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 size_t new_word_size = 3385 heap_word_size(_virtual_space.committed_size()); 3386 MemRegion mr(_cmsSpace->bottom(), new_word_size); 3387 _bts->resize(new_word_size); // resize the block offset shared array 3388 Universe::heap()->barrier_set()->resize_covered_region(mr); 3389 // Hmmmm... why doesn't CFLS::set_end verify locking? 3390 // This is quite ugly; FIX ME XXX 3391 _cmsSpace->assert_locked(freelistLock()); 3392 _cmsSpace->set_end((HeapWord*)_virtual_space.high()); 3393 3394 // update the space and generation capacity counters 3395 if (UsePerfData) { 3396 _space_counters->update_capacity(); 3397 _gen_counters->update_all(); 3398 } 3399 3400 if (Verbose && PrintGC) { 3401 size_t new_mem_size = _virtual_space.committed_size(); 3402 size_t old_mem_size = new_mem_size - bytes; 3403 gclog_or_tty->print_cr("Expanding %s from " SIZE_FORMAT "K by " SIZE_FORMAT "K to " SIZE_FORMAT "K", 3404 name(), old_mem_size/K, bytes/K, new_mem_size/K); 3405 } 3406 } 3407 return result; 3408 } 3409 3410 bool ConcurrentMarkSweepGeneration::grow_to_reserved() { 3411 assert_locked_or_safepoint(Heap_lock); 3412 bool success = true; 3413 const size_t remaining_bytes = _virtual_space.uncommitted_size(); 3414 if (remaining_bytes > 0) { 3415 success = grow_by(remaining_bytes); 3416 DEBUG_ONLY(if (!success) warning("grow to reserved failed");) 3417 } 3418 return success; 3419 } 3420 3421 void ConcurrentMarkSweepGeneration::shrink_free_list_by(size_t bytes) { 3422 assert_locked_or_safepoint(Heap_lock); 3423 assert_lock_strong(freelistLock()); 3424 if (PrintGCDetails && Verbose) { 3425 warning("Shrinking of CMS not yet implemented"); 3426 } 3427 return; 3428 } 3429 3430 3431 // Simple ctor/dtor wrapper for accounting & timer chores around concurrent 3432 // phases. 3433 class CMSPhaseAccounting: public StackObj { 3434 public: 3435 CMSPhaseAccounting(CMSCollector *collector, 3436 const char *phase, 3437 bool print_cr = true); 3438 ~CMSPhaseAccounting(); 3439 3440 private: 3441 CMSCollector *_collector; 3442 const char *_phase; 3443 elapsedTimer _wallclock; 3444 bool _print_cr; 3445 3446 public: 3447 // Not MT-safe; so do not pass around these StackObj's 3448 // where they may be accessed by other threads. 3449 jlong wallclock_millis() { 3450 assert(_wallclock.is_active(), "Wall clock should not stop"); 3451 _wallclock.stop(); // to record time 3452 jlong ret = _wallclock.milliseconds(); 3453 _wallclock.start(); // restart 3454 return ret; 3455 } 3456 }; 3457 3458 CMSPhaseAccounting::CMSPhaseAccounting(CMSCollector *collector, 3459 const char *phase, 3460 bool print_cr) : 3461 _collector(collector), _phase(phase), _print_cr(print_cr) { 3462 3463 if (PrintCMSStatistics != 0) { 3464 _collector->resetYields(); 3465 } 3466 if (PrintGCDetails) { 3467 gclog_or_tty->date_stamp(PrintGCDateStamps); 3468 gclog_or_tty->stamp(PrintGCTimeStamps); 3469 gclog_or_tty->print_cr("[%s-concurrent-%s-start]", 3470 _collector->cmsGen()->short_name(), _phase); 3471 } 3472 _collector->resetTimer(); 3473 _wallclock.start(); 3474 _collector->startTimer(); 3475 } 3476 3477 CMSPhaseAccounting::~CMSPhaseAccounting() { 3478 assert(_wallclock.is_active(), "Wall clock should not have stopped"); 3479 _collector->stopTimer(); 3480 _wallclock.stop(); 3481 if (PrintGCDetails) { 3482 gclog_or_tty->date_stamp(PrintGCDateStamps); 3483 gclog_or_tty->stamp(PrintGCTimeStamps); 3484 gclog_or_tty->print("[%s-concurrent-%s: %3.3f/%3.3f secs]", 3485 _collector->cmsGen()->short_name(), 3486 _phase, _collector->timerValue(), _wallclock.seconds()); 3487 if (_print_cr) { 3488 gclog_or_tty->print_cr(""); 3489 } 3490 if (PrintCMSStatistics != 0) { 3491 gclog_or_tty->print_cr(" (CMS-concurrent-%s yielded %d times)", _phase, 3492 _collector->yields()); 3493 } 3494 } 3495 } 3496 3497 // CMS work 3498 3499 // Checkpoint the roots into this generation from outside 3500 // this generation. [Note this initial checkpoint need only 3501 // be approximate -- we'll do a catch up phase subsequently.] 3502 void CMSCollector::checkpointRootsInitial(bool asynch) { 3503 assert(_collectorState == InitialMarking, "Wrong collector state"); 3504 check_correct_thread_executing(); 3505 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 3506 3507 ReferenceProcessor* rp = ref_processor(); 3508 SpecializationStats::clear(); 3509 assert(_restart_addr == NULL, "Control point invariant"); 3510 if (asynch) { 3511 // acquire locks for subsequent manipulations 3512 MutexLockerEx x(bitMapLock(), 3513 Mutex::_no_safepoint_check_flag); 3514 checkpointRootsInitialWork(asynch); 3515 // enable ("weak") refs discovery 3516 rp->enable_discovery(true /*verify_disabled*/, true /*check_no_refs*/); 3517 _collectorState = Marking; 3518 } else { 3519 // (Weak) Refs discovery: this is controlled from genCollectedHeap::do_collection 3520 // which recognizes if we are a CMS generation, and doesn't try to turn on 3521 // discovery; verify that they aren't meddling. 3522 assert(!rp->discovery_is_atomic(), 3523 "incorrect setting of discovery predicate"); 3524 assert(!rp->discovery_enabled(), "genCollectedHeap shouldn't control " 3525 "ref discovery for this generation kind"); 3526 // already have locks 3527 checkpointRootsInitialWork(asynch); 3528 // now enable ("weak") refs discovery 3529 rp->enable_discovery(true /*verify_disabled*/, false /*verify_no_refs*/); 3530 _collectorState = Marking; 3531 } 3532 SpecializationStats::print(); 3533 } 3534 3535 void CMSCollector::checkpointRootsInitialWork(bool asynch) { 3536 assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped"); 3537 assert(_collectorState == InitialMarking, "just checking"); 3538 3539 // If there has not been a GC[n-1] since last GC[n] cycle completed, 3540 // precede our marking with a collection of all 3541 // younger generations to keep floating garbage to a minimum. 3542 // XXX: we won't do this for now -- it's an optimization to be done later. 3543 3544 // already have locks 3545 assert_lock_strong(bitMapLock()); 3546 assert(_markBitMap.isAllClear(), "was reset at end of previous cycle"); 3547 3548 // Setup the verification and class unloading state for this 3549 // CMS collection cycle. 3550 setup_cms_unloading_and_verification_state(); 3551 3552 NOT_PRODUCT(TraceTime t("\ncheckpointRootsInitialWork", 3553 PrintGCDetails && Verbose, true, gclog_or_tty);) 3554 if (UseAdaptiveSizePolicy) { 3555 size_policy()->checkpoint_roots_initial_begin(); 3556 } 3557 3558 // Reset all the PLAB chunk arrays if necessary. 3559 if (_survivor_plab_array != NULL && !CMSPLABRecordAlways) { 3560 reset_survivor_plab_arrays(); 3561 } 3562 3563 ResourceMark rm; 3564 HandleMark hm; 3565 3566 FalseClosure falseClosure; 3567 // In the case of a synchronous collection, we will elide the 3568 // remark step, so it's important to catch all the nmethod oops 3569 // in this step. 3570 // The final 'true' flag to gen_process_strong_roots will ensure this. 3571 // If 'async' is true, we can relax the nmethod tracing. 3572 MarkRefsIntoClosure notOlder(_span, &_markBitMap); 3573 GenCollectedHeap* gch = GenCollectedHeap::heap(); 3574 3575 verify_work_stacks_empty(); 3576 verify_overflow_empty(); 3577 3578 gch->ensure_parsability(false); // fill TLABs, but no need to retire them 3579 // Update the saved marks which may affect the root scans. 3580 gch->save_marks(); 3581 3582 // weak reference processing has not started yet. 3583 ref_processor()->set_enqueuing_is_done(false); 3584 3585 // Need to remember all newly created CLDs, 3586 // so that we can guarantee that the remark finds them. 3587 ClassLoaderDataGraph::remember_new_clds(true); 3588 3589 // Whenever a CLD is found, it will be claimed before proceeding to mark 3590 // the klasses. The claimed marks need to be cleared before marking starts. 3591 ClassLoaderDataGraph::clear_claimed_marks(); 3592 3593 CMKlassClosure klass_closure(¬Older); 3594 { 3595 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;) 3596 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 3597 gch->gen_process_strong_roots(_cmsGen->level(), 3598 true, // younger gens are roots 3599 true, // activate StrongRootsScope 3600 false, // not scavenging 3601 SharedHeap::ScanningOption(roots_scanning_options()), 3602 ¬Older, 3603 true, // walk all of code cache if (so & SO_CodeCache) 3604 NULL, 3605 &klass_closure); 3606 } 3607 3608 // Clear mod-union table; it will be dirtied in the prologue of 3609 // CMS generation per each younger generation collection. 3610 3611 assert(_modUnionTable.isAllClear(), 3612 "Was cleared in most recent final checkpoint phase" 3613 " or no bits are set in the gc_prologue before the start of the next " 3614 "subsequent marking phase."); 3615 3616 assert(_ct->klass_rem_set()->mod_union_is_clear(), "Must be"); 3617 3618 // Save the end of the used_region of the constituent generations 3619 // to be used to limit the extent of sweep in each generation. 3620 save_sweep_limits(); 3621 if (UseAdaptiveSizePolicy) { 3622 size_policy()->checkpoint_roots_initial_end(gch->gc_cause()); 3623 } 3624 verify_overflow_empty(); 3625 } 3626 3627 bool CMSCollector::markFromRoots(bool asynch) { 3628 // we might be tempted to assert that: 3629 // assert(asynch == !SafepointSynchronize::is_at_safepoint(), 3630 // "inconsistent argument?"); 3631 // However that wouldn't be right, because it's possible that 3632 // a safepoint is indeed in progress as a younger generation 3633 // stop-the-world GC happens even as we mark in this generation. 3634 assert(_collectorState == Marking, "inconsistent state?"); 3635 check_correct_thread_executing(); 3636 verify_overflow_empty(); 3637 3638 bool res; 3639 if (asynch) { 3640 3641 // Start the timers for adaptive size policy for the concurrent phases 3642 // Do it here so that the foreground MS can use the concurrent 3643 // timer since a foreground MS might has the sweep done concurrently 3644 // or STW. 3645 if (UseAdaptiveSizePolicy) { 3646 size_policy()->concurrent_marking_begin(); 3647 } 3648 3649 // Weak ref discovery note: We may be discovering weak 3650 // refs in this generation concurrent (but interleaved) with 3651 // weak ref discovery by a younger generation collector. 3652 3653 CMSTokenSyncWithLocks ts(true, bitMapLock()); 3654 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 3655 CMSPhaseAccounting pa(this, "mark", !PrintGCDetails); 3656 res = markFromRootsWork(asynch); 3657 if (res) { 3658 _collectorState = Precleaning; 3659 } else { // We failed and a foreground collection wants to take over 3660 assert(_foregroundGCIsActive, "internal state inconsistency"); 3661 assert(_restart_addr == NULL, "foreground will restart from scratch"); 3662 if (PrintGCDetails) { 3663 gclog_or_tty->print_cr("bailing out to foreground collection"); 3664 } 3665 } 3666 if (UseAdaptiveSizePolicy) { 3667 size_policy()->concurrent_marking_end(); 3668 } 3669 } else { 3670 assert(SafepointSynchronize::is_at_safepoint(), 3671 "inconsistent with asynch == false"); 3672 if (UseAdaptiveSizePolicy) { 3673 size_policy()->ms_collection_marking_begin(); 3674 } 3675 // already have locks 3676 res = markFromRootsWork(asynch); 3677 _collectorState = FinalMarking; 3678 if (UseAdaptiveSizePolicy) { 3679 GenCollectedHeap* gch = GenCollectedHeap::heap(); 3680 size_policy()->ms_collection_marking_end(gch->gc_cause()); 3681 } 3682 } 3683 verify_overflow_empty(); 3684 return res; 3685 } 3686 3687 bool CMSCollector::markFromRootsWork(bool asynch) { 3688 // iterate over marked bits in bit map, doing a full scan and mark 3689 // from these roots using the following algorithm: 3690 // . if oop is to the right of the current scan pointer, 3691 // mark corresponding bit (we'll process it later) 3692 // . else (oop is to left of current scan pointer) 3693 // push oop on marking stack 3694 // . drain the marking stack 3695 3696 // Note that when we do a marking step we need to hold the 3697 // bit map lock -- recall that direct allocation (by mutators) 3698 // and promotion (by younger generation collectors) is also 3699 // marking the bit map. [the so-called allocate live policy.] 3700 // Because the implementation of bit map marking is not 3701 // robust wrt simultaneous marking of bits in the same word, 3702 // we need to make sure that there is no such interference 3703 // between concurrent such updates. 3704 3705 // already have locks 3706 assert_lock_strong(bitMapLock()); 3707 3708 verify_work_stacks_empty(); 3709 verify_overflow_empty(); 3710 bool result = false; 3711 if (CMSConcurrentMTEnabled && ConcGCThreads > 0) { 3712 result = do_marking_mt(asynch); 3713 } else { 3714 result = do_marking_st(asynch); 3715 } 3716 return result; 3717 } 3718 3719 // Forward decl 3720 class CMSConcMarkingTask; 3721 3722 class CMSConcMarkingTerminator: public ParallelTaskTerminator { 3723 CMSCollector* _collector; 3724 CMSConcMarkingTask* _task; 3725 public: 3726 virtual void yield(); 3727 3728 // "n_threads" is the number of threads to be terminated. 3729 // "queue_set" is a set of work queues of other threads. 3730 // "collector" is the CMS collector associated with this task terminator. 3731 // "yield" indicates whether we need the gang as a whole to yield. 3732 CMSConcMarkingTerminator(int n_threads, TaskQueueSetSuper* queue_set, CMSCollector* collector) : 3733 ParallelTaskTerminator(n_threads, queue_set), 3734 _collector(collector) { } 3735 3736 void set_task(CMSConcMarkingTask* task) { 3737 _task = task; 3738 } 3739 }; 3740 3741 class CMSConcMarkingTerminatorTerminator: public TerminatorTerminator { 3742 CMSConcMarkingTask* _task; 3743 public: 3744 bool should_exit_termination(); 3745 void set_task(CMSConcMarkingTask* task) { 3746 _task = task; 3747 } 3748 }; 3749 3750 // MT Concurrent Marking Task 3751 class CMSConcMarkingTask: public YieldingFlexibleGangTask { 3752 CMSCollector* _collector; 3753 int _n_workers; // requested/desired # workers 3754 bool _asynch; 3755 bool _result; 3756 CompactibleFreeListSpace* _cms_space; 3757 char _pad_front[64]; // padding to ... 3758 HeapWord* _global_finger; // ... avoid sharing cache line 3759 char _pad_back[64]; 3760 HeapWord* _restart_addr; 3761 3762 // Exposed here for yielding support 3763 Mutex* const _bit_map_lock; 3764 3765 // The per thread work queues, available here for stealing 3766 OopTaskQueueSet* _task_queues; 3767 3768 // Termination (and yielding) support 3769 CMSConcMarkingTerminator _term; 3770 CMSConcMarkingTerminatorTerminator _term_term; 3771 3772 public: 3773 CMSConcMarkingTask(CMSCollector* collector, 3774 CompactibleFreeListSpace* cms_space, 3775 bool asynch, 3776 YieldingFlexibleWorkGang* workers, 3777 OopTaskQueueSet* task_queues): 3778 YieldingFlexibleGangTask("Concurrent marking done multi-threaded"), 3779 _collector(collector), 3780 _cms_space(cms_space), 3781 _asynch(asynch), _n_workers(0), _result(true), 3782 _task_queues(task_queues), 3783 _term(_n_workers, task_queues, _collector), 3784 _bit_map_lock(collector->bitMapLock()) 3785 { 3786 _requested_size = _n_workers; 3787 _term.set_task(this); 3788 _term_term.set_task(this); 3789 _restart_addr = _global_finger = _cms_space->bottom(); 3790 } 3791 3792 3793 OopTaskQueueSet* task_queues() { return _task_queues; } 3794 3795 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 3796 3797 HeapWord** global_finger_addr() { return &_global_finger; } 3798 3799 CMSConcMarkingTerminator* terminator() { return &_term; } 3800 3801 virtual void set_for_termination(int active_workers) { 3802 terminator()->reset_for_reuse(active_workers); 3803 } 3804 3805 void work(uint worker_id); 3806 bool should_yield() { 3807 return ConcurrentMarkSweepThread::should_yield() 3808 && !_collector->foregroundGCIsActive() 3809 && _asynch; 3810 } 3811 3812 virtual void coordinator_yield(); // stuff done by coordinator 3813 bool result() { return _result; } 3814 3815 void reset(HeapWord* ra) { 3816 assert(_global_finger >= _cms_space->end(), "Postcondition of ::work(i)"); 3817 _restart_addr = _global_finger = ra; 3818 _term.reset_for_reuse(); 3819 } 3820 3821 static bool get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 3822 OopTaskQueue* work_q); 3823 3824 private: 3825 void do_scan_and_mark(int i, CompactibleFreeListSpace* sp); 3826 void do_work_steal(int i); 3827 void bump_global_finger(HeapWord* f); 3828 }; 3829 3830 bool CMSConcMarkingTerminatorTerminator::should_exit_termination() { 3831 assert(_task != NULL, "Error"); 3832 return _task->yielding(); 3833 // Note that we do not need the disjunct || _task->should_yield() above 3834 // because we want terminating threads to yield only if the task 3835 // is already in the midst of yielding, which happens only after at least one 3836 // thread has yielded. 3837 } 3838 3839 void CMSConcMarkingTerminator::yield() { 3840 if (_task->should_yield()) { 3841 _task->yield(); 3842 } else { 3843 ParallelTaskTerminator::yield(); 3844 } 3845 } 3846 3847 //////////////////////////////////////////////////////////////// 3848 // Concurrent Marking Algorithm Sketch 3849 //////////////////////////////////////////////////////////////// 3850 // Until all tasks exhausted (both spaces): 3851 // -- claim next available chunk 3852 // -- bump global finger via CAS 3853 // -- find first object that starts in this chunk 3854 // and start scanning bitmap from that position 3855 // -- scan marked objects for oops 3856 // -- CAS-mark target, and if successful: 3857 // . if target oop is above global finger (volatile read) 3858 // nothing to do 3859 // . if target oop is in chunk and above local finger 3860 // then nothing to do 3861 // . else push on work-queue 3862 // -- Deal with possible overflow issues: 3863 // . local work-queue overflow causes stuff to be pushed on 3864 // global (common) overflow queue 3865 // . always first empty local work queue 3866 // . then get a batch of oops from global work queue if any 3867 // . then do work stealing 3868 // -- When all tasks claimed (both spaces) 3869 // and local work queue empty, 3870 // then in a loop do: 3871 // . check global overflow stack; steal a batch of oops and trace 3872 // . try to steal from other threads oif GOS is empty 3873 // . if neither is available, offer termination 3874 // -- Terminate and return result 3875 // 3876 void CMSConcMarkingTask::work(uint worker_id) { 3877 elapsedTimer _timer; 3878 ResourceMark rm; 3879 HandleMark hm; 3880 3881 DEBUG_ONLY(_collector->verify_overflow_empty();) 3882 3883 // Before we begin work, our work queue should be empty 3884 assert(work_queue(worker_id)->size() == 0, "Expected to be empty"); 3885 // Scan the bitmap covering _cms_space, tracing through grey objects. 3886 _timer.start(); 3887 do_scan_and_mark(worker_id, _cms_space); 3888 _timer.stop(); 3889 if (PrintCMSStatistics != 0) { 3890 gclog_or_tty->print_cr("Finished cms space scanning in %dth thread: %3.3f sec", 3891 worker_id, _timer.seconds()); 3892 // XXX: need xxx/xxx type of notation, two timers 3893 } 3894 3895 // ... do work stealing 3896 _timer.reset(); 3897 _timer.start(); 3898 do_work_steal(worker_id); 3899 _timer.stop(); 3900 if (PrintCMSStatistics != 0) { 3901 gclog_or_tty->print_cr("Finished work stealing in %dth thread: %3.3f sec", 3902 worker_id, _timer.seconds()); 3903 // XXX: need xxx/xxx type of notation, two timers 3904 } 3905 assert(_collector->_markStack.isEmpty(), "Should have been emptied"); 3906 assert(work_queue(worker_id)->size() == 0, "Should have been emptied"); 3907 // Note that under the current task protocol, the 3908 // following assertion is true even of the spaces 3909 // expanded since the completion of the concurrent 3910 // marking. XXX This will likely change under a strict 3911 // ABORT semantics. 3912 // After perm removal the comparison was changed to 3913 // greater than or equal to from strictly greater than. 3914 // Before perm removal the highest address sweep would 3915 // have been at the end of perm gen but now is at the 3916 // end of the tenured gen. 3917 assert(_global_finger >= _cms_space->end(), 3918 "All tasks have been completed"); 3919 DEBUG_ONLY(_collector->verify_overflow_empty();) 3920 } 3921 3922 void CMSConcMarkingTask::bump_global_finger(HeapWord* f) { 3923 HeapWord* read = _global_finger; 3924 HeapWord* cur = read; 3925 while (f > read) { 3926 cur = read; 3927 read = (HeapWord*) Atomic::cmpxchg_ptr(f, &_global_finger, cur); 3928 if (cur == read) { 3929 // our cas succeeded 3930 assert(_global_finger >= f, "protocol consistency"); 3931 break; 3932 } 3933 } 3934 } 3935 3936 // This is really inefficient, and should be redone by 3937 // using (not yet available) block-read and -write interfaces to the 3938 // stack and the work_queue. XXX FIX ME !!! 3939 bool CMSConcMarkingTask::get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 3940 OopTaskQueue* work_q) { 3941 // Fast lock-free check 3942 if (ovflw_stk->length() == 0) { 3943 return false; 3944 } 3945 assert(work_q->size() == 0, "Shouldn't steal"); 3946 MutexLockerEx ml(ovflw_stk->par_lock(), 3947 Mutex::_no_safepoint_check_flag); 3948 // Grab up to 1/4 the size of the work queue 3949 size_t num = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 3950 (size_t)ParGCDesiredObjsFromOverflowList); 3951 num = MIN2(num, ovflw_stk->length()); 3952 for (int i = (int) num; i > 0; i--) { 3953 oop cur = ovflw_stk->pop(); 3954 assert(cur != NULL, "Counted wrong?"); 3955 work_q->push(cur); 3956 } 3957 return num > 0; 3958 } 3959 3960 void CMSConcMarkingTask::do_scan_and_mark(int i, CompactibleFreeListSpace* sp) { 3961 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 3962 int n_tasks = pst->n_tasks(); 3963 // We allow that there may be no tasks to do here because 3964 // we are restarting after a stack overflow. 3965 assert(pst->valid() || n_tasks == 0, "Uninitialized use?"); 3966 uint nth_task = 0; 3967 3968 HeapWord* aligned_start = sp->bottom(); 3969 if (sp->used_region().contains(_restart_addr)) { 3970 // Align down to a card boundary for the start of 0th task 3971 // for this space. 3972 aligned_start = 3973 (HeapWord*)align_size_down((uintptr_t)_restart_addr, 3974 CardTableModRefBS::card_size); 3975 } 3976 3977 size_t chunk_size = sp->marking_task_size(); 3978 while (!pst->is_task_claimed(/* reference */ nth_task)) { 3979 // Having claimed the nth task in this space, 3980 // compute the chunk that it corresponds to: 3981 MemRegion span = MemRegion(aligned_start + nth_task*chunk_size, 3982 aligned_start + (nth_task+1)*chunk_size); 3983 // Try and bump the global finger via a CAS; 3984 // note that we need to do the global finger bump 3985 // _before_ taking the intersection below, because 3986 // the task corresponding to that region will be 3987 // deemed done even if the used_region() expands 3988 // because of allocation -- as it almost certainly will 3989 // during start-up while the threads yield in the 3990 // closure below. 3991 HeapWord* finger = span.end(); 3992 bump_global_finger(finger); // atomically 3993 // There are null tasks here corresponding to chunks 3994 // beyond the "top" address of the space. 3995 span = span.intersection(sp->used_region()); 3996 if (!span.is_empty()) { // Non-null task 3997 HeapWord* prev_obj; 3998 assert(!span.contains(_restart_addr) || nth_task == 0, 3999 "Inconsistency"); 4000 if (nth_task == 0) { 4001 // For the 0th task, we'll not need to compute a block_start. 4002 if (span.contains(_restart_addr)) { 4003 // In the case of a restart because of stack overflow, 4004 // we might additionally skip a chunk prefix. 4005 prev_obj = _restart_addr; 4006 } else { 4007 prev_obj = span.start(); 4008 } 4009 } else { 4010 // We want to skip the first object because 4011 // the protocol is to scan any object in its entirety 4012 // that _starts_ in this span; a fortiori, any 4013 // object starting in an earlier span is scanned 4014 // as part of an earlier claimed task. 4015 // Below we use the "careful" version of block_start 4016 // so we do not try to navigate uninitialized objects. 4017 prev_obj = sp->block_start_careful(span.start()); 4018 // Below we use a variant of block_size that uses the 4019 // Printezis bits to avoid waiting for allocated 4020 // objects to become initialized/parsable. 4021 while (prev_obj < span.start()) { 4022 size_t sz = sp->block_size_no_stall(prev_obj, _collector); 4023 if (sz > 0) { 4024 prev_obj += sz; 4025 } else { 4026 // In this case we may end up doing a bit of redundant 4027 // scanning, but that appears unavoidable, short of 4028 // locking the free list locks; see bug 6324141. 4029 break; 4030 } 4031 } 4032 } 4033 if (prev_obj < span.end()) { 4034 MemRegion my_span = MemRegion(prev_obj, span.end()); 4035 // Do the marking work within a non-empty span -- 4036 // the last argument to the constructor indicates whether the 4037 // iteration should be incremental with periodic yields. 4038 Par_MarkFromRootsClosure cl(this, _collector, my_span, 4039 &_collector->_markBitMap, 4040 work_queue(i), 4041 &_collector->_markStack, 4042 _asynch); 4043 _collector->_markBitMap.iterate(&cl, my_span.start(), my_span.end()); 4044 } // else nothing to do for this task 4045 } // else nothing to do for this task 4046 } 4047 // We'd be tempted to assert here that since there are no 4048 // more tasks left to claim in this space, the global_finger 4049 // must exceed space->top() and a fortiori space->end(). However, 4050 // that would not quite be correct because the bumping of 4051 // global_finger occurs strictly after the claiming of a task, 4052 // so by the time we reach here the global finger may not yet 4053 // have been bumped up by the thread that claimed the last 4054 // task. 4055 pst->all_tasks_completed(); 4056 } 4057 4058 class Par_ConcMarkingClosure: public CMSOopClosure { 4059 private: 4060 CMSCollector* _collector; 4061 CMSConcMarkingTask* _task; 4062 MemRegion _span; 4063 CMSBitMap* _bit_map; 4064 CMSMarkStack* _overflow_stack; 4065 OopTaskQueue* _work_queue; 4066 protected: 4067 DO_OOP_WORK_DEFN 4068 public: 4069 Par_ConcMarkingClosure(CMSCollector* collector, CMSConcMarkingTask* task, OopTaskQueue* work_queue, 4070 CMSBitMap* bit_map, CMSMarkStack* overflow_stack): 4071 CMSOopClosure(collector->ref_processor()), 4072 _collector(collector), 4073 _task(task), 4074 _span(collector->_span), 4075 _work_queue(work_queue), 4076 _bit_map(bit_map), 4077 _overflow_stack(overflow_stack) 4078 { } 4079 virtual void do_oop(oop* p); 4080 virtual void do_oop(narrowOop* p); 4081 4082 void trim_queue(size_t max); 4083 void handle_stack_overflow(HeapWord* lost); 4084 void do_yield_check() { 4085 if (_task->should_yield()) { 4086 _task->yield(); 4087 } 4088 } 4089 }; 4090 4091 // Grey object scanning during work stealing phase -- 4092 // the salient assumption here is that any references 4093 // that are in these stolen objects being scanned must 4094 // already have been initialized (else they would not have 4095 // been published), so we do not need to check for 4096 // uninitialized objects before pushing here. 4097 void Par_ConcMarkingClosure::do_oop(oop obj) { 4098 assert(obj->is_oop_or_null(true), "expected an oop or NULL"); 4099 HeapWord* addr = (HeapWord*)obj; 4100 // Check if oop points into the CMS generation 4101 // and is not marked 4102 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 4103 // a white object ... 4104 // If we manage to "claim" the object, by being the 4105 // first thread to mark it, then we push it on our 4106 // marking stack 4107 if (_bit_map->par_mark(addr)) { // ... now grey 4108 // push on work queue (grey set) 4109 bool simulate_overflow = false; 4110 NOT_PRODUCT( 4111 if (CMSMarkStackOverflowALot && 4112 _collector->simulate_overflow()) { 4113 // simulate a stack overflow 4114 simulate_overflow = true; 4115 } 4116 ) 4117 if (simulate_overflow || 4118 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) { 4119 // stack overflow 4120 if (PrintCMSStatistics != 0) { 4121 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 4122 SIZE_FORMAT, _overflow_stack->capacity()); 4123 } 4124 // We cannot assert that the overflow stack is full because 4125 // it may have been emptied since. 4126 assert(simulate_overflow || 4127 _work_queue->size() == _work_queue->max_elems(), 4128 "Else push should have succeeded"); 4129 handle_stack_overflow(addr); 4130 } 4131 } // Else, some other thread got there first 4132 do_yield_check(); 4133 } 4134 } 4135 4136 void Par_ConcMarkingClosure::do_oop(oop* p) { Par_ConcMarkingClosure::do_oop_work(p); } 4137 void Par_ConcMarkingClosure::do_oop(narrowOop* p) { Par_ConcMarkingClosure::do_oop_work(p); } 4138 4139 void Par_ConcMarkingClosure::trim_queue(size_t max) { 4140 while (_work_queue->size() > max) { 4141 oop new_oop; 4142 if (_work_queue->pop_local(new_oop)) { 4143 assert(new_oop->is_oop(), "Should be an oop"); 4144 assert(_bit_map->isMarked((HeapWord*)new_oop), "Grey object"); 4145 assert(_span.contains((HeapWord*)new_oop), "Not in span"); 4146 new_oop->oop_iterate(this); // do_oop() above 4147 do_yield_check(); 4148 } 4149 } 4150 } 4151 4152 // Upon stack overflow, we discard (part of) the stack, 4153 // remembering the least address amongst those discarded 4154 // in CMSCollector's _restart_address. 4155 void Par_ConcMarkingClosure::handle_stack_overflow(HeapWord* lost) { 4156 // We need to do this under a mutex to prevent other 4157 // workers from interfering with the work done below. 