1 /* 2 * Copyright (c) 2001, 2013, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "classfile/classLoaderData.hpp" 27 #include "classfile/symbolTable.hpp" 28 #include "classfile/systemDictionary.hpp" 29 #include "code/codeCache.hpp" 30 #include "gc_implementation/concurrentMarkSweep/cmsAdaptiveSizePolicy.hpp" 31 #include "gc_implementation/concurrentMarkSweep/cmsCollectorPolicy.hpp" 32 #include "gc_implementation/concurrentMarkSweep/cmsGCAdaptivePolicyCounters.hpp" 33 #include "gc_implementation/concurrentMarkSweep/cmsOopClosures.inline.hpp" 34 #include "gc_implementation/concurrentMarkSweep/compactibleFreeListSpace.hpp" 35 #include "gc_implementation/concurrentMarkSweep/concurrentMarkSweepGeneration.inline.hpp" 36 #include "gc_implementation/concurrentMarkSweep/concurrentMarkSweepThread.hpp" 37 #include "gc_implementation/concurrentMarkSweep/vmCMSOperations.hpp" 38 #include "gc_implementation/parNew/parNewGeneration.hpp" 39 #include "gc_implementation/shared/collectorCounters.hpp" 40 #include "gc_implementation/shared/isGCActiveMark.hpp" 41 #include "gc_interface/collectedHeap.inline.hpp" 42 #include "memory/cardTableRS.hpp" 43 #include "memory/collectorPolicy.hpp" 44 #include "memory/gcLocker.inline.hpp" 45 #include "memory/genCollectedHeap.hpp" 46 #include "memory/genMarkSweep.hpp" 47 #include "memory/genOopClosures.inline.hpp" 48 #include "memory/iterator.hpp" 49 #include "memory/referencePolicy.hpp" 50 #include "memory/resourceArea.hpp" 51 #include "memory/tenuredGeneration.hpp" 52 #include "oops/oop.inline.hpp" 53 #include "prims/jvmtiExport.hpp" 54 #include "runtime/globals_extension.hpp" 55 #include "runtime/handles.inline.hpp" 56 #include "runtime/java.hpp" 57 #include "runtime/vmThread.hpp" 58 #include "services/memoryService.hpp" 59 #include "services/runtimeService.hpp" 60 61 // statics 62 CMSCollector* ConcurrentMarkSweepGeneration::_collector = NULL; 63 bool CMSCollector::_full_gc_requested = false; 64 65 ////////////////////////////////////////////////////////////////// 66 // In support of CMS/VM thread synchronization 67 ////////////////////////////////////////////////////////////////// 68 // We split use of the CGC_lock into 2 "levels". 69 // The low-level locking is of the usual CGC_lock monitor. We introduce 70 // a higher level "token" (hereafter "CMS token") built on top of the 71 // low level monitor (hereafter "CGC lock"). 72 // The token-passing protocol gives priority to the VM thread. The 73 // CMS-lock doesn't provide any fairness guarantees, but clients 74 // should ensure that it is only held for very short, bounded 75 // durations. 76 // 77 // When either of the CMS thread or the VM thread is involved in 78 // collection operations during which it does not want the other 79 // thread to interfere, it obtains the CMS token. 80 // 81 // If either thread tries to get the token while the other has 82 // it, that thread waits. However, if the VM thread and CMS thread 83 // both want the token, then the VM thread gets priority while the 84 // CMS thread waits. This ensures, for instance, that the "concurrent" 85 // phases of the CMS thread's work do not block out the VM thread 86 // for long periods of time as the CMS thread continues to hog 87 // the token. (See bug 4616232). 88 // 89 // The baton-passing functions are, however, controlled by the 90 // flags _foregroundGCShouldWait and _foregroundGCIsActive, 91 // and here the low-level CMS lock, not the high level token, 92 // ensures mutual exclusion. 93 // 94 // Two important conditions that we have to satisfy: 95 // 1. if a thread does a low-level wait on the CMS lock, then it 96 // relinquishes the CMS token if it were holding that token 97 // when it acquired the low-level CMS lock. 98 // 2. any low-level notifications on the low-level lock 99 // should only be sent when a thread has relinquished the token. 100 // 101 // In the absence of either property, we'd have potential deadlock. 102 // 103 // We protect each of the CMS (concurrent and sequential) phases 104 // with the CMS _token_, not the CMS _lock_. 105 // 106 // The only code protected by CMS lock is the token acquisition code 107 // itself, see ConcurrentMarkSweepThread::[de]synchronize(), and the 108 // baton-passing code. 109 // 110 // Unfortunately, i couldn't come up with a good abstraction to factor and 111 // hide the naked CGC_lock manipulation in the baton-passing code 112 // further below. That's something we should try to do. Also, the proof 113 // of correctness of this 2-level locking scheme is far from obvious, 114 // and potentially quite slippery. We have an uneasy supsicion, for instance, 115 // that there may be a theoretical possibility of delay/starvation in the 116 // low-level lock/wait/notify scheme used for the baton-passing because of 117 // potential intereference with the priority scheme embodied in the 118 // CMS-token-passing protocol. See related comments at a CGC_lock->wait() 119 // invocation further below and marked with "XXX 20011219YSR". 120 // Indeed, as we note elsewhere, this may become yet more slippery 121 // in the presence of multiple CMS and/or multiple VM threads. XXX 122 123 class CMSTokenSync: public StackObj { 124 private: 125 bool _is_cms_thread; 126 public: 127 CMSTokenSync(bool is_cms_thread): 128 _is_cms_thread(is_cms_thread) { 129 assert(is_cms_thread == Thread::current()->is_ConcurrentGC_thread(), 130 "Incorrect argument to constructor"); 131 ConcurrentMarkSweepThread::synchronize(_is_cms_thread); 132 } 133 134 ~CMSTokenSync() { 135 assert(_is_cms_thread ? 136 ConcurrentMarkSweepThread::cms_thread_has_cms_token() : 137 ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 138 "Incorrect state"); 139 ConcurrentMarkSweepThread::desynchronize(_is_cms_thread); 140 } 141 }; 142 143 // Convenience class that does a CMSTokenSync, and then acquires 144 // upto three locks. 145 class CMSTokenSyncWithLocks: public CMSTokenSync { 146 private: 147 // Note: locks are acquired in textual declaration order 148 // and released in the opposite order 149 MutexLockerEx _locker1, _locker2, _locker3; 150 public: 151 CMSTokenSyncWithLocks(bool is_cms_thread, Mutex* mutex1, 152 Mutex* mutex2 = NULL, Mutex* mutex3 = NULL): 153 CMSTokenSync(is_cms_thread), 154 _locker1(mutex1, Mutex::_no_safepoint_check_flag), 155 _locker2(mutex2, Mutex::_no_safepoint_check_flag), 156 _locker3(mutex3, Mutex::_no_safepoint_check_flag) 157 { } 158 }; 159 160 161 // Wrapper class to temporarily disable icms during a foreground cms collection. 162 class ICMSDisabler: public StackObj { 163 public: 164 // The ctor disables icms and wakes up the thread so it notices the change; 165 // the dtor re-enables icms. Note that the CMSCollector methods will check 166 // CMSIncrementalMode. 167 ICMSDisabler() { CMSCollector::disable_icms(); CMSCollector::start_icms(); } 168 ~ICMSDisabler() { CMSCollector::enable_icms(); } 169 }; 170 171 ////////////////////////////////////////////////////////////////// 172 // Concurrent Mark-Sweep Generation ///////////////////////////// 173 ////////////////////////////////////////////////////////////////// 174 175 NOT_PRODUCT(CompactibleFreeListSpace* debug_cms_space;) 176 177 // This struct contains per-thread things necessary to support parallel 178 // young-gen collection. 179 class CMSParGCThreadState: public CHeapObj<mtGC> { 180 public: 181 CFLS_LAB lab; 182 PromotionInfo promo; 183 184 // Constructor. 185 CMSParGCThreadState(CompactibleFreeListSpace* cfls) : lab(cfls) { 186 promo.setSpace(cfls); 187 } 188 }; 189 190 ConcurrentMarkSweepGeneration::ConcurrentMarkSweepGeneration( 191 ReservedSpace rs, size_t initial_byte_size, int level, 192 CardTableRS* ct, bool use_adaptive_freelists, 193 FreeBlockDictionary<FreeChunk>::DictionaryChoice dictionaryChoice) : 194 CardGeneration(rs, initial_byte_size, level, ct), 195 _dilatation_factor(((double)MinChunkSize)/((double)(CollectedHeap::min_fill_size()))), 196 _debug_collection_type(Concurrent_collection_type), 197 _did_compact(false) 198 { 199 HeapWord* bottom = (HeapWord*) _virtual_space.low(); 200 HeapWord* end = (HeapWord*) _virtual_space.high(); 201 202 _direct_allocated_words = 0; 203 NOT_PRODUCT( 204 _numObjectsPromoted = 0; 205 _numWordsPromoted = 0; 206 _numObjectsAllocated = 0; 207 _numWordsAllocated = 0; 208 ) 209 210 _cmsSpace = new CompactibleFreeListSpace(_bts, MemRegion(bottom, end), 211 use_adaptive_freelists, 212 dictionaryChoice); 213 NOT_PRODUCT(debug_cms_space = _cmsSpace;) 214 if (_cmsSpace == NULL) { 215 vm_exit_during_initialization( 216 "CompactibleFreeListSpace allocation failure"); 217 } 218 _cmsSpace->_gen = this; 219 220 _gc_stats = new CMSGCStats(); 221 222 // Verify the assumption that FreeChunk::_prev and OopDesc::_klass 223 // offsets match. The ability to tell free chunks from objects 224 // depends on this property. 225 debug_only( 226 FreeChunk* junk = NULL; 227 assert(UseCompressedKlassPointers || 228 junk->prev_addr() == (void*)(oop(junk)->klass_addr()), 229 "Offset of FreeChunk::_prev within FreeChunk must match" 230 " that of OopDesc::_klass within OopDesc"); 231 ) 232 if (CollectedHeap::use_parallel_gc_threads()) { 233 typedef CMSParGCThreadState* CMSParGCThreadStatePtr; 234 _par_gc_thread_states = 235 NEW_C_HEAP_ARRAY(CMSParGCThreadStatePtr, ParallelGCThreads, mtGC); 236 if (_par_gc_thread_states == NULL) { 237 vm_exit_during_initialization("Could not allocate par gc structs"); 238 } 239 for (uint i = 0; i < ParallelGCThreads; i++) { 240 _par_gc_thread_states[i] = new CMSParGCThreadState(cmsSpace()); 241 if (_par_gc_thread_states[i] == NULL) { 242 vm_exit_during_initialization("Could not allocate par gc structs"); 243 } 244 } 245 } else { 246 _par_gc_thread_states = NULL; 247 } 248 _incremental_collection_failed = false; 249 // The "dilatation_factor" is the expansion that can occur on 250 // account of the fact that the minimum object size in the CMS 251 // generation may be larger than that in, say, a contiguous young 252 // generation. 253 // Ideally, in the calculation below, we'd compute the dilatation 254 // factor as: MinChunkSize/(promoting_gen's min object size) 255 // Since we do not have such a general query interface for the 256 // promoting generation, we'll instead just use the mimimum 257 // object size (which today is a header's worth of space); 258 // note that all arithmetic is in units of HeapWords. 259 assert(MinChunkSize >= CollectedHeap::min_fill_size(), "just checking"); 260 assert(_dilatation_factor >= 1.0, "from previous assert"); 261 } 262 263 264 // The field "_initiating_occupancy" represents the occupancy percentage 265 // at which we trigger a new collection cycle. Unless explicitly specified 266 // via CMSInitiatingOccupancyFraction (argument "io" below), it 267 // is calculated by: 268 // 269 // Let "f" be MinHeapFreeRatio in 270 // 271 // _intiating_occupancy = 100-f + 272 // f * (CMSTriggerRatio/100) 273 // where CMSTriggerRatio is the argument "tr" below. 274 // 275 // That is, if we assume the heap is at its desired maximum occupancy at the 276 // end of a collection, we let CMSTriggerRatio of the (purported) free 277 // space be allocated before initiating a new collection cycle. 278 // 279 void ConcurrentMarkSweepGeneration::init_initiating_occupancy(intx io, uintx tr) { 280 assert(io <= 100 && tr <= 100, "Check the arguments"); 281 if (io >= 0) { 282 _initiating_occupancy = (double)io / 100.0; 283 } else { 284 _initiating_occupancy = ((100 - MinHeapFreeRatio) + 285 (double)(tr * MinHeapFreeRatio) / 100.0) 286 / 100.0; 287 } 288 } 289 290 void ConcurrentMarkSweepGeneration::ref_processor_init() { 291 assert(collector() != NULL, "no collector"); 292 collector()->ref_processor_init(); 293 } 294 295 void CMSCollector::ref_processor_init() { 296 if (_ref_processor == NULL) { 297 // Allocate and initialize a reference processor 298 _ref_processor = 299 new ReferenceProcessor(_span, // span 300 (ParallelGCThreads > 1) && ParallelRefProcEnabled, // mt processing 301 (int) ParallelGCThreads, // mt processing degree 302 _cmsGen->refs_discovery_is_mt(), // mt discovery 303 (int) MAX2(ConcGCThreads, ParallelGCThreads), // mt discovery degree 304 _cmsGen->refs_discovery_is_atomic(), // discovery is not atomic 305 &_is_alive_closure, // closure for liveness info 306 false); // next field updates do not need write barrier 307 // Initialize the _ref_processor field of CMSGen 308 _cmsGen->set_ref_processor(_ref_processor); 309 310 } 311 } 312 313 CMSAdaptiveSizePolicy* CMSCollector::size_policy() { 314 GenCollectedHeap* gch = GenCollectedHeap::heap(); 315 assert(gch->kind() == CollectedHeap::GenCollectedHeap, 316 "Wrong type of heap"); 317 CMSAdaptiveSizePolicy* sp = (CMSAdaptiveSizePolicy*) 318 gch->gen_policy()->size_policy(); 319 assert(sp->is_gc_cms_adaptive_size_policy(), 320 "Wrong type of size policy"); 321 return sp; 322 } 323 324 CMSGCAdaptivePolicyCounters* CMSCollector::gc_adaptive_policy_counters() { 325 CMSGCAdaptivePolicyCounters* results = 326 (CMSGCAdaptivePolicyCounters*) collector_policy()->counters(); 327 assert( 328 results->kind() == GCPolicyCounters::CMSGCAdaptivePolicyCountersKind, 329 "Wrong gc policy counter kind"); 330 return results; 331 } 332 333 334 void ConcurrentMarkSweepGeneration::initialize_performance_counters() { 335 336 const char* gen_name = "old"; 337 338 // Generation Counters - generation 1, 1 subspace 339 _gen_counters = new GenerationCounters(gen_name, 1, 1, &_virtual_space); 340 341 _space_counters = new GSpaceCounters(gen_name, 0, 342 _virtual_space.reserved_size(), 343 this, _gen_counters); 344 } 345 346 CMSStats::CMSStats(ConcurrentMarkSweepGeneration* cms_gen, unsigned int alpha): 347 _cms_gen(cms_gen) 348 { 349 assert(alpha <= 100, "bad value"); 350 _saved_alpha = alpha; 351 352 // Initialize the alphas to the bootstrap value of 100. 353 _gc0_alpha = _cms_alpha = 100; 354 355 _cms_begin_time.update(); 356 _cms_end_time.update(); 357 358 _gc0_duration = 0.0; 359 _gc0_period = 0.0; 360 _gc0_promoted = 0; 361 362 _cms_duration = 0.0; 363 _cms_period = 0.0; 364 _cms_allocated = 0; 365 366 _cms_used_at_gc0_begin = 0; 367 _cms_used_at_gc0_end = 0; 368 _allow_duty_cycle_reduction = false; 369 _valid_bits = 0; 370 _icms_duty_cycle = CMSIncrementalDutyCycle; 371 } 372 373 double CMSStats::cms_free_adjustment_factor(size_t free) const { 374 // TBD: CR 6909490 375 return 1.0; 376 } 377 378 void CMSStats::adjust_cms_free_adjustment_factor(bool fail, size_t free) { 379 } 380 381 // If promotion failure handling is on use 382 // the padded average size of the promotion for each 383 // young generation collection. 384 double CMSStats::time_until_cms_gen_full() const { 385 size_t cms_free = _cms_gen->cmsSpace()->free(); 386 GenCollectedHeap* gch = GenCollectedHeap::heap(); 387 size_t expected_promotion = MIN2(gch->get_gen(0)->capacity(), 388 (size_t) _cms_gen->gc_stats()->avg_promoted()->padded_average()); 389 if (cms_free > expected_promotion) { 390 // Start a cms collection if there isn't enough space to promote 391 // for the next minor collection. Use the padded average as 392 // a safety factor. 393 cms_free -= expected_promotion; 394 395 // Adjust by the safety factor. 396 double cms_free_dbl = (double)cms_free; 397 double cms_adjustment = (100.0 - CMSIncrementalSafetyFactor)/100.0; 398 // Apply a further correction factor which tries to adjust 399 // for recent occurance of concurrent mode failures. 400 cms_adjustment = cms_adjustment * cms_free_adjustment_factor(cms_free); 401 cms_free_dbl = cms_free_dbl * cms_adjustment; 402 403 if (PrintGCDetails && Verbose) { 404 gclog_or_tty->print_cr("CMSStats::time_until_cms_gen_full: cms_free " 405 SIZE_FORMAT " expected_promotion " SIZE_FORMAT, 406 cms_free, expected_promotion); 407 gclog_or_tty->print_cr(" cms_free_dbl %f cms_consumption_rate %f", 408 cms_free_dbl, cms_consumption_rate() + 1.0); 409 } 410 // Add 1 in case the consumption rate goes to zero. 411 return cms_free_dbl / (cms_consumption_rate() + 1.0); 412 } 413 return 0.0; 414 } 415 416 // Compare the duration of the cms collection to the 417 // time remaining before the cms generation is empty. 418 // Note that the time from the start of the cms collection 419 // to the start of the cms sweep (less than the total 420 // duration of the cms collection) can be used. This 421 // has been tried and some applications experienced 422 // promotion failures early in execution. This was 423 // possibly because the averages were not accurate 424 // enough at the beginning. 425 double CMSStats::time_until_cms_start() const { 426 // We add "gc0_period" to the "work" calculation 427 // below because this query is done (mostly) at the 428 // end of a scavenge, so we need to conservatively 429 // account for that much possible delay 430 // in the query so as to avoid concurrent mode failures 431 // due to starting the collection just a wee bit too 432 // late. 433 double work = cms_duration() + gc0_period(); 434 double deadline = time_until_cms_gen_full(); 435 // If a concurrent mode failure occurred recently, we want to be 436 // more conservative and halve our expected time_until_cms_gen_full() 437 if (work > deadline) { 438 if (Verbose && PrintGCDetails) { 439 gclog_or_tty->print( 440 " CMSCollector: collect because of anticipated promotion " 441 "before full %3.7f + %3.7f > %3.7f ", cms_duration(), 442 gc0_period(), time_until_cms_gen_full()); 443 } 444 return 0.0; 445 } 446 return work - deadline; 447 } 448 449 // Return a duty cycle based on old_duty_cycle and new_duty_cycle, limiting the 450 // amount of change to prevent wild oscillation. 451 unsigned int CMSStats::icms_damped_duty_cycle(unsigned int old_duty_cycle, 452 unsigned int new_duty_cycle) { 453 assert(old_duty_cycle <= 100, "bad input value"); 454 assert(new_duty_cycle <= 100, "bad input value"); 455 456 // Note: use subtraction with caution since it may underflow (values are 457 // unsigned). Addition is safe since we're in the range 0-100. 458 unsigned int damped_duty_cycle = new_duty_cycle; 459 if (new_duty_cycle < old_duty_cycle) { 460 const unsigned int largest_delta = MAX2(old_duty_cycle / 4, 5U); 461 if (new_duty_cycle + largest_delta < old_duty_cycle) { 462 damped_duty_cycle = old_duty_cycle - largest_delta; 463 } 464 } else if (new_duty_cycle > old_duty_cycle) { 465 const unsigned int largest_delta = MAX2(old_duty_cycle / 4, 15U); 466 if (new_duty_cycle > old_duty_cycle + largest_delta) { 467 damped_duty_cycle = MIN2(old_duty_cycle + largest_delta, 100U); 468 } 469 } 470 assert(damped_duty_cycle <= 100, "invalid duty cycle computed"); 471 472 if (CMSTraceIncrementalPacing) { 473 gclog_or_tty->print(" [icms_damped_duty_cycle(%d,%d) = %d] ", 474 old_duty_cycle, new_duty_cycle, damped_duty_cycle); 475 } 476 return damped_duty_cycle; 477 } 478 479 unsigned int CMSStats::icms_update_duty_cycle_impl() { 480 assert(CMSIncrementalPacing && valid(), 481 "should be handled in icms_update_duty_cycle()"); 482 483 double cms_time_so_far = cms_timer().seconds(); 484 double scaled_duration = cms_duration_per_mb() * _cms_used_at_gc0_end / M; 485 double scaled_duration_remaining = fabsd(scaled_duration - cms_time_so_far); 486 487 // Avoid division by 0. 488 double time_until_full = MAX2(time_until_cms_gen_full(), 0.01); 489 double duty_cycle_dbl = 100.0 * scaled_duration_remaining / time_until_full; 490 491 unsigned int new_duty_cycle = MIN2((unsigned int)duty_cycle_dbl, 100U); 492 if (new_duty_cycle > _icms_duty_cycle) { 493 // Avoid very small duty cycles (1 or 2); 0 is allowed. 494 if (new_duty_cycle > 2) { 495 _icms_duty_cycle = icms_damped_duty_cycle(_icms_duty_cycle, 496 new_duty_cycle); 497 } 498 } else if (_allow_duty_cycle_reduction) { 499 // The duty cycle is reduced only once per cms cycle (see record_cms_end()). 500 new_duty_cycle = icms_damped_duty_cycle(_icms_duty_cycle, new_duty_cycle); 501 // Respect the minimum duty cycle. 502 unsigned int min_duty_cycle = (unsigned int)CMSIncrementalDutyCycleMin; 503 _icms_duty_cycle = MAX2(new_duty_cycle, min_duty_cycle); 504 } 505 506 if (PrintGCDetails || CMSTraceIncrementalPacing) { 507 gclog_or_tty->print(" icms_dc=%d ", _icms_duty_cycle); 508 } 509 510 _allow_duty_cycle_reduction = false; 511 return _icms_duty_cycle; 512 } 513 514 #ifndef PRODUCT 515 void CMSStats::print_on(outputStream *st) const { 516 st->print(" gc0_alpha=%d,cms_alpha=%d", _gc0_alpha, _cms_alpha); 517 st->print(",gc0_dur=%g,gc0_per=%g,gc0_promo=" SIZE_FORMAT, 518 gc0_duration(), gc0_period(), gc0_promoted()); 519 st->print(",cms_dur=%g,cms_dur_per_mb=%g,cms_per=%g,cms_alloc=" SIZE_FORMAT, 520 cms_duration(), cms_duration_per_mb(), 521 cms_period(), cms_allocated()); 522 st->print(",cms_since_beg=%g,cms_since_end=%g", 523 cms_time_since_begin(), cms_time_since_end()); 524 st->print(",cms_used_beg=" SIZE_FORMAT ",cms_used_end=" SIZE_FORMAT, 525 _cms_used_at_gc0_begin, _cms_used_at_gc0_end); 526 if (CMSIncrementalMode) { 527 st->print(",dc=%d", icms_duty_cycle()); 528 } 529 530 if (valid()) { 531 st->print(",promo_rate=%g,cms_alloc_rate=%g", 532 promotion_rate(), cms_allocation_rate()); 533 st->print(",cms_consumption_rate=%g,time_until_full=%g", 534 cms_consumption_rate(), time_until_cms_gen_full()); 535 } 536 st->print(" "); 537 } 538 #endif // #ifndef PRODUCT 539 540 CMSCollector::CollectorState CMSCollector::_collectorState = 541 CMSCollector::Idling; 542 bool CMSCollector::_foregroundGCIsActive = false; 543 bool CMSCollector::_foregroundGCShouldWait = false; 544 545 CMSCollector::CMSCollector(ConcurrentMarkSweepGeneration* cmsGen, 546 CardTableRS* ct, 547 ConcurrentMarkSweepPolicy* cp): 548 _cmsGen(cmsGen), 549 _ct(ct), 550 _ref_processor(NULL), // will be set later 551 _conc_workers(NULL), // may be set later 552 _abort_preclean(false), 553 _start_sampling(false), 554 _between_prologue_and_epilogue(false), 555 _markBitMap(0, Mutex::leaf + 1, "CMS_markBitMap_lock"), 556 _modUnionTable((CardTableModRefBS::card_shift - LogHeapWordSize), 557 -1 /* lock-free */, "No_lock" /* dummy */), 558 _modUnionClosure(&_modUnionTable), 559 _modUnionClosurePar(&_modUnionTable), 560 // Adjust my span to cover old (cms) gen 561 _span(cmsGen->reserved()), 562 // Construct the is_alive_closure with _span & markBitMap 563 _is_alive_closure(_span, &_markBitMap), 564 _restart_addr(NULL), 565 _overflow_list(NULL), 566 _stats(cmsGen), 567 _eden_chunk_array(NULL), // may be set in ctor body 568 _eden_chunk_capacity(0), // -- ditto -- 569 _eden_chunk_index(0), // -- ditto -- 570 _survivor_plab_array(NULL), // -- ditto -- 571 _survivor_chunk_array(NULL), // -- ditto -- 572 _survivor_chunk_capacity(0), // -- ditto -- 573 _survivor_chunk_index(0), // -- ditto -- 574 _ser_pmc_preclean_ovflw(0), 575 _ser_kac_preclean_ovflw(0), 576 _ser_pmc_remark_ovflw(0), 577 _par_pmc_remark_ovflw(0), 578 _ser_kac_ovflw(0), 579 _par_kac_ovflw(0), 580 #ifndef PRODUCT 581 _num_par_pushes(0), 582 #endif 583 _collection_count_start(0), 584 _verifying(false), 585 _icms_start_limit(NULL), 586 _icms_stop_limit(NULL), 587 _verification_mark_bm(0, Mutex::leaf + 1, "CMS_verification_mark_bm_lock"), 588 _completed_initialization(false), 589 _collector_policy(cp), 590 _should_unload_classes(false), 591 _concurrent_cycles_since_last_unload(0), 592 _roots_scanning_options(0), 593 _inter_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding), 594 _intra_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding) 595 { 596 if (ExplicitGCInvokesConcurrentAndUnloadsClasses) { 597 ExplicitGCInvokesConcurrent = true; 598 } 599 // Now expand the span and allocate the collection support structures 600 // (MUT, marking bit map etc.) to cover both generations subject to 601 // collection. 602 603 // For use by dirty card to oop closures. 604 _cmsGen->cmsSpace()->set_collector(this); 605 606 // Allocate MUT and marking bit map 607 { 608 MutexLockerEx x(_markBitMap.lock(), Mutex::_no_safepoint_check_flag); 609 if (!_markBitMap.allocate(_span)) { 610 warning("Failed to allocate CMS Bit Map"); 611 return; 612 } 613 assert(_markBitMap.covers(_span), "_markBitMap inconsistency?"); 614 } 615 { 616 _modUnionTable.allocate(_span); 617 assert(_modUnionTable.covers(_span), "_modUnionTable inconsistency?"); 618 } 619 620 if (!_markStack.allocate(MarkStackSize)) { 621 warning("Failed to allocate CMS Marking Stack"); 622 return; 623 } 624 625 // Support for multi-threaded concurrent phases 626 if (CMSConcurrentMTEnabled) { 627 if (FLAG_IS_DEFAULT(ConcGCThreads)) { 628 // just for now 629 FLAG_SET_DEFAULT(ConcGCThreads, (ParallelGCThreads + 3)/4); 630 } 631 if (ConcGCThreads > 1) { 632 _conc_workers = new YieldingFlexibleWorkGang("Parallel CMS Threads", 633 ConcGCThreads, true); 634 if (_conc_workers == NULL) { 635 warning("GC/CMS: _conc_workers allocation failure: " 636 "forcing -CMSConcurrentMTEnabled"); 637 CMSConcurrentMTEnabled = false; 638 } else { 639 _conc_workers->initialize_workers(); 640 } 641 } else { 642 CMSConcurrentMTEnabled = false; 643 } 644 } 645 if (!CMSConcurrentMTEnabled) { 646 ConcGCThreads = 0; 647 } else { 648 // Turn off CMSCleanOnEnter optimization temporarily for 649 // the MT case where it's not fixed yet; see 6178663. 650 CMSCleanOnEnter = false; 651 } 652 assert((_conc_workers != NULL) == (ConcGCThreads > 1), 653 "Inconsistency"); 654 655 // Parallel task queues; these are shared for the 656 // concurrent and stop-world phases of CMS, but 657 // are not shared with parallel scavenge (ParNew). 658 { 659 uint i; 660 uint num_queues = (uint) MAX2(ParallelGCThreads, ConcGCThreads); 661 662 if ((CMSParallelRemarkEnabled || CMSConcurrentMTEnabled 663 || ParallelRefProcEnabled) 664 && num_queues > 0) { 665 _task_queues = new OopTaskQueueSet(num_queues); 666 if (_task_queues == NULL) { 667 warning("task_queues allocation failure."); 668 return; 669 } 670 _hash_seed = NEW_C_HEAP_ARRAY(int, num_queues, mtGC); 671 if (_hash_seed == NULL) { 672 warning("_hash_seed array allocation failure"); 673 return; 674 } 675 676 typedef Padded<OopTaskQueue> PaddedOopTaskQueue; 677 for (i = 0; i < num_queues; i++) { 678 PaddedOopTaskQueue *q = new PaddedOopTaskQueue(); 679 if (q == NULL) { 680 warning("work_queue allocation failure."); 681 return; 682 } 683 _task_queues->register_queue(i, q); 684 } 685 for (i = 0; i < num_queues; i++) { 686 _task_queues->queue(i)->initialize(); 687 _hash_seed[i] = 17; // copied from ParNew 688 } 689 } 690 } 691 692 _cmsGen ->init_initiating_occupancy(CMSInitiatingOccupancyFraction, CMSTriggerRatio); 693 694 // Clip CMSBootstrapOccupancy between 0 and 100. 695 _bootstrap_occupancy = ((double)CMSBootstrapOccupancy)/(double)100; 696 697 _full_gcs_since_conc_gc = 0; 698 699 // Now tell CMS generations the identity of their collector 700 ConcurrentMarkSweepGeneration::set_collector(this); 701 702 // Create & start a CMS thread for this CMS collector 703 _cmsThread = ConcurrentMarkSweepThread::start(this); 704 assert(cmsThread() != NULL, "CMS Thread should have been created"); 705 assert(cmsThread()->collector() == this, 706 "CMS Thread should refer to this gen"); 707 assert(CGC_lock != NULL, "Where's the CGC_lock?"); 708 709 // Support for parallelizing young gen rescan 710 GenCollectedHeap* gch = GenCollectedHeap::heap(); 711 _young_gen = gch->prev_gen(_cmsGen); 712 if (gch->supports_inline_contig_alloc()) { 713 _top_addr = gch->top_addr(); 714 _end_addr = gch->end_addr(); 715 assert(_young_gen != NULL, "no _young_gen"); 716 _eden_chunk_index = 0; 717 _eden_chunk_capacity = (_young_gen->max_capacity()+CMSSamplingGrain)/CMSSamplingGrain; 718 _eden_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, _eden_chunk_capacity, mtGC); 719 if (_eden_chunk_array == NULL) { 720 _eden_chunk_capacity = 0; 721 warning("GC/CMS: _eden_chunk_array allocation failure"); 722 } 723 } 724 assert(_eden_chunk_array != NULL || _eden_chunk_capacity == 0, "Error"); 725 726 // Support for parallelizing survivor space rescan 727 if (CMSParallelRemarkEnabled && CMSParallelSurvivorRemarkEnabled) { 728 const size_t max_plab_samples = 729 ((DefNewGeneration*)_young_gen)->max_survivor_size()/MinTLABSize; 730 731 _survivor_plab_array = NEW_C_HEAP_ARRAY(ChunkArray, ParallelGCThreads, mtGC); 732 _survivor_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, 2*max_plab_samples, mtGC); 733 _cursor = NEW_C_HEAP_ARRAY(size_t, ParallelGCThreads, mtGC); 734 if (_survivor_plab_array == NULL || _survivor_chunk_array == NULL 735 || _cursor == NULL) { 736 warning("Failed to allocate survivor plab/chunk array"); 737 if (_survivor_plab_array != NULL) { 738 FREE_C_HEAP_ARRAY(ChunkArray, _survivor_plab_array, mtGC); 739 _survivor_plab_array = NULL; 740 } 741 if (_survivor_chunk_array != NULL) { 742 FREE_C_HEAP_ARRAY(HeapWord*, _survivor_chunk_array, mtGC); 743 _survivor_chunk_array = NULL; 744 } 745 if (_cursor != NULL) { 746 FREE_C_HEAP_ARRAY(size_t, _cursor, mtGC); 747 _cursor = NULL; 748 } 749 } else { 750 _survivor_chunk_capacity = 2*max_plab_samples; 751 for (uint i = 0; i < ParallelGCThreads; i++) { 752 HeapWord** vec = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples, mtGC); 753 if (vec == NULL) { 754 warning("Failed to allocate survivor plab array"); 755 for (int j = i; j > 0; j--) { 756 FREE_C_HEAP_ARRAY(HeapWord*, _survivor_plab_array[j-1].array(), mtGC); 757 } 758 FREE_C_HEAP_ARRAY(ChunkArray, _survivor_plab_array, mtGC); 759 FREE_C_HEAP_ARRAY(HeapWord*, _survivor_chunk_array, mtGC); 760 _survivor_plab_array = NULL; 761 _survivor_chunk_array = NULL; 762 _survivor_chunk_capacity = 0; 763 break; 764 } else { 765 ChunkArray* cur = 766 ::new (&_survivor_plab_array[i]) ChunkArray(vec, 767 max_plab_samples); 768 assert(cur->end() == 0, "Should be 0"); 769 assert(cur->array() == vec, "Should be vec"); 770 assert(cur->capacity() == max_plab_samples, "Error"); 771 } 772 } 773 } 774 } 775 assert( ( _survivor_plab_array != NULL 776 && _survivor_chunk_array != NULL) 777 || ( _survivor_chunk_capacity == 0 778 && _survivor_chunk_index == 0), 779 "Error"); 780 781 // Choose what strong roots should be scanned depending on verification options 782 if (!CMSClassUnloadingEnabled) { 783 // If class unloading is disabled we want to include all classes into the root set. 784 add_root_scanning_option(SharedHeap::SO_AllClasses); 785 } else { 786 add_root_scanning_option(SharedHeap::SO_SystemClasses); 787 } 788 789 NOT_PRODUCT(_overflow_counter = CMSMarkStackOverflowInterval;) 790 _gc_counters = new CollectorCounters("CMS", 1); 791 _completed_initialization = true; 792 _inter_sweep_timer.start(); // start of time 793 } 794 795 const char* ConcurrentMarkSweepGeneration::name() const { 796 return "concurrent mark-sweep generation"; 797 } 798 void ConcurrentMarkSweepGeneration::update_counters() { 799 if (UsePerfData) { 800 _space_counters->update_all(); 801 _gen_counters->update_all(); 802 } 803 } 804 805 // this is an optimized version of update_counters(). it takes the 806 // used value as a parameter rather than computing it. 807 // 808 void ConcurrentMarkSweepGeneration::update_counters(size_t used) { 809 if (UsePerfData) { 810 _space_counters->update_used(used); 811 _space_counters->update_capacity(); 812 _gen_counters->update_all(); 813 } 814 } 815 816 void ConcurrentMarkSweepGeneration::print() const { 817 Generation::print(); 818 cmsSpace()->print(); 819 } 820 821 #ifndef PRODUCT 822 void ConcurrentMarkSweepGeneration::print_statistics() { 823 cmsSpace()->printFLCensus(0); 824 } 825 #endif 826 827 void ConcurrentMarkSweepGeneration::printOccupancy(const char *s) { 828 GenCollectedHeap* gch = GenCollectedHeap::heap(); 829 if (PrintGCDetails) { 830 if (Verbose) { 831 gclog_or_tty->print("[%d %s-%s: "SIZE_FORMAT"("SIZE_FORMAT")]", 832 level(), short_name(), s, used(), capacity()); 833 } else { 834 gclog_or_tty->print("[%d %s-%s: "SIZE_FORMAT"K("SIZE_FORMAT"K)]", 835 level(), short_name(), s, used() / K, capacity() / K); 836 } 837 } 838 if (Verbose) { 839 gclog_or_tty->print(" "SIZE_FORMAT"("SIZE_FORMAT")", 840 gch->used(), gch->capacity()); 841 } else { 842 gclog_or_tty->print(" "SIZE_FORMAT"K("SIZE_FORMAT"K)", 843 gch->used() / K, gch->capacity() / K); 844 } 845 } 846 847 size_t 848 ConcurrentMarkSweepGeneration::contiguous_available() const { 849 // dld proposes an improvement in precision here. If the committed 850 // part of the space ends in a free block we should add that to 851 // uncommitted size in the calculation below. Will make this 852 // change later, staying with the approximation below for the 853 // time being. -- ysr. 854 return MAX2(_virtual_space.uncommitted_size(), unsafe_max_alloc_nogc()); 855 } 856 857 size_t 858 ConcurrentMarkSweepGeneration::unsafe_max_alloc_nogc() const { 859 return _cmsSpace->max_alloc_in_words() * HeapWordSize; 860 } 861 862 size_t ConcurrentMarkSweepGeneration::max_available() const { 863 return free() + _virtual_space.uncommitted_size(); 864 } 865 866 bool ConcurrentMarkSweepGeneration::promotion_attempt_is_safe(size_t max_promotion_in_bytes) const { 867 size_t available = max_available(); 868 size_t av_promo = (size_t)gc_stats()->avg_promoted()->padded_average(); 869 bool res = (available >= av_promo) || (available >= max_promotion_in_bytes); 870 if (Verbose && PrintGCDetails) { 871 gclog_or_tty->print_cr( 872 "CMS: promo attempt is%s safe: available("SIZE_FORMAT") %s av_promo("SIZE_FORMAT")," 873 "max_promo("SIZE_FORMAT")", 874 res? "":" not", available, res? ">=":"<", 875 av_promo, max_promotion_in_bytes); 876 } 877 return res; 878 } 879 880 // At a promotion failure dump information on block layout in heap 881 // (cms old generation). 882 void ConcurrentMarkSweepGeneration::promotion_failure_occurred() { 883 if (CMSDumpAtPromotionFailure) { 884 cmsSpace()->dump_at_safepoint_with_locks(collector(), gclog_or_tty); 885 } 886 } 887 888 CompactibleSpace* 889 ConcurrentMarkSweepGeneration::first_compaction_space() const { 890 return _cmsSpace; 891 } 892 893 void ConcurrentMarkSweepGeneration::reset_after_compaction() { 894 // Clear the promotion information. These pointers can be adjusted 895 // along with all the other pointers into the heap but 896 // compaction is expected to be a rare event with 897 // a heap using cms so don't do it without seeing the need. 898 if (CollectedHeap::use_parallel_gc_threads()) { 899 for (uint i = 0; i < ParallelGCThreads; i++) { 900 _par_gc_thread_states[i]->promo.reset(); 901 } 902 } 903 } 904 905 void ConcurrentMarkSweepGeneration::space_iterate(SpaceClosure* blk, bool usedOnly) { 906 blk->do_space(_cmsSpace); 907 } 908 909 void ConcurrentMarkSweepGeneration::compute_new_size() { 910 assert_locked_or_safepoint(Heap_lock); 911 912 // If incremental collection failed, we just want to expand 913 // to the limit. 914 if (incremental_collection_failed()) { 915 clear_incremental_collection_failed(); 916 grow_to_reserved(); 917 return; 918 } 919 920 // The heap has been compacted but not reset yet. 921 // Any metric such as free() or used() will be incorrect. 922 923 CardGeneration::compute_new_size(); 924 925 // Reset again after a possible resizing 926 if (did_compact()) { 927 cmsSpace()->reset_after_compaction(); 928 } 929 } 930 931 void ConcurrentMarkSweepGeneration::compute_new_size_free_list() { 932 assert_locked_or_safepoint(Heap_lock); 933 934 // If incremental collection failed, we just want to expand 935 // to the limit. 936 if (incremental_collection_failed()) { 937 clear_incremental_collection_failed(); 938 grow_to_reserved(); 939 return; 940 } 941 942 double free_percentage = ((double) free()) / capacity(); 943 double desired_free_percentage = (double) MinHeapFreeRatio / 100; 944 double maximum_free_percentage = (double) MaxHeapFreeRatio / 100; 945 946 // compute expansion delta needed for reaching desired free percentage 947 if (free_percentage < desired_free_percentage) { 948 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 949 assert(desired_capacity >= capacity(), "invalid expansion size"); 950 size_t expand_bytes = MAX2(desired_capacity - capacity(), MinHeapDeltaBytes); 951 if (PrintGCDetails && Verbose) { 952 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 953 gclog_or_tty->print_cr("\nFrom compute_new_size: "); 954 gclog_or_tty->print_cr(" Free fraction %f", free_percentage); 955 gclog_or_tty->print_cr(" Desired free fraction %f", 956 desired_free_percentage); 957 gclog_or_tty->print_cr(" Maximum free fraction %f", 958 maximum_free_percentage); 959 gclog_or_tty->print_cr(" Capactiy "SIZE_FORMAT, capacity()/1000); 960 gclog_or_tty->print_cr(" Desired capacity "SIZE_FORMAT, 961 desired_capacity/1000); 962 int prev_level = level() - 1; 963 if (prev_level >= 0) { 964 size_t prev_size = 0; 965 GenCollectedHeap* gch = GenCollectedHeap::heap(); 966 Generation* prev_gen = gch->_gens[prev_level]; 967 prev_size = prev_gen->capacity(); 968 gclog_or_tty->print_cr(" Younger gen size "SIZE_FORMAT, 969 prev_size/1000); 970 } 971 gclog_or_tty->print_cr(" unsafe_max_alloc_nogc "SIZE_FORMAT, 972 unsafe_max_alloc_nogc()/1000); 973 gclog_or_tty->print_cr(" contiguous available "SIZE_FORMAT, 974 contiguous_available()/1000); 975 gclog_or_tty->print_cr(" Expand by "SIZE_FORMAT" (bytes)", 976 expand_bytes); 977 } 978 // safe if expansion fails 979 expand(expand_bytes, 0, CMSExpansionCause::_satisfy_free_ratio); 980 if (PrintGCDetails && Verbose) { 981 gclog_or_tty->print_cr(" Expanded free fraction %f", 982 ((double) free()) / capacity()); 983 } 984 } else { 985 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 986 assert(desired_capacity <= capacity(), "invalid expansion size"); 987 size_t shrink_bytes = capacity() - desired_capacity; 988 // Don't shrink unless the delta is greater than the minimum shrink we want 989 if (shrink_bytes >= MinHeapDeltaBytes) { 990 shrink_free_list_by(shrink_bytes); 991 } 992 } 993 } 994 995 Mutex* ConcurrentMarkSweepGeneration::freelistLock() const { 996 return cmsSpace()->freelistLock(); 997 } 998 999 HeapWord* ConcurrentMarkSweepGeneration::allocate(size_t size, 1000 bool tlab) { 1001 CMSSynchronousYieldRequest yr; 1002 MutexLockerEx x(freelistLock(), 1003 Mutex::_no_safepoint_check_flag); 1004 return have_lock_and_allocate(size, tlab); 1005 } 1006 1007 HeapWord* ConcurrentMarkSweepGeneration::have_lock_and_allocate(size_t size, 1008 bool tlab /* ignored */) { 1009 assert_lock_strong(freelistLock()); 1010 size_t adjustedSize = CompactibleFreeListSpace::adjustObjectSize(size); 1011 HeapWord* res = cmsSpace()->allocate(adjustedSize); 1012 // Allocate the object live (grey) if the background collector has 1013 // started marking. This is necessary because the marker may 1014 // have passed this address and consequently this object will 1015 // not otherwise be greyed and would be incorrectly swept up. 1016 // Note that if this object contains references, the writing 1017 // of those references will dirty the card containing this object 1018 // allowing the object to be blackened (and its references scanned) 1019 // either during a preclean phase or at the final checkpoint. 1020 if (res != NULL) { 1021 // We may block here with an uninitialized object with 1022 // its mark-bit or P-bits not yet set. Such objects need 1023 // to be safely navigable by block_start(). 1024 assert(oop(res)->klass_or_null() == NULL, "Object should be uninitialized here."); 1025 assert(!((FreeChunk*)res)->is_free(), "Error, block will look free but show wrong size"); 1026 collector()->direct_allocated(res, adjustedSize); 1027 _direct_allocated_words += adjustedSize; 1028 // allocation counters 1029 NOT_PRODUCT( 1030 _numObjectsAllocated++; 1031 _numWordsAllocated += (int)adjustedSize; 1032 ) 1033 } 1034 return res; 1035 } 1036 1037 // In the case of direct allocation by mutators in a generation that 1038 // is being concurrently collected, the object must be allocated 1039 // live (grey) if the background collector has started marking. 1040 // This is necessary because the marker may 1041 // have passed this address and consequently this object will 1042 // not otherwise be greyed and would be incorrectly swept up. 1043 // Note that if this object contains references, the writing 1044 // of those references will dirty the card containing this object 1045 // allowing the object to be blackened (and its references scanned) 1046 // either during a preclean phase or at the final checkpoint. 1047 void CMSCollector::direct_allocated(HeapWord* start, size_t size) { 1048 assert(_markBitMap.covers(start, size), "Out of bounds"); 1049 if (_collectorState >= Marking) { 1050 MutexLockerEx y(_markBitMap.lock(), 1051 Mutex::_no_safepoint_check_flag); 1052 // [see comments preceding SweepClosure::do_blk() below for details] 1053 // 1054 // Can the P-bits be deleted now? JJJ 1055 // 1056 // 1. need to mark the object as live so it isn't collected 1057 // 2. need to mark the 2nd bit to indicate the object may be uninitialized 1058 // 3. need to mark the end of the object so marking, precleaning or sweeping 1059 // can skip over uninitialized or unparsable objects. An allocated 1060 // object is considered uninitialized for our purposes as long as 1061 // its klass word is NULL. All old gen objects are parsable 1062 // as soon as they are initialized.) 1063 _markBitMap.mark(start); // object is live 1064 _markBitMap.mark(start + 1); // object is potentially uninitialized? 1065 _markBitMap.mark(start + size - 1); 1066 // mark end of object 1067 } 1068 // check that oop looks uninitialized 1069 assert(oop(start)->klass_or_null() == NULL, "_klass should be NULL"); 1070 } 1071 1072 void CMSCollector::promoted(bool par, HeapWord* start, 1073 bool is_obj_array, size_t obj_size) { 1074 assert(_markBitMap.covers(start), "Out of bounds"); 1075 // See comment in direct_allocated() about when objects should 1076 // be allocated live. 1077 if (_collectorState >= Marking) { 1078 // we already hold the marking bit map lock, taken in 1079 // the prologue 1080 if (par) { 1081 _markBitMap.par_mark(start); 1082 } else { 1083 _markBitMap.mark(start); 1084 } 1085 // We don't need to mark the object as uninitialized (as 1086 // in direct_allocated above) because this is being done with the 1087 // world stopped and the object will be initialized by the 1088 // time the marking, precleaning or sweeping get to look at it. 1089 // But see the code for copying objects into the CMS generation, 1090 // where we need to ensure that concurrent readers of the 1091 // block offset table are able to safely navigate a block that 1092 // is in flux from being free to being allocated (and in 1093 // transition while being copied into) and subsequently 1094 // becoming a bona-fide object when the copy/promotion is complete. 1095 assert(SafepointSynchronize::is_at_safepoint(), 1096 "expect promotion only at safepoints"); 1097 1098 if (_collectorState < Sweeping) { 1099 // Mark the appropriate cards in the modUnionTable, so that 1100 // this object gets scanned before the sweep. If this is 1101 // not done, CMS generation references in the object might 1102 // not get marked. 1103 // For the case of arrays, which are otherwise precisely 1104 // marked, we need to dirty the entire array, not just its head. 1105 if (is_obj_array) { 1106 // The [par_]mark_range() method expects mr.end() below to 1107 // be aligned to the granularity of a bit's representation 1108 // in the heap. In the case of the MUT below, that's a 1109 // card size. 1110 MemRegion mr(start, 1111 (HeapWord*)round_to((intptr_t)(start + obj_size), 1112 CardTableModRefBS::card_size /* bytes */)); 1113 if (par) { 1114 _modUnionTable.par_mark_range(mr); 1115 } else { 1116 _modUnionTable.mark_range(mr); 1117 } 1118 } else { // not an obj array; we can just mark the head 1119 if (par) { 1120 _modUnionTable.par_mark(start); 1121 } else { 1122 _modUnionTable.mark(start); 1123 } 1124 } 1125 } 1126 } 1127 } 1128 1129 static inline size_t percent_of_space(Space* space, HeapWord* addr) 1130 { 1131 size_t delta = pointer_delta(addr, space->bottom()); 1132 return (size_t)(delta * 100.0 / (space->capacity() / HeapWordSize)); 1133 } 1134 1135 void CMSCollector::icms_update_allocation_limits() 1136 { 1137 Generation* gen0 = GenCollectedHeap::heap()->get_gen(0); 1138 EdenSpace* eden = gen0->as_DefNewGeneration()->eden(); 1139 1140 const unsigned int duty_cycle = stats().icms_update_duty_cycle(); 1141 if (CMSTraceIncrementalPacing) { 1142 stats().print(); 1143 } 1144 1145 assert(duty_cycle <= 100, "invalid duty cycle"); 1146 if (duty_cycle != 0) { 1147 // The duty_cycle is a percentage between 0 and 100; convert to words and 1148 // then compute the offset from the endpoints of the space. 1149 size_t free_words = eden->free() / HeapWordSize; 1150 double free_words_dbl = (double)free_words; 1151 size_t duty_cycle_words = (size_t)(free_words_dbl * duty_cycle / 100.0); 1152 size_t offset_words = (free_words - duty_cycle_words) / 2; 1153 1154 _icms_start_limit = eden->top() + offset_words; 1155 _icms_stop_limit = eden->end() - offset_words; 1156 1157 // The limits may be adjusted (shifted to the right) by 1158 // CMSIncrementalOffset, to allow the application more mutator time after a 1159 // young gen gc (when all mutators were stopped) and before CMS starts and 1160 // takes away one or more cpus. 1161 if (CMSIncrementalOffset != 0) { 1162 double adjustment_dbl = free_words_dbl * CMSIncrementalOffset / 100.0; 1163 size_t adjustment = (size_t)adjustment_dbl; 1164 HeapWord* tmp_stop = _icms_stop_limit + adjustment; 1165 if (tmp_stop > _icms_stop_limit && tmp_stop < eden->end()) { 1166 _icms_start_limit += adjustment; 1167 _icms_stop_limit = tmp_stop; 1168 } 1169 } 1170 } 1171 if (duty_cycle == 0 || (_icms_start_limit == _icms_stop_limit)) { 1172 _icms_start_limit = _icms_stop_limit = eden->end(); 1173 } 1174 1175 // Install the new start limit. 1176 eden->set_soft_end(_icms_start_limit); 1177 1178 if (CMSTraceIncrementalMode) { 1179 gclog_or_tty->print(" icms alloc limits: " 1180 PTR_FORMAT "," PTR_FORMAT 1181 " (" SIZE_FORMAT "%%," SIZE_FORMAT "%%) ", 1182 _icms_start_limit, _icms_stop_limit, 1183 percent_of_space(eden, _icms_start_limit), 1184 percent_of_space(eden, _icms_stop_limit)); 1185 if (Verbose) { 1186 gclog_or_tty->print("eden: "); 1187 eden->print_on(gclog_or_tty); 1188 } 1189 } 1190 } 1191 1192 // Any changes here should try to maintain the invariant 1193 // that if this method is called with _icms_start_limit 1194 // and _icms_stop_limit both NULL, then it should return NULL 1195 // and not notify the icms thread. 