1 /* 2 * Copyright (c) 2001, 2012, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "classfile/symbolTable.hpp" 27 #include "classfile/systemDictionary.hpp" 28 #include "code/codeCache.hpp" 29 #include "gc_implementation/concurrentMarkSweep/cmsAdaptiveSizePolicy.hpp" 30 #include "gc_implementation/concurrentMarkSweep/cmsCollectorPolicy.hpp" 31 #include "gc_implementation/concurrentMarkSweep/cmsGCAdaptivePolicyCounters.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.hpp" 52 #include "memory/referencePolicy.hpp" 53 #include "memory/resourceArea.hpp" 54 #include "oops/oop.inline.hpp" 55 #include "prims/jvmtiExport.hpp" 56 #include "runtime/globals_extension.hpp" 57 #include "runtime/handles.inline.hpp" 58 #include "runtime/java.hpp" 59 #include "runtime/vmThread.hpp" 60 #include "services/memoryService.hpp" 61 #include "services/runtimeService.hpp" 62 63 // statics 64 CMSCollector* ConcurrentMarkSweepGeneration::_collector = NULL; 65 bool CMSCollector::_full_gc_requested = false; 66 GCCause::Cause CMSCollector::_full_gc_cause = GCCause::_no_gc; 67 68 ////////////////////////////////////////////////////////////////// 69 // In support of CMS/VM thread synchronization 70 ////////////////////////////////////////////////////////////////// 71 // We split use of the CGC_lock into 2 "levels". 72 // The low-level locking is of the usual CGC_lock monitor. We introduce 73 // a higher level "token" (hereafter "CMS token") built on top of the 74 // low level monitor (hereafter "CGC lock"). 75 // The token-passing protocol gives priority to the VM thread. The 76 // CMS-lock doesn't provide any fairness guarantees, but clients 77 // should ensure that it is only held for very short, bounded 78 // durations. 79 // 80 // When either of the CMS thread or the VM thread is involved in 81 // collection operations during which it does not want the other 82 // thread to interfere, it obtains the CMS token. 83 // 84 // If either thread tries to get the token while the other has 85 // it, that thread waits. However, if the VM thread and CMS thread 86 // both want the token, then the VM thread gets priority while the 87 // CMS thread waits. This ensures, for instance, that the "concurrent" 88 // phases of the CMS thread's work do not block out the VM thread 89 // for long periods of time as the CMS thread continues to hog 90 // the token. (See bug 4616232). 91 // 92 // The baton-passing functions are, however, controlled by the 93 // flags _foregroundGCShouldWait and _foregroundGCIsActive, 94 // and here the low-level CMS lock, not the high level token, 95 // ensures mutual exclusion. 96 // 97 // Two important conditions that we have to satisfy: 98 // 1. if a thread does a low-level wait on the CMS lock, then it 99 // relinquishes the CMS token if it were holding that token 100 // when it acquired the low-level CMS lock. 101 // 2. any low-level notifications on the low-level lock 102 // should only be sent when a thread has relinquished the token. 103 // 104 // In the absence of either property, we'd have potential deadlock. 105 // 106 // We protect each of the CMS (concurrent and sequential) phases 107 // with the CMS _token_, not the CMS _lock_. 108 // 109 // The only code protected by CMS lock is the token acquisition code 110 // itself, see ConcurrentMarkSweepThread::[de]synchronize(), and the 111 // baton-passing code. 112 // 113 // Unfortunately, i couldn't come up with a good abstraction to factor and 114 // hide the naked CGC_lock manipulation in the baton-passing code 115 // further below. That's something we should try to do. Also, the proof 116 // of correctness of this 2-level locking scheme is far from obvious, 117 // and potentially quite slippery. We have an uneasy supsicion, for instance, 118 // that there may be a theoretical possibility of delay/starvation in the 119 // low-level lock/wait/notify scheme used for the baton-passing because of 120 // potential intereference with the priority scheme embodied in the 121 // CMS-token-passing protocol. See related comments at a CGC_lock->wait() 122 // invocation further below and marked with "XXX 20011219YSR". 123 // Indeed, as we note elsewhere, this may become yet more slippery 124 // in the presence of multiple CMS and/or multiple VM threads. XXX 125 126 class CMSTokenSync: public StackObj { 127 private: 128 bool _is_cms_thread; 129 public: 130 CMSTokenSync(bool is_cms_thread): 131 _is_cms_thread(is_cms_thread) { 132 assert(is_cms_thread == Thread::current()->is_ConcurrentGC_thread(), 133 "Incorrect argument to constructor"); 134 ConcurrentMarkSweepThread::synchronize(_is_cms_thread); 135 } 136 137 ~CMSTokenSync() { 138 assert(_is_cms_thread ? 139 ConcurrentMarkSweepThread::cms_thread_has_cms_token() : 140 ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 141 "Incorrect state"); 142 ConcurrentMarkSweepThread::desynchronize(_is_cms_thread); 143 } 144 }; 145 146 // Convenience class that does a CMSTokenSync, and then acquires 147 // upto three locks. 148 class CMSTokenSyncWithLocks: public CMSTokenSync { 149 private: 150 // Note: locks are acquired in textual declaration order 151 // and released in the opposite order 152 MutexLockerEx _locker1, _locker2, _locker3; 153 public: 154 CMSTokenSyncWithLocks(bool is_cms_thread, Mutex* mutex1, 155 Mutex* mutex2 = NULL, Mutex* mutex3 = NULL): 156 CMSTokenSync(is_cms_thread), 157 _locker1(mutex1, Mutex::_no_safepoint_check_flag), 158 _locker2(mutex2, Mutex::_no_safepoint_check_flag), 159 _locker3(mutex3, Mutex::_no_safepoint_check_flag) 160 { } 161 }; 162 163 164 // Wrapper class to temporarily disable icms during a foreground cms collection. 165 class ICMSDisabler: public StackObj { 166 public: 167 // The ctor disables icms and wakes up the thread so it notices the change; 168 // the dtor re-enables icms. Note that the CMSCollector methods will check 169 // CMSIncrementalMode. 170 ICMSDisabler() { CMSCollector::disable_icms(); CMSCollector::start_icms(); } 171 ~ICMSDisabler() { CMSCollector::enable_icms(); } 172 }; 173 174 ////////////////////////////////////////////////////////////////// 175 // Concurrent Mark-Sweep Generation ///////////////////////////// 176 ////////////////////////////////////////////////////////////////// 177 178 NOT_PRODUCT(CompactibleFreeListSpace* debug_cms_space;) 179 180 // This struct contains per-thread things necessary to support parallel 181 // young-gen collection. 182 class CMSParGCThreadState: public CHeapObj<mtGC> { 183 public: 184 CFLS_LAB lab; 185 PromotionInfo promo; 186 187 // Constructor. 188 CMSParGCThreadState(CompactibleFreeListSpace* cfls) : lab(cfls) { 189 promo.setSpace(cfls); 190 } 191 }; 192 193 ConcurrentMarkSweepGeneration::ConcurrentMarkSweepGeneration( 194 ReservedSpace rs, size_t initial_byte_size, int level, 195 CardTableRS* ct, bool use_adaptive_freelists, 196 FreeBlockDictionary<FreeChunk>::DictionaryChoice dictionaryChoice) : 197 CardGeneration(rs, initial_byte_size, level, ct), 198 _dilatation_factor(((double)MinChunkSize)/((double)(CollectedHeap::min_fill_size()))), 199 _debug_collection_type(Concurrent_collection_type) 200 { 201 HeapWord* bottom = (HeapWord*) _virtual_space.low(); 202 HeapWord* end = (HeapWord*) _virtual_space.high(); 203 204 _direct_allocated_words = 0; 205 NOT_PRODUCT( 206 _numObjectsPromoted = 0; 207 _numWordsPromoted = 0; 208 _numObjectsAllocated = 0; 209 _numWordsAllocated = 0; 210 ) 211 212 _cmsSpace = new CompactibleFreeListSpace(_bts, MemRegion(bottom, end), 213 use_adaptive_freelists, 214 dictionaryChoice); 215 NOT_PRODUCT(debug_cms_space = _cmsSpace;) 216 if (_cmsSpace == NULL) { 217 vm_exit_during_initialization( 218 "CompactibleFreeListSpace allocation failure"); 219 } 220 _cmsSpace->_gen = this; 221 222 _gc_stats = new CMSGCStats(); 223 224 // Verify the assumption that FreeChunk::_prev and OopDesc::_klass 225 // offsets match. The ability to tell free chunks from objects 226 // depends on this property. 227 debug_only( 228 FreeChunk* junk = NULL; 229 assert(UseCompressedOops || 230 junk->prev_addr() == (void*)(oop(junk)->klass_addr()), 231 "Offset of FreeChunk::_prev within FreeChunk must match" 232 " that of OopDesc::_klass within OopDesc"); 233 ) 234 if (CollectedHeap::use_parallel_gc_threads()) { 235 typedef CMSParGCThreadState* CMSParGCThreadStatePtr; 236 _par_gc_thread_states = 237 NEW_C_HEAP_ARRAY(CMSParGCThreadStatePtr, ParallelGCThreads, mtGC); 238 if (_par_gc_thread_states == NULL) { 239 vm_exit_during_initialization("Could not allocate par gc structs"); 240 } 241 for (uint i = 0; i < ParallelGCThreads; i++) { 242 _par_gc_thread_states[i] = new CMSParGCThreadState(cmsSpace()); 243 if (_par_gc_thread_states[i] == NULL) { 244 vm_exit_during_initialization("Could not allocate par gc structs"); 245 } 246 } 247 } else { 248 _par_gc_thread_states = NULL; 249 } 250 _incremental_collection_failed = false; 251 // The "dilatation_factor" is the expansion that can occur on 252 // account of the fact that the minimum object size in the CMS 253 // generation may be larger than that in, say, a contiguous young 254 // generation. 255 // Ideally, in the calculation below, we'd compute the dilatation 256 // factor as: MinChunkSize/(promoting_gen's min object size) 257 // Since we do not have such a general query interface for the 258 // promoting generation, we'll instead just use the mimimum 259 // object size (which today is a header's worth of space); 260 // note that all arithmetic is in units of HeapWords. 261 assert(MinChunkSize >= CollectedHeap::min_fill_size(), "just checking"); 262 assert(_dilatation_factor >= 1.0, "from previous assert"); 263 } 264 265 266 // The field "_initiating_occupancy" represents the occupancy percentage 267 // at which we trigger a new collection cycle. Unless explicitly specified 268 // via CMSInitiating[Perm]OccupancyFraction (argument "io" below), it 269 // is calculated by: 270 // 271 // Let "f" be MinHeapFreeRatio in 272 // 273 // _intiating_occupancy = 100-f + 274 // f * (CMSTrigger[Perm]Ratio/100) 275 // where CMSTrigger[Perm]Ratio is the argument "tr" below. 276 // 277 // That is, if we assume the heap is at its desired maximum occupancy at the 278 // end of a collection, we let CMSTrigger[Perm]Ratio of the (purported) free 279 // space be allocated before initiating a new collection cycle. 280 // 281 void ConcurrentMarkSweepGeneration::init_initiating_occupancy(intx io, intx tr) { 282 assert(io <= 100 && tr >= 0 && tr <= 100, "Check the arguments"); 283 if (io >= 0) { 284 _initiating_occupancy = (double)io / 100.0; 285 } else { 286 _initiating_occupancy = ((100 - MinHeapFreeRatio) + 287 (double)(tr * MinHeapFreeRatio) / 100.0) 288 / 100.0; 289 } 290 } 291 292 void ConcurrentMarkSweepGeneration::ref_processor_init() { 293 assert(collector() != NULL, "no collector"); 294 collector()->ref_processor_init(); 295 } 296 297 void CMSCollector::ref_processor_init() { 298 if (_ref_processor == NULL) { 299 // Allocate and initialize a reference processor 300 _ref_processor = 301 new ReferenceProcessor(_span, // span 302 (ParallelGCThreads > 1) && ParallelRefProcEnabled, // mt processing 303 (int) ParallelGCThreads, // mt processing degree 304 _cmsGen->refs_discovery_is_mt(), // mt discovery 305 (int) MAX2(ConcGCThreads, ParallelGCThreads), // mt discovery degree 306 _cmsGen->refs_discovery_is_atomic(), // discovery is not atomic 307 &_is_alive_closure, // closure for liveness info 308 false); // next field updates do not need write barrier 309 // Initialize the _ref_processor field of CMSGen 310 _cmsGen->set_ref_processor(_ref_processor); 311 312 // Allocate a dummy ref processor for perm gen. 313 ReferenceProcessor* rp2 = new ReferenceProcessor(); 314 if (rp2 == NULL) { 315 vm_exit_during_initialization("Could not allocate ReferenceProcessor object"); 316 } 317 _permGen->set_ref_processor(rp2); 318 } 319 } 320 321 CMSAdaptiveSizePolicy* CMSCollector::size_policy() { 322 GenCollectedHeap* gch = GenCollectedHeap::heap(); 323 assert(gch->kind() == CollectedHeap::GenCollectedHeap, 324 "Wrong type of heap"); 325 CMSAdaptiveSizePolicy* sp = (CMSAdaptiveSizePolicy*) 326 gch->gen_policy()->size_policy(); 327 assert(sp->is_gc_cms_adaptive_size_policy(), 328 "Wrong type of size policy"); 329 return sp; 330 } 331 332 CMSGCAdaptivePolicyCounters* CMSCollector::gc_adaptive_policy_counters() { 333 CMSGCAdaptivePolicyCounters* results = 334 (CMSGCAdaptivePolicyCounters*) collector_policy()->counters(); 335 assert( 336 results->kind() == GCPolicyCounters::CMSGCAdaptivePolicyCountersKind, 337 "Wrong gc policy counter kind"); 338 return results; 339 } 340 341 342 void ConcurrentMarkSweepGeneration::initialize_performance_counters() { 343 344 const char* gen_name = "old"; 345 346 // Generation Counters - generation 1, 1 subspace 347 _gen_counters = new GenerationCounters(gen_name, 1, 1, &_virtual_space); 348 349 _space_counters = new GSpaceCounters(gen_name, 0, 350 _virtual_space.reserved_size(), 351 this, _gen_counters); 352 } 353 354 CMSStats::CMSStats(ConcurrentMarkSweepGeneration* cms_gen, unsigned int alpha): 355 _cms_gen(cms_gen) 356 { 357 assert(alpha <= 100, "bad value"); 358 _saved_alpha = alpha; 359 360 // Initialize the alphas to the bootstrap value of 100. 361 _gc0_alpha = _cms_alpha = 100; 362 363 _cms_begin_time.update(); 364 _cms_end_time.update(); 365 366 _gc0_duration = 0.0; 367 _gc0_period = 0.0; 368 _gc0_promoted = 0; 369 370 _cms_duration = 0.0; 371 _cms_period = 0.0; 372 _cms_allocated = 0; 373 374 _cms_used_at_gc0_begin = 0; 375 _cms_used_at_gc0_end = 0; 376 _allow_duty_cycle_reduction = false; 377 _valid_bits = 0; 378 _icms_duty_cycle = CMSIncrementalDutyCycle; 379 } 380 381 double CMSStats::cms_free_adjustment_factor(size_t free) const { 382 // TBD: CR 6909490 383 return 1.0; 384 } 385 386 void CMSStats::adjust_cms_free_adjustment_factor(bool fail, size_t free) { 387 } 388 389 // If promotion failure handling is on use 390 // the padded average size of the promotion for each 391 // young generation collection. 392 double CMSStats::time_until_cms_gen_full() const { 393 size_t cms_free = _cms_gen->cmsSpace()->free(); 394 GenCollectedHeap* gch = GenCollectedHeap::heap(); 395 size_t expected_promotion = MIN2(gch->get_gen(0)->capacity(), 396 (size_t) _cms_gen->gc_stats()->avg_promoted()->padded_average()); 397 if (cms_free > expected_promotion) { 398 // Start a cms collection if there isn't enough space to promote 399 // for the next minor collection. Use the padded average as 400 // a safety factor. 401 cms_free -= expected_promotion; 402 403 // Adjust by the safety factor. 404 double cms_free_dbl = (double)cms_free; 405 double cms_adjustment = (100.0 - CMSIncrementalSafetyFactor)/100.0; 406 // Apply a further correction factor which tries to adjust 407 // for recent occurance of concurrent mode failures. 408 cms_adjustment = cms_adjustment * cms_free_adjustment_factor(cms_free); 409 cms_free_dbl = cms_free_dbl * cms_adjustment; 410 411 if (PrintGCDetails && Verbose) { 412 gclog_or_tty->print_cr("CMSStats::time_until_cms_gen_full: cms_free " 413 SIZE_FORMAT " expected_promotion " SIZE_FORMAT, 414 cms_free, expected_promotion); 415 gclog_or_tty->print_cr(" cms_free_dbl %f cms_consumption_rate %f", 416 cms_free_dbl, cms_consumption_rate() + 1.0); 417 } 418 // Add 1 in case the consumption rate goes to zero. 419 return cms_free_dbl / (cms_consumption_rate() + 1.0); 420 } 421 return 0.0; 422 } 423 424 // Compare the duration of the cms collection to the 425 // time remaining before the cms generation is empty. 426 // Note that the time from the start of the cms collection 427 // to the start of the cms sweep (less than the total 428 // duration of the cms collection) can be used. This 429 // has been tried and some applications experienced 430 // promotion failures early in execution. This was 431 // possibly because the averages were not accurate 432 // enough at the beginning. 433 double CMSStats::time_until_cms_start() const { 434 // We add "gc0_period" to the "work" calculation 435 // below because this query is done (mostly) at the 436 // end of a scavenge, so we need to conservatively 437 // account for that much possible delay 438 // in the query so as to avoid concurrent mode failures 439 // due to starting the collection just a wee bit too 440 // late. 441 double work = cms_duration() + gc0_period(); 442 double deadline = time_until_cms_gen_full(); 443 // If a concurrent mode failure occurred recently, we want to be 444 // more conservative and halve our expected time_until_cms_gen_full() 445 if (work > deadline) { 446 if (Verbose && PrintGCDetails) { 447 gclog_or_tty->print( 448 " CMSCollector: collect because of anticipated promotion " 449 "before full %3.7f + %3.7f > %3.7f ", cms_duration(), 450 gc0_period(), time_until_cms_gen_full()); 451 } 452 return 0.0; 453 } 454 return work - deadline; 455 } 456 457 // Return a duty cycle based on old_duty_cycle and new_duty_cycle, limiting the 458 // amount of change to prevent wild oscillation. 459 unsigned int CMSStats::icms_damped_duty_cycle(unsigned int old_duty_cycle, 460 unsigned int new_duty_cycle) { 461 assert(old_duty_cycle <= 100, "bad input value"); 462 assert(new_duty_cycle <= 100, "bad input value"); 463 464 // Note: use subtraction with caution since it may underflow (values are 465 // unsigned). Addition is safe since we're in the range 0-100. 466 unsigned int damped_duty_cycle = new_duty_cycle; 467 if (new_duty_cycle < old_duty_cycle) { 468 const unsigned int largest_delta = MAX2(old_duty_cycle / 4, 5U); 469 if (new_duty_cycle + largest_delta < old_duty_cycle) { 470 damped_duty_cycle = old_duty_cycle - largest_delta; 471 } 472 } else if (new_duty_cycle > old_duty_cycle) { 473 const unsigned int largest_delta = MAX2(old_duty_cycle / 4, 15U); 474 if (new_duty_cycle > old_duty_cycle + largest_delta) { 475 damped_duty_cycle = MIN2(old_duty_cycle + largest_delta, 100U); 476 } 477 } 478 assert(damped_duty_cycle <= 100, "invalid duty cycle computed"); 479 480 if (CMSTraceIncrementalPacing) { 481 gclog_or_tty->print(" [icms_damped_duty_cycle(%d,%d) = %d] ", 482 old_duty_cycle, new_duty_cycle, damped_duty_cycle); 483 } 484 return damped_duty_cycle; 485 } 486 487 unsigned int CMSStats::icms_update_duty_cycle_impl() { 488 assert(CMSIncrementalPacing && valid(), 489 "should be handled in icms_update_duty_cycle()"); 490 491 double cms_time_so_far = cms_timer().seconds(); 492 double scaled_duration = cms_duration_per_mb() * _cms_used_at_gc0_end / M; 493 double scaled_duration_remaining = fabsd(scaled_duration - cms_time_so_far); 494 495 // Avoid division by 0. 496 double time_until_full = MAX2(time_until_cms_gen_full(), 0.01); 497 double duty_cycle_dbl = 100.0 * scaled_duration_remaining / time_until_full; 498 499 unsigned int new_duty_cycle = MIN2((unsigned int)duty_cycle_dbl, 100U); 500 if (new_duty_cycle > _icms_duty_cycle) { 501 // Avoid very small duty cycles (1 or 2); 0 is allowed. 502 if (new_duty_cycle > 2) { 503 _icms_duty_cycle = icms_damped_duty_cycle(_icms_duty_cycle, 504 new_duty_cycle); 505 } 506 } else if (_allow_duty_cycle_reduction) { 507 // The duty cycle is reduced only once per cms cycle (see record_cms_end()). 508 new_duty_cycle = icms_damped_duty_cycle(_icms_duty_cycle, new_duty_cycle); 509 // Respect the minimum duty cycle. 510 unsigned int min_duty_cycle = (unsigned int)CMSIncrementalDutyCycleMin; 511 _icms_duty_cycle = MAX2(new_duty_cycle, min_duty_cycle); 512 } 513 514 if (PrintGCDetails || CMSTraceIncrementalPacing) { 515 gclog_or_tty->print(" icms_dc=%d ", _icms_duty_cycle); 516 } 517 518 _allow_duty_cycle_reduction = false; 519 return _icms_duty_cycle; 520 } 521 522 #ifndef PRODUCT 523 void CMSStats::print_on(outputStream *st) const { 524 st->print(" gc0_alpha=%d,cms_alpha=%d", _gc0_alpha, _cms_alpha); 525 st->print(",gc0_dur=%g,gc0_per=%g,gc0_promo=" SIZE_FORMAT, 526 gc0_duration(), gc0_period(), gc0_promoted()); 527 st->print(",cms_dur=%g,cms_dur_per_mb=%g,cms_per=%g,cms_alloc=" SIZE_FORMAT, 528 cms_duration(), cms_duration_per_mb(), 529 cms_period(), cms_allocated()); 530 st->print(",cms_since_beg=%g,cms_since_end=%g", 531 cms_time_since_begin(), cms_time_since_end()); 532 st->print(",cms_used_beg=" SIZE_FORMAT ",cms_used_end=" SIZE_FORMAT, 533 _cms_used_at_gc0_begin, _cms_used_at_gc0_end); 534 if (CMSIncrementalMode) { 535 st->print(",dc=%d", icms_duty_cycle()); 536 } 537 538 if (valid()) { 539 st->print(",promo_rate=%g,cms_alloc_rate=%g", 540 promotion_rate(), cms_allocation_rate()); 541 st->print(",cms_consumption_rate=%g,time_until_full=%g", 542 cms_consumption_rate(), time_until_cms_gen_full()); 543 } 544 st->print(" "); 545 } 546 #endif // #ifndef PRODUCT 547 548 CMSCollector::CollectorState CMSCollector::_collectorState = 549 CMSCollector::Idling; 550 bool CMSCollector::_foregroundGCIsActive = false; 551 bool CMSCollector::_foregroundGCShouldWait = false; 552 553 CMSCollector::CMSCollector(ConcurrentMarkSweepGeneration* cmsGen, 554 ConcurrentMarkSweepGeneration* permGen, 555 CardTableRS* ct, 556 ConcurrentMarkSweepPolicy* cp): 557 _cmsGen(cmsGen), 558 _permGen(permGen), 559 _ct(ct), 560 _ref_processor(NULL), // will be set later 561 _conc_workers(NULL), // may be set later 562 _abort_preclean(false), 563 _start_sampling(false), 564 _between_prologue_and_epilogue(false), 565 _markBitMap(0, Mutex::leaf + 1, "CMS_markBitMap_lock"), 566 _perm_gen_verify_bit_map(0, -1 /* no mutex */, "No_lock"), 567 _modUnionTable((CardTableModRefBS::card_shift - LogHeapWordSize), 568 -1 /* lock-free */, "No_lock" /* dummy */), 569 _modUnionClosure(&_modUnionTable), 570 _modUnionClosurePar(&_modUnionTable), 571 // Adjust my span to cover old (cms) gen and perm gen 572 _span(cmsGen->reserved()._union(permGen->reserved())), 573 // Construct the is_alive_closure with _span & markBitMap 574 _is_alive_closure(_span, &_markBitMap), 575 _restart_addr(NULL), 576 _overflow_list(NULL), 577 _stats(cmsGen), 578 _eden_chunk_lock(new Mutex(Mutex::leaf + 1, "CMS_eden_chunk_lock", true)), 579 _eden_chunk_array(NULL), // may be set in ctor body 580 _eden_chunk_capacity(0), // -- ditto -- 581 _eden_chunk_index(0), // -- ditto -- 582 _survivor_plab_array(NULL), // -- ditto -- 583 _survivor_chunk_array(NULL), // -- ditto -- 584 _survivor_chunk_capacity(0), // -- ditto -- 585 _survivor_chunk_index(0), // -- ditto -- 586 _ser_pmc_preclean_ovflw(0), 587 _ser_kac_preclean_ovflw(0), 588 _ser_pmc_remark_ovflw(0), 589 _par_pmc_remark_ovflw(0), 590 _ser_kac_ovflw(0), 591 _par_kac_ovflw(0), 592 #ifndef PRODUCT 593 _num_par_pushes(0), 594 #endif 595 _collection_count_start(0), 596 _verifying(false), 597 _icms_start_limit(NULL), 598 _icms_stop_limit(NULL), 599 _verification_mark_bm(0, Mutex::leaf + 1, "CMS_verification_mark_bm_lock"), 600 _completed_initialization(false), 601 _collector_policy(cp), 602 _should_unload_classes(false), 603 _concurrent_cycles_since_last_unload(0), 604 _roots_scanning_options(0), 605 _inter_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding), 606 _intra_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding), 607 _gc_tracer_cm(new (ResourceObj::C_HEAP, mtGC) CMSTracer()), 608 _gc_timer_cm(new (ResourceObj::C_HEAP, mtGC) ConcurrentGCTimer()), 609 _cms_start_registered(false) 610 { 611 if (ExplicitGCInvokesConcurrentAndUnloadsClasses) { 612 ExplicitGCInvokesConcurrent = true; 613 } 614 // Now expand the span and allocate the collection support structures 615 // (MUT, marking bit map etc.) to cover both generations subject to 616 // collection. 617 618 // First check that _permGen is adjacent to _cmsGen and above it. 619 assert( _cmsGen->reserved().word_size() > 0 620 && _permGen->reserved().word_size() > 0, 621 "generations should not be of zero size"); 622 assert(_cmsGen->reserved().intersection(_permGen->reserved()).is_empty(), 623 "_cmsGen and _permGen should not overlap"); 624 assert(_cmsGen->reserved().end() == _permGen->reserved().start(), 625 "_cmsGen->end() different from _permGen->start()"); 626 627 // For use by dirty card to oop closures. 628 _cmsGen->cmsSpace()->set_collector(this); 629 _permGen->cmsSpace()->set_collector(this); 630 631 // Allocate MUT and marking bit map 632 { 633 MutexLockerEx x(_markBitMap.lock(), Mutex::_no_safepoint_check_flag); 634 if (!_markBitMap.allocate(_span)) { 635 warning("Failed to allocate CMS Bit Map"); 636 return; 637 } 638 assert(_markBitMap.covers(_span), "_markBitMap inconsistency?"); 639 } 640 { 641 _modUnionTable.allocate(_span); 642 assert(_modUnionTable.covers(_span), "_modUnionTable inconsistency?"); 643 } 644 645 if (!_markStack.allocate(MarkStackSize)) { 646 warning("Failed to allocate CMS Marking Stack"); 647 return; 648 } 649 if (!_revisitStack.allocate(CMSRevisitStackSize)) { 650 warning("Failed to allocate CMS Revisit Stack"); 651 return; 652 } 653 654 // Support for multi-threaded concurrent phases 655 if (CMSConcurrentMTEnabled) { 656 if (FLAG_IS_DEFAULT(ConcGCThreads)) { 657 // just for now 658 FLAG_SET_DEFAULT(ConcGCThreads, (ParallelGCThreads + 3)/4); 659 } 660 if (ConcGCThreads > 1) { 661 _conc_workers = new YieldingFlexibleWorkGang("Parallel CMS Threads", 662 ConcGCThreads, true); 663 if (_conc_workers == NULL) { 664 warning("GC/CMS: _conc_workers allocation failure: " 665 "forcing -CMSConcurrentMTEnabled"); 666 CMSConcurrentMTEnabled = false; 667 } else { 668 _conc_workers->initialize_workers(); 669 } 670 } else { 671 CMSConcurrentMTEnabled = false; 672 } 673 } 674 if (!CMSConcurrentMTEnabled) { 675 ConcGCThreads = 0; 676 } else { 677 // Turn off CMSCleanOnEnter optimization temporarily for 678 // the MT case where it's not fixed yet; see 6178663. 679 CMSCleanOnEnter = false; 680 } 681 assert((_conc_workers != NULL) == (ConcGCThreads > 1), 682 "Inconsistency"); 683 684 // Parallel task queues; these are shared for the 685 // concurrent and stop-world phases of CMS, but 686 // are not shared with parallel scavenge (ParNew). 687 { 688 uint i; 689 uint num_queues = (uint) MAX2(ParallelGCThreads, ConcGCThreads); 690 691 if ((CMSParallelRemarkEnabled || CMSConcurrentMTEnabled 692 || ParallelRefProcEnabled) 693 && num_queues > 0) { 694 _task_queues = new OopTaskQueueSet(num_queues); 695 if (_task_queues == NULL) { 696 warning("task_queues allocation failure."); 697 return; 698 } 699 _hash_seed = NEW_C_HEAP_ARRAY(int, num_queues, mtGC); 700 if (_hash_seed == NULL) { 701 warning("_hash_seed array allocation failure"); 702 return; 703 } 704 705 typedef Padded<OopTaskQueue> PaddedOopTaskQueue; 706 for (i = 0; i < num_queues; i++) { 707 PaddedOopTaskQueue *q = new PaddedOopTaskQueue(); 708 if (q == NULL) { 709 warning("work_queue allocation failure."); 710 return; 711 } 712 _task_queues->register_queue(i, q); 713 } 714 for (i = 0; i < num_queues; i++) { 715 _task_queues->queue(i)->initialize(); 716 _hash_seed[i] = 17; // copied from ParNew 717 } 718 } 719 } 720 721 _cmsGen ->init_initiating_occupancy(CMSInitiatingOccupancyFraction, CMSTriggerRatio); 722 _permGen->init_initiating_occupancy(CMSInitiatingPermOccupancyFraction, CMSTriggerPermRatio); 723 724 // Clip CMSBootstrapOccupancy between 0 and 100. 725 _bootstrap_occupancy = ((double)MIN2((uintx)100, MAX2((uintx)0, CMSBootstrapOccupancy))) 726 /(double)100; 727 728 _full_gcs_since_conc_gc = 0; 729 730 // Now tell CMS generations the identity of their collector 731 ConcurrentMarkSweepGeneration::set_collector(this); 732 733 // Create & start a CMS thread for this CMS collector 734 _cmsThread = ConcurrentMarkSweepThread::start(this); 735 assert(cmsThread() != NULL, "CMS Thread should have been created"); 736 assert(cmsThread()->collector() == this, 737 "CMS Thread should refer to this gen"); 738 assert(CGC_lock != NULL, "Where's the CGC_lock?"); 739 740 // Support for parallelizing young gen rescan 741 GenCollectedHeap* gch = GenCollectedHeap::heap(); 742 _young_gen = gch->prev_gen(_cmsGen); 743 if (gch->supports_inline_contig_alloc()) { 744 _top_addr = gch->top_addr(); 745 _end_addr = gch->end_addr(); 746 assert(_young_gen != NULL, "no _young_gen"); 747 _eden_chunk_index = 0; 748 _eden_chunk_capacity = (_young_gen->max_capacity()+CMSSamplingGrain)/CMSSamplingGrain; 749 _eden_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, _eden_chunk_capacity, mtGC); 750 if (_eden_chunk_array == NULL) { 751 _eden_chunk_capacity = 0; 752 warning("GC/CMS: _eden_chunk_array allocation failure"); 753 } 754 } 755 assert(_eden_chunk_array != NULL || _eden_chunk_capacity == 0, "Error"); 756 757 // Support for parallelizing survivor space rescan 758 if ((CMSParallelRemarkEnabled && CMSParallelSurvivorRemarkEnabled) || CMSParallelInitialMarkEnabled) { 759 const size_t max_plab_samples = 760 ((DefNewGeneration*)_young_gen)->max_survivor_size()/MinTLABSize; 761 762 _survivor_plab_array = NEW_C_HEAP_ARRAY(ChunkArray, ParallelGCThreads, mtGC); 763 _survivor_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, 2*max_plab_samples, mtGC); 764 _cursor = NEW_C_HEAP_ARRAY(size_t, ParallelGCThreads, mtGC); 765 if (_survivor_plab_array == NULL || _survivor_chunk_array == NULL 766 || _cursor == NULL) { 767 warning("Failed to allocate survivor plab/chunk array"); 768 if (_survivor_plab_array != NULL) { 769 FREE_C_HEAP_ARRAY(ChunkArray, _survivor_plab_array, mtGC); 770 _survivor_plab_array = NULL; 771 } 772 if (_survivor_chunk_array != NULL) { 773 FREE_C_HEAP_ARRAY(HeapWord*, _survivor_chunk_array, mtGC); 774 _survivor_chunk_array = NULL; 775 } 776 if (_cursor != NULL) { 777 FREE_C_HEAP_ARRAY(size_t, _cursor, mtGC); 778 _cursor = NULL; 779 } 780 } else { 781 _survivor_chunk_capacity = 2*max_plab_samples; 782 for (uint i = 0; i < ParallelGCThreads; i++) { 783 HeapWord** vec = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples, mtGC); 784 if (vec == NULL) { 785 warning("Failed to allocate survivor plab array"); 786 for (int j = i; j > 0; j--) { 787 FREE_C_HEAP_ARRAY(HeapWord*, _survivor_plab_array[j-1].array(), mtGC); 788 } 789 FREE_C_HEAP_ARRAY(ChunkArray, _survivor_plab_array, mtGC); 790 FREE_C_HEAP_ARRAY(HeapWord*, _survivor_chunk_array, mtGC); 791 _survivor_plab_array = NULL; 792 _survivor_chunk_array = NULL; 793 _survivor_chunk_capacity = 0; 794 break; 795 } else { 796 ChunkArray* cur = 797 ::new (&_survivor_plab_array[i]) ChunkArray(vec, 798 max_plab_samples); 799 assert(cur->end() == 0, "Should be 0"); 800 assert(cur->array() == vec, "Should be vec"); 801 assert(cur->capacity() == max_plab_samples, "Error"); 802 } 803 } 804 } 805 } 806 assert( ( _survivor_plab_array != NULL 807 && _survivor_chunk_array != NULL) 808 || ( _survivor_chunk_capacity == 0 809 && _survivor_chunk_index == 0), 810 "Error"); 811 812 // Choose what strong roots should be scanned depending on verification options 813 // and perm gen collection mode. 814 if (!CMSClassUnloadingEnabled) { 815 // If class unloading is disabled we want to include all classes into the root set. 816 add_root_scanning_option(SharedHeap::SO_AllClasses); 817 } else { 818 add_root_scanning_option(SharedHeap::SO_SystemClasses); 819 } 820 821 NOT_PRODUCT(_overflow_counter = CMSMarkStackOverflowInterval;) 822 _gc_counters = new CollectorCounters("CMS", 1); 823 _completed_initialization = true; 824 _inter_sweep_timer.start(); // start of time 825 } 826 827 const char* ConcurrentMarkSweepGeneration::name() const { 828 return "concurrent mark-sweep generation"; 829 } 830 void ConcurrentMarkSweepGeneration::update_counters() { 831 if (UsePerfData) { 832 _space_counters->update_all(); 833 _gen_counters->update_all(); 834 } 835 } 836 837 // this is an optimized version of update_counters(). it takes the 838 // used value as a parameter rather than computing it. 839 // 840 void ConcurrentMarkSweepGeneration::update_counters(size_t used) { 841 if (UsePerfData) { 842 _space_counters->update_used(used); 843 _space_counters->update_capacity(); 844 _gen_counters->update_all(); 845 } 846 } 847 848 void ConcurrentMarkSweepGeneration::print() const { 849 Generation::print(); 850 cmsSpace()->print(); 851 } 852 853 #ifndef PRODUCT 854 void ConcurrentMarkSweepGeneration::print_statistics() { 855 cmsSpace()->printFLCensus(0); 856 } 857 #endif 858 859 void ConcurrentMarkSweepGeneration::printOccupancy(const char *s) { 860 GenCollectedHeap* gch = GenCollectedHeap::heap(); 861 if (PrintGCDetails) { 862 if (Verbose) { 863 gclog_or_tty->print(" [%d %s-%s: "SIZE_FORMAT"("SIZE_FORMAT")]", 864 level(), short_name(), s, used(), capacity()); 865 } else { 866 gclog_or_tty->print(" [%d %s-%s: "SIZE_FORMAT"K("SIZE_FORMAT"K)]", 867 level(), short_name(), s, used() / K, capacity() / K); 868 } 869 } 870 if (Verbose) { 871 gclog_or_tty->print(" "SIZE_FORMAT"("SIZE_FORMAT")", 872 gch->used(), gch->capacity()); 873 } else { 874 gclog_or_tty->print(" "SIZE_FORMAT"K("SIZE_FORMAT"K)", 875 gch->used() / K, gch->capacity() / K); 876 } 877 } 878 879 size_t 880 ConcurrentMarkSweepGeneration::contiguous_available() const { 881 // dld proposes an improvement in precision here. If the committed 882 // part of the space ends in a free block we should add that to 883 // uncommitted size in the calculation below. Will make this 884 // change later, staying with the approximation below for the 885 // time being. -- ysr. 886 return MAX2(_virtual_space.uncommitted_size(), unsafe_max_alloc_nogc()); 887 } 888 889 size_t 890 ConcurrentMarkSweepGeneration::unsafe_max_alloc_nogc() const { 891 return _cmsSpace->max_alloc_in_words() * HeapWordSize; 892 } 893 894 size_t ConcurrentMarkSweepGeneration::max_available() const { 895 return free() + _virtual_space.uncommitted_size(); 896 } 897 898 bool ConcurrentMarkSweepGeneration::promotion_attempt_is_safe(size_t max_promotion_in_bytes) const { 899 size_t available = max_available(); 900 size_t av_promo = (size_t)gc_stats()->avg_promoted()->padded_average(); 901 bool res = (available >= av_promo) || (available >= max_promotion_in_bytes); 902 if (Verbose && PrintGCDetails) { 903 gclog_or_tty->print_cr( 904 "CMS: promo attempt is%s safe: available("SIZE_FORMAT") %s av_promo("SIZE_FORMAT")," 905 "max_promo("SIZE_FORMAT")", 906 res? "":" not", available, res? ">=":"<", 907 av_promo, max_promotion_in_bytes); 908 } 909 return res; 910 } 911 912 // At a promotion failure dump information on block layout in heap 913 // (cms old generation). 914 void ConcurrentMarkSweepGeneration::promotion_failure_occurred() { 915 if (CMSDumpAtPromotionFailure) { 916 cmsSpace()->dump_at_safepoint_with_locks(collector(), gclog_or_tty); 917 } 918 } 919 920 CompactibleSpace* 921 ConcurrentMarkSweepGeneration::first_compaction_space() const { 922 return _cmsSpace; 923 } 924 925 void ConcurrentMarkSweepGeneration::reset_after_compaction() { 926 // Clear the promotion information. These pointers can be adjusted 927 // along with all the other pointers into the heap but 928 // compaction is expected to be a rare event with 929 // a heap using cms so don't do it without seeing the need. 930 if (CollectedHeap::use_parallel_gc_threads()) { 931 for (uint i = 0; i < ParallelGCThreads; i++) { 932 _par_gc_thread_states[i]->promo.reset(); 933 } 934 } 935 } 936 937 void ConcurrentMarkSweepGeneration::space_iterate(SpaceClosure* blk, bool usedOnly) { 938 blk->do_space(_cmsSpace); 939 } 940 941 void ConcurrentMarkSweepGeneration::compute_new_size() { 942 assert_locked_or_safepoint(Heap_lock); 943 944 // If incremental collection failed, we just want to expand 945 // to the limit. 946 if (incremental_collection_failed()) { 947 clear_incremental_collection_failed(); 948 grow_to_reserved(); 949 return; 950 } 951 952 size_t expand_bytes = 0; 953 double free_percentage = ((double) free()) / capacity(); 954 double desired_free_percentage = (double) MinHeapFreeRatio / 100; 955 double maximum_free_percentage = (double) MaxHeapFreeRatio / 100; 956 957 // compute expansion delta needed for reaching desired free percentage 958 if (free_percentage < desired_free_percentage) { 959 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 960 assert(desired_capacity >= capacity(), "invalid expansion size"); 961 expand_bytes = MAX2(desired_capacity - capacity(), MinHeapDeltaBytes); 962 } 963 if (expand_bytes > 0) { 964 if (PrintGCDetails && Verbose) { 965 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 966 gclog_or_tty->print_cr("\nFrom compute_new_size: "); 967 gclog_or_tty->print_cr(" Free fraction %f", free_percentage); 968 gclog_or_tty->print_cr(" Desired free fraction %f", 969 desired_free_percentage); 970 gclog_or_tty->print_cr(" Maximum free fraction %f", 971 maximum_free_percentage); 972 gclog_or_tty->print_cr(" Capactiy "SIZE_FORMAT, capacity()/1000); 973 gclog_or_tty->print_cr(" Desired capacity "SIZE_FORMAT, 974 desired_capacity/1000); 975 int prev_level = level() - 1; 976 if (prev_level >= 0) { 977 size_t prev_size = 0; 978 GenCollectedHeap* gch = GenCollectedHeap::heap(); 979 Generation* prev_gen = gch->_gens[prev_level]; 980 prev_size = prev_gen->capacity(); 981 gclog_or_tty->print_cr(" Younger gen size "SIZE_FORMAT, 982 prev_size/1000); 983 } 984 gclog_or_tty->print_cr(" unsafe_max_alloc_nogc "SIZE_FORMAT, 985 unsafe_max_alloc_nogc()/1000); 986 gclog_or_tty->print_cr(" contiguous available "SIZE_FORMAT, 987 contiguous_available()/1000); 988 gclog_or_tty->print_cr(" Expand by "SIZE_FORMAT" (bytes)", 989 expand_bytes); 990 } 991 // safe if expansion fails 992 expand(expand_bytes, 0, CMSExpansionCause::_satisfy_free_ratio); 993 if (PrintGCDetails && Verbose) { 994 gclog_or_tty->print_cr(" Expanded free fraction %f", 995 ((double) free()) / capacity()); 996 } 997 } 998 } 999 1000 Mutex* ConcurrentMarkSweepGeneration::freelistLock() const { 1001 return cmsSpace()->freelistLock(); 1002 } 1003 1004 HeapWord* ConcurrentMarkSweepGeneration::allocate(size_t size, 1005 bool tlab) { 1006 CMSSynchronousYieldRequest yr; 1007 MutexLockerEx x(freelistLock(), 1008 Mutex::_no_safepoint_check_flag); 1009 return have_lock_and_allocate(size, tlab); 1010 } 1011 1012 HeapWord* ConcurrentMarkSweepGeneration::have_lock_and_allocate(size_t size, 1013 bool tlab /* ignored */) { 1014 assert_lock_strong(freelistLock()); 1015 size_t adjustedSize = CompactibleFreeListSpace::adjustObjectSize(size); 1016 HeapWord* res = cmsSpace()->allocate(adjustedSize); 1017 // Allocate the object live (grey) if the background collector has 1018 // started marking. This is necessary because the marker may 1019 // have passed this address and consequently this object will 1020 // not otherwise be greyed and would be incorrectly swept up. 1021 // Note that if this object contains references, the writing 1022 // of those references will dirty the card containing this object 1023 // allowing the object to be blackened (and its references scanned) 1024 // either during a preclean phase or at the final checkpoint. 1025 if (res != NULL) { 1026 // We may block here with an uninitialized object with 1027 // its mark-bit or P-bits not yet set. Such objects need 1028 // to be safely navigable by block_start(). 1029 assert(oop(res)->klass_or_null() == NULL, "Object should be uninitialized here."); 1030 assert(!((FreeChunk*)res)->is_free(), "Error, block will look free but show wrong size"); 1031 collector()->direct_allocated(res, adjustedSize); 1032 _direct_allocated_words += adjustedSize; 1033 // allocation counters 1034 NOT_PRODUCT( 1035 _numObjectsAllocated++; 1036 _numWordsAllocated += (int)adjustedSize; 1037 ) 1038 } 1039 return res; 1040 } 1041 1042 // In the case of direct allocation by mutators in a generation that 1043 // is being concurrently collected, the object must be allocated 1044 // live (grey) if the background collector has started marking. 1045 // This is necessary because the marker may 1046 // have passed this address and consequently this object will 1047 // not otherwise be greyed and would be incorrectly swept up. 1048 // Note that if this object contains references, the writing 1049 // of those references will dirty the card containing this object 1050 // allowing the object to be blackened (and its references scanned) 1051 // either during a preclean phase or at the final checkpoint. 1052 void CMSCollector::direct_allocated(HeapWord* start, size_t size) { 1053 assert(_markBitMap.covers(start, size), "Out of bounds"); 1054 if (_collectorState >= Marking) { 1055 MutexLockerEx y(_markBitMap.lock(), 1056 Mutex::_no_safepoint_check_flag); 1057 // [see comments preceding SweepClosure::do_blk() below for details] 1058 // 1. need to mark the object as live so it isn't collected 1059 // 2. need to mark the 2nd bit to indicate the object may be uninitialized 1060 // 3. need to mark the end of the object so marking, precleaning or sweeping 1061 // can skip over uninitialized or unparsable objects. An allocated 1062 // object is considered uninitialized for our purposes as long as 1063 // its klass word is NULL. (Unparsable objects are those which are 1064 // initialized in the sense just described, but whose sizes can still 1065 // not be correctly determined. Note that the class of unparsable objects 1066 // can only occur in the perm gen. All old gen objects are parsable 1067 // as soon as they are initialized.) 1068 _markBitMap.mark(start); // object is live 1069 _markBitMap.mark(start + 1); // object is potentially uninitialized? 1070 _markBitMap.mark(start + size - 1); 1071 // mark end of object 1072 } 1073 // check that oop looks uninitialized 1074 assert(oop(start)->klass_or_null() == NULL, "_klass should be NULL"); 1075 } 1076 1077 void CMSCollector::promoted(bool par, HeapWord* start, 1078 bool is_obj_array, size_t obj_size) { 1079 assert(_markBitMap.covers(start), "Out of bounds"); 1080 // See comment in direct_allocated() about when objects should 1081 // be allocated live. 1082 if (_collectorState >= Marking) { 1083 // we already hold the marking bit map lock, taken in 1084 // the prologue 1085 if (par) { 1086 _markBitMap.par_mark(start); 1087 } else { 1088 _markBitMap.mark(start); 1089 } 1090 // We don't need to mark the object as uninitialized (as 1091 // in direct_allocated above) because this is being done with the 1092 // world stopped and the object will be initialized by the 1093 // time the marking, precleaning or sweeping get to look at it. 1094 // But see the code for copying objects into the CMS generation, 1095 // where we need to ensure that concurrent readers of the 1096 // block offset table are able to safely navigate a block that 1097 // is in flux from being free to being allocated (and in 1098 // transition while being copied into) and subsequently 1099 // becoming a bona-fide object when the copy/promotion is complete. 1100 assert(SafepointSynchronize::is_at_safepoint(), 1101 "expect promotion only at safepoints"); 1102 1103 if (_collectorState < Sweeping) { 1104 // Mark the appropriate cards in the modUnionTable, so that 1105 // this object gets scanned before the sweep. If this is 1106 // not done, CMS generation references in the object might 1107 // not get marked. 1108 // For the case of arrays, which are otherwise precisely 1109 // marked, we need to dirty the entire array, not just its head. 1110 if (is_obj_array) { 1111 // The [par_]mark_range() method expects mr.end() below to 1112 // be aligned to the granularity of a bit's representation 1113 // in the heap. In the case of the MUT below, that's a 1114 // card size. 1115 MemRegion mr(start, 1116 (HeapWord*)round_to((intptr_t)(start + obj_size), 1117 CardTableModRefBS::card_size /* bytes */)); 1118 if (par) { 1119 _modUnionTable.par_mark_range(mr); 1120 } else { 1121 _modUnionTable.mark_range(mr); 1122 } 1123 } else { // not an obj array; we can just mark the head 1124 if (par) { 1125 _modUnionTable.par_mark(start); 1126 } else { 1127 _modUnionTable.mark(start); 1128 } 1129 } 1130 } 1131 } 1132 } 1133 1134 static inline size_t percent_of_space(Space* space, HeapWord* addr) 1135 { 1136 size_t delta = pointer_delta(addr, space->bottom()); 1137 return (size_t)(delta * 100.0 / (space->capacity() / HeapWordSize)); 1138 } 1139 1140 void CMSCollector::icms_update_allocation_limits() 1141 { 1142 Generation* gen0 = GenCollectedHeap::heap()->get_gen(0); 1143 EdenSpace* eden = gen0->as_DefNewGeneration()->eden(); 1144 1145 const unsigned int duty_cycle = stats().icms_update_duty_cycle(); 1146 if (CMSTraceIncrementalPacing) { 1147 stats().print(); 1148 } 1149 1150 assert(duty_cycle <= 100, "invalid duty cycle"); 1151 if (duty_cycle != 0) { 1152 // The duty_cycle is a percentage between 0 and 100; convert to words and 1153 // then compute the offset from the endpoints of the space. 1154 size_t free_words = eden->free() / HeapWordSize; 1155 double free_words_dbl = (double)free_words; 1156 size_t duty_cycle_words = (size_t)(free_words_dbl * duty_cycle / 100.0); 1157 size_t offset_words = (free_words - duty_cycle_words) / 2; 1158 1159 _icms_start_limit = eden->top() + offset_words; 1160 _icms_stop_limit = eden->end() - offset_words; 1161 1162 // The limits may be adjusted (shifted to the right) by 1163 // CMSIncrementalOffset, to allow the application more mutator time after a 1164 // young gen gc (when all mutators were stopped) and before CMS starts and 1165 // takes away one or more cpus. 1166 if (CMSIncrementalOffset != 0) { 1167 double adjustment_dbl = free_words_dbl * CMSIncrementalOffset / 100.0; 1168 size_t adjustment = (size_t)adjustment_dbl; 1169 HeapWord* tmp_stop = _icms_stop_limit + adjustment; 1170 if (tmp_stop > _icms_stop_limit && tmp_stop < eden->end()) { 1171 _icms_start_limit += adjustment; 1172 _icms_stop_limit = tmp_stop; 1173 } 1174 } 1175 } 1176 if (duty_cycle == 0 || (_icms_start_limit == _icms_stop_limit)) { 1177 _icms_start_limit = _icms_stop_limit = eden->end(); 1178 } 1179 1180 // Install the new start limit. 1181 eden->set_soft_end(_icms_start_limit); 1182 1183 if (CMSTraceIncrementalMode) { 1184 gclog_or_tty->print(" icms alloc limits: " 1185 PTR_FORMAT "," PTR_FORMAT 1186 " (" SIZE_FORMAT "%%," SIZE_FORMAT "%%) ", 1187 _icms_start_limit, _icms_stop_limit, 1188 percent_of_space(eden, _icms_start_limit), 1189 percent_of_space(eden, _icms_stop_limit)); 1190 if (Verbose) { 1191 gclog_or_tty->print("eden: "); 1192 eden->print_on(gclog_or_tty); 1193 } 1194 } 1195 } 1196 1197 // Any changes here should try to maintain the invariant 1198 // that if this method is called with _icms_start_limit 1199 // and _icms_stop_limit both NULL, then it should return NULL 1200 // and not notify the icms thread. 1201 HeapWord* 1202 CMSCollector::allocation_limit_reached(Space* space, HeapWord* top, 1203 size_t word_size) 1204 { 1205 // A start_limit equal to end() means the duty cycle is 0, so treat that as a 1206 // nop. 1207 if (CMSIncrementalMode && _icms_start_limit != space->end()) { 1208 if (top <= _icms_start_limit) { 1209 if (CMSTraceIncrementalMode) { 1210 space->print_on(gclog_or_tty); 1211 gclog_or_tty->stamp(); 1212 gclog_or_tty->print_cr(" start limit top=" PTR_FORMAT 1213 ", new limit=" PTR_FORMAT 1214 " (" SIZE_FORMAT "%%)", 1215 top, _icms_stop_limit, 1216 percent_of_space(space, _icms_stop_limit)); 1217 } 1218 ConcurrentMarkSweepThread::start_icms(); 1219 assert(top < _icms_stop_limit, "Tautology"); 1220 if (word_size < pointer_delta(_icms_stop_limit, top)) { 1221 return _icms_stop_limit; 1222 } 1223 1224 // The allocation will cross both the _start and _stop limits, so do the 1225 // stop notification also and return end(). 1226 if (CMSTraceIncrementalMode) { 1227 space->print_on(gclog_or_tty); 1228 gclog_or_tty->stamp(); 1229 gclog_or_tty->print_cr(" +stop limit top=" PTR_FORMAT 1230 ", new limit=" PTR_FORMAT 1231 " (" SIZE_FORMAT "%%)", 1232 top, space->end(), 1233 percent_of_space(space, space->end())); 1234 } 1235 ConcurrentMarkSweepThread::stop_icms(); 1236 return space->end(); 1237 } 1238 1239 if (top <= _icms_stop_limit) { 1240 if (CMSTraceIncrementalMode) { 1241 space->print_on(gclog_or_tty); 1242 gclog_or_tty->stamp(); 1243 gclog_or_tty->print_cr(" stop limit top=" PTR_FORMAT 1244 ", new limit=" PTR_FORMAT 1245 " (" SIZE_FORMAT "%%)", 1246 top, space->end(), 1247 percent_of_space(space, space->end())); 1248 } 1249 ConcurrentMarkSweepThread::stop_icms(); 1250 return space->end(); 1251 } 1252 1253 if (CMSTraceIncrementalMode) { 1254 space->print_on(gclog_or_tty); 1255 gclog_or_tty->stamp(); 1256 gclog_or_tty->print_cr(" end limit top=" PTR_FORMAT 1257 ", new limit=" PTR_FORMAT, 1258 top, NULL); 1259 } 1260 } 1261 1262 return NULL; 1263 } 1264 1265 oop ConcurrentMarkSweepGeneration::promote(oop obj, size_t obj_size) { 1266 assert(obj_size == (size_t)obj->size(), "bad obj_size passed in"); 1267 // allocate, copy and if necessary update promoinfo -- 1268 // delegate to underlying space. 