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 #ifdef SPARC 793 // Issue a stern warning, but allow use for experimentation and debugging. 794 if (VM_Version::is_sun4v() && UseMemSetInBOT) { 795 assert(!FLAG_IS_DEFAULT(UseMemSetInBOT), "Error"); 796 warning("Experimental flag -XX:+UseMemSetInBOT is known to cause instability" 797 " on sun4v; please understand that you are using at your own risk!"); 798 } 799 #endif 800 } 801 802 const char* ConcurrentMarkSweepGeneration::name() const { 803 return "concurrent mark-sweep generation"; 804 } 805 void ConcurrentMarkSweepGeneration::update_counters() { 806 if (UsePerfData) { 807 _space_counters->update_all(); 808 _gen_counters->update_all(); 809 } 810 } 811 812 // this is an optimized version of update_counters(). it takes the 813 // used value as a parameter rather than computing it. 814 // 815 void ConcurrentMarkSweepGeneration::update_counters(size_t used) { 816 if (UsePerfData) { 817 _space_counters->update_used(used); 818 _space_counters->update_capacity(); 819 _gen_counters->update_all(); 820 } 821 } 822 823 void ConcurrentMarkSweepGeneration::print() const { 824 Generation::print(); 825 cmsSpace()->print(); 826 } 827 828 #ifndef PRODUCT 829 void ConcurrentMarkSweepGeneration::print_statistics() { 830 cmsSpace()->printFLCensus(0); 831 } 832 #endif 833 834 void ConcurrentMarkSweepGeneration::printOccupancy(const char *s) { 835 GenCollectedHeap* gch = GenCollectedHeap::heap(); 836 if (PrintGCDetails) { 837 if (Verbose) { 838 gclog_or_tty->print(" [%d %s-%s: "SIZE_FORMAT"("SIZE_FORMAT")]", 839 level(), short_name(), s, used(), capacity()); 840 } else { 841 gclog_or_tty->print(" [%d %s-%s: "SIZE_FORMAT"K("SIZE_FORMAT"K)]", 842 level(), short_name(), s, used() / K, capacity() / K); 843 } 844 } 845 if (Verbose) { 846 gclog_or_tty->print(" "SIZE_FORMAT"("SIZE_FORMAT")", 847 gch->used(), gch->capacity()); 848 } else { 849 gclog_or_tty->print(" "SIZE_FORMAT"K("SIZE_FORMAT"K)", 850 gch->used() / K, gch->capacity() / K); 851 } 852 } 853 854 size_t 855 ConcurrentMarkSweepGeneration::contiguous_available() const { 856 // dld proposes an improvement in precision here. If the committed 857 // part of the space ends in a free block we should add that to 858 // uncommitted size in the calculation below. Will make this 859 // change later, staying with the approximation below for the 860 // time being. -- ysr. 861 return MAX2(_virtual_space.uncommitted_size(), unsafe_max_alloc_nogc()); 862 } 863 864 size_t 865 ConcurrentMarkSweepGeneration::unsafe_max_alloc_nogc() const { 866 return _cmsSpace->max_alloc_in_words() * HeapWordSize; 867 } 868 869 size_t ConcurrentMarkSweepGeneration::max_available() const { 870 return free() + _virtual_space.uncommitted_size(); 871 } 872 873 bool ConcurrentMarkSweepGeneration::promotion_attempt_is_safe(size_t max_promotion_in_bytes) const { 874 size_t available = max_available(); 875 size_t av_promo = (size_t)gc_stats()->avg_promoted()->padded_average(); 876 bool res = (available >= av_promo) || (available >= max_promotion_in_bytes); 877 if (Verbose && PrintGCDetails) { 878 gclog_or_tty->print_cr( 879 "CMS: promo attempt is%s safe: available("SIZE_FORMAT") %s av_promo("SIZE_FORMAT")," 880 "max_promo("SIZE_FORMAT")", 881 res? "":" not", available, res? ">=":"<", 882 av_promo, max_promotion_in_bytes); 883 } 884 return res; 885 } 886 887 // At a promotion failure dump information on block layout in heap 888 // (cms old generation). 889 void ConcurrentMarkSweepGeneration::promotion_failure_occurred() { 890 if (CMSDumpAtPromotionFailure) { 891 cmsSpace()->dump_at_safepoint_with_locks(collector(), gclog_or_tty); 892 } 893 } 894 895 CompactibleSpace* 896 ConcurrentMarkSweepGeneration::first_compaction_space() const { 897 return _cmsSpace; 898 } 899 900 void ConcurrentMarkSweepGeneration::reset_after_compaction() { 901 // Clear the promotion information. These pointers can be adjusted 902 // along with all the other pointers into the heap but 903 // compaction is expected to be a rare event with 904 // a heap using cms so don't do it without seeing the need. 905 if (CollectedHeap::use_parallel_gc_threads()) { 906 for (uint i = 0; i < ParallelGCThreads; i++) { 907 _par_gc_thread_states[i]->promo.reset(); 908 } 909 } 910 } 911 912 void ConcurrentMarkSweepGeneration::space_iterate(SpaceClosure* blk, bool usedOnly) { 913 blk->do_space(_cmsSpace); 914 } 915 916 void ConcurrentMarkSweepGeneration::compute_new_size() { 917 assert_locked_or_safepoint(Heap_lock); 918 919 // If incremental collection failed, we just want to expand 920 // to the limit. 921 if (incremental_collection_failed()) { 922 clear_incremental_collection_failed(); 923 grow_to_reserved(); 924 return; 925 } 926 927 size_t expand_bytes = 0; 928 double free_percentage = ((double) free()) / capacity(); 929 double desired_free_percentage = (double) MinHeapFreeRatio / 100; 930 double maximum_free_percentage = (double) MaxHeapFreeRatio / 100; 931 932 // compute expansion delta needed for reaching desired free percentage 933 if (free_percentage < desired_free_percentage) { 934 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 935 assert(desired_capacity >= capacity(), "invalid expansion size"); 936 expand_bytes = MAX2(desired_capacity - capacity(), MinHeapDeltaBytes); 937 } 938 if (expand_bytes > 0) { 939 if (PrintGCDetails && Verbose) { 940 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 941 gclog_or_tty->print_cr("\nFrom compute_new_size: "); 942 gclog_or_tty->print_cr(" Free fraction %f", free_percentage); 943 gclog_or_tty->print_cr(" Desired free fraction %f", 944 desired_free_percentage); 945 gclog_or_tty->print_cr(" Maximum free fraction %f", 946 maximum_free_percentage); 947 gclog_or_tty->print_cr(" Capactiy "SIZE_FORMAT, capacity()/1000); 948 gclog_or_tty->print_cr(" Desired capacity "SIZE_FORMAT, 949 desired_capacity/1000); 950 int prev_level = level() - 1; 951 if (prev_level >= 0) { 952 size_t prev_size = 0; 953 GenCollectedHeap* gch = GenCollectedHeap::heap(); 954 Generation* prev_gen = gch->_gens[prev_level]; 955 prev_size = prev_gen->capacity(); 956 gclog_or_tty->print_cr(" Younger gen size "SIZE_FORMAT, 957 prev_size/1000); 958 } 959 gclog_or_tty->print_cr(" unsafe_max_alloc_nogc "SIZE_FORMAT, 960 unsafe_max_alloc_nogc()/1000); 961 gclog_or_tty->print_cr(" contiguous available "SIZE_FORMAT, 962 contiguous_available()/1000); 963 gclog_or_tty->print_cr(" Expand by "SIZE_FORMAT" (bytes)", 964 expand_bytes); 965 } 966 // safe if expansion fails 967 expand(expand_bytes, 0, CMSExpansionCause::_satisfy_free_ratio); 968 if (PrintGCDetails && Verbose) { 969 gclog_or_tty->print_cr(" Expanded free fraction %f", 970 ((double) free()) / capacity()); 971 } 972 } 973 } 974 975 Mutex* ConcurrentMarkSweepGeneration::freelistLock() const { 976 return cmsSpace()->freelistLock(); 977 } 978 979 HeapWord* ConcurrentMarkSweepGeneration::allocate(size_t size, 980 bool tlab) { 981 CMSSynchronousYieldRequest yr; 982 MutexLockerEx x(freelistLock(), 983 Mutex::_no_safepoint_check_flag); 984 return have_lock_and_allocate(size, tlab); 985 } 986 987 HeapWord* ConcurrentMarkSweepGeneration::have_lock_and_allocate(size_t size, 988 bool tlab /* ignored */) { 989 assert_lock_strong(freelistLock()); 990 size_t adjustedSize = CompactibleFreeListSpace::adjustObjectSize(size); 991 HeapWord* res = cmsSpace()->allocate(adjustedSize); 992 // Allocate the object live (grey) if the background collector has 993 // started marking. This is necessary because the marker may 994 // have passed this address and consequently this object will 995 // not otherwise be greyed and would be incorrectly swept up. 996 // Note that if this object contains references, the writing 997 // of those references will dirty the card containing this object 998 // allowing the object to be blackened (and its references scanned) 999 // either during a preclean phase or at the final checkpoint. 1000 if (res != NULL) { 1001 // We may block here with an uninitialized object with 1002 // its mark-bit or P-bits not yet set. Such objects need 1003 // to be safely navigable by block_start(). 1004 assert(oop(res)->klass_or_null() == NULL, "Object should be uninitialized here."); 1005 assert(!((FreeChunk*)res)->is_free(), "Error, block will look free but show wrong size"); 1006 collector()->direct_allocated(res, adjustedSize); 1007 _direct_allocated_words += adjustedSize; 1008 // allocation counters 1009 NOT_PRODUCT( 1010 _numObjectsAllocated++; 1011 _numWordsAllocated += (int)adjustedSize; 1012 ) 1013 } 1014 return res; 1015 } 1016 1017 // In the case of direct allocation by mutators in a generation that 1018 // is being concurrently collected, the object must be allocated 1019 // live (grey) if the background collector has started marking. 1020 // This is necessary because the marker may 1021 // have passed this address and consequently this object will 1022 // not otherwise be greyed and would be incorrectly swept up. 1023 // Note that if this object contains references, the writing 1024 // of those references will dirty the card containing this object 1025 // allowing the object to be blackened (and its references scanned) 1026 // either during a preclean phase or at the final checkpoint. 1027 void CMSCollector::direct_allocated(HeapWord* start, size_t size) { 1028 assert(_markBitMap.covers(start, size), "Out of bounds"); 1029 if (_collectorState >= Marking) { 1030 MutexLockerEx y(_markBitMap.lock(), 1031 Mutex::_no_safepoint_check_flag); 1032 // [see comments preceding SweepClosure::do_blk() below for details] 1033 // 1034 // Can the P-bits be deleted now? JJJ 1035 // 1036 // 1. need to mark the object as live so it isn't collected 1037 // 2. need to mark the 2nd bit to indicate the object may be uninitialized 1038 // 3. need to mark the end of the object so marking, precleaning or sweeping 1039 // can skip over uninitialized or unparsable objects. An allocated 1040 // object is considered uninitialized for our purposes as long as 1041 // its klass word is NULL. All old gen objects are parsable 1042 // as soon as they are initialized.) 1043 _markBitMap.mark(start); // object is live 1044 _markBitMap.mark(start + 1); // object is potentially uninitialized? 1045 _markBitMap.mark(start + size - 1); 1046 // mark end of object 1047 } 1048 // check that oop looks uninitialized 1049 assert(oop(start)->klass_or_null() == NULL, "_klass should be NULL"); 1050 } 1051 1052 void CMSCollector::promoted(bool par, HeapWord* start, 1053 bool is_obj_array, size_t obj_size) { 1054 assert(_markBitMap.covers(start), "Out of bounds"); 1055 // See comment in direct_allocated() about when objects should 1056 // be allocated live. 1057 if (_collectorState >= Marking) { 1058 // we already hold the marking bit map lock, taken in 1059 // the prologue 1060 if (par) { 1061 _markBitMap.par_mark(start); 1062 } else { 1063 _markBitMap.mark(start); 1064 } 1065 // We don't need to mark the object as uninitialized (as 1066 // in direct_allocated above) because this is being done with the 1067 // world stopped and the object will be initialized by the 1068 // time the marking, precleaning or sweeping get to look at it. 1069 // But see the code for copying objects into the CMS generation, 1070 // where we need to ensure that concurrent readers of the 1071 // block offset table are able to safely navigate a block that 1072 // is in flux from being free to being allocated (and in 1073 // transition while being copied into) and subsequently 1074 // becoming a bona-fide object when the copy/promotion is complete. 1075 assert(SafepointSynchronize::is_at_safepoint(), 1076 "expect promotion only at safepoints"); 1077 1078 if (_collectorState < Sweeping) { 1079 // Mark the appropriate cards in the modUnionTable, so that 1080 // this object gets scanned before the sweep. If this is 1081 // not done, CMS generation references in the object might 1082 // not get marked. 1083 // For the case of arrays, which are otherwise precisely 1084 // marked, we need to dirty the entire array, not just its head. 1085 if (is_obj_array) { 1086 // The [par_]mark_range() method expects mr.end() below to 1087 // be aligned to the granularity of a bit's representation 1088 // in the heap. In the case of the MUT below, that's a 1089 // card size. 1090 MemRegion mr(start, 1091 (HeapWord*)round_to((intptr_t)(start + obj_size), 1092 CardTableModRefBS::card_size /* bytes */)); 1093 if (par) { 1094 _modUnionTable.par_mark_range(mr); 1095 } else { 1096 _modUnionTable.mark_range(mr); 1097 } 1098 } else { // not an obj array; we can just mark the head 1099 if (par) { 1100 _modUnionTable.par_mark(start); 1101 } else { 1102 _modUnionTable.mark(start); 1103 } 1104 } 1105 } 1106 } 1107 } 1108 1109 static inline size_t percent_of_space(Space* space, HeapWord* addr) 1110 { 1111 size_t delta = pointer_delta(addr, space->bottom()); 1112 return (size_t)(delta * 100.0 / (space->capacity() / HeapWordSize)); 1113 } 1114 1115 void CMSCollector::icms_update_allocation_limits() 1116 { 1117 Generation* gen0 = GenCollectedHeap::heap()->get_gen(0); 1118 EdenSpace* eden = gen0->as_DefNewGeneration()->eden(); 1119 1120 const unsigned int duty_cycle = stats().icms_update_duty_cycle(); 1121 if (CMSTraceIncrementalPacing) { 1122 stats().print(); 1123 } 1124 1125 assert(duty_cycle <= 100, "invalid duty cycle"); 1126 if (duty_cycle != 0) { 1127 // The duty_cycle is a percentage between 0 and 100; convert to words and 1128 // then compute the offset from the endpoints of the space. 1129 size_t free_words = eden->free() / HeapWordSize; 1130 double free_words_dbl = (double)free_words; 1131 size_t duty_cycle_words = (size_t)(free_words_dbl * duty_cycle / 100.0); 1132 size_t offset_words = (free_words - duty_cycle_words) / 2; 1133 1134 _icms_start_limit = eden->top() + offset_words; 1135 _icms_stop_limit = eden->end() - offset_words; 1136 1137 // The limits may be adjusted (shifted to the right) by 1138 // CMSIncrementalOffset, to allow the application more mutator time after a 1139 // young gen gc (when all mutators were stopped) and before CMS starts and 1140 // takes away one or more cpus. 1141 if (CMSIncrementalOffset != 0) { 1142 double adjustment_dbl = free_words_dbl * CMSIncrementalOffset / 100.0; 1143 size_t adjustment = (size_t)adjustment_dbl; 1144 HeapWord* tmp_stop = _icms_stop_limit + adjustment; 1145 if (tmp_stop > _icms_stop_limit && tmp_stop < eden->end()) { 1146 _icms_start_limit += adjustment; 1147 _icms_stop_limit = tmp_stop; 1148 } 1149 } 1150 } 1151 if (duty_cycle == 0 || (_icms_start_limit == _icms_stop_limit)) { 1152 _icms_start_limit = _icms_stop_limit = eden->end(); 1153 } 1154 1155 // Install the new start limit. 1156 eden->set_soft_end(_icms_start_limit); 1157 1158 if (CMSTraceIncrementalMode) { 1159 gclog_or_tty->print(" icms alloc limits: " 1160 PTR_FORMAT "," PTR_FORMAT 1161 " (" SIZE_FORMAT "%%," SIZE_FORMAT "%%) ", 1162 _icms_start_limit, _icms_stop_limit, 1163 percent_of_space(eden, _icms_start_limit), 1164 percent_of_space(eden, _icms_stop_limit)); 1165 if (Verbose) { 1166 gclog_or_tty->print("eden: "); 1167 eden->print_on(gclog_or_tty); 1168 } 1169 } 1170 } 1171 1172 // Any changes here should try to maintain the invariant 1173 // that if this method is called with _icms_start_limit 1174 // and _icms_stop_limit both NULL, then it should return NULL 1175 // and not notify the icms thread. 1176 HeapWord* 1177 CMSCollector::allocation_limit_reached(Space* space, HeapWord* top, 1178 size_t word_size) 1179 { 1180 // A start_limit equal to end() means the duty cycle is 0, so treat that as a 1181 // nop. 1182 if (CMSIncrementalMode && _icms_start_limit != space->end()) { 1183 if (top <= _icms_start_limit) { 1184 if (CMSTraceIncrementalMode) { 1185 space->print_on(gclog_or_tty); 1186 gclog_or_tty->stamp(); 1187 gclog_or_tty->print_cr(" start limit top=" PTR_FORMAT 1188 ", new limit=" PTR_FORMAT 1189 " (" SIZE_FORMAT "%%)", 1190 top, _icms_stop_limit, 1191 percent_of_space(space, _icms_stop_limit)); 1192 } 1193 ConcurrentMarkSweepThread::start_icms(); 1194 assert(top < _icms_stop_limit, "Tautology"); 1195 if (word_size < pointer_delta(_icms_stop_limit, top)) { 1196 return _icms_stop_limit; 1197 } 1198 1199 // The allocation will cross both the _start and _stop limits, so do the 1200 // stop notification also and return end(). 1201 if (CMSTraceIncrementalMode) { 1202 space->print_on(gclog_or_tty); 1203 gclog_or_tty->stamp(); 1204 gclog_or_tty->print_cr(" +stop limit top=" PTR_FORMAT 1205 ", new limit=" PTR_FORMAT 1206 " (" SIZE_FORMAT "%%)", 1207 top, space->end(), 1208 percent_of_space(space, space->end())); 1209 } 1210 ConcurrentMarkSweepThread::stop_icms(); 1211 return space->end(); 1212 } 1213 1214 if (top <= _icms_stop_limit) { 1215 if (CMSTraceIncrementalMode) { 1216 space->print_on(gclog_or_tty); 1217 gclog_or_tty->stamp(); 1218 gclog_or_tty->print_cr(" stop limit top=" PTR_FORMAT 1219 ", new limit=" PTR_FORMAT 1220 " (" SIZE_FORMAT "%%)", 1221 top, space->end(), 1222 percent_of_space(space, space->end())); 1223 } 1224 ConcurrentMarkSweepThread::stop_icms(); 1225 return space->end(); 1226 } 1227 1228 if (CMSTraceIncrementalMode) { 1229 space->print_on(gclog_or_tty); 1230 gclog_or_tty->stamp(); 1231 gclog_or_tty->print_cr(" end limit top=" PTR_FORMAT 1232 ", new limit=" PTR_FORMAT, 1233 top, NULL); 1234 } 1235 } 1236 1237 return NULL; 1238 } 1239 1240 oop ConcurrentMarkSweepGeneration::promote(oop obj, size_t obj_size) { 1241 assert(obj_size == (size_t)obj->size(), "bad obj_size passed in"); 1242 // allocate, copy and if necessary update promoinfo -- 1243 // delegate to underlying space. 1244 assert_lock_strong(freelistLock()); 1245 1246 #ifndef PRODUCT 1247 if (Universe::heap()->promotion_should_fail()) { 1248 return NULL; 1249 } 1250 #endif // #ifndef PRODUCT 1251 1252 oop res = _cmsSpace->promote(obj, obj_size); 1253 if (res == NULL) { 1254 // expand and retry 1255 size_t s = _cmsSpace->expansionSpaceRequired(obj_size); // HeapWords 1256 expand(s*HeapWordSize, MinHeapDeltaBytes, 1257 CMSExpansionCause::_satisfy_promotion); 1258 // Since there's currently no next generation, we don't try to promote 1259 // into a more senior generation. 1260 assert(next_gen() == NULL, "assumption, based upon which no attempt " 1261 "is made to pass on a possibly failing " 1262 "promotion to next generation"); 1263 res = _cmsSpace->promote(obj, obj_size); 1264 } 1265 if (res != NULL) { 1266 // See comment in allocate() about when objects should 1267 // be allocated live. 1268 assert(obj->is_oop(), "Will dereference klass pointer below"); 1269 collector()->promoted(false, // Not parallel 1270 (HeapWord*)res, obj->is_objArray(), obj_size); 1271 // promotion counters 1272 NOT_PRODUCT( 1273 _numObjectsPromoted++; 1274 _numWordsPromoted += 1275 (int)(CompactibleFreeListSpace::adjustObjectSize(obj->size())); 1276 ) 1277 } 1278 return res; 1279 } 1280 1281 1282 HeapWord* 1283 ConcurrentMarkSweepGeneration::allocation_limit_reached(Space* space, 1284 HeapWord* top, 1285 size_t word_sz) 1286 { 1287 return collector()->allocation_limit_reached(space, top, word_sz); 1288 } 1289 1290 // IMPORTANT: Notes on object size recognition in CMS. 1291 // --------------------------------------------------- 1292 // A block of storage in the CMS generation is always in 1293 // one of three states. A free block (FREE), an allocated 1294 // object (OBJECT) whose size() method reports the correct size, 1295 // and an intermediate state (TRANSIENT) in which its size cannot 1296 // be accurately determined. 1297 // STATE IDENTIFICATION: (32 bit and 64 bit w/o COOPS) 1298 // ----------------------------------------------------- 1299 // FREE: klass_word & 1 == 1; mark_word holds block size 1300 // 1301 // OBJECT: klass_word installed; klass_word != 0 && klass_word & 1 == 0; 1302 // obj->size() computes correct size 1303 // 1304 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT 1305 // 1306 // STATE IDENTIFICATION: (64 bit+COOPS) 1307 // ------------------------------------ 1308 // FREE: mark_word & CMS_FREE_BIT == 1; mark_word & ~CMS_FREE_BIT gives block_size 1309 // 1310 // OBJECT: klass_word installed; klass_word != 0; 1311 // obj->size() computes correct size 1312 // 1313 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT 1314 // 1315 // 1316 // STATE TRANSITION DIAGRAM 1317 // 1318 // mut / parnew mut / parnew 1319 // FREE --------------------> TRANSIENT ---------------------> OBJECT --| 1320 // ^ | 1321 // |------------------------ DEAD <------------------------------------| 1322 // sweep mut 1323 // 1324 // While a block is in TRANSIENT state its size cannot be determined 1325 // so readers will either need to come back later or stall until 1326 // the size can be determined. Note that for the case of direct 1327 // allocation, P-bits, when available, may be used to determine the 1328 // size of an object that may not yet have been initialized. 1329 1330 // Things to support parallel young-gen collection. 1331 oop 1332 ConcurrentMarkSweepGeneration::par_promote(int thread_num, 1333 oop old, markOop m, 1334 size_t word_sz) { 1335 #ifndef PRODUCT 1336 if (Universe::heap()->promotion_should_fail()) { 1337 return NULL; 1338 } 1339 #endif // #ifndef PRODUCT 1340 1341 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1342 PromotionInfo* promoInfo = &ps->promo; 1343 // if we are tracking promotions, then first ensure space for 1344 // promotion (including spooling space for saving header if necessary). 1345 // then allocate and copy, then track promoted info if needed. 1346 // When tracking (see PromotionInfo::track()), the mark word may 1347 // be displaced and in this case restoration of the mark word 1348 // occurs in the (oop_since_save_marks_)iterate phase. 1349 if (promoInfo->tracking() && !promoInfo->ensure_spooling_space()) { 1350 // Out of space for allocating spooling buffers; 1351 // try expanding and allocating spooling buffers. 1352 if (!expand_and_ensure_spooling_space(promoInfo)) { 1353 return NULL; 1354 } 1355 } 1356 assert(promoInfo->has_spooling_space(), "Control point invariant"); 1357 const size_t alloc_sz = CompactibleFreeListSpace::adjustObjectSize(word_sz); 1358 HeapWord* obj_ptr = ps->lab.alloc(alloc_sz); 1359 if (obj_ptr == NULL) { 1360 obj_ptr = expand_and_par_lab_allocate(ps, alloc_sz); 1361 if (obj_ptr == NULL) { 1362 return NULL; 1363 } 1364 } 1365 oop obj = oop(obj_ptr); 1366 OrderAccess::storestore(); 1367 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1368 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1369 // IMPORTANT: See note on object initialization for CMS above. 1370 // Otherwise, copy the object. Here we must be careful to insert the 1371 // klass pointer last, since this marks the block as an allocated object. 1372 // Except with compressed oops it's the mark word. 1373 HeapWord* old_ptr = (HeapWord*)old; 1374 // Restore the mark word copied above. 1375 obj->set_mark(m); 1376 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1377 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1378 OrderAccess::storestore(); 1379 1380 if (UseCompressedKlassPointers) { 1381 // Copy gap missed by (aligned) header size calculation below 1382 obj->set_klass_gap(old->klass_gap()); 1383 } 1384 if (word_sz > (size_t)oopDesc::header_size()) { 1385 Copy::aligned_disjoint_words(old_ptr + oopDesc::header_size(), 1386 obj_ptr + oopDesc::header_size(), 1387 word_sz - oopDesc::header_size()); 1388 } 1389 1390 // Now we can track the promoted object, if necessary. We take care 1391 // to delay the transition from uninitialized to full object 1392 // (i.e., insertion of klass pointer) until after, so that it 1393 // atomically becomes a promoted object. 1394 if (promoInfo->tracking()) { 1395 promoInfo->track((PromotedObject*)obj, old->klass()); 1396 } 1397 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1398 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1399 assert(old->is_oop(), "Will use and dereference old klass ptr below"); 1400 1401 // Finally, install the klass pointer (this should be volatile). 1402 OrderAccess::storestore(); 1403 obj->set_klass(old->klass()); 1404 // We should now be able to calculate the right size for this object 1405 assert(obj->is_oop() && obj->size() == (int)word_sz, "Error, incorrect size computed for promoted object"); 1406 1407 collector()->promoted(true, // parallel 1408 obj_ptr, old->is_objArray(), word_sz); 1409 1410 NOT_PRODUCT( 1411 Atomic::inc_ptr(&_numObjectsPromoted); 1412 Atomic::add_ptr(alloc_sz, &_numWordsPromoted); 1413 ) 1414 1415 return obj; 1416 } 1417 1418 void 1419 ConcurrentMarkSweepGeneration:: 1420 par_promote_alloc_undo(int thread_num, 1421 HeapWord* obj, size_t word_sz) { 1422 // CMS does not support promotion undo. 1423 ShouldNotReachHere(); 1424 } 1425 1426 void 1427 ConcurrentMarkSweepGeneration:: 1428 par_promote_alloc_done(int thread_num) { 1429 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1430 ps->lab.retire(thread_num); 1431 } 1432 1433 void 1434 ConcurrentMarkSweepGeneration:: 1435 par_oop_since_save_marks_iterate_done(int thread_num) { 1436 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1437 ParScanWithoutBarrierClosure* dummy_cl = NULL; 1438 ps->promo.promoted_oops_iterate_nv(dummy_cl); 1439 } 1440 1441 bool ConcurrentMarkSweepGeneration::should_collect(bool full, 1442 size_t size, 1443 bool tlab) 1444 { 1445 // We allow a STW collection only if a full 1446 // collection was requested. 1447 return full || should_allocate(size, tlab); // FIX ME !!! 1448 // This and promotion failure handling are connected at the 1449 // hip and should be fixed by untying them. 1450 } 1451 1452 bool CMSCollector::shouldConcurrentCollect() { 1453 if (_full_gc_requested) { 1454 if (Verbose && PrintGCDetails) { 1455 gclog_or_tty->print_cr("CMSCollector: collect because of explicit " 1456 " gc request (or gc_locker)"); 1457 } 1458 return true; 1459 } 1460 1461 // For debugging purposes, change the type of collection. 1462 // If the rotation is not on the concurrent collection 1463 // type, don't start a concurrent collection. 1464 NOT_PRODUCT( 1465 if (RotateCMSCollectionTypes && 1466 (_cmsGen->debug_collection_type() != 1467 ConcurrentMarkSweepGeneration::Concurrent_collection_type)) { 1468 assert(_cmsGen->debug_collection_type() != 1469 ConcurrentMarkSweepGeneration::Unknown_collection_type, 1470 "Bad cms collection type"); 1471 return false; 1472 } 1473 ) 1474 1475 FreelistLocker x(this); 1476 // ------------------------------------------------------------------ 1477 // Print out lots of information which affects the initiation of 1478 // a collection. 1479 if (PrintCMSInitiationStatistics && stats().valid()) { 1480 gclog_or_tty->print("CMSCollector shouldConcurrentCollect: "); 1481 gclog_or_tty->stamp(); 1482 gclog_or_tty->print_cr(""); 1483 stats().print_on(gclog_or_tty); 1484 gclog_or_tty->print_cr("time_until_cms_gen_full %3.7f", 1485 stats().time_until_cms_gen_full()); 1486 gclog_or_tty->print_cr("free="SIZE_FORMAT, _cmsGen->free()); 1487 gclog_or_tty->print_cr("contiguous_available="SIZE_FORMAT, 1488 _cmsGen->contiguous_available()); 1489 gclog_or_tty->print_cr("promotion_rate=%g", stats().promotion_rate()); 1490 gclog_or_tty->print_cr("cms_allocation_rate=%g", stats().cms_allocation_rate()); 1491 gclog_or_tty->print_cr("occupancy=%3.7f", _cmsGen->occupancy()); 1492 gclog_or_tty->print_cr("initiatingOccupancy=%3.7f", _cmsGen->initiating_occupancy()); 1493 gclog_or_tty->print_cr("metadata initialized %d", 1494 MetaspaceGC::should_concurrent_collect()); 1495 } 1496 // ------------------------------------------------------------------ 1497 1498 // If the estimated time to complete a cms collection (cms_duration()) 1499 // is less than the estimated time remaining until the cms generation 1500 // is full, start a collection. 1501 if (!UseCMSInitiatingOccupancyOnly) { 1502 if (stats().valid()) { 1503 if (stats().time_until_cms_start() == 0.0) { 1504 return true; 1505 } 1506 } else { 1507 // We want to conservatively collect somewhat early in order 1508 // to try and "bootstrap" our CMS/promotion statistics; 1509 // this branch will not fire after the first successful CMS 1510 // collection because the stats should then be valid. 1511 if (_cmsGen->occupancy() >= _bootstrap_occupancy) { 1512 if (Verbose && PrintGCDetails) { 1513 gclog_or_tty->print_cr( 1514 " CMSCollector: collect for bootstrapping statistics:" 1515 " occupancy = %f, boot occupancy = %f", _cmsGen->occupancy(), 1516 _bootstrap_occupancy); 1517 } 1518 return true; 1519 } 1520 } 1521 } 1522 1523 // Otherwise, we start a collection cycle if 1524 // old gen want a collection cycle started. Each may use 1525 // an appropriate criterion for making this decision. 1526 // XXX We need to make sure that the gen expansion 1527 // criterion dovetails well with this. XXX NEED TO FIX THIS 1528 if (_cmsGen->should_concurrent_collect()) { 1529 if (Verbose && PrintGCDetails) { 1530 gclog_or_tty->print_cr("CMS old gen initiated"); 1531 } 1532 return true; 1533 } 1534 1535 // We start a collection if we believe an incremental collection may fail; 1536 // this is not likely to be productive in practice because it's probably too 1537 // late anyway. 1538 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1539 assert(gch->collector_policy()->is_two_generation_policy(), 1540 "You may want to check the correctness of the following"); 1541 if (gch->incremental_collection_will_fail(true /* consult_young */)) { 1542 if (Verbose && PrintGCDetails) { 1543 gclog_or_tty->print("CMSCollector: collect because incremental collection will fail "); 1544 } 1545 return true; 1546 } 1547 1548 if (MetaspaceGC::should_concurrent_collect()) { 1549 if (Verbose && PrintGCDetails) { 1550 gclog_or_tty->print("CMSCollector: collect for metadata allocation "); 1551 } 1552 return true; 1553 } 1554 1555 return false; 1556 } 1557 1558 // Clear _expansion_cause fields of constituent generations 1559 void CMSCollector::clear_expansion_cause() { 1560 _cmsGen->clear_expansion_cause(); 1561 } 1562 1563 // We should be conservative in starting a collection cycle. To 1564 // start too eagerly runs the risk of collecting too often in the 1565 // extreme. To collect too rarely falls back on full collections, 1566 // which works, even if not optimum in terms of concurrent work. 1567 // As a work around for too eagerly collecting, use the flag 1568 // UseCMSInitiatingOccupancyOnly. This also has the advantage of 1569 // giving the user an easily understandable way of controlling the 1570 // collections. 1571 // We want to start a new collection cycle if any of the following 1572 // conditions hold: 1573 // . our current occupancy exceeds the configured initiating occupancy 1574 // for this generation, or 1575 // . we recently needed to expand this space and have not, since that 1576 // expansion, done a collection of this generation, or 1577 // . the underlying space believes that it may be a good idea to initiate 1578 // a concurrent collection (this may be based on criteria such as the 1579 // following: the space uses linear allocation and linear allocation is 1580 // going to fail, or there is believed to be excessive fragmentation in 1581 // the generation, etc... or ... 1582 // [.(currently done by CMSCollector::shouldConcurrentCollect() only for 1583 // the case of the old generation; see CR 6543076): 1584 // we may be approaching a point at which allocation requests may fail because 1585 // we will be out of sufficient free space given allocation rate estimates.] 1586 bool ConcurrentMarkSweepGeneration::should_concurrent_collect() const { 1587 1588 assert_lock_strong(freelistLock()); 1589 if (occupancy() > initiating_occupancy()) { 1590 if (PrintGCDetails && Verbose) { 1591 gclog_or_tty->print(" %s: collect because of occupancy %f / %f ", 1592 short_name(), occupancy(), initiating_occupancy()); 1593 } 1594 return true; 1595 } 1596 if (UseCMSInitiatingOccupancyOnly) { 1597 return false; 1598 } 1599 if (expansion_cause() == CMSExpansionCause::_satisfy_allocation) { 1600 if (PrintGCDetails && Verbose) { 1601 gclog_or_tty->print(" %s: collect because expanded for allocation ", 1602 short_name()); 1603 } 1604 return true; 1605 } 1606 if (_cmsSpace->should_concurrent_collect()) { 1607 if (PrintGCDetails && Verbose) { 1608 gclog_or_tty->print(" %s: collect because cmsSpace says so ", 1609 short_name()); 1610 } 1611 return true; 1612 } 1613 return false; 1614 } 1615 1616 void ConcurrentMarkSweepGeneration::collect(bool full, 1617 bool clear_all_soft_refs, 1618 size_t size, 1619 bool tlab) 1620 { 1621 collector()->collect(full, clear_all_soft_refs, size, tlab); 1622 } 1623 1624 void CMSCollector::collect(bool full, 1625 bool clear_all_soft_refs, 1626 size_t size, 1627 bool tlab) 1628 { 1629 if (!UseCMSCollectionPassing && _collectorState > Idling) { 1630 // For debugging purposes skip the collection if the state 1631 // is not currently idle 1632 if (TraceCMSState) { 1633 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " skipped full:%d CMS state %d", 1634 Thread::current(), full, _collectorState); 1635 } 1636 return; 1637 } 1638 1639 // The following "if" branch is present for defensive reasons. 1640 // In the current uses of this interface, it can be replaced with: 1641 // assert(!GC_locker.is_active(), "Can't be called otherwise"); 1642 // But I am not placing that assert here to allow future 1643 // generality in invoking this interface. 1644 if (GC_locker::is_active()) { 1645 // A consistency test for GC_locker 1646 assert(GC_locker::needs_gc(), "Should have been set already"); 1647 // Skip this foreground collection, instead 1648 // expanding the heap if necessary. 1649 // Need the free list locks for the call to free() in compute_new_size() 1650 compute_new_size(); 1651 return; 1652 } 1653 acquire_control_and_collect(full, clear_all_soft_refs); 1654 _full_gcs_since_conc_gc++; 1655 1656 } 1657 1658 void CMSCollector::request_full_gc(unsigned int full_gc_count) { 1659 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1660 unsigned int gc_count = gch->total_full_collections(); 1661 if (gc_count == full_gc_count) { 1662 MutexLockerEx y(CGC_lock, Mutex::_no_safepoint_check_flag); 1663 _full_gc_requested = true; 1664 CGC_lock->notify(); // nudge CMS thread 1665 } else { 1666 assert(gc_count > full_gc_count, "Error: causal loop"); 1667 } 1668 } 1669 1670 1671 // The foreground and background collectors need to coordinate in order 1672 // to make sure that they do not mutually interfere with CMS collections. 1673 // When a background collection is active, 1674 // the foreground collector may need to take over (preempt) and 1675 // synchronously complete an ongoing collection. Depending on the 1676 // frequency of the background collections and the heap usage 1677 // of the application, this preemption can be seldom or frequent. 1678 // There are only certain 1679 // points in the background collection that the "collection-baton" 1680 // can be passed to the foreground collector. 1681 // 1682 // The foreground collector will wait for the baton before 1683 // starting any part of the collection. The foreground collector 1684 // will only wait at one location. 1685 // 1686 // The background collector will yield the baton before starting a new 1687 // phase of the collection (e.g., before initial marking, marking from roots, 1688 // precleaning, final re-mark, sweep etc.) This is normally done at the head 1689 // of the loop which switches the phases. The background collector does some 1690 // of the phases (initial mark, final re-mark) with the world stopped. 1691 // Because of locking involved in stopping the world, 1692 // the foreground collector should not block waiting for the background 1693 // collector when it is doing a stop-the-world phase. The background 1694 // collector will yield the baton at an additional point just before 1695 // it enters a stop-the-world phase. Once the world is stopped, the 1696 // background collector checks the phase of the collection. If the 1697 // phase has not changed, it proceeds with the collection. If the 1698 // phase has changed, it skips that phase of the collection. See 1699 // the comments on the use of the Heap_lock in collect_in_background(). 1700 // 1701 // Variable used in baton passing. 1702 // _foregroundGCIsActive - Set to true by the foreground collector when 1703 // it wants the baton. The foreground clears it when it has finished 1704 // the collection. 1705 // _foregroundGCShouldWait - Set to true by the background collector 1706 // when it is running. The foreground collector waits while 1707 // _foregroundGCShouldWait is true. 1708 // CGC_lock - monitor used to protect access to the above variables 1709 // and to notify the foreground and background collectors. 1710 // _collectorState - current state of the CMS collection. 1711 // 1712 // The foreground collector 1713 // acquires the CGC_lock 1714 // sets _foregroundGCIsActive 1715 // waits on the CGC_lock for _foregroundGCShouldWait to be false 1716 // various locks acquired in preparation for the collection 1717 // are released so as not to block the background collector 1718 // that is in the midst of a collection 1719 // proceeds with the collection 1720 // clears _foregroundGCIsActive 1721 // returns 1722 // 1723 // The background collector in a loop iterating on the phases of the 1724 // collection 1725 // acquires the CGC_lock 1726 // sets _foregroundGCShouldWait 1727 // if _foregroundGCIsActive is set 1728 // clears _foregroundGCShouldWait, notifies _CGC_lock 1729 // waits on _CGC_lock for _foregroundGCIsActive to become false 1730 // and exits the loop. 1731 // otherwise 1732 // proceed with that phase of the collection 1733 // if the phase is a stop-the-world phase, 1734 // yield the baton once more just before enqueueing 1735 // the stop-world CMS operation (executed by the VM thread). 1736 // returns after all phases of the collection are done 1737 // 1738 1739 void CMSCollector::acquire_control_and_collect(bool full, 1740 bool clear_all_soft_refs) { 1741 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint"); 1742 assert(!Thread::current()->is_ConcurrentGC_thread(), 1743 "shouldn't try to acquire control from self!"); 1744 1745 // Start the protocol for acquiring control of the 1746 // collection from the background collector (aka CMS thread). 1747 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 1748 "VM thread should have CMS token"); 1749 // Remember the possibly interrupted state of an ongoing 1750 // concurrent collection 1751 CollectorState first_state = _collectorState; 1752 1753 // Signal to a possibly ongoing concurrent collection that 1754 // we want to do a foreground collection. 1755 _foregroundGCIsActive = true; 1756 1757 // Disable incremental mode during a foreground collection. 1758 ICMSDisabler icms_disabler; 1759 1760 // release locks and wait for a notify from the background collector 1761 // releasing the locks in only necessary for phases which 1762 // do yields to improve the granularity of the collection. 1763 assert_lock_strong(bitMapLock()); 1764 // We need to lock the Free list lock for the space that we are 1765 // currently collecting. 1766 assert(haveFreelistLocks(), "Must be holding free list locks"); 1767 bitMapLock()->unlock(); 1768 releaseFreelistLocks(); 1769 { 1770 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1771 if (_foregroundGCShouldWait) { 1772 // We are going to be waiting for action for the CMS thread; 1773 // it had better not be gone (for instance at shutdown)! 1774 assert(ConcurrentMarkSweepThread::cmst() != NULL, 1775 "CMS thread must be running"); 1776 // Wait here until the background collector gives us the go-ahead 1777 ConcurrentMarkSweepThread::clear_CMS_flag( 1778 ConcurrentMarkSweepThread::CMS_vm_has_token); // release token 1779 // Get a possibly blocked CMS thread going: 1780 // Note that we set _foregroundGCIsActive true above, 1781 // without protection of the CGC_lock. 1782 CGC_lock->notify(); 1783 assert(!ConcurrentMarkSweepThread::vm_thread_wants_cms_token(), 1784 "Possible deadlock"); 1785 while (_foregroundGCShouldWait) { 1786 // wait for notification 1787 CGC_lock->wait(Mutex::_no_safepoint_check_flag); 1788 // Possibility of delay/starvation here, since CMS token does 1789 // not know to give priority to VM thread? Actually, i think 1790 // there wouldn't be any delay/starvation, but the proof of 1791 // that "fact" (?) appears non-trivial. XXX 20011219YSR 1792 } 1793 ConcurrentMarkSweepThread::set_CMS_flag( 1794 ConcurrentMarkSweepThread::CMS_vm_has_token); 1795 } 1796 } 1797 // The CMS_token is already held. Get back the other locks. 1798 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 1799 "VM thread should have CMS token"); 1800 getFreelistLocks(); 1801 bitMapLock()->lock_without_safepoint_check(); 1802 if (TraceCMSState) { 1803 gclog_or_tty->print_cr("CMS foreground collector has asked for control " 1804 INTPTR_FORMAT " with first state %d", Thread::current(), first_state); 1805 gclog_or_tty->print_cr(" gets control with state %d", _collectorState); 1806 } 1807 1808 // Check if we need to do a compaction, or if not, whether 1809 // we need to start the mark-sweep from scratch. 1810 bool should_compact = false; 1811 bool should_start_over = false; 1812 decide_foreground_collection_type(clear_all_soft_refs, 1813 &should_compact, &should_start_over); 1814 1815 NOT_PRODUCT( 1816 if (RotateCMSCollectionTypes) { 1817 if (_cmsGen->debug_collection_type() == 1818 ConcurrentMarkSweepGeneration::MSC_foreground_collection_type) { 1819 should_compact = true; 1820 } else if (_cmsGen->debug_collection_type() == 1821 ConcurrentMarkSweepGeneration::MS_foreground_collection_type) { 1822 should_compact = false; 1823 } 1824 } 1825 ) 1826 1827 if (PrintGCDetails && first_state > Idling) { 1828 GCCause::Cause cause = GenCollectedHeap::heap()->gc_cause(); 1829 if (GCCause::is_user_requested_gc(cause) || 1830 GCCause::is_serviceability_requested_gc(cause)) { 1831 gclog_or_tty->print(" (concurrent mode interrupted)"); 1832 } else { 1833 gclog_or_tty->print(" (concurrent mode failure)"); 1834 } 1835 } 1836 1837 if (should_compact) { 1838 // If the collection is being acquired from the background 1839 // collector, there may be references on the discovered 1840 // references lists that have NULL referents (being those 1841 // that were concurrently cleared by a mutator) or 1842 // that are no longer active (having been enqueued concurrently 1843 // by the mutator). 