1 /* 2 * Copyright (c) 2001, 2012, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "classfile/classLoaderData.hpp" 27 #include "classfile/symbolTable.hpp" 28 #include "classfile/systemDictionary.hpp" 29 #include "code/codeCache.hpp" 30 #include "gc_implementation/concurrentMarkSweep/cmsAdaptiveSizePolicy.hpp" 31 #include "gc_implementation/concurrentMarkSweep/cmsCollectorPolicy.hpp" 32 #include "gc_implementation/concurrentMarkSweep/cmsGCAdaptivePolicyCounters.hpp" 33 #include "gc_implementation/concurrentMarkSweep/cmsOopClosures.inline.hpp" 34 #include "gc_implementation/concurrentMarkSweep/compactibleFreeListSpace.hpp" 35 #include "gc_implementation/concurrentMarkSweep/concurrentMarkSweepGeneration.inline.hpp" 36 #include "gc_implementation/concurrentMarkSweep/concurrentMarkSweepThread.hpp" 37 #include "gc_implementation/concurrentMarkSweep/vmCMSOperations.hpp" 38 #include "gc_implementation/parNew/parNewGeneration.hpp" 39 #include "gc_implementation/shared/collectorCounters.hpp" 40 #include "gc_implementation/shared/isGCActiveMark.hpp" 41 #include "gc_interface/collectedHeap.inline.hpp" 42 #include "memory/cardTableRS.hpp" 43 #include "memory/collectorPolicy.hpp" 44 #include "memory/gcLocker.inline.hpp" 45 #include "memory/genCollectedHeap.hpp" 46 #include "memory/genMarkSweep.hpp" 47 #include "memory/genOopClosures.inline.hpp" 48 #include "memory/iterator.hpp" 49 #include "memory/referencePolicy.hpp" 50 #include "memory/resourceArea.hpp" 51 #include "oops/oop.inline.hpp" 52 #include "prims/jvmtiExport.hpp" 53 #include "runtime/globals_extension.hpp" 54 #include "runtime/handles.inline.hpp" 55 #include "runtime/java.hpp" 56 #include "runtime/vmThread.hpp" 57 #include "services/memoryService.hpp" 58 #include "services/runtimeService.hpp" 59 60 // statics 61 CMSCollector* ConcurrentMarkSweepGeneration::_collector = NULL; 62 bool CMSCollector::_full_gc_requested = false; 63 64 ////////////////////////////////////////////////////////////////// 65 // In support of CMS/VM thread synchronization 66 ////////////////////////////////////////////////////////////////// 67 // We split use of the CGC_lock into 2 "levels". 68 // The low-level locking is of the usual CGC_lock monitor. We introduce 69 // a higher level "token" (hereafter "CMS token") built on top of the 70 // low level monitor (hereafter "CGC lock"). 71 // The token-passing protocol gives priority to the VM thread. The 72 // CMS-lock doesn't provide any fairness guarantees, but clients 73 // should ensure that it is only held for very short, bounded 74 // durations. 75 // 76 // When either of the CMS thread or the VM thread is involved in 77 // collection operations during which it does not want the other 78 // thread to interfere, it obtains the CMS token. 79 // 80 // If either thread tries to get the token while the other has 81 // it, that thread waits. However, if the VM thread and CMS thread 82 // both want the token, then the VM thread gets priority while the 83 // CMS thread waits. This ensures, for instance, that the "concurrent" 84 // phases of the CMS thread's work do not block out the VM thread 85 // for long periods of time as the CMS thread continues to hog 86 // the token. (See bug 4616232). 87 // 88 // The baton-passing functions are, however, controlled by the 89 // flags _foregroundGCShouldWait and _foregroundGCIsActive, 90 // and here the low-level CMS lock, not the high level token, 91 // ensures mutual exclusion. 92 // 93 // Two important conditions that we have to satisfy: 94 // 1. if a thread does a low-level wait on the CMS lock, then it 95 // relinquishes the CMS token if it were holding that token 96 // when it acquired the low-level CMS lock. 97 // 2. any low-level notifications on the low-level lock 98 // should only be sent when a thread has relinquished the token. 99 // 100 // In the absence of either property, we'd have potential deadlock. 101 // 102 // We protect each of the CMS (concurrent and sequential) phases 103 // with the CMS _token_, not the CMS _lock_. 104 // 105 // The only code protected by CMS lock is the token acquisition code 106 // itself, see ConcurrentMarkSweepThread::[de]synchronize(), and the 107 // baton-passing code. 108 // 109 // Unfortunately, i couldn't come up with a good abstraction to factor and 110 // hide the naked CGC_lock manipulation in the baton-passing code 111 // further below. That's something we should try to do. Also, the proof 112 // of correctness of this 2-level locking scheme is far from obvious, 113 // and potentially quite slippery. We have an uneasy supsicion, for instance, 114 // that there may be a theoretical possibility of delay/starvation in the 115 // low-level lock/wait/notify scheme used for the baton-passing because of 116 // potential intereference with the priority scheme embodied in the 117 // CMS-token-passing protocol. See related comments at a CGC_lock->wait() 118 // invocation further below and marked with "XXX 20011219YSR". 119 // Indeed, as we note elsewhere, this may become yet more slippery 120 // in the presence of multiple CMS and/or multiple VM threads. XXX 121 122 class CMSTokenSync: public StackObj { 123 private: 124 bool _is_cms_thread; 125 public: 126 CMSTokenSync(bool is_cms_thread): 127 _is_cms_thread(is_cms_thread) { 128 assert(is_cms_thread == Thread::current()->is_ConcurrentGC_thread(), 129 "Incorrect argument to constructor"); 130 ConcurrentMarkSweepThread::synchronize(_is_cms_thread); 131 } 132 133 ~CMSTokenSync() { 134 assert(_is_cms_thread ? 135 ConcurrentMarkSweepThread::cms_thread_has_cms_token() : 136 ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 137 "Incorrect state"); 138 ConcurrentMarkSweepThread::desynchronize(_is_cms_thread); 139 } 140 }; 141 142 // Convenience class that does a CMSTokenSync, and then acquires 143 // upto three locks. 144 class CMSTokenSyncWithLocks: public CMSTokenSync { 145 private: 146 // Note: locks are acquired in textual declaration order 147 // and released in the opposite order 148 MutexLockerEx _locker1, _locker2, _locker3; 149 public: 150 CMSTokenSyncWithLocks(bool is_cms_thread, Mutex* mutex1, 151 Mutex* mutex2 = NULL, Mutex* mutex3 = NULL): 152 CMSTokenSync(is_cms_thread), 153 _locker1(mutex1, Mutex::_no_safepoint_check_flag), 154 _locker2(mutex2, Mutex::_no_safepoint_check_flag), 155 _locker3(mutex3, Mutex::_no_safepoint_check_flag) 156 { } 157 }; 158 159 160 // Wrapper class to temporarily disable icms during a foreground cms collection. 161 class ICMSDisabler: public StackObj { 162 public: 163 // The ctor disables icms and wakes up the thread so it notices the change; 164 // the dtor re-enables icms. Note that the CMSCollector methods will check 165 // CMSIncrementalMode. 166 ICMSDisabler() { CMSCollector::disable_icms(); CMSCollector::start_icms(); } 167 ~ICMSDisabler() { CMSCollector::enable_icms(); } 168 }; 169 170 ////////////////////////////////////////////////////////////////// 171 // Concurrent Mark-Sweep Generation ///////////////////////////// 172 ////////////////////////////////////////////////////////////////// 173 174 NOT_PRODUCT(CompactibleFreeListSpace* debug_cms_space;) 175 176 // This struct contains per-thread things necessary to support parallel 177 // young-gen collection. 178 class CMSParGCThreadState: public CHeapObj<mtGC> { 179 public: 180 CFLS_LAB lab; 181 PromotionInfo promo; 182 183 // Constructor. 184 CMSParGCThreadState(CompactibleFreeListSpace* cfls) : lab(cfls) { 185 promo.setSpace(cfls); 186 } 187 }; 188 189 ConcurrentMarkSweepGeneration::ConcurrentMarkSweepGeneration( 190 ReservedSpace rs, size_t initial_byte_size, int level, 191 CardTableRS* ct, bool use_adaptive_freelists, 192 FreeBlockDictionary<FreeChunk>::DictionaryChoice dictionaryChoice) : 193 CardGeneration(rs, initial_byte_size, level, ct), 194 _dilatation_factor(((double)MinChunkSize)/((double)(CollectedHeap::min_fill_size()))), 195 _debug_collection_type(Concurrent_collection_type) 196 { 197 HeapWord* bottom = (HeapWord*) _virtual_space.low(); 198 HeapWord* end = (HeapWord*) _virtual_space.high(); 199 200 _direct_allocated_words = 0; 201 NOT_PRODUCT( 202 _numObjectsPromoted = 0; 203 _numWordsPromoted = 0; 204 _numObjectsAllocated = 0; 205 _numWordsAllocated = 0; 206 ) 207 208 _cmsSpace = new CompactibleFreeListSpace(_bts, MemRegion(bottom, end), 209 use_adaptive_freelists, 210 dictionaryChoice); 211 NOT_PRODUCT(debug_cms_space = _cmsSpace;) 212 if (_cmsSpace == NULL) { 213 vm_exit_during_initialization( 214 "CompactibleFreeListSpace allocation failure"); 215 } 216 _cmsSpace->_gen = this; 217 218 _gc_stats = new CMSGCStats(); 219 220 // Verify the assumption that FreeChunk::_prev and OopDesc::_klass 221 // offsets match. The ability to tell free chunks from objects 222 // depends on this property. 223 debug_only( 224 FreeChunk* junk = NULL; 225 assert(UseCompressedOops || 226 junk->prev_addr() == (void*)(oop(junk)->klass_addr()), 227 "Offset of FreeChunk::_prev within FreeChunk must match" 228 " that of OopDesc::_klass within OopDesc"); 229 ) 230 if (CollectedHeap::use_parallel_gc_threads()) { 231 typedef CMSParGCThreadState* CMSParGCThreadStatePtr; 232 _par_gc_thread_states = 233 NEW_C_HEAP_ARRAY(CMSParGCThreadStatePtr, ParallelGCThreads, mtGC); 234 if (_par_gc_thread_states == NULL) { 235 vm_exit_during_initialization("Could not allocate par gc structs"); 236 } 237 for (uint i = 0; i < ParallelGCThreads; i++) { 238 _par_gc_thread_states[i] = new CMSParGCThreadState(cmsSpace()); 239 if (_par_gc_thread_states[i] == NULL) { 240 vm_exit_during_initialization("Could not allocate par gc structs"); 241 } 242 } 243 } else { 244 _par_gc_thread_states = NULL; 245 } 246 _incremental_collection_failed = false; 247 // The "dilatation_factor" is the expansion that can occur on 248 // account of the fact that the minimum object size in the CMS 249 // generation may be larger than that in, say, a contiguous young 250 // generation. 251 // Ideally, in the calculation below, we'd compute the dilatation 252 // factor as: MinChunkSize/(promoting_gen's min object size) 253 // Since we do not have such a general query interface for the 254 // promoting generation, we'll instead just use the mimimum 255 // object size (which today is a header's worth of space); 256 // note that all arithmetic is in units of HeapWords. 257 assert(MinChunkSize >= CollectedHeap::min_fill_size(), "just checking"); 258 assert(_dilatation_factor >= 1.0, "from previous assert"); 259 } 260 261 262 // The field "_initiating_occupancy" represents the occupancy percentage 263 // at which we trigger a new collection cycle. Unless explicitly specified 264 // via CMSInitiatingOccupancyFraction (argument "io" below), it 265 // is calculated by: 266 // 267 // Let "f" be MinHeapFreeRatio in 268 // 269 // _intiating_occupancy = 100-f + 270 // f * (CMSTriggerRatio/100) 271 // where CMSTriggerRatio is the argument "tr" below. 272 // 273 // That is, if we assume the heap is at its desired maximum occupancy at the 274 // end of a collection, we let CMSTriggerRatio of the (purported) free 275 // space be allocated before initiating a new collection cycle. 276 // 277 void ConcurrentMarkSweepGeneration::init_initiating_occupancy(intx io, intx tr) { 278 assert(io <= 100 && tr >= 0 && tr <= 100, "Check the arguments"); 279 if (io >= 0) { 280 _initiating_occupancy = (double)io / 100.0; 281 } else { 282 _initiating_occupancy = ((100 - MinHeapFreeRatio) + 283 (double)(tr * MinHeapFreeRatio) / 100.0) 284 / 100.0; 285 } 286 } 287 288 void ConcurrentMarkSweepGeneration::ref_processor_init() { 289 assert(collector() != NULL, "no collector"); 290 collector()->ref_processor_init(); 291 } 292 293 void CMSCollector::ref_processor_init() { 294 if (_ref_processor == NULL) { 295 // Allocate and initialize a reference processor 296 _ref_processor = 297 new ReferenceProcessor(_span, // span 298 (ParallelGCThreads > 1) && ParallelRefProcEnabled, // mt processing 299 (int) ParallelGCThreads, // mt processing degree 300 _cmsGen->refs_discovery_is_mt(), // mt discovery 301 (int) MAX2(ConcGCThreads, ParallelGCThreads), // mt discovery degree 302 _cmsGen->refs_discovery_is_atomic(), // discovery is not atomic 303 &_is_alive_closure, // closure for liveness info 304 false); // next field updates do not need write barrier 305 // Initialize the _ref_processor field of CMSGen 306 _cmsGen->set_ref_processor(_ref_processor); 307 308 } 309 } 310 311 CMSAdaptiveSizePolicy* CMSCollector::size_policy() { 312 GenCollectedHeap* gch = GenCollectedHeap::heap(); 313 assert(gch->kind() == CollectedHeap::GenCollectedHeap, 314 "Wrong type of heap"); 315 CMSAdaptiveSizePolicy* sp = (CMSAdaptiveSizePolicy*) 316 gch->gen_policy()->size_policy(); 317 assert(sp->is_gc_cms_adaptive_size_policy(), 318 "Wrong type of size policy"); 319 return sp; 320 } 321 322 CMSGCAdaptivePolicyCounters* CMSCollector::gc_adaptive_policy_counters() { 323 CMSGCAdaptivePolicyCounters* results = 324 (CMSGCAdaptivePolicyCounters*) collector_policy()->counters(); 325 assert( 326 results->kind() == GCPolicyCounters::CMSGCAdaptivePolicyCountersKind, 327 "Wrong gc policy counter kind"); 328 return results; 329 } 330 331 332 void ConcurrentMarkSweepGeneration::initialize_performance_counters() { 333 334 const char* gen_name = "old"; 335 336 // Generation Counters - generation 1, 1 subspace 337 _gen_counters = new GenerationCounters(gen_name, 1, 1, &_virtual_space); 338 339 _space_counters = new GSpaceCounters(gen_name, 0, 340 _virtual_space.reserved_size(), 341 this, _gen_counters); 342 } 343 344 CMSStats::CMSStats(ConcurrentMarkSweepGeneration* cms_gen, unsigned int alpha): 345 _cms_gen(cms_gen) 346 { 347 assert(alpha <= 100, "bad value"); 348 _saved_alpha = alpha; 349 350 // Initialize the alphas to the bootstrap value of 100. 351 _gc0_alpha = _cms_alpha = 100; 352 353 _cms_begin_time.update(); 354 _cms_end_time.update(); 355 356 _gc0_duration = 0.0; 357 _gc0_period = 0.0; 358 _gc0_promoted = 0; 359 360 _cms_duration = 0.0; 361 _cms_period = 0.0; 362 _cms_allocated = 0; 363 364 _cms_used_at_gc0_begin = 0; 365 _cms_used_at_gc0_end = 0; 366 _allow_duty_cycle_reduction = false; 367 _valid_bits = 0; 368 _icms_duty_cycle = CMSIncrementalDutyCycle; 369 } 370 371 double CMSStats::cms_free_adjustment_factor(size_t free) const { 372 // TBD: CR 6909490 373 return 1.0; 374 } 375 376 void CMSStats::adjust_cms_free_adjustment_factor(bool fail, size_t free) { 377 } 378 379 // If promotion failure handling is on use 380 // the padded average size of the promotion for each 381 // young generation collection. 382 double CMSStats::time_until_cms_gen_full() const { 383 size_t cms_free = _cms_gen->cmsSpace()->free(); 384 GenCollectedHeap* gch = GenCollectedHeap::heap(); 385 size_t expected_promotion = MIN2(gch->get_gen(0)->capacity(), 386 (size_t) _cms_gen->gc_stats()->avg_promoted()->padded_average()); 387 if (cms_free > expected_promotion) { 388 // Start a cms collection if there isn't enough space to promote 389 // for the next minor collection. Use the padded average as 390 // a safety factor. 391 cms_free -= expected_promotion; 392 393 // Adjust by the safety factor. 394 double cms_free_dbl = (double)cms_free; 395 double cms_adjustment = (100.0 - CMSIncrementalSafetyFactor)/100.0; 396 // Apply a further correction factor which tries to adjust 397 // for recent occurance of concurrent mode failures. 398 cms_adjustment = cms_adjustment * cms_free_adjustment_factor(cms_free); 399 cms_free_dbl = cms_free_dbl * cms_adjustment; 400 401 if (PrintGCDetails && Verbose) { 402 gclog_or_tty->print_cr("CMSStats::time_until_cms_gen_full: cms_free " 403 SIZE_FORMAT " expected_promotion " SIZE_FORMAT, 404 cms_free, expected_promotion); 405 gclog_or_tty->print_cr(" cms_free_dbl %f cms_consumption_rate %f", 406 cms_free_dbl, cms_consumption_rate() + 1.0); 407 } 408 // Add 1 in case the consumption rate goes to zero. 409 return cms_free_dbl / (cms_consumption_rate() + 1.0); 410 } 411 return 0.0; 412 } 413 414 // Compare the duration of the cms collection to the 415 // time remaining before the cms generation is empty. 416 // Note that the time from the start of the cms collection 417 // to the start of the cms sweep (less than the total 418 // duration of the cms collection) can be used. This 419 // has been tried and some applications experienced 420 // promotion failures early in execution. This was 421 // possibly because the averages were not accurate 422 // enough at the beginning. 423 double CMSStats::time_until_cms_start() const { 424 // We add "gc0_period" to the "work" calculation 425 // below because this query is done (mostly) at the 426 // end of a scavenge, so we need to conservatively 427 // account for that much possible delay 428 // in the query so as to avoid concurrent mode failures 429 // due to starting the collection just a wee bit too 430 // late. 431 double work = cms_duration() + gc0_period(); 432 double deadline = time_until_cms_gen_full(); 433 // If a concurrent mode failure occurred recently, we want to be 434 // more conservative and halve our expected time_until_cms_gen_full() 435 if (work > deadline) { 436 if (Verbose && PrintGCDetails) { 437 gclog_or_tty->print( 438 " CMSCollector: collect because of anticipated promotion " 439 "before full %3.7f + %3.7f > %3.7f ", cms_duration(), 440 gc0_period(), time_until_cms_gen_full()); 441 } 442 return 0.0; 443 } 444 return work - deadline; 445 } 446 447 // Return a duty cycle based on old_duty_cycle and new_duty_cycle, limiting the 448 // amount of change to prevent wild oscillation. 449 unsigned int CMSStats::icms_damped_duty_cycle(unsigned int old_duty_cycle, 450 unsigned int new_duty_cycle) { 451 assert(old_duty_cycle <= 100, "bad input value"); 452 assert(new_duty_cycle <= 100, "bad input value"); 453 454 // Note: use subtraction with caution since it may underflow (values are 455 // unsigned). Addition is safe since we're in the range 0-100. 456 unsigned int damped_duty_cycle = new_duty_cycle; 457 if (new_duty_cycle < old_duty_cycle) { 458 const unsigned int largest_delta = MAX2(old_duty_cycle / 4, 5U); 459 if (new_duty_cycle + largest_delta < old_duty_cycle) { 460 damped_duty_cycle = old_duty_cycle - largest_delta; 461 } 462 } else if (new_duty_cycle > old_duty_cycle) { 463 const unsigned int largest_delta = MAX2(old_duty_cycle / 4, 15U); 464 if (new_duty_cycle > old_duty_cycle + largest_delta) { 465 damped_duty_cycle = MIN2(old_duty_cycle + largest_delta, 100U); 466 } 467 } 468 assert(damped_duty_cycle <= 100, "invalid duty cycle computed"); 469 470 if (CMSTraceIncrementalPacing) { 471 gclog_or_tty->print(" [icms_damped_duty_cycle(%d,%d) = %d] ", 472 old_duty_cycle, new_duty_cycle, damped_duty_cycle); 473 } 474 return damped_duty_cycle; 475 } 476 477 unsigned int CMSStats::icms_update_duty_cycle_impl() { 478 assert(CMSIncrementalPacing && valid(), 479 "should be handled in icms_update_duty_cycle()"); 480 481 double cms_time_so_far = cms_timer().seconds(); 482 double scaled_duration = cms_duration_per_mb() * _cms_used_at_gc0_end / M; 483 double scaled_duration_remaining = fabsd(scaled_duration - cms_time_so_far); 484 485 // Avoid division by 0. 486 double time_until_full = MAX2(time_until_cms_gen_full(), 0.01); 487 double duty_cycle_dbl = 100.0 * scaled_duration_remaining / time_until_full; 488 489 unsigned int new_duty_cycle = MIN2((unsigned int)duty_cycle_dbl, 100U); 490 if (new_duty_cycle > _icms_duty_cycle) { 491 // Avoid very small duty cycles (1 or 2); 0 is allowed. 492 if (new_duty_cycle > 2) { 493 _icms_duty_cycle = icms_damped_duty_cycle(_icms_duty_cycle, 494 new_duty_cycle); 495 } 496 } else if (_allow_duty_cycle_reduction) { 497 // The duty cycle is reduced only once per cms cycle (see record_cms_end()). 498 new_duty_cycle = icms_damped_duty_cycle(_icms_duty_cycle, new_duty_cycle); 499 // Respect the minimum duty cycle. 500 unsigned int min_duty_cycle = (unsigned int)CMSIncrementalDutyCycleMin; 501 _icms_duty_cycle = MAX2(new_duty_cycle, min_duty_cycle); 502 } 503 504 if (PrintGCDetails || CMSTraceIncrementalPacing) { 505 gclog_or_tty->print(" icms_dc=%d ", _icms_duty_cycle); 506 } 507 508 _allow_duty_cycle_reduction = false; 509 return _icms_duty_cycle; 510 } 511 512 #ifndef PRODUCT 513 void CMSStats::print_on(outputStream *st) const { 514 st->print(" gc0_alpha=%d,cms_alpha=%d", _gc0_alpha, _cms_alpha); 515 st->print(",gc0_dur=%g,gc0_per=%g,gc0_promo=" SIZE_FORMAT, 516 gc0_duration(), gc0_period(), gc0_promoted()); 517 st->print(",cms_dur=%g,cms_dur_per_mb=%g,cms_per=%g,cms_alloc=" SIZE_FORMAT, 518 cms_duration(), cms_duration_per_mb(), 519 cms_period(), cms_allocated()); 520 st->print(",cms_since_beg=%g,cms_since_end=%g", 521 cms_time_since_begin(), cms_time_since_end()); 522 st->print(",cms_used_beg=" SIZE_FORMAT ",cms_used_end=" SIZE_FORMAT, 523 _cms_used_at_gc0_begin, _cms_used_at_gc0_end); 524 if (CMSIncrementalMode) { 525 st->print(",dc=%d", icms_duty_cycle()); 526 } 527 528 if (valid()) { 529 st->print(",promo_rate=%g,cms_alloc_rate=%g", 530 promotion_rate(), cms_allocation_rate()); 531 st->print(",cms_consumption_rate=%g,time_until_full=%g", 532 cms_consumption_rate(), time_until_cms_gen_full()); 533 } 534 st->print(" "); 535 } 536 #endif // #ifndef PRODUCT 537 538 CMSCollector::CollectorState CMSCollector::_collectorState = 539 CMSCollector::Idling; 540 bool CMSCollector::_foregroundGCIsActive = false; 541 bool CMSCollector::_foregroundGCShouldWait = false; 542 543 CMSCollector::CMSCollector(ConcurrentMarkSweepGeneration* cmsGen, 544 CardTableRS* ct, 545 ConcurrentMarkSweepPolicy* cp): 546 _cmsGen(cmsGen), 547 _ct(ct), 548 _ref_processor(NULL), // will be set later 549 _conc_workers(NULL), // may be set later 550 _abort_preclean(false), 551 _start_sampling(false), 552 _between_prologue_and_epilogue(false), 553 _markBitMap(0, Mutex::leaf + 1, "CMS_markBitMap_lock"), 554 _modUnionTable((CardTableModRefBS::card_shift - LogHeapWordSize), 555 -1 /* lock-free */, "No_lock" /* dummy */), 556 _modUnionClosure(&_modUnionTable), 557 _modUnionClosurePar(&_modUnionTable), 558 // Adjust my span to cover old (cms) gen 559 _span(cmsGen->reserved()), 560 // Construct the is_alive_closure with _span & markBitMap 561 _is_alive_closure(_span, &_markBitMap), 562 _restart_addr(NULL), 563 _overflow_list(NULL), 564 _stats(cmsGen), 565 _eden_chunk_array(NULL), // may be set in ctor body 566 _eden_chunk_capacity(0), // -- ditto -- 567 _eden_chunk_index(0), // -- ditto -- 568 _survivor_plab_array(NULL), // -- ditto -- 569 _survivor_chunk_array(NULL), // -- ditto -- 570 _survivor_chunk_capacity(0), // -- ditto -- 571 _survivor_chunk_index(0), // -- ditto -- 572 _ser_pmc_preclean_ovflw(0), 573 _ser_kac_preclean_ovflw(0), 574 _ser_pmc_remark_ovflw(0), 575 _par_pmc_remark_ovflw(0), 576 _ser_kac_ovflw(0), 577 _par_kac_ovflw(0), 578 #ifndef PRODUCT 579 _num_par_pushes(0), 580 #endif 581 _collection_count_start(0), 582 _verifying(false), 583 _icms_start_limit(NULL), 584 _icms_stop_limit(NULL), 585 _verification_mark_bm(0, Mutex::leaf + 1, "CMS_verification_mark_bm_lock"), 586 _completed_initialization(false), 587 _collector_policy(cp), 588 _should_unload_classes(false), 589 _concurrent_cycles_since_last_unload(0), 590 _roots_scanning_options(0), 591 _inter_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding), 592 _intra_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding) 593 { 594 if (ExplicitGCInvokesConcurrentAndUnloadsClasses) { 595 ExplicitGCInvokesConcurrent = true; 596 } 597 // Now expand the span and allocate the collection support structures 598 // (MUT, marking bit map etc.) to cover both generations subject to 599 // collection. 600 601 // For use by dirty card to oop closures. 602 _cmsGen->cmsSpace()->set_collector(this); 603 604 // Allocate MUT and marking bit map 605 { 606 MutexLockerEx x(_markBitMap.lock(), Mutex::_no_safepoint_check_flag); 607 if (!_markBitMap.allocate(_span)) { 608 warning("Failed to allocate CMS Bit Map"); 609 return; 610 } 611 assert(_markBitMap.covers(_span), "_markBitMap inconsistency?"); 612 } 613 { 614 _modUnionTable.allocate(_span); 615 assert(_modUnionTable.covers(_span), "_modUnionTable inconsistency?"); 616 } 617 618 if (!_markStack.allocate(MarkStackSize)) { 619 warning("Failed to allocate CMS Marking Stack"); 620 return; 621 } 622 623 // Support for multi-threaded concurrent phases 624 if (CMSConcurrentMTEnabled) { 625 if (FLAG_IS_DEFAULT(ConcGCThreads)) { 626 // just for now 627 FLAG_SET_DEFAULT(ConcGCThreads, (ParallelGCThreads + 3)/4); 628 } 629 if (ConcGCThreads > 1) { 630 _conc_workers = new YieldingFlexibleWorkGang("Parallel CMS Threads", 631 ConcGCThreads, true); 632 if (_conc_workers == NULL) { 633 warning("GC/CMS: _conc_workers allocation failure: " 634 "forcing -CMSConcurrentMTEnabled"); 635 CMSConcurrentMTEnabled = false; 636 } else { 637 _conc_workers->initialize_workers(); 638 } 639 } else { 640 CMSConcurrentMTEnabled = false; 641 } 642 } 643 if (!CMSConcurrentMTEnabled) { 644 ConcGCThreads = 0; 645 } else { 646 // Turn off CMSCleanOnEnter optimization temporarily for 647 // the MT case where it's not fixed yet; see 6178663. 648 CMSCleanOnEnter = false; 649 } 650 assert((_conc_workers != NULL) == (ConcGCThreads > 1), 651 "Inconsistency"); 652 653 // Parallel task queues; these are shared for the 654 // concurrent and stop-world phases of CMS, but 655 // are not shared with parallel scavenge (ParNew). 656 { 657 uint i; 658 uint num_queues = (uint) MAX2(ParallelGCThreads, ConcGCThreads); 659 660 if ((CMSParallelRemarkEnabled || CMSConcurrentMTEnabled 661 || ParallelRefProcEnabled) 662 && num_queues > 0) { 663 _task_queues = new OopTaskQueueSet(num_queues); 664 if (_task_queues == NULL) { 665 warning("task_queues allocation failure."); 666 return; 667 } 668 _hash_seed = NEW_C_HEAP_ARRAY(int, num_queues, mtGC); 669 if (_hash_seed == NULL) { 670 warning("_hash_seed array allocation failure"); 671 return; 672 } 673 674 typedef Padded<OopTaskQueue> PaddedOopTaskQueue; 675 for (i = 0; i < num_queues; i++) { 676 PaddedOopTaskQueue *q = new PaddedOopTaskQueue(); 677 if (q == NULL) { 678 warning("work_queue allocation failure."); 679 return; 680 } 681 _task_queues->register_queue(i, q); 682 } 683 for (i = 0; i < num_queues; i++) { 684 _task_queues->queue(i)->initialize(); 685 _hash_seed[i] = 17; // copied from ParNew 686 } 687 } 688 } 689 690 _cmsGen ->init_initiating_occupancy(CMSInitiatingOccupancyFraction, CMSTriggerRatio); 691 692 // Clip CMSBootstrapOccupancy between 0 and 100. 693 _bootstrap_occupancy = ((double)MIN2((uintx)100, MAX2((uintx)0, CMSBootstrapOccupancy))) 694 /(double)100; 695 696 _full_gcs_since_conc_gc = 0; 697 698 // Now tell CMS generations the identity of their collector 699 ConcurrentMarkSweepGeneration::set_collector(this); 700 701 // Create & start a CMS thread for this CMS collector 702 _cmsThread = ConcurrentMarkSweepThread::start(this); 703 assert(cmsThread() != NULL, "CMS Thread should have been created"); 704 assert(cmsThread()->collector() == this, 705 "CMS Thread should refer to this gen"); 706 assert(CGC_lock != NULL, "Where's the CGC_lock?"); 707 708 // Support for parallelizing young gen rescan 709 GenCollectedHeap* gch = GenCollectedHeap::heap(); 710 _young_gen = gch->prev_gen(_cmsGen); 711 if (gch->supports_inline_contig_alloc()) { 712 _top_addr = gch->top_addr(); 713 _end_addr = gch->end_addr(); 714 assert(_young_gen != NULL, "no _young_gen"); 715 _eden_chunk_index = 0; 716 _eden_chunk_capacity = (_young_gen->max_capacity()+CMSSamplingGrain)/CMSSamplingGrain; 717 _eden_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, _eden_chunk_capacity, mtGC); 718 if (_eden_chunk_array == NULL) { 719 _eden_chunk_capacity = 0; 720 warning("GC/CMS: _eden_chunk_array allocation failure"); 721 } 722 } 723 assert(_eden_chunk_array != NULL || _eden_chunk_capacity == 0, "Error"); 724 725 // Support for parallelizing survivor space rescan 726 if (CMSParallelRemarkEnabled && CMSParallelSurvivorRemarkEnabled) { 727 const size_t max_plab_samples = 728 ((DefNewGeneration*)_young_gen)->max_survivor_size()/MinTLABSize; 729 730 _survivor_plab_array = NEW_C_HEAP_ARRAY(ChunkArray, ParallelGCThreads, mtGC); 731 _survivor_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, 2*max_plab_samples, mtGC); 732 _cursor = NEW_C_HEAP_ARRAY(size_t, ParallelGCThreads, mtGC); 733 if (_survivor_plab_array == NULL || _survivor_chunk_array == NULL 734 || _cursor == NULL) { 735 warning("Failed to allocate survivor plab/chunk array"); 736 if (_survivor_plab_array != NULL) { 737 FREE_C_HEAP_ARRAY(ChunkArray, _survivor_plab_array, mtGC); 738 _survivor_plab_array = NULL; 739 } 740 if (_survivor_chunk_array != NULL) { 741 FREE_C_HEAP_ARRAY(HeapWord*, _survivor_chunk_array, mtGC); 742 _survivor_chunk_array = NULL; 743 } 744 if (_cursor != NULL) { 745 FREE_C_HEAP_ARRAY(size_t, _cursor, mtGC); 746 _cursor = NULL; 747 } 748 } else { 749 _survivor_chunk_capacity = 2*max_plab_samples; 750 for (uint i = 0; i < ParallelGCThreads; i++) { 751 HeapWord** vec = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples, mtGC); 752 if (vec == NULL) { 753 warning("Failed to allocate survivor plab array"); 754 for (int j = i; j > 0; j--) { 755 FREE_C_HEAP_ARRAY(HeapWord*, _survivor_plab_array[j-1].array(), mtGC); 756 } 757 FREE_C_HEAP_ARRAY(ChunkArray, _survivor_plab_array, mtGC); 758 FREE_C_HEAP_ARRAY(HeapWord*, _survivor_chunk_array, mtGC); 759 _survivor_plab_array = NULL; 760 _survivor_chunk_array = NULL; 761 _survivor_chunk_capacity = 0; 762 break; 763 } else { 764 ChunkArray* cur = 765 ::new (&_survivor_plab_array[i]) ChunkArray(vec, 766 max_plab_samples); 767 assert(cur->end() == 0, "Should be 0"); 768 assert(cur->array() == vec, "Should be vec"); 769 assert(cur->capacity() == max_plab_samples, "Error"); 770 } 771 } 772 } 773 } 774 assert( ( _survivor_plab_array != NULL 775 && _survivor_chunk_array != NULL) 776 || ( _survivor_chunk_capacity == 0 777 && _survivor_chunk_index == 0), 778 "Error"); 779 780 // Choose what strong roots should be scanned depending on verification options 781 if (!CMSClassUnloadingEnabled) { 782 // If class unloading is disabled we want to include all classes into the root set. 783 add_root_scanning_option(SharedHeap::SO_AllClasses); 784 } else { 785 add_root_scanning_option(SharedHeap::SO_SystemClasses); 786 } 787 788 NOT_PRODUCT(_overflow_counter = CMSMarkStackOverflowInterval;) 789 _gc_counters = new CollectorCounters("CMS", 1); 790 _completed_initialization = true; 791 _inter_sweep_timer.start(); // start of time 792 #ifdef SPARC 793 // Issue a stern warning, but allow use for experimentation and debugging. 794 if (VM_Version::is_sun4v() && UseMemSetInBOT) { 795 assert(!FLAG_IS_DEFAULT(UseMemSetInBOT), "Error"); 796 warning("Experimental flag -XX:+UseMemSetInBOT is known to cause instability" 797 " on sun4v; please understand that you are using at your own risk!"); 798 } 799 #endif 800 } 801 802 const char* ConcurrentMarkSweepGeneration::name() const { 803 return "concurrent mark-sweep generation"; 804 } 805 void ConcurrentMarkSweepGeneration::update_counters() { 806 if (UsePerfData) { 807 _space_counters->update_all(); 808 _gen_counters->update_all(); 809 } 810 } 811 812 // this is an optimized version of update_counters(). it takes the 813 // used value as a parameter rather than computing it. 814 // 815 void ConcurrentMarkSweepGeneration::update_counters(size_t used) { 816 if (UsePerfData) { 817 _space_counters->update_used(used); 818 _space_counters->update_capacity(); 819 _gen_counters->update_all(); 820 } 821 } 822 823 void ConcurrentMarkSweepGeneration::print() const { 824 Generation::print(); 825 cmsSpace()->print(); 826 } 827 828 #ifndef PRODUCT 829 void ConcurrentMarkSweepGeneration::print_statistics() { 830 cmsSpace()->printFLCensus(0); 831 } 832 #endif 833 834 void ConcurrentMarkSweepGeneration::printOccupancy(const char *s) { 835 GenCollectedHeap* gch = GenCollectedHeap::heap(); 836 if (PrintGCDetails) { 837 if (Verbose) { 838 gclog_or_tty->print(" [%d %s-%s: "SIZE_FORMAT"("SIZE_FORMAT")]", 839 level(), short_name(), s, used(), capacity()); 840 } else { 841 gclog_or_tty->print(" [%d %s-%s: "SIZE_FORMAT"K("SIZE_FORMAT"K)]", 842 level(), short_name(), s, used() / K, capacity() / K); 843 } 844 } 845 if (Verbose) { 846 gclog_or_tty->print(" "SIZE_FORMAT"("SIZE_FORMAT")", 847 gch->used(), gch->capacity()); 848 } else { 849 gclog_or_tty->print(" "SIZE_FORMAT"K("SIZE_FORMAT"K)", 850 gch->used() / K, gch->capacity() / K); 851 } 852 } 853 854 size_t 855 ConcurrentMarkSweepGeneration::contiguous_available() const { 856 // dld proposes an improvement in precision here. If the committed 857 // part of the space ends in a free block we should add that to 858 // uncommitted size in the calculation below. Will make this 859 // change later, staying with the approximation below for the 860 // time being. -- ysr. 861 return MAX2(_virtual_space.uncommitted_size(), unsafe_max_alloc_nogc()); 862 } 863 864 size_t 865 ConcurrentMarkSweepGeneration::unsafe_max_alloc_nogc() const { 866 return _cmsSpace->max_alloc_in_words() * HeapWordSize; 867 } 868 869 size_t ConcurrentMarkSweepGeneration::max_available() const { 870 return free() + _virtual_space.uncommitted_size(); 871 } 872 873 bool ConcurrentMarkSweepGeneration::promotion_attempt_is_safe(size_t max_promotion_in_bytes) const { 874 size_t available = max_available(); 875 size_t av_promo = (size_t)gc_stats()->avg_promoted()->padded_average(); 876 bool res = (available >= av_promo) || (available >= max_promotion_in_bytes); 877 if (Verbose && PrintGCDetails) { 878 gclog_or_tty->print_cr( 879 "CMS: promo attempt is%s safe: available("SIZE_FORMAT") %s av_promo("SIZE_FORMAT")," 880 "max_promo("SIZE_FORMAT")", 881 res? "":" not", available, res? ">=":"<", 882 av_promo, max_promotion_in_bytes); 883 } 884 return res; 885 } 886 887 // At a promotion failure dump information on block layout in heap 888 // (cms old generation). 889 void ConcurrentMarkSweepGeneration::promotion_failure_occurred() { 890 if (CMSDumpAtPromotionFailure) { 891 cmsSpace()->dump_at_safepoint_with_locks(collector(), gclog_or_tty); 892 } 893 } 894 895 CompactibleSpace* 896 ConcurrentMarkSweepGeneration::first_compaction_space() const { 897 return _cmsSpace; 898 } 899 900 void ConcurrentMarkSweepGeneration::reset_after_compaction() { 901 // Clear the promotion information. These pointers can be adjusted 902 // along with all the other pointers into the heap but 903 // compaction is expected to be a rare event with 904 // a heap using cms so don't do it without seeing the need. 905 if (CollectedHeap::use_parallel_gc_threads()) { 906 for (uint i = 0; i < ParallelGCThreads; i++) { 907 _par_gc_thread_states[i]->promo.reset(); 908 } 909 } 910 } 911 912 void ConcurrentMarkSweepGeneration::space_iterate(SpaceClosure* blk, bool usedOnly) { 913 blk->do_space(_cmsSpace); 914 } 915 916 void ConcurrentMarkSweepGeneration::compute_new_size() { 917 assert_locked_or_safepoint(Heap_lock); 918 919 // If incremental collection failed, we just want to expand 920 // to the limit. 921 if (incremental_collection_failed()) { 922 clear_incremental_collection_failed(); 923 grow_to_reserved(); 924 return; 925 } 926 927 size_t expand_bytes = 0; 928 double free_percentage = ((double) free()) / capacity(); 929 double desired_free_percentage = (double) MinHeapFreeRatio / 100; 930 double maximum_free_percentage = (double) MaxHeapFreeRatio / 100; 931 932 // compute expansion delta needed for reaching desired free percentage 933 if (free_percentage < desired_free_percentage) { 934 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 935 assert(desired_capacity >= capacity(), "invalid expansion size"); 936 expand_bytes = MAX2(desired_capacity - capacity(), MinHeapDeltaBytes); 937 } 938 if (expand_bytes > 0) { 939 if (PrintGCDetails && Verbose) { 940 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 941 gclog_or_tty->print_cr("\nFrom compute_new_size: "); 942 gclog_or_tty->print_cr(" Free fraction %f", free_percentage); 943 gclog_or_tty->print_cr(" Desired free fraction %f", 944 desired_free_percentage); 945 gclog_or_tty->print_cr(" Maximum free fraction %f", 946 maximum_free_percentage); 947 gclog_or_tty->print_cr(" Capactiy "SIZE_FORMAT, capacity()/1000); 948 gclog_or_tty->print_cr(" Desired capacity "SIZE_FORMAT, 949 desired_capacity/1000); 950 int prev_level = level() - 1; 951 if (prev_level >= 0) { 952 size_t prev_size = 0; 953 GenCollectedHeap* gch = GenCollectedHeap::heap(); 954 Generation* prev_gen = gch->_gens[prev_level]; 955 prev_size = prev_gen->capacity(); 956 gclog_or_tty->print_cr(" Younger gen size "SIZE_FORMAT, 957 prev_size/1000); 958 } 959 gclog_or_tty->print_cr(" unsafe_max_alloc_nogc "SIZE_FORMAT, 960 unsafe_max_alloc_nogc()/1000); 961 gclog_or_tty->print_cr(" contiguous available "SIZE_FORMAT, 962 contiguous_available()/1000); 963 gclog_or_tty->print_cr(" Expand by "SIZE_FORMAT" (bytes)", 964 expand_bytes); 965 } 966 // safe if expansion fails 967 expand(expand_bytes, 0, CMSExpansionCause::_satisfy_free_ratio); 968 if (PrintGCDetails && Verbose) { 969 gclog_or_tty->print_cr(" Expanded free fraction %f", 970 ((double) free()) / capacity()); 971 } 972 } 973 } 974 975 Mutex* ConcurrentMarkSweepGeneration::freelistLock() const { 976 return cmsSpace()->freelistLock(); 977 } 978 979 HeapWord* ConcurrentMarkSweepGeneration::allocate(size_t size, 980 bool tlab) { 981 CMSSynchronousYieldRequest yr; 982 MutexLockerEx x(freelistLock(), 983 Mutex::_no_safepoint_check_flag); 984 return have_lock_and_allocate(size, tlab); 985 } 986 987 HeapWord* ConcurrentMarkSweepGeneration::have_lock_and_allocate(size_t size, 988 bool tlab /* ignored */) { 989 assert_lock_strong(freelistLock()); 990 size_t adjustedSize = CompactibleFreeListSpace::adjustObjectSize(size); 991 HeapWord* res = cmsSpace()->allocate(adjustedSize); 992 // Allocate the object live (grey) if the background collector has 993 // started marking. This is necessary because the marker may 994 // have passed this address and consequently this object will 995 // not otherwise be greyed and would be incorrectly swept up. 996 // Note that if this object contains references, the writing 997 // of those references will dirty the card containing this object 998 // allowing the object to be blackened (and its references scanned) 999 // either during a preclean phase or at the final checkpoint. 1000 if (res != NULL) { 1001 // We may block here with an uninitialized object with 1002 // its mark-bit or P-bits not yet set. Such objects need 1003 // to be safely navigable by block_start(). 1004 assert(oop(res)->klass_or_null() == NULL, "Object should be uninitialized here."); 1005 assert(!((FreeChunk*)res)->is_free(), "Error, block will look free but show wrong size"); 1006 collector()->direct_allocated(res, adjustedSize); 1007 _direct_allocated_words += adjustedSize; 1008 // allocation counters 1009 NOT_PRODUCT( 1010 _numObjectsAllocated++; 1011 _numWordsAllocated += (int)adjustedSize; 1012 ) 1013 } 1014 return res; 1015 } 1016 1017 // In the case of direct allocation by mutators in a generation that 1018 // is being concurrently collected, the object must be allocated 1019 // live (grey) if the background collector has started marking. 1020 // This is necessary because the marker may 1021 // have passed this address and consequently this object will 1022 // not otherwise be greyed and would be incorrectly swept up. 1023 // Note that if this object contains references, the writing 1024 // of those references will dirty the card containing this object 1025 // allowing the object to be blackened (and its references scanned) 1026 // either during a preclean phase or at the final checkpoint. 1027 void CMSCollector::direct_allocated(HeapWord* start, size_t size) { 1028 assert(_markBitMap.covers(start, size), "Out of bounds"); 1029 if (_collectorState >= Marking) { 1030 MutexLockerEx y(_markBitMap.lock(), 1031 Mutex::_no_safepoint_check_flag); 1032 // [see comments preceding SweepClosure::do_blk() below for details] 1033 // 1034 // Can the P-bits be deleted now? JJJ 1035 // 1036 // 1. need to mark the object as live so it isn't collected 1037 // 2. need to mark the 2nd bit to indicate the object may be uninitialized 1038 // 3. need to mark the end of the object so marking, precleaning or sweeping 1039 // can skip over uninitialized or unparsable objects. An allocated 1040 // object is considered uninitialized for our purposes as long as 1041 // its klass word is NULL. All old gen objects are parsable 1042 // as soon as they are initialized.) 1043 _markBitMap.mark(start); // object is live 1044 _markBitMap.mark(start + 1); // object is potentially uninitialized? 1045 _markBitMap.mark(start + size - 1); 1046 // mark end of object 1047 } 1048 // check that oop looks uninitialized 1049 assert(oop(start)->klass_or_null() == NULL, "_klass should be NULL"); 1050 } 1051 1052 void CMSCollector::promoted(bool par, HeapWord* start, 1053 bool is_obj_array, size_t obj_size) { 1054 assert(_markBitMap.covers(start), "Out of bounds"); 1055 // See comment in direct_allocated() about when objects should 1056 // be allocated live. 1057 if (_collectorState >= Marking) { 1058 // we already hold the marking bit map lock, taken in 1059 // the prologue 1060 if (par) { 1061 _markBitMap.par_mark(start); 1062 } else { 1063 _markBitMap.mark(start); 1064 } 1065 // We don't need to mark the object as uninitialized (as 1066 // in direct_allocated above) because this is being done with the 1067 // world stopped and the object will be initialized by the 1068 // time the marking, precleaning or sweeping get to look at it. 1069 // But see the code for copying objects into the CMS generation, 1070 // where we need to ensure that concurrent readers of the 1071 // block offset table are able to safely navigate a block that 1072 // is in flux from being free to being allocated (and in 1073 // transition while being copied into) and subsequently 1074 // becoming a bona-fide object when the copy/promotion is complete. 1075 assert(SafepointSynchronize::is_at_safepoint(), 1076 "expect promotion only at safepoints"); 1077 1078 if (_collectorState < Sweeping) { 1079 // Mark the appropriate cards in the modUnionTable, so that 1080 // this object gets scanned before the sweep. If this is 1081 // not done, CMS generation references in the object might 1082 // not get marked. 1083 // For the case of arrays, which are otherwise precisely 1084 // marked, we need to dirty the entire array, not just its head. 1085 if (is_obj_array) { 1086 // The [par_]mark_range() method expects mr.end() below to 1087 // be aligned to the granularity of a bit's representation 1088 // in the heap. In the case of the MUT below, that's a 1089 // card size. 1090 MemRegion mr(start, 1091 (HeapWord*)round_to((intptr_t)(start + obj_size), 1092 CardTableModRefBS::card_size /* bytes */)); 1093 if (par) { 1094 _modUnionTable.par_mark_range(mr); 1095 } else { 1096 _modUnionTable.mark_range(mr); 1097 } 1098 } else { // not an obj array; we can just mark the head 1099 if (par) { 1100 _modUnionTable.par_mark(start); 1101 } else { 1102 _modUnionTable.mark(start); 1103 } 1104 } 1105 } 1106 } 1107 } 1108 1109 static inline size_t percent_of_space(Space* space, HeapWord* addr) 1110 { 1111 size_t delta = pointer_delta(addr, space->bottom()); 1112 return (size_t)(delta * 100.0 / (space->capacity() / HeapWordSize)); 1113 } 1114 1115 void CMSCollector::icms_update_allocation_limits() 1116 { 1117 Generation* gen0 = GenCollectedHeap::heap()->get_gen(0); 1118 EdenSpace* eden = gen0->as_DefNewGeneration()->eden(); 1119 1120 const unsigned int duty_cycle = stats().icms_update_duty_cycle(); 1121 if (CMSTraceIncrementalPacing) { 1122 stats().print(); 1123 } 1124 1125 assert(duty_cycle <= 100, "invalid duty cycle"); 1126 if (duty_cycle != 0) { 1127 // The duty_cycle is a percentage between 0 and 100; convert to words and 1128 // then compute the offset from the endpoints of the space. 1129 size_t free_words = eden->free() / HeapWordSize; 1130 double free_words_dbl = (double)free_words; 1131 size_t duty_cycle_words = (size_t)(free_words_dbl * duty_cycle / 100.0); 1132 size_t offset_words = (free_words - duty_cycle_words) / 2; 1133 1134 _icms_start_limit = eden->top() + offset_words; 1135 _icms_stop_limit = eden->end() - offset_words; 1136 1137 // The limits may be adjusted (shifted to the right) by 1138 // CMSIncrementalOffset, to allow the application more mutator time after a 1139 // young gen gc (when all mutators were stopped) and before CMS starts and 1140 // takes away one or more cpus. 1141 if (CMSIncrementalOffset != 0) { 1142 double adjustment_dbl = free_words_dbl * CMSIncrementalOffset / 100.0; 1143 size_t adjustment = (size_t)adjustment_dbl; 1144 HeapWord* tmp_stop = _icms_stop_limit + adjustment; 1145 if (tmp_stop > _icms_stop_limit && tmp_stop < eden->end()) { 1146 _icms_start_limit += adjustment; 1147 _icms_stop_limit = tmp_stop; 1148 } 1149 } 1150 } 1151 if (duty_cycle == 0 || (_icms_start_limit == _icms_stop_limit)) { 1152 _icms_start_limit = _icms_stop_limit = eden->end(); 1153 } 1154 1155 // Install the new start limit. 1156 eden->set_soft_end(_icms_start_limit); 1157 1158 if (CMSTraceIncrementalMode) { 1159 gclog_or_tty->print(" icms alloc limits: " 1160 PTR_FORMAT "," PTR_FORMAT 1161 " (" SIZE_FORMAT "%%," SIZE_FORMAT "%%) ", 1162 _icms_start_limit, _icms_stop_limit, 1163 percent_of_space(eden, _icms_start_limit), 1164 percent_of_space(eden, _icms_stop_limit)); 1165 if (Verbose) { 1166 gclog_or_tty->print("eden: "); 1167 eden->print_on(gclog_or_tty); 1168 } 1169 } 1170 } 1171 1172 // Any changes here should try to maintain the invariant 1173 // that if this method is called with _icms_start_limit 1174 // and _icms_stop_limit both NULL, then it should return NULL 1175 // and not notify the icms thread. 1176 HeapWord* 1177 CMSCollector::allocation_limit_reached(Space* space, HeapWord* top, 1178 size_t word_size) 1179 { 1180 // A start_limit equal to end() means the duty cycle is 0, so treat that as a 1181 // nop. 1182 if (CMSIncrementalMode && _icms_start_limit != space->end()) { 1183 if (top <= _icms_start_limit) { 1184 if (CMSTraceIncrementalMode) { 1185 space->print_on(gclog_or_tty); 1186 gclog_or_tty->stamp(); 1187 gclog_or_tty->print_cr(" start limit top=" PTR_FORMAT 1188 ", new limit=" PTR_FORMAT 1189 " (" SIZE_FORMAT "%%)", 1190 top, _icms_stop_limit, 1191 percent_of_space(space, _icms_stop_limit)); 1192 } 1193 ConcurrentMarkSweepThread::start_icms(); 1194 assert(top < _icms_stop_limit, "Tautology"); 1195 if (word_size < pointer_delta(_icms_stop_limit, top)) { 1196 return _icms_stop_limit; 1197 } 1198 1199 // The allocation will cross both the _start and _stop limits, so do the 1200 // stop notification also and return end(). 