1269 assert_lock_strong(freelistLock()); 1270 1271 #ifndef PRODUCT 1272 if (Universe::heap()->promotion_should_fail()) { 1273 return NULL; 1274 } 1275 #endif // #ifndef PRODUCT 1276 1277 oop res = _cmsSpace->promote(obj, obj_size); 1278 if (res == NULL) { 1279 // expand and retry 1280 size_t s = _cmsSpace->expansionSpaceRequired(obj_size); // HeapWords 1281 expand(s*HeapWordSize, MinHeapDeltaBytes, 1282 CMSExpansionCause::_satisfy_promotion); 1283 // Since there's currently no next generation, we don't try to promote 1284 // into a more senior generation. 1285 assert(next_gen() == NULL, "assumption, based upon which no attempt " 1286 "is made to pass on a possibly failing " 1287 "promotion to next generation"); 1288 res = _cmsSpace->promote(obj, obj_size); 1289 } 1290 if (res != NULL) { 1291 // See comment in allocate() about when objects should 1292 // be allocated live. 1293 assert(obj->is_oop(), "Will dereference klass pointer below"); 1294 collector()->promoted(false, // Not parallel 1295 (HeapWord*)res, obj->is_objArray(), obj_size); 1296 // promotion counters 1297 NOT_PRODUCT( 1298 _numObjectsPromoted++; 1299 _numWordsPromoted += 1300 (int)(CompactibleFreeListSpace::adjustObjectSize(obj->size())); 1301 ) 1302 } 1303 return res; 1304 } 1305 1306 1307 HeapWord* 1308 ConcurrentMarkSweepGeneration::allocation_limit_reached(Space* space, 1309 HeapWord* top, 1310 size_t word_sz) 1311 { 1312 return collector()->allocation_limit_reached(space, top, word_sz); 1313 } 1314 1315 // IMPORTANT: Notes on object size recognition in CMS. 1316 // --------------------------------------------------- 1317 // A block of storage in the CMS generation is always in 1318 // one of three states. A free block (FREE), an allocated 1319 // object (OBJECT) whose size() method reports the correct size, 1320 // and an intermediate state (TRANSIENT) in which its size cannot 1321 // be accurately determined. 1322 // STATE IDENTIFICATION: (32 bit and 64 bit w/o COOPS) 1323 // ----------------------------------------------------- 1324 // FREE: klass_word & 1 == 1; mark_word holds block size 1325 // 1326 // OBJECT: klass_word installed; klass_word != 0 && klass_word & 1 == 0; 1327 // obj->size() computes correct size 1328 // [Perm Gen objects needs to be "parsable" before they can be navigated] 1329 // 1330 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT 1331 // 1332 // STATE IDENTIFICATION: (64 bit+COOPS) 1333 // ------------------------------------ 1334 // FREE: mark_word & CMS_FREE_BIT == 1; mark_word & ~CMS_FREE_BIT gives block_size 1335 // 1336 // OBJECT: klass_word installed; klass_word != 0; 1337 // obj->size() computes correct size 1338 // [Perm Gen comment above continues to hold] 1339 // 1340 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT 1341 // 1342 // 1343 // STATE TRANSITION DIAGRAM 1344 // 1345 // mut / parnew mut / parnew 1346 // FREE --------------------> TRANSIENT ---------------------> OBJECT --| 1347 // ^ | 1348 // |------------------------ DEAD <------------------------------------| 1349 // sweep mut 1350 // 1351 // While a block is in TRANSIENT state its size cannot be determined 1352 // so readers will either need to come back later or stall until 1353 // the size can be determined. Note that for the case of direct 1354 // allocation, P-bits, when available, may be used to determine the 1355 // size of an object that may not yet have been initialized. 1356 1357 // Things to support parallel young-gen collection. 1358 oop 1359 ConcurrentMarkSweepGeneration::par_promote(int thread_num, 1360 oop old, markOop m, 1361 size_t word_sz) { 1362 #ifndef PRODUCT 1363 if (Universe::heap()->promotion_should_fail()) { 1364 return NULL; 1365 } 1366 #endif // #ifndef PRODUCT 1367 1368 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1369 PromotionInfo* promoInfo = &ps->promo; 1370 // if we are tracking promotions, then first ensure space for 1371 // promotion (including spooling space for saving header if necessary). 1372 // then allocate and copy, then track promoted info if needed. 1373 // When tracking (see PromotionInfo::track()), the mark word may 1374 // be displaced and in this case restoration of the mark word 1375 // occurs in the (oop_since_save_marks_)iterate phase. 1376 if (promoInfo->tracking() && !promoInfo->ensure_spooling_space()) { 1377 // Out of space for allocating spooling buffers; 1378 // try expanding and allocating spooling buffers. 1379 if (!expand_and_ensure_spooling_space(promoInfo)) { 1380 return NULL; 1381 } 1382 } 1383 assert(promoInfo->has_spooling_space(), "Control point invariant"); 1384 const size_t alloc_sz = CompactibleFreeListSpace::adjustObjectSize(word_sz); 1385 HeapWord* obj_ptr = ps->lab.alloc(alloc_sz); 1386 if (obj_ptr == NULL) { 1387 obj_ptr = expand_and_par_lab_allocate(ps, alloc_sz); 1388 if (obj_ptr == NULL) { 1389 return NULL; 1390 } 1391 } 1392 oop obj = oop(obj_ptr); 1393 OrderAccess::storestore(); 1394 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1395 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1396 // IMPORTANT: See note on object initialization for CMS above. 1397 // Otherwise, copy the object. Here we must be careful to insert the 1398 // klass pointer last, since this marks the block as an allocated object. 1399 // Except with compressed oops it's the mark word. 1400 HeapWord* old_ptr = (HeapWord*)old; 1401 // Restore the mark word copied above. 1402 obj->set_mark(m); 1403 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1404 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1405 OrderAccess::storestore(); 1406 1407 if (UseCompressedOops) { 1408 // Copy gap missed by (aligned) header size calculation below 1409 obj->set_klass_gap(old->klass_gap()); 1410 } 1411 if (word_sz > (size_t)oopDesc::header_size()) { 1412 Copy::aligned_disjoint_words(old_ptr + oopDesc::header_size(), 1413 obj_ptr + oopDesc::header_size(), 1414 word_sz - oopDesc::header_size()); 1415 } 1416 1417 // Now we can track the promoted object, if necessary. We take care 1418 // to delay the transition from uninitialized to full object 1419 // (i.e., insertion of klass pointer) until after, so that it 1420 // atomically becomes a promoted object. 1421 if (promoInfo->tracking()) { 1422 promoInfo->track((PromotedObject*)obj, old->klass()); 1423 } 1424 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1425 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1426 assert(old->is_oop(), "Will use and dereference old klass ptr below"); 1427 1428 // Finally, install the klass pointer (this should be volatile). 1429 OrderAccess::storestore(); 1430 obj->set_klass(old->klass()); 1431 // We should now be able to calculate the right size for this object 1432 assert(obj->is_oop() && obj->size() == (int)word_sz, "Error, incorrect size computed for promoted object"); 1433 1434 collector()->promoted(true, // parallel 1435 obj_ptr, old->is_objArray(), word_sz); 1436 1437 NOT_PRODUCT( 1438 Atomic::inc_ptr(&_numObjectsPromoted); 1439 Atomic::add_ptr(alloc_sz, &_numWordsPromoted); 1440 ) 1441 1442 return obj; 1443 } 1444 1445 void 1446 ConcurrentMarkSweepGeneration:: 1447 par_promote_alloc_undo(int thread_num, 1448 HeapWord* obj, size_t word_sz) { 1449 // CMS does not support promotion undo. 1450 ShouldNotReachHere(); 1451 } 1452 1453 void 1454 ConcurrentMarkSweepGeneration:: 1455 par_promote_alloc_done(int thread_num) { 1456 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1457 ps->lab.retire(thread_num); 1458 } 1459 1460 void 1461 ConcurrentMarkSweepGeneration:: 1462 par_oop_since_save_marks_iterate_done(int thread_num) { 1463 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1464 ParScanWithoutBarrierClosure* dummy_cl = NULL; 1465 ps->promo.promoted_oops_iterate_nv(dummy_cl); 1466 } 1467 1468 // XXXPERM 1469 bool ConcurrentMarkSweepGeneration::should_collect(bool full, 1470 size_t size, 1471 bool tlab) 1472 { 1473 // We allow a STW collection only if a full 1474 // collection was requested. 1475 return full || should_allocate(size, tlab); // FIX ME !!! 1476 // This and promotion failure handling are connected at the 1477 // hip and should be fixed by untying them. 1478 } 1479 1480 bool CMSCollector::shouldConcurrentCollect() { 1481 if (_full_gc_requested) { 1482 if (Verbose && PrintGCDetails) { 1483 gclog_or_tty->print_cr("CMSCollector: collect because of explicit " 1484 " gc request (or gc_locker)"); 1485 } 1486 return true; 1487 } 1488 1489 // For debugging purposes, change the type of collection. 1490 // If the rotation is not on the concurrent collection 1491 // type, don't start a concurrent collection. 1492 NOT_PRODUCT( 1493 if (RotateCMSCollectionTypes && 1494 (_cmsGen->debug_collection_type() != 1495 ConcurrentMarkSweepGeneration::Concurrent_collection_type)) { 1496 assert(_cmsGen->debug_collection_type() != 1497 ConcurrentMarkSweepGeneration::Unknown_collection_type, 1498 "Bad cms collection type"); 1499 return false; 1500 } 1501 ) 1502 1503 FreelistLocker x(this); 1504 // ------------------------------------------------------------------ 1505 // Print out lots of information which affects the initiation of 1506 // a collection. 1507 if (PrintCMSInitiationStatistics && stats().valid()) { 1508 gclog_or_tty->print("CMSCollector shouldConcurrentCollect: "); 1509 gclog_or_tty->stamp(); 1510 gclog_or_tty->print_cr(""); 1511 stats().print_on(gclog_or_tty); 1512 gclog_or_tty->print_cr("time_until_cms_gen_full %3.7f", 1513 stats().time_until_cms_gen_full()); 1514 gclog_or_tty->print_cr("free="SIZE_FORMAT, _cmsGen->free()); 1515 gclog_or_tty->print_cr("contiguous_available="SIZE_FORMAT, 1516 _cmsGen->contiguous_available()); 1517 gclog_or_tty->print_cr("promotion_rate=%g", stats().promotion_rate()); 1518 gclog_or_tty->print_cr("cms_allocation_rate=%g", stats().cms_allocation_rate()); 1519 gclog_or_tty->print_cr("occupancy=%3.7f", _cmsGen->occupancy()); 1520 gclog_or_tty->print_cr("initiatingOccupancy=%3.7f", _cmsGen->initiating_occupancy()); 1521 gclog_or_tty->print_cr("initiatingPermOccupancy=%3.7f", _permGen->initiating_occupancy()); 1522 } 1523 // ------------------------------------------------------------------ 1524 1525 // If the estimated time to complete a cms collection (cms_duration()) 1526 // is less than the estimated time remaining until the cms generation 1527 // is full, start a collection. 1528 if (!UseCMSInitiatingOccupancyOnly) { 1529 if (stats().valid()) { 1530 if (stats().time_until_cms_start() == 0.0) { 1531 return true; 1532 } 1533 } else { 1534 // We want to conservatively collect somewhat early in order 1535 // to try and "bootstrap" our CMS/promotion statistics; 1536 // this branch will not fire after the first successful CMS 1537 // collection because the stats should then be valid. 1538 if (_cmsGen->occupancy() >= _bootstrap_occupancy) { 1539 if (Verbose && PrintGCDetails) { 1540 gclog_or_tty->print_cr( 1541 " CMSCollector: collect for bootstrapping statistics:" 1542 " occupancy = %f, boot occupancy = %f", _cmsGen->occupancy(), 1543 _bootstrap_occupancy); 1544 } 1545 return true; 1546 } 1547 } 1548 } 1549 1550 // Otherwise, we start a collection cycle if either the perm gen or 1551 // old gen want a collection cycle started. Each may use 1552 // an appropriate criterion for making this decision. 1553 // XXX We need to make sure that the gen expansion 1554 // criterion dovetails well with this. XXX NEED TO FIX THIS 1555 if (_cmsGen->should_concurrent_collect()) { 1556 if (Verbose && PrintGCDetails) { 1557 gclog_or_tty->print_cr("CMS old gen initiated"); 1558 } 1559 return true; 1560 } 1561 1562 // We start a collection if we believe an incremental collection may fail; 1563 // this is not likely to be productive in practice because it's probably too 1564 // late anyway. 1565 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1566 assert(gch->collector_policy()->is_two_generation_policy(), 1567 "You may want to check the correctness of the following"); 1568 if (gch->incremental_collection_will_fail(true /* consult_young */)) { 1569 if (Verbose && PrintGCDetails) { 1570 gclog_or_tty->print("CMSCollector: collect because incremental collection will fail "); 1571 } 1572 return true; 1573 } 1574 1575 if (CMSClassUnloadingEnabled && _permGen->should_concurrent_collect()) { 1576 bool res = update_should_unload_classes(); 1577 if (res) { 1578 if (Verbose && PrintGCDetails) { 1579 gclog_or_tty->print_cr("CMS perm gen initiated"); 1580 } 1581 return true; 1582 } 1583 } 1584 return false; 1585 } 1586 1587 // Clear _expansion_cause fields of constituent generations 1588 void CMSCollector::clear_expansion_cause() { 1589 _cmsGen->clear_expansion_cause(); 1590 _permGen->clear_expansion_cause(); 1591 } 1592 1593 // We should be conservative in starting a collection cycle. To 1594 // start too eagerly runs the risk of collecting too often in the 1595 // extreme. To collect too rarely falls back on full collections, 1596 // which works, even if not optimum in terms of concurrent work. 1597 // As a work around for too eagerly collecting, use the flag 1598 // UseCMSInitiatingOccupancyOnly. This also has the advantage of 1599 // giving the user an easily understandable way of controlling the 1600 // collections. 1601 // We want to start a new collection cycle if any of the following 1602 // conditions hold: 1603 // . our current occupancy exceeds the configured initiating occupancy 1604 // for this generation, or 1605 // . we recently needed to expand this space and have not, since that 1606 // expansion, done a collection of this generation, or 1607 // . the underlying space believes that it may be a good idea to initiate 1608 // a concurrent collection (this may be based on criteria such as the 1609 // following: the space uses linear allocation and linear allocation is 1610 // going to fail, or there is believed to be excessive fragmentation in 1611 // the generation, etc... or ... 1612 // [.(currently done by CMSCollector::shouldConcurrentCollect() only for 1613 // the case of the old generation, not the perm generation; see CR 6543076): 1614 // we may be approaching a point at which allocation requests may fail because 1615 // we will be out of sufficient free space given allocation rate estimates.] 1616 bool ConcurrentMarkSweepGeneration::should_concurrent_collect() const { 1617 1618 assert_lock_strong(freelistLock()); 1619 if (occupancy() > initiating_occupancy()) { 1620 if (PrintGCDetails && Verbose) { 1621 gclog_or_tty->print(" %s: collect because of occupancy %f / %f ", 1622 short_name(), occupancy(), initiating_occupancy()); 1623 } 1624 return true; 1625 } 1626 if (UseCMSInitiatingOccupancyOnly) { 1627 return false; 1628 } 1629 if (expansion_cause() == CMSExpansionCause::_satisfy_allocation) { 1630 if (PrintGCDetails && Verbose) { 1631 gclog_or_tty->print(" %s: collect because expanded for allocation ", 1632 short_name()); 1633 } 1634 return true; 1635 } 1636 if (_cmsSpace->should_concurrent_collect()) { 1637 if (PrintGCDetails && Verbose) { 1638 gclog_or_tty->print(" %s: collect because cmsSpace says so ", 1639 short_name()); 1640 } 1641 return true; 1642 } 1643 return false; 1644 } 1645 1646 void ConcurrentMarkSweepGeneration::collect(bool full, 1647 bool clear_all_soft_refs, 1648 size_t size, 1649 bool tlab) 1650 { 1651 collector()->collect(full, clear_all_soft_refs, size, tlab); 1652 } 1653 1654 void CMSCollector::collect(bool full, 1655 bool clear_all_soft_refs, 1656 size_t size, 1657 bool tlab) 1658 { 1659 if (!UseCMSCollectionPassing && _collectorState > Idling) { 1660 // For debugging purposes skip the collection if the state 1661 // is not currently idle 1662 if (TraceCMSState) { 1663 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " skipped full:%d CMS state %d", 1664 Thread::current(), full, _collectorState); 1665 } 1666 return; 1667 } 1668 1669 // The following "if" branch is present for defensive reasons. 1670 // In the current uses of this interface, it can be replaced with: 1671 // assert(!GC_locker.is_active(), "Can't be called otherwise"); 1672 // But I am not placing that assert here to allow future 1673 // generality in invoking this interface. 1674 if (GC_locker::is_active()) { 1675 // A consistency test for GC_locker 1676 assert(GC_locker::needs_gc(), "Should have been set already"); 1677 // Skip this foreground collection, instead 1678 // expanding the heap if necessary. 1679 // Need the free list locks for the call to free() in compute_new_size() 1680 compute_new_size(); 1681 return; 1682 } 1683 acquire_control_and_collect(full, clear_all_soft_refs); 1684 _full_gcs_since_conc_gc++; 1685 1686 } 1687 1688 void CMSCollector::request_full_gc(unsigned int full_gc_count, GCCause::Cause cause) { 1689 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1690 unsigned int gc_count = gch->total_full_collections(); 1691 if (gc_count == full_gc_count) { 1692 MutexLockerEx y(CGC_lock, Mutex::_no_safepoint_check_flag); 1693 _full_gc_requested = true; 1694 _full_gc_cause = cause; 1695 CGC_lock->notify(); // nudge CMS thread 1696 } else { 1697 assert(gc_count > full_gc_count, "Error: causal loop"); 1698 } 1699 } 1700 1701 bool CMSCollector::is_external_interruption() { 1702 GCCause::Cause cause = GenCollectedHeap::heap()->gc_cause(); 1703 return GCCause::is_user_requested_gc(cause) || 1704 GCCause::is_serviceability_requested_gc(cause); 1705 } 1706 1707 void CMSCollector::report_concurrent_mode_interruption() { 1708 if (is_external_interruption()) { 1709 if (PrintGCDetails) { 1710 gclog_or_tty->print(" (concurrent mode interrupted)"); 1711 } 1712 } else { 1713 if (PrintGCDetails) { 1714 gclog_or_tty->print(" (concurrent mode failure)"); 1715 } 1716 _gc_tracer_cm->report_concurrent_mode_failure(); 1717 } 1718 } 1719 1720 // The foreground and background collectors need to coordinate in order 1721 // to make sure that they do not mutually interfere with CMS collections. 1722 // When a background collection is active, 1723 // the foreground collector may need to take over (preempt) and 1724 // synchronously complete an ongoing collection. Depending on the 1725 // frequency of the background collections and the heap usage 1726 // of the application, this preemption can be seldom or frequent. 1727 // There are only certain 1728 // points in the background collection that the "collection-baton" 1729 // can be passed to the foreground collector. 1730 // 1731 // The foreground collector will wait for the baton before 1732 // starting any part of the collection. The foreground collector 1733 // will only wait at one location. 1734 // 1735 // The background collector will yield the baton before starting a new 1736 // phase of the collection (e.g., before initial marking, marking from roots, 1737 // precleaning, final re-mark, sweep etc.) This is normally done at the head 1738 // of the loop which switches the phases. The background collector does some 1739 // of the phases (initial mark, final re-mark) with the world stopped. 1740 // Because of locking involved in stopping the world, 1741 // the foreground collector should not block waiting for the background 1742 // collector when it is doing a stop-the-world phase. The background 1743 // collector will yield the baton at an additional point just before 1744 // it enters a stop-the-world phase. Once the world is stopped, the 1745 // background collector checks the phase of the collection. If the 1746 // phase has not changed, it proceeds with the collection. If the 1747 // phase has changed, it skips that phase of the collection. See 1748 // the comments on the use of the Heap_lock in collect_in_background(). 1749 // 1750 // Variable used in baton passing. 1751 // _foregroundGCIsActive - Set to true by the foreground collector when 1752 // it wants the baton. The foreground clears it when it has finished 1753 // the collection. 1754 // _foregroundGCShouldWait - Set to true by the background collector 1755 // when it is running. The foreground collector waits while 1756 // _foregroundGCShouldWait is true. 1757 // CGC_lock - monitor used to protect access to the above variables 1758 // and to notify the foreground and background collectors. 1759 // _collectorState - current state of the CMS collection. 1760 // 1761 // The foreground collector 1762 // acquires the CGC_lock 1763 // sets _foregroundGCIsActive 1764 // waits on the CGC_lock for _foregroundGCShouldWait to be false 1765 // various locks acquired in preparation for the collection 1766 // are released so as not to block the background collector 1767 // that is in the midst of a collection 1768 // proceeds with the collection 1769 // clears _foregroundGCIsActive 1770 // returns 1771 // 1772 // The background collector in a loop iterating on the phases of the 1773 // collection 1774 // acquires the CGC_lock 1775 // sets _foregroundGCShouldWait 1776 // if _foregroundGCIsActive is set 1777 // clears _foregroundGCShouldWait, notifies _CGC_lock 1778 // waits on _CGC_lock for _foregroundGCIsActive to become false 1779 // and exits the loop. 1780 // otherwise 1781 // proceed with that phase of the collection 1782 // if the phase is a stop-the-world phase, 1783 // yield the baton once more just before enqueueing 1784 // the stop-world CMS operation (executed by the VM thread). 1785 // returns after all phases of the collection are done 1786 // 1787 1788 void CMSCollector::acquire_control_and_collect(bool full, 1789 bool clear_all_soft_refs) { 1790 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint"); 1791 assert(!Thread::current()->is_ConcurrentGC_thread(), 1792 "shouldn't try to acquire control from self!"); 1793 1794 // Start the protocol for acquiring control of the 1795 // collection from the background collector (aka CMS thread). 1796 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 1797 "VM thread should have CMS token"); 1798 // Remember the possibly interrupted state of an ongoing 1799 // concurrent collection 1800 CollectorState first_state = _collectorState; 1801 1802 // Signal to a possibly ongoing concurrent collection that 1803 // we want to do a foreground collection. 1804 _foregroundGCIsActive = true; 1805 1806 // Disable incremental mode during a foreground collection. 1807 ICMSDisabler icms_disabler; 1808 1809 // release locks and wait for a notify from the background collector 1810 // releasing the locks in only necessary for phases which 1811 // do yields to improve the granularity of the collection. 1812 assert_lock_strong(bitMapLock()); 1813 // We need to lock the Free list lock for the space that we are 1814 // currently collecting. 1815 assert(haveFreelistLocks(), "Must be holding free list locks"); 1816 bitMapLock()->unlock(); 1817 releaseFreelistLocks(); 1818 { 1819 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1820 if (_foregroundGCShouldWait) { 1821 // We are going to be waiting for action for the CMS thread; 1822 // it had better not be gone (for instance at shutdown)! 1823 assert(ConcurrentMarkSweepThread::cmst() != NULL, 1824 "CMS thread must be running"); 1825 // Wait here until the background collector gives us the go-ahead 1826 ConcurrentMarkSweepThread::clear_CMS_flag( 1827 ConcurrentMarkSweepThread::CMS_vm_has_token); // release token 1828 // Get a possibly blocked CMS thread going: 1829 // Note that we set _foregroundGCIsActive true above, 1830 // without protection of the CGC_lock. 1831 CGC_lock->notify(); 1832 assert(!ConcurrentMarkSweepThread::vm_thread_wants_cms_token(), 1833 "Possible deadlock"); 1834 while (_foregroundGCShouldWait) { 1835 // wait for notification 1836 CGC_lock->wait(Mutex::_no_safepoint_check_flag); 1837 // Possibility of delay/starvation here, since CMS token does 1838 // not know to give priority to VM thread? Actually, i think 1839 // there wouldn't be any delay/starvation, but the proof of 1840 // that "fact" (?) appears non-trivial. XXX 20011219YSR 1841 } 1842 ConcurrentMarkSweepThread::set_CMS_flag( 1843 ConcurrentMarkSweepThread::CMS_vm_has_token); 1844 } 1845 } 1846 // The CMS_token is already held. Get back the other locks. 1847 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 1848 "VM thread should have CMS token"); 1849 getFreelistLocks(); 1850 bitMapLock()->lock_without_safepoint_check(); 1851 if (TraceCMSState) { 1852 gclog_or_tty->print_cr("CMS foreground collector has asked for control " 1853 INTPTR_FORMAT " with first state %d", Thread::current(), first_state); 1854 gclog_or_tty->print_cr(" gets control with state %d", _collectorState); 1855 } 1856 1857 // Check if we need to do a compaction, or if not, whether 1858 // we need to start the mark-sweep from scratch. 1859 bool should_compact = false; 1860 bool should_start_over = false; 1861 decide_foreground_collection_type(clear_all_soft_refs, 1862 &should_compact, &should_start_over); 1863 1864 NOT_PRODUCT( 1865 if (RotateCMSCollectionTypes) { 1866 if (_cmsGen->debug_collection_type() == 1867 ConcurrentMarkSweepGeneration::MSC_foreground_collection_type) { 1868 should_compact = true; 1869 } else if (_cmsGen->debug_collection_type() == 1870 ConcurrentMarkSweepGeneration::MS_foreground_collection_type) { 1871 should_compact = false; 1872 } 1873 } 1874 ) 1875 1876 if (first_state > Idling) { 1877 report_concurrent_mode_interruption(); 1878 } 1879 1880 if (should_compact) { 1881 // If the collection is being acquired from the background 1882 // collector, there may be references on the discovered 1883 // references lists that have NULL referents (being those 1884 // that were concurrently cleared by a mutator) or 1885 // that are no longer active (having been enqueued concurrently 1886 // by the mutator). 1887 // Scrub the list of those references because Mark-Sweep-Compact 1888 // code assumes referents are not NULL and that all discovered 1889 // Reference objects are active. 1890 ref_processor()->clean_up_discovered_references(); 1891 1892 if (first_state > Idling) { 1893 save_heap_summary(); 1894 } 1895 1896 do_compaction_work(clear_all_soft_refs); 1897 1898 // Has the GC time limit been exceeded? 1899 DefNewGeneration* young_gen = _young_gen->as_DefNewGeneration(); 1900 size_t max_eden_size = young_gen->max_capacity() - 1901 young_gen->to()->capacity() - 1902 young_gen->from()->capacity(); 1903 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1904 GCCause::Cause gc_cause = gch->gc_cause(); 1905 size_policy()->check_gc_overhead_limit(_young_gen->used(), 1906 young_gen->eden()->used(), 1907 _cmsGen->max_capacity(), 1908 max_eden_size, 1909 full, 1910 gc_cause, 1911 gch->collector_policy()); 1912 } else { 1913 do_mark_sweep_work(clear_all_soft_refs, first_state, 1914 should_start_over); 1915 } 1916 // Reset the expansion cause, now that we just completed 1917 // a collection cycle. 1918 clear_expansion_cause(); 1919 _foregroundGCIsActive = false; 1920 return; 1921 } 1922 1923 // Resize the perm generation and the tenured generation 1924 // after obtaining the free list locks for the 1925 // two generations. 1926 void CMSCollector::compute_new_size() { 1927 assert_locked_or_safepoint(Heap_lock); 1928 FreelistLocker z(this); 1929 _permGen->compute_new_size(); 1930 _cmsGen->compute_new_size(); 1931 } 1932 1933 // A work method used by foreground collection to determine 1934 // what type of collection (compacting or not, continuing or fresh) 1935 // it should do. 1936 // NOTE: the intent is to make UseCMSCompactAtFullCollection 1937 // and CMSCompactWhenClearAllSoftRefs the default in the future 1938 // and do away with the flags after a suitable period. 1939 void CMSCollector::decide_foreground_collection_type( 1940 bool clear_all_soft_refs, bool* should_compact, 1941 bool* should_start_over) { 1942 // Normally, we'll compact only if the UseCMSCompactAtFullCollection 1943 // flag is set, and we have either requested a System.gc() or 1944 // the number of full gc's since the last concurrent cycle 1945 // has exceeded the threshold set by CMSFullGCsBeforeCompaction, 1946 // or if an incremental collection has failed 1947 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1948 assert(gch->collector_policy()->is_two_generation_policy(), 1949 "You may want to check the correctness of the following"); 1950 // Inform cms gen if this was due to partial collection failing. 1951 // The CMS gen may use this fact to determine its expansion policy. 1952 if (gch->incremental_collection_will_fail(false /* don't consult_young */)) { 1953 assert(!_cmsGen->incremental_collection_failed(), 1954 "Should have been noticed, reacted to and cleared"); 1955 _cmsGen->set_incremental_collection_failed(); 1956 } 1957 *should_compact = 1958 UseCMSCompactAtFullCollection && 1959 ((_full_gcs_since_conc_gc >= CMSFullGCsBeforeCompaction) || 1960 GCCause::is_user_requested_gc(gch->gc_cause()) || 1961 gch->incremental_collection_will_fail(true /* consult_young */)); 1962 *should_start_over = false; 1963 if (clear_all_soft_refs && !*should_compact) { 1964 // We are about to do a last ditch collection attempt 1965 // so it would normally make sense to do a compaction 1966 // to reclaim as much space as possible. 1967 if (CMSCompactWhenClearAllSoftRefs) { 1968 // Default: The rationale is that in this case either 1969 // we are past the final marking phase, in which case 1970 // we'd have to start over, or so little has been done 1971 // that there's little point in saving that work. Compaction 1972 // appears to be the sensible choice in either case. 1973 *should_compact = true; 1974 } else { 1975 // We have been asked to clear all soft refs, but not to 1976 // compact. Make sure that we aren't past the final checkpoint 1977 // phase, for that is where we process soft refs. If we are already 1978 // past that phase, we'll need to redo the refs discovery phase and 1979 // if necessary clear soft refs that weren't previously 1980 // cleared. We do so by remembering the phase in which 1981 // we came in, and if we are past the refs processing 1982 // phase, we'll choose to just redo the mark-sweep 1983 // collection from scratch. 1984 if (_collectorState > FinalMarking) { 1985 // We are past the refs processing phase; 1986 // start over and do a fresh synchronous CMS cycle 1987 _collectorState = Resetting; // skip to reset to start new cycle 1988 reset(false /* == !asynch */); 1989 *should_start_over = true; 1990 } // else we can continue a possibly ongoing current cycle 1991 } 1992 } 1993 } 1994 1995 // A work method used by the foreground collector to do 1996 // a mark-sweep-compact. 1997 void CMSCollector::do_compaction_work(bool clear_all_soft_refs) { 1998 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1999 2000 STWGCTimer* gc_timer = GenMarkSweep::gc_timer(); 2001 gc_timer->register_gc_start(); 2002 2003 SerialOldTracer* gc_tracer = GenMarkSweep::gc_tracer(); 2004 gc_tracer->report_gc_start(gch->gc_cause(), gc_timer->gc_start()); 2005 2006 GCTraceTime t("CMS:MSC ", PrintGCDetails && Verbose, true, NULL); 2007 if (PrintGC && Verbose && !(GCCause::is_user_requested_gc(gch->gc_cause()))) { 2008 gclog_or_tty->print_cr("Compact ConcurrentMarkSweepGeneration after %d " 2009 "collections passed to foreground collector", _full_gcs_since_conc_gc); 2010 } 2011 2012 // Sample collection interval time and reset for collection pause. 2013 if (UseAdaptiveSizePolicy) { 2014 size_policy()->msc_collection_begin(); 2015 } 2016 2017 // Temporarily widen the span of the weak reference processing to 2018 // the entire heap. 2019 MemRegion new_span(GenCollectedHeap::heap()->reserved_region()); 2020 ReferenceProcessorSpanMutator rp_mut_span(ref_processor(), new_span); 2021 // Temporarily, clear the "is_alive_non_header" field of the 2022 // reference processor. 2023 ReferenceProcessorIsAliveMutator rp_mut_closure(ref_processor(), NULL); 2024 // Temporarily make reference _processing_ single threaded (non-MT). 2025 ReferenceProcessorMTProcMutator rp_mut_mt_processing(ref_processor(), false); 2026 // Temporarily make refs discovery atomic 2027 ReferenceProcessorAtomicMutator rp_mut_atomic(ref_processor(), true); 2028 // Temporarily make reference _discovery_ single threaded (non-MT) 2029 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false); 2030 2031 ref_processor()->set_enqueuing_is_done(false); 2032 ref_processor()->enable_discovery(false /*verify_disabled*/, false /*check_no_refs*/); 2033 ref_processor()->setup_policy(clear_all_soft_refs); 2034 // If an asynchronous collection finishes, the _modUnionTable is 2035 // all clear. If we are assuming the collection from an asynchronous 2036 // collection, clear the _modUnionTable. 2037 assert(_collectorState != Idling || _modUnionTable.isAllClear(), 2038 "_modUnionTable should be clear if the baton was not passed"); 2039 _modUnionTable.clear_all(); 2040 2041 // We must adjust the allocation statistics being maintained 2042 // in the free list space. We do so by reading and clearing 2043 // the sweep timer and updating the block flux rate estimates below. 2044 assert(!_intra_sweep_timer.is_active(), "_intra_sweep_timer should be inactive"); 2045 if (_inter_sweep_timer.is_active()) { 2046 _inter_sweep_timer.stop(); 2047 // Note that we do not use this sample to update the _inter_sweep_estimate. 2048 _cmsGen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()), 2049 _inter_sweep_estimate.padded_average(), 2050 _intra_sweep_estimate.padded_average()); 2051 } 2052 2053 GenMarkSweep::invoke_at_safepoint(_cmsGen->level(), 2054 ref_processor(), clear_all_soft_refs); 2055 #ifdef ASSERT 2056 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 2057 size_t free_size = cms_space->free(); 2058 assert(free_size == 2059 pointer_delta(cms_space->end(), cms_space->compaction_top()) 2060 * HeapWordSize, 2061 "All the free space should be compacted into one chunk at top"); 2062 assert(cms_space->dictionary()->total_chunk_size( 2063 debug_only(cms_space->freelistLock())) == 0 || 2064 cms_space->totalSizeInIndexedFreeLists() == 0, 2065 "All the free space should be in a single chunk"); 2066 size_t num = cms_space->totalCount(); 2067 assert((free_size == 0 && num == 0) || 2068 (free_size > 0 && (num == 1 || num == 2)), 2069 "There should be at most 2 free chunks after compaction"); 2070 #endif // ASSERT 2071 _collectorState = Resetting; 2072 assert(_restart_addr == NULL, 2073 "Should have been NULL'd before baton was passed"); 2074 reset(false /* == !asynch */); 2075 _cmsGen->reset_after_compaction(); 2076 _concurrent_cycles_since_last_unload = 0; 2077 2078 if (verifying() && !should_unload_classes()) { 2079 perm_gen_verify_bit_map()->clear_all(); 2080 } 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 // Sample collection pause time and reset for collection interval. 2094 if (UseAdaptiveSizePolicy) { 2095 size_policy()->msc_collection_end(gch->gc_cause()); 2096 } 2097 2098 gc_timer->register_gc_end(); 2099 2100 gc_tracer->report_gc_end(gc_timer->gc_end(), gc_timer->time_partitions()); 2101 2102 // For a mark-sweep-compact, compute_new_size() will be called 2103 // in the heap's do_collection() method. 2104 } 2105 2106 // A work method used by the foreground collector to do 2107 // a mark-sweep, after taking over from a possibly on-going 2108 // concurrent mark-sweep collection. 2109 void CMSCollector::do_mark_sweep_work(bool clear_all_soft_refs, 2110 CollectorState first_state, bool should_start_over) { 2111 if (PrintGC && Verbose) { 2112 gclog_or_tty->print_cr("Pass concurrent collection to foreground " 2113 "collector with count %d", 2114 _full_gcs_since_conc_gc); 2115 } 2116 switch (_collectorState) { 2117 case Idling: 2118 if (first_state == Idling || should_start_over) { 2119 // The background GC was not active, or should 2120 // restarted from scratch; start the cycle. 2121 _collectorState = InitialMarking; 2122 } 2123 // If first_state was not Idling, then a background GC 2124 // was in progress and has now finished. No need to do it 2125 // again. Leave the state as Idling. 2126 break; 2127 case Precleaning: 2128 // In the foreground case don't do the precleaning since 2129 // it is not done concurrently and there is extra work 2130 // required. 2131 _collectorState = FinalMarking; 2132 } 2133 collect_in_foreground(clear_all_soft_refs, GenCollectedHeap::heap()->gc_cause()); 2134 2135 // For a mark-sweep, compute_new_size() will be called 2136 // in the heap's do_collection() method. 2137 } 2138 2139 2140 void CMSCollector::print_eden_and_survivor_chunk_arrays() { 2141 DefNewGeneration* dng = _young_gen->as_DefNewGeneration(); 2142 EdenSpace* eden_space = dng->eden(); 2143 ContiguousSpace* from_space = dng->from(); 2144 ContiguousSpace* to_space = dng->to(); 2145 // Eden 2146 if (_eden_chunk_array != NULL) { 2147 gclog_or_tty->print_cr("eden " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")", 2148 eden_space->bottom(), eden_space->top(), 2149 eden_space->end(), eden_space->capacity()); 2150 gclog_or_tty->print_cr("_eden_chunk_index=" SIZE_FORMAT ", " 2151 "_eden_chunk_capacity=" SIZE_FORMAT, 2152 _eden_chunk_index, _eden_chunk_capacity); 2153 for (size_t i = 0; i < _eden_chunk_index; i++) { 2154 gclog_or_tty->print_cr("_eden_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT, 2155 i, _eden_chunk_array[i]); 2156 } 2157 } 2158 // Survivor 2159 if (_survivor_chunk_array != NULL) { 2160 gclog_or_tty->print_cr("survivor " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")", 2161 from_space->bottom(), from_space->top(), 2162 from_space->end(), from_space->capacity()); 2163 gclog_or_tty->print_cr("_survivor_chunk_index=" SIZE_FORMAT ", " 2164 "_survivor_chunk_capacity=" SIZE_FORMAT, 2165 _survivor_chunk_index, _survivor_chunk_capacity); 2166 for (size_t i = 0; i < _survivor_chunk_index; i++) { 2167 gclog_or_tty->print_cr("_survivor_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT, 2168 i, _survivor_chunk_array[i]); 2169 } 2170 } 2171 } 2172 2173 void CMSCollector::getFreelistLocks() const { 2174 // Get locks for all free lists in all generations that this 2175 // collector is responsible for 2176 _cmsGen->freelistLock()->lock_without_safepoint_check(); 2177 _permGen->freelistLock()->lock_without_safepoint_check(); 2178 } 2179 2180 void CMSCollector::releaseFreelistLocks() const { 2181 // Release locks for all free lists in all generations that this 2182 // collector is responsible for 2183 _cmsGen->freelistLock()->unlock(); 2184 _permGen->freelistLock()->unlock(); 2185 } 2186 2187 bool CMSCollector::haveFreelistLocks() const { 2188 // Check locks for all free lists in all generations that this 2189 // collector is responsible for 2190 assert_lock_strong(_cmsGen->freelistLock()); 2191 assert_lock_strong(_permGen->freelistLock()); 2192 PRODUCT_ONLY(ShouldNotReachHere()); 2193 return true; 2194 } 2195 2196 // A utility class that is used by the CMS collector to 2197 // temporarily "release" the foreground collector from its 2198 // usual obligation to wait for the background collector to 2199 // complete an ongoing phase before proceeding. 2200 class ReleaseForegroundGC: public StackObj { 2201 private: 2202 CMSCollector* _c; 2203 public: 2204 ReleaseForegroundGC(CMSCollector* c) : _c(c) { 2205 assert(_c->_foregroundGCShouldWait, "Else should not need to call"); 2206 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 2207 // allow a potentially blocked foreground collector to proceed 2208 _c->_foregroundGCShouldWait = false; 2209 if (_c->_foregroundGCIsActive) { 2210 CGC_lock->notify(); 2211 } 2212 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 2213 "Possible deadlock"); 2214 } 2215 2216 ~ReleaseForegroundGC() { 2217 assert(!_c->_foregroundGCShouldWait, "Usage protocol violation?"); 2218 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 2219 _c->_foregroundGCShouldWait = true; 2220 } 2221 }; 2222 2223 // There are separate collect_in_background and collect_in_foreground because of 2224 // the different locking requirements of the background collector and the 2225 // foreground collector. There was originally an attempt to share 2226 // one "collect" method between the background collector and the foreground 2227 // collector but the if-then-else required made it cleaner to have 2228 // separate methods. 2229 void CMSCollector::collect_in_background(bool clear_all_soft_refs, GCCause::Cause cause) { 2230 assert(Thread::current()->is_ConcurrentGC_thread(), 2231 "A CMS asynchronous collection is only allowed on a CMS thread."); 2232 2233 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2234 { 2235 bool safepoint_check = Mutex::_no_safepoint_check_flag; 2236 MutexLockerEx hl(Heap_lock, safepoint_check); 2237 FreelistLocker fll(this); 2238 MutexLockerEx x(CGC_lock, safepoint_check); 2239 if (_foregroundGCIsActive || !UseAsyncConcMarkSweepGC) { 2240 // The foreground collector is active or we're 2241 // not using asynchronous collections. Skip this 2242 // background collection. 2243 assert(!_foregroundGCShouldWait, "Should be clear"); 2244 return; 2245 } else { 2246 assert(_collectorState == Idling, "Should be idling before start."); 2247 _collectorState = InitialMarking; 2248 register_gc_start(cause); 2249 // Reset the expansion cause, now that we are about to begin 2250 // a new cycle. 2251 clear_expansion_cause(); 2252 } 2253 // Decide if we want to enable class unloading as part of the 2254 // ensuing concurrent GC cycle. 2255 update_should_unload_classes(); 2256 _full_gc_requested = false; // acks all outstanding full gc requests 2257 _full_gc_cause = GCCause::_no_gc; 2258 // Signal that we are about to start a collection 2259 gch->increment_total_full_collections(); // ... starting a collection cycle 2260 _collection_count_start = gch->total_full_collections(); 2261 } 2262 2263 // Used for PrintGC 2264 size_t prev_used; 2265 if (PrintGC && Verbose) { 2266 prev_used = _cmsGen->used(); // XXXPERM 2267 } 2268 2269 // The change of the collection state is normally done at this level; 2270 // the exceptions are phases that are executed while the world is 2271 // stopped. For those phases the change of state is done while the 2272 // world is stopped. For baton passing purposes this allows the 2273 // background collector to finish the phase and change state atomically. 2274 // The foreground collector cannot wait on a phase that is done 2275 // while the world is stopped because the foreground collector already 2276 // has the world stopped and would deadlock. 2277 while (_collectorState != Idling) { 2278 if (TraceCMSState) { 2279 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " in CMS state %d", 2280 Thread::current(), _collectorState); 2281 } 2282 // The foreground collector 2283 // holds the Heap_lock throughout its collection. 2284 // holds the CMS token (but not the lock) 2285 // except while it is waiting for the background collector to yield. 2286 // 2287 // The foreground collector should be blocked (not for long) 2288 // if the background collector is about to start a phase 2289 // executed with world stopped. If the background 2290 // collector has already started such a phase, the 2291 // foreground collector is blocked waiting for the 2292 // Heap_lock. The stop-world phases (InitialMarking and FinalMarking) 2293 // are executed in the VM thread. 2294 // 2295 // The locking order is 2296 // PendingListLock (PLL) -- if applicable (FinalMarking) 2297 // Heap_lock (both this & PLL locked in VM_CMS_Operation::prologue()) 2298 // CMS token (claimed in 2299 // stop_world_and_do() --> 2300 // safepoint_synchronize() --> 2301 // CMSThread::synchronize()) 2302 2303 { 2304 // Check if the FG collector wants us to yield. 2305 CMSTokenSync x(true); // is cms thread 2306 if (waitForForegroundGC()) { 2307 // We yielded to a foreground GC, nothing more to be 2308 // done this round. 2309 assert(_foregroundGCShouldWait == false, "We set it to false in " 2310 "waitForForegroundGC()"); 2311 if (TraceCMSState) { 2312 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT 2313 " exiting collection CMS state %d", 2314 Thread::current(), _collectorState); 2315 } 2316 return; 2317 } else { 2318 // The background collector can run but check to see if the 2319 // foreground collector has done a collection while the 2320 // background collector was waiting to get the CGC_lock 2321 // above. If yes, break so that _foregroundGCShouldWait 2322 // is cleared before returning. 2323 if (_collectorState == Idling) { 2324 break; 2325 } 2326 } 2327 } 2328 2329 assert(_foregroundGCShouldWait, "Foreground collector, if active, " 2330 "should be waiting"); 2331 2332 switch (_collectorState) { 2333 case InitialMarking: 2334 { 2335 ReleaseForegroundGC x(this); 2336 stats().record_cms_begin(); 2337 VM_CMS_Initial_Mark initial_mark_op(this); 2338 VMThread::execute(&initial_mark_op); 2339 } 2340 // The collector state may be any legal state at this point 2341 // since the background collector may have yielded to the 2342 // foreground collector. 2343 break; 2344 case Marking: 2345 // initial marking in checkpointRootsInitialWork has been completed 2346 if (markFromRoots(true)) { // we were successful 2347 assert(_collectorState == Precleaning, "Collector state should " 2348 "have changed"); 2349 } else { 2350 assert(_foregroundGCIsActive, "Internal state inconsistency"); 2351 } 2352 break; 2353 case Precleaning: 2354 if (UseAdaptiveSizePolicy) { 2355 size_policy()->concurrent_precleaning_begin(); 2356 } 2357 // marking from roots in markFromRoots has been completed 2358 preclean(); 2359 if (UseAdaptiveSizePolicy) { 2360 size_policy()->concurrent_precleaning_end(); 2361 } 2362 assert(_collectorState == AbortablePreclean || 2363 _collectorState == FinalMarking, 2364 "Collector state should have changed"); 2365 break; 2366 case AbortablePreclean: 2367 if (UseAdaptiveSizePolicy) { 2368 size_policy()->concurrent_phases_resume(); 2369 } 2370 abortable_preclean(); 2371 if (UseAdaptiveSizePolicy) { 2372 size_policy()->concurrent_precleaning_end(); 2373 } 2374 assert(_collectorState == FinalMarking, "Collector state should " 2375 "have changed"); 2376 break; 2377 case FinalMarking: 2378 { 2379 ReleaseForegroundGC x(this); 2380 2381 VM_CMS_Final_Remark final_remark_op(this); 2382 VMThread::execute(&final_remark_op); 2383 } 2384 assert(_foregroundGCShouldWait, "block post-condition"); 2385 break; 2386 case Sweeping: 2387 if (UseAdaptiveSizePolicy) { 2388 size_policy()->concurrent_sweeping_begin(); 2389 } 2390 // final marking in checkpointRootsFinal has been completed 2391 sweep(true); 2392 assert(_collectorState == Resizing, "Collector state change " 2393 "to Resizing must be done under the free_list_lock"); 2394 _full_gcs_since_conc_gc = 0; 2395 2396 // Stop the timers for adaptive size policy for the concurrent phases 2397 if (UseAdaptiveSizePolicy) { 2398 size_policy()->concurrent_sweeping_end(); 2399 size_policy()->concurrent_phases_end(gch->gc_cause(), 2400 gch->prev_gen(_cmsGen)->capacity(), 2401 _cmsGen->free()); 2402 } 2403 2404 case Resizing: { 2405 // Sweeping has been completed... 2406 // At this point the background collection has completed. 2407 // Don't move the call to compute_new_size() down 2408 // into code that might be executed if the background 2409 // collection was preempted. 2410 { 2411 ReleaseForegroundGC x(this); // unblock FG collection 2412 MutexLockerEx y(Heap_lock, Mutex::_no_safepoint_check_flag); 2413 CMSTokenSync z(true); // not strictly needed. 