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_array(NULL), // may be set in ctor body 579 _eden_chunk_capacity(0), // -- ditto -- 580 _eden_chunk_index(0), // -- ditto -- 581 _survivor_plab_array(NULL), // -- ditto -- 582 _survivor_chunk_array(NULL), // -- ditto -- 583 _survivor_chunk_capacity(0), // -- ditto -- 584 _survivor_chunk_index(0), // -- ditto -- 585 _ser_pmc_preclean_ovflw(0), 586 _ser_kac_preclean_ovflw(0), 587 _ser_pmc_remark_ovflw(0), 588 _par_pmc_remark_ovflw(0), 589 _ser_kac_ovflw(0), 590 _par_kac_ovflw(0), 591 #ifndef PRODUCT 592 _num_par_pushes(0), 593 #endif 594 _collection_count_start(0), 595 _verifying(false), 596 _icms_start_limit(NULL), 597 _icms_stop_limit(NULL), 598 _verification_mark_bm(0, Mutex::leaf + 1, "CMS_verification_mark_bm_lock"), 599 _completed_initialization(false), 600 _collector_policy(cp), 601 _should_unload_classes(false), 602 _concurrent_cycles_since_last_unload(0), 603 _roots_scanning_options(0), 604 _inter_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding), 605 _intra_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding), 606 _gc_tracer_cm(new (ResourceObj::C_HEAP, mtGC) CMSTracer()), 607 _gc_timer_cm(new (ResourceObj::C_HEAP, mtGC) ConcurrentGCTimer()), 608 _cms_start_registered(false) 609 { 610 if (ExplicitGCInvokesConcurrentAndUnloadsClasses) { 611 ExplicitGCInvokesConcurrent = true; 612 } 613 // Now expand the span and allocate the collection support structures 614 // (MUT, marking bit map etc.) to cover both generations subject to 615 // collection. 616 617 // First check that _permGen is adjacent to _cmsGen and above it. 618 assert( _cmsGen->reserved().word_size() > 0 619 && _permGen->reserved().word_size() > 0, 620 "generations should not be of zero size"); 621 assert(_cmsGen->reserved().intersection(_permGen->reserved()).is_empty(), 622 "_cmsGen and _permGen should not overlap"); 623 assert(_cmsGen->reserved().end() == _permGen->reserved().start(), 624 "_cmsGen->end() different from _permGen->start()"); 625 626 // For use by dirty card to oop closures. 627 _cmsGen->cmsSpace()->set_collector(this); 628 _permGen->cmsSpace()->set_collector(this); 629 630 // Allocate MUT and marking bit map 631 { 632 MutexLockerEx x(_markBitMap.lock(), Mutex::_no_safepoint_check_flag); 633 if (!_markBitMap.allocate(_span)) { 634 warning("Failed to allocate CMS Bit Map"); 635 return; 636 } 637 assert(_markBitMap.covers(_span), "_markBitMap inconsistency?"); 638 } 639 { 640 _modUnionTable.allocate(_span); 641 assert(_modUnionTable.covers(_span), "_modUnionTable inconsistency?"); 642 } 643 644 if (!_markStack.allocate(MarkStackSize)) { 645 warning("Failed to allocate CMS Marking Stack"); 646 return; 647 } 648 if (!_revisitStack.allocate(CMSRevisitStackSize)) { 649 warning("Failed to allocate CMS Revisit Stack"); 650 return; 651 } 652 653 // Support for multi-threaded concurrent phases 654 if (CMSConcurrentMTEnabled) { 655 if (FLAG_IS_DEFAULT(ConcGCThreads)) { 656 // just for now 657 FLAG_SET_DEFAULT(ConcGCThreads, (ParallelGCThreads + 3)/4); 658 } 659 if (ConcGCThreads > 1) { 660 _conc_workers = new YieldingFlexibleWorkGang("Parallel CMS Threads", 661 ConcGCThreads, true); 662 if (_conc_workers == NULL) { 663 warning("GC/CMS: _conc_workers allocation failure: " 664 "forcing -CMSConcurrentMTEnabled"); 665 CMSConcurrentMTEnabled = false; 666 } else { 667 _conc_workers->initialize_workers(); 668 } 669 } else { 670 CMSConcurrentMTEnabled = false; 671 } 672 } 673 if (!CMSConcurrentMTEnabled) { 674 ConcGCThreads = 0; 675 } else { 676 // Turn off CMSCleanOnEnter optimization temporarily for 677 // the MT case where it's not fixed yet; see 6178663. 678 CMSCleanOnEnter = false; 679 } 680 assert((_conc_workers != NULL) == (ConcGCThreads > 1), 681 "Inconsistency"); 682 683 // Parallel task queues; these are shared for the 684 // concurrent and stop-world phases of CMS, but 685 // are not shared with parallel scavenge (ParNew). 686 { 687 uint i; 688 uint num_queues = (uint) MAX2(ParallelGCThreads, ConcGCThreads); 689 690 if ((CMSParallelRemarkEnabled || CMSConcurrentMTEnabled 691 || ParallelRefProcEnabled) 692 && num_queues > 0) { 693 _task_queues = new OopTaskQueueSet(num_queues); 694 if (_task_queues == NULL) { 695 warning("task_queues allocation failure."); 696 return; 697 } 698 _hash_seed = NEW_C_HEAP_ARRAY(int, num_queues, mtGC); 699 if (_hash_seed == NULL) { 700 warning("_hash_seed array allocation failure"); 701 return; 702 } 703 704 typedef Padded<OopTaskQueue> PaddedOopTaskQueue; 705 for (i = 0; i < num_queues; i++) { 706 PaddedOopTaskQueue *q = new PaddedOopTaskQueue(); 707 if (q == NULL) { 708 warning("work_queue allocation failure."); 709 return; 710 } 711 _task_queues->register_queue(i, q); 712 } 713 for (i = 0; i < num_queues; i++) { 714 _task_queues->queue(i)->initialize(); 715 _hash_seed[i] = 17; // copied from ParNew 716 } 717 } 718 } 719 720 _cmsGen ->init_initiating_occupancy(CMSInitiatingOccupancyFraction, CMSTriggerRatio); 721 _permGen->init_initiating_occupancy(CMSInitiatingPermOccupancyFraction, CMSTriggerPermRatio); 722 723 // Clip CMSBootstrapOccupancy between 0 and 100. 724 _bootstrap_occupancy = ((double)MIN2((uintx)100, MAX2((uintx)0, CMSBootstrapOccupancy))) 725 /(double)100; 726 727 _full_gcs_since_conc_gc = 0; 728 729 // Now tell CMS generations the identity of their collector 730 ConcurrentMarkSweepGeneration::set_collector(this); 731 732 // Create & start a CMS thread for this CMS collector 733 _cmsThread = ConcurrentMarkSweepThread::start(this); 734 assert(cmsThread() != NULL, "CMS Thread should have been created"); 735 assert(cmsThread()->collector() == this, 736 "CMS Thread should refer to this gen"); 737 assert(CGC_lock != NULL, "Where's the CGC_lock?"); 738 739 // Support for parallelizing young gen rescan 740 GenCollectedHeap* gch = GenCollectedHeap::heap(); 741 _young_gen = gch->prev_gen(_cmsGen); 742 if (gch->supports_inline_contig_alloc()) { 743 _top_addr = gch->top_addr(); 744 _end_addr = gch->end_addr(); 745 assert(_young_gen != NULL, "no _young_gen"); 746 _eden_chunk_index = 0; 747 _eden_chunk_capacity = (_young_gen->max_capacity()+CMSSamplingGrain)/CMSSamplingGrain; 748 _eden_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, _eden_chunk_capacity, mtGC); 749 if (_eden_chunk_array == NULL) { 750 _eden_chunk_capacity = 0; 751 warning("GC/CMS: _eden_chunk_array allocation failure"); 752 } 753 } 754 assert(_eden_chunk_array != NULL || _eden_chunk_capacity == 0, "Error"); 755 756 // Support for parallelizing survivor space rescan 757 if (CMSParallelRemarkEnabled && CMSParallelSurvivorRemarkEnabled) { 758 const size_t max_plab_samples = 759 ((DefNewGeneration*)_young_gen)->max_survivor_size()/MinTLABSize; 760 761 _survivor_plab_array = NEW_C_HEAP_ARRAY(ChunkArray, ParallelGCThreads, mtGC); 762 _survivor_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, 2*max_plab_samples, mtGC); 763 _cursor = NEW_C_HEAP_ARRAY(size_t, ParallelGCThreads, mtGC); 764 if (_survivor_plab_array == NULL || _survivor_chunk_array == NULL 765 || _cursor == NULL) { 766 warning("Failed to allocate survivor plab/chunk array"); 767 if (_survivor_plab_array != NULL) { 768 FREE_C_HEAP_ARRAY(ChunkArray, _survivor_plab_array, mtGC); 769 _survivor_plab_array = NULL; 770 } 771 if (_survivor_chunk_array != NULL) { 772 FREE_C_HEAP_ARRAY(HeapWord*, _survivor_chunk_array, mtGC); 773 _survivor_chunk_array = NULL; 774 } 775 if (_cursor != NULL) { 776 FREE_C_HEAP_ARRAY(size_t, _cursor, mtGC); 777 _cursor = NULL; 778 } 779 } else { 780 _survivor_chunk_capacity = 2*max_plab_samples; 781 for (uint i = 0; i < ParallelGCThreads; i++) { 782 HeapWord** vec = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples, mtGC); 783 if (vec == NULL) { 784 warning("Failed to allocate survivor plab array"); 785 for (int j = i; j > 0; j--) { 786 FREE_C_HEAP_ARRAY(HeapWord*, _survivor_plab_array[j-1].array(), mtGC); 787 } 788 FREE_C_HEAP_ARRAY(ChunkArray, _survivor_plab_array, mtGC); 789 FREE_C_HEAP_ARRAY(HeapWord*, _survivor_chunk_array, mtGC); 790 _survivor_plab_array = NULL; 791 _survivor_chunk_array = NULL; 792 _survivor_chunk_capacity = 0; 793 break; 794 } else { 795 ChunkArray* cur = 796 ::new (&_survivor_plab_array[i]) ChunkArray(vec, 797 max_plab_samples); 798 assert(cur->end() == 0, "Should be 0"); 799 assert(cur->array() == vec, "Should be vec"); 800 assert(cur->capacity() == max_plab_samples, "Error"); 801 } 802 } 803 } 804 } 805 assert( ( _survivor_plab_array != NULL 806 && _survivor_chunk_array != NULL) 807 || ( _survivor_chunk_capacity == 0 808 && _survivor_chunk_index == 0), 809 "Error"); 810 811 // Choose what strong roots should be scanned depending on verification options 812 // and perm gen collection mode. 813 if (!CMSClassUnloadingEnabled) { 814 // If class unloading is disabled we want to include all classes into the root set. 815 add_root_scanning_option(SharedHeap::SO_AllClasses); 816 } else { 817 add_root_scanning_option(SharedHeap::SO_SystemClasses); 818 } 819 820 NOT_PRODUCT(_overflow_counter = CMSMarkStackOverflowInterval;) 821 _gc_counters = new CollectorCounters("CMS", 1); 822 _completed_initialization = true; 823 _inter_sweep_timer.start(); // start of time 824 } 825 826 const char* ConcurrentMarkSweepGeneration::name() const { 827 return "concurrent mark-sweep generation"; 828 } 829 void ConcurrentMarkSweepGeneration::update_counters() { 830 if (UsePerfData) { 831 _space_counters->update_all(); 832 _gen_counters->update_all(); 833 } 834 } 835 836 // this is an optimized version of update_counters(). it takes the 837 // used value as a parameter rather than computing it. 838 // 839 void ConcurrentMarkSweepGeneration::update_counters(size_t used) { 840 if (UsePerfData) { 841 _space_counters->update_used(used); 842 _space_counters->update_capacity(); 843 _gen_counters->update_all(); 844 } 845 } 846 847 void ConcurrentMarkSweepGeneration::print() const { 848 Generation::print(); 849 cmsSpace()->print(); 850 } 851 852 #ifndef PRODUCT 853 void ConcurrentMarkSweepGeneration::print_statistics() { 854 cmsSpace()->printFLCensus(0); 855 } 856 #endif 857 858 void ConcurrentMarkSweepGeneration::printOccupancy(const char *s) { 859 GenCollectedHeap* gch = GenCollectedHeap::heap(); 860 if (PrintGCDetails) { 861 if (Verbose) { 862 gclog_or_tty->print(" [%d %s-%s: "SIZE_FORMAT"("SIZE_FORMAT")]", 863 level(), short_name(), s, used(), capacity()); 864 } else { 865 gclog_or_tty->print(" [%d %s-%s: "SIZE_FORMAT"K("SIZE_FORMAT"K)]", 866 level(), short_name(), s, used() / K, capacity() / K); 867 } 868 } 869 if (Verbose) { 870 gclog_or_tty->print(" "SIZE_FORMAT"("SIZE_FORMAT")", 871 gch->used(), gch->capacity()); 872 } else { 873 gclog_or_tty->print(" "SIZE_FORMAT"K("SIZE_FORMAT"K)", 874 gch->used() / K, gch->capacity() / K); 875 } 876 } 877 878 size_t 879 ConcurrentMarkSweepGeneration::contiguous_available() const { 880 // dld proposes an improvement in precision here. If the committed 881 // part of the space ends in a free block we should add that to 882 // uncommitted size in the calculation below. Will make this 883 // change later, staying with the approximation below for the 884 // time being. -- ysr. 885 return MAX2(_virtual_space.uncommitted_size(), unsafe_max_alloc_nogc()); 886 } 887 888 size_t 889 ConcurrentMarkSweepGeneration::unsafe_max_alloc_nogc() const { 890 return _cmsSpace->max_alloc_in_words() * HeapWordSize; 891 } 892 893 size_t ConcurrentMarkSweepGeneration::max_available() const { 894 return free() + _virtual_space.uncommitted_size(); 895 } 896 897 bool ConcurrentMarkSweepGeneration::promotion_attempt_is_safe(size_t max_promotion_in_bytes) const { 898 size_t available = max_available(); 899 size_t av_promo = (size_t)gc_stats()->avg_promoted()->padded_average(); 900 bool res = (available >= av_promo) || (available >= max_promotion_in_bytes); 901 if (Verbose && PrintGCDetails) { 902 gclog_or_tty->print_cr( 903 "CMS: promo attempt is%s safe: available("SIZE_FORMAT") %s av_promo("SIZE_FORMAT")," 904 "max_promo("SIZE_FORMAT")", 905 res? "":" not", available, res? ">=":"<", 906 av_promo, max_promotion_in_bytes); 907 } 908 return res; 909 } 910 911 // At a promotion failure dump information on block layout in heap 912 // (cms old generation). 913 void ConcurrentMarkSweepGeneration::promotion_failure_occurred() { 914 if (CMSDumpAtPromotionFailure) { 915 cmsSpace()->dump_at_safepoint_with_locks(collector(), gclog_or_tty); 916 } 917 } 918 919 CompactibleSpace* 920 ConcurrentMarkSweepGeneration::first_compaction_space() const { 921 return _cmsSpace; 922 } 923 924 void ConcurrentMarkSweepGeneration::reset_after_compaction() { 925 // Clear the promotion information. These pointers can be adjusted 926 // along with all the other pointers into the heap but 927 // compaction is expected to be a rare event with 928 // a heap using cms so don't do it without seeing the need. 929 if (CollectedHeap::use_parallel_gc_threads()) { 930 for (uint i = 0; i < ParallelGCThreads; i++) { 931 _par_gc_thread_states[i]->promo.reset(); 932 } 933 } 934 } 935 936 void ConcurrentMarkSweepGeneration::space_iterate(SpaceClosure* blk, bool usedOnly) { 937 blk->do_space(_cmsSpace); 938 } 939 940 void ConcurrentMarkSweepGeneration::compute_new_size() { 941 assert_locked_or_safepoint(Heap_lock); 942 943 // If incremental collection failed, we just want to expand 944 // to the limit. 945 if (incremental_collection_failed()) { 946 clear_incremental_collection_failed(); 947 grow_to_reserved(); 948 return; 949 } 950 951 size_t expand_bytes = 0; 952 double free_percentage = ((double) free()) / capacity(); 953 double desired_free_percentage = (double) MinHeapFreeRatio / 100; 954 double maximum_free_percentage = (double) MaxHeapFreeRatio / 100; 955 956 // compute expansion delta needed for reaching desired free percentage 957 if (free_percentage < desired_free_percentage) { 958 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 959 assert(desired_capacity >= capacity(), "invalid expansion size"); 960 expand_bytes = MAX2(desired_capacity - capacity(), MinHeapDeltaBytes); 961 } 962 if (expand_bytes > 0) { 963 if (PrintGCDetails && Verbose) { 964 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 965 gclog_or_tty->print_cr("\nFrom compute_new_size: "); 966 gclog_or_tty->print_cr(" Free fraction %f", free_percentage); 967 gclog_or_tty->print_cr(" Desired free fraction %f", 968 desired_free_percentage); 969 gclog_or_tty->print_cr(" Maximum free fraction %f", 970 maximum_free_percentage); 971 gclog_or_tty->print_cr(" Capactiy "SIZE_FORMAT, capacity()/1000); 972 gclog_or_tty->print_cr(" Desired capacity "SIZE_FORMAT, 973 desired_capacity/1000); 974 int prev_level = level() - 1; 975 if (prev_level >= 0) { 976 size_t prev_size = 0; 977 GenCollectedHeap* gch = GenCollectedHeap::heap(); 978 Generation* prev_gen = gch->_gens[prev_level]; 979 prev_size = prev_gen->capacity(); 980 gclog_or_tty->print_cr(" Younger gen size "SIZE_FORMAT, 981 prev_size/1000); 982 } 983 gclog_or_tty->print_cr(" unsafe_max_alloc_nogc "SIZE_FORMAT, 984 unsafe_max_alloc_nogc()/1000); 985 gclog_or_tty->print_cr(" contiguous available "SIZE_FORMAT, 986 contiguous_available()/1000); 987 gclog_or_tty->print_cr(" Expand by "SIZE_FORMAT" (bytes)", 988 expand_bytes); 989 } 990 // safe if expansion fails 991 expand(expand_bytes, 0, CMSExpansionCause::_satisfy_free_ratio); 992 if (PrintGCDetails && Verbose) { 993 gclog_or_tty->print_cr(" Expanded free fraction %f", 994 ((double) free()) / capacity()); 995 } 996 } 997 } 998 999 Mutex* ConcurrentMarkSweepGeneration::freelistLock() const { 1000 return cmsSpace()->freelistLock(); 1001 } 1002 1003 HeapWord* ConcurrentMarkSweepGeneration::allocate(size_t size, 1004 bool tlab) { 1005 CMSSynchronousYieldRequest yr; 1006 MutexLockerEx x(freelistLock(), 1007 Mutex::_no_safepoint_check_flag); 1008 return have_lock_and_allocate(size, tlab); 1009 } 1010 1011 HeapWord* ConcurrentMarkSweepGeneration::have_lock_and_allocate(size_t size, 1012 bool tlab /* ignored */) { 1013 assert_lock_strong(freelistLock()); 1014 size_t adjustedSize = CompactibleFreeListSpace::adjustObjectSize(size); 1015 HeapWord* res = cmsSpace()->allocate(adjustedSize); 1016 // Allocate the object live (grey) if the background collector has 1017 // started marking. This is necessary because the marker may 1018 // have passed this address and consequently this object will 1019 // not otherwise be greyed and would be incorrectly swept up. 1020 // Note that if this object contains references, the writing 1021 // of those references will dirty the card containing this object 1022 // allowing the object to be blackened (and its references scanned) 1023 // either during a preclean phase or at the final checkpoint. 1024 if (res != NULL) { 1025 // We may block here with an uninitialized object with 1026 // its mark-bit or P-bits not yet set. Such objects need 1027 // to be safely navigable by block_start(). 1028 assert(oop(res)->klass_or_null() == NULL, "Object should be uninitialized here."); 1029 assert(!((FreeChunk*)res)->is_free(), "Error, block will look free but show wrong size"); 1030 collector()->direct_allocated(res, adjustedSize); 1031 _direct_allocated_words += adjustedSize; 1032 // allocation counters 1033 NOT_PRODUCT( 1034 _numObjectsAllocated++; 1035 _numWordsAllocated += (int)adjustedSize; 1036 ) 1037 } 1038 return res; 1039 } 1040 1041 // In the case of direct allocation by mutators in a generation that 1042 // is being concurrently collected, the object must be allocated 1043 // live (grey) if the background collector has started marking. 