentMarkSweepGeneration::Unknown_collection_type, 1500 "Bad cms collection type"); 1501 return false; 1502 } 1503 ) 1504 1505 FreelistLocker x(this); 1506 // ------------------------------------------------------------------ 1507 // Print out lots of information which affects the initiation of 1508 // a collection. 1509 if (PrintCMSInitiationStatistics && stats().valid()) { 1510 gclog_or_tty->print("CMSCollector shouldConcurrentCollect: "); 1511 gclog_or_tty->stamp(); 1512 gclog_or_tty->print_cr(""); 1513 stats().print_on(gclog_or_tty); 1514 gclog_or_tty->print_cr("time_until_cms_gen_full %3.7f", 1515 stats().time_until_cms_gen_full()); 1516 gclog_or_tty->print_cr("free="SIZE_FORMAT, _cmsGen->free()); 1517 gclog_or_tty->print_cr("contiguous_available="SIZE_FORMAT, 1518 _cmsGen->contiguous_available()); 1519 gclog_or_tty->print_cr("promotion_rate=%g", stats().promotion_rate()); 1520 gclog_or_tty->print_cr("cms_allocation_rate=%g", stats().cms_allocation_rate()); 1521 gclog_or_tty->print_cr("occupancy=%3.7f", _cmsGen->occupancy()); 1522 gclog_or_tty->print_cr("initiatingOccupancy=%3.7f", _cmsGen->initiating_occupancy()); 1523 gclog_or_tty->print_cr("metadata initialized %d", 1524 MetaspaceGC::should_concurrent_collect()); 1525 } 1526 // ------------------------------------------------------------------ 1527 1528 // If the estimated time to complete a cms collection (cms_duration()) 1529 // is less than the estimated time remaining until the cms generation 1530 // is full, start a collection. 1531 if (!UseCMSInitiatingOccupancyOnly) { 1532 if (stats().valid()) { 1533 if (stats().time_until_cms_start() == 0.0) { 1534 return true; 1535 } 1536 } else { 1537 // We want to conservatively collect somewhat early in order 1538 // to try and "bootstrap" our CMS/promotion statistics; 1539 // this branch will not fire after the first successful CMS 1540 // collection because the stats should then be valid. 1541 if (_cmsGen->occupancy() >= _bootstrap_occupancy) { 1542 if (Verbose && PrintGCDetails) { 1543 gclog_or_tty->print_cr( 1544 " CMSCollector: collect for bootstrapping statistics:" 1545 " occupancy = %f, boot occupancy = %f", _cmsGen->occupancy(), 1546 _bootstrap_occupancy); 1547 } 1548 return true; 1549 } 1550 } 1551 } 1552 1553 // Otherwise, we start a collection cycle if 1554 // old gen want a collection cycle started. Each may use 1555 // an appropriate criterion for making this decision. 1556 // XXX We need to make sure that the gen expansion 1557 // criterion dovetails well with this. XXX NEED TO FIX THIS 1558 if (_cmsGen->should_concurrent_collect()) { 1559 if (Verbose && PrintGCDetails) { 1560 gclog_or_tty->print_cr("CMS old gen initiated"); 1561 } 1562 return true; 1563 } 1564 1565 // We start a collection if we believe an incremental collection may fail; 1566 // this is not likely to be productive in practice because it's probably too 1567 // late anyway. 1568 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1569 assert(gch->collector_policy()->is_two_generation_policy(), 1570 "You may want to check the correctness of the following"); 1571 if (gch->incremental_collection_will_fail(true /* consult_young */)) { 1572 if (Verbose && PrintGCDetails) { 1573 gclog_or_tty->print("CMSCollector: collect because incremental collection will fail "); 1574 } 1575 return true; 1576 } 1577 1578 if (MetaspaceGC::should_concurrent_collect()) { 1579 if (Verbose && PrintGCDetails) { 1580 gclog_or_tty->print("CMSCollector: collect for metadata allocation "); 1581 } 1582 return true; 1583 } 1584 1585 return false; 1586 } 1587 1588 void CMSCollector::set_did_compact(bool v) { _cmsGen->set_did_compact(v); } 1589 1590 // Clear _expansion_cause fields of constituent generations 1591 void CMSCollector::clear_expansion_cause() { 1592 _cmsGen->clear_expansion_cause(); 1593 } 1594 1595 // We should be conservative in starting a collection cycle. To 1596 // start too eagerly runs the risk of collecting too often in the 1597 // extreme. To collect too rarely falls back on full collections, 1598 // which works, even if not optimum in terms of concurrent work. 1599 // As a work around for too eagerly collecting, use the flag 1600 // UseCMSInitiatingOccupancyOnly. This also has the advantage of 1601 // giving the user an easily understandable way of controlling the 1602 // collections. 1603 // We want to start a new collection cycle if any of the following 1604 // conditions hold: 1605 // . our current occupancy exceeds the configured initiating occupancy 1606 // for this generation, or 1607 // . we recently needed to expand this space and have not, since that 1608 // expansion, done a collection of this generation, or 1609 // . the underlying space believes that it may be a good idea to initiate 1610 // a concurrent collection (this may be based on criteria such as the 1611 // following: the space uses linear allocation and linear allocation is 1612 // going to fail, or there is believed to be excessive fragmentation in 1613 // the generation, etc... or ... 1614 // [.(currently done by CMSCollector::shouldConcurrentCollect() only for 1615 // the case of the old generation; see CR 6543076): 1616 // we may be approaching a point at which allocation requests may fail because 1617 // we will be out of sufficient free space given allocation rate estimates.] 1618 bool ConcurrentMarkSweepGeneration::should_concurrent_collect() const { 1619 1620 assert_lock_strong(freelistLock()); 1621 if (occupancy() > initiating_occupancy()) { 1622 if (PrintGCDetails && Verbose) { 1623 gclog_or_tty->print(" %s: collect because of occupancy %f / %f ", 1624 short_name(), occupancy(), initiating_occupancy()); 1625 } 1626 return true; 1627 } 1628 if (UseCMSInitiatingOccupancyOnly) { 1629 return false; 1630 } 1631 if (expansion_cause() == CMSExpansionCause::_satisfy_allocation) { 1632 if (PrintGCDetails && Verbose) { 1633 gclog_or_tty->print(" %s: collect because expanded for allocation ", 1634 short_name()); 1635 } 1636 return true; 1637 } 1638 if (_cmsSpace->should_concurrent_collect()) { 1639 if (PrintGCDetails && Verbose) { 1640 gclog_or_tty->print(" %s: collect because cmsSpace says so ", 1641 short_name()); 1642 } 1643 return true; 1644 } 1645 return false; 1646 } 1647 1648 void ConcurrentMarkSweepGeneration::collect(bool full, 1649 bool clear_all_soft_refs, 1650 size_t size, 1651 bool tlab) 1652 { 1653 collector()->collect(full, clear_all_soft_refs, size, tlab); 1654 } 1655 1656 void CMSCollector::collect(bool full, 1657 bool clear_all_soft_refs, 1658 size_t size, 1659 bool tlab) 1660 { 1661 if (!UseCMSCollectionPassing && _collectorState > Idling) { 1662 // For debugging purposes skip the collection if the state 1663 // is not currently idle 1664 if (TraceCMSState) { 1665 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " skipped full:%d CMS state %d", 1666 Thread::current(), full, _collectorState); 1667 } 1668 return; 1669 } 1670 1671 // The following "if" branch is present for defensive reasons. 1672 // In the current uses of this interface, it can be replaced with: 1673 // assert(!GC_locker.is_active(), "Can't be called otherwise"); 1674 // But I am not placing that assert here to allow future 1675 // generality in invoking this interface. 1676 if (GC_locker::is_active()) { 1677 // A consistency test for GC_locker 1678 assert(GC_locker::needs_gc(), "Should have been set already"); 1679 // Skip this foreground collection, instead 1680 // expanding the heap if necessary. 1681 // Need the free list locks for the call to free() in compute_new_size() 1682 compute_new_size(); 1683 return; 1684 } 1685 acquire_control_and_collect(full, clear_all_soft_refs); 1686 _full_gcs_since_conc_gc++; 1687 } 1688 1689 void CMSCollector::request_full_gc(unsigned int full_gc_count, GCCause::Cause cause) { 1690 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1691 unsigned int gc_count = gch->total_full_collections(); 1692 if (gc_count == full_gc_count) { 1693 MutexLockerEx y(CGC_lock, Mutex::_no_safepoint_check_flag); 1694 _full_gc_requested = true; 1695 _full_gc_cause = cause; 1696 CGC_lock->notify(); // nudge CMS thread 1697 } else { 1698 assert(gc_count > full_gc_count, "Error: causal loop"); 1699 } 1700 } 1701 1702 bool CMSCollector::is_external_interruption() { 1703 GCCause::Cause cause = GenCollectedHeap::heap()->gc_cause(); 1704 return GCCause::is_user_requested_gc(cause) || 1705 GCCause::is_serviceability_requested_gc(cause); 1706 } 1707 1708 void CMSCollector::report_concurrent_mode_interruption() { 1709 if (is_external_interruption()) { 1710 if (PrintGCDetails) { 1711 gclog_or_tty->print(" (concurrent mode interrupted)"); 1712 } 1713 } else { 1714 if (PrintGCDetails) { 1715 gclog_or_tty->print(" (concurrent mode failure)"); 1716 } 1717 _gc_tracer_cm->report_concurrent_mode_failure(); 1718 } 1719 } 1720 1721 1722 // The foreground and background collectors need to coordinate in order 1723 // to make sure that they do not mutually interfere with CMS collections. 1724 // When a background collection is active, 1725 // the foreground collector may need to take over (preempt) and 1726 // synchronously complete an ongoing collection. Depending on the 1727 // frequency of the background collections and the heap usage 1728 // of the application, this preemption can be seldom or frequent. 1729 // There are only certain 1730 // points in the background collection that the "collection-baton" 1731 // can be passed to the foreground collector. 1732 // 1733 // The foreground collector will wait for the baton before 1734 // starting any part of the collection. The foreground collector 1735 // will only wait at one location. 1736 // 1737 // The background collector will yield the baton before starting a new 1738 // phase of the collection (e.g., before initial marking, marking from roots, 1739 // precleaning, final re-mark, sweep etc.) This is normally done at the head 1740 // of the loop which switches the phases. The background collector does some 1741 // of the phases (initial mark, final re-mark) with the world stopped. 1742 // Because of locking involved in stopping the world, 1743 // the foreground collector should not block waiting for the background 1744 // collector when it is doing a stop-the-world phase. The background 1745 // collector will yield the baton at an additional point just before 1746 // it enters a stop-the-world phase. Once the world is stopped, the 1747 // background collector checks the phase of the collection. If the 1748 // phase has not changed, it proceeds with the collection. If the 1749 // phase has changed, it skips that phase of the collection. See 1750 // the comments on the use of the Heap_lock in collect_in_background(). 1751 // 1752 // Variable used in baton passing. 1753 // _foregroundGCIsActive - Set to true by the foreground collector when 1754 // it wants the baton. The foreground clears it when it has finished 1755 // the collection. 1756 // _foregroundGCShouldWait - Set to true by the background collector 1757 // when it is running. The foreground collector waits while 1758 // _foregroundGCShouldWait is true. 1759 // CGC_lock - monitor used to protect access to the above variables 1760 // and to notify the foreground and background collectors. 1761 // _collectorState - current state of the CMS collection. 1762 // 1763 // The foreground collector 1764 // acquires the CGC_lock 1765 // sets _foregroundGCIsActive 1766 // waits on the CGC_lock for _foregroundGCShouldWait to be false 1767 // various locks acquired in preparation for the collection 1768 // are released so as not to block the background collector 1769 // that is in the midst of a collection 1770 // proceeds with the collection 1771 // clears _foregroundGCIsActive 1772 // returns 1773 // 1774 // The background collector in a loop iterating on the phases of the 1775 // collection 1776 // acquires the CGC_lock 1777 // sets _foregroundGCShouldWait 1778 // if _foregroundGCIsActive is set 1779 // clears _foregroundGCShouldWait, notifies _CGC_lock 1780 // waits on _CGC_lock for _foregroundGCIsActive to become false 1781 // and exits the loop. 1782 // otherwise 1783 // proceed with that phase of the collection 1784 // if the phase is a stop-the-world phase, 1785 // yield the baton once more just before enqueueing 1786 // the stop-world CMS operation (executed by the VM thread). 1787 // returns after all phases of the collection are done 1788 // 1789 1790 void CMSCollector::acquire_control_and_collect(bool full, 1791 bool clear_all_soft_refs) { 1792 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint"); 1793 assert(!Thread::current()->is_ConcurrentGC_thread(), 1794 "shouldn't try to acquire control from self!"); 1795 1796 // Start the protocol for acquiring control of the 1797 // collection from the background collector (aka CMS thread). 1798 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 1799 "VM thread should have CMS token"); 1800 // Remember the possibly interrupted state of an ongoing 1801 // concurrent collection 1802 CollectorState first_state = _collectorState; 1803 1804 // Signal to a possibly ongoing concurrent collection that 1805 // we want to do a foreground collection. 1806 _foregroundGCIsActive = true; 1807 1808 // Disable incremental mode during a foreground collection. 1809 ICMSDisabler icms_disabler; 1810 1811 // release locks and wait for a notify from the background collector 1812 // releasing the locks in only necessary for phases which 1813 // do yields to improve the granularity of the collection. 1814 assert_lock_strong(bitMapLock()); 1815 // We need to lock the Free list lock for the space that we are 1816 // currently collecting. 1817 assert(haveFreelistLocks(), "Must be holding free list locks"); 1818 bitMapLock()->unlock(); 1819 releaseFreelistLocks(); 1820 { 1821 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1822 if (_foregroundGCShouldWait) { 1823 // We are going to be waiting for action for the CMS thread; 1824 // it had better not be gone (for instance at shutdown)! 1825 assert(ConcurrentMarkSweepThread::cmst() != NULL, 1826 "CMS thread must be running"); 1827 // Wait here until the background collector gives us the go-ahead 1828 ConcurrentMarkSweepThread::clear_CMS_flag( 1829 ConcurrentMarkSweepThread::CMS_vm_has_token); // release token 1830 // Get a possibly blocked CMS thread going: 1831 // Note that we set _foregroundGCIsActive true above, 1832 // without protection of the CGC_lock. 1833 CGC_lock->notify(); 1834 assert(!ConcurrentMarkSweepThread::vm_thread_wants_cms_token(), 1835 "Possible deadlock"); 1836 while (_foregroundGCShouldWait) { 1837 // wait for notification 1838 CGC_lock->wait(Mutex::_no_safepoint_check_flag); 1839 // Possibility of delay/starvation here, since CMS token does 1840 // not know to give priority to VM thread? Actually, i think 1841 // there wouldn't be any delay/starvation, but the proof of 1842 // that "fact" (?) appears non-trivial. XXX 20011219YSR 1843 } 1844 ConcurrentMarkSweepThread::set_CMS_flag( 1845 ConcurrentMarkSweepThread::CMS_vm_has_token); 1846 } 1847 } 1848 // The CMS_token is already held. Get back the other locks. 1849 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 1850 "VM thread should have CMS token"); 1851 getFreelistLocks(); 1852 bitMapLock()->lock_without_safepoint_check(); 1853 if (TraceCMSState) { 1854 gclog_or_tty->print_cr("CMS foreground collector has asked for control " 1855 INTPTR_FORMAT " with first state %d", Thread::current(), first_state); 1856 gclog_or_tty->print_cr(" gets control with state %d", _collectorState); 1857 } 1858 1859 // Check if we need to do a compaction, or if not, whether 1860 // we need to start the mark-sweep from scratch. 1861 bool should_compact = false; 1862 bool should_start_over = false; 1863 decide_foreground_collection_type(clear_all_soft_refs, 1864 &should_compact, &should_start_over); 1865 1866 NOT_PRODUCT( 1867 if (RotateCMSCollectionTypes) { 1868 if (_cmsGen->debug_collection_type() == 1869 ConcurrentMarkSweepGeneration::MSC_foreground_collection_type) { 1870 should_compact = true; 1871 } else if (_cmsGen->debug_collection_type() == 1872 ConcurrentMarkSweepGeneration::MS_foreground_collection_type) { 1873 should_compact = false; 1874 } 1875 } 1876 ) 1877 1878 if (first_state > Idling) { 1879 report_concurrent_mode_interruption(); 1880 } 1881 1882 set_did_compact(should_compact); 1883 if (should_compact) { 1884 // If the collection is being acquired from the background 1885 // collector, there may be references on the discovered 1886 // references lists that have NULL referents (being those 1887 // that were concurrently cleared by a mutator) or 1888 // that are no longer active (having been enqueued concurrently 1889 // by the mutator). 1890 // Scrub the list of those references because Mark-Sweep-Compact 1891 // code assumes referents are not NULL and that all discovered 1892 // Reference objects are active. 1893 ref_processor()->clean_up_discovered_references(); 1894 1895 if (first_state > Idling) { 1896 save_heap_summary(); 1897 } 1898 1899 do_compaction_work(clear_all_soft_refs); 1900 1901 // Has the GC time limit been exceeded? 1902 DefNewGeneration* young_gen = _young_gen->as_DefNewGeneration(); 1903 size_t max_eden_size = young_gen->max_capacity() - 1904 young_gen->to()->capacity() - 1905 young_gen->from()->capacity(); 1906 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1907 GCCause::Cause gc_cause = gch->gc_cause(); 1908 size_policy()->check_gc_overhead_limit(_young_gen->used(), 1909 young_gen->eden()->used(), 1910 _cmsGen->max_capacity(), 1911 max_eden_size, 1912 full, 1913 gc_cause, 1914 gch->collector_policy()); 1915 } else { 1916 do_mark_sweep_work(clear_all_soft_refs, first_state, 1917 should_start_over); 1918 } 1919 // Reset the expansion cause, now that we just completed 1920 // a collection cycle. 1921 clear_expansion_cause(); 1922 _foregroundGCIsActive = false; 1923 return; 1924 } 1925 1926 // Resize the tenured generation 1927 // after obtaining the free list locks for the 1928 // two generations. 1929 void CMSCollector::compute_new_size() { 1930 assert_locked_or_safepoint(Heap_lock); 1931 FreelistLocker z(this); 1932 MetaspaceGC::compute_new_size(); 1933 _cmsGen->compute_new_size_free_list(); 1934 } 1935 1936 // A work method used by foreground collection to determine 1937 // what type of collection (compacting or not, continuing or fresh) 1938 // it should do. 1939 // NOTE: the intent is to make UseCMSCompactAtFullCollection 1940 // and CMSCompactWhenClearAllSoftRefs the default in the future 1941 // and do away with the flags after a suitable period. 1942 void CMSCollector::decide_foreground_collection_type( 1943 bool clear_all_soft_refs, bool* should_compact, 1944 bool* should_start_over) { 1945 // Normally, we'll compact only if the UseCMSCompactAtFullCollection 1946 // flag is set, and we have either requested a System.gc() or 1947 // the number of full gc's since the last concurrent cycle 1948 // has exceeded the threshold set by CMSFullGCsBeforeCompaction, 1949 // or if an incremental collection has failed 1950 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1951 assert(gch->collector_policy()->is_two_generation_policy(), 1952 "You may want to check the correctness of the following"); 1953 // Inform cms gen if this was due to partial collection failing. 1954 // The CMS gen may use this fact to determine its expansion policy. 1955 if (gch->incremental_collection_will_fail(false /* don't consult_young */)) { 1956 assert(!_cmsGen->incremental_collection_failed(), 1957 "Should have been noticed, reacted to and cleared"); 1958 _cmsGen->set_incremental_collection_failed(); 1959 } 1960 *should_compact = 1961 UseCMSCompactAtFullCollection && 1962 ((_full_gcs_since_conc_gc >= CMSFullGCsBeforeCompaction) || 1963 GCCause::is_user_requested_gc(gch->gc_cause()) || 1964 gch->incremental_collection_will_fail(true /* consult_young */)); 1965 *should_start_over = false; 1966 if (clear_all_soft_refs && !*should_compact) { 1967 // We are about to do a last ditch collection attempt 1968 // so it would normally make sense to do a compaction 1969 // to reclaim as much space as possible. 1970 if (CMSCompactWhenClearAllSoftRefs) { 1971 // Default: The rationale is that in this case either 1972 // we are past the final marking phase, in which case 1973 // we'd have to start over, or so little has been done 1974 // that there's little point in saving that work. Compaction 1975 // appears to be the sensible choice in either case. 1976 *should_compact = true; 1977 } else { 1978 // We have been asked to clear all soft refs, but not to 1979 // compact. Make sure that we aren't past the final checkpoint 1980 // phase, for that is where we process soft refs. If we are already 1981 // past that phase, we'll need to redo the refs discovery phase and 1982 // if necessary clear soft refs that weren't previously 1983 // cleared. We do so by remembering the phase in which 1984 // we came in, and if we are past the refs processing 1985 // phase, we'll choose to just redo the mark-sweep 1986 // collection from scratch. 1987 if (_collectorState > FinalMarking) { 1988 // We are past the refs processing phase; 1989 // start over and do a fresh synchronous CMS cycle 1990 _collectorState = Resetting; // skip to reset to start new cycle 1991 reset(false /* == !asynch */); 1992 *should_start_over = true; 1993 } // else we can continue a possibly ongoing current cycle 1994 } 1995 } 1996 } 1997 1998 // A work method used by the foreground collector to do 1999 // a mark-sweep-compact. 2000 void CMSCollector::do_compaction_work(bool clear_all_soft_refs) { 2001 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2002 2003 STWGCTimer* gc_timer = GenMarkSweep::gc_timer(); 2004 gc_timer->register_gc_start(os::elapsed_counter()); 2005 2006 SerialOldTracer* gc_tracer = GenMarkSweep::gc_tracer(); 2007 gc_tracer->report_gc_start(gch->gc_cause(), gc_timer->gc_start()); 2008 2009 GCTraceTime t("CMS:MSC ", PrintGCDetails && Verbose, true, NULL); 2010 if (PrintGC && Verbose && !(GCCause::is_user_requested_gc(gch->gc_cause()))) { 2011 gclog_or_tty->print_cr("Compact ConcurrentMarkSweepGeneration after %d " 2012 "collections passed to foreground collector", _full_gcs_since_conc_gc); 2013 } 2014 2015 // Sample collection interval time and reset for collection pause. 2016 if (UseAdaptiveSizePolicy) { 2017 size_policy()->msc_collection_begin(); 2018 } 2019 2020 // Temporarily widen the span of the weak reference processing to 2021 // the entire heap. 2022 MemRegion new_span(GenCollectedHeap::heap()->reserved_region()); 2023 ReferenceProcessorSpanMutator rp_mut_span(ref_processor(), new_span); 2024 // Temporarily, clear the "is_alive_non_header" field of the 2025 // reference processor. 2026 ReferenceProcessorIsAliveMutator rp_mut_closure(ref_processor(), NULL); 2027 // Temporarily make reference _processing_ single threaded (non-MT). 2028 ReferenceProcessorMTProcMutator rp_mut_mt_processing(ref_processor(), false); 2029 // Temporarily make refs discovery atomic 2030 ReferenceProcessorAtomicMutator rp_mut_atomic(ref_processor(), true); 2031 // Temporarily make reference _discovery_ single threaded (non-MT) 2032 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false); 2033 2034 ref_processor()->set_enqueuing_is_done(false); 2035 ref_processor()->enable_discovery(false /*verify_disabled*/, false /*check_no_refs*/); 2036 ref_processor()->setup_policy(clear_all_soft_refs); 2037 // If an asynchronous collection finishes, the _modUnionTable is 2038 // all clear. If we are assuming the collection from an asynchronous 2039 // collection, clear the _modUnionTable. 2040 assert(_collectorState != Idling || _modUnionTable.isAllClear(), 2041 "_modUnionTable should be clear if the baton was not passed"); 2042 _modUnionTable.clear_all(); 2043 assert(_collectorState != Idling || _ct->klass_rem_set()->mod_union_is_clear(), 2044 "mod union for klasses should be clear if the baton was passed"); 2045 _ct->klass_rem_set()->clear_mod_union(); 2046 2047 // We must adjust the allocation statistics being maintained 2048 // in the free list space. We do so by reading and clearing 2049 // the sweep timer and updating the block flux rate estimates below. 2050 assert(!_intra_sweep_timer.is_active(), "_intra_sweep_timer should be inactive"); 2051 if (_inter_sweep_timer.is_active()) { 2052 _inter_sweep_timer.stop(); 2053 // Note that we do not use this sample to update the _inter_sweep_estimate. 2054 _cmsGen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()), 2055 _inter_sweep_estimate.padded_average(), 2056 _intra_sweep_estimate.padded_average()); 2057 } 2058 2059 GenMarkSweep::invoke_at_safepoint(_cmsGen->level(), 2060 ref_processor(), clear_all_soft_refs); 2061 #ifdef ASSERT 2062 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 2063 size_t free_size = cms_space->free(); 2064 assert(free_size == 2065 pointer_delta(cms_space->end(), cms_space->compaction_top()) 2066 * HeapWordSize, 2067 "All the free space should be compacted into one chunk at top"); 2068 assert(cms_space->dictionary()->total_chunk_size( 2069 debug_only(cms_space->freelistLock())) == 0 || 2070 cms_space->totalSizeInIndexedFreeLists() == 0, 2071 "All the free space should be in a single chunk"); 2072 size_t num = cms_space->totalCount(); 2073 assert((free_size == 0 && num == 0) || 2074 (free_size > 0 && (num == 1 || num == 2)), 2075 "There should be at most 2 free chunks after compaction"); 2076 #endif // ASSERT 2077 _collectorState = Resetting; 2078 assert(_restart_addr == NULL, 2079 "Should have been NULL'd before baton was passed"); 2080 reset(false /* == !asynch */); 2081 _cmsGen->reset_after_compaction(); 2082 _concurrent_cycles_since_last_unload = 0; 2083 2084 // Clear any data recorded in the PLAB chunk arrays. 2085 if (_survivor_plab_array != NULL) { 2086 reset_survivor_plab_arrays(); 2087 } 2088 2089 // Adjust the per-size allocation stats for the next epoch. 2090 _cmsGen->cmsSpace()->endSweepFLCensus(sweep_count() /* fake */); 2091 // Restart the "inter sweep timer" for the next epoch. 2092 _inter_sweep_timer.reset(); 2093 _inter_sweep_timer.start(); 2094 2095 // Sample collection pause time and reset for collection interval. 2096 if (UseAdaptiveSizePolicy) { 2097 size_policy()->msc_collection_end(gch->gc_cause()); 2098 } 2099 2100 gc_timer->register_gc_end(os::elapsed_counter()); 2101 2102 gc_tracer->report_gc_end(gc_timer->gc_end(), gc_timer->time_partitions()); 2103 2104 // For a mark-sweep-compact, compute_new_size() will be called 2105 // in the heap's do_collection() method. 2106 } 2107 2108 // A work method used by the foreground collector to do 2109 // a mark-sweep, after taking over from a possibly on-going 2110 // concurrent mark-sweep collection. 2111 void CMSCollector::do_mark_sweep_work(bool clear_all_soft_refs, 2112 CollectorState first_state, bool should_start_over) { 2113 if (PrintGC && Verbose) { 2114 gclog_or_tty->print_cr("Pass concurrent collection to foreground " 2115 "collector with count %d", 2116 _full_gcs_since_conc_gc); 2117 } 2118 switch (_collectorState) { 2119 case Idling: 2120 if (first_state == Idling || should_start_over) { 2121 // The background GC was not active, or should 2122 // restarted from scratch; start the cycle. 2123 _collectorState = InitialMarking; 2124 } 2125 // If first_state was not Idling, then a background GC 2126 // was in progress and has now finished. No need to do it 2127 // again. Leave the state as Idling. 2128 break; 2129 case Precleaning: 2130 // In the foreground case don't do the precleaning since 2131 // it is not done concurrently and there is extra work 2132 // required. 2133 _collectorState = FinalMarking; 2134 } 2135 collect_in_foreground(clear_all_soft_refs, GenCollectedHeap::heap()->gc_cause()); 2136 2137 // For a mark-sweep, compute_new_size() will be called 2138 // in the heap's do_collection() method. 2139 } 2140 2141 2142 void CMSCollector::print_eden_and_survivor_chunk_arrays() { 2143 DefNewGeneration* dng = _young_gen->as_DefNewGeneration(); 2144 EdenSpace* eden_space = dng->eden(); 2145 ContiguousSpace* from_space = dng->from(); 2146 ContiguousSpace* to_space = dng->to(); 2147 // Eden 2148 if (_eden_chunk_array != NULL) { 2149 gclog_or_tty->print_cr("eden " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")", 2150 eden_space->bottom(), eden_space->top(), 2151 eden_space->end(), eden_space->capacity()); 2152 gclog_or_tty->print_cr("_eden_chunk_index=" SIZE_FORMAT ", " 2153 "_eden_chunk_capacity=" SIZE_FORMAT, 2154 _eden_chunk_index, _eden_chunk_capacity); 2155 for (size_t i = 0; i < _eden_chunk_index; i++) { 2156 gclog_or_tty->print_cr("_eden_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT, 2157 i, _eden_chunk_array[i]); 2158 } 2159 } 2160 // Survivor 2161 if (_survivor_chunk_array != NULL) { 2162 gclog_or_tty->print_cr("survivor " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")", 2163 from_space->bottom(), from_space->top(), 2164 from_space->end(), from_space->capacity()); 2165 gclog_or_tty->print_cr("_survivor_chunk_index=" SIZE_FORMAT ", " 2166 "_survivor_chunk_capacity=" SIZE_FORMAT, 2167 _survivor_chunk_index, _survivor_chunk_capacity); 2168 for (size_t i = 0; i < _survivor_chunk_index; i++) { 2169 gclog_or_tty->print_cr("_survivor_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT, 2170 i, _survivor_chunk_array[i]); 2171 } 2172 } 2173 } 2174 2175 void CMSCollector::getFreelistLocks() const { 2176 // Get locks for all free lists in all generations that this 2177 // collector is responsible for 2178 _cmsGen->freelistLock()->lock_without_safepoint_check(); 2179 } 2180 2181 void CMSCollector::releaseFreelistLocks() const { 2182 // Release locks for all free lists in all generations that this 2183 // collector is responsible for 2184 _cmsGen->freelistLock()->unlock(); 2185 } 2186 2187 bool CMSCollector::haveFreelistLocks() const { 2188 // Check locks for all free lists in all generations that this 2189 // collector is responsible for 2190 assert_lock_strong(_cmsGen->freelistLock()); 2191 PRODUCT_ONLY(ShouldNotReachHere()); 2192 return true; 2193 } 2194 2195 // A utility class that is used by the CMS collector to 2196 // temporarily "release" the foreground collector from its 2197 // usual obligation to wait for the background collector to 2198 // complete an ongoing phase before proceeding. 2199 class ReleaseForegroundGC: public StackObj { 2200 private: 2201 CMSCollector* _c; 2202 public: 2203 ReleaseForegroundGC(CMSCollector* c) : _c(c) { 2204 assert(_c->_foregroundGCShouldWait, "Else should not need to call"); 2205 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 2206 // allow a potentially blocked foreground collector to proceed 2207 _c->_foregroundGCShouldWait = false; 2208 if (_c->_foregroundGCIsActive) { 2209 CGC_lock->notify(); 2210 } 2211 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 2212 "Possible deadlock"); 2213 } 2214 2215 ~ReleaseForegroundGC() { 2216 assert(!_c->_foregroundGCShouldWait, "Usage protocol violation?"); 2217 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 2218 _c->_foregroundGCShouldWait = true; 2219 } 2220 }; 2221 2222 // There are separate collect_in_background and collect_in_foreground because of 2223 // the different locking requirements of the background collector and the 2224 // foreground collector. There was originally an attempt to share 2225 // one "collect" method between the background collector and the foreground 2226 // collector but the if-then-else required made it cleaner to have 2227 // separate methods. 2228 void CMSCollector::collect_in_background(bool clear_all_soft_refs, GCCause::Cause cause) { 2229 assert(Thread::current()->is_ConcurrentGC_thread(), 2230 "A CMS asynchronous collection is only allowed on a CMS thread."); 2231 2232 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2233 { 2234 bool safepoint_check = Mutex::_no_safepoint_check_flag; 2235 MutexLockerEx hl(Heap_lock, safepoint_check); 2236 FreelistLocker fll(this); 2237 MutexLockerEx x(CGC_lock, safepoint_check); 2238 if (_foregroundGCIsActive || !UseAsyncConcMarkSweepGC) { 2239 // The foreground collector is active or we're 2240 // not using asynchronous collections. Skip this 2241 // background collection. 2242 assert(!_foregroundGCShouldWait, "Should be clear"); 2243 return; 2244 } else { 2245 assert(_collectorState == Idling, "Should be idling before start."); 2246 _collectorState = InitialMarking; 2247 register_gc_start(cause); 2248 // Reset the expansion cause, now that we are about to begin 2249 // a new cycle. 2250 clear_expansion_cause(); 2251 2252 // Clear the MetaspaceGC flag since a concurrent collection 2253 // is starting but also clear it after the collection. 2254 MetaspaceGC::set_should_concurrent_collect(false); 2255 } 2256 // Decide if we want to enable class unloading as part of the 2257 // ensuing concurrent GC cycle. 2258 update_should_unload_classes(); 2259 _full_gc_requested = false; // acks all outstanding full gc requests 2260 _full_gc_cause = GCCause::_no_gc; 2261 // Signal that we are about to start a collection 2262 gch->increment_total_full_collections(); // ... starting a collection cycle 2263 _collection_count_start = gch->total_full_collections(); 2264 } 2265 2266 // Used for PrintGC 2267 size_t prev_used; 2268 if (PrintGC && Verbose) { 2269 prev_used = _cmsGen->used(); // XXXPERM 2270 } 2271 2272 // The change of the collection state is normally done at this level; 2273 // the exceptions are phases that are executed while the world is 2274 // stopped. For those phases the change of state is done while the 2275 // world is stopped. For baton passing purposes this allows the 2276 // background collector to finish the phase and change state atomically. 2277 // The foreground collector cannot wait on a phase that is done 2278 // while the world is stopped because the foreground collector already 2279 // has the world stopped and would deadlock. 2280 while (_collectorState != Idling) { 2281 if (TraceCMSState) { 2282 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " in CMS state %d", 2283 Thread::current(), _collectorState); 2284 } 2285 // The foreground collector 2286 // holds the Heap_lock throughout its collection. 2287 // holds the CMS token (but not the lock) 2288 // except while it is waiting for the background collector to yield. 2289 // 2290 // The foreground collector should be blocked (not for long) 2291 // if the background collector is about to start a phase 2292 // executed with world stopped. If the background 2293 // collector has already started such a phase, the 2294 // foreground collector is blocked waiting for the 2295 // Heap_lock. The stop-world phases (InitialMarking and FinalMarking) 2296 // are executed in the VM thread. 2297 // 2298 // The locking order is 2299 // PendingListLock (PLL) -- if applicable (FinalMarking) 2300 // Heap_lock (both this & PLL locked in VM_CMS_Operation::prologue()) 2301 // CMS token (claimed in 2302 // stop_world_and_do() --> 2303 // safepoint_synchronize() --> 2304 // CMSThread::synchronize()) 2305 2306 { 2307 // Check if the FG collector wants us to yield. 2308 CMSTokenSync x(true); // is cms thread 2309 if (waitForForegroundGC()) { 2310 // We yielded to a foreground GC, nothing more to be 2311 // done this round. 2312 assert(_foregroundGCShouldWait == false, "We set it to false in " 2313 "waitForForegroundGC()"); 2314 if (TraceCMSState) { 2315 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT 2316 " exiting collection CMS state %d", 2317 Thread::current(), _collectorState); 2318 } 2319 return; 2320 } else { 2321 // The background collector can run but check to see if the 2322 // foreground collector has done a collection while the 2323 // background collector was waiting to get the CGC_lock 2324 // above. If yes, break so that _foregroundGCShouldWait 2325 // is cleared before returning. 2326 if (_collectorState == Idling) { 2327 break; 2328 } 2329 } 2330 } 2331 2332 assert(_foregroundGCShouldWait, "Foreground collector, if active, " 2333 "should be waiting"); 2334 2335 switch (_collectorState) { 2336 case InitialMarking: 2337 { 2338 ReleaseForegroundGC x(this); 2339 stats().record_cms_begin(); 2340 VM_CMS_Initial_Mark initial_mark_op(this); 2341 VMThread::execute(&initial_mark_op); 2342 } 2343 // The collector state may be any legal state at this point 2344 // since the background collector may have yielded to the 2345 // foreground collector. 2346 break; 2347 case Marking: 2348 // initial marking in checkpointRootsInitialWork has been completed 2349 if (markFromRoots(true)) { // we were successful 2350 assert(_collectorState == Precleaning, "Collector state should " 2351 "have changed"); 2352 } else { 2353 assert(_foregroundGCIsActive, "Internal state inconsistency"); 2354 } 2355 break; 2356 case Precleaning: 2357 if (UseAdaptiveSizePolicy) { 2358 size_policy()->concurrent_precleaning_begin(); 2359 } 2360 // marking from roots in markFromRoots has been completed 2361 preclean(); 2362 if (UseAdaptiveSizePolicy) { 2363 size_policy()->concurrent_precleaning_end(); 2364 } 2365 assert(_collectorState == AbortablePreclean || 2366 _collectorState == FinalMarking, 2367 "Collector state should have changed"); 2368 break; 2369 case AbortablePreclean: 2370 if (UseAdaptiveSizePolicy) { 2371 size_policy()->concurrent_phases_resume(); 2372 } 2373 abortable_preclean(); 2374 if (UseAdaptiveSizePolicy) { 2375 size_policy()->concurrent_precleaning_end(); 2376 } 2377 assert(_collectorState == FinalMarking, "Collector state should " 2378 "have changed"); 2379 break; 2380 case FinalMarking: 2381 { 2382 ReleaseForegroundGC x(this); 2383 2384 VM_CMS_Final_Remark final_remark_op(this); 2385 VMThread::execute(&final_remark_op); 2386 } 2387 assert(_foregroundGCShouldWait, "block post-condition"); 2388 break; 2389 case Sweeping: 2390 if (UseAdaptiveSizePolicy) { 2391 size_policy()->concurrent_sweeping_begin(); 2392 } 2393 // final marking in checkpointRootsFinal has been completed 2394 sweep(true); 2395 assert(_collectorState == Resizing, "Collector state change " 2396 "to Resizing must be done under the free_list_lock"); 2397 _full_gcs_since_conc_gc = 0; 2398 2399 // Stop the timers for adaptive size policy for the concurrent phases 2400 if (UseAdaptiveSizePolicy) { 2401 size_policy()->concurrent_sweeping_end(); 2402 size_policy()->concurrent_phases_end(gch->gc_cause(), 2403 gch->prev_gen(_cmsGen)->capacity(), 2404 _cmsGen->free()); 2405 } 2406 2407 case Resizing: { 2408 // Sweeping has been completed... 2409 // At this point the background collection has completed. 2410 // Don't move the call to compute_new_size() down 2411 // into code that might be executed if the background 2412 // collection was preempted. 2413 { 2414 ReleaseForegroundGC x(this); // unblock FG collection 2415 MutexLockerEx y(Heap_lock, Mutex::_no_safepoint_check_flag); 2416 CMSTokenSync z(true); // not strictly needed. 2417 if (_collectorState == Resizing) { 2418 compute_new_size(); 2419 save_heap_summary(); 2420 _collectorState = Resetting; 2421 } else { 2422 assert(_collectorState == Idling, "The state should only change" 2423 " because the foreground collector has finished the collection"); 2424 } 2425 } 2426 break; 2427 } 2428 case Resetting: 2429 // CMS heap resizing has been completed 2430 reset(true); 2431 assert(_collectorState == Idling, "Collector state should " 2432 "have changed"); 2433 2434 MetaspaceGC::set_should_concurrent_collect(false); 2435 2436 stats().record_cms_end(); 2437 // Don't move the concurrent_phases_end() and compute_new_size() 2438 // calls to here because a preempted background collection 2439 // has it's state set to "Resetting". 2440 break; 2441 case Idling: 2442 default: 2443 ShouldNotReachHere(); 2444 break; 2445 } 2446 if (TraceCMSState) { 2447 gclog_or_tty->print_cr(" Thread " INTPTR_FORMAT " done - next CMS state %d", 2448 Thread::current(), _collectorState); 2449 } 2450 assert(_foregroundGCShouldWait, "block post-condition"); 2451 } 2452 2453 // Should this be in gc_epilogue? 2454 collector_policy()->counters()->update_counters(); 2455 2456 { 2457 // Clear _foregroundGCShouldWait and, in the event that the 2458 // foreground collector is waiting, notify it, before 2459 // returning. 