2414 if (_collectorState == Resizing) { 2415 compute_new_size(); 2416 save_heap_summary(); 2417 _collectorState = Resetting; 2418 } else { 2419 assert(_collectorState == Idling, "The state should only change" 2420 " because the foreground collector has finished the collection"); 2421 } 2422 } 2423 break; 2424 } 2425 case Resetting: 2426 // CMS heap resizing has been completed 2427 reset(true); 2428 assert(_collectorState == Idling, "Collector state should " 2429 "have changed"); 2430 stats().record_cms_end(); 2431 // Don't move the concurrent_phases_end() and compute_new_size() 2432 // calls to here because a preempted background collection 2433 // has it's state set to "Resetting". 2434 break; 2435 case Idling: 2436 default: 2437 ShouldNotReachHere(); 2438 break; 2439 } 2440 if (TraceCMSState) { 2441 gclog_or_tty->print_cr(" Thread " INTPTR_FORMAT " done - next CMS state %d", 2442 Thread::current(), _collectorState); 2443 } 2444 assert(_foregroundGCShouldWait, "block post-condition"); 2445 } 2446 2447 // Should this be in gc_epilogue? 2448 collector_policy()->counters()->update_counters(); 2449 2450 { 2451 // Clear _foregroundGCShouldWait and, in the event that the 2452 // foreground collector is waiting, notify it, before 2453 // returning. 2454 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 2455 _foregroundGCShouldWait = false; 2456 if (_foregroundGCIsActive) { 2457 CGC_lock->notify(); 2458 } 2459 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 2460 "Possible deadlock"); 2461 } 2462 if (TraceCMSState) { 2463 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT 2464 " exiting collection CMS state %d", 2465 Thread::current(), _collectorState); 2466 } 2467 if (PrintGC && Verbose) { 2468 _cmsGen->print_heap_change(prev_used); 2469 } 2470 } 2471 2472 void CMSCollector::register_foreground_gc_start(GCCause::Cause cause) { 2473 if (!_cms_start_registered) { 2474 register_gc_start(cause); 2475 } 2476 } 2477 2478 void CMSCollector::register_gc_start(GCCause::Cause cause) { 2479 _cms_start_registered = true; 2480 _gc_timer_cm->register_gc_start(); 2481 _gc_tracer_cm->report_gc_start(cause, _gc_timer_cm->gc_start()); 2482 } 2483 2484 void CMSCollector::register_gc_end() { 2485 if (_cms_start_registered) { 2486 report_heap_summary(GCWhen::AfterGC); 2487 2488 _gc_timer_cm->register_gc_end(); 2489 _gc_tracer_cm->report_gc_end(_gc_timer_cm->gc_end(), _gc_timer_cm->time_partitions()); 2490 _cms_start_registered = false; 2491 } 2492 } 2493 2494 void CMSCollector::save_heap_summary() { 2495 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2496 _last_heap_summary = gch->create_heap_summary(); 2497 _last_perm_gen_summary = gch->create_perm_gen_summary(); 2498 } 2499 2500 void CMSCollector::report_heap_summary(GCWhen::Type when) { 2501 _gc_tracer_cm->report_gc_heap_summary(when, _last_heap_summary, _last_perm_gen_summary); 2502 } 2503 2504 void CMSCollector::collect_in_foreground(bool clear_all_soft_refs, GCCause::Cause cause) { 2505 assert(_foregroundGCIsActive && !_foregroundGCShouldWait, 2506 "Foreground collector should be waiting, not executing"); 2507 assert(Thread::current()->is_VM_thread(), "A foreground collection" 2508 "may only be done by the VM Thread with the world stopped"); 2509 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 2510 "VM thread should have CMS token"); 2511 2512 NOT_PRODUCT(GCTraceTime t("CMS:MS (foreground) ", PrintGCDetails && Verbose, 2513 true, NULL);) 2514 if (UseAdaptiveSizePolicy) { 2515 size_policy()->ms_collection_begin(); 2516 } 2517 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact); 2518 2519 HandleMark hm; // Discard invalid handles created during verification 2520 2521 if (VerifyBeforeGC && 2522 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2523 Universe::verify(); 2524 } 2525 2526 // Snapshot the soft reference policy to be used in this collection cycle. 2527 ref_processor()->setup_policy(clear_all_soft_refs); 2528 2529 bool init_mark_was_synchronous = false; // until proven otherwise 2530 while (_collectorState != Idling) { 2531 if (TraceCMSState) { 2532 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " in CMS state %d", 2533 Thread::current(), _collectorState); 2534 } 2535 switch (_collectorState) { 2536 case InitialMarking: 2537 register_foreground_gc_start(cause); 2538 init_mark_was_synchronous = true; // fact to be exploited in re-mark 2539 checkpointRootsInitial(false); 2540 assert(_collectorState == Marking, "Collector state should have changed" 2541 " within checkpointRootsInitial()"); 2542 break; 2543 case Marking: 2544 // initial marking in checkpointRootsInitialWork has been completed 2545 if (VerifyDuringGC && 2546 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2547 Universe::verify("Verify before initial mark: "); 2548 } 2549 { 2550 bool res = markFromRoots(false); 2551 assert(res && _collectorState == FinalMarking, "Collector state should " 2552 "have changed"); 2553 break; 2554 } 2555 case FinalMarking: 2556 if (VerifyDuringGC && 2557 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2558 Universe::verify("Verify before re-mark: "); 2559 } 2560 checkpointRootsFinal(false, clear_all_soft_refs, 2561 init_mark_was_synchronous); 2562 assert(_collectorState == Sweeping, "Collector state should not " 2563 "have changed within checkpointRootsFinal()"); 2564 break; 2565 case Sweeping: 2566 // final marking in checkpointRootsFinal has been completed 2567 if (VerifyDuringGC && 2568 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2569 Universe::verify("Verify before sweep: "); 2570 } 2571 sweep(false); 2572 assert(_collectorState == Resizing, "Incorrect state"); 2573 break; 2574 case Resizing: { 2575 // Sweeping has been completed; the actual resize in this case 2576 // is done separately; nothing to be done in this state. 2577 _collectorState = Resetting; 2578 break; 2579 } 2580 case Resetting: 2581 // The heap has been resized. 2582 if (VerifyDuringGC && 2583 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2584 Universe::verify("Verify before reset: "); 2585 } 2586 save_heap_summary(); 2587 reset(false); 2588 assert(_collectorState == Idling, "Collector state should " 2589 "have changed"); 2590 break; 2591 case Precleaning: 2592 case AbortablePreclean: 2593 // Elide the preclean phase 2594 _collectorState = FinalMarking; 2595 break; 2596 default: 2597 ShouldNotReachHere(); 2598 } 2599 if (TraceCMSState) { 2600 gclog_or_tty->print_cr(" Thread " INTPTR_FORMAT " done - next CMS state %d", 2601 Thread::current(), _collectorState); 2602 } 2603 } 2604 2605 if (UseAdaptiveSizePolicy) { 2606 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2607 size_policy()->ms_collection_end(gch->gc_cause()); 2608 } 2609 2610 if (VerifyAfterGC && 2611 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2612 Universe::verify(); 2613 } 2614 if (TraceCMSState) { 2615 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT 2616 " exiting collection CMS state %d", 2617 Thread::current(), _collectorState); 2618 } 2619 } 2620 2621 bool CMSCollector::waitForForegroundGC() { 2622 bool res = false; 2623 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 2624 "CMS thread should have CMS token"); 2625 // Block the foreground collector until the 2626 // background collectors decides whether to 2627 // yield. 2628 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 2629 _foregroundGCShouldWait = true; 2630 if (_foregroundGCIsActive) { 2631 // The background collector yields to the 2632 // foreground collector and returns a value 2633 // indicating that it has yielded. The foreground 2634 // collector can proceed. 2635 res = true; 2636 _foregroundGCShouldWait = false; 2637 ConcurrentMarkSweepThread::clear_CMS_flag( 2638 ConcurrentMarkSweepThread::CMS_cms_has_token); 2639 ConcurrentMarkSweepThread::set_CMS_flag( 2640 ConcurrentMarkSweepThread::CMS_cms_wants_token); 2641 // Get a possibly blocked foreground thread going 2642 CGC_lock->notify(); 2643 if (TraceCMSState) { 2644 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " waiting at CMS state %d", 2645 Thread::current(), _collectorState); 2646 } 2647 while (_foregroundGCIsActive) { 2648 CGC_lock->wait(Mutex::_no_safepoint_check_flag); 2649 } 2650 ConcurrentMarkSweepThread::set_CMS_flag( 2651 ConcurrentMarkSweepThread::CMS_cms_has_token); 2652 ConcurrentMarkSweepThread::clear_CMS_flag( 2653 ConcurrentMarkSweepThread::CMS_cms_wants_token); 2654 } 2655 if (TraceCMSState) { 2656 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " continuing at CMS state %d", 2657 Thread::current(), _collectorState); 2658 } 2659 return res; 2660 } 2661 2662 // Because of the need to lock the free lists and other structures in 2663 // the collector, common to all the generations that the collector is 2664 // collecting, we need the gc_prologues of individual CMS generations 2665 // delegate to their collector. It may have been simpler had the 2666 // current infrastructure allowed one to call a prologue on a 2667 // collector. In the absence of that we have the generation's 2668 // prologue delegate to the collector, which delegates back 2669 // some "local" work to a worker method in the individual generations 2670 // that it's responsible for collecting, while itself doing any 2671 // work common to all generations it's responsible for. A similar 2672 // comment applies to the gc_epilogue()'s. 2673 // The role of the varaible _between_prologue_and_epilogue is to 2674 // enforce the invocation protocol. 2675 void CMSCollector::gc_prologue(bool full) { 2676 // Call gc_prologue_work() for each CMSGen and PermGen that 2677 // we are responsible for. 2678 2679 // The following locking discipline assumes that we are only called 2680 // when the world is stopped. 2681 assert(SafepointSynchronize::is_at_safepoint(), "world is stopped assumption"); 2682 2683 // The CMSCollector prologue must call the gc_prologues for the 2684 // "generations" (including PermGen if any) that it's responsible 2685 // for. 2686 2687 assert( Thread::current()->is_VM_thread() 2688 || ( CMSScavengeBeforeRemark 2689 && Thread::current()->is_ConcurrentGC_thread()), 2690 "Incorrect thread type for prologue execution"); 2691 2692 if (_between_prologue_and_epilogue) { 2693 // We have already been invoked; this is a gc_prologue delegation 2694 // from yet another CMS generation that we are responsible for, just 2695 // ignore it since all relevant work has already been done. 2696 return; 2697 } 2698 2699 // set a bit saying prologue has been called; cleared in epilogue 2700 _between_prologue_and_epilogue = true; 2701 // Claim locks for common data structures, then call gc_prologue_work() 2702 // for each CMSGen and PermGen that we are responsible for. 2703 2704 getFreelistLocks(); // gets free list locks on constituent spaces 2705 bitMapLock()->lock_without_safepoint_check(); 2706 2707 // Should call gc_prologue_work() for all cms gens we are responsible for 2708 bool registerClosure = _collectorState >= Marking 2709 && _collectorState < Sweeping; 2710 ModUnionClosure* muc = CollectedHeap::use_parallel_gc_threads() ? 2711 &_modUnionClosurePar 2712 : &_modUnionClosure; 2713 _cmsGen->gc_prologue_work(full, registerClosure, muc); 2714 _permGen->gc_prologue_work(full, registerClosure, muc); 2715 2716 if (!full) { 2717 stats().record_gc0_begin(); 2718 } 2719 } 2720 2721 void ConcurrentMarkSweepGeneration::gc_prologue(bool full) { 2722 // Delegate to CMScollector which knows how to coordinate between 2723 // this and any other CMS generations that it is responsible for 2724 // collecting. 2725 collector()->gc_prologue(full); 2726 } 2727 2728 // This is a "private" interface for use by this generation's CMSCollector. 2729 // Not to be called directly by any other entity (for instance, 2730 // GenCollectedHeap, which calls the "public" gc_prologue method above). 2731 void ConcurrentMarkSweepGeneration::gc_prologue_work(bool full, 2732 bool registerClosure, ModUnionClosure* modUnionClosure) { 2733 assert(!incremental_collection_failed(), "Shouldn't be set yet"); 2734 assert(cmsSpace()->preconsumptionDirtyCardClosure() == NULL, 2735 "Should be NULL"); 2736 if (registerClosure) { 2737 cmsSpace()->setPreconsumptionDirtyCardClosure(modUnionClosure); 2738 } 2739 cmsSpace()->gc_prologue(); 2740 // Clear stat counters 2741 NOT_PRODUCT( 2742 assert(_numObjectsPromoted == 0, "check"); 2743 assert(_numWordsPromoted == 0, "check"); 2744 if (Verbose && PrintGC) { 2745 gclog_or_tty->print("Allocated "SIZE_FORMAT" objects, " 2746 SIZE_FORMAT" bytes concurrently", 2747 _numObjectsAllocated, _numWordsAllocated*sizeof(HeapWord)); 2748 } 2749 _numObjectsAllocated = 0; 2750 _numWordsAllocated = 0; 2751 ) 2752 } 2753 2754 void CMSCollector::gc_epilogue(bool full) { 2755 // The following locking discipline assumes that we are only called 2756 // when the world is stopped. 2757 assert(SafepointSynchronize::is_at_safepoint(), 2758 "world is stopped assumption"); 2759 2760 // Currently the CMS epilogue (see CompactibleFreeListSpace) merely checks 2761 // if linear allocation blocks need to be appropriately marked to allow the 2762 // the blocks to be parsable. We also check here whether we need to nudge the 2763 // CMS collector thread to start a new cycle (if it's not already active). 2764 assert( Thread::current()->is_VM_thread() 2765 || ( CMSScavengeBeforeRemark 2766 && Thread::current()->is_ConcurrentGC_thread()), 2767 "Incorrect thread type for epilogue execution"); 2768 2769 if (!_between_prologue_and_epilogue) { 2770 // We have already been invoked; this is a gc_epilogue delegation 2771 // from yet another CMS generation that we are responsible for, just 2772 // ignore it since all relevant work has already been done. 2773 return; 2774 } 2775 assert(haveFreelistLocks(), "must have freelist locks"); 2776 assert_lock_strong(bitMapLock()); 2777 2778 _cmsGen->gc_epilogue_work(full); 2779 _permGen->gc_epilogue_work(full); 2780 2781 if (_collectorState == AbortablePreclean || _collectorState == Precleaning) { 2782 // in case sampling was not already enabled, enable it 2783 _start_sampling = true; 2784 } 2785 // reset _eden_chunk_array so sampling starts afresh 2786 _eden_chunk_index = 0; 2787 2788 size_t cms_used = _cmsGen->cmsSpace()->used(); 2789 size_t perm_used = _permGen->cmsSpace()->used(); 2790 2791 // update performance counters - this uses a special version of 2792 // update_counters() that allows the utilization to be passed as a 2793 // parameter, avoiding multiple calls to used(). 2794 // 2795 _cmsGen->update_counters(cms_used); 2796 _permGen->update_counters(perm_used); 2797 2798 if (CMSIncrementalMode) { 2799 icms_update_allocation_limits(); 2800 } 2801 2802 bitMapLock()->unlock(); 2803 releaseFreelistLocks(); 2804 2805 if (!CleanChunkPoolAsync) { 2806 Chunk::clean_chunk_pool(); 2807 } 2808 2809 _between_prologue_and_epilogue = false; // ready for next cycle 2810 } 2811 2812 void ConcurrentMarkSweepGeneration::gc_epilogue(bool full) { 2813 collector()->gc_epilogue(full); 2814 2815 // Also reset promotion tracking in par gc thread states. 2816 if (CollectedHeap::use_parallel_gc_threads()) { 2817 for (uint i = 0; i < ParallelGCThreads; i++) { 2818 _par_gc_thread_states[i]->promo.stopTrackingPromotions(i); 2819 } 2820 } 2821 } 2822 2823 void ConcurrentMarkSweepGeneration::gc_epilogue_work(bool full) { 2824 assert(!incremental_collection_failed(), "Should have been cleared"); 2825 cmsSpace()->setPreconsumptionDirtyCardClosure(NULL); 2826 cmsSpace()->gc_epilogue(); 2827 // Print stat counters 2828 NOT_PRODUCT( 2829 assert(_numObjectsAllocated == 0, "check"); 2830 assert(_numWordsAllocated == 0, "check"); 2831 if (Verbose && PrintGC) { 2832 gclog_or_tty->print("Promoted "SIZE_FORMAT" objects, " 2833 SIZE_FORMAT" bytes", 2834 _numObjectsPromoted, _numWordsPromoted*sizeof(HeapWord)); 2835 } 2836 _numObjectsPromoted = 0; 2837 _numWordsPromoted = 0; 2838 ) 2839 2840 if (PrintGC && Verbose) { 2841 // Call down the chain in contiguous_available needs the freelistLock 2842 // so print this out before releasing the freeListLock. 2843 gclog_or_tty->print(" Contiguous available "SIZE_FORMAT" bytes ", 2844 contiguous_available()); 2845 } 2846 } 2847 2848 #ifndef PRODUCT 2849 bool CMSCollector::have_cms_token() { 2850 Thread* thr = Thread::current(); 2851 if (thr->is_VM_thread()) { 2852 return ConcurrentMarkSweepThread::vm_thread_has_cms_token(); 2853 } else if (thr->is_ConcurrentGC_thread()) { 2854 return ConcurrentMarkSweepThread::cms_thread_has_cms_token(); 2855 } else if (thr->is_GC_task_thread()) { 2856 return ConcurrentMarkSweepThread::vm_thread_has_cms_token() && 2857 ParGCRareEvent_lock->owned_by_self(); 2858 } 2859 return false; 2860 } 2861 #endif 2862 2863 // Check reachability of the given heap address in CMS generation, 2864 // treating all other generations as roots. 2865 bool CMSCollector::is_cms_reachable(HeapWord* addr) { 2866 // We could "guarantee" below, rather than assert, but i'll 2867 // leave these as "asserts" so that an adventurous debugger 2868 // could try this in the product build provided some subset of 2869 // the conditions were met, provided they were intersted in the 2870 // results and knew that the computation below wouldn't interfere 2871 // with other concurrent computations mutating the structures 2872 // being read or written. 2873 assert(SafepointSynchronize::is_at_safepoint(), 2874 "Else mutations in object graph will make answer suspect"); 2875 assert(have_cms_token(), "Should hold cms token"); 2876 assert(haveFreelistLocks(), "must hold free list locks"); 2877 assert_lock_strong(bitMapLock()); 2878 2879 // Clear the marking bit map array before starting, but, just 2880 // for kicks, first report if the given address is already marked 2881 gclog_or_tty->print_cr("Start: Address 0x%x is%s marked", addr, 2882 _markBitMap.isMarked(addr) ? "" : " not"); 2883 2884 if (verify_after_remark()) { 2885 MutexLockerEx x(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag); 2886 bool result = verification_mark_bm()->isMarked(addr); 2887 gclog_or_tty->print_cr("TransitiveMark: Address 0x%x %s marked", addr, 2888 result ? "IS" : "is NOT"); 2889 return result; 2890 } else { 2891 gclog_or_tty->print_cr("Could not compute result"); 2892 return false; 2893 } 2894 } 2895 2896 //////////////////////////////////////////////////////// 2897 // CMS Verification Support 2898 //////////////////////////////////////////////////////// 2899 // Following the remark phase, the following invariant 2900 // should hold -- each object in the CMS heap which is 2901 // marked in markBitMap() should be marked in the verification_mark_bm(). 2902 2903 class VerifyMarkedClosure: public BitMapClosure { 2904 CMSBitMap* _marks; 2905 bool _failed; 2906 2907 public: 2908 VerifyMarkedClosure(CMSBitMap* bm): _marks(bm), _failed(false) {} 2909 2910 bool do_bit(size_t offset) { 2911 HeapWord* addr = _marks->offsetToHeapWord(offset); 2912 if (!_marks->isMarked(addr)) { 2913 oop(addr)->print_on(gclog_or_tty); 2914 gclog_or_tty->print_cr(" ("INTPTR_FORMAT" should have been marked)", addr); 2915 _failed = true; 2916 } 2917 return true; 2918 } 2919 2920 bool failed() { return _failed; } 2921 }; 2922 2923 bool CMSCollector::verify_after_remark(bool silent) { 2924 if (!silent) gclog_or_tty->print(" [Verifying CMS Marking... "); 2925 MutexLockerEx ml(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag); 2926 static bool init = false; 2927 2928 assert(SafepointSynchronize::is_at_safepoint(), 2929 "Else mutations in object graph will make answer suspect"); 2930 assert(have_cms_token(), 2931 "Else there may be mutual interference in use of " 2932 " verification data structures"); 2933 assert(_collectorState > Marking && _collectorState <= Sweeping, 2934 "Else marking info checked here may be obsolete"); 2935 assert(haveFreelistLocks(), "must hold free list locks"); 2936 assert_lock_strong(bitMapLock()); 2937 2938 2939 // Allocate marking bit map if not already allocated 2940 if (!init) { // first time 2941 if (!verification_mark_bm()->allocate(_span)) { 2942 return false; 2943 } 2944 init = true; 2945 } 2946 2947 assert(verification_mark_stack()->isEmpty(), "Should be empty"); 2948 2949 // Turn off refs discovery -- so we will be tracing through refs. 2950 // This is as intended, because by this time 2951 // GC must already have cleared any refs that need to be cleared, 2952 // and traced those that need to be marked; moreover, 2953 // the marking done here is not going to intefere in any 2954 // way with the marking information used by GC. 2955 NoRefDiscovery no_discovery(ref_processor()); 2956 2957 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;) 2958 2959 // Clear any marks from a previous round 2960 verification_mark_bm()->clear_all(); 2961 assert(verification_mark_stack()->isEmpty(), "markStack should be empty"); 2962 verify_work_stacks_empty(); 2963 2964 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2965 gch->ensure_parsability(false); // fill TLABs, but no need to retire them 2966 // Update the saved marks which may affect the root scans. 2967 gch->save_marks(); 2968 2969 if (CMSRemarkVerifyVariant == 1) { 2970 // In this first variant of verification, we complete 2971 // all marking, then check if the new marks-verctor is 2972 // a subset of the CMS marks-vector. 2973 verify_after_remark_work_1(); 2974 } else if (CMSRemarkVerifyVariant == 2) { 2975 // In this second variant of verification, we flag an error 2976 // (i.e. an object reachable in the new marks-vector not reachable 2977 // in the CMS marks-vector) immediately, also indicating the 2978 // identify of an object (A) that references the unmarked object (B) -- 2979 // presumably, a mutation to A failed to be picked up by preclean/remark? 2980 verify_after_remark_work_2(); 2981 } else { 2982 warning("Unrecognized value %d for CMSRemarkVerifyVariant", 2983 CMSRemarkVerifyVariant); 2984 } 2985 if (!silent) gclog_or_tty->print(" done] "); 2986 return true; 2987 } 2988 2989 void CMSCollector::verify_after_remark_work_1() { 2990 ResourceMark rm; 2991 HandleMark hm; 2992 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2993 2994 // Mark from roots one level into CMS 2995 MarkRefsIntoClosure notOlder(_span, verification_mark_bm()); 2996 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2997 2998 gch->gen_process_strong_roots(_cmsGen->level(), 2999 true, // younger gens are roots 3000 true, // activate StrongRootsScope 3001 true, // collecting perm gen 3002 SharedHeap::ScanningOption(roots_scanning_options()), 3003 ¬Older, 3004 true, // walk code active on stacks 3005 NULL); 3006 3007 // Now mark from the roots 3008 assert(_revisitStack.isEmpty(), "Should be empty"); 3009 MarkFromRootsClosure markFromRootsClosure(this, _span, 3010 verification_mark_bm(), verification_mark_stack(), &_revisitStack, 3011 false /* don't yield */, true /* verifying */); 3012 assert(_restart_addr == NULL, "Expected pre-condition"); 3013 verification_mark_bm()->iterate(&markFromRootsClosure); 3014 while (_restart_addr != NULL) { 3015 // Deal with stack overflow: by restarting at the indicated 3016 // address. 3017 HeapWord* ra = _restart_addr; 3018 markFromRootsClosure.reset(ra); 3019 _restart_addr = NULL; 3020 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end()); 3021 } 3022 assert(verification_mark_stack()->isEmpty(), "Should have been drained"); 3023 verify_work_stacks_empty(); 3024 // Should reset the revisit stack above, since no class tree 3025 // surgery is forthcoming. 3026 _revisitStack.reset(); // throwing away all contents 3027 3028 // Marking completed -- now verify that each bit marked in 3029 // verification_mark_bm() is also marked in markBitMap(); flag all 3030 // errors by printing corresponding objects. 3031 VerifyMarkedClosure vcl(markBitMap()); 3032 verification_mark_bm()->iterate(&vcl); 3033 if (vcl.failed()) { 3034 gclog_or_tty->print("Verification failed"); 3035 Universe::heap()->print_on(gclog_or_tty); 3036 fatal("CMS: failed marking verification after remark"); 3037 } 3038 } 3039 3040 void CMSCollector::verify_after_remark_work_2() { 3041 ResourceMark rm; 3042 HandleMark hm; 3043 GenCollectedHeap* gch = GenCollectedHeap::heap(); 3044 3045 // Mark from roots one level into CMS 3046 MarkRefsIntoVerifyClosure notOlder(_span, verification_mark_bm(), 3047 markBitMap()); 3048 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 3049 gch->gen_process_strong_roots(_cmsGen->level(), 3050 true, // younger gens are roots 3051 true, // activate StrongRootsScope 3052 true, // collecting perm gen 3053 SharedHeap::ScanningOption(roots_scanning_options()), 3054 ¬Older, 3055 true, // walk code active on stacks 3056 NULL); 3057 3058 // Now mark from the roots 3059 assert(_revisitStack.isEmpty(), "Should be empty"); 3060 MarkFromRootsVerifyClosure markFromRootsClosure(this, _span, 3061 verification_mark_bm(), markBitMap(), verification_mark_stack()); 3062 assert(_restart_addr == NULL, "Expected pre-condition"); 3063 verification_mark_bm()->iterate(&markFromRootsClosure); 3064 while (_restart_addr != NULL) { 3065 // Deal with stack overflow: by restarting at the indicated 3066 // address. 3067 HeapWord* ra = _restart_addr; 3068 markFromRootsClosure.reset(ra); 3069 _restart_addr = NULL; 3070 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end()); 3071 } 3072 assert(verification_mark_stack()->isEmpty(), "Should have been drained"); 3073 verify_work_stacks_empty(); 3074 // Should reset the revisit stack above, since no class tree 3075 // surgery is forthcoming. 3076 _revisitStack.reset(); // throwing away all contents 3077 3078 // Marking completed -- now verify that each bit marked in 3079 // verification_mark_bm() is also marked in markBitMap(); flag all 3080 // errors by printing corresponding objects. 3081 VerifyMarkedClosure vcl(markBitMap()); 3082 verification_mark_bm()->iterate(&vcl); 3083 assert(!vcl.failed(), "Else verification above should not have succeeded"); 3084 } 3085 3086 void ConcurrentMarkSweepGeneration::save_marks() { 3087 // delegate to CMS space 3088 cmsSpace()->save_marks(); 3089 for (uint i = 0; i < ParallelGCThreads; i++) { 3090 _par_gc_thread_states[i]->promo.startTrackingPromotions(); 3091 } 3092 } 3093 3094 bool ConcurrentMarkSweepGeneration::no_allocs_since_save_marks() { 3095 return cmsSpace()->no_allocs_since_save_marks(); 3096 } 3097 3098 #define CMS_SINCE_SAVE_MARKS_DEFN(OopClosureType, nv_suffix) \ 3099 \ 3100 void ConcurrentMarkSweepGeneration:: \ 3101 oop_since_save_marks_iterate##nv_suffix(OopClosureType* cl) { \ 3102 cl->set_generation(this); \ 3103 cmsSpace()->oop_since_save_marks_iterate##nv_suffix(cl); \ 3104 cl->reset_generation(); \ 3105 save_marks(); \ 3106 } 3107 3108 ALL_SINCE_SAVE_MARKS_CLOSURES(CMS_SINCE_SAVE_MARKS_DEFN) 3109 3110 void 3111 ConcurrentMarkSweepGeneration::object_iterate_since_last_GC(ObjectClosure* blk) 3112 { 3113 // Not currently implemented; need to do the following. -- ysr. 3114 // dld -- I think that is used for some sort of allocation profiler. So it 3115 // really means the objects allocated by the mutator since the last 3116 // GC. We could potentially implement this cheaply by recording only 3117 // the direct allocations in a side data structure. 3118 // 3119 // I think we probably ought not to be required to support these 3120 // iterations at any arbitrary point; I think there ought to be some 3121 // call to enable/disable allocation profiling in a generation/space, 3122 // and the iterator ought to return the objects allocated in the 3123 // gen/space since the enable call, or the last iterator call (which 3124 // will probably be at a GC.) That way, for gens like CM&S that would 3125 // require some extra data structure to support this, we only pay the 3126 // cost when it's in use... 3127 cmsSpace()->object_iterate_since_last_GC(blk); 3128 } 3129 3130 void 3131 ConcurrentMarkSweepGeneration::younger_refs_iterate(OopsInGenClosure* cl) { 3132 cl->set_generation(this); 3133 younger_refs_in_space_iterate(_cmsSpace, cl); 3134 cl->reset_generation(); 3135 } 3136 3137 void 3138 ConcurrentMarkSweepGeneration::oop_iterate(MemRegion mr, OopClosure* cl) { 3139 if (freelistLock()->owned_by_self()) { 3140 Generation::oop_iterate(mr, cl); 3141 } else { 3142 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 3143 Generation::oop_iterate(mr, cl); 3144 } 3145 } 3146 3147 void 3148 ConcurrentMarkSweepGeneration::oop_iterate(OopClosure* cl) { 3149 if (freelistLock()->owned_by_self()) { 3150 Generation::oop_iterate(cl); 3151 } else { 3152 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 3153 Generation::oop_iterate(cl); 3154 } 3155 } 3156 3157 void 3158 ConcurrentMarkSweepGeneration::object_iterate(ObjectClosure* cl) { 3159 if (freelistLock()->owned_by_self()) { 3160 Generation::object_iterate(cl); 3161 } else { 3162 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 3163 Generation::object_iterate(cl); 3164 } 3165 } 3166 3167 void 3168 ConcurrentMarkSweepGeneration::safe_object_iterate(ObjectClosure* cl) { 3169 if (freelistLock()->owned_by_self()) { 3170 Generation::safe_object_iterate(cl); 3171 } else { 3172 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 3173 Generation::safe_object_iterate(cl); 3174 } 3175 } 3176 3177 void 3178 ConcurrentMarkSweepGeneration::pre_adjust_pointers() { 3179 } 3180 3181 void 3182 ConcurrentMarkSweepGeneration::post_compact() { 3183 } 3184 3185 void 3186 ConcurrentMarkSweepGeneration::prepare_for_verify() { 3187 // Fix the linear allocation blocks to look like free blocks. 3188 3189 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those 3190 // are not called when the heap is verified during universe initialization and 3191 // at vm shutdown. 3192 if (freelistLock()->owned_by_self()) { 3193 cmsSpace()->prepare_for_verify(); 3194 } else { 3195 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag); 3196 cmsSpace()->prepare_for_verify(); 3197 } 3198 } 3199 3200 void 3201 ConcurrentMarkSweepGeneration::verify() { 3202 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those 3203 // are not called when the heap is verified during universe initialization and 3204 // at vm shutdown. 3205 if (freelistLock()->owned_by_self()) { 3206 cmsSpace()->verify(); 3207 } else { 3208 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag); 3209 cmsSpace()->verify(); 3210 } 3211 } 3212 3213 void CMSCollector::verify() { 3214 _cmsGen->verify(); 3215 _permGen->verify(); 3216 } 3217 3218 #ifndef PRODUCT 3219 bool CMSCollector::overflow_list_is_empty() const { 3220 assert(_num_par_pushes >= 0, "Inconsistency"); 3221 if (_overflow_list == NULL) { 3222 assert(_num_par_pushes == 0, "Inconsistency"); 3223 } 3224 return _overflow_list == NULL; 3225 } 3226 3227 // The methods verify_work_stacks_empty() and verify_overflow_empty() 3228 // merely consolidate assertion checks that appear to occur together frequently. 3229 void CMSCollector::verify_work_stacks_empty() const { 3230 assert(_markStack.isEmpty(), "Marking stack should be empty"); 3231 assert(overflow_list_is_empty(), "Overflow list should be empty"); 3232 } 3233 3234 void CMSCollector::verify_overflow_empty() const { 3235 assert(overflow_list_is_empty(), "Overflow list should be empty"); 3236 assert(no_preserved_marks(), "No preserved marks"); 3237 } 3238 #endif // PRODUCT 3239 3240 // Decide if we want to enable class unloading as part of the 3241 // ensuing concurrent GC cycle. We will collect the perm gen and 3242 // unload classes if it's the case that: 3243 // (1) an explicit gc request has been made and the flag 3244 // ExplicitGCInvokesConcurrentAndUnloadsClasses is set, OR 3245 // (2) (a) class unloading is enabled at the command line, and 3246 // (b) (i) perm gen threshold has been crossed, or 3247 // (ii) old gen is getting really full, or 3248 // (iii) the previous N CMS collections did not collect the 3249 // perm gen 3250 // NOTE: Provided there is no change in the state of the heap between 3251 // calls to this method, it should have idempotent results. Moreover, 3252 // its results should be monotonically increasing (i.e. going from 0 to 1, 3253 // but not 1 to 0) between successive calls between which the heap was 3254 // not collected. For the implementation below, it must thus rely on 3255 // the property that concurrent_cycles_since_last_unload() 3256 // will not decrease unless a collection cycle happened and that 3257 // _permGen->should_concurrent_collect() and _cmsGen->is_too_full() are 3258 // themselves also monotonic in that sense. See check_monotonicity() 3259 // below. 3260 bool CMSCollector::update_should_unload_classes() { 3261 _should_unload_classes = false; 3262 // Condition 1 above 3263 if (_full_gc_requested && ExplicitGCInvokesConcurrentAndUnloadsClasses) { 3264 _should_unload_classes = true; 3265 } else if (CMSClassUnloadingEnabled) { // Condition 2.a above 3266 // Disjuncts 2.b.(i,ii,iii) above 3267 _should_unload_classes = (concurrent_cycles_since_last_unload() >= 3268 CMSClassUnloadingMaxInterval) 3269 || _permGen->should_concurrent_collect() 3270 || _cmsGen->is_too_full(); 3271 } 3272 return _should_unload_classes; 3273 } 3274 3275 bool ConcurrentMarkSweepGeneration::is_too_full() const { 3276 bool res = should_concurrent_collect(); 3277 res = res && (occupancy() > (double)CMSIsTooFullPercentage/100.0); 3278 return res; 3279 } 3280 3281 void CMSCollector::setup_cms_unloading_and_verification_state() { 3282 const bool should_verify = VerifyBeforeGC || VerifyAfterGC || VerifyDuringGC 3283 || VerifyBeforeExit; 3284 const int rso = SharedHeap::SO_Strings | SharedHeap::SO_CodeCache; 3285 3286 if (should_unload_classes()) { // Should unload classes this cycle 3287 remove_root_scanning_option(rso); // Shrink the root set appropriately 3288 set_verifying(should_verify); // Set verification state for this cycle 3289 return; // Nothing else needs to be done at this time 3290 } 3291 3292 // Not unloading classes this cycle 3293 assert(!should_unload_classes(), "Inconsitency!"); 3294 if ((!verifying() || unloaded_classes_last_cycle()) && should_verify) { 3295 // We were not verifying, or we _were_ unloading classes in the last cycle, 3296 // AND some verification options are enabled this cycle; in this case, 3297 // we must make sure that the deadness map is allocated if not already so, 3298 // and cleared (if already allocated previously -- 3299 // CMSBitMap::sizeInBits() is used to determine if it's allocated). 3300 if (perm_gen_verify_bit_map()->sizeInBits() == 0) { 3301 if (!perm_gen_verify_bit_map()->allocate(_permGen->reserved())) { 3302 warning("Failed to allocate permanent generation verification CMS Bit Map;\n" 3303 "permanent generation verification disabled"); 3304 return; // Note that we leave verification disabled, so we'll retry this 3305 // allocation next cycle. We _could_ remember this failure 3306 // and skip further attempts and permanently disable verification 3307 // attempts if that is considered more desirable. 3308 } 3309 assert(perm_gen_verify_bit_map()->covers(_permGen->reserved()), 3310 "_perm_gen_ver_bit_map inconsistency?"); 3311 } else { 3312 perm_gen_verify_bit_map()->clear_all(); 3313 } 3314 // Include symbols, strings and code cache elements to prevent their resurrection. 3315 add_root_scanning_option(rso); 3316 set_verifying(true); 3317 } else if (verifying() && !should_verify) { 3318 // We were verifying, but some verification flags got disabled. 3319 set_verifying(false); 3320 // Exclude symbols, strings and code cache elements from root scanning to 3321 // reduce IM and RM pauses. 3322 remove_root_scanning_option(rso); 3323 } 3324 } 3325 3326 3327 #ifndef PRODUCT 3328 HeapWord* CMSCollector::block_start(const void* p) const { 3329 const HeapWord* addr = (HeapWord*)p; 3330 if (_span.contains(p)) { 3331 if (_cmsGen->cmsSpace()->is_in_reserved(addr)) { 3332 return _cmsGen->cmsSpace()->block_start(p); 3333 } else { 3334 assert(_permGen->cmsSpace()->is_in_reserved(addr), 3335 "Inconsistent _span?"); 3336 return _permGen->cmsSpace()->block_start(p); 3337 } 3338 } 3339 return NULL; 3340 } 3341 #endif 3342 3343 HeapWord* 3344 ConcurrentMarkSweepGeneration::expand_and_allocate(size_t word_size, 3345 bool tlab, 3346 bool parallel) { 3347 CMSSynchronousYieldRequest yr; 3348 assert(!tlab, "Can't deal with TLAB allocation"); 3349 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 3350 expand(word_size*HeapWordSize, MinHeapDeltaBytes, 3351 CMSExpansionCause::_satisfy_allocation); 3352 if (GCExpandToAllocateDelayMillis > 0) { 3353 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 3354 } 3355 return have_lock_and_allocate(word_size, tlab); 3356 } 3357 3358 // YSR: All of this generation expansion/shrinking stuff is an exact copy of 3359 // OneContigSpaceCardGeneration, which makes me wonder if we should move this 3360 // to CardGeneration and share it... 3361 bool ConcurrentMarkSweepGeneration::expand(size_t bytes, size_t expand_bytes) { 3362 return CardGeneration::expand(bytes, expand_bytes); 3363 } 3364 3365 void ConcurrentMarkSweepGeneration::expand(size_t bytes, size_t expand_bytes, 3366 CMSExpansionCause::Cause cause) 3367 { 3368 3369 bool success = expand(bytes, expand_bytes); 3370 3371 // remember why we expanded; this information is used 3372 // by shouldConcurrentCollect() when making decisions on whether to start 3373 // a new CMS cycle. 3374 if (success) { 3375 set_expansion_cause(cause); 3376 if (PrintGCDetails && Verbose) { 3377 gclog_or_tty->print_cr("Expanded CMS gen for %s", 3378 CMSExpansionCause::to_string(cause)); 3379 } 3380 } 3381 } 3382 3383 HeapWord* ConcurrentMarkSweepGeneration::expand_and_par_lab_allocate(CMSParGCThreadState* ps, size_t word_sz) { 3384 HeapWord* res = NULL; 3385 MutexLocker x(ParGCRareEvent_lock); 3386 while (true) { 3387 // Expansion by some other thread might make alloc OK now: 3388 res = ps->lab.alloc(word_sz); 3389 if (res != NULL) return res; 3390 // If there's not enough expansion space available, give up. 3391 if (_virtual_space.uncommitted_size() < (word_sz * HeapWordSize)) { 3392 return NULL; 3393 } 3394 // Otherwise, we try expansion. 3395 expand(word_sz*HeapWordSize, MinHeapDeltaBytes, 3396 CMSExpansionCause::_allocate_par_lab); 3397 // Now go around the loop and try alloc again; 3398 // A competing par_promote might beat us to the expansion space, 3399 // so we may go around the loop again if promotion fails agaion. 3400 if (GCExpandToAllocateDelayMillis > 0) { 3401 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 3402 } 3403 } 3404 } 3405 3406 3407 bool ConcurrentMarkSweepGeneration::expand_and_ensure_spooling_space( 3408 PromotionInfo* promo) { 3409 MutexLocker x(ParGCRareEvent_lock); 3410 size_t refill_size_bytes = promo->refillSize() * HeapWordSize; 3411 while (true) { 3412 // Expansion by some other thread might make alloc OK now: 3413 if (promo->ensure_spooling_space()) { 3414 assert(promo->has_spooling_space(), 3415 "Post-condition of successful ensure_spooling_space()"); 3416 return true; 3417 } 3418 // If there's not enough expansion space available, give up. 3419 if (_virtual_space.uncommitted_size() < refill_size_bytes) { 3420 return false; 3421 } 3422 // Otherwise, we try expansion. 3423 expand(refill_size_bytes, MinHeapDeltaBytes, 3424 CMSExpansionCause::_allocate_par_spooling_space); 3425 // Now go around the loop and try alloc again; 3426 // A competing allocation might beat us to the expansion space, 3427 // so we may go around the loop again if allocation fails again. 3428 if (GCExpandToAllocateDelayMillis > 0) { 3429 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 3430 } 3431 } 3432 } 3433 3434 3435 3436 void ConcurrentMarkSweepGeneration::shrink(size_t bytes) { 3437 assert_locked_or_safepoint(Heap_lock); 3438 size_t size = ReservedSpace::page_align_size_down(bytes); 3439 if (size > 0) { 3440 shrink_by(size); 3441 } 3442 } 3443 3444 bool ConcurrentMarkSweepGeneration::grow_by(size_t bytes) { 3445 assert_locked_or_safepoint(Heap_lock); 3446 bool result = _virtual_space.expand_by(bytes); 3447 if (result) { 3448 HeapWord* old_end = _cmsSpace->end(); 3449 size_t new_word_size = 3450 heap_word_size(_virtual_space.committed_size()); 3451 MemRegion mr(_cmsSpace->bottom(), new_word_size); 3452 _bts->resize(new_word_size); // resize the block offset shared array 3453 Universe::heap()->barrier_set()->resize_covered_region(mr); 3454 // Hmmmm... why doesn't CFLS::set_end verify locking? 3455 // This is quite ugly; FIX ME XXX 3456 _cmsSpace->assert_locked(freelistLock()); 3457 _cmsSpace->set_end((HeapWord*)_virtual_space.high()); 3458 3459 // update the space and generation capacity counters 3460 if (UsePerfData) { 3461 _space_counters->update_capacity(); 3462 _gen_counters->update_all(); 3463 } 3464 3465 if (Verbose && PrintGC) { 3466 size_t new_mem_size = _virtual_space.committed_size(); 3467 size_t old_mem_size = new_mem_size - bytes; 3468 gclog_or_tty->print_cr("Expanding %s from %ldK by %ldK to %ldK", 3469 name(), old_mem_size/K, bytes/K, new_mem_size/K); 3470 } 3471 } 3472 return result; 3473 } 3474 3475 bool ConcurrentMarkSweepGeneration::grow_to_reserved() { 3476 assert_locked_or_safepoint(Heap_lock); 3477 bool success = true; 3478 const size_t remaining_bytes = _virtual_space.uncommitted_size(); 3479 if (remaining_bytes > 0) { 3480 success = grow_by(remaining_bytes); 3481 DEBUG_ONLY(if (!success) warning("grow to reserved failed");) 3482 } 3483 return success; 3484 } 3485 3486 void ConcurrentMarkSweepGeneration::shrink_by(size_t bytes) { 3487 assert_locked_or_safepoint(Heap_lock); 3488 assert_lock_strong(freelistLock()); 3489 // XXX Fix when compaction is implemented. 3490 warning("Shrinking of CMS not yet implemented"); 3491 return; 3492 } 3493 3494 3495 // Simple ctor/dtor wrapper for accounting & timer chores around concurrent 3496 // phases. 3497 class CMSPhaseAccounting: public StackObj { 3498 public: 3499 CMSPhaseAccounting(CMSCollector *collector, 3500 const char *phase, 3501 bool print_cr = true); 3502 ~CMSPhaseAccounting(); 3503 3504 private: 3505 CMSCollector *_collector; 3506 const char *_phase; 3507 elapsedTimer _wallclock; 3508 bool _print_cr; 3509 3510 public: 3511 // Not MT-safe; so do not pass around these StackObj's 3512 // where they may be accessed by other threads. 3513 jlong wallclock_millis() { 3514 assert(_wallclock.is_active(), "Wall clock should not stop"); 3515 _wallclock.stop(); // to record time 3516 jlong ret = _wallclock.milliseconds(); 3517 _wallclock.start(); // restart 3518 return ret; 3519 } 3520 }; 3521 3522 CMSPhaseAccounting::CMSPhaseAccounting(CMSCollector *collector, 3523 const char *phase, 3524 bool print_cr) : 3525 _collector(collector), _phase(phase), _print_cr(print_cr) { 3526 3527 if (PrintCMSStatistics != 0) { 3528 _collector->resetYields(); 3529 } 3530 if (PrintGCDetails && PrintGCTimeStamps) { 3531 gclog_or_tty->date_stamp(PrintGCDateStamps); 3532 gclog_or_tty->stamp(); 3533 gclog_or_tty->print_cr(": [%s-concurrent-%s-start]", 3534 _collector->cmsGen()->short_name(), _phase); 3535 } 3536 _collector->resetTimer(); 3537 _wallclock.start(); 3538 _collector->startTimer(); 3539 } 3540 3541 CMSPhaseAccounting::~CMSPhaseAccounting() { 3542 assert(_wallclock.is_active(), "Wall clock should not have stopped"); 3543 _collector->stopTimer(); 3544 _wallclock.stop(); 3545 if (PrintGCDetails) { 3546 gclog_or_tty->date_stamp(PrintGCDateStamps); 3547 gclog_or_tty->stamp(PrintGCTimeStamps); 3548 gclog_or_tty->print("[%s-concurrent-%s: %3.3f/%3.3f secs]", 3549 _collector->cmsGen()->short_name(), 3550 _phase, _collector->timerValue(), _wallclock.seconds()); 3551 if (_print_cr) { 3552 gclog_or_tty->print_cr(""); 3553 } 3554 if (PrintCMSStatistics != 0) { 3555 gclog_or_tty->print_cr(" (CMS-concurrent-%s yielded %d times)", _phase, 3556 _collector->yields()); 3557 } 3558 } 3559 } 3560 3561 // CMS work 3562 3563 // The common parts of CMSParInitialMarkTask and CMSParRemarkTask. 3564 class CMSParMarkTask : public AbstractGangTask { 3565 protected: 3566 CMSCollector* _collector; 3567 int _n_workers; 3568 CMSParMarkTask(const char* name, CMSCollector* collector, int n_workers) : 3569 AbstractGangTask(name), 3570 _collector(collector), 3571 _n_workers(n_workers) {} 3572 // Work method in support of parallel rescan ... of young gen spaces 3573 void do_young_space_rescan(uint worker_id, OopsInGenClosure* cl, 3574 ContiguousSpace* space, 3575 HeapWord** chunk_array, size_t chunk_top); 3576 void work_on_young_gen_roots(uint worker_id, OopsInGenClosure* cl); 3577 }; 3578 3579 // Parallel initial mark task 3580 class CMSParInitialMarkTask: public CMSParMarkTask { 3581 public: 3582 CMSParInitialMarkTask(CMSCollector* collector, int n_workers) : 3583 CMSParMarkTask("Scan roots and young gen for initial mark in parallel", 3584 collector, n_workers) {} 3585 void work(uint worker_id); 3586 }; 3587 3588 // Checkpoint the roots into this generation from outside 3589 // this generation. [Note this initial checkpoint need only 3590 // be approximate -- we'll do a catch up phase subsequently.] 3591 void CMSCollector::checkpointRootsInitial(bool asynch) { 3592 assert(_collectorState == InitialMarking, "Wrong collector state"); 3593 check_correct_thread_executing(); 3594 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 3595 3596 save_heap_summary(); 3597 report_heap_summary(GCWhen::BeforeGC); 3598 3599 ReferenceProcessor* rp = ref_processor(); 3600 SpecializationStats::clear(); 3601 assert(_restart_addr == NULL, "Control point invariant"); 3602 if (asynch) { 3603 // acquire locks for subsequent manipulations 3604 MutexLockerEx x(bitMapLock(), 3605 Mutex::_no_safepoint_check_flag); 3606 checkpointRootsInitialWork(asynch); 3607 // enable ("weak") refs discovery 3608 rp->enable_discovery(true /*verify_disabled*/, true /*check_no_refs*/); 3609 _collectorState = Marking; 3610 } else { 3611 // (Weak) Refs discovery: this is controlled from genCollectedHeap::do_collection 3612 // which recognizes if we are a CMS generation, and doesn't try to turn on 3613 // discovery; verify that they aren't meddling. 3614 assert(!rp->discovery_is_atomic(), 3615 "incorrect setting of discovery predicate"); 3616 assert(!rp->discovery_enabled(), "genCollectedHeap shouldn't control " 3617 "ref discovery for this generation kind"); 3618 // already have locks 3619 checkpointRootsInitialWork(asynch); 3620 // now enable ("weak") refs discovery 3621 rp->enable_discovery(true /*verify_disabled*/, false /*verify_no_refs*/); 3622 _collectorState = Marking; 3623 } 3624 SpecializationStats::print(); 3625 } 3626 3627 void CMSCollector::checkpointRootsInitialWork(bool asynch) { 3628 assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped"); 3629 assert(_collectorState == InitialMarking, "just checking"); 3630 3631 // If there has not been a GC[n-1] since last GC[n] cycle completed, 3632 // precede our marking with a collection of all 3633 // younger generations to keep floating garbage to a minimum. 3634 // XXX: we won't do this for now -- it's an optimization to be done later. 3635 3636 // already have locks 3637 assert_lock_strong(bitMapLock()); 3638 assert(_markBitMap.isAllClear(), "was reset at end of previous cycle"); 3639 3640 // Setup the verification and class unloading state for this 3641 // CMS collection cycle. 3642 setup_cms_unloading_and_verification_state(); 3643 3644 NOT_PRODUCT(GCTraceTime t("\ncheckpointRootsInitialWork", 3645 PrintGCDetails && Verbose, true, _gc_timer_cm);) 3646 if (UseAdaptiveSizePolicy) { 3647 size_policy()->checkpoint_roots_initial_begin(); 3648 } 3649 3650 // Reset all the PLAB chunk arrays if necessary. 3651 if (_survivor_plab_array != NULL && !CMSPLABRecordAlways) { 3652 reset_survivor_plab_arrays(); 3653 } 3654 3655 ResourceMark rm; 3656 HandleMark hm; 3657 3658 FalseClosure falseClosure; 3659 // In the case of a synchronous collection, we will elide the 3660 // remark step, so it's important to catch all the nmethod oops 3661 // in this step. 3662 // The final 'true' flag to gen_process_strong_roots will ensure this. 3663 // If 'async' is true, we can relax the nmethod tracing. 3664 MarkRefsIntoClosure notOlder(_span, &_markBitMap); 3665 GenCollectedHeap* gch = GenCollectedHeap::heap(); 3666 3667 verify_work_stacks_empty(); 3668 verify_overflow_empty(); 3669 3670 gch->ensure_parsability(false); // fill TLABs, but no need to retire them 3671 // Update the saved marks which may affect the root scans. 