1044 // This is necessary because the marker may 1045 // have passed this address and consequently this object will 1046 // not otherwise be greyed and would be incorrectly swept up. 1047 // Note that if this object contains references, the writing 1048 // of those references will dirty the card containing this object 1049 // allowing the object to be blackened (and its references scanned) 1050 // either during a preclean phase or at the final checkpoint. 1051 void CMSCollector::direct_allocated(HeapWord* start, size_t size) { 1052 assert(_markBitMap.covers(start, size), "Out of bounds"); 1053 if (_collectorState >= Marking) { 1054 MutexLockerEx y(_markBitMap.lock(), 1055 Mutex::_no_safepoint_check_flag); 1056 // [see comments preceding SweepClosure::do_blk() below for details] 1057 // 1. need to mark the object as live so it isn't collected 1058 // 2. need to mark the 2nd bit to indicate the object may be uninitialized 1059 // 3. need to mark the end of the object so marking, precleaning or sweeping 1060 // can skip over uninitialized or unparsable objects. An allocated 1061 // object is considered uninitialized for our purposes as long as 1062 // its klass word is NULL. (Unparsable objects are those which are 1063 // initialized in the sense just described, but whose sizes can still 1064 // not be correctly determined. Note that the class of unparsable objects 1065 // can only occur in the perm gen. All old gen objects are parsable 1066 // as soon as they are initialized.) 1067 _markBitMap.mark(start); // object is live 1068 _markBitMap.mark(start + 1); // object is potentially uninitialized? 1069 _markBitMap.mark(start + size - 1); 1070 // mark end of object 1071 } 1072 // check that oop looks uninitialized 1073 assert(oop(start)->klass_or_null() == NULL, "_klass should be NULL"); 1074 } 1075 1076 void CMSCollector::promoted(bool par, HeapWord* start, 1077 bool is_obj_array, size_t obj_size) { 1078 assert(_markBitMap.covers(start), "Out of bounds"); 1079 // See comment in direct_allocated() about when objects should 1080 // be allocated live. 1081 if (_collectorState >= Marking) { 1082 // we already hold the marking bit map lock, taken in 1083 // the prologue 1084 if (par) { 1085 _markBitMap.par_mark(start); 1086 } else { 1087 _markBitMap.mark(start); 1088 } 1089 // We don't need to mark the object as uninitialized (as 1090 // in direct_allocated above) because this is being done with the 1091 // world stopped and the object will be initialized by the 1092 // time the marking, precleaning or sweeping get to look at it. 1093 // But see the code for copying objects into the CMS generation, 1094 // where we need to ensure that concurrent readers of the 1095 // block offset table are able to safely navigate a block that 1096 // is in flux from being free to being allocated (and in 1097 // transition while being copied into) and subsequently 1098 // becoming a bona-fide object when the copy/promotion is complete. 1099 assert(SafepointSynchronize::is_at_safepoint(), 1100 "expect promotion only at safepoints"); 1101 1102 if (_collectorState < Sweeping) { 1103 // Mark the appropriate cards in the modUnionTable, so that 1104 // this object gets scanned before the sweep. If this is 1105 // not done, CMS generation references in the object might 1106 // not get marked. 1107 // For the case of arrays, which are otherwise precisely 1108 // marked, we need to dirty the entire array, not just its head. 1109 if (is_obj_array) { 1110 // The [par_]mark_range() method expects mr.end() below to 1111 // be aligned to the granularity of a bit's representation 1112 // in the heap. In the case of the MUT below, that's a 1113 // card size. 1114 MemRegion mr(start, 1115 (HeapWord*)round_to((intptr_t)(start + obj_size), 1116 CardTableModRefBS::card_size /* bytes */)); 1117 if (par) { 1118 _modUnionTable.par_mark_range(mr); 1119 } else { 1120 _modUnionTable.mark_range(mr); 1121 } 1122 } else { // not an obj array; we can just mark the head 1123 if (par) { 1124 _modUnionTable.par_mark(start); 1125 } else { 1126 _modUnionTable.mark(start); 1127 } 1128 } 1129 } 1130 } 1131 } 1132 1133 static inline size_t percent_of_space(Space* space, HeapWord* addr) 1134 { 1135 size_t delta = pointer_delta(addr, space->bottom()); 1136 return (size_t)(delta * 100.0 / (space->capacity() / HeapWordSize)); 1137 } 1138 1139 void CMSCollector::icms_update_allocation_limits() 1140 { 1141 Generation* gen0 = GenCollectedHeap::heap()->get_gen(0); 1142 EdenSpace* eden = gen0->as_DefNewGeneration()->eden(); 1143 1144 const unsigned int duty_cycle = stats().icms_update_duty_cycle(); 1145 if (CMSTraceIncrementalPacing) { 1146 stats().print(); 1147 } 1148 1149 assert(duty_cycle <= 100, "invalid duty cycle"); 1150 if (duty_cycle != 0) { 1151 // The duty_cycle is a percentage between 0 and 100; convert to words and 1152 // then compute the offset from the endpoints of the space. 1153 size_t free_words = eden->free() / HeapWordSize; 1154 double free_words_dbl = (double)free_words; 1155 size_t duty_cycle_words = (size_t)(free_words_dbl * duty_cycle / 100.0); 1156 size_t offset_words = (free_words - duty_cycle_words) / 2; 1157 1158 _icms_start_limit = eden->top() + offset_words; 1159 _icms_stop_limit = eden->end() - offset_words; 1160 1161 // The limits may be adjusted (shifted to the right) by 1162 // CMSIncrementalOffset, to allow the application more mutator time after a 1163 // young gen gc (when all mutators were stopped) and before CMS starts and 1164 // takes away one or more cpus. 1165 if (CMSIncrementalOffset != 0) { 1166 double adjustment_dbl = free_words_dbl * CMSIncrementalOffset / 100.0; 1167 size_t adjustment = (size_t)adjustment_dbl; 1168 HeapWord* tmp_stop = _icms_stop_limit + adjustment; 1169 if (tmp_stop > _icms_stop_limit && tmp_stop < eden->end()) { 1170 _icms_start_limit += adjustment; 1171 _icms_stop_limit = tmp_stop; 1172 } 1173 } 1174 } 1175 if (duty_cycle == 0 || (_icms_start_limit == _icms_stop_limit)) { 1176 _icms_start_limit = _icms_stop_limit = eden->end(); 1177 } 1178 1179 // Install the new start limit. 1180 eden->set_soft_end(_icms_start_limit); 1181 1182 if (CMSTraceIncrementalMode) { 1183 gclog_or_tty->print(" icms alloc limits: " 1184 PTR_FORMAT "," PTR_FORMAT 1185 " (" SIZE_FORMAT "%%," SIZE_FORMAT "%%) ", 1186 _icms_start_limit, _icms_stop_limit, 1187 percent_of_space(eden, _icms_start_limit), 1188 percent_of_space(eden, _icms_stop_limit)); 1189 if (Verbose) { 1190 gclog_or_tty->print("eden: "); 1191 eden->print_on(gclog_or_tty); 1192 } 1193 } 1194 } 1195 1196 // Any changes here should try to maintain the invariant 1197 // that if this method is called with _icms_start_limit 1198 // and _icms_stop_limit both NULL, then it should return NULL 1199 // and not notify the icms thread. 1200 HeapWord* 1201 CMSCollector::allocation_limit_reached(Space* space, HeapWord* top, 1202 size_t word_size) 1203 { 1204 // A start_limit equal to end() means the duty cycle is 0, so treat that as a 1205 // nop. 1206 if (CMSIncrementalMode && _icms_start_limit != space->end()) { 1207 if (top <= _icms_start_limit) { 1208 if (CMSTraceIncrementalMode) { 1209 space->print_on(gclog_or_tty); 1210 gclog_or_tty->stamp(); 1211 gclog_or_tty->print_cr(" start limit top=" PTR_FORMAT 1212 ", new limit=" PTR_FORMAT 1213 " (" SIZE_FORMAT "%%)", 1214 top, _icms_stop_limit, 1215 percent_of_space(space, _icms_stop_limit)); 1216 } 1217 ConcurrentMarkSweepThread::start_icms(); 1218 assert(top < _icms_stop_limit, "Tautology"); 1219 if (word_size < pointer_delta(_icms_stop_limit, top)) { 1220 return _icms_stop_limit; 1221 } 1222 1223 // The allocation will cross both the _start and _stop limits, so do the 1224 // stop notification also and return end(). 1225 if (CMSTraceIncrementalMode) { 1226 space->print_on(gclog_or_tty); 1227 gclog_or_tty->stamp(); 1228 gclog_or_tty->print_cr(" +stop limit top=" PTR_FORMAT 1229 ", new limit=" PTR_FORMAT 1230 " (" SIZE_FORMAT "%%)", 1231 top, space->end(), 1232 percent_of_space(space, space->end())); 1233 } 1234 ConcurrentMarkSweepThread::stop_icms(); 1235 return space->end(); 1236 } 1237 1238 if (top <= _icms_stop_limit) { 1239 if (CMSTraceIncrementalMode) { 1240 space->print_on(gclog_or_tty); 1241 gclog_or_tty->stamp(); 1242 gclog_or_tty->print_cr(" stop limit top=" PTR_FORMAT 1243 ", new limit=" PTR_FORMAT 1244 " (" SIZE_FORMAT "%%)", 1245 top, space->end(), 1246 percent_of_space(space, space->end())); 1247 } 1248 ConcurrentMarkSweepThread::stop_icms(); 1249 return space->end(); 1250 } 1251 1252 if (CMSTraceIncrementalMode) { 1253 space->print_on(gclog_or_tty); 1254 gclog_or_tty->stamp(); 1255 gclog_or_tty->print_cr(" end limit top=" PTR_FORMAT 1256 ", new limit=" PTR_FORMAT, 1257 top, NULL); 1258 } 1259 } 1260 1261 return NULL; 1262 } 1263 1264 oop ConcurrentMarkSweepGeneration::promote(oop obj, size_t obj_size) { 1265 assert(obj_size == (size_t)obj->size(), "bad obj_size passed in"); 1266 // allocate, copy and if necessary update promoinfo -- 1267 // delegate to underlying space. 1268 assert_lock_strong(freelistLock()); 1269 1270 #ifndef PRODUCT 1271 if (Universe::heap()->promotion_should_fail()) { 1272 return NULL; 1273 } 1274 #endif // #ifndef PRODUCT 1275 1276 oop res = _cmsSpace->promote(obj, obj_size); 1277 if (res == NULL) { 1278 // expand and retry 1279 size_t s = _cmsSpace->expansionSpaceRequired(obj_size); // HeapWords 1280 expand(s*HeapWordSize, MinHeapDeltaBytes, 1281 CMSExpansionCause::_satisfy_promotion); 1282 // Since there's currently no next generation, we don't try to promote 1283 // into a more senior generation. 1284 assert(next_gen() == NULL, "assumption, based upon which no attempt " 1285 "is made to pass on a possibly failing " 1286 "promotion to next generation"); 1287 res = _cmsSpace->promote(obj, obj_size); 1288 } 1289 if (res != NULL) { 1290 // See comment in allocate() about when objects should 1291 // be allocated live. 1292 assert(obj->is_oop(), "Will dereference klass pointer below"); 1293 collector()->promoted(false, // Not parallel 1294 (HeapWord*)res, obj->is_objArray(), obj_size); 1295 // promotion counters 1296 NOT_PRODUCT( 1297 _numObjectsPromoted++; 1298 _numWordsPromoted += 1299 (int)(CompactibleFreeListSpace::adjustObjectSize(obj->size())); 1300 ) 1301 } 1302 return res; 1303 } 1304 1305 1306 HeapWord* 1307 ConcurrentMarkSweepGeneration::allocation_limit_reached(Space* space, 1308 HeapWord* top, 1309 size_t word_sz) 1310 { 1311 return collector()->allocation_limit_reached(space, top, word_sz); 1312 } 1313 1314 // IMPORTANT: Notes on object size recognition in CMS. 1315 // --------------------------------------------------- 1316 // A block of storage in the CMS generation is always in 1317 // one of three states. A free block (FREE), an allocated 1318 // object (OBJECT) whose size() method reports the correct size, 1319 // and an intermediate state (TRANSIENT) in which its size cannot 1320 // be accurately determined. 1321 // STATE IDENTIFICATION: (32 bit and 64 bit w/o COOPS) 1322 // ----------------------------------------------------- 1323 // FREE: klass_word & 1 == 1; mark_word holds block size 1324 // 1325 // OBJECT: klass_word installed; klass_word != 0 && klass_word & 1 == 0; 1326 // obj->size() computes correct size 1327 // [Perm Gen objects needs to be "parsable" before they can be navigated] 1328 // 1329 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT 1330 // 1331 // STATE IDENTIFICATION: (64 bit+COOPS) 1332 // ------------------------------------ 1333 // FREE: mark_word & CMS_FREE_BIT == 1; mark_word & ~CMS_FREE_BIT gives block_size 1334 // 1335 // OBJECT: klass_word installed; klass_word != 0; 1336 // obj->size() computes correct size 1337 // [Perm Gen comment above continues to hold] 1338 // 1339 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT 1340 // 1341 // 1342 // STATE TRANSITION DIAGRAM 1343 // 1344 // mut / parnew mut / parnew 1345 // FREE --------------------> TRANSIENT ---------------------> OBJECT --| 1346 // ^ | 1347 // |------------------------ DEAD <------------------------------------| 1348 // sweep mut 1349 // 1350 // While a block is in TRANSIENT state its size cannot be determined 1351 // so readers will either need to come back later or stall until 1352 // the size can be determined. Note that for the case of direct 1353 // allocation, P-bits, when available, may be used to determine the 1354 // size of an object that may not yet have been initialized. 1355 1356 // Things to support parallel young-gen collection. 1357 oop 1358 ConcurrentMarkSweepGeneration::par_promote(int thread_num, 1359 oop old, markOop m, 1360 size_t word_sz) { 1361 #ifndef PRODUCT 1362 if (Universe::heap()->promotion_should_fail()) { 1363 return NULL; 1364 } 1365 #endif // #ifndef PRODUCT 1366 1367 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1368 PromotionInfo* promoInfo = &ps->promo; 1369 // if we are tracking promotions, then first ensure space for 1370 // promotion (including spooling space for saving header if necessary). 1371 // then allocate and copy, then track promoted info if needed. 1372 // When tracking (see PromotionInfo::track()), the mark word may 1373 // be displaced and in this case restoration of the mark word 1374 // occurs in the (oop_since_save_marks_)iterate phase. 1375 if (promoInfo->tracking() && !promoInfo->ensure_spooling_space()) { 1376 // Out of space for allocating spooling buffers; 1377 // try expanding and allocating spooling buffers. 1378 if (!expand_and_ensure_spooling_space(promoInfo)) { 1379 return NULL; 1380 } 1381 } 1382 assert(promoInfo->has_spooling_space(), "Control point invariant"); 1383 const size_t alloc_sz = CompactibleFreeListSpace::adjustObjectSize(word_sz); 1384 HeapWord* obj_ptr = ps->lab.alloc(alloc_sz); 1385 if (obj_ptr == NULL) { 1386 obj_ptr = expand_and_par_lab_allocate(ps, alloc_sz); 1387 if (obj_ptr == NULL) { 1388 return NULL; 1389 } 1390 } 1391 oop obj = oop(obj_ptr); 1392 OrderAccess::storestore(); 1393 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1394 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1395 // IMPORTANT: See note on object initialization for CMS above. 1396 // Otherwise, copy the object. Here we must be careful to insert the 1397 // klass pointer last, since this marks the block as an allocated object. 1398 // Except with compressed oops it's the mark word. 1399 HeapWord* old_ptr = (HeapWord*)old; 1400 // Restore the mark word copied above. 1401 obj->set_mark(m); 1402 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1403 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1404 OrderAccess::storestore(); 1405 1406 if (UseCompressedOops) { 1407 // Copy gap missed by (aligned) header size calculation below 1408 obj->set_klass_gap(old->klass_gap()); 1409 } 1410 if (word_sz > (size_t)oopDesc::header_size()) { 1411 Copy::aligned_disjoint_words(old_ptr + oopDesc::header_size(), 1412 obj_ptr + oopDesc::header_size(), 1413 word_sz - oopDesc::header_size()); 1414 } 1415 1416 // Now we can track the promoted object, if necessary. We take care 1417 // to delay the transition from uninitialized to full object 1418 // (i.e., insertion of klass pointer) until after, so that it 1419 // atomically becomes a promoted object. 1420 if (promoInfo->tracking()) { 1421 promoInfo->track((PromotedObject*)obj, old->klass()); 1422 } 1423 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1424 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1425 assert(old->is_oop(), "Will use and dereference old klass ptr below"); 1426 1427 // Finally, install the klass pointer (this should be volatile). 1428 OrderAccess::storestore(); 1429 obj->set_klass(old->klass()); 1430 // We should now be able to calculate the right size for this object 1431 assert(obj->is_oop() && obj->size() == (int)word_sz, "Error, incorrect size computed for promoted object"); 1432 1433 collector()->promoted(true, // parallel 1434 obj_ptr, old->is_objArray(), word_sz); 1435 1436 NOT_PRODUCT( 1437 Atomic::inc_ptr(&_numObjectsPromoted); 1438 Atomic::add_ptr(alloc_sz, &_numWordsPromoted); 1439 ) 1440 1441 return obj; 1442 } 1443 1444 void 1445 ConcurrentMarkSweepGeneration:: 1446 par_promote_alloc_undo(int thread_num, 1447 HeapWord* obj, size_t word_sz) { 1448 // CMS does not support promotion undo. 1449 ShouldNotReachHere(); 1450 } 1451 1452 void 1453 ConcurrentMarkSweepGeneration:: 1454 par_promote_alloc_done(int thread_num) { 1455 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1456 ps->lab.retire(thread_num); 1457 } 1458 1459 void 1460 ConcurrentMarkSweepGeneration:: 1461 par_oop_since_save_marks_iterate_done(int thread_num) { 1462 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1463 ParScanWithoutBarrierClosure* dummy_cl = NULL; 1464 ps->promo.promoted_oops_iterate_nv(dummy_cl); 1465 } 1466 1467 // XXXPERM 1468 bool ConcurrentMarkSweepGeneration::should_collect(bool full, 1469 size_t size, 1470 bool tlab) 1471 { 1472 // We allow a STW collection only if a full 1473 // collection was requested. 1474 return full || should_allocate(size, tlab); // FIX ME !!! 1475 // This and promotion failure handling are connected at the 1476 // hip and should be fixed by untying them. 1477 } 1478 1479 bool CMSCollector::shouldConcurrentCollect() { 1480 if (_full_gc_requested) { 1481 if (Verbose && PrintGCDetails) { 1482 gclog_or_tty->print_cr("CMSCollector: collect because of explicit " 1483 " gc request (or gc_locker)"); 1484 } 1485 return true; 1486 } 1487 1488 // For debugging purposes, change the type of collection. 1489 // If the rotation is not on the concurrent collection 1490 // type, don't start a concurrent collection. 1491 NOT_PRODUCT( 1492 if (RotateCMSCollectionTypes && 1493 (_cmsGen->debug_collection_type() != 1494 ConcurrentMarkSweepGeneration::Concurrent_collection_type)) { 1495 assert(_cmsGen->debug_collection_type() != 1496 ConcurrentMarkSweepGeneration::Unknown_collection_type, 1497 "Bad cms collection type"); 1498 return false; 1499 } 1500 ) 1501 1502 FreelistLocker x(this); 1503 // ------------------------------------------------------------------ 1504 // Print out lots of information which affects the initiation of 1505 // a collection. 