4158 MutexLockerEx ml(_overflow_stack->par_lock(), 4159 Mutex::_no_safepoint_check_flag); 4160 // Remember the least grey address discarded 4161 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 4162 _collector->lower_restart_addr(ra); 4163 _overflow_stack->reset(); // discard stack contents 4164 _overflow_stack->expand(); // expand the stack if possible 4165 } 4166 4167 4168 void CMSConcMarkingTask::do_work_steal(int i) { 4169 OopTaskQueue* work_q = work_queue(i); 4170 oop obj_to_scan; 4171 CMSBitMap* bm = &(_collector->_markBitMap); 4172 CMSMarkStack* ovflw = &(_collector->_markStack); 4173 int* seed = _collector->hash_seed(i); 4174 Par_ConcMarkingClosure cl(_collector, this, work_q, bm, ovflw); 4175 while (true) { 4176 cl.trim_queue(0); 4177 assert(work_q->size() == 0, "Should have been emptied above"); 4178 if (get_work_from_overflow_stack(ovflw, work_q)) { 4179 // Can't assert below because the work obtained from the 4180 // overflow stack may already have been stolen from us. 4181 // assert(work_q->size() > 0, "Work from overflow stack"); 4182 continue; 4183 } else if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 4184 assert(obj_to_scan->is_oop(), "Should be an oop"); 4185 assert(bm->isMarked((HeapWord*)obj_to_scan), "Grey object"); 4186 obj_to_scan->oop_iterate(&cl); 4187 } else if (terminator()->offer_termination(&_term_term)) { 4188 assert(work_q->size() == 0, "Impossible!"); 4189 break; 4190 } else if (yielding() || should_yield()) { 4191 yield(); 4192 } 4193 } 4194 } 4195 4196 // This is run by the CMS (coordinator) thread. 4197 void CMSConcMarkingTask::coordinator_yield() { 4198 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 4199 "CMS thread should hold CMS token"); 4200 // First give up the locks, then yield, then re-lock 4201 // We should probably use a constructor/destructor idiom to 4202 // do this unlock/lock or modify the MutexUnlocker class to 4203 // serve our purpose. XXX 4204 assert_lock_strong(_bit_map_lock); 4205 _bit_map_lock->unlock(); 4206 ConcurrentMarkSweepThread::desynchronize(true); 4207 ConcurrentMarkSweepThread::acknowledge_yield_request(); 4208 _collector->stopTimer(); 4209 if (PrintCMSStatistics != 0) { 4210 _collector->incrementYields(); 4211 } 4212 _collector->icms_wait(); 4213 4214 // It is possible for whichever thread initiated the yield request 4215 // not to get a chance to wake up and take the bitmap lock between 4216 // this thread releasing it and reacquiring it. So, while the 4217 // should_yield() flag is on, let's sleep for a bit to give the 4218 // other thread a chance to wake up. The limit imposed on the number 4219 // of iterations is defensive, to avoid any unforseen circumstances 4220 // putting us into an infinite loop. Since it's always been this 4221 // (coordinator_yield()) method that was observed to cause the 4222 // problem, we are using a parameter (CMSCoordinatorYieldSleepCount) 4223 // which is by default non-zero. For the other seven methods that 4224 // also perform the yield operation, as are using a different 4225 // parameter (CMSYieldSleepCount) which is by default zero. This way we 4226 // can enable the sleeping for those methods too, if necessary. 4227 // See 6442774. 4228 // 4229 // We really need to reconsider the synchronization between the GC 4230 // thread and the yield-requesting threads in the future and we 4231 // should really use wait/notify, which is the recommended 4232 // way of doing this type of interaction. Additionally, we should 4233 // consolidate the eight methods that do the yield operation and they 4234 // are almost identical into one for better maintenability and 4235 // readability. See 6445193. 4236 // 4237 // Tony 2006.06.29 4238 for (unsigned i = 0; i < CMSCoordinatorYieldSleepCount && 4239 ConcurrentMarkSweepThread::should_yield() && 4240 !CMSCollector::foregroundGCIsActive(); ++i) { 4241 os::sleep(Thread::current(), 1, false); 4242 ConcurrentMarkSweepThread::acknowledge_yield_request(); 4243 } 4244 4245 ConcurrentMarkSweepThread::synchronize(true); 4246 _bit_map_lock->lock_without_safepoint_check(); 4247 _collector->startTimer(); 4248 } 4249 4250 bool CMSCollector::do_marking_mt(bool asynch) { 4251 assert(ConcGCThreads > 0 && conc_workers() != NULL, "precondition"); 4252 int num_workers = AdaptiveSizePolicy::calc_active_conc_workers( 4253 conc_workers()->total_workers(), 4254 conc_workers()->active_workers(), 4255 Threads::number_of_non_daemon_threads()); 4256 conc_workers()->set_active_workers(num_workers); 4257 4258 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 4259 4260 CMSConcMarkingTask tsk(this, 4261 cms_space, 4262 asynch, 4263 conc_workers(), 4264 task_queues()); 4265 4266 // Since the actual number of workers we get may be different 4267 // from the number we requested above, do we need to do anything different 4268 // below? In particular, may be we need to subclass the SequantialSubTasksDone 4269 // class?? XXX 4270 cms_space ->initialize_sequential_subtasks_for_marking(num_workers); 4271 4272 // Refs discovery is already non-atomic. 4273 assert(!ref_processor()->discovery_is_atomic(), "Should be non-atomic"); 4274 assert(ref_processor()->discovery_is_mt(), "Discovery should be MT"); 4275 conc_workers()->start_task(&tsk); 4276 while (tsk.yielded()) { 4277 tsk.coordinator_yield(); 4278 conc_workers()->continue_task(&tsk); 4279 } 4280 // If the task was aborted, _restart_addr will be non-NULL 4281 assert(tsk.completed() || _restart_addr != NULL, "Inconsistency"); 4282 while (_restart_addr != NULL) { 4283 // XXX For now we do not make use of ABORTED state and have not 4284 // yet implemented the right abort semantics (even in the original 4285 // single-threaded CMS case). That needs some more investigation 4286 // and is deferred for now; see CR# TBF. 07252005YSR. XXX 4287 assert(!CMSAbortSemantics || tsk.aborted(), "Inconsistency"); 4288 // If _restart_addr is non-NULL, a marking stack overflow 4289 // occurred; we need to do a fresh marking iteration from the 4290 // indicated restart address. 4291 if (_foregroundGCIsActive && asynch) { 4292 // We may be running into repeated stack overflows, having 4293 // reached the limit of the stack size, while making very 4294 // slow forward progress. It may be best to bail out and 4295 // let the foreground collector do its job. 4296 // Clear _restart_addr, so that foreground GC 4297 // works from scratch. This avoids the headache of 4298 // a "rescan" which would otherwise be needed because 4299 // of the dirty mod union table & card table. 4300 _restart_addr = NULL; 4301 return false; 4302 } 4303 // Adjust the task to restart from _restart_addr 4304 tsk.reset(_restart_addr); 4305 cms_space ->initialize_sequential_subtasks_for_marking(num_workers, 4306 _restart_addr); 4307 _restart_addr = NULL; 4308 // Get the workers going again 4309 conc_workers()->start_task(&tsk); 4310 while (tsk.yielded()) { 4311 tsk.coordinator_yield(); 4312 conc_workers()->continue_task(&tsk); 4313 } 4314 } 4315 assert(tsk.completed(), "Inconsistency"); 4316 assert(tsk.result() == true, "Inconsistency"); 4317 return true; 4318 } 4319 4320 bool CMSCollector::do_marking_st(bool asynch) { 4321 ResourceMark rm; 4322 HandleMark hm; 4323 4324 // Temporarily make refs discovery single threaded (non-MT) 4325 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false); 4326 MarkFromRootsClosure markFromRootsClosure(this, _span, &_markBitMap, 4327 &_markStack, CMSYield && asynch); 4328 // the last argument to iterate indicates whether the iteration 4329 // should be incremental with periodic yields. 4330 _markBitMap.iterate(&markFromRootsClosure); 4331 // If _restart_addr is non-NULL, a marking stack overflow 4332 // occurred; we need to do a fresh iteration from the 4333 // indicated restart address. 4334 while (_restart_addr != NULL) { 4335 if (_foregroundGCIsActive && asynch) { 4336 // We may be running into repeated stack overflows, having 4337 // reached the limit of the stack size, while making very 4338 // slow forward progress. It may be best to bail out and 4339 // let the foreground collector do its job. 4340 // Clear _restart_addr, so that foreground GC 4341 // works from scratch. This avoids the headache of 4342 // a "rescan" which would otherwise be needed because 4343 // of the dirty mod union table & card table. 4344 _restart_addr = NULL; 4345 return false; // indicating failure to complete marking 4346 } 4347 // Deal with stack overflow: 4348 // we restart marking from _restart_addr 4349 HeapWord* ra = _restart_addr; 4350 markFromRootsClosure.reset(ra); 4351 _restart_addr = NULL; 4352 _markBitMap.iterate(&markFromRootsClosure, ra, _span.end()); 4353 } 4354 return true; 4355 } 4356 4357 void CMSCollector::preclean() { 4358 check_correct_thread_executing(); 4359 assert(Thread::current()->is_ConcurrentGC_thread(), "Wrong thread"); 4360 verify_work_stacks_empty(); 4361 verify_overflow_empty(); 4362 _abort_preclean = false; 4363 if (CMSPrecleaningEnabled) { 4364 _eden_chunk_index = 0; 4365 size_t used = get_eden_used(); 4366 size_t capacity = get_eden_capacity(); 4367 // Don't start sampling unless we will get sufficiently 4368 // many samples. 4369 if (used < (capacity/(CMSScheduleRemarkSamplingRatio * 100) 4370 * CMSScheduleRemarkEdenPenetration)) { 4371 _start_sampling = true; 4372 } else { 4373 _start_sampling = false; 4374 } 4375 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 4376 CMSPhaseAccounting pa(this, "preclean", !PrintGCDetails); 4377 preclean_work(CMSPrecleanRefLists1, CMSPrecleanSurvivors1); 4378 } 4379 CMSTokenSync x(true); // is cms thread 4380 if (CMSPrecleaningEnabled) { 4381 sample_eden(); 4382 _collectorState = AbortablePreclean; 4383 } else { 4384 _collectorState = FinalMarking; 4385 } 4386 verify_work_stacks_empty(); 4387 verify_overflow_empty(); 4388 } 4389 4390 // Try and schedule the remark such that young gen 4391 // occupancy is CMSScheduleRemarkEdenPenetration %. 4392 void CMSCollector::abortable_preclean() { 4393 check_correct_thread_executing(); 4394 assert(CMSPrecleaningEnabled, "Inconsistent control state"); 4395 assert(_collectorState == AbortablePreclean, "Inconsistent control state"); 4396 4397 // If Eden's current occupancy is below this threshold, 4398 // immediately schedule the remark; else preclean 4399 // past the next scavenge in an effort to 4400 // schedule the pause as described avove. By choosing 4401 // CMSScheduleRemarkEdenSizeThreshold >= max eden size 4402 // we will never do an actual abortable preclean cycle. 4403 if (get_eden_used() > CMSScheduleRemarkEdenSizeThreshold) { 4404 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 4405 CMSPhaseAccounting pa(this, "abortable-preclean", !PrintGCDetails); 4406 // We need more smarts in the abortable preclean 4407 // loop below to deal with cases where allocation 4408 // in young gen is very very slow, and our precleaning 4409 // is running a losing race against a horde of 4410 // mutators intent on flooding us with CMS updates 4411 // (dirty cards). 4412 // One, admittedly dumb, strategy is to give up 4413 // after a certain number of abortable precleaning loops 4414 // or after a certain maximum time. We want to make 4415 // this smarter in the next iteration. 4416 // XXX FIX ME!!! YSR 4417 size_t loops = 0, workdone = 0, cumworkdone = 0, waited = 0; 4418 while (!(should_abort_preclean() || 4419 ConcurrentMarkSweepThread::should_terminate())) { 4420 workdone = preclean_work(CMSPrecleanRefLists2, CMSPrecleanSurvivors2); 4421 cumworkdone += workdone; 4422 loops++; 4423 // Voluntarily terminate abortable preclean phase if we have 4424 // been at it for too long. 4425 if ((CMSMaxAbortablePrecleanLoops != 0) && 4426 loops >= CMSMaxAbortablePrecleanLoops) { 4427 if (PrintGCDetails) { 4428 gclog_or_tty->print(" CMS: abort preclean due to loops "); 4429 } 4430 break; 4431 } 4432 if (pa.wallclock_millis() > CMSMaxAbortablePrecleanTime) { 4433 if (PrintGCDetails) { 4434 gclog_or_tty->print(" CMS: abort preclean due to time "); 4435 } 4436 break; 4437 } 4438 // If we are doing little work each iteration, we should 4439 // take a short break. 4440 if (workdone < CMSAbortablePrecleanMinWorkPerIteration) { 4441 // Sleep for some time, waiting for work to accumulate 4442 stopTimer(); 4443 cmsThread()->wait_on_cms_lock(CMSAbortablePrecleanWaitMillis); 4444 startTimer(); 4445 waited++; 4446 } 4447 } 4448 if (PrintCMSStatistics > 0) { 4449 gclog_or_tty->print(" [%d iterations, %d waits, %d cards)] ", 4450 loops, waited, cumworkdone); 4451 } 4452 } 4453 CMSTokenSync x(true); // is cms thread 4454 if (_collectorState != Idling) { 4455 assert(_collectorState == AbortablePreclean, 4456 "Spontaneous state transition?"); 4457 _collectorState = FinalMarking; 4458 } // Else, a foreground collection completed this CMS cycle. 4459 return; 4460 } 4461 4462 // Respond to an Eden sampling opportunity 4463 void CMSCollector::sample_eden() { 4464 // Make sure a young gc cannot sneak in between our 4465 // reading and recording of a sample. 4466 assert(Thread::current()->is_ConcurrentGC_thread(), 4467 "Only the cms thread may collect Eden samples"); 4468 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 4469 "Should collect samples while holding CMS token"); 4470 if (!_start_sampling) { 4471 return; 4472 } 4473 if (_eden_chunk_array) { 4474 if (_eden_chunk_index < _eden_chunk_capacity) { 4475 _eden_chunk_array[_eden_chunk_index] = *_top_addr; // take sample 4476 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr, 4477 "Unexpected state of Eden"); 4478 // We'd like to check that what we just sampled is an oop-start address; 4479 // however, we cannot do that here since the object may not yet have been 4480 // initialized. So we'll instead do the check when we _use_ this sample 4481 // later. 4482 if (_eden_chunk_index == 0 || 4483 (pointer_delta(_eden_chunk_array[_eden_chunk_index], 4484 _eden_chunk_array[_eden_chunk_index-1]) 4485 >= CMSSamplingGrain)) { 4486 _eden_chunk_index++; // commit sample 4487 } 4488 } 4489 } 4490 if ((_collectorState == AbortablePreclean) && !_abort_preclean) { 4491 size_t used = get_eden_used(); 4492 size_t capacity = get_eden_capacity(); 4493 assert(used <= capacity, "Unexpected state of Eden"); 4494 if (used > (capacity/100 * CMSScheduleRemarkEdenPenetration)) { 4495 _abort_preclean = true; 4496 } 4497 } 4498 } 4499 4500 4501 size_t CMSCollector::preclean_work(bool clean_refs, bool clean_survivor) { 4502 assert(_collectorState == Precleaning || 4503 _collectorState == AbortablePreclean, "incorrect state"); 4504 ResourceMark rm; 4505 HandleMark hm; 4506 4507 // Precleaning is currently not MT but the reference processor 4508 // may be set for MT. Disable it temporarily here. 4509 ReferenceProcessor* rp = ref_processor(); 4510 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(rp, false); 4511 4512 // Do one pass of scrubbing the discovered reference lists 4513 // to remove any reference objects with strongly-reachable 4514 // referents. 4515 if (clean_refs) { 4516 CMSPrecleanRefsYieldClosure yield_cl(this); 4517 assert(rp->span().equals(_span), "Spans should be equal"); 4518 CMSKeepAliveClosure keep_alive(this, _span, &_markBitMap, 4519 &_markStack, true /* preclean */); 4520 CMSDrainMarkingStackClosure complete_trace(this, 4521 _span, &_markBitMap, &_markStack, 4522 &keep_alive, true /* preclean */); 4523 4524 // We don't want this step to interfere with a young 4525 // collection because we don't want to take CPU 4526 // or memory bandwidth away from the young GC threads 4527 // (which may be as many as there are CPUs). 4528 // Note that we don't need to protect ourselves from 4529 // interference with mutators because they can't 4530 // manipulate the discovered reference lists nor affect 4531 // the computed reachability of the referents, the 4532 // only properties manipulated by the precleaning 4533 // of these reference lists. 4534 stopTimer(); 4535 CMSTokenSyncWithLocks x(true /* is cms thread */, 4536 bitMapLock()); 4537 startTimer(); 4538 sample_eden(); 4539 4540 // The following will yield to allow foreground 4541 // collection to proceed promptly. XXX YSR: 4542 // The code in this method may need further 4543 // tweaking for better performance and some restructuring 4544 // for cleaner interfaces. 4545 rp->preclean_discovered_references( 4546 rp->is_alive_non_header(), &keep_alive, &complete_trace, &yield_cl); 4547 } 4548 4549 if (clean_survivor) { // preclean the active survivor space(s) 4550 assert(_young_gen->kind() == Generation::DefNew || 4551 _young_gen->kind() == Generation::ParNew || 4552 _young_gen->kind() == Generation::ASParNew, 4553 "incorrect type for cast"); 4554 DefNewGeneration* dng = (DefNewGeneration*)_young_gen; 4555 PushAndMarkClosure pam_cl(this, _span, ref_processor(), 4556 &_markBitMap, &_modUnionTable, 4557 &_markStack, true /* precleaning phase */); 4558 stopTimer(); 4559 CMSTokenSyncWithLocks ts(true /* is cms thread */, 4560 bitMapLock()); 4561 startTimer(); 4562 unsigned int before_count = 4563 GenCollectedHeap::heap()->total_collections(); 4564 SurvivorSpacePrecleanClosure 4565 sss_cl(this, _span, &_markBitMap, &_markStack, 4566 &pam_cl, before_count, CMSYield); 4567 dng->from()->object_iterate_careful(&sss_cl); 4568 dng->to()->object_iterate_careful(&sss_cl); 4569 } 4570 MarkRefsIntoAndScanClosure 4571 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable, 4572 &_markStack, this, CMSYield, 4573 true /* precleaning phase */); 4574 // CAUTION: The following closure has persistent state that may need to 4575 // be reset upon a decrease in the sequence of addresses it 4576 // processes. 4577 ScanMarkedObjectsAgainCarefullyClosure 4578 smoac_cl(this, _span, 4579 &_markBitMap, &_markStack, &mrias_cl, CMSYield); 4580 4581 // Preclean dirty cards in ModUnionTable and CardTable using 4582 // appropriate convergence criterion; 4583 // repeat CMSPrecleanIter times unless we find that 4584 // we are losing. 4585 assert(CMSPrecleanIter < 10, "CMSPrecleanIter is too large"); 4586 assert(CMSPrecleanNumerator < CMSPrecleanDenominator, 4587 "Bad convergence multiplier"); 4588 assert(CMSPrecleanThreshold >= 100, 4589 "Unreasonably low CMSPrecleanThreshold"); 4590 4591 size_t numIter, cumNumCards, lastNumCards, curNumCards; 4592 for (numIter = 0, cumNumCards = lastNumCards = curNumCards = 0; 4593 numIter < CMSPrecleanIter; 4594 numIter++, lastNumCards = curNumCards, cumNumCards += curNumCards) { 4595 curNumCards = preclean_mod_union_table(_cmsGen, &smoac_cl); 4596 if (Verbose && PrintGCDetails) { 4597 gclog_or_tty->print(" (modUnionTable: %d cards)", curNumCards); 4598 } 4599 // Either there are very few dirty cards, so re-mark 4600 // pause will be small anyway, or our pre-cleaning isn't 4601 // that much faster than the rate at which cards are being 4602 // dirtied, so we might as well stop and re-mark since 4603 // precleaning won't improve our re-mark time by much. 4604 if (curNumCards <= CMSPrecleanThreshold || 4605 (numIter > 0 && 4606 (curNumCards * CMSPrecleanDenominator > 4607 lastNumCards * CMSPrecleanNumerator))) { 4608 numIter++; 4609 cumNumCards += curNumCards; 4610 break; 4611 } 4612 } 4613 4614 preclean_klasses(&mrias_cl, _cmsGen->freelistLock()); 4615 4616 curNumCards = preclean_card_table(_cmsGen, &smoac_cl); 4617 cumNumCards += curNumCards; 4618 if (PrintGCDetails && PrintCMSStatistics != 0) { 4619 gclog_or_tty->print_cr(" (cardTable: %d cards, re-scanned %d cards, %d iterations)", 4620 curNumCards, cumNumCards, numIter); 4621 } 4622 return cumNumCards; // as a measure of useful work done 4623 } 4624 4625 // PRECLEANING NOTES: 4626 // Precleaning involves: 4627 // . reading the bits of the modUnionTable and clearing the set bits. 4628 // . For the cards corresponding to the set bits, we scan the 4629 // objects on those cards. This means we need the free_list_lock 4630 // so that we can safely iterate over the CMS space when scanning 4631 // for oops. 4632 // . When we scan the objects, we'll be both reading and setting 4633 // marks in the marking bit map, so we'll need the marking bit map. 4634 // . For protecting _collector_state transitions, we take the CGC_lock. 4635 // Note that any races in the reading of of card table entries by the 4636 // CMS thread on the one hand and the clearing of those entries by the 4637 // VM thread or the setting of those entries by the mutator threads on the 4638 // other are quite benign. However, for efficiency it makes sense to keep 4639 // the VM thread from racing with the CMS thread while the latter is 4640 // dirty card info to the modUnionTable. We therefore also use the 4641 // CGC_lock to protect the reading of the card table and the mod union 4642 // table by the CM thread. 4643 // . We run concurrently with mutator updates, so scanning 4644 // needs to be done carefully -- we should not try to scan 4645 // potentially uninitialized objects. 4646 // 4647 // Locking strategy: While holding the CGC_lock, we scan over and 4648 // reset a maximal dirty range of the mod union / card tables, then lock 4649 // the free_list_lock and bitmap lock to do a full marking, then 4650 // release these locks; and repeat the cycle. This allows for a 4651 // certain amount of fairness in the sharing of these locks between 4652 // the CMS collector on the one hand, and the VM thread and the 4653 // mutators on the other. 4654 4655 // NOTE: preclean_mod_union_table() and preclean_card_table() 4656 // further below are largely identical; if you need to modify 4657 // one of these methods, please check the other method too. 4658 4659 size_t CMSCollector::preclean_mod_union_table( 4660 ConcurrentMarkSweepGeneration* gen, 4661 ScanMarkedObjectsAgainCarefullyClosure* cl) { 4662 verify_work_stacks_empty(); 4663 verify_overflow_empty(); 4664 4665 // strategy: starting with the first card, accumulate contiguous 4666 // ranges of dirty cards; clear these cards, then scan the region 4667 // covered by these cards. 4668 4669 // Since all of the MUT is committed ahead, we can just use 4670 // that, in case the generations expand while we are precleaning. 4671 // It might also be fine to just use the committed part of the 4672 // generation, but we might potentially miss cards when the 4673 // generation is rapidly expanding while we are in the midst 4674 // of precleaning. 4675 HeapWord* startAddr = gen->reserved().start(); 4676 HeapWord* endAddr = gen->reserved().end(); 4677 4678 cl->setFreelistLock(gen->freelistLock()); // needed for yielding 4679 4680 size_t numDirtyCards, cumNumDirtyCards; 4681 HeapWord *nextAddr, *lastAddr; 4682 for (cumNumDirtyCards = numDirtyCards = 0, 4683 nextAddr = lastAddr = startAddr; 4684 nextAddr < endAddr; 4685 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 4686 4687 ResourceMark rm; 4688 HandleMark hm; 4689 4690 MemRegion dirtyRegion; 4691 { 4692 stopTimer(); 4693 // Potential yield point 4694 CMSTokenSync ts(true); 4695 startTimer(); 4696 sample_eden(); 4697 // Get dirty region starting at nextOffset (inclusive), 4698 // simultaneously clearing it. 4699 dirtyRegion = 4700 _modUnionTable.getAndClearMarkedRegion(nextAddr, endAddr); 4701 assert(dirtyRegion.start() >= nextAddr, 4702 "returned region inconsistent?"); 4703 } 4704 // Remember where the next search should begin. 4705 // The returned region (if non-empty) is a right open interval, 4706 // so lastOffset is obtained from the right end of that 4707 // interval. 4708 lastAddr = dirtyRegion.end(); 4709 // Should do something more transparent and less hacky XXX 4710 numDirtyCards = 4711 _modUnionTable.heapWordDiffToOffsetDiff(dirtyRegion.word_size()); 4712 4713 // We'll scan the cards in the dirty region (with periodic 4714 // yields for foreground GC as needed). 4715 if (!dirtyRegion.is_empty()) { 4716 assert(numDirtyCards > 0, "consistency check"); 4717 HeapWord* stop_point = NULL; 4718 stopTimer(); 4719 // Potential yield point 4720 CMSTokenSyncWithLocks ts(true, gen->freelistLock(), 4721 bitMapLock()); 4722 startTimer(); 4723 { 4724 verify_work_stacks_empty(); 4725 verify_overflow_empty(); 4726 sample_eden(); 4727 stop_point = 4728 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 4729 } 4730 if (stop_point != NULL) { 4731 // The careful iteration stopped early either because it found an 4732 // uninitialized object, or because we were in the midst of an 4733 // "abortable preclean", which should now be aborted. Redirty 4734 // the bits corresponding to the partially-scanned or unscanned 4735 // cards. We'll either restart at the next block boundary or 4736 // abort the preclean. 4737 assert((_collectorState == AbortablePreclean && should_abort_preclean()), 4738 "Should only be AbortablePreclean."); 4739 _modUnionTable.mark_range(MemRegion(stop_point, dirtyRegion.end())); 4740 if (should_abort_preclean()) { 4741 break; // out of preclean loop 4742 } else { 4743 // Compute the next address at which preclean should pick up; 4744 // might need bitMapLock in order to read P-bits. 4745 lastAddr = next_card_start_after_block(stop_point); 4746 } 4747 } 4748 } else { 4749 assert(lastAddr == endAddr, "consistency check"); 4750 assert(numDirtyCards == 0, "consistency check"); 4751 break; 4752 } 4753 } 4754 verify_work_stacks_empty(); 4755 verify_overflow_empty(); 4756 return cumNumDirtyCards; 4757 } 4758 4759 // NOTE: preclean_mod_union_table() above and preclean_card_table() 4760 // below are largely identical; if you need to modify 4761 // one of these methods, please check the other method too. 4762 4763 size_t CMSCollector::preclean_card_table(ConcurrentMarkSweepGeneration* gen, 4764 ScanMarkedObjectsAgainCarefullyClosure* cl) { 4765 // strategy: it's similar to precleamModUnionTable above, in that 4766 // we accumulate contiguous ranges of dirty cards, mark these cards 4767 // precleaned, then scan the region covered by these cards. 4768 HeapWord* endAddr = (HeapWord*)(gen->_virtual_space.high()); 4769 HeapWord* startAddr = (HeapWord*)(gen->_virtual_space.low()); 4770 4771 cl->setFreelistLock(gen->freelistLock()); // needed for yielding 4772 4773 size_t numDirtyCards, cumNumDirtyCards; 4774 HeapWord *lastAddr, *nextAddr; 4775 4776 for (cumNumDirtyCards = numDirtyCards = 0, 4777 nextAddr = lastAddr = startAddr; 4778 nextAddr < endAddr; 4779 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 4780 4781 ResourceMark rm; 4782 HandleMark hm; 4783 4784 MemRegion dirtyRegion; 4785 { 4786 // See comments in "Precleaning notes" above on why we 4787 // do this locking. XXX Could the locking overheads be 4788 // too high when dirty cards are sparse? [I don't think so.] 4789 stopTimer(); 4790 CMSTokenSync x(true); // is cms thread 4791 startTimer(); 4792 sample_eden(); 4793 // Get and clear dirty region from card table 4794 dirtyRegion = _ct->ct_bs()->dirty_card_range_after_reset( 4795 MemRegion(nextAddr, endAddr), 4796 true, 4797 CardTableModRefBS::precleaned_card_val()); 4798 4799 assert(dirtyRegion.start() >= nextAddr, 4800 "returned region inconsistent?"); 4801 } 4802 lastAddr = dirtyRegion.end(); 4803 numDirtyCards = 4804 dirtyRegion.word_size()/CardTableModRefBS::card_size_in_words; 4805 4806 if (!dirtyRegion.is_empty()) { 4807 stopTimer(); 4808 CMSTokenSyncWithLocks ts(true, gen->freelistLock(), bitMapLock()); 4809 startTimer(); 4810 sample_eden(); 4811 verify_work_stacks_empty(); 4812 verify_overflow_empty(); 4813 HeapWord* stop_point = 4814 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 4815 if (stop_point != NULL) { 4816 assert((_collectorState == AbortablePreclean && should_abort_preclean()), 4817 "Should only be AbortablePreclean."); 4818 _ct->ct_bs()->invalidate(MemRegion(stop_point, dirtyRegion.end())); 4819 if (should_abort_preclean()) { 4820 break; // out of preclean loop 4821 } else { 4822 // Compute the next address at which preclean should pick up. 4823 lastAddr = next_card_start_after_block(stop_point); 4824 } 4825 } 4826 } else { 4827 break; 4828 } 4829 } 4830 verify_work_stacks_empty(); 4831 verify_overflow_empty(); 4832 return cumNumDirtyCards; 4833 } 4834 4835 class PrecleanKlassClosure : public KlassClosure { 4836 CMKlassClosure _cm_klass_closure; 4837 public: 4838 PrecleanKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {} 4839 void do_klass(Klass* k) { 4840 if (k->has_accumulated_modified_oops()) { 4841 k->clear_accumulated_modified_oops(); 4842 4843 _cm_klass_closure.do_klass(k); 4844 } 4845 } 4846 }; 4847 4848 // The freelist lock is needed to prevent asserts, is it really needed? 4849 void CMSCollector::preclean_klasses(MarkRefsIntoAndScanClosure* cl, Mutex* freelistLock) { 4850 4851 cl->set_freelistLock(freelistLock); 4852 4853 CMSTokenSyncWithLocks ts(true, freelistLock, bitMapLock()); 4854 4855 // SSS: Add equivalent to ScanMarkedObjectsAgainCarefullyClosure::do_yield_check and should_abort_preclean? 4856 // SSS: We should probably check if precleaning should be aborted, at suitable intervals? 