1196 HeapWord* 1197 CMSCollector::allocation_limit_reached(Space* space, HeapWord* top, 1198 size_t word_size) 1199 { 1200 // A start_limit equal to end() means the duty cycle is 0, so treat that as a 1201 // nop. 1202 if (CMSIncrementalMode && _icms_start_limit != space->end()) { 1203 if (top <= _icms_start_limit) { 1204 if (CMSTraceIncrementalMode) { 1205 space->print_on(gclog_or_tty); 1206 gclog_or_tty->stamp(); 1207 gclog_or_tty->print_cr(" start limit top=" PTR_FORMAT 1208 ", new limit=" PTR_FORMAT 1209 " (" SIZE_FORMAT "%%)", 1210 top, _icms_stop_limit, 1211 percent_of_space(space, _icms_stop_limit)); 1212 } 1213 ConcurrentMarkSweepThread::start_icms(); 1214 assert(top < _icms_stop_limit, "Tautology"); 1215 if (word_size < pointer_delta(_icms_stop_limit, top)) { 1216 return _icms_stop_limit; 1217 } 1218 1219 // The allocation will cross both the _start and _stop limits, so do the 1220 // stop notification also and return end(). 1221 if (CMSTraceIncrementalMode) { 1222 space->print_on(gclog_or_tty); 1223 gclog_or_tty->stamp(); 1224 gclog_or_tty->print_cr(" +stop limit top=" PTR_FORMAT 1225 ", new limit=" PTR_FORMAT 1226 " (" SIZE_FORMAT "%%)", 1227 top, space->end(), 1228 percent_of_space(space, space->end())); 1229 } 1230 ConcurrentMarkSweepThread::stop_icms(); 1231 return space->end(); 1232 } 1233 1234 if (top <= _icms_stop_limit) { 1235 if (CMSTraceIncrementalMode) { 1236 space->print_on(gclog_or_tty); 1237 gclog_or_tty->stamp(); 1238 gclog_or_tty->print_cr(" stop limit top=" PTR_FORMAT 1239 ", new limit=" PTR_FORMAT 1240 " (" SIZE_FORMAT "%%)", 1241 top, space->end(), 1242 percent_of_space(space, space->end())); 1243 } 1244 ConcurrentMarkSweepThread::stop_icms(); 1245 return space->end(); 1246 } 1247 1248 if (CMSTraceIncrementalMode) { 1249 space->print_on(gclog_or_tty); 1250 gclog_or_tty->stamp(); 1251 gclog_or_tty->print_cr(" end limit top=" PTR_FORMAT 1252 ", new limit=" PTR_FORMAT, 1253 top, NULL); 1254 } 1255 } 1256 1257 return NULL; 1258 } 1259 1260 oop ConcurrentMarkSweepGeneration::promote(oop obj, size_t obj_size) { 1261 assert(obj_size == (size_t)obj->size(), "bad obj_size passed in"); 1262 // allocate, copy and if necessary update promoinfo -- 1263 // delegate to underlying space. 1264 assert_lock_strong(freelistLock()); 1265 1266 #ifndef PRODUCT 1267 if (Universe::heap()->promotion_should_fail()) { 1268 return NULL; 1269 } 1270 #endif // #ifndef PRODUCT 1271 1272 oop res = _cmsSpace->promote(obj, obj_size); 1273 if (res == NULL) { 1274 // expand and retry 1275 size_t s = _cmsSpace->expansionSpaceRequired(obj_size); // HeapWords 1276 expand(s*HeapWordSize, MinHeapDeltaBytes, 1277 CMSExpansionCause::_satisfy_promotion); 1278 // Since there's currently no next generation, we don't try to promote 1279 // into a more senior generation. 1280 assert(next_gen() == NULL, "assumption, based upon which no attempt " 1281 "is made to pass on a possibly failing " 1282 "promotion to next generation"); 1283 res = _cmsSpace->promote(obj, obj_size); 1284 } 1285 if (res != NULL) { 1286 // See comment in allocate() about when objects should 1287 // be allocated live. 1288 assert(obj->is_oop(), "Will dereference klass pointer below"); 1289 collector()->promoted(false, // Not parallel 1290 (HeapWord*)res, obj->is_objArray(), obj_size); 1291 // promotion counters 1292 NOT_PRODUCT( 1293 _numObjectsPromoted++; 1294 _numWordsPromoted += 1295 (int)(CompactibleFreeListSpace::adjustObjectSize(obj->size())); 1296 ) 1297 } 1298 return res; 1299 } 1300 1301 1302 HeapWord* 1303 ConcurrentMarkSweepGeneration::allocation_limit_reached(Space* space, 1304 HeapWord* top, 1305 size_t word_sz) 1306 { 1307 return collector()->allocation_limit_reached(space, top, word_sz); 1308 } 1309 1310 // IMPORTANT: Notes on object size recognition in CMS. 1311 // --------------------------------------------------- 1312 // A block of storage in the CMS generation is always in 1313 // one of three states. A free block (FREE), an allocated 1314 // object (OBJECT) whose size() method reports the correct size, 1315 // and an intermediate state (TRANSIENT) in which its size cannot 1316 // be accurately determined. 1317 // STATE IDENTIFICATION: (32 bit and 64 bit w/o COOPS) 1318 // ----------------------------------------------------- 1319 // FREE: klass_word & 1 == 1; mark_word holds block size 1320 // 1321 // OBJECT: klass_word installed; klass_word != 0 && klass_word & 1 == 0; 1322 // obj->size() computes correct size 1323 // 1324 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT 1325 // 1326 // STATE IDENTIFICATION: (64 bit+COOPS) 1327 // ------------------------------------ 1328 // FREE: mark_word & CMS_FREE_BIT == 1; mark_word & ~CMS_FREE_BIT gives block_size 1329 // 1330 // OBJECT: klass_word installed; klass_word != 0; 1331 // obj->size() computes correct size 1332 // 1333 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT 1334 // 1335 // 1336 // STATE TRANSITION DIAGRAM 1337 // 1338 // mut / parnew mut / parnew 1339 // FREE --------------------> TRANSIENT ---------------------> OBJECT --| 1340 // ^ | 1341 // |------------------------ DEAD <------------------------------------| 1342 // sweep mut 1343 // 1344 // While a block is in TRANSIENT state its size cannot be determined 1345 // so readers will either need to come back later or stall until 1346 // the size can be determined. Note that for the case of direct 1347 // allocation, P-bits, when available, may be used to determine the 1348 // size of an object that may not yet have been initialized. 1349 1350 // Things to support parallel young-gen collection. 1351 oop 1352 ConcurrentMarkSweepGeneration::par_promote(int thread_num, 1353 oop old, markOop m, 1354 size_t word_sz) { 1355 #ifndef PRODUCT 1356 if (Universe::heap()->promotion_should_fail()) { 1357 return NULL; 1358 } 1359 #endif // #ifndef PRODUCT 1360 1361 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1362 PromotionInfo* promoInfo = &ps->promo; 1363 // if we are tracking promotions, then first ensure space for 1364 // promotion (including spooling space for saving header if necessary). 1365 // then allocate and copy, then track promoted info if needed. 1366 // When tracking (see PromotionInfo::track()), the mark word may 1367 // be displaced and in this case restoration of the mark word 1368 // occurs in the (oop_since_save_marks_)iterate phase. 1369 if (promoInfo->tracking() && !promoInfo->ensure_spooling_space()) { 1370 // Out of space for allocating spooling buffers; 1371 // try expanding and allocating spooling buffers. 1372 if (!expand_and_ensure_spooling_space(promoInfo)) { 1373 return NULL; 1374 } 1375 } 1376 assert(promoInfo->has_spooling_space(), "Control point invariant"); 1377 const size_t alloc_sz = CompactibleFreeListSpace::adjustObjectSize(word_sz); 1378 HeapWord* obj_ptr = ps->lab.alloc(alloc_sz); 1379 if (obj_ptr == NULL) { 1380 obj_ptr = expand_and_par_lab_allocate(ps, alloc_sz); 1381 if (obj_ptr == NULL) { 1382 return NULL; 1383 } 1384 } 1385 oop obj = oop(obj_ptr); 1386 OrderAccess::storestore(); 1387 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1388 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1389 // IMPORTANT: See note on object initialization for CMS above. 1390 // Otherwise, copy the object. Here we must be careful to insert the 1391 // klass pointer last, since this marks the block as an allocated object. 1392 // Except with compressed oops it's the mark word. 1393 HeapWord* old_ptr = (HeapWord*)old; 1394 // Restore the mark word copied above. 1395 obj->set_mark(m); 1396 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1397 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1398 OrderAccess::storestore(); 1399 1400 if (UseCompressedKlassPointers) { 1401 // Copy gap missed by (aligned) header size calculation below 1402 obj->set_klass_gap(old->klass_gap()); 1403 } 1404 if (word_sz > (size_t)oopDesc::header_size()) { 1405 Copy::aligned_disjoint_words(old_ptr + oopDesc::header_size(), 1406 obj_ptr + oopDesc::header_size(), 1407 word_sz - oopDesc::header_size()); 1408 } 1409 1410 // Now we can track the promoted object, if necessary. We take care 1411 // to delay the transition from uninitialized to full object 1412 // (i.e., insertion of klass pointer) until after, so that it 1413 // atomically becomes a promoted object. 1414 if (promoInfo->tracking()) { 1415 promoInfo->track((PromotedObject*)obj, old->klass()); 1416 } 1417 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1418 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1419 assert(old->is_oop(), "Will use and dereference old klass ptr below"); 1420 1421 // Finally, install the klass pointer (this should be volatile). 1422 OrderAccess::storestore(); 1423 obj->set_klass(old->klass()); 1424 // We should now be able to calculate the right size for this object 1425 assert(obj->is_oop() && obj->size() == (int)word_sz, "Error, incorrect size computed for promoted object"); 1426 1427 collector()->promoted(true, // parallel 1428 obj_ptr, old->is_objArray(), word_sz); 1429 1430 NOT_PRODUCT( 1431 Atomic::inc_ptr(&_numObjectsPromoted); 1432 Atomic::add_ptr(alloc_sz, &_numWordsPromoted); 1433 ) 1434 1435 return obj; 1436 } 1437 1438 void 1439 ConcurrentMarkSweepGeneration:: 1440 par_promote_alloc_undo(int thread_num, 1441 HeapWord* obj, size_t word_sz) { 1442 // CMS does not support promotion undo. 1443 ShouldNotReachHere(); 1444 } 1445 1446 void 1447 ConcurrentMarkSweepGeneration:: 1448 par_promote_alloc_done(int thread_num) { 1449 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1450 ps->lab.retire(thread_num); 1451 } 1452 1453 void 1454 ConcurrentMarkSweepGeneration:: 1455 par_oop_since_save_marks_iterate_done(int thread_num) { 1456 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1457 ParScanWithoutBarrierClosure* dummy_cl = NULL; 1458 ps->promo.promoted_oops_iterate_nv(dummy_cl); 1459 } 1460 1461 bool ConcurrentMarkSweepGeneration::should_collect(bool full, 1462 size_t size, 1463 bool tlab) 1464 { 1465 // We allow a STW collection only if a full 1466 // collection was requested. 1467 return full || should_allocate(size, tlab); // FIX ME !!! 1468 // This and promotion failure handling are connected at the 1469 // hip and should be fixed by untying them. 1470 } 1471 1472 bool CMSCollector::shouldConcurrentCollect() { 1473 if (_full_gc_requested) { 1474 if (Verbose && PrintGCDetails) { 1475 gclog_or_tty->print_cr("CMSCollector: collect because of explicit " 1476 " gc request (or gc_locker)"); 1477 } 1478 return true; 1479 } 1480 1481 // For debugging purposes, change the type of collection. 1482 // If the rotation is not on the concurrent collection 1483 // type, don't start a concurrent collection. 1484 NOT_PRODUCT( 1485 if (RotateCMSCollectionTypes && 1486 (_cmsGen->debug_collection_type() != 1487 ConcurrentMarkSweepGeneration::Concurrent_collection_type)) { 1488 assert(_cmsGen->debug_collection_type() != 1489 ConcurrentMarkSweepGeneration::Unknown_collection_type, 1490 "Bad cms collection type"); 1491 return false; 1492 } 1493 ) 1494 1495 FreelistLocker x(this); 1496 // ------------------------------------------------------------------ 1497 // Print out lots of information which affects the initiation of 1498 // a collection. 1499 if (PrintCMSInitiationStatistics && stats().valid()) { 1500 gclog_or_tty->print("CMSCollector shouldConcurrentCollect: "); 1501 gclog_or_tty->stamp(); 1502 gclog_or_tty->print_cr(""); 1503 stats().print_on(gclog_or_tty); 1504 gclog_or_tty->print_cr("time_until_cms_gen_full %3.7f", 1505 stats().time_until_cms_gen_full()); 1506 gclog_or_tty->print_cr("free="SIZE_FORMAT, _cmsGen->free()); 1507 gclog_or_tty->print_cr("contiguous_available="SIZE_FORMAT, 1508 _cmsGen->contiguous_available()); 1509 gclog_or_tty->print_cr("promotion_rate=%g", stats().promotion_rate()); 1510 gclog_or_tty->print_cr("cms_allocation_rate=%g", stats().cms_allocation_rate()); 1511 gclog_or_tty->print_cr("occupancy=%3.7f", _cmsGen->occupancy()); 1512 gclog_or_tty->print_cr("initiatingOccupancy=%3.7f", _cmsGen->initiating_occupancy()); 1513 gclog_or_tty->print_cr("metadata initialized %d", 1514 MetaspaceGC::should_concurrent_collect()); 1515 } 1516 // ------------------------------------------------------------------ 1517 1518 // If the estimated time to complete a cms collection (cms_duration()) 1519 // is less than the estimated time remaining until the cms generation 1520 // is full, start a collection. 1521 if (!UseCMSInitiatingOccupancyOnly) { 1522 if (stats().valid()) { 1523 if (stats().time_until_cms_start() == 0.0) { 1524 return true; 1525 } 1526 } else { 1527 // We want to conservatively collect somewhat early in order 1528 // to try and "bootstrap" our CMS/promotion statistics; 1529 // this branch will not fire after the first successful CMS 1530 // collection because the stats should then be valid. 1531 if (_cmsGen->occupancy() >= _bootstrap_occupancy) { 1532 if (Verbose && PrintGCDetails) { 1533 gclog_or_tty->print_cr( 1534 " CMSCollector: collect for bootstrapping statistics:" 1535 " occupancy = %f, boot occupancy = %f", _cmsGen->occupancy(), 1536 _bootstrap_occupancy); 1537 } 1538 return true; 1539 } 1540 } 1541 } 1542 1543 // Otherwise, we start a collection cycle if 1544 // old gen want a collection cycle started. Each may use 1545 // an appropriate criterion for making this decision. 1546 // XXX We need to make sure that the gen expansion 1547 // criterion dovetails well with this. XXX NEED TO FIX THIS 1548 if (_cmsGen->should_concurrent_collect()) { 1549 if (Verbose && PrintGCDetails) { 1550 gclog_or_tty->print_cr("CMS old gen initiated"); 1551 } 1552 return true; 1553 } 1554 1555 // We start a collection if we believe an incremental collection may fail; 1556 // this is not likely to be productive in practice because it's probably too 1557 // late anyway. 1558 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1559 assert(gch->collector_policy()->is_two_generation_policy(), 1560 "You may want to check the correctness of the following"); 1561 if (gch->incremental_collection_will_fail(true /* consult_young */)) { 1562 if (Verbose && PrintGCDetails) { 1563 gclog_or_tty->print("CMSCollector: collect because incremental collection will fail "); 1564 } 1565 return true; 1566 } 1567 1568 if (MetaspaceGC::should_concurrent_collect()) { 1569 if (Verbose && PrintGCDetails) { 1570 gclog_or_tty->print("CMSCollector: collect for metadata allocation "); 1571 } 1572 return true; 1573 } 1574 1575 return false; 1576 } 1577 1578 void CMSCollector::set_did_compact(bool v) { _cmsGen->set_did_compact(v); } 1579 1580 // Clear _expansion_cause fields of constituent generations 1581 void CMSCollector::clear_expansion_cause() { 1582 _cmsGen->clear_expansion_cause(); 1583 } 1584 1585 // We should be conservative in starting a collection cycle. To 1586 // start too eagerly runs the risk of collecting too often in the 1587 // extreme. To collect too rarely falls back on full collections, 1588 // which works, even if not optimum in terms of concurrent work. 1589 // As a work around for too eagerly collecting, use the flag 1590 // UseCMSInitiatingOccupancyOnly. This also has the advantage of 1591 // giving the user an easily understandable way of controlling the 1592 // collections. 1593 // We want to start a new collection cycle if any of the following 1594 // conditions hold: 1595 // . our current occupancy exceeds the configured initiating occupancy 1596 // for this generation, or 1597 // . we recently needed to expand this space and have not, since that 1598 // expansion, done a collection of this generation, or 1599 // . the underlying space believes that it may be a good idea to initiate 1600 // a concurrent collection (this may be based on criteria such as the 1601 // following: the space uses linear allocation and linear allocation is 1602 // going to fail, or there is believed to be excessive fragmentation in 1603 // the generation, etc... or ... 1604 // [.(currently done by CMSCollector::shouldConcurrentCollect() only for 1605 // the case of the old generation; see CR 6543076): 1606 // we may be approaching a point at which allocation requests may fail because 1607 // we will be out of sufficient free space given allocation rate estimates.] 1608 bool ConcurrentMarkSweepGeneration::should_concurrent_collect() const { 1609 1610 assert_lock_strong(freelistLock()); 1611 if (occupancy() > initiating_occupancy()) { 1612 if (PrintGCDetails && Verbose) { 1613 gclog_or_tty->print(" %s: collect because of occupancy %f / %f ", 1614 short_name(), occupancy(), initiating_occupancy()); 1615 } 1616 return true; 1617 } 1618 if (UseCMSInitiatingOccupancyOnly) { 1619 return false; 1620 } 1621 if (expansion_cause() == CMSExpansionCause::_satisfy_allocation) { 1622 if (PrintGCDetails && Verbose) { 1623 gclog_or_tty->print(" %s: collect because expanded for allocation ", 1624 short_name()); 1625 } 1626 return true; 1627 } 1628 if (_cmsSpace->should_concurrent_collect()) { 1629 if (PrintGCDetails && Verbose) { 1630 gclog_or_tty->print(" %s: collect because cmsSpace says so ", 1631 short_name()); 1632 } 1633 return true; 1634 } 1635 return false; 1636 } 1637 1638 void ConcurrentMarkSweepGeneration::collect(bool full, 1639 bool clear_all_soft_refs, 1640 size_t size, 1641 bool tlab) 1642 { 1643 collector()->collect(full, clear_all_soft_refs, size, tlab); 1644 } 1645 1646 void CMSCollector::collect(bool full, 1647 bool clear_all_soft_refs, 1648 size_t size, 1649 bool tlab) 1650 { 1651 if (!UseCMSCollectionPassing && _collectorState > Idling) { 1652 // For debugging purposes skip the collection if the state 1653 // is not currently idle 1654 if (TraceCMSState) { 1655 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " skipped full:%d CMS state %d", 1656 Thread::current(), full, _collectorState); 1657 } 1658 return; 1659 } 1660 1661 // The following "if" branch is present for defensive reasons. 1662 // In the current uses of this interface, it can be replaced with: 1663 // assert(!GC_locker.is_active(), "Can't be called otherwise"); 1664 // But I am not placing that assert here to allow future 1665 // generality in invoking this interface. 1666 if (GC_locker::is_active()) { 1667 // A consistency test for GC_locker 1668 assert(GC_locker::needs_gc(), "Should have been set already"); 1669 // Skip this foreground collection, instead 1670 // expanding the heap if necessary. 1671 // Need the free list locks for the call to free() in compute_new_size() 1672 compute_new_size(); 1673 return; 1674 } 1675 acquire_control_and_collect(full, clear_all_soft_refs); 1676 _full_gcs_since_conc_gc++; 1677 } 1678 1679 void CMSCollector::request_full_gc(unsigned int full_gc_count) { 1680 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1681 unsigned int gc_count = gch->total_full_collections(); 1682 if (gc_count == full_gc_count) { 1683 MutexLockerEx y(CGC_lock, Mutex::_no_safepoint_check_flag); 1684 _full_gc_requested = true; 1685 CGC_lock->notify(); // nudge CMS thread 1686 } else { 1687 assert(gc_count > full_gc_count, "Error: causal loop"); 1688 } 1689 } 1690 1691 1692 // The foreground and background collectors need to coordinate in order 1693 // to make sure that they do not mutually interfere with CMS collections. 1694 // When a background collection is active, 1695 // the foreground collector may need to take over (preempt) and 1696 // synchronously complete an ongoing collection. Depending on the 1697 // frequency of the background collections and the heap usage 1698 // of the application, this preemption can be seldom or frequent. 1699 // There are only certain 1700 // points in the background collection that the "collection-baton" 1701 // can be passed to the foreground collector. 1702 // 1703 // The foreground collector will wait for the baton before 1704 // starting any part of the collection. The foreground collector 1705 // will only wait at one location. 1706 // 1707 // The background collector will yield the baton before starting a new 1708 // phase of the collection (e.g., before initial marking, marking from roots, 1709 // precleaning, final re-mark, sweep etc.) This is normally done at the head 1710 // of the loop which switches the phases. The background collector does some 1711 // of the phases (initial mark, final re-mark) with the world stopped. 1712 // Because of locking involved in stopping the world, 1713 // the foreground collector should not block waiting for the background 1714 // collector when it is doing a stop-the-world phase. The background 1715 // collector will yield the baton at an additional point just before 1716 // it enters a stop-the-world phase. Once the world is stopped, the 1717 // background collector checks the phase of the collection. If the 1718 // phase has not changed, it proceeds with the collection. If the 1719 // phase has changed, it skips that phase of the collection. See 1720 // the comments on the use of the Heap_lock in collect_in_background(). 1721 // 1722 // Variable used in baton passing. 1723 // _foregroundGCIsActive - Set to true by the foreground collector when 1724 // it wants the baton. The foreground clears it when it has finished 1725 // the collection. 1726 // _foregroundGCShouldWait - Set to true by the background collector 1727 // when it is running. The foreground collector waits while 1728 // _foregroundGCShouldWait is true. 1729 // CGC_lock - monitor used to protect access to the above variables 1730 // and to notify the foreground and background collectors. 1731 // _collectorState - current state of the CMS collection. 1732 // 1733 // The foreground collector 1734 // acquires the CGC_lock 1735 // sets _foregroundGCIsActive 1736 // waits on the CGC_lock for _foregroundGCShouldWait to be false 1737 // various locks acquired in preparation for the collection 1738 // are released so as not to block the background collector 1739 // that is in the midst of a collection 1740 // proceeds with the collection 1741 // clears _foregroundGCIsActive 1742 // returns 1743 // 1744 // The background collector in a loop iterating on the phases of the 1745 // collection 1746 // acquires the CGC_lock 1747 // sets _foregroundGCShouldWait 1748 // if _foregroundGCIsActive is set 1749 // clears _foregroundGCShouldWait, notifies _CGC_lock 1750 // waits on _CGC_lock for _foregroundGCIsActive to become false 1751 // and exits the loop. 1752 // otherwise 1753 // proceed with that phase of the collection 1754 // if the phase is a stop-the-world phase, 1755 // yield the baton once more just before enqueueing 1756 // the stop-world CMS operation (executed by the VM thread). 1757 // returns after all phases of the collection are done 1758 // 1759 1760 void CMSCollector::acquire_control_and_collect(bool full, 1761 bool clear_all_soft_refs) { 1762 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint"); 1763 assert(!Thread::current()->is_ConcurrentGC_thread(), 1764 "shouldn't try to acquire control from self!"); 1765 1766 // Start the protocol for acquiring control of the 1767 // collection from the background collector (aka CMS thread). 1768 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 1769 "VM thread should have CMS token"); 1770 // Remember the possibly interrupted state of an ongoing 1771 // concurrent collection 1772 CollectorState first_state = _collectorState; 1773 1774 // Signal to a possibly ongoing concurrent collection that 1775 // we want to do a foreground collection. 1776 _foregroundGCIsActive = true; 1777 1778 // Disable incremental mode during a foreground collection. 1779 ICMSDisabler icms_disabler; 1780 1781 // release locks and wait for a notify from the background collector 1782 // releasing the locks in only necessary for phases which 1783 // do yields to improve the granularity of the collection. 1784 assert_lock_strong(bitMapLock()); 1785 // We need to lock the Free list lock for the space that we are 1786 // currently collecting. 1787 assert(haveFreelistLocks(), "Must be holding free list locks"); 1788 bitMapLock()->unlock(); 1789 releaseFreelistLocks(); 1790 { 1791 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1792 if (_foregroundGCShouldWait) { 1793 // We are going to be waiting for action for the CMS thread; 1794 // it had better not be gone (for instance at shutdown)! 1795 assert(ConcurrentMarkSweepThread::cmst() != NULL, 1796 "CMS thread must be running"); 1797 // Wait here until the background collector gives us the go-ahead 1798 ConcurrentMarkSweepThread::clear_CMS_flag( 1799 ConcurrentMarkSweepThread::CMS_vm_has_token); // release token 1800 // Get a possibly blocked CMS thread going: 1801 // Note that we set _foregroundGCIsActive true above, 1802 // without protection of the CGC_lock. 1803 CGC_lock->notify(); 1804 assert(!ConcurrentMarkSweepThread::vm_thread_wants_cms_token(), 1805 "Possible deadlock"); 1806 while (_foregroundGCShouldWait) { 1807 // wait for notification 1808 CGC_lock->wait(Mutex::_no_safepoint_check_flag); 1809 // Possibility of delay/starvation here, since CMS token does 1810 // not know to give priority to VM thread? Actually, i think 1811 // there wouldn't be any delay/starvation, but the proof of 1812 // that "fact" (?) appears non-trivial. XXX 20011219YSR 1813 } 1814 ConcurrentMarkSweepThread::set_CMS_flag( 1815 ConcurrentMarkSweepThread::CMS_vm_has_token); 1816 } 1817 } 1818 // The CMS_token is already held. Get back the other locks. 1819 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 1820 "VM thread should have CMS token"); 1821 getFreelistLocks(); 1822 bitMapLock()->lock_without_safepoint_check(); 1823 if (TraceCMSState) { 1824 gclog_or_tty->print_cr("CMS foreground collector has asked for control " 1825 INTPTR_FORMAT " with first state %d", Thread::current(), first_state); 1826 gclog_or_tty->print_cr(" gets control with state %d", _collectorState); 1827 } 1828 1829 // Check if we need to do a compaction, or if not, whether 1830 // we need to start the mark-sweep from scratch. 1831 bool should_compact = false; 1832 bool should_start_over = false; 1833 decide_foreground_collection_type(clear_all_soft_refs, 1834 &should_compact, &should_start_over); 1835 1836 NOT_PRODUCT( 1837 if (RotateCMSCollectionTypes) { 1838 if (_cmsGen->debug_collection_type() == 1839 ConcurrentMarkSweepGeneration::MSC_foreground_collection_type) { 1840 should_compact = true; 1841 } else if (_cmsGen->debug_collection_type() == 1842 ConcurrentMarkSweepGeneration::MS_foreground_collection_type) { 1843 should_compact = false; 1844 } 1845 } 1846 ) 1847 1848 if (PrintGCDetails && first_state > Idling) { 1849 GCCause::Cause cause = GenCollectedHeap::heap()->gc_cause(); 1850 if (GCCause::is_user_requested_gc(cause) || 1851 GCCause::is_serviceability_requested_gc(cause)) { 1852 gclog_or_tty->print(" (concurrent mode interrupted)"); 1853 } else { 1854 gclog_or_tty->print(" (concurrent mode failure)"); 1855 } 1856 } 1857 1858 set_did_compact(should_compact); 1859 if (should_compact) { 1860 // If the collection is being acquired from the background 1861 // collector, there may be references on the discovered 1862 // references lists that have NULL referents (being those 1863 // that were concurrently cleared by a mutator) or 1864 // that are no longer active (having been enqueued concurrently 1865 // by the mutator). 1866 // Scrub the list of those references because Mark-Sweep-Compact 1867 // code assumes referents are not NULL and that all discovered 1868 // Reference objects are active. 1869 ref_processor()->clean_up_discovered_references(); 1870 1871 do_compaction_work(clear_all_soft_refs); 1872 1873 // Has the GC time limit been exceeded? 1874 DefNewGeneration* young_gen = _young_gen->as_DefNewGeneration(); 1875 size_t max_eden_size = young_gen->max_capacity() - 1876 young_gen->to()->capacity() - 1877 young_gen->from()->capacity(); 1878 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1879 GCCause::Cause gc_cause = gch->gc_cause(); 1880 size_policy()->check_gc_overhead_limit(_young_gen->used(), 1881 young_gen->eden()->used(), 1882 _cmsGen->max_capacity(), 1883 max_eden_size, 1884 full, 1885 gc_cause, 1886 gch->collector_policy()); 1887 } else { 1888 do_mark_sweep_work(clear_all_soft_refs, first_state, 1889 should_start_over); 1890 } 1891 // Reset the expansion cause, now that we just completed 1892 // a collection cycle. 1893 clear_expansion_cause(); 1894 _foregroundGCIsActive = false; 1895 return; 1896 } 1897 1898 // Resize the tenured generation 1899 // after obtaining the free list locks for the 1900 // two generations. 1901 void CMSCollector::compute_new_size() { 1902 assert_locked_or_safepoint(Heap_lock); 1903 FreelistLocker z(this); 1904 MetaspaceGC::compute_new_size(); 1905 _cmsGen->compute_new_size_free_list(); 1906 } 1907 1908 // A work method used by foreground collection to determine 1909 // what type of collection (compacting or not, continuing or fresh) 1910 // it should do. 1911 // NOTE: the intent is to make UseCMSCompactAtFullCollection 1912 // and CMSCompactWhenClearAllSoftRefs the default in the future 1913 // and do away with the flags after a suitable period. 1914 void CMSCollector::decide_foreground_collection_type( 1915 bool clear_all_soft_refs, bool* should_compact, 1916 bool* should_start_over) { 1917 // Normally, we'll compact only if the UseCMSCompactAtFullCollection 1918 // flag is set, and we have either requested a System.gc() or 1919 // the number of full gc's since the last concurrent cycle 1920 // has exceeded the threshold set by CMSFullGCsBeforeCompaction, 1921 // or if an incremental collection has failed 1922 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1923 assert(gch->collector_policy()->is_two_generation_policy(), 1924 "You may want to check the correctness of the following"); 1925 // Inform cms gen if this was due to partial collection failing. 1926 // The CMS gen may use this fact to determine its expansion policy. 1927 if (gch->incremental_collection_will_fail(false /* don't consult_young */)) { 1928 assert(!_cmsGen->incremental_collection_failed(), 1929 "Should have been noticed, reacted to and cleared"); 1930 _cmsGen->set_incremental_collection_failed(); 1931 } 1932 *should_compact = 1933 UseCMSCompactAtFullCollection && 1934 ((_full_gcs_since_conc_gc >= CMSFullGCsBeforeCompaction) || 1935 GCCause::is_user_requested_gc(gch->gc_cause()) || 1936 gch->incremental_collection_will_fail(true /* consult_young */)); 1937 *should_start_over = false; 1938 if (clear_all_soft_refs && !*should_compact) { 1939 // We are about to do a last ditch collection attempt 1940 // so it would normally make sense to do a compaction 1941 // to reclaim as much space as possible. 1942 if (CMSCompactWhenClearAllSoftRefs) { 1943 // Default: The rationale is that in this case either 1944 // we are past the final marking phase, in which case 1945 // we'd have to start over, or so little has been done 1946 // that there's little point in saving that work. Compaction 1947 // appears to be the sensible choice in either case. 1948 *should_compact = true; 1949 } else { 1950 // We have been asked to clear all soft refs, but not to 1951 // compact. Make sure that we aren't past the final checkpoint 1952 // phase, for that is where we process soft refs. If we are already 1953 // past that phase, we'll need to redo the refs discovery phase and 1954 // if necessary clear soft refs that weren't previously 1955 // cleared. We do so by remembering the phase in which 1956 // we came in, and if we are past the refs processing 1957 // phase, we'll choose to just redo the mark-sweep 1958 // collection from scratch. 1959 if (_collectorState > FinalMarking) { 1960 // We are past the refs processing phase; 1961 // start over and do a fresh synchronous CMS cycle 1962 _collectorState = Resetting; // skip to reset to start new cycle 1963 reset(false /* == !asynch */); 1964 *should_start_over = true; 1965 } // else we can continue a possibly ongoing current cycle 1966 } 1967 } 1968 } 1969 1970 // A work method used by the foreground collector to do 1971 // a mark-sweep-compact. 1972 void CMSCollector::do_compaction_work(bool clear_all_soft_refs) { 1973 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1974 TraceTime t("CMS:MSC ", PrintGCDetails && Verbose, true, gclog_or_tty); 1975 if (PrintGC && Verbose && !(GCCause::is_user_requested_gc(gch->gc_cause()))) { 1976 gclog_or_tty->print_cr("Compact ConcurrentMarkSweepGeneration after %d " 1977 "collections passed to foreground collector", _full_gcs_since_conc_gc); 1978 } 1979 1980 // Sample collection interval time and reset for collection pause. 1981 if (UseAdaptiveSizePolicy) { 1982 size_policy()->msc_collection_begin(); 1983 } 1984 1985 // Temporarily widen the span of the weak reference processing to 1986 // the entire heap. 1987 MemRegion new_span(GenCollectedHeap::heap()->reserved_region()); 1988 ReferenceProcessorSpanMutator rp_mut_span(ref_processor(), new_span); 1989 // Temporarily, clear the "is_alive_non_header" field of the 1990 // reference processor. 1991 ReferenceProcessorIsAliveMutator rp_mut_closure(ref_processor(), NULL); 1992 // Temporarily make reference _processing_ single threaded (non-MT). 1993 ReferenceProcessorMTProcMutator rp_mut_mt_processing(ref_processor(), false); 1994 // Temporarily make refs discovery atomic 1995 ReferenceProcessorAtomicMutator rp_mut_atomic(ref_processor(), true); 1996 // Temporarily make reference _discovery_ single threaded (non-MT) 1997 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false); 1998 1999 ref_processor()->set_enqueuing_is_done(false); 2000 ref_processor()->enable_discovery(false /*verify_disabled*/, false /*check_no_refs*/); 2001 ref_processor()->setup_policy(clear_all_soft_refs); 2002 // If an asynchronous collection finishes, the _modUnionTable is 2003 // all clear. If we are assuming the collection from an asynchronous 2004 // collection, clear the _modUnionTable. 2005 assert(_collectorState != Idling || _modUnionTable.isAllClear(), 2006 "_modUnionTable should be clear if the baton was not passed"); 2007 _modUnionTable.clear_all(); 2008 assert(_collectorState != Idling || _ct->klass_rem_set()->mod_union_is_clear(), 2009 "mod union for klasses should be clear if the baton was passed"); 2010 _ct->klass_rem_set()->clear_mod_union(); 2011 2012 // We must adjust the allocation statistics being maintained 2013 // in the free list space. We do so by reading and clearing 2014 // the sweep timer and updating the block flux rate estimates below. 2015 assert(!_intra_sweep_timer.is_active(), "_intra_sweep_timer should be inactive"); 2016 if (_inter_sweep_timer.is_active()) { 2017 _inter_sweep_timer.stop(); 2018 // Note that we do not use this sample to update the _inter_sweep_estimate. 2019 _cmsGen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()), 2020 _inter_sweep_estimate.padded_average(), 2021 _intra_sweep_estimate.padded_average()); 2022 } 2023 2024 GenMarkSweep::invoke_at_safepoint(_cmsGen->level(), 2025 ref_processor(), clear_all_soft_refs); 2026 #ifdef ASSERT 2027 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 2028 size_t free_size = cms_space->free(); 2029 assert(free_size == 2030 pointer_delta(cms_space->end(), cms_space->compaction_top()) 2031 * HeapWordSize, 2032 "All the free space should be compacted into one chunk at top"); 2033 assert(cms_space->dictionary()->total_chunk_size( 2034 debug_only(cms_space->freelistLock())) == 0 || 2035 cms_space->totalSizeInIndexedFreeLists() == 0, 2036 "All the free space should be in a single chunk"); 2037 size_t num = cms_space->totalCount(); 2038 assert((free_size == 0 && num == 0) || 2039 (free_size > 0 && (num == 1 || num == 2)), 2040 "There should be at most 2 free chunks after compaction"); 2041 #endif // ASSERT 2042 _collectorState = Resetting; 2043 assert(_restart_addr == NULL, 2044 "Should have been NULL'd before baton was passed"); 2045 reset(false /* == !asynch */); 2046 _cmsGen->reset_after_compaction(); 2047 _concurrent_cycles_since_last_unload = 0; 2048 2049 // Clear any data recorded in the PLAB chunk arrays. 2050 if (_survivor_plab_array != NULL) { 2051 reset_survivor_plab_arrays(); 2052 } 2053 2054 // Adjust the per-size allocation stats for the next epoch. 2055 _cmsGen->cmsSpace()->endSweepFLCensus(sweep_count() /* fake */); 2056 // Restart the "inter sweep timer" for the next epoch. 2057 _inter_sweep_timer.reset(); 2058 _inter_sweep_timer.start(); 2059 2060 // Sample collection pause time and reset for collection interval. 2061 if (UseAdaptiveSizePolicy) { 2062 size_policy()->msc_collection_end(gch->gc_cause()); 2063 } 2064 2065 // For a mark-sweep-compact, compute_new_size() will be called 2066 // in the heap's do_collection() method. 2067 } 2068 2069 // A work method used by the foreground collector to do 2070 // a mark-sweep, after taking over from a possibly on-going 2071 // concurrent mark-sweep collection. 2072 void CMSCollector::do_mark_sweep_work(bool clear_all_soft_refs, 2073 CollectorState first_state, bool should_start_over) { 2074 if (PrintGC && Verbose) { 2075 gclog_or_tty->print_cr("Pass concurrent collection to foreground " 2076 "collector with count %d", 2077 _full_gcs_since_conc_gc); 2078 } 2079 switch (_collectorState) { 2080 case Idling: 2081 if (first_state == Idling || should_start_over) { 2082 // The background GC was not active, or should 2083 // restarted from scratch; start the cycle. 2084 _collectorState = InitialMarking; 2085 } 2086 // If first_state was not Idling, then a background GC 2087 // was in progress and has now finished. No need to do it 2088 // again. Leave the state as Idling. 2089 break; 2090 case Precleaning: 2091 // In the foreground case don't do the precleaning since 2092 // it is not done concurrently and there is extra work 2093 // required. 2094 _collectorState = FinalMarking; 2095 } 2096 collect_in_foreground(clear_all_soft_refs); 2097 2098 // For a mark-sweep, compute_new_size() will be called 2099 // in the heap's do_collection() method. 2100 } 2101 2102 2103 void CMSCollector::getFreelistLocks() const { 2104 // Get locks for all free lists in all generations that this 2105 // collector is responsible for 2106 _cmsGen->freelistLock()->lock_without_safepoint_check(); 2107 } 2108 2109 void CMSCollector::releaseFreelistLocks() const { 2110 // Release locks for all free lists in all generations that this 2111 // collector is responsible for 2112 _cmsGen->freelistLock()->unlock(); 2113 } 2114 2115 bool CMSCollector::haveFreelistLocks() const { 2116 // Check locks for all free lists in all generations that this 2117 // collector is responsible for 2118 assert_lock_strong(_cmsGen->freelistLock()); 2119 PRODUCT_ONLY(ShouldNotReachHere()); 2120 return true; 2121 } 2122 2123 // A utility class that is used by the CMS collector to 2124 // temporarily "release" the foreground collector from its 2125 // usual obligation to wait for the background collector to 2126 // complete an ongoing phase before proceeding. 2127 class ReleaseForegroundGC: public StackObj { 2128 private: 2129 CMSCollector* _c; 2130 public: 2131 ReleaseForegroundGC(CMSCollector* c) : _c(c) { 2132 assert(_c->_foregroundGCShouldWait, "Else should not need to call"); 2133 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 2134 // allow a potentially blocked foreground collector to proceed 2135 _c->_foregroundGCShouldWait = false; 2136 if (_c->_foregroundGCIsActive) { 2137 CGC_lock->notify(); 2138 } 2139 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 2140 "Possible deadlock"); 2141 } 2142 2143 ~ReleaseForegroundGC() { 2144 assert(!_c->_foregroundGCShouldWait, "Usage protocol violation?"); 2145 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 2146 _c->_foregroundGCShouldWait = true; 2147 } 2148 }; 2149 2150 // There are separate collect_in_background and collect_in_foreground because of 2151 // the different locking requirements of the background collector and the 2152 // foreground collector. There was originally an attempt to share 2153 // one "collect" method between the background collector and the foreground 2154 // collector but the if-then-else required made it cleaner to have 2155 // separate methods. 2156 void CMSCollector::collect_in_background(bool clear_all_soft_refs) { 2157 assert(Thread::current()->is_ConcurrentGC_thread(), 2158 "A CMS asynchronous collection is only allowed on a CMS thread."); 2159 2160 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2161 { 2162 bool safepoint_check = Mutex::_no_safepoint_check_flag; 2163 MutexLockerEx hl(Heap_lock, safepoint_check); 2164 FreelistLocker fll(this); 2165 MutexLockerEx x(CGC_lock, safepoint_check); 2166 if (_foregroundGCIsActive || !UseAsyncConcMarkSweepGC) { 2167 // The foreground collector is active or we're 2168 // not using asynchronous collections. Skip this 2169 // background collection. 2170 assert(!_foregroundGCShouldWait, "Should be clear"); 2171 return; 2172 } else { 2173 assert(_collectorState == Idling, "Should be idling before start."); 2174 _collectorState = InitialMarking; 2175 // Reset the expansion cause, now that we are about to begin 2176 // a new cycle. 2177 clear_expansion_cause(); 2178 2179 // Clear the MetaspaceGC flag since a concurrent collection 2180 // is starting but also clear it after the collection. 2181 MetaspaceGC::set_should_concurrent_collect(false); 2182 } 2183 // Decide if we want to enable class unloading as part of the 2184 // ensuing concurrent GC cycle. 2185 update_should_unload_classes(); 2186 _full_gc_requested = false; // acks all outstanding full gc requests 2187 // Signal that we are about to start a collection 2188 gch->increment_total_full_collections(); // ... starting a collection cycle 2189 _collection_count_start = gch->total_full_collections(); 2190 } 2191 2192 // Used for PrintGC 2193 size_t prev_used; 2194 if (PrintGC && Verbose) { 2195 prev_used = _cmsGen->used(); // XXXPERM 2196 } 2197 2198 // The change of the collection state is normally done at this level; 2199 // the exceptions are phases that are executed while the world is 2200 // stopped. For those phases the change of state is done while the 2201 // world is stopped. For baton passing purposes this allows the 2202 // background collector to finish the phase and change state atomically. 2203 // The foreground collector cannot wait on a phase that is done 2204 // while the world is stopped because the foreground collector already 2205 // has the world stopped and would deadlock. 2206 while (_collectorState != Idling) { 2207 if (TraceCMSState) { 2208 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " in CMS state %d", 2209 Thread::current(), _collectorState); 2210 } 2211 // The foreground collector 2212 // holds the Heap_lock throughout its collection. 2213 // holds the CMS token (but not the lock) 2214 // except while it is waiting for the background collector to yield. 2215 // 2216 // The foreground collector should be blocked (not for long) 2217 // if the background collector is about to start a phase 2218 // executed with world stopped. If the background 2219 // collector has already started such a phase, the 2220 // foreground collector is blocked waiting for the 2221 // Heap_lock. The stop-world phases (InitialMarking and FinalMarking) 2222 // are executed in the VM thread. 2223 // 2224 // The locking order is 2225 // PendingListLock (PLL) -- if applicable (FinalMarking) 2226 // Heap_lock (both this & PLL locked in VM_CMS_Operation::prologue()) 2227 // CMS token (claimed in 2228 // stop_world_and_do() --> 2229 // safepoint_synchronize() --> 2230 // CMSThread::synchronize()) 2231 2232 { 2233 // Check if the FG collector wants us to yield. 2234 CMSTokenSync x(true); // is cms thread 2235 if (waitForForegroundGC()) { 2236 // We yielded to a foreground GC, nothing more to be 2237 // done this round. 2238 assert(_foregroundGCShouldWait == false, "We set it to false in " 2239 "waitForForegroundGC()"); 2240 if (TraceCMSState) { 2241 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT 2242 " exiting collection CMS state %d", 2243 Thread::current(), _collectorState); 2244 } 2245 return; 2246 } else { 2247 // The background collector can run but check to see if the 2248 // foreground collector has done a collection while the 2249 // background collector was waiting to get the CGC_lock 2250 // above. If yes, break so that _foregroundGCShouldWait 2251 // is cleared before returning. 2252 if (_collectorState == Idling) { 2253 break; 2254 } 2255 } 2256 } 2257 2258 assert(_foregroundGCShouldWait, "Foreground collector, if active, " 2259 "should be waiting"); 2260 2261 switch (_collectorState) { 2262 case InitialMarking: 2263 { 2264 ReleaseForegroundGC x(this); 2265 stats().record_cms_begin(); 2266 2267 VM_CMS_Initial_Mark initial_mark_op(this); 2268 VMThread::execute(&initial_mark_op); 2269 } 2270 // The collector state may be any legal state at this point 2271 // since the background collector may have yielded to the 2272 // foreground collector. 2273 break; 2274 case Marking: 2275 // initial marking in checkpointRootsInitialWork has been completed 2276 if (markFromRoots(true)) { // we were successful 2277 assert(_collectorState == Precleaning, "Collector state should " 2278 "have changed"); 2279 } else { 2280 assert(_foregroundGCIsActive, "Internal state inconsistency"); 2281 } 2282 break; 2283 case Precleaning: 2284 if (UseAdaptiveSizePolicy) { 2285 size_policy()->concurrent_precleaning_begin(); 2286 } 2287 // marking from roots in markFromRoots has been completed 2288 preclean(); 2289 if (UseAdaptiveSizePolicy) { 2290 size_policy()->concurrent_precleaning_end(); 2291 } 2292 assert(_collectorState == AbortablePreclean || 2293 _collectorState == FinalMarking, 2294 "Collector state should have changed"); 2295 break; 2296 case AbortablePreclean: 2297 if (UseAdaptiveSizePolicy) { 2298 size_policy()->concurrent_phases_resume(); 2299 } 2300 abortable_preclean(); 2301 if (UseAdaptiveSizePolicy) { 2302 size_policy()->concurrent_precleaning_end(); 2303 } 2304 assert(_collectorState == FinalMarking, "Collector state should " 2305 "have changed"); 2306 break; 2307 case FinalMarking: 2308 { 2309 ReleaseForegroundGC x(this); 2310 2311 VM_CMS_Final_Remark final_remark_op(this); 2312 VMThread::execute(&final_remark_op); 2313 } 2314 assert(_foregroundGCShouldWait, "block post-condition"); 2315 break; 2316 case Sweeping: 2317 if (UseAdaptiveSizePolicy) { 2318 size_policy()->concurrent_sweeping_begin(); 2319 } 2320 // final marking in checkpointRootsFinal has been completed 2321 sweep(true); 2322 assert(_collectorState == Resizing, "Collector state change " 2323 "to Resizing must be done under the free_list_lock"); 2324 _full_gcs_since_conc_gc = 0; 2325 2326 // Stop the timers for adaptive size policy for the concurrent phases 2327 if (UseAdaptiveSizePolicy) { 2328 size_policy()->concurrent_sweeping_end(); 2329 size_policy()->concurrent_phases_end(gch->gc_cause(), 2330 gch->prev_gen(_cmsGen)->capacity(), 2331 _cmsGen->free()); 2332 } 2333 2334 case Resizing: { 2335 // Sweeping has been completed... 