1 /* 2 * Copyright (c) 2001, 2014, 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/stringTable.hpp" 28 #include "classfile/systemDictionary.hpp" 29 #include "code/codeCache.hpp" 30 #include "gc_implementation/shared/adaptiveSizePolicy.hpp" 31 #include "gc_implementation/concurrentMarkSweep/cmsCollectorPolicy.hpp" 32 #include "gc_implementation/concurrentMarkSweep/cmsOopClosures.inline.hpp" 33 #include "gc_implementation/concurrentMarkSweep/compactibleFreeListSpace.hpp" 34 #include "gc_implementation/concurrentMarkSweep/concurrentMarkSweepGeneration.inline.hpp" 35 #include "gc_implementation/concurrentMarkSweep/concurrentMarkSweepThread.hpp" 36 #include "gc_implementation/concurrentMarkSweep/vmCMSOperations.hpp" 37 #include "gc_implementation/parNew/parNewGeneration.hpp" 38 #include "gc_implementation/shared/collectorCounters.hpp" 39 #include "gc_implementation/shared/gcTimer.hpp" 40 #include "gc_implementation/shared/gcTrace.hpp" 41 #include "gc_implementation/shared/gcTraceTime.hpp" 42 #include "gc_implementation/shared/isGCActiveMark.hpp" 43 #include "gc_interface/collectedHeap.inline.hpp" 44 #include "memory/allocation.hpp" 45 #include "memory/cardTableRS.hpp" 46 #include "memory/collectorPolicy.hpp" 47 #include "memory/gcLocker.inline.hpp" 48 #include "memory/genCollectedHeap.hpp" 49 #include "memory/genMarkSweep.hpp" 50 #include "memory/genOopClosures.inline.hpp" 51 #include "memory/iterator.inline.hpp" 52 #include "memory/padded.hpp" 53 #include "memory/referencePolicy.hpp" 54 #include "memory/resourceArea.hpp" 55 #include "memory/tenuredGeneration.hpp" 56 #include "oops/oop.inline.hpp" 57 #include "prims/jvmtiExport.hpp" 58 #include "runtime/atomic.inline.hpp" 59 #include "runtime/globals_extension.hpp" 60 #include "runtime/handles.inline.hpp" 61 #include "runtime/java.hpp" 62 #include "runtime/orderAccess.inline.hpp" 63 #include "runtime/vmThread.hpp" 64 #include "services/memoryService.hpp" 65 #include "services/runtimeService.hpp" 66 67 PRAGMA_FORMAT_MUTE_WARNINGS_FOR_GCC 68 69 // statics 70 CMSCollector* ConcurrentMarkSweepGeneration::_collector = NULL; 71 bool CMSCollector::_full_gc_requested = false; 72 GCCause::Cause CMSCollector::_full_gc_cause = GCCause::_no_gc; 73 74 ////////////////////////////////////////////////////////////////// 75 // In support of CMS/VM thread synchronization 76 ////////////////////////////////////////////////////////////////// 77 // We split use of the CGC_lock into 2 "levels". 78 // The low-level locking is of the usual CGC_lock monitor. We introduce 79 // a higher level "token" (hereafter "CMS token") built on top of the 80 // low level monitor (hereafter "CGC lock"). 81 // The token-passing protocol gives priority to the VM thread. The 82 // CMS-lock doesn't provide any fairness guarantees, but clients 83 // should ensure that it is only held for very short, bounded 84 // durations. 85 // 86 // When either of the CMS thread or the VM thread is involved in 87 // collection operations during which it does not want the other 88 // thread to interfere, it obtains the CMS token. 89 // 90 // If either thread tries to get the token while the other has 91 // it, that thread waits. However, if the VM thread and CMS thread 92 // both want the token, then the VM thread gets priority while the 93 // CMS thread waits. This ensures, for instance, that the "concurrent" 94 // phases of the CMS thread's work do not block out the VM thread 95 // for long periods of time as the CMS thread continues to hog 96 // the token. (See bug 4616232). 97 // 98 // The baton-passing functions are, however, controlled by the 99 // flags _foregroundGCShouldWait and _foregroundGCIsActive, 100 // and here the low-level CMS lock, not the high level token, 101 // ensures mutual exclusion. 102 // 103 // Two important conditions that we have to satisfy: 104 // 1. if a thread does a low-level wait on the CMS lock, then it 105 // relinquishes the CMS token if it were holding that token 106 // when it acquired the low-level CMS lock. 107 // 2. any low-level notifications on the low-level lock 108 // should only be sent when a thread has relinquished the token. 109 // 110 // In the absence of either property, we'd have potential deadlock. 111 // 112 // We protect each of the CMS (concurrent and sequential) phases 113 // with the CMS _token_, not the CMS _lock_. 114 // 115 // The only code protected by CMS lock is the token acquisition code 116 // itself, see ConcurrentMarkSweepThread::[de]synchronize(), and the 117 // baton-passing code. 118 // 119 // Unfortunately, i couldn't come up with a good abstraction to factor and 120 // hide the naked CGC_lock manipulation in the baton-passing code 121 // further below. That's something we should try to do. Also, the proof 122 // of correctness of this 2-level locking scheme is far from obvious, 123 // and potentially quite slippery. We have an uneasy suspicion, for instance, 124 // that there may be a theoretical possibility of delay/starvation in the 125 // low-level lock/wait/notify scheme used for the baton-passing because of 126 // potential interference with the priority scheme embodied in the 127 // CMS-token-passing protocol. See related comments at a CGC_lock->wait() 128 // invocation further below and marked with "XXX 20011219YSR". 129 // Indeed, as we note elsewhere, this may become yet more slippery 130 // in the presence of multiple CMS and/or multiple VM threads. XXX 131 132 class CMSTokenSync: public StackObj { 133 private: 134 bool _is_cms_thread; 135 public: 136 CMSTokenSync(bool is_cms_thread): 137 _is_cms_thread(is_cms_thread) { 138 assert(is_cms_thread == Thread::current()->is_ConcurrentGC_thread(), 139 "Incorrect argument to constructor"); 140 ConcurrentMarkSweepThread::synchronize(_is_cms_thread); 141 } 142 143 ~CMSTokenSync() { 144 assert(_is_cms_thread ? 145 ConcurrentMarkSweepThread::cms_thread_has_cms_token() : 146 ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 147 "Incorrect state"); 148 ConcurrentMarkSweepThread::desynchronize(_is_cms_thread); 149 } 150 }; 151 152 // Convenience class that does a CMSTokenSync, and then acquires 153 // upto three locks. 154 class CMSTokenSyncWithLocks: public CMSTokenSync { 155 private: 156 // Note: locks are acquired in textual declaration order 157 // and released in the opposite order 158 MutexLockerEx _locker1, _locker2, _locker3; 159 public: 160 CMSTokenSyncWithLocks(bool is_cms_thread, Mutex* mutex1, 161 Mutex* mutex2 = NULL, Mutex* mutex3 = NULL): 162 CMSTokenSync(is_cms_thread), 163 _locker1(mutex1, Mutex::_no_safepoint_check_flag), 164 _locker2(mutex2, Mutex::_no_safepoint_check_flag), 165 _locker3(mutex3, Mutex::_no_safepoint_check_flag) 166 { } 167 }; 168 169 170 // Wrapper class to temporarily disable icms during a foreground cms collection. 171 class ICMSDisabler: public StackObj { 172 public: 173 // The ctor disables icms and wakes up the thread so it notices the change; 174 // the dtor re-enables icms. Note that the CMSCollector methods will check 175 // CMSIncrementalMode. 176 ICMSDisabler() { CMSCollector::disable_icms(); CMSCollector::start_icms(); } 177 ~ICMSDisabler() { CMSCollector::enable_icms(); } 178 }; 179 180 ////////////////////////////////////////////////////////////////// 181 // Concurrent Mark-Sweep Generation ///////////////////////////// 182 ////////////////////////////////////////////////////////////////// 183 184 NOT_PRODUCT(CompactibleFreeListSpace* debug_cms_space;) 185 186 // This struct contains per-thread things necessary to support parallel 187 // young-gen collection. 188 class CMSParGCThreadState: public CHeapObj<mtGC> { 189 public: 190 CFLS_LAB lab; 191 PromotionInfo promo; 192 193 // Constructor. 194 CMSParGCThreadState(CompactibleFreeListSpace* cfls) : lab(cfls) { 195 promo.setSpace(cfls); 196 } 197 }; 198 199 ConcurrentMarkSweepGeneration::ConcurrentMarkSweepGeneration( 200 ReservedSpace rs, size_t initial_byte_size, int level, 201 CardTableRS* ct, bool use_adaptive_freelists, 202 FreeBlockDictionary<FreeChunk>::DictionaryChoice dictionaryChoice) : 203 CardGeneration(rs, initial_byte_size, level, ct), 204 _dilatation_factor(((double)MinChunkSize)/((double)(CollectedHeap::min_fill_size()))), 205 _debug_collection_type(Concurrent_collection_type), 206 _did_compact(false) 207 { 208 HeapWord* bottom = (HeapWord*) _virtual_space.low(); 209 HeapWord* end = (HeapWord*) _virtual_space.high(); 210 211 _direct_allocated_words = 0; 212 NOT_PRODUCT( 213 _numObjectsPromoted = 0; 214 _numWordsPromoted = 0; 215 _numObjectsAllocated = 0; 216 _numWordsAllocated = 0; 217 ) 218 219 _cmsSpace = new CompactibleFreeListSpace(_bts, MemRegion(bottom, end), 220 use_adaptive_freelists, 221 dictionaryChoice); 222 NOT_PRODUCT(debug_cms_space = _cmsSpace;) 223 if (_cmsSpace == NULL) { 224 vm_exit_during_initialization( 225 "CompactibleFreeListSpace allocation failure"); 226 } 227 _cmsSpace->_gen = this; 228 229 _gc_stats = new CMSGCStats(); 230 231 // Verify the assumption that FreeChunk::_prev and OopDesc::_klass 232 // offsets match. The ability to tell free chunks from objects 233 // depends on this property. 234 debug_only( 235 FreeChunk* junk = NULL; 236 assert(UseCompressedClassPointers || 237 junk->prev_addr() == (void*)(oop(junk)->klass_addr()), 238 "Offset of FreeChunk::_prev within FreeChunk must match" 239 " that of OopDesc::_klass within OopDesc"); 240 ) 241 if (CollectedHeap::use_parallel_gc_threads()) { 242 typedef CMSParGCThreadState* CMSParGCThreadStatePtr; 243 _par_gc_thread_states = 244 NEW_C_HEAP_ARRAY(CMSParGCThreadStatePtr, ParallelGCThreads, mtGC); 245 if (_par_gc_thread_states == NULL) { 246 vm_exit_during_initialization("Could not allocate par gc structs"); 247 } 248 for (uint i = 0; i < ParallelGCThreads; i++) { 249 _par_gc_thread_states[i] = new CMSParGCThreadState(cmsSpace()); 250 if (_par_gc_thread_states[i] == NULL) { 251 vm_exit_during_initialization("Could not allocate par gc structs"); 252 } 253 } 254 } else { 255 _par_gc_thread_states = NULL; 256 } 257 _incremental_collection_failed = false; 258 // The "dilatation_factor" is the expansion that can occur on 259 // account of the fact that the minimum object size in the CMS 260 // generation may be larger than that in, say, a contiguous young 261 // generation. 262 // Ideally, in the calculation below, we'd compute the dilatation 263 // factor as: MinChunkSize/(promoting_gen's min object size) 264 // Since we do not have such a general query interface for the 265 // promoting generation, we'll instead just use the minimum 266 // object size (which today is a header's worth of space); 267 // note that all arithmetic is in units of HeapWords. 268 assert(MinChunkSize >= CollectedHeap::min_fill_size(), "just checking"); 269 assert(_dilatation_factor >= 1.0, "from previous assert"); 270 } 271 272 273 // The field "_initiating_occupancy" represents the occupancy percentage 274 // at which we trigger a new collection cycle. Unless explicitly specified 275 // via CMSInitiatingOccupancyFraction (argument "io" below), it 276 // is calculated by: 277 // 278 // Let "f" be MinHeapFreeRatio in 279 // 280 // _initiating_occupancy = 100-f + 281 // f * (CMSTriggerRatio/100) 282 // where CMSTriggerRatio is the argument "tr" below. 283 // 284 // That is, if we assume the heap is at its desired maximum occupancy at the 285 // end of a collection, we let CMSTriggerRatio of the (purported) free 286 // space be allocated before initiating a new collection cycle. 287 // 288 void ConcurrentMarkSweepGeneration::init_initiating_occupancy(intx io, uintx tr) { 289 assert(io <= 100 && tr <= 100, "Check the arguments"); 290 if (io >= 0) { 291 _initiating_occupancy = (double)io / 100.0; 292 } else { 293 _initiating_occupancy = ((100 - MinHeapFreeRatio) + 294 (double)(tr * MinHeapFreeRatio) / 100.0) 295 / 100.0; 296 } 297 } 298 299 void ConcurrentMarkSweepGeneration::ref_processor_init() { 300 assert(collector() != NULL, "no collector"); 301 collector()->ref_processor_init(); 302 } 303 304 void CMSCollector::ref_processor_init() { 305 if (_ref_processor == NULL) { 306 // Allocate and initialize a reference processor 307 _ref_processor = 308 new ReferenceProcessor(_span, // span 309 (ParallelGCThreads > 1) && ParallelRefProcEnabled, // mt processing 310 (int) ParallelGCThreads, // mt processing degree 311 _cmsGen->refs_discovery_is_mt(), // mt discovery 312 (int) MAX2(ConcGCThreads, ParallelGCThreads), // mt discovery degree 313 _cmsGen->refs_discovery_is_atomic(), // discovery is not atomic 314 &_is_alive_closure); // closure for liveness info 315 // Initialize the _ref_processor field of CMSGen 316 _cmsGen->set_ref_processor(_ref_processor); 317 318 } 319 } 320 321 AdaptiveSizePolicy* CMSCollector::size_policy() { 322 GenCollectedHeap* gch = GenCollectedHeap::heap(); 323 assert(gch->kind() == CollectedHeap::GenCollectedHeap, 324 "Wrong type of heap"); 325 return gch->gen_policy()->size_policy(); 326 } 327 328 void ConcurrentMarkSweepGeneration::initialize_performance_counters() { 329 330 const char* gen_name = "old"; 331 GenCollectorPolicy* gcp = (GenCollectorPolicy*) GenCollectedHeap::heap()->collector_policy(); 332 333 // Generation Counters - generation 1, 1 subspace 334 _gen_counters = new GenerationCounters(gen_name, 1, 1, 335 gcp->min_old_size(), gcp->max_old_size(), &_virtual_space); 336 337 _space_counters = new GSpaceCounters(gen_name, 0, 338 _virtual_space.reserved_size(), 339 this, _gen_counters); 340 } 341 342 CMSStats::CMSStats(ConcurrentMarkSweepGeneration* cms_gen, unsigned int alpha): 343 _cms_gen(cms_gen) 344 { 345 assert(alpha <= 100, "bad value"); 346 _saved_alpha = alpha; 347 348 // Initialize the alphas to the bootstrap value of 100. 349 _gc0_alpha = _cms_alpha = 100; 350 351 _cms_begin_time.update(); 352 _cms_end_time.update(); 353 354 _gc0_duration = 0.0; 355 _gc0_period = 0.0; 356 _gc0_promoted = 0; 357 358 _cms_duration = 0.0; 359 _cms_period = 0.0; 360 _cms_allocated = 0; 361 362 _cms_used_at_gc0_begin = 0; 363 _cms_used_at_gc0_end = 0; 364 _allow_duty_cycle_reduction = false; 365 _valid_bits = 0; 366 _icms_duty_cycle = CMSIncrementalDutyCycle; 367 } 368 369 double CMSStats::cms_free_adjustment_factor(size_t free) const { 370 // TBD: CR 6909490 371 return 1.0; 372 } 373 374 void CMSStats::adjust_cms_free_adjustment_factor(bool fail, size_t free) { 375 } 376 377 // If promotion failure handling is on use 378 // the padded average size of the promotion for each 379 // young generation collection. 380 double CMSStats::time_until_cms_gen_full() const { 381 size_t cms_free = _cms_gen->cmsSpace()->free(); 382 GenCollectedHeap* gch = GenCollectedHeap::heap(); 383 size_t expected_promotion = MIN2(gch->young_gen()->capacity(), 384 (size_t) _cms_gen->gc_stats()->avg_promoted()->padded_average()); 385 if (cms_free > expected_promotion) { 386 // Start a cms collection if there isn't enough space to promote 387 // for the next minor collection. Use the padded average as 388 // a safety factor. 389 cms_free -= expected_promotion; 390 391 // Adjust by the safety factor. 392 double cms_free_dbl = (double)cms_free; 393 double cms_adjustment = (100.0 - CMSIncrementalSafetyFactor)/100.0; 394 // Apply a further correction factor which tries to adjust 395 // for recent occurance of concurrent mode failures. 396 cms_adjustment = cms_adjustment * cms_free_adjustment_factor(cms_free); 397 cms_free_dbl = cms_free_dbl * cms_adjustment; 398 399 if (PrintGCDetails && Verbose) { 400 gclog_or_tty->print_cr("CMSStats::time_until_cms_gen_full: cms_free " 401 SIZE_FORMAT " expected_promotion " SIZE_FORMAT, 402 cms_free, expected_promotion); 403 gclog_or_tty->print_cr(" cms_free_dbl %f cms_consumption_rate %f", 404 cms_free_dbl, cms_consumption_rate() + 1.0); 405 } 406 // Add 1 in case the consumption rate goes to zero. 407 return cms_free_dbl / (cms_consumption_rate() + 1.0); 408 } 409 return 0.0; 410 } 411 412 // Compare the duration of the cms collection to the 413 // time remaining before the cms generation is empty. 414 // Note that the time from the start of the cms collection 415 // to the start of the cms sweep (less than the total 416 // duration of the cms collection) can be used. This 417 // has been tried and some applications experienced 418 // promotion failures early in execution. This was 419 // possibly because the averages were not accurate 420 // enough at the beginning. 421 double CMSStats::time_until_cms_start() const { 422 // We add "gc0_period" to the "work" calculation 423 // below because this query is done (mostly) at the 424 // end of a scavenge, so we need to conservatively 425 // account for that much possible delay 426 // in the query so as to avoid concurrent mode failures 427 // due to starting the collection just a wee bit too 428 // late. 429 double work = cms_duration() + gc0_period(); 430 double deadline = time_until_cms_gen_full(); 431 // If a concurrent mode failure occurred recently, we want to be 432 // more conservative and halve our expected time_until_cms_gen_full() 433 if (work > deadline) { 434 if (Verbose && PrintGCDetails) { 435 gclog_or_tty->print( 436 " CMSCollector: collect because of anticipated promotion " 437 "before full %3.7f + %3.7f > %3.7f ", cms_duration(), 438 gc0_period(), time_until_cms_gen_full()); 439 } 440 return 0.0; 441 } 442 return work - deadline; 443 } 444 445 // Return a duty cycle based on old_duty_cycle and new_duty_cycle, limiting the 446 // amount of change to prevent wild oscillation. 447 unsigned int CMSStats::icms_damped_duty_cycle(unsigned int old_duty_cycle, 448 unsigned int new_duty_cycle) { 449 assert(old_duty_cycle <= 100, "bad input value"); 450 assert(new_duty_cycle <= 100, "bad input value"); 451 452 // Note: use subtraction with caution since it may underflow (values are 453 // unsigned). Addition is safe since we're in the range 0-100. 454 unsigned int damped_duty_cycle = new_duty_cycle; 455 if (new_duty_cycle < old_duty_cycle) { 456 const unsigned int largest_delta = MAX2(old_duty_cycle / 4, 5U); 457 if (new_duty_cycle + largest_delta < old_duty_cycle) { 458 damped_duty_cycle = old_duty_cycle - largest_delta; 459 } 460 } else if (new_duty_cycle > old_duty_cycle) { 461 const unsigned int largest_delta = MAX2(old_duty_cycle / 4, 15U); 462 if (new_duty_cycle > old_duty_cycle + largest_delta) { 463 damped_duty_cycle = MIN2(old_duty_cycle + largest_delta, 100U); 464 } 465 } 466 assert(damped_duty_cycle <= 100, "invalid duty cycle computed"); 467 468 if (CMSTraceIncrementalPacing) { 469 gclog_or_tty->print(" [icms_damped_duty_cycle(%d,%d) = %d] ", 470 old_duty_cycle, new_duty_cycle, damped_duty_cycle); 471 } 472 return damped_duty_cycle; 473 } 474 475 unsigned int CMSStats::icms_update_duty_cycle_impl() { 476 assert(CMSIncrementalPacing && valid(), 477 "should be handled in icms_update_duty_cycle()"); 478 479 double cms_time_so_far = cms_timer().seconds(); 480 double scaled_duration = cms_duration_per_mb() * _cms_used_at_gc0_end / M; 481 double scaled_duration_remaining = fabsd(scaled_duration - cms_time_so_far); 482 483 // Avoid division by 0. 484 double time_until_full = MAX2(time_until_cms_gen_full(), 0.01); 485 double duty_cycle_dbl = 100.0 * scaled_duration_remaining / time_until_full; 486 487 unsigned int new_duty_cycle = MIN2((unsigned int)duty_cycle_dbl, 100U); 488 if (new_duty_cycle > _icms_duty_cycle) { 489 // Avoid very small duty cycles (1 or 2); 0 is allowed. 490 if (new_duty_cycle > 2) { 491 _icms_duty_cycle = icms_damped_duty_cycle(_icms_duty_cycle, 492 new_duty_cycle); 493 } 494 } else if (_allow_duty_cycle_reduction) { 495 // The duty cycle is reduced only once per cms cycle (see record_cms_end()). 496 new_duty_cycle = icms_damped_duty_cycle(_icms_duty_cycle, new_duty_cycle); 497 // Respect the minimum duty cycle. 498 unsigned int min_duty_cycle = (unsigned int)CMSIncrementalDutyCycleMin; 499 _icms_duty_cycle = MAX2(new_duty_cycle, min_duty_cycle); 500 } 501 502 if (PrintGCDetails || CMSTraceIncrementalPacing) { 503 gclog_or_tty->print(" icms_dc=%d ", _icms_duty_cycle); 504 } 505 506 _allow_duty_cycle_reduction = false; 507 return _icms_duty_cycle; 508 } 509 510 #ifndef PRODUCT 511 void CMSStats::print_on(outputStream *st) const { 512 st->print(" gc0_alpha=%d,cms_alpha=%d", _gc0_alpha, _cms_alpha); 513 st->print(",gc0_dur=%g,gc0_per=%g,gc0_promo=" SIZE_FORMAT, 514 gc0_duration(), gc0_period(), gc0_promoted()); 515 st->print(",cms_dur=%g,cms_dur_per_mb=%g,cms_per=%g,cms_alloc=" SIZE_FORMAT, 516 cms_duration(), cms_duration_per_mb(), 517 cms_period(), cms_allocated()); 518 st->print(",cms_since_beg=%g,cms_since_end=%g", 519 cms_time_since_begin(), cms_time_since_end()); 520 st->print(",cms_used_beg=" SIZE_FORMAT ",cms_used_end=" SIZE_FORMAT, 521 _cms_used_at_gc0_begin, _cms_used_at_gc0_end); 522 if (CMSIncrementalMode) { 523 st->print(",dc=%d", icms_duty_cycle()); 524 } 525 526 if (valid()) { 527 st->print(",promo_rate=%g,cms_alloc_rate=%g", 528 promotion_rate(), cms_allocation_rate()); 529 st->print(",cms_consumption_rate=%g,time_until_full=%g", 530 cms_consumption_rate(), time_until_cms_gen_full()); 531 } 532 st->print(" "); 533 } 534 #endif // #ifndef PRODUCT 535 536 CMSCollector::CollectorState CMSCollector::_collectorState = 537 CMSCollector::Idling; 538 bool CMSCollector::_foregroundGCIsActive = false; 539 bool CMSCollector::_foregroundGCShouldWait = false; 540 541 CMSCollector::CMSCollector(ConcurrentMarkSweepGeneration* cmsGen, 542 CardTableRS* ct, 543 ConcurrentMarkSweepPolicy* cp): 544 _cmsGen(cmsGen), 545 _ct(ct), 546 _ref_processor(NULL), // will be set later 547 _conc_workers(NULL), // may be set later 548 _abort_preclean(false), 549 _start_sampling(false), 550 _between_prologue_and_epilogue(false), 551 _markBitMap(0, Mutex::leaf + 1, "CMS_markBitMap_lock"), 552 _modUnionTable((CardTableModRefBS::card_shift - LogHeapWordSize), 553 -1 /* lock-free */, "No_lock" /* dummy */), 554 _modUnionClosure(&_modUnionTable), 555 _modUnionClosurePar(&_modUnionTable), 556 // Adjust my span to cover old (cms) gen 557 _span(cmsGen->reserved()), 558 // Construct the is_alive_closure with _span & markBitMap 559 _is_alive_closure(_span, &_markBitMap), 560 _restart_addr(NULL), 561 _overflow_list(NULL), 562 _stats(cmsGen), 563 _eden_chunk_lock(new Mutex(Mutex::leaf + 1, "CMS_eden_chunk_lock", true)), 564 _eden_chunk_array(NULL), // may be set in ctor body 565 _eden_chunk_capacity(0), // -- ditto -- 566 _eden_chunk_index(0), // -- ditto -- 567 _survivor_plab_array(NULL), // -- ditto -- 568 _survivor_chunk_array(NULL), // -- ditto -- 569 _survivor_chunk_capacity(0), // -- ditto -- 570 _survivor_chunk_index(0), // -- ditto -- 571 _ser_pmc_preclean_ovflw(0), 572 _ser_kac_preclean_ovflw(0), 573 _ser_pmc_remark_ovflw(0), 574 _par_pmc_remark_ovflw(0), 575 _ser_kac_ovflw(0), 576 _par_kac_ovflw(0), 577 #ifndef PRODUCT 578 _num_par_pushes(0), 579 #endif 580 _collection_count_start(0), 581 _verifying(false), 582 _icms_start_limit(NULL), 583 _icms_stop_limit(NULL), 584 _verification_mark_bm(0, Mutex::leaf + 1, "CMS_verification_mark_bm_lock"), 585 _completed_initialization(false), 586 _collector_policy(cp), 587 _should_unload_classes(CMSClassUnloadingEnabled), 588 _concurrent_cycles_since_last_unload(0), 589 _roots_scanning_options(SharedHeap::SO_None), 590 _inter_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding), 591 _intra_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding), 592 _gc_tracer_cm(new (ResourceObj::C_HEAP, mtGC) CMSTracer()), 593 _gc_timer_cm(new (ResourceObj::C_HEAP, mtGC) ConcurrentGCTimer()), 594 _cms_start_registered(false) 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->young_gen(); 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) || CMSParallelInitialMarkEnabled) { 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 NOT_PRODUCT(_overflow_counter = CMSMarkStackOverflowInterval;) 782 _gc_counters = new CollectorCounters("CMS", 1); 783 _completed_initialization = true; 784 _inter_sweep_timer.start(); // start of time 785 } 786 787 const char* ConcurrentMarkSweepGeneration::name() const { 788 return "concurrent mark-sweep generation"; 789 } 790 void ConcurrentMarkSweepGeneration::update_counters() { 791 if (UsePerfData) { 792 _space_counters->update_all(); 793 _gen_counters->update_all(); 794 } 795 } 796 797 // this is an optimized version of update_counters(). it takes the 798 // used value as a parameter rather than computing it. 799 // 800 void ConcurrentMarkSweepGeneration::update_counters(size_t used) { 801 if (UsePerfData) { 802 _space_counters->update_used(used); 803 _space_counters->update_capacity(); 804 _gen_counters->update_all(); 805 } 806 } 807 808 void ConcurrentMarkSweepGeneration::print() const { 809 Generation::print(); 810 cmsSpace()->print(); 811 } 812 813 #ifndef PRODUCT 814 void ConcurrentMarkSweepGeneration::print_statistics() { 815 cmsSpace()->printFLCensus(0); 816 } 817 #endif 818 819 void ConcurrentMarkSweepGeneration::printOccupancy(const char *s) { 820 GenCollectedHeap* gch = GenCollectedHeap::heap(); 821 if (PrintGCDetails) { 822 if (Verbose) { 823 gclog_or_tty->print("[%d %s-%s: "SIZE_FORMAT"("SIZE_FORMAT")]", 824 level(), short_name(), s, used(), capacity()); 825 } else { 826 gclog_or_tty->print("[%d %s-%s: "SIZE_FORMAT"K("SIZE_FORMAT"K)]", 827 level(), short_name(), s, used() / K, capacity() / K); 828 } 829 } 830 if (Verbose) { 831 gclog_or_tty->print(" "SIZE_FORMAT"("SIZE_FORMAT")", 832 gch->used(), gch->capacity()); 833 } else { 834 gclog_or_tty->print(" "SIZE_FORMAT"K("SIZE_FORMAT"K)", 835 gch->used() / K, gch->capacity() / K); 836 } 837 } 838 839 size_t 840 ConcurrentMarkSweepGeneration::contiguous_available() const { 841 // dld proposes an improvement in precision here. If the committed 842 // part of the space ends in a free block we should add that to 843 // uncommitted size in the calculation below. Will make this 844 // change later, staying with the approximation below for the 845 // time being. -- ysr. 846 return MAX2(_virtual_space.uncommitted_size(), unsafe_max_alloc_nogc()); 847 } 848 849 size_t 850 ConcurrentMarkSweepGeneration::unsafe_max_alloc_nogc() const { 851 return _cmsSpace->max_alloc_in_words() * HeapWordSize; 852 } 853 854 size_t ConcurrentMarkSweepGeneration::max_available() const { 855 return free() + _virtual_space.uncommitted_size(); 856 } 857 858 bool ConcurrentMarkSweepGeneration::promotion_attempt_is_safe(size_t max_promotion_in_bytes) const { 859 size_t available = max_available(); 860 size_t av_promo = (size_t)gc_stats()->avg_promoted()->padded_average(); 861 bool res = (available >= av_promo) || (available >= max_promotion_in_bytes); 862 if (Verbose && PrintGCDetails) { 863 gclog_or_tty->print_cr( 864 "CMS: promo attempt is%s safe: available("SIZE_FORMAT") %s av_promo("SIZE_FORMAT")," 865 "max_promo("SIZE_FORMAT")", 866 res? "":" not", available, res? ">=":"<", 867 av_promo, max_promotion_in_bytes); 868 } 869 return res; 870 } 871 872 // At a promotion failure dump information on block layout in heap 873 // (cms old generation). 874 void ConcurrentMarkSweepGeneration::promotion_failure_occurred() { 875 if (CMSDumpAtPromotionFailure) { 876 cmsSpace()->dump_at_safepoint_with_locks(collector(), gclog_or_tty); 877 } 878 } 879 880 CompactibleSpace* 881 ConcurrentMarkSweepGeneration::first_compaction_space() const { 882 return _cmsSpace; 883 } 884 885 void ConcurrentMarkSweepGeneration::reset_after_compaction() { 886 // Clear the promotion information. These pointers can be adjusted 887 // along with all the other pointers into the heap but 888 // compaction is expected to be a rare event with 889 // a heap using cms so don't do it without seeing the need. 890 if (CollectedHeap::use_parallel_gc_threads()) { 891 for (uint i = 0; i < ParallelGCThreads; i++) { 892 _par_gc_thread_states[i]->promo.reset(); 893 } 894 } 895 } 896 897 void ConcurrentMarkSweepGeneration::space_iterate(SpaceClosure* blk, bool usedOnly) { 898 blk->do_space(_cmsSpace); 899 } 900 901 void ConcurrentMarkSweepGeneration::compute_new_size() { 902 assert_locked_or_safepoint(Heap_lock); 903 904 // If incremental collection failed, we just want to expand 905 // to the limit. 906 if (incremental_collection_failed()) { 907 clear_incremental_collection_failed(); 908 grow_to_reserved(); 909 return; 910 } 911 912 // The heap has been compacted but not reset yet. 913 // Any metric such as free() or used() will be incorrect. 914 915 CardGeneration::compute_new_size(); 916 917 // Reset again after a possible resizing 918 if (did_compact()) { 919 cmsSpace()->reset_after_compaction(); 920 } 921 } 922 923 void ConcurrentMarkSweepGeneration::compute_new_size_free_list() { 924 assert_locked_or_safepoint(Heap_lock); 925 926 // If incremental collection failed, we just want to expand 927 // to the limit. 928 if (incremental_collection_failed()) { 929 clear_incremental_collection_failed(); 930 grow_to_reserved(); 931 return; 932 } 933 934 double free_percentage = ((double) free()) / capacity(); 935 double desired_free_percentage = (double) MinHeapFreeRatio / 100; 936 double maximum_free_percentage = (double) MaxHeapFreeRatio / 100; 937 938 // compute expansion delta needed for reaching desired free percentage 939 if (free_percentage < desired_free_percentage) { 940 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 941 assert(desired_capacity >= capacity(), "invalid expansion size"); 942 size_t expand_bytes = MAX2(desired_capacity - capacity(), MinHeapDeltaBytes); 943 if (PrintGCDetails && Verbose) { 944 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 945 gclog_or_tty->print_cr("\nFrom compute_new_size: "); 946 gclog_or_tty->print_cr(" Free fraction %f", free_percentage); 947 gclog_or_tty->print_cr(" Desired free fraction %f", 948 desired_free_percentage); 949 gclog_or_tty->print_cr(" Maximum free fraction %f", 950 maximum_free_percentage); 951 gclog_or_tty->print_cr(" Capacity "SIZE_FORMAT, capacity()/1000); 952 gclog_or_tty->print_cr(" Desired capacity "SIZE_FORMAT, 953 desired_capacity/1000); 954 int prev_level = level() - 1; 955 if (prev_level >= 0) { 956 size_t prev_size = 0; 957 GenCollectedHeap* gch = GenCollectedHeap::heap(); 958 Generation* prev_gen = gch->young_gen(); 959 prev_size = prev_gen->capacity(); 960 gclog_or_tty->print_cr(" Younger gen size "SIZE_FORMAT, 961 prev_size/1000); 962 } 963 gclog_or_tty->print_cr(" unsafe_max_alloc_nogc "SIZE_FORMAT, 964 unsafe_max_alloc_nogc()/1000); 965 gclog_or_tty->print_cr(" contiguous available "SIZE_FORMAT, 966 contiguous_available()/1000); 967 gclog_or_tty->print_cr(" Expand by "SIZE_FORMAT" (bytes)", 968 expand_bytes); 969 } 970 // safe if expansion fails 971 expand(expand_bytes, 0, CMSExpansionCause::_satisfy_free_ratio); 972 if (PrintGCDetails && Verbose) { 973 gclog_or_tty->print_cr(" Expanded free fraction %f", 974 ((double) free()) / capacity()); 975 } 976 } else { 977 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 978 assert(desired_capacity <= capacity(), "invalid expansion size"); 979 size_t shrink_bytes = capacity() - desired_capacity; 980 // Don't shrink unless the delta is greater than the minimum shrink we want 981 if (shrink_bytes >= MinHeapDeltaBytes) { 982 shrink_free_list_by(shrink_bytes); 983 } 984 } 985 } 986 987 Mutex* ConcurrentMarkSweepGeneration::freelistLock() const { 988 return cmsSpace()->freelistLock(); 989 } 990 991 HeapWord* ConcurrentMarkSweepGeneration::allocate(size_t size, 992 bool tlab) { 993 CMSSynchronousYieldRequest yr; 994 MutexLockerEx x(freelistLock(), 995 Mutex::_no_safepoint_check_flag); 996 return have_lock_and_allocate(size, tlab); 997 } 998 999 HeapWord* ConcurrentMarkSweepGeneration::have_lock_and_allocate(size_t size, 1000 bool tlab /* ignored */) { 1001 assert_lock_strong(freelistLock()); 1002 size_t adjustedSize = CompactibleFreeListSpace::adjustObjectSize(size); 1003 HeapWord* res = cmsSpace()->allocate(adjustedSize); 1004 // Allocate the object live (grey) if the background collector has 1005 // started marking. This is necessary because the marker may 1006 // have passed this address and consequently this object will 1007 // not otherwise be greyed and would be incorrectly swept up. 1008 // Note that if this object contains references, the writing 1009 // of those references will dirty the card containing this object 1010 // allowing the object to be blackened (and its references scanned) 1011 // either during a preclean phase or at the final checkpoint. 1012 if (res != NULL) { 1013 // We may block here with an uninitialized object with 1014 // its mark-bit or P-bits not yet set. Such objects need 1015 // to be safely navigable by block_start(). 1016 assert(oop(res)->klass_or_null() == NULL, "Object should be uninitialized here."); 1017 assert(!((FreeChunk*)res)->is_free(), "Error, block will look free but show wrong size"); 1018 collector()->direct_allocated(res, adjustedSize); 1019 _direct_allocated_words += adjustedSize; 1020 // allocation counters 1021 NOT_PRODUCT( 1022 _numObjectsAllocated++; 1023 _numWordsAllocated += (int)adjustedSize; 1024 ) 1025 } 1026 return res; 1027 } 1028 1029 // In the case of direct allocation by mutators in a generation that 1030 // is being concurrently collected, the object must be allocated 1031 // live (grey) if the background collector has started marking. 1032 // This is necessary because the marker may 1033 // have passed this address and consequently this object will 1034 // not otherwise be greyed and would be incorrectly swept up. 1035 // Note that if this object contains references, the writing 1036 // of those references will dirty the card containing this object 1037 // allowing the object to be blackened (and its references scanned) 1038 // either during a preclean phase or at the final checkpoint. 1039 void CMSCollector::direct_allocated(HeapWord* start, size_t size) { 1040 assert(_markBitMap.covers(start, size), "Out of bounds"); 1041 if (_collectorState >= Marking) { 1042 MutexLockerEx y(_markBitMap.lock(), 1043 Mutex::_no_safepoint_check_flag); 1044 // [see comments preceding SweepClosure::do_blk() below for details] 1045 // 1046 // Can the P-bits be deleted now? JJJ 1047 // 1048 // 1. need to mark the object as live so it isn't collected 1049 // 2. need to mark the 2nd bit to indicate the object may be uninitialized 1050 // 3. need to mark the end of the object so marking, precleaning or sweeping 1051 // can skip over uninitialized or unparsable objects. An allocated 1052 // object is considered uninitialized for our purposes as long as 1053 // its klass word is NULL. All old gen objects are parsable 1054 // as soon as they are initialized.) 1055 _markBitMap.mark(start); // object is live 1056 _markBitMap.mark(start + 1); // object is potentially uninitialized? 1057 _markBitMap.mark(start + size - 1); 1058 // mark end of object 1059 } 1060 // check that oop looks uninitialized 1061 assert(oop(start)->klass_or_null() == NULL, "_klass should be NULL"); 1062 } 1063 1064 void CMSCollector::promoted(bool par, HeapWord* start, 1065 bool is_obj_array, size_t obj_size) { 1066 assert(_markBitMap.covers(start), "Out of bounds"); 1067 // See comment in direct_allocated() about when objects should 1068 // be allocated live. 1069 if (_collectorState >= Marking) { 1070 // we already hold the marking bit map lock, taken in 1071 // the prologue 1072 if (par) { 1073 _markBitMap.par_mark(start); 1074 } else { 1075 _markBitMap.mark(start); 1076 } 1077 // We don't need to mark the object as uninitialized (as 1078 // in direct_allocated above) because this is being done with the 1079 // world stopped and the object will be initialized by the 1080 // time the marking, precleaning or sweeping get to look at it. 1081 // But see the code for copying objects into the CMS generation, 1082 // where we need to ensure that concurrent readers of the 1083 // block offset table are able to safely navigate a block that 1084 // is in flux from being free to being allocated (and in 1085 // transition while being copied into) and subsequently 1086 // becoming a bona-fide object when the copy/promotion is complete. 1087 assert(SafepointSynchronize::is_at_safepoint(), 1088 "expect promotion only at safepoints"); 1089 1090 if (_collectorState < Sweeping) { 1091 // Mark the appropriate cards in the modUnionTable, so that 1092 // this object gets scanned before the sweep. If this is 1093 // not done, CMS generation references in the object might 1094 // not get marked. 1095 // For the case of arrays, which are otherwise precisely 1096 // marked, we need to dirty the entire array, not just its head. 1097 if (is_obj_array) { 1098 // The [par_]mark_range() method expects mr.end() below to 1099 // be aligned to the granularity of a bit's representation 1100 // in the heap. In the case of the MUT below, that's a 1101 // card size. 1102 MemRegion mr(start, 1103 (HeapWord*)round_to((intptr_t)(start + obj_size), 1104 CardTableModRefBS::card_size /* bytes */)); 1105 if (par) { 1106 _modUnionTable.par_mark_range(mr); 1107 } else { 1108 _modUnionTable.mark_range(mr); 1109 } 1110 } else { // not an obj array; we can just mark the head 1111 if (par) { 1112 _modUnionTable.par_mark(start); 1113 } else { 1114 _modUnionTable.mark(start); 1115 } 1116 } 1117 } 1118 } 1119 } 1120 1121 static inline size_t percent_of_space(Space* space, HeapWord* addr) 1122 { 1123 size_t delta = pointer_delta(addr, space->bottom()); 1124 return (size_t)(delta * 100.0 / (space->capacity() / HeapWordSize)); 1125 } 1126 1127 void CMSCollector::icms_update_allocation_limits() 1128 { 1129 Generation* young = GenCollectedHeap::heap()->young_gen(); 1130 EdenSpace* eden = young->as_DefNewGeneration()->eden(); 1131 1132 const unsigned int duty_cycle = stats().icms_update_duty_cycle(); 1133 if (CMSTraceIncrementalPacing) { 1134 stats().print(); 1135 } 1136 1137 assert(duty_cycle <= 100, "invalid duty cycle"); 1138 if (duty_cycle != 0) { 1139 // The duty_cycle is a percentage between 0 and 100; convert to words and 1140 // then compute the offset from the endpoints of the space. 1141 size_t free_words = eden->free() / HeapWordSize; 1142 double free_words_dbl = (double)free_words; 1143 size_t duty_cycle_words = (size_t)(free_words_dbl * duty_cycle / 100.0); 1144 size_t offset_words = (free_words - duty_cycle_words) / 2; 1145 1146 _icms_start_limit = eden->top() + offset_words; 1147 _icms_stop_limit = eden->end() - offset_words; 1148 1149 // The limits may be adjusted (shifted to the right) by 1150 // CMSIncrementalOffset, to allow the application more mutator time after a 1151 // young gen gc (when all mutators were stopped) and before CMS starts and 1152 // takes away one or more cpus. 1153 if (CMSIncrementalOffset != 0) { 1154 double adjustment_dbl = free_words_dbl * CMSIncrementalOffset / 100.0; 1155 size_t adjustment = (size_t)adjustment_dbl; 1156 HeapWord* tmp_stop = _icms_stop_limit + adjustment; 1157 if (tmp_stop > _icms_stop_limit && tmp_stop < eden->end()) { 1158 _icms_start_limit += adjustment; 1159 _icms_stop_limit = tmp_stop; 1160 } 1161 } 1162 } 1163 if (duty_cycle == 0 || (_icms_start_limit == _icms_stop_limit)) { 1164 _icms_start_limit = _icms_stop_limit = eden->end(); 1165 } 1166 1167 // Install the new start limit. 1168 eden->set_soft_end(_icms_start_limit); 1169 1170 if (CMSTraceIncrementalMode) { 1171 gclog_or_tty->print(" icms alloc limits: " 1172 PTR_FORMAT "," PTR_FORMAT 1173 " (" SIZE_FORMAT "%%," SIZE_FORMAT "%%) ", 1174 p2i(_icms_start_limit), p2i(_icms_stop_limit), 1175 percent_of_space(eden, _icms_start_limit), 1176 percent_of_space(eden, _icms_stop_limit)); 1177 if (Verbose) { 1178 gclog_or_tty->print("eden: "); 1179 eden->print_on(gclog_or_tty); 1180 } 1181 } 1182 } 1183 1184 // Any changes here should try to maintain the invariant 1185 // that if this method is called with _icms_start_limit 1186 // and _icms_stop_limit both NULL, then it should return NULL 1187 // and not notify the icms thread. 1188 HeapWord* 1189 CMSCollector::allocation_limit_reached(Space* space, HeapWord* top, 1190 size_t word_size) 1191 { 1192 // A start_limit equal to end() means the duty cycle is 0, so treat that as a 1193 // nop. 1194 if (CMSIncrementalMode && _icms_start_limit != space->end()) { 1195 if (top <= _icms_start_limit) { 1196 if (CMSTraceIncrementalMode) { 1197 space->print_on(gclog_or_tty); 1198 gclog_or_tty->stamp(); 1199 gclog_or_tty->print_cr(" start limit top=" PTR_FORMAT 1200 ", new limit=" PTR_FORMAT 1201 " (" SIZE_FORMAT "%%)", 1202 p2i(top), p2i(_icms_stop_limit), 1203 percent_of_space(space, _icms_stop_limit)); 1204 } 1205 ConcurrentMarkSweepThread::start_icms(); 1206 assert(top < _icms_stop_limit, "Tautology"); 1207 if (word_size < pointer_delta(_icms_stop_limit, top)) { 1208 return _icms_stop_limit; 1209 } 1210 1211 // The allocation will cross both the _start and _stop limits, so do the 1212 // stop notification also and return end(). 1213 if (CMSTraceIncrementalMode) { 1214 space->print_on(gclog_or_tty); 1215 gclog_or_tty->stamp(); 1216 gclog_or_tty->print_cr(" +stop limit top=" PTR_FORMAT 1217 ", new limit=" PTR_FORMAT 1218 " (" SIZE_FORMAT "%%)", 1219 p2i(top), p2i(space->end()), 1220 percent_of_space(space, space->end())); 1221 } 1222 ConcurrentMarkSweepThread::stop_icms(); 1223 return space->end(); 1224 } 1225 1226 if (top <= _icms_stop_limit) { 1227 if (CMSTraceIncrementalMode) { 1228 space->print_on(gclog_or_tty); 1229 gclog_or_tty->stamp(); 1230 gclog_or_tty->print_cr(" stop limit top=" PTR_FORMAT 1231 ", new limit=" PTR_FORMAT 1232 " (" SIZE_FORMAT "%%)", 1233 top, space->end(), 1234 percent_of_space(space, space->end())); 1235 } 1236 ConcurrentMarkSweepThread::stop_icms(); 1237 return space->end(); 1238 } 1239 1240 if (CMSTraceIncrementalMode) { 1241 space->print_on(gclog_or_tty); 1242 gclog_or_tty->stamp(); 1243 gclog_or_tty->print_cr(" end limit top=" PTR_FORMAT 1244 ", new limit=" PTR_FORMAT, 1245 top, NULL); 1246 } 1247 } 1248 1249 return NULL; 1250 } 1251 1252 oop ConcurrentMarkSweepGeneration::promote(oop obj, size_t obj_size) { 1253 assert(obj_size == (size_t)obj->size(), "bad obj_size passed in"); 1254 // allocate, copy and if necessary update promoinfo -- 1255 // delegate to underlying space. 1256 assert_lock_strong(freelistLock()); 1257 1258 #ifndef PRODUCT 1259 if (Universe::heap()->promotion_should_fail()) { 1260 return NULL; 1261 } 1262 #endif // #ifndef PRODUCT 1263 1264 oop res = _cmsSpace->promote(obj, obj_size); 1265 if (res == NULL) { 1266 // expand and retry 1267 size_t s = _cmsSpace->expansionSpaceRequired(obj_size); // HeapWords 1268 expand(s*HeapWordSize, MinHeapDeltaBytes, 1269 CMSExpansionCause::_satisfy_promotion); 1270 // Since this is the old generation, we don't try to promote 1271 // into a more senior generation. 1272 res = _cmsSpace->promote(obj, obj_size); 1273 } 1274 if (res != NULL) { 1275 // See comment in allocate() about when objects should 1276 // be allocated live. 1277 assert(obj->is_oop(), "Will dereference klass pointer below"); 1278 collector()->promoted(false, // Not parallel 1279 (HeapWord*)res, obj->is_objArray(), obj_size); 1280 // promotion counters 1281 NOT_PRODUCT( 1282 _numObjectsPromoted++; 1283 _numWordsPromoted += 1284 (int)(CompactibleFreeListSpace::adjustObjectSize(obj->size())); 1285 ) 1286 } 1287 return res; 1288 } 1289 1290 1291 HeapWord* 1292 ConcurrentMarkSweepGeneration::allocation_limit_reached(Space* space, 1293 HeapWord* top, 1294 size_t word_sz) 1295 { 1296 return collector()->allocation_limit_reached(space, top, word_sz); 1297 } 1298 1299 // IMPORTANT: Notes on object size recognition in CMS. 1300 // --------------------------------------------------- 1301 // A block of storage in the CMS generation is always in 1302 // one of three states. A free block (FREE), an allocated 1303 // object (OBJECT) whose size() method reports the correct size, 1304 // and an intermediate state (TRANSIENT) in which its size cannot 1305 // be accurately determined. 1306 // STATE IDENTIFICATION: (32 bit and 64 bit w/o COOPS) 1307 // ----------------------------------------------------- 1308 // FREE: klass_word & 1 == 1; mark_word holds block size 1309 // 1310 // OBJECT: klass_word installed; klass_word != 0 && klass_word & 1 == 0; 1311 // obj->size() computes correct size 1312 // 1313 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT 1314 // 1315 // STATE IDENTIFICATION: (64 bit+COOPS) 1316 // ------------------------------------ 1317 // FREE: mark_word & CMS_FREE_BIT == 1; mark_word & ~CMS_FREE_BIT gives block_size 1318 // 1319 // OBJECT: klass_word installed; klass_word != 0; 1320 // obj->size() computes correct size 1321 // 1322 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT 1323 // 1324 // 1325 // STATE TRANSITION DIAGRAM 1326 // 1327 // mut / parnew mut / parnew 1328 // FREE --------------------> TRANSIENT ---------------------> OBJECT --| 1329 // ^ | 1330 // |------------------------ DEAD <------------------------------------| 1331 // sweep mut 1332 // 1333 // While a block is in TRANSIENT state its size cannot be determined 1334 // so readers will either need to come back later or stall until 1335 // the size can be determined. Note that for the case of direct 1336 // allocation, P-bits, when available, may be used to determine the 1337 // size of an object that may not yet have been initialized. 1338 1339 // Things to support parallel young-gen collection. 1340 oop 1341 ConcurrentMarkSweepGeneration::par_promote(int thread_num, 1342 oop old, markOop m, 1343 size_t word_sz) { 1344 #ifndef PRODUCT 1345 if (Universe::heap()->promotion_should_fail()) { 1346 return NULL; 1347 } 1348 #endif // #ifndef PRODUCT 1349 1350 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1351 PromotionInfo* promoInfo = &ps->promo; 1352 // if we are tracking promotions, then first ensure space for 1353 // promotion (including spooling space for saving header if necessary). 