1 /* 2 * Copyright (c) 2001, 2014, 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/stringTable.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/orderAccess.inline.hpp" 63 #include "runtime/vmThread.hpp" 64 #include "services/memoryService.hpp" 65 #include "services/runtimeService.hpp" 66 67 PRAGMA_FORMAT_MUTE_WARNINGS_FOR_GCC 68 69 // statics 70 CMSCollector* ConcurrentMarkSweepGeneration::_collector = NULL; 71 bool CMSCollector::_full_gc_requested = false; 72 GCCause::Cause CMSCollector::_full_gc_cause = GCCause::_no_gc; 73 74 ////////////////////////////////////////////////////////////////// 75 // In support of CMS/VM thread synchronization 76 ////////////////////////////////////////////////////////////////// 77 // We split use of the CGC_lock into 2 "levels". 78 // The low-level locking is of the usual CGC_lock monitor. We introduce 79 // a higher level "token" (hereafter "CMS token") built on top of the 80 // low level monitor (hereafter "CGC lock"). 81 // The token-passing protocol gives priority to the VM thread. The 82 // CMS-lock doesn't provide any fairness guarantees, but clients 83 // should ensure that it is only held for very short, bounded 84 // durations. 85 // 86 // When either of the CMS thread or the VM thread is involved in 87 // collection operations during which it does not want the other 88 // thread to interfere, it obtains the CMS token. 89 // 90 // If either thread tries to get the token while the other has 91 // it, that thread waits. However, if the VM thread and CMS thread 92 // both want the token, then the VM thread gets priority while the 93 // CMS thread waits. This ensures, for instance, that the "concurrent" 94 // phases of the CMS thread's work do not block out the VM thread 95 // for long periods of time as the CMS thread continues to hog 96 // the token. (See bug 4616232). 97 // 98 // The baton-passing functions are, however, controlled by the 99 // flags _foregroundGCShouldWait and _foregroundGCIsActive, 100 // and here the low-level CMS lock, not the high level token, 101 // ensures mutual exclusion. 102 // 103 // Two important conditions that we have to satisfy: 104 // 1. if a thread does a low-level wait on the CMS lock, then it 105 // relinquishes the CMS token if it were holding that token 106 // when it acquired the low-level CMS lock. 107 // 2. any low-level notifications on the low-level lock 108 // should only be sent when a thread has relinquished the token. 109 // 110 // In the absence of either property, we'd have potential deadlock. 111 // 112 // We protect each of the CMS (concurrent and sequential) phases 113 // with the CMS _token_, not the CMS _lock_. 114 // 115 // The only code protected by CMS lock is the token acquisition code 116 // itself, see ConcurrentMarkSweepThread::[de]synchronize(), and the 117 // baton-passing code. 118 // 119 // Unfortunately, i couldn't come up with a good abstraction to factor and 120 // hide the naked CGC_lock manipulation in the baton-passing code 121 // further below. That's something we should try to do. Also, the proof 122 // of correctness of this 2-level locking scheme is far from obvious, 123 // and potentially quite slippery. We have an uneasy suspicion, for instance, 124 // that there may be a theoretical possibility of delay/starvation in the 125 // low-level lock/wait/notify scheme used for the baton-passing because of 126 // potential interference with the priority scheme embodied in the 127 // CMS-token-passing protocol. See related comments at a CGC_lock->wait() 128 // invocation further below and marked with "XXX 20011219YSR". 129 // Indeed, as we note elsewhere, this may become yet more slippery 130 // in the presence of multiple CMS and/or multiple VM threads. XXX 131 132 class CMSTokenSync: public StackObj { 133 private: 134 bool _is_cms_thread; 135 public: 136 CMSTokenSync(bool is_cms_thread): 137 _is_cms_thread(is_cms_thread) { 138 assert(is_cms_thread == Thread::current()->is_ConcurrentGC_thread(), 139 "Incorrect argument to constructor"); 140 ConcurrentMarkSweepThread::synchronize(_is_cms_thread); 141 } 142 143 ~CMSTokenSync() { 144 assert(_is_cms_thread ? 145 ConcurrentMarkSweepThread::cms_thread_has_cms_token() : 146 ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 147 "Incorrect state"); 148 ConcurrentMarkSweepThread::desynchronize(_is_cms_thread); 149 } 150 }; 151 152 // Convenience class that does a CMSTokenSync, and then acquires 153 // upto three locks. 154 class CMSTokenSyncWithLocks: public CMSTokenSync { 155 private: 156 // Note: locks are acquired in textual declaration order 157 // and released in the opposite order 158 MutexLockerEx _locker1, _locker2, _locker3; 159 public: 160 CMSTokenSyncWithLocks(bool is_cms_thread, Mutex* mutex1, 161 Mutex* mutex2 = NULL, Mutex* mutex3 = NULL): 162 CMSTokenSync(is_cms_thread), 163 _locker1(mutex1, Mutex::_no_safepoint_check_flag), 164 _locker2(mutex2, Mutex::_no_safepoint_check_flag), 165 _locker3(mutex3, Mutex::_no_safepoint_check_flag) 166 { } 167 }; 168 169 170 // Wrapper class to temporarily disable icms during a foreground cms collection. 171 class ICMSDisabler: public StackObj { 172 public: 173 // The ctor disables icms and wakes up the thread so it notices the change; 174 // the dtor re-enables icms. Note that the CMSCollector methods will check 175 // CMSIncrementalMode. 176 ICMSDisabler() { CMSCollector::disable_icms(); CMSCollector::start_icms(); } 177 ~ICMSDisabler() { CMSCollector::enable_icms(); } 178 }; 179 180 ////////////////////////////////////////////////////////////////// 181 // Concurrent Mark-Sweep Generation ///////////////////////////// 182 ////////////////////////////////////////////////////////////////// 183 184 NOT_PRODUCT(CompactibleFreeListSpace* debug_cms_space;) 185 186 // This struct contains per-thread things necessary to support parallel 187 // young-gen collection. 188 class CMSParGCThreadState: public CHeapObj<mtGC> { 189 public: 190 CFLS_LAB lab; 191 PromotionInfo promo; 192 193 // Constructor. 194 CMSParGCThreadState(CompactibleFreeListSpace* cfls) : lab(cfls) { 195 promo.setSpace(cfls); 196 } 197 }; 198 199 ConcurrentMarkSweepGeneration::ConcurrentMarkSweepGeneration( 200 ReservedSpace rs, size_t initial_byte_size, int level, 201 CardTableRS* ct, bool use_adaptive_freelists, 202 FreeBlockDictionary<FreeChunk>::DictionaryChoice dictionaryChoice) : 203 CardGeneration(rs, initial_byte_size, level, ct), 204 _dilatation_factor(((double)MinChunkSize)/((double)(CollectedHeap::min_fill_size()))), 205 _debug_collection_type(Concurrent_collection_type), 206 _did_compact(false) 207 { 208 HeapWord* bottom = (HeapWord*) _virtual_space.low(); 209 HeapWord* end = (HeapWord*) _virtual_space.high(); 210 211 _direct_allocated_words = 0; 212 NOT_PRODUCT( 213 _numObjectsPromoted = 0; 214 _numWordsPromoted = 0; 215 _numObjectsAllocated = 0; 216 _numWordsAllocated = 0; 217 ) 218 219 _cmsSpace = new CompactibleFreeListSpace(_bts, MemRegion(bottom, end), 220 use_adaptive_freelists, 221 dictionaryChoice); 222 NOT_PRODUCT(debug_cms_space = _cmsSpace;) 223 if (_cmsSpace == NULL) { 224 vm_exit_during_initialization( 225 "CompactibleFreeListSpace allocation failure"); 226 } 227 _cmsSpace->_gen = this; 228 229 _gc_stats = new CMSGCStats(); 230 231 // Verify the assumption that FreeChunk::_prev and OopDesc::_klass 232 // offsets match. The ability to tell free chunks from objects 233 // depends on this property. 234 debug_only( 235 FreeChunk* junk = NULL; 236 assert(UseCompressedClassPointers || 237 junk->prev_addr() == (void*)(oop(junk)->klass_addr()), 238 "Offset of FreeChunk::_prev within FreeChunk must match" 239 " that of OopDesc::_klass within OopDesc"); 240 ) 241 if (CollectedHeap::use_parallel_gc_threads()) { 242 typedef CMSParGCThreadState* CMSParGCThreadStatePtr; 243 _par_gc_thread_states = 244 NEW_C_HEAP_ARRAY(CMSParGCThreadStatePtr, ParallelGCThreads, mtGC); 245 if (_par_gc_thread_states == NULL) { 246 vm_exit_during_initialization("Could not allocate par gc structs"); 247 } 248 for (uint i = 0; i < ParallelGCThreads; i++) { 249 _par_gc_thread_states[i] = new CMSParGCThreadState(cmsSpace()); 250 if (_par_gc_thread_states[i] == NULL) { 251 vm_exit_during_initialization("Could not allocate par gc structs"); 252 } 253 } 254 } else { 255 _par_gc_thread_states = NULL; 256 } 257 _incremental_collection_failed = false; 258 // The "dilatation_factor" is the expansion that can occur on 259 // account of the fact that the minimum object size in the CMS 260 // generation may be larger than that in, say, a contiguous young 261 // generation. 262 // Ideally, in the calculation below, we'd compute the dilatation 263 // factor as: MinChunkSize/(promoting_gen's min object size) 264 // Since we do not have such a general query interface for the 265 // promoting generation, we'll instead just use the minimum 266 // object size (which today is a header's worth of space); 267 // note that all arithmetic is in units of HeapWords. 268 assert(MinChunkSize >= CollectedHeap::min_fill_size(), "just checking"); 269 assert(_dilatation_factor >= 1.0, "from previous assert"); 270 } 271 272 273 // The field "_initiating_occupancy" represents the occupancy percentage 274 // at which we trigger a new collection cycle. Unless explicitly specified 275 // via CMSInitiatingOccupancyFraction (argument "io" below), it 276 // is calculated by: 277 // 278 // Let "f" be MinHeapFreeRatio in 279 // 280 // _initiating_occupancy = 100-f + 281 // f * (CMSTriggerRatio/100) 282 // where CMSTriggerRatio is the argument "tr" below. 283 // 284 // That is, if we assume the heap is at its desired maximum occupancy at the 285 // end of a collection, we let CMSTriggerRatio of the (purported) free 286 // space be allocated before initiating a new collection cycle. 287 // 288 void ConcurrentMarkSweepGeneration::init_initiating_occupancy(intx io, uintx tr) { 289 assert(io <= 100 && tr <= 100, "Check the arguments"); 290 if (io >= 0) { 291 _initiating_occupancy = (double)io / 100.0; 292 } else { 293 _initiating_occupancy = ((100 - MinHeapFreeRatio) + 294 (double)(tr * MinHeapFreeRatio) / 100.0) 295 / 100.0; 296 } 297 } 298 299 void ConcurrentMarkSweepGeneration::ref_processor_init() { 300 assert(collector() != NULL, "no collector"); 301 collector()->ref_processor_init(); 302 } 303 304 void CMSCollector::ref_processor_init() { 305 if (_ref_processor == NULL) { 306 // Allocate and initialize a reference processor 307 _ref_processor = 308 new ReferenceProcessor(_span, // span 309 (ParallelGCThreads > 1) && ParallelRefProcEnabled, // mt processing 310 (int) ParallelGCThreads, // mt processing degree 311 _cmsGen->refs_discovery_is_mt(), // mt discovery 312 (int) MAX2(ConcGCThreads, ParallelGCThreads), // mt discovery degree 313 _cmsGen->refs_discovery_is_atomic(), // discovery is not atomic 314 &_is_alive_closure); // closure for liveness info 315 // Initialize the _ref_processor field of CMSGen 316 _cmsGen->set_ref_processor(_ref_processor); 317 318 } 319 } 320 321 CMSAdaptiveSizePolicy* CMSCollector::size_policy() { 322 GenCollectedHeap* gch = GenCollectedHeap::heap(); 323 assert(gch->kind() == CollectedHeap::GenCollectedHeap, 324 "Wrong type of heap"); 325 CMSAdaptiveSizePolicy* sp = (CMSAdaptiveSizePolicy*) 326 gch->gen_policy()->size_policy(); 327 assert(sp->is_gc_cms_adaptive_size_policy(), 328 "Wrong type of size policy"); 329 return sp; 330 } 331 332 CMSGCAdaptivePolicyCounters* CMSCollector::gc_adaptive_policy_counters() { 333 CMSGCAdaptivePolicyCounters* results = 334 (CMSGCAdaptivePolicyCounters*) collector_policy()->counters(); 335 assert( 336 results->kind() == GCPolicyCounters::CMSGCAdaptivePolicyCountersKind, 337 "Wrong gc policy counter kind"); 338 return results; 339 } 340 341 342 void ConcurrentMarkSweepGeneration::initialize_performance_counters() { 343 344 const char* gen_name = "old"; 345 346 // Generation Counters - generation 1, 1 subspace 347 _gen_counters = new GenerationCounters(gen_name, 1, 1, &_virtual_space); 348 349 _space_counters = new GSpaceCounters(gen_name, 0, 350 _virtual_space.reserved_size(), 351 this, _gen_counters); 352 } 353 354 CMSStats::CMSStats(ConcurrentMarkSweepGeneration* cms_gen, unsigned int alpha): 355 _cms_gen(cms_gen) 356 { 357 assert(alpha <= 100, "bad value"); 358 _saved_alpha = alpha; 359 360 // Initialize the alphas to the bootstrap value of 100. 361 _gc0_alpha = _cms_alpha = 100; 362 363 _cms_begin_time.update(); 364 _cms_end_time.update(); 365 366 _gc0_duration = 0.0; 367 _gc0_period = 0.0; 368 _gc0_promoted = 0; 369 370 _cms_duration = 0.0; 371 _cms_period = 0.0; 372 _cms_allocated = 0; 373 374 _cms_used_at_gc0_begin = 0; 375 _cms_used_at_gc0_end = 0; 376 _allow_duty_cycle_reduction = false; 377 _valid_bits = 0; 378 _icms_duty_cycle = CMSIncrementalDutyCycle; 379 } 380 381 double CMSStats::cms_free_adjustment_factor(size_t free) const { 382 // TBD: CR 6909490 383 return 1.0; 384 } 385 386 void CMSStats::adjust_cms_free_adjustment_factor(bool fail, size_t free) { 387 } 388 389 // If promotion failure handling is on use 390 // the padded average size of the promotion for each 391 // young generation collection. 392 double CMSStats::time_until_cms_gen_full() const { 393 size_t cms_free = _cms_gen->cmsSpace()->free(); 394 GenCollectedHeap* gch = GenCollectedHeap::heap(); 395 size_t expected_promotion = MIN2(gch->get_gen(0)->capacity(), 396 (size_t) _cms_gen->gc_stats()->avg_promoted()->padded_average()); 397 if (cms_free > expected_promotion) { 398 // Start a cms collection if there isn't enough space to promote 399 // for the next minor collection. Use the padded average as 400 // a safety factor. 401 cms_free -= expected_promotion; 402 403 // Adjust by the safety factor. 404 double cms_free_dbl = (double)cms_free; 405 double cms_adjustment = (100.0 - CMSIncrementalSafetyFactor)/100.0; 406 // Apply a further correction factor which tries to adjust 407 // for recent occurance of concurrent mode failures. 408 cms_adjustment = cms_adjustment * cms_free_adjustment_factor(cms_free); 409 cms_free_dbl = cms_free_dbl * cms_adjustment; 410 411 if (PrintGCDetails && Verbose) { 412 gclog_or_tty->print_cr("CMSStats::time_until_cms_gen_full: cms_free " 413 SIZE_FORMAT " expected_promotion " SIZE_FORMAT, 414 cms_free, expected_promotion); 415 gclog_or_tty->print_cr(" cms_free_dbl %f cms_consumption_rate %f", 416 cms_free_dbl, cms_consumption_rate() + 1.0); 417 } 418 // Add 1 in case the consumption rate goes to zero. 419 return cms_free_dbl / (cms_consumption_rate() + 1.0); 420 } 421 return 0.0; 422 } 423 424 // Compare the duration of the cms collection to the 425 // time remaining before the cms generation is empty. 426 // Note that the time from the start of the cms collection 427 // to the start of the cms sweep (less than the total 428 // duration of the cms collection) can be used. This 429 // has been tried and some applications experienced 430 // promotion failures early in execution. This was 431 // possibly because the averages were not accurate 432 // enough at the beginning. 433 double CMSStats::time_until_cms_start() const { 434 // We add "gc0_period" to the "work" calculation 435 // below because this query is done (mostly) at the 436 // end of a scavenge, so we need to conservatively 437 // account for that much possible delay 438 // in the query so as to avoid concurrent mode failures 439 // due to starting the collection just a wee bit too 440 // late. 441 double work = cms_duration() + gc0_period(); 442 double deadline = time_until_cms_gen_full(); 443 // If a concurrent mode failure occurred recently, we want to be 444 // more conservative and halve our expected time_until_cms_gen_full() 445 if (work > deadline) { 446 if (Verbose && PrintGCDetails) { 447 gclog_or_tty->print( 448 " CMSCollector: collect because of anticipated promotion " 449 "before full %3.7f + %3.7f > %3.7f ", cms_duration(), 450 gc0_period(), time_until_cms_gen_full()); 451 } 452 return 0.0; 453 } 454 return work - deadline; 455 } 456 457 // Return a duty cycle based on old_duty_cycle and new_duty_cycle, limiting the 458 // amount of change to prevent wild oscillation. 459 unsigned int CMSStats::icms_damped_duty_cycle(unsigned int old_duty_cycle, 460 unsigned int new_duty_cycle) { 461 assert(old_duty_cycle <= 100, "bad input value"); 462 assert(new_duty_cycle <= 100, "bad input value"); 463 464 // Note: use subtraction with caution since it may underflow (values are 465 // unsigned). Addition is safe since we're in the range 0-100. 466 unsigned int damped_duty_cycle = new_duty_cycle; 467 if (new_duty_cycle < old_duty_cycle) { 468 const unsigned int largest_delta = MAX2(old_duty_cycle / 4, 5U); 469 if (new_duty_cycle + largest_delta < old_duty_cycle) { 470 damped_duty_cycle = old_duty_cycle - largest_delta; 471 } 472 } else if (new_duty_cycle > old_duty_cycle) { 473 const unsigned int largest_delta = MAX2(old_duty_cycle / 4, 15U); 474 if (new_duty_cycle > old_duty_cycle + largest_delta) { 475 damped_duty_cycle = MIN2(old_duty_cycle + largest_delta, 100U); 476 } 477 } 478 assert(damped_duty_cycle <= 100, "invalid duty cycle computed"); 479 480 if (CMSTraceIncrementalPacing) { 481 gclog_or_tty->print(" [icms_damped_duty_cycle(%d,%d) = %d] ", 482 old_duty_cycle, new_duty_cycle, damped_duty_cycle); 483 } 484 return damped_duty_cycle; 485 } 486 487 unsigned int CMSStats::icms_update_duty_cycle_impl() { 488 assert(CMSIncrementalPacing && valid(), 489 "should be handled in icms_update_duty_cycle()"); 490 491 double cms_time_so_far = cms_timer().seconds(); 492 double scaled_duration = cms_duration_per_mb() * _cms_used_at_gc0_end / M; 493 double scaled_duration_remaining = fabsd(scaled_duration - cms_time_so_far); 494 495 // Avoid division by 0. 496 double time_until_full = MAX2(time_until_cms_gen_full(), 0.01); 497 double duty_cycle_dbl = 100.0 * scaled_duration_remaining / time_until_full; 498 499 unsigned int new_duty_cycle = MIN2((unsigned int)duty_cycle_dbl, 100U); 500 if (new_duty_cycle > _icms_duty_cycle) { 501 // Avoid very small duty cycles (1 or 2); 0 is allowed. 502 if (new_duty_cycle > 2) { 503 _icms_duty_cycle = icms_damped_duty_cycle(_icms_duty_cycle, 504 new_duty_cycle); 505 } 506 } else if (_allow_duty_cycle_reduction) { 507 // The duty cycle is reduced only once per cms cycle (see record_cms_end()). 508 new_duty_cycle = icms_damped_duty_cycle(_icms_duty_cycle, new_duty_cycle); 509 // Respect the minimum duty cycle. 510 unsigned int min_duty_cycle = (unsigned int)CMSIncrementalDutyCycleMin; 511 _icms_duty_cycle = MAX2(new_duty_cycle, min_duty_cycle); 512 } 513 514 if (PrintGCDetails || CMSTraceIncrementalPacing) { 515 gclog_or_tty->print(" icms_dc=%d ", _icms_duty_cycle); 516 } 517 518 _allow_duty_cycle_reduction = false; 519 return _icms_duty_cycle; 520 } 521 522 #ifndef PRODUCT 523 void CMSStats::print_on(outputStream *st) const { 524 st->print(" gc0_alpha=%d,cms_alpha=%d", _gc0_alpha, _cms_alpha); 525 st->print(",gc0_dur=%g,gc0_per=%g,gc0_promo=" SIZE_FORMAT, 526 gc0_duration(), gc0_period(), gc0_promoted()); 527 st->print(",cms_dur=%g,cms_dur_per_mb=%g,cms_per=%g,cms_alloc=" SIZE_FORMAT, 528 cms_duration(), cms_duration_per_mb(), 529 cms_period(), cms_allocated()); 530 st->print(",cms_since_beg=%g,cms_since_end=%g", 531 cms_time_since_begin(), cms_time_since_end()); 532 st->print(",cms_used_beg=" SIZE_FORMAT ",cms_used_end=" SIZE_FORMAT, 533 _cms_used_at_gc0_begin, _cms_used_at_gc0_end); 534 if (CMSIncrementalMode) { 535 st->print(",dc=%d", icms_duty_cycle()); 536 } 537 538 if (valid()) { 539 st->print(",promo_rate=%g,cms_alloc_rate=%g", 540 promotion_rate(), cms_allocation_rate()); 541 st->print(",cms_consumption_rate=%g,time_until_full=%g", 542 cms_consumption_rate(), time_until_cms_gen_full()); 543 } 544 st->print(" "); 545 } 546 #endif // #ifndef PRODUCT 547 548 CMSCollector::CollectorState CMSCollector::_collectorState = 549 CMSCollector::Idling; 550 bool CMSCollector::_foregroundGCIsActive = false; 551 bool CMSCollector::_foregroundGCShouldWait = false; 552 553 CMSCollector::CMSCollector(ConcurrentMarkSweepGeneration* cmsGen, 554 CardTableRS* ct, 555 ConcurrentMarkSweepPolicy* cp): 556 _cmsGen(cmsGen), 557 _ct(ct), 558 _ref_processor(NULL), // will be set later 559 _conc_workers(NULL), // may be set later 560 _abort_preclean(false), 561 _start_sampling(false), 562 _between_prologue_and_epilogue(false), 563 _markBitMap(0, Mutex::leaf + 1, "CMS_markBitMap_lock"), 564 _modUnionTable((CardTableModRefBS::card_shift - LogHeapWordSize), 565 -1 /* lock-free */, "No_lock" /* dummy */), 566 _modUnionClosure(&_modUnionTable), 567 _modUnionClosurePar(&_modUnionTable), 568 // Adjust my span to cover old (cms) gen 569 _span(cmsGen->reserved()), 570 // Construct the is_alive_closure with _span & markBitMap 571 _is_alive_closure(_span, &_markBitMap), 572 _restart_addr(NULL), 573 _overflow_list(NULL), 574 _stats(cmsGen), 575 _eden_chunk_lock(new Mutex(Mutex::leaf + 1, "CMS_eden_chunk_lock", true)), 576 _eden_chunk_array(NULL), // may be set in ctor body 577 _eden_chunk_capacity(0), // -- ditto -- 578 _eden_chunk_index(0), // -- ditto -- 579 _survivor_plab_array(NULL), // -- ditto -- 580 _survivor_chunk_array(NULL), // -- ditto -- 581 _survivor_chunk_capacity(0), // -- ditto -- 582 _survivor_chunk_index(0), // -- ditto -- 583 _ser_pmc_preclean_ovflw(0), 584 _ser_kac_preclean_ovflw(0), 585 _ser_pmc_remark_ovflw(0), 586 _par_pmc_remark_ovflw(0), 587 _ser_kac_ovflw(0), 588 _par_kac_ovflw(0), 589 #ifndef PRODUCT 590 _num_par_pushes(0), 591 #endif 592 _collection_count_start(0), 593 _verifying(false), 594 _icms_start_limit(NULL), 595 _icms_stop_limit(NULL), 596 _verification_mark_bm(0, Mutex::leaf + 1, "CMS_verification_mark_bm_lock"), 597 _completed_initialization(false), 598 _collector_policy(cp), 599 _should_unload_classes(CMSClassUnloadingEnabled), 600 _concurrent_cycles_since_last_unload(0), 601 _roots_scanning_options(SharedHeap::SO_None), 602 _inter_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding), 603 _intra_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding), 604 _gc_tracer_cm(new (ResourceObj::C_HEAP, mtGC) CMSTracer()), 605 _gc_timer_cm(new (ResourceObj::C_HEAP, mtGC) ConcurrentGCTimer()), 606 _cms_start_registered(false) 607 { 608 if (ExplicitGCInvokesConcurrentAndUnloadsClasses) { 609 ExplicitGCInvokesConcurrent = true; 610 } 611 // Now expand the span and allocate the collection support structures 612 // (MUT, marking bit map etc.) to cover both generations subject to 613 // collection. 614 615 // For use by dirty card to oop closures. 616 _cmsGen->cmsSpace()->set_collector(this); 617 618 // Allocate MUT and marking bit map 619 { 620 MutexLockerEx x(_markBitMap.lock(), Mutex::_no_safepoint_check_flag); 621 if (!_markBitMap.allocate(_span)) { 622 warning("Failed to allocate CMS Bit Map"); 623 return; 624 } 625 assert(_markBitMap.covers(_span), "_markBitMap inconsistency?"); 626 } 627 { 628 _modUnionTable.allocate(_span); 629 assert(_modUnionTable.covers(_span), "_modUnionTable inconsistency?"); 630 } 631 632 if (!_markStack.allocate(MarkStackSize)) { 633 warning("Failed to allocate CMS Marking Stack"); 634 return; 635 } 636 637 // Support for multi-threaded concurrent phases 638 if (CMSConcurrentMTEnabled) { 639 if (FLAG_IS_DEFAULT(ConcGCThreads)) { 640 // just for now 641 FLAG_SET_DEFAULT(ConcGCThreads, (ParallelGCThreads + 3)/4); 642 } 643 if (ConcGCThreads > 1) { 644 _conc_workers = new YieldingFlexibleWorkGang("Parallel CMS Threads", 645 ConcGCThreads, true); 646 if (_conc_workers == NULL) { 647 warning("GC/CMS: _conc_workers allocation failure: " 648 "forcing -CMSConcurrentMTEnabled"); 649 CMSConcurrentMTEnabled = false; 650 } else { 651 _conc_workers->initialize_workers(); 652 } 653 } else { 654 CMSConcurrentMTEnabled = false; 655 } 656 } 657 if (!CMSConcurrentMTEnabled) { 658 ConcGCThreads = 0; 659 } else { 660 // Turn off CMSCleanOnEnter optimization temporarily for 661 // the MT case where it's not fixed yet; see 6178663. 662 CMSCleanOnEnter = false; 663 } 664 assert((_conc_workers != NULL) == (ConcGCThreads > 1), 665 "Inconsistency"); 666 667 // Parallel task queues; these are shared for the 668 // concurrent and stop-world phases of CMS, but 669 // are not shared with parallel scavenge (ParNew). 670 { 671 uint i; 672 uint num_queues = (uint) MAX2(ParallelGCThreads, ConcGCThreads); 673 674 if ((CMSParallelRemarkEnabled || CMSConcurrentMTEnabled 675 || ParallelRefProcEnabled) 676 && num_queues > 0) { 677 _task_queues = new OopTaskQueueSet(num_queues); 678 if (_task_queues == NULL) { 679 warning("task_queues allocation failure."); 680 return; 681 } 682 _hash_seed = NEW_C_HEAP_ARRAY(int, num_queues, mtGC); 683 if (_hash_seed == NULL) { 684 warning("_hash_seed array allocation failure"); 685 return; 686 } 687 688 typedef Padded<OopTaskQueue> PaddedOopTaskQueue; 689 for (i = 0; i < num_queues; i++) { 690 PaddedOopTaskQueue *q = new PaddedOopTaskQueue(); 691 if (q == NULL) { 692 warning("work_queue allocation failure."); 693 return; 694 } 695 _task_queues->register_queue(i, q); 696 } 697 for (i = 0; i < num_queues; i++) { 698 _task_queues->queue(i)->initialize(); 699 _hash_seed[i] = 17; // copied from ParNew 700 } 701 } 702 } 703 704 _cmsGen ->init_initiating_occupancy(CMSInitiatingOccupancyFraction, CMSTriggerRatio); 705 706 // Clip CMSBootstrapOccupancy between 0 and 100. 707 _bootstrap_occupancy = ((double)CMSBootstrapOccupancy)/(double)100; 708 709 _full_gcs_since_conc_gc = 0; 710 711 // Now tell CMS generations the identity of their collector 712 ConcurrentMarkSweepGeneration::set_collector(this); 713 714 // Create & start a CMS thread for this CMS collector 715 _cmsThread = ConcurrentMarkSweepThread::start(this); 716 assert(cmsThread() != NULL, "CMS Thread should have been created"); 717 assert(cmsThread()->collector() == this, 718 "CMS Thread should refer to this gen"); 719 assert(CGC_lock != NULL, "Where's the CGC_lock?"); 720 721 // Support for parallelizing young gen rescan 722 GenCollectedHeap* gch = GenCollectedHeap::heap(); 723 _young_gen = gch->prev_gen(_cmsGen); 724 if (gch->supports_inline_contig_alloc()) { 725 _top_addr = gch->top_addr(); 726 _end_addr = gch->end_addr(); 727 assert(_young_gen != NULL, "no _young_gen"); 728 _eden_chunk_index = 0; 729 _eden_chunk_capacity = (_young_gen->max_capacity()+CMSSamplingGrain)/CMSSamplingGrain; 730 _eden_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, _eden_chunk_capacity, mtGC); 731 if (_eden_chunk_array == NULL) { 732 _eden_chunk_capacity = 0; 733 warning("GC/CMS: _eden_chunk_array allocation failure"); 734 } 735 } 736 assert(_eden_chunk_array != NULL || _eden_chunk_capacity == 0, "Error"); 737 738 // Support for parallelizing survivor space rescan 739 if ((CMSParallelRemarkEnabled && CMSParallelSurvivorRemarkEnabled) || CMSParallelInitialMarkEnabled) { 740 const size_t max_plab_samples = 741 ((DefNewGeneration*)_young_gen)->max_survivor_size()/MinTLABSize; 742 743 _survivor_plab_array = NEW_C_HEAP_ARRAY(ChunkArray, ParallelGCThreads, mtGC); 744 _survivor_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, 2*max_plab_samples, mtGC); 745 _cursor = NEW_C_HEAP_ARRAY(size_t, ParallelGCThreads, mtGC); 746 if (_survivor_plab_array == NULL || _survivor_chunk_array == NULL 747 || _cursor == NULL) { 748 warning("Failed to allocate survivor plab/chunk array"); 749 if (_survivor_plab_array != NULL) { 750 FREE_C_HEAP_ARRAY(ChunkArray, _survivor_plab_array, mtGC); 751 _survivor_plab_array = NULL; 752 } 753 if (_survivor_chunk_array != NULL) { 754 FREE_C_HEAP_ARRAY(HeapWord*, _survivor_chunk_array, mtGC); 755 _survivor_chunk_array = NULL; 756 } 757 if (_cursor != NULL) { 758 FREE_C_HEAP_ARRAY(size_t, _cursor, mtGC); 759 _cursor = NULL; 760 } 761 } else { 762 _survivor_chunk_capacity = 2*max_plab_samples; 763 for (uint i = 0; i < ParallelGCThreads; i++) { 764 HeapWord** vec = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples, mtGC); 765 if (vec == NULL) { 766 warning("Failed to allocate survivor plab array"); 767 for (int j = i; j > 0; j--) { 768 FREE_C_HEAP_ARRAY(HeapWord*, _survivor_plab_array[j-1].array(), mtGC); 769 } 770 FREE_C_HEAP_ARRAY(ChunkArray, _survivor_plab_array, mtGC); 771 FREE_C_HEAP_ARRAY(HeapWord*, _survivor_chunk_array, mtGC); 772 _survivor_plab_array = NULL; 773 _survivor_chunk_array = NULL; 774 _survivor_chunk_capacity = 0; 775 break; 776 } else { 777 ChunkArray* cur = 778 ::new (&_survivor_plab_array[i]) ChunkArray(vec, 779 max_plab_samples); 780 assert(cur->end() == 0, "Should be 0"); 781 assert(cur->array() == vec, "Should be vec"); 782 assert(cur->capacity() == max_plab_samples, "Error"); 783 } 784 } 785 } 786 } 787 assert( ( _survivor_plab_array != NULL 788 && _survivor_chunk_array != NULL) 789 || ( _survivor_chunk_capacity == 0 790 && _survivor_chunk_index == 0), 791 "Error"); 792 793 NOT_PRODUCT(_overflow_counter = CMSMarkStackOverflowInterval;) 794 _gc_counters = new CollectorCounters("CMS", 1); 795 _completed_initialization = true; 796 _inter_sweep_timer.start(); // start of time 797 } 798 799 const char* ConcurrentMarkSweepGeneration::name() const { 800 return "concurrent mark-sweep generation"; 801 } 802 void ConcurrentMarkSweepGeneration::update_counters() { 803 if (UsePerfData) { 804 _space_counters->update_all(); 805 _gen_counters->update_all(); 806 } 807 } 808 809 // this is an optimized version of update_counters(). it takes the 810 // used value as a parameter rather than computing it. 811 // 812 void ConcurrentMarkSweepGeneration::update_counters(size_t used) { 813 if (UsePerfData) { 814 _space_counters->update_used(used); 815 _space_counters->update_capacity(); 816 _gen_counters->update_all(); 817 } 818 } 819 820 void ConcurrentMarkSweepGeneration::print() const { 821 Generation::print(); 822 cmsSpace()->print(); 823 } 824 825 #ifndef PRODUCT 826 void ConcurrentMarkSweepGeneration::print_statistics() { 827 cmsSpace()->printFLCensus(0); 828 } 829 #endif 830 831 void ConcurrentMarkSweepGeneration::printOccupancy(const char *s) { 832 GenCollectedHeap* gch = GenCollectedHeap::heap(); 833 if (PrintGCDetails) { 834 if (Verbose) { 835 gclog_or_tty->print("[%d %s-%s: "SIZE_FORMAT"("SIZE_FORMAT")]", 836 level(), short_name(), s, used(), capacity()); 837 } else { 838 gclog_or_tty->print("[%d %s-%s: "SIZE_FORMAT"K("SIZE_FORMAT"K)]", 839 level(), short_name(), s, used() / K, capacity() / K); 840 } 841 } 842 if (Verbose) { 843 gclog_or_tty->print(" "SIZE_FORMAT"("SIZE_FORMAT")", 844 gch->used(), gch->capacity()); 845 } else { 846 gclog_or_tty->print(" "SIZE_FORMAT"K("SIZE_FORMAT"K)", 847 gch->used() / K, gch->capacity() / K); 848 } 849 } 850 851 size_t 852 ConcurrentMarkSweepGeneration::contiguous_available() const { 853 // dld proposes an improvement in precision here. If the committed 854 // part of the space ends in a free block we should add that to 855 // uncommitted size in the calculation below. Will make this 856 // change later, staying with the approximation below for the 857 // time being. -- ysr. 858 return MAX2(_virtual_space.uncommitted_size(), unsafe_max_alloc_nogc()); 859 } 860 861 size_t 862 ConcurrentMarkSweepGeneration::unsafe_max_alloc_nogc() const { 863 return _cmsSpace->max_alloc_in_words() * HeapWordSize; 864 } 865 866 size_t ConcurrentMarkSweepGeneration::max_available() const { 867 return free() + _virtual_space.uncommitted_size(); 868 } 869 870 bool ConcurrentMarkSweepGeneration::promotion_attempt_is_safe(size_t max_promotion_in_bytes) const { 871 size_t available = max_available(); 872 size_t av_promo = (size_t)gc_stats()->avg_promoted()->padded_average(); 873 bool res = (available >= av_promo) || (available >= max_promotion_in_bytes); 874 if (Verbose && PrintGCDetails) { 875 gclog_or_tty->print_cr( 876 "CMS: promo attempt is%s safe: available("SIZE_FORMAT") %s av_promo("SIZE_FORMAT")," 877 "max_promo("SIZE_FORMAT")", 878 res? "":" not", available, res? ">=":"<", 879 av_promo, max_promotion_in_bytes); 880 } 881 return res; 882 } 883 884 // At a promotion failure dump information on block layout in heap 885 // (cms old generation). 886 void ConcurrentMarkSweepGeneration::promotion_failure_occurred() { 887 if (CMSDumpAtPromotionFailure) { 888 cmsSpace()->dump_at_safepoint_with_locks(collector(), gclog_or_tty); 889 } 890 } 891 892 CompactibleSpace* 893 ConcurrentMarkSweepGeneration::first_compaction_space() const { 894 return _cmsSpace; 895 } 896 897 void ConcurrentMarkSweepGeneration::reset_after_compaction() { 898 // Clear the promotion information. These pointers can be adjusted 899 // along with all the other pointers into the heap but 900 // compaction is expected to be a rare event with 901 // a heap using cms so don't do it without seeing the need. 902 if (CollectedHeap::use_parallel_gc_threads()) { 903 for (uint i = 0; i < ParallelGCThreads; i++) { 904 _par_gc_thread_states[i]->promo.reset(); 905 } 906 } 907 } 908 909 void ConcurrentMarkSweepGeneration::space_iterate(SpaceClosure* blk, bool usedOnly) { 910 blk->do_space(_cmsSpace); 911 } 912 913 void ConcurrentMarkSweepGeneration::compute_new_size() { 914 assert_locked_or_safepoint(Heap_lock); 915 916 // If incremental collection failed, we just want to expand 917 // to the limit. 918 if (incremental_collection_failed()) { 919 clear_incremental_collection_failed(); 920 grow_to_reserved(); 921 return; 922 } 923 924 // The heap has been compacted but not reset yet. 925 // Any metric such as free() or used() will be incorrect. 926 927 CardGeneration::compute_new_size(); 928 929 // Reset again after a possible resizing 930 if (did_compact()) { 931 cmsSpace()->reset_after_compaction(); 932 } 933 } 934 935 void ConcurrentMarkSweepGeneration::compute_new_size_free_list() { 936 assert_locked_or_safepoint(Heap_lock); 937 938 // If incremental collection failed, we just want to expand 939 // to the limit. 940 if (incremental_collection_failed()) { 941 clear_incremental_collection_failed(); 942 grow_to_reserved(); 943 return; 944 } 945 946 double free_percentage = ((double) free()) / capacity(); 947 double desired_free_percentage = (double) MinHeapFreeRatio / 100; 948 double maximum_free_percentage = (double) MaxHeapFreeRatio / 100; 949 950 // compute expansion delta needed for reaching desired free percentage 951 if (free_percentage < desired_free_percentage) { 952 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 953 assert(desired_capacity >= capacity(), "invalid expansion size"); 954 size_t expand_bytes = MAX2(desired_capacity - capacity(), MinHeapDeltaBytes); 955 if (PrintGCDetails && Verbose) { 956 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 957 gclog_or_tty->print_cr("\nFrom compute_new_size: "); 958 gclog_or_tty->print_cr(" Free fraction %f", free_percentage); 959 gclog_or_tty->print_cr(" Desired free fraction %f", 960 desired_free_percentage); 961 gclog_or_tty->print_cr(" Maximum free fraction %f", 962 maximum_free_percentage); 963 gclog_or_tty->print_cr(" Capacity "SIZE_FORMAT, capacity()/1000); 964 gclog_or_tty->print_cr(" Desired capacity "SIZE_FORMAT, 965 desired_capacity/1000); 966 int prev_level = level() - 1; 967 if (prev_level >= 0) { 968 size_t prev_size = 0; 969 GenCollectedHeap* gch = GenCollectedHeap::heap(); 970 Generation* prev_gen = gch->_gens[prev_level]; 971 prev_size = prev_gen->capacity(); 972 gclog_or_tty->print_cr(" Younger gen size "SIZE_FORMAT, 973 prev_size/1000); 974 } 975 gclog_or_tty->print_cr(" unsafe_max_alloc_nogc "SIZE_FORMAT, 976 unsafe_max_alloc_nogc()/1000); 977 gclog_or_tty->print_cr(" contiguous available "SIZE_FORMAT, 978 contiguous_available()/1000); 979 gclog_or_tty->print_cr(" Expand by "SIZE_FORMAT" (bytes)", 980 expand_bytes); 981 } 982 // safe if expansion fails 983 expand(expand_bytes, 0, CMSExpansionCause::_satisfy_free_ratio); 984 if (PrintGCDetails && Verbose) { 985 gclog_or_tty->print_cr(" Expanded free fraction %f", 986 ((double) free()) / capacity()); 987 } 988 } else { 989 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 990 assert(desired_capacity <= capacity(), "invalid expansion size"); 991 size_t shrink_bytes = capacity() - desired_capacity; 992 // Don't shrink unless the delta is greater than the minimum shrink we want 993 if (shrink_bytes >= MinHeapDeltaBytes) { 994 shrink_free_list_by(shrink_bytes); 995 } 996 } 997 } 998 999 Mutex* ConcurrentMarkSweepGeneration::freelistLock() const { 1000 return cmsSpace()->freelistLock(); 1001 } 1002 1003 HeapWord* ConcurrentMarkSweepGeneration::allocate(size_t size, 1004 bool tlab) { 1005 CMSSynchronousYieldRequest yr; 1006 MutexLockerEx x(freelistLock(), 1007 Mutex::_no_safepoint_check_flag); 1008 return have_lock_and_allocate(size, tlab); 1009 } 1010 1011 HeapWord* ConcurrentMarkSweepGeneration::have_lock_and_allocate(size_t size, 1012 bool tlab /* ignored */) { 1013 assert_lock_strong(freelistLock()); 1014 size_t adjustedSize = CompactibleFreeListSpace::adjustObjectSize(size); 1015 HeapWord* res = cmsSpace()->allocate(adjustedSize); 1016 // Allocate the object live (grey) if the background collector has 1017 // started marking. This is necessary because the marker may 1018 // have passed this address and consequently this object will 1019 // not otherwise be greyed and would be incorrectly swept up. 1020 // Note that if this object contains references, the writing 1021 // of those references will dirty the card containing this object 1022 // allowing the object to be blackened (and its references scanned) 1023 // either during a preclean phase or at the final checkpoint. 1024 if (res != NULL) { 1025 // We may block here with an uninitialized object with 1026 // its mark-bit or P-bits not yet set. Such objects need 1027 // to be safely navigable by block_start(). 1028 assert(oop(res)->klass_or_null() == NULL, "Object should be uninitialized here."); 1029 assert(!((FreeChunk*)res)->is_free(), "Error, block will look free but show wrong size"); 1030 collector()->direct_allocated(res, adjustedSize); 1031 _direct_allocated_words += adjustedSize; 1032 // allocation counters 1033 NOT_PRODUCT( 1034 _numObjectsAllocated++; 1035 _numWordsAllocated += (int)adjustedSize; 1036 ) 1037 } 1038 return res; 1039 } 1040 1041 // In the case of direct allocation by mutators in a generation that 1042 // is being concurrently collected, the object must be allocated 1043 // live (grey) if the background collector has started marking. 1044 // This is necessary because the marker may 1045 // have passed this address and consequently this object will 1046 // not otherwise be greyed and would be incorrectly swept up. 1047 // Note that if this object contains references, the writing 1048 // of those references will dirty the card containing this object 1049 // allowing the object to be blackened (and its references scanned) 1050 // either during a preclean phase or at the final checkpoint. 1051 void CMSCollector::direct_allocated(HeapWord* start, size_t size) { 1052 assert(_markBitMap.covers(start, size), "Out of bounds"); 1053 if (_collectorState >= Marking) { 1054 MutexLockerEx y(_markBitMap.lock(), 1055 Mutex::_no_safepoint_check_flag); 1056 // [see comments preceding SweepClosure::do_blk() below for details] 1057 // 1058 // Can the P-bits be deleted now? JJJ 1059 // 1060 // 1. need to mark the object as live so it isn't collected 1061 // 2. need to mark the 2nd bit to indicate the object may be uninitialized 1062 // 3. need to mark the end of the object so marking, precleaning or sweeping 1063 // can skip over uninitialized or unparsable objects. An allocated 1064 // object is considered uninitialized for our purposes as long as 1065 // its klass word is NULL. All old gen objects are parsable 1066 // as soon as they are initialized.) 1067 _markBitMap.mark(start); // object is live 1068 _markBitMap.mark(start + 1); // object is potentially uninitialized? 1069 _markBitMap.mark(start + size - 1); 1070 // mark end of object 1071 } 1072 // check that oop looks uninitialized 1073 assert(oop(start)->klass_or_null() == NULL, "_klass should be NULL"); 1074 } 1075 1076 void CMSCollector::promoted(bool par, HeapWord* start, 1077 bool is_obj_array, size_t obj_size) { 1078 assert(_markBitMap.covers(start), "Out of bounds"); 1079 // See comment in direct_allocated() about when objects should 1080 // be allocated live. 1081 if (_collectorState >= Marking) { 1082 // we already hold the marking bit map lock, taken in 1083 // the prologue 1084 if (par) { 1085 _markBitMap.par_mark(start); 1086 } else { 1087 _markBitMap.mark(start); 1088 } 1089 // We don't need to mark the object as uninitialized (as 1090 // in direct_allocated above) because this is being done with the 1091 // world stopped and the object will be initialized by the 1092 // time the marking, precleaning or sweeping get to look at it. 1093 // But see the code for copying objects into the CMS generation, 1094 // where we need to ensure that concurrent readers of the 1095 // block offset table are able to safely navigate a block that 1096 // is in flux from being free to being allocated (and in 1097 // transition while being copied into) and subsequently 1098 // becoming a bona-fide object when the copy/promotion is complete. 1099 assert(SafepointSynchronize::is_at_safepoint(), 1100 "expect promotion only at safepoints"); 1101 1102 if (_collectorState < Sweeping) { 1103 // Mark the appropriate cards in the modUnionTable, so that 1104 // this object gets scanned before the sweep. If this is 1105 // not done, CMS generation references in the object might 1106 // not get marked. 1107 // For the case of arrays, which are otherwise precisely 1108 // marked, we need to dirty the entire array, not just its head. 1109 if (is_obj_array) { 1110 // The [par_]mark_range() method expects mr.end() below to 1111 // be aligned to the granularity of a bit's representation 1112 // in the heap. In the case of the MUT below, that's a 1113 // card size. 1114 MemRegion mr(start, 1115 (HeapWord*)round_to((intptr_t)(start + obj_size), 1116 CardTableModRefBS::card_size /* bytes */)); 1117 if (par) { 1118 _modUnionTable.par_mark_range(mr); 1119 } else { 1120 _modUnionTable.mark_range(mr); 1121 } 1122 } else { // not an obj array; we can just mark the head 1123 if (par) { 1124 _modUnionTable.par_mark(start); 1125 } else { 1126 _modUnionTable.mark(start); 1127 } 1128 } 1129 } 1130 } 1131 } 1132 1133 static inline size_t percent_of_space(Space* space, HeapWord* addr) 1134 { 1135 size_t delta = pointer_delta(addr, space->bottom()); 1136 return (size_t)(delta * 100.0 / (space->capacity() / HeapWordSize)); 1137 } 1138 1139 void CMSCollector::icms_update_allocation_limits() 1140 { 1141 Generation* gen0 = GenCollectedHeap::heap()->get_gen(0); 1142 EdenSpace* eden = gen0->as_DefNewGeneration()->eden(); 1143 1144 const unsigned int duty_cycle = stats().icms_update_duty_cycle(); 1145 if (CMSTraceIncrementalPacing) { 1146 stats().print(); 1147 } 1148 1149 assert(duty_cycle <= 100, "invalid duty cycle"); 1150 if (duty_cycle != 0) { 1151 // The duty_cycle is a percentage between 0 and 100; convert to words and 1152 // then compute the offset from the endpoints of the space. 1153 size_t free_words = eden->free() / HeapWordSize; 1154 double free_words_dbl = (double)free_words; 1155 size_t duty_cycle_words = (size_t)(free_words_dbl * duty_cycle / 100.0); 1156 size_t offset_words = (free_words - duty_cycle_words) / 2; 1157 1158 _icms_start_limit = eden->top() + offset_words; 1159 _icms_stop_limit = eden->end() - offset_words; 1160 1161 // The limits may be adjusted (shifted to the right) by 1162 // CMSIncrementalOffset, to allow the application more mutator time after a 1163 // young gen gc (when all mutators were stopped) and before CMS starts and 1164 // takes away one or more cpus. 1165 if (CMSIncrementalOffset != 0) { 1166 double adjustment_dbl = free_words_dbl * CMSIncrementalOffset / 100.0; 1167 size_t adjustment = (size_t)adjustment_dbl; 1168 HeapWord* tmp_stop = _icms_stop_limit + adjustment; 1169 if (tmp_stop > _icms_stop_limit && tmp_stop < eden->end()) { 1170 _icms_start_limit += adjustment; 1171 _icms_stop_limit = tmp_stop; 1172 } 1173 } 1174 } 1175 if (duty_cycle == 0 || (_icms_start_limit == _icms_stop_limit)) { 1176 _icms_start_limit = _icms_stop_limit = eden->end(); 1177 } 1178 1179 // Install the new start limit. 1180 eden->set_soft_end(_icms_start_limit); 1181 1182 if (CMSTraceIncrementalMode) { 1183 gclog_or_tty->print(" icms alloc limits: " 1184 PTR_FORMAT "," PTR_FORMAT 1185 " (" SIZE_FORMAT "%%," SIZE_FORMAT "%%) ", 1186 p2i(_icms_start_limit), p2i(_icms_stop_limit), 1187 percent_of_space(eden, _icms_start_limit), 1188 percent_of_space(eden, _icms_stop_limit)); 1189 if (Verbose) { 1190 gclog_or_tty->print("eden: "); 1191 eden->print_on(gclog_or_tty); 1192 } 1193 } 1194 } 1195 1196 // Any changes here should try to maintain the invariant 1197 // that if this method is called with _icms_start_limit 1198 // and _icms_stop_limit both NULL, then it should return NULL 1199 // and not notify the icms thread. 1200 HeapWord* 1201 CMSCollector::allocation_limit_reached(Space* space, HeapWord* top, 1202 size_t word_size) 1203 { 1204 // A start_limit equal to end() means the duty cycle is 0, so treat that as a 1205 // nop. 1206 if (CMSIncrementalMode && _icms_start_limit != space->end()) { 1207 if (top <= _icms_start_limit) { 1208 if (CMSTraceIncrementalMode) { 1209 space->print_on(gclog_or_tty); 1210 gclog_or_tty->stamp(); 1211 gclog_or_tty->print_cr(" start limit top=" PTR_FORMAT 1212 ", new limit=" PTR_FORMAT 1213 " (" SIZE_FORMAT "%%)", 1214 p2i(top), p2i(_icms_stop_limit), 1215 percent_of_space(space, _icms_stop_limit)); 1216 } 1217 ConcurrentMarkSweepThread::start_icms(); 1218 assert(top < _icms_stop_limit, "Tautology"); 1219 if (word_size < pointer_delta(_icms_stop_limit, top)) { 1220 return _icms_stop_limit; 1221 } 1222 1223 // The allocation will cross both the _start and _stop limits, so do the 1224 // stop notification also and return end(). 1225 if (CMSTraceIncrementalMode) { 1226 space->print_on(gclog_or_tty); 1227 gclog_or_tty->stamp(); 1228 gclog_or_tty->print_cr(" +stop limit top=" PTR_FORMAT 1229 ", new limit=" PTR_FORMAT 1230 " (" SIZE_FORMAT "%%)", 1231 p2i(top), p2i(space->end()), 1232 percent_of_space(space, space->end())); 1233 } 1234 ConcurrentMarkSweepThread::stop_icms(); 1235 return space->end(); 1236 } 1237 1238 if (top <= _icms_stop_limit) { 1239 if (CMSTraceIncrementalMode) { 1240 space->print_on(gclog_or_tty); 1241 gclog_or_tty->stamp(); 1242 gclog_or_tty->print_cr(" stop limit top=" PTR_FORMAT 1243 ", new limit=" PTR_FORMAT 1244 " (" SIZE_FORMAT "%%)", 1245 top, space->end(), 1246 percent_of_space(space, space->end())); 1247 } 1248 ConcurrentMarkSweepThread::stop_icms(); 1249 return space->end(); 1250 } 1251 1252 if (CMSTraceIncrementalMode) { 1253 space->print_on(gclog_or_tty); 1254 gclog_or_tty->stamp(); 1255 gclog_or_tty->print_cr(" end limit top=" PTR_FORMAT 1256 ", new limit=" PTR_FORMAT, 1257 top, NULL); 1258 } 1259 } 1260 1261 return NULL; 1262 } 1263 1264 oop ConcurrentMarkSweepGeneration::promote(oop obj, size_t obj_size) { 1265 assert(obj_size == (size_t)obj->size(), "bad obj_size passed in"); 1266 // allocate, copy and if necessary update promoinfo -- 1267 // delegate to underlying space. 