2336 // At this point the background collection has completed. 2337 // Don't move the call to compute_new_size() down 2338 // into code that might be executed if the background 2339 // collection was preempted. 2340 { 2341 ReleaseForegroundGC x(this); // unblock FG collection 2342 MutexLockerEx y(Heap_lock, Mutex::_no_safepoint_check_flag); 2343 CMSTokenSync z(true); // not strictly needed. 2344 if (_collectorState == Resizing) { 2345 compute_new_size(); 2346 _collectorState = Resetting; 2347 } else { 2348 assert(_collectorState == Idling, "The state should only change" 2349 " because the foreground collector has finished the collection"); 2350 } 2351 } 2352 break; 2353 } 2354 case Resetting: 2355 // CMS heap resizing has been completed 2356 reset(true); 2357 assert(_collectorState == Idling, "Collector state should " 2358 "have changed"); 2359 2360 MetaspaceGC::set_should_concurrent_collect(false); 2361 2362 stats().record_cms_end(); 2363 // Don't move the concurrent_phases_end() and compute_new_size() 2364 // calls to here because a preempted background collection 2365 // has it's state set to "Resetting". 2366 break; 2367 case Idling: 2368 default: 2369 ShouldNotReachHere(); 2370 break; 2371 } 2372 if (TraceCMSState) { 2373 gclog_or_tty->print_cr(" Thread " INTPTR_FORMAT " done - next CMS state %d", 2374 Thread::current(), _collectorState); 2375 } 2376 assert(_foregroundGCShouldWait, "block post-condition"); 2377 } 2378 2379 // Should this be in gc_epilogue? 2380 collector_policy()->counters()->update_counters(); 2381 2382 { 2383 // Clear _foregroundGCShouldWait and, in the event that the 2384 // foreground collector is waiting, notify it, before 2385 // returning. 2386 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 2387 _foregroundGCShouldWait = false; 2388 if (_foregroundGCIsActive) { 2389 CGC_lock->notify(); 2390 } 2391 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 2392 "Possible deadlock"); 2393 } 2394 if (TraceCMSState) { 2395 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT 2396 " exiting collection CMS state %d", 2397 Thread::current(), _collectorState); 2398 } 2399 if (PrintGC && Verbose) { 2400 _cmsGen->print_heap_change(prev_used); 2401 } 2402 } 2403 2404 void CMSCollector::collect_in_foreground(bool clear_all_soft_refs) { 2405 assert(_foregroundGCIsActive && !_foregroundGCShouldWait, 2406 "Foreground collector should be waiting, not executing"); 2407 assert(Thread::current()->is_VM_thread(), "A foreground collection" 2408 "may only be done by the VM Thread with the world stopped"); 2409 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 2410 "VM thread should have CMS token"); 2411 2412 NOT_PRODUCT(TraceTime t("CMS:MS (foreground) ", PrintGCDetails && Verbose, 2413 true, gclog_or_tty);) 2414 if (UseAdaptiveSizePolicy) { 2415 size_policy()->ms_collection_begin(); 2416 } 2417 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact); 2418 2419 HandleMark hm; // Discard invalid handles created during verification 2420 2421 if (VerifyBeforeGC && 2422 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2423 Universe::verify(); 2424 } 2425 2426 // Snapshot the soft reference policy to be used in this collection cycle. 2427 ref_processor()->setup_policy(clear_all_soft_refs); 2428 2429 bool init_mark_was_synchronous = false; // until proven otherwise 2430 while (_collectorState != Idling) { 2431 if (TraceCMSState) { 2432 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " in CMS state %d", 2433 Thread::current(), _collectorState); 2434 } 2435 switch (_collectorState) { 2436 case InitialMarking: 2437 init_mark_was_synchronous = true; // fact to be exploited in re-mark 2438 checkpointRootsInitial(false); 2439 assert(_collectorState == Marking, "Collector state should have changed" 2440 " within checkpointRootsInitial()"); 2441 break; 2442 case Marking: 2443 // initial marking in checkpointRootsInitialWork has been completed 2444 if (VerifyDuringGC && 2445 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2446 Universe::verify("Verify before initial mark: "); 2447 } 2448 { 2449 bool res = markFromRoots(false); 2450 assert(res && _collectorState == FinalMarking, "Collector state should " 2451 "have changed"); 2452 break; 2453 } 2454 case FinalMarking: 2455 if (VerifyDuringGC && 2456 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2457 Universe::verify("Verify before re-mark: "); 2458 } 2459 checkpointRootsFinal(false, clear_all_soft_refs, 2460 init_mark_was_synchronous); 2461 assert(_collectorState == Sweeping, "Collector state should not " 2462 "have changed within checkpointRootsFinal()"); 2463 break; 2464 case Sweeping: 2465 // final marking in checkpointRootsFinal has been completed 2466 if (VerifyDuringGC && 2467 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2468 Universe::verify("Verify before sweep: "); 2469 } 2470 sweep(false); 2471 assert(_collectorState == Resizing, "Incorrect state"); 2472 break; 2473 case Resizing: { 2474 // Sweeping has been completed; the actual resize in this case 2475 // is done separately; nothing to be done in this state. 2476 _collectorState = Resetting; 2477 break; 2478 } 2479 case Resetting: 2480 // The heap has been resized. 2481 if (VerifyDuringGC && 2482 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2483 Universe::verify("Verify before reset: "); 2484 } 2485 reset(false); 2486 assert(_collectorState == Idling, "Collector state should " 2487 "have changed"); 2488 break; 2489 case Precleaning: 2490 case AbortablePreclean: 2491 // Elide the preclean phase 2492 _collectorState = FinalMarking; 2493 break; 2494 default: 2495 ShouldNotReachHere(); 2496 } 2497 if (TraceCMSState) { 2498 gclog_or_tty->print_cr(" Thread " INTPTR_FORMAT " done - next CMS state %d", 2499 Thread::current(), _collectorState); 2500 } 2501 } 2502 2503 if (UseAdaptiveSizePolicy) { 2504 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2505 size_policy()->ms_collection_end(gch->gc_cause()); 2506 } 2507 2508 if (VerifyAfterGC && 2509 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2510 Universe::verify(); 2511 } 2512 if (TraceCMSState) { 2513 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT 2514 " exiting collection CMS state %d", 2515 Thread::current(), _collectorState); 2516 } 2517 } 2518 2519 bool CMSCollector::waitForForegroundGC() { 2520 bool res = false; 2521 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 2522 "CMS thread should have CMS token"); 2523 // Block the foreground collector until the 2524 // background collectors decides whether to 2525 // yield. 2526 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 2527 _foregroundGCShouldWait = true; 2528 if (_foregroundGCIsActive) { 2529 // The background collector yields to the 2530 // foreground collector and returns a value 2531 // indicating that it has yielded. The foreground 2532 // collector can proceed. 2533 res = true; 2534 _foregroundGCShouldWait = false; 2535 ConcurrentMarkSweepThread::clear_CMS_flag( 2536 ConcurrentMarkSweepThread::CMS_cms_has_token); 2537 ConcurrentMarkSweepThread::set_CMS_flag( 2538 ConcurrentMarkSweepThread::CMS_cms_wants_token); 2539 // Get a possibly blocked foreground thread going 2540 CGC_lock->notify(); 2541 if (TraceCMSState) { 2542 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " waiting at CMS state %d", 2543 Thread::current(), _collectorState); 2544 } 2545 while (_foregroundGCIsActive) { 2546 CGC_lock->wait(Mutex::_no_safepoint_check_flag); 2547 } 2548 ConcurrentMarkSweepThread::set_CMS_flag( 2549 ConcurrentMarkSweepThread::CMS_cms_has_token); 2550 ConcurrentMarkSweepThread::clear_CMS_flag( 2551 ConcurrentMarkSweepThread::CMS_cms_wants_token); 2552 } 2553 if (TraceCMSState) { 2554 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " continuing at CMS state %d", 2555 Thread::current(), _collectorState); 2556 } 2557 return res; 2558 } 2559 2560 // Because of the need to lock the free lists and other structures in 2561 // the collector, common to all the generations that the collector is 2562 // collecting, we need the gc_prologues of individual CMS generations 2563 // delegate to their collector. It may have been simpler had the 2564 // current infrastructure allowed one to call a prologue on a 2565 // collector. In the absence of that we have the generation's 2566 // prologue delegate to the collector, which delegates back 2567 // some "local" work to a worker method in the individual generations 2568 // that it's responsible for collecting, while itself doing any 2569 // work common to all generations it's responsible for. A similar 2570 // comment applies to the gc_epilogue()'s. 2571 // The role of the varaible _between_prologue_and_epilogue is to 2572 // enforce the invocation protocol. 2573 void CMSCollector::gc_prologue(bool full) { 2574 // Call gc_prologue_work() for the CMSGen 2575 // we are responsible for. 2576 2577 // The following locking discipline assumes that we are only called 2578 // when the world is stopped. 2579 assert(SafepointSynchronize::is_at_safepoint(), "world is stopped assumption"); 2580 2581 // The CMSCollector prologue must call the gc_prologues for the 2582 // "generations" that it's responsible 2583 // for. 2584 2585 assert( Thread::current()->is_VM_thread() 2586 || ( CMSScavengeBeforeRemark 2587 && Thread::current()->is_ConcurrentGC_thread()), 2588 "Incorrect thread type for prologue execution"); 2589 2590 if (_between_prologue_and_epilogue) { 2591 // We have already been invoked; this is a gc_prologue delegation 2592 // from yet another CMS generation that we are responsible for, just 2593 // ignore it since all relevant work has already been done. 2594 return; 2595 } 2596 2597 // set a bit saying prologue has been called; cleared in epilogue 2598 _between_prologue_and_epilogue = true; 2599 // Claim locks for common data structures, then call gc_prologue_work() 2600 // for each CMSGen. 2601 2602 getFreelistLocks(); // gets free list locks on constituent spaces 2603 bitMapLock()->lock_without_safepoint_check(); 2604 2605 // Should call gc_prologue_work() for all cms gens we are responsible for 2606 bool duringMarking = _collectorState >= Marking 2607 && _collectorState < Sweeping; 2608 2609 // The young collections clear the modified oops state, which tells if 2610 // there are any modified oops in the class. The remark phase also needs 2611 // that information. Tell the young collection to save the union of all 2612 // modified klasses. 2613 if (duringMarking) { 2614 _ct->klass_rem_set()->set_accumulate_modified_oops(true); 2615 } 2616 2617 bool registerClosure = duringMarking; 2618 2619 ModUnionClosure* muc = CollectedHeap::use_parallel_gc_threads() ? 2620 &_modUnionClosurePar 2621 : &_modUnionClosure; 2622 _cmsGen->gc_prologue_work(full, registerClosure, muc); 2623 2624 if (!full) { 2625 stats().record_gc0_begin(); 2626 } 2627 } 2628 2629 void ConcurrentMarkSweepGeneration::gc_prologue(bool full) { 2630 2631 _capacity_at_prologue = capacity(); 2632 _used_at_prologue = used(); 2633 2634 // Delegate to CMScollector which knows how to coordinate between 2635 // this and any other CMS generations that it is responsible for 2636 // collecting. 2637 collector()->gc_prologue(full); 2638 } 2639 2640 // This is a "private" interface for use by this generation's CMSCollector. 2641 // Not to be called directly by any other entity (for instance, 2642 // GenCollectedHeap, which calls the "public" gc_prologue method above). 2643 void ConcurrentMarkSweepGeneration::gc_prologue_work(bool full, 2644 bool registerClosure, ModUnionClosure* modUnionClosure) { 2645 assert(!incremental_collection_failed(), "Shouldn't be set yet"); 2646 assert(cmsSpace()->preconsumptionDirtyCardClosure() == NULL, 2647 "Should be NULL"); 2648 if (registerClosure) { 2649 cmsSpace()->setPreconsumptionDirtyCardClosure(modUnionClosure); 2650 } 2651 cmsSpace()->gc_prologue(); 2652 // Clear stat counters 2653 NOT_PRODUCT( 2654 assert(_numObjectsPromoted == 0, "check"); 2655 assert(_numWordsPromoted == 0, "check"); 2656 if (Verbose && PrintGC) { 2657 gclog_or_tty->print("Allocated "SIZE_FORMAT" objects, " 2658 SIZE_FORMAT" bytes concurrently", 2659 _numObjectsAllocated, _numWordsAllocated*sizeof(HeapWord)); 2660 } 2661 _numObjectsAllocated = 0; 2662 _numWordsAllocated = 0; 2663 ) 2664 } 2665 2666 void CMSCollector::gc_epilogue(bool full) { 2667 // The following locking discipline assumes that we are only called 2668 // when the world is stopped. 2669 assert(SafepointSynchronize::is_at_safepoint(), 2670 "world is stopped assumption"); 2671 2672 // Currently the CMS epilogue (see CompactibleFreeListSpace) merely checks 2673 // if linear allocation blocks need to be appropriately marked to allow the 2674 // the blocks to be parsable. We also check here whether we need to nudge the 2675 // CMS collector thread to start a new cycle (if it's not already active). 2676 assert( Thread::current()->is_VM_thread() 2677 || ( CMSScavengeBeforeRemark 2678 && Thread::current()->is_ConcurrentGC_thread()), 2679 "Incorrect thread type for epilogue execution"); 2680 2681 if (!_between_prologue_and_epilogue) { 2682 // We have already been invoked; this is a gc_epilogue delegation 2683 // from yet another CMS generation that we are responsible for, just 2684 // ignore it since all relevant work has already been done. 2685 return; 2686 } 2687 assert(haveFreelistLocks(), "must have freelist locks"); 2688 assert_lock_strong(bitMapLock()); 2689 2690 _ct->klass_rem_set()->set_accumulate_modified_oops(false); 2691 2692 _cmsGen->gc_epilogue_work(full); 2693 2694 if (_collectorState == AbortablePreclean || _collectorState == Precleaning) { 2695 // in case sampling was not already enabled, enable it 2696 _start_sampling = true; 2697 } 2698 // reset _eden_chunk_array so sampling starts afresh 2699 _eden_chunk_index = 0; 2700 2701 size_t cms_used = _cmsGen->cmsSpace()->used(); 2702 2703 // update performance counters - this uses a special version of 2704 // update_counters() that allows the utilization to be passed as a 2705 // parameter, avoiding multiple calls to used(). 2706 // 2707 _cmsGen->update_counters(cms_used); 2708 2709 if (CMSIncrementalMode) { 2710 icms_update_allocation_limits(); 2711 } 2712 2713 bitMapLock()->unlock(); 2714 releaseFreelistLocks(); 2715 2716 if (!CleanChunkPoolAsync) { 2717 Chunk::clean_chunk_pool(); 2718 } 2719 2720 set_did_compact(false); 2721 _between_prologue_and_epilogue = false; // ready for next cycle 2722 } 2723 2724 void ConcurrentMarkSweepGeneration::gc_epilogue(bool full) { 2725 collector()->gc_epilogue(full); 2726 2727 // Also reset promotion tracking in par gc thread states. 2728 if (CollectedHeap::use_parallel_gc_threads()) { 2729 for (uint i = 0; i < ParallelGCThreads; i++) { 2730 _par_gc_thread_states[i]->promo.stopTrackingPromotions(i); 2731 } 2732 } 2733 } 2734 2735 void ConcurrentMarkSweepGeneration::gc_epilogue_work(bool full) { 2736 assert(!incremental_collection_failed(), "Should have been cleared"); 2737 cmsSpace()->setPreconsumptionDirtyCardClosure(NULL); 2738 cmsSpace()->gc_epilogue(); 2739 // Print stat counters 2740 NOT_PRODUCT( 2741 assert(_numObjectsAllocated == 0, "check"); 2742 assert(_numWordsAllocated == 0, "check"); 2743 if (Verbose && PrintGC) { 2744 gclog_or_tty->print("Promoted "SIZE_FORMAT" objects, " 2745 SIZE_FORMAT" bytes", 2746 _numObjectsPromoted, _numWordsPromoted*sizeof(HeapWord)); 2747 } 2748 _numObjectsPromoted = 0; 2749 _numWordsPromoted = 0; 2750 ) 2751 2752 if (PrintGC && Verbose) { 2753 // Call down the chain in contiguous_available needs the freelistLock 2754 // so print this out before releasing the freeListLock. 2755 gclog_or_tty->print(" Contiguous available "SIZE_FORMAT" bytes ", 2756 contiguous_available()); 2757 } 2758 } 2759 2760 #ifndef PRODUCT 2761 bool CMSCollector::have_cms_token() { 2762 Thread* thr = Thread::current(); 2763 if (thr->is_VM_thread()) { 2764 return ConcurrentMarkSweepThread::vm_thread_has_cms_token(); 2765 } else if (thr->is_ConcurrentGC_thread()) { 2766 return ConcurrentMarkSweepThread::cms_thread_has_cms_token(); 2767 } else if (thr->is_GC_task_thread()) { 2768 return ConcurrentMarkSweepThread::vm_thread_has_cms_token() && 2769 ParGCRareEvent_lock->owned_by_self(); 2770 } 2771 return false; 2772 } 2773 #endif 2774 2775 // Check reachability of the given heap address in CMS generation, 2776 // treating all other generations as roots. 2777 bool CMSCollector::is_cms_reachable(HeapWord* addr) { 2778 // We could "guarantee" below, rather than assert, but i'll 2779 // leave these as "asserts" so that an adventurous debugger 2780 // could try this in the product build provided some subset of 2781 // the conditions were met, provided they were intersted in the 2782 // results and knew that the computation below wouldn't interfere 2783 // with other concurrent computations mutating the structures 2784 // being read or written. 2785 assert(SafepointSynchronize::is_at_safepoint(), 2786 "Else mutations in object graph will make answer suspect"); 2787 assert(have_cms_token(), "Should hold cms token"); 2788 assert(haveFreelistLocks(), "must hold free list locks"); 2789 assert_lock_strong(bitMapLock()); 2790 2791 // Clear the marking bit map array before starting, but, just 2792 // for kicks, first report if the given address is already marked 2793 gclog_or_tty->print_cr("Start: Address 0x%x is%s marked", addr, 2794 _markBitMap.isMarked(addr) ? "" : " not"); 2795 2796 if (verify_after_remark()) { 2797 MutexLockerEx x(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag); 2798 bool result = verification_mark_bm()->isMarked(addr); 2799 gclog_or_tty->print_cr("TransitiveMark: Address 0x%x %s marked", addr, 2800 result ? "IS" : "is NOT"); 2801 return result; 2802 } else { 2803 gclog_or_tty->print_cr("Could not compute result"); 2804 return false; 2805 } 2806 } 2807 2808 2809 void 2810 CMSCollector::print_on_error(outputStream* st) { 2811 CMSCollector* collector = ConcurrentMarkSweepGeneration::_collector; 2812 if (collector != NULL) { 2813 CMSBitMap* bitmap = &collector->_markBitMap; 2814 st->print_cr("Marking Bits: (CMSBitMap*) " PTR_FORMAT, bitmap); 2815 bitmap->print_on_error(st, " Bits: "); 2816 2817 st->cr(); 2818 2819 CMSBitMap* mut_bitmap = &collector->_modUnionTable; 2820 st->print_cr("Mod Union Table: (CMSBitMap*) " PTR_FORMAT, mut_bitmap); 2821 mut_bitmap->print_on_error(st, " Bits: "); 2822 } 2823 } 2824 2825 //////////////////////////////////////////////////////// 2826 // CMS Verification Support 2827 //////////////////////////////////////////////////////// 2828 // Following the remark phase, the following invariant 2829 // should hold -- each object in the CMS heap which is 2830 // marked in markBitMap() should be marked in the verification_mark_bm(). 2831 2832 class VerifyMarkedClosure: public BitMapClosure { 2833 CMSBitMap* _marks; 2834 bool _failed; 2835 2836 public: 2837 VerifyMarkedClosure(CMSBitMap* bm): _marks(bm), _failed(false) {} 2838 2839 bool do_bit(size_t offset) { 2840 HeapWord* addr = _marks->offsetToHeapWord(offset); 2841 if (!_marks->isMarked(addr)) { 2842 oop(addr)->print_on(gclog_or_tty); 2843 gclog_or_tty->print_cr(" ("INTPTR_FORMAT" should have been marked)", addr); 2844 _failed = true; 2845 } 2846 return true; 2847 } 2848 2849 bool failed() { return _failed; } 2850 }; 2851 2852 bool CMSCollector::verify_after_remark(bool silent) { 2853 if (!silent) gclog_or_tty->print(" [Verifying CMS Marking... "); 2854 MutexLockerEx ml(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag); 2855 static bool init = false; 2856 2857 assert(SafepointSynchronize::is_at_safepoint(), 2858 "Else mutations in object graph will make answer suspect"); 2859 assert(have_cms_token(), 2860 "Else there may be mutual interference in use of " 2861 " verification data structures"); 2862 assert(_collectorState > Marking && _collectorState <= Sweeping, 2863 "Else marking info checked here may be obsolete"); 2864 assert(haveFreelistLocks(), "must hold free list locks"); 2865 assert_lock_strong(bitMapLock()); 2866 2867 2868 // Allocate marking bit map if not already allocated 2869 if (!init) { // first time 2870 if (!verification_mark_bm()->allocate(_span)) { 2871 return false; 2872 } 2873 init = true; 2874 } 2875 2876 assert(verification_mark_stack()->isEmpty(), "Should be empty"); 2877 2878 // Turn off refs discovery -- so we will be tracing through refs. 2879 // This is as intended, because by this time 2880 // GC must already have cleared any refs that need to be cleared, 2881 // and traced those that need to be marked; moreover, 2882 // the marking done here is not going to intefere in any 2883 // way with the marking information used by GC. 2884 NoRefDiscovery no_discovery(ref_processor()); 2885 2886 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;) 2887 2888 // Clear any marks from a previous round 2889 verification_mark_bm()->clear_all(); 2890 assert(verification_mark_stack()->isEmpty(), "markStack should be empty"); 2891 verify_work_stacks_empty(); 2892 2893 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2894 gch->ensure_parsability(false); // fill TLABs, but no need to retire them 2895 // Update the saved marks which may affect the root scans. 2896 gch->save_marks(); 2897 2898 if (CMSRemarkVerifyVariant == 1) { 2899 // In this first variant of verification, we complete 2900 // all marking, then check if the new marks-verctor is 2901 // a subset of the CMS marks-vector. 2902 verify_after_remark_work_1(); 2903 } else if (CMSRemarkVerifyVariant == 2) { 2904 // In this second variant of verification, we flag an error 2905 // (i.e. an object reachable in the new marks-vector not reachable 2906 // in the CMS marks-vector) immediately, also indicating the 2907 // identify of an object (A) that references the unmarked object (B) -- 2908 // presumably, a mutation to A failed to be picked up by preclean/remark? 2909 verify_after_remark_work_2(); 2910 } else { 2911 warning("Unrecognized value %d for CMSRemarkVerifyVariant", 2912 CMSRemarkVerifyVariant); 2913 } 2914 if (!silent) gclog_or_tty->print(" done] "); 2915 return true; 2916 } 2917 2918 void CMSCollector::verify_after_remark_work_1() { 2919 ResourceMark rm; 2920 HandleMark hm; 2921 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2922 2923 // Get a clear set of claim bits for the strong roots processing to work with. 2924 ClassLoaderDataGraph::clear_claimed_marks(); 2925 2926 // Mark from roots one level into CMS 2927 MarkRefsIntoClosure notOlder(_span, verification_mark_bm()); 2928 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2929 2930 gch->gen_process_strong_roots(_cmsGen->level(), 2931 true, // younger gens are roots 2932 true, // activate StrongRootsScope 2933 false, // not scavenging 2934 SharedHeap::ScanningOption(roots_scanning_options()), 2935 ¬Older, 2936 true, // walk code active on stacks 2937 NULL, 2938 NULL); // SSS: Provide correct closure 2939 2940 // Now mark from the roots 2941 MarkFromRootsClosure markFromRootsClosure(this, _span, 2942 verification_mark_bm(), verification_mark_stack(), 2943 false /* don't yield */, true /* verifying */); 2944 assert(_restart_addr == NULL, "Expected pre-condition"); 2945 verification_mark_bm()->iterate(&markFromRootsClosure); 2946 while (_restart_addr != NULL) { 2947 // Deal with stack overflow: by restarting at the indicated 2948 // address. 2949 HeapWord* ra = _restart_addr; 2950 markFromRootsClosure.reset(ra); 2951 _restart_addr = NULL; 2952 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end()); 2953 } 2954 assert(verification_mark_stack()->isEmpty(), "Should have been drained"); 2955 verify_work_stacks_empty(); 2956 2957 // Marking completed -- now verify that each bit marked in 2958 // verification_mark_bm() is also marked in markBitMap(); flag all 2959 // errors by printing corresponding objects. 2960 VerifyMarkedClosure vcl(markBitMap()); 2961 verification_mark_bm()->iterate(&vcl); 2962 if (vcl.failed()) { 2963 gclog_or_tty->print("Verification failed"); 2964 Universe::heap()->print_on(gclog_or_tty); 2965 fatal("CMS: failed marking verification after remark"); 2966 } 2967 } 2968 2969 class VerifyKlassOopsKlassClosure : public KlassClosure { 2970 class VerifyKlassOopsClosure : public OopClosure { 2971 CMSBitMap* _bitmap; 2972 public: 2973 VerifyKlassOopsClosure(CMSBitMap* bitmap) : _bitmap(bitmap) { } 2974 void do_oop(oop* p) { guarantee(*p == NULL || _bitmap->isMarked((HeapWord*) *p), "Should be marked"); } 2975 void do_oop(narrowOop* p) { ShouldNotReachHere(); } 2976 } _oop_closure; 2977 public: 2978 VerifyKlassOopsKlassClosure(CMSBitMap* bitmap) : _oop_closure(bitmap) {} 2979 void do_klass(Klass* k) { 2980 k->oops_do(&_oop_closure); 2981 } 2982 }; 2983 2984 void CMSCollector::verify_after_remark_work_2() { 2985 ResourceMark rm; 2986 HandleMark hm; 2987 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2988 2989 // Get a clear set of claim bits for the strong roots processing to work with. 2990 ClassLoaderDataGraph::clear_claimed_marks(); 2991 2992 // Mark from roots one level into CMS 2993 MarkRefsIntoVerifyClosure notOlder(_span, verification_mark_bm(), 2994 markBitMap()); 2995 CMKlassClosure klass_closure(¬Older); 2996 2997 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2998 gch->gen_process_strong_roots(_cmsGen->level(), 2999 true, // younger gens are roots 3000 true, // activate StrongRootsScope 3001 false, // not scavenging 3002 SharedHeap::ScanningOption(roots_scanning_options()), 3003 ¬Older, 3004 true, // walk code active on stacks 3005 NULL, 3006 &klass_closure); 3007 3008 // Now mark from the roots 3009 MarkFromRootsVerifyClosure markFromRootsClosure(this, _span, 3010 verification_mark_bm(), markBitMap(), verification_mark_stack()); 3011 assert(_restart_addr == NULL, "Expected pre-condition"); 3012 verification_mark_bm()->iterate(&markFromRootsClosure); 3013 while (_restart_addr != NULL) { 3014 // Deal with stack overflow: by restarting at the indicated 3015 // address. 3016 HeapWord* ra = _restart_addr; 3017 markFromRootsClosure.reset(ra); 3018 _restart_addr = NULL; 3019 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end()); 3020 } 3021 assert(verification_mark_stack()->isEmpty(), "Should have been drained"); 3022 verify_work_stacks_empty(); 3023 3024 VerifyKlassOopsKlassClosure verify_klass_oops(verification_mark_bm()); 3025 ClassLoaderDataGraph::classes_do(&verify_klass_oops); 3026 3027 // Marking completed -- now verify that each bit marked in 3028 // verification_mark_bm() is also marked in markBitMap(); flag all 3029 // errors by printing corresponding objects. 3030 VerifyMarkedClosure vcl(markBitMap()); 3031 verification_mark_bm()->iterate(&vcl); 3032 assert(!vcl.failed(), "Else verification above should not have succeeded"); 3033 } 3034 3035 void ConcurrentMarkSweepGeneration::save_marks() { 3036 // delegate to CMS space 3037 cmsSpace()->save_marks(); 3038 for (uint i = 0; i < ParallelGCThreads; i++) { 3039 _par_gc_thread_states[i]->promo.startTrackingPromotions(); 3040 } 3041 } 3042 3043 bool ConcurrentMarkSweepGeneration::no_allocs_since_save_marks() { 3044 return cmsSpace()->no_allocs_since_save_marks(); 3045 } 3046 3047 #define CMS_SINCE_SAVE_MARKS_DEFN(OopClosureType, nv_suffix) \ 3048 \ 3049 void ConcurrentMarkSweepGeneration:: \ 3050 oop_since_save_marks_iterate##nv_suffix(OopClosureType* cl) { \ 3051 cl->set_generation(this); \ 3052 cmsSpace()->oop_since_save_marks_iterate##nv_suffix(cl); \ 3053 cl->reset_generation(); \ 3054 save_marks(); \ 3055 } 3056 3057 ALL_SINCE_SAVE_MARKS_CLOSURES(CMS_SINCE_SAVE_MARKS_DEFN) 3058 3059 void 3060 ConcurrentMarkSweepGeneration::object_iterate_since_last_GC(ObjectClosure* blk) 3061 { 3062 // Not currently implemented; need to do the following. -- ysr. 3063 // dld -- I think that is used for some sort of allocation profiler. So it 3064 // really means the objects allocated by the mutator since the last 3065 // GC. We could potentially implement this cheaply by recording only 3066 // the direct allocations in a side data structure. 3067 // 3068 // I think we probably ought not to be required to support these 3069 // iterations at any arbitrary point; I think there ought to be some 3070 // call to enable/disable allocation profiling in a generation/space, 3071 // and the iterator ought to return the objects allocated in the 3072 // gen/space since the enable call, or the last iterator call (which 3073 // will probably be at a GC.) That way, for gens like CM&S that would 3074 // require some extra data structure to support this, we only pay the 3075 // cost when it's in use... 3076 cmsSpace()->object_iterate_since_last_GC(blk); 3077 } 3078 3079 void 3080 ConcurrentMarkSweepGeneration::younger_refs_iterate(OopsInGenClosure* cl) { 3081 cl->set_generation(this); 3082 younger_refs_in_space_iterate(_cmsSpace, cl); 3083 cl->reset_generation(); 3084 } 3085 3086 void 3087 ConcurrentMarkSweepGeneration::oop_iterate(MemRegion mr, ExtendedOopClosure* cl) { 3088 if (freelistLock()->owned_by_self()) { 3089 Generation::oop_iterate(mr, cl); 3090 } else { 3091 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 3092 Generation::oop_iterate(mr, cl); 3093 } 3094 } 3095 3096 void 3097 ConcurrentMarkSweepGeneration::oop_iterate(ExtendedOopClosure* cl) { 3098 if (freelistLock()->owned_by_self()) { 3099 Generation::oop_iterate(cl); 3100 } else { 3101 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 3102 Generation::oop_iterate(cl); 3103 } 3104 } 3105 3106 void 3107 ConcurrentMarkSweepGeneration::object_iterate(ObjectClosure* cl) { 3108 if (freelistLock()->owned_by_self()) { 3109 Generation::object_iterate(cl); 3110 } else { 3111 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 3112 Generation::object_iterate(cl); 3113 } 3114 } 3115 3116 void 3117 ConcurrentMarkSweepGeneration::safe_object_iterate(ObjectClosure* cl) { 3118 if (freelistLock()->owned_by_self()) { 3119 Generation::safe_object_iterate(cl); 3120 } else { 3121 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 3122 Generation::safe_object_iterate(cl); 3123 } 3124 } 3125 3126 void 3127 ConcurrentMarkSweepGeneration::post_compact() { 3128 } 3129 3130 void 3131 ConcurrentMarkSweepGeneration::prepare_for_verify() { 3132 // Fix the linear allocation blocks to look like free blocks. 3133 3134 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those 3135 // are not called when the heap is verified during universe initialization and 3136 // at vm shutdown. 3137 if (freelistLock()->owned_by_self()) { 3138 cmsSpace()->prepare_for_verify(); 3139 } else { 3140 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag); 3141 cmsSpace()->prepare_for_verify(); 3142 } 3143 } 3144 3145 void 3146 ConcurrentMarkSweepGeneration::verify() { 3147 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those 3148 // are not called when the heap is verified during universe initialization and 3149 // at vm shutdown. 3150 if (freelistLock()->owned_by_self()) { 3151 cmsSpace()->verify(); 3152 } else { 3153 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag); 3154 cmsSpace()->verify(); 3155 } 3156 } 3157 3158 void CMSCollector::verify() { 3159 _cmsGen->verify(); 3160 } 3161 3162 #ifndef PRODUCT 3163 bool CMSCollector::overflow_list_is_empty() const { 3164 assert(_num_par_pushes >= 0, "Inconsistency"); 3165 if (_overflow_list == NULL) { 3166 assert(_num_par_pushes == 0, "Inconsistency"); 3167 } 3168 return _overflow_list == NULL; 3169 } 3170 3171 // The methods verify_work_stacks_empty() and verify_overflow_empty() 3172 // merely consolidate assertion checks that appear to occur together frequently. 3173 void CMSCollector::verify_work_stacks_empty() const { 3174 assert(_markStack.isEmpty(), "Marking stack should be empty"); 3175 assert(overflow_list_is_empty(), "Overflow list should be empty"); 3176 } 3177 3178 void CMSCollector::verify_overflow_empty() const { 3179 assert(overflow_list_is_empty(), "Overflow list should be empty"); 3180 assert(no_preserved_marks(), "No preserved marks"); 3181 } 3182 #endif // PRODUCT 3183 3184 // Decide if we want to enable class unloading as part of the 3185 // ensuing concurrent GC cycle. We will collect and 3186 // unload classes if it's the case that: 3187 // (1) an explicit gc request has been made and the flag 3188 // ExplicitGCInvokesConcurrentAndUnloadsClasses is set, OR 3189 // (2) (a) class unloading is enabled at the command line, and 3190 // (b) old gen is getting really full 3191 // NOTE: Provided there is no change in the state of the heap between 3192 // calls to this method, it should have idempotent results. Moreover, 3193 // its results should be monotonically increasing (i.e. going from 0 to 1, 3194 // but not 1 to 0) between successive calls between which the heap was 3195 // not collected. For the implementation below, it must thus rely on 3196 // the property that concurrent_cycles_since_last_unload() 3197 // will not decrease unless a collection cycle happened and that 3198 // _cmsGen->is_too_full() are 3199 // themselves also monotonic in that sense. See check_monotonicity() 3200 // below. 3201 void CMSCollector::update_should_unload_classes() { 3202 _should_unload_classes = false; 3203 // Condition 1 above 3204 if (_full_gc_requested && ExplicitGCInvokesConcurrentAndUnloadsClasses) { 3205 _should_unload_classes = true; 3206 } else if (CMSClassUnloadingEnabled) { // Condition 2.a above 3207 // Disjuncts 2.b.(i,ii,iii) above 3208 _should_unload_classes = (concurrent_cycles_since_last_unload() >= 3209 CMSClassUnloadingMaxInterval) 3210 || _cmsGen->is_too_full(); 3211 } 3212 } 3213 3214 bool ConcurrentMarkSweepGeneration::is_too_full() const { 3215 bool res = should_concurrent_collect(); 3216 res = res && (occupancy() > (double)CMSIsTooFullPercentage/100.0); 3217 return res; 3218 } 3219 3220 void CMSCollector::setup_cms_unloading_and_verification_state() { 3221 const bool should_verify = VerifyBeforeGC || VerifyAfterGC || VerifyDuringGC 3222 || VerifyBeforeExit; 3223 const int rso = SharedHeap::SO_Strings | SharedHeap::SO_CodeCache; 3224 3225 if (should_unload_classes()) { // Should unload classes this cycle 3226 remove_root_scanning_option(rso); // Shrink the root set appropriately 3227 set_verifying(should_verify); // Set verification state for this cycle 3228 return; // Nothing else needs to be done at this time 3229 } 3230 3231 // Not unloading classes this cycle 3232 assert(!should_unload_classes(), "Inconsitency!"); 3233 if ((!verifying() || unloaded_classes_last_cycle()) && should_verify) { 3234 // Include symbols, strings and code cache elements to prevent their resurrection. 3235 add_root_scanning_option(rso); 3236 set_verifying(true); 3237 } else if (verifying() && !should_verify) { 3238 // We were verifying, but some verification flags got disabled. 3239 set_verifying(false); 3240 // Exclude symbols, strings and code cache elements from root scanning to 3241 // reduce IM and RM pauses. 3242 remove_root_scanning_option(rso); 3243 } 3244 } 3245 3246 3247 #ifndef PRODUCT 3248 HeapWord* CMSCollector::block_start(const void* p) const { 3249 const HeapWord* addr = (HeapWord*)p; 3250 if (_span.contains(p)) { 3251 if (_cmsGen->cmsSpace()->is_in_reserved(addr)) { 3252 return _cmsGen->cmsSpace()->block_start(p); 3253 } 3254 } 3255 return NULL; 3256 } 3257 #endif 3258 3259 HeapWord* 3260 ConcurrentMarkSweepGeneration::expand_and_allocate(size_t word_size, 3261 bool tlab, 3262 bool parallel) { 3263 CMSSynchronousYieldRequest yr; 3264 assert(!tlab, "Can't deal with TLAB allocation"); 3265 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 3266 expand(word_size*HeapWordSize, MinHeapDeltaBytes, 3267 CMSExpansionCause::_satisfy_allocation); 3268 if (GCExpandToAllocateDelayMillis > 0) { 3269 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 3270 } 3271 return have_lock_and_allocate(word_size, tlab); 3272 } 3273 3274 // YSR: All of this generation expansion/shrinking stuff is an exact copy of 3275 // OneContigSpaceCardGeneration, which makes me wonder if we should move this 3276 // to CardGeneration and share it... 3277 bool ConcurrentMarkSweepGeneration::expand(size_t bytes, size_t expand_bytes) { 3278 return CardGeneration::expand(bytes, expand_bytes); 3279 } 3280 3281 void ConcurrentMarkSweepGeneration::expand(size_t bytes, size_t expand_bytes, 3282 CMSExpansionCause::Cause cause) 3283 { 3284 3285 bool success = expand(bytes, expand_bytes); 3286 3287 // remember why we expanded; this information is used 3288 // by shouldConcurrentCollect() when making decisions on whether to start 3289 // a new CMS cycle. 3290 if (success) { 3291 set_expansion_cause(cause); 3292 if (PrintGCDetails && Verbose) { 3293 gclog_or_tty->print_cr("Expanded CMS gen for %s", 3294 CMSExpansionCause::to_string(cause)); 3295 } 3296 } 3297 } 3298 3299 HeapWord* ConcurrentMarkSweepGeneration::expand_and_par_lab_allocate(CMSParGCThreadState* ps, size_t word_sz) { 3300 HeapWord* res = NULL; 3301 MutexLocker x(ParGCRareEvent_lock); 3302 while (true) { 3303 // Expansion by some other thread might make alloc OK now: 3304 res = ps->lab.alloc(word_sz); 3305 if (res != NULL) return res; 3306 // If there's not enough expansion space available, give up. 3307 if (_virtual_space.uncommitted_size() < (word_sz * HeapWordSize)) { 3308 return NULL; 3309 } 3310 // Otherwise, we try expansion. 3311 expand(word_sz*HeapWordSize, MinHeapDeltaBytes, 3312 CMSExpansionCause::_allocate_par_lab); 3313 // Now go around the loop and try alloc again; 3314 // A competing par_promote might beat us to the expansion space, 3315 // so we may go around the loop again if promotion fails agaion. 3316 if (GCExpandToAllocateDelayMillis > 0) { 3317 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 3318 } 3319 } 3320 } 3321 3322 3323 bool ConcurrentMarkSweepGeneration::expand_and_ensure_spooling_space( 3324 PromotionInfo* promo) { 3325 MutexLocker x(ParGCRareEvent_lock); 3326 size_t refill_size_bytes = promo->refillSize() * HeapWordSize; 3327 while (true) { 3328 // Expansion by some other thread might make alloc OK now: 3329 if (promo->ensure_spooling_space()) { 3330 assert(promo->has_spooling_space(), 3331 "Post-condition of successful ensure_spooling_space()"); 3332 return true; 3333 } 3334 // If there's not enough expansion space available, give up. 3335 if (_virtual_space.uncommitted_size() < refill_size_bytes) { 3336 return false; 3337 } 3338 // Otherwise, we try expansion. 3339 expand(refill_size_bytes, MinHeapDeltaBytes, 3340 CMSExpansionCause::_allocate_par_spooling_space); 3341 // Now go around the loop and try alloc again; 3342 // A competing allocation might beat us to the expansion space, 3343 // so we may go around the loop again if allocation fails again. 3344 if (GCExpandToAllocateDelayMillis > 0) { 3345 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 3346 } 3347 } 3348 } 3349 3350 3351 void ConcurrentMarkSweepGeneration::shrink_by(size_t bytes) { 3352 assert_locked_or_safepoint(ExpandHeap_lock); 3353 // Shrink committed space 3354 _virtual_space.shrink_by(bytes); 3355 // Shrink space; this also shrinks the space's BOT 3356 _cmsSpace->set_end((HeapWord*) _virtual_space.high()); 3357 size_t new_word_size = heap_word_size(_cmsSpace->capacity()); 3358 // Shrink the shared block offset array 3359 _bts->resize(new_word_size); 3360 MemRegion mr(_cmsSpace->bottom(), new_word_size); 3361 // Shrink the card table 3362 Universe::heap()->barrier_set()->resize_covered_region(mr); 3363 3364 if (Verbose && PrintGC) { 3365 size_t new_mem_size = _virtual_space.committed_size(); 3366 size_t old_mem_size = new_mem_size + bytes; 3367 gclog_or_tty->print_cr("Shrinking %s from " SIZE_FORMAT "K to " SIZE_FORMAT "K", 3368 name(), old_mem_size/K, new_mem_size/K); 3369 } 3370 } 3371 3372 void ConcurrentMarkSweepGeneration::shrink(size_t bytes) { 3373 assert_locked_or_safepoint(Heap_lock); 3374 size_t size = ReservedSpace::page_align_size_down(bytes); 3375 if (size > 0) { 3376 shrink_by(size); 3377 } 3378 } 3379 3380 bool ConcurrentMarkSweepGeneration::grow_by(size_t bytes) { 3381 assert_locked_or_safepoint(Heap_lock); 3382 bool result = _virtual_space.expand_by(bytes); 3383 if (result) { 3384 HeapWord* old_end = _cmsSpace->end(); 3385 size_t new_word_size = 3386 heap_word_size(_virtual_space.committed_size()); 3387 MemRegion mr(_cmsSpace->bottom(), new_word_size); 3388 _bts->resize(new_word_size); // resize the block offset shared array 3389 Universe::heap()->barrier_set()->resize_covered_region(mr); 3390 // Hmmmm... why doesn't CFLS::set_end verify locking? 3391 // This is quite ugly; FIX ME XXX 3392 _cmsSpace->assert_locked(freelistLock()); 3393 _cmsSpace->set_end((HeapWord*)_virtual_space.high()); 3394 3395 // update the space and generation capacity counters 3396 if (UsePerfData) { 3397 _space_counters->update_capacity(); 3398 _gen_counters->update_all(); 3399 } 3400 3401 if (Verbose && PrintGC) { 3402 size_t new_mem_size = _virtual_space.committed_size(); 3403 size_t old_mem_size = new_mem_size - bytes; 3404 gclog_or_tty->print_cr("Expanding %s from " SIZE_FORMAT "K by " SIZE_FORMAT "K to " SIZE_FORMAT "K", 3405 name(), old_mem_size/K, bytes/K, new_mem_size/K); 3406 } 3407 } 3408 return result; 3409 } 3410 3411 bool ConcurrentMarkSweepGeneration::grow_to_reserved() { 3412 assert_locked_or_safepoint(Heap_lock); 3413 bool success = true; 3414 const size_t remaining_bytes = _virtual_space.uncommitted_size(); 3415 if (remaining_bytes > 0) { 3416 success = grow_by(remaining_bytes); 3417 DEBUG_ONLY(if (!success) warning("grow to reserved failed");) 3418 } 3419 return success; 3420 } 3421 3422 void ConcurrentMarkSweepGeneration::shrink_free_list_by(size_t bytes) { 3423 assert_locked_or_safepoint(Heap_lock); 3424 assert_lock_strong(freelistLock()); 3425 if (PrintGCDetails && Verbose) { 3426 warning("Shrinking of CMS not yet implemented"); 3427 } 3428 return; 3429 } 3430 3431 3432 // Simple ctor/dtor wrapper for accounting & timer chores around concurrent 3433 // phases. 3434 class CMSPhaseAccounting: public StackObj { 3435 public: 3436 CMSPhaseAccounting(CMSCollector *collector, 3437 const char *phase, 3438 bool print_cr = true); 3439 ~CMSPhaseAccounting(); 3440 3441 private: 3442 CMSCollector *_collector; 3443 const char *_phase; 3444 elapsedTimer _wallclock; 3445 bool _print_cr; 3446 3447 public: 3448 // Not MT-safe; so do not pass around these StackObj's 3449 // where they may be accessed by other threads. 3450 jlong wallclock_millis() { 3451 assert(_wallclock.is_active(), "Wall clock should not stop"); 3452 _wallclock.stop(); // to record time 3453 jlong ret = _wallclock.milliseconds(); 3454 _wallclock.start(); // restart 3455 return ret; 3456 } 3457 }; 3458 3459 CMSPhaseAccounting::CMSPhaseAccounting(CMSCollector *collector, 3460 const char *phase, 3461 bool print_cr) : 3462 _collector(collector), _phase(phase), _print_cr(print_cr) { 3463 3464 if (PrintCMSStatistics != 0) { 3465 _collector->resetYields(); 3466 } 3467 if (PrintGCDetails) { 3468 gclog_or_tty->date_stamp(PrintGCDateStamps); 3469 gclog_or_tty->stamp(PrintGCTimeStamps); 3470 gclog_or_tty->print_cr("[%s-concurrent-%s-start]", 3471 _collector->cmsGen()->short_name(), _phase); 3472 } 3473 _collector->resetTimer(); 3474 _wallclock.start(); 3475 _collector->startTimer(); 3476 } 3477 3478 CMSPhaseAccounting::~CMSPhaseAccounting() { 3479 assert(_wallclock.is_active(), "Wall clock should not have stopped"); 3480 _collector->stopTimer(); 3481 _wallclock.stop(); 3482 if (PrintGCDetails) { 3483 gclog_or_tty->date_stamp(PrintGCDateStamps); 3484 gclog_or_tty->stamp(PrintGCTimeStamps); 3485 gclog_or_tty->print("[%s-concurrent-%s: %3.3f/%3.3f secs]", 3486 _collector->cmsGen()->short_name(), 3487 _phase, _collector->timerValue(), _wallclock.seconds()); 3488 if (_print_cr) { 3489 gclog_or_tty->print_cr(""); 3490 } 3491 if (PrintCMSStatistics != 0) { 3492 gclog_or_tty->print_cr(" (CMS-concurrent-%s yielded %d times)", _phase, 3493 _collector->yields()); 3494 } 3495 } 3496 } 3497 3498 // CMS work 3499 3500 // Checkpoint the roots into this generation from outside 3501 // this generation. [Note this initial checkpoint need only 3502 // be approximate -- we'll do a catch up phase subsequently.] 3503 void CMSCollector::checkpointRootsInitial(bool asynch) { 3504 assert(_collectorState == InitialMarking, "Wrong collector state"); 3505 check_correct_thread_executing(); 3506 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 3507 3508 ReferenceProcessor* rp = ref_processor(); 3509 SpecializationStats::clear(); 3510 assert(_restart_addr == NULL, "Control point invariant"); 3511 if (asynch) { 3512 // acquire locks for subsequent manipulations 3513 MutexLockerEx x(bitMapLock(), 3514 Mutex::_no_safepoint_check_flag); 3515 checkpointRootsInitialWork(asynch); 3516 // enable ("weak") refs discovery 3517 rp->enable_discovery(true /*verify_disabled*/, true /*check_no_refs*/); 3518 _collectorState = Marking; 3519 } else { 3520 // (Weak) Refs discovery: this is controlled from genCollectedHeap::do_collection 3521 // which recognizes if we are a CMS generation, and doesn't try to turn on 3522 // discovery; verify that they aren't meddling. 3523 assert(!rp->discovery_is_atomic(), 3524 "incorrect setting of discovery predicate"); 3525 assert(!rp->discovery_enabled(), "genCollectedHeap shouldn't control " 3526 "ref discovery for this generation kind"); 3527 // already have locks 3528 checkpointRootsInitialWork(asynch); 3529 // now enable ("weak") refs discovery 3530 rp->enable_discovery(true /*verify_disabled*/, false /*verify_no_refs*/); 3531 _collectorState = Marking; 3532 } 3533 SpecializationStats::print(); 3534 } 3535 3536 void CMSCollector::checkpointRootsInitialWork(bool asynch) { 3537 assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped"); 3538 assert(_collectorState == InitialMarking, "just checking"); 3539 3540 // If there has not been a GC[n-1] since last GC[n] cycle completed, 3541 // precede our marking with a collection of all 3542 // younger generations to keep floating garbage to a minimum. 