1 /* 2 * Copyright (c) 2001, 2012, 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/symbolTable.hpp" 27 #include "classfile/systemDictionary.hpp" 28 #include "code/codeCache.hpp" 29 #include "gc_implementation/concurrentMarkSweep/cmsAdaptiveSizePolicy.hpp" 30 #include "gc_implementation/concurrentMarkSweep/cmsCollectorPolicy.hpp" 31 #include "gc_implementation/concurrentMarkSweep/cmsGCAdaptivePolicyCounters.hpp" 32 #include "gc_implementation/concurrentMarkSweep/cmsOopClosures.inline.hpp" 33 #include "gc_implementation/concurrentMarkSweep/compactibleFreeListSpace.hpp" 34 #include "gc_implementation/concurrentMarkSweep/concurrentMarkSweepGeneration.inline.hpp" 35 #include "gc_implementation/concurrentMarkSweep/concurrentMarkSweepThread.hpp" 36 #include "gc_implementation/concurrentMarkSweep/vmCMSOperations.hpp" 37 #include "gc_implementation/parNew/parNewGeneration.hpp" 38 #include "gc_implementation/shared/collectorCounters.hpp" 39 #include "gc_implementation/shared/gcTimer.hpp" 40 #include "gc_implementation/shared/gcTrace.hpp" 41 #include "gc_implementation/shared/gcTraceTime.hpp" 42 #include "gc_implementation/shared/isGCActiveMark.hpp" 43 #include "gc_interface/collectedHeap.inline.hpp" 44 #include "memory/allocation.hpp" 45 #include "memory/cardTableRS.hpp" 46 #include "memory/collectorPolicy.hpp" 47 #include "memory/gcLocker.inline.hpp" 48 #include "memory/genCollectedHeap.hpp" 49 #include "memory/genMarkSweep.hpp" 50 #include "memory/genOopClosures.inline.hpp" 51 #include "memory/iterator.hpp" 52 #include "memory/referencePolicy.hpp" 53 #include "memory/resourceArea.hpp" 54 #include "oops/oop.inline.hpp" 55 #include "prims/jvmtiExport.hpp" 56 #include "runtime/globals_extension.hpp" 57 #include "runtime/handles.inline.hpp" 58 #include "runtime/java.hpp" 59 #include "runtime/vmThread.hpp" 60 #include "services/memoryService.hpp" 61 #include "services/runtimeService.hpp" 62 63 // statics 64 CMSCollector* ConcurrentMarkSweepGeneration::_collector = NULL; 65 bool CMSCollector::_full_gc_requested = false; 66 67 ////////////////////////////////////////////////////////////////// 68 // In support of CMS/VM thread synchronization 69 ////////////////////////////////////////////////////////////////// 70 // We split use of the CGC_lock into 2 "levels". 71 // The low-level locking is of the usual CGC_lock monitor. We introduce 72 // a higher level "token" (hereafter "CMS token") built on top of the 73 // low level monitor (hereafter "CGC lock"). 74 // The token-passing protocol gives priority to the VM thread. The 75 // CMS-lock doesn't provide any fairness guarantees, but clients 76 // should ensure that it is only held for very short, bounded 77 // durations. 78 // 79 // When either of the CMS thread or the VM thread is involved in 80 // collection operations during which it does not want the other 81 // thread to interfere, it obtains the CMS token. 82 // 83 // If either thread tries to get the token while the other has 84 // it, that thread waits. However, if the VM thread and CMS thread 85 // both want the token, then the VM thread gets priority while the 86 // CMS thread waits. This ensures, for instance, that the "concurrent" 87 // phases of the CMS thread's work do not block out the VM thread 88 // for long periods of time as the CMS thread continues to hog 89 // the token. (See bug 4616232). 90 // 91 // The baton-passing functions are, however, controlled by the 92 // flags _foregroundGCShouldWait and _foregroundGCIsActive, 93 // and here the low-level CMS lock, not the high level token, 94 // ensures mutual exclusion. 95 // 96 // Two important conditions that we have to satisfy: 97 // 1. if a thread does a low-level wait on the CMS lock, then it 98 // relinquishes the CMS token if it were holding that token 99 // when it acquired the low-level CMS lock. 100 // 2. any low-level notifications on the low-level lock 101 // should only be sent when a thread has relinquished the token. 102 // 103 // In the absence of either property, we'd have potential deadlock. 104 // 105 // We protect each of the CMS (concurrent and sequential) phases 106 // with the CMS _token_, not the CMS _lock_. 107 // 108 // The only code protected by CMS lock is the token acquisition code 109 // itself, see ConcurrentMarkSweepThread::[de]synchronize(), and the 110 // baton-passing code. 111 // 112 // Unfortunately, i couldn't come up with a good abstraction to factor and 113 // hide the naked CGC_lock manipulation in the baton-passing code 114 // further below. That's something we should try to do. Also, the proof 115 // of correctness of this 2-level locking scheme is far from obvious, 116 // and potentially quite slippery. We have an uneasy supsicion, for instance, 117 // that there may be a theoretical possibility of delay/starvation in the 118 // low-level lock/wait/notify scheme used for the baton-passing because of 119 // potential intereference with the priority scheme embodied in the 120 // CMS-token-passing protocol. See related comments at a CGC_lock->wait() 121 // invocation further below and marked with "XXX 20011219YSR". 122 // Indeed, as we note elsewhere, this may become yet more slippery 123 // in the presence of multiple CMS and/or multiple VM threads. XXX 124 125 class CMSTokenSync: public StackObj { 126 private: 127 bool _is_cms_thread; 128 public: 129 CMSTokenSync(bool is_cms_thread): 130 _is_cms_thread(is_cms_thread) { 131 assert(is_cms_thread == Thread::current()->is_ConcurrentGC_thread(), 132 "Incorrect argument to constructor"); 133 ConcurrentMarkSweepThread::synchronize(_is_cms_thread); 134 } 135 136 ~CMSTokenSync() { 137 assert(_is_cms_thread ? 138 ConcurrentMarkSweepThread::cms_thread_has_cms_token() : 139 ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 140 "Incorrect state"); 141 ConcurrentMarkSweepThread::desynchronize(_is_cms_thread); 142 } 143 }; 144 145 // Convenience class that does a CMSTokenSync, and then acquires 146 // upto three locks. 147 class CMSTokenSyncWithLocks: public CMSTokenSync { 148 private: 149 // Note: locks are acquired in textual declaration order 150 // and released in the opposite order 151 MutexLockerEx _locker1, _locker2, _locker3; 152 public: 153 CMSTokenSyncWithLocks(bool is_cms_thread, Mutex* mutex1, 154 Mutex* mutex2 = NULL, Mutex* mutex3 = NULL): 155 CMSTokenSync(is_cms_thread), 156 _locker1(mutex1, Mutex::_no_safepoint_check_flag), 157 _locker2(mutex2, Mutex::_no_safepoint_check_flag), 158 _locker3(mutex3, Mutex::_no_safepoint_check_flag) 159 { } 160 }; 161 162 163 // Wrapper class to temporarily disable icms during a foreground cms collection. 164 class ICMSDisabler: public StackObj { 165 public: 166 // The ctor disables icms and wakes up the thread so it notices the change; 167 // the dtor re-enables icms. Note that the CMSCollector methods will check 168 // CMSIncrementalMode. 169 ICMSDisabler() { CMSCollector::disable_icms(); CMSCollector::start_icms(); } 170 ~ICMSDisabler() { CMSCollector::enable_icms(); } 171 }; 172 173 ////////////////////////////////////////////////////////////////// 174 // Concurrent Mark-Sweep Generation ///////////////////////////// 175 ////////////////////////////////////////////////////////////////// 176 177 NOT_PRODUCT(CompactibleFreeListSpace* debug_cms_space;) 178 179 // This struct contains per-thread things necessary to support parallel 180 // young-gen collection. 181 class CMSParGCThreadState: public CHeapObj<mtGC> { 182 public: 183 CFLS_LAB lab; 184 PromotionInfo promo; 185 186 // Constructor. 187 CMSParGCThreadState(CompactibleFreeListSpace* cfls) : lab(cfls) { 188 promo.setSpace(cfls); 189 } 190 }; 191 192 ConcurrentMarkSweepGeneration::ConcurrentMarkSweepGeneration( 193 ReservedSpace rs, size_t initial_byte_size, int level, 194 CardTableRS* ct, bool use_adaptive_freelists, 195 FreeBlockDictionary<FreeChunk>::DictionaryChoice dictionaryChoice) : 196 CardGeneration(rs, initial_byte_size, level, ct), 197 _dilatation_factor(((double)MinChunkSize)/((double)(CollectedHeap::min_fill_size()))), 198 _debug_collection_type(Concurrent_collection_type) 199 { 200 HeapWord* bottom = (HeapWord*) _virtual_space.low(); 201 HeapWord* end = (HeapWord*) _virtual_space.high(); 202 203 _direct_allocated_words = 0; 204 NOT_PRODUCT( 205 _numObjectsPromoted = 0; 206 _numWordsPromoted = 0; 207 _numObjectsAllocated = 0; 208 _numWordsAllocated = 0; 209 ) 210 211 _cmsSpace = new CompactibleFreeListSpace(_bts, MemRegion(bottom, end), 212 use_adaptive_freelists, 213 dictionaryChoice); 214 NOT_PRODUCT(debug_cms_space = _cmsSpace;) 215 if (_cmsSpace == NULL) { 216 vm_exit_during_initialization( 217 "CompactibleFreeListSpace allocation failure"); 218 } 219 _cmsSpace->_gen = this; 220 221 _gc_stats = new CMSGCStats(); 222 223 // Verify the assumption that FreeChunk::_prev and OopDesc::_klass 224 // offsets match. The ability to tell free chunks from objects 225 // depends on this property. 226 debug_only( 227 FreeChunk* junk = NULL; 228 assert(UseCompressedOops || 229 junk->prev_addr() == (void*)(oop(junk)->klass_addr()), 230 "Offset of FreeChunk::_prev within FreeChunk must match" 231 " that of OopDesc::_klass within OopDesc"); 232 ) 233 if (CollectedHeap::use_parallel_gc_threads()) { 234 typedef CMSParGCThreadState* CMSParGCThreadStatePtr; 235 _par_gc_thread_states = 236 NEW_C_HEAP_ARRAY(CMSParGCThreadStatePtr, ParallelGCThreads, mtGC); 237 if (_par_gc_thread_states == NULL) { 238 vm_exit_during_initialization("Could not allocate par gc structs"); 239 } 240 for (uint i = 0; i < ParallelGCThreads; i++) { 241 _par_gc_thread_states[i] = new CMSParGCThreadState(cmsSpace()); 242 if (_par_gc_thread_states[i] == NULL) { 243 vm_exit_during_initialization("Could not allocate par gc structs"); 244 } 245 } 246 } else { 247 _par_gc_thread_states = NULL; 248 } 249 _incremental_collection_failed = false; 250 // The "dilatation_factor" is the expansion that can occur on 251 // account of the fact that the minimum object size in the CMS 252 // generation may be larger than that in, say, a contiguous young 253 // generation. 254 // Ideally, in the calculation below, we'd compute the dilatation 255 // factor as: MinChunkSize/(promoting_gen's min object size) 256 // Since we do not have such a general query interface for the 257 // promoting generation, we'll instead just use the mimimum 258 // object size (which today is a header's worth of space); 259 // note that all arithmetic is in units of HeapWords. 260 assert(MinChunkSize >= CollectedHeap::min_fill_size(), "just checking"); 261 assert(_dilatation_factor >= 1.0, "from previous assert"); 262 } 263 264 265 // The field "_initiating_occupancy" represents the occupancy percentage 266 // at which we trigger a new collection cycle. Unless explicitly specified 267 // via CMSInitiating[Perm]OccupancyFraction (argument "io" below), it 268 // is calculated by: 269 // 270 // Let "f" be MinHeapFreeRatio in 271 // 272 // _intiating_occupancy = 100-f + 273 // f * (CMSTrigger[Perm]Ratio/100) 274 // where CMSTrigger[Perm]Ratio is the argument "tr" below. 275 // 276 // That is, if we assume the heap is at its desired maximum occupancy at the 277 // end of a collection, we let CMSTrigger[Perm]Ratio of the (purported) free 278 // space be allocated before initiating a new collection cycle. 279 // 280 void ConcurrentMarkSweepGeneration::init_initiating_occupancy(intx io, intx tr) { 281 assert(io <= 100 && tr >= 0 && tr <= 100, "Check the arguments"); 282 if (io >= 0) { 283 _initiating_occupancy = (double)io / 100.0; 284 } else { 285 _initiating_occupancy = ((100 - MinHeapFreeRatio) + 286 (double)(tr * MinHeapFreeRatio) / 100.0) 287 / 100.0; 288 } 289 } 290 291 void ConcurrentMarkSweepGeneration::ref_processor_init() { 292 assert(collector() != NULL, "no collector"); 293 collector()->ref_processor_init(); 294 } 295 296 void CMSCollector::ref_processor_init() { 297 if (_ref_processor == NULL) { 298 // Allocate and initialize a reference processor 299 _ref_processor = 300 new ReferenceProcessor(_span, // span 301 (ParallelGCThreads > 1) && ParallelRefProcEnabled, // mt processing 302 (int) ParallelGCThreads, // mt processing degree 303 _cmsGen->refs_discovery_is_mt(), // mt discovery 304 (int) MAX2(ConcGCThreads, ParallelGCThreads), // mt discovery degree 305 _cmsGen->refs_discovery_is_atomic(), // discovery is not atomic 306 &_is_alive_closure, // closure for liveness info 307 false); // next field updates do not need write barrier 308 // Initialize the _ref_processor field of CMSGen 309 _cmsGen->set_ref_processor(_ref_processor); 310 311 // Allocate a dummy ref processor for perm gen. 312 ReferenceProcessor* rp2 = new ReferenceProcessor(); 313 if (rp2 == NULL) { 314 vm_exit_during_initialization("Could not allocate ReferenceProcessor object"); 315 } 316 _permGen->set_ref_processor(rp2); 317 } 318 } 319 320 CMSAdaptiveSizePolicy* CMSCollector::size_policy() { 321 GenCollectedHeap* gch = GenCollectedHeap::heap(); 322 assert(gch->kind() == CollectedHeap::GenCollectedHeap, 323 "Wrong type of heap"); 324 CMSAdaptiveSizePolicy* sp = (CMSAdaptiveSizePolicy*) 325 gch->gen_policy()->size_policy(); 326 assert(sp->is_gc_cms_adaptive_size_policy(), 327 "Wrong type of size policy"); 328 return sp; 329 } 330 331 CMSGCAdaptivePolicyCounters* CMSCollector::gc_adaptive_policy_counters() { 332 CMSGCAdaptivePolicyCounters* results = 333 (CMSGCAdaptivePolicyCounters*) collector_policy()->counters(); 334 assert( 335 results->kind() == GCPolicyCounters::CMSGCAdaptivePolicyCountersKind, 336 "Wrong gc policy counter kind"); 337 return results; 338 } 339 340 341 void ConcurrentMarkSweepGeneration::initialize_performance_counters() { 342 343 const char* gen_name = "old"; 344 345 // Generation Counters - generation 1, 1 subspace 346 _gen_counters = new GenerationCounters(gen_name, 1, 1, &_virtual_space); 347 348 _space_counters = new GSpaceCounters(gen_name, 0, 349 _virtual_space.reserved_size(), 350 this, _gen_counters); 351 } 352 353 CMSStats::CMSStats(ConcurrentMarkSweepGeneration* cms_gen, unsigned int alpha): 354 _cms_gen(cms_gen) 355 { 356 assert(alpha <= 100, "bad value"); 357 _saved_alpha = alpha; 358 359 // Initialize the alphas to the bootstrap value of 100. 360 _gc0_alpha = _cms_alpha = 100; 361 362 _cms_begin_time.update(); 363 _cms_end_time.update(); 364 365 _gc0_duration = 0.0; 366 _gc0_period = 0.0; 367 _gc0_promoted = 0; 368 369 _cms_duration = 0.0; 370 _cms_period = 0.0; 371 _cms_allocated = 0; 372 373 _cms_used_at_gc0_begin = 0; 374 _cms_used_at_gc0_end = 0; 375 _allow_duty_cycle_reduction = false; 376 _valid_bits = 0; 377 _icms_duty_cycle = CMSIncrementalDutyCycle; 378 } 379 380 double CMSStats::cms_free_adjustment_factor(size_t free) const { 381 // TBD: CR 6909490 382 return 1.0; 383 } 384 385 void CMSStats::adjust_cms_free_adjustment_factor(bool fail, size_t free) { 386 } 387 388 // If promotion failure handling is on use 389 // the padded average size of the promotion for each 390 // young generation collection. 391 double CMSStats::time_until_cms_gen_full() const { 392 size_t cms_free = _cms_gen->cmsSpace()->free(); 393 GenCollectedHeap* gch = GenCollectedHeap::heap(); 394 size_t expected_promotion = MIN2(gch->get_gen(0)->capacity(), 395 (size_t) _cms_gen->gc_stats()->avg_promoted()->padded_average()); 396 if (cms_free > expected_promotion) { 397 // Start a cms collection if there isn't enough space to promote 398 // for the next minor collection. Use the padded average as 399 // a safety factor. 400 cms_free -= expected_promotion; 401 402 // Adjust by the safety factor. 403 double cms_free_dbl = (double)cms_free; 404 double cms_adjustment = (100.0 - CMSIncrementalSafetyFactor)/100.0; 405 // Apply a further correction factor which tries to adjust 406 // for recent occurance of concurrent mode failures. 407 cms_adjustment = cms_adjustment * cms_free_adjustment_factor(cms_free); 408 cms_free_dbl = cms_free_dbl * cms_adjustment; 409 410 if (PrintGCDetails && Verbose) { 411 gclog_or_tty->print_cr("CMSStats::time_until_cms_gen_full: cms_free " 412 SIZE_FORMAT " expected_promotion " SIZE_FORMAT, 413 cms_free, expected_promotion); 414 gclog_or_tty->print_cr(" cms_free_dbl %f cms_consumption_rate %f", 415 cms_free_dbl, cms_consumption_rate() + 1.0); 416 } 417 // Add 1 in case the consumption rate goes to zero. 418 return cms_free_dbl / (cms_consumption_rate() + 1.0); 419 } 420 return 0.0; 421 } 422 423 // Compare the duration of the cms collection to the 424 // time remaining before the cms generation is empty. 425 // Note that the time from the start of the cms collection 426 // to the start of the cms sweep (less than the total 427 // duration of the cms collection) can be used. This 428 // has been tried and some applications experienced 429 // promotion failures early in execution. This was 430 // possibly because the averages were not accurate 431 // enough at the beginning. 432 double CMSStats::time_until_cms_start() const { 433 // We add "gc0_period" to the "work" calculation 434 // below because this query is done (mostly) at the 435 // end of a scavenge, so we need to conservatively 436 // account for that much possible delay 437 // in the query so as to avoid concurrent mode failures 438 // due to starting the collection just a wee bit too 439 // late. 440 double work = cms_duration() + gc0_period(); 441 double deadline = time_until_cms_gen_full(); 442 // If a concurrent mode failure occurred recently, we want to be 443 // more conservative and halve our expected time_until_cms_gen_full() 444 if (work > deadline) { 445 if (Verbose && PrintGCDetails) { 446 gclog_or_tty->print( 447 " CMSCollector: collect because of anticipated promotion " 448 "before full %3.7f + %3.7f > %3.7f ", cms_duration(), 449 gc0_period(), time_until_cms_gen_full()); 450 } 451 return 0.0; 452 } 453 return work - deadline; 454 } 455 456 // Return a duty cycle based on old_duty_cycle and new_duty_cycle, limiting the 457 // amount of change to prevent wild oscillation. 458 unsigned int CMSStats::icms_damped_duty_cycle(unsigned int old_duty_cycle, 459 unsigned int new_duty_cycle) { 460 assert(old_duty_cycle <= 100, "bad input value"); 461 assert(new_duty_cycle <= 100, "bad input value"); 462 463 // Note: use subtraction with caution since it may underflow (values are 464 // unsigned). Addition is safe since we're in the range 0-100. 465 unsigned int damped_duty_cycle = new_duty_cycle; 466 if (new_duty_cycle < old_duty_cycle) { 467 const unsigned int largest_delta = MAX2(old_duty_cycle / 4, 5U); 468 if (new_duty_cycle + largest_delta < old_duty_cycle) { 469 damped_duty_cycle = old_duty_cycle - largest_delta; 470 } 471 } else if (new_duty_cycle > old_duty_cycle) { 472 const unsigned int largest_delta = MAX2(old_duty_cycle / 4, 15U); 473 if (new_duty_cycle > old_duty_cycle + largest_delta) { 474 damped_duty_cycle = MIN2(old_duty_cycle + largest_delta, 100U); 475 } 476 } 477 assert(damped_duty_cycle <= 100, "invalid duty cycle computed"); 478 479 if (CMSTraceIncrementalPacing) { 480 gclog_or_tty->print(" [icms_damped_duty_cycle(%d,%d) = %d] ", 481 old_duty_cycle, new_duty_cycle, damped_duty_cycle); 482 } 483 return damped_duty_cycle; 484 } 485 486 unsigned int CMSStats::icms_update_duty_cycle_impl() { 487 assert(CMSIncrementalPacing && valid(), 488 "should be handled in icms_update_duty_cycle()"); 489 490 double cms_time_so_far = cms_timer().seconds(); 491 double scaled_duration = cms_duration_per_mb() * _cms_used_at_gc0_end / M; 492 double scaled_duration_remaining = fabsd(scaled_duration - cms_time_so_far); 493 494 // Avoid division by 0. 495 double time_until_full = MAX2(time_until_cms_gen_full(), 0.01); 496 double duty_cycle_dbl = 100.0 * scaled_duration_remaining / time_until_full; 497 498 unsigned int new_duty_cycle = MIN2((unsigned int)duty_cycle_dbl, 100U); 499 if (new_duty_cycle > _icms_duty_cycle) { 500 // Avoid very small duty cycles (1 or 2); 0 is allowed. 501 if (new_duty_cycle > 2) { 502 _icms_duty_cycle = icms_damped_duty_cycle(_icms_duty_cycle, 503 new_duty_cycle); 504 } 505 } else if (_allow_duty_cycle_reduction) { 506 // The duty cycle is reduced only once per cms cycle (see record_cms_end()). 507 new_duty_cycle = icms_damped_duty_cycle(_icms_duty_cycle, new_duty_cycle); 508 // Respect the minimum duty cycle. 509 unsigned int min_duty_cycle = (unsigned int)CMSIncrementalDutyCycleMin; 510 _icms_duty_cycle = MAX2(new_duty_cycle, min_duty_cycle); 511 } 512 513 if (PrintGCDetails || CMSTraceIncrementalPacing) { 514 gclog_or_tty->print(" icms_dc=%d ", _icms_duty_cycle); 515 } 516 517 _allow_duty_cycle_reduction = false; 518 return _icms_duty_cycle; 519 } 520 521 #ifndef PRODUCT 522 void CMSStats::print_on(outputStream *st) const { 523 st->print(" gc0_alpha=%d,cms_alpha=%d", _gc0_alpha, _cms_alpha); 524 st->print(",gc0_dur=%g,gc0_per=%g,gc0_promo=" SIZE_FORMAT, 525 gc0_duration(), gc0_period(), gc0_promoted()); 526 st->print(",cms_dur=%g,cms_dur_per_mb=%g,cms_per=%g,cms_alloc=" SIZE_FORMAT, 527 cms_duration(), cms_duration_per_mb(), 528 cms_period(), cms_allocated()); 529 st->print(",cms_since_beg=%g,cms_since_end=%g", 530 cms_time_since_begin(), cms_time_since_end()); 531 st->print(",cms_used_beg=" SIZE_FORMAT ",cms_used_end=" SIZE_FORMAT, 532 _cms_used_at_gc0_begin, _cms_used_at_gc0_end); 533 if (CMSIncrementalMode) { 534 st->print(",dc=%d", icms_duty_cycle()); 535 } 536 537 if (valid()) { 538 st->print(",promo_rate=%g,cms_alloc_rate=%g", 539 promotion_rate(), cms_allocation_rate()); 540 st->print(",cms_consumption_rate=%g,time_until_full=%g", 541 cms_consumption_rate(), time_until_cms_gen_full()); 542 } 543 st->print(" "); 544 } 545 #endif // #ifndef PRODUCT 546 547 CMSCollector::CollectorState CMSCollector::_collectorState = 548 CMSCollector::Idling; 549 bool CMSCollector::_foregroundGCIsActive = false; 550 bool CMSCollector::_foregroundGCShouldWait = false; 551 552 CMSCollector::CMSCollector(ConcurrentMarkSweepGeneration* cmsGen, 553 ConcurrentMarkSweepGeneration* permGen, 554 CardTableRS* ct, 555 ConcurrentMarkSweepPolicy* cp): 556 _cmsGen(cmsGen), 557 _permGen(permGen), 558 _ct(ct), 559 _ref_processor(NULL), // will be set later 560 _conc_workers(NULL), // may be set later 561 _abort_preclean(false), 562 _start_sampling(false), 563 _between_prologue_and_epilogue(false), 564 _markBitMap(0, Mutex::leaf + 1, "CMS_markBitMap_lock"), 565 _perm_gen_verify_bit_map(0, -1 /* no mutex */, "No_lock"), 566 _modUnionTable((CardTableModRefBS::card_shift - LogHeapWordSize), 567 -1 /* lock-free */, "No_lock" /* dummy */), 568 _modUnionClosure(&_modUnionTable), 569 _modUnionClosurePar(&_modUnionTable), 570 // Adjust my span to cover old (cms) gen and perm gen 571 _span(cmsGen->reserved()._union(permGen->reserved())), 572 // Construct the is_alive_closure with _span & markBitMap 573 _is_alive_closure(_span, &_markBitMap), 574 _restart_addr(NULL), 575 _overflow_list(NULL), 576 _stats(cmsGen), 577 _eden_chunk_array(NULL), // may be set in ctor body 578 _eden_chunk_capacity(0), // -- ditto -- 579 _eden_chunk_index(0), // -- ditto -- 580 _survivor_plab_array(NULL), // -- ditto -- 581 _survivor_chunk_array(NULL), // -- ditto -- 582 _survivor_chunk_capacity(0), // -- ditto -- 583 _survivor_chunk_index(0), // -- ditto -- 584 _ser_pmc_preclean_ovflw(0), 585 _ser_kac_preclean_ovflw(0), 586 _ser_pmc_remark_ovflw(0), 587 _par_pmc_remark_ovflw(0), 588 _ser_kac_ovflw(0), 589 _par_kac_ovflw(0), 590 #ifndef PRODUCT 591 _num_par_pushes(0), 592 #endif 593 _collection_count_start(0), 594 _verifying(false), 595 _icms_start_limit(NULL), 596 _icms_stop_limit(NULL), 597 _verification_mark_bm(0, Mutex::leaf + 1, "CMS_verification_mark_bm_lock"), 598 _completed_initialization(false), 599 _collector_policy(cp), 600 _should_unload_classes(false), 601 _concurrent_cycles_since_last_unload(0), 602 _roots_scanning_options(0), 603 _inter_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding), 604 _intra_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding), 605 _gc_tracer_cm(new (ResourceObj::C_HEAP, mtGC) CMSTracer()), 606 _gc_timer_cm(new (ResourceObj::C_HEAP, mtGC) ConcurrentGCTimer()), 607 _cms_start_registered(false) 608 { 609 if (ExplicitGCInvokesConcurrentAndUnloadsClasses) { 610 ExplicitGCInvokesConcurrent = true; 611 } 612 // Now expand the span and allocate the collection support structures 613 // (MUT, marking bit map etc.) to cover both generations subject to 614 // collection. 615 616 // First check that _permGen is adjacent to _cmsGen and above it. 617 assert( _cmsGen->reserved().word_size() > 0 618 && _permGen->reserved().word_size() > 0, 619 "generations should not be of zero size"); 620 assert(_cmsGen->reserved().intersection(_permGen->reserved()).is_empty(), 621 "_cmsGen and _permGen should not overlap"); 622 assert(_cmsGen->reserved().end() == _permGen->reserved().start(), 623 "_cmsGen->end() different from _permGen->start()"); 624 625 // For use by dirty card to oop closures. 626 _cmsGen->cmsSpace()->set_collector(this); 627 _permGen->cmsSpace()->set_collector(this); 628 629 // Allocate MUT and marking bit map 630 { 631 MutexLockerEx x(_markBitMap.lock(), Mutex::_no_safepoint_check_flag); 632 if (!_markBitMap.allocate(_span)) { 633 warning("Failed to allocate CMS Bit Map"); 634 return; 635 } 636 assert(_markBitMap.covers(_span), "_markBitMap inconsistency?"); 637 } 638 { 639 _modUnionTable.allocate(_span); 640 assert(_modUnionTable.covers(_span), "_modUnionTable inconsistency?"); 641 } 642 643 if (!_markStack.allocate(MarkStackSize)) { 644 warning("Failed to allocate CMS Marking Stack"); 645 return; 646 } 647 if (!_revisitStack.allocate(CMSRevisitStackSize)) { 648 warning("Failed to allocate CMS Revisit Stack"); 649 return; 650 } 651 652 // Support for multi-threaded concurrent phases 653 if (CMSConcurrentMTEnabled) { 654 if (FLAG_IS_DEFAULT(ConcGCThreads)) { 655 // just for now 656 FLAG_SET_DEFAULT(ConcGCThreads, (ParallelGCThreads + 3)/4); 657 } 658 if (ConcGCThreads > 1) { 659 _conc_workers = new YieldingFlexibleWorkGang("Parallel CMS Threads", 660 ConcGCThreads, true); 661 if (_conc_workers == NULL) { 662 warning("GC/CMS: _conc_workers allocation failure: " 663 "forcing -CMSConcurrentMTEnabled"); 664 CMSConcurrentMTEnabled = false; 665 } else { 666 _conc_workers->initialize_workers(); 667 } 668 } else { 669 CMSConcurrentMTEnabled = false; 670 } 671 } 672 if (!CMSConcurrentMTEnabled) { 673 ConcGCThreads = 0; 674 } else { 675 // Turn off CMSCleanOnEnter optimization temporarily for 676 // the MT case where it's not fixed yet; see 6178663. 677 CMSCleanOnEnter = false; 678 } 679 assert((_conc_workers != NULL) == (ConcGCThreads > 1), 680 "Inconsistency"); 681 682 // Parallel task queues; these are shared for the 683 // concurrent and stop-world phases of CMS, but 684 // are not shared with parallel scavenge (ParNew). 685 { 686 uint i; 687 uint num_queues = (uint) MAX2(ParallelGCThreads, ConcGCThreads); 688 689 if ((CMSParallelRemarkEnabled || CMSConcurrentMTEnabled 690 || ParallelRefProcEnabled) 691 && num_queues > 0) { 692 _task_queues = new OopTaskQueueSet(num_queues); 693 if (_task_queues == NULL) { 694 warning("task_queues allocation failure."); 695 return; 696 } 697 _hash_seed = NEW_C_HEAP_ARRAY(int, num_queues, mtGC); 698 if (_hash_seed == NULL) { 699 warning("_hash_seed array allocation failure"); 700 return; 701 } 702 703 typedef Padded<OopTaskQueue> PaddedOopTaskQueue; 704 for (i = 0; i < num_queues; i++) { 705 PaddedOopTaskQueue *q = new PaddedOopTaskQueue(); 706 if (q == NULL) { 707 warning("work_queue allocation failure."); 708 return; 709 } 710 _task_queues->register_queue(i, q); 711 } 712 for (i = 0; i < num_queues; i++) { 713 _task_queues->queue(i)->initialize(); 714 _hash_seed[i] = 17; // copied from ParNew 715 } 716 } 717 } 718 719 _cmsGen ->init_initiating_occupancy(CMSInitiatingOccupancyFraction, CMSTriggerRatio); 720 _permGen->init_initiating_occupancy(CMSInitiatingPermOccupancyFraction, CMSTriggerPermRatio); 721 722 // Clip CMSBootstrapOccupancy between 0 and 100. 723 _bootstrap_occupancy = ((double)MIN2((uintx)100, MAX2((uintx)0, CMSBootstrapOccupancy))) 724 /(double)100; 725 726 _full_gcs_since_conc_gc = 0; 727 728 // Now tell CMS generations the identity of their collector 729 ConcurrentMarkSweepGeneration::set_collector(this); 730 731 // Create & start a CMS thread for this CMS collector 732 _cmsThread = ConcurrentMarkSweepThread::start(this); 733 assert(cmsThread() != NULL, "CMS Thread should have been created"); 734 assert(cmsThread()->collector() == this, 735 "CMS Thread should refer to this gen"); 736 assert(CGC_lock != NULL, "Where's the CGC_lock?"); 737 738 // Support for parallelizing young gen rescan 739 GenCollectedHeap* gch = GenCollectedHeap::heap(); 740 _young_gen = gch->prev_gen(_cmsGen); 741 if (gch->supports_inline_contig_alloc()) { 742 _top_addr = gch->top_addr(); 743 _end_addr = gch->end_addr(); 744 assert(_young_gen != NULL, "no _young_gen"); 745 _eden_chunk_index = 0; 746 _eden_chunk_capacity = (_young_gen->max_capacity()+CMSSamplingGrain)/CMSSamplingGrain; 747 _eden_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, _eden_chunk_capacity, mtGC); 748 if (_eden_chunk_array == NULL) { 749 _eden_chunk_capacity = 0; 750 warning("GC/CMS: _eden_chunk_array allocation failure"); 751 } 752 } 753 assert(_eden_chunk_array != NULL || _eden_chunk_capacity == 0, "Error"); 754 755 // Support for parallelizing survivor space rescan 756 if (CMSParallelRemarkEnabled && CMSParallelSurvivorRemarkEnabled) { 757 const size_t max_plab_samples = 758 ((DefNewGeneration*)_young_gen)->max_survivor_size()/MinTLABSize; 759 760 _survivor_plab_array = NEW_C_HEAP_ARRAY(ChunkArray, ParallelGCThreads, mtGC); 761 _survivor_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, 2*max_plab_samples, mtGC); 762 _cursor = NEW_C_HEAP_ARRAY(size_t, ParallelGCThreads, mtGC); 763 if (_survivor_plab_array == NULL || _survivor_chunk_array == NULL 764 || _cursor == NULL) { 765 warning("Failed to allocate survivor plab/chunk array"); 766 if (_survivor_plab_array != NULL) { 767 FREE_C_HEAP_ARRAY(ChunkArray, _survivor_plab_array, mtGC); 768 _survivor_plab_array = NULL; 769 } 770 if (_survivor_chunk_array != NULL) { 771 FREE_C_HEAP_ARRAY(HeapWord*, _survivor_chunk_array, mtGC); 772 _survivor_chunk_array = NULL; 773 } 774 if (_cursor != NULL) { 775 FREE_C_HEAP_ARRAY(size_t, _cursor, mtGC); 776 _cursor = NULL; 777 } 778 } else { 779 _survivor_chunk_capacity = 2*max_plab_samples; 780 for (uint i = 0; i < ParallelGCThreads; i++) { 781 HeapWord** vec = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples, mtGC); 782 if (vec == NULL) { 783 warning("Failed to allocate survivor plab array"); 784 for (int j = i; j > 0; j--) { 785 FREE_C_HEAP_ARRAY(HeapWord*, _survivor_plab_array[j-1].array(), mtGC); 786 } 787 FREE_C_HEAP_ARRAY(ChunkArray, _survivor_plab_array, mtGC); 788 FREE_C_HEAP_ARRAY(HeapWord*, _survivor_chunk_array, mtGC); 789 _survivor_plab_array = NULL; 790 _survivor_chunk_array = NULL; 791 _survivor_chunk_capacity = 0; 792 break; 793 } else { 794 ChunkArray* cur = 795 ::new (&_survivor_plab_array[i]) ChunkArray(vec, 796 max_plab_samples); 797 assert(cur->end() == 0, "Should be 0"); 798 assert(cur->array() == vec, "Should be vec"); 799 assert(cur->capacity() == max_plab_samples, "Error"); 800 } 801 } 802 } 803 } 804 assert( ( _survivor_plab_array != NULL 805 && _survivor_chunk_array != NULL) 806 || ( _survivor_chunk_capacity == 0 807 && _survivor_chunk_index == 0), 808 "Error"); 809 810 // Choose what strong roots should be scanned depending on verification options 811 // and perm gen collection mode. 812 if (!CMSClassUnloadingEnabled) { 813 // If class unloading is disabled we want to include all classes into the root set. 814 add_root_scanning_option(SharedHeap::SO_AllClasses); 815 } else { 816 add_root_scanning_option(SharedHeap::SO_SystemClasses); 817 } 818 819 NOT_PRODUCT(_overflow_counter = CMSMarkStackOverflowInterval;) 820 _gc_counters = new CollectorCounters("CMS", 1); 821 _completed_initialization = true; 822 _inter_sweep_timer.start(); // start of time 823 } 824 825 const char* ConcurrentMarkSweepGeneration::name() const { 826 return "concurrent mark-sweep generation"; 827 } 828 void ConcurrentMarkSweepGeneration::update_counters() { 829 if (UsePerfData) { 830 _space_counters->update_all(); 831 _gen_counters->update_all(); 832 } 833 } 834 835 // this is an optimized version of update_counters(). it takes the 836 // used value as a parameter rather than computing it. 837 // 838 void ConcurrentMarkSweepGeneration::update_counters(size_t used) { 839 if (UsePerfData) { 840 _space_counters->update_used(used); 841 _space_counters->update_capacity(); 842 _gen_counters->update_all(); 843 } 844 } 845 846 void ConcurrentMarkSweepGeneration::print() const { 847 Generation::print(); 848 cmsSpace()->print(); 849 } 850 851 #ifndef PRODUCT 852 void ConcurrentMarkSweepGeneration::print_statistics() { 853 cmsSpace()->printFLCensus(0); 854 } 855 #endif 856 857 void ConcurrentMarkSweepGeneration::printOccupancy(const char *s) { 858 GenCollectedHeap* gch = GenCollectedHeap::heap(); 859 if (PrintGCDetails) { 860 if (Verbose) { 861 gclog_or_tty->print(" [%d %s-%s: "SIZE_FORMAT"("SIZE_FORMAT")]", 862 level(), short_name(), s, used(), capacity()); 863 } else { 864 gclog_or_tty->print(" [%d %s-%s: "SIZE_FORMAT"K("SIZE_FORMAT"K)]", 865 level(), short_name(), s, used() / K, capacity() / K); 866 } 867 } 868 if (Verbose) { 869 gclog_or_tty->print(" "SIZE_FORMAT"("SIZE_FORMAT")", 870 gch->used(), gch->capacity()); 871 } else { 872 gclog_or_tty->print(" "SIZE_FORMAT"K("SIZE_FORMAT"K)", 873 gch->used() / K, gch->capacity() / K); 874 } 875 } 876 877 size_t 878 ConcurrentMarkSweepGeneration::contiguous_available() const { 879 // dld proposes an improvement in precision here. If the committed 880 // part of the space ends in a free block we should add that to 881 // uncommitted size in the calculation below. Will make this 882 // change later, staying with the approximation below for the 883 // time being. -- ysr. 884 return MAX2(_virtual_space.uncommitted_size(), unsafe_max_alloc_nogc()); 885 } 886 887 size_t 888 ConcurrentMarkSweepGeneration::unsafe_max_alloc_nogc() const { 889 return _cmsSpace->max_alloc_in_words() * HeapWordSize; 890 } 891 892 size_t ConcurrentMarkSweepGeneration::max_available() const { 893 return free() + _virtual_space.uncommitted_size(); 894 } 895 896 bool ConcurrentMarkSweepGeneration::promotion_attempt_is_safe(size_t max_promotion_in_bytes) const { 897 size_t available = max_available(); 898 size_t av_promo = (size_t)gc_stats()->avg_promoted()->padded_average(); 899 bool res = (available >= av_promo) || (available >= max_promotion_in_bytes); 900 if (Verbose && PrintGCDetails) { 901 gclog_or_tty->print_cr( 902 "CMS: promo attempt is%s safe: available("SIZE_FORMAT") %s av_promo("SIZE_FORMAT")," 903 "max_promo("SIZE_FORMAT")", 904 res? "":" not", available, res? ">=":"<", 905 av_promo, max_promotion_in_bytes); 906 } 907 return res; 908 } 909 910 // At a promotion failure dump information on block layout in heap 911 // (cms old generation). 912 void ConcurrentMarkSweepGeneration::promotion_failure_occurred() { 913 if (CMSDumpAtPromotionFailure) { 914 cmsSpace()->dump_at_safepoint_with_locks(collector(), gclog_or_tty); 915 } 916 } 917 918 CompactibleSpace* 919 ConcurrentMarkSweepGeneration::first_compaction_space() const { 920 return _cmsSpace; 921 } 922 923 void ConcurrentMarkSweepGeneration::reset_after_compaction() { 924 // Clear the promotion information. These pointers can be adjusted 925 // along with all the other pointers into the heap but 926 // compaction is expected to be a rare event with 927 // a heap using cms so don't do it without seeing the need. 928 if (CollectedHeap::use_parallel_gc_threads()) { 929 for (uint i = 0; i < ParallelGCThreads; i++) { 930 _par_gc_thread_states[i]->promo.reset(); 931 } 932 } 933 } 934 935 void ConcurrentMarkSweepGeneration::space_iterate(SpaceClosure* blk, bool usedOnly) { 936 blk->do_space(_cmsSpace); 937 } 938 939 void ConcurrentMarkSweepGeneration::compute_new_size() { 940 assert_locked_or_safepoint(Heap_lock); 941 942 // If incremental collection failed, we just want to expand 943 // to the limit. 944 if (incremental_collection_failed()) { 945 clear_incremental_collection_failed(); 946 grow_to_reserved(); 947 return; 948 } 949 950 size_t expand_bytes = 0; 951 double free_percentage = ((double) free()) / capacity(); 952 double desired_free_percentage = (double) MinHeapFreeRatio / 100; 953 double maximum_free_percentage = (double) MaxHeapFreeRatio / 100; 954 955 // compute expansion delta needed for reaching desired free percentage 956 if (free_percentage < desired_free_percentage) { 957 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 958 assert(desired_capacity >= capacity(), "invalid expansion size"); 959 expand_bytes = MAX2(desired_capacity - capacity(), MinHeapDeltaBytes); 960 } 961 if (expand_bytes > 0) { 962 if (PrintGCDetails && Verbose) { 963 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 964 gclog_or_tty->print_cr("\nFrom compute_new_size: "); 965 gclog_or_tty->print_cr(" Free fraction %f", free_percentage); 966 gclog_or_tty->print_cr(" Desired free fraction %f", 967 desired_free_percentage); 968 gclog_or_tty->print_cr(" Maximum free fraction %f", 969 maximum_free_percentage); 970 gclog_or_tty->print_cr(" Capactiy "SIZE_FORMAT, capacity()/1000); 971 gclog_or_tty->print_cr(" Desired capacity "SIZE_FORMAT, 972 desired_capacity/1000); 973 int prev_level = level() - 1; 974 if (prev_level >= 0) { 975 size_t prev_size = 0; 976 GenCollectedHeap* gch = GenCollectedHeap::heap(); 977 Generation* prev_gen = gch->_gens[prev_level]; 978 prev_size = prev_gen->capacity(); 979 gclog_or_tty->print_cr(" Younger gen size "SIZE_FORMAT, 980 prev_size/1000); 981 } 982 gclog_or_tty->print_cr(" unsafe_max_alloc_nogc "SIZE_FORMAT, 983 unsafe_max_alloc_nogc()/1000); 984 gclog_or_tty->print_cr(" contiguous available "SIZE_FORMAT, 985 contiguous_available()/1000); 986 gclog_or_tty->print_cr(" Expand by "SIZE_FORMAT" (bytes)", 987 expand_bytes); 988 } 989 // safe if expansion fails 990 expand(expand_bytes, 0, CMSExpansionCause::_satisfy_free_ratio); 991 if (PrintGCDetails && Verbose) { 992 gclog_or_tty->print_cr(" Expanded free fraction %f", 993 ((double) free()) / capacity()); 994 } 995 } 996 } 997 998 Mutex* ConcurrentMarkSweepGeneration::freelistLock() const { 999 return cmsSpace()->freelistLock(); 1000 } 1001 1002 HeapWord* ConcurrentMarkSweepGeneration::allocate(size_t size, 1003 bool tlab) { 1004 CMSSynchronousYieldRequest yr; 1005 MutexLockerEx x(freelistLock(), 1006 Mutex::_no_safepoint_check_flag); 1007 return have_lock_and_allocate(size, tlab); 1008 } 1009 1010 HeapWord* ConcurrentMarkSweepGeneration::have_lock_and_allocate(size_t size, 1011 bool tlab /* ignored */) { 1012 assert_lock_strong(freelistLock()); 1013 size_t adjustedSize = CompactibleFreeListSpace::adjustObjectSize(size); 1014 HeapWord* res = cmsSpace()->allocate(adjustedSize); 1015 // Allocate the object live (grey) if the background collector has 1016 // started marking. This is necessary because the marker may 1017 // have passed this address and consequently this object will 1018 // not otherwise be greyed and would be incorrectly swept up. 1019 // Note that if this object contains references, the writing 1020 // of those references will dirty the card containing this object 1021 // allowing the object to be blackened (and its references scanned) 1022 // either during a preclean phase or at the final checkpoint. 1023 if (res != NULL) { 1024 // We may block here with an uninitialized object with 1025 // its mark-bit or P-bits not yet set. Such objects need 1026 // to be safely navigable by block_start(). 1027 assert(oop(res)->klass_or_null() == NULL, "Object should be uninitialized here."); 1028 assert(!((FreeChunk*)res)->is_free(), "Error, block will look free but show wrong size"); 1029 collector()->direct_allocated(res, adjustedSize); 1030 _direct_allocated_words += adjustedSize; 1031 // allocation counters 1032 NOT_PRODUCT( 1033 _numObjectsAllocated++; 1034 _numWordsAllocated += (int)adjustedSize; 1035 ) 1036 } 1037 return res; 1038 } 1039 1040 // In the case of direct allocation by mutators in a generation that 1041 // is being concurrently collected, the object must be allocated 1042 // live (grey) if the background collector has started marking. 1043 // This is necessary because the marker may 1044 // have passed this address and consequently this object will 1045 // not otherwise be greyed and would be incorrectly swept up. 1046 // Note that if this object contains references, the writing 1047 // of those references will dirty the card containing this object 1048 // allowing the object to be blackened (and its references scanned) 1049 // either during a preclean phase or at the final checkpoint. 1050 void CMSCollector::direct_allocated(HeapWord* start, size_t size) { 1051 assert(_markBitMap.covers(start, size), "Out of bounds"); 1052 if (_collectorState >= Marking) { 1053 MutexLockerEx y(_markBitMap.lock(), 1054 Mutex::_no_safepoint_check_flag); 1055 // [see comments preceding SweepClosure::do_blk() below for details] 1056 // 1. need to mark the object as live so it isn't collected 1057 // 2. need to mark the 2nd bit to indicate the object may be uninitialized 1058 // 3. need to mark the end of the object so marking, precleaning or sweeping 1059 // can skip over uninitialized or unparsable objects. An allocated 1060 // object is considered uninitialized for our purposes as long as 1061 // its klass word is NULL. (Unparsable objects are those which are 1062 // initialized in the sense just described, but whose sizes can still 1063 // not be correctly determined. Note that the class of unparsable objects 1064 // can only occur in the perm gen. All old gen objects are parsable 1065 // as soon as they are initialized.) 1066 _markBitMap.mark(start); // object is live 1067 _markBitMap.mark(start + 1); // object is potentially uninitialized? 1068 _markBitMap.mark(start + size - 1); 1069 // mark end of object 1070 } 1071 // check that oop looks uninitialized 1072 assert(oop(start)->klass_or_null() == NULL, "_klass should be NULL"); 1073 } 1074 1075 void CMSCollector::promoted(bool par, HeapWord* start, 1076 bool is_obj_array, size_t obj_size) { 1077 assert(_markBitMap.covers(start), "Out of bounds"); 1078 // See comment in direct_allocated() about when objects should 1079 // be allocated live. 1080 if (_collectorState >= Marking) { 1081 // we already hold the marking bit map lock, taken in 1082 // the prologue 1083 if (par) { 1084 _markBitMap.par_mark(start); 1085 } else { 1086 _markBitMap.mark(start); 1087 } 1088 // We don't need to mark the object as uninitialized (as 1089 // in direct_allocated above) because this is being done with the 1090 // world stopped and the object will be initialized by the 1091 // time the marking, precleaning or sweeping get to look at it. 1092 // But see the code for copying objects into the CMS generation, 1093 // where we need to ensure that concurrent readers of the 1094 // block offset table are able to safely navigate a block that 1095 // is in flux from being free to being allocated (and in 1096 // transition while being copied into) and subsequently 1097 // becoming a bona-fide object when the copy/promotion is complete. 1098 assert(SafepointSynchronize::is_at_safepoint(), 1099 "expect promotion only at safepoints"); 1100 1101 if (_collectorState < Sweeping) { 1102 // Mark the appropriate cards in the modUnionTable, so that 1103 // this object gets scanned before the sweep. If this is 1104 // not done, CMS generation references in the object might 1105 // not get marked. 1106 // For the case of arrays, which are otherwise precisely 1107 // marked, we need to dirty the entire array, not just its head. 1108 if (is_obj_array) { 1109 // The [par_]mark_range() method expects mr.end() below to 1110 // be aligned to the granularity of a bit's representation 1111 // in the heap. In the case of the MUT below, that's a 1112 // card size. 1113 MemRegion mr(start, 1114 (HeapWord*)round_to((intptr_t)(start + obj_size), 1115 CardTableModRefBS::card_size /* bytes */)); 1116 if (par) { 1117 _modUnionTable.par_mark_range(mr); 1118 } else { 1119 _modUnionTable.mark_range(mr); 1120 } 1121 } else { // not an obj array; we can just mark the head 1122 if (par) { 1123 _modUnionTable.par_mark(start); 1124 } else { 1125 _modUnionTable.mark(start); 1126 } 1127 } 1128 } 1129 } 1130 } 1131 1132 static inline size_t percent_of_space(Space* space, HeapWord* addr) 1133 { 1134 size_t delta = pointer_delta(addr, space->bottom()); 1135 return (size_t)(delta * 100.0 / (space->capacity() / HeapWordSize)); 1136 } 1137 1138 void CMSCollector::icms_update_allocation_limits() 1139 { 1140 Generation* gen0 = GenCollectedHeap::heap()->get_gen(0); 1141 EdenSpace* eden = gen0->as_DefNewGeneration()->eden(); 1142 1143 const unsigned int duty_cycle = stats().icms_update_duty_cycle(); 1144 if (CMSTraceIncrementalPacing) { 1145 stats().print(); 1146 } 1147 1148 assert(duty_cycle <= 100, "invalid duty cycle"); 1149 if (duty_cycle != 0) { 1150 // The duty_cycle is a percentage between 0 and 100; convert to words and 1151 // then compute the offset from the endpoints of the space. 1152 size_t free_words = eden->free() / HeapWordSize; 1153 double free_words_dbl = (double)free_words; 1154 size_t duty_cycle_words = (size_t)(free_words_dbl * duty_cycle / 100.0); 1155 size_t offset_words = (free_words - duty_cycle_words) / 2; 1156 1157 _icms_start_limit = eden->top() + offset_words; 1158 _icms_stop_limit = eden->end() - offset_words; 1159 1160 // The limits may be adjusted (shifted to the right) by 1161 // CMSIncrementalOffset, to allow the application more mutator time after a 1162 // young gen gc (when all mutators were stopped) and before CMS starts and 1163 // takes away one or more cpus. 1164 if (CMSIncrementalOffset != 0) { 1165 double adjustment_dbl = free_words_dbl * CMSIncrementalOffset / 100.0; 1166 size_t adjustment = (size_t)adjustment_dbl; 1167 HeapWord* tmp_stop = _icms_stop_limit + adjustment; 1168 if (tmp_stop > _icms_stop_limit && tmp_stop < eden->end()) { 1169 _icms_start_limit += adjustment; 1170 _icms_stop_limit = tmp_stop; 1171 } 1172 } 1173 } 1174 if (duty_cycle == 0 || (_icms_start_limit == _icms_stop_limit)) { 1175 _icms_start_limit = _icms_stop_limit = eden->end(); 1176 } 1177 1178 // Install the new start limit. 1179 eden->set_soft_end(_icms_start_limit); 1180 1181 if (CMSTraceIncrementalMode) { 1182 gclog_or_tty->print(" icms alloc limits: " 1183 PTR_FORMAT "," PTR_FORMAT 1184 " (" SIZE_FORMAT "%%," SIZE_FORMAT "%%) ", 1185 _icms_start_limit, _icms_stop_limit, 1186 percent_of_space(eden, _icms_start_limit), 1187 percent_of_space(eden, _icms_stop_limit)); 1188 if (Verbose) { 1189 gclog_or_tty->print("eden: "); 1190 eden->print_on(gclog_or_tty); 1191 } 1192 } 1193 } 1194 1195 // Any changes here should try to maintain the invariant 1196 // that if this method is called with _icms_start_limit 1197 // and _icms_stop_limit both NULL, then it should return NULL 1198 // and not notify the icms thread. 