1506 if (PrintCMSInitiationStatistics && stats().valid()) { 1507 gclog_or_tty->print("CMSCollector shouldConcurrentCollect: "); 1508 gclog_or_tty->stamp(); 1509 gclog_or_tty->print_cr(""); 1510 stats().print_on(gclog_or_tty); 1511 gclog_or_tty->print_cr("time_until_cms_gen_full %3.7f", 1512 stats().time_until_cms_gen_full()); 1513 gclog_or_tty->print_cr("free="SIZE_FORMAT, _cmsGen->free()); 1514 gclog_or_tty->print_cr("contiguous_available="SIZE_FORMAT, 1515 _cmsGen->contiguous_available()); 1516 gclog_or_tty->print_cr("promotion_rate=%g", stats().promotion_rate()); 1517 gclog_or_tty->print_cr("cms_allocation_rate=%g", stats().cms_allocation_rate()); 1518 gclog_or_tty->print_cr("occupancy=%3.7f", _cmsGen->occupancy()); 1519 gclog_or_tty->print_cr("initiatingOccupancy=%3.7f", _cmsGen->initiating_occupancy()); 1520 gclog_or_tty->print_cr("initiatingPermOccupancy=%3.7f", _permGen->initiating_occupancy()); 1521 } 1522 // ------------------------------------------------------------------ 1523 1524 // If the estimated time to complete a cms collection (cms_duration()) 1525 // is less than the estimated time remaining until the cms generation 1526 // is full, start a collection. 1527 if (!UseCMSInitiatingOccupancyOnly) { 1528 if (stats().valid()) { 1529 if (stats().time_until_cms_start() == 0.0) { 1530 return true; 1531 } 1532 } else { 1533 // We want to conservatively collect somewhat early in order 1534 // to try and "bootstrap" our CMS/promotion statistics; 1535 // this branch will not fire after the first successful CMS 1536 // collection because the stats should then be valid. 1537 if (_cmsGen->occupancy() >= _bootstrap_occupancy) { 1538 if (Verbose && PrintGCDetails) { 1539 gclog_or_tty->print_cr( 1540 " CMSCollector: collect for bootstrapping statistics:" 1541 " occupancy = %f, boot occupancy = %f", _cmsGen->occupancy(), 1542 _bootstrap_occupancy); 1543 } 1544 return true; 1545 } 1546 } 1547 } 1548 1549 // Otherwise, we start a collection cycle if either the perm gen or 1550 // old gen want a collection cycle started. Each may use 1551 // an appropriate criterion for making this decision. 1552 // XXX We need to make sure that the gen expansion 1553 // criterion dovetails well with this. XXX NEED TO FIX THIS 1554 if (_cmsGen->should_concurrent_collect()) { 1555 if (Verbose && PrintGCDetails) { 1556 gclog_or_tty->print_cr("CMS old gen initiated"); 1557 } 1558 return true; 1559 } 1560 1561 // We start a collection if we believe an incremental collection may fail; 1562 // this is not likely to be productive in practice because it's probably too 1563 // late anyway. 1564 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1565 assert(gch->collector_policy()->is_two_generation_policy(), 1566 "You may want to check the correctness of the following"); 1567 if (gch->incremental_collection_will_fail(true /* consult_young */)) { 1568 if (Verbose && PrintGCDetails) { 1569 gclog_or_tty->print("CMSCollector: collect because incremental collection will fail "); 1570 } 1571 return true; 1572 } 1573 1574 if (CMSClassUnloadingEnabled && _permGen->should_concurrent_collect()) { 1575 bool res = update_should_unload_classes(); 1576 if (res) { 1577 if (Verbose && PrintGCDetails) { 1578 gclog_or_tty->print_cr("CMS perm gen initiated"); 1579 } 1580 return true; 1581 } 1582 } 1583 return false; 1584 } 1585 1586 // Clear _expansion_cause fields of constituent generations 1587 void CMSCollector::clear_expansion_cause() { 1588 _cmsGen->clear_expansion_cause(); 1589 _permGen->clear_expansion_cause(); 1590 } 1591 1592 // We should be conservative in starting a collection cycle. To 1593 // start too eagerly runs the risk of collecting too often in the 1594 // extreme. To collect too rarely falls back on full collections, 1595 // which works, even if not optimum in terms of concurrent work. 1596 // As a work around for too eagerly collecting, use the flag 1597 // UseCMSInitiatingOccupancyOnly. This also has the advantage of 1598 // giving the user an easily understandable way of controlling the 1599 // collections. 1600 // We want to start a new collection cycle if any of the following 1601 // conditions hold: 1602 // . our current occupancy exceeds the configured initiating occupancy 1603 // for this generation, or 1604 // . we recently needed to expand this space and have not, since that 1605 // expansion, done a collection of this generation, or 1606 // . the underlying space believes that it may be a good idea to initiate 1607 // a concurrent collection (this may be based on criteria such as the 1608 // following: the space uses linear allocation and linear allocation is 1609 // going to fail, or there is believed to be excessive fragmentation in 1610 // the generation, etc... or ... 1611 // [.(currently done by CMSCollector::shouldConcurrentCollect() only for 1612 // the case of the old generation, not the perm generation; see CR 6543076): 1613 // we may be approaching a point at which allocation requests may fail because 1614 // we will be out of sufficient free space given allocation rate estimates.] 1615 bool ConcurrentMarkSweepGeneration::should_concurrent_collect() const { 1616 1617 assert_lock_strong(freelistLock()); 1618 if (occupancy() > initiating_occupancy()) { 1619 if (PrintGCDetails && Verbose) { 1620 gclog_or_tty->print(" %s: collect because of occupancy %f / %f ", 1621 short_name(), occupancy(), initiating_occupancy()); 1622 } 1623 return true; 1624 } 1625 if (UseCMSInitiatingOccupancyOnly) { 1626 return false; 1627 } 1628 if (expansion_cause() == CMSExpansionCause::_satisfy_allocation) { 1629 if (PrintGCDetails && Verbose) { 1630 gclog_or_tty->print(" %s: collect because expanded for allocation ", 1631 short_name()); 1632 } 1633 return true; 1634 } 1635 if (_cmsSpace->should_concurrent_collect()) { 1636 if (PrintGCDetails && Verbose) { 1637 gclog_or_tty->print(" %s: collect because cmsSpace says so ", 1638 short_name()); 1639 } 1640 return true; 1641 } 1642 return false; 1643 } 1644 1645 void ConcurrentMarkSweepGeneration::collect(bool full, 1646 bool clear_all_soft_refs, 1647 size_t size, 1648 bool tlab) 1649 { 1650 collector()->collect(full, clear_all_soft_refs, size, tlab); 1651 } 1652 1653 void CMSCollector::collect(bool full, 1654 bool clear_all_soft_refs, 1655 size_t size, 1656 bool tlab) 1657 { 1658 if (!UseCMSCollectionPassing && _collectorState > Idling) { 1659 // For debugging purposes skip the collection if the state 1660 // is not currently idle 1661 if (TraceCMSState) { 1662 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " skipped full:%d CMS state %d", 1663 Thread::current(), full, _collectorState); 1664 } 1665 return; 1666 } 1667 1668 // The following "if" branch is present for defensive reasons. 1669 // In the current uses of this interface, it can be replaced with: 1670 // assert(!GC_locker.is_active(), "Can't be called otherwise"); 1671 // But I am not placing that assert here to allow future 1672 // generality in invoking this interface. 1673 if (GC_locker::is_active()) { 1674 // A consistency test for GC_locker 1675 assert(GC_locker::needs_gc(), "Should have been set already"); 1676 // Skip this foreground collection, instead 1677 // expanding the heap if necessary. 1678 // Need the free list locks for the call to free() in compute_new_size() 1679 compute_new_size(); 1680 return; 1681 } 1682 acquire_control_and_collect(full, clear_all_soft_refs); 1683 _full_gcs_since_conc_gc++; 1684 1685 } 1686 1687 void CMSCollector::request_full_gc(unsigned int full_gc_count, GCCause::Cause cause) { 1688 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1689 unsigned int gc_count = gch->total_full_collections(); 1690 if (gc_count == full_gc_count) { 1691 MutexLockerEx y(CGC_lock, Mutex::_no_safepoint_check_flag); 1692 _full_gc_requested = true; 1693 _full_gc_cause = cause; 1694 CGC_lock->notify(); // nudge CMS thread 1695 } else { 1696 assert(gc_count > full_gc_count, "Error: causal loop"); 1697 } 1698 } 1699 1700 bool CMSCollector::is_external_interruption() { 1701 GCCause::Cause cause = GenCollectedHeap::heap()->gc_cause(); 1702 return GCCause::is_user_requested_gc(cause) || 1703 GCCause::is_serviceability_requested_gc(cause); 1704 } 1705 1706 void CMSCollector::report_concurrent_mode_interruption() { 1707 if (is_external_interruption()) { 1708 if (PrintGCDetails) { 1709 gclog_or_tty->print(" (concurrent mode interrupted)"); 1710 } 1711 } else { 1712 if (PrintGCDetails) { 1713 gclog_or_tty->print(" (concurrent mode failure)"); 1714 } 1715 _gc_tracer_cm->report_concurrent_mode_failure(); 1716 } 1717 } 1718 1719 // The foreground and background collectors need to coordinate in order 1720 // to make sure that they do not mutually interfere with CMS collections. 1721 // When a background collection is active, 1722 // the foreground collector may need to take over (preempt) and 1723 // synchronously complete an ongoing collection. Depending on the 1724 // frequency of the background collections and the heap usage 1725 // of the application, this preemption can be seldom or frequent. 1726 // There are only certain 1727 // points in the background collection that the "collection-baton" 1728 // can be passed to the foreground collector. 1729 // 1730 // The foreground collector will wait for the baton before 1731 // starting any part of the collection. The foreground collector 1732 // will only wait at one location. 1733 // 1734 // The background collector will yield the baton before starting a new 1735 // phase of the collection (e.g., before initial marking, marking from roots, 1736 // precleaning, final re-mark, sweep etc.) This is normally done at the head 1737 // of the loop which switches the phases. The background collector does some 1738 // of the phases (initial mark, final re-mark) with the world stopped. 1739 // Because of locking involved in stopping the world, 1740 // the foreground collector should not block waiting for the background 1741 // collector when it is doing a stop-the-world phase. The background 1742 // collector will yield the baton at an additional point just before 1743 // it enters a stop-the-world phase. Once the world is stopped, the 1744 // background collector checks the phase of the collection. If the 1745 // phase has not changed, it proceeds with the collection. If the 1746 // phase has changed, it skips that phase of the collection. See 1747 // the comments on the use of the Heap_lock in collect_in_background(). 1748 // 1749 // Variable used in baton passing. 1750 // _foregroundGCIsActive - Set to true by the foreground collector when 1751 // it wants the baton. The foreground clears it when it has finished 1752 // the collection. 1753 // _foregroundGCShouldWait - Set to true by the background collector 1754 // when it is running. The foreground collector waits while 1755 // _foregroundGCShouldWait is true. 1756 // CGC_lock - monitor used to protect access to the above variables 1757 // and to notify the foreground and background collectors. 1758 // _collectorState - current state of the CMS collection. 1759 // 1760 // The foreground collector 1761 // acquires the CGC_lock 1762 // sets _foregroundGCIsActive 1763 // waits on the CGC_lock for _foregroundGCShouldWait to be false 1764 // various locks acquired in preparation for the collection 1765 // are released so as not to block the background collector 1766 // that is in the midst of a collection 1767 // proceeds with the collection 1768 // clears _foregroundGCIsActive 1769 // returns 1770 // 1771 // The background collector in a loop iterating on the phases of the 1772 // collection 1773 // acquires the CGC_lock 1774 // sets _foregroundGCShouldWait 1775 // if _foregroundGCIsActive is set 1776 // clears _foregroundGCShouldWait, notifies _CGC_lock 1777 // waits on _CGC_lock for _foregroundGCIsActive to become false 1778 // and exits the loop. 1779 // otherwise 1780 // proceed with that phase of the collection 1781 // if the phase is a stop-the-world phase, 1782 // yield the baton once more just before enqueueing 1783 // the stop-world CMS operation (executed by the VM thread). 1784 // returns after all phases of the collection are done 1785 // 1786 1787 void CMSCollector::acquire_control_and_collect(bool full, 1788 bool clear_all_soft_refs) { 1789 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint"); 1790 assert(!Thread::current()->is_ConcurrentGC_thread(), 1791 "shouldn't try to acquire control from self!"); 1792 1793 // Start the protocol for acquiring control of the 1794 // collection from the background collector (aka CMS thread). 1795 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 1796 "VM thread should have CMS token"); 1797 // Remember the possibly interrupted state of an ongoing 1798 // concurrent collection 1799 CollectorState first_state = _collectorState; 1800 1801 // Signal to a possibly ongoing concurrent collection that 1802 // we want to do a foreground collection. 1803 _foregroundGCIsActive = true; 1804 1805 // Disable incremental mode during a foreground collection. 1806 ICMSDisabler icms_disabler; 1807 1808 // release locks and wait for a notify from the background collector 1809 // releasing the locks in only necessary for phases which 1810 // do yields to improve the granularity of the collection. 1811 assert_lock_strong(bitMapLock()); 1812 // We need to lock the Free list lock for the space that we are 1813 // currently collecting. 1814 assert(haveFreelistLocks(), "Must be holding free list locks"); 1815 bitMapLock()->unlock(); 1816 releaseFreelistLocks(); 1817 { 1818 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1819 if (_foregroundGCShouldWait) { 1820 // We are going to be waiting for action for the CMS thread; 1821 // it had better not be gone (for instance at shutdown)! 1822 assert(ConcurrentMarkSweepThread::cmst() != NULL, 1823 "CMS thread must be running"); 1824 // Wait here until the background collector gives us the go-ahead 1825 ConcurrentMarkSweepThread::clear_CMS_flag( 1826 ConcurrentMarkSweepThread::CMS_vm_has_token); // release token 1827 // Get a possibly blocked CMS thread going: 1828 // Note that we set _foregroundGCIsActive true above, 1829 // without protection of the CGC_lock. 1830 CGC_lock->notify(); 1831 assert(!ConcurrentMarkSweepThread::vm_thread_wants_cms_token(), 1832 "Possible deadlock"); 1833 while (_foregroundGCShouldWait) { 1834 // wait for notification 1835 CGC_lock->wait(Mutex::_no_safepoint_check_flag); 1836 // Possibility of delay/starvation here, since CMS token does 1837 // not know to give priority to VM thread? Actually, i think 1838 // there wouldn't be any delay/starvation, but the proof of 1839 // that "fact" (?) appears non-trivial. XXX 20011219YSR 1840 } 1841 ConcurrentMarkSweepThread::set_CMS_flag( 1842 ConcurrentMarkSweepThread::CMS_vm_has_token); 1843 } 1844 } 1845 // The CMS_token is already held. Get back the other locks. 1846 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 1847 "VM thread should have CMS token"); 1848 getFreelistLocks(); 1849 bitMapLock()->lock_without_safepoint_check(); 1850 if (TraceCMSState) { 1851 gclog_or_tty->print_cr("CMS foreground collector has asked for control " 1852 INTPTR_FORMAT " with first state %d", Thread::current(), first_state); 1853 gclog_or_tty->print_cr(" gets control with state %d", _collectorState); 1854 } 1855 1856 // Check if we need to do a compaction, or if not, whether 1857 // we need to start the mark-sweep from scratch. 1858 bool should_compact = false; 1859 bool should_start_over = false; 1860 decide_foreground_collection_type(clear_all_soft_refs, 1861 &should_compact, &should_start_over); 1862 1863 NOT_PRODUCT( 1864 if (RotateCMSCollectionTypes) { 1865 if (_cmsGen->debug_collection_type() == 1866 ConcurrentMarkSweepGeneration::MSC_foreground_collection_type) { 1867 should_compact = true; 1868 } else if (_cmsGen->debug_collection_type() == 1869 ConcurrentMarkSweepGeneration::MS_foreground_collection_type) { 1870 should_compact = false; 1871 } 1872 } 1873 ) 1874 1875 if (first_state > Idling) { 1876 report_concurrent_mode_interruption(); 1877 } 1878 1879 if (should_compact) { 1880 // If the collection is being acquired from the background 1881 // collector, there may be references on the discovered 1882 // references lists that have NULL referents (being those 1883 // that were concurrently cleared by a mutator) or 1884 // that are no longer active (having been enqueued concurrently 1885 // by the mutator). 1886 // Scrub the list of those references because Mark-Sweep-Compact 1887 // code assumes referents are not NULL and that all discovered 1888 // Reference objects are active. 1889 ref_processor()->clean_up_discovered_references(); 1890 1891 if (first_state > Idling) { 1892 save_heap_summary(); 1893 } 1894 1895 do_compaction_work(clear_all_soft_refs); 1896 1897 // Has the GC time limit been exceeded? 1898 DefNewGeneration* young_gen = _young_gen->as_DefNewGeneration(); 1899 size_t max_eden_size = young_gen->max_capacity() - 1900 young_gen->to()->capacity() - 1901 young_gen->from()->capacity(); 1902 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1903 GCCause::Cause gc_cause = gch->gc_cause(); 1904 size_policy()->check_gc_overhead_limit(_young_gen->used(), 1905 young_gen->eden()->used(), 1906 _cmsGen->max_capacity(), 1907 max_eden_size, 1908 full, 1909 gc_cause, 1910 gch->collector_policy()); 1911 } else { 1912 do_mark_sweep_work(clear_all_soft_refs, first_state, 1913 should_start_over); 1914 } 1915 // Reset the expansion cause, now that we just completed 1916 // a collection cycle. 