3543 // XXX: we won't do this for now -- it's an optimization to be done later. 3544 3545 // already have locks 3546 assert_lock_strong(bitMapLock()); 3547 assert(_markBitMap.isAllClear(), "was reset at end of previous cycle"); 3548 3549 // Setup the verification and class unloading state for this 3550 // CMS collection cycle. 3551 setup_cms_unloading_and_verification_state(); 3552 3553 NOT_PRODUCT(TraceTime t("\ncheckpointRootsInitialWork", 3554 PrintGCDetails && Verbose, true, gclog_or_tty);) 3555 if (UseAdaptiveSizePolicy) { 3556 size_policy()->checkpoint_roots_initial_begin(); 3557 } 3558 3559 // Reset all the PLAB chunk arrays if necessary. 3560 if (_survivor_plab_array != NULL && !CMSPLABRecordAlways) { 3561 reset_survivor_plab_arrays(); 3562 } 3563 3564 ResourceMark rm; 3565 HandleMark hm; 3566 3567 FalseClosure falseClosure; 3568 // In the case of a synchronous collection, we will elide the 3569 // remark step, so it's important to catch all the nmethod oops 3570 // in this step. 3571 // The final 'true' flag to gen_process_strong_roots will ensure this. 3572 // If 'async' is true, we can relax the nmethod tracing. 3573 MarkRefsIntoClosure notOlder(_span, &_markBitMap); 3574 GenCollectedHeap* gch = GenCollectedHeap::heap(); 3575 3576 verify_work_stacks_empty(); 3577 verify_overflow_empty(); 3578 3579 gch->ensure_parsability(false); // fill TLABs, but no need to retire them 3580 // Update the saved marks which may affect the root scans. 3581 gch->save_marks(); 3582 3583 // weak reference processing has not started yet. 3584 ref_processor()->set_enqueuing_is_done(false); 3585 3586 // Need to remember all newly created CLDs, 3587 // so that we can guarantee that the remark finds them. 3588 ClassLoaderDataGraph::remember_new_clds(true); 3589 3590 // Whenever a CLD is found, it will be claimed before proceeding to mark 3591 // the klasses. The claimed marks need to be cleared before marking starts. 3592 ClassLoaderDataGraph::clear_claimed_marks(); 3593 3594 CMKlassClosure klass_closure(¬Older); 3595 { 3596 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;) 3597 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 3598 gch->gen_process_strong_roots(_cmsGen->level(), 3599 true, // younger gens are roots 3600 true, // activate StrongRootsScope 3601 false, // not scavenging 3602 SharedHeap::ScanningOption(roots_scanning_options()), 3603 ¬Older, 3604 true, // walk all of code cache if (so & SO_CodeCache) 3605 NULL, 3606 &klass_closure); 3607 } 3608 3609 // Clear mod-union table; it will be dirtied in the prologue of 3610 // CMS generation per each younger generation collection. 3611 3612 assert(_modUnionTable.isAllClear(), 3613 "Was cleared in most recent final checkpoint phase" 3614 " or no bits are set in the gc_prologue before the start of the next " 3615 "subsequent marking phase."); 3616 3617 assert(_ct->klass_rem_set()->mod_union_is_clear(), "Must be"); 3618 3619 // Save the end of the used_region of the constituent generations 3620 // to be used to limit the extent of sweep in each generation. 3621 save_sweep_limits(); 3622 if (UseAdaptiveSizePolicy) { 3623 size_policy()->checkpoint_roots_initial_end(gch->gc_cause()); 3624 } 3625 verify_overflow_empty(); 3626 } 3627 3628 bool CMSCollector::markFromRoots(bool asynch) { 3629 // we might be tempted to assert that: 3630 // assert(asynch == !SafepointSynchronize::is_at_safepoint(), 3631 // "inconsistent argument?"); 3632 // However that wouldn't be right, because it's possible that 3633 // a safepoint is indeed in progress as a younger generation 3634 // stop-the-world GC happens even as we mark in this generation. 3635 assert(_collectorState == Marking, "inconsistent state?"); 3636 check_correct_thread_executing(); 3637 verify_overflow_empty(); 3638 3639 bool res; 3640 if (asynch) { 3641 3642 // Start the timers for adaptive size policy for the concurrent phases 3643 // Do it here so that the foreground MS can use the concurrent 3644 // timer since a foreground MS might has the sweep done concurrently 3645 // or STW. 3646 if (UseAdaptiveSizePolicy) { 3647 size_policy()->concurrent_marking_begin(); 3648 } 3649 3650 // Weak ref discovery note: We may be discovering weak 3651 // refs in this generation concurrent (but interleaved) with 3652 // weak ref discovery by a younger generation collector. 3653 3654 CMSTokenSyncWithLocks ts(true, bitMapLock()); 3655 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 3656 CMSPhaseAccounting pa(this, "mark", !PrintGCDetails); 3657 res = markFromRootsWork(asynch); 3658 if (res) { 3659 _collectorState = Precleaning; 3660 } else { // We failed and a foreground collection wants to take over 3661 assert(_foregroundGCIsActive, "internal state inconsistency"); 3662 assert(_restart_addr == NULL, "foreground will restart from scratch"); 3663 if (PrintGCDetails) { 3664 gclog_or_tty->print_cr("bailing out to foreground collection"); 3665 } 3666 } 3667 if (UseAdaptiveSizePolicy) { 3668 size_policy()->concurrent_marking_end(); 3669 } 3670 } else { 3671 assert(SafepointSynchronize::is_at_safepoint(), 3672 "inconsistent with asynch == false"); 3673 if (UseAdaptiveSizePolicy) { 3674 size_policy()->ms_collection_marking_begin(); 3675 } 3676 // already have locks 3677 res = markFromRootsWork(asynch); 3678 _collectorState = FinalMarking; 3679 if (UseAdaptiveSizePolicy) { 3680 GenCollectedHeap* gch = GenCollectedHeap::heap(); 3681 size_policy()->ms_collection_marking_end(gch->gc_cause()); 3682 } 3683 } 3684 verify_overflow_empty(); 3685 return res; 3686 } 3687 3688 bool CMSCollector::markFromRootsWork(bool asynch) { 3689 // iterate over marked bits in bit map, doing a full scan and mark 3690 // from these roots using the following algorithm: 3691 // . if oop is to the right of the current scan pointer, 3692 // mark corresponding bit (we'll process it later) 3693 // . else (oop is to left of current scan pointer) 3694 // push oop on marking stack 3695 // . drain the marking stack 3696 3697 // Note that when we do a marking step we need to hold the 3698 // bit map lock -- recall that direct allocation (by mutators) 3699 // and promotion (by younger generation collectors) is also 3700 // marking the bit map. [the so-called allocate live policy.] 3701 // Because the implementation of bit map marking is not 3702 // robust wrt simultaneous marking of bits in the same word, 3703 // we need to make sure that there is no such interference 3704 // between concurrent such updates. 3705 3706 // already have locks 3707 assert_lock_strong(bitMapLock()); 3708 3709 verify_work_stacks_empty(); 3710 verify_overflow_empty(); 3711 bool result = false; 3712 if (CMSConcurrentMTEnabled && ConcGCThreads > 0) { 3713 result = do_marking_mt(asynch); 3714 } else { 3715 result = do_marking_st(asynch); 3716 } 3717 return result; 3718 } 3719 3720 // Forward decl 3721 class CMSConcMarkingTask; 3722 3723 class CMSConcMarkingTerminator: public ParallelTaskTerminator { 3724 CMSCollector* _collector; 3725 CMSConcMarkingTask* _task; 3726 public: 3727 virtual void yield(); 3728 3729 // "n_threads" is the number of threads to be terminated. 3730 // "queue_set" is a set of work queues of other threads. 3731 // "collector" is the CMS collector associated with this task terminator. 3732 // "yield" indicates whether we need the gang as a whole to yield. 3733 CMSConcMarkingTerminator(int n_threads, TaskQueueSetSuper* queue_set, CMSCollector* collector) : 3734 ParallelTaskTerminator(n_threads, queue_set), 3735 _collector(collector) { } 3736 3737 void set_task(CMSConcMarkingTask* task) { 3738 _task = task; 3739 } 3740 }; 3741 3742 class CMSConcMarkingTerminatorTerminator: public TerminatorTerminator { 3743 CMSConcMarkingTask* _task; 3744 public: 3745 bool should_exit_termination(); 3746 void set_task(CMSConcMarkingTask* task) { 3747 _task = task; 3748 } 3749 }; 3750 3751 // MT Concurrent Marking Task 3752 class CMSConcMarkingTask: public YieldingFlexibleGangTask { 3753 CMSCollector* _collector; 3754 int _n_workers; // requested/desired # workers 3755 bool _asynch; 3756 bool _result; 3757 CompactibleFreeListSpace* _cms_space; 3758 char _pad_front[64]; // padding to ... 3759 HeapWord* _global_finger; // ... avoid sharing cache line 3760 char _pad_back[64]; 3761 HeapWord* _restart_addr; 3762 3763 // Exposed here for yielding support 3764 Mutex* const _bit_map_lock; 3765 3766 // The per thread work queues, available here for stealing 3767 OopTaskQueueSet* _task_queues; 3768 3769 // Termination (and yielding) support 3770 CMSConcMarkingTerminator _term; 3771 CMSConcMarkingTerminatorTerminator _term_term; 3772 3773 public: 3774 CMSConcMarkingTask(CMSCollector* collector, 3775 CompactibleFreeListSpace* cms_space, 3776 bool asynch, 3777 YieldingFlexibleWorkGang* workers, 3778 OopTaskQueueSet* task_queues): 3779 YieldingFlexibleGangTask("Concurrent marking done multi-threaded"), 3780 _collector(collector), 3781 _cms_space(cms_space), 3782 _asynch(asynch), _n_workers(0), _result(true), 3783 _task_queues(task_queues), 3784 _term(_n_workers, task_queues, _collector), 3785 _bit_map_lock(collector->bitMapLock()) 3786 { 3787 _requested_size = _n_workers; 3788 _term.set_task(this); 3789 _term_term.set_task(this); 3790 _restart_addr = _global_finger = _cms_space->bottom(); 3791 } 3792 3793 3794 OopTaskQueueSet* task_queues() { return _task_queues; } 3795 3796 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 3797 3798 HeapWord** global_finger_addr() { return &_global_finger; } 3799 3800 CMSConcMarkingTerminator* terminator() { return &_term; } 3801 3802 virtual void set_for_termination(int active_workers) { 3803 terminator()->reset_for_reuse(active_workers); 3804 } 3805 3806 void work(uint worker_id); 3807 bool should_yield() { 3808 return ConcurrentMarkSweepThread::should_yield() 3809 && !_collector->foregroundGCIsActive() 3810 && _asynch; 3811 } 3812 3813 virtual void coordinator_yield(); // stuff done by coordinator 3814 bool result() { return _result; } 3815 3816 void reset(HeapWord* ra) { 3817 assert(_global_finger >= _cms_space->end(), "Postcondition of ::work(i)"); 3818 _restart_addr = _global_finger = ra; 3819 _term.reset_for_reuse(); 3820 } 3821 3822 static bool get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 3823 OopTaskQueue* work_q); 3824 3825 private: 3826 void do_scan_and_mark(int i, CompactibleFreeListSpace* sp); 3827 void do_work_steal(int i); 3828 void bump_global_finger(HeapWord* f); 3829 }; 3830 3831 bool CMSConcMarkingTerminatorTerminator::should_exit_termination() { 3832 assert(_task != NULL, "Error"); 3833 return _task->yielding(); 3834 // Note that we do not need the disjunct || _task->should_yield() above 3835 // because we want terminating threads to yield only if the task 3836 // is already in the midst of yielding, which happens only after at least one 3837 // thread has yielded. 3838 } 3839 3840 void CMSConcMarkingTerminator::yield() { 3841 if (_task->should_yield()) { 3842 _task->yield(); 3843 } else { 3844 ParallelTaskTerminator::yield(); 3845 } 3846 } 3847 3848 //////////////////////////////////////////////////////////////// 3849 // Concurrent Marking Algorithm Sketch 3850 //////////////////////////////////////////////////////////////// 3851 // Until all tasks exhausted (both spaces): 3852 // -- claim next available chunk 3853 // -- bump global finger via CAS 3854 // -- find first object that starts in this chunk 3855 // and start scanning bitmap from that position 3856 // -- scan marked objects for oops 3857 // -- CAS-mark target, and if successful: 3858 // . if target oop is above global finger (volatile read) 3859 // nothing to do 3860 // . if target oop is in chunk and above local finger 3861 // then nothing to do 3862 // . else push on work-queue 3863 // -- Deal with possible overflow issues: 3864 // . local work-queue overflow causes stuff to be pushed on 3865 // global (common) overflow queue 3866 // . always first empty local work queue 3867 // . then get a batch of oops from global work queue if any 3868 // . then do work stealing 3869 // -- When all tasks claimed (both spaces) 3870 // and local work queue empty, 3871 // then in a loop do: 3872 // . check global overflow stack; steal a batch of oops and trace 3873 // . try to steal from other threads oif GOS is empty 3874 // . if neither is available, offer termination 3875 // -- Terminate and return result 3876 // 3877 void CMSConcMarkingTask::work(uint worker_id) { 3878 elapsedTimer _timer; 3879 ResourceMark rm; 3880 HandleMark hm; 3881 3882 DEBUG_ONLY(_collector->verify_overflow_empty();) 3883 3884 // Before we begin work, our work queue should be empty 3885 assert(work_queue(worker_id)->size() == 0, "Expected to be empty"); 3886 // Scan the bitmap covering _cms_space, tracing through grey objects. 3887 _timer.start(); 3888 do_scan_and_mark(worker_id, _cms_space); 3889 _timer.stop(); 3890 if (PrintCMSStatistics != 0) { 3891 gclog_or_tty->print_cr("Finished cms space scanning in %dth thread: %3.3f sec", 3892 worker_id, _timer.seconds()); 3893 // XXX: need xxx/xxx type of notation, two timers 3894 } 3895 3896 // ... do work stealing 3897 _timer.reset(); 3898 _timer.start(); 3899 do_work_steal(worker_id); 3900 _timer.stop(); 3901 if (PrintCMSStatistics != 0) { 3902 gclog_or_tty->print_cr("Finished work stealing in %dth thread: %3.3f sec", 3903 worker_id, _timer.seconds()); 3904 // XXX: need xxx/xxx type of notation, two timers 3905 } 3906 assert(_collector->_markStack.isEmpty(), "Should have been emptied"); 3907 assert(work_queue(worker_id)->size() == 0, "Should have been emptied"); 3908 // Note that under the current task protocol, the 3909 // following assertion is true even of the spaces 3910 // expanded since the completion of the concurrent 3911 // marking. XXX This will likely change under a strict 3912 // ABORT semantics. 3913 // After perm removal the comparison was changed to 3914 // greater than or equal to from strictly greater than. 3915 // Before perm removal the highest address sweep would 3916 // have been at the end of perm gen but now is at the 3917 // end of the tenured gen. 3918 assert(_global_finger >= _cms_space->end(), 3919 "All tasks have been completed"); 3920 DEBUG_ONLY(_collector->verify_overflow_empty();) 3921 } 3922 3923 void CMSConcMarkingTask::bump_global_finger(HeapWord* f) { 3924 HeapWord* read = _global_finger; 3925 HeapWord* cur = read; 3926 while (f > read) { 3927 cur = read; 3928 read = (HeapWord*) Atomic::cmpxchg_ptr(f, &_global_finger, cur); 3929 if (cur == read) { 3930 // our cas succeeded 3931 assert(_global_finger >= f, "protocol consistency"); 3932 break; 3933 } 3934 } 3935 } 3936 3937 // This is really inefficient, and should be redone by 3938 // using (not yet available) block-read and -write interfaces to the 3939 // stack and the work_queue. XXX FIX ME !!! 3940 bool CMSConcMarkingTask::get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 3941 OopTaskQueue* work_q) { 3942 // Fast lock-free check 3943 if (ovflw_stk->length() == 0) { 3944 return false; 3945 } 3946 assert(work_q->size() == 0, "Shouldn't steal"); 3947 MutexLockerEx ml(ovflw_stk->par_lock(), 3948 Mutex::_no_safepoint_check_flag); 3949 // Grab up to 1/4 the size of the work queue 3950 size_t num = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 3951 (size_t)ParGCDesiredObjsFromOverflowList); 3952 num = MIN2(num, ovflw_stk->length()); 3953 for (int i = (int) num; i > 0; i--) { 3954 oop cur = ovflw_stk->pop(); 3955 assert(cur != NULL, "Counted wrong?"); 3956 work_q->push(cur); 3957 } 3958 return num > 0; 3959 } 3960 3961 void CMSConcMarkingTask::do_scan_and_mark(int i, CompactibleFreeListSpace* sp) { 3962 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 3963 int n_tasks = pst->n_tasks(); 3964 // We allow that there may be no tasks to do here because 3965 // we are restarting after a stack overflow. 3966 assert(pst->valid() || n_tasks == 0, "Uninitialized use?"); 3967 uint nth_task = 0; 3968 3969 HeapWord* aligned_start = sp->bottom(); 3970 if (sp->used_region().contains(_restart_addr)) { 3971 // Align down to a card boundary for the start of 0th task 3972 // for this space. 3973 aligned_start = 3974 (HeapWord*)align_size_down((uintptr_t)_restart_addr, 3975 CardTableModRefBS::card_size); 3976 } 3977 3978 size_t chunk_size = sp->marking_task_size(); 3979 while (!pst->is_task_claimed(/* reference */ nth_task)) { 3980 // Having claimed the nth task in this space, 3981 // compute the chunk that it corresponds to: 3982 MemRegion span = MemRegion(aligned_start + nth_task*chunk_size, 3983 aligned_start + (nth_task+1)*chunk_size); 3984 // Try and bump the global finger via a CAS; 3985 // note that we need to do the global finger bump 3986 // _before_ taking the intersection below, because 3987 // the task corresponding to that region will be 3988 // deemed done even if the used_region() expands 3989 // because of allocation -- as it almost certainly will 3990 // during start-up while the threads yield in the 3991 // closure below. 3992 HeapWord* finger = span.end(); 3993 bump_global_finger(finger); // atomically 3994 // There are null tasks here corresponding to chunks 3995 // beyond the "top" address of the space. 3996 span = span.intersection(sp->used_region()); 3997 if (!span.is_empty()) { // Non-null task 3998 HeapWord* prev_obj; 3999 assert(!span.contains(_restart_addr) || nth_task == 0, 4000 "Inconsistency"); 4001 if (nth_task == 0) { 4002 // For the 0th task, we'll not need to compute a block_start. 4003 if (span.contains(_restart_addr)) { 4004 // In the case of a restart because of stack overflow, 4005 // we might additionally skip a chunk prefix. 4006 prev_obj = _restart_addr; 4007 } else { 4008 prev_obj = span.start(); 4009 } 4010 } else { 4011 // We want to skip the first object because 4012 // the protocol is to scan any object in its entirety 4013 // that _starts_ in this span; a fortiori, any 4014 // object starting in an earlier span is scanned 4015 // as part of an earlier claimed task. 4016 // Below we use the "careful" version of block_start 4017 // so we do not try to navigate uninitialized objects. 4018 prev_obj = sp->block_start_careful(span.start()); 4019 // Below we use a variant of block_size that uses the 4020 // Printezis bits to avoid waiting for allocated 4021 // objects to become initialized/parsable. 4022 while (prev_obj < span.start()) { 4023 size_t sz = sp->block_size_no_stall(prev_obj, _collector); 4024 if (sz > 0) { 4025 prev_obj += sz; 4026 } else { 4027 // In this case we may end up doing a bit of redundant 4028 // scanning, but that appears unavoidable, short of 4029 // locking the free list locks; see bug 6324141. 4030 break; 4031 } 4032 } 4033 } 4034 if (prev_obj < span.end()) { 4035 MemRegion my_span = MemRegion(prev_obj, span.end()); 4036 // Do the marking work within a non-empty span -- 4037 // the last argument to the constructor indicates whether the 4038 // iteration should be incremental with periodic yields. 4039 Par_MarkFromRootsClosure cl(this, _collector, my_span, 4040 &_collector->_markBitMap, 4041 work_queue(i), 4042 &_collector->_markStack, 4043 _asynch); 4044 _collector->_markBitMap.iterate(&cl, my_span.start(), my_span.end()); 4045 } // else nothing to do for this task 4046 } // else nothing to do for this task 4047 } 4048 // We'd be tempted to assert here that since there are no 4049 // more tasks left to claim in this space, the global_finger 4050 // must exceed space->top() and a fortiori space->end(). However, 4051 // that would not quite be correct because the bumping of 4052 // global_finger occurs strictly after the claiming of a task, 4053 // so by the time we reach here the global finger may not yet 4054 // have been bumped up by the thread that claimed the last 4055 // task. 4056 pst->all_tasks_completed(); 4057 } 4058 4059 class Par_ConcMarkingClosure: public CMSOopClosure { 4060 private: 4061 CMSCollector* _collector; 4062 CMSConcMarkingTask* _task; 4063 MemRegion _span; 4064 CMSBitMap* _bit_map; 4065 CMSMarkStack* _overflow_stack; 4066 OopTaskQueue* _work_queue; 4067 protected: 4068 DO_OOP_WORK_DEFN 4069 public: 4070 Par_ConcMarkingClosure(CMSCollector* collector, CMSConcMarkingTask* task, OopTaskQueue* work_queue, 4071 CMSBitMap* bit_map, CMSMarkStack* overflow_stack): 4072 CMSOopClosure(collector->ref_processor()), 4073 _collector(collector), 4074 _task(task), 4075 _span(collector->_span), 4076 _work_queue(work_queue), 4077 _bit_map(bit_map), 4078 _overflow_stack(overflow_stack) 4079 { } 4080 virtual void do_oop(oop* p); 4081 virtual void do_oop(narrowOop* p); 4082 4083 void trim_queue(size_t max); 4084 void handle_stack_overflow(HeapWord* lost); 4085 void do_yield_check() { 4086 if (_task->should_yield()) { 4087 _task->yield(); 4088 } 4089 } 4090 }; 4091 4092 // Grey object scanning during work stealing phase -- 4093 // the salient assumption here is that any references 4094 // that are in these stolen objects being scanned must 4095 // already have been initialized (else they would not have 4096 // been published), so we do not need to check for 4097 // uninitialized objects before pushing here. 4098 void Par_ConcMarkingClosure::do_oop(oop obj) { 4099 assert(obj->is_oop_or_null(true), "expected an oop or NULL"); 4100 HeapWord* addr = (HeapWord*)obj; 4101 // Check if oop points into the CMS generation 4102 // and is not marked 4103 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 4104 // a white object ... 4105 // If we manage to "claim" the object, by being the 4106 // first thread to mark it, then we push it on our 4107 // marking stack 4108 if (_bit_map->par_mark(addr)) { // ... now grey 4109 // push on work queue (grey set) 4110 bool simulate_overflow = false; 4111 NOT_PRODUCT( 4112 if (CMSMarkStackOverflowALot && 4113 _collector->simulate_overflow()) { 4114 // simulate a stack overflow 4115 simulate_overflow = true; 4116 } 4117 ) 4118 if (simulate_overflow || 4119 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) { 4120 // stack overflow 4121 if (PrintCMSStatistics != 0) { 4122 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 4123 SIZE_FORMAT, _overflow_stack->capacity()); 4124 } 4125 // We cannot assert that the overflow stack is full because 4126 // it may have been emptied since. 4127 assert(simulate_overflow || 4128 _work_queue->size() == _work_queue->max_elems(), 4129 "Else push should have succeeded"); 4130 handle_stack_overflow(addr); 4131 } 4132 } // Else, some other thread got there first 4133 do_yield_check(); 4134 } 4135 } 4136 4137 void Par_ConcMarkingClosure::do_oop(oop* p) { Par_ConcMarkingClosure::do_oop_work(p); } 4138 void Par_ConcMarkingClosure::do_oop(narrowOop* p) { Par_ConcMarkingClosure::do_oop_work(p); } 4139 4140 void Par_ConcMarkingClosure::trim_queue(size_t max) { 4141 while (_work_queue->size() > max) { 4142 oop new_oop; 4143 if (_work_queue->pop_local(new_oop)) { 4144 assert(new_oop->is_oop(), "Should be an oop"); 4145 assert(_bit_map->isMarked((HeapWord*)new_oop), "Grey object"); 4146 assert(_span.contains((HeapWord*)new_oop), "Not in span"); 4147 new_oop->oop_iterate(this); // do_oop() above 4148 do_yield_check(); 4149 } 4150 } 4151 } 4152 4153 // Upon stack overflow, we discard (part of) the stack, 4154 // remembering the least address amongst those discarded 4155 // in CMSCollector's _restart_address. 4156 void Par_ConcMarkingClosure::handle_stack_overflow(HeapWord* lost) { 4157 // We need to do this under a mutex to prevent other 4158 // workers from interfering with the work done below. 4159 MutexLockerEx ml(_overflow_stack->par_lock(), 4160 Mutex::_no_safepoint_check_flag); 4161 // Remember the least grey address discarded 4162 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 4163 _collector->lower_restart_addr(ra); 4164 _overflow_stack->reset(); // discard stack contents 4165 _overflow_stack->expand(); // expand the stack if possible 4166 } 4167 4168 4169 void CMSConcMarkingTask::do_work_steal(int i) { 4170 OopTaskQueue* work_q = work_queue(i); 4171 oop obj_to_scan; 4172 CMSBitMap* bm = &(_collector->_markBitMap); 4173 CMSMarkStack* ovflw = &(_collector->_markStack); 4174 int* seed = _collector->hash_seed(i); 4175 Par_ConcMarkingClosure cl(_collector, this, work_q, bm, ovflw); 4176 while (true) { 4177 cl.trim_queue(0); 4178 assert(work_q->size() == 0, "Should have been emptied above"); 4179 if (get_work_from_overflow_stack(ovflw, work_q)) { 4180 // Can't assert below because the work obtained from the 4181 // overflow stack may already have been stolen from us. 4182 // assert(work_q->size() > 0, "Work from overflow stack"); 4183 continue; 4184 } else if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 4185 assert(obj_to_scan->is_oop(), "Should be an oop"); 4186 assert(bm->isMarked((HeapWord*)obj_to_scan), "Grey object"); 4187 obj_to_scan->oop_iterate(&cl); 4188 } else if (terminator()->offer_termination(&_term_term)) { 4189 assert(work_q->size() == 0, "Impossible!"); 4190 break; 4191 } else if (yielding() || should_yield()) { 4192 yield(); 4193 } 4194 } 4195 } 4196 4197 // This is run by the CMS (coordinator) thread. 4198 void CMSConcMarkingTask::coordinator_yield() { 4199 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 4200 "CMS thread should hold CMS token"); 4201 // First give up the locks, then yield, then re-lock 4202 // We should probably use a constructor/destructor idiom to 4203 // do this unlock/lock or modify the MutexUnlocker class to 4204 // serve our purpose. XXX 4205 assert_lock_strong(_bit_map_lock); 4206 _bit_map_lock->unlock(); 4207 ConcurrentMarkSweepThread::desynchronize(true); 4208 ConcurrentMarkSweepThread::acknowledge_yield_request(); 4209 _collector->stopTimer(); 4210 if (PrintCMSStatistics != 0) { 4211 _collector->incrementYields(); 4212 } 4213 _collector->icms_wait(); 4214 4215 // It is possible for whichever thread initiated the yield request 4216 // not to get a chance to wake up and take the bitmap lock between 4217 // this thread releasing it and reacquiring it. So, while the 4218 // should_yield() flag is on, let's sleep for a bit to give the 4219 // other thread a chance to wake up. The limit imposed on the number 4220 // of iterations is defensive, to avoid any unforseen circumstances 4221 // putting us into an infinite loop. Since it's always been this 4222 // (coordinator_yield()) method that was observed to cause the 4223 // problem, we are using a parameter (CMSCoordinatorYieldSleepCount) 4224 // which is by default non-zero. For the other seven methods that 4225 // also perform the yield operation, as are using a different 4226 // parameter (CMSYieldSleepCount) which is by default zero. This way we 4227 // can enable the sleeping for those methods too, if necessary. 4228 // See 6442774. 4229 // 4230 // We really need to reconsider the synchronization between the GC 4231 // thread and the yield-requesting threads in the future and we 4232 // should really use wait/notify, which is the recommended 4233 // way of doing this type of interaction. Additionally, we should 4234 // consolidate the eight methods that do the yield operation and they 4235 // are almost identical into one for better maintenability and 4236 // readability. See 6445193. 4237 // 4238 // Tony 2006.06.29 4239 for (unsigned i = 0; i < CMSCoordinatorYieldSleepCount && 4240 ConcurrentMarkSweepThread::should_yield() && 4241 !CMSCollector::foregroundGCIsActive(); ++i) { 4242 os::sleep(Thread::current(), 1, false); 4243 ConcurrentMarkSweepThread::acknowledge_yield_request(); 4244 } 4245 4246 ConcurrentMarkSweepThread::synchronize(true); 4247 _bit_map_lock->lock_without_safepoint_check(); 4248 _collector->startTimer(); 4249 } 4250 4251 bool CMSCollector::do_marking_mt(bool asynch) { 4252 assert(ConcGCThreads > 0 && conc_workers() != NULL, "precondition"); 4253 int num_workers = AdaptiveSizePolicy::calc_active_conc_workers( 4254 conc_workers()->total_workers(), 4255 conc_workers()->active_workers(), 4256 Threads::number_of_non_daemon_threads()); 4257 conc_workers()->set_active_workers(num_workers); 4258 4259 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 4260 4261 CMSConcMarkingTask tsk(this, 4262 cms_space, 4263 asynch, 4264 conc_workers(), 4265 task_queues()); 4266 4267 // Since the actual number of workers we get may be different 4268 // from the number we requested above, do we need to do anything different 4269 // below? In particular, may be we need to subclass the SequantialSubTasksDone 4270 // class?? XXX 4271 cms_space ->initialize_sequential_subtasks_for_marking(num_workers); 4272 4273 // Refs discovery is already non-atomic. 4274 assert(!ref_processor()->discovery_is_atomic(), "Should be non-atomic"); 4275 assert(ref_processor()->discovery_is_mt(), "Discovery should be MT"); 4276 conc_workers()->start_task(&tsk); 4277 while (tsk.yielded()) { 4278 tsk.coordinator_yield(); 4279 conc_workers()->continue_task(&tsk); 4280 } 4281 // If the task was aborted, _restart_addr will be non-NULL 4282 assert(tsk.completed() || _restart_addr != NULL, "Inconsistency"); 4283 while (_restart_addr != NULL) { 4284 // XXX For now we do not make use of ABORTED state and have not 4285 // yet implemented the right abort semantics (even in the original 4286 // single-threaded CMS case). That needs some more investigation 4287 // and is deferred for now; see CR# TBF. 07252005YSR. XXX 4288 assert(!CMSAbortSemantics || tsk.aborted(), "Inconsistency"); 4289 // If _restart_addr is non-NULL, a marking stack overflow 4290 // occurred; we need to do a fresh marking iteration from the 4291 // indicated restart address. 4292 if (_foregroundGCIsActive && asynch) { 4293 // We may be running into repeated stack overflows, having 4294 // reached the limit of the stack size, while making very 4295 // slow forward progress. It may be best to bail out and 4296 // let the foreground collector do its job. 4297 // Clear _restart_addr, so that foreground GC 4298 // works from scratch. This avoids the headache of 4299 // a "rescan" which would otherwise be needed because 4300 // of the dirty mod union table & card table. 4301 _restart_addr = NULL; 4302 return false; 4303 } 4304 // Adjust the task to restart from _restart_addr 4305 tsk.reset(_restart_addr); 4306 cms_space ->initialize_sequential_subtasks_for_marking(num_workers, 4307 _restart_addr); 4308 _restart_addr = NULL; 4309 // Get the workers going again 4310 conc_workers()->start_task(&tsk); 4311 while (tsk.yielded()) { 4312 tsk.coordinator_yield(); 4313 conc_workers()->continue_task(&tsk); 4314 } 4315 } 4316 assert(tsk.completed(), "Inconsistency"); 4317 assert(tsk.result() == true, "Inconsistency"); 4318 return true; 4319 } 4320 4321 bool CMSCollector::do_marking_st(bool asynch) { 4322 ResourceMark rm; 4323 HandleMark hm; 4324 4325 // Temporarily make refs discovery single threaded (non-MT) 4326 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false); 4327 MarkFromRootsClosure markFromRootsClosure(this, _span, &_markBitMap, 4328 &_markStack, CMSYield && asynch); 4329 // the last argument to iterate indicates whether the iteration 4330 // should be incremental with periodic yields. 4331 _markBitMap.iterate(&markFromRootsClosure); 4332 // If _restart_addr is non-NULL, a marking stack overflow 4333 // occurred; we need to do a fresh iteration from the 4334 // indicated restart address. 4335 while (_restart_addr != NULL) { 4336 if (_foregroundGCIsActive && asynch) { 4337 // We may be running into repeated stack overflows, having 4338 // reached the limit of the stack size, while making very 4339 // slow forward progress. It may be best to bail out and 4340 // let the foreground collector do its job. 4341 // Clear _restart_addr, so that foreground GC 4342 // works from scratch. This avoids the headache of 4343 // a "rescan" which would otherwise be needed because 4344 // of the dirty mod union table & card table. 4345 _restart_addr = NULL; 4346 return false; // indicating failure to complete marking 4347 } 4348 // Deal with stack overflow: 4349 // we restart marking from _restart_addr 4350 HeapWord* ra = _restart_addr; 4351 markFromRootsClosure.reset(ra); 4352 _restart_addr = NULL; 4353 _markBitMap.iterate(&markFromRootsClosure, ra, _span.end()); 4354 } 4355 return true; 4356 } 4357 4358 void CMSCollector::preclean() { 4359 check_correct_thread_executing(); 4360 assert(Thread::current()->is_ConcurrentGC_thread(), "Wrong thread"); 4361 verify_work_stacks_empty(); 4362 verify_overflow_empty(); 4363 _abort_preclean = false; 4364 if (CMSPrecleaningEnabled) { 4365 _eden_chunk_index = 0; 4366 size_t used = get_eden_used(); 4367 size_t capacity = get_eden_capacity(); 4368 // Don't start sampling unless we will get sufficiently 4369 // many samples. 4370 if (used < (capacity/(CMSScheduleRemarkSamplingRatio * 100) 4371 * CMSScheduleRemarkEdenPenetration)) { 4372 _start_sampling = true; 4373 } else { 4374 _start_sampling = false; 4375 } 4376 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 4377 CMSPhaseAccounting pa(this, "preclean", !PrintGCDetails); 4378 preclean_work(CMSPrecleanRefLists1, CMSPrecleanSurvivors1); 4379 } 4380 CMSTokenSync x(true); // is cms thread 4381 if (CMSPrecleaningEnabled) { 4382 sample_eden(); 4383 _collectorState = AbortablePreclean; 4384 } else { 4385 _collectorState = FinalMarking; 4386 } 4387 verify_work_stacks_empty(); 4388 verify_overflow_empty(); 4389 } 4390 4391 // Try and schedule the remark such that young gen 4392 // occupancy is CMSScheduleRemarkEdenPenetration %. 4393 void CMSCollector::abortable_preclean() { 4394 check_correct_thread_executing(); 4395 assert(CMSPrecleaningEnabled, "Inconsistent control state"); 4396 assert(_collectorState == AbortablePreclean, "Inconsistent control state"); 4397 4398 // If Eden's current occupancy is below this threshold, 4399 // immediately schedule the remark; else preclean 4400 // past the next scavenge in an effort to 4401 // schedule the pause as described avove. By choosing 4402 // CMSScheduleRemarkEdenSizeThreshold >= max eden size 4403 // we will never do an actual abortable preclean cycle. 4404 if (get_eden_used() > CMSScheduleRemarkEdenSizeThreshold) { 4405 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 4406 CMSPhaseAccounting pa(this, "abortable-preclean", !PrintGCDetails); 4407 // We need more smarts in the abortable preclean 4408 // loop below to deal with cases where allocation 4409 // in young gen is very very slow, and our precleaning 4410 // is running a losing race against a horde of 4411 // mutators intent on flooding us with CMS updates 4412 // (dirty cards). 4413 // One, admittedly dumb, strategy is to give up 4414 // after a certain number of abortable precleaning loops 4415 // or after a certain maximum time. We want to make 4416 // this smarter in the next iteration. 4417 // XXX FIX ME!!! YSR 4418 size_t loops = 0, workdone = 0, cumworkdone = 0, waited = 0; 4419 while (!(should_abort_preclean() || 4420 ConcurrentMarkSweepThread::should_terminate())) { 4421 workdone = preclean_work(CMSPrecleanRefLists2, CMSPrecleanSurvivors2); 4422 cumworkdone += workdone; 4423 loops++; 4424 // Voluntarily terminate abortable preclean phase if we have 4425 // been at it for too long. 4426 if ((CMSMaxAbortablePrecleanLoops != 0) && 4427 loops >= CMSMaxAbortablePrecleanLoops) { 4428 if (PrintGCDetails) { 4429 gclog_or_tty->print(" CMS: abort preclean due to loops "); 4430 } 4431 break; 4432 } 4433 if (pa.wallclock_millis() > CMSMaxAbortablePrecleanTime) { 4434 if (PrintGCDetails) { 4435 gclog_or_tty->print(" CMS: abort preclean due to time "); 4436 } 4437 break; 4438 } 4439 // If we are doing little work each iteration, we should 4440 // take a short break. 4441 if (workdone < CMSAbortablePrecleanMinWorkPerIteration) { 4442 // Sleep for some time, waiting for work to accumulate 4443 stopTimer(); 4444 cmsThread()->wait_on_cms_lock(CMSAbortablePrecleanWaitMillis); 4445 startTimer(); 4446 waited++; 4447 } 4448 } 4449 if (PrintCMSStatistics > 0) { 4450 gclog_or_tty->print(" [%d iterations, %d waits, %d cards)] ", 4451 loops, waited, cumworkdone); 4452 } 4453 } 4454 CMSTokenSync x(true); // is cms thread 4455 if (_collectorState != Idling) { 4456 assert(_collectorState == AbortablePreclean, 4457 "Spontaneous state transition?"); 4458 _collectorState = FinalMarking; 4459 } // Else, a foreground collection completed this CMS cycle. 4460 return; 4461 } 4462 4463 // Respond to an Eden sampling opportunity 4464 void CMSCollector::sample_eden() { 4465 // Make sure a young gc cannot sneak in between our 4466 // reading and recording of a sample. 4467 assert(Thread::current()->is_ConcurrentGC_thread(), 4468 "Only the cms thread may collect Eden samples"); 4469 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 4470 "Should collect samples while holding CMS token"); 4471 if (!_start_sampling) { 4472 return; 4473 } 4474 if (_eden_chunk_array) { 4475 if (_eden_chunk_index < _eden_chunk_capacity) { 4476 _eden_chunk_array[_eden_chunk_index] = *_top_addr; // take sample 4477 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr, 4478 "Unexpected state of Eden"); 4479 // We'd like to check that what we just sampled is an oop-start address; 4480 // however, we cannot do that here since the object may not yet have been 4481 // initialized. So we'll instead do the check when we _use_ this sample 4482 // later. 4483 if (_eden_chunk_index == 0 || 4484 (pointer_delta(_eden_chunk_array[_eden_chunk_index], 4485 _eden_chunk_array[_eden_chunk_index-1]) 4486 >= CMSSamplingGrain)) { 4487 _eden_chunk_index++; // commit sample 4488 } 4489 } 4490 } 4491 if ((_collectorState == AbortablePreclean) && !_abort_preclean) { 4492 size_t used = get_eden_used(); 4493 size_t capacity = get_eden_capacity(); 4494 assert(used <= capacity, "Unexpected state of Eden"); 4495 if (used > (capacity/100 * CMSScheduleRemarkEdenPenetration)) { 4496 _abort_preclean = true; 4497 } 4498 } 4499 } 4500 4501 4502 size_t CMSCollector::preclean_work(bool clean_refs, bool clean_survivor) { 4503 assert(_collectorState == Precleaning || 4504 _collectorState == AbortablePreclean, "incorrect state"); 4505 ResourceMark rm; 4506 HandleMark hm; 4507 4508 // Precleaning is currently not MT but the reference processor 4509 // may be set for MT. Disable it temporarily here. 4510 ReferenceProcessor* rp = ref_processor(); 4511 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(rp, false); 4512 4513 // Do one pass of scrubbing the discovered reference lists 4514 // to remove any reference objects with strongly-reachable 4515 // referents. 4516 if (clean_refs) { 4517 CMSPrecleanRefsYieldClosure yield_cl(this); 4518 assert(rp->span().equals(_span), "Spans should be equal"); 4519 CMSKeepAliveClosure keep_alive(this, _span, &_markBitMap, 4520 &_markStack, true /* preclean */); 4521 CMSDrainMarkingStackClosure complete_trace(this, 4522 _span, &_markBitMap, &_markStack, 4523 &keep_alive, true /* preclean */); 4524 4525 // We don't want this step to interfere with a young 4526 // collection because we don't want to take CPU 4527 // or memory bandwidth away from the young GC threads 4528 // (which may be as many as there are CPUs). 4529 // Note that we don't need to protect ourselves from 4530 // interference with mutators because they can't 4531 // manipulate the discovered reference lists nor affect 4532 // the computed reachability of the referents, the 4533 // only properties manipulated by the precleaning 4534 // of these reference lists. 4535 stopTimer(); 4536 CMSTokenSyncWithLocks x(true /* is cms thread */, 4537 bitMapLock()); 4538 startTimer(); 4539 sample_eden(); 4540 4541 // The following will yield to allow foreground 4542 // collection to proceed promptly. XXX YSR: 4543 // The code in this method may need further 4544 // tweaking for better performance and some restructuring 4545 // for cleaner interfaces. 4546 rp->preclean_discovered_references( 4547 rp->is_alive_non_header(), &keep_alive, &complete_trace, &yield_cl); 4548 } 4549 4550 if (clean_survivor) { // preclean the active survivor space(s) 4551 assert(_young_gen->kind() == Generation::DefNew || 4552 _young_gen->kind() == Generation::ParNew || 4553 _young_gen->kind() == Generation::ASParNew, 4554 "incorrect type for cast"); 4555 DefNewGeneration* dng = (DefNewGeneration*)_young_gen; 4556 PushAndMarkClosure pam_cl(this, _span, ref_processor(), 4557 &_markBitMap, &_modUnionTable, 4558 &_markStack, true /* precleaning phase */); 4559 stopTimer(); 4560 CMSTokenSyncWithLocks ts(true /* is cms thread */, 4561 bitMapLock()); 4562 startTimer(); 4563 unsigned int before_count = 4564 GenCollectedHeap::heap()->total_collections(); 4565 SurvivorSpacePrecleanClosure 4566 sss_cl(this, _span, &_markBitMap, &_markStack, 4567 &pam_cl, before_count, CMSYield); 4568 dng->from()->object_iterate_careful(&sss_cl); 4569 dng->to()->object_iterate_careful(&sss_cl); 4570 } 4571 MarkRefsIntoAndScanClosure 4572 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable, 4573 &_markStack, this, CMSYield, 4574 true /* precleaning phase */); 4575 // CAUTION: The following closure has persistent state that may need to 4576 // be reset upon a decrease in the sequence of addresses it 4577 // processes. 4578 ScanMarkedObjectsAgainCarefullyClosure 4579 smoac_cl(this, _span, 4580 &_markBitMap, &_markStack, &mrias_cl, CMSYield); 4581 4582 // Preclean dirty cards in ModUnionTable and CardTable using 4583 // appropriate convergence criterion; 4584 // repeat CMSPrecleanIter times unless we find that 4585 // we are losing. 4586 assert(CMSPrecleanIter < 10, "CMSPrecleanIter is too large"); 4587 assert(CMSPrecleanNumerator < CMSPrecleanDenominator, 4588 "Bad convergence multiplier"); 4589 assert(CMSPrecleanThreshold >= 100, 4590 "Unreasonably low CMSPrecleanThreshold"); 4591 4592 size_t numIter, cumNumCards, lastNumCards, curNumCards; 4593 for (numIter = 0, cumNumCards = lastNumCards = curNumCards = 0; 4594 numIter < CMSPrecleanIter; 4595 numIter++, lastNumCards = curNumCards, cumNumCards += curNumCards) { 4596 curNumCards = preclean_mod_union_table(_cmsGen, &smoac_cl); 4597 if (Verbose && PrintGCDetails) { 4598 gclog_or_tty->print(" (modUnionTable: %d cards)", curNumCards); 4599 } 4600 // Either there are very few dirty cards, so re-mark 4601 // pause will be small anyway, or our pre-cleaning isn't 4602 // that much faster than the rate at which cards are being 4603 // dirtied, so we might as well stop and re-mark since 4604 // precleaning won't improve our re-mark time by much. 4605 if (curNumCards <= CMSPrecleanThreshold || 4606 (numIter > 0 && 4607 (curNumCards * CMSPrecleanDenominator > 4608 lastNumCards * CMSPrecleanNumerator))) { 4609 numIter++; 4610 cumNumCards += curNumCards; 4611 break; 4612 } 4613 } 4614 4615 preclean_klasses(&mrias_cl, _cmsGen->freelistLock()); 4616 4617 curNumCards = preclean_card_table(_cmsGen, &smoac_cl); 4618 cumNumCards += curNumCards; 4619 if (PrintGCDetails && PrintCMSStatistics != 0) { 4620 gclog_or_tty->print_cr(" (cardTable: %d cards, re-scanned %d cards, %d iterations)", 4621 curNumCards, cumNumCards, numIter); 4622 } 4623 return cumNumCards; // as a measure of useful work done 4624 } 4625 4626 // PRECLEANING NOTES: 4627 // Precleaning involves: 4628 // . reading the bits of the modUnionTable and clearing the set bits. 4629 // . For the cards corresponding to the set bits, we scan the 4630 // objects on those cards. This means we need the free_list_lock 4631 // so that we can safely iterate over the CMS space when scanning 4632 // for oops. 4633 // . When we scan the objects, we'll be both reading and setting 4634 // marks in the marking bit map, so we'll need the marking bit map. 4635 // . For protecting _collector_state transitions, we take the CGC_lock. 4636 // Note that any races in the reading of of card table entries by the 4637 // CMS thread on the one hand and the clearing of those entries by the 4638 // VM thread or the setting of those entries by the mutator threads on the 4639 // other are quite benign. However, for efficiency it makes sense to keep 4640 // the VM thread from racing with the CMS thread while the latter is 4641 // dirty card info to the modUnionTable. We therefore also use the 4642 // CGC_lock to protect the reading of the card table and the mod union 4643 // table by the CM thread. 4644 // . We run concurrently with mutator updates, so scanning 4645 // needs to be done carefully -- we should not try to scan 4646 // potentially uninitialized objects. 4647 // 4648 // Locking strategy: While holding the CGC_lock, we scan over and 4649 // reset a maximal dirty range of the mod union / card tables, then lock 4650 // the free_list_lock and bitmap lock to do a full marking, then 4651 // release these locks; and repeat the cycle. This allows for a 4652 // certain amount of fairness in the sharing of these locks between 4653 // the CMS collector on the one hand, and the VM thread and the 4654 // mutators on the other. 4655 4656 // NOTE: preclean_mod_union_table() and preclean_card_table() 4657 // further below are largely identical; if you need to modify 4658 // one of these methods, please check the other method too. 4659 4660 size_t CMSCollector::preclean_mod_union_table( 4661 ConcurrentMarkSweepGeneration* gen, 4662 ScanMarkedObjectsAgainCarefullyClosure* cl) { 4663 verify_work_stacks_empty(); 4664 verify_overflow_empty(); 4665 4666 // strategy: starting with the first card, accumulate contiguous 4667 // ranges of dirty cards; clear these cards, then scan the region 4668 // covered by these cards. 4669 4670 // Since all of the MUT is committed ahead, we can just use 4671 // that, in case the generations expand while we are precleaning. 4672 // It might also be fine to just use the committed part of the 4673 // generation, but we might potentially miss cards when the 4674 // generation is rapidly expanding while we are in the midst 4675 // of precleaning. 4676 HeapWord* startAddr = gen->reserved().start(); 4677 HeapWord* endAddr = gen->reserved().end(); 4678 4679 cl->setFreelistLock(gen->freelistLock()); // needed for yielding 4680 4681 size_t numDirtyCards, cumNumDirtyCards; 4682 HeapWord *nextAddr, *lastAddr; 4683 for (cumNumDirtyCards = numDirtyCards = 0, 4684 nextAddr = lastAddr = startAddr; 4685 nextAddr < endAddr; 4686 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 4687 4688 ResourceMark rm; 4689 HandleMark hm; 4690 4691 MemRegion dirtyRegion; 4692 { 4693 stopTimer(); 4694 // Potential yield point 4695 CMSTokenSync ts(true); 4696 startTimer(); 4697 sample_eden(); 4698 // Get dirty region starting at nextOffset (inclusive), 4699 // simultaneously clearing it. 4700 dirtyRegion = 4701 _modUnionTable.getAndClearMarkedRegion(nextAddr, endAddr); 4702 assert(dirtyRegion.start() >= nextAddr, 4703 "returned region inconsistent?"); 4704 } 4705 // Remember where the next search should begin. 