1199 HeapWord* 1200 CMSCollector::allocation_limit_reached(Space* space, HeapWord* top, 1201 size_t word_size) 1202 { 1203 // A start_limit equal to end() means the duty cycle is 0, so treat that as a 1204 // nop. 1205 if (CMSIncrementalMode && _icms_start_limit != space->end()) { 1206 if (top <= _icms_start_limit) { 1207 if (CMSTraceIncrementalMode) { 1208 space->print_on(gclog_or_tty); 1209 gclog_or_tty->stamp(); 1210 gclog_or_tty->print_cr(" start limit top=" PTR_FORMAT 1211 ", new limit=" PTR_FORMAT 1212 " (" SIZE_FORMAT "%%)", 1213 top, _icms_stop_limit, 1214 percent_of_space(space, _icms_stop_limit)); 1215 } 1216 ConcurrentMarkSweepThread::start_icms(); 1217 assert(top < _icms_stop_limit, "Tautology"); 1218 if (word_size < pointer_delta(_icms_stop_limit, top)) { 1219 return _icms_stop_limit; 1220 } 1221 1222 // The allocation will cross both the _start and _stop limits, so do the 1223 // stop notification also and return end(). 1224 if (CMSTraceIncrementalMode) { 1225 space->print_on(gclog_or_tty); 1226 gclog_or_tty->stamp(); 1227 gclog_or_tty->print_cr(" +stop limit top=" PTR_FORMAT 1228 ", new limit=" PTR_FORMAT 1229 " (" SIZE_FORMAT "%%)", 1230 top, space->end(), 1231 percent_of_space(space, space->end())); 1232 } 1233 ConcurrentMarkSweepThread::stop_icms(); 1234 return space->end(); 1235 } 1236 1237 if (top <= _icms_stop_limit) { 1238 if (CMSTraceIncrementalMode) { 1239 space->print_on(gclog_or_tty); 1240 gclog_or_tty->stamp(); 1241 gclog_or_tty->print_cr(" stop limit top=" PTR_FORMAT 1242 ", new limit=" PTR_FORMAT 1243 " (" SIZE_FORMAT "%%)", 1244 top, space->end(), 1245 percent_of_space(space, space->end())); 1246 } 1247 ConcurrentMarkSweepThread::stop_icms(); 1248 return space->end(); 1249 } 1250 1251 if (CMSTraceIncrementalMode) { 1252 space->print_on(gclog_or_tty); 1253 gclog_or_tty->stamp(); 1254 gclog_or_tty->print_cr(" end limit top=" PTR_FORMAT 1255 ", new limit=" PTR_FORMAT, 1256 top, NULL); 1257 } 1258 } 1259 1260 return NULL; 1261 } 1262 1263 oop ConcurrentMarkSweepGeneration::promote(oop obj, size_t obj_size) { 1264 assert(obj_size == (size_t)obj->size(), "bad obj_size passed in"); 1265 // allocate, copy and if necessary update promoinfo -- 1266 // delegate to underlying space. 1267 assert_lock_strong(freelistLock()); 1268 1269 #ifndef PRODUCT 1270 if (Universe::heap()->promotion_should_fail()) { 1271 return NULL; 1272 } 1273 #endif // #ifndef PRODUCT 1274 1275 oop res = _cmsSpace->promote(obj, obj_size); 1276 if (res == NULL) { 1277 // expand and retry 1278 size_t s = _cmsSpace->expansionSpaceRequired(obj_size); // HeapWords 1279 expand(s*HeapWordSize, MinHeapDeltaBytes, 1280 CMSExpansionCause::_satisfy_promotion); 1281 // Since there's currently no next generation, we don't try to promote 1282 // into a more senior generation. 1283 assert(next_gen() == NULL, "assumption, based upon which no attempt " 1284 "is made to pass on a possibly failing " 1285 "promotion to next generation"); 1286 res = _cmsSpace->promote(obj, obj_size); 1287 } 1288 if (res != NULL) { 1289 // See comment in allocate() about when objects should 1290 // be allocated live. 1291 assert(obj->is_oop(), "Will dereference klass pointer below"); 1292 collector()->promoted(false, // Not parallel 1293 (HeapWord*)res, obj->is_objArray(), obj_size); 1294 // promotion counters 1295 NOT_PRODUCT( 1296 _numObjectsPromoted++; 1297 _numWordsPromoted += 1298 (int)(CompactibleFreeListSpace::adjustObjectSize(obj->size())); 1299 ) 1300 } 1301 return res; 1302 } 1303 1304 1305 HeapWord* 1306 ConcurrentMarkSweepGeneration::allocation_limit_reached(Space* space, 1307 HeapWord* top, 1308 size_t word_sz) 1309 { 1310 return collector()->allocation_limit_reached(space, top, word_sz); 1311 } 1312 1313 // IMPORTANT: Notes on object size recognition in CMS. 1314 // --------------------------------------------------- 1315 // A block of storage in the CMS generation is always in 1316 // one of three states. A free block (FREE), an allocated 1317 // object (OBJECT) whose size() method reports the correct size, 1318 // and an intermediate state (TRANSIENT) in which its size cannot 1319 // be accurately determined. 1320 // STATE IDENTIFICATION: (32 bit and 64 bit w/o COOPS) 1321 // ----------------------------------------------------- 1322 // FREE: klass_word & 1 == 1; mark_word holds block size 1323 // 1324 // OBJECT: klass_word installed; klass_word != 0 && klass_word & 1 == 0; 1325 // obj->size() computes correct size 1326 // [Perm Gen objects needs to be "parsable" before they can be navigated] 1327 // 1328 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT 1329 // 1330 // STATE IDENTIFICATION: (64 bit+COOPS) 1331 // ------------------------------------ 1332 // FREE: mark_word & CMS_FREE_BIT == 1; mark_word & ~CMS_FREE_BIT gives block_size 1333 // 1334 // OBJECT: klass_word installed; klass_word != 0; 1335 // obj->size() computes correct size 1336 // [Perm Gen comment above continues to hold] 1337 // 1338 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT 1339 // 1340 // 1341 // STATE TRANSITION DIAGRAM 1342 // 1343 // mut / parnew mut / parnew 1344 // FREE --------------------> TRANSIENT ---------------------> OBJECT --| 1345 // ^ | 1346 // |------------------------ DEAD <------------------------------------| 1347 // sweep mut 1348 // 1349 // While a block is in TRANSIENT state its size cannot be determined 1350 // so readers will either need to come back later or stall until 1351 // the size can be determined. Note that for the case of direct 1352 // allocation, P-bits, when available, may be used to determine the 1353 // size of an object that may not yet have been initialized. 1354 1355 // Things to support parallel young-gen collection. 1356 oop 1357 ConcurrentMarkSweepGeneration::par_promote(int thread_num, 1358 oop old, markOop m, 1359 size_t word_sz) { 1360 #ifndef PRODUCT 1361 if (Universe::heap()->promotion_should_fail()) { 1362 return NULL; 1363 } 1364 #endif // #ifndef PRODUCT 1365 1366 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1367 PromotionInfo* promoInfo = &ps->promo; 1368 // if we are tracking promotions, then first ensure space for 1369 // promotion (including spooling space for saving header if necessary). 1370 // then allocate and copy, then track promoted info if needed. 1371 // When tracking (see PromotionInfo::track()), the mark word may 1372 // be displaced and in this case restoration of the mark word 1373 // occurs in the (oop_since_save_marks_)iterate phase. 1374 if (promoInfo->tracking() && !promoInfo->ensure_spooling_space()) { 1375 // Out of space for allocating spooling buffers; 1376 // try expanding and allocating spooling buffers. 1377 if (!expand_and_ensure_spooling_space(promoInfo)) { 1378 return NULL; 1379 } 1380 } 1381 assert(promoInfo->has_spooling_space(), "Control point invariant"); 1382 const size_t alloc_sz = CompactibleFreeListSpace::adjustObjectSize(word_sz); 1383 HeapWord* obj_ptr = ps->lab.alloc(alloc_sz); 1384 if (obj_ptr == NULL) { 1385 obj_ptr = expand_and_par_lab_allocate(ps, alloc_sz); 1386 if (obj_ptr == NULL) { 1387 return NULL; 1388 } 1389 } 1390 oop obj = oop(obj_ptr); 1391 OrderAccess::storestore(); 1392 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1393 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1394 // IMPORTANT: See note on object initialization for CMS above. 1395 // Otherwise, copy the object. Here we must be careful to insert the 1396 // klass pointer last, since this marks the block as an allocated object. 1397 // Except with compressed oops it's the mark word. 1398 HeapWord* old_ptr = (HeapWord*)old; 1399 // Restore the mark word copied above. 1400 obj->set_mark(m); 1401 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1402 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1403 OrderAccess::storestore(); 1404 1405 if (UseCompressedOops) { 1406 // Copy gap missed by (aligned) header size calculation below 1407 obj->set_klass_gap(old->klass_gap()); 1408 } 1409 if (word_sz > (size_t)oopDesc::header_size()) { 1410 Copy::aligned_disjoint_words(old_ptr + oopDesc::header_size(), 1411 obj_ptr + oopDesc::header_size(), 1412 word_sz - oopDesc::header_size()); 1413 } 1414 1415 // Now we can track the promoted object, if necessary. We take care 1416 // to delay the transition from uninitialized to full object 1417 // (i.e., insertion of klass pointer) until after, so that it 1418 // atomically becomes a promoted object. 1419 if (promoInfo->tracking()) { 1420 promoInfo->track((PromotedObject*)obj, old->klass()); 1421 } 1422 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1423 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1424 assert(old->is_oop(), "Will use and dereference old klass ptr below"); 1425 1426 // Finally, install the klass pointer (this should be volatile). 1427 OrderAccess::storestore(); 1428 obj->set_klass(old->klass()); 1429 // We should now be able to calculate the right size for this object 1430 assert(obj->is_oop() && obj->size() == (int)word_sz, "Error, incorrect size computed for promoted object"); 1431 1432 collector()->promoted(true, // parallel 1433 obj_ptr, old->is_objArray(), word_sz); 1434 1435 NOT_PRODUCT( 1436 Atomic::inc_ptr(&_numObjectsPromoted); 1437 Atomic::add_ptr(alloc_sz, &_numWordsPromoted); 1438 ) 1439 1440 return obj; 1441 } 1442 1443 void 1444 ConcurrentMarkSweepGeneration:: 1445 par_promote_alloc_undo(int thread_num, 1446 HeapWord* obj, size_t word_sz) { 1447 // CMS does not support promotion undo. 1448 ShouldNotReachHere(); 1449 } 1450 1451 void 1452 ConcurrentMarkSweepGeneration:: 1453 par_promote_alloc_done(int thread_num) { 1454 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1455 ps->lab.retire(thread_num); 1456 } 1457 1458 void 1459 ConcurrentMarkSweepGeneration:: 1460 par_oop_since_save_marks_iterate_done(int thread_num) { 1461 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1462 ParScanWithoutBarrierClosure* dummy_cl = NULL; 1463 ps->promo.promoted_oops_iterate_nv(dummy_cl); 1464 } 1465 1466 // XXXPERM 1467 bool ConcurrentMarkSweepGeneration::should_collect(bool full, 1468 size_t size, 1469 bool tlab) 1470 { 1471 // We allow a STW collection only if a full 1472 // collection was requested. 1473 return full || should_allocate(size, tlab); // FIX ME !!! 1474 // This and promotion failure handling are connected at the 1475 // hip and should be fixed by untying them. 1476 } 1477 1478 bool CMSCollector::shouldConcurrentCollect() { 1479 if (_full_gc_requested) { 1480 if (Verbose && PrintGCDetails) { 1481 gclog_or_tty->print_cr("CMSCollector: collect because of explicit " 1482 " gc request (or gc_locker)"); 1483 } 1484 return true; 1485 } 1486 1487 // For debugging purposes, change the type of collection. 1488 // If the rotation is not on the concurrent collection 1489 // type, don't start a concurrent collection. 1490 NOT_PRODUCT( 1491 if (RotateCMSCollectionTypes && 1492 (_cmsGen->debug_collection_type() != 1493 ConcurrentMarkSweepGeneration::Concurrent_collection_type)) { 1494 assert(_cmsGen->debug_collection_type() != 1495 ConcurrentMarkSweepGeneration::Unknown_collection_type, 1496 "Bad cms collection type"); 1497 return false; 1498 } 1499 ) 1500 1501 FreelistLocker x(this); 1502 // ------------------------------------------------------------------ 1503 // Print out lots of information which affects the initiation of 1504 // a collection. 1505 if (PrintCMSInitiationStatistics && stats().valid()) { 1506 gclog_or_tty->print("CMSCollector shouldConcurrentCollect: "); 1507 gclog_or_tty->stamp(); 1508 gclog_or_tty->print_cr(""); 1509 stats().print_on(gclog_or_tty); 1510 gclog_or_tty->print_cr("time_until_cms_gen_full %3.7f", 1511 stats().time_until_cms_gen_full()); 1512 gclog_or_tty->print_cr("free="SIZE_FORMAT, _cmsGen->free()); 1513 gclog_or_tty->print_cr("contiguous_available="SIZE_FORMAT, 1514 _cmsGen->contiguous_available()); 1515 gclog_or_tty->print_cr("promotion_rate=%g", stats().promotion_rate()); 1516 gclog_or_tty->print_cr("cms_allocation_rate=%g", stats().cms_allocation_rate()); 1517 gclog_or_tty->print_cr("occupancy=%3.7f", _cmsGen->occupancy()); 1518 gclog_or_tty->print_cr("initiatingOccupancy=%3.7f", _cmsGen->initiating_occupancy()); 1519 gclog_or_tty->print_cr("initiatingPermOccupancy=%3.7f", _permGen->initiating_occupancy()); 1520 } 1521 // ------------------------------------------------------------------ 1522 1523 // If the estimated time to complete a cms collection (cms_duration()) 1524 // is less than the estimated time remaining until the cms generation 1525 // is full, start a collection. 1526 if (!UseCMSInitiatingOccupancyOnly) { 1527 if (stats().valid()) { 1528 if (stats().time_until_cms_start() == 0.0) { 1529 return true; 1530 } 1531 } else { 1532 // We want to conservatively collect somewhat early in order 1533 // to try and "bootstrap" our CMS/promotion statistics; 1534 // this branch will not fire after the first successful CMS 1535 // collection because the stats should then be valid. 1536 if (_cmsGen->occupancy() >= _bootstrap_occupancy) { 1537 if (Verbose && PrintGCDetails) { 1538 gclog_or_tty->print_cr( 1539 " CMSCollector: collect for bootstrapping statistics:" 1540 " occupancy = %f, boot occupancy = %f", _cmsGen->occupancy(), 1541 _bootstrap_occupancy); 1542 } 1543 return true; 1544 } 1545 } 1546 } 1547 1548 // Otherwise, we start a collection cycle if either the perm gen or 1549 // old gen want a collection cycle started. Each may use 1550 // an appropriate criterion for making this decision. 1551 // XXX We need to make sure that the gen expansion 1552 // criterion dovetails well with this. XXX NEED TO FIX THIS 1553 if (_cmsGen->should_concurrent_collect()) { 1554 if (Verbose && PrintGCDetails) { 1555 gclog_or_tty->print_cr("CMS old gen initiated"); 1556 } 1557 return true; 1558 } 1559 1560 // We start a collection if we believe an incremental collection may fail; 1561 // this is not likely to be productive in practice because it's probably too 1562 // late anyway. 1563 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1564 assert(gch->collector_policy()->is_two_generation_policy(), 1565 "You may want to check the correctness of the following"); 1566 if (gch->incremental_collection_will_fail(true /* consult_young */)) { 1567 if (Verbose && PrintGCDetails) { 1568 gclog_or_tty->print("CMSCollector: collect because incremental collection will fail "); 1569 } 1570 return true; 1571 } 1572 1573 if (CMSClassUnloadingEnabled && _permGen->should_concurrent_collect()) { 1574 bool res = update_should_unload_classes(); 1575 if (res) { 1576 if (Verbose && PrintGCDetails) { 1577 gclog_or_tty->print_cr("CMS perm gen initiated"); 1578 } 1579 return true; 1580 } 1581 } 1582 return false; 1583 } 1584 1585 // Clear _expansion_cause fields of constituent generations 1586 void CMSCollector::clear_expansion_cause() { 1587 _cmsGen->clear_expansion_cause(); 1588 _permGen->clear_expansion_cause(); 1589 } 1590 1591 // We should be conservative in starting a collection cycle. To 1592 // start too eagerly runs the risk of collecting too often in the 1593 // extreme. To collect too rarely falls back on full collections, 1594 // which works, even if not optimum in terms of concurrent work. 1595 // As a work around for too eagerly collecting, use the flag 1596 // UseCMSInitiatingOccupancyOnly. This also has the advantage of 1597 // giving the user an easily understandable way of controlling the 1598 // collections. 1599 // We want to start a new collection cycle if any of the following 1600 // conditions hold: 1601 // . our current occupancy exceeds the configured initiating occupancy 1602 // for this generation, or 1603 // . we recently needed to expand this space and have not, since that 1604 // expansion, done a collection of this generation, or 1605 // . the underlying space believes that it may be a good idea to initiate 1606 // a concurrent collection (this may be based on criteria such as the 1607 // following: the space uses linear allocation and linear allocation is 1608 // going to fail, or there is believed to be excessive fragmentation in 1609 // the generation, etc... or ... 1610 // [.(currently done by CMSCollector::shouldConcurrentCollect() only for 1611 // the case of the old generation, not the perm generation; see CR 6543076): 1612 // we may be approaching a point at which allocation requests may fail because 1613 // we will be out of sufficient free space given allocation rate estimates.] 1614 bool ConcurrentMarkSweepGeneration::should_concurrent_collect() const { 1615 1616 assert_lock_strong(freelistLock()); 1617 if (occupancy() > initiating_occupancy()) { 1618 if (PrintGCDetails && Verbose) { 1619 gclog_or_tty->print(" %s: collect because of occupancy %f / %f ", 1620 short_name(), occupancy(), initiating_occupancy()); 1621 } 1622 return true; 1623 } 1624 if (UseCMSInitiatingOccupancyOnly) { 1625 return false; 1626 } 1627 if (expansion_cause() == CMSExpansionCause::_satisfy_allocation) { 1628 if (PrintGCDetails && Verbose) { 1629 gclog_or_tty->print(" %s: collect because expanded for allocation ", 1630 short_name()); 1631 } 1632 return true; 1633 } 1634 if (_cmsSpace->should_concurrent_collect()) { 1635 if (PrintGCDetails && Verbose) { 1636 gclog_or_tty->print(" %s: collect because cmsSpace says so ", 1637 short_name()); 1638 } 1639 return true; 1640 } 1641 return false; 1642 } 1643 1644 void ConcurrentMarkSweepGeneration::collect(bool full, 1645 bool clear_all_soft_refs, 1646 size_t size, 1647 bool tlab) 1648 { 1649 collector()->collect(full, clear_all_soft_refs, size, tlab); 1650 } 1651 1652 void CMSCollector::collect(bool full, 1653 bool clear_all_soft_refs, 1654 size_t size, 1655 bool tlab) 1656 { 1657 if (!UseCMSCollectionPassing && _collectorState > Idling) { 1658 // For debugging purposes skip the collection if the state 1659 // is not currently idle 1660 if (TraceCMSState) { 1661 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " skipped full:%d CMS state %d", 1662 Thread::current(), full, _collectorState); 1663 } 1664 return; 1665 } 1666 1667 // The following "if" branch is present for defensive reasons. 1668 // In the current uses of this interface, it can be replaced with: 1669 // assert(!GC_locker.is_active(), "Can't be called otherwise"); 1670 // But I am not placing that assert here to allow future 1671 // generality in invoking this interface. 1672 if (GC_locker::is_active()) { 1673 // A consistency test for GC_locker 1674 assert(GC_locker::needs_gc(), "Should have been set already"); 1675 // Skip this foreground collection, instead 1676 // expanding the heap if necessary. 1677 // Need the free list locks for the call to free() in compute_new_size() 1678 compute_new_size(); 1679 return; 1680 } 1681 acquire_control_and_collect(full, clear_all_soft_refs); 1682 _full_gcs_since_conc_gc++; 1683 1684 } 1685 1686 void CMSCollector::request_full_gc(unsigned int full_gc_count) { 1687 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1688 unsigned int gc_count = gch->total_full_collections(); 1689 if (gc_count == full_gc_count) { 1690 MutexLockerEx y(CGC_lock, Mutex::_no_safepoint_check_flag); 1691 _full_gc_requested = true; 1692 CGC_lock->notify(); // nudge CMS thread 1693 } else { 1694 assert(gc_count > full_gc_count, "Error: causal loop"); 1695 } 1696 } 1697 1698 1699 // The foreground and background collectors need to coordinate in order 1700 // to make sure that they do not mutually interfere with CMS collections. 1701 // When a background collection is active, 1702 // the foreground collector may need to take over (preempt) and 1703 // synchronously complete an ongoing collection. Depending on the 1704 // frequency of the background collections and the heap usage 1705 // of the application, this preemption can be seldom or frequent. 1706 // There are only certain 1707 // points in the background collection that the "collection-baton" 1708 // can be passed to the foreground collector. 1709 // 1710 // The foreground collector will wait for the baton before 1711 // starting any part of the collection. The foreground collector 1712 // will only wait at one location. 1713 // 1714 // The background collector will yield the baton before starting a new 1715 // phase of the collection (e.g., before initial marking, marking from roots, 1716 // precleaning, final re-mark, sweep etc.) This is normally done at the head 1717 // of the loop which switches the phases. The background collector does some 1718 // of the phases (initial mark, final re-mark) with the world stopped. 1719 // Because of locking involved in stopping the world, 1720 // the foreground collector should not block waiting for the background 1721 // collector when it is doing a stop-the-world phase. The background 1722 // collector will yield the baton at an additional point just before 1723 // it enters a stop-the-world phase. Once the world is stopped, the 1724 // background collector checks the phase of the collection. If the 1725 // phase has not changed, it proceeds with the collection. If the 1726 // phase has changed, it skips that phase of the collection. See 1727 // the comments on the use of the Heap_lock in collect_in_background(). 1728 // 1729 // Variable used in baton passing. 1730 // _foregroundGCIsActive - Set to true by the foreground collector when 1731 // it wants the baton. The foreground clears it when it has finished 1732 // the collection. 1733 // _foregroundGCShouldWait - Set to true by the background collector 1734 // when it is running. The foreground collector waits while 1735 // _foregroundGCShouldWait is true. 1736 // CGC_lock - monitor used to protect access to the above variables 1737 // and to notify the foreground and background collectors. 1738 // _collectorState - current state of the CMS collection. 1739 // 1740 // The foreground collector 1741 // acquires the CGC_lock 1742 // sets _foregroundGCIsActive 1743 // waits on the CGC_lock for _foregroundGCShouldWait to be false 1744 // various locks acquired in preparation for the collection 1745 // are released so as not to block the background collector 1746 // that is in the midst of a collection 1747 // proceeds with the collection 1748 // clears _foregroundGCIsActive 1749 // returns 1750 // 1751 // The background collector in a loop iterating on the phases of the 1752 // collection 1753 // acquires the CGC_lock 1754 // sets _foregroundGCShouldWait 1755 // if _foregroundGCIsActive is set 1756 // clears _foregroundGCShouldWait, notifies _CGC_lock 1757 // waits on _CGC_lock for _foregroundGCIsActive to become false 1758 // and exits the loop. 1759 // otherwise 1760 // proceed with that phase of the collection 1761 // if the phase is a stop-the-world phase, 1762 // yield the baton once more just before enqueueing 1763 // the stop-world CMS operation (executed by the VM thread). 1764 // returns after all phases of the collection are done 1765 // 1766 1767 void CMSCollector::acquire_control_and_collect(bool full, 1768 bool clear_all_soft_refs) { 1769 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint"); 1770 assert(!Thread::current()->is_ConcurrentGC_thread(), 1771 "shouldn't try to acquire control from self!"); 1772 1773 // Start the protocol for acquiring control of the 1774 // collection from the background collector (aka CMS thread). 1775 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 1776 "VM thread should have CMS token"); 1777 // Remember the possibly interrupted state of an ongoing 1778 // concurrent collection 1779 CollectorState first_state = _collectorState; 1780 1781 // Signal to a possibly ongoing concurrent collection that 1782 // we want to do a foreground collection. 1783 _foregroundGCIsActive = true; 1784 1785 // Disable incremental mode during a foreground collection. 1786 ICMSDisabler icms_disabler; 1787 1788 // release locks and wait for a notify from the background collector 1789 // releasing the locks in only necessary for phases which 1790 // do yields to improve the granularity of the collection. 1791 assert_lock_strong(bitMapLock()); 1792 // We need to lock the Free list lock for the space that we are 1793 // currently collecting. 1794 assert(haveFreelistLocks(), "Must be holding free list locks"); 1795 bitMapLock()->unlock(); 1796 releaseFreelistLocks(); 1797 { 1798 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1799 if (_foregroundGCShouldWait) { 1800 // We are going to be waiting for action for the CMS thread; 1801 // it had better not be gone (for instance at shutdown)! 1802 assert(ConcurrentMarkSweepThread::cmst() != NULL, 1803 "CMS thread must be running"); 1804 // Wait here until the background collector gives us the go-ahead 1805 ConcurrentMarkSweepThread::clear_CMS_flag( 1806 ConcurrentMarkSweepThread::CMS_vm_has_token); // release token 1807 // Get a possibly blocked CMS thread going: 1808 // Note that we set _foregroundGCIsActive true above, 1809 // without protection of the CGC_lock. 1810 CGC_lock->notify(); 1811 assert(!ConcurrentMarkSweepThread::vm_thread_wants_cms_token(), 1812 "Possible deadlock"); 1813 while (_foregroundGCShouldWait) { 1814 // wait for notification 1815 CGC_lock->wait(Mutex::_no_safepoint_check_flag); 1816 // Possibility of delay/starvation here, since CMS token does 1817 // not know to give priority to VM thread? Actually, i think 1818 // there wouldn't be any delay/starvation, but the proof of 1819 // that "fact" (?) appears non-trivial. XXX 20011219YSR 1820 } 1821 ConcurrentMarkSweepThread::set_CMS_flag( 1822 ConcurrentMarkSweepThread::CMS_vm_has_token); 1823 } 1824 } 1825 // The CMS_token is already held. Get back the other locks. 1826 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 1827 "VM thread should have CMS token"); 1828 getFreelistLocks(); 1829 bitMapLock()->lock_without_safepoint_check(); 1830 if (TraceCMSState) { 1831 gclog_or_tty->print_cr("CMS foreground collector has asked for control " 1832 INTPTR_FORMAT " with first state %d", Thread::current(), first_state); 1833 gclog_or_tty->print_cr(" gets control with state %d", _collectorState); 1834 } 1835 1836 // Check if we need to do a compaction, or if not, whether 1837 // we need to start the mark-sweep from scratch. 1838 bool should_compact = false; 1839 bool should_start_over = false; 1840 decide_foreground_collection_type(clear_all_soft_refs, 1841 &should_compact, &should_start_over); 1842 1843 NOT_PRODUCT( 1844 if (RotateCMSCollectionTypes) { 1845 if (_cmsGen->debug_collection_type() == 1846 ConcurrentMarkSweepGeneration::MSC_foreground_collection_type) { 1847 should_compact = true; 1848 } else if (_cmsGen->debug_collection_type() == 1849 ConcurrentMarkSweepGeneration::MS_foreground_collection_type) { 1850 should_compact = false; 1851 } 1852 } 1853 ) 1854 1855 if (PrintGCDetails && first_state > Idling) { 1856 GCCause::Cause cause = GenCollectedHeap::heap()->gc_cause(); 1857 if (GCCause::is_user_requested_gc(cause) || 1858 GCCause::is_serviceability_requested_gc(cause)) { 1859 gclog_or_tty->print(" (concurrent mode interrupted)"); 1860 } else { 1861 gclog_or_tty->print(" (concurrent mode failure)"); 1862 } 1863 } 1864 1865 if (should_compact) { 1866 // If the collection is being acquired from the background 1867 // collector, there may be references on the discovered 1868 // references lists that have NULL referents (being those 1869 // that were concurrently cleared by a mutator) or 1870 // that are no longer active (having been enqueued concurrently 1871 // by the mutator). 1872 // Scrub the list of those references because Mark-Sweep-Compact 1873 // code assumes referents are not NULL and that all discovered 1874 // Reference objects are active. 1875 ref_processor()->clean_up_discovered_references(); 1876 1877 do_compaction_work(clear_all_soft_refs); 1878 1879 // Has the GC time limit been exceeded? 1880 DefNewGeneration* young_gen = _young_gen->as_DefNewGeneration(); 1881 size_t max_eden_size = young_gen->max_capacity() - 1882 young_gen->to()->capacity() - 1883 young_gen->from()->capacity(); 1884 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1885 GCCause::Cause gc_cause = gch->gc_cause(); 1886 size_policy()->check_gc_overhead_limit(_young_gen->used(), 1887 young_gen->eden()->used(), 1888 _cmsGen->max_capacity(), 1889 max_eden_size, 1890 full, 1891 gc_cause, 1892 gch->collector_policy()); 1893 } else { 1894 do_mark_sweep_work(clear_all_soft_refs, first_state, 1895 should_start_over); 1896 } 1897 // Reset the expansion cause, now that we just completed 1898 // a collection cycle. 1899 clear_expansion_cause(); 1900 _foregroundGCIsActive = false; 1901 return; 1902 } 1903 1904 // Resize the perm generation and the tenured generation 1905 // after obtaining the free list locks for the 1906 // two generations. 1907 void CMSCollector::compute_new_size() { 1908 assert_locked_or_safepoint(Heap_lock); 1909 FreelistLocker z(this); 1910 _permGen->compute_new_size(); 1911 _cmsGen->compute_new_size(); 1912 } 1913 1914 // A work method used by foreground collection to determine 1915 // what type of collection (compacting or not, continuing or fresh) 1916 // it should do. 1917 // NOTE: the intent is to make UseCMSCompactAtFullCollection 1918 // and CMSCompactWhenClearAllSoftRefs the default in the future 1919 // and do away with the flags after a suitable period. 1920 void CMSCollector::decide_foreground_collection_type( 1921 bool clear_all_soft_refs, bool* should_compact, 1922 bool* should_start_over) { 1923 // Normally, we'll compact only if the UseCMSCompactAtFullCollection 1924 // flag is set, and we have either requested a System.gc() or 1925 // the number of full gc's since the last concurrent cycle 1926 // has exceeded the threshold set by CMSFullGCsBeforeCompaction, 1927 // or if an incremental collection has failed 1928 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1929 assert(gch->collector_policy()->is_two_generation_policy(), 1930 "You may want to check the correctness of the following"); 1931 // Inform cms gen if this was due to partial collection failing. 1932 // The CMS gen may use this fact to determine its expansion policy. 1933 if (gch->incremental_collection_will_fail(false /* don't consult_young */)) { 1934 assert(!_cmsGen->incremental_collection_failed(), 1935 "Should have been noticed, reacted to and cleared"); 1936 _cmsGen->set_incremental_collection_failed(); 1937 } 1938 *should_compact = 1939 UseCMSCompactAtFullCollection && 1940 ((_full_gcs_since_conc_gc >= CMSFullGCsBeforeCompaction) || 1941 GCCause::is_user_requested_gc(gch->gc_cause()) || 1942 gch->incremental_collection_will_fail(true /* consult_young */)); 1943 *should_start_over = false; 1944 if (clear_all_soft_refs && !*should_compact) { 1945 // We are about to do a last ditch collection attempt 1946 // so it would normally make sense to do a compaction 1947 // to reclaim as much space as possible. 1948 if (CMSCompactWhenClearAllSoftRefs) { 1949 // Default: The rationale is that in this case either 1950 // we are past the final marking phase, in which case 1951 // we'd have to start over, or so little has been done 1952 // that there's little point in saving that work. Compaction 1953 // appears to be the sensible choice in either case. 1954 *should_compact = true; 1955 } else { 1956 // We have been asked to clear all soft refs, but not to 1957 // compact. Make sure that we aren't past the final checkpoint 1958 // phase, for that is where we process soft refs. If we are already 1959 // past that phase, we'll need to redo the refs discovery phase and 1960 // if necessary clear soft refs that weren't previously 1961 // cleared. We do so by remembering the phase in which 1962 // we came in, and if we are past the refs processing 1963 // phase, we'll choose to just redo the mark-sweep 1964 // collection from scratch. 1965 if (_collectorState > FinalMarking) { 1966 // We are past the refs processing phase; 1967 // start over and do a fresh synchronous CMS cycle 1968 _collectorState = Resetting; // skip to reset to start new cycle 1969 reset(false /* == !asynch */); 1970 *should_start_over = true; 1971 } // else we can continue a possibly ongoing current cycle 1972 } 1973 } 1974 } 1975 1976 // A work method used by the foreground collector to do 1977 // a mark-sweep-compact. 1978 void CMSCollector::do_compaction_work(bool clear_all_soft_refs) { 1979 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1980 1981 STWGCTimer* gc_timer = GenMarkSweep::gc_timer(); 1982 gc_timer->register_gc_start(os::elapsed_counter()); 1983 1984 SerialOldTracer* gc_tracer = GenMarkSweep::gc_tracer(); 1985 gc_tracer->report_gc_start(gch->gc_cause(), gc_timer->gc_start()); 1986 1987 GCTraceTime t("CMS:MSC ", PrintGCDetails && Verbose, true, NULL); 1988 if (PrintGC && Verbose && !(GCCause::is_user_requested_gc(gch->gc_cause()))) { 1989 gclog_or_tty->print_cr("Compact ConcurrentMarkSweepGeneration after %d " 1990 "collections passed to foreground collector", _full_gcs_since_conc_gc); 1991 } 1992 1993 // Sample collection interval time and reset for collection pause. 1994 if (UseAdaptiveSizePolicy) { 1995 size_policy()->msc_collection_begin(); 1996 } 1997 1998 // Temporarily widen the span of the weak reference processing to 1999 // the entire heap. 2000 MemRegion new_span(GenCollectedHeap::heap()->reserved_region()); 2001 ReferenceProcessorSpanMutator rp_mut_span(ref_processor(), new_span); 2002 // Temporarily, clear the "is_alive_non_header" field of the 2003 // reference processor. 2004 ReferenceProcessorIsAliveMutator rp_mut_closure(ref_processor(), NULL); 2005 // Temporarily make reference _processing_ single threaded (non-MT). 2006 ReferenceProcessorMTProcMutator rp_mut_mt_processing(ref_processor(), false); 2007 // Temporarily make refs discovery atomic 2008 ReferenceProcessorAtomicMutator rp_mut_atomic(ref_processor(), true); 2009 // Temporarily make reference _discovery_ single threaded (non-MT) 2010 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false); 2011 2012 ref_processor()->set_enqueuing_is_done(false); 2013 ref_processor()->enable_discovery(false /*verify_disabled*/, false /*check_no_refs*/); 2014 ref_processor()->setup_policy(clear_all_soft_refs); 2015 // If an asynchronous collection finishes, the _modUnionTable is 2016 // all clear. If we are assuming the collection from an asynchronous 2017 // collection, clear the _modUnionTable. 2018 assert(_collectorState != Idling || _modUnionTable.isAllClear(), 2019 "_modUnionTable should be clear if the baton was not passed"); 2020 _modUnionTable.clear_all(); 2021 2022 // We must adjust the allocation statistics being maintained 2023 // in the free list space. We do so by reading and clearing 2024 // the sweep timer and updating the block flux rate estimates below. 2025 assert(!_intra_sweep_timer.is_active(), "_intra_sweep_timer should be inactive"); 2026 if (_inter_sweep_timer.is_active()) { 2027 _inter_sweep_timer.stop(); 2028 // Note that we do not use this sample to update the _inter_sweep_estimate. 2029 _cmsGen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()), 2030 _inter_sweep_estimate.padded_average(), 2031 _intra_sweep_estimate.padded_average()); 2032 } 2033 2034 GenMarkSweep::invoke_at_safepoint(_cmsGen->level(), 2035 ref_processor(), clear_all_soft_refs); 2036 #ifdef ASSERT 2037 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 2038 size_t free_size = cms_space->free(); 2039 assert(free_size == 2040 pointer_delta(cms_space->end(), cms_space->compaction_top()) 2041 * HeapWordSize, 2042 "All the free space should be compacted into one chunk at top"); 2043 assert(cms_space->dictionary()->total_chunk_size( 2044 debug_only(cms_space->freelistLock())) == 0 || 2045 cms_space->totalSizeInIndexedFreeLists() == 0, 2046 "All the free space should be in a single chunk"); 2047 size_t num = cms_space->totalCount(); 2048 assert((free_size == 0 && num == 0) || 2049 (free_size > 0 && (num == 1 || num == 2)), 2050 "There should be at most 2 free chunks after compaction"); 2051 #endif // ASSERT 2052 _collectorState = Resetting; 2053 assert(_restart_addr == NULL, 2054 "Should have been NULL'd before baton was passed"); 2055 reset(false /* == !asynch */); 2056 _cmsGen->reset_after_compaction(); 2057 _concurrent_cycles_since_last_unload = 0; 2058 2059 if (verifying() && !should_unload_classes()) { 2060 perm_gen_verify_bit_map()->clear_all(); 2061 } 2062 2063 // Clear any data recorded in the PLAB chunk arrays. 2064 if (_survivor_plab_array != NULL) { 2065 reset_survivor_plab_arrays(); 2066 } 2067 2068 // Adjust the per-size allocation stats for the next epoch. 2069 _cmsGen->cmsSpace()->endSweepFLCensus(sweep_count() /* fake */); 2070 // Restart the "inter sweep timer" for the next epoch. 2071 _inter_sweep_timer.reset(); 2072 _inter_sweep_timer.start(); 2073 2074 // Sample collection pause time and reset for collection interval. 2075 if (UseAdaptiveSizePolicy) { 2076 size_policy()->msc_collection_end(gch->gc_cause()); 2077 } 2078 2079 gc_timer->register_gc_end(os::elapsed_counter()); 2080 2081 gc_tracer->report_gc_end(gc_timer->gc_end(), gc_timer->time_partitions()); 2082 2083 // For a mark-sweep-compact, compute_new_size() will be called 2084 // in the heap's do_collection() method. 2085 } 2086 2087 // A work method used by the foreground collector to do 2088 // a mark-sweep, after taking over from a possibly on-going 2089 // concurrent mark-sweep collection. 2090 void CMSCollector::do_mark_sweep_work(bool clear_all_soft_refs, 2091 CollectorState first_state, bool should_start_over) { 2092 if (PrintGC && Verbose) { 2093 gclog_or_tty->print_cr("Pass concurrent collection to foreground " 2094 "collector with count %d", 2095 _full_gcs_since_conc_gc); 2096 } 2097 switch (_collectorState) { 2098 case Idling: 2099 if (first_state == Idling || should_start_over) { 2100 // The background GC was not active, or should 2101 // restarted from scratch; start the cycle. 2102 _collectorState = InitialMarking; 2103 } 2104 // If first_state was not Idling, then a background GC 2105 // was in progress and has now finished. No need to do it 2106 // again. Leave the state as Idling. 2107 break; 2108 case Precleaning: 2109 // In the foreground case don't do the precleaning since 2110 // it is not done concurrently and there is extra work 2111 // required. 2112 _collectorState = FinalMarking; 2113 } 2114 collect_in_foreground(clear_all_soft_refs); 2115 2116 // For a mark-sweep, compute_new_size() will be called 2117 // in the heap's do_collection() method. 2118 } 2119 2120 2121 void CMSCollector::getFreelistLocks() const { 2122 // Get locks for all free lists in all generations that this 2123 // collector is responsible for 2124 _cmsGen->freelistLock()->lock_without_safepoint_check(); 2125 _permGen->freelistLock()->lock_without_safepoint_check(); 2126 } 2127 2128 void CMSCollector::releaseFreelistLocks() const { 2129 // Release locks for all free lists in all generations that this 2130 // collector is responsible for 2131 _cmsGen->freelistLock()->unlock(); 2132 _permGen->freelistLock()->unlock(); 2133 } 2134 2135 bool CMSCollector::haveFreelistLocks() const { 2136 // Check locks for all free lists in all generations that this 2137 // collector is responsible for 2138 assert_lock_strong(_cmsGen->freelistLock()); 2139 assert_lock_strong(_permGen->freelistLock()); 2140 PRODUCT_ONLY(ShouldNotReachHere()); 2141 return true; 2142 } 2143 2144 // A utility class that is used by the CMS collector to 2145 // temporarily "release" the foreground collector from its 2146 // usual obligation to wait for the background collector to 2147 // complete an ongoing phase before proceeding. 2148 class ReleaseForegroundGC: public StackObj { 2149 private: 2150 CMSCollector* _c; 2151 public: 2152 ReleaseForegroundGC(CMSCollector* c) : _c(c) { 2153 assert(_c->_foregroundGCShouldWait, "Else should not need to call"); 2154 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 2155 // allow a potentially blocked foreground collector to proceed 2156 _c->_foregroundGCShouldWait = false; 2157 if (_c->_foregroundGCIsActive) { 2158 CGC_lock->notify(); 2159 } 2160 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 2161 "Possible deadlock"); 2162 } 2163 2164 ~ReleaseForegroundGC() { 2165 assert(!_c->_foregroundGCShouldWait, "Usage protocol violation?"); 2166 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 2167 _c->_foregroundGCShouldWait = true; 2168 } 2169 }; 2170 2171 // There are separate collect_in_background and collect_in_foreground because of 2172 // the different locking requirements of the background collector and the 2173 // foreground collector. There was originally an attempt to share 2174 // one "collect" method between the background collector and the foreground 2175 // collector but the if-then-else required made it cleaner to have 2176 // separate methods. 2177 void CMSCollector::collect_in_background(bool clear_all_soft_refs) { 2178 assert(Thread::current()->is_ConcurrentGC_thread(), 2179 "A CMS asynchronous collection is only allowed on a CMS thread."); 2180 2181 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2182 { 2183 bool safepoint_check = Mutex::_no_safepoint_check_flag; 2184 MutexLockerEx hl(Heap_lock, safepoint_check); 2185 FreelistLocker fll(this); 2186 MutexLockerEx x(CGC_lock, safepoint_check); 2187 if (_foregroundGCIsActive || !UseAsyncConcMarkSweepGC) { 2188 // The foreground collector is active or we're 2189 // not using asynchronous collections. Skip this 2190 // background collection. 2191 assert(!_foregroundGCShouldWait, "Should be clear"); 2192 return; 2193 } else { 2194 assert(_collectorState == Idling, "Should be idling before start."); 2195 _collectorState = InitialMarking; 2196 // Reset the expansion cause, now that we are about to begin 2197 // a new cycle. 2198 clear_expansion_cause(); 2199 } 2200 // Decide if we want to enable class unloading as part of the 2201 // ensuing concurrent GC cycle. 2202 update_should_unload_classes(); 2203 _full_gc_requested = false; // acks all outstanding full gc requests 2204 // Signal that we are about to start a collection 2205 gch->increment_total_full_collections(); // ... starting a collection cycle 2206 _collection_count_start = gch->total_full_collections(); 2207 } 2208 2209 // Used for PrintGC 2210 size_t prev_used; 2211 if (PrintGC && Verbose) { 2212 prev_used = _cmsGen->used(); // XXXPERM 2213 } 2214 2215 // The change of the collection state is normally done at this level; 2216 // the exceptions are phases that are executed while the world is 2217 // stopped. For those phases the change of state is done while the 2218 // world is stopped. For baton passing purposes this allows the 2219 // background collector to finish the phase and change state atomically. 2220 // The foreground collector cannot wait on a phase that is done 2221 // while the world is stopped because the foreground collector already 2222 // has the world stopped and would deadlock. 2223 while (_collectorState != Idling) { 2224 if (TraceCMSState) { 2225 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " in CMS state %d", 2226 Thread::current(), _collectorState); 2227 } 2228 // The foreground collector 2229 // holds the Heap_lock throughout its collection. 2230 // holds the CMS token (but not the lock) 2231 // except while it is waiting for the background collector to yield. 2232 // 2233 // The foreground collector should be blocked (not for long) 2234 // if the background collector is about to start a phase 2235 // executed with world stopped. If the background 2236 // collector has already started such a phase, the 2237 // foreground collector is blocked waiting for the 2238 // Heap_lock. The stop-world phases (InitialMarking and FinalMarking) 2239 // are executed in the VM thread. 2240 // 2241 // The locking order is 2242 // PendingListLock (PLL) -- if applicable (FinalMarking) 2243 // Heap_lock (both this & PLL locked in VM_CMS_Operation::prologue()) 2244 // CMS token (claimed in 2245 // stop_world_and_do() --> 2246 // safepoint_synchronize() --> 2247 // CMSThread::synchronize()) 2248 2249 { 2250 // Check if the FG collector wants us to yield. 2251 CMSTokenSync x(true); // is cms thread 2252 if (waitForForegroundGC()) { 2253 // We yielded to a foreground GC, nothing more to be 2254 // done this round. 2255 assert(_foregroundGCShouldWait == false, "We set it to false in " 2256 "waitForForegroundGC()"); 2257 if (TraceCMSState) { 2258 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT 2259 " exiting collection CMS state %d", 2260 Thread::current(), _collectorState); 2261 } 2262 return; 2263 } else { 2264 // The background collector can run but check to see if the 2265 // foreground collector has done a collection while the 2266 // background collector was waiting to get the CGC_lock 2267 // above. If yes, break so that _foregroundGCShouldWait 2268 // is cleared before returning. 2269 if (_collectorState == Idling) { 2270 break; 2271 } 2272 } 2273 } 2274 2275 assert(_foregroundGCShouldWait, "Foreground collector, if active, " 2276 "should be waiting"); 2277 2278 switch (_collectorState) { 2279 case InitialMarking: 2280 { 2281 ReleaseForegroundGC x(this); 2282 stats().record_cms_begin(); 2283 VM_CMS_Initial_Mark initial_mark_op(this); 2284 VMThread::execute(&initial_mark_op); 2285 } 2286 // The collector state may be any legal state at this point 2287 // since the background collector may have yielded to the 2288 // foreground collector. 2289 break; 2290 case Marking: 2291 // initial marking in checkpointRootsInitialWork has been completed 2292 if (markFromRoots(true)) { // we were successful 2293 assert(_collectorState == Precleaning, "Collector state should " 2294 "have changed"); 2295 } else { 2296 assert(_foregroundGCIsActive, "Internal state inconsistency"); 2297 } 2298 break; 2299 case Precleaning: 2300 if (UseAdaptiveSizePolicy) { 2301 size_policy()->concurrent_precleaning_begin(); 2302 } 2303 // marking from roots in markFromRoots has been completed 2304 preclean(); 2305 if (UseAdaptiveSizePolicy) { 2306 size_policy()->concurrent_precleaning_end(); 2307 } 2308 assert(_collectorState == AbortablePreclean || 2309 _collectorState == FinalMarking, 2310 "Collector state should have changed"); 2311 break; 2312 case AbortablePreclean: 2313 if (UseAdaptiveSizePolicy) { 2314 size_policy()->concurrent_phases_resume(); 2315 } 2316 abortable_preclean(); 2317 if (UseAdaptiveSizePolicy) { 2318 size_policy()->concurrent_precleaning_end(); 2319 } 2320 assert(_collectorState == FinalMarking, "Collector state should " 2321 "have changed"); 2322 break; 2323 case FinalMarking: 2324 { 2325 ReleaseForegroundGC x(this); 2326 2327 VM_CMS_Final_Remark final_remark_op(this); 2328 VMThread::execute(&final_remark_op); 2329 } 2330 assert(_foregroundGCShouldWait, "block post-condition"); 2331 break; 2332 case Sweeping: 2333 if (UseAdaptiveSizePolicy) { 2334 size_policy()->concurrent_sweeping_begin(); 2335 } 2336 // final marking in checkpointRootsFinal has been completed 2337 sweep(true); 2338 assert(_collectorState == Resizing, "Collector state change " 2339 "to Resizing must be done under the free_list_lock"); 2340 _full_gcs_since_conc_gc = 0; 2341 2342 // Stop the timers for adaptive size policy for the concurrent phases 2343 if (UseAdaptiveSizePolicy) { 2344 size_policy()->concurrent_sweeping_end(); 2345 size_policy()->concurrent_phases_end(gch->gc_cause(), 2346 gch->prev_gen(_cmsGen)->capacity(), 2347 _cmsGen->free()); 2348 } 2349 2350 case Resizing: { 2351 // Sweeping has been completed... 