1 /* 2 * Copyright (c) 2001, 2013, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "classfile/classLoaderData.hpp" 27 #include "classfile/symbolTable.hpp" 28 #include "classfile/systemDictionary.hpp" 29 #include "code/codeCache.hpp" 30 #include "gc_implementation/concurrentMarkSweep/cmsAdaptiveSizePolicy.hpp" 31 #include "gc_implementation/concurrentMarkSweep/cmsCollectorPolicy.hpp" 32 #include "gc_implementation/concurrentMarkSweep/cmsGCAdaptivePolicyCounters.hpp" 33 #include "gc_implementation/concurrentMarkSweep/cmsOopClosures.inline.hpp" 34 #include "gc_implementation/concurrentMarkSweep/compactibleFreeListSpace.hpp" 35 #include "gc_implementation/concurrentMarkSweep/concurrentMarkSweepGeneration.inline.hpp" 36 #include "gc_implementation/concurrentMarkSweep/concurrentMarkSweepThread.hpp" 37 #include "gc_implementation/concurrentMarkSweep/vmCMSOperations.hpp" 38 #include "gc_implementation/parNew/parNewGeneration.hpp" 39 #include "gc_implementation/shared/collectorCounters.hpp" 40 #include "gc_implementation/shared/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 "memory/tenuredGeneration.hpp" 52 #include "oops/oop.inline.hpp" 53 #include "prims/jvmtiExport.hpp" 54 #include "runtime/globals_extension.hpp" 55 #include "runtime/handles.inline.hpp" 56 #include "runtime/java.hpp" 57 #include "runtime/vmThread.hpp" 58 #include "services/memoryService.hpp" 59 #include "services/runtimeService.hpp" 60 61 // statics 62 CMSCollector* ConcurrentMarkSweepGeneration::_collector = NULL; 63 bool CMSCollector::_full_gc_requested = false; 64 65 ////////////////////////////////////////////////////////////////// 66 // In support of CMS/VM thread synchronization 67 ////////////////////////////////////////////////////////////////// 68 // We split use of the CGC_lock into 2 "levels". 69 // The low-level locking is of the usual CGC_lock monitor. We introduce 70 // a higher level "token" (hereafter "CMS token") built on top of the 71 // low level monitor (hereafter "CGC lock"). 72 // The token-passing protocol gives priority to the VM thread. The 73 // CMS-lock doesn't provide any fairness guarantees, but clients 74 // should ensure that it is only held for very short, bounded 75 // durations. 76 // 77 // When either of the CMS thread or the VM thread is involved in 78 // collection operations during which it does not want the other 79 // thread to interfere, it obtains the CMS token. 80 // 81 // If either thread tries to get the token while the other has 82 // it, that thread waits. However, if the VM thread and CMS thread 83 // both want the token, then the VM thread gets priority while the 84 // CMS thread waits. This ensures, for instance, that the "concurrent" 85 // phases of the CMS thread's work do not block out the VM thread 86 // for long periods of time as the CMS thread continues to hog 87 // the token. (See bug 4616232). 88 // 89 // The baton-passing functions are, however, controlled by the 90 // flags _foregroundGCShouldWait and _foregroundGCIsActive, 91 // and here the low-level CMS lock, not the high level token, 92 // ensures mutual exclusion. 93 // 94 // Two important conditions that we have to satisfy: 95 // 1. if a thread does a low-level wait on the CMS lock, then it 96 // relinquishes the CMS token if it were holding that token 97 // when it acquired the low-level CMS lock. 98 // 2. any low-level notifications on the low-level lock 99 // should only be sent when a thread has relinquished the token. 100 // 101 // In the absence of either property, we'd have potential deadlock. 102 // 103 // We protect each of the CMS (concurrent and sequential) phases 104 // with the CMS _token_, not the CMS _lock_. 105 // 106 // The only code protected by CMS lock is the token acquisition code 107 // itself, see ConcurrentMarkSweepThread::[de]synchronize(), and the 108 // baton-passing code. 109 // 110 // Unfortunately, i couldn't come up with a good abstraction to factor and 111 // hide the naked CGC_lock manipulation in the baton-passing code 112 // further below. That's something we should try to do. Also, the proof 113 // of correctness of this 2-level locking scheme is far from obvious, 114 // and potentially quite slippery. We have an uneasy supsicion, for instance, 115 // that there may be a theoretical possibility of delay/starvation in the 116 // low-level lock/wait/notify scheme used for the baton-passing because of 117 // potential intereference with the priority scheme embodied in the 118 // CMS-token-passing protocol. See related comments at a CGC_lock->wait() 119 // invocation further below and marked with "XXX 20011219YSR". 120 // Indeed, as we note elsewhere, this may become yet more slippery 121 // in the presence of multiple CMS and/or multiple VM threads. XXX 122 123 class CMSTokenSync: public StackObj { 124 private: 125 bool _is_cms_thread; 126 public: 127 CMSTokenSync(bool is_cms_thread): 128 _is_cms_thread(is_cms_thread) { 129 assert(is_cms_thread == Thread::current()->is_ConcurrentGC_thread(), 130 "Incorrect argument to constructor"); 131 ConcurrentMarkSweepThread::synchronize(_is_cms_thread); 132 } 133 134 ~CMSTokenSync() { 135 assert(_is_cms_thread ? 136 ConcurrentMarkSweepThread::cms_thread_has_cms_token() : 137 ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 138 "Incorrect state"); 139 ConcurrentMarkSweepThread::desynchronize(_is_cms_thread); 140 } 141 }; 142 143 // Convenience class that does a CMSTokenSync, and then acquires 144 // upto three locks. 145 class CMSTokenSyncWithLocks: public CMSTokenSync { 146 private: 147 // Note: locks are acquired in textual declaration order 148 // and released in the opposite order 149 MutexLockerEx _locker1, _locker2, _locker3; 150 public: 151 CMSTokenSyncWithLocks(bool is_cms_thread, Mutex* mutex1, 152 Mutex* mutex2 = NULL, Mutex* mutex3 = NULL): 153 CMSTokenSync(is_cms_thread), 154 _locker1(mutex1, Mutex::_no_safepoint_check_flag), 155 _locker2(mutex2, Mutex::_no_safepoint_check_flag), 156 _locker3(mutex3, Mutex::_no_safepoint_check_flag) 157 { } 158 }; 159 160 161 // Wrapper class to temporarily disable icms during a foreground cms collection. 162 class ICMSDisabler: public StackObj { 163 public: 164 // The ctor disables icms and wakes up the thread so it notices the change; 165 // the dtor re-enables icms. Note that the CMSCollector methods will check 166 // CMSIncrementalMode. 167 ICMSDisabler() { CMSCollector::disable_icms(); CMSCollector::start_icms(); } 168 ~ICMSDisabler() { CMSCollector::enable_icms(); } 169 }; 170 171 ////////////////////////////////////////////////////////////////// 172 // Concurrent Mark-Sweep Generation ///////////////////////////// 173 ////////////////////////////////////////////////////////////////// 174 175 NOT_PRODUCT(CompactibleFreeListSpace* debug_cms_space;) 176 177 // This struct contains per-thread things necessary to support parallel 178 // young-gen collection. 179 class CMSParGCThreadState: public CHeapObj<mtGC> { 180 public: 181 CFLS_LAB lab; 182 PromotionInfo promo; 183 184 // Constructor. 185 CMSParGCThreadState(CompactibleFreeListSpace* cfls) : lab(cfls) { 186 promo.setSpace(cfls); 187 } 188 }; 189 190 ConcurrentMarkSweepGeneration::ConcurrentMarkSweepGeneration( 191 ReservedSpace rs, size_t initial_byte_size, int level, 192 CardTableRS* ct, bool use_adaptive_freelists, 193 FreeBlockDictionary<FreeChunk>::DictionaryChoice dictionaryChoice) : 194 CardGeneration(rs, initial_byte_size, level, ct), 195 _dilatation_factor(((double)MinChunkSize)/((double)(CollectedHeap::min_fill_size()))), 196 _debug_collection_type(Concurrent_collection_type), 197 _did_compact(false) 198 { 199 HeapWord* bottom = (HeapWord*) _virtual_space.low(); 200 HeapWord* end = (HeapWord*) _virtual_space.high(); 201 202 _direct_allocated_words = 0; 203 NOT_PRODUCT( 204 _numObjectsPromoted = 0; 205 _numWordsPromoted = 0; 206 _numObjectsAllocated = 0; 207 _numWordsAllocated = 0; 208 ) 209 210 _cmsSpace = new CompactibleFreeListSpace(_bts, MemRegion(bottom, end), 211 use_adaptive_freelists, 212 dictionaryChoice); 213 NOT_PRODUCT(debug_cms_space = _cmsSpace;) 214 if (_cmsSpace == NULL) { 215 vm_exit_during_initialization( 216 "CompactibleFreeListSpace allocation failure"); 217 } 218 _cmsSpace->_gen = this; 219 220 _gc_stats = new CMSGCStats(); 221 222 // Verify the assumption that FreeChunk::_prev and OopDesc::_klass 223 // offsets match. The ability to tell free chunks from objects 224 // depends on this property. 225 debug_only( 226 FreeChunk* junk = NULL; 227 assert(UseCompressedKlassPointers || 228 junk->prev_addr() == (void*)(oop(junk)->klass_addr()), 229 "Offset of FreeChunk::_prev within FreeChunk must match" 230 " that of OopDesc::_klass within OopDesc"); 231 ) 232 if (CollectedHeap::use_parallel_gc_threads()) { 233 typedef CMSParGCThreadState* CMSParGCThreadStatePtr; 234 _par_gc_thread_states = 235 NEW_C_HEAP_ARRAY(CMSParGCThreadStatePtr, ParallelGCThreads, mtGC); 236 if (_par_gc_thread_states == NULL) { 237 vm_exit_during_initialization("Could not allocate par gc structs"); 238 } 239 for (uint i = 0; i < ParallelGCThreads; i++) { 240 _par_gc_thread_states[i] = new CMSParGCThreadState(cmsSpace()); 241 if (_par_gc_thread_states[i] == NULL) { 242 vm_exit_during_initialization("Could not allocate par gc structs"); 243 } 244 } 245 } else { 246 _par_gc_thread_states = NULL; 247 } 248 _incremental_collection_failed = false; 249 // The "dilatation_factor" is the expansion that can occur on 250 // account of the fact that the minimum object size in the CMS 251 // generation may be larger than that in, say, a contiguous young 252 // generation. 253 // Ideally, in the calculation below, we'd compute the dilatation 254 // factor as: MinChunkSize/(promoting_gen's min object size) 255 // Since we do not have such a general query interface for the 256 // promoting generation, we'll instead just use the mimimum 257 // object size (which today is a header's worth of space); 258 // note that all arithmetic is in units of HeapWords. 259 assert(MinChunkSize >= CollectedHeap::min_fill_size(), "just checking"); 260 assert(_dilatation_factor >= 1.0, "from previous assert"); 261 } 262 263 264 // The field "_initiating_occupancy" represents the occupancy percentage 265 // at which we trigger a new collection cycle. Unless explicitly specified 266 // via CMSInitiatingOccupancyFraction (argument "io" below), it 267 // is calculated by: 268 // 269 // Let "f" be MinHeapFreeRatio in 270 // 271 // _intiating_occupancy = 100-f + 272 // f * (CMSTriggerRatio/100) 273 // where CMSTriggerRatio is the argument "tr" below. 274 // 275 // That is, if we assume the heap is at its desired maximum occupancy at the 276 // end of a collection, we let CMSTriggerRatio of the (purported) free 277 // space be allocated before initiating a new collection cycle. 278 // 279 void ConcurrentMarkSweepGeneration::init_initiating_occupancy(intx io, uintx tr) { 280 assert(io <= 100 && tr <= 100, "Check the arguments"); 281 if (io >= 0) { 282 _initiating_occupancy = (double)io / 100.0; 283 } else { 284 _initiating_occupancy = ((100 - MinHeapFreeRatio) + 285 (double)(tr * MinHeapFreeRatio) / 100.0) 286 / 100.0; 287 } 288 } 289 290 void ConcurrentMarkSweepGeneration::ref_processor_init() { 291 assert(collector() != NULL, "no collector"); 292 collector()->ref_processor_init(); 293 } 294 295 void CMSCollector::ref_processor_init() { 296 if (_ref_processor == NULL) { 297 // Allocate and initialize a reference processor 298 _ref_processor = 299 new ReferenceProcessor(_span, // span 300 (ParallelGCThreads > 1) && ParallelRefProcEnabled, // mt processing 301 (int) ParallelGCThreads, // mt processing degree 302 _cmsGen->refs_discovery_is_mt(), // mt discovery 303 (int) MAX2(ConcGCThreads, ParallelGCThreads), // mt discovery degree 304 _cmsGen->refs_discovery_is_atomic(), // discovery is not atomic 305 &_is_alive_closure, // closure for liveness info 306 false); // next field updates do not need write barrier 307 // Initialize the _ref_processor field of CMSGen 308 _cmsGen->set_ref_processor(_ref_processor); 309 310 } 311 } 312 313 CMSAdaptiveSizePolicy* CMSCollector::size_policy() { 314 GenCollectedHeap* gch = GenCollectedHeap::heap(); 315 assert(gch->kind() == CollectedHeap::GenCollectedHeap, 316 "Wrong type of heap"); 317 CMSAdaptiveSizePolicy* sp = (CMSAdaptiveSizePolicy*) 318 gch->gen_policy()->size_policy(); 319 assert(sp->is_gc_cms_adaptive_size_policy(), 320 "Wrong type of size policy"); 321 return sp; 322 } 323 324 CMSGCAdaptivePolicyCounters* CMSCollector::gc_adaptive_policy_counters() { 325 CMSGCAdaptivePolicyCounters* results = 326 (CMSGCAdaptivePolicyCounters*) collector_policy()->counters(); 327 assert( 328 results->kind() == GCPolicyCounters::CMSGCAdaptivePolicyCountersKind, 329 "Wrong gc policy counter kind"); 330 return results; 331 } 332 333 334 void ConcurrentMarkSweepGeneration::initialize_performance_counters() { 335 336 const char* gen_name = "old"; 337 338 // Generation Counters - generation 1, 1 subspace 339 _gen_counters = new GenerationCounters(gen_name, 1, 1, &_virtual_space); 340 341 _space_counters = new GSpaceCounters(gen_name, 0, 342 _virtual_space.reserved_size(), 343 this, _gen_counters); 344 } 345 346 CMSStats::CMSStats(ConcurrentMarkSweepGeneration* cms_gen, unsigned int alpha): 347 _cms_gen(cms_gen) 348 { 349 assert(alpha <= 100, "bad value"); 350 _saved_alpha = alpha; 351 352 // Initialize the alphas to the bootstrap value of 100. 353 _gc0_alpha = _cms_alpha = 100; 354 355 _cms_begin_time.update(); 356 _cms_end_time.update(); 357 358 _gc0_duration = 0.0; 359 _gc0_period = 0.0; 360 _gc0_promoted = 0; 361 362 _cms_duration = 0.0; 363 _cms_period = 0.0; 364 _cms_allocated = 0; 365 366 _cms_used_at_gc0_begin = 0; 367 _cms_used_at_gc0_end = 0; 368 _allow_duty_cycle_reduction = false; 369 _valid_bits = 0; 370 _icms_duty_cycle = CMSIncrementalDutyCycle; 371 } 372 373 double CMSStats::cms_free_adjustment_factor(size_t free) const { 374 // TBD: CR 6909490 375 return 1.0; 376 } 377 378 void CMSStats::adjust_cms_free_adjustment_factor(bool fail, size_t free) { 379 } 380 381 // If promotion failure handling is on use 382 // the padded average size of the promotion for each 383 // young generation collection. 384 double CMSStats::time_until_cms_gen_full() const { 385 size_t cms_free = _cms_gen->cmsSpace()->free(); 386 GenCollectedHeap* gch = GenCollectedHeap::heap(); 387 size_t expected_promotion = MIN2(gch->get_gen(0)->capacity(), 388 (size_t) _cms_gen->gc_stats()->avg_promoted()->padded_average()); 389 if (cms_free > expected_promotion) { 390 // Start a cms collection if there isn't enough space to promote 391 // for the next minor collection. Use the padded average as 392 // a safety factor. 393 cms_free -= expected_promotion; 394 395 // Adjust by the safety factor. 396 double cms_free_dbl = (double)cms_free; 397 double cms_adjustment = (100.0 - CMSIncrementalSafetyFactor)/100.0; 398 // Apply a further correction factor which tries to adjust 399 // for recent occurance of concurrent mode failures. 400 cms_adjustment = cms_adjustment * cms_free_adjustment_factor(cms_free); 401 cms_free_dbl = cms_free_dbl * cms_adjustment; 402 403 if (PrintGCDetails && Verbose) { 404 gclog_or_tty->print_cr("CMSStats::time_until_cms_gen_full: cms_free " 405 SIZE_FORMAT " expected_promotion " SIZE_FORMAT, 406 cms_free, expected_promotion); 407 gclog_or_tty->print_cr(" cms_free_dbl %f cms_consumption_rate %f", 408 cms_free_dbl, cms_consumption_rate() + 1.0); 409 } 410 // Add 1 in case the consumption rate goes to zero. 411 return cms_free_dbl / (cms_consumption_rate() + 1.0); 412 } 413 return 0.0; 414 } 415 416 // Compare the duration of the cms collection to the 417 // time remaining before the cms generation is empty. 418 // Note that the time from the start of the cms collection 419 // to the start of the cms sweep (less than the total 420 // duration of the cms collection) can be used. This 421 // has been tried and some applications experienced 422 // promotion failures early in execution. This was 423 // possibly because the averages were not accurate 424 // enough at the beginning. 425 double CMSStats::time_until_cms_start() const { 426 // We add "gc0_period" to the "work" calculation 427 // below because this query is done (mostly) at the 428 // end of a scavenge, so we need to conservatively 429 // account for that much possible delay 430 // in the query so as to avoid concurrent mode failures 431 // due to starting the collection just a wee bit too 432 // late. 433 double work = cms_duration() + gc0_period(); 434 double deadline = time_until_cms_gen_full(); 435 // If a concurrent mode failure occurred recently, we want to be 436 // more conservative and halve our expected time_until_cms_gen_full() 437 if (work > deadline) { 438 if (Verbose && PrintGCDetails) { 439 gclog_or_tty->print( 440 " CMSCollector: collect because of anticipated promotion " 441 "before full %3.7f + %3.7f > %3.7f ", cms_duration(), 442 gc0_period(), time_until_cms_gen_full()); 443 } 444 return 0.0; 445 } 446 return work - deadline; 447 } 448 449 // Return a duty cycle based on old_duty_cycle and new_duty_cycle, limiting the 450 // amount of change to prevent wild oscillation. 451 unsigned int CMSStats::icms_damped_duty_cycle(unsigned int old_duty_cycle, 452 unsigned int new_duty_cycle) { 453 assert(old_duty_cycle <= 100, "bad input value"); 454 assert(new_duty_cycle <= 100, "bad input value"); 455 456 // Note: use subtraction with caution since it may underflow (values are 457 // unsigned). Addition is safe since we're in the range 0-100. 458 unsigned int damped_duty_cycle = new_duty_cycle; 459 if (new_duty_cycle < old_duty_cycle) { 460 const unsigned int largest_delta = MAX2(old_duty_cycle / 4, 5U); 461 if (new_duty_cycle + largest_delta < old_duty_cycle) { 462 damped_duty_cycle = old_duty_cycle - largest_delta; 463 } 464 } else if (new_duty_cycle > old_duty_cycle) { 465 const unsigned int largest_delta = MAX2(old_duty_cycle / 4, 15U); 466 if (new_duty_cycle > old_duty_cycle + largest_delta) { 467 damped_duty_cycle = MIN2(old_duty_cycle + largest_delta, 100U); 468 } 469 } 470 assert(damped_duty_cycle <= 100, "invalid duty cycle computed"); 471 472 if (CMSTraceIncrementalPacing) { 473 gclog_or_tty->print(" [icms_damped_duty_cycle(%d,%d) = %d] ", 474 old_duty_cycle, new_duty_cycle, damped_duty_cycle); 475 } 476 return damped_duty_cycle; 477 } 478 479 unsigned int CMSStats::icms_update_duty_cycle_impl() { 480 assert(CMSIncrementalPacing && valid(), 481 "should be handled in icms_update_duty_cycle()"); 482 483 double cms_time_so_far = cms_timer().seconds(); 484 double scaled_duration = cms_duration_per_mb() * _cms_used_at_gc0_end / M; 485 double scaled_duration_remaining = fabsd(scaled_duration - cms_time_so_far); 486 487 // Avoid division by 0. 488 double time_until_full = MAX2(time_until_cms_gen_full(), 0.01); 489 double duty_cycle_dbl = 100.0 * scaled_duration_remaining / time_until_full; 490 491 unsigned int new_duty_cycle = MIN2((unsigned int)duty_cycle_dbl, 100U); 492 if (new_duty_cycle > _icms_duty_cycle) { 493 // Avoid very small duty cycles (1 or 2); 0 is allowed. 494 if (new_duty_cycle > 2) { 495 _icms_duty_cycle = icms_damped_duty_cycle(_icms_duty_cycle, 496 new_duty_cycle); 497 } 498 } else if (_allow_duty_cycle_reduction) { 499 // The duty cycle is reduced only once per cms cycle (see record_cms_end()). 500 new_duty_cycle = icms_damped_duty_cycle(_icms_duty_cycle, new_duty_cycle); 501 // Respect the minimum duty cycle. 502 unsigned int min_duty_cycle = (unsigned int)CMSIncrementalDutyCycleMin; 503 _icms_duty_cycle = MAX2(new_duty_cycle, min_duty_cycle); 504 } 505 506 if (PrintGCDetails || CMSTraceIncrementalPacing) { 507 gclog_or_tty->print(" icms_dc=%d ", _icms_duty_cycle); 508 } 509 510 _allow_duty_cycle_reduction = false; 511 return _icms_duty_cycle; 512 } 513 514 #ifndef PRODUCT 515 void CMSStats::print_on(outputStream *st) const { 516 st->print(" gc0_alpha=%d,cms_alpha=%d", _gc0_alpha, _cms_alpha); 517 st->print(",gc0_dur=%g,gc0_per=%g,gc0_promo=" SIZE_FORMAT, 518 gc0_duration(), gc0_period(), gc0_promoted()); 519 st->print(",cms_dur=%g,cms_dur_per_mb=%g,cms_per=%g,cms_alloc=" SIZE_FORMAT, 520 cms_duration(), cms_duration_per_mb(), 521 cms_period(), cms_allocated()); 522 st->print(",cms_since_beg=%g,cms_since_end=%g", 523 cms_time_since_begin(), cms_time_since_end()); 524 st->print(",cms_used_beg=" SIZE_FORMAT ",cms_used_end=" SIZE_FORMAT, 525 _cms_used_at_gc0_begin, _cms_used_at_gc0_end); 526 if (CMSIncrementalMode) { 527 st->print(",dc=%d", icms_duty_cycle()); 528 } 529 530 if (valid()) { 531 st->print(",promo_rate=%g,cms_alloc_rate=%g", 532 promotion_rate(), cms_allocation_rate()); 533 st->print(",cms_consumption_rate=%g,time_until_full=%g", 534 cms_consumption_rate(), time_until_cms_gen_full()); 535 } 536 st->print(" "); 537 } 538 #endif // #ifndef PRODUCT 539 540 CMSCollector::CollectorState CMSCollector::_collectorState = 541 CMSCollector::Idling; 542 bool CMSCollector::_foregroundGCIsActive = false; 543 bool CMSCollector::_foregroundGCShouldWait = false; 544 545 CMSCollector::CMSCollector(ConcurrentMarkSweepGeneration* cmsGen, 546 CardTableRS* ct, 547 ConcurrentMarkSweepPolicy* cp): 548 _cmsGen(cmsGen), 549 _ct(ct), 550 _ref_processor(NULL), // will be set later 551 _conc_workers(NULL), // may be set later 552 _abort_preclean(false), 553 _start_sampling(false), 554 _between_prologue_and_epilogue(false), 555 _markBitMap(0, Mutex::leaf + 1, "CMS_markBitMap_lock"), 556 _modUnionTable((CardTableModRefBS::card_shift - LogHeapWordSize), 557 -1 /* lock-free */, "No_lock" /* dummy */), 558 _modUnionClosure(&_modUnionTable), 559 _modUnionClosurePar(&_modUnionTable), 560 // Adjust my span to cover old (cms) gen 561 _span(cmsGen->reserved()), 562 // Construct the is_alive_closure with _span & markBitMap 563 _is_alive_closure(_span, &_markBitMap), 564 _restart_addr(NULL), 565 _overflow_list(NULL), 566 _stats(cmsGen), 567 _eden_chunk_array(NULL), // may be set in ctor body 568 _eden_chunk_capacity(0), // -- ditto -- 569 _eden_chunk_index(0), // -- ditto -- 570 _survivor_plab_array(NULL), // -- ditto -- 571 _survivor_chunk_array(NULL), // -- ditto -- 572 _survivor_chunk_capacity(0), // -- ditto -- 573 _survivor_chunk_index(0), // -- ditto -- 574 _ser_pmc_preclean_ovflw(0), 575 _ser_kac_preclean_ovflw(0), 576 _ser_pmc_remark_ovflw(0), 577 _par_pmc_remark_ovflw(0), 578 _ser_kac_ovflw(0), 579 _par_kac_ovflw(0), 580 #ifndef PRODUCT 581 _num_par_pushes(0), 582 #endif 583 _collection_count_start(0), 584 _verifying(false), 585 _icms_start_limit(NULL), 586 _icms_stop_limit(NULL), 587 _verification_mark_bm(0, Mutex::leaf + 1, "CMS_verification_mark_bm_lock"), 588 _completed_initialization(false), 589 _collector_policy(cp), 590 _should_unload_classes(false), 591 _concurrent_cycles_since_last_unload(0), 592 _roots_scanning_options(0), 593 _inter_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding), 594 _intra_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding) 595 { 596 if (ExplicitGCInvokesConcurrentAndUnloadsClasses) { 597 ExplicitGCInvokesConcurrent = true; 598 } 599 // Now expand the span and allocate the collection support structures 600 // (MUT, marking bit map etc.) to cover both generations subject to 601 // collection. 602 603 // For use by dirty card to oop closures. 604 _cmsGen->cmsSpace()->set_collector(this); 605 606 // Allocate MUT and marking bit map 607 { 608 MutexLockerEx x(_markBitMap.lock(), Mutex::_no_safepoint_check_flag); 609 if (!_markBitMap.allocate(_span)) { 610 warning("Failed to allocate CMS Bit Map"); 611 return; 612 } 613 assert(_markBitMap.covers(_span), "_markBitMap inconsistency?"); 614 } 615 { 616 _modUnionTable.allocate(_span); 617 assert(_modUnionTable.covers(_span), "_modUnionTable inconsistency?"); 618 } 619 620 if (!_markStack.allocate(MarkStackSize)) { 621 warning("Failed to allocate CMS Marking Stack"); 622 return; 623 } 624 625 // Support for multi-threaded concurrent phases 626 if (CMSConcurrentMTEnabled) { 627 if (FLAG_IS_DEFAULT(ConcGCThreads)) { 628 // just for now 629 FLAG_SET_DEFAULT(ConcGCThreads, (ParallelGCThreads + 3)/4); 630 } 631 if (ConcGCThreads > 1) { 632 _conc_workers = new YieldingFlexibleWorkGang("Parallel CMS Threads", 633 ConcGCThreads, true); 634 if (_conc_workers == NULL) { 635 warning("GC/CMS: _conc_workers allocation failure: " 636 "forcing -CMSConcurrentMTEnabled"); 637 CMSConcurrentMTEnabled = false; 638 } else { 639 _conc_workers->initialize_workers(); 640 } 641 } else { 642 CMSConcurrentMTEnabled = false; 643 } 644 } 645 if (!CMSConcurrentMTEnabled) { 646 ConcGCThreads = 0; 647 } else { 648 // Turn off CMSCleanOnEnter optimization temporarily for 649 // the MT case where it's not fixed yet; see 6178663. 650 CMSCleanOnEnter = false; 651 } 652 assert((_conc_workers != NULL) == (ConcGCThreads > 1), 653 "Inconsistency"); 654 655 // Parallel task queues; these are shared for the 656 // concurrent and stop-world phases of CMS, but 657 // are not shared with parallel scavenge (ParNew). 658 { 659 uint i; 660 uint num_queues = (uint) MAX2(ParallelGCThreads, ConcGCThreads); 661 662 if ((CMSParallelRemarkEnabled || CMSConcurrentMTEnabled 663 || ParallelRefProcEnabled) 664 && num_queues > 0) { 665 _task_queues = new OopTaskQueueSet(num_queues); 666 if (_task_queues == NULL) { 667 warning("task_queues allocation failure."); 668 return; 669 } 670 _hash_seed = NEW_C_HEAP_ARRAY(int, num_queues, mtGC); 671 if (_hash_seed == NULL) { 672 warning("_hash_seed array allocation failure"); 673 return; 674 } 675 676 typedef Padded<OopTaskQueue> PaddedOopTaskQueue; 677 for (i = 0; i < num_queues; i++) { 678 PaddedOopTaskQueue *q = new PaddedOopTaskQueue(); 679 if (q == NULL) { 680 warning("work_queue allocation failure."); 681 return; 682 } 683 _task_queues->register_queue(i, q); 684 } 685 for (i = 0; i < num_queues; i++) { 686 _task_queues->queue(i)->initialize(); 687 _hash_seed[i] = 17; // copied from ParNew 688 } 689 } 690 } 691 692 _cmsGen ->init_initiating_occupancy(CMSInitiatingOccupancyFraction, CMSTriggerRatio); 693 694 // Clip CMSBootstrapOccupancy between 0 and 100. 695 _bootstrap_occupancy = ((double)CMSBootstrapOccupancy)/(double)100; 696 697 _full_gcs_since_conc_gc = 0; 698 699 // Now tell CMS generations the identity of their collector 700 ConcurrentMarkSweepGeneration::set_collector(this); 701 702 // Create & start a CMS thread for this CMS collector 703 _cmsThread = ConcurrentMarkSweepThread::start(this); 704 assert(cmsThread() != NULL, "CMS Thread should have been created"); 705 assert(cmsThread()->collector() == this, 706 "CMS Thread should refer to this gen"); 707 assert(CGC_lock != NULL, "Where's the CGC_lock?"); 708 709 // Support for parallelizing young gen rescan 710 GenCollectedHeap* gch = GenCollectedHeap::heap(); 711 _young_gen = gch->prev_gen(_cmsGen); 712 if (gch->supports_inline_contig_alloc()) { 713 _top_addr = gch->top_addr(); 714 _end_addr = gch->end_addr(); 715 assert(_young_gen != NULL, "no _young_gen"); 716 _eden_chunk_index = 0; 717 _eden_chunk_capacity = (_young_gen->max_capacity()+CMSSamplingGrain)/CMSSamplingGrain; 718 _eden_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, _eden_chunk_capacity, mtGC); 719 if (_eden_chunk_array == NULL) { 720 _eden_chunk_capacity = 0; 721 warning("GC/CMS: _eden_chunk_array allocation failure"); 722 } 723 } 724 assert(_eden_chunk_array != NULL || _eden_chunk_capacity == 0, "Error"); 725 726 // Support for parallelizing survivor space rescan 727 if (CMSParallelRemarkEnabled && CMSParallelSurvivorRemarkEnabled) { 728 const size_t max_plab_samples = 729 ((DefNewGeneration*)_young_gen)->max_survivor_size()/MinTLABSize; 730 731 _survivor_plab_array = NEW_C_HEAP_ARRAY(ChunkArray, ParallelGCThreads, mtGC); 732 _survivor_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, 2*max_plab_samples, mtGC); 733 _cursor = NEW_C_HEAP_ARRAY(size_t, ParallelGCThreads, mtGC); 734 if (_survivor_plab_array == NULL || _survivor_chunk_array == NULL 735 || _cursor == NULL) { 736 warning("Failed to allocate survivor plab/chunk array"); 737 if (_survivor_plab_array != NULL) { 738 FREE_C_HEAP_ARRAY(ChunkArray, _survivor_plab_array, mtGC); 739 _survivor_plab_array = NULL; 740 } 741 if (_survivor_chunk_array != NULL) { 742 FREE_C_HEAP_ARRAY(HeapWord*, _survivor_chunk_array, mtGC); 743 _survivor_chunk_array = NULL; 744 } 745 if (_cursor != NULL) { 746 FREE_C_HEAP_ARRAY(size_t, _cursor, mtGC); 747 _cursor = NULL; 748 } 749 } else { 750 _survivor_chunk_capacity = 2*max_plab_samples; 751 for (uint i = 0; i < ParallelGCThreads; i++) { 752 HeapWord** vec = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples, mtGC); 753 if (vec == NULL) { 754 warning("Failed to allocate survivor plab array"); 755 for (int j = i; j > 0; j--) { 756 FREE_C_HEAP_ARRAY(HeapWord*, _survivor_plab_array[j-1].array(), mtGC); 757 } 758 FREE_C_HEAP_ARRAY(ChunkArray, _survivor_plab_array, mtGC); 759 FREE_C_HEAP_ARRAY(HeapWord*, _survivor_chunk_array, mtGC); 760 _survivor_plab_array = NULL; 761 _survivor_chunk_array = NULL; 762 _survivor_chunk_capacity = 0; 763 break; 764 } else { 765 ChunkArray* cur = 766 ::new (&_survivor_plab_array[i]) ChunkArray(vec, 767 max_plab_samples); 768 assert(cur->end() == 0, "Should be 0"); 769 assert(cur->array() == vec, "Should be vec"); 770 assert(cur->capacity() == max_plab_samples, "Error"); 771 } 772 } 773 } 774 } 775 assert( ( _survivor_plab_array != NULL 776 && _survivor_chunk_array != NULL) 777 || ( _survivor_chunk_capacity == 0 778 && _survivor_chunk_index == 0), 779 "Error"); 780 781 // Choose what strong roots should be scanned depending on verification options 782 if (!CMSClassUnloadingEnabled) { 783 // If class unloading is disabled we want to include all classes into the root set. 784 add_root_scanning_option(SharedHeap::SO_AllClasses); 785 } else { 786 add_root_scanning_option(SharedHeap::SO_SystemClasses); 787 } 788 789 NOT_PRODUCT(_overflow_counter = CMSMarkStackOverflowInterval;) 790 _gc_counters = new CollectorCounters("CMS", 1); 791 _completed_initialization = true; 792 _inter_sweep_timer.start(); // start of time 793 } 794 795 const char* ConcurrentMarkSweepGeneration::name() const { 796 return "concurrent mark-sweep generation"; 797 } 798 void ConcurrentMarkSweepGeneration::update_counters() { 799 if (UsePerfData) { 800 _space_counters->update_all(); 801 _gen_counters->update_all(); 802 } 803 } 804 805 // this is an optimized version of update_counters(). it takes the 806 // used value as a parameter rather than computing it. 807 // 808 void ConcurrentMarkSweepGeneration::update_counters(size_t used) { 809 if (UsePerfData) { 810 _space_counters->update_used(used); 811 _space_counters->update_capacity(); 812 _gen_counters->update_all(); 813 } 814 } 815 816 void ConcurrentMarkSweepGeneration::print() const { 817 Generation::print(); 818 cmsSpace()->print(); 819 } 820 821 #ifndef PRODUCT 822 void ConcurrentMarkSweepGeneration::print_statistics() { 823 cmsSpace()->printFLCensus(0); 824 } 825 #endif 826 827 void ConcurrentMarkSweepGeneration::printOccupancy(const char *s) { 828 GenCollectedHeap* gch = GenCollectedHeap::heap(); 829 if (PrintGCDetails) { 830 if (Verbose) { 831 gclog_or_tty->print("[%d %s-%s: "SIZE_FORMAT"("SIZE_FORMAT")]", 832 level(), short_name(), s, used(), capacity()); 833 } else { 834 gclog_or_tty->print("[%d %s-%s: "SIZE_FORMAT"K("SIZE_FORMAT"K)]", 835 level(), short_name(), s, used() / K, capacity() / K); 836 } 837 } 838 if (Verbose) { 839 gclog_or_tty->print(" "SIZE_FORMAT"("SIZE_FORMAT")", 840 gch->used(), gch->capacity()); 841 } else { 842 gclog_or_tty->print(" "SIZE_FORMAT"K("SIZE_FORMAT"K)", 843 gch->used() / K, gch->capacity() / K); 844 } 845 } 846 847 size_t 848 ConcurrentMarkSweepGeneration::contiguous_available() const { 849 // dld proposes an improvement in precision here. If the committed 850 // part of the space ends in a free block we should add that to 851 // uncommitted size in the calculation below. Will make this 852 // change later, staying with the approximation below for the 853 // time being. -- ysr. 854 return MAX2(_virtual_space.uncommitted_size(), unsafe_max_alloc_nogc()); 855 } 856 857 size_t 858 ConcurrentMarkSweepGeneration::unsafe_max_alloc_nogc() const { 859 return _cmsSpace->max_alloc_in_words() * HeapWordSize; 860 } 861 862 size_t ConcurrentMarkSweepGeneration::max_available() const { 863 return free() + _virtual_space.uncommitted_size(); 864 } 865 866 bool ConcurrentMarkSweepGeneration::promotion_attempt_is_safe(size_t max_promotion_in_bytes) const { 867 size_t available = max_available(); 868 size_t av_promo = (size_t)gc_stats()->avg_promoted()->padded_average(); 869 bool res = (available >= av_promo) || (available >= max_promotion_in_bytes); 870 if (Verbose && PrintGCDetails) { 871 gclog_or_tty->print_cr( 872 "CMS: promo attempt is%s safe: available("SIZE_FORMAT") %s av_promo("SIZE_FORMAT")," 873 "max_promo("SIZE_FORMAT")", 874 res? "":" not", available, res? ">=":"<", 875 av_promo, max_promotion_in_bytes); 876 } 877 return res; 878 } 879 880 // At a promotion failure dump information on block layout in heap 881 // (cms old generation). 882 void ConcurrentMarkSweepGeneration::promotion_failure_occurred() { 883 if (CMSDumpAtPromotionFailure) { 884 cmsSpace()->dump_at_safepoint_with_locks(collector(), gclog_or_tty); 885 } 886 } 887 888 CompactibleSpace* 889 ConcurrentMarkSweepGeneration::first_compaction_space() const { 890 return _cmsSpace; 891 } 892 893 void ConcurrentMarkSweepGeneration::reset_after_compaction() { 894 // Clear the promotion information. These pointers can be adjusted 895 // along with all the other pointers into the heap but 896 // compaction is expected to be a rare event with 897 // a heap using cms so don't do it without seeing the need. 898 if (CollectedHeap::use_parallel_gc_threads()) { 899 for (uint i = 0; i < ParallelGCThreads; i++) { 900 _par_gc_thread_states[i]->promo.reset(); 901 } 902 } 903 } 904 905 void ConcurrentMarkSweepGeneration::space_iterate(SpaceClosure* blk, bool usedOnly) { 906 blk->do_space(_cmsSpace); 907 } 908 909 void ConcurrentMarkSweepGeneration::compute_new_size() { 910 assert_locked_or_safepoint(Heap_lock); 911 912 // If incremental collection failed, we just want to expand 913 // to the limit. 914 if (incremental_collection_failed()) { 915 clear_incremental_collection_failed(); 916 grow_to_reserved(); 917 return; 918 } 919 920 // The heap has been compacted but not reset yet. 921 // Any metric such as free() or used() will be incorrect. 922 923 CardGeneration::compute_new_size(); 924 925 // Reset again after a possible resizing 926 if (did_compact()) { 927 cmsSpace()->reset_after_compaction(); 928 } 929 } 930 931 void ConcurrentMarkSweepGeneration::compute_new_size_free_list() { 932 assert_locked_or_safepoint(Heap_lock); 933 934 // If incremental collection failed, we just want to expand 935 // to the limit. 936 if (incremental_collection_failed()) { 937 clear_incremental_collection_failed(); 938 grow_to_reserved(); 939 return; 940 } 941 942 double free_percentage = ((double) free()) / capacity(); 943 double desired_free_percentage = (double) MinHeapFreeRatio / 100; 944 double maximum_free_percentage = (double) MaxHeapFreeRatio / 100; 945 946 // compute expansion delta needed for reaching desired free percentage 947 if (free_percentage < desired_free_percentage) { 948 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 949 assert(desired_capacity >= capacity(), "invalid expansion size"); 950 size_t expand_bytes = MAX2(desired_capacity - capacity(), MinHeapDeltaBytes); 951 if (PrintGCDetails && Verbose) { 952 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 953 gclog_or_tty->print_cr("\nFrom compute_new_size: "); 954 gclog_or_tty->print_cr(" Free fraction %f", free_percentage); 955 gclog_or_tty->print_cr(" Desired free fraction %f", 956 desired_free_percentage); 957 gclog_or_tty->print_cr(" Maximum free fraction %f", 958 maximum_free_percentage); 959 gclog_or_tty->print_cr(" Capactiy "SIZE_FORMAT, capacity()/1000); 960 gclog_or_tty->print_cr(" Desired capacity "SIZE_FORMAT, 961 desired_capacity/1000); 962 int prev_level = level() - 1; 963 if (prev_level >= 0) { 964 size_t prev_size = 0; 965 GenCollectedHeap* gch = GenCollectedHeap::heap(); 966 Generation* prev_gen = gch->_gens[prev_level]; 967 prev_size = prev_gen->capacity(); 968 gclog_or_tty->print_cr(" Younger gen size "SIZE_FORMAT, 969 prev_size/1000); 970 } 971 gclog_or_tty->print_cr(" unsafe_max_alloc_nogc "SIZE_FORMAT, 972 unsafe_max_alloc_nogc()/1000); 973 gclog_or_tty->print_cr(" contiguous available "SIZE_FORMAT, 974 contiguous_available()/1000); 975 gclog_or_tty->print_cr(" Expand by "SIZE_FORMAT" (bytes)", 976 expand_bytes); 977 } 978 // safe if expansion fails 979 expand(expand_bytes, 0, CMSExpansionCause::_satisfy_free_ratio); 980 if (PrintGCDetails && Verbose) { 981 gclog_or_tty->print_cr(" Expanded free fraction %f", 982 ((double) free()) / capacity()); 983 } 984 } else { 985 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 986 assert(desired_capacity <= capacity(), "invalid expansion size"); 987 size_t shrink_bytes = capacity() - desired_capacity; 988 // Don't shrink unless the delta is greater than the minimum shrink we want 989 if (shrink_bytes >= MinHeapDeltaBytes) { 990 shrink_free_list_by(shrink_bytes); 991 } 992 } 993 } 994 995 Mutex* ConcurrentMarkSweepGeneration::freelistLock() const { 996 return cmsSpace()->freelistLock(); 997 } 998 999 HeapWord* ConcurrentMarkSweepGeneration::allocate(size_t size, 1000 bool tlab) { 1001 CMSSynchronousYieldRequest yr; 1002 MutexLockerEx x(freelistLock(), 1003 Mutex::_no_safepoint_check_flag); 1004 return have_lock_and_allocate(size, tlab); 1005 } 1006 1007 HeapWord* ConcurrentMarkSweepGeneration::have_lock_and_allocate(size_t size, 1008 bool tlab /* ignored */) { 1009 assert_lock_strong(freelistLock()); 1010 size_t adjustedSize = CompactibleFreeListSpace::adjustObjectSize(size); 1011 HeapWord* res = cmsSpace()->allocate(adjustedSize); 1012 // Allocate the object live (grey) if the background collector has 1013 // started marking. This is necessary because the marker may 1014 // have passed this address and consequently this object will 1015 // not otherwise be greyed and would be incorrectly swept up. 1016 // Note that if this object contains references, the writing 1017 // of those references will dirty the card containing this object 1018 // allowing the object to be blackened (and its references scanned) 1019 // either during a preclean phase or at the final checkpoint. 1020 if (res != NULL) { 1021 // We may block here with an uninitialized object with 1022 // its mark-bit or P-bits not yet set. Such objects need 1023 // to be safely navigable by block_start(). 1024 assert(oop(res)->klass_or_null() == NULL, "Object should be uninitialized here."); 1025 assert(!((FreeChunk*)res)->is_free(), "Error, block will look free but show wrong size"); 1026 collector()->direct_allocated(res, adjustedSize); 1027 _direct_allocated_words += adjustedSize; 1028 // allocation counters 1029 NOT_PRODUCT( 1030 _numObjectsAllocated++; 1031 _numWordsAllocated += (int)adjustedSize; 1032 ) 1033 } 1034 return res; 1035 } 1036 1037 // In the case of direct allocation by mutators in a generation that 1038 // is being concurrently collected, the object must be allocated 1039 // live (grey) if the background collector has started marking. 1040 // This is necessary because the marker may 1041 // have passed this address and consequently this object will 1042 // not otherwise be greyed and would be incorrectly swept up. 1043 // Note that if this object contains references, the writing 1044 // of those references will dirty the card containing this object 1045 // allowing the object to be blackened (and its references scanned) 1046 // either during a preclean phase or at the final checkpoint. 1047 void CMSCollector::direct_allocated(HeapWord* start, size_t size) { 1048 assert(_markBitMap.covers(start, size), "Out of bounds"); 1049 if (_collectorState >= Marking) { 1050 MutexLockerEx y(_markBitMap.lock(), 1051 Mutex::_no_safepoint_check_flag); 1052 // [see comments preceding SweepClosure::do_blk() below for details] 1053 // 1054 // Can the P-bits be deleted now? JJJ 1055 // 1056 // 1. need to mark the object as live so it isn't collected 1057 // 2. need to mark the 2nd bit to indicate the object may be uninitialized 1058 // 3. need to mark the end of the object so marking, precleaning or sweeping 1059 // can skip over uninitialized or unparsable objects. An allocated 1060 // object is considered uninitialized for our purposes as long as 1061 // its klass word is NULL. All old gen objects are parsable 1062 // as soon as they are initialized.) 1063 _markBitMap.mark(start); // object is live 1064 _markBitMap.mark(start + 1); // object is potentially uninitialized? 1065 _markBitMap.mark(start + size - 1); 1066 // mark end of object 1067 } 1068 // check that oop looks uninitialized 1069 assert(oop(start)->klass_or_null() == NULL, "_klass should be NULL"); 1070 } 1071 1072 void CMSCollector::promoted(bool par, HeapWord* start, 1073 bool is_obj_array, size_t obj_size) { 1074 assert(_markBitMap.covers(start), "Out of bounds"); 1075 // See comment in direct_allocated() about when objects should 1076 // be allocated live. 1077 if (_collectorState >= Marking) { 1078 // we already hold the marking bit map lock, taken in 1079 // the prologue 1080 if (par) { 1081 _markBitMap.par_mark(start); 1082 } else { 1083 _markBitMap.mark(start); 1084 } 1085 // We don't need to mark the object as uninitialized (as 1086 // in direct_allocated above) because this is being done with the 1087 // world stopped and the object will be initialized by the 1088 // time the marking, precleaning or sweeping get to look at it. 1089 // But see the code for copying objects into the CMS generation, 1090 // where we need to ensure that concurrent readers of the 1091 // block offset table are able to safely navigate a block that 1092 // is in flux from being free to being allocated (and in 1093 // transition while being copied into) and subsequently 1094 // becoming a bona-fide object when the copy/promotion is complete. 1095 assert(SafepointSynchronize::is_at_safepoint(), 1096 "expect promotion only at safepoints"); 1097 1098 if (_collectorState < Sweeping) { 1099 // Mark the appropriate cards in the modUnionTable, so that 1100 // this object gets scanned before the sweep. If this is 1101 // not done, CMS generation references in the object might 1102 // not get marked. 1103 // For the case of arrays, which are otherwise precisely 1104 // marked, we need to dirty the entire array, not just its head. 1105 if (is_obj_array) { 1106 // The [par_]mark_range() method expects mr.end() below to 1107 // be aligned to the granularity of a bit's representation 1108 // in the heap. In the case of the MUT below, that's a 1109 // card size. 1110 MemRegion mr(start, 1111 (HeapWord*)round_to((intptr_t)(start + obj_size), 1112 CardTableModRefBS::card_size /* bytes */)); 1113 if (par) { 1114 _modUnionTable.par_mark_range(mr); 1115 } else { 1116 _modUnionTable.mark_range(mr); 1117 } 1118 } else { // not an obj array; we can just mark the head 1119 if (par) { 1120 _modUnionTable.par_mark(start); 1121 } else { 1122 _modUnionTable.mark(start); 1123 } 1124 } 1125 } 1126 } 1127 } 1128 1129 static inline size_t percent_of_space(Space* space, HeapWord* addr) 1130 { 1131 size_t delta = pointer_delta(addr, space->bottom()); 1132 return (size_t)(delta * 100.0 / (space->capacity() / HeapWordSize)); 1133 } 1134 1135 void CMSCollector::icms_update_allocation_limits() 1136 { 1137 Generation* gen0 = GenCollectedHeap::heap()->get_gen(0); 1138 EdenSpace* eden = gen0->as_DefNewGeneration()->eden(); 1139 1140 const unsigned int duty_cycle = stats().icms_update_duty_cycle(); 1141 if (CMSTraceIncrementalPacing) { 1142 stats().print(); 1143 } 1144 1145 assert(duty_cycle <= 100, "invalid duty cycle"); 1146 if (duty_cycle != 0) { 1147 // The duty_cycle is a percentage between 0 and 100; convert to words and 1148 // then compute the offset from the endpoints of the space. 1149 size_t free_words = eden->free() / HeapWordSize; 1150 double free_words_dbl = (double)free_words; 1151 size_t duty_cycle_words = (size_t)(free_words_dbl * duty_cycle / 100.0); 1152 size_t offset_words = (free_words - duty_cycle_words) / 2; 1153 1154 _icms_start_limit = eden->top() + offset_words; 1155 _icms_stop_limit = eden->end() - offset_words; 1156 1157 // The limits may be adjusted (shifted to the right) by 1158 // CMSIncrementalOffset, to allow the application more mutator time after a 1159 // young gen gc (when all mutators were stopped) and before CMS starts and 1160 // takes away one or more cpus. 1161 if (CMSIncrementalOffset != 0) { 1162 double adjustment_dbl = free_words_dbl * CMSIncrementalOffset / 100.0; 1163 size_t adjustment = (size_t)adjustment_dbl; 1164 HeapWord* tmp_stop = _icms_stop_limit + adjustment; 1165 if (tmp_stop > _icms_stop_limit && tmp_stop < eden->end()) { 1166 _icms_start_limit += adjustment; 1167 _icms_stop_limit = tmp_stop; 1168 } 1169 } 1170 } 1171 if (duty_cycle == 0 || (_icms_start_limit == _icms_stop_limit)) { 1172 _icms_start_limit = _icms_stop_limit = eden->end(); 1173 } 1174 1175 // Install the new start limit. 1176 eden->set_soft_end(_icms_start_limit); 1177 1178 if (CMSTraceIncrementalMode) { 1179 gclog_or_tty->print(" icms alloc limits: " 1180 PTR_FORMAT "," PTR_FORMAT 1181 " (" SIZE_FORMAT "%%," SIZE_FORMAT "%%) ", 1182 _icms_start_limit, _icms_stop_limit, 1183 percent_of_space(eden, _icms_start_limit), 1184 percent_of_space(eden, _icms_stop_limit)); 1185 if (Verbose) { 1186 gclog_or_tty->print("eden: "); 1187 eden->print_on(gclog_or_tty); 1188 } 1189 } 1190 } 1191 1192 // Any changes here should try to maintain the invariant 1193 // that if this method is called with _icms_start_limit 1194 // and _icms_stop_limit both NULL, then it should return NULL 1195 // and not notify the icms thread. 