1917 clear_expansion_cause(); 1918 _foregroundGCIsActive = false; 1919 return; 1920 } 1921 1922 // Resize the perm generation and the tenured generation 1923 // after obtaining the free list locks for the 1924 // two generations. 1925 void CMSCollector::compute_new_size() { 1926 assert_locked_or_safepoint(Heap_lock); 1927 FreelistLocker z(this); 1928 _permGen->compute_new_size(); 1929 _cmsGen->compute_new_size(); 1930 } 1931 1932 // A work method used by foreground collection to determine 1933 // what type of collection (compacting or not, continuing or fresh) 1934 // it should do. 1935 // NOTE: the intent is to make UseCMSCompactAtFullCollection 1936 // and CMSCompactWhenClearAllSoftRefs the default in the future 1937 // and do away with the flags after a suitable period. 1938 void CMSCollector::decide_foreground_collection_type( 1939 bool clear_all_soft_refs, bool* should_compact, 1940 bool* should_start_over) { 1941 // Normally, we'll compact only if the UseCMSCompactAtFullCollection 1942 // flag is set, and we have either requested a System.gc() or 1943 // the number of full gc's since the last concurrent cycle 1944 // has exceeded the threshold set by CMSFullGCsBeforeCompaction, 1945 // or if an incremental collection has failed 1946 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1947 assert(gch->collector_policy()->is_two_generation_policy(), 1948 "You may want to check the correctness of the following"); 1949 // Inform cms gen if this was due to partial collection failing. 1950 // The CMS gen may use this fact to determine its expansion policy. 1951 if (gch->incremental_collection_will_fail(false /* don't consult_young */)) { 1952 assert(!_cmsGen->incremental_collection_failed(), 1953 "Should have been noticed, reacted to and cleared"); 1954 _cmsGen->set_incremental_collection_failed(); 1955 } 1956 *should_compact = 1957 UseCMSCompactAtFullCollection && 1958 ((_full_gcs_since_conc_gc >= CMSFullGCsBeforeCompaction) || 1959 GCCause::is_user_requested_gc(gch->gc_cause()) || 1960 gch->incremental_collection_will_fail(true /* consult_young */)); 1961 *should_start_over = false; 1962 if (clear_all_soft_refs && !*should_compact) { 1963 // We are about to do a last ditch collection attempt 1964 // so it would normally make sense to do a compaction 1965 // to reclaim as much space as possible. 1966 if (CMSCompactWhenClearAllSoftRefs) { 1967 // Default: The rationale is that in this case either 1968 // we are past the final marking phase, in which case 1969 // we'd have to start over, or so little has been done 1970 // that there's little point in saving that work. Compaction 1971 // appears to be the sensible choice in either case. 1972 *should_compact = true; 1973 } else { 1974 // We have been asked to clear all soft refs, but not to 1975 // compact. Make sure that we aren't past the final checkpoint 1976 // phase, for that is where we process soft refs. If we are already 1977 // past that phase, we'll need to redo the refs discovery phase and 1978 // if necessary clear soft refs that weren't previously 1979 // cleared. We do so by remembering the phase in which 1980 // we came in, and if we are past the refs processing 1981 // phase, we'll choose to just redo the mark-sweep 1982 // collection from scratch. 1983 if (_collectorState > FinalMarking) { 1984 // We are past the refs processing phase; 1985 // start over and do a fresh synchronous CMS cycle 1986 _collectorState = Resetting; // skip to reset to start new cycle 1987 reset(false /* == !asynch */); 1988 *should_start_over = true; 1989 } // else we can continue a possibly ongoing current cycle 1990 } 1991 } 1992 } 1993 1994 // A work method used by the foreground collector to do 1995 // a mark-sweep-compact. 1996 void CMSCollector::do_compaction_work(bool clear_all_soft_refs) { 1997 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1998 1999 STWGCTimer* gc_timer = GenMarkSweep::gc_timer(); 2000 gc_timer->register_gc_start(os::elapsed_counter()); 2001 2002 SerialOldTracer* gc_tracer = GenMarkSweep::gc_tracer(); 2003 gc_tracer->report_gc_start(gch->gc_cause(), gc_timer->gc_start()); 2004 2005 GCTraceTime t("CMS:MSC ", PrintGCDetails && Verbose, true, NULL); 2006 if (PrintGC && Verbose && !(GCCause::is_user_requested_gc(gch->gc_cause()))) { 2007 gclog_or_tty->print_cr("Compact ConcurrentMarkSweepGeneration after %d " 2008 "collections passed to foreground collector", _full_gcs_since_conc_gc); 2009 } 2010 2011 // Sample collection interval time and reset for collection pause. 2012 if (UseAdaptiveSizePolicy) { 2013 size_policy()->msc_collection_begin(); 2014 } 2015 2016 // Temporarily widen the span of the weak reference processing to 2017 // the entire heap. 2018 MemRegion new_span(GenCollectedHeap::heap()->reserved_region()); 2019 ReferenceProcessorSpanMutator rp_mut_span(ref_processor(), new_span); 2020 // Temporarily, clear the "is_alive_non_header" field of the 2021 // reference processor. 2022 ReferenceProcessorIsAliveMutator rp_mut_closure(ref_processor(), NULL); 2023 // Temporarily make reference _processing_ single threaded (non-MT). 2024 ReferenceProcessorMTProcMutator rp_mut_mt_processing(ref_processor(), false); 2025 // Temporarily make refs discovery atomic 2026 ReferenceProcessorAtomicMutator rp_mut_atomic(ref_processor(), true); 2027 // Temporarily make reference _discovery_ single threaded (non-MT) 2028 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false); 2029 2030 ref_processor()->set_enqueuing_is_done(false); 2031 ref_processor()->enable_discovery(false /*verify_disabled*/, false /*check_no_refs*/); 2032 ref_processor()->setup_policy(clear_all_soft_refs); 2033 // If an asynchronous collection finishes, the _modUnionTable is 2034 // all clear. If we are assuming the collection from an asynchronous 2035 // collection, clear the _modUnionTable. 2036 assert(_collectorState != Idling || _modUnionTable.isAllClear(), 2037 "_modUnionTable should be clear if the baton was not passed"); 2038 _modUnionTable.clear_all(); 2039 2040 // We must adjust the allocation statistics being maintained 2041 // in the free list space. We do so by reading and clearing 2042 // the sweep timer and updating the block flux rate estimates below. 2043 assert(!_intra_sweep_timer.is_active(), "_intra_sweep_timer should be inactive"); 2044 if (_inter_sweep_timer.is_active()) { 2045 _inter_sweep_timer.stop(); 2046 // Note that we do not use this sample to update the _inter_sweep_estimate. 2047 _cmsGen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()), 2048 _inter_sweep_estimate.padded_average(), 2049 _intra_sweep_estimate.padded_average()); 2050 } 2051 2052 GenMarkSweep::invoke_at_safepoint(_cmsGen->level(), 2053 ref_processor(), clear_all_soft_refs); 2054 #ifdef ASSERT 2055 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 2056 size_t free_size = cms_space->free(); 2057 assert(free_size == 2058 pointer_delta(cms_space->end(), cms_space->compaction_top()) 2059 * HeapWordSize, 2060 "All the free space should be compacted into one chunk at top"); 2061 assert(cms_space->dictionary()->total_chunk_size( 2062 debug_only(cms_space->freelistLock())) == 0 || 2063 cms_space->totalSizeInIndexedFreeLists() == 0, 2064 "All the free space should be in a single chunk"); 2065 size_t num = cms_space->totalCount(); 2066 assert((free_size == 0 && num == 0) || 2067 (free_size > 0 && (num == 1 || num == 2)), 2068 "There should be at most 2 free chunks after compaction"); 2069 #endif // ASSERT 2070 _collectorState = Resetting; 2071 assert(_restart_addr == NULL, 2072 "Should have been NULL'd before baton was passed"); 2073 reset(false /* == !asynch */); 2074 _cmsGen->reset_after_compaction(); 2075 _concurrent_cycles_since_last_unload = 0; 2076 2077 if (verifying() && !should_unload_classes()) { 2078 perm_gen_verify_bit_map()->clear_all(); 2079 } 2080 2081 // Clear any data recorded in the PLAB chunk arrays. 2082 if (_survivor_plab_array != NULL) { 2083 reset_survivor_plab_arrays(); 2084 } 2085 2086 // Adjust the per-size allocation stats for the next epoch. 2087 _cmsGen->cmsSpace()->endSweepFLCensus(sweep_count() /* fake */); 2088 // Restart the "inter sweep timer" for the next epoch. 2089 _inter_sweep_timer.reset(); 2090 _inter_sweep_timer.start(); 2091 2092 // Sample collection pause time and reset for collection interval. 2093 if (UseAdaptiveSizePolicy) { 2094 size_policy()->msc_collection_end(gch->gc_cause()); 2095 } 2096 2097 gc_timer->register_gc_end(os::elapsed_counter()); 2098 2099 gc_tracer->report_gc_end(gc_timer->gc_end(), gc_timer->time_partitions()); 2100 2101 // For a mark-sweep-compact, compute_new_size() will be called 2102 // in the heap's do_collection() method. 2103 } 2104 2105 // A work method used by the foreground collector to do 2106 // a mark-sweep, after taking over from a possibly on-going 2107 // concurrent mark-sweep collection. 2108 void CMSCollector::do_mark_sweep_work(bool clear_all_soft_refs, 2109 CollectorState first_state, bool should_start_over) { 2110 if (PrintGC && Verbose) { 2111 gclog_or_tty->print_cr("Pass concurrent collection to foreground " 2112 "collector with count %d", 2113 _full_gcs_since_conc_gc); 2114 } 2115 switch (_collectorState) { 2116 case Idling: 2117 if (first_state == Idling || should_start_over) { 2118 // The background GC was not active, or should 2119 // restarted from scratch; start the cycle. 2120 _collectorState = InitialMarking; 2121 } 2122 // If first_state was not Idling, then a background GC 2123 // was in progress and has now finished. No need to do it 2124 // again. Leave the state as Idling. 2125 break; 2126 case Precleaning: 2127 // In the foreground case don't do the precleaning since 2128 // it is not done concurrently and there is extra work 2129 // required. 2130 _collectorState = FinalMarking; 2131 } 2132 collect_in_foreground(clear_all_soft_refs, GenCollectedHeap::heap()->gc_cause()); 2133 2134 // For a mark-sweep, compute_new_size() will be called 2135 // in the heap's do_collection() method. 2136 } 2137 2138 2139 void CMSCollector::getFreelistLocks() const { 2140 // Get locks for all free lists in all generations that this 2141 // collector is responsible for 2142 _cmsGen->freelistLock()->lock_without_safepoint_check(); 2143 _permGen->freelistLock()->lock_without_safepoint_check(); 2144 } 2145 2146 void CMSCollector::releaseFreelistLocks() const { 2147 // Release locks for all free lists in all generations that this 2148 // collector is responsible for 2149 _cmsGen->freelistLock()->unlock(); 2150 _permGen->freelistLock()->unlock(); 2151 } 2152 2153 bool CMSCollector::haveFreelistLocks() const { 2154 // Check locks for all free lists in all generations that this 2155 // collector is responsible for 2156 assert_lock_strong(_cmsGen->freelistLock()); 2157 assert_lock_strong(_permGen->freelistLock()); 2158 PRODUCT_ONLY(ShouldNotReachHere()); 2159 return true; 2160 } 2161 2162 // A utility class that is used by the CMS collector to 2163 // temporarily "release" the foreground collector from its 2164 // usual obligation to wait for the background collector to 2165 // complete an ongoing phase before proceeding. 2166 class ReleaseForegroundGC: public StackObj { 2167 private: 2168 CMSCollector* _c; 2169 public: 2170 ReleaseForegroundGC(CMSCollector* c) : _c(c) { 2171 assert(_c->_foregroundGCShouldWait, "Else should not need to call"); 2172 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 2173 // allow a potentially blocked foreground collector to proceed 2174 _c->_foregroundGCShouldWait = false; 2175 if (_c->_foregroundGCIsActive) { 2176 CGC_lock->notify(); 2177 } 2178 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 2179 "Possible deadlock"); 2180 } 2181 2182 ~ReleaseForegroundGC() { 2183 assert(!_c->_foregroundGCShouldWait, "Usage protocol violation?"); 2184 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 2185 _c->_foregroundGCShouldWait = true; 2186 } 2187 }; 2188 2189 // There are separate collect_in_background and collect_in_foreground because of 2190 // the different locking requirements of the background collector and the 2191 // foreground collector. There was originally an attempt to share 2192 // one "collect" method between the background collector and the foreground 2193 // collector but the if-then-else required made it cleaner to have 2194 // separate methods. 2195 void CMSCollector::collect_in_background(bool clear_all_soft_refs, GCCause::Cause cause) { 2196 assert(Thread::current()->is_ConcurrentGC_thread(), 2197 "A CMS asynchronous collection is only allowed on a CMS thread."); 2198 2199 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2200 { 2201 bool safepoint_check = Mutex::_no_safepoint_check_flag; 2202 MutexLockerEx hl(Heap_lock, safepoint_check); 2203 FreelistLocker fll(this); 2204 MutexLockerEx x(CGC_lock, safepoint_check); 2205 if (_foregroundGCIsActive || !UseAsyncConcMarkSweepGC) { 2206 // The foreground collector is active or we're 2207 // not using asynchronous collections. Skip this 2208 // background collection. 2209 assert(!_foregroundGCShouldWait, "Should be clear"); 2210 return; 2211 } else { 2212 assert(_collectorState == Idling, "Should be idling before start."); 2213 _collectorState = InitialMarking; 2214 register_gc_start(cause); 2215 // Reset the expansion cause, now that we are about to begin 2216 // a new cycle. 2217 clear_expansion_cause(); 2218 } 2219 // Decide if we want to enable class unloading as part of the 2220 // ensuing concurrent GC cycle. 2221 update_should_unload_classes(); 2222 _full_gc_requested = false; // acks all outstanding full gc requests 2223 _full_gc_cause = GCCause::_no_gc; 2224 // Signal that we are about to start a collection 2225 gch->increment_total_full_collections(); // ... starting a collection cycle 2226 _collection_count_start = gch->total_full_collections(); 2227 } 2228 2229 // Used for PrintGC 2230 size_t prev_used; 2231 if (PrintGC && Verbose) { 2232 prev_used = _cmsGen->used(); // XXXPERM 2233 } 2234 2235 // The change of the collection state is normally done at this level; 2236 // the exceptions are phases that are executed while the world is 2237 // stopped. For those phases the change of state is done while the 2238 // world is stopped. For baton passing purposes this allows the 2239 // background collector to finish the phase and change state atomically. 2240 // The foreground collector cannot wait on a phase that is done 2241 // while the world is stopped because the foreground collector already 2242 // has the world stopped and would deadlock. 2243 while (_collectorState != Idling) { 2244 if (TraceCMSState) { 2245 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " in CMS state %d", 2246 Thread::current(), _collectorState); 2247 } 2248 // The foreground collector 2249 // holds the Heap_lock throughout its collection. 2250 // holds the CMS token (but not the lock) 2251 // except while it is waiting for the background collector to yield. 2252 // 2253 // The foreground collector should be blocked (not for long) 2254 // if the background collector is about to start a phase 2255 // executed with world stopped. If the background 2256 // collector has already started such a phase, the 2257 // foreground collector is blocked waiting for the 2258 // Heap_lock. The stop-world phases (InitialMarking and FinalMarking) 2259 // are executed in the VM thread. 2260 // 2261 // The locking order is 2262 // PendingListLock (PLL) -- if applicable (FinalMarking) 2263 // Heap_lock (both this & PLL locked in VM_CMS_Operation::prologue()) 2264 // CMS token (claimed in 2265 // stop_world_and_do() --> 2266 // safepoint_synchronize() --> 2267 // CMSThread::synchronize()) 2268 2269 { 2270 // Check if the FG collector wants us to yield. 2271 CMSTokenSync x(true); // is cms thread 2272 if (waitForForegroundGC()) { 2273 // We yielded to a foreground GC, nothing more to be 2274 // done this round. 2275 assert(_foregroundGCShouldWait == false, "We set it to false in " 2276 "waitForForegroundGC()"); 2277 if (TraceCMSState) { 2278 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT 2279 " exiting collection CMS state %d", 2280 Thread::current(), _collectorState); 2281 } 2282 return; 2283 } else { 2284 // The background collector can run but check to see if the 2285 // foreground collector has done a collection while the 2286 // background collector was waiting to get the CGC_lock 2287 // above. If yes, break so that _foregroundGCShouldWait 2288 // is cleared before returning. 2289 if (_collectorState == Idling) { 2290 break; 2291 } 2292 } 2293 } 2294 2295 assert(_foregroundGCShouldWait, "Foreground collector, if active, " 2296 "should be waiting"); 2297 2298 switch (_collectorState) { 2299 case InitialMarking: 2300 { 2301 ReleaseForegroundGC x(this); 2302 stats().record_cms_begin(); 2303 VM_CMS_Initial_Mark initial_mark_op(this); 2304 VMThread::execute(&initial_mark_op); 2305 } 2306 // The collector state may be any legal state at this point 2307 // since the background collector may have yielded to the 2308 // foreground collector. 