2352 // At this point the background collection has completed. 2353 // Don't move the call to compute_new_size() down 2354 // into code that might be executed if the background 2355 // collection was preempted. 2356 { 2357 ReleaseForegroundGC x(this); // unblock FG collection 2358 MutexLockerEx y(Heap_lock, Mutex::_no_safepoint_check_flag); 2359 CMSTokenSync z(true); // not strictly needed. 2360 if (_collectorState == Resizing) { 2361 compute_new_size(); 2362 _collectorState = Resetting; 2363 } else { 2364 assert(_collectorState == Idling, "The state should only change" 2365 " because the foreground collector has finished the collection"); 2366 } 2367 } 2368 break; 2369 } 2370 case Resetting: 2371 // CMS heap resizing has been completed 2372 reset(true); 2373 assert(_collectorState == Idling, "Collector state should " 2374 "have changed"); 2375 stats().record_cms_end(); 2376 // Don't move the concurrent_phases_end() and compute_new_size() 2377 // calls to here because a preempted background collection 2378 // has it's state set to "Resetting". 2379 break; 2380 case Idling: 2381 default: 2382 ShouldNotReachHere(); 2383 break; 2384 } 2385 if (TraceCMSState) { 2386 gclog_or_tty->print_cr(" Thread " INTPTR_FORMAT " done - next CMS state %d", 2387 Thread::current(), _collectorState); 2388 } 2389 assert(_foregroundGCShouldWait, "block post-condition"); 2390 } 2391 2392 // Should this be in gc_epilogue? 2393 collector_policy()->counters()->update_counters(); 2394 2395 { 2396 // Clear _foregroundGCShouldWait and, in the event that the 2397 // foreground collector is waiting, notify it, before 2398 // returning. 2399 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 2400 _foregroundGCShouldWait = false; 2401 if (_foregroundGCIsActive) { 2402 CGC_lock->notify(); 2403 } 2404 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 2405 "Possible deadlock"); 2406 } 2407 if (TraceCMSState) { 2408 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT 2409 " exiting collection CMS state %d", 2410 Thread::current(), _collectorState); 2411 } 2412 if (PrintGC && Verbose) { 2413 _cmsGen->print_heap_change(prev_used); 2414 } 2415 } 2416 2417 void CMSCollector::register_gc_start(GCCause::Cause cause) { 2418 _cms_start_registered = true; 2419 _gc_timer_cm->register_gc_start(os::elapsed_counter()); 2420 _gc_tracer_cm->report_gc_start(cause, _gc_timer_cm->gc_start()); 2421 } 2422 2423 void CMSCollector::register_gc_end() { 2424 if (_cms_start_registered) { 2425 _gc_timer_cm->register_gc_end(os::elapsed_counter()); 2426 _gc_tracer_cm->report_gc_end(_gc_timer_cm->gc_end(), _gc_timer_cm->time_partitions()); 2427 _cms_start_registered = false; 2428 } 2429 } 2430 2431 void CMSCollector::collect_in_foreground(bool clear_all_soft_refs) { 2432 assert(_foregroundGCIsActive && !_foregroundGCShouldWait, 2433 "Foreground collector should be waiting, not executing"); 2434 assert(Thread::current()->is_VM_thread(), "A foreground collection" 2435 "may only be done by the VM Thread with the world stopped"); 2436 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 2437 "VM thread should have CMS token"); 2438 2439 NOT_PRODUCT(GCTraceTime t("CMS:MS (foreground) ", PrintGCDetails && Verbose, 2440 true, NULL);) 2441 if (UseAdaptiveSizePolicy) { 2442 size_policy()->ms_collection_begin(); 2443 } 2444 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact); 2445 2446 HandleMark hm; // Discard invalid handles created during verification 2447 2448 if (VerifyBeforeGC && 2449 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2450 Universe::verify(); 2451 } 2452 2453 // Snapshot the soft reference policy to be used in this collection cycle. 2454 ref_processor()->setup_policy(clear_all_soft_refs); 2455 2456 bool init_mark_was_synchronous = false; // until proven otherwise 2457 while (_collectorState != Idling) { 2458 if (TraceCMSState) { 2459 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " in CMS state %d", 2460 Thread::current(), _collectorState); 2461 } 2462 switch (_collectorState) { 2463 case InitialMarking: 2464 register_gc_start(GenCollectedHeap::heap()->gc_cause()); 2465 init_mark_was_synchronous = true; // fact to be exploited in re-mark 2466 checkpointRootsInitial(false); 2467 assert(_collectorState == Marking, "Collector state should have changed" 2468 " within checkpointRootsInitial()"); 2469 break; 2470 case Marking: 2471 // initial marking in checkpointRootsInitialWork has been completed 2472 if (VerifyDuringGC && 2473 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2474 gclog_or_tty->print("Verify before initial mark: "); 2475 Universe::verify(); 2476 } 2477 { 2478 bool res = markFromRoots(false); 2479 assert(res && _collectorState == FinalMarking, "Collector state should " 2480 "have changed"); 2481 break; 2482 } 2483 case FinalMarking: 2484 if (VerifyDuringGC && 2485 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2486 gclog_or_tty->print("Verify before re-mark: "); 2487 Universe::verify(); 2488 } 2489 checkpointRootsFinal(false, clear_all_soft_refs, 2490 init_mark_was_synchronous); 2491 assert(_collectorState == Sweeping, "Collector state should not " 2492 "have changed within checkpointRootsFinal()"); 2493 break; 2494 case Sweeping: 2495 // final marking in checkpointRootsFinal has been completed 2496 if (VerifyDuringGC && 2497 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2498 gclog_or_tty->print("Verify before sweep: "); 2499 Universe::verify(); 2500 } 2501 sweep(false); 2502 assert(_collectorState == Resizing, "Incorrect state"); 2503 break; 2504 case Resizing: { 2505 // Sweeping has been completed; the actual resize in this case 2506 // is done separately; nothing to be done in this state. 2507 _collectorState = Resetting; 2508 break; 2509 } 2510 case Resetting: 2511 // The heap has been resized. 2512 if (VerifyDuringGC && 2513 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2514 gclog_or_tty->print("Verify before reset: "); 2515 Universe::verify(); 2516 } 2517 reset(false); 2518 assert(_collectorState == Idling, "Collector state should " 2519 "have changed"); 2520 break; 2521 case Precleaning: 2522 case AbortablePreclean: 2523 // Elide the preclean phase 2524 _collectorState = FinalMarking; 2525 break; 2526 default: 2527 ShouldNotReachHere(); 2528 } 2529 if (TraceCMSState) { 2530 gclog_or_tty->print_cr(" Thread " INTPTR_FORMAT " done - next CMS state %d", 2531 Thread::current(), _collectorState); 2532 } 2533 } 2534 2535 if (UseAdaptiveSizePolicy) { 2536 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2537 size_policy()->ms_collection_end(gch->gc_cause()); 2538 } 2539 2540 if (VerifyAfterGC && 2541 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2542 Universe::verify(); 2543 } 2544 if (TraceCMSState) { 2545 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT 2546 " exiting collection CMS state %d", 2547 Thread::current(), _collectorState); 2548 } 2549 } 2550 2551 bool CMSCollector::waitForForegroundGC() { 2552 bool res = false; 2553 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 2554 "CMS thread should have CMS token"); 2555 // Block the foreground collector until the 2556 // background collectors decides whether to 2557 // yield. 2558 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 2559 _foregroundGCShouldWait = true; 2560 if (_foregroundGCIsActive) { 2561 // The background collector yields to the 2562 // foreground collector and returns a value 2563 // indicating that it has yielded. The foreground 2564 // collector can proceed. 2565 res = true; 2566 _foregroundGCShouldWait = false; 2567 ConcurrentMarkSweepThread::clear_CMS_flag( 2568 ConcurrentMarkSweepThread::CMS_cms_has_token); 2569 ConcurrentMarkSweepThread::set_CMS_flag( 2570 ConcurrentMarkSweepThread::CMS_cms_wants_token); 2571 // Get a possibly blocked foreground thread going 2572 CGC_lock->notify(); 2573 if (TraceCMSState) { 2574 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " waiting at CMS state %d", 2575 Thread::current(), _collectorState); 2576 } 2577 while (_foregroundGCIsActive) { 2578 CGC_lock->wait(Mutex::_no_safepoint_check_flag); 2579 } 2580 ConcurrentMarkSweepThread::set_CMS_flag( 2581 ConcurrentMarkSweepThread::CMS_cms_has_token); 2582 ConcurrentMarkSweepThread::clear_CMS_flag( 2583 ConcurrentMarkSweepThread::CMS_cms_wants_token); 2584 } 2585 if (TraceCMSState) { 2586 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " continuing at CMS state %d", 2587 Thread::current(), _collectorState); 2588 } 2589 return res; 2590 } 2591 2592 // Because of the need to lock the free lists and other structures in 2593 // the collector, common to all the generations that the collector is 2594 // collecting, we need the gc_prologues of individual CMS generations 2595 // delegate to their collector. It may have been simpler had the 2596 // current infrastructure allowed one to call a prologue on a 2597 // collector. In the absence of that we have the generation's 2598 // prologue delegate to the collector, which delegates back 2599 // some "local" work to a worker method in the individual generations 2600 // that it's responsible for collecting, while itself doing any 2601 // work common to all generations it's responsible for. A similar 2602 // comment applies to the gc_epilogue()'s. 2603 // The role of the varaible _between_prologue_and_epilogue is to 2604 // enforce the invocation protocol. 2605 void CMSCollector::gc_prologue(bool full) { 2606 // Call gc_prologue_work() for each CMSGen and PermGen that 2607 // we are responsible for. 2608 2609 // The following locking discipline assumes that we are only called 2610 // when the world is stopped. 2611 assert(SafepointSynchronize::is_at_safepoint(), "world is stopped assumption"); 2612 2613 // The CMSCollector prologue must call the gc_prologues for the 2614 // "generations" (including PermGen if any) that it's responsible 2615 // for. 2616 2617 assert( Thread::current()->is_VM_thread() 2618 || ( CMSScavengeBeforeRemark 2619 && Thread::current()->is_ConcurrentGC_thread()), 2620 "Incorrect thread type for prologue execution"); 2621 2622 if (_between_prologue_and_epilogue) { 2623 // We have already been invoked; this is a gc_prologue delegation 2624 // from yet another CMS generation that we are responsible for, just 2625 // ignore it since all relevant work has already been done. 2626 return; 2627 } 2628 2629 // set a bit saying prologue has been called; cleared in epilogue 2630 _between_prologue_and_epilogue = true; 2631 // Claim locks for common data structures, then call gc_prologue_work() 2632 // for each CMSGen and PermGen that we are responsible for. 2633 2634 getFreelistLocks(); // gets free list locks on constituent spaces 2635 bitMapLock()->lock_without_safepoint_check(); 2636 2637 // Should call gc_prologue_work() for all cms gens we are responsible for 2638 bool registerClosure = _collectorState >= Marking 2639 && _collectorState < Sweeping; 2640 ModUnionClosure* muc = CollectedHeap::use_parallel_gc_threads() ? 2641 &_modUnionClosurePar 2642 : &_modUnionClosure; 2643 _cmsGen->gc_prologue_work(full, registerClosure, muc); 2644 _permGen->gc_prologue_work(full, registerClosure, muc); 2645 2646 if (!full) { 2647 stats().record_gc0_begin(); 2648 } 2649 } 2650 2651 void ConcurrentMarkSweepGeneration::gc_prologue(bool full) { 2652 // Delegate to CMScollector which knows how to coordinate between 2653 // this and any other CMS generations that it is responsible for 2654 // collecting. 2655 collector()->gc_prologue(full); 2656 } 2657 2658 // This is a "private" interface for use by this generation's CMSCollector. 2659 // Not to be called directly by any other entity (for instance, 2660 // GenCollectedHeap, which calls the "public" gc_prologue method above). 2661 void ConcurrentMarkSweepGeneration::gc_prologue_work(bool full, 2662 bool registerClosure, ModUnionClosure* modUnionClosure) { 2663 assert(!incremental_collection_failed(), "Shouldn't be set yet"); 2664 assert(cmsSpace()->preconsumptionDirtyCardClosure() == NULL, 2665 "Should be NULL"); 2666 if (registerClosure) { 2667 cmsSpace()->setPreconsumptionDirtyCardClosure(modUnionClosure); 2668 } 2669 cmsSpace()->gc_prologue(); 2670 // Clear stat counters 2671 NOT_PRODUCT( 2672 assert(_numObjectsPromoted == 0, "check"); 2673 assert(_numWordsPromoted == 0, "check"); 2674 if (Verbose && PrintGC) { 2675 gclog_or_tty->print("Allocated "SIZE_FORMAT" objects, " 2676 SIZE_FORMAT" bytes concurrently", 2677 _numObjectsAllocated, _numWordsAllocated*sizeof(HeapWord)); 2678 } 2679 _numObjectsAllocated = 0; 2680 _numWordsAllocated = 0; 2681 ) 2682 } 2683 2684 void CMSCollector::gc_epilogue(bool full) { 2685 // The following locking discipline assumes that we are only called 2686 // when the world is stopped. 2687 assert(SafepointSynchronize::is_at_safepoint(), 2688 "world is stopped assumption"); 2689 2690 // Currently the CMS epilogue (see CompactibleFreeListSpace) merely checks 2691 // if linear allocation blocks need to be appropriately marked to allow the 2692 // the blocks to be parsable. We also check here whether we need to nudge the 2693 // CMS collector thread to start a new cycle (if it's not already active). 2694 assert( Thread::current()->is_VM_thread() 2695 || ( CMSScavengeBeforeRemark 2696 && Thread::current()->is_ConcurrentGC_thread()), 2697 "Incorrect thread type for epilogue execution"); 2698 2699 if (!_between_prologue_and_epilogue) { 2700 // We have already been invoked; this is a gc_epilogue delegation 2701 // from yet another CMS generation that we are responsible for, just 2702 // ignore it since all relevant work has already been done. 2703 return; 2704 } 2705 assert(haveFreelistLocks(), "must have freelist locks"); 2706 assert_lock_strong(bitMapLock()); 2707 2708 _cmsGen->gc_epilogue_work(full); 2709 _permGen->gc_epilogue_work(full); 2710 2711 if (_collectorState == AbortablePreclean || _collectorState == Precleaning) { 2712 // in case sampling was not already enabled, enable it 2713 _start_sampling = true; 2714 } 2715 // reset _eden_chunk_array so sampling starts afresh 2716 _eden_chunk_index = 0; 2717 2718 size_t cms_used = _cmsGen->cmsSpace()->used(); 2719 size_t perm_used = _permGen->cmsSpace()->used(); 2720 2721 // update performance counters - this uses a special version of 2722 // update_counters() that allows the utilization to be passed as a 2723 // parameter, avoiding multiple calls to used(). 2724 // 2725 _cmsGen->update_counters(cms_used); 2726 _permGen->update_counters(perm_used); 2727 2728 if (CMSIncrementalMode) { 2729 icms_update_allocation_limits(); 2730 } 2731 2732 bitMapLock()->unlock(); 2733 releaseFreelistLocks(); 2734 2735 if (!CleanChunkPoolAsync) { 2736 Chunk::clean_chunk_pool(); 2737 } 2738 2739 _between_prologue_and_epilogue = false; // ready for next cycle 2740 } 2741 2742 void ConcurrentMarkSweepGeneration::gc_epilogue(bool full) { 2743 collector()->gc_epilogue(full); 2744 2745 // Also reset promotion tracking in par gc thread states. 2746 if (CollectedHeap::use_parallel_gc_threads()) { 2747 for (uint i = 0; i < ParallelGCThreads; i++) { 2748 _par_gc_thread_states[i]->promo.stopTrackingPromotions(i); 2749 } 2750 } 2751 } 2752 2753 void ConcurrentMarkSweepGeneration::gc_epilogue_work(bool full) { 2754 assert(!incremental_collection_failed(), "Should have been cleared"); 2755 cmsSpace()->setPreconsumptionDirtyCardClosure(NULL); 2756 cmsSpace()->gc_epilogue(); 2757 // Print stat counters 2758 NOT_PRODUCT( 2759 assert(_numObjectsAllocated == 0, "check"); 2760 assert(_numWordsAllocated == 0, "check"); 2761 if (Verbose && PrintGC) { 2762 gclog_or_tty->print("Promoted "SIZE_FORMAT" objects, " 2763 SIZE_FORMAT" bytes", 2764 _numObjectsPromoted, _numWordsPromoted*sizeof(HeapWord)); 2765 } 2766 _numObjectsPromoted = 0; 2767 _numWordsPromoted = 0; 2768 ) 2769 2770 if (PrintGC && Verbose) { 2771 // Call down the chain in contiguous_available needs the freelistLock 2772 // so print this out before releasing the freeListLock. 2773 gclog_or_tty->print(" Contiguous available "SIZE_FORMAT" bytes ", 2774 contiguous_available()); 2775 } 2776 } 2777 2778 #ifndef PRODUCT 2779 bool CMSCollector::have_cms_token() { 2780 Thread* thr = Thread::current(); 2781 if (thr->is_VM_thread()) { 2782 return ConcurrentMarkSweepThread::vm_thread_has_cms_token(); 2783 } else if (thr->is_ConcurrentGC_thread()) { 2784 return ConcurrentMarkSweepThread::cms_thread_has_cms_token(); 2785 } else if (thr->is_GC_task_thread()) { 2786 return ConcurrentMarkSweepThread::vm_thread_has_cms_token() && 2787 ParGCRareEvent_lock->owned_by_self(); 2788 } 2789 return false; 2790 } 2791 #endif 2792 2793 // Check reachability of the given heap address in CMS generation, 2794 // treating all other generations as roots. 2795 bool CMSCollector::is_cms_reachable(HeapWord* addr) { 2796 // We could "guarantee" below, rather than assert, but i'll 2797 // leave these as "asserts" so that an adventurous debugger 2798 // could try this in the product build provided some subset of 2799 // the conditions were met, provided they were intersted in the 2800 // results and knew that the computation below wouldn't interfere 2801 // with other concurrent computations mutating the structures 2802 // being read or written. 2803 assert(SafepointSynchronize::is_at_safepoint(), 2804 "Else mutations in object graph will make answer suspect"); 2805 assert(have_cms_token(), "Should hold cms token"); 2806 assert(haveFreelistLocks(), "must hold free list locks"); 2807 assert_lock_strong(bitMapLock()); 2808 2809 // Clear the marking bit map array before starting, but, just 2810 // for kicks, first report if the given address is already marked 2811 gclog_or_tty->print_cr("Start: Address 0x%x is%s marked", addr, 2812 _markBitMap.isMarked(addr) ? "" : " not"); 2813 2814 if (verify_after_remark()) { 2815 MutexLockerEx x(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag); 2816 bool result = verification_mark_bm()->isMarked(addr); 2817 gclog_or_tty->print_cr("TransitiveMark: Address 0x%x %s marked", addr, 2818 result ? "IS" : "is NOT"); 2819 return result; 2820 } else { 2821 gclog_or_tty->print_cr("Could not compute result"); 2822 return false; 2823 } 2824 } 2825 2826 //////////////////////////////////////////////////////// 2827 // CMS Verification Support 2828 //////////////////////////////////////////////////////// 2829 // Following the remark phase, the following invariant 2830 // should hold -- each object in the CMS heap which is 2831 // marked in markBitMap() should be marked in the verification_mark_bm(). 2832 2833 class VerifyMarkedClosure: public BitMapClosure { 2834 CMSBitMap* _marks; 2835 bool _failed; 2836 2837 public: 2838 VerifyMarkedClosure(CMSBitMap* bm): _marks(bm), _failed(false) {} 2839 2840 bool do_bit(size_t offset) { 2841 HeapWord* addr = _marks->offsetToHeapWord(offset); 2842 if (!_marks->isMarked(addr)) { 2843 oop(addr)->print_on(gclog_or_tty); 2844 gclog_or_tty->print_cr(" ("INTPTR_FORMAT" should have been marked)", addr); 2845 _failed = true; 2846 } 2847 return true; 2848 } 2849 2850 bool failed() { return _failed; } 2851 }; 2852 2853 bool CMSCollector::verify_after_remark() { 2854 gclog_or_tty->print(" [Verifying CMS Marking... "); 2855 MutexLockerEx ml(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag); 2856 static bool init = false; 2857 2858 assert(SafepointSynchronize::is_at_safepoint(), 2859 "Else mutations in object graph will make answer suspect"); 2860 assert(have_cms_token(), 2861 "Else there may be mutual interference in use of " 2862 " verification data structures"); 2863 assert(_collectorState > Marking && _collectorState <= Sweeping, 2864 "Else marking info checked here may be obsolete"); 2865 assert(haveFreelistLocks(), "must hold free list locks"); 2866 assert_lock_strong(bitMapLock()); 2867 2868 2869 // Allocate marking bit map if not already allocated 2870 if (!init) { // first time 2871 if (!verification_mark_bm()->allocate(_span)) { 2872 return false; 2873 } 2874 init = true; 2875 } 2876 2877 assert(verification_mark_stack()->isEmpty(), "Should be empty"); 2878 2879 // Turn off refs discovery -- so we will be tracing through refs. 2880 // This is as intended, because by this time 2881 // GC must already have cleared any refs that need to be cleared, 2882 // and traced those that need to be marked; moreover, 2883 // the marking done here is not going to intefere in any 2884 // way with the marking information used by GC. 2885 NoRefDiscovery no_discovery(ref_processor()); 2886 2887 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;) 2888 2889 // Clear any marks from a previous round 2890 verification_mark_bm()->clear_all(); 2891 assert(verification_mark_stack()->isEmpty(), "markStack should be empty"); 2892 verify_work_stacks_empty(); 2893 2894 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2895 gch->ensure_parsability(false); // fill TLABs, but no need to retire them 2896 // Update the saved marks which may affect the root scans. 2897 gch->save_marks(); 2898 2899 if (CMSRemarkVerifyVariant == 1) { 2900 // In this first variant of verification, we complete 2901 // all marking, then check if the new marks-verctor is 2902 // a subset of the CMS marks-vector. 2903 verify_after_remark_work_1(); 2904 } else if (CMSRemarkVerifyVariant == 2) { 2905 // In this second variant of verification, we flag an error 2906 // (i.e. an object reachable in the new marks-vector not reachable 2907 // in the CMS marks-vector) immediately, also indicating the 2908 // identify of an object (A) that references the unmarked object (B) -- 2909 // presumably, a mutation to A failed to be picked up by preclean/remark? 2910 verify_after_remark_work_2(); 2911 } else { 2912 warning("Unrecognized value %d for CMSRemarkVerifyVariant", 2913 CMSRemarkVerifyVariant); 2914 } 2915 gclog_or_tty->print(" done] "); 2916 return true; 2917 } 2918 2919 void CMSCollector::verify_after_remark_work_1() { 2920 ResourceMark rm; 2921 HandleMark hm; 2922 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2923 2924 // Mark from roots one level into CMS 2925 MarkRefsIntoClosure notOlder(_span, verification_mark_bm()); 2926 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2927 2928 gch->gen_process_strong_roots(_cmsGen->level(), 2929 true, // younger gens are roots 2930 true, // activate StrongRootsScope 2931 true, // collecting perm gen 2932 SharedHeap::ScanningOption(roots_scanning_options()), 2933 ¬Older, 2934 true, // walk code active on stacks 2935 NULL); 2936 2937 // Now mark from the roots 2938 assert(_revisitStack.isEmpty(), "Should be empty"); 2939 MarkFromRootsClosure markFromRootsClosure(this, _span, 2940 verification_mark_bm(), verification_mark_stack(), &_revisitStack, 2941 false /* don't yield */, true /* verifying */); 2942 assert(_restart_addr == NULL, "Expected pre-condition"); 2943 verification_mark_bm()->iterate(&markFromRootsClosure); 2944 while (_restart_addr != NULL) { 2945 // Deal with stack overflow: by restarting at the indicated 2946 // address. 2947 HeapWord* ra = _restart_addr; 2948 markFromRootsClosure.reset(ra); 2949 _restart_addr = NULL; 2950 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end()); 2951 } 2952 assert(verification_mark_stack()->isEmpty(), "Should have been drained"); 2953 verify_work_stacks_empty(); 2954 // Should reset the revisit stack above, since no class tree 2955 // surgery is forthcoming. 2956 _revisitStack.reset(); // throwing away all contents 2957 2958 // Marking completed -- now verify that each bit marked in 2959 // verification_mark_bm() is also marked in markBitMap(); flag all 2960 // errors by printing corresponding objects. 2961 VerifyMarkedClosure vcl(markBitMap()); 2962 verification_mark_bm()->iterate(&vcl); 2963 if (vcl.failed()) { 2964 gclog_or_tty->print("Verification failed"); 2965 Universe::heap()->print_on(gclog_or_tty); 2966 fatal("CMS: failed marking verification after remark"); 2967 } 2968 } 2969 2970 void CMSCollector::verify_after_remark_work_2() { 2971 ResourceMark rm; 2972 HandleMark hm; 2973 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2974 2975 // Mark from roots one level into CMS 2976 MarkRefsIntoVerifyClosure notOlder(_span, verification_mark_bm(), 2977 markBitMap()); 2978 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2979 gch->gen_process_strong_roots(_cmsGen->level(), 2980 true, // younger gens are roots 2981 true, // activate StrongRootsScope 2982 true, // collecting perm gen 2983 SharedHeap::ScanningOption(roots_scanning_options()), 2984 ¬Older, 2985 true, // walk code active on stacks 2986 NULL); 2987 2988 // Now mark from the roots 2989 assert(_revisitStack.isEmpty(), "Should be empty"); 2990 MarkFromRootsVerifyClosure markFromRootsClosure(this, _span, 2991 verification_mark_bm(), markBitMap(), verification_mark_stack()); 2992 assert(_restart_addr == NULL, "Expected pre-condition"); 2993 verification_mark_bm()->iterate(&markFromRootsClosure); 2994 while (_restart_addr != NULL) { 2995 // Deal with stack overflow: by restarting at the indicated 2996 // address. 2997 HeapWord* ra = _restart_addr; 2998 markFromRootsClosure.reset(ra); 2999 _restart_addr = NULL; 3000 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end()); 3001 } 3002 assert(verification_mark_stack()->isEmpty(), "Should have been drained"); 3003 verify_work_stacks_empty(); 3004 // Should reset the revisit stack above, since no class tree 3005 // surgery is forthcoming. 3006 _revisitStack.reset(); // throwing away all contents 3007 3008 // Marking completed -- now verify that each bit marked in 3009 // verification_mark_bm() is also marked in markBitMap(); flag all 3010 // errors by printing corresponding objects. 3011 VerifyMarkedClosure vcl(markBitMap()); 3012 verification_mark_bm()->iterate(&vcl); 3013 assert(!vcl.failed(), "Else verification above should not have succeeded"); 3014 } 3015 3016 void ConcurrentMarkSweepGeneration::save_marks() { 3017 // delegate to CMS space 3018 cmsSpace()->save_marks(); 3019 for (uint i = 0; i < ParallelGCThreads; i++) { 3020 _par_gc_thread_states[i]->promo.startTrackingPromotions(); 3021 } 3022 } 3023 3024 bool ConcurrentMarkSweepGeneration::no_allocs_since_save_marks() { 3025 return cmsSpace()->no_allocs_since_save_marks(); 3026 } 3027 3028 #define CMS_SINCE_SAVE_MARKS_DEFN(OopClosureType, nv_suffix) \ 3029 \ 3030 void ConcurrentMarkSweepGeneration:: \ 3031 oop_since_save_marks_iterate##nv_suffix(OopClosureType* cl) { \ 3032 cl->set_generation(this); \ 3033 cmsSpace()->oop_since_save_marks_iterate##nv_suffix(cl); \ 3034 cl->reset_generation(); \ 3035 save_marks(); \ 3036 } 3037 3038 ALL_SINCE_SAVE_MARKS_CLOSURES(CMS_SINCE_SAVE_MARKS_DEFN) 3039 3040 void 3041 ConcurrentMarkSweepGeneration::object_iterate_since_last_GC(ObjectClosure* blk) 3042 { 3043 // Not currently implemented; need to do the following. -- ysr. 3044 // dld -- I think that is used for some sort of allocation profiler. So it 3045 // really means the objects allocated by the mutator since the last 3046 // GC. We could potentially implement this cheaply by recording only 3047 // the direct allocations in a side data structure. 3048 // 3049 // I think we probably ought not to be required to support these 3050 // iterations at any arbitrary point; I think there ought to be some 3051 // call to enable/disable allocation profiling in a generation/space, 3052 // and the iterator ought to return the objects allocated in the 3053 // gen/space since the enable call, or the last iterator call (which 3054 // will probably be at a GC.) That way, for gens like CM&S that would 3055 // require some extra data structure to support this, we only pay the 3056 // cost when it's in use... 3057 cmsSpace()->object_iterate_since_last_GC(blk); 3058 } 3059 3060 void 3061 ConcurrentMarkSweepGeneration::younger_refs_iterate(OopsInGenClosure* cl) { 3062 cl->set_generation(this); 3063 younger_refs_in_space_iterate(_cmsSpace, cl); 3064 cl->reset_generation(); 3065 } 3066 3067 void 3068 ConcurrentMarkSweepGeneration::oop_iterate(MemRegion mr, OopClosure* cl) { 3069 if (freelistLock()->owned_by_self()) { 3070 Generation::oop_iterate(mr, cl); 3071 } else { 3072 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 3073 Generation::oop_iterate(mr, cl); 3074 } 3075 } 3076 3077 void 3078 ConcurrentMarkSweepGeneration::oop_iterate(OopClosure* cl) { 3079 if (freelistLock()->owned_by_self()) { 3080 Generation::oop_iterate(cl); 3081 } else { 3082 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 3083 Generation::oop_iterate(cl); 3084 } 3085 } 3086 3087 void 3088 ConcurrentMarkSweepGeneration::object_iterate(ObjectClosure* cl) { 3089 if (freelistLock()->owned_by_self()) { 3090 Generation::object_iterate(cl); 3091 } else { 3092 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 3093 Generation::object_iterate(cl); 3094 } 3095 } 3096 3097 void 3098 ConcurrentMarkSweepGeneration::safe_object_iterate(ObjectClosure* cl) { 3099 if (freelistLock()->owned_by_self()) { 3100 Generation::safe_object_iterate(cl); 3101 } else { 3102 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 3103 Generation::safe_object_iterate(cl); 3104 } 3105 } 3106 3107 void 3108 ConcurrentMarkSweepGeneration::pre_adjust_pointers() { 3109 } 3110 3111 void 3112 ConcurrentMarkSweepGeneration::post_compact() { 3113 } 3114 3115 void 3116 ConcurrentMarkSweepGeneration::prepare_for_verify() { 3117 // Fix the linear allocation blocks to look like free blocks. 3118 3119 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those 3120 // are not called when the heap is verified during universe initialization and 3121 // at vm shutdown. 3122 if (freelistLock()->owned_by_self()) { 3123 cmsSpace()->prepare_for_verify(); 3124 } else { 3125 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag); 3126 cmsSpace()->prepare_for_verify(); 3127 } 3128 } 3129 3130 void 3131 ConcurrentMarkSweepGeneration::verify() { 3132 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those 3133 // are not called when the heap is verified during universe initialization and 3134 // at vm shutdown. 3135 if (freelistLock()->owned_by_self()) { 3136 cmsSpace()->verify(); 3137 } else { 3138 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag); 3139 cmsSpace()->verify(); 3140 } 3141 } 3142 3143 void CMSCollector::verify() { 3144 _cmsGen->verify(); 3145 _permGen->verify(); 3146 } 3147 3148 #ifndef PRODUCT 3149 bool CMSCollector::overflow_list_is_empty() const { 3150 assert(_num_par_pushes >= 0, "Inconsistency"); 3151 if (_overflow_list == NULL) { 3152 assert(_num_par_pushes == 0, "Inconsistency"); 3153 } 3154 return _overflow_list == NULL; 3155 } 3156 3157 // The methods verify_work_stacks_empty() and verify_overflow_empty() 3158 // merely consolidate assertion checks that appear to occur together frequently. 3159 void CMSCollector::verify_work_stacks_empty() const { 3160 assert(_markStack.isEmpty(), "Marking stack should be empty"); 3161 assert(overflow_list_is_empty(), "Overflow list should be empty"); 3162 } 3163 3164 void CMSCollector::verify_overflow_empty() const { 3165 assert(overflow_list_is_empty(), "Overflow list should be empty"); 3166 assert(no_preserved_marks(), "No preserved marks"); 3167 } 3168 #endif // PRODUCT 3169 3170 // Decide if we want to enable class unloading as part of the 3171 // ensuing concurrent GC cycle. We will collect the perm gen and 3172 // unload classes if it's the case that: 3173 // (1) an explicit gc request has been made and the flag 3174 // ExplicitGCInvokesConcurrentAndUnloadsClasses is set, OR 3175 // (2) (a) class unloading is enabled at the command line, and 3176 // (b) (i) perm gen threshold has been crossed, or 3177 // (ii) old gen is getting really full, or 3178 // (iii) the previous N CMS collections did not collect the 3179 // perm gen 3180 // NOTE: Provided there is no change in the state of the heap between 3181 // calls to this method, it should have idempotent results. Moreover, 3182 // its results should be monotonically increasing (i.e. going from 0 to 1, 3183 // but not 1 to 0) between successive calls between which the heap was 3184 // not collected. For the implementation below, it must thus rely on 3185 // the property that concurrent_cycles_since_last_unload() 3186 // will not decrease unless a collection cycle happened and that 3187 // _permGen->should_concurrent_collect() and _cmsGen->is_too_full() are 3188 // themselves also monotonic in that sense. See check_monotonicity() 3189 // below. 3190 bool CMSCollector::update_should_unload_classes() { 3191 _should_unload_classes = false; 3192 // Condition 1 above 3193 if (_full_gc_requested && ExplicitGCInvokesConcurrentAndUnloadsClasses) { 3194 _should_unload_classes = true; 3195 } else if (CMSClassUnloadingEnabled) { // Condition 2.a above 3196 // Disjuncts 2.b.(i,ii,iii) above 3197 _should_unload_classes = (concurrent_cycles_since_last_unload() >= 3198 CMSClassUnloadingMaxInterval) 3199 || _permGen->should_concurrent_collect() 3200 || _cmsGen->is_too_full(); 3201 } 3202 return _should_unload_classes; 3203 } 3204 3205 bool ConcurrentMarkSweepGeneration::is_too_full() const { 3206 bool res = should_concurrent_collect(); 3207 res = res && (occupancy() > (double)CMSIsTooFullPercentage/100.0); 3208 return res; 3209 } 3210 3211 void CMSCollector::setup_cms_unloading_and_verification_state() { 3212 const bool should_verify = VerifyBeforeGC || VerifyAfterGC || VerifyDuringGC 3213 || VerifyBeforeExit; 3214 const int rso = SharedHeap::SO_Strings | SharedHeap::SO_CodeCache; 3215 3216 if (should_unload_classes()) { // Should unload classes this cycle 3217 remove_root_scanning_option(rso); // Shrink the root set appropriately 3218 set_verifying(should_verify); // Set verification state for this cycle 3219 return; // Nothing else needs to be done at this time 3220 } 3221 3222 // Not unloading classes this cycle 3223 assert(!should_unload_classes(), "Inconsitency!"); 3224 if ((!verifying() || unloaded_classes_last_cycle()) && should_verify) { 3225 // We were not verifying, or we _were_ unloading classes in the last cycle, 3226 // AND some verification options are enabled this cycle; in this case, 3227 // we must make sure that the deadness map is allocated if not already so, 3228 // and cleared (if already allocated previously -- 3229 // CMSBitMap::sizeInBits() is used to determine if it's allocated). 3230 if (perm_gen_verify_bit_map()->sizeInBits() == 0) { 3231 if (!perm_gen_verify_bit_map()->allocate(_permGen->reserved())) { 3232 warning("Failed to allocate permanent generation verification CMS Bit Map;\n" 3233 "permanent generation verification disabled"); 3234 return; // Note that we leave verification disabled, so we'll retry this 3235 // allocation next cycle. We _could_ remember this failure 3236 // and skip further attempts and permanently disable verification 3237 // attempts if that is considered more desirable. 3238 } 3239 assert(perm_gen_verify_bit_map()->covers(_permGen->reserved()), 3240 "_perm_gen_ver_bit_map inconsistency?"); 3241 } else { 3242 perm_gen_verify_bit_map()->clear_all(); 3243 } 3244 // Include symbols, strings and code cache elements to prevent their resurrection. 3245 add_root_scanning_option(rso); 3246 set_verifying(true); 3247 } else if (verifying() && !should_verify) { 3248 // We were verifying, but some verification flags got disabled. 3249 set_verifying(false); 3250 // Exclude symbols, strings and code cache elements from root scanning to 3251 // reduce IM and RM pauses. 3252 remove_root_scanning_option(rso); 3253 } 3254 } 3255 3256 3257 #ifndef PRODUCT 3258 HeapWord* CMSCollector::block_start(const void* p) const { 3259 const HeapWord* addr = (HeapWord*)p; 3260 if (_span.contains(p)) { 3261 if (_cmsGen->cmsSpace()->is_in_reserved(addr)) { 3262 return _cmsGen->cmsSpace()->block_start(p); 3263 } else { 3264 assert(_permGen->cmsSpace()->is_in_reserved(addr), 3265 "Inconsistent _span?"); 3266 return _permGen->cmsSpace()->block_start(p); 3267 } 3268 } 3269 return NULL; 3270 } 3271 #endif 3272 3273 HeapWord* 3274 ConcurrentMarkSweepGeneration::expand_and_allocate(size_t word_size, 3275 bool tlab, 3276 bool parallel) { 3277 CMSSynchronousYieldRequest yr; 3278 assert(!tlab, "Can't deal with TLAB allocation"); 3279 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 3280 expand(word_size*HeapWordSize, MinHeapDeltaBytes, 3281 CMSExpansionCause::_satisfy_allocation); 3282 if (GCExpandToAllocateDelayMillis > 0) { 3283 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 3284 } 3285 return have_lock_and_allocate(word_size, tlab); 3286 } 3287 3288 // YSR: All of this generation expansion/shrinking stuff is an exact copy of 3289 // OneContigSpaceCardGeneration, which makes me wonder if we should move this 3290 // to CardGeneration and share it... 3291 bool ConcurrentMarkSweepGeneration::expand(size_t bytes, size_t expand_bytes) { 3292 return CardGeneration::expand(bytes, expand_bytes); 3293 } 3294 3295 void ConcurrentMarkSweepGeneration::expand(size_t bytes, size_t expand_bytes, 3296 CMSExpansionCause::Cause cause) 3297 { 3298 3299 bool success = expand(bytes, expand_bytes); 3300 3301 // remember why we expanded; this information is used 3302 // by shouldConcurrentCollect() when making decisions on whether to start 3303 // a new CMS cycle. 3304 if (success) { 3305 set_expansion_cause(cause); 3306 if (PrintGCDetails && Verbose) { 3307 gclog_or_tty->print_cr("Expanded CMS gen for %s", 3308 CMSExpansionCause::to_string(cause)); 3309 } 3310 } 3311 } 3312 3313 HeapWord* ConcurrentMarkSweepGeneration::expand_and_par_lab_allocate(CMSParGCThreadState* ps, size_t word_sz) { 3314 HeapWord* res = NULL; 3315 MutexLocker x(ParGCRareEvent_lock); 3316 while (true) { 3317 // Expansion by some other thread might make alloc OK now: 3318 res = ps->lab.alloc(word_sz); 3319 if (res != NULL) return res; 3320 // If there's not enough expansion space available, give up. 3321 if (_virtual_space.uncommitted_size() < (word_sz * HeapWordSize)) { 3322 return NULL; 3323 } 3324 // Otherwise, we try expansion. 3325 expand(word_sz*HeapWordSize, MinHeapDeltaBytes, 3326 CMSExpansionCause::_allocate_par_lab); 3327 // Now go around the loop and try alloc again; 3328 // A competing par_promote might beat us to the expansion space, 3329 // so we may go around the loop again if promotion fails agaion. 3330 if (GCExpandToAllocateDelayMillis > 0) { 3331 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 3332 } 3333 } 3334 } 3335 3336 3337 bool ConcurrentMarkSweepGeneration::expand_and_ensure_spooling_space( 3338 PromotionInfo* promo) { 3339 MutexLocker x(ParGCRareEvent_lock); 3340 size_t refill_size_bytes = promo->refillSize() * HeapWordSize; 3341 while (true) { 3342 // Expansion by some other thread might make alloc OK now: 3343 if (promo->ensure_spooling_space()) { 3344 assert(promo->has_spooling_space(), 3345 "Post-condition of successful ensure_spooling_space()"); 3346 return true; 3347 } 3348 // If there's not enough expansion space available, give up. 3349 if (_virtual_space.uncommitted_size() < refill_size_bytes) { 3350 return false; 3351 } 3352 // Otherwise, we try expansion. 3353 expand(refill_size_bytes, MinHeapDeltaBytes, 3354 CMSExpansionCause::_allocate_par_spooling_space); 3355 // Now go around the loop and try alloc again; 3356 // A competing allocation might beat us to the expansion space, 3357 // so we may go around the loop again if allocation fails again. 3358 if (GCExpandToAllocateDelayMillis > 0) { 3359 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 3360 } 3361 } 3362 } 3363 3364 3365 3366 void ConcurrentMarkSweepGeneration::shrink(size_t bytes) { 3367 assert_locked_or_safepoint(Heap_lock); 3368 size_t size = ReservedSpace::page_align_size_down(bytes); 3369 if (size > 0) { 3370 shrink_by(size); 3371 } 3372 } 3373 3374 bool ConcurrentMarkSweepGeneration::grow_by(size_t bytes) { 3375 assert_locked_or_safepoint(Heap_lock); 3376 bool result = _virtual_space.expand_by(bytes); 3377 if (result) { 3378 HeapWord* old_end = _cmsSpace->end(); 3379 size_t new_word_size = 3380 heap_word_size(_virtual_space.committed_size()); 3381 MemRegion mr(_cmsSpace->bottom(), new_word_size); 3382 _bts->resize(new_word_size); // resize the block offset shared array 3383 Universe::heap()->barrier_set()->resize_covered_region(mr); 3384 // Hmmmm... why doesn't CFLS::set_end verify locking? 3385 // This is quite ugly; FIX ME XXX 3386 _cmsSpace->assert_locked(freelistLock()); 3387 _cmsSpace->set_end((HeapWord*)_virtual_space.high()); 3388 3389 // update the space and generation capacity counters 3390 if (UsePerfData) { 3391 _space_counters->update_capacity(); 3392 _gen_counters->update_all(); 3393 } 3394 3395 if (Verbose && PrintGC) { 3396 size_t new_mem_size = _virtual_space.committed_size(); 3397 size_t old_mem_size = new_mem_size - bytes; 3398 gclog_or_tty->print_cr("Expanding %s from %ldK by %ldK to %ldK", 3399 name(), old_mem_size/K, bytes/K, new_mem_size/K); 3400 } 3401 } 3402 return result; 3403 } 3404 3405 bool ConcurrentMarkSweepGeneration::grow_to_reserved() { 3406 assert_locked_or_safepoint(Heap_lock); 3407 bool success = true; 3408 const size_t remaining_bytes = _virtual_space.uncommitted_size(); 3409 if (remaining_bytes > 0) { 3410 success = grow_by(remaining_bytes); 3411 DEBUG_ONLY(if (!success) warning("grow to reserved failed");) 3412 } 3413 return success; 3414 } 3415 3416 void ConcurrentMarkSweepGeneration::shrink_by(size_t bytes) { 3417 assert_locked_or_safepoint(Heap_lock); 3418 assert_lock_strong(freelistLock()); 3419 // XXX Fix when compaction is implemented. 3420 warning("Shrinking of CMS not yet implemented"); 3421 return; 3422 } 3423 3424 3425 // Simple ctor/dtor wrapper for accounting & timer chores around concurrent 3426 // phases. 3427 class CMSPhaseAccounting: public StackObj { 3428 public: 3429 CMSPhaseAccounting(CMSCollector *collector, 3430 const char *phase, 3431 bool print_cr = true); 3432 ~CMSPhaseAccounting(); 3433 3434 private: 3435 CMSCollector *_collector; 3436 const char *_phase; 3437 elapsedTimer _wallclock; 3438 bool _print_cr; 3439 3440 public: 3441 // Not MT-safe; so do not pass around these StackObj's 3442 // where they may be accessed by other threads. 3443 jlong wallclock_millis() { 3444 assert(_wallclock.is_active(), "Wall clock should not stop"); 3445 _wallclock.stop(); // to record time 3446 jlong ret = _wallclock.milliseconds(); 3447 _wallclock.start(); // restart 3448 return ret; 3449 } 3450 }; 3451 3452 CMSPhaseAccounting::CMSPhaseAccounting(CMSCollector *collector, 3453 const char *phase, 3454 bool print_cr) : 3455 _collector(collector), _phase(phase), _print_cr(print_cr) { 3456 3457 if (PrintCMSStatistics != 0) { 3458 _collector->resetYields(); 3459 } 3460 if (PrintGCDetails && PrintGCTimeStamps) { 3461 gclog_or_tty->date_stamp(PrintGCDateStamps); 3462 gclog_or_tty->stamp(); 3463 gclog_or_tty->print_cr(": [%s-concurrent-%s-start]", 3464 _collector->cmsGen()->short_name(), _phase); 3465 } 3466 _collector->resetTimer(); 3467 _wallclock.start(); 3468 _collector->startTimer(); 3469 } 3470 3471 CMSPhaseAccounting::~CMSPhaseAccounting() { 3472 assert(_wallclock.is_active(), "Wall clock should not have stopped"); 3473 _collector->stopTimer(); 3474 _wallclock.stop(); 3475 if (PrintGCDetails) { 3476 gclog_or_tty->date_stamp(PrintGCDateStamps); 3477 gclog_or_tty->stamp(PrintGCTimeStamps); 3478 gclog_or_tty->print("[%s-concurrent-%s: %3.3f/%3.3f secs]", 3479 _collector->cmsGen()->short_name(), 3480 _phase, _collector->timerValue(), _wallclock.seconds()); 3481 if (_print_cr) { 3482 gclog_or_tty->print_cr(""); 3483 } 3484 if (PrintCMSStatistics != 0) { 3485 gclog_or_tty->print_cr(" (CMS-concurrent-%s yielded %d times)", _phase, 3486 _collector->yields()); 3487 } 3488 } 3489 } 3490 3491 // CMS work 3492 3493 // Checkpoint the roots into this generation from outside 3494 // this generation. [Note this initial checkpoint need only 3495 // be approximate -- we'll do a catch up phase subsequently.] 3496 void CMSCollector::checkpointRootsInitial(bool asynch) { 3497 assert(_collectorState == InitialMarking, "Wrong collector state"); 3498 check_correct_thread_executing(); 3499 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 3500 3501 ReferenceProcessor* rp = ref_processor(); 3502 SpecializationStats::clear(); 3503 assert(_restart_addr == NULL, "Control point invariant"); 3504 if (asynch) { 3505 // acquire locks for subsequent manipulations 3506 MutexLockerEx x(bitMapLock(), 3507 Mutex::_no_safepoint_check_flag); 3508 checkpointRootsInitialWork(asynch); 3509 // enable ("weak") refs discovery 3510 rp->enable_discovery(true /*verify_disabled*/, true /*check_no_refs*/); 3511 _collectorState = Marking; 3512 } else { 3513 // (Weak) Refs discovery: this is controlled from genCollectedHeap::do_collection 3514 // which recognizes if we are a CMS generation, and doesn't try to turn on 3515 // discovery; verify that they aren't meddling. 3516 assert(!rp->discovery_is_atomic(), 3517 "incorrect setting of discovery predicate"); 3518 assert(!rp->discovery_enabled(), "genCollectedHeap shouldn't control " 3519 "ref discovery for this generation kind"); 3520 // already have locks 3521 checkpointRootsInitialWork(asynch); 3522 // now enable ("weak") refs discovery 3523 rp->enable_discovery(true /*verify_disabled*/, false /*verify_no_refs*/); 3524 _collectorState = Marking; 3525 } 3526 SpecializationStats::print(); 3527 } 3528 3529 void CMSCollector::checkpointRootsInitialWork(bool asynch) { 3530 assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped"); 3531 assert(_collectorState == InitialMarking, "just checking"); 3532 3533 // If there has not been a GC[n-1] since last GC[n] cycle completed, 3534 // precede our marking with a collection of all 3535 // younger generations to keep floating garbage to a minimum. 3536 // XXX: we won't do this for now -- it's an optimization to be done later. 3537 3538 // already have locks 3539 assert_lock_strong(bitMapLock()); 3540 assert(_markBitMap.isAllClear(), "was reset at end of previous cycle"); 3541 3542 // Setup the verification and class unloading state for this 3543 // CMS collection cycle. 3544 setup_cms_unloading_and_verification_state(); 3545 3546 NOT_PRODUCT(GCTraceTime t("\ncheckpointRootsInitialWork", 3547 PrintGCDetails && Verbose, true, _gc_timer_cm);) 3548 if (UseAdaptiveSizePolicy) { 3549 size_policy()->checkpoint_roots_initial_begin(); 3550 } 3551 3552 // Reset all the PLAB chunk arrays if necessary. 3553 if (_survivor_plab_array != NULL && !CMSPLABRecordAlways) { 3554 reset_survivor_plab_arrays(); 3555 } 3556 3557 ResourceMark rm; 3558 HandleMark hm; 3559 3560 FalseClosure falseClosure; 3561 // In the case of a synchronous collection, we will elide the 3562 // remark step, so it's important to catch all the nmethod oops 3563 // in this step. 3564 // The final 'true' flag to gen_process_strong_roots will ensure this. 3565 // If 'async' is true, we can relax the nmethod tracing. 3566 MarkRefsIntoClosure notOlder(_span, &_markBitMap); 3567 GenCollectedHeap* gch = GenCollectedHeap::heap(); 3568 3569 verify_work_stacks_empty(); 3570 verify_overflow_empty(); 3571 3572 gch->ensure_parsability(false); // fill TLABs, but no need to retire them 3573 // Update the saved marks which may affect the root scans. 