1201 if (CMSTraceIncrementalMode) { 1202 space->print_on(gclog_or_tty); 1203 gclog_or_tty->stamp(); 1204 gclog_or_tty->print_cr(" +stop limit top=" PTR_FORMAT 1205 ", new limit=" PTR_FORMAT 1206 " (" SIZE_FORMAT "%%)", 1207 top, space->end(), 1208 percent_of_space(space, space->end())); 1209 } 1210 ConcurrentMarkSweepThread::stop_icms(); 1211 return space->end(); 1212 } 1213 1214 if (top <= _icms_stop_limit) { 1215 if (CMSTraceIncrementalMode) { 1216 space->print_on(gclog_or_tty); 1217 gclog_or_tty->stamp(); 1218 gclog_or_tty->print_cr(" stop limit top=" PTR_FORMAT 1219 ", new limit=" PTR_FORMAT 1220 " (" SIZE_FORMAT "%%)", 1221 top, space->end(), 1222 percent_of_space(space, space->end())); 1223 } 1224 ConcurrentMarkSweepThread::stop_icms(); 1225 return space->end(); 1226 } 1227 1228 if (CMSTraceIncrementalMode) { 1229 space->print_on(gclog_or_tty); 1230 gclog_or_tty->stamp(); 1231 gclog_or_tty->print_cr(" end limit top=" PTR_FORMAT 1232 ", new limit=" PTR_FORMAT, 1233 top, NULL); 1234 } 1235 } 1236 1237 return NULL; 1238 } 1239 1240 oop ConcurrentMarkSweepGeneration::promote(oop obj, size_t obj_size) { 1241 assert(obj_size == (size_t)obj->size(), "bad obj_size passed in"); 1242 // allocate, copy and if necessary update promoinfo -- 1243 // delegate to underlying space. 1244 assert_lock_strong(freelistLock()); 1245 1246 #ifndef PRODUCT 1247 if (Universe::heap()->promotion_should_fail()) { 1248 return NULL; 1249 } 1250 #endif // #ifndef PRODUCT 1251 1252 oop res = _cmsSpace->promote(obj, obj_size); 1253 if (res == NULL) { 1254 // expand and retry 1255 size_t s = _cmsSpace->expansionSpaceRequired(obj_size); // HeapWords 1256 expand(s*HeapWordSize, MinHeapDeltaBytes, 1257 CMSExpansionCause::_satisfy_promotion); 1258 // Since there's currently no next generation, we don't try to promote 1259 // into a more senior generation. 1260 assert(next_gen() == NULL, "assumption, based upon which no attempt " 1261 "is made to pass on a possibly failing " 1262 "promotion to next generation"); 1263 res = _cmsSpace->promote(obj, obj_size); 1264 } 1265 if (res != NULL) { 1266 // See comment in allocate() about when objects should 1267 // be allocated live. 1268 assert(obj->is_oop(), "Will dereference klass pointer below"); 1269 collector()->promoted(false, // Not parallel 1270 (HeapWord*)res, obj->is_objArray(), obj_size); 1271 // promotion counters 1272 NOT_PRODUCT( 1273 _numObjectsPromoted++; 1274 _numWordsPromoted += 1275 (int)(CompactibleFreeListSpace::adjustObjectSize(obj->size())); 1276 ) 1277 } 1278 return res; 1279 } 1280 1281 1282 HeapWord* 1283 ConcurrentMarkSweepGeneration::allocation_limit_reached(Space* space, 1284 HeapWord* top, 1285 size_t word_sz) 1286 { 1287 return collector()->allocation_limit_reached(space, top, word_sz); 1288 } 1289 1290 // IMPORTANT: Notes on object size recognition in CMS. 1291 // --------------------------------------------------- 1292 // A block of storage in the CMS generation is always in 1293 // one of three states. A free block (FREE), an allocated 1294 // object (OBJECT) whose size() method reports the correct size, 1295 // and an intermediate state (TRANSIENT) in which its size cannot 1296 // be accurately determined. 1297 // STATE IDENTIFICATION: (32 bit and 64 bit w/o COOPS) 1298 // ----------------------------------------------------- 1299 // FREE: klass_word & 1 == 1; mark_word holds block size 1300 // 1301 // OBJECT: klass_word installed; klass_word != 0 && klass_word & 1 == 0; 1302 // obj->size() computes correct size 1303 // 1304 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT 1305 // 1306 // STATE IDENTIFICATION: (64 bit+COOPS) 1307 // ------------------------------------ 1308 // FREE: mark_word & CMS_FREE_BIT == 1; mark_word & ~CMS_FREE_BIT gives block_size 1309 // 1310 // OBJECT: klass_word installed; klass_word != 0; 1311 // obj->size() computes correct size 1312 // 1313 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT 1314 // 1315 // 1316 // STATE TRANSITION DIAGRAM 1317 // 1318 // mut / parnew mut / parnew 1319 // FREE --------------------> TRANSIENT ---------------------> OBJECT --| 1320 // ^ | 1321 // |------------------------ DEAD <------------------------------------| 1322 // sweep mut 1323 // 1324 // While a block is in TRANSIENT state its size cannot be determined 1325 // so readers will either need to come back later or stall until 1326 // the size can be determined. Note that for the case of direct 1327 // allocation, P-bits, when available, may be used to determine the 1328 // size of an object that may not yet have been initialized. 1329 1330 // Things to support parallel young-gen collection. 1331 oop 1332 ConcurrentMarkSweepGeneration::par_promote(int thread_num, 1333 oop old, markOop m, 1334 size_t word_sz) { 1335 #ifndef PRODUCT 1336 if (Universe::heap()->promotion_should_fail()) { 1337 return NULL; 1338 } 1339 #endif // #ifndef PRODUCT 1340 1341 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1342 PromotionInfo* promoInfo = &ps->promo; 1343 // if we are tracking promotions, then first ensure space for 1344 // promotion (including spooling space for saving header if necessary). 1345 // then allocate and copy, then track promoted info if needed. 1346 // When tracking (see PromotionInfo::track()), the mark word may 1347 // be displaced and in this case restoration of the mark word 1348 // occurs in the (oop_since_save_marks_)iterate phase. 1349 if (promoInfo->tracking() && !promoInfo->ensure_spooling_space()) { 1350 // Out of space for allocating spooling buffers; 1351 // try expanding and allocating spooling buffers. 1352 if (!expand_and_ensure_spooling_space(promoInfo)) { 1353 return NULL; 1354 } 1355 } 1356 assert(promoInfo->has_spooling_space(), "Control point invariant"); 1357 const size_t alloc_sz = CompactibleFreeListSpace::adjustObjectSize(word_sz); 1358 HeapWord* obj_ptr = ps->lab.alloc(alloc_sz); 1359 if (obj_ptr == NULL) { 1360 obj_ptr = expand_and_par_lab_allocate(ps, alloc_sz); 1361 if (obj_ptr == NULL) { 1362 return NULL; 1363 } 1364 } 1365 oop obj = oop(obj_ptr); 1366 OrderAccess::storestore(); 1367 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1368 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1369 // IMPORTANT: See note on object initialization for CMS above. 1370 // Otherwise, copy the object. Here we must be careful to insert the 1371 // klass pointer last, since this marks the block as an allocated object. 1372 // Except with compressed oops it's the mark word. 1373 HeapWord* old_ptr = (HeapWord*)old; 1374 // Restore the mark word copied above. 1375 obj->set_mark(m); 1376 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1377 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1378 OrderAccess::storestore(); 1379 1380 if (UseCompressedKlassPointers) { 1381 // Copy gap missed by (aligned) header size calculation below 1382 obj->set_klass_gap(old->klass_gap()); 1383 } 1384 if (word_sz > (size_t)oopDesc::header_size()) { 1385 Copy::aligned_disjoint_words(old_ptr + oopDesc::header_size(), 1386 obj_ptr + oopDesc::header_size(), 1387 word_sz - oopDesc::header_size()); 1388 } 1389 1390 // Now we can track the promoted object, if necessary. We take care 1391 // to delay the transition from uninitialized to full object 1392 // (i.e., insertion of klass pointer) until after, so that it 1393 // atomically becomes a promoted object. 1394 if (promoInfo->tracking()) { 1395 promoInfo->track((PromotedObject*)obj, old->klass()); 1396 } 1397 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1398 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1399 assert(old->is_oop(), "Will use and dereference old klass ptr below"); 1400 1401 // Finally, install the klass pointer (this should be volatile). 1402 OrderAccess::storestore(); 1403 obj->set_klass(old->klass()); 1404 // We should now be able to calculate the right size for this object 1405 assert(obj->is_oop() && obj->size() == (int)word_sz, "Error, incorrect size computed for promoted object"); 1406 1407 collector()->promoted(true, // parallel 1408 obj_ptr, old->is_objArray(), word_sz); 1409 1410 NOT_PRODUCT( 1411 Atomic::inc_ptr(&_numObjectsPromoted); 1412 Atomic::add_ptr(alloc_sz, &_numWordsPromoted); 1413 ) 1414 1415 return obj; 1416 } 1417 1418 void 1419 ConcurrentMarkSweepGeneration:: 1420 par_promote_alloc_undo(int thread_num, 1421 HeapWord* obj, size_t word_sz) { 1422 // CMS does not support promotion undo. 1423 ShouldNotReachHere(); 1424 } 1425 1426 void 1427 ConcurrentMarkSweepGeneration:: 1428 par_promote_alloc_done(int thread_num) { 1429 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1430 ps->lab.retire(thread_num); 1431 } 1432 1433 void 1434 ConcurrentMarkSweepGeneration:: 1435 par_oop_since_save_marks_iterate_done(int thread_num) { 1436 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1437 ParScanWithoutBarrierClosure* dummy_cl = NULL; 1438 ps->promo.promoted_oops_iterate_nv(dummy_cl); 1439 } 1440 1441 bool ConcurrentMarkSweepGeneration::should_collect(bool full, 1442 size_t size, 1443 bool tlab) 1444 { 1445 // We allow a STW collection only if a full 1446 // collection was requested. 1447 return full || should_allocate(size, tlab); // FIX ME !!! 1448 // This and promotion failure handling are connected at the 1449 // hip and should be fixed by untying them. 1450 } 1451 1452 bool CMSCollector::shouldConcurrentCollect() { 1453 if (_full_gc_requested) { 1454 if (Verbose && PrintGCDetails) { 1455 gclog_or_tty->print_cr("CMSCollector: collect because of explicit " 1456 " gc request (or gc_locker)"); 1457 } 1458 return true; 1459 } 1460 1461 // For debugging purposes, change the type of collection. 1462 // If the rotation is not on the concurrent collection 1463 // type, don't start a concurrent collection. 1464 NOT_PRODUCT( 1465 if (RotateCMSCollectionTypes && 1466 (_cmsGen->debug_collection_type() != 1467 ConcurrentMarkSweepGeneration::Concurrent_collection_type)) { 1468 assert(_cmsGen->debug_collection_type() != 1469 ConcurrentMarkSweepGeneration::Unknown_collection_type, 1470 "Bad cms collection type"); 1471 return false; 1472 } 1473 ) 1474 1475 FreelistLocker x(this); 1476 // ------------------------------------------------------------------ 1477 // Print out lots of information which affects the initiation of 1478 // a collection. 1479 if (PrintCMSInitiationStatistics && stats().valid()) { 1480 gclog_or_tty->print("CMSCollector shouldConcurrentCollect: "); 1481 gclog_or_tty->stamp(); 1482 gclog_or_tty->print_cr(""); 1483 stats().print_on(gclog_or_tty); 1484 gclog_or_tty->print_cr("time_until_cms_gen_full %3.7f", 1485 stats().time_until_cms_gen_full()); 1486 gclog_or_tty->print_cr("free="SIZE_FORMAT, _cmsGen->free()); 1487 gclog_or_tty->print_cr("contiguous_available="SIZE_FORMAT, 1488 _cmsGen->contiguous_available()); 1489 gclog_or_tty->print_cr("promotion_rate=%g", stats().promotion_rate()); 1490 gclog_or_tty->print_cr("cms_allocation_rate=%g", stats().cms_allocation_rate()); 1491 gclog_or_tty->print_cr("occupancy=%3.7f", _cmsGen->occupancy()); 1492 gclog_or_tty->print_cr("initiatingOccupancy=%3.7f", _cmsGen->initiating_occupancy()); 1493 gclog_or_tty->print_cr("metadata initialized %d", 1494 MetaspaceGC::should_concurrent_collect()); 1495 } 1496 // ------------------------------------------------------------------ 1497 1498 // If the estimated time to complete a cms collection (cms_duration()) 1499 // is less than the estimated time remaining until the cms generation 1500 // is full, start a collection. 1501 if (!UseCMSInitiatingOccupancyOnly) { 1502 if (stats().valid()) { 1503 if (stats().time_until_cms_start() == 0.0) { 1504 return true; 1505 } 1506 } else { 1507 // We want to conservatively collect somewhat early in order 1508 // to try and "bootstrap" our CMS/promotion statistics; 1509 // this branch will not fire after the first successful CMS 1510 // collection because the stats should then be valid. 1511 if (_cmsGen->occupancy() >= _bootstrap_occupancy) { 1512 if (Verbose && PrintGCDetails) { 1513 gclog_or_tty->print_cr( 1514 " CMSCollector: collect for bootstrapping statistics:" 1515 " occupancy = %f, boot occupancy = %f", _cmsGen->occupancy(), 1516 _bootstrap_occupancy); 1517 } 1518 return true; 1519 } 1520 } 1521 } 1522 1523 // Otherwise, we start a collection cycle if 1524 // old gen want a collection cycle started. Each may use 1525 // an appropriate criterion for making this decision. 1526 // XXX We need to make sure that the gen expansion 1527 // criterion dovetails well with this. XXX NEED TO FIX THIS 1528 if (_cmsGen->should_concurrent_collect()) { 1529 if (Verbose && PrintGCDetails) { 1530 gclog_or_tty->print_cr("CMS old gen initiated"); 1531 } 1532 return true; 1533 } 1534 1535 // We start a collection if we believe an incremental collection may fail; 1536 // this is not likely to be productive in practice because it's probably too 1537 // late anyway. 1538 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1539 assert(gch->collector_policy()->is_two_generation_policy(), 1540 "You may want to check the correctness of the following"); 1541 if (gch->incremental_collection_will_fail(true /* consult_young */)) { 1542 if (Verbose && PrintGCDetails) { 1543 gclog_or_tty->print("CMSCollector: collect because incremental collection will fail "); 1544 } 1545 return true; 1546 } 1547 1548 if (MetaspaceGC::should_concurrent_collect()) { 1549 if (Verbose && PrintGCDetails) { 1550 gclog_or_tty->print("CMSCollector: collect for metadata allocation "); 1551 } 1552 return true; 1553 } 1554 1555 return false; 1556 } 1557 1558 // Clear _expansion_cause fields of constituent generations 1559 void CMSCollector::clear_expansion_cause() { 1560 _cmsGen->clear_expansion_cause(); 1561 } 1562 1563 // We should be conservative in starting a collection cycle. To 1564 // start too eagerly runs the risk of collecting too often in the 1565 // extreme. To collect too rarely falls back on full collections, 1566 // which works, even if not optimum in terms of concurrent work. 1567 // As a work around for too eagerly collecting, use the flag 1568 // UseCMSInitiatingOccupancyOnly. This also has the advantage of 1569 // giving the user an easily understandable way of controlling the 1570 // collections. 1571 // We want to start a new collection cycle if any of the following 1572 // conditions hold: 1573 // . our current occupancy exceeds the configured initiating occupancy 1574 // for this generation, or 1575 // . we recently needed to expand this space and have not, since that 1576 // expansion, done a collection of this generation, or 1577 // . the underlying space believes that it may be a good idea to initiate 1578 // a concurrent collection (this may be based on criteria such as the 1579 // following: the space uses linear allocation and linear allocation is 1580 // going to fail, or there is believed to be excessive fragmentation in 1581 // the generation, etc... or ... 1582 // [.(currently done by CMSCollector::shouldConcurrentCollect() only for 1583 // the case of the old generation; see CR 6543076): 1584 // we may be approaching a point at which allocation requests may fail because 1585 // we will be out of sufficient free space given allocation rate estimates.] 1586 bool ConcurrentMarkSweepGeneration::should_concurrent_collect() const { 1587 1588 assert_lock_strong(freelistLock()); 1589 if (occupancy() > initiating_occupancy()) { 1590 if (PrintGCDetails && Verbose) { 1591 gclog_or_tty->print(" %s: collect because of occupancy %f / %f ", 1592 short_name(), occupancy(), initiating_occupancy()); 1593 } 1594 return true; 1595 } 1596 if (UseCMSInitiatingOccupancyOnly) { 1597 return false; 1598 } 1599 if (expansion_cause() == CMSExpansionCause::_satisfy_allocation) { 1600 if (PrintGCDetails && Verbose) { 1601 gclog_or_tty->print(" %s: collect because expanded for allocation ", 1602 short_name()); 1603 } 1604 return true; 1605 } 1606 if (_cmsSpace->should_concurrent_collect()) { 1607 if (PrintGCDetails && Verbose) { 1608 gclog_or_tty->print(" %s: collect because cmsSpace says so ", 1609 short_name()); 1610 } 1611 return true; 1612 } 1613 return false; 1614 } 1615 1616 void ConcurrentMarkSweepGeneration::collect(bool full, 1617 bool clear_all_soft_refs, 1618 size_t size, 1619 bool tlab) 1620 { 1621 collector()->collect(full, clear_all_soft_refs, size, tlab); 1622 } 1623 1624 void CMSCollector::collect(bool full, 1625 bool clear_all_soft_refs, 1626 size_t size, 1627 bool tlab) 1628 { 1629 if (!UseCMSCollectionPassing && _collectorState > Idling) { 1630 // For debugging purposes skip the collection if the state 1631 // is not currently idle 1632 if (TraceCMSState) { 1633 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " skipped full:%d CMS state %d", 1634 Thread::current(), full, _collectorState); 1635 } 1636 return; 1637 } 1638 1639 // The following "if" branch is present for defensive reasons. 1640 // In the current uses of this interface, it can be replaced with: 1641 // assert(!GC_locker.is_active(), "Can't be called otherwise"); 1642 // But I am not placing that assert here to allow future 1643 // generality in invoking this interface. 1644 if (GC_locker::is_active()) { 1645 // A consistency test for GC_locker 1646 assert(GC_locker::needs_gc(), "Should have been set already"); 1647 // Skip this foreground collection, instead 1648 // expanding the heap if necessary. 1649 // Need the free list locks for the call to free() in compute_new_size() 1650 compute_new_size(); 1651 return; 1652 } 1653 acquire_control_and_collect(full, clear_all_soft_refs); 1654 _full_gcs_since_conc_gc++; 1655 1656 } 1657 1658 void CMSCollector::request_full_gc(unsigned int full_gc_count) { 1659 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1660 unsigned int gc_count = gch->total_full_collections(); 1661 if (gc_count == full_gc_count) { 1662 MutexLockerEx y(CGC_lock, Mutex::_no_safepoint_check_flag); 1663 _full_gc_requested = true; 1664 CGC_lock->notify(); // nudge CMS thread 1665 } else { 1666 assert(gc_count > full_gc_count, "Error: causal loop"); 1667 } 1668 } 1669 1670 1671 // The foreground and background collectors need to coordinate in order 1672 // to make sure that they do not mutually interfere with CMS collections. 1673 // When a background collection is active, 1674 // the foreground collector may need to take over (preempt) and 1675 // synchronously complete an ongoing collection. Depending on the 1676 // frequency of the background collections and the heap usage 1677 // of the application, this preemption can be seldom or frequent. 1678 // There are only certain 1679 // points in the background collection that the "collection-baton" 1680 // can be passed to the foreground collector. 1681 // 1682 // The foreground collector will wait for the baton before 1683 // starting any part of the collection. The foreground collector 1684 // will only wait at one location. 1685 // 1686 // The background collector will yield the baton before starting a new 1687 // phase of the collection (e.g., before initial marking, marking from roots, 1688 // precleaning, final re-mark, sweep etc.) This is normally done at the head 1689 // of the loop which switches the phases. The background collector does some 1690 // of the phases (initial mark, final re-mark) with the world stopped. 1691 // Because of locking involved in stopping the world, 1692 // the foreground collector should not block waiting for the background 1693 // collector when it is doing a stop-the-world phase. The background 1694 // collector will yield the baton at an additional point just before 1695 // it enters a stop-the-world phase. Once the world is stopped, the 1696 // background collector checks the phase of the collection. If the 1697 // phase has not changed, it proceeds with the collection. If the 1698 // phase has changed, it skips that phase of the collection. See 1699 // the comments on the use of the Heap_lock in collect_in_background(). 1700 // 1701 // Variable used in baton passing. 1702 // _foregroundGCIsActive - Set to true by the foreground collector when 1703 // it wants the baton. The foreground clears it when it has finished 1704 // the collection. 1705 // _foregroundGCShouldWait - Set to true by the background collector 1706 // when it is running. The foreground collector waits while 1707 // _foregroundGCShouldWait is true. 1708 // CGC_lock - monitor used to protect access to the above variables 1709 // and to notify the foreground and background collectors. 1710 // _collectorState - current state of the CMS collection. 1711 // 1712 // The foreground collector 1713 // acquires the CGC_lock 1714 // sets _foregroundGCIsActive 1715 // waits on the CGC_lock for _foregroundGCShouldWait to be false 1716 // various locks acquired in preparation for the collection 1717 // are released so as not to block the background collector 1718 // that is in the midst of a collection 1719 // proceeds with the collection 1720 // clears _foregroundGCIsActive 1721 // returns 1722 // 1723 // The background collector in a loop iterating on the phases of the 1724 // collection 1725 // acquires the CGC_lock 1726 // sets _foregroundGCShouldWait 1727 // if _foregroundGCIsActive is set 1728 // clears _foregroundGCShouldWait, notifies _CGC_lock 1729 // waits on _CGC_lock for _foregroundGCIsActive to become false 1730 // and exits the loop. 1731 // otherwise 1732 // proceed with that phase of the collection 1733 // if the phase is a stop-the-world phase, 1734 // yield the baton once more just before enqueueing 1735 // the stop-world CMS operation (executed by the VM thread). 1736 // returns after all phases of the collection are done 1737 // 1738 1739 void CMSCollector::acquire_control_and_collect(bool full, 1740 bool clear_all_soft_refs) { 1741 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint"); 1742 assert(!Thread::current()->is_ConcurrentGC_thread(), 1743 "shouldn't try to acquire control from self!"); 1744 1745 // Start the protocol for acquiring control of the 1746 // collection from the background collector (aka CMS thread). 1747 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 1748 "VM thread should have CMS token"); 1749 // Remember the possibly interrupted state of an ongoing 1750 // concurrent collection 1751 CollectorState first_state = _collectorState; 1752 1753 // Signal to a possibly ongoing concurrent collection that 1754 // we want to do a foreground collection. 1755 _foregroundGCIsActive = true; 1756 1757 // Disable incremental mode during a foreground collection. 1758 ICMSDisabler icms_disabler; 1759 1760 // release locks and wait for a notify from the background collector 1761 // releasing the locks in only necessary for phases which 1762 // do yields to improve the granularity of the collection. 1763 assert_lock_strong(bitMapLock()); 1764 // We need to lock the Free list lock for the space that we are 1765 // currently collecting. 1766 assert(haveFreelistLocks(), "Must be holding free list locks"); 1767 bitMapLock()->unlock(); 1768 releaseFreelistLocks(); 1769 { 1770 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1771 if (_foregroundGCShouldWait) { 1772 // We are going to be waiting for action for the CMS thread; 1773 // it had better not be gone (for instance at shutdown)! 1774 assert(ConcurrentMarkSweepThread::cmst() != NULL, 1775 "CMS thread must be running"); 1776 // Wait here until the background collector gives us the go-ahead 1777 ConcurrentMarkSweepThread::clear_CMS_flag( 1778 ConcurrentMarkSweepThread::CMS_vm_has_token); // release token 1779 // Get a possibly blocked CMS thread going: 1780 // Note that we set _foregroundGCIsActive true above, 1781 // without protection of the CGC_lock. 1782 CGC_lock->notify(); 1783 assert(!ConcurrentMarkSweepThread::vm_thread_wants_cms_token(), 1784 "Possible deadlock"); 1785 while (_foregroundGCShouldWait) { 1786 // wait for notification 1787 CGC_lock->wait(Mutex::_no_safepoint_check_flag); 1788 // Possibility of delay/starvation here, since CMS token does 1789 // not know to give priority to VM thread? Actually, i think 1790 // there wouldn't be any delay/starvation, but the proof of 1791 // that "fact" (?) appears non-trivial. XXX 20011219YSR 1792 } 1793 ConcurrentMarkSweepThread::set_CMS_flag( 1794 ConcurrentMarkSweepThread::CMS_vm_has_token); 1795 } 1796 } 1797 // The CMS_token is already held. Get back the other locks. 1798 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 1799 "VM thread should have CMS token"); 1800 getFreelistLocks(); 1801 bitMapLock()->lock_without_safepoint_check(); 1802 if (TraceCMSState) { 1803 gclog_or_tty->print_cr("CMS foreground collector has asked for control " 1804 INTPTR_FORMAT " with first state %d", Thread::current(), first_state); 1805 gclog_or_tty->print_cr(" gets control with state %d", _collectorState); 1806 } 1807 1808 // Check if we need to do a compaction, or if not, whether 1809 // we need to start the mark-sweep from scratch. 1810 bool should_compact = false; 1811 bool should_start_over = false; 1812 decide_foreground_collection_type(clear_all_soft_refs, 1813 &should_compact, &should_start_over); 1814 1815 NOT_PRODUCT( 1816 if (RotateCMSCollectionTypes) { 1817 if (_cmsGen->debug_collection_type() == 1818 ConcurrentMarkSweepGeneration::MSC_foreground_collection_type) { 1819 should_compact = true; 1820 } else if (_cmsGen->debug_collection_type() == 1821 ConcurrentMarkSweepGeneration::MS_foreground_collection_type) { 1822 should_compact = false; 1823 } 1824 } 1825 ) 1826 1827 if (PrintGCDetails && first_state > Idling) { 1828 GCCause::Cause cause = GenCollectedHeap::heap()->gc_cause(); 1829 if (GCCause::is_user_requested_gc(cause) || 1830 GCCause::is_serviceability_requested_gc(cause)) { 1831 gclog_or_tty->print(" (concurrent mode interrupted)"); 1832 } else { 1833 gclog_or_tty->print(" (concurrent mode failure)"); 1834 } 1835 } 1836 1837 if (should_compact) { 1838 // If the collection is being acquired from the background 1839 // collector, there may be references on the discovered 1840 // references lists that have NULL referents (being those 1841 // that were concurrently cleared by a mutator) or 1842 // that are no longer active (having been enqueued concurrently 1843 // by the mutator). 1844 // Scrub the list of those references because Mark-Sweep-Compact 1845 // code assumes referents are not NULL and that all discovered 1846 // Reference objects are active. 1847 ref_processor()->clean_up_discovered_references(); 1848 1849 do_compaction_work(clear_all_soft_refs); 1850 1851 // Has the GC time limit been exceeded? 1852 DefNewGeneration* young_gen = _young_gen->as_DefNewGeneration(); 1853 size_t max_eden_size = young_gen->max_capacity() - 1854 young_gen->to()->capacity() - 1855 young_gen->from()->capacity(); 1856 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1857 GCCause::Cause gc_cause = gch->gc_cause(); 1858 size_policy()->check_gc_overhead_limit(_young_gen->used(), 1859 young_gen->eden()->used(), 1860 _cmsGen->max_capacity(), 1861 max_eden_size, 1862 full, 1863 gc_cause, 1864 gch->collector_policy()); 1865 } else { 1866 do_mark_sweep_work(clear_all_soft_refs, first_state, 1867 should_start_over); 1868 } 1869 // Reset the expansion cause, now that we just completed 1870 // a collection cycle. 1871 clear_expansion_cause(); 1872 _foregroundGCIsActive = false; 1873 return; 1874 } 1875 1876 // Resize the tenured generation 1877 // after obtaining the free list locks for the 1878 // two generations. 1879 void CMSCollector::compute_new_size() { 1880 assert_locked_or_safepoint(Heap_lock); 1881 FreelistLocker z(this); 1882 MetaspaceGC::compute_new_size(); 1883 _cmsGen->compute_new_size(); 1884 } 1885 1886 // A work method used by foreground collection to determine 1887 // what type of collection (compacting or not, continuing or fresh) 1888 // it should do. 1889 // NOTE: the intent is to make UseCMSCompactAtFullCollection 1890 // and CMSCompactWhenClearAllSoftRefs the default in the future 1891 // and do away with the flags after a suitable period. 1892 void CMSCollector::decide_foreground_collection_type( 1893 bool clear_all_soft_refs, bool* should_compact, 1894 bool* should_start_over) { 1895 // Normally, we'll compact only if the UseCMSCompactAtFullCollection 1896 // flag is set, and we have either requested a System.gc() or 1897 // the number of full gc's since the last concurrent cycle 1898 // has exceeded the threshold set by CMSFullGCsBeforeCompaction, 1899 // or if an incremental collection has failed 1900 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1901 assert(gch->collector_policy()->is_two_generation_policy(), 1902 "You may want to check the correctness of the following"); 1903 // Inform cms gen if this was due to partial collection failing. 1904 // The CMS gen may use this fact to determine its expansion policy. 1905 if (gch->incremental_collection_will_fail(false /* don't consult_young */)) { 1906 assert(!_cmsGen->incremental_collection_failed(), 1907 "Should have been noticed, reacted to and cleared"); 1908 _cmsGen->set_incremental_collection_failed(); 1909 } 1910 *should_compact = 1911 UseCMSCompactAtFullCollection && 1912 ((_full_gcs_since_conc_gc >= CMSFullGCsBeforeCompaction) || 1913 GCCause::is_user_requested_gc(gch->gc_cause()) || 1914 gch->incremental_collection_will_fail(true /* consult_young */)); 1915 *should_start_over = false; 1916 if (clear_all_soft_refs && !*should_compact) { 1917 // We are about to do a last ditch collection attempt 1918 // so it would normally make sense to do a compaction 1919 // to reclaim as much space as possible. 1920 if (CMSCompactWhenClearAllSoftRefs) { 1921 // Default: The rationale is that in this case either 1922 // we are past the final marking phase, in which case 1923 // we'd have to start over, or so little has been done 1924 // that there's little point in saving that work. Compaction 1925 // appears to be the sensible choice in either case. 1926 *should_compact = true; 1927 } else { 1928 // We have been asked to clear all soft refs, but not to 1929 // compact. Make sure that we aren't past the final checkpoint 1930 // phase, for that is where we process soft refs. If we are already 1931 // past that phase, we'll need to redo the refs discovery phase and 1932 // if necessary clear soft refs that weren't previously 1933 // cleared. We do so by remembering the phase in which 1934 // we came in, and if we are past the refs processing 1935 // phase, we'll choose to just redo the mark-sweep 1936 // collection from scratch. 1937 if (_collectorState > FinalMarking) { 1938 // We are past the refs processing phase; 1939 // start over and do a fresh synchronous CMS cycle 1940 _collectorState = Resetting; // skip to reset to start new cycle 1941 reset(false /* == !asynch */); 1942 *should_start_over = true; 1943 } // else we can continue a possibly ongoing current cycle 1944 } 1945 } 1946 } 1947 1948 // A work method used by the foreground collector to do 1949 // a mark-sweep-compact. 1950 void CMSCollector::do_compaction_work(bool clear_all_soft_refs) { 1951 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1952 TraceTime t("CMS:MSC ", PrintGCDetails && Verbose, true, gclog_or_tty); 1953 if (PrintGC && Verbose && !(GCCause::is_user_requested_gc(gch->gc_cause()))) { 1954 gclog_or_tty->print_cr("Compact ConcurrentMarkSweepGeneration after %d " 1955 "collections passed to foreground collector", _full_gcs_since_conc_gc); 1956 } 1957 1958 // Sample collection interval time and reset for collection pause. 1959 if (UseAdaptiveSizePolicy) { 1960 size_policy()->msc_collection_begin(); 1961 } 1962 1963 // Temporarily widen the span of the weak reference processing to 1964 // the entire heap. 1965 MemRegion new_span(GenCollectedHeap::heap()->reserved_region()); 1966 ReferenceProcessorSpanMutator rp_mut_span(ref_processor(), new_span); 1967 // Temporarily, clear the "is_alive_non_header" field of the 1968 // reference processor. 1969 ReferenceProcessorIsAliveMutator rp_mut_closure(ref_processor(), NULL); 1970 // Temporarily make reference _processing_ single threaded (non-MT). 1971 ReferenceProcessorMTProcMutator rp_mut_mt_processing(ref_processor(), false); 1972 // Temporarily make refs discovery atomic 1973 ReferenceProcessorAtomicMutator rp_mut_atomic(ref_processor(), true); 1974 // Temporarily make reference _discovery_ single threaded (non-MT) 1975 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false); 1976 1977 ref_processor()->set_enqueuing_is_done(false); 1978 ref_processor()->enable_discovery(false /*verify_disabled*/, false /*check_no_refs*/); 1979 ref_processor()->setup_policy(clear_all_soft_refs); 1980 // If an asynchronous collection finishes, the _modUnionTable is 1981 // all clear. If we are assuming the collection from an asynchronous 1982 // collection, clear the _modUnionTable. 1983 assert(_collectorState != Idling || _modUnionTable.isAllClear(), 1984 "_modUnionTable should be clear if the baton was not passed"); 1985 _modUnionTable.clear_all(); 1986 assert(_collectorState != Idling || _ct->klass_rem_set()->mod_union_is_clear(), 1987 "mod union for klasses should be clear if the baton was passed"); 1988 _ct->klass_rem_set()->clear_mod_union(); 1989 1990 // We must adjust the allocation statistics being maintained 1991 // in the free list space. We do so by reading and clearing 1992 // the sweep timer and updating the block flux rate estimates below. 1993 assert(!_intra_sweep_timer.is_active(), "_intra_sweep_timer should be inactive"); 1994 if (_inter_sweep_timer.is_active()) { 1995 _inter_sweep_timer.stop(); 1996 // Note that we do not use this sample to update the _inter_sweep_estimate. 1997 _cmsGen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()), 1998 _inter_sweep_estimate.padded_average(), 1999 _intra_sweep_estimate.padded_average()); 2000 } 2001 2002 GenMarkSweep::invoke_at_safepoint(_cmsGen->level(), 2003 ref_processor(), clear_all_soft_refs); 2004 #ifdef ASSERT 2005 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 2006 size_t free_size = cms_space->free(); 2007 assert(free_size == 2008 pointer_delta(cms_space->end(), cms_space->compaction_top()) 2009 * HeapWordSize, 2010 "All the free space should be compacted into one chunk at top"); 2011 assert(cms_space->dictionary()->total_chunk_size( 2012 debug_only(cms_space->freelistLock())) == 0 || 2013 cms_space->totalSizeInIndexedFreeLists() == 0, 2014 "All the free space should be in a single chunk"); 2015 size_t num = cms_space->totalCount(); 2016 assert((free_size == 0 && num == 0) || 2017 (free_size > 0 && (num == 1 || num == 2)), 2018 "There should be at most 2 free chunks after compaction"); 2019 #endif // ASSERT 2020 _collectorState = Resetting; 2021 assert(_restart_addr == NULL, 2022 "Should have been NULL'd before baton was passed"); 2023 reset(false /* == !asynch */); 2024 _cmsGen->reset_after_compaction(); 2025 _concurrent_cycles_since_last_unload = 0; 2026 2027 // Clear any data recorded in the PLAB chunk arrays. 2028 if (_survivor_plab_array != NULL) { 2029 reset_survivor_plab_arrays(); 2030 } 2031 2032 // Adjust the per-size allocation stats for the next epoch. 2033 _cmsGen->cmsSpace()->endSweepFLCensus(sweep_count() /* fake */); 2034 // Restart the "inter sweep timer" for the next epoch. 2035 _inter_sweep_timer.reset(); 2036 _inter_sweep_timer.start(); 2037 2038 // Sample collection pause time and reset for collection interval. 2039 if (UseAdaptiveSizePolicy) { 2040 size_policy()->msc_collection_end(gch->gc_cause()); 2041 } 2042 2043 // For a mark-sweep-compact, compute_new_size() will be called 2044 // in the heap's do_collection() method. 2045 } 2046 2047 // A work method used by the foreground collector to do 2048 // a mark-sweep, after taking over from a possibly on-going 2049 // concurrent mark-sweep collection. 2050 void CMSCollector::do_mark_sweep_work(bool clear_all_soft_refs, 2051 CollectorState first_state, bool should_start_over) { 2052 if (PrintGC && Verbose) { 2053 gclog_or_tty->print_cr("Pass concurrent collection to foreground " 2054 "collector with count %d", 2055 _full_gcs_since_conc_gc); 2056 } 2057 switch (_collectorState) { 2058 case Idling: 2059 if (first_state == Idling || should_start_over) { 2060 // The background GC was not active, or should 2061 // restarted from scratch; start the cycle. 2062 _collectorState = InitialMarking; 2063 } 2064 // If first_state was not Idling, then a background GC 2065 // was in progress and has now finished. No need to do it 2066 // again. Leave the state as Idling. 2067 break; 2068 case Precleaning: 2069 // In the foreground case don't do the precleaning since 2070 // it is not done concurrently and there is extra work 2071 // required. 2072 _collectorState = FinalMarking; 2073 } 2074 if (PrintGCDetails && 2075 (_collectorState > Idling || 2076 !GCCause::is_user_requested_gc(GenCollectedHeap::heap()->gc_cause()))) { 2077 gclog_or_tty->print(" (concurrent mode failure)"); 2078 } 2079 collect_in_foreground(clear_all_soft_refs); 2080 2081 // For a mark-sweep, compute_new_size() will be called 2082 // in the heap's do_collection() method. 2083 } 2084 2085 2086 void CMSCollector::getFreelistLocks() const { 2087 // Get locks for all free lists in all generations that this 2088 // collector is responsible for 2089 _cmsGen->freelistLock()->lock_without_safepoint_check(); 2090 } 2091 2092 void CMSCollector::releaseFreelistLocks() const { 2093 // Release locks for all free lists in all generations that this 2094 // collector is responsible for 2095 _cmsGen->freelistLock()->unlock(); 2096 } 2097 2098 bool CMSCollector::haveFreelistLocks() const { 2099 // Check locks for all free lists in all generations that this 2100 // collector is responsible for 2101 assert_lock_strong(_cmsGen->freelistLock()); 2102 PRODUCT_ONLY(ShouldNotReachHere()); 2103 return true; 2104 } 2105 2106 // A utility class that is used by the CMS collector to 2107 // temporarily "release" the foreground collector from its 2108 // usual obligation to wait for the background collector to 2109 // complete an ongoing phase before proceeding. 2110 class ReleaseForegroundGC: public StackObj { 2111 private: 2112 CMSCollector* _c; 2113 public: 2114 ReleaseForegroundGC(CMSCollector* c) : _c(c) { 2115 assert(_c->_foregroundGCShouldWait, "Else should not need to call"); 2116 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 2117 // allow a potentially blocked foreground collector to proceed 2118 _c->_foregroundGCShouldWait = false; 2119 if (_c->_foregroundGCIsActive) { 2120 CGC_lock->notify(); 2121 } 2122 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 2123 "Possible deadlock"); 2124 } 2125 2126 ~ReleaseForegroundGC() { 2127 assert(!_c->_foregroundGCShouldWait, "Usage protocol violation?"); 2128 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 2129 _c->_foregroundGCShouldWait = true; 2130 } 2131 }; 2132 2133 // There are separate collect_in_background and collect_in_foreground because of 2134 // the different locking requirements of the background collector and the 2135 // foreground collector. There was originally an attempt to share 2136 // one "collect" method between the background collector and the foreground 2137 // collector but the if-then-else required made it cleaner to have 2138 // separate methods. 2139 void CMSCollector::collect_in_background(bool clear_all_soft_refs) { 2140 assert(Thread::current()->is_ConcurrentGC_thread(), 2141 "A CMS asynchronous collection is only allowed on a CMS thread."); 2142 2143 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2144 { 2145 bool safepoint_check = Mutex::_no_safepoint_check_flag; 2146 MutexLockerEx hl(Heap_lock, safepoint_check); 2147 FreelistLocker fll(this); 2148 MutexLockerEx x(CGC_lock, safepoint_check); 2149 if (_foregroundGCIsActive || !UseAsyncConcMarkSweepGC) { 2150 // The foreground collector is active or we're 2151 // not using asynchronous collections. Skip this 2152 // background collection. 2153 assert(!_foregroundGCShouldWait, "Should be clear"); 2154 return; 2155 } else { 2156 assert(_collectorState == Idling, "Should be idling before start."); 2157 _collectorState = InitialMarking; 2158 // Reset the expansion cause, now that we are about to begin 2159 // a new cycle. 2160 clear_expansion_cause(); 2161 2162 // Clear the MetaspaceGC flag since a concurrent collection 2163 // is starting but also clear it after the collection. 2164 MetaspaceGC::set_should_concurrent_collect(false); 2165 } 2166 // Decide if we want to enable class unloading as part of the 2167 // ensuing concurrent GC cycle. 2168 update_should_unload_classes(); 2169 _full_gc_requested = false; // acks all outstanding full gc requests 2170 // Signal that we are about to start a collection 2171 gch->increment_total_full_collections(); // ... starting a collection cycle 2172 _collection_count_start = gch->total_full_collections(); 2173 } 2174 2175 // Used for PrintGC 2176 size_t prev_used; 2177 if (PrintGC && Verbose) { 2178 prev_used = _cmsGen->used(); // XXXPERM 2179 } 2180 2181 // The change of the collection state is normally done at this level; 2182 // the exceptions are phases that are executed while the world is 2183 // stopped. For those phases the change of state is done while the 2184 // world is stopped. For baton passing purposes this allows the 2185 // background collector to finish the phase and change state atomically. 2186 // The foreground collector cannot wait on a phase that is done 2187 // while the world is stopped because the foreground collector already 2188 // has the world stopped and would deadlock. 2189 while (_collectorState != Idling) { 2190 if (TraceCMSState) { 2191 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " in CMS state %d", 2192 Thread::current(), _collectorState); 2193 } 2194 // The foreground collector 2195 // holds the Heap_lock throughout its collection. 2196 // holds the CMS token (but not the lock) 2197 // except while it is waiting for the background collector to yield. 2198 // 2199 // The foreground collector should be blocked (not for long) 2200 // if the background collector is about to start a phase 2201 // executed with world stopped. If the background 2202 // collector has already started such a phase, the 2203 // foreground collector is blocked waiting for the 2204 // Heap_lock. The stop-world phases (InitialMarking and FinalMarking) 2205 // are executed in the VM thread. 2206 // 2207 // The locking order is 2208 // PendingListLock (PLL) -- if applicable (FinalMarking) 2209 // Heap_lock (both this & PLL locked in VM_CMS_Operation::prologue()) 2210 // CMS token (claimed in 2211 // stop_world_and_do() --> 2212 // safepoint_synchronize() --> 2213 // CMSThread::synchronize()) 2214 2215 { 2216 // Check if the FG collector wants us to yield. 2217 CMSTokenSync x(true); // is cms thread 2218 if (waitForForegroundGC()) { 2219 // We yielded to a foreground GC, nothing more to be 2220 // done this round. 2221 assert(_foregroundGCShouldWait == false, "We set it to false in " 2222 "waitForForegroundGC()"); 2223 if (TraceCMSState) { 2224 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT 2225 " exiting collection CMS state %d", 2226 Thread::current(), _collectorState); 2227 } 2228 return; 2229 } else { 2230 // The background collector can run but check to see if the 2231 // foreground collector has done a collection while the 2232 // background collector was waiting to get the CGC_lock 2233 // above. If yes, break so that _foregroundGCShouldWait 2234 // is cleared before returning. 2235 if (_collectorState == Idling) { 2236 break; 2237 } 2238 } 2239 } 2240 2241 assert(_foregroundGCShouldWait, "Foreground collector, if active, " 2242 "should be waiting"); 2243 2244 switch (_collectorState) { 2245 case InitialMarking: 2246 { 2247 ReleaseForegroundGC x(this); 2248 stats().record_cms_begin(); 2249 2250 VM_CMS_Initial_Mark initial_mark_op(this); 2251 VMThread::execute(&initial_mark_op); 2252 } 2253 // The collector state may be any legal state at this point 2254 // since the background collector may have yielded to the 2255 // foreground collector. 2256 break; 2257 case Marking: 2258 // initial marking in checkpointRootsInitialWork has been completed 2259 if (markFromRoots(true)) { // we were successful 2260 assert(_collectorState == Precleaning, "Collector state should " 2261 "have changed"); 2262 } else { 2263 assert(_foregroundGCIsActive, "Internal state inconsistency"); 2264 } 2265 break; 2266 case Precleaning: 2267 if (UseAdaptiveSizePolicy) { 2268 size_policy()->concurrent_precleaning_begin(); 2269 } 2270 // marking from roots in markFromRoots has been completed 2271 preclean(); 2272 if (UseAdaptiveSizePolicy) { 2273 size_policy()->concurrent_precleaning_end(); 2274 } 2275 assert(_collectorState == AbortablePreclean || 2276 _collectorState == FinalMarking, 2277 "Collector state should have changed"); 2278 break; 2279 case AbortablePreclean: 2280 if (UseAdaptiveSizePolicy) { 2281 size_policy()->concurrent_phases_resume(); 2282 } 2283 abortable_preclean(); 2284 if (UseAdaptiveSizePolicy) { 2285 size_policy()->concurrent_precleaning_end(); 2286 } 2287 assert(_collectorState == FinalMarking, "Collector state should " 2288 "have changed"); 2289 break; 2290 case FinalMarking: 2291 { 2292 ReleaseForegroundGC x(this); 2293 2294 VM_CMS_Final_Remark final_remark_op(this); 2295 VMThread::execute(&final_remark_op); 2296 } 2297 assert(_foregroundGCShouldWait, "block post-condition"); 2298 break; 2299 case Sweeping: 2300 if (UseAdaptiveSizePolicy) { 2301 size_policy()->concurrent_sweeping_begin(); 2302 } 2303 // final marking in checkpointRootsFinal has been completed 2304 sweep(true); 2305 assert(_collectorState == Resizing, "Collector state change " 2306 "to Resizing must be done under the free_list_lock"); 2307 _full_gcs_since_conc_gc = 0; 2308 2309 // Stop the timers for adaptive size policy for the concurrent phases 2310 if (UseAdaptiveSizePolicy) { 2311 size_policy()->concurrent_sweeping_end(); 2312 size_policy()->concurrent_phases_end(gch->gc_cause(), 2313 gch->prev_gen(_cmsGen)->capacity(), 2314 _cmsGen->free()); 2315 } 2316 2317 case Resizing: { 2318 // Sweeping has been completed... 