3672 gch->save_marks(); 3673 3674 // weak reference processing has not started yet. 3675 ref_processor()->set_enqueuing_is_done(false); 3676 3677 if (CMSPrintEdenSurvivorChunks) { 3678 print_eden_and_survivor_chunk_arrays(); 3679 } 3680 3681 { 3682 // This is not needed. DEBUG_ONLY(RememberKlassesChecker imx(true);) 3683 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;) 3684 if (CMSParallelInitialMarkEnabled && CollectedHeap::use_parallel_gc_threads()) { 3685 // The parallel version. 3686 FlexibleWorkGang* workers = gch->workers(); 3687 assert(workers != NULL, "Need parallel worker threads."); 3688 int n_workers = workers->active_workers(); 3689 CMSParInitialMarkTask tsk(this, n_workers); 3690 gch->set_par_threads(n_workers); 3691 initialize_sequential_subtasks_for_young_gen_rescan(n_workers); 3692 if (n_workers > 1) { 3693 GenCollectedHeap::StrongRootsScope srs(gch); 3694 workers->run_task(&tsk); 3695 } else { 3696 GenCollectedHeap::StrongRootsScope srs(gch); 3697 tsk.work(0); 3698 } 3699 gch->set_par_threads(0); 3700 } else { 3701 // The serial version. 3702 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 3703 gch->gen_process_strong_roots(_cmsGen->level(), 3704 true, // younger gens are roots 3705 true, // activate StrongRootsScope 3706 true, // collecting perm gen 3707 SharedHeap::ScanningOption(roots_scanning_options()), 3708 ¬Older, 3709 true, // walk all of code cache if (so & SO_CodeCache) 3710 NULL); 3711 } 3712 } 3713 // Clear mod-union table; it will be dirtied in the prologue of 3714 // CMS generation per each younger generation collection. 3715 3716 assert(_modUnionTable.isAllClear(), 3717 "Was cleared in most recent final checkpoint phase" 3718 " or no bits are set in the gc_prologue before the start of the next " 3719 "subsequent marking phase."); 3720 3721 // Save the end of the used_region of the constituent generations 3722 // to be used to limit the extent of sweep in each generation. 3723 save_sweep_limits(); 3724 if (UseAdaptiveSizePolicy) { 3725 size_policy()->checkpoint_roots_initial_end(gch->gc_cause()); 3726 } 3727 verify_overflow_empty(); 3728 } 3729 3730 bool CMSCollector::markFromRoots(bool asynch) { 3731 // we might be tempted to assert that: 3732 // assert(asynch == !SafepointSynchronize::is_at_safepoint(), 3733 // "inconsistent argument?"); 3734 // However that wouldn't be right, because it's possible that 3735 // a safepoint is indeed in progress as a younger generation 3736 // stop-the-world GC happens even as we mark in this generation. 3737 assert(_collectorState == Marking, "inconsistent state?"); 3738 check_correct_thread_executing(); 3739 verify_overflow_empty(); 3740 3741 bool res; 3742 if (asynch) { 3743 3744 // Start the timers for adaptive size policy for the concurrent phases 3745 // Do it here so that the foreground MS can use the concurrent 3746 // timer since a foreground MS might has the sweep done concurrently 3747 // or STW. 3748 if (UseAdaptiveSizePolicy) { 3749 size_policy()->concurrent_marking_begin(); 3750 } 3751 3752 // Weak ref discovery note: We may be discovering weak 3753 // refs in this generation concurrent (but interleaved) with 3754 // weak ref discovery by a younger generation collector. 3755 3756 CMSTokenSyncWithLocks ts(true, bitMapLock()); 3757 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 3758 CMSPhaseAccounting pa(this, "mark", !PrintGCDetails); 3759 res = markFromRootsWork(asynch); 3760 if (res) { 3761 _collectorState = Precleaning; 3762 } else { // We failed and a foreground collection wants to take over 3763 assert(_foregroundGCIsActive, "internal state inconsistency"); 3764 assert(_restart_addr == NULL, "foreground will restart from scratch"); 3765 if (PrintGCDetails) { 3766 gclog_or_tty->print_cr("bailing out to foreground collection"); 3767 } 3768 } 3769 if (UseAdaptiveSizePolicy) { 3770 size_policy()->concurrent_marking_end(); 3771 } 3772 } else { 3773 assert(SafepointSynchronize::is_at_safepoint(), 3774 "inconsistent with asynch == false"); 3775 if (UseAdaptiveSizePolicy) { 3776 size_policy()->ms_collection_marking_begin(); 3777 } 3778 // already have locks 3779 res = markFromRootsWork(asynch); 3780 _collectorState = FinalMarking; 3781 if (UseAdaptiveSizePolicy) { 3782 GenCollectedHeap* gch = GenCollectedHeap::heap(); 3783 size_policy()->ms_collection_marking_end(gch->gc_cause()); 3784 } 3785 } 3786 verify_overflow_empty(); 3787 return res; 3788 } 3789 3790 bool CMSCollector::markFromRootsWork(bool asynch) { 3791 // iterate over marked bits in bit map, doing a full scan and mark 3792 // from these roots using the following algorithm: 3793 // . if oop is to the right of the current scan pointer, 3794 // mark corresponding bit (we'll process it later) 3795 // . else (oop is to left of current scan pointer) 3796 // push oop on marking stack 3797 // . drain the marking stack 3798 3799 // Note that when we do a marking step we need to hold the 3800 // bit map lock -- recall that direct allocation (by mutators) 3801 // and promotion (by younger generation collectors) is also 3802 // marking the bit map. [the so-called allocate live policy.] 3803 // Because the implementation of bit map marking is not 3804 // robust wrt simultaneous marking of bits in the same word, 3805 // we need to make sure that there is no such interference 3806 // between concurrent such updates. 3807 3808 // already have locks 3809 assert_lock_strong(bitMapLock()); 3810 3811 // Clear the revisit stack, just in case there are any 3812 // obsolete contents from a short-circuited previous CMS cycle. 3813 _revisitStack.reset(); 3814 verify_work_stacks_empty(); 3815 verify_overflow_empty(); 3816 assert(_revisitStack.isEmpty(), "tabula rasa"); 3817 DEBUG_ONLY(RememberKlassesChecker cmx(should_unload_classes());) 3818 bool result = false; 3819 if (CMSConcurrentMTEnabled && ConcGCThreads > 0) { 3820 result = do_marking_mt(asynch); 3821 } else { 3822 result = do_marking_st(asynch); 3823 } 3824 return result; 3825 } 3826 3827 // Forward decl 3828 class CMSConcMarkingTask; 3829 3830 class CMSConcMarkingTerminator: public ParallelTaskTerminator { 3831 CMSCollector* _collector; 3832 CMSConcMarkingTask* _task; 3833 public: 3834 virtual void yield(); 3835 3836 // "n_threads" is the number of threads to be terminated. 3837 // "queue_set" is a set of work queues of other threads. 3838 // "collector" is the CMS collector associated with this task terminator. 3839 // "yield" indicates whether we need the gang as a whole to yield. 3840 CMSConcMarkingTerminator(int n_threads, TaskQueueSetSuper* queue_set, CMSCollector* collector) : 3841 ParallelTaskTerminator(n_threads, queue_set), 3842 _collector(collector) { } 3843 3844 void set_task(CMSConcMarkingTask* task) { 3845 _task = task; 3846 } 3847 }; 3848 3849 class CMSConcMarkingTerminatorTerminator: public TerminatorTerminator { 3850 CMSConcMarkingTask* _task; 3851 public: 3852 bool should_exit_termination(); 3853 void set_task(CMSConcMarkingTask* task) { 3854 _task = task; 3855 } 3856 }; 3857 3858 // MT Concurrent Marking Task 3859 class CMSConcMarkingTask: public YieldingFlexibleGangTask { 3860 CMSCollector* _collector; 3861 int _n_workers; // requested/desired # workers 3862 bool _asynch; 3863 bool _result; 3864 CompactibleFreeListSpace* _cms_space; 3865 CompactibleFreeListSpace* _perm_space; 3866 char _pad_front[64]; // padding to ... 3867 HeapWord* _global_finger; // ... avoid sharing cache line 3868 char _pad_back[64]; 3869 HeapWord* _restart_addr; 3870 3871 // Exposed here for yielding support 3872 Mutex* const _bit_map_lock; 3873 3874 // The per thread work queues, available here for stealing 3875 OopTaskQueueSet* _task_queues; 3876 3877 // Termination (and yielding) support 3878 CMSConcMarkingTerminator _term; 3879 CMSConcMarkingTerminatorTerminator _term_term; 3880 3881 public: 3882 CMSConcMarkingTask(CMSCollector* collector, 3883 CompactibleFreeListSpace* cms_space, 3884 CompactibleFreeListSpace* perm_space, 3885 bool asynch, 3886 YieldingFlexibleWorkGang* workers, 3887 OopTaskQueueSet* task_queues): 3888 YieldingFlexibleGangTask("Concurrent marking done multi-threaded"), 3889 _collector(collector), 3890 _cms_space(cms_space), 3891 _perm_space(perm_space), 3892 _asynch(asynch), _n_workers(0), _result(true), 3893 _task_queues(task_queues), 3894 _term(_n_workers, task_queues, _collector), 3895 _bit_map_lock(collector->bitMapLock()) 3896 { 3897 _requested_size = _n_workers; 3898 _term.set_task(this); 3899 _term_term.set_task(this); 3900 assert(_cms_space->bottom() < _perm_space->bottom(), 3901 "Finger incorrectly initialized below"); 3902 _restart_addr = _global_finger = _cms_space->bottom(); 3903 } 3904 3905 3906 OopTaskQueueSet* task_queues() { return _task_queues; } 3907 3908 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 3909 3910 HeapWord** global_finger_addr() { return &_global_finger; } 3911 3912 CMSConcMarkingTerminator* terminator() { return &_term; } 3913 3914 virtual void set_for_termination(int active_workers) { 3915 terminator()->reset_for_reuse(active_workers); 3916 } 3917 3918 void work(uint worker_id); 3919 bool should_yield() { 3920 return ConcurrentMarkSweepThread::should_yield() 3921 && !_collector->foregroundGCIsActive() 3922 && _asynch; 3923 } 3924 3925 virtual void coordinator_yield(); // stuff done by coordinator 3926 bool result() { return _result; } 3927 3928 void reset(HeapWord* ra) { 3929 assert(_global_finger >= _cms_space->end(), "Postcondition of ::work(i)"); 3930 assert(_global_finger >= _perm_space->end(), "Postcondition of ::work(i)"); 3931 assert(ra < _perm_space->end(), "ra too large"); 3932 _restart_addr = _global_finger = ra; 3933 _term.reset_for_reuse(); 3934 } 3935 3936 static bool get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 3937 OopTaskQueue* work_q); 3938 3939 private: 3940 void do_scan_and_mark(int i, CompactibleFreeListSpace* sp); 3941 void do_work_steal(int i); 3942 void bump_global_finger(HeapWord* f); 3943 }; 3944 3945 bool CMSConcMarkingTerminatorTerminator::should_exit_termination() { 3946 assert(_task != NULL, "Error"); 3947 return _task->yielding(); 3948 // Note that we do not need the disjunct || _task->should_yield() above 3949 // because we want terminating threads to yield only if the task 3950 // is already in the midst of yielding, which happens only after at least one 3951 // thread has yielded. 3952 } 3953 3954 void CMSConcMarkingTerminator::yield() { 3955 if (_task->should_yield()) { 3956 _task->yield(); 3957 } else { 3958 ParallelTaskTerminator::yield(); 3959 } 3960 } 3961 3962 //////////////////////////////////////////////////////////////// 3963 // Concurrent Marking Algorithm Sketch 3964 //////////////////////////////////////////////////////////////// 3965 // Until all tasks exhausted (both spaces): 3966 // -- claim next available chunk 3967 // -- bump global finger via CAS 3968 // -- find first object that starts in this chunk 3969 // and start scanning bitmap from that position 3970 // -- scan marked objects for oops 3971 // -- CAS-mark target, and if successful: 3972 // . if target oop is above global finger (volatile read) 3973 // nothing to do 3974 // . if target oop is in chunk and above local finger 3975 // then nothing to do 3976 // . else push on work-queue 3977 // -- Deal with possible overflow issues: 3978 // . local work-queue overflow causes stuff to be pushed on 3979 // global (common) overflow queue 3980 // . always first empty local work queue 3981 // . then get a batch of oops from global work queue if any 3982 // . then do work stealing 3983 // -- When all tasks claimed (both spaces) 3984 // and local work queue empty, 3985 // then in a loop do: 3986 // . check global overflow stack; steal a batch of oops and trace 3987 // . try to steal from other threads oif GOS is empty 3988 // . if neither is available, offer termination 3989 // -- Terminate and return result 3990 // 3991 void CMSConcMarkingTask::work(uint worker_id) { 3992 elapsedTimer _timer; 3993 ResourceMark rm; 3994 HandleMark hm; 3995 3996 DEBUG_ONLY(_collector->verify_overflow_empty();) 3997 3998 // Before we begin work, our work queue should be empty 3999 assert(work_queue(worker_id)->size() == 0, "Expected to be empty"); 4000 // Scan the bitmap covering _cms_space, tracing through grey objects. 4001 _timer.start(); 4002 do_scan_and_mark(worker_id, _cms_space); 4003 _timer.stop(); 4004 if (PrintCMSStatistics != 0) { 4005 gclog_or_tty->print_cr("Finished cms space scanning in %dth thread: %3.3f sec", 4006 worker_id, _timer.seconds()); 4007 // XXX: need xxx/xxx type of notation, two timers 4008 } 4009 4010 // ... do the same for the _perm_space 4011 _timer.reset(); 4012 _timer.start(); 4013 do_scan_and_mark(worker_id, _perm_space); 4014 _timer.stop(); 4015 if (PrintCMSStatistics != 0) { 4016 gclog_or_tty->print_cr("Finished perm space scanning in %dth thread: %3.3f sec", 4017 worker_id, _timer.seconds()); 4018 // XXX: need xxx/xxx type of notation, two timers 4019 } 4020 4021 // ... do work stealing 4022 _timer.reset(); 4023 _timer.start(); 4024 do_work_steal(worker_id); 4025 _timer.stop(); 4026 if (PrintCMSStatistics != 0) { 4027 gclog_or_tty->print_cr("Finished work stealing in %dth thread: %3.3f sec", 4028 worker_id, _timer.seconds()); 4029 // XXX: need xxx/xxx type of notation, two timers 4030 } 4031 assert(_collector->_markStack.isEmpty(), "Should have been emptied"); 4032 assert(work_queue(worker_id)->size() == 0, "Should have been emptied"); 4033 // Note that under the current task protocol, the 4034 // following assertion is true even of the spaces 4035 // expanded since the completion of the concurrent 4036 // marking. XXX This will likely change under a strict 4037 // ABORT semantics. 4038 assert(_global_finger > _cms_space->end() && 4039 _global_finger >= _perm_space->end(), 4040 "All tasks have been completed"); 4041 DEBUG_ONLY(_collector->verify_overflow_empty();) 4042 } 4043 4044 void CMSConcMarkingTask::bump_global_finger(HeapWord* f) { 4045 HeapWord* read = _global_finger; 4046 HeapWord* cur = read; 4047 while (f > read) { 4048 cur = read; 4049 read = (HeapWord*) Atomic::cmpxchg_ptr(f, &_global_finger, cur); 4050 if (cur == read) { 4051 // our cas succeeded 4052 assert(_global_finger >= f, "protocol consistency"); 4053 break; 4054 } 4055 } 4056 } 4057 4058 // This is really inefficient, and should be redone by 4059 // using (not yet available) block-read and -write interfaces to the 4060 // stack and the work_queue. XXX FIX ME !!! 4061 bool CMSConcMarkingTask::get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 4062 OopTaskQueue* work_q) { 4063 // Fast lock-free check 4064 if (ovflw_stk->length() == 0) { 4065 return false; 4066 } 4067 assert(work_q->size() == 0, "Shouldn't steal"); 4068 MutexLockerEx ml(ovflw_stk->par_lock(), 4069 Mutex::_no_safepoint_check_flag); 4070 // Grab up to 1/4 the size of the work queue 4071 size_t num = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 4072 (size_t)ParGCDesiredObjsFromOverflowList); 4073 num = MIN2(num, ovflw_stk->length()); 4074 for (int i = (int) num; i > 0; i--) { 4075 oop cur = ovflw_stk->pop(); 4076 assert(cur != NULL, "Counted wrong?"); 4077 work_q->push(cur); 4078 } 4079 return num > 0; 4080 } 4081 4082 void CMSConcMarkingTask::do_scan_and_mark(int i, CompactibleFreeListSpace* sp) { 4083 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 4084 int n_tasks = pst->n_tasks(); 4085 // We allow that there may be no tasks to do here because 4086 // we are restarting after a stack overflow. 4087 assert(pst->valid() || n_tasks == 0, "Uninitialized use?"); 4088 uint nth_task = 0; 4089 4090 HeapWord* aligned_start = sp->bottom(); 4091 if (sp->used_region().contains(_restart_addr)) { 4092 // Align down to a card boundary for the start of 0th task 4093 // for this space. 4094 aligned_start = 4095 (HeapWord*)align_size_down((uintptr_t)_restart_addr, 4096 CardTableModRefBS::card_size); 4097 } 4098 4099 size_t chunk_size = sp->marking_task_size(); 4100 while (!pst->is_task_claimed(/* reference */ nth_task)) { 4101 // Having claimed the nth task in this space, 4102 // compute the chunk that it corresponds to: 4103 MemRegion span = MemRegion(aligned_start + nth_task*chunk_size, 4104 aligned_start + (nth_task+1)*chunk_size); 4105 // Try and bump the global finger via a CAS; 4106 // note that we need to do the global finger bump 4107 // _before_ taking the intersection below, because 4108 // the task corresponding to that region will be 4109 // deemed done even if the used_region() expands 4110 // because of allocation -- as it almost certainly will 4111 // during start-up while the threads yield in the 4112 // closure below. 4113 HeapWord* finger = span.end(); 4114 bump_global_finger(finger); // atomically 4115 // There are null tasks here corresponding to chunks 4116 // beyond the "top" address of the space. 4117 span = span.intersection(sp->used_region()); 4118 if (!span.is_empty()) { // Non-null task 4119 HeapWord* prev_obj; 4120 assert(!span.contains(_restart_addr) || nth_task == 0, 4121 "Inconsistency"); 4122 if (nth_task == 0) { 4123 // For the 0th task, we'll not need to compute a block_start. 4124 if (span.contains(_restart_addr)) { 4125 // In the case of a restart because of stack overflow, 4126 // we might additionally skip a chunk prefix. 4127 prev_obj = _restart_addr; 4128 } else { 4129 prev_obj = span.start(); 4130 } 4131 } else { 4132 // We want to skip the first object because 4133 // the protocol is to scan any object in its entirety 4134 // that _starts_ in this span; a fortiori, any 4135 // object starting in an earlier span is scanned 4136 // as part of an earlier claimed task. 4137 // Below we use the "careful" version of block_start 4138 // so we do not try to navigate uninitialized objects. 4139 prev_obj = sp->block_start_careful(span.start()); 4140 // Below we use a variant of block_size that uses the 4141 // Printezis bits to avoid waiting for allocated 4142 // objects to become initialized/parsable. 4143 while (prev_obj < span.start()) { 4144 size_t sz = sp->block_size_no_stall(prev_obj, _collector); 4145 if (sz > 0) { 4146 prev_obj += sz; 4147 } else { 4148 // In this case we may end up doing a bit of redundant 4149 // scanning, but that appears unavoidable, short of 4150 // locking the free list locks; see bug 6324141. 4151 break; 4152 } 4153 } 4154 } 4155 if (prev_obj < span.end()) { 4156 MemRegion my_span = MemRegion(prev_obj, span.end()); 4157 // Do the marking work within a non-empty span -- 4158 // the last argument to the constructor indicates whether the 4159 // iteration should be incremental with periodic yields. 4160 Par_MarkFromRootsClosure cl(this, _collector, my_span, 4161 &_collector->_markBitMap, 4162 work_queue(i), 4163 &_collector->_markStack, 4164 &_collector->_revisitStack, 4165 _asynch); 4166 _collector->_markBitMap.iterate(&cl, my_span.start(), my_span.end()); 4167 } // else nothing to do for this task 4168 } // else nothing to do for this task 4169 } 4170 // We'd be tempted to assert here that since there are no 4171 // more tasks left to claim in this space, the global_finger 4172 // must exceed space->top() and a fortiori space->end(). However, 4173 // that would not quite be correct because the bumping of 4174 // global_finger occurs strictly after the claiming of a task, 4175 // so by the time we reach here the global finger may not yet 4176 // have been bumped up by the thread that claimed the last 4177 // task. 4178 pst->all_tasks_completed(); 4179 } 4180 4181 class Par_ConcMarkingClosure: public Par_KlassRememberingOopClosure { 4182 private: 4183 CMSConcMarkingTask* _task; 4184 MemRegion _span; 4185 CMSBitMap* _bit_map; 4186 CMSMarkStack* _overflow_stack; 4187 OopTaskQueue* _work_queue; 4188 protected: 4189 DO_OOP_WORK_DEFN 4190 public: 4191 Par_ConcMarkingClosure(CMSCollector* collector, CMSConcMarkingTask* task, OopTaskQueue* work_queue, 4192 CMSBitMap* bit_map, CMSMarkStack* overflow_stack, 4193 CMSMarkStack* revisit_stack): 4194 Par_KlassRememberingOopClosure(collector, collector->ref_processor(), revisit_stack), 4195 _task(task), 4196 _span(collector->_span), 4197 _work_queue(work_queue), 4198 _bit_map(bit_map), 4199 _overflow_stack(overflow_stack) 4200 { } 4201 virtual void do_oop(oop* p); 4202 virtual void do_oop(narrowOop* p); 4203 void trim_queue(size_t max); 4204 void handle_stack_overflow(HeapWord* lost); 4205 void do_yield_check() { 4206 if (_task->should_yield()) { 4207 _task->yield(); 4208 } 4209 } 4210 }; 4211 4212 // Grey object scanning during work stealing phase -- 4213 // the salient assumption here is that any references 4214 // that are in these stolen objects being scanned must 4215 // already have been initialized (else they would not have 4216 // been published), so we do not need to check for 4217 // uninitialized objects before pushing here. 4218 void Par_ConcMarkingClosure::do_oop(oop obj) { 4219 assert(obj->is_oop_or_null(true), "expected an oop or NULL"); 4220 HeapWord* addr = (HeapWord*)obj; 4221 // Check if oop points into the CMS generation 4222 // and is not marked 4223 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 4224 // a white object ... 4225 // If we manage to "claim" the object, by being the 4226 // first thread to mark it, then we push it on our 4227 // marking stack 4228 if (_bit_map->par_mark(addr)) { // ... now grey 4229 // push on work queue (grey set) 4230 bool simulate_overflow = false; 4231 NOT_PRODUCT( 4232 if (CMSMarkStackOverflowALot && 4233 _collector->simulate_overflow()) { 4234 // simulate a stack overflow 4235 simulate_overflow = true; 4236 } 4237 ) 4238 if (simulate_overflow || 4239 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) { 4240 // stack overflow 4241 if (PrintCMSStatistics != 0) { 4242 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 4243 SIZE_FORMAT, _overflow_stack->capacity()); 4244 } 4245 // We cannot assert that the overflow stack is full because 4246 // it may have been emptied since. 4247 assert(simulate_overflow || 4248 _work_queue->size() == _work_queue->max_elems(), 4249 "Else push should have succeeded"); 4250 handle_stack_overflow(addr); 4251 } 4252 } // Else, some other thread got there first 4253 do_yield_check(); 4254 } 4255 } 4256 4257 void Par_ConcMarkingClosure::do_oop(oop* p) { Par_ConcMarkingClosure::do_oop_work(p); } 4258 void Par_ConcMarkingClosure::do_oop(narrowOop* p) { Par_ConcMarkingClosure::do_oop_work(p); } 4259 4260 void Par_ConcMarkingClosure::trim_queue(size_t max) { 4261 while (_work_queue->size() > max) { 4262 oop new_oop; 4263 if (_work_queue->pop_local(new_oop)) { 4264 assert(new_oop->is_oop(), "Should be an oop"); 4265 assert(_bit_map->isMarked((HeapWord*)new_oop), "Grey object"); 4266 assert(_span.contains((HeapWord*)new_oop), "Not in span"); 4267 assert(new_oop->is_parsable(), "Should be parsable"); 4268 new_oop->oop_iterate(this); // do_oop() above 4269 do_yield_check(); 4270 } 4271 } 4272 } 4273 4274 // Upon stack overflow, we discard (part of) the stack, 4275 // remembering the least address amongst those discarded 4276 // in CMSCollector's _restart_address. 4277 void Par_ConcMarkingClosure::handle_stack_overflow(HeapWord* lost) { 4278 // We need to do this under a mutex to prevent other 4279 // workers from interfering with the work done below. 4280 MutexLockerEx ml(_overflow_stack->par_lock(), 4281 Mutex::_no_safepoint_check_flag); 4282 // Remember the least grey address discarded 4283 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 4284 _collector->lower_restart_addr(ra); 4285 _overflow_stack->reset(); // discard stack contents 4286 _overflow_stack->expand(); // expand the stack if possible 4287 } 4288 4289 4290 void CMSConcMarkingTask::do_work_steal(int i) { 4291 OopTaskQueue* work_q = work_queue(i); 4292 oop obj_to_scan; 4293 CMSBitMap* bm = &(_collector->_markBitMap); 4294 CMSMarkStack* ovflw = &(_collector->_markStack); 4295 CMSMarkStack* revisit = &(_collector->_revisitStack); 4296 int* seed = _collector->hash_seed(i); 4297 Par_ConcMarkingClosure cl(_collector, this, work_q, bm, ovflw, revisit); 4298 while (true) { 4299 cl.trim_queue(0); 4300 assert(work_q->size() == 0, "Should have been emptied above"); 4301 if (get_work_from_overflow_stack(ovflw, work_q)) { 4302 // Can't assert below because the work obtained from the 4303 // overflow stack may already have been stolen from us. 4304 // assert(work_q->size() > 0, "Work from overflow stack"); 4305 continue; 4306 } else if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 4307 assert(obj_to_scan->is_oop(), "Should be an oop"); 4308 assert(bm->isMarked((HeapWord*)obj_to_scan), "Grey object"); 4309 obj_to_scan->oop_iterate(&cl); 4310 } else if (terminator()->offer_termination(&_term_term)) { 4311 assert(work_q->size() == 0, "Impossible!"); 4312 break; 4313 } else if (yielding() || should_yield()) { 4314 yield(); 4315 } 4316 } 4317 } 4318 4319 // This is run by the CMS (coordinator) thread. 4320 void CMSConcMarkingTask::coordinator_yield() { 4321 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 4322 "CMS thread should hold CMS token"); 4323 DEBUG_ONLY(RememberKlassesChecker mux(false);) 4324 // First give up the locks, then yield, then re-lock 4325 // We should probably use a constructor/destructor idiom to 4326 // do this unlock/lock or modify the MutexUnlocker class to 4327 // serve our purpose. XXX 4328 assert_lock_strong(_bit_map_lock); 4329 _bit_map_lock->unlock(); 4330 ConcurrentMarkSweepThread::desynchronize(true); 4331 ConcurrentMarkSweepThread::acknowledge_yield_request(); 4332 _collector->stopTimer(); 4333 if (PrintCMSStatistics != 0) { 4334 _collector->incrementYields(); 4335 } 4336 _collector->icms_wait(); 4337 4338 // It is possible for whichever thread initiated the yield request 4339 // not to get a chance to wake up and take the bitmap lock between 4340 // this thread releasing it and reacquiring it. So, while the 4341 // should_yield() flag is on, let's sleep for a bit to give the 4342 // other thread a chance to wake up. The limit imposed on the number 4343 // of iterations is defensive, to avoid any unforseen circumstances 4344 // putting us into an infinite loop. Since it's always been this 4345 // (coordinator_yield()) method that was observed to cause the 4346 // problem, we are using a parameter (CMSCoordinatorYieldSleepCount) 4347 // which is by default non-zero. For the other seven methods that 4348 // also perform the yield operation, as are using a different 4349 // parameter (CMSYieldSleepCount) which is by default zero. This way we 4350 // can enable the sleeping for those methods too, if necessary. 4351 // See 6442774. 4352 // 4353 // We really need to reconsider the synchronization between the GC 4354 // thread and the yield-requesting threads in the future and we 4355 // should really use wait/notify, which is the recommended 4356 // way of doing this type of interaction. Additionally, we should 4357 // consolidate the eight methods that do the yield operation and they 4358 // are almost identical into one for better maintenability and 4359 // readability. See 6445193. 4360 // 4361 // Tony 2006.06.29 4362 for (unsigned i = 0; i < CMSCoordinatorYieldSleepCount && 4363 ConcurrentMarkSweepThread::should_yield() && 4364 !CMSCollector::foregroundGCIsActive(); ++i) { 4365 os::sleep(Thread::current(), 1, false); 4366 ConcurrentMarkSweepThread::acknowledge_yield_request(); 4367 } 4368 4369 ConcurrentMarkSweepThread::synchronize(true); 4370 _bit_map_lock->lock_without_safepoint_check(); 4371 _collector->startTimer(); 4372 } 4373 4374 bool CMSCollector::do_marking_mt(bool asynch) { 4375 assert(ConcGCThreads > 0 && conc_workers() != NULL, "precondition"); 4376 int num_workers = AdaptiveSizePolicy::calc_active_conc_workers( 4377 conc_workers()->total_workers(), 4378 conc_workers()->active_workers(), 4379 Threads::number_of_non_daemon_threads()); 4380 conc_workers()->set_active_workers(num_workers); 4381 4382 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 4383 CompactibleFreeListSpace* perm_space = _permGen->cmsSpace(); 4384 4385 CMSConcMarkingTask tsk(this, 4386 cms_space, 4387 perm_space, 4388 asynch, 4389 conc_workers(), 4390 task_queues()); 4391 4392 // Since the actual number of workers we get may be different 4393 // from the number we requested above, do we need to do anything different 4394 // below? In particular, may be we need to subclass the SequantialSubTasksDone 4395 // class?? XXX 4396 cms_space ->initialize_sequential_subtasks_for_marking(num_workers); 4397 perm_space->initialize_sequential_subtasks_for_marking(num_workers); 4398 4399 // Refs discovery is already non-atomic. 4400 assert(!ref_processor()->discovery_is_atomic(), "Should be non-atomic"); 4401 assert(ref_processor()->discovery_is_mt(), "Discovery should be MT"); 4402 DEBUG_ONLY(RememberKlassesChecker cmx(should_unload_classes());) 4403 conc_workers()->start_task(&tsk); 4404 while (tsk.yielded()) { 4405 tsk.coordinator_yield(); 4406 conc_workers()->continue_task(&tsk); 4407 } 4408 // If the task was aborted, _restart_addr will be non-NULL 4409 assert(tsk.completed() || _restart_addr != NULL, "Inconsistency"); 4410 while (_restart_addr != NULL) { 4411 // XXX For now we do not make use of ABORTED state and have not 4412 // yet implemented the right abort semantics (even in the original 4413 // single-threaded CMS case). That needs some more investigation 4414 // and is deferred for now; see CR# TBF. 07252005YSR. XXX 4415 assert(!CMSAbortSemantics || tsk.aborted(), "Inconsistency"); 4416 // If _restart_addr is non-NULL, a marking stack overflow 4417 // occurred; we need to do a fresh marking iteration from the 4418 // indicated restart address. 4419 if (_foregroundGCIsActive && asynch) { 4420 // We may be running into repeated stack overflows, having 4421 // reached the limit of the stack size, while making very 4422 // slow forward progress. It may be best to bail out and 4423 // let the foreground collector do its job. 4424 // Clear _restart_addr, so that foreground GC 4425 // works from scratch. This avoids the headache of 4426 // a "rescan" which would otherwise be needed because 4427 // of the dirty mod union table & card table. 4428 _restart_addr = NULL; 4429 return false; 4430 } 4431 // Adjust the task to restart from _restart_addr 4432 tsk.reset(_restart_addr); 4433 cms_space ->initialize_sequential_subtasks_for_marking(num_workers, 4434 _restart_addr); 4435 perm_space->initialize_sequential_subtasks_for_marking(num_workers, 4436 _restart_addr); 4437 _restart_addr = NULL; 4438 // Get the workers going again 4439 conc_workers()->start_task(&tsk); 4440 while (tsk.yielded()) { 4441 tsk.coordinator_yield(); 4442 conc_workers()->continue_task(&tsk); 4443 } 4444 } 4445 assert(tsk.completed(), "Inconsistency"); 4446 assert(tsk.result() == true, "Inconsistency"); 4447 return true; 4448 } 4449 4450 bool CMSCollector::do_marking_st(bool asynch) { 4451 ResourceMark rm; 4452 HandleMark hm; 4453 4454 // Temporarily make refs discovery single threaded (non-MT) 4455 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false); 4456 MarkFromRootsClosure markFromRootsClosure(this, _span, &_markBitMap, 4457 &_markStack, &_revisitStack, CMSYield && asynch); 4458 // the last argument to iterate indicates whether the iteration 4459 // should be incremental with periodic yields. 4460 _markBitMap.iterate(&markFromRootsClosure); 4461 // If _restart_addr is non-NULL, a marking stack overflow 4462 // occurred; we need to do a fresh iteration from the 4463 // indicated restart address. 4464 while (_restart_addr != NULL) { 4465 if (_foregroundGCIsActive && asynch) { 4466 // We may be running into repeated stack overflows, having 4467 // reached the limit of the stack size, while making very 4468 // slow forward progress. It may be best to bail out and 4469 // let the foreground collector do its job. 4470 // Clear _restart_addr, so that foreground GC 4471 // works from scratch. This avoids the headache of 4472 // a "rescan" which would otherwise be needed because 4473 // of the dirty mod union table & card table. 4474 _restart_addr = NULL; 4475 return false; // indicating failure to complete marking 4476 } 4477 // Deal with stack overflow: 4478 // we restart marking from _restart_addr 4479 HeapWord* ra = _restart_addr; 4480 markFromRootsClosure.reset(ra); 4481 _restart_addr = NULL; 4482 _markBitMap.iterate(&markFromRootsClosure, ra, _span.end()); 4483 } 4484 return true; 4485 } 4486 4487 void CMSCollector::preclean() { 4488 check_correct_thread_executing(); 4489 assert(Thread::current()->is_ConcurrentGC_thread(), "Wrong thread"); 4490 verify_work_stacks_empty(); 4491 verify_overflow_empty(); 4492 _abort_preclean = false; 4493 if (CMSPrecleaningEnabled) { 4494 if (!CMSEdenChunksRecordAlways) { 4495 _eden_chunk_index = 0; 4496 } 4497 size_t used = get_eden_used(); 4498 size_t capacity = get_eden_capacity(); 4499 // Don't start sampling unless we will get sufficiently 4500 // many samples. 4501 if (used < (capacity/(CMSScheduleRemarkSamplingRatio * 100) 4502 * CMSScheduleRemarkEdenPenetration)) { 4503 _start_sampling = true; 4504 } else { 4505 _start_sampling = false; 4506 } 4507 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 4508 CMSPhaseAccounting pa(this, "preclean", !PrintGCDetails); 4509 preclean_work(CMSPrecleanRefLists1, CMSPrecleanSurvivors1); 4510 } 4511 CMSTokenSync x(true); // is cms thread 4512 if (CMSPrecleaningEnabled) { 4513 sample_eden(); 4514 _collectorState = AbortablePreclean; 4515 } else { 4516 _collectorState = FinalMarking; 4517 } 4518 verify_work_stacks_empty(); 4519 verify_overflow_empty(); 4520 } 4521 4522 // Try and schedule the remark such that young gen 4523 // occupancy is CMSScheduleRemarkEdenPenetration %. 4524 void CMSCollector::abortable_preclean() { 4525 check_correct_thread_executing(); 4526 assert(CMSPrecleaningEnabled, "Inconsistent control state"); 4527 assert(_collectorState == AbortablePreclean, "Inconsistent control state"); 4528 4529 // If Eden's current occupancy is below this threshold, 4530 // immediately schedule the remark; else preclean 4531 // past the next scavenge in an effort to 4532 // schedule the pause as described avove. By choosing 4533 // CMSScheduleRemarkEdenSizeThreshold >= max eden size 4534 // we will never do an actual abortable preclean cycle. 4535 if (get_eden_used() > CMSScheduleRemarkEdenSizeThreshold) { 4536 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 4537 CMSPhaseAccounting pa(this, "abortable-preclean", !PrintGCDetails); 4538 // We need more smarts in the abortable preclean 4539 // loop below to deal with cases where allocation 4540 // in young gen is very very slow, and our precleaning 4541 // is running a losing race against a horde of 4542 // mutators intent on flooding us with CMS updates 4543 // (dirty cards). 4544 // One, admittedly dumb, strategy is to give up 4545 // after a certain number of abortable precleaning loops 4546 // or after a certain maximum time. We want to make 4547 // this smarter in the next iteration. 4548 // XXX FIX ME!!! YSR 4549 size_t loops = 0, workdone = 0, cumworkdone = 0, waited = 0; 4550 while (!(should_abort_preclean() || 4551 ConcurrentMarkSweepThread::should_terminate())) { 4552 workdone = preclean_work(CMSPrecleanRefLists2, CMSPrecleanSurvivors2); 4553 cumworkdone += workdone; 4554 loops++; 4555 // Voluntarily terminate abortable preclean phase if we have 4556 // been at it for too long. 4557 if ((CMSMaxAbortablePrecleanLoops != 0) && 4558 loops >= CMSMaxAbortablePrecleanLoops) { 4559 if (PrintGCDetails) { 4560 gclog_or_tty->print(" CMS: abort preclean due to loops "); 4561 } 4562 break; 4563 } 4564 if (pa.wallclock_millis() > CMSMaxAbortablePrecleanTime) { 4565 if (PrintGCDetails) { 4566 gclog_or_tty->print(" CMS: abort preclean due to time "); 4567 } 4568 break; 4569 } 4570 // If we are doing little work each iteration, we should 4571 // take a short break. 4572 if (workdone < CMSAbortablePrecleanMinWorkPerIteration) { 4573 // Sleep for some time, waiting for work to accumulate 4574 stopTimer(); 4575 cmsThread()->wait_on_cms_lock(CMSAbortablePrecleanWaitMillis); 4576 startTimer(); 4577 waited++; 4578 } 4579 } 4580 if (PrintCMSStatistics > 0) { 4581 gclog_or_tty->print(" [%d iterations, %d waits, %d cards)] ", 4582 loops, waited, cumworkdone); 4583 } 4584 } 4585 CMSTokenSync x(true); // is cms thread 4586 if (_collectorState != Idling) { 4587 assert(_collectorState == AbortablePreclean, 4588 "Spontaneous state transition?"); 4589 _collectorState = FinalMarking; 4590 } // Else, a foreground collection completed this CMS cycle. 4591 return; 4592 } 4593 4594 // Respond to an Eden sampling opportunity 4595 void CMSCollector::sample_eden() { 4596 // Make sure a young gc cannot sneak in between our 4597 // reading and recording of a sample. 4598 assert(Thread::current()->is_ConcurrentGC_thread(), 4599 "Only the cms thread may collect Eden samples"); 4600 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 4601 "Should collect samples while holding CMS token"); 4602 if (!_start_sampling) { 4603 return; 4604 } 4605 // When CMSEdenChunksRecordAlways is true, the eden chunk array 4606 // is populated by the young generation. 4607 if (_eden_chunk_array != NULL && !CMSEdenChunksRecordAlways) { 4608 if (_eden_chunk_index < _eden_chunk_capacity) { 4609 _eden_chunk_array[_eden_chunk_index] = *_top_addr; // take sample 4610 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr, 4611 "Unexpected state of Eden"); 4612 // We'd like to check that what we just sampled is an oop-start address; 4613 // however, we cannot do that here since the object may not yet have been 4614 // initialized. So we'll instead do the check when we _use_ this sample 4615 // later. 4616 if (_eden_chunk_index == 0 || 4617 (pointer_delta(_eden_chunk_array[_eden_chunk_index], 4618 _eden_chunk_array[_eden_chunk_index-1]) 4619 >= CMSSamplingGrain)) { 4620 _eden_chunk_index++; // commit sample 4621 } 4622 } 4623 } 4624 if ((_collectorState == AbortablePreclean) && !_abort_preclean) { 4625 size_t used = get_eden_used(); 4626 size_t capacity = get_eden_capacity(); 4627 assert(used <= capacity, "Unexpected state of Eden"); 4628 if (used > (capacity/100 * CMSScheduleRemarkEdenPenetration)) { 4629 _abort_preclean = true; 4630 } 4631 } 4632 } 4633 4634 4635 size_t CMSCollector::preclean_work(bool clean_refs, bool clean_survivor) { 4636 assert(_collectorState == Precleaning || 4637 _collectorState == AbortablePreclean, "incorrect state"); 4638 ResourceMark rm; 4639 HandleMark hm; 4640 4641 // Precleaning is currently not MT but the reference processor 4642 // may be set for MT. Disable it temporarily here. 4643 ReferenceProcessor* rp = ref_processor(); 4644 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(rp, false); 4645 4646 // Do one pass of scrubbing the discovered reference lists 4647 // to remove any reference objects with strongly-reachable 4648 // referents. 4649 if (clean_refs) { 4650 CMSPrecleanRefsYieldClosure yield_cl(this); 4651 assert(rp->span().equals(_span), "Spans should be equal"); 4652 CMSKeepAliveClosure keep_alive(this, _span, &_markBitMap, 4653 &_markStack, &_revisitStack, 4654 true /* preclean */); 4655 CMSDrainMarkingStackClosure complete_trace(this, 4656 _span, &_markBitMap, &_markStack, 4657 &keep_alive, true /* preclean */); 4658 4659 // We don't want this step to interfere with a young 4660 // collection because we don't want to take CPU 4661 // or memory bandwidth away from the young GC threads 4662 // (which may be as many as there are CPUs). 4663 // Note that we don't need to protect ourselves from 4664 // interference with mutators because they can't 4665 // manipulate the discovered reference lists nor affect 4666 // the computed reachability of the referents, the 4667 // only properties manipulated by the precleaning 4668 // of these reference lists. 4669 stopTimer(); 4670 CMSTokenSyncWithLocks x(true /* is cms thread */, 4671 bitMapLock()); 4672 startTimer(); 4673 sample_eden(); 4674 4675 // The following will yield to allow foreground 4676 // collection to proceed promptly. XXX YSR: 4677 // The code in this method may need further 4678 // tweaking for better performance and some restructuring 4679 // for cleaner interfaces. 4680 GCTimer *gc_timer = NULL; // Currently not tracing concurrent phases 4681 rp->preclean_discovered_references( 4682 rp->is_alive_non_header(), &keep_alive, &complete_trace, 4683 &yield_cl, should_unload_classes(), gc_timer); 4684 } 4685 4686 if (clean_survivor) { // preclean the active survivor space(s) 4687 assert(_young_gen->kind() == Generation::DefNew || 4688 _young_gen->kind() == Generation::ParNew || 4689 _young_gen->kind() == Generation::ASParNew, 4690 "incorrect type for cast"); 4691 DefNewGeneration* dng = (DefNewGeneration*)_young_gen; 4692 PushAndMarkClosure pam_cl(this, _span, ref_processor(), 4693 &_markBitMap, &_modUnionTable, 4694 &_markStack, &_revisitStack, 4695 true /* precleaning phase */); 4696 stopTimer(); 4697 CMSTokenSyncWithLocks ts(true /* is cms thread */, 4698 bitMapLock()); 4699 startTimer(); 4700 unsigned int before_count = 4701 GenCollectedHeap::heap()->total_collections(); 4702 SurvivorSpacePrecleanClosure 4703 sss_cl(this, _span, &_markBitMap, &_markStack, 4704 &pam_cl, before_count, CMSYield); 4705 DEBUG_ONLY(RememberKlassesChecker mx(should_unload_classes());) 4706 dng->from()->object_iterate_careful(&sss_cl); 4707 dng->to()->object_iterate_careful(&sss_cl); 4708 } 4709 MarkRefsIntoAndScanClosure 4710 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable, 4711 &_markStack, &_revisitStack, this, CMSYield, 4712 true /* precleaning phase */); 4713 // CAUTION: The following closure has persistent state that may need to 4714 // be reset upon a decrease in the sequence of addresses it 4715 // processes. 4716 ScanMarkedObjectsAgainCarefullyClosure 4717 smoac_cl(this, _span, 4718 &_markBitMap, &_markStack, &_revisitStack, &mrias_cl, CMSYield); 4719 4720 // Preclean dirty cards in ModUnionTable and CardTable using 4721 // appropriate convergence criterion; 4722 // repeat CMSPrecleanIter times unless we find that 4723 // we are losing. 4724 assert(CMSPrecleanIter < 10, "CMSPrecleanIter is too large"); 4725 assert(CMSPrecleanNumerator < CMSPrecleanDenominator, 4726 "Bad convergence multiplier"); 4727 assert(CMSPrecleanThreshold >= 100, 4728 "Unreasonably low CMSPrecleanThreshold"); 4729 4730 size_t numIter, cumNumCards, lastNumCards, curNumCards; 4731 for (numIter = 0, cumNumCards = lastNumCards = curNumCards = 0; 4732 numIter < CMSPrecleanIter; 4733 numIter++, lastNumCards = curNumCards, cumNumCards += curNumCards) { 4734 curNumCards = preclean_mod_union_table(_cmsGen, &smoac_cl); 4735 if (CMSPermGenPrecleaningEnabled) { 4736 curNumCards += preclean_mod_union_table(_permGen, &smoac_cl); 4737 } 4738 if (Verbose && PrintGCDetails) { 4739 gclog_or_tty->print(" (modUnionTable: %d cards)", curNumCards); 4740 } 4741 // Either there are very few dirty cards, so re-mark 4742 // pause will be small anyway, or our pre-cleaning isn't 4743 // that much faster than the rate at which cards are being 4744 // dirtied, so we might as well stop and re-mark since 4745 // precleaning won't improve our re-mark time by much. 4746 if (curNumCards <= CMSPrecleanThreshold || 4747 (numIter > 0 && 4748 (curNumCards * CMSPrecleanDenominator > 4749 lastNumCards * CMSPrecleanNumerator))) { 4750 numIter++; 4751 cumNumCards += curNumCards; 4752 break; 4753 } 4754 } 4755 curNumCards = preclean_card_table(_cmsGen, &smoac_cl); 4756 if (CMSPermGenPrecleaningEnabled) { 4757 curNumCards += preclean_card_table(_permGen, &smoac_cl); 4758 } 4759 cumNumCards += curNumCards; 4760 if (PrintGCDetails && PrintCMSStatistics != 0) { 4761 gclog_or_tty->print_cr(" (cardTable: %d cards, re-scanned %d cards, %d iterations)", 4762 curNumCards, cumNumCards, numIter); 4763 } 4764 return cumNumCards; // as a measure of useful work done 4765 } 4766 4767 // PRECLEANING NOTES: 4768 // Precleaning involves: 4769 // . reading the bits of the modUnionTable and clearing the set bits. 4770 // . For the cards corresponding to the set bits, we scan the 4771 // objects on those cards. This means we need the free_list_lock 4772 // so that we can safely iterate over the CMS space when scanning 4773 // for oops. 4774 // . When we scan the objects, we'll be both reading and setting 4775 // marks in the marking bit map, so we'll need the marking bit map. 4776 // . For protecting _collector_state transitions, we take the CGC_lock. 4777 // Note that any races in the reading of of card table entries by the 4778 // CMS thread on the one hand and the clearing of those entries by the 4779 // VM thread or the setting of those entries by the mutator threads on the 4780 // other are quite benign. However, for efficiency it makes sense to keep 4781 // the VM thread from racing with the CMS thread while the latter is 4782 // dirty card info to the modUnionTable. We therefore also use the 4783 // CGC_lock to protect the reading of the card table and the mod union 4784 // table by the CM thread. 4785 // . We run concurrently with mutator updates, so scanning 4786 // needs to be done carefully -- we should not try to scan 4787 // potentially uninitialized objects. 4788 // 4789 // Locking strategy: While holding the CGC_lock, we scan over and 4790 // reset a maximal dirty range of the mod union / card tables, then lock 4791 // the free_list_lock and bitmap lock to do a full marking, then 4792 // release these locks; and repeat the cycle. This allows for a 4793 // certain amount of fairness in the sharing of these locks between 4794 // the CMS collector on the one hand, and the VM thread and the 4795 // mutators on the other. 4796 4797 // NOTE: preclean_mod_union_table() and preclean_card_table() 4798 // further below are largely identical; if you need to modify 4799 // one of these methods, please check the other method too. 4800 4801 size_t CMSCollector::preclean_mod_union_table( 4802 ConcurrentMarkSweepGeneration* gen, 4803 ScanMarkedObjectsAgainCarefullyClosure* cl) { 4804 verify_work_stacks_empty(); 4805 verify_overflow_empty(); 4806 4807 // Turn off checking for this method but turn it back on 4808 // selectively. There are yield points in this method 4809 // but it is difficult to turn the checking off just around 4810 // the yield points. It is simpler to selectively turn 4811 // it on. 4812 DEBUG_ONLY(RememberKlassesChecker mux(false);) 4813 4814 // strategy: starting with the first card, accumulate contiguous 4815 // ranges of dirty cards; clear these cards, then scan the region 4816 // covered by these cards. 4817 4818 // Since all of the MUT is committed ahead, we can just use 4819 // that, in case the generations expand while we are precleaning. 4820 // It might also be fine to just use the committed part of the 4821 // generation, but we might potentially miss cards when the 4822 // generation is rapidly expanding while we are in the midst 4823 // of precleaning. 4824 HeapWord* startAddr = gen->reserved().start(); 4825 HeapWord* endAddr = gen->reserved().end(); 4826 4827 cl->setFreelistLock(gen->freelistLock()); // needed for yielding 4828 4829 size_t numDirtyCards, cumNumDirtyCards; 4830 HeapWord *nextAddr, *lastAddr; 4831 for (cumNumDirtyCards = numDirtyCards = 0, 4832 nextAddr = lastAddr = startAddr; 4833 nextAddr < endAddr; 4834 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 4835 4836 ResourceMark rm; 4837 HandleMark hm; 4838 4839 MemRegion dirtyRegion; 4840 { 4841 stopTimer(); 4842 // Potential yield point 4843 CMSTokenSync ts(true); 4844 startTimer(); 4845 sample_eden(); 4846 // Get dirty region starting at nextOffset (inclusive), 4847 // simultaneously clearing it. 4848 dirtyRegion = 4849 _modUnionTable.getAndClearMarkedRegion(nextAddr, endAddr); 4850 assert(dirtyRegion.start() >= nextAddr, 4851 "returned region inconsistent?"); 4852 } 4853 // Remember where the next search should begin. 