3574 gch->save_marks(); 3575 3576 // weak reference processing has not started yet. 3577 ref_processor()->set_enqueuing_is_done(false); 3578 3579 { 3580 // This is not needed. DEBUG_ONLY(RememberKlassesChecker imx(true);) 3581 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;) 3582 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 3583 gch->gen_process_strong_roots(_cmsGen->level(), 3584 true, // younger gens are roots 3585 true, // activate StrongRootsScope 3586 true, // collecting perm gen 3587 SharedHeap::ScanningOption(roots_scanning_options()), 3588 ¬Older, 3589 true, // walk all of code cache if (so & SO_CodeCache) 3590 NULL); 3591 } 3592 3593 // Clear mod-union table; it will be dirtied in the prologue of 3594 // CMS generation per each younger generation collection. 3595 3596 assert(_modUnionTable.isAllClear(), 3597 "Was cleared in most recent final checkpoint phase" 3598 " or no bits are set in the gc_prologue before the start of the next " 3599 "subsequent marking phase."); 3600 3601 // Save the end of the used_region of the constituent generations 3602 // to be used to limit the extent of sweep in each generation. 3603 save_sweep_limits(); 3604 if (UseAdaptiveSizePolicy) { 3605 size_policy()->checkpoint_roots_initial_end(gch->gc_cause()); 3606 } 3607 verify_overflow_empty(); 3608 } 3609 3610 bool CMSCollector::markFromRoots(bool asynch) { 3611 // we might be tempted to assert that: 3612 // assert(asynch == !SafepointSynchronize::is_at_safepoint(), 3613 // "inconsistent argument?"); 3614 // However that wouldn't be right, because it's possible that 3615 // a safepoint is indeed in progress as a younger generation 3616 // stop-the-world GC happens even as we mark in this generation. 3617 assert(_collectorState == Marking, "inconsistent state?"); 3618 check_correct_thread_executing(); 3619 verify_overflow_empty(); 3620 3621 bool res; 3622 if (asynch) { 3623 3624 // Start the timers for adaptive size policy for the concurrent phases 3625 // Do it here so that the foreground MS can use the concurrent 3626 // timer since a foreground MS might has the sweep done concurrently 3627 // or STW. 3628 if (UseAdaptiveSizePolicy) { 3629 size_policy()->concurrent_marking_begin(); 3630 } 3631 3632 // Weak ref discovery note: We may be discovering weak 3633 // refs in this generation concurrent (but interleaved) with 3634 // weak ref discovery by a younger generation collector. 3635 3636 CMSTokenSyncWithLocks ts(true, bitMapLock()); 3637 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 3638 CMSPhaseAccounting pa(this, "mark", !PrintGCDetails); 3639 res = markFromRootsWork(asynch); 3640 if (res) { 3641 _collectorState = Precleaning; 3642 } else { // We failed and a foreground collection wants to take over 3643 assert(_foregroundGCIsActive, "internal state inconsistency"); 3644 assert(_restart_addr == NULL, "foreground will restart from scratch"); 3645 if (PrintGCDetails) { 3646 gclog_or_tty->print_cr("bailing out to foreground collection"); 3647 } 3648 } 3649 if (UseAdaptiveSizePolicy) { 3650 size_policy()->concurrent_marking_end(); 3651 } 3652 } else { 3653 assert(SafepointSynchronize::is_at_safepoint(), 3654 "inconsistent with asynch == false"); 3655 if (UseAdaptiveSizePolicy) { 3656 size_policy()->ms_collection_marking_begin(); 3657 } 3658 // already have locks 3659 res = markFromRootsWork(asynch); 3660 _collectorState = FinalMarking; 3661 if (UseAdaptiveSizePolicy) { 3662 GenCollectedHeap* gch = GenCollectedHeap::heap(); 3663 size_policy()->ms_collection_marking_end(gch->gc_cause()); 3664 } 3665 } 3666 verify_overflow_empty(); 3667 return res; 3668 } 3669 3670 bool CMSCollector::markFromRootsWork(bool asynch) { 3671 // iterate over marked bits in bit map, doing a full scan and mark 3672 // from these roots using the following algorithm: 3673 // . if oop is to the right of the current scan pointer, 3674 // mark corresponding bit (we'll process it later) 3675 // . else (oop is to left of current scan pointer) 3676 // push oop on marking stack 3677 // . drain the marking stack 3678 3679 // Note that when we do a marking step we need to hold the 3680 // bit map lock -- recall that direct allocation (by mutators) 3681 // and promotion (by younger generation collectors) is also 3682 // marking the bit map. [the so-called allocate live policy.] 3683 // Because the implementation of bit map marking is not 3684 // robust wrt simultaneous marking of bits in the same word, 3685 // we need to make sure that there is no such interference 3686 // between concurrent such updates. 3687 3688 // already have locks 3689 assert_lock_strong(bitMapLock()); 3690 3691 // Clear the revisit stack, just in case there are any 3692 // obsolete contents from a short-circuited previous CMS cycle. 3693 _revisitStack.reset(); 3694 verify_work_stacks_empty(); 3695 verify_overflow_empty(); 3696 assert(_revisitStack.isEmpty(), "tabula rasa"); 3697 DEBUG_ONLY(RememberKlassesChecker cmx(should_unload_classes());) 3698 bool result = false; 3699 if (CMSConcurrentMTEnabled && ConcGCThreads > 0) { 3700 result = do_marking_mt(asynch); 3701 } else { 3702 result = do_marking_st(asynch); 3703 } 3704 return result; 3705 } 3706 3707 // Forward decl 3708 class CMSConcMarkingTask; 3709 3710 class CMSConcMarkingTerminator: public ParallelTaskTerminator { 3711 CMSCollector* _collector; 3712 CMSConcMarkingTask* _task; 3713 public: 3714 virtual void yield(); 3715 3716 // "n_threads" is the number of threads to be terminated. 3717 // "queue_set" is a set of work queues of other threads. 3718 // "collector" is the CMS collector associated with this task terminator. 3719 // "yield" indicates whether we need the gang as a whole to yield. 3720 CMSConcMarkingTerminator(int n_threads, TaskQueueSetSuper* queue_set, CMSCollector* collector) : 3721 ParallelTaskTerminator(n_threads, queue_set), 3722 _collector(collector) { } 3723 3724 void set_task(CMSConcMarkingTask* task) { 3725 _task = task; 3726 } 3727 }; 3728 3729 class CMSConcMarkingTerminatorTerminator: public TerminatorTerminator { 3730 CMSConcMarkingTask* _task; 3731 public: 3732 bool should_exit_termination(); 3733 void set_task(CMSConcMarkingTask* task) { 3734 _task = task; 3735 } 3736 }; 3737 3738 // MT Concurrent Marking Task 3739 class CMSConcMarkingTask: public YieldingFlexibleGangTask { 3740 CMSCollector* _collector; 3741 int _n_workers; // requested/desired # workers 3742 bool _asynch; 3743 bool _result; 3744 CompactibleFreeListSpace* _cms_space; 3745 CompactibleFreeListSpace* _perm_space; 3746 char _pad_front[64]; // padding to ... 3747 HeapWord* _global_finger; // ... avoid sharing cache line 3748 char _pad_back[64]; 3749 HeapWord* _restart_addr; 3750 3751 // Exposed here for yielding support 3752 Mutex* const _bit_map_lock; 3753 3754 // The per thread work queues, available here for stealing 3755 OopTaskQueueSet* _task_queues; 3756 3757 // Termination (and yielding) support 3758 CMSConcMarkingTerminator _term; 3759 CMSConcMarkingTerminatorTerminator _term_term; 3760 3761 public: 3762 CMSConcMarkingTask(CMSCollector* collector, 3763 CompactibleFreeListSpace* cms_space, 3764 CompactibleFreeListSpace* perm_space, 3765 bool asynch, 3766 YieldingFlexibleWorkGang* workers, 3767 OopTaskQueueSet* task_queues): 3768 YieldingFlexibleGangTask("Concurrent marking done multi-threaded"), 3769 _collector(collector), 3770 _cms_space(cms_space), 3771 _perm_space(perm_space), 3772 _asynch(asynch), _n_workers(0), _result(true), 3773 _task_queues(task_queues), 3774 _term(_n_workers, task_queues, _collector), 3775 _bit_map_lock(collector->bitMapLock()) 3776 { 3777 _requested_size = _n_workers; 3778 _term.set_task(this); 3779 _term_term.set_task(this); 3780 assert(_cms_space->bottom() < _perm_space->bottom(), 3781 "Finger incorrectly initialized below"); 3782 _restart_addr = _global_finger = _cms_space->bottom(); 3783 } 3784 3785 3786 OopTaskQueueSet* task_queues() { return _task_queues; } 3787 3788 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 3789 3790 HeapWord** global_finger_addr() { return &_global_finger; } 3791 3792 CMSConcMarkingTerminator* terminator() { return &_term; } 3793 3794 virtual void set_for_termination(int active_workers) { 3795 terminator()->reset_for_reuse(active_workers); 3796 } 3797 3798 void work(uint worker_id); 3799 bool should_yield() { 3800 return ConcurrentMarkSweepThread::should_yield() 3801 && !_collector->foregroundGCIsActive() 3802 && _asynch; 3803 } 3804 3805 virtual void coordinator_yield(); // stuff done by coordinator 3806 bool result() { return _result; } 3807 3808 void reset(HeapWord* ra) { 3809 assert(_global_finger >= _cms_space->end(), "Postcondition of ::work(i)"); 3810 assert(_global_finger >= _perm_space->end(), "Postcondition of ::work(i)"); 3811 assert(ra < _perm_space->end(), "ra too large"); 3812 _restart_addr = _global_finger = ra; 3813 _term.reset_for_reuse(); 3814 } 3815 3816 static bool get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 3817 OopTaskQueue* work_q); 3818 3819 private: 3820 void do_scan_and_mark(int i, CompactibleFreeListSpace* sp); 3821 void do_work_steal(int i); 3822 void bump_global_finger(HeapWord* f); 3823 }; 3824 3825 bool CMSConcMarkingTerminatorTerminator::should_exit_termination() { 3826 assert(_task != NULL, "Error"); 3827 return _task->yielding(); 3828 // Note that we do not need the disjunct || _task->should_yield() above 3829 // because we want terminating threads to yield only if the task 3830 // is already in the midst of yielding, which happens only after at least one 3831 // thread has yielded. 3832 } 3833 3834 void CMSConcMarkingTerminator::yield() { 3835 if (_task->should_yield()) { 3836 _task->yield(); 3837 } else { 3838 ParallelTaskTerminator::yield(); 3839 } 3840 } 3841 3842 //////////////////////////////////////////////////////////////// 3843 // Concurrent Marking Algorithm Sketch 3844 //////////////////////////////////////////////////////////////// 3845 // Until all tasks exhausted (both spaces): 3846 // -- claim next available chunk 3847 // -- bump global finger via CAS 3848 // -- find first object that starts in this chunk 3849 // and start scanning bitmap from that position 3850 // -- scan marked objects for oops 3851 // -- CAS-mark target, and if successful: 3852 // . if target oop is above global finger (volatile read) 3853 // nothing to do 3854 // . if target oop is in chunk and above local finger 3855 // then nothing to do 3856 // . else push on work-queue 3857 // -- Deal with possible overflow issues: 3858 // . local work-queue overflow causes stuff to be pushed on 3859 // global (common) overflow queue 3860 // . always first empty local work queue 3861 // . then get a batch of oops from global work queue if any 3862 // . then do work stealing 3863 // -- When all tasks claimed (both spaces) 3864 // and local work queue empty, 3865 // then in a loop do: 3866 // . check global overflow stack; steal a batch of oops and trace 3867 // . try to steal from other threads oif GOS is empty 3868 // . if neither is available, offer termination 3869 // -- Terminate and return result 3870 // 3871 void CMSConcMarkingTask::work(uint worker_id) { 3872 elapsedTimer _timer; 3873 ResourceMark rm; 3874 HandleMark hm; 3875 3876 DEBUG_ONLY(_collector->verify_overflow_empty();) 3877 3878 // Before we begin work, our work queue should be empty 3879 assert(work_queue(worker_id)->size() == 0, "Expected to be empty"); 3880 // Scan the bitmap covering _cms_space, tracing through grey objects. 3881 _timer.start(); 3882 do_scan_and_mark(worker_id, _cms_space); 3883 _timer.stop(); 3884 if (PrintCMSStatistics != 0) { 3885 gclog_or_tty->print_cr("Finished cms space scanning in %dth thread: %3.3f sec", 3886 worker_id, _timer.seconds()); 3887 // XXX: need xxx/xxx type of notation, two timers 3888 } 3889 3890 // ... do the same for the _perm_space 3891 _timer.reset(); 3892 _timer.start(); 3893 do_scan_and_mark(worker_id, _perm_space); 3894 _timer.stop(); 3895 if (PrintCMSStatistics != 0) { 3896 gclog_or_tty->print_cr("Finished perm space scanning in %dth thread: %3.3f sec", 3897 worker_id, _timer.seconds()); 3898 // XXX: need xxx/xxx type of notation, two timers 3899 } 3900 3901 // ... do work stealing 3902 _timer.reset(); 3903 _timer.start(); 3904 do_work_steal(worker_id); 3905 _timer.stop(); 3906 if (PrintCMSStatistics != 0) { 3907 gclog_or_tty->print_cr("Finished work stealing in %dth thread: %3.3f sec", 3908 worker_id, _timer.seconds()); 3909 // XXX: need xxx/xxx type of notation, two timers 3910 } 3911 assert(_collector->_markStack.isEmpty(), "Should have been emptied"); 3912 assert(work_queue(worker_id)->size() == 0, "Should have been emptied"); 3913 // Note that under the current task protocol, the 3914 // following assertion is true even of the spaces 3915 // expanded since the completion of the concurrent 3916 // marking. XXX This will likely change under a strict 3917 // ABORT semantics. 3918 assert(_global_finger > _cms_space->end() && 3919 _global_finger >= _perm_space->end(), 3920 "All tasks have been completed"); 3921 DEBUG_ONLY(_collector->verify_overflow_empty();) 3922 } 3923 3924 void CMSConcMarkingTask::bump_global_finger(HeapWord* f) { 3925 HeapWord* read = _global_finger; 3926 HeapWord* cur = read; 3927 while (f > read) { 3928 cur = read; 3929 read = (HeapWord*) Atomic::cmpxchg_ptr(f, &_global_finger, cur); 3930 if (cur == read) { 3931 // our cas succeeded 3932 assert(_global_finger >= f, "protocol consistency"); 3933 break; 3934 } 3935 } 3936 } 3937 3938 // This is really inefficient, and should be redone by 3939 // using (not yet available) block-read and -write interfaces to the 3940 // stack and the work_queue. XXX FIX ME !!! 3941 bool CMSConcMarkingTask::get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 3942 OopTaskQueue* work_q) { 3943 // Fast lock-free check 3944 if (ovflw_stk->length() == 0) { 3945 return false; 3946 } 3947 assert(work_q->size() == 0, "Shouldn't steal"); 3948 MutexLockerEx ml(ovflw_stk->par_lock(), 3949 Mutex::_no_safepoint_check_flag); 3950 // Grab up to 1/4 the size of the work queue 3951 size_t num = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 3952 (size_t)ParGCDesiredObjsFromOverflowList); 3953 num = MIN2(num, ovflw_stk->length()); 3954 for (int i = (int) num; i > 0; i--) { 3955 oop cur = ovflw_stk->pop(); 3956 assert(cur != NULL, "Counted wrong?"); 3957 work_q->push(cur); 3958 } 3959 return num > 0; 3960 } 3961 3962 void CMSConcMarkingTask::do_scan_and_mark(int i, CompactibleFreeListSpace* sp) { 3963 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 3964 int n_tasks = pst->n_tasks(); 3965 // We allow that there may be no tasks to do here because 3966 // we are restarting after a stack overflow. 3967 assert(pst->valid() || n_tasks == 0, "Uninitialized use?"); 3968 uint nth_task = 0; 3969 3970 HeapWord* aligned_start = sp->bottom(); 3971 if (sp->used_region().contains(_restart_addr)) { 3972 // Align down to a card boundary for the start of 0th task 3973 // for this space. 3974 aligned_start = 3975 (HeapWord*)align_size_down((uintptr_t)_restart_addr, 3976 CardTableModRefBS::card_size); 3977 } 3978 3979 size_t chunk_size = sp->marking_task_size(); 3980 while (!pst->is_task_claimed(/* reference */ nth_task)) { 3981 // Having claimed the nth task in this space, 3982 // compute the chunk that it corresponds to: 3983 MemRegion span = MemRegion(aligned_start + nth_task*chunk_size, 3984 aligned_start + (nth_task+1)*chunk_size); 3985 // Try and bump the global finger via a CAS; 3986 // note that we need to do the global finger bump 3987 // _before_ taking the intersection below, because 3988 // the task corresponding to that region will be 3989 // deemed done even if the used_region() expands 3990 // because of allocation -- as it almost certainly will 3991 // during start-up while the threads yield in the 3992 // closure below. 3993 HeapWord* finger = span.end(); 3994 bump_global_finger(finger); // atomically 3995 // There are null tasks here corresponding to chunks 3996 // beyond the "top" address of the space. 3997 span = span.intersection(sp->used_region()); 3998 if (!span.is_empty()) { // Non-null task 3999 HeapWord* prev_obj; 4000 assert(!span.contains(_restart_addr) || nth_task == 0, 4001 "Inconsistency"); 4002 if (nth_task == 0) { 4003 // For the 0th task, we'll not need to compute a block_start. 4004 if (span.contains(_restart_addr)) { 4005 // In the case of a restart because of stack overflow, 4006 // we might additionally skip a chunk prefix. 4007 prev_obj = _restart_addr; 4008 } else { 4009 prev_obj = span.start(); 4010 } 4011 } else { 4012 // We want to skip the first object because 4013 // the protocol is to scan any object in its entirety 4014 // that _starts_ in this span; a fortiori, any 4015 // object starting in an earlier span is scanned 4016 // as part of an earlier claimed task. 4017 // Below we use the "careful" version of block_start 4018 // so we do not try to navigate uninitialized objects. 4019 prev_obj = sp->block_start_careful(span.start()); 4020 // Below we use a variant of block_size that uses the 4021 // Printezis bits to avoid waiting for allocated 4022 // objects to become initialized/parsable. 4023 while (prev_obj < span.start()) { 4024 size_t sz = sp->block_size_no_stall(prev_obj, _collector); 4025 if (sz > 0) { 4026 prev_obj += sz; 4027 } else { 4028 // In this case we may end up doing a bit of redundant 4029 // scanning, but that appears unavoidable, short of 4030 // locking the free list locks; see bug 6324141. 4031 break; 4032 } 4033 } 4034 } 4035 if (prev_obj < span.end()) { 4036 MemRegion my_span = MemRegion(prev_obj, span.end()); 4037 // Do the marking work within a non-empty span -- 4038 // the last argument to the constructor indicates whether the 4039 // iteration should be incremental with periodic yields. 4040 Par_MarkFromRootsClosure cl(this, _collector, my_span, 4041 &_collector->_markBitMap, 4042 work_queue(i), 4043 &_collector->_markStack, 4044 &_collector->_revisitStack, 4045 _asynch); 4046 _collector->_markBitMap.iterate(&cl, my_span.start(), my_span.end()); 4047 } // else nothing to do for this task 4048 } // else nothing to do for this task 4049 } 4050 // We'd be tempted to assert here that since there are no 4051 // more tasks left to claim in this space, the global_finger 4052 // must exceed space->top() and a fortiori space->end(). However, 4053 // that would not quite be correct because the bumping of 4054 // global_finger occurs strictly after the claiming of a task, 4055 // so by the time we reach here the global finger may not yet 4056 // have been bumped up by the thread that claimed the last 4057 // task. 4058 pst->all_tasks_completed(); 4059 } 4060 4061 class Par_ConcMarkingClosure: public Par_KlassRememberingOopClosure { 4062 private: 4063 CMSConcMarkingTask* _task; 4064 MemRegion _span; 4065 CMSBitMap* _bit_map; 4066 CMSMarkStack* _overflow_stack; 4067 OopTaskQueue* _work_queue; 4068 protected: 4069 DO_OOP_WORK_DEFN 4070 public: 4071 Par_ConcMarkingClosure(CMSCollector* collector, CMSConcMarkingTask* task, OopTaskQueue* work_queue, 4072 CMSBitMap* bit_map, CMSMarkStack* overflow_stack, 4073 CMSMarkStack* revisit_stack): 4074 Par_KlassRememberingOopClosure(collector, collector->ref_processor(), revisit_stack), 4075 _task(task), 4076 _span(collector->_span), 4077 _work_queue(work_queue), 4078 _bit_map(bit_map), 4079 _overflow_stack(overflow_stack) 4080 { } 4081 virtual void do_oop(oop* p); 4082 virtual void do_oop(narrowOop* p); 4083 void trim_queue(size_t max); 4084 void handle_stack_overflow(HeapWord* lost); 4085 void do_yield_check() { 4086 if (_task->should_yield()) { 4087 _task->yield(); 4088 } 4089 } 4090 }; 4091 4092 // Grey object scanning during work stealing phase -- 4093 // the salient assumption here is that any references 4094 // that are in these stolen objects being scanned must 4095 // already have been initialized (else they would not have 4096 // been published), so we do not need to check for 4097 // uninitialized objects before pushing here. 4098 void Par_ConcMarkingClosure::do_oop(oop obj) { 4099 assert(obj->is_oop_or_null(true), "expected an oop or NULL"); 4100 HeapWord* addr = (HeapWord*)obj; 4101 // Check if oop points into the CMS generation 4102 // and is not marked 4103 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 4104 // a white object ... 4105 // If we manage to "claim" the object, by being the 4106 // first thread to mark it, then we push it on our 4107 // marking stack 4108 if (_bit_map->par_mark(addr)) { // ... now grey 4109 // push on work queue (grey set) 4110 bool simulate_overflow = false; 4111 NOT_PRODUCT( 4112 if (CMSMarkStackOverflowALot && 4113 _collector->simulate_overflow()) { 4114 // simulate a stack overflow 4115 simulate_overflow = true; 4116 } 4117 ) 4118 if (simulate_overflow || 4119 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) { 4120 // stack overflow 4121 if (PrintCMSStatistics != 0) { 4122 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 4123 SIZE_FORMAT, _overflow_stack->capacity()); 4124 } 4125 // We cannot assert that the overflow stack is full because 4126 // it may have been emptied since. 4127 assert(simulate_overflow || 4128 _work_queue->size() == _work_queue->max_elems(), 4129 "Else push should have succeeded"); 4130 handle_stack_overflow(addr); 4131 } 4132 } // Else, some other thread got there first 4133 do_yield_check(); 4134 } 4135 } 4136 4137 void Par_ConcMarkingClosure::do_oop(oop* p) { Par_ConcMarkingClosure::do_oop_work(p); } 4138 void Par_ConcMarkingClosure::do_oop(narrowOop* p) { Par_ConcMarkingClosure::do_oop_work(p); } 4139 4140 void Par_ConcMarkingClosure::trim_queue(size_t max) { 4141 while (_work_queue->size() > max) { 4142 oop new_oop; 4143 if (_work_queue->pop_local(new_oop)) { 4144 assert(new_oop->is_oop(), "Should be an oop"); 4145 assert(_bit_map->isMarked((HeapWord*)new_oop), "Grey object"); 4146 assert(_span.contains((HeapWord*)new_oop), "Not in span"); 4147 assert(new_oop->is_parsable(), "Should be parsable"); 4148 new_oop->oop_iterate(this); // do_oop() above 4149 do_yield_check(); 4150 } 4151 } 4152 } 4153 4154 // Upon stack overflow, we discard (part of) the stack, 4155 // remembering the least address amongst those discarded 4156 // in CMSCollector's _restart_address. 4157 void Par_ConcMarkingClosure::handle_stack_overflow(HeapWord* lost) { 4158 // We need to do this under a mutex to prevent other 4159 // workers from interfering with the work done below. 4160 MutexLockerEx ml(_overflow_stack->par_lock(), 4161 Mutex::_no_safepoint_check_flag); 4162 // Remember the least grey address discarded 4163 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 4164 _collector->lower_restart_addr(ra); 4165 _overflow_stack->reset(); // discard stack contents 4166 _overflow_stack->expand(); // expand the stack if possible 4167 } 4168 4169 4170 void CMSConcMarkingTask::do_work_steal(int i) { 4171 OopTaskQueue* work_q = work_queue(i); 4172 oop obj_to_scan; 4173 CMSBitMap* bm = &(_collector->_markBitMap); 4174 CMSMarkStack* ovflw = &(_collector->_markStack); 4175 CMSMarkStack* revisit = &(_collector->_revisitStack); 4176 int* seed = _collector->hash_seed(i); 4177 Par_ConcMarkingClosure cl(_collector, this, work_q, bm, ovflw, revisit); 4178 while (true) { 4179 cl.trim_queue(0); 4180 assert(work_q->size() == 0, "Should have been emptied above"); 4181 if (get_work_from_overflow_stack(ovflw, work_q)) { 4182 // Can't assert below because the work obtained from the 4183 // overflow stack may already have been stolen from us. 4184 // assert(work_q->size() > 0, "Work from overflow stack"); 4185 continue; 4186 } else if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 4187 assert(obj_to_scan->is_oop(), "Should be an oop"); 4188 assert(bm->isMarked((HeapWord*)obj_to_scan), "Grey object"); 4189 obj_to_scan->oop_iterate(&cl); 4190 } else if (terminator()->offer_termination(&_term_term)) { 4191 assert(work_q->size() == 0, "Impossible!"); 4192 break; 4193 } else if (yielding() || should_yield()) { 4194 yield(); 4195 } 4196 } 4197 } 4198 4199 // This is run by the CMS (coordinator) thread. 4200 void CMSConcMarkingTask::coordinator_yield() { 4201 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 4202 "CMS thread should hold CMS token"); 4203 DEBUG_ONLY(RememberKlassesChecker mux(false);) 4204 // First give up the locks, then yield, then re-lock 4205 // We should probably use a constructor/destructor idiom to 4206 // do this unlock/lock or modify the MutexUnlocker class to 4207 // serve our purpose. XXX 4208 assert_lock_strong(_bit_map_lock); 4209 _bit_map_lock->unlock(); 4210 ConcurrentMarkSweepThread::desynchronize(true); 4211 ConcurrentMarkSweepThread::acknowledge_yield_request(); 4212 _collector->stopTimer(); 4213 if (PrintCMSStatistics != 0) { 4214 _collector->incrementYields(); 4215 } 4216 _collector->icms_wait(); 4217 4218 // It is possible for whichever thread initiated the yield request 4219 // not to get a chance to wake up and take the bitmap lock between 4220 // this thread releasing it and reacquiring it. So, while the 4221 // should_yield() flag is on, let's sleep for a bit to give the 4222 // other thread a chance to wake up. The limit imposed on the number 4223 // of iterations is defensive, to avoid any unforseen circumstances 4224 // putting us into an infinite loop. Since it's always been this 4225 // (coordinator_yield()) method that was observed to cause the 4226 // problem, we are using a parameter (CMSCoordinatorYieldSleepCount) 4227 // which is by default non-zero. For the other seven methods that 4228 // also perform the yield operation, as are using a different 4229 // parameter (CMSYieldSleepCount) which is by default zero. This way we 4230 // can enable the sleeping for those methods too, if necessary. 4231 // See 6442774. 4232 // 4233 // We really need to reconsider the synchronization between the GC 4234 // thread and the yield-requesting threads in the future and we 4235 // should really use wait/notify, which is the recommended 4236 // way of doing this type of interaction. Additionally, we should 4237 // consolidate the eight methods that do the yield operation and they 4238 // are almost identical into one for better maintenability and 4239 // readability. See 6445193. 4240 // 4241 // Tony 2006.06.29 4242 for (unsigned i = 0; i < CMSCoordinatorYieldSleepCount && 4243 ConcurrentMarkSweepThread::should_yield() && 4244 !CMSCollector::foregroundGCIsActive(); ++i) { 4245 os::sleep(Thread::current(), 1, false); 4246 ConcurrentMarkSweepThread::acknowledge_yield_request(); 4247 } 4248 4249 ConcurrentMarkSweepThread::synchronize(true); 4250 _bit_map_lock->lock_without_safepoint_check(); 4251 _collector->startTimer(); 4252 } 4253 4254 bool CMSCollector::do_marking_mt(bool asynch) { 4255 assert(ConcGCThreads > 0 && conc_workers() != NULL, "precondition"); 4256 int num_workers = AdaptiveSizePolicy::calc_active_conc_workers( 4257 conc_workers()->total_workers(), 4258 conc_workers()->active_workers(), 4259 Threads::number_of_non_daemon_threads()); 4260 conc_workers()->set_active_workers(num_workers); 4261 4262 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 4263 CompactibleFreeListSpace* perm_space = _permGen->cmsSpace(); 4264 4265 CMSConcMarkingTask tsk(this, 4266 cms_space, 4267 perm_space, 4268 asynch, 4269 conc_workers(), 4270 task_queues()); 4271 4272 // Since the actual number of workers we get may be different 4273 // from the number we requested above, do we need to do anything different 4274 // below? In particular, may be we need to subclass the SequantialSubTasksDone 4275 // class?? XXX 4276 cms_space ->initialize_sequential_subtasks_for_marking(num_workers); 4277 perm_space->initialize_sequential_subtasks_for_marking(num_workers); 4278 4279 // Refs discovery is already non-atomic. 4280 assert(!ref_processor()->discovery_is_atomic(), "Should be non-atomic"); 4281 assert(ref_processor()->discovery_is_mt(), "Discovery should be MT"); 4282 DEBUG_ONLY(RememberKlassesChecker cmx(should_unload_classes());) 4283 conc_workers()->start_task(&tsk); 4284 while (tsk.yielded()) { 4285 tsk.coordinator_yield(); 4286 conc_workers()->continue_task(&tsk); 4287 } 4288 // If the task was aborted, _restart_addr will be non-NULL 4289 assert(tsk.completed() || _restart_addr != NULL, "Inconsistency"); 4290 while (_restart_addr != NULL) { 4291 // XXX For now we do not make use of ABORTED state and have not 4292 // yet implemented the right abort semantics (even in the original 4293 // single-threaded CMS case). That needs some more investigation 4294 // and is deferred for now; see CR# TBF. 07252005YSR. XXX 4295 assert(!CMSAbortSemantics || tsk.aborted(), "Inconsistency"); 4296 // If _restart_addr is non-NULL, a marking stack overflow 4297 // occurred; we need to do a fresh marking iteration from the 4298 // indicated restart address. 4299 if (_foregroundGCIsActive && asynch) { 4300 // We may be running into repeated stack overflows, having 4301 // reached the limit of the stack size, while making very 4302 // slow forward progress. It may be best to bail out and 4303 // let the foreground collector do its job. 4304 // Clear _restart_addr, so that foreground GC 4305 // works from scratch. This avoids the headache of 4306 // a "rescan" which would otherwise be needed because 4307 // of the dirty mod union table & card table. 4308 _restart_addr = NULL; 4309 return false; 4310 } 4311 // Adjust the task to restart from _restart_addr 4312 tsk.reset(_restart_addr); 4313 cms_space ->initialize_sequential_subtasks_for_marking(num_workers, 4314 _restart_addr); 4315 perm_space->initialize_sequential_subtasks_for_marking(num_workers, 4316 _restart_addr); 4317 _restart_addr = NULL; 4318 // Get the workers going again 4319 conc_workers()->start_task(&tsk); 4320 while (tsk.yielded()) { 4321 tsk.coordinator_yield(); 4322 conc_workers()->continue_task(&tsk); 4323 } 4324 } 4325 assert(tsk.completed(), "Inconsistency"); 4326 assert(tsk.result() == true, "Inconsistency"); 4327 return true; 4328 } 4329 4330 bool CMSCollector::do_marking_st(bool asynch) { 4331 ResourceMark rm; 4332 HandleMark hm; 4333 4334 // Temporarily make refs discovery single threaded (non-MT) 4335 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false); 4336 MarkFromRootsClosure markFromRootsClosure(this, _span, &_markBitMap, 4337 &_markStack, &_revisitStack, CMSYield && asynch); 4338 // the last argument to iterate indicates whether the iteration 4339 // should be incremental with periodic yields. 4340 _markBitMap.iterate(&markFromRootsClosure); 4341 // If _restart_addr is non-NULL, a marking stack overflow 4342 // occurred; we need to do a fresh iteration from the 4343 // indicated restart address. 4344 while (_restart_addr != NULL) { 4345 if (_foregroundGCIsActive && asynch) { 4346 // We may be running into repeated stack overflows, having 4347 // reached the limit of the stack size, while making very 4348 // slow forward progress. It may be best to bail out and 4349 // let the foreground collector do its job. 4350 // Clear _restart_addr, so that foreground GC 4351 // works from scratch. This avoids the headache of 4352 // a "rescan" which would otherwise be needed because 4353 // of the dirty mod union table & card table. 4354 _restart_addr = NULL; 4355 return false; // indicating failure to complete marking 4356 } 4357 // Deal with stack overflow: 4358 // we restart marking from _restart_addr 4359 HeapWord* ra = _restart_addr; 4360 markFromRootsClosure.reset(ra); 4361 _restart_addr = NULL; 4362 _markBitMap.iterate(&markFromRootsClosure, ra, _span.end()); 4363 } 4364 return true; 4365 } 4366 4367 void CMSCollector::preclean() { 4368 check_correct_thread_executing(); 4369 assert(Thread::current()->is_ConcurrentGC_thread(), "Wrong thread"); 4370 verify_work_stacks_empty(); 4371 verify_overflow_empty(); 4372 _abort_preclean = false; 4373 if (CMSPrecleaningEnabled) { 4374 _eden_chunk_index = 0; 4375 size_t used = get_eden_used(); 4376 size_t capacity = get_eden_capacity(); 4377 // Don't start sampling unless we will get sufficiently 4378 // many samples. 4379 if (used < (capacity/(CMSScheduleRemarkSamplingRatio * 100) 4380 * CMSScheduleRemarkEdenPenetration)) { 4381 _start_sampling = true; 4382 } else { 4383 _start_sampling = false; 4384 } 4385 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 4386 CMSPhaseAccounting pa(this, "preclean", !PrintGCDetails); 4387 preclean_work(CMSPrecleanRefLists1, CMSPrecleanSurvivors1); 4388 } 4389 CMSTokenSync x(true); // is cms thread 4390 if (CMSPrecleaningEnabled) { 4391 sample_eden(); 4392 _collectorState = AbortablePreclean; 4393 } else { 4394 _collectorState = FinalMarking; 4395 } 4396 verify_work_stacks_empty(); 4397 verify_overflow_empty(); 4398 } 4399 4400 // Try and schedule the remark such that young gen 4401 // occupancy is CMSScheduleRemarkEdenPenetration %. 4402 void CMSCollector::abortable_preclean() { 4403 check_correct_thread_executing(); 4404 assert(CMSPrecleaningEnabled, "Inconsistent control state"); 4405 assert(_collectorState == AbortablePreclean, "Inconsistent control state"); 4406 4407 // If Eden's current occupancy is below this threshold, 4408 // immediately schedule the remark; else preclean 4409 // past the next scavenge in an effort to 4410 // schedule the pause as described avove. By choosing 4411 // CMSScheduleRemarkEdenSizeThreshold >= max eden size 4412 // we will never do an actual abortable preclean cycle. 4413 if (get_eden_used() > CMSScheduleRemarkEdenSizeThreshold) { 4414 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 4415 CMSPhaseAccounting pa(this, "abortable-preclean", !PrintGCDetails); 4416 // We need more smarts in the abortable preclean 4417 // loop below to deal with cases where allocation 4418 // in young gen is very very slow, and our precleaning 4419 // is running a losing race against a horde of 4420 // mutators intent on flooding us with CMS updates 4421 // (dirty cards). 4422 // One, admittedly dumb, strategy is to give up 4423 // after a certain number of abortable precleaning loops 4424 // or after a certain maximum time. We want to make 4425 // this smarter in the next iteration. 4426 // XXX FIX ME!!! YSR 4427 size_t loops = 0, workdone = 0, cumworkdone = 0, waited = 0; 4428 while (!(should_abort_preclean() || 4429 ConcurrentMarkSweepThread::should_terminate())) { 4430 workdone = preclean_work(CMSPrecleanRefLists2, CMSPrecleanSurvivors2); 4431 cumworkdone += workdone; 4432 loops++; 4433 // Voluntarily terminate abortable preclean phase if we have 4434 // been at it for too long. 4435 if ((CMSMaxAbortablePrecleanLoops != 0) && 4436 loops >= CMSMaxAbortablePrecleanLoops) { 4437 if (PrintGCDetails) { 4438 gclog_or_tty->print(" CMS: abort preclean due to loops "); 4439 } 4440 break; 4441 } 4442 if (pa.wallclock_millis() > CMSMaxAbortablePrecleanTime) { 4443 if (PrintGCDetails) { 4444 gclog_or_tty->print(" CMS: abort preclean due to time "); 4445 } 4446 break; 4447 } 4448 // If we are doing little work each iteration, we should 4449 // take a short break. 4450 if (workdone < CMSAbortablePrecleanMinWorkPerIteration) { 4451 // Sleep for some time, waiting for work to accumulate 4452 stopTimer(); 4453 cmsThread()->wait_on_cms_lock(CMSAbortablePrecleanWaitMillis); 4454 startTimer(); 4455 waited++; 4456 } 4457 } 4458 if (PrintCMSStatistics > 0) { 4459 gclog_or_tty->print(" [%d iterations, %d waits, %d cards)] ", 4460 loops, waited, cumworkdone); 4461 } 4462 } 4463 CMSTokenSync x(true); // is cms thread 4464 if (_collectorState != Idling) { 4465 assert(_collectorState == AbortablePreclean, 4466 "Spontaneous state transition?"); 4467 _collectorState = FinalMarking; 4468 } // Else, a foreground collection completed this CMS cycle. 4469 return; 4470 } 4471 4472 // Respond to an Eden sampling opportunity 4473 void CMSCollector::sample_eden() { 4474 // Make sure a young gc cannot sneak in between our 4475 // reading and recording of a sample. 4476 assert(Thread::current()->is_ConcurrentGC_thread(), 4477 "Only the cms thread may collect Eden samples"); 4478 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 4479 "Should collect samples while holding CMS token"); 4480 if (!_start_sampling) { 4481 return; 4482 } 4483 if (_eden_chunk_array) { 4484 if (_eden_chunk_index < _eden_chunk_capacity) { 4485 _eden_chunk_array[_eden_chunk_index] = *_top_addr; // take sample 4486 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr, 4487 "Unexpected state of Eden"); 4488 // We'd like to check that what we just sampled is an oop-start address; 4489 // however, we cannot do that here since the object may not yet have been 4490 // initialized. So we'll instead do the check when we _use_ this sample 4491 // later. 4492 if (_eden_chunk_index == 0 || 4493 (pointer_delta(_eden_chunk_array[_eden_chunk_index], 4494 _eden_chunk_array[_eden_chunk_index-1]) 4495 >= CMSSamplingGrain)) { 4496 _eden_chunk_index++; // commit sample 4497 } 4498 } 4499 } 4500 if ((_collectorState == AbortablePreclean) && !_abort_preclean) { 4501 size_t used = get_eden_used(); 4502 size_t capacity = get_eden_capacity(); 4503 assert(used <= capacity, "Unexpected state of Eden"); 4504 if (used > (capacity/100 * CMSScheduleRemarkEdenPenetration)) { 4505 _abort_preclean = true; 4506 } 4507 } 4508 } 4509 4510 4511 size_t CMSCollector::preclean_work(bool clean_refs, bool clean_survivor) { 4512 assert(_collectorState == Precleaning || 4513 _collectorState == AbortablePreclean, "incorrect state"); 4514 ResourceMark rm; 4515 HandleMark hm; 4516 4517 // Precleaning is currently not MT but the reference processor 4518 // may be set for MT. Disable it temporarily here. 4519 ReferenceProcessor* rp = ref_processor(); 4520 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(rp, false); 4521 4522 // Do one pass of scrubbing the discovered reference lists 4523 // to remove any reference objects with strongly-reachable 4524 // referents. 4525 if (clean_refs) { 4526 CMSPrecleanRefsYieldClosure yield_cl(this); 4527 assert(rp->span().equals(_span), "Spans should be equal"); 4528 CMSKeepAliveClosure keep_alive(this, _span, &_markBitMap, 4529 &_markStack, &_revisitStack, 4530 true /* preclean */); 4531 CMSDrainMarkingStackClosure complete_trace(this, 4532 _span, &_markBitMap, &_markStack, 4533 &keep_alive, true /* preclean */); 4534 4535 // We don't want this step to interfere with a young 4536 // collection because we don't want to take CPU 4537 // or memory bandwidth away from the young GC threads 4538 // (which may be as many as there are CPUs). 4539 // Note that we don't need to protect ourselves from 4540 // interference with mutators because they can't 4541 // manipulate the discovered reference lists nor affect 4542 // the computed reachability of the referents, the 4543 // only properties manipulated by the precleaning 4544 // of these reference lists. 4545 stopTimer(); 4546 CMSTokenSyncWithLocks x(true /* is cms thread */, 4547 bitMapLock()); 4548 startTimer(); 4549 sample_eden(); 4550 4551 // The following will yield to allow foreground 4552 // collection to proceed promptly. XXX YSR: 4553 // The code in this method may need further 4554 // tweaking for better performance and some restructuring 4555 // for cleaner interfaces. 4556 GCTimer *gc_timer = NULL; // Currently not tracing concurrent phases 4557 rp->preclean_discovered_references( 4558 rp->is_alive_non_header(), &keep_alive, &complete_trace, 4559 &yield_cl, should_unload_classes(), gc_timer); 4560 } 4561 4562 if (clean_survivor) { // preclean the active survivor space(s) 4563 assert(_young_gen->kind() == Generation::DefNew || 4564 _young_gen->kind() == Generation::ParNew || 4565 _young_gen->kind() == Generation::ASParNew, 4566 "incorrect type for cast"); 4567 DefNewGeneration* dng = (DefNewGeneration*)_young_gen; 4568 PushAndMarkClosure pam_cl(this, _span, ref_processor(), 4569 &_markBitMap, &_modUnionTable, 4570 &_markStack, &_revisitStack, 4571 true /* precleaning phase */); 4572 stopTimer(); 4573 CMSTokenSyncWithLocks ts(true /* is cms thread */, 4574 bitMapLock()); 4575 startTimer(); 4576 unsigned int before_count = 4577 GenCollectedHeap::heap()->total_collections(); 4578 SurvivorSpacePrecleanClosure 4579 sss_cl(this, _span, &_markBitMap, &_markStack, 4580 &pam_cl, before_count, CMSYield); 4581 DEBUG_ONLY(RememberKlassesChecker mx(should_unload_classes());) 4582 dng->from()->object_iterate_careful(&sss_cl); 4583 dng->to()->object_iterate_careful(&sss_cl); 4584 } 4585 MarkRefsIntoAndScanClosure 4586 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable, 4587 &_markStack, &_revisitStack, this, CMSYield, 4588 true /* precleaning phase */); 4589 // CAUTION: The following closure has persistent state that may need to 4590 // be reset upon a decrease in the sequence of addresses it 4591 // processes. 4592 ScanMarkedObjectsAgainCarefullyClosure 4593 smoac_cl(this, _span, 4594 &_markBitMap, &_markStack, &_revisitStack, &mrias_cl, CMSYield); 4595 4596 // Preclean dirty cards in ModUnionTable and CardTable using 4597 // appropriate convergence criterion; 4598 // repeat CMSPrecleanIter times unless we find that 4599 // we are losing. 4600 assert(CMSPrecleanIter < 10, "CMSPrecleanIter is too large"); 4601 assert(CMSPrecleanNumerator < CMSPrecleanDenominator, 4602 "Bad convergence multiplier"); 4603 assert(CMSPrecleanThreshold >= 100, 4604 "Unreasonably low CMSPrecleanThreshold"); 4605 4606 size_t numIter, cumNumCards, lastNumCards, curNumCards; 4607 for (numIter = 0, cumNumCards = lastNumCards = curNumCards = 0; 4608 numIter < CMSPrecleanIter; 4609 numIter++, lastNumCards = curNumCards, cumNumCards += curNumCards) { 4610 curNumCards = preclean_mod_union_table(_cmsGen, &smoac_cl); 4611 if (CMSPermGenPrecleaningEnabled) { 4612 curNumCards += preclean_mod_union_table(_permGen, &smoac_cl); 4613 } 4614 if (Verbose && PrintGCDetails) { 4615 gclog_or_tty->print(" (modUnionTable: %d cards)", curNumCards); 4616 } 4617 // Either there are very few dirty cards, so re-mark 4618 // pause will be small anyway, or our pre-cleaning isn't 4619 // that much faster than the rate at which cards are being 4620 // dirtied, so we might as well stop and re-mark since 4621 // precleaning won't improve our re-mark time by much. 4622 if (curNumCards <= CMSPrecleanThreshold || 4623 (numIter > 0 && 4624 (curNumCards * CMSPrecleanDenominator > 4625 lastNumCards * CMSPrecleanNumerator))) { 4626 numIter++; 4627 cumNumCards += curNumCards; 4628 break; 4629 } 4630 } 4631 curNumCards = preclean_card_table(_cmsGen, &smoac_cl); 4632 if (CMSPermGenPrecleaningEnabled) { 4633 curNumCards += preclean_card_table(_permGen, &smoac_cl); 4634 } 4635 cumNumCards += curNumCards; 4636 if (PrintGCDetails && PrintCMSStatistics != 0) { 4637 gclog_or_tty->print_cr(" (cardTable: %d cards, re-scanned %d cards, %d iterations)", 4638 curNumCards, cumNumCards, numIter); 4639 } 4640 return cumNumCards; // as a measure of useful work done 4641 } 4642 4643 // PRECLEANING NOTES: 4644 // Precleaning involves: 4645 // . reading the bits of the modUnionTable and clearing the set bits. 4646 // . For the cards corresponding to the set bits, we scan the 4647 // objects on those cards. This means we need the free_list_lock 4648 // so that we can safely iterate over the CMS space when scanning 4649 // for oops. 4650 // . When we scan the objects, we'll be both reading and setting 4651 // marks in the marking bit map, so we'll need the marking bit map. 4652 // . For protecting _collector_state transitions, we take the CGC_lock. 4653 // Note that any races in the reading of of card table entries by the 4654 // CMS thread on the one hand and the clearing of those entries by the 4655 // VM thread or the setting of those entries by the mutator threads on the 4656 // other are quite benign. However, for efficiency it makes sense to keep 4657 // the VM thread from racing with the CMS thread while the latter is 4658 // dirty card info to the modUnionTable. We therefore also use the 4659 // CGC_lock to protect the reading of the card table and the mod union 4660 // table by the CM thread. 4661 // . We run concurrently with mutator updates, so scanning 4662 // needs to be done carefully -- we should not try to scan 4663 // potentially uninitialized objects. 4664 // 4665 // Locking strategy: While holding the CGC_lock, we scan over and 4666 // reset a maximal dirty range of the mod union / card tables, then lock 4667 // the free_list_lock and bitmap lock to do a full marking, then 4668 // release these locks; and repeat the cycle. This allows for a 4669 // certain amount of fairness in the sharing of these locks between 4670 // the CMS collector on the one hand, and the VM thread and the 4671 // mutators on the other. 4672 4673 // NOTE: preclean_mod_union_table() and preclean_card_table() 4674 // further below are largely identical; if you need to modify 4675 // one of these methods, please check the other method too. 4676 4677 size_t CMSCollector::preclean_mod_union_table( 4678 ConcurrentMarkSweepGeneration* gen, 4679 ScanMarkedObjectsAgainCarefullyClosure* cl) { 4680 verify_work_stacks_empty(); 4681 verify_overflow_empty(); 4682 4683 // Turn off checking for this method but turn it back on 4684 // selectively. There are yield points in this method 4685 // but it is difficult to turn the checking off just around 4686 // the yield points. It is simpler to selectively turn 4687 // it on. 4688 DEBUG_ONLY(RememberKlassesChecker mux(false);) 4689 4690 // strategy: starting with the first card, accumulate contiguous 4691 // ranges of dirty cards; clear these cards, then scan the region 4692 // covered by these cards. 4693 4694 // Since all of the MUT is committed ahead, we can just use 4695 // that, in case the generations expand while we are precleaning. 4696 // It might also be fine to just use the committed part of the 4697 // generation, but we might potentially miss cards when the 4698 // generation is rapidly expanding while we are in the midst 4699 // of precleaning. 4700 HeapWord* startAddr = gen->reserved().start(); 4701 HeapWord* endAddr = gen->reserved().end(); 4702 4703 cl->setFreelistLock(gen->freelistLock()); // needed for yielding 4704 4705 size_t numDirtyCards, cumNumDirtyCards; 4706 HeapWord *nextAddr, *lastAddr; 4707 for (cumNumDirtyCards = numDirtyCards = 0, 4708 nextAddr = lastAddr = startAddr; 4709 nextAddr < endAddr; 4710 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 4711 4712 ResourceMark rm; 4713 HandleMark hm; 4714 4715 MemRegion dirtyRegion; 4716 { 4717 stopTimer(); 4718 // Potential yield point 4719 CMSTokenSync ts(true); 4720 startTimer(); 4721 sample_eden(); 4722 // Get dirty region starting at nextOffset (inclusive), 4723 // simultaneously clearing it. 4724 dirtyRegion = 4725 _modUnionTable.getAndClearMarkedRegion(nextAddr, endAddr); 4726 assert(dirtyRegion.start() >= nextAddr, 4727 "returned region inconsistent?"); 4728 } 4729 // Remember where the next search should begin. 4730 // The returned region (if non-empty) is a right open interval, 4731 // so lastOffset is obtained from the right end of that 4732 // interval. 