2319 // At this point the background collection has completed. 2320 // Don't move the call to compute_new_size() down 2321 // into code that might be executed if the background 2322 // collection was preempted. 2323 { 2324 ReleaseForegroundGC x(this); // unblock FG collection 2325 MutexLockerEx y(Heap_lock, Mutex::_no_safepoint_check_flag); 2326 CMSTokenSync z(true); // not strictly needed. 2327 if (_collectorState == Resizing) { 2328 compute_new_size(); 2329 _collectorState = Resetting; 2330 } else { 2331 assert(_collectorState == Idling, "The state should only change" 2332 " because the foreground collector has finished the collection"); 2333 } 2334 } 2335 break; 2336 } 2337 case Resetting: 2338 // CMS heap resizing has been completed 2339 reset(true); 2340 assert(_collectorState == Idling, "Collector state should " 2341 "have changed"); 2342 2343 MetaspaceGC::set_should_concurrent_collect(false); 2344 2345 stats().record_cms_end(); 2346 // Don't move the concurrent_phases_end() and compute_new_size() 2347 // calls to here because a preempted background collection 2348 // has it's state set to "Resetting". 2349 break; 2350 case Idling: 2351 default: 2352 ShouldNotReachHere(); 2353 break; 2354 } 2355 if (TraceCMSState) { 2356 gclog_or_tty->print_cr(" Thread " INTPTR_FORMAT " done - next CMS state %d", 2357 Thread::current(), _collectorState); 2358 } 2359 assert(_foregroundGCShouldWait, "block post-condition"); 2360 } 2361 2362 // Should this be in gc_epilogue? 2363 collector_policy()->counters()->update_counters(); 2364 2365 { 2366 // Clear _foregroundGCShouldWait and, in the event that the 2367 // foreground collector is waiting, notify it, before 2368 // returning. 2369 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 2370 _foregroundGCShouldWait = false; 2371 if (_foregroundGCIsActive) { 2372 CGC_lock->notify(); 2373 } 2374 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 2375 "Possible deadlock"); 2376 } 2377 if (TraceCMSState) { 2378 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT 2379 " exiting collection CMS state %d", 2380 Thread::current(), _collectorState); 2381 } 2382 if (PrintGC && Verbose) { 2383 _cmsGen->print_heap_change(prev_used); 2384 } 2385 } 2386 2387 void CMSCollector::collect_in_foreground(bool clear_all_soft_refs) { 2388 assert(_foregroundGCIsActive && !_foregroundGCShouldWait, 2389 "Foreground collector should be waiting, not executing"); 2390 assert(Thread::current()->is_VM_thread(), "A foreground collection" 2391 "may only be done by the VM Thread with the world stopped"); 2392 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 2393 "VM thread should have CMS token"); 2394 2395 NOT_PRODUCT(TraceTime t("CMS:MS (foreground) ", PrintGCDetails && Verbose, 2396 true, gclog_or_tty);) 2397 if (UseAdaptiveSizePolicy) { 2398 size_policy()->ms_collection_begin(); 2399 } 2400 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact); 2401 2402 HandleMark hm; // Discard invalid handles created during verification 2403 2404 if (VerifyBeforeGC && 2405 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2406 Universe::verify(true); 2407 } 2408 2409 // Snapshot the soft reference policy to be used in this collection cycle. 2410 ref_processor()->setup_policy(clear_all_soft_refs); 2411 2412 bool init_mark_was_synchronous = false; // until proven otherwise 2413 while (_collectorState != Idling) { 2414 if (TraceCMSState) { 2415 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " in CMS state %d", 2416 Thread::current(), _collectorState); 2417 } 2418 switch (_collectorState) { 2419 case InitialMarking: 2420 init_mark_was_synchronous = true; // fact to be exploited in re-mark 2421 checkpointRootsInitial(false); 2422 assert(_collectorState == Marking, "Collector state should have changed" 2423 " within checkpointRootsInitial()"); 2424 break; 2425 case Marking: 2426 // initial marking in checkpointRootsInitialWork has been completed 2427 if (VerifyDuringGC && 2428 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2429 gclog_or_tty->print("Verify before initial mark: "); 2430 Universe::verify(true); 2431 } 2432 { 2433 bool res = markFromRoots(false); 2434 assert(res && _collectorState == FinalMarking, "Collector state should " 2435 "have changed"); 2436 break; 2437 } 2438 case FinalMarking: 2439 if (VerifyDuringGC && 2440 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2441 gclog_or_tty->print("Verify before re-mark: "); 2442 Universe::verify(true); 2443 } 2444 checkpointRootsFinal(false, clear_all_soft_refs, 2445 init_mark_was_synchronous); 2446 assert(_collectorState == Sweeping, "Collector state should not " 2447 "have changed within checkpointRootsFinal()"); 2448 break; 2449 case Sweeping: 2450 // final marking in checkpointRootsFinal has been completed 2451 if (VerifyDuringGC && 2452 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2453 gclog_or_tty->print("Verify before sweep: "); 2454 Universe::verify(true); 2455 } 2456 sweep(false); 2457 assert(_collectorState == Resizing, "Incorrect state"); 2458 break; 2459 case Resizing: { 2460 // Sweeping has been completed; the actual resize in this case 2461 // is done separately; nothing to be done in this state. 2462 _collectorState = Resetting; 2463 break; 2464 } 2465 case Resetting: 2466 // The heap has been resized. 2467 if (VerifyDuringGC && 2468 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2469 gclog_or_tty->print("Verify before reset: "); 2470 Universe::verify(true); 2471 } 2472 reset(false); 2473 assert(_collectorState == Idling, "Collector state should " 2474 "have changed"); 2475 break; 2476 case Precleaning: 2477 case AbortablePreclean: 2478 // Elide the preclean phase 2479 _collectorState = FinalMarking; 2480 break; 2481 default: 2482 ShouldNotReachHere(); 2483 } 2484 if (TraceCMSState) { 2485 gclog_or_tty->print_cr(" Thread " INTPTR_FORMAT " done - next CMS state %d", 2486 Thread::current(), _collectorState); 2487 } 2488 } 2489 2490 if (UseAdaptiveSizePolicy) { 2491 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2492 size_policy()->ms_collection_end(gch->gc_cause()); 2493 } 2494 2495 if (VerifyAfterGC && 2496 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2497 Universe::verify(true); 2498 } 2499 if (TraceCMSState) { 2500 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT 2501 " exiting collection CMS state %d", 2502 Thread::current(), _collectorState); 2503 } 2504 } 2505 2506 bool CMSCollector::waitForForegroundGC() { 2507 bool res = false; 2508 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 2509 "CMS thread should have CMS token"); 2510 // Block the foreground collector until the 2511 // background collectors decides whether to 2512 // yield. 2513 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 2514 _foregroundGCShouldWait = true; 2515 if (_foregroundGCIsActive) { 2516 // The background collector yields to the 2517 // foreground collector and returns a value 2518 // indicating that it has yielded. The foreground 2519 // collector can proceed. 2520 res = true; 2521 _foregroundGCShouldWait = false; 2522 ConcurrentMarkSweepThread::clear_CMS_flag( 2523 ConcurrentMarkSweepThread::CMS_cms_has_token); 2524 ConcurrentMarkSweepThread::set_CMS_flag( 2525 ConcurrentMarkSweepThread::CMS_cms_wants_token); 2526 // Get a possibly blocked foreground thread going 2527 CGC_lock->notify(); 2528 if (TraceCMSState) { 2529 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " waiting at CMS state %d", 2530 Thread::current(), _collectorState); 2531 } 2532 while (_foregroundGCIsActive) { 2533 CGC_lock->wait(Mutex::_no_safepoint_check_flag); 2534 } 2535 ConcurrentMarkSweepThread::set_CMS_flag( 2536 ConcurrentMarkSweepThread::CMS_cms_has_token); 2537 ConcurrentMarkSweepThread::clear_CMS_flag( 2538 ConcurrentMarkSweepThread::CMS_cms_wants_token); 2539 } 2540 if (TraceCMSState) { 2541 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " continuing at CMS state %d", 2542 Thread::current(), _collectorState); 2543 } 2544 return res; 2545 } 2546 2547 // Because of the need to lock the free lists and other structures in 2548 // the collector, common to all the generations that the collector is 2549 // collecting, we need the gc_prologues of individual CMS generations 2550 // delegate to their collector. It may have been simpler had the 2551 // current infrastructure allowed one to call a prologue on a 2552 // collector. In the absence of that we have the generation's 2553 // prologue delegate to the collector, which delegates back 2554 // some "local" work to a worker method in the individual generations 2555 // that it's responsible for collecting, while itself doing any 2556 // work common to all generations it's responsible for. A similar 2557 // comment applies to the gc_epilogue()'s. 2558 // The role of the varaible _between_prologue_and_epilogue is to 2559 // enforce the invocation protocol. 2560 void CMSCollector::gc_prologue(bool full) { 2561 // Call gc_prologue_work() for the CMSGen 2562 // we are responsible for. 2563 2564 // The following locking discipline assumes that we are only called 2565 // when the world is stopped. 2566 assert(SafepointSynchronize::is_at_safepoint(), "world is stopped assumption"); 2567 2568 // The CMSCollector prologue must call the gc_prologues for the 2569 // "generations" that it's responsible 2570 // for. 2571 2572 assert( Thread::current()->is_VM_thread() 2573 || ( CMSScavengeBeforeRemark 2574 && Thread::current()->is_ConcurrentGC_thread()), 2575 "Incorrect thread type for prologue execution"); 2576 2577 if (_between_prologue_and_epilogue) { 2578 // We have already been invoked; this is a gc_prologue delegation 2579 // from yet another CMS generation that we are responsible for, just 2580 // ignore it since all relevant work has already been done. 2581 return; 2582 } 2583 2584 // set a bit saying prologue has been called; cleared in epilogue 2585 _between_prologue_and_epilogue = true; 2586 // Claim locks for common data structures, then call gc_prologue_work() 2587 // for each CMSGen. 2588 2589 getFreelistLocks(); // gets free list locks on constituent spaces 2590 bitMapLock()->lock_without_safepoint_check(); 2591 2592 // Should call gc_prologue_work() for all cms gens we are responsible for 2593 bool duringMarking = _collectorState >= Marking 2594 && _collectorState < Sweeping; 2595 2596 // The young collections clear the modified oops state, which tells if 2597 // there are any modified oops in the class. The remark phase also needs 2598 // that information. Tell the young collection to save the union of all 2599 // modified klasses. 2600 if (duringMarking) { 2601 _ct->klass_rem_set()->set_accumulate_modified_oops(true); 2602 } 2603 2604 bool registerClosure = duringMarking; 2605 2606 ModUnionClosure* muc = CollectedHeap::use_parallel_gc_threads() ? 2607 &_modUnionClosurePar 2608 : &_modUnionClosure; 2609 _cmsGen->gc_prologue_work(full, registerClosure, muc); 2610 2611 if (!full) { 2612 stats().record_gc0_begin(); 2613 } 2614 } 2615 2616 void ConcurrentMarkSweepGeneration::gc_prologue(bool full) { 2617 // Delegate to CMScollector which knows how to coordinate between 2618 // this and any other CMS generations that it is responsible for 2619 // collecting. 2620 collector()->gc_prologue(full); 2621 } 2622 2623 // This is a "private" interface for use by this generation's CMSCollector. 2624 // Not to be called directly by any other entity (for instance, 2625 // GenCollectedHeap, which calls the "public" gc_prologue method above). 2626 void ConcurrentMarkSweepGeneration::gc_prologue_work(bool full, 2627 bool registerClosure, ModUnionClosure* modUnionClosure) { 2628 assert(!incremental_collection_failed(), "Shouldn't be set yet"); 2629 assert(cmsSpace()->preconsumptionDirtyCardClosure() == NULL, 2630 "Should be NULL"); 2631 if (registerClosure) { 2632 cmsSpace()->setPreconsumptionDirtyCardClosure(modUnionClosure); 2633 } 2634 cmsSpace()->gc_prologue(); 2635 // Clear stat counters 2636 NOT_PRODUCT( 2637 assert(_numObjectsPromoted == 0, "check"); 2638 assert(_numWordsPromoted == 0, "check"); 2639 if (Verbose && PrintGC) { 2640 gclog_or_tty->print("Allocated "SIZE_FORMAT" objects, " 2641 SIZE_FORMAT" bytes concurrently", 2642 _numObjectsAllocated, _numWordsAllocated*sizeof(HeapWord)); 2643 } 2644 _numObjectsAllocated = 0; 2645 _numWordsAllocated = 0; 2646 ) 2647 } 2648 2649 void CMSCollector::gc_epilogue(bool full) { 2650 // The following locking discipline assumes that we are only called 2651 // when the world is stopped. 2652 assert(SafepointSynchronize::is_at_safepoint(), 2653 "world is stopped assumption"); 2654 2655 // Currently the CMS epilogue (see CompactibleFreeListSpace) merely checks 2656 // if linear allocation blocks need to be appropriately marked to allow the 2657 // the blocks to be parsable. We also check here whether we need to nudge the 2658 // CMS collector thread to start a new cycle (if it's not already active). 2659 assert( Thread::current()->is_VM_thread() 2660 || ( CMSScavengeBeforeRemark 2661 && Thread::current()->is_ConcurrentGC_thread()), 2662 "Incorrect thread type for epilogue execution"); 2663 2664 if (!_between_prologue_and_epilogue) { 2665 // We have already been invoked; this is a gc_epilogue delegation 2666 // from yet another CMS generation that we are responsible for, just 2667 // ignore it since all relevant work has already been done. 2668 return; 2669 } 2670 assert(haveFreelistLocks(), "must have freelist locks"); 2671 assert_lock_strong(bitMapLock()); 2672 2673 _ct->klass_rem_set()->set_accumulate_modified_oops(false); 2674 2675 _cmsGen->gc_epilogue_work(full); 2676 2677 if (_collectorState == AbortablePreclean || _collectorState == Precleaning) { 2678 // in case sampling was not already enabled, enable it 2679 _start_sampling = true; 2680 } 2681 // reset _eden_chunk_array so sampling starts afresh 2682 _eden_chunk_index = 0; 2683 2684 size_t cms_used = _cmsGen->cmsSpace()->used(); 2685 2686 // update performance counters - this uses a special version of 2687 // update_counters() that allows the utilization to be passed as a 2688 // parameter, avoiding multiple calls to used(). 2689 // 2690 _cmsGen->update_counters(cms_used); 2691 2692 if (CMSIncrementalMode) { 2693 icms_update_allocation_limits(); 2694 } 2695 2696 bitMapLock()->unlock(); 2697 releaseFreelistLocks(); 2698 2699 if (!CleanChunkPoolAsync) { 2700 Chunk::clean_chunk_pool(); 2701 } 2702 2703 _between_prologue_and_epilogue = false; // ready for next cycle 2704 } 2705 2706 void ConcurrentMarkSweepGeneration::gc_epilogue(bool full) { 2707 collector()->gc_epilogue(full); 2708 2709 // Also reset promotion tracking in par gc thread states. 2710 if (CollectedHeap::use_parallel_gc_threads()) { 2711 for (uint i = 0; i < ParallelGCThreads; i++) { 2712 _par_gc_thread_states[i]->promo.stopTrackingPromotions(i); 2713 } 2714 } 2715 } 2716 2717 void ConcurrentMarkSweepGeneration::gc_epilogue_work(bool full) { 2718 assert(!incremental_collection_failed(), "Should have been cleared"); 2719 cmsSpace()->setPreconsumptionDirtyCardClosure(NULL); 2720 cmsSpace()->gc_epilogue(); 2721 // Print stat counters 2722 NOT_PRODUCT( 2723 assert(_numObjectsAllocated == 0, "check"); 2724 assert(_numWordsAllocated == 0, "check"); 2725 if (Verbose && PrintGC) { 2726 gclog_or_tty->print("Promoted "SIZE_FORMAT" objects, " 2727 SIZE_FORMAT" bytes", 2728 _numObjectsPromoted, _numWordsPromoted*sizeof(HeapWord)); 2729 } 2730 _numObjectsPromoted = 0; 2731 _numWordsPromoted = 0; 2732 ) 2733 2734 if (PrintGC && Verbose) { 2735 // Call down the chain in contiguous_available needs the freelistLock 2736 // so print this out before releasing the freeListLock. 2737 gclog_or_tty->print(" Contiguous available "SIZE_FORMAT" bytes ", 2738 contiguous_available()); 2739 } 2740 } 2741 2742 #ifndef PRODUCT 2743 bool CMSCollector::have_cms_token() { 2744 Thread* thr = Thread::current(); 2745 if (thr->is_VM_thread()) { 2746 return ConcurrentMarkSweepThread::vm_thread_has_cms_token(); 2747 } else if (thr->is_ConcurrentGC_thread()) { 2748 return ConcurrentMarkSweepThread::cms_thread_has_cms_token(); 2749 } else if (thr->is_GC_task_thread()) { 2750 return ConcurrentMarkSweepThread::vm_thread_has_cms_token() && 2751 ParGCRareEvent_lock->owned_by_self(); 2752 } 2753 return false; 2754 } 2755 #endif 2756 2757 // Check reachability of the given heap address in CMS generation, 2758 // treating all other generations as roots. 2759 bool CMSCollector::is_cms_reachable(HeapWord* addr) { 2760 // We could "guarantee" below, rather than assert, but i'll 2761 // leave these as "asserts" so that an adventurous debugger 2762 // could try this in the product build provided some subset of 2763 // the conditions were met, provided they were intersted in the 2764 // results and knew that the computation below wouldn't interfere 2765 // with other concurrent computations mutating the structures 2766 // being read or written. 2767 assert(SafepointSynchronize::is_at_safepoint(), 2768 "Else mutations in object graph will make answer suspect"); 2769 assert(have_cms_token(), "Should hold cms token"); 2770 assert(haveFreelistLocks(), "must hold free list locks"); 2771 assert_lock_strong(bitMapLock()); 2772 2773 // Clear the marking bit map array before starting, but, just 2774 // for kicks, first report if the given address is already marked 2775 gclog_or_tty->print_cr("Start: Address 0x%x is%s marked", addr, 2776 _markBitMap.isMarked(addr) ? "" : " not"); 2777 2778 if (verify_after_remark()) { 2779 MutexLockerEx x(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag); 2780 bool result = verification_mark_bm()->isMarked(addr); 2781 gclog_or_tty->print_cr("TransitiveMark: Address 0x%x %s marked", addr, 2782 result ? "IS" : "is NOT"); 2783 return result; 2784 } else { 2785 gclog_or_tty->print_cr("Could not compute result"); 2786 return false; 2787 } 2788 } 2789 2790 //////////////////////////////////////////////////////// 2791 // CMS Verification Support 2792 //////////////////////////////////////////////////////// 2793 // Following the remark phase, the following invariant 2794 // should hold -- each object in the CMS heap which is 2795 // marked in markBitMap() should be marked in the verification_mark_bm(). 2796 2797 class VerifyMarkedClosure: public BitMapClosure { 2798 CMSBitMap* _marks; 2799 bool _failed; 2800 2801 public: 2802 VerifyMarkedClosure(CMSBitMap* bm): _marks(bm), _failed(false) {} 2803 2804 bool do_bit(size_t offset) { 2805 HeapWord* addr = _marks->offsetToHeapWord(offset); 2806 if (!_marks->isMarked(addr)) { 2807 oop(addr)->print_on(gclog_or_tty); 2808 gclog_or_tty->print_cr(" ("INTPTR_FORMAT" should have been marked)", addr); 2809 _failed = true; 2810 } 2811 return true; 2812 } 2813 2814 bool failed() { return _failed; } 2815 }; 2816 2817 bool CMSCollector::verify_after_remark() { 2818 gclog_or_tty->print(" [Verifying CMS Marking... "); 2819 MutexLockerEx ml(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag); 2820 static bool init = false; 2821 2822 assert(SafepointSynchronize::is_at_safepoint(), 2823 "Else mutations in object graph will make answer suspect"); 2824 assert(have_cms_token(), 2825 "Else there may be mutual interference in use of " 2826 " verification data structures"); 2827 assert(_collectorState > Marking && _collectorState <= Sweeping, 2828 "Else marking info checked here may be obsolete"); 2829 assert(haveFreelistLocks(), "must hold free list locks"); 2830 assert_lock_strong(bitMapLock()); 2831 2832 2833 // Allocate marking bit map if not already allocated 2834 if (!init) { // first time 2835 if (!verification_mark_bm()->allocate(_span)) { 2836 return false; 2837 } 2838 init = true; 2839 } 2840 2841 assert(verification_mark_stack()->isEmpty(), "Should be empty"); 2842 2843 // Turn off refs discovery -- so we will be tracing through refs. 2844 // This is as intended, because by this time 2845 // GC must already have cleared any refs that need to be cleared, 2846 // and traced those that need to be marked; moreover, 2847 // the marking done here is not going to intefere in any 2848 // way with the marking information used by GC. 2849 NoRefDiscovery no_discovery(ref_processor()); 2850 2851 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;) 2852 2853 // Clear any marks from a previous round 2854 verification_mark_bm()->clear_all(); 2855 assert(verification_mark_stack()->isEmpty(), "markStack should be empty"); 2856 verify_work_stacks_empty(); 2857 2858 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2859 gch->ensure_parsability(false); // fill TLABs, but no need to retire them 2860 // Update the saved marks which may affect the root scans. 2861 gch->save_marks(); 2862 2863 if (CMSRemarkVerifyVariant == 1) { 2864 // In this first variant of verification, we complete 2865 // all marking, then check if the new marks-verctor is 2866 // a subset of the CMS marks-vector. 2867 verify_after_remark_work_1(); 2868 } else if (CMSRemarkVerifyVariant == 2) { 2869 // In this second variant of verification, we flag an error 2870 // (i.e. an object reachable in the new marks-vector not reachable 2871 // in the CMS marks-vector) immediately, also indicating the 2872 // identify of an object (A) that references the unmarked object (B) -- 2873 // presumably, a mutation to A failed to be picked up by preclean/remark? 2874 verify_after_remark_work_2(); 2875 } else { 2876 warning("Unrecognized value %d for CMSRemarkVerifyVariant", 2877 CMSRemarkVerifyVariant); 2878 } 2879 gclog_or_tty->print(" done] "); 2880 return true; 2881 } 2882 2883 void CMSCollector::verify_after_remark_work_1() { 2884 ResourceMark rm; 2885 HandleMark hm; 2886 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2887 2888 // Get a clear set of claim bits for the strong roots processing to work with. 2889 ClassLoaderDataGraph::clear_claimed_marks(); 2890 2891 // Mark from roots one level into CMS 2892 MarkRefsIntoClosure notOlder(_span, verification_mark_bm()); 2893 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2894 2895 gch->gen_process_strong_roots(_cmsGen->level(), 2896 true, // younger gens are roots 2897 true, // activate StrongRootsScope 2898 false, // not scavenging 2899 SharedHeap::ScanningOption(roots_scanning_options()), 2900 ¬Older, 2901 true, // walk code active on stacks 2902 NULL, 2903 NULL); // SSS: Provide correct closure 2904 2905 // Now mark from the roots 2906 MarkFromRootsClosure markFromRootsClosure(this, _span, 2907 verification_mark_bm(), verification_mark_stack(), 2908 false /* don't yield */, true /* verifying */); 2909 assert(_restart_addr == NULL, "Expected pre-condition"); 2910 verification_mark_bm()->iterate(&markFromRootsClosure); 2911 while (_restart_addr != NULL) { 2912 // Deal with stack overflow: by restarting at the indicated 2913 // address. 2914 HeapWord* ra = _restart_addr; 2915 markFromRootsClosure.reset(ra); 2916 _restart_addr = NULL; 2917 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end()); 2918 } 2919 assert(verification_mark_stack()->isEmpty(), "Should have been drained"); 2920 verify_work_stacks_empty(); 2921 2922 // Marking completed -- now verify that each bit marked in 2923 // verification_mark_bm() is also marked in markBitMap(); flag all 2924 // errors by printing corresponding objects. 2925 VerifyMarkedClosure vcl(markBitMap()); 2926 verification_mark_bm()->iterate(&vcl); 2927 if (vcl.failed()) { 2928 gclog_or_tty->print("Verification failed"); 2929 Universe::heap()->print_on(gclog_or_tty); 2930 fatal("CMS: failed marking verification after remark"); 2931 } 2932 } 2933 2934 class VerifyKlassOopsKlassClosure : public KlassClosure { 2935 class VerifyKlassOopsClosure : public OopClosure { 2936 CMSBitMap* _bitmap; 2937 public: 2938 VerifyKlassOopsClosure(CMSBitMap* bitmap) : _bitmap(bitmap) { } 2939 void do_oop(oop* p) { guarantee(*p == NULL || _bitmap->isMarked((HeapWord*) *p), "Should be marked"); } 2940 void do_oop(narrowOop* p) { ShouldNotReachHere(); } 2941 } _oop_closure; 2942 public: 2943 VerifyKlassOopsKlassClosure(CMSBitMap* bitmap) : _oop_closure(bitmap) {} 2944 void do_klass(Klass* k) { 2945 k->oops_do(&_oop_closure); 2946 } 2947 }; 2948 2949 void CMSCollector::verify_after_remark_work_2() { 2950 ResourceMark rm; 2951 HandleMark hm; 2952 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2953 2954 // Get a clear set of claim bits for the strong roots processing to work with. 2955 ClassLoaderDataGraph::clear_claimed_marks(); 2956 2957 // Mark from roots one level into CMS 2958 MarkRefsIntoVerifyClosure notOlder(_span, verification_mark_bm(), 2959 markBitMap()); 2960 CMKlassClosure klass_closure(¬Older); 2961 2962 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2963 gch->gen_process_strong_roots(_cmsGen->level(), 2964 true, // younger gens are roots 2965 true, // activate StrongRootsScope 2966 false, // not scavenging 2967 SharedHeap::ScanningOption(roots_scanning_options()), 2968 ¬Older, 2969 true, // walk code active on stacks 2970 NULL, 2971 &klass_closure); 2972 2973 // Now mark from the roots 2974 MarkFromRootsVerifyClosure markFromRootsClosure(this, _span, 2975 verification_mark_bm(), markBitMap(), verification_mark_stack()); 2976 assert(_restart_addr == NULL, "Expected pre-condition"); 2977 verification_mark_bm()->iterate(&markFromRootsClosure); 2978 while (_restart_addr != NULL) { 2979 // Deal with stack overflow: by restarting at the indicated 2980 // address. 2981 HeapWord* ra = _restart_addr; 2982 markFromRootsClosure.reset(ra); 2983 _restart_addr = NULL; 2984 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end()); 2985 } 2986 assert(verification_mark_stack()->isEmpty(), "Should have been drained"); 2987 verify_work_stacks_empty(); 2988 2989 VerifyKlassOopsKlassClosure verify_klass_oops(verification_mark_bm()); 2990 ClassLoaderDataGraph::classes_do(&verify_klass_oops); 2991 2992 // Marking completed -- now verify that each bit marked in 2993 // verification_mark_bm() is also marked in markBitMap(); flag all 2994 // errors by printing corresponding objects. 2995 VerifyMarkedClosure vcl(markBitMap()); 2996 verification_mark_bm()->iterate(&vcl); 2997 assert(!vcl.failed(), "Else verification above should not have succeeded"); 2998 } 2999 3000 void ConcurrentMarkSweepGeneration::save_marks() { 3001 // delegate to CMS space 3002 cmsSpace()->save_marks(); 3003 for (uint i = 0; i < ParallelGCThreads; i++) { 3004 _par_gc_thread_states[i]->promo.startTrackingPromotions(); 3005 } 3006 } 3007 3008 bool ConcurrentMarkSweepGeneration::no_allocs_since_save_marks() { 3009 return cmsSpace()->no_allocs_since_save_marks(); 3010 } 3011 3012 #define CMS_SINCE_SAVE_MARKS_DEFN(OopClosureType, nv_suffix) \ 3013 \ 3014 void ConcurrentMarkSweepGeneration:: \ 3015 oop_since_save_marks_iterate##nv_suffix(OopClosureType* cl) { \ 3016 cl->set_generation(this); \ 3017 cmsSpace()->oop_since_save_marks_iterate##nv_suffix(cl); \ 3018 cl->reset_generation(); \ 3019 save_marks(); \ 3020 } 3021 3022 ALL_SINCE_SAVE_MARKS_CLOSURES(CMS_SINCE_SAVE_MARKS_DEFN) 3023 3024 void 3025 ConcurrentMarkSweepGeneration::object_iterate_since_last_GC(ObjectClosure* blk) 3026 { 3027 // Not currently implemented; need to do the following. -- ysr. 3028 // dld -- I think that is used for some sort of allocation profiler. So it 3029 // really means the objects allocated by the mutator since the last 3030 // GC. We could potentially implement this cheaply by recording only 3031 // the direct allocations in a side data structure. 3032 // 3033 // I think we probably ought not to be required to support these 3034 // iterations at any arbitrary point; I think there ought to be some 3035 // call to enable/disable allocation profiling in a generation/space, 3036 // and the iterator ought to return the objects allocated in the 3037 // gen/space since the enable call, or the last iterator call (which 3038 // will probably be at a GC.) That way, for gens like CM&S that would 3039 // require some extra data structure to support this, we only pay the 3040 // cost when it's in use... 3041 cmsSpace()->object_iterate_since_last_GC(blk); 3042 } 3043 3044 void 3045 ConcurrentMarkSweepGeneration::younger_refs_iterate(OopsInGenClosure* cl) { 3046 cl->set_generation(this); 3047 younger_refs_in_space_iterate(_cmsSpace, cl); 3048 cl->reset_generation(); 3049 } 3050 3051 void 3052 ConcurrentMarkSweepGeneration::oop_iterate(MemRegion mr, ExtendedOopClosure* cl) { 3053 if (freelistLock()->owned_by_self()) { 3054 Generation::oop_iterate(mr, cl); 3055 } else { 3056 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 3057 Generation::oop_iterate(mr, cl); 3058 } 3059 } 3060 3061 void 3062 ConcurrentMarkSweepGeneration::oop_iterate(ExtendedOopClosure* cl) { 3063 if (freelistLock()->owned_by_self()) { 3064 Generation::oop_iterate(cl); 3065 } else { 3066 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 3067 Generation::oop_iterate(cl); 3068 } 3069 } 3070 3071 void 3072 ConcurrentMarkSweepGeneration::object_iterate(ObjectClosure* cl) { 3073 if (freelistLock()->owned_by_self()) { 3074 Generation::object_iterate(cl); 3075 } else { 3076 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 3077 Generation::object_iterate(cl); 3078 } 3079 } 3080 3081 void 3082 ConcurrentMarkSweepGeneration::safe_object_iterate(ObjectClosure* cl) { 3083 if (freelistLock()->owned_by_self()) { 3084 Generation::safe_object_iterate(cl); 3085 } else { 3086 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 3087 Generation::safe_object_iterate(cl); 3088 } 3089 } 3090 3091 void 3092 ConcurrentMarkSweepGeneration::post_compact() { 3093 } 3094 3095 void 3096 ConcurrentMarkSweepGeneration::prepare_for_verify() { 3097 // Fix the linear allocation blocks to look like free blocks. 3098 3099 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those 3100 // are not called when the heap is verified during universe initialization and 3101 // at vm shutdown. 3102 if (freelistLock()->owned_by_self()) { 3103 cmsSpace()->prepare_for_verify(); 3104 } else { 3105 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag); 3106 cmsSpace()->prepare_for_verify(); 3107 } 3108 } 3109 3110 void 3111 ConcurrentMarkSweepGeneration::verify() { 3112 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those 3113 // are not called when the heap is verified during universe initialization and 3114 // at vm shutdown. 3115 if (freelistLock()->owned_by_self()) { 3116 cmsSpace()->verify(); 3117 } else { 3118 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag); 3119 cmsSpace()->verify(); 3120 } 3121 } 3122 3123 void CMSCollector::verify() { 3124 _cmsGen->verify(); 3125 } 3126 3127 #ifndef PRODUCT 3128 bool CMSCollector::overflow_list_is_empty() const { 3129 assert(_num_par_pushes >= 0, "Inconsistency"); 3130 if (_overflow_list == NULL) { 3131 assert(_num_par_pushes == 0, "Inconsistency"); 3132 } 3133 return _overflow_list == NULL; 3134 } 3135 3136 // The methods verify_work_stacks_empty() and verify_overflow_empty() 3137 // merely consolidate assertion checks that appear to occur together frequently. 3138 void CMSCollector::verify_work_stacks_empty() const { 3139 assert(_markStack.isEmpty(), "Marking stack should be empty"); 3140 assert(overflow_list_is_empty(), "Overflow list should be empty"); 3141 } 3142 3143 void CMSCollector::verify_overflow_empty() const { 3144 assert(overflow_list_is_empty(), "Overflow list should be empty"); 3145 assert(no_preserved_marks(), "No preserved marks"); 3146 } 3147 #endif // PRODUCT 3148 3149 // Decide if we want to enable class unloading as part of the 3150 // ensuing concurrent GC cycle. We will collect and 3151 // unload classes if it's the case that: 3152 // (1) an explicit gc request has been made and the flag 3153 // ExplicitGCInvokesConcurrentAndUnloadsClasses is set, OR 3154 // (2) (a) class unloading is enabled at the command line, and 3155 // (b) old gen is getting really full 3156 // NOTE: Provided there is no change in the state of the heap between 3157 // calls to this method, it should have idempotent results. Moreover, 3158 // its results should be monotonically increasing (i.e. going from 0 to 1, 3159 // but not 1 to 0) between successive calls between which the heap was 3160 // not collected. For the implementation below, it must thus rely on 3161 // the property that concurrent_cycles_since_last_unload() 3162 // will not decrease unless a collection cycle happened and that 3163 // _cmsGen->is_too_full() are 3164 // themselves also monotonic in that sense. See check_monotonicity() 3165 // below. 3166 void CMSCollector::update_should_unload_classes() { 3167 _should_unload_classes = false; 3168 // Condition 1 above 3169 if (_full_gc_requested && ExplicitGCInvokesConcurrentAndUnloadsClasses) { 3170 _should_unload_classes = true; 3171 } else if (CMSClassUnloadingEnabled) { // Condition 2.a above 3172 // Disjuncts 2.b.(i,ii,iii) above 3173 _should_unload_classes = (concurrent_cycles_since_last_unload() >= 3174 CMSClassUnloadingMaxInterval) 3175 || _cmsGen->is_too_full(); 3176 } 3177 } 3178 3179 bool ConcurrentMarkSweepGeneration::is_too_full() const { 3180 bool res = should_concurrent_collect(); 3181 res = res && (occupancy() > (double)CMSIsTooFullPercentage/100.0); 3182 return res; 3183 } 3184 3185 void CMSCollector::setup_cms_unloading_and_verification_state() { 3186 const bool should_verify = VerifyBeforeGC || VerifyAfterGC || VerifyDuringGC 3187 || VerifyBeforeExit; 3188 const int rso = SharedHeap::SO_Strings | SharedHeap::SO_CodeCache; 3189 3190 if (should_unload_classes()) { // Should unload classes this cycle 3191 remove_root_scanning_option(rso); // Shrink the root set appropriately 3192 set_verifying(should_verify); // Set verification state for this cycle 3193 return; // Nothing else needs to be done at this time 3194 } 3195 3196 // Not unloading classes this cycle 3197 assert(!should_unload_classes(), "Inconsitency!"); 3198 if ((!verifying() || unloaded_classes_last_cycle()) && should_verify) { 3199 // Include symbols, strings and code cache elements to prevent their resurrection. 3200 add_root_scanning_option(rso); 3201 set_verifying(true); 3202 } else if (verifying() && !should_verify) { 3203 // We were verifying, but some verification flags got disabled. 3204 set_verifying(false); 3205 // Exclude symbols, strings and code cache elements from root scanning to 3206 // reduce IM and RM pauses. 3207 remove_root_scanning_option(rso); 3208 } 3209 } 3210 3211 3212 #ifndef PRODUCT 3213 HeapWord* CMSCollector::block_start(const void* p) const { 3214 const HeapWord* addr = (HeapWord*)p; 3215 if (_span.contains(p)) { 3216 if (_cmsGen->cmsSpace()->is_in_reserved(addr)) { 3217 return _cmsGen->cmsSpace()->block_start(p); 3218 } 3219 } 3220 return NULL; 3221 } 3222 #endif 3223 3224 HeapWord* 3225 ConcurrentMarkSweepGeneration::expand_and_allocate(size_t word_size, 3226 bool tlab, 3227 bool parallel) { 3228 CMSSynchronousYieldRequest yr; 3229 assert(!tlab, "Can't deal with TLAB allocation"); 3230 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 3231 expand(word_size*HeapWordSize, MinHeapDeltaBytes, 3232 CMSExpansionCause::_satisfy_allocation); 3233 if (GCExpandToAllocateDelayMillis > 0) { 3234 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 3235 } 3236 return have_lock_and_allocate(word_size, tlab); 3237 } 3238 3239 // YSR: All of this generation expansion/shrinking stuff is an exact copy of 3240 // OneContigSpaceCardGeneration, which makes me wonder if we should move this 3241 // to CardGeneration and share it... 3242 bool ConcurrentMarkSweepGeneration::expand(size_t bytes, size_t expand_bytes) { 3243 return CardGeneration::expand(bytes, expand_bytes); 3244 } 3245 3246 void ConcurrentMarkSweepGeneration::expand(size_t bytes, size_t expand_bytes, 3247 CMSExpansionCause::Cause cause) 3248 { 3249 3250 bool success = expand(bytes, expand_bytes); 3251 3252 // remember why we expanded; this information is used 3253 // by shouldConcurrentCollect() when making decisions on whether to start 3254 // a new CMS cycle. 3255 if (success) { 3256 set_expansion_cause(cause); 3257 if (PrintGCDetails && Verbose) { 3258 gclog_or_tty->print_cr("Expanded CMS gen for %s", 3259 CMSExpansionCause::to_string(cause)); 3260 } 3261 } 3262 } 3263 3264 HeapWord* ConcurrentMarkSweepGeneration::expand_and_par_lab_allocate(CMSParGCThreadState* ps, size_t word_sz) { 3265 HeapWord* res = NULL; 3266 MutexLocker x(ParGCRareEvent_lock); 3267 while (true) { 3268 // Expansion by some other thread might make alloc OK now: 3269 res = ps->lab.alloc(word_sz); 3270 if (res != NULL) return res; 3271 // If there's not enough expansion space available, give up. 3272 if (_virtual_space.uncommitted_size() < (word_sz * HeapWordSize)) { 3273 return NULL; 3274 } 3275 // Otherwise, we try expansion. 3276 expand(word_sz*HeapWordSize, MinHeapDeltaBytes, 3277 CMSExpansionCause::_allocate_par_lab); 3278 // Now go around the loop and try alloc again; 3279 // A competing par_promote might beat us to the expansion space, 3280 // so we may go around the loop again if promotion fails agaion. 3281 if (GCExpandToAllocateDelayMillis > 0) { 3282 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 3283 } 3284 } 3285 } 3286 3287 3288 bool ConcurrentMarkSweepGeneration::expand_and_ensure_spooling_space( 3289 PromotionInfo* promo) { 3290 MutexLocker x(ParGCRareEvent_lock); 3291 size_t refill_size_bytes = promo->refillSize() * HeapWordSize; 3292 while (true) { 3293 // Expansion by some other thread might make alloc OK now: 3294 if (promo->ensure_spooling_space()) { 3295 assert(promo->has_spooling_space(), 3296 "Post-condition of successful ensure_spooling_space()"); 3297 return true; 3298 } 3299 // If there's not enough expansion space available, give up. 3300 if (_virtual_space.uncommitted_size() < refill_size_bytes) { 3301 return false; 3302 } 3303 // Otherwise, we try expansion. 3304 expand(refill_size_bytes, MinHeapDeltaBytes, 3305 CMSExpansionCause::_allocate_par_spooling_space); 3306 // Now go around the loop and try alloc again; 3307 // A competing allocation might beat us to the expansion space, 3308 // so we may go around the loop again if allocation fails again. 3309 if (GCExpandToAllocateDelayMillis > 0) { 3310 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 3311 } 3312 } 3313 } 3314 3315 3316 3317 void ConcurrentMarkSweepGeneration::shrink(size_t bytes) { 3318 assert_locked_or_safepoint(Heap_lock); 3319 size_t size = ReservedSpace::page_align_size_down(bytes); 3320 if (size > 0) { 3321 shrink_by(size); 3322 } 3323 } 3324 3325 bool ConcurrentMarkSweepGeneration::grow_by(size_t bytes) { 3326 assert_locked_or_safepoint(Heap_lock); 3327 bool result = _virtual_space.expand_by(bytes); 3328 if (result) { 3329 HeapWord* old_end = _cmsSpace->end(); 3330 size_t new_word_size = 3331 heap_word_size(_virtual_space.committed_size()); 3332 MemRegion mr(_cmsSpace->bottom(), new_word_size); 3333 _bts->resize(new_word_size); // resize the block offset shared array 3334 Universe::heap()->barrier_set()->resize_covered_region(mr); 3335 // Hmmmm... why doesn't CFLS::set_end verify locking? 3336 // This is quite ugly; FIX ME XXX 3337 _cmsSpace->assert_locked(freelistLock()); 3338 _cmsSpace->set_end((HeapWord*)_virtual_space.high()); 3339 3340 // update the space and generation capacity counters 3341 if (UsePerfData) { 3342 _space_counters->update_capacity(); 3343 _gen_counters->update_all(); 3344 } 3345 3346 if (Verbose && PrintGC) { 3347 size_t new_mem_size = _virtual_space.committed_size(); 3348 size_t old_mem_size = new_mem_size - bytes; 3349 gclog_or_tty->print_cr("Expanding %s from %ldK by %ldK to %ldK", 3350 name(), old_mem_size/K, bytes/K, new_mem_size/K); 3351 } 3352 } 3353 return result; 3354 } 3355 3356 bool ConcurrentMarkSweepGeneration::grow_to_reserved() { 3357 assert_locked_or_safepoint(Heap_lock); 3358 bool success = true; 3359 const size_t remaining_bytes = _virtual_space.uncommitted_size(); 3360 if (remaining_bytes > 0) { 3361 success = grow_by(remaining_bytes); 3362 DEBUG_ONLY(if (!success) warning("grow to reserved failed");) 3363 } 3364 return success; 3365 } 3366 3367 void ConcurrentMarkSweepGeneration::shrink_by(size_t bytes) { 3368 assert_locked_or_safepoint(Heap_lock); 3369 assert_lock_strong(freelistLock()); 3370 // XXX Fix when compaction is implemented. 3371 warning("Shrinking of CMS not yet implemented"); 3372 return; 3373 } 3374 3375 3376 // Simple ctor/dtor wrapper for accounting & timer chores around concurrent 3377 // phases. 3378 class CMSPhaseAccounting: public StackObj { 3379 public: 3380 CMSPhaseAccounting(CMSCollector *collector, 3381 const char *phase, 3382 bool print_cr = true); 3383 ~CMSPhaseAccounting(); 3384 3385 private: 3386 CMSCollector *_collector; 3387 const char *_phase; 3388 elapsedTimer _wallclock; 3389 bool _print_cr; 3390 3391 public: 3392 // Not MT-safe; so do not pass around these StackObj's 3393 // where they may be accessed by other threads. 3394 jlong wallclock_millis() { 3395 assert(_wallclock.is_active(), "Wall clock should not stop"); 3396 _wallclock.stop(); // to record time 3397 jlong ret = _wallclock.milliseconds(); 3398 _wallclock.start(); // restart 3399 return ret; 3400 } 3401 }; 3402 3403 CMSPhaseAccounting::CMSPhaseAccounting(CMSCollector *collector, 3404 const char *phase, 3405 bool print_cr) : 3406 _collector(collector), _phase(phase), _print_cr(print_cr) { 3407 3408 if (PrintCMSStatistics != 0) { 3409 _collector->resetYields(); 3410 } 3411 if (PrintGCDetails && PrintGCTimeStamps) { 3412 gclog_or_tty->date_stamp(PrintGCDateStamps); 3413 gclog_or_tty->stamp(); 3414 gclog_or_tty->print_cr(": [%s-concurrent-%s-start]", 3415 _collector->cmsGen()->short_name(), _phase); 3416 } 3417 _collector->resetTimer(); 3418 _wallclock.start(); 3419 _collector->startTimer(); 3420 } 3421 3422 CMSPhaseAccounting::~CMSPhaseAccounting() { 3423 assert(_wallclock.is_active(), "Wall clock should not have stopped"); 3424 _collector->stopTimer(); 3425 _wallclock.stop(); 3426 if (PrintGCDetails) { 3427 gclog_or_tty->date_stamp(PrintGCDateStamps); 3428 gclog_or_tty->stamp(PrintGCTimeStamps); 3429 gclog_or_tty->print("[%s-concurrent-%s: %3.3f/%3.3f secs]", 3430 _collector->cmsGen()->short_name(), 3431 _phase, _collector->timerValue(), _wallclock.seconds()); 3432 if (_print_cr) { 3433 gclog_or_tty->print_cr(""); 3434 } 3435 if (PrintCMSStatistics != 0) { 3436 gclog_or_tty->print_cr(" (CMS-concurrent-%s yielded %d times)", _phase, 3437 _collector->yields()); 3438 } 3439 } 3440 } 3441 3442 // CMS work 3443 3444 // Checkpoint the roots into this generation from outside 3445 // this generation. [Note this initial checkpoint need only 3446 // be approximate -- we'll do a catch up phase subsequently.] 3447 void CMSCollector::checkpointRootsInitial(bool asynch) { 3448 assert(_collectorState == InitialMarking, "Wrong collector state"); 3449 check_correct_thread_executing(); 3450 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 3451 3452 ReferenceProcessor* rp = ref_processor(); 3453 SpecializationStats::clear(); 3454 assert(_restart_addr == NULL, "Control point invariant"); 3455 if (asynch) { 3456 // acquire locks for subsequent manipulations 3457 MutexLockerEx x(bitMapLock(), 3458 Mutex::_no_safepoint_check_flag); 3459 checkpointRootsInitialWork(asynch); 3460 // enable ("weak") refs discovery 3461 rp->enable_discovery(true /*verify_disabled*/, true /*check_no_refs*/); 3462 _collectorState = Marking; 3463 } else { 3464 // (Weak) Refs discovery: this is controlled from genCollectedHeap::do_collection 3465 // which recognizes if we are a CMS generation, and doesn't try to turn on 3466 // discovery; verify that they aren't meddling. 3467 assert(!rp->discovery_is_atomic(), 3468 "incorrect setting of discovery predicate"); 3469 assert(!rp->discovery_enabled(), "genCollectedHeap shouldn't control " 3470 "ref discovery for this generation kind"); 3471 // already have locks 3472 checkpointRootsInitialWork(asynch); 3473 // now enable ("weak") refs discovery 3474 rp->enable_discovery(true /*verify_disabled*/, false /*verify_no_refs*/); 3475 _collectorState = Marking; 3476 } 3477 SpecializationStats::print(); 3478 } 3479 3480 void CMSCollector::checkpointRootsInitialWork(bool asynch) { 3481 assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped"); 3482 assert(_collectorState == InitialMarking, "just checking"); 3483 3484 // If there has not been a GC[n-1] since last GC[n] cycle completed, 3485 // precede our marking with a collection of all 3486 // younger generations to keep floating garbage to a minimum. 3487 // XXX: we won't do this for now -- it's an optimization to be done later. 3488 3489 // already have locks 3490 assert_lock_strong(bitMapLock()); 3491 assert(_markBitMap.isAllClear(), "was reset at end of previous cycle"); 3492 3493 // Setup the verification and class unloading state for this 3494 // CMS collection cycle. 3495 setup_cms_unloading_and_verification_state(); 3496 3497 NOT_PRODUCT(TraceTime t("\ncheckpointRootsInitialWork", 3498 PrintGCDetails && Verbose, true, gclog_or_tty);) 3499 if (UseAdaptiveSizePolicy) { 3500 size_policy()->checkpoint_roots_initial_begin(); 3501 } 3502 3503 // Reset all the PLAB chunk arrays if necessary. 3504 if (_survivor_plab_array != NULL && !CMSPLABRecordAlways) { 3505 reset_survivor_plab_arrays(); 3506 } 3507 3508 ResourceMark rm; 3509 HandleMark hm; 3510 3511 FalseClosure falseClosure; 3512 // In the case of a synchronous collection, we will elide the 3513 // remark step, so it's important to catch all the nmethod oops 3514 // in this step. 3515 // The final 'true' flag to gen_process_strong_roots will ensure this. 3516 // If 'async' is true, we can relax the nmethod tracing. 3517 MarkRefsIntoClosure notOlder(_span, &_markBitMap); 3518 GenCollectedHeap* gch = GenCollectedHeap::heap(); 3519 3520 verify_work_stacks_empty(); 3521 verify_overflow_empty(); 3522 3523 gch->ensure_parsability(false); // fill TLABs, but no need to retire them 3524 // Update the saved marks which may affect the root scans. 3525 gch->save_marks(); 3526 3527 // weak reference processing has not started yet. 3528 ref_processor()->set_enqueuing_is_done(false); 3529 3530 // Need to remember all newly created CLDs, 3531 // so that we can guarantee that the remark finds them. 3532 ClassLoaderDataGraph::remember_new_clds(true); 3533 3534 // Whenever a CLD is found, it will be claimed before proceeding to mark 3535 // the klasses. The claimed marks need to be cleared before marking starts. 