2460 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 2461 _foregroundGCShouldWait = false; 2462 if (_foregroundGCIsActive) { 2463 CGC_lock->notify(); 2464 } 2465 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 2466 "Possible deadlock"); 2467 } 2468 if (TraceCMSState) { 2469 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT 2470 " exiting collection CMS state %d", 2471 Thread::current(), _collectorState); 2472 } 2473 if (PrintGC && Verbose) { 2474 _cmsGen->print_heap_change(prev_used); 2475 } 2476 } 2477 2478 void CMSCollector::register_foreground_gc_start(GCCause::Cause cause) { 2479 if (!_cms_start_registered) { 2480 register_gc_start(cause); 2481 } 2482 } 2483 2484 void CMSCollector::register_gc_start(GCCause::Cause cause) { 2485 _cms_start_registered = true; 2486 _gc_timer_cm->register_gc_start(os::elapsed_counter()); 2487 _gc_tracer_cm->report_gc_start(cause, _gc_timer_cm->gc_start()); 2488 } 2489 2490 void CMSCollector::register_gc_end() { 2491 if (_cms_start_registered) { 2492 report_heap_summary(GCWhen::AfterGC); 2493 2494 _gc_timer_cm->register_gc_end(os::elapsed_counter()); 2495 _gc_tracer_cm->report_gc_end(_gc_timer_cm->gc_end(), _gc_timer_cm->time_partitions()); 2496 _cms_start_registered = false; 2497 } 2498 } 2499 2500 void CMSCollector::save_heap_summary() { 2501 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2502 _last_heap_summary = gch->create_heap_summary(); 2503 _last_metaspace_summary = gch->create_metaspace_summary(); 2504 } 2505 2506 void CMSCollector::report_heap_summary(GCWhen::Type when) { 2507 _gc_tracer_cm->report_gc_heap_summary(when, _last_heap_summary, _last_metaspace_summary); 2508 } 2509 2510 void CMSCollector::collect_in_foreground(bool clear_all_soft_refs, GCCause::Cause cause) { 2511 assert(_foregroundGCIsActive && !_foregroundGCShouldWait, 2512 "Foreground collector should be waiting, not executing"); 2513 assert(Thread::current()->is_VM_thread(), "A foreground collection" 2514 "may only be done by the VM Thread with the world stopped"); 2515 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 2516 "VM thread should have CMS token"); 2517 2518 NOT_PRODUCT(GCTraceTime t("CMS:MS (foreground) ", PrintGCDetails && Verbose, 2519 true, NULL);) 2520 if (UseAdaptiveSizePolicy) { 2521 size_policy()->ms_collection_begin(); 2522 } 2523 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact); 2524 2525 HandleMark hm; // Discard invalid handles created during verification 2526 2527 if (VerifyBeforeGC && 2528 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2529 Universe::verify(); 2530 } 2531 2532 // Snapshot the soft reference policy to be used in this collection cycle. 2533 ref_processor()->setup_policy(clear_all_soft_refs); 2534 2535 bool init_mark_was_synchronous = false; // until proven otherwise 2536 while (_collectorState != Idling) { 2537 if (TraceCMSState) { 2538 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " in CMS state %d", 2539 Thread::current(), _collectorState); 2540 } 2541 switch (_collectorState) { 2542 case InitialMarking: 2543 register_foreground_gc_start(cause); 2544 init_mark_was_synchronous = true; // fact to be exploited in re-mark 2545 checkpointRootsInitial(false); 2546 assert(_collectorState == Marking, "Collector state should have changed" 2547 " within checkpointRootsInitial()"); 2548 break; 2549 case Marking: 2550 // initial marking in checkpointRootsInitialWork has been completed 2551 if (VerifyDuringGC && 2552 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2553 Universe::verify("Verify before initial mark: "); 2554 } 2555 { 2556 bool res = markFromRoots(false); 2557 assert(res && _collectorState == FinalMarking, "Collector state should " 2558 "have changed"); 2559 break; 2560 } 2561 case FinalMarking: 2562 if (VerifyDuringGC && 2563 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2564 Universe::verify("Verify before re-mark: "); 2565 } 2566 checkpointRootsFinal(false, clear_all_soft_refs, 2567 init_mark_was_synchronous); 2568 assert(_collectorState == Sweeping, "Collector state should not " 2569 "have changed within checkpointRootsFinal()"); 2570 break; 2571 case Sweeping: 2572 // final marking in checkpointRootsFinal has been completed 2573 if (VerifyDuringGC && 2574 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2575 Universe::verify("Verify before sweep: "); 2576 } 2577 sweep(false); 2578 assert(_collectorState == Resizing, "Incorrect state"); 2579 break; 2580 case Resizing: { 2581 // Sweeping has been completed; the actual resize in this case 2582 // is done separately; nothing to be done in this state. 2583 _collectorState = Resetting; 2584 break; 2585 } 2586 case Resetting: 2587 // The heap has been resized. 2588 if (VerifyDuringGC && 2589 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2590 Universe::verify("Verify before reset: "); 2591 } 2592 save_heap_summary(); 2593 reset(false); 2594 assert(_collectorState == Idling, "Collector state should " 2595 "have changed"); 2596 break; 2597 case Precleaning: 2598 case AbortablePreclean: 2599 // Elide the preclean phase 2600 _collectorState = FinalMarking; 2601 break; 2602 default: 2603 ShouldNotReachHere(); 2604 } 2605 if (TraceCMSState) { 2606 gclog_or_tty->print_cr(" Thread " INTPTR_FORMAT " done - next CMS state %d", 2607 Thread::current(), _collectorState); 2608 } 2609 } 2610 2611 if (UseAdaptiveSizePolicy) { 2612 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2613 size_policy()->ms_collection_end(gch->gc_cause()); 2614 } 2615 2616 if (VerifyAfterGC && 2617 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2618 Universe::verify(); 2619 } 2620 if (TraceCMSState) { 2621 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT 2622 " exiting collection CMS state %d", 2623 Thread::current(), _collectorState); 2624 } 2625 } 2626 2627 bool CMSCollector::waitForForegroundGC() { 2628 bool res = false; 2629 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 2630 "CMS thread should have CMS token"); 2631 // Block the foreground collector until the 2632 // background collectors decides whether to 2633 // yield. 2634 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 2635 _foregroundGCShouldWait = true; 2636 if (_foregroundGCIsActive) { 2637 // The background collector yields to the 2638 // foreground collector and returns a value 2639 // indicating that it has yielded. The foreground 2640 // collector can proceed. 2641 res = true; 2642 _foregroundGCShouldWait = false; 2643 ConcurrentMarkSweepThread::clear_CMS_flag( 2644 ConcurrentMarkSweepThread::CMS_cms_has_token); 2645 ConcurrentMarkSweepThread::set_CMS_flag( 2646 ConcurrentMarkSweepThread::CMS_cms_wants_token); 2647 // Get a possibly blocked foreground thread going 2648 CGC_lock->notify(); 2649 if (TraceCMSState) { 2650 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " waiting at CMS state %d", 2651 Thread::current(), _collectorState); 2652 } 2653 while (_foregroundGCIsActive) { 2654 CGC_lock->wait(Mutex::_no_safepoint_check_flag); 2655 } 2656 ConcurrentMarkSweepThread::set_CMS_flag( 2657 ConcurrentMarkSweepThread::CMS_cms_has_token); 2658 ConcurrentMarkSweepThread::clear_CMS_flag( 2659 ConcurrentMarkSweepThread::CMS_cms_wants_token); 2660 } 2661 if (TraceCMSState) { 2662 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " continuing at CMS state %d", 2663 Thread::current(), _collectorState); 2664 } 2665 return res; 2666 } 2667 2668 // Because of the need to lock the free lists and other structures in 2669 // the collector, common to all the generations that the collector is 2670 // collecting, we need the gc_prologues of individual CMS generations 2671 // delegate to their collector. It may have been simpler had the 2672 // current infrastructure allowed one to call a prologue on a 2673 // collector. In the absence of that we have the generation's 2674 // prologue delegate to the collector, which delegates back 2675 // some "local" work to a worker method in the individual generations 2676 // that it's responsible for collecting, while itself doing any 2677 // work common to all generations it's responsible for. A similar 2678 // comment applies to the gc_epilogue()'s. 2679 // The role of the varaible _between_prologue_and_epilogue is to 2680 // enforce the invocation protocol. 2681 void CMSCollector::gc_prologue(bool full) { 2682 // Call gc_prologue_work() for the CMSGen 2683 // we are responsible for. 2684 2685 // The following locking discipline assumes that we are only called 2686 // when the world is stopped. 2687 assert(SafepointSynchronize::is_at_safepoint(), "world is stopped assumption"); 2688 2689 // The CMSCollector prologue must call the gc_prologues for the 2690 // "generations" that it's responsible 2691 // for. 2692 2693 assert( Thread::current()->is_VM_thread() 2694 || ( CMSScavengeBeforeRemark 2695 && Thread::current()->is_ConcurrentGC_thread()), 2696 "Incorrect thread type for prologue execution"); 2697 2698 if (_between_prologue_and_epilogue) { 2699 // We have already been invoked; this is a gc_prologue delegation 2700 // from yet another CMS generation that we are responsible for, just 2701 // ignore it since all relevant work has already been done. 2702 return; 2703 } 2704 2705 // set a bit saying prologue has been called; cleared in epilogue 2706 _between_prologue_and_epilogue = true; 2707 // Claim locks for common data structures, then call gc_prologue_work() 2708 // for each CMSGen. 2709 2710 getFreelistLocks(); // gets free list locks on constituent spaces 2711 bitMapLock()->lock_without_safepoint_check(); 2712 2713 // Should call gc_prologue_work() for all cms gens we are responsible for 2714 bool duringMarking = _collectorState >= Marking 2715 && _collectorState < Sweeping; 2716 2717 // The young collections clear the modified oops state, which tells if 2718 // there are any modified oops in the class. The remark phase also needs 2719 // that information. Tell the young collection to save the union of all 2720 // modified klasses. 2721 if (duringMarking) { 2722 _ct->klass_rem_set()->set_accumulate_modified_oops(true); 2723 } 2724 2725 bool registerClosure = duringMarking; 2726 2727 ModUnionClosure* muc = CollectedHeap::use_parallel_gc_threads() ? 2728 &_modUnionClosurePar 2729 : &_modUnionClosure; 2730 _cmsGen->gc_prologue_work(full, registerClosure, muc); 2731 2732 if (!full) { 2733 stats().record_gc0_begin(); 2734 } 2735 } 2736 2737 void ConcurrentMarkSweepGeneration::gc_prologue(bool full) { 2738 2739 _capacity_at_prologue = capacity(); 2740 _used_at_prologue = used(); 2741 2742 // Delegate to CMScollector which knows how to coordinate between 2743 // this and any other CMS generations that it is responsible for 2744 // collecting. 2745 collector()->gc_prologue(full); 2746 } 2747 2748 // This is a "private" interface for use by this generation's CMSCollector. 2749 // Not to be called directly by any other entity (for instance, 2750 // GenCollectedHeap, which calls the "public" gc_prologue method above). 2751 void ConcurrentMarkSweepGeneration::gc_prologue_work(bool full, 2752 bool registerClosure, ModUnionClosure* modUnionClosure) { 2753 assert(!incremental_collection_failed(), "Shouldn't be set yet"); 2754 assert(cmsSpace()->preconsumptionDirtyCardClosure() == NULL, 2755 "Should be NULL"); 2756 if (registerClosure) { 2757 cmsSpace()->setPreconsumptionDirtyCardClosure(modUnionClosure); 2758 } 2759 cmsSpace()->gc_prologue(); 2760 // Clear stat counters 2761 NOT_PRODUCT( 2762 assert(_numObjectsPromoted == 0, "check"); 2763 assert(_numWordsPromoted == 0, "check"); 2764 if (Verbose && PrintGC) { 2765 gclog_or_tty->print("Allocated "SIZE_FORMAT" objects, " 2766 SIZE_FORMAT" bytes concurrently", 2767 _numObjectsAllocated, _numWordsAllocated*sizeof(HeapWord)); 2768 } 2769 _numObjectsAllocated = 0; 2770 _numWordsAllocated = 0; 2771 ) 2772 } 2773 2774 void CMSCollector::gc_epilogue(bool full) { 2775 // The following locking discipline assumes that we are only called 2776 // when the world is stopped. 2777 assert(SafepointSynchronize::is_at_safepoint(), 2778 "world is stopped assumption"); 2779 2780 // Currently the CMS epilogue (see CompactibleFreeListSpace) merely checks 2781 // if linear allocation blocks need to be appropriately marked to allow the 2782 // the blocks to be parsable. We also check here whether we need to nudge the 2783 // CMS collector thread to start a new cycle (if it's not already active). 2784 assert( Thread::current()->is_VM_thread() 2785 || ( CMSScavengeBeforeRemark 2786 && Thread::current()->is_ConcurrentGC_thread()), 2787 "Incorrect thread type for epilogue execution"); 2788 2789 if (!_between_prologue_and_epilogue) { 2790 // We have already been invoked; this is a gc_epilogue delegation 2791 // from yet another CMS generation that we are responsible for, just 2792 // ignore it since all relevant work has already been done. 2793 return; 2794 } 2795 assert(haveFreelistLocks(), "must have freelist locks"); 2796 assert_lock_strong(bitMapLock()); 2797 2798 _ct->klass_rem_set()->set_accumulate_modified_oops(false); 2799 2800 _cmsGen->gc_epilogue_work(full); 2801 2802 if (_collectorState == AbortablePreclean || _collectorState == Precleaning) { 2803 // in case sampling was not already enabled, enable it 2804 _start_sampling = true; 2805 } 2806 // reset _eden_chunk_array so sampling starts afresh 2807 _eden_chunk_index = 0; 2808 2809 size_t cms_used = _cmsGen->cmsSpace()->used(); 2810 2811 // update performance counters - this uses a special version of 2812 // update_counters() that allows the utilization to be passed as a 2813 // parameter, avoiding multiple calls to used(). 2814 // 2815 _cmsGen->update_counters(cms_used); 2816 2817 if (CMSIncrementalMode) { 2818 icms_update_allocation_limits(); 2819 } 2820 2821 bitMapLock()->unlock(); 2822 releaseFreelistLocks(); 2823 2824 if (!CleanChunkPoolAsync) { 2825 Chunk::clean_chunk_pool(); 2826 } 2827 2828 set_did_compact(false); 2829 _between_prologue_and_epilogue = false; // ready for next cycle 2830 } 2831 2832 void ConcurrentMarkSweepGeneration::gc_epilogue(bool full) { 2833 collector()->gc_epilogue(full); 2834 2835 // Also reset promotion tracking in par gc thread states. 2836 if (CollectedHeap::use_parallel_gc_threads()) { 2837 for (uint i = 0; i < ParallelGCThreads; i++) { 2838 _par_gc_thread_states[i]->promo.stopTrackingPromotions(i); 2839 } 2840 } 2841 } 2842 2843 void ConcurrentMarkSweepGeneration::gc_epilogue_work(bool full) { 2844 assert(!incremental_collection_failed(), "Should have been cleared"); 2845 cmsSpace()->setPreconsumptionDirtyCardClosure(NULL); 2846 cmsSpace()->gc_epilogue(); 2847 // Print stat counters 2848 NOT_PRODUCT( 2849 assert(_numObjectsAllocated == 0, "check"); 2850 assert(_numWordsAllocated == 0, "check"); 2851 if (Verbose && PrintGC) { 2852 gclog_or_tty->print("Promoted "SIZE_FORMAT" objects, " 2853 SIZE_FORMAT" bytes", 2854 _numObjectsPromoted, _numWordsPromoted*sizeof(HeapWord)); 2855 } 2856 _numObjectsPromoted = 0; 2857 _numWordsPromoted = 0; 2858 ) 2859 2860 if (PrintGC && Verbose) { 2861 // Call down the chain in contiguous_available needs the freelistLock 2862 // so print this out before releasing the freeListLock. 2863 gclog_or_tty->print(" Contiguous available "SIZE_FORMAT" bytes ", 2864 contiguous_available()); 2865 } 2866 } 2867 2868 #ifndef PRODUCT 2869 bool CMSCollector::have_cms_token() { 2870 Thread* thr = Thread::current(); 2871 if (thr->is_VM_thread()) { 2872 return ConcurrentMarkSweepThread::vm_thread_has_cms_token(); 2873 } else if (thr->is_ConcurrentGC_thread()) { 2874 return ConcurrentMarkSweepThread::cms_thread_has_cms_token(); 2875 } else if (thr->is_GC_task_thread()) { 2876 return ConcurrentMarkSweepThread::vm_thread_has_cms_token() && 2877 ParGCRareEvent_lock->owned_by_self(); 2878 } 2879 return false; 2880 } 2881 #endif 2882 2883 // Check reachability of the given heap address in CMS generation, 2884 // treating all other generations as roots. 2885 bool CMSCollector::is_cms_reachable(HeapWord* addr) { 2886 // We could "guarantee" below, rather than assert, but i'll 2887 // leave these as "asserts" so that an adventurous debugger 2888 // could try this in the product build provided some subset of 2889 // the conditions were met, provided they were intersted in the 2890 // results and knew that the computation below wouldn't interfere 2891 // with other concurrent computations mutating the structures 2892 // being read or written. 2893 assert(SafepointSynchronize::is_at_safepoint(), 2894 "Else mutations in object graph will make answer suspect"); 2895 assert(have_cms_token(), "Should hold cms token"); 2896 assert(haveFreelistLocks(), "must hold free list locks"); 2897 assert_lock_strong(bitMapLock()); 2898 2899 // Clear the marking bit map array before starting, but, just 2900 // for kicks, first report if the given address is already marked 2901 gclog_or_tty->print_cr("Start: Address 0x%x is%s marked", addr, 2902 _markBitMap.isMarked(addr) ? "" : " not"); 2903 2904 if (verify_after_remark()) { 2905 MutexLockerEx x(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag); 2906 bool result = verification_mark_bm()->isMarked(addr); 2907 gclog_or_tty->print_cr("TransitiveMark: Address 0x%x %s marked", addr, 2908 result ? "IS" : "is NOT"); 2909 return result; 2910 } else { 2911 gclog_or_tty->print_cr("Could not compute result"); 2912 return false; 2913 } 2914 } 2915 2916 2917 void 2918 CMSCollector::print_on_error(outputStream* st) { 2919 CMSCollector* collector = ConcurrentMarkSweepGeneration::_collector; 2920 if (collector != NULL) { 2921 CMSBitMap* bitmap = &collector->_markBitMap; 2922 st->print_cr("Marking Bits: (CMSBitMap*) " PTR_FORMAT, bitmap); 2923 bitmap->print_on_error(st, " Bits: "); 2924 2925 st->cr(); 2926 2927 CMSBitMap* mut_bitmap = &collector->_modUnionTable; 2928 st->print_cr("Mod Union Table: (CMSBitMap*) " PTR_FORMAT, mut_bitmap); 2929 mut_bitmap->print_on_error(st, " Bits: "); 2930 } 2931 } 2932 2933 //////////////////////////////////////////////////////// 2934 // CMS Verification Support 2935 //////////////////////////////////////////////////////// 2936 // Following the remark phase, the following invariant 2937 // should hold -- each object in the CMS heap which is 2938 // marked in markBitMap() should be marked in the verification_mark_bm(). 2939 2940 class VerifyMarkedClosure: public BitMapClosure { 2941 CMSBitMap* _marks; 2942 bool _failed; 2943 2944 public: 2945 VerifyMarkedClosure(CMSBitMap* bm): _marks(bm), _failed(false) {} 2946 2947 bool do_bit(size_t offset) { 2948 HeapWord* addr = _marks->offsetToHeapWord(offset); 2949 if (!_marks->isMarked(addr)) { 2950 oop(addr)->print_on(gclog_or_tty); 2951 gclog_or_tty->print_cr(" ("INTPTR_FORMAT" should have been marked)", addr); 2952 _failed = true; 2953 } 2954 return true; 2955 } 2956 2957 bool failed() { return _failed; } 2958 }; 2959 2960 bool CMSCollector::verify_after_remark(bool silent) { 2961 if (!silent) gclog_or_tty->print(" [Verifying CMS Marking... "); 2962 MutexLockerEx ml(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag); 2963 static bool init = false; 2964 2965 assert(SafepointSynchronize::is_at_safepoint(), 2966 "Else mutations in object graph will make answer suspect"); 2967 assert(have_cms_token(), 2968 "Else there may be mutual interference in use of " 2969 " verification data structures"); 2970 assert(_collectorState > Marking && _collectorState <= Sweeping, 2971 "Else marking info checked here may be obsolete"); 2972 assert(haveFreelistLocks(), "must hold free list locks"); 2973 assert_lock_strong(bitMapLock()); 2974 2975 2976 // Allocate marking bit map if not already allocated 2977 if (!init) { // first time 2978 if (!verification_mark_bm()->allocate(_span)) { 2979 return false; 2980 } 2981 init = true; 2982 } 2983 2984 assert(verification_mark_stack()->isEmpty(), "Should be empty"); 2985 2986 // Turn off refs discovery -- so we will be tracing through refs. 2987 // This is as intended, because by this time 2988 // GC must already have cleared any refs that need to be cleared, 2989 // and traced those that need to be marked; moreover, 2990 // the marking done here is not going to intefere in any 2991 // way with the marking information used by GC. 2992 NoRefDiscovery no_discovery(ref_processor()); 2993 2994 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;) 2995 2996 // Clear any marks from a previous round 2997 verification_mark_bm()->clear_all(); 2998 assert(verification_mark_stack()->isEmpty(), "markStack should be empty"); 2999 verify_work_stacks_empty(); 3000 3001 GenCollectedHeap* gch = GenCollectedHeap::heap(); 3002 gch->ensure_parsability(false); // fill TLABs, but no need to retire them 3003 // Update the saved marks which may affect the root scans. 3004 gch->save_marks(); 3005 3006 if (CMSRemarkVerifyVariant == 1) { 3007 // In this first variant of verification, we complete 3008 // all marking, then check if the new marks-verctor is 3009 // a subset of the CMS marks-vector. 3010 verify_after_remark_work_1(); 3011 } else if (CMSRemarkVerifyVariant == 2) { 3012 // In this second variant of verification, we flag an error 3013 // (i.e. an object reachable in the new marks-vector not reachable 3014 // in the CMS marks-vector) immediately, also indicating the 3015 // identify of an object (A) that references the unmarked object (B) -- 3016 // presumably, a mutation to A failed to be picked up by preclean/remark? 3017 verify_after_remark_work_2(); 3018 } else { 3019 warning("Unrecognized value %d for CMSRemarkVerifyVariant", 3020 CMSRemarkVerifyVariant); 3021 } 3022 if (!silent) gclog_or_tty->print(" done] "); 3023 return true; 3024 } 3025 3026 void CMSCollector::verify_after_remark_work_1() { 3027 ResourceMark rm; 3028 HandleMark hm; 3029 GenCollectedHeap* gch = GenCollectedHeap::heap(); 3030 3031 // Get a clear set of claim bits for the strong roots processing to work with. 3032 ClassLoaderDataGraph::clear_claimed_marks(); 3033 3034 // Mark from roots one level into CMS 3035 MarkRefsIntoClosure notOlder(_span, verification_mark_bm()); 3036 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 3037 3038 gch->gen_process_strong_roots(_cmsGen->level(), 3039 true, // younger gens are roots 3040 true, // activate StrongRootsScope 3041 false, // not scavenging 3042 SharedHeap::ScanningOption(roots_scanning_options()), 3043 ¬Older, 3044 true, // walk code active on stacks 3045 NULL, 3046 NULL); // SSS: Provide correct closure 3047 3048 // Now mark from the roots 3049 MarkFromRootsClosure markFromRootsClosure(this, _span, 3050 verification_mark_bm(), verification_mark_stack(), 3051 false /* don't yield */, true /* verifying */); 3052 assert(_restart_addr == NULL, "Expected pre-condition"); 3053 verification_mark_bm()->iterate(&markFromRootsClosure); 3054 while (_restart_addr != NULL) { 3055 // Deal with stack overflow: by restarting at the indicated 3056 // address. 3057 HeapWord* ra = _restart_addr; 3058 markFromRootsClosure.reset(ra); 3059 _restart_addr = NULL; 3060 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end()); 3061 } 3062 assert(verification_mark_stack()->isEmpty(), "Should have been drained"); 3063 verify_work_stacks_empty(); 3064 3065 // Marking completed -- now verify that each bit marked in 3066 // verification_mark_bm() is also marked in markBitMap(); flag all 3067 // errors by printing corresponding objects. 3068 VerifyMarkedClosure vcl(markBitMap()); 3069 verification_mark_bm()->iterate(&vcl); 3070 if (vcl.failed()) { 3071 gclog_or_tty->print("Verification failed"); 3072 Universe::heap()->print_on(gclog_or_tty); 3073 fatal("CMS: failed marking verification after remark"); 3074 } 3075 } 3076 3077 class VerifyKlassOopsKlassClosure : public KlassClosure { 3078 class VerifyKlassOopsClosure : public OopClosure { 3079 CMSBitMap* _bitmap; 3080 public: 3081 VerifyKlassOopsClosure(CMSBitMap* bitmap) : _bitmap(bitmap) { } 3082 void do_oop(oop* p) { guarantee(*p == NULL || _bitmap->isMarked((HeapWord*) *p), "Should be marked"); } 3083 void do_oop(narrowOop* p) { ShouldNotReachHere(); } 3084 } _oop_closure; 3085 public: 3086 VerifyKlassOopsKlassClosure(CMSBitMap* bitmap) : _oop_closure(bitmap) {} 3087 void do_klass(Klass* k) { 3088 k->oops_do(&_oop_closure); 3089 } 3090 }; 3091 3092 void CMSCollector::verify_after_remark_work_2() { 3093 ResourceMark rm; 3094 HandleMark hm; 3095 GenCollectedHeap* gch = GenCollectedHeap::heap(); 3096 3097 // Get a clear set of claim bits for the strong roots processing to work with. 3098 ClassLoaderDataGraph::clear_claimed_marks(); 3099 3100 // Mark from roots one level into CMS 3101 MarkRefsIntoVerifyClosure notOlder(_span, verification_mark_bm(), 3102 markBitMap()); 3103 CMKlassClosure klass_closure(¬Older); 3104 3105 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 3106 gch->gen_process_strong_roots(_cmsGen->level(), 3107 true, // younger gens are roots 3108 true, // activate StrongRootsScope 3109 false, // not scavenging 3110 SharedHeap::ScanningOption(roots_scanning_options()), 3111 ¬Older, 3112 true, // walk code active on stacks 3113 NULL, 3114 &klass_closure); 3115 3116 // Now mark from the roots 3117 MarkFromRootsVerifyClosure markFromRootsClosure(this, _span, 3118 verification_mark_bm(), markBitMap(), verification_mark_stack()); 3119 assert(_restart_addr == NULL, "Expected pre-condition"); 3120 verification_mark_bm()->iterate(&markFromRootsClosure); 3121 while (_restart_addr != NULL) { 3122 // Deal with stack overflow: by restarting at the indicated 3123 // address. 3124 HeapWord* ra = _restart_addr; 3125 markFromRootsClosure.reset(ra); 3126 _restart_addr = NULL; 3127 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end()); 3128 } 3129 assert(verification_mark_stack()->isEmpty(), "Should have been drained"); 3130 verify_work_stacks_empty(); 3131 3132 VerifyKlassOopsKlassClosure verify_klass_oops(verification_mark_bm()); 3133 ClassLoaderDataGraph::classes_do(&verify_klass_oops); 3134 3135 // Marking completed -- now verify that each bit marked in 3136 // verification_mark_bm() is also marked in markBitMap(); flag all 3137 // errors by printing corresponding objects. 3138 VerifyMarkedClosure vcl(markBitMap()); 3139 verification_mark_bm()->iterate(&vcl); 3140 assert(!vcl.failed(), "Else verification above should not have succeeded"); 3141 } 3142 3143 void ConcurrentMarkSweepGeneration::save_marks() { 3144 // delegate to CMS space 3145 cmsSpace()->save_marks(); 3146 for (uint i = 0; i < ParallelGCThreads; i++) { 3147 _par_gc_thread_states[i]->promo.startTrackingPromotions(); 3148 } 3149 } 3150 3151 bool ConcurrentMarkSweepGeneration::no_allocs_since_save_marks() { 3152 return cmsSpace()->no_allocs_since_save_marks(); 3153 } 3154 3155 #define CMS_SINCE_SAVE_MARKS_DEFN(OopClosureType, nv_suffix) \ 3156 \ 3157 void ConcurrentMarkSweepGeneration:: \ 3158 oop_since_save_marks_iterate##nv_suffix(OopClosureType* cl) { \ 3159 cl->set_generation(this); \ 3160 cmsSpace()->oop_since_save_marks_iterate##nv_suffix(cl); \ 3161 cl->reset_generation(); \ 3162 save_marks(); \ 3163 } 3164 3165 ALL_SINCE_SAVE_MARKS_CLOSURES(CMS_SINCE_SAVE_MARKS_DEFN) 3166 3167 void 3168 ConcurrentMarkSweepGeneration::younger_refs_iterate(OopsInGenClosure* cl) { 3169 cl->set_generation(this); 3170 younger_refs_in_space_iterate(_cmsSpace, cl); 3171 cl->reset_generation(); 3172 } 3173 3174 void 3175 ConcurrentMarkSweepGeneration::oop_iterate(MemRegion mr, ExtendedOopClosure* cl) { 3176 if (freelistLock()->owned_by_self()) { 3177 Generation::oop_iterate(mr, cl); 3178 } else { 3179 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 3180 Generation::oop_iterate(mr, cl); 3181 } 3182 } 3183 3184 void 3185 ConcurrentMarkSweepGeneration::oop_iterate(ExtendedOopClosure* cl) { 3186 if (freelistLock()->owned_by_self()) { 3187 Generation::oop_iterate(cl); 3188 } else { 3189 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 3190 Generation::oop_iterate(cl); 3191 } 3192 } 3193 3194 void 3195 ConcurrentMarkSweepGeneration::object_iterate(ObjectClosure* cl) { 3196 if (freelistLock()->owned_by_self()) { 3197 Generation::object_iterate(cl); 3198 } else { 3199 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 3200 Generation::object_iterate(cl); 3201 } 3202 } 3203 3204 void 3205 ConcurrentMarkSweepGeneration::safe_object_iterate(ObjectClosure* cl) { 3206 if (freelistLock()->owned_by_self()) { 3207 Generation::safe_object_iterate(cl); 3208 } else { 3209 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 3210 Generation::safe_object_iterate(cl); 3211 } 3212 } 3213 3214 void 3215 ConcurrentMarkSweepGeneration::post_compact() { 3216 } 3217 3218 void 3219 ConcurrentMarkSweepGeneration::prepare_for_verify() { 3220 // Fix the linear allocation blocks to look like free blocks. 3221 3222 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those 3223 // are not called when the heap is verified during universe initialization and 3224 // at vm shutdown. 3225 if (freelistLock()->owned_by_self()) { 3226 cmsSpace()->prepare_for_verify(); 3227 } else { 3228 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag); 3229 cmsSpace()->prepare_for_verify(); 3230 } 3231 } 3232 3233 void 3234 ConcurrentMarkSweepGeneration::verify() { 3235 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those 3236 // are not called when the heap is verified during universe initialization and 3237 // at vm shutdown. 3238 if (freelistLock()->owned_by_self()) { 3239 cmsSpace()->verify(); 3240 } else { 3241 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag); 3242 cmsSpace()->verify(); 3243 } 3244 } 3245 3246 void CMSCollector::verify() { 3247 _cmsGen->verify(); 3248 } 3249 3250 #ifndef PRODUCT 3251 bool CMSCollector::overflow_list_is_empty() const { 3252 assert(_num_par_pushes >= 0, "Inconsistency"); 3253 if (_overflow_list == NULL) { 3254 assert(_num_par_pushes == 0, "Inconsistency"); 3255 } 3256 return _overflow_list == NULL; 3257 } 3258 3259 // The methods verify_work_stacks_empty() and verify_overflow_empty() 3260 // merely consolidate assertion checks that appear to occur together frequently. 3261 void CMSCollector::verify_work_stacks_empty() const { 3262 assert(_markStack.isEmpty(), "Marking stack should be empty"); 3263 assert(overflow_list_is_empty(), "Overflow list should be empty"); 3264 } 3265 3266 void CMSCollector::verify_overflow_empty() const { 3267 assert(overflow_list_is_empty(), "Overflow list should be empty"); 3268 assert(no_preserved_marks(), "No preserved marks"); 3269 } 3270 #endif // PRODUCT 3271 3272 // Decide if we want to enable class unloading as part of the 3273 // ensuing concurrent GC cycle. We will collect and 3274 // unload classes if it's the case that: 3275 // (1) an explicit gc request has been made and the flag 3276 // ExplicitGCInvokesConcurrentAndUnloadsClasses is set, OR 3277 // (2) (a) class unloading is enabled at the command line, and 3278 // (b) old gen is getting really full 3279 // NOTE: Provided there is no change in the state of the heap between 3280 // calls to this method, it should have idempotent results. Moreover, 3281 // its results should be monotonically increasing (i.e. going from 0 to 1, 3282 // but not 1 to 0) between successive calls between which the heap was 3283 // not collected. For the implementation below, it must thus rely on 3284 // the property that concurrent_cycles_since_last_unload() 3285 // will not decrease unless a collection cycle happened and that 3286 // _cmsGen->is_too_full() are 3287 // themselves also monotonic in that sense. See check_monotonicity() 3288 // below. 3289 void CMSCollector::update_should_unload_classes() { 3290 _should_unload_classes = false; 3291 // Condition 1 above 3292 if (_full_gc_requested && ExplicitGCInvokesConcurrentAndUnloadsClasses) { 3293 _should_unload_classes = true; 3294 } else if (CMSClassUnloadingEnabled) { // Condition 2.a above 3295 // Disjuncts 2.b.(i,ii,iii) above 3296 _should_unload_classes = (concurrent_cycles_since_last_unload() >= 3297 CMSClassUnloadingMaxInterval) 3298 || _cmsGen->is_too_full(); 3299 } 3300 } 3301 3302 bool ConcurrentMarkSweepGeneration::is_too_full() const { 3303 bool res = should_concurrent_collect(); 3304 res = res && (occupancy() > (double)CMSIsTooFullPercentage/100.0); 3305 return res; 3306 } 3307 3308 void CMSCollector::setup_cms_unloading_and_verification_state() { 3309 const bool should_verify = VerifyBeforeGC || VerifyAfterGC || VerifyDuringGC 3310 || VerifyBeforeExit; 3311 const int rso = SharedHeap::SO_Strings | SharedHeap::SO_CodeCache; 3312 3313 // We set the proper root for this CMS cycle here. 3314 // When class unloading is disabled, we set the default root scanning option 3315 // once in the CMSCollector constructor. 3316 if (CMSClassUnloadingEnabled) { 3317 if (should_unload_classes()) { 3318 set_root_scanning_option(SharedHeap::SO_SystemClasses); 3319 } else { 3320 set_root_scanning_option(SharedHeap::SO_AllClasses | rso); 3321 } 3322 } 3323 3324 if (should_unload_classes()) { // Should unload classes this cycle 3325 remove_root_scanning_option(rso); // Shrink the root set appropriately 3326 set_verifying(should_verify); // Set verification state for this cycle 3327 return; // Nothing else needs to be done at this time 3328 } 3329 3330 // Not unloading classes this cycle 3331 assert(!should_unload_classes(), "Inconsitency!"); 3332 if ((!verifying() || unloaded_classes_last_cycle()) && should_verify) { 3333 // Include symbols, strings and code cache elements to prevent their resurrection. 3334 add_root_scanning_option(rso); 3335 set_verifying(true); 3336 } else if (verifying() && !should_verify) { 3337 // We were verifying, but some verification flags got disabled. 3338 set_verifying(false); 3339 // Exclude symbols, strings and code cache elements from root scanning to 3340 // reduce IM and RM pauses. 3341 remove_root_scanning_option(rso); 3342 } 3343 } 3344 3345 3346 #ifndef PRODUCT 3347 HeapWord* CMSCollector::block_start(const void* p) const { 3348 const HeapWord* addr = (HeapWord*)p; 3349 if (_span.contains(p)) { 3350 if (_cmsGen->cmsSpace()->is_in_reserved(addr)) { 3351 return _cmsGen->cmsSpace()->block_start(p); 3352 } 3353 } 3354 return NULL; 3355 } 3356 #endif 3357 3358 HeapWord* 3359 ConcurrentMarkSweepGeneration::expand_and_allocate(size_t word_size, 3360 bool tlab, 3361 bool parallel) { 3362 CMSSynchronousYieldRequest yr; 3363 assert(!tlab, "Can't deal with TLAB allocation"); 3364 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 3365 expand(word_size*HeapWordSize, MinHeapDeltaBytes, 3366 CMSExpansionCause::_satisfy_allocation); 3367 if (GCExpandToAllocateDelayMillis > 0) { 3368 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 3369 } 3370 return have_lock_and_allocate(word_size, tlab); 3371 } 3372 3373 // YSR: All of this generation expansion/shrinking stuff is an exact copy of 3374 // OneContigSpaceCardGeneration, which makes me wonder if we should move this 3375 // to CardGeneration and share it... 3376 bool ConcurrentMarkSweepGeneration::expand(size_t bytes, size_t expand_bytes) { 3377 return CardGeneration::expand(bytes, expand_bytes); 3378 } 3379 3380 void ConcurrentMarkSweepGeneration::expand(size_t bytes, size_t expand_bytes, 3381 CMSExpansionCause::Cause cause) 3382 { 3383 3384 bool success = expand(bytes, expand_bytes); 3385 3386 // remember why we expanded; this information is used 3387 // by shouldConcurrentCollect() when making decisions on whether to start 3388 // a new CMS cycle. 3389 if (success) { 3390 set_expansion_cause(cause); 3391 if (PrintGCDetails && Verbose) { 3392 gclog_or_tty->print_cr("Expanded CMS gen for %s", 3393 CMSExpansionCause::to_string(cause)); 3394 } 3395 } 3396 } 3397 3398 HeapWord* ConcurrentMarkSweepGeneration::expand_and_par_lab_allocate(CMSParGCThreadState* ps, size_t word_sz) { 3399 HeapWord* res = NULL; 3400 MutexLocker x(ParGCRareEvent_lock); 3401 while (true) { 3402 // Expansion by some other thread might make alloc OK now: 3403 res = ps->lab.alloc(word_sz); 3404 if (res != NULL) return res; 3405 // If there's not enough expansion space available, give up. 3406 if (_virtual_space.uncommitted_size() < (word_sz * HeapWordSize)) { 3407 return NULL; 3408 } 3409 // Otherwise, we try expansion. 3410 expand(word_sz*HeapWordSize, MinHeapDeltaBytes, 3411 CMSExpansionCause::_allocate_par_lab); 3412 // Now go around the loop and try alloc again; 3413 // A competing par_promote might beat us to the expansion space, 3414 // so we may go around the loop again if promotion fails agaion. 3415 if (GCExpandToAllocateDelayMillis > 0) { 3416 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 3417 } 3418 } 3419 } 3420 3421 3422 bool ConcurrentMarkSweepGeneration::expand_and_ensure_spooling_space( 3423 PromotionInfo* promo) { 3424 MutexLocker x(ParGCRareEvent_lock); 3425 size_t refill_size_bytes = promo->refillSize() * HeapWordSize; 3426 while (true) { 3427 // Expansion by some other thread might make alloc OK now: 3428 if (promo->ensure_spooling_space()) { 3429 assert(promo->has_spooling_space(), 3430 "Post-condition of successful ensure_spooling_space()"); 3431 return true; 3432 } 3433 // If there's not enough expansion space available, give up. 3434 if (_virtual_space.uncommitted_size() < refill_size_bytes) { 3435 return false; 3436 } 3437 // Otherwise, we try expansion. 3438 expand(refill_size_bytes, MinHeapDeltaBytes, 3439 CMSExpansionCause::_allocate_par_spooling_space); 3440 // Now go around the loop and try alloc again; 3441 // A competing allocation might beat us to the expansion space, 3442 // so we may go around the loop again if allocation fails again. 3443 if (GCExpandToAllocateDelayMillis > 0) { 3444 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 3445 } 3446 } 3447 } 3448 3449 3450 void ConcurrentMarkSweepGeneration::shrink_by(size_t bytes) { 3451 assert_locked_or_safepoint(ExpandHeap_lock); 3452 // Shrink committed space 3453 _virtual_space.shrink_by(bytes); 3454 // Shrink space; this also shrinks the space's BOT 3455 _cmsSpace->set_end((HeapWord*) _virtual_space.high()); 3456 size_t new_word_size = heap_word_size(_cmsSpace->capacity()); 3457 // Shrink the shared block offset array 3458 _bts->resize(new_word_size); 3459 MemRegion mr(_cmsSpace->bottom(), new_word_size); 3460 // Shrink the card table 3461 Universe::heap()->barrier_set()->resize_covered_region(mr); 3462 3463 if (Verbose && PrintGC) { 3464 size_t new_mem_size = _virtual_space.committed_size(); 3465 size_t old_mem_size = new_mem_size + bytes; 3466 gclog_or_tty->print_cr("Shrinking %s from " SIZE_FORMAT "K to " SIZE_FORMAT "K", 3467 name(), old_mem_size/K, new_mem_size/K); 3468 } 3469 } 3470 3471 void ConcurrentMarkSweepGeneration::shrink(size_t bytes) { 3472 assert_locked_or_safepoint(Heap_lock); 3473 size_t size = ReservedSpace::page_align_size_down(bytes); 3474 // Only shrink if a compaction was done so that all the free space 3475 // in the generation is in a contiguous block at the end. 3476 if (size > 0 && did_compact()) { 3477 shrink_by(size); 3478 } 3479 } 3480 3481 bool ConcurrentMarkSweepGeneration::grow_by(size_t bytes) { 3482 assert_locked_or_safepoint(Heap_lock); 3483 bool result = _virtual_space.expand_by(bytes); 3484 if (result) { 3485 size_t new_word_size = 3486 heap_word_size(_virtual_space.committed_size()); 3487 MemRegion mr(_cmsSpace->bottom(), new_word_size); 3488 _bts->resize(new_word_size); // resize the block offset shared array 3489 Universe::heap()->barrier_set()->resize_covered_region(mr); 3490 // Hmmmm... why doesn't CFLS::set_end verify locking? 3491 // This is quite ugly; FIX ME XXX 3492 _cmsSpace->assert_locked(freelistLock()); 3493 _cmsSpace->set_end((HeapWord*)_virtual_space.high()); 3494 3495 // update the space and generation capacity counters 3496 if (UsePerfData) { 3497 _space_counters->update_capacity(); 3498 _gen_counters->update_all(); 3499 } 3500 3501 if (Verbose && PrintGC) { 3502 size_t new_mem_size = _virtual_space.committed_size(); 3503 size_t old_mem_size = new_mem_size - bytes; 3504 gclog_or_tty->print_cr("Expanding %s from " SIZE_FORMAT "K by " SIZE_FORMAT "K to " SIZE_FORMAT "K", 3505 name(), old_mem_size/K, bytes/K, new_mem_size/K); 3506 } 3507 } 3508 return result; 3509 } 3510 3511 bool ConcurrentMarkSweepGeneration::grow_to_reserved() { 3512 assert_locked_or_safepoint(Heap_lock); 3513 bool success = true; 3514 const size_t remaining_bytes = _virtual_space.uncommitted_size(); 3515 if (remaining_bytes > 0) { 3516 success = grow_by(remaining_bytes); 3517 DEBUG_ONLY(if (!success) warning("grow to reserved failed");) 3518 } 3519 return success; 3520 } 3521 3522 void ConcurrentMarkSweepGeneration::shrink_free_list_by(size_t bytes) { 3523 assert_locked_or_safepoint(Heap_lock); 3524 assert_lock_strong(freelistLock()); 3525 if (PrintGCDetails && Verbose) { 3526 warning("Shrinking of CMS not yet implemented"); 3527 } 3528 return; 3529 } 3530 3531 3532 // Simple ctor/dtor wrapper for accounting & timer chores around concurrent 3533 // phases. 