1268 assert_lock_strong(freelistLock()); 1269 1270 #ifndef PRODUCT 1271 if (Universe::heap()->promotion_should_fail()) { 1272 return NULL; 1273 } 1274 #endif // #ifndef PRODUCT 1275 1276 oop res = _cmsSpace->promote(obj, obj_size); 1277 if (res == NULL) { 1278 // expand and retry 1279 size_t s = _cmsSpace->expansionSpaceRequired(obj_size); // HeapWords 1280 expand(s*HeapWordSize, MinHeapDeltaBytes, 1281 CMSExpansionCause::_satisfy_promotion); 1282 // Since there's currently no next generation, we don't try to promote 1283 // into a more senior generation. 1284 assert(next_gen() == NULL, "assumption, based upon which no attempt " 1285 "is made to pass on a possibly failing " 1286 "promotion to next generation"); 1287 res = _cmsSpace->promote(obj, obj_size); 1288 } 1289 if (res != NULL) { 1290 // See comment in allocate() about when objects should 1291 // be allocated live. 1292 assert(obj->is_oop(), "Will dereference klass pointer below"); 1293 collector()->promoted(false, // Not parallel 1294 (HeapWord*)res, obj->is_objArray(), obj_size); 1295 // promotion counters 1296 NOT_PRODUCT( 1297 _numObjectsPromoted++; 1298 _numWordsPromoted += 1299 (int)(CompactibleFreeListSpace::adjustObjectSize(obj->size())); 1300 ) 1301 } 1302 return res; 1303 } 1304 1305 1306 HeapWord* 1307 ConcurrentMarkSweepGeneration::allocation_limit_reached(Space* space, 1308 HeapWord* top, 1309 size_t word_sz) 1310 { 1311 return collector()->allocation_limit_reached(space, top, word_sz); 1312 } 1313 1314 // IMPORTANT: Notes on object size recognition in CMS. 1315 // --------------------------------------------------- 1316 // A block of storage in the CMS generation is always in 1317 // one of three states. A free block (FREE), an allocated 1318 // object (OBJECT) whose size() method reports the correct size, 1319 // and an intermediate state (TRANSIENT) in which its size cannot 1320 // be accurately determined. 1321 // STATE IDENTIFICATION: (32 bit and 64 bit w/o COOPS) 1322 // ----------------------------------------------------- 1323 // FREE: klass_word & 1 == 1; mark_word holds block size 1324 // 1325 // OBJECT: klass_word installed; klass_word != 0 && klass_word & 1 == 0; 1326 // obj->size() computes correct size 1327 // 1328 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT 1329 // 1330 // STATE IDENTIFICATION: (64 bit+COOPS) 1331 // ------------------------------------ 1332 // FREE: mark_word & CMS_FREE_BIT == 1; mark_word & ~CMS_FREE_BIT gives block_size 1333 // 1334 // OBJECT: klass_word installed; klass_word != 0; 1335 // obj->size() computes correct size 1336 // 1337 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT 1338 // 1339 // 1340 // STATE TRANSITION DIAGRAM 1341 // 1342 // mut / parnew mut / parnew 1343 // FREE --------------------> TRANSIENT ---------------------> OBJECT --| 1344 // ^ | 1345 // |------------------------ DEAD <------------------------------------| 1346 // sweep mut 1347 // 1348 // While a block is in TRANSIENT state its size cannot be determined 1349 // so readers will either need to come back later or stall until 1350 // the size can be determined. Note that for the case of direct 1351 // allocation, P-bits, when available, may be used to determine the 1352 // size of an object that may not yet have been initialized. 1353 1354 // Things to support parallel young-gen collection. 1355 oop 1356 ConcurrentMarkSweepGeneration::par_promote(int thread_num, 1357 oop old, markOop m, 1358 size_t word_sz) { 1359 #ifndef PRODUCT 1360 if (Universe::heap()->promotion_should_fail()) { 1361 return NULL; 1362 } 1363 #endif // #ifndef PRODUCT 1364 1365 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1366 PromotionInfo* promoInfo = &ps->promo; 1367 // if we are tracking promotions, then first ensure space for 1368 // promotion (including spooling space for saving header if necessary). 1369 // then allocate and copy, then track promoted info if needed. 1370 // When tracking (see PromotionInfo::track()), the mark word may 1371 // be displaced and in this case restoration of the mark word 1372 // occurs in the (oop_since_save_marks_)iterate phase. 1373 if (promoInfo->tracking() && !promoInfo->ensure_spooling_space()) { 1374 // Out of space for allocating spooling buffers; 1375 // try expanding and allocating spooling buffers. 1376 if (!expand_and_ensure_spooling_space(promoInfo)) { 1377 return NULL; 1378 } 1379 } 1380 assert(promoInfo->has_spooling_space(), "Control point invariant"); 1381 const size_t alloc_sz = CompactibleFreeListSpace::adjustObjectSize(word_sz); 1382 HeapWord* obj_ptr = ps->lab.alloc(alloc_sz); 1383 if (obj_ptr == NULL) { 1384 obj_ptr = expand_and_par_lab_allocate(ps, alloc_sz); 1385 if (obj_ptr == NULL) { 1386 return NULL; 1387 } 1388 } 1389 oop obj = oop(obj_ptr); 1390 OrderAccess::storestore(); 1391 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1392 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1393 // IMPORTANT: See note on object initialization for CMS above. 1394 // Otherwise, copy the object. Here we must be careful to insert the 1395 // klass pointer last, since this marks the block as an allocated object. 1396 // Except with compressed oops it's the mark word. 1397 HeapWord* old_ptr = (HeapWord*)old; 1398 // Restore the mark word copied above. 1399 obj->set_mark(m); 1400 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1401 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1402 OrderAccess::storestore(); 1403 1404 if (UseCompressedClassPointers) { 1405 // Copy gap missed by (aligned) header size calculation below 1406 obj->set_klass_gap(old->klass_gap()); 1407 } 1408 if (word_sz > (size_t)oopDesc::header_size()) { 1409 Copy::aligned_disjoint_words(old_ptr + oopDesc::header_size(), 1410 obj_ptr + oopDesc::header_size(), 1411 word_sz - oopDesc::header_size()); 1412 } 1413 1414 // Now we can track the promoted object, if necessary. We take care 1415 // to delay the transition from uninitialized to full object 1416 // (i.e., insertion of klass pointer) until after, so that it 1417 // atomically becomes a promoted object. 1418 if (promoInfo->tracking()) { 1419 promoInfo->track((PromotedObject*)obj, old->klass()); 1420 } 1421 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1422 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1423 assert(old->is_oop(), "Will use and dereference old klass ptr below"); 1424 1425 // Finally, install the klass pointer (this should be volatile). 1426 OrderAccess::storestore(); 1427 obj->set_klass(old->klass()); 1428 // We should now be able to calculate the right size for this object 1429 assert(obj->is_oop() && obj->size() == (int)word_sz, "Error, incorrect size computed for promoted object"); 1430 1431 collector()->promoted(true, // parallel 1432 obj_ptr, old->is_objArray(), word_sz); 1433 1434 NOT_PRODUCT( 1435 Atomic::inc_ptr(&_numObjectsPromoted); 1436 Atomic::add_ptr(alloc_sz, &_numWordsPromoted); 1437 ) 1438 1439 return obj; 1440 } 1441 1442 void 1443 ConcurrentMarkSweepGeneration:: 1444 par_promote_alloc_undo(int thread_num, 1445 HeapWord* obj, size_t word_sz) { 1446 // CMS does not support promotion undo. 1447 ShouldNotReachHere(); 1448 } 1449 1450 void 1451 ConcurrentMarkSweepGeneration:: 1452 par_promote_alloc_done(int thread_num) { 1453 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1454 ps->lab.retire(thread_num); 1455 } 1456 1457 void 1458 ConcurrentMarkSweepGeneration:: 1459 par_oop_since_save_marks_iterate_done(int thread_num) { 1460 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1461 ParScanWithoutBarrierClosure* dummy_cl = NULL; 1462 ps->promo.promoted_oops_iterate_nv(dummy_cl); 1463 } 1464 1465 bool ConcurrentMarkSweepGeneration::should_collect(bool full, 1466 size_t size, 1467 bool tlab) 1468 { 1469 // We allow a STW collection only if a full 1470 // collection was requested. 1471 return full || should_allocate(size, tlab); // FIX ME !!! 1472 // This and promotion failure handling are connected at the 1473 // hip and should be fixed by untying them. 1474 } 1475 1476 bool CMSCollector::shouldConcurrentCollect() { 1477 if (_full_gc_requested) { 1478 if (Verbose && PrintGCDetails) { 1479 gclog_or_tty->print_cr("CMSCollector: collect because of explicit " 1480 " gc request (or gc_locker)"); 1481 } 1482 return true; 1483 } 1484 1485 // For debugging purposes, change the type of collection. 1486 // If the rotation is not on the concurrent collection 1487 // type, don't start a concurrent collection. 1488 NOT_PRODUCT( 1489 if (RotateCMSCollectionTypes && 1490 (_cmsGen->debug_collection_type() != 1491 ConcurrentMarkSweepGeneration::Concurrent_collection_type)) { 1492 assert(_cmsGen->debug_collection_type() != 1493 ConcurrentMarkSweepGeneration::Unknown_collection_type, 1494 "Bad cms collection type"); 1495 return false; 1496 } 1497 ) 1498 1499 FreelistLocker x(this); 1500 // ------------------------------------------------------------------ 1501 // Print out lots of information which affects the initiation of 1502 // a collection. 1503 if (PrintCMSInitiationStatistics && stats().valid()) { 1504 gclog_or_tty->print("CMSCollector shouldConcurrentCollect: "); 1505 gclog_or_tty->stamp(); 1506 gclog_or_tty->cr(); 1507 stats().print_on(gclog_or_tty); 1508 gclog_or_tty->print_cr("time_until_cms_gen_full %3.7f", 1509 stats().time_until_cms_gen_full()); 1510 gclog_or_tty->print_cr("free="SIZE_FORMAT, _cmsGen->free()); 1511 gclog_or_tty->print_cr("contiguous_available="SIZE_FORMAT, 1512 _cmsGen->contiguous_available()); 1513 gclog_or_tty->print_cr("promotion_rate=%g", stats().promotion_rate()); 1514 gclog_or_tty->print_cr("cms_allocation_rate=%g", stats().cms_allocation_rate()); 1515 gclog_or_tty->print_cr("occupancy=%3.7f", _cmsGen->occupancy()); 1516 gclog_or_tty->print_cr("initiatingOccupancy=%3.7f", _cmsGen->initiating_occupancy()); 1517 gclog_or_tty->print_cr("cms_time_since_begin=%3.7f", stats().cms_time_since_begin()); 1518 gclog_or_tty->print_cr("cms_time_since_end=%3.7f", stats().cms_time_since_end()); 1519 gclog_or_tty->print_cr("metadata initialized %d", 1520 MetaspaceGC::should_concurrent_collect()); 1521 } 1522 // ------------------------------------------------------------------ 1523 1524 // If the estimated time to complete a cms collection (cms_duration()) 1525 // is less than the estimated time remaining until the cms generation 1526 // is full, start a collection. 1527 if (!UseCMSInitiatingOccupancyOnly) { 1528 if (stats().valid()) { 1529 if (stats().time_until_cms_start() == 0.0) { 1530 return true; 1531 } 1532 } else { 1533 // We want to conservatively collect somewhat early in order 1534 // to try and "bootstrap" our CMS/promotion statistics; 1535 // this branch will not fire after the first successful CMS 1536 // collection because the stats should then be valid. 1537 if (_cmsGen->occupancy() >= _bootstrap_occupancy) { 1538 if (Verbose && PrintGCDetails) { 1539 gclog_or_tty->print_cr( 1540 " CMSCollector: collect for bootstrapping statistics:" 1541 " occupancy = %f, boot occupancy = %f", _cmsGen->occupancy(), 1542 _bootstrap_occupancy); 1543 } 1544 return true; 1545 } 1546 } 1547 } 1548 1549 // Otherwise, we start a collection cycle if 1550 // old gen want a collection cycle started. Each may use 1551 // an appropriate criterion for making this decision. 1552 // XXX We need to make sure that the gen expansion 1553 // criterion dovetails well with this. XXX NEED TO FIX THIS 1554 if (_cmsGen->should_concurrent_collect()) { 1555 if (Verbose && PrintGCDetails) { 1556 gclog_or_tty->print_cr("CMS old gen initiated"); 1557 } 1558 return true; 1559 } 1560 1561 // We start a collection if we believe an incremental collection may fail; 1562 // this is not likely to be productive in practice because it's probably too 1563 // late anyway. 1564 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1565 assert(gch->collector_policy()->is_generation_policy(), 1566 "You may want to check the correctness of the following"); 1567 if (gch->incremental_collection_will_fail(true /* consult_young */)) { 1568 if (Verbose && PrintGCDetails) { 1569 gclog_or_tty->print("CMSCollector: collect because incremental collection will fail "); 1570 } 1571 return true; 1572 } 1573 1574 if (MetaspaceGC::should_concurrent_collect()) { 1575 if (Verbose && PrintGCDetails) { 1576 gclog_or_tty->print("CMSCollector: collect for metadata allocation "); 1577 } 1578 return true; 1579 } 1580 1581 // CMSTriggerInterval starts a CMS cycle if enough time has passed. 1582 if (CMSTriggerInterval >= 0) { 1583 if (CMSTriggerInterval == 0) { 1584 // Trigger always 1585 return true; 1586 } 1587 1588 // Check the CMS time since begin (we do not check the stats validity 1589 // as we want to be able to trigger the first CMS cycle as well) 1590 if (stats().cms_time_since_begin() >= (CMSTriggerInterval / ((double) MILLIUNITS))) { 1591 if (Verbose && PrintGCDetails) { 1592 if (stats().valid()) { 1593 gclog_or_tty->print_cr("CMSCollector: collect because of trigger interval (time since last begin %3.7f secs)", 1594 stats().cms_time_since_begin()); 1595 } else { 1596 gclog_or_tty->print_cr("CMSCollector: collect because of trigger interval (first collection)"); 1597 } 1598 } 1599 return true; 1600 } 1601 } 1602 1603 return false; 1604 } 1605 1606 void CMSCollector::set_did_compact(bool v) { _cmsGen->set_did_compact(v); } 1607 1608 // Clear _expansion_cause fields of constituent generations 1609 void CMSCollector::clear_expansion_cause() { 1610 _cmsGen->clear_expansion_cause(); 1611 } 1612 1613 // We should be conservative in starting a collection cycle. To 1614 // start too eagerly runs the risk of collecting too often in the 1615 // extreme. To collect too rarely falls back on full collections, 1616 // which works, even if not optimum in terms of concurrent work. 1617 // As a work around for too eagerly collecting, use the flag 1618 // UseCMSInitiatingOccupancyOnly. This also has the advantage of 1619 // giving the user an easily understandable way of controlling the 1620 // collections. 1621 // We want to start a new collection cycle if any of the following 1622 // conditions hold: 1623 // . our current occupancy exceeds the configured initiating occupancy 1624 // for this generation, or 1625 // . we recently needed to expand this space and have not, since that 1626 // expansion, done a collection of this generation, or 1627 // . the underlying space believes that it may be a good idea to initiate 1628 // a concurrent collection (this may be based on criteria such as the 1629 // following: the space uses linear allocation and linear allocation is 1630 // going to fail, or there is believed to be excessive fragmentation in 1631 // the generation, etc... or ... 1632 // [.(currently done by CMSCollector::shouldConcurrentCollect() only for 1633 // the case of the old generation; see CR 6543076): 1634 // we may be approaching a point at which allocation requests may fail because 1635 // we will be out of sufficient free space given allocation rate estimates.] 1636 bool ConcurrentMarkSweepGeneration::should_concurrent_collect() const { 1637 1638 assert_lock_strong(freelistLock()); 1639 if (occupancy() > initiating_occupancy()) { 1640 if (PrintGCDetails && Verbose) { 1641 gclog_or_tty->print(" %s: collect because of occupancy %f / %f ", 1642 short_name(), occupancy(), initiating_occupancy()); 1643 } 1644 return true; 1645 } 1646 if (UseCMSInitiatingOccupancyOnly) { 1647 return false; 1648 } 1649 if (expansion_cause() == CMSExpansionCause::_satisfy_allocation) { 1650 if (PrintGCDetails && Verbose) { 1651 gclog_or_tty->print(" %s: collect because expanded for allocation ", 1652 short_name()); 1653 } 1654 return true; 1655 } 1656 if (_cmsSpace->should_concurrent_collect()) { 1657 if (PrintGCDetails && Verbose) { 1658 gclog_or_tty->print(" %s: collect because cmsSpace says so ", 1659 short_name()); 1660 } 1661 return true; 1662 } 1663 return false; 1664 } 1665 1666 void ConcurrentMarkSweepGeneration::collect(bool full, 1667 bool clear_all_soft_refs, 1668 size_t size, 1669 bool tlab) 1670 { 1671 collector()->collect(full, clear_all_soft_refs, size, tlab); 1672 } 1673 1674 void CMSCollector::collect(bool full, 1675 bool clear_all_soft_refs, 1676 size_t size, 1677 bool tlab) 1678 { 1679 if (!UseCMSCollectionPassing && _collectorState > Idling) { 1680 // For debugging purposes skip the collection if the state 1681 // is not currently idle 1682 if (TraceCMSState) { 1683 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " skipped full:%d CMS state %d", 1684 Thread::current(), full, _collectorState); 1685 } 1686 return; 1687 } 1688 1689 // The following "if" branch is present for defensive reasons. 1690 // In the current uses of this interface, it can be replaced with: 1691 // assert(!GC_locker.is_active(), "Can't be called otherwise"); 1692 // But I am not placing that assert here to allow future 1693 // generality in invoking this interface. 1694 if (GC_locker::is_active()) { 1695 // A consistency test for GC_locker 1696 assert(GC_locker::needs_gc(), "Should have been set already"); 1697 // Skip this foreground collection, instead 1698 // expanding the heap if necessary. 1699 // Need the free list locks for the call to free() in compute_new_size() 1700 compute_new_size(); 1701 return; 1702 } 1703 acquire_control_and_collect(full, clear_all_soft_refs); 1704 _full_gcs_since_conc_gc++; 1705 } 1706 1707 void CMSCollector::request_full_gc(unsigned int full_gc_count, GCCause::Cause cause) { 1708 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1709 unsigned int gc_count = gch->total_full_collections(); 1710 if (gc_count == full_gc_count) { 1711 MutexLockerEx y(CGC_lock, Mutex::_no_safepoint_check_flag); 1712 _full_gc_requested = true; 1713 _full_gc_cause = cause; 1714 CGC_lock->notify(); // nudge CMS thread 1715 } else { 1716 assert(gc_count > full_gc_count, "Error: causal loop"); 1717 } 1718 } 1719 1720 bool CMSCollector::is_external_interruption() { 1721 GCCause::Cause cause = GenCollectedHeap::heap()->gc_cause(); 1722 return GCCause::is_user_requested_gc(cause) || 1723 GCCause::is_serviceability_requested_gc(cause); 1724 } 1725 1726 void CMSCollector::report_concurrent_mode_interruption() { 1727 if (is_external_interruption()) { 1728 if (PrintGCDetails) { 1729 gclog_or_tty->print(" (concurrent mode interrupted)"); 1730 } 1731 } else { 1732 if (PrintGCDetails) { 1733 gclog_or_tty->print(" (concurrent mode failure)"); 1734 } 1735 _gc_tracer_cm->report_concurrent_mode_failure(); 1736 } 1737 } 1738 1739 1740 // The foreground and background collectors need to coordinate in order 1741 // to make sure that they do not mutually interfere with CMS collections. 1742 // When a background collection is active, 1743 // the foreground collector may need to take over (preempt) and 1744 // synchronously complete an ongoing collection. Depending on the 1745 // frequency of the background collections and the heap usage 1746 // of the application, this preemption can be seldom or frequent. 1747 // There are only certain 1748 // points in the background collection that the "collection-baton" 1749 // can be passed to the foreground collector. 1750 // 1751 // The foreground collector will wait for the baton before 1752 // starting any part of the collection. The foreground collector 1753 // will only wait at one location. 1754 // 1755 // The background collector will yield the baton before starting a new 1756 // phase of the collection (e.g., before initial marking, marking from roots, 1757 // precleaning, final re-mark, sweep etc.) This is normally done at the head 1758 // of the loop which switches the phases. The background collector does some 1759 // of the phases (initial mark, final re-mark) with the world stopped. 1760 // Because of locking involved in stopping the world, 1761 // the foreground collector should not block waiting for the background 1762 // collector when it is doing a stop-the-world phase. The background 1763 // collector will yield the baton at an additional point just before 1764 // it enters a stop-the-world phase. Once the world is stopped, the 1765 // background collector checks the phase of the collection. If the 1766 // phase has not changed, it proceeds with the collection. If the 1767 // phase has changed, it skips that phase of the collection. See 1768 // the comments on the use of the Heap_lock in collect_in_background(). 1769 // 1770 // Variable used in baton passing. 1771 // _foregroundGCIsActive - Set to true by the foreground collector when 1772 // it wants the baton. The foreground clears it when it has finished 1773 // the collection. 1774 // _foregroundGCShouldWait - Set to true by the background collector 1775 // when it is running. The foreground collector waits while 1776 // _foregroundGCShouldWait is true. 1777 // CGC_lock - monitor used to protect access to the above variables 1778 // and to notify the foreground and background collectors. 1779 // _collectorState - current state of the CMS collection. 1780 // 1781 // The foreground collector 1782 // acquires the CGC_lock 1783 // sets _foregroundGCIsActive 1784 // waits on the CGC_lock for _foregroundGCShouldWait to be false 1785 // various locks acquired in preparation for the collection 1786 // are released so as not to block the background collector 1787 // that is in the midst of a collection 1788 // proceeds with the collection 1789 // clears _foregroundGCIsActive 1790 // returns 1791 // 1792 // The background collector in a loop iterating on the phases of the 1793 // collection 1794 // acquires the CGC_lock 1795 // sets _foregroundGCShouldWait 1796 // if _foregroundGCIsActive is set 1797 // clears _foregroundGCShouldWait, notifies _CGC_lock 1798 // waits on _CGC_lock for _foregroundGCIsActive to become false 1799 // and exits the loop. 1800 // otherwise 1801 // proceed with that phase of the collection 1802 // if the phase is a stop-the-world phase, 1803 // yield the baton once more just before enqueueing 1804 // the stop-world CMS operation (executed by the VM thread). 1805 // returns after all phases of the collection are done 1806 // 1807 1808 void CMSCollector::acquire_control_and_collect(bool full, 1809 bool clear_all_soft_refs) { 1810 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint"); 1811 assert(!Thread::current()->is_ConcurrentGC_thread(), 1812 "shouldn't try to acquire control from self!"); 1813 1814 // Start the protocol for acquiring control of the 1815 // collection from the background collector (aka CMS thread). 1816 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 1817 "VM thread should have CMS token"); 1818 // Remember the possibly interrupted state of an ongoing 1819 // concurrent collection 1820 CollectorState first_state = _collectorState; 1821 1822 // Signal to a possibly ongoing concurrent collection that 1823 // we want to do a foreground collection. 1824 _foregroundGCIsActive = true; 1825 1826 // Disable incremental mode during a foreground collection. 1827 ICMSDisabler icms_disabler; 1828 1829 // release locks and wait for a notify from the background collector 1830 // releasing the locks in only necessary for phases which 1831 // do yields to improve the granularity of the collection. 1832 assert_lock_strong(bitMapLock()); 1833 // We need to lock the Free list lock for the space that we are 1834 // currently collecting. 1835 assert(haveFreelistLocks(), "Must be holding free list locks"); 1836 bitMapLock()->unlock(); 1837 releaseFreelistLocks(); 1838 { 1839 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1840 if (_foregroundGCShouldWait) { 1841 // We are going to be waiting for action for the CMS thread; 1842 // it had better not be gone (for instance at shutdown)! 1843 assert(ConcurrentMarkSweepThread::cmst() != NULL, 1844 "CMS thread must be running"); 1845 // Wait here until the background collector gives us the go-ahead 1846 ConcurrentMarkSweepThread::clear_CMS_flag( 1847 ConcurrentMarkSweepThread::CMS_vm_has_token); // release token 1848 // Get a possibly blocked CMS thread going: 1849 // Note that we set _foregroundGCIsActive true above, 1850 // without protection of the CGC_lock. 1851 CGC_lock->notify(); 1852 assert(!ConcurrentMarkSweepThread::vm_thread_wants_cms_token(), 1853 "Possible deadlock"); 1854 while (_foregroundGCShouldWait) { 1855 // wait for notification 1856 CGC_lock->wait(Mutex::_no_safepoint_check_flag); 1857 // Possibility of delay/starvation here, since CMS token does 1858 // not know to give priority to VM thread? Actually, i think 1859 // there wouldn't be any delay/starvation, but the proof of 1860 // that "fact" (?) appears non-trivial. XXX 20011219YSR 1861 } 1862 ConcurrentMarkSweepThread::set_CMS_flag( 1863 ConcurrentMarkSweepThread::CMS_vm_has_token); 1864 } 1865 } 1866 // The CMS_token is already held. Get back the other locks. 1867 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 1868 "VM thread should have CMS token"); 1869 getFreelistLocks(); 1870 bitMapLock()->lock_without_safepoint_check(); 1871 if (TraceCMSState) { 1872 gclog_or_tty->print_cr("CMS foreground collector has asked for control " 1873 INTPTR_FORMAT " with first state %d", Thread::current(), first_state); 1874 gclog_or_tty->print_cr(" gets control with state %d", _collectorState); 1875 } 1876 1877 // Check if we need to do a compaction, or if not, whether 1878 // we need to start the mark-sweep from scratch. 1879 bool should_compact = false; 1880 bool should_start_over = false; 1881 decide_foreground_collection_type(clear_all_soft_refs, 1882 &should_compact, &should_start_over); 1883 1884 NOT_PRODUCT( 1885 if (RotateCMSCollectionTypes) { 1886 if (_cmsGen->debug_collection_type() == 1887 ConcurrentMarkSweepGeneration::MSC_foreground_collection_type) { 1888 should_compact = true; 1889 } else if (_cmsGen->debug_collection_type() == 1890 ConcurrentMarkSweepGeneration::MS_foreground_collection_type) { 1891 should_compact = false; 1892 } 1893 } 1894 ) 1895 1896 if (first_state > Idling) { 1897 report_concurrent_mode_interruption(); 1898 } 1899 1900 set_did_compact(should_compact); 1901 if (should_compact) { 1902 // If the collection is being acquired from the background 1903 // collector, there may be references on the discovered 1904 // references lists that have NULL referents (being those 1905 // that were concurrently cleared by a mutator) or 1906 // that are no longer active (having been enqueued concurrently 1907 // by the mutator). 1908 // Scrub the list of those references because Mark-Sweep-Compact 1909 // code assumes referents are not NULL and that all discovered 1910 // Reference objects are active. 1911 ref_processor()->clean_up_discovered_references(); 1912 1913 if (first_state > Idling) { 1914 save_heap_summary(); 1915 } 1916 1917 do_compaction_work(clear_all_soft_refs); 1918 1919 // Has the GC time limit been exceeded? 1920 DefNewGeneration* young_gen = _young_gen->as_DefNewGeneration(); 1921 size_t max_eden_size = young_gen->max_capacity() - 1922 young_gen->to()->capacity() - 1923 young_gen->from()->capacity(); 1924 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1925 GCCause::Cause gc_cause = gch->gc_cause(); 1926 size_policy()->check_gc_overhead_limit(_young_gen->used(), 1927 young_gen->eden()->used(), 1928 _cmsGen->max_capacity(), 1929 max_eden_size, 1930 full, 1931 gc_cause, 1932 gch->collector_policy()); 1933 } else { 1934 do_mark_sweep_work(clear_all_soft_refs, first_state, 1935 should_start_over); 1936 } 1937 // Reset the expansion cause, now that we just completed 1938 // a collection cycle. 1939 clear_expansion_cause(); 1940 _foregroundGCIsActive = false; 1941 return; 1942 } 1943 1944 // Resize the tenured generation 1945 // after obtaining the free list locks for the 1946 // two generations. 1947 void CMSCollector::compute_new_size() { 1948 assert_locked_or_safepoint(Heap_lock); 1949 FreelistLocker z(this); 1950 MetaspaceGC::compute_new_size(); 1951 _cmsGen->compute_new_size_free_list(); 1952 } 1953 1954 // A work method used by foreground collection to determine 1955 // what type of collection (compacting or not, continuing or fresh) 1956 // it should do. 1957 // NOTE: the intent is to make UseCMSCompactAtFullCollection 1958 // and CMSCompactWhenClearAllSoftRefs the default in the future 1959 // and do away with the flags after a suitable period. 1960 void CMSCollector::decide_foreground_collection_type( 1961 bool clear_all_soft_refs, bool* should_compact, 1962 bool* should_start_over) { 1963 // Normally, we'll compact only if the UseCMSCompactAtFullCollection 1964 // flag is set, and we have either requested a System.gc() or 1965 // the number of full gc's since the last concurrent cycle 1966 // has exceeded the threshold set by CMSFullGCsBeforeCompaction, 1967 // or if an incremental collection has failed 1968 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1969 assert(gch->collector_policy()->is_generation_policy(), 1970 "You may want to check the correctness of the following"); 1971 // Inform cms gen if this was due to partial collection failing. 1972 // The CMS gen may use this fact to determine its expansion policy. 1973 if (gch->incremental_collection_will_fail(false /* don't consult_young */)) { 1974 assert(!_cmsGen->incremental_collection_failed(), 1975 "Should have been noticed, reacted to and cleared"); 1976 _cmsGen->set_incremental_collection_failed(); 1977 } 1978 *should_compact = 1979 UseCMSCompactAtFullCollection && 1980 ((_full_gcs_since_conc_gc >= CMSFullGCsBeforeCompaction) || 1981 GCCause::is_user_requested_gc(gch->gc_cause()) || 1982 gch->incremental_collection_will_fail(true /* consult_young */)); 1983 *should_start_over = false; 1984 if (clear_all_soft_refs && !*should_compact) { 1985 // We are about to do a last ditch collection attempt 1986 // so it would normally make sense to do a compaction 1987 // to reclaim as much space as possible. 1988 if (CMSCompactWhenClearAllSoftRefs) { 1989 // Default: The rationale is that in this case either 1990 // we are past the final marking phase, in which case 1991 // we'd have to start over, or so little has been done 1992 // that there's little point in saving that work. Compaction 1993 // appears to be the sensible choice in either case. 1994 *should_compact = true; 1995 } else { 1996 // We have been asked to clear all soft refs, but not to 1997 // compact. Make sure that we aren't past the final checkpoint 1998 // phase, for that is where we process soft refs. If we are already 1999 // past that phase, we'll need to redo the refs discovery phase and 2000 // if necessary clear soft refs that weren't previously 2001 // cleared. We do so by remembering the phase in which 2002 // we came in, and if we are past the refs processing 2003 // phase, we'll choose to just redo the mark-sweep 2004 // collection from scratch. 2005 if (_collectorState > FinalMarking) { 2006 // We are past the refs processing phase; 2007 // start over and do a fresh synchronous CMS cycle 2008 _collectorState = Resetting; // skip to reset to start new cycle 2009 reset(false /* == !asynch */); 2010 *should_start_over = true; 2011 } // else we can continue a possibly ongoing current cycle 2012 } 2013 } 2014 } 2015 2016 // A work method used by the foreground collector to do 2017 // a mark-sweep-compact. 2018 void CMSCollector::do_compaction_work(bool clear_all_soft_refs) { 2019 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2020 2021 STWGCTimer* gc_timer = GenMarkSweep::gc_timer(); 2022 gc_timer->register_gc_start(); 2023 2024 SerialOldTracer* gc_tracer = GenMarkSweep::gc_tracer(); 2025 gc_tracer->report_gc_start(gch->gc_cause(), gc_timer->gc_start()); 2026 2027 GCTraceTime t("CMS:MSC ", PrintGCDetails && Verbose, true, NULL); 2028 if (PrintGC && Verbose && !(GCCause::is_user_requested_gc(gch->gc_cause()))) { 2029 gclog_or_tty->print_cr("Compact ConcurrentMarkSweepGeneration after %d " 2030 "collections passed to foreground collector", _full_gcs_since_conc_gc); 2031 } 2032 2033 // Sample collection interval time and reset for collection pause. 2034 if (UseAdaptiveSizePolicy) { 2035 size_policy()->msc_collection_begin(); 2036 } 2037 2038 // Temporarily widen the span of the weak reference processing to 2039 // the entire heap. 2040 MemRegion new_span(GenCollectedHeap::heap()->reserved_region()); 2041 ReferenceProcessorSpanMutator rp_mut_span(ref_processor(), new_span); 2042 // Temporarily, clear the "is_alive_non_header" field of the 2043 // reference processor. 2044 ReferenceProcessorIsAliveMutator rp_mut_closure(ref_processor(), NULL); 2045 // Temporarily make reference _processing_ single threaded (non-MT). 2046 ReferenceProcessorMTProcMutator rp_mut_mt_processing(ref_processor(), false); 2047 // Temporarily make refs discovery atomic 2048 ReferenceProcessorAtomicMutator rp_mut_atomic(ref_processor(), true); 2049 // Temporarily make reference _discovery_ single threaded (non-MT) 2050 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false); 2051 2052 ref_processor()->set_enqueuing_is_done(false); 2053 ref_processor()->enable_discovery(false /*verify_disabled*/, false /*check_no_refs*/); 2054 ref_processor()->setup_policy(clear_all_soft_refs); 2055 // If an asynchronous collection finishes, the _modUnionTable is 2056 // all clear. If we are assuming the collection from an asynchronous 2057 // collection, clear the _modUnionTable. 2058 assert(_collectorState != Idling || _modUnionTable.isAllClear(), 2059 "_modUnionTable should be clear if the baton was not passed"); 2060 _modUnionTable.clear_all(); 2061 assert(_collectorState != Idling || _ct->klass_rem_set()->mod_union_is_clear(), 2062 "mod union for klasses should be clear if the baton was passed"); 2063 _ct->klass_rem_set()->clear_mod_union(); 2064 2065 // We must adjust the allocation statistics being maintained 2066 // in the free list space. We do so by reading and clearing 2067 // the sweep timer and updating the block flux rate estimates below. 2068 assert(!_intra_sweep_timer.is_active(), "_intra_sweep_timer should be inactive"); 2069 if (_inter_sweep_timer.is_active()) { 2070 _inter_sweep_timer.stop(); 2071 // Note that we do not use this sample to update the _inter_sweep_estimate. 2072 _cmsGen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()), 2073 _inter_sweep_estimate.padded_average(), 2074 _intra_sweep_estimate.padded_average()); 2075 } 2076 2077 GenMarkSweep::invoke_at_safepoint(_cmsGen->level(), 2078 ref_processor(), clear_all_soft_refs); 2079 #ifdef ASSERT 2080 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 2081 size_t free_size = cms_space->free(); 2082 assert(free_size == 2083 pointer_delta(cms_space->end(), cms_space->compaction_top()) 2084 * HeapWordSize, 2085 "All the free space should be compacted into one chunk at top"); 2086 assert(cms_space->dictionary()->total_chunk_size( 2087 debug_only(cms_space->freelistLock())) == 0 || 2088 cms_space->totalSizeInIndexedFreeLists() == 0, 2089 "All the free space should be in a single chunk"); 2090 size_t num = cms_space->totalCount(); 2091 assert((free_size == 0 && num == 0) || 2092 (free_size > 0 && (num == 1 || num == 2)), 2093 "There should be at most 2 free chunks after compaction"); 2094 #endif // ASSERT 2095 _collectorState = Resetting; 2096 assert(_restart_addr == NULL, 2097 "Should have been NULL'd before baton was passed"); 2098 reset(false /* == !asynch */); 2099 _cmsGen->reset_after_compaction(); 2100 _concurrent_cycles_since_last_unload = 0; 2101 2102 // Clear any data recorded in the PLAB chunk arrays. 2103 if (_survivor_plab_array != NULL) { 2104 reset_survivor_plab_arrays(); 2105 } 2106 2107 // Adjust the per-size allocation stats for the next epoch. 2108 _cmsGen->cmsSpace()->endSweepFLCensus(sweep_count() /* fake */); 2109 // Restart the "inter sweep timer" for the next epoch. 2110 _inter_sweep_timer.reset(); 2111 _inter_sweep_timer.start(); 2112 2113 // Sample collection pause time and reset for collection interval. 2114 if (UseAdaptiveSizePolicy) { 2115 size_policy()->msc_collection_end(gch->gc_cause()); 2116 } 2117 2118 gc_timer->register_gc_end(); 2119 2120 gc_tracer->report_gc_end(gc_timer->gc_end(), gc_timer->time_partitions()); 2121 2122 // For a mark-sweep-compact, compute_new_size() will be called 2123 // in the heap's do_collection() method. 2124 } 2125 2126 // A work method used by the foreground collector to do 2127 // a mark-sweep, after taking over from a possibly on-going 2128 // concurrent mark-sweep collection. 2129 void CMSCollector::do_mark_sweep_work(bool clear_all_soft_refs, 2130 CollectorState first_state, bool should_start_over) { 2131 if (PrintGC && Verbose) { 2132 gclog_or_tty->print_cr("Pass concurrent collection to foreground " 2133 "collector with count %d", 2134 _full_gcs_since_conc_gc); 2135 } 2136 switch (_collectorState) { 2137 case Idling: 2138 if (first_state == Idling || should_start_over) { 2139 // The background GC was not active, or should 2140 // restarted from scratch; start the cycle. 2141 _collectorState = InitialMarking; 2142 } 2143 // If first_state was not Idling, then a background GC 2144 // was in progress and has now finished. No need to do it 2145 // again. Leave the state as Idling. 2146 break; 2147 case Precleaning: 2148 // In the foreground case don't do the precleaning since 2149 // it is not done concurrently and there is extra work 2150 // required. 2151 _collectorState = FinalMarking; 2152 } 2153 collect_in_foreground(clear_all_soft_refs, GenCollectedHeap::heap()->gc_cause()); 2154 2155 // For a mark-sweep, compute_new_size() will be called 2156 // in the heap's do_collection() method. 2157 } 2158 2159 2160 void CMSCollector::print_eden_and_survivor_chunk_arrays() { 2161 DefNewGeneration* dng = _young_gen->as_DefNewGeneration(); 2162 EdenSpace* eden_space = dng->eden(); 2163 ContiguousSpace* from_space = dng->from(); 2164 ContiguousSpace* to_space = dng->to(); 2165 // Eden 2166 if (_eden_chunk_array != NULL) { 2167 gclog_or_tty->print_cr("eden " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")", 2168 eden_space->bottom(), eden_space->top(), 2169 eden_space->end(), eden_space->capacity()); 2170 gclog_or_tty->print_cr("_eden_chunk_index=" SIZE_FORMAT ", " 2171 "_eden_chunk_capacity=" SIZE_FORMAT, 2172 _eden_chunk_index, _eden_chunk_capacity); 2173 for (size_t i = 0; i < _eden_chunk_index; i++) { 2174 gclog_or_tty->print_cr("_eden_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT, 2175 i, _eden_chunk_array[i]); 2176 } 2177 } 2178 // Survivor 2179 if (_survivor_chunk_array != NULL) { 2180 gclog_or_tty->print_cr("survivor " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")", 2181 from_space->bottom(), from_space->top(), 2182 from_space->end(), from_space->capacity()); 2183 gclog_or_tty->print_cr("_survivor_chunk_index=" SIZE_FORMAT ", " 2184 "_survivor_chunk_capacity=" SIZE_FORMAT, 2185 _survivor_chunk_index, _survivor_chunk_capacity); 2186 for (size_t i = 0; i < _survivor_chunk_index; i++) { 2187 gclog_or_tty->print_cr("_survivor_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT, 2188 i, _survivor_chunk_array[i]); 2189 } 2190 } 2191 } 2192 2193 void CMSCollector::getFreelistLocks() const { 2194 // Get locks for all free lists in all generations that this 2195 // collector is responsible for 2196 _cmsGen->freelistLock()->lock_without_safepoint_check(); 2197 } 2198 2199 void CMSCollector::releaseFreelistLocks() const { 2200 // Release locks for all free lists in all generations that this 2201 // collector is responsible for 2202 _cmsGen->freelistLock()->unlock(); 2203 } 2204 2205 bool CMSCollector::haveFreelistLocks() const { 2206 // Check locks for all free lists in all generations that this 2207 // collector is responsible for 2208 assert_lock_strong(_cmsGen->freelistLock()); 2209 PRODUCT_ONLY(ShouldNotReachHere()); 2210 return true; 2211 } 2212 2213 // A utility class that is used by the CMS collector to 2214 // temporarily "release" the foreground collector from its 2215 // usual obligation to wait for the background collector to 2216 // complete an ongoing phase before proceeding. 2217 class ReleaseForegroundGC: public StackObj { 2218 private: 2219 CMSCollector* _c; 2220 public: 2221 ReleaseForegroundGC(CMSCollector* c) : _c(c) { 2222 assert(_c->_foregroundGCShouldWait, "Else should not need to call"); 2223 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 2224 // allow a potentially blocked foreground collector to proceed 2225 _c->_foregroundGCShouldWait = false; 2226 if (_c->_foregroundGCIsActive) { 2227 CGC_lock->notify(); 2228 } 2229 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 2230 "Possible deadlock"); 2231 } 2232 2233 ~ReleaseForegroundGC() { 2234 assert(!_c->_foregroundGCShouldWait, "Usage protocol violation?"); 2235 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 2236 _c->_foregroundGCShouldWait = true; 2237 } 2238 }; 2239 2240 // There are separate collect_in_background and collect_in_foreground because of 2241 // the different locking requirements of the background collector and the 2242 // foreground collector. There was originally an attempt to share 2243 // one "collect" method between the background collector and the foreground 2244 // collector but the if-then-else required made it cleaner to have 2245 // separate methods. 2246 void CMSCollector::collect_in_background(bool clear_all_soft_refs, GCCause::Cause cause) { 2247 assert(Thread::current()->is_ConcurrentGC_thread(), 2248 "A CMS asynchronous collection is only allowed on a CMS thread."); 2249 2250 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2251 { 2252 bool safepoint_check = Mutex::_no_safepoint_check_flag; 2253 MutexLockerEx hl(Heap_lock, safepoint_check); 2254 FreelistLocker fll(this); 2255 MutexLockerEx x(CGC_lock, safepoint_check); 2256 if (_foregroundGCIsActive || !UseAsyncConcMarkSweepGC) { 2257 // The foreground collector is active or we're 2258 // not using asynchronous collections. Skip this 2259 // background collection. 2260 assert(!_foregroundGCShouldWait, "Should be clear"); 2261 return; 2262 } else { 2263 assert(_collectorState == Idling, "Should be idling before start."); 2264 _collectorState = InitialMarking; 2265 register_gc_start(cause); 2266 // Reset the expansion cause, now that we are about to begin 2267 // a new cycle. 2268 clear_expansion_cause(); 2269 2270 // Clear the MetaspaceGC flag since a concurrent collection 2271 // is starting but also clear it after the collection. 2272 MetaspaceGC::set_should_concurrent_collect(false); 2273 } 2274 // Decide if we want to enable class unloading as part of the 2275 // ensuing concurrent GC cycle. 2276 update_should_unload_classes(); 2277 _full_gc_requested = false; // acks all outstanding full gc requests 2278 _full_gc_cause = GCCause::_no_gc; 2279 // Signal that we are about to start a collection 2280 gch->increment_total_full_collections(); // ... starting a collection cycle 2281 _collection_count_start = gch->total_full_collections(); 2282 } 2283 2284 // Used for PrintGC 2285 size_t prev_used; 2286 if (PrintGC && Verbose) { 2287 prev_used = _cmsGen->used(); // XXXPERM 2288 } 2289 2290 // The change of the collection state is normally done at this level; 2291 // the exceptions are phases that are executed while the world is 2292 // stopped. For those phases the change of state is done while the 2293 // world is stopped. For baton passing purposes this allows the 2294 // background collector to finish the phase and change state atomically. 2295 // The foreground collector cannot wait on a phase that is done 2296 // while the world is stopped because the foreground collector already 2297 // has the world stopped and would deadlock. 2298 while (_collectorState != Idling) { 2299 if (TraceCMSState) { 2300 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " in CMS state %d", 2301 Thread::current(), _collectorState); 2302 } 2303 // The foreground collector 2304 // holds the Heap_lock throughout its collection. 2305 // holds the CMS token (but not the lock) 2306 // except while it is waiting for the background collector to yield. 2307 // 2308 // The foreground collector should be blocked (not for long) 2309 // if the background collector is about to start a phase 2310 // executed with world stopped. If the background 2311 // collector has already started such a phase, the 2312 // foreground collector is blocked waiting for the 2313 // Heap_lock. The stop-world phases (InitialMarking and FinalMarking) 2314 // are executed in the VM thread. 2315 // 2316 // The locking order is 2317 // PendingListLock (PLL) -- if applicable (FinalMarking) 2318 // Heap_lock (both this & PLL locked in VM_CMS_Operation::prologue()) 2319 // CMS token (claimed in 2320 // stop_world_and_do() --> 2321 // safepoint_synchronize() --> 2322 // CMSThread::synchronize()) 2323 2324 { 2325 // Check if the FG collector wants us to yield. 2326 CMSTokenSync x(true); // is cms thread 2327 if (waitForForegroundGC()) { 2328 // We yielded to a foreground GC, nothing more to be 2329 // done this round. 2330 assert(_foregroundGCShouldWait == false, "We set it to false in " 2331 "waitForForegroundGC()"); 2332 if (TraceCMSState) { 2333 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT 2334 " exiting collection CMS state %d", 2335 Thread::current(), _collectorState); 2336 } 2337 return; 2338 } else { 2339 // The background collector can run but check to see if the 2340 // foreground collector has done a collection while the 2341 // background collector was waiting to get the CGC_lock 2342 // above. If yes, break so that _foregroundGCShouldWait 2343 // is cleared before returning. 2344 if (_collectorState == Idling) { 2345 break; 2346 } 2347 } 2348 } 2349 2350 assert(_foregroundGCShouldWait, "Foreground collector, if active, " 2351 "should be waiting"); 2352 2353 switch (_collectorState) { 2354 case InitialMarking: 2355 { 2356 ReleaseForegroundGC x(this); 2357 stats().record_cms_begin(); 2358 VM_CMS_Initial_Mark initial_mark_op(this); 2359 VMThread::execute(&initial_mark_op); 2360 } 2361 // The collector state may be any legal state at this point 2362 // since the background collector may have yielded to the 2363 // foreground collector. 2364 break; 2365 case Marking: 2366 // initial marking in checkpointRootsInitialWork has been completed 2367 if (markFromRoots(true)) { // we were successful 2368 assert(_collectorState == Precleaning, "Collector state should " 2369 "have changed"); 2370 } else { 2371 assert(_foregroundGCIsActive, "Internal state inconsistency"); 2372 } 2373 break; 2374 case Precleaning: 2375 if (UseAdaptiveSizePolicy) { 2376 size_policy()->concurrent_precleaning_begin(); 2377 } 2378 // marking from roots in markFromRoots has been completed 2379 preclean(); 2380 if (UseAdaptiveSizePolicy) { 2381 size_policy()->concurrent_precleaning_end(); 2382 } 2383 assert(_collectorState == AbortablePreclean || 2384 _collectorState == FinalMarking, 2385 "Collector state should have changed"); 2386 break; 2387 case AbortablePreclean: 2388 if (UseAdaptiveSizePolicy) { 2389 size_policy()->concurrent_phases_resume(); 2390 } 2391 abortable_preclean(); 2392 if (UseAdaptiveSizePolicy) { 2393 size_policy()->concurrent_precleaning_end(); 2394 } 2395 assert(_collectorState == FinalMarking, "Collector state should " 2396 "have changed"); 2397 break; 2398 case FinalMarking: 2399 { 2400 ReleaseForegroundGC x(this); 2401 2402 VM_CMS_Final_Remark final_remark_op(this); 2403 VMThread::execute(&final_remark_op); 2404 } 2405 assert(_foregroundGCShouldWait, "block post-condition"); 2406 break; 2407 case Sweeping: 2408 if (UseAdaptiveSizePolicy) { 2409 size_policy()->concurrent_sweeping_begin(); 2410 } 2411 // final marking in checkpointRootsFinal has been completed 2412 sweep(true); 2413 assert(_collectorState == Resizing, "Collector state change " 2414 "to Resizing must be done under the free_list_lock"); 2415 _full_gcs_since_conc_gc = 0; 2416 2417 // Stop the timers for adaptive size policy for the concurrent phases 2418 if (UseAdaptiveSizePolicy) { 2419 size_policy()->concurrent_sweeping_end(); 2420 size_policy()->concurrent_phases_end(gch->gc_cause(), 2421 gch->prev_gen(_cmsGen)->capacity(), 2422 _cmsGen->free()); 2423 } 2424 2425 case Resizing: { 2426 // Sweeping has been completed... 