1354 // then allocate and copy, then track promoted info if needed. 1355 // When tracking (see PromotionInfo::track()), the mark word may 1356 // be displaced and in this case restoration of the mark word 1357 // occurs in the (oop_since_save_marks_)iterate phase. 1358 if (promoInfo->tracking() && !promoInfo->ensure_spooling_space()) { 1359 // Out of space for allocating spooling buffers; 1360 // try expanding and allocating spooling buffers. 1361 if (!expand_and_ensure_spooling_space(promoInfo)) { 1362 return NULL; 1363 } 1364 } 1365 assert(promoInfo->has_spooling_space(), "Control point invariant"); 1366 const size_t alloc_sz = CompactibleFreeListSpace::adjustObjectSize(word_sz); 1367 HeapWord* obj_ptr = ps->lab.alloc(alloc_sz); 1368 if (obj_ptr == NULL) { 1369 obj_ptr = expand_and_par_lab_allocate(ps, alloc_sz); 1370 if (obj_ptr == NULL) { 1371 return NULL; 1372 } 1373 } 1374 oop obj = oop(obj_ptr); 1375 OrderAccess::storestore(); 1376 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1377 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1378 // IMPORTANT: See note on object initialization for CMS above. 1379 // Otherwise, copy the object. Here we must be careful to insert the 1380 // klass pointer last, since this marks the block as an allocated object. 1381 // Except with compressed oops it's the mark word. 1382 HeapWord* old_ptr = (HeapWord*)old; 1383 // Restore the mark word copied above. 1384 obj->set_mark(m); 1385 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1386 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1387 OrderAccess::storestore(); 1388 1389 if (UseCompressedClassPointers) { 1390 // Copy gap missed by (aligned) header size calculation below 1391 obj->set_klass_gap(old->klass_gap()); 1392 } 1393 if (word_sz > (size_t)oopDesc::header_size()) { 1394 Copy::aligned_disjoint_words(old_ptr + oopDesc::header_size(), 1395 obj_ptr + oopDesc::header_size(), 1396 word_sz - oopDesc::header_size()); 1397 } 1398 1399 // Now we can track the promoted object, if necessary. We take care 1400 // to delay the transition from uninitialized to full object 1401 // (i.e., insertion of klass pointer) until after, so that it 1402 // atomically becomes a promoted object. 1403 if (promoInfo->tracking()) { 1404 promoInfo->track((PromotedObject*)obj, old->klass()); 1405 } 1406 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1407 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1408 assert(old->is_oop(), "Will use and dereference old klass ptr below"); 1409 1410 // Finally, install the klass pointer (this should be volatile). 1411 OrderAccess::storestore(); 1412 obj->set_klass(old->klass()); 1413 // We should now be able to calculate the right size for this object 1414 assert(obj->is_oop() && obj->size() == (int)word_sz, "Error, incorrect size computed for promoted object"); 1415 1416 collector()->promoted(true, // parallel 1417 obj_ptr, old->is_objArray(), word_sz); 1418 1419 NOT_PRODUCT( 1420 Atomic::inc_ptr(&_numObjectsPromoted); 1421 Atomic::add_ptr(alloc_sz, &_numWordsPromoted); 1422 ) 1423 1424 return obj; 1425 } 1426 1427 void 1428 ConcurrentMarkSweepGeneration:: 1429 par_promote_alloc_undo(int thread_num, 1430 HeapWord* obj, size_t word_sz) { 1431 // CMS does not support promotion undo. 1432 ShouldNotReachHere(); 1433 } 1434 1435 void 1436 ConcurrentMarkSweepGeneration:: 1437 par_promote_alloc_done(int thread_num) { 1438 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1439 ps->lab.retire(thread_num); 1440 } 1441 1442 void 1443 ConcurrentMarkSweepGeneration:: 1444 par_oop_since_save_marks_iterate_done(int thread_num) { 1445 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1446 ParScanWithoutBarrierClosure* dummy_cl = NULL; 1447 ps->promo.promoted_oops_iterate_nv(dummy_cl); 1448 } 1449 1450 bool ConcurrentMarkSweepGeneration::should_collect(bool full, 1451 size_t size, 1452 bool tlab) 1453 { 1454 // We allow a STW collection only if a full 1455 // collection was requested. 1456 return full || should_allocate(size, tlab); // FIX ME !!! 1457 // This and promotion failure handling are connected at the 1458 // hip and should be fixed by untying them. 1459 } 1460 1461 bool CMSCollector::shouldConcurrentCollect() { 1462 if (_full_gc_requested) { 1463 if (Verbose && PrintGCDetails) { 1464 gclog_or_tty->print_cr("CMSCollector: collect because of explicit " 1465 " gc request (or gc_locker)"); 1466 } 1467 return true; 1468 } 1469 1470 // For debugging purposes, change the type of collection. 1471 // If the rotation is not on the concurrent collection 1472 // type, don't start a concurrent collection. 1473 NOT_PRODUCT( 1474 if (RotateCMSCollectionTypes && 1475 (_cmsGen->debug_collection_type() != 1476 ConcurrentMarkSweepGeneration::Concurrent_collection_type)) { 1477 assert(_cmsGen->debug_collection_type() != 1478 ConcurrentMarkSweepGeneration::Unknown_collection_type, 1479 "Bad cms collection type"); 1480 return false; 1481 } 1482 ) 1483 1484 FreelistLocker x(this); 1485 // ------------------------------------------------------------------ 1486 // Print out lots of information which affects the initiation of 1487 // a collection. 1488 if (PrintCMSInitiationStatistics && stats().valid()) { 1489 gclog_or_tty->print("CMSCollector shouldConcurrentCollect: "); 1490 gclog_or_tty->stamp(); 1491 gclog_or_tty->cr(); 1492 stats().print_on(gclog_or_tty); 1493 gclog_or_tty->print_cr("time_until_cms_gen_full %3.7f", 1494 stats().time_until_cms_gen_full()); 1495 gclog_or_tty->print_cr("free="SIZE_FORMAT, _cmsGen->free()); 1496 gclog_or_tty->print_cr("contiguous_available="SIZE_FORMAT, 1497 _cmsGen->contiguous_available()); 1498 gclog_or_tty->print_cr("promotion_rate=%g", stats().promotion_rate()); 1499 gclog_or_tty->print_cr("cms_allocation_rate=%g", stats().cms_allocation_rate()); 1500 gclog_or_tty->print_cr("occupancy=%3.7f", _cmsGen->occupancy()); 1501 gclog_or_tty->print_cr("initiatingOccupancy=%3.7f", _cmsGen->initiating_occupancy()); 1502 gclog_or_tty->print_cr("cms_time_since_begin=%3.7f", stats().cms_time_since_begin()); 1503 gclog_or_tty->print_cr("cms_time_since_end=%3.7f", stats().cms_time_since_end()); 1504 gclog_or_tty->print_cr("metadata initialized %d", 1505 MetaspaceGC::should_concurrent_collect()); 1506 } 1507 // ------------------------------------------------------------------ 1508 1509 // If the estimated time to complete a cms collection (cms_duration()) 1510 // is less than the estimated time remaining until the cms generation 1511 // is full, start a collection. 1512 if (!UseCMSInitiatingOccupancyOnly) { 1513 if (stats().valid()) { 1514 if (stats().time_until_cms_start() == 0.0) { 1515 return true; 1516 } 1517 } else { 1518 // We want to conservatively collect somewhat early in order 1519 // to try and "bootstrap" our CMS/promotion statistics; 1520 // this branch will not fire after the first successful CMS 1521 // collection because the stats should then be valid. 1522 if (_cmsGen->occupancy() >= _bootstrap_occupancy) { 1523 if (Verbose && PrintGCDetails) { 1524 gclog_or_tty->print_cr( 1525 " CMSCollector: collect for bootstrapping statistics:" 1526 " occupancy = %f, boot occupancy = %f", _cmsGen->occupancy(), 1527 _bootstrap_occupancy); 1528 } 1529 return true; 1530 } 1531 } 1532 } 1533 1534 // Otherwise, we start a collection cycle if 1535 // old gen want a collection cycle started. Each may use 1536 // an appropriate criterion for making this decision. 1537 // XXX We need to make sure that the gen expansion 1538 // criterion dovetails well with this. XXX NEED TO FIX THIS 1539 if (_cmsGen->should_concurrent_collect()) { 1540 if (Verbose && PrintGCDetails) { 1541 gclog_or_tty->print_cr("CMS old gen initiated"); 1542 } 1543 return true; 1544 } 1545 1546 // We start a collection if we believe an incremental collection may fail; 1547 // this is not likely to be productive in practice because it's probably too 1548 // late anyway. 1549 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1550 assert(gch->collector_policy()->is_generation_policy(), 1551 "You may want to check the correctness of the following"); 1552 if (gch->incremental_collection_will_fail(true /* consult_young */)) { 1553 if (Verbose && PrintGCDetails) { 1554 gclog_or_tty->print("CMSCollector: collect because incremental collection will fail "); 1555 } 1556 return true; 1557 } 1558 1559 if (MetaspaceGC::should_concurrent_collect()) { 1560 if (Verbose && PrintGCDetails) { 1561 gclog_or_tty->print("CMSCollector: collect for metadata allocation "); 1562 } 1563 return true; 1564 } 1565 1566 // CMSTriggerInterval starts a CMS cycle if enough time has passed. 1567 if (CMSTriggerInterval >= 0) { 1568 if (CMSTriggerInterval == 0) { 1569 // Trigger always 1570 return true; 1571 } 1572 1573 // Check the CMS time since begin (we do not check the stats validity 1574 // as we want to be able to trigger the first CMS cycle as well) 1575 if (stats().cms_time_since_begin() >= (CMSTriggerInterval / ((double) MILLIUNITS))) { 1576 if (Verbose && PrintGCDetails) { 1577 if (stats().valid()) { 1578 gclog_or_tty->print_cr("CMSCollector: collect because of trigger interval (time since last begin %3.7f secs)", 1579 stats().cms_time_since_begin()); 1580 } else { 1581 gclog_or_tty->print_cr("CMSCollector: collect because of trigger interval (first collection)"); 1582 } 1583 } 1584 return true; 1585 } 1586 } 1587 1588 return false; 1589 } 1590 1591 void CMSCollector::set_did_compact(bool v) { _cmsGen->set_did_compact(v); } 1592 1593 // Clear _expansion_cause fields of constituent generations 1594 void CMSCollector::clear_expansion_cause() { 1595 _cmsGen->clear_expansion_cause(); 1596 } 1597 1598 // We should be conservative in starting a collection cycle. To 1599 // start too eagerly runs the risk of collecting too often in the 1600 // extreme. To collect too rarely falls back on full collections, 1601 // which works, even if not optimum in terms of concurrent work. 1602 // As a work around for too eagerly collecting, use the flag 1603 // UseCMSInitiatingOccupancyOnly. This also has the advantage of 1604 // giving the user an easily understandable way of controlling the 1605 // collections. 1606 // We want to start a new collection cycle if any of the following 1607 // conditions hold: 1608 // . our current occupancy exceeds the configured initiating occupancy 1609 // for this generation, or 1610 // . we recently needed to expand this space and have not, since that 1611 // expansion, done a collection of this generation, or 1612 // . the underlying space believes that it may be a good idea to initiate 1613 // a concurrent collection (this may be based on criteria such as the 1614 // following: the space uses linear allocation and linear allocation is 1615 // going to fail, or there is believed to be excessive fragmentation in 1616 // the generation, etc... or ... 1617 // [.(currently done by CMSCollector::shouldConcurrentCollect() only for 1618 // the case of the old generation; see CR 6543076): 1619 // we may be approaching a point at which allocation requests may fail because 1620 // we will be out of sufficient free space given allocation rate estimates.] 1621 bool ConcurrentMarkSweepGeneration::should_concurrent_collect() const { 1622 1623 assert_lock_strong(freelistLock()); 1624 if (occupancy() > initiating_occupancy()) { 1625 if (PrintGCDetails && Verbose) { 1626 gclog_or_tty->print(" %s: collect because of occupancy %f / %f ", 1627 short_name(), occupancy(), initiating_occupancy()); 1628 } 1629 return true; 1630 } 1631 if (UseCMSInitiatingOccupancyOnly) { 1632 return false; 1633 } 1634 if (expansion_cause() == CMSExpansionCause::_satisfy_allocation) { 1635 if (PrintGCDetails && Verbose) { 1636 gclog_or_tty->print(" %s: collect because expanded for allocation ", 1637 short_name()); 1638 } 1639 return true; 1640 } 1641 if (_cmsSpace->should_concurrent_collect()) { 1642 if (PrintGCDetails && Verbose) { 1643 gclog_or_tty->print(" %s: collect because cmsSpace says so ", 1644 short_name()); 1645 } 1646 return true; 1647 } 1648 return false; 1649 } 1650 1651 void ConcurrentMarkSweepGeneration::collect(bool full, 1652 bool clear_all_soft_refs, 1653 size_t size, 1654 bool tlab) 1655 { 1656 collector()->collect(full, clear_all_soft_refs, size, tlab); 1657 } 1658 1659 void CMSCollector::collect(bool full, 1660 bool clear_all_soft_refs, 1661 size_t size, 1662 bool tlab) 1663 { 1664 if (!UseCMSCollectionPassing && _collectorState > Idling) { 1665 // For debugging purposes skip the collection if the state 1666 // is not currently idle 1667 if (TraceCMSState) { 1668 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " skipped full:%d CMS state %d", 1669 Thread::current(), full, _collectorState); 1670 } 1671 return; 1672 } 1673 1674 // The following "if" branch is present for defensive reasons. 1675 // In the current uses of this interface, it can be replaced with: 1676 // assert(!GC_locker.is_active(), "Can't be called otherwise"); 1677 // But I am not placing that assert here to allow future 1678 // generality in invoking this interface. 1679 if (GC_locker::is_active()) { 1680 // A consistency test for GC_locker 1681 assert(GC_locker::needs_gc(), "Should have been set already"); 1682 // Skip this foreground collection, instead 1683 // expanding the heap if necessary. 1684 // Need the free list locks for the call to free() in compute_new_size() 1685 compute_new_size(); 1686 return; 1687 } 1688 acquire_control_and_collect(full, clear_all_soft_refs); 1689 _full_gcs_since_conc_gc++; 1690 } 1691 1692 void CMSCollector::request_full_gc(unsigned int full_gc_count, GCCause::Cause cause) { 1693 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1694 unsigned int gc_count = gch->total_full_collections(); 1695 if (gc_count == full_gc_count) { 1696 MutexLockerEx y(CGC_lock, Mutex::_no_safepoint_check_flag); 1697 _full_gc_requested = true; 1698 _full_gc_cause = cause; 1699 CGC_lock->notify(); // nudge CMS thread 1700 } else { 1701 assert(gc_count > full_gc_count, "Error: causal loop"); 1702 } 1703 } 1704 1705 bool CMSCollector::is_external_interruption() { 1706 GCCause::Cause cause = GenCollectedHeap::heap()->gc_cause(); 1707 return GCCause::is_user_requested_gc(cause) || 1708 GCCause::is_serviceability_requested_gc(cause); 1709 } 1710 1711 void CMSCollector::report_concurrent_mode_interruption() { 1712 if (is_external_interruption()) { 1713 if (PrintGCDetails) { 1714 gclog_or_tty->print(" (concurrent mode interrupted)"); 1715 } 1716 } else { 1717 if (PrintGCDetails) { 1718 gclog_or_tty->print(" (concurrent mode failure)"); 1719 } 1720 _gc_tracer_cm->report_concurrent_mode_failure(); 1721 } 1722 } 1723 1724 1725 // The foreground and background collectors need to coordinate in order 1726 // to make sure that they do not mutually interfere with CMS collections. 1727 // When a background collection is active, 1728 // the foreground collector may need to take over (preempt) and 1729 // synchronously complete an ongoing collection. Depending on the 1730 // frequency of the background collections and the heap usage 1731 // of the application, this preemption can be seldom or frequent. 1732 // There are only certain 1733 // points in the background collection that the "collection-baton" 1734 // can be passed to the foreground collector. 1735 // 1736 // The foreground collector will wait for the baton before 1737 // starting any part of the collection. The foreground collector 1738 // will only wait at one location. 1739 // 1740 // The background collector will yield the baton before starting a new 1741 // phase of the collection (e.g., before initial marking, marking from roots, 1742 // precleaning, final re-mark, sweep etc.) This is normally done at the head 1743 // of the loop which switches the phases. The background collector does some 1744 // of the phases (initial mark, final re-mark) with the world stopped. 1745 // Because of locking involved in stopping the world, 1746 // the foreground collector should not block waiting for the background 1747 // collector when it is doing a stop-the-world phase. The background 1748 // collector will yield the baton at an additional point just before 1749 // it enters a stop-the-world phase. Once the world is stopped, the 1750 // background collector checks the phase of the collection. If the 1751 // phase has not changed, it proceeds with the collection. If the 1752 // phase has changed, it skips that phase of the collection. See 1753 // the comments on the use of the Heap_lock in collect_in_background(). 1754 // 1755 // Variable used in baton passing. 1756 // _foregroundGCIsActive - Set to true by the foreground collector when 1757 // it wants the baton. The foreground clears it when it has finished 1758 // the collection. 1759 // _foregroundGCShouldWait - Set to true by the background collector 1760 // when it is running. The foreground collector waits while 1761 // _foregroundGCShouldWait is true. 1762 // CGC_lock - monitor used to protect access to the above variables 1763 // and to notify the foreground and background collectors. 1764 // _collectorState - current state of the CMS collection. 1765 // 1766 // The foreground collector 1767 // acquires the CGC_lock 1768 // sets _foregroundGCIsActive 1769 // waits on the CGC_lock for _foregroundGCShouldWait to be false 1770 // various locks acquired in preparation for the collection 1771 // are released so as not to block the background collector 1772 // that is in the midst of a collection 1773 // proceeds with the collection 1774 // clears _foregroundGCIsActive 1775 // returns 1776 // 1777 // The background collector in a loop iterating on the phases of the 1778 // collection 1779 // acquires the CGC_lock 1780 // sets _foregroundGCShouldWait 1781 // if _foregroundGCIsActive is set 1782 // clears _foregroundGCShouldWait, notifies _CGC_lock 1783 // waits on _CGC_lock for _foregroundGCIsActive to become false 1784 // and exits the loop. 1785 // otherwise 1786 // proceed with that phase of the collection 1787 // if the phase is a stop-the-world phase, 1788 // yield the baton once more just before enqueueing 1789 // the stop-world CMS operation (executed by the VM thread). 1790 // returns after all phases of the collection are done 1791 // 1792 1793 void CMSCollector::acquire_control_and_collect(bool full, 1794 bool clear_all_soft_refs) { 1795 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint"); 1796 assert(!Thread::current()->is_ConcurrentGC_thread(), 1797 "shouldn't try to acquire control from self!"); 1798 1799 // Start the protocol for acquiring control of the 1800 // collection from the background collector (aka CMS thread). 1801 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 1802 "VM thread should have CMS token"); 1803 // Remember the possibly interrupted state of an ongoing 1804 // concurrent collection 1805 CollectorState first_state = _collectorState; 1806 1807 // Signal to a possibly ongoing concurrent collection that 1808 // we want to do a foreground collection. 1809 _foregroundGCIsActive = true; 1810 1811 // Disable incremental mode during a foreground collection. 1812 ICMSDisabler icms_disabler; 1813 1814 // release locks and wait for a notify from the background collector 1815 // releasing the locks in only necessary for phases which 1816 // do yields to improve the granularity of the collection. 1817 assert_lock_strong(bitMapLock()); 1818 // We need to lock the Free list lock for the space that we are 1819 // currently collecting. 1820 assert(haveFreelistLocks(), "Must be holding free list locks"); 1821 bitMapLock()->unlock(); 1822 releaseFreelistLocks(); 1823 { 1824 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1825 if (_foregroundGCShouldWait) { 1826 // We are going to be waiting for action for the CMS thread; 1827 // it had better not be gone (for instance at shutdown)! 1828 assert(ConcurrentMarkSweepThread::cmst() != NULL, 1829 "CMS thread must be running"); 1830 // Wait here until the background collector gives us the go-ahead 1831 ConcurrentMarkSweepThread::clear_CMS_flag( 1832 ConcurrentMarkSweepThread::CMS_vm_has_token); // release token 1833 // Get a possibly blocked CMS thread going: 1834 // Note that we set _foregroundGCIsActive true above, 1835 // without protection of the CGC_lock. 1836 CGC_lock->notify(); 1837 assert(!ConcurrentMarkSweepThread::vm_thread_wants_cms_token(), 1838 "Possible deadlock"); 1839 while (_foregroundGCShouldWait) { 1840 // wait for notification 1841 CGC_lock->wait(Mutex::_no_safepoint_check_flag); 1842 // Possibility of delay/starvation here, since CMS token does 1843 // not know to give priority to VM thread? Actually, i think 1844 // there wouldn't be any delay/starvation, but the proof of 1845 // that "fact" (?) appears non-trivial. XXX 20011219YSR 1846 } 1847 ConcurrentMarkSweepThread::set_CMS_flag( 1848 ConcurrentMarkSweepThread::CMS_vm_has_token); 1849 } 1850 } 1851 // The CMS_token is already held. Get back the other locks. 1852 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 1853 "VM thread should have CMS token"); 1854 getFreelistLocks(); 1855 bitMapLock()->lock_without_safepoint_check(); 1856 if (TraceCMSState) { 1857 gclog_or_tty->print_cr("CMS foreground collector has asked for control " 1858 INTPTR_FORMAT " with first state %d", Thread::current(), first_state); 1859 gclog_or_tty->print_cr(" gets control with state %d", _collectorState); 1860 } 1861 1862 // Check if we need to do a compaction, or if not, whether 1863 // we need to start the mark-sweep from scratch. 1864 bool should_compact = false; 1865 bool should_start_over = false; 1866 decide_foreground_collection_type(clear_all_soft_refs, 1867 &should_compact, &should_start_over); 1868 1869 NOT_PRODUCT( 1870 if (RotateCMSCollectionTypes) { 1871 if (_cmsGen->debug_collection_type() == 1872 ConcurrentMarkSweepGeneration::MSC_foreground_collection_type) { 1873 should_compact = true; 1874 } else if (_cmsGen->debug_collection_type() == 1875 ConcurrentMarkSweepGeneration::MS_foreground_collection_type) { 1876 should_compact = false; 1877 } 1878 } 1879 ) 1880 1881 if (first_state > Idling) { 1882 report_concurrent_mode_interruption(); 1883 } 1884 1885 set_did_compact(should_compact); 1886 if (should_compact) { 1887 // If the collection is being acquired from the background 1888 // collector, there may be references on the discovered 1889 // references lists that have NULL referents (being those 1890 // that were concurrently cleared by a mutator) or 1891 // that are no longer active (having been enqueued concurrently 1892 // by the mutator). 1893 // Scrub the list of those references because Mark-Sweep-Compact 1894 // code assumes referents are not NULL and that all discovered 1895 // Reference objects are active. 1896 ref_processor()->clean_up_discovered_references(); 1897 1898 if (first_state > Idling) { 1899 save_heap_summary(); 1900 } 1901 1902 do_compaction_work(clear_all_soft_refs); 1903 1904 // Has the GC time limit been exceeded? 1905 DefNewGeneration* young_gen = _young_gen->as_DefNewGeneration(); 1906 size_t max_eden_size = young_gen->max_capacity() - 1907 young_gen->to()->capacity() - 1908 young_gen->from()->capacity(); 1909 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1910 GCCause::Cause gc_cause = gch->gc_cause(); 1911 size_policy()->check_gc_overhead_limit(_young_gen->used(), 1912 young_gen->eden()->used(), 1913 _cmsGen->max_capacity(), 1914 max_eden_size, 1915 full, 1916 gc_cause, 1917 gch->collector_policy()); 1918 } else { 1919 do_mark_sweep_work(clear_all_soft_refs, first_state, 1920 should_start_over); 1921 } 1922 // Reset the expansion cause, now that we just completed 1923 // a collection cycle. 1924 clear_expansion_cause(); 1925 _foregroundGCIsActive = false; 1926 return; 1927 } 1928 1929 // Resize the tenured generation 1930 // after obtaining the free list locks for the 1931 // two generations. 1932 void CMSCollector::compute_new_size() { 1933 assert_locked_or_safepoint(Heap_lock); 1934 FreelistLocker z(this); 1935 MetaspaceGC::compute_new_size(); 1936 _cmsGen->compute_new_size_free_list(); 1937 } 1938 1939 // A work method used by foreground collection to determine 1940 // what type of collection (compacting or not, continuing or fresh) 1941 // it should do. 1942 // NOTE: the intent is to make UseCMSCompactAtFullCollection 1943 // and CMSCompactWhenClearAllSoftRefs the default in the future 1944 // and do away with the flags after a suitable period. 1945 void CMSCollector::decide_foreground_collection_type( 1946 bool clear_all_soft_refs, bool* should_compact, 1947 bool* should_start_over) { 1948 // Normally, we'll compact only if the UseCMSCompactAtFullCollection 1949 // flag is set, and we have either requested a System.gc() or 1950 // the number of full gc's since the last concurrent cycle 1951 // has exceeded the threshold set by CMSFullGCsBeforeCompaction, 1952 // or if an incremental collection has failed 1953 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1954 assert(gch->collector_policy()->is_generation_policy(), 1955 "You may want to check the correctness of the following"); 1956 // Inform cms gen if this was due to partial collection failing. 1957 // The CMS gen may use this fact to determine its expansion policy. 1958 if (gch->incremental_collection_will_fail(false /* don't consult_young */)) { 1959 assert(!_cmsGen->incremental_collection_failed(), 1960 "Should have been noticed, reacted to and cleared"); 1961 _cmsGen->set_incremental_collection_failed(); 1962 } 1963 *should_compact = 1964 UseCMSCompactAtFullCollection && 1965 ((_full_gcs_since_conc_gc >= CMSFullGCsBeforeCompaction) || 1966 GCCause::is_user_requested_gc(gch->gc_cause()) || 1967 gch->incremental_collection_will_fail(true /* consult_young */)); 1968 *should_start_over = false; 1969 if (clear_all_soft_refs && !*should_compact) { 1970 // We are about to do a last ditch collection attempt 1971 // so it would normally make sense to do a compaction 1972 // to reclaim as much space as possible. 1973 if (CMSCompactWhenClearAllSoftRefs) { 1974 // Default: The rationale is that in this case either 1975 // we are past the final marking phase, in which case 1976 // we'd have to start over, or so little has been done 1977 // that there's little point in saving that work. Compaction 1978 // appears to be the sensible choice in either case. 1979 *should_compact = true; 1980 } else { 1981 // We have been asked to clear all soft refs, but not to 1982 // compact. Make sure that we aren't past the final checkpoint 1983 // phase, for that is where we process soft refs. If we are already 1984 // past that phase, we'll need to redo the refs discovery phase and 1985 // if necessary clear soft refs that weren't previously 1986 // cleared. We do so by remembering the phase in which 1987 // we came in, and if we are past the refs processing 1988 // phase, we'll choose to just redo the mark-sweep 1989 // collection from scratch. 1990 if (_collectorState > FinalMarking) { 1991 // We are past the refs processing phase; 1992 // start over and do a fresh synchronous CMS cycle 1993 _collectorState = Resetting; // skip to reset to start new cycle 1994 reset(false /* == !asynch */); 1995 *should_start_over = true; 1996 } // else we can continue a possibly ongoing current cycle 1997 } 1998 } 1999 } 2000 2001 // A work method used by the foreground collector to do 2002 // a mark-sweep-compact. 2003 void CMSCollector::do_compaction_work(bool clear_all_soft_refs) { 2004 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2005 2006 STWGCTimer* gc_timer = GenMarkSweep::gc_timer(); 2007 gc_timer->register_gc_start(); 2008 2009 SerialOldTracer* gc_tracer = GenMarkSweep::gc_tracer(); 2010 gc_tracer->report_gc_start(gch->gc_cause(), gc_timer->gc_start()); 2011 2012 GCTraceTime t("CMS:MSC ", PrintGCDetails && Verbose, true, NULL, gc_tracer->gc_id()); 2013 if (PrintGC && Verbose && !(GCCause::is_user_requested_gc(gch->gc_cause()))) { 2014 gclog_or_tty->print_cr("Compact ConcurrentMarkSweepGeneration after %d " 2015 "collections passed to foreground collector", _full_gcs_since_conc_gc); 2016 } 2017 2018 // Temporarily widen the span of the weak reference processing to 2019 // the entire heap. 2020 MemRegion new_span(GenCollectedHeap::heap()->reserved_region()); 2021 ReferenceProcessorSpanMutator rp_mut_span(ref_processor(), new_span); 2022 // Temporarily, clear the "is_alive_non_header" field of the 2023 // reference processor. 2024 ReferenceProcessorIsAliveMutator rp_mut_closure(ref_processor(), NULL); 2025 // Temporarily make reference _processing_ single threaded (non-MT). 2026 ReferenceProcessorMTProcMutator rp_mut_mt_processing(ref_processor(), false); 2027 // Temporarily make refs discovery atomic 2028 ReferenceProcessorAtomicMutator rp_mut_atomic(ref_processor(), true); 2029 // Temporarily make reference _discovery_ single threaded (non-MT) 2030 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false); 2031 2032 ref_processor()->set_enqueuing_is_done(false); 2033 ref_processor()->enable_discovery(false /*verify_disabled*/, false /*check_no_refs*/); 2034 ref_processor()->setup_policy(clear_all_soft_refs); 2035 // If an asynchronous collection finishes, the _modUnionTable is 2036 // all clear. If we are assuming the collection from an asynchronous 2037 // collection, clear the _modUnionTable. 2038 assert(_collectorState != Idling || _modUnionTable.isAllClear(), 2039 "_modUnionTable should be clear if the baton was not passed"); 2040 _modUnionTable.clear_all(); 2041 assert(_collectorState != Idling || _ct->klass_rem_set()->mod_union_is_clear(), 2042 "mod union for klasses should be clear if the baton was passed"); 2043 _ct->klass_rem_set()->clear_mod_union(); 2044 2045 // We must adjust the allocation statistics being maintained 2046 // in the free list space. We do so by reading and clearing 2047 // the sweep timer and updating the block flux rate estimates below. 2048 assert(!_intra_sweep_timer.is_active(), "_intra_sweep_timer should be inactive"); 2049 if (_inter_sweep_timer.is_active()) { 2050 _inter_sweep_timer.stop(); 2051 // Note that we do not use this sample to update the _inter_sweep_estimate. 2052 _cmsGen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()), 2053 _inter_sweep_estimate.padded_average(), 2054 _intra_sweep_estimate.padded_average()); 2055 } 2056 2057 GenMarkSweep::invoke_at_safepoint(_cmsGen->level(), 2058 ref_processor(), clear_all_soft_refs); 2059 #ifdef ASSERT 2060 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 2061 size_t free_size = cms_space->free(); 2062 assert(free_size == 2063 pointer_delta(cms_space->end(), cms_space->compaction_top()) 2064 * HeapWordSize, 2065 "All the free space should be compacted into one chunk at top"); 2066 assert(cms_space->dictionary()->total_chunk_size( 2067 debug_only(cms_space->freelistLock())) == 0 || 2068 cms_space->totalSizeInIndexedFreeLists() == 0, 2069 "All the free space should be in a single chunk"); 2070 size_t num = cms_space->totalCount(); 2071 assert((free_size == 0 && num == 0) || 2072 (free_size > 0 && (num == 1 || num == 2)), 2073 "There should be at most 2 free chunks after compaction"); 2074 #endif // ASSERT 2075 _collectorState = Resetting; 2076 assert(_restart_addr == NULL, 2077 "Should have been NULL'd before baton was passed"); 2078 reset(false /* == !asynch */); 2079 _cmsGen->reset_after_compaction(); 2080 _concurrent_cycles_since_last_unload = 0; 2081 2082 // Clear any data recorded in the PLAB chunk arrays. 2083 if (_survivor_plab_array != NULL) { 2084 reset_survivor_plab_arrays(); 2085 } 2086 2087 // Adjust the per-size allocation stats for the next epoch. 2088 _cmsGen->cmsSpace()->endSweepFLCensus(sweep_count() /* fake */); 2089 // Restart the "inter sweep timer" for the next epoch. 2090 _inter_sweep_timer.reset(); 2091 _inter_sweep_timer.start(); 2092 2093 gc_timer->register_gc_end(); 2094 2095 gc_tracer->report_gc_end(gc_timer->gc_end(), gc_timer->time_partitions()); 2096 2097 // For a mark-sweep-compact, compute_new_size() will be called 2098 // in the heap's do_collection() method. 2099 } 2100 2101 // A work method used by the foreground collector to do 2102 // a mark-sweep, after taking over from a possibly on-going 2103 // concurrent mark-sweep collection. 2104 void CMSCollector::do_mark_sweep_work(bool clear_all_soft_refs, 2105 CollectorState first_state, bool should_start_over) { 2106 if (PrintGC && Verbose) { 2107 gclog_or_tty->print_cr("Pass concurrent collection to foreground " 2108 "collector with count %d", 2109 _full_gcs_since_conc_gc); 2110 } 2111 switch (_collectorState) { 2112 case Idling: 2113 if (first_state == Idling || should_start_over) { 2114 // The background GC was not active, or should 2115 // restarted from scratch; start the cycle. 2116 _collectorState = InitialMarking; 2117 } 2118 // If first_state was not Idling, then a background GC 2119 // was in progress and has now finished. No need to do it 2120 // again. Leave the state as Idling. 2121 break; 2122 case Precleaning: 2123 // In the foreground case don't do the precleaning since 2124 // it is not done concurrently and there is extra work 2125 // required. 2126 _collectorState = FinalMarking; 2127 } 2128 collect_in_foreground(clear_all_soft_refs, GenCollectedHeap::heap()->gc_cause()); 2129 2130 // For a mark-sweep, compute_new_size() will be called 2131 // in the heap's do_collection() method. 2132 } 2133 2134 2135 void CMSCollector::print_eden_and_survivor_chunk_arrays() { 2136 DefNewGeneration* dng = _young_gen->as_DefNewGeneration(); 2137 EdenSpace* eden_space = dng->eden(); 2138 ContiguousSpace* from_space = dng->from(); 2139 ContiguousSpace* to_space = dng->to(); 2140 // Eden 2141 if (_eden_chunk_array != NULL) { 2142 gclog_or_tty->print_cr("eden " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")", 2143 eden_space->bottom(), eden_space->top(), 2144 eden_space->end(), eden_space->capacity()); 2145 gclog_or_tty->print_cr("_eden_chunk_index=" SIZE_FORMAT ", " 2146 "_eden_chunk_capacity=" SIZE_FORMAT, 2147 _eden_chunk_index, _eden_chunk_capacity); 2148 for (size_t i = 0; i < _eden_chunk_index; i++) { 2149 gclog_or_tty->print_cr("_eden_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT, 2150 i, _eden_chunk_array[i]); 2151 } 2152 } 2153 // Survivor 2154 if (_survivor_chunk_array != NULL) { 2155 gclog_or_tty->print_cr("survivor " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")", 2156 from_space->bottom(), from_space->top(), 2157 from_space->end(), from_space->capacity()); 2158 gclog_or_tty->print_cr("_survivor_chunk_index=" SIZE_FORMAT ", " 2159 "_survivor_chunk_capacity=" SIZE_FORMAT, 2160 _survivor_chunk_index, _survivor_chunk_capacity); 2161 for (size_t i = 0; i < _survivor_chunk_index; i++) { 2162 gclog_or_tty->print_cr("_survivor_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT, 2163 i, _survivor_chunk_array[i]); 2164 } 2165 } 2166 } 2167 2168 void CMSCollector::getFreelistLocks() const { 2169 // Get locks for all free lists in all generations that this 2170 // collector is responsible for 2171 _cmsGen->freelistLock()->lock_without_safepoint_check(); 2172 } 2173 2174 void CMSCollector::releaseFreelistLocks() const { 2175 // Release locks for all free lists in all generations that this 2176 // collector is responsible for 2177 _cmsGen->freelistLock()->unlock(); 2178 } 2179 2180 bool CMSCollector::haveFreelistLocks() const { 2181 // Check locks for all free lists in all generations that this 2182 // collector is responsible for 2183 assert_lock_strong(_cmsGen->freelistLock()); 2184 PRODUCT_ONLY(ShouldNotReachHere()); 2185 return true; 2186 } 2187 2188 // A utility class that is used by the CMS collector to 2189 // temporarily "release" the foreground collector from its 2190 // usual obligation to wait for the background collector to 2191 // complete an ongoing phase before proceeding. 2192 class ReleaseForegroundGC: public StackObj { 2193 private: 2194 CMSCollector* _c; 2195 public: 2196 ReleaseForegroundGC(CMSCollector* c) : _c(c) { 2197 assert(_c->_foregroundGCShouldWait, "Else should not need to call"); 2198 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 2199 // allow a potentially blocked foreground collector to proceed 2200 _c->_foregroundGCShouldWait = false; 2201 if (_c->_foregroundGCIsActive) { 2202 CGC_lock->notify(); 2203 } 2204 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 2205 "Possible deadlock"); 2206 } 2207 2208 ~ReleaseForegroundGC() { 2209 assert(!_c->_foregroundGCShouldWait, "Usage protocol violation?"); 2210 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 2211 _c->_foregroundGCShouldWait = true; 2212 } 2213 }; 2214 2215 // There are separate collect_in_background and collect_in_foreground because of 2216 // the different locking requirements of the background collector and the 2217 // foreground collector. There was originally an attempt to share 2218 // one "collect" method between the background collector and the foreground 2219 // collector but the if-then-else required made it cleaner to have 2220 // separate methods. 2221 void CMSCollector::collect_in_background(bool clear_all_soft_refs, GCCause::Cause cause) { 2222 assert(Thread::current()->is_ConcurrentGC_thread(), 2223 "A CMS asynchronous collection is only allowed on a CMS thread."); 2224 2225 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2226 { 2227 bool safepoint_check = Mutex::_no_safepoint_check_flag; 2228 MutexLockerEx hl(Heap_lock, safepoint_check); 2229 FreelistLocker fll(this); 2230 MutexLockerEx x(CGC_lock, safepoint_check); 2231 if (_foregroundGCIsActive || !UseAsyncConcMarkSweepGC) { 2232 // The foreground collector is active or we're 2233 // not using asynchronous collections. Skip this 2234 // background collection. 2235 assert(!_foregroundGCShouldWait, "Should be clear"); 2236 return; 2237 } else { 2238 assert(_collectorState == Idling, "Should be idling before start."); 2239 _collectorState = InitialMarking; 2240 register_gc_start(cause); 2241 // Reset the expansion cause, now that we are about to begin 2242 // a new cycle. 2243 clear_expansion_cause(); 2244 2245 // Clear the MetaspaceGC flag since a concurrent collection 2246 // is starting but also clear it after the collection. 2247 MetaspaceGC::set_should_concurrent_collect(false); 2248 } 2249 // Decide if we want to enable class unloading as part of the 2250 // ensuing concurrent GC cycle. 2251 update_should_unload_classes(); 2252 _full_gc_requested = false; // acks all outstanding full gc requests 2253 _full_gc_cause = GCCause::_no_gc; 2254 // Signal that we are about to start a collection 2255 gch->increment_total_full_collections(); // ... starting a collection cycle 2256 _collection_count_start = gch->total_full_collections(); 2257 } 2258 2259 // Used for PrintGC 2260 size_t prev_used; 2261 if (PrintGC && Verbose) { 2262 prev_used = _cmsGen->used(); // XXXPERM 2263 } 2264 2265 // The change of the collection state is normally done at this level; 2266 // the exceptions are phases that are executed while the world is 2267 // stopped. For those phases the change of state is done while the 2268 // world is stopped. For baton passing purposes this allows the 2269 // background collector to finish the phase and change state atomically. 2270 // The foreground collector cannot wait on a phase that is done 2271 // while the world is stopped because the foreground collector already 2272 // has the world stopped and would deadlock. 2273 while (_collectorState != Idling) { 2274 if (TraceCMSState) { 2275 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " in CMS state %d", 2276 Thread::current(), _collectorState); 2277 } 2278 // The foreground collector 2279 // holds the Heap_lock throughout its collection. 2280 // holds the CMS token (but not the lock) 2281 // except while it is waiting for the background collector to yield. 2282 // 2283 // The foreground collector should be blocked (not for long) 2284 // if the background collector is about to start a phase 2285 // executed with world stopped. If the background 2286 // collector has already started such a phase, the 2287 // foreground collector is blocked waiting for the 2288 // Heap_lock. The stop-world phases (InitialMarking and FinalMarking) 2289 // are executed in the VM thread. 2290 // 2291 // The locking order is 2292 // PendingListLock (PLL) -- if applicable (FinalMarking) 2293 // Heap_lock (both this & PLL locked in VM_CMS_Operation::prologue()) 2294 // CMS token (claimed in 2295 // stop_world_and_do() --> 2296 // safepoint_synchronize() --> 2297 // CMSThread::synchronize()) 2298 2299 { 2300 // Check if the FG collector wants us to yield. 2301 CMSTokenSync x(true); // is cms thread 2302 if (waitForForegroundGC()) { 2303 // We yielded to a foreground GC, nothing more to be 2304 // done this round. 2305 assert(_foregroundGCShouldWait == false, "We set it to false in " 2306 "waitForForegroundGC()"); 2307 if (TraceCMSState) { 2308 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT 2309 " exiting collection CMS state %d", 2310 Thread::current(), _collectorState); 2311 } 2312 return; 2313 } else { 2314 // The background collector can run but check to see if the 2315 // foreground collector has done a collection while the 2316 // background collector was waiting to get the CGC_lock 2317 // above. If yes, break so that _foregroundGCShouldWait 2318 // is cleared before returning. 2319 if (_collectorState == Idling) { 2320 break; 2321 } 2322 } 2323 } 2324 2325 assert(_foregroundGCShouldWait, "Foreground collector, if active, " 2326 "should be waiting"); 2327 2328 switch (_collectorState) { 2329 case InitialMarking: 2330 { 2331 ReleaseForegroundGC x(this); 2332 stats().record_cms_begin(); 2333 VM_CMS_Initial_Mark initial_mark_op(this); 2334 VMThread::execute(&initial_mark_op); 2335 } 2336 // The collector state may be any legal state at this point 2337 // since the background collector may have yielded to the 2338 // foreground collector. 2339 break; 2340 case Marking: 2341 // initial marking in checkpointRootsInitialWork has been completed 2342 if (markFromRoots(true)) { // we were successful 2343 assert(_collectorState == Precleaning, "Collector state should " 2344 "have changed"); 2345 } else { 2346 assert(_foregroundGCIsActive, "Internal state inconsistency"); 2347 } 2348 break; 2349 case Precleaning: 2350 // marking from roots in markFromRoots has been completed 2351 preclean(); 2352 assert(_collectorState == AbortablePreclean || 2353 _collectorState == FinalMarking, 2354 "Collector state should have changed"); 2355 break; 2356 case AbortablePreclean: 2357 abortable_preclean(); 2358 assert(_collectorState == FinalMarking, "Collector state should " 2359 "have changed"); 2360 break; 2361 case FinalMarking: 2362 { 2363 ReleaseForegroundGC x(this); 2364 2365 VM_CMS_Final_Remark final_remark_op(this); 2366 VMThread::execute(&final_remark_op); 2367 } 2368 assert(_foregroundGCShouldWait, "block post-condition"); 2369 break; 2370 case Sweeping: 2371 // final marking in checkpointRootsFinal has been completed 2372 sweep(true); 2373 assert(_collectorState == Resizing, "Collector state change " 2374 "to Resizing must be done under the free_list_lock"); 2375 _full_gcs_since_conc_gc = 0; 2376 2377 case Resizing: { 2378 // Sweeping has been completed... 2379 // At this point the background collection has completed. 2380 // Don't move the call to compute_new_size() down 2381 // into code that might be executed if the background 2382 // collection was preempted. 2383 { 2384 ReleaseForegroundGC x(this); // unblock FG collection 2385 MutexLockerEx y(Heap_lock, Mutex::_no_safepoint_check_flag); 2386 CMSTokenSync z(true); // not strictly needed. 2387 if (_collectorState == Resizing) { 2388 compute_new_size(); 2389 save_heap_summary(); 2390 _collectorState = Resetting; 2391 } else { 2392 assert(_collectorState == Idling, "The state should only change" 2393 " because the foreground collector has finished the collection"); 2394 } 2395 } 2396 break; 2397 } 2398 case Resetting: 2399 // CMS heap resizing has been completed 2400 reset(true); 2401 assert(_collectorState == Idling, "Collector state should " 2402 "have changed"); 2403 2404 MetaspaceGC::set_should_concurrent_collect(false); 2405 2406 stats().record_cms_end(); 2407 // Don't move the concurrent_phases_end() and compute_new_size() 2408 // calls to here because a preempted background collection 2409 // has it's state set to "Resetting". 2410 break; 2411 case Idling: 2412 default: 2413 ShouldNotReachHere(); 2414 break; 2415 } 2416 if (TraceCMSState) { 2417 gclog_or_tty->print_cr(" Thread " INTPTR_FORMAT " done - next CMS state %d", 2418 Thread::current(), _collectorState); 2419 } 2420 assert(_foregroundGCShouldWait, "block post-condition"); 2421 } 2422 2423 // Should this be in gc_epilogue? 2424 collector_policy()->counters()->update_counters(); 2425 2426 { 2427 // Clear _foregroundGCShouldWait and, in the event that the 2428 // foreground collector is waiting, notify it, before 2429 // returning. 2430 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 2431 _foregroundGCShouldWait = false; 2432 if (_foregroundGCIsActive) { 2433 CGC_lock->notify(); 2434 } 2435 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 2436 "Possible deadlock"); 2437 } 2438 if (TraceCMSState) { 2439 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT 2440 " exiting collection CMS state %d", 2441 Thread::current(), _collectorState); 2442 } 2443 if (PrintGC && Verbose) { 2444 _cmsGen->print_heap_change(prev_used); 2445 } 2446 } 2447 2448 void CMSCollector::register_foreground_gc_start(GCCause::Cause cause) { 2449 if (!