4857 PrecleanKlassClosure preclean_klass_closure(cl); 4858 ClassLoaderDataGraph::classes_do(&preclean_klass_closure); 4859 4860 verify_work_stacks_empty(); 4861 verify_overflow_empty(); 4862 } 4863 4864 void CMSCollector::checkpointRootsFinal(bool asynch, 4865 bool clear_all_soft_refs, bool init_mark_was_synchronous) { 4866 assert(_collectorState == FinalMarking, "incorrect state transition?"); 4867 check_correct_thread_executing(); 4868 // world is stopped at this checkpoint 4869 assert(SafepointSynchronize::is_at_safepoint(), 4870 "world should be stopped"); 4871 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 4872 4873 verify_work_stacks_empty(); 4874 verify_overflow_empty(); 4875 4876 SpecializationStats::clear(); 4877 if (PrintGCDetails) { 4878 gclog_or_tty->print("[YG occupancy: "SIZE_FORMAT" K ("SIZE_FORMAT" K)]", 4879 _young_gen->used() / K, 4880 _young_gen->capacity() / K); 4881 } 4882 if (asynch) { 4883 if (CMSScavengeBeforeRemark) { 4884 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4885 // Temporarily set flag to false, GCH->do_collection will 4886 // expect it to be false and set to true 4887 FlagSetting fl(gch->_is_gc_active, false); 4888 NOT_PRODUCT(TraceTime t("Scavenge-Before-Remark", 4889 PrintGCDetails && Verbose, true, gclog_or_tty);) 4890 int level = _cmsGen->level() - 1; 4891 if (level >= 0) { 4892 gch->do_collection(true, // full (i.e. force, see below) 4893 false, // !clear_all_soft_refs 4894 0, // size 4895 false, // is_tlab 4896 level // max_level 4897 ); 4898 } 4899 } 4900 FreelistLocker x(this); 4901 MutexLockerEx y(bitMapLock(), 4902 Mutex::_no_safepoint_check_flag); 4903 assert(!init_mark_was_synchronous, "but that's impossible!"); 4904 checkpointRootsFinalWork(asynch, clear_all_soft_refs, false); 4905 } else { 4906 // already have all the locks 4907 checkpointRootsFinalWork(asynch, clear_all_soft_refs, 4908 init_mark_was_synchronous); 4909 } 4910 verify_work_stacks_empty(); 4911 verify_overflow_empty(); 4912 SpecializationStats::print(); 4913 } 4914 4915 void CMSCollector::checkpointRootsFinalWork(bool asynch, 4916 bool clear_all_soft_refs, bool init_mark_was_synchronous) { 4917 4918 NOT_PRODUCT(TraceTime tr("checkpointRootsFinalWork", PrintGCDetails, false, gclog_or_tty);) 4919 4920 assert(haveFreelistLocks(), "must have free list locks"); 4921 assert_lock_strong(bitMapLock()); 4922 4923 if (UseAdaptiveSizePolicy) { 4924 size_policy()->checkpoint_roots_final_begin(); 4925 } 4926 4927 ResourceMark rm; 4928 HandleMark hm; 4929 4930 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4931 4932 if (should_unload_classes()) { 4933 CodeCache::gc_prologue(); 4934 } 4935 assert(haveFreelistLocks(), "must have free list locks"); 4936 assert_lock_strong(bitMapLock()); 4937 4938 if (!init_mark_was_synchronous) { 4939 // We might assume that we need not fill TLAB's when 4940 // CMSScavengeBeforeRemark is set, because we may have just done 4941 // a scavenge which would have filled all TLAB's -- and besides 4942 // Eden would be empty. This however may not always be the case -- 4943 // for instance although we asked for a scavenge, it may not have 4944 // happened because of a JNI critical section. We probably need 4945 // a policy for deciding whether we can in that case wait until 4946 // the critical section releases and then do the remark following 4947 // the scavenge, and skip it here. In the absence of that policy, 4948 // or of an indication of whether the scavenge did indeed occur, 4949 // we cannot rely on TLAB's having been filled and must do 4950 // so here just in case a scavenge did not happen. 4951 gch->ensure_parsability(false); // fill TLAB's, but no need to retire them 4952 // Update the saved marks which may affect the root scans. 4953 gch->save_marks(); 4954 4955 { 4956 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;) 4957 4958 // Note on the role of the mod union table: 4959 // Since the marker in "markFromRoots" marks concurrently with 4960 // mutators, it is possible for some reachable objects not to have been 4961 // scanned. For instance, an only reference to an object A was 4962 // placed in object B after the marker scanned B. Unless B is rescanned, 4963 // A would be collected. Such updates to references in marked objects 4964 // are detected via the mod union table which is the set of all cards 4965 // dirtied since the first checkpoint in this GC cycle and prior to 4966 // the most recent young generation GC, minus those cleaned up by the 4967 // concurrent precleaning. 4968 if (CMSParallelRemarkEnabled && CollectedHeap::use_parallel_gc_threads()) { 4969 TraceTime t("Rescan (parallel) ", PrintGCDetails, false, gclog_or_tty); 4970 do_remark_parallel(); 4971 } else { 4972 TraceTime t("Rescan (non-parallel) ", PrintGCDetails, false, 4973 gclog_or_tty); 4974 do_remark_non_parallel(); 4975 } 4976 } 4977 } else { 4978 assert(!asynch, "Can't have init_mark_was_synchronous in asynch mode"); 4979 // The initial mark was stop-world, so there's no rescanning to 4980 // do; go straight on to the next step below. 4981 } 4982 verify_work_stacks_empty(); 4983 verify_overflow_empty(); 4984 4985 { 4986 NOT_PRODUCT(TraceTime ts("refProcessingWork", PrintGCDetails, false, gclog_or_tty);) 4987 refProcessingWork(asynch, clear_all_soft_refs); 4988 } 4989 verify_work_stacks_empty(); 4990 verify_overflow_empty(); 4991 4992 if (should_unload_classes()) { 4993 CodeCache::gc_epilogue(); 4994 } 4995 JvmtiExport::gc_epilogue(); 4996 4997 // If we encountered any (marking stack / work queue) overflow 4998 // events during the current CMS cycle, take appropriate 4999 // remedial measures, where possible, so as to try and avoid 5000 // recurrence of that condition. 5001 assert(_markStack.isEmpty(), "No grey objects"); 5002 size_t ser_ovflw = _ser_pmc_remark_ovflw + _ser_pmc_preclean_ovflw + 5003 _ser_kac_ovflw + _ser_kac_preclean_ovflw; 5004 if (ser_ovflw > 0) { 5005 if (PrintCMSStatistics != 0) { 5006 gclog_or_tty->print_cr("Marking stack overflow (benign) " 5007 "(pmc_pc="SIZE_FORMAT", pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT 5008 ", kac_preclean="SIZE_FORMAT")", 5009 _ser_pmc_preclean_ovflw, _ser_pmc_remark_ovflw, 5010 _ser_kac_ovflw, _ser_kac_preclean_ovflw); 5011 } 5012 _markStack.expand(); 5013 _ser_pmc_remark_ovflw = 0; 5014 _ser_pmc_preclean_ovflw = 0; 5015 _ser_kac_preclean_ovflw = 0; 5016 _ser_kac_ovflw = 0; 5017 } 5018 if (_par_pmc_remark_ovflw > 0 || _par_kac_ovflw > 0) { 5019 if (PrintCMSStatistics != 0) { 5020 gclog_or_tty->print_cr("Work queue overflow (benign) " 5021 "(pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT")", 5022 _par_pmc_remark_ovflw, _par_kac_ovflw); 5023 } 5024 _par_pmc_remark_ovflw = 0; 5025 _par_kac_ovflw = 0; 5026 } 5027 if (PrintCMSStatistics != 0) { 5028 if (_markStack._hit_limit > 0) { 5029 gclog_or_tty->print_cr(" (benign) Hit max stack size limit ("SIZE_FORMAT")", 5030 _markStack._hit_limit); 5031 } 5032 if (_markStack._failed_double > 0) { 5033 gclog_or_tty->print_cr(" (benign) Failed stack doubling ("SIZE_FORMAT")," 5034 " current capacity "SIZE_FORMAT, 5035 _markStack._failed_double, 5036 _markStack.capacity()); 5037 } 5038 } 5039 _markStack._hit_limit = 0; 5040 _markStack._failed_double = 0; 5041 5042 if ((VerifyAfterGC || VerifyDuringGC) && 5043 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 5044 verify_after_remark(); 5045 } 5046 5047 // Change under the freelistLocks. 5048 _collectorState = Sweeping; 5049 // Call isAllClear() under bitMapLock 5050 assert(_modUnionTable.isAllClear(), 5051 "Should be clear by end of the final marking"); 5052 assert(_ct->klass_rem_set()->mod_union_is_clear(), 5053 "Should be clear by end of the final marking"); 5054 if (UseAdaptiveSizePolicy) { 5055 size_policy()->checkpoint_roots_final_end(gch->gc_cause()); 5056 } 5057 } 5058 5059 // Parallel remark task 5060 class CMSParRemarkTask: public AbstractGangTask { 5061 CMSCollector* _collector; 5062 int _n_workers; 5063 CompactibleFreeListSpace* _cms_space; 5064 5065 // The per-thread work queues, available here for stealing. 5066 OopTaskQueueSet* _task_queues; 5067 ParallelTaskTerminator _term; 5068 5069 public: 5070 // A value of 0 passed to n_workers will cause the number of 5071 // workers to be taken from the active workers in the work gang. 5072 CMSParRemarkTask(CMSCollector* collector, 5073 CompactibleFreeListSpace* cms_space, 5074 int n_workers, FlexibleWorkGang* workers, 5075 OopTaskQueueSet* task_queues): 5076 AbstractGangTask("Rescan roots and grey objects in parallel"), 5077 _collector(collector), 5078 _cms_space(cms_space), 5079 _n_workers(n_workers), 5080 _task_queues(task_queues), 5081 _term(n_workers, task_queues) { } 5082 5083 OopTaskQueueSet* task_queues() { return _task_queues; } 5084 5085 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 5086 5087 ParallelTaskTerminator* terminator() { return &_term; } 5088 int n_workers() { return _n_workers; } 5089 5090 void work(uint worker_id); 5091 5092 private: 5093 // Work method in support of parallel rescan ... of young gen spaces 5094 void do_young_space_rescan(int i, Par_MarkRefsIntoAndScanClosure* cl, 5095 ContiguousSpace* space, 5096 HeapWord** chunk_array, size_t chunk_top); 5097 5098 // ... of dirty cards in old space 5099 void do_dirty_card_rescan_tasks(CompactibleFreeListSpace* sp, int i, 5100 Par_MarkRefsIntoAndScanClosure* cl); 5101 5102 // ... work stealing for the above 5103 void do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, int* seed); 5104 }; 5105 5106 class RemarkKlassClosure : public KlassClosure { 5107 CMKlassClosure _cm_klass_closure; 5108 public: 5109 RemarkKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {} 5110 void do_klass(Klass* k) { 5111 // Check if we have modified any oops in the Klass during the concurrent marking. 5112 if (k->has_accumulated_modified_oops()) { 5113 k->clear_accumulated_modified_oops(); 5114 5115 // We could have transfered the current modified marks to the accumulated marks, 5116 // like we do with the Card Table to Mod Union Table. But it's not really necessary. 5117 } else if (k->has_modified_oops()) { 5118 // Don't clear anything, this info is needed by the next young collection. 5119 } else { 5120 // No modified oops in the Klass. 5121 return; 5122 } 5123 5124 // The klass has modified fields, need to scan the klass. 5125 _cm_klass_closure.do_klass(k); 5126 } 5127 }; 5128 5129 // work_queue(i) is passed to the closure 5130 // Par_MarkRefsIntoAndScanClosure. The "i" parameter 5131 // also is passed to do_dirty_card_rescan_tasks() and to 5132 // do_work_steal() to select the i-th task_queue. 5133 5134 void CMSParRemarkTask::work(uint worker_id) { 5135 elapsedTimer _timer; 5136 ResourceMark rm; 5137 HandleMark hm; 5138 5139 // ---------- rescan from roots -------------- 5140 _timer.start(); 5141 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5142 Par_MarkRefsIntoAndScanClosure par_mrias_cl(_collector, 5143 _collector->_span, _collector->ref_processor(), 5144 &(_collector->_markBitMap), 5145 work_queue(worker_id)); 5146 5147 // Rescan young gen roots first since these are likely 5148 // coarsely partitioned and may, on that account, constitute 5149 // the critical path; thus, it's best to start off that 5150 // work first. 5151 // ---------- young gen roots -------------- 5152 { 5153 DefNewGeneration* dng = _collector->_young_gen->as_DefNewGeneration(); 5154 EdenSpace* eden_space = dng->eden(); 5155 ContiguousSpace* from_space = dng->from(); 5156 ContiguousSpace* to_space = dng->to(); 5157 5158 HeapWord** eca = _collector->_eden_chunk_array; 5159 size_t ect = _collector->_eden_chunk_index; 5160 HeapWord** sca = _collector->_survivor_chunk_array; 5161 size_t sct = _collector->_survivor_chunk_index; 5162 5163 assert(ect <= _collector->_eden_chunk_capacity, "out of bounds"); 5164 assert(sct <= _collector->_survivor_chunk_capacity, "out of bounds"); 5165 5166 do_young_space_rescan(worker_id, &par_mrias_cl, to_space, NULL, 0); 5167 do_young_space_rescan(worker_id, &par_mrias_cl, from_space, sca, sct); 5168 do_young_space_rescan(worker_id, &par_mrias_cl, eden_space, eca, ect); 5169 5170 _timer.stop(); 5171 if (PrintCMSStatistics != 0) { 5172 gclog_or_tty->print_cr( 5173 "Finished young gen rescan work in %dth thread: %3.3f sec", 5174 worker_id, _timer.seconds()); 5175 } 5176 } 5177 5178 // ---------- remaining roots -------------- 5179 _timer.reset(); 5180 _timer.start(); 5181 gch->gen_process_strong_roots(_collector->_cmsGen->level(), 5182 false, // yg was scanned above 5183 false, // this is parallel code 5184 false, // not scavenging 5185 SharedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()), 5186 &par_mrias_cl, 5187 true, // walk all of code cache if (so & SO_CodeCache) 5188 NULL, 5189 NULL, // The dirty klasses will be handled below 5190 worker_id); 5191 5192 assert(_collector->should_unload_classes() 5193 || (_collector->CMSCollector::roots_scanning_options() & SharedHeap::SO_CodeCache), 5194 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 5195 _timer.stop(); 5196 if (PrintCMSStatistics != 0) { 5197 gclog_or_tty->print_cr( 5198 "Finished remaining root rescan work in %dth thread: %3.3f sec", 5199 worker_id, _timer.seconds()); 5200 } 5201 5202 // ---------- unhandled CLD scanning ---------- 5203 if (worker_id == 0) { // Single threaded at the moment. 5204 _timer.reset(); 5205 _timer.start(); 5206 5207 // Scan all new class loader data objects and new dependencies that were 5208 // introduced during concurrent marking. 5209 ResourceMark rm; 5210 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds(); 5211 for (int i = 0; i < array->length(); i++) { 5212 par_mrias_cl.do_class_loader_data(array->at(i)); 5213 } 5214 5215 // We don't need to keep track of new CLDs anymore. 5216 ClassLoaderDataGraph::remember_new_clds(false); 5217 5218 _timer.stop(); 5219 if (PrintCMSStatistics != 0) { 5220 gclog_or_tty->print_cr( 5221 "Finished unhandled CLD scanning work in %dth thread: %3.3f sec", 5222 worker_id, _timer.seconds()); 5223 } 5224 } 5225 5226 // ---------- dirty klass scanning ---------- 5227 if (worker_id == 0) { // Single threaded at the moment. 5228 _timer.reset(); 5229 _timer.start(); 5230 5231 // Scan all classes that was dirtied during the concurrent marking phase. 5232 RemarkKlassClosure remark_klass_closure(&par_mrias_cl); 5233 ClassLoaderDataGraph::classes_do(&remark_klass_closure); 5234 5235 _timer.stop(); 5236 if (PrintCMSStatistics != 0) { 5237 gclog_or_tty->print_cr( 5238 "Finished dirty klass scanning work in %dth thread: %3.3f sec", 5239 worker_id, _timer.seconds()); 5240 } 5241 } 5242 5243 // We might have added oops to ClassLoaderData::_handles during the 5244 // concurrent marking phase. These oops point to newly allocated objects 5245 // that are guaranteed to be kept alive. Either by the direct allocation 5246 // code, or when the young collector processes the strong roots. Hence, 5247 // we don't have to revisit the _handles block during the remark phase. 5248 5249 // ---------- rescan dirty cards ------------ 5250 _timer.reset(); 5251 _timer.start(); 5252 5253 // Do the rescan tasks for each of the two spaces 5254 // (cms_space) in turn. 5255 // "worker_id" is passed to select the task_queue for "worker_id" 5256 do_dirty_card_rescan_tasks(_cms_space, worker_id, &par_mrias_cl); 5257 _timer.stop(); 5258 if (PrintCMSStatistics != 0) { 5259 gclog_or_tty->print_cr( 5260 "Finished dirty card rescan work in %dth thread: %3.3f sec", 5261 worker_id, _timer.seconds()); 5262 } 5263 5264 // ---------- steal work from other threads ... 5265 // ---------- ... and drain overflow list. 5266 _timer.reset(); 5267 _timer.start(); 5268 do_work_steal(worker_id, &par_mrias_cl, _collector->hash_seed(worker_id)); 5269 _timer.stop(); 5270 if (PrintCMSStatistics != 0) { 5271 gclog_or_tty->print_cr( 5272 "Finished work stealing in %dth thread: %3.3f sec", 5273 worker_id, _timer.seconds()); 5274 } 5275 } 5276 5277 // Note that parameter "i" is not used. 5278 void 5279 CMSParRemarkTask::do_young_space_rescan(int i, 5280 Par_MarkRefsIntoAndScanClosure* cl, ContiguousSpace* space, 5281 HeapWord** chunk_array, size_t chunk_top) { 5282 // Until all tasks completed: 5283 // . claim an unclaimed task 5284 // . compute region boundaries corresponding to task claimed 5285 // using chunk_array 5286 // . par_oop_iterate(cl) over that region 5287 5288 ResourceMark rm; 5289 HandleMark hm; 5290 5291 SequentialSubTasksDone* pst = space->par_seq_tasks(); 5292 assert(pst->valid(), "Uninitialized use?"); 5293 5294 uint nth_task = 0; 5295 uint n_tasks = pst->n_tasks(); 5296 5297 HeapWord *start, *end; 5298 while (!pst->is_task_claimed(/* reference */ nth_task)) { 5299 // We claimed task # nth_task; compute its boundaries. 5300 if (chunk_top == 0) { // no samples were taken 5301 assert(nth_task == 0 && n_tasks == 1, "Can have only 1 EdenSpace task"); 5302 start = space->bottom(); 5303 end = space->top(); 5304 } else if (nth_task == 0) { 5305 start = space->bottom(); 5306 end = chunk_array[nth_task]; 5307 } else if (nth_task < (uint)chunk_top) { 5308 assert(nth_task >= 1, "Control point invariant"); 5309 start = chunk_array[nth_task - 1]; 5310 end = chunk_array[nth_task]; 5311 } else { 5312 assert(nth_task == (uint)chunk_top, "Control point invariant"); 5313 start = chunk_array[chunk_top - 1]; 5314 end = space->top(); 5315 } 5316 MemRegion mr(start, end); 5317 // Verify that mr is in space 5318 assert(mr.is_empty() || space->used_region().contains(mr), 5319 "Should be in space"); 5320 // Verify that "start" is an object boundary 5321 assert(mr.is_empty() || oop(mr.start())->is_oop(), 5322 "Should be an oop"); 5323 space->par_oop_iterate(mr, cl); 5324 } 5325 pst->all_tasks_completed(); 5326 } 5327 5328 void 5329 CMSParRemarkTask::do_dirty_card_rescan_tasks( 5330 CompactibleFreeListSpace* sp, int i, 5331 Par_MarkRefsIntoAndScanClosure* cl) { 5332 // Until all tasks completed: 5333 // . claim an unclaimed task 5334 // . compute region boundaries corresponding to task claimed 5335 // . transfer dirty bits ct->mut for that region 5336 // . apply rescanclosure to dirty mut bits for that region 5337 5338 ResourceMark rm; 5339 HandleMark hm; 5340 5341 OopTaskQueue* work_q = work_queue(i); 5342 ModUnionClosure modUnionClosure(&(_collector->_modUnionTable)); 5343 // CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! 5344 // CAUTION: This closure has state that persists across calls to 5345 // the work method dirty_range_iterate_clear() in that it has 5346 // imbedded in it a (subtype of) UpwardsObjectClosure. The 5347 // use of that state in the imbedded UpwardsObjectClosure instance 5348 // assumes that the cards are always iterated (even if in parallel 5349 // by several threads) in monotonically increasing order per each 5350 // thread. This is true of the implementation below which picks 5351 // card ranges (chunks) in monotonically increasing order globally 5352 // and, a-fortiori, in monotonically increasing order per thread 5353 // (the latter order being a subsequence of the former). 5354 // If the work code below is ever reorganized into a more chaotic 5355 // work-partitioning form than the current "sequential tasks" 5356 // paradigm, the use of that persistent state will have to be 5357 // revisited and modified appropriately. See also related 5358 // bug 4756801 work on which should examine this code to make 5359 // sure that the changes there do not run counter to the 5360 // assumptions made here and necessary for correctness and 5361 // efficiency. Note also that this code might yield inefficient 5362 // behaviour in the case of very large objects that span one or 5363 // more work chunks. Such objects would potentially be scanned 5364 // several times redundantly. Work on 4756801 should try and 5365 // address that performance anomaly if at all possible. XXX 5366 MemRegion full_span = _collector->_span; 5367 CMSBitMap* bm = &(_collector->_markBitMap); // shared 5368 MarkFromDirtyCardsClosure 5369 greyRescanClosure(_collector, full_span, // entire span of interest 5370 sp, bm, work_q, cl); 5371 5372 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 5373 assert(pst->valid(), "Uninitialized use?"); 5374 uint nth_task = 0; 5375 const int alignment = CardTableModRefBS::card_size * BitsPerWord; 5376 MemRegion span = sp->used_region(); 5377 HeapWord* start_addr = span.start(); 5378 HeapWord* end_addr = (HeapWord*)round_to((intptr_t)span.end(), 5379 alignment); 5380 const size_t chunk_size = sp->rescan_task_size(); // in HeapWord units 5381 assert((HeapWord*)round_to((intptr_t)start_addr, alignment) == 5382 start_addr, "Check alignment"); 5383 assert((size_t)round_to((intptr_t)chunk_size, alignment) == 5384 chunk_size, "Check alignment"); 5385 5386 while (!pst->is_task_claimed(/* reference */ nth_task)) { 5387 // Having claimed the nth_task, compute corresponding mem-region, 5388 // which is a-fortiori aligned correctly (i.e. at a MUT bopundary). 5389 // The alignment restriction ensures that we do not need any 5390 // synchronization with other gang-workers while setting or 5391 // clearing bits in thus chunk of the MUT. 5392 MemRegion this_span = MemRegion(start_addr + nth_task*chunk_size, 5393 start_addr + (nth_task+1)*chunk_size); 5394 // The last chunk's end might be way beyond end of the 5395 // used region. In that case pull back appropriately. 5396 if (this_span.end() > end_addr) { 5397 this_span.set_end(end_addr); 5398 assert(!this_span.is_empty(), "Program logic (calculation of n_tasks)"); 5399 } 5400 // Iterate over the dirty cards covering this chunk, marking them 5401 // precleaned, and setting the corresponding bits in the mod union 5402 // table. Since we have been careful to partition at Card and MUT-word 5403 // boundaries no synchronization is needed between parallel threads. 5404 _collector->_ct->ct_bs()->dirty_card_iterate(this_span, 5405 &modUnionClosure); 5406 5407 // Having transferred these marks into the modUnionTable, 5408 // rescan the marked objects on the dirty cards in the modUnionTable. 5409 // Even if this is at a synchronous collection, the initial marking 5410 // may have been done during an asynchronous collection so there 5411 // may be dirty bits in the mod-union table. 5412 _collector->_modUnionTable.dirty_range_iterate_clear( 5413 this_span, &greyRescanClosure); 5414 _collector->_modUnionTable.verifyNoOneBitsInRange( 5415 this_span.start(), 5416 this_span.end()); 5417 } 5418 pst->all_tasks_completed(); // declare that i am done 5419 } 5420 5421 // . see if we can share work_queues with ParNew? XXX 5422 void 5423 CMSParRemarkTask::do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, 5424 int* seed) { 5425 OopTaskQueue* work_q = work_queue(i); 5426 NOT_PRODUCT(int num_steals = 0;) 5427 oop obj_to_scan; 5428 CMSBitMap* bm = &(_collector->_markBitMap); 5429 5430 while (true) { 5431 // Completely finish any left over work from (an) earlier round(s) 5432 cl->trim_queue(0); 5433 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 5434 (size_t)ParGCDesiredObjsFromOverflowList); 5435 // Now check if there's any work in the overflow list 5436 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads, 5437 // only affects the number of attempts made to get work from the 5438 // overflow list and does not affect the number of workers. Just 5439 // pass ParallelGCThreads so this behavior is unchanged. 5440 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 5441 work_q, 5442 ParallelGCThreads)) { 5443 // found something in global overflow list; 5444 // not yet ready to go stealing work from others. 5445 // We'd like to assert(work_q->size() != 0, ...) 5446 // because we just took work from the overflow list, 5447 // but of course we can't since all of that could have 5448 // been already stolen from us. 5449 // "He giveth and He taketh away." 5450 continue; 5451 } 5452 // Verify that we have no work before we resort to stealing 5453 assert(work_q->size() == 0, "Have work, shouldn't steal"); 5454 // Try to steal from other queues that have work 5455 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 5456 NOT_PRODUCT(num_steals++;) 5457 assert(obj_to_scan->is_oop(), "Oops, not an oop!"); 5458 assert(bm->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 5459 // Do scanning work 5460 obj_to_scan->oop_iterate(cl); 5461 // Loop around, finish this work, and try to steal some more 5462 } else if (terminator()->offer_termination()) { 5463 break; // nirvana from the infinite cycle 5464 } 5465 } 5466 NOT_PRODUCT( 5467 if (PrintCMSStatistics != 0) { 5468 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals); 5469 } 5470 ) 5471 assert(work_q->size() == 0 && _collector->overflow_list_is_empty(), 5472 "Else our work is not yet done"); 5473 } 5474 5475 // Return a thread-local PLAB recording array, as appropriate. 5476 void* CMSCollector::get_data_recorder(int thr_num) { 5477 if (_survivor_plab_array != NULL && 5478 (CMSPLABRecordAlways || 5479 (_collectorState > Marking && _collectorState < FinalMarking))) { 5480 assert(thr_num < (int)ParallelGCThreads, "thr_num is out of bounds"); 5481 ChunkArray* ca = &_survivor_plab_array[thr_num]; 5482 ca->reset(); // clear it so that fresh data is recorded 5483 return (void*) ca; 5484 } else { 5485 return NULL; 5486 } 5487 } 5488 5489 // Reset all the thread-local PLAB recording arrays 5490 void CMSCollector::reset_survivor_plab_arrays() { 5491 for (uint i = 0; i < ParallelGCThreads; i++) { 5492 _survivor_plab_array[i].reset(); 5493 } 5494 } 5495 5496 // Merge the per-thread plab arrays into the global survivor chunk 5497 // array which will provide the partitioning of the survivor space 5498 // for CMS rescan. 5499 void CMSCollector::merge_survivor_plab_arrays(ContiguousSpace* surv, 5500 int no_of_gc_threads) { 5501 assert(_survivor_plab_array != NULL, "Error"); 5502 assert(_survivor_chunk_array != NULL, "Error"); 5503 assert(_collectorState == FinalMarking, "Error"); 5504 for (int j = 0; j < no_of_gc_threads; j++) { 5505 _cursor[j] = 0; 5506 } 5507 HeapWord* top = surv->top(); 5508 size_t i; 5509 for (i = 0; i < _survivor_chunk_capacity; i++) { // all sca entries 5510 HeapWord* min_val = top; // Higher than any PLAB address 5511 uint min_tid = 0; // position of min_val this round 5512 for (int j = 0; j < no_of_gc_threads; j++) { 5513 ChunkArray* cur_sca = &_survivor_plab_array[j]; 5514 if (_cursor[j] == cur_sca->end()) { 5515 continue; 5516 } 5517 assert(_cursor[j] < cur_sca->end(), "ctl pt invariant"); 5518 HeapWord* cur_val = cur_sca->nth(_cursor[j]); 5519 assert(surv->used_region().contains(cur_val), "Out of bounds value"); 5520 if (cur_val < min_val) { 5521 min_tid = j; 5522 min_val = cur_val; 5523 } else { 5524 assert(cur_val < top, "All recorded addresses should be less"); 5525 } 5526 } 5527 // At this point min_val and min_tid are respectively 5528 // the least address in _survivor_plab_array[j]->nth(_cursor[j]) 5529 // and the thread (j) that witnesses that address. 5530 // We record this address in the _survivor_chunk_array[i] 5531 // and increment _cursor[min_tid] prior to the next round i. 5532 if (min_val == top) { 5533 break; 5534 } 5535 _survivor_chunk_array[i] = min_val; 5536 _cursor[min_tid]++; 5537 } 5538 // We are all done; record the size of the _survivor_chunk_array 5539 _survivor_chunk_index = i; // exclusive: [0, i) 5540 if (PrintCMSStatistics > 0) { 5541 gclog_or_tty->print(" (Survivor:" SIZE_FORMAT "chunks) ", i); 5542 } 5543 // Verify that we used up all the recorded entries 5544 #ifdef ASSERT 5545 size_t total = 0; 5546 for (int j = 0; j < no_of_gc_threads; j++) { 5547 assert(_cursor[j] == _survivor_plab_array[j].end(), "Ctl pt invariant"); 5548 total += _cursor[j]; 5549 } 5550 assert(total == _survivor_chunk_index, "Ctl Pt Invariant"); 5551 // Check that the merged array is in sorted order 5552 if (total > 0) { 5553 for (size_t i = 0; i < total - 1; i++) { 5554 if (PrintCMSStatistics > 0) { 5555 gclog_or_tty->print(" (chunk" SIZE_FORMAT ":" INTPTR_FORMAT ") ", 5556 i, _survivor_chunk_array[i]); 5557 } 5558 assert(_survivor_chunk_array[i] < _survivor_chunk_array[i+1], 5559 "Not sorted"); 5560 } 5561 } 5562 #endif // ASSERT 5563 } 5564 5565 // Set up the space's par_seq_tasks structure for work claiming 5566 // for parallel rescan of young gen. 5567 // See ParRescanTask where this is currently used. 5568 void 5569 CMSCollector:: 5570 initialize_sequential_subtasks_for_young_gen_rescan(int n_threads) { 5571 assert(n_threads > 0, "Unexpected n_threads argument"); 5572 DefNewGeneration* dng = (DefNewGeneration*)_young_gen; 5573 5574 // Eden space 5575 { 5576 SequentialSubTasksDone* pst = dng->eden()->par_seq_tasks(); 5577 assert(!pst->valid(), "Clobbering existing data?"); 5578 // Each valid entry in [0, _eden_chunk_index) represents a task. 5579 size_t n_tasks = _eden_chunk_index + 1; 5580 assert(n_tasks == 1 || _eden_chunk_array != NULL, "Error"); 5581 // Sets the condition for completion of the subtask (how many threads 5582 // need to finish in order to be done). 5583 pst->set_n_threads(n_threads); 5584 pst->set_n_tasks((int)n_tasks); 5585 } 5586 5587 // Merge the survivor plab arrays into _survivor_chunk_array 5588 if (_survivor_plab_array != NULL) { 5589 merge_survivor_plab_arrays(dng->from(), n_threads); 5590 } else { 5591 assert(_survivor_chunk_index == 0, "Error"); 5592 } 5593 5594 // To space 5595 { 5596 SequentialSubTasksDone* pst = dng->to()->par_seq_tasks(); 5597 assert(!pst->valid(), "Clobbering existing data?"); 5598 // Sets the condition for completion of the subtask (how many threads 5599 // need to finish in order to be done). 5600 pst->set_n_threads(n_threads); 5601 pst->set_n_tasks(1); 5602 assert(pst->valid(), "Error"); 5603 } 5604 5605 // From space 5606 { 5607 SequentialSubTasksDone* pst = dng->from()->par_seq_tasks(); 5608 assert(!pst->valid(), "Clobbering existing data?"); 5609 size_t n_tasks = _survivor_chunk_index + 1; 5610 assert(n_tasks == 1 || _survivor_chunk_array != NULL, "Error"); 5611 // Sets the condition for completion of the subtask (how many threads 5612 // need to finish in order to be done). 5613 pst->set_n_threads(n_threads); 5614 pst->set_n_tasks((int)n_tasks); 5615 assert(pst->valid(), "Error"); 5616 } 5617 } 5618 5619 // Parallel version of remark 5620 void CMSCollector::do_remark_parallel() { 5621 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5622 FlexibleWorkGang* workers = gch->workers(); 5623 assert(workers != NULL, "Need parallel worker threads."); 5624 // Choose to use the number of GC workers most recently set 5625 // into "active_workers". If active_workers is not set, set it 5626 // to ParallelGCThreads. 