3534 class CMSPhaseAccounting: public StackObj { 3535 public: 3536 CMSPhaseAccounting(CMSCollector *collector, 3537 const char *phase, 3538 bool print_cr = true); 3539 ~CMSPhaseAccounting(); 3540 3541 private: 3542 CMSCollector *_collector; 3543 const char *_phase; 3544 elapsedTimer _wallclock; 3545 bool _print_cr; 3546 3547 public: 3548 // Not MT-safe; so do not pass around these StackObj's 3549 // where they may be accessed by other threads. 3550 jlong wallclock_millis() { 3551 assert(_wallclock.is_active(), "Wall clock should not stop"); 3552 _wallclock.stop(); // to record time 3553 jlong ret = _wallclock.milliseconds(); 3554 _wallclock.start(); // restart 3555 return ret; 3556 } 3557 }; 3558 3559 CMSPhaseAccounting::CMSPhaseAccounting(CMSCollector *collector, 3560 const char *phase, 3561 bool print_cr) : 3562 _collector(collector), _phase(phase), _print_cr(print_cr) { 3563 3564 if (PrintCMSStatistics != 0) { 3565 _collector->resetYields(); 3566 } 3567 if (PrintGCDetails) { 3568 gclog_or_tty->date_stamp(PrintGCDateStamps); 3569 gclog_or_tty->stamp(PrintGCTimeStamps); 3570 gclog_or_tty->print_cr("[%s-concurrent-%s-start]", 3571 _collector->cmsGen()->short_name(), _phase); 3572 } 3573 _collector->resetTimer(); 3574 _wallclock.start(); 3575 _collector->startTimer(); 3576 } 3577 3578 CMSPhaseAccounting::~CMSPhaseAccounting() { 3579 assert(_wallclock.is_active(), "Wall clock should not have stopped"); 3580 _collector->stopTimer(); 3581 _wallclock.stop(); 3582 if (PrintGCDetails) { 3583 gclog_or_tty->date_stamp(PrintGCDateStamps); 3584 gclog_or_tty->stamp(PrintGCTimeStamps); 3585 gclog_or_tty->print("[%s-concurrent-%s: %3.3f/%3.3f secs]", 3586 _collector->cmsGen()->short_name(), 3587 _phase, _collector->timerValue(), _wallclock.seconds()); 3588 if (_print_cr) { 3589 gclog_or_tty->print_cr(""); 3590 } 3591 if (PrintCMSStatistics != 0) { 3592 gclog_or_tty->print_cr(" (CMS-concurrent-%s yielded %d times)", _phase, 3593 _collector->yields()); 3594 } 3595 } 3596 } 3597 3598 // CMS work 3599 3600 // The common parts of CMSParInitialMarkTask and CMSParRemarkTask. 3601 class CMSParMarkTask : public AbstractGangTask { 3602 protected: 3603 CMSCollector* _collector; 3604 int _n_workers; 3605 CMSParMarkTask(const char* name, CMSCollector* collector, int n_workers) : 3606 AbstractGangTask(name), 3607 _collector(collector), 3608 _n_workers(n_workers) {} 3609 // Work method in support of parallel rescan ... of young gen spaces 3610 void do_young_space_rescan(uint worker_id, OopsInGenClosure* cl, 3611 ContiguousSpace* space, 3612 HeapWord** chunk_array, size_t chunk_top); 3613 void work_on_young_gen_roots(uint worker_id, OopsInGenClosure* cl); 3614 }; 3615 3616 // Parallel initial mark task 3617 class CMSParInitialMarkTask: public CMSParMarkTask { 3618 public: 3619 CMSParInitialMarkTask(CMSCollector* collector, int n_workers) : 3620 CMSParMarkTask("Scan roots and young gen for initial mark in parallel", 3621 collector, n_workers) {} 3622 void work(uint worker_id); 3623 }; 3624 3625 // Checkpoint the roots into this generation from outside 3626 // this generation. [Note this initial checkpoint need only 3627 // be approximate -- we'll do a catch up phase subsequently.] 3628 void CMSCollector::checkpointRootsInitial(bool asynch) { 3629 assert(_collectorState == InitialMarking, "Wrong collector state"); 3630 check_correct_thread_executing(); 3631 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 3632 3633 save_heap_summary(); 3634 report_heap_summary(GCWhen::BeforeGC); 3635 3636 ReferenceProcessor* rp = ref_processor(); 3637 SpecializationStats::clear(); 3638 assert(_restart_addr == NULL, "Control point invariant"); 3639 if (asynch) { 3640 // acquire locks for subsequent manipulations 3641 MutexLockerEx x(bitMapLock(), 3642 Mutex::_no_safepoint_check_flag); 3643 checkpointRootsInitialWork(asynch); 3644 // enable ("weak") refs discovery 3645 rp->enable_discovery(true /*verify_disabled*/, true /*check_no_refs*/); 3646 _collectorState = Marking; 3647 } else { 3648 // (Weak) Refs discovery: this is controlled from genCollectedHeap::do_collection 3649 // which recognizes if we are a CMS generation, and doesn't try to turn on 3650 // discovery; verify that they aren't meddling. 3651 assert(!rp->discovery_is_atomic(), 3652 "incorrect setting of discovery predicate"); 3653 assert(!rp->discovery_enabled(), "genCollectedHeap shouldn't control " 3654 "ref discovery for this generation kind"); 3655 // already have locks 3656 checkpointRootsInitialWork(asynch); 3657 // now enable ("weak") refs discovery 3658 rp->enable_discovery(true /*verify_disabled*/, false /*verify_no_refs*/); 3659 _collectorState = Marking; 3660 } 3661 SpecializationStats::print(); 3662 } 3663 3664 void CMSCollector::checkpointRootsInitialWork(bool asynch) { 3665 assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped"); 3666 assert(_collectorState == InitialMarking, "just checking"); 3667 3668 // If there has not been a GC[n-1] since last GC[n] cycle completed, 3669 // precede our marking with a collection of all 3670 // younger generations to keep floating garbage to a minimum. 3671 // XXX: we won't do this for now -- it's an optimization to be done later. 3672 3673 // already have locks 3674 assert_lock_strong(bitMapLock()); 3675 assert(_markBitMap.isAllClear(), "was reset at end of previous cycle"); 3676 3677 // Setup the verification and class unloading state for this 3678 // CMS collection cycle. 3679 setup_cms_unloading_and_verification_state(); 3680 3681 NOT_PRODUCT(GCTraceTime t("\ncheckpointRootsInitialWork", 3682 PrintGCDetails && Verbose, true, _gc_timer_cm);) 3683 if (UseAdaptiveSizePolicy) { 3684 size_policy()->checkpoint_roots_initial_begin(); 3685 } 3686 3687 // Reset all the PLAB chunk arrays if necessary. 3688 if (_survivor_plab_array != NULL && !CMSPLABRecordAlways) { 3689 reset_survivor_plab_arrays(); 3690 } 3691 3692 ResourceMark rm; 3693 HandleMark hm; 3694 3695 FalseClosure falseClosure; 3696 // In the case of a synchronous collection, we will elide the 3697 // remark step, so it's important to catch all the nmethod oops 3698 // in this step. 3699 // The final 'true' flag to gen_process_strong_roots will ensure this. 3700 // If 'async' is true, we can relax the nmethod tracing. 3701 MarkRefsIntoClosure notOlder(_span, &_markBitMap); 3702 GenCollectedHeap* gch = GenCollectedHeap::heap(); 3703 3704 verify_work_stacks_empty(); 3705 verify_overflow_empty(); 3706 3707 gch->ensure_parsability(false); // fill TLABs, but no need to retire them 3708 // Update the saved marks which may affect the root scans. 3709 gch->save_marks(); 3710 3711 // weak reference processing has not started yet. 3712 ref_processor()->set_enqueuing_is_done(false); 3713 3714 // Need to remember all newly created CLDs, 3715 // so that we can guarantee that the remark finds them. 3716 ClassLoaderDataGraph::remember_new_clds(true); 3717 3718 // Whenever a CLD is found, it will be claimed before proceeding to mark 3719 // the klasses. The claimed marks need to be cleared before marking starts. 3720 ClassLoaderDataGraph::clear_claimed_marks(); 3721 3722 if (CMSPrintEdenSurvivorChunks) { 3723 print_eden_and_survivor_chunk_arrays(); 3724 } 3725 3726 { 3727 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;) 3728 if (CMSParallelInitialMarkEnabled && CollectedHeap::use_parallel_gc_threads()) { 3729 // The parallel version. 3730 FlexibleWorkGang* workers = gch->workers(); 3731 assert(workers != NULL, "Need parallel worker threads."); 3732 int n_workers = workers->active_workers(); 3733 CMSParInitialMarkTask tsk(this, n_workers); 3734 gch->set_par_threads(n_workers); 3735 initialize_sequential_subtasks_for_young_gen_rescan(n_workers); 3736 if (n_workers > 1) { 3737 GenCollectedHeap::StrongRootsScope srs(gch); 3738 workers->run_task(&tsk); 3739 } else { 3740 GenCollectedHeap::StrongRootsScope srs(gch); 3741 tsk.work(0); 3742 } 3743 gch->set_par_threads(0); 3744 } else { 3745 // The serial version. 3746 CMKlassClosure klass_closure(¬Older); 3747 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 3748 gch->gen_process_strong_roots(_cmsGen->level(), 3749 true, // younger gens are roots 3750 true, // activate StrongRootsScope 3751 false, // not scavenging 3752 SharedHeap::ScanningOption(roots_scanning_options()), 3753 ¬Older, 3754 true, // walk all of code cache if (so & SO_CodeCache) 3755 NULL, 3756 &klass_closure); 3757 } 3758 } 3759 3760 // Clear mod-union table; it will be dirtied in the prologue of 3761 // CMS generation per each younger generation collection. 3762 3763 assert(_modUnionTable.isAllClear(), 3764 "Was cleared in most recent final checkpoint phase" 3765 " or no bits are set in the gc_prologue before the start of the next " 3766 "subsequent marking phase."); 3767 3768 assert(_ct->klass_rem_set()->mod_union_is_clear(), "Must be"); 3769 3770 // Save the end of the used_region of the constituent generations 3771 // to be used to limit the extent of sweep in each generation. 3772 save_sweep_limits(); 3773 if (UseAdaptiveSizePolicy) { 3774 size_policy()->checkpoint_roots_initial_end(gch->gc_cause()); 3775 } 3776 verify_overflow_empty(); 3777 } 3778 3779 bool CMSCollector::markFromRoots(bool asynch) { 3780 // we might be tempted to assert that: 3781 // assert(asynch == !SafepointSynchronize::is_at_safepoint(), 3782 // "inconsistent argument?"); 3783 // However that wouldn't be right, because it's possible that 3784 // a safepoint is indeed in progress as a younger generation 3785 // stop-the-world GC happens even as we mark in this generation. 3786 assert(_collectorState == Marking, "inconsistent state?"); 3787 check_correct_thread_executing(); 3788 verify_overflow_empty(); 3789 3790 bool res; 3791 if (asynch) { 3792 3793 // Start the timers for adaptive size policy for the concurrent phases 3794 // Do it here so that the foreground MS can use the concurrent 3795 // timer since a foreground MS might has the sweep done concurrently 3796 // or STW. 3797 if (UseAdaptiveSizePolicy) { 3798 size_policy()->concurrent_marking_begin(); 3799 } 3800 3801 // Weak ref discovery note: We may be discovering weak 3802 // refs in this generation concurrent (but interleaved) with 3803 // weak ref discovery by a younger generation collector. 3804 3805 CMSTokenSyncWithLocks ts(true, bitMapLock()); 3806 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 3807 CMSPhaseAccounting pa(this, "mark", !PrintGCDetails); 3808 res = markFromRootsWork(asynch); 3809 if (res) { 3810 _collectorState = Precleaning; 3811 } else { // We failed and a foreground collection wants to take over 3812 assert(_foregroundGCIsActive, "internal state inconsistency"); 3813 assert(_restart_addr == NULL, "foreground will restart from scratch"); 3814 if (PrintGCDetails) { 3815 gclog_or_tty->print_cr("bailing out to foreground collection"); 3816 } 3817 } 3818 if (UseAdaptiveSizePolicy) { 3819 size_policy()->concurrent_marking_end(); 3820 } 3821 } else { 3822 assert(SafepointSynchronize::is_at_safepoint(), 3823 "inconsistent with asynch == false"); 3824 if (UseAdaptiveSizePolicy) { 3825 size_policy()->ms_collection_marking_begin(); 3826 } 3827 // already have locks 3828 res = markFromRootsWork(asynch); 3829 _collectorState = FinalMarking; 3830 if (UseAdaptiveSizePolicy) { 3831 GenCollectedHeap* gch = GenCollectedHeap::heap(); 3832 size_policy()->ms_collection_marking_end(gch->gc_cause()); 3833 } 3834 } 3835 verify_overflow_empty(); 3836 return res; 3837 } 3838 3839 bool CMSCollector::markFromRootsWork(bool asynch) { 3840 // iterate over marked bits in bit map, doing a full scan and mark 3841 // from these roots using the following algorithm: 3842 // . if oop is to the right of the current scan pointer, 3843 // mark corresponding bit (we'll process it later) 3844 // . else (oop is to left of current scan pointer) 3845 // push oop on marking stack 3846 // . drain the marking stack 3847 3848 // Note that when we do a marking step we need to hold the 3849 // bit map lock -- recall that direct allocation (by mutators) 3850 // and promotion (by younger generation collectors) is also 3851 // marking the bit map. [the so-called allocate live policy.] 3852 // Because the implementation of bit map marking is not 3853 // robust wrt simultaneous marking of bits in the same word, 3854 // we need to make sure that there is no such interference 3855 // between concurrent such updates. 3856 3857 // already have locks 3858 assert_lock_strong(bitMapLock()); 3859 3860 verify_work_stacks_empty(); 3861 verify_overflow_empty(); 3862 bool result = false; 3863 if (CMSConcurrentMTEnabled && ConcGCThreads > 0) { 3864 result = do_marking_mt(asynch); 3865 } else { 3866 result = do_marking_st(asynch); 3867 } 3868 return result; 3869 } 3870 3871 // Forward decl 3872 class CMSConcMarkingTask; 3873 3874 class CMSConcMarkingTerminator: public ParallelTaskTerminator { 3875 CMSCollector* _collector; 3876 CMSConcMarkingTask* _task; 3877 public: 3878 virtual void yield(); 3879 3880 // "n_threads" is the number of threads to be terminated. 3881 // "queue_set" is a set of work queues of other threads. 3882 // "collector" is the CMS collector associated with this task terminator. 3883 // "yield" indicates whether we need the gang as a whole to yield. 3884 CMSConcMarkingTerminator(int n_threads, TaskQueueSetSuper* queue_set, CMSCollector* collector) : 3885 ParallelTaskTerminator(n_threads, queue_set), 3886 _collector(collector) { } 3887 3888 void set_task(CMSConcMarkingTask* task) { 3889 _task = task; 3890 } 3891 }; 3892 3893 class CMSConcMarkingTerminatorTerminator: public TerminatorTerminator { 3894 CMSConcMarkingTask* _task; 3895 public: 3896 bool should_exit_termination(); 3897 void set_task(CMSConcMarkingTask* task) { 3898 _task = task; 3899 } 3900 }; 3901 3902 // MT Concurrent Marking Task 3903 class CMSConcMarkingTask: public YieldingFlexibleGangTask { 3904 CMSCollector* _collector; 3905 int _n_workers; // requested/desired # workers 3906 bool _asynch; 3907 bool _result; 3908 CompactibleFreeListSpace* _cms_space; 3909 char _pad_front[64]; // padding to ... 3910 HeapWord* _global_finger; // ... avoid sharing cache line 3911 char _pad_back[64]; 3912 HeapWord* _restart_addr; 3913 3914 // Exposed here for yielding support 3915 Mutex* const _bit_map_lock; 3916 3917 // The per thread work queues, available here for stealing 3918 OopTaskQueueSet* _task_queues; 3919 3920 // Termination (and yielding) support 3921 CMSConcMarkingTerminator _term; 3922 CMSConcMarkingTerminatorTerminator _term_term; 3923 3924 public: 3925 CMSConcMarkingTask(CMSCollector* collector, 3926 CompactibleFreeListSpace* cms_space, 3927 bool asynch, 3928 YieldingFlexibleWorkGang* workers, 3929 OopTaskQueueSet* task_queues): 3930 YieldingFlexibleGangTask("Concurrent marking done multi-threaded"), 3931 _collector(collector), 3932 _cms_space(cms_space), 3933 _asynch(asynch), _n_workers(0), _result(true), 3934 _task_queues(task_queues), 3935 _term(_n_workers, task_queues, _collector), 3936 _bit_map_lock(collector->bitMapLock()) 3937 { 3938 _requested_size = _n_workers; 3939 _term.set_task(this); 3940 _term_term.set_task(this); 3941 _restart_addr = _global_finger = _cms_space->bottom(); 3942 } 3943 3944 3945 OopTaskQueueSet* task_queues() { return _task_queues; } 3946 3947 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 3948 3949 HeapWord** global_finger_addr() { return &_global_finger; } 3950 3951 CMSConcMarkingTerminator* terminator() { return &_term; } 3952 3953 virtual void set_for_termination(int active_workers) { 3954 terminator()->reset_for_reuse(active_workers); 3955 } 3956 3957 void work(uint worker_id); 3958 bool should_yield() { 3959 return ConcurrentMarkSweepThread::should_yield() 3960 && !_collector->foregroundGCIsActive() 3961 && _asynch; 3962 } 3963 3964 virtual void coordinator_yield(); // stuff done by coordinator 3965 bool result() { return _result; } 3966 3967 void reset(HeapWord* ra) { 3968 assert(_global_finger >= _cms_space->end(), "Postcondition of ::work(i)"); 3969 _restart_addr = _global_finger = ra; 3970 _term.reset_for_reuse(); 3971 } 3972 3973 static bool get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 3974 OopTaskQueue* work_q); 3975 3976 private: 3977 void do_scan_and_mark(int i, CompactibleFreeListSpace* sp); 3978 void do_work_steal(int i); 3979 void bump_global_finger(HeapWord* f); 3980 }; 3981 3982 bool CMSConcMarkingTerminatorTerminator::should_exit_termination() { 3983 assert(_task != NULL, "Error"); 3984 return _task->yielding(); 3985 // Note that we do not need the disjunct || _task->should_yield() above 3986 // because we want terminating threads to yield only if the task 3987 // is already in the midst of yielding, which happens only after at least one 3988 // thread has yielded. 3989 } 3990 3991 void CMSConcMarkingTerminator::yield() { 3992 if (_task->should_yield()) { 3993 _task->yield(); 3994 } else { 3995 ParallelTaskTerminator::yield(); 3996 } 3997 } 3998 3999 //////////////////////////////////////////////////////////////// 4000 // Concurrent Marking Algorithm Sketch 4001 //////////////////////////////////////////////////////////////// 4002 // Until all tasks exhausted (both spaces): 4003 // -- claim next available chunk 4004 // -- bump global finger via CAS 4005 // -- find first object that starts in this chunk 4006 // and start scanning bitmap from that position 4007 // -- scan marked objects for oops 4008 // -- CAS-mark target, and if successful: 4009 // . if target oop is above global finger (volatile read) 4010 // nothing to do 4011 // . if target oop is in chunk and above local finger 4012 // then nothing to do 4013 // . else push on work-queue 4014 // -- Deal with possible overflow issues: 4015 // . local work-queue overflow causes stuff to be pushed on 4016 // global (common) overflow queue 4017 // . always first empty local work queue 4018 // . then get a batch of oops from global work queue if any 4019 // . then do work stealing 4020 // -- When all tasks claimed (both spaces) 4021 // and local work queue empty, 4022 // then in a loop do: 4023 // . check global overflow stack; steal a batch of oops and trace 4024 // . try to steal from other threads oif GOS is empty 4025 // . if neither is available, offer termination 4026 // -- Terminate and return result 4027 // 4028 void CMSConcMarkingTask::work(uint worker_id) { 4029 elapsedTimer _timer; 4030 ResourceMark rm; 4031 HandleMark hm; 4032 4033 DEBUG_ONLY(_collector->verify_overflow_empty();) 4034 4035 // Before we begin work, our work queue should be empty 4036 assert(work_queue(worker_id)->size() == 0, "Expected to be empty"); 4037 // Scan the bitmap covering _cms_space, tracing through grey objects. 4038 _timer.start(); 4039 do_scan_and_mark(worker_id, _cms_space); 4040 _timer.stop(); 4041 if (PrintCMSStatistics != 0) { 4042 gclog_or_tty->print_cr("Finished cms space scanning in %dth thread: %3.3f sec", 4043 worker_id, _timer.seconds()); 4044 // XXX: need xxx/xxx type of notation, two timers 4045 } 4046 4047 // ... do work stealing 4048 _timer.reset(); 4049 _timer.start(); 4050 do_work_steal(worker_id); 4051 _timer.stop(); 4052 if (PrintCMSStatistics != 0) { 4053 gclog_or_tty->print_cr("Finished work stealing in %dth thread: %3.3f sec", 4054 worker_id, _timer.seconds()); 4055 // XXX: need xxx/xxx type of notation, two timers 4056 } 4057 assert(_collector->_markStack.isEmpty(), "Should have been emptied"); 4058 assert(work_queue(worker_id)->size() == 0, "Should have been emptied"); 4059 // Note that under the current task protocol, the 4060 // following assertion is true even of the spaces 4061 // expanded since the completion of the concurrent 4062 // marking. XXX This will likely change under a strict 4063 // ABORT semantics. 4064 // After perm removal the comparison was changed to 4065 // greater than or equal to from strictly greater than. 4066 // Before perm removal the highest address sweep would 4067 // have been at the end of perm gen but now is at the 4068 // end of the tenured gen. 4069 assert(_global_finger >= _cms_space->end(), 4070 "All tasks have been completed"); 4071 DEBUG_ONLY(_collector->verify_overflow_empty();) 4072 } 4073 4074 void CMSConcMarkingTask::bump_global_finger(HeapWord* f) { 4075 HeapWord* read = _global_finger; 4076 HeapWord* cur = read; 4077 while (f > read) { 4078 cur = read; 4079 read = (HeapWord*) Atomic::cmpxchg_ptr(f, &_global_finger, cur); 4080 if (cur == read) { 4081 // our cas succeeded 4082 assert(_global_finger >= f, "protocol consistency"); 4083 break; 4084 } 4085 } 4086 } 4087 4088 // This is really inefficient, and should be redone by 4089 // using (not yet available) block-read and -write interfaces to the 4090 // stack and the work_queue. XXX FIX ME !!! 4091 bool CMSConcMarkingTask::get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 4092 OopTaskQueue* work_q) { 4093 // Fast lock-free check 4094 if (ovflw_stk->length() == 0) { 4095 return false; 4096 } 4097 assert(work_q->size() == 0, "Shouldn't steal"); 4098 MutexLockerEx ml(ovflw_stk->par_lock(), 4099 Mutex::_no_safepoint_check_flag); 4100 // Grab up to 1/4 the size of the work queue 4101 size_t num = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 4102 (size_t)ParGCDesiredObjsFromOverflowList); 4103 num = MIN2(num, ovflw_stk->length()); 4104 for (int i = (int) num; i > 0; i--) { 4105 oop cur = ovflw_stk->pop(); 4106 assert(cur != NULL, "Counted wrong?"); 4107 work_q->push(cur); 4108 } 4109 return num > 0; 4110 } 4111 4112 void CMSConcMarkingTask::do_scan_and_mark(int i, CompactibleFreeListSpace* sp) { 4113 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 4114 int n_tasks = pst->n_tasks(); 4115 // We allow that there may be no tasks to do here because 4116 // we are restarting after a stack overflow. 4117 assert(pst->valid() || n_tasks == 0, "Uninitialized use?"); 4118 uint nth_task = 0; 4119 4120 HeapWord* aligned_start = sp->bottom(); 4121 if (sp->used_region().contains(_restart_addr)) { 4122 // Align down to a card boundary for the start of 0th task 4123 // for this space. 4124 aligned_start = 4125 (HeapWord*)align_size_down((uintptr_t)_restart_addr, 4126 CardTableModRefBS::card_size); 4127 } 4128 4129 size_t chunk_size = sp->marking_task_size(); 4130 while (!pst->is_task_claimed(/* reference */ nth_task)) { 4131 // Having claimed the nth task in this space, 4132 // compute the chunk that it corresponds to: 4133 MemRegion span = MemRegion(aligned_start + nth_task*chunk_size, 4134 aligned_start + (nth_task+1)*chunk_size); 4135 // Try and bump the global finger via a CAS; 4136 // note that we need to do the global finger bump 4137 // _before_ taking the intersection below, because 4138 // the task corresponding to that region will be 4139 // deemed done even if the used_region() expands 4140 // because of allocation -- as it almost certainly will 4141 // during start-up while the threads yield in the 4142 // closure below. 4143 HeapWord* finger = span.end(); 4144 bump_global_finger(finger); // atomically 4145 // There are null tasks here corresponding to chunks 4146 // beyond the "top" address of the space. 4147 span = span.intersection(sp->used_region()); 4148 if (!span.is_empty()) { // Non-null task 4149 HeapWord* prev_obj; 4150 assert(!span.contains(_restart_addr) || nth_task == 0, 4151 "Inconsistency"); 4152 if (nth_task == 0) { 4153 // For the 0th task, we'll not need to compute a block_start. 4154 if (span.contains(_restart_addr)) { 4155 // In the case of a restart because of stack overflow, 4156 // we might additionally skip a chunk prefix. 4157 prev_obj = _restart_addr; 4158 } else { 4159 prev_obj = span.start(); 4160 } 4161 } else { 4162 // We want to skip the first object because 4163 // the protocol is to scan any object in its entirety 4164 // that _starts_ in this span; a fortiori, any 4165 // object starting in an earlier span is scanned 4166 // as part of an earlier claimed task. 4167 // Below we use the "careful" version of block_start 4168 // so we do not try to navigate uninitialized objects. 4169 prev_obj = sp->block_start_careful(span.start()); 4170 // Below we use a variant of block_size that uses the 4171 // Printezis bits to avoid waiting for allocated 4172 // objects to become initialized/parsable. 4173 while (prev_obj < span.start()) { 4174 size_t sz = sp->block_size_no_stall(prev_obj, _collector); 4175 if (sz > 0) { 4176 prev_obj += sz; 4177 } else { 4178 // In this case we may end up doing a bit of redundant 4179 // scanning, but that appears unavoidable, short of 4180 // locking the free list locks; see bug 6324141. 4181 break; 4182 } 4183 } 4184 } 4185 if (prev_obj < span.end()) { 4186 MemRegion my_span = MemRegion(prev_obj, span.end()); 4187 // Do the marking work within a non-empty span -- 4188 // the last argument to the constructor indicates whether the 4189 // iteration should be incremental with periodic yields. 4190 Par_MarkFromRootsClosure cl(this, _collector, my_span, 4191 &_collector->_markBitMap, 4192 work_queue(i), 4193 &_collector->_markStack, 4194 _asynch); 4195 _collector->_markBitMap.iterate(&cl, my_span.start(), my_span.end()); 4196 } // else nothing to do for this task 4197 } // else nothing to do for this task 4198 } 4199 // We'd be tempted to assert here that since there are no 4200 // more tasks left to claim in this space, the global_finger 4201 // must exceed space->top() and a fortiori space->end(). However, 4202 // that would not quite be correct because the bumping of 4203 // global_finger occurs strictly after the claiming of a task, 4204 // so by the time we reach here the global finger may not yet 4205 // have been bumped up by the thread that claimed the last 4206 // task. 4207 pst->all_tasks_completed(); 4208 } 4209 4210 class Par_ConcMarkingClosure: public CMSOopClosure { 4211 private: 4212 CMSCollector* _collector; 4213 CMSConcMarkingTask* _task; 4214 MemRegion _span; 4215 CMSBitMap* _bit_map; 4216 CMSMarkStack* _overflow_stack; 4217 OopTaskQueue* _work_queue; 4218 protected: 4219 DO_OOP_WORK_DEFN 4220 public: 4221 Par_ConcMarkingClosure(CMSCollector* collector, CMSConcMarkingTask* task, OopTaskQueue* work_queue, 4222 CMSBitMap* bit_map, CMSMarkStack* overflow_stack): 4223 CMSOopClosure(collector->ref_processor()), 4224 _collector(collector), 4225 _task(task), 4226 _span(collector->_span), 4227 _work_queue(work_queue), 4228 _bit_map(bit_map), 4229 _overflow_stack(overflow_stack) 4230 { } 4231 virtual void do_oop(oop* p); 4232 virtual void do_oop(narrowOop* p); 4233 4234 void trim_queue(size_t max); 4235 void handle_stack_overflow(HeapWord* lost); 4236 void do_yield_check() { 4237 if (_task->should_yield()) { 4238 _task->yield(); 4239 } 4240 } 4241 }; 4242 4243 // Grey object scanning during work stealing phase -- 4244 // the salient assumption here is that any references 4245 // that are in these stolen objects being scanned must 4246 // already have been initialized (else they would not have 4247 // been published), so we do not need to check for 4248 // uninitialized objects before pushing here. 4249 void Par_ConcMarkingClosure::do_oop(oop obj) { 4250 assert(obj->is_oop_or_null(true), "expected an oop or NULL"); 4251 HeapWord* addr = (HeapWord*)obj; 4252 // Check if oop points into the CMS generation 4253 // and is not marked 4254 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 4255 // a white object ... 4256 // If we manage to "claim" the object, by being the 4257 // first thread to mark it, then we push it on our 4258 // marking stack 4259 if (_bit_map->par_mark(addr)) { // ... now grey 4260 // push on work queue (grey set) 4261 bool simulate_overflow = false; 4262 NOT_PRODUCT( 4263 if (CMSMarkStackOverflowALot && 4264 _collector->simulate_overflow()) { 4265 // simulate a stack overflow 4266 simulate_overflow = true; 4267 } 4268 ) 4269 if (simulate_overflow || 4270 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) { 4271 // stack overflow 4272 if (PrintCMSStatistics != 0) { 4273 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 4274 SIZE_FORMAT, _overflow_stack->capacity()); 4275 } 4276 // We cannot assert that the overflow stack is full because 4277 // it may have been emptied since. 4278 assert(simulate_overflow || 4279 _work_queue->size() == _work_queue->max_elems(), 4280 "Else push should have succeeded"); 4281 handle_stack_overflow(addr); 4282 } 4283 } // Else, some other thread got there first 4284 do_yield_check(); 4285 } 4286 } 4287 4288 void Par_ConcMarkingClosure::do_oop(oop* p) { Par_ConcMarkingClosure::do_oop_work(p); } 4289 void Par_ConcMarkingClosure::do_oop(narrowOop* p) { Par_ConcMarkingClosure::do_oop_work(p); } 4290 4291 void Par_ConcMarkingClosure::trim_queue(size_t max) { 4292 while (_work_queue->size() > max) { 4293 oop new_oop; 4294 if (_work_queue->pop_local(new_oop)) { 4295 assert(new_oop->is_oop(), "Should be an oop"); 4296 assert(_bit_map->isMarked((HeapWord*)new_oop), "Grey object"); 4297 assert(_span.contains((HeapWord*)new_oop), "Not in span"); 4298 new_oop->oop_iterate(this); // do_oop() above 4299 do_yield_check(); 4300 } 4301 } 4302 } 4303 4304 // Upon stack overflow, we discard (part of) the stack, 4305 // remembering the least address amongst those discarded 4306 // in CMSCollector's _restart_address. 4307 void Par_ConcMarkingClosure::handle_stack_overflow(HeapWord* lost) { 4308 // We need to do this under a mutex to prevent other 4309 // workers from interfering with the work done below. 4310 MutexLockerEx ml(_overflow_stack->par_lock(), 4311 Mutex::_no_safepoint_check_flag); 4312 // Remember the least grey address discarded 4313 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 4314 _collector->lower_restart_addr(ra); 4315 _overflow_stack->reset(); // discard stack contents 4316 _overflow_stack->expand(); // expand the stack if possible 4317 } 4318 4319 4320 void CMSConcMarkingTask::do_work_steal(int i) { 4321 OopTaskQueue* work_q = work_queue(i); 4322 oop obj_to_scan; 4323 CMSBitMap* bm = &(_collector->_markBitMap); 4324 CMSMarkStack* ovflw = &(_collector->_markStack); 4325 int* seed = _collector->hash_seed(i); 4326 Par_ConcMarkingClosure cl(_collector, this, work_q, bm, ovflw); 4327 while (true) { 4328 cl.trim_queue(0); 4329 assert(work_q->size() == 0, "Should have been emptied above"); 4330 if (get_work_from_overflow_stack(ovflw, work_q)) { 4331 // Can't assert below because the work obtained from the 4332 // overflow stack may already have been stolen from us. 4333 // assert(work_q->size() > 0, "Work from overflow stack"); 4334 continue; 4335 } else if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 4336 assert(obj_to_scan->is_oop(), "Should be an oop"); 4337 assert(bm->isMarked((HeapWord*)obj_to_scan), "Grey object"); 4338 obj_to_scan->oop_iterate(&cl); 4339 } else if (terminator()->offer_termination(&_term_term)) { 4340 assert(work_q->size() == 0, "Impossible!"); 4341 break; 4342 } else if (yielding() || should_yield()) { 4343 yield(); 4344 } 4345 } 4346 } 4347 4348 // This is run by the CMS (coordinator) thread. 4349 void CMSConcMarkingTask::coordinator_yield() { 4350 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 4351 "CMS thread should hold CMS token"); 4352 // First give up the locks, then yield, then re-lock 4353 // We should probably use a constructor/destructor idiom to 4354 // do this unlock/lock or modify the MutexUnlocker class to 4355 // serve our purpose. XXX 4356 assert_lock_strong(_bit_map_lock); 4357 _bit_map_lock->unlock(); 4358 ConcurrentMarkSweepThread::desynchronize(true); 4359 ConcurrentMarkSweepThread::acknowledge_yield_request(); 4360 _collector->stopTimer(); 4361 if (PrintCMSStatistics != 0) { 4362 _collector->incrementYields(); 4363 } 4364 _collector->icms_wait(); 4365 4366 // It is possible for whichever thread initiated the yield request 4367 // not to get a chance to wake up and take the bitmap lock between 4368 // this thread releasing it and reacquiring it. So, while the 4369 // should_yield() flag is on, let's sleep for a bit to give the 4370 // other thread a chance to wake up. The limit imposed on the number 4371 // of iterations is defensive, to avoid any unforseen circumstances 4372 // putting us into an infinite loop. Since it's always been this 4373 // (coordinator_yield()) method that was observed to cause the 4374 // problem, we are using a parameter (CMSCoordinatorYieldSleepCount) 4375 // which is by default non-zero. For the other seven methods that 4376 // also perform the yield operation, as are using a different 4377 // parameter (CMSYieldSleepCount) which is by default zero. This way we 4378 // can enable the sleeping for those methods too, if necessary. 4379 // See 6442774. 4380 // 4381 // We really need to reconsider the synchronization between the GC 4382 // thread and the yield-requesting threads in the future and we 4383 // should really use wait/notify, which is the recommended 4384 // way of doing this type of interaction. Additionally, we should 4385 // consolidate the eight methods that do the yield operation and they 4386 // are almost identical into one for better maintenability and 4387 // readability. See 6445193. 4388 // 4389 // Tony 2006.06.29 4390 for (unsigned i = 0; i < CMSCoordinatorYieldSleepCount && 4391 ConcurrentMarkSweepThread::should_yield() && 4392 !CMSCollector::foregroundGCIsActive(); ++i) { 4393 os::sleep(Thread::current(), 1, false); 4394 ConcurrentMarkSweepThread::acknowledge_yield_request(); 4395 } 4396 4397 ConcurrentMarkSweepThread::synchronize(true); 4398 _bit_map_lock->lock_without_safepoint_check(); 4399 _collector->startTimer(); 4400 } 4401 4402 bool CMSCollector::do_marking_mt(bool asynch) { 4403 assert(ConcGCThreads > 0 && conc_workers() != NULL, "precondition"); 4404 int num_workers = AdaptiveSizePolicy::calc_active_conc_workers( 4405 conc_workers()->total_workers(), 4406 conc_workers()->active_workers(), 4407 Threads::number_of_non_daemon_threads()); 4408 conc_workers()->set_active_workers(num_workers); 4409 4410 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 4411 4412 CMSConcMarkingTask tsk(this, 4413 cms_space, 4414 asynch, 4415 conc_workers(), 4416 task_queues()); 4417 4418 // Since the actual number of workers we get may be different 4419 // from the number we requested above, do we need to do anything different 4420 // below? In particular, may be we need to subclass the SequantialSubTasksDone 4421 // class?? XXX 4422 cms_space ->initialize_sequential_subtasks_for_marking(num_workers); 4423 4424 // Refs discovery is already non-atomic. 4425 assert(!ref_processor()->discovery_is_atomic(), "Should be non-atomic"); 4426 assert(ref_processor()->discovery_is_mt(), "Discovery should be MT"); 4427 conc_workers()->start_task(&tsk); 4428 while (tsk.yielded()) { 4429 tsk.coordinator_yield(); 4430 conc_workers()->continue_task(&tsk); 4431 } 4432 // If the task was aborted, _restart_addr will be non-NULL 4433 assert(tsk.completed() || _restart_addr != NULL, "Inconsistency"); 4434 while (_restart_addr != NULL) { 4435 // XXX For now we do not make use of ABORTED state and have not 4436 // yet implemented the right abort semantics (even in the original 4437 // single-threaded CMS case). That needs some more investigation 4438 // and is deferred for now; see CR# TBF. 07252005YSR. XXX 4439 assert(!CMSAbortSemantics || tsk.aborted(), "Inconsistency"); 4440 // If _restart_addr is non-NULL, a marking stack overflow 4441 // occurred; we need to do a fresh marking iteration from the 4442 // indicated restart address. 4443 if (_foregroundGCIsActive && asynch) { 4444 // We may be running into repeated stack overflows, having 4445 // reached the limit of the stack size, while making very 4446 // slow forward progress. It may be best to bail out and 4447 // let the foreground collector do its job. 4448 // Clear _restart_addr, so that foreground GC 4449 // works from scratch. This avoids the headache of 4450 // a "rescan" which would otherwise be needed because 4451 // of the dirty mod union table & card table. 4452 _restart_addr = NULL; 4453 return false; 4454 } 4455 // Adjust the task to restart from _restart_addr 4456 tsk.reset(_restart_addr); 4457 cms_space ->initialize_sequential_subtasks_for_marking(num_workers, 4458 _restart_addr); 4459 _restart_addr = NULL; 4460 // Get the workers going again 4461 conc_workers()->start_task(&tsk); 4462 while (tsk.yielded()) { 4463 tsk.coordinator_yield(); 4464 conc_workers()->continue_task(&tsk); 4465 } 4466 } 4467 assert(tsk.completed(), "Inconsistency"); 4468 assert(tsk.result() == true, "Inconsistency"); 4469 return true; 4470 } 4471 4472 bool CMSCollector::do_marking_st(bool asynch) { 4473 ResourceMark rm; 4474 HandleMark hm; 4475 4476 // Temporarily make refs discovery single threaded (non-MT) 4477 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false); 4478 MarkFromRootsClosure markFromRootsClosure(this, _span, &_markBitMap, 4479 &_markStack, CMSYield && asynch); 4480 // the last argument to iterate indicates whether the iteration 4481 // should be incremental with periodic yields. 4482 _markBitMap.iterate(&markFromRootsClosure); 4483 // If _restart_addr is non-NULL, a marking stack overflow 4484 // occurred; we need to do a fresh iteration from the 4485 // indicated restart address. 4486 while (_restart_addr != NULL) { 4487 if (_foregroundGCIsActive && asynch) { 4488 // We may be running into repeated stack overflows, having 4489 // reached the limit of the stack size, while making very 4490 // slow forward progress. It may be best to bail out and 4491 // let the foreground collector do its job. 4492 // Clear _restart_addr, so that foreground GC 4493 // works from scratch. This avoids the headache of 4494 // a "rescan" which would otherwise be needed because 4495 // of the dirty mod union table & card table. 4496 _restart_addr = NULL; 4497 return false; // indicating failure to complete marking 4498 } 4499 // Deal with stack overflow: 4500 // we restart marking from _restart_addr 4501 HeapWord* ra = _restart_addr; 4502 markFromRootsClosure.reset(ra); 4503 _restart_addr = NULL; 4504 _markBitMap.iterate(&markFromRootsClosure, ra, _span.end()); 4505 } 4506 return true; 4507 } 4508 4509 void CMSCollector::preclean() { 4510 check_correct_thread_executing(); 4511 assert(Thread::current()->is_ConcurrentGC_thread(), "Wrong thread"); 4512 verify_work_stacks_empty(); 4513 verify_overflow_empty(); 4514 _abort_preclean = false; 4515 if (CMSPrecleaningEnabled) { 4516 if (!CMSEdenChunksRecordAlways) { 4517 _eden_chunk_index = 0; 4518 } 4519 size_t used = get_eden_used(); 4520 size_t capacity = get_eden_capacity(); 4521 // Don't start sampling unless we will get sufficiently 4522 // many samples. 4523 if (used < (capacity/(CMSScheduleRemarkSamplingRatio * 100) 4524 * CMSScheduleRemarkEdenPenetration)) { 4525 _start_sampling = true; 4526 } else { 4527 _start_sampling = false; 4528 } 4529 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 4530 CMSPhaseAccounting pa(this, "preclean", !PrintGCDetails); 4531 preclean_work(CMSPrecleanRefLists1, CMSPrecleanSurvivors1); 4532 } 4533 CMSTokenSync x(true); // is cms thread 4534 if (CMSPrecleaningEnabled) { 4535 sample_eden(); 4536 _collectorState = AbortablePreclean; 4537 } else { 4538 _collectorState = FinalMarking; 4539 } 4540 verify_work_stacks_empty(); 4541 verify_overflow_empty(); 4542 } 4543 4544 // Try and schedule the remark such that young gen 4545 // occupancy is CMSScheduleRemarkEdenPenetration %. 4546 void CMSCollector::abortable_preclean() { 4547 check_correct_thread_executing(); 4548 assert(CMSPrecleaningEnabled, "Inconsistent control state"); 4549 assert(_collectorState == AbortablePreclean, "Inconsistent control state"); 4550 4551 // If Eden's current occupancy is below this threshold, 4552 // immediately schedule the remark; else preclean 4553 // past the next scavenge in an effort to 4554 // schedule the pause as described avove. By choosing 4555 // CMSScheduleRemarkEdenSizeThreshold >= max eden size 4556 // we will never do an actual abortable preclean cycle. 4557 if (get_eden_used() > CMSScheduleRemarkEdenSizeThreshold) { 4558 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 4559 CMSPhaseAccounting pa(this, "abortable-preclean", !PrintGCDetails); 4560 // We need more smarts in the abortable preclean 4561 // loop below to deal with cases where allocation 4562 // in young gen is very very slow, and our precleaning 4563 // is running a losing race against a horde of 4564 // mutators intent on flooding us with CMS updates 4565 // (dirty cards). 