1196 HeapWord* 1197 CMSCollector::allocation_limit_reached(Space* space, HeapWord* top, 1198 size_t word_size) 1199 { 1200 // A start_limit equal to end() means the duty cycle is 0, so treat that as a 1201 // nop. 1202 if (CMSIncrementalMode && _icms_start_limit != space->end()) { 1203 if (top <= _icms_start_limit) { 1204 if (CMSTraceIncrementalMode) { 1205 space->print_on(gclog_or_tty); 1206 gclog_or_tty->stamp(); 1207 gclog_or_tty->print_cr(" start limit top=" PTR_FORMAT 1208 ", new limit=" PTR_FORMAT 1209 " (" SIZE_FORMAT "%%)", 1210 top, _icms_stop_limit, 1211 percent_of_space(space, _icms_stop_limit)); 1212 } 1213 ConcurrentMarkSweepThread::start_icms(); 1214 assert(top < _icms_stop_limit, "Tautology"); 1215 if (word_size < pointer_delta(_icms_stop_limit, top)) { 1216 return _icms_stop_limit; 1217 } 1218 1219 // The allocation will cross both the _start and _stop limits, so do the 1220 // stop notification also and return end(). 1221 if (CMSTraceIncrementalMode) { 1222 space->print_on(gclog_or_tty); 1223 gclog_or_tty->stamp(); 1224 gclog_or_tty->print_cr(" +stop limit top=" PTR_FORMAT 1225 ", new limit=" PTR_FORMAT 1226 " (" SIZE_FORMAT "%%)", 1227 top, space->end(), 1228 percent_of_space(space, space->end())); 1229 } 1230 ConcurrentMarkSweepThread::stop_icms(); 1231 return space->end(); 1232 } 1233 1234 if (top <= _icms_stop_limit) { 1235 if (CMSTraceIncrementalMode) { 1236 space->print_on(gclog_or_tty); 1237 gclog_or_tty->stamp(); 1238 gclog_or_tty->print_cr(" stop limit top=" PTR_FORMAT 1239 ", new limit=" PTR_FORMAT 1240 " (" SIZE_FORMAT "%%)", 1241 top, space->end(), 1242 percent_of_space(space, space->end())); 1243 } 1244 ConcurrentMarkSweepThread::stop_icms(); 1245 return space->end(); 1246 } 1247 1248 if (CMSTraceIncrementalMode) { 1249 space->print_on(gclog_or_tty); 1250 gclog_or_tty->stamp(); 1251 gclog_or_tty->print_cr(" end limit top=" PTR_FORMAT 1252 ", new limit=" PTR_FORMAT, 1253 top, NULL); 1254 } 1255 } 1256 1257 return NULL; 1258 } 1259 1260 oop ConcurrentMarkSweepGeneration::promote(oop obj, size_t obj_size) { 1261 assert(obj_size == (size_t)obj->size(), "bad obj_size passed in"); 1262 // allocate, copy and if necessary update promoinfo -- 1263 // delegate to underlying space. 1264 assert_lock_strong(freelistLock()); 1265 1266 #ifndef PRODUCT 1267 if (Universe::heap()->promotion_should_fail()) { 1268 return NULL; 1269 } 1270 #endif // #ifndef PRODUCT 1271 1272 oop res = _cmsSpace->promote(obj, obj_size); 1273 if (res == NULL) { 1274 // expand and retry 1275 size_t s = _cmsSpace->expansionSpaceRequired(obj_size); // HeapWords 1276 expand(s*HeapWordSize, MinHeapDeltaBytes, 1277 CMSExpansionCause::_satisfy_promotion); 1278 // Since there's currently no next generation, we don't try to promote 1279 // into a more senior generation. 1280 assert(next_gen() == NULL, "assumption, based upon which no attempt " 1281 "is made to pass on a possibly failing " 1282 "promotion to next generation"); 1283 res = _cmsSpace->promote(obj, obj_size); 1284 } 1285 if (res != NULL) { 1286 // See comment in allocate() about when objects should 1287 // be allocated live. 1288 assert(obj->is_oop(), "Will dereference klass pointer below"); 1289 collector()->promoted(false, // Not parallel 1290 (HeapWord*)res, obj->is_objArray(), obj_size); 1291 // promotion counters 1292 NOT_PRODUCT( 1293 _numObjectsPromoted++; 1294 _numWordsPromoted += 1295 (int)(CompactibleFreeListSpace::adjustObjectSize(obj->size())); 1296 ) 1297 } 1298 return res; 1299 } 1300 1301 1302 HeapWord* 1303 ConcurrentMarkSweepGeneration::allocation_limit_reached(Space* space, 1304 HeapWord* top, 1305 size_t word_sz) 1306 { 1307 return collector()->allocation_limit_reached(space, top, word_sz); 1308 } 1309 1310 // IMPORTANT: Notes on object size recognition in CMS. 1311 // --------------------------------------------------- 1312 // A block of storage in the CMS generation is always in 1313 // one of three states. A free block (FREE), an allocated 1314 // object (OBJECT) whose size() method reports the correct size, 1315 // and an intermediate state (TRANSIENT) in which its size cannot 1316 // be accurately determined. 1317 // STATE IDENTIFICATION: (32 bit and 64 bit w/o COOPS) 1318 // ----------------------------------------------------- 1319 // FREE: klass_word & 1 == 1; mark_word holds block size 1320 // 1321 // OBJECT: klass_word installed; klass_word != 0 && klass_word & 1 == 0; 1322 // obj->size() computes correct size 1323 // 1324 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT 1325 // 1326 // STATE IDENTIFICATION: (64 bit+COOPS) 1327 // ------------------------------------ 1328 // FREE: mark_word & CMS_FREE_BIT == 1; mark_word & ~CMS_FREE_BIT gives block_size 1329 // 1330 // OBJECT: klass_word installed; klass_word != 0; 1331 // obj->size() computes correct size 1332 // 1333 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT 1334 // 1335 // 1336 // STATE TRANSITION DIAGRAM 1337 // 1338 // mut / parnew mut / parnew 1339 // FREE --------------------> TRANSIENT ---------------------> OBJECT --| 1340 // ^ | 1341 // |------------------------ DEAD <------------------------------------| 1342 // sweep mut 1343 // 1344 // While a block is in TRANSIENT state its size cannot be determined 1345 // so readers will either need to come back later or stall until 1346 // the size can be determined. Note that for the case of direct 1347 // allocation, P-bits, when available, may be used to determine the 1348 // size of an object that may not yet have been initialized. 1349 1350 // Things to support parallel young-gen collection. 1351 oop 1352 ConcurrentMarkSweepGeneration::par_promote(int thread_num, 1353 oop old, markOop m, 1354 size_t word_sz) { 1355 #ifndef PRODUCT 1356 if (Universe::heap()->promotion_should_fail()) { 1357 return NULL; 1358 } 1359 #endif // #ifndef PRODUCT 1360 1361 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1362 PromotionInfo* promoInfo = &ps->promo; 1363 // if we are tracking promotions, then first ensure space for 1364 // promotion (including spooling space for saving header if necessary). 1365 // then allocate and copy, then track promoted info if needed. 1366 // When tracking (see PromotionInfo::track()), the mark word may 1367 // be displaced and in this case restoration of the mark word 1368 // occurs in the (oop_since_save_marks_)iterate phase. 1369 if (promoInfo->tracking() && !promoInfo->ensure_spooling_space()) { 1370 // Out of space for allocating spooling buffers; 1371 // try expanding and allocating spooling buffers. 1372 if (!expand_and_ensure_spooling_space(promoInfo)) { 1373 return NULL; 1374 } 1375 } 1376 assert(promoInfo->has_spooling_space(), "Control point invariant"); 1377 const size_t alloc_sz = CompactibleFreeListSpace::adjustObjectSize(word_sz); 1378 HeapWord* obj_ptr = ps->lab.alloc(alloc_sz); 1379 if (obj_ptr == NULL) { 1380 obj_ptr = expand_and_par_lab_allocate(ps, alloc_sz); 1381 if (obj_ptr == NULL) { 1382 return NULL; 1383 } 1384 } 1385 oop obj = oop(obj_ptr); 1386 OrderAccess::storestore(); 1387 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1388 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1389 // IMPORTANT: See note on object initialization for CMS above. 1390 // Otherwise, copy the object. Here we must be careful to insert the 1391 // klass pointer last, since this marks the block as an allocated object. 1392 // Except with compressed oops it's the mark word. 1393 HeapWord* old_ptr = (HeapWord*)old; 1394 // Restore the mark word copied above. 1395 obj->set_mark(m); 1396 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1397 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1398 OrderAccess::storestore(); 1399 1400 if (UseCompressedKlassPointers) { 1401 // Copy gap missed by (aligned) header size calculation below 1402 obj->set_klass_gap(old->klass_gap()); 1403 } 1404 if (word_sz > (size_t)oopDesc::header_size()) { 1405 Copy::aligned_disjoint_words(old_ptr + oopDesc::header_size(), 1406 obj_ptr + oopDesc::header_size(), 1407 word_sz - oopDesc::header_size()); 1408 } 1409 1410 // Now we can track the promoted object, if necessary. We take care 1411 // to delay the transition from uninitialized to full object 1412 // (i.e., insertion of klass pointer) until after, so that it 1413 // atomically becomes a promoted object. 1414 if (promoInfo->tracking()) { 1415 promoInfo->track((PromotedObject*)obj, old->klass()); 1416 } 1417 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1418 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1419 assert(old->is_oop(), "Will use and dereference old klass ptr below"); 1420 1421 // Finally, install the klass pointer (this should be volatile). 1422 OrderAccess::storestore(); 1423 obj->set_klass(old->klass()); 1424 // We should now be able to calculate the right size for this object 1425 assert(obj->is_oop() && obj->size() == (int)word_sz, "Error, incorrect size computed for promoted object"); 1426 1427 collector()->promoted(true, // parallel 1428 obj_ptr, old->is_objArray(), word_sz); 1429 1430 NOT_PRODUCT( 1431 Atomic::inc_ptr(&_numObjectsPromoted); 1432 Atomic::add_ptr(alloc_sz, &_numWordsPromoted); 1433 ) 1434 1435 return obj; 1436 } 1437 1438 void 1439 ConcurrentMarkSweepGeneration:: 1440 par_promote_alloc_undo(int thread_num, 1441 HeapWord* obj, size_t word_sz) { 1442 // CMS does not support promotion undo. 1443 ShouldNotReachHere(); 1444 } 1445 1446 void 1447 ConcurrentMarkSweepGeneration:: 1448 par_promote_alloc_done(int thread_num) { 1449 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1450 ps->lab.retire(thread_num); 1451 } 1452 1453 void 1454 ConcurrentMarkSweepGeneration:: 1455 par_oop_since_save_marks_iterate_done(int thread_num) { 1456 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1457 ParScanWithoutBarrierClosure* dummy_cl = NULL; 1458 ps->promo.promoted_oops_iterate_nv(dummy_cl); 1459 } 1460 1461 bool ConcurrentMarkSweepGeneration::should_collect(bool full, 1462 size_t size, 1463 bool tlab) 1464 { 1465 // We allow a STW collection only if a full 1466 // collection was requested. 1467 return full || should_allocate(size, tlab); // FIX ME !!! 1468 // This and promotion failure handling are connected at the 1469 // hip and should be fixed by untying them. 1470 } 1471 1472 bool CMSCollector::shouldConcurrentCollect() { 1473 if (_full_gc_requested) { 1474 if (Verbose && PrintGCDetails) { 1475 gclog_or_tty->print_cr("CMSCollector: collect because of explicit " 1476 " gc request (or gc_locker)"); 1477 } 1478 return true; 1479 } 1480 1481 // For debugging purposes, change the type of collection. 1482 // If the rotation is not on the concurrent collection 1483 // type, don't start a concurrent collection. 1484 NOT_PRODUCT( 1485 if (RotateCMSCollectionTypes && 1486 (_cmsGen->debug_collection_type() != 1487 ConcurrentMarkSweepGeneration::Concurrent_collection_type)) { 1488 assert(_cmsGen->debug_collection_type() != 1489 ConcurrentMarkSweepGeneration::Unknown_collection_type, 1490 "Bad cms collection type"); 1491 return false; 1492 } 1493 ) 1494 1495 FreelistLocker x(this); 1496 // ------------------------------------------------------------------ 1497 // Print out lots of information which affects the initiation of 1498 // a collection. 1499 if (PrintCMSInitiationStatistics && stats().valid()) { 1500 gclog_or_tty->print("CMSCollector shouldConcurrentCollect: "); 1501 gclog_or_tty->stamp(); 1502 gclog_or_tty->print_cr(""); 1503 stats().print_on(gclog_or_tty); 1504 gclog_or_tty->print_cr("time_until_cms_gen_full %3.7f", 1505 stats().time_until_cms_gen_full()); 1506 gclog_or_tty->print_cr("free="SIZE_FORMAT, _cmsGen->free()); 1507 gclog_or_tty->print_cr("contiguous_available="SIZE_FORMAT, 1508 _cmsGen->contiguous_available()); 1509 gclog_or_tty->print_cr("promotion_rate=%g", stats().promotion_rate()); 1510 gclog_or_tty->print_cr("cms_allocation_rate=%g", stats().cms_allocation_rate()); 1511 gclog_or_tty->print_cr("occupancy=%3.7f", _cmsGen->occupancy()); 1512 gclog_or_tty->print_cr("initiatingOccupancy=%3.7f", _cmsGen->initiating_occupancy()); 1513 gclog_or_tty->print_cr("metadata initialized %d", 1514 MetaspaceGC::should_concurrent_collect()); 1515 } 1516 // ------------------------------------------------------------------ 1517 1518 // If the estimated time to complete a cms collection (cms_duration()) 1519 // is less than the estimated time remaining until the cms generation 1520 // is full, start a collection. 1521 if (!UseCMSInitiatingOccupancyOnly) { 1522 if (stats().valid()) { 1523 if (stats().time_until_cms_start() == 0.0) { 1524 return true; 1525 } 1526 } else { 1527 // We want to conservatively collect somewhat early in order 1528 // to try and "bootstrap" our CMS/promotion statistics; 1529 // this branch will not fire after the first successful CMS 1530 // collection because the stats should then be valid. 1531 if (_cmsGen->occupancy() >= _bootstrap_occupancy) { 1532 if (Verbose && PrintGCDetails) { 1533 gclog_or_tty->print_cr( 1534 " CMSCollector: collect for bootstrapping statistics:" 1535 " occupancy = %f, boot occupancy = %f", _cmsGen->occupancy(), 1536 _bootstrap_occupancy); 1537 } 1538 return true; 1539 } 1540 } 1541 } 1542 1543 // Otherwise, we start a collection cycle if 1544 // old gen want a collection cycle started. Each may use 1545 // an appropriate criterion for making this decision. 1546 // XXX We need to make sure that the gen expansion 1547 // criterion dovetails well with this. XXX NEED TO FIX THIS 1548 if (_cmsGen->should_concurrent_collect()) { 1549 if (Verbose && PrintGCDetails) { 1550 gclog_or_tty->print_cr("CMS old gen initiated"); 1551 } 1552 return true; 1553 } 1554 1555 // We start a collection if we believe an incremental collection may fail; 1556 // this is not likely to be productive in practice because it's probably too 1557 // late anyway. 1558 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1559 assert(gch->collector_policy()->is_two_generation_policy(), 1560 "You may want to check the correctness of the following"); 1561 if (gch->incremental_collection_will_fail(true /* consult_young */)) { 1562 if (Verbose && PrintGCDetails) { 1563 gclog_or_tty->print("CMSCollector: collect because incremental collection will fail "); 1564 } 1565 return true; 1566 } 1567 1568 if (MetaspaceGC::should_concurrent_collect()) { 1569 if (Verbose && PrintGCDetails) { 1570 gclog_or_tty->print("CMSCollector: collect for metadata allocation "); 1571 } 1572 return true; 1573 } 1574 1575 return false; 1576 } 1577 1578 void CMSCollector::set_did_compact(bool v) { _cmsGen->set_did_compact(v); } 1579 1580 // Clear _expansion_cause fields of constituent generations 1581 void CMSCollector::clear_expansion_cause() { 1582 _cmsGen->clear_expansion_cause(); 1583 } 1584 1585 // We should be conservative in starting a collection cycle. To 1586 // start too eagerly runs the risk of collecting too often in the 1587 // extreme. To collect too rarely falls back on full collections, 1588 // which works, even if not optimum in terms of concurrent work. 1589 // As a work around for too eagerly collecting, use the flag 1590 // UseCMSInitiatingOccupancyOnly. This also has the advantage of 1591 // giving the user an easily understandable way of controlling the 1592 // collections. 1593 // We want to start a new collection cycle if any of the following 1594 // conditions hold: 1595 // . our current occupancy exceeds the configured initiating occupancy 1596 // for this generation, or 1597 // . we recently needed to expand this space and have not, since that 1598 // expansion, done a collection of this generation, or 1599 // . the underlying space believes that it may be a good idea to initiate 1600 // a concurrent collection (this may be based on criteria such as the 1601 // following: the space uses linear allocation and linear allocation is 1602 // going to fail, or there is believed to be excessive fragmentation in 1603 // the generation, etc... or ... 1604 // [.(currently done by CMSCollector::shouldConcurrentCollect() only for 1605 // the case of the old generation; see CR 6543076): 1606 // we may be approaching a point at which allocation requests may fail because 1607 // we will be out of sufficient free space given allocation rate estimates.] 1608 bool ConcurrentMarkSweepGeneration::should_concurrent_collect() const { 1609 1610 assert_lock_strong(freelistLock()); 1611 if (occupancy() > initiating_occupancy()) { 1612 if (PrintGCDetails && Verbose) { 1613 gclog_or_tty->print(" %s: collect because of occupancy %f / %f ", 1614 short_name(), occupancy(), initiating_occupancy()); 1615 } 1616 return true; 1617 } 1618 if (UseCMSInitiatingOccupancyOnly) { 1619 return false; 1620 } 1621 if (expansion_cause() == CMSExpansionCause::_satisfy_allocation) { 1622 if (PrintGCDetails && Verbose) { 1623 gclog_or_tty->print(" %s: collect because expanded for allocation ", 1624 short_name()); 1625 } 1626 return true; 1627 } 1628 if (_cmsSpace->should_concurrent_collect()) { 1629 if (PrintGCDetails && Verbose) { 1630 gclog_or_tty->print(" %s: collect because cmsSpace says so ", 1631 short_name()); 1632 } 1633 return true; 1634 } 1635 return false; 1636 } 1637 1638 void ConcurrentMarkSweepGeneration::collect(bool full, 1639 bool clear_all_soft_refs, 1640 size_t size, 1641 bool tlab) 1642 { 1643 collector()->collect(full, clear_all_soft_refs, size, tlab); 1644 } 1645 1646 void CMSCollector::collect(bool full, 1647 bool clear_all_soft_refs, 1648 size_t size, 1649 bool tlab) 1650 { 1651 if (!UseCMSCollectionPassing && _collectorState > Idling) { 1652 // For debugging purposes skip the collection if the state 1653 // is not currently idle 1654 if (TraceCMSState) { 1655 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " skipped full:%d CMS state %d", 1656 Thread::current(), full, _collectorState); 1657 } 1658 return; 1659 } 1660 1661 // The following "if" branch is present for defensive reasons. 1662 // In the current uses of this interface, it can be replaced with: 1663 // assert(!GC_locker.is_active(), "Can't be called otherwise"); 1664 // But I am not placing that assert here to allow future 1665 // generality in invoking this interface. 1666 if (GC_locker::is_active()) { 1667 // A consistency test for GC_locker 1668 assert(GC_locker::needs_gc(), "Should have been set already"); 1669 // Skip this foreground collection, instead 1670 // expanding the heap if necessary. 1671 // Need the free list locks for the call to free() in compute_new_size() 1672 compute_new_size(); 1673 return; 1674 } 1675 acquire_control_and_collect(full, clear_all_soft_refs); 1676 _full_gcs_since_conc_gc++; 1677 } 1678 1679 void CMSCollector::request_full_gc(unsigned int full_gc_count) { 1680 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1681 unsigned int gc_count = gch->total_full_collections(); 1682 if (gc_count == full_gc_count) { 1683 MutexLockerEx y(CGC_lock, Mutex::_no_safepoint_check_flag); 1684 _full_gc_requested = true; 1685 CGC_lock->notify(); // nudge CMS thread 1686 } else { 1687 assert(gc_count > full_gc_count, "Error: causal loop"); 1688 } 1689 } 1690 1691 1692 // The foreground and background collectors need to coordinate in order 1693 // to make sure that they do not mutually interfere with CMS collections. 1694 // When a background collection is active, 1695 // the foreground collector may need to take over (preempt) and 1696 // synchronously complete an ongoing collection. Depending on the 1697 // frequency of the background collections and the heap usage 1698 // of the application, this preemption can be seldom or frequent. 1699 // There are only certain 1700 // points in the background collection that the "collection-baton" 1701 // can be passed to the foreground collector. 1702 // 1703 // The foreground collector will wait for the baton before 1704 // starting any part of the collection. The foreground collector 1705 // will only wait at one location. 1706 // 1707 // The background collector will yield the baton before starting a new 1708 // phase of the collection (e.g., before initial marking, marking from roots, 1709 // precleaning, final re-mark, sweep etc.) This is normally done at the head 1710 // of the loop which switches the phases. The background collector does some 1711 // of the phases (initial mark, final re-mark) with the world stopped. 1712 // Because of locking involved in stopping the world, 1713 // the foreground collector should not block waiting for the background 1714 // collector when it is doing a stop-the-world phase. The background 1715 // collector will yield the baton at an additional point just before 1716 // it enters a stop-the-world phase. Once the world is stopped, the 1717 // background collector checks the phase of the collection. If the 1718 // phase has not changed, it proceeds with the collection. If the 1719 // phase has changed, it skips that phase of the collection. See 1720 // the comments on the use of the Heap_lock in collect_in_background(). 1721 // 1722 // Variable used in baton passing. 1723 // _foregroundGCIsActive - Set to true by the foreground collector when 1724 // it wants the baton. The foreground clears it when it has finished 1725 // the collection. 1726 // _foregroundGCShouldWait - Set to true by the background collector 1727 // when it is running. The foreground collector waits while 1728 // _foregroundGCShouldWait is true. 1729 // CGC_lock - monitor used to protect access to the above variables 1730 // and to notify the foreground and background collectors. 1731 // _collectorState - current state of the CMS collection. 1732 // 1733 // The foreground collector 1734 // acquires the CGC_lock 1735 // sets _foregroundGCIsActive 1736 // waits on the CGC_lock for _foregroundGCShouldWait to be false 1737 // various locks acquired in preparation for the collection 1738 // are released so as not to block the background collector 1739 // that is in the midst of a collection 1740 // proceeds with the collection 1741 // clears _foregroundGCIsActive 1742 // returns 1743 // 1744 // The background collector in a loop iterating on the phases of the 1745 // collection 1746 // acquires the CGC_lock 1747 // sets _foregroundGCShouldWait 1748 // if _foregroundGCIsActive is set 1749 // clears _foregroundGCShouldWait, notifies _CGC_lock 1750 // waits on _CGC_lock for _foregroundGCIsActive to become false 1751 // and exits the loop. 1752 // otherwise 1753 // proceed with that phase of the collection 1754 // if the phase is a stop-the-world phase, 1755 // yield the baton once more just before enqueueing 1756 // the stop-world CMS operation (executed by the VM thread). 1757 // returns after all phases of the collection are done 1758 // 1759 1760 void CMSCollector::acquire_control_and_collect(bool full, 1761 bool clear_all_soft_refs) { 1762 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint"); 1763 assert(!Thread::current()->is_ConcurrentGC_thread(), 1764 "shouldn't try to acquire control from self!"); 1765 1766 // Start the protocol for acquiring control of the 1767 // collection from the background collector (aka CMS thread). 1768 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 1769 "VM thread should have CMS token"); 1770 // Remember the possibly interrupted state of an ongoing 1771 // concurrent collection 1772 CollectorState first_state = _collectorState; 1773 1774 // Signal to a possibly ongoing concurrent collection that 1775 // we want to do a foreground collection. 1776 _foregroundGCIsActive = true; 1777 1778 // Disable incremental mode during a foreground collection. 1779 ICMSDisabler icms_disabler; 1780 1781 // release locks and wait for a notify from the background collector 1782 // releasing the locks in only necessary for phases which 1783 // do yields to improve the granularity of the collection. 1784 assert_lock_strong(bitMapLock()); 1785 // We need to lock the Free list lock for the space that we are 1786 // currently collecting. 1787 assert(haveFreelistLocks(), "Must be holding free list locks"); 1788 bitMapLock()->unlock(); 1789 releaseFreelistLocks(); 1790 { 1791 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1792 if (_foregroundGCShouldWait) { 1793 // We are going to be waiting for action for the CMS thread; 1794 // it had better not be gone (for instance at shutdown)! 1795 assert(ConcurrentMarkSweepThread::cmst() != NULL, 1796 "CMS thread must be running"); 1797 // Wait here until the background collector gives us the go-ahead 1798 ConcurrentMarkSweepThread::clear_CMS_flag( 1799 ConcurrentMarkSweepThread::CMS_vm_has_token); // release token 1800 // Get a possibly blocked CMS thread going: 1801 // Note that we set _foregroundGCIsActive true above, 1802 // without protection of the CGC_lock. 1803 CGC_lock->notify(); 1804 assert(!ConcurrentMarkSweepThread::vm_thread_wants_cms_token(), 1805 "Possible deadlock"); 1806 while (_foregroundGCShouldWait) { 1807 // wait for notification 1808 CGC_lock->wait(Mutex::_no_safepoint_check_flag); 1809 // Possibility of delay/starvation here, since CMS token does 1810 // not know to give priority to VM thread? Actually, i think 1811 // there wouldn't be any delay/starvation, but the proof of 1812 // that "fact" (?) appears non-trivial. XXX 20011219YSR 1813 } 1814 ConcurrentMarkSweepThread::set_CMS_flag( 1815 ConcurrentMarkSweepThread::CMS_vm_has_token); 1816 } 1817 } 1818 // The CMS_token is already held. Get back the other locks. 1819 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 1820 "VM thread should have CMS token"); 1821 getFreelistLocks(); 1822 bitMapLock()->lock_without_safepoint_check(); 1823 if (TraceCMSState) { 1824 gclog_or_tty->print_cr("CMS foreground collector has asked for control " 1825 INTPTR_FORMAT " with first state %d", Thread::current(), first_state); 1826 gclog_or_tty->print_cr(" gets control with state %d", _collectorState); 1827 } 1828 1829 // Check if we need to do a compaction, or if not, whether 1830 // we need to start the mark-sweep from scratch. 1831 bool should_compact = false; 1832 bool should_start_over = false; 1833 decide_foreground_collection_type(clear_all_soft_refs, 1834 &should_compact, &should_start_over); 1835 1836 NOT_PRODUCT( 1837 if (RotateCMSCollectionTypes) { 1838 if (_cmsGen->debug_collection_type() == 1839 ConcurrentMarkSweepGeneration::MSC_foreground_collection_type) { 1840 should_compact = true; 1841 } else if (_cmsGen->debug_collection_type() == 1842 ConcurrentMarkSweepGeneration::MS_foreground_collection_type) { 1843 should_compact = false; 1844 } 1845 } 1846 ) 1847 1848 if (PrintGCDetails && first_state > Idling) { 1849 GCCause::Cause cause = GenCollectedHeap::heap()->gc_cause(); 1850 if (GCCause::is_user_requested_gc(cause) || 1851 GCCause::is_serviceability_requested_gc(cause)) { 1852 gclog_or_tty->print(" (concurrent mode interrupted)"); 1853 } else { 1854 gclog_or_tty->print(" (concurrent mode failure)"); 1855 } 1856 } 1857 1858 set_did_compact(should_compact); 1859 if (should_compact) { 1860 // If the collection is being acquired from the background 1861 // collector, there may be references on the discovered 1862 // references lists that have NULL referents (being those 1863 // that were concurrently cleared by a mutator) or 1864 // that are no longer active (having been enqueued concurrently 1865 // by the mutator). 1866 // Scrub the list of those references because Mark-Sweep-Compact 1867 // code assumes referents are not NULL and that all discovered 1868 // Reference objects are active. 1869 ref_processor()->clean_up_discovered_references(); 1870 1871 do_compaction_work(clear_all_soft_refs); 1872 1873 // Has the GC time limit been exceeded? 1874 DefNewGeneration* young_gen = _young_gen->as_DefNewGeneration(); 1875 size_t max_eden_size = young_gen->max_capacity() - 1876 young_gen->to()->capacity() - 1877 young_gen->from()->capacity(); 1878 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1879 GCCause::Cause gc_cause = gch->gc_cause(); 1880 size_policy()->check_gc_overhead_limit(_young_gen->used(), 1881 young_gen->eden()->used(), 1882 _cmsGen->max_capacity(), 1883 max_eden_size, 1884 full, 1885 gc_cause, 1886 gch->collector_policy()); 1887 } else { 1888 do_mark_sweep_work(clear_all_soft_refs, first_state, 1889 should_start_over); 1890 } 1891 // Reset the expansion cause, now that we just completed 1892 // a collection cycle. 1893 clear_expansion_cause(); 1894 _foregroundGCIsActive = false; 1895 return; 1896 } 1897 1898 // Resize the tenured generation 1899 // after obtaining the free list locks for the 1900 // two generations. 1901 void CMSCollector::compute_new_size() { 1902 assert_locked_or_safepoint(Heap_lock); 1903 FreelistLocker z(this); 1904 MetaspaceGC::compute_new_size(); 1905 _cmsGen->compute_new_size_free_list(); 1906 } 1907 1908 // A work method used by foreground collection to determine 1909 // what type of collection (compacting or not, continuing or fresh) 1910 // it should do. 1911 // NOTE: the intent is to make UseCMSCompactAtFullCollection 1912 // and CMSCompactWhenClearAllSoftRefs the default in the future 1913 // and do away with the flags after a suitable period. 1914 void CMSCollector::decide_foreground_collection_type( 1915 bool clear_all_soft_refs, bool* should_compact, 1916 bool* should_start_over) { 1917 // Normally, we'll compact only if the UseCMSCompactAtFullCollection 1918 // flag is set, and we have either requested a System.gc() or 1919 // the number of full gc's since the last concurrent cycle 1920 // has exceeded the threshold set by CMSFullGCsBeforeCompaction, 1921 // or if an incremental collection has failed 1922 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1923 assert(gch->collector_policy()->is_two_generation_policy(), 1924 "You may want to check the correctness of the following"); 1925 // Inform cms gen if this was due to partial collection failing. 1926 // The CMS gen may use this fact to determine its expansion policy. 1927 if (gch->incremental_collection_will_fail(false /* don't consult_young */)) { 1928 assert(!_cmsGen->incremental_collection_failed(), 1929 "Should have been noticed, reacted to and cleared"); 1930 _cmsGen->set_incremental_collection_failed(); 1931 } 1932 *should_compact = 1933 UseCMSCompactAtFullCollection && 1934 ((_full_gcs_since_conc_gc >= CMSFullGCsBeforeCompaction) || 1935 GCCause::is_user_requested_gc(gch->gc_cause()) || 1936 gch->incremental_collection_will_fail(true /* consult_young */)); 1937 *should_start_over = false; 1938 if (clear_all_soft_refs && !*should_compact) { 1939 // We are about to do a last ditch collection attempt 1940 // so it would normally make sense to do a compaction 1941 // to reclaim as much space as possible. 1942 if (CMSCompactWhenClearAllSoftRefs) { 1943 // Default: The rationale is that in this case either 1944 // we are past the final marking phase, in which case 1945 // we'd have to start over, or so little has been done 1946 // that there's little point in saving that work. Compaction 1947 // appears to be the sensible choice in either case. 1948 *should_compact = true; 1949 } else { 1950 // We have been asked to clear all soft refs, but not to 1951 // compact. Make sure that we aren't past the final checkpoint 1952 // phase, for that is where we process soft refs. If we are already 1953 // past that phase, we'll need to redo the refs discovery phase and 1954 // if necessary clear soft refs that weren't previously 1955 // cleared. We do so by remembering the phase in which 1956 // we came in, and if we are past the refs processing 1957 // phase, we'll choose to just redo the mark-sweep 1958 // collection from scratch. 1959 if (_collectorState > FinalMarking) { 1960 // We are past the refs processing phase; 1961 // start over and do a fresh synchronous CMS cycle 1962 _collectorState = Resetting; // skip to reset to start new cycle 1963 reset(false /* == !asynch */); 1964 *should_start_over = true; 1965 } // else we can continue a possibly ongoing current cycle 1966 } 1967 } 1968 } 1969 1970 // A work method used by the foreground collector to do 1971 // a mark-sweep-compact. 1972 void CMSCollector::do_compaction_work(bool clear_all_soft_refs) { 1973 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1974 TraceTime t("CMS:MSC ", PrintGCDetails && Verbose, true, gclog_or_tty); 1975 if (PrintGC && Verbose && !(GCCause::is_user_requested_gc(gch->gc_cause()))) { 1976 gclog_or_tty->print_cr("Compact ConcurrentMarkSweepGeneration after %d " 1977 "collections passed to foreground collector", _full_gcs_since_conc_gc); 1978 } 1979 1980 // Sample collection interval time and reset for collection pause. 1981 if (UseAdaptiveSizePolicy) { 1982 size_policy()->msc_collection_begin(); 1983 } 1984 1985 // Temporarily widen the span of the weak reference processing to 1986 // the entire heap. 1987 MemRegion new_span(GenCollectedHeap::heap()->reserved_region()); 1988 ReferenceProcessorSpanMutator rp_mut_span(ref_processor(), new_span); 1989 // Temporarily, clear the "is_alive_non_header" field of the 1990 // reference processor. 1991 ReferenceProcessorIsAliveMutator rp_mut_closure(ref_processor(), NULL); 1992 // Temporarily make reference _processing_ single threaded (non-MT). 1993 ReferenceProcessorMTProcMutator rp_mut_mt_processing(ref_processor(), false); 1994 // Temporarily make refs discovery atomic 1995 ReferenceProcessorAtomicMutator rp_mut_atomic(ref_processor(), true); 1996 // Temporarily make reference _discovery_ single threaded (non-MT) 1997 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false); 1998 1999 ref_processor()->set_enqueuing_is_done(false); 2000 ref_processor()->enable_discovery(false /*verify_disabled*/, false /*check_no_refs*/); 2001 ref_processor()->setup_policy(clear_all_soft_refs); 2002 // If an asynchronous collection finishes, the _modUnionTable is 2003 // all clear. If we are assuming the collection from an asynchronous 2004 // collection, clear the _modUnionTable. 2005 assert(_collectorState != Idling || _modUnionTable.isAllClear(), 2006 "_modUnionTable should be clear if the baton was not passed"); 2007 _modUnionTable.clear_all(); 2008 assert(_collectorState != Idling || _ct->klass_rem_set()->mod_union_is_clear(), 2009 "mod union for klasses should be clear if the baton was passed"); 2010 _ct->klass_rem_set()->clear_mod_union(); 2011 2012 // We must adjust the allocation statistics being maintained 2013 // in the free list space. We do so by reading and clearing 2014 // the sweep timer and updating the block flux rate estimates below. 2015 assert(!_intra_sweep_timer.is_active(), "_intra_sweep_timer should be inactive"); 2016 if (_inter_sweep_timer.is_active()) { 2017 _inter_sweep_timer.stop(); 2018 // Note that we do not use this sample to update the _inter_sweep_estimate. 2019 _cmsGen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()), 2020 _inter_sweep_estimate.padded_average(), 2021 _intra_sweep_estimate.padded_average()); 2022 } 2023 2024 GenMarkSweep::invoke_at_safepoint(_cmsGen->level(), 2025 ref_processor(), clear_all_soft_refs); 2026 #ifdef ASSERT 2027 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 2028 size_t free_size = cms_space->free(); 2029 assert(free_size == 2030 pointer_delta(cms_space->end(), cms_space->compaction_top()) 2031 * HeapWordSize, 2032 "All the free space should be compacted into one chunk at top"); 2033 assert(cms_space->dictionary()->total_chunk_size( 2034 debug_only(cms_space->freelistLock())) == 0 || 2035 cms_space->totalSizeInIndexedFreeLists() == 0, 2036 "All the free space should be in a single chunk"); 2037 size_t num = cms_space->totalCount(); 2038 assert((free_size == 0 && num == 0) || 2039 (free_size > 0 && (num == 1 || num == 2)), 2040 "There should be at most 2 free chunks after compaction"); 2041 #endif // ASSERT 2042 _collectorState = Resetting; 2043 assert(_restart_addr == NULL, 2044 "Should have been NULL'd before baton was passed"); 2045 reset(false /* == !asynch */); 2046 _cmsGen->reset_after_compaction(); 2047 _concurrent_cycles_since_last_unload = 0; 2048 2049 // Clear any data recorded in the PLAB chunk arrays. 2050 if (_survivor_plab_array != NULL) { 2051 reset_survivor_plab_arrays(); 2052 } 2053 2054 // Adjust the per-size allocation stats for the next epoch. 2055 _cmsGen->cmsSpace()->endSweepFLCensus(sweep_count() /* fake */); 2056 // Restart the "inter sweep timer" for the next epoch. 2057 _inter_sweep_timer.reset(); 2058 _inter_sweep_timer.start(); 2059 2060 // Sample collection pause time and reset for collection interval. 2061 if (UseAdaptiveSizePolicy) { 2062 size_policy()->msc_collection_end(gch->gc_cause()); 2063 } 2064 2065 // For a mark-sweep-compact, compute_new_size() will be called 2066 // in the heap's do_collection() method. 2067 } 2068 2069 // A work method used by the foreground collector to do 2070 // a mark-sweep, after taking over from a possibly on-going 2071 // concurrent mark-sweep collection. 2072 void CMSCollector::do_mark_sweep_work(bool clear_all_soft_refs, 2073 CollectorState first_state, bool should_start_over) { 2074 if (PrintGC && Verbose) { 2075 gclog_or_tty->print_cr("Pass concurrent collection to foreground " 2076 "collector with count %d", 2077 _full_gcs_since_conc_gc); 2078 } 2079 switch (_collectorState) { 2080 case Idling: 2081 if (first_state == Idling || should_start_over) { 2082 // The background GC was not active, or should 2083 // restarted from scratch; start the cycle. 2084 _collectorState = InitialMarking; 2085 } 2086 // If first_state was not Idling, then a background GC 2087 // was in progress and has now finished. No need to do it 2088 // again. Leave the state as Idling. 2089 break; 2090 case Precleaning: 2091 // In the foreground case don't do the precleaning since 2092 // it is not done concurrently and there is extra work 2093 // required. 2094 _collectorState = FinalMarking; 2095 } 2096 collect_in_foreground(clear_all_soft_refs); 2097 2098 // For a mark-sweep, compute_new_size() will be called 2099 // in the heap's do_collection() method. 2100 } 2101 2102 2103 void CMSCollector::getFreelistLocks() const { 2104 // Get locks for all free lists in all generations that this 2105 // collector is responsible for 2106 _cmsGen->freelistLock()->lock_without_safepoint_check(); 2107 } 2108 2109 void CMSCollector::releaseFreelistLocks() const { 2110 // Release locks for all free lists in all generations that this 2111 // collector is responsible for 2112 _cmsGen->freelistLock()->unlock(); 2113 } 2114 2115 bool CMSCollector::haveFreelistLocks() const { 2116 // Check locks for all free lists in all generations that this 2117 // collector is responsible for 2118 assert_lock_strong(_cmsGen->freelistLock()); 2119 PRODUCT_ONLY(ShouldNotReachHere()); 2120 return true; 2121 } 2122 2123 // A utility class that is used by the CMS collector to 2124 // temporarily "release" the foreground collector from its 2125 // usual obligation to wait for the background collector to 2126 // complete an ongoing phase before proceeding. 2127 class ReleaseForegroundGC: public StackObj { 2128 private: 2129 CMSCollector* _c; 2130 public: 2131 ReleaseForegroundGC(CMSCollector* c) : _c(c) { 2132 assert(_c->_foregroundGCShouldWait, "Else should not need to call"); 2133 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 2134 // allow a potentially blocked foreground collector to proceed 2135 _c->_foregroundGCShouldWait = false; 2136 if (_c->_foregroundGCIsActive) { 2137 CGC_lock->notify(); 2138 } 2139 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 2140 "Possible deadlock"); 2141 } 2142 2143 ~ReleaseForegroundGC() { 2144 assert(!_c->_foregroundGCShouldWait, "Usage protocol violation?"); 2145 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 2146 _c->_foregroundGCShouldWait = true; 2147 } 2148 }; 2149 2150 // There are separate collect_in_background and collect_in_foreground because of 2151 // the different locking requirements of the background collector and the 2152 // foreground collector. There was originally an attempt to share 2153 // one "collect" method between the background collector and the foreground 2154 // collector but the if-then-else required made it cleaner to have 2155 // separate methods. 2156 void CMSCollector::collect_in_background(bool clear_all_soft_refs) { 2157 assert(Thread::current()->is_ConcurrentGC_thread(), 2158 "A CMS asynchronous collection is only allowed on a CMS thread."); 2159 2160 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2161 { 2162 bool safepoint_check = Mutex::_no_safepoint_check_flag; 2163 MutexLockerEx hl(Heap_lock, safepoint_check); 2164 FreelistLocker fll(this); 2165 MutexLockerEx x(CGC_lock, safepoint_check); 2166 if (_foregroundGCIsActive || !UseAsyncConcMarkSweepGC) { 2167 // The foreground collector is active or we're 2168 // not using asynchronous collections. Skip this 2169 // background collection. 2170 assert(!_foregroundGCShouldWait, "Should be clear"); 2171 return; 2172 } else { 2173 assert(_collectorState == Idling, "Should be idling before start."); 2174 _collectorState = InitialMarking; 2175 // Reset the expansion cause, now that we are about to begin 2176 // a new cycle. 2177 clear_expansion_cause(); 2178 2179 // Clear the MetaspaceGC flag since a concurrent collection 2180 // is starting but also clear it after the collection. 2181 MetaspaceGC::set_should_concurrent_collect(false); 2182 } 2183 // Decide if we want to enable class unloading as part of the 2184 // ensuing concurrent GC cycle. 2185 update_should_unload_classes(); 2186 _full_gc_requested = false; // acks all outstanding full gc requests 2187 // Signal that we are about to start a collection 2188 gch->increment_total_full_collections(); // ... starting a collection cycle 2189 _collection_count_start = gch->total_full_collections(); 2190 } 2191 2192 // Used for PrintGC 2193 size_t prev_used; 2194 if (PrintGC && Verbose) { 2195 prev_used = _cmsGen->used(); // XXXPERM 2196 } 2197 2198 // The change of the collection state is normally done at this level; 2199 // the exceptions are phases that are executed while the world is 2200 // stopped. For those phases the change of state is done while the 2201 // world is stopped. For baton passing purposes this allows the 2202 // background collector to finish the phase and change state atomically. 2203 // The foreground collector cannot wait on a phase that is done 2204 // while the world is stopped because the foreground collector already 2205 // has the world stopped and would deadlock. 2206 while (_collectorState != Idling) { 2207 if (TraceCMSState) { 2208 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " in CMS state %d", 2209 Thread::current(), _collectorState); 2210 } 2211 // The foreground collector 2212 // holds the Heap_lock throughout its collection. 2213 // holds the CMS token (but not the lock) 2214 // except while it is waiting for the background collector to yield. 2215 // 2216 // The foreground collector should be blocked (not for long) 2217 // if the background collector is about to start a phase 2218 // executed with world stopped. If the background 2219 // collector has already started such a phase, the 2220 // foreground collector is blocked waiting for the 2221 // Heap_lock. The stop-world phases (InitialMarking and FinalMarking) 2222 // are executed in the VM thread. 2223 // 2224 // The locking order is 2225 // PendingListLock (PLL) -- if applicable (FinalMarking) 2226 // Heap_lock (both this & PLL locked in VM_CMS_Operation::prologue()) 2227 // CMS token (claimed in 2228 // stop_world_and_do() --> 2229 // safepoint_synchronize() --> 2230 // CMSThread::synchronize()) 2231 2232 { 2233 // Check if the FG collector wants us to yield. 2234 CMSTokenSync x(true); // is cms thread 2235 if (waitForForegroundGC()) { 2236 // We yielded to a foreground GC, nothing more to be 2237 // done this round. 2238 assert(_foregroundGCShouldWait == false, "We set it to false in " 2239 "waitForForegroundGC()"); 2240 if (TraceCMSState) { 2241 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT 2242 " exiting collection CMS state %d", 2243 Thread::current(), _collectorState); 2244 } 2245 return; 2246 } else { 2247 // The background collector can run but check to see if the 2248 // foreground collector has done a collection while the 2249 // background collector was waiting to get the CGC_lock 2250 // above. If yes, break so that _foregroundGCShouldWait 2251 // is cleared before returning. 2252 if (_collectorState == Idling) { 2253 break; 2254 } 2255 } 2256 } 2257 2258 assert(_foregroundGCShouldWait, "Foreground collector, if active, " 2259 "should be waiting"); 2260 2261 switch (_collectorState) { 2262 case InitialMarking: 2263 { 2264 ReleaseForegroundGC x(this); 2265 stats().record_cms_begin(); 2266 2267 VM_CMS_Initial_Mark initial_mark_op(this); 2268 VMThread::execute(&initial_mark_op); 2269 } 2270 // The collector state may be any legal state at this point 2271 // since the background collector may have yielded to the 2272 // foreground collector. 2273 break; 2274 case Marking: 2275 // initial marking in checkpointRootsInitialWork has been completed 2276 if (markFromRoots(true)) { // we were successful 2277 assert(_collectorState == Precleaning, "Collector state should " 2278 "have changed"); 2279 } else { 2280 assert(_foregroundGCIsActive, "Internal state inconsistency"); 2281 } 2282 break; 2283 case Precleaning: 2284 if (UseAdaptiveSizePolicy) { 2285 size_policy()->concurrent_precleaning_begin(); 2286 } 2287 // marking from roots in markFromRoots has been completed 2288 preclean(); 2289 if (UseAdaptiveSizePolicy) { 2290 size_policy()->concurrent_precleaning_end(); 2291 } 2292 assert(_collectorState == AbortablePreclean || 2293 _collectorState == FinalMarking, 2294 "Collector state should have changed"); 2295 break; 2296 case AbortablePreclean: 2297 if (UseAdaptiveSizePolicy) { 2298 size_policy()->concurrent_phases_resume(); 2299 } 2300 abortable_preclean(); 2301 if (UseAdaptiveSizePolicy) { 2302 size_policy()->concurrent_precleaning_end(); 2303 } 2304 assert(_collectorState == FinalMarking, "Collector state should " 2305 "have changed"); 2306 break; 2307 case FinalMarking: 2308 { 2309 ReleaseForegroundGC x(this); 2310 2311 VM_CMS_Final_Remark final_remark_op(this); 2312 VMThread::execute(&final_remark_op); 2313 } 2314 assert(_foregroundGCShouldWait, "block post-condition"); 2315 break; 2316 case Sweeping: 2317 if (UseAdaptiveSizePolicy) { 2318 size_policy()->concurrent_sweeping_begin(); 2319 } 2320 // final marking in checkpointRootsFinal has been completed 2321 sweep(true); 2322 assert(_collectorState == Resizing, "Collector state change " 2323 "to Resizing must be done under the free_list_lock"); 2324 _full_gcs_since_conc_gc = 0; 2325 2326 // Stop the timers for adaptive size policy for the concurrent phases 2327 if (UseAdaptiveSizePolicy) { 2328 size_policy()->concurrent_sweeping_end(); 2329 size_policy()->concurrent_phases_end(gch->gc_cause(), 2330 gch->prev_gen(_cmsGen)->capacity(), 2331 _cmsGen->free()); 2332 } 2333 2334 case Resizing: { 2335 // Sweeping has been completed... 