4706 // The returned region (if non-empty) is a right open interval, 4707 // so lastOffset is obtained from the right end of that 4708 // interval. 4709 lastAddr = dirtyRegion.end(); 4710 // Should do something more transparent and less hacky XXX 4711 numDirtyCards = 4712 _modUnionTable.heapWordDiffToOffsetDiff(dirtyRegion.word_size()); 4713 4714 // We'll scan the cards in the dirty region (with periodic 4715 // yields for foreground GC as needed). 4716 if (!dirtyRegion.is_empty()) { 4717 assert(numDirtyCards > 0, "consistency check"); 4718 HeapWord* stop_point = NULL; 4719 stopTimer(); 4720 // Potential yield point 4721 CMSTokenSyncWithLocks ts(true, gen->freelistLock(), 4722 bitMapLock()); 4723 startTimer(); 4724 { 4725 verify_work_stacks_empty(); 4726 verify_overflow_empty(); 4727 sample_eden(); 4728 stop_point = 4729 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 4730 } 4731 if (stop_point != NULL) { 4732 // The careful iteration stopped early either because it found an 4733 // uninitialized object, or because we were in the midst of an 4734 // "abortable preclean", which should now be aborted. Redirty 4735 // the bits corresponding to the partially-scanned or unscanned 4736 // cards. We'll either restart at the next block boundary or 4737 // abort the preclean. 4738 assert((_collectorState == AbortablePreclean && should_abort_preclean()), 4739 "Should only be AbortablePreclean."); 4740 _modUnionTable.mark_range(MemRegion(stop_point, dirtyRegion.end())); 4741 if (should_abort_preclean()) { 4742 break; // out of preclean loop 4743 } else { 4744 // Compute the next address at which preclean should pick up; 4745 // might need bitMapLock in order to read P-bits. 4746 lastAddr = next_card_start_after_block(stop_point); 4747 } 4748 } 4749 } else { 4750 assert(lastAddr == endAddr, "consistency check"); 4751 assert(numDirtyCards == 0, "consistency check"); 4752 break; 4753 } 4754 } 4755 verify_work_stacks_empty(); 4756 verify_overflow_empty(); 4757 return cumNumDirtyCards; 4758 } 4759 4760 // NOTE: preclean_mod_union_table() above and preclean_card_table() 4761 // below are largely identical; if you need to modify 4762 // one of these methods, please check the other method too. 4763 4764 size_t CMSCollector::preclean_card_table(ConcurrentMarkSweepGeneration* gen, 4765 ScanMarkedObjectsAgainCarefullyClosure* cl) { 4766 // strategy: it's similar to precleamModUnionTable above, in that 4767 // we accumulate contiguous ranges of dirty cards, mark these cards 4768 // precleaned, then scan the region covered by these cards. 4769 HeapWord* endAddr = (HeapWord*)(gen->_virtual_space.high()); 4770 HeapWord* startAddr = (HeapWord*)(gen->_virtual_space.low()); 4771 4772 cl->setFreelistLock(gen->freelistLock()); // needed for yielding 4773 4774 size_t numDirtyCards, cumNumDirtyCards; 4775 HeapWord *lastAddr, *nextAddr; 4776 4777 for (cumNumDirtyCards = numDirtyCards = 0, 4778 nextAddr = lastAddr = startAddr; 4779 nextAddr < endAddr; 4780 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 4781 4782 ResourceMark rm; 4783 HandleMark hm; 4784 4785 MemRegion dirtyRegion; 4786 { 4787 // See comments in "Precleaning notes" above on why we 4788 // do this locking. XXX Could the locking overheads be 4789 // too high when dirty cards are sparse? [I don't think so.] 4790 stopTimer(); 4791 CMSTokenSync x(true); // is cms thread 4792 startTimer(); 4793 sample_eden(); 4794 // Get and clear dirty region from card table 4795 dirtyRegion = _ct->ct_bs()->dirty_card_range_after_reset( 4796 MemRegion(nextAddr, endAddr), 4797 true, 4798 CardTableModRefBS::precleaned_card_val()); 4799 4800 assert(dirtyRegion.start() >= nextAddr, 4801 "returned region inconsistent?"); 4802 } 4803 lastAddr = dirtyRegion.end(); 4804 numDirtyCards = 4805 dirtyRegion.word_size()/CardTableModRefBS::card_size_in_words; 4806 4807 if (!dirtyRegion.is_empty()) { 4808 stopTimer(); 4809 CMSTokenSyncWithLocks ts(true, gen->freelistLock(), bitMapLock()); 4810 startTimer(); 4811 sample_eden(); 4812 verify_work_stacks_empty(); 4813 verify_overflow_empty(); 4814 HeapWord* stop_point = 4815 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 4816 if (stop_point != NULL) { 4817 assert((_collectorState == AbortablePreclean && should_abort_preclean()), 4818 "Should only be AbortablePreclean."); 4819 _ct->ct_bs()->invalidate(MemRegion(stop_point, dirtyRegion.end())); 4820 if (should_abort_preclean()) { 4821 break; // out of preclean loop 4822 } else { 4823 // Compute the next address at which preclean should pick up. 4824 lastAddr = next_card_start_after_block(stop_point); 4825 } 4826 } 4827 } else { 4828 break; 4829 } 4830 } 4831 verify_work_stacks_empty(); 4832 verify_overflow_empty(); 4833 return cumNumDirtyCards; 4834 } 4835 4836 class PrecleanKlassClosure : public KlassClosure { 4837 CMKlassClosure _cm_klass_closure; 4838 public: 4839 PrecleanKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {} 4840 void do_klass(Klass* k) { 4841 if (k->has_accumulated_modified_oops()) { 4842 k->clear_accumulated_modified_oops(); 4843 4844 _cm_klass_closure.do_klass(k); 4845 } 4846 } 4847 }; 4848 4849 // The freelist lock is needed to prevent asserts, is it really needed? 4850 void CMSCollector::preclean_klasses(MarkRefsIntoAndScanClosure* cl, Mutex* freelistLock) { 4851 4852 cl->set_freelistLock(freelistLock); 4853 4854 CMSTokenSyncWithLocks ts(true, freelistLock, bitMapLock()); 4855 4856 // SSS: Add equivalent to ScanMarkedObjectsAgainCarefullyClosure::do_yield_check and should_abort_preclean? 4857 // SSS: We should probably check if precleaning should be aborted, at suitable intervals? 4858 PrecleanKlassClosure preclean_klass_closure(cl); 4859 ClassLoaderDataGraph::classes_do(&preclean_klass_closure); 4860 4861 verify_work_stacks_empty(); 4862 verify_overflow_empty(); 4863 } 4864 4865 void CMSCollector::checkpointRootsFinal(bool asynch, 4866 bool clear_all_soft_refs, bool init_mark_was_synchronous) { 4867 assert(_collectorState == FinalMarking, "incorrect state transition?"); 4868 check_correct_thread_executing(); 4869 // world is stopped at this checkpoint 4870 assert(SafepointSynchronize::is_at_safepoint(), 4871 "world should be stopped"); 4872 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 4873 4874 verify_work_stacks_empty(); 4875 verify_overflow_empty(); 4876 4877 SpecializationStats::clear(); 4878 if (PrintGCDetails) { 4879 gclog_or_tty->print("[YG occupancy: "SIZE_FORMAT" K ("SIZE_FORMAT" K)]", 4880 _young_gen->used() / K, 4881 _young_gen->capacity() / K); 4882 } 4883 if (asynch) { 4884 if (CMSScavengeBeforeRemark) { 4885 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4886 // Temporarily set flag to false, GCH->do_collection will 4887 // expect it to be false and set to true 4888 FlagSetting fl(gch->_is_gc_active, false); 4889 NOT_PRODUCT(TraceTime t("Scavenge-Before-Remark", 4890 PrintGCDetails && Verbose, true, gclog_or_tty);) 4891 int level = _cmsGen->level() - 1; 4892 if (level >= 0) { 4893 gch->do_collection(true, // full (i.e. force, see below) 4894 false, // !clear_all_soft_refs 4895 0, // size 4896 false, // is_tlab 4897 level // max_level 4898 ); 4899 } 4900 } 4901 FreelistLocker x(this); 4902 MutexLockerEx y(bitMapLock(), 4903 Mutex::_no_safepoint_check_flag); 4904 assert(!init_mark_was_synchronous, "but that's impossible!"); 4905 checkpointRootsFinalWork(asynch, clear_all_soft_refs, false); 4906 } else { 4907 // already have all the locks 4908 checkpointRootsFinalWork(asynch, clear_all_soft_refs, 4909 init_mark_was_synchronous); 4910 } 4911 verify_work_stacks_empty(); 4912 verify_overflow_empty(); 4913 SpecializationStats::print(); 4914 } 4915 4916 void CMSCollector::checkpointRootsFinalWork(bool asynch, 4917 bool clear_all_soft_refs, bool init_mark_was_synchronous) { 4918 4919 NOT_PRODUCT(TraceTime tr("checkpointRootsFinalWork", PrintGCDetails, false, gclog_or_tty);) 4920 4921 assert(haveFreelistLocks(), "must have free list locks"); 4922 assert_lock_strong(bitMapLock()); 4923 4924 if (UseAdaptiveSizePolicy) { 4925 size_policy()->checkpoint_roots_final_begin(); 4926 } 4927 4928 ResourceMark rm; 4929 HandleMark hm; 4930 4931 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4932 4933 if (should_unload_classes()) { 4934 CodeCache::gc_prologue(); 4935 } 4936 assert(haveFreelistLocks(), "must have free list locks"); 4937 assert_lock_strong(bitMapLock()); 4938 4939 if (!init_mark_was_synchronous) { 4940 // We might assume that we need not fill TLAB's when 4941 // CMSScavengeBeforeRemark is set, because we may have just done 4942 // a scavenge which would have filled all TLAB's -- and besides 4943 // Eden would be empty. This however may not always be the case -- 4944 // for instance although we asked for a scavenge, it may not have 4945 // happened because of a JNI critical section. We probably need 4946 // a policy for deciding whether we can in that case wait until 4947 // the critical section releases and then do the remark following 4948 // the scavenge, and skip it here. In the absence of that policy, 4949 // or of an indication of whether the scavenge did indeed occur, 4950 // we cannot rely on TLAB's having been filled and must do 4951 // so here just in case a scavenge did not happen. 4952 gch->ensure_parsability(false); // fill TLAB's, but no need to retire them 4953 // Update the saved marks which may affect the root scans. 4954 gch->save_marks(); 4955 4956 { 4957 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;) 4958 4959 // Note on the role of the mod union table: 4960 // Since the marker in "markFromRoots" marks concurrently with 4961 // mutators, it is possible for some reachable objects not to have been 4962 // scanned. For instance, an only reference to an object A was 4963 // placed in object B after the marker scanned B. Unless B is rescanned, 4964 // A would be collected. Such updates to references in marked objects 4965 // are detected via the mod union table which is the set of all cards 4966 // dirtied since the first checkpoint in this GC cycle and prior to 4967 // the most recent young generation GC, minus those cleaned up by the 4968 // concurrent precleaning. 4969 if (CMSParallelRemarkEnabled && CollectedHeap::use_parallel_gc_threads()) { 4970 TraceTime t("Rescan (parallel) ", PrintGCDetails, false, gclog_or_tty); 4971 do_remark_parallel(); 4972 } else { 4973 TraceTime t("Rescan (non-parallel) ", PrintGCDetails, false, 4974 gclog_or_tty); 4975 do_remark_non_parallel(); 4976 } 4977 } 4978 } else { 4979 assert(!asynch, "Can't have init_mark_was_synchronous in asynch mode"); 4980 // The initial mark was stop-world, so there's no rescanning to 4981 // do; go straight on to the next step below. 4982 } 4983 verify_work_stacks_empty(); 4984 verify_overflow_empty(); 4985 4986 { 4987 NOT_PRODUCT(TraceTime ts("refProcessingWork", PrintGCDetails, false, gclog_or_tty);) 4988 refProcessingWork(asynch, clear_all_soft_refs); 4989 } 4990 verify_work_stacks_empty(); 4991 verify_overflow_empty(); 4992 4993 if (should_unload_classes()) { 4994 CodeCache::gc_epilogue(); 4995 } 4996 JvmtiExport::gc_epilogue(); 4997 4998 // If we encountered any (marking stack / work queue) overflow 4999 // events during the current CMS cycle, take appropriate 5000 // remedial measures, where possible, so as to try and avoid 5001 // recurrence of that condition. 5002 assert(_markStack.isEmpty(), "No grey objects"); 5003 size_t ser_ovflw = _ser_pmc_remark_ovflw + _ser_pmc_preclean_ovflw + 5004 _ser_kac_ovflw + _ser_kac_preclean_ovflw; 5005 if (ser_ovflw > 0) { 5006 if (PrintCMSStatistics != 0) { 5007 gclog_or_tty->print_cr("Marking stack overflow (benign) " 5008 "(pmc_pc="SIZE_FORMAT", pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT 5009 ", kac_preclean="SIZE_FORMAT")", 5010 _ser_pmc_preclean_ovflw, _ser_pmc_remark_ovflw, 5011 _ser_kac_ovflw, _ser_kac_preclean_ovflw); 5012 } 5013 _markStack.expand(); 5014 _ser_pmc_remark_ovflw = 0; 5015 _ser_pmc_preclean_ovflw = 0; 5016 _ser_kac_preclean_ovflw = 0; 5017 _ser_kac_ovflw = 0; 5018 } 5019 if (_par_pmc_remark_ovflw > 0 || _par_kac_ovflw > 0) { 5020 if (PrintCMSStatistics != 0) { 5021 gclog_or_tty->print_cr("Work queue overflow (benign) " 5022 "(pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT")", 5023 _par_pmc_remark_ovflw, _par_kac_ovflw); 5024 } 5025 _par_pmc_remark_ovflw = 0; 5026 _par_kac_ovflw = 0; 5027 } 5028 if (PrintCMSStatistics != 0) { 5029 if (_markStack._hit_limit > 0) { 5030 gclog_or_tty->print_cr(" (benign) Hit max stack size limit ("SIZE_FORMAT")", 5031 _markStack._hit_limit); 5032 } 5033 if (_markStack._failed_double > 0) { 5034 gclog_or_tty->print_cr(" (benign) Failed stack doubling ("SIZE_FORMAT")," 5035 " current capacity "SIZE_FORMAT, 5036 _markStack._failed_double, 5037 _markStack.capacity()); 5038 } 5039 } 5040 _markStack._hit_limit = 0; 5041 _markStack._failed_double = 0; 5042 5043 if ((VerifyAfterGC || VerifyDuringGC) && 5044 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 5045 verify_after_remark(); 5046 } 5047 5048 // Change under the freelistLocks. 5049 _collectorState = Sweeping; 5050 // Call isAllClear() under bitMapLock 5051 assert(_modUnionTable.isAllClear(), 5052 "Should be clear by end of the final marking"); 5053 assert(_ct->klass_rem_set()->mod_union_is_clear(), 5054 "Should be clear by end of the final marking"); 5055 if (UseAdaptiveSizePolicy) { 5056 size_policy()->checkpoint_roots_final_end(gch->gc_cause()); 5057 } 5058 } 5059 5060 // Parallel remark task 5061 class CMSParRemarkTask: public AbstractGangTask { 5062 CMSCollector* _collector; 5063 int _n_workers; 5064 CompactibleFreeListSpace* _cms_space; 5065 5066 // The per-thread work queues, available here for stealing. 5067 OopTaskQueueSet* _task_queues; 5068 ParallelTaskTerminator _term; 5069 5070 public: 5071 // A value of 0 passed to n_workers will cause the number of 5072 // workers to be taken from the active workers in the work gang. 5073 CMSParRemarkTask(CMSCollector* collector, 5074 CompactibleFreeListSpace* cms_space, 5075 int n_workers, FlexibleWorkGang* workers, 5076 OopTaskQueueSet* task_queues): 5077 AbstractGangTask("Rescan roots and grey objects in parallel"), 5078 _collector(collector), 5079 _cms_space(cms_space), 5080 _n_workers(n_workers), 5081 _task_queues(task_queues), 5082 _term(n_workers, task_queues) { } 5083 5084 OopTaskQueueSet* task_queues() { return _task_queues; } 5085 5086 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 5087 5088 ParallelTaskTerminator* terminator() { return &_term; } 5089 int n_workers() { return _n_workers; } 5090 5091 void work(uint worker_id); 5092 5093 private: 5094 // Work method in support of parallel rescan ... of young gen spaces 5095 void do_young_space_rescan(int i, Par_MarkRefsIntoAndScanClosure* cl, 5096 ContiguousSpace* space, 5097 HeapWord** chunk_array, size_t chunk_top); 5098 5099 // ... of dirty cards in old space 5100 void do_dirty_card_rescan_tasks(CompactibleFreeListSpace* sp, int i, 5101 Par_MarkRefsIntoAndScanClosure* cl); 5102 5103 // ... work stealing for the above 5104 void do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, int* seed); 5105 }; 5106 5107 class RemarkKlassClosure : public KlassClosure { 5108 CMKlassClosure _cm_klass_closure; 5109 public: 5110 RemarkKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {} 5111 void do_klass(Klass* k) { 5112 // Check if we have modified any oops in the Klass during the concurrent marking. 5113 if (k->has_accumulated_modified_oops()) { 5114 k->clear_accumulated_modified_oops(); 5115 5116 // We could have transfered the current modified marks to the accumulated marks, 5117 // like we do with the Card Table to Mod Union Table. But it's not really necessary. 5118 } else if (k->has_modified_oops()) { 5119 // Don't clear anything, this info is needed by the next young collection. 5120 } else { 5121 // No modified oops in the Klass. 5122 return; 5123 } 5124 5125 // The klass has modified fields, need to scan the klass. 5126 _cm_klass_closure.do_klass(k); 5127 } 5128 }; 5129 5130 // work_queue(i) is passed to the closure 5131 // Par_MarkRefsIntoAndScanClosure. The "i" parameter 5132 // also is passed to do_dirty_card_rescan_tasks() and to 5133 // do_work_steal() to select the i-th task_queue. 5134 5135 void CMSParRemarkTask::work(uint worker_id) { 5136 elapsedTimer _timer; 5137 ResourceMark rm; 5138 HandleMark hm; 5139 5140 // ---------- rescan from roots -------------- 5141 _timer.start(); 5142 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5143 Par_MarkRefsIntoAndScanClosure par_mrias_cl(_collector, 5144 _collector->_span, _collector->ref_processor(), 5145 &(_collector->_markBitMap), 5146 work_queue(worker_id)); 5147 5148 // Rescan young gen roots first since these are likely 5149 // coarsely partitioned and may, on that account, constitute 5150 // the critical path; thus, it's best to start off that 5151 // work first. 5152 // ---------- young gen roots -------------- 5153 { 5154 DefNewGeneration* dng = _collector->_young_gen->as_DefNewGeneration(); 5155 EdenSpace* eden_space = dng->eden(); 5156 ContiguousSpace* from_space = dng->from(); 5157 ContiguousSpace* to_space = dng->to(); 5158 5159 HeapWord** eca = _collector->_eden_chunk_array; 5160 size_t ect = _collector->_eden_chunk_index; 5161 HeapWord** sca = _collector->_survivor_chunk_array; 5162 size_t sct = _collector->_survivor_chunk_index; 5163 5164 assert(ect <= _collector->_eden_chunk_capacity, "out of bounds"); 5165 assert(sct <= _collector->_survivor_chunk_capacity, "out of bounds"); 5166 5167 do_young_space_rescan(worker_id, &par_mrias_cl, to_space, NULL, 0); 5168 do_young_space_rescan(worker_id, &par_mrias_cl, from_space, sca, sct); 5169 do_young_space_rescan(worker_id, &par_mrias_cl, eden_space, eca, ect); 5170 5171 _timer.stop(); 5172 if (PrintCMSStatistics != 0) { 5173 gclog_or_tty->print_cr( 5174 "Finished young gen rescan work in %dth thread: %3.3f sec", 5175 worker_id, _timer.seconds()); 5176 } 5177 } 5178 5179 // ---------- remaining roots -------------- 5180 _timer.reset(); 5181 _timer.start(); 5182 gch->gen_process_strong_roots(_collector->_cmsGen->level(), 5183 false, // yg was scanned above 5184 false, // this is parallel code 5185 false, // not scavenging 5186 SharedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()), 5187 &par_mrias_cl, 5188 true, // walk all of code cache if (so & SO_CodeCache) 5189 NULL, 5190 NULL); // The dirty klasses will be handled below 5191 assert(_collector->should_unload_classes() 5192 || (_collector->CMSCollector::roots_scanning_options() & SharedHeap::SO_CodeCache), 5193 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 5194 _timer.stop(); 5195 if (PrintCMSStatistics != 0) { 5196 gclog_or_tty->print_cr( 5197 "Finished remaining root rescan work in %dth thread: %3.3f sec", 5198 worker_id, _timer.seconds()); 5199 } 5200 5201 // ---------- unhandled CLD scanning ---------- 5202 if (worker_id == 0) { // Single threaded at the moment. 5203 _timer.reset(); 5204 _timer.start(); 5205 5206 // Scan all new class loader data objects and new dependencies that were 5207 // introduced during concurrent marking. 5208 ResourceMark rm; 5209 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds(); 5210 for (int i = 0; i < array->length(); i++) { 5211 par_mrias_cl.do_class_loader_data(array->at(i)); 5212 } 5213 5214 // We don't need to keep track of new CLDs anymore. 5215 ClassLoaderDataGraph::remember_new_clds(false); 5216 5217 _timer.stop(); 5218 if (PrintCMSStatistics != 0) { 5219 gclog_or_tty->print_cr( 5220 "Finished unhandled CLD scanning work in %dth thread: %3.3f sec", 5221 worker_id, _timer.seconds()); 5222 } 5223 } 5224 5225 // ---------- dirty klass scanning ---------- 5226 if (worker_id == 0) { // Single threaded at the moment. 5227 _timer.reset(); 5228 _timer.start(); 5229 5230 // Scan all classes that was dirtied during the concurrent marking phase. 5231 RemarkKlassClosure remark_klass_closure(&par_mrias_cl); 5232 ClassLoaderDataGraph::classes_do(&remark_klass_closure); 5233 5234 _timer.stop(); 5235 if (PrintCMSStatistics != 0) { 5236 gclog_or_tty->print_cr( 5237 "Finished dirty klass scanning work in %dth thread: %3.3f sec", 5238 worker_id, _timer.seconds()); 5239 } 5240 } 5241 5242 // We might have added oops to ClassLoaderData::_handles during the 5243 // concurrent marking phase. These oops point to newly allocated objects 5244 // that are guaranteed to be kept alive. Either by the direct allocation 5245 // code, or when the young collector processes the strong roots. Hence, 5246 // we don't have to revisit the _handles block during the remark phase. 5247 5248 // ---------- rescan dirty cards ------------ 5249 _timer.reset(); 5250 _timer.start(); 5251 5252 // Do the rescan tasks for each of the two spaces 5253 // (cms_space) in turn. 5254 // "worker_id" is passed to select the task_queue for "worker_id" 5255 do_dirty_card_rescan_tasks(_cms_space, worker_id, &par_mrias_cl); 5256 _timer.stop(); 5257 if (PrintCMSStatistics != 0) { 5258 gclog_or_tty->print_cr( 5259 "Finished dirty card rescan work in %dth thread: %3.3f sec", 5260 worker_id, _timer.seconds()); 5261 } 5262 5263 // ---------- steal work from other threads ... 5264 // ---------- ... and drain overflow list. 5265 _timer.reset(); 5266 _timer.start(); 5267 do_work_steal(worker_id, &par_mrias_cl, _collector->hash_seed(worker_id)); 5268 _timer.stop(); 5269 if (PrintCMSStatistics != 0) { 5270 gclog_or_tty->print_cr( 5271 "Finished work stealing in %dth thread: %3.3f sec", 5272 worker_id, _timer.seconds()); 5273 } 5274 } 5275 5276 // Note that parameter "i" is not used. 5277 void 5278 CMSParRemarkTask::do_young_space_rescan(int i, 5279 Par_MarkRefsIntoAndScanClosure* cl, ContiguousSpace* space, 5280 HeapWord** chunk_array, size_t chunk_top) { 5281 // Until all tasks completed: 5282 // . claim an unclaimed task 5283 // . compute region boundaries corresponding to task claimed 5284 // using chunk_array 5285 // . par_oop_iterate(cl) over that region 5286 5287 ResourceMark rm; 5288 HandleMark hm; 5289 5290 SequentialSubTasksDone* pst = space->par_seq_tasks(); 5291 assert(pst->valid(), "Uninitialized use?"); 5292 5293 uint nth_task = 0; 5294 uint n_tasks = pst->n_tasks(); 5295 5296 HeapWord *start, *end; 5297 while (!pst->is_task_claimed(/* reference */ nth_task)) { 5298 // We claimed task # nth_task; compute its boundaries. 5299 if (chunk_top == 0) { // no samples were taken 5300 assert(nth_task == 0 && n_tasks == 1, "Can have only 1 EdenSpace task"); 5301 start = space->bottom(); 5302 end = space->top(); 5303 } else if (nth_task == 0) { 5304 start = space->bottom(); 5305 end = chunk_array[nth_task]; 5306 } else if (nth_task < (uint)chunk_top) { 5307 assert(nth_task >= 1, "Control point invariant"); 5308 start = chunk_array[nth_task - 1]; 5309 end = chunk_array[nth_task]; 5310 } else { 5311 assert(nth_task == (uint)chunk_top, "Control point invariant"); 5312 start = chunk_array[chunk_top - 1]; 5313 end = space->top(); 5314 } 5315 MemRegion mr(start, end); 5316 // Verify that mr is in space 5317 assert(mr.is_empty() || space->used_region().contains(mr), 5318 "Should be in space"); 5319 // Verify that "start" is an object boundary 5320 assert(mr.is_empty() || oop(mr.start())->is_oop(), 5321 "Should be an oop"); 5322 space->par_oop_iterate(mr, cl); 5323 } 5324 pst->all_tasks_completed(); 5325 } 5326 5327 void 5328 CMSParRemarkTask::do_dirty_card_rescan_tasks( 5329 CompactibleFreeListSpace* sp, int i, 5330 Par_MarkRefsIntoAndScanClosure* cl) { 5331 // Until all tasks completed: 5332 // . claim an unclaimed task 5333 // . compute region boundaries corresponding to task claimed 5334 // . transfer dirty bits ct->mut for that region 5335 // . apply rescanclosure to dirty mut bits for that region 5336 5337 ResourceMark rm; 5338 HandleMark hm; 5339 5340 OopTaskQueue* work_q = work_queue(i); 5341 ModUnionClosure modUnionClosure(&(_collector->_modUnionTable)); 5342 // CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! 5343 // CAUTION: This closure has state that persists across calls to 5344 // the work method dirty_range_iterate_clear() in that it has 5345 // imbedded in it a (subtype of) UpwardsObjectClosure. The 5346 // use of that state in the imbedded UpwardsObjectClosure instance 5347 // assumes that the cards are always iterated (even if in parallel 5348 // by several threads) in monotonically increasing order per each 5349 // thread. This is true of the implementation below which picks 5350 // card ranges (chunks) in monotonically increasing order globally 5351 // and, a-fortiori, in monotonically increasing order per thread 5352 // (the latter order being a subsequence of the former). 5353 // If the work code below is ever reorganized into a more chaotic 5354 // work-partitioning form than the current "sequential tasks" 5355 // paradigm, the use of that persistent state will have to be 5356 // revisited and modified appropriately. See also related 5357 // bug 4756801 work on which should examine this code to make 5358 // sure that the changes there do not run counter to the 5359 // assumptions made here and necessary for correctness and 5360 // efficiency. Note also that this code might yield inefficient 5361 // behaviour in the case of very large objects that span one or 5362 // more work chunks. Such objects would potentially be scanned 5363 // several times redundantly. Work on 4756801 should try and 5364 // address that performance anomaly if at all possible. XXX 5365 MemRegion full_span = _collector->_span; 5366 CMSBitMap* bm = &(_collector->_markBitMap); // shared 5367 MarkFromDirtyCardsClosure 5368 greyRescanClosure(_collector, full_span, // entire span of interest 5369 sp, bm, work_q, cl); 5370 5371 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 5372 assert(pst->valid(), "Uninitialized use?"); 5373 uint nth_task = 0; 5374 const int alignment = CardTableModRefBS::card_size * BitsPerWord; 5375 MemRegion span = sp->used_region(); 5376 HeapWord* start_addr = span.start(); 5377 HeapWord* end_addr = (HeapWord*)round_to((intptr_t)span.end(), 5378 alignment); 5379 const size_t chunk_size = sp->rescan_task_size(); // in HeapWord units 5380 assert((HeapWord*)round_to((intptr_t)start_addr, alignment) == 5381 start_addr, "Check alignment"); 5382 assert((size_t)round_to((intptr_t)chunk_size, alignment) == 5383 chunk_size, "Check alignment"); 5384 5385 while (!pst->is_task_claimed(/* reference */ nth_task)) { 5386 // Having claimed the nth_task, compute corresponding mem-region, 5387 // which is a-fortiori aligned correctly (i.e. at a MUT bopundary). 5388 // The alignment restriction ensures that we do not need any 5389 // synchronization with other gang-workers while setting or 5390 // clearing bits in thus chunk of the MUT. 5391 MemRegion this_span = MemRegion(start_addr + nth_task*chunk_size, 5392 start_addr + (nth_task+1)*chunk_size); 5393 // The last chunk's end might be way beyond end of the 5394 // used region. In that case pull back appropriately. 5395 if (this_span.end() > end_addr) { 5396 this_span.set_end(end_addr); 5397 assert(!this_span.is_empty(), "Program logic (calculation of n_tasks)"); 5398 } 5399 // Iterate over the dirty cards covering this chunk, marking them 5400 // precleaned, and setting the corresponding bits in the mod union 5401 // table. Since we have been careful to partition at Card and MUT-word 5402 // boundaries no synchronization is needed between parallel threads. 5403 _collector->_ct->ct_bs()->dirty_card_iterate(this_span, 5404 &modUnionClosure); 5405 5406 // Having transferred these marks into the modUnionTable, 5407 // rescan the marked objects on the dirty cards in the modUnionTable. 5408 // Even if this is at a synchronous collection, the initial marking 5409 // may have been done during an asynchronous collection so there 5410 // may be dirty bits in the mod-union table. 5411 _collector->_modUnionTable.dirty_range_iterate_clear( 5412 this_span, &greyRescanClosure); 5413 _collector->_modUnionTable.verifyNoOneBitsInRange( 5414 this_span.start(), 5415 this_span.end()); 5416 } 5417 pst->all_tasks_completed(); // declare that i am done 5418 } 5419 5420 // . see if we can share work_queues with ParNew? XXX 5421 void 5422 CMSParRemarkTask::do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, 5423 int* seed) { 5424 OopTaskQueue* work_q = work_queue(i); 5425 NOT_PRODUCT(int num_steals = 0;) 5426 oop obj_to_scan; 5427 CMSBitMap* bm = &(_collector->_markBitMap); 5428 5429 while (true) { 5430 // Completely finish any left over work from (an) earlier round(s) 5431 cl->trim_queue(0); 5432 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 5433 (size_t)ParGCDesiredObjsFromOverflowList); 5434 // Now check if there's any work in the overflow list 5435 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads, 5436 // only affects the number of attempts made to get work from the 5437 // overflow list and does not affect the number of workers. Just 5438 // pass ParallelGCThreads so this behavior is unchanged. 5439 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 5440 work_q, 5441 ParallelGCThreads)) { 5442 // found something in global overflow list; 5443 // not yet ready to go stealing work from others. 5444 // We'd like to assert(work_q->size() != 0, ...) 5445 // because we just took work from the overflow list, 5446 // but of course we can't since all of that could have 5447 // been already stolen from us. 5448 // "He giveth and He taketh away." 5449 continue; 5450 } 5451 // Verify that we have no work before we resort to stealing 5452 assert(work_q->size() == 0, "Have work, shouldn't steal"); 5453 // Try to steal from other queues that have work 5454 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 5455 NOT_PRODUCT(num_steals++;) 5456 assert(obj_to_scan->is_oop(), "Oops, not an oop!"); 5457 assert(bm->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 5458 // Do scanning work 5459 obj_to_scan->oop_iterate(cl); 5460 // Loop around, finish this work, and try to steal some more 5461 } else if (terminator()->offer_termination()) { 5462 break; // nirvana from the infinite cycle 5463 } 5464 } 5465 NOT_PRODUCT( 5466 if (PrintCMSStatistics != 0) { 5467 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals); 5468 } 5469 ) 5470 assert(work_q->size() == 0 && _collector->overflow_list_is_empty(), 5471 "Else our work is not yet done"); 5472 } 5473 5474 // Return a thread-local PLAB recording array, as appropriate. 5475 void* CMSCollector::get_data_recorder(int thr_num) { 5476 if (_survivor_plab_array != NULL && 5477 (CMSPLABRecordAlways || 5478 (_collectorState > Marking && _collectorState < FinalMarking))) { 5479 assert(thr_num < (int)ParallelGCThreads, "thr_num is out of bounds"); 5480 ChunkArray* ca = &_survivor_plab_array[thr_num]; 5481 ca->reset(); // clear it so that fresh data is recorded 5482 return (void*) ca; 5483 } else { 5484 return NULL; 5485 } 5486 } 5487 5488 // Reset all the thread-local PLAB recording arrays 5489 void CMSCollector::reset_survivor_plab_arrays() { 5490 for (uint i = 0; i < ParallelGCThreads; i++) { 5491 _survivor_plab_array[i].reset(); 5492 } 5493 } 5494 5495 // Merge the per-thread plab arrays into the global survivor chunk 5496 // array which will provide the partitioning of the survivor space 5497 // for CMS rescan. 5498 void CMSCollector::merge_survivor_plab_arrays(ContiguousSpace* surv, 5499 int no_of_gc_threads) { 5500 assert(_survivor_plab_array != NULL, "Error"); 5501 assert(_survivor_chunk_array != NULL, "Error"); 5502 assert(_collectorState == FinalMarking, "Error"); 5503 for (int j = 0; j < no_of_gc_threads; j++) { 5504 _cursor[j] = 0; 5505 } 5506 HeapWord* top = surv->top(); 5507 size_t i; 5508 for (i = 0; i < _survivor_chunk_capacity; i++) { // all sca entries 5509 HeapWord* min_val = top; // Higher than any PLAB address 5510 uint min_tid = 0; // position of min_val this round 5511 for (int j = 0; j < no_of_gc_threads; j++) { 5512 ChunkArray* cur_sca = &_survivor_plab_array[j]; 5513 if (_cursor[j] == cur_sca->end()) { 5514 continue; 5515 } 5516 assert(_cursor[j] < cur_sca->end(), "ctl pt invariant"); 5517 HeapWord* cur_val = cur_sca->nth(_cursor[j]); 5518 assert(surv->used_region().contains(cur_val), "Out of bounds value"); 5519 if (cur_val < min_val) { 5520 min_tid = j; 5521 min_val = cur_val; 5522 } else { 5523 assert(cur_val < top, "All recorded addresses should be less"); 5524 } 5525 } 5526 // At this point min_val and min_tid are respectively 5527 // the least address in _survivor_plab_array[j]->nth(_cursor[j]) 5528 // and the thread (j) that witnesses that address. 5529 // We record this address in the _survivor_chunk_array[i] 5530 // and increment _cursor[min_tid] prior to the next round i. 5531 if (min_val == top) { 5532 break; 5533 } 5534 _survivor_chunk_array[i] = min_val; 5535 _cursor[min_tid]++; 5536 } 5537 // We are all done; record the size of the _survivor_chunk_array 5538 _survivor_chunk_index = i; // exclusive: [0, i) 5539 if (PrintCMSStatistics > 0) { 5540 gclog_or_tty->print(" (Survivor:" SIZE_FORMAT "chunks) ", i); 5541 } 5542 // Verify that we used up all the recorded entries 5543 #ifdef ASSERT 5544 size_t total = 0; 5545 for (int j = 0; j < no_of_gc_threads; j++) { 5546 assert(_cursor[j] == _survivor_plab_array[j].end(), "Ctl pt invariant"); 5547 total += _cursor[j]; 5548 } 5549 assert(total == _survivor_chunk_index, "Ctl Pt Invariant"); 5550 // Check that the merged array is in sorted order 5551 if (total > 0) { 5552 for (size_t i = 0; i < total - 1; i++) { 5553 if (PrintCMSStatistics > 0) { 5554 gclog_or_tty->print(" (chunk" SIZE_FORMAT ":" INTPTR_FORMAT ") ", 5555 i, _survivor_chunk_array[i]); 5556 } 5557 assert(_survivor_chunk_array[i] < _survivor_chunk_array[i+1], 5558 "Not sorted"); 5559 } 5560 } 5561 #endif // ASSERT 5562 } 5563 5564 // Set up the space's par_seq_tasks structure for work claiming 5565 // for parallel rescan of young gen. 5566 // See ParRescanTask where this is currently used. 5567 void 5568 CMSCollector:: 5569 initialize_sequential_subtasks_for_young_gen_rescan(int n_threads) { 5570 assert(n_threads > 0, "Unexpected n_threads argument"); 5571 DefNewGeneration* dng = (DefNewGeneration*)_young_gen; 5572 5573 // Eden space 5574 { 5575 SequentialSubTasksDone* pst = dng->eden()->par_seq_tasks(); 5576 assert(!pst->valid(), "Clobbering existing data?"); 5577 // Each valid entry in [0, _eden_chunk_index) represents a task. 5578 size_t n_tasks = _eden_chunk_index + 1; 5579 assert(n_tasks == 1 || _eden_chunk_array != NULL, "Error"); 5580 // Sets the condition for completion of the subtask (how many threads 5581 // need to finish in order to be done). 5582 pst->set_n_threads(n_threads); 5583 pst->set_n_tasks((int)n_tasks); 5584 } 5585 5586 // Merge the survivor plab arrays into _survivor_chunk_array 5587 if (_survivor_plab_array != NULL) { 5588 merge_survivor_plab_arrays(dng->from(), n_threads); 5589 } else { 5590 assert(_survivor_chunk_index == 0, "Error"); 5591 } 5592 5593 // To space 5594 { 5595 SequentialSubTasksDone* pst = dng->to()->par_seq_tasks(); 5596 assert(!pst->valid(), "Clobbering existing data?"); 5597 // Sets the condition for completion of the subtask (how many threads 5598 // need to finish in order to be done). 5599 pst->set_n_threads(n_threads); 5600 pst->set_n_tasks(1); 5601 assert(pst->valid(), "Error"); 5602 } 5603 5604 // From space 5605 { 5606 SequentialSubTasksDone* pst = dng->from()->par_seq_tasks(); 5607 assert(!pst->valid(), "Clobbering existing data?"); 5608 size_t n_tasks = _survivor_chunk_index + 1; 5609 assert(n_tasks == 1 || _survivor_chunk_array != NULL, "Error"); 5610 // Sets the condition for completion of the subtask (how many threads 5611 // need to finish in order to be done). 5612 pst->set_n_threads(n_threads); 5613 pst->set_n_tasks((int)n_tasks); 5614 assert(pst->valid(), "Error"); 5615 } 5616 } 5617 5618 // Parallel version of remark 5619 void CMSCollector::do_remark_parallel() { 5620 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5621 FlexibleWorkGang* workers = gch->workers(); 5622 assert(workers != NULL, "Need parallel worker threads."); 5623 // Choose to use the number of GC workers most recently set 5624 // into "active_workers". If active_workers is not set, set it 5625 // to ParallelGCThreads. 5626 int n_workers = workers->active_workers(); 5627 if (n_workers == 0) { 5628 assert(n_workers > 0, "Should have been set during scavenge"); 5629 n_workers = ParallelGCThreads; 5630 workers->set_active_workers(n_workers); 5631 } 5632 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 5633 5634 CMSParRemarkTask tsk(this, 5635 cms_space, 5636 n_workers, workers, task_queues()); 5637 5638 // Set up for parallel process_strong_roots work. 5639 gch->set_par_threads(n_workers); 5640 // We won't be iterating over the cards in the card table updating 5641 // the younger_gen cards, so we shouldn't call the following else 5642 // the verification code as well as subsequent younger_refs_iterate 5643 // code would get confused. XXX 5644 // gch->rem_set()->prepare_for_younger_refs_iterate(true); // parallel 5645 5646 // The young gen rescan work will not be done as part of 5647 // process_strong_roots (which currently doesn't knw how to 5648 // parallelize such a scan), but rather will be broken up into 5649 // a set of parallel tasks (via the sampling that the [abortable] 5650 // preclean phase did of EdenSpace, plus the [two] tasks of 5651 // scanning the [two] survivor spaces. Further fine-grain 5652 // parallelization of the scanning of the survivor spaces 5653 // themselves, and of precleaning of the younger gen itself 5654 // is deferred to the future. 5655 initialize_sequential_subtasks_for_young_gen_rescan(n_workers); 5656 5657 // The dirty card rescan work is broken up into a "sequence" 5658 // of parallel tasks (per constituent space) that are dynamically 5659 // claimed by the parallel threads. 5660 cms_space->initialize_sequential_subtasks_for_rescan(n_workers); 5661 5662 // It turns out that even when we're using 1 thread, doing the work in a 5663 // separate thread causes wide variance in run times. We can't help this 5664 // in the multi-threaded case, but we special-case n=1 here to get 5665 // repeatable measurements of the 1-thread overhead of the parallel code. 5666 if (n_workers > 1) { 5667 // Make refs discovery MT-safe, if it isn't already: it may not 5668 // necessarily be so, since it's possible that we are doing 5669 // ST marking. 5670 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), true); 5671 GenCollectedHeap::StrongRootsScope srs(gch); 5672 workers->run_task(&tsk); 5673 } else { 5674 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false); 5675 GenCollectedHeap::StrongRootsScope srs(gch); 5676 tsk.work(0); 5677 } 5678 5679 gch->set_par_threads(0); // 0 ==> non-parallel. 5680 // restore, single-threaded for now, any preserved marks 5681 // as a result of work_q overflow 5682 restore_preserved_marks_if_any(); 5683 } 5684 5685 // Non-parallel version of remark 5686 void CMSCollector::do_remark_non_parallel() { 5687 ResourceMark rm; 5688 HandleMark hm; 5689 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5690 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false); 5691 5692 MarkRefsIntoAndScanClosure 5693 mrias_cl(_span, ref_processor(), &_markBitMap, NULL /* not precleaning */, 5694 &_markStack, this, 5695 false /* should_yield */, false /* not precleaning */); 5696 MarkFromDirtyCardsClosure 5697 markFromDirtyCardsClosure(this, _span, 5698 NULL, // space is set further below 5699 &_markBitMap, &_markStack, &mrias_cl); 5700 { 5701 TraceTime t("grey object rescan", PrintGCDetails, false, gclog_or_tty); 5702 // Iterate over the dirty cards, setting the corresponding bits in the 5703 // mod union table. 5704 { 5705 ModUnionClosure modUnionClosure(&_modUnionTable); 5706 _ct->ct_bs()->dirty_card_iterate( 5707 _cmsGen->used_region(), 5708 &modUnionClosure); 5709 } 5710 // Having transferred these marks into the modUnionTable, we just need 5711 // to rescan the marked objects on the dirty cards in the modUnionTable. 5712 // The initial marking may have been done during an asynchronous 5713 // collection so there may be dirty bits in the mod-union table. 5714 const int alignment = 5715 CardTableModRefBS::card_size * BitsPerWord; 5716 { 5717 // ... First handle dirty cards in CMS gen 5718 markFromDirtyCardsClosure.set_space(_cmsGen->cmsSpace()); 5719 MemRegion ur = _cmsGen->used_region(); 5720 HeapWord* lb = ur.start(); 5721 HeapWord* ub = (HeapWord*)round_to((intptr_t)ur.end(), alignment); 5722 MemRegion cms_span(lb, ub); 5723 _modUnionTable.dirty_range_iterate_clear(cms_span, 5724 &markFromDirtyCardsClosure); 5725 verify_work_stacks_empty(); 5726 if (PrintCMSStatistics != 0) { 5727 gclog_or_tty->print(" (re-scanned "SIZE_FORMAT" dirty cards in cms gen) ", 5728 markFromDirtyCardsClosure.num_dirty_cards()); 5729 } 5730 } 5731 } 5732 if (VerifyDuringGC && 5733 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 5734 HandleMark hm; // Discard invalid handles created during verification 5735 Universe::verify(); 5736 } 5737 { 5738 TraceTime t("root rescan", PrintGCDetails, false, gclog_or_tty); 5739 5740 verify_work_stacks_empty(); 5741 5742 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 5743 GenCollectedHeap::StrongRootsScope srs(gch); 5744 gch->gen_process_strong_roots(_cmsGen->level(), 5745 true, // younger gens as roots 5746 false, // use the local StrongRootsScope 5747 false, // not scavenging 5748 SharedHeap::ScanningOption(roots_scanning_options()), 5749 &mrias_cl, 5750 true, // walk code active on stacks 5751 NULL, 5752 NULL); // The dirty klasses will be handled below 5753 5754 assert(should_unload_classes() 5755 || (roots_scanning_options() & SharedHeap::SO_CodeCache), 5756 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 5757 } 5758 5759 { 5760 TraceTime t("visit unhandled CLDs", PrintGCDetails, false, gclog_or_tty); 5761 5762 verify_work_stacks_empty(); 5763 5764 // Scan all class loader data objects that might have been introduced 5765 // during concurrent marking. 5766 ResourceMark rm; 5767 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds(); 5768 for (int i = 0; i < array->length(); i++) { 5769 mrias_cl.do_class_loader_data(array->at(i)); 5770 } 5771 5772 // We don't need to keep track of new CLDs anymore. 5773 ClassLoaderDataGraph::remember_new_clds(false); 5774 5775 verify_work_stacks_empty(); 5776 } 5777 5778 { 5779 TraceTime t("dirty klass scan", PrintGCDetails, false, gclog_or_tty); 5780 5781 verify_work_stacks_empty(); 5782 5783 RemarkKlassClosure remark_klass_closure(&mrias_cl); 5784 ClassLoaderDataGraph::classes_do(&remark_klass_closure); 5785 5786 verify_work_stacks_empty(); 5787 } 5788 5789 // We might have added oops to ClassLoaderData::_handles during the 5790 // concurrent marking phase. These oops point to newly allocated objects 5791 // that are guaranteed to be kept alive. Either by the direct allocation 5792 // code, or when the young collector processes the strong roots. Hence, 5793 // we don't have to revisit the _handles block during the remark phase. 5794 5795 verify_work_stacks_empty(); 5796 // Restore evacuated mark words, if any, used for overflow list links 5797 if (!CMSOverflowEarlyRestoration) { 5798 restore_preserved_marks_if_any(); 5799 } 5800 verify_overflow_empty(); 5801 } 5802 5803 //////////////////////////////////////////////////////// 5804 // Parallel Reference Processing Task Proxy Class 5805 //////////////////////////////////////////////////////// 5806 class CMSRefProcTaskProxy: public AbstractGangTaskWOopQueues { 5807 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask; 5808 CMSCollector* _collector; 5809 CMSBitMap* _mark_bit_map; 5810 const MemRegion _span; 5811 ProcessTask& _task; 5812 5813 public: 5814 CMSRefProcTaskProxy(ProcessTask& task, 5815 CMSCollector* collector, 5816 const MemRegion& span, 5817 CMSBitMap* mark_bit_map, 5818 AbstractWorkGang* workers, 5819 OopTaskQueueSet* task_queues): 5820 // XXX Should superclass AGTWOQ also know about AWG since it knows 5821 // about the task_queues used by the AWG? Then it could initialize 5822 // the terminator() object. See 6984287. The set_for_termination() 5823 // below is a temporary band-aid for the regression in 6984287. 5824 AbstractGangTaskWOopQueues("Process referents by policy in parallel", 5825 task_queues), 5826 _task(task), 5827 _collector(collector), _span(span), _mark_bit_map(mark_bit_map) 5828 { 5829 assert(_collector->_span.equals(_span) && !_span.is_empty(), 5830 "Inconsistency in _span"); 5831 set_for_termination(workers->active_workers()); 5832 } 5833 5834 OopTaskQueueSet* task_queues() { return queues(); } 5835 5836 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 5837 5838 void do_work_steal(int i, 5839 CMSParDrainMarkingStackClosure* drain, 5840 CMSParKeepAliveClosure* keep_alive, 5841 int* seed); 5842 5843 virtual void work(uint worker_id); 5844 }; 5845 5846 void CMSRefProcTaskProxy::work(uint worker_id) { 5847 assert(_collector->_span.equals(_span), "Inconsistency in _span"); 5848 CMSParKeepAliveClosure par_keep_alive(_collector, _span, 5849 _mark_bit_map, 5850 work_queue(worker_id)); 5851 CMSParDrainMarkingStackClosure par_drain_stack(_collector, _span, 5852 _mark_bit_map, 5853 work_queue(worker_id)); 5854 CMSIsAliveClosure is_alive_closure(_span, _mark_bit_map); 5855 _task.work(worker_id, is_alive_closure, par_keep_alive, par_drain_stack); 5856 if (_task.marks_oops_alive()) { 5857 do_work_steal(worker_id, &par_drain_stack, &par_keep_alive, 5858 _collector->hash_seed(worker_id)); 5859 } 5860 assert(work_queue(worker_id)->size() == 0, "work_queue should be empty"); 5861 assert(_collector->_overflow_list == NULL, "non-empty _overflow_list"); 5862 } 5863 5864 class CMSRefEnqueueTaskProxy: public AbstractGangTask { 5865 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask; 5866 EnqueueTask& _task; 5867 5868 public: 5869 CMSRefEnqueueTaskProxy(EnqueueTask& task) 5870 : AbstractGangTask("Enqueue reference objects in parallel"), 5871 _task(task) 5872 { } 5873 5874 virtual void work(uint worker_id) 5875 { 5876 _task.work(worker_id); 5877 } 5878 }; 5879 5880 CMSParKeepAliveClosure::CMSParKeepAliveClosure(CMSCollector* collector, 5881 MemRegion span, CMSBitMap* bit_map, OopTaskQueue* work_queue): 5882 _span(span), 5883 _bit_map(bit_map), 5884 _work_queue(work_queue), 5885 _mark_and_push(collector, span, bit_map, work_queue), 5886 _low_water_mark(MIN2((uint)(work_queue->max_elems()/4), 5887 (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads))) 5888 { } 5889 5890 // . see if we can share work_queues with ParNew? XXX 5891 void CMSRefProcTaskProxy::do_work_steal(int i, 5892 CMSParDrainMarkingStackClosure* drain, 5893 CMSParKeepAliveClosure* keep_alive, 5894 int* seed) { 5895 OopTaskQueue* work_q = work_queue(i); 5896 NOT_PRODUCT(int num_steals = 0;) 5897 oop obj_to_scan; 5898 5899 while (true) { 5900 // Completely finish any left over work from (an) earlier round(s) 5901 drain->trim_queue(0); 5902 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 5903 (size_t)ParGCDesiredObjsFromOverflowList); 5904 // Now check if there's any work in the overflow list 5905 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads, 5906 // only affects the number of attempts made to get work from the 5907 // overflow list and does not affect the number of workers. Just 5908 // pass ParallelGCThreads so this behavior is unchanged. 