2427 // At this point the background collection has completed. 2428 // Don't move the call to compute_new_size() down 2429 // into code that might be executed if the background 2430 // collection was preempted. 2431 { 2432 ReleaseForegroundGC x(this); // unblock FG collection 2433 MutexLockerEx y(Heap_lock, Mutex::_no_safepoint_check_flag); 2434 CMSTokenSync z(true); // not strictly needed. 2435 if (_collectorState == Resizing) { 2436 compute_new_size(); 2437 save_heap_summary(); 2438 _collectorState = Resetting; 2439 } else { 2440 assert(_collectorState == Idling, "The state should only change" 2441 " because the foreground collector has finished the collection"); 2442 } 2443 } 2444 break; 2445 } 2446 case Resetting: 2447 // CMS heap resizing has been completed 2448 reset(true); 2449 assert(_collectorState == Idling, "Collector state should " 2450 "have changed"); 2451 2452 MetaspaceGC::set_should_concurrent_collect(false); 2453 2454 stats().record_cms_end(); 2455 // Don't move the concurrent_phases_end() and compute_new_size() 2456 // calls to here because a preempted background collection 2457 // has it's state set to "Resetting". 2458 break; 2459 case Idling: 2460 default: 2461 ShouldNotReachHere(); 2462 break; 2463 } 2464 if (TraceCMSState) { 2465 gclog_or_tty->print_cr(" Thread " INTPTR_FORMAT " done - next CMS state %d", 2466 Thread::current(), _collectorState); 2467 } 2468 assert(_foregroundGCShouldWait, "block post-condition"); 2469 } 2470 2471 // Should this be in gc_epilogue? 2472 collector_policy()->counters()->update_counters(); 2473 2474 { 2475 // Clear _foregroundGCShouldWait and, in the event that the 2476 // foreground collector is waiting, notify it, before 2477 // returning. 2478 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 2479 _foregroundGCShouldWait = false; 2480 if (_foregroundGCIsActive) { 2481 CGC_lock->notify(); 2482 } 2483 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 2484 "Possible deadlock"); 2485 } 2486 if (TraceCMSState) { 2487 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT 2488 " exiting collection CMS state %d", 2489 Thread::current(), _collectorState); 2490 } 2491 if (PrintGC && Verbose) { 2492 _cmsGen->print_heap_change(prev_used); 2493 } 2494 } 2495 2496 void CMSCollector::register_foreground_gc_start(GCCause::Cause cause) { 2497 if (!_cms_start_registered) { 2498 register_gc_start(cause); 2499 } 2500 } 2501 2502 void CMSCollector::register_gc_start(GCCause::Cause cause) { 2503 _cms_start_registered = true; 2504 _gc_timer_cm->register_gc_start(); 2505 _gc_tracer_cm->report_gc_start(cause, _gc_timer_cm->gc_start()); 2506 } 2507 2508 void CMSCollector::register_gc_end() { 2509 if (_cms_start_registered) { 2510 report_heap_summary(GCWhen::AfterGC); 2511 2512 _gc_timer_cm->register_gc_end(); 2513 _gc_tracer_cm->report_gc_end(_gc_timer_cm->gc_end(), _gc_timer_cm->time_partitions()); 2514 _cms_start_registered = false; 2515 } 2516 } 2517 2518 void CMSCollector::save_heap_summary() { 2519 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2520 _last_heap_summary = gch->create_heap_summary(); 2521 _last_metaspace_summary = gch->create_metaspace_summary(); 2522 } 2523 2524 void CMSCollector::report_heap_summary(GCWhen::Type when) { 2525 _gc_tracer_cm->report_gc_heap_summary(when, _last_heap_summary); 2526 _gc_tracer_cm->report_metaspace_summary(when, _last_metaspace_summary); 2527 } 2528 2529 void CMSCollector::collect_in_foreground(bool clear_all_soft_refs, GCCause::Cause cause) { 2530 assert(_foregroundGCIsActive && !_foregroundGCShouldWait, 2531 "Foreground collector should be waiting, not executing"); 2532 assert(Thread::current()->is_VM_thread(), "A foreground collection" 2533 "may only be done by the VM Thread with the world stopped"); 2534 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 2535 "VM thread should have CMS token"); 2536 2537 NOT_PRODUCT(GCTraceTime t("CMS:MS (foreground) ", PrintGCDetails && Verbose, 2538 true, NULL);) 2539 if (UseAdaptiveSizePolicy) { 2540 size_policy()->ms_collection_begin(); 2541 } 2542 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact); 2543 2544 HandleMark hm; // Discard invalid handles created during verification 2545 2546 if (VerifyBeforeGC && 2547 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2548 Universe::verify(); 2549 } 2550 2551 // Snapshot the soft reference policy to be used in this collection cycle. 2552 ref_processor()->setup_policy(clear_all_soft_refs); 2553 2554 // Decide if class unloading should be done 2555 update_should_unload_classes(); 2556 2557 bool init_mark_was_synchronous = false; // until proven otherwise 2558 while (_collectorState != Idling) { 2559 if (TraceCMSState) { 2560 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " in CMS state %d", 2561 Thread::current(), _collectorState); 2562 } 2563 switch (_collectorState) { 2564 case InitialMarking: 2565 register_foreground_gc_start(cause); 2566 init_mark_was_synchronous = true; // fact to be exploited in re-mark 2567 checkpointRootsInitial(false); 2568 assert(_collectorState == Marking, "Collector state should have changed" 2569 " within checkpointRootsInitial()"); 2570 break; 2571 case Marking: 2572 // initial marking in checkpointRootsInitialWork has been completed 2573 if (VerifyDuringGC && 2574 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2575 Universe::verify("Verify before initial mark: "); 2576 } 2577 { 2578 bool res = markFromRoots(false); 2579 assert(res && _collectorState == FinalMarking, "Collector state should " 2580 "have changed"); 2581 break; 2582 } 2583 case FinalMarking: 2584 if (VerifyDuringGC && 2585 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2586 Universe::verify("Verify before re-mark: "); 2587 } 2588 checkpointRootsFinal(false, clear_all_soft_refs, 2589 init_mark_was_synchronous); 2590 assert(_collectorState == Sweeping, "Collector state should not " 2591 "have changed within checkpointRootsFinal()"); 2592 break; 2593 case Sweeping: 2594 // final marking in checkpointRootsFinal has been completed 2595 if (VerifyDuringGC && 2596 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2597 Universe::verify("Verify before sweep: "); 2598 } 2599 sweep(false); 2600 assert(_collectorState == Resizing, "Incorrect state"); 2601 break; 2602 case Resizing: { 2603 // Sweeping has been completed; the actual resize in this case 2604 // is done separately; nothing to be done in this state. 2605 _collectorState = Resetting; 2606 break; 2607 } 2608 case Resetting: 2609 // The heap has been resized. 2610 if (VerifyDuringGC && 2611 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2612 Universe::verify("Verify before reset: "); 2613 } 2614 save_heap_summary(); 2615 reset(false); 2616 assert(_collectorState == Idling, "Collector state should " 2617 "have changed"); 2618 break; 2619 case Precleaning: 2620 case AbortablePreclean: 2621 // Elide the preclean phase 2622 _collectorState = FinalMarking; 2623 break; 2624 default: 2625 ShouldNotReachHere(); 2626 } 2627 if (TraceCMSState) { 2628 gclog_or_tty->print_cr(" Thread " INTPTR_FORMAT " done - next CMS state %d", 2629 Thread::current(), _collectorState); 2630 } 2631 } 2632 2633 if (UseAdaptiveSizePolicy) { 2634 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2635 size_policy()->ms_collection_end(gch->gc_cause()); 2636 } 2637 2638 if (VerifyAfterGC && 2639 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2640 Universe::verify(); 2641 } 2642 if (TraceCMSState) { 2643 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT 2644 " exiting collection CMS state %d", 2645 Thread::current(), _collectorState); 2646 } 2647 } 2648 2649 bool CMSCollector::waitForForegroundGC() { 2650 bool res = false; 2651 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 2652 "CMS thread should have CMS token"); 2653 // Block the foreground collector until the 2654 // background collectors decides whether to 2655 // yield. 2656 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 2657 _foregroundGCShouldWait = true; 2658 if (_foregroundGCIsActive) { 2659 // The background collector yields to the 2660 // foreground collector and returns a value 2661 // indicating that it has yielded. The foreground 2662 // collector can proceed. 2663 res = true; 2664 _foregroundGCShouldWait = false; 2665 ConcurrentMarkSweepThread::clear_CMS_flag( 2666 ConcurrentMarkSweepThread::CMS_cms_has_token); 2667 ConcurrentMarkSweepThread::set_CMS_flag( 2668 ConcurrentMarkSweepThread::CMS_cms_wants_token); 2669 // Get a possibly blocked foreground thread going 2670 CGC_lock->notify(); 2671 if (TraceCMSState) { 2672 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " waiting at CMS state %d", 2673 Thread::current(), _collectorState); 2674 } 2675 while (_foregroundGCIsActive) { 2676 CGC_lock->wait(Mutex::_no_safepoint_check_flag); 2677 } 2678 ConcurrentMarkSweepThread::set_CMS_flag( 2679 ConcurrentMarkSweepThread::CMS_cms_has_token); 2680 ConcurrentMarkSweepThread::clear_CMS_flag( 2681 ConcurrentMarkSweepThread::CMS_cms_wants_token); 2682 } 2683 if (TraceCMSState) { 2684 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " continuing at CMS state %d", 2685 Thread::current(), _collectorState); 2686 } 2687 return res; 2688 } 2689 2690 // Because of the need to lock the free lists and other structures in 2691 // the collector, common to all the generations that the collector is 2692 // collecting, we need the gc_prologues of individual CMS generations 2693 // delegate to their collector. It may have been simpler had the 2694 // current infrastructure allowed one to call a prologue on a 2695 // collector. In the absence of that we have the generation's 2696 // prologue delegate to the collector, which delegates back 2697 // some "local" work to a worker method in the individual generations 2698 // that it's responsible for collecting, while itself doing any 2699 // work common to all generations it's responsible for. A similar 2700 // comment applies to the gc_epilogue()'s. 2701 // The role of the variable _between_prologue_and_epilogue is to 2702 // enforce the invocation protocol. 2703 void CMSCollector::gc_prologue(bool full) { 2704 // Call gc_prologue_work() for the CMSGen 2705 // we are responsible for. 2706 2707 // The following locking discipline assumes that we are only called 2708 // when the world is stopped. 2709 assert(SafepointSynchronize::is_at_safepoint(), "world is stopped assumption"); 2710 2711 // The CMSCollector prologue must call the gc_prologues for the 2712 // "generations" that it's responsible 2713 // for. 2714 2715 assert( Thread::current()->is_VM_thread() 2716 || ( CMSScavengeBeforeRemark 2717 && Thread::current()->is_ConcurrentGC_thread()), 2718 "Incorrect thread type for prologue execution"); 2719 2720 if (_between_prologue_and_epilogue) { 2721 // We have already been invoked; this is a gc_prologue delegation 2722 // from yet another CMS generation that we are responsible for, just 2723 // ignore it since all relevant work has already been done. 2724 return; 2725 } 2726 2727 // set a bit saying prologue has been called; cleared in epilogue 2728 _between_prologue_and_epilogue = true; 2729 // Claim locks for common data structures, then call gc_prologue_work() 2730 // for each CMSGen. 2731 2732 getFreelistLocks(); // gets free list locks on constituent spaces 2733 bitMapLock()->lock_without_safepoint_check(); 2734 2735 // Should call gc_prologue_work() for all cms gens we are responsible for 2736 bool duringMarking = _collectorState >= Marking 2737 && _collectorState < Sweeping; 2738 2739 // The young collections clear the modified oops state, which tells if 2740 // there are any modified oops in the class. The remark phase also needs 2741 // that information. Tell the young collection to save the union of all 2742 // modified klasses. 2743 if (duringMarking) { 2744 _ct->klass_rem_set()->set_accumulate_modified_oops(true); 2745 } 2746 2747 bool registerClosure = duringMarking; 2748 2749 ModUnionClosure* muc = CollectedHeap::use_parallel_gc_threads() ? 2750 &_modUnionClosurePar 2751 : &_modUnionClosure; 2752 _cmsGen->gc_prologue_work(full, registerClosure, muc); 2753 2754 if (!full) { 2755 stats().record_gc0_begin(); 2756 } 2757 } 2758 2759 void ConcurrentMarkSweepGeneration::gc_prologue(bool full) { 2760 2761 _capacity_at_prologue = capacity(); 2762 _used_at_prologue = used(); 2763 2764 // Delegate to CMScollector which knows how to coordinate between 2765 // this and any other CMS generations that it is responsible for 2766 // collecting. 2767 collector()->gc_prologue(full); 2768 } 2769 2770 // This is a "private" interface for use by this generation's CMSCollector. 2771 // Not to be called directly by any other entity (for instance, 2772 // GenCollectedHeap, which calls the "public" gc_prologue method above). 2773 void ConcurrentMarkSweepGeneration::gc_prologue_work(bool full, 2774 bool registerClosure, ModUnionClosure* modUnionClosure) { 2775 assert(!incremental_collection_failed(), "Shouldn't be set yet"); 2776 assert(cmsSpace()->preconsumptionDirtyCardClosure() == NULL, 2777 "Should be NULL"); 2778 if (registerClosure) { 2779 cmsSpace()->setPreconsumptionDirtyCardClosure(modUnionClosure); 2780 } 2781 cmsSpace()->gc_prologue(); 2782 // Clear stat counters 2783 NOT_PRODUCT( 2784 assert(_numObjectsPromoted == 0, "check"); 2785 assert(_numWordsPromoted == 0, "check"); 2786 if (Verbose && PrintGC) { 2787 gclog_or_tty->print("Allocated "SIZE_FORMAT" objects, " 2788 SIZE_FORMAT" bytes concurrently", 2789 _numObjectsAllocated, _numWordsAllocated*sizeof(HeapWord)); 2790 } 2791 _numObjectsAllocated = 0; 2792 _numWordsAllocated = 0; 2793 ) 2794 } 2795 2796 void CMSCollector::gc_epilogue(bool full) { 2797 // The following locking discipline assumes that we are only called 2798 // when the world is stopped. 2799 assert(SafepointSynchronize::is_at_safepoint(), 2800 "world is stopped assumption"); 2801 2802 // Currently the CMS epilogue (see CompactibleFreeListSpace) merely checks 2803 // if linear allocation blocks need to be appropriately marked to allow the 2804 // the blocks to be parsable. We also check here whether we need to nudge the 2805 // CMS collector thread to start a new cycle (if it's not already active). 2806 assert( Thread::current()->is_VM_thread() 2807 || ( CMSScavengeBeforeRemark 2808 && Thread::current()->is_ConcurrentGC_thread()), 2809 "Incorrect thread type for epilogue execution"); 2810 2811 if (!_between_prologue_and_epilogue) { 2812 // We have already been invoked; this is a gc_epilogue delegation 2813 // from yet another CMS generation that we are responsible for, just 2814 // ignore it since all relevant work has already been done. 2815 return; 2816 } 2817 assert(haveFreelistLocks(), "must have freelist locks"); 2818 assert_lock_strong(bitMapLock()); 2819 2820 _ct->klass_rem_set()->set_accumulate_modified_oops(false); 2821 2822 _cmsGen->gc_epilogue_work(full); 2823 2824 if (_collectorState == AbortablePreclean || _collectorState == Precleaning) { 2825 // in case sampling was not already enabled, enable it 2826 _start_sampling = true; 2827 } 2828 // reset _eden_chunk_array so sampling starts afresh 2829 _eden_chunk_index = 0; 2830 2831 size_t cms_used = _cmsGen->cmsSpace()->used(); 2832 2833 // update performance counters - this uses a special version of 2834 // update_counters() that allows the utilization to be passed as a 2835 // parameter, avoiding multiple calls to used(). 2836 // 2837 _cmsGen->update_counters(cms_used); 2838 2839 if (CMSIncrementalMode) { 2840 icms_update_allocation_limits(); 2841 } 2842 2843 bitMapLock()->unlock(); 2844 releaseFreelistLocks(); 2845 2846 if (!CleanChunkPoolAsync) { 2847 Chunk::clean_chunk_pool(); 2848 } 2849 2850 set_did_compact(false); 2851 _between_prologue_and_epilogue = false; // ready for next cycle 2852 } 2853 2854 void ConcurrentMarkSweepGeneration::gc_epilogue(bool full) { 2855 collector()->gc_epilogue(full); 2856 2857 // Also reset promotion tracking in par gc thread states. 2858 if (CollectedHeap::use_parallel_gc_threads()) { 2859 for (uint i = 0; i < ParallelGCThreads; i++) { 2860 _par_gc_thread_states[i]->promo.stopTrackingPromotions(i); 2861 } 2862 } 2863 } 2864 2865 void ConcurrentMarkSweepGeneration::gc_epilogue_work(bool full) { 2866 assert(!incremental_collection_failed(), "Should have been cleared"); 2867 cmsSpace()->setPreconsumptionDirtyCardClosure(NULL); 2868 cmsSpace()->gc_epilogue(); 2869 // Print stat counters 2870 NOT_PRODUCT( 2871 assert(_numObjectsAllocated == 0, "check"); 2872 assert(_numWordsAllocated == 0, "check"); 2873 if (Verbose && PrintGC) { 2874 gclog_or_tty->print("Promoted "SIZE_FORMAT" objects, " 2875 SIZE_FORMAT" bytes", 2876 _numObjectsPromoted, _numWordsPromoted*sizeof(HeapWord)); 2877 } 2878 _numObjectsPromoted = 0; 2879 _numWordsPromoted = 0; 2880 ) 2881 2882 if (PrintGC && Verbose) { 2883 // Call down the chain in contiguous_available needs the freelistLock 2884 // so print this out before releasing the freeListLock. 2885 gclog_or_tty->print(" Contiguous available "SIZE_FORMAT" bytes ", 2886 contiguous_available()); 2887 } 2888 } 2889 2890 #ifndef PRODUCT 2891 bool CMSCollector::have_cms_token() { 2892 Thread* thr = Thread::current(); 2893 if (thr->is_VM_thread()) { 2894 return ConcurrentMarkSweepThread::vm_thread_has_cms_token(); 2895 } else if (thr->is_ConcurrentGC_thread()) { 2896 return ConcurrentMarkSweepThread::cms_thread_has_cms_token(); 2897 } else if (thr->is_GC_task_thread()) { 2898 return ConcurrentMarkSweepThread::vm_thread_has_cms_token() && 2899 ParGCRareEvent_lock->owned_by_self(); 2900 } 2901 return false; 2902 } 2903 #endif 2904 2905 // Check reachability of the given heap address in CMS generation, 2906 // treating all other generations as roots. 2907 bool CMSCollector::is_cms_reachable(HeapWord* addr) { 2908 // We could "guarantee" below, rather than assert, but I'll 2909 // leave these as "asserts" so that an adventurous debugger 2910 // could try this in the product build provided some subset of 2911 // the conditions were met, provided they were interested in the 2912 // results and knew that the computation below wouldn't interfere 2913 // with other concurrent computations mutating the structures 2914 // being read or written. 2915 assert(SafepointSynchronize::is_at_safepoint(), 2916 "Else mutations in object graph will make answer suspect"); 2917 assert(have_cms_token(), "Should hold cms token"); 2918 assert(haveFreelistLocks(), "must hold free list locks"); 2919 assert_lock_strong(bitMapLock()); 2920 2921 // Clear the marking bit map array before starting, but, just 2922 // for kicks, first report if the given address is already marked 2923 gclog_or_tty->print_cr("Start: Address " PTR_FORMAT " is%s marked", addr, 2924 _markBitMap.isMarked(addr) ? "" : " not"); 2925 2926 if (verify_after_remark()) { 2927 MutexLockerEx x(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag); 2928 bool result = verification_mark_bm()->isMarked(addr); 2929 gclog_or_tty->print_cr("TransitiveMark: Address " PTR_FORMAT " %s marked", addr, 2930 result ? "IS" : "is NOT"); 2931 return result; 2932 } else { 2933 gclog_or_tty->print_cr("Could not compute result"); 2934 return false; 2935 } 2936 } 2937 2938 2939 void 2940 CMSCollector::print_on_error(outputStream* st) { 2941 CMSCollector* collector = ConcurrentMarkSweepGeneration::_collector; 2942 if (collector != NULL) { 2943 CMSBitMap* bitmap = &collector->_markBitMap; 2944 st->print_cr("Marking Bits: (CMSBitMap*) " PTR_FORMAT, bitmap); 2945 bitmap->print_on_error(st, " Bits: "); 2946 2947 st->cr(); 2948 2949 CMSBitMap* mut_bitmap = &collector->_modUnionTable; 2950 st->print_cr("Mod Union Table: (CMSBitMap*) " PTR_FORMAT, mut_bitmap); 2951 mut_bitmap->print_on_error(st, " Bits: "); 2952 } 2953 } 2954 2955 //////////////////////////////////////////////////////// 2956 // CMS Verification Support 2957 //////////////////////////////////////////////////////// 2958 // Following the remark phase, the following invariant 2959 // should hold -- each object in the CMS heap which is 2960 // marked in markBitMap() should be marked in the verification_mark_bm(). 2961 2962 class VerifyMarkedClosure: public BitMapClosure { 2963 CMSBitMap* _marks; 2964 bool _failed; 2965 2966 public: 2967 VerifyMarkedClosure(CMSBitMap* bm): _marks(bm), _failed(false) {} 2968 2969 bool do_bit(size_t offset) { 2970 HeapWord* addr = _marks->offsetToHeapWord(offset); 2971 if (!_marks->isMarked(addr)) { 2972 oop(addr)->print_on(gclog_or_tty); 2973 gclog_or_tty->print_cr(" ("INTPTR_FORMAT" should have been marked)", addr); 2974 _failed = true; 2975 } 2976 return true; 2977 } 2978 2979 bool failed() { return _failed; } 2980 }; 2981 2982 bool CMSCollector::verify_after_remark(bool silent) { 2983 if (!silent) gclog_or_tty->print(" [Verifying CMS Marking... "); 2984 MutexLockerEx ml(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag); 2985 static bool init = false; 2986 2987 assert(SafepointSynchronize::is_at_safepoint(), 2988 "Else mutations in object graph will make answer suspect"); 2989 assert(have_cms_token(), 2990 "Else there may be mutual interference in use of " 2991 " verification data structures"); 2992 assert(_collectorState > Marking && _collectorState <= Sweeping, 2993 "Else marking info checked here may be obsolete"); 2994 assert(haveFreelistLocks(), "must hold free list locks"); 2995 assert_lock_strong(bitMapLock()); 2996 2997 2998 // Allocate marking bit map if not already allocated 2999 if (!init) { // first time 3000 if (!verification_mark_bm()->allocate(_span)) { 3001 return false; 3002 } 3003 init = true; 3004 } 3005 3006 assert(verification_mark_stack()->isEmpty(), "Should be empty"); 3007 3008 // Turn off refs discovery -- so we will be tracing through refs. 3009 // This is as intended, because by this time 3010 // GC must already have cleared any refs that need to be cleared, 3011 // and traced those that need to be marked; moreover, 3012 // the marking done here is not going to interfere in any 3013 // way with the marking information used by GC. 3014 NoRefDiscovery no_discovery(ref_processor()); 3015 3016 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;) 3017 3018 // Clear any marks from a previous round 3019 verification_mark_bm()->clear_all(); 3020 assert(verification_mark_stack()->isEmpty(), "markStack should be empty"); 3021 verify_work_stacks_empty(); 3022 3023 GenCollectedHeap* gch = GenCollectedHeap::heap(); 3024 gch->ensure_parsability(false); // fill TLABs, but no need to retire them 3025 // Update the saved marks which may affect the root scans. 3026 gch->save_marks(); 3027 3028 if (CMSRemarkVerifyVariant == 1) { 3029 // In this first variant of verification, we complete 3030 // all marking, then check if the new marks-vector is 3031 // a subset of the CMS marks-vector. 3032 verify_after_remark_work_1(); 3033 } else if (CMSRemarkVerifyVariant == 2) { 3034 // In this second variant of verification, we flag an error 3035 // (i.e. an object reachable in the new marks-vector not reachable 3036 // in the CMS marks-vector) immediately, also indicating the 3037 // identify of an object (A) that references the unmarked object (B) -- 3038 // presumably, a mutation to A failed to be picked up by preclean/remark? 3039 verify_after_remark_work_2(); 3040 } else { 3041 warning("Unrecognized value %d for CMSRemarkVerifyVariant", 3042 CMSRemarkVerifyVariant); 3043 } 3044 if (!silent) gclog_or_tty->print(" done] "); 3045 return true; 3046 } 3047 3048 void CMSCollector::verify_after_remark_work_1() { 3049 ResourceMark rm; 3050 HandleMark hm; 3051 GenCollectedHeap* gch = GenCollectedHeap::heap(); 3052 3053 // Get a clear set of claim bits for the strong roots processing to work with. 3054 ClassLoaderDataGraph::clear_claimed_marks(); 3055 3056 // Mark from roots one level into CMS 3057 MarkRefsIntoClosure notOlder(_span, verification_mark_bm()); 3058 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 3059 3060 gch->gen_process_strong_roots(_cmsGen->level(), 3061 true, // younger gens are roots 3062 true, // activate StrongRootsScope 3063 SharedHeap::ScanningOption(roots_scanning_options()), 3064 ¬Older, 3065 NULL, 3066 NULL); // SSS: Provide correct closure 3067 3068 // Now mark from the roots 3069 MarkFromRootsClosure markFromRootsClosure(this, _span, 3070 verification_mark_bm(), verification_mark_stack(), 3071 false /* don't yield */, true /* verifying */); 3072 assert(_restart_addr == NULL, "Expected pre-condition"); 3073 verification_mark_bm()->iterate(&markFromRootsClosure); 3074 while (_restart_addr != NULL) { 3075 // Deal with stack overflow: by restarting at the indicated 3076 // address. 3077 HeapWord* ra = _restart_addr; 3078 markFromRootsClosure.reset(ra); 3079 _restart_addr = NULL; 3080 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end()); 3081 } 3082 assert(verification_mark_stack()->isEmpty(), "Should have been drained"); 3083 verify_work_stacks_empty(); 3084 3085 // Marking completed -- now verify that each bit marked in 3086 // verification_mark_bm() is also marked in markBitMap(); flag all 3087 // errors by printing corresponding objects. 3088 VerifyMarkedClosure vcl(markBitMap()); 3089 verification_mark_bm()->iterate(&vcl); 3090 if (vcl.failed()) { 3091 gclog_or_tty->print("Verification failed"); 3092 Universe::heap()->print_on(gclog_or_tty); 3093 fatal("CMS: failed marking verification after remark"); 3094 } 3095 } 3096 3097 class VerifyKlassOopsKlassClosure : public KlassClosure { 3098 class VerifyKlassOopsClosure : public OopClosure { 3099 CMSBitMap* _bitmap; 3100 public: 3101 VerifyKlassOopsClosure(CMSBitMap* bitmap) : _bitmap(bitmap) { } 3102 void do_oop(oop* p) { guarantee(*p == NULL || _bitmap->isMarked((HeapWord*) *p), "Should be marked"); } 3103 void do_oop(narrowOop* p) { ShouldNotReachHere(); } 3104 } _oop_closure; 3105 public: 3106 VerifyKlassOopsKlassClosure(CMSBitMap* bitmap) : _oop_closure(bitmap) {} 3107 void do_klass(Klass* k) { 3108 k->oops_do(&_oop_closure); 3109 } 3110 }; 3111 3112 void CMSCollector::verify_after_remark_work_2() { 3113 ResourceMark rm; 3114 HandleMark hm; 3115 GenCollectedHeap* gch = GenCollectedHeap::heap(); 3116 3117 // Get a clear set of claim bits for the strong roots processing to work with. 3118 ClassLoaderDataGraph::clear_claimed_marks(); 3119 3120 // Mark from roots one level into CMS 3121 MarkRefsIntoVerifyClosure notOlder(_span, verification_mark_bm(), 3122 markBitMap()); 3123 CMKlassClosure klass_closure(¬Older); 3124 3125 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 3126 gch->gen_process_strong_roots(_cmsGen->level(), 3127 true, // younger gens are roots 3128 true, // activate StrongRootsScope 3129 SharedHeap::ScanningOption(roots_scanning_options()), 3130 ¬Older, 3131 NULL, 3132 &klass_closure); 3133 3134 // Now mark from the roots 3135 MarkFromRootsVerifyClosure markFromRootsClosure(this, _span, 3136 verification_mark_bm(), markBitMap(), verification_mark_stack()); 3137 assert(_restart_addr == NULL, "Expected pre-condition"); 3138 verification_mark_bm()->iterate(&markFromRootsClosure); 3139 while (_restart_addr != NULL) { 3140 // Deal with stack overflow: by restarting at the indicated 3141 // address. 3142 HeapWord* ra = _restart_addr; 3143 markFromRootsClosure.reset(ra); 3144 _restart_addr = NULL; 3145 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end()); 3146 } 3147 assert(verification_mark_stack()->isEmpty(), "Should have been drained"); 3148 verify_work_stacks_empty(); 3149 3150 VerifyKlassOopsKlassClosure verify_klass_oops(verification_mark_bm()); 3151 ClassLoaderDataGraph::classes_do(&verify_klass_oops); 3152 3153 // Marking completed -- now verify that each bit marked in 3154 // verification_mark_bm() is also marked in markBitMap(); flag all 3155 // errors by printing corresponding objects. 3156 VerifyMarkedClosure vcl(markBitMap()); 3157 verification_mark_bm()->iterate(&vcl); 3158 assert(!vcl.failed(), "Else verification above should not have succeeded"); 3159 } 3160 3161 void ConcurrentMarkSweepGeneration::save_marks() { 3162 // delegate to CMS space 3163 cmsSpace()->save_marks(); 3164 for (uint i = 0; i < ParallelGCThreads; i++) { 3165 _par_gc_thread_states[i]->promo.startTrackingPromotions(); 3166 } 3167 } 3168 3169 bool ConcurrentMarkSweepGeneration::no_allocs_since_save_marks() { 3170 return cmsSpace()->no_allocs_since_save_marks(); 3171 } 3172 3173 #define CMS_SINCE_SAVE_MARKS_DEFN(OopClosureType, nv_suffix) \ 3174 \ 3175 void ConcurrentMarkSweepGeneration:: \ 3176 oop_since_save_marks_iterate##nv_suffix(OopClosureType* cl) { \ 3177 cl->set_generation(this); \ 3178 cmsSpace()->oop_since_save_marks_iterate##nv_suffix(cl); \ 3179 cl->reset_generation(); \ 3180 save_marks(); \ 3181 } 3182 3183 ALL_SINCE_SAVE_MARKS_CLOSURES(CMS_SINCE_SAVE_MARKS_DEFN) 3184 3185 void 3186 ConcurrentMarkSweepGeneration::younger_refs_iterate(OopsInGenClosure* cl) { 3187 cl->set_generation(this); 3188 younger_refs_in_space_iterate(_cmsSpace, cl); 3189 cl->reset_generation(); 3190 } 3191 3192 void 3193 ConcurrentMarkSweepGeneration::oop_iterate(ExtendedOopClosure* cl) { 3194 if (freelistLock()->owned_by_self()) { 3195 Generation::oop_iterate(cl); 3196 } else { 3197 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 3198 Generation::oop_iterate(cl); 3199 } 3200 } 3201 3202 void 3203 ConcurrentMarkSweepGeneration::object_iterate(ObjectClosure* cl) { 3204 if (freelistLock()->owned_by_self()) { 3205 Generation::object_iterate(cl); 3206 } else { 3207 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 3208 Generation::object_iterate(cl); 3209 } 3210 } 3211 3212 void 3213 ConcurrentMarkSweepGeneration::safe_object_iterate(ObjectClosure* cl) { 3214 if (freelistLock()->owned_by_self()) { 3215 Generation::safe_object_iterate(cl); 3216 } else { 3217 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 3218 Generation::safe_object_iterate(cl); 3219 } 3220 } 3221 3222 void 3223 ConcurrentMarkSweepGeneration::post_compact() { 3224 } 3225 3226 void 3227 ConcurrentMarkSweepGeneration::prepare_for_verify() { 3228 // Fix the linear allocation blocks to look like free blocks. 3229 3230 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those 3231 // are not called when the heap is verified during universe initialization and 3232 // at vm shutdown. 3233 if (freelistLock()->owned_by_self()) { 3234 cmsSpace()->prepare_for_verify(); 3235 } else { 3236 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag); 3237 cmsSpace()->prepare_for_verify(); 3238 } 3239 } 3240 3241 void 3242 ConcurrentMarkSweepGeneration::verify() { 3243 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those 3244 // are not called when the heap is verified during universe initialization and 3245 // at vm shutdown. 3246 if (freelistLock()->owned_by_self()) { 3247 cmsSpace()->verify(); 3248 } else { 3249 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag); 3250 cmsSpace()->verify(); 3251 } 3252 } 3253 3254 void CMSCollector::verify() { 3255 _cmsGen->verify(); 3256 } 3257 3258 #ifndef PRODUCT 3259 bool CMSCollector::overflow_list_is_empty() const { 3260 assert(_num_par_pushes >= 0, "Inconsistency"); 3261 if (_overflow_list == NULL) { 3262 assert(_num_par_pushes == 0, "Inconsistency"); 3263 } 3264 return _overflow_list == NULL; 3265 } 3266 3267 // The methods verify_work_stacks_empty() and verify_overflow_empty() 3268 // merely consolidate assertion checks that appear to occur together frequently. 3269 void CMSCollector::verify_work_stacks_empty() const { 3270 assert(_markStack.isEmpty(), "Marking stack should be empty"); 3271 assert(overflow_list_is_empty(), "Overflow list should be empty"); 3272 } 3273 3274 void CMSCollector::verify_overflow_empty() const { 3275 assert(overflow_list_is_empty(), "Overflow list should be empty"); 3276 assert(no_preserved_marks(), "No preserved marks"); 3277 } 3278 #endif // PRODUCT 3279 3280 // Decide if we want to enable class unloading as part of the 3281 // ensuing concurrent GC cycle. We will collect and 3282 // unload classes if it's the case that: 3283 // (1) an explicit gc request has been made and the flag 3284 // ExplicitGCInvokesConcurrentAndUnloadsClasses is set, OR 3285 // (2) (a) class unloading is enabled at the command line, and 3286 // (b) old gen is getting really full 3287 // NOTE: Provided there is no change in the state of the heap between 3288 // calls to this method, it should have idempotent results. Moreover, 3289 // its results should be monotonically increasing (i.e. going from 0 to 1, 3290 // but not 1 to 0) between successive calls between which the heap was 3291 // not collected. For the implementation below, it must thus rely on 3292 // the property that concurrent_cycles_since_last_unload() 3293 // will not decrease unless a collection cycle happened and that 3294 // _cmsGen->is_too_full() are 3295 // themselves also monotonic in that sense. See check_monotonicity() 3296 // below. 3297 void CMSCollector::update_should_unload_classes() { 3298 _should_unload_classes = false; 3299 // Condition 1 above 3300 if (_full_gc_requested && ExplicitGCInvokesConcurrentAndUnloadsClasses) { 3301 _should_unload_classes = true; 3302 } else if (CMSClassUnloadingEnabled) { // Condition 2.a above 3303 // Disjuncts 2.b.(i,ii,iii) above 3304 _should_unload_classes = (concurrent_cycles_since_last_unload() >= 3305 CMSClassUnloadingMaxInterval) 3306 || _cmsGen->is_too_full(); 3307 } 3308 } 3309 3310 bool ConcurrentMarkSweepGeneration::is_too_full() const { 3311 bool res = should_concurrent_collect(); 3312 res = res && (occupancy() > (double)CMSIsTooFullPercentage/100.0); 3313 return res; 3314 } 3315 3316 void CMSCollector::setup_cms_unloading_and_verification_state() { 3317 const bool should_verify = VerifyBeforeGC || VerifyAfterGC || VerifyDuringGC 3318 || VerifyBeforeExit; 3319 const int rso = SharedHeap::SO_Strings | SharedHeap::SO_AllCodeCache; 3320 3321 // We set the proper root for this CMS cycle here. 3322 if (should_unload_classes()) { // Should unload classes this cycle 3323 remove_root_scanning_option(SharedHeap::SO_AllClasses); 3324 add_root_scanning_option(SharedHeap::SO_SystemClasses); 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(), "Inconsistency!"); 3332 remove_root_scanning_option(SharedHeap::SO_SystemClasses); 3333 add_root_scanning_option(SharedHeap::SO_AllClasses); 3334 3335 if ((!verifying() || unloaded_classes_last_cycle()) && should_verify) { 3336 // Include symbols, strings and code cache elements to prevent their resurrection. 3337 add_root_scanning_option(rso); 3338 set_verifying(true); 3339 } else if (verifying() && !should_verify) { 3340 // We were verifying, but some verification flags got disabled. 3341 set_verifying(false); 3342 // Exclude symbols, strings and code cache elements from root scanning to 3343 // reduce IM and RM pauses. 3344 remove_root_scanning_option(rso); 3345 } 3346 } 3347 3348 3349 #ifndef PRODUCT 3350 HeapWord* CMSCollector::block_start(const void* p) const { 3351 const HeapWord* addr = (HeapWord*)p; 3352 if (_span.contains(p)) { 3353 if (_cmsGen->cmsSpace()->is_in_reserved(addr)) { 3354 return _cmsGen->cmsSpace()->block_start(p); 3355 } 3356 } 3357 return NULL; 3358 } 3359 #endif 3360 3361 HeapWord* 3362 ConcurrentMarkSweepGeneration::expand_and_allocate(size_t word_size, 3363 bool tlab, 3364 bool parallel) { 3365 CMSSynchronousYieldRequest yr; 3366 assert(!tlab, "Can't deal with TLAB allocation"); 3367 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 3368 expand(word_size*HeapWordSize, MinHeapDeltaBytes, 3369 CMSExpansionCause::_satisfy_allocation); 3370 if (GCExpandToAllocateDelayMillis > 0) { 3371 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 3372 } 3373 return have_lock_and_allocate(word_size, tlab); 3374 } 3375 3376 // YSR: All of this generation expansion/shrinking stuff is an exact copy of 3377 // OneContigSpaceCardGeneration, which makes me wonder if we should move this 3378 // to CardGeneration and share it... 3379 bool ConcurrentMarkSweepGeneration::expand(size_t bytes, size_t expand_bytes) { 3380 return CardGeneration::expand(bytes, expand_bytes); 3381 } 3382 3383 void ConcurrentMarkSweepGeneration::expand(size_t bytes, size_t expand_bytes, 3384 CMSExpansionCause::Cause cause) 3385 { 3386 3387 bool success = expand(bytes, expand_bytes); 3388 3389 // remember why we expanded; this information is used 3390 // by shouldConcurrentCollect() when making decisions on whether to start 3391 // a new CMS cycle. 3392 if (success) { 3393 set_expansion_cause(cause); 3394 if (PrintGCDetails && Verbose) { 3395 gclog_or_tty->print_cr("Expanded CMS gen for %s", 3396 CMSExpansionCause::to_string(cause)); 3397 } 3398 } 3399 } 3400 3401 HeapWord* ConcurrentMarkSweepGeneration::expand_and_par_lab_allocate(CMSParGCThreadState* ps, size_t word_sz) { 3402 HeapWord* res = NULL; 3403 MutexLocker x(ParGCRareEvent_lock); 3404 while (true) { 3405 // Expansion by some other thread might make alloc OK now: 3406 res = ps->lab.alloc(word_sz); 3407 if (res != NULL) return res; 3408 // If there's not enough expansion space available, give up. 3409 if (_virtual_space.uncommitted_size() < (word_sz * HeapWordSize)) { 3410 return NULL; 3411 } 3412 // Otherwise, we try expansion. 3413 expand(word_sz*HeapWordSize, MinHeapDeltaBytes, 3414 CMSExpansionCause::_allocate_par_lab); 3415 // Now go around the loop and try alloc again; 3416 // A competing par_promote might beat us to the expansion space, 3417 // so we may go around the loop again if promotion fails again. 3418 if (GCExpandToAllocateDelayMillis > 0) { 3419 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 3420 } 3421 } 3422 } 3423 3424 3425 bool ConcurrentMarkSweepGeneration::expand_and_ensure_spooling_space( 3426 PromotionInfo* promo) { 3427 MutexLocker x(ParGCRareEvent_lock); 3428 size_t refill_size_bytes = promo->refillSize() * HeapWordSize; 3429 while (true) { 3430 // Expansion by some other thread might make alloc OK now: 3431 if (promo->ensure_spooling_space()) { 3432 assert(promo->has_spooling_space(), 3433 "Post-condition of successful ensure_spooling_space()"); 3434 return true; 3435 } 3436 // If there's not enough expansion space available, give up. 3437 if (_virtual_space.uncommitted_size() < refill_size_bytes) { 3438 return false; 3439 } 3440 // Otherwise, we try expansion. 3441 expand(refill_size_bytes, MinHeapDeltaBytes, 3442 CMSExpansionCause::_allocate_par_spooling_space); 3443 // Now go around the loop and try alloc again; 3444 // A competing allocation might beat us to the expansion space, 3445 // so we may go around the loop again if allocation fails again. 3446 if (GCExpandToAllocateDelayMillis > 0) { 3447 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 3448 } 3449 } 3450 } 3451 3452 3453 void ConcurrentMarkSweepGeneration::shrink_by(size_t bytes) { 3454 assert_locked_or_safepoint(ExpandHeap_lock); 3455 // Shrink committed space 3456 _virtual_space.shrink_by(bytes); 3457 // Shrink space; this also shrinks the space's BOT 3458 _cmsSpace->set_end((HeapWord*) _virtual_space.high()); 3459 size_t new_word_size = heap_word_size(_cmsSpace->capacity()); 3460 // Shrink the shared block offset array 3461 _bts->resize(new_word_size); 3462 MemRegion mr(_cmsSpace->bottom(), new_word_size); 3463 // Shrink the card table 3464 Universe::heap()->barrier_set()->resize_covered_region(mr); 3465 3466 if (Verbose && PrintGC) { 3467 size_t new_mem_size = _virtual_space.committed_size(); 3468 size_t old_mem_size = new_mem_size + bytes; 3469 gclog_or_tty->print_cr("Shrinking %s from " SIZE_FORMAT "K to " SIZE_FORMAT "K", 3470 name(), old_mem_size/K, new_mem_size/K); 3471 } 3472 } 3473 3474 void ConcurrentMarkSweepGeneration::shrink(size_t bytes) { 3475 assert_locked_or_safepoint(Heap_lock); 3476 size_t size = ReservedSpace::page_align_size_down(bytes); 3477 // Only shrink if a compaction was done so that all the free space 3478 // in the generation is in a contiguous block at the end. 3479 if (size > 0 && did_compact()) { 3480 shrink_by(size); 3481 } 3482 } 3483 3484 bool ConcurrentMarkSweepGeneration::grow_by(size_t bytes) { 3485 assert_locked_or_safepoint(Heap_lock); 3486 bool result = _virtual_space.expand_by(bytes); 3487 if (result) { 3488 size_t new_word_size = 3489 heap_word_size(_virtual_space.committed_size()); 3490 MemRegion mr(_cmsSpace->bottom(), new_word_size); 3491 _bts->resize(new_word_size); // resize the block offset shared array 3492 Universe::heap()->barrier_set()->resize_covered_region(mr); 3493 // Hmmmm... why doesn't CFLS::set_end verify locking? 3494 // This is quite ugly; FIX ME XXX 3495 _cmsSpace->assert_locked(freelistLock()); 3496 _cmsSpace->set_end((HeapWord*)_virtual_space.high()); 3497 3498 // update the space and generation capacity counters 3499 if (UsePerfData) { 3500 _space_counters->update_capacity(); 3501 _gen_counters->update_all(); 3502 } 3503 3504 if (Verbose && PrintGC) { 3505 size_t new_mem_size = _virtual_space.committed_size(); 3506 size_t old_mem_size = new_mem_size - bytes; 3507 gclog_or_tty->print_cr("Expanding %s from " SIZE_FORMAT "K by " SIZE_FORMAT "K to " SIZE_FORMAT "K", 3508 name(), old_mem_size/K, bytes/K, new_mem_size/K); 3509 } 3510 } 3511 return result; 3512 } 3513 3514 bool ConcurrentMarkSweepGeneration::grow_to_reserved() { 3515 assert_locked_or_safepoint(Heap_lock); 3516 bool success = true; 3517 const size_t remaining_bytes = _virtual_space.uncommitted_size(); 3518 if (remaining_bytes > 0) { 3519 success = grow_by(remaining_bytes); 3520 DEBUG_ONLY(if (!success) warning("grow to reserved failed");) 3521 } 3522 return success; 3523 } 3524 3525 void ConcurrentMarkSweepGeneration::shrink_free_list_by(size_t bytes) { 3526 assert_locked_or_safepoint(Heap_lock); 3527 assert_lock_strong(freelistLock()); 3528 if (PrintGCDetails && Verbose) { 3529 warning("Shrinking of CMS not yet implemented"); 3530 } 3531 return; 3532 } 3533 3534 3535 // Simple ctor/dtor wrapper for accounting & timer chores around concurrent 3536 // phases. 3537 class CMSPhaseAccounting: public StackObj { 3538 public: 3539 CMSPhaseAccounting(CMSCollector *collector, 3540 const char *phase, 3541 bool print_cr = true); 3542 ~CMSPhaseAccounting(); 3543 3544 private: 3545 CMSCollector *_collector; 3546 const char *_phase; 3547 elapsedTimer _wallclock; 3548 bool _print_cr; 3549 3550 public: 3551 // Not MT-safe; so do not pass around these StackObj's 3552 // where they may be accessed by other threads. 3553 jlong wallclock_millis() { 3554 assert(_wallclock.is_active(), "Wall clock should not stop"); 3555 _wallclock.stop(); // to record time 3556 jlong ret = _wallclock.milliseconds(); 3557 _wallclock.start(); // restart 3558 return ret; 3559 } 3560 }; 3561 3562 CMSPhaseAccounting::CMSPhaseAccounting(CMSCollector *collector, 3563 const char *phase, 3564 bool print_cr) : 3565 _collector(collector), _phase(phase), _print_cr(print_cr) { 3566 3567 if (PrintCMSStatistics != 0) { 3568 _collector->resetYields(); 3569 } 3570 if (PrintGCDetails) { 3571 gclog_or_tty->date_stamp(PrintGCDateStamps); 3572 gclog_or_tty->stamp(PrintGCTimeStamps); 3573 gclog_or_tty->print_cr("[%s-concurrent-%s-start]", 3574 _collector->cmsGen()->short_name(), _phase); 3575 } 3576 _collector->resetTimer(); 3577 _wallclock.start(); 3578 _collector->startTimer(); 3579 } 3580 3581 CMSPhaseAccounting::~CMSPhaseAccounting() { 3582 assert(_wallclock.is_active(), "Wall clock should not have stopped"); 3583 _collector->stopTimer(); 3584 _wallclock.stop(); 3585 if (PrintGCDetails) { 3586 gclog_or_tty->date_stamp(PrintGCDateStamps); 3587 gclog_or_tty->stamp(PrintGCTimeStamps); 3588 gclog_or_tty->print("[%s-concurrent-%s: %3.3f/%3.3f secs]", 3589 _collector->cmsGen()->short_name(), 3590 _phase, _collector->timerValue(), _wallclock.seconds()); 3591 if (_print_cr) { 3592 gclog_or_tty->cr(); 3593 } 3594 if (PrintCMSStatistics != 0) { 3595 gclog_or_tty->print_cr(" (CMS-concurrent-%s yielded %d times)", _phase, 3596 _collector->yields()); 3597 } 3598 } 3599 } 3600 3601 // CMS work 3602 3603 // The common parts of CMSParInitialMarkTask and CMSParRemarkTask. 3604 class CMSParMarkTask : public AbstractGangTask { 3605 protected: 3606 CMSCollector* _collector; 3607 int _n_workers; 3608 CMSParMarkTask(const char* name, CMSCollector* collector, int n_workers) : 3609 AbstractGangTask(name), 3610 _collector(collector), 3611 _n_workers(n_workers) {} 3612 // Work method in support of parallel rescan ... of young gen spaces 3613 void do_young_space_rescan(uint worker_id, OopsInGenClosure* cl, 3614 ContiguousSpace* space, 3615 HeapWord** chunk_array, size_t chunk_top); 3616 void work_on_young_gen_roots(uint worker_id, OopsInGenClosure* cl); 3617 }; 3618 3619 // Parallel initial mark task 3620 class CMSParInitialMarkTask: public CMSParMarkTask { 3621 public: 3622 CMSParInitialMarkTask(CMSCollector* collector, int n_workers) : 3623 CMSParMarkTask("Scan roots and young gen for initial mark in parallel", 3624 collector, n_workers) {} 3625 void work(uint worker_id); 3626 }; 3627 3628 // Checkpoint the roots into this generation from outside 3629 // this generation. [Note this initial checkpoint need only 3630 // be approximate -- we'll do a catch up phase subsequently.] 3631 void CMSCollector::checkpointRootsInitial(bool asynch) { 3632 assert(_collectorState == InitialMarking, "Wrong collector state"); 3633 check_correct_thread_executing(); 3634 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 3635 3636 save_heap_summary(); 3637 report_heap_summary(GCWhen::BeforeGC); 3638 3639 ReferenceProcessor* rp = ref_processor(); 3640 SpecializationStats::clear(); 3641 assert(_restart_addr == NULL, "Control point invariant"); 3642 if (asynch) { 3643 // acquire locks for subsequent manipulations 3644 MutexLockerEx x(bitMapLock(), 3645 Mutex::_no_safepoint_check_flag); 3646 checkpointRootsInitialWork(asynch); 3647 // enable ("weak") refs discovery 3648 rp->enable_discovery(true /*verify_disabled*/, true /*check_no_refs*/); 3649 _collectorState = Marking; 3650 } else { 3651 // (Weak) Refs discovery: this is controlled from genCollectedHeap::do_collection 3652 // which recognizes if we are a CMS generation, and doesn't try to turn on 3653 // discovery; verify that they aren't meddling. 3654 assert(!rp->discovery_is_atomic(), 3655 "incorrect setting of discovery predicate"); 3656 assert(!rp->discovery_enabled(), "genCollectedHeap shouldn't control " 3657 "ref discovery for this generation kind"); 3658 // already have locks 3659 checkpointRootsInitialWork(asynch); 3660 // now enable ("weak") refs discovery 3661 rp->enable_discovery(true /*verify_disabled*/, false /*verify_no_refs*/); 3662 _collectorState = Marking; 3663 } 3664 SpecializationStats::print(); 3665 } 3666 3667 void CMSCollector::checkpointRootsInitialWork(bool asynch) { 3668 assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped"); 3669 assert(_collectorState == InitialMarking, "just checking"); 3670 3671 // If there has not been a GC[n-1] since last GC[n] cycle completed, 3672 // precede our marking with a collection of all 3673 // younger generations to keep floating garbage to a minimum. 