1844 // Scrub the list of those references because Mark-Sweep-Compact 1845 // code assumes referents are not NULL and that all discovered 1846 // Reference objects are active. 1847 ref_processor()->clean_up_discovered_references(); 1848 1849 do_compaction_work(clear_all_soft_refs); 1850 1851 // Has the GC time limit been exceeded? 1852 DefNewGeneration* young_gen = _young_gen->as_DefNewGeneration(); 1853 size_t max_eden_size = young_gen->max_capacity() - 1854 young_gen->to()->capacity() - 1855 young_gen->from()->capacity(); 1856 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1857 GCCause::Cause gc_cause = gch->gc_cause(); 1858 size_policy()->check_gc_overhead_limit(_young_gen->used(), 1859 young_gen->eden()->used(), 1860 _cmsGen->max_capacity(), 1861 max_eden_size, 1862 full, 1863 gc_cause, 1864 gch->collector_policy()); 1865 } else { 1866 do_mark_sweep_work(clear_all_soft_refs, first_state, 1867 should_start_over); 1868 } 1869 // Reset the expansion cause, now that we just completed 1870 // a collection cycle. 1871 clear_expansion_cause(); 1872 _foregroundGCIsActive = false; 1873 return; 1874 } 1875 1876 // Resize the tenured generation 1877 // after obtaining the free list locks for the 1878 // two generations. 1879 void CMSCollector::compute_new_size() { 1880 assert_locked_or_safepoint(Heap_lock); 1881 FreelistLocker z(this); 1882 MetaspaceGC::compute_new_size(); 1883 _cmsGen->compute_new_size(); 1884 } 1885 1886 // A work method used by foreground collection to determine 1887 // what type of collection (compacting or not, continuing or fresh) 1888 // it should do. 1889 // NOTE: the intent is to make UseCMSCompactAtFullCollection 1890 // and CMSCompactWhenClearAllSoftRefs the default in the future 1891 // and do away with the flags after a suitable period. 1892 void CMSCollector::decide_foreground_collection_type( 1893 bool clear_all_soft_refs, bool* should_compact, 1894 bool* should_start_over) { 1895 // Normally, we'll compact only if the UseCMSCompactAtFullCollection 1896 // flag is set, and we have either requested a System.gc() or 1897 // the number of full gc's since the last concurrent cycle 1898 // has exceeded the threshold set by CMSFullGCsBeforeCompaction, 1899 // or if an incremental collection has failed 1900 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1901 assert(gch->collector_policy()->is_two_generation_policy(), 1902 "You may want to check the correctness of the following"); 1903 // Inform cms gen if this was due to partial collection failing. 1904 // The CMS gen may use this fact to determine its expansion policy. 1905 if (gch->incremental_collection_will_fail(false /* don't consult_young */)) { 1906 assert(!_cmsGen->incremental_collection_failed(), 1907 "Should have been noticed, reacted to and cleared"); 1908 _cmsGen->set_incremental_collection_failed(); 1909 } 1910 *should_compact = 1911 UseCMSCompactAtFullCollection && 1912 ((_full_gcs_since_conc_gc >= CMSFullGCsBeforeCompaction) || 1913 GCCause::is_user_requested_gc(gch->gc_cause()) || 1914 gch->incremental_collection_will_fail(true /* consult_young */)); 1915 *should_start_over = false; 1916 if (clear_all_soft_refs && !*should_compact) { 1917 // We are about to do a last ditch collection attempt 1918 // so it would normally make sense to do a compaction 1919 // to reclaim as much space as possible. 1920 if (CMSCompactWhenClearAllSoftRefs) { 1921 // Default: The rationale is that in this case either 1922 // we are past the final marking phase, in which case 1923 // we'd have to start over, or so little has been done 1924 // that there's little point in saving that work. Compaction 1925 // appears to be the sensible choice in either case. 1926 *should_compact = true; 1927 } else { 1928 // We have been asked to clear all soft refs, but not to 1929 // compact. Make sure that we aren't past the final checkpoint 1930 // phase, for that is where we process soft refs. If we are already 1931 // past that phase, we'll need to redo the refs discovery phase and 1932 // if necessary clear soft refs that weren't previously 1933 // cleared. We do so by remembering the phase in which 1934 // we came in, and if we are past the refs processing 1935 // phase, we'll choose to just redo the mark-sweep 1936 // collection from scratch. 1937 if (_collectorState > FinalMarking) { 1938 // We are past the refs processing phase; 1939 // start over and do a fresh synchronous CMS cycle 1940 _collectorState = Resetting; // skip to reset to start new cycle 1941 reset(false /* == !asynch */); 1942 *should_start_over = true; 1943 } // else we can continue a possibly ongoing current cycle 1944 } 1945 } 1946 } 1947 1948 // A work method used by the foreground collector to do 1949 // a mark-sweep-compact. 1950 void CMSCollector::do_compaction_work(bool clear_all_soft_refs) { 1951 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1952 TraceTime t("CMS:MSC ", PrintGCDetails && Verbose, true, gclog_or_tty); 1953 if (PrintGC && Verbose && !(GCCause::is_user_requested_gc(gch->gc_cause()))) { 1954 gclog_or_tty->print_cr("Compact ConcurrentMarkSweepGeneration after %d " 1955 "collections passed to foreground collector", _full_gcs_since_conc_gc); 1956 } 1957 1958 // Sample collection interval time and reset for collection pause. 1959 if (UseAdaptiveSizePolicy) { 1960 size_policy()->msc_collection_begin(); 1961 } 1962 1963 // Temporarily widen the span of the weak reference processing to 1964 // the entire heap. 1965 MemRegion new_span(GenCollectedHeap::heap()->reserved_region()); 1966 ReferenceProcessorSpanMutator rp_mut_span(ref_processor(), new_span); 1967 // Temporarily, clear the "is_alive_non_header" field of the 1968 // reference processor. 1969 ReferenceProcessorIsAliveMutator rp_mut_closure(ref_processor(), NULL); 1970 // Temporarily make reference _processing_ single threaded (non-MT). 1971 ReferenceProcessorMTProcMutator rp_mut_mt_processing(ref_processor(), false); 1972 // Temporarily make refs discovery atomic 1973 ReferenceProcessorAtomicMutator rp_mut_atomic(ref_processor(), true); 1974 // Temporarily make reference _discovery_ single threaded (non-MT) 1975 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false); 1976 1977 ref_processor()->set_enqueuing_is_done(false); 1978 ref_processor()->enable_discovery(false /*verify_disabled*/, false /*check_no_refs*/); 1979 ref_processor()->setup_policy(clear_all_soft_refs); 1980 // If an asynchronous collection finishes, the _modUnionTable is 1981 // all clear. If we are assuming the collection from an asynchronous 1982 // collection, clear the _modUnionTable. 1983 assert(_collectorState != Idling || _modUnionTable.isAllClear(), 1984 "_modUnionTable should be clear if the baton was not passed"); 1985 _modUnionTable.clear_all(); 1986 assert(_collectorState != Idling || _ct->klass_rem_set()->mod_union_is_clear(), 1987 "mod union for klasses should be clear if the baton was passed"); 1988 _ct->klass_rem_set()->clear_mod_union(); 1989 1990 // We must adjust the allocation statistics being maintained 1991 // in the free list space. We do so by reading and clearing 1992 // the sweep timer and updating the block flux rate estimates below. 1993 assert(!_intra_sweep_timer.is_active(), "_intra_sweep_timer should be inactive"); 1994 if (_inter_sweep_timer.is_active()) { 1995 _inter_sweep_timer.stop(); 1996 // Note that we do not use this sample to update the _inter_sweep_estimate. 1997 _cmsGen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()), 1998 _inter_sweep_estimate.padded_average(), 1999 _intra_sweep_estimate.padded_average()); 2000 } 2001 2002 GenMarkSweep::invoke_at_safepoint(_cmsGen->level(), 2003 ref_processor(), clear_all_soft_refs); 2004 #ifdef ASSERT 2005 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 2006 size_t free_size = cms_space->free(); 2007 assert(free_size == 2008 pointer_delta(cms_space->end(), cms_space->compaction_top()) 2009 * HeapWordSize, 2010 "All the free space should be compacted into one chunk at top"); 2011 assert(cms_space->dictionary()->total_chunk_size( 2012 debug_only(cms_space->freelistLock())) == 0 || 2013 cms_space->totalSizeInIndexedFreeLists() == 0, 2014 "All the free space should be in a single chunk"); 2015 size_t num = cms_space->totalCount(); 2016 assert((free_size == 0 && num == 0) || 2017 (free_size > 0 && (num == 1 || num == 2)), 2018 "There should be at most 2 free chunks after compaction"); 2019 #endif // ASSERT 2020 _collectorState = Resetting; 2021 assert(_restart_addr == NULL, 2022 "Should have been NULL'd before baton was passed"); 2023 reset(false /* == !asynch */); 2024 _cmsGen->reset_after_compaction(); 2025 _concurrent_cycles_since_last_unload = 0; 2026 2027 // Clear any data recorded in the PLAB chunk arrays. 2028 if (_survivor_plab_array != NULL) { 2029 reset_survivor_plab_arrays(); 2030 } 2031 2032 // Adjust the per-size allocation stats for the next epoch. 2033 _cmsGen->cmsSpace()->endSweepFLCensus(sweep_count() /* fake */); 2034 // Restart the "inter sweep timer" for the next epoch. 2035 _inter_sweep_timer.reset(); 2036 _inter_sweep_timer.start(); 2037 2038 // Sample collection pause time and reset for collection interval. 2039 if (UseAdaptiveSizePolicy) { 2040 size_policy()->msc_collection_end(gch->gc_cause()); 2041 } 2042 2043 // For a mark-sweep-compact, compute_new_size() will be called 2044 // in the heap's do_collection() method. 2045 } 2046 2047 // A work method used by the foreground collector to do 2048 // a mark-sweep, after taking over from a possibly on-going 2049 // concurrent mark-sweep collection. 2050 void CMSCollector::do_mark_sweep_work(bool clear_all_soft_refs, 2051 CollectorState first_state, bool should_start_over) { 2052 if (PrintGC && Verbose) { 2053 gclog_or_tty->print_cr("Pass concurrent collection to foreground " 2054 "collector with count %d", 2055 _full_gcs_since_conc_gc); 2056 } 2057 switch (_collectorState) { 2058 case Idling: 2059 if (first_state == Idling || should_start_over) { 2060 // The background GC was not active, or should 2061 // restarted from scratch; start the cycle. 2062 _collectorState = InitialMarking; 2063 } 2064 // If first_state was not Idling, then a background GC 2065 // was in progress and has now finished. No need to do it 2066 // again. Leave the state as Idling. 2067 break; 2068 case Precleaning: 2069 // In the foreground case don't do the precleaning since 2070 // it is not done concurrently and there is extra work 2071 // required. 2072 _collectorState = FinalMarking; 2073 } 2074 if (PrintGCDetails && 2075 (_collectorState > Idling || 2076 !GCCause::is_user_requested_gc(GenCollectedHeap::heap()->gc_cause()))) { 2077 gclog_or_tty->print(" (concurrent mode failure)"); 2078 } 2079 collect_in_foreground(clear_all_soft_refs); 2080 2081 // For a mark-sweep, compute_new_size() will be called 2082 // in the heap's do_collection() method. 2083 } 2084 2085 2086 void CMSCollector::getFreelistLocks() const { 2087 // Get locks for all free lists in all generations that this 2088 // collector is responsible for 2089 _cmsGen->freelistLock()->lock_without_safepoint_check(); 2090 } 2091 2092 void CMSCollector::releaseFreelistLocks() const { 2093 // Release locks for all free lists in all generations that this 2094 // collector is responsible for 2095 _cmsGen->freelistLock()->unlock(); 2096 } 2097 2098 bool CMSCollector::haveFreelistLocks() const { 2099 // Check locks for all free lists in all generations that this 2100 // collector is responsible for 2101 assert_lock_strong(_cmsGen->freelistLock()); 2102 PRODUCT_ONLY(ShouldNotReachHere()); 2103 return true; 2104 } 2105 2106 // A utility class that is used by the CMS collector to 2107 // temporarily "release" the foreground collector from its 2108 // usual obligation to wait for the background collector to 2109 // complete an ongoing phase before proceeding. 2110 class ReleaseForegroundGC: public StackObj { 2111 private: 2112 CMSCollector* _c; 2113 public: 2114 ReleaseForegroundGC(CMSCollector* c) : _c(c) { 2115 assert(_c->_foregroundGCShouldWait, "Else should not need to call"); 2116 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 2117 // allow a potentially blocked foreground collector to proceed 2118 _c->_foregroundGCShouldWait = false; 2119 if (_c->_foregroundGCIsActive) { 2120 CGC_lock->notify(); 2121 } 2122 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 2123 "Possible deadlock"); 2124 } 2125 2126 ~ReleaseForegroundGC() { 2127 assert(!_c->_foregroundGCShouldWait, "Usage protocol violation?"); 2128 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 2129 _c->_foregroundGCShouldWait = true; 2130 } 2131 }; 2132 2133 // There are separate collect_in_background and collect_in_foreground because of 2134 // the different locking requirements of the background collector and the 2135 // foreground collector. There was originally an attempt to share 2136 // one "collect" method between the background collector and the foreground 2137 // collector but the if-then-else required made it cleaner to have 2138 // separate methods. 2139 void CMSCollector::collect_in_background(bool clear_all_soft_refs) { 2140 assert(Thread::current()->is_ConcurrentGC_thread(), 2141 "A CMS asynchronous collection is only allowed on a CMS thread."); 2142 2143 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2144 { 2145 bool safepoint_check = Mutex::_no_safepoint_check_flag; 2146 MutexLockerEx hl(Heap_lock, safepoint_check); 2147 FreelistLocker fll(this); 2148 MutexLockerEx x(CGC_lock, safepoint_check); 2149 if (_foregroundGCIsActive || !UseAsyncConcMarkSweepGC) { 2150 // The foreground collector is active or we're 2151 // not using asynchronous collections. Skip this 2152 // background collection. 2153 assert(!_foregroundGCShouldWait, "Should be clear"); 2154 return; 2155 } else { 2156 assert(_collectorState == Idling, "Should be idling before start."); 2157 _collectorState = InitialMarking; 2158 // Reset the expansion cause, now that we are about to begin 2159 // a new cycle. 2160 clear_expansion_cause(); 2161 2162 // Clear the MetaspaceGC flag since a concurrent collection 2163 // is starting but also clear it after the collection. 2164 MetaspaceGC::set_should_concurrent_collect(false); 2165 } 2166 // Decide if we want to enable class unloading as part of the 2167 // ensuing concurrent GC cycle. 2168 update_should_unload_classes(); 2169 _full_gc_requested = false; // acks all outstanding full gc requests 2170 // Signal that we are about to start a collection 2171 gch->increment_total_full_collections(); // ... starting a collection cycle 2172 _collection_count_start = gch->total_full_collections(); 2173 } 2174 2175 // Used for PrintGC 2176 size_t prev_used; 2177 if (PrintGC && Verbose) { 2178 prev_used = _cmsGen->used(); // XXXPERM 2179 } 2180 2181 // The change of the collection state is normally done at this level; 2182 // the exceptions are phases that are executed while the world is 2183 // stopped. For those phases the change of state is done while the 2184 // world is stopped. For baton passing purposes this allows the 2185 // background collector to finish the phase and change state atomically. 2186 // The foreground collector cannot wait on a phase that is done 2187 // while the world is stopped because the foreground collector already 2188 // has the world stopped and would deadlock. 2189 while (_collectorState != Idling) { 2190 if (TraceCMSState) { 2191 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " in CMS state %d", 2192 Thread::current(), _collectorState); 2193 } 2194 // The foreground collector 2195 // holds the Heap_lock throughout its collection. 2196 // holds the CMS token (but not the lock) 2197 // except while it is waiting for the background collector to yield. 2198 // 2199 // The foreground collector should be blocked (not for long) 2200 // if the background collector is about to start a phase 2201 // executed with world stopped. If the background 2202 // collector has already started such a phase, the 2203 // foreground collector is blocked waiting for the 2204 // Heap_lock. The stop-world phases (InitialMarking and FinalMarking) 2205 // are executed in the VM thread. 2206 // 2207 // The locking order is 2208 // PendingListLock (PLL) -- if applicable (FinalMarking) 2209 // Heap_lock (both this & PLL locked in VM_CMS_Operation::prologue()) 2210 // CMS token (claimed in 2211 // stop_world_and_do() --> 2212 // safepoint_synchronize() --> 2213 // CMSThread::synchronize()) 2214 2215 { 2216 // Check if the FG collector wants us to yield. 2217 CMSTokenSync x(true); // is cms thread 2218 if (waitForForegroundGC()) { 2219 // We yielded to a foreground GC, nothing more to be 2220 // done this round. 2221 assert(_foregroundGCShouldWait == false, "We set it to false in " 2222 "waitForForegroundGC()"); 2223 if (TraceCMSState) { 2224 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT 2225 " exiting collection CMS state %d", 2226 Thread::current(), _collectorState); 2227 } 2228 return; 2229 } else { 2230 // The background collector can run but check to see if the 2231 // foreground collector has done a collection while the 2232 // background collector was waiting to get the CGC_lock 2233 // above. If yes, break so that _foregroundGCShouldWait 2234 // is cleared before returning. 2235 if (_collectorState == Idling) { 2236 break; 2237 } 2238 } 2239 } 2240 2241 assert(_foregroundGCShouldWait, "Foreground collector, if active, " 2242 "should be waiting"); 2243 2244 switch (_collectorState) { 2245 case InitialMarking: 2246 { 2247 ReleaseForegroundGC x(this); 2248 stats().record_cms_begin(); 2249 2250 VM_CMS_Initial_Mark initial_mark_op(this); 2251 VMThread::execute(&initial_mark_op); 2252 } 2253 // The collector state may be any legal state at this point 2254 // since the background collector may have yielded to the 2255 // foreground collector. 2256 break; 2257 case Marking: 2258 // initial marking in checkpointRootsInitialWork has been completed 2259 if (markFromRoots(true)) { // we were successful 2260 assert(_collectorState == Precleaning, "Collector state should " 2261 "have changed"); 2262 } else { 2263 assert(_foregroundGCIsActive, "Internal state inconsistency"); 2264 } 2265 break; 2266 case Precleaning: 2267 if (UseAdaptiveSizePolicy) { 2268 size_policy()->concurrent_precleaning_begin(); 2269 } 2270 // marking from roots in markFromRoots has been completed 2271 preclean(); 2272 if (UseAdaptiveSizePolicy) { 2273 size_policy()->concurrent_precleaning_end(); 2274 } 2275 assert(_collectorState == AbortablePreclean || 2276 _collectorState == FinalMarking, 2277 "Collector state should have changed"); 2278 break; 2279 case AbortablePreclean: 2280 if (UseAdaptiveSizePolicy) { 2281 size_policy()->concurrent_phases_resume(); 2282 } 2283 abortable_preclean(); 2284 if (UseAdaptiveSizePolicy) { 2285 size_policy()->concurrent_precleaning_end(); 2286 } 2287 assert(_collectorState == FinalMarking, "Collector state should " 2288 "have changed"); 2289 break; 2290 case FinalMarking: 2291 { 2292 ReleaseForegroundGC x(this); 2293 2294 VM_CMS_Final_Remark final_remark_op(this); 2295 VMThread::execute(&final_remark_op); 2296 } 2297 assert(_foregroundGCShouldWait, "block post-condition"); 2298 break; 2299 case Sweeping: 2300 if (UseAdaptiveSizePolicy) { 2301 size_policy()->concurrent_sweeping_begin(); 2302 } 2303 // final marking in checkpointRootsFinal has been completed 2304 sweep(true); 2305 assert(_collectorState == Resizing, "Collector state change " 2306 "to Resizing must be done under the free_list_lock"); 2307 _full_gcs_since_conc_gc = 0; 2308 2309 // Stop the timers for adaptive size policy for the concurrent phases 2310 if (UseAdaptiveSizePolicy) { 2311 size_policy()->concurrent_sweeping_end(); 2312 size_policy()->concurrent_phases_end(gch->gc_cause(), 2313 gch->prev_gen(_cmsGen)->capacity(), 2314 _cmsGen->free()); 2315 } 2316 2317 case Resizing: { 2318 // Sweeping has been completed... 2319 // At this point the background collection has completed. 2320 // Don't move the call to compute_new_size() down 2321 // into code that might be executed if the background 2322 // collection was preempted. 2323 { 2324 ReleaseForegroundGC x(this); // unblock FG collection 2325 MutexLockerEx y(Heap_lock, Mutex::_no_safepoint_check_flag); 2326 CMSTokenSync z(true); // not strictly needed. 2327 if (_collectorState == Resizing) { 2328 compute_new_size(); 2329 _collectorState = Resetting; 2330 } else { 2331 assert(_collectorState == Idling, "The state should only change" 2332 " because the foreground collector has finished the collection"); 2333 } 2334 } 2335 break; 2336 } 2337 case Resetting: 2338 // CMS heap resizing has been completed 2339 reset(true); 2340 assert(_collectorState == Idling, "Collector state should " 2341 "have changed"); 2342 2343 MetaspaceGC::set_should_concurrent_collect(false); 2344 2345 stats().record_cms_end(); 2346 // Don't move the concurrent_phases_end() and compute_new_size() 2347 // calls to here because a preempted background collection 2348 // has it's state set to "Resetting". 2349 break; 2350 case Idling: 2351 default: 2352 ShouldNotReachHere(); 2353 break; 2354 } 2355 if (TraceCMSState) { 2356 gclog_or_tty->print_cr(" Thread " INTPTR_FORMAT " done - next CMS state %d", 2357 Thread::current(), _collectorState); 2358 } 2359 assert(_foregroundGCShouldWait, "block post-condition"); 2360 } 2361 2362 // Should this be in gc_epilogue? 2363 collector_policy()->counters()->update_counters(); 2364 2365 { 2366 // Clear _foregroundGCShouldWait and, in the event that the 2367 // foreground collector is waiting, notify it, before 2368 // returning. 2369 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 2370 _foregroundGCShouldWait = false; 2371 if (_foregroundGCIsActive) { 2372 CGC_lock->notify(); 2373 } 2374 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 2375 "Possible deadlock"); 2376 } 2377 if (TraceCMSState) { 2378 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT 2379 " exiting collection CMS state %d", 2380 Thread::current(), _collectorState); 2381 } 2382 if (PrintGC && Verbose) { 2383 _cmsGen->print_heap_change(prev_used); 2384 } 2385 } 2386 2387 void CMSCollector::collect_in_foreground(bool clear_all_soft_refs) { 2388 assert(_foregroundGCIsActive && !_foregroundGCShouldWait, 2389 "Foreground collector should be waiting, not executing"); 2390 assert(Thread::current()->is_VM_thread(), "A foreground collection" 2391 "may only be done by the VM Thread with the world stopped"); 2392 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 2393 "VM thread should have CMS token"); 2394 2395 NOT_PRODUCT(TraceTime t("CMS:MS (foreground) ", PrintGCDetails && Verbose, 2396 true, gclog_or_tty);) 2397 if (UseAdaptiveSizePolicy) { 2398 size_policy()->ms_collection_begin(); 2399 } 2400 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact); 2401 2402 HandleMark hm; // Discard invalid handles created during verification 2403 2404 if (VerifyBeforeGC && 2405 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2406 Universe::verify(true); 2407 } 2408 2409 // Snapshot the soft reference policy to be used in this collection cycle. 2410 ref_processor()->setup_policy(clear_all_soft_refs); 2411 2412 bool init_mark_was_synchronous = false; // until proven otherwise 2413 while (_collectorState != Idling) { 2414 if (TraceCMSState) { 2415 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " in CMS state %d", 2416 Thread::current(), _collectorState); 2417 } 2418 switch (_collectorState) { 2419 case InitialMarking: 2420 init_mark_was_synchronous = true; // fact to be exploited in re-mark 2421 checkpointRootsInitial(false); 2422 assert(_collectorState == Marking, "Collector state should have changed" 2423 " within checkpointRootsInitial()"); 2424 break; 2425 case Marking: 2426 // initial marking in checkpointRootsInitialWork has been completed 2427 if (VerifyDuringGC && 2428 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2429 gclog_or_tty->print("Verify before initial mark: "); 2430 Universe::verify(true); 2431 } 2432 { 2433 bool res = markFromRoots(false); 2434 assert(res && _collectorState == FinalMarking, "Collector state should " 2435 "have changed"); 2436 break; 2437 } 2438 case FinalMarking: 2439 if (VerifyDuringGC && 2440 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2441 gclog_or_tty->print("Verify before re-mark: "); 2442 Universe::verify(true); 2443 } 2444 checkpointRootsFinal(false, clear_all_soft_refs, 2445 init_mark_was_synchronous); 2446 assert(_collectorState == Sweeping, "Collector state should not " 2447 "have changed within checkpointRootsFinal()"); 2448 break; 2449 case Sweeping: 2450 // final marking in checkpointRootsFinal has been completed 2451 if (VerifyDuringGC && 2452 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2453 gclog_or_tty->print("Verify before sweep: "); 2454 Universe::verify(true); 2455 } 2456 sweep(false); 2457 assert(_collectorState == Resizing, "Incorrect state"); 2458 break; 2459 case Resizing: { 2460 // Sweeping has been completed; the actual resize in this case 2461 // is done separately; nothing to be done in this state. 2462 _collectorState = Resetting; 2463 break; 2464 } 2465 case Resetting: 2466 // The heap has been resized. 2467 if (VerifyDuringGC && 2468 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2469 gclog_or_tty->print("Verify before reset: "); 2470 Universe::verify(true); 2471 } 2472 reset(false); 2473 assert(_collectorState == Idling, "Collector state should " 2474 "have changed"); 2475 break; 2476 case Precleaning: 2477 case AbortablePreclean: 2478 // Elide the preclean phase 2479 _collectorState = FinalMarking; 2480 break; 2481 default: 2482 ShouldNotReachHere(); 2483 } 2484 if (TraceCMSState) { 2485 gclog_or_tty->print_cr(" Thread " INTPTR_FORMAT " done - next CMS state %d", 2486 Thread::current(), _collectorState); 2487 } 2488 } 2489 2490 if (UseAdaptiveSizePolicy) { 2491 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2492 size_policy()->ms_collection_end(gch->gc_cause()); 2493 } 2494 2495 if (VerifyAfterGC && 2496 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2497 Universe::verify(true); 2498 } 2499 if (TraceCMSState) { 2500 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT 2501 " exiting collection CMS state %d", 2502 Thread::current(), _collectorState); 2503 } 2504 } 2505 2506 bool CMSCollector::waitForForegroundGC() { 2507 bool res = false; 2508 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 2509 "CMS thread should have CMS token"); 2510 // Block the foreground collector until the 2511 // background collectors decides whether to 2512 // yield. 2513 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 2514 _foregroundGCShouldWait = true; 2515 if (_foregroundGCIsActive) { 2516 // The background collector yields to the 2517 // foreground collector and returns a value 2518 // indicating that it has yielded. The foreground 2519 // collector can proceed. 2520 res = true; 2521 _foregroundGCShouldWait = false; 2522 ConcurrentMarkSweepThread::clear_CMS_flag( 2523 ConcurrentMarkSweepThread::CMS_cms_has_token); 2524 ConcurrentMarkSweepThread::set_CMS_flag( 2525 ConcurrentMarkSweepThread::CMS_cms_wants_token); 2526 // Get a possibly blocked foreground thread going 2527 CGC_lock->notify(); 2528 if (TraceCMSState) { 2529 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " waiting at CMS state %d", 2530 Thread::current(), _collectorState); 2531 } 2532 while (_foregroundGCIsActive) { 2533 CGC_lock->wait(Mutex::_no_safepoint_check_flag); 2534 } 2535 ConcurrentMarkSweepThread::set_CMS_flag( 2536 ConcurrentMarkSweepThread::CMS_cms_has_token); 2537 ConcurrentMarkSweepThread::clear_CMS_flag( 2538 ConcurrentMarkSweepThread::CMS_cms_wants_token); 2539 } 2540 if (TraceCMSState) { 2541 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " continuing at CMS state %d", 2542 Thread::current(), _collectorState); 2543 } 2544 return res; 2545 } 2546 2547 // Because of the need to lock the free lists and other structures in 2548 // the collector, common to all the generations that the collector is 2549 // collecting, we need the gc_prologues of individual CMS generations 2550 // delegate to their collector. It may have been simpler had the 2551 // current infrastructure allowed one to call a prologue on a 2552 // collector. In the absence of that we have the generation's 2553 // prologue delegate to the collector, which delegates back 2554 // some "local" work to a worker method in the individual generations 2555 // that it's responsible for collecting, while itself doing any 2556 // work common to all generations it's responsible for. A similar 2557 // comment applies to the gc_epilogue()'s. 2558 // The role of the varaible _between_prologue_and_epilogue is to 2559 // enforce the invocation protocol. 2560 void CMSCollector::gc_prologue(bool full) { 2561 // Call gc_prologue_work() for the CMSGen 2562 // we are responsible for. 2563 2564 // The following locking discipline assumes that we are only called 2565 // when the world is stopped. 2566 assert(SafepointSynchronize::is_at_safepoint(), "world is stopped assumption"); 2567 2568 // The CMSCollector prologue must call the gc_prologues for the 2569 // "generations" that it's responsible 2570 // for. 2571 2572 assert( Thread::current()->is_VM_thread() 2573 || ( CMSScavengeBeforeRemark 2574 && Thread::current()->is_ConcurrentGC_thread()), 2575 "Incorrect thread type for prologue execution"); 2576 2577 if (_between_prologue_and_epilogue) { 2578 // We have already been invoked; this is a gc_prologue delegation 2579 // from yet another CMS generation that we are responsible for, just 2580 // ignore it since all relevant work has already been done. 2581 return; 2582 } 2583 2584 // set a bit saying prologue has been called; cleared in epilogue 2585 _between_prologue_and_epilogue = true; 2586 // Claim locks for common data structures, then call gc_prologue_work() 2587 // for each CMSGen. 2588 2589 getFreelistLocks(); // gets free list locks on constituent spaces 2590 bitMapLock()->lock_without_safepoint_check(); 2591 2592 // Should call gc_prologue_work() for all cms gens we are responsible for 2593 bool duringMarking = _collectorState >= Marking 2594 && _collectorState < Sweeping; 2595 2596 // The young collections clear the modified oops state, which tells if 2597 // there are any modified oops in the class. The remark phase also needs 2598 // that information. Tell the young collection to save the union of all 2599 // modified klasses. 2600 if (duringMarking) { 2601 _ct->klass_rem_set()->set_accumulate_modified_oops(true); 2602 } 2603 2604 bool registerClosure = duringMarking; 2605 2606 ModUnionClosure* muc = CollectedHeap::use_parallel_gc_threads() ? 2607 &_modUnionClosurePar 2608 : &_modUnionClosure; 2609 _cmsGen->gc_prologue_work(full, registerClosure, muc); 2610 2611 if (!full) { 2612 stats().record_gc0_begin(); 2613 } 2614 } 2615 2616 void ConcurrentMarkSweepGeneration::gc_prologue(bool full) { 2617 // Delegate to CMScollector which knows how to coordinate between 2618 // this and any other CMS generations that it is responsible for 2619 // collecting. 2620 collector()->gc_prologue(full); 2621 } 2622 2623 // This is a "private" interface for use by this generation's CMSCollector. 2624 // Not to be called directly by any other entity (for instance, 2625 // GenCollectedHeap, which calls the "public" gc_prologue method above). 2626 void ConcurrentMarkSweepGeneration::gc_prologue_work(bool full, 2627 bool registerClosure, ModUnionClosure* modUnionClosure) { 2628 assert(!incremental_collection_failed(), "Shouldn't be set yet"); 2629 assert(cmsSpace()->preconsumptionDirtyCardClosure() == NULL, 2630 "Should be NULL"); 2631 if (registerClosure) { 2632 cmsSpace()->setPreconsumptionDirtyCardClosure(modUnionClosure); 2633 } 2634 cmsSpace()->gc_prologue(); 2635 // Clear stat counters 2636 NOT_PRODUCT( 2637 assert(_numObjectsPromoted == 0, "check"); 2638 assert(_numWordsPromoted == 0, "check"); 2639 if (Verbose && PrintGC) { 2640 gclog_or_tty->print("Allocated "SIZE_FORMAT" objects, " 2641 SIZE_FORMAT" bytes concurrently", 2642 _numObjectsAllocated, _numWordsAllocated*sizeof(HeapWord)); 2643 } 2644 _numObjectsAllocated = 0; 2645 _numWordsAllocated = 0; 2646 ) 2647 } 2648 2649 void CMSCollector::gc_epilogue(bool full) { 2650 // The following locking discipline assumes that we are only called 2651 // when the world is stopped. 2652 assert(SafepointSynchronize::is_at_safepoint(), 2653 "world is stopped assumption"); 2654 2655 // Currently the CMS epilogue (see CompactibleFreeListSpace) merely checks 2656 // if linear allocation blocks need to be appropriately marked to allow the 2657 // the blocks to be parsable. We also check here whether we need to nudge the 2658 // CMS collector thread to start a new cycle (if it's not already active). 2659 assert( Thread::current()->is_VM_thread() 2660 || ( CMSScavengeBeforeRemark 2661 && Thread::current()->is_ConcurrentGC_thread()), 2662 "Incorrect thread type for epilogue execution"); 2663 2664 if (!_between_prologue_and_epilogue) { 2665 // We have already been invoked; this is a gc_epilogue delegation 2666 // from yet another CMS generation that we are responsible for, just 2667 // ignore it since all relevant work has already been done. 2668 return; 2669 } 2670 assert(haveFreelistLocks(), "must have freelist locks"); 2671 assert_lock_strong(bitMapLock()); 2672 2673 _ct->klass_rem_set()->set_accumulate_modified_oops(false); 2674 2675 _cmsGen->gc_epilogue_work(full); 2676 2677 if (_collectorState == AbortablePreclean || _collectorState == Precleaning) { 2678 // in case sampling was not already enabled, enable it 2679 _start_sampling = true; 2680 } 2681 // reset _eden_chunk_array so sampling starts afresh 2682 _eden_chunk_index = 0; 2683 2684 size_t cms_used = _cmsGen->cmsSpace()->used(); 2685 2686 // update performance counters - this uses a special version of 2687 // update_counters() that allows the utilization to be passed as a 2688 // parameter, avoiding multiple calls to used(). 2689 // 2690 _cmsGen->update_counters(cms_used); 2691 2692 if (CMSIncrementalMode) { 2693 icms_update_allocation_limits(); 2694 } 2695 2696 bitMapLock()->unlock(); 2697 releaseFreelistLocks(); 2698 2699 if (!CleanChunkPoolAsync) { 2700 Chunk::clean_chunk_pool(); 2701 } 2702 2703 _between_prologue_and_epilogue = false; // ready for next cycle 2704 } 2705 2706 void ConcurrentMarkSweepGeneration::gc_epilogue(bool full) { 2707 collector()->gc_epilogue(full); 2708 2709 // Also reset promotion tracking in par gc thread states. 2710 if (CollectedHeap::use_parallel_gc_threads()) { 2711 for (uint i = 0; i < ParallelGCThreads; i++) { 2712 _par_gc_thread_states[i]->promo.stopTrackingPromotions(i); 2713 } 2714 } 2715 } 2716 2717 void ConcurrentMarkSweepGeneration::gc_epilogue_work(bool full) { 2718 assert(!incremental_collection_failed(), "Should have been cleared"); 2719 cmsSpace()->setPreconsumptionDirtyCardClosure(NULL); 2720 cmsSpace()->gc_epilogue(); 2721 // Print stat counters 2722 NOT_PRODUCT( 2723 assert(_numObjectsAllocated == 0, "check"); 2724 assert(_numWordsAllocated == 0, "check"); 2725 if (Verbose && PrintGC) { 2726 gclog_or_tty->print("Promoted "SIZE_FORMAT" objects, " 2727 SIZE_FORMAT" bytes", 2728 _numObjectsPromoted, _numWordsPromoted*sizeof(HeapWord)); 2729 } 2730 _numObjectsPromoted = 0; 2731 _numWordsPromoted = 0; 2732 ) 2733 2734 if (PrintGC && Verbose) { 2735 // Call down the chain in contiguous_available needs the freelistLock 2736 // so print this out before releasing the freeListLock. 2737 gclog_or_tty->print(" Contiguous available "SIZE_FORMAT" bytes ", 2738 contiguous_available()); 2739 } 2740 } 2741 2742 #ifndef PRODUCT 2743 bool CMSCollector::have_cms_token() { 2744 Thread* thr = Thread::current(); 2745 if (thr->is_VM_thread()) { 2746 return ConcurrentMarkSweepThread::vm_thread_has_cms_token(); 2747 } else if (thr->is_ConcurrentGC_thread()) { 2748 return ConcurrentMarkSweepThread::cms_thread_has_cms_token(); 2749 } else if (thr->is_GC_task_thread()) { 2750 return ConcurrentMarkSweepThread::vm_thread_has_cms_token() && 2751 ParGCRareEvent_lock->owned_by_self(); 2752 } 2753 return false; 2754 } 2755 #endif 2756 2757 // Check reachability of the given heap address in CMS generation, 2758 // treating all other generations as roots. 2759 bool CMSCollector::is_cms_reachable(HeapWord* addr) { 2760 // We could "guarantee" below, rather than assert, but i'll 2761 // leave these as "asserts" so that an adventurous debugger 2762 // could try this in the product build provided some subset of 2763 // the conditions were met, provided they were intersted in the 2764 // results and knew that the computation below wouldn't interfere 2765 // with other concurrent computations mutating the structures 2766 // being read or written. 2767 assert(SafepointSynchronize::is_at_safepoint(), 2768 "Else mutations in object graph will make answer suspect"); 2769 assert(have_cms_token(), "Should hold cms token"); 2770 assert(haveFreelistLocks(), "must hold free list locks"); 2771 assert_lock_strong(bitMapLock()); 2772 2773 // Clear the marking bit map array before starting, but, just 2774 // for kicks, first report if the given address is already marked 2775 gclog_or_tty->print_cr("Start: Address 0x%x is%s marked", addr, 2776 _markBitMap.isMarked(addr) ? "" : " not"); 2777 2778 if (verify_after_remark()) { 2779 MutexLockerEx x(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag); 2780 bool result = verification_mark_bm()->isMarked(addr); 2781 gclog_or_tty->print_cr("TransitiveMark: Address 0x%x %s marked", addr, 2782 result ? "IS" : "is NOT"); 2783 return result; 2784 } else { 2785 gclog_or_tty->print_cr("Could not compute result"); 2786 return false; 2787 } 2788 } 2789 2790 //////////////////////////////////////////////////////// 2791 // CMS Verification Support 2792 //////////////////////////////////////////////////////// 2793 // Following the remark phase, the following invariant 2794 // should hold -- each object in the CMS heap which is 2795 // marked in markBitMap() should be marked in the verification_mark_bm(). 2796 2797 class VerifyMarkedClosure: public BitMapClosure { 2798 CMSBitMap* _marks; 2799 bool _failed; 2800 2801 public: 2802 VerifyMarkedClosure(CMSBitMap* bm): _marks(bm), _failed(false) {} 2803 2804 bool do_bit(size_t offset) { 2805 HeapWord* addr = _marks->offsetToHeapWord(offset); 2806 if (!_marks->isMarked(addr)) { 2807 oop(addr)->print_on(gclog_or_tty); 2808 gclog_or_tty->print_cr(" ("INTPTR_FORMAT" should have been marked)", addr); 2809 _failed = true; 2810 } 2811 return true; 2812 } 2813 2814 bool failed() { return _failed; } 2815 }; 2816 2817 bool CMSCollector::verify_after_remark() { 2818 gclog_or_tty->print(" [Verifying CMS Marking... "); 2819 MutexLockerEx ml(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag); 2820 static bool init = false; 2821 2822 assert(SafepointSynchronize::is_at_safepoint(), 2823 "Else mutations in object graph will make answer suspect"); 2824 assert(have_cms_token(), 2825 "Else there may be mutual interference in use of " 2826 " verification data structures"); 2827 assert(_collectorState > Marking && _collectorState <= Sweeping, 2828 "Else marking info checked here may be obsolete"); 2829 assert(haveFreelistLocks(), "must hold free list locks"); 2830 assert_lock_strong(bitMapLock()); 2831 2832 2833 // Allocate marking bit map if not already allocated 2834 if (!init) { // first time 2835 if (!verification_mark_bm()->allocate(_span)) { 2836 return false; 2837 } 2838 init = true; 2839 } 2840 2841 assert(verification_mark_stack()->isEmpty(), "Should be empty"); 2842 2843 // Turn off refs discovery -- so we will be tracing through refs. 2844 // This is as intended, because by this time 2845 // GC must already have cleared any refs that need to be cleared, 2846 // and traced those that need to be marked; moreover, 2847 // the marking done here is not going to intefere in any 2848 // way with the marking information used by GC. 2849 NoRefDiscovery no_discovery(ref_processor()); 2850 2851 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;) 2852 2853 // Clear any marks from a previous round 2854 verification_mark_bm()->clear_all(); 2855 assert(verification_mark_stack()->isEmpty(), "markStack should be empty"); 2856 verify_work_stacks_empty(); 2857 2858 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2859 gch->ensure_parsability(false); // fill TLABs, but no need to retire them 2860 // Update the saved marks which may affect the root scans. 2861 gch->save_marks(); 2862 2863 if (CMSRemarkVerifyVariant == 1) { 2864 // In this first variant of verification, we complete 2865 // all marking, then check if the new marks-verctor is 2866 // a subset of the CMS marks-vector. 2867 verify_after_remark_work_1(); 2868 } else if (CMSRemarkVerifyVariant == 2) { 2869 // In this second variant of verification, we flag an error 2870 // (i.e. an object reachable in the new marks-vector not reachable 2871 // in the CMS marks-vector) immediately, also indicating the 2872 // identify of an object (A) that references the unmarked object (B) -- 2873 // presumably, a mutation to A failed to be picked up by preclean/remark? 2874 verify_after_remark_work_2(); 2875 } else { 2876 warning("Unrecognized value %d for CMSRemarkVerifyVariant", 2877 CMSRemarkVerifyVariant); 2878 } 2879 gclog_or_tty->print(" done] "); 2880 return true; 2881 } 2882 2883 void CMSCollector::verify_after_remark_work_1() { 2884 ResourceMark rm; 2885 HandleMark hm; 2886 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2887 2888 // Get a clear set of claim bits for the strong roots processing to work with. 2889 ClassLoaderDataGraph::clear_claimed_marks(); 2890 2891 // Mark from roots one level into CMS 2892 MarkRefsIntoClosure notOlder(_span, verification_mark_bm()); 2893 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2894 2895 gch->gen_process_strong_roots(_cmsGen->level(), 2896 true, // younger gens are roots 2897 true, // activate StrongRootsScope 2898 false, // not scavenging 2899 SharedHeap::ScanningOption(roots_scanning_options()), 2900 ¬Older, 2901 true, // walk code active on stacks 2902 NULL, 2903 NULL); // SSS: Provide correct closure 2904 2905 // Now mark from the roots 2906 MarkFromRootsClosure markFromRootsClosure(this, _span, 2907 verification_mark_bm(), verification_mark_stack(), 2908 false /* don't yield */, true /* verifying */); 2909 assert(_restart_addr == NULL, "Expected pre-condition"); 2910 verification_mark_bm()->iterate(&markFromRootsClosure); 2911 while (_restart_addr != NULL) { 2912 // Deal with stack overflow: by restarting at the indicated 2913 // address. 