4566 // One, admittedly dumb, strategy is to give up 4567 // after a certain number of abortable precleaning loops 4568 // or after a certain maximum time. We want to make 4569 // this smarter in the next iteration. 4570 // XXX FIX ME!!! YSR 4571 size_t loops = 0, workdone = 0, cumworkdone = 0, waited = 0; 4572 while (!(should_abort_preclean() || 4573 ConcurrentMarkSweepThread::should_terminate())) { 4574 workdone = preclean_work(CMSPrecleanRefLists2, CMSPrecleanSurvivors2); 4575 cumworkdone += workdone; 4576 loops++; 4577 // Voluntarily terminate abortable preclean phase if we have 4578 // been at it for too long. 4579 if ((CMSMaxAbortablePrecleanLoops != 0) && 4580 loops >= CMSMaxAbortablePrecleanLoops) { 4581 if (PrintGCDetails) { 4582 gclog_or_tty->print(" CMS: abort preclean due to loops "); 4583 } 4584 break; 4585 } 4586 if (pa.wallclock_millis() > CMSMaxAbortablePrecleanTime) { 4587 if (PrintGCDetails) { 4588 gclog_or_tty->print(" CMS: abort preclean due to time "); 4589 } 4590 break; 4591 } 4592 // If we are doing little work each iteration, we should 4593 // take a short break. 4594 if (workdone < CMSAbortablePrecleanMinWorkPerIteration) { 4595 // Sleep for some time, waiting for work to accumulate 4596 stopTimer(); 4597 cmsThread()->wait_on_cms_lock(CMSAbortablePrecleanWaitMillis); 4598 startTimer(); 4599 waited++; 4600 } 4601 } 4602 if (PrintCMSStatistics > 0) { 4603 gclog_or_tty->print(" [%d iterations, %d waits, %d cards)] ", 4604 loops, waited, cumworkdone); 4605 } 4606 } 4607 CMSTokenSync x(true); // is cms thread 4608 if (_collectorState != Idling) { 4609 assert(_collectorState == AbortablePreclean, 4610 "Spontaneous state transition?"); 4611 _collectorState = FinalMarking; 4612 } // Else, a foreground collection completed this CMS cycle. 4613 return; 4614 } 4615 4616 // Respond to an Eden sampling opportunity 4617 void CMSCollector::sample_eden() { 4618 // Make sure a young gc cannot sneak in between our 4619 // reading and recording of a sample. 4620 assert(Thread::current()->is_ConcurrentGC_thread(), 4621 "Only the cms thread may collect Eden samples"); 4622 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 4623 "Should collect samples while holding CMS token"); 4624 if (!_start_sampling) { 4625 return; 4626 } 4627 // When CMSEdenChunksRecordAlways is true, the eden chunk array 4628 // is populated by the young generation. 4629 if (_eden_chunk_array != NULL && !CMSEdenChunksRecordAlways) { 4630 if (_eden_chunk_index < _eden_chunk_capacity) { 4631 _eden_chunk_array[_eden_chunk_index] = *_top_addr; // take sample 4632 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr, 4633 "Unexpected state of Eden"); 4634 // We'd like to check that what we just sampled is an oop-start address; 4635 // however, we cannot do that here since the object may not yet have been 4636 // initialized. So we'll instead do the check when we _use_ this sample 4637 // later. 4638 if (_eden_chunk_index == 0 || 4639 (pointer_delta(_eden_chunk_array[_eden_chunk_index], 4640 _eden_chunk_array[_eden_chunk_index-1]) 4641 >= CMSSamplingGrain)) { 4642 _eden_chunk_index++; // commit sample 4643 } 4644 } 4645 } 4646 if ((_collectorState == AbortablePreclean) && !_abort_preclean) { 4647 size_t used = get_eden_used(); 4648 size_t capacity = get_eden_capacity(); 4649 assert(used <= capacity, "Unexpected state of Eden"); 4650 if (used > (capacity/100 * CMSScheduleRemarkEdenPenetration)) { 4651 _abort_preclean = true; 4652 } 4653 } 4654 } 4655 4656 4657 size_t CMSCollector::preclean_work(bool clean_refs, bool clean_survivor) { 4658 assert(_collectorState == Precleaning || 4659 _collectorState == AbortablePreclean, "incorrect state"); 4660 ResourceMark rm; 4661 HandleMark hm; 4662 4663 // Precleaning is currently not MT but the reference processor 4664 // may be set for MT. Disable it temporarily here. 4665 ReferenceProcessor* rp = ref_processor(); 4666 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(rp, false); 4667 4668 // Do one pass of scrubbing the discovered reference lists 4669 // to remove any reference objects with strongly-reachable 4670 // referents. 4671 if (clean_refs) { 4672 CMSPrecleanRefsYieldClosure yield_cl(this); 4673 assert(rp->span().equals(_span), "Spans should be equal"); 4674 CMSKeepAliveClosure keep_alive(this, _span, &_markBitMap, 4675 &_markStack, true /* preclean */); 4676 CMSDrainMarkingStackClosure complete_trace(this, 4677 _span, &_markBitMap, &_markStack, 4678 &keep_alive, true /* preclean */); 4679 4680 // We don't want this step to interfere with a young 4681 // collection because we don't want to take CPU 4682 // or memory bandwidth away from the young GC threads 4683 // (which may be as many as there are CPUs). 4684 // Note that we don't need to protect ourselves from 4685 // interference with mutators because they can't 4686 // manipulate the discovered reference lists nor affect 4687 // the computed reachability of the referents, the 4688 // only properties manipulated by the precleaning 4689 // of these reference lists. 4690 stopTimer(); 4691 CMSTokenSyncWithLocks x(true /* is cms thread */, 4692 bitMapLock()); 4693 startTimer(); 4694 sample_eden(); 4695 4696 // The following will yield to allow foreground 4697 // collection to proceed promptly. XXX YSR: 4698 // The code in this method may need further 4699 // tweaking for better performance and some restructuring 4700 // for cleaner interfaces. 4701 GCTimer *gc_timer = NULL; // Currently not tracing concurrent phases 4702 rp->preclean_discovered_references( 4703 rp->is_alive_non_header(), &keep_alive, &complete_trace, &yield_cl, 4704 gc_timer); 4705 } 4706 4707 if (clean_survivor) { // preclean the active survivor space(s) 4708 assert(_young_gen->kind() == Generation::DefNew || 4709 _young_gen->kind() == Generation::ParNew || 4710 _young_gen->kind() == Generation::ASParNew, 4711 "incorrect type for cast"); 4712 DefNewGeneration* dng = (DefNewGeneration*)_young_gen; 4713 PushAndMarkClosure pam_cl(this, _span, ref_processor(), 4714 &_markBitMap, &_modUnionTable, 4715 &_markStack, true /* precleaning phase */); 4716 stopTimer(); 4717 CMSTokenSyncWithLocks ts(true /* is cms thread */, 4718 bitMapLock()); 4719 startTimer(); 4720 unsigned int before_count = 4721 GenCollectedHeap::heap()->total_collections(); 4722 SurvivorSpacePrecleanClosure 4723 sss_cl(this, _span, &_markBitMap, &_markStack, 4724 &pam_cl, before_count, CMSYield); 4725 dng->from()->object_iterate_careful(&sss_cl); 4726 dng->to()->object_iterate_careful(&sss_cl); 4727 } 4728 MarkRefsIntoAndScanClosure 4729 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable, 4730 &_markStack, this, CMSYield, 4731 true /* precleaning phase */); 4732 // CAUTION: The following closure has persistent state that may need to 4733 // be reset upon a decrease in the sequence of addresses it 4734 // processes. 4735 ScanMarkedObjectsAgainCarefullyClosure 4736 smoac_cl(this, _span, 4737 &_markBitMap, &_markStack, &mrias_cl, CMSYield); 4738 4739 // Preclean dirty cards in ModUnionTable and CardTable using 4740 // appropriate convergence criterion; 4741 // repeat CMSPrecleanIter times unless we find that 4742 // we are losing. 4743 assert(CMSPrecleanIter < 10, "CMSPrecleanIter is too large"); 4744 assert(CMSPrecleanNumerator < CMSPrecleanDenominator, 4745 "Bad convergence multiplier"); 4746 assert(CMSPrecleanThreshold >= 100, 4747 "Unreasonably low CMSPrecleanThreshold"); 4748 4749 size_t numIter, cumNumCards, lastNumCards, curNumCards; 4750 for (numIter = 0, cumNumCards = lastNumCards = curNumCards = 0; 4751 numIter < CMSPrecleanIter; 4752 numIter++, lastNumCards = curNumCards, cumNumCards += curNumCards) { 4753 curNumCards = preclean_mod_union_table(_cmsGen, &smoac_cl); 4754 if (Verbose && PrintGCDetails) { 4755 gclog_or_tty->print(" (modUnionTable: %d cards)", curNumCards); 4756 } 4757 // Either there are very few dirty cards, so re-mark 4758 // pause will be small anyway, or our pre-cleaning isn't 4759 // that much faster than the rate at which cards are being 4760 // dirtied, so we might as well stop and re-mark since 4761 // precleaning won't improve our re-mark time by much. 4762 if (curNumCards <= CMSPrecleanThreshold || 4763 (numIter > 0 && 4764 (curNumCards * CMSPrecleanDenominator > 4765 lastNumCards * CMSPrecleanNumerator))) { 4766 numIter++; 4767 cumNumCards += curNumCards; 4768 break; 4769 } 4770 } 4771 4772 preclean_klasses(&mrias_cl, _cmsGen->freelistLock()); 4773 4774 curNumCards = preclean_card_table(_cmsGen, &smoac_cl); 4775 cumNumCards += curNumCards; 4776 if (PrintGCDetails && PrintCMSStatistics != 0) { 4777 gclog_or_tty->print_cr(" (cardTable: %d cards, re-scanned %d cards, %d iterations)", 4778 curNumCards, cumNumCards, numIter); 4779 } 4780 return cumNumCards; // as a measure of useful work done 4781 } 4782 4783 // PRECLEANING NOTES: 4784 // Precleaning involves: 4785 // . reading the bits of the modUnionTable and clearing the set bits. 4786 // . For the cards corresponding to the set bits, we scan the 4787 // objects on those cards. This means we need the free_list_lock 4788 // so that we can safely iterate over the CMS space when scanning 4789 // for oops. 4790 // . When we scan the objects, we'll be both reading and setting 4791 // marks in the marking bit map, so we'll need the marking bit map. 4792 // . For protecting _collector_state transitions, we take the CGC_lock. 4793 // Note that any races in the reading of of card table entries by the 4794 // CMS thread on the one hand and the clearing of those entries by the 4795 // VM thread or the setting of those entries by the mutator threads on the 4796 // other are quite benign. However, for efficiency it makes sense to keep 4797 // the VM thread from racing with the CMS thread while the latter is 4798 // dirty card info to the modUnionTable. We therefore also use the 4799 // CGC_lock to protect the reading of the card table and the mod union 4800 // table by the CM thread. 4801 // . We run concurrently with mutator updates, so scanning 4802 // needs to be done carefully -- we should not try to scan 4803 // potentially uninitialized objects. 4804 // 4805 // Locking strategy: While holding the CGC_lock, we scan over and 4806 // reset a maximal dirty range of the mod union / card tables, then lock 4807 // the free_list_lock and bitmap lock to do a full marking, then 4808 // release these locks; and repeat the cycle. This allows for a 4809 // certain amount of fairness in the sharing of these locks between 4810 // the CMS collector on the one hand, and the VM thread and the 4811 // mutators on the other. 4812 4813 // NOTE: preclean_mod_union_table() and preclean_card_table() 4814 // further below are largely identical; if you need to modify 4815 // one of these methods, please check the other method too. 4816 4817 size_t CMSCollector::preclean_mod_union_table( 4818 ConcurrentMarkSweepGeneration* gen, 4819 ScanMarkedObjectsAgainCarefullyClosure* cl) { 4820 verify_work_stacks_empty(); 4821 verify_overflow_empty(); 4822 4823 // strategy: starting with the first card, accumulate contiguous 4824 // ranges of dirty cards; clear these cards, then scan the region 4825 // covered by these cards. 4826 4827 // Since all of the MUT is committed ahead, we can just use 4828 // that, in case the generations expand while we are precleaning. 4829 // It might also be fine to just use the committed part of the 4830 // generation, but we might potentially miss cards when the 4831 // generation is rapidly expanding while we are in the midst 4832 // of precleaning. 4833 HeapWord* startAddr = gen->reserved().start(); 4834 HeapWord* endAddr = gen->reserved().end(); 4835 4836 cl->setFreelistLock(gen->freelistLock()); // needed for yielding 4837 4838 size_t numDirtyCards, cumNumDirtyCards; 4839 HeapWord *nextAddr, *lastAddr; 4840 for (cumNumDirtyCards = numDirtyCards = 0, 4841 nextAddr = lastAddr = startAddr; 4842 nextAddr < endAddr; 4843 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 4844 4845 ResourceMark rm; 4846 HandleMark hm; 4847 4848 MemRegion dirtyRegion; 4849 { 4850 stopTimer(); 4851 // Potential yield point 4852 CMSTokenSync ts(true); 4853 startTimer(); 4854 sample_eden(); 4855 // Get dirty region starting at nextOffset (inclusive), 4856 // simultaneously clearing it. 4857 dirtyRegion = 4858 _modUnionTable.getAndClearMarkedRegion(nextAddr, endAddr); 4859 assert(dirtyRegion.start() >= nextAddr, 4860 "returned region inconsistent?"); 4861 } 4862 // Remember where the next search should begin. 4863 // The returned region (if non-empty) is a right open interval, 4864 // so lastOffset is obtained from the right end of that 4865 // interval. 4866 lastAddr = dirtyRegion.end(); 4867 // Should do something more transparent and less hacky XXX 4868 numDirtyCards = 4869 _modUnionTable.heapWordDiffToOffsetDiff(dirtyRegion.word_size()); 4870 4871 // We'll scan the cards in the dirty region (with periodic 4872 // yields for foreground GC as needed). 4873 if (!dirtyRegion.is_empty()) { 4874 assert(numDirtyCards > 0, "consistency check"); 4875 HeapWord* stop_point = NULL; 4876 stopTimer(); 4877 // Potential yield point 4878 CMSTokenSyncWithLocks ts(true, gen->freelistLock(), 4879 bitMapLock()); 4880 startTimer(); 4881 { 4882 verify_work_stacks_empty(); 4883 verify_overflow_empty(); 4884 sample_eden(); 4885 stop_point = 4886 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 4887 } 4888 if (stop_point != NULL) { 4889 // The careful iteration stopped early either because it found an 4890 // uninitialized object, or because we were in the midst of an 4891 // "abortable preclean", which should now be aborted. Redirty 4892 // the bits corresponding to the partially-scanned or unscanned 4893 // cards. We'll either restart at the next block boundary or 4894 // abort the preclean. 4895 assert((_collectorState == AbortablePreclean && should_abort_preclean()), 4896 "Should only be AbortablePreclean."); 4897 _modUnionTable.mark_range(MemRegion(stop_point, dirtyRegion.end())); 4898 if (should_abort_preclean()) { 4899 break; // out of preclean loop 4900 } else { 4901 // Compute the next address at which preclean should pick up; 4902 // might need bitMapLock in order to read P-bits. 4903 lastAddr = next_card_start_after_block(stop_point); 4904 } 4905 } 4906 } else { 4907 assert(lastAddr == endAddr, "consistency check"); 4908 assert(numDirtyCards == 0, "consistency check"); 4909 break; 4910 } 4911 } 4912 verify_work_stacks_empty(); 4913 verify_overflow_empty(); 4914 return cumNumDirtyCards; 4915 } 4916 4917 // NOTE: preclean_mod_union_table() above and preclean_card_table() 4918 // below are largely identical; if you need to modify 4919 // one of these methods, please check the other method too. 4920 4921 size_t CMSCollector::preclean_card_table(ConcurrentMarkSweepGeneration* gen, 4922 ScanMarkedObjectsAgainCarefullyClosure* cl) { 4923 // strategy: it's similar to precleamModUnionTable above, in that 4924 // we accumulate contiguous ranges of dirty cards, mark these cards 4925 // precleaned, then scan the region covered by these cards. 4926 HeapWord* endAddr = (HeapWord*)(gen->_virtual_space.high()); 4927 HeapWord* startAddr = (HeapWord*)(gen->_virtual_space.low()); 4928 4929 cl->setFreelistLock(gen->freelistLock()); // needed for yielding 4930 4931 size_t numDirtyCards, cumNumDirtyCards; 4932 HeapWord *lastAddr, *nextAddr; 4933 4934 for (cumNumDirtyCards = numDirtyCards = 0, 4935 nextAddr = lastAddr = startAddr; 4936 nextAddr < endAddr; 4937 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 4938 4939 ResourceMark rm; 4940 HandleMark hm; 4941 4942 MemRegion dirtyRegion; 4943 { 4944 // See comments in "Precleaning notes" above on why we 4945 // do this locking. XXX Could the locking overheads be 4946 // too high when dirty cards are sparse? [I don't think so.] 4947 stopTimer(); 4948 CMSTokenSync x(true); // is cms thread 4949 startTimer(); 4950 sample_eden(); 4951 // Get and clear dirty region from card table 4952 dirtyRegion = _ct->ct_bs()->dirty_card_range_after_reset( 4953 MemRegion(nextAddr, endAddr), 4954 true, 4955 CardTableModRefBS::precleaned_card_val()); 4956 4957 assert(dirtyRegion.start() >= nextAddr, 4958 "returned region inconsistent?"); 4959 } 4960 lastAddr = dirtyRegion.end(); 4961 numDirtyCards = 4962 dirtyRegion.word_size()/CardTableModRefBS::card_size_in_words; 4963 4964 if (!dirtyRegion.is_empty()) { 4965 stopTimer(); 4966 CMSTokenSyncWithLocks ts(true, gen->freelistLock(), bitMapLock()); 4967 startTimer(); 4968 sample_eden(); 4969 verify_work_stacks_empty(); 4970 verify_overflow_empty(); 4971 HeapWord* stop_point = 4972 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 4973 if (stop_point != NULL) { 4974 assert((_collectorState == AbortablePreclean && should_abort_preclean()), 4975 "Should only be AbortablePreclean."); 4976 _ct->ct_bs()->invalidate(MemRegion(stop_point, dirtyRegion.end())); 4977 if (should_abort_preclean()) { 4978 break; // out of preclean loop 4979 } else { 4980 // Compute the next address at which preclean should pick up. 4981 lastAddr = next_card_start_after_block(stop_point); 4982 } 4983 } 4984 } else { 4985 break; 4986 } 4987 } 4988 verify_work_stacks_empty(); 4989 verify_overflow_empty(); 4990 return cumNumDirtyCards; 4991 } 4992 4993 class PrecleanKlassClosure : public KlassClosure { 4994 CMKlassClosure _cm_klass_closure; 4995 public: 4996 PrecleanKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {} 4997 void do_klass(Klass* k) { 4998 if (k->has_accumulated_modified_oops()) { 4999 k->clear_accumulated_modified_oops(); 5000 5001 _cm_klass_closure.do_klass(k); 5002 } 5003 } 5004 }; 5005 5006 // The freelist lock is needed to prevent asserts, is it really needed? 5007 void CMSCollector::preclean_klasses(MarkRefsIntoAndScanClosure* cl, Mutex* freelistLock) { 5008 5009 cl->set_freelistLock(freelistLock); 5010 5011 CMSTokenSyncWithLocks ts(true, freelistLock, bitMapLock()); 5012 5013 // SSS: Add equivalent to ScanMarkedObjectsAgainCarefullyClosure::do_yield_check and should_abort_preclean? 5014 // SSS: We should probably check if precleaning should be aborted, at suitable intervals? 5015 PrecleanKlassClosure preclean_klass_closure(cl); 5016 ClassLoaderDataGraph::classes_do(&preclean_klass_closure); 5017 5018 verify_work_stacks_empty(); 5019 verify_overflow_empty(); 5020 } 5021 5022 void CMSCollector::checkpointRootsFinal(bool asynch, 5023 bool clear_all_soft_refs, bool init_mark_was_synchronous) { 5024 assert(_collectorState == FinalMarking, "incorrect state transition?"); 5025 check_correct_thread_executing(); 5026 // world is stopped at this checkpoint 5027 assert(SafepointSynchronize::is_at_safepoint(), 5028 "world should be stopped"); 5029 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 5030 5031 verify_work_stacks_empty(); 5032 verify_overflow_empty(); 5033 5034 SpecializationStats::clear(); 5035 if (PrintGCDetails) { 5036 gclog_or_tty->print("[YG occupancy: "SIZE_FORMAT" K ("SIZE_FORMAT" K)]", 5037 _young_gen->used() / K, 5038 _young_gen->capacity() / K); 5039 } 5040 if (asynch) { 5041 if (CMSScavengeBeforeRemark) { 5042 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5043 // Temporarily set flag to false, GCH->do_collection will 5044 // expect it to be false and set to true 5045 FlagSetting fl(gch->_is_gc_active, false); 5046 NOT_PRODUCT(GCTraceTime t("Scavenge-Before-Remark", 5047 PrintGCDetails && Verbose, true, _gc_timer_cm);) 5048 int level = _cmsGen->level() - 1; 5049 if (level >= 0) { 5050 gch->do_collection(true, // full (i.e. force, see below) 5051 false, // !clear_all_soft_refs 5052 0, // size 5053 false, // is_tlab 5054 level // max_level 5055 ); 5056 } 5057 } 5058 FreelistLocker x(this); 5059 MutexLockerEx y(bitMapLock(), 5060 Mutex::_no_safepoint_check_flag); 5061 assert(!init_mark_was_synchronous, "but that's impossible!"); 5062 checkpointRootsFinalWork(asynch, clear_all_soft_refs, false); 5063 } else { 5064 // already have all the locks 5065 checkpointRootsFinalWork(asynch, clear_all_soft_refs, 5066 init_mark_was_synchronous); 5067 } 5068 verify_work_stacks_empty(); 5069 verify_overflow_empty(); 5070 SpecializationStats::print(); 5071 } 5072 5073 void CMSCollector::checkpointRootsFinalWork(bool asynch, 5074 bool clear_all_soft_refs, bool init_mark_was_synchronous) { 5075 5076 NOT_PRODUCT(GCTraceTime tr("checkpointRootsFinalWork", PrintGCDetails, false, _gc_timer_cm);) 5077 5078 assert(haveFreelistLocks(), "must have free list locks"); 5079 assert_lock_strong(bitMapLock()); 5080 5081 if (UseAdaptiveSizePolicy) { 5082 size_policy()->checkpoint_roots_final_begin(); 5083 } 5084 5085 ResourceMark rm; 5086 HandleMark hm; 5087 5088 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5089 5090 if (should_unload_classes()) { 5091 CodeCache::gc_prologue(); 5092 } 5093 assert(haveFreelistLocks(), "must have free list locks"); 5094 assert_lock_strong(bitMapLock()); 5095 5096 if (!init_mark_was_synchronous) { 5097 // We might assume that we need not fill TLAB's when 5098 // CMSScavengeBeforeRemark is set, because we may have just done 5099 // a scavenge which would have filled all TLAB's -- and besides 5100 // Eden would be empty. This however may not always be the case -- 5101 // for instance although we asked for a scavenge, it may not have 5102 // happened because of a JNI critical section. We probably need 5103 // a policy for deciding whether we can in that case wait until 5104 // the critical section releases and then do the remark following 5105 // the scavenge, and skip it here. In the absence of that policy, 5106 // or of an indication of whether the scavenge did indeed occur, 5107 // we cannot rely on TLAB's having been filled and must do 5108 // so here just in case a scavenge did not happen. 5109 gch->ensure_parsability(false); // fill TLAB's, but no need to retire them 5110 // Update the saved marks which may affect the root scans. 5111 gch->save_marks(); 5112 5113 if (CMSPrintEdenSurvivorChunks) { 5114 print_eden_and_survivor_chunk_arrays(); 5115 } 5116 5117 { 5118 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;) 5119 5120 // Note on the role of the mod union table: 5121 // Since the marker in "markFromRoots" marks concurrently with 5122 // mutators, it is possible for some reachable objects not to have been 5123 // scanned. For instance, an only reference to an object A was 5124 // placed in object B after the marker scanned B. Unless B is rescanned, 5125 // A would be collected. Such updates to references in marked objects 5126 // are detected via the mod union table which is the set of all cards 5127 // dirtied since the first checkpoint in this GC cycle and prior to 5128 // the most recent young generation GC, minus those cleaned up by the 5129 // concurrent precleaning. 5130 if (CMSParallelRemarkEnabled && CollectedHeap::use_parallel_gc_threads()) { 5131 GCTraceTime t("Rescan (parallel) ", PrintGCDetails, false, _gc_timer_cm); 5132 do_remark_parallel(); 5133 } else { 5134 GCTraceTime t("Rescan (non-parallel) ", PrintGCDetails, false, 5135 _gc_timer_cm); 5136 do_remark_non_parallel(); 5137 } 5138 } 5139 } else { 5140 assert(!asynch, "Can't have init_mark_was_synchronous in asynch mode"); 5141 // The initial mark was stop-world, so there's no rescanning to 5142 // do; go straight on to the next step below. 5143 } 5144 verify_work_stacks_empty(); 5145 verify_overflow_empty(); 5146 5147 { 5148 NOT_PRODUCT(GCTraceTime ts("refProcessingWork", PrintGCDetails, false, _gc_timer_cm);) 5149 refProcessingWork(asynch, clear_all_soft_refs); 5150 } 5151 verify_work_stacks_empty(); 5152 verify_overflow_empty(); 5153 5154 if (should_unload_classes()) { 5155 CodeCache::gc_epilogue(); 5156 } 5157 JvmtiExport::gc_epilogue(); 5158 5159 // If we encountered any (marking stack / work queue) overflow 5160 // events during the current CMS cycle, take appropriate 5161 // remedial measures, where possible, so as to try and avoid 5162 // recurrence of that condition. 5163 assert(_markStack.isEmpty(), "No grey objects"); 5164 size_t ser_ovflw = _ser_pmc_remark_ovflw + _ser_pmc_preclean_ovflw + 5165 _ser_kac_ovflw + _ser_kac_preclean_ovflw; 5166 if (ser_ovflw > 0) { 5167 if (PrintCMSStatistics != 0) { 5168 gclog_or_tty->print_cr("Marking stack overflow (benign) " 5169 "(pmc_pc="SIZE_FORMAT", pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT 5170 ", kac_preclean="SIZE_FORMAT")", 5171 _ser_pmc_preclean_ovflw, _ser_pmc_remark_ovflw, 5172 _ser_kac_ovflw, _ser_kac_preclean_ovflw); 5173 } 5174 _markStack.expand(); 5175 _ser_pmc_remark_ovflw = 0; 5176 _ser_pmc_preclean_ovflw = 0; 5177 _ser_kac_preclean_ovflw = 0; 5178 _ser_kac_ovflw = 0; 5179 } 5180 if (_par_pmc_remark_ovflw > 0 || _par_kac_ovflw > 0) { 5181 if (PrintCMSStatistics != 0) { 5182 gclog_or_tty->print_cr("Work queue overflow (benign) " 5183 "(pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT")", 5184 _par_pmc_remark_ovflw, _par_kac_ovflw); 5185 } 5186 _par_pmc_remark_ovflw = 0; 5187 _par_kac_ovflw = 0; 5188 } 5189 if (PrintCMSStatistics != 0) { 5190 if (_markStack._hit_limit > 0) { 5191 gclog_or_tty->print_cr(" (benign) Hit max stack size limit ("SIZE_FORMAT")", 5192 _markStack._hit_limit); 5193 } 5194 if (_markStack._failed_double > 0) { 5195 gclog_or_tty->print_cr(" (benign) Failed stack doubling ("SIZE_FORMAT")," 5196 " current capacity "SIZE_FORMAT, 5197 _markStack._failed_double, 5198 _markStack.capacity()); 5199 } 5200 } 5201 _markStack._hit_limit = 0; 5202 _markStack._failed_double = 0; 5203 5204 if ((VerifyAfterGC || VerifyDuringGC) && 5205 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 5206 verify_after_remark(); 5207 } 5208 5209 _gc_tracer_cm->report_object_count_after_gc(&_is_alive_closure); 5210 5211 // Change under the freelistLocks. 5212 _collectorState = Sweeping; 5213 // Call isAllClear() under bitMapLock 5214 assert(_modUnionTable.isAllClear(), 5215 "Should be clear by end of the final marking"); 5216 assert(_ct->klass_rem_set()->mod_union_is_clear(), 5217 "Should be clear by end of the final marking"); 5218 if (UseAdaptiveSizePolicy) { 5219 size_policy()->checkpoint_roots_final_end(gch->gc_cause()); 5220 } 5221 } 5222 5223 void CMSParInitialMarkTask::work(uint worker_id) { 5224 elapsedTimer _timer; 5225 ResourceMark rm; 5226 HandleMark hm; 5227 5228 // ---------- scan from roots -------------- 5229 _timer.start(); 5230 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5231 Par_MarkRefsIntoClosure par_mri_cl(_collector->_span, &(_collector->_markBitMap)); 5232 CMKlassClosure klass_closure(&par_mri_cl); 5233 5234 // ---------- young gen roots -------------- 5235 { 5236 work_on_young_gen_roots(worker_id, &par_mri_cl); 5237 _timer.stop(); 5238 if (PrintCMSStatistics != 0) { 5239 gclog_or_tty->print_cr( 5240 "Finished young gen initial mark scan work in %dth thread: %3.3f sec", 5241 worker_id, _timer.seconds()); 5242 } 5243 } 5244 5245 // ---------- remaining roots -------------- 5246 _timer.reset(); 5247 _timer.start(); 5248 gch->gen_process_strong_roots(_collector->_cmsGen->level(), 5249 false, // yg was scanned above 5250 false, // this is parallel code 5251 false, // not scavenging 5252 SharedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()), 5253 &par_mri_cl, 5254 true, // walk all of code cache if (so & SO_CodeCache) 5255 NULL, 5256 &klass_closure); 5257 assert(_collector->should_unload_classes() 5258 || (_collector->CMSCollector::roots_scanning_options() & SharedHeap::SO_CodeCache), 5259 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 5260 _timer.stop(); 5261 if (PrintCMSStatistics != 0) { 5262 gclog_or_tty->print_cr( 5263 "Finished remaining root initial mark scan work in %dth thread: %3.3f sec", 5264 worker_id, _timer.seconds()); 5265 } 5266 } 5267 5268 // Parallel remark task 5269 class CMSParRemarkTask: public CMSParMarkTask { 5270 CompactibleFreeListSpace* _cms_space; 5271 5272 // The per-thread work queues, available here for stealing. 5273 OopTaskQueueSet* _task_queues; 5274 ParallelTaskTerminator _term; 5275 5276 public: 5277 // A value of 0 passed to n_workers will cause the number of 5278 // workers to be taken from the active workers in the work gang. 5279 CMSParRemarkTask(CMSCollector* collector, 5280 CompactibleFreeListSpace* cms_space, 5281 int n_workers, FlexibleWorkGang* workers, 5282 OopTaskQueueSet* task_queues): 5283 CMSParMarkTask("Rescan roots and grey objects in parallel", 5284 collector, n_workers), 5285 _cms_space(cms_space), 5286 _task_queues(task_queues), 5287 _term(n_workers, task_queues) { } 5288 5289 OopTaskQueueSet* task_queues() { return _task_queues; } 5290 5291 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 5292 5293 ParallelTaskTerminator* terminator() { return &_term; } 5294 int n_workers() { return _n_workers; } 5295 5296 void work(uint worker_id); 5297 5298 private: 5299 // ... of dirty cards in old space 5300 void do_dirty_card_rescan_tasks(CompactibleFreeListSpace* sp, int i, 5301 Par_MarkRefsIntoAndScanClosure* cl); 5302 5303 // ... work stealing for the above 5304 void do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, int* seed); 5305 }; 5306 5307 class RemarkKlassClosure : public KlassClosure { 5308 CMKlassClosure _cm_klass_closure; 5309 public: 5310 RemarkKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {} 5311 void do_klass(Klass* k) { 5312 // Check if we have modified any oops in the Klass during the concurrent marking. 5313 if (k->has_accumulated_modified_oops()) { 5314 k->clear_accumulated_modified_oops(); 5315 5316 // We could have transfered the current modified marks to the accumulated marks, 5317 // like we do with the Card Table to Mod Union Table. But it's not really necessary. 5318 } else if (k->has_modified_oops()) { 5319 // Don't clear anything, this info is needed by the next young collection. 5320 } else { 5321 // No modified oops in the Klass. 5322 return; 5323 } 5324 5325 // The klass has modified fields, need to scan the klass. 5326 _cm_klass_closure.do_klass(k); 5327 } 5328 }; 5329 5330 void CMSParMarkTask::work_on_young_gen_roots(uint worker_id, OopsInGenClosure* cl) { 5331 DefNewGeneration* dng = _collector->_young_gen->as_DefNewGeneration(); 5332 EdenSpace* eden_space = dng->eden(); 5333 ContiguousSpace* from_space = dng->from(); 5334 ContiguousSpace* to_space = dng->to(); 5335 5336 HeapWord** eca = _collector->_eden_chunk_array; 5337 size_t ect = _collector->_eden_chunk_index; 5338 HeapWord** sca = _collector->_survivor_chunk_array; 5339 size_t sct = _collector->_survivor_chunk_index; 5340 5341 assert(ect <= _collector->_eden_chunk_capacity, "out of bounds"); 5342 assert(sct <= _collector->_survivor_chunk_capacity, "out of bounds"); 5343 5344 do_young_space_rescan(worker_id, cl, to_space, NULL, 0); 5345 do_young_space_rescan(worker_id, cl, from_space, sca, sct); 5346 do_young_space_rescan(worker_id, cl, eden_space, eca, ect); 5347 } 5348 5349 // work_queue(i) is passed to the closure 5350 // Par_MarkRefsIntoAndScanClosure. The "i" parameter 5351 // also is passed to do_dirty_card_rescan_tasks() and to 5352 // do_work_steal() to select the i-th task_queue. 5353 5354 void CMSParRemarkTask::work(uint worker_id) { 5355 elapsedTimer _timer; 5356 ResourceMark rm; 5357 HandleMark hm; 5358 5359 // ---------- rescan from roots -------------- 5360 _timer.start(); 5361 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5362 Par_MarkRefsIntoAndScanClosure par_mrias_cl(_collector, 5363 _collector->_span, _collector->ref_processor(), 5364 &(_collector->_markBitMap), 5365 work_queue(worker_id)); 5366 5367 // Rescan young gen roots first since these are likely 5368 // coarsely partitioned and may, on that account, constitute 5369 // the critical path; thus, it's best to start off that 5370 // work first. 5371 // ---------- young gen roots -------------- 5372 { 5373 work_on_young_gen_roots(worker_id, &par_mrias_cl); 5374 _timer.stop(); 5375 if (PrintCMSStatistics != 0) { 5376 gclog_or_tty->print_cr( 5377 "Finished young gen rescan work in %dth thread: %3.3f sec", 5378 worker_id, _timer.seconds()); 5379 } 5380 } 5381 5382 // ---------- remaining roots -------------- 5383 _timer.reset(); 5384 _timer.start(); 5385 gch->gen_process_strong_roots(_collector->_cmsGen->level(), 5386 false, // yg was scanned above 5387 false, // this is parallel code 5388 false, // not scavenging 5389 SharedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()), 5390 &par_mrias_cl, 5391 true, // walk all of code cache if (so & SO_CodeCache) 5392 NULL, 5393 NULL); // The dirty klasses will be handled below 5394 assert(_collector->should_unload_classes() 5395 || (_collector->CMSCollector::roots_scanning_options() & SharedHeap::SO_CodeCache), 5396 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 5397 _timer.stop(); 5398 if (PrintCMSStatistics != 0) { 5399 gclog_or_tty->print_cr( 5400 "Finished remaining root rescan work in %dth thread: %3.3f sec", 5401 worker_id, _timer.seconds()); 5402 } 5403 5404 // ---------- unhandled CLD scanning ---------- 5405 if (worker_id == 0) { // Single threaded at the moment. 5406 _timer.reset(); 5407 _timer.start(); 5408 5409 // Scan all new class loader data objects and new dependencies that were 5410 // introduced during concurrent marking. 5411 ResourceMark rm; 5412 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds(); 5413 for (int i = 0; i < array->length(); i++) { 5414 par_mrias_cl.do_class_loader_data(array->at(i)); 5415 } 5416 5417 // We don't need to keep track of new CLDs anymore. 5418 ClassLoaderDataGraph::remember_new_clds(false); 5419 5420 _timer.stop(); 5421 if (PrintCMSStatistics != 0) { 5422 gclog_or_tty->print_cr( 5423 "Finished unhandled CLD scanning work in %dth thread: %3.3f sec", 5424 worker_id, _timer.seconds()); 5425 } 5426 } 5427 5428 // ---------- dirty klass scanning ---------- 5429 if (worker_id == 0) { // Single threaded at the moment. 5430 _timer.reset(); 5431 _timer.start(); 5432 5433 // Scan all classes that was dirtied during the concurrent marking phase. 5434 RemarkKlassClosure remark_klass_closure(&par_mrias_cl); 5435 ClassLoaderDataGraph::classes_do(&remark_klass_closure); 5436 5437 _timer.stop(); 5438 if (PrintCMSStatistics != 0) { 5439 gclog_or_tty->print_cr( 5440 "Finished dirty klass scanning work in %dth thread: %3.3f sec", 5441 worker_id, _timer.seconds()); 5442 } 5443 } 5444 5445 // We might have added oops to ClassLoaderData::_handles during the 5446 // concurrent marking phase. These oops point to newly allocated objects 5447 // that are guaranteed to be kept alive. Either by the direct allocation 5448 // code, or when the young collector processes the strong roots. Hence, 5449 // we don't have to revisit the _handles block during the remark phase. 5450 5451 // ---------- rescan dirty cards ------------ 5452 _timer.reset(); 5453 _timer.start(); 5454 5455 // Do the rescan tasks for each of the two spaces 5456 // (cms_space) in turn. 5457 // "worker_id" is passed to select the task_queue for "worker_id" 5458 do_dirty_card_rescan_tasks(_cms_space, worker_id, &par_mrias_cl); 5459 _timer.stop(); 5460 if (PrintCMSStatistics != 0) { 5461 gclog_or_tty->print_cr( 5462 "Finished dirty card rescan work in %dth thread: %3.3f sec", 5463 worker_id, _timer.seconds()); 5464 } 5465 5466 // ---------- steal work from other threads ... 5467 // ---------- ... and drain overflow list. 5468 _timer.reset(); 5469 _timer.start(); 5470 do_work_steal(worker_id, &par_mrias_cl, _collector->hash_seed(worker_id)); 5471 _timer.stop(); 5472 if (PrintCMSStatistics != 0) { 5473 gclog_or_tty->print_cr( 5474 "Finished work stealing in %dth thread: %3.3f sec", 5475 worker_id, _timer.seconds()); 5476 } 5477 } 5478 5479 // Note that parameter "i" is not used. 5480 void 5481 CMSParMarkTask::do_young_space_rescan(uint worker_id, 5482 OopsInGenClosure* cl, ContiguousSpace* space, 5483 HeapWord** chunk_array, size_t chunk_top) { 5484 // Until all tasks completed: 5485 // . claim an unclaimed task 5486 // . compute region boundaries corresponding to task claimed 5487 // using chunk_array 5488 // . par_oop_iterate(cl) over that region 5489 5490 ResourceMark rm; 5491 HandleMark hm; 5492 5493 SequentialSubTasksDone* pst = space->par_seq_tasks(); 5494 5495 uint nth_task = 0; 5496 uint n_tasks = pst->n_tasks(); 5497 5498 if (n_tasks > 0) { 5499 assert(pst->valid(), "Uninitialized use?"); 5500 HeapWord *start, *end; 5501 while (!pst->is_task_claimed(/* reference */ nth_task)) { 5502 // We claimed task # nth_task; compute its boundaries. 5503 if (chunk_top == 0) { // no samples were taken 5504 assert(nth_task == 0 && n_tasks == 1, "Can have only 1 EdenSpace task"); 5505 start = space->bottom(); 5506 end = space->top(); 5507 } else if (nth_task == 0) { 5508 start = space->bottom(); 5509 end = chunk_array[nth_task]; 5510 } else if (nth_task < (uint)chunk_top) { 5511 assert(nth_task >= 1, "Control point invariant"); 5512 start = chunk_array[nth_task - 1]; 5513 end = chunk_array[nth_task]; 5514 } else { 5515 assert(nth_task == (uint)chunk_top, "Control point invariant"); 5516 start = chunk_array[chunk_top - 1]; 5517 end = space->top(); 5518 } 5519 MemRegion mr(start, end); 5520 // Verify that mr is in space 5521 assert(mr.is_empty() || space->used_region().contains(mr), 5522 "Should be in space"); 5523 // Verify that "start" is an object boundary 5524 assert(mr.is_empty() || oop(mr.start())->is_oop(), 5525 "Should be an oop"); 5526 space->par_oop_iterate(mr, cl); 5527 } 5528 pst->all_tasks_completed(); 5529 } 5530 } 5531 5532 void 5533 CMSParRemarkTask::do_dirty_card_rescan_tasks( 5534 CompactibleFreeListSpace* sp, int i, 5535 Par_MarkRefsIntoAndScanClosure* cl) { 5536 // Until all tasks completed: 5537 // . claim an unclaimed task 5538 // . compute region boundaries corresponding to task claimed 5539 // . transfer dirty bits ct->mut for that region 5540 // . apply rescanclosure to dirty mut bits for that region 5541 5542 ResourceMark rm; 5543 HandleMark hm; 5544 5545 OopTaskQueue* work_q = work_queue(i); 5546 ModUnionClosure modUnionClosure(&(_collector->_modUnionTable)); 5547 // CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! 5548 // CAUTION: This closure has state that persists across calls to 5549 // the work method dirty_range_iterate_clear() in that it has 5550 // imbedded in it a (subtype of) UpwardsObjectClosure. The 5551 // use of that state in the imbedded UpwardsObjectClosure instance 5552 // assumes that the cards are always iterated (even if in parallel 5553 // by several threads) in monotonically increasing order per each 5554 // thread. This is true of the implementation below which picks 5555 // card ranges (chunks) in monotonically increasing order globally 5556 // and, a-fortiori, in monotonically increasing order per thread 5557 // (the latter order being a subsequence of the former). 5558 // If the work code below is ever reorganized into a more chaotic 5559 // work-partitioning form than the current "sequential tasks" 5560 // paradigm, the use of that persistent state will have to be 5561 // revisited and modified appropriately. See also related 5562 // bug 4756801 work on which should examine this code to make 5563 // sure that the changes there do not run counter to the 5564 // assumptions made here and necessary for correctness and 5565 // efficiency. Note also that this code might yield inefficient 5566 // behaviour in the case of very large objects that span one or 5567 // more work chunks. Such objects would potentially be scanned 5568 // several times redundantly. Work on 4756801 should try and 5569 // address that performance anomaly if at all possible. XXX 5570 MemRegion full_span = _collector->_span; 5571 CMSBitMap* bm = &(_collector->_markBitMap); // shared 5572 MarkFromDirtyCardsClosure 5573 greyRescanClosure(_collector, full_span, // entire span of interest 5574 sp, bm, work_q, cl); 5575 5576 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 5577 assert(pst->valid(), "Uninitialized use?"); 5578 uint nth_task = 0; 5579 const int alignment = CardTableModRefBS::card_size * BitsPerWord; 5580 MemRegion span = sp->used_region(); 5581 HeapWord* start_addr = span.start(); 5582 HeapWord* end_addr = (HeapWord*)round_to((intptr_t)span.end(), 5583 alignment); 5584 const size_t chunk_size = sp->rescan_task_size(); // in HeapWord units 5585 assert((HeapWord*)round_to((intptr_t)start_addr, alignment) == 5586 start_addr, "Check alignment"); 5587 assert((size_t)round_to((intptr_t)chunk_size, alignment) == 5588 chunk_size, "Check alignment"); 5589 5590 while (!pst->is_task_claimed(/* reference */ nth_task)) { 5591 // Having claimed the nth_task, compute corresponding mem-region, 5592 // which is a-fortiori aligned correctly (i.e. at a MUT bopundary). 