_cms_start_registered) { 2450 register_gc_start(cause); 2451 } 2452 } 2453 2454 void CMSCollector::register_gc_start(GCCause::Cause cause) { 2455 _cms_start_registered = true; 2456 _gc_timer_cm->register_gc_start(); 2457 _gc_tracer_cm->report_gc_start(cause, _gc_timer_cm->gc_start()); 2458 } 2459 2460 void CMSCollector::register_gc_end() { 2461 if (_cms_start_registered) { 2462 report_heap_summary(GCWhen::AfterGC); 2463 2464 _gc_timer_cm->register_gc_end(); 2465 _gc_tracer_cm->report_gc_end(_gc_timer_cm->gc_end(), _gc_timer_cm->time_partitions()); 2466 _cms_start_registered = false; 2467 } 2468 } 2469 2470 void CMSCollector::save_heap_summary() { 2471 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2472 _last_heap_summary = gch->create_heap_summary(); 2473 _last_metaspace_summary = gch->create_metaspace_summary(); 2474 } 2475 2476 void CMSCollector::report_heap_summary(GCWhen::Type when) { 2477 _gc_tracer_cm->report_gc_heap_summary(when, _last_heap_summary); 2478 _gc_tracer_cm->report_metaspace_summary(when, _last_metaspace_summary); 2479 } 2480 2481 void CMSCollector::collect_in_foreground(bool clear_all_soft_refs, GCCause::Cause cause) { 2482 assert(_foregroundGCIsActive && !_foregroundGCShouldWait, 2483 "Foreground collector should be waiting, not executing"); 2484 assert(Thread::current()->is_VM_thread(), "A foreground collection" 2485 "may only be done by the VM Thread with the world stopped"); 2486 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 2487 "VM thread should have CMS token"); 2488 2489 // The gc id is created in register_foreground_gc_start if this collection is synchronous 2490 const GCId gc_id = _collectorState == InitialMarking ? GCId::peek() : _gc_tracer_cm->gc_id(); 2491 NOT_PRODUCT(GCTraceTime t("CMS:MS (foreground) ", PrintGCDetails && Verbose, 2492 true, NULL, gc_id);) 2493 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact); 2494 2495 HandleMark hm; // Discard invalid handles created during verification 2496 2497 if (VerifyBeforeGC && 2498 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2499 Universe::verify(); 2500 } 2501 2502 // Snapshot the soft reference policy to be used in this collection cycle. 2503 ref_processor()->setup_policy(clear_all_soft_refs); 2504 2505 // Decide if class unloading should be done 2506 update_should_unload_classes(); 2507 2508 bool init_mark_was_synchronous = false; // until proven otherwise 2509 while (_collectorState != Idling) { 2510 if (TraceCMSState) { 2511 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " in CMS state %d", 2512 Thread::current(), _collectorState); 2513 } 2514 switch (_collectorState) { 2515 case InitialMarking: 2516 register_foreground_gc_start(cause); 2517 init_mark_was_synchronous = true; // fact to be exploited in re-mark 2518 checkpointRootsInitial(false); 2519 assert(_collectorState == Marking, "Collector state should have changed" 2520 " within checkpointRootsInitial()"); 2521 break; 2522 case Marking: 2523 // initial marking in checkpointRootsInitialWork has been completed 2524 if (VerifyDuringGC && 2525 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2526 Universe::verify("Verify before initial mark: "); 2527 } 2528 { 2529 bool res = markFromRoots(false); 2530 assert(res && _collectorState == FinalMarking, "Collector state should " 2531 "have changed"); 2532 break; 2533 } 2534 case FinalMarking: 2535 if (VerifyDuringGC && 2536 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2537 Universe::verify("Verify before re-mark: "); 2538 } 2539 checkpointRootsFinal(false, clear_all_soft_refs, 2540 init_mark_was_synchronous); 2541 assert(_collectorState == Sweeping, "Collector state should not " 2542 "have changed within checkpointRootsFinal()"); 2543 break; 2544 case Sweeping: 2545 // final marking in checkpointRootsFinal has been completed 2546 if (VerifyDuringGC && 2547 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2548 Universe::verify("Verify before sweep: "); 2549 } 2550 sweep(false); 2551 assert(_collectorState == Resizing, "Incorrect state"); 2552 break; 2553 case Resizing: { 2554 // Sweeping has been completed; the actual resize in this case 2555 // is done separately; nothing to be done in this state. 2556 _collectorState = Resetting; 2557 break; 2558 } 2559 case Resetting: 2560 // The heap has been resized. 2561 if (VerifyDuringGC && 2562 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2563 Universe::verify("Verify before reset: "); 2564 } 2565 save_heap_summary(); 2566 reset(false); 2567 assert(_collectorState == Idling, "Collector state should " 2568 "have changed"); 2569 break; 2570 case Precleaning: 2571 case AbortablePreclean: 2572 // Elide the preclean phase 2573 _collectorState = FinalMarking; 2574 break; 2575 default: 2576 ShouldNotReachHere(); 2577 } 2578 if (TraceCMSState) { 2579 gclog_or_tty->print_cr(" Thread " INTPTR_FORMAT " done - next CMS state %d", 2580 Thread::current(), _collectorState); 2581 } 2582 } 2583 2584 if (VerifyAfterGC && 2585 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2586 Universe::verify(); 2587 } 2588 if (TraceCMSState) { 2589 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT 2590 " exiting collection CMS state %d", 2591 Thread::current(), _collectorState); 2592 } 2593 } 2594 2595 bool CMSCollector::waitForForegroundGC() { 2596 bool res = false; 2597 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 2598 "CMS thread should have CMS token"); 2599 // Block the foreground collector until the 2600 // background collectors decides whether to 2601 // yield. 2602 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 2603 _foregroundGCShouldWait = true; 2604 if (_foregroundGCIsActive) { 2605 // The background collector yields to the 2606 // foreground collector and returns a value 2607 // indicating that it has yielded. The foreground 2608 // collector can proceed. 2609 res = true; 2610 _foregroundGCShouldWait = false; 2611 ConcurrentMarkSweepThread::clear_CMS_flag( 2612 ConcurrentMarkSweepThread::CMS_cms_has_token); 2613 ConcurrentMarkSweepThread::set_CMS_flag( 2614 ConcurrentMarkSweepThread::CMS_cms_wants_token); 2615 // Get a possibly blocked foreground thread going 2616 CGC_lock->notify(); 2617 if (TraceCMSState) { 2618 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " waiting at CMS state %d", 2619 Thread::current(), _collectorState); 2620 } 2621 while (_foregroundGCIsActive) { 2622 CGC_lock->wait(Mutex::_no_safepoint_check_flag); 2623 } 2624 ConcurrentMarkSweepThread::set_CMS_flag( 2625 ConcurrentMarkSweepThread::CMS_cms_has_token); 2626 ConcurrentMarkSweepThread::clear_CMS_flag( 2627 ConcurrentMarkSweepThread::CMS_cms_wants_token); 2628 } 2629 if (TraceCMSState) { 2630 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " continuing at CMS state %d", 2631 Thread::current(), _collectorState); 2632 } 2633 return res; 2634 } 2635 2636 // Because of the need to lock the free lists and other structures in 2637 // the collector, common to all the generations that the collector is 2638 // collecting, we need the gc_prologues of individual CMS generations 2639 // delegate to their collector. It may have been simpler had the 2640 // current infrastructure allowed one to call a prologue on a 2641 // collector. In the absence of that we have the generation's 2642 // prologue delegate to the collector, which delegates back 2643 // some "local" work to a worker method in the individual generations 2644 // that it's responsible for collecting, while itself doing any 2645 // work common to all generations it's responsible for. A similar 2646 // comment applies to the gc_epilogue()'s. 2647 // The role of the variable _between_prologue_and_epilogue is to 2648 // enforce the invocation protocol. 2649 void CMSCollector::gc_prologue(bool full) { 2650 // Call gc_prologue_work() for the CMSGen 2651 // we are responsible for. 2652 2653 // The following locking discipline assumes that we are only called 2654 // when the world is stopped. 2655 assert(SafepointSynchronize::is_at_safepoint(), "world is stopped assumption"); 2656 2657 // The CMSCollector prologue must call the gc_prologues for the 2658 // "generations" that it's responsible 2659 // for. 2660 2661 assert( Thread::current()->is_VM_thread() 2662 || ( CMSScavengeBeforeRemark 2663 && Thread::current()->is_ConcurrentGC_thread()), 2664 "Incorrect thread type for prologue execution"); 2665 2666 if (_between_prologue_and_epilogue) { 2667 // We have already been invoked; this is a gc_prologue delegation 2668 // from yet another CMS generation that we are responsible for, just 2669 // ignore it since all relevant work has already been done. 2670 return; 2671 } 2672 2673 // set a bit saying prologue has been called; cleared in epilogue 2674 _between_prologue_and_epilogue = true; 2675 // Claim locks for common data structures, then call gc_prologue_work() 2676 // for each CMSGen. 2677 2678 getFreelistLocks(); // gets free list locks on constituent spaces 2679 bitMapLock()->lock_without_safepoint_check(); 2680 2681 // Should call gc_prologue_work() for all cms gens we are responsible for 2682 bool duringMarking = _collectorState >= Marking 2683 && _collectorState < Sweeping; 2684 2685 // The young collections clear the modified oops state, which tells if 2686 // there are any modified oops in the class. The remark phase also needs 2687 // that information. Tell the young collection to save the union of all 2688 // modified klasses. 2689 if (duringMarking) { 2690 _ct->klass_rem_set()->set_accumulate_modified_oops(true); 2691 } 2692 2693 bool registerClosure = duringMarking; 2694 2695 ModUnionClosure* muc = CollectedHeap::use_parallel_gc_threads() ? 2696 &_modUnionClosurePar 2697 : &_modUnionClosure; 2698 _cmsGen->gc_prologue_work(full, registerClosure, muc); 2699 2700 if (!full) { 2701 stats().record_gc0_begin(); 2702 } 2703 } 2704 2705 void ConcurrentMarkSweepGeneration::gc_prologue(bool full) { 2706 2707 _capacity_at_prologue = capacity(); 2708 _used_at_prologue = used(); 2709 2710 // Delegate to CMScollector which knows how to coordinate between 2711 // this and any other CMS generations that it is responsible for 2712 // collecting. 2713 collector()->gc_prologue(full); 2714 } 2715 2716 // This is a "private" interface for use by this generation's CMSCollector. 2717 // Not to be called directly by any other entity (for instance, 2718 // GenCollectedHeap, which calls the "public" gc_prologue method above). 2719 void ConcurrentMarkSweepGeneration::gc_prologue_work(bool full, 2720 bool registerClosure, ModUnionClosure* modUnionClosure) { 2721 assert(!incremental_collection_failed(), "Shouldn't be set yet"); 2722 assert(cmsSpace()->preconsumptionDirtyCardClosure() == NULL, 2723 "Should be NULL"); 2724 if (registerClosure) { 2725 cmsSpace()->setPreconsumptionDirtyCardClosure(modUnionClosure); 2726 } 2727 cmsSpace()->gc_prologue(); 2728 // Clear stat counters 2729 NOT_PRODUCT( 2730 assert(_numObjectsPromoted == 0, "check"); 2731 assert(_numWordsPromoted == 0, "check"); 2732 if (Verbose && PrintGC) { 2733 gclog_or_tty->print("Allocated "SIZE_FORMAT" objects, " 2734 SIZE_FORMAT" bytes concurrently", 2735 _numObjectsAllocated, _numWordsAllocated*sizeof(HeapWord)); 2736 } 2737 _numObjectsAllocated = 0; 2738 _numWordsAllocated = 0; 2739 ) 2740 } 2741 2742 void CMSCollector::gc_epilogue(bool full) { 2743 // The following locking discipline assumes that we are only called 2744 // when the world is stopped. 2745 assert(SafepointSynchronize::is_at_safepoint(), 2746 "world is stopped assumption"); 2747 2748 // Currently the CMS epilogue (see CompactibleFreeListSpace) merely checks 2749 // if linear allocation blocks need to be appropriately marked to allow the 2750 // the blocks to be parsable. We also check here whether we need to nudge the 2751 // CMS collector thread to start a new cycle (if it's not already active). 2752 assert( Thread::current()->is_VM_thread() 2753 || ( CMSScavengeBeforeRemark 2754 && Thread::current()->is_ConcurrentGC_thread()), 2755 "Incorrect thread type for epilogue execution"); 2756 2757 if (!_between_prologue_and_epilogue) { 2758 // We have already been invoked; this is a gc_epilogue delegation 2759 // from yet another CMS generation that we are responsible for, just 2760 // ignore it since all relevant work has already been done. 2761 return; 2762 } 2763 assert(haveFreelistLocks(), "must have freelist locks"); 2764 assert_lock_strong(bitMapLock()); 2765 2766 _ct->klass_rem_set()->set_accumulate_modified_oops(false); 2767 2768 _cmsGen->gc_epilogue_work(full); 2769 2770 if (_collectorState == AbortablePreclean || _collectorState == Precleaning) { 2771 // in case sampling was not already enabled, enable it 2772 _start_sampling = true; 2773 } 2774 // reset _eden_chunk_array so sampling starts afresh 2775 _eden_chunk_index = 0; 2776 2777 size_t cms_used = _cmsGen->cmsSpace()->used(); 2778 2779 // update performance counters - this uses a special version of 2780 // update_counters() that allows the utilization to be passed as a 2781 // parameter, avoiding multiple calls to used(). 2782 // 2783 _cmsGen->update_counters(cms_used); 2784 2785 if (CMSIncrementalMode) { 2786 icms_update_allocation_limits(); 2787 } 2788 2789 bitMapLock()->unlock(); 2790 releaseFreelistLocks(); 2791 2792 if (!CleanChunkPoolAsync) { 2793 Chunk::clean_chunk_pool(); 2794 } 2795 2796 set_did_compact(false); 2797 _between_prologue_and_epilogue = false; // ready for next cycle 2798 } 2799 2800 void ConcurrentMarkSweepGeneration::gc_epilogue(bool full) { 2801 collector()->gc_epilogue(full); 2802 2803 // Also reset promotion tracking in par gc thread states. 2804 if (CollectedHeap::use_parallel_gc_threads()) { 2805 for (uint i = 0; i < ParallelGCThreads; i++) { 2806 _par_gc_thread_states[i]->promo.stopTrackingPromotions(i); 2807 } 2808 } 2809 } 2810 2811 void ConcurrentMarkSweepGeneration::gc_epilogue_work(bool full) { 2812 assert(!incremental_collection_failed(), "Should have been cleared"); 2813 cmsSpace()->setPreconsumptionDirtyCardClosure(NULL); 2814 cmsSpace()->gc_epilogue(); 2815 // Print stat counters 2816 NOT_PRODUCT( 2817 assert(_numObjectsAllocated == 0, "check"); 2818 assert(_numWordsAllocated == 0, "check"); 2819 if (Verbose && PrintGC) { 2820 gclog_or_tty->print("Promoted "SIZE_FORMAT" objects, " 2821 SIZE_FORMAT" bytes", 2822 _numObjectsPromoted, _numWordsPromoted*sizeof(HeapWord)); 2823 } 2824 _numObjectsPromoted = 0; 2825 _numWordsPromoted = 0; 2826 ) 2827 2828 if (PrintGC && Verbose) { 2829 // Call down the chain in contiguous_available needs the freelistLock 2830 // so print this out before releasing the freeListLock. 2831 gclog_or_tty->print(" Contiguous available "SIZE_FORMAT" bytes ", 2832 contiguous_available()); 2833 } 2834 } 2835 2836 #ifndef PRODUCT 2837 bool CMSCollector::have_cms_token() { 2838 Thread* thr = Thread::current(); 2839 if (thr->is_VM_thread()) { 2840 return ConcurrentMarkSweepThread::vm_thread_has_cms_token(); 2841 } else if (thr->is_ConcurrentGC_thread()) { 2842 return ConcurrentMarkSweepThread::cms_thread_has_cms_token(); 2843 } else if (thr->is_GC_task_thread()) { 2844 return ConcurrentMarkSweepThread::vm_thread_has_cms_token() && 2845 ParGCRareEvent_lock->owned_by_self(); 2846 } 2847 return false; 2848 } 2849 #endif 2850 2851 // Check reachability of the given heap address in CMS generation, 2852 // treating all other generations as roots. 2853 bool CMSCollector::is_cms_reachable(HeapWord* addr) { 2854 // We could "guarantee" below, rather than assert, but I'll 2855 // leave these as "asserts" so that an adventurous debugger 2856 // could try this in the product build provided some subset of 2857 // the conditions were met, provided they were interested in the 2858 // results and knew that the computation below wouldn't interfere 2859 // with other concurrent computations mutating the structures 2860 // being read or written. 2861 assert(SafepointSynchronize::is_at_safepoint(), 2862 "Else mutations in object graph will make answer suspect"); 2863 assert(have_cms_token(), "Should hold cms token"); 2864 assert(haveFreelistLocks(), "must hold free list locks"); 2865 assert_lock_strong(bitMapLock()); 2866 2867 // Clear the marking bit map array before starting, but, just 2868 // for kicks, first report if the given address is already marked 2869 gclog_or_tty->print_cr("Start: Address " PTR_FORMAT " is%s marked", addr, 2870 _markBitMap.isMarked(addr) ? "" : " not"); 2871 2872 if (verify_after_remark()) { 2873 MutexLockerEx x(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag); 2874 bool result = verification_mark_bm()->isMarked(addr); 2875 gclog_or_tty->print_cr("TransitiveMark: Address " PTR_FORMAT " %s marked", addr, 2876 result ? "IS" : "is NOT"); 2877 return result; 2878 } else { 2879 gclog_or_tty->print_cr("Could not compute result"); 2880 return false; 2881 } 2882 } 2883 2884 2885 void 2886 CMSCollector::print_on_error(outputStream* st) { 2887 CMSCollector* collector = ConcurrentMarkSweepGeneration::_collector; 2888 if (collector != NULL) { 2889 CMSBitMap* bitmap = &collector->_markBitMap; 2890 st->print_cr("Marking Bits: (CMSBitMap*) " PTR_FORMAT, bitmap); 2891 bitmap->print_on_error(st, " Bits: "); 2892 2893 st->cr(); 2894 2895 CMSBitMap* mut_bitmap = &collector->_modUnionTable; 2896 st->print_cr("Mod Union Table: (CMSBitMap*) " PTR_FORMAT, mut_bitmap); 2897 mut_bitmap->print_on_error(st, " Bits: "); 2898 } 2899 } 2900 2901 //////////////////////////////////////////////////////// 2902 // CMS Verification Support 2903 //////////////////////////////////////////////////////// 2904 // Following the remark phase, the following invariant 2905 // should hold -- each object in the CMS heap which is 2906 // marked in markBitMap() should be marked in the verification_mark_bm(). 2907 2908 class VerifyMarkedClosure: public BitMapClosure { 2909 CMSBitMap* _marks; 2910 bool _failed; 2911 2912 public: 2913 VerifyMarkedClosure(CMSBitMap* bm): _marks(bm), _failed(false) {} 2914 2915 bool do_bit(size_t offset) { 2916 HeapWord* addr = _marks->offsetToHeapWord(offset); 2917 if (!_marks->isMarked(addr)) { 2918 oop(addr)->print_on(gclog_or_tty); 2919 gclog_or_tty->print_cr(" ("INTPTR_FORMAT" should have been marked)", addr); 2920 _failed = true; 2921 } 2922 return true; 2923 } 2924 2925 bool failed() { return _failed; } 2926 }; 2927 2928 bool CMSCollector::verify_after_remark(bool silent) { 2929 if (!silent) gclog_or_tty->print(" [Verifying CMS Marking... "); 2930 MutexLockerEx ml(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag); 2931 static bool init = false; 2932 2933 assert(SafepointSynchronize::is_at_safepoint(), 2934 "Else mutations in object graph will make answer suspect"); 2935 assert(have_cms_token(), 2936 "Else there may be mutual interference in use of " 2937 " verification data structures"); 2938 assert(_collectorState > Marking && _collectorState <= Sweeping, 2939 "Else marking info checked here may be obsolete"); 2940 assert(haveFreelistLocks(), "must hold free list locks"); 2941 assert_lock_strong(bitMapLock()); 2942 2943 2944 // Allocate marking bit map if not already allocated 2945 if (!init) { // first time 2946 if (!verification_mark_bm()->allocate(_span)) { 2947 return false; 2948 } 2949 init = true; 2950 } 2951 2952 assert(verification_mark_stack()->isEmpty(), "Should be empty"); 2953 2954 // Turn off refs discovery -- so we will be tracing through refs. 2955 // This is as intended, because by this time 2956 // GC must already have cleared any refs that need to be cleared, 2957 // and traced those that need to be marked; moreover, 2958 // the marking done here is not going to interfere in any 2959 // way with the marking information used by GC. 2960 NoRefDiscovery no_discovery(ref_processor()); 2961 2962 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;) 2963 2964 // Clear any marks from a previous round 2965 verification_mark_bm()->clear_all(); 2966 assert(verification_mark_stack()->isEmpty(), "markStack should be empty"); 2967 verify_work_stacks_empty(); 2968 2969 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2970 gch->ensure_parsability(false); // fill TLABs, but no need to retire them 2971 // Update the saved marks which may affect the root scans. 2972 gch->save_marks(); 2973 2974 if (CMSRemarkVerifyVariant == 1) { 2975 // In this first variant of verification, we complete 2976 // all marking, then check if the new marks-vector is 2977 // a subset of the CMS marks-vector. 2978 verify_after_remark_work_1(); 2979 } else if (CMSRemarkVerifyVariant == 2) { 2980 // In this second variant of verification, we flag an error 2981 // (i.e. an object reachable in the new marks-vector not reachable 2982 // in the CMS marks-vector) immediately, also indicating the 2983 // identify of an object (A) that references the unmarked object (B) -- 2984 // presumably, a mutation to A failed to be picked up by preclean/remark? 2985 verify_after_remark_work_2(); 2986 } else { 2987 warning("Unrecognized value %d for CMSRemarkVerifyVariant", 2988 CMSRemarkVerifyVariant); 2989 } 2990 if (!silent) gclog_or_tty->print(" done] "); 2991 return true; 2992 } 2993 2994 void CMSCollector::verify_after_remark_work_1() { 2995 ResourceMark rm; 2996 HandleMark hm; 2997 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2998 2999 // Get a clear set of claim bits for the roots processing to work with. 3000 ClassLoaderDataGraph::clear_claimed_marks(); 3001 3002 // Mark from roots one level into CMS 3003 MarkRefsIntoClosure notOlder(_span, verification_mark_bm()); 3004 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 3005 3006 gch->gen_process_roots(_cmsGen->level(), 3007 true, // younger gens are roots 3008 true, // activate StrongRootsScope 3009 SharedHeap::ScanningOption(roots_scanning_options()), 3010 should_unload_classes(), 3011 ¬Older, 3012 NULL, 3013 NULL); // SSS: Provide correct closure 3014 3015 // Now mark from the roots 3016 MarkFromRootsClosure markFromRootsClosure(this, _span, 3017 verification_mark_bm(), verification_mark_stack(), 3018 false /* don't yield */, true /* verifying */); 3019 assert(_restart_addr == NULL, "Expected pre-condition"); 3020 verification_mark_bm()->iterate(&markFromRootsClosure); 3021 while (_restart_addr != NULL) { 3022 // Deal with stack overflow: by restarting at the indicated 3023 // address. 3024 HeapWord* ra = _restart_addr; 3025 markFromRootsClosure.reset(ra); 3026 _restart_addr = NULL; 3027 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end()); 3028 } 3029 assert(verification_mark_stack()->isEmpty(), "Should have been drained"); 3030 verify_work_stacks_empty(); 3031 3032 // Marking completed -- now verify that each bit marked in 3033 // verification_mark_bm() is also marked in markBitMap(); flag all 3034 // errors by printing corresponding objects. 3035 VerifyMarkedClosure vcl(markBitMap()); 3036 verification_mark_bm()->iterate(&vcl); 3037 if (vcl.failed()) { 3038 gclog_or_tty->print("Verification failed"); 3039 Universe::heap()->print_on(gclog_or_tty); 3040 fatal("CMS: failed marking verification after remark"); 3041 } 3042 } 3043 3044 class VerifyKlassOopsKlassClosure : public KlassClosure { 3045 class VerifyKlassOopsClosure : public OopClosure { 3046 CMSBitMap* _bitmap; 3047 public: 3048 VerifyKlassOopsClosure(CMSBitMap* bitmap) : _bitmap(bitmap) { } 3049 void do_oop(oop* p) { guarantee(*p == NULL || _bitmap->isMarked((HeapWord*) *p), "Should be marked"); } 3050 void do_oop(narrowOop* p) { ShouldNotReachHere(); } 3051 } _oop_closure; 3052 public: 3053 VerifyKlassOopsKlassClosure(CMSBitMap* bitmap) : _oop_closure(bitmap) {} 3054 void do_klass(Klass* k) { 3055 k->oops_do(&_oop_closure); 3056 } 3057 }; 3058 3059 void CMSCollector::verify_after_remark_work_2() { 3060 ResourceMark rm; 3061 HandleMark hm; 3062 GenCollectedHeap* gch = GenCollectedHeap::heap(); 3063 3064 // Get a clear set of claim bits for the roots processing to work with. 3065 ClassLoaderDataGraph::clear_claimed_marks(); 3066 3067 // Mark from roots one level into CMS 3068 MarkRefsIntoVerifyClosure notOlder(_span, verification_mark_bm(), 3069 markBitMap()); 3070 CLDToOopClosure cld_closure(¬Older, true); 3071 3072 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 3073 3074 gch->gen_process_roots(_cmsGen->level(), 3075 true, // younger gens are roots 3076 true, // activate StrongRootsScope 3077 SharedHeap::ScanningOption(roots_scanning_options()), 3078 should_unload_classes(), 3079 ¬Older, 3080 NULL, 3081 &cld_closure); 3082 3083 // Now mark from the roots 3084 MarkFromRootsVerifyClosure markFromRootsClosure(this, _span, 3085 verification_mark_bm(), markBitMap(), verification_mark_stack()); 3086 assert(_restart_addr == NULL, "Expected pre-condition"); 3087 verification_mark_bm()->iterate(&markFromRootsClosure); 3088 while (_restart_addr != NULL) { 3089 // Deal with stack overflow: by restarting at the indicated 3090 // address. 3091 HeapWord* ra = _restart_addr; 3092 markFromRootsClosure.reset(ra); 3093 _restart_addr = NULL; 3094 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end()); 3095 } 3096 assert(verification_mark_stack()->isEmpty(), "Should have been drained"); 3097 verify_work_stacks_empty(); 3098 3099 VerifyKlassOopsKlassClosure verify_klass_oops(verification_mark_bm()); 3100 ClassLoaderDataGraph::classes_do(&verify_klass_oops); 3101 3102 // Marking completed -- now verify that each bit marked in 3103 // verification_mark_bm() is also marked in markBitMap(); flag all 3104 // errors by printing corresponding objects. 3105 VerifyMarkedClosure vcl(markBitMap()); 3106 verification_mark_bm()->iterate(&vcl); 3107 assert(!vcl.failed(), "Else verification above should not have succeeded"); 3108 } 3109 3110 void ConcurrentMarkSweepGeneration::save_marks() { 3111 // delegate to CMS space 3112 cmsSpace()->save_marks(); 3113 for (uint i = 0; i < ParallelGCThreads; i++) { 3114 _par_gc_thread_states[i]->promo.startTrackingPromotions(); 3115 } 3116 } 3117 3118 bool ConcurrentMarkSweepGeneration::no_allocs_since_save_marks() { 3119 return cmsSpace()->no_allocs_since_save_marks(); 3120 } 3121 3122 #define CMS_SINCE_SAVE_MARKS_DEFN(OopClosureType, nv_suffix) \ 3123 \ 3124 void ConcurrentMarkSweepGeneration:: \ 3125 oop_since_save_marks_iterate##nv_suffix(OopClosureType* cl) { \ 3126 cl->set_generation(this); \ 3127 cmsSpace()->oop_since_save_marks_iterate##nv_suffix(cl); \ 3128 cl->reset_generation(); \ 3129 save_marks(); \ 3130 } 3131 3132 ALL_SINCE_SAVE_MARKS_CLOSURES(CMS_SINCE_SAVE_MARKS_DEFN) 3133 3134 void 3135 ConcurrentMarkSweepGeneration::younger_refs_iterate(OopsInGenClosure* cl) { 3136 cl->set_generation(this); 3137 younger_refs_in_space_iterate(_cmsSpace, cl); 3138 cl->reset_generation(); 3139 } 3140 3141 void 3142 ConcurrentMarkSweepGeneration::oop_iterate(ExtendedOopClosure* cl) { 3143 if (freelistLock()->owned_by_self()) { 3144 Generation::oop_iterate(cl); 3145 } else { 3146 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 3147 Generation::oop_iterate(cl); 3148 } 3149 } 3150 3151 void 3152 ConcurrentMarkSweepGeneration::object_iterate(ObjectClosure* cl) { 3153 if (freelistLock()->owned_by_self()) { 3154 Generation::object_iterate(cl); 3155 } else { 3156 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 3157 Generation::object_iterate(cl); 3158 } 3159 } 3160 3161 void 3162 ConcurrentMarkSweepGeneration::safe_object_iterate(ObjectClosure* cl) { 3163 if (freelistLock()->owned_by_self()) { 3164 Generation::safe_object_iterate(cl); 3165 } else { 3166 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 3167 Generation::safe_object_iterate(cl); 3168 } 3169 } 3170 3171 void 3172 ConcurrentMarkSweepGeneration::post_compact() { 3173 } 3174 3175 void 3176 ConcurrentMarkSweepGeneration::prepare_for_verify() { 3177 // Fix the linear allocation blocks to look like free blocks. 3178 3179 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those 3180 // are not called when the heap is verified during universe initialization and 3181 // at vm shutdown. 3182 if (freelistLock()->owned_by_self()) { 3183 cmsSpace()->prepare_for_verify(); 3184 } else { 3185 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag); 3186 cmsSpace()->prepare_for_verify(); 3187 } 3188 } 3189 3190 void 3191 ConcurrentMarkSweepGeneration::verify() { 3192 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those 3193 // are not called when the heap is verified during universe initialization and 3194 // at vm shutdown. 3195 if (freelistLock()->owned_by_self()) { 3196 cmsSpace()->verify(); 3197 } else { 3198 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag); 3199 cmsSpace()->verify(); 3200 } 3201 } 3202 3203 void CMSCollector::verify() { 3204 _cmsGen->verify(); 3205 } 3206 3207 #ifndef PRODUCT 3208 bool CMSCollector::overflow_list_is_empty() const { 3209 assert(_num_par_pushes >= 0, "Inconsistency"); 3210 if (_overflow_list == NULL) { 3211 assert(_num_par_pushes == 0, "Inconsistency"); 3212 } 3213 return _overflow_list == NULL; 3214 } 3215 3216 // The methods verify_work_stacks_empty() and verify_overflow_empty() 3217 // merely consolidate assertion checks that appear to occur together frequently. 3218 void CMSCollector::verify_work_stacks_empty() const { 3219 assert(_markStack.isEmpty(), "Marking stack should be empty"); 3220 assert(overflow_list_is_empty(), "Overflow list should be empty"); 3221 } 3222 3223 void CMSCollector::verify_overflow_empty() const { 3224 assert(overflow_list_is_empty(), "Overflow list should be empty"); 3225 assert(no_preserved_marks(), "No preserved marks"); 3226 } 3227 #endif // PRODUCT 3228 3229 // Decide if we want to enable class unloading as part of the 3230 // ensuing concurrent GC cycle. We will collect and 3231 // unload classes if it's the case that: 3232 // (1) an explicit gc request has been made and the flag 3233 // ExplicitGCInvokesConcurrentAndUnloadsClasses is set, OR 3234 // (2) (a) class unloading is enabled at the command line, and 3235 // (b) old gen is getting really full 3236 // NOTE: Provided there is no change in the state of the heap between 3237 // calls to this method, it should have idempotent results. Moreover, 3238 // its results should be monotonically increasing (i.e. going from 0 to 1, 3239 // but not 1 to 0) between successive calls between which the heap was 3240 // not collected. For the implementation below, it must thus rely on 3241 // the property that concurrent_cycles_since_last_unload() 3242 // will not decrease unless a collection cycle happened and that 3243 // _cmsGen->is_too_full() are 3244 // themselves also monotonic in that sense. See check_monotonicity() 3245 // below. 3246 void CMSCollector::update_should_unload_classes() { 3247 _should_unload_classes = false; 3248 // Condition 1 above 3249 if (_full_gc_requested && ExplicitGCInvokesConcurrentAndUnloadsClasses) { 3250 _should_unload_classes = true; 3251 } else if (CMSClassUnloadingEnabled) { // Condition 2.a above 3252 // Disjuncts 2.b.(i,ii,iii) above 3253 _should_unload_classes = (concurrent_cycles_since_last_unload() >= 3254 CMSClassUnloadingMaxInterval) 3255 || _cmsGen->is_too_full(); 3256 } 3257 } 3258 3259 bool ConcurrentMarkSweepGeneration::is_too_full() const { 3260 bool res = should_concurrent_collect(); 3261 res = res && (occupancy() > (double)CMSIsTooFullPercentage/100.0); 3262 return res; 3263 } 3264 3265 void CMSCollector::setup_cms_unloading_and_verification_state() { 3266 const bool should_verify = VerifyBeforeGC || VerifyAfterGC || VerifyDuringGC 3267 || VerifyBeforeExit; 3268 const int rso = SharedHeap::SO_AllCodeCache; 3269 3270 // We set the proper root for this CMS cycle here. 3271 if (should_unload_classes()) { // Should unload classes this cycle 3272 remove_root_scanning_option(rso); // Shrink the root set appropriately 3273 set_verifying(should_verify); // Set verification state for this cycle 3274 return; // Nothing else needs to be done at this time 3275 } 3276 3277 // Not unloading classes this cycle 3278 assert(!should_unload_classes(), "Inconsistency!"); 3279 3280 if ((!verifying() || unloaded_classes_last_cycle()) && should_verify) { 3281 // Include symbols, strings and code cache elements to prevent their resurrection. 3282 add_root_scanning_option(rso); 3283 set_verifying(true); 3284 } else if (verifying() && !should_verify) { 3285 // We were verifying, but some verification flags got disabled. 3286 set_verifying(false); 3287 // Exclude symbols, strings and code cache elements from root scanning to 3288 // reduce IM and RM pauses. 3289 remove_root_scanning_option(rso); 3290 } 3291 } 3292 3293 3294 #ifndef PRODUCT 3295 HeapWord* CMSCollector::block_start(const void* p) const { 3296 const HeapWord* addr = (HeapWord*)p; 3297 if (_span.contains(p)) { 3298 if (_cmsGen->cmsSpace()->is_in_reserved(addr)) { 3299 return _cmsGen->cmsSpace()->block_start(p); 3300 } 3301 } 3302 return NULL; 3303 } 3304 #endif 3305 3306 HeapWord* 3307 ConcurrentMarkSweepGeneration::expand_and_allocate(size_t word_size, 3308 bool tlab, 3309 bool parallel) { 3310 CMSSynchronousYieldRequest yr; 3311 assert(!tlab, "Can't deal with TLAB allocation"); 3312 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 3313 expand(word_size*HeapWordSize, MinHeapDeltaBytes, 3314 CMSExpansionCause::_satisfy_allocation); 3315 if (GCExpandToAllocateDelayMillis > 0) { 3316 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 3317 } 3318 return have_lock_and_allocate(word_size, tlab); 3319 } 3320 3321 // YSR: All of this generation expansion/shrinking stuff is an exact copy of 3322 // OneContigSpaceCardGeneration, which makes me wonder if we should move this 3323 // to CardGeneration and share it... 3324 bool ConcurrentMarkSweepGeneration::expand(size_t bytes, size_t expand_bytes) { 3325 return CardGeneration::expand(bytes, expand_bytes); 3326 } 3327 3328 void ConcurrentMarkSweepGeneration::expand(size_t bytes, size_t expand_bytes, 3329 CMSExpansionCause::Cause cause) 3330 { 3331 3332 bool success = expand(bytes, expand_bytes); 3333 3334 // remember why we expanded; this information is used 3335 // by shouldConcurrentCollect() when making decisions on whether to start 3336 // a new CMS cycle. 3337 if (success) { 3338 set_expansion_cause(cause); 3339 if (PrintGCDetails && Verbose) { 3340 gclog_or_tty->print_cr("Expanded CMS gen for %s", 3341 CMSExpansionCause::to_string(cause)); 3342 } 3343 } 3344 } 3345 3346 HeapWord* ConcurrentMarkSweepGeneration::expand_and_par_lab_allocate(CMSParGCThreadState* ps, size_t word_sz) { 3347 HeapWord* res = NULL; 3348 MutexLocker x(ParGCRareEvent_lock); 3349 while (true) { 3350 // Expansion by some other thread might make alloc OK now: 3351 res = ps->lab.alloc(word_sz); 3352 if (res != NULL) return res; 3353 // If there's not enough expansion space available, give up. 3354 if (_virtual_space.uncommitted_size() < (word_sz * HeapWordSize)) { 3355 return NULL; 3356 } 3357 // Otherwise, we try expansion. 3358 expand(word_sz*HeapWordSize, MinHeapDeltaBytes, 3359 CMSExpansionCause::_allocate_par_lab); 3360 // Now go around the loop and try alloc again; 3361 // A competing par_promote might beat us to the expansion space, 3362 // so we may go around the loop again if promotion fails again. 3363 if (GCExpandToAllocateDelayMillis > 0) { 3364 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 3365 } 3366 } 3367 } 3368 3369 3370 bool ConcurrentMarkSweepGeneration::expand_and_ensure_spooling_space( 3371 PromotionInfo* promo) { 3372 MutexLocker x(ParGCRareEvent_lock); 3373 size_t refill_size_bytes = promo->refillSize() * HeapWordSize; 3374 while (true) { 3375 // Expansion by some other thread might make alloc OK now: 3376 if (promo->ensure_spooling_space()) { 3377 assert(promo->has_spooling_space(), 3378 "Post-condition of successful ensure_spooling_space()"); 3379 return true; 3380 } 3381 // If there's not enough expansion space available, give up. 3382 if (_virtual_space.uncommitted_size() < refill_size_bytes) { 3383 return false; 3384 } 3385 // Otherwise, we try expansion. 3386 expand(refill_size_bytes, MinHeapDeltaBytes, 3387 CMSExpansionCause::_allocate_par_spooling_space); 3388 // Now go around the loop and try alloc again; 3389 // A competing allocation might beat us to the expansion space, 3390 // so we may go around the loop again if allocation fails again. 3391 if (GCExpandToAllocateDelayMillis > 0) { 3392 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 3393 } 3394 } 3395 } 3396 3397 3398 void ConcurrentMarkSweepGeneration::shrink_by(size_t bytes) { 3399 assert_locked_or_safepoint(ExpandHeap_lock); 3400 // Shrink committed space 3401 _virtual_space.shrink_by(bytes); 3402 // Shrink space; this also shrinks the space's BOT 3403 _cmsSpace->set_end((HeapWord*) _virtual_space.high()); 3404 size_t new_word_size = heap_word_size(_cmsSpace->capacity()); 3405 // Shrink the shared block offset array 3406 _bts->resize(new_word_size); 3407 MemRegion mr(_cmsSpace->bottom(), new_word_size); 3408 // Shrink the card table 3409 Universe::heap()->barrier_set()->resize_covered_region(mr); 3410 3411 if (Verbose && PrintGC) { 3412 size_t new_mem_size = _virtual_space.committed_size(); 3413 size_t old_mem_size = new_mem_size + bytes; 3414 gclog_or_tty->print_cr("Shrinking %s from " SIZE_FORMAT "K to " SIZE_FORMAT "K", 3415 name(), old_mem_size/K, new_mem_size/K); 3416 } 3417 } 3418 3419 void ConcurrentMarkSweepGeneration::shrink(size_t bytes) { 3420 assert_locked_or_safepoint(Heap_lock); 3421 size_t size = ReservedSpace::page_align_size_down(bytes); 3422 // Only shrink if a compaction was done so that all the free space 3423 // in the generation is in a contiguous block at the end. 3424 if (size > 0 && did_compact()) { 3425 shrink_by(size); 3426 } 3427 } 3428 3429 bool ConcurrentMarkSweepGeneration::grow_by(size_t bytes) { 3430 assert_locked_or_safepoint(Heap_lock); 3431 bool result = _virtual_space.expand_by(bytes); 3432 if (result) { 3433 size_t new_word_size = 3434 heap_word_size(_virtual_space.committed_size()); 3435 MemRegion mr(_cmsSpace->bottom(), new_word_size); 3436 _bts->resize(new_word_size); // resize the block offset shared array 3437 Universe::heap()->barrier_set()->resize_covered_region(mr); 3438 // Hmmmm... why doesn't CFLS::set_end verify locking? 3439 // This is quite ugly; FIX ME XXX 3440 _cmsSpace->assert_locked(freelistLock()); 3441 _cmsSpace->set_end((HeapWord*)_virtual_space.high()); 3442 3443 // update the space and generation capacity counters 3444 if (UsePerfData) { 3445 _space_counters->update_capacity(); 3446 _gen_counters->update_all(); 3447 } 3448 3449 if (Verbose && PrintGC) { 3450 size_t new_mem_size = _virtual_space.committed_size(); 3451 size_t old_mem_size = new_mem_size - bytes; 3452 gclog_or_tty->print_cr("Expanding %s from " SIZE_FORMAT "K by " SIZE_FORMAT "K to " SIZE_FORMAT "K", 3453 name(), old_mem_size/K, bytes/K, new_mem_size/K); 3454 } 3455 } 3456 return result; 3457 } 3458 3459 bool ConcurrentMarkSweepGeneration::grow_to_reserved() { 3460 assert_locked_or_safepoint(Heap_lock); 3461 bool success = true; 3462 const size_t remaining_bytes = _virtual_space.uncommitted_size(); 3463 if (remaining_bytes > 0) { 3464 success = grow_by(remaining_bytes); 3465 DEBUG_ONLY(if (!success) warning("grow to reserved failed");) 3466 } 3467 return success; 3468 } 3469 3470 void ConcurrentMarkSweepGeneration::shrink_free_list_by(size_t bytes) { 3471 assert_locked_or_safepoint(Heap_lock); 3472 assert_lock_strong(freelistLock()); 3473 if (PrintGCDetails && Verbose) { 3474 warning("Shrinking of CMS not yet implemented"); 3475 } 3476 return; 3477 } 3478 3479 3480 // Simple ctor/dtor wrapper for accounting & timer chores around concurrent 3481 // phases. 3482 class CMSPhaseAccounting: public StackObj { 3483 public: 3484 CMSPhaseAccounting(CMSCollector *collector, 3485 const char *phase, 3486 const GCId gc_id, 3487 bool print_cr = true); 3488 ~CMSPhaseAccounting(); 3489 3490 private: 3491 CMSCollector *_collector; 3492 const char *_phase; 3493 elapsedTimer _wallclock; 3494 bool _print_cr; 3495 const GCId _gc_id; 3496 3497 public: 3498 // Not MT-safe; so do not pass around these StackObj's 3499 // where they may be accessed by other threads. 3500 jlong wallclock_millis() { 3501 assert(_wallclock.is_active(), "Wall clock should not stop"); 3502 _wallclock.stop(); // to record time 3503 jlong ret = _wallclock.milliseconds(); 3504 _wallclock.start(); // restart 3505 return ret; 3506 } 3507 }; 3508 3509 CMSPhaseAccounting::CMSPhaseAccounting(CMSCollector *collector, 3510 const char *phase, 3511 const GCId gc_id, 3512 bool print_cr) : 3513 _collector(collector), _phase(phase), _print_cr(print_cr), _gc_id(gc_id) { 3514 3515 if (PrintCMSStatistics != 0) { 3516 _collector->resetYields(); 3517 } 3518 if (PrintGCDetails) { 3519 gclog_or_tty->gclog_stamp(_gc_id); 3520 gclog_or_tty->print_cr("[%s-concurrent-%s-start]", 3521 _collector->cmsGen()->short_name(), _phase); 3522 } 3523 _collector->resetTimer(); 3524 _wallclock.start(); 3525 _collector->startTimer(); 3526 } 3527 3528 CMSPhaseAccounting::~CMSPhaseAccounting() { 3529 assert(_wallclock.is_active(), "Wall clock should not have stopped"); 3530 _collector->stopTimer(); 3531 _wallclock.stop(); 3532 if (PrintGCDetails) { 3533 gclog_or_tty->gclog_stamp(_gc_id); 3534 gclog_or_tty->print("[%s-concurrent-%s: %3.3f/%3.3f secs]", 3535 _collector->cmsGen()->short_name(), 3536 _phase, _collector->timerValue(), _wallclock.seconds()); 3537 if (_print_cr) { 3538 gclog_or_tty->cr(); 3539 } 3540 if (PrintCMSStatistics != 0) { 3541 gclog_or_tty->print_cr(" (CMS-concurrent-%s yielded %d times)", _phase, 3542 _collector->yields()); 3543 } 3544 } 3545 } 3546 3547 // CMS work 3548 3549 // The common parts of CMSParInitialMarkTask and CMSParRemarkTask. 3550 class CMSParMarkTask : public AbstractGangTask { 3551 protected: 3552 CMSCollector* _collector; 3553 int _n_workers; 3554 CMSParMarkTask(const char* name, CMSCollector* collector, int n_workers) : 3555 AbstractGangTask(name), 3556 _collector(collector), 3557 _n_workers(n_workers) {} 3558 // Work method in support of parallel rescan ... of young gen spaces 3559 void do_young_space_rescan(uint worker_id, OopsInGenClosure* cl, 3560 ContiguousSpace* space, 3561 HeapWord** chunk_array, size_t chunk_top); 3562 void work_on_young_gen_roots(uint worker_id, OopsInGenClosure* cl); 3563 }; 3564 3565 // Parallel initial mark task 3566 class CMSParInitialMarkTask: public CMSParMarkTask { 3567 public: 3568 CMSParInitialMarkTask(CMSCollector* collector, int n_workers) : 3569 CMSParMarkTask("Scan roots and young gen for initial mark in parallel", 3570 collector, n_workers) {} 3571 void work(uint worker_id); 3572 }; 3573 3574 // Checkpoint the roots into this generation from outside 3575 // this generation. [Note this initial checkpoint need only 3576 // be approximate -- we'll do a catch up phase subsequently.] 3577 void CMSCollector::checkpointRootsInitial(bool asynch) { 3578 assert(_collectorState == InitialMarking, "Wrong collector state"); 3579 check_correct_thread_executing(); 3580 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 3581 3582 save_heap_summary(); 3583 report_heap_summary(GCWhen::BeforeGC); 3584 3585 ReferenceProcessor* rp = ref_processor(); 3586 SpecializationStats::clear(); 3587 assert(_restart_addr == NULL, "Control point invariant"); 3588 if (asynch) { 3589 // acquire locks for subsequent manipulations 3590 MutexLockerEx x(bitMapLock(), 3591 Mutex::_no_safepoint_check_flag); 3592 checkpointRootsInitialWork(asynch); 3593 // enable ("weak") refs discovery 3594 rp->enable_discovery(true /*verify_disabled*/, true /*check_no_refs*/); 3595 _collectorState = Marking; 3596 } else { 3597 // (Weak) Refs discovery: this is controlled from genCollectedHeap::do_collection 3598 // which recognizes if we are a CMS generation, and doesn't try to turn on 3599 // discovery; verify that they aren't meddling. 3600 assert(!rp->discovery_is_atomic(), 3601 "incorrect setting of discovery predicate"); 3602 assert(!rp->discovery_enabled(), "genCollectedHeap shouldn't control " 3603 "ref discovery for this generation kind"); 3604 // already have locks 3605 checkpointRootsInitialWork(asynch); 3606 // now enable ("weak") refs discovery 3607 rp->enable_discovery(true /*verify_disabled*/, false /*verify_no_refs*/); 3608 _collectorState = Marking; 3609 } 3610 SpecializationStats::print(); 3611 } 3612 3613 void CMSCollector::checkpointRootsInitialWork(bool asynch) { 3614 assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped"); 3615 assert(_collectorState == InitialMarking, "just checking"); 3616 3617 // If there has not been a GC[n-1] since last GC[n] cycle completed, 3618 // precede our marking with a collection of all 3619 // younger generations to keep floating garbage to a minimum. 3620 // XXX: we won't do this for now -- it's an optimization to be done later. 3621 3622 // already have locks 3623 assert_lock_strong(bitMapLock()); 3624 assert(_markBitMap.isAllClear(), "was reset at end of previous cycle"); 3625 3626 // Setup the verification and class unloading state for this 3627 // CMS collection cycle. 3628 setup_cms_unloading_and_verification_state(); 3629 3630 NOT_PRODUCT(GCTraceTime t("\ncheckpointRootsInitialWork", 3631 PrintGCDetails && Verbose, true, _gc_timer_cm, _gc_tracer_cm->gc_id());) 3632 3633 // Reset all the PLAB chunk arrays if necessary. 