3536 ClassLoaderDataGraph::clear_claimed_marks(); 3537 3538 CMKlassClosure klass_closure(¬Older); 3539 { 3540 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;) 3541 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 3542 gch->gen_process_strong_roots(_cmsGen->level(), 3543 true, // younger gens are roots 3544 true, // activate StrongRootsScope 3545 false, // not scavenging 3546 SharedHeap::ScanningOption(roots_scanning_options()), 3547 ¬Older, 3548 true, // walk all of code cache if (so & SO_CodeCache) 3549 NULL, 3550 &klass_closure); 3551 } 3552 3553 // Clear mod-union table; it will be dirtied in the prologue of 3554 // CMS generation per each younger generation collection. 3555 3556 assert(_modUnionTable.isAllClear(), 3557 "Was cleared in most recent final checkpoint phase" 3558 " or no bits are set in the gc_prologue before the start of the next " 3559 "subsequent marking phase."); 3560 3561 assert(_ct->klass_rem_set()->mod_union_is_clear(), "Must be"); 3562 3563 // Save the end of the used_region of the constituent generations 3564 // to be used to limit the extent of sweep in each generation. 3565 save_sweep_limits(); 3566 if (UseAdaptiveSizePolicy) { 3567 size_policy()->checkpoint_roots_initial_end(gch->gc_cause()); 3568 } 3569 verify_overflow_empty(); 3570 } 3571 3572 bool CMSCollector::markFromRoots(bool asynch) { 3573 // we might be tempted to assert that: 3574 // assert(asynch == !SafepointSynchronize::is_at_safepoint(), 3575 // "inconsistent argument?"); 3576 // However that wouldn't be right, because it's possible that 3577 // a safepoint is indeed in progress as a younger generation 3578 // stop-the-world GC happens even as we mark in this generation. 3579 assert(_collectorState == Marking, "inconsistent state?"); 3580 check_correct_thread_executing(); 3581 verify_overflow_empty(); 3582 3583 bool res; 3584 if (asynch) { 3585 3586 // Start the timers for adaptive size policy for the concurrent phases 3587 // Do it here so that the foreground MS can use the concurrent 3588 // timer since a foreground MS might has the sweep done concurrently 3589 // or STW. 3590 if (UseAdaptiveSizePolicy) { 3591 size_policy()->concurrent_marking_begin(); 3592 } 3593 3594 // Weak ref discovery note: We may be discovering weak 3595 // refs in this generation concurrent (but interleaved) with 3596 // weak ref discovery by a younger generation collector. 3597 3598 CMSTokenSyncWithLocks ts(true, bitMapLock()); 3599 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 3600 CMSPhaseAccounting pa(this, "mark", !PrintGCDetails); 3601 res = markFromRootsWork(asynch); 3602 if (res) { 3603 _collectorState = Precleaning; 3604 } else { // We failed and a foreground collection wants to take over 3605 assert(_foregroundGCIsActive, "internal state inconsistency"); 3606 assert(_restart_addr == NULL, "foreground will restart from scratch"); 3607 if (PrintGCDetails) { 3608 gclog_or_tty->print_cr("bailing out to foreground collection"); 3609 } 3610 } 3611 if (UseAdaptiveSizePolicy) { 3612 size_policy()->concurrent_marking_end(); 3613 } 3614 } else { 3615 assert(SafepointSynchronize::is_at_safepoint(), 3616 "inconsistent with asynch == false"); 3617 if (UseAdaptiveSizePolicy) { 3618 size_policy()->ms_collection_marking_begin(); 3619 } 3620 // already have locks 3621 res = markFromRootsWork(asynch); 3622 _collectorState = FinalMarking; 3623 if (UseAdaptiveSizePolicy) { 3624 GenCollectedHeap* gch = GenCollectedHeap::heap(); 3625 size_policy()->ms_collection_marking_end(gch->gc_cause()); 3626 } 3627 } 3628 verify_overflow_empty(); 3629 return res; 3630 } 3631 3632 bool CMSCollector::markFromRootsWork(bool asynch) { 3633 // iterate over marked bits in bit map, doing a full scan and mark 3634 // from these roots using the following algorithm: 3635 // . if oop is to the right of the current scan pointer, 3636 // mark corresponding bit (we'll process it later) 3637 // . else (oop is to left of current scan pointer) 3638 // push oop on marking stack 3639 // . drain the marking stack 3640 3641 // Note that when we do a marking step we need to hold the 3642 // bit map lock -- recall that direct allocation (by mutators) 3643 // and promotion (by younger generation collectors) is also 3644 // marking the bit map. [the so-called allocate live policy.] 3645 // Because the implementation of bit map marking is not 3646 // robust wrt simultaneous marking of bits in the same word, 3647 // we need to make sure that there is no such interference 3648 // between concurrent such updates. 3649 3650 // already have locks 3651 assert_lock_strong(bitMapLock()); 3652 3653 verify_work_stacks_empty(); 3654 verify_overflow_empty(); 3655 bool result = false; 3656 if (CMSConcurrentMTEnabled && ConcGCThreads > 0) { 3657 result = do_marking_mt(asynch); 3658 } else { 3659 result = do_marking_st(asynch); 3660 } 3661 return result; 3662 } 3663 3664 // Forward decl 3665 class CMSConcMarkingTask; 3666 3667 class CMSConcMarkingTerminator: public ParallelTaskTerminator { 3668 CMSCollector* _collector; 3669 CMSConcMarkingTask* _task; 3670 public: 3671 virtual void yield(); 3672 3673 // "n_threads" is the number of threads to be terminated. 3674 // "queue_set" is a set of work queues of other threads. 3675 // "collector" is the CMS collector associated with this task terminator. 3676 // "yield" indicates whether we need the gang as a whole to yield. 3677 CMSConcMarkingTerminator(int n_threads, TaskQueueSetSuper* queue_set, CMSCollector* collector) : 3678 ParallelTaskTerminator(n_threads, queue_set), 3679 _collector(collector) { } 3680 3681 void set_task(CMSConcMarkingTask* task) { 3682 _task = task; 3683 } 3684 }; 3685 3686 class CMSConcMarkingTerminatorTerminator: public TerminatorTerminator { 3687 CMSConcMarkingTask* _task; 3688 public: 3689 bool should_exit_termination(); 3690 void set_task(CMSConcMarkingTask* task) { 3691 _task = task; 3692 } 3693 }; 3694 3695 // MT Concurrent Marking Task 3696 class CMSConcMarkingTask: public YieldingFlexibleGangTask { 3697 CMSCollector* _collector; 3698 int _n_workers; // requested/desired # workers 3699 bool _asynch; 3700 bool _result; 3701 CompactibleFreeListSpace* _cms_space; 3702 char _pad_front[64]; // padding to ... 3703 HeapWord* _global_finger; // ... avoid sharing cache line 3704 char _pad_back[64]; 3705 HeapWord* _restart_addr; 3706 3707 // Exposed here for yielding support 3708 Mutex* const _bit_map_lock; 3709 3710 // The per thread work queues, available here for stealing 3711 OopTaskQueueSet* _task_queues; 3712 3713 // Termination (and yielding) support 3714 CMSConcMarkingTerminator _term; 3715 CMSConcMarkingTerminatorTerminator _term_term; 3716 3717 public: 3718 CMSConcMarkingTask(CMSCollector* collector, 3719 CompactibleFreeListSpace* cms_space, 3720 bool asynch, 3721 YieldingFlexibleWorkGang* workers, 3722 OopTaskQueueSet* task_queues): 3723 YieldingFlexibleGangTask("Concurrent marking done multi-threaded"), 3724 _collector(collector), 3725 _cms_space(cms_space), 3726 _asynch(asynch), _n_workers(0), _result(true), 3727 _task_queues(task_queues), 3728 _term(_n_workers, task_queues, _collector), 3729 _bit_map_lock(collector->bitMapLock()) 3730 { 3731 _requested_size = _n_workers; 3732 _term.set_task(this); 3733 _term_term.set_task(this); 3734 _restart_addr = _global_finger = _cms_space->bottom(); 3735 } 3736 3737 3738 OopTaskQueueSet* task_queues() { return _task_queues; } 3739 3740 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 3741 3742 HeapWord** global_finger_addr() { return &_global_finger; } 3743 3744 CMSConcMarkingTerminator* terminator() { return &_term; } 3745 3746 virtual void set_for_termination(int active_workers) { 3747 terminator()->reset_for_reuse(active_workers); 3748 } 3749 3750 void work(uint worker_id); 3751 bool should_yield() { 3752 return ConcurrentMarkSweepThread::should_yield() 3753 && !_collector->foregroundGCIsActive() 3754 && _asynch; 3755 } 3756 3757 virtual void coordinator_yield(); // stuff done by coordinator 3758 bool result() { return _result; } 3759 3760 void reset(HeapWord* ra) { 3761 assert(_global_finger >= _cms_space->end(), "Postcondition of ::work(i)"); 3762 _restart_addr = _global_finger = ra; 3763 _term.reset_for_reuse(); 3764 } 3765 3766 static bool get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 3767 OopTaskQueue* work_q); 3768 3769 private: 3770 void do_scan_and_mark(int i, CompactibleFreeListSpace* sp); 3771 void do_work_steal(int i); 3772 void bump_global_finger(HeapWord* f); 3773 }; 3774 3775 bool CMSConcMarkingTerminatorTerminator::should_exit_termination() { 3776 assert(_task != NULL, "Error"); 3777 return _task->yielding(); 3778 // Note that we do not need the disjunct || _task->should_yield() above 3779 // because we want terminating threads to yield only if the task 3780 // is already in the midst of yielding, which happens only after at least one 3781 // thread has yielded. 3782 } 3783 3784 void CMSConcMarkingTerminator::yield() { 3785 if (_task->should_yield()) { 3786 _task->yield(); 3787 } else { 3788 ParallelTaskTerminator::yield(); 3789 } 3790 } 3791 3792 //////////////////////////////////////////////////////////////// 3793 // Concurrent Marking Algorithm Sketch 3794 //////////////////////////////////////////////////////////////// 3795 // Until all tasks exhausted (both spaces): 3796 // -- claim next available chunk 3797 // -- bump global finger via CAS 3798 // -- find first object that starts in this chunk 3799 // and start scanning bitmap from that position 3800 // -- scan marked objects for oops 3801 // -- CAS-mark target, and if successful: 3802 // . if target oop is above global finger (volatile read) 3803 // nothing to do 3804 // . if target oop is in chunk and above local finger 3805 // then nothing to do 3806 // . else push on work-queue 3807 // -- Deal with possible overflow issues: 3808 // . local work-queue overflow causes stuff to be pushed on 3809 // global (common) overflow queue 3810 // . always first empty local work queue 3811 // . then get a batch of oops from global work queue if any 3812 // . then do work stealing 3813 // -- When all tasks claimed (both spaces) 3814 // and local work queue empty, 3815 // then in a loop do: 3816 // . check global overflow stack; steal a batch of oops and trace 3817 // . try to steal from other threads oif GOS is empty 3818 // . if neither is available, offer termination 3819 // -- Terminate and return result 3820 // 3821 void CMSConcMarkingTask::work(uint worker_id) { 3822 elapsedTimer _timer; 3823 ResourceMark rm; 3824 HandleMark hm; 3825 3826 DEBUG_ONLY(_collector->verify_overflow_empty();) 3827 3828 // Before we begin work, our work queue should be empty 3829 assert(work_queue(worker_id)->size() == 0, "Expected to be empty"); 3830 // Scan the bitmap covering _cms_space, tracing through grey objects. 3831 _timer.start(); 3832 do_scan_and_mark(worker_id, _cms_space); 3833 _timer.stop(); 3834 if (PrintCMSStatistics != 0) { 3835 gclog_or_tty->print_cr("Finished cms space scanning in %dth thread: %3.3f sec", 3836 worker_id, _timer.seconds()); 3837 // XXX: need xxx/xxx type of notation, two timers 3838 } 3839 3840 // ... do work stealing 3841 _timer.reset(); 3842 _timer.start(); 3843 do_work_steal(worker_id); 3844 _timer.stop(); 3845 if (PrintCMSStatistics != 0) { 3846 gclog_or_tty->print_cr("Finished work stealing in %dth thread: %3.3f sec", 3847 worker_id, _timer.seconds()); 3848 // XXX: need xxx/xxx type of notation, two timers 3849 } 3850 assert(_collector->_markStack.isEmpty(), "Should have been emptied"); 3851 assert(work_queue(worker_id)->size() == 0, "Should have been emptied"); 3852 // Note that under the current task protocol, the 3853 // following assertion is true even of the spaces 3854 // expanded since the completion of the concurrent 3855 // marking. XXX This will likely change under a strict 3856 // ABORT semantics. 3857 // After perm removal the comparison was changed to 3858 // greater than or equal to from strictly greater than. 3859 // Before perm removal the highest address sweep would 3860 // have been at the end of perm gen but now is at the 3861 // end of the tenured gen. 3862 assert(_global_finger >= _cms_space->end(), 3863 "All tasks have been completed"); 3864 DEBUG_ONLY(_collector->verify_overflow_empty();) 3865 } 3866 3867 void CMSConcMarkingTask::bump_global_finger(HeapWord* f) { 3868 HeapWord* read = _global_finger; 3869 HeapWord* cur = read; 3870 while (f > read) { 3871 cur = read; 3872 read = (HeapWord*) Atomic::cmpxchg_ptr(f, &_global_finger, cur); 3873 if (cur == read) { 3874 // our cas succeeded 3875 assert(_global_finger >= f, "protocol consistency"); 3876 break; 3877 } 3878 } 3879 } 3880 3881 // This is really inefficient, and should be redone by 3882 // using (not yet available) block-read and -write interfaces to the 3883 // stack and the work_queue. XXX FIX ME !!! 3884 bool CMSConcMarkingTask::get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 3885 OopTaskQueue* work_q) { 3886 // Fast lock-free check 3887 if (ovflw_stk->length() == 0) { 3888 return false; 3889 } 3890 assert(work_q->size() == 0, "Shouldn't steal"); 3891 MutexLockerEx ml(ovflw_stk->par_lock(), 3892 Mutex::_no_safepoint_check_flag); 3893 // Grab up to 1/4 the size of the work queue 3894 size_t num = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 3895 (size_t)ParGCDesiredObjsFromOverflowList); 3896 num = MIN2(num, ovflw_stk->length()); 3897 for (int i = (int) num; i > 0; i--) { 3898 oop cur = ovflw_stk->pop(); 3899 assert(cur != NULL, "Counted wrong?"); 3900 work_q->push(cur); 3901 } 3902 return num > 0; 3903 } 3904 3905 void CMSConcMarkingTask::do_scan_and_mark(int i, CompactibleFreeListSpace* sp) { 3906 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 3907 int n_tasks = pst->n_tasks(); 3908 // We allow that there may be no tasks to do here because 3909 // we are restarting after a stack overflow. 3910 assert(pst->valid() || n_tasks == 0, "Uninitialized use?"); 3911 uint nth_task = 0; 3912 3913 HeapWord* aligned_start = sp->bottom(); 3914 if (sp->used_region().contains(_restart_addr)) { 3915 // Align down to a card boundary for the start of 0th task 3916 // for this space. 3917 aligned_start = 3918 (HeapWord*)align_size_down((uintptr_t)_restart_addr, 3919 CardTableModRefBS::card_size); 3920 } 3921 3922 size_t chunk_size = sp->marking_task_size(); 3923 while (!pst->is_task_claimed(/* reference */ nth_task)) { 3924 // Having claimed the nth task in this space, 3925 // compute the chunk that it corresponds to: 3926 MemRegion span = MemRegion(aligned_start + nth_task*chunk_size, 3927 aligned_start + (nth_task+1)*chunk_size); 3928 // Try and bump the global finger via a CAS; 3929 // note that we need to do the global finger bump 3930 // _before_ taking the intersection below, because 3931 // the task corresponding to that region will be 3932 // deemed done even if the used_region() expands 3933 // because of allocation -- as it almost certainly will 3934 // during start-up while the threads yield in the 3935 // closure below. 3936 HeapWord* finger = span.end(); 3937 bump_global_finger(finger); // atomically 3938 // There are null tasks here corresponding to chunks 3939 // beyond the "top" address of the space. 3940 span = span.intersection(sp->used_region()); 3941 if (!span.is_empty()) { // Non-null task 3942 HeapWord* prev_obj; 3943 assert(!span.contains(_restart_addr) || nth_task == 0, 3944 "Inconsistency"); 3945 if (nth_task == 0) { 3946 // For the 0th task, we'll not need to compute a block_start. 3947 if (span.contains(_restart_addr)) { 3948 // In the case of a restart because of stack overflow, 3949 // we might additionally skip a chunk prefix. 3950 prev_obj = _restart_addr; 3951 } else { 3952 prev_obj = span.start(); 3953 } 3954 } else { 3955 // We want to skip the first object because 3956 // the protocol is to scan any object in its entirety 3957 // that _starts_ in this span; a fortiori, any 3958 // object starting in an earlier span is scanned 3959 // as part of an earlier claimed task. 3960 // Below we use the "careful" version of block_start 3961 // so we do not try to navigate uninitialized objects. 3962 prev_obj = sp->block_start_careful(span.start()); 3963 // Below we use a variant of block_size that uses the 3964 // Printezis bits to avoid waiting for allocated 3965 // objects to become initialized/parsable. 3966 while (prev_obj < span.start()) { 3967 size_t sz = sp->block_size_no_stall(prev_obj, _collector); 3968 if (sz > 0) { 3969 prev_obj += sz; 3970 } else { 3971 // In this case we may end up doing a bit of redundant 3972 // scanning, but that appears unavoidable, short of 3973 // locking the free list locks; see bug 6324141. 3974 break; 3975 } 3976 } 3977 } 3978 if (prev_obj < span.end()) { 3979 MemRegion my_span = MemRegion(prev_obj, span.end()); 3980 // Do the marking work within a non-empty span -- 3981 // the last argument to the constructor indicates whether the 3982 // iteration should be incremental with periodic yields. 3983 Par_MarkFromRootsClosure cl(this, _collector, my_span, 3984 &_collector->_markBitMap, 3985 work_queue(i), 3986 &_collector->_markStack, 3987 _asynch); 3988 _collector->_markBitMap.iterate(&cl, my_span.start(), my_span.end()); 3989 } // else nothing to do for this task 3990 } // else nothing to do for this task 3991 } 3992 // We'd be tempted to assert here that since there are no 3993 // more tasks left to claim in this space, the global_finger 3994 // must exceed space->top() and a fortiori space->end(). However, 3995 // that would not quite be correct because the bumping of 3996 // global_finger occurs strictly after the claiming of a task, 3997 // so by the time we reach here the global finger may not yet 3998 // have been bumped up by the thread that claimed the last 3999 // task. 4000 pst->all_tasks_completed(); 4001 } 4002 4003 class Par_ConcMarkingClosure: public CMSOopClosure { 4004 private: 4005 CMSCollector* _collector; 4006 CMSConcMarkingTask* _task; 4007 MemRegion _span; 4008 CMSBitMap* _bit_map; 4009 CMSMarkStack* _overflow_stack; 4010 OopTaskQueue* _work_queue; 4011 protected: 4012 DO_OOP_WORK_DEFN 4013 public: 4014 Par_ConcMarkingClosure(CMSCollector* collector, CMSConcMarkingTask* task, OopTaskQueue* work_queue, 4015 CMSBitMap* bit_map, CMSMarkStack* overflow_stack): 4016 CMSOopClosure(collector->ref_processor()), 4017 _collector(collector), 4018 _task(task), 4019 _span(collector->_span), 4020 _work_queue(work_queue), 4021 _bit_map(bit_map), 4022 _overflow_stack(overflow_stack) 4023 { } 4024 virtual void do_oop(oop* p); 4025 virtual void do_oop(narrowOop* p); 4026 4027 void trim_queue(size_t max); 4028 void handle_stack_overflow(HeapWord* lost); 4029 void do_yield_check() { 4030 if (_task->should_yield()) { 4031 _task->yield(); 4032 } 4033 } 4034 }; 4035 4036 // Grey object scanning during work stealing phase -- 4037 // the salient assumption here is that any references 4038 // that are in these stolen objects being scanned must 4039 // already have been initialized (else they would not have 4040 // been published), so we do not need to check for 4041 // uninitialized objects before pushing here. 4042 void Par_ConcMarkingClosure::do_oop(oop obj) { 4043 assert(obj->is_oop_or_null(true), "expected an oop or NULL"); 4044 HeapWord* addr = (HeapWord*)obj; 4045 // Check if oop points into the CMS generation 4046 // and is not marked 4047 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 4048 // a white object ... 4049 // If we manage to "claim" the object, by being the 4050 // first thread to mark it, then we push it on our 4051 // marking stack 4052 if (_bit_map->par_mark(addr)) { // ... now grey 4053 // push on work queue (grey set) 4054 bool simulate_overflow = false; 4055 NOT_PRODUCT( 4056 if (CMSMarkStackOverflowALot && 4057 _collector->simulate_overflow()) { 4058 // simulate a stack overflow 4059 simulate_overflow = true; 4060 } 4061 ) 4062 if (simulate_overflow || 4063 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) { 4064 // stack overflow 4065 if (PrintCMSStatistics != 0) { 4066 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 4067 SIZE_FORMAT, _overflow_stack->capacity()); 4068 } 4069 // We cannot assert that the overflow stack is full because 4070 // it may have been emptied since. 4071 assert(simulate_overflow || 4072 _work_queue->size() == _work_queue->max_elems(), 4073 "Else push should have succeeded"); 4074 handle_stack_overflow(addr); 4075 } 4076 } // Else, some other thread got there first 4077 do_yield_check(); 4078 } 4079 } 4080 4081 void Par_ConcMarkingClosure::do_oop(oop* p) { Par_ConcMarkingClosure::do_oop_work(p); } 4082 void Par_ConcMarkingClosure::do_oop(narrowOop* p) { Par_ConcMarkingClosure::do_oop_work(p); } 4083 4084 void Par_ConcMarkingClosure::trim_queue(size_t max) { 4085 while (_work_queue->size() > max) { 4086 oop new_oop; 4087 if (_work_queue->pop_local(new_oop)) { 4088 assert(new_oop->is_oop(), "Should be an oop"); 4089 assert(_bit_map->isMarked((HeapWord*)new_oop), "Grey object"); 4090 assert(_span.contains((HeapWord*)new_oop), "Not in span"); 4091 new_oop->oop_iterate(this); // do_oop() above 4092 do_yield_check(); 4093 } 4094 } 4095 } 4096 4097 // Upon stack overflow, we discard (part of) the stack, 4098 // remembering the least address amongst those discarded 4099 // in CMSCollector's _restart_address. 4100 void Par_ConcMarkingClosure::handle_stack_overflow(HeapWord* lost) { 4101 // We need to do this under a mutex to prevent other 4102 // workers from interfering with the work done below. 4103 MutexLockerEx ml(_overflow_stack->par_lock(), 4104 Mutex::_no_safepoint_check_flag); 4105 // Remember the least grey address discarded 4106 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 4107 _collector->lower_restart_addr(ra); 4108 _overflow_stack->reset(); // discard stack contents 4109 _overflow_stack->expand(); // expand the stack if possible 4110 } 4111 4112 4113 void CMSConcMarkingTask::do_work_steal(int i) { 4114 OopTaskQueue* work_q = work_queue(i); 4115 oop obj_to_scan; 4116 CMSBitMap* bm = &(_collector->_markBitMap); 4117 CMSMarkStack* ovflw = &(_collector->_markStack); 4118 int* seed = _collector->hash_seed(i); 4119 Par_ConcMarkingClosure cl(_collector, this, work_q, bm, ovflw); 4120 while (true) { 4121 cl.trim_queue(0); 4122 assert(work_q->size() == 0, "Should have been emptied above"); 4123 if (get_work_from_overflow_stack(ovflw, work_q)) { 4124 // Can't assert below because the work obtained from the 4125 // overflow stack may already have been stolen from us. 4126 // assert(work_q->size() > 0, "Work from overflow stack"); 4127 continue; 4128 } else if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 4129 assert(obj_to_scan->is_oop(), "Should be an oop"); 4130 assert(bm->isMarked((HeapWord*)obj_to_scan), "Grey object"); 4131 obj_to_scan->oop_iterate(&cl); 4132 } else if (terminator()->offer_termination(&_term_term)) { 4133 assert(work_q->size() == 0, "Impossible!"); 4134 break; 4135 } else if (yielding() || should_yield()) { 4136 yield(); 4137 } 4138 } 4139 } 4140 4141 // This is run by the CMS (coordinator) thread. 4142 void CMSConcMarkingTask::coordinator_yield() { 4143 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 4144 "CMS thread should hold CMS token"); 4145 // First give up the locks, then yield, then re-lock 4146 // We should probably use a constructor/destructor idiom to 4147 // do this unlock/lock or modify the MutexUnlocker class to 4148 // serve our purpose. XXX 4149 assert_lock_strong(_bit_map_lock); 4150 _bit_map_lock->unlock(); 4151 ConcurrentMarkSweepThread::desynchronize(true); 4152 ConcurrentMarkSweepThread::acknowledge_yield_request(); 4153 _collector->stopTimer(); 4154 if (PrintCMSStatistics != 0) { 4155 _collector->incrementYields(); 4156 } 4157 _collector->icms_wait(); 4158 4159 // It is possible for whichever thread initiated the yield request 4160 // not to get a chance to wake up and take the bitmap lock between 4161 // this thread releasing it and reacquiring it. So, while the 4162 // should_yield() flag is on, let's sleep for a bit to give the 4163 // other thread a chance to wake up. The limit imposed on the number 4164 // of iterations is defensive, to avoid any unforseen circumstances 4165 // putting us into an infinite loop. Since it's always been this 4166 // (coordinator_yield()) method that was observed to cause the 4167 // problem, we are using a parameter (CMSCoordinatorYieldSleepCount) 4168 // which is by default non-zero. For the other seven methods that 4169 // also perform the yield operation, as are using a different 4170 // parameter (CMSYieldSleepCount) which is by default zero. This way we 4171 // can enable the sleeping for those methods too, if necessary. 4172 // See 6442774. 4173 // 4174 // We really need to reconsider the synchronization between the GC 4175 // thread and the yield-requesting threads in the future and we 4176 // should really use wait/notify, which is the recommended 4177 // way of doing this type of interaction. Additionally, we should 4178 // consolidate the eight methods that do the yield operation and they 4179 // are almost identical into one for better maintenability and 4180 // readability. See 6445193. 4181 // 4182 // Tony 2006.06.29 4183 for (unsigned i = 0; i < CMSCoordinatorYieldSleepCount && 4184 ConcurrentMarkSweepThread::should_yield() && 4185 !CMSCollector::foregroundGCIsActive(); ++i) { 4186 os::sleep(Thread::current(), 1, false); 4187 ConcurrentMarkSweepThread::acknowledge_yield_request(); 4188 } 4189 4190 ConcurrentMarkSweepThread::synchronize(true); 4191 _bit_map_lock->lock_without_safepoint_check(); 4192 _collector->startTimer(); 4193 } 4194 4195 bool CMSCollector::do_marking_mt(bool asynch) { 4196 assert(ConcGCThreads > 0 && conc_workers() != NULL, "precondition"); 4197 int num_workers = AdaptiveSizePolicy::calc_active_conc_workers( 4198 conc_workers()->total_workers(), 4199 conc_workers()->active_workers(), 4200 Threads::number_of_non_daemon_threads()); 4201 conc_workers()->set_active_workers(num_workers); 4202 4203 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 4204 4205 CMSConcMarkingTask tsk(this, 4206 cms_space, 4207 asynch, 4208 conc_workers(), 4209 task_queues()); 4210 4211 // Since the actual number of workers we get may be different 4212 // from the number we requested above, do we need to do anything different 4213 // below? In particular, may be we need to subclass the SequantialSubTasksDone 4214 // class?? XXX 4215 cms_space ->initialize_sequential_subtasks_for_marking(num_workers); 4216 4217 // Refs discovery is already non-atomic. 4218 assert(!ref_processor()->discovery_is_atomic(), "Should be non-atomic"); 4219 assert(ref_processor()->discovery_is_mt(), "Discovery should be MT"); 4220 conc_workers()->start_task(&tsk); 4221 while (tsk.yielded()) { 4222 tsk.coordinator_yield(); 4223 conc_workers()->continue_task(&tsk); 4224 } 4225 // If the task was aborted, _restart_addr will be non-NULL 4226 assert(tsk.completed() || _restart_addr != NULL, "Inconsistency"); 4227 while (_restart_addr != NULL) { 4228 // XXX For now we do not make use of ABORTED state and have not 4229 // yet implemented the right abort semantics (even in the original 4230 // single-threaded CMS case). That needs some more investigation 4231 // and is deferred for now; see CR# TBF. 07252005YSR. XXX 4232 assert(!CMSAbortSemantics || tsk.aborted(), "Inconsistency"); 4233 // If _restart_addr is non-NULL, a marking stack overflow 4234 // occurred; we need to do a fresh marking iteration from the 4235 // indicated restart address. 4236 if (_foregroundGCIsActive && asynch) { 4237 // We may be running into repeated stack overflows, having 4238 // reached the limit of the stack size, while making very 4239 // slow forward progress. It may be best to bail out and 4240 // let the foreground collector do its job. 4241 // Clear _restart_addr, so that foreground GC 4242 // works from scratch. This avoids the headache of 4243 // a "rescan" which would otherwise be needed because 4244 // of the dirty mod union table & card table. 4245 _restart_addr = NULL; 4246 return false; 4247 } 4248 // Adjust the task to restart from _restart_addr 4249 tsk.reset(_restart_addr); 4250 cms_space ->initialize_sequential_subtasks_for_marking(num_workers, 4251 _restart_addr); 4252 _restart_addr = NULL; 4253 // Get the workers going again 4254 conc_workers()->start_task(&tsk); 4255 while (tsk.yielded()) { 4256 tsk.coordinator_yield(); 4257 conc_workers()->continue_task(&tsk); 4258 } 4259 } 4260 assert(tsk.completed(), "Inconsistency"); 4261 assert(tsk.result() == true, "Inconsistency"); 4262 return true; 4263 } 4264 4265 bool CMSCollector::do_marking_st(bool asynch) { 4266 ResourceMark rm; 4267 HandleMark hm; 4268 4269 // Temporarily make refs discovery single threaded (non-MT) 4270 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false); 4271 MarkFromRootsClosure markFromRootsClosure(this, _span, &_markBitMap, 4272 &_markStack, CMSYield && asynch); 4273 // the last argument to iterate indicates whether the iteration 4274 // should be incremental with periodic yields. 4275 _markBitMap.iterate(&markFromRootsClosure); 4276 // If _restart_addr is non-NULL, a marking stack overflow 4277 // occurred; we need to do a fresh iteration from the 4278 // indicated restart address. 4279 while (_restart_addr != NULL) { 4280 if (_foregroundGCIsActive && asynch) { 4281 // We may be running into repeated stack overflows, having 4282 // reached the limit of the stack size, while making very 4283 // slow forward progress. It may be best to bail out and 4284 // let the foreground collector do its job. 4285 // Clear _restart_addr, so that foreground GC 4286 // works from scratch. This avoids the headache of 4287 // a "rescan" which would otherwise be needed because 4288 // of the dirty mod union table & card table. 4289 _restart_addr = NULL; 4290 return false; // indicating failure to complete marking 4291 } 4292 // Deal with stack overflow: 4293 // we restart marking from _restart_addr 4294 HeapWord* ra = _restart_addr; 4295 markFromRootsClosure.reset(ra); 4296 _restart_addr = NULL; 4297 _markBitMap.iterate(&markFromRootsClosure, ra, _span.end()); 4298 } 4299 return true; 4300 } 4301 4302 void CMSCollector::preclean() { 4303 check_correct_thread_executing(); 4304 assert(Thread::current()->is_ConcurrentGC_thread(), "Wrong thread"); 4305 verify_work_stacks_empty(); 4306 verify_overflow_empty(); 4307 _abort_preclean = false; 4308 if (CMSPrecleaningEnabled) { 4309 _eden_chunk_index = 0; 4310 size_t used = get_eden_used(); 4311 size_t capacity = get_eden_capacity(); 4312 // Don't start sampling unless we will get sufficiently 4313 // many samples. 4314 if (used < (capacity/(CMSScheduleRemarkSamplingRatio * 100) 4315 * CMSScheduleRemarkEdenPenetration)) { 4316 _start_sampling = true; 4317 } else { 4318 _start_sampling = false; 4319 } 4320 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 4321 CMSPhaseAccounting pa(this, "preclean", !PrintGCDetails); 4322 preclean_work(CMSPrecleanRefLists1, CMSPrecleanSurvivors1); 4323 } 4324 CMSTokenSync x(true); // is cms thread 4325 if (CMSPrecleaningEnabled) { 4326 sample_eden(); 4327 _collectorState = AbortablePreclean; 4328 } else { 4329 _collectorState = FinalMarking; 4330 } 4331 verify_work_stacks_empty(); 4332 verify_overflow_empty(); 4333 } 4334 4335 // Try and schedule the remark such that young gen 4336 // occupancy is CMSScheduleRemarkEdenPenetration %. 4337 void CMSCollector::abortable_preclean() { 4338 check_correct_thread_executing(); 4339 assert(CMSPrecleaningEnabled, "Inconsistent control state"); 4340 assert(_collectorState == AbortablePreclean, "Inconsistent control state"); 4341 4342 // If Eden's current occupancy is below this threshold, 4343 // immediately schedule the remark; else preclean 4344 // past the next scavenge in an effort to 4345 // schedule the pause as described avove. By choosing 4346 // CMSScheduleRemarkEdenSizeThreshold >= max eden size 4347 // we will never do an actual abortable preclean cycle. 4348 if (get_eden_used() > CMSScheduleRemarkEdenSizeThreshold) { 4349 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 4350 CMSPhaseAccounting pa(this, "abortable-preclean", !PrintGCDetails); 4351 // We need more smarts in the abortable preclean 4352 // loop below to deal with cases where allocation 4353 // in young gen is very very slow, and our precleaning 4354 // is running a losing race against a horde of 4355 // mutators intent on flooding us with CMS updates 4356 // (dirty cards). 4357 // One, admittedly dumb, strategy is to give up 4358 // after a certain number of abortable precleaning loops 4359 // or after a certain maximum time. We want to make 4360 // this smarter in the next iteration. 4361 // XXX FIX ME!!! YSR 4362 size_t loops = 0, workdone = 0, cumworkdone = 0, waited = 0; 4363 while (!(should_abort_preclean() || 4364 ConcurrentMarkSweepThread::should_terminate())) { 4365 workdone = preclean_work(CMSPrecleanRefLists2, CMSPrecleanSurvivors2); 4366 cumworkdone += workdone; 4367 loops++; 4368 // Voluntarily terminate abortable preclean phase if we have 4369 // been at it for too long. 4370 if ((CMSMaxAbortablePrecleanLoops != 0) && 4371 loops >= CMSMaxAbortablePrecleanLoops) { 4372 if (PrintGCDetails) { 4373 gclog_or_tty->print(" CMS: abort preclean due to loops "); 4374 } 4375 break; 4376 } 4377 if (pa.wallclock_millis() > CMSMaxAbortablePrecleanTime) { 4378 if (PrintGCDetails) { 4379 gclog_or_tty->print(" CMS: abort preclean due to time "); 4380 } 4381 break; 4382 } 4383 // If we are doing little work each iteration, we should 4384 // take a short break. 4385 if (workdone < CMSAbortablePrecleanMinWorkPerIteration) { 4386 // Sleep for some time, waiting for work to accumulate 4387 stopTimer(); 4388 cmsThread()->wait_on_cms_lock(CMSAbortablePrecleanWaitMillis); 4389 startTimer(); 4390 waited++; 4391 } 4392 } 4393 if (PrintCMSStatistics > 0) { 4394 gclog_or_tty->print(" [%d iterations, %d waits, %d cards)] ", 4395 loops, waited, cumworkdone); 4396 } 4397 } 4398 CMSTokenSync x(true); // is cms thread 4399 if (_collectorState != Idling) { 4400 assert(_collectorState == AbortablePreclean, 4401 "Spontaneous state transition?"); 4402 _collectorState = FinalMarking; 4403 } // Else, a foreground collection completed this CMS cycle. 4404 return; 4405 } 4406 4407 // Respond to an Eden sampling opportunity 4408 void CMSCollector::sample_eden() { 4409 // Make sure a young gc cannot sneak in between our 4410 // reading and recording of a sample. 4411 assert(Thread::current()->is_ConcurrentGC_thread(), 4412 "Only the cms thread may collect Eden samples"); 4413 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 4414 "Should collect samples while holding CMS token"); 4415 if (!_start_sampling) { 4416 return; 4417 } 4418 if (_eden_chunk_array) { 4419 if (_eden_chunk_index < _eden_chunk_capacity) { 4420 _eden_chunk_array[_eden_chunk_index] = *_top_addr; // take sample 4421 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr, 4422 "Unexpected state of Eden"); 4423 // We'd like to check that what we just sampled is an oop-start address; 4424 // however, we cannot do that here since the object may not yet have been 4425 // initialized. So we'll instead do the check when we _use_ this sample 4426 // later. 4427 if (_eden_chunk_index == 0 || 4428 (pointer_delta(_eden_chunk_array[_eden_chunk_index], 4429 _eden_chunk_array[_eden_chunk_index-1]) 4430 >= CMSSamplingGrain)) { 4431 _eden_chunk_index++; // commit sample 4432 } 4433 } 4434 } 4435 if ((_collectorState == AbortablePreclean) && !_abort_preclean) { 4436 size_t used = get_eden_used(); 4437 size_t capacity = get_eden_capacity(); 4438 assert(used <= capacity, "Unexpected state of Eden"); 4439 if (used > (capacity/100 * CMSScheduleRemarkEdenPenetration)) { 4440 _abort_preclean = true; 4441 } 4442 } 4443 } 4444 4445 4446 size_t CMSCollector::preclean_work(bool clean_refs, bool clean_survivor) { 4447 assert(_collectorState == Precleaning || 4448 _collectorState == AbortablePreclean, "incorrect state"); 4449 ResourceMark rm; 4450 HandleMark hm; 4451 4452 // Precleaning is currently not MT but the reference processor 4453 // may be set for MT. Disable it temporarily here. 4454 ReferenceProcessor* rp = ref_processor(); 4455 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(rp, false); 4456 4457 // Do one pass of scrubbing the discovered reference lists 4458 // to remove any reference objects with strongly-reachable 4459 // referents. 4460 if (clean_refs) { 4461 CMSPrecleanRefsYieldClosure yield_cl(this); 4462 assert(rp->span().equals(_span), "Spans should be equal"); 4463 CMSKeepAliveClosure keep_alive(this, _span, &_markBitMap, 4464 &_markStack, true /* preclean */); 4465 CMSDrainMarkingStackClosure complete_trace(this, 4466 _span, &_markBitMap, &_markStack, 4467 &keep_alive, true /* preclean */); 4468 4469 // We don't want this step to interfere with a young 4470 // collection because we don't want to take CPU 4471 // or memory bandwidth away from the young GC threads 4472 // (which may be as many as there are CPUs). 4473 // Note that we don't need to protect ourselves from 4474 // interference with mutators because they can't 4475 // manipulate the discovered reference lists nor affect 4476 // the computed reachability of the referents, the 4477 // only properties manipulated by the precleaning 4478 // of these reference lists. 4479 stopTimer(); 4480 CMSTokenSyncWithLocks x(true /* is cms thread */, 4481 bitMapLock()); 4482 startTimer(); 4483 sample_eden(); 4484 4485 // The following will yield to allow foreground 4486 // collection to proceed promptly. XXX YSR: 4487 // The code in this method may need further 4488 // tweaking for better performance and some restructuring 4489 // for cleaner interfaces. 4490 rp->preclean_discovered_references( 4491 rp->is_alive_non_header(), &keep_alive, &complete_trace, &yield_cl); 4492 } 4493 4494 if (clean_survivor) { // preclean the active survivor space(s) 4495 assert(_young_gen->kind() == Generation::DefNew || 4496 _young_gen->kind() == Generation::ParNew || 4497 _young_gen->kind() == Generation::ASParNew, 4498 "incorrect type for cast"); 4499 DefNewGeneration* dng = (DefNewGeneration*)_young_gen; 4500 PushAndMarkClosure pam_cl(this, _span, ref_processor(), 4501 &_markBitMap, &_modUnionTable, 4502 &_markStack, true /* precleaning phase */); 4503 stopTimer(); 4504 CMSTokenSyncWithLocks ts(true /* is cms thread */, 4505 bitMapLock()); 4506 startTimer(); 4507 unsigned int before_count = 4508 GenCollectedHeap::heap()->total_collections(); 4509 SurvivorSpacePrecleanClosure 4510 sss_cl(this, _span, &_markBitMap, &_markStack, 4511 &pam_cl, before_count, CMSYield); 4512 dng->from()->object_iterate_careful(&sss_cl); 4513 dng->to()->object_iterate_careful(&sss_cl); 4514 } 4515 MarkRefsIntoAndScanClosure 4516 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable, 4517 &_markStack, this, CMSYield, 4518 true /* precleaning phase */); 4519 // CAUTION: The following closure has persistent state that may need to 4520 // be reset upon a decrease in the sequence of addresses it 4521 // processes. 4522 ScanMarkedObjectsAgainCarefullyClosure 4523 smoac_cl(this, _span, 4524 &_markBitMap, &_markStack, &mrias_cl, CMSYield); 4525 4526 // Preclean dirty cards in ModUnionTable and CardTable using 4527 // appropriate convergence criterion; 4528 // repeat CMSPrecleanIter times unless we find that 4529 // we are losing. 4530 assert(CMSPrecleanIter < 10, "CMSPrecleanIter is too large"); 4531 assert(CMSPrecleanNumerator < CMSPrecleanDenominator, 4532 "Bad convergence multiplier"); 4533 assert(CMSPrecleanThreshold >= 100, 4534 "Unreasonably low CMSPrecleanThreshold"); 4535 4536 size_t numIter, cumNumCards, lastNumCards, curNumCards; 4537 for (numIter = 0, cumNumCards = lastNumCards = curNumCards = 0; 4538 numIter < CMSPrecleanIter; 4539 numIter++, lastNumCards = curNumCards, cumNumCards += curNumCards) { 4540 curNumCards = preclean_mod_union_table(_cmsGen, &smoac_cl); 4541 if (Verbose && PrintGCDetails) { 4542 gclog_or_tty->print(" (modUnionTable: %d cards)", curNumCards); 4543 } 4544 // Either there are very few dirty cards, so re-mark 4545 // pause will be small anyway, or our pre-cleaning isn't 4546 // that much faster than the rate at which cards are being 4547 // dirtied, so we might as well stop and re-mark since 4548 // precleaning won't improve our re-mark time by much. 4549 if (curNumCards <= CMSPrecleanThreshold || 4550 (numIter > 0 && 4551 (curNumCards * CMSPrecleanDenominator > 4552 lastNumCards * CMSPrecleanNumerator))) { 4553 numIter++; 4554 cumNumCards += curNumCards; 4555 break; 4556 } 4557 } 4558 4559 preclean_klasses(&mrias_cl, _cmsGen->freelistLock()); 4560 4561 curNumCards = preclean_card_table(_cmsGen, &smoac_cl); 4562 cumNumCards += curNumCards; 4563 if (PrintGCDetails && PrintCMSStatistics != 0) { 4564 gclog_or_tty->print_cr(" (cardTable: %d cards, re-scanned %d cards, %d iterations)", 4565 curNumCards, cumNumCards, numIter); 4566 } 4567 return cumNumCards; // as a measure of useful work done 4568 } 4569 4570 // PRECLEANING NOTES: 4571 // Precleaning involves: 4572 // . reading the bits of the modUnionTable and clearing the set bits. 4573 // . For the cards corresponding to the set bits, we scan the 4574 // objects on those cards. This means we need the free_list_lock 4575 // so that we can safely iterate over the CMS space when scanning 4576 // for oops. 4577 // . When we scan the objects, we'll be both reading and setting 4578 // marks in the marking bit map, so we'll need the marking bit map. 4579 // . For protecting _collector_state transitions, we take the CGC_lock. 4580 // Note that any races in the reading of of card table entries by the 4581 // CMS thread on the one hand and the clearing of those entries by the 4582 // VM thread or the setting of those entries by the mutator threads on the 4583 // other are quite benign. However, for efficiency it makes sense to keep 4584 // the VM thread from racing with the CMS thread while the latter is 4585 // dirty card info to the modUnionTable. We therefore also use the 4586 // CGC_lock to protect the reading of the card table and the mod union 4587 // table by the CM thread. 4588 // . We run concurrently with mutator updates, so scanning 4589 // needs to be done carefully -- we should not try to scan 4590 // potentially uninitialized objects. 4591 // 4592 // Locking strategy: While holding the CGC_lock, we scan over and 4593 // reset a maximal dirty range of the mod union / card tables, then lock 4594 // the free_list_lock and bitmap lock to do a full marking, then 4595 // release these locks; and repeat the cycle. This allows for a 4596 // certain amount of fairness in the sharing of these locks between 4597 // the CMS collector on the one hand, and the VM thread and the 4598 // mutators on the other. 4599 4600 // NOTE: preclean_mod_union_table() and preclean_card_table() 4601 // further below are largely identical; if you need to modify 4602 // one of these methods, please check the other method too. 4603 4604 size_t CMSCollector::preclean_mod_union_table( 4605 ConcurrentMarkSweepGeneration* gen, 4606 ScanMarkedObjectsAgainCarefullyClosure* cl) { 4607 verify_work_stacks_empty(); 4608 verify_overflow_empty(); 4609 4610 // strategy: starting with the first card, accumulate contiguous 4611 // ranges of dirty cards; clear these cards, then scan the region 4612 // covered by these cards. 4613 4614 // Since all of the MUT is committed ahead, we can just use 4615 // that, in case the generations expand while we are precleaning. 4616 // It might also be fine to just use the committed part of the 4617 // generation, but we might potentially miss cards when the 4618 // generation is rapidly expanding while we are in the midst 4619 // of precleaning. 4620 HeapWord* startAddr = gen->reserved().start(); 4621 HeapWord* endAddr = gen->reserved().end(); 4622 4623 cl->setFreelistLock(gen->freelistLock()); // needed for yielding 4624 4625 size_t numDirtyCards, cumNumDirtyCards; 4626 HeapWord *nextAddr, *lastAddr; 4627 for (cumNumDirtyCards = numDirtyCards = 0, 4628 nextAddr = lastAddr = startAddr; 4629 nextAddr < endAddr; 4630 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 4631 4632 ResourceMark rm; 4633 HandleMark hm; 4634 4635 MemRegion dirtyRegion; 4636 { 4637 stopTimer(); 4638 // Potential yield point 4639 CMSTokenSync ts(true); 4640 startTimer(); 4641 sample_eden(); 4642 // Get dirty region starting at nextOffset (inclusive), 4643 // simultaneously clearing it. 4644 dirtyRegion = 4645 _modUnionTable.getAndClearMarkedRegion(nextAddr, endAddr); 4646 assert(dirtyRegion.start() >= nextAddr, 4647 "returned region inconsistent?"); 4648 } 4649 // Remember where the next search should begin. 4650 // The returned region (if non-empty) is a right open interval, 4651 // so lastOffset is obtained from the right end of that 4652 // interval. 4653 lastAddr = dirtyRegion.end(); 4654 // Should do something more transparent and less hacky XXX 4655 numDirtyCards = 4656 _modUnionTable.heapWordDiffToOffsetDiff(dirtyRegion.word_size()); 4657 4658 // We'll scan the cards in the dirty region (with periodic 4659 // yields for foreground GC as needed). 4660 if (!dirtyRegion.is_empty()) { 4661 assert(numDirtyCards > 0, "consistency check"); 4662 HeapWord* stop_point = NULL; 4663 stopTimer(); 4664 // Potential yield point 4665 CMSTokenSyncWithLocks ts(true, gen->freelistLock(), 4666 bitMapLock()); 4667 startTimer(); 4668 { 4669 verify_work_stacks_empty(); 4670 verify_overflow_empty(); 4671 sample_eden(); 4672 stop_point = 4673 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 4674 } 4675 if (stop_point != NULL) { 4676 // The careful iteration stopped early either because it found an 4677 // uninitialized object, or because we were in the midst of an 4678 // "abortable preclean", which should now be aborted. Redirty 4679 // the bits corresponding to the partially-scanned or unscanned 4680 // cards. We'll either restart at the next block boundary or 4681 // abort the preclean. 