5627 int n_workers = workers->active_workers(); 5628 if (n_workers == 0) { 5629 assert(n_workers > 0, "Should have been set during scavenge"); 5630 n_workers = ParallelGCThreads; 5631 workers->set_active_workers(n_workers); 5632 } 5633 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 5634 5635 CMSParRemarkTask tsk(this, 5636 cms_space, 5637 n_workers, workers, task_queues()); 5638 5639 // Set up for parallel process_strong_roots work. 5640 gch->set_par_threads(n_workers); 5641 // We won't be iterating over the cards in the card table updating 5642 // the younger_gen cards, so we shouldn't call the following else 5643 // the verification code as well as subsequent younger_refs_iterate 5644 // code would get confused. XXX 5645 // gch->rem_set()->prepare_for_younger_refs_iterate(true); // parallel 5646 5647 // The young gen rescan work will not be done as part of 5648 // process_strong_roots (which currently doesn't knw how to 5649 // parallelize such a scan), but rather will be broken up into 5650 // a set of parallel tasks (via the sampling that the [abortable] 5651 // preclean phase did of EdenSpace, plus the [two] tasks of 5652 // scanning the [two] survivor spaces. Further fine-grain 5653 // parallelization of the scanning of the survivor spaces 5654 // themselves, and of precleaning of the younger gen itself 5655 // is deferred to the future. 5656 initialize_sequential_subtasks_for_young_gen_rescan(n_workers); 5657 5658 // The dirty card rescan work is broken up into a "sequence" 5659 // of parallel tasks (per constituent space) that are dynamically 5660 // claimed by the parallel threads. 5661 cms_space->initialize_sequential_subtasks_for_rescan(n_workers); 5662 5663 // It turns out that even when we're using 1 thread, doing the work in a 5664 // separate thread causes wide variance in run times. We can't help this 5665 // in the multi-threaded case, but we special-case n=1 here to get 5666 // repeatable measurements of the 1-thread overhead of the parallel code. 5667 if (n_workers > 1) { 5668 // Make refs discovery MT-safe, if it isn't already: it may not 5669 // necessarily be so, since it's possible that we are doing 5670 // ST marking. 5671 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), true); 5672 GenCollectedHeap::StrongRootsScope srs(gch); 5673 workers->run_task(&tsk); 5674 } else { 5675 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false); 5676 GenCollectedHeap::StrongRootsScope srs(gch); 5677 tsk.work(0); 5678 } 5679 5680 gch->set_par_threads(0); // 0 ==> non-parallel. 5681 // restore, single-threaded for now, any preserved marks 5682 // as a result of work_q overflow 5683 restore_preserved_marks_if_any(); 5684 } 5685 5686 // Non-parallel version of remark 5687 void CMSCollector::do_remark_non_parallel() { 5688 ResourceMark rm; 5689 HandleMark hm; 5690 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5691 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false); 5692 5693 MarkRefsIntoAndScanClosure 5694 mrias_cl(_span, ref_processor(), &_markBitMap, NULL /* not precleaning */, 5695 &_markStack, this, 5696 false /* should_yield */, false /* not precleaning */); 5697 MarkFromDirtyCardsClosure 5698 markFromDirtyCardsClosure(this, _span, 5699 NULL, // space is set further below 5700 &_markBitMap, &_markStack, &mrias_cl); 5701 { 5702 TraceTime t("grey object rescan", PrintGCDetails, false, gclog_or_tty); 5703 // Iterate over the dirty cards, setting the corresponding bits in the 5704 // mod union table. 5705 { 5706 ModUnionClosure modUnionClosure(&_modUnionTable); 5707 _ct->ct_bs()->dirty_card_iterate( 5708 _cmsGen->used_region(), 5709 &modUnionClosure); 5710 } 5711 // Having transferred these marks into the modUnionTable, we just need 5712 // to rescan the marked objects on the dirty cards in the modUnionTable. 5713 // The initial marking may have been done during an asynchronous 5714 // collection so there may be dirty bits in the mod-union table. 5715 const int alignment = 5716 CardTableModRefBS::card_size * BitsPerWord; 5717 { 5718 // ... First handle dirty cards in CMS gen 5719 markFromDirtyCardsClosure.set_space(_cmsGen->cmsSpace()); 5720 MemRegion ur = _cmsGen->used_region(); 5721 HeapWord* lb = ur.start(); 5722 HeapWord* ub = (HeapWord*)round_to((intptr_t)ur.end(), alignment); 5723 MemRegion cms_span(lb, ub); 5724 _modUnionTable.dirty_range_iterate_clear(cms_span, 5725 &markFromDirtyCardsClosure); 5726 verify_work_stacks_empty(); 5727 if (PrintCMSStatistics != 0) { 5728 gclog_or_tty->print(" (re-scanned "SIZE_FORMAT" dirty cards in cms gen) ", 5729 markFromDirtyCardsClosure.num_dirty_cards()); 5730 } 5731 } 5732 } 5733 if (VerifyDuringGC && 5734 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 5735 HandleMark hm; // Discard invalid handles created during verification 5736 Universe::verify(); 5737 } 5738 { 5739 TraceTime t("root rescan", PrintGCDetails, false, gclog_or_tty); 5740 5741 verify_work_stacks_empty(); 5742 5743 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 5744 GenCollectedHeap::StrongRootsScope srs(gch); 5745 gch->gen_process_strong_roots(_cmsGen->level(), 5746 true, // younger gens as roots 5747 false, // use the local StrongRootsScope 5748 false, // not scavenging 5749 SharedHeap::ScanningOption(roots_scanning_options()), 5750 &mrias_cl, 5751 true, // walk code active on stacks 5752 NULL, 5753 NULL); // The dirty klasses will be handled below 5754 5755 assert(should_unload_classes() 5756 || (roots_scanning_options() & SharedHeap::SO_CodeCache), 5757 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 5758 } 5759 5760 { 5761 TraceTime t("visit unhandled CLDs", PrintGCDetails, false, gclog_or_tty); 5762 5763 verify_work_stacks_empty(); 5764 5765 // Scan all class loader data objects that might have been introduced 5766 // during concurrent marking. 5767 ResourceMark rm; 5768 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds(); 5769 for (int i = 0; i < array->length(); i++) { 5770 mrias_cl.do_class_loader_data(array->at(i)); 5771 } 5772 5773 // We don't need to keep track of new CLDs anymore. 5774 ClassLoaderDataGraph::remember_new_clds(false); 5775 5776 verify_work_stacks_empty(); 5777 } 5778 5779 { 5780 TraceTime t("dirty klass scan", PrintGCDetails, false, gclog_or_tty); 5781 5782 verify_work_stacks_empty(); 5783 5784 RemarkKlassClosure remark_klass_closure(&mrias_cl); 5785 ClassLoaderDataGraph::classes_do(&remark_klass_closure); 5786 5787 verify_work_stacks_empty(); 5788 } 5789 5790 // We might have added oops to ClassLoaderData::_handles during the 5791 // concurrent marking phase. These oops point to newly allocated objects 5792 // that are guaranteed to be kept alive. Either by the direct allocation 5793 // code, or when the young collector processes the strong roots. Hence, 5794 // we don't have to revisit the _handles block during the remark phase. 5795 5796 verify_work_stacks_empty(); 5797 // Restore evacuated mark words, if any, used for overflow list links 5798 if (!CMSOverflowEarlyRestoration) { 5799 restore_preserved_marks_if_any(); 5800 } 5801 verify_overflow_empty(); 5802 } 5803 5804 //////////////////////////////////////////////////////// 5805 // Parallel Reference Processing Task Proxy Class 5806 //////////////////////////////////////////////////////// 5807 class CMSRefProcTaskProxy: public AbstractGangTaskWOopQueues { 5808 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask; 5809 CMSCollector* _collector; 5810 CMSBitMap* _mark_bit_map; 5811 const MemRegion _span; 5812 ProcessTask& _task; 5813 5814 public: 5815 CMSRefProcTaskProxy(ProcessTask& task, 5816 CMSCollector* collector, 5817 const MemRegion& span, 5818 CMSBitMap* mark_bit_map, 5819 AbstractWorkGang* workers, 5820 OopTaskQueueSet* task_queues): 5821 // XXX Should superclass AGTWOQ also know about AWG since it knows 5822 // about the task_queues used by the AWG? Then it could initialize 5823 // the terminator() object. See 6984287. The set_for_termination() 5824 // below is a temporary band-aid for the regression in 6984287. 5825 AbstractGangTaskWOopQueues("Process referents by policy in parallel", 5826 task_queues), 5827 _task(task), 5828 _collector(collector), _span(span), _mark_bit_map(mark_bit_map) 5829 { 5830 assert(_collector->_span.equals(_span) && !_span.is_empty(), 5831 "Inconsistency in _span"); 5832 set_for_termination(workers->active_workers()); 5833 } 5834 5835 OopTaskQueueSet* task_queues() { return queues(); } 5836 5837 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 5838 5839 void do_work_steal(int i, 5840 CMSParDrainMarkingStackClosure* drain, 5841 CMSParKeepAliveClosure* keep_alive, 5842 int* seed); 5843 5844 virtual void work(uint worker_id); 5845 }; 5846 5847 void CMSRefProcTaskProxy::work(uint worker_id) { 5848 assert(_collector->_span.equals(_span), "Inconsistency in _span"); 5849 CMSParKeepAliveClosure par_keep_alive(_collector, _span, 5850 _mark_bit_map, 5851 work_queue(worker_id)); 5852 CMSParDrainMarkingStackClosure par_drain_stack(_collector, _span, 5853 _mark_bit_map, 5854 work_queue(worker_id)); 5855 CMSIsAliveClosure is_alive_closure(_span, _mark_bit_map); 5856 _task.work(worker_id, is_alive_closure, par_keep_alive, par_drain_stack); 5857 if (_task.marks_oops_alive()) { 5858 do_work_steal(worker_id, &par_drain_stack, &par_keep_alive, 5859 _collector->hash_seed(worker_id)); 5860 } 5861 assert(work_queue(worker_id)->size() == 0, "work_queue should be empty"); 5862 assert(_collector->_overflow_list == NULL, "non-empty _overflow_list"); 5863 } 5864 5865 class CMSRefEnqueueTaskProxy: public AbstractGangTask { 5866 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask; 5867 EnqueueTask& _task; 5868 5869 public: 5870 CMSRefEnqueueTaskProxy(EnqueueTask& task) 5871 : AbstractGangTask("Enqueue reference objects in parallel"), 5872 _task(task) 5873 { } 5874 5875 virtual void work(uint worker_id) 5876 { 5877 _task.work(worker_id); 5878 } 5879 }; 5880 5881 CMSParKeepAliveClosure::CMSParKeepAliveClosure(CMSCollector* collector, 5882 MemRegion span, CMSBitMap* bit_map, OopTaskQueue* work_queue): 5883 _span(span), 5884 _bit_map(bit_map), 5885 _work_queue(work_queue), 5886 _mark_and_push(collector, span, bit_map, work_queue), 5887 _low_water_mark(MIN2((uint)(work_queue->max_elems()/4), 5888 (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads))) 5889 { } 5890 5891 // . see if we can share work_queues with ParNew? XXX 5892 void CMSRefProcTaskProxy::do_work_steal(int i, 5893 CMSParDrainMarkingStackClosure* drain, 5894 CMSParKeepAliveClosure* keep_alive, 5895 int* seed) { 5896 OopTaskQueue* work_q = work_queue(i); 5897 NOT_PRODUCT(int num_steals = 0;) 5898 oop obj_to_scan; 5899 5900 while (true) { 5901 // Completely finish any left over work from (an) earlier round(s) 5902 drain->trim_queue(0); 5903 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 5904 (size_t)ParGCDesiredObjsFromOverflowList); 5905 // Now check if there's any work in the overflow list 5906 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads, 5907 // only affects the number of attempts made to get work from the 5908 // overflow list and does not affect the number of workers. Just 5909 // pass ParallelGCThreads so this behavior is unchanged. 5910 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 5911 work_q, 5912 ParallelGCThreads)) { 5913 // Found something in global overflow list; 5914 // not yet ready to go stealing work from others. 5915 // We'd like to assert(work_q->size() != 0, ...) 5916 // because we just took work from the overflow list, 5917 // but of course we can't, since all of that might have 5918 // been already stolen from us. 5919 continue; 5920 } 5921 // Verify that we have no work before we resort to stealing 5922 assert(work_q->size() == 0, "Have work, shouldn't steal"); 5923 // Try to steal from other queues that have work 5924 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 5925 NOT_PRODUCT(num_steals++;) 5926 assert(obj_to_scan->is_oop(), "Oops, not an oop!"); 5927 assert(_mark_bit_map->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 5928 // Do scanning work 5929 obj_to_scan->oop_iterate(keep_alive); 5930 // Loop around, finish this work, and try to steal some more 5931 } else if (terminator()->offer_termination()) { 5932 break; // nirvana from the infinite cycle 5933 } 5934 } 5935 NOT_PRODUCT( 5936 if (PrintCMSStatistics != 0) { 5937 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals); 5938 } 5939 ) 5940 } 5941 5942 void CMSRefProcTaskExecutor::execute(ProcessTask& task) 5943 { 5944 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5945 FlexibleWorkGang* workers = gch->workers(); 5946 assert(workers != NULL, "Need parallel worker threads."); 5947 CMSRefProcTaskProxy rp_task(task, &_collector, 5948 _collector.ref_processor()->span(), 5949 _collector.markBitMap(), 5950 workers, _collector.task_queues()); 5951 workers->run_task(&rp_task); 5952 } 5953 5954 void CMSRefProcTaskExecutor::execute(EnqueueTask& task) 5955 { 5956 5957 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5958 FlexibleWorkGang* workers = gch->workers(); 5959 assert(workers != NULL, "Need parallel worker threads."); 5960 CMSRefEnqueueTaskProxy enq_task(task); 5961 workers->run_task(&enq_task); 5962 } 5963 5964 void CMSCollector::refProcessingWork(bool asynch, bool clear_all_soft_refs) { 5965 5966 ResourceMark rm; 5967 HandleMark hm; 5968 5969 ReferenceProcessor* rp = ref_processor(); 5970 assert(rp->span().equals(_span), "Spans should be equal"); 5971 assert(!rp->enqueuing_is_done(), "Enqueuing should not be complete"); 5972 // Process weak references. 5973 rp->setup_policy(clear_all_soft_refs); 5974 verify_work_stacks_empty(); 5975 5976 CMSKeepAliveClosure cmsKeepAliveClosure(this, _span, &_markBitMap, 5977 &_markStack, false /* !preclean */); 5978 CMSDrainMarkingStackClosure cmsDrainMarkingStackClosure(this, 5979 _span, &_markBitMap, &_markStack, 5980 &cmsKeepAliveClosure, false /* !preclean */); 5981 { 5982 TraceTime t("weak refs processing", PrintGCDetails, false, gclog_or_tty); 5983 if (rp->processing_is_mt()) { 5984 // Set the degree of MT here. If the discovery is done MT, there 5985 // may have been a different number of threads doing the discovery 5986 // and a different number of discovered lists may have Ref objects. 5987 // That is OK as long as the Reference lists are balanced (see 5988 // balance_all_queues() and balance_queues()). 5989 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5990 int active_workers = ParallelGCThreads; 5991 FlexibleWorkGang* workers = gch->workers(); 5992 if (workers != NULL) { 5993 active_workers = workers->active_workers(); 5994 // The expectation is that active_workers will have already 5995 // been set to a reasonable value. If it has not been set, 5996 // investigate. 5997 assert(active_workers > 0, "Should have been set during scavenge"); 5998 } 5999 rp->set_active_mt_degree(active_workers); 6000 CMSRefProcTaskExecutor task_executor(*this); 6001 rp->process_discovered_references(&_is_alive_closure, 6002 &cmsKeepAliveClosure, 6003 &cmsDrainMarkingStackClosure, 6004 &task_executor); 6005 } else { 6006 rp->process_discovered_references(&_is_alive_closure, 6007 &cmsKeepAliveClosure, 6008 &cmsDrainMarkingStackClosure, 6009 NULL); 6010 } 6011 } 6012 6013 // This is the point where the entire marking should have completed. 6014 verify_work_stacks_empty(); 6015 6016 if (should_unload_classes()) { 6017 { 6018 TraceTime t("class unloading", PrintGCDetails, false, gclog_or_tty); 6019 6020 // Unload classes and purge the SystemDictionary. 6021 bool purged_class = SystemDictionary::do_unloading(&_is_alive_closure); 6022 6023 // Unload nmethods. 6024 CodeCache::do_unloading(&_is_alive_closure, purged_class); 6025 6026 // Prune dead klasses from subklass/sibling/implementor lists. 6027 Klass::clean_weak_klass_links(&_is_alive_closure); 6028 } 6029 6030 { 6031 TraceTime t("scrub symbol table", PrintGCDetails, false, gclog_or_tty); 6032 // Clean up unreferenced symbols in symbol table. 6033 SymbolTable::unlink(); 6034 } 6035 } 6036 6037 // CMS doesn't use the StringTable as hard roots when class unloading is turned off. 6038 // Need to check if we really scanned the StringTable. 6039 if ((roots_scanning_options() & SharedHeap::SO_Strings) == 0) { 6040 TraceTime t("scrub string table", PrintGCDetails, false, gclog_or_tty); 6041 // Delete entries for dead interned strings. 6042 StringTable::unlink(&_is_alive_closure); 6043 } 6044 6045 // Restore any preserved marks as a result of mark stack or 6046 // work queue overflow 6047 restore_preserved_marks_if_any(); // done single-threaded for now 6048 6049 rp->set_enqueuing_is_done(true); 6050 if (rp->processing_is_mt()) { 6051 rp->balance_all_queues(); 6052 CMSRefProcTaskExecutor task_executor(*this); 6053 rp->enqueue_discovered_references(&task_executor); 6054 } else { 6055 rp->enqueue_discovered_references(NULL); 6056 } 6057 rp->verify_no_references_recorded(); 6058 assert(!rp->discovery_enabled(), "should have been disabled"); 6059 } 6060 6061 #ifndef PRODUCT 6062 void CMSCollector::check_correct_thread_executing() { 6063 Thread* t = Thread::current(); 6064 // Only the VM thread or the CMS thread should be here. 6065 assert(t->is_ConcurrentGC_thread() || t->is_VM_thread(), 6066 "Unexpected thread type"); 6067 // If this is the vm thread, the foreground process 6068 // should not be waiting. Note that _foregroundGCIsActive is 6069 // true while the foreground collector is waiting. 6070 if (_foregroundGCShouldWait) { 6071 // We cannot be the VM thread 6072 assert(t->is_ConcurrentGC_thread(), 6073 "Should be CMS thread"); 6074 } else { 6075 // We can be the CMS thread only if we are in a stop-world 6076 // phase of CMS collection. 6077 if (t->is_ConcurrentGC_thread()) { 6078 assert(_collectorState == InitialMarking || 6079 _collectorState == FinalMarking, 6080 "Should be a stop-world phase"); 6081 // The CMS thread should be holding the CMS_token. 6082 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6083 "Potential interference with concurrently " 6084 "executing VM thread"); 6085 } 6086 } 6087 } 6088 #endif 6089 6090 void CMSCollector::sweep(bool asynch) { 6091 assert(_collectorState == Sweeping, "just checking"); 6092 check_correct_thread_executing(); 6093 verify_work_stacks_empty(); 6094 verify_overflow_empty(); 6095 increment_sweep_count(); 6096 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 6097 6098 _inter_sweep_timer.stop(); 6099 _inter_sweep_estimate.sample(_inter_sweep_timer.seconds()); 6100 size_policy()->avg_cms_free_at_sweep()->sample(_cmsGen->free()); 6101 6102 assert(!_intra_sweep_timer.is_active(), "Should not be active"); 6103 _intra_sweep_timer.reset(); 6104 _intra_sweep_timer.start(); 6105 if (asynch) { 6106 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 6107 CMSPhaseAccounting pa(this, "sweep", !PrintGCDetails); 6108 // First sweep the old gen 6109 { 6110 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(), 6111 bitMapLock()); 6112 sweepWork(_cmsGen, asynch); 6113 } 6114 6115 // Update Universe::_heap_*_at_gc figures. 6116 // We need all the free list locks to make the abstract state 6117 // transition from Sweeping to Resetting. See detailed note 6118 // further below. 6119 { 6120 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock()); 6121 // Update heap occupancy information which is used as 6122 // input to soft ref clearing policy at the next gc. 6123 Universe::update_heap_info_at_gc(); 6124 _collectorState = Resizing; 6125 } 6126 } else { 6127 // already have needed locks 6128 sweepWork(_cmsGen, asynch); 6129 // Update heap occupancy information which is used as 6130 // input to soft ref clearing policy at the next gc. 6131 Universe::update_heap_info_at_gc(); 6132 _collectorState = Resizing; 6133 } 6134 verify_work_stacks_empty(); 6135 verify_overflow_empty(); 6136 6137 if (should_unload_classes()) { 6138 ClassLoaderDataGraph::purge(); 6139 } 6140 6141 _intra_sweep_timer.stop(); 6142 _intra_sweep_estimate.sample(_intra_sweep_timer.seconds()); 6143 6144 _inter_sweep_timer.reset(); 6145 _inter_sweep_timer.start(); 6146 6147 // We need to use a monotonically non-deccreasing time in ms 6148 // or we will see time-warp warnings and os::javaTimeMillis() 6149 // does not guarantee monotonicity. 6150 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC; 6151 update_time_of_last_gc(now); 6152 6153 // NOTE on abstract state transitions: 6154 // Mutators allocate-live and/or mark the mod-union table dirty 6155 // based on the state of the collection. The former is done in 6156 // the interval [Marking, Sweeping] and the latter in the interval 6157 // [Marking, Sweeping). Thus the transitions into the Marking state 6158 // and out of the Sweeping state must be synchronously visible 6159 // globally to the mutators. 6160 // The transition into the Marking state happens with the world 6161 // stopped so the mutators will globally see it. Sweeping is 6162 // done asynchronously by the background collector so the transition 6163 // from the Sweeping state to the Resizing state must be done 6164 // under the freelistLock (as is the check for whether to 6165 // allocate-live and whether to dirty the mod-union table). 6166 assert(_collectorState == Resizing, "Change of collector state to" 6167 " Resizing must be done under the freelistLocks (plural)"); 6168 6169 // Now that sweeping has been completed, we clear 6170 // the incremental_collection_failed flag, 6171 // thus inviting a younger gen collection to promote into 6172 // this generation. If such a promotion may still fail, 6173 // the flag will be set again when a young collection is 6174 // attempted. 6175 GenCollectedHeap* gch = GenCollectedHeap::heap(); 6176 gch->clear_incremental_collection_failed(); // Worth retrying as fresh space may have been freed up 6177 gch->update_full_collections_completed(_collection_count_start); 6178 } 6179 6180 // FIX ME!!! Looks like this belongs in CFLSpace, with 6181 // CMSGen merely delegating to it. 6182 void ConcurrentMarkSweepGeneration::setNearLargestChunk() { 6183 double nearLargestPercent = FLSLargestBlockCoalesceProximity; 6184 HeapWord* minAddr = _cmsSpace->bottom(); 6185 HeapWord* largestAddr = 6186 (HeapWord*) _cmsSpace->dictionary()->find_largest_dict(); 6187 if (largestAddr == NULL) { 6188 // The dictionary appears to be empty. In this case 6189 // try to coalesce at the end of the heap. 6190 largestAddr = _cmsSpace->end(); 6191 } 6192 size_t largestOffset = pointer_delta(largestAddr, minAddr); 6193 size_t nearLargestOffset = 6194 (size_t)((double)largestOffset * nearLargestPercent) - MinChunkSize; 6195 if (PrintFLSStatistics != 0) { 6196 gclog_or_tty->print_cr( 6197 "CMS: Large Block: " PTR_FORMAT ";" 6198 " Proximity: " PTR_FORMAT " -> " PTR_FORMAT, 6199 largestAddr, 6200 _cmsSpace->nearLargestChunk(), minAddr + nearLargestOffset); 6201 } 6202 _cmsSpace->set_nearLargestChunk(minAddr + nearLargestOffset); 6203 } 6204 6205 bool ConcurrentMarkSweepGeneration::isNearLargestChunk(HeapWord* addr) { 6206 return addr >= _cmsSpace->nearLargestChunk(); 6207 } 6208 6209 FreeChunk* ConcurrentMarkSweepGeneration::find_chunk_at_end() { 6210 return _cmsSpace->find_chunk_at_end(); 6211 } 6212 6213 void ConcurrentMarkSweepGeneration::update_gc_stats(int current_level, 6214 bool full) { 6215 // The next lower level has been collected. Gather any statistics 6216 // that are of interest at this point. 6217 if (!full && (current_level + 1) == level()) { 6218 // Gather statistics on the young generation collection. 6219 collector()->stats().record_gc0_end(used()); 6220 } 6221 } 6222 6223 CMSAdaptiveSizePolicy* ConcurrentMarkSweepGeneration::size_policy() { 6224 GenCollectedHeap* gch = GenCollectedHeap::heap(); 6225 assert(gch->kind() == CollectedHeap::GenCollectedHeap, 6226 "Wrong type of heap"); 6227 CMSAdaptiveSizePolicy* sp = (CMSAdaptiveSizePolicy*) 6228 gch->gen_policy()->size_policy(); 6229 assert(sp->is_gc_cms_adaptive_size_policy(), 6230 "Wrong type of size policy"); 6231 return sp; 6232 } 6233 6234 void ConcurrentMarkSweepGeneration::rotate_debug_collection_type() { 6235 if (PrintGCDetails && Verbose) { 6236 gclog_or_tty->print("Rotate from %d ", _debug_collection_type); 6237 } 6238 _debug_collection_type = (CollectionTypes) (_debug_collection_type + 1); 6239 _debug_collection_type = 6240 (CollectionTypes) (_debug_collection_type % Unknown_collection_type); 6241 if (PrintGCDetails && Verbose) { 6242 gclog_or_tty->print_cr("to %d ", _debug_collection_type); 6243 } 6244 } 6245 6246 void CMSCollector::sweepWork(ConcurrentMarkSweepGeneration* gen, 6247 bool asynch) { 6248 // We iterate over the space(s) underlying this generation, 6249 // checking the mark bit map to see if the bits corresponding 6250 // to specific blocks are marked or not. Blocks that are 6251 // marked are live and are not swept up. All remaining blocks 6252 // are swept up, with coalescing on-the-fly as we sweep up 6253 // contiguous free and/or garbage blocks: 6254 // We need to ensure that the sweeper synchronizes with allocators 6255 // and stop-the-world collectors. In particular, the following 6256 // locks are used: 6257 // . CMS token: if this is held, a stop the world collection cannot occur 6258 // . freelistLock: if this is held no allocation can occur from this 6259 // generation by another thread 6260 // . bitMapLock: if this is held, no other thread can access or update 6261 // 6262 6263 // Note that we need to hold the freelistLock if we use 6264 // block iterate below; else the iterator might go awry if 6265 // a mutator (or promotion) causes block contents to change 6266 // (for instance if the allocator divvies up a block). 6267 // If we hold the free list lock, for all practical purposes 6268 // young generation GC's can't occur (they'll usually need to 6269 // promote), so we might as well prevent all young generation 6270 // GC's while we do a sweeping step. For the same reason, we might 6271 // as well take the bit map lock for the entire duration 6272 6273 // check that we hold the requisite locks 6274 assert(have_cms_token(), "Should hold cms token"); 6275 assert( (asynch && ConcurrentMarkSweepThread::cms_thread_has_cms_token()) 6276 || (!asynch && ConcurrentMarkSweepThread::vm_thread_has_cms_token()), 6277 "Should possess CMS token to sweep"); 6278 assert_lock_strong(gen->freelistLock()); 6279 assert_lock_strong(bitMapLock()); 6280 6281 assert(!_inter_sweep_timer.is_active(), "Was switched off in an outer context"); 6282 assert(_intra_sweep_timer.is_active(), "Was switched on in an outer context"); 6283 gen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()), 6284 _inter_sweep_estimate.padded_average(), 6285 _intra_sweep_estimate.padded_average()); 6286 gen->setNearLargestChunk(); 6287 6288 { 6289 SweepClosure sweepClosure(this, gen, &_markBitMap, 6290 CMSYield && asynch); 6291 gen->cmsSpace()->blk_iterate_careful(&sweepClosure); 6292 // We need to free-up/coalesce garbage/blocks from a 6293 // co-terminal free run. This is done in the SweepClosure 6294 // destructor; so, do not remove this scope, else the 6295 // end-of-sweep-census below will be off by a little bit. 6296 } 6297 gen->cmsSpace()->sweep_completed(); 6298 gen->cmsSpace()->endSweepFLCensus(sweep_count()); 6299 if (should_unload_classes()) { // unloaded classes this cycle, 6300 _concurrent_cycles_since_last_unload = 0; // ... reset count 6301 } else { // did not unload classes, 6302 _concurrent_cycles_since_last_unload++; // ... increment count 6303 } 6304 } 6305 6306 // Reset CMS data structures (for now just the marking bit map) 6307 // preparatory for the next cycle. 6308 void CMSCollector::reset(bool asynch) { 6309 GenCollectedHeap* gch = GenCollectedHeap::heap(); 6310 CMSAdaptiveSizePolicy* sp = size_policy(); 6311 AdaptiveSizePolicyOutput(sp, gch->total_collections()); 6312 if (asynch) { 6313 CMSTokenSyncWithLocks ts(true, bitMapLock()); 6314 6315 // If the state is not "Resetting", the foreground thread 6316 // has done a collection and the resetting. 6317 if (_collectorState != Resetting) { 6318 assert(_collectorState == Idling, "The state should only change" 6319 " because the foreground collector has finished the collection"); 6320 return; 6321 } 6322 6323 // Clear the mark bitmap (no grey objects to start with) 6324 // for the next cycle. 6325 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 6326 CMSPhaseAccounting cmspa(this, "reset", !PrintGCDetails); 6327 6328 HeapWord* curAddr = _markBitMap.startWord(); 6329 while (curAddr < _markBitMap.endWord()) { 6330 size_t remaining = pointer_delta(_markBitMap.endWord(), curAddr); 6331 MemRegion chunk(curAddr, MIN2(CMSBitMapYieldQuantum, remaining)); 6332 _markBitMap.clear_large_range(chunk); 6333 if (ConcurrentMarkSweepThread::should_yield() && 6334 !foregroundGCIsActive() && 6335 CMSYield) { 6336 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6337 "CMS thread should hold CMS token"); 6338 assert_lock_strong(bitMapLock()); 6339 bitMapLock()->unlock(); 6340 ConcurrentMarkSweepThread::desynchronize(true); 6341 ConcurrentMarkSweepThread::acknowledge_yield_request(); 6342 stopTimer(); 6343 if (PrintCMSStatistics != 0) { 6344 incrementYields(); 6345 } 6346 icms_wait(); 6347 6348 // See the comment in coordinator_yield() 6349 for (unsigned i = 0; i < CMSYieldSleepCount && 6350 ConcurrentMarkSweepThread::should_yield() && 6351 !CMSCollector::foregroundGCIsActive(); ++i) { 6352 os::sleep(Thread::current(), 1, false); 6353 ConcurrentMarkSweepThread::acknowledge_yield_request(); 6354 } 6355 6356 ConcurrentMarkSweepThread::synchronize(true); 6357 bitMapLock()->lock_without_safepoint_check(); 6358 startTimer(); 6359 } 6360 curAddr = chunk.end(); 6361 } 6362 // A successful mostly concurrent collection has been done. 6363 // Because only the full (i.e., concurrent mode failure) collections 6364 // are being measured for gc overhead limits, clean the "near" flag 6365 // and count. 6366 sp->reset_gc_overhead_limit_count(); 6367 _collectorState = Idling; 6368 } else { 6369 // already have the lock 6370 assert(_collectorState == Resetting, "just checking"); 6371 assert_lock_strong(bitMapLock()); 6372 _markBitMap.clear_all(); 6373 _collectorState = Idling; 6374 } 6375 6376 // Stop incremental mode after a cycle completes, so that any future cycles 6377 // are triggered by allocation. 