2309 break; 2310 case Marking: 2311 // initial marking in checkpointRootsInitialWork has been completed 2312 if (markFromRoots(true)) { // we were successful 2313 assert(_collectorState == Precleaning, "Collector state should " 2314 "have changed"); 2315 } else { 2316 assert(_foregroundGCIsActive, "Internal state inconsistency"); 2317 } 2318 break; 2319 case Precleaning: 2320 if (UseAdaptiveSizePolicy) { 2321 size_policy()->concurrent_precleaning_begin(); 2322 } 2323 // marking from roots in markFromRoots has been completed 2324 preclean(); 2325 if (UseAdaptiveSizePolicy) { 2326 size_policy()->concurrent_precleaning_end(); 2327 } 2328 assert(_collectorState == AbortablePreclean || 2329 _collectorState == FinalMarking, 2330 "Collector state should have changed"); 2331 break; 2332 case AbortablePreclean: 2333 if (UseAdaptiveSizePolicy) { 2334 size_policy()->concurrent_phases_resume(); 2335 } 2336 abortable_preclean(); 2337 if (UseAdaptiveSizePolicy) { 2338 size_policy()->concurrent_precleaning_end(); 2339 } 2340 assert(_collectorState == FinalMarking, "Collector state should " 2341 "have changed"); 2342 break; 2343 case FinalMarking: 2344 { 2345 ReleaseForegroundGC x(this); 2346 2347 VM_CMS_Final_Remark final_remark_op(this); 2348 VMThread::execute(&final_remark_op); 2349 } 2350 assert(_foregroundGCShouldWait, "block post-condition"); 2351 break; 2352 case Sweeping: 2353 if (UseAdaptiveSizePolicy) { 2354 size_policy()->concurrent_sweeping_begin(); 2355 } 2356 // final marking in checkpointRootsFinal has been completed 2357 sweep(true); 2358 assert(_collectorState == Resizing, "Collector state change " 2359 "to Resizing must be done under the free_list_lock"); 2360 _full_gcs_since_conc_gc = 0; 2361 2362 // Stop the timers for adaptive size policy for the concurrent phases 2363 if (UseAdaptiveSizePolicy) { 2364 size_policy()->concurrent_sweeping_end(); 2365 size_policy()->concurrent_phases_end(gch->gc_cause(), 2366 gch->prev_gen(_cmsGen)->capacity(), 2367 _cmsGen->free()); 2368 } 2369 2370 case Resizing: { 2371 // Sweeping has been completed... 2372 // At this point the background collection has completed. 2373 // Don't move the call to compute_new_size() down 2374 // into code that might be executed if the background 2375 // collection was preempted. 2376 { 2377 ReleaseForegroundGC x(this); // unblock FG collection 2378 MutexLockerEx y(Heap_lock, Mutex::_no_safepoint_check_flag); 2379 CMSTokenSync z(true); // not strictly needed. 2380 if (_collectorState == Resizing) { 2381 compute_new_size(); 2382 save_heap_summary(); 2383 _collectorState = Resetting; 2384 } else { 2385 assert(_collectorState == Idling, "The state should only change" 2386 " because the foreground collector has finished the collection"); 2387 } 2388 } 2389 break; 2390 } 2391 case Resetting: 2392 // CMS heap resizing has been completed 2393 reset(true); 2394 assert(_collectorState == Idling, "Collector state should " 2395 "have changed"); 2396 stats().record_cms_end(); 2397 // Don't move the concurrent_phases_end() and compute_new_size() 2398 // calls to here because a preempted background collection 2399 // has it's state set to "Resetting". 2400 break; 2401 case Idling: 2402 default: 2403 ShouldNotReachHere(); 2404 break; 2405 } 2406 if (TraceCMSState) { 2407 gclog_or_tty->print_cr(" Thread " INTPTR_FORMAT " done - next CMS state %d", 2408 Thread::current(), _collectorState); 2409 } 2410 assert(_foregroundGCShouldWait, "block post-condition"); 2411 } 2412 2413 // Should this be in gc_epilogue? 2414 collector_policy()->counters()->update_counters(); 2415 2416 { 2417 // Clear _foregroundGCShouldWait and, in the event that the 2418 // foreground collector is waiting, notify it, before 2419 // returning. 2420 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 2421 _foregroundGCShouldWait = false; 2422 if (_foregroundGCIsActive) { 2423 CGC_lock->notify(); 2424 } 2425 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 2426 "Possible deadlock"); 2427 } 2428 if (TraceCMSState) { 2429 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT 2430 " exiting collection CMS state %d", 2431 Thread::current(), _collectorState); 2432 } 2433 if (PrintGC && Verbose) { 2434 _cmsGen->print_heap_change(prev_used); 2435 } 2436 } 2437 2438 void CMSCollector::register_foreground_gc_start(GCCause::Cause cause) { 2439 if (!_cms_start_registered) { 2440 register_gc_start(cause); 2441 } 2442 } 2443 2444 void CMSCollector::register_gc_start(GCCause::Cause cause) { 2445 _cms_start_registered = true; 2446 _gc_timer_cm->register_gc_start(os::elapsed_counter()); 2447 _gc_tracer_cm->report_gc_start(cause, _gc_timer_cm->gc_start()); 2448 } 2449 2450 void CMSCollector::register_gc_end() { 2451 if (_cms_start_registered) { 2452 report_heap_summary(GCWhen::AfterGC); 2453 2454 _gc_timer_cm->register_gc_end(os::elapsed_counter()); 2455 _gc_tracer_cm->report_gc_end(_gc_timer_cm->gc_end(), _gc_timer_cm->time_partitions()); 2456 _cms_start_registered = false; 2457 } 2458 } 2459 2460 void CMSCollector::save_heap_summary() { 2461 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2462 _last_heap_summary = gch->create_heap_summary(); 2463 _last_perm_gen_summary = gch->create_perm_gen_summary(); 2464 } 2465 2466 void CMSCollector::report_heap_summary(GCWhen::Type when) { 2467 _gc_tracer_cm->report_gc_heap_summary(when, _last_heap_summary, _last_perm_gen_summary); 2468 } 2469 2470 void CMSCollector::collect_in_foreground(bool clear_all_soft_refs, GCCause::Cause cause) { 2471 assert(_foregroundGCIsActive && !_foregroundGCShouldWait, 2472 "Foreground collector should be waiting, not executing"); 2473 assert(Thread::current()->is_VM_thread(), "A foreground collection" 2474 "may only be done by the VM Thread with the world stopped"); 2475 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 2476 "VM thread should have CMS token"); 2477 2478 NOT_PRODUCT(GCTraceTime t("CMS:MS (foreground) ", PrintGCDetails && Verbose, 2479 true, NULL);) 2480 if (UseAdaptiveSizePolicy) { 2481 size_policy()->ms_collection_begin(); 2482 } 2483 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact); 2484 2485 HandleMark hm; // Discard invalid handles created during verification 2486 2487 if (VerifyBeforeGC && 2488 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2489 Universe::verify(); 2490 } 2491 2492 // Snapshot the soft reference policy to be used in this collection cycle. 2493 ref_processor()->setup_policy(clear_all_soft_refs); 2494 2495 bool init_mark_was_synchronous = false; // until proven otherwise 2496 while (_collectorState != Idling) { 2497 if (TraceCMSState) { 2498 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " in CMS state %d", 2499 Thread::current(), _collectorState); 2500 } 2501 switch (_collectorState) { 2502 case InitialMarking: 2503 register_foreground_gc_start(cause); 2504 init_mark_was_synchronous = true; // fact to be exploited in re-mark 2505 checkpointRootsInitial(false); 2506 assert(_collectorState == Marking, "Collector state should have changed" 2507 " within checkpointRootsInitial()"); 2508 break; 2509 case Marking: 2510 // initial marking in checkpointRootsInitialWork has been completed 2511 if (VerifyDuringGC && 2512 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2513 gclog_or_tty->print("Verify before initial mark: "); 2514 Universe::verify(); 2515 } 2516 { 2517 bool res = markFromRoots(false); 2518 assert(res && _collectorState == FinalMarking, "Collector state should " 2519 "have changed"); 2520 break; 2521 } 2522 case FinalMarking: 2523 if (VerifyDuringGC && 2524 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2525 gclog_or_tty->print("Verify before re-mark: "); 2526 Universe::verify(); 2527 } 2528 checkpointRootsFinal(false, clear_all_soft_refs, 2529 init_mark_was_synchronous); 2530 assert(_collectorState == Sweeping, "Collector state should not " 2531 "have changed within checkpointRootsFinal()"); 2532 break; 2533 case Sweeping: 2534 // final marking in checkpointRootsFinal has been completed 2535 if (VerifyDuringGC && 2536 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2537 gclog_or_tty->print("Verify before sweep: "); 2538 Universe::verify(); 2539 } 2540 sweep(false); 2541 assert(_collectorState == Resizing, "Incorrect state"); 2542 break; 2543 case Resizing: { 2544 // Sweeping has been completed; the actual resize in this case 2545 // is done separately; nothing to be done in this state. 2546 _collectorState = Resetting; 2547 break; 2548 } 2549 case Resetting: 2550 // The heap has been resized. 2551 if (VerifyDuringGC && 2552 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2553 gclog_or_tty->print("Verify before reset: "); 2554 Universe::verify(); 2555 } 2556 save_heap_summary(); 2557 reset(false); 2558 assert(_collectorState == Idling, "Collector state should " 2559 "have changed"); 2560 break; 2561 case Precleaning: 2562 case AbortablePreclean: 2563 // Elide the preclean phase 2564 _collectorState = FinalMarking; 2565 break; 2566 default: 2567 ShouldNotReachHere(); 2568 } 2569 if (TraceCMSState) { 2570 gclog_or_tty->print_cr(" Thread " INTPTR_FORMAT " done - next CMS state %d", 2571 Thread::current(), _collectorState); 2572 } 2573 } 2574 2575 if (UseAdaptiveSizePolicy) { 2576 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2577 size_policy()->ms_collection_end(gch->gc_cause()); 2578 } 2579 2580 if (VerifyAfterGC && 2581 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2582 Universe::verify(); 2583 } 2584 if (TraceCMSState) { 2585 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT 2586 " exiting collection CMS state %d", 2587 Thread::current(), _collectorState); 2588 } 2589 } 2590 2591 bool CMSCollector::waitForForegroundGC() { 2592 bool res = false; 2593 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 2594 "CMS thread should have CMS token"); 2595 // Block the foreground collector until the 2596 // background collectors decides whether to 2597 // yield. 2598 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 2599 _foregroundGCShouldWait = true; 2600 if (_foregroundGCIsActive) { 2601 // The background collector yields to the 2602 // foreground collector and returns a value 2603 // indicating that it has yielded. The foreground 2604 // collector can proceed. 2605 res = true; 2606 _foregroundGCShouldWait = false; 2607 ConcurrentMarkSweepThread::clear_CMS_flag( 2608 ConcurrentMarkSweepThread::CMS_cms_has_token); 2609 ConcurrentMarkSweepThread::set_CMS_flag( 2610 ConcurrentMarkSweepThread::CMS_cms_wants_token); 2611 // Get a possibly blocked foreground thread going 2612 CGC_lock->notify(); 2613 if (TraceCMSState) { 2614 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " waiting at CMS state %d", 2615 Thread::current(), _collectorState); 2616 } 2617 while (_foregroundGCIsActive) { 2618 CGC_lock->wait(Mutex::_no_safepoint_check_flag); 2619 } 2620 ConcurrentMarkSweepThread::set_CMS_flag( 2621 ConcurrentMarkSweepThread::CMS_cms_has_token); 2622 ConcurrentMarkSweepThread::clear_CMS_flag( 2623 ConcurrentMarkSweepThread::CMS_cms_wants_token); 2624 } 2625 if (TraceCMSState) { 2626 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " continuing at CMS state %d", 2627 Thread::current(), _collectorState); 2628 } 2629 return res; 2630 } 2631 2632 // Because of the need to lock the free lists and other structures in 2633 // the collector, common to all the generations that the collector is 2634 // collecting, we need the gc_prologues of individual CMS generations 2635 // delegate to their collector. It may have been simpler had the 2636 // current infrastructure allowed one to call a prologue on a 2637 // collector. In the absence of that we have the generation's 2638 // prologue delegate to the collector, which delegates back 2639 // some "local" work to a worker method in the individual generations 2640 // that it's responsible for collecting, while itself doing any 2641 // work common to all generations it's responsible for. A similar 2642 // comment applies to the gc_epilogue()'s. 2643 // The role of the varaible _between_prologue_and_epilogue is to 2644 // enforce the invocation protocol. 2645 void CMSCollector::gc_prologue(bool full) { 2646 // Call gc_prologue_work() for each CMSGen and PermGen that 2647 // we are responsible for. 2648 2649 // The following locking discipline assumes that we are only called 2650 // when the world is stopped. 2651 assert(SafepointSynchronize::is_at_safepoint(), "world is stopped assumption"); 2652 2653 // The CMSCollector prologue must call the gc_prologues for the 2654 // "generations" (including PermGen if any) that it's responsible 2655 // for. 2656 2657 assert( Thread::current()->is_VM_thread() 2658 || ( CMSScavengeBeforeRemark 2659 && Thread::current()->is_ConcurrentGC_thread()), 2660 "Incorrect thread type for prologue execution"); 2661 2662 if (_between_prologue_and_epilogue) { 2663 // We have already been invoked; this is a gc_prologue delegation 2664 // from yet another CMS generation that we are responsible for, just 2665 // ignore it since all relevant work has already been done. 2666 return; 2667 } 2668 2669 // set a bit saying prologue has been called; cleared in epilogue 2670 _between_prologue_and_epilogue = true; 2671 // Claim locks for common data structures, then call gc_prologue_work() 2672 // for each CMSGen and PermGen that we are responsible for. 2673 2674 getFreelistLocks(); // gets free list locks on constituent spaces 2675 bitMapLock()->lock_without_safepoint_check(); 2676 2677 // Should call gc_prologue_work() for all cms gens we are responsible for 2678 bool registerClosure = _collectorState >= Marking 2679 && _collectorState < Sweeping; 2680 ModUnionClosure* muc = CollectedHeap::use_parallel_gc_threads() ? 2681 &_modUnionClosurePar 2682 : &_modUnionClosure; 2683 _cmsGen->gc_prologue_work(full, registerClosure, muc); 2684 _permGen->gc_prologue_work(full, registerClosure, muc); 2685 2686 if (!full) { 2687 stats().record_gc0_begin(); 2688 } 2689 } 2690 2691 void ConcurrentMarkSweepGeneration::gc_prologue(bool full) { 2692 // Delegate to CMScollector which knows how to coordinate between 2693 // this and any other CMS generations that it is responsible for 2694 // collecting. 2695 collector()->gc_prologue(full); 2696 } 2697 2698 // This is a "private" interface for use by this generation's CMSCollector. 2699 // Not to be called directly by any other entity (for instance, 2700 // GenCollectedHeap, which calls the "public" gc_prologue method above). 2701 void ConcurrentMarkSweepGeneration::gc_prologue_work(bool full, 2702 bool registerClosure, ModUnionClosure* modUnionClosure) { 2703 assert(!incremental_collection_failed(), "Shouldn't be set yet"); 2704 assert(cmsSpace()->preconsumptionDirtyCardClosure() == NULL, 2705 "Should be NULL"); 2706 if (registerClosure) { 2707 cmsSpace()->setPreconsumptionDirtyCardClosure(modUnionClosure); 2708 } 2709 cmsSpace()->gc_prologue(); 2710 // Clear stat counters 2711 NOT_PRODUCT( 2712 assert(_numObjectsPromoted == 0, "check"); 2713 assert(_numWordsPromoted == 0, "check"); 2714 if (Verbose && PrintGC) { 2715 gclog_or_tty->print("Allocated "SIZE_FORMAT" objects, " 2716 SIZE_FORMAT" bytes concurrently", 2717 _numObjectsAllocated, _numWordsAllocated*sizeof(HeapWord)); 2718 } 2719 _numObjectsAllocated = 0; 2720 _numWordsAllocated = 0; 2721 ) 2722 } 2723 2724 void CMSCollector::gc_epilogue(bool full) { 2725 // The following locking discipline assumes that we are only called 2726 // when the world is stopped. 2727 assert(SafepointSynchronize::is_at_safepoint(), 2728 "world is stopped assumption"); 2729 2730 // Currently the CMS epilogue (see CompactibleFreeListSpace) merely checks 2731 // if linear allocation blocks need to be appropriately marked to allow the 2732 // the blocks to be parsable. We also check here whether we need to nudge the 2733 // CMS collector thread to start a new cycle (if it's not already active). 2734 assert( Thread::current()->is_VM_thread() 2735 || ( CMSScavengeBeforeRemark 2736 && Thread::current()->is_ConcurrentGC_thread()), 2737 "Incorrect thread type for epilogue execution"); 2738 2739 if (!_between_prologue_and_epilogue) { 2740 // We have already been invoked; this is a gc_epilogue delegation 2741 // from yet another CMS generation that we are responsible for, just 2742 // ignore it since all relevant work has already been done. 2743 return; 2744 } 2745 assert(haveFreelistLocks(), "must have freelist locks"); 2746 assert_lock_strong(bitMapLock()); 2747 2748 _cmsGen->gc_epilogue_work(full); 2749 _permGen->gc_epilogue_work(full); 2750 2751 if (_collectorState == AbortablePreclean || _collectorState == Precleaning) { 2752 // in case sampling was not already enabled, enable it 2753 _start_sampling = true; 2754 } 2755 // reset _eden_chunk_array so sampling starts afresh 2756 _eden_chunk_index = 0; 2757 2758 size_t cms_used = _cmsGen->cmsSpace()->used(); 2759 size_t perm_used = _permGen->cmsSpace()->used(); 2760 2761 // update performance counters - this uses a special version of 2762 // update_counters() that allows the utilization to be passed as a 2763 // parameter, avoiding multiple calls to used(). 2764 // 2765 _cmsGen->update_counters(cms_used); 2766 _permGen->update_counters(perm_used); 2767 2768 if (CMSIncrementalMode) { 2769 icms_update_allocation_limits(); 2770 } 2771 2772 bitMapLock()->unlock(); 2773 releaseFreelistLocks(); 2774 2775 if (!CleanChunkPoolAsync) { 2776 Chunk::clean_chunk_pool(); 2777 } 2778 2779 _between_prologue_and_epilogue = false; // ready for next cycle 2780 } 2781 2782 void ConcurrentMarkSweepGeneration::gc_epilogue(bool full) { 2783 collector()->gc_epilogue(full); 2784 2785 // Also reset promotion tracking in par gc thread states. 2786 if (CollectedHeap::use_parallel_gc_threads()) { 2787 for (uint i = 0; i < ParallelGCThreads; i++) { 2788 _par_gc_thread_states[i]->promo.stopTrackingPromotions(i); 2789 } 2790 } 2791 } 2792 2793 void ConcurrentMarkSweepGeneration::gc_epilogue_work(bool full) { 2794 assert(!incremental_collection_failed(), "Should have been cleared"); 2795 cmsSpace()->setPreconsumptionDirtyCardClosure(NULL); 2796 cmsSpace()->gc_epilogue(); 2797 // Print stat counters 2798 NOT_PRODUCT( 2799 assert(_numObjectsAllocated == 0, "check"); 2800 assert(_numWordsAllocated == 0, "check"); 2801 if (Verbose && PrintGC) { 2802 gclog_or_tty->print("Promoted "SIZE_FORMAT" objects, " 2803 SIZE_FORMAT" bytes", 2804 _numObjectsPromoted, _numWordsPromoted*sizeof(HeapWord)); 2805 } 2806 _numObjectsPromoted = 0; 2807 _numWordsPromoted = 0; 2808 ) 2809 2810 if (PrintGC && Verbose) { 2811 // Call down the chain in contiguous_available needs the freelistLock 2812 // so print this out before releasing the freeListLock. 