4733 lastAddr = dirtyRegion.end(); 4734 // Should do something more transparent and less hacky XXX 4735 numDirtyCards = 4736 _modUnionTable.heapWordDiffToOffsetDiff(dirtyRegion.word_size()); 4737 4738 // We'll scan the cards in the dirty region (with periodic 4739 // yields for foreground GC as needed). 4740 if (!dirtyRegion.is_empty()) { 4741 assert(numDirtyCards > 0, "consistency check"); 4742 HeapWord* stop_point = NULL; 4743 stopTimer(); 4744 // Potential yield point 4745 CMSTokenSyncWithLocks ts(true, gen->freelistLock(), 4746 bitMapLock()); 4747 startTimer(); 4748 { 4749 verify_work_stacks_empty(); 4750 verify_overflow_empty(); 4751 sample_eden(); 4752 DEBUG_ONLY(RememberKlassesChecker mx(should_unload_classes());) 4753 stop_point = 4754 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 4755 } 4756 if (stop_point != NULL) { 4757 // The careful iteration stopped early either because it found an 4758 // uninitialized object, or because we were in the midst of an 4759 // "abortable preclean", which should now be aborted. Redirty 4760 // the bits corresponding to the partially-scanned or unscanned 4761 // cards. We'll either restart at the next block boundary or 4762 // abort the preclean. 4763 assert((CMSPermGenPrecleaningEnabled && (gen == _permGen)) || 4764 (_collectorState == AbortablePreclean && should_abort_preclean()), 4765 "Unparsable objects should only be in perm gen."); 4766 _modUnionTable.mark_range(MemRegion(stop_point, dirtyRegion.end())); 4767 if (should_abort_preclean()) { 4768 break; // out of preclean loop 4769 } else { 4770 // Compute the next address at which preclean should pick up; 4771 // might need bitMapLock in order to read P-bits. 4772 lastAddr = next_card_start_after_block(stop_point); 4773 } 4774 } 4775 } else { 4776 assert(lastAddr == endAddr, "consistency check"); 4777 assert(numDirtyCards == 0, "consistency check"); 4778 break; 4779 } 4780 } 4781 verify_work_stacks_empty(); 4782 verify_overflow_empty(); 4783 return cumNumDirtyCards; 4784 } 4785 4786 // NOTE: preclean_mod_union_table() above and preclean_card_table() 4787 // below are largely identical; if you need to modify 4788 // one of these methods, please check the other method too. 4789 4790 size_t CMSCollector::preclean_card_table(ConcurrentMarkSweepGeneration* gen, 4791 ScanMarkedObjectsAgainCarefullyClosure* cl) { 4792 // strategy: it's similar to precleamModUnionTable above, in that 4793 // we accumulate contiguous ranges of dirty cards, mark these cards 4794 // precleaned, then scan the region covered by these cards. 4795 HeapWord* endAddr = (HeapWord*)(gen->_virtual_space.high()); 4796 HeapWord* startAddr = (HeapWord*)(gen->_virtual_space.low()); 4797 4798 cl->setFreelistLock(gen->freelistLock()); // needed for yielding 4799 4800 size_t numDirtyCards, cumNumDirtyCards; 4801 HeapWord *lastAddr, *nextAddr; 4802 4803 for (cumNumDirtyCards = numDirtyCards = 0, 4804 nextAddr = lastAddr = startAddr; 4805 nextAddr < endAddr; 4806 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 4807 4808 ResourceMark rm; 4809 HandleMark hm; 4810 4811 MemRegion dirtyRegion; 4812 { 4813 // See comments in "Precleaning notes" above on why we 4814 // do this locking. XXX Could the locking overheads be 4815 // too high when dirty cards are sparse? [I don't think so.] 4816 stopTimer(); 4817 CMSTokenSync x(true); // is cms thread 4818 startTimer(); 4819 sample_eden(); 4820 // Get and clear dirty region from card table 4821 dirtyRegion = _ct->ct_bs()->dirty_card_range_after_reset( 4822 MemRegion(nextAddr, endAddr), 4823 true, 4824 CardTableModRefBS::precleaned_card_val()); 4825 4826 assert(dirtyRegion.start() >= nextAddr, 4827 "returned region inconsistent?"); 4828 } 4829 lastAddr = dirtyRegion.end(); 4830 numDirtyCards = 4831 dirtyRegion.word_size()/CardTableModRefBS::card_size_in_words; 4832 4833 if (!dirtyRegion.is_empty()) { 4834 stopTimer(); 4835 CMSTokenSyncWithLocks ts(true, gen->freelistLock(), bitMapLock()); 4836 startTimer(); 4837 sample_eden(); 4838 verify_work_stacks_empty(); 4839 verify_overflow_empty(); 4840 DEBUG_ONLY(RememberKlassesChecker mx(should_unload_classes());) 4841 HeapWord* stop_point = 4842 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 4843 if (stop_point != NULL) { 4844 // The careful iteration stopped early because it found an 4845 // uninitialized object. Redirty the bits corresponding to the 4846 // partially-scanned or unscanned cards, and start again at the 4847 // next block boundary. 4848 assert(CMSPermGenPrecleaningEnabled || 4849 (_collectorState == AbortablePreclean && should_abort_preclean()), 4850 "Unparsable objects should only be in perm gen."); 4851 _ct->ct_bs()->invalidate(MemRegion(stop_point, dirtyRegion.end())); 4852 if (should_abort_preclean()) { 4853 break; // out of preclean loop 4854 } else { 4855 // Compute the next address at which preclean should pick up. 4856 lastAddr = next_card_start_after_block(stop_point); 4857 } 4858 } 4859 } else { 4860 break; 4861 } 4862 } 4863 verify_work_stacks_empty(); 4864 verify_overflow_empty(); 4865 return cumNumDirtyCards; 4866 } 4867 4868 void CMSCollector::checkpointRootsFinal(bool asynch, 4869 bool clear_all_soft_refs, bool init_mark_was_synchronous) { 4870 assert(_collectorState == FinalMarking, "incorrect state transition?"); 4871 check_correct_thread_executing(); 4872 // world is stopped at this checkpoint 4873 assert(SafepointSynchronize::is_at_safepoint(), 4874 "world should be stopped"); 4875 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 4876 4877 verify_work_stacks_empty(); 4878 verify_overflow_empty(); 4879 4880 SpecializationStats::clear(); 4881 if (PrintGCDetails) { 4882 gclog_or_tty->print("[YG occupancy: "SIZE_FORMAT" K ("SIZE_FORMAT" K)]", 4883 _young_gen->used() / K, 4884 _young_gen->capacity() / K); 4885 } 4886 if (asynch) { 4887 if (CMSScavengeBeforeRemark) { 4888 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4889 // Temporarily set flag to false, GCH->do_collection will 4890 // expect it to be false and set to true 4891 FlagSetting fl(gch->_is_gc_active, false); 4892 NOT_PRODUCT(GCTraceTime t("Scavenge-Before-Remark", 4893 PrintGCDetails && Verbose, true, _gc_timer_cm);) 4894 int level = _cmsGen->level() - 1; 4895 if (level >= 0) { 4896 gch->do_collection(true, // full (i.e. force, see below) 4897 false, // !clear_all_soft_refs 4898 0, // size 4899 false, // is_tlab 4900 level // max_level 4901 ); 4902 } 4903 } 4904 FreelistLocker x(this); 4905 MutexLockerEx y(bitMapLock(), 4906 Mutex::_no_safepoint_check_flag); 4907 assert(!init_mark_was_synchronous, "but that's impossible!"); 4908 checkpointRootsFinalWork(asynch, clear_all_soft_refs, false); 4909 } else { 4910 // already have all the locks 4911 checkpointRootsFinalWork(asynch, clear_all_soft_refs, 4912 init_mark_was_synchronous); 4913 } 4914 verify_work_stacks_empty(); 4915 verify_overflow_empty(); 4916 SpecializationStats::print(); 4917 } 4918 4919 void CMSCollector::checkpointRootsFinalWork(bool asynch, 4920 bool clear_all_soft_refs, bool init_mark_was_synchronous) { 4921 4922 NOT_PRODUCT(GCTraceTime tr("checkpointRootsFinalWork", PrintGCDetails, false, _gc_timer_cm);) 4923 4924 assert(haveFreelistLocks(), "must have free list locks"); 4925 assert_lock_strong(bitMapLock()); 4926 4927 if (UseAdaptiveSizePolicy) { 4928 size_policy()->checkpoint_roots_final_begin(); 4929 } 4930 4931 ResourceMark rm; 4932 HandleMark hm; 4933 4934 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4935 4936 if (should_unload_classes()) { 4937 CodeCache::gc_prologue(); 4938 } 4939 assert(haveFreelistLocks(), "must have free list locks"); 4940 assert_lock_strong(bitMapLock()); 4941 4942 DEBUG_ONLY(RememberKlassesChecker fmx(should_unload_classes());) 4943 if (!init_mark_was_synchronous) { 4944 // We might assume that we need not fill TLAB's when 4945 // CMSScavengeBeforeRemark is set, because we may have just done 4946 // a scavenge which would have filled all TLAB's -- and besides 4947 // Eden would be empty. This however may not always be the case -- 4948 // for instance although we asked for a scavenge, it may not have 4949 // happened because of a JNI critical section. We probably need 4950 // a policy for deciding whether we can in that case wait until 4951 // the critical section releases and then do the remark following 4952 // the scavenge, and skip it here. In the absence of that policy, 4953 // or of an indication of whether the scavenge did indeed occur, 4954 // we cannot rely on TLAB's having been filled and must do 4955 // so here just in case a scavenge did not happen. 4956 gch->ensure_parsability(false); // fill TLAB's, but no need to retire them 4957 // Update the saved marks which may affect the root scans. 4958 gch->save_marks(); 4959 4960 { 4961 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;) 4962 4963 // Note on the role of the mod union table: 4964 // Since the marker in "markFromRoots" marks concurrently with 4965 // mutators, it is possible for some reachable objects not to have been 4966 // scanned. For instance, an only reference to an object A was 4967 // placed in object B after the marker scanned B. Unless B is rescanned, 4968 // A would be collected. Such updates to references in marked objects 4969 // are detected via the mod union table which is the set of all cards 4970 // dirtied since the first checkpoint in this GC cycle and prior to 4971 // the most recent young generation GC, minus those cleaned up by the 4972 // concurrent precleaning. 4973 if (CMSParallelRemarkEnabled && CollectedHeap::use_parallel_gc_threads()) { 4974 GCTraceTime t("Rescan (parallel) ", PrintGCDetails, false, _gc_timer_cm); 4975 do_remark_parallel(); 4976 } else { 4977 GCTraceTime t("Rescan (non-parallel) ", PrintGCDetails, false, 4978 _gc_timer_cm); 4979 do_remark_non_parallel(); 4980 } 4981 } 4982 } else { 4983 assert(!asynch, "Can't have init_mark_was_synchronous in asynch mode"); 4984 // The initial mark was stop-world, so there's no rescanning to 4985 // do; go straight on to the next step below. 4986 } 4987 verify_work_stacks_empty(); 4988 verify_overflow_empty(); 4989 4990 { 4991 NOT_PRODUCT(GCTraceTime ts("refProcessingWork", PrintGCDetails, false, _gc_timer_cm);) 4992 refProcessingWork(asynch, clear_all_soft_refs); 4993 } 4994 verify_work_stacks_empty(); 4995 verify_overflow_empty(); 4996 4997 if (should_unload_classes()) { 4998 CodeCache::gc_epilogue(); 4999 } 5000 JvmtiExport::gc_epilogue(); 5001 5002 // If we encountered any (marking stack / work queue) overflow 5003 // events during the current CMS cycle, take appropriate 5004 // remedial measures, where possible, so as to try and avoid 5005 // recurrence of that condition. 5006 assert(_markStack.isEmpty(), "No grey objects"); 5007 size_t ser_ovflw = _ser_pmc_remark_ovflw + _ser_pmc_preclean_ovflw + 5008 _ser_kac_ovflw + _ser_kac_preclean_ovflw; 5009 if (ser_ovflw > 0) { 5010 if (PrintCMSStatistics != 0) { 5011 gclog_or_tty->print_cr("Marking stack overflow (benign) " 5012 "(pmc_pc="SIZE_FORMAT", pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT 5013 ", kac_preclean="SIZE_FORMAT")", 5014 _ser_pmc_preclean_ovflw, _ser_pmc_remark_ovflw, 5015 _ser_kac_ovflw, _ser_kac_preclean_ovflw); 5016 } 5017 _markStack.expand(); 5018 _ser_pmc_remark_ovflw = 0; 5019 _ser_pmc_preclean_ovflw = 0; 5020 _ser_kac_preclean_ovflw = 0; 5021 _ser_kac_ovflw = 0; 5022 } 5023 if (_par_pmc_remark_ovflw > 0 || _par_kac_ovflw > 0) { 5024 if (PrintCMSStatistics != 0) { 5025 gclog_or_tty->print_cr("Work queue overflow (benign) " 5026 "(pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT")", 5027 _par_pmc_remark_ovflw, _par_kac_ovflw); 5028 } 5029 _par_pmc_remark_ovflw = 0; 5030 _par_kac_ovflw = 0; 5031 } 5032 if (PrintCMSStatistics != 0) { 5033 if (_markStack._hit_limit > 0) { 5034 gclog_or_tty->print_cr(" (benign) Hit max stack size limit ("SIZE_FORMAT")", 5035 _markStack._hit_limit); 5036 } 5037 if (_markStack._failed_double > 0) { 5038 gclog_or_tty->print_cr(" (benign) Failed stack doubling ("SIZE_FORMAT")," 5039 " current capacity "SIZE_FORMAT, 5040 _markStack._failed_double, 5041 _markStack.capacity()); 5042 } 5043 } 5044 _markStack._hit_limit = 0; 5045 _markStack._failed_double = 0; 5046 5047 // Check that all the klasses have been checked 5048 assert(_revisitStack.isEmpty(), "Not all klasses revisited"); 5049 5050 if ((VerifyAfterGC || VerifyDuringGC) && 5051 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 5052 verify_after_remark(); 5053 } 5054 5055 // Change under the freelistLocks. 5056 _collectorState = Sweeping; 5057 // Call isAllClear() under bitMapLock 5058 assert(_modUnionTable.isAllClear(), "Should be clear by end of the" 5059 " final marking"); 5060 if (UseAdaptiveSizePolicy) { 5061 size_policy()->checkpoint_roots_final_end(gch->gc_cause()); 5062 } 5063 } 5064 5065 // Parallel remark task 5066 class CMSParRemarkTask: public AbstractGangTask { 5067 CMSCollector* _collector; 5068 int _n_workers; 5069 CompactibleFreeListSpace* _cms_space; 5070 CompactibleFreeListSpace* _perm_space; 5071 5072 // The per-thread work queues, available here for stealing. 5073 OopTaskQueueSet* _task_queues; 5074 ParallelTaskTerminator _term; 5075 5076 public: 5077 // A value of 0 passed to n_workers will cause the number of 5078 // workers to be taken from the active workers in the work gang. 5079 CMSParRemarkTask(CMSCollector* collector, 5080 CompactibleFreeListSpace* cms_space, 5081 CompactibleFreeListSpace* perm_space, 5082 int n_workers, FlexibleWorkGang* workers, 5083 OopTaskQueueSet* task_queues): 5084 AbstractGangTask("Rescan roots and grey objects in parallel"), 5085 _collector(collector), 5086 _cms_space(cms_space), _perm_space(perm_space), 5087 _n_workers(n_workers), 5088 _task_queues(task_queues), 5089 _term(n_workers, task_queues) { } 5090 5091 OopTaskQueueSet* task_queues() { return _task_queues; } 5092 5093 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 5094 5095 ParallelTaskTerminator* terminator() { return &_term; } 5096 int n_workers() { return _n_workers; } 5097 5098 void work(uint worker_id); 5099 5100 private: 5101 // Work method in support of parallel rescan ... of young gen spaces 5102 void do_young_space_rescan(int i, Par_MarkRefsIntoAndScanClosure* cl, 5103 ContiguousSpace* space, 5104 HeapWord** chunk_array, size_t chunk_top); 5105 5106 // ... of dirty cards in old space 5107 void do_dirty_card_rescan_tasks(CompactibleFreeListSpace* sp, int i, 5108 Par_MarkRefsIntoAndScanClosure* cl); 5109 5110 // ... work stealing for the above 5111 void do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, int* seed); 5112 }; 5113 5114 // work_queue(i) is passed to the closure 5115 // Par_MarkRefsIntoAndScanClosure. The "i" parameter 5116 // also is passed to do_dirty_card_rescan_tasks() and to 5117 // do_work_steal() to select the i-th task_queue. 5118 5119 void CMSParRemarkTask::work(uint worker_id) { 5120 elapsedTimer _timer; 5121 ResourceMark rm; 5122 HandleMark hm; 5123 5124 // ---------- rescan from roots -------------- 5125 _timer.start(); 5126 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5127 Par_MarkRefsIntoAndScanClosure par_mrias_cl(_collector, 5128 _collector->_span, _collector->ref_processor(), 5129 &(_collector->_markBitMap), 5130 work_queue(worker_id), &(_collector->_revisitStack)); 5131 5132 // Rescan young gen roots first since these are likely 5133 // coarsely partitioned and may, on that account, constitute 5134 // the critical path; thus, it's best to start off that 5135 // work first. 5136 // ---------- young gen roots -------------- 5137 { 5138 DefNewGeneration* dng = _collector->_young_gen->as_DefNewGeneration(); 5139 EdenSpace* eden_space = dng->eden(); 5140 ContiguousSpace* from_space = dng->from(); 5141 ContiguousSpace* to_space = dng->to(); 5142 5143 HeapWord** eca = _collector->_eden_chunk_array; 5144 size_t ect = _collector->_eden_chunk_index; 5145 HeapWord** sca = _collector->_survivor_chunk_array; 5146 size_t sct = _collector->_survivor_chunk_index; 5147 5148 assert(ect <= _collector->_eden_chunk_capacity, "out of bounds"); 5149 assert(sct <= _collector->_survivor_chunk_capacity, "out of bounds"); 5150 5151 do_young_space_rescan(worker_id, &par_mrias_cl, to_space, NULL, 0); 5152 do_young_space_rescan(worker_id, &par_mrias_cl, from_space, sca, sct); 5153 do_young_space_rescan(worker_id, &par_mrias_cl, eden_space, eca, ect); 5154 5155 _timer.stop(); 5156 if (PrintCMSStatistics != 0) { 5157 gclog_or_tty->print_cr( 5158 "Finished young gen rescan work in %dth thread: %3.3f sec", 5159 worker_id, _timer.seconds()); 5160 } 5161 } 5162 5163 // ---------- remaining roots -------------- 5164 _timer.reset(); 5165 _timer.start(); 5166 gch->gen_process_strong_roots(_collector->_cmsGen->level(), 5167 false, // yg was scanned above 5168 false, // this is parallel code 5169 true, // collecting perm gen 5170 SharedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()), 5171 &par_mrias_cl, 5172 true, // walk all of code cache if (so & SO_CodeCache) 5173 NULL); 5174 assert(_collector->should_unload_classes() 5175 || (_collector->CMSCollector::roots_scanning_options() & SharedHeap::SO_CodeCache), 5176 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 5177 _timer.stop(); 5178 if (PrintCMSStatistics != 0) { 5179 gclog_or_tty->print_cr( 5180 "Finished remaining root rescan work in %dth thread: %3.3f sec", 5181 worker_id, _timer.seconds()); 5182 } 5183 5184 // ---------- rescan dirty cards ------------ 5185 _timer.reset(); 5186 _timer.start(); 5187 5188 // Do the rescan tasks for each of the two spaces 5189 // (cms_space and perm_space) in turn. 5190 // "worker_id" is passed to select the task_queue for "worker_id" 5191 do_dirty_card_rescan_tasks(_cms_space, worker_id, &par_mrias_cl); 5192 do_dirty_card_rescan_tasks(_perm_space, worker_id, &par_mrias_cl); 5193 _timer.stop(); 5194 if (PrintCMSStatistics != 0) { 5195 gclog_or_tty->print_cr( 5196 "Finished dirty card rescan work in %dth thread: %3.3f sec", 5197 worker_id, _timer.seconds()); 5198 } 5199 5200 // ---------- steal work from other threads ... 5201 // ---------- ... and drain overflow list. 5202 _timer.reset(); 5203 _timer.start(); 5204 do_work_steal(worker_id, &par_mrias_cl, _collector->hash_seed(worker_id)); 5205 _timer.stop(); 5206 if (PrintCMSStatistics != 0) { 5207 gclog_or_tty->print_cr( 5208 "Finished work stealing in %dth thread: %3.3f sec", 5209 worker_id, _timer.seconds()); 5210 } 5211 } 5212 5213 // Note that parameter "i" is not used. 5214 void 5215 CMSParRemarkTask::do_young_space_rescan(int i, 5216 Par_MarkRefsIntoAndScanClosure* cl, ContiguousSpace* space, 5217 HeapWord** chunk_array, size_t chunk_top) { 5218 // Until all tasks completed: 5219 // . claim an unclaimed task 5220 // . compute region boundaries corresponding to task claimed 5221 // using chunk_array 5222 // . par_oop_iterate(cl) over that region 5223 5224 ResourceMark rm; 5225 HandleMark hm; 5226 5227 SequentialSubTasksDone* pst = space->par_seq_tasks(); 5228 assert(pst->valid(), "Uninitialized use?"); 5229 5230 uint nth_task = 0; 5231 uint n_tasks = pst->n_tasks(); 5232 5233 HeapWord *start, *end; 5234 while (!pst->is_task_claimed(/* reference */ nth_task)) { 5235 // We claimed task # nth_task; compute its boundaries. 5236 if (chunk_top == 0) { // no samples were taken 5237 assert(nth_task == 0 && n_tasks == 1, "Can have only 1 EdenSpace task"); 5238 start = space->bottom(); 5239 end = space->top(); 5240 } else if (nth_task == 0) { 5241 start = space->bottom(); 5242 end = chunk_array[nth_task]; 5243 } else if (nth_task < (uint)chunk_top) { 5244 assert(nth_task >= 1, "Control point invariant"); 5245 start = chunk_array[nth_task - 1]; 5246 end = chunk_array[nth_task]; 5247 } else { 5248 assert(nth_task == (uint)chunk_top, "Control point invariant"); 5249 start = chunk_array[chunk_top - 1]; 5250 end = space->top(); 5251 } 5252 MemRegion mr(start, end); 5253 // Verify that mr is in space 5254 assert(mr.is_empty() || space->used_region().contains(mr), 5255 "Should be in space"); 5256 // Verify that "start" is an object boundary 5257 assert(mr.is_empty() || oop(mr.start())->is_oop(), 5258 "Should be an oop"); 5259 space->par_oop_iterate(mr, cl); 5260 } 5261 pst->all_tasks_completed(); 5262 } 5263 5264 void 5265 CMSParRemarkTask::do_dirty_card_rescan_tasks( 5266 CompactibleFreeListSpace* sp, int i, 5267 Par_MarkRefsIntoAndScanClosure* cl) { 5268 // Until all tasks completed: 5269 // . claim an unclaimed task 5270 // . compute region boundaries corresponding to task claimed 5271 // . transfer dirty bits ct->mut for that region 5272 // . apply rescanclosure to dirty mut bits for that region 5273 5274 ResourceMark rm; 5275 HandleMark hm; 5276 5277 OopTaskQueue* work_q = work_queue(i); 5278 ModUnionClosure modUnionClosure(&(_collector->_modUnionTable)); 5279 // CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! 5280 // CAUTION: This closure has state that persists across calls to 5281 // the work method dirty_range_iterate_clear() in that it has 5282 // imbedded in it a (subtype of) UpwardsObjectClosure. The 5283 // use of that state in the imbedded UpwardsObjectClosure instance 5284 // assumes that the cards are always iterated (even if in parallel 5285 // by several threads) in monotonically increasing order per each 5286 // thread. This is true of the implementation below which picks 5287 // card ranges (chunks) in monotonically increasing order globally 5288 // and, a-fortiori, in monotonically increasing order per thread 5289 // (the latter order being a subsequence of the former). 5290 // If the work code below is ever reorganized into a more chaotic 5291 // work-partitioning form than the current "sequential tasks" 5292 // paradigm, the use of that persistent state will have to be 5293 // revisited and modified appropriately. See also related 5294 // bug 4756801 work on which should examine this code to make 5295 // sure that the changes there do not run counter to the 5296 // assumptions made here and necessary for correctness and 5297 // efficiency. Note also that this code might yield inefficient 5298 // behaviour in the case of very large objects that span one or 5299 // more work chunks. Such objects would potentially be scanned 5300 // several times redundantly. Work on 4756801 should try and 5301 // address that performance anomaly if at all possible. XXX 5302 MemRegion full_span = _collector->_span; 5303 CMSBitMap* bm = &(_collector->_markBitMap); // shared 5304 CMSMarkStack* rs = &(_collector->_revisitStack); // shared 5305 MarkFromDirtyCardsClosure 5306 greyRescanClosure(_collector, full_span, // entire span of interest 5307 sp, bm, work_q, rs, cl); 5308 5309 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 5310 assert(pst->valid(), "Uninitialized use?"); 5311 uint nth_task = 0; 5312 const int alignment = CardTableModRefBS::card_size * BitsPerWord; 5313 MemRegion span = sp->used_region(); 5314 HeapWord* start_addr = span.start(); 5315 HeapWord* end_addr = (HeapWord*)round_to((intptr_t)span.end(), 5316 alignment); 5317 const size_t chunk_size = sp->rescan_task_size(); // in HeapWord units 5318 assert((HeapWord*)round_to((intptr_t)start_addr, alignment) == 5319 start_addr, "Check alignment"); 5320 assert((size_t)round_to((intptr_t)chunk_size, alignment) == 5321 chunk_size, "Check alignment"); 5322 5323 while (!pst->is_task_claimed(/* reference */ nth_task)) { 5324 // Having claimed the nth_task, compute corresponding mem-region, 5325 // which is a-fortiori aligned correctly (i.e. at a MUT bopundary). 5326 // The alignment restriction ensures that we do not need any 5327 // synchronization with other gang-workers while setting or 5328 // clearing bits in thus chunk of the MUT. 5329 MemRegion this_span = MemRegion(start_addr + nth_task*chunk_size, 5330 start_addr + (nth_task+1)*chunk_size); 5331 // The last chunk's end might be way beyond end of the 5332 // used region. In that case pull back appropriately. 5333 if (this_span.end() > end_addr) { 5334 this_span.set_end(end_addr); 5335 assert(!this_span.is_empty(), "Program logic (calculation of n_tasks)"); 5336 } 5337 // Iterate over the dirty cards covering this chunk, marking them 5338 // precleaned, and setting the corresponding bits in the mod union 5339 // table. Since we have been careful to partition at Card and MUT-word 5340 // boundaries no synchronization is needed between parallel threads. 5341 _collector->_ct->ct_bs()->dirty_card_iterate(this_span, 5342 &modUnionClosure); 5343 5344 // Having transferred these marks into the modUnionTable, 5345 // rescan the marked objects on the dirty cards in the modUnionTable. 5346 // Even if this is at a synchronous collection, the initial marking 5347 // may have been done during an asynchronous collection so there 5348 // may be dirty bits in the mod-union table. 5349 _collector->_modUnionTable.dirty_range_iterate_clear( 5350 this_span, &greyRescanClosure); 5351 _collector->_modUnionTable.verifyNoOneBitsInRange( 5352 this_span.start(), 5353 this_span.end()); 5354 } 5355 pst->all_tasks_completed(); // declare that i am done 5356 } 5357 5358 // . see if we can share work_queues with ParNew? XXX 5359 void 5360 CMSParRemarkTask::do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, 5361 int* seed) { 5362 OopTaskQueue* work_q = work_queue(i); 5363 NOT_PRODUCT(int num_steals = 0;) 5364 oop obj_to_scan; 5365 CMSBitMap* bm = &(_collector->_markBitMap); 5366 5367 while (true) { 5368 // Completely finish any left over work from (an) earlier round(s) 5369 cl->trim_queue(0); 5370 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 5371 (size_t)ParGCDesiredObjsFromOverflowList); 5372 // Now check if there's any work in the overflow list 5373 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads, 5374 // only affects the number of attempts made to get work from the 5375 // overflow list and does not affect the number of workers. Just 5376 // pass ParallelGCThreads so this behavior is unchanged. 5377 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 5378 work_q, 5379 ParallelGCThreads)) { 5380 // found something in global overflow list; 5381 // not yet ready to go stealing work from others. 5382 // We'd like to assert(work_q->size() != 0, ...) 5383 // because we just took work from the overflow list, 5384 // but of course we can't since all of that could have 5385 // been already stolen from us. 5386 // "He giveth and He taketh away." 5387 continue; 5388 } 5389 // Verify that we have no work before we resort to stealing 5390 assert(work_q->size() == 0, "Have work, shouldn't steal"); 5391 // Try to steal from other queues that have work 5392 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 5393 NOT_PRODUCT(num_steals++;) 5394 assert(obj_to_scan->is_oop(), "Oops, not an oop!"); 5395 assert(bm->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 5396 // Do scanning work 5397 obj_to_scan->oop_iterate(cl); 5398 // Loop around, finish this work, and try to steal some more 5399 } else if (terminator()->offer_termination()) { 5400 break; // nirvana from the infinite cycle 5401 } 5402 } 5403 NOT_PRODUCT( 5404 if (PrintCMSStatistics != 0) { 5405 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals); 5406 } 5407 ) 5408 assert(work_q->size() == 0 && _collector->overflow_list_is_empty(), 5409 "Else our work is not yet done"); 5410 } 5411 5412 // Return a thread-local PLAB recording array, as appropriate. 5413 void* CMSCollector::get_data_recorder(int thr_num) { 5414 if (_survivor_plab_array != NULL && 5415 (CMSPLABRecordAlways || 5416 (_collectorState > Marking && _collectorState < FinalMarking))) { 5417 assert(thr_num < (int)ParallelGCThreads, "thr_num is out of bounds"); 5418 ChunkArray* ca = &_survivor_plab_array[thr_num]; 5419 ca->reset(); // clear it so that fresh data is recorded 5420 return (void*) ca; 5421 } else { 5422 return NULL; 5423 } 5424 } 5425 5426 // Reset all the thread-local PLAB recording arrays 5427 void CMSCollector::reset_survivor_plab_arrays() { 5428 for (uint i = 0; i < ParallelGCThreads; i++) { 5429 _survivor_plab_array[i].reset(); 5430 } 5431 } 5432 5433 // Merge the per-thread plab arrays into the global survivor chunk 5434 // array which will provide the partitioning of the survivor space 5435 // for CMS rescan. 5436 void CMSCollector::merge_survivor_plab_arrays(ContiguousSpace* surv, 5437 int no_of_gc_threads) { 5438 assert(_survivor_plab_array != NULL, "Error"); 5439 assert(_survivor_chunk_array != NULL, "Error"); 5440 assert(_collectorState == FinalMarking, "Error"); 5441 for (int j = 0; j < no_of_gc_threads; j++) { 5442 _cursor[j] = 0; 5443 } 5444 HeapWord* top = surv->top(); 5445 size_t i; 5446 for (i = 0; i < _survivor_chunk_capacity; i++) { // all sca entries 5447 HeapWord* min_val = top; // Higher than any PLAB address 5448 uint min_tid = 0; // position of min_val this round 5449 for (int j = 0; j < no_of_gc_threads; j++) { 5450 ChunkArray* cur_sca = &_survivor_plab_array[j]; 5451 if (_cursor[j] == cur_sca->end()) { 5452 continue; 5453 } 5454 assert(_cursor[j] < cur_sca->end(), "ctl pt invariant"); 5455 HeapWord* cur_val = cur_sca->nth(_cursor[j]); 5456 assert(surv->used_region().contains(cur_val), "Out of bounds value"); 5457 if (cur_val < min_val) { 5458 min_tid = j; 5459 min_val = cur_val; 5460 } else { 5461 assert(cur_val < top, "All recorded addresses should be less"); 5462 } 5463 } 5464 // At this point min_val and min_tid are respectively 5465 // the least address in _survivor_plab_array[j]->nth(_cursor[j]) 5466 // and the thread (j) that witnesses that address. 5467 // We record this address in the _survivor_chunk_array[i] 5468 // and increment _cursor[min_tid] prior to the next round i. 5469 if (min_val == top) { 5470 break; 5471 } 5472 _survivor_chunk_array[i] = min_val; 5473 _cursor[min_tid]++; 5474 } 5475 // We are all done; record the size of the _survivor_chunk_array 5476 _survivor_chunk_index = i; // exclusive: [0, i) 5477 if (PrintCMSStatistics > 0) { 5478 gclog_or_tty->print(" (Survivor:" SIZE_FORMAT "chunks) ", i); 5479 } 5480 // Verify that we used up all the recorded entries 5481 #ifdef ASSERT 5482 size_t total = 0; 5483 for (int j = 0; j < no_of_gc_threads; j++) { 5484 assert(_cursor[j] == _survivor_plab_array[j].end(), "Ctl pt invariant"); 5485 total += _cursor[j]; 5486 } 5487 assert(total == _survivor_chunk_index, "Ctl Pt Invariant"); 5488 // Check that the merged array is in sorted order 5489 if (total > 0) { 5490 for (size_t i = 0; i < total - 1; i++) { 5491 if (PrintCMSStatistics > 0) { 5492 gclog_or_tty->print(" (chunk" SIZE_FORMAT ":" INTPTR_FORMAT ") ", 5493 i, _survivor_chunk_array[i]); 5494 } 5495 assert(_survivor_chunk_array[i] < _survivor_chunk_array[i+1], 5496 "Not sorted"); 5497 } 5498 } 5499 #endif // ASSERT 5500 } 5501 5502 // Set up the space's par_seq_tasks structure for work claiming 5503 // for parallel rescan of young gen. 5504 // See ParRescanTask where this is currently used. 5505 void 5506 CMSCollector:: 5507 initialize_sequential_subtasks_for_young_gen_rescan(int n_threads) { 5508 assert(n_threads > 0, "Unexpected n_threads argument"); 5509 DefNewGeneration* dng = (DefNewGeneration*)_young_gen; 5510 5511 // Eden space 5512 { 5513 SequentialSubTasksDone* pst = dng->eden()->par_seq_tasks(); 5514 assert(!pst->valid(), "Clobbering existing data?"); 5515 // Each valid entry in [0, _eden_chunk_index) represents a task. 5516 size_t n_tasks = _eden_chunk_index + 1; 5517 assert(n_tasks == 1 || _eden_chunk_array != NULL, "Error"); 5518 // Sets the condition for completion of the subtask (how many threads 5519 // need to finish in order to be done). 5520 pst->set_n_threads(n_threads); 5521 pst->set_n_tasks((int)n_tasks); 5522 } 5523 5524 // Merge the survivor plab arrays into _survivor_chunk_array 5525 if (_survivor_plab_array != NULL) { 5526 merge_survivor_plab_arrays(dng->from(), n_threads); 5527 } else { 5528 assert(_survivor_chunk_index == 0, "Error"); 5529 } 5530 5531 // To space 5532 { 5533 SequentialSubTasksDone* pst = dng->to()->par_seq_tasks(); 5534 assert(!pst->valid(), "Clobbering existing data?"); 5535 // Sets the condition for completion of the subtask (how many threads 5536 // need to finish in order to be done). 5537 pst->set_n_threads(n_threads); 5538 pst->set_n_tasks(1); 5539 assert(pst->valid(), "Error"); 5540 } 5541 5542 // From space 5543 { 5544 SequentialSubTasksDone* pst = dng->from()->par_seq_tasks(); 5545 assert(!pst->valid(), "Clobbering existing data?"); 5546 size_t n_tasks = _survivor_chunk_index + 1; 5547 assert(n_tasks == 1 || _survivor_chunk_array != NULL, "Error"); 5548 // Sets the condition for completion of the subtask (how many threads 5549 // need to finish in order to be done). 5550 pst->set_n_threads(n_threads); 5551 pst->set_n_tasks((int)n_tasks); 5552 assert(pst->valid(), "Error"); 5553 } 5554 } 5555 5556 // Parallel version of remark 5557 void CMSCollector::do_remark_parallel() { 5558 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5559 FlexibleWorkGang* workers = gch->workers(); 5560 assert(workers != NULL, "Need parallel worker threads."); 5561 // Choose to use the number of GC workers most recently set 5562 // into "active_workers". If active_workers is not set, set it 5563 // to ParallelGCThreads. 5564 int n_workers = workers->active_workers(); 5565 if (n_workers == 0) { 5566 assert(n_workers > 0, "Should have been set during scavenge"); 5567 n_workers = ParallelGCThreads; 5568 workers->set_active_workers(n_workers); 5569 } 5570 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 5571 CompactibleFreeListSpace* perm_space = _permGen->cmsSpace(); 5572 5573 CMSParRemarkTask tsk(this, 5574 cms_space, perm_space, 5575 n_workers, workers, task_queues()); 5576 5577 // Set up for parallel process_strong_roots work. 5578 gch->set_par_threads(n_workers); 5579 // We won't be iterating over the cards in the card table updating 5580 // the younger_gen cards, so we shouldn't call the following else 5581 // the verification code as well as subsequent younger_refs_iterate 5582 // code would get confused. XXX 5583 // gch->rem_set()->prepare_for_younger_refs_iterate(true); // parallel 5584 5585 // The young gen rescan work will not be done as part of 5586 // process_strong_roots (which currently doesn't knw how to 5587 // parallelize such a scan), but rather will be broken up into 5588 // a set of parallel tasks (via the sampling that the [abortable] 5589 // preclean phase did of EdenSpace, plus the [two] tasks of 5590 // scanning the [two] survivor spaces. Further fine-grain 5591 // parallelization of the scanning of the survivor spaces 5592 // themselves, and of precleaning of the younger gen itself 5593 // is deferred to the future. 5594 initialize_sequential_subtasks_for_young_gen_rescan(n_workers); 5595 5596 // The dirty card rescan work is broken up into a "sequence" 5597 // of parallel tasks (per constituent space) that are dynamically 5598 // claimed by the parallel threads. 5599 cms_space->initialize_sequential_subtasks_for_rescan(n_workers); 5600 perm_space->initialize_sequential_subtasks_for_rescan(n_workers); 5601 5602 // It turns out that even when we're using 1 thread, doing the work in a 5603 // separate thread causes wide variance in run times. We can't help this 5604 // in the multi-threaded case, but we special-case n=1 here to get 5605 // repeatable measurements of the 1-thread overhead of the parallel code. 5606 if (n_workers > 1) { 5607 // Make refs discovery MT-safe, if it isn't already: it may not 5608 // necessarily be so, since it's possible that we are doing 5609 // ST marking. 5610 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), true); 5611 GenCollectedHeap::StrongRootsScope srs(gch); 5612 workers->run_task(&tsk); 5613 } else { 5614 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false); 5615 GenCollectedHeap::StrongRootsScope srs(gch); 5616 tsk.work(0); 5617 } 5618 gch->set_par_threads(0); // 0 ==> non-parallel. 5619 // restore, single-threaded for now, any preserved marks 5620 // as a result of work_q overflow 5621 restore_preserved_marks_if_any(); 5622 } 5623 5624 // Non-parallel version of remark 5625 void CMSCollector::do_remark_non_parallel() { 5626 ResourceMark rm; 5627 HandleMark hm; 5628 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5629 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false); 5630 5631 MarkRefsIntoAndScanClosure 5632 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable, 5633 &_markStack, &_revisitStack, this, 5634 false /* should_yield */, false /* not precleaning */); 5635 MarkFromDirtyCardsClosure 5636 markFromDirtyCardsClosure(this, _span, 5637 NULL, // space is set further below 5638 &_markBitMap, &_markStack, &_revisitStack, 5639 &mrias_cl); 5640 { 5641 GCTraceTime t("grey object rescan", PrintGCDetails, false, _gc_timer_cm); 5642 // Iterate over the dirty cards, setting the corresponding bits in the 5643 // mod union table. 5644 { 5645 ModUnionClosure modUnionClosure(&_modUnionTable); 5646 _ct->ct_bs()->dirty_card_iterate( 5647 _cmsGen->used_region(), 5648 &modUnionClosure); 5649 _ct->ct_bs()->dirty_card_iterate( 5650 _permGen->used_region(), 5651 &modUnionClosure); 5652 } 5653 // Having transferred these marks into the modUnionTable, we just need 5654 // to rescan the marked objects on the dirty cards in the modUnionTable. 5655 // The initial marking may have been done during an asynchronous 5656 // collection so there may be dirty bits in the mod-union table. 5657 const int alignment = 5658 CardTableModRefBS::card_size * BitsPerWord; 5659 { 5660 // ... First handle dirty cards in CMS gen 5661 markFromDirtyCardsClosure.set_space(_cmsGen->cmsSpace()); 5662 MemRegion ur = _cmsGen->used_region(); 5663 HeapWord* lb = ur.start(); 5664 HeapWord* ub = (HeapWord*)round_to((intptr_t)ur.end(), alignment); 5665 MemRegion cms_span(lb, ub); 5666 _modUnionTable.dirty_range_iterate_clear(cms_span, 5667 &markFromDirtyCardsClosure); 5668 verify_work_stacks_empty(); 5669 if (PrintCMSStatistics != 0) { 5670 gclog_or_tty->print(" (re-scanned "SIZE_FORMAT" dirty cards in cms gen) ", 5671 markFromDirtyCardsClosure.num_dirty_cards()); 5672 } 5673 } 5674 { 5675 // .. and then repeat for dirty cards in perm gen 5676 markFromDirtyCardsClosure.set_space(_permGen->cmsSpace()); 5677 MemRegion ur = _permGen->used_region(); 5678 HeapWord* lb = ur.start(); 5679 HeapWord* ub = (HeapWord*)round_to((intptr_t)ur.end(), alignment); 5680 MemRegion perm_span(lb, ub); 5681 _modUnionTable.dirty_range_iterate_clear(perm_span, 5682 &markFromDirtyCardsClosure); 5683 verify_work_stacks_empty(); 5684 if (PrintCMSStatistics != 0) { 5685 gclog_or_tty->print(" (re-scanned "SIZE_FORMAT" dirty cards in perm gen) ", 5686 markFromDirtyCardsClosure.num_dirty_cards()); 5687 } 5688 } 5689 } 5690 if (VerifyDuringGC && 5691 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 5692 HandleMark hm; // Discard invalid handles created during verification 5693 Universe::verify(); 5694 } 5695 { 5696 GCTraceTime t("root rescan", PrintGCDetails, false, _gc_timer_cm); 5697 5698 verify_work_stacks_empty(); 5699 5700 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 5701 GenCollectedHeap::StrongRootsScope srs(gch); 5702 gch->gen_process_strong_roots(_cmsGen->level(), 5703 true, // younger gens as roots 5704 false, // use the local StrongRootsScope 5705 true, // collecting perm gen 5706 SharedHeap::ScanningOption(roots_scanning_options()), 5707 &mrias_cl, 5708 true, // walk code active on stacks 5709 NULL); 5710 assert(should_unload_classes() 5711 || (roots_scanning_options() & SharedHeap::SO_CodeCache), 5712 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 5713 } 5714 verify_work_stacks_empty(); 5715 // Restore evacuated mark words, if any, used for overflow list links 5716 if (!CMSOverflowEarlyRestoration) { 5717 restore_preserved_marks_if_any(); 5718 } 5719 verify_overflow_empty(); 5720 } 5721 5722 //////////////////////////////////////////////////////// 5723 // Parallel Reference Processing Task Proxy Class 5724 //////////////////////////////////////////////////////// 5725 class CMSRefProcTaskProxy: public AbstractGangTaskWOopQueues { 5726 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask; 5727 CMSCollector* _collector; 5728 CMSBitMap* _mark_bit_map; 5729 const MemRegion _span; 5730 ProcessTask& _task; 5731 5732 public: 5733 CMSRefProcTaskProxy(ProcessTask& task, 5734 CMSCollector* collector, 5735 const MemRegion& span, 5736 CMSBitMap* mark_bit_map, 5737 AbstractWorkGang* workers, 5738 OopTaskQueueSet* task_queues): 5739 // XXX Should superclass AGTWOQ also know about AWG since it knows 5740 // about the task_queues used by the AWG? Then it could initialize 5741 // the terminator() object. See 6984287. The set_for_termination() 5742 // below is a temporary band-aid for the regression in 6984287. 5743 AbstractGangTaskWOopQueues("Process referents by policy in parallel", 5744 task_queues), 5745 _task(task), 5746 _collector(collector), _span(span), _mark_bit_map(mark_bit_map) 5747 { 5748 assert(_collector->_span.equals(_span) && !_span.is_empty(), 5749 "Inconsistency in _span"); 5750 set_for_termination(workers->active_workers()); 5751 } 5752 5753 OopTaskQueueSet* task_queues() { return queues(); } 5754 5755 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 5756 5757 void do_work_steal(int i, 5758 CMSParDrainMarkingStackClosure* drain, 5759 CMSParKeepAliveClosure* keep_alive, 5760 int* seed); 5761 5762 virtual void work(uint worker_id); 5763 }; 5764 5765 void CMSRefProcTaskProxy::work(uint worker_id) { 5766 assert(_collector->_span.equals(_span), "Inconsistency in _span"); 5767 CMSParKeepAliveClosure par_keep_alive(_collector, _span, 5768 _mark_bit_map, 5769 &_collector->_revisitStack, 5770 work_queue(worker_id)); 5771 CMSParDrainMarkingStackClosure par_drain_stack(_collector, _span, 5772 _mark_bit_map, 5773 &_collector->_revisitStack, 5774 work_queue(worker_id)); 5775 CMSIsAliveClosure is_alive_closure(_span, _mark_bit_map); 5776 _task.work(worker_id, is_alive_closure, par_keep_alive, par_drain_stack); 5777 if (_task.marks_oops_alive()) { 5778 do_work_steal(worker_id, &par_drain_stack, &par_keep_alive, 5779 _collector->hash_seed(worker_id)); 5780 } 5781 assert(work_queue(worker_id)->size() == 0, "work_queue should be empty"); 5782 assert(_collector->_overflow_list == NULL, "non-empty _overflow_list"); 5783 } 5784 5785 class CMSRefEnqueueTaskProxy: public AbstractGangTask { 5786 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask; 5787 EnqueueTask& _task; 5788 5789 public: 5790 CMSRefEnqueueTaskProxy(EnqueueTask& task) 5791 : AbstractGangTask("Enqueue reference objects in parallel"), 5792 _task(task) 5793 { } 5794 5795 virtual void work(uint worker_id) 5796 { 5797 _task.work(worker_id); 5798 } 5799 }; 5800 5801 CMSParKeepAliveClosure::CMSParKeepAliveClosure(CMSCollector* collector, 5802 MemRegion span, CMSBitMap* bit_map, CMSMarkStack* revisit_stack, 5803 OopTaskQueue* work_queue): 5804 Par_KlassRememberingOopClosure(collector, NULL, revisit_stack), 5805 _span(span), 5806 _bit_map(bit_map), 5807 _work_queue(work_queue), 5808 _mark_and_push(collector, span, bit_map, revisit_stack, work_queue), 5809 _low_water_mark(MIN2((uint)(work_queue->max_elems()/4), 5810 (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads))) 5811 { } 5812 5813 // . see if we can share work_queues with ParNew? XXX 5814 void CMSRefProcTaskProxy::do_work_steal(int i, 5815 CMSParDrainMarkingStackClosure* drain, 5816 CMSParKeepAliveClosure* keep_alive, 5817 int* seed) { 5818 OopTaskQueue* work_q = work_queue(i); 5819 NOT_PRODUCT(int num_steals = 0;) 5820 oop obj_to_scan; 5821 5822 while (true) { 5823 // Completely finish any left over work from (an) earlier round(s) 5824 drain->trim_queue(0); 5825 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 5826 (size_t)ParGCDesiredObjsFromOverflowList); 5827 // Now check if there's any work in the overflow list 5828 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads, 5829 // only affects the number of attempts made to get work from the 5830 // overflow list and does not affect the number of workers. Just 5831 // pass ParallelGCThreads so this behavior is unchanged. 5832 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 5833 work_q, 5834 ParallelGCThreads)) { 5835 // Found something in global overflow list; 5836 // not yet ready to go stealing work from others. 5837 // We'd like to assert(work_q->size() != 0, ...) 5838 // because we just took work from the overflow list, 5839 // but of course we can't, since all of that might have 5840 // been already stolen from us. 5841 continue; 5842 } 5843 // Verify that we have no work before we resort to stealing 5844 assert(work_q->size() == 0, "Have work, shouldn't steal"); 5845 // Try to steal from other queues that have work 5846 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 5847 NOT_PRODUCT(num_steals++;) 5848 assert(obj_to_scan->is_oop(), "Oops, not an oop!"); 5849 assert(_mark_bit_map->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 5850 // Do scanning work 5851 obj_to_scan->oop_iterate(keep_alive); 5852 // Loop around, finish this work, and try to steal some more 5853 } else if (terminator()->offer_termination()) { 5854 break; // nirvana from the infinite cycle 5855 } 5856 } 5857 NOT_PRODUCT( 5858 if (PrintCMSStatistics != 0) { 5859 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals); 5860 } 5861 ) 5862 } 5863 5864 void CMSRefProcTaskExecutor::execute(ProcessTask& task) 5865 { 5866 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5867 FlexibleWorkGang* workers = gch->workers(); 5868 assert(workers != NULL, "Need parallel worker threads."); 5869 CMSRefProcTaskProxy rp_task(task, &_collector, 5870 _collector.ref_processor()->span(), 5871 _collector.markBitMap(), 5872 workers, _collector.task_queues()); 5873 workers->run_task(&rp_task); 5874 } 5875 5876 void CMSRefProcTaskExecutor::execute(EnqueueTask& task) 5877 { 5878 5879 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5880 FlexibleWorkGang* workers = gch->workers(); 5881 assert(workers != NULL, "Need parallel worker threads."); 5882 CMSRefEnqueueTaskProxy enq_task(task); 5883 workers->run_task(&enq_task); 5884 } 5885 5886 void CMSCollector::refProcessingWork(bool asynch, bool clear_all_soft_refs) { 5887 5888 ResourceMark rm; 5889 HandleMark hm; 5890 5891 ReferenceProcessor* rp = ref_processor(); 5892 assert(rp->span().equals(_span), "Spans should be equal"); 5893 assert(!rp->enqueuing_is_done(), "Enqueuing should not be complete"); 5894 // Process weak references. 5895 rp->setup_policy(clear_all_soft_refs); 5896 verify_work_stacks_empty(); 5897 5898 CMSKeepAliveClosure cmsKeepAliveClosure(this, _span, &_markBitMap, 5899 &_markStack, &_revisitStack, 5900 false /* !preclean */); 5901 CMSDrainMarkingStackClosure cmsDrainMarkingStackClosure(this, 5902 _span, &_markBitMap, &_markStack, 5903 &cmsKeepAliveClosure, false /* !preclean */); 5904 { 5905 GCTraceTime t("weak refs processing", PrintGCDetails, false, _gc_timer_cm); 5906 5907 ReferenceProcessorStats stats; 5908 if (rp->processing_is_mt()) { 5909 // Set the degree of MT here. If the discovery is done MT, there 5910 // may have been a different number of threads doing the discovery 5911 // and a different number of discovered lists may have Ref objects. 