3674 // XXX: we won't do this for now -- it's an optimization to be done later. 3675 3676 // already have locks 3677 assert_lock_strong(bitMapLock()); 3678 assert(_markBitMap.isAllClear(), "was reset at end of previous cycle"); 3679 3680 // Setup the verification and class unloading state for this 3681 // CMS collection cycle. 3682 setup_cms_unloading_and_verification_state(); 3683 3684 NOT_PRODUCT(GCTraceTime t("\ncheckpointRootsInitialWork", 3685 PrintGCDetails && Verbose, true, _gc_timer_cm);) 3686 if (UseAdaptiveSizePolicy) { 3687 size_policy()->checkpoint_roots_initial_begin(); 3688 } 3689 3690 // Reset all the PLAB chunk arrays if necessary. 3691 if (_survivor_plab_array != NULL && !CMSPLABRecordAlways) { 3692 reset_survivor_plab_arrays(); 3693 } 3694 3695 ResourceMark rm; 3696 HandleMark hm; 3697 3698 MarkRefsIntoClosure notOlder(_span, &_markBitMap); 3699 GenCollectedHeap* gch = GenCollectedHeap::heap(); 3700 3701 verify_work_stacks_empty(); 3702 verify_overflow_empty(); 3703 3704 gch->ensure_parsability(false); // fill TLABs, but no need to retire them 3705 // Update the saved marks which may affect the root scans. 3706 gch->save_marks(); 3707 3708 // weak reference processing has not started yet. 3709 ref_processor()->set_enqueuing_is_done(false); 3710 3711 // Need to remember all newly created CLDs, 3712 // so that we can guarantee that the remark finds them. 3713 ClassLoaderDataGraph::remember_new_clds(true); 3714 3715 // Whenever a CLD is found, it will be claimed before proceeding to mark 3716 // the klasses. The claimed marks need to be cleared before marking starts. 3717 ClassLoaderDataGraph::clear_claimed_marks(); 3718 3719 if (CMSPrintEdenSurvivorChunks) { 3720 print_eden_and_survivor_chunk_arrays(); 3721 } 3722 3723 { 3724 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;) 3725 if (CMSParallelInitialMarkEnabled && CollectedHeap::use_parallel_gc_threads()) { 3726 // The parallel version. 3727 FlexibleWorkGang* workers = gch->workers(); 3728 assert(workers != NULL, "Need parallel worker threads."); 3729 int n_workers = workers->active_workers(); 3730 CMSParInitialMarkTask tsk(this, n_workers); 3731 gch->set_par_threads(n_workers); 3732 initialize_sequential_subtasks_for_young_gen_rescan(n_workers); 3733 if (n_workers > 1) { 3734 GenCollectedHeap::StrongRootsScope srs(gch); 3735 workers->run_task(&tsk); 3736 } else { 3737 GenCollectedHeap::StrongRootsScope srs(gch); 3738 tsk.work(0); 3739 } 3740 gch->set_par_threads(0); 3741 } else { 3742 // The serial version. 3743 CMKlassClosure klass_closure(¬Older); 3744 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 3745 gch->gen_process_strong_roots(_cmsGen->level(), 3746 true, // younger gens are roots 3747 true, // activate StrongRootsScope 3748 SharedHeap::ScanningOption(roots_scanning_options()), 3749 ¬Older, 3750 NULL, 3751 &klass_closure); 3752 } 3753 } 3754 3755 // Clear mod-union table; it will be dirtied in the prologue of 3756 // CMS generation per each younger generation collection. 3757 3758 assert(_modUnionTable.isAllClear(), 3759 "Was cleared in most recent final checkpoint phase" 3760 " or no bits are set in the gc_prologue before the start of the next " 3761 "subsequent marking phase."); 3762 3763 assert(_ct->klass_rem_set()->mod_union_is_clear(), "Must be"); 3764 3765 // Save the end of the used_region of the constituent generations 3766 // to be used to limit the extent of sweep in each generation. 3767 save_sweep_limits(); 3768 if (UseAdaptiveSizePolicy) { 3769 size_policy()->checkpoint_roots_initial_end(gch->gc_cause()); 3770 } 3771 verify_overflow_empty(); 3772 } 3773 3774 bool CMSCollector::markFromRoots(bool asynch) { 3775 // we might be tempted to assert that: 3776 // assert(asynch == !SafepointSynchronize::is_at_safepoint(), 3777 // "inconsistent argument?"); 3778 // However that wouldn't be right, because it's possible that 3779 // a safepoint is indeed in progress as a younger generation 3780 // stop-the-world GC happens even as we mark in this generation. 3781 assert(_collectorState == Marking, "inconsistent state?"); 3782 check_correct_thread_executing(); 3783 verify_overflow_empty(); 3784 3785 bool res; 3786 if (asynch) { 3787 3788 // Start the timers for adaptive size policy for the concurrent phases 3789 // Do it here so that the foreground MS can use the concurrent 3790 // timer since a foreground MS might has the sweep done concurrently 3791 // or STW. 3792 if (UseAdaptiveSizePolicy) { 3793 size_policy()->concurrent_marking_begin(); 3794 } 3795 3796 // Weak ref discovery note: We may be discovering weak 3797 // refs in this generation concurrent (but interleaved) with 3798 // weak ref discovery by a younger generation collector. 3799 3800 CMSTokenSyncWithLocks ts(true, bitMapLock()); 3801 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 3802 CMSPhaseAccounting pa(this, "mark", !PrintGCDetails); 3803 res = markFromRootsWork(asynch); 3804 if (res) { 3805 _collectorState = Precleaning; 3806 } else { // We failed and a foreground collection wants to take over 3807 assert(_foregroundGCIsActive, "internal state inconsistency"); 3808 assert(_restart_addr == NULL, "foreground will restart from scratch"); 3809 if (PrintGCDetails) { 3810 gclog_or_tty->print_cr("bailing out to foreground collection"); 3811 } 3812 } 3813 if (UseAdaptiveSizePolicy) { 3814 size_policy()->concurrent_marking_end(); 3815 } 3816 } else { 3817 assert(SafepointSynchronize::is_at_safepoint(), 3818 "inconsistent with asynch == false"); 3819 if (UseAdaptiveSizePolicy) { 3820 size_policy()->ms_collection_marking_begin(); 3821 } 3822 // already have locks 3823 res = markFromRootsWork(asynch); 3824 _collectorState = FinalMarking; 3825 if (UseAdaptiveSizePolicy) { 3826 GenCollectedHeap* gch = GenCollectedHeap::heap(); 3827 size_policy()->ms_collection_marking_end(gch->gc_cause()); 3828 } 3829 } 3830 verify_overflow_empty(); 3831 return res; 3832 } 3833 3834 bool CMSCollector::markFromRootsWork(bool asynch) { 3835 // iterate over marked bits in bit map, doing a full scan and mark 3836 // from these roots using the following algorithm: 3837 // . if oop is to the right of the current scan pointer, 3838 // mark corresponding bit (we'll process it later) 3839 // . else (oop is to left of current scan pointer) 3840 // push oop on marking stack 3841 // . drain the marking stack 3842 3843 // Note that when we do a marking step we need to hold the 3844 // bit map lock -- recall that direct allocation (by mutators) 3845 // and promotion (by younger generation collectors) is also 3846 // marking the bit map. [the so-called allocate live policy.] 3847 // Because the implementation of bit map marking is not 3848 // robust wrt simultaneous marking of bits in the same word, 3849 // we need to make sure that there is no such interference 3850 // between concurrent such updates. 3851 3852 // already have locks 3853 assert_lock_strong(bitMapLock()); 3854 3855 verify_work_stacks_empty(); 3856 verify_overflow_empty(); 3857 bool result = false; 3858 if (CMSConcurrentMTEnabled && ConcGCThreads > 0) { 3859 result = do_marking_mt(asynch); 3860 } else { 3861 result = do_marking_st(asynch); 3862 } 3863 return result; 3864 } 3865 3866 // Forward decl 3867 class CMSConcMarkingTask; 3868 3869 class CMSConcMarkingTerminator: public ParallelTaskTerminator { 3870 CMSCollector* _collector; 3871 CMSConcMarkingTask* _task; 3872 public: 3873 virtual void yield(); 3874 3875 // "n_threads" is the number of threads to be terminated. 3876 // "queue_set" is a set of work queues of other threads. 3877 // "collector" is the CMS collector associated with this task terminator. 3878 // "yield" indicates whether we need the gang as a whole to yield. 3879 CMSConcMarkingTerminator(int n_threads, TaskQueueSetSuper* queue_set, CMSCollector* collector) : 3880 ParallelTaskTerminator(n_threads, queue_set), 3881 _collector(collector) { } 3882 3883 void set_task(CMSConcMarkingTask* task) { 3884 _task = task; 3885 } 3886 }; 3887 3888 class CMSConcMarkingTerminatorTerminator: public TerminatorTerminator { 3889 CMSConcMarkingTask* _task; 3890 public: 3891 bool should_exit_termination(); 3892 void set_task(CMSConcMarkingTask* task) { 3893 _task = task; 3894 } 3895 }; 3896 3897 // MT Concurrent Marking Task 3898 class CMSConcMarkingTask: public YieldingFlexibleGangTask { 3899 CMSCollector* _collector; 3900 int _n_workers; // requested/desired # workers 3901 bool _asynch; 3902 bool _result; 3903 CompactibleFreeListSpace* _cms_space; 3904 char _pad_front[64]; // padding to ... 3905 HeapWord* _global_finger; // ... avoid sharing cache line 3906 char _pad_back[64]; 3907 HeapWord* _restart_addr; 3908 3909 // Exposed here for yielding support 3910 Mutex* const _bit_map_lock; 3911 3912 // The per thread work queues, available here for stealing 3913 OopTaskQueueSet* _task_queues; 3914 3915 // Termination (and yielding) support 3916 CMSConcMarkingTerminator _term; 3917 CMSConcMarkingTerminatorTerminator _term_term; 3918 3919 public: 3920 CMSConcMarkingTask(CMSCollector* collector, 3921 CompactibleFreeListSpace* cms_space, 3922 bool asynch, 3923 YieldingFlexibleWorkGang* workers, 3924 OopTaskQueueSet* task_queues): 3925 YieldingFlexibleGangTask("Concurrent marking done multi-threaded"), 3926 _collector(collector), 3927 _cms_space(cms_space), 3928 _asynch(asynch), _n_workers(0), _result(true), 3929 _task_queues(task_queues), 3930 _term(_n_workers, task_queues, _collector), 3931 _bit_map_lock(collector->bitMapLock()) 3932 { 3933 _requested_size = _n_workers; 3934 _term.set_task(this); 3935 _term_term.set_task(this); 3936 _restart_addr = _global_finger = _cms_space->bottom(); 3937 } 3938 3939 3940 OopTaskQueueSet* task_queues() { return _task_queues; } 3941 3942 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 3943 3944 HeapWord** global_finger_addr() { return &_global_finger; } 3945 3946 CMSConcMarkingTerminator* terminator() { return &_term; } 3947 3948 virtual void set_for_termination(int active_workers) { 3949 terminator()->reset_for_reuse(active_workers); 3950 } 3951 3952 void work(uint worker_id); 3953 bool should_yield() { 3954 return ConcurrentMarkSweepThread::should_yield() 3955 && !_collector->foregroundGCIsActive() 3956 && _asynch; 3957 } 3958 3959 virtual void coordinator_yield(); // stuff done by coordinator 3960 bool result() { return _result; } 3961 3962 void reset(HeapWord* ra) { 3963 assert(_global_finger >= _cms_space->end(), "Postcondition of ::work(i)"); 3964 _restart_addr = _global_finger = ra; 3965 _term.reset_for_reuse(); 3966 } 3967 3968 static bool get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 3969 OopTaskQueue* work_q); 3970 3971 private: 3972 void do_scan_and_mark(int i, CompactibleFreeListSpace* sp); 3973 void do_work_steal(int i); 3974 void bump_global_finger(HeapWord* f); 3975 }; 3976 3977 bool CMSConcMarkingTerminatorTerminator::should_exit_termination() { 3978 assert(_task != NULL, "Error"); 3979 return _task->yielding(); 3980 // Note that we do not need the disjunct || _task->should_yield() above 3981 // because we want terminating threads to yield only if the task 3982 // is already in the midst of yielding, which happens only after at least one 3983 // thread has yielded. 3984 } 3985 3986 void CMSConcMarkingTerminator::yield() { 3987 if (_task->should_yield()) { 3988 _task->yield(); 3989 } else { 3990 ParallelTaskTerminator::yield(); 3991 } 3992 } 3993 3994 //////////////////////////////////////////////////////////////// 3995 // Concurrent Marking Algorithm Sketch 3996 //////////////////////////////////////////////////////////////// 3997 // Until all tasks exhausted (both spaces): 3998 // -- claim next available chunk 3999 // -- bump global finger via CAS 4000 // -- find first object that starts in this chunk 4001 // and start scanning bitmap from that position 4002 // -- scan marked objects for oops 4003 // -- CAS-mark target, and if successful: 4004 // . if target oop is above global finger (volatile read) 4005 // nothing to do 4006 // . if target oop is in chunk and above local finger 4007 // then nothing to do 4008 // . else push on work-queue 4009 // -- Deal with possible overflow issues: 4010 // . local work-queue overflow causes stuff to be pushed on 4011 // global (common) overflow queue 4012 // . always first empty local work queue 4013 // . then get a batch of oops from global work queue if any 4014 // . then do work stealing 4015 // -- When all tasks claimed (both spaces) 4016 // and local work queue empty, 4017 // then in a loop do: 4018 // . check global overflow stack; steal a batch of oops and trace 4019 // . try to steal from other threads oif GOS is empty 4020 // . if neither is available, offer termination 4021 // -- Terminate and return result 4022 // 4023 void CMSConcMarkingTask::work(uint worker_id) { 4024 elapsedTimer _timer; 4025 ResourceMark rm; 4026 HandleMark hm; 4027 4028 DEBUG_ONLY(_collector->verify_overflow_empty();) 4029 4030 // Before we begin work, our work queue should be empty 4031 assert(work_queue(worker_id)->size() == 0, "Expected to be empty"); 4032 // Scan the bitmap covering _cms_space, tracing through grey objects. 4033 _timer.start(); 4034 do_scan_and_mark(worker_id, _cms_space); 4035 _timer.stop(); 4036 if (PrintCMSStatistics != 0) { 4037 gclog_or_tty->print_cr("Finished cms space scanning in %dth thread: %3.3f sec", 4038 worker_id, _timer.seconds()); 4039 // XXX: need xxx/xxx type of notation, two timers 4040 } 4041 4042 // ... do work stealing 4043 _timer.reset(); 4044 _timer.start(); 4045 do_work_steal(worker_id); 4046 _timer.stop(); 4047 if (PrintCMSStatistics != 0) { 4048 gclog_or_tty->print_cr("Finished work stealing in %dth thread: %3.3f sec", 4049 worker_id, _timer.seconds()); 4050 // XXX: need xxx/xxx type of notation, two timers 4051 } 4052 assert(_collector->_markStack.isEmpty(), "Should have been emptied"); 4053 assert(work_queue(worker_id)->size() == 0, "Should have been emptied"); 4054 // Note that under the current task protocol, the 4055 // following assertion is true even of the spaces 4056 // expanded since the completion of the concurrent 4057 // marking. XXX This will likely change under a strict 4058 // ABORT semantics. 4059 // After perm removal the comparison was changed to 4060 // greater than or equal to from strictly greater than. 4061 // Before perm removal the highest address sweep would 4062 // have been at the end of perm gen but now is at the 4063 // end of the tenured gen. 4064 assert(_global_finger >= _cms_space->end(), 4065 "All tasks have been completed"); 4066 DEBUG_ONLY(_collector->verify_overflow_empty();) 4067 } 4068 4069 void CMSConcMarkingTask::bump_global_finger(HeapWord* f) { 4070 HeapWord* read = _global_finger; 4071 HeapWord* cur = read; 4072 while (f > read) { 4073 cur = read; 4074 read = (HeapWord*) Atomic::cmpxchg_ptr(f, &_global_finger, cur); 4075 if (cur == read) { 4076 // our cas succeeded 4077 assert(_global_finger >= f, "protocol consistency"); 4078 break; 4079 } 4080 } 4081 } 4082 4083 // This is really inefficient, and should be redone by 4084 // using (not yet available) block-read and -write interfaces to the 4085 // stack and the work_queue. XXX FIX ME !!! 4086 bool CMSConcMarkingTask::get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 4087 OopTaskQueue* work_q) { 4088 // Fast lock-free check 4089 if (ovflw_stk->length() == 0) { 4090 return false; 4091 } 4092 assert(work_q->size() == 0, "Shouldn't steal"); 4093 MutexLockerEx ml(ovflw_stk->par_lock(), 4094 Mutex::_no_safepoint_check_flag); 4095 // Grab up to 1/4 the size of the work queue 4096 size_t num = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 4097 (size_t)ParGCDesiredObjsFromOverflowList); 4098 num = MIN2(num, ovflw_stk->length()); 4099 for (int i = (int) num; i > 0; i--) { 4100 oop cur = ovflw_stk->pop(); 4101 assert(cur != NULL, "Counted wrong?"); 4102 work_q->push(cur); 4103 } 4104 return num > 0; 4105 } 4106 4107 void CMSConcMarkingTask::do_scan_and_mark(int i, CompactibleFreeListSpace* sp) { 4108 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 4109 int n_tasks = pst->n_tasks(); 4110 // We allow that there may be no tasks to do here because 4111 // we are restarting after a stack overflow. 4112 assert(pst->valid() || n_tasks == 0, "Uninitialized use?"); 4113 uint nth_task = 0; 4114 4115 HeapWord* aligned_start = sp->bottom(); 4116 if (sp->used_region().contains(_restart_addr)) { 4117 // Align down to a card boundary for the start of 0th task 4118 // for this space. 4119 aligned_start = 4120 (HeapWord*)align_size_down((uintptr_t)_restart_addr, 4121 CardTableModRefBS::card_size); 4122 } 4123 4124 size_t chunk_size = sp->marking_task_size(); 4125 while (!pst->is_task_claimed(/* reference */ nth_task)) { 4126 // Having claimed the nth task in this space, 4127 // compute the chunk that it corresponds to: 4128 MemRegion span = MemRegion(aligned_start + nth_task*chunk_size, 4129 aligned_start + (nth_task+1)*chunk_size); 4130 // Try and bump the global finger via a CAS; 4131 // note that we need to do the global finger bump 4132 // _before_ taking the intersection below, because 4133 // the task corresponding to that region will be 4134 // deemed done even if the used_region() expands 4135 // because of allocation -- as it almost certainly will 4136 // during start-up while the threads yield in the 4137 // closure below. 4138 HeapWord* finger = span.end(); 4139 bump_global_finger(finger); // atomically 4140 // There are null tasks here corresponding to chunks 4141 // beyond the "top" address of the space. 4142 span = span.intersection(sp->used_region()); 4143 if (!span.is_empty()) { // Non-null task 4144 HeapWord* prev_obj; 4145 assert(!span.contains(_restart_addr) || nth_task == 0, 4146 "Inconsistency"); 4147 if (nth_task == 0) { 4148 // For the 0th task, we'll not need to compute a block_start. 4149 if (span.contains(_restart_addr)) { 4150 // In the case of a restart because of stack overflow, 4151 // we might additionally skip a chunk prefix. 4152 prev_obj = _restart_addr; 4153 } else { 4154 prev_obj = span.start(); 4155 } 4156 } else { 4157 // We want to skip the first object because 4158 // the protocol is to scan any object in its entirety 4159 // that _starts_ in this span; a fortiori, any 4160 // object starting in an earlier span is scanned 4161 // as part of an earlier claimed task. 4162 // Below we use the "careful" version of block_start 4163 // so we do not try to navigate uninitialized objects. 4164 prev_obj = sp->block_start_careful(span.start()); 4165 // Below we use a variant of block_size that uses the 4166 // Printezis bits to avoid waiting for allocated 4167 // objects to become initialized/parsable. 4168 while (prev_obj < span.start()) { 4169 size_t sz = sp->block_size_no_stall(prev_obj, _collector); 4170 if (sz > 0) { 4171 prev_obj += sz; 4172 } else { 4173 // In this case we may end up doing a bit of redundant 4174 // scanning, but that appears unavoidable, short of 4175 // locking the free list locks; see bug 6324141. 4176 break; 4177 } 4178 } 4179 } 4180 if (prev_obj < span.end()) { 4181 MemRegion my_span = MemRegion(prev_obj, span.end()); 4182 // Do the marking work within a non-empty span -- 4183 // the last argument to the constructor indicates whether the 4184 // iteration should be incremental with periodic yields. 4185 Par_MarkFromRootsClosure cl(this, _collector, my_span, 4186 &_collector->_markBitMap, 4187 work_queue(i), 4188 &_collector->_markStack, 4189 _asynch); 4190 _collector->_markBitMap.iterate(&cl, my_span.start(), my_span.end()); 4191 } // else nothing to do for this task 4192 } // else nothing to do for this task 4193 } 4194 // We'd be tempted to assert here that since there are no 4195 // more tasks left to claim in this space, the global_finger 4196 // must exceed space->top() and a fortiori space->end(). However, 4197 // that would not quite be correct because the bumping of 4198 // global_finger occurs strictly after the claiming of a task, 4199 // so by the time we reach here the global finger may not yet 4200 // have been bumped up by the thread that claimed the last 4201 // task. 4202 pst->all_tasks_completed(); 4203 } 4204 4205 class Par_ConcMarkingClosure: public CMSOopClosure { 4206 private: 4207 CMSCollector* _collector; 4208 CMSConcMarkingTask* _task; 4209 MemRegion _span; 4210 CMSBitMap* _bit_map; 4211 CMSMarkStack* _overflow_stack; 4212 OopTaskQueue* _work_queue; 4213 protected: 4214 DO_OOP_WORK_DEFN 4215 public: 4216 Par_ConcMarkingClosure(CMSCollector* collector, CMSConcMarkingTask* task, OopTaskQueue* work_queue, 4217 CMSBitMap* bit_map, CMSMarkStack* overflow_stack): 4218 CMSOopClosure(collector->ref_processor()), 4219 _collector(collector), 4220 _task(task), 4221 _span(collector->_span), 4222 _work_queue(work_queue), 4223 _bit_map(bit_map), 4224 _overflow_stack(overflow_stack) 4225 { } 4226 virtual void do_oop(oop* p); 4227 virtual void do_oop(narrowOop* p); 4228 4229 void trim_queue(size_t max); 4230 void handle_stack_overflow(HeapWord* lost); 4231 void do_yield_check() { 4232 if (_task->should_yield()) { 4233 _task->yield(); 4234 } 4235 } 4236 }; 4237 4238 // Grey object scanning during work stealing phase -- 4239 // the salient assumption here is that any references 4240 // that are in these stolen objects being scanned must 4241 // already have been initialized (else they would not have 4242 // been published), so we do not need to check for 4243 // uninitialized objects before pushing here. 4244 void Par_ConcMarkingClosure::do_oop(oop obj) { 4245 assert(obj->is_oop_or_null(true), "expected an oop or NULL"); 4246 HeapWord* addr = (HeapWord*)obj; 4247 // Check if oop points into the CMS generation 4248 // and is not marked 4249 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 4250 // a white object ... 4251 // If we manage to "claim" the object, by being the 4252 // first thread to mark it, then we push it on our 4253 // marking stack 4254 if (_bit_map->par_mark(addr)) { // ... now grey 4255 // push on work queue (grey set) 4256 bool simulate_overflow = false; 4257 NOT_PRODUCT( 4258 if (CMSMarkStackOverflowALot && 4259 _collector->simulate_overflow()) { 4260 // simulate a stack overflow 4261 simulate_overflow = true; 4262 } 4263 ) 4264 if (simulate_overflow || 4265 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) { 4266 // stack overflow 4267 if (PrintCMSStatistics != 0) { 4268 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 4269 SIZE_FORMAT, _overflow_stack->capacity()); 4270 } 4271 // We cannot assert that the overflow stack is full because 4272 // it may have been emptied since. 4273 assert(simulate_overflow || 4274 _work_queue->size() == _work_queue->max_elems(), 4275 "Else push should have succeeded"); 4276 handle_stack_overflow(addr); 4277 } 4278 } // Else, some other thread got there first 4279 do_yield_check(); 4280 } 4281 } 4282 4283 void Par_ConcMarkingClosure::do_oop(oop* p) { Par_ConcMarkingClosure::do_oop_work(p); } 4284 void Par_ConcMarkingClosure::do_oop(narrowOop* p) { Par_ConcMarkingClosure::do_oop_work(p); } 4285 4286 void Par_ConcMarkingClosure::trim_queue(size_t max) { 4287 while (_work_queue->size() > max) { 4288 oop new_oop; 4289 if (_work_queue->pop_local(new_oop)) { 4290 assert(new_oop->is_oop(), "Should be an oop"); 4291 assert(_bit_map->isMarked((HeapWord*)new_oop), "Grey object"); 4292 assert(_span.contains((HeapWord*)new_oop), "Not in span"); 4293 new_oop->oop_iterate(this); // do_oop() above 4294 do_yield_check(); 4295 } 4296 } 4297 } 4298 4299 // Upon stack overflow, we discard (part of) the stack, 4300 // remembering the least address amongst those discarded 4301 // in CMSCollector's _restart_address. 4302 void Par_ConcMarkingClosure::handle_stack_overflow(HeapWord* lost) { 4303 // We need to do this under a mutex to prevent other 4304 // workers from interfering with the work done below. 4305 MutexLockerEx ml(_overflow_stack->par_lock(), 4306 Mutex::_no_safepoint_check_flag); 4307 // Remember the least grey address discarded 4308 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 4309 _collector->lower_restart_addr(ra); 4310 _overflow_stack->reset(); // discard stack contents 4311 _overflow_stack->expand(); // expand the stack if possible 4312 } 4313 4314 4315 void CMSConcMarkingTask::do_work_steal(int i) { 4316 OopTaskQueue* work_q = work_queue(i); 4317 oop obj_to_scan; 4318 CMSBitMap* bm = &(_collector->_markBitMap); 4319 CMSMarkStack* ovflw = &(_collector->_markStack); 4320 int* seed = _collector->hash_seed(i); 4321 Par_ConcMarkingClosure cl(_collector, this, work_q, bm, ovflw); 4322 while (true) { 4323 cl.trim_queue(0); 4324 assert(work_q->size() == 0, "Should have been emptied above"); 4325 if (get_work_from_overflow_stack(ovflw, work_q)) { 4326 // Can't assert below because the work obtained from the 4327 // overflow stack may already have been stolen from us. 4328 // assert(work_q->size() > 0, "Work from overflow stack"); 4329 continue; 4330 } else if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 4331 assert(obj_to_scan->is_oop(), "Should be an oop"); 4332 assert(bm->isMarked((HeapWord*)obj_to_scan), "Grey object"); 4333 obj_to_scan->oop_iterate(&cl); 4334 } else if (terminator()->offer_termination(&_term_term)) { 4335 assert(work_q->size() == 0, "Impossible!"); 4336 break; 4337 } else if (yielding() || should_yield()) { 4338 yield(); 4339 } 4340 } 4341 } 4342 4343 // This is run by the CMS (coordinator) thread. 4344 void CMSConcMarkingTask::coordinator_yield() { 4345 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 4346 "CMS thread should hold CMS token"); 4347 // First give up the locks, then yield, then re-lock 4348 // We should probably use a constructor/destructor idiom to 4349 // do this unlock/lock or modify the MutexUnlocker class to 4350 // serve our purpose. XXX 4351 assert_lock_strong(_bit_map_lock); 4352 _bit_map_lock->unlock(); 4353 ConcurrentMarkSweepThread::desynchronize(true); 4354 ConcurrentMarkSweepThread::acknowledge_yield_request(); 4355 _collector->stopTimer(); 4356 if (PrintCMSStatistics != 0) { 4357 _collector->incrementYields(); 4358 } 4359 _collector->icms_wait(); 4360 4361 // It is possible for whichever thread initiated the yield request 4362 // not to get a chance to wake up and take the bitmap lock between 4363 // this thread releasing it and reacquiring it. So, while the 4364 // should_yield() flag is on, let's sleep for a bit to give the 4365 // other thread a chance to wake up. The limit imposed on the number 4366 // of iterations is defensive, to avoid any unforseen circumstances 4367 // putting us into an infinite loop. Since it's always been this 4368 // (coordinator_yield()) method that was observed to cause the 4369 // problem, we are using a parameter (CMSCoordinatorYieldSleepCount) 4370 // which is by default non-zero. For the other seven methods that 4371 // also perform the yield operation, as are using a different 4372 // parameter (CMSYieldSleepCount) which is by default zero. This way we 4373 // can enable the sleeping for those methods too, if necessary. 4374 // See 6442774. 4375 // 4376 // We really need to reconsider the synchronization between the GC 4377 // thread and the yield-requesting threads in the future and we 4378 // should really use wait/notify, which is the recommended 4379 // way of doing this type of interaction. Additionally, we should 4380 // consolidate the eight methods that do the yield operation and they 4381 // are almost identical into one for better maintainability and 4382 // readability. See 6445193. 4383 // 4384 // Tony 2006.06.29 4385 for (unsigned i = 0; i < CMSCoordinatorYieldSleepCount && 4386 ConcurrentMarkSweepThread::should_yield() && 4387 !CMSCollector::foregroundGCIsActive(); ++i) { 4388 os::sleep(Thread::current(), 1, false); 4389 ConcurrentMarkSweepThread::acknowledge_yield_request(); 4390 } 4391 4392 ConcurrentMarkSweepThread::synchronize(true); 4393 _bit_map_lock->lock_without_safepoint_check(); 4394 _collector->startTimer(); 4395 } 4396 4397 bool CMSCollector::do_marking_mt(bool asynch) { 4398 assert(ConcGCThreads > 0 && conc_workers() != NULL, "precondition"); 4399 int num_workers = AdaptiveSizePolicy::calc_active_conc_workers( 4400 conc_workers()->total_workers(), 4401 conc_workers()->active_workers(), 4402 Threads::number_of_non_daemon_threads()); 4403 conc_workers()->set_active_workers(num_workers); 4404 4405 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 4406 4407 CMSConcMarkingTask tsk(this, 4408 cms_space, 4409 asynch, 4410 conc_workers(), 4411 task_queues()); 4412 4413 // Since the actual number of workers we get may be different 4414 // from the number we requested above, do we need to do anything different 4415 // below? In particular, may be we need to subclass the SequantialSubTasksDone 4416 // class?? XXX 4417 cms_space ->initialize_sequential_subtasks_for_marking(num_workers); 4418 4419 // Refs discovery is already non-atomic. 4420 assert(!ref_processor()->discovery_is_atomic(), "Should be non-atomic"); 4421 assert(ref_processor()->discovery_is_mt(), "Discovery should be MT"); 4422 conc_workers()->start_task(&tsk); 4423 while (tsk.yielded()) { 4424 tsk.coordinator_yield(); 4425 conc_workers()->continue_task(&tsk); 4426 } 4427 // If the task was aborted, _restart_addr will be non-NULL 4428 assert(tsk.completed() || _restart_addr != NULL, "Inconsistency"); 4429 while (_restart_addr != NULL) { 4430 // XXX For now we do not make use of ABORTED state and have not 4431 // yet implemented the right abort semantics (even in the original 4432 // single-threaded CMS case). That needs some more investigation 4433 // and is deferred for now; see CR# TBF. 07252005YSR. XXX 4434 assert(!CMSAbortSemantics || tsk.aborted(), "Inconsistency"); 4435 // If _restart_addr is non-NULL, a marking stack overflow 4436 // occurred; we need to do a fresh marking iteration from the 4437 // indicated restart address. 4438 if (_foregroundGCIsActive && asynch) { 4439 // We may be running into repeated stack overflows, having 4440 // reached the limit of the stack size, while making very 4441 // slow forward progress. It may be best to bail out and 4442 // let the foreground collector do its job. 4443 // Clear _restart_addr, so that foreground GC 4444 // works from scratch. This avoids the headache of 4445 // a "rescan" which would otherwise be needed because 4446 // of the dirty mod union table & card table. 4447 _restart_addr = NULL; 4448 return false; 4449 } 4450 // Adjust the task to restart from _restart_addr 4451 tsk.reset(_restart_addr); 4452 cms_space ->initialize_sequential_subtasks_for_marking(num_workers, 4453 _restart_addr); 4454 _restart_addr = NULL; 4455 // Get the workers going again 4456 conc_workers()->start_task(&tsk); 4457 while (tsk.yielded()) { 4458 tsk.coordinator_yield(); 4459 conc_workers()->continue_task(&tsk); 4460 } 4461 } 4462 assert(tsk.completed(), "Inconsistency"); 4463 assert(tsk.result() == true, "Inconsistency"); 4464 return true; 4465 } 4466 4467 bool CMSCollector::do_marking_st(bool asynch) { 4468 ResourceMark rm; 4469 HandleMark hm; 4470 4471 // Temporarily make refs discovery single threaded (non-MT) 4472 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false); 4473 MarkFromRootsClosure markFromRootsClosure(this, _span, &_markBitMap, 4474 &_markStack, CMSYield && asynch); 4475 // the last argument to iterate indicates whether the iteration 4476 // should be incremental with periodic yields. 4477 _markBitMap.iterate(&markFromRootsClosure); 4478 // If _restart_addr is non-NULL, a marking stack overflow 4479 // occurred; we need to do a fresh iteration from the 4480 // indicated restart address. 4481 while (_restart_addr != NULL) { 4482 if (_foregroundGCIsActive && asynch) { 4483 // We may be running into repeated stack overflows, having 4484 // reached the limit of the stack size, while making very 4485 // slow forward progress. It may be best to bail out and 4486 // let the foreground collector do its job. 4487 // Clear _restart_addr, so that foreground GC 4488 // works from scratch. This avoids the headache of 4489 // a "rescan" which would otherwise be needed because 4490 // of the dirty mod union table & card table. 4491 _restart_addr = NULL; 4492 return false; // indicating failure to complete marking 4493 } 4494 // Deal with stack overflow: 4495 // we restart marking from _restart_addr 4496 HeapWord* ra = _restart_addr; 4497 markFromRootsClosure.reset(ra); 4498 _restart_addr = NULL; 4499 _markBitMap.iterate(&markFromRootsClosure, ra, _span.end()); 4500 } 4501 return true; 4502 } 4503 4504 void CMSCollector::preclean() { 4505 check_correct_thread_executing(); 4506 assert(Thread::current()->is_ConcurrentGC_thread(), "Wrong thread"); 4507 verify_work_stacks_empty(); 4508 verify_overflow_empty(); 4509 _abort_preclean = false; 4510 if (CMSPrecleaningEnabled) { 4511 if (!CMSEdenChunksRecordAlways) { 4512 _eden_chunk_index = 0; 4513 } 4514 size_t used = get_eden_used(); 4515 size_t capacity = get_eden_capacity(); 4516 // Don't start sampling unless we will get sufficiently 4517 // many samples. 4518 if (used < (capacity/(CMSScheduleRemarkSamplingRatio * 100) 4519 * CMSScheduleRemarkEdenPenetration)) { 4520 _start_sampling = true; 4521 } else { 4522 _start_sampling = false; 4523 } 4524 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 4525 CMSPhaseAccounting pa(this, "preclean", !PrintGCDetails); 4526 preclean_work(CMSPrecleanRefLists1, CMSPrecleanSurvivors1); 4527 } 4528 CMSTokenSync x(true); // is cms thread 4529 if (CMSPrecleaningEnabled) { 4530 sample_eden(); 4531 _collectorState = AbortablePreclean; 4532 } else { 4533 _collectorState = FinalMarking; 4534 } 4535 verify_work_stacks_empty(); 4536 verify_overflow_empty(); 4537 } 4538 4539 // Try and schedule the remark such that young gen 4540 // occupancy is CMSScheduleRemarkEdenPenetration %. 4541 void CMSCollector::abortable_preclean() { 4542 check_correct_thread_executing(); 4543 assert(CMSPrecleaningEnabled, "Inconsistent control state"); 4544 assert(_collectorState == AbortablePreclean, "Inconsistent control state"); 4545 4546 // If Eden's current occupancy is below this threshold, 4547 // immediately schedule the remark; else preclean 4548 // past the next scavenge in an effort to 4549 // schedule the pause as described above. By choosing 4550 // CMSScheduleRemarkEdenSizeThreshold >= max eden size 4551 // we will never do an actual abortable preclean cycle. 4552 if (get_eden_used() > CMSScheduleRemarkEdenSizeThreshold) { 4553 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 4554 CMSPhaseAccounting pa(this, "abortable-preclean", !PrintGCDetails); 4555 // We need more smarts in the abortable preclean 4556 // loop below to deal with cases where allocation 4557 // in young gen is very very slow, and our precleaning 4558 // is running a losing race against a horde of 4559 // mutators intent on flooding us with CMS updates 4560 // (dirty cards). 4561 // One, admittedly dumb, strategy is to give up 4562 // after a certain number of abortable precleaning loops 4563 // or after a certain maximum time. We want to make 4564 // this smarter in the next iteration. 4565 // XXX FIX ME!!! YSR 4566 size_t loops = 0, workdone = 0, cumworkdone = 0, waited = 0; 4567 while (!(should_abort_preclean() || 4568 ConcurrentMarkSweepThread::should_terminate())) { 4569 workdone = preclean_work(CMSPrecleanRefLists2, CMSPrecleanSurvivors2); 4570 cumworkdone += workdone; 4571 loops++; 4572 // Voluntarily terminate abortable preclean phase if we have 4573 // been at it for too long. 4574 if ((CMSMaxAbortablePrecleanLoops != 0) && 4575 loops >= CMSMaxAbortablePrecleanLoops) { 4576 if (PrintGCDetails) { 4577 gclog_or_tty->print(" CMS: abort preclean due to loops "); 4578 } 4579 break; 4580 } 4581 if (pa.wallclock_millis() > CMSMaxAbortablePrecleanTime) { 4582 if (PrintGCDetails) { 4583 gclog_or_tty->print(" CMS: abort preclean due to time "); 4584 } 4585 break; 4586 } 4587 // If we are doing little work each iteration, we should 4588 // take a short break. 4589 if (workdone < CMSAbortablePrecleanMinWorkPerIteration) { 4590 // Sleep for some time, waiting for work to accumulate 4591 stopTimer(); 4592 cmsThread()->wait_on_cms_lock(CMSAbortablePrecleanWaitMillis); 4593 startTimer(); 4594 waited++; 4595 } 4596 } 4597 if (PrintCMSStatistics > 0) { 4598 gclog_or_tty->print(" [" SIZE_FORMAT " iterations, " SIZE_FORMAT " waits, " SIZE_FORMAT " cards)] ", 4599 loops, waited, cumworkdone); 4600 } 4601 } 4602 CMSTokenSync x(true); // is cms thread 4603 if (_collectorState != Idling) { 4604 assert(_collectorState == AbortablePreclean, 4605 "Spontaneous state transition?"); 4606 _collectorState = FinalMarking; 4607 } // Else, a foreground collection completed this CMS cycle. 4608 return; 4609 } 4610 4611 // Respond to an Eden sampling opportunity 4612 void CMSCollector::sample_eden() { 4613 // Make sure a young gc cannot sneak in between our 4614 // reading and recording of a sample. 4615 assert(Thread::current()->is_ConcurrentGC_thread(), 4616 "Only the cms thread may collect Eden samples"); 4617 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 4618 "Should collect samples while holding CMS token"); 4619 if (!_start_sampling) { 4620 return; 4621 } 4622 // When CMSEdenChunksRecordAlways is true, the eden chunk array 4623 // is populated by the young generation. 4624 if (_eden_chunk_array != NULL && !CMSEdenChunksRecordAlways) { 4625 if (_eden_chunk_index < _eden_chunk_capacity) { 4626 _eden_chunk_array[_eden_chunk_index] = *_top_addr; // take sample 4627 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr, 4628 "Unexpected state of Eden"); 4629 // We'd like to check that what we just sampled is an oop-start address; 4630 // however, we cannot do that here since the object may not yet have been 4631 // initialized. So we'll instead do the check when we _use_ this sample 4632 // later. 4633 if (_eden_chunk_index == 0 || 4634 (pointer_delta(_eden_chunk_array[_eden_chunk_index], 4635 _eden_chunk_array[_eden_chunk_index-1]) 4636 >= CMSSamplingGrain)) { 4637 _eden_chunk_index++; // commit sample 4638 } 4639 } 4640 } 4641 if ((_collectorState == AbortablePreclean) && !_abort_preclean) { 4642 size_t used = get_eden_used(); 4643 size_t capacity = get_eden_capacity(); 4644 assert(used <= capacity, "Unexpected state of Eden"); 4645 if (used > (capacity/100 * CMSScheduleRemarkEdenPenetration)) { 4646 _abort_preclean = true; 4647 } 4648 } 4649 } 4650 4651 4652 size_t CMSCollector::preclean_work(bool clean_refs, bool clean_survivor) { 4653 assert(_collectorState == Precleaning || 4654 _collectorState == AbortablePreclean, "incorrect state"); 4655 ResourceMark rm; 4656 HandleMark hm; 4657 4658 // Precleaning is currently not MT but the reference processor 4659 // may be set for MT. Disable it temporarily here. 4660 ReferenceProcessor* rp = ref_processor(); 4661 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(rp, false); 4662 4663 // Do one pass of scrubbing the discovered reference lists 4664 // to remove any reference objects with strongly-reachable 4665 // referents. 4666 if (clean_refs) { 4667 CMSPrecleanRefsYieldClosure yield_cl(this); 4668 assert(rp->span().equals(_span), "Spans should be equal"); 4669 CMSKeepAliveClosure keep_alive(this, _span, &_markBitMap, 4670 &_markStack, true /* preclean */); 4671 CMSDrainMarkingStackClosure complete_trace(this, 4672 _span, &_markBitMap, &_markStack, 4673 &keep_alive, true /* preclean */); 4674 4675 // We don't want this step to interfere with a young 4676 // collection because we don't want to take CPU 4677 // or memory bandwidth away from the young GC threads 4678 // (which may be as many as there are CPUs). 4679 // Note that we don't need to protect ourselves from 4680 // interference with mutators because they can't 4681 // manipulate the discovered reference lists nor affect 4682 // the computed reachability of the referents, the 4683 // only properties manipulated by the precleaning 4684 // of these reference lists. 4685 stopTimer(); 4686 CMSTokenSyncWithLocks x(true /* is cms thread */, 4687 bitMapLock()); 4688 startTimer(); 4689 sample_eden(); 4690 4691 // The following will yield to allow foreground 4692 // collection to proceed promptly. XXX YSR: 4693 // The code in this method may need further 4694 // tweaking for better performance and some restructuring 4695 // for cleaner interfaces. 4696 GCTimer *gc_timer = NULL; // Currently not tracing concurrent phases 4697 rp->preclean_discovered_references( 4698 rp->is_alive_non_header(), &keep_alive, &complete_trace, &yield_cl, 4699 gc_timer); 4700 } 4701 4702 if (clean_survivor) { // preclean the active survivor space(s) 4703 assert(_young_gen->kind() == Generation::DefNew || 4704 _young_gen->kind() == Generation::ParNew || 4705 _young_gen->kind() == Generation::ASParNew, 4706 "incorrect type for cast"); 4707 DefNewGeneration* dng = (DefNewGeneration*)_young_gen; 4708 PushAndMarkClosure pam_cl(this, _span, ref_processor(), 4709 &_markBitMap, &_modUnionTable, 4710 &_markStack, true /* precleaning phase */); 4711 stopTimer(); 4712 CMSTokenSyncWithLocks ts(true /* is cms thread */, 4713 bitMapLock()); 4714 startTimer(); 4715 unsigned int before_count = 4716 GenCollectedHeap::heap()->total_collections(); 4717 SurvivorSpacePrecleanClosure 4718 sss_cl(this, _span, &_markBitMap, &_markStack, 4719 &pam_cl, before_count, CMSYield); 4720 dng->from()->object_iterate_careful(&sss_cl); 4721 dng->to()->object_iterate_careful(&sss_cl); 4722 } 4723 MarkRefsIntoAndScanClosure 4724 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable, 4725 &_markStack, this, CMSYield, 4726 true /* precleaning phase */); 4727 // CAUTION: The following closure has persistent state that may need to 4728 // be reset upon a decrease in the sequence of addresses it 4729 // processes. 4730 ScanMarkedObjectsAgainCarefullyClosure 4731 smoac_cl(this, _span, 4732 &_markBitMap, &_markStack, &mrias_cl, CMSYield); 4733 4734 // Preclean dirty cards in ModUnionTable and CardTable using 4735 // appropriate convergence criterion; 4736 // repeat CMSPrecleanIter times unless we find that 4737 // we are losing. 4738 assert(CMSPrecleanIter < 10, "CMSPrecleanIter is too large"); 4739 assert(CMSPrecleanNumerator < CMSPrecleanDenominator, 4740 "Bad convergence multiplier"); 4741 assert(CMSPrecleanThreshold >= 100, 4742 "Unreasonably low CMSPrecleanThreshold"); 4743 4744 size_t numIter, cumNumCards, lastNumCards, curNumCards; 4745 for (numIter = 0, cumNumCards = lastNumCards = curNumCards = 0; 4746 numIter < CMSPrecleanIter; 4747 numIter++, lastNumCards = curNumCards, cumNumCards += curNumCards) { 4748 curNumCards = preclean_mod_union_table(_cmsGen, &smoac_cl); 4749 if (Verbose && PrintGCDetails) { 4750 gclog_or_tty->print(" (modUnionTable: " SIZE_FORMAT " cards)", curNumCards); 4751 } 4752 // Either there are very few dirty cards, so re-mark 4753 // pause will be small anyway, or our pre-cleaning isn't 4754 // that much faster than the rate at which cards are being 4755 // dirtied, so we might as well stop and re-mark since 4756 // precleaning won't improve our re-mark time by much. 4757 if (curNumCards <= CMSPrecleanThreshold || 4758 (numIter > 0 && 4759 (curNumCards * CMSPrecleanDenominator > 4760 lastNumCards * CMSPrecleanNumerator))) { 4761 numIter++; 4762 cumNumCards += curNumCards; 4763 break; 4764 } 4765 } 4766 4767 preclean_klasses(&mrias_cl, _cmsGen->freelistLock()); 4768 4769 curNumCards = preclean_card_table(_cmsGen, &smoac_cl); 4770 cumNumCards += curNumCards; 4771 if (PrintGCDetails && PrintCMSStatistics != 0) { 4772 gclog_or_tty->print_cr(" (cardTable: " SIZE_FORMAT " cards, re-scanned " SIZE_FORMAT " cards, " SIZE_FORMAT " iterations)", 4773 curNumCards, cumNumCards, numIter); 4774 } 4775 return cumNumCards; // as a measure of useful work done 4776 } 4777 4778 // PRECLEANING NOTES: 4779 // Precleaning involves: 4780 // . reading the bits of the modUnionTable and clearing the set bits. 4781 // . For the cards corresponding to the set bits, we scan the 4782 // objects on those cards. This means we need the free_list_lock 4783 // so that we can safely iterate over the CMS space when scanning 4784 // for oops. 4785 // . When we scan the objects, we'll be both reading and setting 4786 // marks in the marking bit map, so we'll need the marking bit map. 4787 // . For protecting _collector_state transitions, we take the CGC_lock. 4788 // Note that any races in the reading of of card table entries by the 4789 // CMS thread on the one hand and the clearing of those entries by the 4790 // VM thread or the setting of those entries by the mutator threads on the 4791 // other are quite benign. However, for efficiency it makes sense to keep 4792 // the VM thread from racing with the CMS thread while the latter is 4793 // dirty card info to the modUnionTable. We therefore also use the 4794 // CGC_lock to protect the reading of the card table and the mod union 4795 // table by the CM thread. 4796 // . We run concurrently with mutator updates, so scanning 4797 // needs to be done carefully -- we should not try to scan 4798 // potentially uninitialized objects. 4799 // 4800 // Locking strategy: While holding the CGC_lock, we scan over and 4801 // reset a maximal dirty range of the mod union / card tables, then lock 4802 // the free_list_lock and bitmap lock to do a full marking, then 4803 // release these locks; and repeat the cycle. This allows for a 4804 // certain amount of fairness in the sharing of these locks between 4805 // the CMS collector on the one hand, and the VM thread and the 4806 // mutators on the other. 4807 4808 // NOTE: preclean_mod_union_table() and preclean_card_table() 4809 // further below are largely identical; if you need to modify 4810 // one of these methods, please check the other method too. 4811 4812 size_t CMSCollector::preclean_mod_union_table( 4813 ConcurrentMarkSweepGeneration* gen, 4814 ScanMarkedObjectsAgainCarefullyClosure* cl) { 4815 verify_work_stacks_empty(); 4816 verify_overflow_empty(); 4817 4818 // strategy: starting with the first card, accumulate contiguous 4819 // ranges of dirty cards; clear these cards, then scan the region 4820 // covered by these cards. 4821 4822 // Since all of the MUT is committed ahead, we can just use 4823 // that, in case the generations expand while we are precleaning. 4824 // It might also be fine to just use the committed part of the 4825 // generation, but we might potentially miss cards when the 4826 // generation is rapidly expanding while we are in the midst 4827 // of precleaning. 4828 HeapWord* startAddr = gen->reserved().start(); 4829 HeapWord* endAddr = gen->reserved().end(); 4830 4831 cl->setFreelistLock(gen->freelistLock()); // needed for yielding 4832 4833 size_t numDirtyCards, cumNumDirtyCards; 4834 HeapWord *nextAddr, *lastAddr; 4835 for (cumNumDirtyCards = numDirtyCards = 0, 4836 nextAddr = lastAddr = startAddr; 4837 nextAddr < endAddr; 4838 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 4839 4840 ResourceMark rm; 4841 HandleMark hm; 4842 4843 MemRegion dirtyRegion; 4844 { 4845 stopTimer(); 4846 // Potential yield point 4847 CMSTokenSync ts(true); 4848 startTimer(); 4849 sample_eden(); 4850 // Get dirty region starting at nextOffset (inclusive), 4851 // simultaneously clearing it. 4852 dirtyRegion = 4853 _modUnionTable.getAndClearMarkedRegion(nextAddr, endAddr); 4854 assert(dirtyRegion.start() >= nextAddr, 4855 "returned region inconsistent?"); 4856 } 4857 // Remember where the next search should begin. 