6378 stop_icms(); 6379 6380 NOT_PRODUCT( 6381 if (RotateCMSCollectionTypes) { 6382 _cmsGen->rotate_debug_collection_type(); 6383 } 6384 ) 6385 } 6386 6387 void CMSCollector::do_CMS_operation(CMS_op_type op, GCCause::Cause gc_cause) { 6388 gclog_or_tty->date_stamp(PrintGC && PrintGCDateStamps); 6389 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 6390 TraceTime t(GCCauseString("GC", gc_cause), PrintGC, !PrintGCDetails, gclog_or_tty); 6391 TraceCollectorStats tcs(counters()); 6392 6393 switch (op) { 6394 case CMS_op_checkpointRootsInitial: { 6395 SvcGCMarker sgcm(SvcGCMarker::OTHER); 6396 checkpointRootsInitial(true); // asynch 6397 if (PrintGC) { 6398 _cmsGen->printOccupancy("initial-mark"); 6399 } 6400 break; 6401 } 6402 case CMS_op_checkpointRootsFinal: { 6403 SvcGCMarker sgcm(SvcGCMarker::OTHER); 6404 checkpointRootsFinal(true, // asynch 6405 false, // !clear_all_soft_refs 6406 false); // !init_mark_was_synchronous 6407 if (PrintGC) { 6408 _cmsGen->printOccupancy("remark"); 6409 } 6410 break; 6411 } 6412 default: 6413 fatal("No such CMS_op"); 6414 } 6415 } 6416 6417 #ifndef PRODUCT 6418 size_t const CMSCollector::skip_header_HeapWords() { 6419 return FreeChunk::header_size(); 6420 } 6421 6422 // Try and collect here conditions that should hold when 6423 // CMS thread is exiting. The idea is that the foreground GC 6424 // thread should not be blocked if it wants to terminate 6425 // the CMS thread and yet continue to run the VM for a while 6426 // after that. 6427 void CMSCollector::verify_ok_to_terminate() const { 6428 assert(Thread::current()->is_ConcurrentGC_thread(), 6429 "should be called by CMS thread"); 6430 assert(!_foregroundGCShouldWait, "should be false"); 6431 // We could check here that all the various low-level locks 6432 // are not held by the CMS thread, but that is overkill; see 6433 // also CMSThread::verify_ok_to_terminate() where the CGC_lock 6434 // is checked. 6435 } 6436 #endif 6437 6438 size_t CMSCollector::block_size_using_printezis_bits(HeapWord* addr) const { 6439 assert(_markBitMap.isMarked(addr) && _markBitMap.isMarked(addr + 1), 6440 "missing Printezis mark?"); 6441 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 6442 size_t size = pointer_delta(nextOneAddr + 1, addr); 6443 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 6444 "alignment problem"); 6445 assert(size >= 3, "Necessary for Printezis marks to work"); 6446 return size; 6447 } 6448 6449 // A variant of the above (block_size_using_printezis_bits()) except 6450 // that we return 0 if the P-bits are not yet set. 6451 size_t CMSCollector::block_size_if_printezis_bits(HeapWord* addr) const { 6452 if (_markBitMap.isMarked(addr + 1)) { 6453 assert(_markBitMap.isMarked(addr), "P-bit can be set only for marked objects"); 6454 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 6455 size_t size = pointer_delta(nextOneAddr + 1, addr); 6456 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 6457 "alignment problem"); 6458 assert(size >= 3, "Necessary for Printezis marks to work"); 6459 return size; 6460 } 6461 return 0; 6462 } 6463 6464 HeapWord* CMSCollector::next_card_start_after_block(HeapWord* addr) const { 6465 size_t sz = 0; 6466 oop p = (oop)addr; 6467 if (p->klass_or_null() != NULL) { 6468 sz = CompactibleFreeListSpace::adjustObjectSize(p->size()); 6469 } else { 6470 sz = block_size_using_printezis_bits(addr); 6471 } 6472 assert(sz > 0, "size must be nonzero"); 6473 HeapWord* next_block = addr + sz; 6474 HeapWord* next_card = (HeapWord*)round_to((uintptr_t)next_block, 6475 CardTableModRefBS::card_size); 6476 assert(round_down((uintptr_t)addr, CardTableModRefBS::card_size) < 6477 round_down((uintptr_t)next_card, CardTableModRefBS::card_size), 6478 "must be different cards"); 6479 return next_card; 6480 } 6481 6482 6483 // CMS Bit Map Wrapper ///////////////////////////////////////// 6484 6485 // Construct a CMS bit map infrastructure, but don't create the 6486 // bit vector itself. That is done by a separate call CMSBitMap::allocate() 6487 // further below. 6488 CMSBitMap::CMSBitMap(int shifter, int mutex_rank, const char* mutex_name): 6489 _bm(), 6490 _shifter(shifter), 6491 _lock(mutex_rank >= 0 ? new Mutex(mutex_rank, mutex_name, true) : NULL) 6492 { 6493 _bmStartWord = 0; 6494 _bmWordSize = 0; 6495 } 6496 6497 bool CMSBitMap::allocate(MemRegion mr) { 6498 _bmStartWord = mr.start(); 6499 _bmWordSize = mr.word_size(); 6500 ReservedSpace brs(ReservedSpace::allocation_align_size_up( 6501 (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1)); 6502 if (!brs.is_reserved()) { 6503 warning("CMS bit map allocation failure"); 6504 return false; 6505 } 6506 // For now we'll just commit all of the bit map up fromt. 6507 // Later on we'll try to be more parsimonious with swap. 6508 if (!_virtual_space.initialize(brs, brs.size())) { 6509 warning("CMS bit map backing store failure"); 6510 return false; 6511 } 6512 assert(_virtual_space.committed_size() == brs.size(), 6513 "didn't reserve backing store for all of CMS bit map?"); 6514 _bm.set_map((BitMap::bm_word_t*)_virtual_space.low()); 6515 assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >= 6516 _bmWordSize, "inconsistency in bit map sizing"); 6517 _bm.set_size(_bmWordSize >> _shifter); 6518 6519 // bm.clear(); // can we rely on getting zero'd memory? verify below 6520 assert(isAllClear(), 6521 "Expected zero'd memory from ReservedSpace constructor"); 6522 assert(_bm.size() == heapWordDiffToOffsetDiff(sizeInWords()), 6523 "consistency check"); 6524 return true; 6525 } 6526 6527 void CMSBitMap::dirty_range_iterate_clear(MemRegion mr, MemRegionClosure* cl) { 6528 HeapWord *next_addr, *end_addr, *last_addr; 6529 assert_locked(); 6530 assert(covers(mr), "out-of-range error"); 6531 // XXX assert that start and end are appropriately aligned 6532 for (next_addr = mr.start(), end_addr = mr.end(); 6533 next_addr < end_addr; next_addr = last_addr) { 6534 MemRegion dirty_region = getAndClearMarkedRegion(next_addr, end_addr); 6535 last_addr = dirty_region.end(); 6536 if (!dirty_region.is_empty()) { 6537 cl->do_MemRegion(dirty_region); 6538 } else { 6539 assert(last_addr == end_addr, "program logic"); 6540 return; 6541 } 6542 } 6543 } 6544 6545 void CMSBitMap::print_on_error(outputStream* st, const char* prefix) const { 6546 _bm.print_on_error(st, prefix); 6547 } 6548 6549 #ifndef PRODUCT 6550 void CMSBitMap::assert_locked() const { 6551 CMSLockVerifier::assert_locked(lock()); 6552 } 6553 6554 bool CMSBitMap::covers(MemRegion mr) const { 6555 // assert(_bm.map() == _virtual_space.low(), "map inconsistency"); 6556 assert((size_t)_bm.size() == (_bmWordSize >> _shifter), 6557 "size inconsistency"); 6558 return (mr.start() >= _bmStartWord) && 6559 (mr.end() <= endWord()); 6560 } 6561 6562 bool CMSBitMap::covers(HeapWord* start, size_t size) const { 6563 return (start >= _bmStartWord && (start + size) <= endWord()); 6564 } 6565 6566 void CMSBitMap::verifyNoOneBitsInRange(HeapWord* left, HeapWord* right) { 6567 // verify that there are no 1 bits in the interval [left, right) 6568 FalseBitMapClosure falseBitMapClosure; 6569 iterate(&falseBitMapClosure, left, right); 6570 } 6571 6572 void CMSBitMap::region_invariant(MemRegion mr) 6573 { 6574 assert_locked(); 6575 // mr = mr.intersection(MemRegion(_bmStartWord, _bmWordSize)); 6576 assert(!mr.is_empty(), "unexpected empty region"); 6577 assert(covers(mr), "mr should be covered by bit map"); 6578 // convert address range into offset range 6579 size_t start_ofs = heapWordToOffset(mr.start()); 6580 // Make sure that end() is appropriately aligned 6581 assert(mr.end() == (HeapWord*)round_to((intptr_t)mr.end(), 6582 (1 << (_shifter+LogHeapWordSize))), 6583 "Misaligned mr.end()"); 6584 size_t end_ofs = heapWordToOffset(mr.end()); 6585 assert(end_ofs > start_ofs, "Should mark at least one bit"); 6586 } 6587 6588 #endif 6589 6590 bool CMSMarkStack::allocate(size_t size) { 6591 // allocate a stack of the requisite depth 6592 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 6593 size * sizeof(oop))); 6594 if (!rs.is_reserved()) { 6595 warning("CMSMarkStack allocation failure"); 6596 return false; 6597 } 6598 if (!_virtual_space.initialize(rs, rs.size())) { 6599 warning("CMSMarkStack backing store failure"); 6600 return false; 6601 } 6602 assert(_virtual_space.committed_size() == rs.size(), 6603 "didn't reserve backing store for all of CMS stack?"); 6604 _base = (oop*)(_virtual_space.low()); 6605 _index = 0; 6606 _capacity = size; 6607 NOT_PRODUCT(_max_depth = 0); 6608 return true; 6609 } 6610 6611 // XXX FIX ME !!! In the MT case we come in here holding a 6612 // leaf lock. For printing we need to take a further lock 6613 // which has lower rank. We need to recallibrate the two 6614 // lock-ranks involved in order to be able to rpint the 6615 // messages below. (Or defer the printing to the caller. 6616 // For now we take the expedient path of just disabling the 6617 // messages for the problematic case.) 6618 void CMSMarkStack::expand() { 6619 assert(_capacity <= MarkStackSizeMax, "stack bigger than permitted"); 6620 if (_capacity == MarkStackSizeMax) { 6621 if (_hit_limit++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) { 6622 // We print a warning message only once per CMS cycle. 6623 gclog_or_tty->print_cr(" (benign) Hit CMSMarkStack max size limit"); 6624 } 6625 return; 6626 } 6627 // Double capacity if possible 6628 size_t new_capacity = MIN2(_capacity*2, MarkStackSizeMax); 6629 // Do not give up existing stack until we have managed to 6630 // get the double capacity that we desired. 6631 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 6632 new_capacity * sizeof(oop))); 6633 if (rs.is_reserved()) { 6634 // Release the backing store associated with old stack 6635 _virtual_space.release(); 6636 // Reinitialize virtual space for new stack 6637 if (!_virtual_space.initialize(rs, rs.size())) { 6638 fatal("Not enough swap for expanded marking stack"); 6639 } 6640 _base = (oop*)(_virtual_space.low()); 6641 _index = 0; 6642 _capacity = new_capacity; 6643 } else if (_failed_double++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) { 6644 // Failed to double capacity, continue; 6645 // we print a detail message only once per CMS cycle. 6646 gclog_or_tty->print(" (benign) Failed to expand marking stack from "SIZE_FORMAT"K to " 6647 SIZE_FORMAT"K", 6648 _capacity / K, new_capacity / K); 6649 } 6650 } 6651 6652 6653 // Closures 6654 // XXX: there seems to be a lot of code duplication here; 6655 // should refactor and consolidate common code. 6656 6657 // This closure is used to mark refs into the CMS generation in 6658 // the CMS bit map. Called at the first checkpoint. This closure 6659 // assumes that we do not need to re-mark dirty cards; if the CMS 6660 // generation on which this is used is not an oldest 6661 // generation then this will lose younger_gen cards! 6662 6663 MarkRefsIntoClosure::MarkRefsIntoClosure( 6664 MemRegion span, CMSBitMap* bitMap): 6665 _span(span), 6666 _bitMap(bitMap) 6667 { 6668 assert(_ref_processor == NULL, "deliberately left NULL"); 6669 assert(_bitMap->covers(_span), "_bitMap/_span mismatch"); 6670 } 6671 6672 void MarkRefsIntoClosure::do_oop(oop obj) { 6673 // if p points into _span, then mark corresponding bit in _markBitMap 6674 assert(obj->is_oop(), "expected an oop"); 6675 HeapWord* addr = (HeapWord*)obj; 6676 if (_span.contains(addr)) { 6677 // this should be made more efficient 6678 _bitMap->mark(addr); 6679 } 6680 } 6681 6682 void MarkRefsIntoClosure::do_oop(oop* p) { MarkRefsIntoClosure::do_oop_work(p); } 6683 void MarkRefsIntoClosure::do_oop(narrowOop* p) { MarkRefsIntoClosure::do_oop_work(p); } 6684 6685 // A variant of the above, used for CMS marking verification. 6686 MarkRefsIntoVerifyClosure::MarkRefsIntoVerifyClosure( 6687 MemRegion span, CMSBitMap* verification_bm, CMSBitMap* cms_bm): 6688 _span(span), 6689 _verification_bm(verification_bm), 6690 _cms_bm(cms_bm) 6691 { 6692 assert(_ref_processor == NULL, "deliberately left NULL"); 6693 assert(_verification_bm->covers(_span), "_verification_bm/_span mismatch"); 6694 } 6695 6696 void MarkRefsIntoVerifyClosure::do_oop(oop obj) { 6697 // if p points into _span, then mark corresponding bit in _markBitMap 6698 assert(obj->is_oop(), "expected an oop"); 6699 HeapWord* addr = (HeapWord*)obj; 6700 if (_span.contains(addr)) { 6701 _verification_bm->mark(addr); 6702 if (!_cms_bm->isMarked(addr)) { 6703 oop(addr)->print(); 6704 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)", addr); 6705 fatal("... aborting"); 6706 } 6707 } 6708 } 6709 6710 void MarkRefsIntoVerifyClosure::do_oop(oop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); } 6711 void MarkRefsIntoVerifyClosure::do_oop(narrowOop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); } 6712 6713 ////////////////////////////////////////////////// 6714 // MarkRefsIntoAndScanClosure 6715 ////////////////////////////////////////////////// 6716 6717 MarkRefsIntoAndScanClosure::MarkRefsIntoAndScanClosure(MemRegion span, 6718 ReferenceProcessor* rp, 6719 CMSBitMap* bit_map, 6720 CMSBitMap* mod_union_table, 6721 CMSMarkStack* mark_stack, 6722 CMSCollector* collector, 6723 bool should_yield, 6724 bool concurrent_precleaning): 6725 _collector(collector), 6726 _span(span), 6727 _bit_map(bit_map), 6728 _mark_stack(mark_stack), 6729 _pushAndMarkClosure(collector, span, rp, bit_map, mod_union_table, 6730 mark_stack, concurrent_precleaning), 6731 _yield(should_yield), 6732 _concurrent_precleaning(concurrent_precleaning), 6733 _freelistLock(NULL) 6734 { 6735 _ref_processor = rp; 6736 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 6737 } 6738 6739 // This closure is used to mark refs into the CMS generation at the 6740 // second (final) checkpoint, and to scan and transitively follow 6741 // the unmarked oops. It is also used during the concurrent precleaning 6742 // phase while scanning objects on dirty cards in the CMS generation. 6743 // The marks are made in the marking bit map and the marking stack is 6744 // used for keeping the (newly) grey objects during the scan. 6745 // The parallel version (Par_...) appears further below. 6746 void MarkRefsIntoAndScanClosure::do_oop(oop obj) { 6747 if (obj != NULL) { 6748 assert(obj->is_oop(), "expected an oop"); 6749 HeapWord* addr = (HeapWord*)obj; 6750 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 6751 assert(_collector->overflow_list_is_empty(), 6752 "overflow list should be empty"); 6753 if (_span.contains(addr) && 6754 !_bit_map->isMarked(addr)) { 6755 // mark bit map (object is now grey) 6756 _bit_map->mark(addr); 6757 // push on marking stack (stack should be empty), and drain the 6758 // stack by applying this closure to the oops in the oops popped 6759 // from the stack (i.e. blacken the grey objects) 6760 bool res = _mark_stack->push(obj); 6761 assert(res, "Should have space to push on empty stack"); 6762 do { 6763 oop new_oop = _mark_stack->pop(); 6764 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 6765 assert(_bit_map->isMarked((HeapWord*)new_oop), 6766 "only grey objects on this stack"); 6767 // iterate over the oops in this oop, marking and pushing 6768 // the ones in CMS heap (i.e. in _span). 6769 new_oop->oop_iterate(&_pushAndMarkClosure); 6770 // check if it's time to yield 6771 do_yield_check(); 6772 } while (!_mark_stack->isEmpty() || 6773 (!_concurrent_precleaning && take_from_overflow_list())); 6774 // if marking stack is empty, and we are not doing this 6775 // during precleaning, then check the overflow list 6776 } 6777 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 6778 assert(_collector->overflow_list_is_empty(), 6779 "overflow list was drained above"); 6780 // We could restore evacuated mark words, if any, used for 6781 // overflow list links here because the overflow list is 6782 // provably empty here. That would reduce the maximum 6783 // size requirements for preserved_{oop,mark}_stack. 6784 // But we'll just postpone it until we are all done 6785 // so we can just stream through. 6786 if (!_concurrent_precleaning && CMSOverflowEarlyRestoration) { 6787 _collector->restore_preserved_marks_if_any(); 6788 assert(_collector->no_preserved_marks(), "No preserved marks"); 6789 } 6790 assert(!CMSOverflowEarlyRestoration || _collector->no_preserved_marks(), 6791 "All preserved marks should have been restored above"); 6792 } 6793 } 6794 6795 void MarkRefsIntoAndScanClosure::do_oop(oop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); } 6796 void MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); } 6797 6798 void MarkRefsIntoAndScanClosure::do_yield_work() { 6799 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6800 "CMS thread should hold CMS token"); 6801 assert_lock_strong(_freelistLock); 6802 assert_lock_strong(_bit_map->lock()); 6803 // relinquish the free_list_lock and bitMaplock() 6804 _bit_map->lock()->unlock(); 6805 _freelistLock->unlock(); 6806 ConcurrentMarkSweepThread::desynchronize(true); 6807 ConcurrentMarkSweepThread::acknowledge_yield_request(); 6808 _collector->stopTimer(); 6809 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 6810 if (PrintCMSStatistics != 0) { 6811 _collector->incrementYields(); 6812 } 6813 _collector->icms_wait(); 6814 6815 // See the comment in coordinator_yield() 6816 for (unsigned i = 0; 6817 i < CMSYieldSleepCount && 6818 ConcurrentMarkSweepThread::should_yield() && 6819 !CMSCollector::foregroundGCIsActive(); 6820 ++i) { 6821 os::sleep(Thread::current(), 1, false); 6822 ConcurrentMarkSweepThread::acknowledge_yield_request(); 6823 } 6824 6825 ConcurrentMarkSweepThread::synchronize(true); 6826 _freelistLock->lock_without_safepoint_check(); 6827 _bit_map->lock()->lock_without_safepoint_check(); 6828 _collector->startTimer(); 6829 } 6830 6831 /////////////////////////////////////////////////////////// 6832 // Par_MarkRefsIntoAndScanClosure: a parallel version of 6833 // MarkRefsIntoAndScanClosure 6834 /////////////////////////////////////////////////////////// 6835 Par_MarkRefsIntoAndScanClosure::Par_MarkRefsIntoAndScanClosure( 6836 CMSCollector* collector, MemRegion span, ReferenceProcessor* rp, 6837 CMSBitMap* bit_map, OopTaskQueue* work_queue): 6838 _span(span), 6839 _bit_map(bit_map), 6840 _work_queue(work_queue), 6841 _low_water_mark(MIN2((uint)(work_queue->max_elems()/4), 6842 (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads))), 6843 _par_pushAndMarkClosure(collector, span, rp, bit_map, work_queue) 6844 { 6845 _ref_processor = rp; 6846 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 6847 } 6848 6849 // This closure is used to mark refs into the CMS generation at the 6850 // second (final) checkpoint, and to scan and transitively follow 6851 // the unmarked oops. The marks are made in the marking bit map and 6852 // the work_queue is used for keeping the (newly) grey objects during 6853 // the scan phase whence they are also available for stealing by parallel 6854 // threads. Since the marking bit map is shared, updates are 6855 // synchronized (via CAS). 6856 void Par_MarkRefsIntoAndScanClosure::do_oop(oop obj) { 6857 if (obj != NULL) { 6858 // Ignore mark word because this could be an already marked oop 6859 // that may be chained at the end of the overflow list. 6860 assert(obj->is_oop(true), "expected an oop"); 6861 HeapWord* addr = (HeapWord*)obj; 6862 if (_span.contains(addr) && 6863 !_bit_map->isMarked(addr)) { 6864 // mark bit map (object will become grey): 6865 // It is possible for several threads to be 6866 // trying to "claim" this object concurrently; 6867 // the unique thread that succeeds in marking the 6868 // object first will do the subsequent push on 6869 // to the work queue (or overflow list). 6870 if (_bit_map->par_mark(addr)) { 6871 // push on work_queue (which may not be empty), and trim the 6872 // queue to an appropriate length by applying this closure to 6873 // the oops in the oops popped from the stack (i.e. blacken the 6874 // grey objects) 6875 bool res = _work_queue->push(obj); 6876 assert(res, "Low water mark should be less than capacity?"); 6877 trim_queue(_low_water_mark); 6878 } // Else, another thread claimed the object 6879 } 6880 } 6881 } 6882 6883 void Par_MarkRefsIntoAndScanClosure::do_oop(oop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); } 6884 void Par_MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); } 6885 6886 // This closure is used to rescan the marked objects on the dirty cards 6887 // in the mod union table and the card table proper. 6888 size_t ScanMarkedObjectsAgainCarefullyClosure::do_object_careful_m( 6889 oop p, MemRegion mr) { 6890 6891 size_t size = 0; 6892 HeapWord* addr = (HeapWord*)p; 6893 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6894 assert(_span.contains(addr), "we are scanning the CMS generation"); 6895 // check if it's time to yield 6896 if (do_yield_check()) { 6897 // We yielded for some foreground stop-world work, 6898 // and we have been asked to abort this ongoing preclean cycle. 6899 return 0; 6900 } 6901 if (_bitMap->isMarked(addr)) { 6902 // it's marked; is it potentially uninitialized? 6903 if (p->klass_or_null() != NULL) { 6904 // an initialized object; ignore mark word in verification below 6905 // since we are running concurrent with mutators 6906 assert(p->is_oop(true), "should be an oop"); 6907 if (p->is_objArray()) { 6908 // objArrays are precisely marked; restrict scanning 6909 // to dirty cards only. 6910 size = CompactibleFreeListSpace::adjustObjectSize( 6911 p->oop_iterate(_scanningClosure, mr)); 6912 } else { 6913 // A non-array may have been imprecisely marked; we need 6914 // to scan object in its entirety. 6915 size = CompactibleFreeListSpace::adjustObjectSize( 6916 p->oop_iterate(_scanningClosure)); 6917 } 6918 #ifdef ASSERT 6919 size_t direct_size = 6920 CompactibleFreeListSpace::adjustObjectSize(p->size()); 6921 assert(size == direct_size, "Inconsistency in size"); 6922 assert(size >= 3, "Necessary for Printezis marks to work"); 6923 if (!_bitMap->isMarked(addr+1)) { 6924 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size); 6925 } else { 6926 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size-1); 6927 assert(_bitMap->isMarked(addr+size-1), 6928 "inconsistent Printezis mark"); 6929 } 6930 #endif // ASSERT 6931 } else { 6932 // an unitialized object 6933 assert(_bitMap->isMarked(addr+1), "missing Printezis mark?"); 6934 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 6935 size = pointer_delta(nextOneAddr + 1, addr); 6936 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 6937 "alignment problem"); 6938 // Note that pre-cleaning needn't redirty the card. OopDesc::set_klass() 6939 // will dirty the card when the klass pointer is installed in the 6940 // object (signalling the completion of initialization). 6941 } 6942 } else { 6943 // Either a not yet marked object or an uninitialized object 6944 if (p->klass_or_null() == NULL) { 6945 // An uninitialized object, skip to the next card, since 6946 // we may not be able to read its P-bits yet. 6947 assert(size == 0, "Initial value"); 6948 } else { 6949 // An object not (yet) reached by marking: we merely need to 6950 // compute its size so as to go look at the next block. 6951 assert(p->is_oop(true), "should be an oop"); 6952 size = CompactibleFreeListSpace::adjustObjectSize(p->size()); 6953 } 6954 } 6955 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6956 return size; 6957 } 6958 6959 void ScanMarkedObjectsAgainCarefullyClosure::do_yield_work() { 6960 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6961 "CMS thread should hold CMS token"); 6962 assert_lock_strong(_freelistLock); 6963 assert_lock_strong(_bitMap->lock()); 6964 // relinquish the free_list_lock and bitMaplock() 6965 _bitMap->lock()->unlock(); 6966 _freelistLock->unlock(); 6967 ConcurrentMarkSweepThread::desynchronize(true); 6968 ConcurrentMarkSweepThread::acknowledge_yield_request(); 6969 _collector->stopTimer(); 6970 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 6971 if (PrintCMSStatistics != 0) { 6972 _collector->incrementYields(); 6973 } 6974 _collector->icms_wait(); 6975 6976 // See the comment in coordinator_yield() 6977 for (unsigned i = 0; i < CMSYieldSleepCount && 6978 ConcurrentMarkSweepThread::should_yield() && 6979 !CMSCollector::foregroundGCIsActive(); ++i) { 6980 os::sleep(Thread::current(), 1, false); 6981 ConcurrentMarkSweepThread::acknowledge_yield_request(); 6982 } 6983 6984 ConcurrentMarkSweepThread::synchronize(true); 6985 _freelistLock->lock_without_safepoint_check(); 6986 _bitMap->lock()->lock_without_safepoint_check(); 6987 _collector->startTimer(); 6988 } 6989 6990 6991 ////////////////////////////////////////////////////////////////// 6992 // SurvivorSpacePrecleanClosure 6993 ////////////////////////////////////////////////////////////////// 6994 // This (single-threaded) closure is used to preclean the oops in 6995 // the survivor spaces. 6996 size_t SurvivorSpacePrecleanClosure::do_object_careful(oop p) { 6997 6998 HeapWord* addr = (HeapWord*)p; 6999 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 7000 assert(!_span.contains(addr), "we are scanning the survivor spaces"); 7001 assert(p->klass_or_null() != NULL, "object should be initializd"); 7002 // an initialized object; ignore mark word in verification below 7003 // since we are running concurrent with mutators 7004 assert(p->is_oop(true), "should be an oop"); 7005 // Note that we do not yield while we iterate over 7006 // the interior oops of p, pushing the relevant ones 7007 // on our marking stack. 7008 size_t size = p->oop_iterate(_scanning_closure); 7009 do_yield_check(); 7010 // Observe that below, we do not abandon the preclean 7011 // phase as soon as we should; rather we empty the 7012 // marking stack before returning. This is to satisfy 7013 // some existing assertions. In general, it may be a 7014 // good idea to abort immediately and complete the marking 7015 // from the grey objects at a later time. 7016 while (!_mark_stack->isEmpty()) { 7017 oop new_oop = _mark_stack->pop(); 7018 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 7019 assert(_bit_map->isMarked((HeapWord*)new_oop), 7020 "only grey objects on this stack"); 7021 // iterate over the oops in this oop, marking and pushing 7022 // the ones in CMS heap (i.e. in _span). 7023 new_oop->oop_iterate(_scanning_closure); 7024 // check if it's time to yield 7025 do_yield_check(); 7026 } 7027 unsigned int after_count = 7028 GenCollectedHeap::heap()->total_collections(); 7029 bool abort = (_before_count != after_count) || 7030 _collector->should_abort_preclean(); 7031 return abort ? 0 : size; 7032 } 7033 7034 void SurvivorSpacePrecleanClosure::do_yield_work() { 7035 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 7036 "CMS thread should hold CMS token"); 7037 assert_lock_strong(_bit_map->lock()); 7038 // Relinquish the bit map lock 7039 _bit_map->lock()->unlock(); 7040 ConcurrentMarkSweepThread::desynchronize(true); 7041 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7042 _collector->stopTimer(); 7043 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 7044 if (PrintCMSStatistics != 0) { 7045 _collector->incrementYields(); 7046 } 7047 _collector->icms_wait(); 7048 7049 // See the comment in coordinator_yield() 7050 for (unsigned i = 0; i < CMSYieldSleepCount && 7051 ConcurrentMarkSweepThread::should_yield() && 7052 !CMSCollector::foregroundGCIsActive(); ++i) { 7053 os::sleep(Thread::current(), 1, false); 7054 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7055 } 7056 7057 ConcurrentMarkSweepThread::synchronize(true); 7058 _bit_map->lock()->lock_without_safepoint_check(); 7059 _collector->startTimer(); 7060 } 7061 7062 // This closure is used to rescan the marked objects on the dirty cards 7063 // in the mod union table and the card table proper. In the parallel 7064 // case, although the bitMap is shared, we do a single read so the 7065 // isMarked() query is "safe". 7066 bool ScanMarkedObjectsAgainClosure::do_object_bm(oop p, MemRegion mr) { 7067 // Ignore mark word because we are running concurrent with mutators 7068 assert(p->is_oop_or_null(true), "expected an oop or null"); 7069 HeapWord* addr = (HeapWord*)p; 7070 assert(_span.contains(addr), "we are scanning the CMS generation"); 7071 bool is_obj_array = false; 7072 #ifdef ASSERT 7073 if (!_parallel) { 7074 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 7075 assert(_collector->overflow_list_is_empty(), 7076 "overflow list should be empty"); 7077 7078 } 7079 #endif // ASSERT 7080 if (_bit_map->isMarked(addr)) { 7081 // Obj arrays are precisely marked, non-arrays are not; 7082 // so we scan objArrays precisely and non-arrays in their 7083 // entirety. 7084 if (p->is_objArray()) { 7085 is_obj_array = true; 7086 if (_parallel) { 7087 p->oop_iterate(_par_scan_closure, mr); 7088 } else { 7089 p->oop_iterate(_scan_closure, mr); 7090 } 7091 } else { 7092 if (_parallel) { 7093 p->oop_iterate(_par_scan_closure); 7094 } else { 7095 p->oop_iterate(_scan_closure); 7096 } 7097 } 7098 } 7099 #ifdef ASSERT 7100 if (!