4854 // The returned region (if non-empty) is a right open interval, 4855 // so lastOffset is obtained from the right end of that 4856 // interval. 4857 lastAddr = dirtyRegion.end(); 4858 // Should do something more transparent and less hacky XXX 4859 numDirtyCards = 4860 _modUnionTable.heapWordDiffToOffsetDiff(dirtyRegion.word_size()); 4861 4862 // We'll scan the cards in the dirty region (with periodic 4863 // yields for foreground GC as needed). 4864 if (!dirtyRegion.is_empty()) { 4865 assert(numDirtyCards > 0, "consistency check"); 4866 HeapWord* stop_point = NULL; 4867 stopTimer(); 4868 // Potential yield point 4869 CMSTokenSyncWithLocks ts(true, gen->freelistLock(), 4870 bitMapLock()); 4871 startTimer(); 4872 { 4873 verify_work_stacks_empty(); 4874 verify_overflow_empty(); 4875 sample_eden(); 4876 DEBUG_ONLY(RememberKlassesChecker mx(should_unload_classes());) 4877 stop_point = 4878 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 4879 } 4880 if (stop_point != NULL) { 4881 // The careful iteration stopped early either because it found an 4882 // uninitialized object, or because we were in the midst of an 4883 // "abortable preclean", which should now be aborted. Redirty 4884 // the bits corresponding to the partially-scanned or unscanned 4885 // cards. We'll either restart at the next block boundary or 4886 // abort the preclean. 4887 assert((CMSPermGenPrecleaningEnabled && (gen == _permGen)) || 4888 (_collectorState == AbortablePreclean && should_abort_preclean()), 4889 "Unparsable objects should only be in perm gen."); 4890 _modUnionTable.mark_range(MemRegion(stop_point, dirtyRegion.end())); 4891 if (should_abort_preclean()) { 4892 break; // out of preclean loop 4893 } else { 4894 // Compute the next address at which preclean should pick up; 4895 // might need bitMapLock in order to read P-bits. 4896 lastAddr = next_card_start_after_block(stop_point); 4897 } 4898 } 4899 } else { 4900 assert(lastAddr == endAddr, "consistency check"); 4901 assert(numDirtyCards == 0, "consistency check"); 4902 break; 4903 } 4904 } 4905 verify_work_stacks_empty(); 4906 verify_overflow_empty(); 4907 return cumNumDirtyCards; 4908 } 4909 4910 // NOTE: preclean_mod_union_table() above and preclean_card_table() 4911 // below are largely identical; if you need to modify 4912 // one of these methods, please check the other method too. 4913 4914 size_t CMSCollector::preclean_card_table(ConcurrentMarkSweepGeneration* gen, 4915 ScanMarkedObjectsAgainCarefullyClosure* cl) { 4916 // strategy: it's similar to precleamModUnionTable above, in that 4917 // we accumulate contiguous ranges of dirty cards, mark these cards 4918 // precleaned, then scan the region covered by these cards. 4919 HeapWord* endAddr = (HeapWord*)(gen->_virtual_space.high()); 4920 HeapWord* startAddr = (HeapWord*)(gen->_virtual_space.low()); 4921 4922 cl->setFreelistLock(gen->freelistLock()); // needed for yielding 4923 4924 size_t numDirtyCards, cumNumDirtyCards; 4925 HeapWord *lastAddr, *nextAddr; 4926 4927 for (cumNumDirtyCards = numDirtyCards = 0, 4928 nextAddr = lastAddr = startAddr; 4929 nextAddr < endAddr; 4930 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 4931 4932 ResourceMark rm; 4933 HandleMark hm; 4934 4935 MemRegion dirtyRegion; 4936 { 4937 // See comments in "Precleaning notes" above on why we 4938 // do this locking. XXX Could the locking overheads be 4939 // too high when dirty cards are sparse? [I don't think so.] 4940 stopTimer(); 4941 CMSTokenSync x(true); // is cms thread 4942 startTimer(); 4943 sample_eden(); 4944 // Get and clear dirty region from card table 4945 dirtyRegion = _ct->ct_bs()->dirty_card_range_after_reset( 4946 MemRegion(nextAddr, endAddr), 4947 true, 4948 CardTableModRefBS::precleaned_card_val()); 4949 4950 assert(dirtyRegion.start() >= nextAddr, 4951 "returned region inconsistent?"); 4952 } 4953 lastAddr = dirtyRegion.end(); 4954 numDirtyCards = 4955 dirtyRegion.word_size()/CardTableModRefBS::card_size_in_words; 4956 4957 if (!dirtyRegion.is_empty()) { 4958 stopTimer(); 4959 CMSTokenSyncWithLocks ts(true, gen->freelistLock(), bitMapLock()); 4960 startTimer(); 4961 sample_eden(); 4962 verify_work_stacks_empty(); 4963 verify_overflow_empty(); 4964 DEBUG_ONLY(RememberKlassesChecker mx(should_unload_classes());) 4965 HeapWord* stop_point = 4966 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 4967 if (stop_point != NULL) { 4968 // The careful iteration stopped early because it found an 4969 // uninitialized object. Redirty the bits corresponding to the 4970 // partially-scanned or unscanned cards, and start again at the 4971 // next block boundary. 4972 assert(CMSPermGenPrecleaningEnabled || 4973 (_collectorState == AbortablePreclean && should_abort_preclean()), 4974 "Unparsable objects should only be in perm gen."); 4975 _ct->ct_bs()->invalidate(MemRegion(stop_point, dirtyRegion.end())); 4976 if (should_abort_preclean()) { 4977 break; // out of preclean loop 4978 } else { 4979 // Compute the next address at which preclean should pick up. 4980 lastAddr = next_card_start_after_block(stop_point); 4981 } 4982 } 4983 } else { 4984 break; 4985 } 4986 } 4987 verify_work_stacks_empty(); 4988 verify_overflow_empty(); 4989 return cumNumDirtyCards; 4990 } 4991 4992 void CMSCollector::checkpointRootsFinal(bool asynch, 4993 bool clear_all_soft_refs, bool init_mark_was_synchronous) { 4994 assert(_collectorState == FinalMarking, "incorrect state transition?"); 4995 check_correct_thread_executing(); 4996 // world is stopped at this checkpoint 4997 assert(SafepointSynchronize::is_at_safepoint(), 4998 "world should be stopped"); 4999 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 5000 5001 verify_work_stacks_empty(); 5002 verify_overflow_empty(); 5003 5004 SpecializationStats::clear(); 5005 if (PrintGCDetails) { 5006 gclog_or_tty->print("[YG occupancy: "SIZE_FORMAT" K ("SIZE_FORMAT" K)]", 5007 _young_gen->used() / K, 5008 _young_gen->capacity() / K); 5009 } 5010 if (asynch) { 5011 if (CMSScavengeBeforeRemark) { 5012 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5013 // Temporarily set flag to false, GCH->do_collection will 5014 // expect it to be false and set to true 5015 FlagSetting fl(gch->_is_gc_active, false); 5016 NOT_PRODUCT(GCTraceTime t("Scavenge-Before-Remark", 5017 PrintGCDetails && Verbose, true, _gc_timer_cm);) 5018 int level = _cmsGen->level() - 1; 5019 if (level >= 0) { 5020 gch->do_collection(true, // full (i.e. force, see below) 5021 false, // !clear_all_soft_refs 5022 0, // size 5023 false, // is_tlab 5024 level // max_level 5025 ); 5026 } 5027 } 5028 FreelistLocker x(this); 5029 MutexLockerEx y(bitMapLock(), 5030 Mutex::_no_safepoint_check_flag); 5031 assert(!init_mark_was_synchronous, "but that's impossible!"); 5032 checkpointRootsFinalWork(asynch, clear_all_soft_refs, false); 5033 } else { 5034 // already have all the locks 5035 checkpointRootsFinalWork(asynch, clear_all_soft_refs, 5036 init_mark_was_synchronous); 5037 } 5038 verify_work_stacks_empty(); 5039 verify_overflow_empty(); 5040 SpecializationStats::print(); 5041 } 5042 5043 void CMSCollector::checkpointRootsFinalWork(bool asynch, 5044 bool clear_all_soft_refs, bool init_mark_was_synchronous) { 5045 5046 NOT_PRODUCT(GCTraceTime tr("checkpointRootsFinalWork", PrintGCDetails, false, _gc_timer_cm);) 5047 5048 assert(haveFreelistLocks(), "must have free list locks"); 5049 assert_lock_strong(bitMapLock()); 5050 5051 if (UseAdaptiveSizePolicy) { 5052 size_policy()->checkpoint_roots_final_begin(); 5053 } 5054 5055 ResourceMark rm; 5056 HandleMark hm; 5057 5058 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5059 5060 if (should_unload_classes()) { 5061 CodeCache::gc_prologue(); 5062 } 5063 assert(haveFreelistLocks(), "must have free list locks"); 5064 assert_lock_strong(bitMapLock()); 5065 5066 DEBUG_ONLY(RememberKlassesChecker fmx(should_unload_classes());) 5067 if (!init_mark_was_synchronous) { 5068 // We might assume that we need not fill TLAB's when 5069 // CMSScavengeBeforeRemark is set, because we may have just done 5070 // a scavenge which would have filled all TLAB's -- and besides 5071 // Eden would be empty. This however may not always be the case -- 5072 // for instance although we asked for a scavenge, it may not have 5073 // happened because of a JNI critical section. We probably need 5074 // a policy for deciding whether we can in that case wait until 5075 // the critical section releases and then do the remark following 5076 // the scavenge, and skip it here. In the absence of that policy, 5077 // or of an indication of whether the scavenge did indeed occur, 5078 // we cannot rely on TLAB's having been filled and must do 5079 // so here just in case a scavenge did not happen. 5080 gch->ensure_parsability(false); // fill TLAB's, but no need to retire them 5081 // Update the saved marks which may affect the root scans. 5082 gch->save_marks(); 5083 5084 if (CMSPrintEdenSurvivorChunks) { 5085 print_eden_and_survivor_chunk_arrays(); 5086 } 5087 5088 { 5089 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;) 5090 5091 // Note on the role of the mod union table: 5092 // Since the marker in "markFromRoots" marks concurrently with 5093 // mutators, it is possible for some reachable objects not to have been 5094 // scanned. For instance, an only reference to an object A was 5095 // placed in object B after the marker scanned B. Unless B is rescanned, 5096 // A would be collected. Such updates to references in marked objects 5097 // are detected via the mod union table which is the set of all cards 5098 // dirtied since the first checkpoint in this GC cycle and prior to 5099 // the most recent young generation GC, minus those cleaned up by the 5100 // concurrent precleaning. 5101 if (CMSParallelRemarkEnabled && CollectedHeap::use_parallel_gc_threads()) { 5102 GCTraceTime t("Rescan (parallel) ", PrintGCDetails, false, _gc_timer_cm); 5103 do_remark_parallel(); 5104 } else { 5105 GCTraceTime t("Rescan (non-parallel) ", PrintGCDetails, false, 5106 _gc_timer_cm); 5107 do_remark_non_parallel(); 5108 } 5109 } 5110 } else { 5111 assert(!asynch, "Can't have init_mark_was_synchronous in asynch mode"); 5112 // The initial mark was stop-world, so there's no rescanning to 5113 // do; go straight on to the next step below. 5114 } 5115 verify_work_stacks_empty(); 5116 verify_overflow_empty(); 5117 5118 { 5119 NOT_PRODUCT(GCTraceTime ts("refProcessingWork", PrintGCDetails, false, _gc_timer_cm);) 5120 refProcessingWork(asynch, clear_all_soft_refs); 5121 } 5122 verify_work_stacks_empty(); 5123 verify_overflow_empty(); 5124 5125 if (should_unload_classes()) { 5126 CodeCache::gc_epilogue(); 5127 } 5128 JvmtiExport::gc_epilogue(); 5129 5130 // If we encountered any (marking stack / work queue) overflow 5131 // events during the current CMS cycle, take appropriate 5132 // remedial measures, where possible, so as to try and avoid 5133 // recurrence of that condition. 5134 assert(_markStack.isEmpty(), "No grey objects"); 5135 size_t ser_ovflw = _ser_pmc_remark_ovflw + _ser_pmc_preclean_ovflw + 5136 _ser_kac_ovflw + _ser_kac_preclean_ovflw; 5137 if (ser_ovflw > 0) { 5138 if (PrintCMSStatistics != 0) { 5139 gclog_or_tty->print_cr("Marking stack overflow (benign) " 5140 "(pmc_pc="SIZE_FORMAT", pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT 5141 ", kac_preclean="SIZE_FORMAT")", 5142 _ser_pmc_preclean_ovflw, _ser_pmc_remark_ovflw, 5143 _ser_kac_ovflw, _ser_kac_preclean_ovflw); 5144 } 5145 _markStack.expand(); 5146 _ser_pmc_remark_ovflw = 0; 5147 _ser_pmc_preclean_ovflw = 0; 5148 _ser_kac_preclean_ovflw = 0; 5149 _ser_kac_ovflw = 0; 5150 } 5151 if (_par_pmc_remark_ovflw > 0 || _par_kac_ovflw > 0) { 5152 if (PrintCMSStatistics != 0) { 5153 gclog_or_tty->print_cr("Work queue overflow (benign) " 5154 "(pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT")", 5155 _par_pmc_remark_ovflw, _par_kac_ovflw); 5156 } 5157 _par_pmc_remark_ovflw = 0; 5158 _par_kac_ovflw = 0; 5159 } 5160 if (PrintCMSStatistics != 0) { 5161 if (_markStack._hit_limit > 0) { 5162 gclog_or_tty->print_cr(" (benign) Hit max stack size limit ("SIZE_FORMAT")", 5163 _markStack._hit_limit); 5164 } 5165 if (_markStack._failed_double > 0) { 5166 gclog_or_tty->print_cr(" (benign) Failed stack doubling ("SIZE_FORMAT")," 5167 " current capacity "SIZE_FORMAT, 5168 _markStack._failed_double, 5169 _markStack.capacity()); 5170 } 5171 } 5172 _markStack._hit_limit = 0; 5173 _markStack._failed_double = 0; 5174 5175 // Check that all the klasses have been checked 5176 assert(_revisitStack.isEmpty(), "Not all klasses revisited"); 5177 5178 if ((VerifyAfterGC || VerifyDuringGC) && 5179 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 5180 verify_after_remark(); 5181 } 5182 5183 _gc_tracer_cm->report_object_count_after_gc(&_is_alive_closure); 5184 5185 // Change under the freelistLocks. 5186 _collectorState = Sweeping; 5187 // Call isAllClear() under bitMapLock 5188 assert(_modUnionTable.isAllClear(), "Should be clear by end of the" 5189 " final marking"); 5190 if (UseAdaptiveSizePolicy) { 5191 size_policy()->checkpoint_roots_final_end(gch->gc_cause()); 5192 } 5193 } 5194 5195 void CMSParInitialMarkTask::work(uint worker_id) { 5196 elapsedTimer _timer; 5197 ResourceMark rm; 5198 HandleMark hm; 5199 5200 // ---------- scan from roots -------------- 5201 _timer.start(); 5202 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5203 Par_MarkRefsIntoClosure par_mri_cl(_collector->_span, &(_collector->_markBitMap)); 5204 5205 // ---------- young gen roots -------------- 5206 { 5207 work_on_young_gen_roots(worker_id, &par_mri_cl); 5208 _timer.stop(); 5209 if (PrintCMSStatistics != 0) { 5210 gclog_or_tty->print_cr( 5211 "Finished young gen initial mark scan work in %dth thread: %3.3f sec", 5212 worker_id, _timer.seconds()); 5213 } 5214 } 5215 5216 // ---------- remaining roots -------------- 5217 _timer.reset(); 5218 _timer.start(); 5219 gch->gen_process_strong_roots(_collector->_cmsGen->level(), 5220 false, // yg was scanned above 5221 false, // this is parallel code 5222 true, // collecting perm gen 5223 SharedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()), 5224 &par_mri_cl, 5225 true, // walk all of code cache if (so & SO_CodeCache) 5226 NULL); 5227 assert(_collector->should_unload_classes() 5228 || (_collector->CMSCollector::roots_scanning_options() & SharedHeap::SO_CodeCache), 5229 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 5230 _timer.stop(); 5231 if (PrintCMSStatistics != 0) { 5232 gclog_or_tty->print_cr( 5233 "Finished remaining root initial mark scan work in %dth thread: %3.3f sec", 5234 worker_id, _timer.seconds()); 5235 } 5236 } 5237 5238 // Parallel remark task 5239 class CMSParRemarkTask: public CMSParMarkTask { 5240 CompactibleFreeListSpace* _cms_space; 5241 CompactibleFreeListSpace* _perm_space; 5242 5243 // The per-thread work queues, available here for stealing. 5244 OopTaskQueueSet* _task_queues; 5245 ParallelTaskTerminator _term; 5246 5247 public: 5248 // A value of 0 passed to n_workers will cause the number of 5249 // workers to be taken from the active workers in the work gang. 5250 CMSParRemarkTask(CMSCollector* collector, 5251 CompactibleFreeListSpace* cms_space, 5252 CompactibleFreeListSpace* perm_space, 5253 int n_workers, FlexibleWorkGang* workers, 5254 OopTaskQueueSet* task_queues): 5255 CMSParMarkTask("Rescan roots and grey objects in parallel", 5256 collector, n_workers), 5257 _cms_space(cms_space), _perm_space(perm_space), 5258 _task_queues(task_queues), 5259 _term(n_workers, task_queues) { } 5260 5261 OopTaskQueueSet* task_queues() { return _task_queues; } 5262 5263 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 5264 5265 ParallelTaskTerminator* terminator() { return &_term; } 5266 int n_workers() { return _n_workers; } 5267 5268 void work(uint worker_id); 5269 5270 private: 5271 // ... of dirty cards in old space 5272 void do_dirty_card_rescan_tasks(CompactibleFreeListSpace* sp, int i, 5273 Par_MarkRefsIntoAndScanClosure* cl); 5274 5275 // ... work stealing for the above 5276 void do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, int* seed); 5277 }; 5278 5279 void CMSParMarkTask::work_on_young_gen_roots(uint worker_id, OopsInGenClosure* cl) { 5280 DefNewGeneration* dng = _collector->_young_gen->as_DefNewGeneration(); 5281 EdenSpace* eden_space = dng->eden(); 5282 ContiguousSpace* from_space = dng->from(); 5283 ContiguousSpace* to_space = dng->to(); 5284 5285 HeapWord** eca = _collector->_eden_chunk_array; 5286 size_t ect = _collector->_eden_chunk_index; 5287 HeapWord** sca = _collector->_survivor_chunk_array; 5288 size_t sct = _collector->_survivor_chunk_index; 5289 5290 assert(ect <= _collector->_eden_chunk_capacity, "out of bounds"); 5291 assert(sct <= _collector->_survivor_chunk_capacity, "out of bounds"); 5292 5293 do_young_space_rescan(worker_id, cl, to_space, NULL, 0); 5294 do_young_space_rescan(worker_id, cl, from_space, sca, sct); 5295 do_young_space_rescan(worker_id, cl, eden_space, eca, ect); 5296 } 5297 5298 // work_queue(i) is passed to the closure 5299 // Par_MarkRefsIntoAndScanClosure. The "i" parameter 5300 // also is passed to do_dirty_card_rescan_tasks() and to 5301 // do_work_steal() to select the i-th task_queue. 5302 5303 void CMSParRemarkTask::work(uint worker_id) { 5304 elapsedTimer _timer; 5305 ResourceMark rm; 5306 HandleMark hm; 5307 5308 // ---------- rescan from roots -------------- 5309 _timer.start(); 5310 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5311 Par_MarkRefsIntoAndScanClosure par_mrias_cl(_collector, 5312 _collector->_span, _collector->ref_processor(), 5313 &(_collector->_markBitMap), 5314 work_queue(worker_id), &(_collector->_revisitStack)); 5315 5316 // Rescan young gen roots first since these are likely 5317 // coarsely partitioned and may, on that account, constitute 5318 // the critical path; thus, it's best to start off that 5319 // work first. 5320 // ---------- young gen roots -------------- 5321 { 5322 work_on_young_gen_roots(worker_id, &par_mrias_cl); 5323 _timer.stop(); 5324 if (PrintCMSStatistics != 0) { 5325 gclog_or_tty->print_cr( 5326 "Finished young gen rescan work in %dth thread: %3.3f sec", 5327 worker_id, _timer.seconds()); 5328 } 5329 } 5330 5331 // ---------- remaining roots -------------- 5332 _timer.reset(); 5333 _timer.start(); 5334 gch->gen_process_strong_roots(_collector->_cmsGen->level(), 5335 false, // yg was scanned above 5336 false, // this is parallel code 5337 true, // collecting perm gen 5338 SharedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()), 5339 &par_mrias_cl, 5340 true, // walk all of code cache if (so & SO_CodeCache) 5341 NULL); 5342 assert(_collector->should_unload_classes() 5343 || (_collector->CMSCollector::roots_scanning_options() & SharedHeap::SO_CodeCache), 5344 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 5345 _timer.stop(); 5346 if (PrintCMSStatistics != 0) { 5347 gclog_or_tty->print_cr( 5348 "Finished remaining root rescan work in %dth thread: %3.3f sec", 5349 worker_id, _timer.seconds()); 5350 } 5351 5352 // ---------- rescan dirty cards ------------ 5353 _timer.reset(); 5354 _timer.start(); 5355 5356 // Do the rescan tasks for each of the two spaces 5357 // (cms_space and perm_space) in turn. 5358 // "worker_id" is passed to select the task_queue for "worker_id" 5359 do_dirty_card_rescan_tasks(_cms_space, worker_id, &par_mrias_cl); 5360 do_dirty_card_rescan_tasks(_perm_space, worker_id, &par_mrias_cl); 5361 _timer.stop(); 5362 if (PrintCMSStatistics != 0) { 5363 gclog_or_tty->print_cr( 5364 "Finished dirty card rescan work in %dth thread: %3.3f sec", 5365 worker_id, _timer.seconds()); 5366 } 5367 5368 // ---------- steal work from other threads ... 5369 // ---------- ... and drain overflow list. 5370 _timer.reset(); 5371 _timer.start(); 5372 do_work_steal(worker_id, &par_mrias_cl, _collector->hash_seed(worker_id)); 5373 _timer.stop(); 5374 if (PrintCMSStatistics != 0) { 5375 gclog_or_tty->print_cr( 5376 "Finished work stealing in %dth thread: %3.3f sec", 5377 worker_id, _timer.seconds()); 5378 } 5379 } 5380 5381 // Note that parameter "i" is not used. 5382 void 5383 CMSParMarkTask::do_young_space_rescan(uint worker_id, 5384 OopsInGenClosure* cl, ContiguousSpace* space, 5385 HeapWord** chunk_array, size_t chunk_top) { 5386 // Until all tasks completed: 5387 // . claim an unclaimed task 5388 // . compute region boundaries corresponding to task claimed 5389 // using chunk_array 5390 // . par_oop_iterate(cl) over that region 5391 5392 ResourceMark rm; 5393 HandleMark hm; 5394 5395 SequentialSubTasksDone* pst = space->par_seq_tasks(); 5396 assert(pst->valid(), "Uninitialized use?"); 5397 5398 uint nth_task = 0; 5399 uint n_tasks = pst->n_tasks(); 5400 5401 HeapWord *start, *end; 5402 while (!pst->is_task_claimed(/* reference */ nth_task)) { 5403 // We claimed task # nth_task; compute its boundaries. 5404 if (chunk_top == 0) { // no samples were taken 5405 assert(nth_task == 0 && n_tasks == 1, "Can have only 1 EdenSpace task"); 5406 start = space->bottom(); 5407 end = space->top(); 5408 } else if (nth_task == 0) { 5409 start = space->bottom(); 5410 end = chunk_array[nth_task]; 5411 } else if (nth_task < (uint)chunk_top) { 5412 assert(nth_task >= 1, "Control point invariant"); 5413 start = chunk_array[nth_task - 1]; 5414 end = chunk_array[nth_task]; 5415 } else { 5416 assert(nth_task == (uint)chunk_top, "Control point invariant"); 5417 start = chunk_array[chunk_top - 1]; 5418 end = space->top(); 5419 } 5420 MemRegion mr(start, end); 5421 // Verify that mr is in space 5422 assert(mr.is_empty() || space->used_region().contains(mr), 5423 "Should be in space"); 5424 // Verify that "start" is an object boundary 5425 assert(mr.is_empty() || oop(mr.start())->is_oop(), 5426 "Should be an oop"); 5427 space->par_oop_iterate(mr, cl); 5428 } 5429 pst->all_tasks_completed(); 5430 } 5431 5432 void 5433 CMSParRemarkTask::do_dirty_card_rescan_tasks( 5434 CompactibleFreeListSpace* sp, int i, 5435 Par_MarkRefsIntoAndScanClosure* cl) { 5436 // Until all tasks completed: 5437 // . claim an unclaimed task 5438 // . compute region boundaries corresponding to task claimed 5439 // . transfer dirty bits ct->mut for that region 5440 // . apply rescanclosure to dirty mut bits for that region 5441 5442 ResourceMark rm; 5443 HandleMark hm; 5444 5445 OopTaskQueue* work_q = work_queue(i); 5446 ModUnionClosure modUnionClosure(&(_collector->_modUnionTable)); 5447 // CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! 5448 // CAUTION: This closure has state that persists across calls to 5449 // the work method dirty_range_iterate_clear() in that it has 5450 // imbedded in it a (subtype of) UpwardsObjectClosure. The 5451 // use of that state in the imbedded UpwardsObjectClosure instance 5452 // assumes that the cards are always iterated (even if in parallel 5453 // by several threads) in monotonically increasing order per each 5454 // thread. This is true of the implementation below which picks 5455 // card ranges (chunks) in monotonically increasing order globally 5456 // and, a-fortiori, in monotonically increasing order per thread 5457 // (the latter order being a subsequence of the former). 5458 // If the work code below is ever reorganized into a more chaotic 5459 // work-partitioning form than the current "sequential tasks" 5460 // paradigm, the use of that persistent state will have to be 5461 // revisited and modified appropriately. See also related 5462 // bug 4756801 work on which should examine this code to make 5463 // sure that the changes there do not run counter to the 5464 // assumptions made here and necessary for correctness and 5465 // efficiency. Note also that this code might yield inefficient 5466 // behaviour in the case of very large objects that span one or 5467 // more work chunks. Such objects would potentially be scanned 5468 // several times redundantly. Work on 4756801 should try and 5469 // address that performance anomaly if at all possible. XXX 5470 MemRegion full_span = _collector->_span; 5471 CMSBitMap* bm = &(_collector->_markBitMap); // shared 5472 CMSMarkStack* rs = &(_collector->_revisitStack); // shared 5473 MarkFromDirtyCardsClosure 5474 greyRescanClosure(_collector, full_span, // entire span of interest 5475 sp, bm, work_q, rs, cl); 5476 5477 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 5478 assert(pst->valid(), "Uninitialized use?"); 5479 uint nth_task = 0; 5480 const int alignment = CardTableModRefBS::card_size * BitsPerWord; 5481 MemRegion span = sp->used_region(); 5482 HeapWord* start_addr = span.start(); 5483 HeapWord* end_addr = (HeapWord*)round_to((intptr_t)span.end(), 5484 alignment); 5485 const size_t chunk_size = sp->rescan_task_size(); // in HeapWord units 5486 assert((HeapWord*)round_to((intptr_t)start_addr, alignment) == 5487 start_addr, "Check alignment"); 5488 assert((size_t)round_to((intptr_t)chunk_size, alignment) == 5489 chunk_size, "Check alignment"); 5490 5491 while (!pst->is_task_claimed(/* reference */ nth_task)) { 5492 // Having claimed the nth_task, compute corresponding mem-region, 5493 // which is a-fortiori aligned correctly (i.e. at a MUT bopundary). 5494 // The alignment restriction ensures that we do not need any 5495 // synchronization with other gang-workers while setting or 5496 // clearing bits in thus chunk of the MUT. 5497 MemRegion this_span = MemRegion(start_addr + nth_task*chunk_size, 5498 start_addr + (nth_task+1)*chunk_size); 5499 // The last chunk's end might be way beyond end of the 5500 // used region. In that case pull back appropriately. 5501 if (this_span.end() > end_addr) { 5502 this_span.set_end(end_addr); 5503 assert(!this_span.is_empty(), "Program logic (calculation of n_tasks)"); 5504 } 5505 // Iterate over the dirty cards covering this chunk, marking them 5506 // precleaned, and setting the corresponding bits in the mod union 5507 // table. Since we have been careful to partition at Card and MUT-word 5508 // boundaries no synchronization is needed between parallel threads. 5509 _collector->_ct->ct_bs()->dirty_card_iterate(this_span, 5510 &modUnionClosure); 5511 5512 // Having transferred these marks into the modUnionTable, 5513 // rescan the marked objects on the dirty cards in the modUnionTable. 5514 // Even if this is at a synchronous collection, the initial marking 5515 // may have been done during an asynchronous collection so there 5516 // may be dirty bits in the mod-union table. 5517 _collector->_modUnionTable.dirty_range_iterate_clear( 5518 this_span, &greyRescanClosure); 5519 _collector->_modUnionTable.verifyNoOneBitsInRange( 5520 this_span.start(), 5521 this_span.end()); 5522 } 5523 pst->all_tasks_completed(); // declare that i am done 5524 } 5525 5526 // . see if we can share work_queues with ParNew? XXX 5527 void 5528 CMSParRemarkTask::do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, 5529 int* seed) { 5530 OopTaskQueue* work_q = work_queue(i); 5531 NOT_PRODUCT(int num_steals = 0;) 5532 oop obj_to_scan; 5533 CMSBitMap* bm = &(_collector->_markBitMap); 5534 5535 while (true) { 5536 // Completely finish any left over work from (an) earlier round(s) 5537 cl->trim_queue(0); 5538 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 5539 (size_t)ParGCDesiredObjsFromOverflowList); 5540 // Now check if there's any work in the overflow list 5541 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads, 5542 // only affects the number of attempts made to get work from the 5543 // overflow list and does not affect the number of workers. Just 5544 // pass ParallelGCThreads so this behavior is unchanged. 5545 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 5546 work_q, 5547 ParallelGCThreads)) { 5548 // found something in global overflow list; 5549 // not yet ready to go stealing work from others. 5550 // We'd like to assert(work_q->size() != 0, ...) 5551 // because we just took work from the overflow list, 5552 // but of course we can't since all of that could have 5553 // been already stolen from us. 5554 // "He giveth and He taketh away." 5555 continue; 5556 } 5557 // Verify that we have no work before we resort to stealing 5558 assert(work_q->size() == 0, "Have work, shouldn't steal"); 5559 // Try to steal from other queues that have work 5560 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 5561 NOT_PRODUCT(num_steals++;) 5562 assert(obj_to_scan->is_oop(), "Oops, not an oop!"); 5563 assert(bm->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 5564 // Do scanning work 5565 obj_to_scan->oop_iterate(cl); 5566 // Loop around, finish this work, and try to steal some more 5567 } else if (terminator()->offer_termination()) { 5568 break; // nirvana from the infinite cycle 5569 } 5570 } 5571 NOT_PRODUCT( 5572 if (PrintCMSStatistics != 0) { 5573 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals); 5574 } 5575 ) 5576 assert(work_q->size() == 0 && _collector->overflow_list_is_empty(), 5577 "Else our work is not yet done"); 5578 } 5579 5580 // Record object boundaries in _eden_chunk_array by sampling the eden 5581 // top in the slow-path eden object allocation code path and record 5582 // the boundaries, if CMSEdenChunksRecordAlways is true. If 5583 // CMSEdenChunksRecordAlways is false, we use the other asynchronous 5584 // sampling in sample_eden() that activates during the part of the 5585 // preclean phase. 5586 void CMSCollector::sample_eden_chunk() { 5587 if (CMSEdenChunksRecordAlways && _eden_chunk_array != NULL) { 5588 if (_eden_chunk_lock->try_lock()) { 5589 // Record a sample. This is the critical section. The contents 5590 // of the _eden_chunk_array have to be non-decreasing in the 5591 // address order. 5592 _eden_chunk_array[_eden_chunk_index] = *_top_addr; 5593 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr, 5594 "Unexpected state of Eden"); 5595 if (_eden_chunk_index == 0 || 5596 ((_eden_chunk_array[_eden_chunk_index] > _eden_chunk_array[_eden_chunk_index-1]) && 5597 (pointer_delta(_eden_chunk_array[_eden_chunk_index], 5598 _eden_chunk_array[_eden_chunk_index-1]) >= CMSSamplingGrain))) { 5599 _eden_chunk_index++; // commit sample 5600 } 5601 _eden_chunk_lock->unlock(); 5602 } 5603 } 5604 } 5605 5606 // Return a thread-local PLAB recording array, as appropriate. 5607 void* CMSCollector::get_data_recorder(int thr_num) { 5608 if (_survivor_plab_array != NULL && 5609 (CMSPLABRecordAlways || 5610 (_collectorState > Marking && _collectorState < FinalMarking))) { 5611 assert(thr_num < (int)ParallelGCThreads, "thr_num is out of bounds"); 5612 ChunkArray* ca = &_survivor_plab_array[thr_num]; 5613 ca->reset(); // clear it so that fresh data is recorded 5614 return (void*) ca; 5615 } else { 5616 return NULL; 5617 } 5618 } 5619 5620 // Reset all the thread-local PLAB recording arrays 5621 void CMSCollector::reset_survivor_plab_arrays() { 5622 for (uint i = 0; i < ParallelGCThreads; i++) { 5623 _survivor_plab_array[i].reset(); 5624 } 5625 } 5626 5627 // Merge the per-thread plab arrays into the global survivor chunk 5628 // array which will provide the partitioning of the survivor space 5629 // for CMS initial scan and rescan. 5630 void CMSCollector::merge_survivor_plab_arrays(ContiguousSpace* surv, 5631 int no_of_gc_threads) { 5632 assert(_survivor_plab_array != NULL, "Error"); 5633 assert(_survivor_chunk_array != NULL, "Error"); 5634 assert(_collectorState == FinalMarking || 5635 (CMSParallelInitialMarkEnabled && _collectorState == InitialMarking), "Error"); 5636 for (int j = 0; j < no_of_gc_threads; j++) { 5637 _cursor[j] = 0; 5638 } 5639 HeapWord* top = surv->top(); 5640 size_t i; 5641 for (i = 0; i < _survivor_chunk_capacity; i++) { // all sca entries 5642 HeapWord* min_val = top; // Higher than any PLAB address 5643 uint min_tid = 0; // position of min_val this round 5644 for (int j = 0; j < no_of_gc_threads; j++) { 5645 ChunkArray* cur_sca = &_survivor_plab_array[j]; 5646 if (_cursor[j] == cur_sca->end()) { 5647 continue; 5648 } 5649 assert(_cursor[j] < cur_sca->end(), "ctl pt invariant"); 5650 HeapWord* cur_val = cur_sca->nth(_cursor[j]); 5651 assert(surv->used_region().contains(cur_val), "Out of bounds value"); 5652 if (cur_val < min_val) { 5653 min_tid = j; 5654 min_val = cur_val; 5655 } else { 5656 assert(cur_val < top, "All recorded addresses should be less"); 5657 } 5658 } 5659 // At this point min_val and min_tid are respectively 5660 // the least address in _survivor_plab_array[j]->nth(_cursor[j]) 5661 // and the thread (j) that witnesses that address. 5662 // We record this address in the _survivor_chunk_array[i] 5663 // and increment _cursor[min_tid] prior to the next round i. 5664 if (min_val == top) { 5665 break; 5666 } 5667 _survivor_chunk_array[i] = min_val; 5668 _cursor[min_tid]++; 5669 } 5670 // We are all done; record the size of the _survivor_chunk_array 5671 _survivor_chunk_index = i; // exclusive: [0, i) 5672 if (PrintCMSStatistics > 0) { 5673 gclog_or_tty->print(" (Survivor:" SIZE_FORMAT "chunks) ", i); 5674 } 5675 // Verify that we used up all the recorded entries 5676 #ifdef ASSERT 5677 size_t total = 0; 5678 for (int j = 0; j < no_of_gc_threads; j++) { 5679 assert(_cursor[j] == _survivor_plab_array[j].end(), "Ctl pt invariant"); 5680 total += _cursor[j]; 5681 } 5682 assert(total == _survivor_chunk_index, "Ctl Pt Invariant"); 5683 // Check that the merged array is in sorted order 5684 if (total > 0) { 5685 for (size_t i = 0; i < total - 1; i++) { 5686 if (PrintCMSStatistics > 0) { 5687 gclog_or_tty->print(" (chunk" SIZE_FORMAT ":" INTPTR_FORMAT ") ", 5688 i, _survivor_chunk_array[i]); 5689 } 5690 assert(_survivor_chunk_array[i] < _survivor_chunk_array[i+1], 5691 "Not sorted"); 5692 } 5693 } 5694 #endif // ASSERT 5695 } 5696 5697 // Set up the space's par_seq_tasks structure for work claiming 5698 // for parallel initial scan and rescan of young gen. 5699 // See ParRescanTask where this is currently used. 5700 void 5701 CMSCollector:: 5702 initialize_sequential_subtasks_for_young_gen_rescan(int n_threads) { 5703 assert(n_threads > 0, "Unexpected n_threads argument"); 5704 DefNewGeneration* dng = (DefNewGeneration*)_young_gen; 5705 5706 // Eden space 5707 { 5708 SequentialSubTasksDone* pst = dng->eden()->par_seq_tasks(); 5709 assert(!pst->valid(), "Clobbering existing data?"); 5710 // Each valid entry in [0, _eden_chunk_index) represents a task. 5711 size_t n_tasks = _eden_chunk_index + 1; 5712 assert(n_tasks == 1 || _eden_chunk_array != NULL, "Error"); 5713 // Sets the condition for completion of the subtask (how many threads 5714 // need to finish in order to be done). 5715 pst->set_n_threads(n_threads); 5716 pst->set_n_tasks((int)n_tasks); 5717 } 5718 5719 // Merge the survivor plab arrays into _survivor_chunk_array 5720 if (_survivor_plab_array != NULL) { 5721 merge_survivor_plab_arrays(dng->from(), n_threads); 5722 } else { 5723 assert(_survivor_chunk_index == 0, "Error"); 5724 } 5725 5726 // To space 5727 { 5728 SequentialSubTasksDone* pst = dng->to()->par_seq_tasks(); 5729 assert(!pst->valid(), "Clobbering existing data?"); 5730 // Sets the condition for completion of the subtask (how many threads 5731 // need to finish in order to be done). 5732 pst->set_n_threads(n_threads); 5733 pst->set_n_tasks(1); 5734 assert(pst->valid(), "Error"); 5735 } 5736 5737 // From space 5738 { 5739 SequentialSubTasksDone* pst = dng->from()->par_seq_tasks(); 5740 assert(!pst->valid(), "Clobbering existing data?"); 5741 size_t n_tasks = _survivor_chunk_index + 1; 5742 assert(n_tasks == 1 || _survivor_chunk_array != NULL, "Error"); 5743 // Sets the condition for completion of the subtask (how many threads 5744 // need to finish in order to be done). 5745 pst->set_n_threads(n_threads); 5746 pst->set_n_tasks((int)n_tasks); 5747 assert(pst->valid(), "Error"); 5748 } 5749 } 5750 5751 // Parallel version of remark 5752 void CMSCollector::do_remark_parallel() { 5753 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5754 FlexibleWorkGang* workers = gch->workers(); 5755 assert(workers != NULL, "Need parallel worker threads."); 5756 // Choose to use the number of GC workers most recently set 5757 // into "active_workers". If active_workers is not set, set it 5758 // to ParallelGCThreads. 5759 int n_workers = workers->active_workers(); 5760 if (n_workers == 0) { 5761 assert(n_workers > 0, "Should have been set during scavenge"); 5762 n_workers = ParallelGCThreads; 5763 workers->set_active_workers(n_workers); 5764 } 5765 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 5766 CompactibleFreeListSpace* perm_space = _permGen->cmsSpace(); 5767 5768 CMSParRemarkTask tsk(this, 5769 cms_space, perm_space, 5770 n_workers, workers, task_queues()); 5771 5772 // Set up for parallel process_strong_roots work. 5773 gch->set_par_threads(n_workers); 5774 // We won't be iterating over the cards in the card table updating 5775 // the younger_gen cards, so we shouldn't call the following else 5776 // the verification code as well as subsequent younger_refs_iterate 5777 // code would get confused. XXX 5778 // gch->rem_set()->prepare_for_younger_refs_iterate(true); // parallel 5779 5780 // The young gen rescan work will not be done as part of 5781 // process_strong_roots (which currently doesn't knw how to 5782 // parallelize such a scan), but rather will be broken up into 5783 // a set of parallel tasks (via the sampling that the [abortable] 5784 // preclean phase did of EdenSpace, plus the [two] tasks of 5785 // scanning the [two] survivor spaces. Further fine-grain 5786 // parallelization of the scanning of the survivor spaces 5787 // themselves, and of precleaning of the younger gen itself 5788 // is deferred to the future. 5789 initialize_sequential_subtasks_for_young_gen_rescan(n_workers); 5790 5791 // The dirty card rescan work is broken up into a "sequence" 5792 // of parallel tasks (per constituent space) that are dynamically 5793 // claimed by the parallel threads. 5794 cms_space->initialize_sequential_subtasks_for_rescan(n_workers); 5795 perm_space->initialize_sequential_subtasks_for_rescan(n_workers); 5796 5797 // It turns out that even when we're using 1 thread, doing the work in a 5798 // separate thread causes wide variance in run times. We can't help this 5799 // in the multi-threaded case, but we special-case n=1 here to get 5800 // repeatable measurements of the 1-thread overhead of the parallel code. 5801 if (n_workers > 1) { 5802 // Make refs discovery MT-safe, if it isn't already: it may not 5803 // necessarily be so, since it's possible that we are doing 5804 // ST marking. 5805 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), true); 5806 GenCollectedHeap::StrongRootsScope srs(gch); 5807 workers->run_task(&tsk); 5808 } else { 5809 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false); 5810 GenCollectedHeap::StrongRootsScope srs(gch); 5811 tsk.work(0); 5812 } 5813 gch->set_par_threads(0); // 0 ==> non-parallel. 5814 // restore, single-threaded for now, any preserved marks 5815 // as a result of work_q overflow 5816 restore_preserved_marks_if_any(); 5817 } 5818 5819 // Non-parallel version of remark 5820 void CMSCollector::do_remark_non_parallel() { 5821 ResourceMark rm; 5822 HandleMark hm; 5823 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5824 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false); 5825 5826 MarkRefsIntoAndScanClosure 5827 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable, 5828 &_markStack, &_revisitStack, this, 5829 false /* should_yield */, false /* not precleaning */); 5830 MarkFromDirtyCardsClosure 5831 markFromDirtyCardsClosure(this, _span, 5832 NULL, // space is set further below 5833 &_markBitMap, &_markStack, &_revisitStack, 5834 &mrias_cl); 5835 { 5836 GCTraceTime t("grey object rescan", PrintGCDetails, false, _gc_timer_cm); 5837 // Iterate over the dirty cards, setting the corresponding bits in the 5838 // mod union table. 5839 { 5840 ModUnionClosure modUnionClosure(&_modUnionTable); 5841 _ct->ct_bs()->dirty_card_iterate( 5842 _cmsGen->used_region(), 5843 &modUnionClosure); 5844 _ct->ct_bs()->dirty_card_iterate( 5845 _permGen->used_region(), 5846 &modUnionClosure); 5847 } 5848 // Having transferred these marks into the modUnionTable, we just need 5849 // to rescan the marked objects on the dirty cards in the modUnionTable. 5850 // The initial marking may have been done during an asynchronous 5851 // collection so there may be dirty bits in the mod-union table. 5852 const int alignment = 5853 CardTableModRefBS::card_size * BitsPerWord; 5854 { 5855 // ... First handle dirty cards in CMS gen 5856 markFromDirtyCardsClosure.set_space(_cmsGen->cmsSpace()); 5857 MemRegion ur = _cmsGen->used_region(); 5858 HeapWord* lb = ur.start(); 5859 HeapWord* ub = (HeapWord*)round_to((intptr_t)ur.end(), alignment); 5860 MemRegion cms_span(lb, ub); 5861 _modUnionTable.dirty_range_iterate_clear(cms_span, 5862 &markFromDirtyCardsClosure); 5863 verify_work_stacks_empty(); 5864 if (PrintCMSStatistics != 0) { 5865 gclog_or_tty->print(" (re-scanned "SIZE_FORMAT" dirty cards in cms gen) ", 5866 markFromDirtyCardsClosure.num_dirty_cards()); 5867 } 5868 } 5869 { 5870 // .. and then repeat for dirty cards in perm gen 5871 markFromDirtyCardsClosure.set_space(_permGen->cmsSpace()); 5872 MemRegion ur = _permGen->used_region(); 5873 HeapWord* lb = ur.start(); 5874 HeapWord* ub = (HeapWord*)round_to((intptr_t)ur.end(), alignment); 5875 MemRegion perm_span(lb, ub); 5876 _modUnionTable.dirty_range_iterate_clear(perm_span, 5877 &markFromDirtyCardsClosure); 5878 verify_work_stacks_empty(); 5879 if (PrintCMSStatistics != 0) { 5880 gclog_or_tty->print(" (re-scanned "SIZE_FORMAT" dirty cards in perm gen) ", 5881 markFromDirtyCardsClosure.num_dirty_cards()); 5882 } 5883 } 5884 } 5885 if (VerifyDuringGC && 5886 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 5887 HandleMark hm; // Discard invalid handles created during verification 5888 Universe::verify(); 5889 } 5890 { 5891 GCTraceTime t("root rescan", PrintGCDetails, false, _gc_timer_cm); 5892 5893 verify_work_stacks_empty(); 5894 5895 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 5896 GenCollectedHeap::StrongRootsScope srs(gch); 5897 gch->gen_process_strong_roots(_cmsGen->level(), 5898 true, // younger gens as roots 5899 false, // use the local StrongRootsScope 5900 true, // collecting perm gen 5901 SharedHeap::ScanningOption(roots_scanning_options()), 5902 &mrias_cl, 5903 true, // walk code active on stacks 5904 NULL); 5905 assert(should_unload_classes() 5906 || (roots_scanning_options() & SharedHeap::SO_CodeCache), 5907 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 5908 } 5909 verify_work_stacks_empty(); 5910 // Restore evacuated mark words, if any, used for overflow list links 5911 if (!CMSOverflowEarlyRestoration) { 5912 restore_preserved_marks_if_any(); 5913 } 5914 verify_overflow_empty(); 5915 } 5916 5917 //////////////////////////////////////////////////////// 5918 // Parallel Reference Processing Task Proxy Class 5919 //////////////////////////////////////////////////////// 5920 class CMSRefProcTaskProxy: public AbstractGangTaskWOopQueues { 5921 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask; 5922 CMSCollector* _collector; 5923 CMSBitMap* _mark_bit_map; 5924 const MemRegion _span; 5925 ProcessTask& _task; 5926 5927 public: 5928 CMSRefProcTaskProxy(ProcessTask& task, 5929 CMSCollector* collector, 5930 const MemRegion& span, 5931 CMSBitMap* mark_bit_map, 5932 AbstractWorkGang* workers, 5933 OopTaskQueueSet* task_queues): 5934 // XXX Should superclass AGTWOQ also know about AWG since it knows 5935 // about the task_queues used by the AWG? Then it could initialize 5936 // the terminator() object. See 6984287. The set_for_termination() 5937 // below is a temporary band-aid for the regression in 6984287. 5938 AbstractGangTaskWOopQueues("Process referents by policy in parallel", 5939 task_queues), 5940 _task(task), 5941 _collector(collector), _span(span), _mark_bit_map(mark_bit_map) 5942 { 5943 assert(_collector->_span.equals(_span) && !_span.is_empty(), 5944 "Inconsistency in _span"); 5945 set_for_termination(workers->active_workers()); 5946 } 5947 5948 OopTaskQueueSet* task_queues() { return queues(); } 5949 5950 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 5951 5952 void do_work_steal(int i, 5953 CMSParDrainMarkingStackClosure* drain, 5954 CMSParKeepAliveClosure* keep_alive, 5955 int* seed); 5956 5957 virtual void work(uint worker_id); 5958 }; 5959 5960 void CMSRefProcTaskProxy::work(uint worker_id) { 5961 assert(_collector->_span.equals(_span), "Inconsistency in _span"); 5962 CMSParKeepAliveClosure par_keep_alive(_collector, _span, 5963 _mark_bit_map, 5964 &_collector->_revisitStack, 5965 work_queue(worker_id)); 5966 CMSParDrainMarkingStackClosure par_drain_stack(_collector, _span, 5967 _mark_bit_map, 5968 &_collector->_revisitStack, 5969 work_queue(worker_id)); 5970 CMSIsAliveClosure is_alive_closure(_span, _mark_bit_map); 5971 _task.work(worker_id, is_alive_closure, par_keep_alive, par_drain_stack); 5972 if (_task.marks_oops_alive()) { 5973 do_work_steal(worker_id, &par_drain_stack, &par_keep_alive, 5974 _collector->hash_seed(worker_id)); 5975 } 5976 assert(work_queue(worker_id)->size() == 0, "work_queue should be empty"); 5977 assert(_collector->_overflow_list == NULL, "non-empty _overflow_list"); 5978 } 5979 5980 class CMSRefEnqueueTaskProxy: public AbstractGangTask { 5981 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask; 5982 EnqueueTask& _task; 5983 5984 public: 5985 CMSRefEnqueueTaskProxy(EnqueueTask& task) 5986 : AbstractGangTask("Enqueue reference objects in parallel"), 5987 _task(task) 5988 { } 5989 5990 virtual void work(uint worker_id) 5991 { 5992 _task.work(worker_id); 5993 } 5994 }; 5995 5996 CMSParKeepAliveClosure::CMSParKeepAliveClosure(CMSCollector* collector, 5997 MemRegion span, CMSBitMap* bit_map, CMSMarkStack* revisit_stack, 5998 OopTaskQueue* work_queue): 5999 Par_KlassRememberingOopClosure(collector, NULL, revisit_stack), 6000 _span(span), 6001 _bit_map(bit_map), 6002 _work_queue(work_queue), 6003 _mark_and_push(collector, span, bit_map, revisit_stack, work_queue), 6004 _low_water_mark(MIN2((uint)(work_queue->max_elems()/4), 6005 (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads))) 6006 { } 6007 6008 // . see if we can share work_queues with ParNew? XXX 6009 void CMSRefProcTaskProxy::do_work_steal(int i, 6010 CMSParDrainMarkingStackClosure* drain, 6011 CMSParKeepAliveClosure* keep_alive, 6012 int* seed) { 6013 OopTaskQueue* work_q = work_queue(i); 6014 NOT_PRODUCT(int num_steals = 0;) 6015 oop obj_to_scan; 6016 6017 while (true) { 6018 // Completely finish any left over work from (an) earlier round(s) 6019 drain->trim_queue(0); 6020 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 6021 (size_t)ParGCDesiredObjsFromOverflowList); 6022 // Now check if there's any work in the overflow list 6023 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads, 6024 // only affects the number of attempts made to get work from the 6025 // overflow list and does not affect the number of workers. Just 6026 // pass ParallelGCThreads so this behavior is unchanged. 6027 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 6028 work_q, 6029 ParallelGCThreads)) { 6030 // Found something in global overflow list; 6031 // not yet ready to go stealing work from others. 6032 // We'd like to assert(work_q->size() != 0, ...) 