2336 // At this point the background collection has completed. 2337 // Don't move the call to compute_new_size() down 2338 // into code that might be executed if the background 2339 // collection was preempted. 2340 { 2341 ReleaseForegroundGC x(this); // unblock FG collection 2342 MutexLockerEx y(Heap_lock, Mutex::_no_safepoint_check_flag); 2343 CMSTokenSync z(true); // not strictly needed. 2344 if (_collectorState == Resizing) { 2345 compute_new_size(); 2346 _collectorState = Resetting; 2347 } else { 2348 assert(_collectorState == Idling, "The state should only change" 2349 " because the foreground collector has finished the collection"); 2350 } 2351 } 2352 break; 2353 } 2354 case Resetting: 2355 // CMS heap resizing has been completed 2356 reset(true); 2357 assert(_collectorState == Idling, "Collector state should " 2358 "have changed"); 2359 2360 MetaspaceGC::set_should_concurrent_collect(false); 2361 2362 stats().record_cms_end(); 2363 // Don't move the concurrent_phases_end() and compute_new_size() 2364 // calls to here because a preempted background collection 2365 // has it's state set to "Resetting". 2366 break; 2367 case Idling: 2368 default: 2369 ShouldNotReachHere(); 2370 break; 2371 } 2372 if (TraceCMSState) { 2373 gclog_or_tty->print_cr(" Thread " INTPTR_FORMAT " done - next CMS state %d", 2374 Thread::current(), _collectorState); 2375 } 2376 assert(_foregroundGCShouldWait, "block post-condition"); 2377 } 2378 2379 // Should this be in gc_epilogue? 2380 collector_policy()->counters()->update_counters(); 2381 2382 { 2383 // Clear _foregroundGCShouldWait and, in the event that the 2384 // foreground collector is waiting, notify it, before 2385 // returning. 2386 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 2387 _foregroundGCShouldWait = false; 2388 if (_foregroundGCIsActive) { 2389 CGC_lock->notify(); 2390 } 2391 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 2392 "Possible deadlock"); 2393 } 2394 if (TraceCMSState) { 2395 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT 2396 " exiting collection CMS state %d", 2397 Thread::current(), _collectorState); 2398 } 2399 if (PrintGC && Verbose) { 2400 _cmsGen->print_heap_change(prev_used); 2401 } 2402 } 2403 2404 void CMSCollector::collect_in_foreground(bool clear_all_soft_refs) { 2405 assert(_foregroundGCIsActive && !_foregroundGCShouldWait, 2406 "Foreground collector should be waiting, not executing"); 2407 assert(Thread::current()->is_VM_thread(), "A foreground collection" 2408 "may only be done by the VM Thread with the world stopped"); 2409 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 2410 "VM thread should have CMS token"); 2411 2412 NOT_PRODUCT(TraceTime t("CMS:MS (foreground) ", PrintGCDetails && Verbose, 2413 true, gclog_or_tty);) 2414 if (UseAdaptiveSizePolicy) { 2415 size_policy()->ms_collection_begin(); 2416 } 2417 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact); 2418 2419 HandleMark hm; // Discard invalid handles created during verification 2420 2421 if (VerifyBeforeGC && 2422 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2423 Universe::verify(); 2424 } 2425 2426 // Snapshot the soft reference policy to be used in this collection cycle. 2427 ref_processor()->setup_policy(clear_all_soft_refs); 2428 2429 bool init_mark_was_synchronous = false; // until proven otherwise 2430 while (_collectorState != Idling) { 2431 if (TraceCMSState) { 2432 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " in CMS state %d", 2433 Thread::current(), _collectorState); 2434 } 2435 switch (_collectorState) { 2436 case InitialMarking: 2437 init_mark_was_synchronous = true; // fact to be exploited in re-mark 2438 checkpointRootsInitial(false); 2439 assert(_collectorState == Marking, "Collector state should have changed" 2440 " within checkpointRootsInitial()"); 2441 break; 2442 case Marking: 2443 // initial marking in checkpointRootsInitialWork has been completed 2444 if (VerifyDuringGC && 2445 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2446 Universe::verify("Verify before initial mark: "); 2447 } 2448 { 2449 bool res = markFromRoots(false); 2450 assert(res && _collectorState == FinalMarking, "Collector state should " 2451 "have changed"); 2452 break; 2453 } 2454 case FinalMarking: 2455 if (VerifyDuringGC && 2456 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2457 Universe::verify("Verify before re-mark: "); 2458 } 2459 checkpointRootsFinal(false, clear_all_soft_refs, 2460 init_mark_was_synchronous); 2461 assert(_collectorState == Sweeping, "Collector state should not " 2462 "have changed within checkpointRootsFinal()"); 2463 break; 2464 case Sweeping: 2465 // final marking in checkpointRootsFinal has been completed 2466 if (VerifyDuringGC && 2467 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2468 Universe::verify("Verify before sweep: "); 2469 } 2470 sweep(false); 2471 assert(_collectorState == Resizing, "Incorrect state"); 2472 break; 2473 case Resizing: { 2474 // Sweeping has been completed; the actual resize in this case 2475 // is done separately; nothing to be done in this state. 2476 _collectorState = Resetting; 2477 break; 2478 } 2479 case Resetting: 2480 // The heap has been resized. 2481 if (VerifyDuringGC && 2482 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2483 Universe::verify("Verify before reset: "); 2484 } 2485 reset(false); 2486 assert(_collectorState == Idling, "Collector state should " 2487 "have changed"); 2488 break; 2489 case Precleaning: 2490 case AbortablePreclean: 2491 // Elide the preclean phase 2492 _collectorState = FinalMarking; 2493 break; 2494 default: 2495 ShouldNotReachHere(); 2496 } 2497 if (TraceCMSState) { 2498 gclog_or_tty->print_cr(" Thread " INTPTR_FORMAT " done - next CMS state %d", 2499 Thread::current(), _collectorState); 2500 } 2501 } 2502 2503 if (UseAdaptiveSizePolicy) { 2504 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2505 size_policy()->ms_collection_end(gch->gc_cause()); 2506 } 2507 2508 if (VerifyAfterGC && 2509 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2510 Universe::verify(); 2511 } 2512 if (TraceCMSState) { 2513 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT 2514 " exiting collection CMS state %d", 2515 Thread::current(), _collectorState); 2516 } 2517 } 2518 2519 bool CMSCollector::waitForForegroundGC() { 2520 bool res = false; 2521 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 2522 "CMS thread should have CMS token"); 2523 // Block the foreground collector until the 2524 // background collectors decides whether to 2525 // yield. 2526 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 2527 _foregroundGCShouldWait = true; 2528 if (_foregroundGCIsActive) { 2529 // The background collector yields to the 2530 // foreground collector and returns a value 2531 // indicating that it has yielded. The foreground 2532 // collector can proceed. 2533 res = true; 2534 _foregroundGCShouldWait = false; 2535 ConcurrentMarkSweepThread::clear_CMS_flag( 2536 ConcurrentMarkSweepThread::CMS_cms_has_token); 2537 ConcurrentMarkSweepThread::set_CMS_flag( 2538 ConcurrentMarkSweepThread::CMS_cms_wants_token); 2539 // Get a possibly blocked foreground thread going 2540 CGC_lock->notify(); 2541 if (TraceCMSState) { 2542 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " waiting at CMS state %d", 2543 Thread::current(), _collectorState); 2544 } 2545 while (_foregroundGCIsActive) { 2546 CGC_lock->wait(Mutex::_no_safepoint_check_flag); 2547 } 2548 ConcurrentMarkSweepThread::set_CMS_flag( 2549 ConcurrentMarkSweepThread::CMS_cms_has_token); 2550 ConcurrentMarkSweepThread::clear_CMS_flag( 2551 ConcurrentMarkSweepThread::CMS_cms_wants_token); 2552 } 2553 if (TraceCMSState) { 2554 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " continuing at CMS state %d", 2555 Thread::current(), _collectorState); 2556 } 2557 return res; 2558 } 2559 2560 // Because of the need to lock the free lists and other structures in 2561 // the collector, common to all the generations that the collector is 2562 // collecting, we need the gc_prologues of individual CMS generations 2563 // delegate to their collector. It may have been simpler had the 2564 // current infrastructure allowed one to call a prologue on a 2565 // collector. In the absence of that we have the generation's 2566 // prologue delegate to the collector, which delegates back 2567 // some "local" work to a worker method in the individual generations 2568 // that it's responsible for collecting, while itself doing any 2569 // work common to all generations it's responsible for. A similar 2570 // comment applies to the gc_epilogue()'s. 2571 // The role of the varaible _between_prologue_and_epilogue is to 2572 // enforce the invocation protocol. 2573 void CMSCollector::gc_prologue(bool full) { 2574 // Call gc_prologue_work() for the CMSGen 2575 // we are responsible for. 2576 2577 // The following locking discipline assumes that we are only called 2578 // when the world is stopped. 2579 assert(SafepointSynchronize::is_at_safepoint(), "world is stopped assumption"); 2580 2581 // The CMSCollector prologue must call the gc_prologues for the 2582 // "generations" that it's responsible 2583 // for. 2584 2585 assert( Thread::current()->is_VM_thread() 2586 || ( CMSScavengeBeforeRemark 2587 && Thread::current()->is_ConcurrentGC_thread()), 2588 "Incorrect thread type for prologue execution"); 2589 2590 if (_between_prologue_and_epilogue) { 2591 // We have already been invoked; this is a gc_prologue delegation 2592 // from yet another CMS generation that we are responsible for, just 2593 // ignore it since all relevant work has already been done. 2594 return; 2595 } 2596 2597 // set a bit saying prologue has been called; cleared in epilogue 2598 _between_prologue_and_epilogue = true; 2599 // Claim locks for common data structures, then call gc_prologue_work() 2600 // for each CMSGen. 2601 2602 getFreelistLocks(); // gets free list locks on constituent spaces 2603 bitMapLock()->lock_without_safepoint_check(); 2604 2605 // Should call gc_prologue_work() for all cms gens we are responsible for 2606 bool duringMarking = _collectorState >= Marking 2607 && _collectorState < Sweeping; 2608 2609 // The young collections clear the modified oops state, which tells if 2610 // there are any modified oops in the class. The remark phase also needs 2611 // that information. Tell the young collection to save the union of all 2612 // modified klasses. 2613 if (duringMarking) { 2614 _ct->klass_rem_set()->set_accumulate_modified_oops(true); 2615 } 2616 2617 bool registerClosure = duringMarking; 2618 2619 ModUnionClosure* muc = CollectedHeap::use_parallel_gc_threads() ? 2620 &_modUnionClosurePar 2621 : &_modUnionClosure; 2622 _cmsGen->gc_prologue_work(full, registerClosure, muc); 2623 2624 if (!full) { 2625 stats().record_gc0_begin(); 2626 } 2627 } 2628 2629 void ConcurrentMarkSweepGeneration::gc_prologue(bool full) { 2630 2631 _capacity_at_prologue = capacity(); 2632 _used_at_prologue = used(); 2633 2634 // Delegate to CMScollector which knows how to coordinate between 2635 // this and any other CMS generations that it is responsible for 2636 // collecting. 2637 collector()->gc_prologue(full); 2638 } 2639 2640 // This is a "private" interface for use by this generation's CMSCollector. 2641 // Not to be called directly by any other entity (for instance, 2642 // GenCollectedHeap, which calls the "public" gc_prologue method above). 2643 void ConcurrentMarkSweepGeneration::gc_prologue_work(bool full, 2644 bool registerClosure, ModUnionClosure* modUnionClosure) { 2645 assert(!incremental_collection_failed(), "Shouldn't be set yet"); 2646 assert(cmsSpace()->preconsumptionDirtyCardClosure() == NULL, 2647 "Should be NULL"); 2648 if (registerClosure) { 2649 cmsSpace()->setPreconsumptionDirtyCardClosure(modUnionClosure); 2650 } 2651 cmsSpace()->gc_prologue(); 2652 // Clear stat counters 2653 NOT_PRODUCT( 2654 assert(_numObjectsPromoted == 0, "check"); 2655 assert(_numWordsPromoted == 0, "check"); 2656 if (Verbose && PrintGC) { 2657 gclog_or_tty->print("Allocated "SIZE_FORMAT" objects, " 2658 SIZE_FORMAT" bytes concurrently", 2659 _numObjectsAllocated, _numWordsAllocated*sizeof(HeapWord)); 2660 } 2661 _numObjectsAllocated = 0; 2662 _numWordsAllocated = 0; 2663 ) 2664 } 2665 2666 void CMSCollector::gc_epilogue(bool full) { 2667 // The following locking discipline assumes that we are only called 2668 // when the world is stopped. 2669 assert(SafepointSynchronize::is_at_safepoint(), 2670 "world is stopped assumption"); 2671 2672 // Currently the CMS epilogue (see CompactibleFreeListSpace) merely checks 2673 // if linear allocation blocks need to be appropriately marked to allow the 2674 // the blocks to be parsable. We also check here whether we need to nudge the 2675 // CMS collector thread to start a new cycle (if it's not already active). 2676 assert( Thread::current()->is_VM_thread() 2677 || ( CMSScavengeBeforeRemark 2678 && Thread::current()->is_ConcurrentGC_thread()), 2679 "Incorrect thread type for epilogue execution"); 2680 2681 if (!_between_prologue_and_epilogue) { 2682 // We have already been invoked; this is a gc_epilogue delegation 2683 // from yet another CMS generation that we are responsible for, just 2684 // ignore it since all relevant work has already been done. 2685 return; 2686 } 2687 assert(haveFreelistLocks(), "must have freelist locks"); 2688 assert_lock_strong(bitMapLock()); 2689 2690 _ct->klass_rem_set()->set_accumulate_modified_oops(false); 2691 2692 _cmsGen->gc_epilogue_work(full); 2693 2694 if (_collectorState == AbortablePreclean || _collectorState == Precleaning) { 2695 // in case sampling was not already enabled, enable it 2696 _start_sampling = true; 2697 } 2698 // reset _eden_chunk_array so sampling starts afresh 2699 _eden_chunk_index = 0; 2700 2701 size_t cms_used = _cmsGen->cmsSpace()->used(); 2702 2703 // update performance counters - this uses a special version of 2704 // update_counters() that allows the utilization to be passed as a 2705 // parameter, avoiding multiple calls to used(). 2706 // 2707 _cmsGen->update_counters(cms_used); 2708 2709 if (CMSIncrementalMode) { 2710 icms_update_allocation_limits(); 2711 } 2712 2713 bitMapLock()->unlock(); 2714 releaseFreelistLocks(); 2715 2716 if (!CleanChunkPoolAsync) { 2717 Chunk::clean_chunk_pool(); 2718 } 2719 2720 set_did_compact(false); 2721 _between_prologue_and_epilogue = false; // ready for next cycle 2722 } 2723 2724 void ConcurrentMarkSweepGeneration::gc_epilogue(bool full) { 2725 collector()->gc_epilogue(full); 2726 2727 // Also reset promotion tracking in par gc thread states. 2728 if (CollectedHeap::use_parallel_gc_threads()) { 2729 for (uint i = 0; i < ParallelGCThreads; i++) { 2730 _par_gc_thread_states[i]->promo.stopTrackingPromotions(i); 2731 } 2732 } 2733 } 2734 2735 void ConcurrentMarkSweepGeneration::gc_epilogue_work(bool full) { 2736 assert(!incremental_collection_failed(), "Should have been cleared"); 2737 cmsSpace()->setPreconsumptionDirtyCardClosure(NULL); 2738 cmsSpace()->gc_epilogue(); 2739 // Print stat counters 2740 NOT_PRODUCT( 2741 assert(_numObjectsAllocated == 0, "check"); 2742 assert(_numWordsAllocated == 0, "check"); 2743 if (Verbose && PrintGC) { 2744 gclog_or_tty->print("Promoted "SIZE_FORMAT" objects, " 2745 SIZE_FORMAT" bytes", 2746 _numObjectsPromoted, _numWordsPromoted*sizeof(HeapWord)); 2747 } 2748 _numObjectsPromoted = 0; 2749 _numWordsPromoted = 0; 2750 ) 2751 2752 if (PrintGC && Verbose) { 2753 // Call down the chain in contiguous_available needs the freelistLock 2754 // so print this out before releasing the freeListLock. 2755 gclog_or_tty->print(" Contiguous available "SIZE_FORMAT" bytes ", 2756 contiguous_available()); 2757 } 2758 } 2759 2760 #ifndef PRODUCT 2761 bool CMSCollector::have_cms_token() { 2762 Thread* thr = Thread::current(); 2763 if (thr->is_VM_thread()) { 2764 return ConcurrentMarkSweepThread::vm_thread_has_cms_token(); 2765 } else if (thr->is_ConcurrentGC_thread()) { 2766 return ConcurrentMarkSweepThread::cms_thread_has_cms_token(); 2767 } else if (thr->is_GC_task_thread()) { 2768 return ConcurrentMarkSweepThread::vm_thread_has_cms_token() && 2769 ParGCRareEvent_lock->owned_by_self(); 2770 } 2771 return false; 2772 } 2773 #endif 2774 2775 // Check reachability of the given heap address in CMS generation, 2776 // treating all other generations as roots. 2777 bool CMSCollector::is_cms_reachable(HeapWord* addr) { 2778 // We could "guarantee" below, rather than assert, but i'll 2779 // leave these as "asserts" so that an adventurous debugger 2780 // could try this in the product build provided some subset of 2781 // the conditions were met, provided they were intersted in the 2782 // results and knew that the computation below wouldn't interfere 2783 // with other concurrent computations mutating the structures 2784 // being read or written. 2785 assert(SafepointSynchronize::is_at_safepoint(), 2786 "Else mutations in object graph will make answer suspect"); 2787 assert(have_cms_token(), "Should hold cms token"); 2788 assert(haveFreelistLocks(), "must hold free list locks"); 2789 assert_lock_strong(bitMapLock()); 2790 2791 // Clear the marking bit map array before starting, but, just 2792 // for kicks, first report if the given address is already marked 2793 gclog_or_tty->print_cr("Start: Address 0x%x is%s marked", addr, 2794 _markBitMap.isMarked(addr) ? "" : " not"); 2795 2796 if (verify_after_remark()) { 2797 MutexLockerEx x(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag); 2798 bool result = verification_mark_bm()->isMarked(addr); 2799 gclog_or_tty->print_cr("TransitiveMark: Address 0x%x %s marked", addr, 2800 result ? "IS" : "is NOT"); 2801 return result; 2802 } else { 2803 gclog_or_tty->print_cr("Could not compute result"); 2804 return false; 2805 } 2806 } 2807 2808 2809 void 2810 CMSCollector::print_on_error(outputStream* st) { 2811 CMSCollector* collector = ConcurrentMarkSweepGeneration::_collector; 2812 if (collector != NULL) { 2813 CMSBitMap* bitmap = &collector->_markBitMap; 2814 st->print_cr("Marking Bits: (CMSBitMap*) " PTR_FORMAT, bitmap); 2815 bitmap->print_on_error(st, " Bits: "); 2816 2817 st->cr(); 2818 2819 CMSBitMap* mut_bitmap = &collector->_modUnionTable; 2820 st->print_cr("Mod Union Table: (CMSBitMap*) " PTR_FORMAT, mut_bitmap); 2821 mut_bitmap->print_on_error(st, " Bits: "); 2822 } 2823 } 2824 2825 //////////////////////////////////////////////////////// 2826 // CMS Verification Support 2827 //////////////////////////////////////////////////////// 2828 // Following the remark phase, the following invariant 2829 // should hold -- each object in the CMS heap which is 2830 // marked in markBitMap() should be marked in the verification_mark_bm(). 2831 2832 class VerifyMarkedClosure: public BitMapClosure { 2833 CMSBitMap* _marks; 2834 bool _failed; 2835 2836 public: 2837 VerifyMarkedClosure(CMSBitMap* bm): _marks(bm), _failed(false) {} 2838 2839 bool do_bit(size_t offset) { 2840 HeapWord* addr = _marks->offsetToHeapWord(offset); 2841 if (!_marks->isMarked(addr)) { 2842 oop(addr)->print_on(gclog_or_tty); 2843 gclog_or_tty->print_cr(" ("INTPTR_FORMAT" should have been marked)", addr); 2844 _failed = true; 2845 } 2846 return true; 2847 } 2848 2849 bool failed() { return _failed; } 2850 }; 2851 2852 bool CMSCollector::verify_after_remark(bool silent) { 2853 if (!silent) gclog_or_tty->print(" [Verifying CMS Marking... "); 2854 MutexLockerEx ml(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag); 2855 static bool init = false; 2856 2857 assert(SafepointSynchronize::is_at_safepoint(), 2858 "Else mutations in object graph will make answer suspect"); 2859 assert(have_cms_token(), 2860 "Else there may be mutual interference in use of " 2861 " verification data structures"); 2862 assert(_collectorState > Marking && _collectorState <= Sweeping, 2863 "Else marking info checked here may be obsolete"); 2864 assert(haveFreelistLocks(), "must hold free list locks"); 2865 assert_lock_strong(bitMapLock()); 2866 2867 2868 // Allocate marking bit map if not already allocated 2869 if (!init) { // first time 2870 if (!verification_mark_bm()->allocate(_span)) { 2871 return false; 2872 } 2873 init = true; 2874 } 2875 2876 assert(verification_mark_stack()->isEmpty(), "Should be empty"); 2877 2878 // Turn off refs discovery -- so we will be tracing through refs. 2879 // This is as intended, because by this time 2880 // GC must already have cleared any refs that need to be cleared, 2881 // and traced those that need to be marked; moreover, 2882 // the marking done here is not going to intefere in any 2883 // way with the marking information used by GC. 2884 NoRefDiscovery no_discovery(ref_processor()); 2885 2886 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;) 2887 2888 // Clear any marks from a previous round 2889 verification_mark_bm()->clear_all(); 2890 assert(verification_mark_stack()->isEmpty(), "markStack should be empty"); 2891 verify_work_stacks_empty(); 2892 2893 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2894 gch->ensure_parsability(false); // fill TLABs, but no need to retire them 2895 // Update the saved marks which may affect the root scans. 2896 gch->save_marks(); 2897 2898 if (CMSRemarkVerifyVariant == 1) { 2899 // In this first variant of verification, we complete 2900 // all marking, then check if the new marks-verctor is 2901 // a subset of the CMS marks-vector. 2902 verify_after_remark_work_1(); 2903 } else if (CMSRemarkVerifyVariant == 2) { 2904 // In this second variant of verification, we flag an error 2905 // (i.e. an object reachable in the new marks-vector not reachable 2906 // in the CMS marks-vector) immediately, also indicating the 2907 // identify of an object (A) that references the unmarked object (B) -- 2908 // presumably, a mutation to A failed to be picked up by preclean/remark? 2909 verify_after_remark_work_2(); 2910 } else { 2911 warning("Unrecognized value %d for CMSRemarkVerifyVariant", 2912 CMSRemarkVerifyVariant); 2913 } 2914 if (!silent) gclog_or_tty->print(" done] "); 2915 return true; 2916 } 2917 2918 void CMSCollector::verify_after_remark_work_1() { 2919 ResourceMark rm; 2920 HandleMark hm; 2921 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2922 2923 // Get a clear set of claim bits for the strong roots processing to work with. 2924 ClassLoaderDataGraph::clear_claimed_marks(); 2925 2926 // Mark from roots one level into CMS 2927 MarkRefsIntoClosure notOlder(_span, verification_mark_bm()); 2928 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2929 2930 gch->gen_process_strong_roots(_cmsGen->level(), 2931 true, // younger gens are roots 2932 true, // activate StrongRootsScope 2933 false, // not scavenging 2934 SharedHeap::ScanningOption(roots_scanning_options()), 2935 ¬Older, 2936 true, // walk code active on stacks 2937 NULL, 2938 NULL); // SSS: Provide correct closure 2939 2940 // Now mark from the roots 2941 MarkFromRootsClosure markFromRootsClosure(this, _span, 2942 verification_mark_bm(), verification_mark_stack(), 2943 false /* don't yield */, true /* verifying */); 2944 assert(_restart_addr == NULL, "Expected pre-condition"); 2945 verification_mark_bm()->iterate(&markFromRootsClosure); 2946 while (_restart_addr != NULL) { 2947 // Deal with stack overflow: by restarting at the indicated 2948 // address. 2949 HeapWord* ra = _restart_addr; 2950 markFromRootsClosure.reset(ra); 2951 _restart_addr = NULL; 2952 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end()); 2953 } 2954 assert(verification_mark_stack()->isEmpty(), "Should have been drained"); 2955 verify_work_stacks_empty(); 2956 2957 // Marking completed -- now verify that each bit marked in 2958 // verification_mark_bm() is also marked in markBitMap(); flag all 2959 // errors by printing corresponding objects. 2960 VerifyMarkedClosure vcl(markBitMap()); 2961 verification_mark_bm()->iterate(&vcl); 2962 if (vcl.failed()) { 2963 gclog_or_tty->print("Verification failed"); 2964 Universe::heap()->print_on(gclog_or_tty); 2965 fatal("CMS: failed marking verification after remark"); 2966 } 2967 } 2968 2969 class VerifyKlassOopsKlassClosure : public KlassClosure { 2970 class VerifyKlassOopsClosure : public OopClosure { 2971 CMSBitMap* _bitmap; 2972 public: 2973 VerifyKlassOopsClosure(CMSBitMap* bitmap) : _bitmap(bitmap) { } 2974 void do_oop(oop* p) { guarantee(*p == NULL || _bitmap->isMarked((HeapWord*) *p), "Should be marked"); } 2975 void do_oop(narrowOop* p) { ShouldNotReachHere(); } 2976 } _oop_closure; 2977 public: 2978 VerifyKlassOopsKlassClosure(CMSBitMap* bitmap) : _oop_closure(bitmap) {} 2979 void do_klass(Klass* k) { 2980 k->oops_do(&_oop_closure); 2981 } 2982 }; 2983 2984 void CMSCollector::verify_after_remark_work_2() { 2985 ResourceMark rm; 2986 HandleMark hm; 2987 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2988 2989 // Get a clear set of claim bits for the strong roots processing to work with. 2990 ClassLoaderDataGraph::clear_claimed_marks(); 2991 2992 // Mark from roots one level into CMS 2993 MarkRefsIntoVerifyClosure notOlder(_span, verification_mark_bm(), 2994 markBitMap()); 2995 CMKlassClosure klass_closure(¬Older); 2996 2997 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2998 gch->gen_process_strong_roots(_cmsGen->level(), 2999 true, // younger gens are roots 3000 true, // activate StrongRootsScope 3001 false, // not scavenging 3002 SharedHeap::ScanningOption(roots_scanning_options()), 3003 ¬Older, 3004 true, // walk code active on stacks 3005 NULL, 3006 &klass_closure); 3007 3008 // Now mark from the roots 3009 MarkFromRootsVerifyClosure markFromRootsClosure(this, _span, 3010 verification_mark_bm(), markBitMap(), verification_mark_stack()); 3011 assert(_restart_addr == NULL, "Expected pre-condition"); 3012 verification_mark_bm()->iterate(&markFromRootsClosure); 3013 while (_restart_addr != NULL) { 3014 // Deal with stack overflow: by restarting at the indicated 3015 // address. 3016 HeapWord* ra = _restart_addr; 3017 markFromRootsClosure.reset(ra); 3018 _restart_addr = NULL; 3019 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end()); 3020 } 3021 assert(verification_mark_stack()->isEmpty(), "Should have been drained"); 3022 verify_work_stacks_empty(); 3023 3024 VerifyKlassOopsKlassClosure verify_klass_oops(verification_mark_bm()); 3025 ClassLoaderDataGraph::classes_do(&verify_klass_oops); 3026 3027 // Marking completed -- now verify that each bit marked in 3028 // verification_mark_bm() is also marked in markBitMap(); flag all 3029 // errors by printing corresponding objects. 3030 VerifyMarkedClosure vcl(markBitMap()); 3031 verification_mark_bm()->iterate(&vcl); 3032 assert(!vcl.failed(), "Else verification above should not have succeeded"); 3033 } 3034 3035 void ConcurrentMarkSweepGeneration::save_marks() { 3036 // delegate to CMS space 3037 cmsSpace()->save_marks(); 3038 for (uint i = 0; i < ParallelGCThreads; i++) { 3039 _par_gc_thread_states[i]->promo.startTrackingPromotions(); 3040 } 3041 } 3042 3043 bool ConcurrentMarkSweepGeneration::no_allocs_since_save_marks() { 3044 return cmsSpace()->no_allocs_since_save_marks(); 3045 } 3046 3047 #define CMS_SINCE_SAVE_MARKS_DEFN(OopClosureType, nv_suffix) \ 3048 \ 3049 void ConcurrentMarkSweepGeneration:: \ 3050 oop_since_save_marks_iterate##nv_suffix(OopClosureType* cl) { \ 3051 cl->set_generation(this); \ 3052 cmsSpace()->oop_since_save_marks_iterate##nv_suffix(cl); \ 3053 cl->reset_generation(); \ 3054 save_marks(); \ 3055 } 3056 3057 ALL_SINCE_SAVE_MARKS_CLOSURES(CMS_SINCE_SAVE_MARKS_DEFN) 3058 3059 void 3060 ConcurrentMarkSweepGeneration::younger_refs_iterate(OopsInGenClosure* cl) { 3061 cl->set_generation(this); 3062 younger_refs_in_space_iterate(_cmsSpace, cl); 3063 cl->reset_generation(); 3064 } 3065 3066 void 3067 ConcurrentMarkSweepGeneration::oop_iterate(MemRegion mr, ExtendedOopClosure* cl) { 3068 if (freelistLock()->owned_by_self()) { 3069 Generation::oop_iterate(mr, cl); 3070 } else { 3071 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 3072 Generation::oop_iterate(mr, cl); 3073 } 3074 } 3075 3076 void 3077 ConcurrentMarkSweepGeneration::oop_iterate(ExtendedOopClosure* cl) { 3078 if (freelistLock()->owned_by_self()) { 3079 Generation::oop_iterate(cl); 3080 } else { 3081 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 3082 Generation::oop_iterate(cl); 3083 } 3084 } 3085 3086 void 3087 ConcurrentMarkSweepGeneration::object_iterate(ObjectClosure* cl) { 3088 if (freelistLock()->owned_by_self()) { 3089 Generation::object_iterate(cl); 3090 } else { 3091 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 3092 Generation::object_iterate(cl); 3093 } 3094 } 3095 3096 void 3097 ConcurrentMarkSweepGeneration::safe_object_iterate(ObjectClosure* cl) { 3098 if (freelistLock()->owned_by_self()) { 3099 Generation::safe_object_iterate(cl); 3100 } else { 3101 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 3102 Generation::safe_object_iterate(cl); 3103 } 3104 } 3105 3106 void 3107 ConcurrentMarkSweepGeneration::post_compact() { 3108 } 3109 3110 void 3111 ConcurrentMarkSweepGeneration::prepare_for_verify() { 3112 // Fix the linear allocation blocks to look like free blocks. 3113 3114 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those 3115 // are not called when the heap is verified during universe initialization and 3116 // at vm shutdown. 3117 if (freelistLock()->owned_by_self()) { 3118 cmsSpace()->prepare_for_verify(); 3119 } else { 3120 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag); 3121 cmsSpace()->prepare_for_verify(); 3122 } 3123 } 3124 3125 void 3126 ConcurrentMarkSweepGeneration::verify() { 3127 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those 3128 // are not called when the heap is verified during universe initialization and 3129 // at vm shutdown. 3130 if (freelistLock()->owned_by_self()) { 3131 cmsSpace()->verify(); 3132 } else { 3133 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag); 3134 cmsSpace()->verify(); 3135 } 3136 } 3137 3138 void CMSCollector::verify() { 3139 _cmsGen->verify(); 3140 } 3141 3142 #ifndef PRODUCT 3143 bool CMSCollector::overflow_list_is_empty() const { 3144 assert(_num_par_pushes >= 0, "Inconsistency"); 3145 if (_overflow_list == NULL) { 3146 assert(_num_par_pushes == 0, "Inconsistency"); 3147 } 3148 return _overflow_list == NULL; 3149 } 3150 3151 // The methods verify_work_stacks_empty() and verify_overflow_empty() 3152 // merely consolidate assertion checks that appear to occur together frequently. 3153 void CMSCollector::verify_work_stacks_empty() const { 3154 assert(_markStack.isEmpty(), "Marking stack should be empty"); 3155 assert(overflow_list_is_empty(), "Overflow list should be empty"); 3156 } 3157 3158 void CMSCollector::verify_overflow_empty() const { 3159 assert(overflow_list_is_empty(), "Overflow list should be empty"); 3160 assert(no_preserved_marks(), "No preserved marks"); 3161 } 3162 #endif // PRODUCT 3163 3164 // Decide if we want to enable class unloading as part of the 3165 // ensuing concurrent GC cycle. We will collect and 3166 // unload classes if it's the case that: 3167 // (1) an explicit gc request has been made and the flag 3168 // ExplicitGCInvokesConcurrentAndUnloadsClasses is set, OR 3169 // (2) (a) class unloading is enabled at the command line, and 3170 // (b) old gen is getting really full 3171 // NOTE: Provided there is no change in the state of the heap between 3172 // calls to this method, it should have idempotent results. Moreover, 3173 // its results should be monotonically increasing (i.e. going from 0 to 1, 3174 // but not 1 to 0) between successive calls between which the heap was 3175 // not collected. For the implementation below, it must thus rely on 3176 // the property that concurrent_cycles_since_last_unload() 3177 // will not decrease unless a collection cycle happened and that 3178 // _cmsGen->is_too_full() are 3179 // themselves also monotonic in that sense. See check_monotonicity() 3180 // below. 3181 void CMSCollector::update_should_unload_classes() { 3182 _should_unload_classes = false; 3183 // Condition 1 above 3184 if (_full_gc_requested && ExplicitGCInvokesConcurrentAndUnloadsClasses) { 3185 _should_unload_classes = true; 3186 } else if (CMSClassUnloadingEnabled) { // Condition 2.a above 3187 // Disjuncts 2.b.(i,ii,iii) above 3188 _should_unload_classes = (concurrent_cycles_since_last_unload() >= 3189 CMSClassUnloadingMaxInterval) 3190 || _cmsGen->is_too_full(); 3191 } 3192 } 3193 3194 bool ConcurrentMarkSweepGeneration::is_too_full() const { 3195 bool res = should_concurrent_collect(); 3196 res = res && (occupancy() > (double)CMSIsTooFullPercentage/100.0); 3197 return res; 3198 } 3199 3200 void CMSCollector::setup_cms_unloading_and_verification_state() { 3201 const bool should_verify = VerifyBeforeGC || VerifyAfterGC || VerifyDuringGC 3202 || VerifyBeforeExit; 3203 const int rso = SharedHeap::SO_Strings | SharedHeap::SO_CodeCache; 3204 3205 if (should_unload_classes()) { // Should unload classes this cycle 3206 remove_root_scanning_option(rso); // Shrink the root set appropriately 3207 set_verifying(should_verify); // Set verification state for this cycle 3208 return; // Nothing else needs to be done at this time 3209 } 3210 3211 // Not unloading classes this cycle 3212 assert(!should_unload_classes(), "Inconsitency!"); 3213 if ((!verifying() || unloaded_classes_last_cycle()) && should_verify) { 3214 // Include symbols, strings and code cache elements to prevent their resurrection. 3215 add_root_scanning_option(rso); 3216 set_verifying(true); 3217 } else if (verifying() && !should_verify) { 3218 // We were verifying, but some verification flags got disabled. 3219 set_verifying(false); 3220 // Exclude symbols, strings and code cache elements from root scanning to 3221 // reduce IM and RM pauses. 3222 remove_root_scanning_option(rso); 3223 } 3224 } 3225 3226 3227 #ifndef PRODUCT 3228 HeapWord* CMSCollector::block_start(const void* p) const { 3229 const HeapWord* addr = (HeapWord*)p; 3230 if (_span.contains(p)) { 3231 if (_cmsGen->cmsSpace()->is_in_reserved(addr)) { 3232 return _cmsGen->cmsSpace()->block_start(p); 3233 } 3234 } 3235 return NULL; 3236 } 3237 #endif 3238 3239 HeapWord* 3240 ConcurrentMarkSweepGeneration::expand_and_allocate(size_t word_size, 3241 bool tlab, 3242 bool parallel) { 3243 CMSSynchronousYieldRequest yr; 3244 assert(!tlab, "Can't deal with TLAB allocation"); 3245 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 3246 expand(word_size*HeapWordSize, MinHeapDeltaBytes, 3247 CMSExpansionCause::_satisfy_allocation); 3248 if (GCExpandToAllocateDelayMillis > 0) { 3249 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 3250 } 3251 return have_lock_and_allocate(word_size, tlab); 3252 } 3253 3254 // YSR: All of this generation expansion/shrinking stuff is an exact copy of 3255 // OneContigSpaceCardGeneration, which makes me wonder if we should move this 3256 // to CardGeneration and share it... 3257 bool ConcurrentMarkSweepGeneration::expand(size_t bytes, size_t expand_bytes) { 3258 return CardGeneration::expand(bytes, expand_bytes); 3259 } 3260 3261 void ConcurrentMarkSweepGeneration::expand(size_t bytes, size_t expand_bytes, 3262 CMSExpansionCause::Cause cause) 3263 { 3264 3265 bool success = expand(bytes, expand_bytes); 3266 3267 // remember why we expanded; this information is used 3268 // by shouldConcurrentCollect() when making decisions on whether to start 3269 // a new CMS cycle. 3270 if (success) { 3271 set_expansion_cause(cause); 3272 if (PrintGCDetails && Verbose) { 3273 gclog_or_tty->print_cr("Expanded CMS gen for %s", 3274 CMSExpansionCause::to_string(cause)); 3275 } 3276 } 3277 } 3278 3279 HeapWord* ConcurrentMarkSweepGeneration::expand_and_par_lab_allocate(CMSParGCThreadState* ps, size_t word_sz) { 3280 HeapWord* res = NULL; 3281 MutexLocker x(ParGCRareEvent_lock); 3282 while (true) { 3283 // Expansion by some other thread might make alloc OK now: 3284 res = ps->lab.alloc(word_sz); 3285 if (res != NULL) return res; 3286 // If there's not enough expansion space available, give up. 3287 if (_virtual_space.uncommitted_size() < (word_sz * HeapWordSize)) { 3288 return NULL; 3289 } 3290 // Otherwise, we try expansion. 3291 expand(word_sz*HeapWordSize, MinHeapDeltaBytes, 3292 CMSExpansionCause::_allocate_par_lab); 3293 // Now go around the loop and try alloc again; 3294 // A competing par_promote might beat us to the expansion space, 3295 // so we may go around the loop again if promotion fails agaion. 3296 if (GCExpandToAllocateDelayMillis > 0) { 3297 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 3298 } 3299 } 3300 } 3301 3302 3303 bool ConcurrentMarkSweepGeneration::expand_and_ensure_spooling_space( 3304 PromotionInfo* promo) { 3305 MutexLocker x(ParGCRareEvent_lock); 3306 size_t refill_size_bytes = promo->refillSize() * HeapWordSize; 3307 while (true) { 3308 // Expansion by some other thread might make alloc OK now: 3309 if (promo->ensure_spooling_space()) { 3310 assert(promo->has_spooling_space(), 3311 "Post-condition of successful ensure_spooling_space()"); 3312 return true; 3313 } 3314 // If there's not enough expansion space available, give up. 3315 if (_virtual_space.uncommitted_size() < refill_size_bytes) { 3316 return false; 3317 } 3318 // Otherwise, we try expansion. 3319 expand(refill_size_bytes, MinHeapDeltaBytes, 3320 CMSExpansionCause::_allocate_par_spooling_space); 3321 // Now go around the loop and try alloc again; 3322 // A competing allocation might beat us to the expansion space, 3323 // so we may go around the loop again if allocation fails again. 3324 if (GCExpandToAllocateDelayMillis > 0) { 3325 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 3326 } 3327 } 3328 } 3329 3330 3331 void ConcurrentMarkSweepGeneration::shrink_by(size_t bytes) { 3332 assert_locked_or_safepoint(ExpandHeap_lock); 3333 // Shrink committed space 3334 _virtual_space.shrink_by(bytes); 3335 // Shrink space; this also shrinks the space's BOT 3336 _cmsSpace->set_end((HeapWord*) _virtual_space.high()); 3337 size_t new_word_size = heap_word_size(_cmsSpace->capacity()); 3338 // Shrink the shared block offset array 3339 _bts->resize(new_word_size); 3340 MemRegion mr(_cmsSpace->bottom(), new_word_size); 3341 // Shrink the card table 3342 Universe::heap()->barrier_set()->resize_covered_region(mr); 3343 3344 if (Verbose && PrintGC) { 3345 size_t new_mem_size = _virtual_space.committed_size(); 3346 size_t old_mem_size = new_mem_size + bytes; 3347 gclog_or_tty->print_cr("Shrinking %s from " SIZE_FORMAT "K to " SIZE_FORMAT "K", 3348 name(), old_mem_size/K, new_mem_size/K); 3349 } 3350 } 3351 3352 void ConcurrentMarkSweepGeneration::shrink(size_t bytes) { 3353 assert_locked_or_safepoint(Heap_lock); 3354 size_t size = ReservedSpace::page_align_size_down(bytes); 3355 if (size > 0) { 3356 shrink_by(size); 3357 } 3358 } 3359 3360 bool ConcurrentMarkSweepGeneration::grow_by(size_t bytes) { 3361 assert_locked_or_safepoint(Heap_lock); 3362 bool result = _virtual_space.expand_by(bytes); 3363 if (result) { 3364 size_t new_word_size = 3365 heap_word_size(_virtual_space.committed_size()); 3366 MemRegion mr(_cmsSpace->bottom(), new_word_size); 3367 _bts->resize(new_word_size); // resize the block offset shared array 3368 Universe::heap()->barrier_set()->resize_covered_region(mr); 3369 // Hmmmm... why doesn't CFLS::set_end verify locking? 3370 // This is quite ugly; FIX ME XXX 3371 _cmsSpace->assert_locked(freelistLock()); 3372 _cmsSpace->set_end((HeapWord*)_virtual_space.high()); 3373 3374 // update the space and generation capacity counters 3375 if (UsePerfData) { 3376 _space_counters->update_capacity(); 3377 _gen_counters->update_all(); 3378 } 3379 3380 if (Verbose && PrintGC) { 3381 size_t new_mem_size = _virtual_space.committed_size(); 3382 size_t old_mem_size = new_mem_size - bytes; 3383 gclog_or_tty->print_cr("Expanding %s from " SIZE_FORMAT "K by " SIZE_FORMAT "K to " SIZE_FORMAT "K", 3384 name(), old_mem_size/K, bytes/K, new_mem_size/K); 3385 } 3386 } 3387 return result; 3388 } 3389 3390 bool ConcurrentMarkSweepGeneration::grow_to_reserved() { 3391 assert_locked_or_safepoint(Heap_lock); 3392 bool success = true; 3393 const size_t remaining_bytes = _virtual_space.uncommitted_size(); 3394 if (remaining_bytes > 0) { 3395 success = grow_by(remaining_bytes); 3396 DEBUG_ONLY(if (!success) warning("grow to reserved failed");) 3397 } 3398 return success; 3399 } 3400 3401 void ConcurrentMarkSweepGeneration::shrink_free_list_by(size_t bytes) { 3402 assert_locked_or_safepoint(Heap_lock); 3403 assert_lock_strong(freelistLock()); 3404 if (PrintGCDetails && Verbose) { 3405 warning("Shrinking of CMS not yet implemented"); 3406 } 3407 return; 3408 } 3409 3410 3411 // Simple ctor/dtor wrapper for accounting & timer chores around concurrent 3412 // phases. 3413 class CMSPhaseAccounting: public StackObj { 3414 public: 3415 CMSPhaseAccounting(CMSCollector *collector, 3416 const char *phase, 3417 bool print_cr = true); 3418 ~CMSPhaseAccounting(); 3419 3420 private: 3421 CMSCollector *_collector; 3422 const char *_phase; 3423 elapsedTimer _wallclock; 3424 bool _print_cr; 3425 3426 public: 3427 // Not MT-safe; so do not pass around these StackObj's 3428 // where they may be accessed by other threads. 3429 jlong wallclock_millis() { 3430 assert(_wallclock.is_active(), "Wall clock should not stop"); 3431 _wallclock.stop(); // to record time 3432 jlong ret = _wallclock.milliseconds(); 3433 _wallclock.start(); // restart 3434 return ret; 3435 } 3436 }; 3437 3438 CMSPhaseAccounting::CMSPhaseAccounting(CMSCollector *collector, 3439 const char *phase, 3440 bool print_cr) : 3441 _collector(collector), _phase(phase), _print_cr(print_cr) { 3442 3443 if (PrintCMSStatistics != 0) { 3444 _collector->resetYields(); 3445 } 3446 if (PrintGCDetails) { 3447 gclog_or_tty->date_stamp(PrintGCDateStamps); 3448 gclog_or_tty->stamp(PrintGCTimeStamps); 3449 gclog_or_tty->print_cr("[%s-concurrent-%s-start]", 3450 _collector->cmsGen()->short_name(), _phase); 3451 } 3452 _collector->resetTimer(); 3453 _wallclock.start(); 3454 _collector->startTimer(); 3455 } 3456 3457 CMSPhaseAccounting::~CMSPhaseAccounting() { 3458 assert(_wallclock.is_active(), "Wall clock should not have stopped"); 3459 _collector->stopTimer(); 3460 _wallclock.stop(); 3461 if (PrintGCDetails) { 3462 gclog_or_tty->date_stamp(PrintGCDateStamps); 3463 gclog_or_tty->stamp(PrintGCTimeStamps); 3464 gclog_or_tty->print("[%s-concurrent-%s: %3.3f/%3.3f secs]", 3465 _collector->cmsGen()->short_name(), 3466 _phase, _collector->timerValue(), _wallclock.seconds()); 3467 if (_print_cr) { 3468 gclog_or_tty->print_cr(""); 3469 } 3470 if (PrintCMSStatistics != 0) { 3471 gclog_or_tty->print_cr(" (CMS-concurrent-%s yielded %d times)", _phase, 3472 _collector->yields()); 3473 } 3474 } 3475 } 3476 3477 // CMS work 3478 3479 // Checkpoint the roots into this generation from outside 3480 // this generation. [Note this initial checkpoint need only 3481 // be approximate -- we'll do a catch up phase subsequently.] 3482 void CMSCollector::checkpointRootsInitial(bool asynch) { 3483 assert(_collectorState == InitialMarking, "Wrong collector state"); 3484 check_correct_thread_executing(); 3485 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 3486 3487 ReferenceProcessor* rp = ref_processor(); 3488 SpecializationStats::clear(); 3489 assert(_restart_addr == NULL, "Control point invariant"); 3490 if (asynch) { 3491 // acquire locks for subsequent manipulations 3492 MutexLockerEx x(bitMapLock(), 3493 Mutex::_no_safepoint_check_flag); 3494 checkpointRootsInitialWork(asynch); 3495 // enable ("weak") refs discovery 3496 rp->enable_discovery(true /*verify_disabled*/, true /*check_no_refs*/); 3497 _collectorState = Marking; 3498 } else { 3499 // (Weak) Refs discovery: this is controlled from genCollectedHeap::do_collection 3500 // which recognizes if we are a CMS generation, and doesn't try to turn on 3501 // discovery; verify that they aren't meddling. 3502 assert(!rp->discovery_is_atomic(), 3503 "incorrect setting of discovery predicate"); 3504 assert(!rp->discovery_enabled(), "genCollectedHeap shouldn't control " 3505 "ref discovery for this generation kind"); 3506 // already have locks 3507 checkpointRootsInitialWork(asynch); 3508 // now enable ("weak") refs discovery 3509 rp->enable_discovery(true /*verify_disabled*/, false /*verify_no_refs*/); 3510 _collectorState = Marking; 3511 } 3512 SpecializationStats::print(); 3513 } 3514 3515 void CMSCollector::checkpointRootsInitialWork(bool asynch) { 3516 assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped"); 3517 assert(_collectorState == InitialMarking, "just checking"); 3518 3519 // If there has not been a GC[n-1] since last GC[n] cycle completed, 3520 // precede our marking with a collection of all 3521 // younger generations to keep floating garbage to a minimum. 3522 // XXX: we won't do this for now -- it's an optimization to be done later. 3523 3524 // already have locks 3525 assert_lock_strong(bitMapLock()); 3526 assert(_markBitMap.isAllClear(), "was reset at end of previous cycle"); 3527 3528 // Setup the verification and class unloading state for this 3529 // CMS collection cycle. 3530 setup_cms_unloading_and_verification_state(); 3531 3532 NOT_PRODUCT(TraceTime t("\ncheckpointRootsInitialWork", 3533 PrintGCDetails && Verbose, true, gclog_or_tty);) 3534 if (UseAdaptiveSizePolicy) { 3535 size_policy()->checkpoint_roots_initial_begin(); 3536 } 3537 3538 // Reset all the PLAB chunk arrays if necessary. 3539 if (_survivor_plab_array != NULL && !CMSPLABRecordAlways) { 3540 reset_survivor_plab_arrays(); 3541 } 3542 3543 ResourceMark rm; 3544 HandleMark hm; 3545 3546 FalseClosure falseClosure; 3547 // In the case of a synchronous collection, we will elide the 3548 // remark step, so it's important to catch all the nmethod oops 3549 // in this step. 3550 // The final 'true' flag to gen_process_strong_roots will ensure this. 3551 // If 'async' is true, we can relax the nmethod tracing. 