_parallel) { 7101 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 7102 assert(_collector->overflow_list_is_empty(), 7103 "overflow list should be empty"); 7104 7105 } 7106 #endif // ASSERT 7107 return is_obj_array; 7108 } 7109 7110 MarkFromRootsClosure::MarkFromRootsClosure(CMSCollector* collector, 7111 MemRegion span, 7112 CMSBitMap* bitMap, CMSMarkStack* markStack, 7113 bool should_yield, bool verifying): 7114 _collector(collector), 7115 _span(span), 7116 _bitMap(bitMap), 7117 _mut(&collector->_modUnionTable), 7118 _markStack(markStack), 7119 _yield(should_yield), 7120 _skipBits(0) 7121 { 7122 assert(_markStack->isEmpty(), "stack should be empty"); 7123 _finger = _bitMap->startWord(); 7124 _threshold = _finger; 7125 assert(_collector->_restart_addr == NULL, "Sanity check"); 7126 assert(_span.contains(_finger), "Out of bounds _finger?"); 7127 DEBUG_ONLY(_verifying = verifying;) 7128 } 7129 7130 void MarkFromRootsClosure::reset(HeapWord* addr) { 7131 assert(_markStack->isEmpty(), "would cause duplicates on stack"); 7132 assert(_span.contains(addr), "Out of bounds _finger?"); 7133 _finger = addr; 7134 _threshold = (HeapWord*)round_to( 7135 (intptr_t)_finger, CardTableModRefBS::card_size); 7136 } 7137 7138 // Should revisit to see if this should be restructured for 7139 // greater efficiency. 7140 bool MarkFromRootsClosure::do_bit(size_t offset) { 7141 if (_skipBits > 0) { 7142 _skipBits--; 7143 return true; 7144 } 7145 // convert offset into a HeapWord* 7146 HeapWord* addr = _bitMap->startWord() + offset; 7147 assert(_bitMap->endWord() && addr < _bitMap->endWord(), 7148 "address out of range"); 7149 assert(_bitMap->isMarked(addr), "tautology"); 7150 if (_bitMap->isMarked(addr+1)) { 7151 // this is an allocated but not yet initialized object 7152 assert(_skipBits == 0, "tautology"); 7153 _skipBits = 2; // skip next two marked bits ("Printezis-marks") 7154 oop p = oop(addr); 7155 if (p->klass_or_null() == NULL) { 7156 DEBUG_ONLY(if (!_verifying) {) 7157 // We re-dirty the cards on which this object lies and increase 7158 // the _threshold so that we'll come back to scan this object 7159 // during the preclean or remark phase. (CMSCleanOnEnter) 7160 if (CMSCleanOnEnter) { 7161 size_t sz = _collector->block_size_using_printezis_bits(addr); 7162 HeapWord* end_card_addr = (HeapWord*)round_to( 7163 (intptr_t)(addr+sz), CardTableModRefBS::card_size); 7164 MemRegion redirty_range = MemRegion(addr, end_card_addr); 7165 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 7166 // Bump _threshold to end_card_addr; note that 7167 // _threshold cannot possibly exceed end_card_addr, anyhow. 7168 // This prevents future clearing of the card as the scan proceeds 7169 // to the right. 7170 assert(_threshold <= end_card_addr, 7171 "Because we are just scanning into this object"); 7172 if (_threshold < end_card_addr) { 7173 _threshold = end_card_addr; 7174 } 7175 if (p->klass_or_null() != NULL) { 7176 // Redirty the range of cards... 7177 _mut->mark_range(redirty_range); 7178 } // ...else the setting of klass will dirty the card anyway. 7179 } 7180 DEBUG_ONLY(}) 7181 return true; 7182 } 7183 } 7184 scanOopsInOop(addr); 7185 return true; 7186 } 7187 7188 // We take a break if we've been at this for a while, 7189 // so as to avoid monopolizing the locks involved. 7190 void MarkFromRootsClosure::do_yield_work() { 7191 // First give up the locks, then yield, then re-lock 7192 // We should probably use a constructor/destructor idiom to 7193 // do this unlock/lock or modify the MutexUnlocker class to 7194 // serve our purpose. XXX 7195 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 7196 "CMS thread should hold CMS token"); 7197 assert_lock_strong(_bitMap->lock()); 7198 _bitMap->lock()->unlock(); 7199 ConcurrentMarkSweepThread::desynchronize(true); 7200 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7201 _collector->stopTimer(); 7202 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 7203 if (PrintCMSStatistics != 0) { 7204 _collector->incrementYields(); 7205 } 7206 _collector->icms_wait(); 7207 7208 // See the comment in coordinator_yield() 7209 for (unsigned i = 0; i < CMSYieldSleepCount && 7210 ConcurrentMarkSweepThread::should_yield() && 7211 !CMSCollector::foregroundGCIsActive(); ++i) { 7212 os::sleep(Thread::current(), 1, false); 7213 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7214 } 7215 7216 ConcurrentMarkSweepThread::synchronize(true); 7217 _bitMap->lock()->lock_without_safepoint_check(); 7218 _collector->startTimer(); 7219 } 7220 7221 void MarkFromRootsClosure::scanOopsInOop(HeapWord* ptr) { 7222 assert(_bitMap->isMarked(ptr), "expected bit to be set"); 7223 assert(_markStack->isEmpty(), 7224 "should drain stack to limit stack usage"); 7225 // convert ptr to an oop preparatory to scanning 7226 oop obj = oop(ptr); 7227 // Ignore mark word in verification below, since we 7228 // may be running concurrent with mutators. 7229 assert(obj->is_oop(true), "should be an oop"); 7230 assert(_finger <= ptr, "_finger runneth ahead"); 7231 // advance the finger to right end of this object 7232 _finger = ptr + obj->size(); 7233 assert(_finger > ptr, "we just incremented it above"); 7234 // On large heaps, it may take us some time to get through 7235 // the marking phase (especially if running iCMS). During 7236 // this time it's possible that a lot of mutations have 7237 // accumulated in the card table and the mod union table -- 7238 // these mutation records are redundant until we have 7239 // actually traced into the corresponding card. 7240 // Here, we check whether advancing the finger would make 7241 // us cross into a new card, and if so clear corresponding 7242 // cards in the MUT (preclean them in the card-table in the 7243 // future). 7244 7245 DEBUG_ONLY(if (!_verifying) {) 7246 // The clean-on-enter optimization is disabled by default, 7247 // until we fix 6178663. 7248 if (CMSCleanOnEnter && (_finger > _threshold)) { 7249 // [_threshold, _finger) represents the interval 7250 // of cards to be cleared in MUT (or precleaned in card table). 7251 // The set of cards to be cleared is all those that overlap 7252 // with the interval [_threshold, _finger); note that 7253 // _threshold is always kept card-aligned but _finger isn't 7254 // always card-aligned. 7255 HeapWord* old_threshold = _threshold; 7256 assert(old_threshold == (HeapWord*)round_to( 7257 (intptr_t)old_threshold, CardTableModRefBS::card_size), 7258 "_threshold should always be card-aligned"); 7259 _threshold = (HeapWord*)round_to( 7260 (intptr_t)_finger, CardTableModRefBS::card_size); 7261 MemRegion mr(old_threshold, _threshold); 7262 assert(!mr.is_empty(), "Control point invariant"); 7263 assert(_span.contains(mr), "Should clear within span"); 7264 _mut->clear_range(mr); 7265 } 7266 DEBUG_ONLY(}) 7267 // Note: the finger doesn't advance while we drain 7268 // the stack below. 7269 PushOrMarkClosure pushOrMarkClosure(_collector, 7270 _span, _bitMap, _markStack, 7271 _finger, this); 7272 bool res = _markStack->push(obj); 7273 assert(res, "Empty non-zero size stack should have space for single push"); 7274 while (!_markStack->isEmpty()) { 7275 oop new_oop = _markStack->pop(); 7276 // Skip verifying header mark word below because we are 7277 // running concurrent with mutators. 7278 assert(new_oop->is_oop(true), "Oops! expected to pop an oop"); 7279 // now scan this oop's oops 7280 new_oop->oop_iterate(&pushOrMarkClosure); 7281 do_yield_check(); 7282 } 7283 assert(_markStack->isEmpty(), "tautology, emphasizing post-condition"); 7284 } 7285 7286 Par_MarkFromRootsClosure::Par_MarkFromRootsClosure(CMSConcMarkingTask* task, 7287 CMSCollector* collector, MemRegion span, 7288 CMSBitMap* bit_map, 7289 OopTaskQueue* work_queue, 7290 CMSMarkStack* overflow_stack, 7291 bool should_yield): 7292 _collector(collector), 7293 _whole_span(collector->_span), 7294 _span(span), 7295 _bit_map(bit_map), 7296 _mut(&collector->_modUnionTable), 7297 _work_queue(work_queue), 7298 _overflow_stack(overflow_stack), 7299 _yield(should_yield), 7300 _skip_bits(0), 7301 _task(task) 7302 { 7303 assert(_work_queue->size() == 0, "work_queue should be empty"); 7304 _finger = span.start(); 7305 _threshold = _finger; // XXX Defer clear-on-enter optimization for now 7306 assert(_span.contains(_finger), "Out of bounds _finger?"); 7307 } 7308 7309 // Should revisit to see if this should be restructured for 7310 // greater efficiency. 7311 bool Par_MarkFromRootsClosure::do_bit(size_t offset) { 7312 if (_skip_bits > 0) { 7313 _skip_bits--; 7314 return true; 7315 } 7316 // convert offset into a HeapWord* 7317 HeapWord* addr = _bit_map->startWord() + offset; 7318 assert(_bit_map->endWord() && addr < _bit_map->endWord(), 7319 "address out of range"); 7320 assert(_bit_map->isMarked(addr), "tautology"); 7321 if (_bit_map->isMarked(addr+1)) { 7322 // this is an allocated object that might not yet be initialized 7323 assert(_skip_bits == 0, "tautology"); 7324 _skip_bits = 2; // skip next two marked bits ("Printezis-marks") 7325 oop p = oop(addr); 7326 if (p->klass_or_null() == NULL) { 7327 // in the case of Clean-on-Enter optimization, redirty card 7328 // and avoid clearing card by increasing the threshold. 7329 return true; 7330 } 7331 } 7332 scan_oops_in_oop(addr); 7333 return true; 7334 } 7335 7336 void Par_MarkFromRootsClosure::scan_oops_in_oop(HeapWord* ptr) { 7337 assert(_bit_map->isMarked(ptr), "expected bit to be set"); 7338 // Should we assert that our work queue is empty or 7339 // below some drain limit? 7340 assert(_work_queue->size() == 0, 7341 "should drain stack to limit stack usage"); 7342 // convert ptr to an oop preparatory to scanning 7343 oop obj = oop(ptr); 7344 // Ignore mark word in verification below, since we 7345 // may be running concurrent with mutators. 7346 assert(obj->is_oop(true), "should be an oop"); 7347 assert(_finger <= ptr, "_finger runneth ahead"); 7348 // advance the finger to right end of this object 7349 _finger = ptr + obj->size(); 7350 assert(_finger > ptr, "we just incremented it above"); 7351 // On large heaps, it may take us some time to get through 7352 // the marking phase (especially if running iCMS). During 7353 // this time it's possible that a lot of mutations have 7354 // accumulated in the card table and the mod union table -- 7355 // these mutation records are redundant until we have 7356 // actually traced into the corresponding card. 7357 // Here, we check whether advancing the finger would make 7358 // us cross into a new card, and if so clear corresponding 7359 // cards in the MUT (preclean them in the card-table in the 7360 // future). 7361 7362 // The clean-on-enter optimization is disabled by default, 7363 // until we fix 6178663. 7364 if (CMSCleanOnEnter && (_finger > _threshold)) { 7365 // [_threshold, _finger) represents the interval 7366 // of cards to be cleared in MUT (or precleaned in card table). 7367 // The set of cards to be cleared is all those that overlap 7368 // with the interval [_threshold, _finger); note that 7369 // _threshold is always kept card-aligned but _finger isn't 7370 // always card-aligned. 7371 HeapWord* old_threshold = _threshold; 7372 assert(old_threshold == (HeapWord*)round_to( 7373 (intptr_t)old_threshold, CardTableModRefBS::card_size), 7374 "_threshold should always be card-aligned"); 7375 _threshold = (HeapWord*)round_to( 7376 (intptr_t)_finger, CardTableModRefBS::card_size); 7377 MemRegion mr(old_threshold, _threshold); 7378 assert(!mr.is_empty(), "Control point invariant"); 7379 assert(_span.contains(mr), "Should clear within span"); // _whole_span ?? 7380 _mut->clear_range(mr); 7381 } 7382 7383 // Note: the local finger doesn't advance while we drain 7384 // the stack below, but the global finger sure can and will. 7385 HeapWord** gfa = _task->global_finger_addr(); 7386 Par_PushOrMarkClosure pushOrMarkClosure(_collector, 7387 _span, _bit_map, 7388 _work_queue, 7389 _overflow_stack, 7390 _finger, 7391 gfa, this); 7392 bool res = _work_queue->push(obj); // overflow could occur here 7393 assert(res, "Will hold once we use workqueues"); 7394 while (true) { 7395 oop new_oop; 7396 if (!_work_queue->pop_local(new_oop)) { 7397 // We emptied our work_queue; check if there's stuff that can 7398 // be gotten from the overflow stack. 7399 if (CMSConcMarkingTask::get_work_from_overflow_stack( 7400 _overflow_stack, _work_queue)) { 7401 do_yield_check(); 7402 continue; 7403 } else { // done 7404 break; 7405 } 7406 } 7407 // Skip verifying header mark word below because we are 7408 // running concurrent with mutators. 7409 assert(new_oop->is_oop(true), "Oops! expected to pop an oop"); 7410 // now scan this oop's oops 7411 new_oop->oop_iterate(&pushOrMarkClosure); 7412 do_yield_check(); 7413 } 7414 assert(_work_queue->size() == 0, "tautology, emphasizing post-condition"); 7415 } 7416 7417 // Yield in response to a request from VM Thread or 7418 // from mutators. 7419 void Par_MarkFromRootsClosure::do_yield_work() { 7420 assert(_task != NULL, "sanity"); 7421 _task->yield(); 7422 } 7423 7424 // A variant of the above used for verifying CMS marking work. 7425 MarkFromRootsVerifyClosure::MarkFromRootsVerifyClosure(CMSCollector* collector, 7426 MemRegion span, 7427 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 7428 CMSMarkStack* mark_stack): 7429 _collector(collector), 7430 _span(span), 7431 _verification_bm(verification_bm), 7432 _cms_bm(cms_bm), 7433 _mark_stack(mark_stack), 7434 _pam_verify_closure(collector, span, verification_bm, cms_bm, 7435 mark_stack) 7436 { 7437 assert(_mark_stack->isEmpty(), "stack should be empty"); 7438 _finger = _verification_bm->startWord(); 7439 assert(_collector->_restart_addr == NULL, "Sanity check"); 7440 assert(_span.contains(_finger), "Out of bounds _finger?"); 7441 } 7442 7443 void MarkFromRootsVerifyClosure::reset(HeapWord* addr) { 7444 assert(_mark_stack->isEmpty(), "would cause duplicates on stack"); 7445 assert(_span.contains(addr), "Out of bounds _finger?"); 7446 _finger = addr; 7447 } 7448 7449 // Should revisit to see if this should be restructured for 7450 // greater efficiency. 7451 bool MarkFromRootsVerifyClosure::do_bit(size_t offset) { 7452 // convert offset into a HeapWord* 7453 HeapWord* addr = _verification_bm->startWord() + offset; 7454 assert(_verification_bm->endWord() && addr < _verification_bm->endWord(), 7455 "address out of range"); 7456 assert(_verification_bm->isMarked(addr), "tautology"); 7457 assert(_cms_bm->isMarked(addr), "tautology"); 7458 7459 assert(_mark_stack->isEmpty(), 7460 "should drain stack to limit stack usage"); 7461 // convert addr to an oop preparatory to scanning 7462 oop obj = oop(addr); 7463 assert(obj->is_oop(), "should be an oop"); 7464 assert(_finger <= addr, "_finger runneth ahead"); 7465 // advance the finger to right end of this object 7466 _finger = addr + obj->size(); 7467 assert(_finger > addr, "we just incremented it above"); 7468 // Note: the finger doesn't advance while we drain 7469 // the stack below. 7470 bool res = _mark_stack->push(obj); 7471 assert(res, "Empty non-zero size stack should have space for single push"); 7472 while (!_mark_stack->isEmpty()) { 7473 oop new_oop = _mark_stack->pop(); 7474 assert(new_oop->is_oop(), "Oops! expected to pop an oop"); 7475 // now scan this oop's oops 7476 new_oop->oop_iterate(&_pam_verify_closure); 7477 } 7478 assert(_mark_stack->isEmpty(), "tautology, emphasizing post-condition"); 7479 return true; 7480 } 7481 7482 PushAndMarkVerifyClosure::PushAndMarkVerifyClosure( 7483 CMSCollector* collector, MemRegion span, 7484 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 7485 CMSMarkStack* mark_stack): 7486 CMSOopClosure(collector->ref_processor()), 7487 _collector(collector), 7488 _span(span), 7489 _verification_bm(verification_bm), 7490 _cms_bm(cms_bm), 7491 _mark_stack(mark_stack) 7492 { } 7493 7494 void PushAndMarkVerifyClosure::do_oop(oop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 7495 void PushAndMarkVerifyClosure::do_oop(narrowOop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 7496 7497 // Upon stack overflow, we discard (part of) the stack, 7498 // remembering the least address amongst those discarded 7499 // in CMSCollector's _restart_address. 7500 void PushAndMarkVerifyClosure::handle_stack_overflow(HeapWord* lost) { 7501 // Remember the least grey address discarded 7502 HeapWord* ra = (HeapWord*)_mark_stack->least_value(lost); 7503 _collector->lower_restart_addr(ra); 7504 _mark_stack->reset(); // discard stack contents 7505 _mark_stack->expand(); // expand the stack if possible 7506 } 7507 7508 void PushAndMarkVerifyClosure::do_oop(oop obj) { 7509 assert(obj->is_oop_or_null(), "expected an oop or NULL"); 7510 HeapWord* addr = (HeapWord*)obj; 7511 if (_span.contains(addr) && !_verification_bm->isMarked(addr)) { 7512 // Oop lies in _span and isn't yet grey or black 7513 _verification_bm->mark(addr); // now grey 7514 if (!_cms_bm->isMarked(addr)) { 7515 oop(addr)->print(); 7516 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)", 7517 addr); 7518 fatal("... aborting"); 7519 } 7520 7521 if (!_mark_stack->push(obj)) { // stack overflow 7522 if (PrintCMSStatistics != 0) { 7523 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 7524 SIZE_FORMAT, _mark_stack->capacity()); 7525 } 7526 assert(_mark_stack->isFull(), "Else push should have succeeded"); 7527 handle_stack_overflow(addr); 7528 } 7529 // anything including and to the right of _finger 7530 // will be scanned as we iterate over the remainder of the 7531 // bit map 7532 } 7533 } 7534 7535 PushOrMarkClosure::PushOrMarkClosure(CMSCollector* collector, 7536 MemRegion span, 7537 CMSBitMap* bitMap, CMSMarkStack* markStack, 7538 HeapWord* finger, MarkFromRootsClosure* parent) : 7539 CMSOopClosure(collector->ref_processor()), 7540 _collector(collector), 7541 _span(span), 7542 _bitMap(bitMap), 7543 _markStack(markStack), 7544 _finger(finger), 7545 _parent(parent) 7546 { } 7547 7548 Par_PushOrMarkClosure::Par_PushOrMarkClosure(CMSCollector* collector, 7549 MemRegion span, 7550 CMSBitMap* bit_map, 7551 OopTaskQueue* work_queue, 7552 CMSMarkStack* overflow_stack, 7553 HeapWord* finger, 7554 HeapWord** global_finger_addr, 7555 Par_MarkFromRootsClosure* parent) : 7556 CMSOopClosure(collector->ref_processor()), 7557 _collector(collector), 7558 _whole_span(collector->_span), 7559 _span(span), 7560 _bit_map(bit_map), 7561 _work_queue(work_queue), 7562 _overflow_stack(overflow_stack), 7563 _finger(finger), 7564 _global_finger_addr(global_finger_addr), 7565 _parent(parent) 7566 { } 7567 7568 // Assumes thread-safe access by callers, who are 7569 // responsible for mutual exclusion. 7570 void CMSCollector::lower_restart_addr(HeapWord* low) { 7571 assert(_span.contains(low), "Out of bounds addr"); 7572 if (_restart_addr == NULL) { 7573 _restart_addr = low; 7574 } else { 7575 _restart_addr = MIN2(_restart_addr, low); 7576 } 7577 } 7578 7579 // Upon stack overflow, we discard (part of) the stack, 7580 // remembering the least address amongst those discarded 7581 // in CMSCollector's _restart_address. 7582 void PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 7583 // Remember the least grey address discarded 7584 HeapWord* ra = (HeapWord*)_markStack->least_value(lost); 7585 _collector->lower_restart_addr(ra); 7586 _markStack->reset(); // discard stack contents 7587 _markStack->expand(); // expand the stack if possible 7588 } 7589 7590 // Upon stack overflow, we discard (part of) the stack, 7591 // remembering the least address amongst those discarded 7592 // in CMSCollector's _restart_address. 7593 void Par_PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 7594 // We need to do this under a mutex to prevent other 7595 // workers from interfering with the work done below. 7596 MutexLockerEx ml(_overflow_stack->par_lock(), 7597 Mutex::_no_safepoint_check_flag); 7598 // Remember the least grey address discarded 7599 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 7600 _collector->lower_restart_addr(ra); 7601 _overflow_stack->reset(); // discard stack contents 7602 _overflow_stack->expand(); // expand the stack if possible 7603 } 7604 7605 void CMKlassClosure::do_klass(Klass* k) { 7606 assert(_oop_closure != NULL, "Not initialized?"); 7607 k->oops_do(_oop_closure); 7608 } 7609 7610 void PushOrMarkClosure::do_oop(oop obj) { 7611 // Ignore mark word because we are running concurrent with mutators. 7612 assert(obj->is_oop_or_null(true), "expected an oop or NULL"); 7613 HeapWord* addr = (HeapWord*)obj; 7614 if (_span.contains(addr) && !_bitMap->isMarked(addr)) { 7615 // Oop lies in _span and isn't yet grey or black 7616 _bitMap->mark(addr); // now grey 7617 if (addr < _finger) { 7618 // the bit map iteration has already either passed, or 7619 // sampled, this bit in the bit map; we'll need to 7620 // use the marking stack to scan this oop's oops. 7621 bool simulate_overflow = false; 7622 NOT_PRODUCT( 7623 if (CMSMarkStackOverflowALot && 7624 _collector->simulate_overflow()) { 7625 // simulate a stack overflow 7626 simulate_overflow = true; 7627 } 7628 ) 7629 if (simulate_overflow || !_markStack->push(obj)) { // stack overflow 7630 if (PrintCMSStatistics != 0) { 7631 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 7632 SIZE_FORMAT, _markStack->capacity()); 7633 } 7634 assert(simulate_overflow || _markStack->isFull(), "Else push should have succeeded"); 7635 handle_stack_overflow(addr); 7636 } 7637 } 7638 // anything including and to the right of _finger 7639 // will be scanned as we iterate over the remainder of the 7640 // bit map 7641 do_yield_check(); 7642 } 7643 } 7644 7645 void PushOrMarkClosure::do_oop(oop* p) { PushOrMarkClosure::do_oop_work(p); } 7646 void PushOrMarkClosure::do_oop(narrowOop* p) { PushOrMarkClosure::do_oop_work(p); } 7647 7648 void Par_PushOrMarkClosure::do_oop(oop obj) { 7649 // Ignore mark word because we are running concurrent with mutators. 7650 assert(obj->is_oop_or_null(true), "expected an oop or NULL"); 7651 HeapWord* addr = (HeapWord*)obj; 7652 if (_whole_span.contains(addr) && !_bit_map->isMarked(addr)) { 7653 // Oop lies in _span and isn't yet grey or black 7654 // We read the global_finger (volatile read) strictly after marking oop 7655 bool res = _bit_map->par_mark(addr); // now grey 7656 volatile HeapWord** gfa = (volatile HeapWord**)_global_finger_addr; 7657 // Should we push this marked oop on our stack? 7658 // -- if someone else marked it, nothing to do 7659 // -- if target oop is above global finger nothing to do 7660 // -- if target oop is in chunk and above local finger 7661 // then nothing to do 7662 // -- else push on work queue 7663 if ( !res // someone else marked it, they will deal with it 7664 || (addr >= *gfa) // will be scanned in a later task 7665 || (_span.contains(addr) && addr >= _finger)) { // later in this chunk 7666 return; 7667 } 7668 // the bit map iteration has already either passed, or 7669 // sampled, this bit in the bit map; we'll need to 7670 // use the marking stack to scan this oop's oops. 7671 bool simulate_overflow = false; 7672 NOT_PRODUCT( 7673 if (CMSMarkStackOverflowALot && 7674 _collector->simulate_overflow()) { 7675 // simulate a stack overflow 7676 simulate_overflow = true; 7677 } 7678 ) 7679 if (simulate_overflow || 7680 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) { 7681 // stack overflow 7682 if (PrintCMSStatistics != 0) { 7683 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 7684 SIZE_FORMAT, _overflow_stack->capacity()); 7685 } 7686 // We cannot assert that the overflow stack is full because 7687 // it may have been emptied since. 7688 assert(simulate_overflow || 7689 _work_queue->size() == _work_queue->max_elems(), 7690 "Else push should have succeeded"); 7691 handle_stack_overflow(addr); 7692 } 7693 do_yield_check(); 7694 } 7695 } 7696 7697 void Par_PushOrMarkClosure::do_oop(oop* p) { Par_PushOrMarkClosure::do_oop_work(p); } 7698 void Par_PushOrMarkClosure::do_oop(narrowOop* p) { Par_PushOrMarkClosure::do_oop_work(p); } 7699 7700 PushAndMarkClosure::PushAndMarkClosure(CMSCollector* collector, 7701 MemRegion span, 7702 ReferenceProcessor* rp, 7703 CMSBitMap* bit_map, 7704 CMSBitMap* mod_union_table, 7705 CMSMarkStack* mark_stack, 7706 bool concurrent_precleaning): 7707 CMSOopClosure(rp), 7708 _collector(collector), 7709 _span(span), 7710 _bit_map(bit_map), 7711 _mod_union_table(mod_union_table), 7712 _mark_stack(mark_stack), 7713 _concurrent_precleaning(concurrent_precleaning) 7714 { 7715 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 7716 } 7717 7718 // Grey object rescan during pre-cleaning and second checkpoint phases -- 7719 // the non-parallel version (the parallel version appears further below.) 7720 void PushAndMarkClosure::do_oop(oop obj) { 7721 // Ignore mark word verification. If during concurrent precleaning, 7722 // the object monitor may be locked. If during the checkpoint 7723 // phases, the object may already have been reached by a different 7724 // path and may be at the end of the global overflow list (so 7725 // the mark word may be NULL). 7726 assert(obj->is_oop_or_null(true /* ignore mark word */), 7727 "expected an oop or NULL"); 7728 HeapWord* addr = (HeapWord*)obj; 7729 // Check if oop points into the CMS generation 7730 // and is not marked 7731 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 7732 // a white object ... 7733 _bit_map->mark(addr); // ... now grey 7734 // push on the marking stack (grey set) 7735 bool simulate_overflow = false; 7736 NOT_PRODUCT( 7737 if (CMSMarkStackOverflowALot && 7738 _collector->simulate_overflow()) { 7739 // simulate a stack overflow 7740 simulate_overflow = true; 7741 } 7742 ) 7743 if (simulate_overflow || !_mark_stack->push(obj)) { 7744 if (_concurrent_precleaning) { 7745 // During precleaning we can just dirty the appropriate card(s) 7746 // in the mod union table, thus ensuring that the object remains 7747 // in the grey set and continue. In the case of object arrays 7748 // we need to dirty all of the cards that the object spans, 7749 // since the rescan of object arrays will be limited to the 7750 // dirty cards. 7751 // Note that no one can be intefering with us in this action 7752 // of dirtying the mod union table, so no locking or atomics 7753 // are required. 7754 if (obj->is_objArray()) { 7755 size_t sz = obj->size(); 7756 HeapWord* end_card_addr = (HeapWord*)round_to( 7757 (intptr_t)(addr+sz), CardTableModRefBS::card_size); 7758 MemRegion redirty_range = MemRegion(addr, end_card_addr); 7759 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 7760 _mod_union_table->mark_range(redirty_range); 7761 } else { 7762 _mod_union_table->mark(addr); 7763 } 7764 _collector->_ser_pmc_preclean_ovflw++; 7765 } else { 7766 // During the remark phase, we need to remember this oop 7767 // in the overflow list. 7768 _collector->push_on_overflow_list(obj); 7769 _collector->_ser_pmc_remark_ovflw++; 7770 } 7771 } 7772 } 7773 } 7774 7775 Par_PushAndMarkClosure::Par_PushAndMarkClosure(CMSCollector* collector, 7776 MemRegion span, 7777 ReferenceProcessor* rp, 7778 CMSBitMap* bit_map, 7779 OopTaskQueue* work_queue): 7780 CMSOopClosure(rp), 7781 _collector(collector), 7782 _span(span), 7783 _bit_map(bit_map), 7784 _work_queue(work_queue) 7785 { 7786 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 7787 } 7788 7789 void PushAndMarkClosure::do_oop(oop* p) { PushAndMarkClosure::do_oop_work(p); } 7790 void PushAndMarkClosure::do_oop(narrowOop* p) { PushAndMarkClosure::do_oop_work(p); } 7791 7792 // Grey object rescan during second checkpoint phase -- 7793 // the parallel version. 7794 void Par_PushAndMarkClosure::do_oop(oop obj) { 7795 // In the assert below, we ignore the mark word because 7796 // this oop may point to an already visited object that is 7797 // on the overflow stack (in which case the mark word has 7798 // been hijacked for chaining into the overflow stack -- 7799 // if this is the last object in the overflow stack then 7800 // its mark word will be NULL). Because this object may 7801 // have been subsequently popped off the global overflow 7802 // stack, and the mark word possibly restored to the prototypical 7803 // value, by the time we get to examined this failing assert in 7804 // the debugger, is_oop_or_null(false) may subsequently start 7805 // to hold. 7806 assert(obj->is_oop_or_null(true), 7807 "expected an oop or NULL"); 7808 HeapWord* addr = (HeapWord*)obj; 7809 // Check if oop points into the CMS generation 7810 // and is not marked 7811 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 7812 // a white object ... 7813 // If we manage to "claim" the object, by being the 7814 // first thread to mark it, then we push it on our 7815 // marking stack 7816 if (_bit_map->par_mark(addr)) { // ... now grey 7817 // push on work queue (grey set) 7818 bool simulate_overflow = false; 7819 NOT_PRODUCT( 7820 if (CMSMarkStackOverflowALot && 7821 _collector->par_simulate_overflow()) { 7822 // simulate a stack overflow 7823 simulate_overflow = true; 7824 } 7825 ) 7826 if (simulate_overflow || !_work_queue->push(obj)) { 7827 _collector->par_push_on_overflow_list(obj); 7828 _collector->_par_pmc_remark_ovflw++; // imprecise OK: no need to CAS 7829 } 7830 } // Else, some other thread got there first 7831 } 7832 } 7833 7834 void Par_PushAndMarkClosure::do_oop(oop* p) { Par_PushAndMarkClosure::do_oop_work(p); } 7835 void Par_PushAndMarkClosure::do_oop(narrowOop* p) { Par_PushAndMarkClosure::do_oop_work(p); } 7836 7837 void CMSPrecleanRefsYieldClosure::do_yield_work() { 7838 Mutex* bml = _collector->bitMapLock(); 7839 assert_lock_strong(bml); 7840 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 7841 "CMS thread should hold CMS token"); 7842 7843 bml->unlock(); 7844 ConcurrentMarkSweepThread::desynchronize(true); 7845 7846 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7847 7848 _collector->stopTimer(); 7849 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 7850 if (PrintCMSStatistics != 0) { 7851 _collector->incrementYields(); 7852 } 7853 _collector->icms_wait(); 7854 7855 // See the comment in coordinator_yield() 7856 for (unsigned i = 0; i < CMSYieldSleepCount && 7857 ConcurrentMarkSweepThread::should_yield() && 7858 !CMSCollector::foregroundGCIsActive(); ++i) { 7859 os::sleep(Thread::current(), 1, false); 7860 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7861 } 7862 7863 ConcurrentMarkSweepThread::synchronize(true); 7864 bml->lock(); 7865 7866 _collector->startTimer(); 7867 } 7868 7869 bool CMSPrecleanRefsYieldClosure::should_return() { 7870 if (ConcurrentMarkSweepThread::should_yield()) { 7871 do_yield_work(); 7872 } 7873 return _collector->foregroundGCIsActive(); 7874 } 7875 7876 void MarkFromDirtyCardsClosure::do_MemRegion(MemRegion mr) { 7877 assert(((size_t)mr.start())%CardTableModRefBS::card_size_in_words == 0, 7878 "mr should be aligned to start at a card boundary"); 7879 // We'd like to assert: 7880 // assert(mr.word_size()%CardTableModRefBS::card_size_in_words == 0, 7881 // "mr should be a range of cards"); 7882 // However, that would be too strong in one case -- the last 7883 // partition ends at _unallocated_block which, in general, can be 7884 // an arbitrary boundary, not necessarily card aligned. 