2813 gclog_or_tty->print(" Contiguous available "SIZE_FORMAT" bytes ", 2814 contiguous_available()); 2815 } 2816 } 2817 2818 #ifndef PRODUCT 2819 bool CMSCollector::have_cms_token() { 2820 Thread* thr = Thread::current(); 2821 if (thr->is_VM_thread()) { 2822 return ConcurrentMarkSweepThread::vm_thread_has_cms_token(); 2823 } else if (thr->is_ConcurrentGC_thread()) { 2824 return ConcurrentMarkSweepThread::cms_thread_has_cms_token(); 2825 } else if (thr->is_GC_task_thread()) { 2826 return ConcurrentMarkSweepThread::vm_thread_has_cms_token() && 2827 ParGCRareEvent_lock->owned_by_self(); 2828 } 2829 return false; 2830 } 2831 #endif 2832 2833 // Check reachability of the given heap address in CMS generation, 2834 // treating all other generations as roots. 2835 bool CMSCollector::is_cms_reachable(HeapWord* addr) { 2836 // We could "guarantee" below, rather than assert, but i'll 2837 // leave these as "asserts" so that an adventurous debugger 2838 // could try this in the product build provided some subset of 2839 // the conditions were met, provided they were intersted in the 2840 // results and knew that the computation below wouldn't interfere 2841 // with other concurrent computations mutating the structures 2842 // being read or written. 2843 assert(SafepointSynchronize::is_at_safepoint(), 2844 "Else mutations in object graph will make answer suspect"); 2845 assert(have_cms_token(), "Should hold cms token"); 2846 assert(haveFreelistLocks(), "must hold free list locks"); 2847 assert_lock_strong(bitMapLock()); 2848 2849 // Clear the marking bit map array before starting, but, just 2850 // for kicks, first report if the given address is already marked 2851 gclog_or_tty->print_cr("Start: Address 0x%x is%s marked", addr, 2852 _markBitMap.isMarked(addr) ? "" : " not"); 2853 2854 if (verify_after_remark()) { 2855 MutexLockerEx x(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag); 2856 bool result = verification_mark_bm()->isMarked(addr); 2857 gclog_or_tty->print_cr("TransitiveMark: Address 0x%x %s marked", addr, 2858 result ? "IS" : "is NOT"); 2859 return result; 2860 } else { 2861 gclog_or_tty->print_cr("Could not compute result"); 2862 return false; 2863 } 2864 } 2865 2866 //////////////////////////////////////////////////////// 2867 // CMS Verification Support 2868 //////////////////////////////////////////////////////// 2869 // Following the remark phase, the following invariant 2870 // should hold -- each object in the CMS heap which is 2871 // marked in markBitMap() should be marked in the verification_mark_bm(). 2872 2873 class VerifyMarkedClosure: public BitMapClosure { 2874 CMSBitMap* _marks; 2875 bool _failed; 2876 2877 public: 2878 VerifyMarkedClosure(CMSBitMap* bm): _marks(bm), _failed(false) {} 2879 2880 bool do_bit(size_t offset) { 2881 HeapWord* addr = _marks->offsetToHeapWord(offset); 2882 if (!_marks->isMarked(addr)) { 2883 oop(addr)->print_on(gclog_or_tty); 2884 gclog_or_tty->print_cr(" ("INTPTR_FORMAT" should have been marked)", addr); 2885 _failed = true; 2886 } 2887 return true; 2888 } 2889 2890 bool failed() { return _failed; } 2891 }; 2892 2893 bool CMSCollector::verify_after_remark() { 2894 gclog_or_tty->print(" [Verifying CMS Marking... "); 2895 MutexLockerEx ml(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag); 2896 static bool init = false; 2897 2898 assert(SafepointSynchronize::is_at_safepoint(), 2899 "Else mutations in object graph will make answer suspect"); 2900 assert(have_cms_token(), 2901 "Else there may be mutual interference in use of " 2902 " verification data structures"); 2903 assert(_collectorState > Marking && _collectorState <= Sweeping, 2904 "Else marking info checked here may be obsolete"); 2905 assert(haveFreelistLocks(), "must hold free list locks"); 2906 assert_lock_strong(bitMapLock()); 2907 2908 2909 // Allocate marking bit map if not already allocated 2910 if (!init) { // first time 2911 if (!verification_mark_bm()->allocate(_span)) { 2912 return false; 2913 } 2914 init = true; 2915 } 2916 2917 assert(verification_mark_stack()->isEmpty(), "Should be empty"); 2918 2919 // Turn off refs discovery -- so we will be tracing through refs. 2920 // This is as intended, because by this time 2921 // GC must already have cleared any refs that need to be cleared, 2922 // and traced those that need to be marked; moreover, 2923 // the marking done here is not going to intefere in any 2924 // way with the marking information used by GC. 2925 NoRefDiscovery no_discovery(ref_processor()); 2926 2927 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;) 2928 2929 // Clear any marks from a previous round 2930 verification_mark_bm()->clear_all(); 2931 assert(verification_mark_stack()->isEmpty(), "markStack should be empty"); 2932 verify_work_stacks_empty(); 2933 2934 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2935 gch->ensure_parsability(false); // fill TLABs, but no need to retire them 2936 // Update the saved marks which may affect the root scans. 2937 gch->save_marks(); 2938 2939 if (CMSRemarkVerifyVariant == 1) { 2940 // In this first variant of verification, we complete 2941 // all marking, then check if the new marks-verctor is 2942 // a subset of the CMS marks-vector. 2943 verify_after_remark_work_1(); 2944 } else if (CMSRemarkVerifyVariant == 2) { 2945 // In this second variant of verification, we flag an error 2946 // (i.e. an object reachable in the new marks-vector not reachable 2947 // in the CMS marks-vector) immediately, also indicating the 2948 // identify of an object (A) that references the unmarked object (B) -- 2949 // presumably, a mutation to A failed to be picked up by preclean/remark? 2950 verify_after_remark_work_2(); 2951 } else { 2952 warning("Unrecognized value %d for CMSRemarkVerifyVariant", 2953 CMSRemarkVerifyVariant); 2954 } 2955 gclog_or_tty->print(" done] "); 2956 return true; 2957 } 2958 2959 void CMSCollector::verify_after_remark_work_1() { 2960 ResourceMark rm; 2961 HandleMark hm; 2962 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2963 2964 // Mark from roots one level into CMS 2965 MarkRefsIntoClosure notOlder(_span, verification_mark_bm()); 2966 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2967 2968 gch->gen_process_strong_roots(_cmsGen->level(), 2969 true, // younger gens are roots 2970 true, // activate StrongRootsScope 2971 true, // collecting perm gen 2972 SharedHeap::ScanningOption(roots_scanning_options()), 2973 ¬Older, 2974 true, // walk code active on stacks 2975 NULL); 2976 2977 // Now mark from the roots 2978 assert(_revisitStack.isEmpty(), "Should be empty"); 2979 MarkFromRootsClosure markFromRootsClosure(this, _span, 2980 verification_mark_bm(), verification_mark_stack(), &_revisitStack, 2981 false /* don't yield */, true /* verifying */); 2982 assert(_restart_addr == NULL, "Expected pre-condition"); 2983 verification_mark_bm()->iterate(&markFromRootsClosure); 2984 while (_restart_addr != NULL) { 2985 // Deal with stack overflow: by restarting at the indicated 2986 // address. 2987 HeapWord* ra = _restart_addr; 2988 markFromRootsClosure.reset(ra); 2989 _restart_addr = NULL; 2990 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end()); 2991 } 2992 assert(verification_mark_stack()->isEmpty(), "Should have been drained"); 2993 verify_work_stacks_empty(); 2994 // Should reset the revisit stack above, since no class tree 2995 // surgery is forthcoming. 2996 _revisitStack.reset(); // throwing away all contents 2997 2998 // Marking completed -- now verify that each bit marked in 2999 // verification_mark_bm() is also marked in markBitMap(); flag all 3000 // errors by printing corresponding objects. 3001 VerifyMarkedClosure vcl(markBitMap()); 3002 verification_mark_bm()->iterate(&vcl); 3003 if (vcl.failed()) { 3004 gclog_or_tty->print("Verification failed"); 3005 Universe::heap()->print_on(gclog_or_tty); 3006 fatal("CMS: failed marking verification after remark"); 3007 } 3008 } 3009 3010 void CMSCollector::verify_after_remark_work_2() { 3011 ResourceMark rm; 3012 HandleMark hm; 3013 GenCollectedHeap* gch = GenCollectedHeap::heap(); 3014 3015 // Mark from roots one level into CMS 3016 MarkRefsIntoVerifyClosure notOlder(_span, verification_mark_bm(), 3017 markBitMap()); 3018 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 3019 gch->gen_process_strong_roots(_cmsGen->level(), 3020 true, // younger gens are roots 3021 true, // activate StrongRootsScope 3022 true, // collecting perm gen 3023 SharedHeap::ScanningOption(roots_scanning_options()), 3024 ¬Older, 3025 true, // walk code active on stacks 3026 NULL); 3027 3028 // Now mark from the roots 3029 assert(_revisitStack.isEmpty(), "Should be empty"); 3030 MarkFromRootsVerifyClosure markFromRootsClosure(this, _span, 3031 verification_mark_bm(), markBitMap(), verification_mark_stack()); 3032 assert(_restart_addr == NULL, "Expected pre-condition"); 3033 verification_mark_bm()->iterate(&markFromRootsClosure); 3034 while (_restart_addr != NULL) { 3035 // Deal with stack overflow: by restarting at the indicated 3036 // address. 3037 HeapWord* ra = _restart_addr; 3038 markFromRootsClosure.reset(ra); 3039 _restart_addr = NULL; 3040 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end()); 3041 } 3042 assert(verification_mark_stack()->isEmpty(), "Should have been drained"); 3043 verify_work_stacks_empty(); 3044 // Should reset the revisit stack above, since no class tree 3045 // surgery is forthcoming. 3046 _revisitStack.reset(); // throwing away all contents 3047 3048 // Marking completed -- now verify that each bit marked in 3049 // verification_mark_bm() is also marked in markBitMap(); flag all 3050 // errors by printing corresponding objects. 3051 VerifyMarkedClosure vcl(markBitMap()); 3052 verification_mark_bm()->iterate(&vcl); 3053 assert(!vcl.failed(), "Else verification above should not have succeeded"); 3054 } 3055 3056 void ConcurrentMarkSweepGeneration::save_marks() { 3057 // delegate to CMS space 3058 cmsSpace()->save_marks(); 3059 for (uint i = 0; i < ParallelGCThreads; i++) { 3060 _par_gc_thread_states[i]->promo.startTrackingPromotions(); 3061 } 3062 } 3063 3064 bool ConcurrentMarkSweepGeneration::no_allocs_since_save_marks() { 3065 return cmsSpace()->no_allocs_since_save_marks(); 3066 } 3067 3068 #define CMS_SINCE_SAVE_MARKS_DEFN(OopClosureType, nv_suffix) \ 3069 \ 3070 void ConcurrentMarkSweepGeneration:: \ 3071 oop_since_save_marks_iterate##nv_suffix(OopClosureType* cl) { \ 3072 cl->set_generation(this); \ 3073 cmsSpace()->oop_since_save_marks_iterate##nv_suffix(cl); \ 3074 cl->reset_generation(); \ 3075 save_marks(); \ 3076 } 3077 3078 ALL_SINCE_SAVE_MARKS_CLOSURES(CMS_SINCE_SAVE_MARKS_DEFN) 3079 3080 void 3081 ConcurrentMarkSweepGeneration::object_iterate_since_last_GC(ObjectClosure* blk) 3082 { 3083 // Not currently implemented; need to do the following. -- ysr. 3084 // dld -- I think that is used for some sort of allocation profiler. So it 3085 // really means the objects allocated by the mutator since the last 3086 // GC. We could potentially implement this cheaply by recording only 3087 // the direct allocations in a side data structure. 3088 // 3089 // I think we probably ought not to be required to support these 3090 // iterations at any arbitrary point; I think there ought to be some 3091 // call to enable/disable allocation profiling in a generation/space, 3092 // and the iterator ought to return the objects allocated in the 3093 // gen/space since the enable call, or the last iterator call (which 3094 // will probably be at a GC.) That way, for gens like CM&S that would 3095 // require some extra data structure to support this, we only pay the 3096 // cost when it's in use... 3097 cmsSpace()->object_iterate_since_last_GC(blk); 3098 } 3099 3100 void 3101 ConcurrentMarkSweepGeneration::younger_refs_iterate(OopsInGenClosure* cl) { 3102 cl->set_generation(this); 3103 younger_refs_in_space_iterate(_cmsSpace, cl); 3104 cl->reset_generation(); 3105 } 3106 3107 void 3108 ConcurrentMarkSweepGeneration::oop_iterate(MemRegion mr, OopClosure* cl) { 3109 if (freelistLock()->owned_by_self()) { 3110 Generation::oop_iterate(mr, cl); 3111 } else { 3112 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 3113 Generation::oop_iterate(mr, cl); 3114 } 3115 } 3116 3117 void 3118 ConcurrentMarkSweepGeneration::oop_iterate(OopClosure* cl) { 3119 if (freelistLock()->owned_by_self()) { 3120 Generation::oop_iterate(cl); 3121 } else { 3122 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 3123 Generation::oop_iterate(cl); 3124 } 3125 } 3126 3127 void 3128 ConcurrentMarkSweepGeneration::object_iterate(ObjectClosure* cl) { 3129 if (freelistLock()->owned_by_self()) { 3130 Generation::object_iterate(cl); 3131 } else { 3132 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 3133 Generation::object_iterate(cl); 3134 } 3135 } 3136 3137 void 3138 ConcurrentMarkSweepGeneration::safe_object_iterate(ObjectClosure* cl) { 3139 if (freelistLock()->owned_by_self()) { 3140 Generation::safe_object_iterate(cl); 3141 } else { 3142 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 3143 Generation::safe_object_iterate(cl); 3144 } 3145 } 3146 3147 void 3148 ConcurrentMarkSweepGeneration::pre_adjust_pointers() { 3149 } 3150 3151 void 3152 ConcurrentMarkSweepGeneration::post_compact() { 3153 } 3154 3155 void 3156 ConcurrentMarkSweepGeneration::prepare_for_verify() { 3157 // Fix the linear allocation blocks to look like free blocks. 3158 3159 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those 3160 // are not called when the heap is verified during universe initialization and 3161 // at vm shutdown. 3162 if (freelistLock()->owned_by_self()) { 3163 cmsSpace()->prepare_for_verify(); 3164 } else { 3165 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag); 3166 cmsSpace()->prepare_for_verify(); 3167 } 3168 } 3169 3170 void 3171 ConcurrentMarkSweepGeneration::verify() { 3172 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those 3173 // are not called when the heap is verified during universe initialization and 3174 // at vm shutdown. 3175 if (freelistLock()->owned_by_self()) { 3176 cmsSpace()->verify(); 3177 } else { 3178 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag); 3179 cmsSpace()->verify(); 3180 } 3181 } 3182 3183 void CMSCollector::verify() { 3184 _cmsGen->verify(); 3185 _permGen->verify(); 3186 } 3187 3188 #ifndef PRODUCT 3189 bool CMSCollector::overflow_list_is_empty() const { 3190 assert(_num_par_pushes >= 0, "Inconsistency"); 3191 if (_overflow_list == NULL) { 3192 assert(_num_par_pushes == 0, "Inconsistency"); 3193 } 3194 return _overflow_list == NULL; 3195 } 3196 3197 // The methods verify_work_stacks_empty() and verify_overflow_empty() 3198 // merely consolidate assertion checks that appear to occur together frequently. 3199 void CMSCollector::verify_work_stacks_empty() const { 3200 assert(_markStack.isEmpty(), "Marking stack should be empty"); 3201 assert(overflow_list_is_empty(), "Overflow list should be empty"); 3202 } 3203 3204 void CMSCollector::verify_overflow_empty() const { 3205 assert(overflow_list_is_empty(), "Overflow list should be empty"); 3206 assert(no_preserved_marks(), "No preserved marks"); 3207 } 3208 #endif // PRODUCT 3209 3210 // Decide if we want to enable class unloading as part of the 3211 // ensuing concurrent GC cycle. We will collect the perm gen and 3212 // unload classes if it's the case that: 3213 // (1) an explicit gc request has been made and the flag 3214 // ExplicitGCInvokesConcurrentAndUnloadsClasses is set, OR 3215 // (2) (a) class unloading is enabled at the command line, and 3216 // (b) (i) perm gen threshold has been crossed, or 3217 // (ii) old gen is getting really full, or 3218 // (iii) the previous N CMS collections did not collect the 3219 // perm gen 3220 // NOTE: Provided there is no change in the state of the heap between 3221 // calls to this method, it should have idempotent results. Moreover, 3222 // its results should be monotonically increasing (i.e. going from 0 to 1, 3223 // but not 1 to 0) between successive calls between which the heap was 3224 // not collected. For the implementation below, it must thus rely on 3225 // the property that concurrent_cycles_since_last_unload() 3226 // will not decrease unless a collection cycle happened and that 3227 // _permGen->should_concurrent_collect() and _cmsGen->is_too_full() are 3228 // themselves also monotonic in that sense. See check_monotonicity() 3229 // below. 3230 bool CMSCollector::update_should_unload_classes() { 3231 _should_unload_classes = false; 3232 // Condition 1 above 3233 if (_full_gc_requested && ExplicitGCInvokesConcurrentAndUnloadsClasses) { 3234 _should_unload_classes = true; 3235 } else if (CMSClassUnloadingEnabled) { // Condition 2.a above 3236 // Disjuncts 2.b.