4682 assert((_collectorState == AbortablePreclean && should_abort_preclean()), 4683 "Should only be AbortablePreclean."); 4684 _modUnionTable.mark_range(MemRegion(stop_point, dirtyRegion.end())); 4685 if (should_abort_preclean()) { 4686 break; // out of preclean loop 4687 } else { 4688 // Compute the next address at which preclean should pick up; 4689 // might need bitMapLock in order to read P-bits. 4690 lastAddr = next_card_start_after_block(stop_point); 4691 } 4692 } 4693 } else { 4694 assert(lastAddr == endAddr, "consistency check"); 4695 assert(numDirtyCards == 0, "consistency check"); 4696 break; 4697 } 4698 } 4699 verify_work_stacks_empty(); 4700 verify_overflow_empty(); 4701 return cumNumDirtyCards; 4702 } 4703 4704 // NOTE: preclean_mod_union_table() above and preclean_card_table() 4705 // below are largely identical; if you need to modify 4706 // one of these methods, please check the other method too. 4707 4708 size_t CMSCollector::preclean_card_table(ConcurrentMarkSweepGeneration* gen, 4709 ScanMarkedObjectsAgainCarefullyClosure* cl) { 4710 // strategy: it's similar to precleamModUnionTable above, in that 4711 // we accumulate contiguous ranges of dirty cards, mark these cards 4712 // precleaned, then scan the region covered by these cards. 4713 HeapWord* endAddr = (HeapWord*)(gen->_virtual_space.high()); 4714 HeapWord* startAddr = (HeapWord*)(gen->_virtual_space.low()); 4715 4716 cl->setFreelistLock(gen->freelistLock()); // needed for yielding 4717 4718 size_t numDirtyCards, cumNumDirtyCards; 4719 HeapWord *lastAddr, *nextAddr; 4720 4721 for (cumNumDirtyCards = numDirtyCards = 0, 4722 nextAddr = lastAddr = startAddr; 4723 nextAddr < endAddr; 4724 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 4725 4726 ResourceMark rm; 4727 HandleMark hm; 4728 4729 MemRegion dirtyRegion; 4730 { 4731 // See comments in "Precleaning notes" above on why we 4732 // do this locking. XXX Could the locking overheads be 4733 // too high when dirty cards are sparse? [I don't think so.] 4734 stopTimer(); 4735 CMSTokenSync x(true); // is cms thread 4736 startTimer(); 4737 sample_eden(); 4738 // Get and clear dirty region from card table 4739 dirtyRegion = _ct->ct_bs()->dirty_card_range_after_reset( 4740 MemRegion(nextAddr, endAddr), 4741 true, 4742 CardTableModRefBS::precleaned_card_val()); 4743 4744 assert(dirtyRegion.start() >= nextAddr, 4745 "returned region inconsistent?"); 4746 } 4747 lastAddr = dirtyRegion.end(); 4748 numDirtyCards = 4749 dirtyRegion.word_size()/CardTableModRefBS::card_size_in_words; 4750 4751 if (!dirtyRegion.is_empty()) { 4752 stopTimer(); 4753 CMSTokenSyncWithLocks ts(true, gen->freelistLock(), bitMapLock()); 4754 startTimer(); 4755 sample_eden(); 4756 verify_work_stacks_empty(); 4757 verify_overflow_empty(); 4758 HeapWord* stop_point = 4759 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 4760 if (stop_point != NULL) { 4761 assert((_collectorState == AbortablePreclean && should_abort_preclean()), 4762 "Should only be AbortablePreclean."); 4763 _ct->ct_bs()->invalidate(MemRegion(stop_point, dirtyRegion.end())); 4764 if (should_abort_preclean()) { 4765 break; // out of preclean loop 4766 } else { 4767 // Compute the next address at which preclean should pick up. 4768 lastAddr = next_card_start_after_block(stop_point); 4769 } 4770 } 4771 } else { 4772 break; 4773 } 4774 } 4775 verify_work_stacks_empty(); 4776 verify_overflow_empty(); 4777 return cumNumDirtyCards; 4778 } 4779 4780 class PrecleanKlassClosure : public KlassClosure { 4781 CMKlassClosure _cm_klass_closure; 4782 public: 4783 PrecleanKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {} 4784 void do_klass(Klass* k) { 4785 if (k->has_accumulated_modified_oops()) { 4786 k->clear_accumulated_modified_oops(); 4787 4788 _cm_klass_closure.do_klass(k); 4789 } 4790 } 4791 }; 4792 4793 // The freelist lock is needed to prevent asserts, is it really needed? 4794 void CMSCollector::preclean_klasses(MarkRefsIntoAndScanClosure* cl, Mutex* freelistLock) { 4795 4796 cl->set_freelistLock(freelistLock); 4797 4798 CMSTokenSyncWithLocks ts(true, freelistLock, bitMapLock()); 4799 4800 // SSS: Add equivalent to ScanMarkedObjectsAgainCarefullyClosure::do_yield_check and should_abort_preclean? 4801 // SSS: We should probably check if precleaning should be aborted, at suitable intervals? 4802 PrecleanKlassClosure preclean_klass_closure(cl); 4803 ClassLoaderDataGraph::classes_do(&preclean_klass_closure); 4804 4805 verify_work_stacks_empty(); 4806 verify_overflow_empty(); 4807 } 4808 4809 void CMSCollector::checkpointRootsFinal(bool asynch, 4810 bool clear_all_soft_refs, bool init_mark_was_synchronous) { 4811 assert(_collectorState == FinalMarking, "incorrect state transition?"); 4812 check_correct_thread_executing(); 4813 // world is stopped at this checkpoint 4814 assert(SafepointSynchronize::is_at_safepoint(), 4815 "world should be stopped"); 4816 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 4817 4818 verify_work_stacks_empty(); 4819 verify_overflow_empty(); 4820 4821 SpecializationStats::clear(); 4822 if (PrintGCDetails) { 4823 gclog_or_tty->print("[YG occupancy: "SIZE_FORMAT" K ("SIZE_FORMAT" K)]", 4824 _young_gen->used() / K, 4825 _young_gen->capacity() / K); 4826 } 4827 if (asynch) { 4828 if (CMSScavengeBeforeRemark) { 4829 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4830 // Temporarily set flag to false, GCH->do_collection will 4831 // expect it to be false and set to true 4832 FlagSetting fl(gch->_is_gc_active, false); 4833 NOT_PRODUCT(TraceTime t("Scavenge-Before-Remark", 4834 PrintGCDetails && Verbose, true, gclog_or_tty);) 4835 int level = _cmsGen->level() - 1; 4836 if (level >= 0) { 4837 gch->do_collection(true, // full (i.e. force, see below) 4838 false, // !clear_all_soft_refs 4839 0, // size 4840 false, // is_tlab 4841 level // max_level 4842 ); 4843 } 4844 } 4845 FreelistLocker x(this); 4846 MutexLockerEx y(bitMapLock(), 4847 Mutex::_no_safepoint_check_flag); 4848 assert(!init_mark_was_synchronous, "but that's impossible!"); 4849 checkpointRootsFinalWork(asynch, clear_all_soft_refs, false); 4850 } else { 4851 // already have all the locks 4852 checkpointRootsFinalWork(asynch, clear_all_soft_refs, 4853 init_mark_was_synchronous); 4854 } 4855 verify_work_stacks_empty(); 4856 verify_overflow_empty(); 4857 SpecializationStats::print(); 4858 } 4859 4860 void CMSCollector::checkpointRootsFinalWork(bool asynch, 4861 bool clear_all_soft_refs, bool init_mark_was_synchronous) { 4862 4863 NOT_PRODUCT(TraceTime tr("checkpointRootsFinalWork", PrintGCDetails, false, gclog_or_tty);) 4864 4865 assert(haveFreelistLocks(), "must have free list locks"); 4866 assert_lock_strong(bitMapLock()); 4867 4868 if (UseAdaptiveSizePolicy) { 4869 size_policy()->checkpoint_roots_final_begin(); 4870 } 4871 4872 ResourceMark rm; 4873 HandleMark hm; 4874 4875 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4876 4877 if (should_unload_classes()) { 4878 CodeCache::gc_prologue(); 4879 } 4880 assert(haveFreelistLocks(), "must have free list locks"); 4881 assert_lock_strong(bitMapLock()); 4882 4883 if (!init_mark_was_synchronous) { 4884 // We might assume that we need not fill TLAB's when 4885 // CMSScavengeBeforeRemark is set, because we may have just done 4886 // a scavenge which would have filled all TLAB's -- and besides 4887 // Eden would be empty. This however may not always be the case -- 4888 // for instance although we asked for a scavenge, it may not have 4889 // happened because of a JNI critical section. We probably need 4890 // a policy for deciding whether we can in that case wait until 4891 // the critical section releases and then do the remark following 4892 // the scavenge, and skip it here. In the absence of that policy, 4893 // or of an indication of whether the scavenge did indeed occur, 4894 // we cannot rely on TLAB's having been filled and must do 4895 // so here just in case a scavenge did not happen. 4896 gch->ensure_parsability(false); // fill TLAB's, but no need to retire them 4897 // Update the saved marks which may affect the root scans. 4898 gch->save_marks(); 4899 4900 { 4901 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;) 4902 4903 // Note on the role of the mod union table: 4904 // Since the marker in "markFromRoots" marks concurrently with 4905 // mutators, it is possible for some reachable objects not to have been 4906 // scanned. For instance, an only reference to an object A was 4907 // placed in object B after the marker scanned B. Unless B is rescanned, 4908 // A would be collected. Such updates to references in marked objects 4909 // are detected via the mod union table which is the set of all cards 4910 // dirtied since the first checkpoint in this GC cycle and prior to 4911 // the most recent young generation GC, minus those cleaned up by the 4912 // concurrent precleaning. 4913 if (CMSParallelRemarkEnabled && CollectedHeap::use_parallel_gc_threads()) { 4914 TraceTime t("Rescan (parallel) ", PrintGCDetails, false, gclog_or_tty); 4915 do_remark_parallel(); 4916 } else { 4917 TraceTime t("Rescan (non-parallel) ", PrintGCDetails, false, 4918 gclog_or_tty); 4919 do_remark_non_parallel(); 4920 } 4921 } 4922 } else { 4923 assert(!asynch, "Can't have init_mark_was_synchronous in asynch mode"); 4924 // The initial mark was stop-world, so there's no rescanning to 4925 // do; go straight on to the next step below. 4926 } 4927 verify_work_stacks_empty(); 4928 verify_overflow_empty(); 4929 4930 { 4931 NOT_PRODUCT(TraceTime ts("refProcessingWork", PrintGCDetails, false, gclog_or_tty);) 4932 refProcessingWork(asynch, clear_all_soft_refs); 4933 } 4934 verify_work_stacks_empty(); 4935 verify_overflow_empty(); 4936 4937 if (should_unload_classes()) { 4938 CodeCache::gc_epilogue(); 4939 } 4940 JvmtiExport::gc_epilogue(); 4941 4942 // If we encountered any (marking stack / work queue) overflow 4943 // events during the current CMS cycle, take appropriate 4944 // remedial measures, where possible, so as to try and avoid 4945 // recurrence of that condition. 4946 assert(_markStack.isEmpty(), "No grey objects"); 4947 size_t ser_ovflw = _ser_pmc_remark_ovflw + _ser_pmc_preclean_ovflw + 4948 _ser_kac_ovflw + _ser_kac_preclean_ovflw; 4949 if (ser_ovflw > 0) { 4950 if (PrintCMSStatistics != 0) { 4951 gclog_or_tty->print_cr("Marking stack overflow (benign) " 4952 "(pmc_pc="SIZE_FORMAT", pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT 4953 ", kac_preclean="SIZE_FORMAT")", 4954 _ser_pmc_preclean_ovflw, _ser_pmc_remark_ovflw, 4955 _ser_kac_ovflw, _ser_kac_preclean_ovflw); 4956 } 4957 _markStack.expand(); 4958 _ser_pmc_remark_ovflw = 0; 4959 _ser_pmc_preclean_ovflw = 0; 4960 _ser_kac_preclean_ovflw = 0; 4961 _ser_kac_ovflw = 0; 4962 } 4963 if (_par_pmc_remark_ovflw > 0 || _par_kac_ovflw > 0) { 4964 if (PrintCMSStatistics != 0) { 4965 gclog_or_tty->print_cr("Work queue overflow (benign) " 4966 "(pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT")", 4967 _par_pmc_remark_ovflw, _par_kac_ovflw); 4968 } 4969 _par_pmc_remark_ovflw = 0; 4970 _par_kac_ovflw = 0; 4971 } 4972 if (PrintCMSStatistics != 0) { 4973 if (_markStack._hit_limit > 0) { 4974 gclog_or_tty->print_cr(" (benign) Hit max stack size limit ("SIZE_FORMAT")", 4975 _markStack._hit_limit); 4976 } 4977 if (_markStack._failed_double > 0) { 4978 gclog_or_tty->print_cr(" (benign) Failed stack doubling ("SIZE_FORMAT")," 4979 " current capacity "SIZE_FORMAT, 4980 _markStack._failed_double, 4981 _markStack.capacity()); 4982 } 4983 } 4984 _markStack._hit_limit = 0; 4985 _markStack._failed_double = 0; 4986 4987 if ((VerifyAfterGC || VerifyDuringGC) && 4988 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 4989 verify_after_remark(); 4990 } 4991 4992 // Change under the freelistLocks. 4993 _collectorState = Sweeping; 4994 // Call isAllClear() under bitMapLock 4995 assert(_modUnionTable.isAllClear(), 4996 "Should be clear by end of the final marking"); 4997 assert(_ct->klass_rem_set()->mod_union_is_clear(), 4998 "Should be clear by end of the final marking"); 4999 if (UseAdaptiveSizePolicy) { 5000 size_policy()->checkpoint_roots_final_end(gch->gc_cause()); 5001 } 5002 } 5003 5004 // Parallel remark task 5005 class CMSParRemarkTask: public AbstractGangTask { 5006 CMSCollector* _collector; 5007 int _n_workers; 5008 CompactibleFreeListSpace* _cms_space; 5009 5010 // The per-thread work queues, available here for stealing. 5011 OopTaskQueueSet* _task_queues; 5012 ParallelTaskTerminator _term; 5013 5014 public: 5015 // A value of 0 passed to n_workers will cause the number of 5016 // workers to be taken from the active workers in the work gang. 5017 CMSParRemarkTask(CMSCollector* collector, 5018 CompactibleFreeListSpace* cms_space, 5019 int n_workers, FlexibleWorkGang* workers, 5020 OopTaskQueueSet* task_queues): 5021 AbstractGangTask("Rescan roots and grey objects in parallel"), 5022 _collector(collector), 5023 _cms_space(cms_space), 5024 _n_workers(n_workers), 5025 _task_queues(task_queues), 5026 _term(n_workers, task_queues) { } 5027 5028 OopTaskQueueSet* task_queues() { return _task_queues; } 5029 5030 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 5031 5032 ParallelTaskTerminator* terminator() { return &_term; } 5033 int n_workers() { return _n_workers; } 5034 5035 void work(uint worker_id); 5036 5037 private: 5038 // Work method in support of parallel rescan ... of young gen spaces 5039 void do_young_space_rescan(int i, Par_MarkRefsIntoAndScanClosure* cl, 5040 ContiguousSpace* space, 5041 HeapWord** chunk_array, size_t chunk_top); 5042 5043 // ... of dirty cards in old space 5044 void do_dirty_card_rescan_tasks(CompactibleFreeListSpace* sp, int i, 5045 Par_MarkRefsIntoAndScanClosure* cl); 5046 5047 // ... work stealing for the above 5048 void do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, int* seed); 5049 }; 5050 5051 class RemarkKlassClosure : public KlassClosure { 5052 CMKlassClosure _cm_klass_closure; 5053 public: 5054 RemarkKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {} 5055 void do_klass(Klass* k) { 5056 // Check if we have modified any oops in the Klass during the concurrent marking. 5057 if (k->has_accumulated_modified_oops()) { 5058 k->clear_accumulated_modified_oops(); 5059 5060 // We could have transfered the current modified marks to the accumulated marks, 5061 // like we do with the Card Table to Mod Union Table. But it's not really necessary. 5062 } else if (k->has_modified_oops()) { 5063 // Don't clear anything, this info is needed by the next young collection. 5064 } else { 5065 // No modified oops in the Klass. 5066 return; 5067 } 5068 5069 // The klass has modified fields, need to scan the klass. 5070 _cm_klass_closure.do_klass(k); 5071 } 5072 }; 5073 5074 // work_queue(i) is passed to the closure 5075 // Par_MarkRefsIntoAndScanClosure. The "i" parameter 5076 // also is passed to do_dirty_card_rescan_tasks() and to 5077 // do_work_steal() to select the i-th task_queue. 5078 5079 void CMSParRemarkTask::work(uint worker_id) { 5080 elapsedTimer _timer; 5081 ResourceMark rm; 5082 HandleMark hm; 5083 5084 // ---------- rescan from roots -------------- 5085 _timer.start(); 5086 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5087 Par_MarkRefsIntoAndScanClosure par_mrias_cl(_collector, 5088 _collector->_span, _collector->ref_processor(), 5089 &(_collector->_markBitMap), 5090 work_queue(worker_id)); 5091 5092 // Rescan young gen roots first since these are likely 5093 // coarsely partitioned and may, on that account, constitute 5094 // the critical path; thus, it's best to start off that 5095 // work first. 5096 // ---------- young gen roots -------------- 5097 { 5098 DefNewGeneration* dng = _collector->_young_gen->as_DefNewGeneration(); 5099 EdenSpace* eden_space = dng->eden(); 5100 ContiguousSpace* from_space = dng->from(); 5101 ContiguousSpace* to_space = dng->to(); 5102 5103 HeapWord** eca = _collector->_eden_chunk_array; 5104 size_t ect = _collector->_eden_chunk_index; 5105 HeapWord** sca = _collector->_survivor_chunk_array; 5106 size_t sct = _collector->_survivor_chunk_index; 5107 5108 assert(ect <= _collector->_eden_chunk_capacity, "out of bounds"); 5109 assert(sct <= _collector->_survivor_chunk_capacity, "out of bounds"); 5110 5111 do_young_space_rescan(worker_id, &par_mrias_cl, to_space, NULL, 0); 5112 do_young_space_rescan(worker_id, &par_mrias_cl, from_space, sca, sct); 5113 do_young_space_rescan(worker_id, &par_mrias_cl, eden_space, eca, ect); 5114 5115 _timer.stop(); 5116 if (PrintCMSStatistics != 0) { 5117 gclog_or_tty->print_cr( 5118 "Finished young gen rescan work in %dth thread: %3.3f sec", 5119 worker_id, _timer.seconds()); 5120 } 5121 } 5122 5123 // ---------- remaining roots -------------- 5124 _timer.reset(); 5125 _timer.start(); 5126 gch->gen_process_strong_roots(_collector->_cmsGen->level(), 5127 false, // yg was scanned above 5128 false, // this is parallel code 5129 false, // not scavenging 5130 SharedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()), 5131 &par_mrias_cl, 5132 true, // walk all of code cache if (so & SO_CodeCache) 5133 NULL, 5134 NULL); // The dirty klasses will be handled below 5135 assert(_collector->should_unload_classes() 5136 || (_collector->CMSCollector::roots_scanning_options() & SharedHeap::SO_CodeCache), 5137 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 5138 _timer.stop(); 5139 if (PrintCMSStatistics != 0) { 5140 gclog_or_tty->print_cr( 5141 "Finished remaining root rescan work in %dth thread: %3.3f sec", 5142 worker_id, _timer.seconds()); 5143 } 5144 5145 // ---------- unhandled CLD scanning ---------- 5146 if (worker_id == 0) { // Single threaded at the moment. 5147 _timer.reset(); 5148 _timer.start(); 5149 5150 // Scan all new class loader data objects and new dependencies that were 5151 // introduced during concurrent marking. 5152 ResourceMark rm; 5153 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds(); 5154 for (int i = 0; i < array->length(); i++) { 5155 par_mrias_cl.do_class_loader_data(array->at(i)); 5156 } 5157 5158 // We don't need to keep track of new CLDs anymore. 5159 ClassLoaderDataGraph::remember_new_clds(false); 5160 5161 _timer.stop(); 5162 if (PrintCMSStatistics != 0) { 5163 gclog_or_tty->print_cr( 5164 "Finished unhandled CLD scanning work in %dth thread: %3.3f sec", 5165 worker_id, _timer.seconds()); 5166 } 5167 } 5168 5169 // ---------- dirty klass scanning ---------- 5170 if (worker_id == 0) { // Single threaded at the moment. 5171 _timer.reset(); 5172 _timer.start(); 5173 5174 // Scan all classes that was dirtied during the concurrent marking phase. 5175 RemarkKlassClosure remark_klass_closure(&par_mrias_cl); 5176 ClassLoaderDataGraph::classes_do(&remark_klass_closure); 5177 5178 _timer.stop(); 5179 if (PrintCMSStatistics != 0) { 5180 gclog_or_tty->print_cr( 5181 "Finished dirty klass scanning work in %dth thread: %3.3f sec", 5182 worker_id, _timer.seconds()); 5183 } 5184 } 5185 5186 // We might have added oops to ClassLoaderData::_handles during the 5187 // concurrent marking phase. These oops point to newly allocated objects 5188 // that are guaranteed to be kept alive. Either by the direct allocation 5189 // code, or when the young collector processes the strong roots. Hence, 5190 // we don't have to revisit the _handles block during the remark phase. 5191 5192 // ---------- rescan dirty cards ------------ 5193 _timer.reset(); 5194 _timer.start(); 5195 5196 // Do the rescan tasks for each of the two spaces 5197 // (cms_space) in turn. 5198 // "worker_id" is passed to select the task_queue for "worker_id" 5199 do_dirty_card_rescan_tasks(_cms_space, worker_id, &par_mrias_cl); 5200 _timer.stop(); 5201 if (PrintCMSStatistics != 0) { 5202 gclog_or_tty->print_cr( 5203 "Finished dirty card rescan work in %dth thread: %3.3f sec", 5204 worker_id, _timer.seconds()); 5205 } 5206 5207 // ---------- steal work from other threads ... 5208 // ---------- ... and drain overflow list. 5209 _timer.reset(); 5210 _timer.start(); 5211 do_work_steal(worker_id, &par_mrias_cl, _collector->hash_seed(worker_id)); 5212 _timer.stop(); 5213 if (PrintCMSStatistics != 0) { 5214 gclog_or_tty->print_cr( 5215 "Finished work stealing in %dth thread: %3.3f sec", 5216 worker_id, _timer.seconds()); 5217 } 5218 } 5219 5220 // Note that parameter "i" is not used. 5221 void 5222 CMSParRemarkTask::do_young_space_rescan(int i, 5223 Par_MarkRefsIntoAndScanClosure* cl, ContiguousSpace* space, 5224 HeapWord** chunk_array, size_t chunk_top) { 5225 // Until all tasks completed: 5226 // . claim an unclaimed task 5227 // . compute region boundaries corresponding to task claimed 5228 // using chunk_array 5229 // . par_oop_iterate(cl) over that region 5230 5231 ResourceMark rm; 5232 HandleMark hm; 5233 5234 SequentialSubTasksDone* pst = space->par_seq_tasks(); 5235 assert(pst->valid(), "Uninitialized use?"); 5236 5237 uint nth_task = 0; 5238 uint n_tasks = pst->n_tasks(); 5239 5240 HeapWord *start, *end; 5241 while (!pst->is_task_claimed(/* reference */ nth_task)) { 5242 // We claimed task # nth_task; compute its boundaries. 5243 if (chunk_top == 0) { // no samples were taken 5244 assert(nth_task == 0 && n_tasks == 1, "Can have only 1 EdenSpace task"); 5245 start = space->bottom(); 5246 end = space->top(); 5247 } else if (nth_task == 0) { 5248 start = space->bottom(); 5249 end = chunk_array[nth_task]; 5250 } else if (nth_task < (uint)chunk_top) { 5251 assert(nth_task >= 1, "Control point invariant"); 5252 start = chunk_array[nth_task - 1]; 5253 end = chunk_array[nth_task]; 5254 } else { 5255 assert(nth_task == (uint)chunk_top, "Control point invariant"); 5256 start = chunk_array[chunk_top - 1]; 5257 end = space->top(); 5258 } 5259 MemRegion mr(start, end); 5260 // Verify that mr is in space 5261 assert(mr.is_empty() || space->used_region().contains(mr), 5262 "Should be in space"); 5263 // Verify that "start" is an object boundary 5264 assert(mr.is_empty() || oop(mr.start())->is_oop(), 5265 "Should be an oop"); 5266 space->par_oop_iterate(mr, cl); 5267 } 5268 pst->all_tasks_completed(); 5269 } 5270 5271 void 5272 CMSParRemarkTask::do_dirty_card_rescan_tasks( 5273 CompactibleFreeListSpace* sp, int i, 5274 Par_MarkRefsIntoAndScanClosure* cl) { 5275 // Until all tasks completed: 5276 // . claim an unclaimed task 5277 // . compute region boundaries corresponding to task claimed 5278 // . transfer dirty bits ct->mut for that region 5279 // . apply rescanclosure to dirty mut bits for that region 5280 5281 ResourceMark rm; 5282 HandleMark hm; 5283 5284 OopTaskQueue* work_q = work_queue(i); 5285 ModUnionClosure modUnionClosure(&(_collector->_modUnionTable)); 5286 // CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! 5287 // CAUTION: This closure has state that persists across calls to 5288 // the work method dirty_range_iterate_clear() in that it has 5289 // imbedded in it a (subtype of) UpwardsObjectClosure. The 5290 // use of that state in the imbedded UpwardsObjectClosure instance 5291 // assumes that the cards are always iterated (even if in parallel 5292 // by several threads) in monotonically increasing order per each 5293 // thread. This is true of the implementation below which picks 5294 // card ranges (chunks) in monotonically increasing order globally 5295 // and, a-fortiori, in monotonically increasing order per thread 5296 // (the latter order being a subsequence of the former). 5297 // If the work code below is ever reorganized into a more chaotic 5298 // work-partitioning form than the current "sequential tasks" 5299 // paradigm, the use of that persistent state will have to be 5300 // revisited and modified appropriately. See also related 5301 // bug 4756801 work on which should examine this code to make 5302 // sure that the changes there do not run counter to the 5303 // assumptions made here and necessary for correctness and 5304 // efficiency. Note also that this code might yield inefficient 5305 // behaviour in the case of very large objects that span one or 5306 // more work chunks. Such objects would potentially be scanned 5307 // several times redundantly. Work on 4756801 should try and 5308 // address that performance anomaly if at all possible. XXX 5309 MemRegion full_span = _collector->_span; 5310 CMSBitMap* bm = &(_collector->_markBitMap); // shared 5311 MarkFromDirtyCardsClosure 5312 greyRescanClosure(_collector, full_span, // entire span of interest 5313 sp, bm, work_q, cl); 5314 5315 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 5316 assert(pst->valid(), "Uninitialized use?"); 5317 uint nth_task = 0; 5318 const int alignment = CardTableModRefBS::card_size * BitsPerWord; 5319 MemRegion span = sp->used_region(); 5320 HeapWord* start_addr = span.start(); 5321 HeapWord* end_addr = (HeapWord*)round_to((intptr_t)span.end(), 5322 alignment); 5323 const size_t chunk_size = sp->rescan_task_size(); // in HeapWord units 5324 assert((HeapWord*)round_to((intptr_t)start_addr, alignment) == 5325 start_addr, "Check alignment"); 5326 assert((size_t)round_to((intptr_t)chunk_size, alignment) == 5327 chunk_size, "Check alignment"); 5328 5329 while (!pst->is_task_claimed(/* reference */ nth_task)) { 5330 // Having claimed the nth_task, compute corresponding mem-region, 5331 // which is a-fortiori aligned correctly (i.e. at a MUT bopundary). 5332 // The alignment restriction ensures that we do not need any 5333 // synchronization with other gang-workers while setting or 5334 // clearing bits in thus chunk of the MUT. 5335 MemRegion this_span = MemRegion(start_addr + nth_task*chunk_size, 5336 start_addr + (nth_task+1)*chunk_size); 5337 // The last chunk's end might be way beyond end of the 5338 // used region. In that case pull back appropriately. 5339 if (this_span.end() > end_addr) { 5340 this_span.set_end(end_addr); 5341 assert(!this_span.is_empty(), "Program logic (calculation of n_tasks)"); 5342 } 5343 // Iterate over the dirty cards covering this chunk, marking them 5344 // precleaned, and setting the corresponding bits in the mod union 5345 // table. Since we have been careful to partition at Card and MUT-word 5346 // boundaries no synchronization is needed between parallel threads. 5347 _collector->_ct->ct_bs()->dirty_card_iterate(this_span, 5348 &modUnionClosure); 5349 5350 // Having transferred these marks into the modUnionTable, 5351 // rescan the marked objects on the dirty cards in the modUnionTable. 5352 // Even if this is at a synchronous collection, the initial marking 5353 // may have been done during an asynchronous collection so there 5354 // may be dirty bits in the mod-union table. 5355 _collector->_modUnionTable.dirty_range_iterate_clear( 5356 this_span, &greyRescanClosure); 5357 _collector->_modUnionTable.verifyNoOneBitsInRange( 5358 this_span.start(), 5359 this_span.end()); 5360 } 5361 pst->all_tasks_completed(); // declare that i am done 5362 } 5363 5364 // . see if we can share work_queues with ParNew? XXX 5365 void 5366 CMSParRemarkTask::do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, 5367 int* seed) { 5368 OopTaskQueue* work_q = work_queue(i); 5369 NOT_PRODUCT(int num_steals = 0;) 5370 oop obj_to_scan; 5371 CMSBitMap* bm = &(_collector->_markBitMap); 5372 5373 while (true) { 5374 // Completely finish any left over work from (an) earlier round(s) 5375 cl->trim_queue(0); 5376 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 5377 (size_t)ParGCDesiredObjsFromOverflowList); 5378 // Now check if there's any work in the overflow list 5379 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads, 5380 // only affects the number of attempts made to get work from the 5381 // overflow list and does not affect the number of workers. Just 5382 // pass ParallelGCThreads so this behavior is unchanged. 5383 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 5384 work_q, 5385 ParallelGCThreads)) { 5386 // found something in global overflow list; 5387 // not yet ready to go stealing work from others. 5388 // We'd like to assert(work_q->size() != 0, ...) 5389 // because we just took work from the overflow list, 5390 // but of course we can't since all of that could have 5391 // been already stolen from us. 5392 // "He giveth and He taketh away." 5393 continue; 5394 } 5395 // Verify that we have no work before we resort to stealing 5396 assert(work_q->size() == 0, "Have work, shouldn't steal"); 5397 // Try to steal from other queues that have work 5398 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 5399 NOT_PRODUCT(num_steals++;) 5400 assert(obj_to_scan->is_oop(), "Oops, not an oop!"); 5401 assert(bm->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 5402 // Do scanning work 5403 obj_to_scan->oop_iterate(cl); 5404 // Loop around, finish this work, and try to steal some more 5405 } else if (terminator()->offer_termination()) { 5406 break; // nirvana from the infinite cycle 5407 } 5408 } 5409 NOT_PRODUCT( 5410 if (PrintCMSStatistics != 0) { 5411 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals); 5412 } 5413 ) 5414 assert(work_q->size() == 0 && _collector->overflow_list_is_empty(), 5415 "Else our work is not yet done"); 5416 } 5417 5418 // Return a thread-local PLAB recording array, as appropriate. 5419 void* CMSCollector::get_data_recorder(int thr_num) { 5420 if (_survivor_plab_array != NULL && 5421 (CMSPLABRecordAlways || 5422 (_collectorState > Marking && _collectorState < FinalMarking))) { 5423 assert(thr_num < (int)ParallelGCThreads, "thr_num is out of bounds"); 5424 ChunkArray* ca = &_survivor_plab_array[thr_num]; 5425 ca->reset(); // clear it so that fresh data is recorded 5426 return (void*) ca; 5427 } else { 5428 return NULL; 5429 } 5430 } 5431 5432 // Reset all the thread-local PLAB recording arrays 5433 void CMSCollector::reset_survivor_plab_arrays() { 5434 for (uint i = 0; i < ParallelGCThreads; i++) { 5435 _survivor_plab_array[i].reset(); 5436 } 5437 } 5438 5439 // Merge the per-thread plab arrays into the global survivor chunk 5440 // array which will provide the partitioning of the survivor space 5441 // for CMS rescan. 5442 void CMSCollector::merge_survivor_plab_arrays(ContiguousSpace* surv, 5443 int no_of_gc_threads) { 5444 assert(_survivor_plab_array != NULL, "Error"); 5445 assert(_survivor_chunk_array != NULL, "Error"); 5446 assert(_collectorState == FinalMarking, "Error"); 5447 for (int j = 0; j < no_of_gc_threads; j++) { 5448 _cursor[j] = 0; 5449 } 5450 HeapWord* top = surv->top(); 5451 size_t i; 5452 for (i = 0; i < _survivor_chunk_capacity; i++) { // all sca entries 5453 HeapWord* min_val = top; // Higher than any PLAB address 5454 uint min_tid = 0; // position of min_val this round 5455 for (int j = 0; j < no_of_gc_threads; j++) { 5456 ChunkArray* cur_sca = &_survivor_plab_array[j]; 5457 if (_cursor[j] == cur_sca->end()) { 5458 continue; 5459 } 5460 assert(_cursor[j] < cur_sca->end(), "ctl pt invariant"); 5461 HeapWord* cur_val = cur_sca->nth(_cursor[j]); 5462 assert(surv->used_region().contains(cur_val), "Out of bounds value"); 5463 if (cur_val < min_val) { 5464 min_tid = j; 5465 min_val = cur_val; 5466 } else { 5467 assert(cur_val < top, "All recorded addresses should be less"); 5468 } 5469 } 5470 // At this point min_val and min_tid are respectively 5471 // the least address in _survivor_plab_array[j]->nth(_cursor[j]) 5472 // and the thread (j) that witnesses that address. 5473 // We record this address in the _survivor_chunk_array[i] 5474 // and increment _cursor[min_tid] prior to the next round i. 5475 if (min_val == top) { 5476 break; 5477 } 5478 _survivor_chunk_array[i] = min_val; 5479 _cursor[min_tid]++; 5480 } 5481 // We are all done; record the size of the _survivor_chunk_array 5482 _survivor_chunk_index = i; // exclusive: [0, i) 5483 if (PrintCMSStatistics > 0) { 5484 gclog_or_tty->print(" (Survivor:" SIZE_FORMAT "chunks) ", i); 5485 } 5486 // Verify that we used up all the recorded entries 5487 #ifdef ASSERT 5488 size_t total = 0; 5489 for (int j = 0; j < no_of_gc_threads; j++) { 5490 assert(_cursor[j] == _survivor_plab_array[j].end(), "Ctl pt invariant"); 5491 total += _cursor[j]; 5492 } 5493 assert(total == _survivor_chunk_index, "Ctl Pt Invariant"); 5494 // Check that the merged array is in sorted order 5495 if (total > 0) { 5496 for (size_t i = 0; i < total - 1; i++) { 5497 if (PrintCMSStatistics > 0) { 5498 gclog_or_tty->print(" (chunk" SIZE_FORMAT ":" INTPTR_FORMAT ") ", 5499 i, _survivor_chunk_array[i]); 5500 } 5501 assert(_survivor_chunk_array[i] < _survivor_chunk_array[i+1], 5502 "Not sorted"); 5503 } 5504 } 5505 #endif // ASSERT 5506 } 5507 5508 // Set up the space's par_seq_tasks structure for work claiming 5509 // for parallel rescan of young gen. 5510 // See ParRescanTask where this is currently used. 5511 void 5512 CMSCollector:: 5513 initialize_sequential_subtasks_for_young_gen_rescan(int n_threads) { 5514 assert(n_threads > 0, "Unexpected n_threads argument"); 5515 DefNewGeneration* dng = (DefNewGeneration*)_young_gen; 5516 5517 // Eden space 5518 { 5519 SequentialSubTasksDone* pst = dng->eden()->par_seq_tasks(); 5520 assert(!pst->valid(), "Clobbering existing data?"); 5521 // Each valid entry in [0, _eden_chunk_index) represents a task. 5522 size_t n_tasks = _eden_chunk_index + 1; 5523 assert(n_tasks == 1 || _eden_chunk_array != NULL, "Error"); 5524 // Sets the condition for completion of the subtask (how many threads 5525 // need to finish in order to be done). 5526 pst->set_n_threads(n_threads); 5527 pst->set_n_tasks((int)n_tasks); 5528 } 5529 5530 // Merge the survivor plab arrays into _survivor_chunk_array 5531 if (_survivor_plab_array != NULL) { 5532 merge_survivor_plab_arrays(dng->from(), n_threads); 5533 } else { 5534 assert(_survivor_chunk_index == 0, "Error"); 5535 } 5536 5537 // To space 5538 { 5539 SequentialSubTasksDone* pst = dng->to()->par_seq_tasks(); 5540 assert(!pst->valid(), "Clobbering existing data?"); 5541 // Sets the condition for completion of the subtask (how many threads 5542 // need to finish in order to be done). 5543 pst->set_n_threads(n_threads); 5544 pst->set_n_tasks(1); 5545 assert(pst->valid(), "Error"); 5546 } 5547 5548 // From space 5549 { 5550 SequentialSubTasksDone* pst = dng->from()->par_seq_tasks(); 5551 assert(!pst->valid(), "Clobbering existing data?"); 5552 size_t n_tasks = _survivor_chunk_index + 1; 5553 assert(n_tasks == 1 || _survivor_chunk_array != NULL, "Error"); 5554 // Sets the condition for completion of the subtask (how many threads 5555 // need to finish in order to be done). 5556 pst->set_n_threads(n_threads); 5557 pst->set_n_tasks((int)n_tasks); 5558 assert(pst->valid(), "Error"); 5559 } 5560 } 5561 5562 // Parallel version of remark 5563 void CMSCollector::do_remark_parallel() { 5564 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5565 FlexibleWorkGang* workers = gch->workers(); 5566 assert(workers != NULL, "Need parallel worker threads."); 5567 // Choose to use the number of GC workers most recently set 5568 // into "active_workers". If active_workers is not set, set it 5569 // to ParallelGCThreads. 5570 int n_workers = workers->active_workers(); 5571 if (n_workers == 0) { 5572 assert(n_workers > 0, "Should have been set during scavenge"); 5573 n_workers = ParallelGCThreads; 5574 workers->set_active_workers(n_workers); 5575 } 5576 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 5577 5578 CMSParRemarkTask tsk(this, 5579 cms_space, 5580 n_workers, workers, task_queues()); 5581 5582 // Set up for parallel process_strong_roots work. 5583 gch->set_par_threads(n_workers); 5584 // We won't be iterating over the cards in the card table updating 5585 // the younger_gen cards, so we shouldn't call the following else 5586 // the verification code as well as subsequent younger_refs_iterate 5587 // code would get confused. XXX 5588 // gch->rem_set()->prepare_for_younger_refs_iterate(true); // parallel 5589 5590 // The young gen rescan work will not be done as part of 5591 // process_strong_roots (which currently doesn't knw how to 5592 // parallelize such a scan), but rather will be broken up into 5593 // a set of parallel tasks (via the sampling that the [abortable] 5594 // preclean phase did of EdenSpace, plus the [two] tasks of 5595 // scanning the [two] survivor spaces. Further fine-grain 5596 // parallelization of the scanning of the survivor spaces 5597 // themselves, and of precleaning of the younger gen itself 5598 // is deferred to the future. 5599 initialize_sequential_subtasks_for_young_gen_rescan(n_workers); 5600 5601 // The dirty card rescan work is broken up into a "sequence" 5602 // of parallel tasks (per constituent space) that are dynamically 5603 // claimed by the parallel threads. 5604 cms_space->initialize_sequential_subtasks_for_rescan(n_workers); 5605 5606 // It turns out that even when we're using 1 thread, doing the work in a 5607 // separate thread causes wide variance in run times. We can't help this 5608 // in the multi-threaded case, but we special-case n=1 here to get 5609 // repeatable measurements of the 1-thread overhead of the parallel code. 5610 if (n_workers > 1) { 5611 // Make refs discovery MT-safe, if it isn't already: it may not 5612 // necessarily be so, since it's possible that we are doing 5613 // ST marking. 5614 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), true); 5615 GenCollectedHeap::StrongRootsScope srs(gch); 5616 workers->run_task(&tsk); 5617 } else { 5618 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false); 5619 GenCollectedHeap::StrongRootsScope srs(gch); 5620 tsk.work(0); 5621 } 5622 5623 gch->set_par_threads(0); // 0 ==> non-parallel. 5624 // restore, single-threaded for now, any preserved marks 5625 // as a result of work_q overflow 5626 restore_preserved_marks_if_any(); 5627 } 5628 5629 // Non-parallel version of remark 5630 void CMSCollector::do_remark_non_parallel() { 5631 ResourceMark rm; 5632 HandleMark hm; 5633 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5634 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false); 5635 5636 MarkRefsIntoAndScanClosure 5637 mrias_cl(_span, ref_processor(), &_markBitMap, NULL /* not precleaning */, 5638 &_markStack, this, 5639 false /* should_yield */, false /* not precleaning */); 5640 MarkFromDirtyCardsClosure 5641 markFromDirtyCardsClosure(this, _span, 5642 NULL, // space is set further below 5643 &_markBitMap, &_markStack, &mrias_cl); 5644 { 5645 TraceTime t("grey object rescan", PrintGCDetails, false, gclog_or_tty); 5646 // Iterate over the dirty cards, setting the corresponding bits in the 5647 // mod union table. 5648 { 5649 ModUnionClosure modUnionClosure(&_modUnionTable); 5650 _ct->ct_bs()->dirty_card_iterate( 5651 _cmsGen->used_region(), 5652 &modUnionClosure); 5653 } 5654 // Having transferred these marks into the modUnionTable, we just need 5655 // to rescan the marked objects on the dirty cards in the modUnionTable. 5656 // The initial marking may have been done during an asynchronous 5657 // collection so there may be dirty bits in the mod-union table. 5658 const int alignment = 5659 CardTableModRefBS::card_size * BitsPerWord; 5660 { 5661 // ... First handle dirty cards in CMS gen 5662 markFromDirtyCardsClosure.set_space(_cmsGen->cmsSpace()); 5663 MemRegion ur = _cmsGen->used_region(); 5664 HeapWord* lb = ur.start(); 5665 HeapWord* ub = (HeapWord*)round_to((intptr_t)ur.end(), alignment); 5666 MemRegion cms_span(lb, ub); 5667 _modUnionTable.dirty_range_iterate_clear(cms_span, 5668 &markFromDirtyCardsClosure); 5669 verify_work_stacks_empty(); 5670 if (PrintCMSStatistics != 0) { 5671 gclog_or_tty->print(" (re-scanned "SIZE_FORMAT" dirty cards in cms gen) ", 5672 markFromDirtyCardsClosure.num_dirty_cards()); 5673 } 5674 } 5675 } 5676 if (VerifyDuringGC && 5677 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 5678 HandleMark hm; // Discard invalid handles created during verification 5679 Universe::verify(true); 5680 } 5681 { 5682 TraceTime t("root rescan", PrintGCDetails, false, gclog_or_tty); 5683 5684 verify_work_stacks_empty(); 5685 5686 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 5687 GenCollectedHeap::StrongRootsScope srs(gch); 5688 gch->gen_process_strong_roots(_cmsGen->level(), 5689 true, // younger gens as roots 5690 false, // use the local StrongRootsScope 5691 false, // not scavenging 5692 SharedHeap::ScanningOption(roots_scanning_options()), 5693 &mrias_cl, 5694 true, // walk code active on stacks 5695 NULL, 5696 NULL); // The dirty klasses will be handled below 5697 5698 assert(should_unload_classes() 5699 || (roots_scanning_options() & SharedHeap::SO_CodeCache), 5700 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 5701 } 5702 5703 { 5704 TraceTime t("visit unhandled CLDs", PrintGCDetails, false, gclog_or_tty); 5705 5706 verify_work_stacks_empty(); 5707 5708 // Scan all class loader data objects that might have been introduced 5709 // during concurrent marking. 5710 ResourceMark rm; 5711 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds(); 5712 for (int i = 0; i < array->length(); i++) { 5713 mrias_cl.do_class_loader_data(array->at(i)); 5714 } 5715 5716 // We don't need to keep track of new CLDs anymore. 5717 ClassLoaderDataGraph::remember_new_clds(false); 5718 5719 verify_work_stacks_empty(); 5720 } 5721 5722 { 5723 TraceTime t("dirty klass scan", PrintGCDetails, false, gclog_or_tty); 5724 5725 verify_work_stacks_empty(); 5726 5727 RemarkKlassClosure remark_klass_closure(&mrias_cl); 5728 ClassLoaderDataGraph::classes_do(&remark_klass_closure); 5729 5730 verify_work_stacks_empty(); 5731 } 5732 5733 // We might have added oops to ClassLoaderData::_handles during the 5734 // concurrent marking phase. These oops point to newly allocated objects 5735 // that are guaranteed to be kept alive. Either by the direct allocation 5736 // code, or when the young collector processes the strong roots. Hence, 5737 // we don't have to revisit the _handles block during the remark phase. 5738 5739 verify_work_stacks_empty(); 5740 // Restore evacuated mark words, if any, used for overflow list links 5741 if (!CMSOverflowEarlyRestoration) { 5742 restore_preserved_marks_if_any(); 5743 } 5744 verify_overflow_empty(); 5745 } 5746 5747 //////////////////////////////////////////////////////// 5748 // Parallel Reference Processing Task Proxy Class 5749 //////////////////////////////////////////////////////// 5750 class CMSRefProcTaskProxy: public AbstractGangTaskWOopQueues { 5751 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask; 5752 CMSCollector* _collector; 5753 CMSBitMap* _mark_bit_map; 5754 const MemRegion _span; 5755 ProcessTask& _task; 5756 5757 public: 5758 CMSRefProcTaskProxy(ProcessTask& task, 5759 CMSCollector* collector, 5760 const MemRegion& span, 5761 CMSBitMap* mark_bit_map, 5762 AbstractWorkGang* workers, 5763 OopTaskQueueSet* task_queues): 5764 // XXX Should superclass AGTWOQ also know about AWG since it knows 5765 // about the task_queues used by the AWG? Then it could initialize 5766 // the terminator() object. See 6984287. The set_for_termination() 5767 // below is a temporary band-aid for the regression in 6984287. 5768 AbstractGangTaskWOopQueues("Process referents by policy in parallel", 5769 task_queues), 5770 _task(task), 5771 _collector(collector), _span(span), _mark_bit_map(mark_bit_map) 5772 { 5773 assert(_collector->_span.equals(_span) && !_span.is_empty(), 5774 "Inconsistency in _span"); 5775 set_for_termination(workers->active_workers()); 5776 } 5777 5778 OopTaskQueueSet* task_queues() { return queues(); } 5779 5780 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 5781 5782 void do_work_steal(int i, 5783 CMSParDrainMarkingStackClosure* drain, 5784 CMSParKeepAliveClosure* keep_alive, 5785 int* seed); 5786 5787 virtual void work(uint worker_id); 5788 }; 5789 5790 void CMSRefProcTaskProxy::work(uint worker_id) { 5791 assert(_collector->_span.equals(_span), "Inconsistency in _span"); 5792 CMSParKeepAliveClosure par_keep_alive(_collector, _span, 5793 _mark_bit_map, 5794 work_queue(worker_id)); 5795 CMSParDrainMarkingStackClosure par_drain_stack(_collector, _span, 5796 _mark_bit_map, 5797 work_queue(worker_id)); 5798 CMSIsAliveClosure is_alive_closure(_span, _mark_bit_map); 5799 _task.work(worker_id, is_alive_closure, par_keep_alive, par_drain_stack); 5800 if (_task.marks_oops_alive()) { 5801 do_work_steal(worker_id, &par_drain_stack, &par_keep_alive, 5802 _collector->hash_seed(worker_id)); 5803 } 5804 assert(work_queue(worker_id)->size() == 0, "work_queue should be empty"); 5805 assert(_collector->_overflow_list == NULL, "non-empty _overflow_list"); 5806 } 5807 5808 class CMSRefEnqueueTaskProxy: public AbstractGangTask { 5809 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask; 5810 EnqueueTask& _task; 5811 5812 public: 5813 CMSRefEnqueueTaskProxy(EnqueueTask& task) 5814 : AbstractGangTask("Enqueue reference objects in parallel"), 5815 _task(task) 5816 { } 5817 5818 virtual void work(uint worker_id) 5819 { 5820 _task.work(worker_id); 5821 } 5822 }; 5823 5824 CMSParKeepAliveClosure::CMSParKeepAliveClosure(CMSCollector* collector, 5825 MemRegion span, CMSBitMap* bit_map, OopTaskQueue* work_queue): 5826 _span(span), 5827 _bit_map(bit_map), 5828 _work_queue(work_queue), 5829 _mark_and_push(collector, span, bit_map, work_queue), 5830 _low_water_mark(MIN2((uint)(work_queue->max_elems()/4), 5831 (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads))) 5832 { } 5833 5834 // . see if we can share work_queues with ParNew? XXX 5835 void CMSRefProcTaskProxy::do_work_steal(int i, 5836 CMSParDrainMarkingStackClosure* drain, 5837 CMSParKeepAliveClosure* keep_alive, 5838 int* seed) { 5839 OopTaskQueue* work_q = work_queue(i); 5840 NOT_PRODUCT(int num_steals = 0;) 5841 oop obj_to_scan; 5842 5843 while (true) { 5844 // Completely finish any left over work from (an) earlier round(s) 5845 drain->trim_queue(0); 5846 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 5847 (size_t)ParGCDesiredObjsFromOverflowList); 5848 // Now check if there's any work in the overflow list 5849 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads, 5850 // only affects the number of attempts made to get work from the 5851 // overflow list and does not affect the number of workers. Just 5852 // pass ParallelGCThreads so this behavior is unchanged. 5853 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 5854 work_q, 5855 ParallelGCThreads)) { 5856 // Found something in global overflow list; 5857 // not yet ready to go stealing work from others. 5858 // We'd like to assert(work_q->size() != 0, ...) 