5909 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 5910 work_q, 5911 ParallelGCThreads)) { 5912 // Found something in global overflow list; 5913 // not yet ready to go stealing work from others. 5914 // We'd like to assert(work_q->size() != 0, ...) 5915 // because we just took work from the overflow list, 5916 // but of course we can't, since all of that might have 5917 // been already stolen from us. 5918 continue; 5919 } 5920 // Verify that we have no work before we resort to stealing 5921 assert(work_q->size() == 0, "Have work, shouldn't steal"); 5922 // Try to steal from other queues that have work 5923 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 5924 NOT_PRODUCT(num_steals++;) 5925 assert(obj_to_scan->is_oop(), "Oops, not an oop!"); 5926 assert(_mark_bit_map->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 5927 // Do scanning work 5928 obj_to_scan->oop_iterate(keep_alive); 5929 // Loop around, finish this work, and try to steal some more 5930 } else if (terminator()->offer_termination()) { 5931 break; // nirvana from the infinite cycle 5932 } 5933 } 5934 NOT_PRODUCT( 5935 if (PrintCMSStatistics != 0) { 5936 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals); 5937 } 5938 ) 5939 } 5940 5941 void CMSRefProcTaskExecutor::execute(ProcessTask& task) 5942 { 5943 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5944 FlexibleWorkGang* workers = gch->workers(); 5945 assert(workers != NULL, "Need parallel worker threads."); 5946 CMSRefProcTaskProxy rp_task(task, &_collector, 5947 _collector.ref_processor()->span(), 5948 _collector.markBitMap(), 5949 workers, _collector.task_queues()); 5950 workers->run_task(&rp_task); 5951 } 5952 5953 void CMSRefProcTaskExecutor::execute(EnqueueTask& task) 5954 { 5955 5956 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5957 FlexibleWorkGang* workers = gch->workers(); 5958 assert(workers != NULL, "Need parallel worker threads."); 5959 CMSRefEnqueueTaskProxy enq_task(task); 5960 workers->run_task(&enq_task); 5961 } 5962 5963 void CMSCollector::refProcessingWork(bool asynch, bool clear_all_soft_refs) { 5964 5965 ResourceMark rm; 5966 HandleMark hm; 5967 5968 ReferenceProcessor* rp = ref_processor(); 5969 assert(rp->span().equals(_span), "Spans should be equal"); 5970 assert(!rp->enqueuing_is_done(), "Enqueuing should not be complete"); 5971 // Process weak references. 5972 rp->setup_policy(clear_all_soft_refs); 5973 verify_work_stacks_empty(); 5974 5975 CMSKeepAliveClosure cmsKeepAliveClosure(this, _span, &_markBitMap, 5976 &_markStack, false /* !preclean */); 5977 CMSDrainMarkingStackClosure cmsDrainMarkingStackClosure(this, 5978 _span, &_markBitMap, &_markStack, 5979 &cmsKeepAliveClosure, false /* !preclean */); 5980 { 5981 TraceTime t("weak refs processing", PrintGCDetails, false, gclog_or_tty); 5982 if (rp->processing_is_mt()) { 5983 // Set the degree of MT here. If the discovery is done MT, there 5984 // may have been a different number of threads doing the discovery 5985 // and a different number of discovered lists may have Ref objects. 5986 // That is OK as long as the Reference lists are balanced (see 5987 // balance_all_queues() and balance_queues()). 5988 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5989 int active_workers = ParallelGCThreads; 5990 FlexibleWorkGang* workers = gch->workers(); 5991 if (workers != NULL) { 5992 active_workers = workers->active_workers(); 5993 // The expectation is that active_workers will have already 5994 // been set to a reasonable value. If it has not been set, 5995 // investigate. 5996 assert(active_workers > 0, "Should have been set during scavenge"); 5997 } 5998 rp->set_active_mt_degree(active_workers); 5999 CMSRefProcTaskExecutor task_executor(*this); 6000 rp->process_discovered_references(&_is_alive_closure, 6001 &cmsKeepAliveClosure, 6002 &cmsDrainMarkingStackClosure, 6003 &task_executor); 6004 } else { 6005 rp->process_discovered_references(&_is_alive_closure, 6006 &cmsKeepAliveClosure, 6007 &cmsDrainMarkingStackClosure, 6008 NULL); 6009 } 6010 } 6011 6012 // This is the point where the entire marking should have completed. 6013 verify_work_stacks_empty(); 6014 6015 if (should_unload_classes()) { 6016 { 6017 TraceTime t("class unloading", PrintGCDetails, false, gclog_or_tty); 6018 6019 // Unload classes and purge the SystemDictionary. 6020 bool purged_class = SystemDictionary::do_unloading(&_is_alive_closure); 6021 6022 // Unload nmethods. 6023 CodeCache::do_unloading(&_is_alive_closure, purged_class); 6024 6025 // Prune dead klasses from subklass/sibling/implementor lists. 6026 Klass::clean_weak_klass_links(&_is_alive_closure); 6027 } 6028 6029 { 6030 TraceTime t("scrub symbol table", PrintGCDetails, false, gclog_or_tty); 6031 // Clean up unreferenced symbols in symbol table. 6032 SymbolTable::unlink(); 6033 } 6034 } 6035 6036 // CMS doesn't use the StringTable as hard roots when class unloading is turned off. 6037 // Need to check if we really scanned the StringTable. 6038 if ((roots_scanning_options() & SharedHeap::SO_Strings) == 0) { 6039 TraceTime t("scrub string table", PrintGCDetails, false, gclog_or_tty); 6040 // Delete entries for dead interned strings. 6041 StringTable::unlink(&_is_alive_closure); 6042 } 6043 6044 // Restore any preserved marks as a result of mark stack or 6045 // work queue overflow 6046 restore_preserved_marks_if_any(); // done single-threaded for now 6047 6048 rp->set_enqueuing_is_done(true); 6049 if (rp->processing_is_mt()) { 6050 rp->balance_all_queues(); 6051 CMSRefProcTaskExecutor task_executor(*this); 6052 rp->enqueue_discovered_references(&task_executor); 6053 } else { 6054 rp->enqueue_discovered_references(NULL); 6055 } 6056 rp->verify_no_references_recorded(); 6057 assert(!rp->discovery_enabled(), "should have been disabled"); 6058 } 6059 6060 #ifndef PRODUCT 6061 void CMSCollector::check_correct_thread_executing() { 6062 Thread* t = Thread::current(); 6063 // Only the VM thread or the CMS thread should be here. 6064 assert(t->is_ConcurrentGC_thread() || t->is_VM_thread(), 6065 "Unexpected thread type"); 6066 // If this is the vm thread, the foreground process 6067 // should not be waiting. Note that _foregroundGCIsActive is 6068 // true while the foreground collector is waiting. 6069 if (_foregroundGCShouldWait) { 6070 // We cannot be the VM thread 6071 assert(t->is_ConcurrentGC_thread(), 6072 "Should be CMS thread"); 6073 } else { 6074 // We can be the CMS thread only if we are in a stop-world 6075 // phase of CMS collection. 6076 if (t->is_ConcurrentGC_thread()) { 6077 assert(_collectorState == InitialMarking || 6078 _collectorState == FinalMarking, 6079 "Should be a stop-world phase"); 6080 // The CMS thread should be holding the CMS_token. 6081 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6082 "Potential interference with concurrently " 6083 "executing VM thread"); 6084 } 6085 } 6086 } 6087 #endif 6088 6089 void CMSCollector::sweep(bool asynch) { 6090 assert(_collectorState == Sweeping, "just checking"); 6091 check_correct_thread_executing(); 6092 verify_work_stacks_empty(); 6093 verify_overflow_empty(); 6094 increment_sweep_count(); 6095 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 6096 6097 _inter_sweep_timer.stop(); 6098 _inter_sweep_estimate.sample(_inter_sweep_timer.seconds()); 6099 size_policy()->avg_cms_free_at_sweep()->sample(_cmsGen->free()); 6100 6101 assert(!_intra_sweep_timer.is_active(), "Should not be active"); 6102 _intra_sweep_timer.reset(); 6103 _intra_sweep_timer.start(); 6104 if (asynch) { 6105 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 6106 CMSPhaseAccounting pa(this, "sweep", !PrintGCDetails); 6107 // First sweep the old gen 6108 { 6109 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(), 6110 bitMapLock()); 6111 sweepWork(_cmsGen, asynch); 6112 } 6113 6114 // Update Universe::_heap_*_at_gc figures. 6115 // We need all the free list locks to make the abstract state 6116 // transition from Sweeping to Resetting. See detailed note 6117 // further below. 6118 { 6119 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock()); 6120 // Update heap occupancy information which is used as 6121 // input to soft ref clearing policy at the next gc. 6122 Universe::update_heap_info_at_gc(); 6123 _collectorState = Resizing; 6124 } 6125 } else { 6126 // already have needed locks 6127 sweepWork(_cmsGen, asynch); 6128 // Update heap occupancy information which is used as 6129 // input to soft ref clearing policy at the next gc. 6130 Universe::update_heap_info_at_gc(); 6131 _collectorState = Resizing; 6132 } 6133 verify_work_stacks_empty(); 6134 verify_overflow_empty(); 6135 6136 if (should_unload_classes()) { 6137 ClassLoaderDataGraph::purge(); 6138 } 6139 6140 _intra_sweep_timer.stop(); 6141 _intra_sweep_estimate.sample(_intra_sweep_timer.seconds()); 6142 6143 _inter_sweep_timer.reset(); 6144 _inter_sweep_timer.start(); 6145 6146 // We need to use a monotonically non-deccreasing time in ms 6147 // or we will see time-warp warnings and os::javaTimeMillis() 6148 // does not guarantee monotonicity. 6149 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC; 6150 update_time_of_last_gc(now); 6151 6152 // NOTE on abstract state transitions: 6153 // Mutators allocate-live and/or mark the mod-union table dirty 6154 // based on the state of the collection. The former is done in 6155 // the interval [Marking, Sweeping] and the latter in the interval 6156 // [Marking, Sweeping). Thus the transitions into the Marking state 6157 // and out of the Sweeping state must be synchronously visible 6158 // globally to the mutators. 6159 // The transition into the Marking state happens with the world 6160 // stopped so the mutators will globally see it. Sweeping is 6161 // done asynchronously by the background collector so the transition 6162 // from the Sweeping state to the Resizing state must be done 6163 // under the freelistLock (as is the check for whether to 6164 // allocate-live and whether to dirty the mod-union table). 6165 assert(_collectorState == Resizing, "Change of collector state to" 6166 " Resizing must be done under the freelistLocks (plural)"); 6167 6168 // Now that sweeping has been completed, we clear 6169 // the incremental_collection_failed flag, 6170 // thus inviting a younger gen collection to promote into 6171 // this generation. If such a promotion may still fail, 6172 // the flag will be set again when a young collection is 6173 // attempted. 6174 GenCollectedHeap* gch = GenCollectedHeap::heap(); 6175 gch->clear_incremental_collection_failed(); // Worth retrying as fresh space may have been freed up 6176 gch->update_full_collections_completed(_collection_count_start); 6177 } 6178 6179 // FIX ME!!! Looks like this belongs in CFLSpace, with 6180 // CMSGen merely delegating to it. 6181 void ConcurrentMarkSweepGeneration::setNearLargestChunk() { 6182 double nearLargestPercent = FLSLargestBlockCoalesceProximity; 6183 HeapWord* minAddr = _cmsSpace->bottom(); 6184 HeapWord* largestAddr = 6185 (HeapWord*) _cmsSpace->dictionary()->find_largest_dict(); 6186 if (largestAddr == NULL) { 6187 // The dictionary appears to be empty. In this case 6188 // try to coalesce at the end of the heap. 6189 largestAddr = _cmsSpace->end(); 6190 } 6191 size_t largestOffset = pointer_delta(largestAddr, minAddr); 6192 size_t nearLargestOffset = 6193 (size_t)((double)largestOffset * nearLargestPercent) - MinChunkSize; 6194 if (PrintFLSStatistics != 0) { 6195 gclog_or_tty->print_cr( 6196 "CMS: Large Block: " PTR_FORMAT ";" 6197 " Proximity: " PTR_FORMAT " -> " PTR_FORMAT, 6198 largestAddr, 6199 _cmsSpace->nearLargestChunk(), minAddr + nearLargestOffset); 6200 } 6201 _cmsSpace->set_nearLargestChunk(minAddr + nearLargestOffset); 6202 } 6203 6204 bool ConcurrentMarkSweepGeneration::isNearLargestChunk(HeapWord* addr) { 6205 return addr >= _cmsSpace->nearLargestChunk(); 6206 } 6207 6208 FreeChunk* ConcurrentMarkSweepGeneration::find_chunk_at_end() { 6209 return _cmsSpace->find_chunk_at_end(); 6210 } 6211 6212 void ConcurrentMarkSweepGeneration::update_gc_stats(int current_level, 6213 bool full) { 6214 // The next lower level has been collected. Gather any statistics 6215 // that are of interest at this point. 6216 if (!full && (current_level + 1) == level()) { 6217 // Gather statistics on the young generation collection. 6218 collector()->stats().record_gc0_end(used()); 6219 } 6220 } 6221 6222 CMSAdaptiveSizePolicy* ConcurrentMarkSweepGeneration::size_policy() { 6223 GenCollectedHeap* gch = GenCollectedHeap::heap(); 6224 assert(gch->kind() == CollectedHeap::GenCollectedHeap, 6225 "Wrong type of heap"); 6226 CMSAdaptiveSizePolicy* sp = (CMSAdaptiveSizePolicy*) 6227 gch->gen_policy()->size_policy(); 6228 assert(sp->is_gc_cms_adaptive_size_policy(), 6229 "Wrong type of size policy"); 6230 return sp; 6231 } 6232 6233 void ConcurrentMarkSweepGeneration::rotate_debug_collection_type() { 6234 if (PrintGCDetails && Verbose) { 6235 gclog_or_tty->print("Rotate from %d ", _debug_collection_type); 6236 } 6237 _debug_collection_type = (CollectionTypes) (_debug_collection_type + 1); 6238 _debug_collection_type = 6239 (CollectionTypes) (_debug_collection_type % Unknown_collection_type); 6240 if (PrintGCDetails && Verbose) { 6241 gclog_or_tty->print_cr("to %d ", _debug_collection_type); 6242 } 6243 } 6244 6245 void CMSCollector::sweepWork(ConcurrentMarkSweepGeneration* gen, 6246 bool asynch) { 6247 // We iterate over the space(s) underlying this generation, 6248 // checking the mark bit map to see if the bits corresponding 6249 // to specific blocks are marked or not. Blocks that are 6250 // marked are live and are not swept up. All remaining blocks 6251 // are swept up, with coalescing on-the-fly as we sweep up 6252 // contiguous free and/or garbage blocks: 6253 // We need to ensure that the sweeper synchronizes with allocators 6254 // and stop-the-world collectors. In particular, the following 6255 // locks are used: 6256 // . CMS token: if this is held, a stop the world collection cannot occur 6257 // . freelistLock: if this is held no allocation can occur from this 6258 // generation by another thread 6259 // . bitMapLock: if this is held, no other thread can access or update 6260 // 6261 6262 // Note that we need to hold the freelistLock if we use 6263 // block iterate below; else the iterator might go awry if 6264 // a mutator (or promotion) causes block contents to change 6265 // (for instance if the allocator divvies up a block). 6266 // If we hold the free list lock, for all practical purposes 6267 // young generation GC's can't occur (they'll usually need to 6268 // promote), so we might as well prevent all young generation 6269 // GC's while we do a sweeping step. For the same reason, we might 6270 // as well take the bit map lock for the entire duration 6271 6272 // check that we hold the requisite locks 6273 assert(have_cms_token(), "Should hold cms token"); 6274 assert( (asynch && ConcurrentMarkSweepThread::cms_thread_has_cms_token()) 6275 || (!asynch && ConcurrentMarkSweepThread::vm_thread_has_cms_token()), 6276 "Should possess CMS token to sweep"); 6277 assert_lock_strong(gen->freelistLock()); 6278 assert_lock_strong(bitMapLock()); 6279 6280 assert(!_inter_sweep_timer.is_active(), "Was switched off in an outer context"); 6281 assert(_intra_sweep_timer.is_active(), "Was switched on in an outer context"); 6282 gen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()), 6283 _inter_sweep_estimate.padded_average(), 6284 _intra_sweep_estimate.padded_average()); 6285 gen->setNearLargestChunk(); 6286 6287 { 6288 SweepClosure sweepClosure(this, gen, &_markBitMap, 6289 CMSYield && asynch); 6290 gen->cmsSpace()->blk_iterate_careful(&sweepClosure); 6291 // We need to free-up/coalesce garbage/blocks from a 6292 // co-terminal free run. This is done in the SweepClosure 6293 // destructor; so, do not remove this scope, else the 6294 // end-of-sweep-census below will be off by a little bit. 6295 } 6296 gen->cmsSpace()->sweep_completed(); 6297 gen->cmsSpace()->endSweepFLCensus(sweep_count()); 6298 if (should_unload_classes()) { // unloaded classes this cycle, 6299 _concurrent_cycles_since_last_unload = 0; // ... reset count 6300 } else { // did not unload classes, 6301 _concurrent_cycles_since_last_unload++; // ... increment count 6302 } 6303 } 6304 6305 // Reset CMS data structures (for now just the marking bit map) 6306 // preparatory for the next cycle. 6307 void CMSCollector::reset(bool asynch) { 6308 GenCollectedHeap* gch = GenCollectedHeap::heap(); 6309 CMSAdaptiveSizePolicy* sp = size_policy(); 6310 AdaptiveSizePolicyOutput(sp, gch->total_collections()); 6311 if (asynch) { 6312 CMSTokenSyncWithLocks ts(true, bitMapLock()); 6313 6314 // If the state is not "Resetting", the foreground thread 6315 // has done a collection and the resetting. 6316 if (_collectorState != Resetting) { 6317 assert(_collectorState == Idling, "The state should only change" 6318 " because the foreground collector has finished the collection"); 6319 return; 6320 } 6321 6322 // Clear the mark bitmap (no grey objects to start with) 6323 // for the next cycle. 6324 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 6325 CMSPhaseAccounting cmspa(this, "reset", !PrintGCDetails); 6326 6327 HeapWord* curAddr = _markBitMap.startWord(); 6328 while (curAddr < _markBitMap.endWord()) { 6329 size_t remaining = pointer_delta(_markBitMap.endWord(), curAddr); 6330 MemRegion chunk(curAddr, MIN2(CMSBitMapYieldQuantum, remaining)); 6331 _markBitMap.clear_large_range(chunk); 6332 if (ConcurrentMarkSweepThread::should_yield() && 6333 !foregroundGCIsActive() && 6334 CMSYield) { 6335 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6336 "CMS thread should hold CMS token"); 6337 assert_lock_strong(bitMapLock()); 6338 bitMapLock()->unlock(); 6339 ConcurrentMarkSweepThread::desynchronize(true); 6340 ConcurrentMarkSweepThread::acknowledge_yield_request(); 6341 stopTimer(); 6342 if (PrintCMSStatistics != 0) { 6343 incrementYields(); 6344 } 6345 icms_wait(); 6346 6347 // See the comment in coordinator_yield() 6348 for (unsigned i = 0; i < CMSYieldSleepCount && 6349 ConcurrentMarkSweepThread::should_yield() && 6350 !CMSCollector::foregroundGCIsActive(); ++i) { 6351 os::sleep(Thread::current(), 1, false); 6352 ConcurrentMarkSweepThread::acknowledge_yield_request(); 6353 } 6354 6355 ConcurrentMarkSweepThread::synchronize(true); 6356 bitMapLock()->lock_without_safepoint_check(); 6357 startTimer(); 6358 } 6359 curAddr = chunk.end(); 6360 } 6361 // A successful mostly concurrent collection has been done. 6362 // Because only the full (i.e., concurrent mode failure) collections 6363 // are being measured for gc overhead limits, clean the "near" flag 6364 // and count. 6365 sp->reset_gc_overhead_limit_count(); 6366 _collectorState = Idling; 6367 } else { 6368 // already have the lock 6369 assert(_collectorState == Resetting, "just checking"); 6370 assert_lock_strong(bitMapLock()); 6371 _markBitMap.clear_all(); 6372 _collectorState = Idling; 6373 } 6374 6375 // Stop incremental mode after a cycle completes, so that any future cycles 6376 // are triggered by allocation. 6377 stop_icms(); 6378 6379 NOT_PRODUCT( 6380 if (RotateCMSCollectionTypes) { 6381 _cmsGen->rotate_debug_collection_type(); 6382 } 6383 ) 6384 } 6385 6386 void CMSCollector::do_CMS_operation(CMS_op_type op, GCCause::Cause gc_cause) { 6387 gclog_or_tty->date_stamp(PrintGC && PrintGCDateStamps); 6388 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 6389 TraceTime t(GCCauseString("GC", gc_cause), PrintGC, !PrintGCDetails, gclog_or_tty); 6390 TraceCollectorStats tcs(counters()); 6391 6392 switch (op) { 6393 case CMS_op_checkpointRootsInitial: { 6394 SvcGCMarker sgcm(SvcGCMarker::OTHER); 6395 checkpointRootsInitial(true); // asynch 6396 if (PrintGC) { 6397 _cmsGen->printOccupancy("initial-mark"); 6398 } 6399 break; 6400 } 6401 case CMS_op_checkpointRootsFinal: { 6402 SvcGCMarker sgcm(SvcGCMarker::OTHER); 6403 checkpointRootsFinal(true, // asynch 6404 false, // !clear_all_soft_refs 6405 false); // !init_mark_was_synchronous 6406 if (PrintGC) { 6407 _cmsGen->printOccupancy("remark"); 6408 } 6409 break; 6410 } 6411 default: 6412 fatal("No such CMS_op"); 6413 } 6414 } 6415 6416 #ifndef PRODUCT 6417 size_t const CMSCollector::skip_header_HeapWords() { 6418 return FreeChunk::header_size(); 6419 } 6420 6421 // Try and collect here conditions that should hold when 6422 // CMS thread is exiting. The idea is that the foreground GC 6423 // thread should not be blocked if it wants to terminate 6424 // the CMS thread and yet continue to run the VM for a while 6425 // after that. 6426 void CMSCollector::verify_ok_to_terminate() const { 6427 assert(Thread::current()->is_ConcurrentGC_thread(), 6428 "should be called by CMS thread"); 6429 assert(!_foregroundGCShouldWait, "should be false"); 6430 // We could check here that all the various low-level locks 6431 // are not held by the CMS thread, but that is overkill; see 6432 // also CMSThread::verify_ok_to_terminate() where the CGC_lock 6433 // is checked. 6434 } 6435 #endif 6436 6437 size_t CMSCollector::block_size_using_printezis_bits(HeapWord* addr) const { 6438 assert(_markBitMap.isMarked(addr) && _markBitMap.isMarked(addr + 1), 6439 "missing Printezis mark?"); 6440 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 6441 size_t size = pointer_delta(nextOneAddr + 1, addr); 6442 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 6443 "alignment problem"); 6444 assert(size >= 3, "Necessary for Printezis marks to work"); 6445 return size; 6446 } 6447 6448 // A variant of the above (block_size_using_printezis_bits()) except 6449 // that we return 0 if the P-bits are not yet set. 6450 size_t CMSCollector::block_size_if_printezis_bits(HeapWord* addr) const { 6451 if (_markBitMap.isMarked(addr + 1)) { 6452 assert(_markBitMap.isMarked(addr), "P-bit can be set only for marked objects"); 6453 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 6454 size_t size = pointer_delta(nextOneAddr + 1, addr); 6455 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 6456 "alignment problem"); 6457 assert(size >= 3, "Necessary for Printezis marks to work"); 6458 return size; 6459 } 6460 return 0; 6461 } 6462 6463 HeapWord* CMSCollector::next_card_start_after_block(HeapWord* addr) const { 6464 size_t sz = 0; 6465 oop p = (oop)addr; 6466 if (p->klass_or_null() != NULL) { 6467 sz = CompactibleFreeListSpace::adjustObjectSize(p->size()); 6468 } else { 6469 sz = block_size_using_printezis_bits(addr); 6470 } 6471 assert(sz > 0, "size must be nonzero"); 6472 HeapWord* next_block = addr + sz; 6473 HeapWord* next_card = (HeapWord*)round_to((uintptr_t)next_block, 6474 CardTableModRefBS::card_size); 6475 assert(round_down((uintptr_t)addr, CardTableModRefBS::card_size) < 6476 round_down((uintptr_t)next_card, CardTableModRefBS::card_size), 6477 "must be different cards"); 6478 return next_card; 6479 } 6480 6481 6482 // CMS Bit Map Wrapper ///////////////////////////////////////// 6483 6484 // Construct a CMS bit map infrastructure, but don't create the 6485 // bit vector itself. That is done by a separate call CMSBitMap::allocate() 6486 // further below. 6487 CMSBitMap::CMSBitMap(int shifter, int mutex_rank, const char* mutex_name): 6488 _bm(), 6489 _shifter(shifter), 6490 _lock(mutex_rank >= 0 ? new Mutex(mutex_rank, mutex_name, true) : NULL) 6491 { 6492 _bmStartWord = 0; 6493 _bmWordSize = 0; 6494 } 6495 6496 bool CMSBitMap::allocate(MemRegion mr) { 6497 _bmStartWord = mr.start(); 6498 _bmWordSize = mr.word_size(); 6499 ReservedSpace brs(ReservedSpace::allocation_align_size_up( 6500 (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1)); 6501 if (!brs.is_reserved()) { 6502 warning("CMS bit map allocation failure"); 6503 return false; 6504 } 6505 // For now we'll just commit all of the bit map up fromt. 6506 // Later on we'll try to be more parsimonious with swap. 6507 if (!_virtual_space.initialize(brs, brs.size())) { 6508 warning("CMS bit map backing store failure"); 6509 return false; 6510 } 6511 assert(_virtual_space.committed_size() == brs.size(), 6512 "didn't reserve backing store for all of CMS bit map?"); 6513 _bm.set_map((BitMap::bm_word_t*)_virtual_space.low()); 6514 assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >= 6515 _bmWordSize, "inconsistency in bit map sizing"); 6516 _bm.set_size(_bmWordSize >> _shifter); 6517 6518 // bm.clear(); // can we rely on getting zero'd memory? verify below 6519 assert(isAllClear(), 6520 "Expected zero'd memory from ReservedSpace constructor"); 6521 assert(_bm.size() == heapWordDiffToOffsetDiff(sizeInWords()), 6522 "consistency check"); 6523 return true; 6524 } 6525 6526 void CMSBitMap::dirty_range_iterate_clear(MemRegion mr, MemRegionClosure* cl) { 6527 HeapWord *next_addr, *end_addr, *last_addr; 6528 assert_locked(); 6529 assert(covers(mr), "out-of-range error"); 6530 // XXX assert that start and end are appropriately aligned 6531 for (next_addr = mr.start(), end_addr = mr.end(); 6532 next_addr < end_addr; next_addr = last_addr) { 6533 MemRegion dirty_region = getAndClearMarkedRegion(next_addr, end_addr); 6534 last_addr = dirty_region.end(); 6535 if (!dirty_region.is_empty()) { 6536 cl->do_MemRegion(dirty_region); 6537 } else { 6538 assert(last_addr == end_addr, "program logic"); 6539 return; 6540 } 6541 } 6542 } 6543 6544 void CMSBitMap::print_on_error(outputStream* st, const char* prefix) const { 6545 _bm.print_on_error(st, prefix); 6546 } 6547 6548 #ifndef PRODUCT 6549 void CMSBitMap::assert_locked() const { 6550 CMSLockVerifier::assert_locked(lock()); 6551 } 6552 6553 bool CMSBitMap::covers(MemRegion mr) const { 6554 // assert(_bm.map() == _virtual_space.low(), "map inconsistency"); 6555 assert((size_t)_bm.size() == (_bmWordSize >> _shifter), 6556 "size inconsistency"); 6557 return (mr.start() >= _bmStartWord) && 6558 (mr.end() <= endWord()); 6559 } 6560 6561 bool CMSBitMap::covers(HeapWord* start, size_t size) const { 6562 return (start >= _bmStartWord && (start + size) <= endWord()); 6563 } 6564 6565 void CMSBitMap::verifyNoOneBitsInRange(HeapWord* left, HeapWord* right) { 6566 // verify that there are no 1 bits in the interval [left, right) 6567 FalseBitMapClosure falseBitMapClosure; 6568 iterate(&falseBitMapClosure, left, right); 6569 } 6570 6571 void CMSBitMap::region_invariant(MemRegion mr) 6572 { 6573 assert_locked(); 6574 // mr = mr.intersection(MemRegion(_bmStartWord, _bmWordSize)); 6575 assert(!mr.is_empty(), "unexpected empty region"); 6576 assert(covers(mr), "mr should be covered by bit map"); 6577 // convert address range into offset range 6578 size_t start_ofs = heapWordToOffset(mr.start()); 6579 // Make sure that end() is appropriately aligned 6580 assert(mr.end() == (HeapWord*)round_to((intptr_t)mr.end(), 6581 (1 << (_shifter+LogHeapWordSize))), 6582 "Misaligned mr.end()"); 6583 size_t end_ofs = heapWordToOffset(mr.end()); 6584 assert(end_ofs > start_ofs, "Should mark at least one bit"); 6585 } 6586 6587 #endif 6588 6589 bool CMSMarkStack::allocate(size_t size) { 6590 // allocate a stack of the requisite depth 6591 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 6592 size * sizeof(oop))); 6593 if (!rs.is_reserved()) { 6594 warning("CMSMarkStack allocation failure"); 6595 return false; 6596 } 6597 if (!_virtual_space.initialize(rs, rs.size())) { 6598 warning("CMSMarkStack backing store failure"); 6599 return false; 6600 } 6601 assert(_virtual_space.committed_size() == rs.size(), 6602 "didn't reserve backing store for all of CMS stack?"); 6603 _base = (oop*)(_virtual_space.low()); 6604 _index = 0; 6605 _capacity = size; 6606 NOT_PRODUCT(_max_depth = 0); 6607 return true; 6608 } 6609 6610 // XXX FIX ME !!! In the MT case we come in here holding a 6611 // leaf lock. For printing we need to take a further lock 6612 // which has lower rank. We need to recallibrate the two 6613 // lock-ranks involved in order to be able to rpint the 6614 // messages below. (Or defer the printing to the caller. 6615 // For now we take the expedient path of just disabling the 6616 // messages for the problematic case.) 6617 void CMSMarkStack::expand() { 6618 assert(_capacity <= MarkStackSizeMax, "stack bigger than permitted"); 6619 if (_capacity == MarkStackSizeMax) { 6620 if (_hit_limit++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) { 6621 // We print a warning message only once per CMS cycle. 6622 gclog_or_tty->print_cr(" (benign) Hit CMSMarkStack max size limit"); 6623 } 6624 return; 6625 } 6626 // Double capacity if possible 6627 size_t new_capacity = MIN2(_capacity*2, MarkStackSizeMax); 6628 // Do not give up existing stack until we have managed to 6629 // get the double capacity that we desired. 6630 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 6631 new_capacity * sizeof(oop))); 6632 if (rs.is_reserved()) { 6633 // Release the backing store associated with old stack 6634 _virtual_space.release(); 6635 // Reinitialize virtual space for new stack 6636 if (!_virtual_space.initialize(rs, rs.size())) { 6637 fatal("Not enough swap for expanded marking stack"); 6638 } 6639 _base = (oop*)(_virtual_space.low()); 6640 _index = 0; 6641 _capacity = new_capacity; 6642 } else if (_failed_double++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) { 6643 // Failed to double capacity, continue; 6644 // we print a detail message only once per CMS cycle. 6645 gclog_or_tty->print(" (benign) Failed to expand marking stack from "SIZE_FORMAT"K to " 6646 SIZE_FORMAT"K", 6647 _capacity / K, new_capacity / K); 6648 } 6649 } 6650 6651 6652 // Closures 6653 // XXX: there seems to be a lot of code duplication here; 6654 // should refactor and consolidate common code. 6655 6656 // This closure is used to mark refs into the CMS generation in 6657 // the CMS bit map. Called at the first checkpoint. This closure 6658 // assumes that we do not need to re-mark dirty cards; if the CMS 6659 // generation on which this is used is not an oldest 6660 // generation then this will lose younger_gen cards! 6661 6662 MarkRefsIntoClosure::MarkRefsIntoClosure( 6663 MemRegion span, CMSBitMap* bitMap): 6664 _span(span), 6665 _bitMap(bitMap) 6666 { 6667 assert(_ref_processor == NULL, "deliberately left NULL"); 6668 assert(_bitMap->covers(_span), "_bitMap/_span mismatch"); 6669 } 6670 6671 void MarkRefsIntoClosure::do_oop(oop obj) { 6672 // if p points into _span, then mark corresponding bit in _markBitMap 6673 assert(obj->is_oop(), "expected an oop"); 6674 HeapWord* addr = (HeapWord*)obj; 6675 if (_span.contains(addr)) { 6676 // this should be made more efficient 6677 _bitMap->mark(addr); 6678 } 6679 } 6680 6681 void MarkRefsIntoClosure::do_oop(oop* p) { MarkRefsIntoClosure::do_oop_work(p); } 6682 void MarkRefsIntoClosure::do_oop(narrowOop* p) { MarkRefsIntoClosure::do_oop_work(p); } 6683 6684 // A variant of the above, used for CMS marking verification. 6685 MarkRefsIntoVerifyClosure::MarkRefsIntoVerifyClosure( 6686 MemRegion span, CMSBitMap* verification_bm, CMSBitMap* cms_bm): 6687 _span(span), 6688 _verification_bm(verification_bm), 6689 _cms_bm(cms_bm) 6690 { 6691 assert(_ref_processor == NULL, "deliberately left NULL"); 6692 assert(_verification_bm->covers(_span), "_verification_bm/_span mismatch"); 6693 } 6694 6695 void MarkRefsIntoVerifyClosure::do_oop(oop obj) { 6696 // if p points into _span, then mark corresponding bit in _markBitMap 6697 assert(obj->is_oop(), "expected an oop"); 6698 HeapWord* addr = (HeapWord*)obj; 6699 if (_span.contains(addr)) { 6700 _verification_bm->mark(addr); 6701 if (!_cms_bm->isMarked(addr)) { 6702 oop(addr)->print(); 6703 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)", addr); 6704 fatal("... aborting"); 6705 } 6706 } 6707 } 6708 6709 void MarkRefsIntoVerifyClosure::do_oop(oop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); } 6710 void MarkRefsIntoVerifyClosure::do_oop(narrowOop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); } 6711 6712 ////////////////////////////////////////////////// 6713 // MarkRefsIntoAndScanClosure 6714 ////////////////////////////////////////////////// 6715 6716 MarkRefsIntoAndScanClosure::MarkRefsIntoAndScanClosure(MemRegion span, 6717 ReferenceProcessor* rp, 6718 CMSBitMap* bit_map, 6719 CMSBitMap* mod_union_table, 6720 CMSMarkStack* mark_stack, 6721 CMSCollector* collector, 6722 bool should_yield, 6723 bool concurrent_precleaning): 6724 _collector(collector), 6725 _span(span), 6726 _bit_map(bit_map), 6727 _mark_stack(mark_stack), 6728 _pushAndMarkClosure(collector, span, rp, bit_map, mod_union_table, 6729 mark_stack, concurrent_precleaning), 6730 _yield(should_yield), 6731 _concurrent_precleaning(concurrent_precleaning), 6732 _freelistLock(NULL) 6733 { 6734 _ref_processor = rp; 6735 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 6736 } 6737 6738 // This closure is used to mark refs into the CMS generation at the 6739 // second (final) checkpoint, and to scan and transitively follow 6740 // the unmarked oops. It is also used during the concurrent precleaning 6741 // phase while scanning objects on dirty cards in the CMS generation. 6742 // The marks are made in the marking bit map and the marking stack is 6743 // used for keeping the (newly) grey objects during the scan. 6744 // The parallel version (Par_...) appears further below. 6745 void MarkRefsIntoAndScanClosure::do_oop(oop obj) { 6746 if (obj != NULL) { 6747 assert(obj->is_oop(), "expected an oop"); 6748 HeapWord* addr = (HeapWord*)obj; 6749 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 6750 assert(_collector->overflow_list_is_empty(), 6751 "overflow list should be empty"); 6752 if (_span.contains(addr) && 6753 !_bit_map->isMarked(addr)) { 6754 // mark bit map (object is now grey) 6755 _bit_map->mark(addr); 6756 // push on marking stack (stack should be empty), and drain the 6757 // stack by applying this closure to the oops in the oops popped 6758 // from the stack (i.e. blacken the grey objects) 6759 bool res = _mark_stack->push(obj); 6760 assert(res, "Should have space to push on empty stack"); 6761 do { 6762 oop new_oop = _mark_stack->pop(); 6763 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 6764 assert(_bit_map->isMarked((HeapWord*)new_oop), 6765 "only grey objects on this stack"); 6766 // iterate over the oops in this oop, marking and pushing 6767 // the ones in CMS heap (i.e. in _span). 6768 new_oop->oop_iterate(&_pushAndMarkClosure); 6769 // check if it's time to yield 6770 do_yield_check(); 6771 } while (!_mark_stack->isEmpty() || 6772 (!_concurrent_precleaning && take_from_overflow_list())); 6773 // if marking stack is empty, and we are not doing this 6774 // during precleaning, then check the overflow list 6775 } 6776 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 6777 assert(_collector->overflow_list_is_empty(), 6778 "overflow list was drained above"); 6779 // We could restore evacuated mark words, if any, used for 6780 // overflow list links here because the overflow list is 6781 // provably empty here. That would reduce the maximum 6782 // size requirements for preserved_{oop,mark}_stack. 6783 // But we'll just postpone it until we are all done 6784 // so we can just stream through. 6785 if (!_concurrent_precleaning && CMSOverflowEarlyRestoration) { 6786 _collector->restore_preserved_marks_if_any(); 6787 assert(_collector->no_preserved_marks(), "No preserved marks"); 6788 } 6789 assert(!CMSOverflowEarlyRestoration || _collector->no_preserved_marks(), 6790 "All preserved marks should have been restored above"); 6791 } 6792 } 6793 6794 void MarkRefsIntoAndScanClosure::do_oop(oop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); } 6795 void MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); } 6796 6797 void MarkRefsIntoAndScanClosure::do_yield_work() { 6798 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6799 "CMS thread should hold CMS token"); 6800 assert_lock_strong(_freelistLock); 6801 assert_lock_strong(_bit_map->lock()); 6802 // relinquish the free_list_lock and bitMaplock() 6803 _bit_map->lock()->unlock(); 6804 _freelistLock->unlock(); 6805 ConcurrentMarkSweepThread::desynchronize(true); 6806 ConcurrentMarkSweepThread::acknowledge_yield_request(); 6807 _collector->stopTimer(); 6808 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 6809 if (PrintCMSStatistics != 0) { 6810 _collector->incrementYields(); 6811 } 6812 _collector->icms_wait(); 6813 6814 // See the comment in coordinator_yield() 6815 for (unsigned i = 0; 6816 i < CMSYieldSleepCount && 6817 ConcurrentMarkSweepThread::should_yield() && 6818 !CMSCollector::foregroundGCIsActive(); 6819 ++i) { 6820 os::sleep(Thread::current(), 1, false); 6821 ConcurrentMarkSweepThread::acknowledge_yield_request(); 6822 } 6823 6824 ConcurrentMarkSweepThread::synchronize(true); 6825 _freelistLock->lock_without_safepoint_check(); 6826 _bit_map->lock()->lock_without_safepoint_check(); 6827 _collector->startTimer(); 6828 } 6829 6830 /////////////////////////////////////////////////////////// 6831 // Par_MarkRefsIntoAndScanClosure: a parallel version of 6832 // MarkRefsIntoAndScanClosure 6833 /////////////////////////////////////////////////////////// 6834 Par_MarkRefsIntoAndScanClosure::Par_MarkRefsIntoAndScanClosure( 6835 CMSCollector* collector, MemRegion span, ReferenceProcessor* rp, 6836 CMSBitMap* bit_map, OopTaskQueue* work_queue): 6837 _span(span), 6838 _bit_map(bit_map), 6839 _work_queue(work_queue), 6840 _low_water_mark(MIN2((uint)(work_queue->max_elems()/4), 6841 (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads))), 6842 _par_pushAndMarkClosure(collector, span, rp, bit_map, work_queue) 6843 { 6844 _ref_processor = rp; 6845 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 6846 } 6847 6848 // This closure is used to mark refs into the CMS generation at the 6849 // second (final) checkpoint, and to scan and transitively follow 6850 // the unmarked oops. The marks are made in the marking bit map and 6851 // the work_queue is used for keeping the (newly) grey objects during 6852 // the scan phase whence they are also available for stealing by parallel 6853 // threads. Since the marking bit map is shared, updates are 6854 // synchronized (via CAS). 6855 void Par_MarkRefsIntoAndScanClosure::do_oop(oop obj) { 6856 if (obj != NULL) { 6857 // Ignore mark word because this could be an already marked oop 6858 // that may be chained at the end of the overflow list. 6859 assert(obj->is_oop(true), "expected an oop"); 6860 HeapWord* addr = (HeapWord*)obj; 6861 if (_span.contains(addr) && 6862 !_bit_map->isMarked(addr)) { 6863 // mark bit map (object will become grey): 6864 // It is possible for several threads to be 6865 // trying to "claim" this object concurrently; 6866 // the unique thread that succeeds in marking the 6867 // object first will do the subsequent push on 6868 // to the work queue (or overflow list). 6869 if (_bit_map->par_mark(addr)) { 6870 // push on work_queue (which may not be empty), and trim the 6871 // queue to an appropriate length by applying this closure to 6872 // the oops in the oops popped from the stack (i.e. blacken the 6873 // grey objects) 6874 bool res = _work_queue->push(obj); 6875 assert(res, "Low water mark should be less than capacity?"); 6876 trim_queue(_low_water_mark); 6877 } // Else, another thread claimed the object 6878 } 6879 } 6880 } 6881 6882 void Par_MarkRefsIntoAndScanClosure::do_oop(oop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); } 6883 void Par_MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); } 6884 6885 // This closure is used to rescan the marked objects on the dirty cards 6886 // in the mod union table and the card table proper. 6887 size_t ScanMarkedObjectsAgainCarefullyClosure::do_object_careful_m( 6888 oop p, MemRegion mr) { 6889 6890 size_t size = 0; 6891 HeapWord* addr = (HeapWord*)p; 6892 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6893 assert(_span.contains(addr), "we are scanning the CMS generation"); 6894 // check if it's time to yield 6895 if (do_yield_check()) { 6896 // We yielded for some foreground stop-world work, 6897 // and we have been asked to abort this ongoing preclean cycle. 6898 return 0; 6899 } 6900 if (_bitMap->isMarked(addr)) { 6901 // it's marked; is it potentially uninitialized? 6902 if (p->klass_or_null() != NULL) { 6903 // an initialized object; ignore mark word in verification below 6904 // since we are running concurrent with mutators 6905 assert(p->is_oop(true), "should be an oop"); 6906 if (p->is_objArray()) { 6907 // objArrays are precisely marked; restrict scanning 6908 // to dirty cards only. 6909 size = CompactibleFreeListSpace::adjustObjectSize( 6910 p->oop_iterate(_scanningClosure, mr)); 6911 } else { 6912 // A non-array may have been imprecisely marked; we need 6913 // to scan object in its entirety. 6914 size = CompactibleFreeListSpace::adjustObjectSize( 6915 p->oop_iterate(_scanningClosure)); 6916 } 6917 #ifdef ASSERT 6918 size_t direct_size = 6919 CompactibleFreeListSpace::adjustObjectSize(p->size()); 6920 assert(size == direct_size, "Inconsistency in size"); 6921 assert(size >= 3, "Necessary for Printezis marks to work"); 6922 if (!_bitMap->isMarked(addr+1)) { 6923 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size); 6924 } else { 6925 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size-1); 6926 assert(_bitMap->isMarked(addr+size-1), 6927 "inconsistent Printezis mark"); 6928 } 6929 #endif // ASSERT 6930 } else { 6931 // an unitialized object 6932 assert(_bitMap->isMarked(addr+1), "missing Printezis mark?"); 6933 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 6934 size = pointer_delta(nextOneAddr + 1, addr); 6935 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 6936 "alignment problem"); 6937 // Note that pre-cleaning needn't redirty the card. OopDesc::set_klass() 6938 // will dirty the card when the klass pointer is installed in the 6939 // object (signalling the completion of initialization). 6940 } 6941 } else { 6942 // Either a not yet marked object or an uninitialized object 6943 if (p->klass_or_null() == NULL) { 6944 // An uninitialized object, skip to the next card, since 6945 // we may not be able to read its P-bits yet. 6946 assert(size == 0, "Initial value"); 6947 } else { 6948 // An object not (yet) reached by marking: we merely need to 6949 // compute its size so as to go look at the next block. 6950 assert(p->is_oop(true), "should be an oop"); 6951 size = CompactibleFreeListSpace::adjustObjectSize(p->size()); 6952 } 6953 } 6954 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6955 return size; 6956 } 6957 6958 void ScanMarkedObjectsAgainCarefullyClosure::do_yield_work() { 6959 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6960 "CMS thread should hold CMS token"); 6961 assert_lock_strong(_freelistLock); 6962 assert_lock_strong(_bitMap->lock()); 6963 // relinquish the free_list_lock and bitMaplock() 6964 _bitMap->lock()->unlock(); 6965 _freelistLock->unlock(); 6966 ConcurrentMarkSweepThread::desynchronize(true); 6967 ConcurrentMarkSweepThread::acknowledge_yield_request(); 6968 _collector->stopTimer(); 6969 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 6970 if (PrintCMSStatistics != 0) { 6971 _collector->incrementYields(); 6972 } 6973 _collector->icms_wait(); 6974 6975 // See the comment in coordinator_yield() 6976 for (unsigned i = 0; i < CMSYieldSleepCount && 6977 ConcurrentMarkSweepThread::should_yield() && 6978 !CMSCollector::foregroundGCIsActive(); ++i) { 6979 os::sleep(Thread::current(), 1, false); 6980 ConcurrentMarkSweepThread::acknowledge_yield_request(); 6981 } 6982 6983 ConcurrentMarkSweepThread::synchronize(true); 6984 _freelistLock->lock_without_safepoint_check(); 6985 _bitMap->lock()->lock_without_safepoint_check(); 6986 _collector->startTimer(); 6987 } 6988 6989 6990 ////////////////////////////////////////////////////////////////// 6991 // SurvivorSpacePrecleanClosure 6992 ////////////////////////////////////////////////////////////////// 6993 // This (single-threaded) closure is used to preclean the oops in 6994 // the survivor spaces. 6995 size_t SurvivorSpacePrecleanClosure::do_object_careful(oop p) { 6996 6997 HeapWord* addr = (HeapWord*)p; 6998 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6999 assert(!_span.contains(addr), "we are scanning the survivor spaces"); 7000 assert(p->klass_or_null() != NULL, "object should be initializd"); 7001 // an initialized object; ignore mark word in verification below 7002 // since we are running concurrent with mutators 7003 assert(p->is_oop(true), "should be an oop"); 7004 // Note that we do not yield while we iterate over 7005 // the interior oops of p, pushing the relevant ones 7006 // on our marking stack. 7007 size_t size = p->oop_iterate(_scanning_closure); 7008 do_yield_check(); 7009 // Observe that below, we do not abandon the preclean 7010 // phase as soon as we should; rather we empty the 7011 // marking stack before returning. This is to satisfy 7012 // some existing assertions. In general, it may be a 7013 // good idea to abort immediately and complete the marking 7014 // from the grey objects at a later time. 7015 while (!