4858 // The returned region (if non-empty) is a right open interval, 4859 // so lastOffset is obtained from the right end of that 4860 // interval. 4861 lastAddr = dirtyRegion.end(); 4862 // Should do something more transparent and less hacky XXX 4863 numDirtyCards = 4864 _modUnionTable.heapWordDiffToOffsetDiff(dirtyRegion.word_size()); 4865 4866 // We'll scan the cards in the dirty region (with periodic 4867 // yields for foreground GC as needed). 4868 if (!dirtyRegion.is_empty()) { 4869 assert(numDirtyCards > 0, "consistency check"); 4870 HeapWord* stop_point = NULL; 4871 stopTimer(); 4872 // Potential yield point 4873 CMSTokenSyncWithLocks ts(true, gen->freelistLock(), 4874 bitMapLock()); 4875 startTimer(); 4876 { 4877 verify_work_stacks_empty(); 4878 verify_overflow_empty(); 4879 sample_eden(); 4880 stop_point = 4881 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 4882 } 4883 if (stop_point != NULL) { 4884 // The careful iteration stopped early either because it found an 4885 // uninitialized object, or because we were in the midst of an 4886 // "abortable preclean", which should now be aborted. Redirty 4887 // the bits corresponding to the partially-scanned or unscanned 4888 // cards. We'll either restart at the next block boundary or 4889 // abort the preclean. 4890 assert((_collectorState == AbortablePreclean && should_abort_preclean()), 4891 "Should only be AbortablePreclean."); 4892 _modUnionTable.mark_range(MemRegion(stop_point, dirtyRegion.end())); 4893 if (should_abort_preclean()) { 4894 break; // out of preclean loop 4895 } else { 4896 // Compute the next address at which preclean should pick up; 4897 // might need bitMapLock in order to read P-bits. 4898 lastAddr = next_card_start_after_block(stop_point); 4899 } 4900 } 4901 } else { 4902 assert(lastAddr == endAddr, "consistency check"); 4903 assert(numDirtyCards == 0, "consistency check"); 4904 break; 4905 } 4906 } 4907 verify_work_stacks_empty(); 4908 verify_overflow_empty(); 4909 return cumNumDirtyCards; 4910 } 4911 4912 // NOTE: preclean_mod_union_table() above and preclean_card_table() 4913 // below are largely identical; if you need to modify 4914 // one of these methods, please check the other method too. 4915 4916 size_t CMSCollector::preclean_card_table(ConcurrentMarkSweepGeneration* gen, 4917 ScanMarkedObjectsAgainCarefullyClosure* cl) { 4918 // strategy: it's similar to precleamModUnionTable above, in that 4919 // we accumulate contiguous ranges of dirty cards, mark these cards 4920 // precleaned, then scan the region covered by these cards. 4921 HeapWord* endAddr = (HeapWord*)(gen->_virtual_space.high()); 4922 HeapWord* startAddr = (HeapWord*)(gen->_virtual_space.low()); 4923 4924 cl->setFreelistLock(gen->freelistLock()); // needed for yielding 4925 4926 size_t numDirtyCards, cumNumDirtyCards; 4927 HeapWord *lastAddr, *nextAddr; 4928 4929 for (cumNumDirtyCards = numDirtyCards = 0, 4930 nextAddr = lastAddr = startAddr; 4931 nextAddr < endAddr; 4932 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 4933 4934 ResourceMark rm; 4935 HandleMark hm; 4936 4937 MemRegion dirtyRegion; 4938 { 4939 // See comments in "Precleaning notes" above on why we 4940 // do this locking. XXX Could the locking overheads be 4941 // too high when dirty cards are sparse? [I don't think so.] 4942 stopTimer(); 4943 CMSTokenSync x(true); // is cms thread 4944 startTimer(); 4945 sample_eden(); 4946 // Get and clear dirty region from card table 4947 dirtyRegion = _ct->ct_bs()->dirty_card_range_after_reset( 4948 MemRegion(nextAddr, endAddr), 4949 true, 4950 CardTableModRefBS::precleaned_card_val()); 4951 4952 assert(dirtyRegion.start() >= nextAddr, 4953 "returned region inconsistent?"); 4954 } 4955 lastAddr = dirtyRegion.end(); 4956 numDirtyCards = 4957 dirtyRegion.word_size()/CardTableModRefBS::card_size_in_words; 4958 4959 if (!dirtyRegion.is_empty()) { 4960 stopTimer(); 4961 CMSTokenSyncWithLocks ts(true, gen->freelistLock(), bitMapLock()); 4962 startTimer(); 4963 sample_eden(); 4964 verify_work_stacks_empty(); 4965 verify_overflow_empty(); 4966 HeapWord* stop_point = 4967 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 4968 if (stop_point != NULL) { 4969 assert((_collectorState == AbortablePreclean && should_abort_preclean()), 4970 "Should only be AbortablePreclean."); 4971 _ct->ct_bs()->invalidate(MemRegion(stop_point, dirtyRegion.end())); 4972 if (should_abort_preclean()) { 4973 break; // out of preclean loop 4974 } else { 4975 // Compute the next address at which preclean should pick up. 4976 lastAddr = next_card_start_after_block(stop_point); 4977 } 4978 } 4979 } else { 4980 break; 4981 } 4982 } 4983 verify_work_stacks_empty(); 4984 verify_overflow_empty(); 4985 return cumNumDirtyCards; 4986 } 4987 4988 class PrecleanKlassClosure : public KlassClosure { 4989 CMKlassClosure _cm_klass_closure; 4990 public: 4991 PrecleanKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {} 4992 void do_klass(Klass* k) { 4993 if (k->has_accumulated_modified_oops()) { 4994 k->clear_accumulated_modified_oops(); 4995 4996 _cm_klass_closure.do_klass(k); 4997 } 4998 } 4999 }; 5000 5001 // The freelist lock is needed to prevent asserts, is it really needed? 5002 void CMSCollector::preclean_klasses(MarkRefsIntoAndScanClosure* cl, Mutex* freelistLock) { 5003 5004 cl->set_freelistLock(freelistLock); 5005 5006 CMSTokenSyncWithLocks ts(true, freelistLock, bitMapLock()); 5007 5008 // SSS: Add equivalent to ScanMarkedObjectsAgainCarefullyClosure::do_yield_check and should_abort_preclean? 5009 // SSS: We should probably check if precleaning should be aborted, at suitable intervals? 5010 PrecleanKlassClosure preclean_klass_closure(cl); 5011 ClassLoaderDataGraph::classes_do(&preclean_klass_closure); 5012 5013 verify_work_stacks_empty(); 5014 verify_overflow_empty(); 5015 } 5016 5017 void CMSCollector::checkpointRootsFinal(bool asynch, 5018 bool clear_all_soft_refs, bool init_mark_was_synchronous) { 5019 assert(_collectorState == FinalMarking, "incorrect state transition?"); 5020 check_correct_thread_executing(); 5021 // world is stopped at this checkpoint 5022 assert(SafepointSynchronize::is_at_safepoint(), 5023 "world should be stopped"); 5024 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 5025 5026 verify_work_stacks_empty(); 5027 verify_overflow_empty(); 5028 5029 SpecializationStats::clear(); 5030 if (PrintGCDetails) { 5031 gclog_or_tty->print("[YG occupancy: "SIZE_FORMAT" K ("SIZE_FORMAT" K)]", 5032 _young_gen->used() / K, 5033 _young_gen->capacity() / K); 5034 } 5035 if (asynch) { 5036 if (CMSScavengeBeforeRemark) { 5037 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5038 // Temporarily set flag to false, GCH->do_collection will 5039 // expect it to be false and set to true 5040 FlagSetting fl(gch->_is_gc_active, false); 5041 NOT_PRODUCT(GCTraceTime t("Scavenge-Before-Remark", 5042 PrintGCDetails && Verbose, true, _gc_timer_cm);) 5043 int level = _cmsGen->level() - 1; 5044 if (level >= 0) { 5045 gch->do_collection(true, // full (i.e. force, see below) 5046 false, // !clear_all_soft_refs 5047 0, // size 5048 false, // is_tlab 5049 level // max_level 5050 ); 5051 } 5052 } 5053 FreelistLocker x(this); 5054 MutexLockerEx y(bitMapLock(), 5055 Mutex::_no_safepoint_check_flag); 5056 assert(!init_mark_was_synchronous, "but that's impossible!"); 5057 checkpointRootsFinalWork(asynch, clear_all_soft_refs, false); 5058 } else { 5059 // already have all the locks 5060 checkpointRootsFinalWork(asynch, clear_all_soft_refs, 5061 init_mark_was_synchronous); 5062 } 5063 verify_work_stacks_empty(); 5064 verify_overflow_empty(); 5065 SpecializationStats::print(); 5066 } 5067 5068 void CMSCollector::checkpointRootsFinalWork(bool asynch, 5069 bool clear_all_soft_refs, bool init_mark_was_synchronous) { 5070 5071 NOT_PRODUCT(GCTraceTime tr("checkpointRootsFinalWork", PrintGCDetails, false, _gc_timer_cm);) 5072 5073 assert(haveFreelistLocks(), "must have free list locks"); 5074 assert_lock_strong(bitMapLock()); 5075 5076 if (UseAdaptiveSizePolicy) { 5077 size_policy()->checkpoint_roots_final_begin(); 5078 } 5079 5080 ResourceMark rm; 5081 HandleMark hm; 5082 5083 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5084 5085 if (should_unload_classes()) { 5086 CodeCache::gc_prologue(); 5087 } 5088 assert(haveFreelistLocks(), "must have free list locks"); 5089 assert_lock_strong(bitMapLock()); 5090 5091 if (!init_mark_was_synchronous) { 5092 // We might assume that we need not fill TLAB's when 5093 // CMSScavengeBeforeRemark is set, because we may have just done 5094 // a scavenge which would have filled all TLAB's -- and besides 5095 // Eden would be empty. This however may not always be the case -- 5096 // for instance although we asked for a scavenge, it may not have 5097 // happened because of a JNI critical section. We probably need 5098 // a policy for deciding whether we can in that case wait until 5099 // the critical section releases and then do the remark following 5100 // the scavenge, and skip it here. In the absence of that policy, 5101 // or of an indication of whether the scavenge did indeed occur, 5102 // we cannot rely on TLAB's having been filled and must do 5103 // so here just in case a scavenge did not happen. 5104 gch->ensure_parsability(false); // fill TLAB's, but no need to retire them 5105 // Update the saved marks which may affect the root scans. 5106 gch->save_marks(); 5107 5108 if (CMSPrintEdenSurvivorChunks) { 5109 print_eden_and_survivor_chunk_arrays(); 5110 } 5111 5112 { 5113 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;) 5114 5115 // Note on the role of the mod union table: 5116 // Since the marker in "markFromRoots" marks concurrently with 5117 // mutators, it is possible for some reachable objects not to have been 5118 // scanned. For instance, an only reference to an object A was 5119 // placed in object B after the marker scanned B. Unless B is rescanned, 5120 // A would be collected. Such updates to references in marked objects 5121 // are detected via the mod union table which is the set of all cards 5122 // dirtied since the first checkpoint in this GC cycle and prior to 5123 // the most recent young generation GC, minus those cleaned up by the 5124 // concurrent precleaning. 5125 if (CMSParallelRemarkEnabled && CollectedHeap::use_parallel_gc_threads()) { 5126 GCTraceTime t("Rescan (parallel) ", PrintGCDetails, false, _gc_timer_cm); 5127 do_remark_parallel(); 5128 } else { 5129 GCTraceTime t("Rescan (non-parallel) ", PrintGCDetails, false, 5130 _gc_timer_cm); 5131 do_remark_non_parallel(); 5132 } 5133 } 5134 } else { 5135 assert(!asynch, "Can't have init_mark_was_synchronous in asynch mode"); 5136 // The initial mark was stop-world, so there's no rescanning to 5137 // do; go straight on to the next step below. 5138 } 5139 verify_work_stacks_empty(); 5140 verify_overflow_empty(); 5141 5142 { 5143 NOT_PRODUCT(GCTraceTime ts("refProcessingWork", PrintGCDetails, false, _gc_timer_cm);) 5144 refProcessingWork(asynch, clear_all_soft_refs); 5145 } 5146 verify_work_stacks_empty(); 5147 verify_overflow_empty(); 5148 5149 if (should_unload_classes()) { 5150 CodeCache::gc_epilogue(); 5151 } 5152 JvmtiExport::gc_epilogue(); 5153 5154 // If we encountered any (marking stack / work queue) overflow 5155 // events during the current CMS cycle, take appropriate 5156 // remedial measures, where possible, so as to try and avoid 5157 // recurrence of that condition. 5158 assert(_markStack.isEmpty(), "No grey objects"); 5159 size_t ser_ovflw = _ser_pmc_remark_ovflw + _ser_pmc_preclean_ovflw + 5160 _ser_kac_ovflw + _ser_kac_preclean_ovflw; 5161 if (ser_ovflw > 0) { 5162 if (PrintCMSStatistics != 0) { 5163 gclog_or_tty->print_cr("Marking stack overflow (benign) " 5164 "(pmc_pc="SIZE_FORMAT", pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT 5165 ", kac_preclean="SIZE_FORMAT")", 5166 _ser_pmc_preclean_ovflw, _ser_pmc_remark_ovflw, 5167 _ser_kac_ovflw, _ser_kac_preclean_ovflw); 5168 } 5169 _markStack.expand(); 5170 _ser_pmc_remark_ovflw = 0; 5171 _ser_pmc_preclean_ovflw = 0; 5172 _ser_kac_preclean_ovflw = 0; 5173 _ser_kac_ovflw = 0; 5174 } 5175 if (_par_pmc_remark_ovflw > 0 || _par_kac_ovflw > 0) { 5176 if (PrintCMSStatistics != 0) { 5177 gclog_or_tty->print_cr("Work queue overflow (benign) " 5178 "(pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT")", 5179 _par_pmc_remark_ovflw, _par_kac_ovflw); 5180 } 5181 _par_pmc_remark_ovflw = 0; 5182 _par_kac_ovflw = 0; 5183 } 5184 if (PrintCMSStatistics != 0) { 5185 if (_markStack._hit_limit > 0) { 5186 gclog_or_tty->print_cr(" (benign) Hit max stack size limit ("SIZE_FORMAT")", 5187 _markStack._hit_limit); 5188 } 5189 if (_markStack._failed_double > 0) { 5190 gclog_or_tty->print_cr(" (benign) Failed stack doubling ("SIZE_FORMAT")," 5191 " current capacity "SIZE_FORMAT, 5192 _markStack._failed_double, 5193 _markStack.capacity()); 5194 } 5195 } 5196 _markStack._hit_limit = 0; 5197 _markStack._failed_double = 0; 5198 5199 if ((VerifyAfterGC || VerifyDuringGC) && 5200 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 5201 verify_after_remark(); 5202 } 5203 5204 _gc_tracer_cm->report_object_count_after_gc(&_is_alive_closure); 5205 5206 // Change under the freelistLocks. 5207 _collectorState = Sweeping; 5208 // Call isAllClear() under bitMapLock 5209 assert(_modUnionTable.isAllClear(), 5210 "Should be clear by end of the final marking"); 5211 assert(_ct->klass_rem_set()->mod_union_is_clear(), 5212 "Should be clear by end of the final marking"); 5213 if (UseAdaptiveSizePolicy) { 5214 size_policy()->checkpoint_roots_final_end(gch->gc_cause()); 5215 } 5216 } 5217 5218 void CMSParInitialMarkTask::work(uint worker_id) { 5219 elapsedTimer _timer; 5220 ResourceMark rm; 5221 HandleMark hm; 5222 5223 // ---------- scan from roots -------------- 5224 _timer.start(); 5225 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5226 Par_MarkRefsIntoClosure par_mri_cl(_collector->_span, &(_collector->_markBitMap)); 5227 CMKlassClosure klass_closure(&par_mri_cl); 5228 5229 // ---------- young gen roots -------------- 5230 { 5231 work_on_young_gen_roots(worker_id, &par_mri_cl); 5232 _timer.stop(); 5233 if (PrintCMSStatistics != 0) { 5234 gclog_or_tty->print_cr( 5235 "Finished young gen initial mark scan work in %dth thread: %3.3f sec", 5236 worker_id, _timer.seconds()); 5237 } 5238 } 5239 5240 // ---------- remaining roots -------------- 5241 _timer.reset(); 5242 _timer.start(); 5243 gch->gen_process_strong_roots(_collector->_cmsGen->level(), 5244 false, // yg was scanned above 5245 false, // this is parallel code 5246 SharedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()), 5247 &par_mri_cl, 5248 NULL, 5249 &klass_closure); 5250 assert(_collector->should_unload_classes() 5251 || (_collector->CMSCollector::roots_scanning_options() & SharedHeap::SO_AllCodeCache), 5252 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 5253 _timer.stop(); 5254 if (PrintCMSStatistics != 0) { 5255 gclog_or_tty->print_cr( 5256 "Finished remaining root initial mark scan work in %dth thread: %3.3f sec", 5257 worker_id, _timer.seconds()); 5258 } 5259 } 5260 5261 // Parallel remark task 5262 class CMSParRemarkTask: public CMSParMarkTask { 5263 CompactibleFreeListSpace* _cms_space; 5264 5265 // The per-thread work queues, available here for stealing. 5266 OopTaskQueueSet* _task_queues; 5267 ParallelTaskTerminator _term; 5268 5269 public: 5270 // A value of 0 passed to n_workers will cause the number of 5271 // workers to be taken from the active workers in the work gang. 5272 CMSParRemarkTask(CMSCollector* collector, 5273 CompactibleFreeListSpace* cms_space, 5274 int n_workers, FlexibleWorkGang* workers, 5275 OopTaskQueueSet* task_queues): 5276 CMSParMarkTask("Rescan roots and grey objects in parallel", 5277 collector, n_workers), 5278 _cms_space(cms_space), 5279 _task_queues(task_queues), 5280 _term(n_workers, task_queues) { } 5281 5282 OopTaskQueueSet* task_queues() { return _task_queues; } 5283 5284 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 5285 5286 ParallelTaskTerminator* terminator() { return &_term; } 5287 int n_workers() { return _n_workers; } 5288 5289 void work(uint worker_id); 5290 5291 private: 5292 // ... of dirty cards in old space 5293 void do_dirty_card_rescan_tasks(CompactibleFreeListSpace* sp, int i, 5294 Par_MarkRefsIntoAndScanClosure* cl); 5295 5296 // ... work stealing for the above 5297 void do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, int* seed); 5298 }; 5299 5300 class RemarkKlassClosure : public KlassClosure { 5301 CMKlassClosure _cm_klass_closure; 5302 public: 5303 RemarkKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {} 5304 void do_klass(Klass* k) { 5305 // Check if we have modified any oops in the Klass during the concurrent marking. 5306 if (k->has_accumulated_modified_oops()) { 5307 k->clear_accumulated_modified_oops(); 5308 5309 // We could have transfered the current modified marks to the accumulated marks, 5310 // like we do with the Card Table to Mod Union Table. But it's not really necessary. 5311 } else if (k->has_modified_oops()) { 5312 // Don't clear anything, this info is needed by the next young collection. 5313 } else { 5314 // No modified oops in the Klass. 5315 return; 5316 } 5317 5318 // The klass has modified fields, need to scan the klass. 5319 _cm_klass_closure.do_klass(k); 5320 } 5321 }; 5322 5323 void CMSParMarkTask::work_on_young_gen_roots(uint worker_id, OopsInGenClosure* cl) { 5324 DefNewGeneration* dng = _collector->_young_gen->as_DefNewGeneration(); 5325 EdenSpace* eden_space = dng->eden(); 5326 ContiguousSpace* from_space = dng->from(); 5327 ContiguousSpace* to_space = dng->to(); 5328 5329 HeapWord** eca = _collector->_eden_chunk_array; 5330 size_t ect = _collector->_eden_chunk_index; 5331 HeapWord** sca = _collector->_survivor_chunk_array; 5332 size_t sct = _collector->_survivor_chunk_index; 5333 5334 assert(ect <= _collector->_eden_chunk_capacity, "out of bounds"); 5335 assert(sct <= _collector->_survivor_chunk_capacity, "out of bounds"); 5336 5337 do_young_space_rescan(worker_id, cl, to_space, NULL, 0); 5338 do_young_space_rescan(worker_id, cl, from_space, sca, sct); 5339 do_young_space_rescan(worker_id, cl, eden_space, eca, ect); 5340 } 5341 5342 // work_queue(i) is passed to the closure 5343 // Par_MarkRefsIntoAndScanClosure. The "i" parameter 5344 // also is passed to do_dirty_card_rescan_tasks() and to 5345 // do_work_steal() to select the i-th task_queue. 5346 5347 void CMSParRemarkTask::work(uint worker_id) { 5348 elapsedTimer _timer; 5349 ResourceMark rm; 5350 HandleMark hm; 5351 5352 // ---------- rescan from roots -------------- 5353 _timer.start(); 5354 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5355 Par_MarkRefsIntoAndScanClosure par_mrias_cl(_collector, 5356 _collector->_span, _collector->ref_processor(), 5357 &(_collector->_markBitMap), 5358 work_queue(worker_id)); 5359 5360 // Rescan young gen roots first since these are likely 5361 // coarsely partitioned and may, on that account, constitute 5362 // the critical path; thus, it's best to start off that 5363 // work first. 5364 // ---------- young gen roots -------------- 5365 { 5366 work_on_young_gen_roots(worker_id, &par_mrias_cl); 5367 _timer.stop(); 5368 if (PrintCMSStatistics != 0) { 5369 gclog_or_tty->print_cr( 5370 "Finished young gen rescan work in %dth thread: %3.3f sec", 5371 worker_id, _timer.seconds()); 5372 } 5373 } 5374 5375 // ---------- remaining roots -------------- 5376 _timer.reset(); 5377 _timer.start(); 5378 gch->gen_process_strong_roots(_collector->_cmsGen->level(), 5379 false, // yg was scanned above 5380 false, // this is parallel code 5381 SharedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()), 5382 &par_mrias_cl, 5383 NULL, 5384 NULL); // The dirty klasses will be handled below 5385 assert(_collector->should_unload_classes() 5386 || (_collector->CMSCollector::roots_scanning_options() & SharedHeap::SO_AllCodeCache), 5387 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 5388 _timer.stop(); 5389 if (PrintCMSStatistics != 0) { 5390 gclog_or_tty->print_cr( 5391 "Finished remaining root rescan work in %dth thread: %3.3f sec", 5392 worker_id, _timer.seconds()); 5393 } 5394 5395 // ---------- unhandled CLD scanning ---------- 5396 if (worker_id == 0) { // Single threaded at the moment. 5397 _timer.reset(); 5398 _timer.start(); 5399 5400 // Scan all new class loader data objects and new dependencies that were 5401 // introduced during concurrent marking. 5402 ResourceMark rm; 5403 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds(); 5404 for (int i = 0; i < array->length(); i++) { 5405 par_mrias_cl.do_class_loader_data(array->at(i)); 5406 } 5407 5408 // We don't need to keep track of new CLDs anymore. 5409 ClassLoaderDataGraph::remember_new_clds(false); 5410 5411 _timer.stop(); 5412 if (PrintCMSStatistics != 0) { 5413 gclog_or_tty->print_cr( 5414 "Finished unhandled CLD scanning work in %dth thread: %3.3f sec", 5415 worker_id, _timer.seconds()); 5416 } 5417 } 5418 5419 // ---------- dirty klass scanning ---------- 5420 if (worker_id == 0) { // Single threaded at the moment. 5421 _timer.reset(); 5422 _timer.start(); 5423 5424 // Scan all classes that was dirtied during the concurrent marking phase. 5425 RemarkKlassClosure remark_klass_closure(&par_mrias_cl); 5426 ClassLoaderDataGraph::classes_do(&remark_klass_closure); 5427 5428 _timer.stop(); 5429 if (PrintCMSStatistics != 0) { 5430 gclog_or_tty->print_cr( 5431 "Finished dirty klass scanning work in %dth thread: %3.3f sec", 5432 worker_id, _timer.seconds()); 5433 } 5434 } 5435 5436 // We might have added oops to ClassLoaderData::_handles during the 5437 // concurrent marking phase. These oops point to newly allocated objects 5438 // that are guaranteed to be kept alive. Either by the direct allocation 5439 // code, or when the young collector processes the strong roots. Hence, 5440 // we don't have to revisit the _handles block during the remark phase. 5441 5442 // ---------- rescan dirty cards ------------ 5443 _timer.reset(); 5444 _timer.start(); 5445 5446 // Do the rescan tasks for each of the two spaces 5447 // (cms_space) in turn. 5448 // "worker_id" is passed to select the task_queue for "worker_id" 5449 do_dirty_card_rescan_tasks(_cms_space, worker_id, &par_mrias_cl); 5450 _timer.stop(); 5451 if (PrintCMSStatistics != 0) { 5452 gclog_or_tty->print_cr( 5453 "Finished dirty card rescan work in %dth thread: %3.3f sec", 5454 worker_id, _timer.seconds()); 5455 } 5456 5457 // ---------- steal work from other threads ... 5458 // ---------- ... and drain overflow list. 5459 _timer.reset(); 5460 _timer.start(); 5461 do_work_steal(worker_id, &par_mrias_cl, _collector->hash_seed(worker_id)); 5462 _timer.stop(); 5463 if (PrintCMSStatistics != 0) { 5464 gclog_or_tty->print_cr( 5465 "Finished work stealing in %dth thread: %3.3f sec", 5466 worker_id, _timer.seconds()); 5467 } 5468 } 5469 5470 // Note that parameter "i" is not used. 5471 void 5472 CMSParMarkTask::do_young_space_rescan(uint worker_id, 5473 OopsInGenClosure* cl, ContiguousSpace* space, 5474 HeapWord** chunk_array, size_t chunk_top) { 5475 // Until all tasks completed: 5476 // . claim an unclaimed task 5477 // . compute region boundaries corresponding to task claimed 5478 // using chunk_array 5479 // . par_oop_iterate(cl) over that region 5480 5481 ResourceMark rm; 5482 HandleMark hm; 5483 5484 SequentialSubTasksDone* pst = space->par_seq_tasks(); 5485 5486 uint nth_task = 0; 5487 uint n_tasks = pst->n_tasks(); 5488 5489 if (n_tasks > 0) { 5490 assert(pst->valid(), "Uninitialized use?"); 5491 HeapWord *start, *end; 5492 while (!pst->is_task_claimed(/* reference */ nth_task)) { 5493 // We claimed task # nth_task; compute its boundaries. 5494 if (chunk_top == 0) { // no samples were taken 5495 assert(nth_task == 0 && n_tasks == 1, "Can have only 1 EdenSpace task"); 5496 start = space->bottom(); 5497 end = space->top(); 5498 } else if (nth_task == 0) { 5499 start = space->bottom(); 5500 end = chunk_array[nth_task]; 5501 } else if (nth_task < (uint)chunk_top) { 5502 assert(nth_task >= 1, "Control point invariant"); 5503 start = chunk_array[nth_task - 1]; 5504 end = chunk_array[nth_task]; 5505 } else { 5506 assert(nth_task == (uint)chunk_top, "Control point invariant"); 5507 start = chunk_array[chunk_top - 1]; 5508 end = space->top(); 5509 } 5510 MemRegion mr(start, end); 5511 // Verify that mr is in space 5512 assert(mr.is_empty() || space->used_region().contains(mr), 5513 "Should be in space"); 5514 // Verify that "start" is an object boundary 5515 assert(mr.is_empty() || oop(mr.start())->is_oop(), 5516 "Should be an oop"); 5517 space->par_oop_iterate(mr, cl); 5518 } 5519 pst->all_tasks_completed(); 5520 } 5521 } 5522 5523 void 5524 CMSParRemarkTask::do_dirty_card_rescan_tasks( 5525 CompactibleFreeListSpace* sp, int i, 5526 Par_MarkRefsIntoAndScanClosure* cl) { 5527 // Until all tasks completed: 5528 // . claim an unclaimed task 5529 // . compute region boundaries corresponding to task claimed 5530 // . transfer dirty bits ct->mut for that region 5531 // . apply rescanclosure to dirty mut bits for that region 5532 5533 ResourceMark rm; 5534 HandleMark hm; 5535 5536 OopTaskQueue* work_q = work_queue(i); 5537 ModUnionClosure modUnionClosure(&(_collector->_modUnionTable)); 5538 // CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! 5539 // CAUTION: This closure has state that persists across calls to 5540 // the work method dirty_range_iterate_clear() in that it has 5541 // embedded in it a (subtype of) UpwardsObjectClosure. The 5542 // use of that state in the embedded UpwardsObjectClosure instance 5543 // assumes that the cards are always iterated (even if in parallel 5544 // by several threads) in monotonically increasing order per each 5545 // thread. This is true of the implementation below which picks 5546 // card ranges (chunks) in monotonically increasing order globally 5547 // and, a-fortiori, in monotonically increasing order per thread 5548 // (the latter order being a subsequence of the former). 5549 // If the work code below is ever reorganized into a more chaotic 5550 // work-partitioning form than the current "sequential tasks" 5551 // paradigm, the use of that persistent state will have to be 5552 // revisited and modified appropriately. See also related 5553 // bug 4756801 work on which should examine this code to make 5554 // sure that the changes there do not run counter to the 5555 // assumptions made here and necessary for correctness and 5556 // efficiency. Note also that this code might yield inefficient 5557 // behavior in the case of very large objects that span one or 5558 // more work chunks. Such objects would potentially be scanned 5559 // several times redundantly. Work on 4756801 should try and 5560 // address that performance anomaly if at all possible. XXX 5561 MemRegion full_span = _collector->_span; 5562 CMSBitMap* bm = &(_collector->_markBitMap); // shared 5563 MarkFromDirtyCardsClosure 5564 greyRescanClosure(_collector, full_span, // entire span of interest 5565 sp, bm, work_q, cl); 5566 5567 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 5568 assert(pst->valid(), "Uninitialized use?"); 5569 uint nth_task = 0; 5570 const int alignment = CardTableModRefBS::card_size * BitsPerWord; 5571 MemRegion span = sp->used_region(); 5572 HeapWord* start_addr = span.start(); 5573 HeapWord* end_addr = (HeapWord*)round_to((intptr_t)span.end(), 5574 alignment); 5575 const size_t chunk_size = sp->rescan_task_size(); // in HeapWord units 5576 assert((HeapWord*)round_to((intptr_t)start_addr, alignment) == 5577 start_addr, "Check alignment"); 5578 assert((size_t)round_to((intptr_t)chunk_size, alignment) == 5579 chunk_size, "Check alignment"); 5580 5581 while (!pst->is_task_claimed(/* reference */ nth_task)) { 5582 // Having claimed the nth_task, compute corresponding mem-region, 5583 // which is a-fortiori aligned correctly (i.e. at a MUT boundary). 5584 // The alignment restriction ensures that we do not need any 5585 // synchronization with other gang-workers while setting or 5586 // clearing bits in thus chunk of the MUT. 5587 MemRegion this_span = MemRegion(start_addr + nth_task*chunk_size, 5588 start_addr + (nth_task+1)*chunk_size); 5589 // The last chunk's end might be way beyond end of the 5590 // used region. In that case pull back appropriately. 5591 if (this_span.end() > end_addr) { 5592 this_span.set_end(end_addr); 5593 assert(!this_span.is_empty(), "Program logic (calculation of n_tasks)"); 5594 } 5595 // Iterate over the dirty cards covering this chunk, marking them 5596 // precleaned, and setting the corresponding bits in the mod union 5597 // table. Since we have been careful to partition at Card and MUT-word 5598 // boundaries no synchronization is needed between parallel threads. 5599 _collector->_ct->ct_bs()->dirty_card_iterate(this_span, 5600 &modUnionClosure); 5601 5602 // Having transferred these marks into the modUnionTable, 5603 // rescan the marked objects on the dirty cards in the modUnionTable. 5604 // Even if this is at a synchronous collection, the initial marking 5605 // may have been done during an asynchronous collection so there 5606 // may be dirty bits in the mod-union table. 5607 _collector->_modUnionTable.dirty_range_iterate_clear( 5608 this_span, &greyRescanClosure); 5609 _collector->_modUnionTable.verifyNoOneBitsInRange( 5610 this_span.start(), 5611 this_span.end()); 5612 } 5613 pst->all_tasks_completed(); // declare that i am done 5614 } 5615 5616 // . see if we can share work_queues with ParNew? XXX 5617 void 5618 CMSParRemarkTask::do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, 5619 int* seed) { 5620 OopTaskQueue* work_q = work_queue(i); 5621 NOT_PRODUCT(int num_steals = 0;) 5622 oop obj_to_scan; 5623 CMSBitMap* bm = &(_collector->_markBitMap); 5624 5625 while (true) { 5626 // Completely finish any left over work from (an) earlier round(s) 5627 cl->trim_queue(0); 5628 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 5629 (size_t)ParGCDesiredObjsFromOverflowList); 5630 // Now check if there's any work in the overflow list 5631 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads, 5632 // only affects the number of attempts made to get work from the 5633 // overflow list and does not affect the number of workers. Just 5634 // pass ParallelGCThreads so this behavior is unchanged. 5635 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 5636 work_q, 5637 ParallelGCThreads)) { 5638 // found something in global overflow list; 5639 // not yet ready to go stealing work from others. 5640 // We'd like to assert(work_q->size() != 0, ...) 5641 // because we just took work from the overflow list, 5642 // but of course we can't since all of that could have 5643 // been already stolen from us. 5644 // "He giveth and He taketh away." 5645 continue; 5646 } 5647 // Verify that we have no work before we resort to stealing 5648 assert(work_q->size() == 0, "Have work, shouldn't steal"); 5649 // Try to steal from other queues that have work 5650 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 5651 NOT_PRODUCT(num_steals++;) 5652 assert(obj_to_scan->is_oop(), "Oops, not an oop!"); 5653 assert(bm->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 5654 // Do scanning work 5655 obj_to_scan->oop_iterate(cl); 5656 // Loop around, finish this work, and try to steal some more 5657 } else if (terminator()->offer_termination()) { 5658 break; // nirvana from the infinite cycle 5659 } 5660 } 5661 NOT_PRODUCT( 5662 if (PrintCMSStatistics != 0) { 5663 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals); 5664 } 5665 ) 5666 assert(work_q->size() == 0 && _collector->overflow_list_is_empty(), 5667 "Else our work is not yet done"); 5668 } 5669 5670 // Record object boundaries in _eden_chunk_array by sampling the eden 5671 // top in the slow-path eden object allocation code path and record 5672 // the boundaries, if CMSEdenChunksRecordAlways is true. If 5673 // CMSEdenChunksRecordAlways is false, we use the other asynchronous 5674 // sampling in sample_eden() that activates during the part of the 5675 // preclean phase. 5676 void CMSCollector::sample_eden_chunk() { 5677 if (CMSEdenChunksRecordAlways && _eden_chunk_array != NULL) { 5678 if (_eden_chunk_lock->try_lock()) { 5679 // Record a sample. This is the critical section. The contents 5680 // of the _eden_chunk_array have to be non-decreasing in the 5681 // address order. 5682 _eden_chunk_array[_eden_chunk_index] = *_top_addr; 5683 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr, 5684 "Unexpected state of Eden"); 5685 if (_eden_chunk_index == 0 || 5686 ((_eden_chunk_array[_eden_chunk_index] > _eden_chunk_array[_eden_chunk_index-1]) && 5687 (pointer_delta(_eden_chunk_array[_eden_chunk_index], 5688 _eden_chunk_array[_eden_chunk_index-1]) >= CMSSamplingGrain))) { 5689 _eden_chunk_index++; // commit sample 5690 } 5691 _eden_chunk_lock->unlock(); 5692 } 5693 } 5694 } 5695 5696 // Return a thread-local PLAB recording array, as appropriate. 5697 void* CMSCollector::get_data_recorder(int thr_num) { 5698 if (_survivor_plab_array != NULL && 5699 (CMSPLABRecordAlways || 5700 (_collectorState > Marking && _collectorState < FinalMarking))) { 5701 assert(thr_num < (int)ParallelGCThreads, "thr_num is out of bounds"); 5702 ChunkArray* ca = &_survivor_plab_array[thr_num]; 5703 ca->reset(); // clear it so that fresh data is recorded 5704 return (void*) ca; 5705 } else { 5706 return NULL; 5707 } 5708 } 5709 5710 // Reset all the thread-local PLAB recording arrays 5711 void CMSCollector::reset_survivor_plab_arrays() { 5712 for (uint i = 0; i < ParallelGCThreads; i++) { 5713 _survivor_plab_array[i].reset(); 5714 } 5715 } 5716 5717 // Merge the per-thread plab arrays into the global survivor chunk 5718 // array which will provide the partitioning of the survivor space 5719 // for CMS initial scan and rescan. 5720 void CMSCollector::merge_survivor_plab_arrays(ContiguousSpace* surv, 5721 int no_of_gc_threads) { 5722 assert(_survivor_plab_array != NULL, "Error"); 5723 assert(_survivor_chunk_array != NULL, "Error"); 5724 assert(_collectorState == FinalMarking || 5725 (CMSParallelInitialMarkEnabled && _collectorState == InitialMarking), "Error"); 5726 for (int j = 0; j < no_of_gc_threads; j++) { 5727 _cursor[j] = 0; 5728 } 5729 HeapWord* top = surv->top(); 5730 size_t i; 5731 for (i = 0; i < _survivor_chunk_capacity; i++) { // all sca entries 5732 HeapWord* min_val = top; // Higher than any PLAB address 5733 uint min_tid = 0; // position of min_val this round 5734 for (int j = 0; j < no_of_gc_threads; j++) { 5735 ChunkArray* cur_sca = &_survivor_plab_array[j]; 5736 if (_cursor[j] == cur_sca->end()) { 5737 continue; 5738 } 5739 assert(_cursor[j] < cur_sca->end(), "ctl pt invariant"); 5740 HeapWord* cur_val = cur_sca->nth(_cursor[j]); 5741 assert(surv->used_region().contains(cur_val), "Out of bounds value"); 5742 if (cur_val < min_val) { 5743 min_tid = j; 5744 min_val = cur_val; 5745 } else { 5746 assert(cur_val < top, "All recorded addresses should be less"); 5747 } 5748 } 5749 // At this point min_val and min_tid are respectively 5750 // the least address in _survivor_plab_array[j]->nth(_cursor[j]) 5751 // and the thread (j) that witnesses that address. 5752 // We record this address in the _survivor_chunk_array[i] 5753 // and increment _cursor[min_tid] prior to the next round i. 5754 if (min_val == top) { 5755 break; 5756 } 5757 _survivor_chunk_array[i] = min_val; 5758 _cursor[min_tid]++; 5759 } 5760 // We are all done; record the size of the _survivor_chunk_array 5761 _survivor_chunk_index = i; // exclusive: [0, i) 5762 if (PrintCMSStatistics > 0) { 5763 gclog_or_tty->print(" (Survivor:" SIZE_FORMAT "chunks) ", i); 5764 } 5765 // Verify that we used up all the recorded entries 5766 #ifdef ASSERT 5767 size_t total = 0; 5768 for (int j = 0; j < no_of_gc_threads; j++) { 5769 assert(_cursor[j] == _survivor_plab_array[j].end(), "Ctl pt invariant"); 5770 total += _cursor[j]; 5771 } 5772 assert(total == _survivor_chunk_index, "Ctl Pt Invariant"); 5773 // Check that the merged array is in sorted order 5774 if (total > 0) { 5775 for (size_t i = 0; i < total - 1; i++) { 5776 if (PrintCMSStatistics > 0) { 5777 gclog_or_tty->print(" (chunk" SIZE_FORMAT ":" INTPTR_FORMAT ") ", 5778 i, _survivor_chunk_array[i]); 5779 } 5780 assert(_survivor_chunk_array[i] < _survivor_chunk_array[i+1], 5781 "Not sorted"); 5782 } 5783 } 5784 #endif // ASSERT 5785 } 5786 5787 // Set up the space's par_seq_tasks structure for work claiming 5788 // for parallel initial scan and rescan of young gen. 5789 // See ParRescanTask where this is currently used. 5790 void 5791 CMSCollector:: 5792 initialize_sequential_subtasks_for_young_gen_rescan(int n_threads) { 5793 assert(n_threads > 0, "Unexpected n_threads argument"); 5794 DefNewGeneration* dng = (DefNewGeneration*)_young_gen; 5795 5796 // Eden space 5797 if (!dng->eden()->is_empty()) { 5798 SequentialSubTasksDone* pst = dng->eden()->par_seq_tasks(); 5799 assert(!pst->valid(), "Clobbering existing data?"); 5800 // Each valid entry in [0, _eden_chunk_index) represents a task. 5801 size_t n_tasks = _eden_chunk_index + 1; 5802 assert(n_tasks == 1 || _eden_chunk_array != NULL, "Error"); 5803 // Sets the condition for completion of the subtask (how many threads 5804 // need to finish in order to be done). 5805 pst->set_n_threads(n_threads); 5806 pst->set_n_tasks((int)n_tasks); 5807 } 5808 5809 // Merge the survivor plab arrays into _survivor_chunk_array 5810 if (_survivor_plab_array != NULL) { 5811 merge_survivor_plab_arrays(dng->from(), n_threads); 5812 } else { 5813 assert(_survivor_chunk_index == 0, "Error"); 5814 } 5815 5816 // To space 5817 { 5818 SequentialSubTasksDone* pst = dng->to()->par_seq_tasks(); 5819 assert(!pst->valid(), "Clobbering existing data?"); 5820 // Sets the condition for completion of the subtask (how many threads 5821 // need to finish in order to be done). 5822 pst->set_n_threads(n_threads); 5823 pst->set_n_tasks(1); 5824 assert(pst->valid(), "Error"); 5825 } 5826 5827 // From space 5828 { 5829 SequentialSubTasksDone* pst = dng->from()->par_seq_tasks(); 5830 assert(!pst->valid(), "Clobbering existing data?"); 5831 size_t n_tasks = _survivor_chunk_index + 1; 5832 assert(n_tasks == 1 || _survivor_chunk_array != NULL, "Error"); 5833 // Sets the condition for completion of the subtask (how many threads 5834 // need to finish in order to be done). 5835 pst->set_n_threads(n_threads); 5836 pst->set_n_tasks((int)n_tasks); 5837 assert(pst->valid(), "Error"); 5838 } 5839 } 5840 5841 // Parallel version of remark 5842 void CMSCollector::do_remark_parallel() { 5843 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5844 FlexibleWorkGang* workers = gch->workers(); 5845 assert(workers != NULL, "Need parallel worker threads."); 5846 // Choose to use the number of GC workers most recently set 5847 // into "active_workers". If active_workers is not set, set it 5848 // to ParallelGCThreads. 5849 int n_workers = workers->active_workers(); 5850 if (n_workers == 0) { 5851 assert(n_workers > 0, "Should have been set during scavenge"); 5852 n_workers = ParallelGCThreads; 5853 workers->set_active_workers(n_workers); 5854 } 5855 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 5856 5857 CMSParRemarkTask tsk(this, 5858 cms_space, 5859 n_workers, workers, task_queues()); 5860 5861 // Set up for parallel process_strong_roots work. 5862 gch->set_par_threads(n_workers); 5863 // We won't be iterating over the cards in the card table updating 5864 // the younger_gen cards, so we shouldn't call the following else 5865 // the verification code as well as subsequent younger_refs_iterate 5866 // code would get confused. XXX 5867 // gch->rem_set()->prepare_for_younger_refs_iterate(true); // parallel 5868 5869 // The young gen rescan work will not be done as part of 5870 // process_strong_roots (which currently doesn't knw how to 5871 // parallelize such a scan), but rather will be broken up into 5872 // a set of parallel tasks (via the sampling that the [abortable] 5873 // preclean phase did of EdenSpace, plus the [two] tasks of 5874 // scanning the [two] survivor spaces. Further fine-grain 5875 // parallelization of the scanning of the survivor spaces 5876 // themselves, and of precleaning of the younger gen itself 5877 // is deferred to the future. 5878 initialize_sequential_subtasks_for_young_gen_rescan(n_workers); 5879 5880 // The dirty card rescan work is broken up into a "sequence" 5881 // of parallel tasks (per constituent space) that are dynamically 5882 // claimed by the parallel threads. 5883 cms_space->initialize_sequential_subtasks_for_rescan(n_workers); 5884 5885 // It turns out that even when we're using 1 thread, doing the work in a 5886 // separate thread causes wide variance in run times. We can't help this 5887 // in the multi-threaded case, but we special-case n=1 here to get 5888 // repeatable measurements of the 1-thread overhead of the parallel code. 5889 if (n_workers > 1) { 5890 // Make refs discovery MT-safe, if it isn't already: it may not 5891 // necessarily be so, since it's possible that we are doing 5892 // ST marking. 5893 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), true); 5894 GenCollectedHeap::StrongRootsScope srs(gch); 5895 workers->run_task(&tsk); 5896 } else { 5897 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false); 5898 GenCollectedHeap::StrongRootsScope srs(gch); 5899 tsk.work(0); 5900 } 5901 5902 gch->set_par_threads(0); // 0 ==> non-parallel. 5903 // restore, single-threaded for now, any preserved marks 5904 // as a result of work_q overflow 5905 restore_preserved_marks_if_any(); 5906 } 5907 5908 // Non-parallel version of remark 5909 void CMSCollector::do_remark_non_parallel() { 5910 ResourceMark rm; 5911 HandleMark hm; 5912 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5913 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false); 5914 5915 MarkRefsIntoAndScanClosure 5916 mrias_cl(_span, ref_processor(), &_markBitMap, NULL /* not precleaning */, 5917 &_markStack, this, 5918 false /* should_yield */, false /* not precleaning */); 5919 MarkFromDirtyCardsClosure 5920 markFromDirtyCardsClosure(this, _span, 5921 NULL, // space is set further below 5922 &_markBitMap, &_markStack, &mrias_cl); 5923 { 5924 GCTraceTime t("grey object rescan", PrintGCDetails, false, _gc_timer_cm); 5925 // Iterate over the dirty cards, setting the corresponding bits in the 5926 // mod union table. 5927 { 5928 ModUnionClosure modUnionClosure(&_modUnionTable); 5929 _ct->ct_bs()->dirty_card_iterate( 5930 _cmsGen->used_region(), 5931 &modUnionClosure); 5932 } 5933 // Having transferred these marks into the modUnionTable, we just need 5934 // to rescan the marked objects on the dirty cards in the modUnionTable. 5935 // The initial marking may have been done during an asynchronous 5936 // collection so there may be dirty bits in the mod-union table. 5937 const int alignment = 5938 CardTableModRefBS::card_size * BitsPerWord; 5939 { 5940 // ... First handle dirty cards in CMS gen 5941 markFromDirtyCardsClosure.set_space(_cmsGen->cmsSpace()); 5942 MemRegion ur = _cmsGen->used_region(); 5943 HeapWord* lb = ur.start(); 5944 HeapWord* ub = (HeapWord*)round_to((intptr_t)ur.end(), alignment); 5945 MemRegion cms_span(lb, ub); 5946 _modUnionTable.dirty_range_iterate_clear(cms_span, 5947 &markFromDirtyCardsClosure); 5948 verify_work_stacks_empty(); 5949 if (PrintCMSStatistics != 0) { 5950 gclog_or_tty->print(" (re-scanned "SIZE_FORMAT" dirty cards in cms gen) ", 5951 markFromDirtyCardsClosure.num_dirty_cards()); 5952 } 5953 } 5954 } 5955 if (VerifyDuringGC && 5956 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 5957 HandleMark hm; // Discard invalid handles created during verification 5958 Universe::verify(); 5959 } 5960 { 5961 GCTraceTime t("root rescan", PrintGCDetails, false, _gc_timer_cm); 5962 5963 verify_work_stacks_empty(); 5964 5965 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 5966 GenCollectedHeap::StrongRootsScope srs(gch); 5967 gch->gen_process_strong_roots(_cmsGen->level(), 5968 true, // younger gens as roots 5969 false, // use the local StrongRootsScope 5970 SharedHeap::ScanningOption(roots_scanning_options()), 5971 &mrias_cl, 5972 NULL, 5973 NULL); // The dirty klasses will be handled below 5974 5975 assert(should_unload_classes() 5976 || (roots_scanning_options() & SharedHeap::SO_AllCodeCache), 5977 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 5978 } 5979 5980 { 5981 GCTraceTime t("visit unhandled CLDs", PrintGCDetails, false, _gc_timer_cm); 5982 5983 verify_work_stacks_empty(); 5984 5985 // Scan all class loader data objects that might have been introduced 5986 // during concurrent marking. 5987 ResourceMark rm; 5988 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds(); 5989 for (int i = 0; i < array->length(); i++) { 5990 mrias_cl.do_class_loader_data(array->at(i)); 5991 } 5992 5993 // We don't need to keep track of new CLDs anymore. 5994 ClassLoaderDataGraph::remember_new_clds(false); 5995 5996 verify_work_stacks_empty(); 5997 } 5998 5999 { 6000 GCTraceTime t("dirty klass scan", PrintGCDetails, false, _gc_timer_cm); 6001 6002 verify_work_stacks_empty(); 6003 6004 RemarkKlassClosure remark_klass_closure(&mrias_cl); 6005 ClassLoaderDataGraph::classes_do(&remark_klass_closure); 6006 6007 verify_work_stacks_empty(); 6008 } 6009 6010 // We might have added oops to ClassLoaderData::_handles during the 6011 // concurrent marking phase. These oops point to newly allocated objects 6012 // that are guaranteed to be kept alive. Either by the direct allocation 6013 // code, or when the young collector processes the strong roots. Hence, 6014 // we don't have to revisit the _handles block during the remark phase. 6015 6016 verify_work_stacks_empty(); 6017 // Restore evacuated mark words, if any, used for overflow list links 6018 if (!CMSOverflowEarlyRestoration) { 6019 restore_preserved_marks_if_any(); 6020 } 6021 verify_overflow_empty(); 6022 } 6023 6024 //////////////////////////////////////////////////////// 6025 // Parallel Reference Processing Task Proxy Class 6026 //////////////////////////////////////////////////////// 6027 class CMSRefProcTaskProxy: public AbstractGangTaskWOopQueues { 6028 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask; 6029 CMSCollector* _collector; 6030 CMSBitMap* _mark_bit_map; 6031 const MemRegion _span; 6032 ProcessTask& _task; 6033 6034 public: 6035 CMSRefProcTaskProxy(ProcessTask& task, 6036 CMSCollector* collector, 6037 const MemRegion& span, 6038 CMSBitMap* mark_bit_map, 6039 AbstractWorkGang* workers, 6040 OopTaskQueueSet* task_queues): 6041 // XXX Should superclass AGTWOQ also know about AWG since it knows 6042 // about the task_queues used by the AWG? Then it could initialize 6043 // the terminator() object. See 6984287. The set_for_termination() 6044 // below is a temporary band-aid for the regression in 6984287. 6045 AbstractGangTaskWOopQueues("Process referents by policy in parallel", 6046 task_queues), 6047 _task(task), 6048 _collector(collector), _span(span), _mark_bit_map(mark_bit_map) 6049 { 6050 assert(_collector->_span.equals(_span) && !