2914 HeapWord* ra = _restart_addr; 2915 markFromRootsClosure.reset(ra); 2916 _restart_addr = NULL; 2917 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end()); 2918 } 2919 assert(verification_mark_stack()->isEmpty(), "Should have been drained"); 2920 verify_work_stacks_empty(); 2921 2922 // Marking completed -- now verify that each bit marked in 2923 // verification_mark_bm() is also marked in markBitMap(); flag all 2924 // errors by printing corresponding objects. 2925 VerifyMarkedClosure vcl(markBitMap()); 2926 verification_mark_bm()->iterate(&vcl); 2927 if (vcl.failed()) { 2928 gclog_or_tty->print("Verification failed"); 2929 Universe::heap()->print_on(gclog_or_tty); 2930 fatal("CMS: failed marking verification after remark"); 2931 } 2932 } 2933 2934 class VerifyKlassOopsKlassClosure : public KlassClosure { 2935 class VerifyKlassOopsClosure : public OopClosure { 2936 CMSBitMap* _bitmap; 2937 public: 2938 VerifyKlassOopsClosure(CMSBitMap* bitmap) : _bitmap(bitmap) { } 2939 void do_oop(oop* p) { guarantee(*p == NULL || _bitmap->isMarked((HeapWord*) *p), "Should be marked"); } 2940 void do_oop(narrowOop* p) { ShouldNotReachHere(); } 2941 } _oop_closure; 2942 public: 2943 VerifyKlassOopsKlassClosure(CMSBitMap* bitmap) : _oop_closure(bitmap) {} 2944 void do_klass(Klass* k) { 2945 k->oops_do(&_oop_closure); 2946 } 2947 }; 2948 2949 void CMSCollector::verify_after_remark_work_2() { 2950 ResourceMark rm; 2951 HandleMark hm; 2952 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2953 2954 // Get a clear set of claim bits for the strong roots processing to work with. 2955 ClassLoaderDataGraph::clear_claimed_marks(); 2956 2957 // Mark from roots one level into CMS 2958 MarkRefsIntoVerifyClosure notOlder(_span, verification_mark_bm(), 2959 markBitMap()); 2960 CMKlassClosure klass_closure(¬Older); 2961 2962 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2963 gch->gen_process_strong_roots(_cmsGen->level(), 2964 true, // younger gens are roots 2965 true, // activate StrongRootsScope 2966 false, // not scavenging 2967 SharedHeap::ScanningOption(roots_scanning_options()), 2968 ¬Older, 2969 true, // walk code active on stacks 2970 NULL, 2971 &klass_closure); 2972 2973 // Now mark from the roots 2974 MarkFromRootsVerifyClosure markFromRootsClosure(this, _span, 2975 verification_mark_bm(), markBitMap(), verification_mark_stack()); 2976 assert(_restart_addr == NULL, "Expected pre-condition"); 2977 verification_mark_bm()->iterate(&markFromRootsClosure); 2978 while (_restart_addr != NULL) { 2979 // Deal with stack overflow: by restarting at the indicated 2980 // address. 2981 HeapWord* ra = _restart_addr; 2982 markFromRootsClosure.reset(ra); 2983 _restart_addr = NULL; 2984 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end()); 2985 } 2986 assert(verification_mark_stack()->isEmpty(), "Should have been drained"); 2987 verify_work_stacks_empty(); 2988 2989 VerifyKlassOopsKlassClosure verify_klass_oops(verification_mark_bm()); 2990 ClassLoaderDataGraph::classes_do(&verify_klass_oops); 2991 2992 // Marking completed -- now verify that each bit marked in 2993 // verification_mark_bm() is also marked in markBitMap(); flag all 2994 // errors by printing corresponding objects. 2995 VerifyMarkedClosure vcl(markBitMap()); 2996 verification_mark_bm()->iterate(&vcl); 2997 assert(!vcl.failed(), "Else verification above should not have succeeded"); 2998 } 2999 3000 void ConcurrentMarkSweepGeneration::save_marks() { 3001 // delegate to CMS space 3002 cmsSpace()->save_marks(); 3003 for (uint i = 0; i < ParallelGCThreads; i++) { 3004 _par_gc_thread_states[i]->promo.startTrackingPromotions(); 3005 } 3006 } 3007 3008 bool ConcurrentMarkSweepGeneration::no_allocs_since_save_marks() { 3009 return cmsSpace()->no_allocs_since_save_marks(); 3010 } 3011 3012 #define CMS_SINCE_SAVE_MARKS_DEFN(OopClosureType, nv_suffix) \ 3013 \ 3014 void ConcurrentMarkSweepGeneration:: \ 3015 oop_since_save_marks_iterate##nv_suffix(OopClosureType* cl) { \ 3016 cl->set_generation(this); \ 3017 cmsSpace()->oop_since_save_marks_iterate##nv_suffix(cl); \ 3018 cl->reset_generation(); \ 3019 save_marks(); \ 3020 } 3021 3022 ALL_SINCE_SAVE_MARKS_CLOSURES(CMS_SINCE_SAVE_MARKS_DEFN) 3023 3024 void 3025 ConcurrentMarkSweepGeneration::object_iterate_since_last_GC(ObjectClosure* blk) 3026 { 3027 // Not currently implemented; need to do the following. -- ysr. 3028 // dld -- I think that is used for some sort of allocation profiler. So it 3029 // really means the objects allocated by the mutator since the last 3030 // GC. We could potentially implement this cheaply by recording only 3031 // the direct allocations in a side data structure. 3032 // 3033 // I think we probably ought not to be required to support these 3034 // iterations at any arbitrary point; I think there ought to be some 3035 // call to enable/disable allocation profiling in a generation/space, 3036 // and the iterator ought to return the objects allocated in the 3037 // gen/space since the enable call, or the last iterator call (which 3038 // will probably be at a GC.) That way, for gens like CM&S that would 3039 // require some extra data structure to support this, we only pay the 3040 // cost when it's in use... 3041 cmsSpace()->object_iterate_since_last_GC(blk); 3042 } 3043 3044 void 3045 ConcurrentMarkSweepGeneration::younger_refs_iterate(OopsInGenClosure* cl) { 3046 cl->set_generation(this); 3047 younger_refs_in_space_iterate(_cmsSpace, cl); 3048 cl->reset_generation(); 3049 } 3050 3051 void 3052 ConcurrentMarkSweepGeneration::oop_iterate(MemRegion mr, ExtendedOopClosure* cl) { 3053 if (freelistLock()->owned_by_self()) { 3054 Generation::oop_iterate(mr, cl); 3055 } else { 3056 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 3057 Generation::oop_iterate(mr, cl); 3058 } 3059 } 3060 3061 void 3062 ConcurrentMarkSweepGeneration::oop_iterate(ExtendedOopClosure* cl) { 3063 if (freelistLock()->owned_by_self()) { 3064 Generation::oop_iterate(cl); 3065 } else { 3066 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 3067 Generation::oop_iterate(cl); 3068 } 3069 } 3070 3071 void 3072 ConcurrentMarkSweepGeneration::object_iterate(ObjectClosure* cl) { 3073 if (freelistLock()->owned_by_self()) { 3074 Generation::object_iterate(cl); 3075 } else { 3076 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 3077 Generation::object_iterate(cl); 3078 } 3079 } 3080 3081 void 3082 ConcurrentMarkSweepGeneration::safe_object_iterate(ObjectClosure* cl) { 3083 if (freelistLock()->owned_by_self()) { 3084 Generation::safe_object_iterate(cl); 3085 } else { 3086 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 3087 Generation::safe_object_iterate(cl); 3088 } 3089 } 3090 3091 void 3092 ConcurrentMarkSweepGeneration::post_compact() { 3093 } 3094 3095 void 3096 ConcurrentMarkSweepGeneration::prepare_for_verify() { 3097 // Fix the linear allocation blocks to look like free blocks. 3098 3099 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those 3100 // are not called when the heap is verified during universe initialization and 3101 // at vm shutdown. 3102 if (freelistLock()->owned_by_self()) { 3103 cmsSpace()->prepare_for_verify(); 3104 } else { 3105 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag); 3106 cmsSpace()->prepare_for_verify(); 3107 } 3108 } 3109 3110 void 3111 ConcurrentMarkSweepGeneration::verify() { 3112 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those 3113 // are not called when the heap is verified during universe initialization and 3114 // at vm shutdown. 3115 if (freelistLock()->owned_by_self()) { 3116 cmsSpace()->verify(); 3117 } else { 3118 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag); 3119 cmsSpace()->verify(); 3120 } 3121 } 3122 3123 void CMSCollector::verify() { 3124 _cmsGen->verify(); 3125 } 3126 3127 #ifndef PRODUCT 3128 bool CMSCollector::overflow_list_is_empty() const { 3129 assert(_num_par_pushes >= 0, "Inconsistency"); 3130 if (_overflow_list == NULL) { 3131 assert(_num_par_pushes == 0, "Inconsistency"); 3132 } 3133 return _overflow_list == NULL; 3134 } 3135 3136 // The methods verify_work_stacks_empty() and verify_overflow_empty() 3137 // merely consolidate assertion checks that appear to occur together frequently. 3138 void CMSCollector::verify_work_stacks_empty() const { 3139 assert(_markStack.isEmpty(), "Marking stack should be empty"); 3140 assert(overflow_list_is_empty(), "Overflow list should be empty"); 3141 } 3142 3143 void CMSCollector::verify_overflow_empty() const { 3144 assert(overflow_list_is_empty(), "Overflow list should be empty"); 3145 assert(no_preserved_marks(), "No preserved marks"); 3146 } 3147 #endif // PRODUCT 3148 3149 // Decide if we want to enable class unloading as part of the 3150 // ensuing concurrent GC cycle. We will collect and 3151 // unload classes if it's the case that: 3152 // (1) an explicit gc request has been made and the flag 3153 // ExplicitGCInvokesConcurrentAndUnloadsClasses is set, OR 3154 // (2) (a) class unloading is enabled at the command line, and 3155 // (b) old gen is getting really full 3156 // NOTE: Provided there is no change in the state of the heap between 3157 // calls to this method, it should have idempotent results. Moreover, 3158 // its results should be monotonically increasing (i.e. going from 0 to 1, 3159 // but not 1 to 0) between successive calls between which the heap was 3160 // not collected. For the implementation below, it must thus rely on 3161 // the property that concurrent_cycles_since_last_unload() 3162 // will not decrease unless a collection cycle happened and that 3163 // _cmsGen->is_too_full() are 3164 // themselves also monotonic in that sense. See check_monotonicity() 3165 // below. 3166 void CMSCollector::update_should_unload_classes() { 3167 _should_unload_classes = false; 3168 // Condition 1 above 3169 if (_full_gc_requested && ExplicitGCInvokesConcurrentAndUnloadsClasses) { 3170 _should_unload_classes = true; 3171 } else if (CMSClassUnloadingEnabled) { // Condition 2.a above 3172 // Disjuncts 2.b.(i,ii,iii) above 3173 _should_unload_classes = (concurrent_cycles_since_last_unload() >= 3174 CMSClassUnloadingMaxInterval) 3175 || _cmsGen->is_too_full(); 3176 } 3177 } 3178 3179 bool ConcurrentMarkSweepGeneration::is_too_full() const { 3180 bool res = should_concurrent_collect(); 3181 res = res && (occupancy() > (double)CMSIsTooFullPercentage/100.0); 3182 return res; 3183 } 3184 3185 void CMSCollector::setup_cms_unloading_and_verification_state() { 3186 const bool should_verify = VerifyBeforeGC || VerifyAfterGC || VerifyDuringGC 3187 || VerifyBeforeExit; 3188 const int rso = SharedHeap::SO_Strings | SharedHeap::SO_CodeCache; 3189 3190 if (should_unload_classes()) { // Should unload classes this cycle 3191 remove_root_scanning_option(rso); // Shrink the root set appropriately 3192 set_verifying(should_verify); // Set verification state for this cycle 3193 return; // Nothing else needs to be done at this time 3194 } 3195 3196 // Not unloading classes this cycle 3197 assert(!should_unload_classes(), "Inconsitency!"); 3198 if ((!verifying() || unloaded_classes_last_cycle()) && should_verify) { 3199 // Include symbols, strings and code cache elements to prevent their resurrection. 3200 add_root_scanning_option(rso); 3201 set_verifying(true); 3202 } else if (verifying() && !should_verify) { 3203 // We were verifying, but some verification flags got disabled. 3204 set_verifying(false); 3205 // Exclude symbols, strings and code cache elements from root scanning to 3206 // reduce IM and RM pauses. 3207 remove_root_scanning_option(rso); 3208 } 3209 } 3210 3211 3212 #ifndef PRODUCT 3213 HeapWord* CMSCollector::block_start(const void* p) const { 3214 const HeapWord* addr = (HeapWord*)p; 3215 if (_span.contains(p)) { 3216 if (_cmsGen->cmsSpace()->is_in_reserved(addr)) { 3217 return _cmsGen->cmsSpace()->block_start(p); 3218 } 3219 } 3220 return NULL; 3221 } 3222 #endif 3223 3224 HeapWord* 3225 ConcurrentMarkSweepGeneration::expand_and_allocate(size_t word_size, 3226 bool tlab, 3227 bool parallel) { 3228 CMSSynchronousYieldRequest yr; 3229 assert(!tlab, "Can't deal with TLAB allocation"); 3230 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 3231 expand(word_size*HeapWordSize, MinHeapDeltaBytes, 3232 CMSExpansionCause::_satisfy_allocation); 3233 if (GCExpandToAllocateDelayMillis > 0) { 3234 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 3235 } 3236 return have_lock_and_allocate(word_size, tlab); 3237 } 3238 3239 // YSR: All of this generation expansion/shrinking stuff is an exact copy of 3240 // OneContigSpaceCardGeneration, which makes me wonder if we should move this 3241 // to CardGeneration and share it... 3242 bool ConcurrentMarkSweepGeneration::expand(size_t bytes, size_t expand_bytes) { 3243 return CardGeneration::expand(bytes, expand_bytes); 3244 } 3245 3246 void ConcurrentMarkSweepGeneration::expand(size_t bytes, size_t expand_bytes, 3247 CMSExpansionCause::Cause cause) 3248 { 3249 3250 bool success = expand(bytes, expand_bytes); 3251 3252 // remember why we expanded; this information is used 3253 // by shouldConcurrentCollect() when making decisions on whether to start 3254 // a new CMS cycle. 3255 if (success) { 3256 set_expansion_cause(cause); 3257 if (PrintGCDetails && Verbose) { 3258 gclog_or_tty->print_cr("Expanded CMS gen for %s", 3259 CMSExpansionCause::to_string(cause)); 3260 } 3261 } 3262 } 3263 3264 HeapWord* ConcurrentMarkSweepGeneration::expand_and_par_lab_allocate(CMSParGCThreadState* ps, size_t word_sz) { 3265 HeapWord* res = NULL; 3266 MutexLocker x(ParGCRareEvent_lock); 3267 while (true) { 3268 // Expansion by some other thread might make alloc OK now: 3269 res = ps->lab.alloc(word_sz); 3270 if (res != NULL) return res; 3271 // If there's not enough expansion space available, give up. 3272 if (_virtual_space.uncommitted_size() < (word_sz * HeapWordSize)) { 3273 return NULL; 3274 } 3275 // Otherwise, we try expansion. 3276 expand(word_sz*HeapWordSize, MinHeapDeltaBytes, 3277 CMSExpansionCause::_allocate_par_lab); 3278 // Now go around the loop and try alloc again; 3279 // A competing par_promote might beat us to the expansion space, 3280 // so we may go around the loop again if promotion fails agaion. 3281 if (GCExpandToAllocateDelayMillis > 0) { 3282 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 3283 } 3284 } 3285 } 3286 3287 3288 bool ConcurrentMarkSweepGeneration::expand_and_ensure_spooling_space( 3289 PromotionInfo* promo) { 3290 MutexLocker x(ParGCRareEvent_lock); 3291 size_t refill_size_bytes = promo->refillSize() * HeapWordSize; 3292 while (true) { 3293 // Expansion by some other thread might make alloc OK now: 3294 if (promo->ensure_spooling_space()) { 3295 assert(promo->has_spooling_space(), 3296 "Post-condition of successful ensure_spooling_space()"); 3297 return true; 3298 } 3299 // If there's not enough expansion space available, give up. 3300 if (_virtual_space.uncommitted_size() < refill_size_bytes) { 3301 return false; 3302 } 3303 // Otherwise, we try expansion. 3304 expand(refill_size_bytes, MinHeapDeltaBytes, 3305 CMSExpansionCause::_allocate_par_spooling_space); 3306 // Now go around the loop and try alloc again; 3307 // A competing allocation might beat us to the expansion space, 3308 // so we may go around the loop again if allocation fails again. 3309 if (GCExpandToAllocateDelayMillis > 0) { 3310 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 3311 } 3312 } 3313 } 3314 3315 3316 3317 void ConcurrentMarkSweepGeneration::shrink(size_t bytes) { 3318 assert_locked_or_safepoint(Heap_lock); 3319 size_t size = ReservedSpace::page_align_size_down(bytes); 3320 if (size > 0) { 3321 shrink_by(size); 3322 } 3323 } 3324 3325 bool ConcurrentMarkSweepGeneration::grow_by(size_t bytes) { 3326 assert_locked_or_safepoint(Heap_lock); 3327 bool result = _virtual_space.expand_by(bytes); 3328 if (result) { 3329 HeapWord* old_end = _cmsSpace->end(); 3330 size_t new_word_size = 3331 heap_word_size(_virtual_space.committed_size()); 3332 MemRegion mr(_cmsSpace->bottom(), new_word_size); 3333 _bts->resize(new_word_size); // resize the block offset shared array 3334 Universe::heap()->barrier_set()->resize_covered_region(mr); 3335 // Hmmmm... why doesn't CFLS::set_end verify locking? 3336 // This is quite ugly; FIX ME XXX 3337 _cmsSpace->assert_locked(freelistLock()); 3338 _cmsSpace->set_end((HeapWord*)_virtual_space.high()); 3339 3340 // update the space and generation capacity counters 3341 if (UsePerfData) { 3342 _space_counters->update_capacity(); 3343 _gen_counters->update_all(); 3344 } 3345 3346 if (Verbose && PrintGC) { 3347 size_t new_mem_size = _virtual_space.committed_size(); 3348 size_t old_mem_size = new_mem_size - bytes; 3349 gclog_or_tty->print_cr("Expanding %s from %ldK by %ldK to %ldK", 3350 name(), old_mem_size/K, bytes/K, new_mem_size/K); 3351 } 3352 } 3353 return result; 3354 } 3355 3356 bool ConcurrentMarkSweepGeneration::grow_to_reserved() { 3357 assert_locked_or_safepoint(Heap_lock); 3358 bool success = true; 3359 const size_t remaining_bytes = _virtual_space.uncommitted_size(); 3360 if (remaining_bytes > 0) { 3361 success = grow_by(remaining_bytes); 3362 DEBUG_ONLY(if (!success) warning("grow to reserved failed");) 3363 } 3364 return success; 3365 } 3366 3367 void ConcurrentMarkSweepGeneration::shrink_by(size_t bytes) { 3368 assert_locked_or_safepoint(Heap_lock); 3369 assert_lock_strong(freelistLock()); 3370 // XXX Fix when compaction is implemented. 3371 warning("Shrinking of CMS not yet implemented"); 3372 return; 3373 } 3374 3375 3376 // Simple ctor/dtor wrapper for accounting & timer chores around concurrent 3377 // phases. 3378 class CMSPhaseAccounting: public StackObj { 3379 public: 3380 CMSPhaseAccounting(CMSCollector *collector, 3381 const char *phase, 3382 bool print_cr = true); 3383 ~CMSPhaseAccounting(); 3384 3385 private: 3386 CMSCollector *_collector; 3387 const char *_phase; 3388 elapsedTimer _wallclock; 3389 bool _print_cr; 3390 3391 public: 3392 // Not MT-safe; so do not pass around these StackObj's 3393 // where they may be accessed by other threads. 3394 jlong wallclock_millis() { 3395 assert(_wallclock.is_active(), "Wall clock should not stop"); 3396 _wallclock.stop(); // to record time 3397 jlong ret = _wallclock.milliseconds(); 3398 _wallclock.start(); // restart 3399 return ret; 3400 } 3401 }; 3402 3403 CMSPhaseAccounting::CMSPhaseAccounting(CMSCollector *collector, 3404 const char *phase, 3405 bool print_cr) : 3406 _collector(collector), _phase(phase), _print_cr(print_cr) { 3407 3408 if (PrintCMSStatistics != 0) { 3409 _collector->resetYields(); 3410 } 3411 if (PrintGCDetails && PrintGCTimeStamps) { 3412 gclog_or_tty->date_stamp(PrintGCDateStamps); 3413 gclog_or_tty->stamp(); 3414 gclog_or_tty->print_cr(": [%s-concurrent-%s-start]", 3415 _collector->cmsGen()->short_name(), _phase); 3416 } 3417 _collector->resetTimer(); 3418 _wallclock.start(); 3419 _collector->startTimer(); 3420 } 3421 3422 CMSPhaseAccounting::~CMSPhaseAccounting() { 3423 assert(_wallclock.is_active(), "Wall clock should not have stopped"); 3424 _collector->stopTimer(); 3425 _wallclock.stop(); 3426 if (PrintGCDetails) { 3427 gclog_or_tty->date_stamp(PrintGCDateStamps); 3428 if (PrintGCTimeStamps) { 3429 gclog_or_tty->stamp(); 3430 gclog_or_tty->print(": "); 3431 } 3432 gclog_or_tty->print("[%s-concurrent-%s: %3.3f/%3.3f secs]", 3433 _collector->cmsGen()->short_name(), 3434 _phase, _collector->timerValue(), _wallclock.seconds()); 3435 if (_print_cr) { 3436 gclog_or_tty->print_cr(""); 3437 } 3438 if (PrintCMSStatistics != 0) { 3439 gclog_or_tty->print_cr(" (CMS-concurrent-%s yielded %d times)", _phase, 3440 _collector->yields()); 3441 } 3442 } 3443 } 3444 3445 // CMS work 3446 3447 // Checkpoint the roots into this generation from outside 3448 // this generation. [Note this initial checkpoint need only 3449 // be approximate -- we'll do a catch up phase subsequently.] 3450 void CMSCollector::checkpointRootsInitial(bool asynch) { 3451 assert(_collectorState == InitialMarking, "Wrong collector state"); 3452 check_correct_thread_executing(); 3453 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 3454 3455 ReferenceProcessor* rp = ref_processor(); 3456 SpecializationStats::clear(); 3457 assert(_restart_addr == NULL, "Control point invariant"); 3458 if (asynch) { 3459 // acquire locks for subsequent manipulations 3460 MutexLockerEx x(bitMapLock(), 3461 Mutex::_no_safepoint_check_flag); 3462 checkpointRootsInitialWork(asynch); 3463 // enable ("weak") refs discovery 3464 rp->enable_discovery(true /*verify_disabled*/, true /*check_no_refs*/); 3465 _collectorState = Marking; 3466 } else { 3467 // (Weak) Refs discovery: this is controlled from genCollectedHeap::do_collection 3468 // which recognizes if we are a CMS generation, and doesn't try to turn on 3469 // discovery; verify that they aren't meddling. 3470 assert(!rp->discovery_is_atomic(), 3471 "incorrect setting of discovery predicate"); 3472 assert(!rp->discovery_enabled(), "genCollectedHeap shouldn't control " 3473 "ref discovery for this generation kind"); 3474 // already have locks 3475 checkpointRootsInitialWork(asynch); 3476 // now enable ("weak") refs discovery 3477 rp->enable_discovery(true /*verify_disabled*/, false /*verify_no_refs*/); 3478 _collectorState = Marking; 3479 } 3480 SpecializationStats::print(); 3481 } 3482 3483 void CMSCollector::checkpointRootsInitialWork(bool asynch) { 3484 assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped"); 3485 assert(_collectorState == InitialMarking, "just checking"); 3486 3487 // If there has not been a GC[n-1] since last GC[n] cycle completed, 3488 // precede our marking with a collection of all 3489 // younger generations to keep floating garbage to a minimum. 3490 // XXX: we won't do this for now -- it's an optimization to be done later. 3491 3492 // already have locks 3493 assert_lock_strong(bitMapLock()); 3494 assert(_markBitMap.isAllClear(), "was reset at end of previous cycle"); 3495 3496 // Setup the verification and class unloading state for this 3497 // CMS collection cycle. 3498 setup_cms_unloading_and_verification_state(); 3499 3500 NOT_PRODUCT(TraceTime t("\ncheckpointRootsInitialWork", 3501 PrintGCDetails && Verbose, true, gclog_or_tty);) 3502 if (UseAdaptiveSizePolicy) { 3503 size_policy()->checkpoint_roots_initial_begin(); 3504 } 3505 3506 // Reset all the PLAB chunk arrays if necessary. 3507 if (_survivor_plab_array != NULL && !CMSPLABRecordAlways) { 3508 reset_survivor_plab_arrays(); 3509 } 3510 3511 ResourceMark rm; 3512 HandleMark hm; 3513 3514 FalseClosure falseClosure; 3515 // In the case of a synchronous collection, we will elide the 3516 // remark step, so it's important to catch all the nmethod oops 3517 // in this step. 3518 // The final 'true' flag to gen_process_strong_roots will ensure this. 3519 // If 'async' is true, we can relax the nmethod tracing. 3520 MarkRefsIntoClosure notOlder(_span, &_markBitMap); 3521 GenCollectedHeap* gch = GenCollectedHeap::heap(); 3522 3523 verify_work_stacks_empty(); 3524 verify_overflow_empty(); 3525 3526 gch->ensure_parsability(false); // fill TLABs, but no need to retire them 3527 // Update the saved marks which may affect the root scans. 3528 gch->save_marks(); 3529 3530 // weak reference processing has not started yet. 3531 ref_processor()->set_enqueuing_is_done(false); 3532 3533 // Need to remember all newly created CLDs, 3534 // so that we can guarantee that the remark finds them. 3535 ClassLoaderDataGraph::remember_new_clds(true); 3536 3537 // Whenever a CLD is found, it will be claimed before proceeding to mark 3538 // the klasses. The claimed marks need to be cleared before marking starts. 3539 ClassLoaderDataGraph::clear_claimed_marks(); 3540 3541 CMKlassClosure klass_closure(¬Older); 3542 { 3543 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;) 3544 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 3545 gch->gen_process_strong_roots(_cmsGen->level(), 3546 true, // younger gens are roots 3547 true, // activate StrongRootsScope 3548 false, // not scavenging 3549 SharedHeap::ScanningOption(roots_scanning_options()), 3550 ¬Older, 3551 true, // walk all of code cache if (so & SO_CodeCache) 3552 NULL, 3553 &klass_closure); 3554 } 3555 3556 // Clear mod-union table; it will be dirtied in the prologue of 3557 // CMS generation per each younger generation collection. 3558 3559 assert(_modUnionTable.isAllClear(), 3560 "Was cleared in most recent final checkpoint phase" 3561 " or no bits are set in the gc_prologue before the start of the next " 3562 "subsequent marking phase."); 3563 3564 assert(_ct->klass_rem_set()->mod_union_is_clear(), "Must be"); 3565 3566 // Save the end of the used_region of the constituent generations 3567 // to be used to limit the extent of sweep in each generation. 3568 save_sweep_limits(); 3569 if (UseAdaptiveSizePolicy) { 3570 size_policy()->checkpoint_roots_initial_end(gch->gc_cause()); 3571 } 3572 verify_overflow_empty(); 3573 } 3574 3575 bool CMSCollector::markFromRoots(bool asynch) { 3576 // we might be tempted to assert that: 3577 // assert(asynch == !SafepointSynchronize::is_at_safepoint(), 3578 // "inconsistent argument?"); 3579 // However that wouldn't be right, because it's possible that 3580 // a safepoint is indeed in progress as a younger generation 3581 // stop-the-world GC happens even as we mark in this generation. 3582 assert(_collectorState == Marking, "inconsistent state?"); 3583 check_correct_thread_executing(); 3584 verify_overflow_empty(); 3585 3586 bool res; 3587 if (asynch) { 3588 3589 // Start the timers for adaptive size policy for the concurrent phases 3590 // Do it here so that the foreground MS can use the concurrent 3591 // timer since a foreground MS might has the sweep done concurrently 3592 // or STW. 3593 if (UseAdaptiveSizePolicy) { 3594 size_policy()->concurrent_marking_begin(); 3595 } 3596 3597 // Weak ref discovery note: We may be discovering weak 3598 // refs in this generation concurrent (but interleaved) with 3599 // weak ref discovery by a younger generation collector. 3600 3601 CMSTokenSyncWithLocks ts(true, bitMapLock()); 3602 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 3603 CMSPhaseAccounting pa(this, "mark", !PrintGCDetails); 3604 res = markFromRootsWork(asynch); 3605 if (res) { 3606 _collectorState = Precleaning; 3607 } else { // We failed and a foreground collection wants to take over 3608 assert(_foregroundGCIsActive, "internal state inconsistency"); 3609 assert(_restart_addr == NULL, "foreground will restart from scratch"); 3610 if (PrintGCDetails) { 3611 gclog_or_tty->print_cr("bailing out to foreground collection"); 3612 } 3613 } 3614 if (UseAdaptiveSizePolicy) { 3615 size_policy()->concurrent_marking_end(); 3616 } 3617 } else { 3618 assert(SafepointSynchronize::is_at_safepoint(), 3619 "inconsistent with asynch == false"); 3620 if (UseAdaptiveSizePolicy) { 3621 size_policy()->ms_collection_marking_begin(); 3622 } 3623 // already have locks 3624 res = markFromRootsWork(asynch); 3625 _collectorState = FinalMarking; 3626 if (UseAdaptiveSizePolicy) { 3627 GenCollectedHeap* gch = GenCollectedHeap::heap(); 3628 size_policy()->ms_collection_marking_end(gch->gc_cause()); 3629 } 3630 } 3631 verify_overflow_empty(); 3632 return res; 3633 } 3634 3635 bool CMSCollector::markFromRootsWork(bool asynch) { 3636 // iterate over marked bits in bit map, doing a full scan and mark 3637 // from these roots using the following algorithm: 3638 // . if oop is to the right of the current scan pointer, 3639 // mark corresponding bit (we'll process it later) 3640 // . else (oop is to left of current scan pointer) 3641 // push oop on marking stack 3642 // . drain the marking stack 3643 3644 // Note that when we do a marking step we need to hold the 3645 // bit map lock -- recall that direct allocation (by mutators) 3646 // and promotion (by younger generation collectors) is also 3647 // marking the bit map. [the so-called allocate live policy.] 3648 // Because the implementation of bit map marking is not 3649 // robust wrt simultaneous marking of bits in the same word, 3650 // we need to make sure that there is no such interference 3651 // between concurrent such updates. 3652 3653 // already have locks 3654 assert_lock_strong(bitMapLock()); 3655 3656 verify_work_stacks_empty(); 3657 verify_overflow_empty(); 3658 bool result = false; 3659 if (CMSConcurrentMTEnabled && ConcGCThreads > 0) { 3660 result = do_marking_mt(asynch); 3661 } else { 3662 result = do_marking_st(asynch); 3663 } 3664 return result; 3665 } 3666 3667 // Forward decl 3668 class CMSConcMarkingTask; 3669 3670 class CMSConcMarkingTerminator: public ParallelTaskTerminator { 3671 CMSCollector* _collector; 3672 CMSConcMarkingTask* _task; 3673 public: 3674 virtual void yield(); 3675 3676 // "n_threads" is the number of threads to be terminated. 3677 // "queue_set" is a set of work queues of other threads. 3678 // "collector" is the CMS collector associated with this task terminator. 3679 // "yield" indicates whether we need the gang as a whole to yield. 3680 CMSConcMarkingTerminator(int n_threads, TaskQueueSetSuper* queue_set, CMSCollector* collector) : 3681 ParallelTaskTerminator(n_threads, queue_set), 3682 _collector(collector) { } 3683 3684 void set_task(CMSConcMarkingTask* task) { 3685 _task = task; 3686 } 3687 }; 3688 3689 class CMSConcMarkingTerminatorTerminator: public TerminatorTerminator { 3690 CMSConcMarkingTask* _task; 3691 public: 3692 bool should_exit_termination(); 3693 void set_task(CMSConcMarkingTask* task) { 3694 _task = task; 3695 } 3696 }; 3697 3698 // MT Concurrent Marking Task 3699 class CMSConcMarkingTask: public YieldingFlexibleGangTask { 3700 CMSCollector* _collector; 3701 int _n_workers; // requested/desired # workers 3702 bool _asynch; 3703 bool _result; 3704 CompactibleFreeListSpace* _cms_space; 3705 char _pad_front[64]; // padding to ... 3706 HeapWord* _global_finger; // ... avoid sharing cache line 3707 char _pad_back[64]; 3708 HeapWord* _restart_addr; 3709 3710 // Exposed here for yielding support 3711 Mutex* const _bit_map_lock; 3712 3713 // The per thread work queues, available here for stealing 3714 OopTaskQueueSet* _task_queues; 3715 3716 // Termination (and yielding) support 3717 CMSConcMarkingTerminator _term; 3718 CMSConcMarkingTerminatorTerminator _term_term; 3719 3720 public: 3721 CMSConcMarkingTask(CMSCollector* collector, 3722 CompactibleFreeListSpace* cms_space, 3723 bool asynch, 3724 YieldingFlexibleWorkGang* workers, 3725 OopTaskQueueSet* task_queues): 3726 YieldingFlexibleGangTask("Concurrent marking done multi-threaded"), 3727 _collector(collector), 3728 _cms_space(cms_space), 3729 _asynch(asynch), _n_workers(0), _result(true), 3730 _task_queues(task_queues), 3731 _term(_n_workers, task_queues, _collector), 3732 _bit_map_lock(collector->bitMapLock()) 3733 { 3734 _requested_size = _n_workers; 3735 _term.set_task(this); 3736 _term_term.set_task(this); 3737 _restart_addr = _global_finger = _cms_space->bottom(); 3738 } 3739 3740 3741 OopTaskQueueSet* task_queues() { return _task_queues; } 3742 3743 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 3744 3745 HeapWord** global_finger_addr() { return &_global_finger; } 3746 3747 CMSConcMarkingTerminator* terminator() { return &_term; } 3748 3749 virtual void set_for_termination(int active_workers) { 3750 terminator()->reset_for_reuse(active_workers); 3751 } 3752 3753 void work(uint worker_id); 3754 bool should_yield() { 3755 return ConcurrentMarkSweepThread::should_yield() 3756 && !_collector->foregroundGCIsActive() 3757 && _asynch; 3758 } 3759 3760 virtual void coordinator_yield(); // stuff done by coordinator 3761 bool result() { return _result; } 3762 3763 void reset(HeapWord* ra) { 3764 assert(_global_finger >= _cms_space->end(), "Postcondition of ::work(i)"); 3765 _restart_addr = _global_finger = ra; 3766 _term.reset_for_reuse(); 3767 } 3768 3769 static bool get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 3770 OopTaskQueue* work_q); 3771 3772 private: 3773 void do_scan_and_mark(int i, CompactibleFreeListSpace* sp); 3774 void do_work_steal(int i); 3775 void bump_global_finger(HeapWord* f); 3776 }; 3777 3778 bool CMSConcMarkingTerminatorTerminator::should_exit_termination() { 3779 assert(_task != NULL, "Error"); 3780 return _task->yielding(); 3781 // Note that we do not need the disjunct || _task->should_yield() above 3782 // because we want terminating threads to yield only if the task 3783 // is already in the midst of yielding, which happens only after at least one 3784 // thread has yielded. 3785 } 3786 3787 void CMSConcMarkingTerminator::yield() { 3788 if (_task->should_yield()) { 3789 _task->yield(); 3790 } else { 3791 ParallelTaskTerminator::yield(); 3792 } 3793 } 3794 3795 //////////////////////////////////////////////////////////////// 3796 // Concurrent Marking Algorithm Sketch 3797 //////////////////////////////////////////////////////////////// 3798 // Until all tasks exhausted (both spaces): 3799 // -- claim next available chunk 3800 // -- bump global finger via CAS 3801 // -- find first object that starts in this chunk 3802 // and start scanning bitmap from that position 3803 // -- scan marked objects for oops 3804 // -- CAS-mark target, and if successful: 3805 // . if target oop is above global finger (volatile read) 3806 // nothing to do 3807 // . if target oop is in chunk and above local finger 3808 // then nothing to do 3809 // . else push on work-queue 3810 // -- Deal with possible overflow issues: 3811 // . local work-queue overflow causes stuff to be pushed on 3812 // global (common) overflow queue 3813 // . always first empty local work queue 3814 // . then get a batch of oops from global work queue if any 3815 // . then do work stealing 3816 // -- When all tasks claimed (both spaces) 3817 // and local work queue empty, 3818 // then in a loop do: 3819 // . check global overflow stack; steal a batch of oops and trace 3820 // . try to steal from other threads oif GOS is empty 3821 // . if neither is available, offer termination 3822 // -- Terminate and return result 3823 // 3824 void CMSConcMarkingTask::work(uint worker_id) { 3825 elapsedTimer _timer; 3826 ResourceMark rm; 3827 HandleMark hm; 3828 3829 DEBUG_ONLY(_collector->verify_overflow_empty();) 3830 3831 // Before we begin work, our work queue should be empty 3832 assert(work_queue(worker_id)->size() == 0, "Expected to be empty"); 3833 // Scan the bitmap covering _cms_space, tracing through grey objects. 3834 _timer.start(); 3835 do_scan_and_mark(worker_id, _cms_space); 3836 _timer.stop(); 3837 if (PrintCMSStatistics != 0) { 3838 gclog_or_tty->print_cr("Finished cms space scanning in %dth thread: %3.3f sec", 3839 worker_id, _timer.seconds()); 3840 // XXX: need xxx/xxx type of notation, two timers 3841 } 3842 3843 // ... do work stealing 3844 _timer.reset(); 3845 _timer.start(); 3846 do_work_steal(worker_id); 3847 _timer.stop(); 3848 if (PrintCMSStatistics != 0) { 3849 gclog_or_tty->print_cr("Finished work stealing in %dth thread: %3.3f sec", 3850 worker_id, _timer.seconds()); 3851 // XXX: need xxx/xxx type of notation, two timers 3852 } 3853 assert(_collector->_markStack.isEmpty(), "Should have been emptied"); 3854 assert(work_queue(worker_id)->size() == 0, "Should have been emptied"); 3855 // Note that under the current task protocol, the 3856 // following assertion is true even of the spaces 3857 // expanded since the completion of the concurrent 3858 // marking. XXX This will likely change under a strict 3859 // ABORT semantics. 3860 // After perm removal the comparison was changed to 3861 // greater than or equal to from strictly greater than. 3862 // Before perm removal the highest address sweep would 3863 // have been at the end of perm gen but now is at the 3864 // end of the tenured gen. 3865 assert(_global_finger >= _cms_space->end(), 3866 "All tasks have been completed"); 3867 DEBUG_ONLY(_collector->verify_overflow_empty();) 3868 } 3869 3870 void CMSConcMarkingTask::bump_global_finger(HeapWord* f) { 3871 HeapWord* read = _global_finger; 3872 HeapWord* cur = read; 3873 while (f > read) { 3874 cur = read; 3875 read = (HeapWord*) Atomic::cmpxchg_ptr(f, &_global_finger, cur); 3876 if (cur == read) { 3877 // our cas succeeded 3878 assert(_global_finger >= f, "protocol consistency"); 3879 break; 3880 } 3881 } 3882 } 3883 3884 // This is really inefficient, and should be redone by 3885 // using (not yet available) block-read and -write interfaces to the 3886 // stack and the work_queue. XXX FIX ME !!! 3887 bool CMSConcMarkingTask::get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 3888 OopTaskQueue* work_q) { 3889 // Fast lock-free check 3890 if (ovflw_stk->length() == 0) { 3891 return false; 3892 } 3893 assert(work_q->size() == 0, "Shouldn't steal"); 3894 MutexLockerEx ml(ovflw_stk->par_lock(), 3895 Mutex::_no_safepoint_check_flag); 3896 // Grab up to 1/4 the size of the work queue 3897 size_t num = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 3898 (size_t)ParGCDesiredObjsFromOverflowList); 3899 num = MIN2(num, ovflw_stk->length()); 3900 for (int i = (int) num; i > 0; i--) { 3901 oop cur = ovflw_stk->pop(); 3902 assert(cur != NULL, "Counted wrong?"); 3903 work_q->push(cur); 3904 } 3905 return num > 0; 3906 } 3907 3908 void CMSConcMarkingTask::do_scan_and_mark(int i, CompactibleFreeListSpace* sp) { 3909 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 3910 int n_tasks = pst->n_tasks(); 3911 // We allow that there may be no tasks to do here because 3912 // we are restarting after a stack overflow. 3913 assert(pst->valid() || n_tasks == 0, "Uninitialized use?"); 3914 uint nth_task = 0; 3915 3916 HeapWord* aligned_start = sp->bottom(); 3917 if (sp->used_region().contains(_restart_addr)) { 3918 // Align down to a card boundary for the start of 0th task 3919 // for this space. 3920 aligned_start = 3921 (HeapWord*)align_size_down((uintptr_t)_restart_addr, 3922 CardTableModRefBS::card_size); 3923 } 3924 3925 size_t chunk_size = sp->marking_task_size(); 3926 while (!pst->is_task_claimed(/* reference */ nth_task)) { 3927 // Having claimed the nth task in this space, 3928 // compute the chunk that it corresponds to: 3929 MemRegion span = MemRegion(aligned_start + nth_task*chunk_size, 3930 aligned_start + (nth_task+1)*chunk_size); 3931 // Try and bump the global finger via a CAS; 3932 // note that we need to do the global finger bump 3933 // _before_ taking the intersection below, because 3934 // the task corresponding to that region will be 3935 // deemed done even if the used_region() expands 3936 // because of allocation -- as it almost certainly will 3937 // during start-up while the threads yield in the 3938 // closure below. 3939 HeapWord* finger = span.end(); 3940 bump_global_finger(finger); // atomically 3941 // There are null tasks here corresponding to chunks 3942 // beyond the "top" address of the space. 3943 span = span.intersection(sp->used_region()); 3944 if (!span.is_empty()) { // Non-null task 3945 HeapWord* prev_obj; 3946 assert(!span.contains(_restart_addr) || nth_task == 0, 3947 "Inconsistency"); 3948 if (nth_task == 0) { 3949 // For the 0th task, we'll not need to compute a block_start. 3950 if (span.contains(_restart_addr)) { 3951 // In the case of a restart because of stack overflow, 3952 // we might additionally skip a chunk prefix. 3953 prev_obj = _restart_addr; 3954 } else { 3955 prev_obj = span.start(); 3956 } 3957 } else { 3958 // We want to skip the first object because 3959 // the protocol is to scan any object in its entirety 3960 // that _starts_ in this span; a fortiori, any 3961 // object starting in an earlier span is scanned 3962 // as part of an earlier claimed task. 3963 // Below we use the "careful" version of block_start 3964 // so we do not try to navigate uninitialized objects. 3965 prev_obj = sp->block_start_careful(span.start()); 3966 // Below we use a variant of block_size that uses the 3967 // Printezis bits to avoid waiting for allocated 3968 // objects to become initialized/parsable. 3969 while (prev_obj < span.start()) { 3970 size_t sz = sp->block_size_no_stall(prev_obj, _collector); 3971 if (sz > 0) { 3972 prev_obj += sz; 3973 } else { 3974 // In this case we may end up doing a bit of redundant 3975 // scanning, but that appears unavoidable, short of 3976 // locking the free list locks; see bug 6324141. 3977 break; 3978 } 3979 } 3980 } 3981 if (prev_obj < span.end()) { 3982 MemRegion my_span = MemRegion(prev_obj, span.end()); 3983 // Do the marking work within a non-empty span -- 3984 // the last argument to the constructor indicates whether the 3985 // iteration should be incremental with periodic yields. 3986 Par_MarkFromRootsClosure cl(this, _collector, my_span, 3987 &_collector->_markBitMap, 3988 work_queue(i), 3989 &_collector->_markStack, 3990 _asynch); 3991 _collector->_markBitMap.iterate(&cl, my_span.start(), my_span.end()); 3992 } // else nothing to do for this task 3993 } // else nothing to do for this task 3994 } 3995 // We'd be tempted to assert here that since there are no 3996 // more tasks left to claim in this space, the global_finger 3997 // must exceed space->top() and a fortiori space->end(). However, 3998 // that would not quite be correct because the bumping of 3999 // global_finger occurs strictly after the claiming of a task, 4000 // so by the time we reach here the global finger may not yet 4001 // have been bumped up by the thread that claimed the last 4002 // task. 4003 pst->all_tasks_completed(); 4004 } 4005 4006 class Par_ConcMarkingClosure: public CMSOopClosure { 4007 private: 4008 CMSCollector* _collector; 4009 CMSConcMarkingTask* _task; 4010 MemRegion _span; 4011 CMSBitMap* _bit_map; 4012 CMSMarkStack* _overflow_stack; 4013 OopTaskQueue* _work_queue; 4014 protected: 4015 DO_OOP_WORK_DEFN 4016 public: 4017 Par_ConcMarkingClosure(CMSCollector* collector, CMSConcMarkingTask* task, OopTaskQueue* work_queue, 4018 CMSBitMap* bit_map, CMSMarkStack* overflow_stack): 4019 CMSOopClosure(collector->ref_processor()), 4020 _collector(collector), 4021 _task(task), 4022 _span(collector->_span), 4023 _work_queue(work_queue), 4024 _bit_map(bit_map), 4025 _overflow_stack(overflow_stack) 4026 { } 4027 virtual void do_oop(oop* p); 4028 virtual void do_oop(narrowOop* p); 4029 4030 void trim_queue(size_t max); 4031 void handle_stack_overflow(HeapWord* lost); 4032 void do_yield_check() { 4033 if (_task->should_yield()) { 4034 _task->yield(); 4035 } 4036 } 4037 }; 4038 4039 // Grey object scanning during work stealing phase -- 4040 // the salient assumption here is that any references 4041 // that are in these stolen objects being scanned must 4042 // already have been initialized (else they would not have 4043 // been published), so we do not need to check for 4044 // uninitialized objects before pushing here. 4045 void Par_ConcMarkingClosure::do_oop(oop obj) { 4046 assert(obj->is_oop_or_null(true), "expected an oop or NULL"); 4047 HeapWord* addr = (HeapWord*)obj; 4048 // Check if oop points into the CMS generation 4049 // and is not marked 4050 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 4051 // a white object ... 