_mark_stack->isEmpty()) { 7016 oop new_oop = _mark_stack->pop(); 7017 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 7018 assert(_bit_map->isMarked((HeapWord*)new_oop), 7019 "only grey objects on this stack"); 7020 // iterate over the oops in this oop, marking and pushing 7021 // the ones in CMS heap (i.e. in _span). 7022 new_oop->oop_iterate(_scanning_closure); 7023 // check if it's time to yield 7024 do_yield_check(); 7025 } 7026 unsigned int after_count = 7027 GenCollectedHeap::heap()->total_collections(); 7028 bool abort = (_before_count != after_count) || 7029 _collector->should_abort_preclean(); 7030 return abort ? 0 : size; 7031 } 7032 7033 void SurvivorSpacePrecleanClosure::do_yield_work() { 7034 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 7035 "CMS thread should hold CMS token"); 7036 assert_lock_strong(_bit_map->lock()); 7037 // Relinquish the bit map lock 7038 _bit_map->lock()->unlock(); 7039 ConcurrentMarkSweepThread::desynchronize(true); 7040 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7041 _collector->stopTimer(); 7042 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 7043 if (PrintCMSStatistics != 0) { 7044 _collector->incrementYields(); 7045 } 7046 _collector->icms_wait(); 7047 7048 // See the comment in coordinator_yield() 7049 for (unsigned i = 0; i < CMSYieldSleepCount && 7050 ConcurrentMarkSweepThread::should_yield() && 7051 !CMSCollector::foregroundGCIsActive(); ++i) { 7052 os::sleep(Thread::current(), 1, false); 7053 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7054 } 7055 7056 ConcurrentMarkSweepThread::synchronize(true); 7057 _bit_map->lock()->lock_without_safepoint_check(); 7058 _collector->startTimer(); 7059 } 7060 7061 // This closure is used to rescan the marked objects on the dirty cards 7062 // in the mod union table and the card table proper. In the parallel 7063 // case, although the bitMap is shared, we do a single read so the 7064 // isMarked() query is "safe". 7065 bool ScanMarkedObjectsAgainClosure::do_object_bm(oop p, MemRegion mr) { 7066 // Ignore mark word because we are running concurrent with mutators 7067 assert(p->is_oop_or_null(true), "expected an oop or null"); 7068 HeapWord* addr = (HeapWord*)p; 7069 assert(_span.contains(addr), "we are scanning the CMS generation"); 7070 bool is_obj_array = false; 7071 #ifdef ASSERT 7072 if (!_parallel) { 7073 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 7074 assert(_collector->overflow_list_is_empty(), 7075 "overflow list should be empty"); 7076 7077 } 7078 #endif // ASSERT 7079 if (_bit_map->isMarked(addr)) { 7080 // Obj arrays are precisely marked, non-arrays are not; 7081 // so we scan objArrays precisely and non-arrays in their 7082 // entirety. 7083 if (p->is_objArray()) { 7084 is_obj_array = true; 7085 if (_parallel) { 7086 p->oop_iterate(_par_scan_closure, mr); 7087 } else { 7088 p->oop_iterate(_scan_closure, mr); 7089 } 7090 } else { 7091 if (_parallel) { 7092 p->oop_iterate(_par_scan_closure); 7093 } else { 7094 p->oop_iterate(_scan_closure); 7095 } 7096 } 7097 } 7098 #ifdef ASSERT 7099 if (!_parallel) { 7100 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 7101 assert(_collector->overflow_list_is_empty(), 7102 "overflow list should be empty"); 7103 7104 } 7105 #endif // ASSERT 7106 return is_obj_array; 7107 } 7108 7109 MarkFromRootsClosure::MarkFromRootsClosure(CMSCollector* collector, 7110 MemRegion span, 7111 CMSBitMap* bitMap, CMSMarkStack* markStack, 7112 bool should_yield, bool verifying): 7113 _collector(collector), 7114 _span(span), 7115 _bitMap(bitMap), 7116 _mut(&collector->_modUnionTable), 7117 _markStack(markStack), 7118 _yield(should_yield), 7119 _skipBits(0) 7120 { 7121 assert(_markStack->isEmpty(), "stack should be empty"); 7122 _finger = _bitMap->startWord(); 7123 _threshold = _finger; 7124 assert(_collector->_restart_addr == NULL, "Sanity check"); 7125 assert(_span.contains(_finger), "Out of bounds _finger?"); 7126 DEBUG_ONLY(_verifying = verifying;) 7127 } 7128 7129 void MarkFromRootsClosure::reset(HeapWord* addr) { 7130 assert(_markStack->isEmpty(), "would cause duplicates on stack"); 7131 assert(_span.contains(addr), "Out of bounds _finger?"); 7132 _finger = addr; 7133 _threshold = (HeapWord*)round_to( 7134 (intptr_t)_finger, CardTableModRefBS::card_size); 7135 } 7136 7137 // Should revisit to see if this should be restructured for 7138 // greater efficiency. 7139 bool MarkFromRootsClosure::do_bit(size_t offset) { 7140 if (_skipBits > 0) { 7141 _skipBits--; 7142 return true; 7143 } 7144 // convert offset into a HeapWord* 7145 HeapWord* addr = _bitMap->startWord() + offset; 7146 assert(_bitMap->endWord() && addr < _bitMap->endWord(), 7147 "address out of range"); 7148 assert(_bitMap->isMarked(addr), "tautology"); 7149 if (_bitMap->isMarked(addr+1)) { 7150 // this is an allocated but not yet initialized object 7151 assert(_skipBits == 0, "tautology"); 7152 _skipBits = 2; // skip next two marked bits ("Printezis-marks") 7153 oop p = oop(addr); 7154 if (p->klass_or_null() == NULL) { 7155 DEBUG_ONLY(if (!_verifying) {) 7156 // We re-dirty the cards on which this object lies and increase 7157 // the _threshold so that we'll come back to scan this object 7158 // during the preclean or remark phase. (CMSCleanOnEnter) 7159 if (CMSCleanOnEnter) { 7160 size_t sz = _collector->block_size_using_printezis_bits(addr); 7161 HeapWord* end_card_addr = (HeapWord*)round_to( 7162 (intptr_t)(addr+sz), CardTableModRefBS::card_size); 7163 MemRegion redirty_range = MemRegion(addr, end_card_addr); 7164 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 7165 // Bump _threshold to end_card_addr; note that 7166 // _threshold cannot possibly exceed end_card_addr, anyhow. 7167 // This prevents future clearing of the card as the scan proceeds 7168 // to the right. 7169 assert(_threshold <= end_card_addr, 7170 "Because we are just scanning into this object"); 7171 if (_threshold < end_card_addr) { 7172 _threshold = end_card_addr; 7173 } 7174 if (p->klass_or_null() != NULL) { 7175 // Redirty the range of cards... 7176 _mut->mark_range(redirty_range); 7177 } // ...else the setting of klass will dirty the card anyway. 7178 } 7179 DEBUG_ONLY(}) 7180 return true; 7181 } 7182 } 7183 scanOopsInOop(addr); 7184 return true; 7185 } 7186 7187 // We take a break if we've been at this for a while, 7188 // so as to avoid monopolizing the locks involved. 7189 void MarkFromRootsClosure::do_yield_work() { 7190 // First give up the locks, then yield, then re-lock 7191 // We should probably use a constructor/destructor idiom to 7192 // do this unlock/lock or modify the MutexUnlocker class to 7193 // serve our purpose. XXX 7194 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 7195 "CMS thread should hold CMS token"); 7196 assert_lock_strong(_bitMap->lock()); 7197 _bitMap->lock()->unlock(); 7198 ConcurrentMarkSweepThread::desynchronize(true); 7199 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7200 _collector->stopTimer(); 7201 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 7202 if (PrintCMSStatistics != 0) { 7203 _collector->incrementYields(); 7204 } 7205 _collector->icms_wait(); 7206 7207 // See the comment in coordinator_yield() 7208 for (unsigned i = 0; i < CMSYieldSleepCount && 7209 ConcurrentMarkSweepThread::should_yield() && 7210 !CMSCollector::foregroundGCIsActive(); ++i) { 7211 os::sleep(Thread::current(), 1, false); 7212 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7213 } 7214 7215 ConcurrentMarkSweepThread::synchronize(true); 7216 _bitMap->lock()->lock_without_safepoint_check(); 7217 _collector->startTimer(); 7218 } 7219 7220 void MarkFromRootsClosure::scanOopsInOop(HeapWord* ptr) { 7221 assert(_bitMap->isMarked(ptr), "expected bit to be set"); 7222 assert(_markStack->isEmpty(), 7223 "should drain stack to limit stack usage"); 7224 // convert ptr to an oop preparatory to scanning 7225 oop obj = oop(ptr); 7226 // Ignore mark word in verification below, since we 7227 // may be running concurrent with mutators. 7228 assert(obj->is_oop(true), "should be an oop"); 7229 assert(_finger <= ptr, "_finger runneth ahead"); 7230 // advance the finger to right end of this object 7231 _finger = ptr + obj->size(); 7232 assert(_finger > ptr, "we just incremented it above"); 7233 // On large heaps, it may take us some time to get through 7234 // the marking phase (especially if running iCMS). During 7235 // this time it's possible that a lot of mutations have 7236 // accumulated in the card table and the mod union table -- 7237 // these mutation records are redundant until we have 7238 // actually traced into the corresponding card. 7239 // Here, we check whether advancing the finger would make 7240 // us cross into a new card, and if so clear corresponding 7241 // cards in the MUT (preclean them in the card-table in the 7242 // future). 7243 7244 DEBUG_ONLY(if (!_verifying) {) 7245 // The clean-on-enter optimization is disabled by default, 7246 // until we fix 6178663. 7247 if (CMSCleanOnEnter && (_finger > _threshold)) { 7248 // [_threshold, _finger) represents the interval 7249 // of cards to be cleared in MUT (or precleaned in card table). 7250 // The set of cards to be cleared is all those that overlap 7251 // with the interval [_threshold, _finger); note that 7252 // _threshold is always kept card-aligned but _finger isn't 7253 // always card-aligned. 7254 HeapWord* old_threshold = _threshold; 7255 assert(old_threshold == (HeapWord*)round_to( 7256 (intptr_t)old_threshold, CardTableModRefBS::card_size), 7257 "_threshold should always be card-aligned"); 7258 _threshold = (HeapWord*)round_to( 7259 (intptr_t)_finger, CardTableModRefBS::card_size); 7260 MemRegion mr(old_threshold, _threshold); 7261 assert(!mr.is_empty(), "Control point invariant"); 7262 assert(_span.contains(mr), "Should clear within span"); 7263 _mut->clear_range(mr); 7264 } 7265 DEBUG_ONLY(}) 7266 // Note: the finger doesn't advance while we drain 7267 // the stack below. 7268 PushOrMarkClosure pushOrMarkClosure(_collector, 7269 _span, _bitMap, _markStack, 7270 _finger, this); 7271 bool res = _markStack->push(obj); 7272 assert(res, "Empty non-zero size stack should have space for single push"); 7273 while (!_markStack->isEmpty()) { 7274 oop new_oop = _markStack->pop(); 7275 // Skip verifying header mark word below because we are 7276 // running concurrent with mutators. 7277 assert(new_oop->is_oop(true), "Oops! expected to pop an oop"); 7278 // now scan this oop's oops 7279 new_oop->oop_iterate(&pushOrMarkClosure); 7280 do_yield_check(); 7281 } 7282 assert(_markStack->isEmpty(), "tautology, emphasizing post-condition"); 7283 } 7284 7285 Par_MarkFromRootsClosure::Par_MarkFromRootsClosure(CMSConcMarkingTask* task, 7286 CMSCollector* collector, MemRegion span, 7287 CMSBitMap* bit_map, 7288 OopTaskQueue* work_queue, 7289 CMSMarkStack* overflow_stack, 7290 bool should_yield): 7291 _collector(collector), 7292 _whole_span(collector->_span), 7293 _span(span), 7294 _bit_map(bit_map), 7295 _mut(&collector->_modUnionTable), 7296 _work_queue(work_queue), 7297 _overflow_stack(overflow_stack), 7298 _yield(should_yield), 7299 _skip_bits(0), 7300 _task(task) 7301 { 7302 assert(_work_queue->size() == 0, "work_queue should be empty"); 7303 _finger = span.start(); 7304 _threshold = _finger; // XXX Defer clear-on-enter optimization for now 7305 assert(_span.contains(_finger), "Out of bounds _finger?"); 7306 } 7307 7308 // Should revisit to see if this should be restructured for 7309 // greater efficiency. 7310 bool Par_MarkFromRootsClosure::do_bit(size_t offset) { 7311 if (_skip_bits > 0) { 7312 _skip_bits--; 7313 return true; 7314 } 7315 // convert offset into a HeapWord* 7316 HeapWord* addr = _bit_map->startWord() + offset; 7317 assert(_bit_map->endWord() && addr < _bit_map->endWord(), 7318 "address out of range"); 7319 assert(_bit_map->isMarked(addr), "tautology"); 7320 if (_bit_map->isMarked(addr+1)) { 7321 // this is an allocated object that might not yet be initialized 7322 assert(_skip_bits == 0, "tautology"); 7323 _skip_bits = 2; // skip next two marked bits ("Printezis-marks") 7324 oop p = oop(addr); 7325 if (p->klass_or_null() == NULL) { 7326 // in the case of Clean-on-Enter optimization, redirty card 7327 // and avoid clearing card by increasing the threshold. 7328 return true; 7329 } 7330 } 7331 scan_oops_in_oop(addr); 7332 return true; 7333 } 7334 7335 void Par_MarkFromRootsClosure::scan_oops_in_oop(HeapWord* ptr) { 7336 assert(_bit_map->isMarked(ptr), "expected bit to be set"); 7337 // Should we assert that our work queue is empty or 7338 // below some drain limit? 7339 assert(_work_queue->size() == 0, 7340 "should drain stack to limit stack usage"); 7341 // convert ptr to an oop preparatory to scanning 7342 oop obj = oop(ptr); 7343 // Ignore mark word in verification below, since we 7344 // may be running concurrent with mutators. 7345 assert(obj->is_oop(true), "should be an oop"); 7346 assert(_finger <= ptr, "_finger runneth ahead"); 7347 // advance the finger to right end of this object 7348 _finger = ptr + obj->size(); 7349 assert(_finger > ptr, "we just incremented it above"); 7350 // On large heaps, it may take us some time to get through 7351 // the marking phase (especially if running iCMS). During 7352 // this time it's possible that a lot of mutations have 7353 // accumulated in the card table and the mod union table -- 7354 // these mutation records are redundant until we have 7355 // actually traced into the corresponding card. 7356 // Here, we check whether advancing the finger would make 7357 // us cross into a new card, and if so clear corresponding 7358 // cards in the MUT (preclean them in the card-table in the 7359 // future). 7360 7361 // The clean-on-enter optimization is disabled by default, 7362 // until we fix 6178663. 7363 if (CMSCleanOnEnter && (_finger > _threshold)) { 7364 // [_threshold, _finger) represents the interval 7365 // of cards to be cleared in MUT (or precleaned in card table). 7366 // The set of cards to be cleared is all those that overlap 7367 // with the interval [_threshold, _finger); note that 7368 // _threshold is always kept card-aligned but _finger isn't 7369 // always card-aligned. 7370 HeapWord* old_threshold = _threshold; 7371 assert(old_threshold == (HeapWord*)round_to( 7372 (intptr_t)old_threshold, CardTableModRefBS::card_size), 7373 "_threshold should always be card-aligned"); 7374 _threshold = (HeapWord*)round_to( 7375 (intptr_t)_finger, CardTableModRefBS::card_size); 7376 MemRegion mr(old_threshold, _threshold); 7377 assert(!mr.is_empty(), "Control point invariant"); 7378 assert(_span.contains(mr), "Should clear within span"); // _whole_span ?? 7379 _mut->clear_range(mr); 7380 } 7381 7382 // Note: the local finger doesn't advance while we drain 7383 // the stack below, but the global finger sure can and will. 7384 HeapWord** gfa = _task->global_finger_addr(); 7385 Par_PushOrMarkClosure pushOrMarkClosure(_collector, 7386 _span, _bit_map, 7387 _work_queue, 7388 _overflow_stack, 7389 _finger, 7390 gfa, this); 7391 bool res = _work_queue->push(obj); // overflow could occur here 7392 assert(res, "Will hold once we use workqueues"); 7393 while (true) { 7394 oop new_oop; 7395 if (!_work_queue->pop_local(new_oop)) { 7396 // We emptied our work_queue; check if there's stuff that can 7397 // be gotten from the overflow stack. 7398 if (CMSConcMarkingTask::get_work_from_overflow_stack( 7399 _overflow_stack, _work_queue)) { 7400 do_yield_check(); 7401 continue; 7402 } else { // done 7403 break; 7404 } 7405 } 7406 // Skip verifying header mark word below because we are 7407 // running concurrent with mutators. 7408 assert(new_oop->is_oop(true), "Oops! expected to pop an oop"); 7409 // now scan this oop's oops 7410 new_oop->oop_iterate(&pushOrMarkClosure); 7411 do_yield_check(); 7412 } 7413 assert(_work_queue->size() == 0, "tautology, emphasizing post-condition"); 7414 } 7415 7416 // Yield in response to a request from VM Thread or 7417 // from mutators. 7418 void Par_MarkFromRootsClosure::do_yield_work() { 7419 assert(_task != NULL, "sanity"); 7420 _task->yield(); 7421 } 7422 7423 // A variant of the above used for verifying CMS marking work. 7424 MarkFromRootsVerifyClosure::MarkFromRootsVerifyClosure(CMSCollector* collector, 7425 MemRegion span, 7426 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 7427 CMSMarkStack* mark_stack): 7428 _collector(collector), 7429 _span(span), 7430 _verification_bm(verification_bm), 7431 _cms_bm(cms_bm), 7432 _mark_stack(mark_stack), 7433 _pam_verify_closure(collector, span, verification_bm, cms_bm, 7434 mark_stack) 7435 { 7436 assert(_mark_stack->isEmpty(), "stack should be empty"); 7437 _finger = _verification_bm->startWord(); 7438 assert(_collector->_restart_addr == NULL, "Sanity check"); 7439 assert(_span.contains(_finger), "Out of bounds _finger?"); 7440 } 7441 7442 void MarkFromRootsVerifyClosure::reset(HeapWord* addr) { 7443 assert(_mark_stack->isEmpty(), "would cause duplicates on stack"); 7444 assert(_span.contains(addr), "Out of bounds _finger?"); 7445 _finger = addr; 7446 } 7447 7448 // Should revisit to see if this should be restructured for 7449 // greater efficiency. 7450 bool MarkFromRootsVerifyClosure::do_bit(size_t offset) { 7451 // convert offset into a HeapWord* 7452 HeapWord* addr = _verification_bm->startWord() + offset; 7453 assert(_verification_bm->endWord() && addr < _verification_bm->endWord(), 7454 "address out of range"); 7455 assert(_verification_bm->isMarked(addr), "tautology"); 7456 assert(_cms_bm->isMarked(addr), "tautology"); 7457 7458 assert(_mark_stack->isEmpty(), 7459 "should drain stack to limit stack usage"); 7460 // convert addr to an oop preparatory to scanning 7461 oop obj = oop(addr); 7462 assert(obj->is_oop(), "should be an oop"); 7463 assert(_finger <= addr, "_finger runneth ahead"); 7464 // advance the finger to right end of this object 7465 _finger = addr + obj->size(); 7466 assert(_finger > addr, "we just incremented it above"); 7467 // Note: the finger doesn't advance while we drain 7468 // the stack below. 7469 bool res = _mark_stack->push(obj); 7470 assert(res, "Empty non-zero size stack should have space for single push"); 7471 while (!_mark_stack->isEmpty()) { 7472 oop new_oop = _mark_stack->pop(); 7473 assert(new_oop->is_oop(), "Oops! expected to pop an oop"); 7474 // now scan this oop's oops 7475 new_oop->oop_iterate(&_pam_verify_closure); 7476 } 7477 assert(_mark_stack->isEmpty(), "tautology, emphasizing post-condition"); 7478 return true; 7479 } 7480 7481 PushAndMarkVerifyClosure::PushAndMarkVerifyClosure( 7482 CMSCollector* collector, MemRegion span, 7483 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 7484 CMSMarkStack* mark_stack): 7485 CMSOopClosure(collector->ref_processor()), 7486 _collector(collector), 7487 _span(span), 7488 _verification_bm(verification_bm), 7489 _cms_bm(cms_bm), 7490 _mark_stack(mark_stack) 7491 { } 7492 7493 void PushAndMarkVerifyClosure::do_oop(oop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 7494 void PushAndMarkVerifyClosure::do_oop(narrowOop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 7495 7496 // Upon stack overflow, we discard (part of) the stack, 7497 // remembering the least address amongst those discarded 7498 // in CMSCollector's _restart_address. 7499 void PushAndMarkVerifyClosure::handle_stack_overflow(HeapWord* lost) { 7500 // Remember the least grey address discarded 7501 HeapWord* ra = (HeapWord*)_mark_stack->least_value(lost); 7502 _collector->lower_restart_addr(ra); 7503 _mark_stack->reset(); // discard stack contents 7504 _mark_stack->expand(); // expand the stack if possible 7505 } 7506 7507 void PushAndMarkVerifyClosure::do_oop(oop obj) { 7508 assert(obj->is_oop_or_null(), "expected an oop or NULL"); 7509 HeapWord* addr = (HeapWord*)obj; 7510 if (_span.contains(addr) && !_verification_bm->isMarked(addr)) { 7511 // Oop lies in _span and isn't yet grey or black 7512 _verification_bm->mark(addr); // now grey 7513 if (!_cms_bm->isMarked(addr)) { 7514 oop(addr)->print(); 7515 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)", 7516 addr); 7517 fatal("... aborting"); 7518 } 7519 7520 if (!_mark_stack->push(obj)) { // stack overflow 7521 if (PrintCMSStatistics != 0) { 7522 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 7523 SIZE_FORMAT, _mark_stack->capacity()); 7524 } 7525 assert(_mark_stack->isFull(), "Else push should have succeeded"); 7526 handle_stack_overflow(addr); 7527 } 7528 // anything including and to the right of _finger 7529 // will be scanned as we iterate over the remainder of the 7530 // bit map 7531 } 7532 } 7533 7534 PushOrMarkClosure::PushOrMarkClosure(CMSCollector* collector, 7535 MemRegion span, 7536 CMSBitMap* bitMap, CMSMarkStack* markStack, 7537 HeapWord* finger, MarkFromRootsClosure* parent) : 7538 CMSOopClosure(collector->ref_processor()), 7539 _collector(collector), 7540 _span(span), 7541 _bitMap(bitMap), 7542 _markStack(markStack), 7543 _finger(finger), 7544 _parent(parent) 7545 { } 7546 7547 Par_PushOrMarkClosure::Par_PushOrMarkClosure(CMSCollector* collector, 7548 MemRegion span, 7549 CMSBitMap* bit_map, 7550 OopTaskQueue* work_queue, 7551 CMSMarkStack* overflow_stack, 7552 HeapWord* finger, 7553 HeapWord** global_finger_addr, 7554 Par_MarkFromRootsClosure* parent) : 7555 CMSOopClosure(collector->ref_processor()), 7556 _collector(collector), 7557 _whole_span(collector->_span), 7558 _span(span), 7559 _bit_map(bit_map), 7560 _work_queue(work_queue), 7561 _overflow_stack(overflow_stack), 7562 _finger(finger), 7563 _global_finger_addr(global_finger_addr), 7564 _parent(parent) 7565 { } 7566 7567 // Assumes thread-safe access by callers, who are 7568 // responsible for mutual exclusion. 7569 void CMSCollector::lower_restart_addr(HeapWord* low) { 7570 assert(_span.contains(low), "Out of bounds addr"); 7571 if (_restart_addr == NULL) { 7572 _restart_addr = low; 7573 } else { 7574 _restart_addr = MIN2(_restart_addr, low); 7575 } 7576 } 7577 7578 // Upon stack overflow, we discard (part of) the stack, 7579 // remembering the least address amongst those discarded 7580 // in CMSCollector's _restart_address. 7581 void PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 7582 // Remember the least grey address discarded 7583 HeapWord* ra = (HeapWord*)_markStack->least_value(lost); 7584 _collector->lower_restart_addr(ra); 7585 _markStack->reset(); // discard stack contents 7586 _markStack->expand(); // expand the stack if possible 7587 } 7588 7589 // Upon stack overflow, we discard (part of) the stack, 7590 // remembering the least address amongst those discarded 7591 // in CMSCollector's _restart_address. 7592 void Par_PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 7593 // We need to do this under a mutex to prevent other 7594 // workers from interfering with the work done below. 7595 MutexLockerEx ml(_overflow_stack->par_lock(), 7596 Mutex::_no_safepoint_check_flag); 7597 // Remember the least grey address discarded 7598 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 7599 _collector->lower_restart_addr(ra); 7600 _overflow_stack->reset(); // discard stack contents 7601 _overflow_stack->expand(); // expand the stack if possible 7602 } 7603 7604 void CMKlassClosure::do_klass(Klass* k) { 7605 assert(_oop_closure != NULL, "Not initialized?"); 7606 k->oops_do(_oop_closure); 7607 } 7608 7609 void PushOrMarkClosure::do_oop(oop obj) { 7610 // Ignore mark word because we are running concurrent with mutators. 7611 assert(obj->is_oop_or_null(true), "expected an oop or NULL"); 7612 HeapWord* addr = (HeapWord*)obj; 7613 if (_span.contains(addr) && !_bitMap->isMarked(addr)) { 7614 // Oop lies in _span and isn't yet grey or black 7615 _bitMap->mark(addr); // now grey 7616 if (addr < _finger) { 7617 // the bit map iteration has already either passed, or 7618 // sampled, this bit in the bit map; we'll need to 7619 // use the marking stack to scan this oop's oops. 7620 bool simulate_overflow = false; 7621 NOT_PRODUCT( 7622 if (CMSMarkStackOverflowALot && 7623 _collector->simulate_overflow()) { 7624 // simulate a stack overflow 7625 simulate_overflow = true; 7626 } 7627 ) 7628 if (simulate_overflow || !_markStack->push(obj)) { // stack overflow 7629 if (PrintCMSStatistics != 0) { 7630 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 7631 SIZE_FORMAT, _markStack->capacity()); 7632 } 7633 assert(simulate_overflow || _markStack->isFull(), "Else push should have succeeded"); 7634 handle_stack_overflow(addr); 7635 } 7636 } 7637 // anything including and to the right of _finger 7638 // will be scanned as we iterate over the remainder of the 7639 // bit map 7640 do_yield_check(); 7641 } 7642 } 7643 7644 void PushOrMarkClosure::do_oop(oop* p) { PushOrMarkClosure::do_oop_work(p); } 7645 void PushOrMarkClosure::do_oop(narrowOop* p) { PushOrMarkClosure::do_oop_work(p); } 7646 7647 void Par_PushOrMarkClosure::do_oop(oop obj) { 7648 // Ignore mark word because we are running concurrent with mutators. 7649 assert(obj->is_oop_or_null(true), "expected an oop or NULL"); 7650 HeapWord* addr = (HeapWord*)obj; 7651 if (_whole_span.contains(addr) && !_bit_map->isMarked(addr)) { 7652 // Oop lies in _span and isn't yet grey or black 7653 // We read the global_finger (volatile read) strictly after marking oop 7654 bool res = _bit_map->par_mark(addr); // now grey 7655 volatile HeapWord** gfa = (volatile HeapWord**)_global_finger_addr; 7656 // Should we push this marked oop on our stack? 7657 // -- if someone else marked it, nothing to do 7658 // -- if target oop is above global finger nothing to do 7659 // -- if target oop is in chunk and above local finger 7660 // then nothing to do 7661 // -- else push on work queue 7662 if ( !res // someone else marked it, they will deal with it 7663 || (addr >= *gfa) // will be scanned in a later task 7664 || (_span.contains(addr) && addr >= _finger)) { // later in this chunk 7665 return; 7666 } 7667 // the bit map iteration has already either passed, or 7668 // sampled, this bit in the bit map; we'll need to 7669 // use the marking stack to scan this oop's oops. 7670 bool simulate_overflow = false; 7671 NOT_PRODUCT( 7672 if (CMSMarkStackOverflowALot && 7673 _collector->simulate_overflow()) { 7674 // simulate a stack overflow 7675 simulate_overflow = true; 7676 } 7677 ) 7678 if (simulate_overflow || 7679 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) { 7680 // stack overflow 7681 if (PrintCMSStatistics != 0) { 7682 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 7683 SIZE_FORMAT, _overflow_stack->capacity()); 7684 } 7685 // We cannot assert that the overflow stack is full because 7686 // it may have been emptied since. 7687 assert(simulate_overflow || 7688 _work_queue->size() == _work_queue->max_elems(), 7689 "Else push should have succeeded"); 7690 handle_stack_overflow(addr); 7691 } 7692 do_yield_check(); 7693 } 7694 } 7695 7696 void Par_PushOrMarkClosure::do_oop(oop* p) { Par_PushOrMarkClosure::do_oop_work(p); } 7697 void Par_PushOrMarkClosure::do_oop(narrowOop* p) { Par_PushOrMarkClosure::do_oop_work(p); } 7698 7699 PushAndMarkClosure::PushAndMarkClosure(CMSCollector* collector, 7700 MemRegion span, 7701 ReferenceProcessor* rp, 7702 CMSBitMap* bit_map, 7703 CMSBitMap* mod_union_table, 7704 CMSMarkStack* mark_stack, 7705 bool concurrent_precleaning): 7706 CMSOopClosure(rp), 7707 _collector(collector), 7708 _span(span), 7709 _bit_map(bit_map), 7710 _mod_union_table(mod_union_table), 7711 _mark_stack(mark_stack), 7712 _concurrent_precleaning(concurrent_precleaning) 7713 { 7714 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 7715 } 7716 7717 // Grey object rescan during pre-cleaning and second checkpoint phases -- 7718 // the non-parallel version (the parallel version appears further below.) 7719 void PushAndMarkClosure::do_oop(oop obj) { 7720 // Ignore mark word verification. If during concurrent precleaning, 7721 // the object monitor may be locked. If during the checkpoint 7722 // phases, the object may already have been reached by a different 7723 // path and may be at the end of the global overflow list (so 7724 // the mark word may be NULL). 7725 assert(obj->is_oop_or_null(true /* ignore mark word */), 7726 "expected an oop or NULL"); 7727 HeapWord* addr = (HeapWord*)obj; 7728 // Check if oop points into the CMS generation 7729 // and is not marked 7730 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 7731 // a white object ... 7732 _bit_map->mark(addr); // ... now grey 7733 // push on the marking stack (grey set) 7734 bool simulate_overflow = false; 7735 NOT_PRODUCT( 7736 if (CMSMarkStackOverflowALot && 7737 _collector->simulate_overflow()) { 7738 // simulate a stack overflow 7739 simulate_overflow = true; 7740 } 7741 ) 7742 if (simulate_overflow || !_mark_stack->push(obj)) { 7743 if (_concurrent_precleaning) { 7744 // During precleaning we can just dirty the appropriate card(s) 7745 // in the mod union table, thus ensuring that the object remains 7746 // in the grey set and continue. In the case of object arrays 7747 // we need to dirty all of the cards that the object spans, 7748 // since the rescan of object arrays will be limited to the 7749 // dirty cards. 7750 // Note that no one can be intefering with us in this action 7751 // of dirtying the mod union table, so no locking or atomics 7752 // are required. 7753 if (obj->is_objArray()) { 7754 size_t sz = obj->size(); 7755 HeapWord* end_card_addr = (HeapWord*)round_to( 7756 (intptr_t)(addr+sz), CardTableModRefBS::card_size); 7757 MemRegion redirty_range = MemRegion(addr, end_card_addr); 7758 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 7759 _mod_union_table->mark_range(redirty_range); 7760 } else { 7761 _mod_union_table->mark(addr); 7762 } 7763 _collector->_ser_pmc_preclean_ovflw++; 7764 } else { 7765 // During the remark phase, we need to remember this oop 7766 // in the overflow list. 7767 _collector->push_on_overflow_list(obj); 7768 _collector->_ser_pmc_remark_ovflw++; 7769 } 7770 } 7771 } 7772 } 7773 7774 Par_PushAndMarkClosure::Par_PushAndMarkClosure(CMSCollector* collector, 7775 MemRegion span, 7776 ReferenceProcessor* rp, 7777 CMSBitMap* bit_map, 7778 OopTaskQueue* work_queue): 7779 CMSOopClosure(rp), 7780 _collector(collector), 7781 _span(span), 7782 _bit_map(bit_map), 7783 _work_queue(work_queue) 7784 { 7785 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 7786 } 7787 7788 void PushAndMarkClosure::do_oop(oop* p) { PushAndMarkClosure::do_oop_work(p); } 7789 void PushAndMarkClosure::do_oop(narrowOop* p) { PushAndMarkClosure::do_oop_work(p); } 7790 7791 // Grey object rescan during second checkpoint phase -- 7792 // the parallel version. 7793 void Par_PushAndMarkClosure::do_oop(oop obj) { 7794 // In the assert below, we ignore the mark word because 7795 // this oop may point to an already visited object that is 7796 // on the overflow stack (in which case the mark word has 7797 // been hijacked for chaining into the overflow stack -- 7798 // if this is the last object in the overflow stack then 7799 // its mark word will be NULL). Because this object may 7800 // have been subsequently popped off the global overflow 7801 // stack, and the mark word possibly restored to the prototypical 7802 // value, by the time we get to examined this failing assert in 7803 // the debugger, is_oop_or_null(false) may subsequently start 7804 // to hold. 7805 assert(obj->is_oop_or_null(true), 7806 "expected an oop or NULL"); 7807 HeapWord* addr = (HeapWord*)obj; 7808 // Check if oop points into the CMS generation 7809 // and is not marked 7810 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 7811 // a white object ... 7812 // If we manage to "claim" the object, by being the 7813 // first thread to mark it, then we push it on our 7814 // marking stack 7815 if (_bit_map->par_mark(addr)) { // ... now grey 7816 // push on work queue (grey set) 7817 bool simulate_overflow = false; 7818 NOT_PRODUCT( 7819 if (CMSMarkStackOverflowALot && 7820 _collector->par_simulate_overflow()) { 7821 // simulate a stack overflow 7822 simulate_overflow = true; 7823 } 7824 ) 7825 if (simulate_overflow || !_work_queue->push(obj)) { 7826 _collector->par_push_on_overflow_list(obj); 7827 _collector->_par_pmc_remark_ovflw++; // imprecise OK: no need to CAS 7828 } 7829 } // Else, some other thread got there first 7830 } 7831 } 7832 7833 void Par_PushAndMarkClosure::do_oop(oop* p) { Par_PushAndMarkClosure::do_oop_work(p); } 7834 void Par_PushAndMarkClosure::do_oop(narrowOop* p) { Par_PushAndMarkClosure::do_oop_work(p); } 7835 7836 void CMSPrecleanRefsYieldClosure::do_yield_work() { 7837 Mutex* bml = _collector->bitMapLock(); 7838 assert_lock_strong(bml); 7839 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 7840 "CMS thread should hold CMS token"); 7841 7842 bml->unlock(); 7843 ConcurrentMarkSweepThread::desynchronize(true); 7844 7845 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7846 7847 _collector->stopTimer(); 7848 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 7849 if (PrintCMSStatistics != 0) { 7850 _collector->incrementYields(); 7851 } 7852 _collector->icms_wait(); 7853 7854 // See the comment in coordinator_yield() 7855 for (unsigned i = 0; i < CMSYieldSleepCount && 7856 ConcurrentMarkSweepThread::should_yield() && 7857 !CMSCollector::foregroundGCIsActive(); ++i) { 7858 os::sleep(Thread::current(), 1, false); 7859 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7860 } 7861 7862 ConcurrentMarkSweepThread::synchronize(true); 7863 bml->lock(); 7864 7865 _collector->startTimer(); 7866 } 7867 7868 bool CMSPrecleanRefsYieldClosure::should_return() { 7869 if (ConcurrentMarkSweepThread::should_yield()) { 7870 do_yield_work(); 7871 } 7872 return _collector->foregroundGCIsActive(); 7873 } 7874 7875 void MarkFromDirtyCardsClosure::do_MemRegion(MemRegion mr) { 7876 assert(((size_t)mr.start())%CardTableModRefBS::card_size_in_words == 0, 7877 "mr should be aligned to start at a card boundary"); 7878 // We'd like to assert: 7879 // assert(mr.word_size()%CardTableModRefBS::card_size_in_words == 0, 7880 // "mr should be a range of cards"); 7881 // However, that would be too strong in one case -- the last 7882 // partition ends at _unallocated_block which, in general, can be 7883 // an arbitrary boundary, not necessarily card aligned. 7884 if (PrintCMSStatistics != 0) { 7885 _num_dirty_cards += 7886 mr.word_size()/CardTableModRefBS::card_size_in_words; 7887 } 7888 _space->object_iterate_mem(mr, &_scan_cl); 7889 } 7890 7891 SweepClosure::SweepClosure(CMSCollector* collector, 7892 ConcurrentMarkSweepGeneration* g, 7893 CMSBitMap* bitMap, bool should_yield) : 7894 _collector(collector), 7895 _g(g), 7896 _sp(g->cmsSpace()), 7897 _limit(_sp->sweep_limit()), 7898 _freelistLock(_sp->freelistLock()), 7899 _bitMap(bitMap), 7900 _yield(should_yield), 7901 _inFreeRange(false), // No free range at beginning of sweep 7902 _freeRangeInFreeLists(false), // No free range at beginning of sweep 7903 _lastFreeRangeCoalesced(false), 7904 _freeFinger(g->used_region().start()) 7905 { 7906 NOT_PRODUCT( 7907 _numObjectsFreed = 0; 7908 _numWordsFreed = 0; 7909 _numObjectsLive = 0; 7910 _numWordsLive = 0; 7911 _numObjectsAlreadyFree = 0; 7912 _numWordsAlreadyFree = 0; 7913 _last_fc = NULL; 7914 7915 _sp->initializeIndexedFreeListArrayReturnedBytes(); 7916 _sp->dictionary()->initialize_dict_returned_bytes(); 7917 ) 7918 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7919 "sweep _limit out of bounds"); 7920 if (CMSTraceSweeper) { 7921 gclog_or_tty->print_cr("\n====================\nStarting new sweep with limit " PTR_FORMAT, 7922 _limit); 7923 } 7924 } 7925 7926 void SweepClosure::print_on(outputStream* st) const { 7927 tty->print_cr("_sp = [" PTR_FORMAT "," PTR_FORMAT ")", 7928 _sp->bottom(), _sp->end()); 7929 tty->print_cr("_limit = " PTR_FORMAT, _limit); 7930 tty->print_cr("_freeFinger = " PTR_FORMAT, _freeFinger); 7931 NOT_PRODUCT(tty->print_cr("_last_fc = " PTR_FORMAT, _last_fc);) 7932 tty->print_cr("_inFreeRange = %d, _freeRangeInFreeLists = %d, _lastFreeRangeCoalesced = %d", 7933 _inFreeRange, _freeRangeInFreeLists, _lastFreeRangeCoalesced); 7934 } 7935 7936 #ifndef PRODUCT 7937 // Assertion checking only: no useful work in product mode -- 7938 // however, if any of the flags below become product flags, 7939 // you may need to review this code to see if it needs to be 7940 // enabled in product mode. 7941 SweepClosure::~SweepClosure() { 7942 assert_lock_strong(_freelistLock); 7943 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7944 "sweep _limit out of bounds"); 7945 if (inFreeRange()) { 7946 warning("inFreeRange() should have been reset; dumping state of SweepClosure"); 7947 print(); 7948 ShouldNotReachHere(); 7949 } 7950 if (Verbose && PrintGC) { 7951 gclog_or_tty->print("Collected "SIZE_FORMAT" objects, " SIZE_FORMAT " bytes", 7952 _numObjectsFreed, _numWordsFreed*sizeof(HeapWord)); 7953 gclog_or_tty->print_cr("\nLive "SIZE_FORMAT" objects, " 7954 SIZE_FORMAT" bytes " 7955 "Already free "SIZE_FORMAT" objects, "SIZE_FORMAT" bytes", 7956 _numObjectsLive, _numWordsLive*sizeof(HeapWord), 7957 _numObjectsAlreadyFree, _numWordsAlreadyFree*sizeof(HeapWord)); 7958 size_t totalBytes = (_numWordsFreed + _numWordsLive + _numWordsAlreadyFree) 7959 * sizeof(HeapWord); 7960 gclog_or_tty->print_cr("Total sweep: "SIZE_FORMAT" bytes", totalBytes); 7961 7962 if (PrintCMSStatistics && CMSVerifyReturnedBytes) { 7963 size_t indexListReturnedBytes = _sp->sumIndexedFreeListArrayReturnedBytes(); 7964 size_t dict_returned_bytes = _sp->dictionary()->sum_dict_returned_bytes(); 7965 size_t returned_bytes = indexListReturnedBytes + dict_returned_bytes; 7966 gclog_or_tty->print("Returned "SIZE_FORMAT" bytes", returned_bytes); 7967 gclog_or_tty->print(" Indexed List Returned "SIZE_FORMAT" bytes", 7968 indexListReturnedBytes); 7969 gclog_or_tty->print_cr(" Dictionary Returned "SIZE_FORMAT" bytes", 7970 dict_returned_bytes); 7971 } 7972 } 7973 if (CMSTraceSweeper) { 7974 gclog_or_tty->print_cr("end of sweep with _limit = " PTR_FORMAT "\n================", 7975 _limit); 7976 } 7977 } 7978 #endif // PRODUCT 7979 7980 void SweepClosure::initialize_free_range(HeapWord* freeFinger, 7981 bool freeRangeInFreeLists) { 7982 if (CMSTraceSweeper) { 7983 gclog_or_tty->print("---- Start free range at 0x%x with free block (%d)\n", 7984 freeFinger, freeRangeInFreeLists); 7985 } 7986 assert(!inFreeRange(), "Trampling existing free range"); 7987 set_inFreeRange(true); 7988 set_lastFreeRangeCoalesced(false); 7989 7990 set_freeFinger(freeFinger); 7991 set_freeRangeInFreeLists(freeRangeInFreeLists); 7992 if (CMSTestInFreeList) { 7993 if (freeRangeInFreeLists) { 7994 FreeChunk* fc = (FreeChunk*) freeFinger; 7995 assert(fc->is_free(), "A chunk on the free list should be free."); 7996 assert(fc->size() > 0, "Free range should have a size"); 7997 assert(_sp->verify_chunk_in_free_list(fc), "Chunk is not in free lists"); 7998 } 7999 } 8000 } 8001 8002 // Note that the sweeper runs concurrently with mutators. Thus, 8003 // it is possible for direct allocation in this generation to happen 8004 // in the middle of the sweep. Note that the sweeper also coalesces 8005 // contiguous free blocks. Thus, unless the sweeper and the allocator 8006 // synchronize appropriately freshly allocated blocks may get swept up. 8007 // This is accomplished by the sweeper locking the free lists while 8008 // it is sweeping. Thus blocks that are determined to be free are 8009 // indeed free. There is however one additional complication: 8010 // blocks that have been allocated since the final checkpoint and 8011 // mark, will not have been marked and so would be treated as 8012 // unreachable and swept up. To prevent this, the allocator marks 8013 // the bit map when allocating during the sweep phase. This leads, 8014 // however, to a further complication -- objects may have been allocated 8015 // but not yet initialized -- in the sense that the header isn't yet 8016 // installed. The sweeper can not then determine the size of the block 8017 // in order to skip over it. To deal with this case, we use a technique 8018 // (due to Printezis) to encode such uninitialized block sizes in the 8019 // bit map. Since the bit map uses a bit per every HeapWord, but the 8020 // CMS generation has a minimum object size of 3 HeapWords, it follows 8021 // that "normal marks" won't be adjacent in the bit map (there will 8022 // always be at least two 0 bits between successive 1 bits). We make use 8023 // of these "unused" bits to represent uninitialized blocks -- the bit 8024 // corresponding to the start of the uninitialized object and the next 8025 // bit are both set. Finally, a 1 bit marks the end of the object that 8026 // started with the two consecutive 1 bits to indicate its potentially 8027 // uninitialized state. 8028 8029 size_t SweepClosure::do_blk_careful(HeapWord* addr) { 8030 FreeChunk* fc = (FreeChunk*)addr; 8031 size_t res; 8032 8033 // Check if we are done sweeping. Below we check "addr >= _limit" rather 8034 // than "addr == _limit" because although _limit was a block boundary when 8035 // we started the sweep, it may no longer be one because heap expansion 8036 // may have caused us to coalesce the block ending at the address _limit 8037 // with a newly expanded chunk (this happens when _limit was set to the 8038 // previous _end of the space), so we may have stepped past _limit: 8039 // see the following Zeno-like trail of CRs 6977970, 7008136, 7042740. 8040 if (addr >= _limit) { // we have swept up to or past the limit: finish up 8041 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 8042 "sweep _limit out of bounds"); 8043 assert(addr < _sp->end(), "addr out of bounds"); 8044 // Flush any free range we might be holding as a single 8045 // coalesced chunk to the appropriate free list. 8046 if (inFreeRange()) { 8047 assert(freeFinger() >= _sp->bottom() && freeFinger() < _limit, 8048 err_msg("freeFinger() " PTR_FORMAT" is out-of-bounds", freeFinger())); 8049 flush_cur_free_chunk(freeFinger(), 8050 pointer_delta(addr, freeFinger())); 8051 if (CMSTraceSweeper) { 8052 gclog_or_tty->print("Sweep: last chunk: "); 8053 gclog_or_tty->print("put_free_blk 0x%x ("SIZE_FORMAT") " 8054 "[coalesced:"SIZE_FORMAT"]\n", 8055 freeFinger(), pointer_delta(addr, freeFinger()), 8056 lastFreeRangeCoalesced()); 8057 } 8058 } 8059 8060 // help the iterator loop finish 8061 return pointer_delta(_sp->end(), addr); 8062 } 8063 8064 assert(addr < _limit, "sweep invariant"); 8065 // check if we should yield 8066 do_yield_check(addr); 8067 if (fc->is_free()) { 8068 // Chunk that is already free 8069 res = fc->size(); 8070 do_already_free_chunk(fc); 8071 debug_only(_sp->verifyFreeLists()); 8072 // If we flush the chunk at hand in lookahead_and_flush() 8073 // and it's coalesced with a preceding chunk, then the 8074 // process of "mangling" the payload of the coalesced block 8075 // will cause erasure of the size information from the 8076 // (erstwhile) header of all the coalesced blocks but the 8077 // first, so the first disjunct in the assert will not hold 8078 // in that specific case (in which case the second disjunct 8079 // will hold). 8080 assert(res == fc->size() || ((HeapWord*)fc) + res >= _limit, 8081 "Otherwise the size info doesn't change at this step"); 8082 NOT_PRODUCT( 8083 _numObjectsAlreadyFree++; 8084 _numWordsAlreadyFree += res; 8085 ) 8086 NOT_PRODUCT(_last_fc = fc;) 8087 } else if (!_bitMap->isMarked(addr)) { 8088 // Chunk is fresh garbage 8089 res = do_garbage_chunk(fc); 8090 debug_only(_sp->verifyFreeLists()); 8091 NOT_PRODUCT( 8092 _numObjectsFreed++; 8093 _numWordsFreed += res; 8094 ) 8095 } else { 8096 // Chunk that is alive. 8097 res = do_live_chunk(fc); 8098 debug_only(_sp->verifyFreeLists()); 8099 NOT_PRODUCT( 8100 _numObjectsLive++; 8101 _numWordsLive += res; 8102 ) 8103 } 8104 return res; 8105 } 8106 8107 // For the smart allocation, record following 8108 // split deaths - a free chunk is removed from its free list because 8109 // it is being split into two or more chunks. 8110 // split birth - a free chunk is being added to its free list because 8111 // a larger free chunk has been split and resulted in this free chunk. 8112 // coal death - a free chunk is being removed from its free list because 8113 // it is being coalesced into a large free chunk. 8114 // coal birth - a free chunk is being added to its free list because 8115 // it was created when two or more free chunks where coalesced into 8116 // this free chunk. 8117 // 8118 // These statistics are used to determine the desired number of free 8119 // chunks of a given size. The desired number is chosen to be relative 8120 // to the end of a CMS sweep. The desired number at the end of a sweep 8121 // is the 8122 // count-at-end-of-previous-sweep (an amount that was enough) 8123 // - count-at-beginning-of-current-sweep (the excess) 8124 // + split-births (gains in this size during interval) 8125 // - split-deaths (demands on this size during interval) 8126 // where the interval is from the end of one sweep to the end of the 8127 // next. 8128 // 8129 // When sweeping the sweeper maintains an accumulated chunk which is 8130 // the chunk that is made up of chunks that have been coalesced. That 8131 // will be termed the left-hand chunk. A new chunk of garbage that 8132 // is being considered for coalescing will be referred to as the 8133 // right-hand chunk. 8134 // 8135 // When making a decision on whether to coalesce a right-hand chunk with 8136 // the current left-hand chunk, the current count vs. the desired count 8137 // of the left-hand chunk is considered. Also if the right-hand chunk 8138 // is near the large chunk at the end of the heap (see 8139 // ConcurrentMarkSweepGeneration::isNearLargestChunk()), then the 8140 // left-hand chunk is coalesced. 8141 // 8142 // When making a decision about whether to split a chunk, the desired count 8143 // vs. the current count of the candidate to be split is also considered. 8144 // If the candidate is underpopulated (currently fewer chunks than desired) 8145 // a chunk of an overpopulated (currently more chunks than desired) size may 8146 // be chosen. The "hint" associated with a free list, if non-null, points 8147 // to a free list which may be overpopulated. 8148 // 8149 8150 void SweepClosure::do_already_free_chunk(FreeChunk* fc) { 8151 const size_t size = fc->size(); 8152 // Chunks that cannot be coalesced are not in the 8153 // free lists. 8154 if (CMSTestInFreeList && !fc->cantCoalesce()) { 8155 assert(_sp->verify_chunk_in_free_list(fc), 8156 "free chunk should be in free lists"); 8157 } 8158 // a chunk that is already free, should not have been 8159 // marked in the bit map 8160 HeapWord* const addr = (HeapWord*) fc; 8161 assert(!