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