3552 MarkRefsIntoClosure notOlder(_span, &_markBitMap); 3553 GenCollectedHeap* gch = GenCollectedHeap::heap(); 3554 3555 verify_work_stacks_empty(); 3556 verify_overflow_empty(); 3557 3558 gch->ensure_parsability(false); // fill TLABs, but no need to retire them 3559 // Update the saved marks which may affect the root scans. 3560 gch->save_marks(); 3561 3562 // weak reference processing has not started yet. 3563 ref_processor()->set_enqueuing_is_done(false); 3564 3565 // Need to remember all newly created CLDs, 3566 // so that we can guarantee that the remark finds them. 3567 ClassLoaderDataGraph::remember_new_clds(true); 3568 3569 // Whenever a CLD is found, it will be claimed before proceeding to mark 3570 // the klasses. The claimed marks need to be cleared before marking starts. 3571 ClassLoaderDataGraph::clear_claimed_marks(); 3572 3573 CMKlassClosure klass_closure(¬Older); 3574 { 3575 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;) 3576 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 3577 gch->gen_process_strong_roots(_cmsGen->level(), 3578 true, // younger gens are roots 3579 true, // activate StrongRootsScope 3580 false, // not scavenging 3581 SharedHeap::ScanningOption(roots_scanning_options()), 3582 ¬Older, 3583 true, // walk all of code cache if (so & SO_CodeCache) 3584 NULL, 3585 &klass_closure); 3586 } 3587 3588 // Clear mod-union table; it will be dirtied in the prologue of 3589 // CMS generation per each younger generation collection. 3590 3591 assert(_modUnionTable.isAllClear(), 3592 "Was cleared in most recent final checkpoint phase" 3593 " or no bits are set in the gc_prologue before the start of the next " 3594 "subsequent marking phase."); 3595 3596 assert(_ct->klass_rem_set()->mod_union_is_clear(), "Must be"); 3597 3598 // Save the end of the used_region of the constituent generations 3599 // to be used to limit the extent of sweep in each generation. 3600 save_sweep_limits(); 3601 if (UseAdaptiveSizePolicy) { 3602 size_policy()->checkpoint_roots_initial_end(gch->gc_cause()); 3603 } 3604 verify_overflow_empty(); 3605 } 3606 3607 bool CMSCollector::markFromRoots(bool asynch) { 3608 // we might be tempted to assert that: 3609 // assert(asynch == !SafepointSynchronize::is_at_safepoint(), 3610 // "inconsistent argument?"); 3611 // However that wouldn't be right, because it's possible that 3612 // a safepoint is indeed in progress as a younger generation 3613 // stop-the-world GC happens even as we mark in this generation. 3614 assert(_collectorState == Marking, "inconsistent state?"); 3615 check_correct_thread_executing(); 3616 verify_overflow_empty(); 3617 3618 bool res; 3619 if (asynch) { 3620 3621 // Start the timers for adaptive size policy for the concurrent phases 3622 // Do it here so that the foreground MS can use the concurrent 3623 // timer since a foreground MS might has the sweep done concurrently 3624 // or STW. 3625 if (UseAdaptiveSizePolicy) { 3626 size_policy()->concurrent_marking_begin(); 3627 } 3628 3629 // Weak ref discovery note: We may be discovering weak 3630 // refs in this generation concurrent (but interleaved) with 3631 // weak ref discovery by a younger generation collector. 3632 3633 CMSTokenSyncWithLocks ts(true, bitMapLock()); 3634 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 3635 CMSPhaseAccounting pa(this, "mark", !PrintGCDetails); 3636 res = markFromRootsWork(asynch); 3637 if (res) { 3638 _collectorState = Precleaning; 3639 } else { // We failed and a foreground collection wants to take over 3640 assert(_foregroundGCIsActive, "internal state inconsistency"); 3641 assert(_restart_addr == NULL, "foreground will restart from scratch"); 3642 if (PrintGCDetails) { 3643 gclog_or_tty->print_cr("bailing out to foreground collection"); 3644 } 3645 } 3646 if (UseAdaptiveSizePolicy) { 3647 size_policy()->concurrent_marking_end(); 3648 } 3649 } else { 3650 assert(SafepointSynchronize::is_at_safepoint(), 3651 "inconsistent with asynch == false"); 3652 if (UseAdaptiveSizePolicy) { 3653 size_policy()->ms_collection_marking_begin(); 3654 } 3655 // already have locks 3656 res = markFromRootsWork(asynch); 3657 _collectorState = FinalMarking; 3658 if (UseAdaptiveSizePolicy) { 3659 GenCollectedHeap* gch = GenCollectedHeap::heap(); 3660 size_policy()->ms_collection_marking_end(gch->gc_cause()); 3661 } 3662 } 3663 verify_overflow_empty(); 3664 return res; 3665 } 3666 3667 bool CMSCollector::markFromRootsWork(bool asynch) { 3668 // iterate over marked bits in bit map, doing a full scan and mark 3669 // from these roots using the following algorithm: 3670 // . if oop is to the right of the current scan pointer, 3671 // mark corresponding bit (we'll process it later) 3672 // . else (oop is to left of current scan pointer) 3673 // push oop on marking stack 3674 // . drain the marking stack 3675 3676 // Note that when we do a marking step we need to hold the 3677 // bit map lock -- recall that direct allocation (by mutators) 3678 // and promotion (by younger generation collectors) is also 3679 // marking the bit map. [the so-called allocate live policy.] 3680 // Because the implementation of bit map marking is not 3681 // robust wrt simultaneous marking of bits in the same word, 3682 // we need to make sure that there is no such interference 3683 // between concurrent such updates. 3684 3685 // already have locks 3686 assert_lock_strong(bitMapLock()); 3687 3688 verify_work_stacks_empty(); 3689 verify_overflow_empty(); 3690 bool result = false; 3691 if (CMSConcurrentMTEnabled && ConcGCThreads > 0) { 3692 result = do_marking_mt(asynch); 3693 } else { 3694 result = do_marking_st(asynch); 3695 } 3696 return result; 3697 } 3698 3699 // Forward decl 3700 class CMSConcMarkingTask; 3701 3702 class CMSConcMarkingTerminator: public ParallelTaskTerminator { 3703 CMSCollector* _collector; 3704 CMSConcMarkingTask* _task; 3705 public: 3706 virtual void yield(); 3707 3708 // "n_threads" is the number of threads to be terminated. 3709 // "queue_set" is a set of work queues of other threads. 3710 // "collector" is the CMS collector associated with this task terminator. 3711 // "yield" indicates whether we need the gang as a whole to yield. 3712 CMSConcMarkingTerminator(int n_threads, TaskQueueSetSuper* queue_set, CMSCollector* collector) : 3713 ParallelTaskTerminator(n_threads, queue_set), 3714 _collector(collector) { } 3715 3716 void set_task(CMSConcMarkingTask* task) { 3717 _task = task; 3718 } 3719 }; 3720 3721 class CMSConcMarkingTerminatorTerminator: public TerminatorTerminator { 3722 CMSConcMarkingTask* _task; 3723 public: 3724 bool should_exit_termination(); 3725 void set_task(CMSConcMarkingTask* task) { 3726 _task = task; 3727 } 3728 }; 3729 3730 // MT Concurrent Marking Task 3731 class CMSConcMarkingTask: public YieldingFlexibleGangTask { 3732 CMSCollector* _collector; 3733 int _n_workers; // requested/desired # workers 3734 bool _asynch; 3735 bool _result; 3736 CompactibleFreeListSpace* _cms_space; 3737 char _pad_front[64]; // padding to ... 3738 HeapWord* _global_finger; // ... avoid sharing cache line 3739 char _pad_back[64]; 3740 HeapWord* _restart_addr; 3741 3742 // Exposed here for yielding support 3743 Mutex* const _bit_map_lock; 3744 3745 // The per thread work queues, available here for stealing 3746 OopTaskQueueSet* _task_queues; 3747 3748 // Termination (and yielding) support 3749 CMSConcMarkingTerminator _term; 3750 CMSConcMarkingTerminatorTerminator _term_term; 3751 3752 public: 3753 CMSConcMarkingTask(CMSCollector* collector, 3754 CompactibleFreeListSpace* cms_space, 3755 bool asynch, 3756 YieldingFlexibleWorkGang* workers, 3757 OopTaskQueueSet* task_queues): 3758 YieldingFlexibleGangTask("Concurrent marking done multi-threaded"), 3759 _collector(collector), 3760 _cms_space(cms_space), 3761 _asynch(asynch), _n_workers(0), _result(true), 3762 _task_queues(task_queues), 3763 _term(_n_workers, task_queues, _collector), 3764 _bit_map_lock(collector->bitMapLock()) 3765 { 3766 _requested_size = _n_workers; 3767 _term.set_task(this); 3768 _term_term.set_task(this); 3769 _restart_addr = _global_finger = _cms_space->bottom(); 3770 } 3771 3772 3773 OopTaskQueueSet* task_queues() { return _task_queues; } 3774 3775 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 3776 3777 HeapWord** global_finger_addr() { return &_global_finger; } 3778 3779 CMSConcMarkingTerminator* terminator() { return &_term; } 3780 3781 virtual void set_for_termination(int active_workers) { 3782 terminator()->reset_for_reuse(active_workers); 3783 } 3784 3785 void work(uint worker_id); 3786 bool should_yield() { 3787 return ConcurrentMarkSweepThread::should_yield() 3788 && !_collector->foregroundGCIsActive() 3789 && _asynch; 3790 } 3791 3792 virtual void coordinator_yield(); // stuff done by coordinator 3793 bool result() { return _result; } 3794 3795 void reset(HeapWord* ra) { 3796 assert(_global_finger >= _cms_space->end(), "Postcondition of ::work(i)"); 3797 _restart_addr = _global_finger = ra; 3798 _term.reset_for_reuse(); 3799 } 3800 3801 static bool get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 3802 OopTaskQueue* work_q); 3803 3804 private: 3805 void do_scan_and_mark(int i, CompactibleFreeListSpace* sp); 3806 void do_work_steal(int i); 3807 void bump_global_finger(HeapWord* f); 3808 }; 3809 3810 bool CMSConcMarkingTerminatorTerminator::should_exit_termination() { 3811 assert(_task != NULL, "Error"); 3812 return _task->yielding(); 3813 // Note that we do not need the disjunct || _task->should_yield() above 3814 // because we want terminating threads to yield only if the task 3815 // is already in the midst of yielding, which happens only after at least one 3816 // thread has yielded. 3817 } 3818 3819 void CMSConcMarkingTerminator::yield() { 3820 if (_task->should_yield()) { 3821 _task->yield(); 3822 } else { 3823 ParallelTaskTerminator::yield(); 3824 } 3825 } 3826 3827 //////////////////////////////////////////////////////////////// 3828 // Concurrent Marking Algorithm Sketch 3829 //////////////////////////////////////////////////////////////// 3830 // Until all tasks exhausted (both spaces): 3831 // -- claim next available chunk 3832 // -- bump global finger via CAS 3833 // -- find first object that starts in this chunk 3834 // and start scanning bitmap from that position 3835 // -- scan marked objects for oops 3836 // -- CAS-mark target, and if successful: 3837 // . if target oop is above global finger (volatile read) 3838 // nothing to do 3839 // . if target oop is in chunk and above local finger 3840 // then nothing to do 3841 // . else push on work-queue 3842 // -- Deal with possible overflow issues: 3843 // . local work-queue overflow causes stuff to be pushed on 3844 // global (common) overflow queue 3845 // . always first empty local work queue 3846 // . then get a batch of oops from global work queue if any 3847 // . then do work stealing 3848 // -- When all tasks claimed (both spaces) 3849 // and local work queue empty, 3850 // then in a loop do: 3851 // . check global overflow stack; steal a batch of oops and trace 3852 // . try to steal from other threads oif GOS is empty 3853 // . if neither is available, offer termination 3854 // -- Terminate and return result 3855 // 3856 void CMSConcMarkingTask::work(uint worker_id) { 3857 elapsedTimer _timer; 3858 ResourceMark rm; 3859 HandleMark hm; 3860 3861 DEBUG_ONLY(_collector->verify_overflow_empty();) 3862 3863 // Before we begin work, our work queue should be empty 3864 assert(work_queue(worker_id)->size() == 0, "Expected to be empty"); 3865 // Scan the bitmap covering _cms_space, tracing through grey objects. 3866 _timer.start(); 3867 do_scan_and_mark(worker_id, _cms_space); 3868 _timer.stop(); 3869 if (PrintCMSStatistics != 0) { 3870 gclog_or_tty->print_cr("Finished cms space scanning in %dth thread: %3.3f sec", 3871 worker_id, _timer.seconds()); 3872 // XXX: need xxx/xxx type of notation, two timers 3873 } 3874 3875 // ... do work stealing 3876 _timer.reset(); 3877 _timer.start(); 3878 do_work_steal(worker_id); 3879 _timer.stop(); 3880 if (PrintCMSStatistics != 0) { 3881 gclog_or_tty->print_cr("Finished work stealing in %dth thread: %3.3f sec", 3882 worker_id, _timer.seconds()); 3883 // XXX: need xxx/xxx type of notation, two timers 3884 } 3885 assert(_collector->_markStack.isEmpty(), "Should have been emptied"); 3886 assert(work_queue(worker_id)->size() == 0, "Should have been emptied"); 3887 // Note that under the current task protocol, the 3888 // following assertion is true even of the spaces 3889 // expanded since the completion of the concurrent 3890 // marking. XXX This will likely change under a strict 3891 // ABORT semantics. 3892 // After perm removal the comparison was changed to 3893 // greater than or equal to from strictly greater than. 3894 // Before perm removal the highest address sweep would 3895 // have been at the end of perm gen but now is at the 3896 // end of the tenured gen. 3897 assert(_global_finger >= _cms_space->end(), 3898 "All tasks have been completed"); 3899 DEBUG_ONLY(_collector->verify_overflow_empty();) 3900 } 3901 3902 void CMSConcMarkingTask::bump_global_finger(HeapWord* f) { 3903 HeapWord* read = _global_finger; 3904 HeapWord* cur = read; 3905 while (f > read) { 3906 cur = read; 3907 read = (HeapWord*) Atomic::cmpxchg_ptr(f, &_global_finger, cur); 3908 if (cur == read) { 3909 // our cas succeeded 3910 assert(_global_finger >= f, "protocol consistency"); 3911 break; 3912 } 3913 } 3914 } 3915 3916 // This is really inefficient, and should be redone by 3917 // using (not yet available) block-read and -write interfaces to the 3918 // stack and the work_queue. XXX FIX ME !!! 3919 bool CMSConcMarkingTask::get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 3920 OopTaskQueue* work_q) { 3921 // Fast lock-free check 3922 if (ovflw_stk->length() == 0) { 3923 return false; 3924 } 3925 assert(work_q->size() == 0, "Shouldn't steal"); 3926 MutexLockerEx ml(ovflw_stk->par_lock(), 3927 Mutex::_no_safepoint_check_flag); 3928 // Grab up to 1/4 the size of the work queue 3929 size_t num = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 3930 (size_t)ParGCDesiredObjsFromOverflowList); 3931 num = MIN2(num, ovflw_stk->length()); 3932 for (int i = (int) num; i > 0; i--) { 3933 oop cur = ovflw_stk->pop(); 3934 assert(cur != NULL, "Counted wrong?"); 3935 work_q->push(cur); 3936 } 3937 return num > 0; 3938 } 3939 3940 void CMSConcMarkingTask::do_scan_and_mark(int i, CompactibleFreeListSpace* sp) { 3941 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 3942 int n_tasks = pst->n_tasks(); 3943 // We allow that there may be no tasks to do here because 3944 // we are restarting after a stack overflow. 3945 assert(pst->valid() || n_tasks == 0, "Uninitialized use?"); 3946 uint nth_task = 0; 3947 3948 HeapWord* aligned_start = sp->bottom(); 3949 if (sp->used_region().contains(_restart_addr)) { 3950 // Align down to a card boundary for the start of 0th task 3951 // for this space. 3952 aligned_start = 3953 (HeapWord*)align_size_down((uintptr_t)_restart_addr, 3954 CardTableModRefBS::card_size); 3955 } 3956 3957 size_t chunk_size = sp->marking_task_size(); 3958 while (!pst->is_task_claimed(/* reference */ nth_task)) { 3959 // Having claimed the nth task in this space, 3960 // compute the chunk that it corresponds to: 3961 MemRegion span = MemRegion(aligned_start + nth_task*chunk_size, 3962 aligned_start + (nth_task+1)*chunk_size); 3963 // Try and bump the global finger via a CAS; 3964 // note that we need to do the global finger bump 3965 // _before_ taking the intersection below, because 3966 // the task corresponding to that region will be 3967 // deemed done even if the used_region() expands 3968 // because of allocation -- as it almost certainly will 3969 // during start-up while the threads yield in the 3970 // closure below. 3971 HeapWord* finger = span.end(); 3972 bump_global_finger(finger); // atomically 3973 // There are null tasks here corresponding to chunks 3974 // beyond the "top" address of the space. 3975 span = span.intersection(sp->used_region()); 3976 if (!span.is_empty()) { // Non-null task 3977 HeapWord* prev_obj; 3978 assert(!span.contains(_restart_addr) || nth_task == 0, 3979 "Inconsistency"); 3980 if (nth_task == 0) { 3981 // For the 0th task, we'll not need to compute a block_start. 3982 if (span.contains(_restart_addr)) { 3983 // In the case of a restart because of stack overflow, 3984 // we might additionally skip a chunk prefix. 3985 prev_obj = _restart_addr; 3986 } else { 3987 prev_obj = span.start(); 3988 } 3989 } else { 3990 // We want to skip the first object because 3991 // the protocol is to scan any object in its entirety 3992 // that _starts_ in this span; a fortiori, any 3993 // object starting in an earlier span is scanned 3994 // as part of an earlier claimed task. 3995 // Below we use the "careful" version of block_start 3996 // so we do not try to navigate uninitialized objects. 3997 prev_obj = sp->block_start_careful(span.start()); 3998 // Below we use a variant of block_size that uses the 3999 // Printezis bits to avoid waiting for allocated 4000 // objects to become initialized/parsable. 4001 while (prev_obj < span.start()) { 4002 size_t sz = sp->block_size_no_stall(prev_obj, _collector); 4003 if (sz > 0) { 4004 prev_obj += sz; 4005 } else { 4006 // In this case we may end up doing a bit of redundant 4007 // scanning, but that appears unavoidable, short of 4008 // locking the free list locks; see bug 6324141. 4009 break; 4010 } 4011 } 4012 } 4013 if (prev_obj < span.end()) { 4014 MemRegion my_span = MemRegion(prev_obj, span.end()); 4015 // Do the marking work within a non-empty span -- 4016 // the last argument to the constructor indicates whether the 4017 // iteration should be incremental with periodic yields. 4018 Par_MarkFromRootsClosure cl(this, _collector, my_span, 4019 &_collector->_markBitMap, 4020 work_queue(i), 4021 &_collector->_markStack, 4022 _asynch); 4023 _collector->_markBitMap.iterate(&cl, my_span.start(), my_span.end()); 4024 } // else nothing to do for this task 4025 } // else nothing to do for this task 4026 } 4027 // We'd be tempted to assert here that since there are no 4028 // more tasks left to claim in this space, the global_finger 4029 // must exceed space->top() and a fortiori space->end(). However, 4030 // that would not quite be correct because the bumping of 4031 // global_finger occurs strictly after the claiming of a task, 4032 // so by the time we reach here the global finger may not yet 4033 // have been bumped up by the thread that claimed the last 4034 // task. 4035 pst->all_tasks_completed(); 4036 } 4037 4038 class Par_ConcMarkingClosure: public CMSOopClosure { 4039 private: 4040 CMSCollector* _collector; 4041 CMSConcMarkingTask* _task; 4042 MemRegion _span; 4043 CMSBitMap* _bit_map; 4044 CMSMarkStack* _overflow_stack; 4045 OopTaskQueue* _work_queue; 4046 protected: 4047 DO_OOP_WORK_DEFN 4048 public: 4049 Par_ConcMarkingClosure(CMSCollector* collector, CMSConcMarkingTask* task, OopTaskQueue* work_queue, 4050 CMSBitMap* bit_map, CMSMarkStack* overflow_stack): 4051 CMSOopClosure(collector->ref_processor()), 4052 _collector(collector), 4053 _task(task), 4054 _span(collector->_span), 4055 _work_queue(work_queue), 4056 _bit_map(bit_map), 4057 _overflow_stack(overflow_stack) 4058 { } 4059 virtual void do_oop(oop* p); 4060 virtual void do_oop(narrowOop* p); 4061 4062 void trim_queue(size_t max); 4063 void handle_stack_overflow(HeapWord* lost); 4064 void do_yield_check() { 4065 if (_task->should_yield()) { 4066 _task->yield(); 4067 } 4068 } 4069 }; 4070 4071 // Grey object scanning during work stealing phase -- 4072 // the salient assumption here is that any references 4073 // that are in these stolen objects being scanned must 4074 // already have been initialized (else they would not have 4075 // been published), so we do not need to check for 4076 // uninitialized objects before pushing here. 4077 void Par_ConcMarkingClosure::do_oop(oop obj) { 4078 assert(obj->is_oop_or_null(true), "expected an oop or NULL"); 4079 HeapWord* addr = (HeapWord*)obj; 4080 // Check if oop points into the CMS generation 4081 // and is not marked 4082 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 4083 // a white object ... 4084 // If we manage to "claim" the object, by being the 4085 // first thread to mark it, then we push it on our 4086 // marking stack 4087 if (_bit_map->par_mark(addr)) { // ... now grey 4088 // push on work queue (grey set) 4089 bool simulate_overflow = false; 4090 NOT_PRODUCT( 4091 if (CMSMarkStackOverflowALot && 4092 _collector->simulate_overflow()) { 4093 // simulate a stack overflow 4094 simulate_overflow = true; 4095 } 4096 ) 4097 if (simulate_overflow || 4098 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) { 4099 // stack overflow 4100 if (PrintCMSStatistics != 0) { 4101 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 4102 SIZE_FORMAT, _overflow_stack->capacity()); 4103 } 4104 // We cannot assert that the overflow stack is full because 4105 // it may have been emptied since. 4106 assert(simulate_overflow || 4107 _work_queue->size() == _work_queue->max_elems(), 4108 "Else push should have succeeded"); 4109 handle_stack_overflow(addr); 4110 } 4111 } // Else, some other thread got there first 4112 do_yield_check(); 4113 } 4114 } 4115 4116 void Par_ConcMarkingClosure::do_oop(oop* p) { Par_ConcMarkingClosure::do_oop_work(p); } 4117 void Par_ConcMarkingClosure::do_oop(narrowOop* p) { Par_ConcMarkingClosure::do_oop_work(p); } 4118 4119 void Par_ConcMarkingClosure::trim_queue(size_t max) { 4120 while (_work_queue->size() > max) { 4121 oop new_oop; 4122 if (_work_queue->pop_local(new_oop)) { 4123 assert(new_oop->is_oop(), "Should be an oop"); 4124 assert(_bit_map->isMarked((HeapWord*)new_oop), "Grey object"); 4125 assert(_span.contains((HeapWord*)new_oop), "Not in span"); 4126 new_oop->oop_iterate(this); // do_oop() above 4127 do_yield_check(); 4128 } 4129 } 4130 } 4131 4132 // Upon stack overflow, we discard (part of) the stack, 4133 // remembering the least address amongst those discarded 4134 // in CMSCollector's _restart_address. 4135 void Par_ConcMarkingClosure::handle_stack_overflow(HeapWord* lost) { 4136 // We need to do this under a mutex to prevent other 4137 // workers from interfering with the work done below. 4138 MutexLockerEx ml(_overflow_stack->par_lock(), 4139 Mutex::_no_safepoint_check_flag); 4140 // Remember the least grey address discarded 4141 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 4142 _collector->lower_restart_addr(ra); 4143 _overflow_stack->reset(); // discard stack contents 4144 _overflow_stack->expand(); // expand the stack if possible 4145 } 4146 4147 4148 void CMSConcMarkingTask::do_work_steal(int i) { 4149 OopTaskQueue* work_q = work_queue(i); 4150 oop obj_to_scan; 4151 CMSBitMap* bm = &(_collector->_markBitMap); 4152 CMSMarkStack* ovflw = &(_collector->_markStack); 4153 int* seed = _collector->hash_seed(i); 4154 Par_ConcMarkingClosure cl(_collector, this, work_q, bm, ovflw); 4155 while (true) { 4156 cl.trim_queue(0); 4157 assert(work_q->size() == 0, "Should have been emptied above"); 4158 if (get_work_from_overflow_stack(ovflw, work_q)) { 4159 // Can't assert below because the work obtained from the 4160 // overflow stack may already have been stolen from us. 4161 // assert(work_q->size() > 0, "Work from overflow stack"); 4162 continue; 4163 } else if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 4164 assert(obj_to_scan->is_oop(), "Should be an oop"); 4165 assert(bm->isMarked((HeapWord*)obj_to_scan), "Grey object"); 4166 obj_to_scan->oop_iterate(&cl); 4167 } else if (terminator()->offer_termination(&_term_term)) { 4168 assert(work_q->size() == 0, "Impossible!"); 4169 break; 4170 } else if (yielding() || should_yield()) { 4171 yield(); 4172 } 4173 } 4174 } 4175 4176 // This is run by the CMS (coordinator) thread. 4177 void CMSConcMarkingTask::coordinator_yield() { 4178 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 4179 "CMS thread should hold CMS token"); 4180 // First give up the locks, then yield, then re-lock 4181 // We should probably use a constructor/destructor idiom to 4182 // do this unlock/lock or modify the MutexUnlocker class to 4183 // serve our purpose. XXX 4184 assert_lock_strong(_bit_map_lock); 4185 _bit_map_lock->unlock(); 4186 ConcurrentMarkSweepThread::desynchronize(true); 4187 ConcurrentMarkSweepThread::acknowledge_yield_request(); 4188 _collector->stopTimer(); 4189 if (PrintCMSStatistics != 0) { 4190 _collector->incrementYields(); 4191 } 4192 _collector->icms_wait(); 4193 4194 // It is possible for whichever thread initiated the yield request 4195 // not to get a chance to wake up and take the bitmap lock between 4196 // this thread releasing it and reacquiring it. So, while the 4197 // should_yield() flag is on, let's sleep for a bit to give the 4198 // other thread a chance to wake up. The limit imposed on the number 4199 // of iterations is defensive, to avoid any unforseen circumstances 4200 // putting us into an infinite loop. Since it's always been this 4201 // (coordinator_yield()) method that was observed to cause the 4202 // problem, we are using a parameter (CMSCoordinatorYieldSleepCount) 4203 // which is by default non-zero. For the other seven methods that 4204 // also perform the yield operation, as are using a different 4205 // parameter (CMSYieldSleepCount) which is by default zero. This way we 4206 // can enable the sleeping for those methods too, if necessary. 4207 // See 6442774. 4208 // 4209 // We really need to reconsider the synchronization between the GC 4210 // thread and the yield-requesting threads in the future and we 4211 // should really use wait/notify, which is the recommended 4212 // way of doing this type of interaction. Additionally, we should 4213 // consolidate the eight methods that do the yield operation and they 4214 // are almost identical into one for better maintenability and 4215 // readability. See 6445193. 4216 // 4217 // Tony 2006.06.29 4218 for (unsigned i = 0; i < CMSCoordinatorYieldSleepCount && 4219 ConcurrentMarkSweepThread::should_yield() && 4220 !CMSCollector::foregroundGCIsActive(); ++i) { 4221 os::sleep(Thread::current(), 1, false); 4222 ConcurrentMarkSweepThread::acknowledge_yield_request(); 4223 } 4224 4225 ConcurrentMarkSweepThread::synchronize(true); 4226 _bit_map_lock->lock_without_safepoint_check(); 4227 _collector->startTimer(); 4228 } 4229 4230 bool CMSCollector::do_marking_mt(bool asynch) { 4231 assert(ConcGCThreads > 0 && conc_workers() != NULL, "precondition"); 4232 int num_workers = AdaptiveSizePolicy::calc_active_conc_workers( 4233 conc_workers()->total_workers(), 4234 conc_workers()->active_workers(), 4235 Threads::number_of_non_daemon_threads()); 4236 conc_workers()->set_active_workers(num_workers); 4237 4238 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 4239 4240 CMSConcMarkingTask tsk(this, 4241 cms_space, 4242 asynch, 4243 conc_workers(), 4244 task_queues()); 4245 4246 // Since the actual number of workers we get may be different 4247 // from the number we requested above, do we need to do anything different 4248 // below? In particular, may be we need to subclass the SequantialSubTasksDone 4249 // class?? XXX 4250 cms_space ->initialize_sequential_subtasks_for_marking(num_workers); 4251 4252 // Refs discovery is already non-atomic. 4253 assert(!ref_processor()->discovery_is_atomic(), "Should be non-atomic"); 4254 assert(ref_processor()->discovery_is_mt(), "Discovery should be MT"); 4255 conc_workers()->start_task(&tsk); 4256 while (tsk.yielded()) { 4257 tsk.coordinator_yield(); 4258 conc_workers()->continue_task(&tsk); 4259 } 4260 // If the task was aborted, _restart_addr will be non-NULL 4261 assert(tsk.completed() || _restart_addr != NULL, "Inconsistency"); 4262 while (_restart_addr != NULL) { 4263 // XXX For now we do not make use of ABORTED state and have not 4264 // yet implemented the right abort semantics (even in the original 4265 // single-threaded CMS case). That needs some more investigation 4266 // and is deferred for now; see CR# TBF. 07252005YSR. XXX 4267 assert(!CMSAbortSemantics || tsk.aborted(), "Inconsistency"); 4268 // If _restart_addr is non-NULL, a marking stack overflow 4269 // occurred; we need to do a fresh marking iteration from the 4270 // indicated restart address. 4271 if (_foregroundGCIsActive && asynch) { 4272 // We may be running into repeated stack overflows, having 4273 // reached the limit of the stack size, while making very 4274 // slow forward progress. It may be best to bail out and 4275 // let the foreground collector do its job. 4276 // Clear _restart_addr, so that foreground GC 4277 // works from scratch. This avoids the headache of 4278 // a "rescan" which would otherwise be needed because 4279 // of the dirty mod union table & card table. 4280 _restart_addr = NULL; 4281 return false; 4282 } 4283 // Adjust the task to restart from _restart_addr 4284 tsk.reset(_restart_addr); 4285 cms_space ->initialize_sequential_subtasks_for_marking(num_workers, 4286 _restart_addr); 4287 _restart_addr = NULL; 4288 // Get the workers going again 4289 conc_workers()->start_task(&tsk); 4290 while (tsk.yielded()) { 4291 tsk.coordinator_yield(); 4292 conc_workers()->continue_task(&tsk); 4293 } 4294 } 4295 assert(tsk.completed(), "Inconsistency"); 4296 assert(tsk.result() == true, "Inconsistency"); 4297 return true; 4298 } 4299 4300 bool CMSCollector::do_marking_st(bool asynch) { 4301 ResourceMark rm; 4302 HandleMark hm; 4303 4304 // Temporarily make refs discovery single threaded (non-MT) 4305 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false); 4306 MarkFromRootsClosure markFromRootsClosure(this, _span, &_markBitMap, 4307 &_markStack, CMSYield && asynch); 4308 // the last argument to iterate indicates whether the iteration 4309 // should be incremental with periodic yields. 4310 _markBitMap.iterate(&markFromRootsClosure); 4311 // If _restart_addr is non-NULL, a marking stack overflow 4312 // occurred; we need to do a fresh iteration from the 4313 // indicated restart address. 4314 while (_restart_addr != NULL) { 4315 if (_foregroundGCIsActive && asynch) { 4316 // We may be running into repeated stack overflows, having 4317 // reached the limit of the stack size, while making very 4318 // slow forward progress. It may be best to bail out and 4319 // let the foreground collector do its job. 4320 // Clear _restart_addr, so that foreground GC 4321 // works from scratch. This avoids the headache of 4322 // a "rescan" which would otherwise be needed because 4323 // of the dirty mod union table & card table. 4324 _restart_addr = NULL; 4325 return false; // indicating failure to complete marking 4326 } 4327 // Deal with stack overflow: 4328 // we restart marking from _restart_addr 4329 HeapWord* ra = _restart_addr; 4330 markFromRootsClosure.reset(ra); 4331 _restart_addr = NULL; 4332 _markBitMap.iterate(&markFromRootsClosure, ra, _span.end()); 4333 } 4334 return true; 4335 } 4336 4337 void CMSCollector::preclean() { 4338 check_correct_thread_executing(); 4339 assert(Thread::current()->is_ConcurrentGC_thread(), "Wrong thread"); 4340 verify_work_stacks_empty(); 4341 verify_overflow_empty(); 4342 _abort_preclean = false; 4343 if (CMSPrecleaningEnabled) { 4344 _eden_chunk_index = 0; 4345 size_t used = get_eden_used(); 4346 size_t capacity = get_eden_capacity(); 4347 // Don't start sampling unless we will get sufficiently 4348 // many samples. 4349 if (used < (capacity/(CMSScheduleRemarkSamplingRatio * 100) 4350 * CMSScheduleRemarkEdenPenetration)) { 4351 _start_sampling = true; 4352 } else { 4353 _start_sampling = false; 4354 } 4355 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 4356 CMSPhaseAccounting pa(this, "preclean", !PrintGCDetails); 4357 preclean_work(CMSPrecleanRefLists1, CMSPrecleanSurvivors1); 4358 } 4359 CMSTokenSync x(true); // is cms thread 4360 if (CMSPrecleaningEnabled) { 4361 sample_eden(); 4362 _collectorState = AbortablePreclean; 4363 } else { 4364 _collectorState = FinalMarking; 4365 } 4366 verify_work_stacks_empty(); 4367 verify_overflow_empty(); 4368 } 4369 4370 // Try and schedule the remark such that young gen 4371 // occupancy is CMSScheduleRemarkEdenPenetration %. 4372 void CMSCollector::abortable_preclean() { 4373 check_correct_thread_executing(); 4374 assert(CMSPrecleaningEnabled, "Inconsistent control state"); 4375 assert(_collectorState == AbortablePreclean, "Inconsistent control state"); 4376 4377 // If Eden's current occupancy is below this threshold, 4378 // immediately schedule the remark; else preclean 4379 // past the next scavenge in an effort to 4380 // schedule the pause as described avove. By choosing 4381 // CMSScheduleRemarkEdenSizeThreshold >= max eden size 4382 // we will never do an actual abortable preclean cycle. 4383 if (get_eden_used() > CMSScheduleRemarkEdenSizeThreshold) { 4384 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 4385 CMSPhaseAccounting pa(this, "abortable-preclean", !PrintGCDetails); 4386 // We need more smarts in the abortable preclean 4387 // loop below to deal with cases where allocation 4388 // in young gen is very very slow, and our precleaning 4389 // is running a losing race against a horde of 4390 // mutators intent on flooding us with CMS updates 4391 // (dirty cards). 4392 // One, admittedly dumb, strategy is to give up 4393 // after a certain number of abortable precleaning loops 4394 // or after a certain maximum time. We want to make 4395 // this smarter in the next iteration. 4396 // XXX FIX ME!!! YSR 4397 size_t loops = 0, workdone = 0, cumworkdone = 0, waited = 0; 4398 while (!(should_abort_preclean() || 4399 ConcurrentMarkSweepThread::should_terminate())) { 4400 workdone = preclean_work(CMSPrecleanRefLists2, CMSPrecleanSurvivors2); 4401 cumworkdone += workdone; 4402 loops++; 4403 // Voluntarily terminate abortable preclean phase if we have 4404 // been at it for too long. 4405 if ((CMSMaxAbortablePrecleanLoops != 0) && 4406 loops >= CMSMaxAbortablePrecleanLoops) { 4407 if (PrintGCDetails) { 4408 gclog_or_tty->print(" CMS: abort preclean due to loops "); 4409 } 4410 break; 4411 } 4412 if (pa.wallclock_millis() > CMSMaxAbortablePrecleanTime) { 4413 if (PrintGCDetails) { 4414 gclog_or_tty->print(" CMS: abort preclean due to time "); 4415 } 4416 break; 4417 } 4418 // If we are doing little work each iteration, we should 4419 // take a short break. 4420 if (workdone < CMSAbortablePrecleanMinWorkPerIteration) { 4421 // Sleep for some time, waiting for work to accumulate 4422 stopTimer(); 4423 cmsThread()->wait_on_cms_lock(CMSAbortablePrecleanWaitMillis); 4424 startTimer(); 4425 waited++; 4426 } 4427 } 4428 if (PrintCMSStatistics > 0) { 4429 gclog_or_tty->print(" [%d iterations, %d waits, %d cards)] ", 4430 loops, waited, cumworkdone); 4431 } 4432 } 4433 CMSTokenSync x(true); // is cms thread 4434 if (_collectorState != Idling) { 4435 assert(_collectorState == AbortablePreclean, 4436 "Spontaneous state transition?"); 4437 _collectorState = FinalMarking; 4438 } // Else, a foreground collection completed this CMS cycle. 4439 return; 4440 } 4441 4442 // Respond to an Eden sampling opportunity 4443 void CMSCollector::sample_eden() { 4444 // Make sure a young gc cannot sneak in between our 4445 // reading and recording of a sample. 4446 assert(Thread::current()->is_ConcurrentGC_thread(), 4447 "Only the cms thread may collect Eden samples"); 4448 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 4449 "Should collect samples while holding CMS token"); 4450 if (!_start_sampling) { 4451 return; 4452 } 4453 if (_eden_chunk_array) { 4454 if (_eden_chunk_index < _eden_chunk_capacity) { 4455 _eden_chunk_array[_eden_chunk_index] = *_top_addr; // take sample 4456 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr, 4457 "Unexpected state of Eden"); 4458 // We'd like to check that what we just sampled is an oop-start address; 4459 // however, we cannot do that here since the object may not yet have been 4460 // initialized. So we'll instead do the check when we _use_ this sample 4461 // later. 4462 if (_eden_chunk_index == 0 || 4463 (pointer_delta(_eden_chunk_array[_eden_chunk_index], 4464 _eden_chunk_array[_eden_chunk_index-1]) 4465 >= CMSSamplingGrain)) { 4466 _eden_chunk_index++; // commit sample 4467 } 4468 } 4469 } 4470 if ((_collectorState == AbortablePreclean) && !_abort_preclean) { 4471 size_t used = get_eden_used(); 4472 size_t capacity = get_eden_capacity(); 4473 assert(used <= capacity, "Unexpected state of Eden"); 4474 if (used > (capacity/100 * CMSScheduleRemarkEdenPenetration)) { 4475 _abort_preclean = true; 4476 } 4477 } 4478 } 4479 4480 4481 size_t CMSCollector::preclean_work(bool clean_refs, bool clean_survivor) { 4482 assert(_collectorState == Precleaning || 4483 _collectorState == AbortablePreclean, "incorrect state"); 4484 ResourceMark rm; 4485 HandleMark hm; 4486 4487 // Precleaning is currently not MT but the reference processor 4488 // may be set for MT. Disable it temporarily here. 4489 ReferenceProcessor* rp = ref_processor(); 4490 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(rp, false); 4491 4492 // Do one pass of scrubbing the discovered reference lists 4493 // to remove any reference objects with strongly-reachable 4494 // referents. 4495 if (clean_refs) { 4496 CMSPrecleanRefsYieldClosure yield_cl(this); 4497 assert(rp->span().equals(_span), "Spans should be equal"); 4498 CMSKeepAliveClosure keep_alive(this, _span, &_markBitMap, 4499 &_markStack, true /* preclean */); 4500 CMSDrainMarkingStackClosure complete_trace(this, 4501 _span, &_markBitMap, &_markStack, 4502 &keep_alive, true /* preclean */); 4503 4504 // We don't want this step to interfere with a young 4505 // collection because we don't want to take CPU 4506 // or memory bandwidth away from the young GC threads 4507 // (which may be as many as there are CPUs). 4508 // Note that we don't need to protect ourselves from 4509 // interference with mutators because they can't 4510 // manipulate the discovered reference lists nor affect 4511 // the computed reachability of the referents, the 4512 // only properties manipulated by the precleaning 4513 // of these reference lists. 4514 stopTimer(); 4515 CMSTokenSyncWithLocks x(true /* is cms thread */, 4516 bitMapLock()); 4517 startTimer(); 4518 sample_eden(); 4519 4520 // The following will yield to allow foreground 4521 // collection to proceed promptly. XXX YSR: 4522 // The code in this method may need further 4523 // tweaking for better performance and some restructuring 4524 // for cleaner interfaces. 4525 rp->preclean_discovered_references( 4526 rp->is_alive_non_header(), &keep_alive, &complete_trace, &yield_cl); 4527 } 4528 4529 if (clean_survivor) { // preclean the active survivor space(s) 4530 assert(_young_gen->kind() == Generation::DefNew || 4531 _young_gen->kind() == Generation::ParNew || 4532 _young_gen->kind() == Generation::ASParNew, 4533 "incorrect type for cast"); 4534 DefNewGeneration* dng = (DefNewGeneration*)_young_gen; 4535 PushAndMarkClosure pam_cl(this, _span, ref_processor(), 4536 &_markBitMap, &_modUnionTable, 4537 &_markStack, true /* precleaning phase */); 4538 stopTimer(); 4539 CMSTokenSyncWithLocks ts(true /* is cms thread */, 4540 bitMapLock()); 4541 startTimer(); 4542 unsigned int before_count = 4543 GenCollectedHeap::heap()->total_collections(); 4544 SurvivorSpacePrecleanClosure 4545 sss_cl(this, _span, &_markBitMap, &_markStack, 4546 &pam_cl, before_count, CMSYield); 4547 dng->from()->object_iterate_careful(&sss_cl); 4548 dng->to()->object_iterate_careful(&sss_cl); 4549 } 4550 MarkRefsIntoAndScanClosure 4551 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable, 4552 &_markStack, this, CMSYield, 4553 true /* precleaning phase */); 4554 // CAUTION: The following closure has persistent state that may need to 4555 // be reset upon a decrease in the sequence of addresses it 4556 // processes. 4557 ScanMarkedObjectsAgainCarefullyClosure 4558 smoac_cl(this, _span, 4559 &_markBitMap, &_markStack, &mrias_cl, CMSYield); 4560 4561 // Preclean dirty cards in ModUnionTable and CardTable using 4562 // appropriate convergence criterion; 4563 // repeat CMSPrecleanIter times unless we find that 4564 // we are losing. 4565 assert(CMSPrecleanIter < 10, "CMSPrecleanIter is too large"); 4566 assert(CMSPrecleanNumerator < CMSPrecleanDenominator, 4567 "Bad convergence multiplier"); 4568 assert(CMSPrecleanThreshold >= 100, 4569 "Unreasonably low CMSPrecleanThreshold"); 4570 4571 size_t numIter, cumNumCards, lastNumCards, curNumCards; 4572 for (numIter = 0, cumNumCards = lastNumCards = curNumCards = 0; 4573 numIter < CMSPrecleanIter; 4574 numIter++, lastNumCards = curNumCards, cumNumCards += curNumCards) { 4575 curNumCards = preclean_mod_union_table(_cmsGen, &smoac_cl); 4576 if (Verbose && PrintGCDetails) { 4577 gclog_or_tty->print(" (modUnionTable: %d cards)", curNumCards); 4578 } 4579 // Either there are very few dirty cards, so re-mark 4580 // pause will be small anyway, or our pre-cleaning isn't 4581 // that much faster than the rate at which cards are being 4582 // dirtied, so we might as well stop and re-mark since 4583 // precleaning won't improve our re-mark time by much. 4584 if (curNumCards <= CMSPrecleanThreshold || 4585 (numIter > 0 && 4586 (curNumCards * CMSPrecleanDenominator > 4587 lastNumCards * CMSPrecleanNumerator))) { 4588 numIter++; 4589 cumNumCards += curNumCards; 4590 break; 4591 } 4592 } 4593 4594 preclean_klasses(&mrias_cl, _cmsGen->freelistLock()); 4595 4596 curNumCards = preclean_card_table(_cmsGen, &smoac_cl); 4597 cumNumCards += curNumCards; 4598 if (PrintGCDetails && PrintCMSStatistics != 0) { 4599 gclog_or_tty->print_cr(" (cardTable: %d cards, re-scanned %d cards, %d iterations)", 4600 curNumCards, cumNumCards, numIter); 4601 } 4602 return cumNumCards; // as a measure of useful work done 4603 } 4604 4605 // PRECLEANING NOTES: 4606 // Precleaning involves: 4607 // . reading the bits of the modUnionTable and clearing the set bits. 4608 // . For the cards corresponding to the set bits, we scan the 4609 // objects on those cards. This means we need the free_list_lock 4610 // so that we can safely iterate over the CMS space when scanning 4611 // for oops. 4612 // . When we scan the objects, we'll be both reading and setting 4613 // marks in the marking bit map, so we'll need the marking bit map. 4614 // . For protecting _collector_state transitions, we take the CGC_lock. 4615 // Note that any races in the reading of of card table entries by the 4616 // CMS thread on the one hand and the clearing of those entries by the 4617 // VM thread or the setting of those entries by the mutator threads on the 4618 // other are quite benign. However, for efficiency it makes sense to keep 4619 // the VM thread from racing with the CMS thread while the latter is 4620 // dirty card info to the modUnionTable. We therefore also use the 4621 // CGC_lock to protect the reading of the card table and the mod union 4622 // table by the CM thread. 4623 // . We run concurrently with mutator updates, so scanning 4624 // needs to be done carefully -- we should not try to scan 4625 // potentially uninitialized objects. 4626 // 4627 // Locking strategy: While holding the CGC_lock, we scan over and 4628 // reset a maximal dirty range of the mod union / card tables, then lock 4629 // the free_list_lock and bitmap lock to do a full marking, then 4630 // release these locks; and repeat the cycle. This allows for a 4631 // certain amount of fairness in the sharing of these locks between 4632 // the CMS collector on the one hand, and the VM thread and the 4633 // mutators on the other. 4634 4635 // NOTE: preclean_mod_union_table() and preclean_card_table() 4636 // further below are largely identical; if you need to modify 4637 // one of these methods, please check the other method too. 4638 4639 size_t CMSCollector::preclean_mod_union_table( 4640 ConcurrentMarkSweepGeneration* gen, 4641 ScanMarkedObjectsAgainCarefullyClosure* cl) { 4642 verify_work_stacks_empty(); 4643 verify_overflow_empty(); 4644 4645 // strategy: starting with the first card, accumulate contiguous 4646 // ranges of dirty cards; clear these cards, then scan the region 4647 // covered by these cards. 4648 4649 // Since all of the MUT is committed ahead, we can just use 4650 // that, in case the generations expand while we are precleaning. 4651 // It might also be fine to just use the committed part of the 4652 // generation, but we might potentially miss cards when the 4653 // generation is rapidly expanding while we are in the midst 4654 // of precleaning. 4655 HeapWord* startAddr = gen->reserved().start(); 4656 HeapWord* endAddr = gen->reserved().end(); 4657 4658 cl->setFreelistLock(gen->freelistLock()); // needed for yielding 4659 4660 size_t numDirtyCards, cumNumDirtyCards; 4661 HeapWord *nextAddr, *lastAddr; 4662 for (cumNumDirtyCards = numDirtyCards = 0, 4663 nextAddr = lastAddr = startAddr; 4664 nextAddr < endAddr; 4665 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 4666 4667 ResourceMark rm; 4668 HandleMark hm; 4669 4670 MemRegion dirtyRegion; 4671 { 4672 stopTimer(); 4673 // Potential yield point 4674 CMSTokenSync ts(true); 4675 startTimer(); 4676 sample_eden(); 4677 // Get dirty region starting at nextOffset (inclusive), 4678 // simultaneously clearing it. 4679 dirtyRegion = 4680 _modUnionTable.getAndClearMarkedRegion(nextAddr, endAddr); 4681 assert(dirtyRegion.start() >= nextAddr, 4682 "returned region inconsistent?"); 4683 } 4684 // Remember where the next search should begin. 4685 // The returned region (if non-empty) is a right open interval, 4686 // so lastOffset is obtained from the right end of that 4687 // interval. 4688 lastAddr = dirtyRegion.end(); 4689 // Should do something more transparent and less hacky XXX 4690 numDirtyCards = 4691 _modUnionTable.heapWordDiffToOffsetDiff(dirtyRegion.word_size()); 4692 4693 // We'll scan the cards in the dirty region (with periodic 4694 // yields for foreground GC as needed). 