5912 // That is OK as long as the Reference lists are balanced (see 5913 // balance_all_queues() and balance_queues()). 5914 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5915 int active_workers = ParallelGCThreads; 5916 FlexibleWorkGang* workers = gch->workers(); 5917 if (workers != NULL) { 5918 active_workers = workers->active_workers(); 5919 // The expectation is that active_workers will have already 5920 // been set to a reasonable value. If it has not been set, 5921 // investigate. 5922 assert(active_workers > 0, "Should have been set during scavenge"); 5923 } 5924 rp->set_active_mt_degree(active_workers); 5925 CMSRefProcTaskExecutor task_executor(*this); 5926 stats = rp->process_discovered_references(&_is_alive_closure, 5927 &cmsKeepAliveClosure, 5928 &cmsDrainMarkingStackClosure, 5929 &task_executor, 5930 _gc_timer_cm); 5931 } else { 5932 stats = rp->process_discovered_references(&_is_alive_closure, 5933 &cmsKeepAliveClosure, 5934 &cmsDrainMarkingStackClosure, 5935 NULL, 5936 _gc_timer_cm); 5937 } 5938 _gc_tracer_cm->report_gc_reference_stats(stats); 5939 5940 verify_work_stacks_empty(); 5941 } 5942 5943 if (should_unload_classes()) { 5944 { 5945 GCTraceTime t("class unloading", PrintGCDetails, false, _gc_timer_cm); 5946 5947 // Follow SystemDictionary roots and unload classes 5948 bool purged_class = SystemDictionary::do_unloading(&_is_alive_closure); 5949 5950 // Follow CodeCache roots and unload any methods marked for unloading 5951 CodeCache::do_unloading(&_is_alive_closure, 5952 &cmsKeepAliveClosure, 5953 purged_class); 5954 5955 cmsDrainMarkingStackClosure.do_void(); 5956 verify_work_stacks_empty(); 5957 5958 // Update subklass/sibling/implementor links in KlassKlass descendants 5959 assert(!_revisitStack.isEmpty(), "revisit stack should not be empty"); 5960 oop k; 5961 while ((k = _revisitStack.pop()) != NULL) { 5962 ((Klass*)(oopDesc*)k)->follow_weak_klass_links( 5963 &_is_alive_closure, 5964 &cmsKeepAliveClosure); 5965 } 5966 assert(!ClassUnloading || 5967 (_markStack.isEmpty() && overflow_list_is_empty()), 5968 "Should not have found new reachable objects"); 5969 assert(_revisitStack.isEmpty(), "revisit stack should have been drained"); 5970 cmsDrainMarkingStackClosure.do_void(); 5971 verify_work_stacks_empty(); 5972 } 5973 5974 { 5975 GCTraceTime t("scrub symbol table", PrintGCDetails, false, _gc_timer_cm); 5976 // Clean up unreferenced symbols in symbol table. 5977 SymbolTable::unlink(); 5978 } 5979 } 5980 5981 if (should_unload_classes() || !JavaObjectsInPerm) { 5982 GCTraceTime t("scrub string table", PrintGCDetails, false, _gc_timer_cm); 5983 // Now clean up stale oops in StringTable 5984 StringTable::unlink(&_is_alive_closure); 5985 } 5986 5987 verify_work_stacks_empty(); 5988 // Restore any preserved marks as a result of mark stack or 5989 // work queue overflow 5990 restore_preserved_marks_if_any(); // done single-threaded for now 5991 5992 rp->set_enqueuing_is_done(true); 5993 if (rp->processing_is_mt()) { 5994 rp->balance_all_queues(); 5995 CMSRefProcTaskExecutor task_executor(*this); 5996 rp->enqueue_discovered_references(&task_executor); 5997 } else { 5998 rp->enqueue_discovered_references(NULL); 5999 } 6000 rp->verify_no_references_recorded(); 6001 assert(!rp->discovery_enabled(), "should have been disabled"); 6002 } 6003 6004 #ifndef PRODUCT 6005 void CMSCollector::check_correct_thread_executing() { 6006 Thread* t = Thread::current(); 6007 // Only the VM thread or the CMS thread should be here. 6008 assert(t->is_ConcurrentGC_thread() || t->is_VM_thread(), 6009 "Unexpected thread type"); 6010 // If this is the vm thread, the foreground process 6011 // should not be waiting. Note that _foregroundGCIsActive is 6012 // true while the foreground collector is waiting. 6013 if (_foregroundGCShouldWait) { 6014 // We cannot be the VM thread 6015 assert(t->is_ConcurrentGC_thread(), 6016 "Should be CMS thread"); 6017 } else { 6018 // We can be the CMS thread only if we are in a stop-world 6019 // phase of CMS collection. 6020 if (t->is_ConcurrentGC_thread()) { 6021 assert(_collectorState == InitialMarking || 6022 _collectorState == FinalMarking, 6023 "Should be a stop-world phase"); 6024 // The CMS thread should be holding the CMS_token. 6025 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6026 "Potential interference with concurrently " 6027 "executing VM thread"); 6028 } 6029 } 6030 } 6031 #endif 6032 6033 void CMSCollector::sweep(bool asynch) { 6034 assert(_collectorState == Sweeping, "just checking"); 6035 check_correct_thread_executing(); 6036 verify_work_stacks_empty(); 6037 verify_overflow_empty(); 6038 increment_sweep_count(); 6039 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 6040 6041 _inter_sweep_timer.stop(); 6042 _inter_sweep_estimate.sample(_inter_sweep_timer.seconds()); 6043 size_policy()->avg_cms_free_at_sweep()->sample(_cmsGen->free()); 6044 6045 // PermGen verification support: If perm gen sweeping is disabled in 6046 // this cycle, we preserve the perm gen object "deadness" information 6047 // in the perm_gen_verify_bit_map. In order to do that we traverse 6048 // all blocks in perm gen and mark all dead objects. 6049 if (verifying() && !should_unload_classes()) { 6050 assert(perm_gen_verify_bit_map()->sizeInBits() != 0, 6051 "Should have already been allocated"); 6052 MarkDeadObjectsClosure mdo(this, _permGen->cmsSpace(), 6053 markBitMap(), perm_gen_verify_bit_map()); 6054 if (asynch) { 6055 CMSTokenSyncWithLocks ts(true, _permGen->freelistLock(), 6056 bitMapLock()); 6057 _permGen->cmsSpace()->blk_iterate(&mdo); 6058 } else { 6059 // In the case of synchronous sweep, we already have 6060 // the requisite locks/tokens. 6061 _permGen->cmsSpace()->blk_iterate(&mdo); 6062 } 6063 } 6064 6065 assert(!_intra_sweep_timer.is_active(), "Should not be active"); 6066 _intra_sweep_timer.reset(); 6067 _intra_sweep_timer.start(); 6068 if (asynch) { 6069 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 6070 CMSPhaseAccounting pa(this, "sweep", !PrintGCDetails); 6071 // First sweep the old gen then the perm gen 6072 { 6073 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(), 6074 bitMapLock()); 6075 sweepWork(_cmsGen, asynch); 6076 } 6077 6078 // Now repeat for perm gen 6079 if (should_unload_classes()) { 6080 CMSTokenSyncWithLocks ts(true, _permGen->freelistLock(), 6081 bitMapLock()); 6082 sweepWork(_permGen, asynch); 6083 } 6084 6085 // Update Universe::_heap_*_at_gc figures. 6086 // We need all the free list locks to make the abstract state 6087 // transition from Sweeping to Resetting. See detailed note 6088 // further below. 6089 { 6090 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(), 6091 _permGen->freelistLock()); 6092 // Update heap occupancy information which is used as 6093 // input to soft ref clearing policy at the next gc. 6094 Universe::update_heap_info_at_gc(); 6095 _collectorState = Resizing; 6096 } 6097 } else { 6098 // already have needed locks 6099 sweepWork(_cmsGen, asynch); 6100 6101 if (should_unload_classes()) { 6102 sweepWork(_permGen, asynch); 6103 } 6104 // Update heap occupancy information which is used as 6105 // input to soft ref clearing policy at the next gc. 6106 Universe::update_heap_info_at_gc(); 6107 _collectorState = Resizing; 6108 } 6109 verify_work_stacks_empty(); 6110 verify_overflow_empty(); 6111 6112 _intra_sweep_timer.stop(); 6113 _intra_sweep_estimate.sample(_intra_sweep_timer.seconds()); 6114 6115 _inter_sweep_timer.reset(); 6116 _inter_sweep_timer.start(); 6117 6118 // We need to use a monotonically non-deccreasing time in ms 6119 // or we will see time-warp warnings and os::javaTimeMillis() 6120 // does not guarantee monotonicity. 6121 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC; 6122 update_time_of_last_gc(now); 6123 6124 // NOTE on abstract state transitions: 6125 // Mutators allocate-live and/or mark the mod-union table dirty 6126 // based on the state of the collection. The former is done in 6127 // the interval [Marking, Sweeping] and the latter in the interval 6128 // [Marking, Sweeping). Thus the transitions into the Marking state 6129 // and out of the Sweeping state must be synchronously visible 6130 // globally to the mutators. 6131 // The transition into the Marking state happens with the world 6132 // stopped so the mutators will globally see it. Sweeping is 6133 // done asynchronously by the background collector so the transition 6134 // from the Sweeping state to the Resizing state must be done 6135 // under the freelistLock (as is the check for whether to 6136 // allocate-live and whether to dirty the mod-union table). 6137 assert(_collectorState == Resizing, "Change of collector state to" 6138 " Resizing must be done under the freelistLocks (plural)"); 6139 6140 // Now that sweeping has been completed, we clear 6141 // the incremental_collection_failed flag, 6142 // thus inviting a younger gen collection to promote into 6143 // this generation. If such a promotion may still fail, 6144 // the flag will be set again when a young collection is 6145 // attempted. 6146 GenCollectedHeap* gch = GenCollectedHeap::heap(); 6147 gch->clear_incremental_collection_failed(); // Worth retrying as fresh space may have been freed up 6148 gch->update_full_collections_completed(_collection_count_start); 6149 } 6150 6151 // FIX ME!!! Looks like this belongs in CFLSpace, with 6152 // CMSGen merely delegating to it. 6153 void ConcurrentMarkSweepGeneration::setNearLargestChunk() { 6154 double nearLargestPercent = FLSLargestBlockCoalesceProximity; 6155 HeapWord* minAddr = _cmsSpace->bottom(); 6156 HeapWord* largestAddr = 6157 (HeapWord*) _cmsSpace->dictionary()->find_largest_dict(); 6158 if (largestAddr == NULL) { 6159 // The dictionary appears to be empty. In this case 6160 // try to coalesce at the end of the heap. 6161 largestAddr = _cmsSpace->end(); 6162 } 6163 size_t largestOffset = pointer_delta(largestAddr, minAddr); 6164 size_t nearLargestOffset = 6165 (size_t)((double)largestOffset * nearLargestPercent) - MinChunkSize; 6166 if (PrintFLSStatistics != 0) { 6167 gclog_or_tty->print_cr( 6168 "CMS: Large Block: " PTR_FORMAT ";" 6169 " Proximity: " PTR_FORMAT " -> " PTR_FORMAT, 6170 largestAddr, 6171 _cmsSpace->nearLargestChunk(), minAddr + nearLargestOffset); 6172 } 6173 _cmsSpace->set_nearLargestChunk(minAddr + nearLargestOffset); 6174 } 6175 6176 bool ConcurrentMarkSweepGeneration::isNearLargestChunk(HeapWord* addr) { 6177 return addr >= _cmsSpace->nearLargestChunk(); 6178 } 6179 6180 FreeChunk* ConcurrentMarkSweepGeneration::find_chunk_at_end() { 6181 return _cmsSpace->find_chunk_at_end(); 6182 } 6183 6184 void ConcurrentMarkSweepGeneration::update_gc_stats(int current_level, 6185 bool full) { 6186 // The next lower level has been collected. Gather any statistics 6187 // that are of interest at this point. 6188 if (!full && (current_level + 1) == level()) { 6189 // Gather statistics on the young generation collection. 6190 collector()->stats().record_gc0_end(used()); 6191 } 6192 } 6193 6194 CMSAdaptiveSizePolicy* ConcurrentMarkSweepGeneration::size_policy() { 6195 GenCollectedHeap* gch = GenCollectedHeap::heap(); 6196 assert(gch->kind() == CollectedHeap::GenCollectedHeap, 6197 "Wrong type of heap"); 6198 CMSAdaptiveSizePolicy* sp = (CMSAdaptiveSizePolicy*) 6199 gch->gen_policy()->size_policy(); 6200 assert(sp->is_gc_cms_adaptive_size_policy(), 6201 "Wrong type of size policy"); 6202 return sp; 6203 } 6204 6205 void ConcurrentMarkSweepGeneration::rotate_debug_collection_type() { 6206 if (PrintGCDetails && Verbose) { 6207 gclog_or_tty->print("Rotate from %d ", _debug_collection_type); 6208 } 6209 _debug_collection_type = (CollectionTypes) (_debug_collection_type + 1); 6210 _debug_collection_type = 6211 (CollectionTypes) (_debug_collection_type % Unknown_collection_type); 6212 if (PrintGCDetails && Verbose) { 6213 gclog_or_tty->print_cr("to %d ", _debug_collection_type); 6214 } 6215 } 6216 6217 void CMSCollector::sweepWork(ConcurrentMarkSweepGeneration* gen, 6218 bool asynch) { 6219 // We iterate over the space(s) underlying this generation, 6220 // checking the mark bit map to see if the bits corresponding 6221 // to specific blocks are marked or not. Blocks that are 6222 // marked are live and are not swept up. All remaining blocks 6223 // are swept up, with coalescing on-the-fly as we sweep up 6224 // contiguous free and/or garbage blocks: 6225 // We need to ensure that the sweeper synchronizes with allocators 6226 // and stop-the-world collectors. In particular, the following 6227 // locks are used: 6228 // . CMS token: if this is held, a stop the world collection cannot occur 6229 // . freelistLock: if this is held no allocation can occur from this 6230 // generation by another thread 6231 // . bitMapLock: if this is held, no other thread can access or update 6232 // 6233 6234 // Note that we need to hold the freelistLock if we use 6235 // block iterate below; else the iterator might go awry if 6236 // a mutator (or promotion) causes block contents to change 6237 // (for instance if the allocator divvies up a block). 6238 // If we hold the free list lock, for all practical purposes 6239 // young generation GC's can't occur (they'll usually need to 6240 // promote), so we might as well prevent all young generation 6241 // GC's while we do a sweeping step. For the same reason, we might 6242 // as well take the bit map lock for the entire duration 6243 6244 // check that we hold the requisite locks 6245 assert(have_cms_token(), "Should hold cms token"); 6246 assert( (asynch && ConcurrentMarkSweepThread::cms_thread_has_cms_token()) 6247 || (!asynch && ConcurrentMarkSweepThread::vm_thread_has_cms_token()), 6248 "Should possess CMS token to sweep"); 6249 assert_lock_strong(gen->freelistLock()); 6250 assert_lock_strong(bitMapLock()); 6251 6252 assert(!_inter_sweep_timer.is_active(), "Was switched off in an outer context"); 6253 assert(_intra_sweep_timer.is_active(), "Was switched on in an outer context"); 6254 gen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()), 6255 _inter_sweep_estimate.padded_average(), 6256 _intra_sweep_estimate.padded_average()); 6257 gen->setNearLargestChunk(); 6258 6259 { 6260 SweepClosure sweepClosure(this, gen, &_markBitMap, 6261 CMSYield && asynch); 6262 gen->cmsSpace()->blk_iterate_careful(&sweepClosure); 6263 // We need to free-up/coalesce garbage/blocks from a 6264 // co-terminal free run. This is done in the SweepClosure 6265 // destructor; so, do not remove this scope, else the 6266 // end-of-sweep-census below will be off by a little bit. 6267 } 6268 gen->cmsSpace()->sweep_completed(); 6269 gen->cmsSpace()->endSweepFLCensus(sweep_count()); 6270 if (should_unload_classes()) { // unloaded classes this cycle, 6271 _concurrent_cycles_since_last_unload = 0; // ... reset count 6272 } else { // did not unload classes, 6273 _concurrent_cycles_since_last_unload++; // ... increment count 6274 } 6275 } 6276 6277 // Reset CMS data structures (for now just the marking bit map) 6278 // preparatory for the next cycle. 6279 void CMSCollector::reset(bool asynch) { 6280 GenCollectedHeap* gch = GenCollectedHeap::heap(); 6281 CMSAdaptiveSizePolicy* sp = size_policy(); 6282 AdaptiveSizePolicyOutput(sp, gch->total_collections()); 6283 if (asynch) { 6284 CMSTokenSyncWithLocks ts(true, bitMapLock()); 6285 6286 // If the state is not "Resetting", the foreground thread 6287 // has done a collection and the resetting. 6288 if (_collectorState != Resetting) { 6289 assert(_collectorState == Idling, "The state should only change" 6290 " because the foreground collector has finished the collection"); 6291 return; 6292 } 6293 6294 // Clear the mark bitmap (no grey objects to start with) 6295 // for the next cycle. 6296 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 6297 CMSPhaseAccounting cmspa(this, "reset", !PrintGCDetails); 6298 6299 HeapWord* curAddr = _markBitMap.startWord(); 6300 while (curAddr < _markBitMap.endWord()) { 6301 size_t remaining = pointer_delta(_markBitMap.endWord(), curAddr); 6302 MemRegion chunk(curAddr, MIN2(CMSBitMapYieldQuantum, remaining)); 6303 _markBitMap.clear_large_range(chunk); 6304 if (ConcurrentMarkSweepThread::should_yield() && 6305 !foregroundGCIsActive() && 6306 CMSYield) { 6307 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6308 "CMS thread should hold CMS token"); 6309 assert_lock_strong(bitMapLock()); 6310 bitMapLock()->unlock(); 6311 ConcurrentMarkSweepThread::desynchronize(true); 6312 ConcurrentMarkSweepThread::acknowledge_yield_request(); 6313 stopTimer(); 6314 if (PrintCMSStatistics != 0) { 6315 incrementYields(); 6316 } 6317 icms_wait(); 6318 6319 // See the comment in coordinator_yield() 6320 for (unsigned i = 0; i < CMSYieldSleepCount && 6321 ConcurrentMarkSweepThread::should_yield() && 6322 !CMSCollector::foregroundGCIsActive(); ++i) { 6323 os::sleep(Thread::current(), 1, false); 6324 ConcurrentMarkSweepThread::acknowledge_yield_request(); 6325 } 6326 6327 ConcurrentMarkSweepThread::synchronize(true); 6328 bitMapLock()->lock_without_safepoint_check(); 6329 startTimer(); 6330 } 6331 curAddr = chunk.end(); 6332 } 6333 // A successful mostly concurrent collection has been done. 6334 // Because only the full (i.e., concurrent mode failure) collections 6335 // are being measured for gc overhead limits, clean the "near" flag 6336 // and count. 6337 sp->reset_gc_overhead_limit_count(); 6338 _collectorState = Idling; 6339 } else { 6340 // already have the lock 6341 assert(_collectorState == Resetting, "just checking"); 6342 assert_lock_strong(bitMapLock()); 6343 _markBitMap.clear_all(); 6344 _collectorState = Idling; 6345 } 6346 6347 // Stop incremental mode after a cycle completes, so that any future cycles 6348 // are triggered by allocation. 6349 stop_icms(); 6350 6351 NOT_PRODUCT( 6352 if (RotateCMSCollectionTypes) { 6353 _cmsGen->rotate_debug_collection_type(); 6354 } 6355 ) 6356 6357 register_gc_end(); 6358 } 6359 6360 void CMSCollector::do_CMS_operation(CMS_op_type op, GCCause::Cause gc_cause) { 6361 gclog_or_tty->date_stamp(PrintGC && PrintGCDateStamps); 6362 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 6363 GCTraceTime t(GCCauseString("GC", gc_cause), PrintGC, !PrintGCDetails, NULL); 6364 TraceCollectorStats tcs(counters()); 6365 6366 switch (op) { 6367 case CMS_op_checkpointRootsInitial: { 6368 SvcGCMarker sgcm(SvcGCMarker::OTHER); 6369 checkpointRootsInitial(true); // asynch 6370 if (PrintGC) { 6371 _cmsGen->printOccupancy("initial-mark"); 6372 } 6373 break; 6374 } 6375 case CMS_op_checkpointRootsFinal: { 6376 SvcGCMarker sgcm(SvcGCMarker::OTHER); 6377 checkpointRootsFinal(true, // asynch 6378 false, // !clear_all_soft_refs 6379 false); // !init_mark_was_synchronous 6380 if (PrintGC) { 6381 _cmsGen->printOccupancy("remark"); 6382 } 6383 break; 6384 } 6385 default: 6386 fatal("No such CMS_op"); 6387 } 6388 } 6389 6390 #ifndef PRODUCT 6391 size_t const CMSCollector::skip_header_HeapWords() { 6392 return FreeChunk::header_size(); 6393 } 6394 6395 // Try and collect here conditions that should hold when 6396 // CMS thread is exiting. The idea is that the foreground GC 6397 // thread should not be blocked if it wants to terminate 6398 // the CMS thread and yet continue to run the VM for a while 6399 // after that. 6400 void CMSCollector::verify_ok_to_terminate() const { 6401 assert(Thread::current()->is_ConcurrentGC_thread(), 6402 "should be called by CMS thread"); 6403 assert(!_foregroundGCShouldWait, "should be false"); 6404 // We could check here that all the various low-level locks 6405 // are not held by the CMS thread, but that is overkill; see 6406 // also CMSThread::verify_ok_to_terminate() where the CGC_lock 6407 // is checked. 6408 } 6409 #endif 6410 6411 size_t CMSCollector::block_size_using_printezis_bits(HeapWord* addr) const { 6412 assert(_markBitMap.isMarked(addr) && _markBitMap.isMarked(addr + 1), 6413 "missing Printezis mark?"); 6414 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 6415 size_t size = pointer_delta(nextOneAddr + 1, addr); 6416 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 6417 "alignment problem"); 6418 assert(size >= 3, "Necessary for Printezis marks to work"); 6419 return size; 6420 } 6421 6422 // A variant of the above (block_size_using_printezis_bits()) except 6423 // that we return 0 if the P-bits are not yet set. 6424 size_t CMSCollector::block_size_if_printezis_bits(HeapWord* addr) const { 6425 if (_markBitMap.isMarked(addr + 1)) { 6426 assert(_markBitMap.isMarked(addr), "P-bit can be set only for marked objects"); 6427 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 6428 size_t size = pointer_delta(nextOneAddr + 1, addr); 6429 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 6430 "alignment problem"); 6431 assert(size >= 3, "Necessary for Printezis marks to work"); 6432 return size; 6433 } 6434 return 0; 6435 } 6436 6437 HeapWord* CMSCollector::next_card_start_after_block(HeapWord* addr) const { 6438 size_t sz = 0; 6439 oop p = (oop)addr; 6440 if (p->klass_or_null() != NULL && p->is_parsable()) { 6441 sz = CompactibleFreeListSpace::adjustObjectSize(p->size()); 6442 } else { 6443 sz = block_size_using_printezis_bits(addr); 6444 } 6445 assert(sz > 0, "size must be nonzero"); 6446 HeapWord* next_block = addr + sz; 6447 HeapWord* next_card = (HeapWord*)round_to((uintptr_t)next_block, 6448 CardTableModRefBS::card_size); 6449 assert(round_down((uintptr_t)addr, CardTableModRefBS::card_size) < 6450 round_down((uintptr_t)next_card, CardTableModRefBS::card_size), 6451 "must be different cards"); 6452 return next_card; 6453 } 6454 6455 6456 // CMS Bit Map Wrapper ///////////////////////////////////////// 6457 6458 // Construct a CMS bit map infrastructure, but don't create the 6459 // bit vector itself. That is done by a separate call CMSBitMap::allocate() 6460 // further below. 6461 CMSBitMap::CMSBitMap(int shifter, int mutex_rank, const char* mutex_name): 6462 _bm(), 6463 _shifter(shifter), 6464 _lock(mutex_rank >= 0 ? new Mutex(mutex_rank, mutex_name, true) : NULL) 6465 { 6466 _bmStartWord = 0; 6467 _bmWordSize = 0; 6468 } 6469 6470 bool CMSBitMap::allocate(MemRegion mr) { 6471 _bmStartWord = mr.start(); 6472 _bmWordSize = mr.word_size(); 6473 ReservedSpace brs(ReservedSpace::allocation_align_size_up( 6474 (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1)); 6475 if (!brs.is_reserved()) { 6476 warning("CMS bit map allocation failure"); 6477 return false; 6478 } 6479 // For now we'll just commit all of the bit map up fromt. 6480 // Later on we'll try to be more parsimonious with swap. 6481 if (!_virtual_space.initialize(brs, brs.size())) { 6482 warning("CMS bit map backing store failure"); 6483 return false; 6484 } 6485 assert(_virtual_space.committed_size() == brs.size(), 6486 "didn't reserve backing store for all of CMS bit map?"); 6487 _bm.set_map((BitMap::bm_word_t*)_virtual_space.low()); 6488 assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >= 6489 _bmWordSize, "inconsistency in bit map sizing"); 6490 _bm.set_size(_bmWordSize >> _shifter); 6491 6492 // bm.clear(); // can we rely on getting zero'd memory? verify below 6493 assert(isAllClear(), 6494 "Expected zero'd memory from ReservedSpace constructor"); 6495 assert(_bm.size() == heapWordDiffToOffsetDiff(sizeInWords()), 6496 "consistency check"); 6497 return true; 6498 } 6499 6500 void CMSBitMap::dirty_range_iterate_clear(MemRegion mr, MemRegionClosure* cl) { 6501 HeapWord *next_addr, *end_addr, *last_addr; 6502 assert_locked(); 6503 assert(covers(mr), "out-of-range error"); 6504 // XXX assert that start and end are appropriately aligned 6505 for (next_addr = mr.start(), end_addr = mr.end(); 6506 next_addr < end_addr; next_addr = last_addr) { 6507 MemRegion dirty_region = getAndClearMarkedRegion(next_addr, end_addr); 6508 last_addr = dirty_region.end(); 6509 if (!dirty_region.is_empty()) { 6510 cl->do_MemRegion(dirty_region); 6511 } else { 6512 assert(last_addr == end_addr, "program logic"); 6513 return; 6514 } 6515 } 6516 } 6517 6518 #ifndef PRODUCT 6519 void CMSBitMap::assert_locked() const { 6520 CMSLockVerifier::assert_locked(lock()); 6521 } 6522 6523 bool CMSBitMap::covers(MemRegion mr) const { 6524 // assert(_bm.map() == _virtual_space.low(), "map inconsistency"); 6525 assert((size_t)_bm.size() == (_bmWordSize >> _shifter), 6526 "size inconsistency"); 6527 return (mr.start() >= _bmStartWord) && 6528 (mr.end() <= endWord()); 6529 } 6530 6531 bool CMSBitMap::covers(HeapWord* start, size_t size) const { 6532 return (start >= _bmStartWord && (start + size) <= endWord()); 6533 } 6534 6535 void CMSBitMap::verifyNoOneBitsInRange(HeapWord* left, HeapWord* right) { 6536 // verify that there are no 1 bits in the interval [left, right) 6537 FalseBitMapClosure falseBitMapClosure; 6538 iterate(&falseBitMapClosure, left, right); 6539 } 6540 6541 void CMSBitMap::region_invariant(MemRegion mr) 6542 { 6543 assert_locked(); 6544 // mr = mr.intersection(MemRegion(_bmStartWord, _bmWordSize)); 6545 assert(!mr.is_empty(), "unexpected empty region"); 6546 assert(covers(mr), "mr should be covered by bit map"); 6547 // convert address range into offset range 6548 size_t start_ofs = heapWordToOffset(mr.start()); 6549 // Make sure that end() is appropriately aligned 6550 assert(mr.end() == (HeapWord*)round_to((intptr_t)mr.end(), 6551 (1 << (_shifter+LogHeapWordSize))), 6552 "Misaligned mr.end()"); 6553 size_t end_ofs = heapWordToOffset(mr.end()); 6554 assert(end_ofs > start_ofs, "Should mark at least one bit"); 6555 } 6556 6557 #endif 6558 6559 bool CMSMarkStack::allocate(size_t size) { 6560 // allocate a stack of the requisite depth 6561 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 6562 size * sizeof(oop))); 6563 if (!rs.is_reserved()) { 6564 warning("CMSMarkStack allocation failure"); 6565 return false; 6566 } 6567 if (!_virtual_space.initialize(rs, rs.size())) { 6568 warning("CMSMarkStack backing store failure"); 6569 return false; 6570 } 6571 assert(_virtual_space.committed_size() == rs.size(), 6572 "didn't reserve backing store for all of CMS stack?"); 6573 _base = (oop*)(_virtual_space.low()); 6574 _index = 0; 6575 _capacity = size; 6576 NOT_PRODUCT(_max_depth = 0); 6577 return true; 6578 } 6579 6580 // XXX FIX ME !!! In the MT case we come in here holding a 6581 // leaf lock. For printing we need to take a further lock 6582 // which has lower rank. We need to recallibrate the two 6583 // lock-ranks involved in order to be able to rpint the 6584 // messages below. (Or defer the printing to the caller. 6585 // For now we take the expedient path of just disabling the 6586 // messages for the problematic case.) 6587 void CMSMarkStack::expand() { 6588 assert(_capacity <= MarkStackSizeMax, "stack bigger than permitted"); 6589 if (_capacity == MarkStackSizeMax) { 6590 if (_hit_limit++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) { 6591 // We print a warning message only once per CMS cycle. 6592 gclog_or_tty->print_cr(" (benign) Hit CMSMarkStack max size limit"); 6593 } 6594 return; 6595 } 6596 // Double capacity if possible 6597 size_t new_capacity = MIN2(_capacity*2, MarkStackSizeMax); 6598 // Do not give up existing stack until we have managed to 6599 // get the double capacity that we desired. 6600 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 6601 new_capacity * sizeof(oop))); 6602 if (rs.is_reserved()) { 6603 // Release the backing store associated with old stack 6604 _virtual_space.release(); 6605 // Reinitialize virtual space for new stack 6606 if (!_virtual_space.initialize(rs, rs.size())) { 6607 fatal("Not enough swap for expanded marking stack"); 6608 } 6609 _base = (oop*)(_virtual_space.low()); 6610 _index = 0; 6611 _capacity = new_capacity; 6612 } else if (_failed_double++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) { 6613 // Failed to double capacity, continue; 6614 // we print a detail message only once per CMS cycle. 6615 gclog_or_tty->print(" (benign) Failed to expand marking stack from "SIZE_FORMAT"K to " 6616 SIZE_FORMAT"K", 6617 _capacity / K, new_capacity / K); 6618 } 6619 } 6620 6621 6622 // Closures 6623 // XXX: there seems to be a lot of code duplication here; 6624 // should refactor and consolidate common code. 6625 6626 // This closure is used to mark refs into the CMS generation in 6627 // the CMS bit map. Called at the first checkpoint. This closure 6628 // assumes that we do not need to re-mark dirty cards; if the CMS 6629 // generation on which this is used is not an oldest (modulo perm gen) 6630 // generation then this will lose younger_gen cards! 6631 6632 MarkRefsIntoClosure::MarkRefsIntoClosure( 6633 MemRegion span, CMSBitMap* bitMap): 6634 _span(span), 6635 _bitMap(bitMap) 6636 { 6637 assert(_ref_processor == NULL, "deliberately left NULL"); 6638 assert(_bitMap->covers(_span), "_bitMap/_span mismatch"); 6639 } 6640 6641 void MarkRefsIntoClosure::do_oop(oop obj) { 6642 // if p points into _span, then mark corresponding bit in _markBitMap 6643 assert(obj->is_oop(), "expected an oop"); 6644 HeapWord* addr = (HeapWord*)obj; 6645 if (_span.contains(addr)) { 6646 // this should be made more efficient 6647 _bitMap->mark(addr); 6648 } 6649 } 6650 6651 void MarkRefsIntoClosure::do_oop(oop* p) { MarkRefsIntoClosure::do_oop_work(p); } 6652 void MarkRefsIntoClosure::do_oop(narrowOop* p) { MarkRefsIntoClosure::do_oop_work(p); } 6653 6654 // A variant of the above, used for CMS marking verification. 6655 MarkRefsIntoVerifyClosure::MarkRefsIntoVerifyClosure( 6656 MemRegion span, CMSBitMap* verification_bm, CMSBitMap* cms_bm): 6657 _span(span), 6658 _verification_bm(verification_bm), 6659 _cms_bm(cms_bm) 6660 { 6661 assert(_ref_processor == NULL, "deliberately left NULL"); 6662 assert(_verification_bm->covers(_span), "_verification_bm/_span mismatch"); 6663 } 6664 6665 void MarkRefsIntoVerifyClosure::do_oop(oop obj) { 6666 // if p points into _span, then mark corresponding bit in _markBitMap 6667 assert(obj->is_oop(), "expected an oop"); 6668 HeapWord* addr = (HeapWord*)obj; 6669 if (_span.contains(addr)) { 6670 _verification_bm->mark(addr); 6671 if (!_cms_bm->isMarked(addr)) { 6672 oop(addr)->print(); 6673 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)", addr); 6674 fatal("... aborting"); 6675 } 6676 } 6677 } 6678 6679 void MarkRefsIntoVerifyClosure::do_oop(oop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); } 6680 void MarkRefsIntoVerifyClosure::do_oop(narrowOop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); } 6681 6682 ////////////////////////////////////////////////// 6683 // MarkRefsIntoAndScanClosure 6684 ////////////////////////////////////////////////// 6685 6686 MarkRefsIntoAndScanClosure::MarkRefsIntoAndScanClosure(MemRegion span, 6687 ReferenceProcessor* rp, 6688 CMSBitMap* bit_map, 6689 CMSBitMap* mod_union_table, 6690 CMSMarkStack* mark_stack, 6691 CMSMarkStack* revisit_stack, 6692 CMSCollector* collector, 6693 bool should_yield, 6694 bool concurrent_precleaning): 6695 _collector(collector), 6696 _span(span), 6697 _bit_map(bit_map), 6698 _mark_stack(mark_stack), 6699 _pushAndMarkClosure(collector, span, rp, bit_map, mod_union_table, 6700 mark_stack, revisit_stack, concurrent_precleaning), 6701 _yield(should_yield), 6702 _concurrent_precleaning(concurrent_precleaning), 6703 _freelistLock(NULL) 6704 { 6705 _ref_processor = rp; 6706 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 6707 } 6708 6709 // This closure is used to mark refs into the CMS generation at the 6710 // second (final) checkpoint, and to scan and transitively follow 6711 // the unmarked oops. It is also used during the concurrent precleaning 6712 // phase while scanning objects on dirty cards in the CMS generation. 6713 // The marks are made in the marking bit map and the marking stack is 6714 // used for keeping the (newly) grey objects during the scan. 6715 // The parallel version (Par_...) appears further below. 6716 void MarkRefsIntoAndScanClosure::do_oop(oop obj) { 6717 if (obj != NULL) { 6718 assert(obj->is_oop(), "expected an oop"); 6719 HeapWord* addr = (HeapWord*)obj; 6720 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 6721 assert(_collector->overflow_list_is_empty(), 6722 "overflow list should be empty"); 6723 if (_span.contains(addr) && 6724 !_bit_map->isMarked(addr)) { 6725 // mark bit map (object is now grey) 6726 _bit_map->mark(addr); 6727 // push on marking stack (stack should be empty), and drain the 6728 // stack by applying this closure to the oops in the oops popped 6729 // from the stack (i.e. blacken the grey objects) 6730 bool res = _mark_stack->push(obj); 6731 assert(res, "Should have space to push on empty stack"); 6732 do { 6733 oop new_oop = _mark_stack->pop(); 6734 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 6735 assert(new_oop->is_parsable(), "Found unparsable oop"); 6736 assert(_bit_map->isMarked((HeapWord*)new_oop), 6737 "only grey objects on this stack"); 6738 // iterate over the oops in this oop, marking and pushing 6739 // the ones in CMS heap (i.e. in _span). 6740 new_oop->oop_iterate(&_pushAndMarkClosure); 6741 // check if it's time to yield 6742 do_yield_check(); 6743 } while (!_mark_stack->isEmpty() || 6744 (!_concurrent_precleaning && take_from_overflow_list())); 6745 // if marking stack is empty, and we are not doing this 6746 // during precleaning, then check the overflow list 6747 } 6748 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 6749 assert(_collector->overflow_list_is_empty(), 6750 "overflow list was drained above"); 6751 // We could restore evacuated mark words, if any, used for 6752 // overflow list links here because the overflow list is 6753 // provably empty here. That would reduce the maximum 6754 // size requirements for preserved_{oop,mark}_stack. 6755 // But we'll just postpone it until we are all done 6756 // so we can just stream through. 6757 if (!_concurrent_precleaning && CMSOverflowEarlyRestoration) { 6758 _collector->restore_preserved_marks_if_any(); 6759 assert(_collector->no_preserved_marks(), "No preserved marks"); 6760 } 6761 assert(!CMSOverflowEarlyRestoration || _collector->no_preserved_marks(), 6762 "All preserved marks should have been restored above"); 6763 } 6764 } 6765 6766 void MarkRefsIntoAndScanClosure::do_oop(oop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); } 6767 void MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); } 6768 6769 void MarkRefsIntoAndScanClosure::do_yield_work() { 6770 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6771 "CMS thread should hold CMS token"); 6772 assert_lock_strong(_freelistLock); 6773 assert_lock_strong(_bit_map->lock()); 6774 // relinquish the free_list_lock and bitMaplock() 6775 DEBUG_ONLY(RememberKlassesChecker mux(false);) 6776 _bit_map->lock()->unlock(); 6777 _freelistLock->unlock(); 6778 ConcurrentMarkSweepThread::desynchronize(true); 6779 ConcurrentMarkSweepThread::acknowledge_yield_request(); 6780 _collector->stopTimer(); 6781 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 6782 if (PrintCMSStatistics != 0) { 6783 _collector->incrementYields(); 6784 } 6785 _collector->icms_wait(); 6786 6787 // See the comment in coordinator_yield() 6788 for (unsigned i = 0; 6789 i < CMSYieldSleepCount && 6790 ConcurrentMarkSweepThread::should_yield() && 6791 !CMSCollector::foregroundGCIsActive(); 6792 ++i) { 6793 os::sleep(Thread::current(), 1, false); 6794 ConcurrentMarkSweepThread::acknowledge_yield_request(); 6795 } 6796 6797 ConcurrentMarkSweepThread::synchronize(true); 6798 _freelistLock->lock_without_safepoint_check(); 6799 _bit_map->lock()->lock_without_safepoint_check(); 6800 _collector->startTimer(); 6801 } 6802 6803 /////////////////////////////////////////////////////////// 6804 // Par_MarkRefsIntoAndScanClosure: a parallel version of 6805 // MarkRefsIntoAndScanClosure 6806 /////////////////////////////////////////////////////////// 6807 Par_MarkRefsIntoAndScanClosure::Par_MarkRefsIntoAndScanClosure( 6808 CMSCollector* collector, MemRegion span, ReferenceProcessor* rp, 6809 CMSBitMap* bit_map, OopTaskQueue* work_queue, CMSMarkStack* revisit_stack): 6810 _span(span), 6811 _bit_map(bit_map), 6812 _work_queue(work_queue), 6813 _low_water_mark(MIN2((uint)(work_queue->max_elems()/4), 6814 (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads))), 6815 _par_pushAndMarkClosure(collector, span, rp, bit_map, work_queue, 6816 revisit_stack) 6817 { 6818 _ref_processor = rp; 6819 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 6820 } 6821 6822 // This closure is used to mark refs into the CMS generation at the 6823 // second (final) checkpoint, and to scan and transitively follow 6824 // the unmarked oops. The marks are made in the marking bit map and 6825 // the work_queue is used for keeping the (newly) grey objects during 6826 // the scan phase whence they are also available for stealing by parallel 6827 // threads. Since the marking bit map is shared, updates are 6828 // synchronized (via CAS). 6829 void Par_MarkRefsIntoAndScanClosure::do_oop(oop obj) { 6830 if (obj != NULL) { 6831 // Ignore mark word because this could be an already marked oop 6832 // that may be chained at the end of the overflow list. 6833 assert(obj->is_oop(true), "expected an oop"); 6834 HeapWord* addr = (HeapWord*)obj; 6835 if (_span.contains(addr) && 6836 !_bit_map->isMarked(addr)) { 6837 // mark bit map (object will become grey): 6838 // It is possible for several threads to be 6839 // trying to "claim" this object concurrently; 6840 // the unique thread that succeeds in marking the 6841 // object first will do the subsequent push on 6842 // to the work queue (or overflow list). 6843 if (_bit_map->par_mark(addr)) { 6844 // push on work_queue (which may not be empty), and trim the 6845 // queue to an appropriate length by applying this closure to 6846 // the oops in the oops popped from the stack (i.e. blacken the 6847 // grey objects) 6848 bool res = _work_queue->push(obj); 6849 assert(res, "Low water mark should be less than capacity?"); 6850 trim_queue(_low_water_mark); 6851 } // Else, another thread claimed the object 6852 } 6853 } 6854 } 6855 6856 void Par_MarkRefsIntoAndScanClosure::do_oop(oop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); } 6857 void Par_MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); } 6858 6859 // This closure is used to rescan the marked objects on the dirty cards 6860 // in the mod union table and the card table proper. 6861 size_t ScanMarkedObjectsAgainCarefullyClosure::do_object_careful_m( 6862 oop p, MemRegion mr) { 6863 6864 size_t size = 0; 6865 HeapWord* addr = (HeapWord*)p; 6866 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6867 assert(_span.contains(addr), "we are scanning the CMS generation"); 6868 // check if it's time to yield 6869 if (do_yield_check()) { 6870 // We yielded for some foreground stop-world work, 6871 // and we have been asked to abort this ongoing preclean cycle. 6872 return 0; 6873 } 6874 if (_bitMap->isMarked(addr)) { 6875 // it's marked; is it potentially uninitialized? 6876 if (p->klass_or_null() != NULL) { 6877 // If is_conc_safe is false, the object may be undergoing 6878 // change by the VM outside a safepoint. Don't try to 6879 // scan it, but rather leave it for the remark phase. 6880 if (CMSPermGenPrecleaningEnabled && 6881 (!p->is_conc_safe() || !p->is_parsable())) { 6882 // Signal precleaning to redirty the card since 6883 // the klass pointer is already installed. 6884 assert(size == 0, "Initial value"); 6885 } else { 6886 assert(p->is_parsable(), "must be parsable."); 6887 // an initialized object; ignore mark word in verification below 6888 // since we are running concurrent with mutators 6889 assert(p->is_oop(true), "should be an oop"); 6890 if (p->is_objArray()) { 6891 // objArrays are precisely marked; restrict scanning 6892 // to dirty cards only. 6893 size = CompactibleFreeListSpace::adjustObjectSize( 6894 p->oop_iterate(_scanningClosure, mr)); 6895 } else { 6896 // A non-array may have been imprecisely marked; we need 6897 // to scan object in its entirety. 6898 size = CompactibleFreeListSpace::adjustObjectSize( 6899 p->oop_iterate(_scanningClosure)); 6900 } 6901 #ifdef DEBUG 6902 size_t direct_size = 6903 CompactibleFreeListSpace::adjustObjectSize(p->size()); 6904 assert(size == direct_size, "Inconsistency in size"); 6905 assert(size >= 3, "Necessary for Printezis marks to work"); 6906 if (!_bitMap->isMarked(addr+1)) { 6907 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size); 6908 } else { 6909 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size-1); 6910 assert(_bitMap->isMarked(addr+size-1), 6911 "inconsistent Printezis mark"); 6912 } 6913 #endif // DEBUG 6914 } 6915 } else { 6916 // an unitialized object 6917 assert(_bitMap->isMarked(addr+1), "missing Printezis mark?"); 6918 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 6919 size = pointer_delta(nextOneAddr + 1, addr); 6920 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 6921 "alignment problem"); 6922 // Note that pre-cleaning needn't redirty the card. OopDesc::set_klass() 6923 // will dirty the card when the klass pointer is installed in the 6924 // object (signalling the completion of initialization). 6925 } 6926 } else { 6927 // Either a not yet marked object or an uninitialized object 6928 if (p->klass_or_null() == NULL || !p->is_parsable()) { 6929 // An uninitialized object, skip to the next card, since 6930 // we may not be able to read its P-bits yet. 6931 assert(size == 0, "Initial value"); 6932 } else { 6933 // An object not (yet) reached by marking: we merely need to 6934 // compute its size so as to go look at the next block. 6935 assert(p->is_oop(true), "should be an oop"); 6936 size = CompactibleFreeListSpace::adjustObjectSize(p->size()); 6937 } 6938 } 6939 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6940 return size; 6941 } 6942 6943 void ScanMarkedObjectsAgainCarefullyClosure::do_yield_work() { 6944 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6945 "CMS thread should hold CMS token"); 6946 assert_lock_strong(_freelistLock); 6947 assert_lock_strong(_bitMap->lock()); 6948 DEBUG_ONLY(RememberKlassesChecker mux(false);) 6949 // relinquish the free_list_lock and bitMaplock() 6950 _bitMap->lock()->unlock(); 6951 _freelistLock->unlock(); 6952 ConcurrentMarkSweepThread::desynchronize(true); 6953 ConcurrentMarkSweepThread::acknowledge_yield_request(); 6954 _collector->stopTimer(); 6955 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 6956 if (PrintCMSStatistics != 0) { 6957 _collector->incrementYields(); 6958 } 6959 _collector->icms_wait(); 6960 6961 // See the comment in coordinator_yield() 6962 for (unsigned i = 0; i < CMSYieldSleepCount && 6963 ConcurrentMarkSweepThread::should_yield() && 6964 !CMSCollector::foregroundGCIsActive(); ++i) { 6965 os::sleep(Thread::current(), 1, false); 6966 ConcurrentMarkSweepThread::acknowledge_yield_request(); 6967 } 6968 6969 ConcurrentMarkSweepThread::synchronize(true); 6970 _freelistLock->lock_without_safepoint_check(); 6971 _bitMap->lock()->lock_without_safepoint_check(); 6972 _collector->startTimer(); 6973 } 6974 6975 6976 ////////////////////////////////////////////////////////////////// 6977 // SurvivorSpacePrecleanClosure 6978 ////////////////////////////////////////////////////////////////// 6979 // This (single-threaded) closure is used to preclean the oops in 6980 // the survivor spaces. 6981 size_t SurvivorSpacePrecleanClosure::do_object_careful(oop p) { 6982 6983 HeapWord* addr = (HeapWord*)p; 6984 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6985 assert(!_span.contains(addr), "we are scanning the survivor spaces"); 6986 assert(p->klass_or_null() != NULL, "object should be initializd"); 6987 assert(p->is_parsable(), "must be parsable."); 6988 // an initialized object; ignore mark word in verification below 6989 // since we are running concurrent with mutators 6990 assert(p->is_oop(true), "should be an oop"); 6991 // Note that we do not yield while we iterate over 6992 // the interior oops of p, pushing the relevant ones 6993 // on our marking stack. 6994 size_t size = p->oop_iterate(_scanning_closure); 6995 do_yield_check(); 6996 // Observe that below, we do not abandon the preclean 6997 // phase as soon as we should; rather we empty the 6998 // marking stack before returning. This is to satisfy 6999 // some existing assertions. In general, it may be a 7000 // good idea to abort immediately and complete the marking 7001 // from the grey objects at a later time. 7002 while (!_mark_stack->isEmpty()) { 7003 oop new_oop = _mark_stack->pop(); 7004 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 7005 assert(new_oop->is_parsable(), "Found unparsable oop"); 7006 assert(_bit_map->isMarked((HeapWord*)new_oop), 7007 "only grey objects on this stack"); 7008 // iterate over the oops in this oop, marking and pushing 7009 // the ones in CMS heap (i.e. in _span). 7010 new_oop->oop_iterate(_scanning_closure); 7011 // check if it's time to yield 7012 do_yield_check(); 7013 } 7014 unsigned int after_count = 7015 GenCollectedHeap::heap()->total_collections(); 7016 bool abort = (_before_count != after_count) || 7017 _collector->should_abort_preclean(); 7018 return abort ? 0 : size; 7019 } 7020 7021 void SurvivorSpacePrecleanClosure::do_yield_work() { 7022 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 7023 "CMS thread should hold CMS token"); 7024 assert_lock_strong(_bit_map->lock()); 7025 DEBUG_ONLY(RememberKlassesChecker smx(false);) 7026 // Relinquish the bit map lock 7027 _bit_map->lock()->unlock(); 7028 ConcurrentMarkSweepThread::desynchronize(true); 7029 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7030 _collector->stopTimer(); 7031 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 7032 if (PrintCMSStatistics != 0) { 7033 _collector->incrementYields(); 7034 } 7035 _collector->icms_wait(); 7036 7037 // See the comment in coordinator_yield() 7038 for (unsigned i = 0; i < CMSYieldSleepCount && 7039 ConcurrentMarkSweepThread::should_yield() && 7040 !CMSCollector::foregroundGCIsActive(); ++i) { 7041 os::sleep(Thread::current(), 1, false); 7042 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7043 } 7044 7045 ConcurrentMarkSweepThread::synchronize(true); 7046 _bit_map->lock()->lock_without_safepoint_check(); 7047 _collector->startTimer(); 7048 } 7049 7050 // This closure is used to rescan the marked objects on the dirty cards 7051 // in the mod union table and the card table proper. In the parallel 7052 // case, although the bitMap is shared, we do a single read so the 7053 // isMarked() query is "safe". 7054 bool ScanMarkedObjectsAgainClosure::do_object_bm(oop p, MemRegion mr) { 7055 // Ignore mark word because we are running concurrent with mutators 7056 assert(p->is_oop_or_null(true), "expected an oop or null"); 7057 HeapWord* addr = (HeapWord*)p; 7058 assert(_span.contains(addr), "we are scanning the CMS generation"); 7059 bool is_obj_array = false; 7060 #ifdef DEBUG 7061 if (!_parallel) { 7062 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 7063 assert(_collector->overflow_list_is_empty(), 7064 "overflow list should be empty"); 7065 7066 } 7067 #endif // DEBUG 7068 if (_bit_map->isMarked(addr)) { 7069 // Obj arrays are precisely marked, non-arrays are not; 7070 // so we scan objArrays precisely and non-arrays in their 7071 // entirety. 