5859 // because we just took work from the overflow list, 5860 // but of course we can't, since all of that might have 5861 // been already stolen from us. 5862 continue; 5863 } 5864 // Verify that we have no work before we resort to stealing 5865 assert(work_q->size() == 0, "Have work, shouldn't steal"); 5866 // Try to steal from other queues that have work 5867 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 5868 NOT_PRODUCT(num_steals++;) 5869 assert(obj_to_scan->is_oop(), "Oops, not an oop!"); 5870 assert(_mark_bit_map->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 5871 // Do scanning work 5872 obj_to_scan->oop_iterate(keep_alive); 5873 // Loop around, finish this work, and try to steal some more 5874 } else if (terminator()->offer_termination()) { 5875 break; // nirvana from the infinite cycle 5876 } 5877 } 5878 NOT_PRODUCT( 5879 if (PrintCMSStatistics != 0) { 5880 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals); 5881 } 5882 ) 5883 } 5884 5885 void CMSRefProcTaskExecutor::execute(ProcessTask& task) 5886 { 5887 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5888 FlexibleWorkGang* workers = gch->workers(); 5889 assert(workers != NULL, "Need parallel worker threads."); 5890 CMSRefProcTaskProxy rp_task(task, &_collector, 5891 _collector.ref_processor()->span(), 5892 _collector.markBitMap(), 5893 workers, _collector.task_queues()); 5894 workers->run_task(&rp_task); 5895 } 5896 5897 void CMSRefProcTaskExecutor::execute(EnqueueTask& task) 5898 { 5899 5900 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5901 FlexibleWorkGang* workers = gch->workers(); 5902 assert(workers != NULL, "Need parallel worker threads."); 5903 CMSRefEnqueueTaskProxy enq_task(task); 5904 workers->run_task(&enq_task); 5905 } 5906 5907 void CMSCollector::refProcessingWork(bool asynch, bool clear_all_soft_refs) { 5908 5909 ResourceMark rm; 5910 HandleMark hm; 5911 5912 ReferenceProcessor* rp = ref_processor(); 5913 assert(rp->span().equals(_span), "Spans should be equal"); 5914 assert(!rp->enqueuing_is_done(), "Enqueuing should not be complete"); 5915 // Process weak references. 5916 rp->setup_policy(clear_all_soft_refs); 5917 verify_work_stacks_empty(); 5918 5919 CMSKeepAliveClosure cmsKeepAliveClosure(this, _span, &_markBitMap, 5920 &_markStack, false /* !preclean */); 5921 CMSDrainMarkingStackClosure cmsDrainMarkingStackClosure(this, 5922 _span, &_markBitMap, &_markStack, 5923 &cmsKeepAliveClosure, false /* !preclean */); 5924 { 5925 TraceTime t("weak refs processing", PrintGCDetails, false, gclog_or_tty); 5926 if (rp->processing_is_mt()) { 5927 // Set the degree of MT here. If the discovery is done MT, there 5928 // may have been a different number of threads doing the discovery 5929 // and a different number of discovered lists may have Ref objects. 5930 // That is OK as long as the Reference lists are balanced (see 5931 // balance_all_queues() and balance_queues()). 5932 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5933 int active_workers = ParallelGCThreads; 5934 FlexibleWorkGang* workers = gch->workers(); 5935 if (workers != NULL) { 5936 active_workers = workers->active_workers(); 5937 // The expectation is that active_workers will have already 5938 // been set to a reasonable value. If it has not been set, 5939 // investigate. 5940 assert(active_workers > 0, "Should have been set during scavenge"); 5941 } 5942 rp->set_active_mt_degree(active_workers); 5943 CMSRefProcTaskExecutor task_executor(*this); 5944 rp->process_discovered_references(&_is_alive_closure, 5945 &cmsKeepAliveClosure, 5946 &cmsDrainMarkingStackClosure, 5947 &task_executor); 5948 } else { 5949 rp->process_discovered_references(&_is_alive_closure, 5950 &cmsKeepAliveClosure, 5951 &cmsDrainMarkingStackClosure, 5952 NULL); 5953 } 5954 verify_work_stacks_empty(); 5955 } 5956 5957 if (should_unload_classes()) { 5958 { 5959 TraceTime t("class unloading", PrintGCDetails, false, gclog_or_tty); 5960 5961 // Follow SystemDictionary roots and unload classes 5962 bool purged_class = SystemDictionary::do_unloading(&_is_alive_closure); 5963 5964 // Follow CodeCache roots and unload any methods marked for unloading 5965 CodeCache::do_unloading(&_is_alive_closure, 5966 &cmsKeepAliveClosure, 5967 purged_class); 5968 5969 cmsDrainMarkingStackClosure.do_void(); 5970 verify_work_stacks_empty(); 5971 5972 // Update subklass/sibling/implementor links in KlassKlass descendants 5973 Klass::clean_weak_klass_links(&_is_alive_closure); 5974 // Nothing should have been pushed onto the working stacks. 5975 verify_work_stacks_empty(); 5976 } 5977 5978 { 5979 TraceTime t("scrub symbol table", PrintGCDetails, false, gclog_or_tty); 5980 // Clean up unreferenced symbols in symbol table. 5981 SymbolTable::unlink(); 5982 } 5983 } 5984 5985 // CMS doesn't use the StringTable as hard roots when class unloading is turned off. 5986 // Need to check if we really scanned the StringTable. 5987 if ((roots_scanning_options() & SharedHeap::SO_Strings) == 0) { 5988 TraceTime t("scrub string table", PrintGCDetails, false, gclog_or_tty); 5989 // Now clean up stale oops in StringTable 5990 StringTable::unlink(&_is_alive_closure); 5991 } 5992 5993 verify_work_stacks_empty(); 5994 // Restore any preserved marks as a result of mark stack or 5995 // work queue overflow 5996 restore_preserved_marks_if_any(); // done single-threaded for now 5997 5998 rp->set_enqueuing_is_done(true); 5999 if (rp->processing_is_mt()) { 6000 rp->balance_all_queues(); 6001 CMSRefProcTaskExecutor task_executor(*this); 6002 rp->enqueue_discovered_references(&task_executor); 6003 } else { 6004 rp->enqueue_discovered_references(NULL); 6005 } 6006 rp->verify_no_references_recorded(); 6007 assert(!rp->discovery_enabled(), "should have been disabled"); 6008 } 6009 6010 #ifndef PRODUCT 6011 void CMSCollector::check_correct_thread_executing() { 6012 Thread* t = Thread::current(); 6013 // Only the VM thread or the CMS thread should be here. 6014 assert(t->is_ConcurrentGC_thread() || t->is_VM_thread(), 6015 "Unexpected thread type"); 6016 // If this is the vm thread, the foreground process 6017 // should not be waiting. Note that _foregroundGCIsActive is 6018 // true while the foreground collector is waiting. 6019 if (_foregroundGCShouldWait) { 6020 // We cannot be the VM thread 6021 assert(t->is_ConcurrentGC_thread(), 6022 "Should be CMS thread"); 6023 } else { 6024 // We can be the CMS thread only if we are in a stop-world 6025 // phase of CMS collection. 6026 if (t->is_ConcurrentGC_thread()) { 6027 assert(_collectorState == InitialMarking || 6028 _collectorState == FinalMarking, 6029 "Should be a stop-world phase"); 6030 // The CMS thread should be holding the CMS_token. 6031 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6032 "Potential interference with concurrently " 6033 "executing VM thread"); 6034 } 6035 } 6036 } 6037 #endif 6038 6039 void CMSCollector::sweep(bool asynch) { 6040 assert(_collectorState == Sweeping, "just checking"); 6041 check_correct_thread_executing(); 6042 verify_work_stacks_empty(); 6043 verify_overflow_empty(); 6044 increment_sweep_count(); 6045 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 6046 6047 _inter_sweep_timer.stop(); 6048 _inter_sweep_estimate.sample(_inter_sweep_timer.seconds()); 6049 size_policy()->avg_cms_free_at_sweep()->sample(_cmsGen->free()); 6050 6051 assert(!_intra_sweep_timer.is_active(), "Should not be active"); 6052 _intra_sweep_timer.reset(); 6053 _intra_sweep_timer.start(); 6054 if (asynch) { 6055 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 6056 CMSPhaseAccounting pa(this, "sweep", !PrintGCDetails); 6057 // First sweep the old gen 6058 { 6059 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(), 6060 bitMapLock()); 6061 sweepWork(_cmsGen, asynch); 6062 } 6063 6064 // Update Universe::_heap_*_at_gc figures. 6065 // We need all the free list locks to make the abstract state 6066 // transition from Sweeping to Resetting. See detailed note 6067 // further below. 6068 { 6069 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock()); 6070 // Update heap occupancy information which is used as 6071 // input to soft ref clearing policy at the next gc. 6072 Universe::update_heap_info_at_gc(); 6073 _collectorState = Resizing; 6074 } 6075 } else { 6076 // already have needed locks 6077 sweepWork(_cmsGen, asynch); 6078 // Update heap occupancy information which is used as 6079 // input to soft ref clearing policy at the next gc. 6080 Universe::update_heap_info_at_gc(); 6081 _collectorState = Resizing; 6082 } 6083 verify_work_stacks_empty(); 6084 verify_overflow_empty(); 6085 6086 _intra_sweep_timer.stop(); 6087 _intra_sweep_estimate.sample(_intra_sweep_timer.seconds()); 6088 6089 _inter_sweep_timer.reset(); 6090 _inter_sweep_timer.start(); 6091 6092 // We need to use a monotonically non-deccreasing time in ms 6093 // or we will see time-warp warnings and os::javaTimeMillis() 6094 // does not guarantee monotonicity. 6095 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC; 6096 update_time_of_last_gc(now); 6097 6098 // NOTE on abstract state transitions: 6099 // Mutators allocate-live and/or mark the mod-union table dirty 6100 // based on the state of the collection. The former is done in 6101 // the interval [Marking, Sweeping] and the latter in the interval 6102 // [Marking, Sweeping). Thus the transitions into the Marking state 6103 // and out of the Sweeping state must be synchronously visible 6104 // globally to the mutators. 6105 // The transition into the Marking state happens with the world 6106 // stopped so the mutators will globally see it. Sweeping is 6107 // done asynchronously by the background collector so the transition 6108 // from the Sweeping state to the Resizing state must be done 6109 // under the freelistLock (as is the check for whether to 6110 // allocate-live and whether to dirty the mod-union table). 6111 assert(_collectorState == Resizing, "Change of collector state to" 6112 " Resizing must be done under the freelistLocks (plural)"); 6113 6114 // Now that sweeping has been completed, we clear 6115 // the incremental_collection_failed flag, 6116 // thus inviting a younger gen collection to promote into 6117 // this generation. If such a promotion may still fail, 6118 // the flag will be set again when a young collection is 6119 // attempted. 6120 GenCollectedHeap* gch = GenCollectedHeap::heap(); 6121 gch->clear_incremental_collection_failed(); // Worth retrying as fresh space may have been freed up 6122 gch->update_full_collections_completed(_collection_count_start); 6123 } 6124 6125 // FIX ME!!! Looks like this belongs in CFLSpace, with 6126 // CMSGen merely delegating to it. 6127 void ConcurrentMarkSweepGeneration::setNearLargestChunk() { 6128 double nearLargestPercent = FLSLargestBlockCoalesceProximity; 6129 HeapWord* minAddr = _cmsSpace->bottom(); 6130 HeapWord* largestAddr = 6131 (HeapWord*) _cmsSpace->dictionary()->find_largest_dict(); 6132 if (largestAddr == NULL) { 6133 // The dictionary appears to be empty. In this case 6134 // try to coalesce at the end of the heap. 6135 largestAddr = _cmsSpace->end(); 6136 } 6137 size_t largestOffset = pointer_delta(largestAddr, minAddr); 6138 size_t nearLargestOffset = 6139 (size_t)((double)largestOffset * nearLargestPercent) - MinChunkSize; 6140 if (PrintFLSStatistics != 0) { 6141 gclog_or_tty->print_cr( 6142 "CMS: Large Block: " PTR_FORMAT ";" 6143 " Proximity: " PTR_FORMAT " -> " PTR_FORMAT, 6144 largestAddr, 6145 _cmsSpace->nearLargestChunk(), minAddr + nearLargestOffset); 6146 } 6147 _cmsSpace->set_nearLargestChunk(minAddr + nearLargestOffset); 6148 } 6149 6150 bool ConcurrentMarkSweepGeneration::isNearLargestChunk(HeapWord* addr) { 6151 return addr >= _cmsSpace->nearLargestChunk(); 6152 } 6153 6154 FreeChunk* ConcurrentMarkSweepGeneration::find_chunk_at_end() { 6155 return _cmsSpace->find_chunk_at_end(); 6156 } 6157 6158 void ConcurrentMarkSweepGeneration::update_gc_stats(int current_level, 6159 bool full) { 6160 // The next lower level has been collected. Gather any statistics 6161 // that are of interest at this point. 6162 if (!full && (current_level + 1) == level()) { 6163 // Gather statistics on the young generation collection. 6164 collector()->stats().record_gc0_end(used()); 6165 } 6166 } 6167 6168 CMSAdaptiveSizePolicy* ConcurrentMarkSweepGeneration::size_policy() { 6169 GenCollectedHeap* gch = GenCollectedHeap::heap(); 6170 assert(gch->kind() == CollectedHeap::GenCollectedHeap, 6171 "Wrong type of heap"); 6172 CMSAdaptiveSizePolicy* sp = (CMSAdaptiveSizePolicy*) 6173 gch->gen_policy()->size_policy(); 6174 assert(sp->is_gc_cms_adaptive_size_policy(), 6175 "Wrong type of size policy"); 6176 return sp; 6177 } 6178 6179 void ConcurrentMarkSweepGeneration::rotate_debug_collection_type() { 6180 if (PrintGCDetails && Verbose) { 6181 gclog_or_tty->print("Rotate from %d ", _debug_collection_type); 6182 } 6183 _debug_collection_type = (CollectionTypes) (_debug_collection_type + 1); 6184 _debug_collection_type = 6185 (CollectionTypes) (_debug_collection_type % Unknown_collection_type); 6186 if (PrintGCDetails && Verbose) { 6187 gclog_or_tty->print_cr("to %d ", _debug_collection_type); 6188 } 6189 } 6190 6191 void CMSCollector::sweepWork(ConcurrentMarkSweepGeneration* gen, 6192 bool asynch) { 6193 // We iterate over the space(s) underlying this generation, 6194 // checking the mark bit map to see if the bits corresponding 6195 // to specific blocks are marked or not. Blocks that are 6196 // marked are live and are not swept up. All remaining blocks 6197 // are swept up, with coalescing on-the-fly as we sweep up 6198 // contiguous free and/or garbage blocks: 6199 // We need to ensure that the sweeper synchronizes with allocators 6200 // and stop-the-world collectors. In particular, the following 6201 // locks are used: 6202 // . CMS token: if this is held, a stop the world collection cannot occur 6203 // . freelistLock: if this is held no allocation can occur from this 6204 // generation by another thread 6205 // . bitMapLock: if this is held, no other thread can access or update 6206 // 6207 6208 // Note that we need to hold the freelistLock if we use 6209 // block iterate below; else the iterator might go awry if 6210 // a mutator (or promotion) causes block contents to change 6211 // (for instance if the allocator divvies up a block). 6212 // If we hold the free list lock, for all practical purposes 6213 // young generation GC's can't occur (they'll usually need to 6214 // promote), so we might as well prevent all young generation 6215 // GC's while we do a sweeping step. For the same reason, we might 6216 // as well take the bit map lock for the entire duration 6217 6218 // check that we hold the requisite locks 6219 assert(have_cms_token(), "Should hold cms token"); 6220 assert( (asynch && ConcurrentMarkSweepThread::cms_thread_has_cms_token()) 6221 || (!asynch && ConcurrentMarkSweepThread::vm_thread_has_cms_token()), 6222 "Should possess CMS token to sweep"); 6223 assert_lock_strong(gen->freelistLock()); 6224 assert_lock_strong(bitMapLock()); 6225 6226 assert(!_inter_sweep_timer.is_active(), "Was switched off in an outer context"); 6227 assert(_intra_sweep_timer.is_active(), "Was switched on in an outer context"); 6228 gen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()), 6229 _inter_sweep_estimate.padded_average(), 6230 _intra_sweep_estimate.padded_average()); 6231 gen->setNearLargestChunk(); 6232 6233 { 6234 SweepClosure sweepClosure(this, gen, &_markBitMap, 6235 CMSYield && asynch); 6236 gen->cmsSpace()->blk_iterate_careful(&sweepClosure); 6237 // We need to free-up/coalesce garbage/blocks from a 6238 // co-terminal free run. This is done in the SweepClosure 6239 // destructor; so, do not remove this scope, else the 6240 // end-of-sweep-census below will be off by a little bit. 6241 } 6242 gen->cmsSpace()->sweep_completed(); 6243 gen->cmsSpace()->endSweepFLCensus(sweep_count()); 6244 if (should_unload_classes()) { // unloaded classes this cycle, 6245 _concurrent_cycles_since_last_unload = 0; // ... reset count 6246 } else { // did not unload classes, 6247 _concurrent_cycles_since_last_unload++; // ... increment count 6248 } 6249 } 6250 6251 // Reset CMS data structures (for now just the marking bit map) 6252 // preparatory for the next cycle. 6253 void CMSCollector::reset(bool asynch) { 6254 GenCollectedHeap* gch = GenCollectedHeap::heap(); 6255 CMSAdaptiveSizePolicy* sp = size_policy(); 6256 AdaptiveSizePolicyOutput(sp, gch->total_collections()); 6257 if (asynch) { 6258 CMSTokenSyncWithLocks ts(true, bitMapLock()); 6259 6260 // If the state is not "Resetting", the foreground thread 6261 // has done a collection and the resetting. 6262 if (_collectorState != Resetting) { 6263 assert(_collectorState == Idling, "The state should only change" 6264 " because the foreground collector has finished the collection"); 6265 return; 6266 } 6267 6268 // Clear the mark bitmap (no grey objects to start with) 6269 // for the next cycle. 6270 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 6271 CMSPhaseAccounting cmspa(this, "reset", !PrintGCDetails); 6272 6273 HeapWord* curAddr = _markBitMap.startWord(); 6274 while (curAddr < _markBitMap.endWord()) { 6275 size_t remaining = pointer_delta(_markBitMap.endWord(), curAddr); 6276 MemRegion chunk(curAddr, MIN2(CMSBitMapYieldQuantum, remaining)); 6277 _markBitMap.clear_large_range(chunk); 6278 if (ConcurrentMarkSweepThread::should_yield() && 6279 !foregroundGCIsActive() && 6280 CMSYield) { 6281 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6282 "CMS thread should hold CMS token"); 6283 assert_lock_strong(bitMapLock()); 6284 bitMapLock()->unlock(); 6285 ConcurrentMarkSweepThread::desynchronize(true); 6286 ConcurrentMarkSweepThread::acknowledge_yield_request(); 6287 stopTimer(); 6288 if (PrintCMSStatistics != 0) { 6289 incrementYields(); 6290 } 6291 icms_wait(); 6292 6293 // See the comment in coordinator_yield() 6294 for (unsigned i = 0; i < CMSYieldSleepCount && 6295 ConcurrentMarkSweepThread::should_yield() && 6296 !CMSCollector::foregroundGCIsActive(); ++i) { 6297 os::sleep(Thread::current(), 1, false); 6298 ConcurrentMarkSweepThread::acknowledge_yield_request(); 6299 } 6300 6301 ConcurrentMarkSweepThread::synchronize(true); 6302 bitMapLock()->lock_without_safepoint_check(); 6303 startTimer(); 6304 } 6305 curAddr = chunk.end(); 6306 } 6307 // A successful mostly concurrent collection has been done. 6308 // Because only the full (i.e., concurrent mode failure) collections 6309 // are being measured for gc overhead limits, clean the "near" flag 6310 // and count. 6311 sp->reset_gc_overhead_limit_count(); 6312 _collectorState = Idling; 6313 } else { 6314 // already have the lock 6315 assert(_collectorState == Resetting, "just checking"); 6316 assert_lock_strong(bitMapLock()); 6317 _markBitMap.clear_all(); 6318 _collectorState = Idling; 6319 } 6320 6321 // Stop incremental mode after a cycle completes, so that any future cycles 6322 // are triggered by allocation. 6323 stop_icms(); 6324 6325 NOT_PRODUCT( 6326 if (RotateCMSCollectionTypes) { 6327 _cmsGen->rotate_debug_collection_type(); 6328 } 6329 ) 6330 } 6331 6332 void CMSCollector::do_CMS_operation(CMS_op_type op, GCCause::Cause gc_cause) { 6333 gclog_or_tty->date_stamp(PrintGC && PrintGCDateStamps); 6334 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 6335 TraceTime t(GCCauseString("GC", gc_cause), PrintGC, !PrintGCDetails, gclog_or_tty); 6336 TraceCollectorStats tcs(counters()); 6337 6338 switch (op) { 6339 case CMS_op_checkpointRootsInitial: { 6340 SvcGCMarker sgcm(SvcGCMarker::OTHER); 6341 checkpointRootsInitial(true); // asynch 6342 if (PrintGC) { 6343 _cmsGen->printOccupancy("initial-mark"); 6344 } 6345 break; 6346 } 6347 case CMS_op_checkpointRootsFinal: { 6348 SvcGCMarker sgcm(SvcGCMarker::OTHER); 6349 checkpointRootsFinal(true, // asynch 6350 false, // !clear_all_soft_refs 6351 false); // !init_mark_was_synchronous 6352 if (PrintGC) { 6353 _cmsGen->printOccupancy("remark"); 6354 } 6355 break; 6356 } 6357 default: 6358 fatal("No such CMS_op"); 6359 } 6360 } 6361 6362 #ifndef PRODUCT 6363 size_t const CMSCollector::skip_header_HeapWords() { 6364 return FreeChunk::header_size(); 6365 } 6366 6367 // Try and collect here conditions that should hold when 6368 // CMS thread is exiting. The idea is that the foreground GC 6369 // thread should not be blocked if it wants to terminate 6370 // the CMS thread and yet continue to run the VM for a while 6371 // after that. 6372 void CMSCollector::verify_ok_to_terminate() const { 6373 assert(Thread::current()->is_ConcurrentGC_thread(), 6374 "should be called by CMS thread"); 6375 assert(!_foregroundGCShouldWait, "should be false"); 6376 // We could check here that all the various low-level locks 6377 // are not held by the CMS thread, but that is overkill; see 6378 // also CMSThread::verify_ok_to_terminate() where the CGC_lock 6379 // is checked. 6380 } 6381 #endif 6382 6383 size_t CMSCollector::block_size_using_printezis_bits(HeapWord* addr) const { 6384 assert(_markBitMap.isMarked(addr) && _markBitMap.isMarked(addr + 1), 6385 "missing Printezis mark?"); 6386 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 6387 size_t size = pointer_delta(nextOneAddr + 1, addr); 6388 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 6389 "alignment problem"); 6390 assert(size >= 3, "Necessary for Printezis marks to work"); 6391 return size; 6392 } 6393 6394 // A variant of the above (block_size_using_printezis_bits()) except 6395 // that we return 0 if the P-bits are not yet set. 6396 size_t CMSCollector::block_size_if_printezis_bits(HeapWord* addr) const { 6397 if (_markBitMap.isMarked(addr + 1)) { 6398 assert(_markBitMap.isMarked(addr), "P-bit can be set only for marked objects"); 6399 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 6400 size_t size = pointer_delta(nextOneAddr + 1, addr); 6401 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 6402 "alignment problem"); 6403 assert(size >= 3, "Necessary for Printezis marks to work"); 6404 return size; 6405 } 6406 return 0; 6407 } 6408 6409 HeapWord* CMSCollector::next_card_start_after_block(HeapWord* addr) const { 6410 size_t sz = 0; 6411 oop p = (oop)addr; 6412 if (p->klass_or_null() != NULL) { 6413 sz = CompactibleFreeListSpace::adjustObjectSize(p->size()); 6414 } else { 6415 sz = block_size_using_printezis_bits(addr); 6416 } 6417 assert(sz > 0, "size must be nonzero"); 6418 HeapWord* next_block = addr + sz; 6419 HeapWord* next_card = (HeapWord*)round_to((uintptr_t)next_block, 6420 CardTableModRefBS::card_size); 6421 assert(round_down((uintptr_t)addr, CardTableModRefBS::card_size) < 6422 round_down((uintptr_t)next_card, CardTableModRefBS::card_size), 6423 "must be different cards"); 6424 return next_card; 6425 } 6426 6427 6428 // CMS Bit Map Wrapper ///////////////////////////////////////// 6429 6430 // Construct a CMS bit map infrastructure, but don't create the 6431 // bit vector itself. That is done by a separate call CMSBitMap::allocate() 6432 // further below. 6433 CMSBitMap::CMSBitMap(int shifter, int mutex_rank, const char* mutex_name): 6434 _bm(), 6435 _shifter(shifter), 6436 _lock(mutex_rank >= 0 ? new Mutex(mutex_rank, mutex_name, true) : NULL) 6437 { 6438 _bmStartWord = 0; 6439 _bmWordSize = 0; 6440 } 6441 6442 bool CMSBitMap::allocate(MemRegion mr) { 6443 _bmStartWord = mr.start(); 6444 _bmWordSize = mr.word_size(); 6445 ReservedSpace brs(ReservedSpace::allocation_align_size_up( 6446 (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1)); 6447 if (!brs.is_reserved()) { 6448 warning("CMS bit map allocation failure"); 6449 return false; 6450 } 6451 // For now we'll just commit all of the bit map up fromt. 6452 // Later on we'll try to be more parsimonious with swap. 6453 if (!_virtual_space.initialize(brs, brs.size())) { 6454 warning("CMS bit map backing store failure"); 6455 return false; 6456 } 6457 assert(_virtual_space.committed_size() == brs.size(), 6458 "didn't reserve backing store for all of CMS bit map?"); 6459 _bm.set_map((BitMap::bm_word_t*)_virtual_space.low()); 6460 assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >= 6461 _bmWordSize, "inconsistency in bit map sizing"); 6462 _bm.set_size(_bmWordSize >> _shifter); 6463 6464 // bm.clear(); // can we rely on getting zero'd memory? verify below 6465 assert(isAllClear(), 6466 "Expected zero'd memory from ReservedSpace constructor"); 6467 assert(_bm.size() == heapWordDiffToOffsetDiff(sizeInWords()), 6468 "consistency check"); 6469 return true; 6470 } 6471 6472 void CMSBitMap::dirty_range_iterate_clear(MemRegion mr, MemRegionClosure* cl) { 6473 HeapWord *next_addr, *end_addr, *last_addr; 6474 assert_locked(); 6475 assert(covers(mr), "out-of-range error"); 6476 // XXX assert that start and end are appropriately aligned 6477 for (next_addr = mr.start(), end_addr = mr.end(); 6478 next_addr < end_addr; next_addr = last_addr) { 6479 MemRegion dirty_region = getAndClearMarkedRegion(next_addr, end_addr); 6480 last_addr = dirty_region.end(); 6481 if (!dirty_region.is_empty()) { 6482 cl->do_MemRegion(dirty_region); 6483 } else { 6484 assert(last_addr == end_addr, "program logic"); 6485 return; 6486 } 6487 } 6488 } 6489 6490 #ifndef PRODUCT 6491 void CMSBitMap::assert_locked() const { 6492 CMSLockVerifier::assert_locked(lock()); 6493 } 6494 6495 bool CMSBitMap::covers(MemRegion mr) const { 6496 // assert(_bm.map() == _virtual_space.low(), "map inconsistency"); 6497 assert((size_t)_bm.size() == (_bmWordSize >> _shifter), 6498 "size inconsistency"); 6499 return (mr.start() >= _bmStartWord) && 6500 (mr.end() <= endWord()); 6501 } 6502 6503 bool CMSBitMap::covers(HeapWord* start, size_t size) const { 6504 return (start >= _bmStartWord && (start + size) <= endWord()); 6505 } 6506 6507 void CMSBitMap::verifyNoOneBitsInRange(HeapWord* left, HeapWord* right) { 6508 // verify that there are no 1 bits in the interval [left, right) 6509 FalseBitMapClosure falseBitMapClosure; 6510 iterate(&falseBitMapClosure, left, right); 6511 } 6512 6513 void CMSBitMap::region_invariant(MemRegion mr) 6514 { 6515 assert_locked(); 6516 // mr = mr.intersection(MemRegion(_bmStartWord, _bmWordSize)); 6517 assert(!mr.is_empty(), "unexpected empty region"); 6518 assert(covers(mr), "mr should be covered by bit map"); 6519 // convert address range into offset range 6520 size_t start_ofs = heapWordToOffset(mr.start()); 6521 // Make sure that end() is appropriately aligned 6522 assert(mr.end() == (HeapWord*)round_to((intptr_t)mr.end(), 6523 (1 << (_shifter+LogHeapWordSize))), 6524 "Misaligned mr.end()"); 6525 size_t end_ofs = heapWordToOffset(mr.end()); 6526 assert(end_ofs > start_ofs, "Should mark at least one bit"); 6527 } 6528 6529 #endif 6530 6531 bool CMSMarkStack::allocate(size_t size) { 6532 // allocate a stack of the requisite depth 6533 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 6534 size * sizeof(oop))); 6535 if (!rs.is_reserved()) { 6536 warning("CMSMarkStack allocation failure"); 6537 return false; 6538 } 6539 if (!_virtual_space.initialize(rs, rs.size())) { 6540 warning("CMSMarkStack backing store failure"); 6541 return false; 6542 } 6543 assert(_virtual_space.committed_size() == rs.size(), 6544 "didn't reserve backing store for all of CMS stack?"); 6545 _base = (oop*)(_virtual_space.low()); 6546 _index = 0; 6547 _capacity = size; 6548 NOT_PRODUCT(_max_depth = 0); 6549 return true; 6550 } 6551 6552 // XXX FIX ME !!! In the MT case we come in here holding a 6553 // leaf lock. For printing we need to take a further lock 6554 // which has lower rank. We need to recallibrate the two 6555 // lock-ranks involved in order to be able to rpint the 6556 // messages below. (Or defer the printing to the caller. 6557 // For now we take the expedient path of just disabling the 6558 // messages for the problematic case.) 6559 void CMSMarkStack::expand() { 6560 assert(_capacity <= MarkStackSizeMax, "stack bigger than permitted"); 6561 if (_capacity == MarkStackSizeMax) { 6562 if (_hit_limit++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) { 6563 // We print a warning message only once per CMS cycle. 6564 gclog_or_tty->print_cr(" (benign) Hit CMSMarkStack max size limit"); 6565 } 6566 return; 6567 } 6568 // Double capacity if possible 6569 size_t new_capacity = MIN2(_capacity*2, MarkStackSizeMax); 6570 // Do not give up existing stack until we have managed to 6571 // get the double capacity that we desired. 6572 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 6573 new_capacity * sizeof(oop))); 6574 if (rs.is_reserved()) { 6575 // Release the backing store associated with old stack 6576 _virtual_space.release(); 6577 // Reinitialize virtual space for new stack 6578 if (!_virtual_space.initialize(rs, rs.size())) { 6579 fatal("Not enough swap for expanded marking stack"); 6580 } 6581 _base = (oop*)(_virtual_space.low()); 6582 _index = 0; 6583 _capacity = new_capacity; 6584 } else if (_failed_double++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) { 6585 // Failed to double capacity, continue; 6586 // we print a detail message only once per CMS cycle. 6587 gclog_or_tty->print(" (benign) Failed to expand marking stack from "SIZE_FORMAT"K to " 6588 SIZE_FORMAT"K", 6589 _capacity / K, new_capacity / K); 6590 } 6591 } 6592 6593 6594 // Closures 6595 // XXX: there seems to be a lot of code duplication here; 6596 // should refactor and consolidate common code. 6597 6598 // This closure is used to mark refs into the CMS generation in 6599 // the CMS bit map. Called at the first checkpoint. This closure 6600 // assumes that we do not need to re-mark dirty cards; if the CMS 6601 // generation on which this is used is not an oldest 6602 // generation then this will lose younger_gen cards! 6603 6604 MarkRefsIntoClosure::MarkRefsIntoClosure( 6605 MemRegion span, CMSBitMap* bitMap): 6606 _span(span), 6607 _bitMap(bitMap) 6608 { 6609 assert(_ref_processor == NULL, "deliberately left NULL"); 6610 assert(_bitMap->covers(_span), "_bitMap/_span mismatch"); 6611 } 6612 6613 void MarkRefsIntoClosure::do_oop(oop obj) { 6614 // if p points into _span, then mark corresponding bit in _markBitMap 6615 assert(obj->is_oop(), "expected an oop"); 6616 HeapWord* addr = (HeapWord*)obj; 6617 if (_span.contains(addr)) { 6618 // this should be made more efficient 6619 _bitMap->mark(addr); 6620 } 6621 } 6622 6623 void MarkRefsIntoClosure::do_oop(oop* p) { MarkRefsIntoClosure::do_oop_work(p); } 6624 void MarkRefsIntoClosure::do_oop(narrowOop* p) { MarkRefsIntoClosure::do_oop_work(p); } 6625 6626 // A variant of the above, used for CMS marking verification. 6627 MarkRefsIntoVerifyClosure::MarkRefsIntoVerifyClosure( 6628 MemRegion span, CMSBitMap* verification_bm, CMSBitMap* cms_bm): 6629 _span(span), 6630 _verification_bm(verification_bm), 6631 _cms_bm(cms_bm) 6632 { 6633 assert(_ref_processor == NULL, "deliberately left NULL"); 6634 assert(_verification_bm->covers(_span), "_verification_bm/_span mismatch"); 6635 } 6636 6637 void MarkRefsIntoVerifyClosure::do_oop(oop obj) { 6638 // if p points into _span, then mark corresponding bit in _markBitMap 6639 assert(obj->is_oop(), "expected an oop"); 6640 HeapWord* addr = (HeapWord*)obj; 6641 if (_span.contains(addr)) { 6642 _verification_bm->mark(addr); 6643 if (!_cms_bm->isMarked(addr)) { 6644 oop(addr)->print(); 6645 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)", addr); 6646 fatal("... aborting"); 6647 } 6648 } 6649 } 6650 6651 void MarkRefsIntoVerifyClosure::do_oop(oop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); } 6652 void MarkRefsIntoVerifyClosure::do_oop(narrowOop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); } 6653 6654 ////////////////////////////////////////////////// 6655 // MarkRefsIntoAndScanClosure 6656 ////////////////////////////////////////////////// 6657 6658 MarkRefsIntoAndScanClosure::MarkRefsIntoAndScanClosure(MemRegion span, 6659 ReferenceProcessor* rp, 6660 CMSBitMap* bit_map, 6661 CMSBitMap* mod_union_table, 6662 CMSMarkStack* mark_stack, 6663 CMSCollector* collector, 6664 bool should_yield, 6665 bool concurrent_precleaning): 6666 _collector(collector), 6667 _span(span), 6668 _bit_map(bit_map), 6669 _mark_stack(mark_stack), 6670 _pushAndMarkClosure(collector, span, rp, bit_map, mod_union_table, 6671 mark_stack, concurrent_precleaning), 6672 _yield(should_yield), 6673 _concurrent_precleaning(concurrent_precleaning), 6674 _freelistLock(NULL) 6675 { 6676 _ref_processor = rp; 6677 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 6678 } 6679 6680 // This closure is used to mark refs into the CMS generation at the 6681 // second (final) checkpoint, and to scan and transitively follow 6682 // the unmarked oops. It is also used during the concurrent precleaning 6683 // phase while scanning objects on dirty cards in the CMS generation. 6684 // The marks are made in the marking bit map and the marking stack is 6685 // used for keeping the (newly) grey objects during the scan. 6686 // The parallel version (Par_...) appears further below. 6687 void MarkRefsIntoAndScanClosure::do_oop(oop obj) { 6688 if (obj != NULL) { 6689 assert(obj->is_oop(), "expected an oop"); 6690 HeapWord* addr = (HeapWord*)obj; 6691 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 6692 assert(_collector->overflow_list_is_empty(), 6693 "overflow list should be empty"); 6694 if (_span.contains(addr) && 6695 !_bit_map->isMarked(addr)) { 6696 // mark bit map (object is now grey) 6697 _bit_map->mark(addr); 6698 // push on marking stack (stack should be empty), and drain the 6699 // stack by applying this closure to the oops in the oops popped 6700 // from the stack (i.e. blacken the grey objects) 6701 bool res = _mark_stack->push(obj); 6702 assert(res, "Should have space to push on empty stack"); 6703 do { 6704 oop new_oop = _mark_stack->pop(); 6705 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 6706 assert(_bit_map->isMarked((HeapWord*)new_oop), 6707 "only grey objects on this stack"); 6708 // iterate over the oops in this oop, marking and pushing 6709 // the ones in CMS heap (i.e. in _span). 6710 new_oop->oop_iterate(&_pushAndMarkClosure); 6711 // check if it's time to yield 6712 do_yield_check(); 6713 } while (!_mark_stack->isEmpty() || 6714 (!_concurrent_precleaning && take_from_overflow_list())); 6715 // if marking stack is empty, and we are not doing this 6716 // during precleaning, then check the overflow list 6717 } 6718 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 6719 assert(_collector->overflow_list_is_empty(), 6720 "overflow list was drained above"); 6721 // We could restore evacuated mark words, if any, used for 6722 // overflow list links here because the overflow list is 6723 // provably empty here. That would reduce the maximum 6724 // size requirements for preserved_{oop,mark}_stack. 6725 // But we'll just postpone it until we are all done 6726 // so we can just stream through. 6727 if (!_concurrent_precleaning && CMSOverflowEarlyRestoration) { 6728 _collector->restore_preserved_marks_if_any(); 6729 assert(_collector->no_preserved_marks(), "No preserved marks"); 6730 } 6731 assert(!CMSOverflowEarlyRestoration || _collector->no_preserved_marks(), 6732 "All preserved marks should have been restored above"); 6733 } 6734 } 6735 6736 void MarkRefsIntoAndScanClosure::do_oop(oop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); } 6737 void MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); } 6738 6739 void MarkRefsIntoAndScanClosure::do_yield_work() { 6740 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6741 "CMS thread should hold CMS token"); 6742 assert_lock_strong(_freelistLock); 6743 assert_lock_strong(_bit_map->lock()); 6744 // relinquish the free_list_lock and bitMaplock() 6745 _bit_map->lock()->unlock(); 6746 _freelistLock->unlock(); 6747 ConcurrentMarkSweepThread::desynchronize(true); 6748 ConcurrentMarkSweepThread::acknowledge_yield_request(); 6749 _collector->stopTimer(); 6750 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 6751 if (PrintCMSStatistics != 0) { 6752 _collector->incrementYields(); 6753 } 6754 _collector->icms_wait(); 6755 6756 // See the comment in coordinator_yield() 6757 for (unsigned i = 0; 6758 i < CMSYieldSleepCount && 6759 ConcurrentMarkSweepThread::should_yield() && 6760 !CMSCollector::foregroundGCIsActive(); 6761 ++i) { 6762 os::sleep(Thread::current(), 1, false); 6763 ConcurrentMarkSweepThread::acknowledge_yield_request(); 6764 } 6765 6766 ConcurrentMarkSweepThread::synchronize(true); 6767 _freelistLock->lock_without_safepoint_check(); 6768 _bit_map->lock()->lock_without_safepoint_check(); 6769 _collector->startTimer(); 6770 } 6771 6772 /////////////////////////////////////////////////////////// 6773 // Par_MarkRefsIntoAndScanClosure: a parallel version of 6774 // MarkRefsIntoAndScanClosure 6775 /////////////////////////////////////////////////////////// 6776 Par_MarkRefsIntoAndScanClosure::Par_MarkRefsIntoAndScanClosure( 6777 CMSCollector* collector, MemRegion span, ReferenceProcessor* rp, 6778 CMSBitMap* bit_map, OopTaskQueue* work_queue): 6779 _span(span), 6780 _bit_map(bit_map), 6781 _work_queue(work_queue), 6782 _low_water_mark(MIN2((uint)(work_queue->max_elems()/4), 6783 (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads))), 6784 _par_pushAndMarkClosure(collector, span, rp, bit_map, work_queue) 6785 { 6786 _ref_processor = rp; 6787 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 6788 } 6789 6790 // This closure is used to mark refs into the CMS generation at the 6791 // second (final) checkpoint, and to scan and transitively follow 6792 // the unmarked oops. The marks are made in the marking bit map and 6793 // the work_queue is used for keeping the (newly) grey objects during 6794 // the scan phase whence they are also available for stealing by parallel 6795 // threads. Since the marking bit map is shared, updates are 6796 // synchronized (via CAS). 6797 void Par_MarkRefsIntoAndScanClosure::do_oop(oop obj) { 6798 if (obj != NULL) { 6799 // Ignore mark word because this could be an already marked oop 6800 // that may be chained at the end of the overflow list. 6801 assert(obj->is_oop(true), "expected an oop"); 6802 HeapWord* addr = (HeapWord*)obj; 6803 if (_span.contains(addr) && 6804 !_bit_map->isMarked(addr)) { 6805 // mark bit map (object will become grey): 6806 // It is possible for several threads to be 6807 // trying to "claim" this object concurrently; 6808 // the unique thread that succeeds in marking the 6809 // object first will do the subsequent push on 6810 // to the work queue (or overflow list). 6811 if (_bit_map->par_mark(addr)) { 6812 // push on work_queue (which may not be empty), and trim the 6813 // queue to an appropriate length by applying this closure to 6814 // the oops in the oops popped from the stack (i.e. blacken the 6815 // grey objects) 6816 bool res = _work_queue->push(obj); 6817 assert(res, "Low water mark should be less than capacity?"); 6818 trim_queue(_low_water_mark); 6819 } // Else, another thread claimed the object 6820 } 6821 } 6822 } 6823 6824 void Par_MarkRefsIntoAndScanClosure::do_oop(oop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); } 6825 void Par_MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); } 6826 6827 // This closure is used to rescan the marked objects on the dirty cards 6828 // in the mod union table and the card table proper. 6829 size_t ScanMarkedObjectsAgainCarefullyClosure::do_object_careful_m( 6830 oop p, MemRegion mr) { 6831 6832 size_t size = 0; 6833 HeapWord* addr = (HeapWord*)p; 6834 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6835 assert(_span.contains(addr), "we are scanning the CMS generation"); 6836 // check if it's time to yield 6837 if (do_yield_check()) { 6838 // We yielded for some foreground stop-world work, 6839 // and we have been asked to abort this ongoing preclean cycle. 6840 return 0; 6841 } 6842 if (_bitMap->isMarked(addr)) { 6843 // it's marked; is it potentially uninitialized? 6844 if (p->klass_or_null() != NULL) { 6845 // an initialized object; ignore mark word in verification below 6846 // since we are running concurrent with mutators 6847 assert(p->is_oop(true), "should be an oop"); 6848 if (p->is_objArray()) { 6849 // objArrays are precisely marked; restrict scanning 6850 // to dirty cards only. 6851 size = CompactibleFreeListSpace::adjustObjectSize( 6852 p->oop_iterate(_scanningClosure, mr)); 6853 } else { 6854 // A non-array may have been imprecisely marked; we need 6855 // to scan object in its entirety. 6856 size = CompactibleFreeListSpace::adjustObjectSize( 6857 p->oop_iterate(_scanningClosure)); 6858 } 6859 #ifdef DEBUG 6860 size_t direct_size = 6861 CompactibleFreeListSpace::adjustObjectSize(p->size()); 6862 assert(size == direct_size, "Inconsistency in size"); 6863 assert(size >= 3, "Necessary for Printezis marks to work"); 6864 if (!_bitMap->isMarked(addr+1)) { 6865 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size); 6866 } else { 6867 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size-1); 6868 assert(_bitMap->isMarked(addr+size-1), 6869 "inconsistent Printezis mark"); 6870 } 6871 #endif // DEBUG 6872 } else { 6873 // an unitialized object 6874 assert(_bitMap->isMarked(addr+1), "missing Printezis mark?"); 6875 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 6876 size = pointer_delta(nextOneAddr + 1, addr); 6877 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 6878 "alignment problem"); 6879 // Note that pre-cleaning needn't redirty the card. OopDesc::set_klass() 6880 // will dirty the card when the klass pointer is installed in the 6881 // object (signalling the completion of initialization). 6882 } 6883 } else { 6884 // Either a not yet marked object or an uninitialized object 6885 if (p->klass_or_null() == NULL) { 6886 // An uninitialized object, skip to the next card, since 6887 // we may not be able to read its P-bits yet. 6888 assert(size == 0, "Initial value"); 6889 } else { 6890 // An object not (yet) reached by marking: we merely need to 6891 // compute its size so as to go look at the next block. 6892 assert(p->is_oop(true), "should be an oop"); 6893 size = CompactibleFreeListSpace::adjustObjectSize(p->size()); 6894 } 6895 } 6896 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6897 return size; 6898 } 6899 6900 void ScanMarkedObjectsAgainCarefullyClosure::do_yield_work() { 6901 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6902 "CMS thread should hold CMS token"); 6903 assert_lock_strong(_freelistLock); 6904 assert_lock_strong(_bitMap->lock()); 6905 // relinquish the free_list_lock and bitMaplock() 6906 _bitMap->lock()->unlock(); 6907 _freelistLock->unlock(); 6908 ConcurrentMarkSweepThread::desynchronize(true); 6909 ConcurrentMarkSweepThread::acknowledge_yield_request(); 6910 _collector->stopTimer(); 6911 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 6912 if (PrintCMSStatistics != 0) { 6913 _collector->incrementYields(); 6914 } 6915 _collector->icms_wait(); 6916 6917 // See the comment in coordinator_yield() 6918 for (unsigned i = 0; i < CMSYieldSleepCount && 6919 ConcurrentMarkSweepThread::should_yield() && 6920 !CMSCollector::foregroundGCIsActive(); ++i) { 6921 os::sleep(Thread::current(), 1, false); 6922 ConcurrentMarkSweepThread::acknowledge_yield_request(); 6923 } 6924 6925 ConcurrentMarkSweepThread::synchronize(true); 6926 _freelistLock->lock_without_safepoint_check(); 6927 _bitMap->lock()->lock_without_safepoint_check(); 6928 _collector->startTimer(); 6929 } 6930 6931 6932 ////////////////////////////////////////////////////////////////// 6933 // SurvivorSpacePrecleanClosure 6934 ////////////////////////////////////////////////////////////////// 6935 // This (single-threaded) closure is used to preclean the oops in 6936 // the survivor spaces. 6937 size_t SurvivorSpacePrecleanClosure::do_object_careful(oop p) { 6938 6939 HeapWord* addr = (HeapWord*)p; 6940 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6941 assert(!_span.contains(addr), "we are scanning the survivor spaces"); 6942 assert(p->klass_or_null() != NULL, "object should be initializd"); 6943 // an initialized object; ignore mark word in verification below 6944 // since we are running concurrent with mutators 6945 assert(p->is_oop(true), "should be an oop"); 6946 // Note that we do not yield while we iterate over 6947 // the interior oops of p, pushing the relevant ones 6948 // on our marking stack. 6949 size_t size = p->oop_iterate(_scanning_closure); 6950 do_yield_check(); 6951 // Observe that below, we do not abandon the preclean 6952 // phase as soon as we should; rather we empty the 6953 // marking stack before returning. This is to satisfy 6954 // some existing assertions. In general, it may be a 6955 // good idea to abort immediately and complete the marking 6956 // from the grey objects at a later time. 6957 while (!_mark_stack->isEmpty()) { 6958 oop new_oop = _mark_stack->pop(); 6959 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 6960 assert(_bit_map->isMarked((HeapWord*)new_oop), 6961 "only grey objects on this stack"); 6962 // iterate over the oops in this oop, marking and pushing 6963 // the ones in CMS heap (i.e. in _span). 6964 new_oop->oop_iterate(_scanning_closure); 6965 // check if it's time to yield 6966 do_yield_check(); 6967 } 6968 unsigned int after_count = 6969 GenCollectedHeap::heap()->total_collections(); 6970 bool abort = (_before_count != after_count) || 6971 _collector->should_abort_preclean(); 6972 return abort ? 0 : size; 6973 } 6974 6975 void SurvivorSpacePrecleanClosure::do_yield_work() { 6976 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6977 "CMS thread should hold CMS token"); 6978 assert_lock_strong(_bit_map->lock()); 6979 // Relinquish the bit map lock 6980 _bit_map->lock()->unlock(); 6981 ConcurrentMarkSweepThread::desynchronize(true); 6982 ConcurrentMarkSweepThread::acknowledge_yield_request(); 6983 _collector->stopTimer(); 6984 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 6985 if (PrintCMSStatistics != 0) { 6986 _collector->incrementYields(); 6987 } 6988 _collector->icms_wait(); 6989 6990 // See the comment in coordinator_yield() 6991 for (unsigned i = 0; i < CMSYieldSleepCount && 6992 ConcurrentMarkSweepThread::should_yield() && 6993 !CMSCollector::foregroundGCIsActive(); ++i) { 6994 os::sleep(Thread::current(), 1, false); 6995 ConcurrentMarkSweepThread::acknowledge_yield_request(); 6996 } 6997 6998 ConcurrentMarkSweepThread::synchronize(true); 6999 _bit_map->lock()->lock_without_safepoint_check(); 7000 _collector->startTimer(); 7001 } 7002 7003 // This closure is used to rescan the marked objects on the dirty cards 7004 // in the mod union table and the card table proper. In the parallel 7005 // case, although the bitMap is shared, we do a single read so the 7006 // isMarked() query is "safe". 