4695 if (!dirtyRegion.is_empty()) { 4696 assert(numDirtyCards > 0, "consistency check"); 4697 HeapWord* stop_point = NULL; 4698 stopTimer(); 4699 // Potential yield point 4700 CMSTokenSyncWithLocks ts(true, gen->freelistLock(), 4701 bitMapLock()); 4702 startTimer(); 4703 { 4704 verify_work_stacks_empty(); 4705 verify_overflow_empty(); 4706 sample_eden(); 4707 stop_point = 4708 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 4709 } 4710 if (stop_point != NULL) { 4711 // The careful iteration stopped early either because it found an 4712 // uninitialized object, or because we were in the midst of an 4713 // "abortable preclean", which should now be aborted. Redirty 4714 // the bits corresponding to the partially-scanned or unscanned 4715 // cards. We'll either restart at the next block boundary or 4716 // abort the preclean. 4717 assert((_collectorState == AbortablePreclean && should_abort_preclean()), 4718 "Should only be AbortablePreclean."); 4719 _modUnionTable.mark_range(MemRegion(stop_point, dirtyRegion.end())); 4720 if (should_abort_preclean()) { 4721 break; // out of preclean loop 4722 } else { 4723 // Compute the next address at which preclean should pick up; 4724 // might need bitMapLock in order to read P-bits. 4725 lastAddr = next_card_start_after_block(stop_point); 4726 } 4727 } 4728 } else { 4729 assert(lastAddr == endAddr, "consistency check"); 4730 assert(numDirtyCards == 0, "consistency check"); 4731 break; 4732 } 4733 } 4734 verify_work_stacks_empty(); 4735 verify_overflow_empty(); 4736 return cumNumDirtyCards; 4737 } 4738 4739 // NOTE: preclean_mod_union_table() above and preclean_card_table() 4740 // below are largely identical; if you need to modify 4741 // one of these methods, please check the other method too. 4742 4743 size_t CMSCollector::preclean_card_table(ConcurrentMarkSweepGeneration* gen, 4744 ScanMarkedObjectsAgainCarefullyClosure* cl) { 4745 // strategy: it's similar to precleamModUnionTable above, in that 4746 // we accumulate contiguous ranges of dirty cards, mark these cards 4747 // precleaned, then scan the region covered by these cards. 4748 HeapWord* endAddr = (HeapWord*)(gen->_virtual_space.high()); 4749 HeapWord* startAddr = (HeapWord*)(gen->_virtual_space.low()); 4750 4751 cl->setFreelistLock(gen->freelistLock()); // needed for yielding 4752 4753 size_t numDirtyCards, cumNumDirtyCards; 4754 HeapWord *lastAddr, *nextAddr; 4755 4756 for (cumNumDirtyCards = numDirtyCards = 0, 4757 nextAddr = lastAddr = startAddr; 4758 nextAddr < endAddr; 4759 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 4760 4761 ResourceMark rm; 4762 HandleMark hm; 4763 4764 MemRegion dirtyRegion; 4765 { 4766 // See comments in "Precleaning notes" above on why we 4767 // do this locking. XXX Could the locking overheads be 4768 // too high when dirty cards are sparse? [I don't think so.] 4769 stopTimer(); 4770 CMSTokenSync x(true); // is cms thread 4771 startTimer(); 4772 sample_eden(); 4773 // Get and clear dirty region from card table 4774 dirtyRegion = _ct->ct_bs()->dirty_card_range_after_reset( 4775 MemRegion(nextAddr, endAddr), 4776 true, 4777 CardTableModRefBS::precleaned_card_val()); 4778 4779 assert(dirtyRegion.start() >= nextAddr, 4780 "returned region inconsistent?"); 4781 } 4782 lastAddr = dirtyRegion.end(); 4783 numDirtyCards = 4784 dirtyRegion.word_size()/CardTableModRefBS::card_size_in_words; 4785 4786 if (!dirtyRegion.is_empty()) { 4787 stopTimer(); 4788 CMSTokenSyncWithLocks ts(true, gen->freelistLock(), bitMapLock()); 4789 startTimer(); 4790 sample_eden(); 4791 verify_work_stacks_empty(); 4792 verify_overflow_empty(); 4793 HeapWord* stop_point = 4794 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 4795 if (stop_point != NULL) { 4796 assert((_collectorState == AbortablePreclean && should_abort_preclean()), 4797 "Should only be AbortablePreclean."); 4798 _ct->ct_bs()->invalidate(MemRegion(stop_point, dirtyRegion.end())); 4799 if (should_abort_preclean()) { 4800 break; // out of preclean loop 4801 } else { 4802 // Compute the next address at which preclean should pick up. 4803 lastAddr = next_card_start_after_block(stop_point); 4804 } 4805 } 4806 } else { 4807 break; 4808 } 4809 } 4810 verify_work_stacks_empty(); 4811 verify_overflow_empty(); 4812 return cumNumDirtyCards; 4813 } 4814 4815 class PrecleanKlassClosure : public KlassClosure { 4816 CMKlassClosure _cm_klass_closure; 4817 public: 4818 PrecleanKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {} 4819 void do_klass(Klass* k) { 4820 if (k->has_accumulated_modified_oops()) { 4821 k->clear_accumulated_modified_oops(); 4822 4823 _cm_klass_closure.do_klass(k); 4824 } 4825 } 4826 }; 4827 4828 // The freelist lock is needed to prevent asserts, is it really needed? 4829 void CMSCollector::preclean_klasses(MarkRefsIntoAndScanClosure* cl, Mutex* freelistLock) { 4830 4831 cl->set_freelistLock(freelistLock); 4832 4833 CMSTokenSyncWithLocks ts(true, freelistLock, bitMapLock()); 4834 4835 // SSS: Add equivalent to ScanMarkedObjectsAgainCarefullyClosure::do_yield_check and should_abort_preclean? 4836 // SSS: We should probably check if precleaning should be aborted, at suitable intervals? 4837 PrecleanKlassClosure preclean_klass_closure(cl); 4838 ClassLoaderDataGraph::classes_do(&preclean_klass_closure); 4839 4840 verify_work_stacks_empty(); 4841 verify_overflow_empty(); 4842 } 4843 4844 void CMSCollector::checkpointRootsFinal(bool asynch, 4845 bool clear_all_soft_refs, bool init_mark_was_synchronous) { 4846 assert(_collectorState == FinalMarking, "incorrect state transition?"); 4847 check_correct_thread_executing(); 4848 // world is stopped at this checkpoint 4849 assert(SafepointSynchronize::is_at_safepoint(), 4850 "world should be stopped"); 4851 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 4852 4853 verify_work_stacks_empty(); 4854 verify_overflow_empty(); 4855 4856 SpecializationStats::clear(); 4857 if (PrintGCDetails) { 4858 gclog_or_tty->print("[YG occupancy: "SIZE_FORMAT" K ("SIZE_FORMAT" K)]", 4859 _young_gen->used() / K, 4860 _young_gen->capacity() / K); 4861 } 4862 if (asynch) { 4863 if (CMSScavengeBeforeRemark) { 4864 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4865 // Temporarily set flag to false, GCH->do_collection will 4866 // expect it to be false and set to true 4867 FlagSetting fl(gch->_is_gc_active, false); 4868 NOT_PRODUCT(TraceTime t("Scavenge-Before-Remark", 4869 PrintGCDetails && Verbose, true, gclog_or_tty);) 4870 int level = _cmsGen->level() - 1; 4871 if (level >= 0) { 4872 gch->do_collection(true, // full (i.e. force, see below) 4873 false, // !clear_all_soft_refs 4874 0, // size 4875 false, // is_tlab 4876 level // max_level 4877 ); 4878 } 4879 } 4880 FreelistLocker x(this); 4881 MutexLockerEx y(bitMapLock(), 4882 Mutex::_no_safepoint_check_flag); 4883 assert(!init_mark_was_synchronous, "but that's impossible!"); 4884 checkpointRootsFinalWork(asynch, clear_all_soft_refs, false); 4885 } else { 4886 // already have all the locks 4887 checkpointRootsFinalWork(asynch, clear_all_soft_refs, 4888 init_mark_was_synchronous); 4889 } 4890 verify_work_stacks_empty(); 4891 verify_overflow_empty(); 4892 SpecializationStats::print(); 4893 } 4894 4895 void CMSCollector::checkpointRootsFinalWork(bool asynch, 4896 bool clear_all_soft_refs, bool init_mark_was_synchronous) { 4897 4898 NOT_PRODUCT(TraceTime tr("checkpointRootsFinalWork", PrintGCDetails, false, gclog_or_tty);) 4899 4900 assert(haveFreelistLocks(), "must have free list locks"); 4901 assert_lock_strong(bitMapLock()); 4902 4903 if (UseAdaptiveSizePolicy) { 4904 size_policy()->checkpoint_roots_final_begin(); 4905 } 4906 4907 ResourceMark rm; 4908 HandleMark hm; 4909 4910 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4911 4912 if (should_unload_classes()) { 4913 CodeCache::gc_prologue(); 4914 } 4915 assert(haveFreelistLocks(), "must have free list locks"); 4916 assert_lock_strong(bitMapLock()); 4917 4918 if (!init_mark_was_synchronous) { 4919 // We might assume that we need not fill TLAB's when 4920 // CMSScavengeBeforeRemark is set, because we may have just done 4921 // a scavenge which would have filled all TLAB's -- and besides 4922 // Eden would be empty. This however may not always be the case -- 4923 // for instance although we asked for a scavenge, it may not have 4924 // happened because of a JNI critical section. We probably need 4925 // a policy for deciding whether we can in that case wait until 4926 // the critical section releases and then do the remark following 4927 // the scavenge, and skip it here. In the absence of that policy, 4928 // or of an indication of whether the scavenge did indeed occur, 4929 // we cannot rely on TLAB's having been filled and must do 4930 // so here just in case a scavenge did not happen. 4931 gch->ensure_parsability(false); // fill TLAB's, but no need to retire them 4932 // Update the saved marks which may affect the root scans. 4933 gch->save_marks(); 4934 4935 { 4936 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;) 4937 4938 // Note on the role of the mod union table: 4939 // Since the marker in "markFromRoots" marks concurrently with 4940 // mutators, it is possible for some reachable objects not to have been 4941 // scanned. For instance, an only reference to an object A was 4942 // placed in object B after the marker scanned B. Unless B is rescanned, 4943 // A would be collected. Such updates to references in marked objects 4944 // are detected via the mod union table which is the set of all cards 4945 // dirtied since the first checkpoint in this GC cycle and prior to 4946 // the most recent young generation GC, minus those cleaned up by the 4947 // concurrent precleaning. 4948 if (CMSParallelRemarkEnabled && CollectedHeap::use_parallel_gc_threads()) { 4949 TraceTime t("Rescan (parallel) ", PrintGCDetails, false, gclog_or_tty); 4950 do_remark_parallel(); 4951 } else { 4952 TraceTime t("Rescan (non-parallel) ", PrintGCDetails, false, 4953 gclog_or_tty); 4954 do_remark_non_parallel(); 4955 } 4956 } 4957 } else { 4958 assert(!asynch, "Can't have init_mark_was_synchronous in asynch mode"); 4959 // The initial mark was stop-world, so there's no rescanning to 4960 // do; go straight on to the next step below. 4961 } 4962 verify_work_stacks_empty(); 4963 verify_overflow_empty(); 4964 4965 { 4966 NOT_PRODUCT(TraceTime ts("refProcessingWork", PrintGCDetails, false, gclog_or_tty);) 4967 refProcessingWork(asynch, clear_all_soft_refs); 4968 } 4969 verify_work_stacks_empty(); 4970 verify_overflow_empty(); 4971 4972 if (should_unload_classes()) { 4973 CodeCache::gc_epilogue(); 4974 } 4975 JvmtiExport::gc_epilogue(); 4976 4977 // If we encountered any (marking stack / work queue) overflow 4978 // events during the current CMS cycle, take appropriate 4979 // remedial measures, where possible, so as to try and avoid 4980 // recurrence of that condition. 4981 assert(_markStack.isEmpty(), "No grey objects"); 4982 size_t ser_ovflw = _ser_pmc_remark_ovflw + _ser_pmc_preclean_ovflw + 4983 _ser_kac_ovflw + _ser_kac_preclean_ovflw; 4984 if (ser_ovflw > 0) { 4985 if (PrintCMSStatistics != 0) { 4986 gclog_or_tty->print_cr("Marking stack overflow (benign) " 4987 "(pmc_pc="SIZE_FORMAT", pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT 4988 ", kac_preclean="SIZE_FORMAT")", 4989 _ser_pmc_preclean_ovflw, _ser_pmc_remark_ovflw, 4990 _ser_kac_ovflw, _ser_kac_preclean_ovflw); 4991 } 4992 _markStack.expand(); 4993 _ser_pmc_remark_ovflw = 0; 4994 _ser_pmc_preclean_ovflw = 0; 4995 _ser_kac_preclean_ovflw = 0; 4996 _ser_kac_ovflw = 0; 4997 } 4998 if (_par_pmc_remark_ovflw > 0 || _par_kac_ovflw > 0) { 4999 if (PrintCMSStatistics != 0) { 5000 gclog_or_tty->print_cr("Work queue overflow (benign) " 5001 "(pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT")", 5002 _par_pmc_remark_ovflw, _par_kac_ovflw); 5003 } 5004 _par_pmc_remark_ovflw = 0; 5005 _par_kac_ovflw = 0; 5006 } 5007 if (PrintCMSStatistics != 0) { 5008 if (_markStack._hit_limit > 0) { 5009 gclog_or_tty->print_cr(" (benign) Hit max stack size limit ("SIZE_FORMAT")", 5010 _markStack._hit_limit); 5011 } 5012 if (_markStack._failed_double > 0) { 5013 gclog_or_tty->print_cr(" (benign) Failed stack doubling ("SIZE_FORMAT")," 5014 " current capacity "SIZE_FORMAT, 5015 _markStack._failed_double, 5016 _markStack.capacity()); 5017 } 5018 } 5019 _markStack._hit_limit = 0; 5020 _markStack._failed_double = 0; 5021 5022 if ((VerifyAfterGC || VerifyDuringGC) && 5023 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 5024 verify_after_remark(); 5025 } 5026 5027 // Change under the freelistLocks. 5028 _collectorState = Sweeping; 5029 // Call isAllClear() under bitMapLock 5030 assert(_modUnionTable.isAllClear(), 5031 "Should be clear by end of the final marking"); 5032 assert(_ct->klass_rem_set()->mod_union_is_clear(), 5033 "Should be clear by end of the final marking"); 5034 if (UseAdaptiveSizePolicy) { 5035 size_policy()->checkpoint_roots_final_end(gch->gc_cause()); 5036 } 5037 } 5038 5039 // Parallel remark task 5040 class CMSParRemarkTask: public AbstractGangTask { 5041 CMSCollector* _collector; 5042 int _n_workers; 5043 CompactibleFreeListSpace* _cms_space; 5044 5045 // The per-thread work queues, available here for stealing. 5046 OopTaskQueueSet* _task_queues; 5047 ParallelTaskTerminator _term; 5048 5049 public: 5050 // A value of 0 passed to n_workers will cause the number of 5051 // workers to be taken from the active workers in the work gang. 5052 CMSParRemarkTask(CMSCollector* collector, 5053 CompactibleFreeListSpace* cms_space, 5054 int n_workers, FlexibleWorkGang* workers, 5055 OopTaskQueueSet* task_queues): 5056 AbstractGangTask("Rescan roots and grey objects in parallel"), 5057 _collector(collector), 5058 _cms_space(cms_space), 5059 _n_workers(n_workers), 5060 _task_queues(task_queues), 5061 _term(n_workers, task_queues) { } 5062 5063 OopTaskQueueSet* task_queues() { return _task_queues; } 5064 5065 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 5066 5067 ParallelTaskTerminator* terminator() { return &_term; } 5068 int n_workers() { return _n_workers; } 5069 5070 void work(uint worker_id); 5071 5072 private: 5073 // Work method in support of parallel rescan ... of young gen spaces 5074 void do_young_space_rescan(int i, Par_MarkRefsIntoAndScanClosure* cl, 5075 ContiguousSpace* space, 5076 HeapWord** chunk_array, size_t chunk_top); 5077 5078 // ... of dirty cards in old space 5079 void do_dirty_card_rescan_tasks(CompactibleFreeListSpace* sp, int i, 5080 Par_MarkRefsIntoAndScanClosure* cl); 5081 5082 // ... work stealing for the above 5083 void do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, int* seed); 5084 }; 5085 5086 class RemarkKlassClosure : public KlassClosure { 5087 CMKlassClosure _cm_klass_closure; 5088 public: 5089 RemarkKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {} 5090 void do_klass(Klass* k) { 5091 // Check if we have modified any oops in the Klass during the concurrent marking. 5092 if (k->has_accumulated_modified_oops()) { 5093 k->clear_accumulated_modified_oops(); 5094 5095 // We could have transfered the current modified marks to the accumulated marks, 5096 // like we do with the Card Table to Mod Union Table. But it's not really necessary. 5097 } else if (k->has_modified_oops()) { 5098 // Don't clear anything, this info is needed by the next young collection. 5099 } else { 5100 // No modified oops in the Klass. 5101 return; 5102 } 5103 5104 // The klass has modified fields, need to scan the klass. 5105 _cm_klass_closure.do_klass(k); 5106 } 5107 }; 5108 5109 // work_queue(i) is passed to the closure 5110 // Par_MarkRefsIntoAndScanClosure. The "i" parameter 5111 // also is passed to do_dirty_card_rescan_tasks() and to 5112 // do_work_steal() to select the i-th task_queue. 5113 5114 void CMSParRemarkTask::work(uint worker_id) { 5115 elapsedTimer _timer; 5116 ResourceMark rm; 5117 HandleMark hm; 5118 5119 // ---------- rescan from roots -------------- 5120 _timer.start(); 5121 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5122 Par_MarkRefsIntoAndScanClosure par_mrias_cl(_collector, 5123 _collector->_span, _collector->ref_processor(), 5124 &(_collector->_markBitMap), 5125 work_queue(worker_id)); 5126 5127 // Rescan young gen roots first since these are likely 5128 // coarsely partitioned and may, on that account, constitute 5129 // the critical path; thus, it's best to start off that 5130 // work first. 5131 // ---------- young gen roots -------------- 5132 { 5133 DefNewGeneration* dng = _collector->_young_gen->as_DefNewGeneration(); 5134 EdenSpace* eden_space = dng->eden(); 5135 ContiguousSpace* from_space = dng->from(); 5136 ContiguousSpace* to_space = dng->to(); 5137 5138 HeapWord** eca = _collector->_eden_chunk_array; 5139 size_t ect = _collector->_eden_chunk_index; 5140 HeapWord** sca = _collector->_survivor_chunk_array; 5141 size_t sct = _collector->_survivor_chunk_index; 5142 5143 assert(ect <= _collector->_eden_chunk_capacity, "out of bounds"); 5144 assert(sct <= _collector->_survivor_chunk_capacity, "out of bounds"); 5145 5146 do_young_space_rescan(worker_id, &par_mrias_cl, to_space, NULL, 0); 5147 do_young_space_rescan(worker_id, &par_mrias_cl, from_space, sca, sct); 5148 do_young_space_rescan(worker_id, &par_mrias_cl, eden_space, eca, ect); 5149 5150 _timer.stop(); 5151 if (PrintCMSStatistics != 0) { 5152 gclog_or_tty->print_cr( 5153 "Finished young gen rescan work in %dth thread: %3.3f sec", 5154 worker_id, _timer.seconds()); 5155 } 5156 } 5157 5158 // ---------- remaining roots -------------- 5159 _timer.reset(); 5160 _timer.start(); 5161 gch->gen_process_strong_roots(_collector->_cmsGen->level(), 5162 false, // yg was scanned above 5163 false, // this is parallel code 5164 false, // not scavenging 5165 SharedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()), 5166 &par_mrias_cl, 5167 true, // walk all of code cache if (so & SO_CodeCache) 5168 NULL, 5169 NULL); // The dirty klasses will be handled below 5170 assert(_collector->should_unload_classes() 5171 || (_collector->CMSCollector::roots_scanning_options() & SharedHeap::SO_CodeCache), 5172 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 5173 _timer.stop(); 5174 if (PrintCMSStatistics != 0) { 5175 gclog_or_tty->print_cr( 5176 "Finished remaining root rescan work in %dth thread: %3.3f sec", 5177 worker_id, _timer.seconds()); 5178 } 5179 5180 // ---------- unhandled CLD scanning ---------- 5181 if (worker_id == 0) { // Single threaded at the moment. 5182 _timer.reset(); 5183 _timer.start(); 5184 5185 // Scan all new class loader data objects and new dependencies that were 5186 // introduced during concurrent marking. 5187 ResourceMark rm; 5188 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds(); 5189 for (int i = 0; i < array->length(); i++) { 5190 par_mrias_cl.do_class_loader_data(array->at(i)); 5191 } 5192 5193 // We don't need to keep track of new CLDs anymore. 5194 ClassLoaderDataGraph::remember_new_clds(false); 5195 5196 _timer.stop(); 5197 if (PrintCMSStatistics != 0) { 5198 gclog_or_tty->print_cr( 5199 "Finished unhandled CLD scanning work in %dth thread: %3.3f sec", 5200 worker_id, _timer.seconds()); 5201 } 5202 } 5203 5204 // ---------- dirty klass scanning ---------- 5205 if (worker_id == 0) { // Single threaded at the moment. 5206 _timer.reset(); 5207 _timer.start(); 5208 5209 // Scan all classes that was dirtied during the concurrent marking phase. 5210 RemarkKlassClosure remark_klass_closure(&par_mrias_cl); 5211 ClassLoaderDataGraph::classes_do(&remark_klass_closure); 5212 5213 _timer.stop(); 5214 if (PrintCMSStatistics != 0) { 5215 gclog_or_tty->print_cr( 5216 "Finished dirty klass scanning work in %dth thread: %3.3f sec", 5217 worker_id, _timer.seconds()); 5218 } 5219 } 5220 5221 // We might have added oops to ClassLoaderData::_handles during the 5222 // concurrent marking phase. These oops point to newly allocated objects 5223 // that are guaranteed to be kept alive. Either by the direct allocation 5224 // code, or when the young collector processes the strong roots. Hence, 5225 // we don't have to revisit the _handles block during the remark phase. 5226 5227 // ---------- rescan dirty cards ------------ 5228 _timer.reset(); 5229 _timer.start(); 5230 5231 // Do the rescan tasks for each of the two spaces 5232 // (cms_space) in turn. 5233 // "worker_id" is passed to select the task_queue for "worker_id" 5234 do_dirty_card_rescan_tasks(_cms_space, worker_id, &par_mrias_cl); 5235 _timer.stop(); 5236 if (PrintCMSStatistics != 0) { 5237 gclog_or_tty->print_cr( 5238 "Finished dirty card rescan work in %dth thread: %3.3f sec", 5239 worker_id, _timer.seconds()); 5240 } 5241 5242 // ---------- steal work from other threads ... 5243 // ---------- ... and drain overflow list. 5244 _timer.reset(); 5245 _timer.start(); 5246 do_work_steal(worker_id, &par_mrias_cl, _collector->hash_seed(worker_id)); 5247 _timer.stop(); 5248 if (PrintCMSStatistics != 0) { 5249 gclog_or_tty->print_cr( 5250 "Finished work stealing in %dth thread: %3.3f sec", 5251 worker_id, _timer.seconds()); 5252 } 5253 } 5254 5255 // Note that parameter "i" is not used. 5256 void 5257 CMSParRemarkTask::do_young_space_rescan(int i, 5258 Par_MarkRefsIntoAndScanClosure* cl, ContiguousSpace* space, 5259 HeapWord** chunk_array, size_t chunk_top) { 5260 // Until all tasks completed: 5261 // . claim an unclaimed task 5262 // . compute region boundaries corresponding to task claimed 5263 // using chunk_array 5264 // . par_oop_iterate(cl) over that region 5265 5266 ResourceMark rm; 5267 HandleMark hm; 5268 5269 SequentialSubTasksDone* pst = space->par_seq_tasks(); 5270 assert(pst->valid(), "Uninitialized use?"); 5271 5272 uint nth_task = 0; 5273 uint n_tasks = pst->n_tasks(); 5274 5275 HeapWord *start, *end; 5276 while (!pst->is_task_claimed(/* reference */ nth_task)) { 5277 // We claimed task # nth_task; compute its boundaries. 5278 if (chunk_top == 0) { // no samples were taken 5279 assert(nth_task == 0 && n_tasks == 1, "Can have only 1 EdenSpace task"); 5280 start = space->bottom(); 5281 end = space->top(); 5282 } else if (nth_task == 0) { 5283 start = space->bottom(); 5284 end = chunk_array[nth_task]; 5285 } else if (nth_task < (uint)chunk_top) { 5286 assert(nth_task >= 1, "Control point invariant"); 5287 start = chunk_array[nth_task - 1]; 5288 end = chunk_array[nth_task]; 5289 } else { 5290 assert(nth_task == (uint)chunk_top, "Control point invariant"); 5291 start = chunk_array[chunk_top - 1]; 5292 end = space->top(); 5293 } 5294 MemRegion mr(start, end); 5295 // Verify that mr is in space 5296 assert(mr.is_empty() || space->used_region().contains(mr), 5297 "Should be in space"); 5298 // Verify that "start" is an object boundary 5299 assert(mr.is_empty() || oop(mr.start())->is_oop(), 5300 "Should be an oop"); 5301 space->par_oop_iterate(mr, cl); 5302 } 5303 pst->all_tasks_completed(); 5304 } 5305 5306 void 5307 CMSParRemarkTask::do_dirty_card_rescan_tasks( 5308 CompactibleFreeListSpace* sp, int i, 5309 Par_MarkRefsIntoAndScanClosure* cl) { 5310 // Until all tasks completed: 5311 // . claim an unclaimed task 5312 // . compute region boundaries corresponding to task claimed 5313 // . transfer dirty bits ct->mut for that region 5314 // . apply rescanclosure to dirty mut bits for that region 5315 5316 ResourceMark rm; 5317 HandleMark hm; 5318 5319 OopTaskQueue* work_q = work_queue(i); 5320 ModUnionClosure modUnionClosure(&(_collector->_modUnionTable)); 5321 // CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! 5322 // CAUTION: This closure has state that persists across calls to 5323 // the work method dirty_range_iterate_clear() in that it has 5324 // imbedded in it a (subtype of) UpwardsObjectClosure. The 5325 // use of that state in the imbedded UpwardsObjectClosure instance 5326 // assumes that the cards are always iterated (even if in parallel 5327 // by several threads) in monotonically increasing order per each 5328 // thread. This is true of the implementation below which picks 5329 // card ranges (chunks) in monotonically increasing order globally 5330 // and, a-fortiori, in monotonically increasing order per thread 5331 // (the latter order being a subsequence of the former). 5332 // If the work code below is ever reorganized into a more chaotic 5333 // work-partitioning form than the current "sequential tasks" 5334 // paradigm, the use of that persistent state will have to be 5335 // revisited and modified appropriately. See also related 5336 // bug 4756801 work on which should examine this code to make 5337 // sure that the changes there do not run counter to the 5338 // assumptions made here and necessary for correctness and 5339 // efficiency. Note also that this code might yield inefficient 5340 // behaviour in the case of very large objects that span one or 5341 // more work chunks. Such objects would potentially be scanned 5342 // several times redundantly. Work on 4756801 should try and 5343 // address that performance anomaly if at all possible. XXX 5344 MemRegion full_span = _collector->_span; 5345 CMSBitMap* bm = &(_collector->_markBitMap); // shared 5346 MarkFromDirtyCardsClosure 5347 greyRescanClosure(_collector, full_span, // entire span of interest 5348 sp, bm, work_q, cl); 5349 5350 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 5351 assert(pst->valid(), "Uninitialized use?"); 5352 uint nth_task = 0; 5353 const int alignment = CardTableModRefBS::card_size * BitsPerWord; 5354 MemRegion span = sp->used_region(); 5355 HeapWord* start_addr = span.start(); 5356 HeapWord* end_addr = (HeapWord*)round_to((intptr_t)span.end(), 5357 alignment); 5358 const size_t chunk_size = sp->rescan_task_size(); // in HeapWord units 5359 assert((HeapWord*)round_to((intptr_t)start_addr, alignment) == 5360 start_addr, "Check alignment"); 5361 assert((size_t)round_to((intptr_t)chunk_size, alignment) == 5362 chunk_size, "Check alignment"); 5363 5364 while (!pst->is_task_claimed(/* reference */ nth_task)) { 5365 // Having claimed the nth_task, compute corresponding mem-region, 5366 // which is a-fortiori aligned correctly (i.e. at a MUT bopundary). 5367 // The alignment restriction ensures that we do not need any 5368 // synchronization with other gang-workers while setting or 5369 // clearing bits in thus chunk of the MUT. 5370 MemRegion this_span = MemRegion(start_addr + nth_task*chunk_size, 5371 start_addr + (nth_task+1)*chunk_size); 5372 // The last chunk's end might be way beyond end of the 5373 // used region. In that case pull back appropriately. 5374 if (this_span.end() > end_addr) { 5375 this_span.set_end(end_addr); 5376 assert(!this_span.is_empty(), "Program logic (calculation of n_tasks)"); 5377 } 5378 // Iterate over the dirty cards covering this chunk, marking them 5379 // precleaned, and setting the corresponding bits in the mod union 5380 // table. Since we have been careful to partition at Card and MUT-word 5381 // boundaries no synchronization is needed between parallel threads. 5382 _collector->_ct->ct_bs()->dirty_card_iterate(this_span, 5383 &modUnionClosure); 5384 5385 // Having transferred these marks into the modUnionTable, 5386 // rescan the marked objects on the dirty cards in the modUnionTable. 5387 // Even if this is at a synchronous collection, the initial marking 5388 // may have been done during an asynchronous collection so there 5389 // may be dirty bits in the mod-union table. 5390 _collector->_modUnionTable.dirty_range_iterate_clear( 5391 this_span, &greyRescanClosure); 5392 _collector->_modUnionTable.verifyNoOneBitsInRange( 5393 this_span.start(), 5394 this_span.end()); 5395 } 5396 pst->all_tasks_completed(); // declare that i am done 5397 } 5398 5399 // . see if we can share work_queues with ParNew? XXX 5400 void 5401 CMSParRemarkTask::do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, 5402 int* seed) { 5403 OopTaskQueue* work_q = work_queue(i); 5404 NOT_PRODUCT(int num_steals = 0;) 5405 oop obj_to_scan; 5406 CMSBitMap* bm = &(_collector->_markBitMap); 5407 5408 while (true) { 5409 // Completely finish any left over work from (an) earlier round(s) 5410 cl->trim_queue(0); 5411 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 5412 (size_t)ParGCDesiredObjsFromOverflowList); 5413 // Now check if there's any work in the overflow list 5414 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads, 5415 // only affects the number of attempts made to get work from the 5416 // overflow list and does not affect the number of workers. Just 5417 // pass ParallelGCThreads so this behavior is unchanged. 5418 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 5419 work_q, 5420 ParallelGCThreads)) { 5421 // found something in global overflow list; 5422 // not yet ready to go stealing work from others. 5423 // We'd like to assert(work_q->size() != 0, ...) 5424 // because we just took work from the overflow list, 5425 // but of course we can't since all of that could have 5426 // been already stolen from us. 5427 // "He giveth and He taketh away." 5428 continue; 5429 } 5430 // Verify that we have no work before we resort to stealing 5431 assert(work_q->size() == 0, "Have work, shouldn't steal"); 5432 // Try to steal from other queues that have work 5433 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 5434 NOT_PRODUCT(num_steals++;) 5435 assert(obj_to_scan->is_oop(), "Oops, not an oop!"); 5436 assert(bm->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 5437 // Do scanning work 5438 obj_to_scan->oop_iterate(cl); 5439 // Loop around, finish this work, and try to steal some more 5440 } else if (terminator()->offer_termination()) { 5441 break; // nirvana from the infinite cycle 5442 } 5443 } 5444 NOT_PRODUCT( 5445 if (PrintCMSStatistics != 0) { 5446 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals); 5447 } 5448 ) 5449 assert(work_q->size() == 0 && _collector->overflow_list_is_empty(), 5450 "Else our work is not yet done"); 5451 } 5452 5453 // Return a thread-local PLAB recording array, as appropriate. 5454 void* CMSCollector::get_data_recorder(int thr_num) { 5455 if (_survivor_plab_array != NULL && 5456 (CMSPLABRecordAlways || 5457 (_collectorState > Marking && _collectorState < FinalMarking))) { 5458 assert(thr_num < (int)ParallelGCThreads, "thr_num is out of bounds"); 5459 ChunkArray* ca = &_survivor_plab_array[thr_num]; 5460 ca->reset(); // clear it so that fresh data is recorded 5461 return (void*) ca; 5462 } else { 5463 return NULL; 5464 } 5465 } 5466 5467 // Reset all the thread-local PLAB recording arrays 5468 void CMSCollector::reset_survivor_plab_arrays() { 5469 for (uint i = 0; i < ParallelGCThreads; i++) { 5470 _survivor_plab_array[i].reset(); 5471 } 5472 } 5473 5474 // Merge the per-thread plab arrays into the global survivor chunk 5475 // array which will provide the partitioning of the survivor space 5476 // for CMS rescan. 5477 void CMSCollector::merge_survivor_plab_arrays(ContiguousSpace* surv, 5478 int no_of_gc_threads) { 5479 assert(_survivor_plab_array != NULL, "Error"); 5480 assert(_survivor_chunk_array != NULL, "Error"); 5481 assert(_collectorState == FinalMarking, "Error"); 5482 for (int j = 0; j < no_of_gc_threads; j++) { 5483 _cursor[j] = 0; 5484 } 5485 HeapWord* top = surv->top(); 5486 size_t i; 5487 for (i = 0; i < _survivor_chunk_capacity; i++) { // all sca entries 5488 HeapWord* min_val = top; // Higher than any PLAB address 5489 uint min_tid = 0; // position of min_val this round 5490 for (int j = 0; j < no_of_gc_threads; j++) { 5491 ChunkArray* cur_sca = &_survivor_plab_array[j]; 5492 if (_cursor[j] == cur_sca->end()) { 5493 continue; 5494 } 5495 assert(_cursor[j] < cur_sca->end(), "ctl pt invariant"); 5496 HeapWord* cur_val = cur_sca->nth(_cursor[j]); 5497 assert(surv->used_region().contains(cur_val), "Out of bounds value"); 5498 if (cur_val < min_val) { 5499 min_tid = j; 5500 min_val = cur_val; 5501 } else { 5502 assert(cur_val < top, "All recorded addresses should be less"); 5503 } 5504 } 5505 // At this point min_val and min_tid are respectively 5506 // the least address in _survivor_plab_array[j]->nth(_cursor[j]) 5507 // and the thread (j) that witnesses that address. 5508 // We record this address in the _survivor_chunk_array[i] 5509 // and increment _cursor[min_tid] prior to the next round i. 5510 if (min_val == top) { 5511 break; 5512 } 5513 _survivor_chunk_array[i] = min_val; 5514 _cursor[min_tid]++; 5515 } 5516 // We are all done; record the size of the _survivor_chunk_array 5517 _survivor_chunk_index = i; // exclusive: [0, i) 5518 if (PrintCMSStatistics > 0) { 5519 gclog_or_tty->print(" (Survivor:" SIZE_FORMAT "chunks) ", i); 5520 } 5521 // Verify that we used up all the recorded entries 5522 #ifdef ASSERT 5523 size_t total = 0; 5524 for (int j = 0; j < no_of_gc_threads; j++) { 5525 assert(_cursor[j] == _survivor_plab_array[j].end(), "Ctl pt invariant"); 5526 total += _cursor[j]; 5527 } 5528 assert(total == _survivor_chunk_index, "Ctl Pt Invariant"); 5529 // Check that the merged array is in sorted order 5530 if (total > 0) { 5531 for (size_t i = 0; i < total - 1; i++) { 5532 if (PrintCMSStatistics > 0) { 5533 gclog_or_tty->print(" (chunk" SIZE_FORMAT ":" INTPTR_FORMAT ") ", 5534 i, _survivor_chunk_array[i]); 5535 } 5536 assert(_survivor_chunk_array[i] < _survivor_chunk_array[i+1], 5537 "Not sorted"); 5538 } 5539 } 5540 #endif // ASSERT 5541 } 5542 5543 // Set up the space's par_seq_tasks structure for work claiming 5544 // for parallel rescan of young gen. 5545 // See ParRescanTask where this is currently used. 5546 void 5547 CMSCollector:: 5548 initialize_sequential_subtasks_for_young_gen_rescan(int n_threads) { 5549 assert(n_threads > 0, "Unexpected n_threads argument"); 5550 DefNewGeneration* dng = (DefNewGeneration*)_young_gen; 5551 5552 // Eden space 5553 { 5554 SequentialSubTasksDone* pst = dng->eden()->par_seq_tasks(); 5555 assert(!pst->valid(), "Clobbering existing data?"); 5556 // Each valid entry in [0, _eden_chunk_index) represents a task. 5557 size_t n_tasks = _eden_chunk_index + 1; 5558 assert(n_tasks == 1 || _eden_chunk_array != NULL, "Error"); 5559 // Sets the condition for completion of the subtask (how many threads 5560 // need to finish in order to be done). 5561 pst->set_n_threads(n_threads); 5562 pst->set_n_tasks((int)n_tasks); 5563 } 5564 5565 // Merge the survivor plab arrays into _survivor_chunk_array 5566 if (_survivor_plab_array != NULL) { 5567 merge_survivor_plab_arrays(dng->from(), n_threads); 5568 } else { 5569 assert(_survivor_chunk_index == 0, "Error"); 5570 } 5571 5572 // To space 5573 { 5574 SequentialSubTasksDone* pst = dng->to()->par_seq_tasks(); 5575 assert(!pst->valid(), "Clobbering existing data?"); 5576 // Sets the condition for completion of the subtask (how many threads 5577 // need to finish in order to be done). 5578 pst->set_n_threads(n_threads); 5579 pst->set_n_tasks(1); 5580 assert(pst->valid(), "Error"); 5581 } 5582 5583 // From space 5584 { 5585 SequentialSubTasksDone* pst = dng->from()->par_seq_tasks(); 5586 assert(!pst->valid(), "Clobbering existing data?"); 5587 size_t n_tasks = _survivor_chunk_index + 1; 5588 assert(n_tasks == 1 || _survivor_chunk_array != NULL, "Error"); 5589 // Sets the condition for completion of the subtask (how many threads 5590 // need to finish in order to be done). 5591 pst->set_n_threads(n_threads); 5592 pst->set_n_tasks((int)n_tasks); 5593 assert(pst->valid(), "Error"); 5594 } 5595 } 5596 5597 // Parallel version of remark 5598 void CMSCollector::do_remark_parallel() { 5599 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5600 FlexibleWorkGang* workers = gch->workers(); 5601 assert(workers != NULL, "Need parallel worker threads."); 5602 // Choose to use the number of GC workers most recently set 5603 // into "active_workers". If active_workers is not set, set it 5604 // to ParallelGCThreads. 5605 int n_workers = workers->active_workers(); 5606 if (n_workers == 0) { 5607 assert(n_workers > 0, "Should have been set during scavenge"); 5608 n_workers = ParallelGCThreads; 5609 workers->set_active_workers(n_workers); 5610 } 5611 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 5612 5613 CMSParRemarkTask tsk(this, 5614 cms_space, 5615 n_workers, workers, task_queues()); 5616 5617 // Set up for parallel process_strong_roots work. 5618 gch->set_par_threads(n_workers); 5619 // We won't be iterating over the cards in the card table updating 5620 // the younger_gen cards, so we shouldn't call the following else 5621 // the verification code as well as subsequent younger_refs_iterate 5622 // code would get confused. XXX 5623 // gch->rem_set()->prepare_for_younger_refs_iterate(true); // parallel 5624 5625 // The young gen rescan work will not be done as part of 5626 // process_strong_roots (which currently doesn't knw how to 5627 // parallelize such a scan), but rather will be broken up into 5628 // a set of parallel tasks (via the sampling that the [abortable] 5629 // preclean phase did of EdenSpace, plus the [two] tasks of 5630 // scanning the [two] survivor spaces. Further fine-grain 5631 // parallelization of the scanning of the survivor spaces 5632 // themselves, and of precleaning of the younger gen itself 5633 // is deferred to the future. 5634 initialize_sequential_subtasks_for_young_gen_rescan(n_workers); 5635 5636 // The dirty card rescan work is broken up into a "sequence" 5637 // of parallel tasks (per constituent space) that are dynamically 5638 // claimed by the parallel threads. 5639 cms_space->initialize_sequential_subtasks_for_rescan(n_workers); 5640 5641 // It turns out that even when we're using 1 thread, doing the work in a 5642 // separate thread causes wide variance in run times. We can't help this 5643 // in the multi-threaded case, but we special-case n=1 here to get 5644 // repeatable measurements of the 1-thread overhead of the parallel code. 5645 if (n_workers > 1) { 5646 // Make refs discovery MT-safe, if it isn't already: it may not 5647 // necessarily be so, since it's possible that we are doing 5648 // ST marking. 5649 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), true); 5650 GenCollectedHeap::StrongRootsScope srs(gch); 5651 workers->run_task(&tsk); 5652 } else { 5653 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false); 5654 GenCollectedHeap::StrongRootsScope srs(gch); 5655 tsk.work(0); 5656 } 5657 5658 gch->set_par_threads(0); // 0 ==> non-parallel. 5659 // restore, single-threaded for now, any preserved marks 5660 // as a result of work_q overflow 5661 restore_preserved_marks_if_any(); 5662 } 5663 5664 // Non-parallel version of remark 5665 void CMSCollector::do_remark_non_parallel() { 5666 ResourceMark rm; 5667 HandleMark hm; 5668 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5669 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false); 5670 5671 MarkRefsIntoAndScanClosure 5672 mrias_cl(_span, ref_processor(), &_markBitMap, NULL /* not precleaning */, 5673 &_markStack, this, 5674 false /* should_yield */, false /* not precleaning */); 5675 MarkFromDirtyCardsClosure 5676 markFromDirtyCardsClosure(this, _span, 5677 NULL, // space is set further below 5678 &_markBitMap, &_markStack, &mrias_cl); 5679 { 5680 TraceTime t("grey object rescan", PrintGCDetails, false, gclog_or_tty); 5681 // Iterate over the dirty cards, setting the corresponding bits in the 5682 // mod union table. 5683 { 5684 ModUnionClosure modUnionClosure(&_modUnionTable); 5685 _ct->ct_bs()->dirty_card_iterate( 5686 _cmsGen->used_region(), 5687 &modUnionClosure); 5688 } 5689 // Having transferred these marks into the modUnionTable, we just need 5690 // to rescan the marked objects on the dirty cards in the modUnionTable. 5691 // The initial marking may have been done during an asynchronous 5692 // collection so there may be dirty bits in the mod-union table. 5693 const int alignment = 5694 CardTableModRefBS::card_size * BitsPerWord; 5695 { 5696 // ... First handle dirty cards in CMS gen 5697 markFromDirtyCardsClosure.set_space(_cmsGen->cmsSpace()); 5698 MemRegion ur = _cmsGen->used_region(); 5699 HeapWord* lb = ur.start(); 5700 HeapWord* ub = (HeapWord*)round_to((intptr_t)ur.end(), alignment); 5701 MemRegion cms_span(lb, ub); 5702 _modUnionTable.dirty_range_iterate_clear(cms_span, 5703 &markFromDirtyCardsClosure); 5704 verify_work_stacks_empty(); 5705 if (PrintCMSStatistics != 0) { 5706 gclog_or_tty->print(" (re-scanned "SIZE_FORMAT" dirty cards in cms gen) ", 5707 markFromDirtyCardsClosure.num_dirty_cards()); 5708 } 5709 } 5710 } 5711 if (VerifyDuringGC && 5712 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 5713 HandleMark hm; // Discard invalid handles created during verification 5714 Universe::verify(); 5715 } 5716 { 5717 TraceTime t("root rescan", PrintGCDetails, false, gclog_or_tty); 5718 5719 verify_work_stacks_empty(); 5720 5721 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 5722 GenCollectedHeap::StrongRootsScope srs(gch); 5723 gch->gen_process_strong_roots(_cmsGen->level(), 5724 true, // younger gens as roots 5725 false, // use the local StrongRootsScope 5726 false, // not scavenging 5727 SharedHeap::ScanningOption(roots_scanning_options()), 5728 &mrias_cl, 5729 true, // walk code active on stacks 5730 NULL, 5731 NULL); // The dirty klasses will be handled below 5732 5733 assert(should_unload_classes() 5734 || (roots_scanning_options() & SharedHeap::SO_CodeCache), 5735 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 5736 } 5737 5738 { 5739 TraceTime t("visit unhandled CLDs", PrintGCDetails, false, gclog_or_tty); 5740 5741 verify_work_stacks_empty(); 5742 5743 // Scan all class loader data objects that might have been introduced 5744 // during concurrent marking. 5745 ResourceMark rm; 5746 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds(); 5747 for (int i = 0; i < array->length(); i++) { 5748 mrias_cl.do_class_loader_data(array->at(i)); 5749 } 5750 5751 // We don't need to keep track of new CLDs anymore. 5752 ClassLoaderDataGraph::remember_new_clds(false); 5753 5754 verify_work_stacks_empty(); 5755 } 5756 5757 { 5758 TraceTime t("dirty klass scan", PrintGCDetails, false, gclog_or_tty); 5759 5760 verify_work_stacks_empty(); 5761 5762 RemarkKlassClosure remark_klass_closure(&mrias_cl); 5763 ClassLoaderDataGraph::classes_do(&remark_klass_closure); 5764 5765 verify_work_stacks_empty(); 5766 } 5767 5768 // We might have added oops to ClassLoaderData::_handles during the 5769 // concurrent marking phase. These oops point to newly allocated objects 5770 // that are guaranteed to be kept alive. Either by the direct allocation 5771 // code, or when the young collector processes the strong roots. Hence, 5772 // we don't have to revisit the _handles block during the remark phase. 5773 5774 verify_work_stacks_empty(); 5775 // Restore evacuated mark words, if any, used for overflow list links 5776 if (!CMSOverflowEarlyRestoration) { 5777 restore_preserved_marks_if_any(); 5778 } 5779 verify_overflow_empty(); 5780 } 5781 5782 //////////////////////////////////////////////////////// 5783 // Parallel Reference Processing Task Proxy Class 5784 //////////////////////////////////////////////////////// 5785 class CMSRefProcTaskProxy: public AbstractGangTaskWOopQueues { 5786 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask; 5787 CMSCollector* _collector; 5788 CMSBitMap* _mark_bit_map; 5789 const MemRegion _span; 5790 ProcessTask& _task; 5791 5792 public: 5793 CMSRefProcTaskProxy(ProcessTask& task, 5794 CMSCollector* collector, 5795 const MemRegion& span, 5796 CMSBitMap* mark_bit_map, 5797 AbstractWorkGang* workers, 5798 OopTaskQueueSet* task_queues): 5799 // XXX Should superclass AGTWOQ also know about AWG since it knows 5800 // about the task_queues used by the AWG? Then it could initialize 5801 // the terminator() object. See 6984287. The set_for_termination() 5802 // below is a temporary band-aid for the regression in 6984287. 5803 AbstractGangTaskWOopQueues("Process referents by policy in parallel", 5804 task_queues), 5805 _task(task), 5806 _collector(collector), _span(span), _mark_bit_map(mark_bit_map) 5807 { 5808 assert(_collector->_span.equals(_span) && !_span.is_empty(), 5809 "Inconsistency in _span"); 5810 set_for_termination(workers->active_workers()); 5811 } 5812 5813 OopTaskQueueSet* task_queues() { return queues(); } 5814 5815 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 5816 5817 void do_work_steal(int i, 5818 CMSParDrainMarkingStackClosure* drain, 5819 CMSParKeepAliveClosure* keep_alive, 5820 int* seed); 5821 5822 virtual void work(uint worker_id); 5823 }; 5824 5825 void CMSRefProcTaskProxy::work(uint worker_id) { 5826 assert(_collector->_span.equals(_span), "Inconsistency in _span"); 5827 CMSParKeepAliveClosure par_keep_alive(_collector, _span, 5828 _mark_bit_map, 5829 work_queue(worker_id)); 5830 CMSParDrainMarkingStackClosure par_drain_stack(_collector, _span, 5831 _mark_bit_map, 5832 work_queue(worker_id)); 5833 CMSIsAliveClosure is_alive_closure(_span, _mark_bit_map); 5834 _task.work(worker_id, is_alive_closure, par_keep_alive, par_drain_stack); 5835 if (_task.marks_oops_alive()) { 5836 do_work_steal(worker_id, &par_drain_stack, &par_keep_alive, 5837 _collector->hash_seed(worker_id)); 5838 } 5839 assert(work_queue(worker_id)->size() == 0, "work_queue should be empty"); 5840 assert(_collector->_overflow_list == NULL, "non-empty _overflow_list"); 5841 } 5842 5843 class CMSRefEnqueueTaskProxy: public AbstractGangTask { 5844 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask; 5845 EnqueueTask& _task; 5846 5847 public: 5848 CMSRefEnqueueTaskProxy(EnqueueTask& task) 5849 : AbstractGangTask("Enqueue reference objects in parallel"), 5850 _task(task) 5851 { } 5852 5853 virtual void work(uint worker_id) 5854 { 5855 _task.work(worker_id); 5856 } 5857 }; 5858 5859 CMSParKeepAliveClosure::CMSParKeepAliveClosure(CMSCollector* collector, 5860 MemRegion span, CMSBitMap* bit_map, OopTaskQueue* work_queue): 5861 _span(span), 5862 _bit_map(bit_map), 5863 _work_queue(work_queue), 5864 _mark_and_push(collector, span, bit_map, work_queue), 5865 _low_water_mark(MIN2((uint)(work_queue->max_elems()/4), 5866 (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads))) 5867 { } 5868 5869 // . see if we can share work_queues with ParNew? XXX 5870 void CMSRefProcTaskProxy::do_work_steal(int i, 5871 CMSParDrainMarkingStackClosure* drain, 5872 CMSParKeepAliveClosure* keep_alive, 5873 int* seed) { 5874 OopTaskQueue* work_q = work_queue(i); 5875 NOT_PRODUCT(int num_steals = 0;) 5876 oop obj_to_scan; 5877 5878 while (true) { 5879 // Completely finish any left over work from (an) earlier round(s) 5880 drain->trim_queue(0); 5881 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 5882 (size_t)ParGCDesiredObjsFromOverflowList); 5883 // Now check if there's any work in the overflow list 5884 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads, 5885 // only affects the number of attempts made to get work from the 5886 // overflow list and does not affect the number of workers. Just 5887 // pass ParallelGCThreads so this behavior is unchanged. 