_span.is_empty(), 6051 "Inconsistency in _span"); 6052 set_for_termination(workers->active_workers()); 6053 } 6054 6055 OopTaskQueueSet* task_queues() { return queues(); } 6056 6057 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 6058 6059 void do_work_steal(int i, 6060 CMSParDrainMarkingStackClosure* drain, 6061 CMSParKeepAliveClosure* keep_alive, 6062 int* seed); 6063 6064 virtual void work(uint worker_id); 6065 }; 6066 6067 void CMSRefProcTaskProxy::work(uint worker_id) { 6068 assert(_collector->_span.equals(_span), "Inconsistency in _span"); 6069 CMSParKeepAliveClosure par_keep_alive(_collector, _span, 6070 _mark_bit_map, 6071 work_queue(worker_id)); 6072 CMSParDrainMarkingStackClosure par_drain_stack(_collector, _span, 6073 _mark_bit_map, 6074 work_queue(worker_id)); 6075 CMSIsAliveClosure is_alive_closure(_span, _mark_bit_map); 6076 _task.work(worker_id, is_alive_closure, par_keep_alive, par_drain_stack); 6077 if (_task.marks_oops_alive()) { 6078 do_work_steal(worker_id, &par_drain_stack, &par_keep_alive, 6079 _collector->hash_seed(worker_id)); 6080 } 6081 assert(work_queue(worker_id)->size() == 0, "work_queue should be empty"); 6082 assert(_collector->_overflow_list == NULL, "non-empty _overflow_list"); 6083 } 6084 6085 class CMSRefEnqueueTaskProxy: public AbstractGangTask { 6086 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask; 6087 EnqueueTask& _task; 6088 6089 public: 6090 CMSRefEnqueueTaskProxy(EnqueueTask& task) 6091 : AbstractGangTask("Enqueue reference objects in parallel"), 6092 _task(task) 6093 { } 6094 6095 virtual void work(uint worker_id) 6096 { 6097 _task.work(worker_id); 6098 } 6099 }; 6100 6101 CMSParKeepAliveClosure::CMSParKeepAliveClosure(CMSCollector* collector, 6102 MemRegion span, CMSBitMap* bit_map, OopTaskQueue* work_queue): 6103 _span(span), 6104 _bit_map(bit_map), 6105 _work_queue(work_queue), 6106 _mark_and_push(collector, span, bit_map, work_queue), 6107 _low_water_mark(MIN2((uint)(work_queue->max_elems()/4), 6108 (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads))) 6109 { } 6110 6111 // . see if we can share work_queues with ParNew? XXX 6112 void CMSRefProcTaskProxy::do_work_steal(int i, 6113 CMSParDrainMarkingStackClosure* drain, 6114 CMSParKeepAliveClosure* keep_alive, 6115 int* seed) { 6116 OopTaskQueue* work_q = work_queue(i); 6117 NOT_PRODUCT(int num_steals = 0;) 6118 oop obj_to_scan; 6119 6120 while (true) { 6121 // Completely finish any left over work from (an) earlier round(s) 6122 drain->trim_queue(0); 6123 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 6124 (size_t)ParGCDesiredObjsFromOverflowList); 6125 // Now check if there's any work in the overflow list 6126 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads, 6127 // only affects the number of attempts made to get work from the 6128 // overflow list and does not affect the number of workers. Just 6129 // pass ParallelGCThreads so this behavior is unchanged. 6130 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 6131 work_q, 6132 ParallelGCThreads)) { 6133 // Found something in global overflow list; 6134 // not yet ready to go stealing work from others. 6135 // We'd like to assert(work_q->size() != 0, ...) 6136 // because we just took work from the overflow list, 6137 // but of course we can't, since all of that might have 6138 // been already stolen from us. 6139 continue; 6140 } 6141 // Verify that we have no work before we resort to stealing 6142 assert(work_q->size() == 0, "Have work, shouldn't steal"); 6143 // Try to steal from other queues that have work 6144 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 6145 NOT_PRODUCT(num_steals++;) 6146 assert(obj_to_scan->is_oop(), "Oops, not an oop!"); 6147 assert(_mark_bit_map->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 6148 // Do scanning work 6149 obj_to_scan->oop_iterate(keep_alive); 6150 // Loop around, finish this work, and try to steal some more 6151 } else if (terminator()->offer_termination()) { 6152 break; // nirvana from the infinite cycle 6153 } 6154 } 6155 NOT_PRODUCT( 6156 if (PrintCMSStatistics != 0) { 6157 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals); 6158 } 6159 ) 6160 } 6161 6162 void CMSRefProcTaskExecutor::execute(ProcessTask& task) 6163 { 6164 GenCollectedHeap* gch = GenCollectedHeap::heap(); 6165 FlexibleWorkGang* workers = gch->workers(); 6166 assert(workers != NULL, "Need parallel worker threads."); 6167 CMSRefProcTaskProxy rp_task(task, &_collector, 6168 _collector.ref_processor()->span(), 6169 _collector.markBitMap(), 6170 workers, _collector.task_queues()); 6171 workers->run_task(&rp_task); 6172 } 6173 6174 void CMSRefProcTaskExecutor::execute(EnqueueTask& task) 6175 { 6176 6177 GenCollectedHeap* gch = GenCollectedHeap::heap(); 6178 FlexibleWorkGang* workers = gch->workers(); 6179 assert(workers != NULL, "Need parallel worker threads."); 6180 CMSRefEnqueueTaskProxy enq_task(task); 6181 workers->run_task(&enq_task); 6182 } 6183 6184 void CMSCollector::refProcessingWork(bool asynch, bool clear_all_soft_refs) { 6185 6186 ResourceMark rm; 6187 HandleMark hm; 6188 6189 ReferenceProcessor* rp = ref_processor(); 6190 assert(rp->span().equals(_span), "Spans should be equal"); 6191 assert(!rp->enqueuing_is_done(), "Enqueuing should not be complete"); 6192 // Process weak references. 6193 rp->setup_policy(clear_all_soft_refs); 6194 verify_work_stacks_empty(); 6195 6196 CMSKeepAliveClosure cmsKeepAliveClosure(this, _span, &_markBitMap, 6197 &_markStack, false /* !preclean */); 6198 CMSDrainMarkingStackClosure cmsDrainMarkingStackClosure(this, 6199 _span, &_markBitMap, &_markStack, 6200 &cmsKeepAliveClosure, false /* !preclean */); 6201 { 6202 GCTraceTime t("weak refs processing", PrintGCDetails, false, _gc_timer_cm); 6203 6204 ReferenceProcessorStats stats; 6205 if (rp->processing_is_mt()) { 6206 // Set the degree of MT here. If the discovery is done MT, there 6207 // may have been a different number of threads doing the discovery 6208 // and a different number of discovered lists may have Ref objects. 6209 // That is OK as long as the Reference lists are balanced (see 6210 // balance_all_queues() and balance_queues()). 6211 GenCollectedHeap* gch = GenCollectedHeap::heap(); 6212 int active_workers = ParallelGCThreads; 6213 FlexibleWorkGang* workers = gch->workers(); 6214 if (workers != NULL) { 6215 active_workers = workers->active_workers(); 6216 // The expectation is that active_workers will have already 6217 // been set to a reasonable value. If it has not been set, 6218 // investigate. 6219 assert(active_workers > 0, "Should have been set during scavenge"); 6220 } 6221 rp->set_active_mt_degree(active_workers); 6222 CMSRefProcTaskExecutor task_executor(*this); 6223 stats = rp->process_discovered_references(&_is_alive_closure, 6224 &cmsKeepAliveClosure, 6225 &cmsDrainMarkingStackClosure, 6226 &task_executor, 6227 _gc_timer_cm); 6228 } else { 6229 stats = rp->process_discovered_references(&_is_alive_closure, 6230 &cmsKeepAliveClosure, 6231 &cmsDrainMarkingStackClosure, 6232 NULL, 6233 _gc_timer_cm); 6234 } 6235 _gc_tracer_cm->report_gc_reference_stats(stats); 6236 6237 } 6238 6239 // This is the point where the entire marking should have completed. 6240 verify_work_stacks_empty(); 6241 6242 if (should_unload_classes()) { 6243 { 6244 GCTraceTime t("class unloading", PrintGCDetails, false, _gc_timer_cm); 6245 6246 // Unload classes and purge the SystemDictionary. 6247 bool purged_class = SystemDictionary::do_unloading(&_is_alive_closure); 6248 6249 // Unload nmethods. 6250 CodeCache::do_unloading(&_is_alive_closure, purged_class); 6251 6252 // Prune dead klasses from subklass/sibling/implementor lists. 6253 Klass::clean_weak_klass_links(&_is_alive_closure); 6254 } 6255 6256 { 6257 GCTraceTime t("scrub symbol table", PrintGCDetails, false, _gc_timer_cm); 6258 // Clean up unreferenced symbols in symbol table. 6259 SymbolTable::unlink(); 6260 } 6261 } 6262 6263 // CMS doesn't use the StringTable as hard roots when class unloading is turned off. 6264 // Need to check if we really scanned the StringTable. 6265 if ((roots_scanning_options() & SharedHeap::SO_Strings) == 0) { 6266 GCTraceTime t("scrub string table", PrintGCDetails, false, _gc_timer_cm); 6267 // Delete entries for dead interned strings. 6268 StringTable::unlink(&_is_alive_closure); 6269 } 6270 6271 // Restore any preserved marks as a result of mark stack or 6272 // work queue overflow 6273 restore_preserved_marks_if_any(); // done single-threaded for now 6274 6275 rp->set_enqueuing_is_done(true); 6276 if (rp->processing_is_mt()) { 6277 rp->balance_all_queues(); 6278 CMSRefProcTaskExecutor task_executor(*this); 6279 rp->enqueue_discovered_references(&task_executor); 6280 } else { 6281 rp->enqueue_discovered_references(NULL); 6282 } 6283 rp->verify_no_references_recorded(); 6284 assert(!rp->discovery_enabled(), "should have been disabled"); 6285 } 6286 6287 #ifndef PRODUCT 6288 void CMSCollector::check_correct_thread_executing() { 6289 Thread* t = Thread::current(); 6290 // Only the VM thread or the CMS thread should be here. 6291 assert(t->is_ConcurrentGC_thread() || t->is_VM_thread(), 6292 "Unexpected thread type"); 6293 // If this is the vm thread, the foreground process 6294 // should not be waiting. Note that _foregroundGCIsActive is 6295 // true while the foreground collector is waiting. 6296 if (_foregroundGCShouldWait) { 6297 // We cannot be the VM thread 6298 assert(t->is_ConcurrentGC_thread(), 6299 "Should be CMS thread"); 6300 } else { 6301 // We can be the CMS thread only if we are in a stop-world 6302 // phase of CMS collection. 6303 if (t->is_ConcurrentGC_thread()) { 6304 assert(_collectorState == InitialMarking || 6305 _collectorState == FinalMarking, 6306 "Should be a stop-world phase"); 6307 // The CMS thread should be holding the CMS_token. 6308 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6309 "Potential interference with concurrently " 6310 "executing VM thread"); 6311 } 6312 } 6313 } 6314 #endif 6315 6316 void CMSCollector::sweep(bool asynch) { 6317 assert(_collectorState == Sweeping, "just checking"); 6318 check_correct_thread_executing(); 6319 verify_work_stacks_empty(); 6320 verify_overflow_empty(); 6321 increment_sweep_count(); 6322 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 6323 6324 _inter_sweep_timer.stop(); 6325 _inter_sweep_estimate.sample(_inter_sweep_timer.seconds()); 6326 size_policy()->avg_cms_free_at_sweep()->sample(_cmsGen->free()); 6327 6328 assert(!_intra_sweep_timer.is_active(), "Should not be active"); 6329 _intra_sweep_timer.reset(); 6330 _intra_sweep_timer.start(); 6331 if (asynch) { 6332 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 6333 CMSPhaseAccounting pa(this, "sweep", !PrintGCDetails); 6334 // First sweep the old gen 6335 { 6336 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(), 6337 bitMapLock()); 6338 sweepWork(_cmsGen, asynch); 6339 } 6340 6341 // Update Universe::_heap_*_at_gc figures. 6342 // We need all the free list locks to make the abstract state 6343 // transition from Sweeping to Resetting. See detailed note 6344 // further below. 6345 { 6346 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock()); 6347 // Update heap occupancy information which is used as 6348 // input to soft ref clearing policy at the next gc. 6349 Universe::update_heap_info_at_gc(); 6350 _collectorState = Resizing; 6351 } 6352 } else { 6353 // already have needed locks 6354 sweepWork(_cmsGen, asynch); 6355 // Update heap occupancy information which is used as 6356 // input to soft ref clearing policy at the next gc. 6357 Universe::update_heap_info_at_gc(); 6358 _collectorState = Resizing; 6359 } 6360 verify_work_stacks_empty(); 6361 verify_overflow_empty(); 6362 6363 if (should_unload_classes()) { 6364 ClassLoaderDataGraph::purge(); 6365 } 6366 6367 _intra_sweep_timer.stop(); 6368 _intra_sweep_estimate.sample(_intra_sweep_timer.seconds()); 6369 6370 _inter_sweep_timer.reset(); 6371 _inter_sweep_timer.start(); 6372 6373 // We need to use a monotonically non-decreasing time in ms 6374 // or we will see time-warp warnings and os::javaTimeMillis() 6375 // does not guarantee monotonicity. 6376 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC; 6377 update_time_of_last_gc(now); 6378 6379 // NOTE on abstract state transitions: 6380 // Mutators allocate-live and/or mark the mod-union table dirty 6381 // based on the state of the collection. The former is done in 6382 // the interval [Marking, Sweeping] and the latter in the interval 6383 // [Marking, Sweeping). Thus the transitions into the Marking state 6384 // and out of the Sweeping state must be synchronously visible 6385 // globally to the mutators. 6386 // The transition into the Marking state happens with the world 6387 // stopped so the mutators will globally see it. Sweeping is 6388 // done asynchronously by the background collector so the transition 6389 // from the Sweeping state to the Resizing state must be done 6390 // under the freelistLock (as is the check for whether to 6391 // allocate-live and whether to dirty the mod-union table). 6392 assert(_collectorState == Resizing, "Change of collector state to" 6393 " Resizing must be done under the freelistLocks (plural)"); 6394 6395 // Now that sweeping has been completed, we clear 6396 // the incremental_collection_failed flag, 6397 // thus inviting a younger gen collection to promote into 6398 // this generation. If such a promotion may still fail, 6399 // the flag will be set again when a young collection is 6400 // attempted. 6401 GenCollectedHeap* gch = GenCollectedHeap::heap(); 6402 gch->clear_incremental_collection_failed(); // Worth retrying as fresh space may have been freed up 6403 gch->update_full_collections_completed(_collection_count_start); 6404 } 6405 6406 // FIX ME!!! Looks like this belongs in CFLSpace, with 6407 // CMSGen merely delegating to it. 6408 void ConcurrentMarkSweepGeneration::setNearLargestChunk() { 6409 double nearLargestPercent = FLSLargestBlockCoalesceProximity; 6410 HeapWord* minAddr = _cmsSpace->bottom(); 6411 HeapWord* largestAddr = 6412 (HeapWord*) _cmsSpace->dictionary()->find_largest_dict(); 6413 if (largestAddr == NULL) { 6414 // The dictionary appears to be empty. In this case 6415 // try to coalesce at the end of the heap. 6416 largestAddr = _cmsSpace->end(); 6417 } 6418 size_t largestOffset = pointer_delta(largestAddr, minAddr); 6419 size_t nearLargestOffset = 6420 (size_t)((double)largestOffset * nearLargestPercent) - MinChunkSize; 6421 if (PrintFLSStatistics != 0) { 6422 gclog_or_tty->print_cr( 6423 "CMS: Large Block: " PTR_FORMAT ";" 6424 " Proximity: " PTR_FORMAT " -> " PTR_FORMAT, 6425 largestAddr, 6426 _cmsSpace->nearLargestChunk(), minAddr + nearLargestOffset); 6427 } 6428 _cmsSpace->set_nearLargestChunk(minAddr + nearLargestOffset); 6429 } 6430 6431 bool ConcurrentMarkSweepGeneration::isNearLargestChunk(HeapWord* addr) { 6432 return addr >= _cmsSpace->nearLargestChunk(); 6433 } 6434 6435 FreeChunk* ConcurrentMarkSweepGeneration::find_chunk_at_end() { 6436 return _cmsSpace->find_chunk_at_end(); 6437 } 6438 6439 void ConcurrentMarkSweepGeneration::update_gc_stats(int current_level, 6440 bool full) { 6441 // The next lower level has been collected. Gather any statistics 6442 // that are of interest at this point. 6443 if (!full && (current_level + 1) == level()) { 6444 // Gather statistics on the young generation collection. 6445 collector()->stats().record_gc0_end(used()); 6446 } 6447 } 6448 6449 CMSAdaptiveSizePolicy* ConcurrentMarkSweepGeneration::size_policy() { 6450 GenCollectedHeap* gch = GenCollectedHeap::heap(); 6451 assert(gch->kind() == CollectedHeap::GenCollectedHeap, 6452 "Wrong type of heap"); 6453 CMSAdaptiveSizePolicy* sp = (CMSAdaptiveSizePolicy*) 6454 gch->gen_policy()->size_policy(); 6455 assert(sp->is_gc_cms_adaptive_size_policy(), 6456 "Wrong type of size policy"); 6457 return sp; 6458 } 6459 6460 void ConcurrentMarkSweepGeneration::rotate_debug_collection_type() { 6461 if (PrintGCDetails && Verbose) { 6462 gclog_or_tty->print("Rotate from %d ", _debug_collection_type); 6463 } 6464 _debug_collection_type = (CollectionTypes) (_debug_collection_type + 1); 6465 _debug_collection_type = 6466 (CollectionTypes) (_debug_collection_type % Unknown_collection_type); 6467 if (PrintGCDetails && Verbose) { 6468 gclog_or_tty->print_cr("to %d ", _debug_collection_type); 6469 } 6470 } 6471 6472 void CMSCollector::sweepWork(ConcurrentMarkSweepGeneration* gen, 6473 bool asynch) { 6474 // We iterate over the space(s) underlying this generation, 6475 // checking the mark bit map to see if the bits corresponding 6476 // to specific blocks are marked or not. Blocks that are 6477 // marked are live and are not swept up. All remaining blocks 6478 // are swept up, with coalescing on-the-fly as we sweep up 6479 // contiguous free and/or garbage blocks: 6480 // We need to ensure that the sweeper synchronizes with allocators 6481 // and stop-the-world collectors. In particular, the following 6482 // locks are used: 6483 // . CMS token: if this is held, a stop the world collection cannot occur 6484 // . freelistLock: if this is held no allocation can occur from this 6485 // generation by another thread 6486 // . bitMapLock: if this is held, no other thread can access or update 6487 // 6488 6489 // Note that we need to hold the freelistLock if we use 6490 // block iterate below; else the iterator might go awry if 6491 // a mutator (or promotion) causes block contents to change 6492 // (for instance if the allocator divvies up a block). 6493 // If we hold the free list lock, for all practical purposes 6494 // young generation GC's can't occur (they'll usually need to 6495 // promote), so we might as well prevent all young generation 6496 // GC's while we do a sweeping step. For the same reason, we might 6497 // as well take the bit map lock for the entire duration 6498 6499 // check that we hold the requisite locks 6500 assert(have_cms_token(), "Should hold cms token"); 6501 assert( (asynch && ConcurrentMarkSweepThread::cms_thread_has_cms_token()) 6502 || (!asynch && ConcurrentMarkSweepThread::vm_thread_has_cms_token()), 6503 "Should possess CMS token to sweep"); 6504 assert_lock_strong(gen->freelistLock()); 6505 assert_lock_strong(bitMapLock()); 6506 6507 assert(!_inter_sweep_timer.is_active(), "Was switched off in an outer context"); 6508 assert(_intra_sweep_timer.is_active(), "Was switched on in an outer context"); 6509 gen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()), 6510 _inter_sweep_estimate.padded_average(), 6511 _intra_sweep_estimate.padded_average()); 6512 gen->setNearLargestChunk(); 6513 6514 { 6515 SweepClosure sweepClosure(this, gen, &_markBitMap, 6516 CMSYield && asynch); 6517 gen->cmsSpace()->blk_iterate_careful(&sweepClosure); 6518 // We need to free-up/coalesce garbage/blocks from a 6519 // co-terminal free run. This is done in the SweepClosure 6520 // destructor; so, do not remove this scope, else the 6521 // end-of-sweep-census below will be off by a little bit. 6522 } 6523 gen->cmsSpace()->sweep_completed(); 6524 gen->cmsSpace()->endSweepFLCensus(sweep_count()); 6525 if (should_unload_classes()) { // unloaded classes this cycle, 6526 _concurrent_cycles_since_last_unload = 0; // ... reset count 6527 } else { // did not unload classes, 6528 _concurrent_cycles_since_last_unload++; // ... increment count 6529 } 6530 } 6531 6532 // Reset CMS data structures (for now just the marking bit map) 6533 // preparatory for the next cycle. 6534 void CMSCollector::reset(bool asynch) { 6535 GenCollectedHeap* gch = GenCollectedHeap::heap(); 6536 CMSAdaptiveSizePolicy* sp = size_policy(); 6537 AdaptiveSizePolicyOutput(sp, gch->total_collections()); 6538 if (asynch) { 6539 CMSTokenSyncWithLocks ts(true, bitMapLock()); 6540 6541 // If the state is not "Resetting", the foreground thread 6542 // has done a collection and the resetting. 6543 if (_collectorState != Resetting) { 6544 assert(_collectorState == Idling, "The state should only change" 6545 " because the foreground collector has finished the collection"); 6546 return; 6547 } 6548 6549 // Clear the mark bitmap (no grey objects to start with) 6550 // for the next cycle. 6551 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 6552 CMSPhaseAccounting cmspa(this, "reset", !PrintGCDetails); 6553 6554 HeapWord* curAddr = _markBitMap.startWord(); 6555 while (curAddr < _markBitMap.endWord()) { 6556 size_t remaining = pointer_delta(_markBitMap.endWord(), curAddr); 6557 MemRegion chunk(curAddr, MIN2(CMSBitMapYieldQuantum, remaining)); 6558 _markBitMap.clear_large_range(chunk); 6559 if (ConcurrentMarkSweepThread::should_yield() && 6560 !foregroundGCIsActive() && 6561 CMSYield) { 6562 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6563 "CMS thread should hold CMS token"); 6564 assert_lock_strong(bitMapLock()); 6565 bitMapLock()->unlock(); 6566 ConcurrentMarkSweepThread::desynchronize(true); 6567 ConcurrentMarkSweepThread::acknowledge_yield_request(); 6568 stopTimer(); 6569 if (PrintCMSStatistics != 0) { 6570 incrementYields(); 6571 } 6572 icms_wait(); 6573 6574 // See the comment in coordinator_yield() 6575 for (unsigned i = 0; i < CMSYieldSleepCount && 6576 ConcurrentMarkSweepThread::should_yield() && 6577 !CMSCollector::foregroundGCIsActive(); ++i) { 6578 os::sleep(Thread::current(), 1, false); 6579 ConcurrentMarkSweepThread::acknowledge_yield_request(); 6580 } 6581 6582 ConcurrentMarkSweepThread::synchronize(true); 6583 bitMapLock()->lock_without_safepoint_check(); 6584 startTimer(); 6585 } 6586 curAddr = chunk.end(); 6587 } 6588 // A successful mostly concurrent collection has been done. 6589 // Because only the full (i.e., concurrent mode failure) collections 6590 // are being measured for gc overhead limits, clean the "near" flag 6591 // and count. 6592 sp->reset_gc_overhead_limit_count(); 6593 _collectorState = Idling; 6594 } else { 6595 // already have the lock 6596 assert(_collectorState == Resetting, "just checking"); 6597 assert_lock_strong(bitMapLock()); 6598 _markBitMap.clear_all(); 6599 _collectorState = Idling; 6600 } 6601 6602 // Stop incremental mode after a cycle completes, so that any future cycles 6603 // are triggered by allocation. 6604 stop_icms(); 6605 6606 NOT_PRODUCT( 6607 if (RotateCMSCollectionTypes) { 6608 _cmsGen->rotate_debug_collection_type(); 6609 } 6610 ) 6611 6612 register_gc_end(); 6613 } 6614 6615 void CMSCollector::do_CMS_operation(CMS_op_type op, GCCause::Cause gc_cause) { 6616 gclog_or_tty->date_stamp(PrintGC && PrintGCDateStamps); 6617 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 6618 GCTraceTime t(GCCauseString("GC", gc_cause), PrintGC, !PrintGCDetails, NULL); 6619 TraceCollectorStats tcs(counters()); 6620 6621 switch (op) { 6622 case CMS_op_checkpointRootsInitial: { 6623 SvcGCMarker sgcm(SvcGCMarker::OTHER); 6624 checkpointRootsInitial(true); // asynch 6625 if (PrintGC) { 6626 _cmsGen->printOccupancy("initial-mark"); 6627 } 6628 break; 6629 } 6630 case CMS_op_checkpointRootsFinal: { 6631 SvcGCMarker sgcm(SvcGCMarker::OTHER); 6632 checkpointRootsFinal(true, // asynch 6633 false, // !clear_all_soft_refs 6634 false); // !init_mark_was_synchronous 6635 if (PrintGC) { 6636 _cmsGen->printOccupancy("remark"); 6637 } 6638 break; 6639 } 6640 default: 6641 fatal("No such CMS_op"); 6642 } 6643 } 6644 6645 #ifndef PRODUCT 6646 size_t const CMSCollector::skip_header_HeapWords() { 6647 return FreeChunk::header_size(); 6648 } 6649 6650 // Try and collect here conditions that should hold when 6651 // CMS thread is exiting. The idea is that the foreground GC 6652 // thread should not be blocked if it wants to terminate 6653 // the CMS thread and yet continue to run the VM for a while 6654 // after that. 6655 void CMSCollector::verify_ok_to_terminate() const { 6656 assert(Thread::current()->is_ConcurrentGC_thread(), 6657 "should be called by CMS thread"); 6658 assert(!_foregroundGCShouldWait, "should be false"); 6659 // We could check here that all the various low-level locks 6660 // are not held by the CMS thread, but that is overkill; see 6661 // also CMSThread::verify_ok_to_terminate() where the CGC_lock 6662 // is checked. 6663 } 6664 #endif 6665 6666 size_t CMSCollector::block_size_using_printezis_bits(HeapWord* addr) const { 6667 assert(_markBitMap.isMarked(addr) && _markBitMap.isMarked(addr + 1), 6668 "missing Printezis mark?"); 6669 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 6670 size_t size = pointer_delta(nextOneAddr + 1, addr); 6671 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 6672 "alignment problem"); 6673 assert(size >= 3, "Necessary for Printezis marks to work"); 6674 return size; 6675 } 6676 6677 // A variant of the above (block_size_using_printezis_bits()) except 6678 // that we return 0 if the P-bits are not yet set. 6679 size_t CMSCollector::block_size_if_printezis_bits(HeapWord* addr) const { 6680 if (_markBitMap.isMarked(addr + 1)) { 6681 assert(_markBitMap.isMarked(addr), "P-bit can be set only for marked objects"); 6682 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 6683 size_t size = pointer_delta(nextOneAddr + 1, addr); 6684 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 6685 "alignment problem"); 6686 assert(size >= 3, "Necessary for Printezis marks to work"); 6687 return size; 6688 } 6689 return 0; 6690 } 6691 6692 HeapWord* CMSCollector::next_card_start_after_block(HeapWord* addr) const { 6693 size_t sz = 0; 6694 oop p = (oop)addr; 6695 if (p->klass_or_null() != NULL) { 6696 sz = CompactibleFreeListSpace::adjustObjectSize(p->size()); 6697 } else { 6698 sz = block_size_using_printezis_bits(addr); 6699 } 6700 assert(sz > 0, "size must be nonzero"); 6701 HeapWord* next_block = addr + sz; 6702 HeapWord* next_card = (HeapWord*)round_to((uintptr_t)next_block, 6703 CardTableModRefBS::card_size); 6704 assert(round_down((uintptr_t)addr, CardTableModRefBS::card_size) < 6705 round_down((uintptr_t)next_card, CardTableModRefBS::card_size), 6706 "must be different cards"); 6707 return next_card; 6708 } 6709 6710 6711 // CMS Bit Map Wrapper ///////////////////////////////////////// 6712 6713 // Construct a CMS bit map infrastructure, but don't create the 6714 // bit vector itself. That is done by a separate call CMSBitMap::allocate() 6715 // further below. 6716 CMSBitMap::CMSBitMap(int shifter, int mutex_rank, const char* mutex_name): 6717 _bm(), 6718 _shifter(shifter), 6719 _lock(mutex_rank >= 0 ? new Mutex(mutex_rank, mutex_name, true) : NULL) 6720 { 6721 _bmStartWord = 0; 6722 _bmWordSize = 0; 6723 } 6724 6725 bool CMSBitMap::allocate(MemRegion mr) { 6726 _bmStartWord = mr.start(); 6727 _bmWordSize = mr.word_size(); 6728 ReservedSpace brs(ReservedSpace::allocation_align_size_up( 6729 (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1)); 6730 if (!brs.is_reserved()) { 6731 warning("CMS bit map allocation failure"); 6732 return false; 6733 } 6734 // For now we'll just commit all of the bit map up front. 6735 // Later on we'll try to be more parsimonious with swap. 6736 if (!_virtual_space.initialize(brs, brs.size())) { 6737 warning("CMS bit map backing store failure"); 6738 return false; 6739 } 6740 assert(_virtual_space.committed_size() == brs.size(), 6741 "didn't reserve backing store for all of CMS bit map?"); 6742 _bm.set_map((BitMap::bm_word_t*)_virtual_space.low()); 6743 assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >= 6744 _bmWordSize, "inconsistency in bit map sizing"); 6745 _bm.set_size(_bmWordSize >> _shifter); 6746 6747 // bm.clear(); // can we rely on getting zero'd memory? verify below 6748 assert(isAllClear(), 6749 "Expected zero'd memory from ReservedSpace constructor"); 6750 assert(_bm.size() == heapWordDiffToOffsetDiff(sizeInWords()), 6751 "consistency check"); 6752 return true; 6753 } 6754 6755 void CMSBitMap::dirty_range_iterate_clear(MemRegion mr, MemRegionClosure* cl) { 6756 HeapWord *next_addr, *end_addr, *last_addr; 6757 assert_locked(); 6758 assert(covers(mr), "out-of-range error"); 6759 // XXX assert that start and end are appropriately aligned 6760 for (next_addr = mr.start(), end_addr = mr.end(); 6761 next_addr < end_addr; next_addr = last_addr) { 6762 MemRegion dirty_region = getAndClearMarkedRegion(next_addr, end_addr); 6763 last_addr = dirty_region.end(); 6764 if (!dirty_region.is_empty()) { 6765 cl->do_MemRegion(dirty_region); 6766 } else { 6767 assert(last_addr == end_addr, "program logic"); 6768 return; 6769 } 6770 } 6771 } 6772 6773 void CMSBitMap::print_on_error(outputStream* st, const char* prefix) const { 6774 _bm.print_on_error(st, prefix); 6775 } 6776 6777 #ifndef PRODUCT 6778 void CMSBitMap::assert_locked() const { 6779 CMSLockVerifier::assert_locked(lock()); 6780 } 6781 6782 bool CMSBitMap::covers(MemRegion mr) const { 6783 // assert(_bm.map() == _virtual_space.low(), "map inconsistency"); 6784 assert((size_t)_bm.size() == (_bmWordSize >> _shifter), 6785 "size inconsistency"); 6786 return (mr.start() >= _bmStartWord) && 6787 (mr.end() <= endWord()); 6788 } 6789 6790 bool CMSBitMap::covers(HeapWord* start, size_t size) const { 6791 return (start >= _bmStartWord && (start + size) <= endWord()); 6792 } 6793 6794 void CMSBitMap::verifyNoOneBitsInRange(HeapWord* left, HeapWord* right) { 6795 // verify that there are no 1 bits in the interval [left, right) 6796 FalseBitMapClosure falseBitMapClosure; 6797 iterate(&falseBitMapClosure, left, right); 6798 } 6799 6800 void CMSBitMap::region_invariant(MemRegion mr) 6801 { 6802 assert_locked(); 6803 // mr = mr.intersection(MemRegion(_bmStartWord, _bmWordSize)); 6804 assert(!mr.is_empty(), "unexpected empty region"); 6805 assert(covers(mr), "mr should be covered by bit map"); 6806 // convert address range into offset range 6807 size_t start_ofs = heapWordToOffset(mr.start()); 6808 // Make sure that end() is appropriately aligned 6809 assert(mr.end() == (HeapWord*)round_to((intptr_t)mr.end(), 6810 (1 << (_shifter+LogHeapWordSize))), 6811 "Misaligned mr.end()"); 6812 size_t end_ofs = heapWordToOffset(mr.end()); 6813 assert(end_ofs > start_ofs, "Should mark at least one bit"); 6814 } 6815 6816 #endif 6817 6818 bool CMSMarkStack::allocate(size_t size) { 6819 // allocate a stack of the requisite depth 6820 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 6821 size * sizeof(oop))); 6822 if (!rs.is_reserved()) { 6823 warning("CMSMarkStack allocation failure"); 6824 return false; 6825 } 6826 if (!_virtual_space.initialize(rs, rs.size())) { 6827 warning("CMSMarkStack backing store failure"); 6828 return false; 6829 } 6830 assert(_virtual_space.committed_size() == rs.size(), 6831 "didn't reserve backing store for all of CMS stack?"); 6832 _base = (oop*)(_virtual_space.low()); 6833 _index = 0; 6834 _capacity = size; 6835 NOT_PRODUCT(_max_depth = 0); 6836 return true; 6837 } 6838 6839 // XXX FIX ME !!! In the MT case we come in here holding a 6840 // leaf lock. For printing we need to take a further lock 6841 // which has lower rank. We need to recalibrate the two 6842 // lock-ranks involved in order to be able to print the 6843 // messages below. (Or defer the printing to the caller. 6844 // For now we take the expedient path of just disabling the 6845 // messages for the problematic case.) 6846 void CMSMarkStack::expand() { 6847 assert(_capacity <= MarkStackSizeMax, "stack bigger than permitted"); 6848 if (_capacity == MarkStackSizeMax) { 6849 if (_hit_limit++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) { 6850 // We print a warning message only once per CMS cycle. 6851 gclog_or_tty->print_cr(" (benign) Hit CMSMarkStack max size limit"); 6852 } 6853 return; 6854 } 6855 // Double capacity if possible 6856 size_t new_capacity = MIN2(_capacity*2, MarkStackSizeMax); 6857 // Do not give up existing stack until we have managed to 6858 // get the double capacity that we desired. 6859 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 6860 new_capacity * sizeof(oop))); 6861 if (rs.is_reserved()) { 6862 // Release the backing store associated with old stack 6863 _virtual_space.release(); 6864 // Reinitialize virtual space for new stack 6865 if (!_virtual_space.initialize(rs, rs.size())) { 6866 fatal("Not enough swap for expanded marking stack"); 6867 } 6868 _base = (oop*)(_virtual_space.low()); 6869 _index = 0; 6870 _capacity = new_capacity; 6871 } else if (_failed_double++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) { 6872 // Failed to double capacity, continue; 6873 // we print a detail message only once per CMS cycle. 6874 gclog_or_tty->print(" (benign) Failed to expand marking stack from "SIZE_FORMAT"K to " 6875 SIZE_FORMAT"K", 6876 _capacity / K, new_capacity / K); 6877 } 6878 } 6879 6880 6881 // Closures 6882 // XXX: there seems to be a lot of code duplication here; 6883 // should refactor and consolidate common code. 6884 6885 // This closure is used to mark refs into the CMS generation in 6886 // the CMS bit map. Called at the first checkpoint. This closure 6887 // assumes that we do not need to re-mark dirty cards; if the CMS 6888 // generation on which this is used is not an oldest 6889 // generation then this will lose younger_gen cards! 6890 6891 MarkRefsIntoClosure::MarkRefsIntoClosure( 6892 MemRegion span, CMSBitMap* bitMap): 6893 _span(span), 6894 _bitMap(bitMap) 6895 { 6896 assert(_ref_processor == NULL, "deliberately left NULL"); 6897 assert(_bitMap->covers(_span), "_bitMap/_span mismatch"); 6898 } 6899 6900 void MarkRefsIntoClosure::do_oop(oop obj) { 6901 // if p points into _span, then mark corresponding bit in _markBitMap 6902 assert(obj->is_oop(), "expected an oop"); 6903 HeapWord* addr = (HeapWord*)obj; 6904 if (_span.contains(addr)) { 6905 // this should be made more efficient 6906 _bitMap->mark(addr); 6907 } 6908 } 6909 6910 void MarkRefsIntoClosure::do_oop(oop* p) { MarkRefsIntoClosure::do_oop_work(p); } 6911 void MarkRefsIntoClosure::do_oop(narrowOop* p) { MarkRefsIntoClosure::do_oop_work(p); } 6912 6913 Par_MarkRefsIntoClosure::Par_MarkRefsIntoClosure( 6914 MemRegion span, CMSBitMap* bitMap): 6915 _span(span), 6916 _bitMap(bitMap) 6917 { 6918 assert(_ref_processor == NULL, "deliberately left NULL"); 6919 assert(_bitMap->covers(_span), "_bitMap/_span mismatch"); 6920 } 6921 6922 void Par_MarkRefsIntoClosure::do_oop(oop obj) { 6923 // if p points into _span, then mark corresponding bit in _markBitMap 6924 assert(obj->is_oop(), "expected an oop"); 6925 HeapWord* addr = (HeapWord*)obj; 6926 if (_span.contains(addr)) { 6927 // this should be made more efficient 6928 _bitMap->par_mark(addr); 6929 } 6930 } 6931 6932 void Par_MarkRefsIntoClosure::do_oop(oop* p) { Par_MarkRefsIntoClosure::do_oop_work(p); } 6933 void Par_MarkRefsIntoClosure::do_oop(narrowOop* p) { Par_MarkRefsIntoClosure::do_oop_work(p); } 6934 6935 // A variant of the above, used for CMS marking verification. 6936 MarkRefsIntoVerifyClosure::MarkRefsIntoVerifyClosure( 6937 MemRegion span, CMSBitMap* verification_bm, CMSBitMap* cms_bm): 6938 _span(span), 6939 _verification_bm(verification_bm), 6940 _cms_bm(cms_bm) 6941 { 6942 assert(_ref_processor == NULL, "deliberately left NULL"); 6943 assert(_verification_bm->covers(_span), "_verification_bm/_span mismatch"); 6944 } 6945 6946 void MarkRefsIntoVerifyClosure::do_oop(oop obj) { 6947 // if p points into _span, then mark corresponding bit in _markBitMap 6948 assert(obj->is_oop(), "expected an oop"); 6949 HeapWord* addr = (HeapWord*)obj; 6950 if (_span.contains(addr)) { 6951 _verification_bm->mark(addr); 6952 if (!_cms_bm->isMarked(addr)) { 6953 oop(addr)->print(); 6954 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)", addr); 6955 fatal("... aborting"); 6956 } 6957 } 6958 } 6959 6960 void MarkRefsIntoVerifyClosure::do_oop(oop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); } 6961 void MarkRefsIntoVerifyClosure::do_oop(narrowOop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); } 6962 6963 ////////////////////////////////////////////////// 6964 // MarkRefsIntoAndScanClosure 6965 ////////////////////////////////////////////////// 6966 6967 MarkRefsIntoAndScanClosure::MarkRefsIntoAndScanClosure(MemRegion span, 6968 ReferenceProcessor* rp, 6969 CMSBitMap* bit_map, 6970 CMSBitMap* mod_union_table, 6971 CMSMarkStack* mark_stack, 6972 CMSCollector* collector, 6973 bool should_yield, 6974 bool concurrent_precleaning): 6975 _collector(collector), 6976 _span(span), 6977 _bit_map(bit_map), 6978 _mark_stack(mark_stack), 6979 _pushAndMarkClosure(collector, span, rp, bit_map, mod_union_table, 6980 mark_stack, concurrent_precleaning), 6981 _yield(should_yield), 6982 _concurrent_precleaning(concurrent_precleaning), 6983 _freelistLock(NULL) 6984 { 6985 _ref_processor = rp; 6986 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 6987 } 6988 6989 // This closure is used to mark refs into the CMS generation at the 6990 // second (final) checkpoint, and to scan and transitively follow 6991 // the unmarked oops. It is also used during the concurrent precleaning 6992 // phase while scanning objects on dirty cards in the CMS generation. 6993 // The marks are made in the marking bit map and the marking stack is 6994 // used for keeping the (newly) grey objects during the scan. 6995 // The parallel version (Par_...) appears further below. 6996 void MarkRefsIntoAndScanClosure::do_oop(oop obj) { 6997 if (obj != NULL) { 6998 assert(obj->is_oop(), "expected an oop"); 6999 HeapWord* addr = (HeapWord*)obj; 7000 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 7001 assert(_collector->overflow_list_is_empty(), 7002 "overflow list should be empty"); 7003 if (_span.contains(addr) && 7004 !_bit_map->isMarked(addr)) { 7005 // mark bit map (object is now grey) 7006 _bit_map->mark(addr); 7007 // push on marking stack (stack should be empty), and drain the 7008 // stack by applying this closure to the oops in the oops popped 7009 // from the stack (i.e. blacken the grey objects) 7010 bool res = _mark_stack->push(obj); 7011 assert(res, "Should have space to push on empty stack"); 7012 do { 7013 oop new_oop = _mark_stack->pop(); 7014 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 7015 assert(_bit_map->isMarked((HeapWord*)new_oop), 7016 "only grey objects on this stack"); 7017 // iterate over the oops in this oop, marking and pushing 7018 // the ones in CMS heap (i.e. in _span). 7019 new_oop->oop_iterate(&_pushAndMarkClosure); 7020 // check if it's time to yield 7021 do_yield_check(); 7022 } while (!_mark_stack->isEmpty() || 7023 (!_concurrent_precleaning && take_from_overflow_list())); 7024 // if marking stack is empty, and we are not doing this 7025 // during precleaning, then check the overflow list 7026 } 7027 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 7028 assert(_collector->overflow_list_is_empty(), 7029 "overflow list was drained above"); 7030 // We could restore evacuated mark words, if any, used for 7031 // overflow list links here because the overflow list is 7032 // provably empty here. That would reduce the maximum 7033 // size requirements for preserved_{oop,mark}_stack. 7034 // But we'll just postpone it until we are all done 7035 // so we can just stream through. 7036 if (!_concurrent_precleaning && CMSOverflowEarlyRestoration) { 7037 _collector->restore_preserved_marks_if_any(); 7038 assert(_collector->no_preserved_marks(), "No preserved marks"); 7039 } 7040 assert(!CMSOverflowEarlyRestoration || _collector->no_preserved_marks(), 7041 "All preserved marks should have been restored above"); 7042 } 7043 } 7044 7045 void MarkRefsIntoAndScanClosure::do_oop(oop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); } 7046 void MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); } 7047 7048 void MarkRefsIntoAndScanClosure::do_yield_work() { 7049 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 7050 "CMS thread should hold CMS token"); 7051 assert_lock_strong(_freelistLock); 7052 assert_lock_strong(_bit_map->lock()); 7053 // relinquish the free_list_lock and bitMaplock() 7054 _bit_map->lock()->unlock(); 7055 _freelistLock->unlock(); 7056 ConcurrentMarkSweepThread::desynchronize(true); 7057 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7058 _collector->stopTimer(); 7059 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 7060 if (PrintCMSStatistics != 0) { 7061 _collector->incrementYields(); 7062 } 7063 _collector->icms_wait(); 7064 7065 // See the comment in coordinator_yield() 7066 for (unsigned i = 0; 7067 i < CMSYieldSleepCount && 7068 ConcurrentMarkSweepThread::should_yield() && 7069 !CMSCollector::foregroundGCIsActive(); 7070 ++i) { 7071 os::sleep(Thread::current(), 1, false); 7072 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7073 } 7074 7075 ConcurrentMarkSweepThread::synchronize(true); 7076 _freelistLock->lock_without_safepoint_check(); 7077 _bit_map->lock()->lock_without_safepoint_check(); 7078 _collector->startTimer(); 7079 } 7080 7081 /////////////////////////////////////////////////////////// 7082 // Par_MarkRefsIntoAndScanClosure: a parallel version of 7083 // MarkRefsIntoAndScanClosure 7084 /////////////////////////////////////////////////////////// 7085 Par_MarkRefsIntoAndScanClosure::Par_MarkRefsIntoAndScanClosure( 7086 CMSCollector* collector, MemRegion span, ReferenceProcessor* rp, 7087 CMSBitMap* bit_map, OopTaskQueue* work_queue): 7088 _span(span), 7089 _bit_map(bit_map), 7090 _work_queue(work_queue), 7091 _low_water_mark(MIN2((uint)(work_queue->max_elems()/4), 7092 (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads))), 7093 _par_pushAndMarkClosure(collector, span, rp, bit_map, work_queue) 7094 { 7095 _ref_processor = rp; 7096 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 7097 } 7098 7099 // This closure is used to mark refs into the CMS generation at the 7100 // second (final) checkpoint, and to scan and transitively follow 7101 // the unmarked oops. The marks are made in the marking bit map and 7102 // the work_queue is used for keeping the (newly) grey objects during 7103 // the scan phase whence they are also available for stealing by parallel 7104 // threads. Since the marking bit map is shared, updates are 7105 // synchronized (via CAS). 7106 void Par_MarkRefsIntoAndScanClosure::do_oop(oop obj) { 7107 if (obj != NULL) { 7108 // Ignore mark word because this could be an already marked oop 7109 // that may be chained at the end of the overflow list. 7110 assert(obj->is_oop(true), "expected an oop"); 7111 HeapWord* addr = (HeapWord*)obj; 7112 if (_span.contains(addr) && 7113 !_bit_map->isMarked(addr)) { 7114 // mark bit map (object will become grey): 7115 // It is possible for several threads to be 7116 // trying to "claim" this object concurrently; 7117 // the unique thread that succeeds in marking the 7118 // object first will do the subsequent push on 7119 // to the work queue (or overflow list). 7120 if (_bit_map->par_mark(addr)) { 7121 // push on work_queue (which may not be empty), and trim the 7122 // queue to an appropriate length by applying this closure to 7123 // the oops in the oops popped from the stack (i.e. blacken the 7124 // grey objects) 7125 bool res = _work_queue->push(obj); 7126 assert(res, "Low water mark should be less than capacity?"); 7127 trim_queue(_low_water_mark); 7128 } // Else, another thread claimed the object 7129 } 7130 } 7131 } 7132 7133 void Par_MarkRefsIntoAndScanClosure::do_oop(oop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); } 7134 void Par_MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); } 7135 7136 // This closure is used to rescan the marked objects on the dirty cards 7137 // in the mod union table and the card table proper. 7138 size_t ScanMarkedObjectsAgainCarefullyClosure::do_object_careful_m( 7139 oop p, MemRegion mr) { 7140 7141 size_t size = 0; 7142 HeapWord* addr = (HeapWord*)p; 7143 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 7144 assert(_span.contains(addr), "we are scanning the CMS generation"); 7145 // check if it's time to yield 7146 if (do_yield_check()) { 7147 // We yielded for some foreground stop-world work, 7148 // and we have been asked to abort this ongoing preclean cycle. 7149 return 0; 7150 } 7151 if (_bitMap->isMarked(addr)) { 7152 // it's marked; is it potentially uninitialized? 7153 if (p->klass_or_null() != NULL) { 7154 // an initialized object; ignore mark word in verification below 7155 // since we are running concurrent with mutators 7156 assert(p->is_oop(true), "should be an oop"); 7157 if (p->is_objArray()) { 7158 // objArrays are precisely marked; restrict scanning 7159 // to dirty cards only. 7160 size = CompactibleFreeListSpace::adjustObjectSize( 7161 p->oop_iterate(_scanningClosure, mr)); 7162 } else { 7163 // A non-array may have been imprecisely marked; we need 7164 // to scan object in its entirety. 7165 size = CompactibleFreeListSpace::adjustObjectSize( 7166 p->oop_iterate(_scanningClosure)); 7167 } 7168 #ifdef ASSERT 7169 size_t direct_size = 7170 CompactibleFreeListSpace::adjustObjectSize(p->size()); 7171 assert(size == direct_size, "Inconsistency in size"); 7172 assert(size >= 3, "Necessary for Printezis marks to work"); 7173 if (!_bitMap->isMarked(addr+1)) { 7174 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size); 7175 } else { 7176 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size-1); 7177 assert(_bitMap->isMarked(addr+size-1), 7178 "inconsistent Printezis mark"); 7179 } 7180 #endif // ASSERT 7181 } else { 7182 // An uninitialized object. 7183 assert(_bitMap->isMarked(addr+1), "missing Printezis mark?"); 7184 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 7185 size = pointer_delta(nextOneAddr + 1, addr); 7186 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 7187 "alignment problem"); 7188 // Note that pre-cleaning needn't redirty the card. OopDesc::set_klass() 7189 // will dirty the card when the klass pointer is installed in the 7190 // object (signaling the completion of initialization). 7191 } 7192 } else { 7193 // Either a not yet marked object or an uninitialized object 7194 if (p->klass_or_null() == NULL) { 7195 // An uninitialized object, skip to the next card, since 7196 // we may not be able to read its P-bits yet. 7197 assert(size == 0, "Initial value"); 7198 } else { 7199 // An object not (yet) reached by marking: we merely need to 7200 // compute its size so as to go look at the next block. 7201 assert(p->is_oop(true), "should be an oop"); 7202 size = CompactibleFreeListSpace::adjustObjectSize(p->size()); 7203 } 7204 } 7205 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 7206 return size; 7207 } 7208 7209 void ScanMarkedObjectsAgainCarefullyClosure::do_yield_work() { 7210 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 7211 "CMS thread should hold CMS token"); 7212 assert_lock_strong(_freelistLock); 7213 assert_lock_strong(_bitMap->lock()); 7214 // relinquish the free_list_lock and bitMaplock() 7215 _bitMap->lock()->unlock(); 7216 _freelistLock->unlock(); 7217 ConcurrentMarkSweepThread::desynchronize(true); 7218 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7219 _collector->stopTimer(); 7220 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 7221 if (PrintCMSStatistics != 0) { 7222 _collector->incrementYields(); 7223 } 7224 _collector->icms_wait(); 7225 7226 // See the comment in coordinator_yield() 7227 for (unsigned i = 0; i < CMSYieldSleepCount && 7228 ConcurrentMarkSweepThread::should_yield() && 7229 !CMSCollector::foregroundGCIsActive(); ++i) { 7230 os::sleep(Thread::current(), 1, false); 7231 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7232 } 7233 7234 ConcurrentMarkSweepThread::synchronize(true); 7235 _freelistLock->lock_without_safepoint_check(); 7236 _bitMap->lock()->lock_without_safepoint_check(); 7237 _collector->startTimer(); 7238 } 7239 7240 7241 ////////////////////////////////////////////////////////////////// 7242 // SurvivorSpacePrecleanClosure 7243 ////////////////////////////////////////////////////////////////// 7244 // This (single-threaded) closure is used to preclean the oops in 7245 // the survivor spaces. 