7885 if (PrintCMSStatistics != 0) { 7886 _num_dirty_cards += 7887 mr.word_size()/CardTableModRefBS::card_size_in_words; 7888 } 7889 _space->object_iterate_mem(mr, &_scan_cl); 7890 } 7891 7892 SweepClosure::SweepClosure(CMSCollector* collector, 7893 ConcurrentMarkSweepGeneration* g, 7894 CMSBitMap* bitMap, bool should_yield) : 7895 _collector(collector), 7896 _g(g), 7897 _sp(g->cmsSpace()), 7898 _limit(_sp->sweep_limit()), 7899 _freelistLock(_sp->freelistLock()), 7900 _bitMap(bitMap), 7901 _yield(should_yield), 7902 _inFreeRange(false), // No free range at beginning of sweep 7903 _freeRangeInFreeLists(false), // No free range at beginning of sweep 7904 _lastFreeRangeCoalesced(false), 7905 _freeFinger(g->used_region().start()) 7906 { 7907 NOT_PRODUCT( 7908 _numObjectsFreed = 0; 7909 _numWordsFreed = 0; 7910 _numObjectsLive = 0; 7911 _numWordsLive = 0; 7912 _numObjectsAlreadyFree = 0; 7913 _numWordsAlreadyFree = 0; 7914 _last_fc = NULL; 7915 7916 _sp->initializeIndexedFreeListArrayReturnedBytes(); 7917 _sp->dictionary()->initialize_dict_returned_bytes(); 7918 ) 7919 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7920 "sweep _limit out of bounds"); 7921 if (CMSTraceSweeper) { 7922 gclog_or_tty->print_cr("\n====================\nStarting new sweep with limit " PTR_FORMAT, 7923 _limit); 7924 } 7925 } 7926 7927 void SweepClosure::print_on(outputStream* st) const { 7928 tty->print_cr("_sp = [" PTR_FORMAT "," PTR_FORMAT ")", 7929 _sp->bottom(), _sp->end()); 7930 tty->print_cr("_limit = " PTR_FORMAT, _limit); 7931 tty->print_cr("_freeFinger = " PTR_FORMAT, _freeFinger); 7932 NOT_PRODUCT(tty->print_cr("_last_fc = " PTR_FORMAT, _last_fc);) 7933 tty->print_cr("_inFreeRange = %d, _freeRangeInFreeLists = %d, _lastFreeRangeCoalesced = %d", 7934 _inFreeRange, _freeRangeInFreeLists, _lastFreeRangeCoalesced); 7935 } 7936 7937 #ifndef PRODUCT 7938 // Assertion checking only: no useful work in product mode -- 7939 // however, if any of the flags below become product flags, 7940 // you may need to review this code to see if it needs to be 7941 // enabled in product mode. 7942 SweepClosure::~SweepClosure() { 7943 assert_lock_strong(_freelistLock); 7944 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7945 "sweep _limit out of bounds"); 7946 if (inFreeRange()) { 7947 warning("inFreeRange() should have been reset; dumping state of SweepClosure"); 7948 print(); 7949 ShouldNotReachHere(); 7950 } 7951 if (Verbose && PrintGC) { 7952 gclog_or_tty->print("Collected "SIZE_FORMAT" objects, " SIZE_FORMAT " bytes", 7953 _numObjectsFreed, _numWordsFreed*sizeof(HeapWord)); 7954 gclog_or_tty->print_cr("\nLive "SIZE_FORMAT" objects, " 7955 SIZE_FORMAT" bytes " 7956 "Already free "SIZE_FORMAT" objects, "SIZE_FORMAT" bytes", 7957 _numObjectsLive, _numWordsLive*sizeof(HeapWord), 7958 _numObjectsAlreadyFree, _numWordsAlreadyFree*sizeof(HeapWord)); 7959 size_t totalBytes = (_numWordsFreed + _numWordsLive + _numWordsAlreadyFree) 7960 * sizeof(HeapWord); 7961 gclog_or_tty->print_cr("Total sweep: "SIZE_FORMAT" bytes", totalBytes); 7962 7963 if (PrintCMSStatistics && CMSVerifyReturnedBytes) { 7964 size_t indexListReturnedBytes = _sp->sumIndexedFreeListArrayReturnedBytes(); 7965 size_t dict_returned_bytes = _sp->dictionary()->sum_dict_returned_bytes(); 7966 size_t returned_bytes = indexListReturnedBytes + dict_returned_bytes; 7967 gclog_or_tty->print("Returned "SIZE_FORMAT" bytes", returned_bytes); 7968 gclog_or_tty->print(" Indexed List Returned "SIZE_FORMAT" bytes", 7969 indexListReturnedBytes); 7970 gclog_or_tty->print_cr(" Dictionary Returned "SIZE_FORMAT" bytes", 7971 dict_returned_bytes); 7972 } 7973 } 7974 if (CMSTraceSweeper) { 7975 gclog_or_tty->print_cr("end of sweep with _limit = " PTR_FORMAT "\n================", 7976 _limit); 7977 } 7978 } 7979 #endif // PRODUCT 7980 7981 void SweepClosure::initialize_free_range(HeapWord* freeFinger, 7982 bool freeRangeInFreeLists) { 7983 if (CMSTraceSweeper) { 7984 gclog_or_tty->print("---- Start free range at 0x%x with free block (%d)\n", 7985 freeFinger, freeRangeInFreeLists); 7986 } 7987 assert(!inFreeRange(), "Trampling existing free range"); 7988 set_inFreeRange(true); 7989 set_lastFreeRangeCoalesced(false); 7990 7991 set_freeFinger(freeFinger); 7992 set_freeRangeInFreeLists(freeRangeInFreeLists); 7993 if (CMSTestInFreeList) { 7994 if (freeRangeInFreeLists) { 7995 FreeChunk* fc = (FreeChunk*) freeFinger; 7996 assert(fc->is_free(), "A chunk on the free list should be free."); 7997 assert(fc->size() > 0, "Free range should have a size"); 7998 assert(_sp->verify_chunk_in_free_list(fc), "Chunk is not in free lists"); 7999 } 8000 } 8001 } 8002 8003 // Note that the sweeper runs concurrently with mutators. Thus, 8004 // it is possible for direct allocation in this generation to happen 8005 // in the middle of the sweep. Note that the sweeper also coalesces 8006 // contiguous free blocks. Thus, unless the sweeper and the allocator 8007 // synchronize appropriately freshly allocated blocks may get swept up. 8008 // This is accomplished by the sweeper locking the free lists while 8009 // it is sweeping. Thus blocks that are determined to be free are 8010 // indeed free. There is however one additional complication: 8011 // blocks that have been allocated since the final checkpoint and 8012 // mark, will not have been marked and so would be treated as 8013 // unreachable and swept up. To prevent this, the allocator marks 8014 // the bit map when allocating during the sweep phase. This leads, 8015 // however, to a further complication -- objects may have been allocated 8016 // but not yet initialized -- in the sense that the header isn't yet 8017 // installed. The sweeper can not then determine the size of the block 8018 // in order to skip over it. To deal with this case, we use a technique 8019 // (due to Printezis) to encode such uninitialized block sizes in the 8020 // bit map. Since the bit map uses a bit per every HeapWord, but the 8021 // CMS generation has a minimum object size of 3 HeapWords, it follows 8022 // that "normal marks" won't be adjacent in the bit map (there will 8023 // always be at least two 0 bits between successive 1 bits). We make use 8024 // of these "unused" bits to represent uninitialized blocks -- the bit 8025 // corresponding to the start of the uninitialized object and the next 8026 // bit are both set. Finally, a 1 bit marks the end of the object that 8027 // started with the two consecutive 1 bits to indicate its potentially 8028 // uninitialized state. 8029 8030 size_t SweepClosure::do_blk_careful(HeapWord* addr) { 8031 FreeChunk* fc = (FreeChunk*)addr; 8032 size_t res; 8033 8034 // Check if we are done sweeping. Below we check "addr >= _limit" rather 8035 // than "addr == _limit" because although _limit was a block boundary when 8036 // we started the sweep, it may no longer be one because heap expansion 8037 // may have caused us to coalesce the block ending at the address _limit 8038 // with a newly expanded chunk (this happens when _limit was set to the 8039 // previous _end of the space), so we may have stepped past _limit: 8040 // see the following Zeno-like trail of CRs 6977970, 7008136, 7042740. 8041 if (addr >= _limit) { // we have swept up to or past the limit: finish up 8042 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 8043 "sweep _limit out of bounds"); 8044 assert(addr < _sp->end(), "addr out of bounds"); 8045 // Flush any free range we might be holding as a single 8046 // coalesced chunk to the appropriate free list. 8047 if (inFreeRange()) { 8048 assert(freeFinger() >= _sp->bottom() && freeFinger() < _limit, 8049 err_msg("freeFinger() " PTR_FORMAT" is out-of-bounds", freeFinger())); 8050 flush_cur_free_chunk(freeFinger(), 8051 pointer_delta(addr, freeFinger())); 8052 if (CMSTraceSweeper) { 8053 gclog_or_tty->print("Sweep: last chunk: "); 8054 gclog_or_tty->print("put_free_blk 0x%x ("SIZE_FORMAT") " 8055 "[coalesced:"SIZE_FORMAT"]\n", 8056 freeFinger(), pointer_delta(addr, freeFinger()), 8057 lastFreeRangeCoalesced()); 8058 } 8059 } 8060 8061 // help the iterator loop finish 8062 return pointer_delta(_sp->end(), addr); 8063 } 8064 8065 assert(addr < _limit, "sweep invariant"); 8066 // check if we should yield 8067 do_yield_check(addr); 8068 if (fc->is_free()) { 8069 // Chunk that is already free 8070 res = fc->size(); 8071 do_already_free_chunk(fc); 8072 debug_only(_sp->verifyFreeLists()); 8073 // If we flush the chunk at hand in lookahead_and_flush() 8074 // and it's coalesced with a preceding chunk, then the 8075 // process of "mangling" the payload of the coalesced block 8076 // will cause erasure of the size information from the 8077 // (erstwhile) header of all the coalesced blocks but the 8078 // first, so the first disjunct in the assert will not hold 8079 // in that specific case (in which case the second disjunct 8080 // will hold). 8081 assert(res == fc->size() || ((HeapWord*)fc) + res >= _limit, 8082 "Otherwise the size info doesn't change at this step"); 8083 NOT_PRODUCT( 8084 _numObjectsAlreadyFree++; 8085 _numWordsAlreadyFree += res; 8086 ) 8087 NOT_PRODUCT(_last_fc = fc;) 8088 } else if (!_bitMap->isMarked(addr)) { 8089 // Chunk is fresh garbage 8090 res = do_garbage_chunk(fc); 8091 debug_only(_sp->verifyFreeLists()); 8092 NOT_PRODUCT( 8093 _numObjectsFreed++; 8094 _numWordsFreed += res; 8095 ) 8096 } else { 8097 // Chunk that is alive. 8098 res = do_live_chunk(fc); 8099 debug_only(_sp->verifyFreeLists()); 8100 NOT_PRODUCT( 8101 _numObjectsLive++; 8102 _numWordsLive += res; 8103 ) 8104 } 8105 return res; 8106 } 8107 8108 // For the smart allocation, record following 8109 // split deaths - a free chunk is removed from its free list because 8110 // it is being split into two or more chunks. 8111 // split birth - a free chunk is being added to its free list because 8112 // a larger free chunk has been split and resulted in this free chunk. 8113 // coal death - a free chunk is being removed from its free list because 8114 // it is being coalesced into a large free chunk. 8115 // coal birth - a free chunk is being added to its free list because 8116 // it was created when two or more free chunks where coalesced into 8117 // this free chunk. 8118 // 8119 // These statistics are used to determine the desired number of free 8120 // chunks of a given size. The desired number is chosen to be relative 8121 // to the end of a CMS sweep. The desired number at the end of a sweep 8122 // is the 8123 // count-at-end-of-previous-sweep (an amount that was enough) 8124 // - count-at-beginning-of-current-sweep (the excess) 8125 // + split-births (gains in this size during interval) 8126 // - split-deaths (demands on this size during interval) 8127 // where the interval is from the end of one sweep to the end of the 8128 // next. 8129 // 8130 // When sweeping the sweeper maintains an accumulated chunk which is 8131 // the chunk that is made up of chunks that have been coalesced. That 8132 // will be termed the left-hand chunk. A new chunk of garbage that 8133 // is being considered for coalescing will be referred to as the 8134 // right-hand chunk. 8135 // 8136 // When making a decision on whether to coalesce a right-hand chunk with 8137 // the current left-hand chunk, the current count vs. the desired count 8138 // of the left-hand chunk is considered. Also if the right-hand chunk 8139 // is near the large chunk at the end of the heap (see 8140 // ConcurrentMarkSweepGeneration::isNearLargestChunk()), then the 8141 // left-hand chunk is coalesced. 8142 // 8143 // When making a decision about whether to split a chunk, the desired count 8144 // vs. the current count of the candidate to be split is also considered. 8145 // If the candidate is underpopulated (currently fewer chunks than desired) 8146 // a chunk of an overpopulated (currently more chunks than desired) size may 8147 // be chosen. The "hint" associated with a free list, if non-null, points 8148 // to a free list which may be overpopulated. 8149 // 8150 8151 void SweepClosure::do_already_free_chunk(FreeChunk* fc) { 8152 const size_t size = fc->size(); 8153 // Chunks that cannot be coalesced are not in the 8154 // free lists. 8155 if (CMSTestInFreeList && !fc->cantCoalesce()) { 8156 assert(_sp->verify_chunk_in_free_list(fc), 8157 "free chunk should be in free lists"); 8158 } 8159 // a chunk that is already free, should not have been 8160 // marked in the bit map 8161 HeapWord* const addr = (HeapWord*) fc; 8162 assert(!_bitMap->isMarked(addr), "free chunk should be unmarked"); 8163 // Verify that the bit map has no bits marked between 8164 // addr and purported end of this block. 8165 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 8166 8167 // Some chunks cannot be coalesced under any circumstances. 8168 // See the definition of cantCoalesce(). 8169 if (!fc->cantCoalesce()) { 8170 // This chunk can potentially be coalesced. 8171 if (_sp->adaptive_freelists()) { 8172 // All the work is done in 8173 do_post_free_or_garbage_chunk(fc, size); 8174 } else { // Not adaptive free lists 8175 // this is a free chunk that can potentially be coalesced by the sweeper; 8176 if (!inFreeRange()) { 8177 // if the next chunk is a free block that can't be coalesced 8178 // it doesn't make sense to remove this chunk from the free lists 8179 FreeChunk* nextChunk = (FreeChunk*)(addr + size); 8180 assert((HeapWord*)nextChunk <= _sp->end(), "Chunk size out of bounds?"); 8181 if ((HeapWord*)nextChunk < _sp->end() && // There is another free chunk to the right ... 8182 nextChunk->is_free() && // ... which is free... 8183 nextChunk->cantCoalesce()) { // ... but can't be coalesced 8184 // nothing to do 8185 } else { 8186 // Potentially the start of a new free range: 8187 // Don't eagerly remove it from the free lists. 8188 // No need to remove it if it will just be put 8189 // back again. (Also from a pragmatic point of view 8190 // if it is a free block in a region that is beyond 8191 // any allocated blocks, an assertion will fail) 8192 // Remember the start of a free run. 8193 initialize_free_range(addr, true); 8194 // end - can coalesce with next chunk 8195 } 8196 } else { 8197 // the midst of a free range, we are coalescing 8198 print_free_block_coalesced(fc); 8199 if (CMSTraceSweeper) { 8200 gclog_or_tty->print(" -- pick up free block 0x%x (%d)\n", fc, size); 8201 } 8202 // remove it from the free lists 8203 _sp->removeFreeChunkFromFreeLists(fc); 8204 set_lastFreeRangeCoalesced(true); 8205 // If the chunk is being coalesced and the current free range is 8206 // in the free lists, remove the current free range so that it 8207 // will be returned to the free lists in its entirety - all 8208 // the coalesced pieces included. 8209 if (freeRangeInFreeLists()) { 8210 FreeChunk* ffc = (FreeChunk*) freeFinger(); 8211 assert(ffc->size() == pointer_delta(addr, freeFinger()), 8212 "Size of free range is inconsistent with chunk size."); 8213 if (CMSTestInFreeList) { 8214 assert(_sp->verify_chunk_in_free_list(ffc), 8215 "free range is not in free lists"); 8216 } 8217 _sp->removeFreeChunkFromFreeLists(ffc); 8218 set_freeRangeInFreeLists(false); 8219 } 8220 } 8221 } 8222 // Note that if the chunk is not coalescable (the else arm 8223 // below), we unconditionally flush, without needing to do 8224 // a "lookahead," as we do below. 8225 if (inFreeRange()) lookahead_and_flush(fc, size); 8226 } else { 8227 // Code path common to both original and adaptive free lists. 8228 8229 // cant coalesce with previous block; this should be treated 8230 // as the end of a free run if any 8231 if (inFreeRange()) { 8232 // we kicked some butt; time to pick up the garbage 8233 assert(freeFinger() < addr, "freeFinger points too high"); 8234 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 8235 } 8236 // else, nothing to do, just continue 8237 } 8238 } 8239 8240 size_t SweepClosure::do_garbage_chunk(FreeChunk* fc) { 8241 // This is a chunk of garbage. It is not in any free list. 8242 // Add it to a free list or let it possibly be coalesced into 8243 // a larger chunk. 8244 HeapWord* const addr = (HeapWord*) fc; 8245 const size_t size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 8246 8247 if (_sp->adaptive_freelists()) { 8248 // Verify that the bit map has no bits marked between 8249 // addr and purported end of just dead object. 8250 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 8251 8252 do_post_free_or_garbage_chunk(fc, size); 8253 } else { 8254 if (!inFreeRange()) { 8255 // start of a new free range 8256 assert(size > 0, "A free range should have a size"); 8257 initialize_free_range(addr, false); 8258 } else { 8259 // this will be swept up when we hit the end of the 8260 // free range 8261 if (CMSTraceSweeper) { 8262 gclog_or_tty->print(" -- pick up garbage 0x%x (%d) \n", fc, size); 8263 } 8264 // If the chunk is being coalesced and the current free range is 8265 // in the free lists, remove the current free range so that it 8266 // will be returned to the free lists in its entirety - all 8267 // the coalesced pieces included. 8268 if (freeRangeInFreeLists()) { 8269 FreeChunk* ffc = (FreeChunk*)freeFinger(); 8270 assert(ffc->size() == pointer_delta(addr, freeFinger()), 8271 "Size of free range is inconsistent with chunk size."); 8272 if (CMSTestInFreeList) { 8273 assert(_sp->verify_chunk_in_free_list(ffc), 8274 "free range is not in free lists"); 8275 } 8276 _sp->removeFreeChunkFromFreeLists(ffc); 8277 set_freeRangeInFreeLists(false); 8278 } 8279 set_lastFreeRangeCoalesced(true); 8280 } 8281 // this will be swept up when we hit the end of the free range 8282 8283 // Verify that the bit map has no bits marked between 8284 // addr and purported end of just dead object. 8285 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 8286 } 8287 assert(_limit >= addr + size, 8288 "A freshly garbage chunk can't possibly straddle over _limit"); 8289 if (inFreeRange()) lookahead_and_flush(fc, size); 8290 return size; 8291 } 8292 8293 size_t SweepClosure::do_live_chunk(FreeChunk* fc) { 8294 HeapWord* addr = (HeapWord*) fc; 8295 // The sweeper has just found a live object. Return any accumulated 8296 // left hand chunk to the free lists. 8297 if (inFreeRange()) { 8298 assert(freeFinger() < addr, "freeFinger points too high"); 8299 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 8300 } 8301 8302 // This object is live: we'd normally expect this to be 8303 // an oop, and like to assert the following: 8304 // assert(oop(addr)->is_oop(), "live block should be an oop"); 8305 // However, as we commented above, this may be an object whose 8306 // header hasn't yet been initialized. 8307 size_t size; 8308 assert(_bitMap->isMarked(addr), "Tautology for this control point"); 8309 if (_bitMap->isMarked(addr + 1)) { 8310 // Determine the size from the bit map, rather than trying to 8311 // compute it from the object header. 8312 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 8313 size = pointer_delta(nextOneAddr + 1, addr); 8314 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 8315 "alignment problem"); 8316 8317 #ifdef ASSERT 8318 if (oop(addr)->klass_or_null() != NULL) { 8319 // Ignore mark word because we are running concurrent with mutators 8320 assert(oop(addr)->is_oop(true), "live block should be an oop"); 8321 assert(size == 8322 CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()), 8323 "P-mark and computed size do not agree"); 8324 } 8325 #endif 8326 8327 } else { 8328 // This should be an initialized object that's alive. 8329 assert(oop(addr)->klass_or_null() != NULL, 8330 "Should be an initialized object"); 8331 // Ignore mark word because we are running concurrent with mutators 8332 assert(oop(addr)->is_oop(true), "live block should be an oop"); 8333 // Verify that the bit map has no bits marked between 8334 // addr and purported end of this block. 8335 size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 8336 assert(size >= 3, "Necessary for Printezis marks to work"); 8337 assert(!_bitMap->isMarked(addr+1), "Tautology for this control point"); 8338 DEBUG_ONLY(_bitMap->verifyNoOneBitsInRange(addr+2, addr+size);) 8339 } 8340 return size; 8341 } 8342 8343 void SweepClosure::do_post_free_or_garbage_chunk(FreeChunk* fc, 8344 size_t chunkSize) { 8345 // do_post_free_or_garbage_chunk() should only be called in the case 8346 // of the adaptive free list allocator. 8347 const bool fcInFreeLists = fc->is_free(); 8348 assert(_sp->adaptive_freelists(), "Should only be used in this case."); 8349 assert((HeapWord*)fc <= _limit, "sweep invariant"); 8350 if (CMSTestInFreeList && fcInFreeLists) { 8351 assert(_sp->verify_chunk_in_free_list(fc), "free chunk is not in free lists"); 8352 } 8353 8354 if (CMSTraceSweeper) { 8355 gclog_or_tty->print_cr(" -- pick up another chunk at 0x%x (%d)", fc, chunkSize); 8356 } 8357 8358 HeapWord* const fc_addr = (HeapWord*) fc; 8359 8360 bool coalesce; 8361 const size_t left = pointer_delta(fc_addr, freeFinger()); 8362 const size_t right = chunkSize; 8363 switch (FLSCoalescePolicy) { 8364 // numeric value forms a coalition aggressiveness metric 8365 case 0: { // never coalesce 8366 coalesce = false; 8367 break; 8368 } 8369 case 1: { // coalesce if left & right chunks on overpopulated lists 8370 coalesce = _sp->coalOverPopulated(left) && 8371 _sp->coalOverPopulated(right); 8372 break; 8373 } 8374 case 2: { // coalesce if left chunk on overpopulated list (default) 8375 coalesce = _sp->coalOverPopulated(left); 8376 break; 8377 } 8378 case 3: { // coalesce if left OR right chunk on overpopulated list 8379 coalesce = _sp->coalOverPopulated(left) || 8380 _sp->coalOverPopulated(right); 8381 break; 8382 } 8383 case 4: { // always coalesce 8384 coalesce = true; 8385 break; 8386 } 8387 default: 8388 ShouldNotReachHere(); 8389 } 8390 8391 // Should the current free range be coalesced? 8392 // If the chunk is in a free range and either we decided to coalesce above 8393 // or the chunk is near the large block at the end of the heap 8394 // (isNearLargestChunk() returns true), then coalesce this chunk. 8395 const bool doCoalesce = inFreeRange() 8396 && (coalesce || _g->isNearLargestChunk(fc_addr)); 8397 if (doCoalesce) { 8398 // Coalesce the current free range on the left with the new 8399 // chunk on the right. If either is on a free list, 8400 // it must be removed from the list and stashed in the closure. 8401 if (freeRangeInFreeLists()) { 8402 FreeChunk* const ffc = (FreeChunk*)freeFinger(); 8403 assert(ffc->size() == pointer_delta(fc_addr, freeFinger()), 8404 "Size of free range is inconsistent with chunk size."); 8405 if (CMSTestInFreeList) { 8406 assert(_sp->verify_chunk_in_free_list(ffc), 8407 "Chunk is not in free lists"); 8408 } 8409 _sp->coalDeath(ffc->size()); 8410 _sp->removeFreeChunkFromFreeLists(ffc); 8411 set_freeRangeInFreeLists(false); 8412 } 8413 if (fcInFreeLists) { 8414 _sp->coalDeath(chunkSize); 8415 assert(fc->size() == chunkSize, 8416 "The chunk has the wrong size or is not in the free lists"); 8417 _sp->removeFreeChunkFromFreeLists(fc); 8418 } 8419 set_lastFreeRangeCoalesced(true); 8420 print_free_block_coalesced(fc); 8421 } else { // not in a free range and/or should not coalesce 8422 // Return the current free range and start a new one. 8423 if (inFreeRange()) { 8424 // In a free range but cannot coalesce with the right hand chunk. 8425 // Put the current free range into the free lists. 8426 flush_cur_free_chunk(freeFinger(), 8427 pointer_delta(fc_addr, freeFinger())); 8428 } 8429 // Set up for new free range. Pass along whether the right hand 8430 // chunk is in the free lists. 8431 initialize_free_range((HeapWord*)fc, fcInFreeLists); 8432 } 8433 } 8434 8435 // Lookahead flush: 8436 // If we are tracking a free range, and this is the last chunk that 8437 // we'll look at because its end crosses past _limit, we'll preemptively 8438 // flush it along with any free range we may be holding on to. Note that 8439 // this can be the case only for an already free or freshly garbage 8440 // chunk. If this block is an object, it can never straddle 8441 // over _limit. The "straddling" occurs when _limit is set at 8442 // the previous end of the space when this cycle started, and 8443 // a subsequent heap expansion caused the previously co-terminal 8444 // free block to be coalesced with the newly expanded portion, 8445 // thus rendering _limit a non-block-boundary making it dangerous 8446 // for the sweeper to step over and examine. 8447 void SweepClosure::lookahead_and_flush(FreeChunk* fc, size_t chunk_size) { 8448 assert(inFreeRange(), "Should only be called if currently in a free range."); 8449 HeapWord* const eob = ((HeapWord*)fc) + chunk_size; 8450 assert(_sp->used_region().contains(eob - 1), 8451 err_msg("eob = " PTR_FORMAT " out of bounds wrt _sp = [" PTR_FORMAT "," PTR_FORMAT ")" 8452 " when examining fc = " PTR_FORMAT "(" SIZE_FORMAT ")", 8453 _limit, _sp->bottom(), _sp->end(), fc, chunk_size)); 8454 if (eob >= _limit) { 8455 assert(eob == _limit || fc->is_free(), "Only a free chunk should allow us to cross over the limit"); 8456 if (CMSTraceSweeper) { 8457 gclog_or_tty->print_cr("_limit " PTR_FORMAT " reached or crossed by block " 8458 "[" PTR_FORMAT "," PTR_FORMAT ") in space " 8459 "[" PTR_FORMAT "," PTR_FORMAT ")", 8460 _limit, fc, eob, _sp->bottom(), _sp->end()); 8461 } 8462 // Return the storage we are tracking back into the free lists. 8463 if (CMSTraceSweeper) { 8464 gclog_or_tty->print_cr("Flushing ... "); 8465 } 8466 assert(freeFinger() < eob, "Error"); 8467 flush_cur_free_chunk( freeFinger(), pointer_delta(eob, freeFinger())); 8468 } 8469 } 8470 8471 void SweepClosure::flush_cur_free_chunk(HeapWord* chunk, size_t size) { 8472 assert(inFreeRange(), "Should only be called if currently in a free range."); 8473 assert(size > 0, 8474 "A zero sized chunk cannot be added to the free lists."); 8475 if (!freeRangeInFreeLists()) { 8476 if (CMSTestInFreeList) { 8477 FreeChunk* fc = (FreeChunk*) chunk; 8478 fc->set_size(size); 8479 assert(!_sp->verify_chunk_in_free_list(fc), 8480 "chunk should not be in free lists yet"); 8481 } 8482 if (CMSTraceSweeper) { 8483 gclog_or_tty->print_cr(" -- add free block 0x%x (%d) to free lists", 8484 chunk, size); 8485 } 8486 // A new free range is going to be starting. The current 8487 // free range has not been added to the free lists yet or 8488 // was removed so add it back. 8489 // If the current free range was coalesced, then the death 8490 // of the free range was recorded. Record a birth now. 8491 if (lastFreeRangeCoalesced()) { 8492 _sp->coalBirth(size); 8493 } 8494 _sp->addChunkAndRepairOffsetTable(chunk, size, 8495 lastFreeRangeCoalesced()); 8496 } else if (CMSTraceSweeper) { 8497 gclog_or_tty->print_cr("Already in free list: nothing to flush"); 8498 } 8499 set_inFreeRange(false); 8500 set_freeRangeInFreeLists(false); 8501 } 8502 8503 // We take a break if we've been at this for a while, 8504 // so as to avoid monopolizing the locks involved. 8505 void SweepClosure::do_yield_work(HeapWord* addr) { 8506 // Return current free chunk being used for coalescing (if any) 8507 // to the appropriate freelist. After yielding, the next 8508 // free block encountered will start a coalescing range of 8509 // free blocks. If the next free block is adjacent to the 8510 // chunk just flushed, they will need to wait for the next 8511 // sweep to be coalesced. 8512 if (inFreeRange()) { 8513 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 8514 } 8515 8516 // First give up the locks, then yield, then re-lock. 8517 // We should probably use a constructor/destructor idiom to 8518 // do this unlock/lock or modify the MutexUnlocker class to 8519 // serve our purpose. XXX 8520 assert_lock_strong(_bitMap->lock()); 8521 assert_lock_strong(_freelistLock); 8522 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 8523 "CMS thread should hold CMS token"); 8524 _bitMap->lock()->unlock(); 8525 _freelistLock->unlock(); 8526 ConcurrentMarkSweepThread::desynchronize(true); 8527 ConcurrentMarkSweepThread::acknowledge_yield_request(); 8528 _collector->stopTimer(); 8529 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 8530 if (PrintCMSStatistics != 0) { 8531 _collector->incrementYields(); 8532 } 8533 _collector->icms_wait(); 8534 8535 // See the comment in coordinator_yield() 8536 for (unsigned i = 0; i < CMSYieldSleepCount && 8537 ConcurrentMarkSweepThread::should_yield() && 8538 !CMSCollector::foregroundGCIsActive(); ++i) { 8539 os::sleep(Thread::current(), 1, false); 8540 ConcurrentMarkSweepThread::acknowledge_yield_request(); 8541 } 8542 8543 ConcurrentMarkSweepThread::synchronize(true); 8544 _freelistLock->lock(); 8545 _bitMap->lock()->lock_without_safepoint_check(); 8546 _collector->startTimer(); 8547 } 8548 8549 #ifndef PRODUCT 8550 // This is actually very useful in a product build if it can 8551 // be called from the debugger. Compile it into the product 8552 // as needed. 