5888 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 5889 work_q, 5890 ParallelGCThreads)) { 5891 // Found something in global overflow list; 5892 // not yet ready to go stealing work from others. 5893 // We'd like to assert(work_q->size() != 0, ...) 5894 // because we just took work from the overflow list, 5895 // but of course we can't, since all of that might have 5896 // been already stolen from us. 5897 continue; 5898 } 5899 // Verify that we have no work before we resort to stealing 5900 assert(work_q->size() == 0, "Have work, shouldn't steal"); 5901 // Try to steal from other queues that have work 5902 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 5903 NOT_PRODUCT(num_steals++;) 5904 assert(obj_to_scan->is_oop(), "Oops, not an oop!"); 5905 assert(_mark_bit_map->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 5906 // Do scanning work 5907 obj_to_scan->oop_iterate(keep_alive); 5908 // Loop around, finish this work, and try to steal some more 5909 } else if (terminator()->offer_termination()) { 5910 break; // nirvana from the infinite cycle 5911 } 5912 } 5913 NOT_PRODUCT( 5914 if (PrintCMSStatistics != 0) { 5915 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals); 5916 } 5917 ) 5918 } 5919 5920 void CMSRefProcTaskExecutor::execute(ProcessTask& task) 5921 { 5922 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5923 FlexibleWorkGang* workers = gch->workers(); 5924 assert(workers != NULL, "Need parallel worker threads."); 5925 CMSRefProcTaskProxy rp_task(task, &_collector, 5926 _collector.ref_processor()->span(), 5927 _collector.markBitMap(), 5928 workers, _collector.task_queues()); 5929 workers->run_task(&rp_task); 5930 } 5931 5932 void CMSRefProcTaskExecutor::execute(EnqueueTask& task) 5933 { 5934 5935 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5936 FlexibleWorkGang* workers = gch->workers(); 5937 assert(workers != NULL, "Need parallel worker threads."); 5938 CMSRefEnqueueTaskProxy enq_task(task); 5939 workers->run_task(&enq_task); 5940 } 5941 5942 void CMSCollector::refProcessingWork(bool asynch, bool clear_all_soft_refs) { 5943 5944 ResourceMark rm; 5945 HandleMark hm; 5946 5947 ReferenceProcessor* rp = ref_processor(); 5948 assert(rp->span().equals(_span), "Spans should be equal"); 5949 assert(!rp->enqueuing_is_done(), "Enqueuing should not be complete"); 5950 // Process weak references. 5951 rp->setup_policy(clear_all_soft_refs); 5952 verify_work_stacks_empty(); 5953 5954 CMSKeepAliveClosure cmsKeepAliveClosure(this, _span, &_markBitMap, 5955 &_markStack, false /* !preclean */); 5956 CMSDrainMarkingStackClosure cmsDrainMarkingStackClosure(this, 5957 _span, &_markBitMap, &_markStack, 5958 &cmsKeepAliveClosure, false /* !preclean */); 5959 { 5960 TraceTime t("weak refs processing", PrintGCDetails, false, gclog_or_tty); 5961 if (rp->processing_is_mt()) { 5962 // Set the degree of MT here. If the discovery is done MT, there 5963 // may have been a different number of threads doing the discovery 5964 // and a different number of discovered lists may have Ref objects. 5965 // That is OK as long as the Reference lists are balanced (see 5966 // balance_all_queues() and balance_queues()). 5967 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5968 int active_workers = ParallelGCThreads; 5969 FlexibleWorkGang* workers = gch->workers(); 5970 if (workers != NULL) { 5971 active_workers = workers->active_workers(); 5972 // The expectation is that active_workers will have already 5973 // been set to a reasonable value. If it has not been set, 5974 // investigate. 5975 assert(active_workers > 0, "Should have been set during scavenge"); 5976 } 5977 rp->set_active_mt_degree(active_workers); 5978 CMSRefProcTaskExecutor task_executor(*this); 5979 rp->process_discovered_references(&_is_alive_closure, 5980 &cmsKeepAliveClosure, 5981 &cmsDrainMarkingStackClosure, 5982 &task_executor); 5983 } else { 5984 rp->process_discovered_references(&_is_alive_closure, 5985 &cmsKeepAliveClosure, 5986 &cmsDrainMarkingStackClosure, 5987 NULL); 5988 } 5989 } 5990 5991 // This is the point where the entire marking should have completed. 5992 verify_work_stacks_empty(); 5993 5994 if (should_unload_classes()) { 5995 { 5996 TraceTime t("class unloading", PrintGCDetails, false, gclog_or_tty); 5997 5998 // Unload classes and purge the SystemDictionary. 5999 bool purged_class = SystemDictionary::do_unloading(&_is_alive_closure); 6000 6001 // Unload nmethods. 6002 CodeCache::do_unloading(&_is_alive_closure, purged_class); 6003 6004 // Prune dead klasses from subklass/sibling/implementor lists. 6005 Klass::clean_weak_klass_links(&_is_alive_closure); 6006 } 6007 6008 { 6009 TraceTime t("scrub symbol table", PrintGCDetails, false, gclog_or_tty); 6010 // Clean up unreferenced symbols in symbol table. 6011 SymbolTable::unlink(); 6012 } 6013 } 6014 6015 // CMS doesn't use the StringTable as hard roots when class unloading is turned off. 6016 // Need to check if we really scanned the StringTable. 6017 if ((roots_scanning_options() & SharedHeap::SO_Strings) == 0) { 6018 TraceTime t("scrub string table", PrintGCDetails, false, gclog_or_tty); 6019 // Delete entries for dead interned strings. 6020 StringTable::unlink(&_is_alive_closure); 6021 } 6022 6023 // Restore any preserved marks as a result of mark stack or 6024 // work queue overflow 6025 restore_preserved_marks_if_any(); // done single-threaded for now 6026 6027 rp->set_enqueuing_is_done(true); 6028 if (rp->processing_is_mt()) { 6029 rp->balance_all_queues(); 6030 CMSRefProcTaskExecutor task_executor(*this); 6031 rp->enqueue_discovered_references(&task_executor); 6032 } else { 6033 rp->enqueue_discovered_references(NULL); 6034 } 6035 rp->verify_no_references_recorded(); 6036 assert(!rp->discovery_enabled(), "should have been disabled"); 6037 } 6038 6039 #ifndef PRODUCT 6040 void CMSCollector::check_correct_thread_executing() { 6041 Thread* t = Thread::current(); 6042 // Only the VM thread or the CMS thread should be here. 6043 assert(t->is_ConcurrentGC_thread() || t->is_VM_thread(), 6044 "Unexpected thread type"); 6045 // If this is the vm thread, the foreground process 6046 // should not be waiting. Note that _foregroundGCIsActive is 6047 // true while the foreground collector is waiting. 6048 if (_foregroundGCShouldWait) { 6049 // We cannot be the VM thread 6050 assert(t->is_ConcurrentGC_thread(), 6051 "Should be CMS thread"); 6052 } else { 6053 // We can be the CMS thread only if we are in a stop-world 6054 // phase of CMS collection. 6055 if (t->is_ConcurrentGC_thread()) { 6056 assert(_collectorState == InitialMarking || 6057 _collectorState == FinalMarking, 6058 "Should be a stop-world phase"); 6059 // The CMS thread should be holding the CMS_token. 6060 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6061 "Potential interference with concurrently " 6062 "executing VM thread"); 6063 } 6064 } 6065 } 6066 #endif 6067 6068 void CMSCollector::sweep(bool asynch) { 6069 assert(_collectorState == Sweeping, "just checking"); 6070 check_correct_thread_executing(); 6071 verify_work_stacks_empty(); 6072 verify_overflow_empty(); 6073 increment_sweep_count(); 6074 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 6075 6076 _inter_sweep_timer.stop(); 6077 _inter_sweep_estimate.sample(_inter_sweep_timer.seconds()); 6078 size_policy()->avg_cms_free_at_sweep()->sample(_cmsGen->free()); 6079 6080 assert(!_intra_sweep_timer.is_active(), "Should not be active"); 6081 _intra_sweep_timer.reset(); 6082 _intra_sweep_timer.start(); 6083 if (asynch) { 6084 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 6085 CMSPhaseAccounting pa(this, "sweep", !PrintGCDetails); 6086 // First sweep the old gen 6087 { 6088 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(), 6089 bitMapLock()); 6090 sweepWork(_cmsGen, asynch); 6091 } 6092 6093 // Update Universe::_heap_*_at_gc figures. 6094 // We need all the free list locks to make the abstract state 6095 // transition from Sweeping to Resetting. See detailed note 6096 // further below. 6097 { 6098 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock()); 6099 // Update heap occupancy information which is used as 6100 // input to soft ref clearing policy at the next gc. 6101 Universe::update_heap_info_at_gc(); 6102 _collectorState = Resizing; 6103 } 6104 } else { 6105 // already have needed locks 6106 sweepWork(_cmsGen, asynch); 6107 // Update heap occupancy information which is used as 6108 // input to soft ref clearing policy at the next gc. 6109 Universe::update_heap_info_at_gc(); 6110 _collectorState = Resizing; 6111 } 6112 verify_work_stacks_empty(); 6113 verify_overflow_empty(); 6114 6115 if (should_unload_classes()) { 6116 ClassLoaderDataGraph::purge(); 6117 } 6118 6119 _intra_sweep_timer.stop(); 6120 _intra_sweep_estimate.sample(_intra_sweep_timer.seconds()); 6121 6122 _inter_sweep_timer.reset(); 6123 _inter_sweep_timer.start(); 6124 6125 // We need to use a monotonically non-deccreasing time in ms 6126 // or we will see time-warp warnings and os::javaTimeMillis() 6127 // does not guarantee monotonicity. 6128 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC; 6129 update_time_of_last_gc(now); 6130 6131 // NOTE on abstract state transitions: 6132 // Mutators allocate-live and/or mark the mod-union table dirty 6133 // based on the state of the collection. The former is done in 6134 // the interval [Marking, Sweeping] and the latter in the interval 6135 // [Marking, Sweeping). Thus the transitions into the Marking state 6136 // and out of the Sweeping state must be synchronously visible 6137 // globally to the mutators. 6138 // The transition into the Marking state happens with the world 6139 // stopped so the mutators will globally see it. Sweeping is 6140 // done asynchronously by the background collector so the transition 6141 // from the Sweeping state to the Resizing state must be done 6142 // under the freelistLock (as is the check for whether to 6143 // allocate-live and whether to dirty the mod-union table). 6144 assert(_collectorState == Resizing, "Change of collector state to" 6145 " Resizing must be done under the freelistLocks (plural)"); 6146 6147 // Now that sweeping has been completed, we clear 6148 // the incremental_collection_failed flag, 6149 // thus inviting a younger gen collection to promote into 6150 // this generation. If such a promotion may still fail, 6151 // the flag will be set again when a young collection is 6152 // attempted. 6153 GenCollectedHeap* gch = GenCollectedHeap::heap(); 6154 gch->clear_incremental_collection_failed(); // Worth retrying as fresh space may have been freed up 6155 gch->update_full_collections_completed(_collection_count_start); 6156 } 6157 6158 // FIX ME!!! Looks like this belongs in CFLSpace, with 6159 // CMSGen merely delegating to it. 6160 void ConcurrentMarkSweepGeneration::setNearLargestChunk() { 6161 double nearLargestPercent = FLSLargestBlockCoalesceProximity; 6162 HeapWord* minAddr = _cmsSpace->bottom(); 6163 HeapWord* largestAddr = 6164 (HeapWord*) _cmsSpace->dictionary()->find_largest_dict(); 6165 if (largestAddr == NULL) { 6166 // The dictionary appears to be empty. In this case 6167 // try to coalesce at the end of the heap. 6168 largestAddr = _cmsSpace->end(); 6169 } 6170 size_t largestOffset = pointer_delta(largestAddr, minAddr); 6171 size_t nearLargestOffset = 6172 (size_t)((double)largestOffset * nearLargestPercent) - MinChunkSize; 6173 if (PrintFLSStatistics != 0) { 6174 gclog_or_tty->print_cr( 6175 "CMS: Large Block: " PTR_FORMAT ";" 6176 " Proximity: " PTR_FORMAT " -> " PTR_FORMAT, 6177 largestAddr, 6178 _cmsSpace->nearLargestChunk(), minAddr + nearLargestOffset); 6179 } 6180 _cmsSpace->set_nearLargestChunk(minAddr + nearLargestOffset); 6181 } 6182 6183 bool ConcurrentMarkSweepGeneration::isNearLargestChunk(HeapWord* addr) { 6184 return addr >= _cmsSpace->nearLargestChunk(); 6185 } 6186 6187 FreeChunk* ConcurrentMarkSweepGeneration::find_chunk_at_end() { 6188 return _cmsSpace->find_chunk_at_end(); 6189 } 6190 6191 void ConcurrentMarkSweepGeneration::update_gc_stats(int current_level, 6192 bool full) { 6193 // The next lower level has been collected. Gather any statistics 6194 // that are of interest at this point. 6195 if (!full && (current_level + 1) == level()) { 6196 // Gather statistics on the young generation collection. 6197 collector()->stats().record_gc0_end(used()); 6198 } 6199 } 6200 6201 CMSAdaptiveSizePolicy* ConcurrentMarkSweepGeneration::size_policy() { 6202 GenCollectedHeap* gch = GenCollectedHeap::heap(); 6203 assert(gch->kind() == CollectedHeap::GenCollectedHeap, 6204 "Wrong type of heap"); 6205 CMSAdaptiveSizePolicy* sp = (CMSAdaptiveSizePolicy*) 6206 gch->gen_policy()->size_policy(); 6207 assert(sp->is_gc_cms_adaptive_size_policy(), 6208 "Wrong type of size policy"); 6209 return sp; 6210 } 6211 6212 void ConcurrentMarkSweepGeneration::rotate_debug_collection_type() { 6213 if (PrintGCDetails && Verbose) { 6214 gclog_or_tty->print("Rotate from %d ", _debug_collection_type); 6215 } 6216 _debug_collection_type = (CollectionTypes) (_debug_collection_type + 1); 6217 _debug_collection_type = 6218 (CollectionTypes) (_debug_collection_type % Unknown_collection_type); 6219 if (PrintGCDetails && Verbose) { 6220 gclog_or_tty->print_cr("to %d ", _debug_collection_type); 6221 } 6222 } 6223 6224 void CMSCollector::sweepWork(ConcurrentMarkSweepGeneration* gen, 6225 bool asynch) { 6226 // We iterate over the space(s) underlying this generation, 6227 // checking the mark bit map to see if the bits corresponding 6228 // to specific blocks are marked or not. Blocks that are 6229 // marked are live and are not swept up. All remaining blocks 6230 // are swept up, with coalescing on-the-fly as we sweep up 6231 // contiguous free and/or garbage blocks: 6232 // We need to ensure that the sweeper synchronizes with allocators 6233 // and stop-the-world collectors. In particular, the following 6234 // locks are used: 6235 // . CMS token: if this is held, a stop the world collection cannot occur 6236 // . freelistLock: if this is held no allocation can occur from this 6237 // generation by another thread 6238 // . bitMapLock: if this is held, no other thread can access or update 6239 // 6240 6241 // Note that we need to hold the freelistLock if we use 6242 // block iterate below; else the iterator might go awry if 6243 // a mutator (or promotion) causes block contents to change 6244 // (for instance if the allocator divvies up a block). 6245 // If we hold the free list lock, for all practical purposes 6246 // young generation GC's can't occur (they'll usually need to 6247 // promote), so we might as well prevent all young generation 6248 // GC's while we do a sweeping step. For the same reason, we might 6249 // as well take the bit map lock for the entire duration 6250 6251 // check that we hold the requisite locks 6252 assert(have_cms_token(), "Should hold cms token"); 6253 assert( (asynch && ConcurrentMarkSweepThread::cms_thread_has_cms_token()) 6254 || (!asynch && ConcurrentMarkSweepThread::vm_thread_has_cms_token()), 6255 "Should possess CMS token to sweep"); 6256 assert_lock_strong(gen->freelistLock()); 6257 assert_lock_strong(bitMapLock()); 6258 6259 assert(!_inter_sweep_timer.is_active(), "Was switched off in an outer context"); 6260 assert(_intra_sweep_timer.is_active(), "Was switched on in an outer context"); 6261 gen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()), 6262 _inter_sweep_estimate.padded_average(), 6263 _intra_sweep_estimate.padded_average()); 6264 gen->setNearLargestChunk(); 6265 6266 { 6267 SweepClosure sweepClosure(this, gen, &_markBitMap, 6268 CMSYield && asynch); 6269 gen->cmsSpace()->blk_iterate_careful(&sweepClosure); 6270 // We need to free-up/coalesce garbage/blocks from a 6271 // co-terminal free run. This is done in the SweepClosure 6272 // destructor; so, do not remove this scope, else the 6273 // end-of-sweep-census below will be off by a little bit. 6274 } 6275 gen->cmsSpace()->sweep_completed(); 6276 gen->cmsSpace()->endSweepFLCensus(sweep_count()); 6277 if (should_unload_classes()) { // unloaded classes this cycle, 6278 _concurrent_cycles_since_last_unload = 0; // ... reset count 6279 } else { // did not unload classes, 6280 _concurrent_cycles_since_last_unload++; // ... increment count 6281 } 6282 } 6283 6284 // Reset CMS data structures (for now just the marking bit map) 6285 // preparatory for the next cycle. 6286 void CMSCollector::reset(bool asynch) { 6287 GenCollectedHeap* gch = GenCollectedHeap::heap(); 6288 CMSAdaptiveSizePolicy* sp = size_policy(); 6289 AdaptiveSizePolicyOutput(sp, gch->total_collections()); 6290 if (asynch) { 6291 CMSTokenSyncWithLocks ts(true, bitMapLock()); 6292 6293 // If the state is not "Resetting", the foreground thread 6294 // has done a collection and the resetting. 6295 if (_collectorState != Resetting) { 6296 assert(_collectorState == Idling, "The state should only change" 6297 " because the foreground collector has finished the collection"); 6298 return; 6299 } 6300 6301 // Clear the mark bitmap (no grey objects to start with) 6302 // for the next cycle. 6303 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 6304 CMSPhaseAccounting cmspa(this, "reset", !PrintGCDetails); 6305 6306 HeapWord* curAddr = _markBitMap.startWord(); 6307 while (curAddr < _markBitMap.endWord()) { 6308 size_t remaining = pointer_delta(_markBitMap.endWord(), curAddr); 6309 MemRegion chunk(curAddr, MIN2(CMSBitMapYieldQuantum, remaining)); 6310 _markBitMap.clear_large_range(chunk); 6311 if (ConcurrentMarkSweepThread::should_yield() && 6312 !foregroundGCIsActive() && 6313 CMSYield) { 6314 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6315 "CMS thread should hold CMS token"); 6316 assert_lock_strong(bitMapLock()); 6317 bitMapLock()->unlock(); 6318 ConcurrentMarkSweepThread::desynchronize(true); 6319 ConcurrentMarkSweepThread::acknowledge_yield_request(); 6320 stopTimer(); 6321 if (PrintCMSStatistics != 0) { 6322 incrementYields(); 6323 } 6324 icms_wait(); 6325 6326 // See the comment in coordinator_yield() 6327 for (unsigned i = 0; i < CMSYieldSleepCount && 6328 ConcurrentMarkSweepThread::should_yield() && 6329 !CMSCollector::foregroundGCIsActive(); ++i) { 6330 os::sleep(Thread::current(), 1, false); 6331 ConcurrentMarkSweepThread::acknowledge_yield_request(); 6332 } 6333 6334 ConcurrentMarkSweepThread::synchronize(true); 6335 bitMapLock()->lock_without_safepoint_check(); 6336 startTimer(); 6337 } 6338 curAddr = chunk.end(); 6339 } 6340 // A successful mostly concurrent collection has been done. 6341 // Because only the full (i.e., concurrent mode failure) collections 6342 // are being measured for gc overhead limits, clean the "near" flag 6343 // and count. 6344 sp->reset_gc_overhead_limit_count(); 6345 _collectorState = Idling; 6346 } else { 6347 // already have the lock 6348 assert(_collectorState == Resetting, "just checking"); 6349 assert_lock_strong(bitMapLock()); 6350 _markBitMap.clear_all(); 6351 _collectorState = Idling; 6352 } 6353 6354 // Stop incremental mode after a cycle completes, so that any future cycles 6355 // are triggered by allocation. 6356 stop_icms(); 6357 6358 NOT_PRODUCT( 6359 if (RotateCMSCollectionTypes) { 6360 _cmsGen->rotate_debug_collection_type(); 6361 } 6362 ) 6363 } 6364 6365 void CMSCollector::do_CMS_operation(CMS_op_type op, GCCause::Cause gc_cause) { 6366 gclog_or_tty->date_stamp(PrintGC && PrintGCDateStamps); 6367 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 6368 TraceTime t(GCCauseString("GC", gc_cause), PrintGC, !PrintGCDetails, gclog_or_tty); 6369 TraceCollectorStats tcs(counters()); 6370 6371 switch (op) { 6372 case CMS_op_checkpointRootsInitial: { 6373 SvcGCMarker sgcm(SvcGCMarker::OTHER); 6374 checkpointRootsInitial(true); // asynch 6375 if (PrintGC) { 6376 _cmsGen->printOccupancy("initial-mark"); 6377 } 6378 break; 6379 } 6380 case CMS_op_checkpointRootsFinal: { 6381 SvcGCMarker sgcm(SvcGCMarker::OTHER); 6382 checkpointRootsFinal(true, // asynch 6383 false, // !clear_all_soft_refs 6384 false); // !init_mark_was_synchronous 6385 if (PrintGC) { 6386 _cmsGen->printOccupancy("remark"); 6387 } 6388 break; 6389 } 6390 default: 6391 fatal("No such CMS_op"); 6392 } 6393 } 6394 6395 #ifndef PRODUCT 6396 size_t const CMSCollector::skip_header_HeapWords() { 6397 return FreeChunk::header_size(); 6398 } 6399 6400 // Try and collect here conditions that should hold when 6401 // CMS thread is exiting. The idea is that the foreground GC 6402 // thread should not be blocked if it wants to terminate 6403 // the CMS thread and yet continue to run the VM for a while 6404 // after that. 6405 void CMSCollector::verify_ok_to_terminate() const { 6406 assert(Thread::current()->is_ConcurrentGC_thread(), 6407 "should be called by CMS thread"); 6408 assert(!_foregroundGCShouldWait, "should be false"); 6409 // We could check here that all the various low-level locks 6410 // are not held by the CMS thread, but that is overkill; see 6411 // also CMSThread::verify_ok_to_terminate() where the CGC_lock 6412 // is checked. 6413 } 6414 #endif 6415 6416 size_t CMSCollector::block_size_using_printezis_bits(HeapWord* addr) const { 6417 assert(_markBitMap.isMarked(addr) && _markBitMap.isMarked(addr + 1), 6418 "missing Printezis mark?"); 6419 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 6420 size_t size = pointer_delta(nextOneAddr + 1, addr); 6421 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 6422 "alignment problem"); 6423 assert(size >= 3, "Necessary for Printezis marks to work"); 6424 return size; 6425 } 6426 6427 // A variant of the above (block_size_using_printezis_bits()) except 6428 // that we return 0 if the P-bits are not yet set. 6429 size_t CMSCollector::block_size_if_printezis_bits(HeapWord* addr) const { 6430 if (_markBitMap.isMarked(addr + 1)) { 6431 assert(_markBitMap.isMarked(addr), "P-bit can be set only for marked objects"); 6432 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 6433 size_t size = pointer_delta(nextOneAddr + 1, addr); 6434 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 6435 "alignment problem"); 6436 assert(size >= 3, "Necessary for Printezis marks to work"); 6437 return size; 6438 } 6439 return 0; 6440 } 6441 6442 HeapWord* CMSCollector::next_card_start_after_block(HeapWord* addr) const { 6443 size_t sz = 0; 6444 oop p = (oop)addr; 6445 if (p->klass_or_null() != NULL) { 6446 sz = CompactibleFreeListSpace::adjustObjectSize(p->size()); 6447 } else { 6448 sz = block_size_using_printezis_bits(addr); 6449 } 6450 assert(sz > 0, "size must be nonzero"); 6451 HeapWord* next_block = addr + sz; 6452 HeapWord* next_card = (HeapWord*)round_to((uintptr_t)next_block, 6453 CardTableModRefBS::card_size); 6454 assert(round_down((uintptr_t)addr, CardTableModRefBS::card_size) < 6455 round_down((uintptr_t)next_card, CardTableModRefBS::card_size), 6456 "must be different cards"); 6457 return next_card; 6458 } 6459 6460 6461 // CMS Bit Map Wrapper ///////////////////////////////////////// 6462 6463 // Construct a CMS bit map infrastructure, but don't create the 6464 // bit vector itself. That is done by a separate call CMSBitMap::allocate() 6465 // further below. 6466 CMSBitMap::CMSBitMap(int shifter, int mutex_rank, const char* mutex_name): 6467 _bm(), 6468 _shifter(shifter), 6469 _lock(mutex_rank >= 0 ? new Mutex(mutex_rank, mutex_name, true) : NULL) 6470 { 6471 _bmStartWord = 0; 6472 _bmWordSize = 0; 6473 } 6474 6475 bool CMSBitMap::allocate(MemRegion mr) { 6476 _bmStartWord = mr.start(); 6477 _bmWordSize = mr.word_size(); 6478 ReservedSpace brs(ReservedSpace::allocation_align_size_up( 6479 (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1)); 6480 if (!brs.is_reserved()) { 6481 warning("CMS bit map allocation failure"); 6482 return false; 6483 } 6484 // For now we'll just commit all of the bit map up fromt. 6485 // Later on we'll try to be more parsimonious with swap. 6486 if (!_virtual_space.initialize(brs, brs.size())) { 6487 warning("CMS bit map backing store failure"); 6488 return false; 6489 } 6490 assert(_virtual_space.committed_size() == brs.size(), 6491 "didn't reserve backing store for all of CMS bit map?"); 6492 _bm.set_map((BitMap::bm_word_t*)_virtual_space.low()); 6493 assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >= 6494 _bmWordSize, "inconsistency in bit map sizing"); 6495 _bm.set_size(_bmWordSize >> _shifter); 6496 6497 // bm.clear(); // can we rely on getting zero'd memory? verify below 6498 assert(isAllClear(), 6499 "Expected zero'd memory from ReservedSpace constructor"); 6500 assert(_bm.size() == heapWordDiffToOffsetDiff(sizeInWords()), 6501 "consistency check"); 6502 return true; 6503 } 6504 6505 void CMSBitMap::dirty_range_iterate_clear(MemRegion mr, MemRegionClosure* cl) { 6506 HeapWord *next_addr, *end_addr, *last_addr; 6507 assert_locked(); 6508 assert(covers(mr), "out-of-range error"); 6509 // XXX assert that start and end are appropriately aligned 6510 for (next_addr = mr.start(), end_addr = mr.end(); 6511 next_addr < end_addr; next_addr = last_addr) { 6512 MemRegion dirty_region = getAndClearMarkedRegion(next_addr, end_addr); 6513 last_addr = dirty_region.end(); 6514 if (!dirty_region.is_empty()) { 6515 cl->do_MemRegion(dirty_region); 6516 } else { 6517 assert(last_addr == end_addr, "program logic"); 6518 return; 6519 } 6520 } 6521 } 6522 6523 void CMSBitMap::print_on_error(outputStream* st, const char* prefix) const { 6524 _bm.print_on_error(st, prefix); 6525 } 6526 6527 #ifndef PRODUCT 6528 void CMSBitMap::assert_locked() const { 6529 CMSLockVerifier::assert_locked(lock()); 6530 } 6531 6532 bool CMSBitMap::covers(MemRegion mr) const { 6533 // assert(_bm.map() == _virtual_space.low(), "map inconsistency"); 6534 assert((size_t)_bm.size() == (_bmWordSize >> _shifter), 6535 "size inconsistency"); 6536 return (mr.start() >= _bmStartWord) && 6537 (mr.end() <= endWord()); 6538 } 6539 6540 bool CMSBitMap::covers(HeapWord* start, size_t size) const { 6541 return (start >= _bmStartWord && (start + size) <= endWord()); 6542 } 6543 6544 void CMSBitMap::verifyNoOneBitsInRange(HeapWord* left, HeapWord* right) { 6545 // verify that there are no 1 bits in the interval [left, right) 6546 FalseBitMapClosure falseBitMapClosure; 6547 iterate(&falseBitMapClosure, left, right); 6548 } 6549 6550 void CMSBitMap::region_invariant(MemRegion mr) 6551 { 6552 assert_locked(); 6553 // mr = mr.intersection(MemRegion(_bmStartWord, _bmWordSize)); 6554 assert(!mr.is_empty(), "unexpected empty region"); 6555 assert(covers(mr), "mr should be covered by bit map"); 6556 // convert address range into offset range 6557 size_t start_ofs = heapWordToOffset(mr.start()); 6558 // Make sure that end() is appropriately aligned 6559 assert(mr.end() == (HeapWord*)round_to((intptr_t)mr.end(), 6560 (1 << (_shifter+LogHeapWordSize))), 6561 "Misaligned mr.end()"); 6562 size_t end_ofs = heapWordToOffset(mr.end()); 6563 assert(end_ofs > start_ofs, "Should mark at least one bit"); 6564 } 6565 6566 #endif 6567 6568 bool CMSMarkStack::allocate(size_t size) { 6569 // allocate a stack of the requisite depth 6570 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 6571 size * sizeof(oop))); 6572 if (!rs.is_reserved()) { 6573 warning("CMSMarkStack allocation failure"); 6574 return false; 6575 } 6576 if (!_virtual_space.initialize(rs, rs.size())) { 6577 warning("CMSMarkStack backing store failure"); 6578 return false; 6579 } 6580 assert(_virtual_space.committed_size() == rs.size(), 6581 "didn't reserve backing store for all of CMS stack?"); 6582 _base = (oop*)(_virtual_space.low()); 6583 _index = 0; 6584 _capacity = size; 6585 NOT_PRODUCT(_max_depth = 0); 6586 return true; 6587 } 6588 6589 // XXX FIX ME !!! In the MT case we come in here holding a 6590 // leaf lock. For printing we need to take a further lock 6591 // which has lower rank. We need to recallibrate the two 6592 // lock-ranks involved in order to be able to rpint the 6593 // messages below. (Or defer the printing to the caller. 6594 // For now we take the expedient path of just disabling the 6595 // messages for the problematic case.) 6596 void CMSMarkStack::expand() { 6597 assert(_capacity <= MarkStackSizeMax, "stack bigger than permitted"); 6598 if (_capacity == MarkStackSizeMax) { 6599 if (_hit_limit++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) { 6600 // We print a warning message only once per CMS cycle. 6601 gclog_or_tty->print_cr(" (benign) Hit CMSMarkStack max size limit"); 6602 } 6603 return; 6604 } 6605 // Double capacity if possible 6606 size_t new_capacity = MIN2(_capacity*2, MarkStackSizeMax); 6607 // Do not give up existing stack until we have managed to 6608 // get the double capacity that we desired. 6609 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 6610 new_capacity * sizeof(oop))); 6611 if (rs.is_reserved()) { 6612 // Release the backing store associated with old stack 6613 _virtual_space.release(); 6614 // Reinitialize virtual space for new stack 6615 if (!_virtual_space.initialize(rs, rs.size())) { 6616 fatal("Not enough swap for expanded marking stack"); 6617 } 6618 _base = (oop*)(_virtual_space.low()); 6619 _index = 0; 6620 _capacity = new_capacity; 6621 } else if (_failed_double++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) { 6622 // Failed to double capacity, continue; 6623 // we print a detail message only once per CMS cycle. 6624 gclog_or_tty->print(" (benign) Failed to expand marking stack from "SIZE_FORMAT"K to " 6625 SIZE_FORMAT"K", 6626 _capacity / K, new_capacity / K); 6627 } 6628 } 6629 6630 6631 // Closures 6632 // XXX: there seems to be a lot of code duplication here; 6633 // should refactor and consolidate common code. 6634 6635 // This closure is used to mark refs into the CMS generation in 6636 // the CMS bit map. Called at the first checkpoint. This closure 6637 // assumes that we do not need to re-mark dirty cards; if the CMS 6638 // generation on which this is used is not an oldest 6639 // generation then this will lose younger_gen cards! 6640 6641 MarkRefsIntoClosure::MarkRefsIntoClosure( 6642 MemRegion span, CMSBitMap* bitMap): 6643 _span(span), 6644 _bitMap(bitMap) 6645 { 6646 assert(_ref_processor == NULL, "deliberately left NULL"); 6647 assert(_bitMap->covers(_span), "_bitMap/_span mismatch"); 6648 } 6649 6650 void MarkRefsIntoClosure::do_oop(oop obj) { 6651 // if p points into _span, then mark corresponding bit in _markBitMap 6652 assert(obj->is_oop(), "expected an oop"); 6653 HeapWord* addr = (HeapWord*)obj; 6654 if (_span.contains(addr)) { 6655 // this should be made more efficient 6656 _bitMap->mark(addr); 6657 } 6658 } 6659 6660 void MarkRefsIntoClosure::do_oop(oop* p) { MarkRefsIntoClosure::do_oop_work(p); } 6661 void MarkRefsIntoClosure::do_oop(narrowOop* p) { MarkRefsIntoClosure::do_oop_work(p); } 6662 6663 // A variant of the above, used for CMS marking verification. 6664 MarkRefsIntoVerifyClosure::MarkRefsIntoVerifyClosure( 6665 MemRegion span, CMSBitMap* verification_bm, CMSBitMap* cms_bm): 6666 _span(span), 6667 _verification_bm(verification_bm), 6668 _cms_bm(cms_bm) 6669 { 6670 assert(_ref_processor == NULL, "deliberately left NULL"); 6671 assert(_verification_bm->covers(_span), "_verification_bm/_span mismatch"); 6672 } 6673 6674 void MarkRefsIntoVerifyClosure::do_oop(oop obj) { 6675 // if p points into _span, then mark corresponding bit in _markBitMap 6676 assert(obj->is_oop(), "expected an oop"); 6677 HeapWord* addr = (HeapWord*)obj; 6678 if (_span.contains(addr)) { 6679 _verification_bm->mark(addr); 6680 if (!_cms_bm->isMarked(addr)) { 6681 oop(addr)->print(); 6682 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)", addr); 6683 fatal("... aborting"); 6684 } 6685 } 6686 } 6687 6688 void MarkRefsIntoVerifyClosure::do_oop(oop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); } 6689 void MarkRefsIntoVerifyClosure::do_oop(narrowOop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); } 6690 6691 ////////////////////////////////////////////////// 6692 // MarkRefsIntoAndScanClosure 6693 ////////////////////////////////////////////////// 6694 6695 MarkRefsIntoAndScanClosure::MarkRefsIntoAndScanClosure(MemRegion span, 6696 ReferenceProcessor* rp, 6697 CMSBitMap* bit_map, 6698 CMSBitMap* mod_union_table, 6699 CMSMarkStack* mark_stack, 6700 CMSCollector* collector, 6701 bool should_yield, 6702 bool concurrent_precleaning): 6703 _collector(collector), 6704 _span(span), 6705 _bit_map(bit_map), 6706 _mark_stack(mark_stack), 6707 _pushAndMarkClosure(collector, span, rp, bit_map, mod_union_table, 6708 mark_stack, concurrent_precleaning), 6709 _yield(should_yield), 6710 _concurrent_precleaning(concurrent_precleaning), 6711 _freelistLock(NULL) 6712 { 6713 _ref_processor = rp; 6714 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 6715 } 6716 6717 // This closure is used to mark refs into the CMS generation at the 6718 // second (final) checkpoint, and to scan and transitively follow 6719 // the unmarked oops. It is also used during the concurrent precleaning 6720 // phase while scanning objects on dirty cards in the CMS generation. 6721 // The marks are made in the marking bit map and the marking stack is 6722 // used for keeping the (newly) grey objects during the scan. 6723 // The parallel version (Par_...) appears further below. 6724 void MarkRefsIntoAndScanClosure::do_oop(oop obj) { 6725 if (obj != NULL) { 6726 assert(obj->is_oop(), "expected an oop"); 6727 HeapWord* addr = (HeapWord*)obj; 6728 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 6729 assert(_collector->overflow_list_is_empty(), 6730 "overflow list should be empty"); 6731 if (_span.contains(addr) && 6732 !_bit_map->isMarked(addr)) { 6733 // mark bit map (object is now grey) 6734 _bit_map->mark(addr); 6735 // push on marking stack (stack should be empty), and drain the 6736 // stack by applying this closure to the oops in the oops popped 6737 // from the stack (i.e. blacken the grey objects) 6738 bool res = _mark_stack->push(obj); 6739 assert(res, "Should have space to push on empty stack"); 6740 do { 6741 oop new_oop = _mark_stack->pop(); 6742 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 6743 assert(_bit_map->isMarked((HeapWord*)new_oop), 6744 "only grey objects on this stack"); 6745 // iterate over the oops in this oop, marking and pushing 6746 // the ones in CMS heap (i.e. in _span). 6747 new_oop->oop_iterate(&_pushAndMarkClosure); 6748 // check if it's time to yield 6749 do_yield_check(); 6750 } while (!_mark_stack->isEmpty() || 6751 (!_concurrent_precleaning && take_from_overflow_list())); 6752 // if marking stack is empty, and we are not doing this 6753 // during precleaning, then check the overflow list 6754 } 6755 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 6756 assert(_collector->overflow_list_is_empty(), 6757 "overflow list was drained above"); 6758 // We could restore evacuated mark words, if any, used for 6759 // overflow list links here because the overflow list is 6760 // provably empty here. That would reduce the maximum 6761 // size requirements for preserved_{oop,mark}_stack. 6762 // But we'll just postpone it until we are all done 6763 // so we can just stream through. 6764 if (!_concurrent_precleaning && CMSOverflowEarlyRestoration) { 6765 _collector->restore_preserved_marks_if_any(); 6766 assert(_collector->no_preserved_marks(), "No preserved marks"); 6767 } 6768 assert(!CMSOverflowEarlyRestoration || _collector->no_preserved_marks(), 6769 "All preserved marks should have been restored above"); 6770 } 6771 } 6772 6773 void MarkRefsIntoAndScanClosure::do_oop(oop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); } 6774 void MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); } 6775 6776 void MarkRefsIntoAndScanClosure::do_yield_work() { 6777 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6778 "CMS thread should hold CMS token"); 6779 assert_lock_strong(_freelistLock); 6780 assert_lock_strong(_bit_map->lock()); 6781 // relinquish the free_list_lock and bitMaplock() 6782 _bit_map->lock()->unlock(); 6783 _freelistLock->unlock(); 6784 ConcurrentMarkSweepThread::desynchronize(true); 6785 ConcurrentMarkSweepThread::acknowledge_yield_request(); 6786 _collector->stopTimer(); 6787 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 6788 if (PrintCMSStatistics != 0) { 6789 _collector->incrementYields(); 6790 } 6791 _collector->icms_wait(); 6792 6793 // See the comment in coordinator_yield() 6794 for (unsigned i = 0; 6795 i < CMSYieldSleepCount && 6796 ConcurrentMarkSweepThread::should_yield() && 6797 !CMSCollector::foregroundGCIsActive(); 6798 ++i) { 6799 os::sleep(Thread::current(), 1, false); 6800 ConcurrentMarkSweepThread::acknowledge_yield_request(); 6801 } 6802 6803 ConcurrentMarkSweepThread::synchronize(true); 6804 _freelistLock->lock_without_safepoint_check(); 6805 _bit_map->lock()->lock_without_safepoint_check(); 6806 _collector->startTimer(); 6807 } 6808 6809 /////////////////////////////////////////////////////////// 6810 // Par_MarkRefsIntoAndScanClosure: a parallel version of 6811 // MarkRefsIntoAndScanClosure 6812 /////////////////////////////////////////////////////////// 6813 Par_MarkRefsIntoAndScanClosure::Par_MarkRefsIntoAndScanClosure( 6814 CMSCollector* collector, MemRegion span, ReferenceProcessor* rp, 6815 CMSBitMap* bit_map, OopTaskQueue* work_queue): 6816 _span(span), 6817 _bit_map(bit_map), 6818 _work_queue(work_queue), 6819 _low_water_mark(MIN2((uint)(work_queue->max_elems()/4), 6820 (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads))), 6821 _par_pushAndMarkClosure(collector, span, rp, bit_map, work_queue) 6822 { 6823 _ref_processor = rp; 6824 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 6825 } 6826 6827 // This closure is used to mark refs into the CMS generation at the 6828 // second (final) checkpoint, and to scan and transitively follow 6829 // the unmarked oops. The marks are made in the marking bit map and 6830 // the work_queue is used for keeping the (newly) grey objects during 6831 // the scan phase whence they are also available for stealing by parallel 6832 // threads. Since the marking bit map is shared, updates are 6833 // synchronized (via CAS). 6834 void Par_MarkRefsIntoAndScanClosure::do_oop(oop obj) { 6835 if (obj != NULL) { 6836 // Ignore mark word because this could be an already marked oop 6837 // that may be chained at the end of the overflow list. 6838 assert(obj->is_oop(true), "expected an oop"); 6839 HeapWord* addr = (HeapWord*)obj; 6840 if (_span.contains(addr) && 6841 !_bit_map->isMarked(addr)) { 6842 // mark bit map (object will become grey): 6843 // It is possible for several threads to be 6844 // trying to "claim" this object concurrently; 6845 // the unique thread that succeeds in marking the 6846 // object first will do the subsequent push on 6847 // to the work queue (or overflow list). 6848 if (_bit_map->par_mark(addr)) { 6849 // push on work_queue (which may not be empty), and trim the 6850 // queue to an appropriate length by applying this closure to 6851 // the oops in the oops popped from the stack (i.e. blacken the 6852 // grey objects) 6853 bool res = _work_queue->push(obj); 6854 assert(res, "Low water mark should be less than capacity?"); 6855 trim_queue(_low_water_mark); 6856 } // Else, another thread claimed the object 6857 } 6858 } 6859 } 6860 6861 void Par_MarkRefsIntoAndScanClosure::do_oop(oop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); } 6862 void Par_MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); } 6863 6864 // This closure is used to rescan the marked objects on the dirty cards 6865 // in the mod union table and the card table proper. 6866 size_t ScanMarkedObjectsAgainCarefullyClosure::do_object_careful_m( 6867 oop p, MemRegion mr) { 6868 6869 size_t size = 0; 6870 HeapWord* addr = (HeapWord*)p; 6871 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6872 assert(_span.contains(addr), "we are scanning the CMS generation"); 6873 // check if it's time to yield 6874 if (do_yield_check()) { 6875 // We yielded for some foreground stop-world work, 6876 // and we have been asked to abort this ongoing preclean cycle. 6877 return 0; 6878 } 6879 if (_bitMap->isMarked(addr)) { 6880 // it's marked; is it potentially uninitialized? 6881 if (p->klass_or_null() != NULL) { 6882 // an initialized object; ignore mark word in verification below 6883 // since we are running concurrent with mutators 6884 assert(p->is_oop(true), "should be an oop"); 6885 if (p->is_objArray()) { 6886 // objArrays are precisely marked; restrict scanning 6887 // to dirty cards only. 6888 size = CompactibleFreeListSpace::adjustObjectSize( 6889 p->oop_iterate(_scanningClosure, mr)); 6890 } else { 6891 // A non-array may have been imprecisely marked; we need 6892 // to scan object in its entirety. 6893 size = CompactibleFreeListSpace::adjustObjectSize( 6894 p->oop_iterate(_scanningClosure)); 6895 } 6896 #ifdef ASSERT 6897 size_t direct_size = 6898 CompactibleFreeListSpace::adjustObjectSize(p->size()); 6899 assert(size == direct_size, "Inconsistency in size"); 6900 assert(size >= 3, "Necessary for Printezis marks to work"); 6901 if (!_bitMap->isMarked(addr+1)) { 6902 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size); 6903 } else { 6904 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size-1); 6905 assert(_bitMap->isMarked(addr+size-1), 6906 "inconsistent Printezis mark"); 6907 } 6908 #endif // ASSERT 6909 } else { 6910 // an unitialized object 6911 assert(_bitMap->isMarked(addr+1), "missing Printezis mark?"); 6912 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 6913 size = pointer_delta(nextOneAddr + 1, addr); 6914 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 6915 "alignment problem"); 6916 // Note that pre-cleaning needn't redirty the card. OopDesc::set_klass() 6917 // will dirty the card when the klass pointer is installed in the 6918 // object (signalling the completion of initialization). 6919 } 6920 } else { 6921 // Either a not yet marked object or an uninitialized object 6922 if (p->klass_or_null() == NULL) { 6923 // An uninitialized object, skip to the next card, since 6924 // we may not be able to read its P-bits yet. 6925 assert(size == 0, "Initial value"); 6926 } else { 6927 // An object not (yet) reached by marking: we merely need to 6928 // compute its size so as to go look at the next block. 6929 assert(p->is_oop(true), "should be an oop"); 6930 size = CompactibleFreeListSpace::adjustObjectSize(p->size()); 6931 } 6932 } 6933 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6934 return size; 6935 } 6936 6937 void ScanMarkedObjectsAgainCarefullyClosure::do_yield_work() { 6938 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6939 "CMS thread should hold CMS token"); 6940 assert_lock_strong(_freelistLock); 6941 assert_lock_strong(_bitMap->lock()); 6942 // relinquish the free_list_lock and bitMaplock() 6943 _bitMap->lock()->unlock(); 6944 _freelistLock->unlock(); 6945 ConcurrentMarkSweepThread::desynchronize(true); 6946 ConcurrentMarkSweepThread::acknowledge_yield_request(); 6947 _collector->stopTimer(); 6948 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 6949 if (PrintCMSStatistics != 0) { 6950 _collector->incrementYields(); 6951 } 6952 _collector->icms_wait(); 6953 6954 // See the comment in coordinator_yield() 6955 for (unsigned i = 0; i < CMSYieldSleepCount && 6956 ConcurrentMarkSweepThread::should_yield() && 6957 !CMSCollector::foregroundGCIsActive(); ++i) { 6958 os::sleep(Thread::current(), 1, false); 6959 ConcurrentMarkSweepThread::acknowledge_yield_request(); 6960 } 6961 6962 ConcurrentMarkSweepThread::synchronize(true); 6963 _freelistLock->lock_without_safepoint_check(); 6964 _bitMap->lock()->lock_without_safepoint_check(); 6965 _collector->startTimer(); 6966 } 6967 6968 6969 ////////////////////////////////////////////////////////////////// 6970 // SurvivorSpacePrecleanClosure 6971 ////////////////////////////////////////////////////////////////// 6972 // This (single-threaded) closure is used to preclean the oops in 6973 // the survivor spaces. 6974 size_t SurvivorSpacePrecleanClosure::do_object_careful(oop p) { 6975 6976 HeapWord* addr = (HeapWord*)p; 6977 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6978 assert(!_span.contains(addr), "we are scanning the survivor spaces"); 6979 assert(p->klass_or_null() != NULL, "object should be initializd"); 6980 // an initialized object; ignore mark word in verification below 6981 // since we are running concurrent with mutators 6982 assert(p->is_oop(true), "should be an oop"); 6983 // Note that we do not yield while we iterate over 6984 // the interior oops of p, pushing the relevant ones 6985 // on our marking stack. 6986 size_t size = p->oop_iterate(_scanning_closure); 6987 do_yield_check(); 6988 // Observe that below, we do not abandon the preclean 6989 // phase as soon as we should; rather we empty the 6990 // marking stack before returning. This is to satisfy 6991 // some existing assertions. In general, it may be a 6992 // good idea to abort immediately and complete the marking 6993 // from the grey objects at a later time. 6994 while (!_mark_stack->isEmpty()) { 6995 oop new_oop = _mark_stack->pop(); 6996 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 6997 assert(_bit_map->isMarked((HeapWord*)new_oop), 6998 "only grey objects on this stack"); 6999 // iterate over the oops in this oop, marking and pushing 7000 // the ones in CMS heap (i.e. in _span). 7001 new_oop->oop_iterate(_scanning_closure); 7002 // check if it's time to yield 7003 do_yield_check(); 7004 } 7005 unsigned int after_count = 7006 GenCollectedHeap::heap()->total_collections(); 7007 bool abort = (_before_count != after_count) || 7008 _collector->should_abort_preclean(); 7009 return abort ? 