4052 // If we manage to "claim" the object, by being the 4053 // first thread to mark it, then we push it on our 4054 // marking stack 4055 if (_bit_map->par_mark(addr)) { // ... now grey 4056 // push on work queue (grey set) 4057 bool simulate_overflow = false; 4058 NOT_PRODUCT( 4059 if (CMSMarkStackOverflowALot && 4060 _collector->simulate_overflow()) { 4061 // simulate a stack overflow 4062 simulate_overflow = true; 4063 } 4064 ) 4065 if (simulate_overflow || 4066 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) { 4067 // stack overflow 4068 if (PrintCMSStatistics != 0) { 4069 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 4070 SIZE_FORMAT, _overflow_stack->capacity()); 4071 } 4072 // We cannot assert that the overflow stack is full because 4073 // it may have been emptied since. 4074 assert(simulate_overflow || 4075 _work_queue->size() == _work_queue->max_elems(), 4076 "Else push should have succeeded"); 4077 handle_stack_overflow(addr); 4078 } 4079 } // Else, some other thread got there first 4080 do_yield_check(); 4081 } 4082 } 4083 4084 void Par_ConcMarkingClosure::do_oop(oop* p) { Par_ConcMarkingClosure::do_oop_work(p); } 4085 void Par_ConcMarkingClosure::do_oop(narrowOop* p) { Par_ConcMarkingClosure::do_oop_work(p); } 4086 4087 void Par_ConcMarkingClosure::trim_queue(size_t max) { 4088 while (_work_queue->size() > max) { 4089 oop new_oop; 4090 if (_work_queue->pop_local(new_oop)) { 4091 assert(new_oop->is_oop(), "Should be an oop"); 4092 assert(_bit_map->isMarked((HeapWord*)new_oop), "Grey object"); 4093 assert(_span.contains((HeapWord*)new_oop), "Not in span"); 4094 new_oop->oop_iterate(this); // do_oop() above 4095 do_yield_check(); 4096 } 4097 } 4098 } 4099 4100 // Upon stack overflow, we discard (part of) the stack, 4101 // remembering the least address amongst those discarded 4102 // in CMSCollector's _restart_address. 4103 void Par_ConcMarkingClosure::handle_stack_overflow(HeapWord* lost) { 4104 // We need to do this under a mutex to prevent other 4105 // workers from interfering with the work done below. 4106 MutexLockerEx ml(_overflow_stack->par_lock(), 4107 Mutex::_no_safepoint_check_flag); 4108 // Remember the least grey address discarded 4109 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 4110 _collector->lower_restart_addr(ra); 4111 _overflow_stack->reset(); // discard stack contents 4112 _overflow_stack->expand(); // expand the stack if possible 4113 } 4114 4115 4116 void CMSConcMarkingTask::do_work_steal(int i) { 4117 OopTaskQueue* work_q = work_queue(i); 4118 oop obj_to_scan; 4119 CMSBitMap* bm = &(_collector->_markBitMap); 4120 CMSMarkStack* ovflw = &(_collector->_markStack); 4121 int* seed = _collector->hash_seed(i); 4122 Par_ConcMarkingClosure cl(_collector, this, work_q, bm, ovflw); 4123 while (true) { 4124 cl.trim_queue(0); 4125 assert(work_q->size() == 0, "Should have been emptied above"); 4126 if (get_work_from_overflow_stack(ovflw, work_q)) { 4127 // Can't assert below because the work obtained from the 4128 // overflow stack may already have been stolen from us. 4129 // assert(work_q->size() > 0, "Work from overflow stack"); 4130 continue; 4131 } else if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 4132 assert(obj_to_scan->is_oop(), "Should be an oop"); 4133 assert(bm->isMarked((HeapWord*)obj_to_scan), "Grey object"); 4134 obj_to_scan->oop_iterate(&cl); 4135 } else if (terminator()->offer_termination(&_term_term)) { 4136 assert(work_q->size() == 0, "Impossible!"); 4137 break; 4138 } else if (yielding() || should_yield()) { 4139 yield(); 4140 } 4141 } 4142 } 4143 4144 // This is run by the CMS (coordinator) thread. 4145 void CMSConcMarkingTask::coordinator_yield() { 4146 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 4147 "CMS thread should hold CMS token"); 4148 // First give up the locks, then yield, then re-lock 4149 // We should probably use a constructor/destructor idiom to 4150 // do this unlock/lock or modify the MutexUnlocker class to 4151 // serve our purpose. XXX 4152 assert_lock_strong(_bit_map_lock); 4153 _bit_map_lock->unlock(); 4154 ConcurrentMarkSweepThread::desynchronize(true); 4155 ConcurrentMarkSweepThread::acknowledge_yield_request(); 4156 _collector->stopTimer(); 4157 if (PrintCMSStatistics != 0) { 4158 _collector->incrementYields(); 4159 } 4160 _collector->icms_wait(); 4161 4162 // It is possible for whichever thread initiated the yield request 4163 // not to get a chance to wake up and take the bitmap lock between 4164 // this thread releasing it and reacquiring it. So, while the 4165 // should_yield() flag is on, let's sleep for a bit to give the 4166 // other thread a chance to wake up. The limit imposed on the number 4167 // of iterations is defensive, to avoid any unforseen circumstances 4168 // putting us into an infinite loop. Since it's always been this 4169 // (coordinator_yield()) method that was observed to cause the 4170 // problem, we are using a parameter (CMSCoordinatorYieldSleepCount) 4171 // which is by default non-zero. For the other seven methods that 4172 // also perform the yield operation, as are using a different 4173 // parameter (CMSYieldSleepCount) which is by default zero. This way we 4174 // can enable the sleeping for those methods too, if necessary. 4175 // See 6442774. 4176 // 4177 // We really need to reconsider the synchronization between the GC 4178 // thread and the yield-requesting threads in the future and we 4179 // should really use wait/notify, which is the recommended 4180 // way of doing this type of interaction. Additionally, we should 4181 // consolidate the eight methods that do the yield operation and they 4182 // are almost identical into one for better maintenability and 4183 // readability. See 6445193. 4184 // 4185 // Tony 2006.06.29 4186 for (unsigned i = 0; i < CMSCoordinatorYieldSleepCount && 4187 ConcurrentMarkSweepThread::should_yield() && 4188 !CMSCollector::foregroundGCIsActive(); ++i) { 4189 os::sleep(Thread::current(), 1, false); 4190 ConcurrentMarkSweepThread::acknowledge_yield_request(); 4191 } 4192 4193 ConcurrentMarkSweepThread::synchronize(true); 4194 _bit_map_lock->lock_without_safepoint_check(); 4195 _collector->startTimer(); 4196 } 4197 4198 bool CMSCollector::do_marking_mt(bool asynch) { 4199 assert(ConcGCThreads > 0 && conc_workers() != NULL, "precondition"); 4200 int num_workers = AdaptiveSizePolicy::calc_active_conc_workers( 4201 conc_workers()->total_workers(), 4202 conc_workers()->active_workers(), 4203 Threads::number_of_non_daemon_threads()); 4204 conc_workers()->set_active_workers(num_workers); 4205 4206 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 4207 4208 CMSConcMarkingTask tsk(this, 4209 cms_space, 4210 asynch, 4211 conc_workers(), 4212 task_queues()); 4213 4214 // Since the actual number of workers we get may be different 4215 // from the number we requested above, do we need to do anything different 4216 // below? In particular, may be we need to subclass the SequantialSubTasksDone 4217 // class?? XXX 4218 cms_space ->initialize_sequential_subtasks_for_marking(num_workers); 4219 4220 // Refs discovery is already non-atomic. 4221 assert(!ref_processor()->discovery_is_atomic(), "Should be non-atomic"); 4222 assert(ref_processor()->discovery_is_mt(), "Discovery should be MT"); 4223 conc_workers()->start_task(&tsk); 4224 while (tsk.yielded()) { 4225 tsk.coordinator_yield(); 4226 conc_workers()->continue_task(&tsk); 4227 } 4228 // If the task was aborted, _restart_addr will be non-NULL 4229 assert(tsk.completed() || _restart_addr != NULL, "Inconsistency"); 4230 while (_restart_addr != NULL) { 4231 // XXX For now we do not make use of ABORTED state and have not 4232 // yet implemented the right abort semantics (even in the original 4233 // single-threaded CMS case). That needs some more investigation 4234 // and is deferred for now; see CR# TBF. 07252005YSR. XXX 4235 assert(!CMSAbortSemantics || tsk.aborted(), "Inconsistency"); 4236 // If _restart_addr is non-NULL, a marking stack overflow 4237 // occurred; we need to do a fresh marking iteration from the 4238 // indicated restart address. 4239 if (_foregroundGCIsActive && asynch) { 4240 // We may be running into repeated stack overflows, having 4241 // reached the limit of the stack size, while making very 4242 // slow forward progress. It may be best to bail out and 4243 // let the foreground collector do its job. 4244 // Clear _restart_addr, so that foreground GC 4245 // works from scratch. This avoids the headache of 4246 // a "rescan" which would otherwise be needed because 4247 // of the dirty mod union table & card table. 4248 _restart_addr = NULL; 4249 return false; 4250 } 4251 // Adjust the task to restart from _restart_addr 4252 tsk.reset(_restart_addr); 4253 cms_space ->initialize_sequential_subtasks_for_marking(num_workers, 4254 _restart_addr); 4255 _restart_addr = NULL; 4256 // Get the workers going again 4257 conc_workers()->start_task(&tsk); 4258 while (tsk.yielded()) { 4259 tsk.coordinator_yield(); 4260 conc_workers()->continue_task(&tsk); 4261 } 4262 } 4263 assert(tsk.completed(), "Inconsistency"); 4264 assert(tsk.result() == true, "Inconsistency"); 4265 return true; 4266 } 4267 4268 bool CMSCollector::do_marking_st(bool asynch) { 4269 ResourceMark rm; 4270 HandleMark hm; 4271 4272 // Temporarily make refs discovery single threaded (non-MT) 4273 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false); 4274 MarkFromRootsClosure markFromRootsClosure(this, _span, &_markBitMap, 4275 &_markStack, CMSYield && asynch); 4276 // the last argument to iterate indicates whether the iteration 4277 // should be incremental with periodic yields. 4278 _markBitMap.iterate(&markFromRootsClosure); 4279 // If _restart_addr is non-NULL, a marking stack overflow 4280 // occurred; we need to do a fresh iteration from the 4281 // indicated restart address. 4282 while (_restart_addr != NULL) { 4283 if (_foregroundGCIsActive && asynch) { 4284 // We may be running into repeated stack overflows, having 4285 // reached the limit of the stack size, while making very 4286 // slow forward progress. It may be best to bail out and 4287 // let the foreground collector do its job. 4288 // Clear _restart_addr, so that foreground GC 4289 // works from scratch. This avoids the headache of 4290 // a "rescan" which would otherwise be needed because 4291 // of the dirty mod union table & card table. 4292 _restart_addr = NULL; 4293 return false; // indicating failure to complete marking 4294 } 4295 // Deal with stack overflow: 4296 // we restart marking from _restart_addr 4297 HeapWord* ra = _restart_addr; 4298 markFromRootsClosure.reset(ra); 4299 _restart_addr = NULL; 4300 _markBitMap.iterate(&markFromRootsClosure, ra, _span.end()); 4301 } 4302 return true; 4303 } 4304 4305 void CMSCollector::preclean() { 4306 check_correct_thread_executing(); 4307 assert(Thread::current()->is_ConcurrentGC_thread(), "Wrong thread"); 4308 verify_work_stacks_empty(); 4309 verify_overflow_empty(); 4310 _abort_preclean = false; 4311 if (CMSPrecleaningEnabled) { 4312 _eden_chunk_index = 0; 4313 size_t used = get_eden_used(); 4314 size_t capacity = get_eden_capacity(); 4315 // Don't start sampling unless we will get sufficiently 4316 // many samples. 4317 if (used < (capacity/(CMSScheduleRemarkSamplingRatio * 100) 4318 * CMSScheduleRemarkEdenPenetration)) { 4319 _start_sampling = true; 4320 } else { 4321 _start_sampling = false; 4322 } 4323 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 4324 CMSPhaseAccounting pa(this, "preclean", !PrintGCDetails); 4325 preclean_work(CMSPrecleanRefLists1, CMSPrecleanSurvivors1); 4326 } 4327 CMSTokenSync x(true); // is cms thread 4328 if (CMSPrecleaningEnabled) { 4329 sample_eden(); 4330 _collectorState = AbortablePreclean; 4331 } else { 4332 _collectorState = FinalMarking; 4333 } 4334 verify_work_stacks_empty(); 4335 verify_overflow_empty(); 4336 } 4337 4338 // Try and schedule the remark such that young gen 4339 // occupancy is CMSScheduleRemarkEdenPenetration %. 4340 void CMSCollector::abortable_preclean() { 4341 check_correct_thread_executing(); 4342 assert(CMSPrecleaningEnabled, "Inconsistent control state"); 4343 assert(_collectorState == AbortablePreclean, "Inconsistent control state"); 4344 4345 // If Eden's current occupancy is below this threshold, 4346 // immediately schedule the remark; else preclean 4347 // past the next scavenge in an effort to 4348 // schedule the pause as described avove. By choosing 4349 // CMSScheduleRemarkEdenSizeThreshold >= max eden size 4350 // we will never do an actual abortable preclean cycle. 4351 if (get_eden_used() > CMSScheduleRemarkEdenSizeThreshold) { 4352 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 4353 CMSPhaseAccounting pa(this, "abortable-preclean", !PrintGCDetails); 4354 // We need more smarts in the abortable preclean 4355 // loop below to deal with cases where allocation 4356 // in young gen is very very slow, and our precleaning 4357 // is running a losing race against a horde of 4358 // mutators intent on flooding us with CMS updates 4359 // (dirty cards). 4360 // One, admittedly dumb, strategy is to give up 4361 // after a certain number of abortable precleaning loops 4362 // or after a certain maximum time. We want to make 4363 // this smarter in the next iteration. 4364 // XXX FIX ME!!! YSR 4365 size_t loops = 0, workdone = 0, cumworkdone = 0, waited = 0; 4366 while (!(should_abort_preclean() || 4367 ConcurrentMarkSweepThread::should_terminate())) { 4368 workdone = preclean_work(CMSPrecleanRefLists2, CMSPrecleanSurvivors2); 4369 cumworkdone += workdone; 4370 loops++; 4371 // Voluntarily terminate abortable preclean phase if we have 4372 // been at it for too long. 4373 if ((CMSMaxAbortablePrecleanLoops != 0) && 4374 loops >= CMSMaxAbortablePrecleanLoops) { 4375 if (PrintGCDetails) { 4376 gclog_or_tty->print(" CMS: abort preclean due to loops "); 4377 } 4378 break; 4379 } 4380 if (pa.wallclock_millis() > CMSMaxAbortablePrecleanTime) { 4381 if (PrintGCDetails) { 4382 gclog_or_tty->print(" CMS: abort preclean due to time "); 4383 } 4384 break; 4385 } 4386 // If we are doing little work each iteration, we should 4387 // take a short break. 4388 if (workdone < CMSAbortablePrecleanMinWorkPerIteration) { 4389 // Sleep for some time, waiting for work to accumulate 4390 stopTimer(); 4391 cmsThread()->wait_on_cms_lock(CMSAbortablePrecleanWaitMillis); 4392 startTimer(); 4393 waited++; 4394 } 4395 } 4396 if (PrintCMSStatistics > 0) { 4397 gclog_or_tty->print(" [%d iterations, %d waits, %d cards)] ", 4398 loops, waited, cumworkdone); 4399 } 4400 } 4401 CMSTokenSync x(true); // is cms thread 4402 if (_collectorState != Idling) { 4403 assert(_collectorState == AbortablePreclean, 4404 "Spontaneous state transition?"); 4405 _collectorState = FinalMarking; 4406 } // Else, a foreground collection completed this CMS cycle. 4407 return; 4408 } 4409 4410 // Respond to an Eden sampling opportunity 4411 void CMSCollector::sample_eden() { 4412 // Make sure a young gc cannot sneak in between our 4413 // reading and recording of a sample. 4414 assert(Thread::current()->is_ConcurrentGC_thread(), 4415 "Only the cms thread may collect Eden samples"); 4416 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 4417 "Should collect samples while holding CMS token"); 4418 if (!_start_sampling) { 4419 return; 4420 } 4421 if (_eden_chunk_array) { 4422 if (_eden_chunk_index < _eden_chunk_capacity) { 4423 _eden_chunk_array[_eden_chunk_index] = *_top_addr; // take sample 4424 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr, 4425 "Unexpected state of Eden"); 4426 // We'd like to check that what we just sampled is an oop-start address; 4427 // however, we cannot do that here since the object may not yet have been 4428 // initialized. So we'll instead do the check when we _use_ this sample 4429 // later. 4430 if (_eden_chunk_index == 0 || 4431 (pointer_delta(_eden_chunk_array[_eden_chunk_index], 4432 _eden_chunk_array[_eden_chunk_index-1]) 4433 >= CMSSamplingGrain)) { 4434 _eden_chunk_index++; // commit sample 4435 } 4436 } 4437 } 4438 if ((_collectorState == AbortablePreclean) && !_abort_preclean) { 4439 size_t used = get_eden_used(); 4440 size_t capacity = get_eden_capacity(); 4441 assert(used <= capacity, "Unexpected state of Eden"); 4442 if (used > (capacity/100 * CMSScheduleRemarkEdenPenetration)) { 4443 _abort_preclean = true; 4444 } 4445 } 4446 } 4447 4448 4449 size_t CMSCollector::preclean_work(bool clean_refs, bool clean_survivor) { 4450 assert(_collectorState == Precleaning || 4451 _collectorState == AbortablePreclean, "incorrect state"); 4452 ResourceMark rm; 4453 HandleMark hm; 4454 4455 // Precleaning is currently not MT but the reference processor 4456 // may be set for MT. Disable it temporarily here. 4457 ReferenceProcessor* rp = ref_processor(); 4458 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(rp, false); 4459 4460 // Do one pass of scrubbing the discovered reference lists 4461 // to remove any reference objects with strongly-reachable 4462 // referents. 4463 if (clean_refs) { 4464 CMSPrecleanRefsYieldClosure yield_cl(this); 4465 assert(rp->span().equals(_span), "Spans should be equal"); 4466 CMSKeepAliveClosure keep_alive(this, _span, &_markBitMap, 4467 &_markStack, true /* preclean */); 4468 CMSDrainMarkingStackClosure complete_trace(this, 4469 _span, &_markBitMap, &_markStack, 4470 &keep_alive, true /* preclean */); 4471 4472 // We don't want this step to interfere with a young 4473 // collection because we don't want to take CPU 4474 // or memory bandwidth away from the young GC threads 4475 // (which may be as many as there are CPUs). 4476 // Note that we don't need to protect ourselves from 4477 // interference with mutators because they can't 4478 // manipulate the discovered reference lists nor affect 4479 // the computed reachability of the referents, the 4480 // only properties manipulated by the precleaning 4481 // of these reference lists. 4482 stopTimer(); 4483 CMSTokenSyncWithLocks x(true /* is cms thread */, 4484 bitMapLock()); 4485 startTimer(); 4486 sample_eden(); 4487 4488 // The following will yield to allow foreground 4489 // collection to proceed promptly. XXX YSR: 4490 // The code in this method may need further 4491 // tweaking for better performance and some restructuring 4492 // for cleaner interfaces. 4493 rp->preclean_discovered_references( 4494 rp->is_alive_non_header(), &keep_alive, &complete_trace, &yield_cl); 4495 } 4496 4497 if (clean_survivor) { // preclean the active survivor space(s) 4498 assert(_young_gen->kind() == Generation::DefNew || 4499 _young_gen->kind() == Generation::ParNew || 4500 _young_gen->kind() == Generation::ASParNew, 4501 "incorrect type for cast"); 4502 DefNewGeneration* dng = (DefNewGeneration*)_young_gen; 4503 PushAndMarkClosure pam_cl(this, _span, ref_processor(), 4504 &_markBitMap, &_modUnionTable, 4505 &_markStack, true /* precleaning phase */); 4506 stopTimer(); 4507 CMSTokenSyncWithLocks ts(true /* is cms thread */, 4508 bitMapLock()); 4509 startTimer(); 4510 unsigned int before_count = 4511 GenCollectedHeap::heap()->total_collections(); 4512 SurvivorSpacePrecleanClosure 4513 sss_cl(this, _span, &_markBitMap, &_markStack, 4514 &pam_cl, before_count, CMSYield); 4515 dng->from()->object_iterate_careful(&sss_cl); 4516 dng->to()->object_iterate_careful(&sss_cl); 4517 } 4518 MarkRefsIntoAndScanClosure 4519 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable, 4520 &_markStack, this, CMSYield, 4521 true /* precleaning phase */); 4522 // CAUTION: The following closure has persistent state that may need to 4523 // be reset upon a decrease in the sequence of addresses it 4524 // processes. 4525 ScanMarkedObjectsAgainCarefullyClosure 4526 smoac_cl(this, _span, 4527 &_markBitMap, &_markStack, &mrias_cl, CMSYield); 4528 4529 // Preclean dirty cards in ModUnionTable and CardTable using 4530 // appropriate convergence criterion; 4531 // repeat CMSPrecleanIter times unless we find that 4532 // we are losing. 4533 assert(CMSPrecleanIter < 10, "CMSPrecleanIter is too large"); 4534 assert(CMSPrecleanNumerator < CMSPrecleanDenominator, 4535 "Bad convergence multiplier"); 4536 assert(CMSPrecleanThreshold >= 100, 4537 "Unreasonably low CMSPrecleanThreshold"); 4538 4539 size_t numIter, cumNumCards, lastNumCards, curNumCards; 4540 for (numIter = 0, cumNumCards = lastNumCards = curNumCards = 0; 4541 numIter < CMSPrecleanIter; 4542 numIter++, lastNumCards = curNumCards, cumNumCards += curNumCards) { 4543 curNumCards = preclean_mod_union_table(_cmsGen, &smoac_cl); 4544 if (Verbose && PrintGCDetails) { 4545 gclog_or_tty->print(" (modUnionTable: %d cards)", curNumCards); 4546 } 4547 // Either there are very few dirty cards, so re-mark 4548 // pause will be small anyway, or our pre-cleaning isn't 4549 // that much faster than the rate at which cards are being 4550 // dirtied, so we might as well stop and re-mark since 4551 // precleaning won't improve our re-mark time by much. 4552 if (curNumCards <= CMSPrecleanThreshold || 4553 (numIter > 0 && 4554 (curNumCards * CMSPrecleanDenominator > 4555 lastNumCards * CMSPrecleanNumerator))) { 4556 numIter++; 4557 cumNumCards += curNumCards; 4558 break; 4559 } 4560 } 4561 4562 preclean_klasses(&mrias_cl, _cmsGen->freelistLock()); 4563 4564 curNumCards = preclean_card_table(_cmsGen, &smoac_cl); 4565 cumNumCards += curNumCards; 4566 if (PrintGCDetails && PrintCMSStatistics != 0) { 4567 gclog_or_tty->print_cr(" (cardTable: %d cards, re-scanned %d cards, %d iterations)", 4568 curNumCards, cumNumCards, numIter); 4569 } 4570 return cumNumCards; // as a measure of useful work done 4571 } 4572 4573 // PRECLEANING NOTES: 4574 // Precleaning involves: 4575 // . reading the bits of the modUnionTable and clearing the set bits. 4576 // . For the cards corresponding to the set bits, we scan the 4577 // objects on those cards. This means we need the free_list_lock 4578 // so that we can safely iterate over the CMS space when scanning 4579 // for oops. 4580 // . When we scan the objects, we'll be both reading and setting 4581 // marks in the marking bit map, so we'll need the marking bit map. 4582 // . For protecting _collector_state transitions, we take the CGC_lock. 4583 // Note that any races in the reading of of card table entries by the 4584 // CMS thread on the one hand and the clearing of those entries by the 4585 // VM thread or the setting of those entries by the mutator threads on the 4586 // other are quite benign. However, for efficiency it makes sense to keep 4587 // the VM thread from racing with the CMS thread while the latter is 4588 // dirty card info to the modUnionTable. We therefore also use the 4589 // CGC_lock to protect the reading of the card table and the mod union 4590 // table by the CM thread. 4591 // . We run concurrently with mutator updates, so scanning 4592 // needs to be done carefully -- we should not try to scan 4593 // potentially uninitialized objects. 4594 // 4595 // Locking strategy: While holding the CGC_lock, we scan over and 4596 // reset a maximal dirty range of the mod union / card tables, then lock 4597 // the free_list_lock and bitmap lock to do a full marking, then 4598 // release these locks; and repeat the cycle. This allows for a 4599 // certain amount of fairness in the sharing of these locks between 4600 // the CMS collector on the one hand, and the VM thread and the 4601 // mutators on the other. 4602 4603 // NOTE: preclean_mod_union_table() and preclean_card_table() 4604 // further below are largely identical; if you need to modify 4605 // one of these methods, please check the other method too. 4606 4607 size_t CMSCollector::preclean_mod_union_table( 4608 ConcurrentMarkSweepGeneration* gen, 4609 ScanMarkedObjectsAgainCarefullyClosure* cl) { 4610 verify_work_stacks_empty(); 4611 verify_overflow_empty(); 4612 4613 // strategy: starting with the first card, accumulate contiguous 4614 // ranges of dirty cards; clear these cards, then scan the region 4615 // covered by these cards. 4616 4617 // Since all of the MUT is committed ahead, we can just use 4618 // that, in case the generations expand while we are precleaning. 4619 // It might also be fine to just use the committed part of the 4620 // generation, but we might potentially miss cards when the 4621 // generation is rapidly expanding while we are in the midst 4622 // of precleaning. 4623 HeapWord* startAddr = gen->reserved().start(); 4624 HeapWord* endAddr = gen->reserved().end(); 4625 4626 cl->setFreelistLock(gen->freelistLock()); // needed for yielding 4627 4628 size_t numDirtyCards, cumNumDirtyCards; 4629 HeapWord *nextAddr, *lastAddr; 4630 for (cumNumDirtyCards = numDirtyCards = 0, 4631 nextAddr = lastAddr = startAddr; 4632 nextAddr < endAddr; 4633 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 4634 4635 ResourceMark rm; 4636 HandleMark hm; 4637 4638 MemRegion dirtyRegion; 4639 { 4640 stopTimer(); 4641 // Potential yield point 4642 CMSTokenSync ts(true); 4643 startTimer(); 4644 sample_eden(); 4645 // Get dirty region starting at nextOffset (inclusive), 4646 // simultaneously clearing it. 4647 dirtyRegion = 4648 _modUnionTable.getAndClearMarkedRegion(nextAddr, endAddr); 4649 assert(dirtyRegion.start() >= nextAddr, 4650 "returned region inconsistent?"); 4651 } 4652 // Remember where the next search should begin. 4653 // The returned region (if non-empty) is a right open interval, 4654 // so lastOffset is obtained from the right end of that 4655 // interval. 4656 lastAddr = dirtyRegion.end(); 4657 // Should do something more transparent and less hacky XXX 4658 numDirtyCards = 4659 _modUnionTable.heapWordDiffToOffsetDiff(dirtyRegion.word_size()); 4660 4661 // We'll scan the cards in the dirty region (with periodic 4662 // yields for foreground GC as needed). 4663 if (!dirtyRegion.is_empty()) { 4664 assert(numDirtyCards > 0, "consistency check"); 4665 HeapWord* stop_point = NULL; 4666 stopTimer(); 4667 // Potential yield point 4668 CMSTokenSyncWithLocks ts(true, gen->freelistLock(), 4669 bitMapLock()); 4670 startTimer(); 4671 { 4672 verify_work_stacks_empty(); 4673 verify_overflow_empty(); 4674 sample_eden(); 4675 stop_point = 4676 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 4677 } 4678 if (stop_point != NULL) { 4679 // The careful iteration stopped early either because it found an 4680 // uninitialized object, or because we were in the midst of an 4681 // "abortable preclean", which should now be aborted. Redirty 4682 // the bits corresponding to the partially-scanned or unscanned 4683 // cards. We'll either restart at the next block boundary or 4684 // abort the preclean. 4685 assert((_collectorState == AbortablePreclean && should_abort_preclean()), 4686 "Should only be AbortablePreclean."); 4687 _modUnionTable.mark_range(MemRegion(stop_point, dirtyRegion.end())); 4688 if (should_abort_preclean()) { 4689 break; // out of preclean loop 4690 } else { 4691 // Compute the next address at which preclean should pick up; 4692 // might need bitMapLock in order to read P-bits. 4693 lastAddr = next_card_start_after_block(stop_point); 4694 } 4695 } 4696 } else { 4697 assert(lastAddr == endAddr, "consistency check"); 4698 assert(numDirtyCards == 0, "consistency check"); 4699 break; 4700 } 4701 } 4702 verify_work_stacks_empty(); 4703 verify_overflow_empty(); 4704 return cumNumDirtyCards; 4705 } 4706 4707 // NOTE: preclean_mod_union_table() above and preclean_card_table() 4708 // below are largely identical; if you need to modify 4709 // one of these methods, please check the other method too. 4710 4711 size_t CMSCollector::preclean_card_table(ConcurrentMarkSweepGeneration* gen, 4712 ScanMarkedObjectsAgainCarefullyClosure* cl) { 4713 // strategy: it's similar to precleamModUnionTable above, in that 4714 // we accumulate contiguous ranges of dirty cards, mark these cards 4715 // precleaned, then scan the region covered by these cards. 4716 HeapWord* endAddr = (HeapWord*)(gen->_virtual_space.high()); 4717 HeapWord* startAddr = (HeapWord*)(gen->_virtual_space.low()); 4718 4719 cl->setFreelistLock(gen->freelistLock()); // needed for yielding 4720 4721 size_t numDirtyCards, cumNumDirtyCards; 4722 HeapWord *lastAddr, *nextAddr; 4723 4724 for (cumNumDirtyCards = numDirtyCards = 0, 4725 nextAddr = lastAddr = startAddr; 4726 nextAddr < endAddr; 4727 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 4728 4729 ResourceMark rm; 4730 HandleMark hm; 4731 4732 MemRegion dirtyRegion; 4733 { 4734 // See comments in "Precleaning notes" above on why we 4735 // do this locking. XXX Could the locking overheads be 4736 // too high when dirty cards are sparse? [I don't think so.] 4737 stopTimer(); 4738 CMSTokenSync x(true); // is cms thread 4739 startTimer(); 4740 sample_eden(); 4741 // Get and clear dirty region from card table 4742 dirtyRegion = _ct->ct_bs()->dirty_card_range_after_reset( 4743 MemRegion(nextAddr, endAddr), 4744 true, 4745 CardTableModRefBS::precleaned_card_val()); 4746 4747 assert(dirtyRegion.start() >= nextAddr, 4748 "returned region inconsistent?"); 4749 } 4750 lastAddr = dirtyRegion.end(); 4751 numDirtyCards = 4752 dirtyRegion.word_size()/CardTableModRefBS::card_size_in_words; 4753 4754 if (!dirtyRegion.is_empty()) { 4755 stopTimer(); 4756 CMSTokenSyncWithLocks ts(true, gen->freelistLock(), bitMapLock()); 4757 startTimer(); 4758 sample_eden(); 4759 verify_work_stacks_empty(); 4760 verify_overflow_empty(); 4761 HeapWord* stop_point = 4762 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 4763 if (stop_point != NULL) { 4764 assert((_collectorState == AbortablePreclean && should_abort_preclean()), 4765 "Should only be AbortablePreclean."); 4766 _ct->ct_bs()->invalidate(MemRegion(stop_point, dirtyRegion.end())); 4767 if (should_abort_preclean()) { 4768 break; // out of preclean loop 4769 } else { 4770 // Compute the next address at which preclean should pick up. 4771 lastAddr = next_card_start_after_block(stop_point); 4772 } 4773 } 4774 } else { 4775 break; 4776 } 4777 } 4778 verify_work_stacks_empty(); 4779 verify_overflow_empty(); 4780 return cumNumDirtyCards; 4781 } 4782 4783 class PrecleanKlassClosure : public KlassClosure { 4784 CMKlassClosure _cm_klass_closure; 4785 public: 4786 PrecleanKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {} 4787 void do_klass(Klass* k) { 4788 if (k->has_accumulated_modified_oops()) { 4789 k->clear_accumulated_modified_oops(); 4790 4791 _cm_klass_closure.do_klass(k); 4792 } 4793 } 4794 }; 4795 4796 // The freelist lock is needed to prevent asserts, is it really needed? 4797 void CMSCollector::preclean_klasses(MarkRefsIntoAndScanClosure* cl, Mutex* freelistLock) { 4798 4799 cl->set_freelistLock(freelistLock); 4800 4801 CMSTokenSyncWithLocks ts(true, freelistLock, bitMapLock()); 4802 4803 // SSS: Add equivalent to ScanMarkedObjectsAgainCarefullyClosure::do_yield_check and should_abort_preclean? 4804 // SSS: We should probably check if precleaning should be aborted, at suitable intervals? 4805 PrecleanKlassClosure preclean_klass_closure(cl); 4806 ClassLoaderDataGraph::classes_do(&preclean_klass_closure); 4807 4808 verify_work_stacks_empty(); 4809 verify_overflow_empty(); 4810 } 4811 4812 void CMSCollector::checkpointRootsFinal(bool asynch, 4813 bool clear_all_soft_refs, bool init_mark_was_synchronous) { 4814 assert(_collectorState == FinalMarking, "incorrect state transition?"); 4815 check_correct_thread_executing(); 4816 // world is stopped at this checkpoint 4817 assert(SafepointSynchronize::is_at_safepoint(), 4818 "world should be stopped"); 4819 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 4820 4821 verify_work_stacks_empty(); 4822 verify_overflow_empty(); 4823 4824 SpecializationStats::clear(); 4825 if (PrintGCDetails) { 4826 gclog_or_tty->print("[YG occupancy: "SIZE_FORMAT" K ("SIZE_FORMAT" K)]", 4827 _young_gen->used() / K, 4828 _young_gen->capacity() / K); 4829 } 4830 if (asynch) { 4831 if (CMSScavengeBeforeRemark) { 4832 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4833 // Temporarily set flag to false, GCH->do_collection will 4834 // expect it to be false and set to true 4835 FlagSetting fl(gch->_is_gc_active, false); 4836 NOT_PRODUCT(TraceTime t("Scavenge-Before-Remark", 4837 PrintGCDetails && Verbose, true, gclog_or_tty);) 4838 int level = _cmsGen->level() - 1; 4839 if (level >= 0) { 4840 gch->do_collection(true, // full (i.e. force, see below) 4841 false, // !clear_all_soft_refs 4842 0, // size 4843 false, // is_tlab 4844 level // max_level 4845 ); 4846 } 4847 } 4848 FreelistLocker x(this); 4849 MutexLockerEx y(bitMapLock(), 4850 Mutex::_no_safepoint_check_flag); 4851 assert(!init_mark_was_synchronous, "but that's impossible!"); 4852 checkpointRootsFinalWork(asynch, clear_all_soft_refs, false); 4853 } else { 4854 // already have all the locks 4855 checkpointRootsFinalWork(asynch, clear_all_soft_refs, 4856 init_mark_was_synchronous); 4857 } 4858 verify_work_stacks_empty(); 4859 verify_overflow_empty(); 4860 SpecializationStats::print(); 4861 } 4862 4863 void CMSCollector::checkpointRootsFinalWork(bool asynch, 4864 bool clear_all_soft_refs, bool init_mark_was_synchronous) { 4865 4866 NOT_PRODUCT(TraceTime tr("checkpointRootsFinalWork", PrintGCDetails, false, gclog_or_tty);) 4867 4868 assert(haveFreelistLocks(), "must have free list locks"); 4869 assert_lock_strong(bitMapLock()); 4870 4871 if (UseAdaptiveSizePolicy) { 4872 size_policy()->checkpoint_roots_final_begin(); 4873 } 4874 4875 ResourceMark rm; 4876 HandleMark hm; 4877 4878 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4879 4880 if (should_unload_classes()) { 4881 CodeCache::gc_prologue(); 4882 } 4883 assert(haveFreelistLocks(), "must have free list locks"); 4884 assert_lock_strong(bitMapLock()); 4885 4886 if (!init_mark_was_synchronous) { 4887 // We might assume that we need not fill TLAB's when 4888 // CMSScavengeBeforeRemark is set, because we may have just done 4889 // a scavenge which would have filled all TLAB's -- and besides 4890 // Eden would be empty. This however may not always be the case -- 4891 // for instance although we asked for a scavenge, it may not have 4892 // happened because of a JNI critical section. We probably need 4893 // a policy for deciding whether we can in that case wait until 4894 // the critical section releases and then do the remark following 4895 // the scavenge, and skip it here. In the absence of that policy, 4896 // or of an indication of whether the scavenge did indeed occur, 4897 // we cannot rely on TLAB's having been filled and must do 4898 // so here just in case a scavenge did not happen. 4899 gch->ensure_parsability(false); // fill TLAB's, but no need to retire them 4900 // Update the saved marks which may affect the root scans. 4901 gch->save_marks(); 4902 4903 { 4904 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;) 4905 4906 // Note on the role of the mod union table: 4907 // Since the marker in "markFromRoots" marks concurrently with 4908 // mutators, it is possible for some reachable objects not to have been 4909 // scanned. For instance, an only reference to an object A was 4910 // placed in object B after the marker scanned B. Unless B is rescanned, 4911 // A would be collected. Such updates to references in marked objects 4912 // are detected via the mod union table which is the set of all cards 4913 // dirtied since the first checkpoint in this GC cycle and prior to 4914 // the most recent young generation GC, minus those cleaned up by the 4915 // concurrent precleaning. 4916 if (CMSParallelRemarkEnabled && CollectedHeap::use_parallel_gc_threads()) { 4917 TraceTime t("Rescan (parallel) ", PrintGCDetails, false, gclog_or_tty); 4918 do_remark_parallel(); 4919 } else { 4920 TraceTime t("Rescan (non-parallel) ", PrintGCDetails, false, 4921 gclog_or_tty); 4922 do_remark_non_parallel(); 4923 } 4924 } 4925 } else { 4926 assert(!asynch, "Can't have init_mark_was_synchronous in asynch mode"); 4927 // The initial mark was stop-world, so there's no rescanning to 4928 // do; go straight on to the next step below. 4929 } 4930 verify_work_stacks_empty(); 4931 verify_overflow_empty(); 4932 4933 { 4934 NOT_PRODUCT(TraceTime ts("refProcessingWork", PrintGCDetails, false, gclog_or_tty);) 4935 refProcessingWork(asynch, clear_all_soft_refs); 4936 } 4937 verify_work_stacks_empty(); 4938 verify_overflow_empty(); 4939 4940 if (should_unload_classes()) { 4941 CodeCache::gc_epilogue(); 4942 } 4943 JvmtiExport::gc_epilogue(); 4944 4945 // If we encountered any (marking stack / work queue) overflow 4946 // events during the current CMS cycle, take appropriate 4947 // remedial measures, where possible, so as to try and avoid 4948 // recurrence of that condition. 4949 assert(_markStack.isEmpty(), "No grey objects"); 4950 size_t ser_ovflw = _ser_pmc_remark_ovflw + _ser_pmc_preclean_ovflw + 4951 _ser_kac_ovflw + _ser_kac_preclean_ovflw; 4952 if (ser_ovflw > 0) { 4953 if (PrintCMSStatistics != 0) { 4954 gclog_or_tty->print_cr("Marking stack overflow (benign) " 4955 "(pmc_pc="SIZE_FORMAT", pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT 4956 ", kac_preclean="SIZE_FORMAT")", 4957 _ser_pmc_preclean_ovflw, _ser_pmc_remark_ovflw, 4958 _ser_kac_ovflw, _ser_kac_preclean_ovflw); 4959 } 4960 _markStack.expand(); 4961 _ser_pmc_remark_ovflw = 0; 4962 _ser_pmc_preclean_ovflw = 0; 4963 _ser_kac_preclean_ovflw = 0; 4964 _ser_kac_ovflw = 0; 4965 } 4966 if (_par_pmc_remark_ovflw > 0 || _par_kac_ovflw > 0) { 4967 if (PrintCMSStatistics != 0) { 4968 gclog_or_tty->print_cr("Work queue overflow (benign) " 4969 "(pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT")", 4970 _par_pmc_remark_ovflw, _par_kac_ovflw); 4971 } 4972 _par_pmc_remark_ovflw = 0; 4973 _par_kac_ovflw = 0; 4974 } 4975 if (PrintCMSStatistics != 0) { 4976 if (_markStack._hit_limit > 0) { 4977 gclog_or_tty->print_cr(" (benign) Hit max stack size limit ("SIZE_FORMAT")", 4978 _markStack._hit_limit); 4979 } 4980 if (_markStack._failed_double > 0) { 4981 gclog_or_tty->print_cr(" (benign) Failed stack doubling ("SIZE_FORMAT")," 4982 " current capacity "SIZE_FORMAT, 4983 _markStack._failed_double, 4984 _markStack.capacity()); 4985 } 4986 } 4987 _markStack._hit_limit = 0; 4988 _markStack._failed_double = 0; 4989 4990 if ((VerifyAfterGC || VerifyDuringGC) && 4991 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 4992 verify_after_remark(); 4993 } 4994 4995 // Change under the freelistLocks. 4996 _collectorState = Sweeping; 4997 // Call isAllClear() under bitMapLock 4998 assert(_modUnionTable.isAllClear(), 4999 "Should be clear by end of the final marking"); 5000 assert(_ct->klass_rem_set()->mod_union_is_clear(), 5001 "Should be clear by end of the final marking"); 5002 if (UseAdaptiveSizePolicy) { 5003 size_policy()->checkpoint_roots_final_end(gch->gc_cause()); 5004 } 5005 } 5006 5007 // Parallel remark task 5008 class CMSParRemarkTask: public AbstractGangTask { 5009 CMSCollector* _collector; 5010 int _n_workers; 5011 CompactibleFreeListSpace* _cms_space; 5012 5013 // The per-thread work queues, available here for stealing. 5014 OopTaskQueueSet* _task_queues; 5015 ParallelTaskTerminator _term; 5016 5017 public: 5018 // A value of 0 passed to n_workers will cause the number of 5019 // workers to be taken from the active workers in the work gang. 5020 CMSParRemarkTask(CMSCollector* collector, 5021 CompactibleFreeListSpace* cms_space, 5022 int n_workers, FlexibleWorkGang* workers, 5023 OopTaskQueueSet* task_queues): 5024 AbstractGangTask("Rescan roots and grey objects in parallel"), 5025 _collector(collector), 5026 _cms_space(cms_space), 5027 _n_workers(n_workers), 5028 _task_queues(task_queues), 5029 _term(n_workers, task_queues) { } 5030 5031 OopTaskQueueSet* task_queues() { return _task_queues; } 5032 5033 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 5034 5035 ParallelTaskTerminator* terminator() { return &_term; } 5036 int n_workers() { return _n_workers; } 5037 5038 void work(uint worker_id); 5039 5040 private: 5041 // Work method in support of parallel rescan ... of young gen spaces 5042 void do_young_space_rescan(int i, Par_MarkRefsIntoAndScanClosure* cl, 5043 ContiguousSpace* space, 5044 HeapWord** chunk_array, size_t chunk_top); 5045 5046 // ... of dirty cards in old space 5047 void do_dirty_card_rescan_tasks(CompactibleFreeListSpace* sp, int i, 5048 Par_MarkRefsIntoAndScanClosure* cl); 5049 5050 // ... work stealing for the above 5051 void do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, int* seed); 5052 }; 5053 5054 class RemarkKlassClosure : public KlassClosure { 5055 CMKlassClosure _cm_klass_closure; 5056 public: 5057 RemarkKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {} 5058 void do_klass(Klass* k) { 5059 // Check if we have modified any oops in the Klass during the concurrent marking. 5060 if (k->has_accumulated_modified_oops()) { 5061 k->clear_accumulated_modified_oops(); 5062 5063 // We could have transfered the current modified marks to the accumulated marks, 5064 // like we do with the Card Table to Mod Union Table. But it's not really necessary. 5065 } else if (k->has_modified_oops()) { 5066 // Don't clear anything, this info is needed by the next young collection. 5067 } else { 5068 // No modified oops in the Klass. 5069 return; 5070 } 5071 5072 // The klass has modified fields, need to scan the klass. 5073 _cm_klass_closure.do_klass(k); 5074 } 5075 }; 5076 5077 // work_queue(i) is passed to the closure 5078 // Par_MarkRefsIntoAndScanClosure. The "i" parameter 5079 // also is passed to do_dirty_card_rescan_tasks() and to 5080 // do_work_steal() to select the i-th task_queue. 5081 5082 void CMSParRemarkTask::work(uint worker_id) { 5083 elapsedTimer _timer; 5084 ResourceMark rm; 5085 HandleMark hm; 5086 5087 // ---------- rescan from roots -------------- 5088 _timer.start(); 5089 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5090 Par_MarkRefsIntoAndScanClosure par_mrias_cl(_collector, 5091 _collector->_span, _collector->ref_processor(), 5092 &(_collector->_markBitMap), 5093 work_queue(worker_id)); 5094 5095 // Rescan young gen roots first since these are likely 5096 // coarsely partitioned and may, on that account, constitute 5097 // the critical path; thus, it's best to start off that 5098 // work first. 