7072 if (p->is_objArray()) { 7073 is_obj_array = true; 7074 if (_parallel) { 7075 p->oop_iterate(_par_scan_closure, mr); 7076 } else { 7077 p->oop_iterate(_scan_closure, mr); 7078 } 7079 } else { 7080 if (_parallel) { 7081 p->oop_iterate(_par_scan_closure); 7082 } else { 7083 p->oop_iterate(_scan_closure); 7084 } 7085 } 7086 } 7087 #ifdef DEBUG 7088 if (!_parallel) { 7089 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 7090 assert(_collector->overflow_list_is_empty(), 7091 "overflow list should be empty"); 7092 7093 } 7094 #endif // DEBUG 7095 return is_obj_array; 7096 } 7097 7098 MarkFromRootsClosure::MarkFromRootsClosure(CMSCollector* collector, 7099 MemRegion span, 7100 CMSBitMap* bitMap, CMSMarkStack* markStack, 7101 CMSMarkStack* revisitStack, 7102 bool should_yield, bool verifying): 7103 _collector(collector), 7104 _span(span), 7105 _bitMap(bitMap), 7106 _mut(&collector->_modUnionTable), 7107 _markStack(markStack), 7108 _revisitStack(revisitStack), 7109 _yield(should_yield), 7110 _skipBits(0) 7111 { 7112 assert(_markStack->isEmpty(), "stack should be empty"); 7113 _finger = _bitMap->startWord(); 7114 _threshold = _finger; 7115 assert(_collector->_restart_addr == NULL, "Sanity check"); 7116 assert(_span.contains(_finger), "Out of bounds _finger?"); 7117 DEBUG_ONLY(_verifying = verifying;) 7118 } 7119 7120 void MarkFromRootsClosure::reset(HeapWord* addr) { 7121 assert(_markStack->isEmpty(), "would cause duplicates on stack"); 7122 assert(_span.contains(addr), "Out of bounds _finger?"); 7123 _finger = addr; 7124 _threshold = (HeapWord*)round_to( 7125 (intptr_t)_finger, CardTableModRefBS::card_size); 7126 } 7127 7128 // Should revisit to see if this should be restructured for 7129 // greater efficiency. 7130 bool MarkFromRootsClosure::do_bit(size_t offset) { 7131 if (_skipBits > 0) { 7132 _skipBits--; 7133 return true; 7134 } 7135 // convert offset into a HeapWord* 7136 HeapWord* addr = _bitMap->startWord() + offset; 7137 assert(_bitMap->endWord() && addr < _bitMap->endWord(), 7138 "address out of range"); 7139 assert(_bitMap->isMarked(addr), "tautology"); 7140 if (_bitMap->isMarked(addr+1)) { 7141 // this is an allocated but not yet initialized object 7142 assert(_skipBits == 0, "tautology"); 7143 _skipBits = 2; // skip next two marked bits ("Printezis-marks") 7144 oop p = oop(addr); 7145 if (p->klass_or_null() == NULL || !p->is_parsable()) { 7146 DEBUG_ONLY(if (!_verifying) {) 7147 // We re-dirty the cards on which this object lies and increase 7148 // the _threshold so that we'll come back to scan this object 7149 // during the preclean or remark phase. (CMSCleanOnEnter) 7150 if (CMSCleanOnEnter) { 7151 size_t sz = _collector->block_size_using_printezis_bits(addr); 7152 HeapWord* end_card_addr = (HeapWord*)round_to( 7153 (intptr_t)(addr+sz), CardTableModRefBS::card_size); 7154 MemRegion redirty_range = MemRegion(addr, end_card_addr); 7155 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 7156 // Bump _threshold to end_card_addr; note that 7157 // _threshold cannot possibly exceed end_card_addr, anyhow. 7158 // This prevents future clearing of the card as the scan proceeds 7159 // to the right. 7160 assert(_threshold <= end_card_addr, 7161 "Because we are just scanning into this object"); 7162 if (_threshold < end_card_addr) { 7163 _threshold = end_card_addr; 7164 } 7165 if (p->klass_or_null() != NULL) { 7166 // Redirty the range of cards... 7167 _mut->mark_range(redirty_range); 7168 } // ...else the setting of klass will dirty the card anyway. 7169 } 7170 DEBUG_ONLY(}) 7171 return true; 7172 } 7173 } 7174 scanOopsInOop(addr); 7175 return true; 7176 } 7177 7178 // We take a break if we've been at this for a while, 7179 // so as to avoid monopolizing the locks involved. 7180 void MarkFromRootsClosure::do_yield_work() { 7181 // First give up the locks, then yield, then re-lock 7182 // We should probably use a constructor/destructor idiom to 7183 // do this unlock/lock or modify the MutexUnlocker class to 7184 // serve our purpose. XXX 7185 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 7186 "CMS thread should hold CMS token"); 7187 assert_lock_strong(_bitMap->lock()); 7188 DEBUG_ONLY(RememberKlassesChecker mux(false);) 7189 _bitMap->lock()->unlock(); 7190 ConcurrentMarkSweepThread::desynchronize(true); 7191 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7192 _collector->stopTimer(); 7193 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 7194 if (PrintCMSStatistics != 0) { 7195 _collector->incrementYields(); 7196 } 7197 _collector->icms_wait(); 7198 7199 // See the comment in coordinator_yield() 7200 for (unsigned i = 0; i < CMSYieldSleepCount && 7201 ConcurrentMarkSweepThread::should_yield() && 7202 !CMSCollector::foregroundGCIsActive(); ++i) { 7203 os::sleep(Thread::current(), 1, false); 7204 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7205 } 7206 7207 ConcurrentMarkSweepThread::synchronize(true); 7208 _bitMap->lock()->lock_without_safepoint_check(); 7209 _collector->startTimer(); 7210 } 7211 7212 void MarkFromRootsClosure::scanOopsInOop(HeapWord* ptr) { 7213 assert(_bitMap->isMarked(ptr), "expected bit to be set"); 7214 assert(_markStack->isEmpty(), 7215 "should drain stack to limit stack usage"); 7216 // convert ptr to an oop preparatory to scanning 7217 oop obj = oop(ptr); 7218 // Ignore mark word in verification below, since we 7219 // may be running concurrent with mutators. 7220 assert(obj->is_oop(true), "should be an oop"); 7221 assert(_finger <= ptr, "_finger runneth ahead"); 7222 // advance the finger to right end of this object 7223 _finger = ptr + obj->size(); 7224 assert(_finger > ptr, "we just incremented it above"); 7225 // On large heaps, it may take us some time to get through 7226 // the marking phase (especially if running iCMS). During 7227 // this time it's possible that a lot of mutations have 7228 // accumulated in the card table and the mod union table -- 7229 // these mutation records are redundant until we have 7230 // actually traced into the corresponding card. 7231 // Here, we check whether advancing the finger would make 7232 // us cross into a new card, and if so clear corresponding 7233 // cards in the MUT (preclean them in the card-table in the 7234 // future). 7235 7236 DEBUG_ONLY(if (!_verifying) {) 7237 // The clean-on-enter optimization is disabled by default, 7238 // until we fix 6178663. 7239 if (CMSCleanOnEnter && (_finger > _threshold)) { 7240 // [_threshold, _finger) represents the interval 7241 // of cards to be cleared in MUT (or precleaned in card table). 7242 // The set of cards to be cleared is all those that overlap 7243 // with the interval [_threshold, _finger); note that 7244 // _threshold is always kept card-aligned but _finger isn't 7245 // always card-aligned. 7246 HeapWord* old_threshold = _threshold; 7247 assert(old_threshold == (HeapWord*)round_to( 7248 (intptr_t)old_threshold, CardTableModRefBS::card_size), 7249 "_threshold should always be card-aligned"); 7250 _threshold = (HeapWord*)round_to( 7251 (intptr_t)_finger, CardTableModRefBS::card_size); 7252 MemRegion mr(old_threshold, _threshold); 7253 assert(!mr.is_empty(), "Control point invariant"); 7254 assert(_span.contains(mr), "Should clear within span"); 7255 // XXX When _finger crosses from old gen into perm gen 7256 // we may be doing unnecessary cleaning; do better in the 7257 // future by detecting that condition and clearing fewer 7258 // MUT/CT entries. 7259 _mut->clear_range(mr); 7260 } 7261 DEBUG_ONLY(}) 7262 // Note: the finger doesn't advance while we drain 7263 // the stack below. 7264 PushOrMarkClosure pushOrMarkClosure(_collector, 7265 _span, _bitMap, _markStack, 7266 _revisitStack, 7267 _finger, this); 7268 bool res = _markStack->push(obj); 7269 assert(res, "Empty non-zero size stack should have space for single push"); 7270 while (!_markStack->isEmpty()) { 7271 oop new_oop = _markStack->pop(); 7272 // Skip verifying header mark word below because we are 7273 // running concurrent with mutators. 7274 assert(new_oop->is_oop(true), "Oops! expected to pop an oop"); 7275 // now scan this oop's oops 7276 new_oop->oop_iterate(&pushOrMarkClosure); 7277 do_yield_check(); 7278 } 7279 assert(_markStack->isEmpty(), "tautology, emphasizing post-condition"); 7280 } 7281 7282 Par_MarkFromRootsClosure::Par_MarkFromRootsClosure(CMSConcMarkingTask* task, 7283 CMSCollector* collector, MemRegion span, 7284 CMSBitMap* bit_map, 7285 OopTaskQueue* work_queue, 7286 CMSMarkStack* overflow_stack, 7287 CMSMarkStack* revisit_stack, 7288 bool should_yield): 7289 _collector(collector), 7290 _whole_span(collector->_span), 7291 _span(span), 7292 _bit_map(bit_map), 7293 _mut(&collector->_modUnionTable), 7294 _work_queue(work_queue), 7295 _overflow_stack(overflow_stack), 7296 _revisit_stack(revisit_stack), 7297 _yield(should_yield), 7298 _skip_bits(0), 7299 _task(task) 7300 { 7301 assert(_work_queue->size() == 0, "work_queue should be empty"); 7302 _finger = span.start(); 7303 _threshold = _finger; // XXX Defer clear-on-enter optimization for now 7304 assert(_span.contains(_finger), "Out of bounds _finger?"); 7305 } 7306 7307 // Should revisit to see if this should be restructured for 7308 // greater efficiency. 7309 bool Par_MarkFromRootsClosure::do_bit(size_t offset) { 7310 if (_skip_bits > 0) { 7311 _skip_bits--; 7312 return true; 7313 } 7314 // convert offset into a HeapWord* 7315 HeapWord* addr = _bit_map->startWord() + offset; 7316 assert(_bit_map->endWord() && addr < _bit_map->endWord(), 7317 "address out of range"); 7318 assert(_bit_map->isMarked(addr), "tautology"); 7319 if (_bit_map->isMarked(addr+1)) { 7320 // this is an allocated object that might not yet be initialized 7321 assert(_skip_bits == 0, "tautology"); 7322 _skip_bits = 2; // skip next two marked bits ("Printezis-marks") 7323 oop p = oop(addr); 7324 if (p->klass_or_null() == NULL || !p->is_parsable()) { 7325 // in the case of Clean-on-Enter optimization, redirty card 7326 // and avoid clearing card by increasing the threshold. 7327 return true; 7328 } 7329 } 7330 scan_oops_in_oop(addr); 7331 return true; 7332 } 7333 7334 void Par_MarkFromRootsClosure::scan_oops_in_oop(HeapWord* ptr) { 7335 assert(_bit_map->isMarked(ptr), "expected bit to be set"); 7336 // Should we assert that our work queue is empty or 7337 // below some drain limit? 7338 assert(_work_queue->size() == 0, 7339 "should drain stack to limit stack usage"); 7340 // convert ptr to an oop preparatory to scanning 7341 oop obj = oop(ptr); 7342 // Ignore mark word in verification below, since we 7343 // may be running concurrent with mutators. 7344 assert(obj->is_oop(true), "should be an oop"); 7345 assert(_finger <= ptr, "_finger runneth ahead"); 7346 // advance the finger to right end of this object 7347 _finger = ptr + obj->size(); 7348 assert(_finger > ptr, "we just incremented it above"); 7349 // On large heaps, it may take us some time to get through 7350 // the marking phase (especially if running iCMS). During 7351 // this time it's possible that a lot of mutations have 7352 // accumulated in the card table and the mod union table -- 7353 // these mutation records are redundant until we have 7354 // actually traced into the corresponding card. 7355 // Here, we check whether advancing the finger would make 7356 // us cross into a new card, and if so clear corresponding 7357 // cards in the MUT (preclean them in the card-table in the 7358 // future). 7359 7360 // The clean-on-enter optimization is disabled by default, 7361 // until we fix 6178663. 7362 if (CMSCleanOnEnter && (_finger > _threshold)) { 7363 // [_threshold, _finger) represents the interval 7364 // of cards to be cleared in MUT (or precleaned in card table). 7365 // The set of cards to be cleared is all those that overlap 7366 // with the interval [_threshold, _finger); note that 7367 // _threshold is always kept card-aligned but _finger isn't 7368 // always card-aligned. 7369 HeapWord* old_threshold = _threshold; 7370 assert(old_threshold == (HeapWord*)round_to( 7371 (intptr_t)old_threshold, CardTableModRefBS::card_size), 7372 "_threshold should always be card-aligned"); 7373 _threshold = (HeapWord*)round_to( 7374 (intptr_t)_finger, CardTableModRefBS::card_size); 7375 MemRegion mr(old_threshold, _threshold); 7376 assert(!mr.is_empty(), "Control point invariant"); 7377 assert(_span.contains(mr), "Should clear within span"); // _whole_span ?? 7378 // XXX When _finger crosses from old gen into perm gen 7379 // we may be doing unnecessary cleaning; do better in the 7380 // future by detecting that condition and clearing fewer 7381 // MUT/CT entries. 7382 _mut->clear_range(mr); 7383 } 7384 7385 // Note: the local finger doesn't advance while we drain 7386 // the stack below, but the global finger sure can and will. 7387 HeapWord** gfa = _task->global_finger_addr(); 7388 Par_PushOrMarkClosure pushOrMarkClosure(_collector, 7389 _span, _bit_map, 7390 _work_queue, 7391 _overflow_stack, 7392 _revisit_stack, 7393 _finger, 7394 gfa, this); 7395 bool res = _work_queue->push(obj); // overflow could occur here 7396 assert(res, "Will hold once we use workqueues"); 7397 while (true) { 7398 oop new_oop; 7399 if (!_work_queue->pop_local(new_oop)) { 7400 // We emptied our work_queue; check if there's stuff that can 7401 // be gotten from the overflow stack. 7402 if (CMSConcMarkingTask::get_work_from_overflow_stack( 7403 _overflow_stack, _work_queue)) { 7404 do_yield_check(); 7405 continue; 7406 } else { // done 7407 break; 7408 } 7409 } 7410 // Skip verifying header mark word below because we are 7411 // running concurrent with mutators. 7412 assert(new_oop->is_oop(true), "Oops! expected to pop an oop"); 7413 // now scan this oop's oops 7414 new_oop->oop_iterate(&pushOrMarkClosure); 7415 do_yield_check(); 7416 } 7417 assert(_work_queue->size() == 0, "tautology, emphasizing post-condition"); 7418 } 7419 7420 // Yield in response to a request from VM Thread or 7421 // from mutators. 7422 void Par_MarkFromRootsClosure::do_yield_work() { 7423 assert(_task != NULL, "sanity"); 7424 _task->yield(); 7425 } 7426 7427 // A variant of the above used for verifying CMS marking work. 7428 MarkFromRootsVerifyClosure::MarkFromRootsVerifyClosure(CMSCollector* collector, 7429 MemRegion span, 7430 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 7431 CMSMarkStack* mark_stack): 7432 _collector(collector), 7433 _span(span), 7434 _verification_bm(verification_bm), 7435 _cms_bm(cms_bm), 7436 _mark_stack(mark_stack), 7437 _pam_verify_closure(collector, span, verification_bm, cms_bm, 7438 mark_stack) 7439 { 7440 assert(_mark_stack->isEmpty(), "stack should be empty"); 7441 _finger = _verification_bm->startWord(); 7442 assert(_collector->_restart_addr == NULL, "Sanity check"); 7443 assert(_span.contains(_finger), "Out of bounds _finger?"); 7444 } 7445 7446 void MarkFromRootsVerifyClosure::reset(HeapWord* addr) { 7447 assert(_mark_stack->isEmpty(), "would cause duplicates on stack"); 7448 assert(_span.contains(addr), "Out of bounds _finger?"); 7449 _finger = addr; 7450 } 7451 7452 // Should revisit to see if this should be restructured for 7453 // greater efficiency. 7454 bool MarkFromRootsVerifyClosure::do_bit(size_t offset) { 7455 // convert offset into a HeapWord* 7456 HeapWord* addr = _verification_bm->startWord() + offset; 7457 assert(_verification_bm->endWord() && addr < _verification_bm->endWord(), 7458 "address out of range"); 7459 assert(_verification_bm->isMarked(addr), "tautology"); 7460 assert(_cms_bm->isMarked(addr), "tautology"); 7461 7462 assert(_mark_stack->isEmpty(), 7463 "should drain stack to limit stack usage"); 7464 // convert addr to an oop preparatory to scanning 7465 oop obj = oop(addr); 7466 assert(obj->is_oop(), "should be an oop"); 7467 assert(_finger <= addr, "_finger runneth ahead"); 7468 // advance the finger to right end of this object 7469 _finger = addr + obj->size(); 7470 assert(_finger > addr, "we just incremented it above"); 7471 // Note: the finger doesn't advance while we drain 7472 // the stack below. 7473 bool res = _mark_stack->push(obj); 7474 assert(res, "Empty non-zero size stack should have space for single push"); 7475 while (!_mark_stack->isEmpty()) { 7476 oop new_oop = _mark_stack->pop(); 7477 assert(new_oop->is_oop(), "Oops! expected to pop an oop"); 7478 // now scan this oop's oops 7479 new_oop->oop_iterate(&_pam_verify_closure); 7480 } 7481 assert(_mark_stack->isEmpty(), "tautology, emphasizing post-condition"); 7482 return true; 7483 } 7484 7485 PushAndMarkVerifyClosure::PushAndMarkVerifyClosure( 7486 CMSCollector* collector, MemRegion span, 7487 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 7488 CMSMarkStack* mark_stack): 7489 OopClosure(collector->ref_processor()), 7490 _collector(collector), 7491 _span(span), 7492 _verification_bm(verification_bm), 7493 _cms_bm(cms_bm), 7494 _mark_stack(mark_stack) 7495 { } 7496 7497 void PushAndMarkVerifyClosure::do_oop(oop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 7498 void PushAndMarkVerifyClosure::do_oop(narrowOop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 7499 7500 // Upon stack overflow, we discard (part of) the stack, 7501 // remembering the least address amongst those discarded 7502 // in CMSCollector's _restart_address. 7503 void PushAndMarkVerifyClosure::handle_stack_overflow(HeapWord* lost) { 7504 // Remember the least grey address discarded 7505 HeapWord* ra = (HeapWord*)_mark_stack->least_value(lost); 7506 _collector->lower_restart_addr(ra); 7507 _mark_stack->reset(); // discard stack contents 7508 _mark_stack->expand(); // expand the stack if possible 7509 } 7510 7511 void PushAndMarkVerifyClosure::do_oop(oop obj) { 7512 assert(obj->is_oop_or_null(), "expected an oop or NULL"); 7513 HeapWord* addr = (HeapWord*)obj; 7514 if (_span.contains(addr) && !_verification_bm->isMarked(addr)) { 7515 // Oop lies in _span and isn't yet grey or black 7516 _verification_bm->mark(addr); // now grey 7517 if (!_cms_bm->isMarked(addr)) { 7518 oop(addr)->print(); 7519 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)", 7520 addr); 7521 fatal("... aborting"); 7522 } 7523 7524 if (!_mark_stack->push(obj)) { // stack overflow 7525 if (PrintCMSStatistics != 0) { 7526 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 7527 SIZE_FORMAT, _mark_stack->capacity()); 7528 } 7529 assert(_mark_stack->isFull(), "Else push should have succeeded"); 7530 handle_stack_overflow(addr); 7531 } 7532 // anything including and to the right of _finger 7533 // will be scanned as we iterate over the remainder of the 7534 // bit map 7535 } 7536 } 7537 7538 PushOrMarkClosure::PushOrMarkClosure(CMSCollector* collector, 7539 MemRegion span, 7540 CMSBitMap* bitMap, CMSMarkStack* markStack, 7541 CMSMarkStack* revisitStack, 7542 HeapWord* finger, MarkFromRootsClosure* parent) : 7543 KlassRememberingOopClosure(collector, collector->ref_processor(), revisitStack), 7544 _span(span), 7545 _bitMap(bitMap), 7546 _markStack(markStack), 7547 _finger(finger), 7548 _parent(parent) 7549 { } 7550 7551 Par_PushOrMarkClosure::Par_PushOrMarkClosure(CMSCollector* collector, 7552 MemRegion span, 7553 CMSBitMap* bit_map, 7554 OopTaskQueue* work_queue, 7555 CMSMarkStack* overflow_stack, 7556 CMSMarkStack* revisit_stack, 7557 HeapWord* finger, 7558 HeapWord** global_finger_addr, 7559 Par_MarkFromRootsClosure* parent) : 7560 Par_KlassRememberingOopClosure(collector, 7561 collector->ref_processor(), 7562 revisit_stack), 7563 _whole_span(collector->_span), 7564 _span(span), 7565 _bit_map(bit_map), 7566 _work_queue(work_queue), 7567 _overflow_stack(overflow_stack), 7568 _finger(finger), 7569 _global_finger_addr(global_finger_addr), 7570 _parent(parent) 7571 { } 7572 7573 // Assumes thread-safe access by callers, who are 7574 // responsible for mutual exclusion. 7575 void CMSCollector::lower_restart_addr(HeapWord* low) { 7576 assert(_span.contains(low), "Out of bounds addr"); 7577 if (_restart_addr == NULL) { 7578 _restart_addr = low; 7579 } else { 7580 _restart_addr = MIN2(_restart_addr, low); 7581 } 7582 } 7583 7584 // Upon stack overflow, we discard (part of) the stack, 7585 // remembering the least address amongst those discarded 7586 // in CMSCollector's _restart_address. 7587 void PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 7588 // Remember the least grey address discarded 7589 HeapWord* ra = (HeapWord*)_markStack->least_value(lost); 7590 _collector->lower_restart_addr(ra); 7591 _markStack->reset(); // discard stack contents 7592 _markStack->expand(); // expand the stack if possible 7593 } 7594 7595 // Upon stack overflow, we discard (part of) the stack, 7596 // remembering the least address amongst those discarded 7597 // in CMSCollector's _restart_address. 7598 void Par_PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 7599 // We need to do this under a mutex to prevent other 7600 // workers from interfering with the work done below. 7601 MutexLockerEx ml(_overflow_stack->par_lock(), 7602 Mutex::_no_safepoint_check_flag); 7603 // Remember the least grey address discarded 7604 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 7605 _collector->lower_restart_addr(ra); 7606 _overflow_stack->reset(); // discard stack contents 7607 _overflow_stack->expand(); // expand the stack if possible 7608 } 7609 7610 void PushOrMarkClosure::do_oop(oop obj) { 7611 // Ignore mark word because we are running concurrent with mutators. 7612 assert(obj->is_oop_or_null(true), "expected an oop or NULL"); 7613 HeapWord* addr = (HeapWord*)obj; 7614 if (_span.contains(addr) && !_bitMap->isMarked(addr)) { 7615 // Oop lies in _span and isn't yet grey or black 7616 _bitMap->mark(addr); // now grey 7617 if (addr < _finger) { 7618 // the bit map iteration has already either passed, or 7619 // sampled, this bit in the bit map; we'll need to 7620 // use the marking stack to scan this oop's oops. 7621 bool simulate_overflow = false; 7622 NOT_PRODUCT( 7623 if (CMSMarkStackOverflowALot && 7624 _collector->simulate_overflow()) { 7625 // simulate a stack overflow 7626 simulate_overflow = true; 7627 } 7628 ) 7629 if (simulate_overflow || !_markStack->push(obj)) { // stack overflow 7630 if (PrintCMSStatistics != 0) { 7631 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 7632 SIZE_FORMAT, _markStack->capacity()); 7633 } 7634 assert(simulate_overflow || _markStack->isFull(), "Else push should have succeeded"); 7635 handle_stack_overflow(addr); 7636 } 7637 } 7638 // anything including and to the right of _finger 7639 // will be scanned as we iterate over the remainder of the 7640 // bit map 7641 do_yield_check(); 7642 } 7643 } 7644 7645 void PushOrMarkClosure::do_oop(oop* p) { PushOrMarkClosure::do_oop_work(p); } 7646 void PushOrMarkClosure::do_oop(narrowOop* p) { PushOrMarkClosure::do_oop_work(p); } 7647 7648 void Par_PushOrMarkClosure::do_oop(oop obj) { 7649 // Ignore mark word because we are running concurrent with mutators. 7650 assert(obj->is_oop_or_null(true), "expected an oop or NULL"); 7651 HeapWord* addr = (HeapWord*)obj; 7652 if (_whole_span.contains(addr) && !_bit_map->isMarked(addr)) { 7653 // Oop lies in _span and isn't yet grey or black 7654 // We read the global_finger (volatile read) strictly after marking oop 7655 bool res = _bit_map->par_mark(addr); // now grey 7656 volatile HeapWord** gfa = (volatile HeapWord**)_global_finger_addr; 7657 // Should we push this marked oop on our stack? 7658 // -- if someone else marked it, nothing to do 7659 // -- if target oop is above global finger nothing to do 7660 // -- if target oop is in chunk and above local finger 7661 // then nothing to do 7662 // -- else push on work queue 7663 if ( !res // someone else marked it, they will deal with it 7664 || (addr >= *gfa) // will be scanned in a later task 7665 || (_span.contains(addr) && addr >= _finger)) { // later in this chunk 7666 return; 7667 } 7668 // the bit map iteration has already either passed, or 7669 // sampled, this bit in the bit map; we'll need to 7670 // use the marking stack to scan this oop's oops. 7671 bool simulate_overflow = false; 7672 NOT_PRODUCT( 7673 if (CMSMarkStackOverflowALot && 7674 _collector->simulate_overflow()) { 7675 // simulate a stack overflow 7676 simulate_overflow = true; 7677 } 7678 ) 7679 if (simulate_overflow || 7680 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) { 7681 // stack overflow 7682 if (PrintCMSStatistics != 0) { 7683 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 7684 SIZE_FORMAT, _overflow_stack->capacity()); 7685 } 7686 // We cannot assert that the overflow stack is full because 7687 // it may have been emptied since. 7688 assert(simulate_overflow || 7689 _work_queue->size() == _work_queue->max_elems(), 7690 "Else push should have succeeded"); 7691 handle_stack_overflow(addr); 7692 } 7693 do_yield_check(); 7694 } 7695 } 7696 7697 void Par_PushOrMarkClosure::do_oop(oop* p) { Par_PushOrMarkClosure::do_oop_work(p); } 7698 void Par_PushOrMarkClosure::do_oop(narrowOop* p) { Par_PushOrMarkClosure::do_oop_work(p); } 7699 7700 KlassRememberingOopClosure::KlassRememberingOopClosure(CMSCollector* collector, 7701 ReferenceProcessor* rp, 7702 CMSMarkStack* revisit_stack) : 7703 OopClosure(rp), 7704 _collector(collector), 7705 _revisit_stack(revisit_stack), 7706 _should_remember_klasses(collector->should_unload_classes()) {} 7707 7708 PushAndMarkClosure::PushAndMarkClosure(CMSCollector* collector, 7709 MemRegion span, 7710 ReferenceProcessor* rp, 7711 CMSBitMap* bit_map, 7712 CMSBitMap* mod_union_table, 7713 CMSMarkStack* mark_stack, 7714 CMSMarkStack* revisit_stack, 7715 bool concurrent_precleaning): 7716 KlassRememberingOopClosure(collector, rp, revisit_stack), 7717 _span(span), 7718 _bit_map(bit_map), 7719 _mod_union_table(mod_union_table), 7720 _mark_stack(mark_stack), 7721 _concurrent_precleaning(concurrent_precleaning) 7722 { 7723 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 7724 } 7725 7726 // Grey object rescan during pre-cleaning and second checkpoint phases -- 7727 // the non-parallel version (the parallel version appears further below.) 7728 void PushAndMarkClosure::do_oop(oop obj) { 7729 // Ignore mark word verification. If during concurrent precleaning, 7730 // the object monitor may be locked. If during the checkpoint 7731 // phases, the object may already have been reached by a different 7732 // path and may be at the end of the global overflow list (so 7733 // the mark word may be NULL). 7734 assert(obj->is_oop_or_null(true /* ignore mark word */), 7735 "expected an oop or NULL"); 7736 HeapWord* addr = (HeapWord*)obj; 7737 // Check if oop points into the CMS generation 7738 // and is not marked 7739 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 7740 // a white object ... 7741 _bit_map->mark(addr); // ... now grey 7742 // push on the marking stack (grey set) 7743 bool simulate_overflow = false; 7744 NOT_PRODUCT( 7745 if (CMSMarkStackOverflowALot && 7746 _collector->simulate_overflow()) { 7747 // simulate a stack overflow 7748 simulate_overflow = true; 7749 } 7750 ) 7751 if (simulate_overflow || !_mark_stack->push(obj)) { 7752 if (_concurrent_precleaning) { 7753 // During precleaning we can just dirty the appropriate card(s) 7754 // in the mod union table, thus ensuring that the object remains 7755 // in the grey set and continue. In the case of object arrays 7756 // we need to dirty all of the cards that the object spans, 7757 // since the rescan of object arrays will be limited to the 7758 // dirty cards. 7759 // Note that no one can be intefering with us in this action 7760 // of dirtying the mod union table, so no locking or atomics 7761 // are required. 7762 if (obj->is_objArray()) { 7763 size_t sz = obj->size(); 7764 HeapWord* end_card_addr = (HeapWord*)round_to( 7765 (intptr_t)(addr+sz), CardTableModRefBS::card_size); 7766 MemRegion redirty_range = MemRegion(addr, end_card_addr); 7767 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 7768 _mod_union_table->mark_range(redirty_range); 7769 } else { 7770 _mod_union_table->mark(addr); 7771 } 7772 _collector->_ser_pmc_preclean_ovflw++; 7773 } else { 7774 // During the remark phase, we need to remember this oop 7775 // in the overflow list. 7776 _collector->push_on_overflow_list(obj); 7777 _collector->_ser_pmc_remark_ovflw++; 7778 } 7779 } 7780 } 7781 } 7782 7783 Par_PushAndMarkClosure::Par_PushAndMarkClosure(CMSCollector* collector, 7784 MemRegion span, 7785 ReferenceProcessor* rp, 7786 CMSBitMap* bit_map, 7787 OopTaskQueue* work_queue, 7788 CMSMarkStack* revisit_stack): 7789 Par_KlassRememberingOopClosure(collector, rp, revisit_stack), 7790 _span(span), 7791 _bit_map(bit_map), 7792 _work_queue(work_queue) 7793 { 7794 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 7795 } 7796 7797 void PushAndMarkClosure::do_oop(oop* p) { PushAndMarkClosure::do_oop_work(p); } 7798 void PushAndMarkClosure::do_oop(narrowOop* p) { PushAndMarkClosure::do_oop_work(p); } 7799 7800 // Grey object rescan during second checkpoint phase -- 7801 // the parallel version. 7802 void Par_PushAndMarkClosure::do_oop(oop obj) { 7803 // In the assert below, we ignore the mark word because 7804 // this oop may point to an already visited object that is 7805 // on the overflow stack (in which case the mark word has 7806 // been hijacked for chaining into the overflow stack -- 7807 // if this is the last object in the overflow stack then 7808 // its mark word will be NULL). Because this object may 7809 // have been subsequently popped off the global overflow 7810 // stack, and the mark word possibly restored to the prototypical 7811 // value, by the time we get to examined this failing assert in 7812 // the debugger, is_oop_or_null(false) may subsequently start 7813 // to hold. 7814 assert(obj->is_oop_or_null(true), 7815 "expected an oop or NULL"); 7816 HeapWord* addr = (HeapWord*)obj; 7817 // Check if oop points into the CMS generation 7818 // and is not marked 7819 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 7820 // a white object ... 7821 // If we manage to "claim" the object, by being the 7822 // first thread to mark it, then we push it on our 7823 // marking stack 7824 if (_bit_map->par_mark(addr)) { // ... now grey 7825 // push on work queue (grey set) 7826 bool simulate_overflow = false; 7827 NOT_PRODUCT( 7828 if (CMSMarkStackOverflowALot && 7829 _collector->par_simulate_overflow()) { 7830 // simulate a stack overflow 7831 simulate_overflow = true; 7832 } 7833 ) 7834 if (simulate_overflow || !_work_queue->push(obj)) { 7835 _collector->par_push_on_overflow_list(obj); 7836 _collector->_par_pmc_remark_ovflw++; // imprecise OK: no need to CAS 7837 } 7838 } // Else, some other thread got there first 7839 } 7840 } 7841 7842 void Par_PushAndMarkClosure::do_oop(oop* p) { Par_PushAndMarkClosure::do_oop_work(p); } 7843 void Par_PushAndMarkClosure::do_oop(narrowOop* p) { Par_PushAndMarkClosure::do_oop_work(p); } 7844 7845 void PushAndMarkClosure::remember_mdo(DataLayout* v) { 7846 // TBD 7847 } 7848 7849 void Par_PushAndMarkClosure::remember_mdo(DataLayout* v) { 7850 // TBD 7851 } 7852 7853 void CMSPrecleanRefsYieldClosure::do_yield_work() { 7854 DEBUG_ONLY(RememberKlassesChecker mux(false);) 7855 Mutex* bml = _collector->bitMapLock(); 7856 assert_lock_strong(bml); 7857 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 7858 "CMS thread should hold CMS token"); 7859 7860 bml->unlock(); 7861 ConcurrentMarkSweepThread::desynchronize(true); 7862 7863 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7864 7865 _collector->stopTimer(); 7866 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 7867 if (PrintCMSStatistics != 0) { 7868 _collector->incrementYields(); 7869 } 7870 _collector->icms_wait(); 7871 7872 // See the comment in coordinator_yield() 7873 for (unsigned i = 0; i < CMSYieldSleepCount && 7874 ConcurrentMarkSweepThread::should_yield() && 7875 !CMSCollector::foregroundGCIsActive(); ++i) { 7876 os::sleep(Thread::current(), 1, false); 7877 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7878 } 7879 7880 ConcurrentMarkSweepThread::synchronize(true); 7881 bml->lock(); 7882 7883 _collector->startTimer(); 7884 } 7885 7886 bool CMSPrecleanRefsYieldClosure::should_return() { 7887 if (ConcurrentMarkSweepThread::should_yield()) { 7888 do_yield_work(); 7889 } 7890 return _collector->foregroundGCIsActive(); 7891 } 7892 7893 void MarkFromDirtyCardsClosure::do_MemRegion(MemRegion mr) { 7894 assert(((size_t)mr.start())%CardTableModRefBS::card_size_in_words == 0, 7895 "mr should be aligned to start at a card boundary"); 7896 // We'd like to assert: 7897 // assert(mr.word_size()%CardTableModRefBS::card_size_in_words == 0, 7898 // "mr should be a range of cards"); 7899 // However, that would be too strong in one case -- the last 7900 // partition ends at _unallocated_block which, in general, can be 7901 // an arbitrary boundary, not necessarily card aligned. 7902 if (PrintCMSStatistics != 0) { 7903 _num_dirty_cards += 7904 mr.word_size()/CardTableModRefBS::card_size_in_words; 7905 } 7906 _space->object_iterate_mem(mr, &_scan_cl); 7907 } 7908 7909 SweepClosure::SweepClosure(CMSCollector* collector, 7910 ConcurrentMarkSweepGeneration* g, 7911 CMSBitMap* bitMap, bool should_yield) : 7912 _collector(collector), 7913 _g(g), 7914 _sp(g->cmsSpace()), 7915 _limit(_sp->sweep_limit()), 7916 _freelistLock(_sp->freelistLock()), 7917 _bitMap(bitMap), 7918 _yield(should_yield), 7919 _inFreeRange(false), // No free range at beginning of sweep 7920 _freeRangeInFreeLists(false), // No free range at beginning of sweep 7921 _lastFreeRangeCoalesced(false), 7922 _freeFinger(g->used_region().start()) 7923 { 7924 NOT_PRODUCT( 7925 _numObjectsFreed = 0; 7926 _numWordsFreed = 0; 7927 _numObjectsLive = 0; 7928 _numWordsLive = 0; 7929 _numObjectsAlreadyFree = 0; 7930 _numWordsAlreadyFree = 0; 7931 _last_fc = NULL; 7932 7933 _sp->initializeIndexedFreeListArrayReturnedBytes(); 7934 _sp->dictionary()->initialize_dict_returned_bytes(); 7935 ) 7936 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7937 "sweep _limit out of bounds"); 7938 if (CMSTraceSweeper) { 7939 gclog_or_tty->print_cr("\n====================\nStarting new sweep with limit " PTR_FORMAT, 7940 _limit); 7941 } 7942 } 7943 7944 void SweepClosure::print_on(outputStream* st) const { 7945 tty->print_cr("_sp = [" PTR_FORMAT "," PTR_FORMAT ")", 7946 _sp->bottom(), _sp->end()); 7947 tty->print_cr("_limit = " PTR_FORMAT, _limit); 7948 tty->print_cr("_freeFinger = " PTR_FORMAT, _freeFinger); 7949 NOT_PRODUCT(tty->print_cr("_last_fc = " PTR_FORMAT, _last_fc);) 7950 tty->print_cr("_inFreeRange = %d, _freeRangeInFreeLists = %d, _lastFreeRangeCoalesced = %d", 7951 _inFreeRange, _freeRangeInFreeLists, _lastFreeRangeCoalesced); 7952 } 7953 7954 #ifndef PRODUCT 7955 // Assertion checking only: no useful work in product mode -- 7956 // however, if any of the flags below become product flags, 7957 // you may need to review this code to see if it needs to be 7958 // enabled in product mode. 7959 SweepClosure::~SweepClosure() { 7960 assert_lock_strong(_freelistLock); 7961 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7962 "sweep _limit out of bounds"); 7963 if (inFreeRange()) { 7964 warning("inFreeRange() should have been reset; dumping state of SweepClosure"); 7965 print(); 7966 ShouldNotReachHere(); 7967 } 7968 if (Verbose && PrintGC) { 7969 gclog_or_tty->print("Collected "SIZE_FORMAT" objects, " SIZE_FORMAT " bytes", 7970 _numObjectsFreed, _numWordsFreed*sizeof(HeapWord)); 7971 gclog_or_tty->print_cr("\nLive "SIZE_FORMAT" objects, " 7972 SIZE_FORMAT" bytes " 7973 "Already free "SIZE_FORMAT" objects, "SIZE_FORMAT" bytes", 7974 _numObjectsLive, _numWordsLive*sizeof(HeapWord), 7975 _numObjectsAlreadyFree, _numWordsAlreadyFree*sizeof(HeapWord)); 7976 size_t totalBytes = (_numWordsFreed + _numWordsLive + _numWordsAlreadyFree) 7977 * sizeof(HeapWord); 7978 gclog_or_tty->print_cr("Total sweep: "SIZE_FORMAT" bytes", totalBytes); 7979 7980 if (PrintCMSStatistics && CMSVerifyReturnedBytes) { 7981 size_t indexListReturnedBytes = _sp->sumIndexedFreeListArrayReturnedBytes(); 7982 size_t dict_returned_bytes = _sp->dictionary()->sum_dict_returned_bytes(); 7983 size_t returned_bytes = indexListReturnedBytes + dict_returned_bytes; 7984 gclog_or_tty->print("Returned "SIZE_FORMAT" bytes", returned_bytes); 7985 gclog_or_tty->print(" Indexed List Returned "SIZE_FORMAT" bytes", 7986 indexListReturnedBytes); 7987 gclog_or_tty->print_cr(" Dictionary Returned "SIZE_FORMAT" bytes", 7988 dict_returned_bytes); 7989 } 7990 } 7991 if (CMSTraceSweeper) { 7992 gclog_or_tty->print_cr("end of sweep with _limit = " PTR_FORMAT "\n================", 7993 _limit); 7994 } 7995 } 7996 #endif // PRODUCT 7997 7998 void SweepClosure::initialize_free_range(HeapWord* freeFinger, 7999 bool freeRangeInFreeLists) { 8000 if (CMSTraceSweeper) { 8001 gclog_or_tty->print("---- Start free range at 0x%x with free block (%d)\n", 8002 freeFinger, freeRangeInFreeLists); 8003 } 8004 assert(!inFreeRange(), "Trampling existing free range"); 8005 set_inFreeRange(true); 8006 set_lastFreeRangeCoalesced(false); 8007 8008 set_freeFinger(freeFinger); 8009 set_freeRangeInFreeLists(freeRangeInFreeLists); 8010 if (CMSTestInFreeList) { 8011 if (freeRangeInFreeLists) { 8012 FreeChunk* fc = (FreeChunk*) freeFinger; 8013 assert(fc->is_free(), "A chunk on the free list should be free."); 8014 assert(fc->size() > 0, "Free range should have a size"); 8015 assert(_sp->verify_chunk_in_free_list(fc), "Chunk is not in free lists"); 8016 } 8017 } 8018 } 8019 8020 // Note that the sweeper runs concurrently with mutators. Thus, 8021 // it is possible for direct allocation in this generation to happen 8022 // in the middle of the sweep. Note that the sweeper also coalesces 8023 // contiguous free blocks. Thus, unless the sweeper and the allocator 8024 // synchronize appropriately freshly allocated blocks may get swept up. 8025 // This is accomplished by the sweeper locking the free lists while 8026 // it is sweeping. Thus blocks that are determined to be free are 8027 // indeed free. There is however one additional complication: 8028 // blocks that have been allocated since the final checkpoint and 8029 // mark, will not have been marked and so would be treated as 8030 // unreachable and swept up. To prevent this, the allocator marks 8031 // the bit map when allocating during the sweep phase. This leads, 8032 // however, to a further complication -- objects may have been allocated 8033 // but not yet initialized -- in the sense that the header isn't yet 8034 // installed. The sweeper can not then determine the size of the block 8035 // in order to skip over it. To deal with this case, we use a technique 8036 // (due to Printezis) to encode such uninitialized block sizes in the 8037 // bit map. Since the bit map uses a bit per every HeapWord, but the 8038 // CMS generation has a minimum object size of 3 HeapWords, it follows 8039 // that "normal marks" won't be adjacent in the bit map (there will 8040 // always be at least two 0 bits between successive 1 bits). We make use 8041 // of these "unused" bits to represent uninitialized blocks -- the bit 8042 // corresponding to the start of the uninitialized object and the next 8043 // bit are both set. Finally, a 1 bit marks the end of the object that 8044 // started with the two consecutive 1 bits to indicate its potentially 8045 // uninitialized state. 8046 8047 size_t SweepClosure::do_blk_careful(HeapWord* addr) { 8048 FreeChunk* fc = (FreeChunk*)addr; 8049 size_t res; 8050 8051 // Check if we are done sweeping. Below we check "addr >= _limit" rather 8052 // than "addr == _limit" because although _limit was a block boundary when 8053 // we started the sweep, it may no longer be one because heap expansion 8054 // may have caused us to coalesce the block ending at the address _limit 8055 // with a newly expanded chunk (this happens when _limit was set to the 8056 // previous _end of the space), so we may have stepped past _limit: 8057 // see the following Zeno-like trail of CRs 6977970, 7008136, 7042740. 8058 if (addr >= _limit) { // we have swept up to or past the limit: finish up 8059 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 8060 "sweep _limit out of bounds"); 8061 assert(addr < _sp->end(), "addr out of bounds"); 8062 // Flush any free range we might be holding as a single 8063 // coalesced chunk to the appropriate free list. 8064 if (inFreeRange()) { 8065 assert(freeFinger() >= _sp->bottom() && freeFinger() < _limit, 8066 err_msg("freeFinger() " PTR_FORMAT" is out-of-bounds", freeFinger())); 8067 flush_cur_free_chunk(freeFinger(), 8068 pointer_delta(addr, freeFinger())); 8069 if (CMSTraceSweeper) { 8070 gclog_or_tty->print("Sweep: last chunk: "); 8071 gclog_or_tty->print("put_free_blk 0x%x ("SIZE_FORMAT") " 8072 "[coalesced:"SIZE_FORMAT"]\n", 8073 freeFinger(), pointer_delta(addr, freeFinger()), 8074 lastFreeRangeCoalesced()); 8075 } 8076 } 8077 8078 // help the iterator loop finish 8079 return pointer_delta(_sp->end(), addr); 8080 } 8081 8082 assert(addr < _limit, "sweep invariant"); 8083 // check if we should yield 8084 do_yield_check(addr); 8085 if (fc->is_free()) { 8086 // Chunk that is already free 8087 res = fc->size(); 8088 do_already_free_chunk(fc); 8089 debug_only(_sp->verifyFreeLists()); 8090 // If we flush the chunk at hand in lookahead_and_flush() 8091 // and it's coalesced with a preceding chunk, then the 8092 // process of "mangling" the payload of the coalesced block 8093 // will cause erasure of the size information from the 8094 // (erstwhile) header of all the coalesced blocks but the 8095 // first, so the first disjunct in the assert will not hold 8096 // in that specific case (in which case the second disjunct 8097 // will hold). 8098 assert(res == fc->size() || ((HeapWord*)fc) + res >= _limit, 8099 "Otherwise the size info doesn't change at this step"); 8100 NOT_PRODUCT( 8101 _numObjectsAlreadyFree++; 8102 _numWordsAlreadyFree += res; 8103 ) 8104 NOT_PRODUCT(_last_fc = fc;) 8105 } else if (!_bitMap->isMarked(addr)) { 8106 // Chunk is fresh garbage 8107 res = do_garbage_chunk(fc); 8108 debug_only(_sp->verifyFreeLists()); 8109 NOT_PRODUCT( 8110 _numObjectsFreed++; 8111 _numWordsFreed += res; 8112 ) 8113 } else { 8114 // Chunk that is alive. 8115 res = do_live_chunk(fc); 8116 debug_only(_sp->verifyFreeLists()); 8117 NOT_PRODUCT( 8118 _numObjectsLive++; 8119 _numWordsLive += res; 8120 ) 8121 } 8122 return res; 8123 } 8124 8125 // For the smart allocation, record following 8126 // split deaths - a free chunk is removed from its free list because 8127 // it is being split into two or more chunks. 8128 // split birth - a free chunk is being added to its free list because 8129 // a larger free chunk has been split and resulted in this free chunk. 8130 // coal death - a free chunk is being removed from its free list because 8131 // it is being coalesced into a large free chunk. 8132 // coal birth - a free chunk is being added to its free list because 8133 // it was created when two or more free chunks where coalesced into 8134 // this free chunk. 8135 // 8136 // These statistics are used to determine the desired number of free 8137 // chunks of a given size. The desired number is chosen to be relative 8138 // to the end of a CMS sweep. The desired number at the end of a sweep 8139 // is the 8140 // count-at-end-of-previous-sweep (an amount that was enough) 8141 // - count-at-beginning-of-current-sweep (the excess) 8142 // + split-births (gains in this size during interval) 8143 // - split-deaths (demands on this size during interval) 8144 // where the interval is from the end of one sweep to the end of the 8145 // next. 8146 // 8147 // When sweeping the sweeper maintains an accumulated chunk which is 8148 // the chunk that is made up of chunks that have been coalesced. That 8149 // will be termed the left-hand chunk. A new chunk of garbage that 8150 // is being considered for coalescing will be referred to as the 8151 // right-hand chunk. 8152 // 8153 // When making a decision on whether to coalesce a right-hand chunk with 8154 // the current left-hand chunk, the current count vs. the desired count 8155 // of the left-hand chunk is considered. Also if the right-hand chunk 8156 // is near the large chunk at the end of the heap (see 8157 // ConcurrentMarkSweepGeneration::isNearLargestChunk()), then the 8158 // left-hand chunk is coalesced. 8159 // 8160 // When making a decision about whether to split a chunk, the desired count 8161 // vs. the current count of the candidate to be split is also considered. 8162 // If the candidate is underpopulated (currently fewer chunks than desired) 8163 // a chunk of an overpopulated (currently more chunks than desired) size may 8164 // be chosen. The "hint" associated with a free list, if non-null, points 8165 // to a free list which may be overpopulated. 8166 // 8167 8168 void SweepClosure::do_already_free_chunk(FreeChunk* fc) { 8169 const size_t size = fc->size(); 8170 // Chunks that cannot be coalesced are not in the 8171 // free lists. 8172 if (CMSTestInFreeList && !fc->cantCoalesce()) { 8173 assert(_sp->verify_chunk_in_free_list(fc), 8174 "free chunk should be in free lists"); 8175 } 8176 // a chunk that is already free, should not have been 8177 // marked in the bit map 8178 HeapWord* const addr = (HeapWord*) fc; 8179 assert(!_bitMap->isMarked(addr), "free chunk should be unmarked"); 8180 // Verify that the bit map has no bits marked between 8181 // addr and purported end of this block. 8182 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 8183 8184 // Some chunks cannot be coalesced under any circumstances. 8185 // See the definition of cantCoalesce(). 8186 if (!fc->cantCoalesce()) { 8187 // This chunk can potentially be coalesced. 8188 if (_sp->adaptive_freelists()) { 8189 // All the work is done in 8190 do_post_free_or_garbage_chunk(fc, size); 8191 } else { // Not adaptive free lists 8192 // this is a free chunk that can potentially be coalesced by the sweeper; 8193 if (!inFreeRange()) { 8194 // if the next chunk is a free block that can't be coalesced 8195 // it doesn't make sense to remove this chunk from the free lists 8196 FreeChunk* nextChunk = (FreeChunk*)(addr + size); 8197 assert((HeapWord*)nextChunk <= _sp->end(), "Chunk size out of bounds?"); 8198 if ((HeapWord*)nextChunk < _sp->end() && // There is another free chunk to the right ... 8199 nextChunk->is_free() && // ... which is free... 8200 nextChunk->cantCoalesce()) { // ... but can't be coalesced 8201 // nothing to do 8202 } else { 8203 // Potentially the start of a new free range: 8204 // Don't eagerly remove it from the free lists. 