3634 if (_survivor_plab_array != NULL && !CMSPLABRecordAlways) { 3635 reset_survivor_plab_arrays(); 3636 } 3637 3638 ResourceMark rm; 3639 HandleMark hm; 3640 3641 MarkRefsIntoClosure notOlder(_span, &_markBitMap); 3642 GenCollectedHeap* gch = GenCollectedHeap::heap(); 3643 3644 verify_work_stacks_empty(); 3645 verify_overflow_empty(); 3646 3647 gch->ensure_parsability(false); // fill TLABs, but no need to retire them 3648 // Update the saved marks which may affect the root scans. 3649 gch->save_marks(); 3650 3651 // weak reference processing has not started yet. 3652 ref_processor()->set_enqueuing_is_done(false); 3653 3654 // Need to remember all newly created CLDs, 3655 // so that we can guarantee that the remark finds them. 3656 ClassLoaderDataGraph::remember_new_clds(true); 3657 3658 // Whenever a CLD is found, it will be claimed before proceeding to mark 3659 // the klasses. The claimed marks need to be cleared before marking starts. 3660 ClassLoaderDataGraph::clear_claimed_marks(); 3661 3662 if (CMSPrintEdenSurvivorChunks) { 3663 print_eden_and_survivor_chunk_arrays(); 3664 } 3665 3666 { 3667 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;) 3668 if (CMSParallelInitialMarkEnabled && CollectedHeap::use_parallel_gc_threads()) { 3669 // The parallel version. 3670 FlexibleWorkGang* workers = gch->workers(); 3671 assert(workers != NULL, "Need parallel worker threads."); 3672 int n_workers = workers->active_workers(); 3673 CMSParInitialMarkTask tsk(this, n_workers); 3674 gch->set_par_threads(n_workers); 3675 initialize_sequential_subtasks_for_young_gen_rescan(n_workers); 3676 if (n_workers > 1) { 3677 GenCollectedHeap::StrongRootsScope srs(gch); 3678 workers->run_task(&tsk); 3679 } else { 3680 GenCollectedHeap::StrongRootsScope srs(gch); 3681 tsk.work(0); 3682 } 3683 gch->set_par_threads(0); 3684 } else { 3685 // The serial version. 3686 CLDToOopClosure cld_closure(¬Older, true); 3687 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 3688 gch->gen_process_roots(_cmsGen->level(), 3689 true, // younger gens are roots 3690 true, // activate StrongRootsScope 3691 SharedHeap::ScanningOption(roots_scanning_options()), 3692 should_unload_classes(), 3693 ¬Older, 3694 NULL, 3695 &cld_closure); 3696 } 3697 } 3698 3699 // Clear mod-union table; it will be dirtied in the prologue of 3700 // CMS generation per each younger generation collection. 3701 3702 assert(_modUnionTable.isAllClear(), 3703 "Was cleared in most recent final checkpoint phase" 3704 " or no bits are set in the gc_prologue before the start of the next " 3705 "subsequent marking phase."); 3706 3707 assert(_ct->klass_rem_set()->mod_union_is_clear(), "Must be"); 3708 3709 // Save the end of the used_region of the constituent generations 3710 // to be used to limit the extent of sweep in each generation. 3711 save_sweep_limits(); 3712 verify_overflow_empty(); 3713 } 3714 3715 bool CMSCollector::markFromRoots(bool asynch) { 3716 // we might be tempted to assert that: 3717 // assert(asynch == !SafepointSynchronize::is_at_safepoint(), 3718 // "inconsistent argument?"); 3719 // However that wouldn't be right, because it's possible that 3720 // a safepoint is indeed in progress as a younger generation 3721 // stop-the-world GC happens even as we mark in this generation. 3722 assert(_collectorState == Marking, "inconsistent state?"); 3723 check_correct_thread_executing(); 3724 verify_overflow_empty(); 3725 3726 bool res; 3727 if (asynch) { 3728 // Weak ref discovery note: We may be discovering weak 3729 // refs in this generation concurrent (but interleaved) with 3730 // weak ref discovery by a younger generation collector. 3731 3732 CMSTokenSyncWithLocks ts(true, bitMapLock()); 3733 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 3734 CMSPhaseAccounting pa(this, "mark", _gc_tracer_cm->gc_id(), !PrintGCDetails); 3735 res = markFromRootsWork(asynch); 3736 if (res) { 3737 _collectorState = Precleaning; 3738 } else { // We failed and a foreground collection wants to take over 3739 assert(_foregroundGCIsActive, "internal state inconsistency"); 3740 assert(_restart_addr == NULL, "foreground will restart from scratch"); 3741 if (PrintGCDetails) { 3742 gclog_or_tty->print_cr("bailing out to foreground collection"); 3743 } 3744 } 3745 } else { 3746 assert(SafepointSynchronize::is_at_safepoint(), 3747 "inconsistent with asynch == false"); 3748 // already have locks 3749 res = markFromRootsWork(asynch); 3750 _collectorState = FinalMarking; 3751 } 3752 verify_overflow_empty(); 3753 return res; 3754 } 3755 3756 bool CMSCollector::markFromRootsWork(bool asynch) { 3757 // iterate over marked bits in bit map, doing a full scan and mark 3758 // from these roots using the following algorithm: 3759 // . if oop is to the right of the current scan pointer, 3760 // mark corresponding bit (we'll process it later) 3761 // . else (oop is to left of current scan pointer) 3762 // push oop on marking stack 3763 // . drain the marking stack 3764 3765 // Note that when we do a marking step we need to hold the 3766 // bit map lock -- recall that direct allocation (by mutators) 3767 // and promotion (by younger generation collectors) is also 3768 // marking the bit map. [the so-called allocate live policy.] 3769 // Because the implementation of bit map marking is not 3770 // robust wrt simultaneous marking of bits in the same word, 3771 // we need to make sure that there is no such interference 3772 // between concurrent such updates. 3773 3774 // already have locks 3775 assert_lock_strong(bitMapLock()); 3776 3777 verify_work_stacks_empty(); 3778 verify_overflow_empty(); 3779 bool result = false; 3780 if (CMSConcurrentMTEnabled && ConcGCThreads > 0) { 3781 result = do_marking_mt(asynch); 3782 } else { 3783 result = do_marking_st(asynch); 3784 } 3785 return result; 3786 } 3787 3788 // Forward decl 3789 class CMSConcMarkingTask; 3790 3791 class CMSConcMarkingTerminator: public ParallelTaskTerminator { 3792 CMSCollector* _collector; 3793 CMSConcMarkingTask* _task; 3794 public: 3795 virtual void yield(); 3796 3797 // "n_threads" is the number of threads to be terminated. 3798 // "queue_set" is a set of work queues of other threads. 3799 // "collector" is the CMS collector associated with this task terminator. 3800 // "yield" indicates whether we need the gang as a whole to yield. 3801 CMSConcMarkingTerminator(int n_threads, TaskQueueSetSuper* queue_set, CMSCollector* collector) : 3802 ParallelTaskTerminator(n_threads, queue_set), 3803 _collector(collector) { } 3804 3805 void set_task(CMSConcMarkingTask* task) { 3806 _task = task; 3807 } 3808 }; 3809 3810 class CMSConcMarkingTerminatorTerminator: public TerminatorTerminator { 3811 CMSConcMarkingTask* _task; 3812 public: 3813 bool should_exit_termination(); 3814 void set_task(CMSConcMarkingTask* task) { 3815 _task = task; 3816 } 3817 }; 3818 3819 // MT Concurrent Marking Task 3820 class CMSConcMarkingTask: public YieldingFlexibleGangTask { 3821 CMSCollector* _collector; 3822 int _n_workers; // requested/desired # workers 3823 bool _asynch; 3824 bool _result; 3825 CompactibleFreeListSpace* _cms_space; 3826 char _pad_front[64]; // padding to ... 3827 HeapWord* _global_finger; // ... avoid sharing cache line 3828 char _pad_back[64]; 3829 HeapWord* _restart_addr; 3830 3831 // Exposed here for yielding support 3832 Mutex* const _bit_map_lock; 3833 3834 // The per thread work queues, available here for stealing 3835 OopTaskQueueSet* _task_queues; 3836 3837 // Termination (and yielding) support 3838 CMSConcMarkingTerminator _term; 3839 CMSConcMarkingTerminatorTerminator _term_term; 3840 3841 public: 3842 CMSConcMarkingTask(CMSCollector* collector, 3843 CompactibleFreeListSpace* cms_space, 3844 bool asynch, 3845 YieldingFlexibleWorkGang* workers, 3846 OopTaskQueueSet* task_queues): 3847 YieldingFlexibleGangTask("Concurrent marking done multi-threaded"), 3848 _collector(collector), 3849 _cms_space(cms_space), 3850 _asynch(asynch), _n_workers(0), _result(true), 3851 _task_queues(task_queues), 3852 _term(_n_workers, task_queues, _collector), 3853 _bit_map_lock(collector->bitMapLock()) 3854 { 3855 _requested_size = _n_workers; 3856 _term.set_task(this); 3857 _term_term.set_task(this); 3858 _restart_addr = _global_finger = _cms_space->bottom(); 3859 } 3860 3861 3862 OopTaskQueueSet* task_queues() { return _task_queues; } 3863 3864 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 3865 3866 HeapWord** global_finger_addr() { return &_global_finger; } 3867 3868 CMSConcMarkingTerminator* terminator() { return &_term; } 3869 3870 virtual void set_for_termination(int active_workers) { 3871 terminator()->reset_for_reuse(active_workers); 3872 } 3873 3874 void work(uint worker_id); 3875 bool should_yield() { 3876 return ConcurrentMarkSweepThread::should_yield() 3877 && !_collector->foregroundGCIsActive() 3878 && _asynch; 3879 } 3880 3881 virtual void coordinator_yield(); // stuff done by coordinator 3882 bool result() { return _result; } 3883 3884 void reset(HeapWord* ra) { 3885 assert(_global_finger >= _cms_space->end(), "Postcondition of ::work(i)"); 3886 _restart_addr = _global_finger = ra; 3887 _term.reset_for_reuse(); 3888 } 3889 3890 static bool get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 3891 OopTaskQueue* work_q); 3892 3893 private: 3894 void do_scan_and_mark(int i, CompactibleFreeListSpace* sp); 3895 void do_work_steal(int i); 3896 void bump_global_finger(HeapWord* f); 3897 }; 3898 3899 bool CMSConcMarkingTerminatorTerminator::should_exit_termination() { 3900 assert(_task != NULL, "Error"); 3901 return _task->yielding(); 3902 // Note that we do not need the disjunct || _task->should_yield() above 3903 // because we want terminating threads to yield only if the task 3904 // is already in the midst of yielding, which happens only after at least one 3905 // thread has yielded. 3906 } 3907 3908 void CMSConcMarkingTerminator::yield() { 3909 if (_task->should_yield()) { 3910 _task->yield(); 3911 } else { 3912 ParallelTaskTerminator::yield(); 3913 } 3914 } 3915 3916 //////////////////////////////////////////////////////////////// 3917 // Concurrent Marking Algorithm Sketch 3918 //////////////////////////////////////////////////////////////// 3919 // Until all tasks exhausted (both spaces): 3920 // -- claim next available chunk 3921 // -- bump global finger via CAS 3922 // -- find first object that starts in this chunk 3923 // and start scanning bitmap from that position 3924 // -- scan marked objects for oops 3925 // -- CAS-mark target, and if successful: 3926 // . if target oop is above global finger (volatile read) 3927 // nothing to do 3928 // . if target oop is in chunk and above local finger 3929 // then nothing to do 3930 // . else push on work-queue 3931 // -- Deal with possible overflow issues: 3932 // . local work-queue overflow causes stuff to be pushed on 3933 // global (common) overflow queue 3934 // . always first empty local work queue 3935 // . then get a batch of oops from global work queue if any 3936 // . then do work stealing 3937 // -- When all tasks claimed (both spaces) 3938 // and local work queue empty, 3939 // then in a loop do: 3940 // . check global overflow stack; steal a batch of oops and trace 3941 // . try to steal from other threads oif GOS is empty 3942 // . if neither is available, offer termination 3943 // -- Terminate and return result 3944 // 3945 void CMSConcMarkingTask::work(uint worker_id) { 3946 elapsedTimer _timer; 3947 ResourceMark rm; 3948 HandleMark hm; 3949 3950 DEBUG_ONLY(_collector->verify_overflow_empty();) 3951 3952 // Before we begin work, our work queue should be empty 3953 assert(work_queue(worker_id)->size() == 0, "Expected to be empty"); 3954 // Scan the bitmap covering _cms_space, tracing through grey objects. 3955 _timer.start(); 3956 do_scan_and_mark(worker_id, _cms_space); 3957 _timer.stop(); 3958 if (PrintCMSStatistics != 0) { 3959 gclog_or_tty->print_cr("Finished cms space scanning in %dth thread: %3.3f sec", 3960 worker_id, _timer.seconds()); 3961 // XXX: need xxx/xxx type of notation, two timers 3962 } 3963 3964 // ... do work stealing 3965 _timer.reset(); 3966 _timer.start(); 3967 do_work_steal(worker_id); 3968 _timer.stop(); 3969 if (PrintCMSStatistics != 0) { 3970 gclog_or_tty->print_cr("Finished work stealing in %dth thread: %3.3f sec", 3971 worker_id, _timer.seconds()); 3972 // XXX: need xxx/xxx type of notation, two timers 3973 } 3974 assert(_collector->_markStack.isEmpty(), "Should have been emptied"); 3975 assert(work_queue(worker_id)->size() == 0, "Should have been emptied"); 3976 // Note that under the current task protocol, the 3977 // following assertion is true even of the spaces 3978 // expanded since the completion of the concurrent 3979 // marking. XXX This will likely change under a strict 3980 // ABORT semantics. 3981 // After perm removal the comparison was changed to 3982 // greater than or equal to from strictly greater than. 3983 // Before perm removal the highest address sweep would 3984 // have been at the end of perm gen but now is at the 3985 // end of the tenured gen. 3986 assert(_global_finger >= _cms_space->end(), 3987 "All tasks have been completed"); 3988 DEBUG_ONLY(_collector->verify_overflow_empty();) 3989 } 3990 3991 void CMSConcMarkingTask::bump_global_finger(HeapWord* f) { 3992 HeapWord* read = _global_finger; 3993 HeapWord* cur = read; 3994 while (f > read) { 3995 cur = read; 3996 read = (HeapWord*) Atomic::cmpxchg_ptr(f, &_global_finger, cur); 3997 if (cur == read) { 3998 // our cas succeeded 3999 assert(_global_finger >= f, "protocol consistency"); 4000 break; 4001 } 4002 } 4003 } 4004 4005 // This is really inefficient, and should be redone by 4006 // using (not yet available) block-read and -write interfaces to the 4007 // stack and the work_queue. XXX FIX ME !!! 4008 bool CMSConcMarkingTask::get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 4009 OopTaskQueue* work_q) { 4010 // Fast lock-free check 4011 if (ovflw_stk->length() == 0) { 4012 return false; 4013 } 4014 assert(work_q->size() == 0, "Shouldn't steal"); 4015 MutexLockerEx ml(ovflw_stk->par_lock(), 4016 Mutex::_no_safepoint_check_flag); 4017 // Grab up to 1/4 the size of the work queue 4018 size_t num = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 4019 (size_t)ParGCDesiredObjsFromOverflowList); 4020 num = MIN2(num, ovflw_stk->length()); 4021 for (int i = (int) num; i > 0; i--) { 4022 oop cur = ovflw_stk->pop(); 4023 assert(cur != NULL, "Counted wrong?"); 4024 work_q->push(cur); 4025 } 4026 return num > 0; 4027 } 4028 4029 void CMSConcMarkingTask::do_scan_and_mark(int i, CompactibleFreeListSpace* sp) { 4030 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 4031 int n_tasks = pst->n_tasks(); 4032 // We allow that there may be no tasks to do here because 4033 // we are restarting after a stack overflow. 4034 assert(pst->valid() || n_tasks == 0, "Uninitialized use?"); 4035 uint nth_task = 0; 4036 4037 HeapWord* aligned_start = sp->bottom(); 4038 if (sp->used_region().contains(_restart_addr)) { 4039 // Align down to a card boundary for the start of 0th task 4040 // for this space. 4041 aligned_start = 4042 (HeapWord*)align_size_down((uintptr_t)_restart_addr, 4043 CardTableModRefBS::card_size); 4044 } 4045 4046 size_t chunk_size = sp->marking_task_size(); 4047 while (!pst->is_task_claimed(/* reference */ nth_task)) { 4048 // Having claimed the nth task in this space, 4049 // compute the chunk that it corresponds to: 4050 MemRegion span = MemRegion(aligned_start + nth_task*chunk_size, 4051 aligned_start + (nth_task+1)*chunk_size); 4052 // Try and bump the global finger via a CAS; 4053 // note that we need to do the global finger bump 4054 // _before_ taking the intersection below, because 4055 // the task corresponding to that region will be 4056 // deemed done even if the used_region() expands 4057 // because of allocation -- as it almost certainly will 4058 // during start-up while the threads yield in the 4059 // closure below. 4060 HeapWord* finger = span.end(); 4061 bump_global_finger(finger); // atomically 4062 // There are null tasks here corresponding to chunks 4063 // beyond the "top" address of the space. 4064 span = span.intersection(sp->used_region()); 4065 if (!span.is_empty()) { // Non-null task 4066 HeapWord* prev_obj; 4067 assert(!span.contains(_restart_addr) || nth_task == 0, 4068 "Inconsistency"); 4069 if (nth_task == 0) { 4070 // For the 0th task, we'll not need to compute a block_start. 4071 if (span.contains(_restart_addr)) { 4072 // In the case of a restart because of stack overflow, 4073 // we might additionally skip a chunk prefix. 4074 prev_obj = _restart_addr; 4075 } else { 4076 prev_obj = span.start(); 4077 } 4078 } else { 4079 // We want to skip the first object because 4080 // the protocol is to scan any object in its entirety 4081 // that _starts_ in this span; a fortiori, any 4082 // object starting in an earlier span is scanned 4083 // as part of an earlier claimed task. 4084 // Below we use the "careful" version of block_start 4085 // so we do not try to navigate uninitialized objects. 4086 prev_obj = sp->block_start_careful(span.start()); 4087 // Below we use a variant of block_size that uses the 4088 // Printezis bits to avoid waiting for allocated 4089 // objects to become initialized/parsable. 4090 while (prev_obj < span.start()) { 4091 size_t sz = sp->block_size_no_stall(prev_obj, _collector); 4092 if (sz > 0) { 4093 prev_obj += sz; 4094 } else { 4095 // In this case we may end up doing a bit of redundant 4096 // scanning, but that appears unavoidable, short of 4097 // locking the free list locks; see bug 6324141. 4098 break; 4099 } 4100 } 4101 } 4102 if (prev_obj < span.end()) { 4103 MemRegion my_span = MemRegion(prev_obj, span.end()); 4104 // Do the marking work within a non-empty span -- 4105 // the last argument to the constructor indicates whether the 4106 // iteration should be incremental with periodic yields. 4107 Par_MarkFromRootsClosure cl(this, _collector, my_span, 4108 &_collector->_markBitMap, 4109 work_queue(i), 4110 &_collector->_markStack, 4111 _asynch); 4112 _collector->_markBitMap.iterate(&cl, my_span.start(), my_span.end()); 4113 } // else nothing to do for this task 4114 } // else nothing to do for this task 4115 } 4116 // We'd be tempted to assert here that since there are no 4117 // more tasks left to claim in this space, the global_finger 4118 // must exceed space->top() and a fortiori space->end(). However, 4119 // that would not quite be correct because the bumping of 4120 // global_finger occurs strictly after the claiming of a task, 4121 // so by the time we reach here the global finger may not yet 4122 // have been bumped up by the thread that claimed the last 4123 // task. 4124 pst->all_tasks_completed(); 4125 } 4126 4127 class Par_ConcMarkingClosure: public MetadataAwareOopClosure { 4128 private: 4129 CMSCollector* _collector; 4130 CMSConcMarkingTask* _task; 4131 MemRegion _span; 4132 CMSBitMap* _bit_map; 4133 CMSMarkStack* _overflow_stack; 4134 OopTaskQueue* _work_queue; 4135 protected: 4136 DO_OOP_WORK_DEFN 4137 public: 4138 Par_ConcMarkingClosure(CMSCollector* collector, CMSConcMarkingTask* task, OopTaskQueue* work_queue, 4139 CMSBitMap* bit_map, CMSMarkStack* overflow_stack): 4140 MetadataAwareOopClosure(collector->ref_processor()), 4141 _collector(collector), 4142 _task(task), 4143 _span(collector->_span), 4144 _work_queue(work_queue), 4145 _bit_map(bit_map), 4146 _overflow_stack(overflow_stack) 4147 { } 4148 virtual void do_oop(oop* p); 4149 virtual void do_oop(narrowOop* p); 4150 4151 void trim_queue(size_t max); 4152 void handle_stack_overflow(HeapWord* lost); 4153 void do_yield_check() { 4154 if (_task->should_yield()) { 4155 _task->yield(); 4156 } 4157 } 4158 }; 4159 4160 // Grey object scanning during work stealing phase -- 4161 // the salient assumption here is that any references 4162 // that are in these stolen objects being scanned must 4163 // already have been initialized (else they would not have 4164 // been published), so we do not need to check for 4165 // uninitialized objects before pushing here. 4166 void Par_ConcMarkingClosure::do_oop(oop obj) { 4167 assert(obj->is_oop_or_null(true), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj))); 4168 HeapWord* addr = (HeapWord*)obj; 4169 // Check if oop points into the CMS generation 4170 // and is not marked 4171 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 4172 // a white object ... 4173 // If we manage to "claim" the object, by being the 4174 // first thread to mark it, then we push it on our 4175 // marking stack 4176 if (_bit_map->par_mark(addr)) { // ... now grey 4177 // push on work queue (grey set) 4178 bool simulate_overflow = false; 4179 NOT_PRODUCT( 4180 if (CMSMarkStackOverflowALot && 4181 _collector->simulate_overflow()) { 4182 // simulate a stack overflow 4183 simulate_overflow = true; 4184 } 4185 ) 4186 if (simulate_overflow || 4187 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) { 4188 // stack overflow 4189 if (PrintCMSStatistics != 0) { 4190 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 4191 SIZE_FORMAT, _overflow_stack->capacity()); 4192 } 4193 // We cannot assert that the overflow stack is full because 4194 // it may have been emptied since. 4195 assert(simulate_overflow || 4196 _work_queue->size() == _work_queue->max_elems(), 4197 "Else push should have succeeded"); 4198 handle_stack_overflow(addr); 4199 } 4200 } // Else, some other thread got there first 4201 do_yield_check(); 4202 } 4203 } 4204 4205 void Par_ConcMarkingClosure::do_oop(oop* p) { Par_ConcMarkingClosure::do_oop_work(p); } 4206 void Par_ConcMarkingClosure::do_oop(narrowOop* p) { Par_ConcMarkingClosure::do_oop_work(p); } 4207 4208 void Par_ConcMarkingClosure::trim_queue(size_t max) { 4209 while (_work_queue->size() > max) { 4210 oop new_oop; 4211 if (_work_queue->pop_local(new_oop)) { 4212 assert(new_oop->is_oop(), "Should be an oop"); 4213 assert(_bit_map->isMarked((HeapWord*)new_oop), "Grey object"); 4214 assert(_span.contains((HeapWord*)new_oop), "Not in span"); 4215 new_oop->oop_iterate(this); // do_oop() above 4216 do_yield_check(); 4217 } 4218 } 4219 } 4220 4221 // Upon stack overflow, we discard (part of) the stack, 4222 // remembering the least address amongst those discarded 4223 // in CMSCollector's _restart_address. 4224 void Par_ConcMarkingClosure::handle_stack_overflow(HeapWord* lost) { 4225 // We need to do this under a mutex to prevent other 4226 // workers from interfering with the work done below. 4227 MutexLockerEx ml(_overflow_stack->par_lock(), 4228 Mutex::_no_safepoint_check_flag); 4229 // Remember the least grey address discarded 4230 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 4231 _collector->lower_restart_addr(ra); 4232 _overflow_stack->reset(); // discard stack contents 4233 _overflow_stack->expand(); // expand the stack if possible 4234 } 4235 4236 4237 void CMSConcMarkingTask::do_work_steal(int i) { 4238 OopTaskQueue* work_q = work_queue(i); 4239 oop obj_to_scan; 4240 CMSBitMap* bm = &(_collector->_markBitMap); 4241 CMSMarkStack* ovflw = &(_collector->_markStack); 4242 int* seed = _collector->hash_seed(i); 4243 Par_ConcMarkingClosure cl(_collector, this, work_q, bm, ovflw); 4244 while (true) { 4245 cl.trim_queue(0); 4246 assert(work_q->size() == 0, "Should have been emptied above"); 4247 if (get_work_from_overflow_stack(ovflw, work_q)) { 4248 // Can't assert below because the work obtained from the 4249 // overflow stack may already have been stolen from us. 4250 // assert(work_q->size() > 0, "Work from overflow stack"); 4251 continue; 4252 } else if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 4253 assert(obj_to_scan->is_oop(), "Should be an oop"); 4254 assert(bm->isMarked((HeapWord*)obj_to_scan), "Grey object"); 4255 obj_to_scan->oop_iterate(&cl); 4256 } else if (terminator()->offer_termination(&_term_term)) { 4257 assert(work_q->size() == 0, "Impossible!"); 4258 break; 4259 } else if (yielding() || should_yield()) { 4260 yield(); 4261 } 4262 } 4263 } 4264 4265 // This is run by the CMS (coordinator) thread. 4266 void CMSConcMarkingTask::coordinator_yield() { 4267 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 4268 "CMS thread should hold CMS token"); 4269 // First give up the locks, then yield, then re-lock 4270 // We should probably use a constructor/destructor idiom to 4271 // do this unlock/lock or modify the MutexUnlocker class to 4272 // serve our purpose. XXX 4273 assert_lock_strong(_bit_map_lock); 4274 _bit_map_lock->unlock(); 4275 ConcurrentMarkSweepThread::desynchronize(true); 4276 ConcurrentMarkSweepThread::acknowledge_yield_request(); 4277 _collector->stopTimer(); 4278 if (PrintCMSStatistics != 0) { 4279 _collector->incrementYields(); 4280 } 4281 _collector->icms_wait(); 4282 4283 // It is possible for whichever thread initiated the yield request 4284 // not to get a chance to wake up and take the bitmap lock between 4285 // this thread releasing it and reacquiring it. So, while the 4286 // should_yield() flag is on, let's sleep for a bit to give the 4287 // other thread a chance to wake up. The limit imposed on the number 4288 // of iterations is defensive, to avoid any unforseen circumstances 4289 // putting us into an infinite loop. Since it's always been this 4290 // (coordinator_yield()) method that was observed to cause the 4291 // problem, we are using a parameter (CMSCoordinatorYieldSleepCount) 4292 // which is by default non-zero. For the other seven methods that 4293 // also perform the yield operation, as are using a different 4294 // parameter (CMSYieldSleepCount) which is by default zero. This way we 4295 // can enable the sleeping for those methods too, if necessary. 4296 // See 6442774. 4297 // 4298 // We really need to reconsider the synchronization between the GC 4299 // thread and the yield-requesting threads in the future and we 4300 // should really use wait/notify, which is the recommended 4301 // way of doing this type of interaction. Additionally, we should 4302 // consolidate the eight methods that do the yield operation and they 4303 // are almost identical into one for better maintainability and 4304 // readability. See 6445193. 4305 // 4306 // Tony 2006.06.29 4307 for (unsigned i = 0; i < CMSCoordinatorYieldSleepCount && 4308 ConcurrentMarkSweepThread::should_yield() && 4309 !CMSCollector::foregroundGCIsActive(); ++i) { 4310 os::sleep(Thread::current(), 1, false); 4311 ConcurrentMarkSweepThread::acknowledge_yield_request(); 4312 } 4313 4314 ConcurrentMarkSweepThread::synchronize(true); 4315 _bit_map_lock->lock_without_safepoint_check(); 4316 _collector->startTimer(); 4317 } 4318 4319 bool CMSCollector::do_marking_mt(bool asynch) { 4320 assert(ConcGCThreads > 0 && conc_workers() != NULL, "precondition"); 4321 int num_workers = AdaptiveSizePolicy::calc_active_conc_workers( 4322 conc_workers()->total_workers(), 4323 conc_workers()->active_workers(), 4324 Threads::number_of_non_daemon_threads()); 4325 conc_workers()->set_active_workers(num_workers); 4326 4327 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 4328 4329 CMSConcMarkingTask tsk(this, 4330 cms_space, 4331 asynch, 4332 conc_workers(), 4333 task_queues()); 4334 4335 // Since the actual number of workers we get may be different 4336 // from the number we requested above, do we need to do anything different 4337 // below? In particular, may be we need to subclass the SequantialSubTasksDone 4338 // class?? XXX 4339 cms_space ->initialize_sequential_subtasks_for_marking(num_workers); 4340 4341 // Refs discovery is already non-atomic. 4342 assert(!ref_processor()->discovery_is_atomic(), "Should be non-atomic"); 4343 assert(ref_processor()->discovery_is_mt(), "Discovery should be MT"); 4344 conc_workers()->start_task(&tsk); 4345 while (tsk.yielded()) { 4346 tsk.coordinator_yield(); 4347 conc_workers()->continue_task(&tsk); 4348 } 4349 // If the task was aborted, _restart_addr will be non-NULL 4350 assert(tsk.completed() || _restart_addr != NULL, "Inconsistency"); 4351 while (_restart_addr != NULL) { 4352 // XXX For now we do not make use of ABORTED state and have not 4353 // yet implemented the right abort semantics (even in the original 4354 // single-threaded CMS case). That needs some more investigation 4355 // and is deferred for now; see CR# TBF. 07252005YSR. XXX 4356 assert(!CMSAbortSemantics || tsk.aborted(), "Inconsistency"); 4357 // If _restart_addr is non-NULL, a marking stack overflow 4358 // occurred; we need to do a fresh marking iteration from the 4359 // indicated restart address. 4360 if (_foregroundGCIsActive && asynch) { 4361 // We may be running into repeated stack overflows, having 4362 // reached the limit of the stack size, while making very 4363 // slow forward progress. It may be best to bail out and 4364 // let the foreground collector do its job. 4365 // Clear _restart_addr, so that foreground GC 4366 // works from scratch. This avoids the headache of 4367 // a "rescan" which would otherwise be needed because 4368 // of the dirty mod union table & card table. 4369 _restart_addr = NULL; 4370 return false; 4371 } 4372 // Adjust the task to restart from _restart_addr 4373 tsk.reset(_restart_addr); 4374 cms_space ->initialize_sequential_subtasks_for_marking(num_workers, 4375 _restart_addr); 4376 _restart_addr = NULL; 4377 // Get the workers going again 4378 conc_workers()->start_task(&tsk); 4379 while (tsk.yielded()) { 4380 tsk.coordinator_yield(); 4381 conc_workers()->continue_task(&tsk); 4382 } 4383 } 4384 assert(tsk.completed(), "Inconsistency"); 4385 assert(tsk.result() == true, "Inconsistency"); 4386 return true; 4387 } 4388 4389 bool CMSCollector::do_marking_st(bool asynch) { 4390 ResourceMark rm; 4391 HandleMark hm; 4392 4393 // Temporarily make refs discovery single threaded (non-MT) 4394 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false); 4395 MarkFromRootsClosure markFromRootsClosure(this, _span, &_markBitMap, 4396 &_markStack, CMSYield && asynch); 4397 // the last argument to iterate indicates whether the iteration 4398 // should be incremental with periodic yields. 4399 _markBitMap.iterate(&markFromRootsClosure); 4400 // If _restart_addr is non-NULL, a marking stack overflow 4401 // occurred; we need to do a fresh iteration from the 4402 // indicated restart address. 4403 while (_restart_addr != NULL) { 4404 if (_foregroundGCIsActive && asynch) { 4405 // We may be running into repeated stack overflows, having 4406 // reached the limit of the stack size, while making very 4407 // slow forward progress. It may be best to bail out and 4408 // let the foreground collector do its job. 4409 // Clear _restart_addr, so that foreground GC 4410 // works from scratch. This avoids the headache of 4411 // a "rescan" which would otherwise be needed because 4412 // of the dirty mod union table & card table. 4413 _restart_addr = NULL; 4414 return false; // indicating failure to complete marking 4415 } 4416 // Deal with stack overflow: 4417 // we restart marking from _restart_addr 4418 HeapWord* ra = _restart_addr; 4419 markFromRootsClosure.reset(ra); 4420 _restart_addr = NULL; 4421 _markBitMap.iterate(&markFromRootsClosure, ra, _span.end()); 4422 } 4423 return true; 4424 } 4425 4426 void CMSCollector::preclean() { 4427 check_correct_thread_executing(); 4428 assert(Thread::current()->is_ConcurrentGC_thread(), "Wrong thread"); 4429 verify_work_stacks_empty(); 4430 verify_overflow_empty(); 4431 _abort_preclean = false; 4432 if (CMSPrecleaningEnabled) { 4433 if (!CMSEdenChunksRecordAlways) { 4434 _eden_chunk_index = 0; 4435 } 4436 size_t used = get_eden_used(); 4437 size_t capacity = get_eden_capacity(); 4438 // Don't start sampling unless we will get sufficiently 4439 // many samples. 4440 if (used < (capacity/(CMSScheduleRemarkSamplingRatio * 100) 4441 * CMSScheduleRemarkEdenPenetration)) { 4442 _start_sampling = true; 4443 } else { 4444 _start_sampling = false; 4445 } 4446 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 4447 CMSPhaseAccounting pa(this, "preclean", _gc_tracer_cm->gc_id(), !PrintGCDetails); 4448 preclean_work(CMSPrecleanRefLists1, CMSPrecleanSurvivors1); 4449 } 4450 CMSTokenSync x(true); // is cms thread 4451 if (CMSPrecleaningEnabled) { 4452 sample_eden(); 4453 _collectorState = AbortablePreclean; 4454 } else { 4455 _collectorState = FinalMarking; 4456 } 4457 verify_work_stacks_empty(); 4458 verify_overflow_empty(); 4459 } 4460 4461 // Try and schedule the remark such that young gen 4462 // occupancy is CMSScheduleRemarkEdenPenetration %. 4463 void CMSCollector::abortable_preclean() { 4464 check_correct_thread_executing(); 4465 assert(CMSPrecleaningEnabled, "Inconsistent control state"); 4466 assert(_collectorState == AbortablePreclean, "Inconsistent control state"); 4467 4468 // If Eden's current occupancy is below this threshold, 4469 // immediately schedule the remark; else preclean 4470 // past the next scavenge in an effort to 4471 // schedule the pause as described above. By choosing 4472 // CMSScheduleRemarkEdenSizeThreshold >= max eden size 4473 // we will never do an actual abortable preclean cycle. 4474 if (get_eden_used() > CMSScheduleRemarkEdenSizeThreshold) { 4475 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 4476 CMSPhaseAccounting pa(this, "abortable-preclean", _gc_tracer_cm->gc_id(), !PrintGCDetails); 4477 // We need more smarts in the abortable preclean 4478 // loop below to deal with cases where allocation 4479 // in young gen is very very slow, and our precleaning 4480 // is running a losing race against a horde of 4481 // mutators intent on flooding us with CMS updates 4482 // (dirty cards). 4483 // One, admittedly dumb, strategy is to give up 4484 // after a certain number of abortable precleaning loops 4485 // or after a certain maximum time. We want to make 4486 // this smarter in the next iteration. 4487 // XXX FIX ME!!! YSR 4488 size_t loops = 0, workdone = 0, cumworkdone = 0, waited = 0; 4489 while (!(should_abort_preclean() || 4490 ConcurrentMarkSweepThread::should_terminate())) { 4491 workdone = preclean_work(CMSPrecleanRefLists2, CMSPrecleanSurvivors2); 4492 cumworkdone += workdone; 4493 loops++; 4494 // Voluntarily terminate abortable preclean phase if we have 4495 // been at it for too long. 4496 if ((CMSMaxAbortablePrecleanLoops != 0) && 4497 loops >= CMSMaxAbortablePrecleanLoops) { 4498 if (PrintGCDetails) { 4499 gclog_or_tty->print(" CMS: abort preclean due to loops "); 4500 } 4501 break; 4502 } 4503 if (pa.wallclock_millis() > CMSMaxAbortablePrecleanTime) { 4504 if (PrintGCDetails) { 4505 gclog_or_tty->print(" CMS: abort preclean due to time "); 4506 } 4507 break; 4508 } 4509 // If we are doing little work each iteration, we should 4510 // take a short break. 4511 if (workdone < CMSAbortablePrecleanMinWorkPerIteration) { 4512 // Sleep for some time, waiting for work to accumulate 4513 stopTimer(); 4514 cmsThread()->wait_on_cms_lock(CMSAbortablePrecleanWaitMillis); 4515 startTimer(); 4516 waited++; 4517 } 4518 } 4519 if (PrintCMSStatistics > 0) { 4520 gclog_or_tty->print(" [" SIZE_FORMAT " iterations, " SIZE_FORMAT " waits, " SIZE_FORMAT " cards)] ", 4521 loops, waited, cumworkdone); 4522 } 4523 } 4524 CMSTokenSync x(true); // is cms thread 4525 if (_collectorState != Idling) { 4526 assert(_collectorState == AbortablePreclean, 4527 "Spontaneous state transition?"); 4528 _collectorState = FinalMarking; 4529 } // Else, a foreground collection completed this CMS cycle. 4530 return; 4531 } 4532 4533 // Respond to an Eden sampling opportunity 4534 void CMSCollector::sample_eden() { 4535 // Make sure a young gc cannot sneak in between our 4536 // reading and recording of a sample. 4537 assert(Thread::current()->is_ConcurrentGC_thread(), 4538 "Only the cms thread may collect Eden samples"); 4539 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 4540 "Should collect samples while holding CMS token"); 4541 if (!_start_sampling) { 4542 return; 4543 } 4544 // When CMSEdenChunksRecordAlways is true, the eden chunk array 4545 // is populated by the young generation. 4546 if (_eden_chunk_array != NULL && !CMSEdenChunksRecordAlways) { 4547 if (_eden_chunk_index < _eden_chunk_capacity) { 4548 _eden_chunk_array[_eden_chunk_index] = *_top_addr; // take sample 4549 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr, 4550 "Unexpected state of Eden"); 4551 // We'd like to check that what we just sampled is an oop-start address; 4552 // however, we cannot do that here since the object may not yet have been 4553 // initialized. So we'll instead do the check when we _use_ this sample 4554 // later. 4555 if (_eden_chunk_index == 0 || 4556 (pointer_delta(_eden_chunk_array[_eden_chunk_index], 4557 _eden_chunk_array[_eden_chunk_index-1]) 4558 >= CMSSamplingGrain)) { 4559 _eden_chunk_index++; // commit sample 4560 } 4561 } 4562 } 4563 if ((_collectorState == AbortablePreclean) && !_abort_preclean) { 4564 size_t used = get_eden_used(); 4565 size_t capacity = get_eden_capacity(); 4566 assert(used <= capacity, "Unexpected state of Eden"); 4567 if (used > (capacity/100 * CMSScheduleRemarkEdenPenetration)) { 4568 _abort_preclean = true; 4569 } 4570 } 4571 } 4572 4573 4574 size_t CMSCollector::preclean_work(bool clean_refs, bool clean_survivor) { 4575 assert(_collectorState == Precleaning || 4576 _collectorState == AbortablePreclean, "incorrect state"); 4577 ResourceMark rm; 4578 HandleMark hm; 4579 4580 // Precleaning is currently not MT but the reference processor 4581 // may be set for MT. Disable it temporarily here. 4582 ReferenceProcessor* rp = ref_processor(); 4583 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(rp, false); 4584 4585 // Do one pass of scrubbing the discovered reference lists 4586 // to remove any reference objects with strongly-reachable 4587 // referents. 4588 if (clean_refs) { 4589 CMSPrecleanRefsYieldClosure yield_cl(this); 4590 assert(rp->span().equals(_span), "Spans should be equal"); 4591 CMSKeepAliveClosure keep_alive(this, _span, &_markBitMap, 4592 &_markStack, true /* preclean */); 4593 CMSDrainMarkingStackClosure complete_trace(this, 4594 _span, &_markBitMap, &_markStack, 4595 &keep_alive, true /* preclean */); 4596 4597 // We don't want this step to interfere with a young 4598 // collection because we don't want to take CPU 4599 // or memory bandwidth away from the young GC threads 4600 // (which may be as many as there are CPUs). 4601 // Note that we don't need to protect ourselves from 4602 // interference with mutators because they can't 4603 // manipulate the discovered reference lists nor affect 4604 // the computed reachability of the referents, the 4605 // only properties manipulated by the precleaning 4606 // of these reference lists. 4607 stopTimer(); 4608 CMSTokenSyncWithLocks x(true /* is cms thread */, 4609 bitMapLock()); 4610 startTimer(); 4611 sample_eden(); 4612 4613 // The following will yield to allow foreground 4614 // collection to proceed promptly. XXX YSR: 4615 // The code in this method may need further 4616 // tweaking for better performance and some restructuring 4617 // for cleaner interfaces. 4618 GCTimer *gc_timer = NULL; // Currently not tracing concurrent phases 4619 rp->preclean_discovered_references( 4620 rp->is_alive_non_header(), &keep_alive, &complete_trace, &yield_cl, 4621 gc_timer, _gc_tracer_cm->gc_id()); 4622 } 4623 4624 if (clean_survivor) { // preclean the active survivor space(s) 4625 assert(_young_gen->kind() == Generation::DefNew || 4626 _young_gen->kind() == Generation::ParNew, 4627 "incorrect type for cast"); 4628 DefNewGeneration* dng = (DefNewGeneration*)_young_gen; 4629 PushAndMarkClosure pam_cl(this, _span, ref_processor(), 4630 &_markBitMap, &_modUnionTable, 4631 &_markStack, true /* precleaning phase */); 4632 stopTimer(); 4633 CMSTokenSyncWithLocks ts(true /* is cms thread */, 4634 bitMapLock()); 4635 startTimer(); 4636 unsigned int before_count = 4637 GenCollectedHeap::heap()->total_collections(); 4638 SurvivorSpacePrecleanClosure 4639 sss_cl(this, _span, &_markBitMap, &_markStack, 4640 &pam_cl, before_count, CMSYield); 4641 dng->from()->object_iterate_careful(&sss_cl); 4642 dng->to()->object_iterate_careful(&sss_cl); 4643 } 4644 MarkRefsIntoAndScanClosure 4645 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable, 4646 &_markStack, this, CMSYield, 4647 true /* precleaning phase */); 4648 // CAUTION: The following closure has persistent state that may need to 4649 // be reset upon a decrease in the sequence of addresses it 4650 // processes. 4651 ScanMarkedObjectsAgainCarefullyClosure 4652 smoac_cl(this, _span, 4653 &_markBitMap, &_markStack, &mrias_cl, CMSYield); 4654 4655 // Preclean dirty cards in ModUnionTable and CardTable using 4656 // appropriate convergence criterion; 4657 // repeat CMSPrecleanIter times unless we find that 4658 // we are losing. 4659 assert(CMSPrecleanIter < 10, "CMSPrecleanIter is too large"); 4660 assert(CMSPrecleanNumerator < CMSPrecleanDenominator, 4661 "Bad convergence multiplier"); 4662 assert(CMSPrecleanThreshold >= 100, 4663 "Unreasonably low CMSPrecleanThreshold"); 4664 4665 size_t numIter, cumNumCards, lastNumCards, curNumCards; 4666 for (numIter = 0, cumNumCards = lastNumCards = curNumCards = 0; 4667 numIter < CMSPrecleanIter; 4668 numIter++, lastNumCards = curNumCards, cumNumCards += curNumCards) { 4669 curNumCards = preclean_mod_union_table(_cmsGen, &smoac_cl); 4670 if (Verbose && PrintGCDetails) { 4671 gclog_or_tty->print(" (modUnionTable: " SIZE_FORMAT " cards)", curNumCards); 4672 } 4673 // Either there are very few dirty cards, so re-mark 4674 // pause will be small anyway, or our pre-cleaning isn't 4675 // that much faster than the rate at which cards are being 4676 // dirtied, so we might as well stop and re-mark since 4677 // precleaning won't improve our re-mark time by much. 4678 if (curNumCards <= CMSPrecleanThreshold || 4679 (numIter > 0 && 4680 (curNumCards * CMSPrecleanDenominator > 4681 lastNumCards * CMSPrecleanNumerator))) { 4682 numIter++; 4683 cumNumCards += curNumCards; 4684 break; 4685 } 4686 } 4687 4688 preclean_klasses(&mrias_cl, _cmsGen->freelistLock()); 4689 4690 curNumCards = preclean_card_table(_cmsGen, &smoac_cl); 4691 cumNumCards += curNumCards; 4692 if (PrintGCDetails && PrintCMSStatistics != 0) { 4693 gclog_or_tty->print_cr(" (cardTable: " SIZE_FORMAT " cards, re-scanned " SIZE_FORMAT " cards, " SIZE_FORMAT " iterations)", 4694 curNumCards, cumNumCards, numIter); 4695 } 4696 return cumNumCards; // as a measure of useful work done 4697 } 4698 4699 // PRECLEANING NOTES: 4700 // Precleaning involves: 4701 // . reading the bits of the modUnionTable and clearing the set bits. 4702 // . For the cards corresponding to the set bits, we scan the 4703 // objects on those cards. This means we need the free_list_lock 4704 // so that we can safely iterate over the CMS space when scanning 4705 // for oops. 4706 // . When we scan the objects, we'll be both reading and setting 4707 // marks in the marking bit map, so we'll need the marking bit map. 4708 // . For protecting _collector_state transitions, we take the CGC_lock. 4709 // Note that any races in the reading of of card table entries by the 4710 // CMS thread on the one hand and the clearing of those entries by the 4711 // VM thread or the setting of those entries by the mutator threads on the 4712 // other are quite benign. However, for efficiency it makes sense to keep 4713 // the VM thread from racing with the CMS thread while the latter is 4714 // dirty card info to the modUnionTable. We therefore also use the 4715 // CGC_lock to protect the reading of the card table and the mod union 4716 // table by the CM thread. 4717 // . We run concurrently with mutator updates, so scanning 4718 // needs to be done carefully -- we should not try to scan 4719 // potentially uninitialized objects. 4720 // 4721 // Locking strategy: While holding the CGC_lock, we scan over and 4722 // reset a maximal dirty range of the mod union / card tables, then lock 4723 // the free_list_lock and bitmap lock to do a full marking, then 4724 // release these locks; and repeat the cycle. This allows for a 4725 // certain amount of fairness in the sharing of these locks between 4726 // the CMS collector on the one hand, and the VM thread and the 4727 // mutators on the other. 4728 4729 // NOTE: preclean_mod_union_table() and preclean_card_table() 4730 // further below are largely identical; if you need to modify 4731 // one of these methods, please check the other method too. 4732 4733 size_t CMSCollector::preclean_mod_union_table( 4734 ConcurrentMarkSweepGeneration* gen, 4735 ScanMarkedObjectsAgainCarefullyClosure* cl) { 4736 verify_work_stacks_empty(); 4737 verify_overflow_empty(); 4738 4739 // strategy: starting with the first card, accumulate contiguous 4740 // ranges of dirty cards; clear these cards, then scan the region 4741 // covered by these cards. 4742 4743 // Since all of the MUT is committed ahead, we can just use 4744 // that, in case the generations expand while we are precleaning. 4745 // It might also be fine to just use the committed part of the 4746 // generation, but we might potentially miss cards when the 4747 // generation is rapidly expanding while we are in the midst 4748 // of precleaning. 4749 HeapWord* startAddr = gen->reserved().start(); 4750 HeapWord* endAddr = gen->reserved().end(); 4751 4752 cl->setFreelistLock(gen->freelistLock()); // needed for yielding 4753 4754 size_t numDirtyCards, cumNumDirtyCards; 4755 HeapWord *nextAddr, *lastAddr; 4756 for (cumNumDirtyCards = numDirtyCards = 0, 4757 nextAddr = lastAddr = startAddr; 4758 nextAddr < endAddr; 4759 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 4760 4761 ResourceMark rm; 4762 HandleMark hm; 4763 4764 MemRegion dirtyRegion; 4765 { 4766 stopTimer(); 4767 // Potential yield point 4768 CMSTokenSync ts(true); 4769 startTimer(); 4770 sample_eden(); 4771 // Get dirty region starting at nextOffset (inclusive), 4772 // simultaneously clearing it. 4773 dirtyRegion = 4774 _modUnionTable.getAndClearMarkedRegion(nextAddr, endAddr); 4775 assert(dirtyRegion.start() >= nextAddr, 4776 "returned region inconsistent?"); 4777 } 4778 // Remember where the next search should begin. 