5593 // The alignment restriction ensures that we do not need any 5594 // synchronization with other gang-workers while setting or 5595 // clearing bits in thus chunk of the MUT. 5596 MemRegion this_span = MemRegion(start_addr + nth_task*chunk_size, 5597 start_addr + (nth_task+1)*chunk_size); 5598 // The last chunk's end might be way beyond end of the 5599 // used region. In that case pull back appropriately. 5600 if (this_span.end() > end_addr) { 5601 this_span.set_end(end_addr); 5602 assert(!this_span.is_empty(), "Program logic (calculation of n_tasks)"); 5603 } 5604 // Iterate over the dirty cards covering this chunk, marking them 5605 // precleaned, and setting the corresponding bits in the mod union 5606 // table. Since we have been careful to partition at Card and MUT-word 5607 // boundaries no synchronization is needed between parallel threads. 5608 _collector->_ct->ct_bs()->dirty_card_iterate(this_span, 5609 &modUnionClosure); 5610 5611 // Having transferred these marks into the modUnionTable, 5612 // rescan the marked objects on the dirty cards in the modUnionTable. 5613 // Even if this is at a synchronous collection, the initial marking 5614 // may have been done during an asynchronous collection so there 5615 // may be dirty bits in the mod-union table. 5616 _collector->_modUnionTable.dirty_range_iterate_clear( 5617 this_span, &greyRescanClosure); 5618 _collector->_modUnionTable.verifyNoOneBitsInRange( 5619 this_span.start(), 5620 this_span.end()); 5621 } 5622 pst->all_tasks_completed(); // declare that i am done 5623 } 5624 5625 // . see if we can share work_queues with ParNew? XXX 5626 void 5627 CMSParRemarkTask::do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, 5628 int* seed) { 5629 OopTaskQueue* work_q = work_queue(i); 5630 NOT_PRODUCT(int num_steals = 0;) 5631 oop obj_to_scan; 5632 CMSBitMap* bm = &(_collector->_markBitMap); 5633 5634 while (true) { 5635 // Completely finish any left over work from (an) earlier round(s) 5636 cl->trim_queue(0); 5637 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 5638 (size_t)ParGCDesiredObjsFromOverflowList); 5639 // Now check if there's any work in the overflow list 5640 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads, 5641 // only affects the number of attempts made to get work from the 5642 // overflow list and does not affect the number of workers. Just 5643 // pass ParallelGCThreads so this behavior is unchanged. 5644 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 5645 work_q, 5646 ParallelGCThreads)) { 5647 // found something in global overflow list; 5648 // not yet ready to go stealing work from others. 5649 // We'd like to assert(work_q->size() != 0, ...) 5650 // because we just took work from the overflow list, 5651 // but of course we can't since all of that could have 5652 // been already stolen from us. 5653 // "He giveth and He taketh away." 5654 continue; 5655 } 5656 // Verify that we have no work before we resort to stealing 5657 assert(work_q->size() == 0, "Have work, shouldn't steal"); 5658 // Try to steal from other queues that have work 5659 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 5660 NOT_PRODUCT(num_steals++;) 5661 assert(obj_to_scan->is_oop(), "Oops, not an oop!"); 5662 assert(bm->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 5663 // Do scanning work 5664 obj_to_scan->oop_iterate(cl); 5665 // Loop around, finish this work, and try to steal some more 5666 } else if (terminator()->offer_termination()) { 5667 break; // nirvana from the infinite cycle 5668 } 5669 } 5670 NOT_PRODUCT( 5671 if (PrintCMSStatistics != 0) { 5672 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals); 5673 } 5674 ) 5675 assert(work_q->size() == 0 && _collector->overflow_list_is_empty(), 5676 "Else our work is not yet done"); 5677 } 5678 5679 // Record object boundaries in _eden_chunk_array by sampling the eden 5680 // top in the slow-path eden object allocation code path and record 5681 // the boundaries, if CMSEdenChunksRecordAlways is true. If 5682 // CMSEdenChunksRecordAlways is false, we use the other asynchronous 5683 // sampling in sample_eden() that activates during the part of the 5684 // preclean phase. 5685 void CMSCollector::sample_eden_chunk() { 5686 if (CMSEdenChunksRecordAlways && _eden_chunk_array != NULL) { 5687 if (_eden_chunk_lock->try_lock()) { 5688 // Record a sample. This is the critical section. The contents 5689 // of the _eden_chunk_array have to be non-decreasing in the 5690 // address order. 5691 _eden_chunk_array[_eden_chunk_index] = *_top_addr; 5692 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr, 5693 "Unexpected state of Eden"); 5694 if (_eden_chunk_index == 0 || 5695 ((_eden_chunk_array[_eden_chunk_index] > _eden_chunk_array[_eden_chunk_index-1]) && 5696 (pointer_delta(_eden_chunk_array[_eden_chunk_index], 5697 _eden_chunk_array[_eden_chunk_index-1]) >= CMSSamplingGrain))) { 5698 _eden_chunk_index++; // commit sample 5699 } 5700 _eden_chunk_lock->unlock(); 5701 } 5702 } 5703 } 5704 5705 // Return a thread-local PLAB recording array, as appropriate. 5706 void* CMSCollector::get_data_recorder(int thr_num) { 5707 if (_survivor_plab_array != NULL && 5708 (CMSPLABRecordAlways || 5709 (_collectorState > Marking && _collectorState < FinalMarking))) { 5710 assert(thr_num < (int)ParallelGCThreads, "thr_num is out of bounds"); 5711 ChunkArray* ca = &_survivor_plab_array[thr_num]; 5712 ca->reset(); // clear it so that fresh data is recorded 5713 return (void*) ca; 5714 } else { 5715 return NULL; 5716 } 5717 } 5718 5719 // Reset all the thread-local PLAB recording arrays 5720 void CMSCollector::reset_survivor_plab_arrays() { 5721 for (uint i = 0; i < ParallelGCThreads; i++) { 5722 _survivor_plab_array[i].reset(); 5723 } 5724 } 5725 5726 // Merge the per-thread plab arrays into the global survivor chunk 5727 // array which will provide the partitioning of the survivor space 5728 // for CMS initial scan and rescan. 5729 void CMSCollector::merge_survivor_plab_arrays(ContiguousSpace* surv, 5730 int no_of_gc_threads) { 5731 assert(_survivor_plab_array != NULL, "Error"); 5732 assert(_survivor_chunk_array != NULL, "Error"); 5733 assert(_collectorState == FinalMarking || 5734 (CMSParallelInitialMarkEnabled && _collectorState == InitialMarking), "Error"); 5735 for (int j = 0; j < no_of_gc_threads; j++) { 5736 _cursor[j] = 0; 5737 } 5738 HeapWord* top = surv->top(); 5739 size_t i; 5740 for (i = 0; i < _survivor_chunk_capacity; i++) { // all sca entries 5741 HeapWord* min_val = top; // Higher than any PLAB address 5742 uint min_tid = 0; // position of min_val this round 5743 for (int j = 0; j < no_of_gc_threads; j++) { 5744 ChunkArray* cur_sca = &_survivor_plab_array[j]; 5745 if (_cursor[j] == cur_sca->end()) { 5746 continue; 5747 } 5748 assert(_cursor[j] < cur_sca->end(), "ctl pt invariant"); 5749 HeapWord* cur_val = cur_sca->nth(_cursor[j]); 5750 assert(surv->used_region().contains(cur_val), "Out of bounds value"); 5751 if (cur_val < min_val) { 5752 min_tid = j; 5753 min_val = cur_val; 5754 } else { 5755 assert(cur_val < top, "All recorded addresses should be less"); 5756 } 5757 } 5758 // At this point min_val and min_tid are respectively 5759 // the least address in _survivor_plab_array[j]->nth(_cursor[j]) 5760 // and the thread (j) that witnesses that address. 5761 // We record this address in the _survivor_chunk_array[i] 5762 // and increment _cursor[min_tid] prior to the next round i. 5763 if (min_val == top) { 5764 break; 5765 } 5766 _survivor_chunk_array[i] = min_val; 5767 _cursor[min_tid]++; 5768 } 5769 // We are all done; record the size of the _survivor_chunk_array 5770 _survivor_chunk_index = i; // exclusive: [0, i) 5771 if (PrintCMSStatistics > 0) { 5772 gclog_or_tty->print(" (Survivor:" SIZE_FORMAT "chunks) ", i); 5773 } 5774 // Verify that we used up all the recorded entries 5775 #ifdef ASSERT 5776 size_t total = 0; 5777 for (int j = 0; j < no_of_gc_threads; j++) { 5778 assert(_cursor[j] == _survivor_plab_array[j].end(), "Ctl pt invariant"); 5779 total += _cursor[j]; 5780 } 5781 assert(total == _survivor_chunk_index, "Ctl Pt Invariant"); 5782 // Check that the merged array is in sorted order 5783 if (total > 0) { 5784 for (size_t i = 0; i < total - 1; i++) { 5785 if (PrintCMSStatistics > 0) { 5786 gclog_or_tty->print(" (chunk" SIZE_FORMAT ":" INTPTR_FORMAT ") ", 5787 i, _survivor_chunk_array[i]); 5788 } 5789 assert(_survivor_chunk_array[i] < _survivor_chunk_array[i+1], 5790 "Not sorted"); 5791 } 5792 } 5793 #endif // ASSERT 5794 } 5795 5796 // Set up the space's par_seq_tasks structure for work claiming 5797 // for parallel initial scan and rescan of young gen. 5798 // See ParRescanTask where this is currently used. 5799 void 5800 CMSCollector:: 5801 initialize_sequential_subtasks_for_young_gen_rescan(int n_threads) { 5802 assert(n_threads > 0, "Unexpected n_threads argument"); 5803 DefNewGeneration* dng = (DefNewGeneration*)_young_gen; 5804 5805 // Eden space 5806 if (!dng->eden()->is_empty()) { 5807 SequentialSubTasksDone* pst = dng->eden()->par_seq_tasks(); 5808 assert(!pst->valid(), "Clobbering existing data?"); 5809 // Each valid entry in [0, _eden_chunk_index) represents a task. 5810 size_t n_tasks = _eden_chunk_index + 1; 5811 assert(n_tasks == 1 || _eden_chunk_array != NULL, "Error"); 5812 // Sets the condition for completion of the subtask (how many threads 5813 // need to finish in order to be done). 5814 pst->set_n_threads(n_threads); 5815 pst->set_n_tasks((int)n_tasks); 5816 } 5817 5818 // Merge the survivor plab arrays into _survivor_chunk_array 5819 if (_survivor_plab_array != NULL) { 5820 merge_survivor_plab_arrays(dng->from(), n_threads); 5821 } else { 5822 assert(_survivor_chunk_index == 0, "Error"); 5823 } 5824 5825 // To space 5826 { 5827 SequentialSubTasksDone* pst = dng->to()->par_seq_tasks(); 5828 assert(!pst->valid(), "Clobbering existing data?"); 5829 // Sets the condition for completion of the subtask (how many threads 5830 // need to finish in order to be done). 5831 pst->set_n_threads(n_threads); 5832 pst->set_n_tasks(1); 5833 assert(pst->valid(), "Error"); 5834 } 5835 5836 // From space 5837 { 5838 SequentialSubTasksDone* pst = dng->from()->par_seq_tasks(); 5839 assert(!pst->valid(), "Clobbering existing data?"); 5840 size_t n_tasks = _survivor_chunk_index + 1; 5841 assert(n_tasks == 1 || _survivor_chunk_array != NULL, "Error"); 5842 // Sets the condition for completion of the subtask (how many threads 5843 // need to finish in order to be done). 5844 pst->set_n_threads(n_threads); 5845 pst->set_n_tasks((int)n_tasks); 5846 assert(pst->valid(), "Error"); 5847 } 5848 } 5849 5850 // Parallel version of remark 5851 void CMSCollector::do_remark_parallel() { 5852 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5853 FlexibleWorkGang* workers = gch->workers(); 5854 assert(workers != NULL, "Need parallel worker threads."); 5855 // Choose to use the number of GC workers most recently set 5856 // into "active_workers". If active_workers is not set, set it 5857 // to ParallelGCThreads. 5858 int n_workers = workers->active_workers(); 5859 if (n_workers == 0) { 5860 assert(n_workers > 0, "Should have been set during scavenge"); 5861 n_workers = ParallelGCThreads; 5862 workers->set_active_workers(n_workers); 5863 } 5864 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 5865 5866 CMSParRemarkTask tsk(this, 5867 cms_space, 5868 n_workers, workers, task_queues()); 5869 5870 // Set up for parallel process_strong_roots work. 5871 gch->set_par_threads(n_workers); 5872 // We won't be iterating over the cards in the card table updating 5873 // the younger_gen cards, so we shouldn't call the following else 5874 // the verification code as well as subsequent younger_refs_iterate 5875 // code would get confused. XXX 5876 // gch->rem_set()->prepare_for_younger_refs_iterate(true); // parallel 5877 5878 // The young gen rescan work will not be done as part of 5879 // process_strong_roots (which currently doesn't knw how to 5880 // parallelize such a scan), but rather will be broken up into 5881 // a set of parallel tasks (via the sampling that the [abortable] 5882 // preclean phase did of EdenSpace, plus the [two] tasks of 5883 // scanning the [two] survivor spaces. Further fine-grain 5884 // parallelization of the scanning of the survivor spaces 5885 // themselves, and of precleaning of the younger gen itself 5886 // is deferred to the future. 5887 initialize_sequential_subtasks_for_young_gen_rescan(n_workers); 5888 5889 // The dirty card rescan work is broken up into a "sequence" 5890 // of parallel tasks (per constituent space) that are dynamically 5891 // claimed by the parallel threads. 5892 cms_space->initialize_sequential_subtasks_for_rescan(n_workers); 5893 5894 // It turns out that even when we're using 1 thread, doing the work in a 5895 // separate thread causes wide variance in run times. We can't help this 5896 // in the multi-threaded case, but we special-case n=1 here to get 5897 // repeatable measurements of the 1-thread overhead of the parallel code. 5898 if (n_workers > 1) { 5899 // Make refs discovery MT-safe, if it isn't already: it may not 5900 // necessarily be so, since it's possible that we are doing 5901 // ST marking. 5902 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), true); 5903 GenCollectedHeap::StrongRootsScope srs(gch); 5904 workers->run_task(&tsk); 5905 } else { 5906 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false); 5907 GenCollectedHeap::StrongRootsScope srs(gch); 5908 tsk.work(0); 5909 } 5910 5911 gch->set_par_threads(0); // 0 ==> non-parallel. 5912 // restore, single-threaded for now, any preserved marks 5913 // as a result of work_q overflow 5914 restore_preserved_marks_if_any(); 5915 } 5916 5917 // Non-parallel version of remark 5918 void CMSCollector::do_remark_non_parallel() { 5919 ResourceMark rm; 5920 HandleMark hm; 5921 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5922 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false); 5923 5924 MarkRefsIntoAndScanClosure 5925 mrias_cl(_span, ref_processor(), &_markBitMap, NULL /* not precleaning */, 5926 &_markStack, this, 5927 false /* should_yield */, false /* not precleaning */); 5928 MarkFromDirtyCardsClosure 5929 markFromDirtyCardsClosure(this, _span, 5930 NULL, // space is set further below 5931 &_markBitMap, &_markStack, &mrias_cl); 5932 { 5933 GCTraceTime t("grey object rescan", PrintGCDetails, false, _gc_timer_cm); 5934 // Iterate over the dirty cards, setting the corresponding bits in the 5935 // mod union table. 5936 { 5937 ModUnionClosure modUnionClosure(&_modUnionTable); 5938 _ct->ct_bs()->dirty_card_iterate( 5939 _cmsGen->used_region(), 5940 &modUnionClosure); 5941 } 5942 // Having transferred these marks into the modUnionTable, we just need 5943 // to rescan the marked objects on the dirty cards in the modUnionTable. 5944 // The initial marking may have been done during an asynchronous 5945 // collection so there may be dirty bits in the mod-union table. 5946 const int alignment = 5947 CardTableModRefBS::card_size * BitsPerWord; 5948 { 5949 // ... First handle dirty cards in CMS gen 5950 markFromDirtyCardsClosure.set_space(_cmsGen->cmsSpace()); 5951 MemRegion ur = _cmsGen->used_region(); 5952 HeapWord* lb = ur.start(); 5953 HeapWord* ub = (HeapWord*)round_to((intptr_t)ur.end(), alignment); 5954 MemRegion cms_span(lb, ub); 5955 _modUnionTable.dirty_range_iterate_clear(cms_span, 5956 &markFromDirtyCardsClosure); 5957 verify_work_stacks_empty(); 5958 if (PrintCMSStatistics != 0) { 5959 gclog_or_tty->print(" (re-scanned "SIZE_FORMAT" dirty cards in cms gen) ", 5960 markFromDirtyCardsClosure.num_dirty_cards()); 5961 } 5962 } 5963 } 5964 if (VerifyDuringGC && 5965 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 5966 HandleMark hm; // Discard invalid handles created during verification 5967 Universe::verify(); 5968 } 5969 { 5970 GCTraceTime t("root rescan", PrintGCDetails, false, _gc_timer_cm); 5971 5972 verify_work_stacks_empty(); 5973 5974 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 5975 GenCollectedHeap::StrongRootsScope srs(gch); 5976 gch->gen_process_strong_roots(_cmsGen->level(), 5977 true, // younger gens as roots 5978 false, // use the local StrongRootsScope 5979 false, // not scavenging 5980 SharedHeap::ScanningOption(roots_scanning_options()), 5981 &mrias_cl, 5982 true, // walk code active on stacks 5983 NULL, 5984 NULL); // The dirty klasses will be handled below 5985 5986 assert(should_unload_classes() 5987 || (roots_scanning_options() & SharedHeap::SO_CodeCache), 5988 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 5989 } 5990 5991 { 5992 GCTraceTime t("visit unhandled CLDs", PrintGCDetails, false, _gc_timer_cm); 5993 5994 verify_work_stacks_empty(); 5995 5996 // Scan all class loader data objects that might have been introduced 5997 // during concurrent marking. 5998 ResourceMark rm; 5999 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds(); 6000 for (int i = 0; i < array->length(); i++) { 6001 mrias_cl.do_class_loader_data(array->at(i)); 6002 } 6003 6004 // We don't need to keep track of new CLDs anymore. 6005 ClassLoaderDataGraph::remember_new_clds(false); 6006 6007 verify_work_stacks_empty(); 6008 } 6009 6010 { 6011 GCTraceTime t("dirty klass scan", PrintGCDetails, false, _gc_timer_cm); 6012 6013 verify_work_stacks_empty(); 6014 6015 RemarkKlassClosure remark_klass_closure(&mrias_cl); 6016 ClassLoaderDataGraph::classes_do(&remark_klass_closure); 6017 6018 verify_work_stacks_empty(); 6019 } 6020 6021 // We might have added oops to ClassLoaderData::_handles during the 6022 // concurrent marking phase. These oops point to newly allocated objects 6023 // that are guaranteed to be kept alive. Either by the direct allocation 6024 // code, or when the young collector processes the strong roots. Hence, 6025 // we don't have to revisit the _handles block during the remark phase. 6026 6027 verify_work_stacks_empty(); 6028 // Restore evacuated mark words, if any, used for overflow list links 6029 if (!CMSOverflowEarlyRestoration) { 6030 restore_preserved_marks_if_any(); 6031 } 6032 verify_overflow_empty(); 6033 } 6034 6035 //////////////////////////////////////////////////////// 6036 // Parallel Reference Processing Task Proxy Class 6037 //////////////////////////////////////////////////////// 6038 class CMSRefProcTaskProxy: public AbstractGangTaskWOopQueues { 6039 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask; 6040 CMSCollector* _collector; 6041 CMSBitMap* _mark_bit_map; 6042 const MemRegion _span; 6043 ProcessTask& _task; 6044 6045 public: 6046 CMSRefProcTaskProxy(ProcessTask& task, 6047 CMSCollector* collector, 6048 const MemRegion& span, 6049 CMSBitMap* mark_bit_map, 6050 AbstractWorkGang* workers, 6051 OopTaskQueueSet* task_queues): 6052 // XXX Should superclass AGTWOQ also know about AWG since it knows 6053 // about the task_queues used by the AWG? Then it could initialize 6054 // the terminator() object. See 6984287. The set_for_termination() 6055 // below is a temporary band-aid for the regression in 6984287. 6056 AbstractGangTaskWOopQueues("Process referents by policy in parallel", 6057 task_queues), 6058 _task(task), 6059 _collector(collector), _span(span), _mark_bit_map(mark_bit_map) 6060 { 6061 assert(_collector->_span.equals(_span) && !_span.is_empty(), 6062 "Inconsistency in _span"); 6063 set_for_termination(workers->active_workers()); 6064 } 6065 6066 OopTaskQueueSet* task_queues() { return queues(); } 6067 6068 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 6069 6070 void do_work_steal(int i, 6071 CMSParDrainMarkingStackClosure* drain, 6072 CMSParKeepAliveClosure* keep_alive, 6073 int* seed); 6074 6075 virtual void work(uint worker_id); 6076 }; 6077 6078 void CMSRefProcTaskProxy::work(uint worker_id) { 6079 assert(_collector->_span.equals(_span), "Inconsistency in _span"); 6080 CMSParKeepAliveClosure par_keep_alive(_collector, _span, 6081 _mark_bit_map, 6082 work_queue(worker_id)); 6083 CMSParDrainMarkingStackClosure par_drain_stack(_collector, _span, 6084 _mark_bit_map, 6085 work_queue(worker_id)); 6086 CMSIsAliveClosure is_alive_closure(_span, _mark_bit_map); 6087 _task.work(worker_id, is_alive_closure, par_keep_alive, par_drain_stack); 6088 if (_task.marks_oops_alive()) { 6089 do_work_steal(worker_id, &par_drain_stack, &par_keep_alive, 6090 _collector->hash_seed(worker_id)); 6091 } 6092 assert(work_queue(worker_id)->size() == 0, "work_queue should be empty"); 6093 assert(_collector->_overflow_list == NULL, "non-empty _overflow_list"); 6094 } 6095 6096 class CMSRefEnqueueTaskProxy: public AbstractGangTask { 6097 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask; 6098 EnqueueTask& _task; 6099 6100 public: 6101 CMSRefEnqueueTaskProxy(EnqueueTask& task) 6102 : AbstractGangTask("Enqueue reference objects in parallel"), 6103 _task(task) 6104 { } 6105 6106 virtual void work(uint worker_id) 6107 { 6108 _task.work(worker_id); 6109 } 6110 }; 6111 6112 CMSParKeepAliveClosure::CMSParKeepAliveClosure(CMSCollector* collector, 6113 MemRegion span, CMSBitMap* bit_map, OopTaskQueue* work_queue): 6114 _span(span), 6115 _bit_map(bit_map), 6116 _work_queue(work_queue), 6117 _mark_and_push(collector, span, bit_map, work_queue), 6118 _low_water_mark(MIN2((uint)(work_queue->max_elems()/4), 6119 (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads))) 6120 { } 6121 6122 // . see if we can share work_queues with ParNew? XXX 6123 void CMSRefProcTaskProxy::do_work_steal(int i, 6124 CMSParDrainMarkingStackClosure* drain, 6125 CMSParKeepAliveClosure* keep_alive, 6126 int* seed) { 6127 OopTaskQueue* work_q = work_queue(i); 6128 NOT_PRODUCT(int num_steals = 0;) 6129 oop obj_to_scan; 6130 6131 while (true) { 6132 // Completely finish any left over work from (an) earlier round(s) 6133 drain->trim_queue(0); 6134 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 6135 (size_t)ParGCDesiredObjsFromOverflowList); 6136 // Now check if there's any work in the overflow list 6137 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads, 6138 // only affects the number of attempts made to get work from the 6139 // overflow list and does not affect the number of workers. Just 6140 // pass ParallelGCThreads so this behavior is unchanged. 6141 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 6142 work_q, 6143 ParallelGCThreads)) { 6144 // Found something in global overflow list; 6145 // not yet ready to go stealing work from others. 6146 // We'd like to assert(work_q->size() != 0, ...) 6147 // because we just took work from the overflow list, 6148 // but of course we can't, since all of that might have 6149 // been already stolen from us. 6150 continue; 6151 } 6152 // Verify that we have no work before we resort to stealing 6153 assert(work_q->size() == 0, "Have work, shouldn't steal"); 6154 // Try to steal from other queues that have work 6155 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 6156 NOT_PRODUCT(num_steals++;) 6157 assert(obj_to_scan->is_oop(), "Oops, not an oop!"); 6158 assert(_mark_bit_map->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 6159 // Do scanning work 6160 obj_to_scan->oop_iterate(keep_alive); 6161 // Loop around, finish this work, and try to steal some more 6162 } else if (terminator()->offer_termination()) { 6163 break; // nirvana from the infinite cycle 6164 } 6165 } 6166 NOT_PRODUCT( 6167 if (PrintCMSStatistics != 0) { 6168 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals); 6169 } 6170 ) 6171 } 6172 6173 void CMSRefProcTaskExecutor::execute(ProcessTask& task) 6174 { 6175 GenCollectedHeap* gch = GenCollectedHeap::heap(); 6176 FlexibleWorkGang* workers = gch->workers(); 6177 assert(workers != NULL, "Need parallel worker threads."); 6178 CMSRefProcTaskProxy rp_task(task, &_collector, 6179 _collector.ref_processor()->span(), 6180 _collector.markBitMap(), 6181 workers, _collector.task_queues()); 6182 workers->run_task(&rp_task); 6183 } 6184 6185 void CMSRefProcTaskExecutor::execute(EnqueueTask& task) 6186 { 6187 6188 GenCollectedHeap* gch = GenCollectedHeap::heap(); 6189 FlexibleWorkGang* workers = gch->workers(); 6190 assert(workers != NULL, "Need parallel worker threads."); 6191 CMSRefEnqueueTaskProxy enq_task(task); 6192 workers->run_task(&enq_task); 6193 } 6194 6195 void CMSCollector::refProcessingWork(bool asynch, bool clear_all_soft_refs) { 6196 6197 ResourceMark rm; 6198 HandleMark hm; 6199 6200 ReferenceProcessor* rp = ref_processor(); 6201 assert(rp->span().equals(_span), "Spans should be equal"); 6202 assert(!rp->enqueuing_is_done(), "Enqueuing should not be complete"); 6203 // Process weak references. 6204 rp->setup_policy(clear_all_soft_refs); 6205 verify_work_stacks_empty(); 6206 6207 CMSKeepAliveClosure cmsKeepAliveClosure(this, _span, &_markBitMap, 6208 &_markStack, false /* !preclean */); 6209 CMSDrainMarkingStackClosure cmsDrainMarkingStackClosure(this, 6210 _span, &_markBitMap, &_markStack, 6211 &cmsKeepAliveClosure, false /* !preclean */); 6212 { 6213 GCTraceTime t("weak refs processing", PrintGCDetails, false, _gc_timer_cm); 6214 6215 ReferenceProcessorStats stats; 6216 if (rp->processing_is_mt()) { 6217 // Set the degree of MT here. If the discovery is done MT, there 6218 // may have been a different number of threads doing the discovery 6219 // and a different number of discovered lists may have Ref objects. 6220 // That is OK as long as the Reference lists are balanced (see 6221 // balance_all_queues() and balance_queues()). 6222 GenCollectedHeap* gch = GenCollectedHeap::heap(); 6223 int active_workers = ParallelGCThreads; 6224 FlexibleWorkGang* workers = gch->workers(); 6225 if (workers != NULL) { 6226 active_workers = workers->active_workers(); 6227 // The expectation is that active_workers will have already 6228 // been set to a reasonable value. If it has not been set, 6229 // investigate. 6230 assert(active_workers > 0, "Should have been set during scavenge"); 6231 } 6232 rp->set_active_mt_degree(active_workers); 6233 CMSRefProcTaskExecutor task_executor(*this); 6234 stats = rp->process_discovered_references(&_is_alive_closure, 6235 &cmsKeepAliveClosure, 6236 &cmsDrainMarkingStackClosure, 6237 &task_executor, 6238 _gc_timer_cm); 6239 } else { 6240 stats = rp->process_discovered_references(&_is_alive_closure, 6241 &cmsKeepAliveClosure, 6242 &cmsDrainMarkingStackClosure, 6243 NULL, 6244 _gc_timer_cm); 6245 } 6246 _gc_tracer_cm->report_gc_reference_stats(stats); 6247 6248 } 6249 6250 // This is the point where the entire marking should have completed. 6251 verify_work_stacks_empty(); 6252 6253 if (should_unload_classes()) { 6254 { 6255 GCTraceTime t("class unloading", PrintGCDetails, false, _gc_timer_cm); 6256 6257 // Unload classes and purge the SystemDictionary. 6258 bool purged_class = SystemDictionary::do_unloading(&_is_alive_closure); 6259 6260 // Unload nmethods. 6261 CodeCache::do_unloading(&_is_alive_closure, purged_class); 6262 6263 // Prune dead klasses from subklass/sibling/implementor lists. 6264 Klass::clean_weak_klass_links(&_is_alive_closure); 6265 } 6266 6267 { 6268 GCTraceTime t("scrub symbol table", PrintGCDetails, false, _gc_timer_cm); 6269 // Clean up unreferenced symbols in symbol table. 6270 SymbolTable::unlink(); 6271 } 6272 } 6273 6274 // CMS doesn't use the StringTable as hard roots when class unloading is turned off. 6275 // Need to check if we really scanned the StringTable. 6276 if ((roots_scanning_options() & SharedHeap::SO_Strings) == 0) { 6277 GCTraceTime t("scrub string table", PrintGCDetails, false, _gc_timer_cm); 6278 // Delete entries for dead interned strings. 6279 StringTable::unlink(&_is_alive_closure); 6280 } 6281 6282 // Restore any preserved marks as a result of mark stack or 6283 // work queue overflow 6284 restore_preserved_marks_if_any(); // done single-threaded for now 6285 6286 rp->set_enqueuing_is_done(true); 6287 if (rp->processing_is_mt()) { 6288 rp->balance_all_queues(); 6289 CMSRefProcTaskExecutor task_executor(*this); 6290 rp->enqueue_discovered_references(&task_executor); 6291 } else { 6292 rp->enqueue_discovered_references(NULL); 6293 } 6294 rp->verify_no_references_recorded(); 6295 assert(!rp->discovery_enabled(), "should have been disabled"); 6296 } 6297 6298 #ifndef PRODUCT 6299 void CMSCollector::check_correct_thread_executing() { 6300 Thread* t = Thread::current(); 6301 // Only the VM thread or the CMS thread should be here. 6302 assert(t->is_ConcurrentGC_thread() || t->is_VM_thread(), 6303 "Unexpected thread type"); 6304 // If this is the vm thread, the foreground process 6305 // should not be waiting. Note that _foregroundGCIsActive is 6306 // true while the foreground collector is waiting. 6307 if (_foregroundGCShouldWait) { 6308 // We cannot be the VM thread 6309 assert(t->is_ConcurrentGC_thread(), 6310 "Should be CMS thread"); 6311 } else { 6312 // We can be the CMS thread only if we are in a stop-world 6313 // phase of CMS collection. 6314 if (t->is_ConcurrentGC_thread()) { 6315 assert(_collectorState == InitialMarking || 6316 _collectorState == FinalMarking, 6317 "Should be a stop-world phase"); 6318 // The CMS thread should be holding the CMS_token. 6319 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6320 "Potential interference with concurrently " 6321 "executing VM thread"); 6322 } 6323 } 6324 } 6325 #endif 6326 6327 void CMSCollector::sweep(bool asynch) { 6328 assert(_collectorState == Sweeping, "just checking"); 6329 check_correct_thread_executing(); 6330 verify_work_stacks_empty(); 6331 verify_overflow_empty(); 6332 increment_sweep_count(); 6333 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 6334 6335 _inter_sweep_timer.stop(); 6336 _inter_sweep_estimate.sample(_inter_sweep_timer.seconds()); 6337 size_policy()->avg_cms_free_at_sweep()->sample(_cmsGen->free()); 6338 6339 assert(!_intra_sweep_timer.is_active(), "Should not be active"); 6340 _intra_sweep_timer.reset(); 6341 _intra_sweep_timer.start(); 6342 if (asynch) { 6343 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 6344 CMSPhaseAccounting pa(this, "sweep", !PrintGCDetails); 6345 // First sweep the old gen 6346 { 6347 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(), 6348 bitMapLock()); 6349 sweepWork(_cmsGen, asynch); 6350 } 6351 6352 // Update Universe::_heap_*_at_gc figures. 6353 // We need all the free list locks to make the abstract state 6354 // transition from Sweeping to Resetting. See detailed note 6355 // further below. 6356 { 6357 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock()); 6358 // Update heap occupancy information which is used as 6359 // input to soft ref clearing policy at the next gc. 6360 Universe::update_heap_info_at_gc(); 6361 _collectorState = Resizing; 6362 } 6363 } else { 6364 // already have needed locks 6365 sweepWork(_cmsGen, asynch); 6366 // Update heap occupancy information which is used as 6367 // input to soft ref clearing policy at the next gc. 6368 Universe::update_heap_info_at_gc(); 6369 _collectorState = Resizing; 6370 } 6371 verify_work_stacks_empty(); 6372 verify_overflow_empty(); 6373 6374 if (should_unload_classes()) { 6375 ClassLoaderDataGraph::purge(); 6376 } 6377 6378 _intra_sweep_timer.stop(); 6379 _intra_sweep_estimate.sample(_intra_sweep_timer.seconds()); 6380 6381 _inter_sweep_timer.reset(); 6382 _inter_sweep_timer.start(); 6383 6384 // We need to use a monotonically non-deccreasing time in ms 6385 // or we will see time-warp warnings and os::javaTimeMillis() 6386 // does not guarantee monotonicity. 6387 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC; 6388 update_time_of_last_gc(now); 6389 6390 // NOTE on abstract state transitions: 6391 // Mutators allocate-live and/or mark the mod-union table dirty 6392 // based on the state of the collection. The former is done in 6393 // the interval [Marking, Sweeping] and the latter in the interval 6394 // [Marking, Sweeping). Thus the transitions into the Marking state 6395 // and out of the Sweeping state must be synchronously visible 6396 // globally to the mutators. 6397 // The transition into the Marking state happens with the world 6398 // stopped so the mutators will globally see it. Sweeping is 6399 // done asynchronously by the background collector so the transition 6400 // from the Sweeping state to the Resizing state must be done 6401 // under the freelistLock (as is the check for whether to 6402 // allocate-live and whether to dirty the mod-union table). 6403 assert(_collectorState == Resizing, "Change of collector state to" 6404 " Resizing must be done under the freelistLocks (plural)"); 6405 6406 // Now that sweeping has been completed, we clear 6407 // the incremental_collection_failed flag, 6408 // thus inviting a younger gen collection to promote into 6409 // this generation. If such a promotion may still fail, 6410 // the flag will be set again when a young collection is 6411 // attempted. 6412 GenCollectedHeap* gch = GenCollectedHeap::heap(); 6413 gch->clear_incremental_collection_failed(); // Worth retrying as fresh space may have been freed up 6414 gch->update_full_collections_completed(_collection_count_start); 6415 } 6416 6417 // FIX ME!!! Looks like this belongs in CFLSpace, with 6418 // CMSGen merely delegating to it. 6419 void ConcurrentMarkSweepGeneration::setNearLargestChunk() { 6420 double nearLargestPercent = FLSLargestBlockCoalesceProximity; 6421 HeapWord* minAddr = _cmsSpace->bottom(); 6422 HeapWord* largestAddr = 6423 (HeapWord*) _cmsSpace->dictionary()->find_largest_dict(); 6424 if (largestAddr == NULL) { 6425 // The dictionary appears to be empty. In this case 6426 // try to coalesce at the end of the heap. 6427 largestAddr = _cmsSpace->end(); 6428 } 6429 size_t largestOffset = pointer_delta(largestAddr, minAddr); 6430 size_t nearLargestOffset = 6431 (size_t)((double)largestOffset * nearLargestPercent) - MinChunkSize; 6432 if (PrintFLSStatistics != 0) { 6433 gclog_or_tty->print_cr( 6434 "CMS: Large Block: " PTR_FORMAT ";" 6435 " Proximity: " PTR_FORMAT " -> " PTR_FORMAT, 6436 largestAddr, 6437 _cmsSpace->nearLargestChunk(), minAddr + nearLargestOffset); 6438 } 6439 _cmsSpace->set_nearLargestChunk(minAddr + nearLargestOffset); 6440 } 6441 6442 bool ConcurrentMarkSweepGeneration::isNearLargestChunk(HeapWord* addr) { 6443 return addr >= _cmsSpace->nearLargestChunk(); 6444 } 6445 6446 FreeChunk* ConcurrentMarkSweepGeneration::find_chunk_at_end() { 6447 return _cmsSpace->find_chunk_at_end(); 6448 } 6449 6450 void ConcurrentMarkSweepGeneration::update_gc_stats(int current_level, 6451 bool full) { 6452 // The next lower level has been collected. Gather any statistics 6453 // that are of interest at this point. 6454 if (!full && (current_level + 1) == level()) { 6455 // Gather statistics on the young generation collection. 6456 collector()->stats().record_gc0_end(used()); 6457 } 6458 } 6459 6460 CMSAdaptiveSizePolicy* ConcurrentMarkSweepGeneration::size_policy() { 6461 GenCollectedHeap* gch = GenCollectedHeap::heap(); 6462 assert(gch->kind() == CollectedHeap::GenCollectedHeap, 6463 "Wrong type of heap"); 6464 CMSAdaptiveSizePolicy* sp = (CMSAdaptiveSizePolicy*) 6465 gch->gen_policy()->size_policy(); 6466 assert(sp->is_gc_cms_adaptive_size_policy(), 6467 "Wrong type of size policy"); 6468 return sp; 6469 } 6470 6471 void ConcurrentMarkSweepGeneration::rotate_debug_collection_type() { 6472 if (PrintGCDetails && Verbose) { 6473 gclog_or_tty->print("Rotate from %d ", _debug_collection_type); 6474 } 6475 _debug_collection_type = (CollectionTypes) (_debug_collection_type + 1); 6476 _debug_collection_type = 6477 (CollectionTypes) (_debug_collection_type % Unknown_collection_type); 6478 if (PrintGCDetails && Verbose) { 6479 gclog_or_tty->print_cr("to %d ", _debug_collection_type); 6480 } 6481 } 6482 6483 void CMSCollector::sweepWork(ConcurrentMarkSweepGeneration* gen, 6484 bool asynch) { 6485 // We iterate over the space(s) underlying this generation, 6486 // checking the mark bit map to see if the bits corresponding 6487 // to specific blocks are marked or not. Blocks that are 6488 // marked are live and are not swept up. All remaining blocks 6489 // are swept up, with coalescing on-the-fly as we sweep up 6490 // contiguous free and/or garbage blocks: 6491 // We need to ensure that the sweeper synchronizes with allocators 6492 // and stop-the-world collectors. In particular, the following 6493 // locks are used: 6494 // . CMS token: if this is held, a stop the world collection cannot occur 6495 // . freelistLock: if this is held no allocation can occur from this 6496 // generation by another thread 6497 // . bitMapLock: if this is held, no other thread can access or update 6498 // 6499 6500 // Note that we need to hold the freelistLock if we use 6501 // block iterate below; else the iterator might go awry if 6502 // a mutator (or promotion) causes block contents to change 6503 // (for instance if the allocator divvies up a block). 6504 // If we hold the free list lock, for all practical purposes 6505 // young generation GC's can't occur (they'll usually need to 6506 // promote), so we might as well prevent all young generation 6507 // GC's while we do a sweeping step. For the same reason, we might 6508 // as well take the bit map lock for the entire duration 6509 6510 // check that we hold the requisite locks 6511 assert(have_cms_token(), "Should hold cms token"); 6512 assert( (asynch && ConcurrentMarkSweepThread::cms_thread_has_cms_token()) 6513 || (!asynch && ConcurrentMarkSweepThread::vm_thread_has_cms_token()), 6514 "Should possess CMS token to sweep"); 6515 assert_lock_strong(gen->freelistLock()); 6516 assert_lock_strong(bitMapLock()); 6517 6518 assert(!_inter_sweep_timer.is_active(), "Was switched off in an outer context"); 6519 assert(_intra_sweep_timer.is_active(), "Was switched on in an outer context"); 6520 gen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()), 6521 _inter_sweep_estimate.padded_average(), 6522 _intra_sweep_estimate.padded_average()); 6523 gen->setNearLargestChunk(); 6524 6525 { 6526 SweepClosure sweepClosure(this, gen, &_markBitMap, 6527 CMSYield && asynch); 6528 gen->cmsSpace()->blk_iterate_careful(&sweepClosure); 6529 // We need to free-up/coalesce garbage/blocks from a 6530 // co-terminal free run. This is done in the SweepClosure 6531 // destructor; so, do not remove this scope, else the 6532 // end-of-sweep-census below will be off by a little bit. 6533 } 6534 gen->cmsSpace()->sweep_completed(); 6535 gen->cmsSpace()->endSweepFLCensus(sweep_count()); 6536 if (should_unload_classes()) { // unloaded classes this cycle, 6537 _concurrent_cycles_since_last_unload = 0; // ... reset count 6538 } else { // did not unload classes, 6539 _concurrent_cycles_since_last_unload++; // ... increment count 6540 } 6541 } 6542 6543 // Reset CMS data structures (for now just the marking bit map) 6544 // preparatory for the next cycle. 6545 void CMSCollector::reset(bool asynch) { 6546 GenCollectedHeap* gch = GenCollectedHeap::heap(); 6547 CMSAdaptiveSizePolicy* sp = size_policy(); 6548 AdaptiveSizePolicyOutput(sp, gch->total_collections()); 6549 if (asynch) { 6550 CMSTokenSyncWithLocks ts(true, bitMapLock()); 6551 6552 // If the state is not "Resetting", the foreground thread 6553 // has done a collection and the resetting. 6554 if (_collectorState != Resetting) { 6555 assert(_collectorState == Idling, "The state should only change" 6556 " because the foreground collector has finished the collection"); 6557 return; 6558 } 6559 6560 // Clear the mark bitmap (no grey objects to start with) 6561 // for the next cycle. 