8205 // No need to remove it if it will just be put 8206 // back again. (Also from a pragmatic point of view 8207 // if it is a free block in a region that is beyond 8208 // any allocated blocks, an assertion will fail) 8209 // Remember the start of a free run. 8210 initialize_free_range(addr, true); 8211 // end - can coalesce with next chunk 8212 } 8213 } else { 8214 // the midst of a free range, we are coalescing 8215 print_free_block_coalesced(fc); 8216 if (CMSTraceSweeper) { 8217 gclog_or_tty->print(" -- pick up free block 0x%x (%d)\n", fc, size); 8218 } 8219 // remove it from the free lists 8220 _sp->removeFreeChunkFromFreeLists(fc); 8221 set_lastFreeRangeCoalesced(true); 8222 // If the chunk is being coalesced and the current free range is 8223 // in the free lists, remove the current free range so that it 8224 // will be returned to the free lists in its entirety - all 8225 // the coalesced pieces included. 8226 if (freeRangeInFreeLists()) { 8227 FreeChunk* ffc = (FreeChunk*) freeFinger(); 8228 assert(ffc->size() == pointer_delta(addr, freeFinger()), 8229 "Size of free range is inconsistent with chunk size."); 8230 if (CMSTestInFreeList) { 8231 assert(_sp->verify_chunk_in_free_list(ffc), 8232 "free range is not in free lists"); 8233 } 8234 _sp->removeFreeChunkFromFreeLists(ffc); 8235 set_freeRangeInFreeLists(false); 8236 } 8237 } 8238 } 8239 // Note that if the chunk is not coalescable (the else arm 8240 // below), we unconditionally flush, without needing to do 8241 // a "lookahead," as we do below. 8242 if (inFreeRange()) lookahead_and_flush(fc, size); 8243 } else { 8244 // Code path common to both original and adaptive free lists. 8245 8246 // cant coalesce with previous block; this should be treated 8247 // as the end of a free run if any 8248 if (inFreeRange()) { 8249 // we kicked some butt; time to pick up the garbage 8250 assert(freeFinger() < addr, "freeFinger points too high"); 8251 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 8252 } 8253 // else, nothing to do, just continue 8254 } 8255 } 8256 8257 size_t SweepClosure::do_garbage_chunk(FreeChunk* fc) { 8258 // This is a chunk of garbage. It is not in any free list. 8259 // Add it to a free list or let it possibly be coalesced into 8260 // a larger chunk. 8261 HeapWord* const addr = (HeapWord*) fc; 8262 const size_t size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 8263 8264 if (_sp->adaptive_freelists()) { 8265 // Verify that the bit map has no bits marked between 8266 // addr and purported end of just dead object. 8267 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 8268 8269 do_post_free_or_garbage_chunk(fc, size); 8270 } else { 8271 if (!inFreeRange()) { 8272 // start of a new free range 8273 assert(size > 0, "A free range should have a size"); 8274 initialize_free_range(addr, false); 8275 } else { 8276 // this will be swept up when we hit the end of the 8277 // free range 8278 if (CMSTraceSweeper) { 8279 gclog_or_tty->print(" -- pick up garbage 0x%x (%d) \n", fc, size); 8280 } 8281 // If the chunk is being coalesced and the current free range is 8282 // in the free lists, remove the current free range so that it 8283 // will be returned to the free lists in its entirety - all 8284 // the coalesced pieces included. 8285 if (freeRangeInFreeLists()) { 8286 FreeChunk* ffc = (FreeChunk*)freeFinger(); 8287 assert(ffc->size() == pointer_delta(addr, freeFinger()), 8288 "Size of free range is inconsistent with chunk size."); 8289 if (CMSTestInFreeList) { 8290 assert(_sp->verify_chunk_in_free_list(ffc), 8291 "free range is not in free lists"); 8292 } 8293 _sp->removeFreeChunkFromFreeLists(ffc); 8294 set_freeRangeInFreeLists(false); 8295 } 8296 set_lastFreeRangeCoalesced(true); 8297 } 8298 // this will be swept up when we hit the end of the free range 8299 8300 // Verify that the bit map has no bits marked between 8301 // addr and purported end of just dead object. 8302 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 8303 } 8304 assert(_limit >= addr + size, 8305 "A freshly garbage chunk can't possibly straddle over _limit"); 8306 if (inFreeRange()) lookahead_and_flush(fc, size); 8307 return size; 8308 } 8309 8310 size_t SweepClosure::do_live_chunk(FreeChunk* fc) { 8311 HeapWord* addr = (HeapWord*) fc; 8312 // The sweeper has just found a live object. Return any accumulated 8313 // left hand chunk to the free lists. 8314 if (inFreeRange()) { 8315 assert(freeFinger() < addr, "freeFinger points too high"); 8316 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 8317 } 8318 8319 // This object is live: we'd normally expect this to be 8320 // an oop, and like to assert the following: 8321 // assert(oop(addr)->is_oop(), "live block should be an oop"); 8322 // However, as we commented above, this may be an object whose 8323 // header hasn't yet been initialized. 8324 size_t size; 8325 assert(_bitMap->isMarked(addr), "Tautology for this control point"); 8326 if (_bitMap->isMarked(addr + 1)) { 8327 // Determine the size from the bit map, rather than trying to 8328 // compute it from the object header. 8329 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 8330 size = pointer_delta(nextOneAddr + 1, addr); 8331 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 8332 "alignment problem"); 8333 8334 #ifdef DEBUG 8335 if (oop(addr)->klass_or_null() != NULL && 8336 ( !_collector->should_unload_classes() 8337 || (oop(addr)->is_parsable()) && 8338 oop(addr)->is_conc_safe())) { 8339 // Ignore mark word because we are running concurrent with mutators 8340 assert(oop(addr)->is_oop(true), "live block should be an oop"); 8341 // is_conc_safe is checked before performing this assertion 8342 // because an object that is not is_conc_safe may yet have 8343 // the return from size() correct. 8344 assert(size == 8345 CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()), 8346 "P-mark and computed size do not agree"); 8347 } 8348 #endif 8349 8350 } else { 8351 // This should be an initialized object that's alive. 8352 assert(oop(addr)->klass_or_null() != NULL && 8353 (!_collector->should_unload_classes() 8354 || oop(addr)->is_parsable()), 8355 "Should be an initialized object"); 8356 // Note that there are objects used during class redefinition, 8357 // e.g. merge_cp in VM_RedefineClasses::merge_cp_and_rewrite(), 8358 // which are discarded with their is_conc_safe state still 8359 // false. These object may be floating garbage so may be 8360 // seen here. If they are floating garbage their size 8361 // should be attainable from their klass. Do not that 8362 // is_conc_safe() is true for oop(addr). 8363 // Ignore mark word because we are running concurrent with mutators 8364 assert(oop(addr)->is_oop(true), "live block should be an oop"); 8365 // Verify that the bit map has no bits marked between 8366 // addr and purported end of this block. 8367 size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 8368 assert(size >= 3, "Necessary for Printezis marks to work"); 8369 assert(!_bitMap->isMarked(addr+1), "Tautology for this control point"); 8370 DEBUG_ONLY(_bitMap->verifyNoOneBitsInRange(addr+2, addr+size);) 8371 } 8372 return size; 8373 } 8374 8375 void SweepClosure::do_post_free_or_garbage_chunk(FreeChunk* fc, 8376 size_t chunkSize) { 8377 // do_post_free_or_garbage_chunk() should only be called in the case 8378 // of the adaptive free list allocator. 8379 const bool fcInFreeLists = fc->is_free(); 8380 assert(_sp->adaptive_freelists(), "Should only be used in this case."); 8381 assert((HeapWord*)fc <= _limit, "sweep invariant"); 8382 if (CMSTestInFreeList && fcInFreeLists) { 8383 assert(_sp->verify_chunk_in_free_list(fc), "free chunk is not in free lists"); 8384 } 8385 8386 if (CMSTraceSweeper) { 8387 gclog_or_tty->print_cr(" -- pick up another chunk at 0x%x (%d)", fc, chunkSize); 8388 } 8389 8390 HeapWord* const fc_addr = (HeapWord*) fc; 8391 8392 bool coalesce; 8393 const size_t left = pointer_delta(fc_addr, freeFinger()); 8394 const size_t right = chunkSize; 8395 switch (FLSCoalescePolicy) { 8396 // numeric value forms a coalition aggressiveness metric 8397 case 0: { // never coalesce 8398 coalesce = false; 8399 break; 8400 } 8401 case 1: { // coalesce if left & right chunks on overpopulated lists 8402 coalesce = _sp->coalOverPopulated(left) && 8403 _sp->coalOverPopulated(right); 8404 break; 8405 } 8406 case 2: { // coalesce if left chunk on overpopulated list (default) 8407 coalesce = _sp->coalOverPopulated(left); 8408 break; 8409 } 8410 case 3: { // coalesce if left OR right chunk on overpopulated list 8411 coalesce = _sp->coalOverPopulated(left) || 8412 _sp->coalOverPopulated(right); 8413 break; 8414 } 8415 case 4: { // always coalesce 8416 coalesce = true; 8417 break; 8418 } 8419 default: 8420 ShouldNotReachHere(); 8421 } 8422 8423 // Should the current free range be coalesced? 8424 // If the chunk is in a free range and either we decided to coalesce above 8425 // or the chunk is near the large block at the end of the heap 8426 // (isNearLargestChunk() returns true), then coalesce this chunk. 8427 const bool doCoalesce = inFreeRange() 8428 && (coalesce || _g->isNearLargestChunk(fc_addr)); 8429 if (doCoalesce) { 8430 // Coalesce the current free range on the left with the new 8431 // chunk on the right. If either is on a free list, 8432 // it must be removed from the list and stashed in the closure. 8433 if (freeRangeInFreeLists()) { 8434 FreeChunk* const ffc = (FreeChunk*)freeFinger(); 8435 assert(ffc->size() == pointer_delta(fc_addr, freeFinger()), 8436 "Size of free range is inconsistent with chunk size."); 8437 if (CMSTestInFreeList) { 8438 assert(_sp->verify_chunk_in_free_list(ffc), 8439 "Chunk is not in free lists"); 8440 } 8441 _sp->coalDeath(ffc->size()); 8442 _sp->removeFreeChunkFromFreeLists(ffc); 8443 set_freeRangeInFreeLists(false); 8444 } 8445 if (fcInFreeLists) { 8446 _sp->coalDeath(chunkSize); 8447 assert(fc->size() == chunkSize, 8448 "The chunk has the wrong size or is not in the free lists"); 8449 _sp->removeFreeChunkFromFreeLists(fc); 8450 } 8451 set_lastFreeRangeCoalesced(true); 8452 print_free_block_coalesced(fc); 8453 } else { // not in a free range and/or should not coalesce 8454 // Return the current free range and start a new one. 8455 if (inFreeRange()) { 8456 // In a free range but cannot coalesce with the right hand chunk. 8457 // Put the current free range into the free lists. 8458 flush_cur_free_chunk(freeFinger(), 8459 pointer_delta(fc_addr, freeFinger())); 8460 } 8461 // Set up for new free range. Pass along whether the right hand 8462 // chunk is in the free lists. 8463 initialize_free_range((HeapWord*)fc, fcInFreeLists); 8464 } 8465 } 8466 8467 // Lookahead flush: 8468 // If we are tracking a free range, and this is the last chunk that 8469 // we'll look at because its end crosses past _limit, we'll preemptively 8470 // flush it along with any free range we may be holding on to. Note that 8471 // this can be the case only for an already free or freshly garbage 8472 // chunk. If this block is an object, it can never straddle 8473 // over _limit. The "straddling" occurs when _limit is set at 8474 // the previous end of the space when this cycle started, and 8475 // a subsequent heap expansion caused the previously co-terminal 8476 // free block to be coalesced with the newly expanded portion, 8477 // thus rendering _limit a non-block-boundary making it dangerous 8478 // for the sweeper to step over and examine. 8479 void SweepClosure::lookahead_and_flush(FreeChunk* fc, size_t chunk_size) { 8480 assert(inFreeRange(), "Should only be called if currently in a free range."); 8481 HeapWord* const eob = ((HeapWord*)fc) + chunk_size; 8482 assert(_sp->used_region().contains(eob - 1), 8483 err_msg("eob = " PTR_FORMAT " out of bounds wrt _sp = [" PTR_FORMAT "," PTR_FORMAT ")" 8484 " when examining fc = " PTR_FORMAT "(" SIZE_FORMAT ")", 8485 _limit, _sp->bottom(), _sp->end(), fc, chunk_size)); 8486 if (eob >= _limit) { 8487 assert(eob == _limit || fc->is_free(), "Only a free chunk should allow us to cross over the limit"); 8488 if (CMSTraceSweeper) { 8489 gclog_or_tty->print_cr("_limit " PTR_FORMAT " reached or crossed by block " 8490 "[" PTR_FORMAT "," PTR_FORMAT ") in space " 8491 "[" PTR_FORMAT "," PTR_FORMAT ")", 8492 _limit, fc, eob, _sp->bottom(), _sp->end()); 8493 } 8494 // Return the storage we are tracking back into the free lists. 8495 if (CMSTraceSweeper) { 8496 gclog_or_tty->print_cr("Flushing ... "); 8497 } 8498 assert(freeFinger() < eob, "Error"); 8499 flush_cur_free_chunk( freeFinger(), pointer_delta(eob, freeFinger())); 8500 } 8501 } 8502 8503 void SweepClosure::flush_cur_free_chunk(HeapWord* chunk, size_t size) { 8504 assert(inFreeRange(), "Should only be called if currently in a free range."); 8505 assert(size > 0, 8506 "A zero sized chunk cannot be added to the free lists."); 8507 if (!freeRangeInFreeLists()) { 8508 if (CMSTestInFreeList) { 8509 FreeChunk* fc = (FreeChunk*) chunk; 8510 fc->set_size(size); 8511 assert(!_sp->verify_chunk_in_free_list(fc), 8512 "chunk should not be in free lists yet"); 8513 } 8514 if (CMSTraceSweeper) { 8515 gclog_or_tty->print_cr(" -- add free block 0x%x (%d) to free lists", 8516 chunk, size); 8517 } 8518 // A new free range is going to be starting. The current 8519 // free range has not been added to the free lists yet or 8520 // was removed so add it back. 8521 // If the current free range was coalesced, then the death 8522 // of the free range was recorded. Record a birth now. 8523 if (lastFreeRangeCoalesced()) { 8524 _sp->coalBirth(size); 8525 } 8526 _sp->addChunkAndRepairOffsetTable(chunk, size, 8527 lastFreeRangeCoalesced()); 8528 } else if (CMSTraceSweeper) { 8529 gclog_or_tty->print_cr("Already in free list: nothing to flush"); 8530 } 8531 set_inFreeRange(false); 8532 set_freeRangeInFreeLists(false); 8533 } 8534 8535 // We take a break if we've been at this for a while, 8536 // so as to avoid monopolizing the locks involved. 8537 void SweepClosure::do_yield_work(HeapWord* addr) { 8538 // Return current free chunk being used for coalescing (if any) 8539 // to the appropriate freelist. After yielding, the next 8540 // free block encountered will start a coalescing range of 8541 // free blocks. If the next free block is adjacent to the 8542 // chunk just flushed, they will need to wait for the next 8543 // sweep to be coalesced. 8544 if (inFreeRange()) { 8545 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 8546 } 8547 8548 // First give up the locks, then yield, then re-lock. 8549 // We should probably use a constructor/destructor idiom to 8550 // do this unlock/lock or modify the MutexUnlocker class to 8551 // serve our purpose. XXX 8552 assert_lock_strong(_bitMap->lock()); 8553 assert_lock_strong(_freelistLock); 8554 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 8555 "CMS thread should hold CMS token"); 8556 _bitMap->lock()->unlock(); 8557 _freelistLock->unlock(); 8558 ConcurrentMarkSweepThread::desynchronize(true); 8559 ConcurrentMarkSweepThread::acknowledge_yield_request(); 8560 _collector->stopTimer(); 8561 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 8562 if (PrintCMSStatistics != 0) { 8563 _collector->incrementYields(); 8564 } 8565 _collector->icms_wait(); 8566 8567 // See the comment in coordinator_yield() 8568 for (unsigned i = 0; i < CMSYieldSleepCount && 8569 ConcurrentMarkSweepThread::should_yield() && 8570 !CMSCollector::foregroundGCIsActive(); ++i) { 8571 os::sleep(Thread::current(), 1, false); 8572 ConcurrentMarkSweepThread::acknowledge_yield_request(); 8573 } 8574 8575 ConcurrentMarkSweepThread::synchronize(true); 8576 _freelistLock->lock(); 8577 _bitMap->lock()->lock_without_safepoint_check(); 8578 _collector->startTimer(); 8579 } 8580 8581 #ifndef PRODUCT 8582 // This is actually very useful in a product build if it can 8583 // be called from the debugger. Compile it into the product 8584 // as needed. 8585 bool debug_verify_chunk_in_free_list(FreeChunk* fc) { 8586 return debug_cms_space->verify_chunk_in_free_list(fc); 8587 } 8588 #endif 8589 8590 void SweepClosure::print_free_block_coalesced(FreeChunk* fc) const { 8591 if (CMSTraceSweeper) { 8592 gclog_or_tty->print_cr("Sweep:coal_free_blk " PTR_FORMAT " (" SIZE_FORMAT ")", 8593 fc, fc->size()); 8594 } 8595 } 8596 8597 // CMSIsAliveClosure 8598 bool CMSIsAliveClosure::do_object_b(oop obj) { 8599 HeapWord* addr = (HeapWord*)obj; 8600 return addr != NULL && 8601 (!_span.contains(addr) || _bit_map->isMarked(addr)); 8602 } 8603 8604 CMSKeepAliveClosure::CMSKeepAliveClosure( CMSCollector* collector, 8605 MemRegion span, 8606 CMSBitMap* bit_map, CMSMarkStack* mark_stack, 8607 CMSMarkStack* revisit_stack, bool cpc): 8608 KlassRememberingOopClosure(collector, NULL, revisit_stack), 8609 _span(span), 8610 _bit_map(bit_map), 8611 _mark_stack(mark_stack), 8612 _concurrent_precleaning(cpc) { 8613 assert(!_span.is_empty(), "Empty span could spell trouble"); 8614 } 8615 8616 8617 // CMSKeepAliveClosure: the serial version 8618 void CMSKeepAliveClosure::do_oop(oop obj) { 8619 HeapWord* addr = (HeapWord*)obj; 8620 if (_span.contains(addr) && 8621 !_bit_map->isMarked(addr)) { 8622 _bit_map->mark(addr); 8623 bool simulate_overflow = false; 8624 NOT_PRODUCT( 8625 if (CMSMarkStackOverflowALot && 8626 _collector->simulate_overflow()) { 8627 // simulate a stack overflow 8628 simulate_overflow = true; 8629 } 8630 ) 8631 if (simulate_overflow || !_mark_stack->push(obj)) { 8632 if (_concurrent_precleaning) { 8633 // We dirty the overflown object and let the remark 8634 // phase deal with it. 8635 assert(_collector->overflow_list_is_empty(), "Error"); 8636 // In the case of object arrays, we need to dirty all of 8637 // the cards that the object spans. No locking or atomics 8638 // are needed since no one else can be mutating the mod union 8639 // table. 8640 if (obj->is_objArray()) { 8641 size_t sz = obj->size(); 8642 HeapWord* end_card_addr = 8643 (HeapWord*)round_to((intptr_t)(addr+sz), CardTableModRefBS::card_size); 8644 MemRegion redirty_range = MemRegion(addr, end_card_addr); 8645 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 8646 _collector->_modUnionTable.mark_range(redirty_range); 8647 } else { 8648 _collector->_modUnionTable.mark(addr); 8649 } 8650 _collector->_ser_kac_preclean_ovflw++; 8651 } else { 8652 _collector->push_on_overflow_list(obj); 8653 _collector->_ser_kac_ovflw++; 8654 } 8655 } 8656 } 8657 } 8658 8659 void CMSKeepAliveClosure::do_oop(oop* p) { CMSKeepAliveClosure::do_oop_work(p); } 8660 void CMSKeepAliveClosure::do_oop(narrowOop* p) { CMSKeepAliveClosure::do_oop_work(p); } 8661 8662 // CMSParKeepAliveClosure: a parallel version of the above. 8663 // The work queues are private to each closure (thread), 8664 // but (may be) available for stealing by other threads. 8665 void CMSParKeepAliveClosure::do_oop(oop obj) { 8666 HeapWord* addr = (HeapWord*)obj; 8667 if (_span.contains(addr) && 8668 !_bit_map->isMarked(addr)) { 8669 // In general, during recursive tracing, several threads 8670 // may be concurrently getting here; the first one to 8671 // "tag" it, claims it. 8672 if (_bit_map->par_mark(addr)) { 8673 bool res = _work_queue->push(obj); 8674 assert(res, "Low water mark should be much less than capacity"); 8675 // Do a recursive trim in the hope that this will keep 8676 // stack usage lower, but leave some oops for potential stealers 8677 trim_queue(_low_water_mark); 8678 } // Else, another thread got there first 8679 } 8680 } 8681 8682 void CMSParKeepAliveClosure::do_oop(oop* p) { CMSParKeepAliveClosure::do_oop_work(p); } 8683 void CMSParKeepAliveClosure::do_oop(narrowOop* p) { CMSParKeepAliveClosure::do_oop_work(p); } 8684 8685 void CMSParKeepAliveClosure::trim_queue(uint max) { 8686 while (_work_queue->size() > max) { 8687 oop new_oop; 8688 if (_work_queue->pop_local(new_oop)) { 8689 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 8690 assert(_bit_map->isMarked((HeapWord*)new_oop), 8691 "no white objects on this stack!"); 8692 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 8693 // iterate over the oops in this oop, marking and pushing 8694 // the ones in CMS heap (i.e. in _span). 8695 new_oop->oop_iterate(&_mark_and_push); 8696 } 8697 } 8698 } 8699 8700 CMSInnerParMarkAndPushClosure::CMSInnerParMarkAndPushClosure( 8701 CMSCollector* collector, 8702 MemRegion span, CMSBitMap* bit_map, 8703 CMSMarkStack* revisit_stack, 8704 OopTaskQueue* work_queue): 8705 Par_KlassRememberingOopClosure(collector, NULL, revisit_stack), 8706 _span(span), 8707 _bit_map(bit_map), 8708 _work_queue(work_queue) { } 8709 8710 void CMSInnerParMarkAndPushClosure::do_oop(oop obj) { 8711 HeapWord* addr = (HeapWord*)obj; 8712 if (_span.contains(addr) && 8713 !_bit_map->isMarked(addr)) { 8714 if (_bit_map->par_mark(addr)) { 8715 bool simulate_overflow = false; 8716 NOT_PRODUCT( 8717 if (CMSMarkStackOverflowALot && 8718 _collector->par_simulate_overflow()) { 8719 // simulate a stack overflow 8720 simulate_overflow = true; 8721 } 8722 ) 8723 if (simulate_overflow || !_work_queue->push(obj)) { 8724 _collector->par_push_on_overflow_list(obj); 8725 _collector->_par_kac_ovflw++; 8726 } 8727 } // Else another thread got there already 8728 } 8729 } 8730 8731 void CMSInnerParMarkAndPushClosure::do_oop(oop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); } 8732 void CMSInnerParMarkAndPushClosure::do_oop(narrowOop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); } 8733 8734 ////////////////////////////////////////////////////////////////// 8735 // CMSExpansionCause ///////////////////////////// 8736 ////////////////////////////////////////////////////////////////// 8737 const char* CMSExpansionCause::to_string(CMSExpansionCause::Cause cause) { 8738 switch (cause) { 8739 case _no_expansion: 8740 return "No expansion"; 8741 case _satisfy_free_ratio: 8742 return "Free ratio"; 8743 case _satisfy_promotion: 8744 return "Satisfy promotion"; 8745 case _satisfy_allocation: 8746 return "allocation"; 8747 case _allocate_par_lab: 8748 return "Par LAB"; 8749 case _allocate_par_spooling_space: 8750 return "Par Spooling Space"; 8751 case _adaptive_size_policy: 8752 return "Ergonomics"; 8753 default: 8754 return "unknown"; 8755 } 8756 } 8757 8758 void CMSDrainMarkingStackClosure::do_void() { 8759 // the max number to take from overflow list at a time 8760 const size_t num = _mark_stack->capacity()/4; 8761 assert(!_concurrent_precleaning || _collector->overflow_list_is_empty(), 8762 "Overflow list should be NULL during concurrent phases"); 8763 while (!_mark_stack->isEmpty() || 8764 // if stack is empty, check the overflow list 8765 _collector->take_from_overflow_list(num, _mark_stack)) { 8766 oop obj = _mark_stack->pop(); 8767 HeapWord* addr = (HeapWord*)obj; 8768 assert(_span.contains(addr), "Should be within span"); 8769 assert(_bit_map->isMarked(addr), "Should be marked"); 8770 assert(obj->is_oop(), "Should be an oop"); 8771 obj->oop_iterate(_keep_alive); 8772 } 8773 } 8774 8775 void CMSParDrainMarkingStackClosure::do_void() { 8776 // drain queue 8777 trim_queue(0); 8778 } 8779 8780 // Trim our work_queue so its length is below max at return 8781 void CMSParDrainMarkingStackClosure::trim_queue(uint max) { 8782 while (_work_queue->size() > max) { 8783 oop new_oop; 8784 if (_work_queue->pop_local(new_oop)) { 8785 assert(new_oop->is_oop(), "Expected an oop"); 8786 assert(_bit_map->isMarked((HeapWord*)new_oop), 8787 "no white objects on this stack!"); 8788 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 8789 // iterate over the oops in this oop, marking and pushing 8790 // the ones in CMS heap (i.e. in _span). 8791 new_oop->oop_iterate(&_mark_and_push); 8792 } 8793 } 8794 } 8795 8796 //////////////////////////////////////////////////////////////////// 8797 // Support for Marking Stack Overflow list handling and related code 8798 //////////////////////////////////////////////////////////////////// 8799 // Much of the following code is similar in shape and spirit to the 8800 // code used in ParNewGC. We should try and share that code 8801 // as much as possible in the future. 8802 8803 #ifndef PRODUCT 8804 // Debugging support for CMSStackOverflowALot 8805 8806 // It's OK to call this multi-threaded; the worst thing 8807 // that can happen is that we'll get a bunch of closely 8808 // spaced simulated oveflows, but that's OK, in fact 8809 // probably good as it would exercise the overflow code 8810 // under contention. 8811 bool CMSCollector::simulate_overflow() { 8812 if (_overflow_counter-- <= 0) { // just being defensive 8813 _overflow_counter = CMSMarkStackOverflowInterval; 8814 return true; 8815 } else { 8816 return false; 8817 } 8818 } 8819 8820 bool CMSCollector::par_simulate_overflow() { 8821 return simulate_overflow(); 8822 } 8823 #endif 8824 8825 // Single-threaded 8826 bool CMSCollector::take_from_overflow_list(size_t num, CMSMarkStack* stack) { 8827 assert(stack->isEmpty(), "Expected precondition"); 8828 assert(stack->capacity() > num, "Shouldn't bite more than can chew"); 8829 size_t i = num; 8830 oop cur = _overflow_list; 8831 const markOop proto = markOopDesc::prototype(); 8832 NOT_PRODUCT(ssize_t n = 0;) 8833 for (oop next; i > 0 && cur != NULL; cur = next, i--) { 8834 next = oop(cur->mark()); 8835 cur->set_mark(proto); // until proven otherwise 8836 assert(cur->is_oop(), "Should be an oop"); 8837 bool res = stack->push(cur); 8838 assert(res, "Bit off more than can chew?"); 8839 NOT_PRODUCT(n++;) 8840 } 8841 _overflow_list = cur; 8842 #ifndef PRODUCT 8843 assert(_num_par_pushes >= n, "Too many pops?"); 8844 _num_par_pushes -=n; 8845 #endif 8846 return !stack->isEmpty(); 8847 } 8848 8849 #define BUSY (oop(0x1aff1aff)) 8850 // (MT-safe) Get a prefix of at most "num" from the list. 8851 // The overflow list is chained through the mark word of 8852 // each object in the list. We fetch the entire list, 8853 // break off a prefix of the right size and return the 8854 // remainder. If other threads try to take objects from 8855 // the overflow list at that time, they will wait for 8856 // some time to see if data becomes available. If (and 8857 // only if) another thread places one or more object(s) 8858 // on the global list before we have returned the suffix 8859 // to the global list, we will walk down our local list 8860 // to find its end and append the global list to 8861 // our suffix before returning it. This suffix walk can 8862 // prove to be expensive (quadratic in the amount of traffic) 8863 // when there are many objects in the overflow list and 8864 // there is much producer-consumer contention on the list. 8865 // *NOTE*: The overflow list manipulation code here and 8866 // in ParNewGeneration:: are very similar in shape, 8867 // except that in the ParNew case we use the old (from/eden) 8868 // copy of the object to thread the list via its klass word. 8869 // Because of the common code, if you make any changes in 8870 // the code below, please check the ParNew version to see if 8871 // similar changes might be needed. 8872 // CR 6797058 has been filed to consolidate the common code. 8873 bool CMSCollector::par_take_from_overflow_list(size_t num, 8874 OopTaskQueue* work_q, 8875 int no_of_gc_threads) { 8876 assert(work_q->size() == 0, "First empty local work queue"); 8877 assert(num < work_q->max_elems(), "Can't bite more than we can chew"); 8878 if (_overflow_list == NULL) { 8879 return false; 8880 } 8881 // Grab the entire list; we'll put back a suffix 8882 oop prefix = (oop)Atomic::xchg_ptr(BUSY, &_overflow_list); 8883 Thread* tid = Thread::current(); 8884 // Before "no_of_gc_threads" was introduced CMSOverflowSpinCount was 8885 // set to ParallelGCThreads. 8886 size_t CMSOverflowSpinCount = (size_t) no_of_gc_threads; // was ParallelGCThreads; 8887 size_t sleep_time_millis = MAX2((size_t)1, num/100); 8888 // If the list is busy, we spin for a short while, 8889 // sleeping between attempts to get the list. 8890 for (size_t spin = 0; prefix == BUSY && spin < CMSOverflowSpinCount; spin++) { 8891 os::sleep(tid, sleep_time_millis, false); 8892 if (_overflow_list == NULL) { 8893 // Nothing left to take 8894 return false; 8895 } else if (_overflow_list != BUSY) { 8896 // Try and grab the prefix 8897 prefix = (oop)Atomic::xchg_ptr(BUSY, &_overflow_list); 8898 } 8899 } 8900 // If the list was found to be empty, or we spun long 8901 // enough, we give up and return empty-handed. If we leave 8902 // the list in the BUSY state below, it must be the case that 8903 // some other thread holds the overflow list and will set it 8904 // to a non-BUSY state in the future. 8905 if (prefix == NULL || prefix == BUSY) { 8906 // Nothing to take or waited long enough 8907 if (prefix == NULL) { 8908 // Write back the NULL in case we overwrote it with BUSY above 8909 // and it is still the same value. 8910 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY); 8911 } 8912 return false; 8913 } 8914 assert(prefix != NULL && prefix != BUSY, "Error"); 8915 size_t i = num; 8916 oop cur = prefix; 8917 // Walk down the first "num" objects, unless we reach the end. 8918 for (; i > 1 && cur->mark() != NULL; cur = oop(cur->mark()), i--); 8919 if (cur->mark() == NULL) { 8920 // We have "num" or fewer elements in the list, so there 8921 // is nothing to return to the global list. 8922 // Write back the NULL in lieu of the BUSY we wrote 8923 // above, if it is still the same value. 8924 if (_overflow_list == BUSY) { 8925 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY); 8926 } 8927 } else { 8928 // Chop off the suffix and rerturn it to the global list. 8929 assert(cur->mark() != BUSY, "Error"); 8930 oop suffix_head = cur->mark(); // suffix will be put back on global list 8931 cur->set_mark(NULL); // break off suffix 8932 // It's possible that the list is still in the empty(busy) state 8933 // we left it in a short while ago; in that case we may be 8934 // able to place back the suffix without incurring the cost 8935 // of a walk down the list. 8936 oop observed_overflow_list = _overflow_list; 8937 oop cur_overflow_list = observed_overflow_list; 8938 bool attached = false; 8939 while (observed_overflow_list == BUSY || observed_overflow_list == NULL) { 8940 observed_overflow_list = 8941 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list); 8942 if (cur_overflow_list == observed_overflow_list) { 8943 attached = true; 8944 break; 8945 } else cur_overflow_list = observed_overflow_list; 8946 } 8947 if (!attached) { 8948 // Too bad, someone else sneaked in (at least) an element; we'll need 8949 // to do a splice. Find tail of suffix so we can prepend suffix to global 8950 // list. 8951 for (cur = suffix_head; cur->mark() != NULL; cur = (oop)(cur->mark())); 8952 oop suffix_tail = cur; 8953 assert(suffix_tail != NULL && suffix_tail->mark() == NULL, 8954 "Tautology"); 8955 observed_overflow_list = _overflow_list; 8956 do { 8957 cur_overflow_list = observed_overflow_list; 8958 if (cur_overflow_list != BUSY) { 8959 // Do the splice ... 8960 suffix_tail->set_mark(markOop(cur_overflow_list)); 8961 } else { // cur_overflow_list == BUSY 8962 suffix_tail->set_mark(NULL); 8963 } 8964 // ... and try to place spliced list back on overflow_list ... 8965 observed_overflow_list = 8966 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list); 8967 } while (cur_overflow_list != observed_overflow_list); 8968 // ... until we have succeeded in doing so. 8969 } 8970 } 8971 8972 // Push the prefix elements on work_q 8973 assert(prefix != NULL, "control point invariant"); 8974 const markOop proto = markOopDesc::prototype(); 8975 oop next; 8976 NOT_PRODUCT(ssize_t n = 0;) 8977 for (cur = prefix; cur != NULL; cur = next) { 8978 next = oop(cur->mark()); 8979 cur->set_mark(proto); // until proven otherwise 8980 assert(cur->is_oop(), "Should be an oop"); 8981 bool res = work_q->push(cur); 8982 assert(res, "Bit off more than we can chew?"); 8983 NOT_PRODUCT(n++;) 8984 } 8985 #ifndef PRODUCT 8986 assert(_num_par_pushes >= n, "Too many pops?"); 8987 Atomic::add_ptr(-(intptr_t)n, &_num_par_pushes); 8988 #endif 8989 return true; 8990 } 8991 8992 // Single-threaded 8993 void CMSCollector::push_on_overflow_list(oop p) { 8994 NOT_PRODUCT(_num_par_pushes++;) 8995 assert(p->is_oop(), "Not an oop"); 8996 preserve_mark_if_necessary(p); 8997 p->set_mark((markOop)_overflow_list); 8998 _overflow_list = p; 8999 } 9000 9001 // Multi-threaded; use CAS to prepend to overflow list 9002 void CMSCollector::par_push_on_overflow_list(oop p) { 9003 NOT_PRODUCT(Atomic::inc_ptr(&_num_par_pushes);) 9004 assert(p->is_oop(), "Not an oop"); 9005 par_preserve_mark_if_necessary(p); 9006 oop observed_overflow_list = _overflow_list; 9007 oop cur_overflow_list; 9008 do { 9009 cur_overflow_list = observed_overflow_list; 9010 if (cur_overflow_list != BUSY) { 9011 p->set_mark(markOop(cur_overflow_list)); 9012 } else { 9013 p->set_mark(NULL); 9014 } 9015 observed_overflow_list = 9016 (oop) Atomic::cmpxchg_ptr(p, &_overflow_list, cur_overflow_list); 9017 } while (cur_overflow_list != observed_overflow_list); 9018 } 9019 #undef BUSY 9020 9021 // Single threaded 9022 // General Note on GrowableArray: pushes may silently fail 9023 // because we are (temporarily) out of C-heap for expanding 9024 // the stack. The problem is quite ubiquitous and affects 9025 // a lot of code in the JVM. The prudent thing for GrowableArray 9026 // to do (for now) is to exit with an error. However, that may 9027 // be too draconian in some cases because the caller may be 9028 // able to recover without much harm. For such cases, we 9029 // should probably introduce a "soft_push" method which returns 9030 // an indication of success or failure with the assumption that 9031 // the caller may be able to recover from a failure; code in 9032 // the VM can then be changed, incrementally, to deal with such 9033 // failures where possible, thus, incrementally hardening the VM 9034 // in such low resource situations. 9035 void CMSCollector::preserve_mark_work(oop p, markOop m) { 9036 _preserved_oop_stack.push(p); 9037 _preserved_mark_stack.push(m); 9038 assert(m == p->mark(), "Mark word changed"); 9039 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), 9040 "bijection"); 9041 } 9042 9043 // Single threaded 9044 void CMSCollector::preserve_mark_if_necessary(oop p) { 9045 markOop m = p->mark(); 9046 if (m->must_be_preserved(p)) { 9047 preserve_mark_work(p, m); 9048 } 9049 } 9050 9051 void CMSCollector::par_preserve_mark_if_necessary(oop p) { 9052 markOop m = p->mark(); 9053 if (m->must_be_preserved(p)) { 9054 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag); 9055 // Even though we read the mark word without holding 9056 // the lock, we are assured that it will not change 9057 // because we "own" this oop, so no other thread can 9058 // be trying to push it on the overflow list; see 9059 // the assertion in preserve_mark_work() that checks 9060 // that m == p->mark(). 9061 preserve_mark_work(p, m); 9062 } 9063 } 9064 9065 // We should be able to do this multi-threaded, 9066 // a chunk of stack being a task (this is 9067 // correct because each oop only ever appears 9068 // once in the overflow list. However, it's 9069 // not very easy to completely overlap this with 9070 // other operations, so will generally not be done 9071 // until all work's been completed. Because we 9072 // expect the preserved oop stack (set) to be small, 9073 // it's probably fine to do this single-threaded. 9074 // We can explore cleverer concurrent/overlapped/parallel 9075 // processing of preserved marks if we feel the 9076 // need for this in the future. Stack overflow should 9077 // be so rare in practice and, when it happens, its 9078 // effect on performance so great that this will 9079 // likely just be in the noise anyway. 9080 void CMSCollector::restore_preserved_marks_if_any() { 9081 assert(SafepointSynchronize::is_at_safepoint(), 9082 "world should be stopped"); 9083 assert(Thread::current()->is_ConcurrentGC_thread() || 9084 Thread::current()->is_VM_thread(), 9085 "should be single-threaded"); 9086 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), 9087 "bijection"); 9088 9089 while (!_preserved_oop_stack.is_empty()) { 9090 oop p = _preserved_oop_stack.pop(); 9091 assert(p->is_oop(), "Should be an oop"); 9092 assert(_span.contains(p), "oop should be in _span"); 9093 assert(p->mark() == markOopDesc::prototype(), 9094 "Set when taken from overflow list"); 9095 markOop m = _preserved_mark_stack.pop(); 9096 p->set_mark(m); 9097 } 9098 assert(_preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(), 9099 "stacks were cleared above"); 9100 } 9101 9102 #ifndef PRODUCT 9103 bool CMSCollector::no_preserved_marks() const { 9104 return _preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(); 9105 } 9106 #endif 9107 9108 CMSAdaptiveSizePolicy* ASConcurrentMarkSweepGeneration::cms_size_policy() const 9109 { 9110 GenCollectedHeap* gch = (GenCollectedHeap*) GenCollectedHeap::heap(); 9111 CMSAdaptiveSizePolicy* size_policy = 9112 (CMSAdaptiveSizePolicy*) gch->gen_policy()->size_policy(); 9113 assert(size_policy->is_gc_cms_adaptive_size_policy(), 9114 "Wrong type for size policy"); 9115 return size_policy; 9116 } 9117 9118 void ASConcurrentMarkSweepGeneration::resize(size_t cur_promo_size, 9119 size_t desired_promo_size) { 9120 if (cur_promo_size < desired_promo_size) { 9121 size_t expand_bytes = desired_promo_size - cur_promo_size; 9122 if (PrintAdaptiveSizePolicy && Verbose) { 9123 gclog_or_tty->print_cr(" ASConcurrentMarkSweepGeneration::resize " 9124 "Expanding tenured generation by " SIZE_FORMAT " (bytes)", 9125 expand_bytes); 9126 } 9127 expand(expand_bytes, 9128 MinHeapDeltaBytes, 9129 CMSExpansionCause::_adaptive_size_policy); 9130 } else if (desired_promo_size < cur_promo_size) { 9131 size_t shrink_bytes = cur_promo_size - desired_promo_size; 9132 if (PrintAdaptiveSizePolicy && Verbose) { 9133 gclog_or_tty->print_cr(" ASConcurrentMarkSweepGeneration::resize " 9134 "Shrinking tenured generation by " SIZE_FORMAT " (bytes)", 9135 shrink_bytes); 9136 } 9137 shrink(shrink_bytes); 9138 } 9139 } 9140 9141 CMSGCAdaptivePolicyCounters* ASConcurrentMarkSweepGeneration::gc_adaptive_policy_counters() { 9142 GenCollectedHeap* gch = GenCollectedHeap::heap(); 9143 CMSGCAdaptivePolicyCounters* counters = 9144 (CMSGCAdaptivePolicyCounters*) gch->collector_policy()->counters(); 9145 assert(counters->kind() == GCPolicyCounters::CMSGCAdaptivePolicyCountersKind, 9146 "Wrong kind of counters"); 9147 return counters; 9148 } 9149 9150 9151 void ASConcurrentMarkSweepGeneration::update_counters() { 9152 if (UsePerfData) { 9153 _space_counters->update_all(); 9154 _gen_counters->update_all(); 9155 CMSGCAdaptivePolicyCounters* counters = gc_adaptive_policy_counters(); 9156 GenCollectedHeap* gch = GenCollectedHeap::heap(); 9157 CMSGCStats* gc_stats_l = (CMSGCStats*) gc_stats(); 9158 assert(gc_stats_l->kind() == GCStats::CMSGCStatsKind, 9159 "Wrong gc statistics type"); 9160 counters->update_counters(gc_stats_l); 9161 } 9162 } 9163 9164 void ASConcurrentMarkSweepGeneration::update_counters(size_t used) { 9165 if (UsePerfData) { 9166 _space_counters->update_used(used); 9167 _space_counters->update_capacity(); 9168 _gen_counters->update_all(); 9169 9170 CMSGCAdaptivePolicyCounters* counters = gc_adaptive_policy_counters(); 9171 GenCollectedHeap* gch = GenCollectedHeap::heap(); 9172 CMSGCStats* gc_stats_l = (CMSGCStats*) gc_stats(); 9173 assert(gc_stats_l->kind() == GCStats::CMSGCStatsKind, 9174 "Wrong gc statistics type"); 9175 counters->update_counters(gc_stats_l); 9176 } 9177 } 9178 9179 // The desired expansion delta is computed so that: 9180 // . desired free percentage or greater is used 9181 void ASConcurrentMarkSweepGeneration::compute_new_size() { 9182 assert_locked_or_safepoint(Heap_lock); 9183 9184 GenCollectedHeap* gch = (GenCollectedHeap*) GenCollectedHeap::heap(); 9185 9186 // If incremental collection failed, we just want to expand 9187 // to the limit. 9188 if (incremental_collection_failed()) { 9189 clear_incremental_collection_failed(); 9190 grow_to_reserved(); 9191 return; 9192 } 9193 9194 assert(UseAdaptiveSizePolicy, "Should be using adaptive sizing"); 9195 9196 assert(gch->kind() == CollectedHeap::GenCollectedHeap, 9197 "Wrong type of heap"); 9198 int prev_level = level() - 1; 9199 assert(prev_level >= 0, "The cms generation is the lowest generation"); 9200 Generation* prev_gen = gch->get_gen(prev_level); 9201 assert(prev_gen->kind() == Generation::ASParNew, 9202 "Wrong type of young generation"); 9203 ParNewGeneration* younger_gen = (ParNewGeneration*) prev_gen; 9204 size_t cur_eden = younger_gen->eden()->capacity(); 9205 CMSAdaptiveSizePolicy* size_policy = cms_size_policy(); 9206 size_t cur_promo = free(); 9207 size_policy->compute_tenured_generation_free_space(cur_promo, 9208 max_available(), 9209 cur_eden); 9210 resize(cur_promo, size_policy->promo_size()); 9211 9212 // Record the new size of the space in the cms generation 9213 // that is available for promotions. This is temporary. 9214 // It should be the desired promo size. 9215 size_policy->avg_cms_promo()->sample(free()); 9216 size_policy->avg_old_live()->sample(used()); 9217 9218 if (UsePerfData) { 9219 CMSGCAdaptivePolicyCounters* counters = gc_adaptive_policy_counters(); 9220 counters->update_cms_capacity_counter(capacity()); 9221 } 9222 } 9223 9224 void ASConcurrentMarkSweepGeneration::shrink_by(size_t desired_bytes) { 9225 assert_locked_or_safepoint(Heap_lock); 9226 assert_lock_strong(freelistLock()); 9227 HeapWord* old_end = _cmsSpace->end(); 9228 HeapWord* unallocated_start = _cmsSpace->unallocated_block(); 9229 assert(old_end >= unallocated_start, "Miscalculation of unallocated_start"); 9230 FreeChunk* chunk_at_end = find_chunk_at_end(); 9231 if (chunk_at_end == NULL) { 9232 // No room to shrink 9233 if (PrintGCDetails && Verbose) { 9234 gclog_or_tty->print_cr("No room to shrink: old_end " 9235 PTR_FORMAT " unallocated_start " PTR_FORMAT 9236 " chunk_at_end " PTR_FORMAT, 9237 old_end, unallocated_start, chunk_at_end); 9238 } 9239 return; 9240 } else { 9241 9242 // Find the chunk at the end of the space and determine 9243 // how much it can be shrunk. 9244 size_t shrinkable_size_in_bytes = chunk_at_end->size(); 9245 size_t aligned_shrinkable_size_in_bytes = 9246 align_size_down(shrinkable_size_in_bytes, os::vm_page_size()); 9247 assert(unallocated_start <= chunk_at_end->end(), 9248 "Inconsistent chunk at end of space"); 9249 size_t bytes = MIN2(desired_bytes, aligned_shrinkable_size_in_bytes); 9250 size_t word_size_before = heap_word_size(_virtual_space.committed_size()); 9251 9252 // Shrink the underlying space 9253 _virtual_space.shrink_by(bytes); 9254 if (PrintGCDetails && Verbose) { 9255 gclog_or_tty->print_cr("ConcurrentMarkSweepGeneration::shrink_by:" 9256 " desired_bytes " SIZE_FORMAT 9257 " shrinkable_size_in_bytes " SIZE_FORMAT 9258 " aligned_shrinkable_size_in_bytes " SIZE_FORMAT 9259 " bytes " SIZE_FORMAT, 9260 desired_bytes, shrinkable_size_in_bytes, 9261 aligned_shrinkable_size_in_bytes, bytes); 9262 gclog_or_tty->print_cr(" old_end " SIZE_FORMAT 9263 " unallocated_start " SIZE_FORMAT, 9264 old_end, unallocated_start); 9265 } 9266 9267 // If the space did shrink (shrinking is not guaranteed), 9268 // shrink the chunk at the end by the appropriate amount. 9269 if (((HeapWord*)_virtual_space.high()) < old_end) { 9270 size_t new_word_size = 9271 heap_word_size(_virtual_space.committed_size()); 9272 9273 // Have to remove the chunk from the dictionary because it is changing 9274 // size and might be someplace elsewhere in the dictionary. 9275 9276 // Get the chunk at end, shrink it, and put it 9277 // back. 9278 _cmsSpace->removeChunkFromDictionary(chunk_at_end); 9279 size_t word_size_change = word_size_before - new_word_size; 9280 size_t chunk_at_end_old_size = chunk_at_end->size(); 9281 assert(chunk_at_end_old_size >= word_size_change, 9282 "Shrink is too large"); 9283 chunk_at_end->set_size(chunk_at_end_old_size - 9284 word_size_change); 9285 _cmsSpace->freed((HeapWord*) chunk_at_end->end(), 9286 word_size_change); 9287 9288 _cmsSpace->returnChunkToDictionary(chunk_at_end); 9289 9290 MemRegion mr(_cmsSpace->bottom(), new_word_size); 9291 _bts->resize(new_word_size); // resize the block offset shared array 9292 Universe::heap()->barrier_set()->resize_covered_region(mr); 9293 _cmsSpace->assert_locked(); 9294 _cmsSpace->set_end((HeapWord*)_virtual_space.high()); 9295 9296 NOT_PRODUCT(_cmsSpace->dictionary()->verify()); 9297 9298 // update the space and generation capacity counters 9299 if (UsePerfData) { 9300 _space_counters->update_capacity(); 9301 _gen_counters->update_all(); 9302 } 9303 9304 if (Verbose && PrintGCDetails) { 9305 size_t new_mem_size = _virtual_space.committed_size(); 9306 size_t old_mem_size = new_mem_size + bytes; 9307 gclog_or_tty->print_cr("Shrinking %s from %ldK by %ldK to %ldK", 9308 name(), old_mem_size/K, bytes/K, new_mem_size/K); 9309 } 9310 } 9311 9312 assert(_cmsSpace->unallocated_block() <= _cmsSpace->end(), 9313 "Inconsistency at end of space"); 9314 assert(chunk_at_end->end() == _cmsSpace->end(), 9315 "Shrinking is inconsistent"); 9316 return; 9317 } 9318 } 9319 9320 // Transfer some number of overflown objects to usual marking 9321 // stack. Return true if some objects were transferred. 9322 bool MarkRefsIntoAndScanClosure::take_from_overflow_list() { 9323 size_t num = MIN2((size_t)(_mark_stack->capacity() - _mark_stack->length())/4, 9324 (size_t)ParGCDesiredObjsFromOverflowList); 9325 9326 bool res = _collector->take_from_overflow_list(num, _mark_stack); 9327 assert(_collector->overflow_list_is_empty() || res, 9328 "If list is not empty, we should have taken something"); 9329 assert(!res || !_mark_stack->isEmpty(), 9330 "If we took something, it should now be on our stack"); 9331 return res; 9332 } 9333 9334 size_t MarkDeadObjectsClosure::do_blk(HeapWord* addr) { 9335 size_t res = _sp->block_size_no_stall(addr, _collector); 9336 if (_sp->block_is_obj(addr)) { 9337 if (_live_bit_map->isMarked(addr)) { 9338 // It can't have been dead in a previous cycle 9339 guarantee(!_dead_bit_map->isMarked(addr), "No resurrection!"); 9340 } else { 9341 _dead_bit_map->mark(addr); // mark the dead object 9342 } 9343 } 9344 // Could be 0, if the block size could not be computed without stalling. 9345 return res; 9346 } 9347 9348 TraceCMSMemoryManagerStats::TraceCMSMemoryManagerStats(CMSCollector::CollectorState phase, GCCause::Cause cause): TraceMemoryManagerStats() { 9349 9350 switch (phase) { 9351 case CMSCollector::InitialMarking: 9352 initialize(true /* fullGC */ , 9353 cause /* cause of the GC */, 9354 true /* recordGCBeginTime */, 9355 true /* recordPreGCUsage */, 9356 false /* recordPeakUsage */, 9357 false /* recordPostGCusage */, 9358 true /* recordAccumulatedGCTime */, 9359 false /* recordGCEndTime */, 9360 false /* countCollection */ ); 9361 break; 9362 9363 case CMSCollector::FinalMarking: 9364 initialize(true /* fullGC */ , 9365 cause /* cause of the GC */, 9366 false /* recordGCBeginTime */, 9367 false /* recordPreGCUsage */, 9368 false /* recordPeakUsage */, 9369 false /* recordPostGCusage */, 9370 true /* recordAccumulatedGCTime */, 9371 false /* recordGCEndTime */, 9372 false /* countCollection */ ); 9373 break; 9374 9375 case CMSCollector::Sweeping: 9376 initialize(true /* fullGC */ , 9377 cause /* cause of the GC */, 9378 false /* recordGCBeginTime */, 9379 false /* recordPreGCUsage */, 9380 true /* recordPeakUsage */, 9381 true /* recordPostGCusage */, 9382 false /* recordAccumulatedGCTime */, 9383 true /* recordGCEndTime */, 9384 true /* countCollection */ ); 9385 break; 9386 9387 default: 9388 ShouldNotReachHere(); 9389 } 9390 } 9391