4779 // The returned region (if non-empty) is a right open interval, 4780 // so lastOffset is obtained from the right end of that 4781 // interval. 4782 lastAddr = dirtyRegion.end(); 4783 // Should do something more transparent and less hacky XXX 4784 numDirtyCards = 4785 _modUnionTable.heapWordDiffToOffsetDiff(dirtyRegion.word_size()); 4786 4787 // We'll scan the cards in the dirty region (with periodic 4788 // yields for foreground GC as needed). 4789 if (!dirtyRegion.is_empty()) { 4790 assert(numDirtyCards > 0, "consistency check"); 4791 HeapWord* stop_point = NULL; 4792 stopTimer(); 4793 // Potential yield point 4794 CMSTokenSyncWithLocks ts(true, gen->freelistLock(), 4795 bitMapLock()); 4796 startTimer(); 4797 { 4798 verify_work_stacks_empty(); 4799 verify_overflow_empty(); 4800 sample_eden(); 4801 stop_point = 4802 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 4803 } 4804 if (stop_point != NULL) { 4805 // The careful iteration stopped early either because it found an 4806 // uninitialized object, or because we were in the midst of an 4807 // "abortable preclean", which should now be aborted. Redirty 4808 // the bits corresponding to the partially-scanned or unscanned 4809 // cards. We'll either restart at the next block boundary or 4810 // abort the preclean. 4811 assert((_collectorState == AbortablePreclean && should_abort_preclean()), 4812 "Should only be AbortablePreclean."); 4813 _modUnionTable.mark_range(MemRegion(stop_point, dirtyRegion.end())); 4814 if (should_abort_preclean()) { 4815 break; // out of preclean loop 4816 } else { 4817 // Compute the next address at which preclean should pick up; 4818 // might need bitMapLock in order to read P-bits. 4819 lastAddr = next_card_start_after_block(stop_point); 4820 } 4821 } 4822 } else { 4823 assert(lastAddr == endAddr, "consistency check"); 4824 assert(numDirtyCards == 0, "consistency check"); 4825 break; 4826 } 4827 } 4828 verify_work_stacks_empty(); 4829 verify_overflow_empty(); 4830 return cumNumDirtyCards; 4831 } 4832 4833 // NOTE: preclean_mod_union_table() above and preclean_card_table() 4834 // below are largely identical; if you need to modify 4835 // one of these methods, please check the other method too. 4836 4837 size_t CMSCollector::preclean_card_table(ConcurrentMarkSweepGeneration* gen, 4838 ScanMarkedObjectsAgainCarefullyClosure* cl) { 4839 // strategy: it's similar to precleamModUnionTable above, in that 4840 // we accumulate contiguous ranges of dirty cards, mark these cards 4841 // precleaned, then scan the region covered by these cards. 4842 HeapWord* endAddr = (HeapWord*)(gen->_virtual_space.high()); 4843 HeapWord* startAddr = (HeapWord*)(gen->_virtual_space.low()); 4844 4845 cl->setFreelistLock(gen->freelistLock()); // needed for yielding 4846 4847 size_t numDirtyCards, cumNumDirtyCards; 4848 HeapWord *lastAddr, *nextAddr; 4849 4850 for (cumNumDirtyCards = numDirtyCards = 0, 4851 nextAddr = lastAddr = startAddr; 4852 nextAddr < endAddr; 4853 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 4854 4855 ResourceMark rm; 4856 HandleMark hm; 4857 4858 MemRegion dirtyRegion; 4859 { 4860 // See comments in "Precleaning notes" above on why we 4861 // do this locking. XXX Could the locking overheads be 4862 // too high when dirty cards are sparse? [I don't think so.] 4863 stopTimer(); 4864 CMSTokenSync x(true); // is cms thread 4865 startTimer(); 4866 sample_eden(); 4867 // Get and clear dirty region from card table 4868 dirtyRegion = _ct->ct_bs()->dirty_card_range_after_reset( 4869 MemRegion(nextAddr, endAddr), 4870 true, 4871 CardTableModRefBS::precleaned_card_val()); 4872 4873 assert(dirtyRegion.start() >= nextAddr, 4874 "returned region inconsistent?"); 4875 } 4876 lastAddr = dirtyRegion.end(); 4877 numDirtyCards = 4878 dirtyRegion.word_size()/CardTableModRefBS::card_size_in_words; 4879 4880 if (!dirtyRegion.is_empty()) { 4881 stopTimer(); 4882 CMSTokenSyncWithLocks ts(true, gen->freelistLock(), bitMapLock()); 4883 startTimer(); 4884 sample_eden(); 4885 verify_work_stacks_empty(); 4886 verify_overflow_empty(); 4887 HeapWord* stop_point = 4888 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 4889 if (stop_point != NULL) { 4890 assert((_collectorState == AbortablePreclean && should_abort_preclean()), 4891 "Should only be AbortablePreclean."); 4892 _ct->ct_bs()->invalidate(MemRegion(stop_point, dirtyRegion.end())); 4893 if (should_abort_preclean()) { 4894 break; // out of preclean loop 4895 } else { 4896 // Compute the next address at which preclean should pick up. 4897 lastAddr = next_card_start_after_block(stop_point); 4898 } 4899 } 4900 } else { 4901 break; 4902 } 4903 } 4904 verify_work_stacks_empty(); 4905 verify_overflow_empty(); 4906 return cumNumDirtyCards; 4907 } 4908 4909 class PrecleanKlassClosure : public KlassClosure { 4910 KlassToOopClosure _cm_klass_closure; 4911 public: 4912 PrecleanKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {} 4913 void do_klass(Klass* k) { 4914 if (k->has_accumulated_modified_oops()) { 4915 k->clear_accumulated_modified_oops(); 4916 4917 _cm_klass_closure.do_klass(k); 4918 } 4919 } 4920 }; 4921 4922 // The freelist lock is needed to prevent asserts, is it really needed? 4923 void CMSCollector::preclean_klasses(MarkRefsIntoAndScanClosure* cl, Mutex* freelistLock) { 4924 4925 cl->set_freelistLock(freelistLock); 4926 4927 CMSTokenSyncWithLocks ts(true, freelistLock, bitMapLock()); 4928 4929 // SSS: Add equivalent to ScanMarkedObjectsAgainCarefullyClosure::do_yield_check and should_abort_preclean? 4930 // SSS: We should probably check if precleaning should be aborted, at suitable intervals? 4931 PrecleanKlassClosure preclean_klass_closure(cl); 4932 ClassLoaderDataGraph::classes_do(&preclean_klass_closure); 4933 4934 verify_work_stacks_empty(); 4935 verify_overflow_empty(); 4936 } 4937 4938 void CMSCollector::checkpointRootsFinal(bool asynch, 4939 bool clear_all_soft_refs, bool init_mark_was_synchronous) { 4940 assert(_collectorState == FinalMarking, "incorrect state transition?"); 4941 check_correct_thread_executing(); 4942 // world is stopped at this checkpoint 4943 assert(SafepointSynchronize::is_at_safepoint(), 4944 "world should be stopped"); 4945 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 4946 4947 verify_work_stacks_empty(); 4948 verify_overflow_empty(); 4949 4950 SpecializationStats::clear(); 4951 if (PrintGCDetails) { 4952 gclog_or_tty->print("[YG occupancy: "SIZE_FORMAT" K ("SIZE_FORMAT" K)]", 4953 _young_gen->used() / K, 4954 _young_gen->capacity() / K); 4955 } 4956 if (asynch) { 4957 if (CMSScavengeBeforeRemark) { 4958 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4959 // Temporarily set flag to false, GCH->do_collection will 4960 // expect it to be false and set to true 4961 FlagSetting fl(gch->_is_gc_active, false); 4962 NOT_PRODUCT(GCTraceTime t("Scavenge-Before-Remark", 4963 PrintGCDetails && Verbose, true, _gc_timer_cm, _gc_tracer_cm->gc_id());) 4964 int level = _cmsGen->level() - 1; 4965 if (level >= 0) { 4966 gch->do_collection(true, // full (i.e. force, see below) 4967 false, // !clear_all_soft_refs 4968 0, // size 4969 false, // is_tlab 4970 level // max_level 4971 ); 4972 } 4973 } 4974 FreelistLocker x(this); 4975 MutexLockerEx y(bitMapLock(), 4976 Mutex::_no_safepoint_check_flag); 4977 assert(!init_mark_was_synchronous, "but that's impossible!"); 4978 checkpointRootsFinalWork(asynch, clear_all_soft_refs, false); 4979 } else { 4980 // already have all the locks 4981 checkpointRootsFinalWork(asynch, clear_all_soft_refs, 4982 init_mark_was_synchronous); 4983 } 4984 verify_work_stacks_empty(); 4985 verify_overflow_empty(); 4986 SpecializationStats::print(); 4987 } 4988 4989 void CMSCollector::checkpointRootsFinalWork(bool asynch, 4990 bool clear_all_soft_refs, bool init_mark_was_synchronous) { 4991 4992 NOT_PRODUCT(GCTraceTime tr("checkpointRootsFinalWork", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());) 4993 4994 assert(haveFreelistLocks(), "must have free list locks"); 4995 assert_lock_strong(bitMapLock()); 4996 4997 ResourceMark rm; 4998 HandleMark hm; 4999 5000 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5001 5002 if (should_unload_classes()) { 5003 CodeCache::gc_prologue(); 5004 } 5005 assert(haveFreelistLocks(), "must have free list locks"); 5006 assert_lock_strong(bitMapLock()); 5007 5008 if (!init_mark_was_synchronous) { 5009 // We might assume that we need not fill TLAB's when 5010 // CMSScavengeBeforeRemark is set, because we may have just done 5011 // a scavenge which would have filled all TLAB's -- and besides 5012 // Eden would be empty. This however may not always be the case -- 5013 // for instance although we asked for a scavenge, it may not have 5014 // happened because of a JNI critical section. We probably need 5015 // a policy for deciding whether we can in that case wait until 5016 // the critical section releases and then do the remark following 5017 // the scavenge, and skip it here. In the absence of that policy, 5018 // or of an indication of whether the scavenge did indeed occur, 5019 // we cannot rely on TLAB's having been filled and must do 5020 // so here just in case a scavenge did not happen. 5021 gch->ensure_parsability(false); // fill TLAB's, but no need to retire them 5022 // Update the saved marks which may affect the root scans. 5023 gch->save_marks(); 5024 5025 if (CMSPrintEdenSurvivorChunks) { 5026 print_eden_and_survivor_chunk_arrays(); 5027 } 5028 5029 { 5030 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;) 5031 5032 // Note on the role of the mod union table: 5033 // Since the marker in "markFromRoots" marks concurrently with 5034 // mutators, it is possible for some reachable objects not to have been 5035 // scanned. For instance, an only reference to an object A was 5036 // placed in object B after the marker scanned B. Unless B is rescanned, 5037 // A would be collected. Such updates to references in marked objects 5038 // are detected via the mod union table which is the set of all cards 5039 // dirtied since the first checkpoint in this GC cycle and prior to 5040 // the most recent young generation GC, minus those cleaned up by the 5041 // concurrent precleaning. 5042 if (CMSParallelRemarkEnabled && CollectedHeap::use_parallel_gc_threads()) { 5043 GCTraceTime t("Rescan (parallel) ", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 5044 do_remark_parallel(); 5045 } else { 5046 GCTraceTime t("Rescan (non-parallel) ", PrintGCDetails, false, 5047 _gc_timer_cm, _gc_tracer_cm->gc_id()); 5048 do_remark_non_parallel(); 5049 } 5050 } 5051 } else { 5052 assert(!asynch, "Can't have init_mark_was_synchronous in asynch mode"); 5053 // The initial mark was stop-world, so there's no rescanning to 5054 // do; go straight on to the next step below. 5055 } 5056 verify_work_stacks_empty(); 5057 verify_overflow_empty(); 5058 5059 { 5060 NOT_PRODUCT(GCTraceTime ts("refProcessingWork", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());) 5061 refProcessingWork(asynch, clear_all_soft_refs); 5062 } 5063 verify_work_stacks_empty(); 5064 verify_overflow_empty(); 5065 5066 if (should_unload_classes()) { 5067 CodeCache::gc_epilogue(); 5068 } 5069 JvmtiExport::gc_epilogue(); 5070 5071 // If we encountered any (marking stack / work queue) overflow 5072 // events during the current CMS cycle, take appropriate 5073 // remedial measures, where possible, so as to try and avoid 5074 // recurrence of that condition. 5075 assert(_markStack.isEmpty(), "No grey objects"); 5076 size_t ser_ovflw = _ser_pmc_remark_ovflw + _ser_pmc_preclean_ovflw + 5077 _ser_kac_ovflw + _ser_kac_preclean_ovflw; 5078 if (ser_ovflw > 0) { 5079 if (PrintCMSStatistics != 0) { 5080 gclog_or_tty->print_cr("Marking stack overflow (benign) " 5081 "(pmc_pc="SIZE_FORMAT", pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT 5082 ", kac_preclean="SIZE_FORMAT")", 5083 _ser_pmc_preclean_ovflw, _ser_pmc_remark_ovflw, 5084 _ser_kac_ovflw, _ser_kac_preclean_ovflw); 5085 } 5086 _markStack.expand(); 5087 _ser_pmc_remark_ovflw = 0; 5088 _ser_pmc_preclean_ovflw = 0; 5089 _ser_kac_preclean_ovflw = 0; 5090 _ser_kac_ovflw = 0; 5091 } 5092 if (_par_pmc_remark_ovflw > 0 || _par_kac_ovflw > 0) { 5093 if (PrintCMSStatistics != 0) { 5094 gclog_or_tty->print_cr("Work queue overflow (benign) " 5095 "(pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT")", 5096 _par_pmc_remark_ovflw, _par_kac_ovflw); 5097 } 5098 _par_pmc_remark_ovflw = 0; 5099 _par_kac_ovflw = 0; 5100 } 5101 if (PrintCMSStatistics != 0) { 5102 if (_markStack._hit_limit > 0) { 5103 gclog_or_tty->print_cr(" (benign) Hit max stack size limit ("SIZE_FORMAT")", 5104 _markStack._hit_limit); 5105 } 5106 if (_markStack._failed_double > 0) { 5107 gclog_or_tty->print_cr(" (benign) Failed stack doubling ("SIZE_FORMAT")," 5108 " current capacity "SIZE_FORMAT, 5109 _markStack._failed_double, 5110 _markStack.capacity()); 5111 } 5112 } 5113 _markStack._hit_limit = 0; 5114 _markStack._failed_double = 0; 5115 5116 if ((VerifyAfterGC || VerifyDuringGC) && 5117 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 5118 verify_after_remark(); 5119 } 5120 5121 _gc_tracer_cm->report_object_count_after_gc(&_is_alive_closure); 5122 5123 // Change under the freelistLocks. 5124 _collectorState = Sweeping; 5125 // Call isAllClear() under bitMapLock 5126 assert(_modUnionTable.isAllClear(), 5127 "Should be clear by end of the final marking"); 5128 assert(_ct->klass_rem_set()->mod_union_is_clear(), 5129 "Should be clear by end of the final marking"); 5130 } 5131 5132 void CMSParInitialMarkTask::work(uint worker_id) { 5133 elapsedTimer _timer; 5134 ResourceMark rm; 5135 HandleMark hm; 5136 5137 // ---------- scan from roots -------------- 5138 _timer.start(); 5139 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5140 Par_MarkRefsIntoClosure par_mri_cl(_collector->_span, &(_collector->_markBitMap)); 5141 5142 // ---------- young gen roots -------------- 5143 { 5144 work_on_young_gen_roots(worker_id, &par_mri_cl); 5145 _timer.stop(); 5146 if (PrintCMSStatistics != 0) { 5147 gclog_or_tty->print_cr( 5148 "Finished young gen initial mark scan work in %dth thread: %3.3f sec", 5149 worker_id, _timer.seconds()); 5150 } 5151 } 5152 5153 // ---------- remaining roots -------------- 5154 _timer.reset(); 5155 _timer.start(); 5156 5157 CLDToOopClosure cld_closure(&par_mri_cl, true); 5158 5159 gch->gen_process_roots(_collector->_cmsGen->level(), 5160 false, // yg was scanned above 5161 false, // this is parallel code 5162 SharedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()), 5163 _collector->should_unload_classes(), 5164 &par_mri_cl, 5165 NULL, 5166 &cld_closure); 5167 assert(_collector->should_unload_classes() 5168 || (_collector->CMSCollector::roots_scanning_options() & SharedHeap::SO_AllCodeCache), 5169 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 5170 _timer.stop(); 5171 if (PrintCMSStatistics != 0) { 5172 gclog_or_tty->print_cr( 5173 "Finished remaining root initial mark scan work in %dth thread: %3.3f sec", 5174 worker_id, _timer.seconds()); 5175 } 5176 } 5177 5178 // Parallel remark task 5179 class CMSParRemarkTask: public CMSParMarkTask { 5180 CompactibleFreeListSpace* _cms_space; 5181 5182 // The per-thread work queues, available here for stealing. 5183 OopTaskQueueSet* _task_queues; 5184 ParallelTaskTerminator _term; 5185 5186 public: 5187 // A value of 0 passed to n_workers will cause the number of 5188 // workers to be taken from the active workers in the work gang. 5189 CMSParRemarkTask(CMSCollector* collector, 5190 CompactibleFreeListSpace* cms_space, 5191 int n_workers, FlexibleWorkGang* workers, 5192 OopTaskQueueSet* task_queues): 5193 CMSParMarkTask("Rescan roots and grey objects in parallel", 5194 collector, n_workers), 5195 _cms_space(cms_space), 5196 _task_queues(task_queues), 5197 _term(n_workers, task_queues) { } 5198 5199 OopTaskQueueSet* task_queues() { return _task_queues; } 5200 5201 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 5202 5203 ParallelTaskTerminator* terminator() { return &_term; } 5204 int n_workers() { return _n_workers; } 5205 5206 void work(uint worker_id); 5207 5208 private: 5209 // ... of dirty cards in old space 5210 void do_dirty_card_rescan_tasks(CompactibleFreeListSpace* sp, int i, 5211 Par_MarkRefsIntoAndScanClosure* cl); 5212 5213 // ... work stealing for the above 5214 void do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, int* seed); 5215 }; 5216 5217 class RemarkKlassClosure : public KlassClosure { 5218 KlassToOopClosure _cm_klass_closure; 5219 public: 5220 RemarkKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {} 5221 void do_klass(Klass* k) { 5222 // Check if we have modified any oops in the Klass during the concurrent marking. 5223 if (k->has_accumulated_modified_oops()) { 5224 k->clear_accumulated_modified_oops(); 5225 5226 // We could have transfered the current modified marks to the accumulated marks, 5227 // like we do with the Card Table to Mod Union Table. But it's not really necessary. 5228 } else if (k->has_modified_oops()) { 5229 // Don't clear anything, this info is needed by the next young collection. 5230 } else { 5231 // No modified oops in the Klass. 5232 return; 5233 } 5234 5235 // The klass has modified fields, need to scan the klass. 5236 _cm_klass_closure.do_klass(k); 5237 } 5238 }; 5239 5240 void CMSParMarkTask::work_on_young_gen_roots(uint worker_id, OopsInGenClosure* cl) { 5241 DefNewGeneration* dng = _collector->_young_gen->as_DefNewGeneration(); 5242 EdenSpace* eden_space = dng->eden(); 5243 ContiguousSpace* from_space = dng->from(); 5244 ContiguousSpace* to_space = dng->to(); 5245 5246 HeapWord** eca = _collector->_eden_chunk_array; 5247 size_t ect = _collector->_eden_chunk_index; 5248 HeapWord** sca = _collector->_survivor_chunk_array; 5249 size_t sct = _collector->_survivor_chunk_index; 5250 5251 assert(ect <= _collector->_eden_chunk_capacity, "out of bounds"); 5252 assert(sct <= _collector->_survivor_chunk_capacity, "out of bounds"); 5253 5254 do_young_space_rescan(worker_id, cl, to_space, NULL, 0); 5255 do_young_space_rescan(worker_id, cl, from_space, sca, sct); 5256 do_young_space_rescan(worker_id, cl, eden_space, eca, ect); 5257 } 5258 5259 // work_queue(i) is passed to the closure 5260 // Par_MarkRefsIntoAndScanClosure. The "i" parameter 5261 // also is passed to do_dirty_card_rescan_tasks() and to 5262 // do_work_steal() to select the i-th task_queue. 5263 5264 void CMSParRemarkTask::work(uint worker_id) { 5265 elapsedTimer _timer; 5266 ResourceMark rm; 5267 HandleMark hm; 5268 5269 // ---------- rescan from roots -------------- 5270 _timer.start(); 5271 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5272 Par_MarkRefsIntoAndScanClosure par_mrias_cl(_collector, 5273 _collector->_span, _collector->ref_processor(), 5274 &(_collector->_markBitMap), 5275 work_queue(worker_id)); 5276 5277 // Rescan young gen roots first since these are likely 5278 // coarsely partitioned and may, on that account, constitute 5279 // the critical path; thus, it's best to start off that 5280 // work first. 5281 // ---------- young gen roots -------------- 5282 { 5283 work_on_young_gen_roots(worker_id, &par_mrias_cl); 5284 _timer.stop(); 5285 if (PrintCMSStatistics != 0) { 5286 gclog_or_tty->print_cr( 5287 "Finished young gen rescan work in %dth thread: %3.3f sec", 5288 worker_id, _timer.seconds()); 5289 } 5290 } 5291 5292 // ---------- remaining roots -------------- 5293 _timer.reset(); 5294 _timer.start(); 5295 gch->gen_process_roots(_collector->_cmsGen->level(), 5296 false, // yg was scanned above 5297 false, // this is parallel code 5298 SharedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()), 5299 _collector->should_unload_classes(), 5300 &par_mrias_cl, 5301 NULL, 5302 NULL); // The dirty klasses will be handled below 5303 5304 assert(_collector->should_unload_classes() 5305 || (_collector->CMSCollector::roots_scanning_options() & SharedHeap::SO_AllCodeCache), 5306 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 5307 _timer.stop(); 5308 if (PrintCMSStatistics != 0) { 5309 gclog_or_tty->print_cr( 5310 "Finished remaining root rescan work in %dth thread: %3.3f sec", 5311 worker_id, _timer.seconds()); 5312 } 5313 5314 // ---------- unhandled CLD scanning ---------- 5315 if (worker_id == 0) { // Single threaded at the moment. 5316 _timer.reset(); 5317 _timer.start(); 5318 5319 // Scan all new class loader data objects and new dependencies that were 5320 // introduced during concurrent marking. 5321 ResourceMark rm; 5322 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds(); 5323 for (int i = 0; i < array->length(); i++) { 5324 par_mrias_cl.do_class_loader_data(array->at(i)); 5325 } 5326 5327 // We don't need to keep track of new CLDs anymore. 5328 ClassLoaderDataGraph::remember_new_clds(false); 5329 5330 _timer.stop(); 5331 if (PrintCMSStatistics != 0) { 5332 gclog_or_tty->print_cr( 5333 "Finished unhandled CLD scanning work in %dth thread: %3.3f sec", 5334 worker_id, _timer.seconds()); 5335 } 5336 } 5337 5338 // ---------- dirty klass scanning ---------- 5339 if (worker_id == 0) { // Single threaded at the moment. 5340 _timer.reset(); 5341 _timer.start(); 5342 5343 // Scan all classes that was dirtied during the concurrent marking phase. 5344 RemarkKlassClosure remark_klass_closure(&par_mrias_cl); 5345 ClassLoaderDataGraph::classes_do(&remark_klass_closure); 5346 5347 _timer.stop(); 5348 if (PrintCMSStatistics != 0) { 5349 gclog_or_tty->print_cr( 5350 "Finished dirty klass scanning work in %dth thread: %3.3f sec", 5351 worker_id, _timer.seconds()); 5352 } 5353 } 5354 5355 // We might have added oops to ClassLoaderData::_handles during the 5356 // concurrent marking phase. These oops point to newly allocated objects 5357 // that are guaranteed to be kept alive. Either by the direct allocation 5358 // code, or when the young collector processes the roots. Hence, 5359 // we don't have to revisit the _handles block during the remark phase. 5360 5361 // ---------- rescan dirty cards ------------ 5362 _timer.reset(); 5363 _timer.start(); 5364 5365 // Do the rescan tasks for each of the two spaces 5366 // (cms_space) in turn. 5367 // "worker_id" is passed to select the task_queue for "worker_id" 5368 do_dirty_card_rescan_tasks(_cms_space, worker_id, &par_mrias_cl); 5369 _timer.stop(); 5370 if (PrintCMSStatistics != 0) { 5371 gclog_or_tty->print_cr( 5372 "Finished dirty card rescan work in %dth thread: %3.3f sec", 5373 worker_id, _timer.seconds()); 5374 } 5375 5376 // ---------- steal work from other threads ... 5377 // ---------- ... and drain overflow list. 5378 _timer.reset(); 5379 _timer.start(); 5380 do_work_steal(worker_id, &par_mrias_cl, _collector->hash_seed(worker_id)); 5381 _timer.stop(); 5382 if (PrintCMSStatistics != 0) { 5383 gclog_or_tty->print_cr( 5384 "Finished work stealing in %dth thread: %3.3f sec", 5385 worker_id, _timer.seconds()); 5386 } 5387 } 5388 5389 // Note that parameter "i" is not used. 5390 void 5391 CMSParMarkTask::do_young_space_rescan(uint worker_id, 5392 OopsInGenClosure* cl, ContiguousSpace* space, 5393 HeapWord** chunk_array, size_t chunk_top) { 5394 // Until all tasks completed: 5395 // . claim an unclaimed task 5396 // . compute region boundaries corresponding to task claimed 5397 // using chunk_array 5398 // . par_oop_iterate(cl) over that region 5399 5400 ResourceMark rm; 5401 HandleMark hm; 5402 5403 SequentialSubTasksDone* pst = space->par_seq_tasks(); 5404 5405 uint nth_task = 0; 5406 uint n_tasks = pst->n_tasks(); 5407 5408 if (n_tasks > 0) { 5409 assert(pst->valid(), "Uninitialized use?"); 5410 HeapWord *start, *end; 5411 while (!pst->is_task_claimed(/* reference */ nth_task)) { 5412 // We claimed task # nth_task; compute its boundaries. 5413 if (chunk_top == 0) { // no samples were taken 5414 assert(nth_task == 0 && n_tasks == 1, "Can have only 1 EdenSpace task"); 5415 start = space->bottom(); 5416 end = space->top(); 5417 } else if (nth_task == 0) { 5418 start = space->bottom(); 5419 end = chunk_array[nth_task]; 5420 } else if (nth_task < (uint)chunk_top) { 5421 assert(nth_task >= 1, "Control point invariant"); 5422 start = chunk_array[nth_task - 1]; 5423 end = chunk_array[nth_task]; 5424 } else { 5425 assert(nth_task == (uint)chunk_top, "Control point invariant"); 5426 start = chunk_array[chunk_top - 1]; 5427 end = space->top(); 5428 } 5429 MemRegion mr(start, end); 5430 // Verify that mr is in space 5431 assert(mr.is_empty() || space->used_region().contains(mr), 5432 "Should be in space"); 5433 // Verify that "start" is an object boundary 5434 assert(mr.is_empty() || oop(mr.start())->is_oop(), 5435 "Should be an oop"); 5436 space->par_oop_iterate(mr, cl); 5437 } 5438 pst->all_tasks_completed(); 5439 } 5440 } 5441 5442 void 5443 CMSParRemarkTask::do_dirty_card_rescan_tasks( 5444 CompactibleFreeListSpace* sp, int i, 5445 Par_MarkRefsIntoAndScanClosure* cl) { 5446 // Until all tasks completed: 5447 // . claim an unclaimed task 5448 // . compute region boundaries corresponding to task claimed 5449 // . transfer dirty bits ct->mut for that region 5450 // . apply rescanclosure to dirty mut bits for that region 5451 5452 ResourceMark rm; 5453 HandleMark hm; 5454 5455 OopTaskQueue* work_q = work_queue(i); 5456 ModUnionClosure modUnionClosure(&(_collector->_modUnionTable)); 5457 // CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! 5458 // CAUTION: This closure has state that persists across calls to 5459 // the work method dirty_range_iterate_clear() in that it has 5460 // embedded in it a (subtype of) UpwardsObjectClosure. The 5461 // use of that state in the embedded UpwardsObjectClosure instance 5462 // assumes that the cards are always iterated (even if in parallel 5463 // by several threads) in monotonically increasing order per each 5464 // thread. This is true of the implementation below which picks 5465 // card ranges (chunks) in monotonically increasing order globally 5466 // and, a-fortiori, in monotonically increasing order per thread 5467 // (the latter order being a subsequence of the former). 5468 // If the work code below is ever reorganized into a more chaotic 5469 // work-partitioning form than the current "sequential tasks" 5470 // paradigm, the use of that persistent state will have to be 5471 // revisited and modified appropriately. See also related 5472 // bug 4756801 work on which should examine this code to make 5473 // sure that the changes there do not run counter to the 5474 // assumptions made here and necessary for correctness and 5475 // efficiency. Note also that this code might yield inefficient 5476 // behavior in the case of very large objects that span one or 5477 // more work chunks. Such objects would potentially be scanned 5478 // several times redundantly. Work on 4756801 should try and 5479 // address that performance anomaly if at all possible. XXX 5480 MemRegion full_span = _collector->_span; 5481 CMSBitMap* bm = &(_collector->_markBitMap); // shared 5482 MarkFromDirtyCardsClosure 5483 greyRescanClosure(_collector, full_span, // entire span of interest 5484 sp, bm, work_q, cl); 5485 5486 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 5487 assert(pst->valid(), "Uninitialized use?"); 5488 uint nth_task = 0; 5489 const int alignment = CardTableModRefBS::card_size * BitsPerWord; 5490 MemRegion span = sp->used_region(); 5491 HeapWord* start_addr = span.start(); 5492 HeapWord* end_addr = (HeapWord*)round_to((intptr_t)span.end(), 5493 alignment); 5494 const size_t chunk_size = sp->rescan_task_size(); // in HeapWord units 5495 assert((HeapWord*)round_to((intptr_t)start_addr, alignment) == 5496 start_addr, "Check alignment"); 5497 assert((size_t)round_to((intptr_t)chunk_size, alignment) == 5498 chunk_size, "Check alignment"); 5499 5500 while (!pst->is_task_claimed(/* reference */ nth_task)) { 5501 // Having claimed the nth_task, compute corresponding mem-region, 5502 // which is a-fortiori aligned correctly (i.e. at a MUT boundary). 5503 // The alignment restriction ensures that we do not need any 5504 // synchronization with other gang-workers while setting or 5505 // clearing bits in thus chunk of the MUT. 5506 MemRegion this_span = MemRegion(start_addr + nth_task*chunk_size, 5507 start_addr + (nth_task+1)*chunk_size); 5508 // The last chunk's end might be way beyond end of the 5509 // used region. In that case pull back appropriately. 5510 if (this_span.end() > end_addr) { 5511 this_span.set_end(end_addr); 5512 assert(!this_span.is_empty(), "Program logic (calculation of n_tasks)"); 5513 } 5514 // Iterate over the dirty cards covering this chunk, marking them 5515 // precleaned, and setting the corresponding bits in the mod union 5516 // table. Since we have been careful to partition at Card and MUT-word 5517 // boundaries no synchronization is needed between parallel threads. 5518 _collector->_ct->ct_bs()->dirty_card_iterate(this_span, 5519 &modUnionClosure); 5520 5521 // Having transferred these marks into the modUnionTable, 5522 // rescan the marked objects on the dirty cards in the modUnionTable. 5523 // Even if this is at a synchronous collection, the initial marking 5524 // may have been done during an asynchronous collection so there 5525 // may be dirty bits in the mod-union table. 5526 _collector->_modUnionTable.dirty_range_iterate_clear( 5527 this_span, &greyRescanClosure); 5528 _collector->_modUnionTable.verifyNoOneBitsInRange( 5529 this_span.start(), 5530 this_span.end()); 5531 } 5532 pst->all_tasks_completed(); // declare that i am done 5533 } 5534 5535 // . see if we can share work_queues with ParNew? XXX 5536 void 5537 CMSParRemarkTask::do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, 5538 int* seed) { 5539 OopTaskQueue* work_q = work_queue(i); 5540 NOT_PRODUCT(int num_steals = 0;) 5541 oop obj_to_scan; 5542 CMSBitMap* bm = &(_collector->_markBitMap); 5543 5544 while (true) { 5545 // Completely finish any left over work from (an) earlier round(s) 5546 cl->trim_queue(0); 5547 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 5548 (size_t)ParGCDesiredObjsFromOverflowList); 5549 // Now check if there's any work in the overflow list 5550 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads, 5551 // only affects the number of attempts made to get work from the 5552 // overflow list and does not affect the number of workers. Just 5553 // pass ParallelGCThreads so this behavior is unchanged. 5554 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 5555 work_q, 5556 ParallelGCThreads)) { 5557 // found something in global overflow list; 5558 // not yet ready to go stealing work from others. 5559 // We'd like to assert(work_q->size() != 0, ...) 5560 // because we just took work from the overflow list, 5561 // but of course we can't since all of that could have 5562 // been already stolen from us. 5563 // "He giveth and He taketh away." 5564 continue; 5565 } 5566 // Verify that we have no work before we resort to stealing 5567 assert(work_q->size() == 0, "Have work, shouldn't steal"); 5568 // Try to steal from other queues that have work 5569 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 5570 NOT_PRODUCT(num_steals++;) 5571 assert(obj_to_scan->is_oop(), "Oops, not an oop!"); 5572 assert(bm->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 5573 // Do scanning work 5574 obj_to_scan->oop_iterate(cl); 5575 // Loop around, finish this work, and try to steal some more 5576 } else if (terminator()->offer_termination()) { 5577 break; // nirvana from the infinite cycle 5578 } 5579 } 5580 NOT_PRODUCT( 5581 if (PrintCMSStatistics != 0) { 5582 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals); 5583 } 5584 ) 5585 assert(work_q->size() == 0 && _collector->overflow_list_is_empty(), 5586 "Else our work is not yet done"); 5587 } 5588 5589 // Record object boundaries in _eden_chunk_array by sampling the eden 5590 // top in the slow-path eden object allocation code path and record 5591 // the boundaries, if CMSEdenChunksRecordAlways is true. If 5592 // CMSEdenChunksRecordAlways is false, we use the other asynchronous 5593 // sampling in sample_eden() that activates during the part of the 5594 // preclean phase. 5595 void CMSCollector::sample_eden_chunk() { 5596 if (CMSEdenChunksRecordAlways && _eden_chunk_array != NULL) { 5597 if (_eden_chunk_lock->try_lock()) { 5598 // Record a sample. This is the critical section. The contents 5599 // of the _eden_chunk_array have to be non-decreasing in the 5600 // address order. 5601 _eden_chunk_array[_eden_chunk_index] = *_top_addr; 5602 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr, 5603 "Unexpected state of Eden"); 5604 if (_eden_chunk_index == 0 || 5605 ((_eden_chunk_array[_eden_chunk_index] > _eden_chunk_array[_eden_chunk_index-1]) && 5606 (pointer_delta(_eden_chunk_array[_eden_chunk_index], 5607 _eden_chunk_array[_eden_chunk_index-1]) >= CMSSamplingGrain))) { 5608 _eden_chunk_index++; // commit sample 5609 } 5610 _eden_chunk_lock->unlock(); 5611 } 5612 } 5613 } 5614 5615 // Return a thread-local PLAB recording array, as appropriate. 5616 void* CMSCollector::get_data_recorder(int thr_num) { 5617 if (_survivor_plab_array != NULL && 5618 (CMSPLABRecordAlways || 5619 (_collectorState > Marking && _collectorState < FinalMarking))) { 5620 assert(thr_num < (int)ParallelGCThreads, "thr_num is out of bounds"); 5621 ChunkArray* ca = &_survivor_plab_array[thr_num]; 5622 ca->reset(); // clear it so that fresh data is recorded 5623 return (void*) ca; 5624 } else { 5625 return NULL; 5626 } 5627 } 5628 5629 // Reset all the thread-local PLAB recording arrays 5630 void CMSCollector::reset_survivor_plab_arrays() { 5631 for (uint i = 0; i < ParallelGCThreads; i++) { 5632 _survivor_plab_array[i].reset(); 5633 } 5634 } 5635 5636 // Merge the per-thread plab arrays into the global survivor chunk 5637 // array which will provide the partitioning of the survivor space 5638 // for CMS initial scan and rescan. 5639 void CMSCollector::merge_survivor_plab_arrays(ContiguousSpace* surv, 5640 int no_of_gc_threads) { 5641 assert(_survivor_plab_array != NULL, "Error"); 5642 assert(_survivor_chunk_array != NULL, "Error"); 5643 assert(_collectorState == FinalMarking || 5644 (CMSParallelInitialMarkEnabled && _collectorState == InitialMarking), "Error"); 5645 for (int j = 0; j < no_of_gc_threads; j++) { 5646 _cursor[j] = 0; 5647 } 5648 HeapWord* top = surv->top(); 5649 size_t i; 5650 for (i = 0; i < _survivor_chunk_capacity; i++) { // all sca entries 5651 HeapWord* min_val = top; // Higher than any PLAB address 5652 uint min_tid = 0; // position of min_val this round 5653 for (int j = 0; j < no_of_gc_threads; j++) { 5654 ChunkArray* cur_sca = &_survivor_plab_array[j]; 5655 if (_cursor[j] == cur_sca->end()) { 5656 continue; 5657 } 5658 assert(_cursor[j] < cur_sca->end(), "ctl pt invariant"); 5659 HeapWord* cur_val = cur_sca->nth(_cursor[j]); 5660 assert(surv->used_region().contains(cur_val), "Out of bounds value"); 5661 if (cur_val < min_val) { 5662 min_tid = j; 5663 min_val = cur_val; 5664 } else { 5665 assert(cur_val < top, "All recorded addresses should be less"); 5666 } 5667 } 5668 // At this point min_val and min_tid are respectively 5669 // the least address in _survivor_plab_array[j]->nth(_cursor[j]) 5670 // and the thread (j) that witnesses that address. 5671 // We record this address in the _survivor_chunk_array[i] 5672 // and increment _cursor[min_tid] prior to the next round i. 5673 if (min_val == top) { 5674 break; 5675 } 5676 _survivor_chunk_array[i] = min_val; 5677 _cursor[min_tid]++; 5678 } 5679 // We are all done; record the size of the _survivor_chunk_array 5680 _survivor_chunk_index = i; // exclusive: [0, i) 5681 if (PrintCMSStatistics > 0) { 5682 gclog_or_tty->print(" (Survivor:" SIZE_FORMAT "chunks) ", i); 5683 } 5684 // Verify that we used up all the recorded entries 5685 #ifdef ASSERT 5686 size_t total = 0; 5687 for (int j = 0; j < no_of_gc_threads; j++) { 5688 assert(_cursor[j] == _survivor_plab_array[j].end(), "Ctl pt invariant"); 5689 total += _cursor[j]; 5690 } 5691 assert(total == _survivor_chunk_index, "Ctl Pt Invariant"); 5692 // Check that the merged array is in sorted order 5693 if (total > 0) { 5694 for (size_t i = 0; i < total - 1; i++) { 5695 if (PrintCMSStatistics > 0) { 5696 gclog_or_tty->print(" (chunk" SIZE_FORMAT ":" INTPTR_FORMAT ") ", 5697 i, _survivor_chunk_array[i]); 5698 } 5699 assert(_survivor_chunk_array[i] < _survivor_chunk_array[i+1], 5700 "Not sorted"); 5701 } 5702 } 5703 #endif // ASSERT 5704 } 5705 5706 // Set up the space's par_seq_tasks structure for work claiming 5707 // for parallel initial scan and rescan of young gen. 5708 // See ParRescanTask where this is currently used. 5709 void 5710 CMSCollector:: 5711 initialize_sequential_subtasks_for_young_gen_rescan(int n_threads) { 5712 assert(n_threads > 0, "Unexpected n_threads argument"); 5713 DefNewGeneration* dng = (DefNewGeneration*)_young_gen; 5714 5715 // Eden space 5716 if (!dng->eden()->is_empty()) { 5717 SequentialSubTasksDone* pst = dng->eden()->par_seq_tasks(); 5718 assert(!pst->valid(), "Clobbering existing data?"); 5719 // Each valid entry in [0, _eden_chunk_index) represents a task. 5720 size_t n_tasks = _eden_chunk_index + 1; 5721 assert(n_tasks == 1 || _eden_chunk_array != NULL, "Error"); 5722 // Sets the condition for completion of the subtask (how many threads 5723 // need to finish in order to be done). 5724 pst->set_n_threads(n_threads); 5725 pst->set_n_tasks((int)n_tasks); 5726 } 5727 5728 // Merge the survivor plab arrays into _survivor_chunk_array 5729 if (_survivor_plab_array != NULL) { 5730 merge_survivor_plab_arrays(dng->from(), n_threads); 5731 } else { 5732 assert(_survivor_chunk_index == 0, "Error"); 5733 } 5734 5735 // To space 5736 { 5737 SequentialSubTasksDone* pst = dng->to()->par_seq_tasks(); 5738 assert(!pst->valid(), "Clobbering existing data?"); 5739 // Sets the condition for completion of the subtask (how many threads 5740 // need to finish in order to be done). 5741 pst->set_n_threads(n_threads); 5742 pst->set_n_tasks(1); 5743 assert(pst->valid(), "Error"); 5744 } 5745 5746 // From space 5747 { 5748 SequentialSubTasksDone* pst = dng->from()->par_seq_tasks(); 5749 assert(!pst->valid(), "Clobbering existing data?"); 5750 size_t n_tasks = _survivor_chunk_index + 1; 5751 assert(n_tasks == 1 || _survivor_chunk_array != NULL, "Error"); 5752 // Sets the condition for completion of the subtask (how many threads 5753 // need to finish in order to be done). 5754 pst->set_n_threads(n_threads); 5755 pst->set_n_tasks((int)n_tasks); 5756 assert(pst->valid(), "Error"); 5757 } 5758 } 5759 5760 // Parallel version of remark 5761 void CMSCollector::do_remark_parallel() { 5762 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5763 FlexibleWorkGang* workers = gch->workers(); 5764 assert(workers != NULL, "Need parallel worker threads."); 5765 // Choose to use the number of GC workers most recently set 5766 // into "active_workers". If active_workers is not set, set it 5767 // to ParallelGCThreads. 5768 int n_workers = workers->active_workers(); 5769 if (n_workers == 0) { 5770 assert(n_workers > 0, "Should have been set during scavenge"); 5771 n_workers = ParallelGCThreads; 5772 workers->set_active_workers(n_workers); 5773 } 5774 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 5775 5776 CMSParRemarkTask tsk(this, 5777 cms_space, 5778 n_workers, workers, task_queues()); 5779 5780 // Set up for parallel process_roots work. 5781 gch->set_par_threads(n_workers); 5782 // We won't be iterating over the cards in the card table updating 5783 // the younger_gen cards, so we shouldn't call the following else 5784 // the verification code as well as subsequent younger_refs_iterate 5785 // code would get confused. XXX 5786 // gch->rem_set()->prepare_for_younger_refs_iterate(true); // parallel 5787 5788 // The young gen rescan work will not be done as part of 5789 // process_roots (which currently doesn't know how to 5790 // parallelize such a scan), but rather will be broken up into 5791 // a set of parallel tasks (via the sampling that the [abortable] 5792 // preclean phase did of EdenSpace, plus the [two] tasks of 5793 // scanning the [two] survivor spaces. Further fine-grain 5794 // parallelization of the scanning of the survivor spaces 5795 // themselves, and of precleaning of the younger gen itself 5796 // is deferred to the future. 5797 initialize_sequential_subtasks_for_young_gen_rescan(n_workers); 5798 5799 // The dirty card rescan work is broken up into a "sequence" 5800 // of parallel tasks (per constituent space) that are dynamically 5801 // claimed by the parallel threads. 5802 cms_space->initialize_sequential_subtasks_for_rescan(n_workers); 5803 5804 // It turns out that even when we're using 1 thread, doing the work in a 5805 // separate thread causes wide variance in run times. We can't help this 5806 // in the multi-threaded case, but we special-case n=1 here to get 5807 // repeatable measurements of the 1-thread overhead of the parallel code. 5808 if (n_workers > 1) { 5809 // Make refs discovery MT-safe, if it isn't already: it may not 5810 // necessarily be so, since it's possible that we are doing 5811 // ST marking. 5812 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), true); 5813 GenCollectedHeap::StrongRootsScope srs(gch); 5814 workers->run_task(&tsk); 5815 } else { 5816 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false); 5817 GenCollectedHeap::StrongRootsScope srs(gch); 5818 tsk.work(0); 5819 } 5820 5821 gch->set_par_threads(0); // 0 ==> non-parallel. 5822 // restore, single-threaded for now, any preserved marks 5823 // as a result of work_q overflow 5824 restore_preserved_marks_if_any(); 5825 } 5826 5827 // Non-parallel version of remark 5828 void CMSCollector::do_remark_non_parallel() { 5829 ResourceMark rm; 5830 HandleMark hm; 5831 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5832 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false); 5833 5834 MarkRefsIntoAndScanClosure 5835 mrias_cl(_span, ref_processor(), &_markBitMap, NULL /* not precleaning */, 5836 &_markStack, this, 5837 false /* should_yield */, false /* not precleaning */); 5838 MarkFromDirtyCardsClosure 5839 markFromDirtyCardsClosure(this, _span, 5840 NULL, // space is set further below 5841 &_markBitMap, &_markStack, &mrias_cl); 5842 { 5843 GCTraceTime t("grey object rescan", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 5844 // Iterate over the dirty cards, setting the corresponding bits in the 5845 // mod union table. 5846 { 5847 ModUnionClosure modUnionClosure(&_modUnionTable); 5848 _ct->ct_bs()->dirty_card_iterate( 5849 _cmsGen->used_region(), 5850 &modUnionClosure); 5851 } 5852 // Having transferred these marks into the modUnionTable, we just need 5853 // to rescan the marked objects on the dirty cards in the modUnionTable. 5854 // The initial marking may have been done during an asynchronous 5855 // collection so there may be dirty bits in the mod-union table. 5856 const int alignment = 5857 CardTableModRefBS::card_size * BitsPerWord; 5858 { 5859 // ... First handle dirty cards in CMS gen 5860 markFromDirtyCardsClosure.set_space(_cmsGen->cmsSpace()); 5861 MemRegion ur = _cmsGen->used_region(); 5862 HeapWord* lb = ur.start(); 5863 HeapWord* ub = (HeapWord*)round_to((intptr_t)ur.end(), alignment); 5864 MemRegion cms_span(lb, ub); 5865 _modUnionTable.dirty_range_iterate_clear(cms_span, 5866 &markFromDirtyCardsClosure); 5867 verify_work_stacks_empty(); 5868 if (PrintCMSStatistics != 0) { 5869 gclog_or_tty->print(" (re-scanned "SIZE_FORMAT" dirty cards in cms gen) ", 5870 markFromDirtyCardsClosure.num_dirty_cards()); 5871 } 5872 } 5873 } 5874 if (VerifyDuringGC && 5875 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 5876 HandleMark hm; // Discard invalid handles created during verification 5877 Universe::verify(); 5878 } 5879 { 5880 GCTraceTime t("root rescan", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 5881 5882 verify_work_stacks_empty(); 5883 5884 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 5885 GenCollectedHeap::StrongRootsScope srs(gch); 5886 5887 gch->gen_process_roots(_cmsGen->level(), 5888 true, // younger gens as roots 5889 false, // use the local StrongRootsScope 5890 SharedHeap::ScanningOption(roots_scanning_options()), 5891 should_unload_classes(), 5892 &mrias_cl, 5893 NULL, 5894 NULL); // The dirty klasses will be handled below 5895 5896 assert(should_unload_classes() 5897 || (roots_scanning_options() & SharedHeap::SO_AllCodeCache), 5898 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 5899 } 5900 5901 { 5902 GCTraceTime t("visit unhandled CLDs", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 5903 5904 verify_work_stacks_empty(); 5905 5906 // Scan all class loader data objects that might have been introduced 5907 // during concurrent marking. 5908 ResourceMark rm; 5909 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds(); 5910 for (int i = 0; i < array->length(); i++) { 5911 mrias_cl.do_class_loader_data(array->at(i)); 5912 } 5913 5914 // We don't need to keep track of new CLDs anymore. 