5099 // ---------- young gen roots -------------- 5100 { 5101 DefNewGeneration* dng = _collector->_young_gen->as_DefNewGeneration(); 5102 EdenSpace* eden_space = dng->eden(); 5103 ContiguousSpace* from_space = dng->from(); 5104 ContiguousSpace* to_space = dng->to(); 5105 5106 HeapWord** eca = _collector->_eden_chunk_array; 5107 size_t ect = _collector->_eden_chunk_index; 5108 HeapWord** sca = _collector->_survivor_chunk_array; 5109 size_t sct = _collector->_survivor_chunk_index; 5110 5111 assert(ect <= _collector->_eden_chunk_capacity, "out of bounds"); 5112 assert(sct <= _collector->_survivor_chunk_capacity, "out of bounds"); 5113 5114 do_young_space_rescan(worker_id, &par_mrias_cl, to_space, NULL, 0); 5115 do_young_space_rescan(worker_id, &par_mrias_cl, from_space, sca, sct); 5116 do_young_space_rescan(worker_id, &par_mrias_cl, eden_space, eca, ect); 5117 5118 _timer.stop(); 5119 if (PrintCMSStatistics != 0) { 5120 gclog_or_tty->print_cr( 5121 "Finished young gen rescan work in %dth thread: %3.3f sec", 5122 worker_id, _timer.seconds()); 5123 } 5124 } 5125 5126 // ---------- remaining roots -------------- 5127 _timer.reset(); 5128 _timer.start(); 5129 gch->gen_process_strong_roots(_collector->_cmsGen->level(), 5130 false, // yg was scanned above 5131 false, // this is parallel code 5132 false, // not scavenging 5133 SharedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()), 5134 &par_mrias_cl, 5135 true, // walk all of code cache if (so & SO_CodeCache) 5136 NULL, 5137 NULL); // The dirty klasses will be handled below 5138 assert(_collector->should_unload_classes() 5139 || (_collector->CMSCollector::roots_scanning_options() & SharedHeap::SO_CodeCache), 5140 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 5141 _timer.stop(); 5142 if (PrintCMSStatistics != 0) { 5143 gclog_or_tty->print_cr( 5144 "Finished remaining root rescan work in %dth thread: %3.3f sec", 5145 worker_id, _timer.seconds()); 5146 } 5147 5148 // ---------- unhandled CLD scanning ---------- 5149 if (worker_id == 0) { // Single threaded at the moment. 5150 _timer.reset(); 5151 _timer.start(); 5152 5153 // Scan all new class loader data objects and new dependencies that were 5154 // introduced during concurrent marking. 5155 ResourceMark rm; 5156 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds(); 5157 for (int i = 0; i < array->length(); i++) { 5158 par_mrias_cl.do_class_loader_data(array->at(i)); 5159 } 5160 5161 // We don't need to keep track of new CLDs anymore. 5162 ClassLoaderDataGraph::remember_new_clds(false); 5163 5164 _timer.stop(); 5165 if (PrintCMSStatistics != 0) { 5166 gclog_or_tty->print_cr( 5167 "Finished unhandled CLD scanning work in %dth thread: %3.3f sec", 5168 worker_id, _timer.seconds()); 5169 } 5170 } 5171 5172 // ---------- dirty klass scanning ---------- 5173 if (worker_id == 0) { // Single threaded at the moment. 5174 _timer.reset(); 5175 _timer.start(); 5176 5177 // Scan all classes that was dirtied during the concurrent marking phase. 5178 RemarkKlassClosure remark_klass_closure(&par_mrias_cl); 5179 ClassLoaderDataGraph::classes_do(&remark_klass_closure); 5180 5181 _timer.stop(); 5182 if (PrintCMSStatistics != 0) { 5183 gclog_or_tty->print_cr( 5184 "Finished dirty klass scanning work in %dth thread: %3.3f sec", 5185 worker_id, _timer.seconds()); 5186 } 5187 } 5188 5189 // We might have added oops to ClassLoaderData::_handles during the 5190 // concurrent marking phase. These oops point to newly allocated objects 5191 // that are guaranteed to be kept alive. Either by the direct allocation 5192 // code, or when the young collector processes the strong roots. Hence, 5193 // we don't have to revisit the _handles block during the remark phase. 5194 5195 // ---------- rescan dirty cards ------------ 5196 _timer.reset(); 5197 _timer.start(); 5198 5199 // Do the rescan tasks for each of the two spaces 5200 // (cms_space) in turn. 5201 // "worker_id" is passed to select the task_queue for "worker_id" 5202 do_dirty_card_rescan_tasks(_cms_space, worker_id, &par_mrias_cl); 5203 _timer.stop(); 5204 if (PrintCMSStatistics != 0) { 5205 gclog_or_tty->print_cr( 5206 "Finished dirty card rescan work in %dth thread: %3.3f sec", 5207 worker_id, _timer.seconds()); 5208 } 5209 5210 // ---------- steal work from other threads ... 5211 // ---------- ... and drain overflow list. 5212 _timer.reset(); 5213 _timer.start(); 5214 do_work_steal(worker_id, &par_mrias_cl, _collector->hash_seed(worker_id)); 5215 _timer.stop(); 5216 if (PrintCMSStatistics != 0) { 5217 gclog_or_tty->print_cr( 5218 "Finished work stealing in %dth thread: %3.3f sec", 5219 worker_id, _timer.seconds()); 5220 } 5221 } 5222 5223 // Note that parameter "i" is not used. 5224 void 5225 CMSParRemarkTask::do_young_space_rescan(int i, 5226 Par_MarkRefsIntoAndScanClosure* cl, ContiguousSpace* space, 5227 HeapWord** chunk_array, size_t chunk_top) { 5228 // Until all tasks completed: 5229 // . claim an unclaimed task 5230 // . compute region boundaries corresponding to task claimed 5231 // using chunk_array 5232 // . par_oop_iterate(cl) over that region 5233 5234 ResourceMark rm; 5235 HandleMark hm; 5236 5237 SequentialSubTasksDone* pst = space->par_seq_tasks(); 5238 assert(pst->valid(), "Uninitialized use?"); 5239 5240 uint nth_task = 0; 5241 uint n_tasks = pst->n_tasks(); 5242 5243 HeapWord *start, *end; 5244 while (!pst->is_task_claimed(/* reference */ nth_task)) { 5245 // We claimed task # nth_task; compute its boundaries. 5246 if (chunk_top == 0) { // no samples were taken 5247 assert(nth_task == 0 && n_tasks == 1, "Can have only 1 EdenSpace task"); 5248 start = space->bottom(); 5249 end = space->top(); 5250 } else if (nth_task == 0) { 5251 start = space->bottom(); 5252 end = chunk_array[nth_task]; 5253 } else if (nth_task < (uint)chunk_top) { 5254 assert(nth_task >= 1, "Control point invariant"); 5255 start = chunk_array[nth_task - 1]; 5256 end = chunk_array[nth_task]; 5257 } else { 5258 assert(nth_task == (uint)chunk_top, "Control point invariant"); 5259 start = chunk_array[chunk_top - 1]; 5260 end = space->top(); 5261 } 5262 MemRegion mr(start, end); 5263 // Verify that mr is in space 5264 assert(mr.is_empty() || space->used_region().contains(mr), 5265 "Should be in space"); 5266 // Verify that "start" is an object boundary 5267 assert(mr.is_empty() || oop(mr.start())->is_oop(), 5268 "Should be an oop"); 5269 space->par_oop_iterate(mr, cl); 5270 } 5271 pst->all_tasks_completed(); 5272 } 5273 5274 void 5275 CMSParRemarkTask::do_dirty_card_rescan_tasks( 5276 CompactibleFreeListSpace* sp, int i, 5277 Par_MarkRefsIntoAndScanClosure* cl) { 5278 // Until all tasks completed: 5279 // . claim an unclaimed task 5280 // . compute region boundaries corresponding to task claimed 5281 // . transfer dirty bits ct->mut for that region 5282 // . apply rescanclosure to dirty mut bits for that region 5283 5284 ResourceMark rm; 5285 HandleMark hm; 5286 5287 OopTaskQueue* work_q = work_queue(i); 5288 ModUnionClosure modUnionClosure(&(_collector->_modUnionTable)); 5289 // CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! 5290 // CAUTION: This closure has state that persists across calls to 5291 // the work method dirty_range_iterate_clear() in that it has 5292 // imbedded in it a (subtype of) UpwardsObjectClosure. The 5293 // use of that state in the imbedded UpwardsObjectClosure instance 5294 // assumes that the cards are always iterated (even if in parallel 5295 // by several threads) in monotonically increasing order per each 5296 // thread. This is true of the implementation below which picks 5297 // card ranges (chunks) in monotonically increasing order globally 5298 // and, a-fortiori, in monotonically increasing order per thread 5299 // (the latter order being a subsequence of the former). 5300 // If the work code below is ever reorganized into a more chaotic 5301 // work-partitioning form than the current "sequential tasks" 5302 // paradigm, the use of that persistent state will have to be 5303 // revisited and modified appropriately. See also related 5304 // bug 4756801 work on which should examine this code to make 5305 // sure that the changes there do not run counter to the 5306 // assumptions made here and necessary for correctness and 5307 // efficiency. Note also that this code might yield inefficient 5308 // behaviour in the case of very large objects that span one or 5309 // more work chunks. Such objects would potentially be scanned 5310 // several times redundantly. Work on 4756801 should try and 5311 // address that performance anomaly if at all possible. XXX 5312 MemRegion full_span = _collector->_span; 5313 CMSBitMap* bm = &(_collector->_markBitMap); // shared 5314 MarkFromDirtyCardsClosure 5315 greyRescanClosure(_collector, full_span, // entire span of interest 5316 sp, bm, work_q, cl); 5317 5318 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 5319 assert(pst->valid(), "Uninitialized use?"); 5320 uint nth_task = 0; 5321 const int alignment = CardTableModRefBS::card_size * BitsPerWord; 5322 MemRegion span = sp->used_region(); 5323 HeapWord* start_addr = span.start(); 5324 HeapWord* end_addr = (HeapWord*)round_to((intptr_t)span.end(), 5325 alignment); 5326 const size_t chunk_size = sp->rescan_task_size(); // in HeapWord units 5327 assert((HeapWord*)round_to((intptr_t)start_addr, alignment) == 5328 start_addr, "Check alignment"); 5329 assert((size_t)round_to((intptr_t)chunk_size, alignment) == 5330 chunk_size, "Check alignment"); 5331 5332 while (!pst->is_task_claimed(/* reference */ nth_task)) { 5333 // Having claimed the nth_task, compute corresponding mem-region, 5334 // which is a-fortiori aligned correctly (i.e. at a MUT bopundary). 5335 // The alignment restriction ensures that we do not need any 5336 // synchronization with other gang-workers while setting or 5337 // clearing bits in thus chunk of the MUT. 5338 MemRegion this_span = MemRegion(start_addr + nth_task*chunk_size, 5339 start_addr + (nth_task+1)*chunk_size); 5340 // The last chunk's end might be way beyond end of the 5341 // used region. In that case pull back appropriately. 5342 if (this_span.end() > end_addr) { 5343 this_span.set_end(end_addr); 5344 assert(!this_span.is_empty(), "Program logic (calculation of n_tasks)"); 5345 } 5346 // Iterate over the dirty cards covering this chunk, marking them 5347 // precleaned, and setting the corresponding bits in the mod union 5348 // table. Since we have been careful to partition at Card and MUT-word 5349 // boundaries no synchronization is needed between parallel threads. 5350 _collector->_ct->ct_bs()->dirty_card_iterate(this_span, 5351 &modUnionClosure); 5352 5353 // Having transferred these marks into the modUnionTable, 5354 // rescan the marked objects on the dirty cards in the modUnionTable. 5355 // Even if this is at a synchronous collection, the initial marking 5356 // may have been done during an asynchronous collection so there 5357 // may be dirty bits in the mod-union table. 5358 _collector->_modUnionTable.dirty_range_iterate_clear( 5359 this_span, &greyRescanClosure); 5360 _collector->_modUnionTable.verifyNoOneBitsInRange( 5361 this_span.start(), 5362 this_span.end()); 5363 } 5364 pst->all_tasks_completed(); // declare that i am done 5365 } 5366 5367 // . see if we can share work_queues with ParNew? XXX 5368 void 5369 CMSParRemarkTask::do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, 5370 int* seed) { 5371 OopTaskQueue* work_q = work_queue(i); 5372 NOT_PRODUCT(int num_steals = 0;) 5373 oop obj_to_scan; 5374 CMSBitMap* bm = &(_collector->_markBitMap); 5375 5376 while (true) { 5377 // Completely finish any left over work from (an) earlier round(s) 5378 cl->trim_queue(0); 5379 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 5380 (size_t)ParGCDesiredObjsFromOverflowList); 5381 // Now check if there's any work in the overflow list 5382 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads, 5383 // only affects the number of attempts made to get work from the 5384 // overflow list and does not affect the number of workers. Just 5385 // pass ParallelGCThreads so this behavior is unchanged. 5386 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 5387 work_q, 5388 ParallelGCThreads)) { 5389 // found something in global overflow list; 5390 // not yet ready to go stealing work from others. 5391 // We'd like to assert(work_q->size() != 0, ...) 5392 // because we just took work from the overflow list, 5393 // but of course we can't since all of that could have 5394 // been already stolen from us. 5395 // "He giveth and He taketh away." 5396 continue; 5397 } 5398 // Verify that we have no work before we resort to stealing 5399 assert(work_q->size() == 0, "Have work, shouldn't steal"); 5400 // Try to steal from other queues that have work 5401 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 5402 NOT_PRODUCT(num_steals++;) 5403 assert(obj_to_scan->is_oop(), "Oops, not an oop!"); 5404 assert(bm->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 5405 // Do scanning work 5406 obj_to_scan->oop_iterate(cl); 5407 // Loop around, finish this work, and try to steal some more 5408 } else if (terminator()->offer_termination()) { 5409 break; // nirvana from the infinite cycle 5410 } 5411 } 5412 NOT_PRODUCT( 5413 if (PrintCMSStatistics != 0) { 5414 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals); 5415 } 5416 ) 5417 assert(work_q->size() == 0 && _collector->overflow_list_is_empty(), 5418 "Else our work is not yet done"); 5419 } 5420 5421 // Return a thread-local PLAB recording array, as appropriate. 5422 void* CMSCollector::get_data_recorder(int thr_num) { 5423 if (_survivor_plab_array != NULL && 5424 (CMSPLABRecordAlways || 5425 (_collectorState > Marking && _collectorState < FinalMarking))) { 5426 assert(thr_num < (int)ParallelGCThreads, "thr_num is out of bounds"); 5427 ChunkArray* ca = &_survivor_plab_array[thr_num]; 5428 ca->reset(); // clear it so that fresh data is recorded 5429 return (void*) ca; 5430 } else { 5431 return NULL; 5432 } 5433 } 5434 5435 // Reset all the thread-local PLAB recording arrays 5436 void CMSCollector::reset_survivor_plab_arrays() { 5437 for (uint i = 0; i < ParallelGCThreads; i++) { 5438 _survivor_plab_array[i].reset(); 5439 } 5440 } 5441 5442 // Merge the per-thread plab arrays into the global survivor chunk 5443 // array which will provide the partitioning of the survivor space 5444 // for CMS rescan. 5445 void CMSCollector::merge_survivor_plab_arrays(ContiguousSpace* surv, 5446 int no_of_gc_threads) { 5447 assert(_survivor_plab_array != NULL, "Error"); 5448 assert(_survivor_chunk_array != NULL, "Error"); 5449 assert(_collectorState == FinalMarking, "Error"); 5450 for (int j = 0; j < no_of_gc_threads; j++) { 5451 _cursor[j] = 0; 5452 } 5453 HeapWord* top = surv->top(); 5454 size_t i; 5455 for (i = 0; i < _survivor_chunk_capacity; i++) { // all sca entries 5456 HeapWord* min_val = top; // Higher than any PLAB address 5457 uint min_tid = 0; // position of min_val this round 5458 for (int j = 0; j < no_of_gc_threads; j++) { 5459 ChunkArray* cur_sca = &_survivor_plab_array[j]; 5460 if (_cursor[j] == cur_sca->end()) { 5461 continue; 5462 } 5463 assert(_cursor[j] < cur_sca->end(), "ctl pt invariant"); 5464 HeapWord* cur_val = cur_sca->nth(_cursor[j]); 5465 assert(surv->used_region().contains(cur_val), "Out of bounds value"); 5466 if (cur_val < min_val) { 5467 min_tid = j; 5468 min_val = cur_val; 5469 } else { 5470 assert(cur_val < top, "All recorded addresses should be less"); 5471 } 5472 } 5473 // At this point min_val and min_tid are respectively 5474 // the least address in _survivor_plab_array[j]->nth(_cursor[j]) 5475 // and the thread (j) that witnesses that address. 5476 // We record this address in the _survivor_chunk_array[i] 5477 // and increment _cursor[min_tid] prior to the next round i. 5478 if (min_val == top) { 5479 break; 5480 } 5481 _survivor_chunk_array[i] = min_val; 5482 _cursor[min_tid]++; 5483 } 5484 // We are all done; record the size of the _survivor_chunk_array 5485 _survivor_chunk_index = i; // exclusive: [0, i) 5486 if (PrintCMSStatistics > 0) { 5487 gclog_or_tty->print(" (Survivor:" SIZE_FORMAT "chunks) ", i); 5488 } 5489 // Verify that we used up all the recorded entries 5490 #ifdef ASSERT 5491 size_t total = 0; 5492 for (int j = 0; j < no_of_gc_threads; j++) { 5493 assert(_cursor[j] == _survivor_plab_array[j].end(), "Ctl pt invariant"); 5494 total += _cursor[j]; 5495 } 5496 assert(total == _survivor_chunk_index, "Ctl Pt Invariant"); 5497 // Check that the merged array is in sorted order 5498 if (total > 0) { 5499 for (size_t i = 0; i < total - 1; i++) { 5500 if (PrintCMSStatistics > 0) { 5501 gclog_or_tty->print(" (chunk" SIZE_FORMAT ":" INTPTR_FORMAT ") ", 5502 i, _survivor_chunk_array[i]); 5503 } 5504 assert(_survivor_chunk_array[i] < _survivor_chunk_array[i+1], 5505 "Not sorted"); 5506 } 5507 } 5508 #endif // ASSERT 5509 } 5510 5511 // Set up the space's par_seq_tasks structure for work claiming 5512 // for parallel rescan of young gen. 5513 // See ParRescanTask where this is currently used. 5514 void 5515 CMSCollector:: 5516 initialize_sequential_subtasks_for_young_gen_rescan(int n_threads) { 5517 assert(n_threads > 0, "Unexpected n_threads argument"); 5518 DefNewGeneration* dng = (DefNewGeneration*)_young_gen; 5519 5520 // Eden space 5521 { 5522 SequentialSubTasksDone* pst = dng->eden()->par_seq_tasks(); 5523 assert(!pst->valid(), "Clobbering existing data?"); 5524 // Each valid entry in [0, _eden_chunk_index) represents a task. 5525 size_t n_tasks = _eden_chunk_index + 1; 5526 assert(n_tasks == 1 || _eden_chunk_array != NULL, "Error"); 5527 // Sets the condition for completion of the subtask (how many threads 5528 // need to finish in order to be done). 5529 pst->set_n_threads(n_threads); 5530 pst->set_n_tasks((int)n_tasks); 5531 } 5532 5533 // Merge the survivor plab arrays into _survivor_chunk_array 5534 if (_survivor_plab_array != NULL) { 5535 merge_survivor_plab_arrays(dng->from(), n_threads); 5536 } else { 5537 assert(_survivor_chunk_index == 0, "Error"); 5538 } 5539 5540 // To space 5541 { 5542 SequentialSubTasksDone* pst = dng->to()->par_seq_tasks(); 5543 assert(!pst->valid(), "Clobbering existing data?"); 5544 // Sets the condition for completion of the subtask (how many threads 5545 // need to finish in order to be done). 5546 pst->set_n_threads(n_threads); 5547 pst->set_n_tasks(1); 5548 assert(pst->valid(), "Error"); 5549 } 5550 5551 // From space 5552 { 5553 SequentialSubTasksDone* pst = dng->from()->par_seq_tasks(); 5554 assert(!pst->valid(), "Clobbering existing data?"); 5555 size_t n_tasks = _survivor_chunk_index + 1; 5556 assert(n_tasks == 1 || _survivor_chunk_array != NULL, "Error"); 5557 // Sets the condition for completion of the subtask (how many threads 5558 // need to finish in order to be done). 5559 pst->set_n_threads(n_threads); 5560 pst->set_n_tasks((int)n_tasks); 5561 assert(pst->valid(), "Error"); 5562 } 5563 } 5564 5565 // Parallel version of remark 5566 void CMSCollector::do_remark_parallel() { 5567 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5568 FlexibleWorkGang* workers = gch->workers(); 5569 assert(workers != NULL, "Need parallel worker threads."); 5570 // Choose to use the number of GC workers most recently set 5571 // into "active_workers". If active_workers is not set, set it 5572 // to ParallelGCThreads. 5573 int n_workers = workers->active_workers(); 5574 if (n_workers == 0) { 5575 assert(n_workers > 0, "Should have been set during scavenge"); 5576 n_workers = ParallelGCThreads; 5577 workers->set_active_workers(n_workers); 5578 } 5579 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 5580 5581 CMSParRemarkTask tsk(this, 5582 cms_space, 5583 n_workers, workers, task_queues()); 5584 5585 // Set up for parallel process_strong_roots work. 5586 gch->set_par_threads(n_workers); 5587 // We won't be iterating over the cards in the card table updating 5588 // the younger_gen cards, so we shouldn't call the following else 5589 // the verification code as well as subsequent younger_refs_iterate 5590 // code would get confused. XXX 5591 // gch->rem_set()->prepare_for_younger_refs_iterate(true); // parallel 5592 5593 // The young gen rescan work will not be done as part of 5594 // process_strong_roots (which currently doesn't knw how to 5595 // parallelize such a scan), but rather will be broken up into 5596 // a set of parallel tasks (via the sampling that the [abortable] 5597 // preclean phase did of EdenSpace, plus the [two] tasks of 5598 // scanning the [two] survivor spaces. Further fine-grain 5599 // parallelization of the scanning of the survivor spaces 5600 // themselves, and of precleaning of the younger gen itself 5601 // is deferred to the future. 5602 initialize_sequential_subtasks_for_young_gen_rescan(n_workers); 5603 5604 // The dirty card rescan work is broken up into a "sequence" 5605 // of parallel tasks (per constituent space) that are dynamically 5606 // claimed by the parallel threads. 5607 cms_space->initialize_sequential_subtasks_for_rescan(n_workers); 5608 5609 // It turns out that even when we're using 1 thread, doing the work in a 5610 // separate thread causes wide variance in run times. We can't help this 5611 // in the multi-threaded case, but we special-case n=1 here to get 5612 // repeatable measurements of the 1-thread overhead of the parallel code. 5613 if (n_workers > 1) { 5614 // Make refs discovery MT-safe, if it isn't already: it may not 5615 // necessarily be so, since it's possible that we are doing 5616 // ST marking. 5617 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), true); 5618 GenCollectedHeap::StrongRootsScope srs(gch); 5619 workers->run_task(&tsk); 5620 } else { 5621 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false); 5622 GenCollectedHeap::StrongRootsScope srs(gch); 5623 tsk.work(0); 5624 } 5625 5626 gch->set_par_threads(0); // 0 ==> non-parallel. 5627 // restore, single-threaded for now, any preserved marks 5628 // as a result of work_q overflow 5629 restore_preserved_marks_if_any(); 5630 } 5631 5632 // Non-parallel version of remark 5633 void CMSCollector::do_remark_non_parallel() { 5634 ResourceMark rm; 5635 HandleMark hm; 5636 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5637 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false); 5638 5639 MarkRefsIntoAndScanClosure 5640 mrias_cl(_span, ref_processor(), &_markBitMap, NULL /* not precleaning */, 5641 &_markStack, this, 5642 false /* should_yield */, false /* not precleaning */); 5643 MarkFromDirtyCardsClosure 5644 markFromDirtyCardsClosure(this, _span, 5645 NULL, // space is set further below 5646 &_markBitMap, &_markStack, &mrias_cl); 5647 { 5648 TraceTime t("grey object rescan", PrintGCDetails, false, gclog_or_tty); 5649 // Iterate over the dirty cards, setting the corresponding bits in the 5650 // mod union table. 5651 { 5652 ModUnionClosure modUnionClosure(&_modUnionTable); 5653 _ct->ct_bs()->dirty_card_iterate( 5654 _cmsGen->used_region(), 5655 &modUnionClosure); 5656 } 5657 // Having transferred these marks into the modUnionTable, we just need 5658 // to rescan the marked objects on the dirty cards in the modUnionTable. 5659 // The initial marking may have been done during an asynchronous 5660 // collection so there may be dirty bits in the mod-union table. 5661 const int alignment = 5662 CardTableModRefBS::card_size * BitsPerWord; 5663 { 5664 // ... First handle dirty cards in CMS gen 5665 markFromDirtyCardsClosure.set_space(_cmsGen->cmsSpace()); 5666 MemRegion ur = _cmsGen->used_region(); 5667 HeapWord* lb = ur.start(); 5668 HeapWord* ub = (HeapWord*)round_to((intptr_t)ur.end(), alignment); 5669 MemRegion cms_span(lb, ub); 5670 _modUnionTable.dirty_range_iterate_clear(cms_span, 5671 &markFromDirtyCardsClosure); 5672 verify_work_stacks_empty(); 5673 if (PrintCMSStatistics != 0) { 5674 gclog_or_tty->print(" (re-scanned "SIZE_FORMAT" dirty cards in cms gen) ", 5675 markFromDirtyCardsClosure.num_dirty_cards()); 5676 } 5677 } 5678 } 5679 if (VerifyDuringGC && 5680 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 5681 HandleMark hm; // Discard invalid handles created during verification 5682 Universe::verify(true); 5683 } 5684 { 5685 TraceTime t("root rescan", PrintGCDetails, false, gclog_or_tty); 5686 5687 verify_work_stacks_empty(); 5688 5689 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 5690 GenCollectedHeap::StrongRootsScope srs(gch); 5691 gch->gen_process_strong_roots(_cmsGen->level(), 5692 true, // younger gens as roots 5693 false, // use the local StrongRootsScope 5694 false, // not scavenging 5695 SharedHeap::ScanningOption(roots_scanning_options()), 5696 &mrias_cl, 5697 true, // walk code active on stacks 5698 NULL, 5699 NULL); // The dirty klasses will be handled below 5700 5701 assert(should_unload_classes() 5702 || (roots_scanning_options() & SharedHeap::SO_CodeCache), 5703 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 5704 } 5705 5706 { 5707 TraceTime t("visit unhandled CLDs", PrintGCDetails, false, gclog_or_tty); 5708 5709 verify_work_stacks_empty(); 5710 5711 // Scan all class loader data objects that might have been introduced 5712 // during concurrent marking. 5713 ResourceMark rm; 5714 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds(); 5715 for (int i = 0; i < array->length(); i++) { 5716 mrias_cl.do_class_loader_data(array->at(i)); 5717 } 5718 5719 // We don't need to keep track of new CLDs anymore. 5720 ClassLoaderDataGraph::remember_new_clds(false); 5721 5722 verify_work_stacks_empty(); 5723 } 5724 5725 { 5726 TraceTime t("dirty klass scan", PrintGCDetails, false, gclog_or_tty); 5727 5728 verify_work_stacks_empty(); 5729 5730 RemarkKlassClosure remark_klass_closure(&mrias_cl); 5731 ClassLoaderDataGraph::classes_do(&remark_klass_closure); 5732 5733 verify_work_stacks_empty(); 5734 } 5735 5736 // We might have added oops to ClassLoaderData::_handles during the 5737 // concurrent marking phase. These oops point to newly allocated objects 5738 // that are guaranteed to be kept alive. Either by the direct allocation 5739 // code, or when the young collector processes the strong roots. Hence, 5740 // we don't have to revisit the _handles block during the remark phase. 5741 5742 verify_work_stacks_empty(); 5743 // Restore evacuated mark words, if any, used for overflow list links 5744 if (!CMSOverflowEarlyRestoration) { 5745 restore_preserved_marks_if_any(); 5746 } 5747 verify_overflow_empty(); 5748 } 5749 5750 //////////////////////////////////////////////////////// 5751 // Parallel Reference Processing Task Proxy Class 5752 //////////////////////////////////////////////////////// 5753 class CMSRefProcTaskProxy: public AbstractGangTaskWOopQueues { 5754 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask; 5755 CMSCollector* _collector; 5756 CMSBitMap* _mark_bit_map; 5757 const MemRegion _span; 5758 ProcessTask& _task; 5759 5760 public: 5761 CMSRefProcTaskProxy(ProcessTask& task, 5762 CMSCollector* collector, 5763 const MemRegion& span, 5764 CMSBitMap* mark_bit_map, 5765 AbstractWorkGang* workers, 5766 OopTaskQueueSet* task_queues): 5767 // XXX Should superclass AGTWOQ also know about AWG since it knows 5768 // about the task_queues used by the AWG? Then it could initialize 5769 // the terminator() object. See 6984287. The set_for_termination() 5770 // below is a temporary band-aid for the regression in 6984287. 5771 AbstractGangTaskWOopQueues("Process referents by policy in parallel", 5772 task_queues), 5773 _task(task), 5774 _collector(collector), _span(span), _mark_bit_map(mark_bit_map) 5775 { 5776 assert(_collector->_span.equals(_span) && !_span.is_empty(), 5777 "Inconsistency in _span"); 5778 set_for_termination(workers->active_workers()); 5779 } 5780 5781 OopTaskQueueSet* task_queues() { return queues(); } 5782 5783 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 5784 5785 void do_work_steal(int i, 5786 CMSParDrainMarkingStackClosure* drain, 5787 CMSParKeepAliveClosure* keep_alive, 5788 int* seed); 5789 5790 virtual void work(uint worker_id); 5791 }; 5792 5793 void CMSRefProcTaskProxy::work(uint worker_id) { 5794 assert(_collector->_span.equals(_span), "Inconsistency in _span"); 5795 CMSParKeepAliveClosure par_keep_alive(_collector, _span, 5796 _mark_bit_map, 5797 work_queue(worker_id)); 5798 CMSParDrainMarkingStackClosure par_drain_stack(_collector, _span, 5799 _mark_bit_map, 5800 work_queue(worker_id)); 5801 CMSIsAliveClosure is_alive_closure(_span, _mark_bit_map); 5802 _task.work(worker_id, is_alive_closure, par_keep_alive, par_drain_stack); 5803 if (_task.marks_oops_alive()) { 5804 do_work_steal(worker_id, &par_drain_stack, &par_keep_alive, 5805 _collector->hash_seed(worker_id)); 5806 } 5807 assert(work_queue(worker_id)->size() == 0, "work_queue should be empty"); 5808 assert(_collector->_overflow_list == NULL, "non-empty _overflow_list"); 5809 } 5810 5811 class CMSRefEnqueueTaskProxy: public AbstractGangTask { 5812 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask; 5813 EnqueueTask& _task; 5814 5815 public: 5816 CMSRefEnqueueTaskProxy(EnqueueTask& task) 5817 : AbstractGangTask("Enqueue reference objects in parallel"), 5818 _task(task) 5819 { } 5820 5821 virtual void work(uint worker_id) 5822 { 5823 _task.work(worker_id); 5824 } 5825 }; 5826 5827 CMSParKeepAliveClosure::CMSParKeepAliveClosure(CMSCollector* collector, 5828 MemRegion span, CMSBitMap* bit_map, OopTaskQueue* work_queue): 5829 _span(span), 5830 _bit_map(bit_map), 5831 _work_queue(work_queue), 5832 _mark_and_push(collector, span, bit_map, work_queue), 5833 _low_water_mark(MIN2((uint)(work_queue->max_elems()/4), 5834 (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads))) 5835 { } 5836 5837 // . see if we can share work_queues with ParNew? XXX 5838 void CMSRefProcTaskProxy::do_work_steal(int i, 5839 CMSParDrainMarkingStackClosure* drain, 5840 CMSParKeepAliveClosure* keep_alive, 5841 int* seed) { 5842 OopTaskQueue* work_q = work_queue(i); 5843 NOT_PRODUCT(int num_steals = 0;) 5844 oop obj_to_scan; 5845 5846 while (true) { 5847 // Completely finish any left over work from (an) earlier round(s) 5848 drain->trim_queue(0); 5849 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 5850 (size_t)ParGCDesiredObjsFromOverflowList); 5851 // Now check if there's any work in the overflow list 5852 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads, 5853 // only affects the number of attempts made to get work from the 5854 // overflow list and does not affect the number of workers. Just 5855 // pass ParallelGCThreads so this behavior is unchanged. 5856 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 5857 work_q, 5858 ParallelGCThreads)) { 5859 // Found something in global overflow list; 5860 // not yet ready to go stealing work from others. 5861 // We'd like to assert(work_q->size() != 0, ...) 5862 // because we just took work from the overflow list, 5863 // but of course we can't, since all of that might have 5864 // been already stolen from us. 5865 continue; 5866 } 5867 // Verify that we have no work before we resort to stealing 5868 assert(work_q->size() == 0, "Have work, shouldn't steal"); 5869 // Try to steal from other queues that have work 5870 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 5871 NOT_PRODUCT(num_steals++;) 5872 assert(obj_to_scan->is_oop(), "Oops, not an oop!"); 5873 assert(_mark_bit_map->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 5874 // Do scanning work 5875 obj_to_scan->oop_iterate(keep_alive); 5876 // Loop around, finish this work, and try to steal some more 5877 } else if (terminator()->offer_termination()) { 5878 break; // nirvana from the infinite cycle 5879 } 5880 } 5881 NOT_PRODUCT( 5882 if (PrintCMSStatistics != 0) { 5883 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals); 5884 } 5885 ) 5886 } 5887 5888 void CMSRefProcTaskExecutor::execute(ProcessTask& task) 5889 { 5890 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5891 FlexibleWorkGang* workers = gch->workers(); 5892 assert(workers != NULL, "Need parallel worker threads."); 5893 CMSRefProcTaskProxy rp_task(task, &_collector, 5894 _collector.ref_processor()->span(), 5895 _collector.markBitMap(), 5896 workers, _collector.task_queues()); 5897 workers->run_task(&rp_task); 5898 } 5899 5900 void CMSRefProcTaskExecutor::execute(EnqueueTask& task) 5901 { 5902 5903 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5904 FlexibleWorkGang* workers = gch->workers(); 5905 assert(workers != NULL, "Need parallel worker threads."); 5906 CMSRefEnqueueTaskProxy enq_task(task); 5907 workers->run_task(&enq_task); 5908 } 5909 5910 void CMSCollector::refProcessingWork(bool asynch, bool clear_all_soft_refs) { 5911 5912 ResourceMark rm; 5913 HandleMark hm; 5914 5915 ReferenceProcessor* rp = ref_processor(); 5916 assert(rp->span().equals(_span), "Spans should be equal"); 5917 assert(!rp->enqueuing_is_done(), "Enqueuing should not be complete"); 5918 // Process weak references. 5919 rp->setup_policy(clear_all_soft_refs); 5920 verify_work_stacks_empty(); 5921 5922 CMSKeepAliveClosure cmsKeepAliveClosure(this, _span, &_markBitMap, 5923 &_markStack, false /* !preclean */); 5924 CMSDrainMarkingStackClosure cmsDrainMarkingStackClosure(this, 5925 _span, &_markBitMap, &_markStack, 5926 &cmsKeepAliveClosure, false /* !preclean */); 5927 { 5928 TraceTime t("weak refs processing", PrintGCDetails, false, gclog_or_tty); 5929 if (rp->processing_is_mt()) { 5930 // Set the degree of MT here. If the discovery is done MT, there 5931 // may have been a different number of threads doing the discovery 5932 // and a different number of discovered lists may have Ref objects. 5933 // That is OK as long as the Reference lists are balanced (see 5934 // balance_all_queues() and balance_queues()). 5935 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5936 int active_workers = ParallelGCThreads; 5937 FlexibleWorkGang* workers = gch->workers(); 5938 if (workers != NULL) { 5939 active_workers = workers->active_workers(); 5940 // The expectation is that active_workers will have already 5941 // been set to a reasonable value. If it has not been set, 5942 // investigate. 5943 assert(active_workers > 0, "Should have been set during scavenge"); 5944 } 5945 rp->set_active_mt_degree(active_workers); 5946 CMSRefProcTaskExecutor task_executor(*this); 5947 rp->process_discovered_references(&_is_alive_closure, 5948 &cmsKeepAliveClosure, 5949 &cmsDrainMarkingStackClosure, 5950 &task_executor); 5951 } else { 5952 rp->process_discovered_references(&_is_alive_closure, 5953 &cmsKeepAliveClosure, 5954 &cmsDrainMarkingStackClosure, 5955 NULL); 5956 } 5957 verify_work_stacks_empty(); 5958 } 5959 5960 if (should_unload_classes()) { 5961 { 5962 TraceTime t("class unloading", PrintGCDetails, false, gclog_or_tty); 5963 5964 // Follow SystemDictionary roots and unload classes 5965 bool purged_class = SystemDictionary::do_unloading(&_is_alive_closure); 5966 5967 // Follow CodeCache roots and unload any methods marked for unloading 5968 CodeCache::do_unloading(&_is_alive_closure, 5969 &cmsKeepAliveClosure, 5970 purged_class); 5971 5972 cmsDrainMarkingStackClosure.do_void(); 5973 verify_work_stacks_empty(); 5974 5975 // Update subklass/sibling/implementor links in KlassKlass descendants 5976 Klass::clean_weak_klass_links(&_is_alive_closure); 5977 // Nothing should have been pushed onto the working stacks. 5978 verify_work_stacks_empty(); 5979 } 5980 5981 { 5982 TraceTime t("scrub symbol table", PrintGCDetails, false, gclog_or_tty); 5983 // Clean up unreferenced symbols in symbol table. 5984 SymbolTable::unlink(); 5985 } 5986 } 5987 5988 // CMS doesn't use the StringTable as hard roots when class unloading is turned off. 5989 // Need to check if we really scanned the StringTable. 5990 if ((roots_scanning_options() & SharedHeap::SO_Strings) == 0) { 5991 TraceTime t("scrub string table", PrintGCDetails, false, gclog_or_tty); 5992 // Now clean up stale oops in StringTable 5993 StringTable::unlink(&_is_alive_closure); 5994 } 5995 5996 verify_work_stacks_empty(); 5997 // Restore any preserved marks as a result of mark stack or 5998 // work queue overflow 5999 restore_preserved_marks_if_any(); // done single-threaded for now 6000 6001 rp->set_enqueuing_is_done(true); 6002 if (rp->processing_is_mt()) { 6003 rp->balance_all_queues(); 6004 CMSRefProcTaskExecutor task_executor(*this); 6005 rp->enqueue_discovered_references(&task_executor); 6006 } else { 6007 rp->enqueue_discovered_references(NULL); 6008 } 6009 rp->verify_no_references_recorded(); 6010 assert(!rp->discovery_enabled(), "should have been disabled"); 6011 } 6012 6013 #ifndef PRODUCT 6014 void CMSCollector::check_correct_thread_executing() { 6015 Thread* t = Thread::current(); 6016 // Only the VM thread or the CMS thread should be here. 6017 assert(t->is_ConcurrentGC_thread() || t->is_VM_thread(), 6018 "Unexpected thread type"); 6019 // If this is the vm thread, the foreground process 6020 // should not be waiting. Note that _foregroundGCIsActive is 6021 // true while the foreground collector is waiting. 6022 if (_foregroundGCShouldWait) { 6023 // We cannot be the VM thread 6024 assert(t->is_ConcurrentGC_thread(), 6025 "Should be CMS thread"); 6026 } else { 6027 // We can be the CMS thread only if we are in a stop-world 6028 // phase of CMS collection. 6029 if (t->is_ConcurrentGC_thread()) { 6030 assert(_collectorState == InitialMarking || 6031 _collectorState == FinalMarking, 6032 "Should be a stop-world phase"); 6033 // The CMS thread should be holding the CMS_token. 6034 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6035 "Potential interference with concurrently " 6036 "executing VM thread"); 6037 } 6038 } 6039 } 6040 #endif 6041 6042 void CMSCollector::sweep(bool asynch) { 6043 assert(_collectorState == Sweeping, "just checking"); 6044 check_correct_thread_executing(); 6045 verify_work_stacks_empty(); 6046 verify_overflow_empty(); 6047 increment_sweep_count(); 6048 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 6049 6050 _inter_sweep_timer.stop(); 6051 _inter_sweep_estimate.sample(_inter_sweep_timer.seconds()); 6052 size_policy()->avg_cms_free_at_sweep()->sample(_cmsGen->free()); 6053 6054 assert(!_intra_sweep_timer.is_active(), "Should not be active"); 6055 _intra_sweep_timer.reset(); 6056 _intra_sweep_timer.start(); 6057 if (asynch) { 6058 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 6059 CMSPhaseAccounting pa(this, "sweep", !PrintGCDetails); 6060 // First sweep the old gen 6061 { 6062 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(), 6063 bitMapLock()); 6064 sweepWork(_cmsGen, asynch); 6065 } 6066 6067 // Update Universe::_heap_*_at_gc figures. 6068 // We need all the free list locks to make the abstract state 6069 // transition from Sweeping to Resetting. See detailed note 6070 // further below. 6071 { 6072 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock()); 6073 // Update heap occupancy information which is used as 6074 // input to soft ref clearing policy at the next gc. 6075 Universe::update_heap_info_at_gc(); 6076 _collectorState = Resizing; 6077 } 6078 } else { 6079 // already have needed locks 6080 sweepWork(_cmsGen, asynch); 6081 // Update heap occupancy information which is used as 6082 // input to soft ref clearing policy at the next gc. 6083 Universe::update_heap_info_at_gc(); 6084 _collectorState = Resizing; 6085 } 6086 verify_work_stacks_empty(); 6087 verify_overflow_empty(); 6088 6089 _intra_sweep_timer.stop(); 6090 _intra_sweep_estimate.sample(_intra_sweep_timer.seconds()); 6091 6092 _inter_sweep_timer.reset(); 6093 _inter_sweep_timer.start(); 6094 6095 // We need to use a monotonically non-deccreasing time in ms 6096 // or we will see time-warp warnings and os::javaTimeMillis() 6097 // does not guarantee monotonicity. 6098 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC; 6099 update_time_of_last_gc(now); 6100 6101 // NOTE on abstract state transitions: 6102 // Mutators allocate-live and/or mark the mod-union table dirty 6103 // based on the state of the collection. The former is done in 6104 // the interval [Marking, Sweeping] and the latter in the interval 6105 // [Marking, Sweeping). Thus the transitions into the Marking state 6106 // and out of the Sweeping state must be synchronously visible 6107 // globally to the mutators. 6108 // The transition into the Marking state happens with the world 6109 // stopped so the mutators will globally see it. Sweeping is 6110 // done asynchronously by the background collector so the transition 6111 // from the Sweeping state to the Resizing state must be done 6112 // under the freelistLock (as is the check for whether to 6113 // allocate-live and whether to dirty the mod-union table). 6114 assert(_collectorState == Resizing, "Change of collector state to" 6115 " Resizing must be done under the freelistLocks (plural)"); 6116 6117 // Now that sweeping has been completed, we clear 6118 // the incremental_collection_failed flag, 6119 // thus inviting a younger gen collection to promote into 6120 // this generation. If such a promotion may still fail, 6121 // the flag will be set again when a young collection is 6122 // attempted. 6123 GenCollectedHeap* gch = GenCollectedHeap::heap(); 6124 gch->clear_incremental_collection_failed(); // Worth retrying as fresh space may have been freed up 6125 gch->update_full_collections_completed(_collection_count_start); 6126 } 6127 6128 // FIX ME!!! Looks like this belongs in CFLSpace, with 6129 // CMSGen merely delegating to it. 6130 void ConcurrentMarkSweepGeneration::setNearLargestChunk() { 6131 double nearLargestPercent = FLSLargestBlockCoalesceProximity; 6132 HeapWord* minAddr = _cmsSpace->bottom(); 6133 HeapWord* largestAddr = 6134 (HeapWord*) _cmsSpace->dictionary()->find_largest_dict(); 6135 if (largestAddr == NULL) { 6136 // The dictionary appears to be empty. In this case 6137 // try to coalesce at the end of the heap. 6138 largestAddr = _cmsSpace->end(); 6139 } 6140 size_t largestOffset = pointer_delta(largestAddr, minAddr); 6141 size_t nearLargestOffset = 6142 (size_t)((double)largestOffset * nearLargestPercent) - MinChunkSize; 6143 if (PrintFLSStatistics != 0) { 6144 gclog_or_tty->print_cr( 6145 "CMS: Large Block: " PTR_FORMAT ";" 6146 " Proximity: " PTR_FORMAT " -> " PTR_FORMAT, 6147 largestAddr, 6148 _cmsSpace->nearLargestChunk(), minAddr + nearLargestOffset); 6149 } 6150 _cmsSpace->set_nearLargestChunk(minAddr + nearLargestOffset); 6151 } 6152 6153 bool ConcurrentMarkSweepGeneration::isNearLargestChunk(HeapWord* addr) { 6154 return addr >= _cmsSpace->nearLargestChunk(); 6155 } 6156 6157 FreeChunk* ConcurrentMarkSweepGeneration::find_chunk_at_end() { 6158 return _cmsSpace->find_chunk_at_end(); 6159 } 6160 6161 void ConcurrentMarkSweepGeneration::update_gc_stats(int current_level, 6162 bool full) { 6163 // The next lower level has been collected. Gather any statistics 6164 // that are of interest at this point. 