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