5915 ClassLoaderDataGraph::remember_new_clds(false); 5916 5917 verify_work_stacks_empty(); 5918 } 5919 5920 { 5921 GCTraceTime t("dirty klass scan", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 5922 5923 verify_work_stacks_empty(); 5924 5925 RemarkKlassClosure remark_klass_closure(&mrias_cl); 5926 ClassLoaderDataGraph::classes_do(&remark_klass_closure); 5927 5928 verify_work_stacks_empty(); 5929 } 5930 5931 // We might have added oops to ClassLoaderData::_handles during the 5932 // concurrent marking phase. These oops point to newly allocated objects 5933 // that are guaranteed to be kept alive. Either by the direct allocation 5934 // code, or when the young collector processes the roots. Hence, 5935 // we don't have to revisit the _handles block during the remark phase. 5936 5937 verify_work_stacks_empty(); 5938 // Restore evacuated mark words, if any, used for overflow list links 5939 if (!CMSOverflowEarlyRestoration) { 5940 restore_preserved_marks_if_any(); 5941 } 5942 verify_overflow_empty(); 5943 } 5944 5945 //////////////////////////////////////////////////////// 5946 // Parallel Reference Processing Task Proxy Class 5947 //////////////////////////////////////////////////////// 5948 class CMSRefProcTaskProxy: public AbstractGangTaskWOopQueues { 5949 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask; 5950 CMSCollector* _collector; 5951 CMSBitMap* _mark_bit_map; 5952 const MemRegion _span; 5953 ProcessTask& _task; 5954 5955 public: 5956 CMSRefProcTaskProxy(ProcessTask& task, 5957 CMSCollector* collector, 5958 const MemRegion& span, 5959 CMSBitMap* mark_bit_map, 5960 AbstractWorkGang* workers, 5961 OopTaskQueueSet* task_queues): 5962 // XXX Should superclass AGTWOQ also know about AWG since it knows 5963 // about the task_queues used by the AWG? Then it could initialize 5964 // the terminator() object. See 6984287. The set_for_termination() 5965 // below is a temporary band-aid for the regression in 6984287. 5966 AbstractGangTaskWOopQueues("Process referents by policy in parallel", 5967 task_queues), 5968 _task(task), 5969 _collector(collector), _span(span), _mark_bit_map(mark_bit_map) 5970 { 5971 assert(_collector->_span.equals(_span) && !_span.is_empty(), 5972 "Inconsistency in _span"); 5973 set_for_termination(workers->active_workers()); 5974 } 5975 5976 OopTaskQueueSet* task_queues() { return queues(); } 5977 5978 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 5979 5980 void do_work_steal(int i, 5981 CMSParDrainMarkingStackClosure* drain, 5982 CMSParKeepAliveClosure* keep_alive, 5983 int* seed); 5984 5985 virtual void work(uint worker_id); 5986 }; 5987 5988 void CMSRefProcTaskProxy::work(uint worker_id) { 5989 ResourceMark rm; 5990 HandleMark hm; 5991 assert(_collector->_span.equals(_span), "Inconsistency in _span"); 5992 CMSParKeepAliveClosure par_keep_alive(_collector, _span, 5993 _mark_bit_map, 5994 work_queue(worker_id)); 5995 CMSParDrainMarkingStackClosure par_drain_stack(_collector, _span, 5996 _mark_bit_map, 5997 work_queue(worker_id)); 5998 CMSIsAliveClosure is_alive_closure(_span, _mark_bit_map); 5999 _task.work(worker_id, is_alive_closure, par_keep_alive, par_drain_stack); 6000 if (_task.marks_oops_alive()) { 6001 do_work_steal(worker_id, &par_drain_stack, &par_keep_alive, 6002 _collector->hash_seed(worker_id)); 6003 } 6004 assert(work_queue(worker_id)->size() == 0, "work_queue should be empty"); 6005 assert(_collector->_overflow_list == NULL, "non-empty _overflow_list"); 6006 } 6007 6008 class CMSRefEnqueueTaskProxy: public AbstractGangTask { 6009 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask; 6010 EnqueueTask& _task; 6011 6012 public: 6013 CMSRefEnqueueTaskProxy(EnqueueTask& task) 6014 : AbstractGangTask("Enqueue reference objects in parallel"), 6015 _task(task) 6016 { } 6017 6018 virtual void work(uint worker_id) 6019 { 6020 _task.work(worker_id); 6021 } 6022 }; 6023 6024 CMSParKeepAliveClosure::CMSParKeepAliveClosure(CMSCollector* collector, 6025 MemRegion span, CMSBitMap* bit_map, OopTaskQueue* work_queue): 6026 _span(span), 6027 _bit_map(bit_map), 6028 _work_queue(work_queue), 6029 _mark_and_push(collector, span, bit_map, work_queue), 6030 _low_water_mark(MIN2((uint)(work_queue->max_elems()/4), 6031 (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads))) 6032 { } 6033 6034 // . see if we can share work_queues with ParNew? XXX 6035 void CMSRefProcTaskProxy::do_work_steal(int i, 6036 CMSParDrainMarkingStackClosure* drain, 6037 CMSParKeepAliveClosure* keep_alive, 6038 int* seed) { 6039 OopTaskQueue* work_q = work_queue(i); 6040 NOT_PRODUCT(int num_steals = 0;) 6041 oop obj_to_scan; 6042 6043 while (true) { 6044 // Completely finish any left over work from (an) earlier round(s) 6045 drain->trim_queue(0); 6046 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 6047 (size_t)ParGCDesiredObjsFromOverflowList); 6048 // Now check if there's any work in the overflow list 6049 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads, 6050 // only affects the number of attempts made to get work from the 6051 // overflow list and does not affect the number of workers. Just 6052 // pass ParallelGCThreads so this behavior is unchanged. 6053 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 6054 work_q, 6055 ParallelGCThreads)) { 6056 // Found something in global overflow list; 6057 // not yet ready to go stealing work from others. 6058 // We'd like to assert(work_q->size() != 0, ...) 6059 // because we just took work from the overflow list, 6060 // but of course we can't, since all of that might have 6061 // been already stolen from us. 6062 continue; 6063 } 6064 // Verify that we have no work before we resort to stealing 6065 assert(work_q->size() == 0, "Have work, shouldn't steal"); 6066 // Try to steal from other queues that have work 6067 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 6068 NOT_PRODUCT(num_steals++;) 6069 assert(obj_to_scan->is_oop(), "Oops, not an oop!"); 6070 assert(_mark_bit_map->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 6071 // Do scanning work 6072 obj_to_scan->oop_iterate(keep_alive); 6073 // Loop around, finish this work, and try to steal some more 6074 } else if (terminator()->offer_termination()) { 6075 break; // nirvana from the infinite cycle 6076 } 6077 } 6078 NOT_PRODUCT( 6079 if (PrintCMSStatistics != 0) { 6080 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals); 6081 } 6082 ) 6083 } 6084 6085 void CMSRefProcTaskExecutor::execute(ProcessTask& task) 6086 { 6087 GenCollectedHeap* gch = GenCollectedHeap::heap(); 6088 FlexibleWorkGang* workers = gch->workers(); 6089 assert(workers != NULL, "Need parallel worker threads."); 6090 CMSRefProcTaskProxy rp_task(task, &_collector, 6091 _collector.ref_processor()->span(), 6092 _collector.markBitMap(), 6093 workers, _collector.task_queues()); 6094 workers->run_task(&rp_task); 6095 } 6096 6097 void CMSRefProcTaskExecutor::execute(EnqueueTask& task) 6098 { 6099 6100 GenCollectedHeap* gch = GenCollectedHeap::heap(); 6101 FlexibleWorkGang* workers = gch->workers(); 6102 assert(workers != NULL, "Need parallel worker threads."); 6103 CMSRefEnqueueTaskProxy enq_task(task); 6104 workers->run_task(&enq_task); 6105 } 6106 6107 void CMSCollector::refProcessingWork(bool asynch, bool clear_all_soft_refs) { 6108 6109 ResourceMark rm; 6110 HandleMark hm; 6111 6112 ReferenceProcessor* rp = ref_processor(); 6113 assert(rp->span().equals(_span), "Spans should be equal"); 6114 assert(!rp->enqueuing_is_done(), "Enqueuing should not be complete"); 6115 // Process weak references. 6116 rp->setup_policy(clear_all_soft_refs); 6117 verify_work_stacks_empty(); 6118 6119 CMSKeepAliveClosure cmsKeepAliveClosure(this, _span, &_markBitMap, 6120 &_markStack, false /* !preclean */); 6121 CMSDrainMarkingStackClosure cmsDrainMarkingStackClosure(this, 6122 _span, &_markBitMap, &_markStack, 6123 &cmsKeepAliveClosure, false /* !preclean */); 6124 { 6125 GCTraceTime t("weak refs processing", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 6126 6127 ReferenceProcessorStats stats; 6128 if (rp->processing_is_mt()) { 6129 // Set the degree of MT here. If the discovery is done MT, there 6130 // may have been a different number of threads doing the discovery 6131 // and a different number of discovered lists may have Ref objects. 6132 // That is OK as long as the Reference lists are balanced (see 6133 // balance_all_queues() and balance_queues()). 6134 GenCollectedHeap* gch = GenCollectedHeap::heap(); 6135 int active_workers = ParallelGCThreads; 6136 FlexibleWorkGang* workers = gch->workers(); 6137 if (workers != NULL) { 6138 active_workers = workers->active_workers(); 6139 // The expectation is that active_workers will have already 6140 // been set to a reasonable value. If it has not been set, 6141 // investigate. 6142 assert(active_workers > 0, "Should have been set during scavenge"); 6143 } 6144 rp->set_active_mt_degree(active_workers); 6145 CMSRefProcTaskExecutor task_executor(*this); 6146 stats = rp->process_discovered_references(&_is_alive_closure, 6147 &cmsKeepAliveClosure, 6148 &cmsDrainMarkingStackClosure, 6149 &task_executor, 6150 _gc_timer_cm, 6151 _gc_tracer_cm->gc_id()); 6152 } else { 6153 stats = rp->process_discovered_references(&_is_alive_closure, 6154 &cmsKeepAliveClosure, 6155 &cmsDrainMarkingStackClosure, 6156 NULL, 6157 _gc_timer_cm, 6158 _gc_tracer_cm->gc_id()); 6159 } 6160 _gc_tracer_cm->report_gc_reference_stats(stats); 6161 6162 } 6163 6164 // This is the point where the entire marking should have completed. 6165 verify_work_stacks_empty(); 6166 6167 if (should_unload_classes()) { 6168 { 6169 GCTraceTime t("class unloading", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 6170 6171 // Unload classes and purge the SystemDictionary. 6172 bool purged_class = SystemDictionary::do_unloading(&_is_alive_closure); 6173 6174 // Unload nmethods. 6175 CodeCache::do_unloading(&_is_alive_closure, purged_class); 6176 6177 // Prune dead klasses from subklass/sibling/implementor lists. 6178 Klass::clean_weak_klass_links(&_is_alive_closure); 6179 } 6180 6181 { 6182 GCTraceTime t("scrub symbol table", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 6183 // Clean up unreferenced symbols in symbol table. 6184 SymbolTable::unlink(); 6185 } 6186 6187 { 6188 GCTraceTime t("scrub string table", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 6189 // Delete entries for dead interned strings. 6190 StringTable::unlink(&_is_alive_closure); 6191 } 6192 } 6193 6194 6195 // Restore any preserved marks as a result of mark stack or 6196 // work queue overflow 6197 restore_preserved_marks_if_any(); // done single-threaded for now 6198 6199 rp->set_enqueuing_is_done(true); 6200 if (rp->processing_is_mt()) { 6201 rp->balance_all_queues(); 6202 CMSRefProcTaskExecutor task_executor(*this); 6203 rp->enqueue_discovered_references(&task_executor); 6204 } else { 6205 rp->enqueue_discovered_references(NULL); 6206 } 6207 rp->verify_no_references_recorded(); 6208 assert(!rp->discovery_enabled(), "should have been disabled"); 6209 } 6210 6211 #ifndef PRODUCT 6212 void CMSCollector::check_correct_thread_executing() { 6213 Thread* t = Thread::current(); 6214 // Only the VM thread or the CMS thread should be here. 6215 assert(t->is_ConcurrentGC_thread() || t->is_VM_thread(), 6216 "Unexpected thread type"); 6217 // If this is the vm thread, the foreground process 6218 // should not be waiting. Note that _foregroundGCIsActive is 6219 // true while the foreground collector is waiting. 6220 if (_foregroundGCShouldWait) { 6221 // We cannot be the VM thread 6222 assert(t->is_ConcurrentGC_thread(), 6223 "Should be CMS thread"); 6224 } else { 6225 // We can be the CMS thread only if we are in a stop-world 6226 // phase of CMS collection. 6227 if (t->is_ConcurrentGC_thread()) { 6228 assert(_collectorState == InitialMarking || 6229 _collectorState == FinalMarking, 6230 "Should be a stop-world phase"); 6231 // The CMS thread should be holding the CMS_token. 6232 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6233 "Potential interference with concurrently " 6234 "executing VM thread"); 6235 } 6236 } 6237 } 6238 #endif 6239 6240 void CMSCollector::sweep(bool asynch) { 6241 assert(_collectorState == Sweeping, "just checking"); 6242 check_correct_thread_executing(); 6243 verify_work_stacks_empty(); 6244 verify_overflow_empty(); 6245 increment_sweep_count(); 6246 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 6247 6248 _inter_sweep_timer.stop(); 6249 _inter_sweep_estimate.sample(_inter_sweep_timer.seconds()); 6250 6251 assert(!_intra_sweep_timer.is_active(), "Should not be active"); 6252 _intra_sweep_timer.reset(); 6253 _intra_sweep_timer.start(); 6254 if (asynch) { 6255 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 6256 CMSPhaseAccounting pa(this, "sweep", _gc_tracer_cm->gc_id(), !PrintGCDetails); 6257 // First sweep the old gen 6258 { 6259 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(), 6260 bitMapLock()); 6261 sweepWork(_cmsGen, asynch); 6262 } 6263 6264 // Update Universe::_heap_*_at_gc figures. 6265 // We need all the free list locks to make the abstract state 6266 // transition from Sweeping to Resetting. See detailed note 6267 // further below. 6268 { 6269 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock()); 6270 // Update heap occupancy information which is used as 6271 // input to soft ref clearing policy at the next gc. 6272 Universe::update_heap_info_at_gc(); 6273 _collectorState = Resizing; 6274 } 6275 } else { 6276 // already have needed locks 6277 sweepWork(_cmsGen, asynch); 6278 // Update heap occupancy information which is used as 6279 // input to soft ref clearing policy at the next gc. 6280 Universe::update_heap_info_at_gc(); 6281 _collectorState = Resizing; 6282 } 6283 verify_work_stacks_empty(); 6284 verify_overflow_empty(); 6285 6286 if (should_unload_classes()) { 6287 // Delay purge to the beginning of the next safepoint. Metaspace::contains 6288 // requires that the virtual spaces are stable and not deleted. 6289 ClassLoaderDataGraph::set_should_purge(true); 6290 } 6291 6292 _intra_sweep_timer.stop(); 6293 _intra_sweep_estimate.sample(_intra_sweep_timer.seconds()); 6294 6295 _inter_sweep_timer.reset(); 6296 _inter_sweep_timer.start(); 6297 6298 // We need to use a monotonically non-decreasing time in ms 6299 // or we will see time-warp warnings and os::javaTimeMillis() 6300 // does not guarantee monotonicity. 6301 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC; 6302 update_time_of_last_gc(now); 6303 6304 // NOTE on abstract state transitions: 6305 // Mutators allocate-live and/or mark the mod-union table dirty 6306 // based on the state of the collection. The former is done in 6307 // the interval [Marking, Sweeping] and the latter in the interval 6308 // [Marking, Sweeping). Thus the transitions into the Marking state 6309 // and out of the Sweeping state must be synchronously visible 6310 // globally to the mutators. 6311 // The transition into the Marking state happens with the world 6312 // stopped so the mutators will globally see it. Sweeping is 6313 // done asynchronously by the background collector so the transition 6314 // from the Sweeping state to the Resizing state must be done 6315 // under the freelistLock (as is the check for whether to 6316 // allocate-live and whether to dirty the mod-union table). 6317 assert(_collectorState == Resizing, "Change of collector state to" 6318 " Resizing must be done under the freelistLocks (plural)"); 6319 6320 // Now that sweeping has been completed, we clear 6321 // the incremental_collection_failed flag, 6322 // thus inviting a younger gen collection to promote into 6323 // this generation. If such a promotion may still fail, 6324 // the flag will be set again when a young collection is 6325 // attempted. 6326 GenCollectedHeap* gch = GenCollectedHeap::heap(); 6327 gch->clear_incremental_collection_failed(); // Worth retrying as fresh space may have been freed up 6328 gch->update_full_collections_completed(_collection_count_start); 6329 } 6330 6331 // FIX ME!!! Looks like this belongs in CFLSpace, with 6332 // CMSGen merely delegating to it. 6333 void ConcurrentMarkSweepGeneration::setNearLargestChunk() { 6334 double nearLargestPercent = FLSLargestBlockCoalesceProximity; 6335 HeapWord* minAddr = _cmsSpace->bottom(); 6336 HeapWord* largestAddr = 6337 (HeapWord*) _cmsSpace->dictionary()->find_largest_dict(); 6338 if (largestAddr == NULL) { 6339 // The dictionary appears to be empty. In this case 6340 // try to coalesce at the end of the heap. 6341 largestAddr = _cmsSpace->end(); 6342 } 6343 size_t largestOffset = pointer_delta(largestAddr, minAddr); 6344 size_t nearLargestOffset = 6345 (size_t)((double)largestOffset * nearLargestPercent) - MinChunkSize; 6346 if (PrintFLSStatistics != 0) { 6347 gclog_or_tty->print_cr( 6348 "CMS: Large Block: " PTR_FORMAT ";" 6349 " Proximity: " PTR_FORMAT " -> " PTR_FORMAT, 6350 largestAddr, 6351 _cmsSpace->nearLargestChunk(), minAddr + nearLargestOffset); 6352 } 6353 _cmsSpace->set_nearLargestChunk(minAddr + nearLargestOffset); 6354 } 6355 6356 bool ConcurrentMarkSweepGeneration::isNearLargestChunk(HeapWord* addr) { 6357 return addr >= _cmsSpace->nearLargestChunk(); 6358 } 6359 6360 FreeChunk* ConcurrentMarkSweepGeneration::find_chunk_at_end() { 6361 return _cmsSpace->find_chunk_at_end(); 6362 } 6363 6364 void ConcurrentMarkSweepGeneration::update_gc_stats(int current_level, 6365 bool full) { 6366 // The next lower level has been collected. Gather any statistics 6367 // that are of interest at this point. 6368 if (!full && (current_level + 1) == level()) { 6369 // Gather statistics on the young generation collection. 6370 collector()->stats().record_gc0_end(used()); 6371 } 6372 } 6373 6374 void ConcurrentMarkSweepGeneration::rotate_debug_collection_type() { 6375 if (PrintGCDetails && Verbose) { 6376 gclog_or_tty->print("Rotate from %d ", _debug_collection_type); 6377 } 6378 _debug_collection_type = (CollectionTypes) (_debug_collection_type + 1); 6379 _debug_collection_type = 6380 (CollectionTypes) (_debug_collection_type % Unknown_collection_type); 6381 if (PrintGCDetails && Verbose) { 6382 gclog_or_tty->print_cr("to %d ", _debug_collection_type); 6383 } 6384 } 6385 6386 void CMSCollector::sweepWork(ConcurrentMarkSweepGeneration* gen, 6387 bool asynch) { 6388 // We iterate over the space(s) underlying this generation, 6389 // checking the mark bit map to see if the bits corresponding 6390 // to specific blocks are marked or not. Blocks that are 6391 // marked are live and are not swept up. All remaining blocks 6392 // are swept up, with coalescing on-the-fly as we sweep up 6393 // contiguous free and/or garbage blocks: 6394 // We need to ensure that the sweeper synchronizes with allocators 6395 // and stop-the-world collectors. In particular, the following 6396 // locks are used: 6397 // . CMS token: if this is held, a stop the world collection cannot occur 6398 // . freelistLock: if this is held no allocation can occur from this 6399 // generation by another thread 6400 // . bitMapLock: if this is held, no other thread can access or update 6401 // 6402 6403 // Note that we need to hold the freelistLock if we use 6404 // block iterate below; else the iterator might go awry if 6405 // a mutator (or promotion) causes block contents to change 6406 // (for instance if the allocator divvies up a block). 6407 // If we hold the free list lock, for all practical purposes 6408 // young generation GC's can't occur (they'll usually need to 6409 // promote), so we might as well prevent all young generation 6410 // GC's while we do a sweeping step. For the same reason, we might 6411 // as well take the bit map lock for the entire duration 6412 6413 // check that we hold the requisite locks 6414 assert(have_cms_token(), "Should hold cms token"); 6415 assert( (asynch && ConcurrentMarkSweepThread::cms_thread_has_cms_token()) 6416 || (!asynch && ConcurrentMarkSweepThread::vm_thread_has_cms_token()), 6417 "Should possess CMS token to sweep"); 6418 assert_lock_strong(gen->freelistLock()); 6419 assert_lock_strong(bitMapLock()); 6420 6421 assert(!_inter_sweep_timer.is_active(), "Was switched off in an outer context"); 6422 assert(_intra_sweep_timer.is_active(), "Was switched on in an outer context"); 6423 gen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()), 6424 _inter_sweep_estimate.padded_average(), 6425 _intra_sweep_estimate.padded_average()); 6426 gen->setNearLargestChunk(); 6427 6428 { 6429 SweepClosure sweepClosure(this, gen, &_markBitMap, 6430 CMSYield && asynch); 6431 gen->cmsSpace()->blk_iterate_careful(&sweepClosure); 6432 // We need to free-up/coalesce garbage/blocks from a 6433 // co-terminal free run. This is done in the SweepClosure 6434 // destructor; so, do not remove this scope, else the 6435 // end-of-sweep-census below will be off by a little bit. 6436 } 6437 gen->cmsSpace()->sweep_completed(); 6438 gen->cmsSpace()->endSweepFLCensus(sweep_count()); 6439 if (should_unload_classes()) { // unloaded classes this cycle, 6440 _concurrent_cycles_since_last_unload = 0; // ... reset count 6441 } else { // did not unload classes, 6442 _concurrent_cycles_since_last_unload++; // ... increment count 6443 } 6444 } 6445 6446 // Reset CMS data structures (for now just the marking bit map) 6447 // preparatory for the next cycle. 6448 void CMSCollector::reset(bool asynch) { 6449 if (asynch) { 6450 CMSTokenSyncWithLocks ts(true, bitMapLock()); 6451 6452 // If the state is not "Resetting", the foreground thread 6453 // has done a collection and the resetting. 6454 if (_collectorState != Resetting) { 6455 assert(_collectorState == Idling, "The state should only change" 6456 " because the foreground collector has finished the collection"); 6457 return; 6458 } 6459 6460 // Clear the mark bitmap (no grey objects to start with) 6461 // for the next cycle. 6462 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 6463 CMSPhaseAccounting cmspa(this, "reset", _gc_tracer_cm->gc_id(), !PrintGCDetails); 6464 6465 HeapWord* curAddr = _markBitMap.startWord(); 6466 while (curAddr < _markBitMap.endWord()) { 6467 size_t remaining = pointer_delta(_markBitMap.endWord(), curAddr); 6468 MemRegion chunk(curAddr, MIN2(CMSBitMapYieldQuantum, remaining)); 6469 _markBitMap.clear_large_range(chunk); 6470 if (ConcurrentMarkSweepThread::should_yield() && 6471 !foregroundGCIsActive() && 6472 CMSYield) { 6473 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6474 "CMS thread should hold CMS token"); 6475 assert_lock_strong(bitMapLock()); 6476 bitMapLock()->unlock(); 6477 ConcurrentMarkSweepThread::desynchronize(true); 6478 ConcurrentMarkSweepThread::acknowledge_yield_request(); 6479 stopTimer(); 6480 if (PrintCMSStatistics != 0) { 6481 incrementYields(); 6482 } 6483 icms_wait(); 6484 6485 // See the comment in coordinator_yield() 6486 for (unsigned i = 0; i < CMSYieldSleepCount && 6487 ConcurrentMarkSweepThread::should_yield() && 6488 !CMSCollector::foregroundGCIsActive(); ++i) { 6489 os::sleep(Thread::current(), 1, false); 6490 ConcurrentMarkSweepThread::acknowledge_yield_request(); 6491 } 6492 6493 ConcurrentMarkSweepThread::synchronize(true); 6494 bitMapLock()->lock_without_safepoint_check(); 6495 startTimer(); 6496 } 6497 curAddr = chunk.end(); 6498 } 6499 // A successful mostly concurrent collection has been done. 6500 // Because only the full (i.e., concurrent mode failure) collections 6501 // are being measured for gc overhead limits, clean the "near" flag 6502 // and count. 6503 size_policy()->reset_gc_overhead_limit_count(); 6504 _collectorState = Idling; 6505 } else { 6506 // already have the lock 6507 assert(_collectorState == Resetting, "just checking"); 6508 assert_lock_strong(bitMapLock()); 6509 _markBitMap.clear_all(); 6510 _collectorState = Idling; 6511 } 6512 6513 // Stop incremental mode after a cycle completes, so that any future cycles 6514 // are triggered by allocation. 6515 stop_icms(); 6516 6517 NOT_PRODUCT( 6518 if (RotateCMSCollectionTypes) { 6519 _cmsGen->rotate_debug_collection_type(); 6520 } 6521 ) 6522 6523 register_gc_end(); 6524 } 6525 6526 void CMSCollector::do_CMS_operation(CMS_op_type op, GCCause::Cause gc_cause) { 6527 gclog_or_tty->date_stamp(PrintGC && PrintGCDateStamps); 6528 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 6529 GCTraceTime t(GCCauseString("GC", gc_cause), PrintGC, !PrintGCDetails, NULL, _gc_tracer_cm->gc_id()); 6530 TraceCollectorStats tcs(counters()); 6531 6532 switch (op) { 6533 case CMS_op_checkpointRootsInitial: { 6534 SvcGCMarker sgcm(SvcGCMarker::OTHER); 6535 checkpointRootsInitial(true); // asynch 6536 if (PrintGC) { 6537 _cmsGen->printOccupancy("initial-mark"); 6538 } 6539 break; 6540 } 6541 case CMS_op_checkpointRootsFinal: { 6542 SvcGCMarker sgcm(SvcGCMarker::OTHER); 6543 checkpointRootsFinal(true, // asynch 6544 false, // !clear_all_soft_refs 6545 false); // !init_mark_was_synchronous 6546 if (PrintGC) { 6547 _cmsGen->printOccupancy("remark"); 6548 } 6549 break; 6550 } 6551 default: 6552 fatal("No such CMS_op"); 6553 } 6554 } 6555 6556 #ifndef PRODUCT 6557 size_t const CMSCollector::skip_header_HeapWords() { 6558 return FreeChunk::header_size(); 6559 } 6560 6561 // Try and collect here conditions that should hold when 6562 // CMS thread is exiting. The idea is that the foreground GC 6563 // thread should not be blocked if it wants to terminate 6564 // the CMS thread and yet continue to run the VM for a while 6565 // after that. 6566 void CMSCollector::verify_ok_to_terminate() const { 6567 assert(Thread::current()->is_ConcurrentGC_thread(), 6568 "should be called by CMS thread"); 6569 assert(!_foregroundGCShouldWait, "should be false"); 6570 // We could check here that all the various low-level locks 6571 // are not held by the CMS thread, but that is overkill; see 6572 // also CMSThread::verify_ok_to_terminate() where the CGC_lock 6573 // is checked. 6574 } 6575 #endif 6576 6577 size_t CMSCollector::block_size_using_printezis_bits(HeapWord* addr) const { 6578 assert(_markBitMap.isMarked(addr) && _markBitMap.isMarked(addr + 1), 6579 "missing Printezis mark?"); 6580 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 6581 size_t size = pointer_delta(nextOneAddr + 1, addr); 6582 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 6583 "alignment problem"); 6584 assert(size >= 3, "Necessary for Printezis marks to work"); 6585 return size; 6586 } 6587 6588 // A variant of the above (block_size_using_printezis_bits()) except 6589 // that we return 0 if the P-bits are not yet set. 6590 size_t CMSCollector::block_size_if_printezis_bits(HeapWord* addr) const { 6591 if (_markBitMap.isMarked(addr + 1)) { 6592 assert(_markBitMap.isMarked(addr), "P-bit can be set only for marked objects"); 6593 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 6594 size_t size = pointer_delta(nextOneAddr + 1, addr); 6595 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 6596 "alignment problem"); 6597 assert(size >= 3, "Necessary for Printezis marks to work"); 6598 return size; 6599 } 6600 return 0; 6601 } 6602 6603 HeapWord* CMSCollector::next_card_start_after_block(HeapWord* addr) const { 6604 size_t sz = 0; 6605 oop p = (oop)addr; 6606 if (p->klass_or_null() != NULL) { 6607 sz = CompactibleFreeListSpace::adjustObjectSize(p->size()); 6608 } else { 6609 sz = block_size_using_printezis_bits(addr); 6610 } 6611 assert(sz > 0, "size must be nonzero"); 6612 HeapWord* next_block = addr + sz; 6613 HeapWord* next_card = (HeapWord*)round_to((uintptr_t)next_block, 6614 CardTableModRefBS::card_size); 6615 assert(round_down((uintptr_t)addr, CardTableModRefBS::card_size) < 6616 round_down((uintptr_t)next_card, CardTableModRefBS::card_size), 6617 "must be different cards"); 6618 return next_card; 6619 } 6620 6621 6622 // CMS Bit Map Wrapper ///////////////////////////////////////// 6623 6624 // Construct a CMS bit map infrastructure, but don't create the 6625 // bit vector itself. That is done by a separate call CMSBitMap::allocate() 6626 // further below. 6627 CMSBitMap::CMSBitMap(int shifter, int mutex_rank, const char* mutex_name): 6628 _bm(), 6629 _shifter(shifter), 6630 _lock(mutex_rank >= 0 ? new Mutex(mutex_rank, mutex_name, true) : NULL) 6631 { 6632 _bmStartWord = 0; 6633 _bmWordSize = 0; 6634 } 6635 6636 bool CMSBitMap::allocate(MemRegion mr) { 6637 _bmStartWord = mr.start(); 6638 _bmWordSize = mr.word_size(); 6639 ReservedSpace brs(ReservedSpace::allocation_align_size_up( 6640 (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1)); 6641 if (!brs.is_reserved()) { 6642 warning("CMS bit map allocation failure"); 6643 return false; 6644 } 6645 // For now we'll just commit all of the bit map up front. 6646 // Later on we'll try to be more parsimonious with swap. 6647 if (!_virtual_space.initialize(brs, brs.size())) { 6648 warning("CMS bit map backing store failure"); 6649 return false; 6650 } 6651 assert(_virtual_space.committed_size() == brs.size(), 6652 "didn't reserve backing store for all of CMS bit map?"); 6653 _bm.set_map((BitMap::bm_word_t*)_virtual_space.low()); 6654 assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >= 6655 _bmWordSize, "inconsistency in bit map sizing"); 6656 _bm.set_size(_bmWordSize >> _shifter); 6657 6658 // bm.clear(); // can we rely on getting zero'd memory? verify below 6659 assert(isAllClear(), 6660 "Expected zero'd memory from ReservedSpace constructor"); 6661 assert(_bm.size() == heapWordDiffToOffsetDiff(sizeInWords()), 6662 "consistency check"); 6663 return true; 6664 } 6665 6666 void CMSBitMap::dirty_range_iterate_clear(MemRegion mr, MemRegionClosure* cl) { 6667 HeapWord *next_addr, *end_addr, *last_addr; 6668 assert_locked(); 6669 assert(covers(mr), "out-of-range error"); 6670 // XXX assert that start and end are appropriately aligned 6671 for (next_addr = mr.start(), end_addr = mr.end(); 6672 next_addr < end_addr; next_addr = last_addr) { 6673 MemRegion dirty_region = getAndClearMarkedRegion(next_addr, end_addr); 6674 last_addr = dirty_region.end(); 6675 if (!dirty_region.is_empty()) { 6676 cl->do_MemRegion(dirty_region); 6677 } else { 6678 assert(last_addr == end_addr, "program logic"); 6679 return; 6680 } 6681 } 6682 } 6683 6684 void CMSBitMap::print_on_error(outputStream* st, const char* prefix) const { 6685 _bm.print_on_error(st, prefix); 6686 } 6687 6688 #ifndef PRODUCT 6689 void CMSBitMap::assert_locked() const { 6690 CMSLockVerifier::assert_locked(lock()); 6691 } 6692 6693 bool CMSBitMap::covers(MemRegion mr) const { 6694 // assert(_bm.map() == _virtual_space.low(), "map inconsistency"); 6695 assert((size_t)_bm.size() == (_bmWordSize >> _shifter), 6696 "size inconsistency"); 6697 return (mr.start() >= _bmStartWord) && 6698 (mr.end() <= endWord()); 6699 } 6700 6701 bool CMSBitMap::covers(HeapWord* start, size_t size) const { 6702 return (start >= _bmStartWord && (start + size) <= endWord()); 6703 } 6704 6705 void CMSBitMap::verifyNoOneBitsInRange(HeapWord* left, HeapWord* right) { 6706 // verify that there are no 1 bits in the interval [left, right) 6707 FalseBitMapClosure falseBitMapClosure; 6708 iterate(&falseBitMapClosure, left, right); 6709 } 6710 6711 void CMSBitMap::region_invariant(MemRegion mr) 6712 { 6713 assert_locked(); 6714 // mr = mr.intersection(MemRegion(_bmStartWord, _bmWordSize)); 6715 assert(!mr.is_empty(), "unexpected empty region"); 6716 assert(covers(mr), "mr should be covered by bit map"); 6717 // convert address range into offset range 6718 size_t start_ofs = heapWordToOffset(mr.start()); 6719 // Make sure that end() is appropriately aligned 6720 assert(mr.end() == (HeapWord*)round_to((intptr_t)mr.end(), 6721 (1 << (_shifter+LogHeapWordSize))), 6722 "Misaligned mr.end()"); 6723 size_t end_ofs = heapWordToOffset(mr.end()); 6724 assert(end_ofs > start_ofs, "Should mark at least one bit"); 6725 } 6726 6727 #endif 6728 6729 bool CMSMarkStack::allocate(size_t size) { 6730 // allocate a stack of the requisite depth 6731 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 6732 size * sizeof(oop))); 6733 if (!rs.is_reserved()) { 6734 warning("CMSMarkStack allocation failure"); 6735 return false; 6736 } 6737 if (!_virtual_space.initialize(rs, rs.size())) { 6738 warning("CMSMarkStack backing store failure"); 6739 return false; 6740 } 6741 assert(_virtual_space.committed_size() == rs.size(), 6742 "didn't reserve backing store for all of CMS stack?"); 6743 _base = (oop*)(_virtual_space.low()); 6744 _index = 0; 6745 _capacity = size; 6746 NOT_PRODUCT(_max_depth = 0); 6747 return true; 6748 } 6749 6750 // XXX FIX ME !!! In the MT case we come in here holding a 6751 // leaf lock. For printing we need to take a further lock 6752 // which has lower rank. We need to recalibrate the two 6753 // lock-ranks involved in order to be able to print the 6754 // messages below. (Or defer the printing to the caller. 6755 // For now we take the expedient path of just disabling the 6756 // messages for the problematic case.) 6757 void CMSMarkStack::expand() { 6758 assert(_capacity <= MarkStackSizeMax, "stack bigger than permitted"); 6759 if (_capacity == MarkStackSizeMax) { 6760 if (_hit_limit++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) { 6761 // We print a warning message only once per CMS cycle. 6762 gclog_or_tty->print_cr(" (benign) Hit CMSMarkStack max size limit"); 6763 } 6764 return; 6765 } 6766 // Double capacity if possible 6767 size_t new_capacity = MIN2(_capacity*2, MarkStackSizeMax); 6768 // Do not give up existing stack until we have managed to 6769 // get the double capacity that we desired. 6770 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 6771 new_capacity * sizeof(oop))); 6772 if (rs.is_reserved()) { 6773 // Release the backing store associated with old stack 6774 _virtual_space.release(); 6775 // Reinitialize virtual space for new stack 6776 if (!_virtual_space.initialize(rs, rs.size())) { 6777 fatal("Not enough swap for expanded marking stack"); 6778 } 6779 _base = (oop*)(_virtual_space.low()); 6780 _index = 0; 6781 _capacity = new_capacity; 6782 } else if (_failed_double++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) { 6783 // Failed to double capacity, continue; 6784 // we print a detail message only once per CMS cycle. 6785 gclog_or_tty->print(" (benign) Failed to expand marking stack from "SIZE_FORMAT"K to " 6786 SIZE_FORMAT"K", 6787 _capacity / K, new_capacity / K); 6788 } 6789 } 6790 6791 6792 // Closures 6793 // XXX: there seems to be a lot of code duplication here; 6794 // should refactor and consolidate common code. 6795 6796 // This closure is used to mark refs into the CMS generation in 6797 // the CMS bit map. Called at the first checkpoint. This closure 6798 // assumes that we do not need to re-mark dirty cards; if the CMS 6799 // generation on which this is used is not an oldest 6800 // generation then this will lose younger_gen cards! 6801 6802 MarkRefsIntoClosure::MarkRefsIntoClosure( 6803 MemRegion span, CMSBitMap* bitMap): 6804 _span(span), 6805 _bitMap(bitMap) 6806 { 6807 assert(_ref_processor == NULL, "deliberately left NULL"); 6808 assert(_bitMap->covers(_span), "_bitMap/_span mismatch"); 6809 } 6810 6811 void MarkRefsIntoClosure::do_oop(oop obj) { 6812 // if p points into _span, then mark corresponding bit in _markBitMap 6813 assert(obj->is_oop(), "expected an oop"); 6814 HeapWord* addr = (HeapWord*)obj; 6815 if (_span.contains(addr)) { 6816 // this should be made more efficient 6817 _bitMap->mark(addr); 6818 } 6819 } 6820 6821 void MarkRefsIntoClosure::do_oop(oop* p) { MarkRefsIntoClosure::do_oop_work(p); } 6822 void MarkRefsIntoClosure::do_oop(narrowOop* p) { MarkRefsIntoClosure::do_oop_work(p); } 6823 6824 Par_MarkRefsIntoClosure::Par_MarkRefsIntoClosure( 6825 MemRegion span, CMSBitMap* bitMap): 6826 _span(span), 6827 _bitMap(bitMap) 6828 { 6829 assert(_ref_processor == NULL, "deliberately left NULL"); 6830 assert(_bitMap->covers(_span), "_bitMap/_span mismatch"); 6831 } 6832 6833 void Par_MarkRefsIntoClosure::do_oop(oop obj) { 6834 // if p points into _span, then mark corresponding bit in _markBitMap 6835 assert(obj->is_oop(), "expected an oop"); 6836 HeapWord* addr = (HeapWord*)obj; 6837 if (_span.contains(addr)) { 6838 // this should be made more efficient 6839 _bitMap->par_mark(addr); 6840 } 6841 } 6842 6843 void Par_MarkRefsIntoClosure::do_oop(oop* p) { Par_MarkRefsIntoClosure::do_oop_work(p); } 6844 void Par_MarkRefsIntoClosure::do_oop(narrowOop* p) { Par_MarkRefsIntoClosure::do_oop_work(p); } 6845 6846 // A variant of the above, used for CMS marking verification. 6847 MarkRefsIntoVerifyClosure::MarkRefsIntoVerifyClosure( 6848 MemRegion span, CMSBitMap* verification_bm, CMSBitMap* cms_bm): 6849 _span(span), 6850 _verification_bm(verification_bm), 6851 _cms_bm(cms_bm) 6852 { 6853 assert(_ref_processor == NULL, "deliberately left NULL"); 6854 assert(_verification_bm->covers(_span), "_verification_bm/_span mismatch"); 6855 } 6856 6857 void MarkRefsIntoVerifyClosure::do_oop(oop obj) { 6858 // if p points into _span, then mark corresponding bit in _markBitMap 6859 assert(obj->is_oop(), "expected an oop"); 6860 HeapWord* addr = (HeapWord*)obj; 6861 if (_span.contains(addr)) { 6862 _verification_bm->mark(addr); 6863 if (!_cms_bm->isMarked(addr)) { 6864 oop(addr)->print(); 6865 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)", addr); 6866 fatal("... aborting"); 6867 } 6868 } 6869 } 6870 6871 void MarkRefsIntoVerifyClosure::do_oop(oop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); } 6872 void MarkRefsIntoVerifyClosure::do_oop(narrowOop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); } 6873 6874 ////////////////////////////////////////////////// 6875 // MarkRefsIntoAndScanClosure 6876 ////////////////////////////////////////////////// 6877 6878 MarkRefsIntoAndScanClosure::MarkRefsIntoAndScanClosure(MemRegion span, 6879 ReferenceProcessor* rp, 6880 CMSBitMap* bit_map, 6881 CMSBitMap* mod_union_table, 6882 CMSMarkStack* mark_stack, 6883 CMSCollector* collector, 6884 bool should_yield, 6885 bool concurrent_precleaning): 6886 _collector(collector), 6887 _span(span), 6888 _bit_map(bit_map), 6889 _mark_stack(mark_stack), 6890 _pushAndMarkClosure(collector, span, rp, bit_map, mod_union_table, 6891 mark_stack, concurrent_precleaning), 6892 _yield(should_yield), 6893 _concurrent_precleaning(concurrent_precleaning), 6894 _freelistLock(NULL) 6895 { 6896 _ref_processor = rp; 6897 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 6898 } 6899 6900 // This closure is used to mark refs into the CMS generation at the 6901 // second (final) checkpoint, and to scan and transitively follow 6902 // the unmarked oops. It is also used during the concurrent precleaning 6903 // phase while scanning objects on dirty cards in the CMS generation. 6904 // The marks are made in the marking bit map and the marking stack is 6905 // used for keeping the (newly) grey objects during the scan. 6906 // The parallel version (Par_...) appears further below. 6907 void MarkRefsIntoAndScanClosure::do_oop(oop obj) { 6908 if (obj != NULL) { 6909 assert(obj->is_oop(), "expected an oop"); 6910 HeapWord* addr = (HeapWord*)obj; 6911 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 6912 assert(_collector->overflow_list_is_empty(), 6913 "overflow list should be empty"); 6914 if (_span.contains(addr) && 6915 !_bit_map->isMarked(addr)) { 6916 // mark bit map (object is now grey) 6917 _bit_map->mark(addr); 6918 // push on marking stack (stack should be empty), and drain the 6919 // stack by applying this closure to the oops in the oops popped 6920 // from the stack (i.e. blacken the grey objects) 6921 bool res = _mark_stack->push(obj); 6922 assert(res, "Should have space to push on empty stack"); 6923 do { 6924 oop new_oop = _mark_stack->pop(); 6925 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 6926 assert(_bit_map->isMarked((HeapWord*)new_oop), 6927 "only grey objects on this stack"); 6928 // iterate over the oops in this oop, marking and pushing 6929 // the ones in CMS heap (i.e. in _span). 6930 new_oop->oop_iterate(&_pushAndMarkClosure); 6931 // check if it's time to yield 6932 do_yield_check(); 6933 } while (!_mark_stack->isEmpty() || 6934 (!_concurrent_precleaning && take_from_overflow_list())); 6935 // if marking stack is empty, and we are not doing this 6936 // during precleaning, then check the overflow list 6937 } 6938 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 6939 assert(_collector->overflow_list_is_empty(), 6940 "overflow list was drained above"); 6941 // We could restore evacuated mark words, if any, used for 6942 // overflow list links here because the overflow list is 6943 // provably empty here. That would reduce the maximum 6944 // size requirements for preserved_{oop,mark}_stack. 6945 // But we'll just postpone it until we are all done 6946 // so we can just stream through. 6947 if (!_concurrent_precleaning && CMSOverflowEarlyRestoration) { 6948 _collector->restore_preserved_marks_if_any(); 6949 assert(_collector->no_preserved_marks(), "No preserved marks"); 6950 } 6951 assert(!CMSOverflowEarlyRestoration || _collector->no_preserved_marks(), 6952 "All preserved marks should have been restored above"); 6953 } 6954 } 6955 6956 void MarkRefsIntoAndScanClosure::do_oop(oop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); } 6957 void MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); } 6958 6959 void MarkRefsIntoAndScanClosure::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(_bit_map->lock()); 6964 // relinquish the free_list_lock and bitMaplock() 6965 _bit_map->lock()->unlock(); 6966 _freelistLock->unlock(); 6967 ConcurrentMarkSweepThread::desynchronize(true); 6968 ConcurrentMarkSweepThread::acknowledge_yield_request(); 6969 _collector->stopTimer(); 6970 if (PrintCMSStatistics != 0) { 6971 _collector->incrementYields(); 6972 } 6973 _collector->icms_wait(); 6974 6975 // See the comment in coordinator_yield() 6976 for (unsigned i = 0; 6977 i < CMSYieldSleepCount && 6978 ConcurrentMarkSweepThread::should_yield() && 6979 !CMSCollector::foregroundGCIsActive(); 6980 ++i) { 6981 os::sleep(Thread::current(), 1, false); 6982 ConcurrentMarkSweepThread::acknowledge_yield_request(); 6983 } 6984 6985 ConcurrentMarkSweepThread::synchronize(true); 6986 _freelistLock->lock_without_safepoint_check(); 6987 _bit_map->lock()->lock_without_safepoint_check(); 6988 _collector->startTimer(); 6989 } 6990 6991 /////////////////////////////////////////////////////////// 6992 // Par_MarkRefsIntoAndScanClosure: a parallel version of 6993 // MarkRefsIntoAndScanClosure 6994 /////////////////////////////////////////////////////////// 6995 Par_MarkRefsIntoAndScanClosure::Par_MarkRefsIntoAndScanClosure( 6996 CMSCollector* collector, MemRegion span, ReferenceProcessor* rp, 6997 CMSBitMap* bit_map, OopTaskQueue* work_queue): 6998 _span(span), 6999 _bit_map(bit_map), 7000 _work_queue(work_queue), 7001 _low_water_mark(MIN2((uint)(work_queue->max_elems()/4), 7002 (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads))), 7003 _par_pushAndMarkClosure(collector, span, rp, bit_map, work_queue) 7004 { 7005 _ref_processor = rp; 7006 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 7007 } 7008 7009 // This closure is used to mark refs into the CMS generation at the 7010 // second (final) checkpoint, and to scan and transitively follow 7011 // the unmarked oops. The marks are made in the marking bit map and 7012 // the work_queue is used for keeping the (newly) grey objects during 7013 // the scan phase whence they are also available for stealing by parallel 7014 // threads. Since the marking bit map is shared, updates are 7015 // synchronized (via CAS). 7016 void Par_MarkRefsIntoAndScanClosure::do_oop(oop obj) { 7017 if (obj != NULL) { 7018 // Ignore mark word because this could be an already marked oop 7019 // that may be chained at the end of the overflow list. 7020 assert(obj->is_oop(true), "expected an oop"); 7021 HeapWord* addr = (HeapWord*)obj; 7022 if (_span.contains(addr) && 7023 !_bit_map->isMarked(addr)) { 7024 // mark bit map (object will become grey): 7025 // It is possible for several threads to be 7026 // trying to "claim" this object concurrently; 7027 // the unique thread that succeeds in marking the 7028 // object first will do the subsequent push on 7029 // to the work queue (or overflow list). 7030 if (_bit_map->par_mark(addr)) { 7031 // push on work_queue (which may not be empty), and trim the 7032 // queue to an appropriate length by applying this closure to 7033 // the oops in the oops popped from the stack (i.e. blacken the 7034 // grey objects) 7035 bool res = _work_queue->push(obj); 7036 assert(res, "Low water mark should be less than capacity?"); 7037 trim_queue(_low_water_mark); 7038 } // Else, another thread claimed the object 7039 } 7040 } 7041 } 7042 7043 void Par_MarkRefsIntoAndScanClosure::do_oop(oop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); } 7044 void Par_MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); } 7045 7046 // This closure is used to rescan the marked objects on the dirty cards 7047 // in the mod union table and the card table proper. 7048 size_t ScanMarkedObjectsAgainCarefullyClosure::do_object_careful_m( 7049 oop p, MemRegion mr) { 7050 7051 size_t size = 0; 7052 HeapWord* addr = (HeapWord*)p; 7053 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 7054 assert(_span.contains(addr), "we are scanning the CMS generation"); 7055 // check if it's time to yield 7056 if (do_yield_check()) { 7057 // We yielded for some foreground stop-world work, 7058 // and we have been asked to abort this ongoing preclean cycle. 7059 return 0; 7060 } 7061 if (_bitMap->isMarked(addr)) { 7062 // it's marked; is it potentially uninitialized? 