0 : size; 7010 } 7011 7012 void SurvivorSpacePrecleanClosure::do_yield_work() { 7013 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 7014 "CMS thread should hold CMS token"); 7015 assert_lock_strong(_bit_map->lock()); 7016 // Relinquish the bit map lock 7017 _bit_map->lock()->unlock(); 7018 ConcurrentMarkSweepThread::desynchronize(true); 7019 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7020 _collector->stopTimer(); 7021 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 7022 if (PrintCMSStatistics != 0) { 7023 _collector->incrementYields(); 7024 } 7025 _collector->icms_wait(); 7026 7027 // See the comment in coordinator_yield() 7028 for (unsigned i = 0; i < CMSYieldSleepCount && 7029 ConcurrentMarkSweepThread::should_yield() && 7030 !CMSCollector::foregroundGCIsActive(); ++i) { 7031 os::sleep(Thread::current(), 1, false); 7032 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7033 } 7034 7035 ConcurrentMarkSweepThread::synchronize(true); 7036 _bit_map->lock()->lock_without_safepoint_check(); 7037 _collector->startTimer(); 7038 } 7039 7040 // This closure is used to rescan the marked objects on the dirty cards 7041 // in the mod union table and the card table proper. In the parallel 7042 // case, although the bitMap is shared, we do a single read so the 7043 // isMarked() query is "safe". 7044 bool ScanMarkedObjectsAgainClosure::do_object_bm(oop p, MemRegion mr) { 7045 // Ignore mark word because we are running concurrent with mutators 7046 assert(p->is_oop_or_null(true), "expected an oop or null"); 7047 HeapWord* addr = (HeapWord*)p; 7048 assert(_span.contains(addr), "we are scanning the CMS generation"); 7049 bool is_obj_array = false; 7050 #ifdef ASSERT 7051 if (!_parallel) { 7052 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 7053 assert(_collector->overflow_list_is_empty(), 7054 "overflow list should be empty"); 7055 7056 } 7057 #endif // ASSERT 7058 if (_bit_map->isMarked(addr)) { 7059 // Obj arrays are precisely marked, non-arrays are not; 7060 // so we scan objArrays precisely and non-arrays in their 7061 // entirety. 7062 if (p->is_objArray()) { 7063 is_obj_array = true; 7064 if (_parallel) { 7065 p->oop_iterate(_par_scan_closure, mr); 7066 } else { 7067 p->oop_iterate(_scan_closure, mr); 7068 } 7069 } else { 7070 if (_parallel) { 7071 p->oop_iterate(_par_scan_closure); 7072 } else { 7073 p->oop_iterate(_scan_closure); 7074 } 7075 } 7076 } 7077 #ifdef ASSERT 7078 if (!_parallel) { 7079 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 7080 assert(_collector->overflow_list_is_empty(), 7081 "overflow list should be empty"); 7082 7083 } 7084 #endif // ASSERT 7085 return is_obj_array; 7086 } 7087 7088 MarkFromRootsClosure::MarkFromRootsClosure(CMSCollector* collector, 7089 MemRegion span, 7090 CMSBitMap* bitMap, CMSMarkStack* markStack, 7091 bool should_yield, bool verifying): 7092 _collector(collector), 7093 _span(span), 7094 _bitMap(bitMap), 7095 _mut(&collector->_modUnionTable), 7096 _markStack(markStack), 7097 _yield(should_yield), 7098 _skipBits(0) 7099 { 7100 assert(_markStack->isEmpty(), "stack should be empty"); 7101 _finger = _bitMap->startWord(); 7102 _threshold = _finger; 7103 assert(_collector->_restart_addr == NULL, "Sanity check"); 7104 assert(_span.contains(_finger), "Out of bounds _finger?"); 7105 DEBUG_ONLY(_verifying = verifying;) 7106 } 7107 7108 void MarkFromRootsClosure::reset(HeapWord* addr) { 7109 assert(_markStack->isEmpty(), "would cause duplicates on stack"); 7110 assert(_span.contains(addr), "Out of bounds _finger?"); 7111 _finger = addr; 7112 _threshold = (HeapWord*)round_to( 7113 (intptr_t)_finger, CardTableModRefBS::card_size); 7114 } 7115 7116 // Should revisit to see if this should be restructured for 7117 // greater efficiency. 7118 bool MarkFromRootsClosure::do_bit(size_t offset) { 7119 if (_skipBits > 0) { 7120 _skipBits--; 7121 return true; 7122 } 7123 // convert offset into a HeapWord* 7124 HeapWord* addr = _bitMap->startWord() + offset; 7125 assert(_bitMap->endWord() && addr < _bitMap->endWord(), 7126 "address out of range"); 7127 assert(_bitMap->isMarked(addr), "tautology"); 7128 if (_bitMap->isMarked(addr+1)) { 7129 // this is an allocated but not yet initialized object 7130 assert(_skipBits == 0, "tautology"); 7131 _skipBits = 2; // skip next two marked bits ("Printezis-marks") 7132 oop p = oop(addr); 7133 if (p->klass_or_null() == NULL) { 7134 DEBUG_ONLY(if (!_verifying) {) 7135 // We re-dirty the cards on which this object lies and increase 7136 // the _threshold so that we'll come back to scan this object 7137 // during the preclean or remark phase. (CMSCleanOnEnter) 7138 if (CMSCleanOnEnter) { 7139 size_t sz = _collector->block_size_using_printezis_bits(addr); 7140 HeapWord* end_card_addr = (HeapWord*)round_to( 7141 (intptr_t)(addr+sz), CardTableModRefBS::card_size); 7142 MemRegion redirty_range = MemRegion(addr, end_card_addr); 7143 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 7144 // Bump _threshold to end_card_addr; note that 7145 // _threshold cannot possibly exceed end_card_addr, anyhow. 7146 // This prevents future clearing of the card as the scan proceeds 7147 // to the right. 7148 assert(_threshold <= end_card_addr, 7149 "Because we are just scanning into this object"); 7150 if (_threshold < end_card_addr) { 7151 _threshold = end_card_addr; 7152 } 7153 if (p->klass_or_null() != NULL) { 7154 // Redirty the range of cards... 7155 _mut->mark_range(redirty_range); 7156 } // ...else the setting of klass will dirty the card anyway. 7157 } 7158 DEBUG_ONLY(}) 7159 return true; 7160 } 7161 } 7162 scanOopsInOop(addr); 7163 return true; 7164 } 7165 7166 // We take a break if we've been at this for a while, 7167 // so as to avoid monopolizing the locks involved. 7168 void MarkFromRootsClosure::do_yield_work() { 7169 // First give up the locks, then yield, then re-lock 7170 // We should probably use a constructor/destructor idiom to 7171 // do this unlock/lock or modify the MutexUnlocker class to 7172 // serve our purpose. XXX 7173 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 7174 "CMS thread should hold CMS token"); 7175 assert_lock_strong(_bitMap->lock()); 7176 _bitMap->lock()->unlock(); 7177 ConcurrentMarkSweepThread::desynchronize(true); 7178 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7179 _collector->stopTimer(); 7180 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 7181 if (PrintCMSStatistics != 0) { 7182 _collector->incrementYields(); 7183 } 7184 _collector->icms_wait(); 7185 7186 // See the comment in coordinator_yield() 7187 for (unsigned i = 0; i < CMSYieldSleepCount && 7188 ConcurrentMarkSweepThread::should_yield() && 7189 !CMSCollector::foregroundGCIsActive(); ++i) { 7190 os::sleep(Thread::current(), 1, false); 7191 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7192 } 7193 7194 ConcurrentMarkSweepThread::synchronize(true); 7195 _bitMap->lock()->lock_without_safepoint_check(); 7196 _collector->startTimer(); 7197 } 7198 7199 void MarkFromRootsClosure::scanOopsInOop(HeapWord* ptr) { 7200 assert(_bitMap->isMarked(ptr), "expected bit to be set"); 7201 assert(_markStack->isEmpty(), 7202 "should drain stack to limit stack usage"); 7203 // convert ptr to an oop preparatory to scanning 7204 oop obj = oop(ptr); 7205 // Ignore mark word in verification below, since we 7206 // may be running concurrent with mutators. 7207 assert(obj->is_oop(true), "should be an oop"); 7208 assert(_finger <= ptr, "_finger runneth ahead"); 7209 // advance the finger to right end of this object 7210 _finger = ptr + obj->size(); 7211 assert(_finger > ptr, "we just incremented it above"); 7212 // On large heaps, it may take us some time to get through 7213 // the marking phase (especially if running iCMS). During 7214 // this time it's possible that a lot of mutations have 7215 // accumulated in the card table and the mod union table -- 7216 // these mutation records are redundant until we have 7217 // actually traced into the corresponding card. 7218 // Here, we check whether advancing the finger would make 7219 // us cross into a new card, and if so clear corresponding 7220 // cards in the MUT (preclean them in the card-table in the 7221 // future). 7222 7223 DEBUG_ONLY(if (!_verifying) {) 7224 // The clean-on-enter optimization is disabled by default, 7225 // until we fix 6178663. 7226 if (CMSCleanOnEnter && (_finger > _threshold)) { 7227 // [_threshold, _finger) represents the interval 7228 // of cards to be cleared in MUT (or precleaned in card table). 7229 // The set of cards to be cleared is all those that overlap 7230 // with the interval [_threshold, _finger); note that 7231 // _threshold is always kept card-aligned but _finger isn't 7232 // always card-aligned. 7233 HeapWord* old_threshold = _threshold; 7234 assert(old_threshold == (HeapWord*)round_to( 7235 (intptr_t)old_threshold, CardTableModRefBS::card_size), 7236 "_threshold should always be card-aligned"); 7237 _threshold = (HeapWord*)round_to( 7238 (intptr_t)_finger, CardTableModRefBS::card_size); 7239 MemRegion mr(old_threshold, _threshold); 7240 assert(!mr.is_empty(), "Control point invariant"); 7241 assert(_span.contains(mr), "Should clear within span"); 7242 _mut->clear_range(mr); 7243 } 7244 DEBUG_ONLY(}) 7245 // Note: the finger doesn't advance while we drain 7246 // the stack below. 7247 PushOrMarkClosure pushOrMarkClosure(_collector, 7248 _span, _bitMap, _markStack, 7249 _finger, this); 7250 bool res = _markStack->push(obj); 7251 assert(res, "Empty non-zero size stack should have space for single push"); 7252 while (!_markStack->isEmpty()) { 7253 oop new_oop = _markStack->pop(); 7254 // Skip verifying header mark word below because we are 7255 // running concurrent with mutators. 7256 assert(new_oop->is_oop(true), "Oops! expected to pop an oop"); 7257 // now scan this oop's oops 7258 new_oop->oop_iterate(&pushOrMarkClosure); 7259 do_yield_check(); 7260 } 7261 assert(_markStack->isEmpty(), "tautology, emphasizing post-condition"); 7262 } 7263 7264 Par_MarkFromRootsClosure::Par_MarkFromRootsClosure(CMSConcMarkingTask* task, 7265 CMSCollector* collector, MemRegion span, 7266 CMSBitMap* bit_map, 7267 OopTaskQueue* work_queue, 7268 CMSMarkStack* overflow_stack, 7269 bool should_yield): 7270 _collector(collector), 7271 _whole_span(collector->_span), 7272 _span(span), 7273 _bit_map(bit_map), 7274 _mut(&collector->_modUnionTable), 7275 _work_queue(work_queue), 7276 _overflow_stack(overflow_stack), 7277 _yield(should_yield), 7278 _skip_bits(0), 7279 _task(task) 7280 { 7281 assert(_work_queue->size() == 0, "work_queue should be empty"); 7282 _finger = span.start(); 7283 _threshold = _finger; // XXX Defer clear-on-enter optimization for now 7284 assert(_span.contains(_finger), "Out of bounds _finger?"); 7285 } 7286 7287 // Should revisit to see if this should be restructured for 7288 // greater efficiency. 7289 bool Par_MarkFromRootsClosure::do_bit(size_t offset) { 7290 if (_skip_bits > 0) { 7291 _skip_bits--; 7292 return true; 7293 } 7294 // convert offset into a HeapWord* 7295 HeapWord* addr = _bit_map->startWord() + offset; 7296 assert(_bit_map->endWord() && addr < _bit_map->endWord(), 7297 "address out of range"); 7298 assert(_bit_map->isMarked(addr), "tautology"); 7299 if (_bit_map->isMarked(addr+1)) { 7300 // this is an allocated object that might not yet be initialized 7301 assert(_skip_bits == 0, "tautology"); 7302 _skip_bits = 2; // skip next two marked bits ("Printezis-marks") 7303 oop p = oop(addr); 7304 if (p->klass_or_null() == NULL) { 7305 // in the case of Clean-on-Enter optimization, redirty card 7306 // and avoid clearing card by increasing the threshold. 7307 return true; 7308 } 7309 } 7310 scan_oops_in_oop(addr); 7311 return true; 7312 } 7313 7314 void Par_MarkFromRootsClosure::scan_oops_in_oop(HeapWord* ptr) { 7315 assert(_bit_map->isMarked(ptr), "expected bit to be set"); 7316 // Should we assert that our work queue is empty or 7317 // below some drain limit? 7318 assert(_work_queue->size() == 0, 7319 "should drain stack to limit stack usage"); 7320 // convert ptr to an oop preparatory to scanning 7321 oop obj = oop(ptr); 7322 // Ignore mark word in verification below, since we 7323 // may be running concurrent with mutators. 7324 assert(obj->is_oop(true), "should be an oop"); 7325 assert(_finger <= ptr, "_finger runneth ahead"); 7326 // advance the finger to right end of this object 7327 _finger = ptr + obj->size(); 7328 assert(_finger > ptr, "we just incremented it above"); 7329 // On large heaps, it may take us some time to get through 7330 // the marking phase (especially if running iCMS). During 7331 // this time it's possible that a lot of mutations have 7332 // accumulated in the card table and the mod union table -- 7333 // these mutation records are redundant until we have 7334 // actually traced into the corresponding card. 7335 // Here, we check whether advancing the finger would make 7336 // us cross into a new card, and if so clear corresponding 7337 // cards in the MUT (preclean them in the card-table in the 7338 // future). 7339 7340 // The clean-on-enter optimization is disabled by default, 7341 // until we fix 6178663. 7342 if (CMSCleanOnEnter && (_finger > _threshold)) { 7343 // [_threshold, _finger) represents the interval 7344 // of cards to be cleared in MUT (or precleaned in card table). 7345 // The set of cards to be cleared is all those that overlap 7346 // with the interval [_threshold, _finger); note that 7347 // _threshold is always kept card-aligned but _finger isn't 7348 // always card-aligned. 7349 HeapWord* old_threshold = _threshold; 7350 assert(old_threshold == (HeapWord*)round_to( 7351 (intptr_t)old_threshold, CardTableModRefBS::card_size), 7352 "_threshold should always be card-aligned"); 7353 _threshold = (HeapWord*)round_to( 7354 (intptr_t)_finger, CardTableModRefBS::card_size); 7355 MemRegion mr(old_threshold, _threshold); 7356 assert(!mr.is_empty(), "Control point invariant"); 7357 assert(_span.contains(mr), "Should clear within span"); // _whole_span ?? 7358 _mut->clear_range(mr); 7359 } 7360 7361 // Note: the local finger doesn't advance while we drain 7362 // the stack below, but the global finger sure can and will. 7363 HeapWord** gfa = _task->global_finger_addr(); 7364 Par_PushOrMarkClosure pushOrMarkClosure(_collector, 7365 _span, _bit_map, 7366 _work_queue, 7367 _overflow_stack, 7368 _finger, 7369 gfa, this); 7370 bool res = _work_queue->push(obj); // overflow could occur here 7371 assert(res, "Will hold once we use workqueues"); 7372 while (true) { 7373 oop new_oop; 7374 if (!_work_queue->pop_local(new_oop)) { 7375 // We emptied our work_queue; check if there's stuff that can 7376 // be gotten from the overflow stack. 7377 if (CMSConcMarkingTask::get_work_from_overflow_stack( 7378 _overflow_stack, _work_queue)) { 7379 do_yield_check(); 7380 continue; 7381 } else { // done 7382 break; 7383 } 7384 } 7385 // Skip verifying header mark word below because we are 7386 // running concurrent with mutators. 7387 assert(new_oop->is_oop(true), "Oops! expected to pop an oop"); 7388 // now scan this oop's oops 7389 new_oop->oop_iterate(&pushOrMarkClosure); 7390 do_yield_check(); 7391 } 7392 assert(_work_queue->size() == 0, "tautology, emphasizing post-condition"); 7393 } 7394 7395 // Yield in response to a request from VM Thread or 7396 // from mutators. 7397 void Par_MarkFromRootsClosure::do_yield_work() { 7398 assert(_task != NULL, "sanity"); 7399 _task->yield(); 7400 } 7401 7402 // A variant of the above used for verifying CMS marking work. 7403 MarkFromRootsVerifyClosure::MarkFromRootsVerifyClosure(CMSCollector* collector, 7404 MemRegion span, 7405 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 7406 CMSMarkStack* mark_stack): 7407 _collector(collector), 7408 _span(span), 7409 _verification_bm(verification_bm), 7410 _cms_bm(cms_bm), 7411 _mark_stack(mark_stack), 7412 _pam_verify_closure(collector, span, verification_bm, cms_bm, 7413 mark_stack) 7414 { 7415 assert(_mark_stack->isEmpty(), "stack should be empty"); 7416 _finger = _verification_bm->startWord(); 7417 assert(_collector->_restart_addr == NULL, "Sanity check"); 7418 assert(_span.contains(_finger), "Out of bounds _finger?"); 7419 } 7420 7421 void MarkFromRootsVerifyClosure::reset(HeapWord* addr) { 7422 assert(_mark_stack->isEmpty(), "would cause duplicates on stack"); 7423 assert(_span.contains(addr), "Out of bounds _finger?"); 7424 _finger = addr; 7425 } 7426 7427 // Should revisit to see if this should be restructured for 7428 // greater efficiency. 7429 bool MarkFromRootsVerifyClosure::do_bit(size_t offset) { 7430 // convert offset into a HeapWord* 7431 HeapWord* addr = _verification_bm->startWord() + offset; 7432 assert(_verification_bm->endWord() && addr < _verification_bm->endWord(), 7433 "address out of range"); 7434 assert(_verification_bm->isMarked(addr), "tautology"); 7435 assert(_cms_bm->isMarked(addr), "tautology"); 7436 7437 assert(_mark_stack->isEmpty(), 7438 "should drain stack to limit stack usage"); 7439 // convert addr to an oop preparatory to scanning 7440 oop obj = oop(addr); 7441 assert(obj->is_oop(), "should be an oop"); 7442 assert(_finger <= addr, "_finger runneth ahead"); 7443 // advance the finger to right end of this object 7444 _finger = addr + obj->size(); 7445 assert(_finger > addr, "we just incremented it above"); 7446 // Note: the finger doesn't advance while we drain 7447 // the stack below. 7448 bool res = _mark_stack->push(obj); 7449 assert(res, "Empty non-zero size stack should have space for single push"); 7450 while (!_mark_stack->isEmpty()) { 7451 oop new_oop = _mark_stack->pop(); 7452 assert(new_oop->is_oop(), "Oops! expected to pop an oop"); 7453 // now scan this oop's oops 7454 new_oop->oop_iterate(&_pam_verify_closure); 7455 } 7456 assert(_mark_stack->isEmpty(), "tautology, emphasizing post-condition"); 7457 return true; 7458 } 7459 7460 PushAndMarkVerifyClosure::PushAndMarkVerifyClosure( 7461 CMSCollector* collector, MemRegion span, 7462 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 7463 CMSMarkStack* mark_stack): 7464 CMSOopClosure(collector->ref_processor()), 7465 _collector(collector), 7466 _span(span), 7467 _verification_bm(verification_bm), 7468 _cms_bm(cms_bm), 7469 _mark_stack(mark_stack) 7470 { } 7471 7472 void PushAndMarkVerifyClosure::do_oop(oop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 7473 void PushAndMarkVerifyClosure::do_oop(narrowOop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 7474 7475 // Upon stack overflow, we discard (part of) the stack, 7476 // remembering the least address amongst those discarded 7477 // in CMSCollector's _restart_address. 7478 void PushAndMarkVerifyClosure::handle_stack_overflow(HeapWord* lost) { 7479 // Remember the least grey address discarded 7480 HeapWord* ra = (HeapWord*)_mark_stack->least_value(lost); 7481 _collector->lower_restart_addr(ra); 7482 _mark_stack->reset(); // discard stack contents 7483 _mark_stack->expand(); // expand the stack if possible 7484 } 7485 7486 void PushAndMarkVerifyClosure::do_oop(oop obj) { 7487 assert(obj->is_oop_or_null(), "expected an oop or NULL"); 7488 HeapWord* addr = (HeapWord*)obj; 7489 if (_span.contains(addr) && !_verification_bm->isMarked(addr)) { 7490 // Oop lies in _span and isn't yet grey or black 7491 _verification_bm->mark(addr); // now grey 7492 if (!_cms_bm->isMarked(addr)) { 7493 oop(addr)->print(); 7494 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)", 7495 addr); 7496 fatal("... aborting"); 7497 } 7498 7499 if (!_mark_stack->push(obj)) { // stack overflow 7500 if (PrintCMSStatistics != 0) { 7501 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 7502 SIZE_FORMAT, _mark_stack->capacity()); 7503 } 7504 assert(_mark_stack->isFull(), "Else push should have succeeded"); 7505 handle_stack_overflow(addr); 7506 } 7507 // anything including and to the right of _finger 7508 // will be scanned as we iterate over the remainder of the 7509 // bit map 7510 } 7511 } 7512 7513 PushOrMarkClosure::PushOrMarkClosure(CMSCollector* collector, 7514 MemRegion span, 7515 CMSBitMap* bitMap, CMSMarkStack* markStack, 7516 HeapWord* finger, MarkFromRootsClosure* parent) : 7517 CMSOopClosure(collector->ref_processor()), 7518 _collector(collector), 7519 _span(span), 7520 _bitMap(bitMap), 7521 _markStack(markStack), 7522 _finger(finger), 7523 _parent(parent) 7524 { } 7525 7526 Par_PushOrMarkClosure::Par_PushOrMarkClosure(CMSCollector* collector, 7527 MemRegion span, 7528 CMSBitMap* bit_map, 7529 OopTaskQueue* work_queue, 7530 CMSMarkStack* overflow_stack, 7531 HeapWord* finger, 7532 HeapWord** global_finger_addr, 7533 Par_MarkFromRootsClosure* parent) : 7534 CMSOopClosure(collector->ref_processor()), 7535 _collector(collector), 7536 _whole_span(collector->_span), 7537 _span(span), 7538 _bit_map(bit_map), 7539 _work_queue(work_queue), 7540 _overflow_stack(overflow_stack), 7541 _finger(finger), 7542 _global_finger_addr(global_finger_addr), 7543 _parent(parent) 7544 { } 7545 7546 // Assumes thread-safe access by callers, who are 7547 // responsible for mutual exclusion. 7548 void CMSCollector::lower_restart_addr(HeapWord* low) { 7549 assert(_span.contains(low), "Out of bounds addr"); 7550 if (_restart_addr == NULL) { 7551 _restart_addr = low; 7552 } else { 7553 _restart_addr = MIN2(_restart_addr, low); 7554 } 7555 } 7556 7557 // Upon stack overflow, we discard (part of) the stack, 7558 // remembering the least address amongst those discarded 7559 // in CMSCollector's _restart_address. 7560 void PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 7561 // Remember the least grey address discarded 7562 HeapWord* ra = (HeapWord*)_markStack->least_value(lost); 7563 _collector->lower_restart_addr(ra); 7564 _markStack->reset(); // discard stack contents 7565 _markStack->expand(); // expand the stack if possible 7566 } 7567 7568 // Upon stack overflow, we discard (part of) the stack, 7569 // remembering the least address amongst those discarded 7570 // in CMSCollector's _restart_address. 7571 void Par_PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 7572 // We need to do this under a mutex to prevent other 7573 // workers from interfering with the work done below. 7574 MutexLockerEx ml(_overflow_stack->par_lock(), 7575 Mutex::_no_safepoint_check_flag); 7576 // Remember the least grey address discarded 7577 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 7578 _collector->lower_restart_addr(ra); 7579 _overflow_stack->reset(); // discard stack contents 7580 _overflow_stack->expand(); // expand the stack if possible 7581 } 7582 7583 void CMKlassClosure::do_klass(Klass* k) { 7584 assert(_oop_closure != NULL, "Not initialized?"); 7585 k->oops_do(_oop_closure); 7586 } 7587 7588 void PushOrMarkClosure::do_oop(oop obj) { 7589 // Ignore mark word because we are running concurrent with mutators. 7590 assert(obj->is_oop_or_null(true), "expected an oop or NULL"); 7591 HeapWord* addr = (HeapWord*)obj; 7592 if (_span.contains(addr) && !_bitMap->isMarked(addr)) { 7593 // Oop lies in _span and isn't yet grey or black 7594 _bitMap->mark(addr); // now grey 7595 if (addr < _finger) { 7596 // the bit map iteration has already either passed, or 7597 // sampled, this bit in the bit map; we'll need to 7598 // use the marking stack to scan this oop's oops. 7599 bool simulate_overflow = false; 7600 NOT_PRODUCT( 7601 if (CMSMarkStackOverflowALot && 7602 _collector->simulate_overflow()) { 7603 // simulate a stack overflow 7604 simulate_overflow = true; 7605 } 7606 ) 7607 if (simulate_overflow || !_markStack->push(obj)) { // stack overflow 7608 if (PrintCMSStatistics != 0) { 7609 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 7610 SIZE_FORMAT, _markStack->capacity()); 7611 } 7612 assert(simulate_overflow || _markStack->isFull(), "Else push should have succeeded"); 7613 handle_stack_overflow(addr); 7614 } 7615 } 7616 // anything including and to the right of _finger 7617 // will be scanned as we iterate over the remainder of the 7618 // bit map 7619 do_yield_check(); 7620 } 7621 } 7622 7623 void PushOrMarkClosure::do_oop(oop* p) { PushOrMarkClosure::do_oop_work(p); } 7624 void PushOrMarkClosure::do_oop(narrowOop* p) { PushOrMarkClosure::do_oop_work(p); } 7625 7626 void Par_PushOrMarkClosure::do_oop(oop obj) { 7627 // Ignore mark word because we are running concurrent with mutators. 7628 assert(obj->is_oop_or_null(true), "expected an oop or NULL"); 7629 HeapWord* addr = (HeapWord*)obj; 7630 if (_whole_span.contains(addr) && !_bit_map->isMarked(addr)) { 7631 // Oop lies in _span and isn't yet grey or black 7632 // We read the global_finger (volatile read) strictly after marking oop 7633 bool res = _bit_map->par_mark(addr); // now grey 7634 volatile HeapWord** gfa = (volatile HeapWord**)_global_finger_addr; 7635 // Should we push this marked oop on our stack? 7636 // -- if someone else marked it, nothing to do 7637 // -- if target oop is above global finger nothing to do 7638 // -- if target oop is in chunk and above local finger 7639 // then nothing to do 7640 // -- else push on work queue 7641 if ( !res // someone else marked it, they will deal with it 7642 || (addr >= *gfa) // will be scanned in a later task 7643 || (_span.contains(addr) && addr >= _finger)) { // later in this chunk 7644 return; 7645 } 7646 // the bit map iteration has already either passed, or 7647 // sampled, this bit in the bit map; we'll need to 7648 // use the marking stack to scan this oop's oops. 7649 bool simulate_overflow = false; 7650 NOT_PRODUCT( 7651 if (CMSMarkStackOverflowALot && 7652 _collector->simulate_overflow()) { 7653 // simulate a stack overflow 7654 simulate_overflow = true; 7655 } 7656 ) 7657 if (simulate_overflow || 7658 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) { 7659 // stack overflow 7660 if (PrintCMSStatistics != 0) { 7661 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 7662 SIZE_FORMAT, _overflow_stack->capacity()); 7663 } 7664 // We cannot assert that the overflow stack is full because 7665 // it may have been emptied since. 7666 assert(simulate_overflow || 7667 _work_queue->size() == _work_queue->max_elems(), 7668 "Else push should have succeeded"); 7669 handle_stack_overflow(addr); 7670 } 7671 do_yield_check(); 7672 } 7673 } 7674 7675 void Par_PushOrMarkClosure::do_oop(oop* p) { Par_PushOrMarkClosure::do_oop_work(p); } 7676 void Par_PushOrMarkClosure::do_oop(narrowOop* p) { Par_PushOrMarkClosure::do_oop_work(p); } 7677 7678 PushAndMarkClosure::PushAndMarkClosure(CMSCollector* collector, 7679 MemRegion span, 7680 ReferenceProcessor* rp, 7681 CMSBitMap* bit_map, 7682 CMSBitMap* mod_union_table, 7683 CMSMarkStack* mark_stack, 7684 bool concurrent_precleaning): 7685 CMSOopClosure(rp), 7686 _collector(collector), 7687 _span(span), 7688 _bit_map(bit_map), 7689 _mod_union_table(mod_union_table), 7690 _mark_stack(mark_stack), 7691 _concurrent_precleaning(concurrent_precleaning) 7692 { 7693 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 7694 } 7695 7696 // Grey object rescan during pre-cleaning and second checkpoint phases -- 7697 // the non-parallel version (the parallel version appears further below.) 7698 void PushAndMarkClosure::do_oop(oop obj) { 7699 // Ignore mark word verification. If during concurrent precleaning, 7700 // the object monitor may be locked. If during the checkpoint 7701 // phases, the object may already have been reached by a different 7702 // path and may be at the end of the global overflow list (so 7703 // the mark word may be NULL). 7704 assert(obj->is_oop_or_null(true /* ignore mark word */), 7705 "expected an oop or NULL"); 7706 HeapWord* addr = (HeapWord*)obj; 7707 // Check if oop points into the CMS generation 7708 // and is not marked 7709 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 7710 // a white object ... 7711 _bit_map->mark(addr); // ... now grey 7712 // push on the marking stack (grey set) 7713 bool simulate_overflow = false; 7714 NOT_PRODUCT( 7715 if (CMSMarkStackOverflowALot && 7716 _collector->simulate_overflow()) { 7717 // simulate a stack overflow 7718 simulate_overflow = true; 7719 } 7720 ) 7721 if (simulate_overflow || !_mark_stack->push(obj)) { 7722 if (_concurrent_precleaning) { 7723 // During precleaning we can just dirty the appropriate card(s) 7724 // in the mod union table, thus ensuring that the object remains 7725 // in the grey set and continue. In the case of object arrays 7726 // we need to dirty all of the cards that the object spans, 7727 // since the rescan of object arrays will be limited to the 7728 // dirty cards. 7729 // Note that no one can be intefering with us in this action 7730 // of dirtying the mod union table, so no locking or atomics 7731 // are required. 7732 if (obj->is_objArray()) { 7733 size_t sz = obj->size(); 7734 HeapWord* end_card_addr = (HeapWord*)round_to( 7735 (intptr_t)(addr+sz), CardTableModRefBS::card_size); 7736 MemRegion redirty_range = MemRegion(addr, end_card_addr); 7737 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 7738 _mod_union_table->mark_range(redirty_range); 7739 } else { 7740 _mod_union_table->mark(addr); 7741 } 7742 _collector->_ser_pmc_preclean_ovflw++; 7743 } else { 7744 // During the remark phase, we need to remember this oop 7745 // in the overflow list. 7746 _collector->push_on_overflow_list(obj); 7747 _collector->_ser_pmc_remark_ovflw++; 7748 } 7749 } 7750 } 7751 } 7752 7753 Par_PushAndMarkClosure::Par_PushAndMarkClosure(CMSCollector* collector, 7754 MemRegion span, 7755 ReferenceProcessor* rp, 7756 CMSBitMap* bit_map, 7757 OopTaskQueue* work_queue): 7758 CMSOopClosure(rp), 7759 _collector(collector), 7760 _span(span), 7761 _bit_map(bit_map), 7762 _work_queue(work_queue) 7763 { 7764 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 7765 } 7766 7767 void PushAndMarkClosure::do_oop(oop* p) { PushAndMarkClosure::do_oop_work(p); } 7768 void PushAndMarkClosure::do_oop(narrowOop* p) { PushAndMarkClosure::do_oop_work(p); } 7769 7770 // Grey object rescan during second checkpoint phase -- 7771 // the parallel version. 7772 void Par_PushAndMarkClosure::do_oop(oop obj) { 7773 // In the assert below, we ignore the mark word because 7774 // this oop may point to an already visited object that is 7775 // on the overflow stack (in which case the mark word has 7776 // been hijacked for chaining into the overflow stack -- 7777 // if this is the last object in the overflow stack then 7778 // its mark word will be NULL). Because this object may 7779 // have been subsequently popped off the global overflow 7780 // stack, and the mark word possibly restored to the prototypical 7781 // value, by the time we get to examined this failing assert in 7782 // the debugger, is_oop_or_null(false) may subsequently start 7783 // to hold. 7784 assert(obj->is_oop_or_null(true), 7785 "expected an oop or NULL"); 7786 HeapWord* addr = (HeapWord*)obj; 7787 // Check if oop points into the CMS generation 7788 // and is not marked 7789 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 7790 // a white object ... 7791 // If we manage to "claim" the object, by being the 7792 // first thread to mark it, then we push it on our 7793 // marking stack 7794 if (_bit_map->par_mark(addr)) { // ... now grey 7795 // push on work queue (grey set) 7796 bool simulate_overflow = false; 7797 NOT_PRODUCT( 7798 if (CMSMarkStackOverflowALot && 7799 _collector->par_simulate_overflow()) { 7800 // simulate a stack overflow 7801 simulate_overflow = true; 7802 } 7803 ) 7804 if (simulate_overflow || !_work_queue->push(obj)) { 7805 _collector->par_push_on_overflow_list(obj); 7806 _collector->_par_pmc_remark_ovflw++; // imprecise OK: no need to CAS 7807 } 7808 } // Else, some other thread got there first 7809 } 7810 } 7811 7812 void Par_PushAndMarkClosure::do_oop(oop* p) { Par_PushAndMarkClosure::do_oop_work(p); } 7813 void Par_PushAndMarkClosure::do_oop(narrowOop* p) { Par_PushAndMarkClosure::do_oop_work(p); } 7814 7815 void CMSPrecleanRefsYieldClosure::do_yield_work() { 7816 Mutex* bml = _collector->bitMapLock(); 7817 assert_lock_strong(bml); 7818 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 7819 "CMS thread should hold CMS token"); 7820 7821 bml->unlock(); 7822 ConcurrentMarkSweepThread::desynchronize(true); 7823 7824 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7825 7826 _collector->stopTimer(); 7827 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 7828 if (PrintCMSStatistics != 0) { 7829 _collector->incrementYields(); 7830 } 7831 _collector->icms_wait(); 7832 7833 // See the comment in coordinator_yield() 7834 for (unsigned i = 0; i < CMSYieldSleepCount && 7835 ConcurrentMarkSweepThread::should_yield() && 7836 !CMSCollector::foregroundGCIsActive(); ++i) { 7837 os::sleep(Thread::current(), 1, false); 7838 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7839 } 7840 7841 ConcurrentMarkSweepThread::synchronize(true); 7842 bml->lock(); 7843 7844 _collector->startTimer(); 7845 } 7846 7847 bool CMSPrecleanRefsYieldClosure::should_return() { 7848 if (ConcurrentMarkSweepThread::should_yield()) { 7849 do_yield_work(); 7850 } 7851 return _collector->foregroundGCIsActive(); 7852 } 7853 7854 void MarkFromDirtyCardsClosure::do_MemRegion(MemRegion mr) { 7855 assert(((size_t)mr.start())%CardTableModRefBS::card_size_in_words == 0, 7856 "mr should be aligned to start at a card boundary"); 7857 // We'd like to assert: 7858 // assert(mr.word_size()%CardTableModRefBS::card_size_in_words == 0, 7859 // "mr should be a range of cards"); 7860 // However, that would be too strong in one case -- the last 7861 // partition ends at _unallocated_block which, in general, can be 7862 // an arbitrary boundary, not necessarily card aligned. 7863 if (PrintCMSStatistics != 0) { 7864 _num_dirty_cards += 7865 mr.word_size()/CardTableModRefBS::card_size_in_words; 7866 } 7867 _space->object_iterate_mem(mr, &_scan_cl); 7868 } 7869 7870 SweepClosure::SweepClosure(CMSCollector* collector, 7871 ConcurrentMarkSweepGeneration* g, 7872 CMSBitMap* bitMap, bool should_yield) : 7873 _collector(collector), 7874 _g(g), 7875 _sp(g->cmsSpace()), 7876 _limit(_sp->sweep_limit()), 7877 _freelistLock(_sp->freelistLock()), 7878 _bitMap(bitMap), 7879 _yield(should_yield), 7880 _inFreeRange(false), // No free range at beginning of sweep 7881 _freeRangeInFreeLists(false), // No free range at beginning of sweep 7882 _lastFreeRangeCoalesced(false), 7883 _freeFinger(g->used_region().start()) 7884 { 7885 NOT_PRODUCT( 7886 _numObjectsFreed = 0; 7887 _numWordsFreed = 0; 7888 _numObjectsLive = 0; 7889 _numWordsLive = 0; 7890 _numObjectsAlreadyFree = 0; 7891 _numWordsAlreadyFree = 0; 7892 _last_fc = NULL; 7893 7894 _sp->initializeIndexedFreeListArrayReturnedBytes(); 7895 _sp->dictionary()->initialize_dict_returned_bytes(); 7896 ) 7897 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7898 "sweep _limit out of bounds"); 7899 if (CMSTraceSweeper) { 7900 gclog_or_tty->print_cr("\n====================\nStarting new sweep with limit " PTR_FORMAT, 7901 _limit); 7902 } 7903 } 7904 7905 void SweepClosure::print_on(outputStream* st) const { 7906 tty->print_cr("_sp = [" PTR_FORMAT "," PTR_FORMAT ")", 7907 _sp->bottom(), _sp->end()); 7908 tty->print_cr("_limit = " PTR_FORMAT, _limit); 7909 tty->print_cr("_freeFinger = " PTR_FORMAT, _freeFinger); 7910 NOT_PRODUCT(tty->print_cr("_last_fc = " PTR_FORMAT, _last_fc);) 7911 tty->print_cr("_inFreeRange = %d, _freeRangeInFreeLists = %d, _lastFreeRangeCoalesced = %d", 7912 _inFreeRange, _freeRangeInFreeLists, _lastFreeRangeCoalesced); 7913 } 7914 7915 #ifndef PRODUCT 7916 // Assertion checking only: no useful work in product mode -- 7917 // however, if any of the flags below become product flags, 7918 // you may need to review this code to see if it needs to be 7919 // enabled in product mode. 7920 SweepClosure::~SweepClosure() { 7921 assert_lock_strong(_freelistLock); 7922 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7923 "sweep _limit out of bounds"); 7924 if (inFreeRange()) { 7925 warning("inFreeRange() should have been reset; dumping state of SweepClosure"); 7926 print(); 7927 ShouldNotReachHere(); 7928 } 7929 if (Verbose && PrintGC) { 7930 gclog_or_tty->print("Collected "SIZE_FORMAT" objects, " SIZE_FORMAT " bytes", 7931 _numObjectsFreed, _numWordsFreed*sizeof(HeapWord)); 7932 gclog_or_tty->print_cr("\nLive "SIZE_FORMAT" objects, " 7933 SIZE_FORMAT" bytes " 7934 "Already free "SIZE_FORMAT" objects, "SIZE_FORMAT" bytes", 7935 _numObjectsLive, _numWordsLive*sizeof(HeapWord), 7936 _numObjectsAlreadyFree, _numWordsAlreadyFree*sizeof(HeapWord)); 7937 size_t totalBytes = (_numWordsFreed + _numWordsLive + _numWordsAlreadyFree) 7938 * sizeof(HeapWord); 7939 gclog_or_tty->print_cr("Total sweep: "SIZE_FORMAT" bytes", totalBytes); 7940 7941 if (PrintCMSStatistics && CMSVerifyReturnedBytes) { 7942 size_t indexListReturnedBytes = _sp->sumIndexedFreeListArrayReturnedBytes(); 7943 size_t dict_returned_bytes = _sp->dictionary()->sum_dict_returned_bytes(); 7944 size_t returned_bytes = indexListReturnedBytes + dict_returned_bytes; 7945 gclog_or_tty->print("Returned "SIZE_FORMAT" bytes", returned_bytes); 7946 gclog_or_tty->print(" Indexed List Returned "SIZE_FORMAT" bytes", 7947 indexListReturnedBytes); 7948 gclog_or_tty->print_cr(" Dictionary Returned "SIZE_FORMAT" bytes", 7949 dict_returned_bytes); 7950 } 7951 } 7952 if (CMSTraceSweeper) { 7953 gclog_or_tty->print_cr("end of sweep with _limit = " PTR_FORMAT "\n================", 7954 _limit); 7955 } 7956 } 7957 #endif // PRODUCT 7958 7959 void SweepClosure::initialize_free_range(HeapWord* freeFinger, 7960 bool freeRangeInFreeLists) { 7961 if (CMSTraceSweeper) { 7962 gclog_or_tty->print("---- Start free range at 0x%x with free block (%d)\n", 7963 freeFinger, freeRangeInFreeLists); 7964 } 7965 assert(!inFreeRange(), "Trampling existing free range"); 7966 set_inFreeRange(true); 7967 set_lastFreeRangeCoalesced(false); 7968 7969 set_freeFinger(freeFinger); 7970 set_freeRangeInFreeLists(freeRangeInFreeLists); 7971 if (CMSTestInFreeList) { 7972 if (freeRangeInFreeLists) { 7973 FreeChunk* fc = (FreeChunk*) freeFinger; 7974 assert(fc->is_free(), "A chunk on the free list should be free."); 7975 assert(fc->size() > 0, "Free range should have a size"); 7976 assert(_sp->verify_chunk_in_free_list(fc), "Chunk is not in free lists"); 7977 } 7978 } 7979 } 7980 7981 // Note that the sweeper runs concurrently with mutators. Thus, 7982 // it is possible for direct allocation in this generation to happen 7983 // in the middle of the sweep. Note that the sweeper also coalesces 7984 // contiguous free blocks. Thus, unless the sweeper and the allocator 7985 // synchronize appropriately freshly allocated blocks may get swept up. 7986 // This is accomplished by the sweeper locking the free lists while 7987 // it is sweeping. Thus blocks that are determined to be free are 7988 // indeed free. There is however one additional complication: 7989 // blocks that have been allocated since the final checkpoint and 7990 // mark, will not have been marked and so would be treated as 7991 // unreachable and swept up. To prevent this, the allocator marks 7992 // the bit map when allocating during the sweep phase. This leads, 7993 // however, to a further complication -- objects may have been allocated 7994 // but not yet initialized -- in the sense that the header isn't yet 7995 // installed. The sweeper can not then determine the size of the block 7996 // in order to skip over it. To deal with this case, we use a technique 7997 // (due to Printezis) to encode such uninitialized block sizes in the 7998 // bit map. Since the bit map uses a bit per every HeapWord, but the 7999 // CMS generation has a minimum object size of 3 HeapWords, it follows 8000 // that "normal marks" won't be adjacent in the bit map (there will 8001 // always be at least two 0 bits between successive 1 bits). We make use 8002 // of these "unused" bits to represent uninitialized blocks -- the bit 8003 // corresponding to the start of the uninitialized object and the next 8004 // bit are both set. Finally, a 1 bit marks the end of the object that 8005 // started with the two consecutive 1 bits to indicate its potentially 8006 // uninitialized state. 8007 8008 size_t SweepClosure::do_blk_careful(HeapWord* addr) { 8009 FreeChunk* fc = (FreeChunk*)addr; 8010 size_t res; 8011 8012 // Check if we are done sweeping. Below we check "addr >= _limit" rather 8013 // than "addr == _limit" because although _limit was a block boundary when 8014 // we started the sweep, it may no longer be one because heap expansion 8015 // may have caused us to coalesce the block ending at the address _limit 8016 // with a newly expanded chunk (this happens when _limit was set to the 8017 // previous _end of the space), so we may have stepped past _limit: 8018 // see the following Zeno-like trail of CRs 6977970, 7008136, 7042740. 8019 if (addr >= _limit) { // we have swept up to or past the limit: finish up 8020 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 8021 "sweep _limit out of bounds"); 8022 assert(addr < _sp->end(), "addr out of bounds"); 8023 // Flush any free range we might be holding as a single 8024 // coalesced chunk to the appropriate free list. 8025 if (inFreeRange()) { 8026 assert(freeFinger() >= _sp->bottom() && freeFinger() < _limit, 8027 err_msg("freeFinger() " PTR_FORMAT" is out-of-bounds", freeFinger())); 8028 flush_cur_free_chunk(freeFinger(), 8029 pointer_delta(addr, freeFinger())); 8030 if (CMSTraceSweeper) { 8031 gclog_or_tty->print("Sweep: last chunk: "); 8032 gclog_or_tty->print("put_free_blk 0x%x ("SIZE_FORMAT") " 8033 "[coalesced:"SIZE_FORMAT"]\n", 8034 freeFinger(), pointer_delta(addr, freeFinger()), 8035 lastFreeRangeCoalesced()); 8036 } 8037 } 8038 8039 // help the iterator loop finish 8040 return pointer_delta(_sp->end(), addr); 8041 } 8042 8043 assert(addr < _limit, "sweep invariant"); 8044 // check if we should yield 8045 do_yield_check(addr); 8046 if (fc->is_free()) { 8047 // Chunk that is already free 8048 res = fc->size(); 8049 do_already_free_chunk(fc); 8050 debug_only(_sp->verifyFreeLists()); 8051 // If we flush the chunk at hand in lookahead_and_flush() 8052 // and it's coalesced with a preceding chunk, then the 8053 // process of "mangling" the payload of the coalesced block 8054 // will cause erasure of the size information from the 8055 // (erstwhile) header of all the coalesced blocks but the 8056 // first, so the first disjunct in the assert will not hold 8057 // in that specific case (in which case the second disjunct 8058 // will hold). 8059 assert(res == fc->size() || ((HeapWord*)fc) + res >= _limit, 8060 "Otherwise the size info doesn't change at this step"); 8061 NOT_PRODUCT( 8062 _numObjectsAlreadyFree++; 8063 _numWordsAlreadyFree += res; 8064 ) 8065 NOT_PRODUCT(_last_fc = fc;) 8066 } else if (!_bitMap->isMarked(addr)) { 8067 // Chunk is fresh garbage 8068 res = do_garbage_chunk(fc); 8069 debug_only(_sp->verifyFreeLists()); 8070 NOT_PRODUCT( 8071 _numObjectsFreed++; 8072 _numWordsFreed += res; 8073 ) 8074 } else { 8075 // Chunk that is alive. 8076 res = do_live_chunk(fc); 8077 debug_only(_sp->verifyFreeLists()); 8078 NOT_PRODUCT( 8079 _numObjectsLive++; 8080 _numWordsLive += res; 8081 ) 8082 } 8083 return res; 8084 } 8085 8086 // For the smart allocation, record following 8087 // split deaths - a free chunk is removed from its free list because 8088 // it is being split into two or more chunks. 8089 // split birth - a free chunk is being added to its free list because 8090 // a larger free chunk has been split and resulted in this free chunk. 8091 // coal death - a free chunk is being removed from its free list because 8092 // it is being coalesced into a large free chunk. 8093 // coal birth - a free chunk is being added to its free list because 8094 // it was created when two or more free chunks where coalesced into 8095 // this free chunk. 8096 // 8097 // These statistics are used to determine the desired number of free 8098 // chunks of a given size. The desired number is chosen to be relative 8099 // to the end of a CMS sweep. The desired number at the end of a sweep 8100 // is the 8101 // count-at-end-of-previous-sweep (an amount that was enough) 8102 // - count-at-beginning-of-current-sweep (the excess) 8103 // + split-births (gains in this size during interval) 8104 // - split-deaths (demands on this size during interval) 8105 // where the interval is from the end of one sweep to the end of the 8106 // next. 8107 // 8108 // When sweeping the sweeper maintains an accumulated chunk which is 8109 // the chunk that is made up of chunks that have been coalesced. That 8110 // will be termed the left-hand chunk. A new chunk of garbage that 8111 // is being considered for coalescing will be referred to as the 8112 // right-hand chunk. 8113 // 8114 // When making a decision on whether to coalesce a right-hand chunk with 8115 // the current left-hand chunk, the current count vs. the desired count 8116 // of the left-hand chunk is considered. Also if the right-hand chunk 8117 // is near the large chunk at the end of the heap (see 8118 // ConcurrentMarkSweepGeneration::isNearLargestChunk()), then the 8119 // left-hand chunk is coalesced. 8120 // 8121 // When making a decision about whether to split a chunk, the desired count 8122 // vs. the current count of the candidate to be split is also considered. 8123 // If the candidate is underpopulated (currently fewer chunks than desired) 8124 // a chunk of an overpopulated (currently more chunks than desired) size may 8125 // be chosen. The "hint" associated with a free list, if non-null, points 8126 // to a free list which may be overpopulated. 8127 // 8128 8129 void SweepClosure::do_already_free_chunk(FreeChunk* fc) { 8130 const size_t size = fc->size(); 8131 // Chunks that cannot be coalesced are not in the 8132 // free lists. 8133 if (CMSTestInFreeList && !fc->cantCoalesce()) { 8134 assert(_sp->verify_chunk_in_free_list(fc), 8135 "free chunk should be in free lists"); 8136 } 8137 // a chunk that is already free, should not have been 8138 // marked in the bit map 8139 HeapWord* const addr = (HeapWord*) fc; 8140 assert(!