8553 bool debug_verify_chunk_in_free_list(FreeChunk* fc) { 8554 return debug_cms_space->verify_chunk_in_free_list(fc); 8555 } 8556 #endif 8557 8558 void SweepClosure::print_free_block_coalesced(FreeChunk* fc) const { 8559 if (CMSTraceSweeper) { 8560 gclog_or_tty->print_cr("Sweep:coal_free_blk " PTR_FORMAT " (" SIZE_FORMAT ")", 8561 fc, fc->size()); 8562 } 8563 } 8564 8565 // CMSIsAliveClosure 8566 bool CMSIsAliveClosure::do_object_b(oop obj) { 8567 HeapWord* addr = (HeapWord*)obj; 8568 return addr != NULL && 8569 (!_span.contains(addr) || _bit_map->isMarked(addr)); 8570 } 8571 8572 8573 CMSKeepAliveClosure::CMSKeepAliveClosure( CMSCollector* collector, 8574 MemRegion span, 8575 CMSBitMap* bit_map, CMSMarkStack* mark_stack, 8576 bool cpc): 8577 _collector(collector), 8578 _span(span), 8579 _bit_map(bit_map), 8580 _mark_stack(mark_stack), 8581 _concurrent_precleaning(cpc) { 8582 assert(!_span.is_empty(), "Empty span could spell trouble"); 8583 } 8584 8585 8586 // CMSKeepAliveClosure: the serial version 8587 void CMSKeepAliveClosure::do_oop(oop obj) { 8588 HeapWord* addr = (HeapWord*)obj; 8589 if (_span.contains(addr) && 8590 !_bit_map->isMarked(addr)) { 8591 _bit_map->mark(addr); 8592 bool simulate_overflow = false; 8593 NOT_PRODUCT( 8594 if (CMSMarkStackOverflowALot && 8595 _collector->simulate_overflow()) { 8596 // simulate a stack overflow 8597 simulate_overflow = true; 8598 } 8599 ) 8600 if (simulate_overflow || !_mark_stack->push(obj)) { 8601 if (_concurrent_precleaning) { 8602 // We dirty the overflown object and let the remark 8603 // phase deal with it. 8604 assert(_collector->overflow_list_is_empty(), "Error"); 8605 // In the case of object arrays, we need to dirty all of 8606 // the cards that the object spans. No locking or atomics 8607 // are needed since no one else can be mutating the mod union 8608 // table. 8609 if (obj->is_objArray()) { 8610 size_t sz = obj->size(); 8611 HeapWord* end_card_addr = 8612 (HeapWord*)round_to((intptr_t)(addr+sz), CardTableModRefBS::card_size); 8613 MemRegion redirty_range = MemRegion(addr, end_card_addr); 8614 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 8615 _collector->_modUnionTable.mark_range(redirty_range); 8616 } else { 8617 _collector->_modUnionTable.mark(addr); 8618 } 8619 _collector->_ser_kac_preclean_ovflw++; 8620 } else { 8621 _collector->push_on_overflow_list(obj); 8622 _collector->_ser_kac_ovflw++; 8623 } 8624 } 8625 } 8626 } 8627 8628 void CMSKeepAliveClosure::do_oop(oop* p) { CMSKeepAliveClosure::do_oop_work(p); } 8629 void CMSKeepAliveClosure::do_oop(narrowOop* p) { CMSKeepAliveClosure::do_oop_work(p); } 8630 8631 // CMSParKeepAliveClosure: a parallel version of the above. 8632 // The work queues are private to each closure (thread), 8633 // but (may be) available for stealing by other threads. 8634 void CMSParKeepAliveClosure::do_oop(oop obj) { 8635 HeapWord* addr = (HeapWord*)obj; 8636 if (_span.contains(addr) && 8637 !_bit_map->isMarked(addr)) { 8638 // In general, during recursive tracing, several threads 8639 // may be concurrently getting here; the first one to 8640 // "tag" it, claims it. 8641 if (_bit_map->par_mark(addr)) { 8642 bool res = _work_queue->push(obj); 8643 assert(res, "Low water mark should be much less than capacity"); 8644 // Do a recursive trim in the hope that this will keep 8645 // stack usage lower, but leave some oops for potential stealers 8646 trim_queue(_low_water_mark); 8647 } // Else, another thread got there first 8648 } 8649 } 8650 8651 void CMSParKeepAliveClosure::do_oop(oop* p) { CMSParKeepAliveClosure::do_oop_work(p); } 8652 void CMSParKeepAliveClosure::do_oop(narrowOop* p) { CMSParKeepAliveClosure::do_oop_work(p); } 8653 8654 void CMSParKeepAliveClosure::trim_queue(uint max) { 8655 while (_work_queue->size() > max) { 8656 oop new_oop; 8657 if (_work_queue->pop_local(new_oop)) { 8658 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 8659 assert(_bit_map->isMarked((HeapWord*)new_oop), 8660 "no white objects on this stack!"); 8661 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 8662 // iterate over the oops in this oop, marking and pushing 8663 // the ones in CMS heap (i.e. in _span). 8664 new_oop->oop_iterate(&_mark_and_push); 8665 } 8666 } 8667 } 8668 8669 CMSInnerParMarkAndPushClosure::CMSInnerParMarkAndPushClosure( 8670 CMSCollector* collector, 8671 MemRegion span, CMSBitMap* bit_map, 8672 OopTaskQueue* work_queue): 8673 _collector(collector), 8674 _span(span), 8675 _bit_map(bit_map), 8676 _work_queue(work_queue) { } 8677 8678 void CMSInnerParMarkAndPushClosure::do_oop(oop obj) { 8679 HeapWord* addr = (HeapWord*)obj; 8680 if (_span.contains(addr) && 8681 !_bit_map->isMarked(addr)) { 8682 if (_bit_map->par_mark(addr)) { 8683 bool simulate_overflow = false; 8684 NOT_PRODUCT( 8685 if (CMSMarkStackOverflowALot && 8686 _collector->par_simulate_overflow()) { 8687 // simulate a stack overflow 8688 simulate_overflow = true; 8689 } 8690 ) 8691 if (simulate_overflow || !_work_queue->push(obj)) { 8692 _collector->par_push_on_overflow_list(obj); 8693 _collector->_par_kac_ovflw++; 8694 } 8695 } // Else another thread got there already 8696 } 8697 } 8698 8699 void CMSInnerParMarkAndPushClosure::do_oop(oop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); } 8700 void CMSInnerParMarkAndPushClosure::do_oop(narrowOop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); } 8701 8702 ////////////////////////////////////////////////////////////////// 8703 // CMSExpansionCause ///////////////////////////// 8704 ////////////////////////////////////////////////////////////////// 8705 const char* CMSExpansionCause::to_string(CMSExpansionCause::Cause cause) { 8706 switch (cause) { 8707 case _no_expansion: 8708 return "No expansion"; 8709 case _satisfy_free_ratio: 8710 return "Free ratio"; 8711 case _satisfy_promotion: 8712 return "Satisfy promotion"; 8713 case _satisfy_allocation: 8714 return "allocation"; 8715 case _allocate_par_lab: 8716 return "Par LAB"; 8717 case _allocate_par_spooling_space: 8718 return "Par Spooling Space"; 8719 case _adaptive_size_policy: 8720 return "Ergonomics"; 8721 default: 8722 return "unknown"; 8723 } 8724 } 8725 8726 void CMSDrainMarkingStackClosure::do_void() { 8727 // the max number to take from overflow list at a time 8728 const size_t num = _mark_stack->capacity()/4; 8729 assert(!_concurrent_precleaning || _collector->overflow_list_is_empty(), 8730 "Overflow list should be NULL during concurrent phases"); 8731 while (!_mark_stack->isEmpty() || 8732 // if stack is empty, check the overflow list 8733 _collector->take_from_overflow_list(num, _mark_stack)) { 8734 oop obj = _mark_stack->pop(); 8735 HeapWord* addr = (HeapWord*)obj; 8736 assert(_span.contains(addr), "Should be within span"); 8737 assert(_bit_map->isMarked(addr), "Should be marked"); 8738 assert(obj->is_oop(), "Should be an oop"); 8739 obj->oop_iterate(_keep_alive); 8740 } 8741 } 8742 8743 void CMSParDrainMarkingStackClosure::do_void() { 8744 // drain queue 8745 trim_queue(0); 8746 } 8747 8748 // Trim our work_queue so its length is below max at return 8749 void CMSParDrainMarkingStackClosure::trim_queue(uint max) { 8750 while (_work_queue->size() > max) { 8751 oop new_oop; 8752 if (_work_queue->pop_local(new_oop)) { 8753 assert(new_oop->is_oop(), "Expected an oop"); 8754 assert(_bit_map->isMarked((HeapWord*)new_oop), 8755 "no white objects on this stack!"); 8756 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 8757 // iterate over the oops in this oop, marking and pushing 8758 // the ones in CMS heap (i.e. in _span). 8759 new_oop->oop_iterate(&_mark_and_push); 8760 } 8761 } 8762 } 8763 8764 //////////////////////////////////////////////////////////////////// 8765 // Support for Marking Stack Overflow list handling and related code 8766 //////////////////////////////////////////////////////////////////// 8767 // Much of the following code is similar in shape and spirit to the 8768 // code used in ParNewGC. We should try and share that code 8769 // as much as possible in the future. 8770 8771 #ifndef PRODUCT 8772 // Debugging support for CMSStackOverflowALot 8773 8774 // It's OK to call this multi-threaded; the worst thing 8775 // that can happen is that we'll get a bunch of closely 8776 // spaced simulated oveflows, but that's OK, in fact 8777 // probably good as it would exercise the overflow code 8778 // under contention. 8779 bool CMSCollector::simulate_overflow() { 8780 if (_overflow_counter-- <= 0) { // just being defensive 8781 _overflow_counter = CMSMarkStackOverflowInterval; 8782 return true; 8783 } else { 8784 return false; 8785 } 8786 } 8787 8788 bool CMSCollector::par_simulate_overflow() { 8789 return simulate_overflow(); 8790 } 8791 #endif 8792 8793 // Single-threaded 8794 bool CMSCollector::take_from_overflow_list(size_t num, CMSMarkStack* stack) { 8795 assert(stack->isEmpty(), "Expected precondition"); 8796 assert(stack->capacity() > num, "Shouldn't bite more than can chew"); 8797 size_t i = num; 8798 oop cur = _overflow_list; 8799 const markOop proto = markOopDesc::prototype(); 8800 NOT_PRODUCT(ssize_t n = 0;) 8801 for (oop next; i > 0 && cur != NULL; cur = next, i--) { 8802 next = oop(cur->mark()); 8803 cur->set_mark(proto); // until proven otherwise 8804 assert(cur->is_oop(), "Should be an oop"); 8805 bool res = stack->push(cur); 8806 assert(res, "Bit off more than can chew?"); 8807 NOT_PRODUCT(n++;) 8808 } 8809 _overflow_list = cur; 8810 #ifndef PRODUCT 8811 assert(_num_par_pushes >= n, "Too many pops?"); 8812 _num_par_pushes -=n; 8813 #endif 8814 return !stack->isEmpty(); 8815 } 8816 8817 #define BUSY (oop(0x1aff1aff)) 8818 // (MT-safe) Get a prefix of at most "num" from the list. 8819 // The overflow list is chained through the mark word of 8820 // each object in the list. We fetch the entire list, 8821 // break off a prefix of the right size and return the 8822 // remainder. If other threads try to take objects from 8823 // the overflow list at that time, they will wait for 8824 // some time to see if data becomes available. If (and 8825 // only if) another thread places one or more object(s) 8826 // on the global list before we have returned the suffix 8827 // to the global list, we will walk down our local list 8828 // to find its end and append the global list to 8829 // our suffix before returning it. This suffix walk can 8830 // prove to be expensive (quadratic in the amount of traffic) 8831 // when there are many objects in the overflow list and 8832 // there is much producer-consumer contention on the list. 8833 // *NOTE*: The overflow list manipulation code here and 8834 // in ParNewGeneration:: are very similar in shape, 8835 // except that in the ParNew case we use the old (from/eden) 8836 // copy of the object to thread the list via its klass word. 8837 // Because of the common code, if you make any changes in 8838 // the code below, please check the ParNew version to see if 8839 // similar changes might be needed. 8840 // CR 6797058 has been filed to consolidate the common code. 8841 bool CMSCollector::par_take_from_overflow_list(size_t num, 8842 OopTaskQueue* work_q, 8843 int no_of_gc_threads) { 8844 assert(work_q->size() == 0, "First empty local work queue"); 8845 assert(num < work_q->max_elems(), "Can't bite more than we can chew"); 8846 if (_overflow_list == NULL) { 8847 return false; 8848 } 8849 // Grab the entire list; we'll put back a suffix 8850 oop prefix = (oop)Atomic::xchg_ptr(BUSY, &_overflow_list); 8851 Thread* tid = Thread::current(); 8852 // Before "no_of_gc_threads" was introduced CMSOverflowSpinCount was 8853 // set to ParallelGCThreads. 8854 size_t CMSOverflowSpinCount = (size_t) no_of_gc_threads; // was ParallelGCThreads; 8855 size_t sleep_time_millis = MAX2((size_t)1, num/100); 8856 // If the list is busy, we spin for a short while, 8857 // sleeping between attempts to get the list. 8858 for (size_t spin = 0; prefix == BUSY && spin < CMSOverflowSpinCount; spin++) { 8859 os::sleep(tid, sleep_time_millis, false); 8860 if (_overflow_list == NULL) { 8861 // Nothing left to take 8862 return false; 8863 } else if (_overflow_list != BUSY) { 8864 // Try and grab the prefix 8865 prefix = (oop)Atomic::xchg_ptr(BUSY, &_overflow_list); 8866 } 8867 } 8868 // If the list was found to be empty, or we spun long 8869 // enough, we give up and return empty-handed. If we leave 8870 // the list in the BUSY state below, it must be the case that 8871 // some other thread holds the overflow list and will set it 8872 // to a non-BUSY state in the future. 8873 if (prefix == NULL || prefix == BUSY) { 8874 // Nothing to take or waited long enough 8875 if (prefix == NULL) { 8876 // Write back the NULL in case we overwrote it with BUSY above 8877 // and it is still the same value. 8878 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY); 8879 } 8880 return false; 8881 } 8882 assert(prefix != NULL && prefix != BUSY, "Error"); 8883 size_t i = num; 8884 oop cur = prefix; 8885 // Walk down the first "num" objects, unless we reach the end. 8886 for (; i > 1 && cur->mark() != NULL; cur = oop(cur->mark()), i--); 8887 if (cur->mark() == NULL) { 8888 // We have "num" or fewer elements in the list, so there 8889 // is nothing to return to the global list. 8890 // Write back the NULL in lieu of the BUSY we wrote 8891 // above, if it is still the same value. 8892 if (_overflow_list == BUSY) { 8893 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY); 8894 } 8895 } else { 8896 // Chop off the suffix and rerturn it to the global list. 8897 assert(cur->mark() != BUSY, "Error"); 8898 oop suffix_head = cur->mark(); // suffix will be put back on global list 8899 cur->set_mark(NULL); // break off suffix 8900 // It's possible that the list is still in the empty(busy) state 8901 // we left it in a short while ago; in that case we may be 8902 // able to place back the suffix without incurring the cost 8903 // of a walk down the list. 8904 oop observed_overflow_list = _overflow_list; 8905 oop cur_overflow_list = observed_overflow_list; 8906 bool attached = false; 8907 while (observed_overflow_list == BUSY || observed_overflow_list == NULL) { 8908 observed_overflow_list = 8909 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list); 8910 if (cur_overflow_list == observed_overflow_list) { 8911 attached = true; 8912 break; 8913 } else cur_overflow_list = observed_overflow_list; 8914 } 8915 if (!attached) { 8916 // Too bad, someone else sneaked in (at least) an element; we'll need 8917 // to do a splice. Find tail of suffix so we can prepend suffix to global 8918 // list. 8919 for (cur = suffix_head; cur->mark() != NULL; cur = (oop)(cur->mark())); 8920 oop suffix_tail = cur; 8921 assert(suffix_tail != NULL && suffix_tail->mark() == NULL, 8922 "Tautology"); 8923 observed_overflow_list = _overflow_list; 8924 do { 8925 cur_overflow_list = observed_overflow_list; 8926 if (cur_overflow_list != BUSY) { 8927 // Do the splice ... 8928 suffix_tail->set_mark(markOop(cur_overflow_list)); 8929 } else { // cur_overflow_list == BUSY 8930 suffix_tail->set_mark(NULL); 8931 } 8932 // ... and try to place spliced list back on overflow_list ... 8933 observed_overflow_list = 8934 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list); 8935 } while (cur_overflow_list != observed_overflow_list); 8936 // ... until we have succeeded in doing so. 8937 } 8938 } 8939 8940 // Push the prefix elements on work_q 8941 assert(prefix != NULL, "control point invariant"); 8942 const markOop proto = markOopDesc::prototype(); 8943 oop next; 8944 NOT_PRODUCT(ssize_t n = 0;) 8945 for (cur = prefix; cur != NULL; cur = next) { 8946 next = oop(cur->mark()); 8947 cur->set_mark(proto); // until proven otherwise 8948 assert(cur->is_oop(), "Should be an oop"); 8949 bool res = work_q->push(cur); 8950 assert(res, "Bit off more than we can chew?"); 8951 NOT_PRODUCT(n++;) 8952 } 8953 #ifndef PRODUCT 8954 assert(_num_par_pushes >= n, "Too many pops?"); 8955 Atomic::add_ptr(-(intptr_t)n, &_num_par_pushes); 8956 #endif 8957 return true; 8958 } 8959 8960 // Single-threaded 8961 void CMSCollector::push_on_overflow_list(oop p) { 8962 NOT_PRODUCT(_num_par_pushes++;) 8963 assert(p->is_oop(), "Not an oop"); 8964 preserve_mark_if_necessary(p); 8965 p->set_mark((markOop)_overflow_list); 8966 _overflow_list = p; 8967 } 8968 8969 // Multi-threaded; use CAS to prepend to overflow list 8970 void CMSCollector::par_push_on_overflow_list(oop p) { 8971 NOT_PRODUCT(Atomic::inc_ptr(&_num_par_pushes);) 8972 assert(p->is_oop(), "Not an oop"); 8973 par_preserve_mark_if_necessary(p); 8974 oop observed_overflow_list = _overflow_list; 8975 oop cur_overflow_list; 8976 do { 8977 cur_overflow_list = observed_overflow_list; 8978 if (cur_overflow_list != BUSY) { 8979 p->set_mark(markOop(cur_overflow_list)); 8980 } else { 8981 p->set_mark(NULL); 8982 } 8983 observed_overflow_list = 8984 (oop) Atomic::cmpxchg_ptr(p, &_overflow_list, cur_overflow_list); 8985 } while (cur_overflow_list != observed_overflow_list); 8986 } 8987 #undef BUSY 8988 8989 // Single threaded 8990 // General Note on GrowableArray: pushes may silently fail 8991 // because we are (temporarily) out of C-heap for expanding 8992 // the stack. The problem is quite ubiquitous and affects 8993 // a lot of code in the JVM. The prudent thing for GrowableArray 8994 // to do (for now) is to exit with an error. However, that may 8995 // be too draconian in some cases because the caller may be 8996 // able to recover without much harm. For such cases, we 8997 // should probably introduce a "soft_push" method which returns 8998 // an indication of success or failure with the assumption that 8999 // the caller may be able to recover from a failure; code in 9000 // the VM can then be changed, incrementally, to deal with such 9001 // failures where possible, thus, incrementally hardening the VM 9002 // in such low resource situations. 9003 void CMSCollector::preserve_mark_work(oop p, markOop m) { 9004 _preserved_oop_stack.push(p); 9005 _preserved_mark_stack.push(m); 9006 assert(m == p->mark(), "Mark word changed"); 9007 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), 9008 "bijection"); 9009 } 9010 9011 // Single threaded 9012 void CMSCollector::preserve_mark_if_necessary(oop p) { 9013 markOop m = p->mark(); 9014 if (m->must_be_preserved(p)) { 9015 preserve_mark_work(p, m); 9016 } 9017 } 9018 9019 void CMSCollector::par_preserve_mark_if_necessary(oop p) { 9020 markOop m = p->mark(); 9021 if (m->must_be_preserved(p)) { 9022 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag); 9023 // Even though we read the mark word without holding 9024 // the lock, we are assured that it will not change 9025 // because we "own" this oop, so no other thread can 9026 // be trying to push it on the overflow list; see 9027 // the assertion in preserve_mark_work() that checks 9028 // that m == p->mark(). 9029 preserve_mark_work(p, m); 9030 } 9031 } 9032 9033 // We should be able to do this multi-threaded, 9034 // a chunk of stack being a task (this is 9035 // correct because each oop only ever appears 9036 // once in the overflow list. However, it's 9037 // not very easy to completely overlap this with 9038 // other operations, so will generally not be done 9039 // until all work's been completed. Because we 9040 // expect the preserved oop stack (set) to be small, 9041 // it's probably fine to do this single-threaded. 9042 // We can explore cleverer concurrent/overlapped/parallel 9043 // processing of preserved marks if we feel the 9044 // need for this in the future. Stack overflow should 9045 // be so rare in practice and, when it happens, its 9046 // effect on performance so great that this will 9047 // likely just be in the noise anyway. 9048 void CMSCollector::restore_preserved_marks_if_any() { 9049 assert(SafepointSynchronize::is_at_safepoint(), 9050 "world should be stopped"); 9051 assert(Thread::current()->is_ConcurrentGC_thread() || 9052 Thread::current()->is_VM_thread(), 9053 "should be single-threaded"); 9054 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), 9055 "bijection"); 9056 9057 while (!_preserved_oop_stack.is_empty()) { 9058 oop p = _preserved_oop_stack.pop(); 9059 assert(p->is_oop(), "Should be an oop"); 9060 assert(_span.contains(p), "oop should be in _span"); 9061 assert(p->mark() == markOopDesc::prototype(), 9062 "Set when taken from overflow list"); 9063 markOop m = _preserved_mark_stack.pop(); 9064 p->set_mark(m); 9065 } 9066 assert(_preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(), 9067 "stacks were cleared above"); 9068 } 9069 9070 #ifndef PRODUCT 9071 bool CMSCollector::no_preserved_marks() const { 9072 return _preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(); 9073 } 9074 #endif 9075 9076 CMSAdaptiveSizePolicy* ASConcurrentMarkSweepGeneration::cms_size_policy() const 9077 { 9078 GenCollectedHeap* gch = (GenCollectedHeap*) GenCollectedHeap::heap(); 9079 CMSAdaptiveSizePolicy* size_policy = 9080 (CMSAdaptiveSizePolicy*) gch->gen_policy()->size_policy(); 9081 assert(size_policy->is_gc_cms_adaptive_size_policy(), 9082 "Wrong type for size policy"); 9083 return size_policy; 9084 } 9085 9086 void ASConcurrentMarkSweepGeneration::resize(size_t cur_promo_size, 9087 size_t desired_promo_size) { 9088 if (cur_promo_size < desired_promo_size) { 9089 size_t expand_bytes = desired_promo_size - cur_promo_size; 9090 if (PrintAdaptiveSizePolicy && Verbose) { 9091 gclog_or_tty->print_cr(" ASConcurrentMarkSweepGeneration::resize " 9092 "Expanding tenured generation by " SIZE_FORMAT " (bytes)", 9093 expand_bytes); 9094 } 9095 expand(expand_bytes, 9096 MinHeapDeltaBytes, 9097 CMSExpansionCause::_adaptive_size_policy); 9098 } else if (desired_promo_size < cur_promo_size) { 9099 size_t shrink_bytes = cur_promo_size - desired_promo_size; 9100 if (PrintAdaptiveSizePolicy && Verbose) { 9101 gclog_or_tty->print_cr(" ASConcurrentMarkSweepGeneration::resize " 9102 "Shrinking tenured generation by " SIZE_FORMAT " (bytes)", 9103 shrink_bytes); 9104 } 9105 shrink(shrink_bytes); 9106 } 9107 } 9108 9109 CMSGCAdaptivePolicyCounters* ASConcurrentMarkSweepGeneration::gc_adaptive_policy_counters() { 9110 GenCollectedHeap* gch = GenCollectedHeap::heap(); 9111 CMSGCAdaptivePolicyCounters* counters = 9112 (CMSGCAdaptivePolicyCounters*) gch->collector_policy()->counters(); 9113 assert(counters->kind() == GCPolicyCounters::CMSGCAdaptivePolicyCountersKind, 9114 "Wrong kind of counters"); 9115 return counters; 9116 } 9117 9118 9119 void ASConcurrentMarkSweepGeneration::update_counters() { 9120 if (UsePerfData) { 9121 _space_counters->update_all(); 9122 _gen_counters->update_all(); 9123 CMSGCAdaptivePolicyCounters* counters = gc_adaptive_policy_counters(); 9124 GenCollectedHeap* gch = GenCollectedHeap::heap(); 9125 CMSGCStats* gc_stats_l = (CMSGCStats*) gc_stats(); 9126 assert(gc_stats_l->kind() == GCStats::CMSGCStatsKind, 9127 "Wrong gc statistics type"); 9128 counters->update_counters(gc_stats_l); 9129 } 9130 } 9131 9132 void ASConcurrentMarkSweepGeneration::update_counters(size_t used) { 9133 if (UsePerfData) { 9134 _space_counters->update_used(used); 9135 _space_counters->update_capacity(); 9136 _gen_counters->update_all(); 9137 9138 CMSGCAdaptivePolicyCounters* counters = gc_adaptive_policy_counters(); 9139 GenCollectedHeap* gch = GenCollectedHeap::heap(); 9140 CMSGCStats* gc_stats_l = (CMSGCStats*) gc_stats(); 9141 assert(gc_stats_l->kind() == GCStats::CMSGCStatsKind, 9142 "Wrong gc statistics type"); 9143 counters->update_counters(gc_stats_l); 9144 } 9145 } 9146 9147 void ASConcurrentMarkSweepGeneration::shrink_by(size_t desired_bytes) { 9148 assert_locked_or_safepoint(Heap_lock); 9149 assert_lock_strong(freelistLock()); 9150 HeapWord* old_end = _cmsSpace->end(); 9151 HeapWord* unallocated_start = _cmsSpace->unallocated_block(); 9152 assert(old_end >= unallocated_start, "Miscalculation of unallocated_start"); 9153 FreeChunk* chunk_at_end = find_chunk_at_end(); 9154 if (chunk_at_end == NULL) { 9155 // No room to shrink 9156 if (PrintGCDetails && Verbose) { 9157 gclog_or_tty->print_cr("No room to shrink: old_end " 9158 PTR_FORMAT " unallocated_start " PTR_FORMAT 9159 " chunk_at_end " PTR_FORMAT, 9160 old_end, unallocated_start, chunk_at_end); 9161 } 9162 return; 9163 } else { 9164 9165 // Find the chunk at the end of the space and determine 9166 // how much it can be shrunk. 9167 size_t shrinkable_size_in_bytes = chunk_at_end->size(); 9168 size_t aligned_shrinkable_size_in_bytes = 9169 align_size_down(shrinkable_size_in_bytes, os::vm_page_size()); 9170 assert(unallocated_start <= (HeapWord*) chunk_at_end->end(), 9171 "Inconsistent chunk at end of space"); 9172 size_t bytes = MIN2(desired_bytes, aligned_shrinkable_size_in_bytes); 9173 size_t word_size_before = heap_word_size(_virtual_space.committed_size()); 9174 9175 // Shrink the underlying space 9176 _virtual_space.shrink_by(bytes); 9177 if (PrintGCDetails && Verbose) { 9178 gclog_or_tty->print_cr("ConcurrentMarkSweepGeneration::shrink_by:" 9179 " desired_bytes " SIZE_FORMAT 9180 " shrinkable_size_in_bytes " SIZE_FORMAT 9181 " aligned_shrinkable_size_in_bytes " SIZE_FORMAT 9182 " bytes " SIZE_FORMAT, 9183 desired_bytes, shrinkable_size_in_bytes, 9184 aligned_shrinkable_size_in_bytes, bytes); 9185 gclog_or_tty->print_cr(" old_end " SIZE_FORMAT 9186 " unallocated_start " SIZE_FORMAT, 9187 old_end, unallocated_start); 9188 } 9189 9190 // If the space did shrink (shrinking is not guaranteed), 9191 // shrink the chunk at the end by the appropriate amount. 9192 if (((HeapWord*)_virtual_space.high()) < old_end) { 9193 size_t new_word_size = 9194 heap_word_size(_virtual_space.committed_size()); 9195 9196 // Have to remove the chunk from the dictionary because it is changing 9197 // size and might be someplace elsewhere in the dictionary. 9198 9199 // Get the chunk at end, shrink it, and put it 9200 // back. 9201 _cmsSpace->removeChunkFromDictionary(chunk_at_end); 9202 size_t word_size_change = word_size_before - new_word_size; 9203 size_t chunk_at_end_old_size = chunk_at_end->size(); 9204 assert(chunk_at_end_old_size >= word_size_change, 9205 "Shrink is too large"); 9206 chunk_at_end->set_size(chunk_at_end_old_size - 9207 word_size_change); 9208 _cmsSpace->freed((HeapWord*) chunk_at_end->end(), 9209 word_size_change); 9210 9211 _cmsSpace->returnChunkToDictionary(chunk_at_end); 9212 9213 MemRegion mr(_cmsSpace->bottom(), new_word_size); 9214 _bts->resize(new_word_size); // resize the block offset shared array 9215 Universe::heap()->barrier_set()->resize_covered_region(mr); 9216 _cmsSpace->assert_locked(); 9217 _cmsSpace->set_end((HeapWord*)_virtual_space.high()); 9218 9219 NOT_PRODUCT(_cmsSpace->dictionary()->verify()); 9220 9221 // update the space and generation capacity counters 9222 if (UsePerfData) { 9223 _space_counters->update_capacity(); 9224 _gen_counters->update_all(); 9225 } 9226 9227 if (Verbose && PrintGCDetails) { 9228 size_t new_mem_size = _virtual_space.committed_size(); 9229 size_t old_mem_size = new_mem_size + bytes; 9230 gclog_or_tty->print_cr("Shrinking %s from " SIZE_FORMAT "K by " SIZE_FORMAT "K to " SIZE_FORMAT "K", 9231 name(), old_mem_size/K, bytes/K, new_mem_size/K); 9232 } 9233 } 9234 9235 assert(_cmsSpace->unallocated_block() <= _cmsSpace->end(), 9236 "Inconsistency at end of space"); 9237 assert(chunk_at_end->end() == (uintptr_t*) _cmsSpace->end(), 9238 "Shrinking is inconsistent"); 9239 return; 9240 } 9241 } 9242 9243 // Transfer some number of overflown objects to usual marking 9244 // stack. Return true if some objects were transferred. 9245 bool MarkRefsIntoAndScanClosure::take_from_overflow_list() { 9246 size_t num = MIN2((size_t)(_mark_stack->capacity() - _mark_stack->length())/4, 9247 (size_t)ParGCDesiredObjsFromOverflowList); 9248 9249 bool res = _collector->take_from_overflow_list(num, _mark_stack); 9250 assert(_collector->overflow_list_is_empty() || res, 9251 "If list is not empty, we should have taken something"); 9252 assert(!res || !_mark_stack->isEmpty(), 9253 "If we took something, it should now be on our stack"); 9254 return res; 9255 } 9256 9257 size_t MarkDeadObjectsClosure::do_blk(HeapWord* addr) { 9258 size_t res = _sp->block_size_no_stall(addr, _collector); 9259 if (_sp->block_is_obj(addr)) { 9260 if (_live_bit_map->isMarked(addr)) { 9261 // It can't have been dead in a previous cycle 9262 guarantee(!_dead_bit_map->isMarked(addr), "No resurrection!"); 9263 } else { 9264 _dead_bit_map->mark(addr); // mark the dead object 9265 } 9266 } 9267 // Could be 0, if the block size could not be computed without stalling. 9268 return res; 9269 } 9270 9271 TraceCMSMemoryManagerStats::TraceCMSMemoryManagerStats(CMSCollector::CollectorState phase, GCCause::Cause cause): TraceMemoryManagerStats() { 9272 9273 switch (phase) { 9274 case CMSCollector::InitialMarking: 9275 initialize(true /* fullGC */ , 9276 cause /* cause of the GC */, 9277 true /* recordGCBeginTime */, 9278 true /* recordPreGCUsage */, 9279 false /* recordPeakUsage */, 9280 false /* recordPostGCusage */, 9281 true /* recordAccumulatedGCTime */, 9282 false /* recordGCEndTime */, 9283 false /* countCollection */ ); 9284 break; 9285 9286 case CMSCollector::FinalMarking: 9287 initialize(true /* fullGC */ , 9288 cause /* cause of the GC */, 9289 false /* recordGCBeginTime */, 9290 false /* recordPreGCUsage */, 9291 false /* recordPeakUsage */, 9292 false /* recordPostGCusage */, 9293 true /* recordAccumulatedGCTime */, 9294 false /* recordGCEndTime */, 9295 false /* countCollection */ ); 9296 break; 9297 9298 case CMSCollector::Sweeping: 9299 initialize(true /* fullGC */ , 9300 cause /* cause of the GC */, 9301 false /* recordGCBeginTime */, 9302 false /* recordPreGCUsage */, 9303 true /* recordPeakUsage */, 9304 true /* recordPostGCusage */, 9305 false /* recordAccumulatedGCTime */, 9306 true /* recordGCEndTime */, 9307 true /* countCollection */ ); 9308 break; 9309 9310 default: 9311 ShouldNotReachHere(); 9312 } 9313 } 9314