7063 if (p->klass_or_null() != NULL) { 7064 // an initialized object; ignore mark word in verification below 7065 // since we are running concurrent with mutators 7066 assert(p->is_oop(true), "should be an oop"); 7067 if (p->is_objArray()) { 7068 // objArrays are precisely marked; restrict scanning 7069 // to dirty cards only. 7070 size = CompactibleFreeListSpace::adjustObjectSize( 7071 p->oop_iterate(_scanningClosure, mr)); 7072 } else { 7073 // A non-array may have been imprecisely marked; we need 7074 // to scan object in its entirety. 7075 size = CompactibleFreeListSpace::adjustObjectSize( 7076 p->oop_iterate(_scanningClosure)); 7077 } 7078 #ifdef ASSERT 7079 size_t direct_size = 7080 CompactibleFreeListSpace::adjustObjectSize(p->size()); 7081 assert(size == direct_size, "Inconsistency in size"); 7082 assert(size >= 3, "Necessary for Printezis marks to work"); 7083 if (!_bitMap->isMarked(addr+1)) { 7084 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size); 7085 } else { 7086 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size-1); 7087 assert(_bitMap->isMarked(addr+size-1), 7088 "inconsistent Printezis mark"); 7089 } 7090 #endif // ASSERT 7091 } else { 7092 // An uninitialized object. 7093 assert(_bitMap->isMarked(addr+1), "missing Printezis mark?"); 7094 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 7095 size = pointer_delta(nextOneAddr + 1, addr); 7096 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 7097 "alignment problem"); 7098 // Note that pre-cleaning needn't redirty the card. OopDesc::set_klass() 7099 // will dirty the card when the klass pointer is installed in the 7100 // object (signaling the completion of initialization). 7101 } 7102 } else { 7103 // Either a not yet marked object or an uninitialized object 7104 if (p->klass_or_null() == NULL) { 7105 // An uninitialized object, skip to the next card, since 7106 // we may not be able to read its P-bits yet. 7107 assert(size == 0, "Initial value"); 7108 } else { 7109 // An object not (yet) reached by marking: we merely need to 7110 // compute its size so as to go look at the next block. 7111 assert(p->is_oop(true), "should be an oop"); 7112 size = CompactibleFreeListSpace::adjustObjectSize(p->size()); 7113 } 7114 } 7115 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 7116 return size; 7117 } 7118 7119 void ScanMarkedObjectsAgainCarefullyClosure::do_yield_work() { 7120 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 7121 "CMS thread should hold CMS token"); 7122 assert_lock_strong(_freelistLock); 7123 assert_lock_strong(_bitMap->lock()); 7124 // relinquish the free_list_lock and bitMaplock() 7125 _bitMap->lock()->unlock(); 7126 _freelistLock->unlock(); 7127 ConcurrentMarkSweepThread::desynchronize(true); 7128 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7129 _collector->stopTimer(); 7130 if (PrintCMSStatistics != 0) { 7131 _collector->incrementYields(); 7132 } 7133 _collector->icms_wait(); 7134 7135 // See the comment in coordinator_yield() 7136 for (unsigned i = 0; i < CMSYieldSleepCount && 7137 ConcurrentMarkSweepThread::should_yield() && 7138 !CMSCollector::foregroundGCIsActive(); ++i) { 7139 os::sleep(Thread::current(), 1, false); 7140 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7141 } 7142 7143 ConcurrentMarkSweepThread::synchronize(true); 7144 _freelistLock->lock_without_safepoint_check(); 7145 _bitMap->lock()->lock_without_safepoint_check(); 7146 _collector->startTimer(); 7147 } 7148 7149 7150 ////////////////////////////////////////////////////////////////// 7151 // SurvivorSpacePrecleanClosure 7152 ////////////////////////////////////////////////////////////////// 7153 // This (single-threaded) closure is used to preclean the oops in 7154 // the survivor spaces. 7155 size_t SurvivorSpacePrecleanClosure::do_object_careful(oop p) { 7156 7157 HeapWord* addr = (HeapWord*)p; 7158 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 7159 assert(!_span.contains(addr), "we are scanning the survivor spaces"); 7160 assert(p->klass_or_null() != NULL, "object should be initialized"); 7161 // an initialized object; ignore mark word in verification below 7162 // since we are running concurrent with mutators 7163 assert(p->is_oop(true), "should be an oop"); 7164 // Note that we do not yield while we iterate over 7165 // the interior oops of p, pushing the relevant ones 7166 // on our marking stack. 7167 size_t size = p->oop_iterate(_scanning_closure); 7168 do_yield_check(); 7169 // Observe that below, we do not abandon the preclean 7170 // phase as soon as we should; rather we empty the 7171 // marking stack before returning. This is to satisfy 7172 // some existing assertions. In general, it may be a 7173 // good idea to abort immediately and complete the marking 7174 // from the grey objects at a later time. 7175 while (!_mark_stack->isEmpty()) { 7176 oop new_oop = _mark_stack->pop(); 7177 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 7178 assert(_bit_map->isMarked((HeapWord*)new_oop), 7179 "only grey objects on this stack"); 7180 // iterate over the oops in this oop, marking and pushing 7181 // the ones in CMS heap (i.e. in _span). 7182 new_oop->oop_iterate(_scanning_closure); 7183 // check if it's time to yield 7184 do_yield_check(); 7185 } 7186 unsigned int after_count = 7187 GenCollectedHeap::heap()->total_collections(); 7188 bool abort = (_before_count != after_count) || 7189 _collector->should_abort_preclean(); 7190 return abort ? 0 : size; 7191 } 7192 7193 void SurvivorSpacePrecleanClosure::do_yield_work() { 7194 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 7195 "CMS thread should hold CMS token"); 7196 assert_lock_strong(_bit_map->lock()); 7197 // Relinquish the bit map lock 7198 _bit_map->lock()->unlock(); 7199 ConcurrentMarkSweepThread::desynchronize(true); 7200 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7201 _collector->stopTimer(); 7202 if (PrintCMSStatistics != 0) { 7203 _collector->incrementYields(); 7204 } 7205 _collector->icms_wait(); 7206 7207 // See the comment in coordinator_yield() 7208 for (unsigned i = 0; i < CMSYieldSleepCount && 7209 ConcurrentMarkSweepThread::should_yield() && 7210 !CMSCollector::foregroundGCIsActive(); ++i) { 7211 os::sleep(Thread::current(), 1, false); 7212 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7213 } 7214 7215 ConcurrentMarkSweepThread::synchronize(true); 7216 _bit_map->lock()->lock_without_safepoint_check(); 7217 _collector->startTimer(); 7218 } 7219 7220 // This closure is used to rescan the marked objects on the dirty cards 7221 // in the mod union table and the card table proper. In the parallel 7222 // case, although the bitMap is shared, we do a single read so the 7223 // isMarked() query is "safe". 7224 bool ScanMarkedObjectsAgainClosure::do_object_bm(oop p, MemRegion mr) { 7225 // Ignore mark word because we are running concurrent with mutators 7226 assert(p->is_oop_or_null(true), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(p))); 7227 HeapWord* addr = (HeapWord*)p; 7228 assert(_span.contains(addr), "we are scanning the CMS generation"); 7229 bool is_obj_array = false; 7230 #ifdef ASSERT 7231 if (!_parallel) { 7232 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 7233 assert(_collector->overflow_list_is_empty(), 7234 "overflow list should be empty"); 7235 7236 } 7237 #endif // ASSERT 7238 if (_bit_map->isMarked(addr)) { 7239 // Obj arrays are precisely marked, non-arrays are not; 7240 // so we scan objArrays precisely and non-arrays in their 7241 // entirety. 7242 if (p->is_objArray()) { 7243 is_obj_array = true; 7244 if (_parallel) { 7245 p->oop_iterate(_par_scan_closure, mr); 7246 } else { 7247 p->oop_iterate(_scan_closure, mr); 7248 } 7249 } else { 7250 if (_parallel) { 7251 p->oop_iterate(_par_scan_closure); 7252 } else { 7253 p->oop_iterate(_scan_closure); 7254 } 7255 } 7256 } 7257 #ifdef ASSERT 7258 if (!_parallel) { 7259 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 7260 assert(_collector->overflow_list_is_empty(), 7261 "overflow list should be empty"); 7262 7263 } 7264 #endif // ASSERT 7265 return is_obj_array; 7266 } 7267 7268 MarkFromRootsClosure::MarkFromRootsClosure(CMSCollector* collector, 7269 MemRegion span, 7270 CMSBitMap* bitMap, CMSMarkStack* markStack, 7271 bool should_yield, bool verifying): 7272 _collector(collector), 7273 _span(span), 7274 _bitMap(bitMap), 7275 _mut(&collector->_modUnionTable), 7276 _markStack(markStack), 7277 _yield(should_yield), 7278 _skipBits(0) 7279 { 7280 assert(_markStack->isEmpty(), "stack should be empty"); 7281 _finger = _bitMap->startWord(); 7282 _threshold = _finger; 7283 assert(_collector->_restart_addr == NULL, "Sanity check"); 7284 assert(_span.contains(_finger), "Out of bounds _finger?"); 7285 DEBUG_ONLY(_verifying = verifying;) 7286 } 7287 7288 void MarkFromRootsClosure::reset(HeapWord* addr) { 7289 assert(_markStack->isEmpty(), "would cause duplicates on stack"); 7290 assert(_span.contains(addr), "Out of bounds _finger?"); 7291 _finger = addr; 7292 _threshold = (HeapWord*)round_to( 7293 (intptr_t)_finger, CardTableModRefBS::card_size); 7294 } 7295 7296 // Should revisit to see if this should be restructured for 7297 // greater efficiency. 7298 bool MarkFromRootsClosure::do_bit(size_t offset) { 7299 if (_skipBits > 0) { 7300 _skipBits--; 7301 return true; 7302 } 7303 // convert offset into a HeapWord* 7304 HeapWord* addr = _bitMap->startWord() + offset; 7305 assert(_bitMap->endWord() && addr < _bitMap->endWord(), 7306 "address out of range"); 7307 assert(_bitMap->isMarked(addr), "tautology"); 7308 if (_bitMap->isMarked(addr+1)) { 7309 // this is an allocated but not yet initialized object 7310 assert(_skipBits == 0, "tautology"); 7311 _skipBits = 2; // skip next two marked bits ("Printezis-marks") 7312 oop p = oop(addr); 7313 if (p->klass_or_null() == NULL) { 7314 DEBUG_ONLY(if (!_verifying) {) 7315 // We re-dirty the cards on which this object lies and increase 7316 // the _threshold so that we'll come back to scan this object 7317 // during the preclean or remark phase. (CMSCleanOnEnter) 7318 if (CMSCleanOnEnter) { 7319 size_t sz = _collector->block_size_using_printezis_bits(addr); 7320 HeapWord* end_card_addr = (HeapWord*)round_to( 7321 (intptr_t)(addr+sz), CardTableModRefBS::card_size); 7322 MemRegion redirty_range = MemRegion(addr, end_card_addr); 7323 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 7324 // Bump _threshold to end_card_addr; note that 7325 // _threshold cannot possibly exceed end_card_addr, anyhow. 7326 // This prevents future clearing of the card as the scan proceeds 7327 // to the right. 7328 assert(_threshold <= end_card_addr, 7329 "Because we are just scanning into this object"); 7330 if (_threshold < end_card_addr) { 7331 _threshold = end_card_addr; 7332 } 7333 if (p->klass_or_null() != NULL) { 7334 // Redirty the range of cards... 7335 _mut->mark_range(redirty_range); 7336 } // ...else the setting of klass will dirty the card anyway. 7337 } 7338 DEBUG_ONLY(}) 7339 return true; 7340 } 7341 } 7342 scanOopsInOop(addr); 7343 return true; 7344 } 7345 7346 // We take a break if we've been at this for a while, 7347 // so as to avoid monopolizing the locks involved. 7348 void MarkFromRootsClosure::do_yield_work() { 7349 // First give up the locks, then yield, then re-lock 7350 // We should probably use a constructor/destructor idiom to 7351 // do this unlock/lock or modify the MutexUnlocker class to 7352 // serve our purpose. XXX 7353 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 7354 "CMS thread should hold CMS token"); 7355 assert_lock_strong(_bitMap->lock()); 7356 _bitMap->lock()->unlock(); 7357 ConcurrentMarkSweepThread::desynchronize(true); 7358 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7359 _collector->stopTimer(); 7360 if (PrintCMSStatistics != 0) { 7361 _collector->incrementYields(); 7362 } 7363 _collector->icms_wait(); 7364 7365 // See the comment in coordinator_yield() 7366 for (unsigned i = 0; i < CMSYieldSleepCount && 7367 ConcurrentMarkSweepThread::should_yield() && 7368 !CMSCollector::foregroundGCIsActive(); ++i) { 7369 os::sleep(Thread::current(), 1, false); 7370 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7371 } 7372 7373 ConcurrentMarkSweepThread::synchronize(true); 7374 _bitMap->lock()->lock_without_safepoint_check(); 7375 _collector->startTimer(); 7376 } 7377 7378 void MarkFromRootsClosure::scanOopsInOop(HeapWord* ptr) { 7379 assert(_bitMap->isMarked(ptr), "expected bit to be set"); 7380 assert(_markStack->isEmpty(), 7381 "should drain stack to limit stack usage"); 7382 // convert ptr to an oop preparatory to scanning 7383 oop obj = oop(ptr); 7384 // Ignore mark word in verification below, since we 7385 // may be running concurrent with mutators. 7386 assert(obj->is_oop(true), "should be an oop"); 7387 assert(_finger <= ptr, "_finger runneth ahead"); 7388 // advance the finger to right end of this object 7389 _finger = ptr + obj->size(); 7390 assert(_finger > ptr, "we just incremented it above"); 7391 // On large heaps, it may take us some time to get through 7392 // the marking phase (especially if running iCMS). During 7393 // this time it's possible that a lot of mutations have 7394 // accumulated in the card table and the mod union table -- 7395 // these mutation records are redundant until we have 7396 // actually traced into the corresponding card. 7397 // Here, we check whether advancing the finger would make 7398 // us cross into a new card, and if so clear corresponding 7399 // cards in the MUT (preclean them in the card-table in the 7400 // future). 7401 7402 DEBUG_ONLY(if (!_verifying) {) 7403 // The clean-on-enter optimization is disabled by default, 7404 // until we fix 6178663. 7405 if (CMSCleanOnEnter && (_finger > _threshold)) { 7406 // [_threshold, _finger) represents the interval 7407 // of cards to be cleared in MUT (or precleaned in card table). 7408 // The set of cards to be cleared is all those that overlap 7409 // with the interval [_threshold, _finger); note that 7410 // _threshold is always kept card-aligned but _finger isn't 7411 // always card-aligned. 7412 HeapWord* old_threshold = _threshold; 7413 assert(old_threshold == (HeapWord*)round_to( 7414 (intptr_t)old_threshold, CardTableModRefBS::card_size), 7415 "_threshold should always be card-aligned"); 7416 _threshold = (HeapWord*)round_to( 7417 (intptr_t)_finger, CardTableModRefBS::card_size); 7418 MemRegion mr(old_threshold, _threshold); 7419 assert(!mr.is_empty(), "Control point invariant"); 7420 assert(_span.contains(mr), "Should clear within span"); 7421 _mut->clear_range(mr); 7422 } 7423 DEBUG_ONLY(}) 7424 // Note: the finger doesn't advance while we drain 7425 // the stack below. 7426 PushOrMarkClosure pushOrMarkClosure(_collector, 7427 _span, _bitMap, _markStack, 7428 _finger, this); 7429 bool res = _markStack->push(obj); 7430 assert(res, "Empty non-zero size stack should have space for single push"); 7431 while (!_markStack->isEmpty()) { 7432 oop new_oop = _markStack->pop(); 7433 // Skip verifying header mark word below because we are 7434 // running concurrent with mutators. 7435 assert(new_oop->is_oop(true), "Oops! expected to pop an oop"); 7436 // now scan this oop's oops 7437 new_oop->oop_iterate(&pushOrMarkClosure); 7438 do_yield_check(); 7439 } 7440 assert(_markStack->isEmpty(), "tautology, emphasizing post-condition"); 7441 } 7442 7443 Par_MarkFromRootsClosure::Par_MarkFromRootsClosure(CMSConcMarkingTask* task, 7444 CMSCollector* collector, MemRegion span, 7445 CMSBitMap* bit_map, 7446 OopTaskQueue* work_queue, 7447 CMSMarkStack* overflow_stack, 7448 bool should_yield): 7449 _collector(collector), 7450 _whole_span(collector->_span), 7451 _span(span), 7452 _bit_map(bit_map), 7453 _mut(&collector->_modUnionTable), 7454 _work_queue(work_queue), 7455 _overflow_stack(overflow_stack), 7456 _yield(should_yield), 7457 _skip_bits(0), 7458 _task(task) 7459 { 7460 assert(_work_queue->size() == 0, "work_queue should be empty"); 7461 _finger = span.start(); 7462 _threshold = _finger; // XXX Defer clear-on-enter optimization for now 7463 assert(_span.contains(_finger), "Out of bounds _finger?"); 7464 } 7465 7466 // Should revisit to see if this should be restructured for 7467 // greater efficiency. 7468 bool Par_MarkFromRootsClosure::do_bit(size_t offset) { 7469 if (_skip_bits > 0) { 7470 _skip_bits--; 7471 return true; 7472 } 7473 // convert offset into a HeapWord* 7474 HeapWord* addr = _bit_map->startWord() + offset; 7475 assert(_bit_map->endWord() && addr < _bit_map->endWord(), 7476 "address out of range"); 7477 assert(_bit_map->isMarked(addr), "tautology"); 7478 if (_bit_map->isMarked(addr+1)) { 7479 // this is an allocated object that might not yet be initialized 7480 assert(_skip_bits == 0, "tautology"); 7481 _skip_bits = 2; // skip next two marked bits ("Printezis-marks") 7482 oop p = oop(addr); 7483 if (p->klass_or_null() == NULL) { 7484 // in the case of Clean-on-Enter optimization, redirty card 7485 // and avoid clearing card by increasing the threshold. 7486 return true; 7487 } 7488 } 7489 scan_oops_in_oop(addr); 7490 return true; 7491 } 7492 7493 void Par_MarkFromRootsClosure::scan_oops_in_oop(HeapWord* ptr) { 7494 assert(_bit_map->isMarked(ptr), "expected bit to be set"); 7495 // Should we assert that our work queue is empty or 7496 // below some drain limit? 7497 assert(_work_queue->size() == 0, 7498 "should drain stack to limit stack usage"); 7499 // convert ptr to an oop preparatory to scanning 7500 oop obj = oop(ptr); 7501 // Ignore mark word in verification below, since we 7502 // may be running concurrent with mutators. 7503 assert(obj->is_oop(true), "should be an oop"); 7504 assert(_finger <= ptr, "_finger runneth ahead"); 7505 // advance the finger to right end of this object 7506 _finger = ptr + obj->size(); 7507 assert(_finger > ptr, "we just incremented it above"); 7508 // On large heaps, it may take us some time to get through 7509 // the marking phase (especially if running iCMS). During 7510 // this time it's possible that a lot of mutations have 7511 // accumulated in the card table and the mod union table -- 7512 // these mutation records are redundant until we have 7513 // actually traced into the corresponding card. 7514 // Here, we check whether advancing the finger would make 7515 // us cross into a new card, and if so clear corresponding 7516 // cards in the MUT (preclean them in the card-table in the 7517 // future). 7518 7519 // The clean-on-enter optimization is disabled by default, 7520 // until we fix 6178663. 7521 if (CMSCleanOnEnter && (_finger > _threshold)) { 7522 // [_threshold, _finger) represents the interval 7523 // of cards to be cleared in MUT (or precleaned in card table). 7524 // The set of cards to be cleared is all those that overlap 7525 // with the interval [_threshold, _finger); note that 7526 // _threshold is always kept card-aligned but _finger isn't 7527 // always card-aligned. 7528 HeapWord* old_threshold = _threshold; 7529 assert(old_threshold == (HeapWord*)round_to( 7530 (intptr_t)old_threshold, CardTableModRefBS::card_size), 7531 "_threshold should always be card-aligned"); 7532 _threshold = (HeapWord*)round_to( 7533 (intptr_t)_finger, CardTableModRefBS::card_size); 7534 MemRegion mr(old_threshold, _threshold); 7535 assert(!mr.is_empty(), "Control point invariant"); 7536 assert(_span.contains(mr), "Should clear within span"); // _whole_span ?? 7537 _mut->clear_range(mr); 7538 } 7539 7540 // Note: the local finger doesn't advance while we drain 7541 // the stack below, but the global finger sure can and will. 7542 HeapWord** gfa = _task->global_finger_addr(); 7543 Par_PushOrMarkClosure pushOrMarkClosure(_collector, 7544 _span, _bit_map, 7545 _work_queue, 7546 _overflow_stack, 7547 _finger, 7548 gfa, this); 7549 bool res = _work_queue->push(obj); // overflow could occur here 7550 assert(res, "Will hold once we use workqueues"); 7551 while (true) { 7552 oop new_oop; 7553 if (!_work_queue->pop_local(new_oop)) { 7554 // We emptied our work_queue; check if there's stuff that can 7555 // be gotten from the overflow stack. 7556 if (CMSConcMarkingTask::get_work_from_overflow_stack( 7557 _overflow_stack, _work_queue)) { 7558 do_yield_check(); 7559 continue; 7560 } else { // done 7561 break; 7562 } 7563 } 7564 // Skip verifying header mark word below because we are 7565 // running concurrent with mutators. 7566 assert(new_oop->is_oop(true), "Oops! expected to pop an oop"); 7567 // now scan this oop's oops 7568 new_oop->oop_iterate(&pushOrMarkClosure); 7569 do_yield_check(); 7570 } 7571 assert(_work_queue->size() == 0, "tautology, emphasizing post-condition"); 7572 } 7573 7574 // Yield in response to a request from VM Thread or 7575 // from mutators. 7576 void Par_MarkFromRootsClosure::do_yield_work() { 7577 assert(_task != NULL, "sanity"); 7578 _task->yield(); 7579 } 7580 7581 // A variant of the above used for verifying CMS marking work. 7582 MarkFromRootsVerifyClosure::MarkFromRootsVerifyClosure(CMSCollector* collector, 7583 MemRegion span, 7584 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 7585 CMSMarkStack* mark_stack): 7586 _collector(collector), 7587 _span(span), 7588 _verification_bm(verification_bm), 7589 _cms_bm(cms_bm), 7590 _mark_stack(mark_stack), 7591 _pam_verify_closure(collector, span, verification_bm, cms_bm, 7592 mark_stack) 7593 { 7594 assert(_mark_stack->isEmpty(), "stack should be empty"); 7595 _finger = _verification_bm->startWord(); 7596 assert(_collector->_restart_addr == NULL, "Sanity check"); 7597 assert(_span.contains(_finger), "Out of bounds _finger?"); 7598 } 7599 7600 void MarkFromRootsVerifyClosure::reset(HeapWord* addr) { 7601 assert(_mark_stack->isEmpty(), "would cause duplicates on stack"); 7602 assert(_span.contains(addr), "Out of bounds _finger?"); 7603 _finger = addr; 7604 } 7605 7606 // Should revisit to see if this should be restructured for 7607 // greater efficiency. 7608 bool MarkFromRootsVerifyClosure::do_bit(size_t offset) { 7609 // convert offset into a HeapWord* 7610 HeapWord* addr = _verification_bm->startWord() + offset; 7611 assert(_verification_bm->endWord() && addr < _verification_bm->endWord(), 7612 "address out of range"); 7613 assert(_verification_bm->isMarked(addr), "tautology"); 7614 assert(_cms_bm->isMarked(addr), "tautology"); 7615 7616 assert(_mark_stack->isEmpty(), 7617 "should drain stack to limit stack usage"); 7618 // convert addr to an oop preparatory to scanning 7619 oop obj = oop(addr); 7620 assert(obj->is_oop(), "should be an oop"); 7621 assert(_finger <= addr, "_finger runneth ahead"); 7622 // advance the finger to right end of this object 7623 _finger = addr + obj->size(); 7624 assert(_finger > addr, "we just incremented it above"); 7625 // Note: the finger doesn't advance while we drain 7626 // the stack below. 7627 bool res = _mark_stack->push(obj); 7628 assert(res, "Empty non-zero size stack should have space for single push"); 7629 while (!_mark_stack->isEmpty()) { 7630 oop new_oop = _mark_stack->pop(); 7631 assert(new_oop->is_oop(), "Oops! expected to pop an oop"); 7632 // now scan this oop's oops 7633 new_oop->oop_iterate(&_pam_verify_closure); 7634 } 7635 assert(_mark_stack->isEmpty(), "tautology, emphasizing post-condition"); 7636 return true; 7637 } 7638 7639 PushAndMarkVerifyClosure::PushAndMarkVerifyClosure( 7640 CMSCollector* collector, MemRegion span, 7641 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 7642 CMSMarkStack* mark_stack): 7643 MetadataAwareOopClosure(collector->ref_processor()), 7644 _collector(collector), 7645 _span(span), 7646 _verification_bm(verification_bm), 7647 _cms_bm(cms_bm), 7648 _mark_stack(mark_stack) 7649 { } 7650 7651 void PushAndMarkVerifyClosure::do_oop(oop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 7652 void PushAndMarkVerifyClosure::do_oop(narrowOop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 7653 7654 // Upon stack overflow, we discard (part of) the stack, 7655 // remembering the least address amongst those discarded 7656 // in CMSCollector's _restart_address. 7657 void PushAndMarkVerifyClosure::handle_stack_overflow(HeapWord* lost) { 7658 // Remember the least grey address discarded 7659 HeapWord* ra = (HeapWord*)_mark_stack->least_value(lost); 7660 _collector->lower_restart_addr(ra); 7661 _mark_stack->reset(); // discard stack contents 7662 _mark_stack->expand(); // expand the stack if possible 7663 } 7664 7665 void PushAndMarkVerifyClosure::do_oop(oop obj) { 7666 assert(obj->is_oop_or_null(), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj))); 7667 HeapWord* addr = (HeapWord*)obj; 7668 if (_span.contains(addr) && !_verification_bm->isMarked(addr)) { 7669 // Oop lies in _span and isn't yet grey or black 7670 _verification_bm->mark(addr); // now grey 7671 if (!_cms_bm->isMarked(addr)) { 7672 oop(addr)->print(); 7673 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)", 7674 addr); 7675 fatal("... aborting"); 7676 } 7677 7678 if (!_mark_stack->push(obj)) { // stack overflow 7679 if (PrintCMSStatistics != 0) { 7680 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 7681 SIZE_FORMAT, _mark_stack->capacity()); 7682 } 7683 assert(_mark_stack->isFull(), "Else push should have succeeded"); 7684 handle_stack_overflow(addr); 7685 } 7686 // anything including and to the right of _finger 7687 // will be scanned as we iterate over the remainder of the 7688 // bit map 7689 } 7690 } 7691 7692 PushOrMarkClosure::PushOrMarkClosure(CMSCollector* collector, 7693 MemRegion span, 7694 CMSBitMap* bitMap, CMSMarkStack* markStack, 7695 HeapWord* finger, MarkFromRootsClosure* parent) : 7696 MetadataAwareOopClosure(collector->ref_processor()), 7697 _collector(collector), 7698 _span(span), 7699 _bitMap(bitMap), 7700 _markStack(markStack), 7701 _finger(finger), 7702 _parent(parent) 7703 { } 7704 7705 Par_PushOrMarkClosure::Par_PushOrMarkClosure(CMSCollector* collector, 7706 MemRegion span, 7707 CMSBitMap* bit_map, 7708 OopTaskQueue* work_queue, 7709 CMSMarkStack* overflow_stack, 7710 HeapWord* finger, 7711 HeapWord** global_finger_addr, 7712 Par_MarkFromRootsClosure* parent) : 7713 MetadataAwareOopClosure(collector->ref_processor()), 7714 _collector(collector), 7715 _whole_span(collector->_span), 7716 _span(span), 7717 _bit_map(bit_map), 7718 _work_queue(work_queue), 7719 _overflow_stack(overflow_stack), 7720 _finger(finger), 7721 _global_finger_addr(global_finger_addr), 7722 _parent(parent) 7723 { } 7724 7725 // Assumes thread-safe access by callers, who are 7726 // responsible for mutual exclusion. 7727 void CMSCollector::lower_restart_addr(HeapWord* low) { 7728 assert(_span.contains(low), "Out of bounds addr"); 7729 if (_restart_addr == NULL) { 7730 _restart_addr = low; 7731 } else { 7732 _restart_addr = MIN2(_restart_addr, low); 7733 } 7734 } 7735 7736 // Upon stack overflow, we discard (part of) the stack, 7737 // remembering the least address amongst those discarded 7738 // in CMSCollector's _restart_address. 7739 void PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 7740 // Remember the least grey address discarded 7741 HeapWord* ra = (HeapWord*)_markStack->least_value(lost); 7742 _collector->lower_restart_addr(ra); 7743 _markStack->reset(); // discard stack contents 7744 _markStack->expand(); // expand the stack if possible 7745 } 7746 7747 // Upon stack overflow, we discard (part of) the stack, 7748 // remembering the least address amongst those discarded 7749 // in CMSCollector's _restart_address. 7750 void Par_PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 7751 // We need to do this under a mutex to prevent other 7752 // workers from interfering with the work done below. 7753 MutexLockerEx ml(_overflow_stack->par_lock(), 7754 Mutex::_no_safepoint_check_flag); 7755 // Remember the least grey address discarded 7756 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 7757 _collector->lower_restart_addr(ra); 7758 _overflow_stack->reset(); // discard stack contents 7759 _overflow_stack->expand(); // expand the stack if possible 7760 } 7761 7762 void PushOrMarkClosure::do_oop(oop obj) { 7763 // Ignore mark word because we are running concurrent with mutators. 7764 assert(obj->is_oop_or_null(true), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj))); 7765 HeapWord* addr = (HeapWord*)obj; 7766 if (_span.contains(addr) && !_bitMap->isMarked(addr)) { 7767 // Oop lies in _span and isn't yet grey or black 7768 _bitMap->mark(addr); // now grey 7769 if (addr < _finger) { 7770 // the bit map iteration has already either passed, or 7771 // sampled, this bit in the bit map; we'll need to 7772 // use the marking stack to scan this oop's oops. 7773 bool simulate_overflow = false; 7774 NOT_PRODUCT( 7775 if (CMSMarkStackOverflowALot && 7776 _collector->simulate_overflow()) { 7777 // simulate a stack overflow 7778 simulate_overflow = true; 7779 } 7780 ) 7781 if (simulate_overflow || !_markStack->push(obj)) { // stack overflow 7782 if (PrintCMSStatistics != 0) { 7783 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 7784 SIZE_FORMAT, _markStack->capacity()); 7785 } 7786 assert(simulate_overflow || _markStack->isFull(), "Else push should have succeeded"); 7787 handle_stack_overflow(addr); 7788 } 7789 } 7790 // anything including and to the right of _finger 7791 // will be scanned as we iterate over the remainder of the 7792 // bit map 7793 do_yield_check(); 7794 } 7795 } 7796 7797 void PushOrMarkClosure::do_oop(oop* p) { PushOrMarkClosure::do_oop_work(p); } 7798 void PushOrMarkClosure::do_oop(narrowOop* p) { PushOrMarkClosure::do_oop_work(p); } 7799 7800 void Par_PushOrMarkClosure::do_oop(oop obj) { 7801 // Ignore mark word because we are running concurrent with mutators. 7802 assert(obj->is_oop_or_null(true), err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj))); 7803 HeapWord* addr = (HeapWord*)obj; 7804 if (_whole_span.contains(addr) && !_bit_map->isMarked(addr)) { 7805 // Oop lies in _span and isn't yet grey or black 7806 // We read the global_finger (volatile read) strictly after marking oop 7807 bool res = _bit_map->par_mark(addr); // now grey 7808 volatile HeapWord** gfa = (volatile HeapWord**)_global_finger_addr; 7809 // Should we push this marked oop on our stack? 7810 // -- if someone else marked it, nothing to do 7811 // -- if target oop is above global finger nothing to do 7812 // -- if target oop is in chunk and above local finger 7813 // then nothing to do 7814 // -- else push on work queue 7815 if ( !res // someone else marked it, they will deal with it 7816 || (addr >= *gfa) // will be scanned in a later task 7817 || (_span.contains(addr) && addr >= _finger)) { // later in this chunk 7818 return; 7819 } 7820 // the bit map iteration has already either passed, or 7821 // sampled, this bit in the bit map; we'll need to 7822 // use the marking stack to scan this oop's oops. 7823 bool simulate_overflow = false; 7824 NOT_PRODUCT( 7825 if (CMSMarkStackOverflowALot && 7826 _collector->simulate_overflow()) { 7827 // simulate a stack overflow 7828 simulate_overflow = true; 7829 } 7830 ) 7831 if (simulate_overflow || 7832 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) { 7833 // stack overflow 7834 if (PrintCMSStatistics != 0) { 7835 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 7836 SIZE_FORMAT, _overflow_stack->capacity()); 7837 } 7838 // We cannot assert that the overflow stack is full because 7839 // it may have been emptied since. 7840 assert(simulate_overflow || 7841 _work_queue->size() == _work_queue->max_elems(), 7842 "Else push should have succeeded"); 7843 handle_stack_overflow(addr); 7844 } 7845 do_yield_check(); 7846 } 7847 } 7848 7849 void Par_PushOrMarkClosure::do_oop(oop* p) { Par_PushOrMarkClosure::do_oop_work(p); } 7850 void Par_PushOrMarkClosure::do_oop(narrowOop* p) { Par_PushOrMarkClosure::do_oop_work(p); } 7851 7852 PushAndMarkClosure::PushAndMarkClosure(CMSCollector* collector, 7853 MemRegion span, 7854 ReferenceProcessor* rp, 7855 CMSBitMap* bit_map, 7856 CMSBitMap* mod_union_table, 7857 CMSMarkStack* mark_stack, 7858 bool concurrent_precleaning): 7859 MetadataAwareOopClosure(rp), 7860 _collector(collector), 7861 _span(span), 7862 _bit_map(bit_map), 7863 _mod_union_table(mod_union_table), 7864 _mark_stack(mark_stack), 7865 _concurrent_precleaning(concurrent_precleaning) 7866 { 7867 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 7868 } 7869 7870 // Grey object rescan during pre-cleaning and second checkpoint phases -- 7871 // the non-parallel version (the parallel version appears further below.) 7872 void PushAndMarkClosure::do_oop(oop obj) { 7873 // Ignore mark word verification. If during concurrent precleaning, 7874 // the object monitor may be locked. If during the checkpoint 7875 // phases, the object may already have been reached by a different 7876 // path and may be at the end of the global overflow list (so 7877 // the mark word may be NULL). 7878 assert(obj->is_oop_or_null(true /* ignore mark word */), 7879 err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj))); 7880 HeapWord* addr = (HeapWord*)obj; 7881 // Check if oop points into the CMS generation 7882 // and is not marked 7883 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 7884 // a white object ... 7885 _bit_map->mark(addr); // ... now grey 7886 // push on the marking stack (grey set) 7887 bool simulate_overflow = false; 7888 NOT_PRODUCT( 7889 if (CMSMarkStackOverflowALot && 7890 _collector->simulate_overflow()) { 7891 // simulate a stack overflow 7892 simulate_overflow = true; 7893 } 7894 ) 7895 if (simulate_overflow || !_mark_stack->push(obj)) { 7896 if (_concurrent_precleaning) { 7897 // During precleaning we can just dirty the appropriate card(s) 7898 // in the mod union table, thus ensuring that the object remains 7899 // in the grey set and continue. In the case of object arrays 7900 // we need to dirty all of the cards that the object spans, 7901 // since the rescan of object arrays will be limited to the 7902 // dirty cards. 7903 // Note that no one can be interfering with us in this action 7904 // of dirtying the mod union table, so no locking or atomics 7905 // are required. 7906 if (obj->is_objArray()) { 7907 size_t sz = obj->size(); 7908 HeapWord* end_card_addr = (HeapWord*)round_to( 7909 (intptr_t)(addr+sz), CardTableModRefBS::card_size); 7910 MemRegion redirty_range = MemRegion(addr, end_card_addr); 7911 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 7912 _mod_union_table->mark_range(redirty_range); 7913 } else { 7914 _mod_union_table->mark(addr); 7915 } 7916 _collector->_ser_pmc_preclean_ovflw++; 7917 } else { 7918 // During the remark phase, we need to remember this oop 7919 // in the overflow list. 7920 _collector->push_on_overflow_list(obj); 7921 _collector->_ser_pmc_remark_ovflw++; 7922 } 7923 } 7924 } 7925 } 7926 7927 Par_PushAndMarkClosure::Par_PushAndMarkClosure(CMSCollector* collector, 7928 MemRegion span, 7929 ReferenceProcessor* rp, 7930 CMSBitMap* bit_map, 7931 OopTaskQueue* work_queue): 7932 MetadataAwareOopClosure(rp), 7933 _collector(collector), 7934 _span(span), 7935 _bit_map(bit_map), 7936 _work_queue(work_queue) 7937 { 7938 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 7939 } 7940 7941 void PushAndMarkClosure::do_oop(oop* p) { PushAndMarkClosure::do_oop_work(p); } 7942 void PushAndMarkClosure::do_oop(narrowOop* p) { PushAndMarkClosure::do_oop_work(p); } 7943 7944 // Grey object rescan during second checkpoint phase -- 7945 // the parallel version. 7946 void Par_PushAndMarkClosure::do_oop(oop obj) { 7947 // In the assert below, we ignore the mark word because 7948 // this oop may point to an already visited object that is 7949 // on the overflow stack (in which case the mark word has 7950 // been hijacked for chaining into the overflow stack -- 7951 // if this is the last object in the overflow stack then 7952 // its mark word will be NULL). Because this object may 7953 // have been subsequently popped off the global overflow 7954 // stack, and the mark word possibly restored to the prototypical 7955 // value, by the time we get to examined this failing assert in 7956 // the debugger, is_oop_or_null(false) may subsequently start 7957 // to hold. 7958 assert(obj->is_oop_or_null(true), 7959 err_msg("Expected an oop or NULL at " PTR_FORMAT, p2i(obj))); 7960 HeapWord* addr = (HeapWord*)obj; 7961 // Check if oop points into the CMS generation 7962 // and is not marked 7963 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 7964 // a white object ... 7965 // If we manage to "claim" the object, by being the 7966 // first thread to mark it, then we push it on our 7967 // marking stack 7968 if (_bit_map->par_mark(addr)) { // ... now grey 7969 // push on work queue (grey set) 7970 bool simulate_overflow = false; 7971 NOT_PRODUCT( 7972 if (CMSMarkStackOverflowALot && 7973 _collector->par_simulate_overflow()) { 7974 // simulate a stack overflow 7975 simulate_overflow = true; 7976 } 7977 ) 7978 if (simulate_overflow || !_work_queue->push(obj)) { 7979 _collector->par_push_on_overflow_list(obj); 7980 _collector->_par_pmc_remark_ovflw++; // imprecise OK: no need to CAS 7981 } 7982 } // Else, some other thread got there first 7983 } 7984 } 7985 7986 void Par_PushAndMarkClosure::do_oop(oop* p) { Par_PushAndMarkClosure::do_oop_work(p); } 7987 void Par_PushAndMarkClosure::do_oop(narrowOop* p) { Par_PushAndMarkClosure::do_oop_work(p); } 7988 7989 void CMSPrecleanRefsYieldClosure::do_yield_work() { 7990 Mutex* bml = _collector->bitMapLock(); 7991 assert_lock_strong(bml); 7992 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 7993 "CMS thread should hold CMS token"); 7994 7995 bml->unlock(); 7996 ConcurrentMarkSweepThread::desynchronize(true); 7997 7998 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7999 8000 _collector->stopTimer(); 8001 if (PrintCMSStatistics != 0) { 8002 _collector->incrementYields(); 8003 } 8004 _collector->icms_wait(); 8005 8006 // See the comment in coordinator_yield() 8007 for (unsigned i = 0; i < CMSYieldSleepCount && 8008 ConcurrentMarkSweepThread::should_yield() && 8009 !CMSCollector::foregroundGCIsActive(); ++i) { 8010 os::sleep(Thread::current(), 1, false); 8011 ConcurrentMarkSweepThread::acknowledge_yield_request(); 8012 } 8013 8014 ConcurrentMarkSweepThread::synchronize(true); 8015 bml->lock(); 8016 8017 _collector->startTimer(); 8018 } 8019 8020 bool CMSPrecleanRefsYieldClosure::should_return() { 8021 if (ConcurrentMarkSweepThread::should_yield()) { 8022 do_yield_work(); 8023 } 8024 return _collector->foregroundGCIsActive(); 8025 } 8026 8027 void MarkFromDirtyCardsClosure::do_MemRegion(MemRegion mr) { 8028 assert(((size_t)mr.start())%CardTableModRefBS::card_size_in_words == 0, 8029 "mr should be aligned to start at a card boundary"); 8030 // We'd like to assert: 8031 // assert(mr.word_size()%CardTableModRefBS::card_size_in_words == 0, 8032 // "mr should be a range of cards"); 8033 // However, that would be too strong in one case -- the last 8034 // partition ends at _unallocated_block which, in general, can be 8035 // an arbitrary boundary, not necessarily card aligned. 8036 if (PrintCMSStatistics != 0) { 8037 _num_dirty_cards += 8038 mr.word_size()/CardTableModRefBS::card_size_in_words; 8039 } 8040 _space->object_iterate_mem(mr, &_scan_cl); 8041 } 8042 8043 SweepClosure::SweepClosure(CMSCollector* collector, 8044 ConcurrentMarkSweepGeneration* g, 8045 CMSBitMap* bitMap, bool should_yield) : 8046 _collector(collector), 8047 _g(g), 8048 _sp(g->cmsSpace()), 8049 _limit(_sp->sweep_limit()), 8050 _freelistLock(_sp->freelistLock()), 8051 _bitMap(bitMap), 8052 _yield(should_yield), 8053 _inFreeRange(false), // No free range at beginning of sweep 8054 _freeRangeInFreeLists(false), // No free range at beginning of sweep 8055 _lastFreeRangeCoalesced(false), 8056 _freeFinger(g->used_region().start()) 8057 { 8058 NOT_PRODUCT( 8059 _numObjectsFreed = 0; 8060 _numWordsFreed = 0; 8061 _numObjectsLive = 0; 8062 _numWordsLive = 0; 8063 _numObjectsAlreadyFree = 0; 8064 _numWordsAlreadyFree = 0; 8065 _last_fc = NULL; 8066 8067 _sp->initializeIndexedFreeListArrayReturnedBytes(); 8068 _sp->dictionary()->initialize_dict_returned_bytes(); 8069 ) 8070 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 8071 "sweep _limit out of bounds"); 8072 if (CMSTraceSweeper) { 8073 gclog_or_tty->print_cr("\n====================\nStarting new sweep with limit " PTR_FORMAT, 8074 _limit); 8075 } 8076 } 8077 8078 void SweepClosure::print_on(outputStream* st) const { 8079 tty->print_cr("_sp = [" PTR_FORMAT "," PTR_FORMAT ")", 8080 _sp->bottom(), _sp->end()); 8081 tty->print_cr("_limit = " PTR_FORMAT, _limit); 8082 tty->print_cr("_freeFinger = " PTR_FORMAT, _freeFinger); 8083 NOT_PRODUCT(tty->print_cr("_last_fc = " PTR_FORMAT, _last_fc);) 8084 tty->print_cr("_inFreeRange = %d, _freeRangeInFreeLists = %d, _lastFreeRangeCoalesced = %d", 8085 _inFreeRange, _freeRangeInFreeLists, _lastFreeRangeCoalesced); 8086 } 8087 8088 #ifndef PRODUCT 8089 // Assertion checking only: no useful work in product mode -- 8090 // however, if any of the flags below become product flags, 8091 // you may need to review this code to see if it needs to be 8092 // enabled in product mode. 8093 SweepClosure::~SweepClosure() { 8094 assert_lock_strong(_freelistLock); 8095 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 8096 "sweep _limit out of bounds"); 8097 if (inFreeRange()) { 8098 warning("inFreeRange() should have been reset; dumping state of SweepClosure"); 8099 print(); 8100 ShouldNotReachHere(); 8101 } 8102 if (Verbose && PrintGC) { 8103 gclog_or_tty->print("Collected "SIZE_FORMAT" objects, " SIZE_FORMAT " bytes", 8104 _numObjectsFreed, _numWordsFreed*sizeof(HeapWord)); 8105 gclog_or_tty->print_cr("\nLive "SIZE_FORMAT" objects, " 8106 SIZE_FORMAT" bytes " 8107 "Already free "SIZE_FORMAT" objects, "SIZE_FORMAT" bytes", 8108 _numObjectsLive, _numWordsLive*sizeof(HeapWord), 8109 _numObjectsAlreadyFree, _numWordsAlreadyFree*sizeof(HeapWord)); 8110 size_t totalBytes = (_numWordsFreed + _numWordsLive + _numWordsAlreadyFree) 8111 * sizeof(HeapWord); 8112 gclog_or_tty->print_cr("Total sweep: "SIZE_FORMAT" bytes", totalBytes); 8113 8114 if (PrintCMSStatistics && CMSVerifyReturnedBytes) { 8115 size_t indexListReturnedBytes = _sp->sumIndexedFreeListArrayReturnedBytes(); 8116 size_t dict_returned_bytes = _sp->dictionary()->sum_dict_returned_bytes(); 8117 size_t returned_bytes = indexListReturnedBytes + dict_returned_bytes; 8118 gclog_or_tty->print("Returned "SIZE_FORMAT" bytes", returned_bytes); 8119 gclog_or_tty->print(" Indexed List Returned "SIZE_FORMAT" bytes", 8120 indexListReturnedBytes); 8121 gclog_or_tty->print_cr(" Dictionary Returned "SIZE_FORMAT" bytes", 8122 dict_returned_bytes); 8123 } 8124 } 8125 if (CMSTraceSweeper) { 8126 gclog_or_tty->print_cr("end of sweep with _limit = " PTR_FORMAT "\n================", 8127 _limit); 8128 } 8129 } 8130 #endif // PRODUCT 8131 8132 void SweepClosure::initialize_free_range(HeapWord* freeFinger, 8133 bool freeRangeInFreeLists) { 8134 if (CMSTraceSweeper) { 8135 gclog_or_tty->print("---- Start free range at " PTR_FORMAT " with free block (%d)\n", 8136 freeFinger, freeRangeInFreeLists); 8137 } 8138 assert(!inFreeRange(), "Trampling existing free range"); 8139 set_inFreeRange(true); 8140 set_lastFreeRangeCoalesced(false); 8141 8142 set_freeFinger(freeFinger); 8143 set_freeRangeInFreeLists(freeRangeInFreeLists); 8144 if (CMSTestInFreeList) { 8145 if (freeRangeInFreeLists) { 8146 FreeChunk* fc = (FreeChunk*) freeFinger; 8147 assert(fc->is_free(), "A chunk on the free list should be free."); 8148 assert(fc->size() > 0, "Free range should have a size"); 8149 assert(_sp->verify_chunk_in_free_list(fc), "Chunk is not in free lists"); 8150 } 8151 } 8152 } 8153 8154 // Note that the sweeper runs concurrently with mutators. Thus, 8155 // it is possible for direct allocation in this generation to happen 8156 // in the middle of the sweep. Note that the sweeper also coalesces 8157 // contiguous free blocks. Thus, unless the sweeper and the allocator 8158 // synchronize appropriately freshly allocated blocks may get swept up. 8159 // This is accomplished by the sweeper locking the free lists while 8160 // it is sweeping. Thus blocks that are determined to be free are 8161 // indeed free. There is however one additional complication: 8162 // blocks that have been allocated since the final checkpoint and 8163 // mark, will not have been marked and so would be treated as 8164 // unreachable and swept up. To prevent this, the allocator marks 8165 // the bit map when allocating during the sweep phase. This leads, 8166 // however, to a further complication -- objects may have been allocated 8167 // but not yet initialized -- in the sense that the header isn't yet 8168 // installed. The sweeper can not then determine the size of the block 8169 // in order to skip over it. To deal with this case, we use a technique 8170 // (due to Printezis) to encode such uninitialized block sizes in the 8171 // bit map. Since the bit map uses a bit per every HeapWord, but the 8172 // CMS generation has a minimum object size of 3 HeapWords, it follows 8173 // that "normal marks" won't be adjacent in the bit map (there will 8174 // always be at least two 0 bits between successive 1 bits). We make use 8175 // of these "unused" bits to represent uninitialized blocks -- the bit 8176 // corresponding to the start of the uninitialized object and the next 8177 // bit are both set. Finally, a 1 bit marks the end of the object that 8178 // started with the two consecutive 1 bits to indicate its potentially 8179 // uninitialized state. 