6165 if (!full && (current_level + 1) == level()) { 6166 // Gather statistics on the young generation collection. 6167 collector()->stats().record_gc0_end(used()); 6168 } 6169 } 6170 6171 CMSAdaptiveSizePolicy* ConcurrentMarkSweepGeneration::size_policy() { 6172 GenCollectedHeap* gch = GenCollectedHeap::heap(); 6173 assert(gch->kind() == CollectedHeap::GenCollectedHeap, 6174 "Wrong type of heap"); 6175 CMSAdaptiveSizePolicy* sp = (CMSAdaptiveSizePolicy*) 6176 gch->gen_policy()->size_policy(); 6177 assert(sp->is_gc_cms_adaptive_size_policy(), 6178 "Wrong type of size policy"); 6179 return sp; 6180 } 6181 6182 void ConcurrentMarkSweepGeneration::rotate_debug_collection_type() { 6183 if (PrintGCDetails && Verbose) { 6184 gclog_or_tty->print("Rotate from %d ", _debug_collection_type); 6185 } 6186 _debug_collection_type = (CollectionTypes) (_debug_collection_type + 1); 6187 _debug_collection_type = 6188 (CollectionTypes) (_debug_collection_type % Unknown_collection_type); 6189 if (PrintGCDetails && Verbose) { 6190 gclog_or_tty->print_cr("to %d ", _debug_collection_type); 6191 } 6192 } 6193 6194 void CMSCollector::sweepWork(ConcurrentMarkSweepGeneration* gen, 6195 bool asynch) { 6196 // We iterate over the space(s) underlying this generation, 6197 // checking the mark bit map to see if the bits corresponding 6198 // to specific blocks are marked or not. Blocks that are 6199 // marked are live and are not swept up. All remaining blocks 6200 // are swept up, with coalescing on-the-fly as we sweep up 6201 // contiguous free and/or garbage blocks: 6202 // We need to ensure that the sweeper synchronizes with allocators 6203 // and stop-the-world collectors. In particular, the following 6204 // locks are used: 6205 // . CMS token: if this is held, a stop the world collection cannot occur 6206 // . freelistLock: if this is held no allocation can occur from this 6207 // generation by another thread 6208 // . bitMapLock: if this is held, no other thread can access or update 6209 // 6210 6211 // Note that we need to hold the freelistLock if we use 6212 // block iterate below; else the iterator might go awry if 6213 // a mutator (or promotion) causes block contents to change 6214 // (for instance if the allocator divvies up a block). 6215 // If we hold the free list lock, for all practical purposes 6216 // young generation GC's can't occur (they'll usually need to 6217 // promote), so we might as well prevent all young generation 6218 // GC's while we do a sweeping step. For the same reason, we might 6219 // as well take the bit map lock for the entire duration 6220 6221 // check that we hold the requisite locks 6222 assert(have_cms_token(), "Should hold cms token"); 6223 assert( (asynch && ConcurrentMarkSweepThread::cms_thread_has_cms_token()) 6224 || (!asynch && ConcurrentMarkSweepThread::vm_thread_has_cms_token()), 6225 "Should possess CMS token to sweep"); 6226 assert_lock_strong(gen->freelistLock()); 6227 assert_lock_strong(bitMapLock()); 6228 6229 assert(!_inter_sweep_timer.is_active(), "Was switched off in an outer context"); 6230 assert(_intra_sweep_timer.is_active(), "Was switched on in an outer context"); 6231 gen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()), 6232 _inter_sweep_estimate.padded_average(), 6233 _intra_sweep_estimate.padded_average()); 6234 gen->setNearLargestChunk(); 6235 6236 { 6237 SweepClosure sweepClosure(this, gen, &_markBitMap, 6238 CMSYield && asynch); 6239 gen->cmsSpace()->blk_iterate_careful(&sweepClosure); 6240 // We need to free-up/coalesce garbage/blocks from a 6241 // co-terminal free run. This is done in the SweepClosure 6242 // destructor; so, do not remove this scope, else the 6243 // end-of-sweep-census below will be off by a little bit. 6244 } 6245 gen->cmsSpace()->sweep_completed(); 6246 gen->cmsSpace()->endSweepFLCensus(sweep_count()); 6247 if (should_unload_classes()) { // unloaded classes this cycle, 6248 _concurrent_cycles_since_last_unload = 0; // ... reset count 6249 } else { // did not unload classes, 6250 _concurrent_cycles_since_last_unload++; // ... increment count 6251 } 6252 } 6253 6254 // Reset CMS data structures (for now just the marking bit map) 6255 // preparatory for the next cycle. 6256 void CMSCollector::reset(bool asynch) { 6257 GenCollectedHeap* gch = GenCollectedHeap::heap(); 6258 CMSAdaptiveSizePolicy* sp = size_policy(); 6259 AdaptiveSizePolicyOutput(sp, gch->total_collections()); 6260 if (asynch) { 6261 CMSTokenSyncWithLocks ts(true, bitMapLock()); 6262 6263 // If the state is not "Resetting", the foreground thread 6264 // has done a collection and the resetting. 6265 if (_collectorState != Resetting) { 6266 assert(_collectorState == Idling, "The state should only change" 6267 " because the foreground collector has finished the collection"); 6268 return; 6269 } 6270 6271 // Clear the mark bitmap (no grey objects to start with) 6272 // for the next cycle. 6273 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 6274 CMSPhaseAccounting cmspa(this, "reset", !PrintGCDetails); 6275 6276 HeapWord* curAddr = _markBitMap.startWord(); 6277 while (curAddr < _markBitMap.endWord()) { 6278 size_t remaining = pointer_delta(_markBitMap.endWord(), curAddr); 6279 MemRegion chunk(curAddr, MIN2(CMSBitMapYieldQuantum, remaining)); 6280 _markBitMap.clear_large_range(chunk); 6281 if (ConcurrentMarkSweepThread::should_yield() && 6282 !foregroundGCIsActive() && 6283 CMSYield) { 6284 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6285 "CMS thread should hold CMS token"); 6286 assert_lock_strong(bitMapLock()); 6287 bitMapLock()->unlock(); 6288 ConcurrentMarkSweepThread::desynchronize(true); 6289 ConcurrentMarkSweepThread::acknowledge_yield_request(); 6290 stopTimer(); 6291 if (PrintCMSStatistics != 0) { 6292 incrementYields(); 6293 } 6294 icms_wait(); 6295 6296 // See the comment in coordinator_yield() 6297 for (unsigned i = 0; i < CMSYieldSleepCount && 6298 ConcurrentMarkSweepThread::should_yield() && 6299 !CMSCollector::foregroundGCIsActive(); ++i) { 6300 os::sleep(Thread::current(), 1, false); 6301 ConcurrentMarkSweepThread::acknowledge_yield_request(); 6302 } 6303 6304 ConcurrentMarkSweepThread::synchronize(true); 6305 bitMapLock()->lock_without_safepoint_check(); 6306 startTimer(); 6307 } 6308 curAddr = chunk.end(); 6309 } 6310 // A successful mostly concurrent collection has been done. 6311 // Because only the full (i.e., concurrent mode failure) collections 6312 // are being measured for gc overhead limits, clean the "near" flag 6313 // and count. 6314 sp->reset_gc_overhead_limit_count(); 6315 _collectorState = Idling; 6316 } else { 6317 // already have the lock 6318 assert(_collectorState == Resetting, "just checking"); 6319 assert_lock_strong(bitMapLock()); 6320 _markBitMap.clear_all(); 6321 _collectorState = Idling; 6322 } 6323 6324 // Stop incremental mode after a cycle completes, so that any future cycles 6325 // are triggered by allocation. 6326 stop_icms(); 6327 6328 NOT_PRODUCT( 6329 if (RotateCMSCollectionTypes) { 6330 _cmsGen->rotate_debug_collection_type(); 6331 } 6332 ) 6333 } 6334 6335 void CMSCollector::do_CMS_operation(CMS_op_type op, GCCause::Cause gc_cause) { 6336 gclog_or_tty->date_stamp(PrintGC && PrintGCDateStamps); 6337 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 6338 TraceTime t(GCCauseString("GC", gc_cause), PrintGC, !PrintGCDetails, gclog_or_tty); 6339 TraceCollectorStats tcs(counters()); 6340 6341 switch (op) { 6342 case CMS_op_checkpointRootsInitial: { 6343 SvcGCMarker sgcm(SvcGCMarker::OTHER); 6344 checkpointRootsInitial(true); // asynch 6345 if (PrintGC) { 6346 _cmsGen->printOccupancy("initial-mark"); 6347 } 6348 break; 6349 } 6350 case CMS_op_checkpointRootsFinal: { 6351 SvcGCMarker sgcm(SvcGCMarker::OTHER); 6352 checkpointRootsFinal(true, // asynch 6353 false, // !clear_all_soft_refs 6354 false); // !init_mark_was_synchronous 6355 if (PrintGC) { 6356 _cmsGen->printOccupancy("remark"); 6357 } 6358 break; 6359 } 6360 default: 6361 fatal("No such CMS_op"); 6362 } 6363 } 6364 6365 #ifndef PRODUCT 6366 size_t const CMSCollector::skip_header_HeapWords() { 6367 return FreeChunk::header_size(); 6368 } 6369 6370 // Try and collect here conditions that should hold when 6371 // CMS thread is exiting. The idea is that the foreground GC 6372 // thread should not be blocked if it wants to terminate 6373 // the CMS thread and yet continue to run the VM for a while 6374 // after that. 6375 void CMSCollector::verify_ok_to_terminate() const { 6376 assert(Thread::current()->is_ConcurrentGC_thread(), 6377 "should be called by CMS thread"); 6378 assert(!_foregroundGCShouldWait, "should be false"); 6379 // We could check here that all the various low-level locks 6380 // are not held by the CMS thread, but that is overkill; see 6381 // also CMSThread::verify_ok_to_terminate() where the CGC_lock 6382 // is checked. 6383 } 6384 #endif 6385 6386 size_t CMSCollector::block_size_using_printezis_bits(HeapWord* addr) const { 6387 assert(_markBitMap.isMarked(addr) && _markBitMap.isMarked(addr + 1), 6388 "missing Printezis mark?"); 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 6397 // A variant of the above (block_size_using_printezis_bits()) except 6398 // that we return 0 if the P-bits are not yet set. 6399 size_t CMSCollector::block_size_if_printezis_bits(HeapWord* addr) const { 6400 if (_markBitMap.isMarked(addr + 1)) { 6401 assert(_markBitMap.isMarked(addr), "P-bit can be set only for marked objects"); 6402 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 6403 size_t size = pointer_delta(nextOneAddr + 1, addr); 6404 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 6405 "alignment problem"); 6406 assert(size >= 3, "Necessary for Printezis marks to work"); 6407 return size; 6408 } 6409 return 0; 6410 } 6411 6412 HeapWord* CMSCollector::next_card_start_after_block(HeapWord* addr) const { 6413 size_t sz = 0; 6414 oop p = (oop)addr; 6415 if (p->klass_or_null() != NULL) { 6416 sz = CompactibleFreeListSpace::adjustObjectSize(p->size()); 6417 } else { 6418 sz = block_size_using_printezis_bits(addr); 6419 } 6420 assert(sz > 0, "size must be nonzero"); 6421 HeapWord* next_block = addr + sz; 6422 HeapWord* next_card = (HeapWord*)round_to((uintptr_t)next_block, 6423 CardTableModRefBS::card_size); 6424 assert(round_down((uintptr_t)addr, CardTableModRefBS::card_size) < 6425 round_down((uintptr_t)next_card, CardTableModRefBS::card_size), 6426 "must be different cards"); 6427 return next_card; 6428 } 6429 6430 6431 // CMS Bit Map Wrapper ///////////////////////////////////////// 6432 6433 // Construct a CMS bit map infrastructure, but don't create the 6434 // bit vector itself. That is done by a separate call CMSBitMap::allocate() 6435 // further below. 6436 CMSBitMap::CMSBitMap(int shifter, int mutex_rank, const char* mutex_name): 6437 _bm(), 6438 _shifter(shifter), 6439 _lock(mutex_rank >= 0 ? new Mutex(mutex_rank, mutex_name, true) : NULL) 6440 { 6441 _bmStartWord = 0; 6442 _bmWordSize = 0; 6443 } 6444 6445 bool CMSBitMap::allocate(MemRegion mr) { 6446 _bmStartWord = mr.start(); 6447 _bmWordSize = mr.word_size(); 6448 ReservedSpace brs(ReservedSpace::allocation_align_size_up( 6449 (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1)); 6450 if (!brs.is_reserved()) { 6451 warning("CMS bit map allocation failure"); 6452 return false; 6453 } 6454 // For now we'll just commit all of the bit map up fromt. 6455 // Later on we'll try to be more parsimonious with swap. 6456 if (!_virtual_space.initialize(brs, brs.size())) { 6457 warning("CMS bit map backing store failure"); 6458 return false; 6459 } 6460 assert(_virtual_space.committed_size() == brs.size(), 6461 "didn't reserve backing store for all of CMS bit map?"); 6462 _bm.set_map((BitMap::bm_word_t*)_virtual_space.low()); 6463 assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >= 6464 _bmWordSize, "inconsistency in bit map sizing"); 6465 _bm.set_size(_bmWordSize >> _shifter); 6466 6467 // bm.clear(); // can we rely on getting zero'd memory? verify below 6468 assert(isAllClear(), 6469 "Expected zero'd memory from ReservedSpace constructor"); 6470 assert(_bm.size() == heapWordDiffToOffsetDiff(sizeInWords()), 6471 "consistency check"); 6472 return true; 6473 } 6474 6475 void CMSBitMap::dirty_range_iterate_clear(MemRegion mr, MemRegionClosure* cl) { 6476 HeapWord *next_addr, *end_addr, *last_addr; 6477 assert_locked(); 6478 assert(covers(mr), "out-of-range error"); 6479 // XXX assert that start and end are appropriately aligned 6480 for (next_addr = mr.start(), end_addr = mr.end(); 6481 next_addr < end_addr; next_addr = last_addr) { 6482 MemRegion dirty_region = getAndClearMarkedRegion(next_addr, end_addr); 6483 last_addr = dirty_region.end(); 6484 if (!dirty_region.is_empty()) { 6485 cl->do_MemRegion(dirty_region); 6486 } else { 6487 assert(last_addr == end_addr, "program logic"); 6488 return; 6489 } 6490 } 6491 } 6492 6493 #ifndef PRODUCT 6494 void CMSBitMap::assert_locked() const { 6495 CMSLockVerifier::assert_locked(lock()); 6496 } 6497 6498 bool CMSBitMap::covers(MemRegion mr) const { 6499 // assert(_bm.map() == _virtual_space.low(), "map inconsistency"); 6500 assert((size_t)_bm.size() == (_bmWordSize >> _shifter), 6501 "size inconsistency"); 6502 return (mr.start() >= _bmStartWord) && 6503 (mr.end() <= endWord()); 6504 } 6505 6506 bool CMSBitMap::covers(HeapWord* start, size_t size) const { 6507 return (start >= _bmStartWord && (start + size) <= endWord()); 6508 } 6509 6510 void CMSBitMap::verifyNoOneBitsInRange(HeapWord* left, HeapWord* right) { 6511 // verify that there are no 1 bits in the interval [left, right) 6512 FalseBitMapClosure falseBitMapClosure; 6513 iterate(&falseBitMapClosure, left, right); 6514 } 6515 6516 void CMSBitMap::region_invariant(MemRegion mr) 6517 { 6518 assert_locked(); 6519 // mr = mr.intersection(MemRegion(_bmStartWord, _bmWordSize)); 6520 assert(!mr.is_empty(), "unexpected empty region"); 6521 assert(covers(mr), "mr should be covered by bit map"); 6522 // convert address range into offset range 6523 size_t start_ofs = heapWordToOffset(mr.start()); 6524 // Make sure that end() is appropriately aligned 6525 assert(mr.end() == (HeapWord*)round_to((intptr_t)mr.end(), 6526 (1 << (_shifter+LogHeapWordSize))), 6527 "Misaligned mr.end()"); 6528 size_t end_ofs = heapWordToOffset(mr.end()); 6529 assert(end_ofs > start_ofs, "Should mark at least one bit"); 6530 } 6531 6532 #endif 6533 6534 bool CMSMarkStack::allocate(size_t size) { 6535 // allocate a stack of the requisite depth 6536 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 6537 size * sizeof(oop))); 6538 if (!rs.is_reserved()) { 6539 warning("CMSMarkStack allocation failure"); 6540 return false; 6541 } 6542 if (!_virtual_space.initialize(rs, rs.size())) { 6543 warning("CMSMarkStack backing store failure"); 6544 return false; 6545 } 6546 assert(_virtual_space.committed_size() == rs.size(), 6547 "didn't reserve backing store for all of CMS stack?"); 6548 _base = (oop*)(_virtual_space.low()); 6549 _index = 0; 6550 _capacity = size; 6551 NOT_PRODUCT(_max_depth = 0); 6552 return true; 6553 } 6554 6555 // XXX FIX ME !!! In the MT case we come in here holding a 6556 // leaf lock. For printing we need to take a further lock 6557 // which has lower rank. We need to recallibrate the two 6558 // lock-ranks involved in order to be able to rpint the 6559 // messages below. (Or defer the printing to the caller. 6560 // For now we take the expedient path of just disabling the 6561 // messages for the problematic case.) 6562 void CMSMarkStack::expand() { 6563 assert(_capacity <= MarkStackSizeMax, "stack bigger than permitted"); 6564 if (_capacity == MarkStackSizeMax) { 6565 if (_hit_limit++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) { 6566 // We print a warning message only once per CMS cycle. 6567 gclog_or_tty->print_cr(" (benign) Hit CMSMarkStack max size limit"); 6568 } 6569 return; 6570 } 6571 // Double capacity if possible 6572 size_t new_capacity = MIN2(_capacity*2, MarkStackSizeMax); 6573 // Do not give up existing stack until we have managed to 6574 // get the double capacity that we desired. 6575 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 6576 new_capacity * sizeof(oop))); 6577 if (rs.is_reserved()) { 6578 // Release the backing store associated with old stack 6579 _virtual_space.release(); 6580 // Reinitialize virtual space for new stack 6581 if (!_virtual_space.initialize(rs, rs.size())) { 6582 fatal("Not enough swap for expanded marking stack"); 6583 } 6584 _base = (oop*)(_virtual_space.low()); 6585 _index = 0; 6586 _capacity = new_capacity; 6587 } else if (_failed_double++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) { 6588 // Failed to double capacity, continue; 6589 // we print a detail message only once per CMS cycle. 6590 gclog_or_tty->print(" (benign) Failed to expand marking stack from "SIZE_FORMAT"K to " 6591 SIZE_FORMAT"K", 6592 _capacity / K, new_capacity / K); 6593 } 6594 } 6595 6596 6597 // Closures 6598 // XXX: there seems to be a lot of code duplication here; 6599 // should refactor and consolidate common code. 6600 6601 // This closure is used to mark refs into the CMS generation in 6602 // the CMS bit map. Called at the first checkpoint. This closure 6603 // assumes that we do not need to re-mark dirty cards; if the CMS 6604 // generation on which this is used is not an oldest 6605 // generation then this will lose younger_gen cards! 6606 6607 MarkRefsIntoClosure::MarkRefsIntoClosure( 6608 MemRegion span, CMSBitMap* bitMap): 6609 _span(span), 6610 _bitMap(bitMap) 6611 { 6612 assert(_ref_processor == NULL, "deliberately left NULL"); 6613 assert(_bitMap->covers(_span), "_bitMap/_span mismatch"); 6614 } 6615 6616 void MarkRefsIntoClosure::do_oop(oop obj) { 6617 // if p points into _span, then mark corresponding bit in _markBitMap 6618 assert(obj->is_oop(), "expected an oop"); 6619 HeapWord* addr = (HeapWord*)obj; 6620 if (_span.contains(addr)) { 6621 // this should be made more efficient 6622 _bitMap->mark(addr); 6623 } 6624 } 6625 6626 void MarkRefsIntoClosure::do_oop(oop* p) { MarkRefsIntoClosure::do_oop_work(p); } 6627 void MarkRefsIntoClosure::do_oop(narrowOop* p) { MarkRefsIntoClosure::do_oop_work(p); } 6628 6629 // A variant of the above, used for CMS marking verification. 6630 MarkRefsIntoVerifyClosure::MarkRefsIntoVerifyClosure( 6631 MemRegion span, CMSBitMap* verification_bm, CMSBitMap* cms_bm): 6632 _span(span), 6633 _verification_bm(verification_bm), 6634 _cms_bm(cms_bm) 6635 { 6636 assert(_ref_processor == NULL, "deliberately left NULL"); 6637 assert(_verification_bm->covers(_span), "_verification_bm/_span mismatch"); 6638 } 6639 6640 void MarkRefsIntoVerifyClosure::do_oop(oop obj) { 6641 // if p points into _span, then mark corresponding bit in _markBitMap 6642 assert(obj->is_oop(), "expected an oop"); 6643 HeapWord* addr = (HeapWord*)obj; 6644 if (_span.contains(addr)) { 6645 _verification_bm->mark(addr); 6646 if (!_cms_bm->isMarked(addr)) { 6647 oop(addr)->print(); 6648 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)", addr); 6649 fatal("... aborting"); 6650 } 6651 } 6652 } 6653 6654 void MarkRefsIntoVerifyClosure::do_oop(oop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); } 6655 void MarkRefsIntoVerifyClosure::do_oop(narrowOop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); } 6656 6657 ////////////////////////////////////////////////// 6658 // MarkRefsIntoAndScanClosure 6659 ////////////////////////////////////////////////// 6660 6661 MarkRefsIntoAndScanClosure::MarkRefsIntoAndScanClosure(MemRegion span, 6662 ReferenceProcessor* rp, 6663 CMSBitMap* bit_map, 6664 CMSBitMap* mod_union_table, 6665 CMSMarkStack* mark_stack, 6666 CMSCollector* collector, 6667 bool should_yield, 6668 bool concurrent_precleaning): 6669 _collector(collector), 6670 _span(span), 6671 _bit_map(bit_map), 6672 _mark_stack(mark_stack), 6673 _pushAndMarkClosure(collector, span, rp, bit_map, mod_union_table, 6674 mark_stack, concurrent_precleaning), 6675 _yield(should_yield), 6676 _concurrent_precleaning(concurrent_precleaning), 6677 _freelistLock(NULL) 6678 { 6679 _ref_processor = rp; 6680 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 6681 } 6682 6683 // This closure is used to mark refs into the CMS generation at the 6684 // second (final) checkpoint, and to scan and transitively follow 6685 // the unmarked oops. It is also used during the concurrent precleaning 6686 // phase while scanning objects on dirty cards in the CMS generation. 6687 // The marks are made in the marking bit map and the marking stack is 6688 // used for keeping the (newly) grey objects during the scan. 6689 // The parallel version (Par_...) appears further below. 6690 void MarkRefsIntoAndScanClosure::do_oop(oop obj) { 6691 if (obj != NULL) { 6692 assert(obj->is_oop(), "expected an oop"); 6693 HeapWord* addr = (HeapWord*)obj; 6694 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 6695 assert(_collector->overflow_list_is_empty(), 6696 "overflow list should be empty"); 6697 if (_span.contains(addr) && 6698 !_bit_map->isMarked(addr)) { 6699 // mark bit map (object is now grey) 6700 _bit_map->mark(addr); 6701 // push on marking stack (stack should be empty), and drain the 6702 // stack by applying this closure to the oops in the oops popped 6703 // from the stack (i.e. blacken the grey objects) 6704 bool res = _mark_stack->push(obj); 6705 assert(res, "Should have space to push on empty stack"); 6706 do { 6707 oop new_oop = _mark_stack->pop(); 6708 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 6709 assert(_bit_map->isMarked((HeapWord*)new_oop), 6710 "only grey objects on this stack"); 6711 // iterate over the oops in this oop, marking and pushing 6712 // the ones in CMS heap (i.e. in _span). 6713 new_oop->oop_iterate(&_pushAndMarkClosure); 6714 // check if it's time to yield 6715 do_yield_check(); 6716 } while (!_mark_stack->isEmpty() || 6717 (!_concurrent_precleaning && take_from_overflow_list())); 6718 // if marking stack is empty, and we are not doing this 6719 // during precleaning, then check the overflow list 6720 } 6721 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 6722 assert(_collector->overflow_list_is_empty(), 6723 "overflow list was drained above"); 6724 // We could restore evacuated mark words, if any, used for 6725 // overflow list links here because the overflow list is 6726 // provably empty here. That would reduce the maximum 6727 // size requirements for preserved_{oop,mark}_stack. 6728 // But we'll just postpone it until we are all done 6729 // so we can just stream through. 6730 if (!_concurrent_precleaning && CMSOverflowEarlyRestoration) { 6731 _collector->restore_preserved_marks_if_any(); 6732 assert(_collector->no_preserved_marks(), "No preserved marks"); 6733 } 6734 assert(!CMSOverflowEarlyRestoration || _collector->no_preserved_marks(), 6735 "All preserved marks should have been restored above"); 6736 } 6737 } 6738 6739 void MarkRefsIntoAndScanClosure::do_oop(oop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); } 6740 void MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); } 6741 6742 void MarkRefsIntoAndScanClosure::do_yield_work() { 6743 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6744 "CMS thread should hold CMS token"); 6745 assert_lock_strong(_freelistLock); 6746 assert_lock_strong(_bit_map->lock()); 6747 // relinquish the free_list_lock and bitMaplock() 6748 _bit_map->lock()->unlock(); 6749 _freelistLock->unlock(); 6750 ConcurrentMarkSweepThread::desynchronize(true); 6751 ConcurrentMarkSweepThread::acknowledge_yield_request(); 6752 _collector->stopTimer(); 6753 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 6754 if (PrintCMSStatistics != 0) { 6755 _collector->incrementYields(); 6756 } 6757 _collector->icms_wait(); 6758 6759 // See the comment in coordinator_yield() 6760 for (unsigned i = 0; 6761 i < CMSYieldSleepCount && 6762 ConcurrentMarkSweepThread::should_yield() && 6763 !CMSCollector::foregroundGCIsActive(); 6764 ++i) { 6765 os::sleep(Thread::current(), 1, false); 6766 ConcurrentMarkSweepThread::acknowledge_yield_request(); 6767 } 6768 6769 ConcurrentMarkSweepThread::synchronize(true); 6770 _freelistLock->lock_without_safepoint_check(); 6771 _bit_map->lock()->lock_without_safepoint_check(); 6772 _collector->startTimer(); 6773 } 6774 6775 /////////////////////////////////////////////////////////// 6776 // Par_MarkRefsIntoAndScanClosure: a parallel version of 6777 // MarkRefsIntoAndScanClosure 6778 /////////////////////////////////////////////////////////// 6779 Par_MarkRefsIntoAndScanClosure::Par_MarkRefsIntoAndScanClosure( 6780 CMSCollector* collector, MemRegion span, ReferenceProcessor* rp, 6781 CMSBitMap* bit_map, OopTaskQueue* work_queue): 6782 _span(span), 6783 _bit_map(bit_map), 6784 _work_queue(work_queue), 6785 _low_water_mark(MIN2((uint)(work_queue->max_elems()/4), 6786 (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads))), 6787 _par_pushAndMarkClosure(collector, span, rp, bit_map, work_queue) 6788 { 6789 _ref_processor = rp; 6790 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 6791 } 6792 6793 // This closure is used to mark refs into the CMS generation at the 6794 // second (final) checkpoint, and to scan and transitively follow 6795 // the unmarked oops. The marks are made in the marking bit map and 6796 // the work_queue is used for keeping the (newly) grey objects during 6797 // the scan phase whence they are also available for stealing by parallel 6798 // threads. Since the marking bit map is shared, updates are 6799 // synchronized (via CAS). 6800 void Par_MarkRefsIntoAndScanClosure::do_oop(oop obj) { 6801 if (obj != NULL) { 6802 // Ignore mark word because this could be an already marked oop 6803 // that may be chained at the end of the overflow list. 6804 assert(obj->is_oop(true), "expected an oop"); 6805 HeapWord* addr = (HeapWord*)obj; 6806 if (_span.contains(addr) && 6807 !_bit_map->isMarked(addr)) { 6808 // mark bit map (object will become grey): 6809 // It is possible for several threads to be 6810 // trying to "claim" this object concurrently; 6811 // the unique thread that succeeds in marking the 6812 // object first will do the subsequent push on 6813 // to the work queue (or overflow list). 6814 if (_bit_map->par_mark(addr)) { 6815 // push on work_queue (which may not be empty), and trim the 6816 // queue to an appropriate length by applying this closure to 6817 // the oops in the oops popped from the stack (i.e. blacken the 6818 // grey objects) 6819 bool res = _work_queue->push(obj); 6820 assert(res, "Low water mark should be less than capacity?"); 6821 trim_queue(_low_water_mark); 6822 } // Else, another thread claimed the object 6823 } 6824 } 6825 } 6826 6827 void Par_MarkRefsIntoAndScanClosure::do_oop(oop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); } 6828 void Par_MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); } 6829 6830 // This closure is used to rescan the marked objects on the dirty cards 6831 // in the mod union table and the card table proper. 6832 size_t ScanMarkedObjectsAgainCarefullyClosure::do_object_careful_m( 6833 oop p, MemRegion mr) { 6834 6835 size_t size = 0; 6836 HeapWord* addr = (HeapWord*)p; 6837 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6838 assert(_span.contains(addr), "we are scanning the CMS generation"); 6839 // check if it's time to yield 6840 if (do_yield_check()) { 6841 // We yielded for some foreground stop-world work, 6842 // and we have been asked to abort this ongoing preclean cycle. 6843 return 0; 6844 } 6845 if (_bitMap->isMarked(addr)) { 6846 // it's marked; is it potentially uninitialized? 6847 if (p->klass_or_null() != NULL) { 6848 // an initialized object; ignore mark word in verification below 6849 // since we are running concurrent with mutators 6850 assert(p->is_oop(true), "should be an oop"); 6851 if (p->is_objArray()) { 6852 // objArrays are precisely marked; restrict scanning 6853 // to dirty cards only. 6854 size = CompactibleFreeListSpace::adjustObjectSize( 6855 p->oop_iterate(_scanningClosure, mr)); 6856 } else { 6857 // A non-array may have been imprecisely marked; we need 6858 // to scan object in its entirety. 6859 size = CompactibleFreeListSpace::adjustObjectSize( 6860 p->oop_iterate(_scanningClosure)); 6861 } 6862 #ifdef DEBUG 6863 size_t direct_size = 6864 CompactibleFreeListSpace::adjustObjectSize(p->size()); 6865 assert(size == direct_size, "Inconsistency in size"); 6866 assert(size >= 3, "Necessary for Printezis marks to work"); 6867 if (!_bitMap->isMarked(addr+1)) { 6868 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size); 6869 } else { 6870 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size-1); 6871 assert(_bitMap->isMarked(addr+size-1), 6872 "inconsistent Printezis mark"); 6873 } 6874 #endif // DEBUG 6875 } else { 6876 // an unitialized object 6877 assert(_bitMap->isMarked(addr+1), "missing Printezis mark?"); 6878 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 6879 size = pointer_delta(nextOneAddr + 1, addr); 6880 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 6881 "alignment problem"); 6882 // Note that pre-cleaning needn't redirty the card. OopDesc::set_klass() 6883 // will dirty the card when the klass pointer is installed in the 6884 // object (signalling the completion of initialization). 6885 } 6886 } else { 6887 // Either a not yet marked object or an uninitialized object 6888 if (p->klass_or_null() == NULL) { 6889 // An uninitialized object, skip to the next card, since 6890 // we may not be able to read its P-bits yet. 6891 assert(size == 0, "Initial value"); 6892 } else { 6893 // An object not (yet) reached by marking: we merely need to 6894 // compute its size so as to go look at the next block. 6895 assert(p->is_oop(true), "should be an oop"); 6896 size = CompactibleFreeListSpace::adjustObjectSize(p->size()); 6897 } 6898 } 6899 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6900 return size; 6901 } 6902 6903 void ScanMarkedObjectsAgainCarefullyClosure::do_yield_work() { 6904 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6905 "CMS thread should hold CMS token"); 6906 assert_lock_strong(_freelistLock); 6907 assert_lock_strong(_bitMap->lock()); 6908 // relinquish the free_list_lock and bitMaplock() 6909 _bitMap->lock()->unlock(); 6910 _freelistLock->unlock(); 6911 ConcurrentMarkSweepThread::desynchronize(true); 6912 ConcurrentMarkSweepThread::acknowledge_yield_request(); 6913 _collector->stopTimer(); 6914 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 6915 if (PrintCMSStatistics != 0) { 6916 _collector->incrementYields(); 6917 } 6918 _collector->icms_wait(); 6919 6920 // See the comment in coordinator_yield() 6921 for (unsigned i = 0; i < CMSYieldSleepCount && 6922 ConcurrentMarkSweepThread::should_yield() && 6923 !CMSCollector::foregroundGCIsActive(); ++i) { 6924 os::sleep(Thread::current(), 1, false); 6925 ConcurrentMarkSweepThread::acknowledge_yield_request(); 6926 } 6927 6928 ConcurrentMarkSweepThread::synchronize(true); 6929 _freelistLock->lock_without_safepoint_check(); 6930 _bitMap->lock()->lock_without_safepoint_check(); 6931 _collector->startTimer(); 6932 } 6933 6934 6935 ////////////////////////////////////////////////////////////////// 6936 // SurvivorSpacePrecleanClosure 6937 ////////////////////////////////////////////////////////////////// 6938 // This (single-threaded) closure is used to preclean the oops in 6939 // the survivor spaces. 6940 size_t SurvivorSpacePrecleanClosure::do_object_careful(oop p) { 6941 6942 HeapWord* addr = (HeapWord*)p; 6943 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6944 assert(!_span.contains(addr), "we are scanning the survivor spaces"); 6945 assert(p->klass_or_null() != NULL, "object should be initializd"); 6946 // an initialized object; ignore mark word in verification below 6947 // since we are running concurrent with mutators 6948 assert(p->is_oop(true), "should be an oop"); 6949 // Note that we do not yield while we iterate over 6950 // the interior oops of p, pushing the relevant ones 6951 // on our marking stack. 6952 size_t size = p->oop_iterate(_scanning_closure); 6953 do_yield_check(); 6954 // Observe that below, we do not abandon the preclean 6955 // phase as soon as we should; rather we empty the 6956 // marking stack before returning. This is to satisfy 6957 // some existing assertions. In general, it may be a 6958 // good idea to abort immediately and complete the marking 6959 // from the grey objects at a later time. 6960 while (!_mark_stack->isEmpty()) { 6961 oop new_oop = _mark_stack->pop(); 6962 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 6963 assert(_bit_map->isMarked((HeapWord*)new_oop), 6964 "only grey objects on this stack"); 6965 // iterate over the oops in this oop, marking and pushing 6966 // the ones in CMS heap (i.e. in _span). 6967 new_oop->oop_iterate(_scanning_closure); 6968 // check if it's time to yield 6969 do_yield_check(); 6970 } 6971 unsigned int after_count = 6972 GenCollectedHeap::heap()->total_collections(); 6973 bool abort = (_before_count != after_count) || 6974 _collector->should_abort_preclean(); 6975 return abort ? 0 : size; 6976 } 6977 6978 void SurvivorSpacePrecleanClosure::do_yield_work() { 6979 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6980 "CMS thread should hold CMS token"); 6981 assert_lock_strong(_bit_map->lock()); 6982 // Relinquish the bit map lock 6983 _bit_map->lock()->unlock(); 6984 ConcurrentMarkSweepThread::desynchronize(true); 6985 ConcurrentMarkSweepThread::acknowledge_yield_request(); 6986 _collector->stopTimer(); 6987 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 6988 if (PrintCMSStatistics != 0) { 6989 _collector->incrementYields(); 6990 } 6991 _collector->icms_wait(); 6992 6993 // See the comment in coordinator_yield() 6994 for (unsigned i = 0; i < CMSYieldSleepCount && 6995 ConcurrentMarkSweepThread::should_yield() && 6996 !CMSCollector::foregroundGCIsActive(); ++i) { 6997 os::sleep(Thread::current(), 1, false); 6998 ConcurrentMarkSweepThread::acknowledge_yield_request(); 6999 } 7000 7001 ConcurrentMarkSweepThread::synchronize(true); 7002 _bit_map->lock()->lock_without_safepoint_check(); 7003 _collector->startTimer(); 7004 } 7005 7006 // This closure is used to rescan the marked objects on the dirty cards 7007 // in the mod union table and the card table proper. In the parallel 7008 // case, although the bitMap is shared, we do a single read so the 7009 // isMarked() query is "safe". 7010 bool ScanMarkedObjectsAgainClosure::do_object_bm(oop p, MemRegion mr) { 7011 // Ignore mark word because we are running concurrent with mutators 7012 assert(p->is_oop_or_null(true), "expected an oop or null"); 7013 HeapWord* addr = (HeapWord*)p; 7014 assert(_span.contains(addr), "we are scanning the CMS generation"); 7015 bool is_obj_array = false; 7016 #ifdef DEBUG 7017 if (!_parallel) { 7018 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 7019 assert(_collector->overflow_list_is_empty(), 7020 "overflow list should be empty"); 7021 7022 } 7023 #endif // DEBUG 7024 if (_bit_map->isMarked(addr)) { 7025 // Obj arrays are precisely marked, non-arrays are not; 7026 // so we scan objArrays precisely and non-arrays in their 7027 // entirety. 7028 if (p->is_objArray()) { 7029 is_obj_array = true; 7030 if (_parallel) { 7031 p->oop_iterate(_par_scan_closure, mr); 7032 } else { 7033 p->oop_iterate(_scan_closure, mr); 7034 } 7035 } else { 7036 if (_parallel) { 7037 p->oop_iterate(_par_scan_closure); 7038 } else { 7039 p->oop_iterate(_scan_closure); 7040 } 7041 } 7042 } 7043 #ifdef DEBUG 7044 if (!_parallel) { 7045 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 7046 assert(_collector->overflow_list_is_empty(), 7047 "overflow list should be empty"); 7048 7049 } 7050 #endif // DEBUG 7051 return is_obj_array; 7052 } 7053 7054 MarkFromRootsClosure::MarkFromRootsClosure(CMSCollector* collector, 7055 MemRegion span, 7056 CMSBitMap* bitMap, CMSMarkStack* markStack, 7057 bool should_yield, bool verifying): 7058 _collector(collector), 7059 _span(span), 7060 _bitMap(bitMap), 7061 _mut(&collector->_modUnionTable), 7062 _markStack(markStack), 7063 _yield(should_yield), 7064 _skipBits(0) 7065 { 7066 assert(_markStack->isEmpty(), "stack should be empty"); 7067 _finger = _bitMap->startWord(); 7068 _threshold = _finger; 7069 assert(_collector->_restart_addr == NULL, "Sanity check"); 7070 assert(_span.contains(_finger), "Out of bounds _finger?"); 7071 DEBUG_ONLY(_verifying = verifying;) 7072 } 7073 7074 void MarkFromRootsClosure::reset(HeapWord* addr) { 7075 assert(_markStack->isEmpty(), "would cause duplicates on stack"); 7076 assert(_span.contains(addr), "Out of bounds _finger?"); 7077 _finger = addr; 7078 _threshold = (HeapWord*)round_to( 7079 (intptr_t)_finger, CardTableModRefBS::card_size); 7080 } 7081 7082 // Should revisit to see if this should be restructured for 7083 // greater efficiency. 7084 bool MarkFromRootsClosure::do_bit(size_t offset) { 7085 if (_skipBits > 0) { 7086 _skipBits--; 7087 return true; 7088 } 7089 // convert offset into a HeapWord* 7090 HeapWord* addr = _bitMap->startWord() + offset; 7091 assert(_bitMap->endWord() && addr < _bitMap->endWord(), 7092 "address out of range"); 7093 assert(_bitMap->isMarked(addr), "tautology"); 7094 if (_bitMap->isMarked(addr+1)) { 7095 // this is an allocated but not yet initialized object 7096 assert(_skipBits == 0, "tautology"); 7097 _skipBits = 2; // skip next two marked bits ("Printezis-marks") 7098 oop p = oop(addr); 7099 if (p->klass_or_null() == NULL) { 7100 DEBUG_ONLY(if (!_verifying) {) 7101 // We re-dirty the cards on which this object lies and increase 7102 // the _threshold so that we'll come back to scan this object 7103 // during the preclean or remark phase. (CMSCleanOnEnter) 7104 if (CMSCleanOnEnter) { 7105 size_t sz = _collector->block_size_using_printezis_bits(addr); 7106 HeapWord* end_card_addr = (HeapWord*)round_to( 7107 (intptr_t)(addr+sz), CardTableModRefBS::card_size); 7108 MemRegion redirty_range = MemRegion(addr, end_card_addr); 7109 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 7110 // Bump _threshold to end_card_addr; note that 7111 // _threshold cannot possibly exceed end_card_addr, anyhow. 7112 // This prevents future clearing of the card as the scan proceeds 7113 // to the right. 7114 assert(_threshold <= end_card_addr, 7115 "Because we are just scanning into this object"); 7116 if (_threshold < end_card_addr) { 7117 _threshold = end_card_addr; 7118 } 7119 if (p->klass_or_null() != NULL) { 7120 // Redirty the range of cards... 7121 _mut->mark_range(redirty_range); 7122 } // ...else the setting of klass will dirty the card anyway. 7123 } 7124 DEBUG_ONLY(}) 7125 return true; 7126 } 7127 } 7128 scanOopsInOop(addr); 7129 return true; 7130 } 7131 7132 // We take a break if we've been at this for a while, 7133 // so as to avoid monopolizing the locks involved. 7134 void MarkFromRootsClosure::do_yield_work() { 7135 // First give up the locks, then yield, then re-lock 7136 // We should probably use a constructor/destructor idiom to 7137 // do this unlock/lock or modify the MutexUnlocker class to 7138 // serve our purpose. XXX 7139 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 7140 "CMS thread should hold CMS token"); 7141 assert_lock_strong(_bitMap->lock()); 7142 _bitMap->lock()->unlock(); 7143 ConcurrentMarkSweepThread::desynchronize(true); 7144 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7145 _collector->stopTimer(); 7146 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 7147 if (PrintCMSStatistics != 0) { 7148 _collector->incrementYields(); 7149 } 7150 _collector->icms_wait(); 7151 7152 // See the comment in coordinator_yield() 7153 for (unsigned i = 0; i < CMSYieldSleepCount && 7154 ConcurrentMarkSweepThread::should_yield() && 7155 !CMSCollector::foregroundGCIsActive(); ++i) { 7156 os::sleep(Thread::current(), 1, false); 7157 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7158 } 7159 7160 ConcurrentMarkSweepThread::synchronize(true); 7161 _bitMap->lock()->lock_without_safepoint_check(); 7162 _collector->startTimer(); 7163 } 7164 7165 void MarkFromRootsClosure::scanOopsInOop(HeapWord* ptr) { 7166 assert(_bitMap->isMarked(ptr), "expected bit to be set"); 7167 assert(_markStack->isEmpty(), 7168 "should drain stack to limit stack usage"); 7169 // convert ptr to an oop preparatory to scanning 7170 oop obj = oop(ptr); 7171 // Ignore mark word in verification below, since we 7172 // may be running concurrent with mutators. 7173 assert(obj->is_oop(true), "should be an oop"); 7174 assert(_finger <= ptr, "_finger runneth ahead"); 7175 // advance the finger to right end of this object 7176 _finger = ptr + obj->size(); 7177 assert(_finger > ptr, "we just incremented it above"); 7178 // On large heaps, it may take us some time to get through 7179 // the marking phase (especially if running iCMS). During 7180 // this time it's possible that a lot of mutations have 7181 // accumulated in the card table and the mod union table -- 7182 // these mutation records are redundant until we have 7183 // actually traced into the corresponding card. 7184 // Here, we check whether advancing the finger would make 7185 // us cross into a new card, and if so clear corresponding 7186 // cards in the MUT (preclean them in the card-table in the 7187 // future). 7188 7189 DEBUG_ONLY(if (!_verifying) {) 7190 // The clean-on-enter optimization is disabled by default, 7191 // until we fix 6178663. 7192 if (CMSCleanOnEnter && (_finger > _threshold)) { 7193 // [_threshold, _finger) represents the interval 7194 // of cards to be cleared in MUT (or precleaned in card table). 7195 // The set of cards to be cleared is all those that overlap 7196 // with the interval [_threshold, _finger); note that 7197 // _threshold is always kept card-aligned but _finger isn't 7198 // always card-aligned. 