6033 // because we just took work from the overflow list, 6034 // but of course we can't, since all of that might have 6035 // been already stolen from us. 6036 continue; 6037 } 6038 // Verify that we have no work before we resort to stealing 6039 assert(work_q->size() == 0, "Have work, shouldn't steal"); 6040 // Try to steal from other queues that have work 6041 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 6042 NOT_PRODUCT(num_steals++;) 6043 assert(obj_to_scan->is_oop(), "Oops, not an oop!"); 6044 assert(_mark_bit_map->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 6045 // Do scanning work 6046 obj_to_scan->oop_iterate(keep_alive); 6047 // Loop around, finish this work, and try to steal some more 6048 } else if (terminator()->offer_termination()) { 6049 break; // nirvana from the infinite cycle 6050 } 6051 } 6052 NOT_PRODUCT( 6053 if (PrintCMSStatistics != 0) { 6054 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals); 6055 } 6056 ) 6057 } 6058 6059 void CMSRefProcTaskExecutor::execute(ProcessTask& task) 6060 { 6061 GenCollectedHeap* gch = GenCollectedHeap::heap(); 6062 FlexibleWorkGang* workers = gch->workers(); 6063 assert(workers != NULL, "Need parallel worker threads."); 6064 CMSRefProcTaskProxy rp_task(task, &_collector, 6065 _collector.ref_processor()->span(), 6066 _collector.markBitMap(), 6067 workers, _collector.task_queues()); 6068 workers->run_task(&rp_task); 6069 } 6070 6071 void CMSRefProcTaskExecutor::execute(EnqueueTask& task) 6072 { 6073 6074 GenCollectedHeap* gch = GenCollectedHeap::heap(); 6075 FlexibleWorkGang* workers = gch->workers(); 6076 assert(workers != NULL, "Need parallel worker threads."); 6077 CMSRefEnqueueTaskProxy enq_task(task); 6078 workers->run_task(&enq_task); 6079 } 6080 6081 void CMSCollector::refProcessingWork(bool asynch, bool clear_all_soft_refs) { 6082 6083 ResourceMark rm; 6084 HandleMark hm; 6085 6086 ReferenceProcessor* rp = ref_processor(); 6087 assert(rp->span().equals(_span), "Spans should be equal"); 6088 assert(!rp->enqueuing_is_done(), "Enqueuing should not be complete"); 6089 // Process weak references. 6090 rp->setup_policy(clear_all_soft_refs); 6091 verify_work_stacks_empty(); 6092 6093 CMSKeepAliveClosure cmsKeepAliveClosure(this, _span, &_markBitMap, 6094 &_markStack, &_revisitStack, 6095 false /* !preclean */); 6096 CMSDrainMarkingStackClosure cmsDrainMarkingStackClosure(this, 6097 _span, &_markBitMap, &_markStack, 6098 &cmsKeepAliveClosure, false /* !preclean */); 6099 { 6100 GCTraceTime t("weak refs processing", PrintGCDetails, false, _gc_timer_cm); 6101 6102 ReferenceProcessorStats stats; 6103 if (rp->processing_is_mt()) { 6104 // Set the degree of MT here. If the discovery is done MT, there 6105 // may have been a different number of threads doing the discovery 6106 // and a different number of discovered lists may have Ref objects. 6107 // That is OK as long as the Reference lists are balanced (see 6108 // balance_all_queues() and balance_queues()). 6109 GenCollectedHeap* gch = GenCollectedHeap::heap(); 6110 int active_workers = ParallelGCThreads; 6111 FlexibleWorkGang* workers = gch->workers(); 6112 if (workers != NULL) { 6113 active_workers = workers->active_workers(); 6114 // The expectation is that active_workers will have already 6115 // been set to a reasonable value. If it has not been set, 6116 // investigate. 6117 assert(active_workers > 0, "Should have been set during scavenge"); 6118 } 6119 rp->set_active_mt_degree(active_workers); 6120 CMSRefProcTaskExecutor task_executor(*this); 6121 stats = rp->process_discovered_references(&_is_alive_closure, 6122 &cmsKeepAliveClosure, 6123 &cmsDrainMarkingStackClosure, 6124 &task_executor, 6125 _gc_timer_cm); 6126 } else { 6127 stats = rp->process_discovered_references(&_is_alive_closure, 6128 &cmsKeepAliveClosure, 6129 &cmsDrainMarkingStackClosure, 6130 NULL, 6131 _gc_timer_cm); 6132 } 6133 _gc_tracer_cm->report_gc_reference_stats(stats); 6134 6135 verify_work_stacks_empty(); 6136 } 6137 6138 if (should_unload_classes()) { 6139 { 6140 GCTraceTime t("class unloading", PrintGCDetails, false, _gc_timer_cm); 6141 6142 // Follow SystemDictionary roots and unload classes 6143 bool purged_class = SystemDictionary::do_unloading(&_is_alive_closure); 6144 6145 // Follow CodeCache roots and unload any methods marked for unloading 6146 CodeCache::do_unloading(&_is_alive_closure, 6147 &cmsKeepAliveClosure, 6148 purged_class); 6149 6150 cmsDrainMarkingStackClosure.do_void(); 6151 verify_work_stacks_empty(); 6152 6153 // Update subklass/sibling/implementor links in KlassKlass descendants 6154 assert(!_revisitStack.isEmpty(), "revisit stack should not be empty"); 6155 oop k; 6156 while ((k = _revisitStack.pop()) != NULL) { 6157 ((Klass*)(oopDesc*)k)->follow_weak_klass_links( 6158 &_is_alive_closure, 6159 &cmsKeepAliveClosure); 6160 } 6161 assert(!ClassUnloading || 6162 (_markStack.isEmpty() && overflow_list_is_empty()), 6163 "Should not have found new reachable objects"); 6164 assert(_revisitStack.isEmpty(), "revisit stack should have been drained"); 6165 cmsDrainMarkingStackClosure.do_void(); 6166 verify_work_stacks_empty(); 6167 } 6168 6169 { 6170 GCTraceTime t("scrub symbol table", PrintGCDetails, false, _gc_timer_cm); 6171 // Clean up unreferenced symbols in symbol table. 6172 SymbolTable::unlink(); 6173 } 6174 } 6175 6176 if (should_unload_classes() || !JavaObjectsInPerm) { 6177 GCTraceTime t("scrub string table", PrintGCDetails, false, _gc_timer_cm); 6178 // Now clean up stale oops in StringTable 6179 StringTable::unlink(&_is_alive_closure); 6180 } 6181 6182 verify_work_stacks_empty(); 6183 // Restore any preserved marks as a result of mark stack or 6184 // work queue overflow 6185 restore_preserved_marks_if_any(); // done single-threaded for now 6186 6187 rp->set_enqueuing_is_done(true); 6188 if (rp->processing_is_mt()) { 6189 rp->balance_all_queues(); 6190 CMSRefProcTaskExecutor task_executor(*this); 6191 rp->enqueue_discovered_references(&task_executor); 6192 } else { 6193 rp->enqueue_discovered_references(NULL); 6194 } 6195 rp->verify_no_references_recorded(); 6196 assert(!rp->discovery_enabled(), "should have been disabled"); 6197 } 6198 6199 #ifndef PRODUCT 6200 void CMSCollector::check_correct_thread_executing() { 6201 Thread* t = Thread::current(); 6202 // Only the VM thread or the CMS thread should be here. 6203 assert(t->is_ConcurrentGC_thread() || t->is_VM_thread(), 6204 "Unexpected thread type"); 6205 // If this is the vm thread, the foreground process 6206 // should not be waiting. Note that _foregroundGCIsActive is 6207 // true while the foreground collector is waiting. 6208 if (_foregroundGCShouldWait) { 6209 // We cannot be the VM thread 6210 assert(t->is_ConcurrentGC_thread(), 6211 "Should be CMS thread"); 6212 } else { 6213 // We can be the CMS thread only if we are in a stop-world 6214 // phase of CMS collection. 6215 if (t->is_ConcurrentGC_thread()) { 6216 assert(_collectorState == InitialMarking || 6217 _collectorState == FinalMarking, 6218 "Should be a stop-world phase"); 6219 // The CMS thread should be holding the CMS_token. 6220 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6221 "Potential interference with concurrently " 6222 "executing VM thread"); 6223 } 6224 } 6225 } 6226 #endif 6227 6228 void CMSCollector::sweep(bool asynch) { 6229 assert(_collectorState == Sweeping, "just checking"); 6230 check_correct_thread_executing(); 6231 verify_work_stacks_empty(); 6232 verify_overflow_empty(); 6233 increment_sweep_count(); 6234 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 6235 6236 _inter_sweep_timer.stop(); 6237 _inter_sweep_estimate.sample(_inter_sweep_timer.seconds()); 6238 size_policy()->avg_cms_free_at_sweep()->sample(_cmsGen->free()); 6239 6240 // PermGen verification support: If perm gen sweeping is disabled in 6241 // this cycle, we preserve the perm gen object "deadness" information 6242 // in the perm_gen_verify_bit_map. In order to do that we traverse 6243 // all blocks in perm gen and mark all dead objects. 6244 if (verifying() && !should_unload_classes()) { 6245 assert(perm_gen_verify_bit_map()->sizeInBits() != 0, 6246 "Should have already been allocated"); 6247 MarkDeadObjectsClosure mdo(this, _permGen->cmsSpace(), 6248 markBitMap(), perm_gen_verify_bit_map()); 6249 if (asynch) { 6250 CMSTokenSyncWithLocks ts(true, _permGen->freelistLock(), 6251 bitMapLock()); 6252 _permGen->cmsSpace()->blk_iterate(&mdo); 6253 } else { 6254 // In the case of synchronous sweep, we already have 6255 // the requisite locks/tokens. 6256 _permGen->cmsSpace()->blk_iterate(&mdo); 6257 } 6258 } 6259 6260 assert(!_intra_sweep_timer.is_active(), "Should not be active"); 6261 _intra_sweep_timer.reset(); 6262 _intra_sweep_timer.start(); 6263 if (asynch) { 6264 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 6265 CMSPhaseAccounting pa(this, "sweep", !PrintGCDetails); 6266 // First sweep the old gen then the perm gen 6267 { 6268 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(), 6269 bitMapLock()); 6270 sweepWork(_cmsGen, asynch); 6271 } 6272 6273 // Now repeat for perm gen 6274 if (should_unload_classes()) { 6275 CMSTokenSyncWithLocks ts(true, _permGen->freelistLock(), 6276 bitMapLock()); 6277 sweepWork(_permGen, asynch); 6278 } 6279 6280 // Update Universe::_heap_*_at_gc figures. 6281 // We need all the free list locks to make the abstract state 6282 // transition from Sweeping to Resetting. See detailed note 6283 // further below. 6284 { 6285 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(), 6286 _permGen->freelistLock()); 6287 // Update heap occupancy information which is used as 6288 // input to soft ref clearing policy at the next gc. 6289 Universe::update_heap_info_at_gc(); 6290 _collectorState = Resizing; 6291 } 6292 } else { 6293 // already have needed locks 6294 sweepWork(_cmsGen, asynch); 6295 6296 if (should_unload_classes()) { 6297 sweepWork(_permGen, asynch); 6298 } 6299 // Update heap occupancy information which is used as 6300 // input to soft ref clearing policy at the next gc. 6301 Universe::update_heap_info_at_gc(); 6302 _collectorState = Resizing; 6303 } 6304 verify_work_stacks_empty(); 6305 verify_overflow_empty(); 6306 6307 _intra_sweep_timer.stop(); 6308 _intra_sweep_estimate.sample(_intra_sweep_timer.seconds()); 6309 6310 _inter_sweep_timer.reset(); 6311 _inter_sweep_timer.start(); 6312 6313 // We need to use a monotonically non-deccreasing time in ms 6314 // or we will see time-warp warnings and os::javaTimeMillis() 6315 // does not guarantee monotonicity. 6316 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC; 6317 update_time_of_last_gc(now); 6318 6319 // NOTE on abstract state transitions: 6320 // Mutators allocate-live and/or mark the mod-union table dirty 6321 // based on the state of the collection. The former is done in 6322 // the interval [Marking, Sweeping] and the latter in the interval 6323 // [Marking, Sweeping). Thus the transitions into the Marking state 6324 // and out of the Sweeping state must be synchronously visible 6325 // globally to the mutators. 6326 // The transition into the Marking state happens with the world 6327 // stopped so the mutators will globally see it. Sweeping is 6328 // done asynchronously by the background collector so the transition 6329 // from the Sweeping state to the Resizing state must be done 6330 // under the freelistLock (as is the check for whether to 6331 // allocate-live and whether to dirty the mod-union table). 6332 assert(_collectorState == Resizing, "Change of collector state to" 6333 " Resizing must be done under the freelistLocks (plural)"); 6334 6335 // Now that sweeping has been completed, we clear 6336 // the incremental_collection_failed flag, 6337 // thus inviting a younger gen collection to promote into 6338 // this generation. If such a promotion may still fail, 6339 // the flag will be set again when a young collection is 6340 // attempted. 6341 GenCollectedHeap* gch = GenCollectedHeap::heap(); 6342 gch->clear_incremental_collection_failed(); // Worth retrying as fresh space may have been freed up 6343 gch->update_full_collections_completed(_collection_count_start); 6344 } 6345 6346 // FIX ME!!! Looks like this belongs in CFLSpace, with 6347 // CMSGen merely delegating to it. 6348 void ConcurrentMarkSweepGeneration::setNearLargestChunk() { 6349 double nearLargestPercent = FLSLargestBlockCoalesceProximity; 6350 HeapWord* minAddr = _cmsSpace->bottom(); 6351 HeapWord* largestAddr = 6352 (HeapWord*) _cmsSpace->dictionary()->find_largest_dict(); 6353 if (largestAddr == NULL) { 6354 // The dictionary appears to be empty. In this case 6355 // try to coalesce at the end of the heap. 6356 largestAddr = _cmsSpace->end(); 6357 } 6358 size_t largestOffset = pointer_delta(largestAddr, minAddr); 6359 size_t nearLargestOffset = 6360 (size_t)((double)largestOffset * nearLargestPercent) - MinChunkSize; 6361 if (PrintFLSStatistics != 0) { 6362 gclog_or_tty->print_cr( 6363 "CMS: Large Block: " PTR_FORMAT ";" 6364 " Proximity: " PTR_FORMAT " -> " PTR_FORMAT, 6365 largestAddr, 6366 _cmsSpace->nearLargestChunk(), minAddr + nearLargestOffset); 6367 } 6368 _cmsSpace->set_nearLargestChunk(minAddr + nearLargestOffset); 6369 } 6370 6371 bool ConcurrentMarkSweepGeneration::isNearLargestChunk(HeapWord* addr) { 6372 return addr >= _cmsSpace->nearLargestChunk(); 6373 } 6374 6375 FreeChunk* ConcurrentMarkSweepGeneration::find_chunk_at_end() { 6376 return _cmsSpace->find_chunk_at_end(); 6377 } 6378 6379 void ConcurrentMarkSweepGeneration::update_gc_stats(int current_level, 6380 bool full) { 6381 // The next lower level has been collected. Gather any statistics 6382 // that are of interest at this point. 6383 if (!full && (current_level + 1) == level()) { 6384 // Gather statistics on the young generation collection. 6385 collector()->stats().record_gc0_end(used()); 6386 } 6387 } 6388 6389 CMSAdaptiveSizePolicy* ConcurrentMarkSweepGeneration::size_policy() { 6390 GenCollectedHeap* gch = GenCollectedHeap::heap(); 6391 assert(gch->kind() == CollectedHeap::GenCollectedHeap, 6392 "Wrong type of heap"); 6393 CMSAdaptiveSizePolicy* sp = (CMSAdaptiveSizePolicy*) 6394 gch->gen_policy()->size_policy(); 6395 assert(sp->is_gc_cms_adaptive_size_policy(), 6396 "Wrong type of size policy"); 6397 return sp; 6398 } 6399 6400 void ConcurrentMarkSweepGeneration::rotate_debug_collection_type() { 6401 if (PrintGCDetails && Verbose) { 6402 gclog_or_tty->print("Rotate from %d ", _debug_collection_type); 6403 } 6404 _debug_collection_type = (CollectionTypes) (_debug_collection_type + 1); 6405 _debug_collection_type = 6406 (CollectionTypes) (_debug_collection_type % Unknown_collection_type); 6407 if (PrintGCDetails && Verbose) { 6408 gclog_or_tty->print_cr("to %d ", _debug_collection_type); 6409 } 6410 } 6411 6412 void CMSCollector::sweepWork(ConcurrentMarkSweepGeneration* gen, 6413 bool asynch) { 6414 // We iterate over the space(s) underlying this generation, 6415 // checking the mark bit map to see if the bits corresponding 6416 // to specific blocks are marked or not. Blocks that are 6417 // marked are live and are not swept up. All remaining blocks 6418 // are swept up, with coalescing on-the-fly as we sweep up 6419 // contiguous free and/or garbage blocks: 6420 // We need to ensure that the sweeper synchronizes with allocators 6421 // and stop-the-world collectors. In particular, the following 6422 // locks are used: 6423 // . CMS token: if this is held, a stop the world collection cannot occur 6424 // . freelistLock: if this is held no allocation can occur from this 6425 // generation by another thread 6426 // . bitMapLock: if this is held, no other thread can access or update 6427 // 6428 6429 // Note that we need to hold the freelistLock if we use 6430 // block iterate below; else the iterator might go awry if 6431 // a mutator (or promotion) causes block contents to change 6432 // (for instance if the allocator divvies up a block). 6433 // If we hold the free list lock, for all practical purposes 6434 // young generation GC's can't occur (they'll usually need to 6435 // promote), so we might as well prevent all young generation 6436 // GC's while we do a sweeping step. For the same reason, we might 6437 // as well take the bit map lock for the entire duration 6438 6439 // check that we hold the requisite locks 6440 assert(have_cms_token(), "Should hold cms token"); 6441 assert( (asynch && ConcurrentMarkSweepThread::cms_thread_has_cms_token()) 6442 || (!asynch && ConcurrentMarkSweepThread::vm_thread_has_cms_token()), 6443 "Should possess CMS token to sweep"); 6444 assert_lock_strong(gen->freelistLock()); 6445 assert_lock_strong(bitMapLock()); 6446 6447 assert(!_inter_sweep_timer.is_active(), "Was switched off in an outer context"); 6448 assert(_intra_sweep_timer.is_active(), "Was switched on in an outer context"); 6449 gen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()), 6450 _inter_sweep_estimate.padded_average(), 6451 _intra_sweep_estimate.padded_average()); 6452 gen->setNearLargestChunk(); 6453 6454 { 6455 SweepClosure sweepClosure(this, gen, &_markBitMap, 6456 CMSYield && asynch); 6457 gen->cmsSpace()->blk_iterate_careful(&sweepClosure); 6458 // We need to free-up/coalesce garbage/blocks from a 6459 // co-terminal free run. This is done in the SweepClosure 6460 // destructor; so, do not remove this scope, else the 6461 // end-of-sweep-census below will be off by a little bit. 6462 } 6463 gen->cmsSpace()->sweep_completed(); 6464 gen->cmsSpace()->endSweepFLCensus(sweep_count()); 6465 if (should_unload_classes()) { // unloaded classes this cycle, 6466 _concurrent_cycles_since_last_unload = 0; // ... reset count 6467 } else { // did not unload classes, 6468 _concurrent_cycles_since_last_unload++; // ... increment count 6469 } 6470 } 6471 6472 // Reset CMS data structures (for now just the marking bit map) 6473 // preparatory for the next cycle. 6474 void CMSCollector::reset(bool asynch) { 6475 GenCollectedHeap* gch = GenCollectedHeap::heap(); 6476 CMSAdaptiveSizePolicy* sp = size_policy(); 6477 AdaptiveSizePolicyOutput(sp, gch->total_collections()); 6478 if (asynch) { 6479 CMSTokenSyncWithLocks ts(true, bitMapLock()); 6480 6481 // If the state is not "Resetting", the foreground thread 6482 // has done a collection and the resetting. 6483 if (_collectorState != Resetting) { 6484 assert(_collectorState == Idling, "The state should only change" 6485 " because the foreground collector has finished the collection"); 6486 return; 6487 } 6488 6489 // Clear the mark bitmap (no grey objects to start with) 6490 // for the next cycle. 6491 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 6492 CMSPhaseAccounting cmspa(this, "reset", !PrintGCDetails); 6493 6494 HeapWord* curAddr = _markBitMap.startWord(); 6495 while (curAddr < _markBitMap.endWord()) { 6496 size_t remaining = pointer_delta(_markBitMap.endWord(), curAddr); 6497 MemRegion chunk(curAddr, MIN2(CMSBitMapYieldQuantum, remaining)); 6498 _markBitMap.clear_large_range(chunk); 6499 if (ConcurrentMarkSweepThread::should_yield() && 6500 !foregroundGCIsActive() && 6501 CMSYield) { 6502 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6503 "CMS thread should hold CMS token"); 6504 assert_lock_strong(bitMapLock()); 6505 bitMapLock()->unlock(); 6506 ConcurrentMarkSweepThread::desynchronize(true); 6507 ConcurrentMarkSweepThread::acknowledge_yield_request(); 6508 stopTimer(); 6509 if (PrintCMSStatistics != 0) { 6510 incrementYields(); 6511 } 6512 icms_wait(); 6513 6514 // See the comment in coordinator_yield() 6515 for (unsigned i = 0; i < CMSYieldSleepCount && 6516 ConcurrentMarkSweepThread::should_yield() && 6517 !CMSCollector::foregroundGCIsActive(); ++i) { 6518 os::sleep(Thread::current(), 1, false); 6519 ConcurrentMarkSweepThread::acknowledge_yield_request(); 6520 } 6521 6522 ConcurrentMarkSweepThread::synchronize(true); 6523 bitMapLock()->lock_without_safepoint_check(); 6524 startTimer(); 6525 } 6526 curAddr = chunk.end(); 6527 } 6528 // A successful mostly concurrent collection has been done. 6529 // Because only the full (i.e., concurrent mode failure) collections 6530 // are being measured for gc overhead limits, clean the "near" flag 6531 // and count. 6532 sp->reset_gc_overhead_limit_count(); 6533 _collectorState = Idling; 6534 } else { 6535 // already have the lock 6536 assert(_collectorState == Resetting, "just checking"); 6537 assert_lock_strong(bitMapLock()); 6538 _markBitMap.clear_all(); 6539 _collectorState = Idling; 6540 } 6541 6542 // Stop incremental mode after a cycle completes, so that any future cycles 6543 // are triggered by allocation. 6544 stop_icms(); 6545 6546 NOT_PRODUCT( 6547 if (RotateCMSCollectionTypes) { 6548 _cmsGen->rotate_debug_collection_type(); 6549 } 6550 ) 6551 6552 register_gc_end(); 6553 } 6554 6555 void CMSCollector::do_CMS_operation(CMS_op_type op, GCCause::Cause gc_cause) { 6556 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 6557 GCTraceTime t(GCCauseString("GC", gc_cause), PrintGC, !PrintGCDetails, NULL); 6558 TraceCollectorStats tcs(counters()); 6559 6560 switch (op) { 6561 case CMS_op_checkpointRootsInitial: { 6562 SvcGCMarker sgcm(SvcGCMarker::OTHER); 6563 checkpointRootsInitial(true); // asynch 6564 if (PrintGC) { 6565 _cmsGen->printOccupancy("initial-mark"); 6566 } 6567 break; 6568 } 6569 case CMS_op_checkpointRootsFinal: { 6570 SvcGCMarker sgcm(SvcGCMarker::OTHER); 6571 checkpointRootsFinal(true, // asynch 6572 false, // !clear_all_soft_refs 6573 false); // !init_mark_was_synchronous 6574 if (PrintGC) { 6575 _cmsGen->printOccupancy("remark"); 6576 } 6577 break; 6578 } 6579 default: 6580 fatal("No such CMS_op"); 6581 } 6582 } 6583 6584 #ifndef PRODUCT 6585 size_t const CMSCollector::skip_header_HeapWords() { 6586 return FreeChunk::header_size(); 6587 } 6588 6589 // Try and collect here conditions that should hold when 6590 // CMS thread is exiting. The idea is that the foreground GC 6591 // thread should not be blocked if it wants to terminate 6592 // the CMS thread and yet continue to run the VM for a while 6593 // after that. 6594 void CMSCollector::verify_ok_to_terminate() const { 6595 assert(Thread::current()->is_ConcurrentGC_thread(), 6596 "should be called by CMS thread"); 6597 assert(!_foregroundGCShouldWait, "should be false"); 6598 // We could check here that all the various low-level locks 6599 // are not held by the CMS thread, but that is overkill; see 6600 // also CMSThread::verify_ok_to_terminate() where the CGC_lock 6601 // is checked. 6602 } 6603 #endif 6604 6605 size_t CMSCollector::block_size_using_printezis_bits(HeapWord* addr) const { 6606 assert(_markBitMap.isMarked(addr) && _markBitMap.isMarked(addr + 1), 6607 "missing Printezis mark?"); 6608 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 6609 size_t size = pointer_delta(nextOneAddr + 1, addr); 6610 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 6611 "alignment problem"); 6612 assert(size >= 3, "Necessary for Printezis marks to work"); 6613 return size; 6614 } 6615 6616 // A variant of the above (block_size_using_printezis_bits()) except 6617 // that we return 0 if the P-bits are not yet set. 6618 size_t CMSCollector::block_size_if_printezis_bits(HeapWord* addr) const { 6619 if (_markBitMap.isMarked(addr + 1)) { 6620 assert(_markBitMap.isMarked(addr), "P-bit can be set only for marked objects"); 6621 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 6622 size_t size = pointer_delta(nextOneAddr + 1, addr); 6623 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 6624 "alignment problem"); 6625 assert(size >= 3, "Necessary for Printezis marks to work"); 6626 return size; 6627 } 6628 return 0; 6629 } 6630 6631 HeapWord* CMSCollector::next_card_start_after_block(HeapWord* addr) const { 6632 size_t sz = 0; 6633 oop p = (oop)addr; 6634 if (p->klass_or_null() != NULL && p->is_parsable()) { 6635 sz = CompactibleFreeListSpace::adjustObjectSize(p->size()); 6636 } else { 6637 sz = block_size_using_printezis_bits(addr); 6638 } 6639 assert(sz > 0, "size must be nonzero"); 6640 HeapWord* next_block = addr + sz; 6641 HeapWord* next_card = (HeapWord*)round_to((uintptr_t)next_block, 6642 CardTableModRefBS::card_size); 6643 assert(round_down((uintptr_t)addr, CardTableModRefBS::card_size) < 6644 round_down((uintptr_t)next_card, CardTableModRefBS::card_size), 6645 "must be different cards"); 6646 return next_card; 6647 } 6648 6649 6650 // CMS Bit Map Wrapper ///////////////////////////////////////// 6651 6652 // Construct a CMS bit map infrastructure, but don't create the 6653 // bit vector itself. That is done by a separate call CMSBitMap::allocate() 6654 // further below. 6655 CMSBitMap::CMSBitMap(int shifter, int mutex_rank, const char* mutex_name): 6656 _bm(), 6657 _shifter(shifter), 6658 _lock(mutex_rank >= 0 ? new Mutex(mutex_rank, mutex_name, true) : NULL) 6659 { 6660 _bmStartWord = 0; 6661 _bmWordSize = 0; 6662 } 6663 6664 bool CMSBitMap::allocate(MemRegion mr) { 6665 _bmStartWord = mr.start(); 6666 _bmWordSize = mr.word_size(); 6667 ReservedSpace brs(ReservedSpace::allocation_align_size_up( 6668 (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1)); 6669 if (!brs.is_reserved()) { 6670 warning("CMS bit map allocation failure"); 6671 return false; 6672 } 6673 // For now we'll just commit all of the bit map up fromt. 6674 // Later on we'll try to be more parsimonious with swap. 6675 if (!_virtual_space.initialize(brs, brs.size())) { 6676 warning("CMS bit map backing store failure"); 6677 return false; 6678 } 6679 assert(_virtual_space.committed_size() == brs.size(), 6680 "didn't reserve backing store for all of CMS bit map?"); 6681 _bm.set_map((BitMap::bm_word_t*)_virtual_space.low()); 6682 assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >= 6683 _bmWordSize, "inconsistency in bit map sizing"); 6684 _bm.set_size(_bmWordSize >> _shifter); 6685 6686 // bm.clear(); // can we rely on getting zero'd memory? verify below 6687 assert(isAllClear(), 6688 "Expected zero'd memory from ReservedSpace constructor"); 6689 assert(_bm.size() == heapWordDiffToOffsetDiff(sizeInWords()), 6690 "consistency check"); 6691 return true; 6692 } 6693 6694 void CMSBitMap::dirty_range_iterate_clear(MemRegion mr, MemRegionClosure* cl) { 6695 HeapWord *next_addr, *end_addr, *last_addr; 6696 assert_locked(); 6697 assert(covers(mr), "out-of-range error"); 6698 // XXX assert that start and end are appropriately aligned 6699 for (next_addr = mr.start(), end_addr = mr.end(); 6700 next_addr < end_addr; next_addr = last_addr) { 6701 MemRegion dirty_region = getAndClearMarkedRegion(next_addr, end_addr); 6702 last_addr = dirty_region.end(); 6703 if (!dirty_region.is_empty()) { 6704 cl->do_MemRegion(dirty_region); 6705 } else { 6706 assert(last_addr == end_addr, "program logic"); 6707 return; 6708 } 6709 } 6710 } 6711 6712 #ifndef PRODUCT 6713 void CMSBitMap::assert_locked() const { 6714 CMSLockVerifier::assert_locked(lock()); 6715 } 6716 6717 bool CMSBitMap::covers(MemRegion mr) const { 6718 // assert(_bm.map() == _virtual_space.low(), "map inconsistency"); 6719 assert((size_t)_bm.size() == (_bmWordSize >> _shifter), 6720 "size inconsistency"); 6721 return (mr.start() >= _bmStartWord) && 6722 (mr.end() <= endWord()); 6723 } 6724 6725 bool CMSBitMap::covers(HeapWord* start, size_t size) const { 6726 return (start >= _bmStartWord && (start + size) <= endWord()); 6727 } 6728 6729 void CMSBitMap::verifyNoOneBitsInRange(HeapWord* left, HeapWord* right) { 6730 // verify that there are no 1 bits in the interval [left, right) 6731 FalseBitMapClosure falseBitMapClosure; 6732 iterate(&falseBitMapClosure, left, right); 6733 } 6734 6735 void CMSBitMap::region_invariant(MemRegion mr) 6736 { 6737 assert_locked(); 6738 // mr = mr.intersection(MemRegion(_bmStartWord, _bmWordSize)); 6739 assert(!mr.is_empty(), "unexpected empty region"); 6740 assert(covers(mr), "mr should be covered by bit map"); 6741 // convert address range into offset range 6742 size_t start_ofs = heapWordToOffset(mr.start()); 6743 // Make sure that end() is appropriately aligned 6744 assert(mr.end() == (HeapWord*)round_to((intptr_t)mr.end(), 6745 (1 << (_shifter+LogHeapWordSize))), 6746 "Misaligned mr.end()"); 6747 size_t end_ofs = heapWordToOffset(mr.end()); 6748 assert(end_ofs > start_ofs, "Should mark at least one bit"); 6749 } 6750 6751 #endif 6752 6753 bool CMSMarkStack::allocate(size_t size) { 6754 // allocate a stack of the requisite depth 6755 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 6756 size * sizeof(oop))); 6757 if (!rs.is_reserved()) { 6758 warning("CMSMarkStack allocation failure"); 6759 return false; 6760 } 6761 if (!_virtual_space.initialize(rs, rs.size())) { 6762 warning("CMSMarkStack backing store failure"); 6763 return false; 6764 } 6765 assert(_virtual_space.committed_size() == rs.size(), 6766 "didn't reserve backing store for all of CMS stack?"); 6767 _base = (oop*)(_virtual_space.low()); 6768 _index = 0; 6769 _capacity = size; 6770 NOT_PRODUCT(_max_depth = 0); 6771 return true; 6772 } 6773 6774 // XXX FIX ME !!! In the MT case we come in here holding a 6775 // leaf lock. For printing we need to take a further lock 6776 // which has lower rank. We need to recallibrate the two 6777 // lock-ranks involved in order to be able to rpint the 6778 // messages below. (Or defer the printing to the caller. 6779 // For now we take the expedient path of just disabling the 6780 // messages for the problematic case.) 6781 void CMSMarkStack::expand() { 6782 assert(_capacity <= MarkStackSizeMax, "stack bigger than permitted"); 6783 if (_capacity == MarkStackSizeMax) { 6784 if (_hit_limit++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) { 6785 // We print a warning message only once per CMS cycle. 6786 gclog_or_tty->print_cr(" (benign) Hit CMSMarkStack max size limit"); 6787 } 6788 return; 6789 } 6790 // Double capacity if possible 6791 size_t new_capacity = MIN2(_capacity*2, MarkStackSizeMax); 6792 // Do not give up existing stack until we have managed to 6793 // get the double capacity that we desired. 6794 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 6795 new_capacity * sizeof(oop))); 6796 if (rs.is_reserved()) { 6797 // Release the backing store associated with old stack 6798 _virtual_space.release(); 6799 // Reinitialize virtual space for new stack 6800 if (!_virtual_space.initialize(rs, rs.size())) { 6801 fatal("Not enough swap for expanded marking stack"); 6802 } 6803 _base = (oop*)(_virtual_space.low()); 6804 _index = 0; 6805 _capacity = new_capacity; 6806 } else if (_failed_double++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) { 6807 // Failed to double capacity, continue; 6808 // we print a detail message only once per CMS cycle. 6809 gclog_or_tty->print(" (benign) Failed to expand marking stack from "SIZE_FORMAT"K to " 6810 SIZE_FORMAT"K", 6811 _capacity / K, new_capacity / K); 6812 } 6813 } 6814 6815 6816 // Closures 6817 // XXX: there seems to be a lot of code duplication here; 6818 // should refactor and consolidate common code. 6819 6820 // This closure is used to mark refs into the CMS generation in 6821 // the CMS bit map. Called at the first checkpoint. This closure 6822 // assumes that we do not need to re-mark dirty cards; if the CMS 6823 // generation on which this is used is not an oldest (modulo perm gen) 6824 // generation then this will lose younger_gen cards! 6825 6826 MarkRefsIntoClosure::MarkRefsIntoClosure( 6827 MemRegion span, CMSBitMap* bitMap): 6828 _span(span), 6829 _bitMap(bitMap) 6830 { 6831 assert(_ref_processor == NULL, "deliberately left NULL"); 6832 assert(_bitMap->covers(_span), "_bitMap/_span mismatch"); 6833 } 6834 6835 void MarkRefsIntoClosure::do_oop(oop obj) { 6836 // if p points into _span, then mark corresponding bit in _markBitMap 6837 assert(obj->is_oop(), "expected an oop"); 6838 HeapWord* addr = (HeapWord*)obj; 6839 if (_span.contains(addr)) { 6840 // this should be made more efficient 6841 _bitMap->mark(addr); 6842 } 6843 } 6844 6845 void MarkRefsIntoClosure::do_oop(oop* p) { MarkRefsIntoClosure::do_oop_work(p); } 6846 void MarkRefsIntoClosure::do_oop(narrowOop* p) { MarkRefsIntoClosure::do_oop_work(p); } 6847 6848 Par_MarkRefsIntoClosure::Par_MarkRefsIntoClosure( 6849 MemRegion span, CMSBitMap* bitMap): 6850 _span(span), 6851 _bitMap(bitMap) 6852 { 6853 assert(_ref_processor == NULL, "deliberately left NULL"); 6854 assert(_bitMap->covers(_span), "_bitMap/_span mismatch"); 6855 } 6856 6857 void Par_MarkRefsIntoClosure::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 // this should be made more efficient 6863 _bitMap->par_mark(addr); 6864 } 6865 } 6866 6867 void Par_MarkRefsIntoClosure::do_oop(oop* p) { Par_MarkRefsIntoClosure::do_oop_work(p); } 6868 void Par_MarkRefsIntoClosure::do_oop(narrowOop* p) { Par_MarkRefsIntoClosure::do_oop_work(p); } 6869 6870 // A variant of the above, used for CMS marking verification. 6871 MarkRefsIntoVerifyClosure::MarkRefsIntoVerifyClosure( 6872 MemRegion span, CMSBitMap* verification_bm, CMSBitMap* cms_bm): 6873 _span(span), 6874 _verification_bm(verification_bm), 6875 _cms_bm(cms_bm) 6876 { 6877 assert(_ref_processor == NULL, "deliberately left NULL"); 6878 assert(_verification_bm->covers(_span), "_verification_bm/_span mismatch"); 6879 } 6880 6881 void MarkRefsIntoVerifyClosure::do_oop(oop obj) { 6882 // if p points into _span, then mark corresponding bit in _markBitMap 6883 assert(obj->is_oop(), "expected an oop"); 6884 HeapWord* addr = (HeapWord*)obj; 6885 if (_span.contains(addr)) { 6886 _verification_bm->mark(addr); 6887 if (!_cms_bm->isMarked(addr)) { 6888 oop(addr)->print(); 6889 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)", addr); 6890 fatal("... aborting"); 6891 } 6892 } 6893 } 6894 6895 void MarkRefsIntoVerifyClosure::do_oop(oop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); } 6896 void MarkRefsIntoVerifyClosure::do_oop(narrowOop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); } 6897 6898 ////////////////////////////////////////////////// 6899 // MarkRefsIntoAndScanClosure 6900 ////////////////////////////////////////////////// 6901 6902 MarkRefsIntoAndScanClosure::MarkRefsIntoAndScanClosure(MemRegion span, 6903 ReferenceProcessor* rp, 6904 CMSBitMap* bit_map, 6905 CMSBitMap* mod_union_table, 6906 CMSMarkStack* mark_stack, 6907 CMSMarkStack* revisit_stack, 6908 CMSCollector* collector, 6909 bool should_yield, 6910 bool concurrent_precleaning): 6911 _collector(collector), 6912 _span(span), 6913 _bit_map(bit_map), 6914 _mark_stack(mark_stack), 6915 _pushAndMarkClosure(collector, span, rp, bit_map, mod_union_table, 6916 mark_stack, revisit_stack, concurrent_precleaning), 6917 _yield(should_yield), 6918 _concurrent_precleaning(concurrent_precleaning), 6919 _freelistLock(NULL) 6920 { 6921 _ref_processor = rp; 6922 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 6923 } 6924 6925 // This closure is used to mark refs into the CMS generation at the 6926 // second (final) checkpoint, and to scan and transitively follow 6927 // the unmarked oops. It is also used during the concurrent precleaning 6928 // phase while scanning objects on dirty cards in the CMS generation. 6929 // The marks are made in the marking bit map and the marking stack is 6930 // used for keeping the (newly) grey objects during the scan. 6931 // The parallel version (Par_...) appears further below. 6932 void MarkRefsIntoAndScanClosure::do_oop(oop obj) { 6933 if (obj != NULL) { 6934 assert(obj->is_oop(), "expected an oop"); 6935 HeapWord* addr = (HeapWord*)obj; 6936 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 6937 assert(_collector->overflow_list_is_empty(), 6938 "overflow list should be empty"); 6939 if (_span.contains(addr) && 6940 !_bit_map->isMarked(addr)) { 6941 // mark bit map (object is now grey) 6942 _bit_map->mark(addr); 6943 // push on marking stack (stack should be empty), and drain the 6944 // stack by applying this closure to the oops in the oops popped 6945 // from the stack (i.e. blacken the grey objects) 6946 bool res = _mark_stack->push(obj); 6947 assert(res, "Should have space to push on empty stack"); 6948 do { 6949 oop new_oop = _mark_stack->pop(); 6950 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 6951 assert(new_oop->is_parsable(), "Found unparsable oop"); 6952 assert(_bit_map->isMarked((HeapWord*)new_oop), 6953 "only grey objects on this stack"); 6954 // iterate over the oops in this oop, marking and pushing 6955 // the ones in CMS heap (i.e. in _span). 6956 new_oop->oop_iterate(&_pushAndMarkClosure); 6957 // check if it's time to yield 6958 do_yield_check(); 6959 } while (!_mark_stack->isEmpty() || 6960 (!_concurrent_precleaning && take_from_overflow_list())); 6961 // if marking stack is empty, and we are not doing this 6962 // during precleaning, then check the overflow list 6963 } 6964 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 6965 assert(_collector->overflow_list_is_empty(), 6966 "overflow list was drained above"); 6967 // We could restore evacuated mark words, if any, used for 6968 // overflow list links here because the overflow list is 6969 // provably empty here. That would reduce the maximum 6970 // size requirements for preserved_{oop,mark}_stack. 6971 // But we'll just postpone it until we are all done 6972 // so we can just stream through. 6973 if (!_concurrent_precleaning && CMSOverflowEarlyRestoration) { 6974 _collector->restore_preserved_marks_if_any(); 6975 assert(_collector->no_preserved_marks(), "No preserved marks"); 6976 } 6977 assert(!CMSOverflowEarlyRestoration || _collector->no_preserved_marks(), 6978 "All preserved marks should have been restored above"); 6979 } 6980 } 6981 6982 void MarkRefsIntoAndScanClosure::do_oop(oop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); } 6983 void MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); } 6984 6985 void MarkRefsIntoAndScanClosure::do_yield_work() { 6986 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6987 "CMS thread should hold CMS token"); 6988 assert_lock_strong(_freelistLock); 6989 assert_lock_strong(_bit_map->lock()); 6990 // relinquish the free_list_lock and bitMaplock() 6991 DEBUG_ONLY(RememberKlassesChecker mux(false);) 6992 _bit_map->lock()->unlock(); 6993 _freelistLock->unlock(); 6994 ConcurrentMarkSweepThread::desynchronize(true); 6995 ConcurrentMarkSweepThread::acknowledge_yield_request(); 6996 _collector->stopTimer(); 6997 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 6998 if (PrintCMSStatistics != 0) { 6999 _collector->incrementYields(); 7000 } 7001 _collector->icms_wait(); 7002 7003 // See the comment in coordinator_yield() 7004 for (unsigned i = 0; 7005 i < CMSYieldSleepCount && 7006 ConcurrentMarkSweepThread::should_yield() && 7007 !CMSCollector::foregroundGCIsActive(); 7008 ++i) { 7009 os::sleep(Thread::current(), 1, false); 7010 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7011 } 7012 7013 ConcurrentMarkSweepThread::synchronize(true); 7014 _freelistLock->lock_without_safepoint_check(); 7015 _bit_map->lock()->lock_without_safepoint_check(); 7016 _collector->startTimer(); 7017 } 7018 7019 /////////////////////////////////////////////////////////// 7020 // Par_MarkRefsIntoAndScanClosure: a parallel version of 7021 // MarkRefsIntoAndScanClosure 7022 /////////////////////////////////////////////////////////// 7023 Par_MarkRefsIntoAndScanClosure::Par_MarkRefsIntoAndScanClosure( 7024 CMSCollector* collector, MemRegion span, ReferenceProcessor* rp, 7025 CMSBitMap* bit_map, OopTaskQueue* work_queue, CMSMarkStack* revisit_stack): 7026 _span(span), 7027 _bit_map(bit_map), 7028 _work_queue(work_queue), 7029 _low_water_mark(MIN2((uint)(work_queue->max_elems()/4), 7030 (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads))), 7031 _par_pushAndMarkClosure(collector, span, rp, bit_map, work_queue, 7032 revisit_stack) 7033 { 7034 _ref_processor = rp; 7035 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 7036 } 7037 7038 // This closure is used to mark refs into the CMS generation at the 7039 // second (final) checkpoint, and to scan and transitively follow 7040 // the unmarked oops. The marks are made in the marking bit map and 7041 // the work_queue is used for keeping the (newly) grey objects during 7042 // the scan phase whence they are also available for stealing by parallel 7043 // threads. Since the marking bit map is shared, updates are 7044 // synchronized (via CAS). 7045 void Par_MarkRefsIntoAndScanClosure::do_oop(oop obj) { 7046 if (obj != NULL) { 7047 // Ignore mark word because this could be an already marked oop 7048 // that may be chained at the end of the overflow list. 7049 assert(obj->is_oop(true), "expected an oop"); 7050 HeapWord* addr = (HeapWord*)obj; 7051 if (_span.contains(addr) && 7052 !_bit_map->isMarked(addr)) { 7053 // mark bit map (object will become grey): 7054 // It is possible for several threads to be 7055 // trying to "claim" this object concurrently; 7056 // the unique thread that succeeds in marking the 7057 // object first will do the subsequent push on 7058 // to the work queue (or overflow list). 7059 if (_bit_map->par_mark(addr)) { 7060 // push on work_queue (which may not be empty), and trim the 7061 // queue to an appropriate length by applying this closure to 7062 // the oops in the oops popped from the stack (i.e. blacken the 7063 // grey objects) 7064 bool res = _work_queue->push(obj); 7065 assert(res, "Low water mark should be less than capacity?"); 7066 trim_queue(_low_water_mark); 7067 } // Else, another thread claimed the object 7068 } 7069 } 7070 } 7071 7072 void Par_MarkRefsIntoAndScanClosure::do_oop(oop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); } 7073 void Par_MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); } 7074 7075 // This closure is used to rescan the marked objects on the dirty cards 7076 // in the mod union table and the card table proper. 7077 size_t ScanMarkedObjectsAgainCarefullyClosure::do_object_careful_m( 7078 oop p, MemRegion mr) { 7079 7080 size_t size = 0; 7081 HeapWord* addr = (HeapWord*)p; 7082 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 7083 assert(_span.contains(addr), "we are scanning the CMS generation"); 7084 // check if it's time to yield 7085 if (do_yield_check()) { 7086 // We yielded for some foreground stop-world work, 7087 // and we have been asked to abort this ongoing preclean cycle. 7088 return 0; 7089 } 7090 if (_bitMap->isMarked(addr)) { 7091 // it's marked; is it potentially uninitialized? 7092 if (p->klass_or_null() != NULL) { 7093 // If is_conc_safe is false, the object may be undergoing 7094 // change by the VM outside a safepoint. Don't try to 7095 // scan it, but rather leave it for the remark phase. 7096 if (CMSPermGenPrecleaningEnabled && 7097 (!p->is_conc_safe() || !p->is_parsable())) { 7098 // Signal precleaning to redirty the card since 7099 // the klass pointer is already installed. 7100 assert(size == 0, "Initial value"); 7101 } else { 7102 assert(p->is_parsable(), "must be parsable."); 7103 // an initialized object; ignore mark word in verification below 7104 // since we are running concurrent with mutators 7105 assert(p->is_oop(true), "should be an oop"); 7106 if (p->is_objArray()) { 7107 // objArrays are precisely marked; restrict scanning 7108 // to dirty cards only. 7109 size = CompactibleFreeListSpace::adjustObjectSize( 7110 p->oop_iterate(_scanningClosure, mr)); 7111 } else { 7112 // A non-array may have been imprecisely marked; we need 7113 // to scan object in its entirety. 7114 size = CompactibleFreeListSpace::adjustObjectSize( 7115 p->oop_iterate(_scanningClosure)); 7116 } 7117 #ifdef DEBUG 7118 size_t direct_size = 7119 CompactibleFreeListSpace::adjustObjectSize(p->size()); 7120 assert(size == direct_size, "Inconsistency in size"); 7121 assert(size >= 3, "Necessary for Printezis marks to work"); 7122 if (!_bitMap->isMarked(addr+1)) { 7123 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size); 7124 } else { 7125 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size-1); 7126 assert(_bitMap->isMarked(addr+size-1), 7127 "inconsistent Printezis mark"); 7128 } 7129 #endif // DEBUG 7130 } 7131 } else { 7132 // an unitialized object 7133 assert(_bitMap->isMarked(addr+1), "missing Printezis mark?"); 7134 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 7135 size = pointer_delta(nextOneAddr + 1, addr); 7136 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 7137 "alignment problem"); 7138 // Note that pre-cleaning needn't redirty the card. OopDesc::set_klass() 7139 // will dirty the card when the klass pointer is installed in the 7140 // object (signalling the completion of initialization). 7141 } 7142 } else { 7143 // Either a not yet marked object or an uninitialized object 7144 if (p->klass_or_null() == NULL || !p->is_parsable()) { 7145 // An uninitialized object, skip to the next card, since 7146 // we may not be able to read its P-bits yet. 7147 assert(size == 0, "Initial value"); 7148 } else { 7149 // An object not (yet) reached by marking: we merely need to 7150 // compute its size so as to go look at the next block. 7151 assert(p->is_oop(true), "should be an oop"); 7152 size = CompactibleFreeListSpace::adjustObjectSize(p->size()); 7153 } 7154 } 7155 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 7156 return size; 7157 } 7158 7159 void ScanMarkedObjectsAgainCarefullyClosure::do_yield_work() { 7160 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 7161 "CMS thread should hold CMS token"); 7162 assert_lock_strong(_freelistLock); 7163 assert_lock_strong(_bitMap->lock()); 7164 DEBUG_ONLY(RememberKlassesChecker mux(false);) 7165 // relinquish the free_list_lock and bitMaplock() 7166 _bitMap->lock()->unlock(); 7167 _freelistLock->unlock(); 7168 ConcurrentMarkSweepThread::desynchronize(true); 7169 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7170 _collector->stopTimer(); 7171 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 7172 if (PrintCMSStatistics != 0) { 7173 _collector->incrementYields(); 7174 } 7175 _collector->icms_wait(); 7176 7177 // See the comment in coordinator_yield() 7178 for (unsigned i = 0; i < CMSYieldSleepCount && 7179 ConcurrentMarkSweepThread::should_yield() && 7180 !CMSCollector::foregroundGCIsActive(); ++i) { 7181 os::sleep(Thread::current(), 1, false); 7182 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7183 } 7184 7185 ConcurrentMarkSweepThread::synchronize(true); 7186 _freelistLock->lock_without_safepoint_check(); 7187 _bitMap->lock()->lock_without_safepoint_check(); 7188 _collector->startTimer(); 7189 } 7190 7191 7192 ////////////////////////////////////////////////////////////////// 7193 // SurvivorSpacePrecleanClosure 7194 ////////////////////////////////////////////////////////////////// 7195 // This (single-threaded) closure is used to preclean the oops in 7196 // the survivor spaces. 7197 size_t SurvivorSpacePrecleanClosure::do_object_careful(oop p) { 7198 7199 HeapWord* addr = (HeapWord*)p; 7200 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 7201 assert(!_span.contains(addr), "we are scanning the survivor spaces"); 7202 assert(p->klass_or_null() != NULL, "object should be initializd"); 7203 assert(p->is_parsable(), "must be parsable."); 7204 // an initialized object; ignore mark word in verification below 7205 // since we are running concurrent with mutators 7206 assert(p->is_oop(true), "should be an oop"); 7207 // Note that we do not yield while we iterate over 7208 // the interior oops of p, pushing the relevant ones 7209 // on our marking stack. 7210 size_t size = p->oop_iterate(_scanning_closure); 7211 do_yield_check(); 7212 // Observe that below, we do not abandon the preclean 7213 // phase as soon as we should; rather we empty the 7214 // marking stack before returning. This is to satisfy 7215 // some existing assertions. In general, it may be a 7216 // good idea to abort immediately and complete the marking 7217 // from the grey objects at a later time. 7218 while (!_mark_stack->isEmpty()) { 7219 oop new_oop = _mark_stack->pop(); 7220 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 7221 assert(new_oop->is_parsable(), "Found unparsable oop"); 7222 assert(_bit_map->isMarked((HeapWord*)new_oop), 7223 "only grey objects on this stack"); 7224 // iterate over the oops in this oop, marking and pushing 7225 // the ones in CMS heap (i.e. in _span). 7226 new_oop->oop_iterate(_scanning_closure); 7227 // check if it's time to yield 7228 do_yield_check(); 7229 } 7230 unsigned int after_count = 7231 GenCollectedHeap::heap()->total_collections(); 7232 bool abort = (_before_count != after_count) || 7233 _collector->should_abort_preclean(); 7234 return abort ? 