_bitMap->isMarked(addr), "free chunk should be unmarked"); 8162 // Verify that the bit map has no bits marked between 8163 // addr and purported end of this block. 8164 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 8165 8166 // Some chunks cannot be coalesced under any circumstances. 8167 // See the definition of cantCoalesce(). 8168 if (!fc->cantCoalesce()) { 8169 // This chunk can potentially be coalesced. 8170 if (_sp->adaptive_freelists()) { 8171 // All the work is done in 8172 do_post_free_or_garbage_chunk(fc, size); 8173 } else { // Not adaptive free lists 8174 // this is a free chunk that can potentially be coalesced by the sweeper; 8175 if (!inFreeRange()) { 8176 // if the next chunk is a free block that can't be coalesced 8177 // it doesn't make sense to remove this chunk from the free lists 8178 FreeChunk* nextChunk = (FreeChunk*)(addr + size); 8179 assert((HeapWord*)nextChunk <= _sp->end(), "Chunk size out of bounds?"); 8180 if ((HeapWord*)nextChunk < _sp->end() && // There is another free chunk to the right ... 8181 nextChunk->is_free() && // ... which is free... 8182 nextChunk->cantCoalesce()) { // ... but can't be coalesced 8183 // nothing to do 8184 } else { 8185 // Potentially the start of a new free range: 8186 // Don't eagerly remove it from the free lists. 8187 // No need to remove it if it will just be put 8188 // back again. (Also from a pragmatic point of view 8189 // if it is a free block in a region that is beyond 8190 // any allocated blocks, an assertion will fail) 8191 // Remember the start of a free run. 8192 initialize_free_range(addr, true); 8193 // end - can coalesce with next chunk 8194 } 8195 } else { 8196 // the midst of a free range, we are coalescing 8197 print_free_block_coalesced(fc); 8198 if (CMSTraceSweeper) { 8199 gclog_or_tty->print(" -- pick up free block 0x%x (%d)\n", fc, size); 8200 } 8201 // remove it from the free lists 8202 _sp->removeFreeChunkFromFreeLists(fc); 8203 set_lastFreeRangeCoalesced(true); 8204 // If the chunk is being coalesced and the current free range is 8205 // in the free lists, remove the current free range so that it 8206 // will be returned to the free lists in its entirety - all 8207 // the coalesced pieces included. 8208 if (freeRangeInFreeLists()) { 8209 FreeChunk* ffc = (FreeChunk*) freeFinger(); 8210 assert(ffc->size() == pointer_delta(addr, freeFinger()), 8211 "Size of free range is inconsistent with chunk size."); 8212 if (CMSTestInFreeList) { 8213 assert(_sp->verify_chunk_in_free_list(ffc), 8214 "free range is not in free lists"); 8215 } 8216 _sp->removeFreeChunkFromFreeLists(ffc); 8217 set_freeRangeInFreeLists(false); 8218 } 8219 } 8220 } 8221 // Note that if the chunk is not coalescable (the else arm 8222 // below), we unconditionally flush, without needing to do 8223 // a "lookahead," as we do below. 8224 if (inFreeRange()) lookahead_and_flush(fc, size); 8225 } else { 8226 // Code path common to both original and adaptive free lists. 8227 8228 // cant coalesce with previous block; this should be treated 8229 // as the end of a free run if any 8230 if (inFreeRange()) { 8231 // we kicked some butt; time to pick up the garbage 8232 assert(freeFinger() < addr, "freeFinger points too high"); 8233 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 8234 } 8235 // else, nothing to do, just continue 8236 } 8237 } 8238 8239 size_t SweepClosure::do_garbage_chunk(FreeChunk* fc) { 8240 // This is a chunk of garbage. It is not in any free list. 8241 // Add it to a free list or let it possibly be coalesced into 8242 // a larger chunk. 8243 HeapWord* const addr = (HeapWord*) fc; 8244 const size_t size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 8245 8246 if (_sp->adaptive_freelists()) { 8247 // Verify that the bit map has no bits marked between 8248 // addr and purported end of just dead object. 8249 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 8250 8251 do_post_free_or_garbage_chunk(fc, size); 8252 } else { 8253 if (!inFreeRange()) { 8254 // start of a new free range 8255 assert(size > 0, "A free range should have a size"); 8256 initialize_free_range(addr, false); 8257 } else { 8258 // this will be swept up when we hit the end of the 8259 // free range 8260 if (CMSTraceSweeper) { 8261 gclog_or_tty->print(" -- pick up garbage 0x%x (%d) \n", fc, size); 8262 } 8263 // If the chunk is being coalesced and the current free range is 8264 // in the free lists, remove the current free range so that it 8265 // will be returned to the free lists in its entirety - all 8266 // the coalesced pieces included. 8267 if (freeRangeInFreeLists()) { 8268 FreeChunk* ffc = (FreeChunk*)freeFinger(); 8269 assert(ffc->size() == pointer_delta(addr, freeFinger()), 8270 "Size of free range is inconsistent with chunk size."); 8271 if (CMSTestInFreeList) { 8272 assert(_sp->verify_chunk_in_free_list(ffc), 8273 "free range is not in free lists"); 8274 } 8275 _sp->removeFreeChunkFromFreeLists(ffc); 8276 set_freeRangeInFreeLists(false); 8277 } 8278 set_lastFreeRangeCoalesced(true); 8279 } 8280 // this will be swept up when we hit the end of the free range 8281 8282 // Verify that the bit map has no bits marked between 8283 // addr and purported end of just dead object. 8284 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 8285 } 8286 assert(_limit >= addr + size, 8287 "A freshly garbage chunk can't possibly straddle over _limit"); 8288 if (inFreeRange()) lookahead_and_flush(fc, size); 8289 return size; 8290 } 8291 8292 size_t SweepClosure::do_live_chunk(FreeChunk* fc) { 8293 HeapWord* addr = (HeapWord*) fc; 8294 // The sweeper has just found a live object. Return any accumulated 8295 // left hand chunk to the free lists. 8296 if (inFreeRange()) { 8297 assert(freeFinger() < addr, "freeFinger points too high"); 8298 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 8299 } 8300 8301 // This object is live: we'd normally expect this to be 8302 // an oop, and like to assert the following: 8303 // assert(oop(addr)->is_oop(), "live block should be an oop"); 8304 // However, as we commented above, this may be an object whose 8305 // header hasn't yet been initialized. 8306 size_t size; 8307 assert(_bitMap->isMarked(addr), "Tautology for this control point"); 8308 if (_bitMap->isMarked(addr + 1)) { 8309 // Determine the size from the bit map, rather than trying to 8310 // compute it from the object header. 8311 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 8312 size = pointer_delta(nextOneAddr + 1, addr); 8313 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 8314 "alignment problem"); 8315 8316 #ifdef ASSERT 8317 if (oop(addr)->klass_or_null() != NULL) { 8318 // Ignore mark word because we are running concurrent with mutators 8319 assert(oop(addr)->is_oop(true), "live block should be an oop"); 8320 assert(size == 8321 CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()), 8322 "P-mark and computed size do not agree"); 8323 } 8324 #endif 8325 8326 } else { 8327 // This should be an initialized object that's alive. 8328 assert(oop(addr)->klass_or_null() != NULL, 8329 "Should be an initialized object"); 8330 // Ignore mark word because we are running concurrent with mutators 8331 assert(oop(addr)->is_oop(true), "live block should be an oop"); 8332 // Verify that the bit map has no bits marked between 8333 // addr and purported end of this block. 8334 size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 8335 assert(size >= 3, "Necessary for Printezis marks to work"); 8336 assert(!_bitMap->isMarked(addr+1), "Tautology for this control point"); 8337 DEBUG_ONLY(_bitMap->verifyNoOneBitsInRange(addr+2, addr+size);) 8338 } 8339 return size; 8340 } 8341 8342 void SweepClosure::do_post_free_or_garbage_chunk(FreeChunk* fc, 8343 size_t chunkSize) { 8344 // do_post_free_or_garbage_chunk() should only be called in the case 8345 // of the adaptive free list allocator. 8346 const bool fcInFreeLists = fc->is_free(); 8347 assert(_sp->adaptive_freelists(), "Should only be used in this case."); 8348 assert((HeapWord*)fc <= _limit, "sweep invariant"); 8349 if (CMSTestInFreeList && fcInFreeLists) { 8350 assert(_sp->verify_chunk_in_free_list(fc), "free chunk is not in free lists"); 8351 } 8352 8353 if (CMSTraceSweeper) { 8354 gclog_or_tty->print_cr(" -- pick up another chunk at 0x%x (%d)", fc, chunkSize); 8355 } 8356 8357 HeapWord* const fc_addr = (HeapWord*) fc; 8358 8359 bool coalesce; 8360 const size_t left = pointer_delta(fc_addr, freeFinger()); 8361 const size_t right = chunkSize; 8362 switch (FLSCoalescePolicy) { 8363 // numeric value forms a coalition aggressiveness metric 8364 case 0: { // never coalesce 8365 coalesce = false; 8366 break; 8367 } 8368 case 1: { // coalesce if left & right chunks on overpopulated lists 8369 coalesce = _sp->coalOverPopulated(left) && 8370 _sp->coalOverPopulated(right); 8371 break; 8372 } 8373 case 2: { // coalesce if left chunk on overpopulated list (default) 8374 coalesce = _sp->coalOverPopulated(left); 8375 break; 8376 } 8377 case 3: { // coalesce if left OR right chunk on overpopulated list 8378 coalesce = _sp->coalOverPopulated(left) || 8379 _sp->coalOverPopulated(right); 8380 break; 8381 } 8382 case 4: { // always coalesce 8383 coalesce = true; 8384 break; 8385 } 8386 default: 8387 ShouldNotReachHere(); 8388 } 8389 8390 // Should the current free range be coalesced? 8391 // If the chunk is in a free range and either we decided to coalesce above 8392 // or the chunk is near the large block at the end of the heap 8393 // (isNearLargestChunk() returns true), then coalesce this chunk. 8394 const bool doCoalesce = inFreeRange() 8395 && (coalesce || _g->isNearLargestChunk(fc_addr)); 8396 if (doCoalesce) { 8397 // Coalesce the current free range on the left with the new 8398 // chunk on the right. If either is on a free list, 8399 // it must be removed from the list and stashed in the closure. 8400 if (freeRangeInFreeLists()) { 8401 FreeChunk* const ffc = (FreeChunk*)freeFinger(); 8402 assert(ffc->size() == pointer_delta(fc_addr, freeFinger()), 8403 "Size of free range is inconsistent with chunk size."); 8404 if (CMSTestInFreeList) { 8405 assert(_sp->verify_chunk_in_free_list(ffc), 8406 "Chunk is not in free lists"); 8407 } 8408 _sp->coalDeath(ffc->size()); 8409 _sp->removeFreeChunkFromFreeLists(ffc); 8410 set_freeRangeInFreeLists(false); 8411 } 8412 if (fcInFreeLists) { 8413 _sp->coalDeath(chunkSize); 8414 assert(fc->size() == chunkSize, 8415 "The chunk has the wrong size or is not in the free lists"); 8416 _sp->removeFreeChunkFromFreeLists(fc); 8417 } 8418 set_lastFreeRangeCoalesced(true); 8419 print_free_block_coalesced(fc); 8420 } else { // not in a free range and/or should not coalesce 8421 // Return the current free range and start a new one. 8422 if (inFreeRange()) { 8423 // In a free range but cannot coalesce with the right hand chunk. 8424 // Put the current free range into the free lists. 8425 flush_cur_free_chunk(freeFinger(), 8426 pointer_delta(fc_addr, freeFinger())); 8427 } 8428 // Set up for new free range. Pass along whether the right hand 8429 // chunk is in the free lists. 8430 initialize_free_range((HeapWord*)fc, fcInFreeLists); 8431 } 8432 } 8433 8434 // Lookahead flush: 8435 // If we are tracking a free range, and this is the last chunk that 8436 // we'll look at because its end crosses past _limit, we'll preemptively 8437 // flush it along with any free range we may be holding on to. Note that 8438 // this can be the case only for an already free or freshly garbage 8439 // chunk. If this block is an object, it can never straddle 8440 // over _limit. The "straddling" occurs when _limit is set at 8441 // the previous end of the space when this cycle started, and 8442 // a subsequent heap expansion caused the previously co-terminal 8443 // free block to be coalesced with the newly expanded portion, 8444 // thus rendering _limit a non-block-boundary making it dangerous 8445 // for the sweeper to step over and examine. 8446 void SweepClosure::lookahead_and_flush(FreeChunk* fc, size_t chunk_size) { 8447 assert(inFreeRange(), "Should only be called if currently in a free range."); 8448 HeapWord* const eob = ((HeapWord*)fc) + chunk_size; 8449 assert(_sp->used_region().contains(eob - 1), 8450 err_msg("eob = " PTR_FORMAT " out of bounds wrt _sp = [" PTR_FORMAT "," PTR_FORMAT ")" 8451 " when examining fc = " PTR_FORMAT "(" SIZE_FORMAT ")", 8452 _limit, _sp->bottom(), _sp->end(), fc, chunk_size)); 8453 if (eob >= _limit) { 8454 assert(eob == _limit || fc->is_free(), "Only a free chunk should allow us to cross over the limit"); 8455 if (CMSTraceSweeper) { 8456 gclog_or_tty->print_cr("_limit " PTR_FORMAT " reached or crossed by block " 8457 "[" PTR_FORMAT "," PTR_FORMAT ") in space " 8458 "[" PTR_FORMAT "," PTR_FORMAT ")", 8459 _limit, fc, eob, _sp->bottom(), _sp->end()); 8460 } 8461 // Return the storage we are tracking back into the free lists. 8462 if (CMSTraceSweeper) { 8463 gclog_or_tty->print_cr("Flushing ... "); 8464 } 8465 assert(freeFinger() < eob, "Error"); 8466 flush_cur_free_chunk( freeFinger(), pointer_delta(eob, freeFinger())); 8467 } 8468 } 8469 8470 void SweepClosure::flush_cur_free_chunk(HeapWord* chunk, size_t size) { 8471 assert(inFreeRange(), "Should only be called if currently in a free range."); 8472 assert(size > 0, 8473 "A zero sized chunk cannot be added to the free lists."); 8474 if (!freeRangeInFreeLists()) { 8475 if (CMSTestInFreeList) { 8476 FreeChunk* fc = (FreeChunk*) chunk; 8477 fc->set_size(size); 8478 assert(!_sp->verify_chunk_in_free_list(fc), 8479 "chunk should not be in free lists yet"); 8480 } 8481 if (CMSTraceSweeper) { 8482 gclog_or_tty->print_cr(" -- add free block 0x%x (%d) to free lists", 8483 chunk, size); 8484 } 8485 // A new free range is going to be starting. The current 8486 // free range has not been added to the free lists yet or 8487 // was removed so add it back. 8488 // If the current free range was coalesced, then the death 8489 // of the free range was recorded. Record a birth now. 8490 if (lastFreeRangeCoalesced()) { 8491 _sp->coalBirth(size); 8492 } 8493 _sp->addChunkAndRepairOffsetTable(chunk, size, 8494 lastFreeRangeCoalesced()); 8495 } else if (CMSTraceSweeper) { 8496 gclog_or_tty->print_cr("Already in free list: nothing to flush"); 8497 } 8498 set_inFreeRange(false); 8499 set_freeRangeInFreeLists(false); 8500 } 8501 8502 // We take a break if we've been at this for a while, 8503 // so as to avoid monopolizing the locks involved. 8504 void SweepClosure::do_yield_work(HeapWord* addr) { 8505 // Return current free chunk being used for coalescing (if any) 8506 // to the appropriate freelist. After yielding, the next 8507 // free block encountered will start a coalescing range of 8508 // free blocks. If the next free block is adjacent to the 8509 // chunk just flushed, they will need to wait for the next 8510 // sweep to be coalesced. 8511 if (inFreeRange()) { 8512 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 8513 } 8514 8515 // First give up the locks, then yield, then re-lock. 8516 // We should probably use a constructor/destructor idiom to 8517 // do this unlock/lock or modify the MutexUnlocker class to 8518 // serve our purpose. XXX 8519 assert_lock_strong(_bitMap->lock()); 8520 assert_lock_strong(_freelistLock); 8521 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 8522 "CMS thread should hold CMS token"); 8523 _bitMap->lock()->unlock(); 8524 _freelistLock->unlock(); 8525 ConcurrentMarkSweepThread::desynchronize(true); 8526 ConcurrentMarkSweepThread::acknowledge_yield_request(); 8527 _collector->stopTimer(); 8528 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 8529 if (PrintCMSStatistics != 0) { 8530 _collector->incrementYields(); 8531 } 8532 _collector->icms_wait(); 8533 8534 // See the comment in coordinator_yield() 8535 for (unsigned i = 0; i < CMSYieldSleepCount && 8536 ConcurrentMarkSweepThread::should_yield() && 8537 !CMSCollector::foregroundGCIsActive(); ++i) { 8538 os::sleep(Thread::current(), 1, false); 8539 ConcurrentMarkSweepThread::acknowledge_yield_request(); 8540 } 8541 8542 ConcurrentMarkSweepThread::synchronize(true); 8543 _freelistLock->lock(); 8544 _bitMap->lock()->lock_without_safepoint_check(); 8545 _collector->startTimer(); 8546 } 8547 8548 #ifndef PRODUCT 8549 // This is actually very useful in a product build if it can 8550 // be called from the debugger. Compile it into the product 8551 // as needed. 8552 bool debug_verify_chunk_in_free_list(FreeChunk* fc) { 8553 return debug_cms_space->verify_chunk_in_free_list(fc); 8554 } 8555 #endif 8556 8557 void SweepClosure::print_free_block_coalesced(FreeChunk* fc) const { 8558 if (CMSTraceSweeper) { 8559 gclog_or_tty->print_cr("Sweep:coal_free_blk " PTR_FORMAT " (" SIZE_FORMAT ")", 8560 fc, fc->size()); 8561 } 8562 } 8563 8564 // CMSIsAliveClosure 8565 bool CMSIsAliveClosure::do_object_b(oop obj) { 8566 HeapWord* addr = (HeapWord*)obj; 8567 return addr != NULL && 8568 (!_span.contains(addr) || _bit_map->isMarked(addr)); 8569 } 8570 8571 8572 CMSKeepAliveClosure::CMSKeepAliveClosure( CMSCollector* collector, 8573 MemRegion span, 8574 CMSBitMap* bit_map, CMSMarkStack* mark_stack, 8575 bool cpc): 8576 _collector(collector), 8577 _span(span), 8578 _bit_map(bit_map), 8579 _mark_stack(mark_stack), 8580 _concurrent_precleaning(cpc) { 8581 assert(!_span.is_empty(), "Empty span could spell trouble"); 8582 } 8583 8584 8585 // CMSKeepAliveClosure: the serial version 8586 void CMSKeepAliveClosure::do_oop(oop obj) { 8587 HeapWord* addr = (HeapWord*)obj; 8588 if (_span.contains(addr) && 8589 !_bit_map->isMarked(addr)) { 8590 _bit_map->mark(addr); 8591 bool simulate_overflow = false; 8592 NOT_PRODUCT( 8593 if (CMSMarkStackOverflowALot && 8594 _collector->simulate_overflow()) { 8595 // simulate a stack overflow 8596 simulate_overflow = true; 8597 } 8598 ) 8599 if (simulate_overflow || !_mark_stack->push(obj)) { 8600 if (_concurrent_precleaning) { 8601 // We dirty the overflown object and let the remark 8602 // phase deal with it. 8603 assert(_collector->overflow_list_is_empty(), "Error"); 8604 // In the case of object arrays, we need to dirty all of 8605 // the cards that the object spans. No locking or atomics 8606 // are needed since no one else can be mutating the mod union 8607 // table. 8608 if (obj->is_objArray()) { 8609 size_t sz = obj->size(); 8610 HeapWord* end_card_addr = 8611 (HeapWord*)round_to((intptr_t)(addr+sz), CardTableModRefBS::card_size); 8612 MemRegion redirty_range = MemRegion(addr, end_card_addr); 8613 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 8614 _collector->_modUnionTable.mark_range(redirty_range); 8615 } else { 8616 _collector->_modUnionTable.mark(addr); 8617 } 8618 _collector->_ser_kac_preclean_ovflw++; 8619 } else { 8620 _collector->push_on_overflow_list(obj); 8621 _collector->_ser_kac_ovflw++; 8622 } 8623 } 8624 } 8625 } 8626 8627 void CMSKeepAliveClosure::do_oop(oop* p) { CMSKeepAliveClosure::do_oop_work(p); } 8628 void CMSKeepAliveClosure::do_oop(narrowOop* p) { CMSKeepAliveClosure::do_oop_work(p); } 8629 8630 // CMSParKeepAliveClosure: a parallel version of the above. 8631 // The work queues are private to each closure (thread), 8632 // but (may be) available for stealing by other threads. 8633 void CMSParKeepAliveClosure::do_oop(oop obj) { 8634 HeapWord* addr = (HeapWord*)obj; 8635 if (_span.contains(addr) && 8636 !_bit_map->isMarked(addr)) { 8637 // In general, during recursive tracing, several threads 8638 // may be concurrently getting here; the first one to 8639 // "tag" it, claims it. 8640 if (_bit_map->par_mark(addr)) { 8641 bool res = _work_queue->push(obj); 8642 assert(res, "Low water mark should be much less than capacity"); 8643 // Do a recursive trim in the hope that this will keep 8644 // stack usage lower, but leave some oops for potential stealers 8645 trim_queue(_low_water_mark); 8646 } // Else, another thread got there first 8647 } 8648 } 8649 8650 void CMSParKeepAliveClosure::do_oop(oop* p) { CMSParKeepAliveClosure::do_oop_work(p); } 8651 void CMSParKeepAliveClosure::do_oop(narrowOop* p) { CMSParKeepAliveClosure::do_oop_work(p); } 8652 8653 void CMSParKeepAliveClosure::trim_queue(uint max) { 8654 while (_work_queue->size() > max) { 8655 oop new_oop; 8656 if (_work_queue->pop_local(new_oop)) { 8657 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 8658 assert(_bit_map->isMarked((HeapWord*)new_oop), 8659 "no white objects on this stack!"); 8660 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 8661 // iterate over the oops in this oop, marking and pushing 8662 // the ones in CMS heap (i.e. in _span). 8663 new_oop->oop_iterate(&_mark_and_push); 8664 } 8665 } 8666 } 8667 8668 CMSInnerParMarkAndPushClosure::CMSInnerParMarkAndPushClosure( 8669 CMSCollector* collector, 8670 MemRegion span, CMSBitMap* bit_map, 8671 OopTaskQueue* work_queue): 8672 _collector(collector), 8673 _span(span), 8674 _bit_map(bit_map), 8675 _work_queue(work_queue) { } 8676 8677 void CMSInnerParMarkAndPushClosure::do_oop(oop obj) { 8678 HeapWord* addr = (HeapWord*)obj; 8679 if (_span.contains(addr) && 8680 !_bit_map->isMarked(addr)) { 8681 if (_bit_map->par_mark(addr)) { 8682 bool simulate_overflow = false; 8683 NOT_PRODUCT( 8684 if (CMSMarkStackOverflowALot && 8685 _collector->par_simulate_overflow()) { 8686 // simulate a stack overflow 8687 simulate_overflow = true; 8688 } 8689 ) 8690 if (simulate_overflow || !_work_queue->push(obj)) { 8691 _collector->par_push_on_overflow_list(obj); 8692 _collector->_par_kac_ovflw++; 8693 } 8694 } // Else another thread got there already 8695 } 8696 } 8697 8698 void CMSInnerParMarkAndPushClosure::do_oop(oop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); } 8699 void CMSInnerParMarkAndPushClosure::do_oop(narrowOop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); } 8700 8701 ////////////////////////////////////////////////////////////////// 8702 // CMSExpansionCause ///////////////////////////// 8703 ////////////////////////////////////////////////////////////////// 8704 const char* CMSExpansionCause::to_string(CMSExpansionCause::Cause cause) { 8705 switch (cause) { 8706 case _no_expansion: 8707 return "No expansion"; 8708 case _satisfy_free_ratio: 8709 return "Free ratio"; 8710 case _satisfy_promotion: 8711 return "Satisfy promotion"; 8712 case _satisfy_allocation: 8713 return "allocation"; 8714 case _allocate_par_lab: 8715 return "Par LAB"; 8716 case _allocate_par_spooling_space: 8717 return "Par Spooling Space"; 8718 case _adaptive_size_policy: 8719 return "Ergonomics"; 8720 default: 8721 return "unknown"; 8722 } 8723 } 8724 8725 void CMSDrainMarkingStackClosure::do_void() { 8726 // the max number to take from overflow list at a time 8727 const size_t num = _mark_stack->capacity()/4; 8728 assert(!_concurrent_precleaning || _collector->overflow_list_is_empty(), 8729 "Overflow list should be NULL during concurrent phases"); 8730 while (!_mark_stack->isEmpty() || 8731 // if stack is empty, check the overflow list 8732 _collector->take_from_overflow_list(num, _mark_stack)) { 8733 oop obj = _mark_stack->pop(); 8734 HeapWord* addr = (HeapWord*)obj; 8735 assert(_span.contains(addr), "Should be within span"); 8736 assert(_bit_map->isMarked(addr), "Should be marked"); 8737 assert(obj->is_oop(), "Should be an oop"); 8738 obj->oop_iterate(_keep_alive); 8739 } 8740 } 8741 8742 void CMSParDrainMarkingStackClosure::do_void() { 8743 // drain queue 8744 trim_queue(0); 8745 } 8746 8747 // Trim our work_queue so its length is below max at return 8748 void CMSParDrainMarkingStackClosure::trim_queue(uint max) { 8749 while (_work_queue->size() > max) { 8750 oop new_oop; 8751 if (_work_queue->pop_local(new_oop)) { 8752 assert(new_oop->is_oop(), "Expected an oop"); 8753 assert(_bit_map->isMarked((HeapWord*)new_oop), 8754 "no white objects on this stack!"); 8755 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 8756 // iterate over the oops in this oop, marking and pushing 8757 // the ones in CMS heap (i.e. in _span). 8758 new_oop->oop_iterate(&_mark_and_push); 8759 } 8760 } 8761 } 8762 8763 //////////////////////////////////////////////////////////////////// 8764 // Support for Marking Stack Overflow list handling and related code 8765 //////////////////////////////////////////////////////////////////// 8766 // Much of the following code is similar in shape and spirit to the 8767 // code used in ParNewGC. We should try and share that code 8768 // as much as possible in the future. 8769 8770 #ifndef PRODUCT 8771 // Debugging support for CMSStackOverflowALot 8772 8773 // It's OK to call this multi-threaded; the worst thing 8774 // that can happen is that we'll get a bunch of closely 8775 // spaced simulated oveflows, but that's OK, in fact 8776 // probably good as it would exercise the overflow code 8777 // under contention. 8778 bool CMSCollector::simulate_overflow() { 8779 if (_overflow_counter-- <= 0) { // just being defensive 8780 _overflow_counter = CMSMarkStackOverflowInterval; 8781 return true; 8782 } else { 8783 return false; 8784 } 8785 } 8786 8787 bool CMSCollector::par_simulate_overflow() { 8788 return simulate_overflow(); 8789 } 8790 #endif 8791 8792 // Single-threaded 8793 bool CMSCollector::take_from_overflow_list(size_t num, CMSMarkStack* stack) { 8794 assert(stack->isEmpty(), "Expected precondition"); 8795 assert(stack->capacity() > num, "Shouldn't bite more than can chew"); 8796 size_t i = num; 8797 oop cur = _overflow_list; 8798 const markOop proto = markOopDesc::prototype(); 8799 NOT_PRODUCT(ssize_t n = 0;) 8800 for (oop next; i > 0 && cur != NULL; cur = next, i--) { 8801 next = oop(cur->mark()); 8802 cur->set_mark(proto); // until proven otherwise 8803 assert(cur->is_oop(), "Should be an oop"); 8804 bool res = stack->push(cur); 8805 assert(res, "Bit off more than can chew?"); 8806 NOT_PRODUCT(n++;) 8807 } 8808 _overflow_list = cur; 8809 #ifndef PRODUCT 8810 assert(_num_par_pushes >= n, "Too many pops?"); 8811 _num_par_pushes -=n; 8812 #endif 8813 return !stack->isEmpty(); 8814 } 8815 8816 #define BUSY (oop(0x1aff1aff)) 8817 // (MT-safe) Get a prefix of at most "num" from the list. 8818 // The overflow list is chained through the mark word of 8819 // each object in the list. We fetch the entire list, 8820 // break off a prefix of the right size and return the 8821 // remainder. If other threads try to take objects from 8822 // the overflow list at that time, they will wait for 8823 // some time to see if data becomes available. If (and 8824 // only if) another thread places one or more object(s) 8825 // on the global list before we have returned the suffix 8826 // to the global list, we will walk down our local list 8827 // to find its end and append the global list to 8828 // our suffix before returning it. This suffix walk can 8829 // prove to be expensive (quadratic in the amount of traffic) 8830 // when there are many objects in the overflow list and 8831 // there is much producer-consumer contention on the list. 8832 // *NOTE*: The overflow list manipulation code here and 8833 // in ParNewGeneration:: are very similar in shape, 8834 // except that in the ParNew case we use the old (from/eden) 8835 // copy of the object to thread the list via its klass word. 8836 // Because of the common code, if you make any changes in 8837 // the code below, please check the ParNew version to see if 8838 // similar changes might be needed. 8839 // CR 6797058 has been filed to consolidate the common code. 8840 bool CMSCollector::par_take_from_overflow_list(size_t num, 8841 OopTaskQueue* work_q, 8842 int no_of_gc_threads) { 8843 assert(work_q->size() == 0, "First empty local work queue"); 8844 assert(num < work_q->max_elems(), "Can't bite more than we can chew"); 8845 if (_overflow_list == NULL) { 8846 return false; 8847 } 8848 // Grab the entire list; we'll put back a suffix 8849 oop prefix = (oop)Atomic::xchg_ptr(BUSY, &_overflow_list); 8850 Thread* tid = Thread::current(); 8851 // Before "no_of_gc_threads" was introduced CMSOverflowSpinCount was 8852 // set to ParallelGCThreads. 8853 size_t CMSOverflowSpinCount = (size_t) no_of_gc_threads; // was ParallelGCThreads; 8854 size_t sleep_time_millis = MAX2((size_t)1, num/100); 8855 // If the list is busy, we spin for a short while, 8856 // sleeping between attempts to get the list. 8857 for (size_t spin = 0; prefix == BUSY && spin < CMSOverflowSpinCount; spin++) { 8858 os::sleep(tid, sleep_time_millis, false); 8859 if (_overflow_list == NULL) { 8860 // Nothing left to take 8861 return false; 8862 } else if (_overflow_list != BUSY) { 8863 // Try and grab the prefix 8864 prefix = (oop)Atomic::xchg_ptr(BUSY, &_overflow_list); 8865 } 8866 } 8867 // If the list was found to be empty, or we spun long 8868 // enough, we give up and return empty-handed. If we leave 8869 // the list in the BUSY state below, it must be the case that 8870 // some other thread holds the overflow list and will set it 8871 // to a non-BUSY state in the future. 8872 if (prefix == NULL || prefix == BUSY) { 8873 // Nothing to take or waited long enough 8874 if (prefix == NULL) { 8875 // Write back the NULL in case we overwrote it with BUSY above 8876 // and it is still the same value. 8877 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY); 8878 } 8879 return false; 8880 } 8881 assert(prefix != NULL && prefix != BUSY, "Error"); 8882 size_t i = num; 8883 oop cur = prefix; 8884 // Walk down the first "num" objects, unless we reach the end. 8885 for (; i > 1 && cur->mark() != NULL; cur = oop(cur->mark()), i--); 8886 if (cur->mark() == NULL) { 8887 // We have "num" or fewer elements in the list, so there 8888 // is nothing to return to the global list. 8889 // Write back the NULL in lieu of the BUSY we wrote 8890 // above, if it is still the same value. 8891 if (_overflow_list == BUSY) { 8892 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY); 8893 } 8894 } else { 8895 // Chop off the suffix and rerturn it to the global list. 8896 assert(cur->mark() != BUSY, "Error"); 8897 oop suffix_head = cur->mark(); // suffix will be put back on global list 8898 cur->set_mark(NULL); // break off suffix 8899 // It's possible that the list is still in the empty(busy) state 8900 // we left it in a short while ago; in that case we may be 8901 // able to place back the suffix without incurring the cost 8902 // of a walk down the list. 8903 oop observed_overflow_list = _overflow_list; 8904 oop cur_overflow_list = observed_overflow_list; 8905 bool attached = false; 8906 while (observed_overflow_list == BUSY || observed_overflow_list == NULL) { 8907 observed_overflow_list = 8908 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list); 8909 if (cur_overflow_list == observed_overflow_list) { 8910 attached = true; 8911 break; 8912 } else cur_overflow_list = observed_overflow_list; 8913 } 8914 if (!attached) { 8915 // Too bad, someone else sneaked in (at least) an element; we'll need 8916 // to do a splice. Find tail of suffix so we can prepend suffix to global 8917 // list. 8918 for (cur = suffix_head; cur->mark() != NULL; cur = (oop)(cur->mark())); 8919 oop suffix_tail = cur; 8920 assert(suffix_tail != NULL && suffix_tail->mark() == NULL, 8921 "Tautology"); 8922 observed_overflow_list = _overflow_list; 8923 do { 8924 cur_overflow_list = observed_overflow_list; 8925 if (cur_overflow_list != BUSY) { 8926 // Do the splice ... 8927 suffix_tail->set_mark(markOop(cur_overflow_list)); 8928 } else { // cur_overflow_list == BUSY 8929 suffix_tail->set_mark(NULL); 8930 } 8931 // ... and try to place spliced list back on overflow_list ... 8932 observed_overflow_list = 8933 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list); 8934 } while (cur_overflow_list != observed_overflow_list); 8935 // ... until we have succeeded in doing so. 8936 } 8937 } 8938 8939 // Push the prefix elements on work_q 8940 assert(prefix != NULL, "control point invariant"); 8941 const markOop proto = markOopDesc::prototype(); 8942 oop next; 8943 NOT_PRODUCT(ssize_t n = 0;) 8944 for (cur = prefix; cur != NULL; cur = next) { 8945 next = oop(cur->mark()); 8946 cur->set_mark(proto); // until proven otherwise 8947 assert(cur->is_oop(), "Should be an oop"); 8948 bool res = work_q->push(cur); 8949 assert(res, "Bit off more than we can chew?"); 8950 NOT_PRODUCT(n++;) 8951 } 8952 #ifndef PRODUCT 8953 assert(_num_par_pushes >= n, "Too many pops?"); 8954 Atomic::add_ptr(-(intptr_t)n, &_num_par_pushes); 8955 #endif 8956 return true; 8957 } 8958 8959 // Single-threaded 8960 void CMSCollector::push_on_overflow_list(oop p) { 8961 NOT_PRODUCT(_num_par_pushes++;) 8962 assert(p->is_oop(), "Not an oop"); 8963 preserve_mark_if_necessary(p); 8964 p->set_mark((markOop)_overflow_list); 8965 _overflow_list = p; 8966 } 8967 8968 // Multi-threaded; use CAS to prepend to overflow list 8969 void CMSCollector::par_push_on_overflow_list(oop p) { 8970 NOT_PRODUCT(Atomic::inc_ptr(&_num_par_pushes);) 8971 assert(p->is_oop(), "Not an oop"); 8972 par_preserve_mark_if_necessary(p); 8973 oop observed_overflow_list = _overflow_list; 8974 oop cur_overflow_list; 8975 do { 8976 cur_overflow_list = observed_overflow_list; 8977 if (cur_overflow_list != BUSY) { 8978 p->set_mark(markOop(cur_overflow_list)); 8979 } else { 8980 p->set_mark(NULL); 8981 } 8982 observed_overflow_list = 8983 (oop) Atomic::cmpxchg_ptr(p, &_overflow_list, cur_overflow_list); 8984 } while (cur_overflow_list != observed_overflow_list); 8985 } 8986 #undef BUSY 8987 8988 // Single threaded 8989 // General Note on GrowableArray: pushes may silently fail 8990 // because we are (temporarily) out of C-heap for expanding 8991 // the stack. The problem is quite ubiquitous and affects 8992 // a lot of code in the JVM. The prudent thing for GrowableArray 8993 // to do (for now) is to exit with an error. However, that may 8994 // be too draconian in some cases because the caller may be 8995 // able to recover without much harm. For such cases, we 8996 // should probably introduce a "soft_push" method which returns 8997 // an indication of success or failure with the assumption that 8998 // the caller may be able to recover from a failure; code in 8999 // the VM can then be changed, incrementally, to deal with such 9000 // failures where possible, thus, incrementally hardening the VM 9001 // in such low resource situations. 9002 void CMSCollector::preserve_mark_work(oop p, markOop m) { 9003 _preserved_oop_stack.push(p); 9004 _preserved_mark_stack.push(m); 9005 assert(m == p->mark(), "Mark word changed"); 9006 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), 9007 "bijection"); 9008 } 9009 9010 // Single threaded 9011 void CMSCollector::preserve_mark_if_necessary(oop p) { 9012 markOop m = p->mark(); 9013 if (m->must_be_preserved(p)) { 9014 preserve_mark_work(p, m); 9015 } 9016 } 9017 9018 void CMSCollector::par_preserve_mark_if_necessary(oop p) { 9019 markOop m = p->mark(); 9020 if (m->must_be_preserved(p)) { 9021 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag); 9022 // Even though we read the mark word without holding 9023 // the lock, we are assured that it will not change 9024 // because we "own" this oop, so no other thread can 9025 // be trying to push it on the overflow list; see 9026 // the assertion in preserve_mark_work() that checks 9027 // that m == p->mark(). 9028 preserve_mark_work(p, m); 9029 } 9030 } 9031 9032 // We should be able to do this multi-threaded, 9033 // a chunk of stack being a task (this is 9034 // correct because each oop only ever appears 9035 // once in the overflow list. However, it's 9036 // not very easy to completely overlap this with 9037 // other operations, so will generally not be done 9038 // until all work's been completed. Because we 9039 // expect the preserved oop stack (set) to be small, 9040 // it's probably fine to do this single-threaded. 9041 // We can explore cleverer concurrent/overlapped/parallel 9042 // processing of preserved marks if we feel the 9043 // need for this in the future. Stack overflow should 9044 // be so rare in practice and, when it happens, its 9045 // effect on performance so great that this will 9046 // likely just be in the noise anyway. 9047 void CMSCollector::restore_preserved_marks_if_any() { 9048 assert(SafepointSynchronize::is_at_safepoint(), 9049 "world should be stopped"); 9050 assert(Thread::current()->is_ConcurrentGC_thread() || 9051 Thread::current()->is_VM_thread(), 9052 "should be single-threaded"); 9053 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), 9054 "bijection"); 9055 9056 while (!_preserved_oop_stack.is_empty()) { 9057 oop p = _preserved_oop_stack.pop(); 9058 assert(p->is_oop(), "Should be an oop"); 9059 assert(_span.contains(p), "oop should be in _span"); 9060 assert(p->mark() == markOopDesc::prototype(), 9061 "Set when taken from overflow list"); 9062 markOop m = _preserved_mark_stack.pop(); 9063 p->set_mark(m); 9064 } 9065 assert(_preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(), 9066 "stacks were cleared above"); 9067 } 9068 9069 #ifndef PRODUCT 9070 bool CMSCollector::no_preserved_marks() const { 9071 return _preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(); 9072 } 9073 #endif 9074 9075 CMSAdaptiveSizePolicy* ASConcurrentMarkSweepGeneration::cms_size_policy() const 9076 { 9077 GenCollectedHeap* gch = (GenCollectedHeap*) GenCollectedHeap::heap(); 9078 CMSAdaptiveSizePolicy* size_policy = 9079 (CMSAdaptiveSizePolicy*) gch->gen_policy()->size_policy(); 9080 assert(size_policy->is_gc_cms_adaptive_size_policy(), 9081 "Wrong type for size policy"); 9082 return size_policy; 9083 } 9084 9085 void ASConcurrentMarkSweepGeneration::resize(size_t cur_promo_size, 9086 size_t desired_promo_size) { 9087 if (cur_promo_size < desired_promo_size) { 9088 size_t expand_bytes = desired_promo_size - cur_promo_size; 9089 if (PrintAdaptiveSizePolicy && Verbose) { 9090 gclog_or_tty->print_cr(" ASConcurrentMarkSweepGeneration::resize " 9091 "Expanding tenured generation by " SIZE_FORMAT " (bytes)", 9092 expand_bytes); 9093 } 9094 expand(expand_bytes, 9095 MinHeapDeltaBytes, 9096 CMSExpansionCause::_adaptive_size_policy); 9097 } else if (desired_promo_size < cur_promo_size) { 9098 size_t shrink_bytes = cur_promo_size - desired_promo_size; 9099 if (PrintAdaptiveSizePolicy && Verbose) { 9100 gclog_or_tty->print_cr(" ASConcurrentMarkSweepGeneration::resize " 9101 "Shrinking tenured generation by " SIZE_FORMAT " (bytes)", 9102 shrink_bytes); 9103 } 9104 shrink(shrink_bytes); 9105 } 9106 } 9107 9108 CMSGCAdaptivePolicyCounters* ASConcurrentMarkSweepGeneration::gc_adaptive_policy_counters() { 9109 GenCollectedHeap* gch = GenCollectedHeap::heap(); 9110 CMSGCAdaptivePolicyCounters* counters = 9111 (CMSGCAdaptivePolicyCounters*) gch->collector_policy()->counters(); 9112 assert(counters->kind() == GCPolicyCounters::CMSGCAdaptivePolicyCountersKind, 9113 "Wrong kind of counters"); 9114 return counters; 9115 } 9116 9117 9118 void ASConcurrentMarkSweepGeneration::update_counters() { 9119 if (UsePerfData) { 9120 _space_counters->update_all(); 9121 _gen_counters->update_all(); 9122 CMSGCAdaptivePolicyCounters* counters = gc_adaptive_policy_counters(); 9123 GenCollectedHeap* gch = GenCollectedHeap::heap(); 9124 CMSGCStats* gc_stats_l = (CMSGCStats*) gc_stats(); 9125 assert(gc_stats_l->kind() == GCStats::CMSGCStatsKind, 9126 "Wrong gc statistics type"); 9127 counters->update_counters(gc_stats_l); 9128 } 9129 } 9130 9131 void ASConcurrentMarkSweepGeneration::update_counters(size_t used) { 9132 if (UsePerfData) { 9133 _space_counters->update_used(used); 9134 _space_counters->update_capacity(); 9135 _gen_counters->update_all(); 9136 9137 CMSGCAdaptivePolicyCounters* counters = gc_adaptive_policy_counters(); 9138 GenCollectedHeap* gch = GenCollectedHeap::heap(); 9139 CMSGCStats* gc_stats_l = (CMSGCStats*) gc_stats(); 9140 assert(gc_stats_l->kind() == GCStats::CMSGCStatsKind, 9141 "Wrong gc statistics type"); 9142 counters->update_counters(gc_stats_l); 9143 } 9144 } 9145 9146 void ASConcurrentMarkSweepGeneration::shrink_by(size_t desired_bytes) { 9147 assert_locked_or_safepoint(Heap_lock); 9148 assert_lock_strong(freelistLock()); 9149 HeapWord* old_end = _cmsSpace->end(); 9150 HeapWord* unallocated_start = _cmsSpace->unallocated_block(); 9151 assert(old_end >= unallocated_start, "Miscalculation of unallocated_start"); 9152 FreeChunk* chunk_at_end = find_chunk_at_end(); 9153 if (chunk_at_end == NULL) { 9154 // No room to shrink 9155 if (PrintGCDetails && Verbose) { 9156 gclog_or_tty->print_cr("No room to shrink: old_end " 9157 PTR_FORMAT " unallocated_start " PTR_FORMAT 9158 " chunk_at_end " PTR_FORMAT, 9159 old_end, unallocated_start, chunk_at_end); 9160 } 9161 return; 9162 } else { 9163 9164 // Find the chunk at the end of the space and determine 9165 // how much it can be shrunk. 9166 size_t shrinkable_size_in_bytes = chunk_at_end->size(); 9167 size_t aligned_shrinkable_size_in_bytes = 9168 align_size_down(shrinkable_size_in_bytes, os::vm_page_size()); 9169 assert(unallocated_start <= (HeapWord*) chunk_at_end->end(), 9170 "Inconsistent chunk at end of space"); 9171 size_t bytes = MIN2(desired_bytes, aligned_shrinkable_size_in_bytes); 9172 size_t word_size_before = heap_word_size(_virtual_space.committed_size()); 9173 9174 // Shrink the underlying space 9175 _virtual_space.shrink_by(bytes); 9176 if (PrintGCDetails && Verbose) { 9177 gclog_or_tty->print_cr("ConcurrentMarkSweepGeneration::shrink_by:" 9178 " desired_bytes " SIZE_FORMAT 9179 " shrinkable_size_in_bytes " SIZE_FORMAT 9180 " aligned_shrinkable_size_in_bytes " SIZE_FORMAT 9181 " bytes " SIZE_FORMAT, 9182 desired_bytes, shrinkable_size_in_bytes, 9183 aligned_shrinkable_size_in_bytes, bytes); 9184 gclog_or_tty->print_cr(" old_end " SIZE_FORMAT 9185 " unallocated_start " SIZE_FORMAT, 9186 old_end, unallocated_start); 9187 } 9188 9189 // If the space did shrink (shrinking is not guaranteed), 9190 // shrink the chunk at the end by the appropriate amount. 9191 if (((HeapWord*)_virtual_space.high()) < old_end) { 9192 size_t new_word_size = 9193 heap_word_size(_virtual_space.committed_size()); 9194 9195 // Have to remove the chunk from the dictionary because it is changing 9196 // size and might be someplace elsewhere in the dictionary. 9197 9198 // Get the chunk at end, shrink it, and put it 9199 // back. 9200 _cmsSpace->removeChunkFromDictionary(chunk_at_end); 9201 size_t word_size_change = word_size_before - new_word_size; 9202 size_t chunk_at_end_old_size = chunk_at_end->size(); 9203 assert(chunk_at_end_old_size >= word_size_change, 9204 "Shrink is too large"); 9205 chunk_at_end->set_size(chunk_at_end_old_size - 9206 word_size_change); 9207 _cmsSpace->freed((HeapWord*) chunk_at_end->end(), 9208 word_size_change); 9209 9210 _cmsSpace->returnChunkToDictionary(chunk_at_end); 9211 9212 MemRegion mr(_cmsSpace->bottom(), new_word_size); 9213 _bts->resize(new_word_size); // resize the block offset shared array 9214 Universe::heap()->barrier_set()->resize_covered_region(mr); 9215 _cmsSpace->assert_locked(); 9216 _cmsSpace->set_end((HeapWord*)_virtual_space.high()); 9217 9218 NOT_PRODUCT(_cmsSpace->dictionary()->verify()); 9219 9220 // update the space and generation capacity counters 9221 if (UsePerfData) { 9222 _space_counters->update_capacity(); 9223 _gen_counters->update_all(); 9224 } 9225 9226 if (Verbose && PrintGCDetails) { 9227 size_t new_mem_size = _virtual_space.committed_size(); 9228 size_t old_mem_size = new_mem_size + bytes; 9229 gclog_or_tty->print_cr("Shrinking %s from " SIZE_FORMAT "K by " SIZE_FORMAT "K to " SIZE_FORMAT "K", 9230 name(), old_mem_size/K, bytes/K, new_mem_size/K); 9231 } 9232 } 9233 9234 assert(_cmsSpace->unallocated_block() <= _cmsSpace->end(), 9235 "Inconsistency at end of space"); 9236 assert(chunk_at_end->end() == (uintptr_t*) _cmsSpace->end(), 9237 "Shrinking is inconsistent"); 9238 return; 9239 } 9240 } 9241 9242 // Transfer some number of overflown objects to usual marking 9243 // stack. Return true if some objects were transferred. 9244 bool MarkRefsIntoAndScanClosure::take_from_overflow_list() { 9245 size_t num = MIN2((size_t)(_mark_stack->capacity() - _mark_stack->length())/4, 9246 (size_t)ParGCDesiredObjsFromOverflowList); 9247 9248 bool res = _collector->take_from_overflow_list(num, _mark_stack); 9249 assert(_collector->overflow_list_is_empty() || res, 9250 "If list is not empty, we should have taken something"); 9251 assert(!res || !_mark_stack->isEmpty(), 9252 "If we took something, it should now be on our stack"); 9253 return res; 9254 } 9255 9256 size_t MarkDeadObjectsClosure::do_blk(HeapWord* addr) { 9257 size_t res = _sp->block_size_no_stall(addr, _collector); 9258 if (_sp->block_is_obj(addr)) { 9259 if (_live_bit_map->isMarked(addr)) { 9260 // It can't have been dead in a previous cycle 9261 guarantee(!_dead_bit_map->isMarked(addr), "No resurrection!"); 9262 } else { 9263 _dead_bit_map->mark(addr); // mark the dead object 9264 } 9265 } 9266 // Could be 0, if the block size could not be computed without stalling. 9267 return res; 9268 } 9269 9270 TraceCMSMemoryManagerStats::TraceCMSMemoryManagerStats(CMSCollector::CollectorState phase, GCCause::Cause cause): TraceMemoryManagerStats() { 9271 9272 switch (phase) { 9273 case CMSCollector::InitialMarking: 9274 initialize(true /* fullGC */ , 9275 cause /* cause of the GC */, 9276 true /* recordGCBeginTime */, 9277 true /* recordPreGCUsage */, 9278 false /* recordPeakUsage */, 9279 false /* recordPostGCusage */, 9280 true /* recordAccumulatedGCTime */, 9281 false /* recordGCEndTime */, 9282 false /* countCollection */ ); 9283 break; 9284 9285 case CMSCollector::FinalMarking: 9286 initialize(true /* fullGC */ , 9287 cause /* cause of the GC */, 9288 false /* recordGCBeginTime */, 9289 false /* recordPreGCUsage */, 9290 false /* recordPeakUsage */, 9291 false /* recordPostGCusage */, 9292 true /* recordAccumulatedGCTime */, 9293 false /* recordGCEndTime */, 9294 false /* countCollection */ ); 9295 break; 9296 9297 case CMSCollector::Sweeping: 9298 initialize(true /* fullGC */ , 9299 cause /* cause of the GC */, 9300 false /* recordGCBeginTime */, 9301 false /* recordPreGCUsage */, 9302 true /* recordPeakUsage */, 9303 true /* recordPostGCusage */, 9304 false /* recordAccumulatedGCTime */, 9305 true /* recordGCEndTime */, 9306 true /* countCollection */ ); 9307 break; 9308 9309 default: 9310 ShouldNotReachHere(); 9311 } 9312 } 9313