7199 HeapWord* old_threshold = _threshold; 7200 assert(old_threshold == (HeapWord*)round_to( 7201 (intptr_t)old_threshold, CardTableModRefBS::card_size), 7202 "_threshold should always be card-aligned"); 7203 _threshold = (HeapWord*)round_to( 7204 (intptr_t)_finger, CardTableModRefBS::card_size); 7205 MemRegion mr(old_threshold, _threshold); 7206 assert(!mr.is_empty(), "Control point invariant"); 7207 assert(_span.contains(mr), "Should clear within span"); 7208 _mut->clear_range(mr); 7209 } 7210 DEBUG_ONLY(}) 7211 // Note: the finger doesn't advance while we drain 7212 // the stack below. 7213 PushOrMarkClosure pushOrMarkClosure(_collector, 7214 _span, _bitMap, _markStack, 7215 _finger, this); 7216 bool res = _markStack->push(obj); 7217 assert(res, "Empty non-zero size stack should have space for single push"); 7218 while (!_markStack->isEmpty()) { 7219 oop new_oop = _markStack->pop(); 7220 // Skip verifying header mark word below because we are 7221 // running concurrent with mutators. 7222 assert(new_oop->is_oop(true), "Oops! expected to pop an oop"); 7223 // now scan this oop's oops 7224 new_oop->oop_iterate(&pushOrMarkClosure); 7225 do_yield_check(); 7226 } 7227 assert(_markStack->isEmpty(), "tautology, emphasizing post-condition"); 7228 } 7229 7230 Par_MarkFromRootsClosure::Par_MarkFromRootsClosure(CMSConcMarkingTask* task, 7231 CMSCollector* collector, MemRegion span, 7232 CMSBitMap* bit_map, 7233 OopTaskQueue* work_queue, 7234 CMSMarkStack* overflow_stack, 7235 bool should_yield): 7236 _collector(collector), 7237 _whole_span(collector->_span), 7238 _span(span), 7239 _bit_map(bit_map), 7240 _mut(&collector->_modUnionTable), 7241 _work_queue(work_queue), 7242 _overflow_stack(overflow_stack), 7243 _yield(should_yield), 7244 _skip_bits(0), 7245 _task(task) 7246 { 7247 assert(_work_queue->size() == 0, "work_queue should be empty"); 7248 _finger = span.start(); 7249 _threshold = _finger; // XXX Defer clear-on-enter optimization for now 7250 assert(_span.contains(_finger), "Out of bounds _finger?"); 7251 } 7252 7253 // Should revisit to see if this should be restructured for 7254 // greater efficiency. 7255 bool Par_MarkFromRootsClosure::do_bit(size_t offset) { 7256 if (_skip_bits > 0) { 7257 _skip_bits--; 7258 return true; 7259 } 7260 // convert offset into a HeapWord* 7261 HeapWord* addr = _bit_map->startWord() + offset; 7262 assert(_bit_map->endWord() && addr < _bit_map->endWord(), 7263 "address out of range"); 7264 assert(_bit_map->isMarked(addr), "tautology"); 7265 if (_bit_map->isMarked(addr+1)) { 7266 // this is an allocated object that might not yet be initialized 7267 assert(_skip_bits == 0, "tautology"); 7268 _skip_bits = 2; // skip next two marked bits ("Printezis-marks") 7269 oop p = oop(addr); 7270 if (p->klass_or_null() == NULL) { 7271 // in the case of Clean-on-Enter optimization, redirty card 7272 // and avoid clearing card by increasing the threshold. 7273 return true; 7274 } 7275 } 7276 scan_oops_in_oop(addr); 7277 return true; 7278 } 7279 7280 void Par_MarkFromRootsClosure::scan_oops_in_oop(HeapWord* ptr) { 7281 assert(_bit_map->isMarked(ptr), "expected bit to be set"); 7282 // Should we assert that our work queue is empty or 7283 // below some drain limit? 7284 assert(_work_queue->size() == 0, 7285 "should drain stack to limit stack usage"); 7286 // convert ptr to an oop preparatory to scanning 7287 oop obj = oop(ptr); 7288 // Ignore mark word in verification below, since we 7289 // may be running concurrent with mutators. 7290 assert(obj->is_oop(true), "should be an oop"); 7291 assert(_finger <= ptr, "_finger runneth ahead"); 7292 // advance the finger to right end of this object 7293 _finger = ptr + obj->size(); 7294 assert(_finger > ptr, "we just incremented it above"); 7295 // On large heaps, it may take us some time to get through 7296 // the marking phase (especially if running iCMS). During 7297 // this time it's possible that a lot of mutations have 7298 // accumulated in the card table and the mod union table -- 7299 // these mutation records are redundant until we have 7300 // actually traced into the corresponding card. 7301 // Here, we check whether advancing the finger would make 7302 // us cross into a new card, and if so clear corresponding 7303 // cards in the MUT (preclean them in the card-table in the 7304 // future). 7305 7306 // The clean-on-enter optimization is disabled by default, 7307 // until we fix 6178663. 7308 if (CMSCleanOnEnter && (_finger > _threshold)) { 7309 // [_threshold, _finger) represents the interval 7310 // of cards to be cleared in MUT (or precleaned in card table). 7311 // The set of cards to be cleared is all those that overlap 7312 // with the interval [_threshold, _finger); note that 7313 // _threshold is always kept card-aligned but _finger isn't 7314 // always card-aligned. 7315 HeapWord* old_threshold = _threshold; 7316 assert(old_threshold == (HeapWord*)round_to( 7317 (intptr_t)old_threshold, CardTableModRefBS::card_size), 7318 "_threshold should always be card-aligned"); 7319 _threshold = (HeapWord*)round_to( 7320 (intptr_t)_finger, CardTableModRefBS::card_size); 7321 MemRegion mr(old_threshold, _threshold); 7322 assert(!mr.is_empty(), "Control point invariant"); 7323 assert(_span.contains(mr), "Should clear within span"); // _whole_span ?? 7324 _mut->clear_range(mr); 7325 } 7326 7327 // Note: the local finger doesn't advance while we drain 7328 // the stack below, but the global finger sure can and will. 7329 HeapWord** gfa = _task->global_finger_addr(); 7330 Par_PushOrMarkClosure pushOrMarkClosure(_collector, 7331 _span, _bit_map, 7332 _work_queue, 7333 _overflow_stack, 7334 _finger, 7335 gfa, this); 7336 bool res = _work_queue->push(obj); // overflow could occur here 7337 assert(res, "Will hold once we use workqueues"); 7338 while (true) { 7339 oop new_oop; 7340 if (!_work_queue->pop_local(new_oop)) { 7341 // We emptied our work_queue; check if there's stuff that can 7342 // be gotten from the overflow stack. 7343 if (CMSConcMarkingTask::get_work_from_overflow_stack( 7344 _overflow_stack, _work_queue)) { 7345 do_yield_check(); 7346 continue; 7347 } else { // done 7348 break; 7349 } 7350 } 7351 // Skip verifying header mark word below because we are 7352 // running concurrent with mutators. 7353 assert(new_oop->is_oop(true), "Oops! expected to pop an oop"); 7354 // now scan this oop's oops 7355 new_oop->oop_iterate(&pushOrMarkClosure); 7356 do_yield_check(); 7357 } 7358 assert(_work_queue->size() == 0, "tautology, emphasizing post-condition"); 7359 } 7360 7361 // Yield in response to a request from VM Thread or 7362 // from mutators. 7363 void Par_MarkFromRootsClosure::do_yield_work() { 7364 assert(_task != NULL, "sanity"); 7365 _task->yield(); 7366 } 7367 7368 // A variant of the above used for verifying CMS marking work. 7369 MarkFromRootsVerifyClosure::MarkFromRootsVerifyClosure(CMSCollector* collector, 7370 MemRegion span, 7371 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 7372 CMSMarkStack* mark_stack): 7373 _collector(collector), 7374 _span(span), 7375 _verification_bm(verification_bm), 7376 _cms_bm(cms_bm), 7377 _mark_stack(mark_stack), 7378 _pam_verify_closure(collector, span, verification_bm, cms_bm, 7379 mark_stack) 7380 { 7381 assert(_mark_stack->isEmpty(), "stack should be empty"); 7382 _finger = _verification_bm->startWord(); 7383 assert(_collector->_restart_addr == NULL, "Sanity check"); 7384 assert(_span.contains(_finger), "Out of bounds _finger?"); 7385 } 7386 7387 void MarkFromRootsVerifyClosure::reset(HeapWord* addr) { 7388 assert(_mark_stack->isEmpty(), "would cause duplicates on stack"); 7389 assert(_span.contains(addr), "Out of bounds _finger?"); 7390 _finger = addr; 7391 } 7392 7393 // Should revisit to see if this should be restructured for 7394 // greater efficiency. 7395 bool MarkFromRootsVerifyClosure::do_bit(size_t offset) { 7396 // convert offset into a HeapWord* 7397 HeapWord* addr = _verification_bm->startWord() + offset; 7398 assert(_verification_bm->endWord() && addr < _verification_bm->endWord(), 7399 "address out of range"); 7400 assert(_verification_bm->isMarked(addr), "tautology"); 7401 assert(_cms_bm->isMarked(addr), "tautology"); 7402 7403 assert(_mark_stack->isEmpty(), 7404 "should drain stack to limit stack usage"); 7405 // convert addr to an oop preparatory to scanning 7406 oop obj = oop(addr); 7407 assert(obj->is_oop(), "should be an oop"); 7408 assert(_finger <= addr, "_finger runneth ahead"); 7409 // advance the finger to right end of this object 7410 _finger = addr + obj->size(); 7411 assert(_finger > addr, "we just incremented it above"); 7412 // Note: the finger doesn't advance while we drain 7413 // the stack below. 7414 bool res = _mark_stack->push(obj); 7415 assert(res, "Empty non-zero size stack should have space for single push"); 7416 while (!_mark_stack->isEmpty()) { 7417 oop new_oop = _mark_stack->pop(); 7418 assert(new_oop->is_oop(), "Oops! expected to pop an oop"); 7419 // now scan this oop's oops 7420 new_oop->oop_iterate(&_pam_verify_closure); 7421 } 7422 assert(_mark_stack->isEmpty(), "tautology, emphasizing post-condition"); 7423 return true; 7424 } 7425 7426 PushAndMarkVerifyClosure::PushAndMarkVerifyClosure( 7427 CMSCollector* collector, MemRegion span, 7428 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 7429 CMSMarkStack* mark_stack): 7430 CMSOopClosure(collector->ref_processor()), 7431 _collector(collector), 7432 _span(span), 7433 _verification_bm(verification_bm), 7434 _cms_bm(cms_bm), 7435 _mark_stack(mark_stack) 7436 { } 7437 7438 void PushAndMarkVerifyClosure::do_oop(oop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 7439 void PushAndMarkVerifyClosure::do_oop(narrowOop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 7440 7441 // Upon stack overflow, we discard (part of) the stack, 7442 // remembering the least address amongst those discarded 7443 // in CMSCollector's _restart_address. 7444 void PushAndMarkVerifyClosure::handle_stack_overflow(HeapWord* lost) { 7445 // Remember the least grey address discarded 7446 HeapWord* ra = (HeapWord*)_mark_stack->least_value(lost); 7447 _collector->lower_restart_addr(ra); 7448 _mark_stack->reset(); // discard stack contents 7449 _mark_stack->expand(); // expand the stack if possible 7450 } 7451 7452 void PushAndMarkVerifyClosure::do_oop(oop obj) { 7453 assert(obj->is_oop_or_null(), "expected an oop or NULL"); 7454 HeapWord* addr = (HeapWord*)obj; 7455 if (_span.contains(addr) && !_verification_bm->isMarked(addr)) { 7456 // Oop lies in _span and isn't yet grey or black 7457 _verification_bm->mark(addr); // now grey 7458 if (!_cms_bm->isMarked(addr)) { 7459 oop(addr)->print(); 7460 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)", 7461 addr); 7462 fatal("... aborting"); 7463 } 7464 7465 if (!_mark_stack->push(obj)) { // stack overflow 7466 if (PrintCMSStatistics != 0) { 7467 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 7468 SIZE_FORMAT, _mark_stack->capacity()); 7469 } 7470 assert(_mark_stack->isFull(), "Else push should have succeeded"); 7471 handle_stack_overflow(addr); 7472 } 7473 // anything including and to the right of _finger 7474 // will be scanned as we iterate over the remainder of the 7475 // bit map 7476 } 7477 } 7478 7479 PushOrMarkClosure::PushOrMarkClosure(CMSCollector* collector, 7480 MemRegion span, 7481 CMSBitMap* bitMap, CMSMarkStack* markStack, 7482 HeapWord* finger, MarkFromRootsClosure* parent) : 7483 CMSOopClosure(collector->ref_processor()), 7484 _collector(collector), 7485 _span(span), 7486 _bitMap(bitMap), 7487 _markStack(markStack), 7488 _finger(finger), 7489 _parent(parent) 7490 { } 7491 7492 Par_PushOrMarkClosure::Par_PushOrMarkClosure(CMSCollector* collector, 7493 MemRegion span, 7494 CMSBitMap* bit_map, 7495 OopTaskQueue* work_queue, 7496 CMSMarkStack* overflow_stack, 7497 HeapWord* finger, 7498 HeapWord** global_finger_addr, 7499 Par_MarkFromRootsClosure* parent) : 7500 CMSOopClosure(collector->ref_processor()), 7501 _collector(collector), 7502 _whole_span(collector->_span), 7503 _span(span), 7504 _bit_map(bit_map), 7505 _work_queue(work_queue), 7506 _overflow_stack(overflow_stack), 7507 _finger(finger), 7508 _global_finger_addr(global_finger_addr), 7509 _parent(parent) 7510 { } 7511 7512 // Assumes thread-safe access by callers, who are 7513 // responsible for mutual exclusion. 7514 void CMSCollector::lower_restart_addr(HeapWord* low) { 7515 assert(_span.contains(low), "Out of bounds addr"); 7516 if (_restart_addr == NULL) { 7517 _restart_addr = low; 7518 } else { 7519 _restart_addr = MIN2(_restart_addr, low); 7520 } 7521 } 7522 7523 // Upon stack overflow, we discard (part of) the stack, 7524 // remembering the least address amongst those discarded 7525 // in CMSCollector's _restart_address. 7526 void PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 7527 // Remember the least grey address discarded 7528 HeapWord* ra = (HeapWord*)_markStack->least_value(lost); 7529 _collector->lower_restart_addr(ra); 7530 _markStack->reset(); // discard stack contents 7531 _markStack->expand(); // expand the stack if possible 7532 } 7533 7534 // Upon stack overflow, we discard (part of) the stack, 7535 // remembering the least address amongst those discarded 7536 // in CMSCollector's _restart_address. 7537 void Par_PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 7538 // We need to do this under a mutex to prevent other 7539 // workers from interfering with the work done below. 7540 MutexLockerEx ml(_overflow_stack->par_lock(), 7541 Mutex::_no_safepoint_check_flag); 7542 // Remember the least grey address discarded 7543 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 7544 _collector->lower_restart_addr(ra); 7545 _overflow_stack->reset(); // discard stack contents 7546 _overflow_stack->expand(); // expand the stack if possible 7547 } 7548 7549 void CMKlassClosure::do_klass(Klass* k) { 7550 assert(_oop_closure != NULL, "Not initialized?"); 7551 k->oops_do(_oop_closure); 7552 } 7553 7554 void PushOrMarkClosure::do_oop(oop obj) { 7555 // Ignore mark word because we are running concurrent with mutators. 7556 assert(obj->is_oop_or_null(true), "expected an oop or NULL"); 7557 HeapWord* addr = (HeapWord*)obj; 7558 if (_span.contains(addr) && !_bitMap->isMarked(addr)) { 7559 // Oop lies in _span and isn't yet grey or black 7560 _bitMap->mark(addr); // now grey 7561 if (addr < _finger) { 7562 // the bit map iteration has already either passed, or 7563 // sampled, this bit in the bit map; we'll need to 7564 // use the marking stack to scan this oop's oops. 7565 bool simulate_overflow = false; 7566 NOT_PRODUCT( 7567 if (CMSMarkStackOverflowALot && 7568 _collector->simulate_overflow()) { 7569 // simulate a stack overflow 7570 simulate_overflow = true; 7571 } 7572 ) 7573 if (simulate_overflow || !_markStack->push(obj)) { // stack overflow 7574 if (PrintCMSStatistics != 0) { 7575 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 7576 SIZE_FORMAT, _markStack->capacity()); 7577 } 7578 assert(simulate_overflow || _markStack->isFull(), "Else push should have succeeded"); 7579 handle_stack_overflow(addr); 7580 } 7581 } 7582 // anything including and to the right of _finger 7583 // will be scanned as we iterate over the remainder of the 7584 // bit map 7585 do_yield_check(); 7586 } 7587 } 7588 7589 void PushOrMarkClosure::do_oop(oop* p) { PushOrMarkClosure::do_oop_work(p); } 7590 void PushOrMarkClosure::do_oop(narrowOop* p) { PushOrMarkClosure::do_oop_work(p); } 7591 7592 void Par_PushOrMarkClosure::do_oop(oop obj) { 7593 // Ignore mark word because we are running concurrent with mutators. 7594 assert(obj->is_oop_or_null(true), "expected an oop or NULL"); 7595 HeapWord* addr = (HeapWord*)obj; 7596 if (_whole_span.contains(addr) && !_bit_map->isMarked(addr)) { 7597 // Oop lies in _span and isn't yet grey or black 7598 // We read the global_finger (volatile read) strictly after marking oop 7599 bool res = _bit_map->par_mark(addr); // now grey 7600 volatile HeapWord** gfa = (volatile HeapWord**)_global_finger_addr; 7601 // Should we push this marked oop on our stack? 7602 // -- if someone else marked it, nothing to do 7603 // -- if target oop is above global finger nothing to do 7604 // -- if target oop is in chunk and above local finger 7605 // then nothing to do 7606 // -- else push on work queue 7607 if ( !res // someone else marked it, they will deal with it 7608 || (addr >= *gfa) // will be scanned in a later task 7609 || (_span.contains(addr) && addr >= _finger)) { // later in this chunk 7610 return; 7611 } 7612 // the bit map iteration has already either passed, or 7613 // sampled, this bit in the bit map; we'll need to 7614 // use the marking stack to scan this oop's oops. 7615 bool simulate_overflow = false; 7616 NOT_PRODUCT( 7617 if (CMSMarkStackOverflowALot && 7618 _collector->simulate_overflow()) { 7619 // simulate a stack overflow 7620 simulate_overflow = true; 7621 } 7622 ) 7623 if (simulate_overflow || 7624 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) { 7625 // stack overflow 7626 if (PrintCMSStatistics != 0) { 7627 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 7628 SIZE_FORMAT, _overflow_stack->capacity()); 7629 } 7630 // We cannot assert that the overflow stack is full because 7631 // it may have been emptied since. 7632 assert(simulate_overflow || 7633 _work_queue->size() == _work_queue->max_elems(), 7634 "Else push should have succeeded"); 7635 handle_stack_overflow(addr); 7636 } 7637 do_yield_check(); 7638 } 7639 } 7640 7641 void Par_PushOrMarkClosure::do_oop(oop* p) { Par_PushOrMarkClosure::do_oop_work(p); } 7642 void Par_PushOrMarkClosure::do_oop(narrowOop* p) { Par_PushOrMarkClosure::do_oop_work(p); } 7643 7644 PushAndMarkClosure::PushAndMarkClosure(CMSCollector* collector, 7645 MemRegion span, 7646 ReferenceProcessor* rp, 7647 CMSBitMap* bit_map, 7648 CMSBitMap* mod_union_table, 7649 CMSMarkStack* mark_stack, 7650 bool concurrent_precleaning): 7651 CMSOopClosure(rp), 7652 _collector(collector), 7653 _span(span), 7654 _bit_map(bit_map), 7655 _mod_union_table(mod_union_table), 7656 _mark_stack(mark_stack), 7657 _concurrent_precleaning(concurrent_precleaning) 7658 { 7659 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 7660 } 7661 7662 // Grey object rescan during pre-cleaning and second checkpoint phases -- 7663 // the non-parallel version (the parallel version appears further below.) 7664 void PushAndMarkClosure::do_oop(oop obj) { 7665 // Ignore mark word verification. If during concurrent precleaning, 7666 // the object monitor may be locked. If during the checkpoint 7667 // phases, the object may already have been reached by a different 7668 // path and may be at the end of the global overflow list (so 7669 // the mark word may be NULL). 7670 assert(obj->is_oop_or_null(true /* ignore mark word */), 7671 "expected an oop or NULL"); 7672 HeapWord* addr = (HeapWord*)obj; 7673 // Check if oop points into the CMS generation 7674 // and is not marked 7675 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 7676 // a white object ... 7677 _bit_map->mark(addr); // ... now grey 7678 // push on the marking stack (grey set) 7679 bool simulate_overflow = false; 7680 NOT_PRODUCT( 7681 if (CMSMarkStackOverflowALot && 7682 _collector->simulate_overflow()) { 7683 // simulate a stack overflow 7684 simulate_overflow = true; 7685 } 7686 ) 7687 if (simulate_overflow || !_mark_stack->push(obj)) { 7688 if (_concurrent_precleaning) { 7689 // During precleaning we can just dirty the appropriate card(s) 7690 // in the mod union table, thus ensuring that the object remains 7691 // in the grey set and continue. In the case of object arrays 7692 // we need to dirty all of the cards that the object spans, 7693 // since the rescan of object arrays will be limited to the 7694 // dirty cards. 7695 // Note that no one can be intefering with us in this action 7696 // of dirtying the mod union table, so no locking or atomics 7697 // are required. 7698 if (obj->is_objArray()) { 7699 size_t sz = obj->size(); 7700 HeapWord* end_card_addr = (HeapWord*)round_to( 7701 (intptr_t)(addr+sz), CardTableModRefBS::card_size); 7702 MemRegion redirty_range = MemRegion(addr, end_card_addr); 7703 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 7704 _mod_union_table->mark_range(redirty_range); 7705 } else { 7706 _mod_union_table->mark(addr); 7707 } 7708 _collector->_ser_pmc_preclean_ovflw++; 7709 } else { 7710 // During the remark phase, we need to remember this oop 7711 // in the overflow list. 7712 _collector->push_on_overflow_list(obj); 7713 _collector->_ser_pmc_remark_ovflw++; 7714 } 7715 } 7716 } 7717 } 7718 7719 Par_PushAndMarkClosure::Par_PushAndMarkClosure(CMSCollector* collector, 7720 MemRegion span, 7721 ReferenceProcessor* rp, 7722 CMSBitMap* bit_map, 7723 OopTaskQueue* work_queue): 7724 CMSOopClosure(rp), 7725 _collector(collector), 7726 _span(span), 7727 _bit_map(bit_map), 7728 _work_queue(work_queue) 7729 { 7730 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 7731 } 7732 7733 void PushAndMarkClosure::do_oop(oop* p) { PushAndMarkClosure::do_oop_work(p); } 7734 void PushAndMarkClosure::do_oop(narrowOop* p) { PushAndMarkClosure::do_oop_work(p); } 7735 7736 // Grey object rescan during second checkpoint phase -- 7737 // the parallel version. 7738 void Par_PushAndMarkClosure::do_oop(oop obj) { 7739 // In the assert below, we ignore the mark word because 7740 // this oop may point to an already visited object that is 7741 // on the overflow stack (in which case the mark word has 7742 // been hijacked for chaining into the overflow stack -- 7743 // if this is the last object in the overflow stack then 7744 // its mark word will be NULL). Because this object may 7745 // have been subsequently popped off the global overflow 7746 // stack, and the mark word possibly restored to the prototypical 7747 // value, by the time we get to examined this failing assert in 7748 // the debugger, is_oop_or_null(false) may subsequently start 7749 // to hold. 7750 assert(obj->is_oop_or_null(true), 7751 "expected an oop or NULL"); 7752 HeapWord* addr = (HeapWord*)obj; 7753 // Check if oop points into the CMS generation 7754 // and is not marked 7755 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 7756 // a white object ... 7757 // If we manage to "claim" the object, by being the 7758 // first thread to mark it, then we push it on our 7759 // marking stack 7760 if (_bit_map->par_mark(addr)) { // ... now grey 7761 // push on work queue (grey set) 7762 bool simulate_overflow = false; 7763 NOT_PRODUCT( 7764 if (CMSMarkStackOverflowALot && 7765 _collector->par_simulate_overflow()) { 7766 // simulate a stack overflow 7767 simulate_overflow = true; 7768 } 7769 ) 7770 if (simulate_overflow || !_work_queue->push(obj)) { 7771 _collector->par_push_on_overflow_list(obj); 7772 _collector->_par_pmc_remark_ovflw++; // imprecise OK: no need to CAS 7773 } 7774 } // Else, some other thread got there first 7775 } 7776 } 7777 7778 void Par_PushAndMarkClosure::do_oop(oop* p) { Par_PushAndMarkClosure::do_oop_work(p); } 7779 void Par_PushAndMarkClosure::do_oop(narrowOop* p) { Par_PushAndMarkClosure::do_oop_work(p); } 7780 7781 void CMSPrecleanRefsYieldClosure::do_yield_work() { 7782 Mutex* bml = _collector->bitMapLock(); 7783 assert_lock_strong(bml); 7784 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 7785 "CMS thread should hold CMS token"); 7786 7787 bml->unlock(); 7788 ConcurrentMarkSweepThread::desynchronize(true); 7789 7790 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7791 7792 _collector->stopTimer(); 7793 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 7794 if (PrintCMSStatistics != 0) { 7795 _collector->incrementYields(); 7796 } 7797 _collector->icms_wait(); 7798 7799 // See the comment in coordinator_yield() 7800 for (unsigned i = 0; i < CMSYieldSleepCount && 7801 ConcurrentMarkSweepThread::should_yield() && 7802 !CMSCollector::foregroundGCIsActive(); ++i) { 7803 os::sleep(Thread::current(), 1, false); 7804 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7805 } 7806 7807 ConcurrentMarkSweepThread::synchronize(true); 7808 bml->lock(); 7809 7810 _collector->startTimer(); 7811 } 7812 7813 bool CMSPrecleanRefsYieldClosure::should_return() { 7814 if (ConcurrentMarkSweepThread::should_yield()) { 7815 do_yield_work(); 7816 } 7817 return _collector->foregroundGCIsActive(); 7818 } 7819 7820 void MarkFromDirtyCardsClosure::do_MemRegion(MemRegion mr) { 7821 assert(((size_t)mr.start())%CardTableModRefBS::card_size_in_words == 0, 7822 "mr should be aligned to start at a card boundary"); 7823 // We'd like to assert: 7824 // assert(mr.word_size()%CardTableModRefBS::card_size_in_words == 0, 7825 // "mr should be a range of cards"); 7826 // However, that would be too strong in one case -- the last 7827 // partition ends at _unallocated_block which, in general, can be 7828 // an arbitrary boundary, not necessarily card aligned. 7829 if (PrintCMSStatistics != 0) { 7830 _num_dirty_cards += 7831 mr.word_size()/CardTableModRefBS::card_size_in_words; 7832 } 7833 _space->object_iterate_mem(mr, &_scan_cl); 7834 } 7835 7836 SweepClosure::SweepClosure(CMSCollector* collector, 7837 ConcurrentMarkSweepGeneration* g, 7838 CMSBitMap* bitMap, bool should_yield) : 7839 _collector(collector), 7840 _g(g), 7841 _sp(g->cmsSpace()), 7842 _limit(_sp->sweep_limit()), 7843 _freelistLock(_sp->freelistLock()), 7844 _bitMap(bitMap), 7845 _yield(should_yield), 7846 _inFreeRange(false), // No free range at beginning of sweep 7847 _freeRangeInFreeLists(false), // No free range at beginning of sweep 7848 _lastFreeRangeCoalesced(false), 7849 _freeFinger(g->used_region().start()) 7850 { 7851 NOT_PRODUCT( 7852 _numObjectsFreed = 0; 7853 _numWordsFreed = 0; 7854 _numObjectsLive = 0; 7855 _numWordsLive = 0; 7856 _numObjectsAlreadyFree = 0; 7857 _numWordsAlreadyFree = 0; 7858 _last_fc = NULL; 7859 7860 _sp->initializeIndexedFreeListArrayReturnedBytes(); 7861 _sp->dictionary()->initialize_dict_returned_bytes(); 7862 ) 7863 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7864 "sweep _limit out of bounds"); 7865 if (CMSTraceSweeper) { 7866 gclog_or_tty->print_cr("\n====================\nStarting new sweep with limit " PTR_FORMAT, 7867 _limit); 7868 } 7869 } 7870 7871 void SweepClosure::print_on(outputStream* st) const { 7872 tty->print_cr("_sp = [" PTR_FORMAT "," PTR_FORMAT ")", 7873 _sp->bottom(), _sp->end()); 7874 tty->print_cr("_limit = " PTR_FORMAT, _limit); 7875 tty->print_cr("_freeFinger = " PTR_FORMAT, _freeFinger); 7876 NOT_PRODUCT(tty->print_cr("_last_fc = " PTR_FORMAT, _last_fc);) 7877 tty->print_cr("_inFreeRange = %d, _freeRangeInFreeLists = %d, _lastFreeRangeCoalesced = %d", 7878 _inFreeRange, _freeRangeInFreeLists, _lastFreeRangeCoalesced); 7879 } 7880 7881 #ifndef PRODUCT 7882 // Assertion checking only: no useful work in product mode -- 7883 // however, if any of the flags below become product flags, 7884 // you may need to review this code to see if it needs to be 7885 // enabled in product mode. 7886 SweepClosure::~SweepClosure() { 7887 assert_lock_strong(_freelistLock); 7888 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7889 "sweep _limit out of bounds"); 7890 if (inFreeRange()) { 7891 warning("inFreeRange() should have been reset; dumping state of SweepClosure"); 7892 print(); 7893 ShouldNotReachHere(); 7894 } 7895 if (Verbose && PrintGC) { 7896 gclog_or_tty->print("Collected "SIZE_FORMAT" objects, " SIZE_FORMAT " bytes", 7897 _numObjectsFreed, _numWordsFreed*sizeof(HeapWord)); 7898 gclog_or_tty->print_cr("\nLive "SIZE_FORMAT" objects, " 7899 SIZE_FORMAT" bytes " 7900 "Already free "SIZE_FORMAT" objects, "SIZE_FORMAT" bytes", 7901 _numObjectsLive, _numWordsLive*sizeof(HeapWord), 7902 _numObjectsAlreadyFree, _numWordsAlreadyFree*sizeof(HeapWord)); 7903 size_t totalBytes = (_numWordsFreed + _numWordsLive + _numWordsAlreadyFree) 7904 * sizeof(HeapWord); 7905 gclog_or_tty->print_cr("Total sweep: "SIZE_FORMAT" bytes", totalBytes); 7906 7907 if (PrintCMSStatistics && CMSVerifyReturnedBytes) { 7908 size_t indexListReturnedBytes = _sp->sumIndexedFreeListArrayReturnedBytes(); 7909 size_t dict_returned_bytes = _sp->dictionary()->sum_dict_returned_bytes(); 7910 size_t returned_bytes = indexListReturnedBytes + dict_returned_bytes; 7911 gclog_or_tty->print("Returned "SIZE_FORMAT" bytes", returned_bytes); 7912 gclog_or_tty->print(" Indexed List Returned "SIZE_FORMAT" bytes", 7913 indexListReturnedBytes); 7914 gclog_or_tty->print_cr(" Dictionary Returned "SIZE_FORMAT" bytes", 7915 dict_returned_bytes); 7916 } 7917 } 7918 if (CMSTraceSweeper) { 7919 gclog_or_tty->print_cr("end of sweep with _limit = " PTR_FORMAT "\n================", 7920 _limit); 7921 } 7922 } 7923 #endif // PRODUCT 7924 7925 void SweepClosure::initialize_free_range(HeapWord* freeFinger, 7926 bool freeRangeInFreeLists) { 7927 if (CMSTraceSweeper) { 7928 gclog_or_tty->print("---- Start free range at 0x%x with free block (%d)\n", 7929 freeFinger, freeRangeInFreeLists); 7930 } 7931 assert(!inFreeRange(), "Trampling existing free range"); 7932 set_inFreeRange(true); 7933 set_lastFreeRangeCoalesced(false); 7934 7935 set_freeFinger(freeFinger); 7936 set_freeRangeInFreeLists(freeRangeInFreeLists); 7937 if (CMSTestInFreeList) { 7938 if (freeRangeInFreeLists) { 7939 FreeChunk* fc = (FreeChunk*) freeFinger; 7940 assert(fc->is_free(), "A chunk on the free list should be free."); 7941 assert(fc->size() > 0, "Free range should have a size"); 7942 assert(_sp->verify_chunk_in_free_list(fc), "Chunk is not in free lists"); 7943 } 7944 } 7945 } 7946 7947 // Note that the sweeper runs concurrently with mutators. Thus, 7948 // it is possible for direct allocation in this generation to happen 7949 // in the middle of the sweep. Note that the sweeper also coalesces 7950 // contiguous free blocks. Thus, unless the sweeper and the allocator 7951 // synchronize appropriately freshly allocated blocks may get swept up. 7952 // This is accomplished by the sweeper locking the free lists while 7953 // it is sweeping. Thus blocks that are determined to be free are 7954 // indeed free. There is however one additional complication: 7955 // blocks that have been allocated since the final checkpoint and 7956 // mark, will not have been marked and so would be treated as 7957 // unreachable and swept up. To prevent this, the allocator marks 7958 // the bit map when allocating during the sweep phase. This leads, 7959 // however, to a further complication -- objects may have been allocated 7960 // but not yet initialized -- in the sense that the header isn't yet 7961 // installed. The sweeper can not then determine the size of the block 7962 // in order to skip over it. To deal with this case, we use a technique 7963 // (due to Printezis) to encode such uninitialized block sizes in the 7964 // bit map. Since the bit map uses a bit per every HeapWord, but the 7965 // CMS generation has a minimum object size of 3 HeapWords, it follows 7966 // that "normal marks" won't be adjacent in the bit map (there will 7967 // always be at least two 0 bits between successive 1 bits). We make use 7968 // of these "unused" bits to represent uninitialized blocks -- the bit 7969 // corresponding to the start of the uninitialized object and the next 7970 // bit are both set. Finally, a 1 bit marks the end of the object that 7971 // started with the two consecutive 1 bits to indicate its potentially 7972 // uninitialized state. 7973 7974 size_t SweepClosure::do_blk_careful(HeapWord* addr) { 7975 FreeChunk* fc = (FreeChunk*)addr; 7976 size_t res; 7977 7978 // Check if we are done sweeping. Below we check "addr >= _limit" rather 7979 // than "addr == _limit" because although _limit was a block boundary when 7980 // we started the sweep, it may no longer be one because heap expansion 7981 // may have caused us to coalesce the block ending at the address _limit 7982 // with a newly expanded chunk (this happens when _limit was set to the 7983 // previous _end of the space), so we may have stepped past _limit: 7984 // see the following Zeno-like trail of CRs 6977970, 7008136, 7042740. 7985 if (addr >= _limit) { // we have swept up to or past the limit: finish up 7986 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7987 "sweep _limit out of bounds"); 7988 assert(addr < _sp->end(), "addr out of bounds"); 7989 // Flush any free range we might be holding as a single 7990 // coalesced chunk to the appropriate free list. 7991 if (inFreeRange()) { 7992 assert(freeFinger() >= _sp->bottom() && freeFinger() < _limit, 7993 err_msg("freeFinger() " PTR_FORMAT" is out-of-bounds", freeFinger())); 7994 flush_cur_free_chunk(freeFinger(), 7995 pointer_delta(addr, freeFinger())); 7996 if (CMSTraceSweeper) { 7997 gclog_or_tty->print("Sweep: last chunk: "); 7998 gclog_or_tty->print("put_free_blk 0x%x ("SIZE_FORMAT") " 7999 "[coalesced:"SIZE_FORMAT"]\n", 8000 freeFinger(), pointer_delta(addr, freeFinger()), 8001 lastFreeRangeCoalesced()); 8002 } 8003 } 8004 8005 // help the iterator loop finish 8006 return pointer_delta(_sp->end(), addr); 8007 } 8008 8009 assert(addr < _limit, "sweep invariant"); 8010 // check if we should yield 8011 do_yield_check(addr); 8012 if (fc->is_free()) { 8013 // Chunk that is already free 8014 res = fc->size(); 8015 do_already_free_chunk(fc); 8016 debug_only(_sp->verifyFreeLists()); 8017 // If we flush the chunk at hand in lookahead_and_flush() 8018 // and it's coalesced with a preceding chunk, then the 8019 // process of "mangling" the payload of the coalesced block 8020 // will cause erasure of the size information from the 8021 // (erstwhile) header of all the coalesced blocks but the 8022 // first, so the first disjunct in the assert will not hold 8023 // in that specific case (in which case the second disjunct 8024 // will hold). 8025 assert(res == fc->size() || ((HeapWord*)fc) + res >= _limit, 8026 "Otherwise the size info doesn't change at this step"); 8027 NOT_PRODUCT( 8028 _numObjectsAlreadyFree++; 8029 _numWordsAlreadyFree += res; 8030 ) 8031 NOT_PRODUCT(_last_fc = fc;) 8032 } else if (!_bitMap->isMarked(addr)) { 8033 // Chunk is fresh garbage 8034 res = do_garbage_chunk(fc); 8035 debug_only(_sp->verifyFreeLists()); 8036 NOT_PRODUCT( 8037 _numObjectsFreed++; 8038 _numWordsFreed += res; 8039 ) 8040 } else { 8041 // Chunk that is alive. 8042 res = do_live_chunk(fc); 8043 debug_only(_sp->verifyFreeLists()); 8044 NOT_PRODUCT( 8045 _numObjectsLive++; 8046 _numWordsLive += res; 8047 ) 8048 } 8049 return res; 8050 } 8051 8052 // For the smart allocation, record following 8053 // split deaths - a free chunk is removed from its free list because 8054 // it is being split into two or more chunks. 8055 // split birth - a free chunk is being added to its free list because 8056 // a larger free chunk has been split and resulted in this free chunk. 8057 // coal death - a free chunk is being removed from its free list because 8058 // it is being coalesced into a large free chunk. 8059 // coal birth - a free chunk is being added to its free list because 8060 // it was created when two or more free chunks where coalesced into 8061 // this free chunk. 8062 // 8063 // These statistics are used to determine the desired number of free 8064 // chunks of a given size. The desired number is chosen to be relative 8065 // to the end of a CMS sweep. The desired number at the end of a sweep 8066 // is the 8067 // count-at-end-of-previous-sweep (an amount that was enough) 8068 // - count-at-beginning-of-current-sweep (the excess) 8069 // + split-births (gains in this size during interval) 8070 // - split-deaths (demands on this size during interval) 8071 // where the interval is from the end of one sweep to the end of the 8072 // next. 8073 // 8074 // When sweeping the sweeper maintains an accumulated chunk which is 8075 // the chunk that is made up of chunks that have been coalesced. That 8076 // will be termed the left-hand chunk. A new chunk of garbage that 8077 // is being considered for coalescing will be referred to as the 8078 // right-hand chunk. 8079 // 8080 // When making a decision on whether to coalesce a right-hand chunk with 8081 // the current left-hand chunk, the current count vs. the desired count 8082 // of the left-hand chunk is considered. Also if the right-hand chunk 8083 // is near the large chunk at the end of the heap (see 8084 // ConcurrentMarkSweepGeneration::isNearLargestChunk()), then the 8085 // left-hand chunk is coalesced. 8086 // 8087 // When making a decision about whether to split a chunk, the desired count 8088 // vs. the current count of the candidate to be split is also considered. 8089 // If the candidate is underpopulated (currently fewer chunks than desired) 8090 // a chunk of an overpopulated (currently more chunks than desired) size may 8091 // be chosen. The "hint" associated with a free list, if non-null, points 8092 // to a free list which may be overpopulated. 8093 // 8094 8095 void SweepClosure::do_already_free_chunk(FreeChunk* fc) { 8096 const size_t size = fc->size(); 8097 // Chunks that cannot be coalesced are not in the 8098 // free lists. 8099 if (CMSTestInFreeList && !fc->cantCoalesce()) { 8100 assert(_sp->verify_chunk_in_free_list(fc), 8101 "free chunk should be in free lists"); 8102 } 8103 // a chunk that is already free, should not have been 8104 // marked in the bit map 8105 HeapWord* const addr = (HeapWord*) fc; 8106 assert(!_bitMap->isMarked(addr), "free chunk should be unmarked"); 8107 // Verify that the bit map has no bits marked between 8108 // addr and purported end of this block. 8109 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 8110 8111 // Some chunks cannot be coalesced under any circumstances. 8112 // See the definition of cantCoalesce(). 8113 if (!fc->cantCoalesce()) { 8114 // This chunk can potentially be coalesced. 8115 if (_sp->adaptive_freelists()) { 8116 // All the work is done in 8117 do_post_free_or_garbage_chunk(fc, size); 8118 } else { // Not adaptive free lists 8119 // this is a free chunk that can potentially be coalesced by the sweeper; 8120 if (!inFreeRange()) { 8121 // if the next chunk is a free block that can't be coalesced 8122 // it doesn't make sense to remove this chunk from the free lists 8123 FreeChunk* nextChunk = (FreeChunk*)(addr + size); 8124 assert((HeapWord*)nextChunk <= _sp->end(), "Chunk size out of bounds?"); 8125 if ((HeapWord*)nextChunk < _sp->end() && // There is another free chunk to the right ... 8126 nextChunk->is_free() && // ... which is free... 8127 nextChunk->cantCoalesce()) { // ... but can't be coalesced 8128 // nothing to do 8129 } else { 8130 // Potentially the start of a new free range: 8131 // Don't eagerly remove it from the free lists. 8132 // No need to remove it if it will just be put 8133 // back again. (Also from a pragmatic point of view 8134 // if it is a free block in a region that is beyond 8135 // any allocated blocks, an assertion will fail) 8136 // Remember the start of a free run. 8137 initialize_free_range(addr, true); 8138 // end - can coalesce with next chunk 8139 } 8140 } else { 8141 // the midst of a free range, we are coalescing 8142 print_free_block_coalesced(fc); 8143 if (CMSTraceSweeper) { 8144 gclog_or_tty->print(" -- pick up free block 0x%x (%d)\n", fc, size); 8145 } 8146 // remove it from the free lists 8147 _sp->removeFreeChunkFromFreeLists(fc); 8148 set_lastFreeRangeCoalesced(true); 8149 // If the chunk is being coalesced and the current free range is 8150 // in the free lists, remove the current free range so that it 8151 // will be returned to the free lists in its entirety - all 8152 // the coalesced pieces included. 8153 if (freeRangeInFreeLists()) { 8154 FreeChunk* ffc = (FreeChunk*) freeFinger(); 8155 assert(ffc->size() == pointer_delta(addr, freeFinger()), 8156 "Size of free range is inconsistent with chunk size."); 8157 if (CMSTestInFreeList) { 8158 assert(_sp->verify_chunk_in_free_list(ffc), 8159 "free range is not in free lists"); 8160 } 8161 _sp->removeFreeChunkFromFreeLists(ffc); 8162 set_freeRangeInFreeLists(false); 8163 } 8164 } 8165 } 8166 // Note that if the chunk is not coalescable (the else arm 8167 // below), we unconditionally flush, without needing to do 8168 // a "lookahead," as we do below. 8169 if (inFreeRange()) lookahead_and_flush(fc, size); 8170 } else { 8171 // Code path common to both original and adaptive free lists. 8172 8173 // cant coalesce with previous block; this should be treated 8174 // as the end of a free run if any 8175 if (inFreeRange()) { 8176 // we kicked some butt; time to pick up the garbage 8177 assert(freeFinger() < addr, "freeFinger points too high"); 8178 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 8179 } 8180 // else, nothing to do, just continue 8181 } 8182 } 8183 8184 size_t SweepClosure::do_garbage_chunk(FreeChunk* fc) { 8185 // This is a chunk of garbage. It is not in any free list. 8186 // Add it to a free list or let it possibly be coalesced into 8187 // a larger chunk. 8188 HeapWord* const addr = (HeapWord*) fc; 8189 const size_t size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 8190 8191 if (_sp->adaptive_freelists()) { 8192 // Verify that the bit map has no bits marked between 8193 // addr and purported end of just dead object. 8194 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 8195 8196 do_post_free_or_garbage_chunk(fc, size); 8197 } else { 8198 if (!inFreeRange()) { 8199 // start of a new free range 8200 assert(size > 0, "A free range should have a size"); 8201 initialize_free_range(addr, false); 8202 } else { 8203 // this will be swept up when we hit the end of the 8204 // free range 8205 if (CMSTraceSweeper) { 8206 gclog_or_tty->print(" -- pick up garbage 0x%x (%d) \n", fc, size); 8207 } 8208 // If the chunk is being coalesced and the current free range is 8209 // in the free lists, remove the current free range so that it 8210 // will be returned to the free lists in its entirety - all 8211 // the coalesced pieces included. 8212 if (freeRangeInFreeLists()) { 8213 FreeChunk* ffc = (FreeChunk*)freeFinger(); 8214 assert(ffc->size() == pointer_delta(addr, freeFinger()), 8215 "Size of free range is inconsistent with chunk size."); 8216 if (CMSTestInFreeList) { 8217 assert(_sp->verify_chunk_in_free_list(ffc), 8218 "free range is not in free lists"); 8219 } 8220 _sp->removeFreeChunkFromFreeLists(ffc); 8221 set_freeRangeInFreeLists(false); 8222 } 8223 set_lastFreeRangeCoalesced(true); 8224 } 8225 // this will be swept up when we hit the end of the free range 8226 8227 // Verify that the bit map has no bits marked between 8228 // addr and purported end of just dead object. 