7007 bool ScanMarkedObjectsAgainClosure::do_object_bm(oop p, MemRegion mr) { 7008 // Ignore mark word because we are running concurrent with mutators 7009 assert(p->is_oop_or_null(true), "expected an oop or null"); 7010 HeapWord* addr = (HeapWord*)p; 7011 assert(_span.contains(addr), "we are scanning the CMS generation"); 7012 bool is_obj_array = false; 7013 #ifdef DEBUG 7014 if (!_parallel) { 7015 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 7016 assert(_collector->overflow_list_is_empty(), 7017 "overflow list should be empty"); 7018 7019 } 7020 #endif // DEBUG 7021 if (_bit_map->isMarked(addr)) { 7022 // Obj arrays are precisely marked, non-arrays are not; 7023 // so we scan objArrays precisely and non-arrays in their 7024 // entirety. 7025 if (p->is_objArray()) { 7026 is_obj_array = true; 7027 if (_parallel) { 7028 p->oop_iterate(_par_scan_closure, mr); 7029 } else { 7030 p->oop_iterate(_scan_closure, mr); 7031 } 7032 } else { 7033 if (_parallel) { 7034 p->oop_iterate(_par_scan_closure); 7035 } else { 7036 p->oop_iterate(_scan_closure); 7037 } 7038 } 7039 } 7040 #ifdef DEBUG 7041 if (!_parallel) { 7042 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 7043 assert(_collector->overflow_list_is_empty(), 7044 "overflow list should be empty"); 7045 7046 } 7047 #endif // DEBUG 7048 return is_obj_array; 7049 } 7050 7051 MarkFromRootsClosure::MarkFromRootsClosure(CMSCollector* collector, 7052 MemRegion span, 7053 CMSBitMap* bitMap, CMSMarkStack* markStack, 7054 bool should_yield, bool verifying): 7055 _collector(collector), 7056 _span(span), 7057 _bitMap(bitMap), 7058 _mut(&collector->_modUnionTable), 7059 _markStack(markStack), 7060 _yield(should_yield), 7061 _skipBits(0) 7062 { 7063 assert(_markStack->isEmpty(), "stack should be empty"); 7064 _finger = _bitMap->startWord(); 7065 _threshold = _finger; 7066 assert(_collector->_restart_addr == NULL, "Sanity check"); 7067 assert(_span.contains(_finger), "Out of bounds _finger?"); 7068 DEBUG_ONLY(_verifying = verifying;) 7069 } 7070 7071 void MarkFromRootsClosure::reset(HeapWord* addr) { 7072 assert(_markStack->isEmpty(), "would cause duplicates on stack"); 7073 assert(_span.contains(addr), "Out of bounds _finger?"); 7074 _finger = addr; 7075 _threshold = (HeapWord*)round_to( 7076 (intptr_t)_finger, CardTableModRefBS::card_size); 7077 } 7078 7079 // Should revisit to see if this should be restructured for 7080 // greater efficiency. 7081 bool MarkFromRootsClosure::do_bit(size_t offset) { 7082 if (_skipBits > 0) { 7083 _skipBits--; 7084 return true; 7085 } 7086 // convert offset into a HeapWord* 7087 HeapWord* addr = _bitMap->startWord() + offset; 7088 assert(_bitMap->endWord() && addr < _bitMap->endWord(), 7089 "address out of range"); 7090 assert(_bitMap->isMarked(addr), "tautology"); 7091 if (_bitMap->isMarked(addr+1)) { 7092 // this is an allocated but not yet initialized object 7093 assert(_skipBits == 0, "tautology"); 7094 _skipBits = 2; // skip next two marked bits ("Printezis-marks") 7095 oop p = oop(addr); 7096 if (p->klass_or_null() == NULL) { 7097 DEBUG_ONLY(if (!_verifying) {) 7098 // We re-dirty the cards on which this object lies and increase 7099 // the _threshold so that we'll come back to scan this object 7100 // during the preclean or remark phase. (CMSCleanOnEnter) 7101 if (CMSCleanOnEnter) { 7102 size_t sz = _collector->block_size_using_printezis_bits(addr); 7103 HeapWord* end_card_addr = (HeapWord*)round_to( 7104 (intptr_t)(addr+sz), CardTableModRefBS::card_size); 7105 MemRegion redirty_range = MemRegion(addr, end_card_addr); 7106 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 7107 // Bump _threshold to end_card_addr; note that 7108 // _threshold cannot possibly exceed end_card_addr, anyhow. 7109 // This prevents future clearing of the card as the scan proceeds 7110 // to the right. 7111 assert(_threshold <= end_card_addr, 7112 "Because we are just scanning into this object"); 7113 if (_threshold < end_card_addr) { 7114 _threshold = end_card_addr; 7115 } 7116 if (p->klass_or_null() != NULL) { 7117 // Redirty the range of cards... 7118 _mut->mark_range(redirty_range); 7119 } // ...else the setting of klass will dirty the card anyway. 7120 } 7121 DEBUG_ONLY(}) 7122 return true; 7123 } 7124 } 7125 scanOopsInOop(addr); 7126 return true; 7127 } 7128 7129 // We take a break if we've been at this for a while, 7130 // so as to avoid monopolizing the locks involved. 7131 void MarkFromRootsClosure::do_yield_work() { 7132 // First give up the locks, then yield, then re-lock 7133 // We should probably use a constructor/destructor idiom to 7134 // do this unlock/lock or modify the MutexUnlocker class to 7135 // serve our purpose. XXX 7136 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 7137 "CMS thread should hold CMS token"); 7138 assert_lock_strong(_bitMap->lock()); 7139 _bitMap->lock()->unlock(); 7140 ConcurrentMarkSweepThread::desynchronize(true); 7141 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7142 _collector->stopTimer(); 7143 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 7144 if (PrintCMSStatistics != 0) { 7145 _collector->incrementYields(); 7146 } 7147 _collector->icms_wait(); 7148 7149 // See the comment in coordinator_yield() 7150 for (unsigned i = 0; i < CMSYieldSleepCount && 7151 ConcurrentMarkSweepThread::should_yield() && 7152 !CMSCollector::foregroundGCIsActive(); ++i) { 7153 os::sleep(Thread::current(), 1, false); 7154 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7155 } 7156 7157 ConcurrentMarkSweepThread::synchronize(true); 7158 _bitMap->lock()->lock_without_safepoint_check(); 7159 _collector->startTimer(); 7160 } 7161 7162 void MarkFromRootsClosure::scanOopsInOop(HeapWord* ptr) { 7163 assert(_bitMap->isMarked(ptr), "expected bit to be set"); 7164 assert(_markStack->isEmpty(), 7165 "should drain stack to limit stack usage"); 7166 // convert ptr to an oop preparatory to scanning 7167 oop obj = oop(ptr); 7168 // Ignore mark word in verification below, since we 7169 // may be running concurrent with mutators. 7170 assert(obj->is_oop(true), "should be an oop"); 7171 assert(_finger <= ptr, "_finger runneth ahead"); 7172 // advance the finger to right end of this object 7173 _finger = ptr + obj->size(); 7174 assert(_finger > ptr, "we just incremented it above"); 7175 // On large heaps, it may take us some time to get through 7176 // the marking phase (especially if running iCMS). During 7177 // this time it's possible that a lot of mutations have 7178 // accumulated in the card table and the mod union table -- 7179 // these mutation records are redundant until we have 7180 // actually traced into the corresponding card. 7181 // Here, we check whether advancing the finger would make 7182 // us cross into a new card, and if so clear corresponding 7183 // cards in the MUT (preclean them in the card-table in the 7184 // future). 7185 7186 DEBUG_ONLY(if (!_verifying) {) 7187 // The clean-on-enter optimization is disabled by default, 7188 // until we fix 6178663. 7189 if (CMSCleanOnEnter && (_finger > _threshold)) { 7190 // [_threshold, _finger) represents the interval 7191 // of cards to be cleared in MUT (or precleaned in card table). 7192 // The set of cards to be cleared is all those that overlap 7193 // with the interval [_threshold, _finger); note that 7194 // _threshold is always kept card-aligned but _finger isn't 7195 // always card-aligned. 7196 HeapWord* old_threshold = _threshold; 7197 assert(old_threshold == (HeapWord*)round_to( 7198 (intptr_t)old_threshold, CardTableModRefBS::card_size), 7199 "_threshold should always be card-aligned"); 7200 _threshold = (HeapWord*)round_to( 7201 (intptr_t)_finger, CardTableModRefBS::card_size); 7202 MemRegion mr(old_threshold, _threshold); 7203 assert(!mr.is_empty(), "Control point invariant"); 7204 assert(_span.contains(mr), "Should clear within span"); 7205 _mut->clear_range(mr); 7206 } 7207 DEBUG_ONLY(}) 7208 // Note: the finger doesn't advance while we drain 7209 // the stack below. 7210 PushOrMarkClosure pushOrMarkClosure(_collector, 7211 _span, _bitMap, _markStack, 7212 _finger, this); 7213 bool res = _markStack->push(obj); 7214 assert(res, "Empty non-zero size stack should have space for single push"); 7215 while (!_markStack->isEmpty()) { 7216 oop new_oop = _markStack->pop(); 7217 // Skip verifying header mark word below because we are 7218 // running concurrent with mutators. 7219 assert(new_oop->is_oop(true), "Oops! expected to pop an oop"); 7220 // now scan this oop's oops 7221 new_oop->oop_iterate(&pushOrMarkClosure); 7222 do_yield_check(); 7223 } 7224 assert(_markStack->isEmpty(), "tautology, emphasizing post-condition"); 7225 } 7226 7227 Par_MarkFromRootsClosure::Par_MarkFromRootsClosure(CMSConcMarkingTask* task, 7228 CMSCollector* collector, MemRegion span, 7229 CMSBitMap* bit_map, 7230 OopTaskQueue* work_queue, 7231 CMSMarkStack* overflow_stack, 7232 bool should_yield): 7233 _collector(collector), 7234 _whole_span(collector->_span), 7235 _span(span), 7236 _bit_map(bit_map), 7237 _mut(&collector->_modUnionTable), 7238 _work_queue(work_queue), 7239 _overflow_stack(overflow_stack), 7240 _yield(should_yield), 7241 _skip_bits(0), 7242 _task(task) 7243 { 7244 assert(_work_queue->size() == 0, "work_queue should be empty"); 7245 _finger = span.start(); 7246 _threshold = _finger; // XXX Defer clear-on-enter optimization for now 7247 assert(_span.contains(_finger), "Out of bounds _finger?"); 7248 } 7249 7250 // Should revisit to see if this should be restructured for 7251 // greater efficiency. 7252 bool Par_MarkFromRootsClosure::do_bit(size_t offset) { 7253 if (_skip_bits > 0) { 7254 _skip_bits--; 7255 return true; 7256 } 7257 // convert offset into a HeapWord* 7258 HeapWord* addr = _bit_map->startWord() + offset; 7259 assert(_bit_map->endWord() && addr < _bit_map->endWord(), 7260 "address out of range"); 7261 assert(_bit_map->isMarked(addr), "tautology"); 7262 if (_bit_map->isMarked(addr+1)) { 7263 // this is an allocated object that might not yet be initialized 7264 assert(_skip_bits == 0, "tautology"); 7265 _skip_bits = 2; // skip next two marked bits ("Printezis-marks") 7266 oop p = oop(addr); 7267 if (p->klass_or_null() == NULL) { 7268 // in the case of Clean-on-Enter optimization, redirty card 7269 // and avoid clearing card by increasing the threshold. 7270 return true; 7271 } 7272 } 7273 scan_oops_in_oop(addr); 7274 return true; 7275 } 7276 7277 void Par_MarkFromRootsClosure::scan_oops_in_oop(HeapWord* ptr) { 7278 assert(_bit_map->isMarked(ptr), "expected bit to be set"); 7279 // Should we assert that our work queue is empty or 7280 // below some drain limit? 7281 assert(_work_queue->size() == 0, 7282 "should drain stack to limit stack usage"); 7283 // convert ptr to an oop preparatory to scanning 7284 oop obj = oop(ptr); 7285 // Ignore mark word in verification below, since we 7286 // may be running concurrent with mutators. 7287 assert(obj->is_oop(true), "should be an oop"); 7288 assert(_finger <= ptr, "_finger runneth ahead"); 7289 // advance the finger to right end of this object 7290 _finger = ptr + obj->size(); 7291 assert(_finger > ptr, "we just incremented it above"); 7292 // On large heaps, it may take us some time to get through 7293 // the marking phase (especially if running iCMS). During 7294 // this time it's possible that a lot of mutations have 7295 // accumulated in the card table and the mod union table -- 7296 // these mutation records are redundant until we have 7297 // actually traced into the corresponding card. 7298 // Here, we check whether advancing the finger would make 7299 // us cross into a new card, and if so clear corresponding 7300 // cards in the MUT (preclean them in the card-table in the 7301 // future). 7302 7303 // The clean-on-enter optimization is disabled by default, 7304 // until we fix 6178663. 7305 if (CMSCleanOnEnter && (_finger > _threshold)) { 7306 // [_threshold, _finger) represents the interval 7307 // of cards to be cleared in MUT (or precleaned in card table). 7308 // The set of cards to be cleared is all those that overlap 7309 // with the interval [_threshold, _finger); note that 7310 // _threshold is always kept card-aligned but _finger isn't 7311 // always card-aligned. 7312 HeapWord* old_threshold = _threshold; 7313 assert(old_threshold == (HeapWord*)round_to( 7314 (intptr_t)old_threshold, CardTableModRefBS::card_size), 7315 "_threshold should always be card-aligned"); 7316 _threshold = (HeapWord*)round_to( 7317 (intptr_t)_finger, CardTableModRefBS::card_size); 7318 MemRegion mr(old_threshold, _threshold); 7319 assert(!mr.is_empty(), "Control point invariant"); 7320 assert(_span.contains(mr), "Should clear within span"); // _whole_span ?? 7321 _mut->clear_range(mr); 7322 } 7323 7324 // Note: the local finger doesn't advance while we drain 7325 // the stack below, but the global finger sure can and will. 7326 HeapWord** gfa = _task->global_finger_addr(); 7327 Par_PushOrMarkClosure pushOrMarkClosure(_collector, 7328 _span, _bit_map, 7329 _work_queue, 7330 _overflow_stack, 7331 _finger, 7332 gfa, this); 7333 bool res = _work_queue->push(obj); // overflow could occur here 7334 assert(res, "Will hold once we use workqueues"); 7335 while (true) { 7336 oop new_oop; 7337 if (!_work_queue->pop_local(new_oop)) { 7338 // We emptied our work_queue; check if there's stuff that can 7339 // be gotten from the overflow stack. 7340 if (CMSConcMarkingTask::get_work_from_overflow_stack( 7341 _overflow_stack, _work_queue)) { 7342 do_yield_check(); 7343 continue; 7344 } else { // done 7345 break; 7346 } 7347 } 7348 // Skip verifying header mark word below because we are 7349 // running concurrent with mutators. 7350 assert(new_oop->is_oop(true), "Oops! expected to pop an oop"); 7351 // now scan this oop's oops 7352 new_oop->oop_iterate(&pushOrMarkClosure); 7353 do_yield_check(); 7354 } 7355 assert(_work_queue->size() == 0, "tautology, emphasizing post-condition"); 7356 } 7357 7358 // Yield in response to a request from VM Thread or 7359 // from mutators. 7360 void Par_MarkFromRootsClosure::do_yield_work() { 7361 assert(_task != NULL, "sanity"); 7362 _task->yield(); 7363 } 7364 7365 // A variant of the above used for verifying CMS marking work. 7366 MarkFromRootsVerifyClosure::MarkFromRootsVerifyClosure(CMSCollector* collector, 7367 MemRegion span, 7368 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 7369 CMSMarkStack* mark_stack): 7370 _collector(collector), 7371 _span(span), 7372 _verification_bm(verification_bm), 7373 _cms_bm(cms_bm), 7374 _mark_stack(mark_stack), 7375 _pam_verify_closure(collector, span, verification_bm, cms_bm, 7376 mark_stack) 7377 { 7378 assert(_mark_stack->isEmpty(), "stack should be empty"); 7379 _finger = _verification_bm->startWord(); 7380 assert(_collector->_restart_addr == NULL, "Sanity check"); 7381 assert(_span.contains(_finger), "Out of bounds _finger?"); 7382 } 7383 7384 void MarkFromRootsVerifyClosure::reset(HeapWord* addr) { 7385 assert(_mark_stack->isEmpty(), "would cause duplicates on stack"); 7386 assert(_span.contains(addr), "Out of bounds _finger?"); 7387 _finger = addr; 7388 } 7389 7390 // Should revisit to see if this should be restructured for 7391 // greater efficiency. 7392 bool MarkFromRootsVerifyClosure::do_bit(size_t offset) { 7393 // convert offset into a HeapWord* 7394 HeapWord* addr = _verification_bm->startWord() + offset; 7395 assert(_verification_bm->endWord() && addr < _verification_bm->endWord(), 7396 "address out of range"); 7397 assert(_verification_bm->isMarked(addr), "tautology"); 7398 assert(_cms_bm->isMarked(addr), "tautology"); 7399 7400 assert(_mark_stack->isEmpty(), 7401 "should drain stack to limit stack usage"); 7402 // convert addr to an oop preparatory to scanning 7403 oop obj = oop(addr); 7404 assert(obj->is_oop(), "should be an oop"); 7405 assert(_finger <= addr, "_finger runneth ahead"); 7406 // advance the finger to right end of this object 7407 _finger = addr + obj->size(); 7408 assert(_finger > addr, "we just incremented it above"); 7409 // Note: the finger doesn't advance while we drain 7410 // the stack below. 7411 bool res = _mark_stack->push(obj); 7412 assert(res, "Empty non-zero size stack should have space for single push"); 7413 while (!_mark_stack->isEmpty()) { 7414 oop new_oop = _mark_stack->pop(); 7415 assert(new_oop->is_oop(), "Oops! expected to pop an oop"); 7416 // now scan this oop's oops 7417 new_oop->oop_iterate(&_pam_verify_closure); 7418 } 7419 assert(_mark_stack->isEmpty(), "tautology, emphasizing post-condition"); 7420 return true; 7421 } 7422 7423 PushAndMarkVerifyClosure::PushAndMarkVerifyClosure( 7424 CMSCollector* collector, MemRegion span, 7425 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 7426 CMSMarkStack* mark_stack): 7427 CMSOopClosure(collector->ref_processor()), 7428 _collector(collector), 7429 _span(span), 7430 _verification_bm(verification_bm), 7431 _cms_bm(cms_bm), 7432 _mark_stack(mark_stack) 7433 { } 7434 7435 void PushAndMarkVerifyClosure::do_oop(oop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 7436 void PushAndMarkVerifyClosure::do_oop(narrowOop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 7437 7438 // Upon stack overflow, we discard (part of) the stack, 7439 // remembering the least address amongst those discarded 7440 // in CMSCollector's _restart_address. 7441 void PushAndMarkVerifyClosure::handle_stack_overflow(HeapWord* lost) { 7442 // Remember the least grey address discarded 7443 HeapWord* ra = (HeapWord*)_mark_stack->least_value(lost); 7444 _collector->lower_restart_addr(ra); 7445 _mark_stack->reset(); // discard stack contents 7446 _mark_stack->expand(); // expand the stack if possible 7447 } 7448 7449 void PushAndMarkVerifyClosure::do_oop(oop obj) { 7450 assert(obj->is_oop_or_null(), "expected an oop or NULL"); 7451 HeapWord* addr = (HeapWord*)obj; 7452 if (_span.contains(addr) && !_verification_bm->isMarked(addr)) { 7453 // Oop lies in _span and isn't yet grey or black 7454 _verification_bm->mark(addr); // now grey 7455 if (!_cms_bm->isMarked(addr)) { 7456 oop(addr)->print(); 7457 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)", 7458 addr); 7459 fatal("... aborting"); 7460 } 7461 7462 if (!_mark_stack->push(obj)) { // stack overflow 7463 if (PrintCMSStatistics != 0) { 7464 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 7465 SIZE_FORMAT, _mark_stack->capacity()); 7466 } 7467 assert(_mark_stack->isFull(), "Else push should have succeeded"); 7468 handle_stack_overflow(addr); 7469 } 7470 // anything including and to the right of _finger 7471 // will be scanned as we iterate over the remainder of the 7472 // bit map 7473 } 7474 } 7475 7476 PushOrMarkClosure::PushOrMarkClosure(CMSCollector* collector, 7477 MemRegion span, 7478 CMSBitMap* bitMap, CMSMarkStack* markStack, 7479 HeapWord* finger, MarkFromRootsClosure* parent) : 7480 CMSOopClosure(collector->ref_processor()), 7481 _collector(collector), 7482 _span(span), 7483 _bitMap(bitMap), 7484 _markStack(markStack), 7485 _finger(finger), 7486 _parent(parent) 7487 { } 7488 7489 Par_PushOrMarkClosure::Par_PushOrMarkClosure(CMSCollector* collector, 7490 MemRegion span, 7491 CMSBitMap* bit_map, 7492 OopTaskQueue* work_queue, 7493 CMSMarkStack* overflow_stack, 7494 HeapWord* finger, 7495 HeapWord** global_finger_addr, 7496 Par_MarkFromRootsClosure* parent) : 7497 CMSOopClosure(collector->ref_processor()), 7498 _collector(collector), 7499 _whole_span(collector->_span), 7500 _span(span), 7501 _bit_map(bit_map), 7502 _work_queue(work_queue), 7503 _overflow_stack(overflow_stack), 7504 _finger(finger), 7505 _global_finger_addr(global_finger_addr), 7506 _parent(parent) 7507 { } 7508 7509 // Assumes thread-safe access by callers, who are 7510 // responsible for mutual exclusion. 7511 void CMSCollector::lower_restart_addr(HeapWord* low) { 7512 assert(_span.contains(low), "Out of bounds addr"); 7513 if (_restart_addr == NULL) { 7514 _restart_addr = low; 7515 } else { 7516 _restart_addr = MIN2(_restart_addr, low); 7517 } 7518 } 7519 7520 // Upon stack overflow, we discard (part of) the stack, 7521 // remembering the least address amongst those discarded 7522 // in CMSCollector's _restart_address. 7523 void PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 7524 // Remember the least grey address discarded 7525 HeapWord* ra = (HeapWord*)_markStack->least_value(lost); 7526 _collector->lower_restart_addr(ra); 7527 _markStack->reset(); // discard stack contents 7528 _markStack->expand(); // expand the stack if possible 7529 } 7530 7531 // Upon stack overflow, we discard (part of) the stack, 7532 // remembering the least address amongst those discarded 7533 // in CMSCollector's _restart_address. 7534 void Par_PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 7535 // We need to do this under a mutex to prevent other 7536 // workers from interfering with the work done below. 7537 MutexLockerEx ml(_overflow_stack->par_lock(), 7538 Mutex::_no_safepoint_check_flag); 7539 // Remember the least grey address discarded 7540 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 7541 _collector->lower_restart_addr(ra); 7542 _overflow_stack->reset(); // discard stack contents 7543 _overflow_stack->expand(); // expand the stack if possible 7544 } 7545 7546 void CMKlassClosure::do_klass(Klass* k) { 7547 assert(_oop_closure != NULL, "Not initialized?"); 7548 k->oops_do(_oop_closure); 7549 } 7550 7551 void PushOrMarkClosure::do_oop(oop obj) { 7552 // Ignore mark word because we are running concurrent with mutators. 7553 assert(obj->is_oop_or_null(true), "expected an oop or NULL"); 7554 HeapWord* addr = (HeapWord*)obj; 7555 if (_span.contains(addr) && !_bitMap->isMarked(addr)) { 7556 // Oop lies in _span and isn't yet grey or black 7557 _bitMap->mark(addr); // now grey 7558 if (addr < _finger) { 7559 // the bit map iteration has already either passed, or 7560 // sampled, this bit in the bit map; we'll need to 7561 // use the marking stack to scan this oop's oops. 7562 bool simulate_overflow = false; 7563 NOT_PRODUCT( 7564 if (CMSMarkStackOverflowALot && 7565 _collector->simulate_overflow()) { 7566 // simulate a stack overflow 7567 simulate_overflow = true; 7568 } 7569 ) 7570 if (simulate_overflow || !_markStack->push(obj)) { // stack overflow 7571 if (PrintCMSStatistics != 0) { 7572 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 7573 SIZE_FORMAT, _markStack->capacity()); 7574 } 7575 assert(simulate_overflow || _markStack->isFull(), "Else push should have succeeded"); 7576 handle_stack_overflow(addr); 7577 } 7578 } 7579 // anything including and to the right of _finger 7580 // will be scanned as we iterate over the remainder of the 7581 // bit map 7582 do_yield_check(); 7583 } 7584 } 7585 7586 void PushOrMarkClosure::do_oop(oop* p) { PushOrMarkClosure::do_oop_work(p); } 7587 void PushOrMarkClosure::do_oop(narrowOop* p) { PushOrMarkClosure::do_oop_work(p); } 7588 7589 void Par_PushOrMarkClosure::do_oop(oop obj) { 7590 // Ignore mark word because we are running concurrent with mutators. 7591 assert(obj->is_oop_or_null(true), "expected an oop or NULL"); 7592 HeapWord* addr = (HeapWord*)obj; 7593 if (_whole_span.contains(addr) && !_bit_map->isMarked(addr)) { 7594 // Oop lies in _span and isn't yet grey or black 7595 // We read the global_finger (volatile read) strictly after marking oop 7596 bool res = _bit_map->par_mark(addr); // now grey 7597 volatile HeapWord** gfa = (volatile HeapWord**)_global_finger_addr; 7598 // Should we push this marked oop on our stack? 7599 // -- if someone else marked it, nothing to do 7600 // -- if target oop is above global finger nothing to do 7601 // -- if target oop is in chunk and above local finger 7602 // then nothing to do 7603 // -- else push on work queue 7604 if ( !res // someone else marked it, they will deal with it 7605 || (addr >= *gfa) // will be scanned in a later task 7606 || (_span.contains(addr) && addr >= _finger)) { // later in this chunk 7607 return; 7608 } 7609 // the bit map iteration has already either passed, or 7610 // sampled, this bit in the bit map; we'll need to 7611 // use the marking stack to scan this oop's oops. 7612 bool simulate_overflow = false; 7613 NOT_PRODUCT( 7614 if (CMSMarkStackOverflowALot && 7615 _collector->simulate_overflow()) { 7616 // simulate a stack overflow 7617 simulate_overflow = true; 7618 } 7619 ) 7620 if (simulate_overflow || 7621 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) { 7622 // stack overflow 7623 if (PrintCMSStatistics != 0) { 7624 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 7625 SIZE_FORMAT, _overflow_stack->capacity()); 7626 } 7627 // We cannot assert that the overflow stack is full because 7628 // it may have been emptied since. 7629 assert(simulate_overflow || 7630 _work_queue->size() == _work_queue->max_elems(), 7631 "Else push should have succeeded"); 7632 handle_stack_overflow(addr); 7633 } 7634 do_yield_check(); 7635 } 7636 } 7637 7638 void Par_PushOrMarkClosure::do_oop(oop* p) { Par_PushOrMarkClosure::do_oop_work(p); } 7639 void Par_PushOrMarkClosure::do_oop(narrowOop* p) { Par_PushOrMarkClosure::do_oop_work(p); } 7640 7641 PushAndMarkClosure::PushAndMarkClosure(CMSCollector* collector, 7642 MemRegion span, 7643 ReferenceProcessor* rp, 7644 CMSBitMap* bit_map, 7645 CMSBitMap* mod_union_table, 7646 CMSMarkStack* mark_stack, 7647 bool concurrent_precleaning): 7648 CMSOopClosure(rp), 7649 _collector(collector), 7650 _span(span), 7651 _bit_map(bit_map), 7652 _mod_union_table(mod_union_table), 7653 _mark_stack(mark_stack), 7654 _concurrent_precleaning(concurrent_precleaning) 7655 { 7656 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 7657 } 7658 7659 // Grey object rescan during pre-cleaning and second checkpoint phases -- 7660 // the non-parallel version (the parallel version appears further below.) 7661 void PushAndMarkClosure::do_oop(oop obj) { 7662 // Ignore mark word verification. If during concurrent precleaning, 7663 // the object monitor may be locked. If during the checkpoint 7664 // phases, the object may already have been reached by a different 7665 // path and may be at the end of the global overflow list (so 7666 // the mark word may be NULL). 7667 assert(obj->is_oop_or_null(true /* ignore mark word */), 7668 "expected an oop or NULL"); 7669 HeapWord* addr = (HeapWord*)obj; 7670 // Check if oop points into the CMS generation 7671 // and is not marked 7672 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 7673 // a white object ... 7674 _bit_map->mark(addr); // ... now grey 7675 // push on the marking stack (grey set) 7676 bool simulate_overflow = false; 7677 NOT_PRODUCT( 7678 if (CMSMarkStackOverflowALot && 7679 _collector->simulate_overflow()) { 7680 // simulate a stack overflow 7681 simulate_overflow = true; 7682 } 7683 ) 7684 if (simulate_overflow || !_mark_stack->push(obj)) { 7685 if (_concurrent_precleaning) { 7686 // During precleaning we can just dirty the appropriate card(s) 7687 // in the mod union table, thus ensuring that the object remains 7688 // in the grey set and continue. In the case of object arrays 7689 // we need to dirty all of the cards that the object spans, 7690 // since the rescan of object arrays will be limited to the 7691 // dirty cards. 7692 // Note that no one can be intefering with us in this action 7693 // of dirtying the mod union table, so no locking or atomics 7694 // are required. 7695 if (obj->is_objArray()) { 7696 size_t sz = obj->size(); 7697 HeapWord* end_card_addr = (HeapWord*)round_to( 7698 (intptr_t)(addr+sz), CardTableModRefBS::card_size); 7699 MemRegion redirty_range = MemRegion(addr, end_card_addr); 7700 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 7701 _mod_union_table->mark_range(redirty_range); 7702 } else { 7703 _mod_union_table->mark(addr); 7704 } 7705 _collector->_ser_pmc_preclean_ovflw++; 7706 } else { 7707 // During the remark phase, we need to remember this oop 7708 // in the overflow list. 7709 _collector->push_on_overflow_list(obj); 7710 _collector->_ser_pmc_remark_ovflw++; 7711 } 7712 } 7713 } 7714 } 7715 7716 Par_PushAndMarkClosure::Par_PushAndMarkClosure(CMSCollector* collector, 7717 MemRegion span, 7718 ReferenceProcessor* rp, 7719 CMSBitMap* bit_map, 7720 OopTaskQueue* work_queue): 7721 CMSOopClosure(rp), 7722 _collector(collector), 7723 _span(span), 7724 _bit_map(bit_map), 7725 _work_queue(work_queue) 7726 { 7727 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 7728 } 7729 7730 void PushAndMarkClosure::do_oop(oop* p) { PushAndMarkClosure::do_oop_work(p); } 7731 void PushAndMarkClosure::do_oop(narrowOop* p) { PushAndMarkClosure::do_oop_work(p); } 7732 7733 // Grey object rescan during second checkpoint phase -- 7734 // the parallel version. 7735 void Par_PushAndMarkClosure::do_oop(oop obj) { 7736 // In the assert below, we ignore the mark word because 7737 // this oop may point to an already visited object that is 7738 // on the overflow stack (in which case the mark word has 7739 // been hijacked for chaining into the overflow stack -- 7740 // if this is the last object in the overflow stack then 7741 // its mark word will be NULL). Because this object may 7742 // have been subsequently popped off the global overflow 7743 // stack, and the mark word possibly restored to the prototypical 7744 // value, by the time we get to examined this failing assert in 7745 // the debugger, is_oop_or_null(false) may subsequently start 7746 // to hold. 7747 assert(obj->is_oop_or_null(true), 7748 "expected an oop or NULL"); 7749 HeapWord* addr = (HeapWord*)obj; 7750 // Check if oop points into the CMS generation 7751 // and is not marked 7752 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 7753 // a white object ... 7754 // If we manage to "claim" the object, by being the 7755 // first thread to mark it, then we push it on our 7756 // marking stack 7757 if (_bit_map->par_mark(addr)) { // ... now grey 7758 // push on work queue (grey set) 7759 bool simulate_overflow = false; 7760 NOT_PRODUCT( 7761 if (CMSMarkStackOverflowALot && 7762 _collector->par_simulate_overflow()) { 7763 // simulate a stack overflow 7764 simulate_overflow = true; 7765 } 7766 ) 7767 if (simulate_overflow || !_work_queue->push(obj)) { 7768 _collector->par_push_on_overflow_list(obj); 7769 _collector->_par_pmc_remark_ovflw++; // imprecise OK: no need to CAS 7770 } 7771 } // Else, some other thread got there first 7772 } 7773 } 7774 7775 void Par_PushAndMarkClosure::do_oop(oop* p) { Par_PushAndMarkClosure::do_oop_work(p); } 7776 void Par_PushAndMarkClosure::do_oop(narrowOop* p) { Par_PushAndMarkClosure::do_oop_work(p); } 7777 7778 void CMSPrecleanRefsYieldClosure::do_yield_work() { 7779 Mutex* bml = _collector->bitMapLock(); 7780 assert_lock_strong(bml); 7781 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 7782 "CMS thread should hold CMS token"); 7783 7784 bml->unlock(); 7785 ConcurrentMarkSweepThread::desynchronize(true); 7786 7787 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7788 7789 _collector->stopTimer(); 7790 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 7791 if (PrintCMSStatistics != 0) { 7792 _collector->incrementYields(); 7793 } 7794 _collector->icms_wait(); 7795 7796 // See the comment in coordinator_yield() 7797 for (unsigned i = 0; i < CMSYieldSleepCount && 7798 ConcurrentMarkSweepThread::should_yield() && 7799 !CMSCollector::foregroundGCIsActive(); ++i) { 7800 os::sleep(Thread::current(), 1, false); 7801 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7802 } 7803 7804 ConcurrentMarkSweepThread::synchronize(true); 7805 bml->lock(); 7806 7807 _collector->startTimer(); 7808 } 7809 7810 bool CMSPrecleanRefsYieldClosure::should_return() { 7811 if (ConcurrentMarkSweepThread::should_yield()) { 7812 do_yield_work(); 7813 } 7814 return _collector->foregroundGCIsActive(); 7815 } 7816 7817 void MarkFromDirtyCardsClosure::do_MemRegion(MemRegion mr) { 7818 assert(((size_t)mr.start())%CardTableModRefBS::card_size_in_words == 0, 7819 "mr should be aligned to start at a card boundary"); 7820 // We'd like to assert: 7821 // assert(mr.word_size()%CardTableModRefBS::card_size_in_words == 0, 7822 // "mr should be a range of cards"); 7823 // However, that would be too strong in one case -- the last 7824 // partition ends at _unallocated_block which, in general, can be 7825 // an arbitrary boundary, not necessarily card aligned. 7826 if (PrintCMSStatistics != 0) { 7827 _num_dirty_cards += 7828 mr.word_size()/CardTableModRefBS::card_size_in_words; 7829 } 7830 _space->object_iterate_mem(mr, &_scan_cl); 7831 } 7832 7833 SweepClosure::SweepClosure(CMSCollector* collector, 7834 ConcurrentMarkSweepGeneration* g, 7835 CMSBitMap* bitMap, bool should_yield) : 7836 _collector(collector), 7837 _g(g), 7838 _sp(g->cmsSpace()), 7839 _limit(_sp->sweep_limit()), 7840 _freelistLock(_sp->freelistLock()), 7841 _bitMap(bitMap), 7842 _yield(should_yield), 7843 _inFreeRange(false), // No free range at beginning of sweep 7844 _freeRangeInFreeLists(false), // No free range at beginning of sweep 7845 _lastFreeRangeCoalesced(false), 7846 _freeFinger(g->used_region().start()) 7847 { 7848 NOT_PRODUCT( 7849 _numObjectsFreed = 0; 7850 _numWordsFreed = 0; 7851 _numObjectsLive = 0; 7852 _numWordsLive = 0; 7853 _numObjectsAlreadyFree = 0; 7854 _numWordsAlreadyFree = 0; 7855 _last_fc = NULL; 7856 7857 _sp->initializeIndexedFreeListArrayReturnedBytes(); 7858 _sp->dictionary()->initialize_dict_returned_bytes(); 7859 ) 7860 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7861 "sweep _limit out of bounds"); 7862 if (CMSTraceSweeper) { 7863 gclog_or_tty->print_cr("\n====================\nStarting new sweep with limit " PTR_FORMAT, 7864 _limit); 7865 } 7866 } 7867 7868 void SweepClosure::print_on(outputStream* st) const { 7869 tty->print_cr("_sp = [" PTR_FORMAT "," PTR_FORMAT ")", 7870 _sp->bottom(), _sp->end()); 7871 tty->print_cr("_limit = " PTR_FORMAT, _limit); 7872 tty->print_cr("_freeFinger = " PTR_FORMAT, _freeFinger); 7873 NOT_PRODUCT(tty->print_cr("_last_fc = " PTR_FORMAT, _last_fc);) 7874 tty->print_cr("_inFreeRange = %d, _freeRangeInFreeLists = %d, _lastFreeRangeCoalesced = %d", 7875 _inFreeRange, _freeRangeInFreeLists, _lastFreeRangeCoalesced); 7876 } 7877 7878 #ifndef PRODUCT 7879 // Assertion checking only: no useful work in product mode -- 7880 // however, if any of the flags below become product flags, 7881 // you may need to review this code to see if it needs to be 7882 // enabled in product mode. 7883 SweepClosure::~SweepClosure() { 7884 assert_lock_strong(_freelistLock); 7885 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7886 "sweep _limit out of bounds"); 7887 if (inFreeRange()) { 7888 warning("inFreeRange() should have been reset; dumping state of SweepClosure"); 7889 print(); 7890 ShouldNotReachHere(); 7891 } 7892 if (Verbose && PrintGC) { 7893 gclog_or_tty->print("Collected "SIZE_FORMAT" objects, " SIZE_FORMAT " bytes", 7894 _numObjectsFreed, _numWordsFreed*sizeof(HeapWord)); 7895 gclog_or_tty->print_cr("\nLive "SIZE_FORMAT" objects, " 7896 SIZE_FORMAT" bytes " 7897 "Already free "SIZE_FORMAT" objects, "SIZE_FORMAT" bytes", 7898 _numObjectsLive, _numWordsLive*sizeof(HeapWord), 7899 _numObjectsAlreadyFree, _numWordsAlreadyFree*sizeof(HeapWord)); 7900 size_t totalBytes = (_numWordsFreed + _numWordsLive + _numWordsAlreadyFree) 7901 * sizeof(HeapWord); 7902 gclog_or_tty->print_cr("Total sweep: "SIZE_FORMAT" bytes", totalBytes); 7903 7904 if (PrintCMSStatistics && CMSVerifyReturnedBytes) { 7905 size_t indexListReturnedBytes = _sp->sumIndexedFreeListArrayReturnedBytes(); 7906 size_t dict_returned_bytes = _sp->dictionary()->sum_dict_returned_bytes(); 7907 size_t returned_bytes = indexListReturnedBytes + dict_returned_bytes; 7908 gclog_or_tty->print("Returned "SIZE_FORMAT" bytes", returned_bytes); 7909 gclog_or_tty->print(" Indexed List Returned "SIZE_FORMAT" bytes", 7910 indexListReturnedBytes); 7911 gclog_or_tty->print_cr(" Dictionary Returned "SIZE_FORMAT" bytes", 7912 dict_returned_bytes); 7913 } 7914 } 7915 if (CMSTraceSweeper) { 7916 gclog_or_tty->print_cr("end of sweep with _limit = " PTR_FORMAT "\n================", 7917 _limit); 7918 } 7919 } 7920 #endif // PRODUCT 7921 7922 void SweepClosure::initialize_free_range(HeapWord* freeFinger, 7923 bool freeRangeInFreeLists) { 7924 if (CMSTraceSweeper) { 7925 gclog_or_tty->print("---- Start free range at 0x%x with free block (%d)\n", 7926 freeFinger, freeRangeInFreeLists); 7927 } 7928 assert(!inFreeRange(), "Trampling existing free range"); 7929 set_inFreeRange(true); 7930 set_lastFreeRangeCoalesced(false); 7931 7932 set_freeFinger(freeFinger); 7933 set_freeRangeInFreeLists(freeRangeInFreeLists); 7934 if (CMSTestInFreeList) { 7935 if (freeRangeInFreeLists) { 7936 FreeChunk* fc = (FreeChunk*) freeFinger; 7937 assert(fc->is_free(), "A chunk on the free list should be free."); 7938 assert(fc->size() > 0, "Free range should have a size"); 7939 assert(_sp->verify_chunk_in_free_list(fc), "Chunk is not in free lists"); 7940 } 7941 } 7942 } 7943 7944 // Note that the sweeper runs concurrently with mutators. Thus, 7945 // it is possible for direct allocation in this generation to happen 7946 // in the middle of the sweep. Note that the sweeper also coalesces 7947 // contiguous free blocks. Thus, unless the sweeper and the allocator 7948 // synchronize appropriately freshly allocated blocks may get swept up. 7949 // This is accomplished by the sweeper locking the free lists while 7950 // it is sweeping. Thus blocks that are determined to be free are 7951 // indeed free. There is however one additional complication: 7952 // blocks that have been allocated since the final checkpoint and 7953 // mark, will not have been marked and so would be treated as 7954 // unreachable and swept up. To prevent this, the allocator marks 7955 // the bit map when allocating during the sweep phase. This leads, 7956 // however, to a further complication -- objects may have been allocated 7957 // but not yet initialized -- in the sense that the header isn't yet 7958 // installed. The sweeper can not then determine the size of the block 7959 // in order to skip over it. To deal with this case, we use a technique 7960 // (due to Printezis) to encode such uninitialized block sizes in the 7961 // bit map. Since the bit map uses a bit per every HeapWord, but the 7962 // CMS generation has a minimum object size of 3 HeapWords, it follows 7963 // that "normal marks" won't be adjacent in the bit map (there will 7964 // always be at least two 0 bits between successive 1 bits). We make use 7965 // of these "unused" bits to represent uninitialized blocks -- the bit 7966 // corresponding to the start of the uninitialized object and the next 7967 // bit are both set. Finally, a 1 bit marks the end of the object that 7968 // started with the two consecutive 1 bits to indicate its potentially 7969 // uninitialized state. 7970 7971 size_t SweepClosure::do_blk_careful(HeapWord* addr) { 7972 FreeChunk* fc = (FreeChunk*)addr; 7973 size_t res; 7974 7975 // Check if we are done sweeping. Below we check "addr >= _limit" rather 7976 // than "addr == _limit" because although _limit was a block boundary when 7977 // we started the sweep, it may no longer be one because heap expansion 7978 // may have caused us to coalesce the block ending at the address _limit 7979 // with a newly expanded chunk (this happens when _limit was set to the 7980 // previous _end of the space), so we may have stepped past _limit: 7981 // see the following Zeno-like trail of CRs 6977970, 7008136, 7042740. 7982 if (addr >= _limit) { // we have swept up to or past the limit: finish up 7983 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7984 "sweep _limit out of bounds"); 7985 assert(addr < _sp->end(), "addr out of bounds"); 7986 // Flush any free range we might be holding as a single 7987 // coalesced chunk to the appropriate free list. 7988 if (inFreeRange()) { 7989 assert(freeFinger() >= _sp->bottom() && freeFinger() < _limit, 7990 err_msg("freeFinger() " PTR_FORMAT" is out-of-bounds", freeFinger())); 7991 flush_cur_free_chunk(freeFinger(), 7992 pointer_delta(addr, freeFinger())); 7993 if (CMSTraceSweeper) { 7994 gclog_or_tty->print("Sweep: last chunk: "); 7995 gclog_or_tty->print("put_free_blk 0x%x ("SIZE_FORMAT") " 7996 "[coalesced:"SIZE_FORMAT"]\n", 7997 freeFinger(), pointer_delta(addr, freeFinger()), 7998 lastFreeRangeCoalesced()); 7999 } 8000 } 8001 8002 // help the iterator loop finish 8003 return pointer_delta(_sp->end(), addr); 8004 } 8005 8006 assert(addr < _limit, "sweep invariant"); 8007 // check if we should yield 8008 do_yield_check(addr); 8009 if (fc->is_free()) { 8010 // Chunk that is already free 8011 res = fc->size(); 8012 do_already_free_chunk(fc); 8013 debug_only(_sp->verifyFreeLists()); 8014 // If we flush the chunk at hand in lookahead_and_flush() 8015 // and it's coalesced with a preceding chunk, then the 8016 // process of "mangling" the payload of the coalesced block 8017 // will cause erasure of the size information from the 8018 // (erstwhile) header of all the coalesced blocks but the 8019 // first, so the first disjunct in the assert will not hold 8020 // in that specific case (in which case the second disjunct 8021 // will hold). 8022 assert(res == fc->size() || ((HeapWord*)fc) + res >= _limit, 8023 "Otherwise the size info doesn't change at this step"); 8024 NOT_PRODUCT( 8025 _numObjectsAlreadyFree++; 8026 _numWordsAlreadyFree += res; 8027 ) 8028 NOT_PRODUCT(_last_fc = fc;) 8029 } else if (!_bitMap->isMarked(addr)) { 8030 // Chunk is fresh garbage 8031 res = do_garbage_chunk(fc); 8032 debug_only(_sp->verifyFreeLists()); 8033 NOT_PRODUCT( 8034 _numObjectsFreed++; 8035 _numWordsFreed += res; 8036 ) 8037 } else { 8038 // Chunk that is alive. 8039 res = do_live_chunk(fc); 8040 debug_only(_sp->verifyFreeLists()); 8041 NOT_PRODUCT( 8042 _numObjectsLive++; 8043 _numWordsLive += res; 8044 ) 8045 } 8046 return res; 8047 } 8048 8049 // For the smart allocation, record following 8050 // split deaths - a free chunk is removed from its free list because 8051 // it is being split into two or more chunks. 8052 // split birth - a free chunk is being added to its free list because 8053 // a larger free chunk has been split and resulted in this free chunk. 8054 // coal death - a free chunk is being removed from its free list because 8055 // it is being coalesced into a large free chunk. 8056 // coal birth - a free chunk is being added to its free list because 8057 // it was created when two or more free chunks where coalesced into 8058 // this free chunk. 8059 // 8060 // These statistics are used to determine the desired number of free 8061 // chunks of a given size. The desired number is chosen to be relative 8062 // to the end of a CMS sweep. The desired number at the end of a sweep 8063 // is the 8064 // count-at-end-of-previous-sweep (an amount that was enough) 8065 // - count-at-beginning-of-current-sweep (the excess) 8066 // + split-births (gains in this size during interval) 8067 // - split-deaths (demands on this size during interval) 8068 // where the interval is from the end of one sweep to the end of the 8069 // next. 8070 // 8071 // When sweeping the sweeper maintains an accumulated chunk which is 8072 // the chunk that is made up of chunks that have been coalesced. That 8073 // will be termed the left-hand chunk. A new chunk of garbage that 8074 // is being considered for coalescing will be referred to as the 8075 // right-hand chunk. 8076 // 8077 // When making a decision on whether to coalesce a right-hand chunk with 8078 // the current left-hand chunk, the current count vs. the desired count 8079 // of the left-hand chunk is considered. Also if the right-hand chunk 8080 // is near the large chunk at the end of the heap (see 8081 // ConcurrentMarkSweepGeneration::isNearLargestChunk()), then the 8082 // left-hand chunk is coalesced. 8083 // 8084 // When making a decision about whether to split a chunk, the desired count 8085 // vs. the current count of the candidate to be split is also considered. 8086 // If the candidate is underpopulated (currently fewer chunks than desired) 8087 // a chunk of an overpopulated (currently more chunks than desired) size may 8088 // be chosen. The "hint" associated with a free list, if non-null, points 8089 // to a free list which may be overpopulated. 8090 // 8091 8092 void SweepClosure::do_already_free_chunk(FreeChunk* fc) { 8093 const size_t size = fc->size(); 8094 // Chunks that cannot be coalesced are not in the 8095 // free lists. 8096 if (CMSTestInFreeList && !fc->cantCoalesce()) { 8097 assert(_sp->verify_chunk_in_free_list(fc), 8098 "free chunk should be in free lists"); 8099 } 8100 // a chunk that is already free, should not have been 8101 // marked in the bit map 8102 HeapWord* const addr = (HeapWord*) fc; 8103 assert(!_bitMap->isMarked(addr), "free chunk should be unmarked"); 8104 // Verify that the bit map has no bits marked between 8105 // addr and purported end of this block. 8106 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 8107 8108 // Some chunks cannot be coalesced under any circumstances. 8109 // See the definition of cantCoalesce(). 8110 if (!fc->cantCoalesce()) { 8111 // This chunk can potentially be coalesced. 8112 if (_sp->adaptive_freelists()) { 8113 // All the work is done in 8114 do_post_free_or_garbage_chunk(fc, size); 8115 } else { // Not adaptive free lists 8116 // this is a free chunk that can potentially be coalesced by the sweeper; 8117 if (!inFreeRange()) { 8118 // if the next chunk is a free block that can't be coalesced 8119 // it doesn't make sense to remove this chunk from the free lists 8120 FreeChunk* nextChunk = (FreeChunk*)(addr + size); 8121 assert((HeapWord*)nextChunk <= _sp->end(), "Chunk size out of bounds?"); 8122 if ((HeapWord*)nextChunk < _sp->end() && // There is another free chunk to the right ... 8123 nextChunk->is_free() && // ... which is free... 8124 nextChunk->cantCoalesce()) { // ... but can't be coalesced 8125 // nothing to do 8126 } else { 8127 // Potentially the start of a new free range: 8128 // Don't eagerly remove it from the free lists. 