6562 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 6563 CMSPhaseAccounting cmspa(this, "reset", !PrintGCDetails); 6564 6565 HeapWord* curAddr = _markBitMap.startWord(); 6566 while (curAddr < _markBitMap.endWord()) { 6567 size_t remaining = pointer_delta(_markBitMap.endWord(), curAddr); 6568 MemRegion chunk(curAddr, MIN2(CMSBitMapYieldQuantum, remaining)); 6569 _markBitMap.clear_large_range(chunk); 6570 if (ConcurrentMarkSweepThread::should_yield() && 6571 !foregroundGCIsActive() && 6572 CMSYield) { 6573 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6574 "CMS thread should hold CMS token"); 6575 assert_lock_strong(bitMapLock()); 6576 bitMapLock()->unlock(); 6577 ConcurrentMarkSweepThread::desynchronize(true); 6578 ConcurrentMarkSweepThread::acknowledge_yield_request(); 6579 stopTimer(); 6580 if (PrintCMSStatistics != 0) { 6581 incrementYields(); 6582 } 6583 icms_wait(); 6584 6585 // See the comment in coordinator_yield() 6586 for (unsigned i = 0; i < CMSYieldSleepCount && 6587 ConcurrentMarkSweepThread::should_yield() && 6588 !CMSCollector::foregroundGCIsActive(); ++i) { 6589 os::sleep(Thread::current(), 1, false); 6590 ConcurrentMarkSweepThread::acknowledge_yield_request(); 6591 } 6592 6593 ConcurrentMarkSweepThread::synchronize(true); 6594 bitMapLock()->lock_without_safepoint_check(); 6595 startTimer(); 6596 } 6597 curAddr = chunk.end(); 6598 } 6599 // A successful mostly concurrent collection has been done. 6600 // Because only the full (i.e., concurrent mode failure) collections 6601 // are being measured for gc overhead limits, clean the "near" flag 6602 // and count. 6603 sp->reset_gc_overhead_limit_count(); 6604 _collectorState = Idling; 6605 } else { 6606 // already have the lock 6607 assert(_collectorState == Resetting, "just checking"); 6608 assert_lock_strong(bitMapLock()); 6609 _markBitMap.clear_all(); 6610 _collectorState = Idling; 6611 } 6612 6613 // Stop incremental mode after a cycle completes, so that any future cycles 6614 // are triggered by allocation. 6615 stop_icms(); 6616 6617 NOT_PRODUCT( 6618 if (RotateCMSCollectionTypes) { 6619 _cmsGen->rotate_debug_collection_type(); 6620 } 6621 ) 6622 6623 register_gc_end(); 6624 } 6625 6626 void CMSCollector::do_CMS_operation(CMS_op_type op, GCCause::Cause gc_cause) { 6627 gclog_or_tty->date_stamp(PrintGC && PrintGCDateStamps); 6628 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 6629 GCTraceTime t(GCCauseString("GC", gc_cause), PrintGC, !PrintGCDetails, NULL); 6630 TraceCollectorStats tcs(counters()); 6631 6632 switch (op) { 6633 case CMS_op_checkpointRootsInitial: { 6634 SvcGCMarker sgcm(SvcGCMarker::OTHER); 6635 checkpointRootsInitial(true); // asynch 6636 if (PrintGC) { 6637 _cmsGen->printOccupancy("initial-mark"); 6638 } 6639 break; 6640 } 6641 case CMS_op_checkpointRootsFinal: { 6642 SvcGCMarker sgcm(SvcGCMarker::OTHER); 6643 checkpointRootsFinal(true, // asynch 6644 false, // !clear_all_soft_refs 6645 false); // !init_mark_was_synchronous 6646 if (PrintGC) { 6647 _cmsGen->printOccupancy("remark"); 6648 } 6649 break; 6650 } 6651 default: 6652 fatal("No such CMS_op"); 6653 } 6654 } 6655 6656 #ifndef PRODUCT 6657 size_t const CMSCollector::skip_header_HeapWords() { 6658 return FreeChunk::header_size(); 6659 } 6660 6661 // Try and collect here conditions that should hold when 6662 // CMS thread is exiting. The idea is that the foreground GC 6663 // thread should not be blocked if it wants to terminate 6664 // the CMS thread and yet continue to run the VM for a while 6665 // after that. 6666 void CMSCollector::verify_ok_to_terminate() const { 6667 assert(Thread::current()->is_ConcurrentGC_thread(), 6668 "should be called by CMS thread"); 6669 assert(!_foregroundGCShouldWait, "should be false"); 6670 // We could check here that all the various low-level locks 6671 // are not held by the CMS thread, but that is overkill; see 6672 // also CMSThread::verify_ok_to_terminate() where the CGC_lock 6673 // is checked. 6674 } 6675 #endif 6676 6677 size_t CMSCollector::block_size_using_printezis_bits(HeapWord* addr) const { 6678 assert(_markBitMap.isMarked(addr) && _markBitMap.isMarked(addr + 1), 6679 "missing Printezis mark?"); 6680 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 6681 size_t size = pointer_delta(nextOneAddr + 1, addr); 6682 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 6683 "alignment problem"); 6684 assert(size >= 3, "Necessary for Printezis marks to work"); 6685 return size; 6686 } 6687 6688 // A variant of the above (block_size_using_printezis_bits()) except 6689 // that we return 0 if the P-bits are not yet set. 6690 size_t CMSCollector::block_size_if_printezis_bits(HeapWord* addr) const { 6691 if (_markBitMap.isMarked(addr + 1)) { 6692 assert(_markBitMap.isMarked(addr), "P-bit can be set only for marked objects"); 6693 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 6694 size_t size = pointer_delta(nextOneAddr + 1, addr); 6695 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 6696 "alignment problem"); 6697 assert(size >= 3, "Necessary for Printezis marks to work"); 6698 return size; 6699 } 6700 return 0; 6701 } 6702 6703 HeapWord* CMSCollector::next_card_start_after_block(HeapWord* addr) const { 6704 size_t sz = 0; 6705 oop p = (oop)addr; 6706 if (p->klass_or_null() != NULL) { 6707 sz = CompactibleFreeListSpace::adjustObjectSize(p->size()); 6708 } else { 6709 sz = block_size_using_printezis_bits(addr); 6710 } 6711 assert(sz > 0, "size must be nonzero"); 6712 HeapWord* next_block = addr + sz; 6713 HeapWord* next_card = (HeapWord*)round_to((uintptr_t)next_block, 6714 CardTableModRefBS::card_size); 6715 assert(round_down((uintptr_t)addr, CardTableModRefBS::card_size) < 6716 round_down((uintptr_t)next_card, CardTableModRefBS::card_size), 6717 "must be different cards"); 6718 return next_card; 6719 } 6720 6721 6722 // CMS Bit Map Wrapper ///////////////////////////////////////// 6723 6724 // Construct a CMS bit map infrastructure, but don't create the 6725 // bit vector itself. That is done by a separate call CMSBitMap::allocate() 6726 // further below. 6727 CMSBitMap::CMSBitMap(int shifter, int mutex_rank, const char* mutex_name): 6728 _bm(), 6729 _shifter(shifter), 6730 _lock(mutex_rank >= 0 ? new Mutex(mutex_rank, mutex_name, true) : NULL) 6731 { 6732 _bmStartWord = 0; 6733 _bmWordSize = 0; 6734 } 6735 6736 bool CMSBitMap::allocate(MemRegion mr) { 6737 _bmStartWord = mr.start(); 6738 _bmWordSize = mr.word_size(); 6739 ReservedSpace brs(ReservedSpace::allocation_align_size_up( 6740 (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1)); 6741 if (!brs.is_reserved()) { 6742 warning("CMS bit map allocation failure"); 6743 return false; 6744 } 6745 // For now we'll just commit all of the bit map up fromt. 6746 // Later on we'll try to be more parsimonious with swap. 6747 if (!_virtual_space.initialize(brs, brs.size())) { 6748 warning("CMS bit map backing store failure"); 6749 return false; 6750 } 6751 assert(_virtual_space.committed_size() == brs.size(), 6752 "didn't reserve backing store for all of CMS bit map?"); 6753 _bm.set_map((BitMap::bm_word_t*)_virtual_space.low()); 6754 assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >= 6755 _bmWordSize, "inconsistency in bit map sizing"); 6756 _bm.set_size(_bmWordSize >> _shifter); 6757 6758 // bm.clear(); // can we rely on getting zero'd memory? verify below 6759 assert(isAllClear(), 6760 "Expected zero'd memory from ReservedSpace constructor"); 6761 assert(_bm.size() == heapWordDiffToOffsetDiff(sizeInWords()), 6762 "consistency check"); 6763 return true; 6764 } 6765 6766 void CMSBitMap::dirty_range_iterate_clear(MemRegion mr, MemRegionClosure* cl) { 6767 HeapWord *next_addr, *end_addr, *last_addr; 6768 assert_locked(); 6769 assert(covers(mr), "out-of-range error"); 6770 // XXX assert that start and end are appropriately aligned 6771 for (next_addr = mr.start(), end_addr = mr.end(); 6772 next_addr < end_addr; next_addr = last_addr) { 6773 MemRegion dirty_region = getAndClearMarkedRegion(next_addr, end_addr); 6774 last_addr = dirty_region.end(); 6775 if (!dirty_region.is_empty()) { 6776 cl->do_MemRegion(dirty_region); 6777 } else { 6778 assert(last_addr == end_addr, "program logic"); 6779 return; 6780 } 6781 } 6782 } 6783 6784 void CMSBitMap::print_on_error(outputStream* st, const char* prefix) const { 6785 _bm.print_on_error(st, prefix); 6786 } 6787 6788 #ifndef PRODUCT 6789 void CMSBitMap::assert_locked() const { 6790 CMSLockVerifier::assert_locked(lock()); 6791 } 6792 6793 bool CMSBitMap::covers(MemRegion mr) const { 6794 // assert(_bm.map() == _virtual_space.low(), "map inconsistency"); 6795 assert((size_t)_bm.size() == (_bmWordSize >> _shifter), 6796 "size inconsistency"); 6797 return (mr.start() >= _bmStartWord) && 6798 (mr.end() <= endWord()); 6799 } 6800 6801 bool CMSBitMap::covers(HeapWord* start, size_t size) const { 6802 return (start >= _bmStartWord && (start + size) <= endWord()); 6803 } 6804 6805 void CMSBitMap::verifyNoOneBitsInRange(HeapWord* left, HeapWord* right) { 6806 // verify that there are no 1 bits in the interval [left, right) 6807 FalseBitMapClosure falseBitMapClosure; 6808 iterate(&falseBitMapClosure, left, right); 6809 } 6810 6811 void CMSBitMap::region_invariant(MemRegion mr) 6812 { 6813 assert_locked(); 6814 // mr = mr.intersection(MemRegion(_bmStartWord, _bmWordSize)); 6815 assert(!mr.is_empty(), "unexpected empty region"); 6816 assert(covers(mr), "mr should be covered by bit map"); 6817 // convert address range into offset range 6818 size_t start_ofs = heapWordToOffset(mr.start()); 6819 // Make sure that end() is appropriately aligned 6820 assert(mr.end() == (HeapWord*)round_to((intptr_t)mr.end(), 6821 (1 << (_shifter+LogHeapWordSize))), 6822 "Misaligned mr.end()"); 6823 size_t end_ofs = heapWordToOffset(mr.end()); 6824 assert(end_ofs > start_ofs, "Should mark at least one bit"); 6825 } 6826 6827 #endif 6828 6829 bool CMSMarkStack::allocate(size_t size) { 6830 // allocate a stack of the requisite depth 6831 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 6832 size * sizeof(oop))); 6833 if (!rs.is_reserved()) { 6834 warning("CMSMarkStack allocation failure"); 6835 return false; 6836 } 6837 if (!_virtual_space.initialize(rs, rs.size())) { 6838 warning("CMSMarkStack backing store failure"); 6839 return false; 6840 } 6841 assert(_virtual_space.committed_size() == rs.size(), 6842 "didn't reserve backing store for all of CMS stack?"); 6843 _base = (oop*)(_virtual_space.low()); 6844 _index = 0; 6845 _capacity = size; 6846 NOT_PRODUCT(_max_depth = 0); 6847 return true; 6848 } 6849 6850 // XXX FIX ME !!! In the MT case we come in here holding a 6851 // leaf lock. For printing we need to take a further lock 6852 // which has lower rank. We need to recallibrate the two 6853 // lock-ranks involved in order to be able to rpint the 6854 // messages below. (Or defer the printing to the caller. 6855 // For now we take the expedient path of just disabling the 6856 // messages for the problematic case.) 6857 void CMSMarkStack::expand() { 6858 assert(_capacity <= MarkStackSizeMax, "stack bigger than permitted"); 6859 if (_capacity == MarkStackSizeMax) { 6860 if (_hit_limit++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) { 6861 // We print a warning message only once per CMS cycle. 6862 gclog_or_tty->print_cr(" (benign) Hit CMSMarkStack max size limit"); 6863 } 6864 return; 6865 } 6866 // Double capacity if possible 6867 size_t new_capacity = MIN2(_capacity*2, MarkStackSizeMax); 6868 // Do not give up existing stack until we have managed to 6869 // get the double capacity that we desired. 6870 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 6871 new_capacity * sizeof(oop))); 6872 if (rs.is_reserved()) { 6873 // Release the backing store associated with old stack 6874 _virtual_space.release(); 6875 // Reinitialize virtual space for new stack 6876 if (!_virtual_space.initialize(rs, rs.size())) { 6877 fatal("Not enough swap for expanded marking stack"); 6878 } 6879 _base = (oop*)(_virtual_space.low()); 6880 _index = 0; 6881 _capacity = new_capacity; 6882 } else if (_failed_double++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) { 6883 // Failed to double capacity, continue; 6884 // we print a detail message only once per CMS cycle. 6885 gclog_or_tty->print(" (benign) Failed to expand marking stack from "SIZE_FORMAT"K to " 6886 SIZE_FORMAT"K", 6887 _capacity / K, new_capacity / K); 6888 } 6889 } 6890 6891 6892 // Closures 6893 // XXX: there seems to be a lot of code duplication here; 6894 // should refactor and consolidate common code. 6895 6896 // This closure is used to mark refs into the CMS generation in 6897 // the CMS bit map. Called at the first checkpoint. This closure 6898 // assumes that we do not need to re-mark dirty cards; if the CMS 6899 // generation on which this is used is not an oldest 6900 // generation then this will lose younger_gen cards! 6901 6902 MarkRefsIntoClosure::MarkRefsIntoClosure( 6903 MemRegion span, CMSBitMap* bitMap): 6904 _span(span), 6905 _bitMap(bitMap) 6906 { 6907 assert(_ref_processor == NULL, "deliberately left NULL"); 6908 assert(_bitMap->covers(_span), "_bitMap/_span mismatch"); 6909 } 6910 6911 void MarkRefsIntoClosure::do_oop(oop obj) { 6912 // if p points into _span, then mark corresponding bit in _markBitMap 6913 assert(obj->is_oop(), "expected an oop"); 6914 HeapWord* addr = (HeapWord*)obj; 6915 if (_span.contains(addr)) { 6916 // this should be made more efficient 6917 _bitMap->mark(addr); 6918 } 6919 } 6920 6921 void MarkRefsIntoClosure::do_oop(oop* p) { MarkRefsIntoClosure::do_oop_work(p); } 6922 void MarkRefsIntoClosure::do_oop(narrowOop* p) { MarkRefsIntoClosure::do_oop_work(p); } 6923 6924 Par_MarkRefsIntoClosure::Par_MarkRefsIntoClosure( 6925 MemRegion span, CMSBitMap* bitMap): 6926 _span(span), 6927 _bitMap(bitMap) 6928 { 6929 assert(_ref_processor == NULL, "deliberately left NULL"); 6930 assert(_bitMap->covers(_span), "_bitMap/_span mismatch"); 6931 } 6932 6933 void Par_MarkRefsIntoClosure::do_oop(oop obj) { 6934 // if p points into _span, then mark corresponding bit in _markBitMap 6935 assert(obj->is_oop(), "expected an oop"); 6936 HeapWord* addr = (HeapWord*)obj; 6937 if (_span.contains(addr)) { 6938 // this should be made more efficient 6939 _bitMap->par_mark(addr); 6940 } 6941 } 6942 6943 void Par_MarkRefsIntoClosure::do_oop(oop* p) { Par_MarkRefsIntoClosure::do_oop_work(p); } 6944 void Par_MarkRefsIntoClosure::do_oop(narrowOop* p) { Par_MarkRefsIntoClosure::do_oop_work(p); } 6945 6946 // A variant of the above, used for CMS marking verification. 6947 MarkRefsIntoVerifyClosure::MarkRefsIntoVerifyClosure( 6948 MemRegion span, CMSBitMap* verification_bm, CMSBitMap* cms_bm): 6949 _span(span), 6950 _verification_bm(verification_bm), 6951 _cms_bm(cms_bm) 6952 { 6953 assert(_ref_processor == NULL, "deliberately left NULL"); 6954 assert(_verification_bm->covers(_span), "_verification_bm/_span mismatch"); 6955 } 6956 6957 void MarkRefsIntoVerifyClosure::do_oop(oop obj) { 6958 // if p points into _span, then mark corresponding bit in _markBitMap 6959 assert(obj->is_oop(), "expected an oop"); 6960 HeapWord* addr = (HeapWord*)obj; 6961 if (_span.contains(addr)) { 6962 _verification_bm->mark(addr); 6963 if (!_cms_bm->isMarked(addr)) { 6964 oop(addr)->print(); 6965 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)", addr); 6966 fatal("... aborting"); 6967 } 6968 } 6969 } 6970 6971 void MarkRefsIntoVerifyClosure::do_oop(oop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); } 6972 void MarkRefsIntoVerifyClosure::do_oop(narrowOop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); } 6973 6974 ////////////////////////////////////////////////// 6975 // MarkRefsIntoAndScanClosure 6976 ////////////////////////////////////////////////// 6977 6978 MarkRefsIntoAndScanClosure::MarkRefsIntoAndScanClosure(MemRegion span, 6979 ReferenceProcessor* rp, 6980 CMSBitMap* bit_map, 6981 CMSBitMap* mod_union_table, 6982 CMSMarkStack* mark_stack, 6983 CMSCollector* collector, 6984 bool should_yield, 6985 bool concurrent_precleaning): 6986 _collector(collector), 6987 _span(span), 6988 _bit_map(bit_map), 6989 _mark_stack(mark_stack), 6990 _pushAndMarkClosure(collector, span, rp, bit_map, mod_union_table, 6991 mark_stack, concurrent_precleaning), 6992 _yield(should_yield), 6993 _concurrent_precleaning(concurrent_precleaning), 6994 _freelistLock(NULL) 6995 { 6996 _ref_processor = rp; 6997 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 6998 } 6999 7000 // This closure is used to mark refs into the CMS generation at the 7001 // second (final) checkpoint, and to scan and transitively follow 7002 // the unmarked oops. It is also used during the concurrent precleaning 7003 // phase while scanning objects on dirty cards in the CMS generation. 7004 // The marks are made in the marking bit map and the marking stack is 7005 // used for keeping the (newly) grey objects during the scan. 7006 // The parallel version (Par_...) appears further below. 7007 void MarkRefsIntoAndScanClosure::do_oop(oop obj) { 7008 if (obj != NULL) { 7009 assert(obj->is_oop(), "expected an oop"); 7010 HeapWord* addr = (HeapWord*)obj; 7011 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 7012 assert(_collector->overflow_list_is_empty(), 7013 "overflow list should be empty"); 7014 if (_span.contains(addr) && 7015 !_bit_map->isMarked(addr)) { 7016 // mark bit map (object is now grey) 7017 _bit_map->mark(addr); 7018 // push on marking stack (stack should be empty), and drain the 7019 // stack by applying this closure to the oops in the oops popped 7020 // from the stack (i.e. blacken the grey objects) 7021 bool res = _mark_stack->push(obj); 7022 assert(res, "Should have space to push on empty stack"); 7023 do { 7024 oop new_oop = _mark_stack->pop(); 7025 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 7026 assert(_bit_map->isMarked((HeapWord*)new_oop), 7027 "only grey objects on this stack"); 7028 // iterate over the oops in this oop, marking and pushing 7029 // the ones in CMS heap (i.e. in _span). 7030 new_oop->oop_iterate(&_pushAndMarkClosure); 7031 // check if it's time to yield 7032 do_yield_check(); 7033 } while (!_mark_stack->isEmpty() || 7034 (!_concurrent_precleaning && take_from_overflow_list())); 7035 // if marking stack is empty, and we are not doing this 7036 // during precleaning, then check the overflow list 7037 } 7038 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 7039 assert(_collector->overflow_list_is_empty(), 7040 "overflow list was drained above"); 7041 // We could restore evacuated mark words, if any, used for 7042 // overflow list links here because the overflow list is 7043 // provably empty here. That would reduce the maximum 7044 // size requirements for preserved_{oop,mark}_stack. 7045 // But we'll just postpone it until we are all done 7046 // so we can just stream through. 7047 if (!_concurrent_precleaning && CMSOverflowEarlyRestoration) { 7048 _collector->restore_preserved_marks_if_any(); 7049 assert(_collector->no_preserved_marks(), "No preserved marks"); 7050 } 7051 assert(!CMSOverflowEarlyRestoration || _collector->no_preserved_marks(), 7052 "All preserved marks should have been restored above"); 7053 } 7054 } 7055 7056 void MarkRefsIntoAndScanClosure::do_oop(oop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); } 7057 void MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); } 7058 7059 void MarkRefsIntoAndScanClosure::do_yield_work() { 7060 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 7061 "CMS thread should hold CMS token"); 7062 assert_lock_strong(_freelistLock); 7063 assert_lock_strong(_bit_map->lock()); 7064 // relinquish the free_list_lock and bitMaplock() 7065 _bit_map->lock()->unlock(); 7066 _freelistLock->unlock(); 7067 ConcurrentMarkSweepThread::desynchronize(true); 7068 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7069 _collector->stopTimer(); 7070 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 7071 if (PrintCMSStatistics != 0) { 7072 _collector->incrementYields(); 7073 } 7074 _collector->icms_wait(); 7075 7076 // See the comment in coordinator_yield() 7077 for (unsigned i = 0; 7078 i < CMSYieldSleepCount && 7079 ConcurrentMarkSweepThread::should_yield() && 7080 !CMSCollector::foregroundGCIsActive(); 7081 ++i) { 7082 os::sleep(Thread::current(), 1, false); 7083 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7084 } 7085 7086 ConcurrentMarkSweepThread::synchronize(true); 7087 _freelistLock->lock_without_safepoint_check(); 7088 _bit_map->lock()->lock_without_safepoint_check(); 7089 _collector->startTimer(); 7090 } 7091 7092 /////////////////////////////////////////////////////////// 7093 // Par_MarkRefsIntoAndScanClosure: a parallel version of 7094 // MarkRefsIntoAndScanClosure 7095 /////////////////////////////////////////////////////////// 7096 Par_MarkRefsIntoAndScanClosure::Par_MarkRefsIntoAndScanClosure( 7097 CMSCollector* collector, MemRegion span, ReferenceProcessor* rp, 7098 CMSBitMap* bit_map, OopTaskQueue* work_queue): 7099 _span(span), 7100 _bit_map(bit_map), 7101 _work_queue(work_queue), 7102 _low_water_mark(MIN2((uint)(work_queue->max_elems()/4), 7103 (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads))), 7104 _par_pushAndMarkClosure(collector, span, rp, bit_map, work_queue) 7105 { 7106 _ref_processor = rp; 7107 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 7108 } 7109 7110 // This closure is used to mark refs into the CMS generation at the 7111 // second (final) checkpoint, and to scan and transitively follow 7112 // the unmarked oops. The marks are made in the marking bit map and 7113 // the work_queue is used for keeping the (newly) grey objects during 7114 // the scan phase whence they are also available for stealing by parallel 7115 // threads. Since the marking bit map is shared, updates are 7116 // synchronized (via CAS). 7117 void Par_MarkRefsIntoAndScanClosure::do_oop(oop obj) { 7118 if (obj != NULL) { 7119 // Ignore mark word because this could be an already marked oop 7120 // that may be chained at the end of the overflow list. 7121 assert(obj->is_oop(true), "expected an oop"); 7122 HeapWord* addr = (HeapWord*)obj; 7123 if (_span.contains(addr) && 7124 !_bit_map->isMarked(addr)) { 7125 // mark bit map (object will become grey): 7126 // It is possible for several threads to be 7127 // trying to "claim" this object concurrently; 7128 // the unique thread that succeeds in marking the 7129 // object first will do the subsequent push on 7130 // to the work queue (or overflow list). 7131 if (_bit_map->par_mark(addr)) { 7132 // push on work_queue (which may not be empty), and trim the 7133 // queue to an appropriate length by applying this closure to 7134 // the oops in the oops popped from the stack (i.e. blacken the 7135 // grey objects) 7136 bool res = _work_queue->push(obj); 7137 assert(res, "Low water mark should be less than capacity?"); 7138 trim_queue(_low_water_mark); 7139 } // Else, another thread claimed the object 7140 } 7141 } 7142 } 7143 7144 void Par_MarkRefsIntoAndScanClosure::do_oop(oop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); } 7145 void Par_MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); } 7146 7147 // This closure is used to rescan the marked objects on the dirty cards 7148 // in the mod union table and the card table proper. 7149 size_t ScanMarkedObjectsAgainCarefullyClosure::do_object_careful_m( 7150 oop p, MemRegion mr) { 7151 7152 size_t size = 0; 7153 HeapWord* addr = (HeapWord*)p; 7154 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 7155 assert(_span.contains(addr), "we are scanning the CMS generation"); 7156 // check if it's time to yield 7157 if (do_yield_check()) { 7158 // We yielded for some foreground stop-world work, 7159 // and we have been asked to abort this ongoing preclean cycle. 7160 return 0; 7161 } 7162 if (_bitMap->isMarked(addr)) { 7163 // it's marked; is it potentially uninitialized? 7164 if (p->klass_or_null() != NULL) { 7165 // an initialized object; ignore mark word in verification below 7166 // since we are running concurrent with mutators 7167 assert(p->is_oop(true), "should be an oop"); 7168 if (p->is_objArray()) { 7169 // objArrays are precisely marked; restrict scanning 7170 // to dirty cards only. 7171 size = CompactibleFreeListSpace::adjustObjectSize( 7172 p->oop_iterate(_scanningClosure, mr)); 7173 } else { 7174 // A non-array may have been imprecisely marked; we need 7175 // to scan object in its entirety. 7176 size = CompactibleFreeListSpace::adjustObjectSize( 7177 p->oop_iterate(_scanningClosure)); 7178 } 7179 #ifdef ASSERT 7180 size_t direct_size = 7181 CompactibleFreeListSpace::adjustObjectSize(p->size()); 7182 assert(size == direct_size, "Inconsistency in size"); 7183 assert(size >= 3, "Necessary for Printezis marks to work"); 7184 if (!_bitMap->isMarked(addr+1)) { 7185 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size); 7186 } else { 7187 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size-1); 7188 assert(_bitMap->isMarked(addr+size-1), 7189 "inconsistent Printezis mark"); 7190 } 7191 #endif // ASSERT 7192 } else { 7193 // an unitialized object 7194 assert(_bitMap->isMarked(addr+1), "missing Printezis mark?"); 7195 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 7196 size = pointer_delta(nextOneAddr + 1, addr); 7197 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 7198 "alignment problem"); 7199 // Note that pre-cleaning needn't redirty the card. OopDesc::set_klass() 7200 // will dirty the card when the klass pointer is installed in the 7201 // object (signalling the completion of initialization). 7202 } 7203 } else { 7204 // Either a not yet marked object or an uninitialized object 7205 if (p->klass_or_null() == NULL) { 7206 // An uninitialized object, skip to the next card, since 7207 // we may not be able to read its P-bits yet. 7208 assert(size == 0, "Initial value"); 7209 } else { 7210 // An object not (yet) reached by marking: we merely need to 7211 // compute its size so as to go look at the next block. 7212 assert(p->is_oop(true), "should be an oop"); 7213 size = CompactibleFreeListSpace::adjustObjectSize(p->size()); 7214 } 7215 } 7216 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 7217 return size; 7218 } 7219 7220 void ScanMarkedObjectsAgainCarefullyClosure::do_yield_work() { 7221 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 7222 "CMS thread should hold CMS token"); 7223 assert_lock_strong(_freelistLock); 7224 assert_lock_strong(_bitMap->lock()); 7225 // relinquish the free_list_lock and bitMaplock() 7226 _bitMap->lock()->unlock(); 7227 _freelistLock->unlock(); 7228 ConcurrentMarkSweepThread::desynchronize(true); 7229 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7230 _collector->stopTimer(); 7231 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 7232 if (PrintCMSStatistics != 0) { 7233 _collector->incrementYields(); 7234 } 7235 _collector->icms_wait(); 7236 7237 // See the comment in coordinator_yield() 7238 for (unsigned i = 0; i < CMSYieldSleepCount && 7239 ConcurrentMarkSweepThread::should_yield() && 7240 !CMSCollector::foregroundGCIsActive(); ++i) { 7241 os::sleep(Thread::current(), 1, false); 7242 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7243 } 7244 7245 ConcurrentMarkSweepThread::synchronize(true); 7246 _freelistLock->lock_without_safepoint_check(); 7247 _bitMap->lock()->lock_without_safepoint_check(); 7248 _collector->startTimer(); 7249 } 7250 7251 7252 ////////////////////////////////////////////////////////////////// 7253 // SurvivorSpacePrecleanClosure 7254 ////////////////////////////////////////////////////////////////// 7255 // This (single-threaded) closure is used to preclean the oops in 7256 // the survivor spaces. 7257 size_t SurvivorSpacePrecleanClosure::do_object_careful(oop p) { 7258 7259 HeapWord* addr = (HeapWord*)p; 7260 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 7261 assert(!_span.contains(addr), "we are scanning the survivor spaces"); 7262 assert(p->klass_or_null() != NULL, "object should be initializd"); 7263 // an initialized object; ignore mark word in verification below 7264 // since we are running concurrent with mutators 7265 assert(p->is_oop(true), "should be an oop"); 7266 // Note that we do not yield while we iterate over 7267 // the interior oops of p, pushing the relevant ones 7268 // on our marking stack. 7269 size_t size = p->oop_iterate(_scanning_closure); 7270 do_yield_check(); 7271 // Observe that below, we do not abandon the preclean 7272 // phase as soon as we should; rather we empty the 7273 // marking stack before returning. This is to satisfy 7274 // some existing assertions. In general, it may be a 7275 // good idea to abort immediately and complete the marking 7276 // from the grey objects at a later time. 7277 while (!_mark_stack->isEmpty()) { 7278 oop new_oop = _mark_stack->pop(); 7279 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 7280 assert(_bit_map->isMarked((HeapWord*)new_oop), 7281 "only grey objects on this stack"); 7282 // iterate over the oops in this oop, marking and pushing 7283 // the ones in CMS heap (i.e. in _span). 7284 new_oop->oop_iterate(_scanning_closure); 7285 // check if it's time to yield 7286 do_yield_check(); 7287 } 7288 unsigned int after_count = 7289 GenCollectedHeap::heap()->total_collections(); 7290 bool abort = (_before_count != after_count) || 7291 _collector->should_abort_preclean(); 7292 return abort ? 0 : size; 7293 } 7294 7295 void SurvivorSpacePrecleanClosure::do_yield_work() { 7296 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 7297 "CMS thread should hold CMS token"); 7298 assert_lock_strong(_bit_map->lock()); 7299 // Relinquish the bit map lock 7300 _bit_map->lock()->unlock(); 7301 ConcurrentMarkSweepThread::desynchronize(true); 7302 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7303 _collector->stopTimer(); 7304 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 7305 if (PrintCMSStatistics != 0) { 7306 _collector->incrementYields(); 7307 } 7308 _collector->icms_wait(); 7309 7310 // See the comment in coordinator_yield() 7311 for (unsigned i = 0; i < CMSYieldSleepCount && 7312 ConcurrentMarkSweepThread::should_yield() && 7313 !CMSCollector::foregroundGCIsActive(); ++i) { 7314 os::sleep(Thread::current(), 1, false); 7315 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7316 } 7317 7318 ConcurrentMarkSweepThread::synchronize(true); 7319 _bit_map->lock()->lock_without_safepoint_check(); 7320 _collector->startTimer(); 7321 } 7322 7323 // This closure is used to rescan the marked objects on the dirty cards 7324 // in the mod union table and the card table proper. In the parallel 7325 // case, although the bitMap is shared, we do a single read so the 7326 // isMarked() query is "safe". 7327 bool ScanMarkedObjectsAgainClosure::do_object_bm(oop p, MemRegion mr) { 7328 // Ignore mark word because we are running concurrent with mutators 7329 assert(p->is_oop_or_null(true), "expected an oop or null"); 7330 HeapWord* addr = (HeapWord*)p; 7331 assert(_span.contains(addr), "we are scanning the CMS generation"); 7332 bool is_obj_array = false; 7333 #ifdef ASSERT 7334 if (!_parallel) { 7335 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 7336 assert(_collector->overflow_list_is_empty(), 7337 "overflow list should be empty"); 7338 7339 } 7340 #endif // ASSERT 7341 if (_bit_map->isMarked(addr)) { 7342 // Obj arrays are precisely marked, non-arrays are not; 7343 // so we scan objArrays precisely and non-arrays in their 7344 // entirety. 7345 if (p->is_objArray()) { 7346 is_obj_array = true; 7347 if (_parallel) { 7348 p->oop_iterate(_par_scan_closure, mr); 7349 } else { 7350 p->oop_iterate(_scan_closure, mr); 7351 } 7352 } else { 7353 if (_parallel) { 7354 p->oop_iterate(_par_scan_closure); 7355 } else { 7356 p->oop_iterate(_scan_closure); 7357 } 7358 } 7359 } 7360 #ifdef ASSERT 7361 if (!_parallel) { 7362 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 7363 assert(_collector->overflow_list_is_empty(), 7364 "overflow list should be empty"); 7365 7366 } 7367 #endif // ASSERT 7368 return is_obj_array; 7369 } 7370 7371 MarkFromRootsClosure::MarkFromRootsClosure(CMSCollector* collector, 7372 MemRegion span, 7373 CMSBitMap* bitMap, CMSMarkStack* markStack, 7374 bool should_yield, bool verifying): 7375 _collector(collector), 7376 _span(span), 7377 _bitMap(bitMap), 7378 _mut(&collector->_modUnionTable), 7379 _markStack(markStack), 7380 _yield(should_yield), 7381 _skipBits(0) 7382 { 7383 assert(_markStack->isEmpty(), "stack should be empty"); 7384 _finger = _bitMap->startWord(); 7385 _threshold = _finger; 7386 assert(_collector->_restart_addr == NULL, "Sanity check"); 7387 assert(_span.contains(_finger), "Out of bounds _finger?"); 7388 DEBUG_ONLY(_verifying = verifying;) 7389 } 7390 7391 void MarkFromRootsClosure::reset(HeapWord* addr) { 7392 assert(_markStack->isEmpty(), "would cause duplicates on stack"); 7393 assert(_span.contains(addr), "Out of bounds _finger?"); 7394 _finger = addr; 7395 _threshold = (HeapWord*)round_to( 7396 (intptr_t)_finger, CardTableModRefBS::card_size); 7397 } 7398 7399 // Should revisit to see if this should be restructured for 7400 // greater efficiency. 7401 bool MarkFromRootsClosure::do_bit(size_t offset) { 7402 if (_skipBits > 0) { 7403 _skipBits--; 7404 return true; 7405 } 7406 // convert offset into a HeapWord* 7407 HeapWord* addr = _bitMap->startWord() + offset; 7408 assert(_bitMap->endWord() && addr < _bitMap->endWord(), 7409 "address out of range"); 7410 assert(_bitMap->isMarked(addr), "tautology"); 7411 if (_bitMap->isMarked(addr+1)) { 7412 // this is an allocated but not yet initialized object 7413 assert(_skipBits == 0, "tautology"); 7414 _skipBits = 2; // skip next two marked bits ("Printezis-marks") 7415 oop p = oop(addr); 7416 if (p->klass_or_null() == NULL) { 7417 DEBUG_ONLY(if (!_verifying) {) 7418 // We re-dirty the cards on which this object lies and increase 7419 // the _threshold so that we'll come back to scan this object 7420 // during the preclean or remark phase. (CMSCleanOnEnter) 7421 if (CMSCleanOnEnter) { 7422 size_t sz = _collector->block_size_using_printezis_bits(addr); 7423 HeapWord* end_card_addr = (HeapWord*)round_to( 7424 (intptr_t)(addr+sz), CardTableModRefBS::card_size); 7425 MemRegion redirty_range = MemRegion(addr, end_card_addr); 7426 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 7427 // Bump _threshold to end_card_addr; note that 7428 // _threshold cannot possibly exceed end_card_addr, anyhow. 7429 // This prevents future clearing of the card as the scan proceeds 7430 // to the right. 7431 assert(_threshold <= end_card_addr, 7432 "Because we are just scanning into this object"); 7433 if (_threshold < end_card_addr) { 7434 _threshold = end_card_addr; 7435 } 7436 if (p->klass_or_null() != NULL) { 7437 // Redirty the range of cards... 7438 _mut->mark_range(redirty_range); 7439 } // ...else the setting of klass will dirty the card anyway. 7440 } 7441 DEBUG_ONLY(}) 7442 return true; 7443 } 7444 } 7445 scanOopsInOop(addr); 7446 return true; 7447 } 7448 7449 // We take a break if we've been at this for a while, 7450 // so as to avoid monopolizing the locks involved. 7451 void MarkFromRootsClosure::do_yield_work() { 7452 // First give up the locks, then yield, then re-lock 7453 // We should probably use a constructor/destructor idiom to 7454 // do this unlock/lock or modify the MutexUnlocker class to 7455 // serve our purpose. XXX 7456 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 7457 "CMS thread should hold CMS token"); 7458 assert_lock_strong(_bitMap->lock()); 7459 _bitMap->lock()->unlock(); 7460 ConcurrentMarkSweepThread::desynchronize(true); 7461 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7462 _collector->stopTimer(); 7463 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 7464 if (PrintCMSStatistics != 0) { 7465 _collector->incrementYields(); 7466 } 7467 _collector->icms_wait(); 7468 7469 // See the comment in coordinator_yield() 7470 for (unsigned i = 0; i < CMSYieldSleepCount && 7471 ConcurrentMarkSweepThread::should_yield() && 7472 !CMSCollector::foregroundGCIsActive(); ++i) { 7473 os::sleep(Thread::current(), 1, false); 7474 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7475 } 7476 7477 ConcurrentMarkSweepThread::synchronize(true); 7478 _bitMap->lock()->lock_without_safepoint_check(); 7479 _collector->startTimer(); 7480 } 7481 7482 void MarkFromRootsClosure::scanOopsInOop(HeapWord* ptr) { 7483 assert(_bitMap->isMarked(ptr), "expected bit to be set"); 7484 assert(_markStack->isEmpty(), 7485 "should drain stack to limit stack usage"); 7486 // convert ptr to an oop preparatory to scanning 7487 oop obj = oop(ptr); 7488 // Ignore mark word in verification below, since we 7489 // may be running concurrent with mutators. 7490 assert(obj->is_oop(true), "should be an oop"); 7491 assert(_finger <= ptr, "_finger runneth ahead"); 7492 // advance the finger to right end of this object 7493 _finger = ptr + obj->size(); 7494 assert(_finger > ptr, "we just incremented it above"); 7495 // On large heaps, it may take us some time to get through 7496 // the marking phase (especially if running iCMS). During 7497 // this time it's possible that a lot of mutations have 7498 // accumulated in the card table and the mod union table -- 7499 // these mutation records are redundant until we have 7500 // actually traced into the corresponding card. 7501 // Here, we check whether advancing the finger would make 7502 // us cross into a new card, and if so clear corresponding 7503 // cards in the MUT (preclean them in the card-table in the 7504 // future). 7505 7506 DEBUG_ONLY(if (!_verifying) {) 7507 // The clean-on-enter optimization is disabled by default, 7508 // until we fix 6178663. 7509 if (CMSCleanOnEnter && (_finger > _threshold)) { 7510 // [_threshold, _finger) represents the interval 7511 // of cards to be cleared in MUT (or precleaned in card table). 7512 // The set of cards to be cleared is all those that overlap 7513 // with the interval [_threshold, _finger); note that 7514 // _threshold is always kept card-aligned but _finger isn't 7515 // always card-aligned. 7516 HeapWord* old_threshold = _threshold; 7517 assert(old_threshold == (HeapWord*)round_to( 7518 (intptr_t)old_threshold, CardTableModRefBS::card_size), 7519 "_threshold should always be card-aligned"); 7520 _threshold = (HeapWord*)round_to( 7521 (intptr_t)_finger, CardTableModRefBS::card_size); 7522 MemRegion mr(old_threshold, _threshold); 7523 assert(!mr.is_empty(), "Control point invariant"); 7524 assert(_span.contains(mr), "Should clear within span"); 7525 _mut->clear_range(mr); 7526 } 7527 DEBUG_ONLY(}) 7528 // Note: the finger doesn't advance while we drain 7529 // the stack below. 7530 PushOrMarkClosure pushOrMarkClosure(_collector, 7531 _span, _bitMap, _markStack, 7532 _finger, this); 7533 bool res = _markStack->push(obj); 7534 assert(res, "Empty non-zero size stack should have space for single push"); 7535 while (!_markStack->isEmpty()) { 7536 oop new_oop = _markStack->pop(); 7537 // Skip verifying header mark word below because we are 7538 // running concurrent with mutators. 7539 assert(new_oop->is_oop(true), "Oops! expected to pop an oop"); 7540 // now scan this oop's oops 7541 new_oop->oop_iterate(&pushOrMarkClosure); 7542 do_yield_check(); 7543 } 7544 assert(_markStack->isEmpty(), "tautology, emphasizing post-condition"); 7545 } 7546 7547 Par_MarkFromRootsClosure::Par_MarkFromRootsClosure(CMSConcMarkingTask* task, 7548 CMSCollector* collector, MemRegion span, 7549 CMSBitMap* bit_map, 7550 OopTaskQueue* work_queue, 7551 CMSMarkStack* overflow_stack, 7552 bool should_yield): 7553 _collector(collector), 7554 _whole_span(collector->_span), 7555 _span(span), 7556 _bit_map(bit_map), 7557 _mut(&collector->_modUnionTable), 7558 _work_queue(work_queue), 7559 _overflow_stack(overflow_stack), 7560 _yield(should_yield), 7561 _skip_bits(0), 7562 _task(task) 7563 { 7564 assert(_work_queue->size() == 0, "work_queue should be empty"); 7565 _finger = span.start(); 7566 _threshold = _finger; // XXX Defer clear-on-enter optimization for now 7567 assert(_span.contains(_finger), "Out of bounds _finger?"); 7568 } 7569 7570 // Should revisit to see if this should be restructured for 7571 // greater efficiency. 7572 bool Par_MarkFromRootsClosure::do_bit(size_t offset) { 7573 if (_skip_bits > 0) { 7574 _skip_bits--; 7575 return true; 7576 } 7577 // convert offset into a HeapWord* 7578 HeapWord* addr = _bit_map->startWord() + offset; 7579 assert(_bit_map->endWord() && addr < _bit_map->endWord(), 7580 "address out of range"); 7581 assert(_bit_map->isMarked(addr), "tautology"); 7582 if (_bit_map->isMarked(addr+1)) { 7583 // this is an allocated object that might not yet be initialized 7584 assert(_skip_bits == 0, "tautology"); 7585 _skip_bits = 2; // skip next two marked bits ("Printezis-marks") 7586 oop p = oop(addr); 7587 if (p->klass_or_null() == NULL) { 7588 // in the case of Clean-on-Enter optimization, redirty card 7589 // and avoid clearing card by increasing the threshold. 