0 : size; 7235 } 7236 7237 void SurvivorSpacePrecleanClosure::do_yield_work() { 7238 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 7239 "CMS thread should hold CMS token"); 7240 assert_lock_strong(_bit_map->lock()); 7241 DEBUG_ONLY(RememberKlassesChecker smx(false);) 7242 // Relinquish the bit map lock 7243 _bit_map->lock()->unlock(); 7244 ConcurrentMarkSweepThread::desynchronize(true); 7245 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7246 _collector->stopTimer(); 7247 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 7248 if (PrintCMSStatistics != 0) { 7249 _collector->incrementYields(); 7250 } 7251 _collector->icms_wait(); 7252 7253 // See the comment in coordinator_yield() 7254 for (unsigned i = 0; i < CMSYieldSleepCount && 7255 ConcurrentMarkSweepThread::should_yield() && 7256 !CMSCollector::foregroundGCIsActive(); ++i) { 7257 os::sleep(Thread::current(), 1, false); 7258 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7259 } 7260 7261 ConcurrentMarkSweepThread::synchronize(true); 7262 _bit_map->lock()->lock_without_safepoint_check(); 7263 _collector->startTimer(); 7264 } 7265 7266 // This closure is used to rescan the marked objects on the dirty cards 7267 // in the mod union table and the card table proper. In the parallel 7268 // case, although the bitMap is shared, we do a single read so the 7269 // isMarked() query is "safe". 7270 bool ScanMarkedObjectsAgainClosure::do_object_bm(oop p, MemRegion mr) { 7271 // Ignore mark word because we are running concurrent with mutators 7272 assert(p->is_oop_or_null(true), "expected an oop or null"); 7273 HeapWord* addr = (HeapWord*)p; 7274 assert(_span.contains(addr), "we are scanning the CMS generation"); 7275 bool is_obj_array = false; 7276 #ifdef DEBUG 7277 if (!_parallel) { 7278 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 7279 assert(_collector->overflow_list_is_empty(), 7280 "overflow list should be empty"); 7281 7282 } 7283 #endif // DEBUG 7284 if (_bit_map->isMarked(addr)) { 7285 // Obj arrays are precisely marked, non-arrays are not; 7286 // so we scan objArrays precisely and non-arrays in their 7287 // entirety. 7288 if (p->is_objArray()) { 7289 is_obj_array = true; 7290 if (_parallel) { 7291 p->oop_iterate(_par_scan_closure, mr); 7292 } else { 7293 p->oop_iterate(_scan_closure, mr); 7294 } 7295 } else { 7296 if (_parallel) { 7297 p->oop_iterate(_par_scan_closure); 7298 } else { 7299 p->oop_iterate(_scan_closure); 7300 } 7301 } 7302 } 7303 #ifdef DEBUG 7304 if (!_parallel) { 7305 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 7306 assert(_collector->overflow_list_is_empty(), 7307 "overflow list should be empty"); 7308 7309 } 7310 #endif // DEBUG 7311 return is_obj_array; 7312 } 7313 7314 MarkFromRootsClosure::MarkFromRootsClosure(CMSCollector* collector, 7315 MemRegion span, 7316 CMSBitMap* bitMap, CMSMarkStack* markStack, 7317 CMSMarkStack* revisitStack, 7318 bool should_yield, bool verifying): 7319 _collector(collector), 7320 _span(span), 7321 _bitMap(bitMap), 7322 _mut(&collector->_modUnionTable), 7323 _markStack(markStack), 7324 _revisitStack(revisitStack), 7325 _yield(should_yield), 7326 _skipBits(0) 7327 { 7328 assert(_markStack->isEmpty(), "stack should be empty"); 7329 _finger = _bitMap->startWord(); 7330 _threshold = _finger; 7331 assert(_collector->_restart_addr == NULL, "Sanity check"); 7332 assert(_span.contains(_finger), "Out of bounds _finger?"); 7333 DEBUG_ONLY(_verifying = verifying;) 7334 } 7335 7336 void MarkFromRootsClosure::reset(HeapWord* addr) { 7337 assert(_markStack->isEmpty(), "would cause duplicates on stack"); 7338 assert(_span.contains(addr), "Out of bounds _finger?"); 7339 _finger = addr; 7340 _threshold = (HeapWord*)round_to( 7341 (intptr_t)_finger, CardTableModRefBS::card_size); 7342 } 7343 7344 // Should revisit to see if this should be restructured for 7345 // greater efficiency. 7346 bool MarkFromRootsClosure::do_bit(size_t offset) { 7347 if (_skipBits > 0) { 7348 _skipBits--; 7349 return true; 7350 } 7351 // convert offset into a HeapWord* 7352 HeapWord* addr = _bitMap->startWord() + offset; 7353 assert(_bitMap->endWord() && addr < _bitMap->endWord(), 7354 "address out of range"); 7355 assert(_bitMap->isMarked(addr), "tautology"); 7356 if (_bitMap->isMarked(addr+1)) { 7357 // this is an allocated but not yet initialized object 7358 assert(_skipBits == 0, "tautology"); 7359 _skipBits = 2; // skip next two marked bits ("Printezis-marks") 7360 oop p = oop(addr); 7361 if (p->klass_or_null() == NULL || !p->is_parsable()) { 7362 DEBUG_ONLY(if (!_verifying) {) 7363 // We re-dirty the cards on which this object lies and increase 7364 // the _threshold so that we'll come back to scan this object 7365 // during the preclean or remark phase. (CMSCleanOnEnter) 7366 if (CMSCleanOnEnter) { 7367 size_t sz = _collector->block_size_using_printezis_bits(addr); 7368 HeapWord* end_card_addr = (HeapWord*)round_to( 7369 (intptr_t)(addr+sz), CardTableModRefBS::card_size); 7370 MemRegion redirty_range = MemRegion(addr, end_card_addr); 7371 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 7372 // Bump _threshold to end_card_addr; note that 7373 // _threshold cannot possibly exceed end_card_addr, anyhow. 7374 // This prevents future clearing of the card as the scan proceeds 7375 // to the right. 7376 assert(_threshold <= end_card_addr, 7377 "Because we are just scanning into this object"); 7378 if (_threshold < end_card_addr) { 7379 _threshold = end_card_addr; 7380 } 7381 if (p->klass_or_null() != NULL) { 7382 // Redirty the range of cards... 7383 _mut->mark_range(redirty_range); 7384 } // ...else the setting of klass will dirty the card anyway. 7385 } 7386 DEBUG_ONLY(}) 7387 return true; 7388 } 7389 } 7390 scanOopsInOop(addr); 7391 return true; 7392 } 7393 7394 // We take a break if we've been at this for a while, 7395 // so as to avoid monopolizing the locks involved. 7396 void MarkFromRootsClosure::do_yield_work() { 7397 // First give up the locks, then yield, then re-lock 7398 // We should probably use a constructor/destructor idiom to 7399 // do this unlock/lock or modify the MutexUnlocker class to 7400 // serve our purpose. XXX 7401 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 7402 "CMS thread should hold CMS token"); 7403 assert_lock_strong(_bitMap->lock()); 7404 DEBUG_ONLY(RememberKlassesChecker mux(false);) 7405 _bitMap->lock()->unlock(); 7406 ConcurrentMarkSweepThread::desynchronize(true); 7407 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7408 _collector->stopTimer(); 7409 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 7410 if (PrintCMSStatistics != 0) { 7411 _collector->incrementYields(); 7412 } 7413 _collector->icms_wait(); 7414 7415 // See the comment in coordinator_yield() 7416 for (unsigned i = 0; i < CMSYieldSleepCount && 7417 ConcurrentMarkSweepThread::should_yield() && 7418 !CMSCollector::foregroundGCIsActive(); ++i) { 7419 os::sleep(Thread::current(), 1, false); 7420 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7421 } 7422 7423 ConcurrentMarkSweepThread::synchronize(true); 7424 _bitMap->lock()->lock_without_safepoint_check(); 7425 _collector->startTimer(); 7426 } 7427 7428 void MarkFromRootsClosure::scanOopsInOop(HeapWord* ptr) { 7429 assert(_bitMap->isMarked(ptr), "expected bit to be set"); 7430 assert(_markStack->isEmpty(), 7431 "should drain stack to limit stack usage"); 7432 // convert ptr to an oop preparatory to scanning 7433 oop obj = oop(ptr); 7434 // Ignore mark word in verification below, since we 7435 // may be running concurrent with mutators. 7436 assert(obj->is_oop(true), "should be an oop"); 7437 assert(_finger <= ptr, "_finger runneth ahead"); 7438 // advance the finger to right end of this object 7439 _finger = ptr + obj->size(); 7440 assert(_finger > ptr, "we just incremented it above"); 7441 // On large heaps, it may take us some time to get through 7442 // the marking phase (especially if running iCMS). During 7443 // this time it's possible that a lot of mutations have 7444 // accumulated in the card table and the mod union table -- 7445 // these mutation records are redundant until we have 7446 // actually traced into the corresponding card. 7447 // Here, we check whether advancing the finger would make 7448 // us cross into a new card, and if so clear corresponding 7449 // cards in the MUT (preclean them in the card-table in the 7450 // future). 7451 7452 DEBUG_ONLY(if (!_verifying) {) 7453 // The clean-on-enter optimization is disabled by default, 7454 // until we fix 6178663. 7455 if (CMSCleanOnEnter && (_finger > _threshold)) { 7456 // [_threshold, _finger) represents the interval 7457 // of cards to be cleared in MUT (or precleaned in card table). 7458 // The set of cards to be cleared is all those that overlap 7459 // with the interval [_threshold, _finger); note that 7460 // _threshold is always kept card-aligned but _finger isn't 7461 // always card-aligned. 7462 HeapWord* old_threshold = _threshold; 7463 assert(old_threshold == (HeapWord*)round_to( 7464 (intptr_t)old_threshold, CardTableModRefBS::card_size), 7465 "_threshold should always be card-aligned"); 7466 _threshold = (HeapWord*)round_to( 7467 (intptr_t)_finger, CardTableModRefBS::card_size); 7468 MemRegion mr(old_threshold, _threshold); 7469 assert(!mr.is_empty(), "Control point invariant"); 7470 assert(_span.contains(mr), "Should clear within span"); 7471 // XXX When _finger crosses from old gen into perm gen 7472 // we may be doing unnecessary cleaning; do better in the 7473 // future by detecting that condition and clearing fewer 7474 // MUT/CT entries. 7475 _mut->clear_range(mr); 7476 } 7477 DEBUG_ONLY(}) 7478 // Note: the finger doesn't advance while we drain 7479 // the stack below. 7480 PushOrMarkClosure pushOrMarkClosure(_collector, 7481 _span, _bitMap, _markStack, 7482 _revisitStack, 7483 _finger, this); 7484 bool res = _markStack->push(obj); 7485 assert(res, "Empty non-zero size stack should have space for single push"); 7486 while (!_markStack->isEmpty()) { 7487 oop new_oop = _markStack->pop(); 7488 // Skip verifying header mark word below because we are 7489 // running concurrent with mutators. 7490 assert(new_oop->is_oop(true), "Oops! expected to pop an oop"); 7491 // now scan this oop's oops 7492 new_oop->oop_iterate(&pushOrMarkClosure); 7493 do_yield_check(); 7494 } 7495 assert(_markStack->isEmpty(), "tautology, emphasizing post-condition"); 7496 } 7497 7498 Par_MarkFromRootsClosure::Par_MarkFromRootsClosure(CMSConcMarkingTask* task, 7499 CMSCollector* collector, MemRegion span, 7500 CMSBitMap* bit_map, 7501 OopTaskQueue* work_queue, 7502 CMSMarkStack* overflow_stack, 7503 CMSMarkStack* revisit_stack, 7504 bool should_yield): 7505 _collector(collector), 7506 _whole_span(collector->_span), 7507 _span(span), 7508 _bit_map(bit_map), 7509 _mut(&collector->_modUnionTable), 7510 _work_queue(work_queue), 7511 _overflow_stack(overflow_stack), 7512 _revisit_stack(revisit_stack), 7513 _yield(should_yield), 7514 _skip_bits(0), 7515 _task(task) 7516 { 7517 assert(_work_queue->size() == 0, "work_queue should be empty"); 7518 _finger = span.start(); 7519 _threshold = _finger; // XXX Defer clear-on-enter optimization for now 7520 assert(_span.contains(_finger), "Out of bounds _finger?"); 7521 } 7522 7523 // Should revisit to see if this should be restructured for 7524 // greater efficiency. 7525 bool Par_MarkFromRootsClosure::do_bit(size_t offset) { 7526 if (_skip_bits > 0) { 7527 _skip_bits--; 7528 return true; 7529 } 7530 // convert offset into a HeapWord* 7531 HeapWord* addr = _bit_map->startWord() + offset; 7532 assert(_bit_map->endWord() && addr < _bit_map->endWord(), 7533 "address out of range"); 7534 assert(_bit_map->isMarked(addr), "tautology"); 7535 if (_bit_map->isMarked(addr+1)) { 7536 // this is an allocated object that might not yet be initialized 7537 assert(_skip_bits == 0, "tautology"); 7538 _skip_bits = 2; // skip next two marked bits ("Printezis-marks") 7539 oop p = oop(addr); 7540 if (p->klass_or_null() == NULL || !p->is_parsable()) { 7541 // in the case of Clean-on-Enter optimization, redirty card 7542 // and avoid clearing card by increasing the threshold. 7543 return true; 7544 } 7545 } 7546 scan_oops_in_oop(addr); 7547 return true; 7548 } 7549 7550 void Par_MarkFromRootsClosure::scan_oops_in_oop(HeapWord* ptr) { 7551 assert(_bit_map->isMarked(ptr), "expected bit to be set"); 7552 // Should we assert that our work queue is empty or 7553 // below some drain limit? 7554 assert(_work_queue->size() == 0, 7555 "should drain stack to limit stack usage"); 7556 // convert ptr to an oop preparatory to scanning 7557 oop obj = oop(ptr); 7558 // Ignore mark word in verification below, since we 7559 // may be running concurrent with mutators. 7560 assert(obj->is_oop(true), "should be an oop"); 7561 assert(_finger <= ptr, "_finger runneth ahead"); 7562 // advance the finger to right end of this object 7563 _finger = ptr + obj->size(); 7564 assert(_finger > ptr, "we just incremented it above"); 7565 // On large heaps, it may take us some time to get through 7566 // the marking phase (especially if running iCMS). During 7567 // this time it's possible that a lot of mutations have 7568 // accumulated in the card table and the mod union table -- 7569 // these mutation records are redundant until we have 7570 // actually traced into the corresponding card. 7571 // Here, we check whether advancing the finger would make 7572 // us cross into a new card, and if so clear corresponding 7573 // cards in the MUT (preclean them in the card-table in the 7574 // future). 7575 7576 // The clean-on-enter optimization is disabled by default, 7577 // until we fix 6178663. 7578 if (CMSCleanOnEnter && (_finger > _threshold)) { 7579 // [_threshold, _finger) represents the interval 7580 // of cards to be cleared in MUT (or precleaned in card table). 7581 // The set of cards to be cleared is all those that overlap 7582 // with the interval [_threshold, _finger); note that 7583 // _threshold is always kept card-aligned but _finger isn't 7584 // always card-aligned. 7585 HeapWord* old_threshold = _threshold; 7586 assert(old_threshold == (HeapWord*)round_to( 7587 (intptr_t)old_threshold, CardTableModRefBS::card_size), 7588 "_threshold should always be card-aligned"); 7589 _threshold = (HeapWord*)round_to( 7590 (intptr_t)_finger, CardTableModRefBS::card_size); 7591 MemRegion mr(old_threshold, _threshold); 7592 assert(!mr.is_empty(), "Control point invariant"); 7593 assert(_span.contains(mr), "Should clear within span"); // _whole_span ?? 7594 // XXX When _finger crosses from old gen into perm gen 7595 // we may be doing unnecessary cleaning; do better in the 7596 // future by detecting that condition and clearing fewer 7597 // MUT/CT entries. 7598 _mut->clear_range(mr); 7599 } 7600 7601 // Note: the local finger doesn't advance while we drain 7602 // the stack below, but the global finger sure can and will. 7603 HeapWord** gfa = _task->global_finger_addr(); 7604 Par_PushOrMarkClosure pushOrMarkClosure(_collector, 7605 _span, _bit_map, 7606 _work_queue, 7607 _overflow_stack, 7608 _revisit_stack, 7609 _finger, 7610 gfa, this); 7611 bool res = _work_queue->push(obj); // overflow could occur here 7612 assert(res, "Will hold once we use workqueues"); 7613 while (true) { 7614 oop new_oop; 7615 if (!_work_queue->pop_local(new_oop)) { 7616 // We emptied our work_queue; check if there's stuff that can 7617 // be gotten from the overflow stack. 7618 if (CMSConcMarkingTask::get_work_from_overflow_stack( 7619 _overflow_stack, _work_queue)) { 7620 do_yield_check(); 7621 continue; 7622 } else { // done 7623 break; 7624 } 7625 } 7626 // Skip verifying header mark word below because we are 7627 // running concurrent with mutators. 7628 assert(new_oop->is_oop(true), "Oops! expected to pop an oop"); 7629 // now scan this oop's oops 7630 new_oop->oop_iterate(&pushOrMarkClosure); 7631 do_yield_check(); 7632 } 7633 assert(_work_queue->size() == 0, "tautology, emphasizing post-condition"); 7634 } 7635 7636 // Yield in response to a request from VM Thread or 7637 // from mutators. 7638 void Par_MarkFromRootsClosure::do_yield_work() { 7639 assert(_task != NULL, "sanity"); 7640 _task->yield(); 7641 } 7642 7643 // A variant of the above used for verifying CMS marking work. 7644 MarkFromRootsVerifyClosure::MarkFromRootsVerifyClosure(CMSCollector* collector, 7645 MemRegion span, 7646 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 7647 CMSMarkStack* mark_stack): 7648 _collector(collector), 7649 _span(span), 7650 _verification_bm(verification_bm), 7651 _cms_bm(cms_bm), 7652 _mark_stack(mark_stack), 7653 _pam_verify_closure(collector, span, verification_bm, cms_bm, 7654 mark_stack) 7655 { 7656 assert(_mark_stack->isEmpty(), "stack should be empty"); 7657 _finger = _verification_bm->startWord(); 7658 assert(_collector->_restart_addr == NULL, "Sanity check"); 7659 assert(_span.contains(_finger), "Out of bounds _finger?"); 7660 } 7661 7662 void MarkFromRootsVerifyClosure::reset(HeapWord* addr) { 7663 assert(_mark_stack->isEmpty(), "would cause duplicates on stack"); 7664 assert(_span.contains(addr), "Out of bounds _finger?"); 7665 _finger = addr; 7666 } 7667 7668 // Should revisit to see if this should be restructured for 7669 // greater efficiency. 7670 bool MarkFromRootsVerifyClosure::do_bit(size_t offset) { 7671 // convert offset into a HeapWord* 7672 HeapWord* addr = _verification_bm->startWord() + offset; 7673 assert(_verification_bm->endWord() && addr < _verification_bm->endWord(), 7674 "address out of range"); 7675 assert(_verification_bm->isMarked(addr), "tautology"); 7676 assert(_cms_bm->isMarked(addr), "tautology"); 7677 7678 assert(_mark_stack->isEmpty(), 7679 "should drain stack to limit stack usage"); 7680 // convert addr to an oop preparatory to scanning 7681 oop obj = oop(addr); 7682 assert(obj->is_oop(), "should be an oop"); 7683 assert(_finger <= addr, "_finger runneth ahead"); 7684 // advance the finger to right end of this object 7685 _finger = addr + obj->size(); 7686 assert(_finger > addr, "we just incremented it above"); 7687 // Note: the finger doesn't advance while we drain 7688 // the stack below. 7689 bool res = _mark_stack->push(obj); 7690 assert(res, "Empty non-zero size stack should have space for single push"); 7691 while (!_mark_stack->isEmpty()) { 7692 oop new_oop = _mark_stack->pop(); 7693 assert(new_oop->is_oop(), "Oops! expected to pop an oop"); 7694 // now scan this oop's oops 7695 new_oop->oop_iterate(&_pam_verify_closure); 7696 } 7697 assert(_mark_stack->isEmpty(), "tautology, emphasizing post-condition"); 7698 return true; 7699 } 7700 7701 PushAndMarkVerifyClosure::PushAndMarkVerifyClosure( 7702 CMSCollector* collector, MemRegion span, 7703 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 7704 CMSMarkStack* mark_stack): 7705 OopClosure(collector->ref_processor()), 7706 _collector(collector), 7707 _span(span), 7708 _verification_bm(verification_bm), 7709 _cms_bm(cms_bm), 7710 _mark_stack(mark_stack) 7711 { } 7712 7713 void PushAndMarkVerifyClosure::do_oop(oop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 7714 void PushAndMarkVerifyClosure::do_oop(narrowOop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 7715 7716 // Upon stack overflow, we discard (part of) the stack, 7717 // remembering the least address amongst those discarded 7718 // in CMSCollector's _restart_address. 7719 void PushAndMarkVerifyClosure::handle_stack_overflow(HeapWord* lost) { 7720 // Remember the least grey address discarded 7721 HeapWord* ra = (HeapWord*)_mark_stack->least_value(lost); 7722 _collector->lower_restart_addr(ra); 7723 _mark_stack->reset(); // discard stack contents 7724 _mark_stack->expand(); // expand the stack if possible 7725 } 7726 7727 void PushAndMarkVerifyClosure::do_oop(oop obj) { 7728 assert(obj->is_oop_or_null(), "expected an oop or NULL"); 7729 HeapWord* addr = (HeapWord*)obj; 7730 if (_span.contains(addr) && !_verification_bm->isMarked(addr)) { 7731 // Oop lies in _span and isn't yet grey or black 7732 _verification_bm->mark(addr); // now grey 7733 if (!_cms_bm->isMarked(addr)) { 7734 oop(addr)->print(); 7735 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)", 7736 addr); 7737 fatal("... aborting"); 7738 } 7739 7740 if (!_mark_stack->push(obj)) { // stack overflow 7741 if (PrintCMSStatistics != 0) { 7742 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 7743 SIZE_FORMAT, _mark_stack->capacity()); 7744 } 7745 assert(_mark_stack->isFull(), "Else push should have succeeded"); 7746 handle_stack_overflow(addr); 7747 } 7748 // anything including and to the right of _finger 7749 // will be scanned as we iterate over the remainder of the 7750 // bit map 7751 } 7752 } 7753 7754 PushOrMarkClosure::PushOrMarkClosure(CMSCollector* collector, 7755 MemRegion span, 7756 CMSBitMap* bitMap, CMSMarkStack* markStack, 7757 CMSMarkStack* revisitStack, 7758 HeapWord* finger, MarkFromRootsClosure* parent) : 7759 KlassRememberingOopClosure(collector, collector->ref_processor(), revisitStack), 7760 _span(span), 7761 _bitMap(bitMap), 7762 _markStack(markStack), 7763 _finger(finger), 7764 _parent(parent) 7765 { } 7766 7767 Par_PushOrMarkClosure::Par_PushOrMarkClosure(CMSCollector* collector, 7768 MemRegion span, 7769 CMSBitMap* bit_map, 7770 OopTaskQueue* work_queue, 7771 CMSMarkStack* overflow_stack, 7772 CMSMarkStack* revisit_stack, 7773 HeapWord* finger, 7774 HeapWord** global_finger_addr, 7775 Par_MarkFromRootsClosure* parent) : 7776 Par_KlassRememberingOopClosure(collector, 7777 collector->ref_processor(), 7778 revisit_stack), 7779 _whole_span(collector->_span), 7780 _span(span), 7781 _bit_map(bit_map), 7782 _work_queue(work_queue), 7783 _overflow_stack(overflow_stack), 7784 _finger(finger), 7785 _global_finger_addr(global_finger_addr), 7786 _parent(parent) 7787 { } 7788 7789 // Assumes thread-safe access by callers, who are 7790 // responsible for mutual exclusion. 7791 void CMSCollector::lower_restart_addr(HeapWord* low) { 7792 assert(_span.contains(low), "Out of bounds addr"); 7793 if (_restart_addr == NULL) { 7794 _restart_addr = low; 7795 } else { 7796 _restart_addr = MIN2(_restart_addr, low); 7797 } 7798 } 7799 7800 // Upon stack overflow, we discard (part of) the stack, 7801 // remembering the least address amongst those discarded 7802 // in CMSCollector's _restart_address. 7803 void PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 7804 // Remember the least grey address discarded 7805 HeapWord* ra = (HeapWord*)_markStack->least_value(lost); 7806 _collector->lower_restart_addr(ra); 7807 _markStack->reset(); // discard stack contents 7808 _markStack->expand(); // expand the stack if possible 7809 } 7810 7811 // Upon stack overflow, we discard (part of) the stack, 7812 // remembering the least address amongst those discarded 7813 // in CMSCollector's _restart_address. 7814 void Par_PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 7815 // We need to do this under a mutex to prevent other 7816 // workers from interfering with the work done below. 7817 MutexLockerEx ml(_overflow_stack->par_lock(), 7818 Mutex::_no_safepoint_check_flag); 7819 // Remember the least grey address discarded 7820 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 7821 _collector->lower_restart_addr(ra); 7822 _overflow_stack->reset(); // discard stack contents 7823 _overflow_stack->expand(); // expand the stack if possible 7824 } 7825 7826 void PushOrMarkClosure::do_oop(oop obj) { 7827 // Ignore mark word because we are running concurrent with mutators. 7828 assert(obj->is_oop_or_null(true), "expected an oop or NULL"); 7829 HeapWord* addr = (HeapWord*)obj; 7830 if (_span.contains(addr) && !_bitMap->isMarked(addr)) { 7831 // Oop lies in _span and isn't yet grey or black 7832 _bitMap->mark(addr); // now grey 7833 if (addr < _finger) { 7834 // the bit map iteration has already either passed, or 7835 // sampled, this bit in the bit map; we'll need to 7836 // use the marking stack to scan this oop's oops. 7837 bool simulate_overflow = false; 7838 NOT_PRODUCT( 7839 if (CMSMarkStackOverflowALot && 7840 _collector->simulate_overflow()) { 7841 // simulate a stack overflow 7842 simulate_overflow = true; 7843 } 7844 ) 7845 if (simulate_overflow || !_markStack->push(obj)) { // stack overflow 7846 if (PrintCMSStatistics != 0) { 7847 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 7848 SIZE_FORMAT, _markStack->capacity()); 7849 } 7850 assert(simulate_overflow || _markStack->isFull(), "Else push should have succeeded"); 7851 handle_stack_overflow(addr); 7852 } 7853 } 7854 // anything including and to the right of _finger 7855 // will be scanned as we iterate over the remainder of the 7856 // bit map 7857 do_yield_check(); 7858 } 7859 } 7860 7861 void PushOrMarkClosure::do_oop(oop* p) { PushOrMarkClosure::do_oop_work(p); } 7862 void PushOrMarkClosure::do_oop(narrowOop* p) { PushOrMarkClosure::do_oop_work(p); } 7863 7864 void Par_PushOrMarkClosure::do_oop(oop obj) { 7865 // Ignore mark word because we are running concurrent with mutators. 7866 assert(obj->is_oop_or_null(true), "expected an oop or NULL"); 7867 HeapWord* addr = (HeapWord*)obj; 7868 if (_whole_span.contains(addr) && !_bit_map->isMarked(addr)) { 7869 // Oop lies in _span and isn't yet grey or black 7870 // We read the global_finger (volatile read) strictly after marking oop 7871 bool res = _bit_map->par_mark(addr); // now grey 7872 volatile HeapWord** gfa = (volatile HeapWord**)_global_finger_addr; 7873 // Should we push this marked oop on our stack? 7874 // -- if someone else marked it, nothing to do 7875 // -- if target oop is above global finger nothing to do 7876 // -- if target oop is in chunk and above local finger 7877 // then nothing to do 7878 // -- else push on work queue 7879 if ( !res // someone else marked it, they will deal with it 7880 || (addr >= *gfa) // will be scanned in a later task 7881 || (_span.contains(addr) && addr >= _finger)) { // later in this chunk 7882 return; 7883 } 7884 // the bit map iteration has already either passed, or 7885 // sampled, this bit in the bit map; we'll need to 7886 // use the marking stack to scan this oop's oops. 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 || 7896 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) { 7897 // stack overflow 7898 if (PrintCMSStatistics != 0) { 7899 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 7900 SIZE_FORMAT, _overflow_stack->capacity()); 7901 } 7902 // We cannot assert that the overflow stack is full because 7903 // it may have been emptied since. 7904 assert(simulate_overflow || 7905 _work_queue->size() == _work_queue->max_elems(), 7906 "Else push should have succeeded"); 7907 handle_stack_overflow(addr); 7908 } 7909 do_yield_check(); 7910 } 7911 } 7912 7913 void Par_PushOrMarkClosure::do_oop(oop* p) { Par_PushOrMarkClosure::do_oop_work(p); } 7914 void Par_PushOrMarkClosure::do_oop(narrowOop* p) { Par_PushOrMarkClosure::do_oop_work(p); } 7915 7916 KlassRememberingOopClosure::KlassRememberingOopClosure(CMSCollector* collector, 7917 ReferenceProcessor* rp, 7918 CMSMarkStack* revisit_stack) : 7919 OopClosure(rp), 7920 _collector(collector), 7921 _revisit_stack(revisit_stack), 7922 _should_remember_klasses(collector->should_unload_classes()) {} 7923 7924 PushAndMarkClosure::PushAndMarkClosure(CMSCollector* collector, 7925 MemRegion span, 7926 ReferenceProcessor* rp, 7927 CMSBitMap* bit_map, 7928 CMSBitMap* mod_union_table, 7929 CMSMarkStack* mark_stack, 7930 CMSMarkStack* revisit_stack, 7931 bool concurrent_precleaning): 7932 KlassRememberingOopClosure(collector, rp, revisit_stack), 7933 _span(span), 7934 _bit_map(bit_map), 7935 _mod_union_table(mod_union_table), 7936 _mark_stack(mark_stack), 7937 _concurrent_precleaning(concurrent_precleaning) 7938 { 7939 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 7940 } 7941 7942 // Grey object rescan during pre-cleaning and second checkpoint phases -- 7943 // the non-parallel version (the parallel version appears further below.) 7944 void PushAndMarkClosure::do_oop(oop obj) { 7945 // Ignore mark word verification. If during concurrent precleaning, 7946 // the object monitor may be locked. If during the checkpoint 7947 // phases, the object may already have been reached by a different 7948 // path and may be at the end of the global overflow list (so 7949 // the mark word may be NULL). 7950 assert(obj->is_oop_or_null(true /* ignore mark word */), 7951 "expected an oop or NULL"); 7952 HeapWord* addr = (HeapWord*)obj; 7953 // Check if oop points into the CMS generation 7954 // and is not marked 7955 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 7956 // a white object ... 7957 _bit_map->mark(addr); // ... now grey 7958 // push on the marking stack (grey set) 7959 bool simulate_overflow = false; 7960 NOT_PRODUCT( 7961 if (CMSMarkStackOverflowALot && 7962 _collector->simulate_overflow()) { 7963 // simulate a stack overflow 7964 simulate_overflow = true; 7965 } 7966 ) 7967 if (simulate_overflow || !_mark_stack->push(obj)) { 7968 if (_concurrent_precleaning) { 7969 // During precleaning we can just dirty the appropriate card(s) 7970 // in the mod union table, thus ensuring that the object remains 7971 // in the grey set and continue. In the case of object arrays 7972 // we need to dirty all of the cards that the object spans, 7973 // since the rescan of object arrays will be limited to the 7974 // dirty cards. 7975 // Note that no one can be intefering with us in this action 7976 // of dirtying the mod union table, so no locking or atomics 7977 // are required. 7978 if (obj->is_objArray()) { 7979 size_t sz = obj->size(); 7980 HeapWord* end_card_addr = (HeapWord*)round_to( 7981 (intptr_t)(addr+sz), CardTableModRefBS::card_size); 7982 MemRegion redirty_range = MemRegion(addr, end_card_addr); 7983 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 7984 _mod_union_table->mark_range(redirty_range); 7985 } else { 7986 _mod_union_table->mark(addr); 7987 } 7988 _collector->_ser_pmc_preclean_ovflw++; 7989 } else { 7990 // During the remark phase, we need to remember this oop 7991 // in the overflow list. 7992 _collector->push_on_overflow_list(obj); 7993 _collector->_ser_pmc_remark_ovflw++; 7994 } 7995 } 7996 } 7997 } 7998 7999 Par_PushAndMarkClosure::Par_PushAndMarkClosure(CMSCollector* collector, 8000 MemRegion span, 8001 ReferenceProcessor* rp, 8002 CMSBitMap* bit_map, 8003 OopTaskQueue* work_queue, 8004 CMSMarkStack* revisit_stack): 8005 Par_KlassRememberingOopClosure(collector, rp, revisit_stack), 8006 _span(span), 8007 _bit_map(bit_map), 8008 _work_queue(work_queue) 8009 { 8010 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 8011 } 8012 8013 void PushAndMarkClosure::do_oop(oop* p) { PushAndMarkClosure::do_oop_work(p); } 8014 void PushAndMarkClosure::do_oop(narrowOop* p) { PushAndMarkClosure::do_oop_work(p); } 8015 8016 // Grey object rescan during second checkpoint phase -- 8017 // the parallel version. 8018 void Par_PushAndMarkClosure::do_oop(oop obj) { 8019 // In the assert below, we ignore the mark word because 8020 // this oop may point to an already visited object that is 8021 // on the overflow stack (in which case the mark word has 8022 // been hijacked for chaining into the overflow stack -- 8023 // if this is the last object in the overflow stack then 8024 // its mark word will be NULL). Because this object may 8025 // have been subsequently popped off the global overflow 8026 // stack, and the mark word possibly restored to the prototypical 8027 // value, by the time we get to examined this failing assert in 8028 // the debugger, is_oop_or_null(false) may subsequently start 8029 // to hold. 8030 assert(obj->is_oop_or_null(true), 8031 "expected an oop or NULL"); 8032 HeapWord* addr = (HeapWord*)obj; 8033 // Check if oop points into the CMS generation 8034 // and is not marked 8035 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 8036 // a white object ... 8037 // If we manage to "claim" the object, by being the 8038 // first thread to mark it, then we push it on our 8039 // marking stack 8040 if (_bit_map->par_mark(addr)) { // ... now grey 8041 // push on work queue (grey set) 8042 bool simulate_overflow = false; 8043 NOT_PRODUCT( 8044 if (CMSMarkStackOverflowALot && 8045 _collector->par_simulate_overflow()) { 8046 // simulate a stack overflow 8047 simulate_overflow = true; 8048 } 8049 ) 8050 if (simulate_overflow || !_work_queue->push(obj)) { 8051 _collector->par_push_on_overflow_list(obj); 8052 _collector->_par_pmc_remark_ovflw++; // imprecise OK: no need to CAS 8053 } 8054 } // Else, some other thread got there first 8055 } 8056 } 8057 8058 void Par_PushAndMarkClosure::do_oop(oop* p) { Par_PushAndMarkClosure::do_oop_work(p); } 8059 void Par_PushAndMarkClosure::do_oop(narrowOop* p) { Par_PushAndMarkClosure::do_oop_work(p); } 8060 8061 void PushAndMarkClosure::remember_mdo(DataLayout* v) { 8062 // TBD 8063 } 8064 8065 void Par_PushAndMarkClosure::remember_mdo(DataLayout* v) { 8066 // TBD 8067 } 8068 8069 void CMSPrecleanRefsYieldClosure::do_yield_work() { 8070 DEBUG_ONLY(RememberKlassesChecker mux(false);) 8071 Mutex* bml = _collector->bitMapLock(); 8072 assert_lock_strong(bml); 8073 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 8074 "CMS thread should hold CMS token"); 8075 8076 bml->unlock(); 8077 ConcurrentMarkSweepThread::desynchronize(true); 8078 8079 ConcurrentMarkSweepThread::acknowledge_yield_request(); 8080 8081 _collector->stopTimer(); 8082 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 8083 if (PrintCMSStatistics != 0) { 8084 _collector->incrementYields(); 8085 } 8086 _collector->icms_wait(); 8087 8088 // See the comment in coordinator_yield() 8089 for (unsigned i = 0; i < CMSYieldSleepCount && 8090 ConcurrentMarkSweepThread::should_yield() && 8091 !CMSCollector::foregroundGCIsActive(); ++i) { 8092 os::sleep(Thread::current(), 1, false); 8093 ConcurrentMarkSweepThread::acknowledge_yield_request(); 8094 } 8095 8096 ConcurrentMarkSweepThread::synchronize(true); 8097 bml->lock(); 8098 8099 _collector->startTimer(); 8100 } 8101 8102 bool CMSPrecleanRefsYieldClosure::should_return() { 8103 if (ConcurrentMarkSweepThread::should_yield()) { 8104 do_yield_work(); 8105 } 8106 return _collector->foregroundGCIsActive(); 8107 } 8108 8109 void MarkFromDirtyCardsClosure::do_MemRegion(MemRegion mr) { 8110 assert(((size_t)mr.start())%CardTableModRefBS::card_size_in_words == 0, 8111 "mr should be aligned to start at a card boundary"); 8112 // We'd like to assert: 8113 // assert(mr.word_size()%CardTableModRefBS::card_size_in_words == 0, 8114 // "mr should be a range of cards"); 8115 // However, that would be too strong in one case -- the last 8116 // partition ends at _unallocated_block which, in general, can be 8117 // an arbitrary boundary, not necessarily card aligned. 8118 if (PrintCMSStatistics != 0) { 8119 _num_dirty_cards += 8120 mr.word_size()/CardTableModRefBS::card_size_in_words; 8121 } 8122 _space->object_iterate_mem(mr, &_scan_cl); 8123 } 8124 8125 SweepClosure::SweepClosure(CMSCollector* collector, 8126 ConcurrentMarkSweepGeneration* g, 8127 CMSBitMap* bitMap, bool should_yield) : 8128 _collector(collector), 8129 _g(g), 8130 _sp(g->cmsSpace()), 8131 _limit(_sp->sweep_limit()), 8132 _freelistLock(_sp->freelistLock()), 8133 _bitMap(bitMap), 8134 _yield(should_yield), 8135 _inFreeRange(false), // No free range at beginning of sweep 8136 _freeRangeInFreeLists(false), // No free range at beginning of sweep 8137 _lastFreeRangeCoalesced(false), 8138 _freeFinger(g->used_region().start()) 8139 { 8140 NOT_PRODUCT( 8141 _numObjectsFreed = 0; 8142 _numWordsFreed = 0; 8143 _numObjectsLive = 0; 8144 _numWordsLive = 0; 8145 _numObjectsAlreadyFree = 0; 8146 _numWordsAlreadyFree = 0; 8147 _last_fc = NULL; 8148 8149 _sp->initializeIndexedFreeListArrayReturnedBytes(); 8150 _sp->dictionary()->initialize_dict_returned_bytes(); 8151 ) 8152 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 8153 "sweep _limit out of bounds"); 8154 if (CMSTraceSweeper) { 8155 gclog_or_tty->print_cr("\n====================\nStarting new sweep with limit " PTR_FORMAT, 8156 _limit); 8157 } 8158 } 8159 8160 void SweepClosure::print_on(outputStream* st) const { 8161 tty->print_cr("_sp = [" PTR_FORMAT "," PTR_FORMAT ")", 8162 _sp->bottom(), _sp->end()); 8163 tty->print_cr("_limit = " PTR_FORMAT, _limit); 8164 tty->print_cr("_freeFinger = " PTR_FORMAT, _freeFinger); 8165 NOT_PRODUCT(tty->print_cr("_last_fc = " PTR_FORMAT, _last_fc);) 8166 tty->print_cr("_inFreeRange = %d, _freeRangeInFreeLists = %d, _lastFreeRangeCoalesced = %d", 8167 _inFreeRange, _freeRangeInFreeLists, _lastFreeRangeCoalesced); 8168 } 8169 8170 #ifndef PRODUCT 8171 // Assertion checking only: no useful work in product mode -- 8172 // however, if any of the flags below become product flags, 8173 // you may need to review this code to see if it needs to be 8174 // enabled in product mode. 8175 SweepClosure::~SweepClosure() { 8176 assert_lock_strong(_freelistLock); 8177 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 8178 "sweep _limit out of bounds"); 8179 if (inFreeRange()) { 8180 warning("inFreeRange() should have been reset; dumping state of SweepClosure"); 8181 print(); 8182 ShouldNotReachHere(); 8183 } 8184 if (Verbose && PrintGC) { 8185 gclog_or_tty->print("Collected "SIZE_FORMAT" objects, " SIZE_FORMAT " bytes", 8186 _numObjectsFreed, _numWordsFreed*sizeof(HeapWord)); 8187 gclog_or_tty->print_cr("\nLive "SIZE_FORMAT" objects, " 8188 SIZE_FORMAT" bytes " 8189 "Already free "SIZE_FORMAT" objects, "SIZE_FORMAT" bytes", 8190 _numObjectsLive, _numWordsLive*sizeof(HeapWord), 8191 _numObjectsAlreadyFree, _numWordsAlreadyFree*sizeof(HeapWord)); 8192 size_t totalBytes = (_numWordsFreed + _numWordsLive + _numWordsAlreadyFree) 8193 * sizeof(HeapWord); 8194 gclog_or_tty->print_cr("Total sweep: "SIZE_FORMAT" bytes", totalBytes); 8195 8196 if (PrintCMSStatistics && CMSVerifyReturnedBytes) { 8197 size_t indexListReturnedBytes = _sp->sumIndexedFreeListArrayReturnedBytes(); 8198 size_t dict_returned_bytes = _sp->dictionary()->sum_dict_returned_bytes(); 8199 size_t returned_bytes = indexListReturnedBytes + dict_returned_bytes; 8200 gclog_or_tty->print("Returned "SIZE_FORMAT" bytes", returned_bytes); 8201 gclog_or_tty->print(" Indexed List Returned "SIZE_FORMAT" bytes", 8202 indexListReturnedBytes); 8203 gclog_or_tty->print_cr(" Dictionary Returned "SIZE_FORMAT" bytes", 8204 dict_returned_bytes); 8205 } 8206 } 8207 if (CMSTraceSweeper) { 8208 gclog_or_tty->print_cr("end of sweep with _limit = " PTR_FORMAT "\n================", 8209 _limit); 8210 } 8211 } 8212 #endif // PRODUCT 8213 8214 void SweepClosure::initialize_free_range(HeapWord* freeFinger, 8215 bool freeRangeInFreeLists) { 8216 if (CMSTraceSweeper) { 8217 gclog_or_tty->print("---- Start free range at 0x%x with free block (%d)\n", 8218 freeFinger, freeRangeInFreeLists); 8219 } 8220 assert(!inFreeRange(), "Trampling existing free range"); 8221 set_inFreeRange(true); 8222 set_lastFreeRangeCoalesced(false); 8223 8224 set_freeFinger(freeFinger); 8225 set_freeRangeInFreeLists(freeRangeInFreeLists); 8226 if (CMSTestInFreeList) { 8227 if (freeRangeInFreeLists) { 8228 FreeChunk* fc = (FreeChunk*) freeFinger; 8229 assert(fc->is_free(), "A chunk on the free list should be free."); 8230 assert(fc->size() > 0, "Free range should have a size"); 8231 assert(_sp->verify_chunk_in_free_list(fc), "Chunk is not in free lists"); 8232 } 8233 } 8234 } 8235 8236 // Note that the sweeper runs concurrently with mutators. Thus, 8237 // it is possible for direct allocation in this generation to happen 8238 // in the middle of the sweep. Note that the sweeper also coalesces 8239 // contiguous free blocks. Thus, unless the sweeper and the allocator 8240 // synchronize appropriately freshly allocated blocks may get swept up. 8241 // This is accomplished by the sweeper locking the free lists while 8242 // it is sweeping. Thus blocks that are determined to be free are 8243 // indeed free. There is however one additional complication: 8244 // blocks that have been allocated since the final checkpoint and 8245 // mark, will not have been marked and so would be treated as 8246 // unreachable and swept up. To prevent this, the allocator marks 8247 // the bit map when allocating during the sweep phase. This leads, 8248 // however, to a further complication -- objects may have been allocated 8249 // but not yet initialized -- in the sense that the header isn't yet 8250 // installed. The sweeper can not then determine the size of the block 8251 // in order to skip over it. To deal with this case, we use a technique 8252 // (due to Printezis) to encode such uninitialized block sizes in the 8253 // bit map. Since the bit map uses a bit per every HeapWord, but the 8254 // CMS generation has a minimum object size of 3 HeapWords, it follows 8255 // that "normal marks" won't be adjacent in the bit map (there will 8256 // always be at least two 0 bits between successive 1 bits). We make use 8257 // of these "unused" bits to represent uninitialized blocks -- the bit 8258 // corresponding to the start of the uninitialized object and the next 8259 // bit are both set. Finally, a 1 bit marks the end of the object that 8260 // started with the two consecutive 1 bits to indicate its potentially 8261 // uninitialized state. 8262 8263 size_t SweepClosure::do_blk_careful(HeapWord* addr) { 8264 FreeChunk* fc = (FreeChunk*)addr; 8265 size_t res; 8266 8267 // Check if we are done sweeping. Below we check "addr >= _limit" rather 8268 // than "addr == _limit" because although _limit was a block boundary when 8269 // we started the sweep, it may no longer be one because heap expansion 8270 // may have caused us to coalesce the block ending at the address _limit 8271 // with a newly expanded chunk (this happens when _limit was set to the 8272 // previous _end of the space), so we may have stepped past _limit: 8273 // see the following Zeno-like trail of CRs 6977970, 7008136, 7042740. 8274 if (addr >= _limit) { // we have swept up to or past the limit: finish up 8275 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 8276 "sweep _limit out of bounds"); 8277 assert(addr < _sp->end(), "addr out of bounds"); 8278 // Flush any free range we might be holding as a single 8279 // coalesced chunk to the appropriate free list. 8280 if (inFreeRange()) { 8281 assert(freeFinger() >= _sp->bottom() && freeFinger() < _limit, 8282 err_msg("freeFinger() " PTR_FORMAT" is out-of-bounds", freeFinger())); 8283 flush_cur_free_chunk(freeFinger(), 8284 pointer_delta(addr, freeFinger())); 8285 if (CMSTraceSweeper) { 8286 gclog_or_tty->print("Sweep: last chunk: "); 8287 gclog_or_tty->print("put_free_blk 0x%x ("SIZE_FORMAT") " 8288 "[coalesced:"SIZE_FORMAT"]\n", 8289 freeFinger(), pointer_delta(addr, freeFinger()), 8290 lastFreeRangeCoalesced()); 8291 } 8292 } 8293 8294 // help the iterator loop finish 8295 return pointer_delta(_sp->end(), addr); 8296 } 8297 8298 assert(addr < _limit, "sweep invariant"); 8299 // check if we should yield 8300 do_yield_check(addr); 8301 if (fc->is_free()) { 8302 // Chunk that is already free 8303 res = fc->size(); 8304 do_already_free_chunk(fc); 8305 debug_only(_sp->verifyFreeLists()); 8306 // If we flush the chunk at hand in lookahead_and_flush() 8307 // and it's coalesced with a preceding chunk, then the 8308 // process of "mangling" the payload of the coalesced block 8309 // will cause erasure of the size information from the 8310 // (erstwhile) header of all the coalesced blocks but the 8311 // first, so the first disjunct in the assert will not hold 8312 // in that specific case (in which case the second disjunct 8313 // will hold). 8314 assert(res == fc->size() || ((HeapWord*)fc) + res >= _limit, 8315 "Otherwise the size info doesn't change at this step"); 8316 NOT_PRODUCT( 8317 _numObjectsAlreadyFree++; 8318 _numWordsAlreadyFree += res; 8319 ) 8320 NOT_PRODUCT(_last_fc = fc;) 8321 } else if (!_bitMap->isMarked(addr)) { 8322 // Chunk is fresh garbage 8323 res = do_garbage_chunk(fc); 8324 debug_only(_sp->verifyFreeLists()); 8325 NOT_PRODUCT( 8326 _numObjectsFreed++; 8327 _numWordsFreed += res; 8328 ) 8329 } else { 8330 // Chunk that is alive. 8331 res = do_live_chunk(fc); 8332 debug_only(_sp->verifyFreeLists()); 8333 NOT_PRODUCT( 8334 _numObjectsLive++; 8335 _numWordsLive += res; 8336 ) 8337 } 8338 return res; 8339 } 8340 8341 // For the smart allocation, record following 8342 // split deaths - a free chunk is removed from its free list because 8343 // it is being split into two or more chunks. 8344 // split birth - a free chunk is being added to its free list because 8345 // a larger free chunk has been split and resulted in this free chunk. 8346 // coal death - a free chunk is being removed from its free list because 8347 // it is being coalesced into a large free chunk. 8348 // coal birth - a free chunk is being added to its free list because 8349 // it was created when two or more free chunks where coalesced into 8350 // this free chunk. 8351 // 8352 // These statistics are used to determine the desired number of free 8353 // chunks of a given size. The desired number is chosen to be relative 8354 // to the end of a CMS sweep. The desired number at the end of a sweep 8355 // is the 8356 // count-at-end-of-previous-sweep (an amount that was enough) 8357 // - count-at-beginning-of-current-sweep (the excess) 8358 // + split-births (gains in this size during interval) 8359 // - split-deaths (demands on this size during interval) 8360 // where the interval is from the end of one sweep to the end of the 8361 // next. 8362 // 8363 // When sweeping the sweeper maintains an accumulated chunk which is 8364 // the chunk that is made up of chunks that have been coalesced. That 8365 // will be termed the left-hand chunk. A new chunk of garbage that 8366 // is being considered for coalescing will be referred to as the 8367 // right-hand chunk. 8368 // 8369 // When making a decision on whether to coalesce a right-hand chunk with 8370 // the current left-hand chunk, the current count vs. the desired count 8371 // of the left-hand chunk is considered. Also if the right-hand chunk 8372 // is near the large chunk at the end of the heap (see 8373 // ConcurrentMarkSweepGeneration::isNearLargestChunk()), then the 8374 // left-hand chunk is coalesced. 8375 // 8376 // When making a decision about whether to split a chunk, the desired count 8377 // vs. the current count of the candidate to be split is also considered. 8378 // If the candidate is underpopulated (currently fewer chunks than desired) 8379 // a chunk of an overpopulated (currently more chunks than desired) size may 8380 // be chosen. The "hint" associated with a free list, if non-null, points 8381 // to a free list which may be overpopulated. 8382 // 8383 8384 void SweepClosure::do_already_free_chunk(FreeChunk* fc) { 8385 const size_t size = fc->size(); 8386 // Chunks that cannot be coalesced are not in the 8387 // free lists. 8388 if (CMSTestInFreeList && !fc->cantCoalesce()) { 8389 assert(_sp->verify_chunk_in_free_list(fc), 8390 "free chunk should be in free lists"); 8391 } 8392 // a chunk that is already free, should not have been 8393 // marked in the bit map 8394 HeapWord* const addr = (HeapWord*) fc; 8395 assert(!_bitMap->isMarked(addr), "free chunk should be unmarked"); 8396 // Verify that the bit map has no bits marked between 8397 // addr and purported end of this block. 8398 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 8399 8400 // Some chunks cannot be coalesced under any circumstances. 8401 // See the definition of cantCoalesce(). 8402 if (!fc->cantCoalesce()) { 8403 // This chunk can potentially be coalesced. 8404 if (_sp->adaptive_freelists()) { 8405 // All the work is done in 8406 do_post_free_or_garbage_chunk(fc, size); 8407 } else { // Not adaptive free lists 8408 // this is a free chunk that can potentially be coalesced by the sweeper; 8409 if (!inFreeRange()) { 8410 // if the next chunk is a free block that can't be coalesced 8411 // it doesn't make sense to remove this chunk from the free lists 8412 FreeChunk* nextChunk = (FreeChunk*)(addr + size); 8413 assert((HeapWord*)nextChunk <= _sp->end(), "Chunk size out of bounds?"); 8414 if ((HeapWord*)nextChunk < _sp->end() && // There is another free chunk to the right ... 