_bitMap->isMarked(addr), "free chunk should be unmarked"); 8141 // Verify that the bit map has no bits marked between 8142 // addr and purported end of this block. 8143 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 8144 8145 // Some chunks cannot be coalesced under any circumstances. 8146 // See the definition of cantCoalesce(). 8147 if (!fc->cantCoalesce()) { 8148 // This chunk can potentially be coalesced. 8149 if (_sp->adaptive_freelists()) { 8150 // All the work is done in 8151 do_post_free_or_garbage_chunk(fc, size); 8152 } else { // Not adaptive free lists 8153 // this is a free chunk that can potentially be coalesced by the sweeper; 8154 if (!inFreeRange()) { 8155 // if the next chunk is a free block that can't be coalesced 8156 // it doesn't make sense to remove this chunk from the free lists 8157 FreeChunk* nextChunk = (FreeChunk*)(addr + size); 8158 assert((HeapWord*)nextChunk <= _sp->end(), "Chunk size out of bounds?"); 8159 if ((HeapWord*)nextChunk < _sp->end() && // There is another free chunk to the right ... 8160 nextChunk->is_free() && // ... which is free... 8161 nextChunk->cantCoalesce()) { // ... but can't be coalesced 8162 // nothing to do 8163 } else { 8164 // Potentially the start of a new free range: 8165 // Don't eagerly remove it from the free lists. 8166 // No need to remove it if it will just be put 8167 // back again. (Also from a pragmatic point of view 8168 // if it is a free block in a region that is beyond 8169 // any allocated blocks, an assertion will fail) 8170 // Remember the start of a free run. 8171 initialize_free_range(addr, true); 8172 // end - can coalesce with next chunk 8173 } 8174 } else { 8175 // the midst of a free range, we are coalescing 8176 print_free_block_coalesced(fc); 8177 if (CMSTraceSweeper) { 8178 gclog_or_tty->print(" -- pick up free block 0x%x (%d)\n", fc, size); 8179 } 8180 // remove it from the free lists 8181 _sp->removeFreeChunkFromFreeLists(fc); 8182 set_lastFreeRangeCoalesced(true); 8183 // If the chunk is being coalesced and the current free range is 8184 // in the free lists, remove the current free range so that it 8185 // will be returned to the free lists in its entirety - all 8186 // the coalesced pieces included. 8187 if (freeRangeInFreeLists()) { 8188 FreeChunk* ffc = (FreeChunk*) freeFinger(); 8189 assert(ffc->size() == pointer_delta(addr, freeFinger()), 8190 "Size of free range is inconsistent with chunk size."); 8191 if (CMSTestInFreeList) { 8192 assert(_sp->verify_chunk_in_free_list(ffc), 8193 "free range is not in free lists"); 8194 } 8195 _sp->removeFreeChunkFromFreeLists(ffc); 8196 set_freeRangeInFreeLists(false); 8197 } 8198 } 8199 } 8200 // Note that if the chunk is not coalescable (the else arm 8201 // below), we unconditionally flush, without needing to do 8202 // a "lookahead," as we do below. 8203 if (inFreeRange()) lookahead_and_flush(fc, size); 8204 } else { 8205 // Code path common to both original and adaptive free lists. 8206 8207 // cant coalesce with previous block; this should be treated 8208 // as the end of a free run if any 8209 if (inFreeRange()) { 8210 // we kicked some butt; time to pick up the garbage 8211 assert(freeFinger() < addr, "freeFinger points too high"); 8212 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 8213 } 8214 // else, nothing to do, just continue 8215 } 8216 } 8217 8218 size_t SweepClosure::do_garbage_chunk(FreeChunk* fc) { 8219 // This is a chunk of garbage. It is not in any free list. 8220 // Add it to a free list or let it possibly be coalesced into 8221 // a larger chunk. 8222 HeapWord* const addr = (HeapWord*) fc; 8223 const size_t size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 8224 8225 if (_sp->adaptive_freelists()) { 8226 // Verify that the bit map has no bits marked between 8227 // addr and purported end of just dead object. 8228 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 8229 8230 do_post_free_or_garbage_chunk(fc, size); 8231 } else { 8232 if (!inFreeRange()) { 8233 // start of a new free range 8234 assert(size > 0, "A free range should have a size"); 8235 initialize_free_range(addr, false); 8236 } else { 8237 // this will be swept up when we hit the end of the 8238 // free range 8239 if (CMSTraceSweeper) { 8240 gclog_or_tty->print(" -- pick up garbage 0x%x (%d) \n", fc, size); 8241 } 8242 // If the chunk is being coalesced and the current free range is 8243 // in the free lists, remove the current free range so that it 8244 // will be returned to the free lists in its entirety - all 8245 // the coalesced pieces included. 8246 if (freeRangeInFreeLists()) { 8247 FreeChunk* ffc = (FreeChunk*)freeFinger(); 8248 assert(ffc->size() == pointer_delta(addr, freeFinger()), 8249 "Size of free range is inconsistent with chunk size."); 8250 if (CMSTestInFreeList) { 8251 assert(_sp->verify_chunk_in_free_list(ffc), 8252 "free range is not in free lists"); 8253 } 8254 _sp->removeFreeChunkFromFreeLists(ffc); 8255 set_freeRangeInFreeLists(false); 8256 } 8257 set_lastFreeRangeCoalesced(true); 8258 } 8259 // this will be swept up when we hit the end of the free range 8260 8261 // Verify that the bit map has no bits marked between 8262 // addr and purported end of just dead object. 8263 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 8264 } 8265 assert(_limit >= addr + size, 8266 "A freshly garbage chunk can't possibly straddle over _limit"); 8267 if (inFreeRange()) lookahead_and_flush(fc, size); 8268 return size; 8269 } 8270 8271 size_t SweepClosure::do_live_chunk(FreeChunk* fc) { 8272 HeapWord* addr = (HeapWord*) fc; 8273 // The sweeper has just found a live object. Return any accumulated 8274 // left hand chunk to the free lists. 8275 if (inFreeRange()) { 8276 assert(freeFinger() < addr, "freeFinger points too high"); 8277 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 8278 } 8279 8280 // This object is live: we'd normally expect this to be 8281 // an oop, and like to assert the following: 8282 // assert(oop(addr)->is_oop(), "live block should be an oop"); 8283 // However, as we commented above, this may be an object whose 8284 // header hasn't yet been initialized. 8285 size_t size; 8286 assert(_bitMap->isMarked(addr), "Tautology for this control point"); 8287 if (_bitMap->isMarked(addr + 1)) { 8288 // Determine the size from the bit map, rather than trying to 8289 // compute it from the object header. 8290 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 8291 size = pointer_delta(nextOneAddr + 1, addr); 8292 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 8293 "alignment problem"); 8294 8295 #ifdef ASSERT 8296 if (oop(addr)->klass_or_null() != NULL) { 8297 // Ignore mark word because we are running concurrent with mutators 8298 assert(oop(addr)->is_oop(true), "live block should be an oop"); 8299 assert(size == 8300 CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()), 8301 "P-mark and computed size do not agree"); 8302 } 8303 #endif 8304 8305 } else { 8306 // This should be an initialized object that's alive. 8307 assert(oop(addr)->klass_or_null() != NULL, 8308 "Should be an initialized object"); 8309 // Ignore mark word because we are running concurrent with mutators 8310 assert(oop(addr)->is_oop(true), "live block should be an oop"); 8311 // Verify that the bit map has no bits marked between 8312 // addr and purported end of this block. 8313 size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 8314 assert(size >= 3, "Necessary for Printezis marks to work"); 8315 assert(!_bitMap->isMarked(addr+1), "Tautology for this control point"); 8316 DEBUG_ONLY(_bitMap->verifyNoOneBitsInRange(addr+2, addr+size);) 8317 } 8318 return size; 8319 } 8320 8321 void SweepClosure::do_post_free_or_garbage_chunk(FreeChunk* fc, 8322 size_t chunkSize) { 8323 // do_post_free_or_garbage_chunk() should only be called in the case 8324 // of the adaptive free list allocator. 8325 const bool fcInFreeLists = fc->is_free(); 8326 assert(_sp->adaptive_freelists(), "Should only be used in this case."); 8327 assert((HeapWord*)fc <= _limit, "sweep invariant"); 8328 if (CMSTestInFreeList && fcInFreeLists) { 8329 assert(_sp->verify_chunk_in_free_list(fc), "free chunk is not in free lists"); 8330 } 8331 8332 if (CMSTraceSweeper) { 8333 gclog_or_tty->print_cr(" -- pick up another chunk at 0x%x (%d)", fc, chunkSize); 8334 } 8335 8336 HeapWord* const fc_addr = (HeapWord*) fc; 8337 8338 bool coalesce; 8339 const size_t left = pointer_delta(fc_addr, freeFinger()); 8340 const size_t right = chunkSize; 8341 switch (FLSCoalescePolicy) { 8342 // numeric value forms a coalition aggressiveness metric 8343 case 0: { // never coalesce 8344 coalesce = false; 8345 break; 8346 } 8347 case 1: { // coalesce if left & right chunks on overpopulated lists 8348 coalesce = _sp->coalOverPopulated(left) && 8349 _sp->coalOverPopulated(right); 8350 break; 8351 } 8352 case 2: { // coalesce if left chunk on overpopulated list (default) 8353 coalesce = _sp->coalOverPopulated(left); 8354 break; 8355 } 8356 case 3: { // coalesce if left OR right chunk on overpopulated list 8357 coalesce = _sp->coalOverPopulated(left) || 8358 _sp->coalOverPopulated(right); 8359 break; 8360 } 8361 case 4: { // always coalesce 8362 coalesce = true; 8363 break; 8364 } 8365 default: 8366 ShouldNotReachHere(); 8367 } 8368 8369 // Should the current free range be coalesced? 8370 // If the chunk is in a free range and either we decided to coalesce above 8371 // or the chunk is near the large block at the end of the heap 8372 // (isNearLargestChunk() returns true), then coalesce this chunk. 8373 const bool doCoalesce = inFreeRange() 8374 && (coalesce || _g->isNearLargestChunk(fc_addr)); 8375 if (doCoalesce) { 8376 // Coalesce the current free range on the left with the new 8377 // chunk on the right. If either is on a free list, 8378 // it must be removed from the list and stashed in the closure. 8379 if (freeRangeInFreeLists()) { 8380 FreeChunk* const ffc = (FreeChunk*)freeFinger(); 8381 assert(ffc->size() == pointer_delta(fc_addr, freeFinger()), 8382 "Size of free range is inconsistent with chunk size."); 8383 if (CMSTestInFreeList) { 8384 assert(_sp->verify_chunk_in_free_list(ffc), 8385 "Chunk is not in free lists"); 8386 } 8387 _sp->coalDeath(ffc->size()); 8388 _sp->removeFreeChunkFromFreeLists(ffc); 8389 set_freeRangeInFreeLists(false); 8390 } 8391 if (fcInFreeLists) { 8392 _sp->coalDeath(chunkSize); 8393 assert(fc->size() == chunkSize, 8394 "The chunk has the wrong size or is not in the free lists"); 8395 _sp->removeFreeChunkFromFreeLists(fc); 8396 } 8397 set_lastFreeRangeCoalesced(true); 8398 print_free_block_coalesced(fc); 8399 } else { // not in a free range and/or should not coalesce 8400 // Return the current free range and start a new one. 8401 if (inFreeRange()) { 8402 // In a free range but cannot coalesce with the right hand chunk. 8403 // Put the current free range into the free lists. 8404 flush_cur_free_chunk(freeFinger(), 8405 pointer_delta(fc_addr, freeFinger())); 8406 } 8407 // Set up for new free range. Pass along whether the right hand 8408 // chunk is in the free lists. 8409 initialize_free_range((HeapWord*)fc, fcInFreeLists); 8410 } 8411 } 8412 8413 // Lookahead flush: 8414 // If we are tracking a free range, and this is the last chunk that 8415 // we'll look at because its end crosses past _limit, we'll preemptively 8416 // flush it along with any free range we may be holding on to. Note that 8417 // this can be the case only for an already free or freshly garbage 8418 // chunk. If this block is an object, it can never straddle 8419 // over _limit. The "straddling" occurs when _limit is set at 8420 // the previous end of the space when this cycle started, and 8421 // a subsequent heap expansion caused the previously co-terminal 8422 // free block to be coalesced with the newly expanded portion, 8423 // thus rendering _limit a non-block-boundary making it dangerous 8424 // for the sweeper to step over and examine. 8425 void SweepClosure::lookahead_and_flush(FreeChunk* fc, size_t chunk_size) { 8426 assert(inFreeRange(), "Should only be called if currently in a free range."); 8427 HeapWord* const eob = ((HeapWord*)fc) + chunk_size; 8428 assert(_sp->used_region().contains(eob - 1), 8429 err_msg("eob = " PTR_FORMAT " out of bounds wrt _sp = [" PTR_FORMAT "," PTR_FORMAT ")" 8430 " when examining fc = " PTR_FORMAT "(" SIZE_FORMAT ")", 8431 _limit, _sp->bottom(), _sp->end(), fc, chunk_size)); 8432 if (eob >= _limit) { 8433 assert(eob == _limit || fc->is_free(), "Only a free chunk should allow us to cross over the limit"); 8434 if (CMSTraceSweeper) { 8435 gclog_or_tty->print_cr("_limit " PTR_FORMAT " reached or crossed by block " 8436 "[" PTR_FORMAT "," PTR_FORMAT ") in space " 8437 "[" PTR_FORMAT "," PTR_FORMAT ")", 8438 _limit, fc, eob, _sp->bottom(), _sp->end()); 8439 } 8440 // Return the storage we are tracking back into the free lists. 8441 if (CMSTraceSweeper) { 8442 gclog_or_tty->print_cr("Flushing ... "); 8443 } 8444 assert(freeFinger() < eob, "Error"); 8445 flush_cur_free_chunk( freeFinger(), pointer_delta(eob, freeFinger())); 8446 } 8447 } 8448 8449 void SweepClosure::flush_cur_free_chunk(HeapWord* chunk, size_t size) { 8450 assert(inFreeRange(), "Should only be called if currently in a free range."); 8451 assert(size > 0, 8452 "A zero sized chunk cannot be added to the free lists."); 8453 if (!freeRangeInFreeLists()) { 8454 if (CMSTestInFreeList) { 8455 FreeChunk* fc = (FreeChunk*) chunk; 8456 fc->set_size(size); 8457 assert(!_sp->verify_chunk_in_free_list(fc), 8458 "chunk should not be in free lists yet"); 8459 } 8460 if (CMSTraceSweeper) { 8461 gclog_or_tty->print_cr(" -- add free block 0x%x (%d) to free lists", 8462 chunk, size); 8463 } 8464 // A new free range is going to be starting. The current 8465 // free range has not been added to the free lists yet or 8466 // was removed so add it back. 8467 // If the current free range was coalesced, then the death 8468 // of the free range was recorded. Record a birth now. 8469 if (lastFreeRangeCoalesced()) { 8470 _sp->coalBirth(size); 8471 } 8472 _sp->addChunkAndRepairOffsetTable(chunk, size, 8473 lastFreeRangeCoalesced()); 8474 } else if (CMSTraceSweeper) { 8475 gclog_or_tty->print_cr("Already in free list: nothing to flush"); 8476 } 8477 set_inFreeRange(false); 8478 set_freeRangeInFreeLists(false); 8479 } 8480 8481 // We take a break if we've been at this for a while, 8482 // so as to avoid monopolizing the locks involved. 8483 void SweepClosure::do_yield_work(HeapWord* addr) { 8484 // Return current free chunk being used for coalescing (if any) 8485 // to the appropriate freelist. After yielding, the next 8486 // free block encountered will start a coalescing range of 8487 // free blocks. If the next free block is adjacent to the 8488 // chunk just flushed, they will need to wait for the next 8489 // sweep to be coalesced. 8490 if (inFreeRange()) { 8491 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 8492 } 8493 8494 // First give up the locks, then yield, then re-lock. 8495 // We should probably use a constructor/destructor idiom to 8496 // do this unlock/lock or modify the MutexUnlocker class to 8497 // serve our purpose. XXX 8498 assert_lock_strong(_bitMap->lock()); 8499 assert_lock_strong(_freelistLock); 8500 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 8501 "CMS thread should hold CMS token"); 8502 _bitMap->lock()->unlock(); 8503 _freelistLock->unlock(); 8504 ConcurrentMarkSweepThread::desynchronize(true); 8505 ConcurrentMarkSweepThread::acknowledge_yield_request(); 8506 _collector->stopTimer(); 8507 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 8508 if (PrintCMSStatistics != 0) { 8509 _collector->incrementYields(); 8510 } 8511 _collector->icms_wait(); 8512 8513 // See the comment in coordinator_yield() 8514 for (unsigned i = 0; i < CMSYieldSleepCount && 8515 ConcurrentMarkSweepThread::should_yield() && 8516 !CMSCollector::foregroundGCIsActive(); ++i) { 8517 os::sleep(Thread::current(), 1, false); 8518 ConcurrentMarkSweepThread::acknowledge_yield_request(); 8519 } 8520 8521 ConcurrentMarkSweepThread::synchronize(true); 8522 _freelistLock->lock(); 8523 _bitMap->lock()->lock_without_safepoint_check(); 8524 _collector->startTimer(); 8525 } 8526 8527 #ifndef PRODUCT 8528 // This is actually very useful in a product build if it can 8529 // be called from the debugger. Compile it into the product 8530 // as needed. 8531 bool debug_verify_chunk_in_free_list(FreeChunk* fc) { 8532 return debug_cms_space->verify_chunk_in_free_list(fc); 8533 } 8534 #endif 8535 8536 void SweepClosure::print_free_block_coalesced(FreeChunk* fc) const { 8537 if (CMSTraceSweeper) { 8538 gclog_or_tty->print_cr("Sweep:coal_free_blk " PTR_FORMAT " (" SIZE_FORMAT ")", 8539 fc, fc->size()); 8540 } 8541 } 8542 8543 // CMSIsAliveClosure 8544 bool CMSIsAliveClosure::do_object_b(oop obj) { 8545 HeapWord* addr = (HeapWord*)obj; 8546 return addr != NULL && 8547 (!_span.contains(addr) || _bit_map->isMarked(addr)); 8548 } 8549 8550 8551 CMSKeepAliveClosure::CMSKeepAliveClosure( CMSCollector* collector, 8552 MemRegion span, 8553 CMSBitMap* bit_map, CMSMarkStack* mark_stack, 8554 bool cpc): 8555 _collector(collector), 8556 _span(span), 8557 _bit_map(bit_map), 8558 _mark_stack(mark_stack), 8559 _concurrent_precleaning(cpc) { 8560 assert(!_span.is_empty(), "Empty span could spell trouble"); 8561 } 8562 8563 8564 // CMSKeepAliveClosure: the serial version 8565 void CMSKeepAliveClosure::do_oop(oop obj) { 8566 HeapWord* addr = (HeapWord*)obj; 8567 if (_span.contains(addr) && 8568 !_bit_map->isMarked(addr)) { 8569 _bit_map->mark(addr); 8570 bool simulate_overflow = false; 8571 NOT_PRODUCT( 8572 if (CMSMarkStackOverflowALot && 8573 _collector->simulate_overflow()) { 8574 // simulate a stack overflow 8575 simulate_overflow = true; 8576 } 8577 ) 8578 if (simulate_overflow || !_mark_stack->push(obj)) { 8579 if (_concurrent_precleaning) { 8580 // We dirty the overflown object and let the remark 8581 // phase deal with it. 8582 assert(_collector->overflow_list_is_empty(), "Error"); 8583 // In the case of object arrays, we need to dirty all of 8584 // the cards that the object spans. No locking or atomics 8585 // are needed since no one else can be mutating the mod union 8586 // table. 8587 if (obj->is_objArray()) { 8588 size_t sz = obj->size(); 8589 HeapWord* end_card_addr = 8590 (HeapWord*)round_to((intptr_t)(addr+sz), CardTableModRefBS::card_size); 8591 MemRegion redirty_range = MemRegion(addr, end_card_addr); 8592 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 8593 _collector->_modUnionTable.mark_range(redirty_range); 8594 } else { 8595 _collector->_modUnionTable.mark(addr); 8596 } 8597 _collector->_ser_kac_preclean_ovflw++; 8598 } else { 8599 _collector->push_on_overflow_list(obj); 8600 _collector->_ser_kac_ovflw++; 8601 } 8602 } 8603 } 8604 } 8605 8606 void CMSKeepAliveClosure::do_oop(oop* p) { CMSKeepAliveClosure::do_oop_work(p); } 8607 void CMSKeepAliveClosure::do_oop(narrowOop* p) { CMSKeepAliveClosure::do_oop_work(p); } 8608 8609 // CMSParKeepAliveClosure: a parallel version of the above. 8610 // The work queues are private to each closure (thread), 8611 // but (may be) available for stealing by other threads. 8612 void CMSParKeepAliveClosure::do_oop(oop obj) { 8613 HeapWord* addr = (HeapWord*)obj; 8614 if (_span.contains(addr) && 8615 !_bit_map->isMarked(addr)) { 8616 // In general, during recursive tracing, several threads 8617 // may be concurrently getting here; the first one to 8618 // "tag" it, claims it. 8619 if (_bit_map->par_mark(addr)) { 8620 bool res = _work_queue->push(obj); 8621 assert(res, "Low water mark should be much less than capacity"); 8622 // Do a recursive trim in the hope that this will keep 8623 // stack usage lower, but leave some oops for potential stealers 8624 trim_queue(_low_water_mark); 8625 } // Else, another thread got there first 8626 } 8627 } 8628 8629 void CMSParKeepAliveClosure::do_oop(oop* p) { CMSParKeepAliveClosure::do_oop_work(p); } 8630 void CMSParKeepAliveClosure::do_oop(narrowOop* p) { CMSParKeepAliveClosure::do_oop_work(p); } 8631 8632 void CMSParKeepAliveClosure::trim_queue(uint max) { 8633 while (_work_queue->size() > max) { 8634 oop new_oop; 8635 if (_work_queue->pop_local(new_oop)) { 8636 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 8637 assert(_bit_map->isMarked((HeapWord*)new_oop), 8638 "no white objects on this stack!"); 8639 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 8640 // iterate over the oops in this oop, marking and pushing 8641 // the ones in CMS heap (i.e. in _span). 8642 new_oop->oop_iterate(&_mark_and_push); 8643 } 8644 } 8645 } 8646 8647 CMSInnerParMarkAndPushClosure::CMSInnerParMarkAndPushClosure( 8648 CMSCollector* collector, 8649 MemRegion span, CMSBitMap* bit_map, 8650 OopTaskQueue* work_queue): 8651 _collector(collector), 8652 _span(span), 8653 _bit_map(bit_map), 8654 _work_queue(work_queue) { } 8655 8656 void CMSInnerParMarkAndPushClosure::do_oop(oop obj) { 8657 HeapWord* addr = (HeapWord*)obj; 8658 if (_span.contains(addr) && 8659 !_bit_map->isMarked(addr)) { 8660 if (_bit_map->par_mark(addr)) { 8661 bool simulate_overflow = false; 8662 NOT_PRODUCT( 8663 if (CMSMarkStackOverflowALot && 8664 _collector->par_simulate_overflow()) { 8665 // simulate a stack overflow 8666 simulate_overflow = true; 8667 } 8668 ) 8669 if (simulate_overflow || !_work_queue->push(obj)) { 8670 _collector->par_push_on_overflow_list(obj); 8671 _collector->_par_kac_ovflw++; 8672 } 8673 } // Else another thread got there already 8674 } 8675 } 8676 8677 void CMSInnerParMarkAndPushClosure::do_oop(oop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); } 8678 void CMSInnerParMarkAndPushClosure::do_oop(narrowOop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); } 8679 8680 ////////////////////////////////////////////////////////////////// 8681 // CMSExpansionCause ///////////////////////////// 8682 ////////////////////////////////////////////////////////////////// 8683 const char* CMSExpansionCause::to_string(CMSExpansionCause::Cause cause) { 8684 switch (cause) { 8685 case _no_expansion: 8686 return "No expansion"; 8687 case _satisfy_free_ratio: 8688 return "Free ratio"; 8689 case _satisfy_promotion: 8690 return "Satisfy promotion"; 8691 case _satisfy_allocation: 8692 return "allocation"; 8693 case _allocate_par_lab: 8694 return "Par LAB"; 8695 case _allocate_par_spooling_space: 8696 return "Par Spooling Space"; 8697 case _adaptive_size_policy: 8698 return "Ergonomics"; 8699 default: 8700 return "unknown"; 8701 } 8702 } 8703 8704 void CMSDrainMarkingStackClosure::do_void() { 8705 // the max number to take from overflow list at a time 8706 const size_t num = _mark_stack->capacity()/4; 8707 assert(!_concurrent_precleaning || _collector->overflow_list_is_empty(), 8708 "Overflow list should be NULL during concurrent phases"); 8709 while (!_mark_stack->isEmpty() || 8710 // if stack is empty, check the overflow list 8711 _collector->take_from_overflow_list(num, _mark_stack)) { 8712 oop obj = _mark_stack->pop(); 8713 HeapWord* addr = (HeapWord*)obj; 8714 assert(_span.contains(addr), "Should be within span"); 8715 assert(_bit_map->isMarked(addr), "Should be marked"); 8716 assert(obj->is_oop(), "Should be an oop"); 8717 obj->oop_iterate(_keep_alive); 8718 } 8719 } 8720 8721 void CMSParDrainMarkingStackClosure::do_void() { 8722 // drain queue 8723 trim_queue(0); 8724 } 8725 8726 // Trim our work_queue so its length is below max at return 8727 void CMSParDrainMarkingStackClosure::trim_queue(uint max) { 8728 while (_work_queue->size() > max) { 8729 oop new_oop; 8730 if (_work_queue->pop_local(new_oop)) { 8731 assert(new_oop->is_oop(), "Expected an oop"); 8732 assert(_bit_map->isMarked((HeapWord*)new_oop), 8733 "no white objects on this stack!"); 8734 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 8735 // iterate over the oops in this oop, marking and pushing 8736 // the ones in CMS heap (i.e. in _span). 8737 new_oop->oop_iterate(&_mark_and_push); 8738 } 8739 } 8740 } 8741 8742 //////////////////////////////////////////////////////////////////// 8743 // Support for Marking Stack Overflow list handling and related code 8744 //////////////////////////////////////////////////////////////////// 8745 // Much of the following code is similar in shape and spirit to the 8746 // code used in ParNewGC. We should try and share that code 8747 // as much as possible in the future. 8748 8749 #ifndef PRODUCT 8750 // Debugging support for CMSStackOverflowALot 8751 8752 // It's OK to call this multi-threaded; the worst thing 8753 // that can happen is that we'll get a bunch of closely 8754 // spaced simulated oveflows, but that's OK, in fact 8755 // probably good as it would exercise the overflow code 8756 // under contention. 8757 bool CMSCollector::simulate_overflow() { 8758 if (_overflow_counter-- <= 0) { // just being defensive 8759 _overflow_counter = CMSMarkStackOverflowInterval; 8760 return true; 8761 } else { 8762 return false; 8763 } 8764 } 8765 8766 bool CMSCollector::par_simulate_overflow() { 8767 return simulate_overflow(); 8768 } 8769 #endif 8770 8771 // Single-threaded 8772 bool CMSCollector::take_from_overflow_list(size_t num, CMSMarkStack* stack) { 8773 assert(stack->isEmpty(), "Expected precondition"); 8774 assert(stack->capacity() > num, "Shouldn't bite more than can chew"); 8775 size_t i = num; 8776 oop cur = _overflow_list; 8777 const markOop proto = markOopDesc::prototype(); 8778 NOT_PRODUCT(ssize_t n = 0;) 8779 for (oop next; i > 0 && cur != NULL; cur = next, i--) { 8780 next = oop(cur->mark()); 8781 cur->set_mark(proto); // until proven otherwise 8782 assert(cur->is_oop(), "Should be an oop"); 8783 bool res = stack->push(cur); 8784 assert(res, "Bit off more than can chew?"); 8785 NOT_PRODUCT(n++;) 8786 } 8787 _overflow_list = cur; 8788 #ifndef PRODUCT 8789 assert(_num_par_pushes >= n, "Too many pops?"); 8790 _num_par_pushes -=n; 8791 #endif 8792 return !stack->isEmpty(); 8793 } 8794 8795 #define BUSY (oop(0x1aff1aff)) 8796 // (MT-safe) Get a prefix of at most "num" from the list. 8797 // The overflow list is chained through the mark word of 8798 // each object in the list. We fetch the entire list, 8799 // break off a prefix of the right size and return the 8800 // remainder. If other threads try to take objects from 8801 // the overflow list at that time, they will wait for 8802 // some time to see if data becomes available. If (and 8803 // only if) another thread places one or more object(s) 8804 // on the global list before we have returned the suffix 8805 // to the global list, we will walk down our local list 8806 // to find its end and append the global list to 8807 // our suffix before returning it. This suffix walk can 8808 // prove to be expensive (quadratic in the amount of traffic) 8809 // when there are many objects in the overflow list and 8810 // there is much producer-consumer contention on the list. 8811 // *NOTE*: The overflow list manipulation code here and 8812 // in ParNewGeneration:: are very similar in shape, 8813 // except that in the ParNew case we use the old (from/eden) 8814 // copy of the object to thread the list via its klass word. 8815 // Because of the common code, if you make any changes in 8816 // the code below, please check the ParNew version to see if 8817 // similar changes might be needed. 8818 // CR 6797058 has been filed to consolidate the common code. 8819 bool CMSCollector::par_take_from_overflow_list(size_t num, 8820 OopTaskQueue* work_q, 8821 int no_of_gc_threads) { 8822 assert(work_q->size() == 0, "First empty local work queue"); 8823 assert(num < work_q->max_elems(), "Can't bite more than we can chew"); 8824 if (_overflow_list == NULL) { 8825 return false; 8826 } 8827 // Grab the entire list; we'll put back a suffix 8828 oop prefix = (oop)Atomic::xchg_ptr(BUSY, &_overflow_list); 8829 Thread* tid = Thread::current(); 8830 // Before "no_of_gc_threads" was introduced CMSOverflowSpinCount was 8831 // set to ParallelGCThreads. 8832 size_t CMSOverflowSpinCount = (size_t) no_of_gc_threads; // was ParallelGCThreads; 8833 size_t sleep_time_millis = MAX2((size_t)1, num/100); 8834 // If the list is busy, we spin for a short while, 8835 // sleeping between attempts to get the list. 8836 for (size_t spin = 0; prefix == BUSY && spin < CMSOverflowSpinCount; spin++) { 8837 os::sleep(tid, sleep_time_millis, false); 8838 if (_overflow_list == NULL) { 8839 // Nothing left to take 8840 return false; 8841 } else if (_overflow_list != BUSY) { 8842 // Try and grab the prefix 8843 prefix = (oop)Atomic::xchg_ptr(BUSY, &_overflow_list); 8844 } 8845 } 8846 // If the list was found to be empty, or we spun long 8847 // enough, we give up and return empty-handed. If we leave 8848 // the list in the BUSY state below, it must be the case that 8849 // some other thread holds the overflow list and will set it 8850 // to a non-BUSY state in the future. 8851 if (prefix == NULL || prefix == BUSY) { 8852 // Nothing to take or waited long enough 8853 if (prefix == NULL) { 8854 // Write back the NULL in case we overwrote it with BUSY above 8855 // and it is still the same value. 8856 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY); 8857 } 8858 return false; 8859 } 8860 assert(prefix != NULL && prefix != BUSY, "Error"); 8861 size_t i = num; 8862 oop cur = prefix; 8863 // Walk down the first "num" objects, unless we reach the end. 8864 for (; i > 1 && cur->mark() != NULL; cur = oop(cur->mark()), i--); 8865 if (cur->mark() == NULL) { 8866 // We have "num" or fewer elements in the list, so there 8867 // is nothing to return to the global list. 8868 // Write back the NULL in lieu of the BUSY we wrote 8869 // above, if it is still the same value. 8870 if (_overflow_list == BUSY) { 8871 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY); 8872 } 8873 } else { 8874 // Chop off the suffix and rerturn it to the global list. 8875 assert(cur->mark() != BUSY, "Error"); 8876 oop suffix_head = cur->mark(); // suffix will be put back on global list 8877 cur->set_mark(NULL); // break off suffix 8878 // It's possible that the list is still in the empty(busy) state 8879 // we left it in a short while ago; in that case we may be 8880 // able to place back the suffix without incurring the cost 8881 // of a walk down the list. 8882 oop observed_overflow_list = _overflow_list; 8883 oop cur_overflow_list = observed_overflow_list; 8884 bool attached = false; 8885 while (observed_overflow_list == BUSY || observed_overflow_list == NULL) { 8886 observed_overflow_list = 8887 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list); 8888 if (cur_overflow_list == observed_overflow_list) { 8889 attached = true; 8890 break; 8891 } else cur_overflow_list = observed_overflow_list; 8892 } 8893 if (!attached) { 8894 // Too bad, someone else sneaked in (at least) an element; we'll need 8895 // to do a splice. Find tail of suffix so we can prepend suffix to global 8896 // list. 8897 for (cur = suffix_head; cur->mark() != NULL; cur = (oop)(cur->mark())); 8898 oop suffix_tail = cur; 8899 assert(suffix_tail != NULL && suffix_tail->mark() == NULL, 8900 "Tautology"); 8901 observed_overflow_list = _overflow_list; 8902 do { 8903 cur_overflow_list = observed_overflow_list; 8904 if (cur_overflow_list != BUSY) { 8905 // Do the splice ... 8906 suffix_tail->set_mark(markOop(cur_overflow_list)); 8907 } else { // cur_overflow_list == BUSY 8908 suffix_tail->set_mark(NULL); 8909 } 8910 // ... and try to place spliced list back on overflow_list ... 8911 observed_overflow_list = 8912 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list); 8913 } while (cur_overflow_list != observed_overflow_list); 8914 // ... until we have succeeded in doing so. 8915 } 8916 } 8917 8918 // Push the prefix elements on work_q 8919 assert(prefix != NULL, "control point invariant"); 8920 const markOop proto = markOopDesc::prototype(); 8921 oop next; 8922 NOT_PRODUCT(ssize_t n = 0;) 8923 for (cur = prefix; cur != NULL; cur = next) { 8924 next = oop(cur->mark()); 8925 cur->set_mark(proto); // until proven otherwise 8926 assert(cur->is_oop(), "Should be an oop"); 8927 bool res = work_q->push(cur); 8928 assert(res, "Bit off more than we can chew?"); 8929 NOT_PRODUCT(n++;) 8930 } 8931 #ifndef PRODUCT 8932 assert(_num_par_pushes >= n, "Too many pops?"); 8933 Atomic::add_ptr(-(intptr_t)n, &_num_par_pushes); 8934 #endif 8935 return true; 8936 } 8937 8938 // Single-threaded 8939 void CMSCollector::push_on_overflow_list(oop p) { 8940 NOT_PRODUCT(_num_par_pushes++;) 8941 assert(p->is_oop(), "Not an oop"); 8942 preserve_mark_if_necessary(p); 8943 p->set_mark((markOop)_overflow_list); 8944 _overflow_list = p; 8945 } 8946 8947 // Multi-threaded; use CAS to prepend to overflow list 8948 void CMSCollector::par_push_on_overflow_list(oop p) { 8949 NOT_PRODUCT(Atomic::inc_ptr(&_num_par_pushes);) 8950 assert(p->is_oop(), "Not an oop"); 8951 par_preserve_mark_if_necessary(p); 8952 oop observed_overflow_list = _overflow_list; 8953 oop cur_overflow_list; 8954 do { 8955 cur_overflow_list = observed_overflow_list; 8956 if (cur_overflow_list != BUSY) { 8957 p->set_mark(markOop(cur_overflow_list)); 8958 } else { 8959 p->set_mark(NULL); 8960 } 8961 observed_overflow_list = 8962 (oop) Atomic::cmpxchg_ptr(p, &_overflow_list, cur_overflow_list); 8963 } while (cur_overflow_list != observed_overflow_list); 8964 } 8965 #undef BUSY 8966 8967 // Single threaded 8968 // General Note on GrowableArray: pushes may silently fail 8969 // because we are (temporarily) out of C-heap for expanding 8970 // the stack. The problem is quite ubiquitous and affects 8971 // a lot of code in the JVM. The prudent thing for GrowableArray 8972 // to do (for now) is to exit with an error. However, that may 8973 // be too draconian in some cases because the caller may be 8974 // able to recover without much harm. For such cases, we 8975 // should probably introduce a "soft_push" method which returns 8976 // an indication of success or failure with the assumption that 8977 // the caller may be able to recover from a failure; code in 8978 // the VM can then be changed, incrementally, to deal with such 8979 // failures where possible, thus, incrementally hardening the VM 8980 // in such low resource situations. 8981 void CMSCollector::preserve_mark_work(oop p, markOop m) { 8982 _preserved_oop_stack.push(p); 8983 _preserved_mark_stack.push(m); 8984 assert(m == p->mark(), "Mark word changed"); 8985 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), 8986 "bijection"); 8987 } 8988 8989 // Single threaded 8990 void CMSCollector::preserve_mark_if_necessary(oop p) { 8991 markOop m = p->mark(); 8992 if (m->must_be_preserved(p)) { 8993 preserve_mark_work(p, m); 8994 } 8995 } 8996 8997 void CMSCollector::par_preserve_mark_if_necessary(oop p) { 8998 markOop m = p->mark(); 8999 if (m->must_be_preserved(p)) { 9000 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag); 9001 // Even though we read the mark word without holding 9002 // the lock, we are assured that it will not change 9003 // because we "own" this oop, so no other thread can 9004 // be trying to push it on the overflow list; see 9005 // the assertion in preserve_mark_work() that checks 9006 // that m == p->mark(). 9007 preserve_mark_work(p, m); 9008 } 9009 } 9010 9011 // We should be able to do this multi-threaded, 9012 // a chunk of stack being a task (this is 9013 // correct because each oop only ever appears 9014 // once in the overflow list. However, it's 9015 // not very easy to completely overlap this with 9016 // other operations, so will generally not be done 9017 // until all work's been completed. Because we 9018 // expect the preserved oop stack (set) to be small, 9019 // it's probably fine to do this single-threaded. 9020 // We can explore cleverer concurrent/overlapped/parallel 9021 // processing of preserved marks if we feel the 9022 // need for this in the future. Stack overflow should 9023 // be so rare in practice and, when it happens, its 9024 // effect on performance so great that this will 9025 // likely just be in the noise anyway. 9026 void CMSCollector::restore_preserved_marks_if_any() { 9027 assert(SafepointSynchronize::is_at_safepoint(), 9028 "world should be stopped"); 9029 assert(Thread::current()->is_ConcurrentGC_thread() || 9030 Thread::current()->is_VM_thread(), 9031 "should be single-threaded"); 9032 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), 9033 "bijection"); 9034 9035 while (!_preserved_oop_stack.is_empty()) { 9036 oop p = _preserved_oop_stack.pop(); 9037 assert(p->is_oop(), "Should be an oop"); 9038 assert(_span.contains(p), "oop should be in _span"); 9039 assert(p->mark() == markOopDesc::prototype(), 9040 "Set when taken from overflow list"); 9041 markOop m = _preserved_mark_stack.pop(); 9042 p->set_mark(m); 9043 } 9044 assert(_preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(), 9045 "stacks were cleared above"); 9046 } 9047 9048 #ifndef PRODUCT 9049 bool CMSCollector::no_preserved_marks() const { 9050 return _preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(); 9051 } 9052 #endif 9053 9054 CMSAdaptiveSizePolicy* ASConcurrentMarkSweepGeneration::cms_size_policy() const 9055 { 9056 GenCollectedHeap* gch = (GenCollectedHeap*) GenCollectedHeap::heap(); 9057 CMSAdaptiveSizePolicy* size_policy = 9058 (CMSAdaptiveSizePolicy*) gch->gen_policy()->size_policy(); 9059 assert(size_policy->is_gc_cms_adaptive_size_policy(), 9060 "Wrong type for size policy"); 9061 return size_policy; 9062 } 9063 9064 void ASConcurrentMarkSweepGeneration::resize(size_t cur_promo_size, 9065 size_t desired_promo_size) { 9066 if (cur_promo_size < desired_promo_size) { 9067 size_t expand_bytes = desired_promo_size - cur_promo_size; 9068 if (PrintAdaptiveSizePolicy && Verbose) { 9069 gclog_or_tty->print_cr(" ASConcurrentMarkSweepGeneration::resize " 9070 "Expanding tenured generation by " SIZE_FORMAT " (bytes)", 9071 expand_bytes); 9072 } 9073 expand(expand_bytes, 9074 MinHeapDeltaBytes, 9075 CMSExpansionCause::_adaptive_size_policy); 9076 } else if (desired_promo_size < cur_promo_size) { 9077 size_t shrink_bytes = cur_promo_size - desired_promo_size; 9078 if (PrintAdaptiveSizePolicy && Verbose) { 9079 gclog_or_tty->print_cr(" ASConcurrentMarkSweepGeneration::resize " 9080 "Shrinking tenured generation by " SIZE_FORMAT " (bytes)", 9081 shrink_bytes); 9082 } 9083 shrink(shrink_bytes); 9084 } 9085 } 9086 9087 CMSGCAdaptivePolicyCounters* ASConcurrentMarkSweepGeneration::gc_adaptive_policy_counters() { 9088 GenCollectedHeap* gch = GenCollectedHeap::heap(); 9089 CMSGCAdaptivePolicyCounters* counters = 9090 (CMSGCAdaptivePolicyCounters*) gch->collector_policy()->counters(); 9091 assert(counters->kind() == GCPolicyCounters::CMSGCAdaptivePolicyCountersKind, 9092 "Wrong kind of counters"); 9093 return counters; 9094 } 9095 9096 9097 void ASConcurrentMarkSweepGeneration::update_counters() { 9098 if (UsePerfData) { 9099 _space_counters->update_all(); 9100 _gen_counters->update_all(); 9101 CMSGCAdaptivePolicyCounters* counters = gc_adaptive_policy_counters(); 9102 GenCollectedHeap* gch = GenCollectedHeap::heap(); 9103 CMSGCStats* gc_stats_l = (CMSGCStats*) gc_stats(); 9104 assert(gc_stats_l->kind() == GCStats::CMSGCStatsKind, 9105 "Wrong gc statistics type"); 9106 counters->update_counters(gc_stats_l); 9107 } 9108 } 9109 9110 void ASConcurrentMarkSweepGeneration::update_counters(size_t used) { 9111 if (UsePerfData) { 9112 _space_counters->update_used(used); 9113 _space_counters->update_capacity(); 9114 _gen_counters->update_all(); 9115 9116 CMSGCAdaptivePolicyCounters* counters = gc_adaptive_policy_counters(); 9117 GenCollectedHeap* gch = GenCollectedHeap::heap(); 9118 CMSGCStats* gc_stats_l = (CMSGCStats*) gc_stats(); 9119 assert(gc_stats_l->kind() == GCStats::CMSGCStatsKind, 9120 "Wrong gc statistics type"); 9121 counters->update_counters(gc_stats_l); 9122 } 9123 } 9124 9125 void ASConcurrentMarkSweepGeneration::shrink_by(size_t desired_bytes) { 9126 assert_locked_or_safepoint(Heap_lock); 9127 assert_lock_strong(freelistLock()); 9128 HeapWord* old_end = _cmsSpace->end(); 9129 HeapWord* unallocated_start = _cmsSpace->unallocated_block(); 9130 assert(old_end >= unallocated_start, "Miscalculation of unallocated_start"); 9131 FreeChunk* chunk_at_end = find_chunk_at_end(); 9132 if (chunk_at_end == NULL) { 9133 // No room to shrink 9134 if (PrintGCDetails && Verbose) { 9135 gclog_or_tty->print_cr("No room to shrink: old_end " 9136 PTR_FORMAT " unallocated_start " PTR_FORMAT 9137 " chunk_at_end " PTR_FORMAT, 9138 old_end, unallocated_start, chunk_at_end); 9139 } 9140 return; 9141 } else { 9142 9143 // Find the chunk at the end of the space and determine 9144 // how much it can be shrunk. 9145 size_t shrinkable_size_in_bytes = chunk_at_end->size(); 9146 size_t aligned_shrinkable_size_in_bytes = 9147 align_size_down(shrinkable_size_in_bytes, os::vm_page_size()); 9148 assert(unallocated_start <= (HeapWord*) chunk_at_end->end(), 9149 "Inconsistent chunk at end of space"); 9150 size_t bytes = MIN2(desired_bytes, aligned_shrinkable_size_in_bytes); 9151 size_t word_size_before = heap_word_size(_virtual_space.committed_size()); 9152 9153 // Shrink the underlying space 9154 _virtual_space.shrink_by(bytes); 9155 if (PrintGCDetails && Verbose) { 9156 gclog_or_tty->print_cr("ConcurrentMarkSweepGeneration::shrink_by:" 9157 " desired_bytes " SIZE_FORMAT 9158 " shrinkable_size_in_bytes " SIZE_FORMAT 9159 " aligned_shrinkable_size_in_bytes " SIZE_FORMAT 9160 " bytes " SIZE_FORMAT, 9161 desired_bytes, shrinkable_size_in_bytes, 9162 aligned_shrinkable_size_in_bytes, bytes); 9163 gclog_or_tty->print_cr(" old_end " SIZE_FORMAT 9164 " unallocated_start " SIZE_FORMAT, 9165 old_end, unallocated_start); 9166 } 9167 9168 // If the space did shrink (shrinking is not guaranteed), 9169 // shrink the chunk at the end by the appropriate amount. 9170 if (((HeapWord*)_virtual_space.high()) < old_end) { 9171 size_t new_word_size = 9172 heap_word_size(_virtual_space.committed_size()); 9173 9174 // Have to remove the chunk from the dictionary because it is changing 9175 // size and might be someplace elsewhere in the dictionary. 9176 9177 // Get the chunk at end, shrink it, and put it 9178 // back. 9179 _cmsSpace->removeChunkFromDictionary(chunk_at_end); 9180 size_t word_size_change = word_size_before - new_word_size; 9181 size_t chunk_at_end_old_size = chunk_at_end->size(); 9182 assert(chunk_at_end_old_size >= word_size_change, 9183 "Shrink is too large"); 9184 chunk_at_end->set_size(chunk_at_end_old_size - 9185 word_size_change); 9186 _cmsSpace->freed((HeapWord*) chunk_at_end->end(), 9187 word_size_change); 9188 9189 _cmsSpace->returnChunkToDictionary(chunk_at_end); 9190 9191 MemRegion mr(_cmsSpace->bottom(), new_word_size); 9192 _bts->resize(new_word_size); // resize the block offset shared array 9193 Universe::heap()->barrier_set()->resize_covered_region(mr); 9194 _cmsSpace->assert_locked(); 9195 _cmsSpace->set_end((HeapWord*)_virtual_space.high()); 9196 9197 NOT_PRODUCT(_cmsSpace->dictionary()->verify()); 9198 9199 // update the space and generation capacity counters 9200 if (UsePerfData) { 9201 _space_counters->update_capacity(); 9202 _gen_counters->update_all(); 9203 } 9204 9205 if (Verbose && PrintGCDetails) { 9206 size_t new_mem_size = _virtual_space.committed_size(); 9207 size_t old_mem_size = new_mem_size + bytes; 9208 gclog_or_tty->print_cr("Shrinking %s from " SIZE_FORMAT "K by " SIZE_FORMAT "K to " SIZE_FORMAT "K", 9209 name(), old_mem_size/K, bytes/K, new_mem_size/K); 9210 } 9211 } 9212 9213 assert(_cmsSpace->unallocated_block() <= _cmsSpace->end(), 9214 "Inconsistency at end of space"); 9215 assert(chunk_at_end->end() == (uintptr_t*) _cmsSpace->end(), 9216 "Shrinking is inconsistent"); 9217 return; 9218 } 9219 } 9220 9221 // Transfer some number of overflown objects to usual marking 9222 // stack. Return true if some objects were transferred. 9223 bool MarkRefsIntoAndScanClosure::take_from_overflow_list() { 9224 size_t num = MIN2((size_t)(_mark_stack->capacity() - _mark_stack->length())/4, 9225 (size_t)ParGCDesiredObjsFromOverflowList); 9226 9227 bool res = _collector->take_from_overflow_list(num, _mark_stack); 9228 assert(_collector->overflow_list_is_empty() || res, 9229 "If list is not empty, we should have taken something"); 9230 assert(!res || !_mark_stack->isEmpty(), 9231 "If we took something, it should now be on our stack"); 9232 return res; 9233 } 9234 9235 size_t MarkDeadObjectsClosure::do_blk(HeapWord* addr) { 9236 size_t res = _sp->block_size_no_stall(addr, _collector); 9237 if (_sp->block_is_obj(addr)) { 9238 if (_live_bit_map->isMarked(addr)) { 9239 // It can't have been dead in a previous cycle 9240 guarantee(!_dead_bit_map->isMarked(addr), "No resurrection!"); 9241 } else { 9242 _dead_bit_map->mark(addr); // mark the dead object 9243 } 9244 } 9245 // Could be 0, if the block size could not be computed without stalling. 9246 return res; 9247 } 9248 9249 TraceCMSMemoryManagerStats::TraceCMSMemoryManagerStats(CMSCollector::CollectorState phase, GCCause::Cause cause): TraceMemoryManagerStats() { 9250 9251 switch (phase) { 9252 case CMSCollector::InitialMarking: 9253 initialize(true /* fullGC */ , 9254 cause /* cause of the GC */, 9255 true /* recordGCBeginTime */, 9256 true /* recordPreGCUsage */, 9257 false /* recordPeakUsage */, 9258 false /* recordPostGCusage */, 9259 true /* recordAccumulatedGCTime */, 9260 false /* recordGCEndTime */, 9261 false /* countCollection */ ); 9262 break; 9263 9264 case CMSCollector::FinalMarking: 9265 initialize(true /* fullGC */ , 9266 cause /* cause of the GC */, 9267 false /* recordGCBeginTime */, 9268 false /* recordPreGCUsage */, 9269 false /* recordPeakUsage */, 9270 false /* recordPostGCusage */, 9271 true /* recordAccumulatedGCTime */, 9272 false /* recordGCEndTime */, 9273 false /* countCollection */ ); 9274 break; 9275 9276 case CMSCollector::Sweeping: 9277 initialize(true /* fullGC */ , 9278 cause /* cause of the GC */, 9279 false /* recordGCBeginTime */, 9280 false /* recordPreGCUsage */, 9281 true /* recordPeakUsage */, 9282 true /* recordPostGCusage */, 9283 false /* recordAccumulatedGCTime */, 9284 true /* recordGCEndTime */, 9285 true /* countCollection */ ); 9286 break; 9287 9288 default: 9289 ShouldNotReachHere(); 9290 } 9291 } 9292