(i,ii,iii) above 3237 _should_unload_classes = (concurrent_cycles_since_last_unload() >= 3238 CMSClassUnloadingMaxInterval) 3239 || _permGen->should_concurrent_collect() 3240 || _cmsGen->is_too_full(); 3241 } 3242 return _should_unload_classes; 3243 } 3244 3245 bool ConcurrentMarkSweepGeneration::is_too_full() const { 3246 bool res = should_concurrent_collect(); 3247 res = res && (occupancy() > (double)CMSIsTooFullPercentage/100.0); 3248 return res; 3249 } 3250 3251 void CMSCollector::setup_cms_unloading_and_verification_state() { 3252 const bool should_verify = VerifyBeforeGC || VerifyAfterGC || VerifyDuringGC 3253 || VerifyBeforeExit; 3254 const int rso = SharedHeap::SO_Strings | SharedHeap::SO_CodeCache; 3255 3256 if (should_unload_classes()) { // Should unload classes this cycle 3257 remove_root_scanning_option(rso); // Shrink the root set appropriately 3258 set_verifying(should_verify); // Set verification state for this cycle 3259 return; // Nothing else needs to be done at this time 3260 } 3261 3262 // Not unloading classes this cycle 3263 assert(!should_unload_classes(), "Inconsitency!"); 3264 if ((!verifying() || unloaded_classes_last_cycle()) && should_verify) { 3265 // We were not verifying, or we _were_ unloading classes in the last cycle, 3266 // AND some verification options are enabled this cycle; in this case, 3267 // we must make sure that the deadness map is allocated if not already so, 3268 // and cleared (if already allocated previously -- 3269 // CMSBitMap::sizeInBits() is used to determine if it's allocated). 3270 if (perm_gen_verify_bit_map()->sizeInBits() == 0) { 3271 if (!perm_gen_verify_bit_map()->allocate(_permGen->reserved())) { 3272 warning("Failed to allocate permanent generation verification CMS Bit Map;\n" 3273 "permanent generation verification disabled"); 3274 return; // Note that we leave verification disabled, so we'll retry this 3275 // allocation next cycle. We _could_ remember this failure 3276 // and skip further attempts and permanently disable verification 3277 // attempts if that is considered more desirable. 3278 } 3279 assert(perm_gen_verify_bit_map()->covers(_permGen->reserved()), 3280 "_perm_gen_ver_bit_map inconsistency?"); 3281 } else { 3282 perm_gen_verify_bit_map()->clear_all(); 3283 } 3284 // Include symbols, strings and code cache elements to prevent their resurrection. 3285 add_root_scanning_option(rso); 3286 set_verifying(true); 3287 } else if (verifying() && !should_verify) { 3288 // We were verifying, but some verification flags got disabled. 3289 set_verifying(false); 3290 // Exclude symbols, strings and code cache elements from root scanning to 3291 // reduce IM and RM pauses. 3292 remove_root_scanning_option(rso); 3293 } 3294 } 3295 3296 3297 #ifndef PRODUCT 3298 HeapWord* CMSCollector::block_start(const void* p) const { 3299 const HeapWord* addr = (HeapWord*)p; 3300 if (_span.contains(p)) { 3301 if (_cmsGen->cmsSpace()->is_in_reserved(addr)) { 3302 return _cmsGen->cmsSpace()->block_start(p); 3303 } else { 3304 assert(_permGen->cmsSpace()->is_in_reserved(addr), 3305 "Inconsistent _span?"); 3306 return _permGen->cmsSpace()->block_start(p); 3307 } 3308 } 3309 return NULL; 3310 } 3311 #endif 3312 3313 HeapWord* 3314 ConcurrentMarkSweepGeneration::expand_and_allocate(size_t word_size, 3315 bool tlab, 3316 bool parallel) { 3317 CMSSynchronousYieldRequest yr; 3318 assert(!tlab, "Can't deal with TLAB allocation"); 3319 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 3320 expand(word_size*HeapWordSize, MinHeapDeltaBytes, 3321 CMSExpansionCause::_satisfy_allocation); 3322 if (GCExpandToAllocateDelayMillis > 0) { 3323 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 3324 } 3325 return have_lock_and_allocate(word_size, tlab); 3326 } 3327 3328 // YSR: All of this generation expansion/shrinking stuff is an exact copy of 3329 // OneContigSpaceCardGeneration, which makes me wonder if we should move this 3330 // to CardGeneration and share it... 3331 bool ConcurrentMarkSweepGeneration::expand(size_t bytes, size_t expand_bytes) { 3332 return CardGeneration::expand(bytes, expand_bytes); 3333 } 3334 3335 void ConcurrentMarkSweepGeneration::expand(size_t bytes, size_t expand_bytes, 3336 CMSExpansionCause::Cause cause) 3337 { 3338 3339 bool success = expand(bytes, expand_bytes); 3340 3341 // remember why we expanded; this information is used 3342 // by shouldConcurrentCollect() when making decisions on whether to start 3343 // a new CMS cycle. 3344 if (success) { 3345 set_expansion_cause(cause); 3346 if (PrintGCDetails && Verbose) { 3347 gclog_or_tty->print_cr("Expanded CMS gen for %s", 3348 CMSExpansionCause::to_string(cause)); 3349 } 3350 } 3351 } 3352 3353 HeapWord* ConcurrentMarkSweepGeneration::expand_and_par_lab_allocate(CMSParGCThreadState* ps, size_t word_sz) { 3354 HeapWord* res = NULL; 3355 MutexLocker x(ParGCRareEvent_lock); 3356 while (true) { 3357 // Expansion by some other thread might make alloc OK now: 3358 res = ps->lab.alloc(word_sz); 3359 if (res != NULL) return res; 3360 // If there's not enough expansion space available, give up. 3361 if (_virtual_space.uncommitted_size() < (word_sz * HeapWordSize)) { 3362 return NULL; 3363 } 3364 // Otherwise, we try expansion. 3365 expand(word_sz*HeapWordSize, MinHeapDeltaBytes, 3366 CMSExpansionCause::_allocate_par_lab); 3367 // Now go around the loop and try alloc again; 3368 // A competing par_promote might beat us to the expansion space, 3369 // so we may go around the loop again if promotion fails agaion. 3370 if (GCExpandToAllocateDelayMillis > 0) { 3371 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 3372 } 3373 } 3374 } 3375 3376 3377 bool ConcurrentMarkSweepGeneration::expand_and_ensure_spooling_space( 3378 PromotionInfo* promo) { 3379 MutexLocker x(ParGCRareEvent_lock); 3380 size_t refill_size_bytes = promo->refillSize() * HeapWordSize; 3381 while (true) { 3382 // Expansion by some other thread might make alloc OK now: 3383 if (promo->ensure_spooling_space()) { 3384 assert(promo->has_spooling_space(), 3385 "Post-condition of successful ensure_spooling_space()"); 3386 return true; 3387 } 3388 // If there's not enough expansion space available, give up. 3389 if (_virtual_space.uncommitted_size() < refill_size_bytes) { 3390 return false; 3391 } 3392 // Otherwise, we try expansion. 3393 expand(refill_size_bytes, MinHeapDeltaBytes, 3394 CMSExpansionCause::_allocate_par_spooling_space); 3395 // Now go around the loop and try alloc again; 3396 // A competing allocation might beat us to the expansion space, 3397 // so we may go around the loop again if allocation fails again. 3398 if (GCExpandToAllocateDelayMillis > 0) { 3399 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 3400 } 3401 } 3402 } 3403 3404 3405 3406 void ConcurrentMarkSweepGeneration::shrink(size_t bytes) { 3407 assert_locked_or_safepoint(Heap_lock); 3408 size_t size = ReservedSpace::page_align_size_down(bytes); 3409 if (size > 0) { 3410 shrink_by(size); 3411 } 3412 } 3413 3414 bool ConcurrentMarkSweepGeneration::grow_by(size_t bytes) { 3415 assert_locked_or_safepoint(Heap_lock); 3416 bool result = _virtual_space.expand_by(bytes); 3417 if (result) { 3418 HeapWord* old_end = _cmsSpace->end(); 3419 size_t new_word_size = 3420 heap_word_size(_virtual_space.committed_size()); 3421 MemRegion mr(_cmsSpace->bottom(), new_word_size); 3422 _bts->resize(new_word_size); // resize the block offset shared array 3423 Universe::heap()->barrier_set()->resize_covered_region(mr); 3424 // Hmmmm... why doesn't CFLS::set_end verify locking? 3425 // This is quite ugly; FIX ME XXX 3426 _cmsSpace->assert_locked(freelistLock()); 3427 _cmsSpace->set_end((HeapWord*)_virtual_space.high()); 3428 3429 // update the space and generation capacity counters 3430 if (UsePerfData) { 3431 _space_counters->update_capacity(); 3432 _gen_counters->update_all(); 3433 } 3434 3435 if (Verbose && PrintGC) { 3436 size_t new_mem_size = _virtual_space.committed_size(); 3437 size_t old_mem_size = new_mem_size - bytes; 3438 gclog_or_tty->print_cr("Expanding %s from %ldK by %ldK to %ldK", 3439 name(), old_mem_size/K, bytes/K, new_mem_size/K); 3440 } 3441 } 3442 return result; 3443 } 3444 3445 bool ConcurrentMarkSweepGeneration::grow_to_reserved() { 3446 assert_locked_or_safepoint(Heap_lock); 3447 bool success = true; 3448 const size_t remaining_bytes = _virtual_space.uncommitted_size(); 3449 if (remaining_bytes > 0) { 3450 success = grow_by(remaining_bytes); 3451 DEBUG_ONLY(if (!success) warning("grow to reserved failed");) 3452 } 3453 return success; 3454 } 3455 3456 void ConcurrentMarkSweepGeneration::shrink_by(size_t bytes) { 3457 assert_locked_or_safepoint(Heap_lock); 3458 assert_lock_strong(freelistLock()); 3459 // XXX Fix when compaction is implemented. 3460 warning("Shrinking of CMS not yet implemented"); 3461 return; 3462 } 3463 3464 3465 // Simple ctor/dtor wrapper for accounting & timer chores around concurrent 3466 // phases. 3467 class CMSPhaseAccounting: public StackObj { 3468 public: 3469 CMSPhaseAccounting(CMSCollector *collector, 3470 const char *phase, 3471 bool print_cr = true); 3472 ~CMSPhaseAccounting(); 3473 3474 private: 3475 CMSCollector *_collector; 3476 const char *_phase; 3477 elapsedTimer _wallclock; 3478 bool _print_cr; 3479 3480 public: 3481 // Not MT-safe; so do not pass around these StackObj's 3482 // where they may be accessed by other threads. 3483 jlong wallclock_millis() { 3484 assert(_wallclock.is_active(), "Wall clock should not stop"); 3485 _wallclock.stop(); // to record time 3486 jlong ret = _wallclock.milliseconds(); 3487 _wallclock.start(); // restart 3488 return ret; 3489 } 3490 }; 3491 3492 CMSPhaseAccounting::CMSPhaseAccounting(CMSCollector *collector, 3493 const char *phase, 3494 bool print_cr) : 3495 _collector(collector), _phase(phase), _print_cr(print_cr) { 3496 3497 if (PrintCMSStatistics != 0) { 3498 _collector->resetYields(); 3499 } 3500 if (PrintGCDetails && PrintGCTimeStamps) { 3501 gclog_or_tty->date_stamp(PrintGCDateStamps); 3502 gclog_or_tty->stamp(); 3503 gclog_or_tty->print_cr(": [%s-concurrent-%s-start]", 3504 _collector->cmsGen()->short_name(), _phase); 3505 } 3506 _collector->resetTimer(); 3507 _wallclock.start(); 3508 _collector->startTimer(); 3509 } 3510 3511 CMSPhaseAccounting::~CMSPhaseAccounting() { 3512 assert(_wallclock.is_active(), "Wall clock should not have stopped"); 3513 _collector->stopTimer(); 3514 _wallclock.stop(); 3515 if (PrintGCDetails) { 3516 gclog_or_tty->date_stamp(PrintGCDateStamps); 3517 gclog_or_tty->stamp(PrintGCTimeStamps); 3518 gclog_or_tty->print("[%s-concurrent-%s: %3.3f/%3.3f secs]", 3519 _collector->cmsGen()->short_name(), 3520 _phase, _collector->timerValue(), _wallclock.seconds()); 3521 if (_print_cr) { 3522 gclog_or_tty->print_cr(""); 3523 } 3524 if (PrintCMSStatistics != 0) { 3525 gclog_or_tty->print_cr(" (CMS-concurrent-%s yielded %d times)", _phase, 3526 _collector->yields()); 3527 } 3528 } 3529 } 3530 3531 // CMS work 3532 3533 // Checkpoint the roots into this generation from outside 3534 // this generation. [Note this initial checkpoint need only 3535 // be approximate -- we'll do a catch up phase subsequently.] 3536 void CMSCollector::checkpointRootsInitial(bool asynch) { 3537 assert(_collectorState == InitialMarking, "Wrong collector state"); 3538 check_correct_thread_executing(); 3539 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 3540 3541 save_heap_summary(); 3542 report_heap_summary(GCWhen::BeforeGC); 3543 3544 ReferenceProcessor* rp = ref_processor(); 3545 SpecializationStats::clear(); 3546 assert(_restart_addr == NULL, "Control point invariant"); 3547 if (asynch) { 3548 // acquire locks for subsequent manipulations 3549 MutexLockerEx x(bitMapLock(), 3550 Mutex::_no_safepoint_check_flag); 3551 checkpointRootsInitialWork(asynch); 3552 // enable ("weak") refs discovery 3553 rp->enable_discovery(true /*verify_disabled*/, true /*check_no_refs*/); 3554 _collectorState = Marking; 3555 } else { 3556 // (Weak) Refs discovery: this is controlled from genCollectedHeap::do_collection 3557 // which recognizes if we are a CMS generation, and doesn't try to turn on 3558 // discovery; verify that they aren't meddling. 3559 assert(!rp->discovery_is_atomic(), 3560 "incorrect setting of discovery predicate"); 3561 assert(!rp->discovery_enabled(), "genCollectedHeap shouldn't control " 3562 "ref discovery for this generation kind"); 3563 // already have locks 3564 checkpointRootsInitialWork(asynch); 3565 // now enable ("weak") refs discovery 3566 rp->enable_discovery(true /*verify_disabled*/, false /*verify_no_refs*/); 3567 _collectorState = Marking; 3568 } 3569 SpecializationStats::print(); 3570 } 3571 3572 void CMSCollector::checkpointRootsInitialWork(bool asynch) { 3573 assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped"); 3574 assert(_collectorState == InitialMarking, "just checking"); 3575 3576 // If there has not been a GC[n-1] since last GC[n] cycle completed, 3577 // precede our marking with a collection of all 3578 // younger generations to keep floating garbage to a minimum. 3579 // XXX: we won't do this for now -- it's an optimization to be done later. 3580 3581 // already have locks 3582 assert_lock_strong(bitMapLock()); 3583 assert(_markBitMap.isAllClear(), "was reset at end of previous cycle"); 3584 3585 // Setup the verification and class unloading state for this 3586 // CMS collection cycle. 3587 setup_cms_unloading_and_verification_state(); 3588 3589 NOT_PRODUCT(GCTraceTime t("\ncheckpointRootsInitialWork", 3590 PrintGCDetails && Verbose, true, _gc_timer_cm);) 3591 if (UseAdaptiveSizePolicy) { 3592 size_policy()->checkpoint_roots_initial_begin(); 3593 } 3594 3595 // Reset all the PLAB chunk arrays if necessary. 3596 if (_survivor_plab_array != NULL && !CMSPLABRecordAlways) { 3597 reset_survivor_plab_arrays(); 3598 } 3599 3600 ResourceMark rm; 3601 HandleMark hm; 3602 3603 FalseClosure falseClosure; 3604 // In the case of a synchronous collection, we will elide the 3605 // remark step, so it's important to catch all the nmethod oops 3606 // in this step. 3607 // The final 'true' flag to gen_process_strong_roots will ensure this. 3608 // If 'async' is true, we can relax the nmethod tracing. 3609 MarkRefsIntoClosure notOlder(_span, &_markBitMap); 3610 GenCollectedHeap* gch = GenCollectedHeap::heap(); 3611 3612 verify_work_stacks_empty(); 3613 verify_overflow_empty(); 3614 3615 gch->ensure_parsability(false); // fill TLABs, but no need to retire them 3616 // Update the saved marks which may affect the root scans. 3617 gch->save_marks(); 3618 3619 // weak reference processing has not started yet. 3620 ref_processor()->set_enqueuing_is_done(false); 3621 3622 { 3623 // This is not needed. DEBUG_ONLY(RememberKlassesChecker imx(true);) 3624 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;) 3625 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 3626 gch->gen_process_strong_roots(_cmsGen->level(), 3627 true, // younger gens are roots 3628 true, // activate StrongRootsScope 3629 true, // collecting perm gen 3630 SharedHeap::ScanningOption(roots_scanning_options()), 3631 ¬Older, 3632 true, // walk all of code cache if (so & SO_CodeCache) 3633 NULL); 3634 } 3635 3636 // Clear mod-union table; it will be dirtied in the prologue of 3637 // CMS generation per each younger generation collection. 3638 3639 assert(_modUnionTable.isAllClear(), 3640 "Was cleared in most recent final checkpoint phase" 3641 " or no bits are set in the gc_prologue before the start of the next " 3642 "subsequent marking phase."); 3643 3644 // Save the end of the used_region of the constituent generations 3645 // to be used to limit the extent of sweep in each generation. 3646 save_sweep_limits(); 3647 if (UseAdaptiveSizePolicy) { 3648 size_policy()->checkpoint_roots_initial_end(gch->gc_cause()); 3649 } 3650 verify_overflow_empty(); 3651 } 3652 3653 bool CMSCollector::markFromRoots(bool asynch) { 3654 // we might be tempted to assert that: 3655 // assert(asynch == !SafepointSynchronize::is_at_safepoint(), 3656 // "inconsistent argument?"); 3657 // However that wouldn't be right, because it's possible that 3658 // a safepoint is indeed in progress as a younger generation 3659 // stop-the-world GC happens even as we mark in this generation. 3660 assert(_collectorState == Marking, "inconsistent state?"); 3661 check_correct_thread_executing(); 3662 verify_overflow_empty(); 3663 3664 bool res; 3665 if (asynch) { 3666 3667 // Start the timers for adaptive size policy for the concurrent phases 3668 // Do it here so that the foreground MS can use the concurrent 3669 // timer since a foreground MS might has the sweep done concurrently 3670 // or STW. 3671 if (UseAdaptiveSizePolicy) { 3672 size_policy()->concurrent_marking_begin(); 3673 } 3674 3675 // Weak ref discovery note: We may be discovering weak 3676 // refs in this generation concurrent (but interleaved) with 3677 // weak ref discovery by a younger generation collector. 3678 3679 CMSTokenSyncWithLocks ts(true, bitMapLock()); 3680 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 3681 CMSPhaseAccounting pa(this, "mark", !PrintGCDetails); 3682 res = markFromRootsWork(asynch); 3683 if (res) { 3684 _collectorState = Precleaning; 3685 } else { // We failed and a foreground collection wants to take over 3686 assert(_foregroundGCIsActive, "internal state inconsistency"); 3687 assert(_restart_addr == NULL, "foreground will restart from scratch"); 3688 if (PrintGCDetails) { 3689 gclog_or_tty->print_cr("bailing out to foreground collection"); 3690 } 3691 } 3692 if (UseAdaptiveSizePolicy) { 3693 size_policy()->concurrent_marking_end(); 3694 } 3695 } else { 3696 assert(SafepointSynchronize::is_at_safepoint(), 3697 "inconsistent with asynch == false"); 3698 if (UseAdaptiveSizePolicy) { 3699 size_policy()->ms_collection_marking_begin(); 3700 } 3701 // already have locks 3702 res = markFromRootsWork(asynch); 3703 _collectorState = FinalMarking; 3704 if (UseAdaptiveSizePolicy) { 3705 GenCollectedHeap* gch = GenCollectedHeap::heap(); 3706 size_policy()->ms_collection_marking_end(gch->gc_cause()); 3707 } 3708 } 3709 verify_overflow_empty(); 3710 return res; 3711 } 3712 3713 bool CMSCollector::markFromRootsWork(bool asynch) { 3714 // iterate over marked bits in bit map, doing a full scan and mark 3715 // from these roots using the following algorithm: 3716 // . if oop is to the right of the current scan pointer, 3717 // mark corresponding bit (we'll process it later) 3718 // . else (oop is to left of current scan pointer) 3719 // push oop on marking stack 3720 // . drain the marking stack 3721 3722 // Note that when we do a marking step we need to hold the 3723 // bit map lock -- recall that direct allocation (by mutators) 3724 // and promotion (by younger generation collectors) is also 3725 // marking the bit map. [the so-called allocate live policy.] 