8415 nextChunk->is_free() && // ... which is free... 8416 nextChunk->cantCoalesce()) { // ... but can't be coalesced 8417 // nothing to do 8418 } else { 8419 // Potentially the start of a new free range: 8420 // Don't eagerly remove it from the free lists. 8421 // No need to remove it if it will just be put 8422 // back again. (Also from a pragmatic point of view 8423 // if it is a free block in a region that is beyond 8424 // any allocated blocks, an assertion will fail) 8425 // Remember the start of a free run. 8426 initialize_free_range(addr, true); 8427 // end - can coalesce with next chunk 8428 } 8429 } else { 8430 // the midst of a free range, we are coalescing 8431 print_free_block_coalesced(fc); 8432 if (CMSTraceSweeper) { 8433 gclog_or_tty->print(" -- pick up free block 0x%x (%d)\n", fc, size); 8434 } 8435 // remove it from the free lists 8436 _sp->removeFreeChunkFromFreeLists(fc); 8437 set_lastFreeRangeCoalesced(true); 8438 // If the chunk is being coalesced and the current free range is 8439 // in the free lists, remove the current free range so that it 8440 // will be returned to the free lists in its entirety - all 8441 // the coalesced pieces included. 8442 if (freeRangeInFreeLists()) { 8443 FreeChunk* ffc = (FreeChunk*) freeFinger(); 8444 assert(ffc->size() == pointer_delta(addr, freeFinger()), 8445 "Size of free range is inconsistent with chunk size."); 8446 if (CMSTestInFreeList) { 8447 assert(_sp->verify_chunk_in_free_list(ffc), 8448 "free range is not in free lists"); 8449 } 8450 _sp->removeFreeChunkFromFreeLists(ffc); 8451 set_freeRangeInFreeLists(false); 8452 } 8453 } 8454 } 8455 // Note that if the chunk is not coalescable (the else arm 8456 // below), we unconditionally flush, without needing to do 8457 // a "lookahead," as we do below. 8458 if (inFreeRange()) lookahead_and_flush(fc, size); 8459 } else { 8460 // Code path common to both original and adaptive free lists. 8461 8462 // cant coalesce with previous block; this should be treated 8463 // as the end of a free run if any 8464 if (inFreeRange()) { 8465 // we kicked some butt; time to pick up the garbage 8466 assert(freeFinger() < addr, "freeFinger points too high"); 8467 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 8468 } 8469 // else, nothing to do, just continue 8470 } 8471 } 8472 8473 size_t SweepClosure::do_garbage_chunk(FreeChunk* fc) { 8474 // This is a chunk of garbage. It is not in any free list. 8475 // Add it to a free list or let it possibly be coalesced into 8476 // a larger chunk. 8477 HeapWord* const addr = (HeapWord*) fc; 8478 const size_t size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 8479 8480 if (_sp->adaptive_freelists()) { 8481 // Verify that the bit map has no bits marked between 8482 // addr and purported end of just dead object. 8483 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 8484 8485 do_post_free_or_garbage_chunk(fc, size); 8486 } else { 8487 if (!inFreeRange()) { 8488 // start of a new free range 8489 assert(size > 0, "A free range should have a size"); 8490 initialize_free_range(addr, false); 8491 } else { 8492 // this will be swept up when we hit the end of the 8493 // free range 8494 if (CMSTraceSweeper) { 8495 gclog_or_tty->print(" -- pick up garbage 0x%x (%d) \n", fc, size); 8496 } 8497 // If the chunk is being coalesced and the current free range is 8498 // in the free lists, remove the current free range so that it 8499 // will be returned to the free lists in its entirety - all 8500 // the coalesced pieces included. 8501 if (freeRangeInFreeLists()) { 8502 FreeChunk* ffc = (FreeChunk*)freeFinger(); 8503 assert(ffc->size() == pointer_delta(addr, freeFinger()), 8504 "Size of free range is inconsistent with chunk size."); 8505 if (CMSTestInFreeList) { 8506 assert(_sp->verify_chunk_in_free_list(ffc), 8507 "free range is not in free lists"); 8508 } 8509 _sp->removeFreeChunkFromFreeLists(ffc); 8510 set_freeRangeInFreeLists(false); 8511 } 8512 set_lastFreeRangeCoalesced(true); 8513 } 8514 // this will be swept up when we hit the end of the free range 8515 8516 // Verify that the bit map has no bits marked between 8517 // addr and purported end of just dead object. 8518 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 8519 } 8520 assert(_limit >= addr + size, 8521 "A freshly garbage chunk can't possibly straddle over _limit"); 8522 if (inFreeRange()) lookahead_and_flush(fc, size); 8523 return size; 8524 } 8525 8526 size_t SweepClosure::do_live_chunk(FreeChunk* fc) { 8527 HeapWord* addr = (HeapWord*) fc; 8528 // The sweeper has just found a live object. Return any accumulated 8529 // left hand chunk to the free lists. 8530 if (inFreeRange()) { 8531 assert(freeFinger() < addr, "freeFinger points too high"); 8532 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 8533 } 8534 8535 // This object is live: we'd normally expect this to be 8536 // an oop, and like to assert the following: 8537 // assert(oop(addr)->is_oop(), "live block should be an oop"); 8538 // However, as we commented above, this may be an object whose 8539 // header hasn't yet been initialized. 8540 size_t size; 8541 assert(_bitMap->isMarked(addr), "Tautology for this control point"); 8542 if (_bitMap->isMarked(addr + 1)) { 8543 // Determine the size from the bit map, rather than trying to 8544 // compute it from the object header. 8545 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 8546 size = pointer_delta(nextOneAddr + 1, addr); 8547 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 8548 "alignment problem"); 8549 8550 #ifdef DEBUG 8551 if (oop(addr)->klass_or_null() != NULL && 8552 ( !_collector->should_unload_classes() 8553 || (oop(addr)->is_parsable()) && 8554 oop(addr)->is_conc_safe())) { 8555 // Ignore mark word because we are running concurrent with mutators 8556 assert(oop(addr)->is_oop(true), "live block should be an oop"); 8557 // is_conc_safe is checked before performing this assertion 8558 // because an object that is not is_conc_safe may yet have 8559 // the return from size() correct. 8560 assert(size == 8561 CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()), 8562 "P-mark and computed size do not agree"); 8563 } 8564 #endif 8565 8566 } else { 8567 // This should be an initialized object that's alive. 8568 assert(oop(addr)->klass_or_null() != NULL && 8569 (!_collector->should_unload_classes() 8570 || oop(addr)->is_parsable()), 8571 "Should be an initialized object"); 8572 // Note that there are objects used during class redefinition, 8573 // e.g. merge_cp in VM_RedefineClasses::merge_cp_and_rewrite(), 8574 // which are discarded with their is_conc_safe state still 8575 // false. These object may be floating garbage so may be 8576 // seen here. If they are floating garbage their size 8577 // should be attainable from their klass. Do not that 8578 // is_conc_safe() is true for oop(addr). 8579 // Ignore mark word because we are running concurrent with mutators 8580 assert(oop(addr)->is_oop(true), "live block should be an oop"); 8581 // Verify that the bit map has no bits marked between 8582 // addr and purported end of this block. 8583 size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 8584 assert(size >= 3, "Necessary for Printezis marks to work"); 8585 assert(!_bitMap->isMarked(addr+1), "Tautology for this control point"); 8586 DEBUG_ONLY(_bitMap->verifyNoOneBitsInRange(addr+2, addr+size);) 8587 } 8588 return size; 8589 } 8590 8591 void SweepClosure::do_post_free_or_garbage_chunk(FreeChunk* fc, 8592 size_t chunkSize) { 8593 // do_post_free_or_garbage_chunk() should only be called in the case 8594 // of the adaptive free list allocator. 8595 const bool fcInFreeLists = fc->is_free(); 8596 assert(_sp->adaptive_freelists(), "Should only be used in this case."); 8597 assert((HeapWord*)fc <= _limit, "sweep invariant"); 8598 if (CMSTestInFreeList && fcInFreeLists) { 8599 assert(_sp->verify_chunk_in_free_list(fc), "free chunk is not in free lists"); 8600 } 8601 8602 if (CMSTraceSweeper) { 8603 gclog_or_tty->print_cr(" -- pick up another chunk at 0x%x (%d)", fc, chunkSize); 8604 } 8605 8606 HeapWord* const fc_addr = (HeapWord*) fc; 8607 8608 bool coalesce; 8609 const size_t left = pointer_delta(fc_addr, freeFinger()); 8610 const size_t right = chunkSize; 8611 switch (FLSCoalescePolicy) { 8612 // numeric value forms a coalition aggressiveness metric 8613 case 0: { // never coalesce 8614 coalesce = false; 8615 break; 8616 } 8617 case 1: { // coalesce if left & right chunks on overpopulated lists 8618 coalesce = _sp->coalOverPopulated(left) && 8619 _sp->coalOverPopulated(right); 8620 break; 8621 } 8622 case 2: { // coalesce if left chunk on overpopulated list (default) 8623 coalesce = _sp->coalOverPopulated(left); 8624 break; 8625 } 8626 case 3: { // coalesce if left OR right chunk on overpopulated list 8627 coalesce = _sp->coalOverPopulated(left) || 8628 _sp->coalOverPopulated(right); 8629 break; 8630 } 8631 case 4: { // always coalesce 8632 coalesce = true; 8633 break; 8634 } 8635 default: 8636 ShouldNotReachHere(); 8637 } 8638 8639 // Should the current free range be coalesced? 8640 // If the chunk is in a free range and either we decided to coalesce above 8641 // or the chunk is near the large block at the end of the heap 8642 // (isNearLargestChunk() returns true), then coalesce this chunk. 8643 const bool doCoalesce = inFreeRange() 8644 && (coalesce || _g->isNearLargestChunk(fc_addr)); 8645 if (doCoalesce) { 8646 // Coalesce the current free range on the left with the new 8647 // chunk on the right. If either is on a free list, 8648 // it must be removed from the list and stashed in the closure. 8649 if (freeRangeInFreeLists()) { 8650 FreeChunk* const ffc = (FreeChunk*)freeFinger(); 8651 assert(ffc->size() == pointer_delta(fc_addr, freeFinger()), 8652 "Size of free range is inconsistent with chunk size."); 8653 if (CMSTestInFreeList) { 8654 assert(_sp->verify_chunk_in_free_list(ffc), 8655 "Chunk is not in free lists"); 8656 } 8657 _sp->coalDeath(ffc->size()); 8658 _sp->removeFreeChunkFromFreeLists(ffc); 8659 set_freeRangeInFreeLists(false); 8660 } 8661 if (fcInFreeLists) { 8662 _sp->coalDeath(chunkSize); 8663 assert(fc->size() == chunkSize, 8664 "The chunk has the wrong size or is not in the free lists"); 8665 _sp->removeFreeChunkFromFreeLists(fc); 8666 } 8667 set_lastFreeRangeCoalesced(true); 8668 print_free_block_coalesced(fc); 8669 } else { // not in a free range and/or should not coalesce 8670 // Return the current free range and start a new one. 8671 if (inFreeRange()) { 8672 // In a free range but cannot coalesce with the right hand chunk. 8673 // Put the current free range into the free lists. 8674 flush_cur_free_chunk(freeFinger(), 8675 pointer_delta(fc_addr, freeFinger())); 8676 } 8677 // Set up for new free range. Pass along whether the right hand 8678 // chunk is in the free lists. 8679 initialize_free_range((HeapWord*)fc, fcInFreeLists); 8680 } 8681 } 8682 8683 // Lookahead flush: 8684 // If we are tracking a free range, and this is the last chunk that 8685 // we'll look at because its end crosses past _limit, we'll preemptively 8686 // flush it along with any free range we may be holding on to. Note that 8687 // this can be the case only for an already free or freshly garbage 8688 // chunk. If this block is an object, it can never straddle 8689 // over _limit. The "straddling" occurs when _limit is set at 8690 // the previous end of the space when this cycle started, and 8691 // a subsequent heap expansion caused the previously co-terminal 8692 // free block to be coalesced with the newly expanded portion, 8693 // thus rendering _limit a non-block-boundary making it dangerous 8694 // for the sweeper to step over and examine. 8695 void SweepClosure::lookahead_and_flush(FreeChunk* fc, size_t chunk_size) { 8696 assert(inFreeRange(), "Should only be called if currently in a free range."); 8697 HeapWord* const eob = ((HeapWord*)fc) + chunk_size; 8698 assert(_sp->used_region().contains(eob - 1), 8699 err_msg("eob = " PTR_FORMAT " out of bounds wrt _sp = [" PTR_FORMAT "," PTR_FORMAT ")" 8700 " when examining fc = " PTR_FORMAT "(" SIZE_FORMAT ")", 8701 _limit, _sp->bottom(), _sp->end(), fc, chunk_size)); 8702 if (eob >= _limit) { 8703 assert(eob == _limit || fc->is_free(), "Only a free chunk should allow us to cross over the limit"); 8704 if (CMSTraceSweeper) { 8705 gclog_or_tty->print_cr("_limit " PTR_FORMAT " reached or crossed by block " 8706 "[" PTR_FORMAT "," PTR_FORMAT ") in space " 8707 "[" PTR_FORMAT "," PTR_FORMAT ")", 8708 _limit, fc, eob, _sp->bottom(), _sp->end()); 8709 } 8710 // Return the storage we are tracking back into the free lists. 8711 if (CMSTraceSweeper) { 8712 gclog_or_tty->print_cr("Flushing ... "); 8713 } 8714 assert(freeFinger() < eob, "Error"); 8715 flush_cur_free_chunk( freeFinger(), pointer_delta(eob, freeFinger())); 8716 } 8717 } 8718 8719 void SweepClosure::flush_cur_free_chunk(HeapWord* chunk, size_t size) { 8720 assert(inFreeRange(), "Should only be called if currently in a free range."); 8721 assert(size > 0, 8722 "A zero sized chunk cannot be added to the free lists."); 8723 if (!freeRangeInFreeLists()) { 8724 if (CMSTestInFreeList) { 8725 FreeChunk* fc = (FreeChunk*) chunk; 8726 fc->set_size(size); 8727 assert(!_sp->verify_chunk_in_free_list(fc), 8728 "chunk should not be in free lists yet"); 8729 } 8730 if (CMSTraceSweeper) { 8731 gclog_or_tty->print_cr(" -- add free block 0x%x (%d) to free lists", 8732 chunk, size); 8733 } 8734 // A new free range is going to be starting. The current 8735 // free range has not been added to the free lists yet or 8736 // was removed so add it back. 8737 // If the current free range was coalesced, then the death 8738 // of the free range was recorded. Record a birth now. 8739 if (lastFreeRangeCoalesced()) { 8740 _sp->coalBirth(size); 8741 } 8742 _sp->addChunkAndRepairOffsetTable(chunk, size, 8743 lastFreeRangeCoalesced()); 8744 } else if (CMSTraceSweeper) { 8745 gclog_or_tty->print_cr("Already in free list: nothing to flush"); 8746 } 8747 set_inFreeRange(false); 8748 set_freeRangeInFreeLists(false); 8749 } 8750 8751 // We take a break if we've been at this for a while, 8752 // so as to avoid monopolizing the locks involved. 8753 void SweepClosure::do_yield_work(HeapWord* addr) { 8754 // Return current free chunk being used for coalescing (if any) 8755 // to the appropriate freelist. After yielding, the next 8756 // free block encountered will start a coalescing range of 8757 // free blocks. If the next free block is adjacent to the 8758 // chunk just flushed, they will need to wait for the next 8759 // sweep to be coalesced. 8760 if (inFreeRange()) { 8761 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 8762 } 8763 8764 // First give up the locks, then yield, then re-lock. 8765 // We should probably use a constructor/destructor idiom to 8766 // do this unlock/lock or modify the MutexUnlocker class to 8767 // serve our purpose. XXX 8768 assert_lock_strong(_bitMap->lock()); 8769 assert_lock_strong(_freelistLock); 8770 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 8771 "CMS thread should hold CMS token"); 8772 _bitMap->lock()->unlock(); 8773 _freelistLock->unlock(); 8774 ConcurrentMarkSweepThread::desynchronize(true); 8775 ConcurrentMarkSweepThread::acknowledge_yield_request(); 8776 _collector->stopTimer(); 8777 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 8778 if (PrintCMSStatistics != 0) { 8779 _collector->incrementYields(); 8780 } 8781 _collector->icms_wait(); 8782 8783 // See the comment in coordinator_yield() 8784 for (unsigned i = 0; i < CMSYieldSleepCount && 8785 ConcurrentMarkSweepThread::should_yield() && 8786 !CMSCollector::foregroundGCIsActive(); ++i) { 8787 os::sleep(Thread::current(), 1, false); 8788 ConcurrentMarkSweepThread::acknowledge_yield_request(); 8789 } 8790 8791 ConcurrentMarkSweepThread::synchronize(true); 8792 _freelistLock->lock(); 8793 _bitMap->lock()->lock_without_safepoint_check(); 8794 _collector->startTimer(); 8795 } 8796 8797 #ifndef PRODUCT 8798 // This is actually very useful in a product build if it can 8799 // be called from the debugger. Compile it into the product 8800 // as needed. 8801 bool debug_verify_chunk_in_free_list(FreeChunk* fc) { 8802 return debug_cms_space->verify_chunk_in_free_list(fc); 8803 } 8804 #endif 8805 8806 void SweepClosure::print_free_block_coalesced(FreeChunk* fc) const { 8807 if (CMSTraceSweeper) { 8808 gclog_or_tty->print_cr("Sweep:coal_free_blk " PTR_FORMAT " (" SIZE_FORMAT ")", 8809 fc, fc->size()); 8810 } 8811 } 8812 8813 // CMSIsAliveClosure 8814 bool CMSIsAliveClosure::do_object_b(oop obj) { 8815 HeapWord* addr = (HeapWord*)obj; 8816 return addr != NULL && 8817 (!_span.contains(addr) || _bit_map->isMarked(addr)); 8818 } 8819 8820 CMSKeepAliveClosure::CMSKeepAliveClosure( CMSCollector* collector, 8821 MemRegion span, 8822 CMSBitMap* bit_map, CMSMarkStack* mark_stack, 8823 CMSMarkStack* revisit_stack, bool cpc): 8824 KlassRememberingOopClosure(collector, NULL, revisit_stack), 8825 _span(span), 8826 _bit_map(bit_map), 8827 _mark_stack(mark_stack), 8828 _concurrent_precleaning(cpc) { 8829 assert(!_span.is_empty(), "Empty span could spell trouble"); 8830 } 8831 8832 8833 // CMSKeepAliveClosure: the serial version 8834 void CMSKeepAliveClosure::do_oop(oop obj) { 8835 HeapWord* addr = (HeapWord*)obj; 8836 if (_span.contains(addr) && 8837 !_bit_map->isMarked(addr)) { 8838 _bit_map->mark(addr); 8839 bool simulate_overflow = false; 8840 NOT_PRODUCT( 8841 if (CMSMarkStackOverflowALot && 8842 _collector->simulate_overflow()) { 8843 // simulate a stack overflow 8844 simulate_overflow = true; 8845 } 8846 ) 8847 if (simulate_overflow || !_mark_stack->push(obj)) { 8848 if (_concurrent_precleaning) { 8849 // We dirty the overflown object and let the remark 8850 // phase deal with it. 8851 assert(_collector->overflow_list_is_empty(), "Error"); 8852 // In the case of object arrays, we need to dirty all of 8853 // the cards that the object spans. No locking or atomics 8854 // are needed since no one else can be mutating the mod union 8855 // table. 8856 if (obj->is_objArray()) { 8857 size_t sz = obj->size(); 8858 HeapWord* end_card_addr = 8859 (HeapWord*)round_to((intptr_t)(addr+sz), CardTableModRefBS::card_size); 8860 MemRegion redirty_range = MemRegion(addr, end_card_addr); 8861 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 8862 _collector->_modUnionTable.mark_range(redirty_range); 8863 } else { 8864 _collector->_modUnionTable.mark(addr); 8865 } 8866 _collector->_ser_kac_preclean_ovflw++; 8867 } else { 8868 _collector->push_on_overflow_list(obj); 8869 _collector->_ser_kac_ovflw++; 8870 } 8871 } 8872 } 8873 } 8874 8875 void CMSKeepAliveClosure::do_oop(oop* p) { CMSKeepAliveClosure::do_oop_work(p); } 8876 void CMSKeepAliveClosure::do_oop(narrowOop* p) { CMSKeepAliveClosure::do_oop_work(p); } 8877 8878 // CMSParKeepAliveClosure: a parallel version of the above. 8879 // The work queues are private to each closure (thread), 8880 // but (may be) available for stealing by other threads. 8881 void CMSParKeepAliveClosure::do_oop(oop obj) { 8882 HeapWord* addr = (HeapWord*)obj; 8883 if (_span.contains(addr) && 8884 !_bit_map->isMarked(addr)) { 8885 // In general, during recursive tracing, several threads 8886 // may be concurrently getting here; the first one to 8887 // "tag" it, claims it. 8888 if (_bit_map->par_mark(addr)) { 8889 bool res = _work_queue->push(obj); 8890 assert(res, "Low water mark should be much less than capacity"); 8891 // Do a recursive trim in the hope that this will keep 8892 // stack usage lower, but leave some oops for potential stealers 8893 trim_queue(_low_water_mark); 8894 } // Else, another thread got there first 8895 } 8896 } 8897 8898 void CMSParKeepAliveClosure::do_oop(oop* p) { CMSParKeepAliveClosure::do_oop_work(p); } 8899 void CMSParKeepAliveClosure::do_oop(narrowOop* p) { CMSParKeepAliveClosure::do_oop_work(p); } 8900 8901 void CMSParKeepAliveClosure::trim_queue(uint max) { 8902 while (_work_queue->size() > max) { 8903 oop new_oop; 8904 if (_work_queue->pop_local(new_oop)) { 8905 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 8906 assert(_bit_map->isMarked((HeapWord*)new_oop), 8907 "no white objects on this stack!"); 8908 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 8909 // iterate over the oops in this oop, marking and pushing 8910 // the ones in CMS heap (i.e. in _span). 8911 new_oop->oop_iterate(&_mark_and_push); 8912 } 8913 } 8914 } 8915 8916 CMSInnerParMarkAndPushClosure::CMSInnerParMarkAndPushClosure( 8917 CMSCollector* collector, 8918 MemRegion span, CMSBitMap* bit_map, 8919 CMSMarkStack* revisit_stack, 8920 OopTaskQueue* work_queue): 8921 Par_KlassRememberingOopClosure(collector, NULL, revisit_stack), 8922 _span(span), 8923 _bit_map(bit_map), 8924 _work_queue(work_queue) { } 8925 8926 void CMSInnerParMarkAndPushClosure::do_oop(oop obj) { 8927 HeapWord* addr = (HeapWord*)obj; 8928 if (_span.contains(addr) && 8929 !_bit_map->isMarked(addr)) { 8930 if (_bit_map->par_mark(addr)) { 8931 bool simulate_overflow = false; 8932 NOT_PRODUCT( 8933 if (CMSMarkStackOverflowALot && 8934 _collector->par_simulate_overflow()) { 8935 // simulate a stack overflow 8936 simulate_overflow = true; 8937 } 8938 ) 8939 if (simulate_overflow || !_work_queue->push(obj)) { 8940 _collector->par_push_on_overflow_list(obj); 8941 _collector->_par_kac_ovflw++; 8942 } 8943 } // Else another thread got there already 8944 } 8945 } 8946 8947 void CMSInnerParMarkAndPushClosure::do_oop(oop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); } 8948 void CMSInnerParMarkAndPushClosure::do_oop(narrowOop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); } 8949 8950 ////////////////////////////////////////////////////////////////// 8951 // CMSExpansionCause ///////////////////////////// 8952 ////////////////////////////////////////////////////////////////// 8953 const char* CMSExpansionCause::to_string(CMSExpansionCause::Cause cause) { 8954 switch (cause) { 8955 case _no_expansion: 8956 return "No expansion"; 8957 case _satisfy_free_ratio: 8958 return "Free ratio"; 8959 case _satisfy_promotion: 8960 return "Satisfy promotion"; 8961 case _satisfy_allocation: 8962 return "allocation"; 8963 case _allocate_par_lab: 8964 return "Par LAB"; 8965 case _allocate_par_spooling_space: 8966 return "Par Spooling Space"; 8967 case _adaptive_size_policy: 8968 return "Ergonomics"; 8969 default: 8970 return "unknown"; 8971 } 8972 } 8973 8974 void CMSDrainMarkingStackClosure::do_void() { 8975 // the max number to take from overflow list at a time 8976 const size_t num = _mark_stack->capacity()/4; 8977 assert(!_concurrent_precleaning || _collector->overflow_list_is_empty(), 8978 "Overflow list should be NULL during concurrent phases"); 8979 while (!_mark_stack->isEmpty() || 8980 // if stack is empty, check the overflow list 8981 _collector->take_from_overflow_list(num, _mark_stack)) { 8982 oop obj = _mark_stack->pop(); 8983 HeapWord* addr = (HeapWord*)obj; 8984 assert(_span.contains(addr), "Should be within span"); 8985 assert(_bit_map->isMarked(addr), "Should be marked"); 8986 assert(obj->is_oop(), "Should be an oop"); 8987 obj->oop_iterate(_keep_alive); 8988 } 8989 } 8990 8991 void CMSParDrainMarkingStackClosure::do_void() { 8992 // drain queue 8993 trim_queue(0); 8994 } 8995 8996 // Trim our work_queue so its length is below max at return 8997 void CMSParDrainMarkingStackClosure::trim_queue(uint max) { 8998 while (_work_queue->size() > max) { 8999 oop new_oop; 9000 if (_work_queue->pop_local(new_oop)) { 9001 assert(new_oop->is_oop(), "Expected an oop"); 9002 assert(_bit_map->isMarked((HeapWord*)new_oop), 9003 "no white objects on this stack!"); 9004 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 9005 // iterate over the oops in this oop, marking and pushing 9006 // the ones in CMS heap (i.e. in _span). 9007 new_oop->oop_iterate(&_mark_and_push); 9008 } 9009 } 9010 } 9011 9012 //////////////////////////////////////////////////////////////////// 9013 // Support for Marking Stack Overflow list handling and related code 9014 //////////////////////////////////////////////////////////////////// 9015 // Much of the following code is similar in shape and spirit to the 9016 // code used in ParNewGC. We should try and share that code 9017 // as much as possible in the future. 9018 9019 #ifndef PRODUCT 9020 // Debugging support for CMSStackOverflowALot 9021 9022 // It's OK to call this multi-threaded; the worst thing 9023 // that can happen is that we'll get a bunch of closely 9024 // spaced simulated oveflows, but that's OK, in fact 9025 // probably good as it would exercise the overflow code 9026 // under contention. 9027 bool CMSCollector::simulate_overflow() { 9028 if (_overflow_counter-- <= 0) { // just being defensive 9029 _overflow_counter = CMSMarkStackOverflowInterval; 9030 return true; 9031 } else { 9032 return false; 9033 } 9034 } 9035 9036 bool CMSCollector::par_simulate_overflow() { 9037 return simulate_overflow(); 9038 } 9039 #endif 9040 9041 // Single-threaded 9042 bool CMSCollector::take_from_overflow_list(size_t num, CMSMarkStack* stack) { 9043 assert(stack->isEmpty(), "Expected precondition"); 9044 assert(stack->capacity() > num, "Shouldn't bite more than can chew"); 9045 size_t i = num; 9046 oop cur = _overflow_list; 9047 const markOop proto = markOopDesc::prototype(); 9048 NOT_PRODUCT(ssize_t n = 0;) 9049 for (oop next; i > 0 && cur != NULL; cur = next, i--) { 9050 next = oop(cur->mark()); 9051 cur->set_mark(proto); // until proven otherwise 9052 assert(cur->is_oop(), "Should be an oop"); 9053 bool res = stack->push(cur); 9054 assert(res, "Bit off more than can chew?"); 9055 NOT_PRODUCT(n++;) 9056 } 9057 _overflow_list = cur; 9058 #ifndef PRODUCT 9059 assert(_num_par_pushes >= n, "Too many pops?"); 9060 _num_par_pushes -=n; 9061 #endif 9062 return !stack->isEmpty(); 9063 } 9064 9065 #define BUSY (oop(0x1aff1aff)) 9066 // (MT-safe) Get a prefix of at most "num" from the list. 9067 // The overflow list is chained through the mark word of 9068 // each object in the list. We fetch the entire list, 9069 // break off a prefix of the right size and return the 9070 // remainder. If other threads try to take objects from 9071 // the overflow list at that time, they will wait for 9072 // some time to see if data becomes available. If (and 9073 // only if) another thread places one or more object(s) 9074 // on the global list before we have returned the suffix 9075 // to the global list, we will walk down our local list 9076 // to find its end and append the global list to 9077 // our suffix before returning it. This suffix walk can 9078 // prove to be expensive (quadratic in the amount of traffic) 9079 // when there are many objects in the overflow list and 9080 // there is much producer-consumer contention on the list. 9081 // *NOTE*: The overflow list manipulation code here and 9082 // in ParNewGeneration:: are very similar in shape, 9083 // except that in the ParNew case we use the old (from/eden) 9084 // copy of the object to thread the list via its klass word. 9085 // Because of the common code, if you make any changes in 9086 // the code below, please check the ParNew version to see if 9087 // similar changes might be needed. 9088 // CR 6797058 has been filed to consolidate the common code. 9089 bool CMSCollector::par_take_from_overflow_list(size_t num, 9090 OopTaskQueue* work_q, 9091 int no_of_gc_threads) { 9092 assert(work_q->size() == 0, "First empty local work queue"); 9093 assert(num < work_q->max_elems(), "Can't bite more than we can chew"); 9094 if (_overflow_list == NULL) { 9095 return false; 9096 } 9097 // Grab the entire list; we'll put back a suffix 9098 oop prefix = (oop)Atomic::xchg_ptr(BUSY, &_overflow_list); 9099 Thread* tid = Thread::current(); 9100 // Before "no_of_gc_threads" was introduced CMSOverflowSpinCount was 9101 // set to ParallelGCThreads. 9102 size_t CMSOverflowSpinCount = (size_t) no_of_gc_threads; // was ParallelGCThreads; 9103 size_t sleep_time_millis = MAX2((size_t)1, num/100); 9104 // If the list is busy, we spin for a short while, 9105 // sleeping between attempts to get the list. 9106 for (size_t spin = 0; prefix == BUSY && spin < CMSOverflowSpinCount; spin++) { 9107 os::sleep(tid, sleep_time_millis, false); 9108 if (_overflow_list == NULL) { 9109 // Nothing left to take 9110 return false; 9111 } else if (_overflow_list != BUSY) { 9112 // Try and grab the prefix 9113 prefix = (oop)Atomic::xchg_ptr(BUSY, &_overflow_list); 9114 } 9115 } 9116 // If the list was found to be empty, or we spun long 9117 // enough, we give up and return empty-handed. If we leave 9118 // the list in the BUSY state below, it must be the case that 9119 // some other thread holds the overflow list and will set it 9120 // to a non-BUSY state in the future. 9121 if (prefix == NULL || prefix == BUSY) { 9122 // Nothing to take or waited long enough 9123 if (prefix == NULL) { 9124 // Write back the NULL in case we overwrote it with BUSY above 9125 // and it is still the same value. 9126 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY); 9127 } 9128 return false; 9129 } 9130 assert(prefix != NULL && prefix != BUSY, "Error"); 9131 size_t i = num; 9132 oop cur = prefix; 9133 // Walk down the first "num" objects, unless we reach the end. 9134 for (; i > 1 && cur->mark() != NULL; cur = oop(cur->mark()), i--); 9135 if (cur->mark() == NULL) { 9136 // We have "num" or fewer elements in the list, so there 9137 // is nothing to return to the global list. 9138 // Write back the NULL in lieu of the BUSY we wrote 9139 // above, if it is still the same value. 9140 if (_overflow_list == BUSY) { 9141 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY); 9142 } 9143 } else { 9144 // Chop off the suffix and rerturn it to the global list. 9145 assert(cur->mark() != BUSY, "Error"); 9146 oop suffix_head = cur->mark(); // suffix will be put back on global list 9147 cur->set_mark(NULL); // break off suffix 9148 // It's possible that the list is still in the empty(busy) state 9149 // we left it in a short while ago; in that case we may be 9150 // able to place back the suffix without incurring the cost 9151 // of a walk down the list. 9152 oop observed_overflow_list = _overflow_list; 9153 oop cur_overflow_list = observed_overflow_list; 9154 bool attached = false; 9155 while (observed_overflow_list == BUSY || observed_overflow_list == NULL) { 9156 observed_overflow_list = 9157 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list); 9158 if (cur_overflow_list == observed_overflow_list) { 9159 attached = true; 9160 break; 9161 } else cur_overflow_list = observed_overflow_list; 9162 } 9163 if (!attached) { 9164 // Too bad, someone else sneaked in (at least) an element; we'll need 9165 // to do a splice. Find tail of suffix so we can prepend suffix to global 9166 // list. 9167 for (cur = suffix_head; cur->mark() != NULL; cur = (oop)(cur->mark())); 9168 oop suffix_tail = cur; 9169 assert(suffix_tail != NULL && suffix_tail->mark() == NULL, 9170 "Tautology"); 9171 observed_overflow_list = _overflow_list; 9172 do { 9173 cur_overflow_list = observed_overflow_list; 9174 if (cur_overflow_list != BUSY) { 9175 // Do the splice ... 9176 suffix_tail->set_mark(markOop(cur_overflow_list)); 9177 } else { // cur_overflow_list == BUSY 9178 suffix_tail->set_mark(NULL); 9179 } 9180 // ... and try to place spliced list back on overflow_list ... 9181 observed_overflow_list = 9182 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list); 9183 } while (cur_overflow_list != observed_overflow_list); 9184 // ... until we have succeeded in doing so. 9185 } 9186 } 9187 9188 // Push the prefix elements on work_q 9189 assert(prefix != NULL, "control point invariant"); 9190 const markOop proto = markOopDesc::prototype(); 9191 oop next; 9192 NOT_PRODUCT(ssize_t n = 0;) 9193 for (cur = prefix; cur != NULL; cur = next) { 9194 next = oop(cur->mark()); 9195 cur->set_mark(proto); // until proven otherwise 9196 assert(cur->is_oop(), "Should be an oop"); 9197 bool res = work_q->push(cur); 9198 assert(res, "Bit off more than we can chew?"); 9199 NOT_PRODUCT(n++;) 9200 } 9201 #ifndef PRODUCT 9202 assert(_num_par_pushes >= n, "Too many pops?"); 9203 Atomic::add_ptr(-(intptr_t)n, &_num_par_pushes); 9204 #endif 9205 return true; 9206 } 9207 9208 // Single-threaded 9209 void CMSCollector::push_on_overflow_list(oop p) { 9210 NOT_PRODUCT(_num_par_pushes++;) 9211 assert(p->is_oop(), "Not an oop"); 9212 preserve_mark_if_necessary(p); 9213 p->set_mark((markOop)_overflow_list); 9214 _overflow_list = p; 9215 } 9216 9217 // Multi-threaded; use CAS to prepend to overflow list 9218 void CMSCollector::par_push_on_overflow_list(oop p) { 9219 NOT_PRODUCT(Atomic::inc_ptr(&_num_par_pushes);) 9220 assert(p->is_oop(), "Not an oop"); 9221 par_preserve_mark_if_necessary(p); 9222 oop observed_overflow_list = _overflow_list; 9223 oop cur_overflow_list; 9224 do { 9225 cur_overflow_list = observed_overflow_list; 9226 if (cur_overflow_list != BUSY) { 9227 p->set_mark(markOop(cur_overflow_list)); 9228 } else { 9229 p->set_mark(NULL); 9230 } 9231 observed_overflow_list = 9232 (oop) Atomic::cmpxchg_ptr(p, &_overflow_list, cur_overflow_list); 9233 } while (cur_overflow_list != observed_overflow_list); 9234 } 9235 #undef BUSY 9236 9237 // Single threaded 9238 // General Note on GrowableArray: pushes may silently fail 9239 // because we are (temporarily) out of C-heap for expanding 9240 // the stack. The problem is quite ubiquitous and affects 9241 // a lot of code in the JVM. The prudent thing for GrowableArray 9242 // to do (for now) is to exit with an error. However, that may 9243 // be too draconian in some cases because the caller may be 9244 // able to recover without much harm. For such cases, we 9245 // should probably introduce a "soft_push" method which returns 9246 // an indication of success or failure with the assumption that 9247 // the caller may be able to recover from a failure; code in 9248 // the VM can then be changed, incrementally, to deal with such 9249 // failures where possible, thus, incrementally hardening the VM 9250 // in such low resource situations. 9251 void CMSCollector::preserve_mark_work(oop p, markOop m) { 9252 _preserved_oop_stack.push(p); 9253 _preserved_mark_stack.push(m); 9254 assert(m == p->mark(), "Mark word changed"); 9255 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), 9256 "bijection"); 9257 } 9258 9259 // Single threaded 9260 void CMSCollector::preserve_mark_if_necessary(oop p) { 9261 markOop m = p->mark(); 9262 if (m->must_be_preserved(p)) { 9263 preserve_mark_work(p, m); 9264 } 9265 } 9266 9267 void CMSCollector::par_preserve_mark_if_necessary(oop p) { 9268 markOop m = p->mark(); 9269 if (m->must_be_preserved(p)) { 9270 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag); 9271 // Even though we read the mark word without holding 9272 // the lock, we are assured that it will not change 9273 // because we "own" this oop, so no other thread can 9274 // be trying to push it on the overflow list; see 9275 // the assertion in preserve_mark_work() that checks 9276 // that m == p->mark(). 9277 preserve_mark_work(p, m); 9278 } 9279 } 9280 9281 // We should be able to do this multi-threaded, 9282 // a chunk of stack being a task (this is 9283 // correct because each oop only ever appears 9284 // once in the overflow list. However, it's 9285 // not very easy to completely overlap this with 9286 // other operations, so will generally not be done 9287 // until all work's been completed. Because we 9288 // expect the preserved oop stack (set) to be small, 9289 // it's probably fine to do this single-threaded. 9290 // We can explore cleverer concurrent/overlapped/parallel 9291 // processing of preserved marks if we feel the 9292 // need for this in the future. Stack overflow should 9293 // be so rare in practice and, when it happens, its 9294 // effect on performance so great that this will 9295 // likely just be in the noise anyway. 9296 void CMSCollector::restore_preserved_marks_if_any() { 9297 assert(SafepointSynchronize::is_at_safepoint(), 9298 "world should be stopped"); 9299 assert(Thread::current()->is_ConcurrentGC_thread() || 9300 Thread::current()->is_VM_thread(), 9301 "should be single-threaded"); 9302 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), 9303 "bijection"); 9304 9305 while (!_preserved_oop_stack.is_empty()) { 9306 oop p = _preserved_oop_stack.pop(); 9307 assert(p->is_oop(), "Should be an oop"); 9308 assert(_span.contains(p), "oop should be in _span"); 9309 assert(p->mark() == markOopDesc::prototype(), 9310 "Set when taken from overflow list"); 9311 markOop m = _preserved_mark_stack.pop(); 9312 p->set_mark(m); 9313 } 9314 assert(_preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(), 9315 "stacks were cleared above"); 9316 } 9317 9318 #ifndef PRODUCT 9319 bool CMSCollector::no_preserved_marks() const { 9320 return _preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(); 9321 } 9322 #endif 9323 9324 CMSAdaptiveSizePolicy* ASConcurrentMarkSweepGeneration::cms_size_policy() const 9325 { 9326 GenCollectedHeap* gch = (GenCollectedHeap*) GenCollectedHeap::heap(); 9327 CMSAdaptiveSizePolicy* size_policy = 9328 (CMSAdaptiveSizePolicy*) gch->gen_policy()->size_policy(); 9329 assert(size_policy->is_gc_cms_adaptive_size_policy(), 9330 "Wrong type for size policy"); 9331 return size_policy; 9332 } 9333 9334 void ASConcurrentMarkSweepGeneration::resize(size_t cur_promo_size, 9335 size_t desired_promo_size) { 9336 if (cur_promo_size < desired_promo_size) { 9337 size_t expand_bytes = desired_promo_size - cur_promo_size; 9338 if (PrintAdaptiveSizePolicy && Verbose) { 9339 gclog_or_tty->print_cr(" ASConcurrentMarkSweepGeneration::resize " 9340 "Expanding tenured generation by " SIZE_FORMAT " (bytes)", 9341 expand_bytes); 9342 } 9343 expand(expand_bytes, 9344 MinHeapDeltaBytes, 9345 CMSExpansionCause::_adaptive_size_policy); 9346 } else if (desired_promo_size < cur_promo_size) { 9347 size_t shrink_bytes = cur_promo_size - desired_promo_size; 9348 if (PrintAdaptiveSizePolicy && Verbose) { 9349 gclog_or_tty->print_cr(" ASConcurrentMarkSweepGeneration::resize " 9350 "Shrinking tenured generation by " SIZE_FORMAT " (bytes)", 9351 shrink_bytes); 9352 } 9353 shrink(shrink_bytes); 9354 } 9355 } 9356 9357 CMSGCAdaptivePolicyCounters* ASConcurrentMarkSweepGeneration::gc_adaptive_policy_counters() { 9358 GenCollectedHeap* gch = GenCollectedHeap::heap(); 9359 CMSGCAdaptivePolicyCounters* counters = 9360 (CMSGCAdaptivePolicyCounters*) gch->collector_policy()->counters(); 9361 assert(counters->kind() == GCPolicyCounters::CMSGCAdaptivePolicyCountersKind, 9362 "Wrong kind of counters"); 9363 return counters; 9364 } 9365 9366 9367 void ASConcurrentMarkSweepGeneration::update_counters() { 9368 if (UsePerfData) { 9369 _space_counters->update_all(); 9370 _gen_counters->update_all(); 9371 CMSGCAdaptivePolicyCounters* counters = gc_adaptive_policy_counters(); 9372 GenCollectedHeap* gch = GenCollectedHeap::heap(); 9373 CMSGCStats* gc_stats_l = (CMSGCStats*) gc_stats(); 9374 assert(gc_stats_l->kind() == GCStats::CMSGCStatsKind, 9375 "Wrong gc statistics type"); 9376 counters->update_counters(gc_stats_l); 9377 } 9378 } 9379 9380 void ASConcurrentMarkSweepGeneration::update_counters(size_t used) { 9381 if (UsePerfData) { 9382 _space_counters->update_used(used); 9383 _space_counters->update_capacity(); 9384 _gen_counters->update_all(); 9385 9386 CMSGCAdaptivePolicyCounters* counters = gc_adaptive_policy_counters(); 9387 GenCollectedHeap* gch = GenCollectedHeap::heap(); 9388 CMSGCStats* gc_stats_l = (CMSGCStats*) gc_stats(); 9389 assert(gc_stats_l->kind() == GCStats::CMSGCStatsKind, 9390 "Wrong gc statistics type"); 9391 counters->update_counters(gc_stats_l); 9392 } 9393 } 9394 9395 // The desired expansion delta is computed so that: 9396 // . desired free percentage or greater is used 9397 void ASConcurrentMarkSweepGeneration::compute_new_size() { 9398 assert_locked_or_safepoint(Heap_lock); 9399 9400 GenCollectedHeap* gch = (GenCollectedHeap*) GenCollectedHeap::heap(); 9401 9402 // If incremental collection failed, we just want to expand 9403 // to the limit. 9404 if (incremental_collection_failed()) { 9405 clear_incremental_collection_failed(); 9406 grow_to_reserved(); 9407 return; 9408 } 9409 9410 assert(UseAdaptiveSizePolicy, "Should be using adaptive sizing"); 9411 9412 assert(gch->kind() == CollectedHeap::GenCollectedHeap, 9413 "Wrong type of heap"); 9414 int prev_level = level() - 1; 9415 assert(prev_level >= 0, "The cms generation is the lowest generation"); 9416 Generation* prev_gen = gch->get_gen(prev_level); 9417 assert(prev_gen->kind() == Generation::ASParNew, 9418 "Wrong type of young generation"); 9419 ParNewGeneration* younger_gen = (ParNewGeneration*) prev_gen; 9420 size_t cur_eden = younger_gen->eden()->capacity(); 9421 CMSAdaptiveSizePolicy* size_policy = cms_size_policy(); 9422 size_t cur_promo = free(); 9423 size_policy->compute_tenured_generation_free_space(cur_promo, 9424 max_available(), 9425 cur_eden); 9426 resize(cur_promo, size_policy->promo_size()); 9427 9428 // Record the new size of the space in the cms generation 9429 // that is available for promotions. This is temporary. 9430 // It should be the desired promo size. 9431 size_policy->avg_cms_promo()->sample(free()); 9432 size_policy->avg_old_live()->sample(used()); 9433 9434 if (UsePerfData) { 9435 CMSGCAdaptivePolicyCounters* counters = gc_adaptive_policy_counters(); 9436 counters->update_cms_capacity_counter(capacity()); 9437 } 9438 } 9439 9440 void ASConcurrentMarkSweepGeneration::shrink_by(size_t desired_bytes) { 9441 assert_locked_or_safepoint(Heap_lock); 9442 assert_lock_strong(freelistLock()); 9443 HeapWord* old_end = _cmsSpace->end(); 9444 HeapWord* unallocated_start = _cmsSpace->unallocated_block(); 9445 assert(old_end >= unallocated_start, "Miscalculation of unallocated_start"); 9446 FreeChunk* chunk_at_end = find_chunk_at_end(); 9447 if (chunk_at_end == NULL) { 9448 // No room to shrink 9449 if (PrintGCDetails && Verbose) { 9450 gclog_or_tty->print_cr("No room to shrink: old_end " 9451 PTR_FORMAT " unallocated_start " PTR_FORMAT 9452 " chunk_at_end " PTR_FORMAT, 9453 old_end, unallocated_start, chunk_at_end); 9454 } 9455 return; 9456 } else { 9457 9458 // Find the chunk at the end of the space and determine 9459 // how much it can be shrunk. 9460 size_t shrinkable_size_in_bytes = chunk_at_end->size(); 9461 size_t aligned_shrinkable_size_in_bytes = 9462 align_size_down(shrinkable_size_in_bytes, os::vm_page_size()); 9463 assert(unallocated_start <= chunk_at_end->end(), 9464 "Inconsistent chunk at end of space"); 9465 size_t bytes = MIN2(desired_bytes, aligned_shrinkable_size_in_bytes); 9466 size_t word_size_before = heap_word_size(_virtual_space.committed_size()); 9467 9468 // Shrink the underlying space 9469 _virtual_space.shrink_by(bytes); 9470 if (PrintGCDetails && Verbose) { 9471 gclog_or_tty->print_cr("ConcurrentMarkSweepGeneration::shrink_by:" 9472 " desired_bytes " SIZE_FORMAT 9473 " shrinkable_size_in_bytes " SIZE_FORMAT 9474 " aligned_shrinkable_size_in_bytes " SIZE_FORMAT 9475 " bytes " SIZE_FORMAT, 9476 desired_bytes, shrinkable_size_in_bytes, 9477 aligned_shrinkable_size_in_bytes, bytes); 9478 gclog_or_tty->print_cr(" old_end " SIZE_FORMAT 9479 " unallocated_start " SIZE_FORMAT, 9480 old_end, unallocated_start); 9481 } 9482 9483 // If the space did shrink (shrinking is not guaranteed), 9484 // shrink the chunk at the end by the appropriate amount. 9485 if (((HeapWord*)_virtual_space.high()) < old_end) { 9486 size_t new_word_size = 9487 heap_word_size(_virtual_space.committed_size()); 9488 9489 // Have to remove the chunk from the dictionary because it is changing 9490 // size and might be someplace elsewhere in the dictionary. 9491 9492 // Get the chunk at end, shrink it, and put it 9493 // back. 9494 _cmsSpace->removeChunkFromDictionary(chunk_at_end); 9495 size_t word_size_change = word_size_before - new_word_size; 9496 size_t chunk_at_end_old_size = chunk_at_end->size(); 9497 assert(chunk_at_end_old_size >= word_size_change, 9498 "Shrink is too large"); 9499 chunk_at_end->set_size(chunk_at_end_old_size - 9500 word_size_change); 9501 _cmsSpace->freed((HeapWord*) chunk_at_end->end(), 9502 word_size_change); 9503 9504 _cmsSpace->returnChunkToDictionary(chunk_at_end); 9505 9506 MemRegion mr(_cmsSpace->bottom(), new_word_size); 9507 _bts->resize(new_word_size); // resize the block offset shared array 9508 Universe::heap()->barrier_set()->resize_covered_region(mr); 9509 _cmsSpace->assert_locked(); 9510 _cmsSpace->set_end((HeapWord*)_virtual_space.high()); 9511 9512 NOT_PRODUCT(_cmsSpace->dictionary()->verify()); 9513 9514 // update the space and generation capacity counters 9515 if (UsePerfData) { 9516 _space_counters->update_capacity(); 9517 _gen_counters->update_all(); 9518 } 9519 9520 if (Verbose && PrintGCDetails) { 9521 size_t new_mem_size = _virtual_space.committed_size(); 9522 size_t old_mem_size = new_mem_size + bytes; 9523 gclog_or_tty->print_cr("Shrinking %s from %ldK by %ldK to %ldK", 9524 name(), old_mem_size/K, bytes/K, new_mem_size/K); 9525 } 9526 } 9527 9528 assert(_cmsSpace->unallocated_block() <= _cmsSpace->end(), 9529 "Inconsistency at end of space"); 9530 assert(chunk_at_end->end() == _cmsSpace->end(), 9531 "Shrinking is inconsistent"); 9532 return; 9533 } 9534 } 9535 // Transfer some number of overflown objects to usual marking 9536 // stack. Return true if some objects were transferred. 9537 bool MarkRefsIntoAndScanClosure::take_from_overflow_list() { 9538 size_t num = MIN2((size_t)(_mark_stack->capacity() - _mark_stack->length())/4, 9539 (size_t)ParGCDesiredObjsFromOverflowList); 9540 9541 bool res = _collector->take_from_overflow_list(num, _mark_stack); 9542 assert(_collector->overflow_list_is_empty() || res, 9543 "If list is not empty, we should have taken something"); 9544 assert(!res || !_mark_stack->isEmpty(), 9545 "If we took something, it should now be on our stack"); 9546 return res; 9547 } 9548 9549 size_t MarkDeadObjectsClosure::do_blk(HeapWord* addr) { 9550 size_t res = _sp->block_size_no_stall(addr, _collector); 9551 if (_sp->block_is_obj(addr)) { 9552 if (_live_bit_map->isMarked(addr)) { 9553 // It can't have been dead in a previous cycle 9554 guarantee(!_dead_bit_map->isMarked(addr), "No resurrection!"); 9555 } else { 9556 _dead_bit_map->mark(addr); // mark the dead object 9557 } 9558 } 9559 // Could be 0, if the block size could not be computed without stalling. 9560 return res; 9561 } 9562 9563 TraceCMSMemoryManagerStats::TraceCMSMemoryManagerStats(CMSCollector::CollectorState phase, GCCause::Cause cause): TraceMemoryManagerStats() { 9564 9565 switch (phase) { 9566 case CMSCollector::InitialMarking: 9567 initialize(true /* fullGC */ , 9568 cause /* cause of the GC */, 9569 true /* recordGCBeginTime */, 9570 true /* recordPreGCUsage */, 9571 false /* recordPeakUsage */, 9572 false /* recordPostGCusage */, 9573 true /* recordAccumulatedGCTime */, 9574 false /* recordGCEndTime */, 9575 false /* countCollection */ ); 9576 break; 9577 9578 case CMSCollector::FinalMarking: 9579 initialize(true /* fullGC */ , 9580 cause /* cause of the GC */, 9581 false /* recordGCBeginTime */, 9582 false /* recordPreGCUsage */, 9583 false /* recordPeakUsage */, 9584 false /* recordPostGCusage */, 9585 true /* recordAccumulatedGCTime */, 9586 false /* recordGCEndTime */, 9587 false /* countCollection */ ); 9588 break; 9589 9590 case CMSCollector::Sweeping: 9591 initialize(true /* fullGC */ , 9592 cause /* cause of the GC */, 9593 false /* recordGCBeginTime */, 9594 false /* recordPreGCUsage */, 9595 true /* recordPeakUsage */, 9596 true /* recordPostGCusage */, 9597 false /* recordAccumulatedGCTime */, 9598 true /* recordGCEndTime */, 9599 true /* countCollection */ ); 9600 break; 9601 9602 default: 9603 ShouldNotReachHere(); 9604 } 9605 }