8229 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 8230 } 8231 assert(_limit >= addr + size, 8232 "A freshly garbage chunk can't possibly straddle over _limit"); 8233 if (inFreeRange()) lookahead_and_flush(fc, size); 8234 return size; 8235 } 8236 8237 size_t SweepClosure::do_live_chunk(FreeChunk* fc) { 8238 HeapWord* addr = (HeapWord*) fc; 8239 // The sweeper has just found a live object. Return any accumulated 8240 // left hand chunk to the free lists. 8241 if (inFreeRange()) { 8242 assert(freeFinger() < addr, "freeFinger points too high"); 8243 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 8244 } 8245 8246 // This object is live: we'd normally expect this to be 8247 // an oop, and like to assert the following: 8248 // assert(oop(addr)->is_oop(), "live block should be an oop"); 8249 // However, as we commented above, this may be an object whose 8250 // header hasn't yet been initialized. 8251 size_t size; 8252 assert(_bitMap->isMarked(addr), "Tautology for this control point"); 8253 if (_bitMap->isMarked(addr + 1)) { 8254 // Determine the size from the bit map, rather than trying to 8255 // compute it from the object header. 8256 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 8257 size = pointer_delta(nextOneAddr + 1, addr); 8258 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 8259 "alignment problem"); 8260 8261 #ifdef DEBUG 8262 if (oop(addr)->klass_or_null() != NULL) { 8263 // Ignore mark word because we are running concurrent with mutators 8264 assert(oop(addr)->is_oop(true), "live block should be an oop"); 8265 assert(size == 8266 CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()), 8267 "P-mark and computed size do not agree"); 8268 } 8269 #endif 8270 8271 } else { 8272 // This should be an initialized object that's alive. 8273 assert(oop(addr)->klass_or_null() != NULL, 8274 "Should be an initialized object"); 8275 // Ignore mark word because we are running concurrent with mutators 8276 assert(oop(addr)->is_oop(true), "live block should be an oop"); 8277 // Verify that the bit map has no bits marked between 8278 // addr and purported end of this block. 8279 size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 8280 assert(size >= 3, "Necessary for Printezis marks to work"); 8281 assert(!_bitMap->isMarked(addr+1), "Tautology for this control point"); 8282 DEBUG_ONLY(_bitMap->verifyNoOneBitsInRange(addr+2, addr+size);) 8283 } 8284 return size; 8285 } 8286 8287 void SweepClosure::do_post_free_or_garbage_chunk(FreeChunk* fc, 8288 size_t chunkSize) { 8289 // do_post_free_or_garbage_chunk() should only be called in the case 8290 // of the adaptive free list allocator. 8291 const bool fcInFreeLists = fc->is_free(); 8292 assert(_sp->adaptive_freelists(), "Should only be used in this case."); 8293 assert((HeapWord*)fc <= _limit, "sweep invariant"); 8294 if (CMSTestInFreeList && fcInFreeLists) { 8295 assert(_sp->verify_chunk_in_free_list(fc), "free chunk is not in free lists"); 8296 } 8297 8298 if (CMSTraceSweeper) { 8299 gclog_or_tty->print_cr(" -- pick up another chunk at 0x%x (%d)", fc, chunkSize); 8300 } 8301 8302 HeapWord* const fc_addr = (HeapWord*) fc; 8303 8304 bool coalesce; 8305 const size_t left = pointer_delta(fc_addr, freeFinger()); 8306 const size_t right = chunkSize; 8307 switch (FLSCoalescePolicy) { 8308 // numeric value forms a coalition aggressiveness metric 8309 case 0: { // never coalesce 8310 coalesce = false; 8311 break; 8312 } 8313 case 1: { // coalesce if left & right chunks on overpopulated lists 8314 coalesce = _sp->coalOverPopulated(left) && 8315 _sp->coalOverPopulated(right); 8316 break; 8317 } 8318 case 2: { // coalesce if left chunk on overpopulated list (default) 8319 coalesce = _sp->coalOverPopulated(left); 8320 break; 8321 } 8322 case 3: { // coalesce if left OR right chunk on overpopulated list 8323 coalesce = _sp->coalOverPopulated(left) || 8324 _sp->coalOverPopulated(right); 8325 break; 8326 } 8327 case 4: { // always coalesce 8328 coalesce = true; 8329 break; 8330 } 8331 default: 8332 ShouldNotReachHere(); 8333 } 8334 8335 // Should the current free range be coalesced? 8336 // If the chunk is in a free range and either we decided to coalesce above 8337 // or the chunk is near the large block at the end of the heap 8338 // (isNearLargestChunk() returns true), then coalesce this chunk. 8339 const bool doCoalesce = inFreeRange() 8340 && (coalesce || _g->isNearLargestChunk(fc_addr)); 8341 if (doCoalesce) { 8342 // Coalesce the current free range on the left with the new 8343 // chunk on the right. If either is on a free list, 8344 // it must be removed from the list and stashed in the closure. 8345 if (freeRangeInFreeLists()) { 8346 FreeChunk* const ffc = (FreeChunk*)freeFinger(); 8347 assert(ffc->size() == pointer_delta(fc_addr, freeFinger()), 8348 "Size of free range is inconsistent with chunk size."); 8349 if (CMSTestInFreeList) { 8350 assert(_sp->verify_chunk_in_free_list(ffc), 8351 "Chunk is not in free lists"); 8352 } 8353 _sp->coalDeath(ffc->size()); 8354 _sp->removeFreeChunkFromFreeLists(ffc); 8355 set_freeRangeInFreeLists(false); 8356 } 8357 if (fcInFreeLists) { 8358 _sp->coalDeath(chunkSize); 8359 assert(fc->size() == chunkSize, 8360 "The chunk has the wrong size or is not in the free lists"); 8361 _sp->removeFreeChunkFromFreeLists(fc); 8362 } 8363 set_lastFreeRangeCoalesced(true); 8364 print_free_block_coalesced(fc); 8365 } else { // not in a free range and/or should not coalesce 8366 // Return the current free range and start a new one. 8367 if (inFreeRange()) { 8368 // In a free range but cannot coalesce with the right hand chunk. 8369 // Put the current free range into the free lists. 8370 flush_cur_free_chunk(freeFinger(), 8371 pointer_delta(fc_addr, freeFinger())); 8372 } 8373 // Set up for new free range. Pass along whether the right hand 8374 // chunk is in the free lists. 8375 initialize_free_range((HeapWord*)fc, fcInFreeLists); 8376 } 8377 } 8378 8379 // Lookahead flush: 8380 // If we are tracking a free range, and this is the last chunk that 8381 // we'll look at because its end crosses past _limit, we'll preemptively 8382 // flush it along with any free range we may be holding on to. Note that 8383 // this can be the case only for an already free or freshly garbage 8384 // chunk. If this block is an object, it can never straddle 8385 // over _limit. The "straddling" occurs when _limit is set at 8386 // the previous end of the space when this cycle started, and 8387 // a subsequent heap expansion caused the previously co-terminal 8388 // free block to be coalesced with the newly expanded portion, 8389 // thus rendering _limit a non-block-boundary making it dangerous 8390 // for the sweeper to step over and examine. 8391 void SweepClosure::lookahead_and_flush(FreeChunk* fc, size_t chunk_size) { 8392 assert(inFreeRange(), "Should only be called if currently in a free range."); 8393 HeapWord* const eob = ((HeapWord*)fc) + chunk_size; 8394 assert(_sp->used_region().contains(eob - 1), 8395 err_msg("eob = " PTR_FORMAT " out of bounds wrt _sp = [" PTR_FORMAT "," PTR_FORMAT ")" 8396 " when examining fc = " PTR_FORMAT "(" SIZE_FORMAT ")", 8397 _limit, _sp->bottom(), _sp->end(), fc, chunk_size)); 8398 if (eob >= _limit) { 8399 assert(eob == _limit || fc->is_free(), "Only a free chunk should allow us to cross over the limit"); 8400 if (CMSTraceSweeper) { 8401 gclog_or_tty->print_cr("_limit " PTR_FORMAT " reached or crossed by block " 8402 "[" PTR_FORMAT "," PTR_FORMAT ") in space " 8403 "[" PTR_FORMAT "," PTR_FORMAT ")", 8404 _limit, fc, eob, _sp->bottom(), _sp->end()); 8405 } 8406 // Return the storage we are tracking back into the free lists. 8407 if (CMSTraceSweeper) { 8408 gclog_or_tty->print_cr("Flushing ... "); 8409 } 8410 assert(freeFinger() < eob, "Error"); 8411 flush_cur_free_chunk( freeFinger(), pointer_delta(eob, freeFinger())); 8412 } 8413 } 8414 8415 void SweepClosure::flush_cur_free_chunk(HeapWord* chunk, size_t size) { 8416 assert(inFreeRange(), "Should only be called if currently in a free range."); 8417 assert(size > 0, 8418 "A zero sized chunk cannot be added to the free lists."); 8419 if (!freeRangeInFreeLists()) { 8420 if (CMSTestInFreeList) { 8421 FreeChunk* fc = (FreeChunk*) chunk; 8422 fc->set_size(size); 8423 assert(!_sp->verify_chunk_in_free_list(fc), 8424 "chunk should not be in free lists yet"); 8425 } 8426 if (CMSTraceSweeper) { 8427 gclog_or_tty->print_cr(" -- add free block 0x%x (%d) to free lists", 8428 chunk, size); 8429 } 8430 // A new free range is going to be starting. The current 8431 // free range has not been added to the free lists yet or 8432 // was removed so add it back. 8433 // If the current free range was coalesced, then the death 8434 // of the free range was recorded. Record a birth now. 8435 if (lastFreeRangeCoalesced()) { 8436 _sp->coalBirth(size); 8437 } 8438 _sp->addChunkAndRepairOffsetTable(chunk, size, 8439 lastFreeRangeCoalesced()); 8440 } else if (CMSTraceSweeper) { 8441 gclog_or_tty->print_cr("Already in free list: nothing to flush"); 8442 } 8443 set_inFreeRange(false); 8444 set_freeRangeInFreeLists(false); 8445 } 8446 8447 // We take a break if we've been at this for a while, 8448 // so as to avoid monopolizing the locks involved. 8449 void SweepClosure::do_yield_work(HeapWord* addr) { 8450 // Return current free chunk being used for coalescing (if any) 8451 // to the appropriate freelist. After yielding, the next 8452 // free block encountered will start a coalescing range of 8453 // free blocks. If the next free block is adjacent to the 8454 // chunk just flushed, they will need to wait for the next 8455 // sweep to be coalesced. 8456 if (inFreeRange()) { 8457 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 8458 } 8459 8460 // First give up the locks, then yield, then re-lock. 8461 // We should probably use a constructor/destructor idiom to 8462 // do this unlock/lock or modify the MutexUnlocker class to 8463 // serve our purpose. XXX 8464 assert_lock_strong(_bitMap->lock()); 8465 assert_lock_strong(_freelistLock); 8466 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 8467 "CMS thread should hold CMS token"); 8468 _bitMap->lock()->unlock(); 8469 _freelistLock->unlock(); 8470 ConcurrentMarkSweepThread::desynchronize(true); 8471 ConcurrentMarkSweepThread::acknowledge_yield_request(); 8472 _collector->stopTimer(); 8473 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 8474 if (PrintCMSStatistics != 0) { 8475 _collector->incrementYields(); 8476 } 8477 _collector->icms_wait(); 8478 8479 // See the comment in coordinator_yield() 8480 for (unsigned i = 0; i < CMSYieldSleepCount && 8481 ConcurrentMarkSweepThread::should_yield() && 8482 !CMSCollector::foregroundGCIsActive(); ++i) { 8483 os::sleep(Thread::current(), 1, false); 8484 ConcurrentMarkSweepThread::acknowledge_yield_request(); 8485 } 8486 8487 ConcurrentMarkSweepThread::synchronize(true); 8488 _freelistLock->lock(); 8489 _bitMap->lock()->lock_without_safepoint_check(); 8490 _collector->startTimer(); 8491 } 8492 8493 #ifndef PRODUCT 8494 // This is actually very useful in a product build if it can 8495 // be called from the debugger. Compile it into the product 8496 // as needed. 8497 bool debug_verify_chunk_in_free_list(FreeChunk* fc) { 8498 return debug_cms_space->verify_chunk_in_free_list(fc); 8499 } 8500 #endif 8501 8502 void SweepClosure::print_free_block_coalesced(FreeChunk* fc) const { 8503 if (CMSTraceSweeper) { 8504 gclog_or_tty->print_cr("Sweep:coal_free_blk " PTR_FORMAT " (" SIZE_FORMAT ")", 8505 fc, fc->size()); 8506 } 8507 } 8508 8509 // CMSIsAliveClosure 8510 bool CMSIsAliveClosure::do_object_b(oop obj) { 8511 HeapWord* addr = (HeapWord*)obj; 8512 return addr != NULL && 8513 (!_span.contains(addr) || _bit_map->isMarked(addr)); 8514 } 8515 8516 8517 CMSKeepAliveClosure::CMSKeepAliveClosure( CMSCollector* collector, 8518 MemRegion span, 8519 CMSBitMap* bit_map, CMSMarkStack* mark_stack, 8520 bool cpc): 8521 _collector(collector), 8522 _span(span), 8523 _bit_map(bit_map), 8524 _mark_stack(mark_stack), 8525 _concurrent_precleaning(cpc) { 8526 assert(!_span.is_empty(), "Empty span could spell trouble"); 8527 } 8528 8529 8530 // CMSKeepAliveClosure: the serial version 8531 void CMSKeepAliveClosure::do_oop(oop obj) { 8532 HeapWord* addr = (HeapWord*)obj; 8533 if (_span.contains(addr) && 8534 !_bit_map->isMarked(addr)) { 8535 _bit_map->mark(addr); 8536 bool simulate_overflow = false; 8537 NOT_PRODUCT( 8538 if (CMSMarkStackOverflowALot && 8539 _collector->simulate_overflow()) { 8540 // simulate a stack overflow 8541 simulate_overflow = true; 8542 } 8543 ) 8544 if (simulate_overflow || !_mark_stack->push(obj)) { 8545 if (_concurrent_precleaning) { 8546 // We dirty the overflown object and let the remark 8547 // phase deal with it. 8548 assert(_collector->overflow_list_is_empty(), "Error"); 8549 // In the case of object arrays, we need to dirty all of 8550 // the cards that the object spans. No locking or atomics 8551 // are needed since no one else can be mutating the mod union 8552 // table. 8553 if (obj->is_objArray()) { 8554 size_t sz = obj->size(); 8555 HeapWord* end_card_addr = 8556 (HeapWord*)round_to((intptr_t)(addr+sz), CardTableModRefBS::card_size); 8557 MemRegion redirty_range = MemRegion(addr, end_card_addr); 8558 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 8559 _collector->_modUnionTable.mark_range(redirty_range); 8560 } else { 8561 _collector->_modUnionTable.mark(addr); 8562 } 8563 _collector->_ser_kac_preclean_ovflw++; 8564 } else { 8565 _collector->push_on_overflow_list(obj); 8566 _collector->_ser_kac_ovflw++; 8567 } 8568 } 8569 } 8570 } 8571 8572 void CMSKeepAliveClosure::do_oop(oop* p) { CMSKeepAliveClosure::do_oop_work(p); } 8573 void CMSKeepAliveClosure::do_oop(narrowOop* p) { CMSKeepAliveClosure::do_oop_work(p); } 8574 8575 // CMSParKeepAliveClosure: a parallel version of the above. 8576 // The work queues are private to each closure (thread), 8577 // but (may be) available for stealing by other threads. 8578 void CMSParKeepAliveClosure::do_oop(oop obj) { 8579 HeapWord* addr = (HeapWord*)obj; 8580 if (_span.contains(addr) && 8581 !_bit_map->isMarked(addr)) { 8582 // In general, during recursive tracing, several threads 8583 // may be concurrently getting here; the first one to 8584 // "tag" it, claims it. 8585 if (_bit_map->par_mark(addr)) { 8586 bool res = _work_queue->push(obj); 8587 assert(res, "Low water mark should be much less than capacity"); 8588 // Do a recursive trim in the hope that this will keep 8589 // stack usage lower, but leave some oops for potential stealers 8590 trim_queue(_low_water_mark); 8591 } // Else, another thread got there first 8592 } 8593 } 8594 8595 void CMSParKeepAliveClosure::do_oop(oop* p) { CMSParKeepAliveClosure::do_oop_work(p); } 8596 void CMSParKeepAliveClosure::do_oop(narrowOop* p) { CMSParKeepAliveClosure::do_oop_work(p); } 8597 8598 void CMSParKeepAliveClosure::trim_queue(uint max) { 8599 while (_work_queue->size() > max) { 8600 oop new_oop; 8601 if (_work_queue->pop_local(new_oop)) { 8602 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 8603 assert(_bit_map->isMarked((HeapWord*)new_oop), 8604 "no white objects on this stack!"); 8605 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 8606 // iterate over the oops in this oop, marking and pushing 8607 // the ones in CMS heap (i.e. in _span). 8608 new_oop->oop_iterate(&_mark_and_push); 8609 } 8610 } 8611 } 8612 8613 CMSInnerParMarkAndPushClosure::CMSInnerParMarkAndPushClosure( 8614 CMSCollector* collector, 8615 MemRegion span, CMSBitMap* bit_map, 8616 OopTaskQueue* work_queue): 8617 _collector(collector), 8618 _span(span), 8619 _bit_map(bit_map), 8620 _work_queue(work_queue) { } 8621 8622 void CMSInnerParMarkAndPushClosure::do_oop(oop obj) { 8623 HeapWord* addr = (HeapWord*)obj; 8624 if (_span.contains(addr) && 8625 !_bit_map->isMarked(addr)) { 8626 if (_bit_map->par_mark(addr)) { 8627 bool simulate_overflow = false; 8628 NOT_PRODUCT( 8629 if (CMSMarkStackOverflowALot && 8630 _collector->par_simulate_overflow()) { 8631 // simulate a stack overflow 8632 simulate_overflow = true; 8633 } 8634 ) 8635 if (simulate_overflow || !_work_queue->push(obj)) { 8636 _collector->par_push_on_overflow_list(obj); 8637 _collector->_par_kac_ovflw++; 8638 } 8639 } // Else another thread got there already 8640 } 8641 } 8642 8643 void CMSInnerParMarkAndPushClosure::do_oop(oop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); } 8644 void CMSInnerParMarkAndPushClosure::do_oop(narrowOop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); } 8645 8646 ////////////////////////////////////////////////////////////////// 8647 // CMSExpansionCause ///////////////////////////// 8648 ////////////////////////////////////////////////////////////////// 8649 const char* CMSExpansionCause::to_string(CMSExpansionCause::Cause cause) { 8650 switch (cause) { 8651 case _no_expansion: 8652 return "No expansion"; 8653 case _satisfy_free_ratio: 8654 return "Free ratio"; 8655 case _satisfy_promotion: 8656 return "Satisfy promotion"; 8657 case _satisfy_allocation: 8658 return "allocation"; 8659 case _allocate_par_lab: 8660 return "Par LAB"; 8661 case _allocate_par_spooling_space: 8662 return "Par Spooling Space"; 8663 case _adaptive_size_policy: 8664 return "Ergonomics"; 8665 default: 8666 return "unknown"; 8667 } 8668 } 8669 8670 void CMSDrainMarkingStackClosure::do_void() { 8671 // the max number to take from overflow list at a time 8672 const size_t num = _mark_stack->capacity()/4; 8673 assert(!_concurrent_precleaning || _collector->overflow_list_is_empty(), 8674 "Overflow list should be NULL during concurrent phases"); 8675 while (!_mark_stack->isEmpty() || 8676 // if stack is empty, check the overflow list 8677 _collector->take_from_overflow_list(num, _mark_stack)) { 8678 oop obj = _mark_stack->pop(); 8679 HeapWord* addr = (HeapWord*)obj; 8680 assert(_span.contains(addr), "Should be within span"); 8681 assert(_bit_map->isMarked(addr), "Should be marked"); 8682 assert(obj->is_oop(), "Should be an oop"); 8683 obj->oop_iterate(_keep_alive); 8684 } 8685 } 8686 8687 void CMSParDrainMarkingStackClosure::do_void() { 8688 // drain queue 8689 trim_queue(0); 8690 } 8691 8692 // Trim our work_queue so its length is below max at return 8693 void CMSParDrainMarkingStackClosure::trim_queue(uint max) { 8694 while (_work_queue->size() > max) { 8695 oop new_oop; 8696 if (_work_queue->pop_local(new_oop)) { 8697 assert(new_oop->is_oop(), "Expected an oop"); 8698 assert(_bit_map->isMarked((HeapWord*)new_oop), 8699 "no white objects on this stack!"); 8700 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 8701 // iterate over the oops in this oop, marking and pushing 8702 // the ones in CMS heap (i.e. in _span). 8703 new_oop->oop_iterate(&_mark_and_push); 8704 } 8705 } 8706 } 8707 8708 //////////////////////////////////////////////////////////////////// 8709 // Support for Marking Stack Overflow list handling and related code 8710 //////////////////////////////////////////////////////////////////// 8711 // Much of the following code is similar in shape and spirit to the 8712 // code used in ParNewGC. We should try and share that code 8713 // as much as possible in the future. 8714 8715 #ifndef PRODUCT 8716 // Debugging support for CMSStackOverflowALot 8717 8718 // It's OK to call this multi-threaded; the worst thing 8719 // that can happen is that we'll get a bunch of closely 8720 // spaced simulated oveflows, but that's OK, in fact 8721 // probably good as it would exercise the overflow code 8722 // under contention. 8723 bool CMSCollector::simulate_overflow() { 8724 if (_overflow_counter-- <= 0) { // just being defensive 8725 _overflow_counter = CMSMarkStackOverflowInterval; 8726 return true; 8727 } else { 8728 return false; 8729 } 8730 } 8731 8732 bool CMSCollector::par_simulate_overflow() { 8733 return simulate_overflow(); 8734 } 8735 #endif 8736 8737 // Single-threaded 8738 bool CMSCollector::take_from_overflow_list(size_t num, CMSMarkStack* stack) { 8739 assert(stack->isEmpty(), "Expected precondition"); 8740 assert(stack->capacity() > num, "Shouldn't bite more than can chew"); 8741 size_t i = num; 8742 oop cur = _overflow_list; 8743 const markOop proto = markOopDesc::prototype(); 8744 NOT_PRODUCT(ssize_t n = 0;) 8745 for (oop next; i > 0 && cur != NULL; cur = next, i--) { 8746 next = oop(cur->mark()); 8747 cur->set_mark(proto); // until proven otherwise 8748 assert(cur->is_oop(), "Should be an oop"); 8749 bool res = stack->push(cur); 8750 assert(res, "Bit off more than can chew?"); 8751 NOT_PRODUCT(n++;) 8752 } 8753 _overflow_list = cur; 8754 #ifndef PRODUCT 8755 assert(_num_par_pushes >= n, "Too many pops?"); 8756 _num_par_pushes -=n; 8757 #endif 8758 return !stack->isEmpty(); 8759 } 8760 8761 #define BUSY (oop(0x1aff1aff)) 8762 // (MT-safe) Get a prefix of at most "num" from the list. 8763 // The overflow list is chained through the mark word of 8764 // each object in the list. We fetch the entire list, 8765 // break off a prefix of the right size and return the 8766 // remainder. If other threads try to take objects from 8767 // the overflow list at that time, they will wait for 8768 // some time to see if data becomes available. If (and 8769 // only if) another thread places one or more object(s) 8770 // on the global list before we have returned the suffix 8771 // to the global list, we will walk down our local list 8772 // to find its end and append the global list to 8773 // our suffix before returning it. This suffix walk can 8774 // prove to be expensive (quadratic in the amount of traffic) 8775 // when there are many objects in the overflow list and 8776 // there is much producer-consumer contention on the list. 8777 // *NOTE*: The overflow list manipulation code here and 8778 // in ParNewGeneration:: are very similar in shape, 8779 // except that in the ParNew case we use the old (from/eden) 8780 // copy of the object to thread the list via its klass word. 8781 // Because of the common code, if you make any changes in 8782 // the code below, please check the ParNew version to see if 8783 // similar changes might be needed. 8784 // CR 6797058 has been filed to consolidate the common code. 8785 bool CMSCollector::par_take_from_overflow_list(size_t num, 8786 OopTaskQueue* work_q, 8787 int no_of_gc_threads) { 8788 assert(work_q->size() == 0, "First empty local work queue"); 8789 assert(num < work_q->max_elems(), "Can't bite more than we can chew"); 8790 if (_overflow_list == NULL) { 8791 return false; 8792 } 8793 // Grab the entire list; we'll put back a suffix 8794 oop prefix = (oop)Atomic::xchg_ptr(BUSY, &_overflow_list); 8795 Thread* tid = Thread::current(); 8796 // Before "no_of_gc_threads" was introduced CMSOverflowSpinCount was 8797 // set to ParallelGCThreads. 8798 size_t CMSOverflowSpinCount = (size_t) no_of_gc_threads; // was ParallelGCThreads; 8799 size_t sleep_time_millis = MAX2((size_t)1, num/100); 8800 // If the list is busy, we spin for a short while, 8801 // sleeping between attempts to get the list. 8802 for (size_t spin = 0; prefix == BUSY && spin < CMSOverflowSpinCount; spin++) { 8803 os::sleep(tid, sleep_time_millis, false); 8804 if (_overflow_list == NULL) { 8805 // Nothing left to take 8806 return false; 8807 } else if (_overflow_list != BUSY) { 8808 // Try and grab the prefix 8809 prefix = (oop)Atomic::xchg_ptr(BUSY, &_overflow_list); 8810 } 8811 } 8812 // If the list was found to be empty, or we spun long 8813 // enough, we give up and return empty-handed. If we leave 8814 // the list in the BUSY state below, it must be the case that 8815 // some other thread holds the overflow list and will set it 8816 // to a non-BUSY state in the future. 8817 if (prefix == NULL || prefix == BUSY) { 8818 // Nothing to take or waited long enough 8819 if (prefix == NULL) { 8820 // Write back the NULL in case we overwrote it with BUSY above 8821 // and it is still the same value. 8822 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY); 8823 } 8824 return false; 8825 } 8826 assert(prefix != NULL && prefix != BUSY, "Error"); 8827 size_t i = num; 8828 oop cur = prefix; 8829 // Walk down the first "num" objects, unless we reach the end. 8830 for (; i > 1 && cur->mark() != NULL; cur = oop(cur->mark()), i--); 8831 if (cur->mark() == NULL) { 8832 // We have "num" or fewer elements in the list, so there 8833 // is nothing to return to the global list. 8834 // Write back the NULL in lieu of the BUSY we wrote 8835 // above, if it is still the same value. 8836 if (_overflow_list == BUSY) { 8837 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY); 8838 } 8839 } else { 8840 // Chop off the suffix and rerturn it to the global list. 8841 assert(cur->mark() != BUSY, "Error"); 8842 oop suffix_head = cur->mark(); // suffix will be put back on global list 8843 cur->set_mark(NULL); // break off suffix 8844 // It's possible that the list is still in the empty(busy) state 8845 // we left it in a short while ago; in that case we may be 8846 // able to place back the suffix without incurring the cost 8847 // of a walk down the list. 8848 oop observed_overflow_list = _overflow_list; 8849 oop cur_overflow_list = observed_overflow_list; 8850 bool attached = false; 8851 while (observed_overflow_list == BUSY || observed_overflow_list == NULL) { 8852 observed_overflow_list = 8853 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list); 8854 if (cur_overflow_list == observed_overflow_list) { 8855 attached = true; 8856 break; 8857 } else cur_overflow_list = observed_overflow_list; 8858 } 8859 if (!attached) { 8860 // Too bad, someone else sneaked in (at least) an element; we'll need 8861 // to do a splice. Find tail of suffix so we can prepend suffix to global 8862 // list. 8863 for (cur = suffix_head; cur->mark() != NULL; cur = (oop)(cur->mark())); 8864 oop suffix_tail = cur; 8865 assert(suffix_tail != NULL && suffix_tail->mark() == NULL, 8866 "Tautology"); 8867 observed_overflow_list = _overflow_list; 8868 do { 8869 cur_overflow_list = observed_overflow_list; 8870 if (cur_overflow_list != BUSY) { 8871 // Do the splice ... 8872 suffix_tail->set_mark(markOop(cur_overflow_list)); 8873 } else { // cur_overflow_list == BUSY 8874 suffix_tail->set_mark(NULL); 8875 } 8876 // ... and try to place spliced list back on overflow_list ... 8877 observed_overflow_list = 8878 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list); 8879 } while (cur_overflow_list != observed_overflow_list); 8880 // ... until we have succeeded in doing so. 8881 } 8882 } 8883 8884 // Push the prefix elements on work_q 8885 assert(prefix != NULL, "control point invariant"); 8886 const markOop proto = markOopDesc::prototype(); 8887 oop next; 8888 NOT_PRODUCT(ssize_t n = 0;) 8889 for (cur = prefix; cur != NULL; cur = next) { 8890 next = oop(cur->mark()); 8891 cur->set_mark(proto); // until proven otherwise 8892 assert(cur->is_oop(), "Should be an oop"); 8893 bool res = work_q->push(cur); 8894 assert(res, "Bit off more than we can chew?"); 8895 NOT_PRODUCT(n++;) 8896 } 8897 #ifndef PRODUCT 8898 assert(_num_par_pushes >= n, "Too many pops?"); 8899 Atomic::add_ptr(-(intptr_t)n, &_num_par_pushes); 8900 #endif 8901 return true; 8902 } 8903 8904 // Single-threaded 8905 void CMSCollector::push_on_overflow_list(oop p) { 8906 NOT_PRODUCT(_num_par_pushes++;) 8907 assert(p->is_oop(), "Not an oop"); 8908 preserve_mark_if_necessary(p); 8909 p->set_mark((markOop)_overflow_list); 8910 _overflow_list = p; 8911 } 8912 8913 // Multi-threaded; use CAS to prepend to overflow list 8914 void CMSCollector::par_push_on_overflow_list(oop p) { 8915 NOT_PRODUCT(Atomic::inc_ptr(&_num_par_pushes);) 8916 assert(p->is_oop(), "Not an oop"); 8917 par_preserve_mark_if_necessary(p); 8918 oop observed_overflow_list = _overflow_list; 8919 oop cur_overflow_list; 8920 do { 8921 cur_overflow_list = observed_overflow_list; 8922 if (cur_overflow_list != BUSY) { 8923 p->set_mark(markOop(cur_overflow_list)); 8924 } else { 8925 p->set_mark(NULL); 8926 } 8927 observed_overflow_list = 8928 (oop) Atomic::cmpxchg_ptr(p, &_overflow_list, cur_overflow_list); 8929 } while (cur_overflow_list != observed_overflow_list); 8930 } 8931 #undef BUSY 8932 8933 // Single threaded 8934 // General Note on GrowableArray: pushes may silently fail 8935 // because we are (temporarily) out of C-heap for expanding 8936 // the stack. The problem is quite ubiquitous and affects 8937 // a lot of code in the JVM. The prudent thing for GrowableArray 8938 // to do (for now) is to exit with an error. However, that may 8939 // be too draconian in some cases because the caller may be 8940 // able to recover without much harm. For such cases, we 8941 // should probably introduce a "soft_push" method which returns 8942 // an indication of success or failure with the assumption that 8943 // the caller may be able to recover from a failure; code in 8944 // the VM can then be changed, incrementally, to deal with such 8945 // failures where possible, thus, incrementally hardening the VM 8946 // in such low resource situations. 8947 void CMSCollector::preserve_mark_work(oop p, markOop m) { 8948 _preserved_oop_stack.push(p); 8949 _preserved_mark_stack.push(m); 8950 assert(m == p->mark(), "Mark word changed"); 8951 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), 8952 "bijection"); 8953 } 8954 8955 // Single threaded 8956 void CMSCollector::preserve_mark_if_necessary(oop p) { 8957 markOop m = p->mark(); 8958 if (m->must_be_preserved(p)) { 8959 preserve_mark_work(p, m); 8960 } 8961 } 8962 8963 void CMSCollector::par_preserve_mark_if_necessary(oop p) { 8964 markOop m = p->mark(); 8965 if (m->must_be_preserved(p)) { 8966 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag); 8967 // Even though we read the mark word without holding 8968 // the lock, we are assured that it will not change 8969 // because we "own" this oop, so no other thread can 8970 // be trying to push it on the overflow list; see 8971 // the assertion in preserve_mark_work() that checks 8972 // that m == p->mark(). 8973 preserve_mark_work(p, m); 8974 } 8975 } 8976 8977 // We should be able to do this multi-threaded, 8978 // a chunk of stack being a task (this is 8979 // correct because each oop only ever appears 8980 // once in the overflow list. However, it's 8981 // not very easy to completely overlap this with 8982 // other operations, so will generally not be done 8983 // until all work's been completed. Because we 8984 // expect the preserved oop stack (set) to be small, 8985 // it's probably fine to do this single-threaded. 8986 // We can explore cleverer concurrent/overlapped/parallel 8987 // processing of preserved marks if we feel the 8988 // need for this in the future. Stack overflow should 8989 // be so rare in practice and, when it happens, its 8990 // effect on performance so great that this will 8991 // likely just be in the noise anyway. 8992 void CMSCollector::restore_preserved_marks_if_any() { 8993 assert(SafepointSynchronize::is_at_safepoint(), 8994 "world should be stopped"); 8995 assert(Thread::current()->is_ConcurrentGC_thread() || 8996 Thread::current()->is_VM_thread(), 8997 "should be single-threaded"); 8998 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), 8999 "bijection"); 9000 9001 while (!_preserved_oop_stack.is_empty()) { 9002 oop p = _preserved_oop_stack.pop(); 9003 assert(p->is_oop(), "Should be an oop"); 9004 assert(_span.contains(p), "oop should be in _span"); 9005 assert(p->mark() == markOopDesc::prototype(), 9006 "Set when taken from overflow list"); 9007 markOop m = _preserved_mark_stack.pop(); 9008 p->set_mark(m); 9009 } 9010 assert(_preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(), 9011 "stacks were cleared above"); 9012 } 9013 9014 #ifndef PRODUCT 9015 bool CMSCollector::no_preserved_marks() const { 9016 return _preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(); 9017 } 9018 #endif 9019 9020 CMSAdaptiveSizePolicy* ASConcurrentMarkSweepGeneration::cms_size_policy() const 9021 { 9022 GenCollectedHeap* gch = (GenCollectedHeap*) GenCollectedHeap::heap(); 9023 CMSAdaptiveSizePolicy* size_policy = 9024 (CMSAdaptiveSizePolicy*) gch->gen_policy()->size_policy(); 9025 assert(size_policy->is_gc_cms_adaptive_size_policy(), 9026 "Wrong type for size policy"); 9027 return size_policy; 9028 } 9029 9030 void ASConcurrentMarkSweepGeneration::resize(size_t cur_promo_size, 9031 size_t desired_promo_size) { 9032 if (cur_promo_size < desired_promo_size) { 9033 size_t expand_bytes = desired_promo_size - cur_promo_size; 9034 if (PrintAdaptiveSizePolicy && Verbose) { 9035 gclog_or_tty->print_cr(" ASConcurrentMarkSweepGeneration::resize " 9036 "Expanding tenured generation by " SIZE_FORMAT " (bytes)", 9037 expand_bytes); 9038 } 9039 expand(expand_bytes, 9040 MinHeapDeltaBytes, 9041 CMSExpansionCause::_adaptive_size_policy); 9042 } else if (desired_promo_size < cur_promo_size) { 9043 size_t shrink_bytes = cur_promo_size - desired_promo_size; 9044 if (PrintAdaptiveSizePolicy && Verbose) { 9045 gclog_or_tty->print_cr(" ASConcurrentMarkSweepGeneration::resize " 9046 "Shrinking tenured generation by " SIZE_FORMAT " (bytes)", 9047 shrink_bytes); 9048 } 9049 shrink(shrink_bytes); 9050 } 9051 } 9052 9053 CMSGCAdaptivePolicyCounters* ASConcurrentMarkSweepGeneration::gc_adaptive_policy_counters() { 9054 GenCollectedHeap* gch = GenCollectedHeap::heap(); 9055 CMSGCAdaptivePolicyCounters* counters = 9056 (CMSGCAdaptivePolicyCounters*) gch->collector_policy()->counters(); 9057 assert(counters->kind() == GCPolicyCounters::CMSGCAdaptivePolicyCountersKind, 9058 "Wrong kind of counters"); 9059 return counters; 9060 } 9061 9062 9063 void ASConcurrentMarkSweepGeneration::update_counters() { 9064 if (UsePerfData) { 9065 _space_counters->update_all(); 9066 _gen_counters->update_all(); 9067 CMSGCAdaptivePolicyCounters* counters = gc_adaptive_policy_counters(); 9068 GenCollectedHeap* gch = GenCollectedHeap::heap(); 9069 CMSGCStats* gc_stats_l = (CMSGCStats*) gc_stats(); 9070 assert(gc_stats_l->kind() == GCStats::CMSGCStatsKind, 9071 "Wrong gc statistics type"); 9072 counters->update_counters(gc_stats_l); 9073 } 9074 } 9075 9076 void ASConcurrentMarkSweepGeneration::update_counters(size_t used) { 9077 if (UsePerfData) { 9078 _space_counters->update_used(used); 9079 _space_counters->update_capacity(); 9080 _gen_counters->update_all(); 9081 9082 CMSGCAdaptivePolicyCounters* counters = gc_adaptive_policy_counters(); 9083 GenCollectedHeap* gch = GenCollectedHeap::heap(); 9084 CMSGCStats* gc_stats_l = (CMSGCStats*) gc_stats(); 9085 assert(gc_stats_l->kind() == GCStats::CMSGCStatsKind, 9086 "Wrong gc statistics type"); 9087 counters->update_counters(gc_stats_l); 9088 } 9089 } 9090 9091 // The desired expansion delta is computed so that: 9092 // . desired free percentage or greater is used 9093 void ASConcurrentMarkSweepGeneration::compute_new_size() { 9094 assert_locked_or_safepoint(Heap_lock); 9095 9096 GenCollectedHeap* gch = (GenCollectedHeap*) GenCollectedHeap::heap(); 9097 9098 // If incremental collection failed, we just want to expand 9099 // to the limit. 9100 if (incremental_collection_failed()) { 9101 clear_incremental_collection_failed(); 9102 grow_to_reserved(); 9103 return; 9104 } 9105 9106 assert(UseAdaptiveSizePolicy, "Should be using adaptive sizing"); 9107 9108 assert(gch->kind() == CollectedHeap::GenCollectedHeap, 9109 "Wrong type of heap"); 9110 int prev_level = level() - 1; 9111 assert(prev_level >= 0, "The cms generation is the lowest generation"); 9112 Generation* prev_gen = gch->get_gen(prev_level); 9113 assert(prev_gen->kind() == Generation::ASParNew, 9114 "Wrong type of young generation"); 9115 ParNewGeneration* younger_gen = (ParNewGeneration*) prev_gen; 9116 size_t cur_eden = younger_gen->eden()->capacity(); 9117 CMSAdaptiveSizePolicy* size_policy = cms_size_policy(); 9118 size_t cur_promo = free(); 9119 size_policy->compute_tenured_generation_free_space(cur_promo, 9120 max_available(), 9121 cur_eden); 9122 resize(cur_promo, size_policy->promo_size()); 9123 9124 // Record the new size of the space in the cms generation 9125 // that is available for promotions. This is temporary. 9126 // It should be the desired promo size. 9127 size_policy->avg_cms_promo()->sample(free()); 9128 size_policy->avg_old_live()->sample(used()); 9129 9130 if (UsePerfData) { 9131 CMSGCAdaptivePolicyCounters* counters = gc_adaptive_policy_counters(); 9132 counters->update_cms_capacity_counter(capacity()); 9133 } 9134 } 9135 9136 void ASConcurrentMarkSweepGeneration::shrink_by(size_t desired_bytes) { 9137 assert_locked_or_safepoint(Heap_lock); 9138 assert_lock_strong(freelistLock()); 9139 HeapWord* old_end = _cmsSpace->end(); 9140 HeapWord* unallocated_start = _cmsSpace->unallocated_block(); 9141 assert(old_end >= unallocated_start, "Miscalculation of unallocated_start"); 9142 FreeChunk* chunk_at_end = find_chunk_at_end(); 9143 if (chunk_at_end == NULL) { 9144 // No room to shrink 9145 if (PrintGCDetails && Verbose) { 9146 gclog_or_tty->print_cr("No room to shrink: old_end " 9147 PTR_FORMAT " unallocated_start " PTR_FORMAT 9148 " chunk_at_end " PTR_FORMAT, 9149 old_end, unallocated_start, chunk_at_end); 9150 } 9151 return; 9152 } else { 9153 9154 // Find the chunk at the end of the space and determine 9155 // how much it can be shrunk. 9156 size_t shrinkable_size_in_bytes = chunk_at_end->size(); 9157 size_t aligned_shrinkable_size_in_bytes = 9158 align_size_down(shrinkable_size_in_bytes, os::vm_page_size()); 9159 assert(unallocated_start <= chunk_at_end->end(), 9160 "Inconsistent chunk at end of space"); 9161 size_t bytes = MIN2(desired_bytes, aligned_shrinkable_size_in_bytes); 9162 size_t word_size_before = heap_word_size(_virtual_space.committed_size()); 9163 9164 // Shrink the underlying space 9165 _virtual_space.shrink_by(bytes); 9166 if (PrintGCDetails && Verbose) { 9167 gclog_or_tty->print_cr("ConcurrentMarkSweepGeneration::shrink_by:" 9168 " desired_bytes " SIZE_FORMAT 9169 " shrinkable_size_in_bytes " SIZE_FORMAT 9170 " aligned_shrinkable_size_in_bytes " SIZE_FORMAT 9171 " bytes " SIZE_FORMAT, 9172 desired_bytes, shrinkable_size_in_bytes, 9173 aligned_shrinkable_size_in_bytes, bytes); 9174 gclog_or_tty->print_cr(" old_end " SIZE_FORMAT 9175 " unallocated_start " SIZE_FORMAT, 9176 old_end, unallocated_start); 9177 } 9178 9179 // If the space did shrink (shrinking is not guaranteed), 9180 // shrink the chunk at the end by the appropriate amount. 9181 if (((HeapWord*)_virtual_space.high()) < old_end) { 9182 size_t new_word_size = 9183 heap_word_size(_virtual_space.committed_size()); 9184 9185 // Have to remove the chunk from the dictionary because it is changing 9186 // size and might be someplace elsewhere in the dictionary. 9187 9188 // Get the chunk at end, shrink it, and put it 9189 // back. 9190 _cmsSpace->removeChunkFromDictionary(chunk_at_end); 9191 size_t word_size_change = word_size_before - new_word_size; 9192 size_t chunk_at_end_old_size = chunk_at_end->size(); 9193 assert(chunk_at_end_old_size >= word_size_change, 9194 "Shrink is too large"); 9195 chunk_at_end->set_size(chunk_at_end_old_size - 9196 word_size_change); 9197 _cmsSpace->freed((HeapWord*) chunk_at_end->end(), 9198 word_size_change); 9199 9200 _cmsSpace->returnChunkToDictionary(chunk_at_end); 9201 9202 MemRegion mr(_cmsSpace->bottom(), new_word_size); 9203 _bts->resize(new_word_size); // resize the block offset shared array 9204 Universe::heap()->barrier_set()->resize_covered_region(mr); 9205 _cmsSpace->assert_locked(); 9206 _cmsSpace->set_end((HeapWord*)_virtual_space.high()); 9207 9208 NOT_PRODUCT(_cmsSpace->dictionary()->verify()); 9209 9210 // update the space and generation capacity counters 9211 if (UsePerfData) { 9212 _space_counters->update_capacity(); 9213 _gen_counters->update_all(); 9214 } 9215 9216 if (Verbose && PrintGCDetails) { 9217 size_t new_mem_size = _virtual_space.committed_size(); 9218 size_t old_mem_size = new_mem_size + bytes; 9219 gclog_or_tty->print_cr("Shrinking %s from %ldK by %ldK to %ldK", 9220 name(), old_mem_size/K, bytes/K, new_mem_size/K); 9221 } 9222 } 9223 9224 assert(_cmsSpace->unallocated_block() <= _cmsSpace->end(), 9225 "Inconsistency at end of space"); 9226 assert(chunk_at_end->end() == _cmsSpace->end(), 9227 "Shrinking is inconsistent"); 9228 return; 9229 } 9230 } 9231 9232 // Transfer some number of overflown objects to usual marking 9233 // stack. Return true if some objects were transferred. 9234 bool MarkRefsIntoAndScanClosure::take_from_overflow_list() { 9235 size_t num = MIN2((size_t)(_mark_stack->capacity() - _mark_stack->length())/4, 9236 (size_t)ParGCDesiredObjsFromOverflowList); 9237 9238 bool res = _collector->take_from_overflow_list(num, _mark_stack); 9239 assert(_collector->overflow_list_is_empty() || res, 9240 "If list is not empty, we should have taken something"); 9241 assert(!res || !_mark_stack->isEmpty(), 9242 "If we took something, it should now be on our stack"); 9243 return res; 9244 } 9245 9246 size_t MarkDeadObjectsClosure::do_blk(HeapWord* addr) { 9247 size_t res = _sp->block_size_no_stall(addr, _collector); 9248 if (_sp->block_is_obj(addr)) { 9249 if (_live_bit_map->isMarked(addr)) { 9250 // It can't have been dead in a previous cycle 9251 guarantee(!_dead_bit_map->isMarked(addr), "No resurrection!"); 9252 } else { 9253 _dead_bit_map->mark(addr); // mark the dead object 9254 } 9255 } 9256 // Could be 0, if the block size could not be computed without stalling. 9257 return res; 9258 } 9259 9260 TraceCMSMemoryManagerStats::TraceCMSMemoryManagerStats(CMSCollector::CollectorState phase, GCCause::Cause cause): TraceMemoryManagerStats() { 9261 9262 switch (phase) { 9263 case CMSCollector::InitialMarking: 9264 initialize(true /* fullGC */ , 9265 cause /* cause of the GC */, 9266 true /* recordGCBeginTime */, 9267 true /* recordPreGCUsage */, 9268 false /* recordPeakUsage */, 9269 false /* recordPostGCusage */, 9270 true /* recordAccumulatedGCTime */, 9271 false /* recordGCEndTime */, 9272 false /* countCollection */ ); 9273 break; 9274 9275 case CMSCollector::FinalMarking: 9276 initialize(true /* fullGC */ , 9277 cause /* cause of the GC */, 9278 false /* recordGCBeginTime */, 9279 false /* recordPreGCUsage */, 9280 false /* recordPeakUsage */, 9281 false /* recordPostGCusage */, 9282 true /* recordAccumulatedGCTime */, 9283 false /* recordGCEndTime */, 9284 false /* countCollection */ ); 9285 break; 9286 9287 case CMSCollector::Sweeping: 9288 initialize(true /* fullGC */ , 9289 cause /* cause of the GC */, 9290 false /* recordGCBeginTime */, 9291 false /* recordPreGCUsage */, 9292 true /* recordPeakUsage */, 9293 true /* recordPostGCusage */, 9294 false /* recordAccumulatedGCTime */, 9295 true /* recordGCEndTime */, 9296 true /* countCollection */ ); 9297 break; 9298 9299 default: 9300 ShouldNotReachHere(); 9301 } 9302 } 9303