7246 size_t SurvivorSpacePrecleanClosure::do_object_careful(oop p) { 7247 7248 HeapWord* addr = (HeapWord*)p; 7249 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 7250 assert(!_span.contains(addr), "we are scanning the survivor spaces"); 7251 assert(p->klass_or_null() != NULL, "object should be initialized"); 7252 // an initialized object; ignore mark word in verification below 7253 // since we are running concurrent with mutators 7254 assert(p->is_oop(true), "should be an oop"); 7255 // Note that we do not yield while we iterate over 7256 // the interior oops of p, pushing the relevant ones 7257 // on our marking stack. 7258 size_t size = p->oop_iterate(_scanning_closure); 7259 do_yield_check(); 7260 // Observe that below, we do not abandon the preclean 7261 // phase as soon as we should; rather we empty the 7262 // marking stack before returning. This is to satisfy 7263 // some existing assertions. In general, it may be a 7264 // good idea to abort immediately and complete the marking 7265 // from the grey objects at a later time. 7266 while (!_mark_stack->isEmpty()) { 7267 oop new_oop = _mark_stack->pop(); 7268 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 7269 assert(_bit_map->isMarked((HeapWord*)new_oop), 7270 "only grey objects on this stack"); 7271 // iterate over the oops in this oop, marking and pushing 7272 // the ones in CMS heap (i.e. in _span). 7273 new_oop->oop_iterate(_scanning_closure); 7274 // check if it's time to yield 7275 do_yield_check(); 7276 } 7277 unsigned int after_count = 7278 GenCollectedHeap::heap()->total_collections(); 7279 bool abort = (_before_count != after_count) || 7280 _collector->should_abort_preclean(); 7281 return abort ? 0 : size; 7282 } 7283 7284 void SurvivorSpacePrecleanClosure::do_yield_work() { 7285 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 7286 "CMS thread should hold CMS token"); 7287 assert_lock_strong(_bit_map->lock()); 7288 // Relinquish the bit map lock 7289 _bit_map->lock()->unlock(); 7290 ConcurrentMarkSweepThread::desynchronize(true); 7291 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7292 _collector->stopTimer(); 7293 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 7294 if (PrintCMSStatistics != 0) { 7295 _collector->incrementYields(); 7296 } 7297 _collector->icms_wait(); 7298 7299 // See the comment in coordinator_yield() 7300 for (unsigned i = 0; i < CMSYieldSleepCount && 7301 ConcurrentMarkSweepThread::should_yield() && 7302 !CMSCollector::foregroundGCIsActive(); ++i) { 7303 os::sleep(Thread::current(), 1, false); 7304 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7305 } 7306 7307 ConcurrentMarkSweepThread::synchronize(true); 7308 _bit_map->lock()->lock_without_safepoint_check(); 7309 _collector->startTimer(); 7310 } 7311 7312 // This closure is used to rescan the marked objects on the dirty cards 7313 // in the mod union table and the card table proper. In the parallel 7314 // case, although the bitMap is shared, we do a single read so the 7315 // isMarked() query is "safe". 7316 bool ScanMarkedObjectsAgainClosure::do_object_bm(oop p, MemRegion mr) { 7317 // Ignore mark word because we are running concurrent with mutators 7318 assert(p->is_oop_or_null(true), "expected an oop or null"); 7319 HeapWord* addr = (HeapWord*)p; 7320 assert(_span.contains(addr), "we are scanning the CMS generation"); 7321 bool is_obj_array = false; 7322 #ifdef ASSERT 7323 if (!_parallel) { 7324 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 7325 assert(_collector->overflow_list_is_empty(), 7326 "overflow list should be empty"); 7327 7328 } 7329 #endif // ASSERT 7330 if (_bit_map->isMarked(addr)) { 7331 // Obj arrays are precisely marked, non-arrays are not; 7332 // so we scan objArrays precisely and non-arrays in their 7333 // entirety. 7334 if (p->is_objArray()) { 7335 is_obj_array = true; 7336 if (_parallel) { 7337 p->oop_iterate(_par_scan_closure, mr); 7338 } else { 7339 p->oop_iterate(_scan_closure, mr); 7340 } 7341 } else { 7342 if (_parallel) { 7343 p->oop_iterate(_par_scan_closure); 7344 } else { 7345 p->oop_iterate(_scan_closure); 7346 } 7347 } 7348 } 7349 #ifdef ASSERT 7350 if (!_parallel) { 7351 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 7352 assert(_collector->overflow_list_is_empty(), 7353 "overflow list should be empty"); 7354 7355 } 7356 #endif // ASSERT 7357 return is_obj_array; 7358 } 7359 7360 MarkFromRootsClosure::MarkFromRootsClosure(CMSCollector* collector, 7361 MemRegion span, 7362 CMSBitMap* bitMap, CMSMarkStack* markStack, 7363 bool should_yield, bool verifying): 7364 _collector(collector), 7365 _span(span), 7366 _bitMap(bitMap), 7367 _mut(&collector->_modUnionTable), 7368 _markStack(markStack), 7369 _yield(should_yield), 7370 _skipBits(0) 7371 { 7372 assert(_markStack->isEmpty(), "stack should be empty"); 7373 _finger = _bitMap->startWord(); 7374 _threshold = _finger; 7375 assert(_collector->_restart_addr == NULL, "Sanity check"); 7376 assert(_span.contains(_finger), "Out of bounds _finger?"); 7377 DEBUG_ONLY(_verifying = verifying;) 7378 } 7379 7380 void MarkFromRootsClosure::reset(HeapWord* addr) { 7381 assert(_markStack->isEmpty(), "would cause duplicates on stack"); 7382 assert(_span.contains(addr), "Out of bounds _finger?"); 7383 _finger = addr; 7384 _threshold = (HeapWord*)round_to( 7385 (intptr_t)_finger, CardTableModRefBS::card_size); 7386 } 7387 7388 // Should revisit to see if this should be restructured for 7389 // greater efficiency. 7390 bool MarkFromRootsClosure::do_bit(size_t offset) { 7391 if (_skipBits > 0) { 7392 _skipBits--; 7393 return true; 7394 } 7395 // convert offset into a HeapWord* 7396 HeapWord* addr = _bitMap->startWord() + offset; 7397 assert(_bitMap->endWord() && addr < _bitMap->endWord(), 7398 "address out of range"); 7399 assert(_bitMap->isMarked(addr), "tautology"); 7400 if (_bitMap->isMarked(addr+1)) { 7401 // this is an allocated but not yet initialized object 7402 assert(_skipBits == 0, "tautology"); 7403 _skipBits = 2; // skip next two marked bits ("Printezis-marks") 7404 oop p = oop(addr); 7405 if (p->klass_or_null() == NULL) { 7406 DEBUG_ONLY(if (!_verifying) {) 7407 // We re-dirty the cards on which this object lies and increase 7408 // the _threshold so that we'll come back to scan this object 7409 // during the preclean or remark phase. (CMSCleanOnEnter) 7410 if (CMSCleanOnEnter) { 7411 size_t sz = _collector->block_size_using_printezis_bits(addr); 7412 HeapWord* end_card_addr = (HeapWord*)round_to( 7413 (intptr_t)(addr+sz), CardTableModRefBS::card_size); 7414 MemRegion redirty_range = MemRegion(addr, end_card_addr); 7415 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 7416 // Bump _threshold to end_card_addr; note that 7417 // _threshold cannot possibly exceed end_card_addr, anyhow. 7418 // This prevents future clearing of the card as the scan proceeds 7419 // to the right. 7420 assert(_threshold <= end_card_addr, 7421 "Because we are just scanning into this object"); 7422 if (_threshold < end_card_addr) { 7423 _threshold = end_card_addr; 7424 } 7425 if (p->klass_or_null() != NULL) { 7426 // Redirty the range of cards... 7427 _mut->mark_range(redirty_range); 7428 } // ...else the setting of klass will dirty the card anyway. 7429 } 7430 DEBUG_ONLY(}) 7431 return true; 7432 } 7433 } 7434 scanOopsInOop(addr); 7435 return true; 7436 } 7437 7438 // We take a break if we've been at this for a while, 7439 // so as to avoid monopolizing the locks involved. 7440 void MarkFromRootsClosure::do_yield_work() { 7441 // First give up the locks, then yield, then re-lock 7442 // We should probably use a constructor/destructor idiom to 7443 // do this unlock/lock or modify the MutexUnlocker class to 7444 // serve our purpose. XXX 7445 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 7446 "CMS thread should hold CMS token"); 7447 assert_lock_strong(_bitMap->lock()); 7448 _bitMap->lock()->unlock(); 7449 ConcurrentMarkSweepThread::desynchronize(true); 7450 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7451 _collector->stopTimer(); 7452 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 7453 if (PrintCMSStatistics != 0) { 7454 _collector->incrementYields(); 7455 } 7456 _collector->icms_wait(); 7457 7458 // See the comment in coordinator_yield() 7459 for (unsigned i = 0; i < CMSYieldSleepCount && 7460 ConcurrentMarkSweepThread::should_yield() && 7461 !CMSCollector::foregroundGCIsActive(); ++i) { 7462 os::sleep(Thread::current(), 1, false); 7463 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7464 } 7465 7466 ConcurrentMarkSweepThread::synchronize(true); 7467 _bitMap->lock()->lock_without_safepoint_check(); 7468 _collector->startTimer(); 7469 } 7470 7471 void MarkFromRootsClosure::scanOopsInOop(HeapWord* ptr) { 7472 assert(_bitMap->isMarked(ptr), "expected bit to be set"); 7473 assert(_markStack->isEmpty(), 7474 "should drain stack to limit stack usage"); 7475 // convert ptr to an oop preparatory to scanning 7476 oop obj = oop(ptr); 7477 // Ignore mark word in verification below, since we 7478 // may be running concurrent with mutators. 7479 assert(obj->is_oop(true), "should be an oop"); 7480 assert(_finger <= ptr, "_finger runneth ahead"); 7481 // advance the finger to right end of this object 7482 _finger = ptr + obj->size(); 7483 assert(_finger > ptr, "we just incremented it above"); 7484 // On large heaps, it may take us some time to get through 7485 // the marking phase (especially if running iCMS). During 7486 // this time it's possible that a lot of mutations have 7487 // accumulated in the card table and the mod union table -- 7488 // these mutation records are redundant until we have 7489 // actually traced into the corresponding card. 7490 // Here, we check whether advancing the finger would make 7491 // us cross into a new card, and if so clear corresponding 7492 // cards in the MUT (preclean them in the card-table in the 7493 // future). 7494 7495 DEBUG_ONLY(if (!_verifying) {) 7496 // The clean-on-enter optimization is disabled by default, 7497 // until we fix 6178663. 7498 if (CMSCleanOnEnter && (_finger > _threshold)) { 7499 // [_threshold, _finger) represents the interval 7500 // of cards to be cleared in MUT (or precleaned in card table). 7501 // The set of cards to be cleared is all those that overlap 7502 // with the interval [_threshold, _finger); note that 7503 // _threshold is always kept card-aligned but _finger isn't 7504 // always card-aligned. 7505 HeapWord* old_threshold = _threshold; 7506 assert(old_threshold == (HeapWord*)round_to( 7507 (intptr_t)old_threshold, CardTableModRefBS::card_size), 7508 "_threshold should always be card-aligned"); 7509 _threshold = (HeapWord*)round_to( 7510 (intptr_t)_finger, CardTableModRefBS::card_size); 7511 MemRegion mr(old_threshold, _threshold); 7512 assert(!mr.is_empty(), "Control point invariant"); 7513 assert(_span.contains(mr), "Should clear within span"); 7514 _mut->clear_range(mr); 7515 } 7516 DEBUG_ONLY(}) 7517 // Note: the finger doesn't advance while we drain 7518 // the stack below. 7519 PushOrMarkClosure pushOrMarkClosure(_collector, 7520 _span, _bitMap, _markStack, 7521 _finger, this); 7522 bool res = _markStack->push(obj); 7523 assert(res, "Empty non-zero size stack should have space for single push"); 7524 while (!_markStack->isEmpty()) { 7525 oop new_oop = _markStack->pop(); 7526 // Skip verifying header mark word below because we are 7527 // running concurrent with mutators. 7528 assert(new_oop->is_oop(true), "Oops! expected to pop an oop"); 7529 // now scan this oop's oops 7530 new_oop->oop_iterate(&pushOrMarkClosure); 7531 do_yield_check(); 7532 } 7533 assert(_markStack->isEmpty(), "tautology, emphasizing post-condition"); 7534 } 7535 7536 Par_MarkFromRootsClosure::Par_MarkFromRootsClosure(CMSConcMarkingTask* task, 7537 CMSCollector* collector, MemRegion span, 7538 CMSBitMap* bit_map, 7539 OopTaskQueue* work_queue, 7540 CMSMarkStack* overflow_stack, 7541 bool should_yield): 7542 _collector(collector), 7543 _whole_span(collector->_span), 7544 _span(span), 7545 _bit_map(bit_map), 7546 _mut(&collector->_modUnionTable), 7547 _work_queue(work_queue), 7548 _overflow_stack(overflow_stack), 7549 _yield(should_yield), 7550 _skip_bits(0), 7551 _task(task) 7552 { 7553 assert(_work_queue->size() == 0, "work_queue should be empty"); 7554 _finger = span.start(); 7555 _threshold = _finger; // XXX Defer clear-on-enter optimization for now 7556 assert(_span.contains(_finger), "Out of bounds _finger?"); 7557 } 7558 7559 // Should revisit to see if this should be restructured for 7560 // greater efficiency. 7561 bool Par_MarkFromRootsClosure::do_bit(size_t offset) { 7562 if (_skip_bits > 0) { 7563 _skip_bits--; 7564 return true; 7565 } 7566 // convert offset into a HeapWord* 7567 HeapWord* addr = _bit_map->startWord() + offset; 7568 assert(_bit_map->endWord() && addr < _bit_map->endWord(), 7569 "address out of range"); 7570 assert(_bit_map->isMarked(addr), "tautology"); 7571 if (_bit_map->isMarked(addr+1)) { 7572 // this is an allocated object that might not yet be initialized 7573 assert(_skip_bits == 0, "tautology"); 7574 _skip_bits = 2; // skip next two marked bits ("Printezis-marks") 7575 oop p = oop(addr); 7576 if (p->klass_or_null() == NULL) { 7577 // in the case of Clean-on-Enter optimization, redirty card 7578 // and avoid clearing card by increasing the threshold. 7579 return true; 7580 } 7581 } 7582 scan_oops_in_oop(addr); 7583 return true; 7584 } 7585 7586 void Par_MarkFromRootsClosure::scan_oops_in_oop(HeapWord* ptr) { 7587 assert(_bit_map->isMarked(ptr), "expected bit to be set"); 7588 // Should we assert that our work queue is empty or 7589 // below some drain limit? 7590 assert(_work_queue->size() == 0, 7591 "should drain stack to limit stack usage"); 7592 // convert ptr to an oop preparatory to scanning 7593 oop obj = oop(ptr); 7594 // Ignore mark word in verification below, since we 7595 // may be running concurrent with mutators. 7596 assert(obj->is_oop(true), "should be an oop"); 7597 assert(_finger <= ptr, "_finger runneth ahead"); 7598 // advance the finger to right end of this object 7599 _finger = ptr + obj->size(); 7600 assert(_finger > ptr, "we just incremented it above"); 7601 // On large heaps, it may take us some time to get through 7602 // the marking phase (especially if running iCMS). During 7603 // this time it's possible that a lot of mutations have 7604 // accumulated in the card table and the mod union table -- 7605 // these mutation records are redundant until we have 7606 // actually traced into the corresponding card. 7607 // Here, we check whether advancing the finger would make 7608 // us cross into a new card, and if so clear corresponding 7609 // cards in the MUT (preclean them in the card-table in the 7610 // future). 7611 7612 // The clean-on-enter optimization is disabled by default, 7613 // until we fix 6178663. 7614 if (CMSCleanOnEnter && (_finger > _threshold)) { 7615 // [_threshold, _finger) represents the interval 7616 // of cards to be cleared in MUT (or precleaned in card table). 7617 // The set of cards to be cleared is all those that overlap 7618 // with the interval [_threshold, _finger); note that 7619 // _threshold is always kept card-aligned but _finger isn't 7620 // always card-aligned. 7621 HeapWord* old_threshold = _threshold; 7622 assert(old_threshold == (HeapWord*)round_to( 7623 (intptr_t)old_threshold, CardTableModRefBS::card_size), 7624 "_threshold should always be card-aligned"); 7625 _threshold = (HeapWord*)round_to( 7626 (intptr_t)_finger, CardTableModRefBS::card_size); 7627 MemRegion mr(old_threshold, _threshold); 7628 assert(!mr.is_empty(), "Control point invariant"); 7629 assert(_span.contains(mr), "Should clear within span"); // _whole_span ?? 7630 _mut->clear_range(mr); 7631 } 7632 7633 // Note: the local finger doesn't advance while we drain 7634 // the stack below, but the global finger sure can and will. 7635 HeapWord** gfa = _task->global_finger_addr(); 7636 Par_PushOrMarkClosure pushOrMarkClosure(_collector, 7637 _span, _bit_map, 7638 _work_queue, 7639 _overflow_stack, 7640 _finger, 7641 gfa, this); 7642 bool res = _work_queue->push(obj); // overflow could occur here 7643 assert(res, "Will hold once we use workqueues"); 7644 while (true) { 7645 oop new_oop; 7646 if (!_work_queue->pop_local(new_oop)) { 7647 // We emptied our work_queue; check if there's stuff that can 7648 // be gotten from the overflow stack. 7649 if (CMSConcMarkingTask::get_work_from_overflow_stack( 7650 _overflow_stack, _work_queue)) { 7651 do_yield_check(); 7652 continue; 7653 } else { // done 7654 break; 7655 } 7656 } 7657 // Skip verifying header mark word below because we are 7658 // running concurrent with mutators. 7659 assert(new_oop->is_oop(true), "Oops! expected to pop an oop"); 7660 // now scan this oop's oops 7661 new_oop->oop_iterate(&pushOrMarkClosure); 7662 do_yield_check(); 7663 } 7664 assert(_work_queue->size() == 0, "tautology, emphasizing post-condition"); 7665 } 7666 7667 // Yield in response to a request from VM Thread or 7668 // from mutators. 7669 void Par_MarkFromRootsClosure::do_yield_work() { 7670 assert(_task != NULL, "sanity"); 7671 _task->yield(); 7672 } 7673 7674 // A variant of the above used for verifying CMS marking work. 7675 MarkFromRootsVerifyClosure::MarkFromRootsVerifyClosure(CMSCollector* collector, 7676 MemRegion span, 7677 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 7678 CMSMarkStack* mark_stack): 7679 _collector(collector), 7680 _span(span), 7681 _verification_bm(verification_bm), 7682 _cms_bm(cms_bm), 7683 _mark_stack(mark_stack), 7684 _pam_verify_closure(collector, span, verification_bm, cms_bm, 7685 mark_stack) 7686 { 7687 assert(_mark_stack->isEmpty(), "stack should be empty"); 7688 _finger = _verification_bm->startWord(); 7689 assert(_collector->_restart_addr == NULL, "Sanity check"); 7690 assert(_span.contains(_finger), "Out of bounds _finger?"); 7691 } 7692 7693 void MarkFromRootsVerifyClosure::reset(HeapWord* addr) { 7694 assert(_mark_stack->isEmpty(), "would cause duplicates on stack"); 7695 assert(_span.contains(addr), "Out of bounds _finger?"); 7696 _finger = addr; 7697 } 7698 7699 // Should revisit to see if this should be restructured for 7700 // greater efficiency. 7701 bool MarkFromRootsVerifyClosure::do_bit(size_t offset) { 7702 // convert offset into a HeapWord* 7703 HeapWord* addr = _verification_bm->startWord() + offset; 7704 assert(_verification_bm->endWord() && addr < _verification_bm->endWord(), 7705 "address out of range"); 7706 assert(_verification_bm->isMarked(addr), "tautology"); 7707 assert(_cms_bm->isMarked(addr), "tautology"); 7708 7709 assert(_mark_stack->isEmpty(), 7710 "should drain stack to limit stack usage"); 7711 // convert addr to an oop preparatory to scanning 7712 oop obj = oop(addr); 7713 assert(obj->is_oop(), "should be an oop"); 7714 assert(_finger <= addr, "_finger runneth ahead"); 7715 // advance the finger to right end of this object 7716 _finger = addr + obj->size(); 7717 assert(_finger > addr, "we just incremented it above"); 7718 // Note: the finger doesn't advance while we drain 7719 // the stack below. 7720 bool res = _mark_stack->push(obj); 7721 assert(res, "Empty non-zero size stack should have space for single push"); 7722 while (!_mark_stack->isEmpty()) { 7723 oop new_oop = _mark_stack->pop(); 7724 assert(new_oop->is_oop(), "Oops! expected to pop an oop"); 7725 // now scan this oop's oops 7726 new_oop->oop_iterate(&_pam_verify_closure); 7727 } 7728 assert(_mark_stack->isEmpty(), "tautology, emphasizing post-condition"); 7729 return true; 7730 } 7731 7732 PushAndMarkVerifyClosure::PushAndMarkVerifyClosure( 7733 CMSCollector* collector, MemRegion span, 7734 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 7735 CMSMarkStack* mark_stack): 7736 CMSOopClosure(collector->ref_processor()), 7737 _collector(collector), 7738 _span(span), 7739 _verification_bm(verification_bm), 7740 _cms_bm(cms_bm), 7741 _mark_stack(mark_stack) 7742 { } 7743 7744 void PushAndMarkVerifyClosure::do_oop(oop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 7745 void PushAndMarkVerifyClosure::do_oop(narrowOop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 7746 7747 // Upon stack overflow, we discard (part of) the stack, 7748 // remembering the least address amongst those discarded 7749 // in CMSCollector's _restart_address. 7750 void PushAndMarkVerifyClosure::handle_stack_overflow(HeapWord* lost) { 7751 // Remember the least grey address discarded 7752 HeapWord* ra = (HeapWord*)_mark_stack->least_value(lost); 7753 _collector->lower_restart_addr(ra); 7754 _mark_stack->reset(); // discard stack contents 7755 _mark_stack->expand(); // expand the stack if possible 7756 } 7757 7758 void PushAndMarkVerifyClosure::do_oop(oop obj) { 7759 assert(obj->is_oop_or_null(), "expected an oop or NULL"); 7760 HeapWord* addr = (HeapWord*)obj; 7761 if (_span.contains(addr) && !_verification_bm->isMarked(addr)) { 7762 // Oop lies in _span and isn't yet grey or black 7763 _verification_bm->mark(addr); // now grey 7764 if (!_cms_bm->isMarked(addr)) { 7765 oop(addr)->print(); 7766 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)", 7767 addr); 7768 fatal("... aborting"); 7769 } 7770 7771 if (!_mark_stack->push(obj)) { // stack overflow 7772 if (PrintCMSStatistics != 0) { 7773 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 7774 SIZE_FORMAT, _mark_stack->capacity()); 7775 } 7776 assert(_mark_stack->isFull(), "Else push should have succeeded"); 7777 handle_stack_overflow(addr); 7778 } 7779 // anything including and to the right of _finger 7780 // will be scanned as we iterate over the remainder of the 7781 // bit map 7782 } 7783 } 7784 7785 PushOrMarkClosure::PushOrMarkClosure(CMSCollector* collector, 7786 MemRegion span, 7787 CMSBitMap* bitMap, CMSMarkStack* markStack, 7788 HeapWord* finger, MarkFromRootsClosure* parent) : 7789 CMSOopClosure(collector->ref_processor()), 7790 _collector(collector), 7791 _span(span), 7792 _bitMap(bitMap), 7793 _markStack(markStack), 7794 _finger(finger), 7795 _parent(parent) 7796 { } 7797 7798 Par_PushOrMarkClosure::Par_PushOrMarkClosure(CMSCollector* collector, 7799 MemRegion span, 7800 CMSBitMap* bit_map, 7801 OopTaskQueue* work_queue, 7802 CMSMarkStack* overflow_stack, 7803 HeapWord* finger, 7804 HeapWord** global_finger_addr, 7805 Par_MarkFromRootsClosure* parent) : 7806 CMSOopClosure(collector->ref_processor()), 7807 _collector(collector), 7808 _whole_span(collector->_span), 7809 _span(span), 7810 _bit_map(bit_map), 7811 _work_queue(work_queue), 7812 _overflow_stack(overflow_stack), 7813 _finger(finger), 7814 _global_finger_addr(global_finger_addr), 7815 _parent(parent) 7816 { } 7817 7818 // Assumes thread-safe access by callers, who are 7819 // responsible for mutual exclusion. 7820 void CMSCollector::lower_restart_addr(HeapWord* low) { 7821 assert(_span.contains(low), "Out of bounds addr"); 7822 if (_restart_addr == NULL) { 7823 _restart_addr = low; 7824 } else { 7825 _restart_addr = MIN2(_restart_addr, low); 7826 } 7827 } 7828 7829 // Upon stack overflow, we discard (part of) the stack, 7830 // remembering the least address amongst those discarded 7831 // in CMSCollector's _restart_address. 7832 void PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 7833 // Remember the least grey address discarded 7834 HeapWord* ra = (HeapWord*)_markStack->least_value(lost); 7835 _collector->lower_restart_addr(ra); 7836 _markStack->reset(); // discard stack contents 7837 _markStack->expand(); // expand the stack if possible 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 Par_PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 7844 // We need to do this under a mutex to prevent other 7845 // workers from interfering with the work done below. 7846 MutexLockerEx ml(_overflow_stack->par_lock(), 7847 Mutex::_no_safepoint_check_flag); 7848 // Remember the least grey address discarded 7849 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 7850 _collector->lower_restart_addr(ra); 7851 _overflow_stack->reset(); // discard stack contents 7852 _overflow_stack->expand(); // expand the stack if possible 7853 } 7854 7855 void CMKlassClosure::do_klass(Klass* k) { 7856 assert(_oop_closure != NULL, "Not initialized?"); 7857 k->oops_do(_oop_closure); 7858 } 7859 7860 void PushOrMarkClosure::do_oop(oop obj) { 7861 // Ignore mark word because we are running concurrent with mutators. 7862 assert(obj->is_oop_or_null(true), "expected an oop or NULL"); 7863 HeapWord* addr = (HeapWord*)obj; 7864 if (_span.contains(addr) && !_bitMap->isMarked(addr)) { 7865 // Oop lies in _span and isn't yet grey or black 7866 _bitMap->mark(addr); // now grey 7867 if (addr < _finger) { 7868 // the bit map iteration has already either passed, or 7869 // sampled, this bit in the bit map; we'll need to 7870 // use the marking stack to scan this oop's oops. 7871 bool simulate_overflow = false; 7872 NOT_PRODUCT( 7873 if (CMSMarkStackOverflowALot && 7874 _collector->simulate_overflow()) { 7875 // simulate a stack overflow 7876 simulate_overflow = true; 7877 } 7878 ) 7879 if (simulate_overflow || !_markStack->push(obj)) { // stack overflow 7880 if (PrintCMSStatistics != 0) { 7881 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 7882 SIZE_FORMAT, _markStack->capacity()); 7883 } 7884 assert(simulate_overflow || _markStack->isFull(), "Else push should have succeeded"); 7885 handle_stack_overflow(addr); 7886 } 7887 } 7888 // anything including and to the right of _finger 7889 // will be scanned as we iterate over the remainder of the 7890 // bit map 7891 do_yield_check(); 7892 } 7893 } 7894 7895 void PushOrMarkClosure::do_oop(oop* p) { PushOrMarkClosure::do_oop_work(p); } 7896 void PushOrMarkClosure::do_oop(narrowOop* p) { PushOrMarkClosure::do_oop_work(p); } 7897 7898 void Par_PushOrMarkClosure::do_oop(oop obj) { 7899 // Ignore mark word because we are running concurrent with mutators. 7900 assert(obj->is_oop_or_null(true), "expected an oop or NULL"); 7901 HeapWord* addr = (HeapWord*)obj; 7902 if (_whole_span.contains(addr) && !_bit_map->isMarked(addr)) { 7903 // Oop lies in _span and isn't yet grey or black 7904 // We read the global_finger (volatile read) strictly after marking oop 7905 bool res = _bit_map->par_mark(addr); // now grey 7906 volatile HeapWord** gfa = (volatile HeapWord**)_global_finger_addr; 7907 // Should we push this marked oop on our stack? 7908 // -- if someone else marked it, nothing to do 7909 // -- if target oop is above global finger nothing to do 7910 // -- if target oop is in chunk and above local finger 7911 // then nothing to do 7912 // -- else push on work queue 7913 if ( !res // someone else marked it, they will deal with it 7914 || (addr >= *gfa) // will be scanned in a later task 7915 || (_span.contains(addr) && addr >= _finger)) { // later in this chunk 7916 return; 7917 } 7918 // the bit map iteration has already either passed, or 7919 // sampled, this bit in the bit map; we'll need to 7920 // use the marking stack to scan this oop's oops. 7921 bool simulate_overflow = false; 7922 NOT_PRODUCT( 7923 if (CMSMarkStackOverflowALot && 7924 _collector->simulate_overflow()) { 7925 // simulate a stack overflow 7926 simulate_overflow = true; 7927 } 7928 ) 7929 if (simulate_overflow || 7930 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) { 7931 // stack overflow 7932 if (PrintCMSStatistics != 0) { 7933 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 7934 SIZE_FORMAT, _overflow_stack->capacity()); 7935 } 7936 // We cannot assert that the overflow stack is full because 7937 // it may have been emptied since. 7938 assert(simulate_overflow || 7939 _work_queue->size() == _work_queue->max_elems(), 7940 "Else push should have succeeded"); 7941 handle_stack_overflow(addr); 7942 } 7943 do_yield_check(); 7944 } 7945 } 7946 7947 void Par_PushOrMarkClosure::do_oop(oop* p) { Par_PushOrMarkClosure::do_oop_work(p); } 7948 void Par_PushOrMarkClosure::do_oop(narrowOop* p) { Par_PushOrMarkClosure::do_oop_work(p); } 7949 7950 PushAndMarkClosure::PushAndMarkClosure(CMSCollector* collector, 7951 MemRegion span, 7952 ReferenceProcessor* rp, 7953 CMSBitMap* bit_map, 7954 CMSBitMap* mod_union_table, 7955 CMSMarkStack* mark_stack, 7956 bool concurrent_precleaning): 7957 CMSOopClosure(rp), 7958 _collector(collector), 7959 _span(span), 7960 _bit_map(bit_map), 7961 _mod_union_table(mod_union_table), 7962 _mark_stack(mark_stack), 7963 _concurrent_precleaning(concurrent_precleaning) 7964 { 7965 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 7966 } 7967 7968 // Grey object rescan during pre-cleaning and second checkpoint phases -- 7969 // the non-parallel version (the parallel version appears further below.) 7970 void PushAndMarkClosure::do_oop(oop obj) { 7971 // Ignore mark word verification. If during concurrent precleaning, 7972 // the object monitor may be locked. If during the checkpoint 7973 // phases, the object may already have been reached by a different 7974 // path and may be at the end of the global overflow list (so 7975 // the mark word may be NULL). 7976 assert(obj->is_oop_or_null(true /* ignore mark word */), 7977 "expected an oop or NULL"); 7978 HeapWord* addr = (HeapWord*)obj; 7979 // Check if oop points into the CMS generation 7980 // and is not marked 7981 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 7982 // a white object ... 7983 _bit_map->mark(addr); // ... now grey 7984 // push on the marking stack (grey set) 7985 bool simulate_overflow = false; 7986 NOT_PRODUCT( 7987 if (CMSMarkStackOverflowALot && 7988 _collector->simulate_overflow()) { 7989 // simulate a stack overflow 7990 simulate_overflow = true; 7991 } 7992 ) 7993 if (simulate_overflow || !_mark_stack->push(obj)) { 7994 if (_concurrent_precleaning) { 7995 // During precleaning we can just dirty the appropriate card(s) 7996 // in the mod union table, thus ensuring that the object remains 7997 // in the grey set and continue. In the case of object arrays 7998 // we need to dirty all of the cards that the object spans, 7999 // since the rescan of object arrays will be limited to the 8000 // dirty cards. 8001 // Note that no one can be interfering with us in this action 8002 // of dirtying the mod union table, so no locking or atomics 8003 // are required. 8004 if (obj->is_objArray()) { 8005 size_t sz = obj->size(); 8006 HeapWord* end_card_addr = (HeapWord*)round_to( 8007 (intptr_t)(addr+sz), CardTableModRefBS::card_size); 8008 MemRegion redirty_range = MemRegion(addr, end_card_addr); 8009 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 8010 _mod_union_table->mark_range(redirty_range); 8011 } else { 8012 _mod_union_table->mark(addr); 8013 } 8014 _collector->_ser_pmc_preclean_ovflw++; 8015 } else { 8016 // During the remark phase, we need to remember this oop 8017 // in the overflow list. 8018 _collector->push_on_overflow_list(obj); 8019 _collector->_ser_pmc_remark_ovflw++; 8020 } 8021 } 8022 } 8023 } 8024 8025 Par_PushAndMarkClosure::Par_PushAndMarkClosure(CMSCollector* collector, 8026 MemRegion span, 8027 ReferenceProcessor* rp, 8028 CMSBitMap* bit_map, 8029 OopTaskQueue* work_queue): 8030 CMSOopClosure(rp), 8031 _collector(collector), 8032 _span(span), 8033 _bit_map(bit_map), 8034 _work_queue(work_queue) 8035 { 8036 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 8037 } 8038 8039 void PushAndMarkClosure::do_oop(oop* p) { PushAndMarkClosure::do_oop_work(p); } 8040 void PushAndMarkClosure::do_oop(narrowOop* p) { PushAndMarkClosure::do_oop_work(p); } 8041 8042 // Grey object rescan during second checkpoint phase -- 8043 // the parallel version. 8044 void Par_PushAndMarkClosure::do_oop(oop obj) { 8045 // In the assert below, we ignore the mark word because 8046 // this oop may point to an already visited object that is 8047 // on the overflow stack (in which case the mark word has 8048 // been hijacked for chaining into the overflow stack -- 8049 // if this is the last object in the overflow stack then 8050 // its mark word will be NULL). Because this object may 8051 // have been subsequently popped off the global overflow 8052 // stack, and the mark word possibly restored to the prototypical 8053 // value, by the time we get to examined this failing assert in 8054 // the debugger, is_oop_or_null(false) may subsequently start 8055 // to hold. 8056 assert(obj->is_oop_or_null(true), 8057 "expected an oop or NULL"); 8058 HeapWord* addr = (HeapWord*)obj; 8059 // Check if oop points into the CMS generation 8060 // and is not marked 8061 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 8062 // a white object ... 8063 // If we manage to "claim" the object, by being the 8064 // first thread to mark it, then we push it on our 8065 // marking stack 8066 if (_bit_map->par_mark(addr)) { // ... now grey 8067 // push on work queue (grey set) 8068 bool simulate_overflow = false; 8069 NOT_PRODUCT( 8070 if (CMSMarkStackOverflowALot && 8071 _collector->par_simulate_overflow()) { 8072 // simulate a stack overflow 8073 simulate_overflow = true; 8074 } 8075 ) 8076 if (simulate_overflow || !_work_queue->push(obj)) { 8077 _collector->par_push_on_overflow_list(obj); 8078 _collector->_par_pmc_remark_ovflw++; // imprecise OK: no need to CAS 8079 } 8080 } // Else, some other thread got there first 8081 } 8082 } 8083 8084 void Par_PushAndMarkClosure::do_oop(oop* p) { Par_PushAndMarkClosure::do_oop_work(p); } 8085 void Par_PushAndMarkClosure::do_oop(narrowOop* p) { Par_PushAndMarkClosure::do_oop_work(p); } 8086 8087 void CMSPrecleanRefsYieldClosure::do_yield_work() { 8088 Mutex* bml = _collector->bitMapLock(); 8089 assert_lock_strong(bml); 8090 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 8091 "CMS thread should hold CMS token"); 8092 8093 bml->unlock(); 8094 ConcurrentMarkSweepThread::desynchronize(true); 8095 8096 ConcurrentMarkSweepThread::acknowledge_yield_request(); 8097 8098 _collector->stopTimer(); 8099 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 8100 if (PrintCMSStatistics != 0) { 8101 _collector->incrementYields(); 8102 } 8103 _collector->icms_wait(); 8104 8105 // See the comment in coordinator_yield() 8106 for (unsigned i = 0; i < CMSYieldSleepCount && 8107 ConcurrentMarkSweepThread::should_yield() && 8108 !CMSCollector::foregroundGCIsActive(); ++i) { 8109 os::sleep(Thread::current(), 1, false); 8110 ConcurrentMarkSweepThread::acknowledge_yield_request(); 8111 } 8112 8113 ConcurrentMarkSweepThread::synchronize(true); 8114 bml->lock(); 8115 8116 _collector->startTimer(); 8117 } 8118 8119 bool CMSPrecleanRefsYieldClosure::should_return() { 8120 if (ConcurrentMarkSweepThread::should_yield()) { 8121 do_yield_work(); 8122 } 8123 return _collector->foregroundGCIsActive(); 8124 } 8125 8126 void MarkFromDirtyCardsClosure::do_MemRegion(MemRegion mr) { 8127 assert(((size_t)mr.start())%CardTableModRefBS::card_size_in_words == 0, 8128 "mr should be aligned to start at a card boundary"); 8129 // We'd like to assert: 8130 // assert(mr.word_size()%CardTableModRefBS::card_size_in_words == 0, 8131 // "mr should be a range of cards"); 8132 // However, that would be too strong in one case -- the last 8133 // partition ends at _unallocated_block which, in general, can be 8134 // an arbitrary boundary, not necessarily card aligned. 8135 if (PrintCMSStatistics != 0) { 8136 _num_dirty_cards += 8137 mr.word_size()/CardTableModRefBS::card_size_in_words; 8138 } 8139 _space->object_iterate_mem(mr, &_scan_cl); 8140 } 8141 8142 SweepClosure::SweepClosure(CMSCollector* collector, 8143 ConcurrentMarkSweepGeneration* g, 8144 CMSBitMap* bitMap, bool should_yield) : 8145 _collector(collector), 8146 _g(g), 8147 _sp(g->cmsSpace()), 8148 _limit(_sp->sweep_limit()), 8149 _freelistLock(_sp->freelistLock()), 8150 _bitMap(bitMap), 8151 _yield(should_yield), 8152 _inFreeRange(false), // No free range at beginning of sweep 8153 _freeRangeInFreeLists(false), // No free range at beginning of sweep 8154 _lastFreeRangeCoalesced(false), 8155 _freeFinger(g->used_region().start()) 8156 { 8157 NOT_PRODUCT( 8158 _numObjectsFreed = 0; 8159 _numWordsFreed = 0; 8160 _numObjectsLive = 0; 8161 _numWordsLive = 0; 8162 _numObjectsAlreadyFree = 0; 8163 _numWordsAlreadyFree = 0; 8164 _last_fc = NULL; 8165 8166 _sp->initializeIndexedFreeListArrayReturnedBytes(); 8167 _sp->dictionary()->initialize_dict_returned_bytes(); 8168 ) 8169 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 8170 "sweep _limit out of bounds"); 8171 if (CMSTraceSweeper) { 8172 gclog_or_tty->print_cr("\n====================\nStarting new sweep with limit " PTR_FORMAT, 8173 _limit); 8174 } 8175 } 8176 8177 void SweepClosure::print_on(outputStream* st) const { 8178 tty->print_cr("_sp = [" PTR_FORMAT "," PTR_FORMAT ")", 8179 _sp->bottom(), _sp->end()); 8180 tty->print_cr("_limit = " PTR_FORMAT, _limit); 8181 tty->print_cr("_freeFinger = " PTR_FORMAT, _freeFinger); 8182 NOT_PRODUCT(tty->print_cr("_last_fc = " PTR_FORMAT, _last_fc);) 8183 tty->print_cr("_inFreeRange = %d, _freeRangeInFreeLists = %d, _lastFreeRangeCoalesced = %d", 8184 _inFreeRange, _freeRangeInFreeLists, _lastFreeRangeCoalesced); 8185 } 8186 8187 #ifndef PRODUCT 8188 // Assertion checking only: no useful work in product mode -- 8189 // however, if any of the flags below become product flags, 8190 // you may need to review this code to see if it needs to be 8191 // enabled in product mode. 8192 SweepClosure::~SweepClosure() { 8193 assert_lock_strong(_freelistLock); 8194 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 8195 "sweep _limit out of bounds"); 8196 if (inFreeRange()) { 8197 warning("inFreeRange() should have been reset; dumping state of SweepClosure"); 8198 print(); 8199 ShouldNotReachHere(); 8200 } 8201 if (Verbose && PrintGC) { 8202 gclog_or_tty->print("Collected "SIZE_FORMAT" objects, " SIZE_FORMAT " bytes", 8203 _numObjectsFreed, _numWordsFreed*sizeof(HeapWord)); 8204 gclog_or_tty->print_cr("\nLive "SIZE_FORMAT" objects, " 8205 SIZE_FORMAT" bytes " 8206 "Already free "SIZE_FORMAT" objects, "SIZE_FORMAT" bytes", 8207 _numObjectsLive, _numWordsLive*sizeof(HeapWord), 8208 _numObjectsAlreadyFree, _numWordsAlreadyFree*sizeof(HeapWord)); 8209 size_t totalBytes = (_numWordsFreed + _numWordsLive + _numWordsAlreadyFree) 8210 * sizeof(HeapWord); 8211 gclog_or_tty->print_cr("Total sweep: "SIZE_FORMAT" bytes", totalBytes); 8212 8213 if (PrintCMSStatistics && CMSVerifyReturnedBytes) { 8214 size_t indexListReturnedBytes = _sp->sumIndexedFreeListArrayReturnedBytes(); 8215 size_t dict_returned_bytes = _sp->dictionary()->sum_dict_returned_bytes(); 8216 size_t returned_bytes = indexListReturnedBytes + dict_returned_bytes; 8217 gclog_or_tty->print("Returned "SIZE_FORMAT" bytes", returned_bytes); 8218 gclog_or_tty->print(" Indexed List Returned "SIZE_FORMAT" bytes", 8219 indexListReturnedBytes); 8220 gclog_or_tty->print_cr(" Dictionary Returned "SIZE_FORMAT" bytes", 8221 dict_returned_bytes); 8222 } 8223 } 8224 if (CMSTraceSweeper) { 8225 gclog_or_tty->print_cr("end of sweep with _limit = " PTR_FORMAT "\n================", 8226 _limit); 8227 } 8228 } 8229 #endif // PRODUCT 8230 8231 void SweepClosure::initialize_free_range(HeapWord* freeFinger, 8232 bool freeRangeInFreeLists) { 8233 if (CMSTraceSweeper) { 8234 gclog_or_tty->print("---- Start free range at " PTR_FORMAT " with free block (%d)\n", 8235 freeFinger, freeRangeInFreeLists); 8236 } 8237 assert(!inFreeRange(), "Trampling existing free range"); 8238 set_inFreeRange(true); 8239 set_lastFreeRangeCoalesced(false); 8240 8241 set_freeFinger(freeFinger); 8242 set_freeRangeInFreeLists(freeRangeInFreeLists); 8243 if (CMSTestInFreeList) { 8244 if (freeRangeInFreeLists) { 8245 FreeChunk* fc = (FreeChunk*) freeFinger; 8246 assert(fc->is_free(), "A chunk on the free list should be free."); 8247 assert(fc->size() > 0, "Free range should have a size"); 8248 assert(_sp->verify_chunk_in_free_list(fc), "Chunk is not in free lists"); 8249 } 8250 } 8251 } 8252 8253 // Note that the sweeper runs concurrently with mutators. Thus, 8254 // it is possible for direct allocation in this generation to happen 8255 // in the middle of the sweep. Note that the sweeper also coalesces 8256 // contiguous free blocks. Thus, unless the sweeper and the allocator 8257 // synchronize appropriately freshly allocated blocks may get swept up. 8258 // This is accomplished by the sweeper locking the free lists while 8259 // it is sweeping. Thus blocks that are determined to be free are 8260 // indeed free. There is however one additional complication: 8261 // blocks that have been allocated since the final checkpoint and 8262 // mark, will not have been marked and so would be treated as 8263 // unreachable and swept up. To prevent this, the allocator marks 8264 // the bit map when allocating during the sweep phase. This leads, 8265 // however, to a further complication -- objects may have been allocated 8266 // but not yet initialized -- in the sense that the header isn't yet 8267 // installed. The sweeper can not then determine the size of the block 8268 // in order to skip over it. To deal with this case, we use a technique 8269 // (due to Printezis) to encode such uninitialized block sizes in the 8270 // bit map. Since the bit map uses a bit per every HeapWord, but the 8271 // CMS generation has a minimum object size of 3 HeapWords, it follows 8272 // that "normal marks" won't be adjacent in the bit map (there will 8273 // always be at least two 0 bits between successive 1 bits). We make use 8274 // of these "unused" bits to represent uninitialized blocks -- the bit 8275 // corresponding to the start of the uninitialized object and the next 8276 // bit are both set. Finally, a 1 bit marks the end of the object that 8277 // started with the two consecutive 1 bits to indicate its potentially 8278 // uninitialized state. 8279 8280 size_t SweepClosure::do_blk_careful(HeapWord* addr) { 8281 FreeChunk* fc = (FreeChunk*)addr; 8282 size_t res; 8283 8284 // Check if we are done sweeping. Below we check "addr >= _limit" rather 8285 // than "addr == _limit" because although _limit was a block boundary when 8286 // we started the sweep, it may no longer be one because heap expansion 8287 // may have caused us to coalesce the block ending at the address _limit 8288 // with a newly expanded chunk (this happens when _limit was set to the 8289 // previous _end of the space), so we may have stepped past _limit: 8290 // see the following Zeno-like trail of CRs 6977970, 7008136, 7042740. 8291 if (addr >= _limit) { // we have swept up to or past the limit: finish up 8292 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 8293 "sweep _limit out of bounds"); 8294 assert(addr < _sp->end(), "addr out of bounds"); 8295 // Flush any free range we might be holding as a single 8296 // coalesced chunk to the appropriate free list. 8297 if (inFreeRange()) { 8298 assert(freeFinger() >= _sp->bottom() && freeFinger() < _limit, 8299 err_msg("freeFinger() " PTR_FORMAT" is out-of-bounds", freeFinger())); 8300 flush_cur_free_chunk(freeFinger(), 8301 pointer_delta(addr, freeFinger())); 8302 if (CMSTraceSweeper) { 8303 gclog_or_tty->print("Sweep: last chunk: "); 8304 gclog_or_tty->print("put_free_blk " PTR_FORMAT " ("SIZE_FORMAT") " 8305 "[coalesced:%d]\n", 8306 freeFinger(), pointer_delta(addr, freeFinger()), 8307 lastFreeRangeCoalesced() ? 1 : 0); 8308 } 8309 } 8310 8311 // help the iterator loop finish 8312 return pointer_delta(_sp->end(), addr); 8313 } 8314 8315 assert(addr < _limit, "sweep invariant"); 8316 // check if we should yield 8317 do_yield_check(addr); 8318 if (fc->is_free()) { 8319 // Chunk that is already free 8320 res = fc->size(); 8321 do_already_free_chunk(fc); 8322 debug_only(_sp->verifyFreeLists()); 8323 // If we flush the chunk at hand in lookahead_and_flush() 8324 // and it's coalesced with a preceding chunk, then the 8325 // process of "mangling" the payload of the coalesced block 8326 // will cause erasure of the size information from the 8327 // (erstwhile) header of all the coalesced blocks but the 8328 // first, so the first disjunct in the assert will not hold 8329 // in that specific case (in which case the second disjunct 8330 // will hold). 8331 assert(res == fc->size() || ((HeapWord*)fc) + res >= _limit, 8332 "Otherwise the size info doesn't change at this step"); 8333 NOT_PRODUCT( 8334 _numObjectsAlreadyFree++; 8335 _numWordsAlreadyFree += res; 8336 ) 8337 NOT_PRODUCT(_last_fc = fc;) 8338 } else if (!_bitMap->isMarked(addr)) { 8339 // Chunk is fresh garbage 8340 res = do_garbage_chunk(fc); 8341 debug_only(_sp->verifyFreeLists()); 8342 NOT_PRODUCT( 8343 _numObjectsFreed++; 8344 _numWordsFreed += res; 8345 ) 8346 } else { 8347 // Chunk that is alive. 8348 res = do_live_chunk(fc); 8349 debug_only(_sp->verifyFreeLists()); 8350 NOT_PRODUCT( 8351 _numObjectsLive++; 8352 _numWordsLive += res; 8353 ) 8354 } 8355 return res; 8356 } 8357 8358 // For the smart allocation, record following 8359 // split deaths - a free chunk is removed from its free list because 8360 // it is being split into two or more chunks. 8361 // split birth - a free chunk is being added to its free list because 8362 // a larger free chunk has been split and resulted in this free chunk. 8363 // coal death - a free chunk is being removed from its free list because 8364 // it is being coalesced into a large free chunk. 8365 // coal birth - a free chunk is being added to its free list because 8366 // it was created when two or more free chunks where coalesced into 8367 // this free chunk. 8368 // 8369 // These statistics are used to determine the desired number of free 8370 // chunks of a given size. The desired number is chosen to be relative 8371 // to the end of a CMS sweep. The desired number at the end of a sweep 8372 // is the 8373 // count-at-end-of-previous-sweep (an amount that was enough) 8374 // - count-at-beginning-of-current-sweep (the excess) 8375 // + split-births (gains in this size during interval) 8376 // - split-deaths (demands on this size during interval) 8377 // where the interval is from the end of one sweep to the end of the 8378 // next. 8379 // 8380 // When sweeping the sweeper maintains an accumulated chunk which is 8381 // the chunk that is made up of chunks that have been coalesced. That 8382 // will be termed the left-hand chunk. A new chunk of garbage that 8383 // is being considered for coalescing will be referred to as the 8384 // right-hand chunk. 8385 // 8386 // When making a decision on whether to coalesce a right-hand chunk with 8387 // the current left-hand chunk, the current count vs. the desired count 8388 // of the left-hand chunk is considered. Also if the right-hand chunk 8389 // is near the large chunk at the end of the heap (see 8390 // ConcurrentMarkSweepGeneration::isNearLargestChunk()), then the 8391 // left-hand chunk is coalesced. 