8180 8181 size_t SweepClosure::do_blk_careful(HeapWord* addr) { 8182 FreeChunk* fc = (FreeChunk*)addr; 8183 size_t res; 8184 8185 // Check if we are done sweeping. Below we check "addr >= _limit" rather 8186 // than "addr == _limit" because although _limit was a block boundary when 8187 // we started the sweep, it may no longer be one because heap expansion 8188 // may have caused us to coalesce the block ending at the address _limit 8189 // with a newly expanded chunk (this happens when _limit was set to the 8190 // previous _end of the space), so we may have stepped past _limit: 8191 // see the following Zeno-like trail of CRs 6977970, 7008136, 7042740. 8192 if (addr >= _limit) { // we have swept up to or past the limit: finish up 8193 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 8194 "sweep _limit out of bounds"); 8195 assert(addr < _sp->end(), "addr out of bounds"); 8196 // Flush any free range we might be holding as a single 8197 // coalesced chunk to the appropriate free list. 8198 if (inFreeRange()) { 8199 assert(freeFinger() >= _sp->bottom() && freeFinger() < _limit, 8200 err_msg("freeFinger() " PTR_FORMAT" is out-of-bounds", freeFinger())); 8201 flush_cur_free_chunk(freeFinger(), 8202 pointer_delta(addr, freeFinger())); 8203 if (CMSTraceSweeper) { 8204 gclog_or_tty->print("Sweep: last chunk: "); 8205 gclog_or_tty->print("put_free_blk " PTR_FORMAT " ("SIZE_FORMAT") " 8206 "[coalesced:%d]\n", 8207 freeFinger(), pointer_delta(addr, freeFinger()), 8208 lastFreeRangeCoalesced() ? 1 : 0); 8209 } 8210 } 8211 8212 // help the iterator loop finish 8213 return pointer_delta(_sp->end(), addr); 8214 } 8215 8216 assert(addr < _limit, "sweep invariant"); 8217 // check if we should yield 8218 do_yield_check(addr); 8219 if (fc->is_free()) { 8220 // Chunk that is already free 8221 res = fc->size(); 8222 do_already_free_chunk(fc); 8223 debug_only(_sp->verifyFreeLists()); 8224 // If we flush the chunk at hand in lookahead_and_flush() 8225 // and it's coalesced with a preceding chunk, then the 8226 // process of "mangling" the payload of the coalesced block 8227 // will cause erasure of the size information from the 8228 // (erstwhile) header of all the coalesced blocks but the 8229 // first, so the first disjunct in the assert will not hold 8230 // in that specific case (in which case the second disjunct 8231 // will hold). 8232 assert(res == fc->size() || ((HeapWord*)fc) + res >= _limit, 8233 "Otherwise the size info doesn't change at this step"); 8234 NOT_PRODUCT( 8235 _numObjectsAlreadyFree++; 8236 _numWordsAlreadyFree += res; 8237 ) 8238 NOT_PRODUCT(_last_fc = fc;) 8239 } else if (!_bitMap->isMarked(addr)) { 8240 // Chunk is fresh garbage 8241 res = do_garbage_chunk(fc); 8242 debug_only(_sp->verifyFreeLists()); 8243 NOT_PRODUCT( 8244 _numObjectsFreed++; 8245 _numWordsFreed += res; 8246 ) 8247 } else { 8248 // Chunk that is alive. 8249 res = do_live_chunk(fc); 8250 debug_only(_sp->verifyFreeLists()); 8251 NOT_PRODUCT( 8252 _numObjectsLive++; 8253 _numWordsLive += res; 8254 ) 8255 } 8256 return res; 8257 } 8258 8259 // For the smart allocation, record following 8260 // split deaths - a free chunk is removed from its free list because 8261 // it is being split into two or more chunks. 8262 // split birth - a free chunk is being added to its free list because 8263 // a larger free chunk has been split and resulted in this free chunk. 8264 // coal death - a free chunk is being removed from its free list because 8265 // it is being coalesced into a large free chunk. 8266 // coal birth - a free chunk is being added to its free list because 8267 // it was created when two or more free chunks where coalesced into 8268 // this free chunk. 8269 // 8270 // These statistics are used to determine the desired number of free 8271 // chunks of a given size. The desired number is chosen to be relative 8272 // to the end of a CMS sweep. The desired number at the end of a sweep 8273 // is the 8274 // count-at-end-of-previous-sweep (an amount that was enough) 8275 // - count-at-beginning-of-current-sweep (the excess) 8276 // + split-births (gains in this size during interval) 8277 // - split-deaths (demands on this size during interval) 8278 // where the interval is from the end of one sweep to the end of the 8279 // next. 8280 // 8281 // When sweeping the sweeper maintains an accumulated chunk which is 8282 // the chunk that is made up of chunks that have been coalesced. That 8283 // will be termed the left-hand chunk. A new chunk of garbage that 8284 // is being considered for coalescing will be referred to as the 8285 // right-hand chunk. 8286 // 8287 // When making a decision on whether to coalesce a right-hand chunk with 8288 // the current left-hand chunk, the current count vs. the desired count 8289 // of the left-hand chunk is considered. Also if the right-hand chunk 8290 // is near the large chunk at the end of the heap (see 8291 // ConcurrentMarkSweepGeneration::isNearLargestChunk()), then the 8292 // left-hand chunk is coalesced. 8293 // 8294 // When making a decision about whether to split a chunk, the desired count 8295 // vs. the current count of the candidate to be split is also considered. 8296 // If the candidate is underpopulated (currently fewer chunks than desired) 8297 // a chunk of an overpopulated (currently more chunks than desired) size may 8298 // be chosen. The "hint" associated with a free list, if non-null, points 8299 // to a free list which may be overpopulated. 8300 // 8301 8302 void SweepClosure::do_already_free_chunk(FreeChunk* fc) { 8303 const size_t size = fc->size(); 8304 // Chunks that cannot be coalesced are not in the 8305 // free lists. 8306 if (CMSTestInFreeList && !fc->cantCoalesce()) { 8307 assert(_sp->verify_chunk_in_free_list(fc), 8308 "free chunk should be in free lists"); 8309 } 8310 // a chunk that is already free, should not have been 8311 // marked in the bit map 8312 HeapWord* const addr = (HeapWord*) fc; 8313 assert(!_bitMap->isMarked(addr), "free chunk should be unmarked"); 8314 // Verify that the bit map has no bits marked between 8315 // addr and purported end of this block. 8316 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 8317 8318 // Some chunks cannot be coalesced under any circumstances. 8319 // See the definition of cantCoalesce(). 8320 if (!fc->cantCoalesce()) { 8321 // This chunk can potentially be coalesced. 8322 if (_sp->adaptive_freelists()) { 8323 // All the work is done in 8324 do_post_free_or_garbage_chunk(fc, size); 8325 } else { // Not adaptive free lists 8326 // this is a free chunk that can potentially be coalesced by the sweeper; 8327 if (!inFreeRange()) { 8328 // if the next chunk is a free block that can't be coalesced 8329 // it doesn't make sense to remove this chunk from the free lists 8330 FreeChunk* nextChunk = (FreeChunk*)(addr + size); 8331 assert((HeapWord*)nextChunk <= _sp->end(), "Chunk size out of bounds?"); 8332 if ((HeapWord*)nextChunk < _sp->end() && // There is another free chunk to the right ... 8333 nextChunk->is_free() && // ... which is free... 8334 nextChunk->cantCoalesce()) { // ... but can't be coalesced 8335 // nothing to do 8336 } else { 8337 // Potentially the start of a new free range: 8338 // Don't eagerly remove it from the free lists. 8339 // No need to remove it if it will just be put 8340 // back again. (Also from a pragmatic point of view 8341 // if it is a free block in a region that is beyond 8342 // any allocated blocks, an assertion will fail) 8343 // Remember the start of a free run. 8344 initialize_free_range(addr, true); 8345 // end - can coalesce with next chunk 8346 } 8347 } else { 8348 // the midst of a free range, we are coalescing 8349 print_free_block_coalesced(fc); 8350 if (CMSTraceSweeper) { 8351 gclog_or_tty->print(" -- pick up free block " PTR_FORMAT " (" SIZE_FORMAT ")\n", fc, size); 8352 } 8353 // remove it from the free lists 8354 _sp->removeFreeChunkFromFreeLists(fc); 8355 set_lastFreeRangeCoalesced(true); 8356 // If the chunk is being coalesced and the current free range is 8357 // in the free lists, remove the current free range so that it 8358 // will be returned to the free lists in its entirety - all 8359 // the coalesced pieces included. 8360 if (freeRangeInFreeLists()) { 8361 FreeChunk* ffc = (FreeChunk*) freeFinger(); 8362 assert(ffc->size() == pointer_delta(addr, freeFinger()), 8363 "Size of free range is inconsistent with chunk size."); 8364 if (CMSTestInFreeList) { 8365 assert(_sp->verify_chunk_in_free_list(ffc), 8366 "free range is not in free lists"); 8367 } 8368 _sp->removeFreeChunkFromFreeLists(ffc); 8369 set_freeRangeInFreeLists(false); 8370 } 8371 } 8372 } 8373 // Note that if the chunk is not coalescable (the else arm 8374 // below), we unconditionally flush, without needing to do 8375 // a "lookahead," as we do below. 8376 if (inFreeRange()) lookahead_and_flush(fc, size); 8377 } else { 8378 // Code path common to both original and adaptive free lists. 8379 8380 // cant coalesce with previous block; this should be treated 8381 // as the end of a free run if any 8382 if (inFreeRange()) { 8383 // we kicked some butt; time to pick up the garbage 8384 assert(freeFinger() < addr, "freeFinger points too high"); 8385 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 8386 } 8387 // else, nothing to do, just continue 8388 } 8389 } 8390 8391 size_t SweepClosure::do_garbage_chunk(FreeChunk* fc) { 8392 // This is a chunk of garbage. It is not in any free list. 8393 // Add it to a free list or let it possibly be coalesced into 8394 // a larger chunk. 8395 HeapWord* const addr = (HeapWord*) fc; 8396 const size_t size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 8397 8398 if (_sp->adaptive_freelists()) { 8399 // Verify that the bit map has no bits marked between 8400 // addr and purported end of just dead object. 8401 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 8402 8403 do_post_free_or_garbage_chunk(fc, size); 8404 } else { 8405 if (!inFreeRange()) { 8406 // start of a new free range 8407 assert(size > 0, "A free range should have a size"); 8408 initialize_free_range(addr, false); 8409 } else { 8410 // this will be swept up when we hit the end of the 8411 // free range 8412 if (CMSTraceSweeper) { 8413 gclog_or_tty->print(" -- pick up garbage " PTR_FORMAT " (" SIZE_FORMAT ")\n", fc, size); 8414 } 8415 // If the chunk is being coalesced and the current free range is 8416 // in the free lists, remove the current free range so that it 8417 // will be returned to the free lists in its entirety - all 8418 // the coalesced pieces included. 8419 if (freeRangeInFreeLists()) { 8420 FreeChunk* ffc = (FreeChunk*)freeFinger(); 8421 assert(ffc->size() == pointer_delta(addr, freeFinger()), 8422 "Size of free range is inconsistent with chunk size."); 8423 if (CMSTestInFreeList) { 8424 assert(_sp->verify_chunk_in_free_list(ffc), 8425 "free range is not in free lists"); 8426 } 8427 _sp->removeFreeChunkFromFreeLists(ffc); 8428 set_freeRangeInFreeLists(false); 8429 } 8430 set_lastFreeRangeCoalesced(true); 8431 } 8432 // this will be swept up when we hit the end of the free range 8433 8434 // Verify that the bit map has no bits marked between 8435 // addr and purported end of just dead object. 8436 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 8437 } 8438 assert(_limit >= addr + size, 8439 "A freshly garbage chunk can't possibly straddle over _limit"); 8440 if (inFreeRange()) lookahead_and_flush(fc, size); 8441 return size; 8442 } 8443 8444 size_t SweepClosure::do_live_chunk(FreeChunk* fc) { 8445 HeapWord* addr = (HeapWord*) fc; 8446 // The sweeper has just found a live object. Return any accumulated 8447 // left hand chunk to the free lists. 8448 if (inFreeRange()) { 8449 assert(freeFinger() < addr, "freeFinger points too high"); 8450 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 8451 } 8452 8453 // This object is live: we'd normally expect this to be 8454 // an oop, and like to assert the following: 8455 // assert(oop(addr)->is_oop(), "live block should be an oop"); 8456 // However, as we commented above, this may be an object whose 8457 // header hasn't yet been initialized. 8458 size_t size; 8459 assert(_bitMap->isMarked(addr), "Tautology for this control point"); 8460 if (_bitMap->isMarked(addr + 1)) { 8461 // Determine the size from the bit map, rather than trying to 8462 // compute it from the object header. 8463 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 8464 size = pointer_delta(nextOneAddr + 1, addr); 8465 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 8466 "alignment problem"); 8467 8468 #ifdef ASSERT 8469 if (oop(addr)->klass_or_null() != NULL) { 8470 // Ignore mark word because we are running concurrent with mutators 8471 assert(oop(addr)->is_oop(true), "live block should be an oop"); 8472 assert(size == 8473 CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()), 8474 "P-mark and computed size do not agree"); 8475 } 8476 #endif 8477 8478 } else { 8479 // This should be an initialized object that's alive. 8480 assert(oop(addr)->klass_or_null() != NULL, 8481 "Should be an initialized object"); 8482 // Ignore mark word because we are running concurrent with mutators 8483 assert(oop(addr)->is_oop(true), "live block should be an oop"); 8484 // Verify that the bit map has no bits marked between 8485 // addr and purported end of this block. 8486 size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 8487 assert(size >= 3, "Necessary for Printezis marks to work"); 8488 assert(!_bitMap->isMarked(addr+1), "Tautology for this control point"); 8489 DEBUG_ONLY(_bitMap->verifyNoOneBitsInRange(addr+2, addr+size);) 8490 } 8491 return size; 8492 } 8493 8494 void SweepClosure::do_post_free_or_garbage_chunk(FreeChunk* fc, 8495 size_t chunkSize) { 8496 // do_post_free_or_garbage_chunk() should only be called in the case 8497 // of the adaptive free list allocator. 8498 const bool fcInFreeLists = fc->is_free(); 8499 assert(_sp->adaptive_freelists(), "Should only be used in this case."); 8500 assert((HeapWord*)fc <= _limit, "sweep invariant"); 8501 if (CMSTestInFreeList && fcInFreeLists) { 8502 assert(_sp->verify_chunk_in_free_list(fc), "free chunk is not in free lists"); 8503 } 8504 8505 if (CMSTraceSweeper) { 8506 gclog_or_tty->print_cr(" -- pick up another chunk at " PTR_FORMAT " (" SIZE_FORMAT ")", fc, chunkSize); 8507 } 8508 8509 HeapWord* const fc_addr = (HeapWord*) fc; 8510 8511 bool coalesce; 8512 const size_t left = pointer_delta(fc_addr, freeFinger()); 8513 const size_t right = chunkSize; 8514 switch (FLSCoalescePolicy) { 8515 // numeric value forms a coalition aggressiveness metric 8516 case 0: { // never coalesce 8517 coalesce = false; 8518 break; 8519 } 8520 case 1: { // coalesce if left & right chunks on overpopulated lists 8521 coalesce = _sp->coalOverPopulated(left) && 8522 _sp->coalOverPopulated(right); 8523 break; 8524 } 8525 case 2: { // coalesce if left chunk on overpopulated list (default) 8526 coalesce = _sp->coalOverPopulated(left); 8527 break; 8528 } 8529 case 3: { // coalesce if left OR right chunk on overpopulated list 8530 coalesce = _sp->coalOverPopulated(left) || 8531 _sp->coalOverPopulated(right); 8532 break; 8533 } 8534 case 4: { // always coalesce 8535 coalesce = true; 8536 break; 8537 } 8538 default: 8539 ShouldNotReachHere(); 8540 } 8541 8542 // Should the current free range be coalesced? 8543 // If the chunk is in a free range and either we decided to coalesce above 8544 // or the chunk is near the large block at the end of the heap 8545 // (isNearLargestChunk() returns true), then coalesce this chunk. 8546 const bool doCoalesce = inFreeRange() 8547 && (coalesce || _g->isNearLargestChunk(fc_addr)); 8548 if (doCoalesce) { 8549 // Coalesce the current free range on the left with the new 8550 // chunk on the right. If either is on a free list, 8551 // it must be removed from the list and stashed in the closure. 8552 if (freeRangeInFreeLists()) { 8553 FreeChunk* const ffc = (FreeChunk*)freeFinger(); 8554 assert(ffc->size() == pointer_delta(fc_addr, freeFinger()), 8555 "Size of free range is inconsistent with chunk size."); 8556 if (CMSTestInFreeList) { 8557 assert(_sp->verify_chunk_in_free_list(ffc), 8558 "Chunk is not in free lists"); 8559 } 8560 _sp->coalDeath(ffc->size()); 8561 _sp->removeFreeChunkFromFreeLists(ffc); 8562 set_freeRangeInFreeLists(false); 8563 } 8564 if (fcInFreeLists) { 8565 _sp->coalDeath(chunkSize); 8566 assert(fc->size() == chunkSize, 8567 "The chunk has the wrong size or is not in the free lists"); 8568 _sp->removeFreeChunkFromFreeLists(fc); 8569 } 8570 set_lastFreeRangeCoalesced(true); 8571 print_free_block_coalesced(fc); 8572 } else { // not in a free range and/or should not coalesce 8573 // Return the current free range and start a new one. 8574 if (inFreeRange()) { 8575 // In a free range but cannot coalesce with the right hand chunk. 8576 // Put the current free range into the free lists. 8577 flush_cur_free_chunk(freeFinger(), 8578 pointer_delta(fc_addr, freeFinger())); 8579 } 8580 // Set up for new free range. Pass along whether the right hand 8581 // chunk is in the free lists. 8582 initialize_free_range((HeapWord*)fc, fcInFreeLists); 8583 } 8584 } 8585 8586 // Lookahead flush: 8587 // If we are tracking a free range, and this is the last chunk that 8588 // we'll look at because its end crosses past _limit, we'll preemptively 8589 // flush it along with any free range we may be holding on to. Note that 8590 // this can be the case only for an already free or freshly garbage 8591 // chunk. If this block is an object, it can never straddle 8592 // over _limit. The "straddling" occurs when _limit is set at 8593 // the previous end of the space when this cycle started, and 8594 // a subsequent heap expansion caused the previously co-terminal 8595 // free block to be coalesced with the newly expanded portion, 8596 // thus rendering _limit a non-block-boundary making it dangerous 8597 // for the sweeper to step over and examine. 8598 void SweepClosure::lookahead_and_flush(FreeChunk* fc, size_t chunk_size) { 8599 assert(inFreeRange(), "Should only be called if currently in a free range."); 8600 HeapWord* const eob = ((HeapWord*)fc) + chunk_size; 8601 assert(_sp->used_region().contains(eob - 1), 8602 err_msg("eob = " PTR_FORMAT " eob-1 = " PTR_FORMAT " _limit = " PTR_FORMAT 8603 " out of bounds wrt _sp = [" PTR_FORMAT "," PTR_FORMAT ")" 8604 " when examining fc = " PTR_FORMAT "(" SIZE_FORMAT ")", 8605 eob, eob-1, _limit, _sp->bottom(), _sp->end(), fc, chunk_size)); 8606 if (eob >= _limit) { 8607 assert(eob == _limit || fc->is_free(), "Only a free chunk should allow us to cross over the limit"); 8608 if (CMSTraceSweeper) { 8609 gclog_or_tty->print_cr("_limit " PTR_FORMAT " reached or crossed by block " 8610 "[" PTR_FORMAT "," PTR_FORMAT ") in space " 8611 "[" PTR_FORMAT "," PTR_FORMAT ")", 8612 _limit, fc, eob, _sp->bottom(), _sp->end()); 8613 } 8614 // Return the storage we are tracking back into the free lists. 8615 if (CMSTraceSweeper) { 8616 gclog_or_tty->print_cr("Flushing ... "); 8617 } 8618 assert(freeFinger() < eob, "Error"); 8619 flush_cur_free_chunk( freeFinger(), pointer_delta(eob, freeFinger())); 8620 } 8621 } 8622 8623 void SweepClosure::flush_cur_free_chunk(HeapWord* chunk, size_t size) { 8624 assert(inFreeRange(), "Should only be called if currently in a free range."); 8625 assert(size > 0, 8626 "A zero sized chunk cannot be added to the free lists."); 8627 if (!freeRangeInFreeLists()) { 8628 if (CMSTestInFreeList) { 8629 FreeChunk* fc = (FreeChunk*) chunk; 8630 fc->set_size(size); 8631 assert(!_sp->verify_chunk_in_free_list(fc), 8632 "chunk should not be in free lists yet"); 8633 } 8634 if (CMSTraceSweeper) { 8635 gclog_or_tty->print_cr(" -- add free block " PTR_FORMAT " (" SIZE_FORMAT ") to free lists", 8636 chunk, size); 8637 } 8638 // A new free range is going to be starting. The current 8639 // free range has not been added to the free lists yet or 8640 // was removed so add it back. 8641 // If the current free range was coalesced, then the death 8642 // of the free range was recorded. Record a birth now. 8643 if (lastFreeRangeCoalesced()) { 8644 _sp->coalBirth(size); 8645 } 8646 _sp->addChunkAndRepairOffsetTable(chunk, size, 8647 lastFreeRangeCoalesced()); 8648 } else if (CMSTraceSweeper) { 8649 gclog_or_tty->print_cr("Already in free list: nothing to flush"); 8650 } 8651 set_inFreeRange(false); 8652 set_freeRangeInFreeLists(false); 8653 } 8654 8655 // We take a break if we've been at this for a while, 8656 // so as to avoid monopolizing the locks involved. 8657 void SweepClosure::do_yield_work(HeapWord* addr) { 8658 // Return current free chunk being used for coalescing (if any) 8659 // to the appropriate freelist. After yielding, the next 8660 // free block encountered will start a coalescing range of 8661 // free blocks. If the next free block is adjacent to the 8662 // chunk just flushed, they will need to wait for the next 8663 // sweep to be coalesced. 8664 if (inFreeRange()) { 8665 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 8666 } 8667 8668 // First give up the locks, then yield, then re-lock. 8669 // We should probably use a constructor/destructor idiom to 8670 // do this unlock/lock or modify the MutexUnlocker class to 8671 // serve our purpose. XXX 8672 assert_lock_strong(_bitMap->lock()); 8673 assert_lock_strong(_freelistLock); 8674 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 8675 "CMS thread should hold CMS token"); 8676 _bitMap->lock()->unlock(); 8677 _freelistLock->unlock(); 8678 ConcurrentMarkSweepThread::desynchronize(true); 8679 ConcurrentMarkSweepThread::acknowledge_yield_request(); 8680 _collector->stopTimer(); 8681 if (PrintCMSStatistics != 0) { 8682 _collector->incrementYields(); 8683 } 8684 _collector->icms_wait(); 8685 8686 // See the comment in coordinator_yield() 8687 for (unsigned i = 0; i < CMSYieldSleepCount && 8688 ConcurrentMarkSweepThread::should_yield() && 8689 !CMSCollector::foregroundGCIsActive(); ++i) { 8690 os::sleep(Thread::current(), 1, false); 8691 ConcurrentMarkSweepThread::acknowledge_yield_request(); 8692 } 8693 8694 ConcurrentMarkSweepThread::synchronize(true); 8695 _freelistLock->lock(); 8696 _bitMap->lock()->lock_without_safepoint_check(); 8697 _collector->startTimer(); 8698 } 8699 8700 #ifndef PRODUCT 8701 // This is actually very useful in a product build if it can 8702 // be called from the debugger. Compile it into the product 8703 // as needed. 8704 bool debug_verify_chunk_in_free_list(FreeChunk* fc) { 8705 return debug_cms_space->verify_chunk_in_free_list(fc); 8706 } 8707 #endif 8708 8709 void SweepClosure::print_free_block_coalesced(FreeChunk* fc) const { 8710 if (CMSTraceSweeper) { 8711 gclog_or_tty->print_cr("Sweep:coal_free_blk " PTR_FORMAT " (" SIZE_FORMAT ")", 8712 fc, fc->size()); 8713 } 8714 } 8715 8716 // CMSIsAliveClosure 8717 bool CMSIsAliveClosure::do_object_b(oop obj) { 8718 HeapWord* addr = (HeapWord*)obj; 8719 return addr != NULL && 8720 (!_span.contains(addr) || _bit_map->isMarked(addr)); 8721 } 8722 8723 8724 CMSKeepAliveClosure::CMSKeepAliveClosure( CMSCollector* collector, 8725 MemRegion span, 8726 CMSBitMap* bit_map, CMSMarkStack* mark_stack, 8727 bool cpc): 8728 _collector(collector), 8729 _span(span), 8730 _bit_map(bit_map), 8731 _mark_stack(mark_stack), 8732 _concurrent_precleaning(cpc) { 8733 assert(!_span.is_empty(), "Empty span could spell trouble"); 8734 } 8735 8736 8737 // CMSKeepAliveClosure: the serial version 8738 void CMSKeepAliveClosure::do_oop(oop obj) { 8739 HeapWord* addr = (HeapWord*)obj; 8740 if (_span.contains(addr) && 8741 !_bit_map->isMarked(addr)) { 8742 _bit_map->mark(addr); 8743 bool simulate_overflow = false; 8744 NOT_PRODUCT( 8745 if (CMSMarkStackOverflowALot && 8746 _collector->simulate_overflow()) { 8747 // simulate a stack overflow 8748 simulate_overflow = true; 8749 } 8750 ) 8751 if (simulate_overflow || !_mark_stack->push(obj)) { 8752 if (_concurrent_precleaning) { 8753 // We dirty the overflown object and let the remark 8754 // phase deal with it. 8755 assert(_collector->overflow_list_is_empty(), "Error"); 8756 // In the case of object arrays, we need to dirty all of 8757 // the cards that the object spans. No locking or atomics 8758 // are needed since no one else can be mutating the mod union 8759 // table. 8760 if (obj->is_objArray()) { 8761 size_t sz = obj->size(); 8762 HeapWord* end_card_addr = 8763 (HeapWord*)round_to((intptr_t)(addr+sz), CardTableModRefBS::card_size); 8764 MemRegion redirty_range = MemRegion(addr, end_card_addr); 8765 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 8766 _collector->_modUnionTable.mark_range(redirty_range); 8767 } else { 8768 _collector->_modUnionTable.mark(addr); 8769 } 8770 _collector->_ser_kac_preclean_ovflw++; 8771 } else { 8772 _collector->push_on_overflow_list(obj); 8773 _collector->_ser_kac_ovflw++; 8774 } 8775 } 8776 } 8777 } 8778 8779 void CMSKeepAliveClosure::do_oop(oop* p) { CMSKeepAliveClosure::do_oop_work(p); } 8780 void CMSKeepAliveClosure::do_oop(narrowOop* p) { CMSKeepAliveClosure::do_oop_work(p); } 8781 8782 // CMSParKeepAliveClosure: a parallel version of the above. 8783 // The work queues are private to each closure (thread), 8784 // but (may be) available for stealing by other threads. 8785 void CMSParKeepAliveClosure::do_oop(oop obj) { 8786 HeapWord* addr = (HeapWord*)obj; 8787 if (_span.contains(addr) && 8788 !_bit_map->isMarked(addr)) { 8789 // In general, during recursive tracing, several threads 8790 // may be concurrently getting here; the first one to 8791 // "tag" it, claims it. 8792 if (_bit_map->par_mark(addr)) { 8793 bool res = _work_queue->push(obj); 8794 assert(res, "Low water mark should be much less than capacity"); 8795 // Do a recursive trim in the hope that this will keep 8796 // stack usage lower, but leave some oops for potential stealers 8797 trim_queue(_low_water_mark); 8798 } // Else, another thread got there first 8799 } 8800 } 8801 8802 void CMSParKeepAliveClosure::do_oop(oop* p) { CMSParKeepAliveClosure::do_oop_work(p); } 8803 void CMSParKeepAliveClosure::do_oop(narrowOop* p) { CMSParKeepAliveClosure::do_oop_work(p); } 8804 8805 void CMSParKeepAliveClosure::trim_queue(uint max) { 8806 while (_work_queue->size() > max) { 8807 oop new_oop; 8808 if (_work_queue->pop_local(new_oop)) { 8809 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 8810 assert(_bit_map->isMarked((HeapWord*)new_oop), 8811 "no white objects on this stack!"); 8812 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 8813 // iterate over the oops in this oop, marking and pushing 8814 // the ones in CMS heap (i.e. in _span). 8815 new_oop->oop_iterate(&_mark_and_push); 8816 } 8817 } 8818 } 8819 8820 CMSInnerParMarkAndPushClosure::CMSInnerParMarkAndPushClosure( 8821 CMSCollector* collector, 8822 MemRegion span, CMSBitMap* bit_map, 8823 OopTaskQueue* work_queue): 8824 _collector(collector), 8825 _span(span), 8826 _bit_map(bit_map), 8827 _work_queue(work_queue) { } 8828 8829 void CMSInnerParMarkAndPushClosure::do_oop(oop obj) { 8830 HeapWord* addr = (HeapWord*)obj; 8831 if (_span.contains(addr) && 8832 !_bit_map->isMarked(addr)) { 8833 if (_bit_map->par_mark(addr)) { 8834 bool simulate_overflow = false; 8835 NOT_PRODUCT( 8836 if (CMSMarkStackOverflowALot && 8837 _collector->par_simulate_overflow()) { 8838 // simulate a stack overflow 8839 simulate_overflow = true; 8840 } 8841 ) 8842 if (simulate_overflow || !_work_queue->push(obj)) { 8843 _collector->par_push_on_overflow_list(obj); 8844 _collector->_par_kac_ovflw++; 8845 } 8846 } // Else another thread got there already 8847 } 8848 } 8849 8850 void CMSInnerParMarkAndPushClosure::do_oop(oop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); } 8851 void CMSInnerParMarkAndPushClosure::do_oop(narrowOop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); } 8852 8853 ////////////////////////////////////////////////////////////////// 8854 // CMSExpansionCause ///////////////////////////// 8855 ////////////////////////////////////////////////////////////////// 8856 const char* CMSExpansionCause::to_string(CMSExpansionCause::Cause cause) { 8857 switch (cause) { 8858 case _no_expansion: 8859 return "No expansion"; 8860 case _satisfy_free_ratio: 8861 return "Free ratio"; 8862 case _satisfy_promotion: 8863 return "Satisfy promotion"; 8864 case _satisfy_allocation: 8865 return "allocation"; 8866 case _allocate_par_lab: 8867 return "Par LAB"; 8868 case _allocate_par_spooling_space: 8869 return "Par Spooling Space"; 8870 case _adaptive_size_policy: 8871 return "Ergonomics"; 8872 default: 8873 return "unknown"; 8874 } 8875 } 8876 8877 void CMSDrainMarkingStackClosure::do_void() { 8878 // the max number to take from overflow list at a time 8879 const size_t num = _mark_stack->capacity()/4; 8880 assert(!_concurrent_precleaning || _collector->overflow_list_is_empty(), 8881 "Overflow list should be NULL during concurrent phases"); 8882 while (!_mark_stack->isEmpty() || 8883 // if stack is empty, check the overflow list 8884 _collector->take_from_overflow_list(num, _mark_stack)) { 8885 oop obj = _mark_stack->pop(); 8886 HeapWord* addr = (HeapWord*)obj; 8887 assert(_span.contains(addr), "Should be within span"); 8888 assert(_bit_map->isMarked(addr), "Should be marked"); 8889 assert(obj->is_oop(), "Should be an oop"); 8890 obj->oop_iterate(_keep_alive); 8891 } 8892 } 8893 8894 void CMSParDrainMarkingStackClosure::do_void() { 8895 // drain queue 8896 trim_queue(0); 8897 } 8898 8899 // Trim our work_queue so its length is below max at return 8900 void CMSParDrainMarkingStackClosure::trim_queue(uint max) { 8901 while (_work_queue->size() > max) { 8902 oop new_oop; 8903 if (_work_queue->pop_local(new_oop)) { 8904 assert(new_oop->is_oop(), "Expected an oop"); 8905 assert(_bit_map->isMarked((HeapWord*)new_oop), 8906 "no white objects on this stack!"); 8907 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 8908 // iterate over the oops in this oop, marking and pushing 8909 // the ones in CMS heap (i.e. in _span). 8910 new_oop->oop_iterate(&_mark_and_push); 8911 } 8912 } 8913 } 8914 8915 //////////////////////////////////////////////////////////////////// 8916 // Support for Marking Stack Overflow list handling and related code 8917 //////////////////////////////////////////////////////////////////// 8918 // Much of the following code is similar in shape and spirit to the 8919 // code used in ParNewGC. We should try and share that code 8920 // as much as possible in the future. 8921 8922 #ifndef PRODUCT 8923 // Debugging support for CMSStackOverflowALot 8924 8925 // It's OK to call this multi-threaded; the worst thing 8926 // that can happen is that we'll get a bunch of closely 8927 // spaced simulated overflows, but that's OK, in fact 8928 // probably good as it would exercise the overflow code 8929 // under contention. 8930 bool CMSCollector::simulate_overflow() { 8931 if (_overflow_counter-- <= 0) { // just being defensive 8932 _overflow_counter = CMSMarkStackOverflowInterval; 8933 return true; 8934 } else { 8935 return false; 8936 } 8937 } 8938 8939 bool CMSCollector::par_simulate_overflow() { 8940 return simulate_overflow(); 8941 } 8942 #endif 8943 8944 // Single-threaded 8945 bool CMSCollector::take_from_overflow_list(size_t num, CMSMarkStack* stack) { 8946 assert(stack->isEmpty(), "Expected precondition"); 8947 assert(stack->capacity() > num, "Shouldn't bite more than can chew"); 8948 size_t i = num; 8949 oop cur = _overflow_list; 8950 const markOop proto = markOopDesc::prototype(); 8951 NOT_PRODUCT(ssize_t n = 0;) 8952 for (oop next; i > 0 && cur != NULL; cur = next, i--) { 8953 next = oop(cur->mark()); 8954 cur->set_mark(proto); // until proven otherwise 8955 assert(cur->is_oop(), "Should be an oop"); 8956 bool res = stack->push(cur); 8957 assert(res, "Bit off more than can chew?"); 8958 NOT_PRODUCT(n++;) 8959 } 8960 _overflow_list = cur; 8961 #ifndef PRODUCT 8962 assert(_num_par_pushes >= n, "Too many pops?"); 8963 _num_par_pushes -=n; 8964 #endif 8965 return !stack->isEmpty(); 8966 } 8967 8968 #define BUSY (cast_to_oop<intptr_t>(0x1aff1aff)) 8969 // (MT-safe) Get a prefix of at most "num" from the list. 8970 // The overflow list is chained through the mark word of 8971 // each object in the list. We fetch the entire list, 8972 // break off a prefix of the right size and return the 8973 // remainder. If other threads try to take objects from 8974 // the overflow list at that time, they will wait for 8975 // some time to see if data becomes available. If (and 8976 // only if) another thread places one or more object(s) 8977 // on the global list before we have returned the suffix 8978 // to the global list, we will walk down our local list 8979 // to find its end and append the global list to 8980 // our suffix before returning it. This suffix walk can 8981 // prove to be expensive (quadratic in the amount of traffic) 8982 // when there are many objects in the overflow list and 8983 // there is much producer-consumer contention on the list. 8984 // *NOTE*: The overflow list manipulation code here and 8985 // in ParNewGeneration:: are very similar in shape, 8986 // except that in the ParNew case we use the old (from/eden) 8987 // copy of the object to thread the list via its klass word. 8988 // Because of the common code, if you make any changes in 8989 // the code below, please check the ParNew version to see if 8990 // similar changes might be needed. 8991 // CR 6797058 has been filed to consolidate the common code. 8992 bool CMSCollector::par_take_from_overflow_list(size_t num, 8993 OopTaskQueue* work_q, 8994 int no_of_gc_threads) { 8995 assert(work_q->size() == 0, "First empty local work queue"); 8996 assert(num < work_q->max_elems(), "Can't bite more than we can chew"); 8997 if (_overflow_list == NULL) { 8998 return false; 8999 } 9000 // Grab the entire list; we'll put back a suffix 9001 oop prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list)); 9002 Thread* tid = Thread::current(); 9003 // Before "no_of_gc_threads" was introduced CMSOverflowSpinCount was 9004 // set to ParallelGCThreads. 9005 size_t CMSOverflowSpinCount = (size_t) no_of_gc_threads; // was ParallelGCThreads; 9006 size_t sleep_time_millis = MAX2((size_t)1, num/100); 9007 // If the list is busy, we spin for a short while, 9008 // sleeping between attempts to get the list. 9009 for (size_t spin = 0; prefix == BUSY && spin < CMSOverflowSpinCount; spin++) { 9010 os::sleep(tid, sleep_time_millis, false); 9011 if (_overflow_list == NULL) { 9012 // Nothing left to take 9013 return false; 9014 } else if (_overflow_list != BUSY) { 9015 // Try and grab the prefix 9016 prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list)); 9017 } 9018 } 9019 // If the list was found to be empty, or we spun long 9020 // enough, we give up and return empty-handed. If we leave 9021 // the list in the BUSY state below, it must be the case that 9022 // some other thread holds the overflow list and will set it 9023 // to a non-BUSY state in the future. 9024 if (prefix == NULL || prefix == BUSY) { 9025 // Nothing to take or waited long enough 9026 if (prefix == NULL) { 9027 // Write back the NULL in case we overwrote it with BUSY above 9028 // and it is still the same value. 9029 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY); 9030 } 9031 return false; 9032 } 9033 assert(prefix != NULL && prefix != BUSY, "Error"); 9034 size_t i = num; 9035 oop cur = prefix; 9036 // Walk down the first "num" objects, unless we reach the end. 9037 for (; i > 1 && cur->mark() != NULL; cur = oop(cur->mark()), i--); 9038 if (cur->mark() == NULL) { 9039 // We have "num" or fewer elements in the list, so there 9040 // is nothing to return to the global list. 9041 // Write back the NULL in lieu of the BUSY we wrote 9042 // above, if it is still the same value. 9043 if (_overflow_list == BUSY) { 9044 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY); 9045 } 9046 } else { 9047 // Chop off the suffix and return it to the global list. 9048 assert(cur->mark() != BUSY, "Error"); 9049 oop suffix_head = cur->mark(); // suffix will be put back on global list 9050 cur->set_mark(NULL); // break off suffix 9051 // It's possible that the list is still in the empty(busy) state 9052 // we left it in a short while ago; in that case we may be 9053 // able to place back the suffix without incurring the cost 9054 // of a walk down the list. 9055 oop observed_overflow_list = _overflow_list; 9056 oop cur_overflow_list = observed_overflow_list; 9057 bool attached = false; 9058 while (observed_overflow_list == BUSY || observed_overflow_list == NULL) { 9059 observed_overflow_list = 9060 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list); 9061 if (cur_overflow_list == observed_overflow_list) { 9062 attached = true; 9063 break; 9064 } else cur_overflow_list = observed_overflow_list; 9065 } 9066 if (!attached) { 9067 // Too bad, someone else sneaked in (at least) an element; we'll need 9068 // to do a splice. Find tail of suffix so we can prepend suffix to global 9069 // list. 9070 for (cur = suffix_head; cur->mark() != NULL; cur = (oop)(cur->mark())); 9071 oop suffix_tail = cur; 9072 assert(suffix_tail != NULL && suffix_tail->mark() == NULL, 9073 "Tautology"); 9074 observed_overflow_list = _overflow_list; 9075 do { 9076 cur_overflow_list = observed_overflow_list; 9077 if (cur_overflow_list != BUSY) { 9078 // Do the splice ... 9079 suffix_tail->set_mark(markOop(cur_overflow_list)); 9080 } else { // cur_overflow_list == BUSY 9081 suffix_tail->set_mark(NULL); 9082 } 9083 // ... and try to place spliced list back on overflow_list ... 9084 observed_overflow_list = 9085 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list); 9086 } while (cur_overflow_list != observed_overflow_list); 9087 // ... until we have succeeded in doing so. 9088 } 9089 } 9090 9091 // Push the prefix elements on work_q 9092 assert(prefix != NULL, "control point invariant"); 9093 const markOop proto = markOopDesc::prototype(); 9094 oop next; 9095 NOT_PRODUCT(ssize_t n = 0;) 9096 for (cur = prefix; cur != NULL; cur = next) { 9097 next = oop(cur->mark()); 9098 cur->set_mark(proto); // until proven otherwise 9099 assert(cur->is_oop(), "Should be an oop"); 9100 bool res = work_q->push(cur); 9101 assert(res, "Bit off more than we can chew?"); 9102 NOT_PRODUCT(n++;) 9103 } 9104 #ifndef PRODUCT 9105 assert(_num_par_pushes >= n, "Too many pops?"); 9106 Atomic::add_ptr(-(intptr_t)n, &_num_par_pushes); 9107 #endif 9108 return true; 9109 } 9110 9111 // Single-threaded 9112 void CMSCollector::push_on_overflow_list(oop p) { 9113 NOT_PRODUCT(_num_par_pushes++;) 9114 assert(p->is_oop(), "Not an oop"); 9115 preserve_mark_if_necessary(p); 9116 p->set_mark((markOop)_overflow_list); 9117 _overflow_list = p; 9118 } 9119 9120 // Multi-threaded; use CAS to prepend to overflow list 9121 void CMSCollector::par_push_on_overflow_list(oop p) { 9122 NOT_PRODUCT(Atomic::inc_ptr(&_num_par_pushes);) 9123 assert(p->is_oop(), "Not an oop"); 9124 par_preserve_mark_if_necessary(p); 9125 oop observed_overflow_list = _overflow_list; 9126 oop cur_overflow_list; 9127 do { 9128 cur_overflow_list = observed_overflow_list; 9129 if (cur_overflow_list != BUSY) { 9130 p->set_mark(markOop(cur_overflow_list)); 9131 } else { 9132 p->set_mark(NULL); 9133 } 9134 observed_overflow_list = 9135 (oop) Atomic::cmpxchg_ptr(p, &_overflow_list, cur_overflow_list); 9136 } while (cur_overflow_list != observed_overflow_list); 9137 } 9138 #undef BUSY 9139 9140 // Single threaded 9141 // General Note on GrowableArray: pushes may silently fail 9142 // because we are (temporarily) out of C-heap for expanding 9143 // the stack. The problem is quite ubiquitous and affects 9144 // a lot of code in the JVM. The prudent thing for GrowableArray 9145 // to do (for now) is to exit with an error. However, that may 9146 // be too draconian in some cases because the caller may be 9147 // able to recover without much harm. For such cases, we 9148 // should probably introduce a "soft_push" method which returns 9149 // an indication of success or failure with the assumption that 9150 // the caller may be able to recover from a failure; code in 9151 // the VM can then be changed, incrementally, to deal with such 9152 // failures where possible, thus, incrementally hardening the VM 9153 // in such low resource situations. 9154 void CMSCollector::preserve_mark_work(oop p, markOop m) { 9155 _preserved_oop_stack.push(p); 9156 _preserved_mark_stack.push(m); 9157 assert(m == p->mark(), "Mark word changed"); 9158 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), 9159 "bijection"); 9160 } 9161 9162 // Single threaded 9163 void CMSCollector::preserve_mark_if_necessary(oop p) { 9164 markOop m = p->mark(); 9165 if (m->must_be_preserved(p)) { 9166 preserve_mark_work(p, m); 9167 } 9168 } 9169 9170 void CMSCollector::par_preserve_mark_if_necessary(oop p) { 9171 markOop m = p->mark(); 9172 if (m->must_be_preserved(p)) { 9173 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag); 9174 // Even though we read the mark word without holding 9175 // the lock, we are assured that it will not change 9176 // because we "own" this oop, so no other thread can 9177 // be trying to push it on the overflow list; see 9178 // the assertion in preserve_mark_work() that checks 9179 // that m == p->mark(). 9180 preserve_mark_work(p, m); 9181 } 9182 } 9183 9184 // We should be able to do this multi-threaded, 9185 // a chunk of stack being a task (this is 9186 // correct because each oop only ever appears 9187 // once in the overflow list. However, it's 9188 // not very easy to completely overlap this with 9189 // other operations, so will generally not be done 9190 // until all work's been completed. Because we 9191 // expect the preserved oop stack (set) to be small, 9192 // it's probably fine to do this single-threaded. 9193 // We can explore cleverer concurrent/overlapped/parallel 9194 // processing of preserved marks if we feel the 9195 // need for this in the future. Stack overflow should 9196 // be so rare in practice and, when it happens, its 9197 // effect on performance so great that this will 9198 // likely just be in the noise anyway. 9199 void CMSCollector::restore_preserved_marks_if_any() { 9200 assert(SafepointSynchronize::is_at_safepoint(), 9201 "world should be stopped"); 9202 assert(Thread::current()->is_ConcurrentGC_thread() || 9203 Thread::current()->is_VM_thread(), 9204 "should be single-threaded"); 9205 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), 9206 "bijection"); 9207 9208 while (!_preserved_oop_stack.is_empty()) { 9209 oop p = _preserved_oop_stack.pop(); 9210 assert(p->is_oop(), "Should be an oop"); 9211 assert(_span.contains(p), "oop should be in _span"); 9212 assert(p->mark() == markOopDesc::prototype(), 9213 "Set when taken from overflow list"); 9214 markOop m = _preserved_mark_stack.pop(); 9215 p->set_mark(m); 9216 } 9217 assert(_preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(), 9218 "stacks were cleared above"); 9219 } 9220 9221 #ifndef PRODUCT 9222 bool CMSCollector::no_preserved_marks() const { 9223 return _preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(); 9224 } 9225 #endif 9226 9227 // Transfer some number of overflown objects to usual marking 9228 // stack. Return true if some objects were transferred. 9229 bool MarkRefsIntoAndScanClosure::take_from_overflow_list() { 9230 size_t num = MIN2((size_t)(_mark_stack->capacity() - _mark_stack->length())/4, 9231 (size_t)ParGCDesiredObjsFromOverflowList); 9232 9233 bool res = _collector->take_from_overflow_list(num, _mark_stack); 9234 assert(_collector->overflow_list_is_empty() || res, 9235 "If list is not empty, we should have taken something"); 9236 assert(!res || !_mark_stack->isEmpty(), 9237 "If we took something, it should now be on our stack"); 9238 return res; 9239 } 9240 9241 size_t MarkDeadObjectsClosure::do_blk(HeapWord* addr) { 9242 size_t res = _sp->block_size_no_stall(addr, _collector); 9243 if (_sp->block_is_obj(addr)) { 9244 if (_live_bit_map->isMarked(addr)) { 9245 // It can't have been dead in a previous cycle 9246 guarantee(!_dead_bit_map->isMarked(addr), "No resurrection!"); 9247 } else { 9248 _dead_bit_map->mark(addr); // mark the dead object 9249 } 9250 } 9251 // Could be 0, if the block size could not be computed without stalling. 9252 return res; 9253 } 9254 9255 TraceCMSMemoryManagerStats::TraceCMSMemoryManagerStats(CMSCollector::CollectorState phase, GCCause::Cause cause): TraceMemoryManagerStats() { 9256 9257 switch (phase) { 9258 case CMSCollector::InitialMarking: 9259 initialize(true /* fullGC */ , 9260 cause /* cause of the GC */, 9261 true /* recordGCBeginTime */, 9262 true /* recordPreGCUsage */, 9263 false /* recordPeakUsage */, 9264 false /* recordPostGCusage */, 9265 true /* recordAccumulatedGCTime */, 9266 false /* recordGCEndTime */, 9267 false /* countCollection */ ); 9268 break; 9269 9270 case CMSCollector::FinalMarking: 9271 initialize(true /* fullGC */ , 9272 cause /* cause of the GC */, 9273 false /* recordGCBeginTime */, 9274 false /* recordPreGCUsage */, 9275 false /* recordPeakUsage */, 9276 false /* recordPostGCusage */, 9277 true /* recordAccumulatedGCTime */, 9278 false /* recordGCEndTime */, 9279 false /* countCollection */ ); 9280 break; 9281 9282 case CMSCollector::Sweeping: 9283 initialize(true /* fullGC */ , 9284 cause /* cause of the GC */, 9285 false /* recordGCBeginTime */, 9286 false /* recordPreGCUsage */, 9287 true /* recordPeakUsage */, 9288 true /* recordPostGCusage */, 9289 false /* recordAccumulatedGCTime */, 9290 true /* recordGCEndTime */, 9291 true /* countCollection */ ); 9292 break; 9293 9294 default: 9295 ShouldNotReachHere(); 9296 } 9297 }