7590 return true; 7591 } 7592 } 7593 scan_oops_in_oop(addr); 7594 return true; 7595 } 7596 7597 void Par_MarkFromRootsClosure::scan_oops_in_oop(HeapWord* ptr) { 7598 assert(_bit_map->isMarked(ptr), "expected bit to be set"); 7599 // Should we assert that our work queue is empty or 7600 // below some drain limit? 7601 assert(_work_queue->size() == 0, 7602 "should drain stack to limit stack usage"); 7603 // convert ptr to an oop preparatory to scanning 7604 oop obj = oop(ptr); 7605 // Ignore mark word in verification below, since we 7606 // may be running concurrent with mutators. 7607 assert(obj->is_oop(true), "should be an oop"); 7608 assert(_finger <= ptr, "_finger runneth ahead"); 7609 // advance the finger to right end of this object 7610 _finger = ptr + obj->size(); 7611 assert(_finger > ptr, "we just incremented it above"); 7612 // On large heaps, it may take us some time to get through 7613 // the marking phase (especially if running iCMS). During 7614 // this time it's possible that a lot of mutations have 7615 // accumulated in the card table and the mod union table -- 7616 // these mutation records are redundant until we have 7617 // actually traced into the corresponding card. 7618 // Here, we check whether advancing the finger would make 7619 // us cross into a new card, and if so clear corresponding 7620 // cards in the MUT (preclean them in the card-table in the 7621 // future). 7622 7623 // The clean-on-enter optimization is disabled by default, 7624 // until we fix 6178663. 7625 if (CMSCleanOnEnter && (_finger > _threshold)) { 7626 // [_threshold, _finger) represents the interval 7627 // of cards to be cleared in MUT (or precleaned in card table). 7628 // The set of cards to be cleared is all those that overlap 7629 // with the interval [_threshold, _finger); note that 7630 // _threshold is always kept card-aligned but _finger isn't 7631 // always card-aligned. 7632 HeapWord* old_threshold = _threshold; 7633 assert(old_threshold == (HeapWord*)round_to( 7634 (intptr_t)old_threshold, CardTableModRefBS::card_size), 7635 "_threshold should always be card-aligned"); 7636 _threshold = (HeapWord*)round_to( 7637 (intptr_t)_finger, CardTableModRefBS::card_size); 7638 MemRegion mr(old_threshold, _threshold); 7639 assert(!mr.is_empty(), "Control point invariant"); 7640 assert(_span.contains(mr), "Should clear within span"); // _whole_span ?? 7641 _mut->clear_range(mr); 7642 } 7643 7644 // Note: the local finger doesn't advance while we drain 7645 // the stack below, but the global finger sure can and will. 7646 HeapWord** gfa = _task->global_finger_addr(); 7647 Par_PushOrMarkClosure pushOrMarkClosure(_collector, 7648 _span, _bit_map, 7649 _work_queue, 7650 _overflow_stack, 7651 _finger, 7652 gfa, this); 7653 bool res = _work_queue->push(obj); // overflow could occur here 7654 assert(res, "Will hold once we use workqueues"); 7655 while (true) { 7656 oop new_oop; 7657 if (!_work_queue->pop_local(new_oop)) { 7658 // We emptied our work_queue; check if there's stuff that can 7659 // be gotten from the overflow stack. 7660 if (CMSConcMarkingTask::get_work_from_overflow_stack( 7661 _overflow_stack, _work_queue)) { 7662 do_yield_check(); 7663 continue; 7664 } else { // done 7665 break; 7666 } 7667 } 7668 // Skip verifying header mark word below because we are 7669 // running concurrent with mutators. 7670 assert(new_oop->is_oop(true), "Oops! expected to pop an oop"); 7671 // now scan this oop's oops 7672 new_oop->oop_iterate(&pushOrMarkClosure); 7673 do_yield_check(); 7674 } 7675 assert(_work_queue->size() == 0, "tautology, emphasizing post-condition"); 7676 } 7677 7678 // Yield in response to a request from VM Thread or 7679 // from mutators. 7680 void Par_MarkFromRootsClosure::do_yield_work() { 7681 assert(_task != NULL, "sanity"); 7682 _task->yield(); 7683 } 7684 7685 // A variant of the above used for verifying CMS marking work. 7686 MarkFromRootsVerifyClosure::MarkFromRootsVerifyClosure(CMSCollector* collector, 7687 MemRegion span, 7688 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 7689 CMSMarkStack* mark_stack): 7690 _collector(collector), 7691 _span(span), 7692 _verification_bm(verification_bm), 7693 _cms_bm(cms_bm), 7694 _mark_stack(mark_stack), 7695 _pam_verify_closure(collector, span, verification_bm, cms_bm, 7696 mark_stack) 7697 { 7698 assert(_mark_stack->isEmpty(), "stack should be empty"); 7699 _finger = _verification_bm->startWord(); 7700 assert(_collector->_restart_addr == NULL, "Sanity check"); 7701 assert(_span.contains(_finger), "Out of bounds _finger?"); 7702 } 7703 7704 void MarkFromRootsVerifyClosure::reset(HeapWord* addr) { 7705 assert(_mark_stack->isEmpty(), "would cause duplicates on stack"); 7706 assert(_span.contains(addr), "Out of bounds _finger?"); 7707 _finger = addr; 7708 } 7709 7710 // Should revisit to see if this should be restructured for 7711 // greater efficiency. 7712 bool MarkFromRootsVerifyClosure::do_bit(size_t offset) { 7713 // convert offset into a HeapWord* 7714 HeapWord* addr = _verification_bm->startWord() + offset; 7715 assert(_verification_bm->endWord() && addr < _verification_bm->endWord(), 7716 "address out of range"); 7717 assert(_verification_bm->isMarked(addr), "tautology"); 7718 assert(_cms_bm->isMarked(addr), "tautology"); 7719 7720 assert(_mark_stack->isEmpty(), 7721 "should drain stack to limit stack usage"); 7722 // convert addr to an oop preparatory to scanning 7723 oop obj = oop(addr); 7724 assert(obj->is_oop(), "should be an oop"); 7725 assert(_finger <= addr, "_finger runneth ahead"); 7726 // advance the finger to right end of this object 7727 _finger = addr + obj->size(); 7728 assert(_finger > addr, "we just incremented it above"); 7729 // Note: the finger doesn't advance while we drain 7730 // the stack below. 7731 bool res = _mark_stack->push(obj); 7732 assert(res, "Empty non-zero size stack should have space for single push"); 7733 while (!_mark_stack->isEmpty()) { 7734 oop new_oop = _mark_stack->pop(); 7735 assert(new_oop->is_oop(), "Oops! expected to pop an oop"); 7736 // now scan this oop's oops 7737 new_oop->oop_iterate(&_pam_verify_closure); 7738 } 7739 assert(_mark_stack->isEmpty(), "tautology, emphasizing post-condition"); 7740 return true; 7741 } 7742 7743 PushAndMarkVerifyClosure::PushAndMarkVerifyClosure( 7744 CMSCollector* collector, MemRegion span, 7745 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 7746 CMSMarkStack* mark_stack): 7747 CMSOopClosure(collector->ref_processor()), 7748 _collector(collector), 7749 _span(span), 7750 _verification_bm(verification_bm), 7751 _cms_bm(cms_bm), 7752 _mark_stack(mark_stack) 7753 { } 7754 7755 void PushAndMarkVerifyClosure::do_oop(oop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 7756 void PushAndMarkVerifyClosure::do_oop(narrowOop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 7757 7758 // Upon stack overflow, we discard (part of) the stack, 7759 // remembering the least address amongst those discarded 7760 // in CMSCollector's _restart_address. 7761 void PushAndMarkVerifyClosure::handle_stack_overflow(HeapWord* lost) { 7762 // Remember the least grey address discarded 7763 HeapWord* ra = (HeapWord*)_mark_stack->least_value(lost); 7764 _collector->lower_restart_addr(ra); 7765 _mark_stack->reset(); // discard stack contents 7766 _mark_stack->expand(); // expand the stack if possible 7767 } 7768 7769 void PushAndMarkVerifyClosure::do_oop(oop obj) { 7770 assert(obj->is_oop_or_null(), "expected an oop or NULL"); 7771 HeapWord* addr = (HeapWord*)obj; 7772 if (_span.contains(addr) && !_verification_bm->isMarked(addr)) { 7773 // Oop lies in _span and isn't yet grey or black 7774 _verification_bm->mark(addr); // now grey 7775 if (!_cms_bm->isMarked(addr)) { 7776 oop(addr)->print(); 7777 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)", 7778 addr); 7779 fatal("... aborting"); 7780 } 7781 7782 if (!_mark_stack->push(obj)) { // stack overflow 7783 if (PrintCMSStatistics != 0) { 7784 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 7785 SIZE_FORMAT, _mark_stack->capacity()); 7786 } 7787 assert(_mark_stack->isFull(), "Else push should have succeeded"); 7788 handle_stack_overflow(addr); 7789 } 7790 // anything including and to the right of _finger 7791 // will be scanned as we iterate over the remainder of the 7792 // bit map 7793 } 7794 } 7795 7796 PushOrMarkClosure::PushOrMarkClosure(CMSCollector* collector, 7797 MemRegion span, 7798 CMSBitMap* bitMap, CMSMarkStack* markStack, 7799 HeapWord* finger, MarkFromRootsClosure* parent) : 7800 CMSOopClosure(collector->ref_processor()), 7801 _collector(collector), 7802 _span(span), 7803 _bitMap(bitMap), 7804 _markStack(markStack), 7805 _finger(finger), 7806 _parent(parent) 7807 { } 7808 7809 Par_PushOrMarkClosure::Par_PushOrMarkClosure(CMSCollector* collector, 7810 MemRegion span, 7811 CMSBitMap* bit_map, 7812 OopTaskQueue* work_queue, 7813 CMSMarkStack* overflow_stack, 7814 HeapWord* finger, 7815 HeapWord** global_finger_addr, 7816 Par_MarkFromRootsClosure* parent) : 7817 CMSOopClosure(collector->ref_processor()), 7818 _collector(collector), 7819 _whole_span(collector->_span), 7820 _span(span), 7821 _bit_map(bit_map), 7822 _work_queue(work_queue), 7823 _overflow_stack(overflow_stack), 7824 _finger(finger), 7825 _global_finger_addr(global_finger_addr), 7826 _parent(parent) 7827 { } 7828 7829 // Assumes thread-safe access by callers, who are 7830 // responsible for mutual exclusion. 7831 void CMSCollector::lower_restart_addr(HeapWord* low) { 7832 assert(_span.contains(low), "Out of bounds addr"); 7833 if (_restart_addr == NULL) { 7834 _restart_addr = low; 7835 } else { 7836 _restart_addr = MIN2(_restart_addr, low); 7837 } 7838 } 7839 7840 // Upon stack overflow, we discard (part of) the stack, 7841 // remembering the least address amongst those discarded 7842 // in CMSCollector's _restart_address. 7843 void PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 7844 // Remember the least grey address discarded 7845 HeapWord* ra = (HeapWord*)_markStack->least_value(lost); 7846 _collector->lower_restart_addr(ra); 7847 _markStack->reset(); // discard stack contents 7848 _markStack->expand(); // expand the stack if possible 7849 } 7850 7851 // Upon stack overflow, we discard (part of) the stack, 7852 // remembering the least address amongst those discarded 7853 // in CMSCollector's _restart_address. 7854 void Par_PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 7855 // We need to do this under a mutex to prevent other 7856 // workers from interfering with the work done below. 7857 MutexLockerEx ml(_overflow_stack->par_lock(), 7858 Mutex::_no_safepoint_check_flag); 7859 // Remember the least grey address discarded 7860 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 7861 _collector->lower_restart_addr(ra); 7862 _overflow_stack->reset(); // discard stack contents 7863 _overflow_stack->expand(); // expand the stack if possible 7864 } 7865 7866 void CMKlassClosure::do_klass(Klass* k) { 7867 assert(_oop_closure != NULL, "Not initialized?"); 7868 k->oops_do(_oop_closure); 7869 } 7870 7871 void PushOrMarkClosure::do_oop(oop obj) { 7872 // Ignore mark word because we are running concurrent with mutators. 7873 assert(obj->is_oop_or_null(true), "expected an oop or NULL"); 7874 HeapWord* addr = (HeapWord*)obj; 7875 if (_span.contains(addr) && !_bitMap->isMarked(addr)) { 7876 // Oop lies in _span and isn't yet grey or black 7877 _bitMap->mark(addr); // now grey 7878 if (addr < _finger) { 7879 // the bit map iteration has already either passed, or 7880 // sampled, this bit in the bit map; we'll need to 7881 // use the marking stack to scan this oop's oops. 7882 bool simulate_overflow = false; 7883 NOT_PRODUCT( 7884 if (CMSMarkStackOverflowALot && 7885 _collector->simulate_overflow()) { 7886 // simulate a stack overflow 7887 simulate_overflow = true; 7888 } 7889 ) 7890 if (simulate_overflow || !_markStack->push(obj)) { // stack overflow 7891 if (PrintCMSStatistics != 0) { 7892 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 7893 SIZE_FORMAT, _markStack->capacity()); 7894 } 7895 assert(simulate_overflow || _markStack->isFull(), "Else push should have succeeded"); 7896 handle_stack_overflow(addr); 7897 } 7898 } 7899 // anything including and to the right of _finger 7900 // will be scanned as we iterate over the remainder of the 7901 // bit map 7902 do_yield_check(); 7903 } 7904 } 7905 7906 void PushOrMarkClosure::do_oop(oop* p) { PushOrMarkClosure::do_oop_work(p); } 7907 void PushOrMarkClosure::do_oop(narrowOop* p) { PushOrMarkClosure::do_oop_work(p); } 7908 7909 void Par_PushOrMarkClosure::do_oop(oop obj) { 7910 // Ignore mark word because we are running concurrent with mutators. 7911 assert(obj->is_oop_or_null(true), "expected an oop or NULL"); 7912 HeapWord* addr = (HeapWord*)obj; 7913 if (_whole_span.contains(addr) && !_bit_map->isMarked(addr)) { 7914 // Oop lies in _span and isn't yet grey or black 7915 // We read the global_finger (volatile read) strictly after marking oop 7916 bool res = _bit_map->par_mark(addr); // now grey 7917 volatile HeapWord** gfa = (volatile HeapWord**)_global_finger_addr; 7918 // Should we push this marked oop on our stack? 7919 // -- if someone else marked it, nothing to do 7920 // -- if target oop is above global finger nothing to do 7921 // -- if target oop is in chunk and above local finger 7922 // then nothing to do 7923 // -- else push on work queue 7924 if ( !res // someone else marked it, they will deal with it 7925 || (addr >= *gfa) // will be scanned in a later task 7926 || (_span.contains(addr) && addr >= _finger)) { // later in this chunk 7927 return; 7928 } 7929 // the bit map iteration has already either passed, or 7930 // sampled, this bit in the bit map; we'll need to 7931 // use the marking stack to scan this oop's oops. 7932 bool simulate_overflow = false; 7933 NOT_PRODUCT( 7934 if (CMSMarkStackOverflowALot && 7935 _collector->simulate_overflow()) { 7936 // simulate a stack overflow 7937 simulate_overflow = true; 7938 } 7939 ) 7940 if (simulate_overflow || 7941 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) { 7942 // stack overflow 7943 if (PrintCMSStatistics != 0) { 7944 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 7945 SIZE_FORMAT, _overflow_stack->capacity()); 7946 } 7947 // We cannot assert that the overflow stack is full because 7948 // it may have been emptied since. 7949 assert(simulate_overflow || 7950 _work_queue->size() == _work_queue->max_elems(), 7951 "Else push should have succeeded"); 7952 handle_stack_overflow(addr); 7953 } 7954 do_yield_check(); 7955 } 7956 } 7957 7958 void Par_PushOrMarkClosure::do_oop(oop* p) { Par_PushOrMarkClosure::do_oop_work(p); } 7959 void Par_PushOrMarkClosure::do_oop(narrowOop* p) { Par_PushOrMarkClosure::do_oop_work(p); } 7960 7961 PushAndMarkClosure::PushAndMarkClosure(CMSCollector* collector, 7962 MemRegion span, 7963 ReferenceProcessor* rp, 7964 CMSBitMap* bit_map, 7965 CMSBitMap* mod_union_table, 7966 CMSMarkStack* mark_stack, 7967 bool concurrent_precleaning): 7968 CMSOopClosure(rp), 7969 _collector(collector), 7970 _span(span), 7971 _bit_map(bit_map), 7972 _mod_union_table(mod_union_table), 7973 _mark_stack(mark_stack), 7974 _concurrent_precleaning(concurrent_precleaning) 7975 { 7976 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 7977 } 7978 7979 // Grey object rescan during pre-cleaning and second checkpoint phases -- 7980 // the non-parallel version (the parallel version appears further below.) 7981 void PushAndMarkClosure::do_oop(oop obj) { 7982 // Ignore mark word verification. If during concurrent precleaning, 7983 // the object monitor may be locked. If during the checkpoint 7984 // phases, the object may already have been reached by a different 7985 // path and may be at the end of the global overflow list (so 7986 // the mark word may be NULL). 7987 assert(obj->is_oop_or_null(true /* ignore mark word */), 7988 "expected an oop or NULL"); 7989 HeapWord* addr = (HeapWord*)obj; 7990 // Check if oop points into the CMS generation 7991 // and is not marked 7992 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 7993 // a white object ... 7994 _bit_map->mark(addr); // ... now grey 7995 // push on the marking stack (grey set) 7996 bool simulate_overflow = false; 7997 NOT_PRODUCT( 7998 if (CMSMarkStackOverflowALot && 7999 _collector->simulate_overflow()) { 8000 // simulate a stack overflow 8001 simulate_overflow = true; 8002 } 8003 ) 8004 if (simulate_overflow || !_mark_stack->push(obj)) { 8005 if (_concurrent_precleaning) { 8006 // During precleaning we can just dirty the appropriate card(s) 8007 // in the mod union table, thus ensuring that the object remains 8008 // in the grey set and continue. In the case of object arrays 8009 // we need to dirty all of the cards that the object spans, 8010 // since the rescan of object arrays will be limited to the 8011 // dirty cards. 8012 // Note that no one can be intefering with us in this action 8013 // of dirtying the mod union table, so no locking or atomics 8014 // are required. 8015 if (obj->is_objArray()) { 8016 size_t sz = obj->size(); 8017 HeapWord* end_card_addr = (HeapWord*)round_to( 8018 (intptr_t)(addr+sz), CardTableModRefBS::card_size); 8019 MemRegion redirty_range = MemRegion(addr, end_card_addr); 8020 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 8021 _mod_union_table->mark_range(redirty_range); 8022 } else { 8023 _mod_union_table->mark(addr); 8024 } 8025 _collector->_ser_pmc_preclean_ovflw++; 8026 } else { 8027 // During the remark phase, we need to remember this oop 8028 // in the overflow list. 8029 _collector->push_on_overflow_list(obj); 8030 _collector->_ser_pmc_remark_ovflw++; 8031 } 8032 } 8033 } 8034 } 8035 8036 Par_PushAndMarkClosure::Par_PushAndMarkClosure(CMSCollector* collector, 8037 MemRegion span, 8038 ReferenceProcessor* rp, 8039 CMSBitMap* bit_map, 8040 OopTaskQueue* work_queue): 8041 CMSOopClosure(rp), 8042 _collector(collector), 8043 _span(span), 8044 _bit_map(bit_map), 8045 _work_queue(work_queue) 8046 { 8047 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 8048 } 8049 8050 void PushAndMarkClosure::do_oop(oop* p) { PushAndMarkClosure::do_oop_work(p); } 8051 void PushAndMarkClosure::do_oop(narrowOop* p) { PushAndMarkClosure::do_oop_work(p); } 8052 8053 // Grey object rescan during second checkpoint phase -- 8054 // the parallel version. 8055 void Par_PushAndMarkClosure::do_oop(oop obj) { 8056 // In the assert below, we ignore the mark word because 8057 // this oop may point to an already visited object that is 8058 // on the overflow stack (in which case the mark word has 8059 // been hijacked for chaining into the overflow stack -- 8060 // if this is the last object in the overflow stack then 8061 // its mark word will be NULL). Because this object may 8062 // have been subsequently popped off the global overflow 8063 // stack, and the mark word possibly restored to the prototypical 8064 // value, by the time we get to examined this failing assert in 8065 // the debugger, is_oop_or_null(false) may subsequently start 8066 // to hold. 8067 assert(obj->is_oop_or_null(true), 8068 "expected an oop or NULL"); 8069 HeapWord* addr = (HeapWord*)obj; 8070 // Check if oop points into the CMS generation 8071 // and is not marked 8072 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 8073 // a white object ... 8074 // If we manage to "claim" the object, by being the 8075 // first thread to mark it, then we push it on our 8076 // marking stack 8077 if (_bit_map->par_mark(addr)) { // ... now grey 8078 // push on work queue (grey set) 8079 bool simulate_overflow = false; 8080 NOT_PRODUCT( 8081 if (CMSMarkStackOverflowALot && 8082 _collector->par_simulate_overflow()) { 8083 // simulate a stack overflow 8084 simulate_overflow = true; 8085 } 8086 ) 8087 if (simulate_overflow || !_work_queue->push(obj)) { 8088 _collector->par_push_on_overflow_list(obj); 8089 _collector->_par_pmc_remark_ovflw++; // imprecise OK: no need to CAS 8090 } 8091 } // Else, some other thread got there first 8092 } 8093 } 8094 8095 void Par_PushAndMarkClosure::do_oop(oop* p) { Par_PushAndMarkClosure::do_oop_work(p); } 8096 void Par_PushAndMarkClosure::do_oop(narrowOop* p) { Par_PushAndMarkClosure::do_oop_work(p); } 8097 8098 void CMSPrecleanRefsYieldClosure::do_yield_work() { 8099 Mutex* bml = _collector->bitMapLock(); 8100 assert_lock_strong(bml); 8101 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 8102 "CMS thread should hold CMS token"); 8103 8104 bml->unlock(); 8105 ConcurrentMarkSweepThread::desynchronize(true); 8106 8107 ConcurrentMarkSweepThread::acknowledge_yield_request(); 8108 8109 _collector->stopTimer(); 8110 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 8111 if (PrintCMSStatistics != 0) { 8112 _collector->incrementYields(); 8113 } 8114 _collector->icms_wait(); 8115 8116 // See the comment in coordinator_yield() 8117 for (unsigned i = 0; i < CMSYieldSleepCount && 8118 ConcurrentMarkSweepThread::should_yield() && 8119 !CMSCollector::foregroundGCIsActive(); ++i) { 8120 os::sleep(Thread::current(), 1, false); 8121 ConcurrentMarkSweepThread::acknowledge_yield_request(); 8122 } 8123 8124 ConcurrentMarkSweepThread::synchronize(true); 8125 bml->lock(); 8126 8127 _collector->startTimer(); 8128 } 8129 8130 bool CMSPrecleanRefsYieldClosure::should_return() { 8131 if (ConcurrentMarkSweepThread::should_yield()) { 8132 do_yield_work(); 8133 } 8134 return _collector->foregroundGCIsActive(); 8135 } 8136 8137 void MarkFromDirtyCardsClosure::do_MemRegion(MemRegion mr) { 8138 assert(((size_t)mr.start())%CardTableModRefBS::card_size_in_words == 0, 8139 "mr should be aligned to start at a card boundary"); 8140 // We'd like to assert: 8141 // assert(mr.word_size()%CardTableModRefBS::card_size_in_words == 0, 8142 // "mr should be a range of cards"); 8143 // However, that would be too strong in one case -- the last 8144 // partition ends at _unallocated_block which, in general, can be 8145 // an arbitrary boundary, not necessarily card aligned. 8146 if (PrintCMSStatistics != 0) { 8147 _num_dirty_cards += 8148 mr.word_size()/CardTableModRefBS::card_size_in_words; 8149 } 8150 _space->object_iterate_mem(mr, &_scan_cl); 8151 } 8152 8153 SweepClosure::SweepClosure(CMSCollector* collector, 8154 ConcurrentMarkSweepGeneration* g, 8155 CMSBitMap* bitMap, bool should_yield) : 8156 _collector(collector), 8157 _g(g), 8158 _sp(g->cmsSpace()), 8159 _limit(_sp->sweep_limit()), 8160 _freelistLock(_sp->freelistLock()), 8161 _bitMap(bitMap), 8162 _yield(should_yield), 8163 _inFreeRange(false), // No free range at beginning of sweep 8164 _freeRangeInFreeLists(false), // No free range at beginning of sweep 8165 _lastFreeRangeCoalesced(false), 8166 _freeFinger(g->used_region().start()) 8167 { 8168 NOT_PRODUCT( 8169 _numObjectsFreed = 0; 8170 _numWordsFreed = 0; 8171 _numObjectsLive = 0; 8172 _numWordsLive = 0; 8173 _numObjectsAlreadyFree = 0; 8174 _numWordsAlreadyFree = 0; 8175 _last_fc = NULL; 8176 8177 _sp->initializeIndexedFreeListArrayReturnedBytes(); 8178 _sp->dictionary()->initialize_dict_returned_bytes(); 8179 ) 8180 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 8181 "sweep _limit out of bounds"); 8182 if (CMSTraceSweeper) { 8183 gclog_or_tty->print_cr("\n====================\nStarting new sweep with limit " PTR_FORMAT, 8184 _limit); 8185 } 8186 } 8187 8188 void SweepClosure::print_on(outputStream* st) const { 8189 tty->print_cr("_sp = [" PTR_FORMAT "," PTR_FORMAT ")", 8190 _sp->bottom(), _sp->end()); 8191 tty->print_cr("_limit = " PTR_FORMAT, _limit); 8192 tty->print_cr("_freeFinger = " PTR_FORMAT, _freeFinger); 8193 NOT_PRODUCT(tty->print_cr("_last_fc = " PTR_FORMAT, _last_fc);) 8194 tty->print_cr("_inFreeRange = %d, _freeRangeInFreeLists = %d, _lastFreeRangeCoalesced = %d", 8195 _inFreeRange, _freeRangeInFreeLists, _lastFreeRangeCoalesced); 8196 } 8197 8198 #ifndef PRODUCT 8199 // Assertion checking only: no useful work in product mode -- 8200 // however, if any of the flags below become product flags, 8201 // you may need to review this code to see if it needs to be 8202 // enabled in product mode. 8203 SweepClosure::~SweepClosure() { 8204 assert_lock_strong(_freelistLock); 8205 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 8206 "sweep _limit out of bounds"); 8207 if (inFreeRange()) { 8208 warning("inFreeRange() should have been reset; dumping state of SweepClosure"); 8209 print(); 8210 ShouldNotReachHere(); 8211 } 8212 if (Verbose && PrintGC) { 8213 gclog_or_tty->print("Collected "SIZE_FORMAT" objects, " SIZE_FORMAT " bytes", 8214 _numObjectsFreed, _numWordsFreed*sizeof(HeapWord)); 8215 gclog_or_tty->print_cr("\nLive "SIZE_FORMAT" objects, " 8216 SIZE_FORMAT" bytes " 8217 "Already free "SIZE_FORMAT" objects, "SIZE_FORMAT" bytes", 8218 _numObjectsLive, _numWordsLive*sizeof(HeapWord), 8219 _numObjectsAlreadyFree, _numWordsAlreadyFree*sizeof(HeapWord)); 8220 size_t totalBytes = (_numWordsFreed + _numWordsLive + _numWordsAlreadyFree) 8221 * sizeof(HeapWord); 8222 gclog_or_tty->print_cr("Total sweep: "SIZE_FORMAT" bytes", totalBytes); 8223 8224 if (PrintCMSStatistics && CMSVerifyReturnedBytes) { 8225 size_t indexListReturnedBytes = _sp->sumIndexedFreeListArrayReturnedBytes(); 8226 size_t dict_returned_bytes = _sp->dictionary()->sum_dict_returned_bytes(); 8227 size_t returned_bytes = indexListReturnedBytes + dict_returned_bytes; 8228 gclog_or_tty->print("Returned "SIZE_FORMAT" bytes", returned_bytes); 8229 gclog_or_tty->print(" Indexed List Returned "SIZE_FORMAT" bytes", 8230 indexListReturnedBytes); 8231 gclog_or_tty->print_cr(" Dictionary Returned "SIZE_FORMAT" bytes", 8232 dict_returned_bytes); 8233 } 8234 } 8235 if (CMSTraceSweeper) { 8236 gclog_or_tty->print_cr("end of sweep with _limit = " PTR_FORMAT "\n================", 8237 _limit); 8238 } 8239 } 8240 #endif // PRODUCT 8241 8242 void SweepClosure::initialize_free_range(HeapWord* freeFinger, 8243 bool freeRangeInFreeLists) { 8244 if (CMSTraceSweeper) { 8245 gclog_or_tty->print("---- Start free range at 0x%x with free block (%d)\n", 8246 freeFinger, freeRangeInFreeLists); 8247 } 8248 assert(!inFreeRange(), "Trampling existing free range"); 8249 set_inFreeRange(true); 8250 set_lastFreeRangeCoalesced(false); 8251 8252 set_freeFinger(freeFinger); 8253 set_freeRangeInFreeLists(freeRangeInFreeLists); 8254 if (CMSTestInFreeList) { 8255 if (freeRangeInFreeLists) { 8256 FreeChunk* fc = (FreeChunk*) freeFinger; 8257 assert(fc->is_free(), "A chunk on the free list should be free."); 8258 assert(fc->size() > 0, "Free range should have a size"); 8259 assert(_sp->verify_chunk_in_free_list(fc), "Chunk is not in free lists"); 8260 } 8261 } 8262 } 8263 8264 // Note that the sweeper runs concurrently with mutators. Thus, 8265 // it is possible for direct allocation in this generation to happen 8266 // in the middle of the sweep. Note that the sweeper also coalesces 8267 // contiguous free blocks. Thus, unless the sweeper and the allocator 8268 // synchronize appropriately freshly allocated blocks may get swept up. 8269 // This is accomplished by the sweeper locking the free lists while 8270 // it is sweeping. Thus blocks that are determined to be free are 8271 // indeed free. There is however one additional complication: 8272 // blocks that have been allocated since the final checkpoint and 8273 // mark, will not have been marked and so would be treated as 8274 // unreachable and swept up. To prevent this, the allocator marks 8275 // the bit map when allocating during the sweep phase. This leads, 8276 // however, to a further complication -- objects may have been allocated 8277 // but not yet initialized -- in the sense that the header isn't yet 8278 // installed. The sweeper can not then determine the size of the block 8279 // in order to skip over it. To deal with this case, we use a technique 8280 // (due to Printezis) to encode such uninitialized block sizes in the 8281 // bit map. Since the bit map uses a bit per every HeapWord, but the 8282 // CMS generation has a minimum object size of 3 HeapWords, it follows 8283 // that "normal marks" won't be adjacent in the bit map (there will 8284 // always be at least two 0 bits between successive 1 bits). We make use 8285 // of these "unused" bits to represent uninitialized blocks -- the bit 8286 // corresponding to the start of the uninitialized object and the next 8287 // bit are both set. Finally, a 1 bit marks the end of the object that 8288 // started with the two consecutive 1 bits to indicate its potentially 8289 // uninitialized state. 8290 8291 size_t SweepClosure::do_blk_careful(HeapWord* addr) { 8292 FreeChunk* fc = (FreeChunk*)addr; 8293 size_t res; 8294 8295 // Check if we are done sweeping. Below we check "addr >= _limit" rather 8296 // than "addr == _limit" because although _limit was a block boundary when 8297 // we started the sweep, it may no longer be one because heap expansion 8298 // may have caused us to coalesce the block ending at the address _limit 8299 // with a newly expanded chunk (this happens when _limit was set to the 8300 // previous _end of the space), so we may have stepped past _limit: 8301 // see the following Zeno-like trail of CRs 6977970, 7008136, 7042740. 8302 if (addr >= _limit) { // we have swept up to or past the limit: finish up 8303 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 8304 "sweep _limit out of bounds"); 8305 assert(addr < _sp->end(), "addr out of bounds"); 8306 // Flush any free range we might be holding as a single 8307 // coalesced chunk to the appropriate free list. 8308 if (inFreeRange()) { 8309 assert(freeFinger() >= _sp->bottom() && freeFinger() < _limit, 8310 err_msg("freeFinger() " PTR_FORMAT" is out-of-bounds", freeFinger())); 8311 flush_cur_free_chunk(freeFinger(), 8312 pointer_delta(addr, freeFinger())); 8313 if (CMSTraceSweeper) { 8314 gclog_or_tty->print("Sweep: last chunk: "); 8315 gclog_or_tty->print("put_free_blk 0x%x ("SIZE_FORMAT") " 8316 "[coalesced:"SIZE_FORMAT"]\n", 8317 freeFinger(), pointer_delta(addr, freeFinger()), 8318 lastFreeRangeCoalesced()); 8319 } 8320 } 8321 8322 // help the iterator loop finish 8323 return pointer_delta(_sp->end(), addr); 8324 } 8325 8326 assert(addr < _limit, "sweep invariant"); 8327 // check if we should yield 8328 do_yield_check(addr); 8329 if (fc->is_free()) { 8330 // Chunk that is already free 8331 res = fc->size(); 8332 do_already_free_chunk(fc); 8333 debug_only(_sp->verifyFreeLists()); 8334 // If we flush the chunk at hand in lookahead_and_flush() 8335 // and it's coalesced with a preceding chunk, then the 8336 // process of "mangling" the payload of the coalesced block 8337 // will cause erasure of the size information from the 8338 // (erstwhile) header of all the coalesced blocks but the 8339 // first, so the first disjunct in the assert will not hold 8340 // in that specific case (in which case the second disjunct 8341 // will hold). 8342 assert(res == fc->size() || ((HeapWord*)fc) + res >= _limit, 8343 "Otherwise the size info doesn't change at this step"); 8344 NOT_PRODUCT( 8345 _numObjectsAlreadyFree++; 8346 _numWordsAlreadyFree += res; 8347 ) 8348 NOT_PRODUCT(_last_fc = fc;) 8349 } else if (!_bitMap->isMarked(addr)) { 8350 // Chunk is fresh garbage 8351 res = do_garbage_chunk(fc); 8352 debug_only(_sp->verifyFreeLists()); 8353 NOT_PRODUCT( 8354 _numObjectsFreed++; 8355 _numWordsFreed += res; 8356 ) 8357 } else { 8358 // Chunk that is alive. 8359 res = do_live_chunk(fc); 8360 debug_only(_sp->verifyFreeLists()); 8361 NOT_PRODUCT( 8362 _numObjectsLive++; 8363 _numWordsLive += res; 8364 ) 8365 } 8366 return res; 8367 } 8368 8369 // For the smart allocation, record following 8370 // split deaths - a free chunk is removed from its free list because 8371 // it is being split into two or more chunks. 8372 // split birth - a free chunk is being added to its free list because 8373 // a larger free chunk has been split and resulted in this free chunk. 8374 // coal death - a free chunk is being removed from its free list because 8375 // it is being coalesced into a large free chunk. 8376 // coal birth - a free chunk is being added to its free list because 8377 // it was created when two or more free chunks where coalesced into 8378 // this free chunk. 8379 // 8380 // These statistics are used to determine the desired number of free 8381 // chunks of a given size. The desired number is chosen to be relative 8382 // to the end of a CMS sweep. The desired number at the end of a sweep 8383 // is the 8384 // count-at-end-of-previous-sweep (an amount that was enough) 8385 // - count-at-beginning-of-current-sweep (the excess) 8386 // + split-births (gains in this size during interval) 8387 // - split-deaths (demands on this size during interval) 8388 // where the interval is from the end of one sweep to the end of the 8389 // next. 8390 // 8391 // When sweeping the sweeper maintains an accumulated chunk which is 8392 // the chunk that is made up of chunks that have been coalesced. That 8393 // will be termed the left-hand chunk. A new chunk of garbage that 8394 // is being considered for coalescing will be referred to as the 8395 // right-hand chunk. 8396 // 8397 // When making a decision on whether to coalesce a right-hand chunk with 8398 // the current left-hand chunk, the current count vs. the desired count 8399 // of the left-hand chunk is considered. Also if the right-hand chunk 8400 // is near the large chunk at the end of the heap (see 8401 // ConcurrentMarkSweepGeneration::isNearLargestChunk()), then the 8402 // left-hand chunk is coalesced. 8403 // 8404 // When making a decision about whether to split a chunk, the desired count 8405 // vs. the current count of the candidate to be split is also considered. 8406 // If the candidate is underpopulated (currently fewer chunks than desired) 8407 // a chunk of an overpopulated (currently more chunks than desired) size may 8408 // be chosen. The "hint" associated with a free list, if non-null, points 8409 // to a free list which may be overpopulated. 8410 // 8411 8412 void SweepClosure::do_already_free_chunk(FreeChunk* fc) { 8413 const size_t size = fc->size(); 8414 // Chunks that cannot be coalesced are not in the 8415 // free lists. 8416 if (CMSTestInFreeList && !fc->cantCoalesce()) { 8417 assert(_sp->verify_chunk_in_free_list(fc), 8418 "free chunk should be in free lists"); 8419 } 8420 // a chunk that is already free, should not have been 8421 // marked in the bit map 8422 HeapWord* const addr = (HeapWord*) fc; 8423 assert(!_bitMap->isMarked(addr), "free chunk should be unmarked"); 8424 // Verify that the bit map has no bits marked between 8425 // addr and purported end of this block. 8426 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 8427 8428 // Some chunks cannot be coalesced under any circumstances. 8429 // See the definition of cantCoalesce(). 8430 if (!fc->cantCoalesce()) { 8431 // This chunk can potentially be coalesced. 8432 if (_sp->adaptive_freelists()) { 8433 // All the work is done in 8434 do_post_free_or_garbage_chunk(fc, size); 8435 } else { // Not adaptive free lists 8436 // this is a free chunk that can potentially be coalesced by the sweeper; 8437 if (!inFreeRange()) { 8438 // if the next chunk is a free block that can't be coalesced 8439 // it doesn't make sense to remove this chunk from the free lists 8440 FreeChunk* nextChunk = (FreeChunk*)(addr + size); 8441 assert((HeapWord*)nextChunk <= _sp->end(), "Chunk size out of bounds?"); 8442 if ((HeapWord*)nextChunk < _sp->end() && // There is another free chunk to the right ... 8443 nextChunk->is_free() && // ... which is free... 8444 nextChunk->cantCoalesce()) { // ... but can't be coalesced 8445 // nothing to do 8446 } else { 8447 // Potentially the start of a new free range: 8448 // Don't eagerly remove it from the free lists. 8449 // No need to remove it if it will just be put 8450 // back again. (Also from a pragmatic point of view 8451 // if it is a free block in a region that is beyond 8452 // any allocated blocks, an assertion will fail) 8453 // Remember the start of a free run. 8454 initialize_free_range(addr, true); 8455 // end - can coalesce with next chunk 8456 } 8457 } else { 8458 // the midst of a free range, we are coalescing 8459 print_free_block_coalesced(fc); 8460 if (CMSTraceSweeper) { 8461 gclog_or_tty->print(" -- pick up free block 0x%x (%d)\n", fc, size); 8462 } 8463 // remove it from the free lists 8464 _sp->removeFreeChunkFromFreeLists(fc); 8465 set_lastFreeRangeCoalesced(true); 8466 // If the chunk is being coalesced and the current free range is 8467 // in the free lists, remove the current free range so that it 8468 // will be returned to the free lists in its entirety - all 8469 // the coalesced pieces included. 8470 if (freeRangeInFreeLists()) { 8471 FreeChunk* ffc = (FreeChunk*) freeFinger(); 8472 assert(ffc->size() == pointer_delta(addr, freeFinger()), 8473 "Size of free range is inconsistent with chunk size."); 8474 if (CMSTestInFreeList) { 8475 assert(_sp->verify_chunk_in_free_list(ffc), 8476 "free range is not in free lists"); 8477 } 8478 _sp->removeFreeChunkFromFreeLists(ffc); 8479 set_freeRangeInFreeLists(false); 8480 } 8481 } 8482 } 8483 // Note that if the chunk is not coalescable (the else arm 8484 // below), we unconditionally flush, without needing to do 8485 // a "lookahead," as we do below. 8486 if (inFreeRange()) lookahead_and_flush(fc, size); 8487 } else { 8488 // Code path common to both original and adaptive free lists. 8489 8490 // cant coalesce with previous block; this should be treated 8491 // as the end of a free run if any 8492 if (inFreeRange()) { 8493 // we kicked some butt; time to pick up the garbage 8494 assert(freeFinger() < addr, "freeFinger points too high"); 8495 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 8496 } 8497 // else, nothing to do, just continue 8498 } 8499 } 8500 8501 size_t SweepClosure::do_garbage_chunk(FreeChunk* fc) { 8502 // This is a chunk of garbage. It is not in any free list. 8503 // Add it to a free list or let it possibly be coalesced into 8504 // a larger chunk. 8505 HeapWord* const addr = (HeapWord*) fc; 8506 const size_t size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 8507 8508 if (_sp->adaptive_freelists()) { 8509 // Verify that the bit map has no bits marked between 8510 // addr and purported end of just dead object. 8511 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 8512 8513 do_post_free_or_garbage_chunk(fc, size); 8514 } else { 8515 if (!inFreeRange()) { 8516 // start of a new free range 8517 assert(size > 0, "A free range should have a size"); 8518 initialize_free_range(addr, false); 8519 } else { 8520 // this will be swept up when we hit the end of the 8521 // free range 8522 if (CMSTraceSweeper) { 8523 gclog_or_tty->print(" -- pick up garbage 0x%x (%d) \n", fc, size); 8524 } 8525 // If the chunk is being coalesced and the current free range is 8526 // in the free lists, remove the current free range so that it 8527 // will be returned to the free lists in its entirety - all 8528 // the coalesced pieces included. 8529 if (freeRangeInFreeLists()) { 8530 FreeChunk* ffc = (FreeChunk*)freeFinger(); 8531 assert(ffc->size() == pointer_delta(addr, freeFinger()), 8532 "Size of free range is inconsistent with chunk size."); 8533 if (CMSTestInFreeList) { 8534 assert(_sp->verify_chunk_in_free_list(ffc), 8535 "free range is not in free lists"); 8536 } 8537 _sp->removeFreeChunkFromFreeLists(ffc); 8538 set_freeRangeInFreeLists(false); 8539 } 8540 set_lastFreeRangeCoalesced(true); 8541 } 8542 // this will be swept up when we hit the end of the free range 8543 8544 // Verify that the bit map has no bits marked between 8545 // addr and purported end of just dead object. 8546 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 8547 } 8548 assert(_limit >= addr + size, 8549 "A freshly garbage chunk can't possibly straddle over _limit"); 8550 if (inFreeRange()) lookahead_and_flush(fc, size); 8551 return size; 8552 } 8553 8554 size_t SweepClosure::do_live_chunk(FreeChunk* fc) { 8555 HeapWord* addr = (HeapWord*) fc; 8556 // The sweeper has just found a live object. Return any accumulated 8557 // left hand chunk to the free lists. 