8129 // No need to remove it if it will just be put 8130 // back again. (Also from a pragmatic point of view 8131 // if it is a free block in a region that is beyond 8132 // any allocated blocks, an assertion will fail) 8133 // Remember the start of a free run. 8134 initialize_free_range(addr, true); 8135 // end - can coalesce with next chunk 8136 } 8137 } else { 8138 // the midst of a free range, we are coalescing 8139 print_free_block_coalesced(fc); 8140 if (CMSTraceSweeper) { 8141 gclog_or_tty->print(" -- pick up free block 0x%x (%d)\n", fc, size); 8142 } 8143 // remove it from the free lists 8144 _sp->removeFreeChunkFromFreeLists(fc); 8145 set_lastFreeRangeCoalesced(true); 8146 // If the chunk is being coalesced and the current free range is 8147 // in the free lists, remove the current free range so that it 8148 // will be returned to the free lists in its entirety - all 8149 // the coalesced pieces included. 8150 if (freeRangeInFreeLists()) { 8151 FreeChunk* ffc = (FreeChunk*) freeFinger(); 8152 assert(ffc->size() == pointer_delta(addr, freeFinger()), 8153 "Size of free range is inconsistent with chunk size."); 8154 if (CMSTestInFreeList) { 8155 assert(_sp->verify_chunk_in_free_list(ffc), 8156 "free range is not in free lists"); 8157 } 8158 _sp->removeFreeChunkFromFreeLists(ffc); 8159 set_freeRangeInFreeLists(false); 8160 } 8161 } 8162 } 8163 // Note that if the chunk is not coalescable (the else arm 8164 // below), we unconditionally flush, without needing to do 8165 // a "lookahead," as we do below. 8166 if (inFreeRange()) lookahead_and_flush(fc, size); 8167 } else { 8168 // Code path common to both original and adaptive free lists. 8169 8170 // cant coalesce with previous block; this should be treated 8171 // as the end of a free run if any 8172 if (inFreeRange()) { 8173 // we kicked some butt; time to pick up the garbage 8174 assert(freeFinger() < addr, "freeFinger points too high"); 8175 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 8176 } 8177 // else, nothing to do, just continue 8178 } 8179 } 8180 8181 size_t SweepClosure::do_garbage_chunk(FreeChunk* fc) { 8182 // This is a chunk of garbage. It is not in any free list. 8183 // Add it to a free list or let it possibly be coalesced into 8184 // a larger chunk. 8185 HeapWord* const addr = (HeapWord*) fc; 8186 const size_t size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 8187 8188 if (_sp->adaptive_freelists()) { 8189 // Verify that the bit map has no bits marked between 8190 // addr and purported end of just dead object. 8191 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 8192 8193 do_post_free_or_garbage_chunk(fc, size); 8194 } else { 8195 if (!inFreeRange()) { 8196 // start of a new free range 8197 assert(size > 0, "A free range should have a size"); 8198 initialize_free_range(addr, false); 8199 } else { 8200 // this will be swept up when we hit the end of the 8201 // free range 8202 if (CMSTraceSweeper) { 8203 gclog_or_tty->print(" -- pick up garbage 0x%x (%d) \n", fc, size); 8204 } 8205 // If the chunk is being coalesced and the current free range is 8206 // in the free lists, remove the current free range so that it 8207 // will be returned to the free lists in its entirety - all 8208 // the coalesced pieces included. 8209 if (freeRangeInFreeLists()) { 8210 FreeChunk* ffc = (FreeChunk*)freeFinger(); 8211 assert(ffc->size() == pointer_delta(addr, freeFinger()), 8212 "Size of free range is inconsistent with chunk size."); 8213 if (CMSTestInFreeList) { 8214 assert(_sp->verify_chunk_in_free_list(ffc), 8215 "free range is not in free lists"); 8216 } 8217 _sp->removeFreeChunkFromFreeLists(ffc); 8218 set_freeRangeInFreeLists(false); 8219 } 8220 set_lastFreeRangeCoalesced(true); 8221 } 8222 // this will be swept up when we hit the end of the free range 8223 8224 // Verify that the bit map has no bits marked between 8225 // addr and purported end of just dead object. 8226 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 8227 } 8228 assert(_limit >= addr + size, 8229 "A freshly garbage chunk can't possibly straddle over _limit"); 8230 if (inFreeRange()) lookahead_and_flush(fc, size); 8231 return size; 8232 } 8233 8234 size_t SweepClosure::do_live_chunk(FreeChunk* fc) { 8235 HeapWord* addr = (HeapWord*) fc; 8236 // The sweeper has just found a live object. Return any accumulated 8237 // left hand chunk to the free lists. 8238 if (inFreeRange()) { 8239 assert(freeFinger() < addr, "freeFinger points too high"); 8240 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 8241 } 8242 8243 // This object is live: we'd normally expect this to be 8244 // an oop, and like to assert the following: 8245 // assert(oop(addr)->is_oop(), "live block should be an oop"); 8246 // However, as we commented above, this may be an object whose 8247 // header hasn't yet been initialized. 8248 size_t size; 8249 assert(_bitMap->isMarked(addr), "Tautology for this control point"); 8250 if (_bitMap->isMarked(addr + 1)) { 8251 // Determine the size from the bit map, rather than trying to 8252 // compute it from the object header. 8253 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 8254 size = pointer_delta(nextOneAddr + 1, addr); 8255 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 8256 "alignment problem"); 8257 8258 #ifdef DEBUG 8259 if (oop(addr)->klass_or_null() != NULL) { 8260 // Ignore mark word because we are running concurrent with mutators 8261 assert(oop(addr)->is_oop(true), "live block should be an oop"); 8262 assert(size == 8263 CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()), 8264 "P-mark and computed size do not agree"); 8265 } 8266 #endif 8267 8268 } else { 8269 // This should be an initialized object that's alive. 8270 assert(oop(addr)->klass_or_null() != NULL, 8271 "Should be an initialized object"); 8272 // Ignore mark word because we are running concurrent with mutators 8273 assert(oop(addr)->is_oop(true), "live block should be an oop"); 8274 // Verify that the bit map has no bits marked between 8275 // addr and purported end of this block. 8276 size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 8277 assert(size >= 3, "Necessary for Printezis marks to work"); 8278 assert(!_bitMap->isMarked(addr+1), "Tautology for this control point"); 8279 DEBUG_ONLY(_bitMap->verifyNoOneBitsInRange(addr+2, addr+size);) 8280 } 8281 return size; 8282 } 8283 8284 void SweepClosure::do_post_free_or_garbage_chunk(FreeChunk* fc, 8285 size_t chunkSize) { 8286 // do_post_free_or_garbage_chunk() should only be called in the case 8287 // of the adaptive free list allocator. 8288 const bool fcInFreeLists = fc->is_free(); 8289 assert(_sp->adaptive_freelists(), "Should only be used in this case."); 8290 assert((HeapWord*)fc <= _limit, "sweep invariant"); 8291 if (CMSTestInFreeList && fcInFreeLists) { 8292 assert(_sp->verify_chunk_in_free_list(fc), "free chunk is not in free lists"); 8293 } 8294 8295 if (CMSTraceSweeper) { 8296 gclog_or_tty->print_cr(" -- pick up another chunk at 0x%x (%d)", fc, chunkSize); 8297 } 8298 8299 HeapWord* const fc_addr = (HeapWord*) fc; 8300 8301 bool coalesce; 8302 const size_t left = pointer_delta(fc_addr, freeFinger()); 8303 const size_t right = chunkSize; 8304 switch (FLSCoalescePolicy) { 8305 // numeric value forms a coalition aggressiveness metric 8306 case 0: { // never coalesce 8307 coalesce = false; 8308 break; 8309 } 8310 case 1: { // coalesce if left & right chunks on overpopulated lists 8311 coalesce = _sp->coalOverPopulated(left) && 8312 _sp->coalOverPopulated(right); 8313 break; 8314 } 8315 case 2: { // coalesce if left chunk on overpopulated list (default) 8316 coalesce = _sp->coalOverPopulated(left); 8317 break; 8318 } 8319 case 3: { // coalesce if left OR right chunk on overpopulated list 8320 coalesce = _sp->coalOverPopulated(left) || 8321 _sp->coalOverPopulated(right); 8322 break; 8323 } 8324 case 4: { // always coalesce 8325 coalesce = true; 8326 break; 8327 } 8328 default: 8329 ShouldNotReachHere(); 8330 } 8331 8332 // Should the current free range be coalesced? 8333 // If the chunk is in a free range and either we decided to coalesce above 8334 // or the chunk is near the large block at the end of the heap 8335 // (isNearLargestChunk() returns true), then coalesce this chunk. 8336 const bool doCoalesce = inFreeRange() 8337 && (coalesce || _g->isNearLargestChunk(fc_addr)); 8338 if (doCoalesce) { 8339 // Coalesce the current free range on the left with the new 8340 // chunk on the right. If either is on a free list, 8341 // it must be removed from the list and stashed in the closure. 8342 if (freeRangeInFreeLists()) { 8343 FreeChunk* const ffc = (FreeChunk*)freeFinger(); 8344 assert(ffc->size() == pointer_delta(fc_addr, freeFinger()), 8345 "Size of free range is inconsistent with chunk size."); 8346 if (CMSTestInFreeList) { 8347 assert(_sp->verify_chunk_in_free_list(ffc), 8348 "Chunk is not in free lists"); 8349 } 8350 _sp->coalDeath(ffc->size()); 8351 _sp->removeFreeChunkFromFreeLists(ffc); 8352 set_freeRangeInFreeLists(false); 8353 } 8354 if (fcInFreeLists) { 8355 _sp->coalDeath(chunkSize); 8356 assert(fc->size() == chunkSize, 8357 "The chunk has the wrong size or is not in the free lists"); 8358 _sp->removeFreeChunkFromFreeLists(fc); 8359 } 8360 set_lastFreeRangeCoalesced(true); 8361 print_free_block_coalesced(fc); 8362 } else { // not in a free range and/or should not coalesce 8363 // Return the current free range and start a new one. 8364 if (inFreeRange()) { 8365 // In a free range but cannot coalesce with the right hand chunk. 8366 // Put the current free range into the free lists. 8367 flush_cur_free_chunk(freeFinger(), 8368 pointer_delta(fc_addr, freeFinger())); 8369 } 8370 // Set up for new free range. Pass along whether the right hand 8371 // chunk is in the free lists. 8372 initialize_free_range((HeapWord*)fc, fcInFreeLists); 8373 } 8374 } 8375 8376 // Lookahead flush: 8377 // If we are tracking a free range, and this is the last chunk that 8378 // we'll look at because its end crosses past _limit, we'll preemptively 8379 // flush it along with any free range we may be holding on to. Note that 8380 // this can be the case only for an already free or freshly garbage 8381 // chunk. If this block is an object, it can never straddle 8382 // over _limit. The "straddling" occurs when _limit is set at 8383 // the previous end of the space when this cycle started, and 8384 // a subsequent heap expansion caused the previously co-terminal 8385 // free block to be coalesced with the newly expanded portion, 8386 // thus rendering _limit a non-block-boundary making it dangerous 8387 // for the sweeper to step over and examine. 8388 void SweepClosure::lookahead_and_flush(FreeChunk* fc, size_t chunk_size) { 8389 assert(inFreeRange(), "Should only be called if currently in a free range."); 8390 HeapWord* const eob = ((HeapWord*)fc) + chunk_size; 8391 assert(_sp->used_region().contains(eob - 1), 8392 err_msg("eob = " PTR_FORMAT " out of bounds wrt _sp = [" PTR_FORMAT "," PTR_FORMAT ")" 8393 " when examining fc = " PTR_FORMAT "(" SIZE_FORMAT ")", 8394 _limit, _sp->bottom(), _sp->end(), fc, chunk_size)); 8395 if (eob >= _limit) { 8396 assert(eob == _limit || fc->is_free(), "Only a free chunk should allow us to cross over the limit"); 8397 if (CMSTraceSweeper) { 8398 gclog_or_tty->print_cr("_limit " PTR_FORMAT " reached or crossed by block " 8399 "[" PTR_FORMAT "," PTR_FORMAT ") in space " 8400 "[" PTR_FORMAT "," PTR_FORMAT ")", 8401 _limit, fc, eob, _sp->bottom(), _sp->end()); 8402 } 8403 // Return the storage we are tracking back into the free lists. 8404 if (CMSTraceSweeper) { 8405 gclog_or_tty->print_cr("Flushing ... "); 8406 } 8407 assert(freeFinger() < eob, "Error"); 8408 flush_cur_free_chunk( freeFinger(), pointer_delta(eob, freeFinger())); 8409 } 8410 } 8411 8412 void SweepClosure::flush_cur_free_chunk(HeapWord* chunk, size_t size) { 8413 assert(inFreeRange(), "Should only be called if currently in a free range."); 8414 assert(size > 0, 8415 "A zero sized chunk cannot be added to the free lists."); 8416 if (!freeRangeInFreeLists()) { 8417 if (CMSTestInFreeList) { 8418 FreeChunk* fc = (FreeChunk*) chunk; 8419 fc->set_size(size); 8420 assert(!_sp->verify_chunk_in_free_list(fc), 8421 "chunk should not be in free lists yet"); 8422 } 8423 if (CMSTraceSweeper) { 8424 gclog_or_tty->print_cr(" -- add free block 0x%x (%d) to free lists", 8425 chunk, size); 8426 } 8427 // A new free range is going to be starting. The current 8428 // free range has not been added to the free lists yet or 8429 // was removed so add it back. 8430 // If the current free range was coalesced, then the death 8431 // of the free range was recorded. Record a birth now. 8432 if (lastFreeRangeCoalesced()) { 8433 _sp->coalBirth(size); 8434 } 8435 _sp->addChunkAndRepairOffsetTable(chunk, size, 8436 lastFreeRangeCoalesced()); 8437 } else if (CMSTraceSweeper) { 8438 gclog_or_tty->print_cr("Already in free list: nothing to flush"); 8439 } 8440 set_inFreeRange(false); 8441 set_freeRangeInFreeLists(false); 8442 } 8443 8444 // We take a break if we've been at this for a while, 8445 // so as to avoid monopolizing the locks involved. 8446 void SweepClosure::do_yield_work(HeapWord* addr) { 8447 // Return current free chunk being used for coalescing (if any) 8448 // to the appropriate freelist. After yielding, the next 8449 // free block encountered will start a coalescing range of 8450 // free blocks. If the next free block is adjacent to the 8451 // chunk just flushed, they will need to wait for the next 8452 // sweep to be coalesced. 8453 if (inFreeRange()) { 8454 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 8455 } 8456 8457 // First give up the locks, then yield, then re-lock. 8458 // We should probably use a constructor/destructor idiom to 8459 // do this unlock/lock or modify the MutexUnlocker class to 8460 // serve our purpose. XXX 8461 assert_lock_strong(_bitMap->lock()); 8462 assert_lock_strong(_freelistLock); 8463 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 8464 "CMS thread should hold CMS token"); 8465 _bitMap->lock()->unlock(); 8466 _freelistLock->unlock(); 8467 ConcurrentMarkSweepThread::desynchronize(true); 8468 ConcurrentMarkSweepThread::acknowledge_yield_request(); 8469 _collector->stopTimer(); 8470 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 8471 if (PrintCMSStatistics != 0) { 8472 _collector->incrementYields(); 8473 } 8474 _collector->icms_wait(); 8475 8476 // See the comment in coordinator_yield() 8477 for (unsigned i = 0; i < CMSYieldSleepCount && 8478 ConcurrentMarkSweepThread::should_yield() && 8479 !CMSCollector::foregroundGCIsActive(); ++i) { 8480 os::sleep(Thread::current(), 1, false); 8481 ConcurrentMarkSweepThread::acknowledge_yield_request(); 8482 } 8483 8484 ConcurrentMarkSweepThread::synchronize(true); 8485 _freelistLock->lock(); 8486 _bitMap->lock()->lock_without_safepoint_check(); 8487 _collector->startTimer(); 8488 } 8489 8490 #ifndef PRODUCT 8491 // This is actually very useful in a product build if it can 8492 // be called from the debugger. Compile it into the product 8493 // as needed. 8494 bool debug_verify_chunk_in_free_list(FreeChunk* fc) { 8495 return debug_cms_space->verify_chunk_in_free_list(fc); 8496 } 8497 #endif 8498 8499 void SweepClosure::print_free_block_coalesced(FreeChunk* fc) const { 8500 if (CMSTraceSweeper) { 8501 gclog_or_tty->print_cr("Sweep:coal_free_blk " PTR_FORMAT " (" SIZE_FORMAT ")", 8502 fc, fc->size()); 8503 } 8504 } 8505 8506 // CMSIsAliveClosure 8507 bool CMSIsAliveClosure::do_object_b(oop obj) { 8508 HeapWord* addr = (HeapWord*)obj; 8509 return addr != NULL && 8510 (!_span.contains(addr) || _bit_map->isMarked(addr)); 8511 } 8512 8513 8514 CMSKeepAliveClosure::CMSKeepAliveClosure( CMSCollector* collector, 8515 MemRegion span, 8516 CMSBitMap* bit_map, CMSMarkStack* mark_stack, 8517 bool cpc): 8518 _collector(collector), 8519 _span(span), 8520 _bit_map(bit_map), 8521 _mark_stack(mark_stack), 8522 _concurrent_precleaning(cpc) { 8523 assert(!_span.is_empty(), "Empty span could spell trouble"); 8524 } 8525 8526 8527 // CMSKeepAliveClosure: the serial version 8528 void CMSKeepAliveClosure::do_oop(oop obj) { 8529 HeapWord* addr = (HeapWord*)obj; 8530 if (_span.contains(addr) && 8531 !_bit_map->isMarked(addr)) { 8532 _bit_map->mark(addr); 8533 bool simulate_overflow = false; 8534 NOT_PRODUCT( 8535 if (CMSMarkStackOverflowALot && 8536 _collector->simulate_overflow()) { 8537 // simulate a stack overflow 8538 simulate_overflow = true; 8539 } 8540 ) 8541 if (simulate_overflow || !_mark_stack->push(obj)) { 8542 if (_concurrent_precleaning) { 8543 // We dirty the overflown object and let the remark 8544 // phase deal with it. 8545 assert(_collector->overflow_list_is_empty(), "Error"); 8546 // In the case of object arrays, we need to dirty all of 8547 // the cards that the object spans. No locking or atomics 8548 // are needed since no one else can be mutating the mod union 8549 // table. 8550 if (obj->is_objArray()) { 8551 size_t sz = obj->size(); 8552 HeapWord* end_card_addr = 8553 (HeapWord*)round_to((intptr_t)(addr+sz), CardTableModRefBS::card_size); 8554 MemRegion redirty_range = MemRegion(addr, end_card_addr); 8555 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 8556 _collector->_modUnionTable.mark_range(redirty_range); 8557 } else { 8558 _collector->_modUnionTable.mark(addr); 8559 } 8560 _collector->_ser_kac_preclean_ovflw++; 8561 } else { 8562 _collector->push_on_overflow_list(obj); 8563 _collector->_ser_kac_ovflw++; 8564 } 8565 } 8566 } 8567 } 8568 8569 void CMSKeepAliveClosure::do_oop(oop* p) { CMSKeepAliveClosure::do_oop_work(p); } 8570 void CMSKeepAliveClosure::do_oop(narrowOop* p) { CMSKeepAliveClosure::do_oop_work(p); } 8571 8572 // CMSParKeepAliveClosure: a parallel version of the above. 8573 // The work queues are private to each closure (thread), 8574 // but (may be) available for stealing by other threads. 8575 void CMSParKeepAliveClosure::do_oop(oop obj) { 8576 HeapWord* addr = (HeapWord*)obj; 8577 if (_span.contains(addr) && 8578 !_bit_map->isMarked(addr)) { 8579 // In general, during recursive tracing, several threads 8580 // may be concurrently getting here; the first one to 8581 // "tag" it, claims it. 8582 if (_bit_map->par_mark(addr)) { 8583 bool res = _work_queue->push(obj); 8584 assert(res, "Low water mark should be much less than capacity"); 8585 // Do a recursive trim in the hope that this will keep 8586 // stack usage lower, but leave some oops for potential stealers 8587 trim_queue(_low_water_mark); 8588 } // Else, another thread got there first 8589 } 8590 } 8591 8592 void CMSParKeepAliveClosure::do_oop(oop* p) { CMSParKeepAliveClosure::do_oop_work(p); } 8593 void CMSParKeepAliveClosure::do_oop(narrowOop* p) { CMSParKeepAliveClosure::do_oop_work(p); } 8594 8595 void CMSParKeepAliveClosure::trim_queue(uint max) { 8596 while (_work_queue->size() > max) { 8597 oop new_oop; 8598 if (_work_queue->pop_local(new_oop)) { 8599 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 8600 assert(_bit_map->isMarked((HeapWord*)new_oop), 8601 "no white objects on this stack!"); 8602 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 8603 // iterate over the oops in this oop, marking and pushing 8604 // the ones in CMS heap (i.e. in _span). 8605 new_oop->oop_iterate(&_mark_and_push); 8606 } 8607 } 8608 } 8609 8610 CMSInnerParMarkAndPushClosure::CMSInnerParMarkAndPushClosure( 8611 CMSCollector* collector, 8612 MemRegion span, CMSBitMap* bit_map, 8613 OopTaskQueue* work_queue): 8614 _collector(collector), 8615 _span(span), 8616 _bit_map(bit_map), 8617 _work_queue(work_queue) { } 8618 8619 void CMSInnerParMarkAndPushClosure::do_oop(oop obj) { 8620 HeapWord* addr = (HeapWord*)obj; 8621 if (_span.contains(addr) && 8622 !_bit_map->isMarked(addr)) { 8623 if (_bit_map->par_mark(addr)) { 8624 bool simulate_overflow = false; 8625 NOT_PRODUCT( 8626 if (CMSMarkStackOverflowALot && 8627 _collector->par_simulate_overflow()) { 8628 // simulate a stack overflow 8629 simulate_overflow = true; 8630 } 8631 ) 8632 if (simulate_overflow || !_work_queue->push(obj)) { 8633 _collector->par_push_on_overflow_list(obj); 8634 _collector->_par_kac_ovflw++; 8635 } 8636 } // Else another thread got there already 8637 } 8638 } 8639 8640 void CMSInnerParMarkAndPushClosure::do_oop(oop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); } 8641 void CMSInnerParMarkAndPushClosure::do_oop(narrowOop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); } 8642 8643 ////////////////////////////////////////////////////////////////// 8644 // CMSExpansionCause ///////////////////////////// 8645 ////////////////////////////////////////////////////////////////// 8646 const char* CMSExpansionCause::to_string(CMSExpansionCause::Cause cause) { 8647 switch (cause) { 8648 case _no_expansion: 8649 return "No expansion"; 8650 case _satisfy_free_ratio: 8651 return "Free ratio"; 8652 case _satisfy_promotion: 8653 return "Satisfy promotion"; 8654 case _satisfy_allocation: 8655 return "allocation"; 8656 case _allocate_par_lab: 8657 return "Par LAB"; 8658 case _allocate_par_spooling_space: 8659 return "Par Spooling Space"; 8660 case _adaptive_size_policy: 8661 return "Ergonomics"; 8662 default: 8663 return "unknown"; 8664 } 8665 } 8666 8667 void CMSDrainMarkingStackClosure::do_void() { 8668 // the max number to take from overflow list at a time 8669 const size_t num = _mark_stack->capacity()/4; 8670 assert(!_concurrent_precleaning || _collector->overflow_list_is_empty(), 8671 "Overflow list should be NULL during concurrent phases"); 8672 while (!_mark_stack->isEmpty() || 8673 // if stack is empty, check the overflow list 8674 _collector->take_from_overflow_list(num, _mark_stack)) { 8675 oop obj = _mark_stack->pop(); 8676 HeapWord* addr = (HeapWord*)obj; 8677 assert(_span.contains(addr), "Should be within span"); 8678 assert(_bit_map->isMarked(addr), "Should be marked"); 8679 assert(obj->is_oop(), "Should be an oop"); 8680 obj->oop_iterate(_keep_alive); 8681 } 8682 } 8683 8684 void CMSParDrainMarkingStackClosure::do_void() { 8685 // drain queue 8686 trim_queue(0); 8687 } 8688 8689 // Trim our work_queue so its length is below max at return 8690 void CMSParDrainMarkingStackClosure::trim_queue(uint max) { 8691 while (_work_queue->size() > max) { 8692 oop new_oop; 8693 if (_work_queue->pop_local(new_oop)) { 8694 assert(new_oop->is_oop(), "Expected an oop"); 8695 assert(_bit_map->isMarked((HeapWord*)new_oop), 8696 "no white objects on this stack!"); 8697 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 8698 // iterate over the oops in this oop, marking and pushing 8699 // the ones in CMS heap (i.e. in _span). 8700 new_oop->oop_iterate(&_mark_and_push); 8701 } 8702 } 8703 } 8704 8705 //////////////////////////////////////////////////////////////////// 8706 // Support for Marking Stack Overflow list handling and related code 8707 //////////////////////////////////////////////////////////////////// 8708 // Much of the following code is similar in shape and spirit to the 8709 // code used in ParNewGC. We should try and share that code 8710 // as much as possible in the future. 8711 8712 #ifndef PRODUCT 8713 // Debugging support for CMSStackOverflowALot 8714 8715 // It's OK to call this multi-threaded; the worst thing 8716 // that can happen is that we'll get a bunch of closely 8717 // spaced simulated oveflows, but that's OK, in fact 8718 // probably good as it would exercise the overflow code 8719 // under contention. 8720 bool CMSCollector::simulate_overflow() { 8721 if (_overflow_counter-- <= 0) { // just being defensive 8722 _overflow_counter = CMSMarkStackOverflowInterval; 8723 return true; 8724 } else { 8725 return false; 8726 } 8727 } 8728 8729 bool CMSCollector::par_simulate_overflow() { 8730 return simulate_overflow(); 8731 } 8732 #endif 8733 8734 // Single-threaded 8735 bool CMSCollector::take_from_overflow_list(size_t num, CMSMarkStack* stack) { 8736 assert(stack->isEmpty(), "Expected precondition"); 8737 assert(stack->capacity() > num, "Shouldn't bite more than can chew"); 8738 size_t i = num; 8739 oop cur = _overflow_list; 8740 const markOop proto = markOopDesc::prototype(); 8741 NOT_PRODUCT(ssize_t n = 0;) 8742 for (oop next; i > 0 && cur != NULL; cur = next, i--) { 8743 next = oop(cur->mark()); 8744 cur->set_mark(proto); // until proven otherwise 8745 assert(cur->is_oop(), "Should be an oop"); 8746 bool res = stack->push(cur); 8747 assert(res, "Bit off more than can chew?"); 8748 NOT_PRODUCT(n++;) 8749 } 8750 _overflow_list = cur; 8751 #ifndef PRODUCT 8752 assert(_num_par_pushes >= n, "Too many pops?"); 8753 _num_par_pushes -=n; 8754 #endif 8755 return !stack->isEmpty(); 8756 } 8757 8758 #define BUSY (oop(0x1aff1aff)) 8759 // (MT-safe) Get a prefix of at most "num" from the list. 8760 // The overflow list is chained through the mark word of 8761 // each object in the list. We fetch the entire list, 8762 // break off a prefix of the right size and return the 8763 // remainder. If other threads try to take objects from 8764 // the overflow list at that time, they will wait for 8765 // some time to see if data becomes available. If (and 8766 // only if) another thread places one or more object(s) 8767 // on the global list before we have returned the suffix 8768 // to the global list, we will walk down our local list 8769 // to find its end and append the global list to 8770 // our suffix before returning it. This suffix walk can 8771 // prove to be expensive (quadratic in the amount of traffic) 8772 // when there are many objects in the overflow list and 8773 // there is much producer-consumer contention on the list. 8774 // *NOTE*: The overflow list manipulation code here and 8775 // in ParNewGeneration:: are very similar in shape, 8776 // except that in the ParNew case we use the old (from/eden) 8777 // copy of the object to thread the list via its klass word. 8778 // Because of the common code, if you make any changes in 8779 // the code below, please check the ParNew version to see if 8780 // similar changes might be needed. 8781 // CR 6797058 has been filed to consolidate the common code. 8782 bool CMSCollector::par_take_from_overflow_list(size_t num, 8783 OopTaskQueue* work_q, 8784 int no_of_gc_threads) { 8785 assert(work_q->size() == 0, "First empty local work queue"); 8786 assert(num < work_q->max_elems(), "Can't bite more than we can chew"); 8787 if (_overflow_list == NULL) { 8788 return false; 8789 } 8790 // Grab the entire list; we'll put back a suffix 8791 oop prefix = (oop)Atomic::xchg_ptr(BUSY, &_overflow_list); 8792 Thread* tid = Thread::current(); 8793 // Before "no_of_gc_threads" was introduced CMSOverflowSpinCount was 8794 // set to ParallelGCThreads. 8795 size_t CMSOverflowSpinCount = (size_t) no_of_gc_threads; // was ParallelGCThreads; 8796 size_t sleep_time_millis = MAX2((size_t)1, num/100); 8797 // If the list is busy, we spin for a short while, 8798 // sleeping between attempts to get the list. 8799 for (size_t spin = 0; prefix == BUSY && spin < CMSOverflowSpinCount; spin++) { 8800 os::sleep(tid, sleep_time_millis, false); 8801 if (_overflow_list == NULL) { 8802 // Nothing left to take 8803 return false; 8804 } else if (_overflow_list != BUSY) { 8805 // Try and grab the prefix 8806 prefix = (oop)Atomic::xchg_ptr(BUSY, &_overflow_list); 8807 } 8808 } 8809 // If the list was found to be empty, or we spun long 8810 // enough, we give up and return empty-handed. If we leave 8811 // the list in the BUSY state below, it must be the case that 8812 // some other thread holds the overflow list and will set it 8813 // to a non-BUSY state in the future. 8814 if (prefix == NULL || prefix == BUSY) { 8815 // Nothing to take or waited long enough 8816 if (prefix == NULL) { 8817 // Write back the NULL in case we overwrote it with BUSY above 8818 // and it is still the same value. 8819 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY); 8820 } 8821 return false; 8822 } 8823 assert(prefix != NULL && prefix != BUSY, "Error"); 8824 size_t i = num; 8825 oop cur = prefix; 8826 // Walk down the first "num" objects, unless we reach the end. 8827 for (; i > 1 && cur->mark() != NULL; cur = oop(cur->mark()), i--); 8828 if (cur->mark() == NULL) { 8829 // We have "num" or fewer elements in the list, so there 8830 // is nothing to return to the global list. 8831 // Write back the NULL in lieu of the BUSY we wrote 8832 // above, if it is still the same value. 8833 if (_overflow_list == BUSY) { 8834 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY); 8835 } 8836 } else { 8837 // Chop off the suffix and rerturn it to the global list. 8838 assert(cur->mark() != BUSY, "Error"); 8839 oop suffix_head = cur->mark(); // suffix will be put back on global list 8840 cur->set_mark(NULL); // break off suffix 8841 // It's possible that the list is still in the empty(busy) state 8842 // we left it in a short while ago; in that case we may be 8843 // able to place back the suffix without incurring the cost 8844 // of a walk down the list. 8845 oop observed_overflow_list = _overflow_list; 8846 oop cur_overflow_list = observed_overflow_list; 8847 bool attached = false; 8848 while (observed_overflow_list == BUSY || observed_overflow_list == NULL) { 8849 observed_overflow_list = 8850 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list); 8851 if (cur_overflow_list == observed_overflow_list) { 8852 attached = true; 8853 break; 8854 } else cur_overflow_list = observed_overflow_list; 8855 } 8856 if (!attached) { 8857 // Too bad, someone else sneaked in (at least) an element; we'll need 8858 // to do a splice. Find tail of suffix so we can prepend suffix to global 8859 // list. 8860 for (cur = suffix_head; cur->mark() != NULL; cur = (oop)(cur->mark())); 8861 oop suffix_tail = cur; 8862 assert(suffix_tail != NULL && suffix_tail->mark() == NULL, 8863 "Tautology"); 8864 observed_overflow_list = _overflow_list; 8865 do { 8866 cur_overflow_list = observed_overflow_list; 8867 if (cur_overflow_list != BUSY) { 8868 // Do the splice ... 8869 suffix_tail->set_mark(markOop(cur_overflow_list)); 8870 } else { // cur_overflow_list == BUSY 8871 suffix_tail->set_mark(NULL); 8872 } 8873 // ... and try to place spliced list back on overflow_list ... 8874 observed_overflow_list = 8875 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list); 8876 } while (cur_overflow_list != observed_overflow_list); 8877 // ... until we have succeeded in doing so. 8878 } 8879 } 8880 8881 // Push the prefix elements on work_q 8882 assert(prefix != NULL, "control point invariant"); 8883 const markOop proto = markOopDesc::prototype(); 8884 oop next; 8885 NOT_PRODUCT(ssize_t n = 0;) 8886 for (cur = prefix; cur != NULL; cur = next) { 8887 next = oop(cur->mark()); 8888 cur->set_mark(proto); // until proven otherwise 8889 assert(cur->is_oop(), "Should be an oop"); 8890 bool res = work_q->push(cur); 8891 assert(res, "Bit off more than we can chew?"); 8892 NOT_PRODUCT(n++;) 8893 } 8894 #ifndef PRODUCT 8895 assert(_num_par_pushes >= n, "Too many pops?"); 8896 Atomic::add_ptr(-(intptr_t)n, &_num_par_pushes); 8897 #endif 8898 return true; 8899 } 8900 8901 // Single-threaded 8902 void CMSCollector::push_on_overflow_list(oop p) { 8903 NOT_PRODUCT(_num_par_pushes++;) 8904 assert(p->is_oop(), "Not an oop"); 8905 preserve_mark_if_necessary(p); 8906 p->set_mark((markOop)_overflow_list); 8907 _overflow_list = p; 8908 } 8909 8910 // Multi-threaded; use CAS to prepend to overflow list 8911 void CMSCollector::par_push_on_overflow_list(oop p) { 8912 NOT_PRODUCT(Atomic::inc_ptr(&_num_par_pushes);) 8913 assert(p->is_oop(), "Not an oop"); 8914 par_preserve_mark_if_necessary(p); 8915 oop observed_overflow_list = _overflow_list; 8916 oop cur_overflow_list; 8917 do { 8918 cur_overflow_list = observed_overflow_list; 8919 if (cur_overflow_list != BUSY) { 8920 p->set_mark(markOop(cur_overflow_list)); 8921 } else { 8922 p->set_mark(NULL); 8923 } 8924 observed_overflow_list = 8925 (oop) Atomic::cmpxchg_ptr(p, &_overflow_list, cur_overflow_list); 8926 } while (cur_overflow_list != observed_overflow_list); 8927 } 8928 #undef BUSY 8929 8930 // Single threaded 8931 // General Note on GrowableArray: pushes may silently fail 8932 // because we are (temporarily) out of C-heap for expanding 8933 // the stack. The problem is quite ubiquitous and affects 8934 // a lot of code in the JVM. The prudent thing for GrowableArray 8935 // to do (for now) is to exit with an error. However, that may 8936 // be too draconian in some cases because the caller may be 8937 // able to recover without much harm. For such cases, we 8938 // should probably introduce a "soft_push" method which returns 8939 // an indication of success or failure with the assumption that 8940 // the caller may be able to recover from a failure; code in 8941 // the VM can then be changed, incrementally, to deal with such 8942 // failures where possible, thus, incrementally hardening the VM 8943 // in such low resource situations. 8944 void CMSCollector::preserve_mark_work(oop p, markOop m) { 8945 _preserved_oop_stack.push(p); 8946 _preserved_mark_stack.push(m); 8947 assert(m == p->mark(), "Mark word changed"); 8948 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), 8949 "bijection"); 8950 } 8951 8952 // Single threaded 8953 void CMSCollector::preserve_mark_if_necessary(oop p) { 8954 markOop m = p->mark(); 8955 if (m->must_be_preserved(p)) { 8956 preserve_mark_work(p, m); 8957 } 8958 } 8959 8960 void CMSCollector::par_preserve_mark_if_necessary(oop p) { 8961 markOop m = p->mark(); 8962 if (m->must_be_preserved(p)) { 8963 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag); 8964 // Even though we read the mark word without holding 8965 // the lock, we are assured that it will not change 8966 // because we "own" this oop, so no other thread can 8967 // be trying to push it on the overflow list; see 8968 // the assertion in preserve_mark_work() that checks 8969 // that m == p->mark(). 8970 preserve_mark_work(p, m); 8971 } 8972 } 8973 8974 // We should be able to do this multi-threaded, 8975 // a chunk of stack being a task (this is 8976 // correct because each oop only ever appears 8977 // once in the overflow list. However, it's 8978 // not very easy to completely overlap this with 8979 // other operations, so will generally not be done 8980 // until all work's been completed. Because we 8981 // expect the preserved oop stack (set) to be small, 8982 // it's probably fine to do this single-threaded. 8983 // We can explore cleverer concurrent/overlapped/parallel 8984 // processing of preserved marks if we feel the 8985 // need for this in the future. Stack overflow should 8986 // be so rare in practice and, when it happens, its 8987 // effect on performance so great that this will 8988 // likely just be in the noise anyway. 8989 void CMSCollector::restore_preserved_marks_if_any() { 8990 assert(SafepointSynchronize::is_at_safepoint(), 8991 "world should be stopped"); 8992 assert(Thread::current()->is_ConcurrentGC_thread() || 8993 Thread::current()->is_VM_thread(), 8994 "should be single-threaded"); 8995 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), 8996 "bijection"); 8997 8998 while (!_preserved_oop_stack.is_empty()) { 8999 oop p = _preserved_oop_stack.pop(); 9000 assert(p->is_oop(), "Should be an oop"); 9001 assert(_span.contains(p), "oop should be in _span"); 9002 assert(p->mark() == markOopDesc::prototype(), 9003 "Set when taken from overflow list"); 9004 markOop m = _preserved_mark_stack.pop(); 9005 p->set_mark(m); 9006 } 9007 assert(_preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(), 9008 "stacks were cleared above"); 9009 } 9010 9011 #ifndef PRODUCT 9012 bool CMSCollector::no_preserved_marks() const { 9013 return _preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(); 9014 } 9015 #endif 9016 9017 CMSAdaptiveSizePolicy* ASConcurrentMarkSweepGeneration::cms_size_policy() const 9018 { 9019 GenCollectedHeap* gch = (GenCollectedHeap*) GenCollectedHeap::heap(); 9020 CMSAdaptiveSizePolicy* size_policy = 9021 (CMSAdaptiveSizePolicy*) gch->gen_policy()->size_policy(); 9022 assert(size_policy->is_gc_cms_adaptive_size_policy(), 9023 "Wrong type for size policy"); 9024 return size_policy; 9025 } 9026 9027 void ASConcurrentMarkSweepGeneration::resize(size_t cur_promo_size, 9028 size_t desired_promo_size) { 9029 if (cur_promo_size < desired_promo_size) { 9030 size_t expand_bytes = desired_promo_size - cur_promo_size; 9031 if (PrintAdaptiveSizePolicy && Verbose) { 9032 gclog_or_tty->print_cr(" ASConcurrentMarkSweepGeneration::resize " 9033 "Expanding tenured generation by " SIZE_FORMAT " (bytes)", 9034 expand_bytes); 9035 } 9036 expand(expand_bytes, 9037 MinHeapDeltaBytes, 9038 CMSExpansionCause::_adaptive_size_policy); 9039 } else if (desired_promo_size < cur_promo_size) { 9040 size_t shrink_bytes = cur_promo_size - desired_promo_size; 9041 if (PrintAdaptiveSizePolicy && Verbose) { 9042 gclog_or_tty->print_cr(" ASConcurrentMarkSweepGeneration::resize " 9043 "Shrinking tenured generation by " SIZE_FORMAT " (bytes)", 9044 shrink_bytes); 9045 } 9046 shrink(shrink_bytes); 9047 } 9048 } 9049 9050 CMSGCAdaptivePolicyCounters* ASConcurrentMarkSweepGeneration::gc_adaptive_policy_counters() { 9051 GenCollectedHeap* gch = GenCollectedHeap::heap(); 9052 CMSGCAdaptivePolicyCounters* counters = 9053 (CMSGCAdaptivePolicyCounters*) gch->collector_policy()->counters(); 9054 assert(counters->kind() == GCPolicyCounters::CMSGCAdaptivePolicyCountersKind, 9055 "Wrong kind of counters"); 9056 return counters; 9057 } 9058 9059 9060 void ASConcurrentMarkSweepGeneration::update_counters() { 9061 if (UsePerfData) { 9062 _space_counters->update_all(); 9063 _gen_counters->update_all(); 9064 CMSGCAdaptivePolicyCounters* counters = gc_adaptive_policy_counters(); 9065 GenCollectedHeap* gch = GenCollectedHeap::heap(); 9066 CMSGCStats* gc_stats_l = (CMSGCStats*) gc_stats(); 9067 assert(gc_stats_l->kind() == GCStats::CMSGCStatsKind, 9068 "Wrong gc statistics type"); 9069 counters->update_counters(gc_stats_l); 9070 } 9071 } 9072 9073 void ASConcurrentMarkSweepGeneration::update_counters(size_t used) { 9074 if (UsePerfData) { 9075 _space_counters->update_used(used); 9076 _space_counters->update_capacity(); 9077 _gen_counters->update_all(); 9078 9079 CMSGCAdaptivePolicyCounters* counters = gc_adaptive_policy_counters(); 9080 GenCollectedHeap* gch = GenCollectedHeap::heap(); 9081 CMSGCStats* gc_stats_l = (CMSGCStats*) gc_stats(); 9082 assert(gc_stats_l->kind() == GCStats::CMSGCStatsKind, 9083 "Wrong gc statistics type"); 9084 counters->update_counters(gc_stats_l); 9085 } 9086 } 9087 9088 // The desired expansion delta is computed so that: 9089 // . desired free percentage or greater is used 9090 void ASConcurrentMarkSweepGeneration::compute_new_size() { 9091 assert_locked_or_safepoint(Heap_lock); 9092 9093 GenCollectedHeap* gch = (GenCollectedHeap*) GenCollectedHeap::heap(); 9094 9095 // If incremental collection failed, we just want to expand 9096 // to the limit. 9097 if (incremental_collection_failed()) { 9098 clear_incremental_collection_failed(); 9099 grow_to_reserved(); 9100 return; 9101 } 9102 9103 assert(UseAdaptiveSizePolicy, "Should be using adaptive sizing"); 9104 9105 assert(gch->kind() == CollectedHeap::GenCollectedHeap, 9106 "Wrong type of heap"); 9107 int prev_level = level() - 1; 9108 assert(prev_level >= 0, "The cms generation is the lowest generation"); 9109 Generation* prev_gen = gch->get_gen(prev_level); 9110 assert(prev_gen->kind() == Generation::ASParNew, 9111 "Wrong type of young generation"); 9112 ParNewGeneration* younger_gen = (ParNewGeneration*) prev_gen; 9113 size_t cur_eden = younger_gen->eden()->capacity(); 9114 CMSAdaptiveSizePolicy* size_policy = cms_size_policy(); 9115 size_t cur_promo = free(); 9116 size_policy->compute_tenured_generation_free_space(cur_promo, 9117 max_available(), 9118 cur_eden); 9119 resize(cur_promo, size_policy->promo_size()); 9120 9121 // Record the new size of the space in the cms generation 9122 // that is available for promotions. This is temporary. 9123 // It should be the desired promo size. 9124 size_policy->avg_cms_promo()->sample(free()); 9125 size_policy->avg_old_live()->sample(used()); 9126 9127 if (UsePerfData) { 9128 CMSGCAdaptivePolicyCounters* counters = gc_adaptive_policy_counters(); 9129 counters->update_cms_capacity_counter(capacity()); 9130 } 9131 } 9132 9133 void ASConcurrentMarkSweepGeneration::shrink_by(size_t desired_bytes) { 9134 assert_locked_or_safepoint(Heap_lock); 9135 assert_lock_strong(freelistLock()); 9136 HeapWord* old_end = _cmsSpace->end(); 9137 HeapWord* unallocated_start = _cmsSpace->unallocated_block(); 9138 assert(old_end >= unallocated_start, "Miscalculation of unallocated_start"); 9139 FreeChunk* chunk_at_end = find_chunk_at_end(); 9140 if (chunk_at_end == NULL) { 9141 // No room to shrink 9142 if (PrintGCDetails && Verbose) { 9143 gclog_or_tty->print_cr("No room to shrink: old_end " 9144 PTR_FORMAT " unallocated_start " PTR_FORMAT 9145 " chunk_at_end " PTR_FORMAT, 9146 old_end, unallocated_start, chunk_at_end); 9147 } 9148 return; 9149 } else { 9150 9151 // Find the chunk at the end of the space and determine 9152 // how much it can be shrunk. 9153 size_t shrinkable_size_in_bytes = chunk_at_end->size(); 9154 size_t aligned_shrinkable_size_in_bytes = 9155 align_size_down(shrinkable_size_in_bytes, os::vm_page_size()); 9156 assert(unallocated_start <= chunk_at_end->end(), 9157 "Inconsistent chunk at end of space"); 9158 size_t bytes = MIN2(desired_bytes, aligned_shrinkable_size_in_bytes); 9159 size_t word_size_before = heap_word_size(_virtual_space.committed_size()); 9160 9161 // Shrink the underlying space 9162 _virtual_space.shrink_by(bytes); 9163 if (PrintGCDetails && Verbose) { 9164 gclog_or_tty->print_cr("ConcurrentMarkSweepGeneration::shrink_by:" 9165 " desired_bytes " SIZE_FORMAT 9166 " shrinkable_size_in_bytes " SIZE_FORMAT 9167 " aligned_shrinkable_size_in_bytes " SIZE_FORMAT 9168 " bytes " SIZE_FORMAT, 9169 desired_bytes, shrinkable_size_in_bytes, 9170 aligned_shrinkable_size_in_bytes, bytes); 9171 gclog_or_tty->print_cr(" old_end " SIZE_FORMAT 9172 " unallocated_start " SIZE_FORMAT, 9173 old_end, unallocated_start); 9174 } 9175 9176 // If the space did shrink (shrinking is not guaranteed), 9177 // shrink the chunk at the end by the appropriate amount. 9178 if (((HeapWord*)_virtual_space.high()) < old_end) { 9179 size_t new_word_size = 9180 heap_word_size(_virtual_space.committed_size()); 9181 9182 // Have to remove the chunk from the dictionary because it is changing 9183 // size and might be someplace elsewhere in the dictionary. 9184 9185 // Get the chunk at end, shrink it, and put it 9186 // back. 9187 _cmsSpace->removeChunkFromDictionary(chunk_at_end); 9188 size_t word_size_change = word_size_before - new_word_size; 9189 size_t chunk_at_end_old_size = chunk_at_end->size(); 9190 assert(chunk_at_end_old_size >= word_size_change, 9191 "Shrink is too large"); 9192 chunk_at_end->set_size(chunk_at_end_old_size - 9193 word_size_change); 9194 _cmsSpace->freed((HeapWord*) chunk_at_end->end(), 9195 word_size_change); 9196 9197 _cmsSpace->returnChunkToDictionary(chunk_at_end); 9198 9199 MemRegion mr(_cmsSpace->bottom(), new_word_size); 9200 _bts->resize(new_word_size); // resize the block offset shared array 9201 Universe::heap()->barrier_set()->resize_covered_region(mr); 9202 _cmsSpace->assert_locked(); 9203 _cmsSpace->set_end((HeapWord*)_virtual_space.high()); 9204 9205 NOT_PRODUCT(_cmsSpace->dictionary()->verify()); 9206 9207 // update the space and generation capacity counters 9208 if (UsePerfData) { 9209 _space_counters->update_capacity(); 9210 _gen_counters->update_all(); 9211 } 9212 9213 if (Verbose && PrintGCDetails) { 9214 size_t new_mem_size = _virtual_space.committed_size(); 9215 size_t old_mem_size = new_mem_size + bytes; 9216 gclog_or_tty->print_cr("Shrinking %s from %ldK by %ldK to %ldK", 9217 name(), old_mem_size/K, bytes/K, new_mem_size/K); 9218 } 9219 } 9220 9221 assert(_cmsSpace->unallocated_block() <= _cmsSpace->end(), 9222 "Inconsistency at end of space"); 9223 assert(chunk_at_end->end() == _cmsSpace->end(), 9224 "Shrinking is inconsistent"); 9225 return; 9226 } 9227 } 9228 9229 // Transfer some number of overflown objects to usual marking 9230 // stack. Return true if some objects were transferred. 9231 bool MarkRefsIntoAndScanClosure::take_from_overflow_list() { 9232 size_t num = MIN2((size_t)(_mark_stack->capacity() - _mark_stack->length())/4, 9233 (size_t)ParGCDesiredObjsFromOverflowList); 9234 9235 bool res = _collector->take_from_overflow_list(num, _mark_stack); 9236 assert(_collector->overflow_list_is_empty() || res, 9237 "If list is not empty, we should have taken something"); 9238 assert(!res || !_mark_stack->isEmpty(), 9239 "If we took something, it should now be on our stack"); 9240 return res; 9241 } 9242 9243 size_t MarkDeadObjectsClosure::do_blk(HeapWord* addr) { 9244 size_t res = _sp->block_size_no_stall(addr, _collector); 9245 if (_sp->block_is_obj(addr)) { 9246 if (_live_bit_map->isMarked(addr)) { 9247 // It can't have been dead in a previous cycle 9248 guarantee(!_dead_bit_map->isMarked(addr), "No resurrection!"); 9249 } else { 9250 _dead_bit_map->mark(addr); // mark the dead object 9251 } 9252 } 9253 // Could be 0, if the block size could not be computed without stalling. 9254 return res; 9255 } 9256 9257 TraceCMSMemoryManagerStats::TraceCMSMemoryManagerStats(CMSCollector::CollectorState phase, GCCause::Cause cause): TraceMemoryManagerStats() { 9258 9259 switch (phase) { 9260 case CMSCollector::InitialMarking: 9261 initialize(true /* fullGC */ , 9262 cause /* cause of the GC */, 9263 true /* recordGCBeginTime */, 9264 true /* recordPreGCUsage */, 9265 false /* recordPeakUsage */, 9266 false /* recordPostGCusage */, 9267 true /* recordAccumulatedGCTime */, 9268 false /* recordGCEndTime */, 9269 false /* countCollection */ ); 9270 break; 9271 9272 case CMSCollector::FinalMarking: 9273 initialize(true /* fullGC */ , 9274 cause /* cause of the GC */, 9275 false /* recordGCBeginTime */, 9276 false /* recordPreGCUsage */, 9277 false /* recordPeakUsage */, 9278 false /* recordPostGCusage */, 9279 true /* recordAccumulatedGCTime */, 9280 false /* recordGCEndTime */, 9281 false /* countCollection */ ); 9282 break; 9283 9284 case CMSCollector::Sweeping: 9285 initialize(true /* fullGC */ , 9286 cause /* cause of the GC */, 9287 false /* recordGCBeginTime */, 9288 false /* recordPreGCUsage */, 9289 true /* recordPeakUsage */, 9290 true /* recordPostGCusage */, 9291 false /* recordAccumulatedGCTime */, 9292 true /* recordGCEndTime */, 9293 true /* countCollection */ ); 9294 break; 9295 9296 default: 9297 ShouldNotReachHere(); 9298 } 9299 } 9300