8392 // 8393 // When making a decision about whether to split a chunk, the desired count 8394 // vs. the current count of the candidate to be split is also considered. 8395 // If the candidate is underpopulated (currently fewer chunks than desired) 8396 // a chunk of an overpopulated (currently more chunks than desired) size may 8397 // be chosen. The "hint" associated with a free list, if non-null, points 8398 // to a free list which may be overpopulated. 8399 // 8400 8401 void SweepClosure::do_already_free_chunk(FreeChunk* fc) { 8402 const size_t size = fc->size(); 8403 // Chunks that cannot be coalesced are not in the 8404 // free lists. 8405 if (CMSTestInFreeList && !fc->cantCoalesce()) { 8406 assert(_sp->verify_chunk_in_free_list(fc), 8407 "free chunk should be in free lists"); 8408 } 8409 // a chunk that is already free, should not have been 8410 // marked in the bit map 8411 HeapWord* const addr = (HeapWord*) fc; 8412 assert(!_bitMap->isMarked(addr), "free chunk should be unmarked"); 8413 // Verify that the bit map has no bits marked between 8414 // addr and purported end of this block. 8415 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 8416 8417 // Some chunks cannot be coalesced under any circumstances. 8418 // See the definition of cantCoalesce(). 8419 if (!fc->cantCoalesce()) { 8420 // This chunk can potentially be coalesced. 8421 if (_sp->adaptive_freelists()) { 8422 // All the work is done in 8423 do_post_free_or_garbage_chunk(fc, size); 8424 } else { // Not adaptive free lists 8425 // this is a free chunk that can potentially be coalesced by the sweeper; 8426 if (!inFreeRange()) { 8427 // if the next chunk is a free block that can't be coalesced 8428 // it doesn't make sense to remove this chunk from the free lists 8429 FreeChunk* nextChunk = (FreeChunk*)(addr + size); 8430 assert((HeapWord*)nextChunk <= _sp->end(), "Chunk size out of bounds?"); 8431 if ((HeapWord*)nextChunk < _sp->end() && // There is another free chunk to the right ... 8432 nextChunk->is_free() && // ... which is free... 8433 nextChunk->cantCoalesce()) { // ... but can't be coalesced 8434 // nothing to do 8435 } else { 8436 // Potentially the start of a new free range: 8437 // Don't eagerly remove it from the free lists. 8438 // No need to remove it if it will just be put 8439 // back again. (Also from a pragmatic point of view 8440 // if it is a free block in a region that is beyond 8441 // any allocated blocks, an assertion will fail) 8442 // Remember the start of a free run. 8443 initialize_free_range(addr, true); 8444 // end - can coalesce with next chunk 8445 } 8446 } else { 8447 // the midst of a free range, we are coalescing 8448 print_free_block_coalesced(fc); 8449 if (CMSTraceSweeper) { 8450 gclog_or_tty->print(" -- pick up free block " PTR_FORMAT " (" SIZE_FORMAT ")\n", fc, size); 8451 } 8452 // remove it from the free lists 8453 _sp->removeFreeChunkFromFreeLists(fc); 8454 set_lastFreeRangeCoalesced(true); 8455 // If the chunk is being coalesced and the current free range is 8456 // in the free lists, remove the current free range so that it 8457 // will be returned to the free lists in its entirety - all 8458 // the coalesced pieces included. 8459 if (freeRangeInFreeLists()) { 8460 FreeChunk* ffc = (FreeChunk*) freeFinger(); 8461 assert(ffc->size() == pointer_delta(addr, freeFinger()), 8462 "Size of free range is inconsistent with chunk size."); 8463 if (CMSTestInFreeList) { 8464 assert(_sp->verify_chunk_in_free_list(ffc), 8465 "free range is not in free lists"); 8466 } 8467 _sp->removeFreeChunkFromFreeLists(ffc); 8468 set_freeRangeInFreeLists(false); 8469 } 8470 } 8471 } 8472 // Note that if the chunk is not coalescable (the else arm 8473 // below), we unconditionally flush, without needing to do 8474 // a "lookahead," as we do below. 8475 if (inFreeRange()) lookahead_and_flush(fc, size); 8476 } else { 8477 // Code path common to both original and adaptive free lists. 8478 8479 // cant coalesce with previous block; this should be treated 8480 // as the end of a free run if any 8481 if (inFreeRange()) { 8482 // we kicked some butt; time to pick up the garbage 8483 assert(freeFinger() < addr, "freeFinger points too high"); 8484 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 8485 } 8486 // else, nothing to do, just continue 8487 } 8488 } 8489 8490 size_t SweepClosure::do_garbage_chunk(FreeChunk* fc) { 8491 // This is a chunk of garbage. It is not in any free list. 8492 // Add it to a free list or let it possibly be coalesced into 8493 // a larger chunk. 8494 HeapWord* const addr = (HeapWord*) fc; 8495 const size_t size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 8496 8497 if (_sp->adaptive_freelists()) { 8498 // Verify that the bit map has no bits marked between 8499 // addr and purported end of just dead object. 8500 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 8501 8502 do_post_free_or_garbage_chunk(fc, size); 8503 } else { 8504 if (!inFreeRange()) { 8505 // start of a new free range 8506 assert(size > 0, "A free range should have a size"); 8507 initialize_free_range(addr, false); 8508 } else { 8509 // this will be swept up when we hit the end of the 8510 // free range 8511 if (CMSTraceSweeper) { 8512 gclog_or_tty->print(" -- pick up garbage " PTR_FORMAT " (" SIZE_FORMAT ")\n", fc, size); 8513 } 8514 // If the chunk is being coalesced and the current free range is 8515 // in the free lists, remove the current free range so that it 8516 // will be returned to the free lists in its entirety - all 8517 // the coalesced pieces included. 8518 if (freeRangeInFreeLists()) { 8519 FreeChunk* ffc = (FreeChunk*)freeFinger(); 8520 assert(ffc->size() == pointer_delta(addr, freeFinger()), 8521 "Size of free range is inconsistent with chunk size."); 8522 if (CMSTestInFreeList) { 8523 assert(_sp->verify_chunk_in_free_list(ffc), 8524 "free range is not in free lists"); 8525 } 8526 _sp->removeFreeChunkFromFreeLists(ffc); 8527 set_freeRangeInFreeLists(false); 8528 } 8529 set_lastFreeRangeCoalesced(true); 8530 } 8531 // this will be swept up when we hit the end of the free range 8532 8533 // Verify that the bit map has no bits marked between 8534 // addr and purported end of just dead object. 8535 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 8536 } 8537 assert(_limit >= addr + size, 8538 "A freshly garbage chunk can't possibly straddle over _limit"); 8539 if (inFreeRange()) lookahead_and_flush(fc, size); 8540 return size; 8541 } 8542 8543 size_t SweepClosure::do_live_chunk(FreeChunk* fc) { 8544 HeapWord* addr = (HeapWord*) fc; 8545 // The sweeper has just found a live object. Return any accumulated 8546 // left hand chunk to the free lists. 8547 if (inFreeRange()) { 8548 assert(freeFinger() < addr, "freeFinger points too high"); 8549 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 8550 } 8551 8552 // This object is live: we'd normally expect this to be 8553 // an oop, and like to assert the following: 8554 // assert(oop(addr)->is_oop(), "live block should be an oop"); 8555 // However, as we commented above, this may be an object whose 8556 // header hasn't yet been initialized. 8557 size_t size; 8558 assert(_bitMap->isMarked(addr), "Tautology for this control point"); 8559 if (_bitMap->isMarked(addr + 1)) { 8560 // Determine the size from the bit map, rather than trying to 8561 // compute it from the object header. 8562 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 8563 size = pointer_delta(nextOneAddr + 1, addr); 8564 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 8565 "alignment problem"); 8566 8567 #ifdef ASSERT 8568 if (oop(addr)->klass_or_null() != NULL) { 8569 // Ignore mark word because we are running concurrent with mutators 8570 assert(oop(addr)->is_oop(true), "live block should be an oop"); 8571 assert(size == 8572 CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()), 8573 "P-mark and computed size do not agree"); 8574 } 8575 #endif 8576 8577 } else { 8578 // This should be an initialized object that's alive. 8579 assert(oop(addr)->klass_or_null() != NULL, 8580 "Should be an initialized object"); 8581 // Ignore mark word because we are running concurrent with mutators 8582 assert(oop(addr)->is_oop(true), "live block should be an oop"); 8583 // Verify that the bit map has no bits marked between 8584 // addr and purported end of this block. 8585 size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 8586 assert(size >= 3, "Necessary for Printezis marks to work"); 8587 assert(!_bitMap->isMarked(addr+1), "Tautology for this control point"); 8588 DEBUG_ONLY(_bitMap->verifyNoOneBitsInRange(addr+2, addr+size);) 8589 } 8590 return size; 8591 } 8592 8593 void SweepClosure::do_post_free_or_garbage_chunk(FreeChunk* fc, 8594 size_t chunkSize) { 8595 // do_post_free_or_garbage_chunk() should only be called in the case 8596 // of the adaptive free list allocator. 8597 const bool fcInFreeLists = fc->is_free(); 8598 assert(_sp->adaptive_freelists(), "Should only be used in this case."); 8599 assert((HeapWord*)fc <= _limit, "sweep invariant"); 8600 if (CMSTestInFreeList && fcInFreeLists) { 8601 assert(_sp->verify_chunk_in_free_list(fc), "free chunk is not in free lists"); 8602 } 8603 8604 if (CMSTraceSweeper) { 8605 gclog_or_tty->print_cr(" -- pick up another chunk at " PTR_FORMAT " (" SIZE_FORMAT ")", fc, chunkSize); 8606 } 8607 8608 HeapWord* const fc_addr = (HeapWord*) fc; 8609 8610 bool coalesce; 8611 const size_t left = pointer_delta(fc_addr, freeFinger()); 8612 const size_t right = chunkSize; 8613 switch (FLSCoalescePolicy) { 8614 // numeric value forms a coalition aggressiveness metric 8615 case 0: { // never coalesce 8616 coalesce = false; 8617 break; 8618 } 8619 case 1: { // coalesce if left & right chunks on overpopulated lists 8620 coalesce = _sp->coalOverPopulated(left) && 8621 _sp->coalOverPopulated(right); 8622 break; 8623 } 8624 case 2: { // coalesce if left chunk on overpopulated list (default) 8625 coalesce = _sp->coalOverPopulated(left); 8626 break; 8627 } 8628 case 3: { // coalesce if left OR right chunk on overpopulated list 8629 coalesce = _sp->coalOverPopulated(left) || 8630 _sp->coalOverPopulated(right); 8631 break; 8632 } 8633 case 4: { // always coalesce 8634 coalesce = true; 8635 break; 8636 } 8637 default: 8638 ShouldNotReachHere(); 8639 } 8640 8641 // Should the current free range be coalesced? 8642 // If the chunk is in a free range and either we decided to coalesce above 8643 // or the chunk is near the large block at the end of the heap 8644 // (isNearLargestChunk() returns true), then coalesce this chunk. 8645 const bool doCoalesce = inFreeRange() 8646 && (coalesce || _g->isNearLargestChunk(fc_addr)); 8647 if (doCoalesce) { 8648 // Coalesce the current free range on the left with the new 8649 // chunk on the right. If either is on a free list, 8650 // it must be removed from the list and stashed in the closure. 8651 if (freeRangeInFreeLists()) { 8652 FreeChunk* const ffc = (FreeChunk*)freeFinger(); 8653 assert(ffc->size() == pointer_delta(fc_addr, freeFinger()), 8654 "Size of free range is inconsistent with chunk size."); 8655 if (CMSTestInFreeList) { 8656 assert(_sp->verify_chunk_in_free_list(ffc), 8657 "Chunk is not in free lists"); 8658 } 8659 _sp->coalDeath(ffc->size()); 8660 _sp->removeFreeChunkFromFreeLists(ffc); 8661 set_freeRangeInFreeLists(false); 8662 } 8663 if (fcInFreeLists) { 8664 _sp->coalDeath(chunkSize); 8665 assert(fc->size() == chunkSize, 8666 "The chunk has the wrong size or is not in the free lists"); 8667 _sp->removeFreeChunkFromFreeLists(fc); 8668 } 8669 set_lastFreeRangeCoalesced(true); 8670 print_free_block_coalesced(fc); 8671 } else { // not in a free range and/or should not coalesce 8672 // Return the current free range and start a new one. 8673 if (inFreeRange()) { 8674 // In a free range but cannot coalesce with the right hand chunk. 8675 // Put the current free range into the free lists. 8676 flush_cur_free_chunk(freeFinger(), 8677 pointer_delta(fc_addr, freeFinger())); 8678 } 8679 // Set up for new free range. Pass along whether the right hand 8680 // chunk is in the free lists. 8681 initialize_free_range((HeapWord*)fc, fcInFreeLists); 8682 } 8683 } 8684 8685 // Lookahead flush: 8686 // If we are tracking a free range, and this is the last chunk that 8687 // we'll look at because its end crosses past _limit, we'll preemptively 8688 // flush it along with any free range we may be holding on to. Note that 8689 // this can be the case only for an already free or freshly garbage 8690 // chunk. If this block is an object, it can never straddle 8691 // over _limit. The "straddling" occurs when _limit is set at 8692 // the previous end of the space when this cycle started, and 8693 // a subsequent heap expansion caused the previously co-terminal 8694 // free block to be coalesced with the newly expanded portion, 8695 // thus rendering _limit a non-block-boundary making it dangerous 8696 // for the sweeper to step over and examine. 8697 void SweepClosure::lookahead_and_flush(FreeChunk* fc, size_t chunk_size) { 8698 assert(inFreeRange(), "Should only be called if currently in a free range."); 8699 HeapWord* const eob = ((HeapWord*)fc) + chunk_size; 8700 assert(_sp->used_region().contains(eob - 1), 8701 err_msg("eob = " PTR_FORMAT " eob-1 = " PTR_FORMAT " _limit = " PTR_FORMAT 8702 " out of bounds wrt _sp = [" PTR_FORMAT "," PTR_FORMAT ")" 8703 " when examining fc = " PTR_FORMAT "(" SIZE_FORMAT ")", 8704 eob, eob-1, _limit, _sp->bottom(), _sp->end(), fc, chunk_size)); 8705 if (eob >= _limit) { 8706 assert(eob == _limit || fc->is_free(), "Only a free chunk should allow us to cross over the limit"); 8707 if (CMSTraceSweeper) { 8708 gclog_or_tty->print_cr("_limit " PTR_FORMAT " reached or crossed by block " 8709 "[" PTR_FORMAT "," PTR_FORMAT ") in space " 8710 "[" PTR_FORMAT "," PTR_FORMAT ")", 8711 _limit, fc, eob, _sp->bottom(), _sp->end()); 8712 } 8713 // Return the storage we are tracking back into the free lists. 8714 if (CMSTraceSweeper) { 8715 gclog_or_tty->print_cr("Flushing ... "); 8716 } 8717 assert(freeFinger() < eob, "Error"); 8718 flush_cur_free_chunk( freeFinger(), pointer_delta(eob, freeFinger())); 8719 } 8720 } 8721 8722 void SweepClosure::flush_cur_free_chunk(HeapWord* chunk, size_t size) { 8723 assert(inFreeRange(), "Should only be called if currently in a free range."); 8724 assert(size > 0, 8725 "A zero sized chunk cannot be added to the free lists."); 8726 if (!freeRangeInFreeLists()) { 8727 if (CMSTestInFreeList) { 8728 FreeChunk* fc = (FreeChunk*) chunk; 8729 fc->set_size(size); 8730 assert(!_sp->verify_chunk_in_free_list(fc), 8731 "chunk should not be in free lists yet"); 8732 } 8733 if (CMSTraceSweeper) { 8734 gclog_or_tty->print_cr(" -- add free block " PTR_FORMAT " (" SIZE_FORMAT ") to free lists", 8735 chunk, size); 8736 } 8737 // A new free range is going to be starting. The current 8738 // free range has not been added to the free lists yet or 8739 // was removed so add it back. 8740 // If the current free range was coalesced, then the death 8741 // of the free range was recorded. Record a birth now. 8742 if (lastFreeRangeCoalesced()) { 8743 _sp->coalBirth(size); 8744 } 8745 _sp->addChunkAndRepairOffsetTable(chunk, size, 8746 lastFreeRangeCoalesced()); 8747 } else if (CMSTraceSweeper) { 8748 gclog_or_tty->print_cr("Already in free list: nothing to flush"); 8749 } 8750 set_inFreeRange(false); 8751 set_freeRangeInFreeLists(false); 8752 } 8753 8754 // We take a break if we've been at this for a while, 8755 // so as to avoid monopolizing the locks involved. 8756 void SweepClosure::do_yield_work(HeapWord* addr) { 8757 // Return current free chunk being used for coalescing (if any) 8758 // to the appropriate freelist. After yielding, the next 8759 // free block encountered will start a coalescing range of 8760 // free blocks. If the next free block is adjacent to the 8761 // chunk just flushed, they will need to wait for the next 8762 // sweep to be coalesced. 8763 if (inFreeRange()) { 8764 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 8765 } 8766 8767 // First give up the locks, then yield, then re-lock. 8768 // We should probably use a constructor/destructor idiom to 8769 // do this unlock/lock or modify the MutexUnlocker class to 8770 // serve our purpose. XXX 8771 assert_lock_strong(_bitMap->lock()); 8772 assert_lock_strong(_freelistLock); 8773 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 8774 "CMS thread should hold CMS token"); 8775 _bitMap->lock()->unlock(); 8776 _freelistLock->unlock(); 8777 ConcurrentMarkSweepThread::desynchronize(true); 8778 ConcurrentMarkSweepThread::acknowledge_yield_request(); 8779 _collector->stopTimer(); 8780 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 8781 if (PrintCMSStatistics != 0) { 8782 _collector->incrementYields(); 8783 } 8784 _collector->icms_wait(); 8785 8786 // See the comment in coordinator_yield() 8787 for (unsigned i = 0; i < CMSYieldSleepCount && 8788 ConcurrentMarkSweepThread::should_yield() && 8789 !CMSCollector::foregroundGCIsActive(); ++i) { 8790 os::sleep(Thread::current(), 1, false); 8791 ConcurrentMarkSweepThread::acknowledge_yield_request(); 8792 } 8793 8794 ConcurrentMarkSweepThread::synchronize(true); 8795 _freelistLock->lock(); 8796 _bitMap->lock()->lock_without_safepoint_check(); 8797 _collector->startTimer(); 8798 } 8799 8800 #ifndef PRODUCT 8801 // This is actually very useful in a product build if it can 8802 // be called from the debugger. Compile it into the product 8803 // as needed. 8804 bool debug_verify_chunk_in_free_list(FreeChunk* fc) { 8805 return debug_cms_space->verify_chunk_in_free_list(fc); 8806 } 8807 #endif 8808 8809 void SweepClosure::print_free_block_coalesced(FreeChunk* fc) const { 8810 if (CMSTraceSweeper) { 8811 gclog_or_tty->print_cr("Sweep:coal_free_blk " PTR_FORMAT " (" SIZE_FORMAT ")", 8812 fc, fc->size()); 8813 } 8814 } 8815 8816 // CMSIsAliveClosure 8817 bool CMSIsAliveClosure::do_object_b(oop obj) { 8818 HeapWord* addr = (HeapWord*)obj; 8819 return addr != NULL && 8820 (!_span.contains(addr) || _bit_map->isMarked(addr)); 8821 } 8822 8823 8824 CMSKeepAliveClosure::CMSKeepAliveClosure( CMSCollector* collector, 8825 MemRegion span, 8826 CMSBitMap* bit_map, CMSMarkStack* mark_stack, 8827 bool cpc): 8828 _collector(collector), 8829 _span(span), 8830 _bit_map(bit_map), 8831 _mark_stack(mark_stack), 8832 _concurrent_precleaning(cpc) { 8833 assert(!_span.is_empty(), "Empty span could spell trouble"); 8834 } 8835 8836 8837 // CMSKeepAliveClosure: the serial version 8838 void CMSKeepAliveClosure::do_oop(oop obj) { 8839 HeapWord* addr = (HeapWord*)obj; 8840 if (_span.contains(addr) && 8841 !_bit_map->isMarked(addr)) { 8842 _bit_map->mark(addr); 8843 bool simulate_overflow = false; 8844 NOT_PRODUCT( 8845 if (CMSMarkStackOverflowALot && 8846 _collector->simulate_overflow()) { 8847 // simulate a stack overflow 8848 simulate_overflow = true; 8849 } 8850 ) 8851 if (simulate_overflow || !_mark_stack->push(obj)) { 8852 if (_concurrent_precleaning) { 8853 // We dirty the overflown object and let the remark 8854 // phase deal with it. 8855 assert(_collector->overflow_list_is_empty(), "Error"); 8856 // In the case of object arrays, we need to dirty all of 8857 // the cards that the object spans. No locking or atomics 8858 // are needed since no one else can be mutating the mod union 8859 // table. 8860 if (obj->is_objArray()) { 8861 size_t sz = obj->size(); 8862 HeapWord* end_card_addr = 8863 (HeapWord*)round_to((intptr_t)(addr+sz), CardTableModRefBS::card_size); 8864 MemRegion redirty_range = MemRegion(addr, end_card_addr); 8865 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 8866 _collector->_modUnionTable.mark_range(redirty_range); 8867 } else { 8868 _collector->_modUnionTable.mark(addr); 8869 } 8870 _collector->_ser_kac_preclean_ovflw++; 8871 } else { 8872 _collector->push_on_overflow_list(obj); 8873 _collector->_ser_kac_ovflw++; 8874 } 8875 } 8876 } 8877 } 8878 8879 void CMSKeepAliveClosure::do_oop(oop* p) { CMSKeepAliveClosure::do_oop_work(p); } 8880 void CMSKeepAliveClosure::do_oop(narrowOop* p) { CMSKeepAliveClosure::do_oop_work(p); } 8881 8882 // CMSParKeepAliveClosure: a parallel version of the above. 8883 // The work queues are private to each closure (thread), 8884 // but (may be) available for stealing by other threads. 8885 void CMSParKeepAliveClosure::do_oop(oop obj) { 8886 HeapWord* addr = (HeapWord*)obj; 8887 if (_span.contains(addr) && 8888 !_bit_map->isMarked(addr)) { 8889 // In general, during recursive tracing, several threads 8890 // may be concurrently getting here; the first one to 8891 // "tag" it, claims it. 8892 if (_bit_map->par_mark(addr)) { 8893 bool res = _work_queue->push(obj); 8894 assert(res, "Low water mark should be much less than capacity"); 8895 // Do a recursive trim in the hope that this will keep 8896 // stack usage lower, but leave some oops for potential stealers 8897 trim_queue(_low_water_mark); 8898 } // Else, another thread got there first 8899 } 8900 } 8901 8902 void CMSParKeepAliveClosure::do_oop(oop* p) { CMSParKeepAliveClosure::do_oop_work(p); } 8903 void CMSParKeepAliveClosure::do_oop(narrowOop* p) { CMSParKeepAliveClosure::do_oop_work(p); } 8904 8905 void CMSParKeepAliveClosure::trim_queue(uint max) { 8906 while (_work_queue->size() > max) { 8907 oop new_oop; 8908 if (_work_queue->pop_local(new_oop)) { 8909 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 8910 assert(_bit_map->isMarked((HeapWord*)new_oop), 8911 "no white objects on this stack!"); 8912 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 8913 // iterate over the oops in this oop, marking and pushing 8914 // the ones in CMS heap (i.e. in _span). 8915 new_oop->oop_iterate(&_mark_and_push); 8916 } 8917 } 8918 } 8919 8920 CMSInnerParMarkAndPushClosure::CMSInnerParMarkAndPushClosure( 8921 CMSCollector* collector, 8922 MemRegion span, CMSBitMap* bit_map, 8923 OopTaskQueue* work_queue): 8924 _collector(collector), 8925 _span(span), 8926 _bit_map(bit_map), 8927 _work_queue(work_queue) { } 8928 8929 void CMSInnerParMarkAndPushClosure::do_oop(oop obj) { 8930 HeapWord* addr = (HeapWord*)obj; 8931 if (_span.contains(addr) && 8932 !_bit_map->isMarked(addr)) { 8933 if (_bit_map->par_mark(addr)) { 8934 bool simulate_overflow = false; 8935 NOT_PRODUCT( 8936 if (CMSMarkStackOverflowALot && 8937 _collector->par_simulate_overflow()) { 8938 // simulate a stack overflow 8939 simulate_overflow = true; 8940 } 8941 ) 8942 if (simulate_overflow || !_work_queue->push(obj)) { 8943 _collector->par_push_on_overflow_list(obj); 8944 _collector->_par_kac_ovflw++; 8945 } 8946 } // Else another thread got there already 8947 } 8948 } 8949 8950 void CMSInnerParMarkAndPushClosure::do_oop(oop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); } 8951 void CMSInnerParMarkAndPushClosure::do_oop(narrowOop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); } 8952 8953 ////////////////////////////////////////////////////////////////// 8954 // CMSExpansionCause ///////////////////////////// 8955 ////////////////////////////////////////////////////////////////// 8956 const char* CMSExpansionCause::to_string(CMSExpansionCause::Cause cause) { 8957 switch (cause) { 8958 case _no_expansion: 8959 return "No expansion"; 8960 case _satisfy_free_ratio: 8961 return "Free ratio"; 8962 case _satisfy_promotion: 8963 return "Satisfy promotion"; 8964 case _satisfy_allocation: 8965 return "allocation"; 8966 case _allocate_par_lab: 8967 return "Par LAB"; 8968 case _allocate_par_spooling_space: 8969 return "Par Spooling Space"; 8970 case _adaptive_size_policy: 8971 return "Ergonomics"; 8972 default: 8973 return "unknown"; 8974 } 8975 } 8976 8977 void CMSDrainMarkingStackClosure::do_void() { 8978 // the max number to take from overflow list at a time 8979 const size_t num = _mark_stack->capacity()/4; 8980 assert(!_concurrent_precleaning || _collector->overflow_list_is_empty(), 8981 "Overflow list should be NULL during concurrent phases"); 8982 while (!_mark_stack->isEmpty() || 8983 // if stack is empty, check the overflow list 8984 _collector->take_from_overflow_list(num, _mark_stack)) { 8985 oop obj = _mark_stack->pop(); 8986 HeapWord* addr = (HeapWord*)obj; 8987 assert(_span.contains(addr), "Should be within span"); 8988 assert(_bit_map->isMarked(addr), "Should be marked"); 8989 assert(obj->is_oop(), "Should be an oop"); 8990 obj->oop_iterate(_keep_alive); 8991 } 8992 } 8993 8994 void CMSParDrainMarkingStackClosure::do_void() { 8995 // drain queue 8996 trim_queue(0); 8997 } 8998 8999 // Trim our work_queue so its length is below max at return 9000 void CMSParDrainMarkingStackClosure::trim_queue(uint max) { 9001 while (_work_queue->size() > max) { 9002 oop new_oop; 9003 if (_work_queue->pop_local(new_oop)) { 9004 assert(new_oop->is_oop(), "Expected an oop"); 9005 assert(_bit_map->isMarked((HeapWord*)new_oop), 9006 "no white objects on this stack!"); 9007 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 9008 // iterate over the oops in this oop, marking and pushing 9009 // the ones in CMS heap (i.e. in _span). 9010 new_oop->oop_iterate(&_mark_and_push); 9011 } 9012 } 9013 } 9014 9015 //////////////////////////////////////////////////////////////////// 9016 // Support for Marking Stack Overflow list handling and related code 9017 //////////////////////////////////////////////////////////////////// 9018 // Much of the following code is similar in shape and spirit to the 9019 // code used in ParNewGC. We should try and share that code 9020 // as much as possible in the future. 9021 9022 #ifndef PRODUCT 9023 // Debugging support for CMSStackOverflowALot 9024 9025 // It's OK to call this multi-threaded; the worst thing 9026 // that can happen is that we'll get a bunch of closely 9027 // spaced simulated overflows, but that's OK, in fact 9028 // probably good as it would exercise the overflow code 9029 // under contention. 9030 bool CMSCollector::simulate_overflow() { 9031 if (_overflow_counter-- <= 0) { // just being defensive 9032 _overflow_counter = CMSMarkStackOverflowInterval; 9033 return true; 9034 } else { 9035 return false; 9036 } 9037 } 9038 9039 bool CMSCollector::par_simulate_overflow() { 9040 return simulate_overflow(); 9041 } 9042 #endif 9043 9044 // Single-threaded 9045 bool CMSCollector::take_from_overflow_list(size_t num, CMSMarkStack* stack) { 9046 assert(stack->isEmpty(), "Expected precondition"); 9047 assert(stack->capacity() > num, "Shouldn't bite more than can chew"); 9048 size_t i = num; 9049 oop cur = _overflow_list; 9050 const markOop proto = markOopDesc::prototype(); 9051 NOT_PRODUCT(ssize_t n = 0;) 9052 for (oop next; i > 0 && cur != NULL; cur = next, i--) { 9053 next = oop(cur->mark()); 9054 cur->set_mark(proto); // until proven otherwise 9055 assert(cur->is_oop(), "Should be an oop"); 9056 bool res = stack->push(cur); 9057 assert(res, "Bit off more than can chew?"); 9058 NOT_PRODUCT(n++;) 9059 } 9060 _overflow_list = cur; 9061 #ifndef PRODUCT 9062 assert(_num_par_pushes >= n, "Too many pops?"); 9063 _num_par_pushes -=n; 9064 #endif 9065 return !stack->isEmpty(); 9066 } 9067 9068 #define BUSY (cast_to_oop<intptr_t>(0x1aff1aff)) 9069 // (MT-safe) Get a prefix of at most "num" from the list. 9070 // The overflow list is chained through the mark word of 9071 // each object in the list. We fetch the entire list, 9072 // break off a prefix of the right size and return the 9073 // remainder. If other threads try to take objects from 9074 // the overflow list at that time, they will wait for 9075 // some time to see if data becomes available. If (and 9076 // only if) another thread places one or more object(s) 9077 // on the global list before we have returned the suffix 9078 // to the global list, we will walk down our local list 9079 // to find its end and append the global list to 9080 // our suffix before returning it. This suffix walk can 9081 // prove to be expensive (quadratic in the amount of traffic) 9082 // when there are many objects in the overflow list and 9083 // there is much producer-consumer contention on the list. 9084 // *NOTE*: The overflow list manipulation code here and 9085 // in ParNewGeneration:: are very similar in shape, 9086 // except that in the ParNew case we use the old (from/eden) 9087 // copy of the object to thread the list via its klass word. 9088 // Because of the common code, if you make any changes in 9089 // the code below, please check the ParNew version to see if 9090 // similar changes might be needed. 9091 // CR 6797058 has been filed to consolidate the common code. 9092 bool CMSCollector::par_take_from_overflow_list(size_t num, 9093 OopTaskQueue* work_q, 9094 int no_of_gc_threads) { 9095 assert(work_q->size() == 0, "First empty local work queue"); 9096 assert(num < work_q->max_elems(), "Can't bite more than we can chew"); 9097 if (_overflow_list == NULL) { 9098 return false; 9099 } 9100 // Grab the entire list; we'll put back a suffix 9101 oop prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list)); 9102 Thread* tid = Thread::current(); 9103 // Before "no_of_gc_threads" was introduced CMSOverflowSpinCount was 9104 // set to ParallelGCThreads. 9105 size_t CMSOverflowSpinCount = (size_t) no_of_gc_threads; // was ParallelGCThreads; 9106 size_t sleep_time_millis = MAX2((size_t)1, num/100); 9107 // If the list is busy, we spin for a short while, 9108 // sleeping between attempts to get the list. 9109 for (size_t spin = 0; prefix == BUSY && spin < CMSOverflowSpinCount; spin++) { 9110 os::sleep(tid, sleep_time_millis, false); 9111 if (_overflow_list == NULL) { 9112 // Nothing left to take 9113 return false; 9114 } else if (_overflow_list != BUSY) { 9115 // Try and grab the prefix 9116 prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list)); 9117 } 9118 } 9119 // If the list was found to be empty, or we spun long 9120 // enough, we give up and return empty-handed. If we leave 9121 // the list in the BUSY state below, it must be the case that 9122 // some other thread holds the overflow list and will set it 9123 // to a non-BUSY state in the future. 9124 if (prefix == NULL || prefix == BUSY) { 9125 // Nothing to take or waited long enough 9126 if (prefix == NULL) { 9127 // Write back the NULL in case we overwrote it with BUSY above 9128 // and it is still the same value. 9129 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY); 9130 } 9131 return false; 9132 } 9133 assert(prefix != NULL && prefix != BUSY, "Error"); 9134 size_t i = num; 9135 oop cur = prefix; 9136 // Walk down the first "num" objects, unless we reach the end. 9137 for (; i > 1 && cur->mark() != NULL; cur = oop(cur->mark()), i--); 9138 if (cur->mark() == NULL) { 9139 // We have "num" or fewer elements in the list, so there 9140 // is nothing to return to the global list. 9141 // Write back the NULL in lieu of the BUSY we wrote 9142 // above, if it is still the same value. 9143 if (_overflow_list == BUSY) { 9144 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY); 9145 } 9146 } else { 9147 // Chop off the suffix and return it to the global list. 9148 assert(cur->mark() != BUSY, "Error"); 9149 oop suffix_head = cur->mark(); // suffix will be put back on global list 9150 cur->set_mark(NULL); // break off suffix 9151 // It's possible that the list is still in the empty(busy) state 9152 // we left it in a short while ago; in that case we may be 9153 // able to place back the suffix without incurring the cost 9154 // of a walk down the list. 9155 oop observed_overflow_list = _overflow_list; 9156 oop cur_overflow_list = observed_overflow_list; 9157 bool attached = false; 9158 while (observed_overflow_list == BUSY || observed_overflow_list == NULL) { 9159 observed_overflow_list = 9160 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list); 9161 if (cur_overflow_list == observed_overflow_list) { 9162 attached = true; 9163 break; 9164 } else cur_overflow_list = observed_overflow_list; 9165 } 9166 if (!attached) { 9167 // Too bad, someone else sneaked in (at least) an element; we'll need 9168 // to do a splice. Find tail of suffix so we can prepend suffix to global 9169 // list. 9170 for (cur = suffix_head; cur->mark() != NULL; cur = (oop)(cur->mark())); 9171 oop suffix_tail = cur; 9172 assert(suffix_tail != NULL && suffix_tail->mark() == NULL, 9173 "Tautology"); 9174 observed_overflow_list = _overflow_list; 9175 do { 9176 cur_overflow_list = observed_overflow_list; 9177 if (cur_overflow_list != BUSY) { 9178 // Do the splice ... 9179 suffix_tail->set_mark(markOop(cur_overflow_list)); 9180 } else { // cur_overflow_list == BUSY 9181 suffix_tail->set_mark(NULL); 9182 } 9183 // ... and try to place spliced list back on overflow_list ... 9184 observed_overflow_list = 9185 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list); 9186 } while (cur_overflow_list != observed_overflow_list); 9187 // ... until we have succeeded in doing so. 9188 } 9189 } 9190 9191 // Push the prefix elements on work_q 9192 assert(prefix != NULL, "control point invariant"); 9193 const markOop proto = markOopDesc::prototype(); 9194 oop next; 9195 NOT_PRODUCT(ssize_t n = 0;) 9196 for (cur = prefix; cur != NULL; cur = next) { 9197 next = oop(cur->mark()); 9198 cur->set_mark(proto); // until proven otherwise 9199 assert(cur->is_oop(), "Should be an oop"); 9200 bool res = work_q->push(cur); 9201 assert(res, "Bit off more than we can chew?"); 9202 NOT_PRODUCT(n++;) 9203 } 9204 #ifndef PRODUCT 9205 assert(_num_par_pushes >= n, "Too many pops?"); 9206 Atomic::add_ptr(-(intptr_t)n, &_num_par_pushes); 9207 #endif 9208 return true; 9209 } 9210 9211 // Single-threaded 9212 void CMSCollector::push_on_overflow_list(oop p) { 9213 NOT_PRODUCT(_num_par_pushes++;) 9214 assert(p->is_oop(), "Not an oop"); 9215 preserve_mark_if_necessary(p); 9216 p->set_mark((markOop)_overflow_list); 9217 _overflow_list = p; 9218 } 9219 9220 // Multi-threaded; use CAS to prepend to overflow list 9221 void CMSCollector::par_push_on_overflow_list(oop p) { 9222 NOT_PRODUCT(Atomic::inc_ptr(&_num_par_pushes);) 9223 assert(p->is_oop(), "Not an oop"); 9224 par_preserve_mark_if_necessary(p); 9225 oop observed_overflow_list = _overflow_list; 9226 oop cur_overflow_list; 9227 do { 9228 cur_overflow_list = observed_overflow_list; 9229 if (cur_overflow_list != BUSY) { 9230 p->set_mark(markOop(cur_overflow_list)); 9231 } else { 9232 p->set_mark(NULL); 9233 } 9234 observed_overflow_list = 9235 (oop) Atomic::cmpxchg_ptr(p, &_overflow_list, cur_overflow_list); 9236 } while (cur_overflow_list != observed_overflow_list); 9237 } 9238 #undef BUSY 9239 9240 // Single threaded 9241 // General Note on GrowableArray: pushes may silently fail 9242 // because we are (temporarily) out of C-heap for expanding 9243 // the stack. The problem is quite ubiquitous and affects 9244 // a lot of code in the JVM. The prudent thing for GrowableArray 9245 // to do (for now) is to exit with an error. However, that may 9246 // be too draconian in some cases because the caller may be 9247 // able to recover without much harm. For such cases, we 9248 // should probably introduce a "soft_push" method which returns 9249 // an indication of success or failure with the assumption that 9250 // the caller may be able to recover from a failure; code in 9251 // the VM can then be changed, incrementally, to deal with such 9252 // failures where possible, thus, incrementally hardening the VM 9253 // in such low resource situations. 9254 void CMSCollector::preserve_mark_work(oop p, markOop m) { 9255 _preserved_oop_stack.push(p); 9256 _preserved_mark_stack.push(m); 9257 assert(m == p->mark(), "Mark word changed"); 9258 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), 9259 "bijection"); 9260 } 9261 9262 // Single threaded 9263 void CMSCollector::preserve_mark_if_necessary(oop p) { 9264 markOop m = p->mark(); 9265 if (m->must_be_preserved(p)) { 9266 preserve_mark_work(p, m); 9267 } 9268 } 9269 9270 void CMSCollector::par_preserve_mark_if_necessary(oop p) { 9271 markOop m = p->mark(); 9272 if (m->must_be_preserved(p)) { 9273 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag); 9274 // Even though we read the mark word without holding 9275 // the lock, we are assured that it will not change 9276 // because we "own" this oop, so no other thread can 9277 // be trying to push it on the overflow list; see 9278 // the assertion in preserve_mark_work() that checks 9279 // that m == p->mark(). 9280 preserve_mark_work(p, m); 9281 } 9282 } 9283 9284 // We should be able to do this multi-threaded, 9285 // a chunk of stack being a task (this is 9286 // correct because each oop only ever appears 9287 // once in the overflow list. However, it's 9288 // not very easy to completely overlap this with 9289 // other operations, so will generally not be done 9290 // until all work's been completed. Because we 9291 // expect the preserved oop stack (set) to be small, 9292 // it's probably fine to do this single-threaded. 9293 // We can explore cleverer concurrent/overlapped/parallel 9294 // processing of preserved marks if we feel the 9295 // need for this in the future. Stack overflow should 9296 // be so rare in practice and, when it happens, its 9297 // effect on performance so great that this will 9298 // likely just be in the noise anyway. 9299 void CMSCollector::restore_preserved_marks_if_any() { 9300 assert(SafepointSynchronize::is_at_safepoint(), 9301 "world should be stopped"); 9302 assert(Thread::current()->is_ConcurrentGC_thread() || 9303 Thread::current()->is_VM_thread(), 9304 "should be single-threaded"); 9305 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), 9306 "bijection"); 9307 9308 while (!_preserved_oop_stack.is_empty()) { 9309 oop p = _preserved_oop_stack.pop(); 9310 assert(p->is_oop(), "Should be an oop"); 9311 assert(_span.contains(p), "oop should be in _span"); 9312 assert(p->mark() == markOopDesc::prototype(), 9313 "Set when taken from overflow list"); 9314 markOop m = _preserved_mark_stack.pop(); 9315 p->set_mark(m); 9316 } 9317 assert(_preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(), 9318 "stacks were cleared above"); 9319 } 9320 9321 #ifndef PRODUCT 9322 bool CMSCollector::no_preserved_marks() const { 9323 return _preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(); 9324 } 9325 #endif 9326 9327 CMSAdaptiveSizePolicy* ASConcurrentMarkSweepGeneration::cms_size_policy() const 9328 { 9329 GenCollectedHeap* gch = (GenCollectedHeap*) GenCollectedHeap::heap(); 9330 CMSAdaptiveSizePolicy* size_policy = 9331 (CMSAdaptiveSizePolicy*) gch->gen_policy()->size_policy(); 9332 assert(size_policy->is_gc_cms_adaptive_size_policy(), 9333 "Wrong type for size policy"); 9334 return size_policy; 9335 } 9336 9337 void ASConcurrentMarkSweepGeneration::resize(size_t cur_promo_size, 9338 size_t desired_promo_size) { 9339 if (cur_promo_size < desired_promo_size) { 9340 size_t expand_bytes = desired_promo_size - cur_promo_size; 9341 if (PrintAdaptiveSizePolicy && Verbose) { 9342 gclog_or_tty->print_cr(" ASConcurrentMarkSweepGeneration::resize " 9343 "Expanding tenured generation by " SIZE_FORMAT " (bytes)", 9344 expand_bytes); 9345 } 9346 expand(expand_bytes, 9347 MinHeapDeltaBytes, 9348 CMSExpansionCause::_adaptive_size_policy); 9349 } else if (desired_promo_size < cur_promo_size) { 9350 size_t shrink_bytes = cur_promo_size - desired_promo_size; 9351 if (PrintAdaptiveSizePolicy && Verbose) { 9352 gclog_or_tty->print_cr(" ASConcurrentMarkSweepGeneration::resize " 9353 "Shrinking tenured generation by " SIZE_FORMAT " (bytes)", 9354 shrink_bytes); 9355 } 9356 shrink(shrink_bytes); 9357 } 9358 } 9359 9360 CMSGCAdaptivePolicyCounters* ASConcurrentMarkSweepGeneration::gc_adaptive_policy_counters() { 9361 GenCollectedHeap* gch = GenCollectedHeap::heap(); 9362 CMSGCAdaptivePolicyCounters* counters = 9363 (CMSGCAdaptivePolicyCounters*) gch->collector_policy()->counters(); 9364 assert(counters->kind() == GCPolicyCounters::CMSGCAdaptivePolicyCountersKind, 9365 "Wrong kind of counters"); 9366 return counters; 9367 } 9368 9369 9370 void ASConcurrentMarkSweepGeneration::update_counters() { 9371 if (UsePerfData) { 9372 _space_counters->update_all(); 9373 _gen_counters->update_all(); 9374 CMSGCAdaptivePolicyCounters* counters = gc_adaptive_policy_counters(); 9375 GenCollectedHeap* gch = GenCollectedHeap::heap(); 9376 CMSGCStats* gc_stats_l = (CMSGCStats*) gc_stats(); 9377 assert(gc_stats_l->kind() == GCStats::CMSGCStatsKind, 9378 "Wrong gc statistics type"); 9379 counters->update_counters(gc_stats_l); 9380 } 9381 } 9382 9383 void ASConcurrentMarkSweepGeneration::update_counters(size_t used) { 9384 if (UsePerfData) { 9385 _space_counters->update_used(used); 9386 _space_counters->update_capacity(); 9387 _gen_counters->update_all(); 9388 9389 CMSGCAdaptivePolicyCounters* counters = gc_adaptive_policy_counters(); 9390 GenCollectedHeap* gch = GenCollectedHeap::heap(); 9391 CMSGCStats* gc_stats_l = (CMSGCStats*) gc_stats(); 9392 assert(gc_stats_l->kind() == GCStats::CMSGCStatsKind, 9393 "Wrong gc statistics type"); 9394 counters->update_counters(gc_stats_l); 9395 } 9396 } 9397 9398 void ASConcurrentMarkSweepGeneration::shrink_by(size_t desired_bytes) { 9399 assert_locked_or_safepoint(Heap_lock); 9400 assert_lock_strong(freelistLock()); 9401 HeapWord* old_end = _cmsSpace->end(); 9402 HeapWord* unallocated_start = _cmsSpace->unallocated_block(); 9403 assert(old_end >= unallocated_start, "Miscalculation of unallocated_start"); 9404 FreeChunk* chunk_at_end = find_chunk_at_end(); 9405 if (chunk_at_end == NULL) { 9406 // No room to shrink 9407 if (PrintGCDetails && Verbose) { 9408 gclog_or_tty->print_cr("No room to shrink: old_end " 9409 PTR_FORMAT " unallocated_start " PTR_FORMAT 9410 " chunk_at_end " PTR_FORMAT, 9411 old_end, unallocated_start, chunk_at_end); 9412 } 9413 return; 9414 } else { 9415 9416 // Find the chunk at the end of the space and determine 9417 // how much it can be shrunk. 9418 size_t shrinkable_size_in_bytes = chunk_at_end->size(); 9419 size_t aligned_shrinkable_size_in_bytes = 9420 align_size_down(shrinkable_size_in_bytes, os::vm_page_size()); 9421 assert(unallocated_start <= (HeapWord*) chunk_at_end->end(), 9422 "Inconsistent chunk at end of space"); 9423 size_t bytes = MIN2(desired_bytes, aligned_shrinkable_size_in_bytes); 9424 size_t word_size_before = heap_word_size(_virtual_space.committed_size()); 9425 9426 // Shrink the underlying space 9427 _virtual_space.shrink_by(bytes); 9428 if (PrintGCDetails && Verbose) { 9429 gclog_or_tty->print_cr("ConcurrentMarkSweepGeneration::shrink_by:" 9430 " desired_bytes " SIZE_FORMAT 9431 " shrinkable_size_in_bytes " SIZE_FORMAT 9432 " aligned_shrinkable_size_in_bytes " SIZE_FORMAT 9433 " bytes " SIZE_FORMAT, 9434 desired_bytes, shrinkable_size_in_bytes, 9435 aligned_shrinkable_size_in_bytes, bytes); 9436 gclog_or_tty->print_cr(" old_end " SIZE_FORMAT 9437 " unallocated_start " SIZE_FORMAT, 9438 old_end, unallocated_start); 9439 } 9440 9441 // If the space did shrink (shrinking is not guaranteed), 9442 // shrink the chunk at the end by the appropriate amount. 9443 if (((HeapWord*)_virtual_space.high()) < old_end) { 9444 size_t new_word_size = 9445 heap_word_size(_virtual_space.committed_size()); 9446 9447 // Have to remove the chunk from the dictionary because it is changing 9448 // size and might be someplace elsewhere in the dictionary. 9449 9450 // Get the chunk at end, shrink it, and put it 9451 // back. 9452 _cmsSpace->removeChunkFromDictionary(chunk_at_end); 9453 size_t word_size_change = word_size_before - new_word_size; 9454 size_t chunk_at_end_old_size = chunk_at_end->size(); 9455 assert(chunk_at_end_old_size >= word_size_change, 9456 "Shrink is too large"); 9457 chunk_at_end->set_size(chunk_at_end_old_size - 9458 word_size_change); 9459 _cmsSpace->freed((HeapWord*) chunk_at_end->end(), 9460 word_size_change); 9461 9462 _cmsSpace->returnChunkToDictionary(chunk_at_end); 9463 9464 MemRegion mr(_cmsSpace->bottom(), new_word_size); 9465 _bts->resize(new_word_size); // resize the block offset shared array 9466 Universe::heap()->barrier_set()->resize_covered_region(mr); 9467 _cmsSpace->assert_locked(); 9468 _cmsSpace->set_end((HeapWord*)_virtual_space.high()); 9469 9470 NOT_PRODUCT(_cmsSpace->dictionary()->verify()); 9471 9472 // update the space and generation capacity counters 9473 if (UsePerfData) { 9474 _space_counters->update_capacity(); 9475 _gen_counters->update_all(); 9476 } 9477 9478 if (Verbose && PrintGCDetails) { 9479 size_t new_mem_size = _virtual_space.committed_size(); 9480 size_t old_mem_size = new_mem_size + bytes; 9481 gclog_or_tty->print_cr("Shrinking %s from " SIZE_FORMAT "K by " SIZE_FORMAT "K to " SIZE_FORMAT "K", 9482 name(), old_mem_size/K, bytes/K, new_mem_size/K); 9483 } 9484 } 9485 9486 assert(_cmsSpace->unallocated_block() <= _cmsSpace->end(), 9487 "Inconsistency at end of space"); 9488 assert(chunk_at_end->end() == (uintptr_t*) _cmsSpace->end(), 9489 "Shrinking is inconsistent"); 9490 return; 9491 } 9492 } 9493 // Transfer some number of overflown objects to usual marking 9494 // stack. Return true if some objects were transferred. 9495 bool MarkRefsIntoAndScanClosure::take_from_overflow_list() { 9496 size_t num = MIN2((size_t)(_mark_stack->capacity() - _mark_stack->length())/4, 9497 (size_t)ParGCDesiredObjsFromOverflowList); 9498 9499 bool res = _collector->take_from_overflow_list(num, _mark_stack); 9500 assert(_collector->overflow_list_is_empty() || res, 9501 "If list is not empty, we should have taken something"); 9502 assert(!res || !_mark_stack->isEmpty(), 9503 "If we took something, it should now be on our stack"); 9504 return res; 9505 } 9506 9507 size_t MarkDeadObjectsClosure::do_blk(HeapWord* addr) { 9508 size_t res = _sp->block_size_no_stall(addr, _collector); 9509 if (_sp->block_is_obj(addr)) { 9510 if (_live_bit_map->isMarked(addr)) { 9511 // It can't have been dead in a previous cycle 9512 guarantee(!_dead_bit_map->isMarked(addr), "No resurrection!"); 9513 } else { 9514 _dead_bit_map->mark(addr); // mark the dead object 9515 } 9516 } 9517 // Could be 0, if the block size could not be computed without stalling. 9518 return res; 9519 } 9520 9521 TraceCMSMemoryManagerStats::TraceCMSMemoryManagerStats(CMSCollector::CollectorState phase, GCCause::Cause cause): TraceMemoryManagerStats() { 9522 9523 switch (phase) { 9524 case CMSCollector::InitialMarking: 9525 initialize(true /* fullGC */ , 9526 cause /* cause of the GC */, 9527 true /* recordGCBeginTime */, 9528 true /* recordPreGCUsage */, 9529 false /* recordPeakUsage */, 9530 false /* recordPostGCusage */, 9531 true /* recordAccumulatedGCTime */, 9532 false /* recordGCEndTime */, 9533 false /* countCollection */ ); 9534 break; 9535 9536 case CMSCollector::FinalMarking: 9537 initialize(true /* fullGC */ , 9538 cause /* cause of the GC */, 9539 false /* recordGCBeginTime */, 9540 false /* recordPreGCUsage */, 9541 false /* recordPeakUsage */, 9542 false /* recordPostGCusage */, 9543 true /* recordAccumulatedGCTime */, 9544 false /* recordGCEndTime */, 9545 false /* countCollection */ ); 9546 break; 9547 9548 case CMSCollector::Sweeping: 9549 initialize(true /* fullGC */ , 9550 cause /* cause of the GC */, 9551 false /* recordGCBeginTime */, 9552 false /* recordPreGCUsage */, 9553 true /* recordPeakUsage */, 9554 true /* recordPostGCusage */, 9555 false /* recordAccumulatedGCTime */, 9556 true /* recordGCEndTime */, 9557 true /* countCollection */ ); 9558 break; 9559 9560 default: 9561 ShouldNotReachHere(); 9562 } 9563 }