8558 if (inFreeRange()) { 8559 assert(freeFinger() < addr, "freeFinger points too high"); 8560 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 8561 } 8562 8563 // This object is live: we'd normally expect this to be 8564 // an oop, and like to assert the following: 8565 // assert(oop(addr)->is_oop(), "live block should be an oop"); 8566 // However, as we commented above, this may be an object whose 8567 // header hasn't yet been initialized. 8568 size_t size; 8569 assert(_bitMap->isMarked(addr), "Tautology for this control point"); 8570 if (_bitMap->isMarked(addr + 1)) { 8571 // Determine the size from the bit map, rather than trying to 8572 // compute it from the object header. 8573 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 8574 size = pointer_delta(nextOneAddr + 1, addr); 8575 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 8576 "alignment problem"); 8577 8578 #ifdef ASSERT 8579 if (oop(addr)->klass_or_null() != NULL) { 8580 // Ignore mark word because we are running concurrent with mutators 8581 assert(oop(addr)->is_oop(true), "live block should be an oop"); 8582 assert(size == 8583 CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()), 8584 "P-mark and computed size do not agree"); 8585 } 8586 #endif 8587 8588 } else { 8589 // This should be an initialized object that's alive. 8590 assert(oop(addr)->klass_or_null() != NULL, 8591 "Should be an initialized object"); 8592 // Ignore mark word because we are running concurrent with mutators 8593 assert(oop(addr)->is_oop(true), "live block should be an oop"); 8594 // Verify that the bit map has no bits marked between 8595 // addr and purported end of this block. 8596 size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 8597 assert(size >= 3, "Necessary for Printezis marks to work"); 8598 assert(!_bitMap->isMarked(addr+1), "Tautology for this control point"); 8599 DEBUG_ONLY(_bitMap->verifyNoOneBitsInRange(addr+2, addr+size);) 8600 } 8601 return size; 8602 } 8603 8604 void SweepClosure::do_post_free_or_garbage_chunk(FreeChunk* fc, 8605 size_t chunkSize) { 8606 // do_post_free_or_garbage_chunk() should only be called in the case 8607 // of the adaptive free list allocator. 8608 const bool fcInFreeLists = fc->is_free(); 8609 assert(_sp->adaptive_freelists(), "Should only be used in this case."); 8610 assert((HeapWord*)fc <= _limit, "sweep invariant"); 8611 if (CMSTestInFreeList && fcInFreeLists) { 8612 assert(_sp->verify_chunk_in_free_list(fc), "free chunk is not in free lists"); 8613 } 8614 8615 if (CMSTraceSweeper) { 8616 gclog_or_tty->print_cr(" -- pick up another chunk at 0x%x (%d)", fc, chunkSize); 8617 } 8618 8619 HeapWord* const fc_addr = (HeapWord*) fc; 8620 8621 bool coalesce; 8622 const size_t left = pointer_delta(fc_addr, freeFinger()); 8623 const size_t right = chunkSize; 8624 switch (FLSCoalescePolicy) { 8625 // numeric value forms a coalition aggressiveness metric 8626 case 0: { // never coalesce 8627 coalesce = false; 8628 break; 8629 } 8630 case 1: { // coalesce if left & right chunks on overpopulated lists 8631 coalesce = _sp->coalOverPopulated(left) && 8632 _sp->coalOverPopulated(right); 8633 break; 8634 } 8635 case 2: { // coalesce if left chunk on overpopulated list (default) 8636 coalesce = _sp->coalOverPopulated(left); 8637 break; 8638 } 8639 case 3: { // coalesce if left OR right chunk on overpopulated list 8640 coalesce = _sp->coalOverPopulated(left) || 8641 _sp->coalOverPopulated(right); 8642 break; 8643 } 8644 case 4: { // always coalesce 8645 coalesce = true; 8646 break; 8647 } 8648 default: 8649 ShouldNotReachHere(); 8650 } 8651 8652 // Should the current free range be coalesced? 8653 // If the chunk is in a free range and either we decided to coalesce above 8654 // or the chunk is near the large block at the end of the heap 8655 // (isNearLargestChunk() returns true), then coalesce this chunk. 8656 const bool doCoalesce = inFreeRange() 8657 && (coalesce || _g->isNearLargestChunk(fc_addr)); 8658 if (doCoalesce) { 8659 // Coalesce the current free range on the left with the new 8660 // chunk on the right. If either is on a free list, 8661 // it must be removed from the list and stashed in the closure. 8662 if (freeRangeInFreeLists()) { 8663 FreeChunk* const ffc = (FreeChunk*)freeFinger(); 8664 assert(ffc->size() == pointer_delta(fc_addr, freeFinger()), 8665 "Size of free range is inconsistent with chunk size."); 8666 if (CMSTestInFreeList) { 8667 assert(_sp->verify_chunk_in_free_list(ffc), 8668 "Chunk is not in free lists"); 8669 } 8670 _sp->coalDeath(ffc->size()); 8671 _sp->removeFreeChunkFromFreeLists(ffc); 8672 set_freeRangeInFreeLists(false); 8673 } 8674 if (fcInFreeLists) { 8675 _sp->coalDeath(chunkSize); 8676 assert(fc->size() == chunkSize, 8677 "The chunk has the wrong size or is not in the free lists"); 8678 _sp->removeFreeChunkFromFreeLists(fc); 8679 } 8680 set_lastFreeRangeCoalesced(true); 8681 print_free_block_coalesced(fc); 8682 } else { // not in a free range and/or should not coalesce 8683 // Return the current free range and start a new one. 8684 if (inFreeRange()) { 8685 // In a free range but cannot coalesce with the right hand chunk. 8686 // Put the current free range into the free lists. 8687 flush_cur_free_chunk(freeFinger(), 8688 pointer_delta(fc_addr, freeFinger())); 8689 } 8690 // Set up for new free range. Pass along whether the right hand 8691 // chunk is in the free lists. 8692 initialize_free_range((HeapWord*)fc, fcInFreeLists); 8693 } 8694 } 8695 8696 // Lookahead flush: 8697 // If we are tracking a free range, and this is the last chunk that 8698 // we'll look at because its end crosses past _limit, we'll preemptively 8699 // flush it along with any free range we may be holding on to. Note that 8700 // this can be the case only for an already free or freshly garbage 8701 // chunk. If this block is an object, it can never straddle 8702 // over _limit. The "straddling" occurs when _limit is set at 8703 // the previous end of the space when this cycle started, and 8704 // a subsequent heap expansion caused the previously co-terminal 8705 // free block to be coalesced with the newly expanded portion, 8706 // thus rendering _limit a non-block-boundary making it dangerous 8707 // for the sweeper to step over and examine. 8708 void SweepClosure::lookahead_and_flush(FreeChunk* fc, size_t chunk_size) { 8709 assert(inFreeRange(), "Should only be called if currently in a free range."); 8710 HeapWord* const eob = ((HeapWord*)fc) + chunk_size; 8711 assert(_sp->used_region().contains(eob - 1), 8712 err_msg("eob = " PTR_FORMAT " eob-1 = " PTR_FORMAT " _limit = " PTR_FORMAT 8713 " out of bounds wrt _sp = [" PTR_FORMAT "," PTR_FORMAT ")" 8714 " when examining fc = " PTR_FORMAT "(" SIZE_FORMAT ")", 8715 eob, eob-1, _limit, _sp->bottom(), _sp->end(), fc, chunk_size)); 8716 if (eob >= _limit) { 8717 assert(eob == _limit || fc->is_free(), "Only a free chunk should allow us to cross over the limit"); 8718 if (CMSTraceSweeper) { 8719 gclog_or_tty->print_cr("_limit " PTR_FORMAT " reached or crossed by block " 8720 "[" PTR_FORMAT "," PTR_FORMAT ") in space " 8721 "[" PTR_FORMAT "," PTR_FORMAT ")", 8722 _limit, fc, eob, _sp->bottom(), _sp->end()); 8723 } 8724 // Return the storage we are tracking back into the free lists. 8725 if (CMSTraceSweeper) { 8726 gclog_or_tty->print_cr("Flushing ... "); 8727 } 8728 assert(freeFinger() < eob, "Error"); 8729 flush_cur_free_chunk( freeFinger(), pointer_delta(eob, freeFinger())); 8730 } 8731 } 8732 8733 void SweepClosure::flush_cur_free_chunk(HeapWord* chunk, size_t size) { 8734 assert(inFreeRange(), "Should only be called if currently in a free range."); 8735 assert(size > 0, 8736 "A zero sized chunk cannot be added to the free lists."); 8737 if (!freeRangeInFreeLists()) { 8738 if (CMSTestInFreeList) { 8739 FreeChunk* fc = (FreeChunk*) chunk; 8740 fc->set_size(size); 8741 assert(!_sp->verify_chunk_in_free_list(fc), 8742 "chunk should not be in free lists yet"); 8743 } 8744 if (CMSTraceSweeper) { 8745 gclog_or_tty->print_cr(" -- add free block 0x%x (%d) to free lists", 8746 chunk, size); 8747 } 8748 // A new free range is going to be starting. The current 8749 // free range has not been added to the free lists yet or 8750 // was removed so add it back. 8751 // If the current free range was coalesced, then the death 8752 // of the free range was recorded. Record a birth now. 8753 if (lastFreeRangeCoalesced()) { 8754 _sp->coalBirth(size); 8755 } 8756 _sp->addChunkAndRepairOffsetTable(chunk, size, 8757 lastFreeRangeCoalesced()); 8758 } else if (CMSTraceSweeper) { 8759 gclog_or_tty->print_cr("Already in free list: nothing to flush"); 8760 } 8761 set_inFreeRange(false); 8762 set_freeRangeInFreeLists(false); 8763 } 8764 8765 // We take a break if we've been at this for a while, 8766 // so as to avoid monopolizing the locks involved. 8767 void SweepClosure::do_yield_work(HeapWord* addr) { 8768 // Return current free chunk being used for coalescing (if any) 8769 // to the appropriate freelist. After yielding, the next 8770 // free block encountered will start a coalescing range of 8771 // free blocks. If the next free block is adjacent to the 8772 // chunk just flushed, they will need to wait for the next 8773 // sweep to be coalesced. 8774 if (inFreeRange()) { 8775 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 8776 } 8777 8778 // First give up the locks, then yield, then re-lock. 8779 // We should probably use a constructor/destructor idiom to 8780 // do this unlock/lock or modify the MutexUnlocker class to 8781 // serve our purpose. XXX 8782 assert_lock_strong(_bitMap->lock()); 8783 assert_lock_strong(_freelistLock); 8784 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 8785 "CMS thread should hold CMS token"); 8786 _bitMap->lock()->unlock(); 8787 _freelistLock->unlock(); 8788 ConcurrentMarkSweepThread::desynchronize(true); 8789 ConcurrentMarkSweepThread::acknowledge_yield_request(); 8790 _collector->stopTimer(); 8791 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 8792 if (PrintCMSStatistics != 0) { 8793 _collector->incrementYields(); 8794 } 8795 _collector->icms_wait(); 8796 8797 // See the comment in coordinator_yield() 8798 for (unsigned i = 0; i < CMSYieldSleepCount && 8799 ConcurrentMarkSweepThread::should_yield() && 8800 !CMSCollector::foregroundGCIsActive(); ++i) { 8801 os::sleep(Thread::current(), 1, false); 8802 ConcurrentMarkSweepThread::acknowledge_yield_request(); 8803 } 8804 8805 ConcurrentMarkSweepThread::synchronize(true); 8806 _freelistLock->lock(); 8807 _bitMap->lock()->lock_without_safepoint_check(); 8808 _collector->startTimer(); 8809 } 8810 8811 #ifndef PRODUCT 8812 // This is actually very useful in a product build if it can 8813 // be called from the debugger. Compile it into the product 8814 // as needed. 8815 bool debug_verify_chunk_in_free_list(FreeChunk* fc) { 8816 return debug_cms_space->verify_chunk_in_free_list(fc); 8817 } 8818 #endif 8819 8820 void SweepClosure::print_free_block_coalesced(FreeChunk* fc) const { 8821 if (CMSTraceSweeper) { 8822 gclog_or_tty->print_cr("Sweep:coal_free_blk " PTR_FORMAT " (" SIZE_FORMAT ")", 8823 fc, fc->size()); 8824 } 8825 } 8826 8827 // CMSIsAliveClosure 8828 bool CMSIsAliveClosure::do_object_b(oop obj) { 8829 HeapWord* addr = (HeapWord*)obj; 8830 return addr != NULL && 8831 (!_span.contains(addr) || _bit_map->isMarked(addr)); 8832 } 8833 8834 8835 CMSKeepAliveClosure::CMSKeepAliveClosure( CMSCollector* collector, 8836 MemRegion span, 8837 CMSBitMap* bit_map, CMSMarkStack* mark_stack, 8838 bool cpc): 8839 _collector(collector), 8840 _span(span), 8841 _bit_map(bit_map), 8842 _mark_stack(mark_stack), 8843 _concurrent_precleaning(cpc) { 8844 assert(!_span.is_empty(), "Empty span could spell trouble"); 8845 } 8846 8847 8848 // CMSKeepAliveClosure: the serial version 8849 void CMSKeepAliveClosure::do_oop(oop obj) { 8850 HeapWord* addr = (HeapWord*)obj; 8851 if (_span.contains(addr) && 8852 !_bit_map->isMarked(addr)) { 8853 _bit_map->mark(addr); 8854 bool simulate_overflow = false; 8855 NOT_PRODUCT( 8856 if (CMSMarkStackOverflowALot && 8857 _collector->simulate_overflow()) { 8858 // simulate a stack overflow 8859 simulate_overflow = true; 8860 } 8861 ) 8862 if (simulate_overflow || !_mark_stack->push(obj)) { 8863 if (_concurrent_precleaning) { 8864 // We dirty the overflown object and let the remark 8865 // phase deal with it. 8866 assert(_collector->overflow_list_is_empty(), "Error"); 8867 // In the case of object arrays, we need to dirty all of 8868 // the cards that the object spans. No locking or atomics 8869 // are needed since no one else can be mutating the mod union 8870 // table. 8871 if (obj->is_objArray()) { 8872 size_t sz = obj->size(); 8873 HeapWord* end_card_addr = 8874 (HeapWord*)round_to((intptr_t)(addr+sz), CardTableModRefBS::card_size); 8875 MemRegion redirty_range = MemRegion(addr, end_card_addr); 8876 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 8877 _collector->_modUnionTable.mark_range(redirty_range); 8878 } else { 8879 _collector->_modUnionTable.mark(addr); 8880 } 8881 _collector->_ser_kac_preclean_ovflw++; 8882 } else { 8883 _collector->push_on_overflow_list(obj); 8884 _collector->_ser_kac_ovflw++; 8885 } 8886 } 8887 } 8888 } 8889 8890 void CMSKeepAliveClosure::do_oop(oop* p) { CMSKeepAliveClosure::do_oop_work(p); } 8891 void CMSKeepAliveClosure::do_oop(narrowOop* p) { CMSKeepAliveClosure::do_oop_work(p); } 8892 8893 // CMSParKeepAliveClosure: a parallel version of the above. 8894 // The work queues are private to each closure (thread), 8895 // but (may be) available for stealing by other threads. 8896 void CMSParKeepAliveClosure::do_oop(oop obj) { 8897 HeapWord* addr = (HeapWord*)obj; 8898 if (_span.contains(addr) && 8899 !_bit_map->isMarked(addr)) { 8900 // In general, during recursive tracing, several threads 8901 // may be concurrently getting here; the first one to 8902 // "tag" it, claims it. 8903 if (_bit_map->par_mark(addr)) { 8904 bool res = _work_queue->push(obj); 8905 assert(res, "Low water mark should be much less than capacity"); 8906 // Do a recursive trim in the hope that this will keep 8907 // stack usage lower, but leave some oops for potential stealers 8908 trim_queue(_low_water_mark); 8909 } // Else, another thread got there first 8910 } 8911 } 8912 8913 void CMSParKeepAliveClosure::do_oop(oop* p) { CMSParKeepAliveClosure::do_oop_work(p); } 8914 void CMSParKeepAliveClosure::do_oop(narrowOop* p) { CMSParKeepAliveClosure::do_oop_work(p); } 8915 8916 void CMSParKeepAliveClosure::trim_queue(uint max) { 8917 while (_work_queue->size() > max) { 8918 oop new_oop; 8919 if (_work_queue->pop_local(new_oop)) { 8920 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 8921 assert(_bit_map->isMarked((HeapWord*)new_oop), 8922 "no white objects on this stack!"); 8923 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 8924 // iterate over the oops in this oop, marking and pushing 8925 // the ones in CMS heap (i.e. in _span). 8926 new_oop->oop_iterate(&_mark_and_push); 8927 } 8928 } 8929 } 8930 8931 CMSInnerParMarkAndPushClosure::CMSInnerParMarkAndPushClosure( 8932 CMSCollector* collector, 8933 MemRegion span, CMSBitMap* bit_map, 8934 OopTaskQueue* work_queue): 8935 _collector(collector), 8936 _span(span), 8937 _bit_map(bit_map), 8938 _work_queue(work_queue) { } 8939 8940 void CMSInnerParMarkAndPushClosure::do_oop(oop obj) { 8941 HeapWord* addr = (HeapWord*)obj; 8942 if (_span.contains(addr) && 8943 !_bit_map->isMarked(addr)) { 8944 if (_bit_map->par_mark(addr)) { 8945 bool simulate_overflow = false; 8946 NOT_PRODUCT( 8947 if (CMSMarkStackOverflowALot && 8948 _collector->par_simulate_overflow()) { 8949 // simulate a stack overflow 8950 simulate_overflow = true; 8951 } 8952 ) 8953 if (simulate_overflow || !_work_queue->push(obj)) { 8954 _collector->par_push_on_overflow_list(obj); 8955 _collector->_par_kac_ovflw++; 8956 } 8957 } // Else another thread got there already 8958 } 8959 } 8960 8961 void CMSInnerParMarkAndPushClosure::do_oop(oop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); } 8962 void CMSInnerParMarkAndPushClosure::do_oop(narrowOop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); } 8963 8964 ////////////////////////////////////////////////////////////////// 8965 // CMSExpansionCause ///////////////////////////// 8966 ////////////////////////////////////////////////////////////////// 8967 const char* CMSExpansionCause::to_string(CMSExpansionCause::Cause cause) { 8968 switch (cause) { 8969 case _no_expansion: 8970 return "No expansion"; 8971 case _satisfy_free_ratio: 8972 return "Free ratio"; 8973 case _satisfy_promotion: 8974 return "Satisfy promotion"; 8975 case _satisfy_allocation: 8976 return "allocation"; 8977 case _allocate_par_lab: 8978 return "Par LAB"; 8979 case _allocate_par_spooling_space: 8980 return "Par Spooling Space"; 8981 case _adaptive_size_policy: 8982 return "Ergonomics"; 8983 default: 8984 return "unknown"; 8985 } 8986 } 8987 8988 void CMSDrainMarkingStackClosure::do_void() { 8989 // the max number to take from overflow list at a time 8990 const size_t num = _mark_stack->capacity()/4; 8991 assert(!_concurrent_precleaning || _collector->overflow_list_is_empty(), 8992 "Overflow list should be NULL during concurrent phases"); 8993 while (!_mark_stack->isEmpty() || 8994 // if stack is empty, check the overflow list 8995 _collector->take_from_overflow_list(num, _mark_stack)) { 8996 oop obj = _mark_stack->pop(); 8997 HeapWord* addr = (HeapWord*)obj; 8998 assert(_span.contains(addr), "Should be within span"); 8999 assert(_bit_map->isMarked(addr), "Should be marked"); 9000 assert(obj->is_oop(), "Should be an oop"); 9001 obj->oop_iterate(_keep_alive); 9002 } 9003 } 9004 9005 void CMSParDrainMarkingStackClosure::do_void() { 9006 // drain queue 9007 trim_queue(0); 9008 } 9009 9010 // Trim our work_queue so its length is below max at return 9011 void CMSParDrainMarkingStackClosure::trim_queue(uint max) { 9012 while (_work_queue->size() > max) { 9013 oop new_oop; 9014 if (_work_queue->pop_local(new_oop)) { 9015 assert(new_oop->is_oop(), "Expected an oop"); 9016 assert(_bit_map->isMarked((HeapWord*)new_oop), 9017 "no white objects on this stack!"); 9018 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 9019 // iterate over the oops in this oop, marking and pushing 9020 // the ones in CMS heap (i.e. in _span). 9021 new_oop->oop_iterate(&_mark_and_push); 9022 } 9023 } 9024 } 9025 9026 //////////////////////////////////////////////////////////////////// 9027 // Support for Marking Stack Overflow list handling and related code 9028 //////////////////////////////////////////////////////////////////// 9029 // Much of the following code is similar in shape and spirit to the 9030 // code used in ParNewGC. We should try and share that code 9031 // as much as possible in the future. 9032 9033 #ifndef PRODUCT 9034 // Debugging support for CMSStackOverflowALot 9035 9036 // It's OK to call this multi-threaded; the worst thing 9037 // that can happen is that we'll get a bunch of closely 9038 // spaced simulated oveflows, but that's OK, in fact 9039 // probably good as it would exercise the overflow code 9040 // under contention. 9041 bool CMSCollector::simulate_overflow() { 9042 if (_overflow_counter-- <= 0) { // just being defensive 9043 _overflow_counter = CMSMarkStackOverflowInterval; 9044 return true; 9045 } else { 9046 return false; 9047 } 9048 } 9049 9050 bool CMSCollector::par_simulate_overflow() { 9051 return simulate_overflow(); 9052 } 9053 #endif 9054 9055 // Single-threaded 9056 bool CMSCollector::take_from_overflow_list(size_t num, CMSMarkStack* stack) { 9057 assert(stack->isEmpty(), "Expected precondition"); 9058 assert(stack->capacity() > num, "Shouldn't bite more than can chew"); 9059 size_t i = num; 9060 oop cur = _overflow_list; 9061 const markOop proto = markOopDesc::prototype(); 9062 NOT_PRODUCT(ssize_t n = 0;) 9063 for (oop next; i > 0 && cur != NULL; cur = next, i--) { 9064 next = oop(cur->mark()); 9065 cur->set_mark(proto); // until proven otherwise 9066 assert(cur->is_oop(), "Should be an oop"); 9067 bool res = stack->push(cur); 9068 assert(res, "Bit off more than can chew?"); 9069 NOT_PRODUCT(n++;) 9070 } 9071 _overflow_list = cur; 9072 #ifndef PRODUCT 9073 assert(_num_par_pushes >= n, "Too many pops?"); 9074 _num_par_pushes -=n; 9075 #endif 9076 return !stack->isEmpty(); 9077 } 9078 9079 #define BUSY (oop(0x1aff1aff)) 9080 // (MT-safe) Get a prefix of at most "num" from the list. 9081 // The overflow list is chained through the mark word of 9082 // each object in the list. We fetch the entire list, 9083 // break off a prefix of the right size and return the 9084 // remainder. If other threads try to take objects from 9085 // the overflow list at that time, they will wait for 9086 // some time to see if data becomes available. If (and 9087 // only if) another thread places one or more object(s) 9088 // on the global list before we have returned the suffix 9089 // to the global list, we will walk down our local list 9090 // to find its end and append the global list to 9091 // our suffix before returning it. This suffix walk can 9092 // prove to be expensive (quadratic in the amount of traffic) 9093 // when there are many objects in the overflow list and 9094 // there is much producer-consumer contention on the list. 9095 // *NOTE*: The overflow list manipulation code here and 9096 // in ParNewGeneration:: are very similar in shape, 9097 // except that in the ParNew case we use the old (from/eden) 9098 // copy of the object to thread the list via its klass word. 9099 // Because of the common code, if you make any changes in 9100 // the code below, please check the ParNew version to see if 9101 // similar changes might be needed. 9102 // CR 6797058 has been filed to consolidate the common code. 9103 bool CMSCollector::par_take_from_overflow_list(size_t num, 9104 OopTaskQueue* work_q, 9105 int no_of_gc_threads) { 9106 assert(work_q->size() == 0, "First empty local work queue"); 9107 assert(num < work_q->max_elems(), "Can't bite more than we can chew"); 9108 if (_overflow_list == NULL) { 9109 return false; 9110 } 9111 // Grab the entire list; we'll put back a suffix 9112 oop prefix = (oop)Atomic::xchg_ptr(BUSY, &_overflow_list); 9113 Thread* tid = Thread::current(); 9114 // Before "no_of_gc_threads" was introduced CMSOverflowSpinCount was 9115 // set to ParallelGCThreads. 9116 size_t CMSOverflowSpinCount = (size_t) no_of_gc_threads; // was ParallelGCThreads; 9117 size_t sleep_time_millis = MAX2((size_t)1, num/100); 9118 // If the list is busy, we spin for a short while, 9119 // sleeping between attempts to get the list. 9120 for (size_t spin = 0; prefix == BUSY && spin < CMSOverflowSpinCount; spin++) { 9121 os::sleep(tid, sleep_time_millis, false); 9122 if (_overflow_list == NULL) { 9123 // Nothing left to take 9124 return false; 9125 } else if (_overflow_list != BUSY) { 9126 // Try and grab the prefix 9127 prefix = (oop)Atomic::xchg_ptr(BUSY, &_overflow_list); 9128 } 9129 } 9130 // If the list was found to be empty, or we spun long 9131 // enough, we give up and return empty-handed. If we leave 9132 // the list in the BUSY state below, it must be the case that 9133 // some other thread holds the overflow list and will set it 9134 // to a non-BUSY state in the future. 9135 if (prefix == NULL || prefix == BUSY) { 9136 // Nothing to take or waited long enough 9137 if (prefix == NULL) { 9138 // Write back the NULL in case we overwrote it with BUSY above 9139 // and it is still the same value. 9140 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY); 9141 } 9142 return false; 9143 } 9144 assert(prefix != NULL && prefix != BUSY, "Error"); 9145 size_t i = num; 9146 oop cur = prefix; 9147 // Walk down the first "num" objects, unless we reach the end. 9148 for (; i > 1 && cur->mark() != NULL; cur = oop(cur->mark()), i--); 9149 if (cur->mark() == NULL) { 9150 // We have "num" or fewer elements in the list, so there 9151 // is nothing to return to the global list. 9152 // Write back the NULL in lieu of the BUSY we wrote 9153 // above, if it is still the same value. 9154 if (_overflow_list == BUSY) { 9155 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY); 9156 } 9157 } else { 9158 // Chop off the suffix and rerturn it to the global list. 9159 assert(cur->mark() != BUSY, "Error"); 9160 oop suffix_head = cur->mark(); // suffix will be put back on global list 9161 cur->set_mark(NULL); // break off suffix 9162 // It's possible that the list is still in the empty(busy) state 9163 // we left it in a short while ago; in that case we may be 9164 // able to place back the suffix without incurring the cost 9165 // of a walk down the list. 9166 oop observed_overflow_list = _overflow_list; 9167 oop cur_overflow_list = observed_overflow_list; 9168 bool attached = false; 9169 while (observed_overflow_list == BUSY || observed_overflow_list == NULL) { 9170 observed_overflow_list = 9171 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list); 9172 if (cur_overflow_list == observed_overflow_list) { 9173 attached = true; 9174 break; 9175 } else cur_overflow_list = observed_overflow_list; 9176 } 9177 if (!attached) { 9178 // Too bad, someone else sneaked in (at least) an element; we'll need 9179 // to do a splice. Find tail of suffix so we can prepend suffix to global 9180 // list. 9181 for (cur = suffix_head; cur->mark() != NULL; cur = (oop)(cur->mark())); 9182 oop suffix_tail = cur; 9183 assert(suffix_tail != NULL && suffix_tail->mark() == NULL, 9184 "Tautology"); 9185 observed_overflow_list = _overflow_list; 9186 do { 9187 cur_overflow_list = observed_overflow_list; 9188 if (cur_overflow_list != BUSY) { 9189 // Do the splice ... 9190 suffix_tail->set_mark(markOop(cur_overflow_list)); 9191 } else { // cur_overflow_list == BUSY 9192 suffix_tail->set_mark(NULL); 9193 } 9194 // ... and try to place spliced list back on overflow_list ... 9195 observed_overflow_list = 9196 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list); 9197 } while (cur_overflow_list != observed_overflow_list); 9198 // ... until we have succeeded in doing so. 9199 } 9200 } 9201 9202 // Push the prefix elements on work_q 9203 assert(prefix != NULL, "control point invariant"); 9204 const markOop proto = markOopDesc::prototype(); 9205 oop next; 9206 NOT_PRODUCT(ssize_t n = 0;) 9207 for (cur = prefix; cur != NULL; cur = next) { 9208 next = oop(cur->mark()); 9209 cur->set_mark(proto); // until proven otherwise 9210 assert(cur->is_oop(), "Should be an oop"); 9211 bool res = work_q->push(cur); 9212 assert(res, "Bit off more than we can chew?"); 9213 NOT_PRODUCT(n++;) 9214 } 9215 #ifndef PRODUCT 9216 assert(_num_par_pushes >= n, "Too many pops?"); 9217 Atomic::add_ptr(-(intptr_t)n, &_num_par_pushes); 9218 #endif 9219 return true; 9220 } 9221 9222 // Single-threaded 9223 void CMSCollector::push_on_overflow_list(oop p) { 9224 NOT_PRODUCT(_num_par_pushes++;) 9225 assert(p->is_oop(), "Not an oop"); 9226 preserve_mark_if_necessary(p); 9227 p->set_mark((markOop)_overflow_list); 9228 _overflow_list = p; 9229 } 9230 9231 // Multi-threaded; use CAS to prepend to overflow list 9232 void CMSCollector::par_push_on_overflow_list(oop p) { 9233 NOT_PRODUCT(Atomic::inc_ptr(&_num_par_pushes);) 9234 assert(p->is_oop(), "Not an oop"); 9235 par_preserve_mark_if_necessary(p); 9236 oop observed_overflow_list = _overflow_list; 9237 oop cur_overflow_list; 9238 do { 9239 cur_overflow_list = observed_overflow_list; 9240 if (cur_overflow_list != BUSY) { 9241 p->set_mark(markOop(cur_overflow_list)); 9242 } else { 9243 p->set_mark(NULL); 9244 } 9245 observed_overflow_list = 9246 (oop) Atomic::cmpxchg_ptr(p, &_overflow_list, cur_overflow_list); 9247 } while (cur_overflow_list != observed_overflow_list); 9248 } 9249 #undef BUSY 9250 9251 // Single threaded 9252 // General Note on GrowableArray: pushes may silently fail 9253 // because we are (temporarily) out of C-heap for expanding 9254 // the stack. The problem is quite ubiquitous and affects 9255 // a lot of code in the JVM. The prudent thing for GrowableArray 9256 // to do (for now) is to exit with an error. However, that may 9257 // be too draconian in some cases because the caller may be 9258 // able to recover without much harm. For such cases, we 9259 // should probably introduce a "soft_push" method which returns 9260 // an indication of success or failure with the assumption that 9261 // the caller may be able to recover from a failure; code in 9262 // the VM can then be changed, incrementally, to deal with such 9263 // failures where possible, thus, incrementally hardening the VM 9264 // in such low resource situations. 9265 void CMSCollector::preserve_mark_work(oop p, markOop m) { 9266 _preserved_oop_stack.push(p); 9267 _preserved_mark_stack.push(m); 9268 assert(m == p->mark(), "Mark word changed"); 9269 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), 9270 "bijection"); 9271 } 9272 9273 // Single threaded 9274 void CMSCollector::preserve_mark_if_necessary(oop p) { 9275 markOop m = p->mark(); 9276 if (m->must_be_preserved(p)) { 9277 preserve_mark_work(p, m); 9278 } 9279 } 9280 9281 void CMSCollector::par_preserve_mark_if_necessary(oop p) { 9282 markOop m = p->mark(); 9283 if (m->must_be_preserved(p)) { 9284 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag); 9285 // Even though we read the mark word without holding 9286 // the lock, we are assured that it will not change 9287 // because we "own" this oop, so no other thread can 9288 // be trying to push it on the overflow list; see 9289 // the assertion in preserve_mark_work() that checks 9290 // that m == p->mark(). 9291 preserve_mark_work(p, m); 9292 } 9293 } 9294 9295 // We should be able to do this multi-threaded, 9296 // a chunk of stack being a task (this is 9297 // correct because each oop only ever appears 9298 // once in the overflow list. However, it's 9299 // not very easy to completely overlap this with 9300 // other operations, so will generally not be done 9301 // until all work's been completed. Because we 9302 // expect the preserved oop stack (set) to be small, 9303 // it's probably fine to do this single-threaded. 9304 // We can explore cleverer concurrent/overlapped/parallel 9305 // processing of preserved marks if we feel the 9306 // need for this in the future. Stack overflow should 9307 // be so rare in practice and, when it happens, its 9308 // effect on performance so great that this will 9309 // likely just be in the noise anyway. 9310 void CMSCollector::restore_preserved_marks_if_any() { 9311 assert(SafepointSynchronize::is_at_safepoint(), 9312 "world should be stopped"); 9313 assert(Thread::current()->is_ConcurrentGC_thread() || 9314 Thread::current()->is_VM_thread(), 9315 "should be single-threaded"); 9316 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), 9317 "bijection"); 9318 9319 while (!_preserved_oop_stack.is_empty()) { 9320 oop p = _preserved_oop_stack.pop(); 9321 assert(p->is_oop(), "Should be an oop"); 9322 assert(_span.contains(p), "oop should be in _span"); 9323 assert(p->mark() == markOopDesc::prototype(), 9324 "Set when taken from overflow list"); 9325 markOop m = _preserved_mark_stack.pop(); 9326 p->set_mark(m); 9327 } 9328 assert(_preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(), 9329 "stacks were cleared above"); 9330 } 9331 9332 #ifndef PRODUCT 9333 bool CMSCollector::no_preserved_marks() const { 9334 return _preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(); 9335 } 9336 #endif 9337 9338 CMSAdaptiveSizePolicy* ASConcurrentMarkSweepGeneration::cms_size_policy() const 9339 { 9340 GenCollectedHeap* gch = (GenCollectedHeap*) GenCollectedHeap::heap(); 9341 CMSAdaptiveSizePolicy* size_policy = 9342 (CMSAdaptiveSizePolicy*) gch->gen_policy()->size_policy(); 9343 assert(size_policy->is_gc_cms_adaptive_size_policy(), 9344 "Wrong type for size policy"); 9345 return size_policy; 9346 } 9347 9348 void ASConcurrentMarkSweepGeneration::resize(size_t cur_promo_size, 9349 size_t desired_promo_size) { 9350 if (cur_promo_size < desired_promo_size) { 9351 size_t expand_bytes = desired_promo_size - cur_promo_size; 9352 if (PrintAdaptiveSizePolicy && Verbose) { 9353 gclog_or_tty->print_cr(" ASConcurrentMarkSweepGeneration::resize " 9354 "Expanding tenured generation by " SIZE_FORMAT " (bytes)", 9355 expand_bytes); 9356 } 9357 expand(expand_bytes, 9358 MinHeapDeltaBytes, 9359 CMSExpansionCause::_adaptive_size_policy); 9360 } else if (desired_promo_size < cur_promo_size) { 9361 size_t shrink_bytes = cur_promo_size - desired_promo_size; 9362 if (PrintAdaptiveSizePolicy && Verbose) { 9363 gclog_or_tty->print_cr(" ASConcurrentMarkSweepGeneration::resize " 9364 "Shrinking tenured generation by " SIZE_FORMAT " (bytes)", 9365 shrink_bytes); 9366 } 9367 shrink(shrink_bytes); 9368 } 9369 } 9370 9371 CMSGCAdaptivePolicyCounters* ASConcurrentMarkSweepGeneration::gc_adaptive_policy_counters() { 9372 GenCollectedHeap* gch = GenCollectedHeap::heap(); 9373 CMSGCAdaptivePolicyCounters* counters = 9374 (CMSGCAdaptivePolicyCounters*) gch->collector_policy()->counters(); 9375 assert(counters->kind() == GCPolicyCounters::CMSGCAdaptivePolicyCountersKind, 9376 "Wrong kind of counters"); 9377 return counters; 9378 } 9379 9380 9381 void ASConcurrentMarkSweepGeneration::update_counters() { 9382 if (UsePerfData) { 9383 _space_counters->update_all(); 9384 _gen_counters->update_all(); 9385 CMSGCAdaptivePolicyCounters* counters = gc_adaptive_policy_counters(); 9386 GenCollectedHeap* gch = GenCollectedHeap::heap(); 9387 CMSGCStats* gc_stats_l = (CMSGCStats*) gc_stats(); 9388 assert(gc_stats_l->kind() == GCStats::CMSGCStatsKind, 9389 "Wrong gc statistics type"); 9390 counters->update_counters(gc_stats_l); 9391 } 9392 } 9393 9394 void ASConcurrentMarkSweepGeneration::update_counters(size_t used) { 9395 if (UsePerfData) { 9396 _space_counters->update_used(used); 9397 _space_counters->update_capacity(); 9398 _gen_counters->update_all(); 9399 9400 CMSGCAdaptivePolicyCounters* counters = gc_adaptive_policy_counters(); 9401 GenCollectedHeap* gch = GenCollectedHeap::heap(); 9402 CMSGCStats* gc_stats_l = (CMSGCStats*) gc_stats(); 9403 assert(gc_stats_l->kind() == GCStats::CMSGCStatsKind, 9404 "Wrong gc statistics type"); 9405 counters->update_counters(gc_stats_l); 9406 } 9407 } 9408 9409 void ASConcurrentMarkSweepGeneration::shrink_by(size_t desired_bytes) { 9410 assert_locked_or_safepoint(Heap_lock); 9411 assert_lock_strong(freelistLock()); 9412 HeapWord* old_end = _cmsSpace->end(); 9413 HeapWord* unallocated_start = _cmsSpace->unallocated_block(); 9414 assert(old_end >= unallocated_start, "Miscalculation of unallocated_start"); 9415 FreeChunk* chunk_at_end = find_chunk_at_end(); 9416 if (chunk_at_end == NULL) { 9417 // No room to shrink 9418 if (PrintGCDetails && Verbose) { 9419 gclog_or_tty->print_cr("No room to shrink: old_end " 9420 PTR_FORMAT " unallocated_start " PTR_FORMAT 9421 " chunk_at_end " PTR_FORMAT, 9422 old_end, unallocated_start, chunk_at_end); 9423 } 9424 return; 9425 } else { 9426 9427 // Find the chunk at the end of the space and determine 9428 // how much it can be shrunk. 9429 size_t shrinkable_size_in_bytes = chunk_at_end->size(); 9430 size_t aligned_shrinkable_size_in_bytes = 9431 align_size_down(shrinkable_size_in_bytes, os::vm_page_size()); 9432 assert(unallocated_start <= (HeapWord*) chunk_at_end->end(), 9433 "Inconsistent chunk at end of space"); 9434 size_t bytes = MIN2(desired_bytes, aligned_shrinkable_size_in_bytes); 9435 size_t word_size_before = heap_word_size(_virtual_space.committed_size()); 9436 9437 // Shrink the underlying space 9438 _virtual_space.shrink_by(bytes); 9439 if (PrintGCDetails && Verbose) { 9440 gclog_or_tty->print_cr("ConcurrentMarkSweepGeneration::shrink_by:" 9441 " desired_bytes " SIZE_FORMAT 9442 " shrinkable_size_in_bytes " SIZE_FORMAT 9443 " aligned_shrinkable_size_in_bytes " SIZE_FORMAT 9444 " bytes " SIZE_FORMAT, 9445 desired_bytes, shrinkable_size_in_bytes, 9446 aligned_shrinkable_size_in_bytes, bytes); 9447 gclog_or_tty->print_cr(" old_end " SIZE_FORMAT 9448 " unallocated_start " SIZE_FORMAT, 9449 old_end, unallocated_start); 9450 } 9451 9452 // If the space did shrink (shrinking is not guaranteed), 9453 // shrink the chunk at the end by the appropriate amount. 9454 if (((HeapWord*)_virtual_space.high()) < old_end) { 9455 size_t new_word_size = 9456 heap_word_size(_virtual_space.committed_size()); 9457 9458 // Have to remove the chunk from the dictionary because it is changing 9459 // size and might be someplace elsewhere in the dictionary. 9460 9461 // Get the chunk at end, shrink it, and put it 9462 // back. 9463 _cmsSpace->removeChunkFromDictionary(chunk_at_end); 9464 size_t word_size_change = word_size_before - new_word_size; 9465 size_t chunk_at_end_old_size = chunk_at_end->size(); 9466 assert(chunk_at_end_old_size >= word_size_change, 9467 "Shrink is too large"); 9468 chunk_at_end->set_size(chunk_at_end_old_size - 9469 word_size_change); 9470 _cmsSpace->freed((HeapWord*) chunk_at_end->end(), 9471 word_size_change); 9472 9473 _cmsSpace->returnChunkToDictionary(chunk_at_end); 9474 9475 MemRegion mr(_cmsSpace->bottom(), new_word_size); 9476 _bts->resize(new_word_size); // resize the block offset shared array 9477 Universe::heap()->barrier_set()->resize_covered_region(mr); 9478 _cmsSpace->assert_locked(); 9479 _cmsSpace->set_end((HeapWord*)_virtual_space.high()); 9480 9481 NOT_PRODUCT(_cmsSpace->dictionary()->verify()); 9482 9483 // update the space and generation capacity counters 9484 if (UsePerfData) { 9485 _space_counters->update_capacity(); 9486 _gen_counters->update_all(); 9487 } 9488 9489 if (Verbose && PrintGCDetails) { 9490 size_t new_mem_size = _virtual_space.committed_size(); 9491 size_t old_mem_size = new_mem_size + bytes; 9492 gclog_or_tty->print_cr("Shrinking %s from " SIZE_FORMAT "K by " SIZE_FORMAT "K to " SIZE_FORMAT "K", 9493 name(), old_mem_size/K, bytes/K, new_mem_size/K); 9494 } 9495 } 9496 9497 assert(_cmsSpace->unallocated_block() <= _cmsSpace->end(), 9498 "Inconsistency at end of space"); 9499 assert(chunk_at_end->end() == (uintptr_t*) _cmsSpace->end(), 9500 "Shrinking is inconsistent"); 9501 return; 9502 } 9503 } 9504 // Transfer some number of overflown objects to usual marking 9505 // stack. Return true if some objects were transferred. 9506 bool MarkRefsIntoAndScanClosure::take_from_overflow_list() { 9507 size_t num = MIN2((size_t)(_mark_stack->capacity() - _mark_stack->length())/4, 9508 (size_t)ParGCDesiredObjsFromOverflowList); 9509 9510 bool res = _collector->take_from_overflow_list(num, _mark_stack); 9511 assert(_collector->overflow_list_is_empty() || res, 9512 "If list is not empty, we should have taken something"); 9513 assert(!res || !_mark_stack->isEmpty(), 9514 "If we took something, it should now be on our stack"); 9515 return res; 9516 } 9517 9518 size_t MarkDeadObjectsClosure::do_blk(HeapWord* addr) { 9519 size_t res = _sp->block_size_no_stall(addr, _collector); 9520 if (_sp->block_is_obj(addr)) { 9521 if (_live_bit_map->isMarked(addr)) { 9522 // It can't have been dead in a previous cycle 9523 guarantee(!_dead_bit_map->isMarked(addr), "No resurrection!"); 9524 } else { 9525 _dead_bit_map->mark(addr); // mark the dead object 9526 } 9527 } 9528 // Could be 0, if the block size could not be computed without stalling. 9529 return res; 9530 } 9531 9532 TraceCMSMemoryManagerStats::TraceCMSMemoryManagerStats(CMSCollector::CollectorState phase, GCCause::Cause cause): TraceMemoryManagerStats() { 9533 9534 switch (phase) { 9535 case CMSCollector::InitialMarking: 9536 initialize(true /* fullGC */ , 9537 cause /* cause of the GC */, 9538 true /* recordGCBeginTime */, 9539 true /* recordPreGCUsage */, 9540 false /* recordPeakUsage */, 9541 false /* recordPostGCusage */, 9542 true /* recordAccumulatedGCTime */, 9543 false /* recordGCEndTime */, 9544 false /* countCollection */ ); 9545 break; 9546 9547 case CMSCollector::FinalMarking: 9548 initialize(true /* fullGC */ , 9549 cause /* cause of the GC */, 9550 false /* recordGCBeginTime */, 9551 false /* recordPreGCUsage */, 9552 false /* recordPeakUsage */, 9553 false /* recordPostGCusage */, 9554 true /* recordAccumulatedGCTime */, 9555 false /* recordGCEndTime */, 9556 false /* countCollection */ ); 9557 break; 9558 9559 case CMSCollector::Sweeping: 9560 initialize(true /* fullGC */ , 9561 cause /* cause of the GC */, 9562 false /* recordGCBeginTime */, 9563 false /* recordPreGCUsage */, 9564 true /* recordPeakUsage */, 9565 true /* recordPostGCusage */, 9566 false /* recordAccumulatedGCTime */, 9567 true /* recordGCEndTime */, 9568 true /* countCollection */ ); 9569 break; 9570 9571 default: 9572 ShouldNotReachHere(); 9573 } 9574 }