3726 // Because the implementation of bit map marking is not 3727 // robust wrt simultaneous marking of bits in the same word, 3728 // we need to make sure that there is no such interference 3729 // between concurrent such updates. 3730 3731 // already have locks 3732 assert_lock_strong(bitMapLock()); 3733 3734 // Clear the revisit stack, just in case there are any 3735 // obsolete contents from a short-circuited previous CMS cycle. 3736 _revisitStack.reset(); 3737 verify_work_stacks_empty(); 3738 verify_overflow_empty(); 3739 assert(_revisitStack.isEmpty(), "tabula rasa"); 3740 DEBUG_ONLY(RememberKlassesChecker cmx(should_unload_classes());) 3741 bool result = false; 3742 if (CMSConcurrentMTEnabled && ConcGCThreads > 0) { 3743 result = do_marking_mt(asynch); 3744 } else { 3745 result = do_marking_st(asynch); 3746 } 3747 return result; 3748 } 3749 3750 // Forward decl 3751 class CMSConcMarkingTask; 3752 3753 class CMSConcMarkingTerminator: public ParallelTaskTerminator { 3754 CMSCollector* _collector; 3755 CMSConcMarkingTask* _task; 3756 public: 3757 virtual void yield(); 3758 3759 // "n_threads" is the number of threads to be terminated. 3760 // "queue_set" is a set of work queues of other threads. 3761 // "collector" is the CMS collector associated with this task terminator. 3762 // "yield" indicates whether we need the gang as a whole to yield. 3763 CMSConcMarkingTerminator(int n_threads, TaskQueueSetSuper* queue_set, CMSCollector* collector) : 3764 ParallelTaskTerminator(n_threads, queue_set), 3765 _collector(collector) { } 3766 3767 void set_task(CMSConcMarkingTask* task) { 3768 _task = task; 3769 } 3770 }; 3771 3772 class CMSConcMarkingTerminatorTerminator: public TerminatorTerminator { 3773 CMSConcMarkingTask* _task; 3774 public: 3775 bool should_exit_termination(); 3776 void set_task(CMSConcMarkingTask* task) { 3777 _task = task; 3778 } 3779 }; 3780 3781 // MT Concurrent Marking Task 3782 class CMSConcMarkingTask: public YieldingFlexibleGangTask { 3783 CMSCollector* _collector; 3784 int _n_workers; // requested/desired # workers 3785 bool _asynch; 3786 bool _result; 3787 CompactibleFreeListSpace* _cms_space; 3788 CompactibleFreeListSpace* _perm_space; 3789 char _pad_front[64]; // padding to ... 3790 HeapWord* _global_finger; // ... avoid sharing cache line 3791 char _pad_back[64]; 3792 HeapWord* _restart_addr; 3793 3794 // Exposed here for yielding support 3795 Mutex* const _bit_map_lock; 3796 3797 // The per thread work queues, available here for stealing 3798 OopTaskQueueSet* _task_queues; 3799 3800 // Termination (and yielding) support 3801 CMSConcMarkingTerminator _term; 3802 CMSConcMarkingTerminatorTerminator _term_term; 3803 3804 public: 3805 CMSConcMarkingTask(CMSCollector* collector, 3806 CompactibleFreeListSpace* cms_space, 3807 CompactibleFreeListSpace* perm_space, 3808 bool asynch, 3809 YieldingFlexibleWorkGang* workers, 3810 OopTaskQueueSet* task_queues): 3811 YieldingFlexibleGangTask("Concurrent marking done multi-threaded"), 3812 _collector(collector), 3813 _cms_space(cms_space), 3814 _perm_space(perm_space), 3815 _asynch(asynch), _n_workers(0), _result(true), 3816 _task_queues(task_queues), 3817 _term(_n_workers, task_queues, _collector), 3818 _bit_map_lock(collector->bitMapLock()) 3819 { 3820 _requested_size = _n_workers; 3821 _term.set_task(this); 3822 _term_term.set_task(this); 3823 assert(_cms_space->bottom() < _perm_space->bottom(), 3824 "Finger incorrectly initialized below"); 3825 _restart_addr = _global_finger = _cms_space->bottom(); 3826 } 3827 3828 3829 OopTaskQueueSet* task_queues() { return _task_queues; } 3830 3831 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 3832 3833 HeapWord** global_finger_addr() { return &_global_finger; } 3834 3835 CMSConcMarkingTerminator* terminator() { return &_term; } 3836 3837 virtual void set_for_termination(int active_workers) { 3838 terminator()->reset_for_reuse(active_workers); 3839 } 3840 3841 void work(uint worker_id); 3842 bool should_yield() { 3843 return ConcurrentMarkSweepThread::should_yield() 3844 && !_collector->foregroundGCIsActive() 3845 && _asynch; 3846 } 3847 3848 virtual void coordinator_yield(); // stuff done by coordinator 3849 bool result() { return _result; } 3850 3851 void reset(HeapWord* ra) { 3852 assert(_global_finger >= _cms_space->end(), "Postcondition of ::work(i)"); 3853 assert(_global_finger >= _perm_space->end(), "Postcondition of ::work(i)"); 3854 assert(ra < _perm_space->end(), "ra too large"); 3855 _restart_addr = _global_finger = ra; 3856 _term.reset_for_reuse(); 3857 } 3858 3859 static bool get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 3860 OopTaskQueue* work_q); 3861 3862 private: 3863 void do_scan_and_mark(int i, CompactibleFreeListSpace* sp); 3864 void do_work_steal(int i); 3865 void bump_global_finger(HeapWord* f); 3866 }; 3867 3868 bool CMSConcMarkingTerminatorTerminator::should_exit_termination() { 3869 assert(_task != NULL, "Error"); 3870 return _task->yielding(); 3871 // Note that we do not need the disjunct || _task->should_yield() above 3872 // because we want terminating threads to yield only if the task 3873 // is already in the midst of yielding, which happens only after at least one 3874 // thread has yielded. 3875 } 3876 3877 void CMSConcMarkingTerminator::yield() { 3878 if (_task->should_yield()) { 3879 _task->yield(); 3880 } else { 3881 ParallelTaskTerminator::yield(); 3882 } 3883 } 3884 3885 //////////////////////////////////////////////////////////////// 3886 // Concurrent Marking Algorithm Sketch 3887 //////////////////////////////////////////////////////////////// 3888 // Until all tasks exhausted (both spaces): 3889 // -- claim next available chunk 3890 // -- bump global finger via CAS 3891 // -- find first object that starts in this chunk 3892 // and start scanning bitmap from that position 3893 // -- scan marked objects for oops 3894 // -- CAS-mark target, and if successful: 3895 // . if target oop is above global finger (volatile read) 3896 // nothing to do 3897 // . if target oop is in chunk and above local finger 3898 // then nothing to do 3899 // . else push on work-queue 3900 // -- Deal with possible overflow issues: 3901 // . local work-queue overflow causes stuff to be pushed on 3902 // global (common) overflow queue 3903 // . always first empty local work queue 3904 // . then get a batch of oops from global work queue if any 3905 // . then do work stealing 3906 // -- When all tasks claimed (both spaces) 3907 // and local work queue empty, 3908 // then in a loop do: 3909 // . check global overflow stack; steal a batch of oops and trace 3910 // . try to steal from other threads oif GOS is empty 3911 // . if neither is available, offer termination 3912 // -- Terminate and return result 3913 // 3914 void CMSConcMarkingTask::work(uint worker_id) { 3915 elapsedTimer _timer; 3916 ResourceMark rm; 3917 HandleMark hm; 3918 3919 DEBUG_ONLY(_collector->verify_overflow_empty();) 3920 3921 // Before we begin work, our work queue should be empty 3922 assert(work_queue(worker_id)->size() == 0, "Expected to be empty"); 3923 // Scan the bitmap covering _cms_space, tracing through grey objects. 3924 _timer.start(); 3925 do_scan_and_mark(worker_id, _cms_space); 3926 _timer.stop(); 3927 if (PrintCMSStatistics != 0) { 3928 gclog_or_tty->print_cr("Finished cms space scanning in %dth thread: %3.3f sec", 3929 worker_id, _timer.seconds()); 3930 // XXX: need xxx/xxx type of notation, two timers 3931 } 3932 3933 // ... do the same for the _perm_space 3934 _timer.reset(); 3935 _timer.start(); 3936 do_scan_and_mark(worker_id, _perm_space); 3937 _timer.stop(); 3938 if (PrintCMSStatistics != 0) { 3939 gclog_or_tty->print_cr("Finished perm space scanning in %dth thread: %3.3f sec", 3940 worker_id, _timer.seconds()); 3941 // XXX: need xxx/xxx type of notation, two timers 3942 } 3943 3944 // ... do work stealing 3945 _timer.reset(); 3946 _timer.start(); 3947 do_work_steal(worker_id); 3948 _timer.stop(); 3949 if (PrintCMSStatistics != 0) { 3950 gclog_or_tty->print_cr("Finished work stealing in %dth thread: %3.3f sec", 3951 worker_id, _timer.seconds()); 3952 // XXX: need xxx/xxx type of notation, two timers 3953 } 3954 assert(_collector->_markStack.isEmpty(), "Should have been emptied"); 3955 assert(work_queue(worker_id)->size() == 0, "Should have been emptied"); 3956 // Note that under the current task protocol, the 3957 // following assertion is true even of the spaces 3958 // expanded since the completion of the concurrent 3959 // marking. XXX This will likely change under a strict 3960 // ABORT semantics. 3961 assert(_global_finger > _cms_space->end() && 3962 _global_finger >= _perm_space->end(), 3963 "All tasks have been completed"); 3964 DEBUG_ONLY(_collector->verify_overflow_empty();) 3965 } 3966 3967 void CMSConcMarkingTask::bump_global_finger(HeapWord* f) { 3968 HeapWord* read = _global_finger; 3969 HeapWord* cur = read; 3970 while (f > read) { 3971 cur = read; 3972 read = (HeapWord*) Atomic::cmpxchg_ptr(f, &_global_finger, cur); 3973 if (cur == read) { 3974 // our cas succeeded 3975 assert(_global_finger >= f, "protocol consistency"); 3976 break; 3977 } 3978 } 3979 } 3980 3981 // This is really inefficient, and should be redone by 3982 // using (not yet available) block-read and -write interfaces to the 3983 // stack and the work_queue. XXX FIX ME !!! 3984 bool CMSConcMarkingTask::get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 3985 OopTaskQueue* work_q) { 3986 // Fast lock-free check 3987 if (ovflw_stk->length() == 0) { 3988 return false; 3989 } 3990 assert(work_q->size() == 0, "Shouldn't steal"); 3991 MutexLockerEx ml(ovflw_stk->par_lock(), 3992 Mutex::_no_safepoint_check_flag); 3993 // Grab up to 1/4 the size of the work queue 3994 size_t num = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 3995 (size_t)ParGCDesiredObjsFromOverflowList); 3996 num = MIN2(num, ovflw_stk->length()); 3997 for (int i = (int) num; i > 0; i--) { 3998 oop cur = ovflw_stk->pop(); 3999 assert(cur != NULL, "Counted wrong?"); 4000 work_q->push(cur); 4001 } 4002 return num > 0; 4003 } 4004 4005 void CMSConcMarkingTask::do_scan_and_mark(int i, CompactibleFreeListSpace* sp) { 4006 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 4007 int n_tasks = pst->n_tasks(); 4008 // We allow that there may be no tasks to do here because 4009 // we are restarting after a stack overflow. 4010 assert(pst->valid() || n_tasks == 0, "Uninitialized use?"); 4011 uint nth_task = 0; 4012 4013 HeapWord* aligned_start = sp->bottom(); 4014 if (sp->used_region().contains(_restart_addr)) { 4015 // Align down to a card boundary for the start of 0th task 4016 // for this space. 4017 aligned_start = 4018 (HeapWord*)align_size_down((uintptr_t)_restart_addr, 4019 CardTableModRefBS::card_size); 4020 } 4021 4022 size_t chunk_size = sp->marking_task_size(); 4023 while (!pst->is_task_claimed(/* reference */ nth_task)) { 4024 // Having claimed the nth task in this space, 4025 // compute the chunk that it corresponds to: 4026 MemRegion span = MemRegion(aligned_start + nth_task*chunk_size, 4027 aligned_start + (nth_task+1)*chunk_size); 4028 // Try and bump the global finger via a CAS; 4029 // note that we need to do the global finger bump 4030 // _before_ taking the intersection below, because 4031 // the task corresponding to that region will be 4032 // deemed done even if the used_region() expands 4033 // because of allocation -- as it almost certainly will 4034 // during start-up while the threads yield in the 4035 // closure below. 4036 HeapWord* finger = span.end(); 4037 bump_global_finger(finger); // atomically 4038 // There are null tasks here corresponding to chunks 4039 // beyond the "top" address of the space. 4040 span = span.intersection(sp->used_region()); 4041 if (!span.is_empty()) { // Non-null task 4042 HeapWord* prev_obj; 4043 assert(!span.contains(_restart_addr) || nth_task == 0, 4044 "Inconsistency"); 4045 if (nth_task == 0) { 4046 // For the 0th task, we'll not need to compute a block_start. 4047 if (span.contains(_restart_addr)) { 4048 // In the case of a restart because of stack overflow, 4049 // we might additionally skip a chunk prefix. 4050 prev_obj = _restart_addr; 4051 } else { 4052 prev_obj = span.start(); 4053 } 4054 } else { 4055 // We want to skip the first object because 4056 // the protocol is to scan any object in its entirety 4057 // that _starts_ in this span; a fortiori, any 4058 // object starting in an earlier span is scanned 4059 // as part of an earlier claimed task. 4060 // Below we use the "careful" version of block_start 4061 // so we do not try to navigate uninitialized objects. 4062 prev_obj = sp->block_start_careful(span.start()); 4063 // Below we use a variant of block_size that uses the 4064 // Printezis bits to avoid waiting for allocated 4065 // objects to become initialized/parsable. 4066 while (prev_obj < span.start()) { 4067 size_t sz = sp->block_size_no_stall(prev_obj, _collector); 4068 if (sz > 0) { 4069 prev_obj += sz; 4070 } else { 4071 // In this case we may end up doing a bit of redundant 4072 // scanning, but that appears unavoidable, short of 4073 // locking the free list locks; see bug 6324141. 4074 break; 4075 } 4076 } 4077 } 4078 if (prev_obj < span.end()) { 4079 MemRegion my_span = MemRegion(prev_obj, span.end()); 4080 // Do the marking work within a non-empty span -- 4081 // the last argument to the constructor indicates whether the 4082 // iteration should be incremental with periodic yields. 4083 Par_MarkFromRootsClosure cl(this, _collector, my_span, 4084 &_collector->_markBitMap, 4085 work_queue(i), 4086 &_collector->_markStack, 4087 &_collector->_revisitStack, 4088 _asynch); 4089 _collector->_markBitMap.iterate(&cl, my_span.start(), my_span.end()); 4090 } // else nothing to do for this task 4091 } // else nothing to do for this task 4092 } 4093 // We'd be tempted to assert here that since there are no 4094 // more tasks left to claim in this space, the global_finger 4095 // must exceed space->top() and a fortiori space->end(). However, 4096 // that would not quite be correct because the bumping of 4097 // global_finger occurs strictly after the claiming of a task, 4098 // so by the time we reach here the global finger may not yet 4099 // have been bumped up by the thread that claimed the last 4100 // task. 4101 pst->all_tasks_completed(); 4102 } 4103 4104 class Par_ConcMarkingClosure: public Par_KlassRememberingOopClosure { 4105 private: 4106 CMSConcMarkingTask* _task; 4107 MemRegion _span; 4108 CMSBitMap* _bit_map; 4109 CMSMarkStack* _overflow_stack; 4110 OopTaskQueue* _work_queue; 4111 protected: 4112 DO_OOP_WORK_DEFN 4113 public: 4114 Par_ConcMarkingClosure(CMSCollector* collector, CMSConcMarkingTask* task, OopTaskQueue* work_queue, 4115 CMSBitMap* bit_map, CMSMarkStack* overflow_stack, 4116 CMSMarkStack* revisit_stack): 4117 Par_KlassRememberingOopClosure(collector, collector->ref_processor(), revisit_stack), 4118 _task(task), 4119 _span(collector->_span), 4120 _work_queue(work_queue), 4121 _bit_map(bit_map), 4122 _overflow_stack(overflow_stack) 4123 { } 4124 virtual void do_oop(oop* p); 4125 virtual void do_oop(narrowOop* p); 4126 void trim_queue(size_t max); 4127 void handle_stack_overflow(HeapWord* lost); 4128 void do_yield_check() { 4129 if (_task->should_yield()) { 4130 _task->yield(); 4131 } 4132 } 4133 }; 4134 4135 // Grey object scanning during work stealing phase -- 4136 // the salient assumption here is that any references 4137 // that are in these stolen objects being scanned must 4138 // already have been initialized (else they would not have 4139 // been published), so we do not need to check for 4140 // uninitialized objects before pushing here. 4141 void Par_ConcMarkingClosure::do_oop(oop obj) { 4142 assert(obj->is_oop_or_null(true), "expected an oop or NULL"); 4143 HeapWord* addr = (HeapWord*)obj; 4144 // Check if oop points into the CMS generation 4145 // and is not marked 4146 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 4147 // a white object ... 4148 // If we manage to "claim" the object, by being the 4149 // first thread to mark it, then we push it on our 4150 // marking stack 4151 if (_bit_map->par_mark(addr)) { // ... now grey 4152 // push on work queue (grey set) 4153 bool simulate_overflow = false; 4154 NOT_PRODUCT( 4155 if (CMSMarkStackOverflowALot && 4156 _collector->simulate_overflow()) { 4157 // simulate a stack overflow 4158 simulate_overflow = true; 4159 } 4160 ) 4161 if (simulate_overflow || 4162 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) { 4163 // stack overflow 4164 if (PrintCMSStatistics != 0) { 4165 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 4166 SIZE_FORMAT, _overflow_stack->capacity()); 4167 } 4168 // We cannot assert that the overflow stack is full because 4169 // it may have been emptied since. 4170 assert(simulate_overflow || 4171 _work_queue->size() == _work_queue->max_elems(), 4172 "Else push should have succeeded"); 4173 handle_stack_overflow(addr); 4174 } 4175 } // Else, some other thread got there first 4176 do_yield_check(); 4177 } 4178 } 4179 4180 void Par_ConcMarkingClosure::do_oop(oop* p) { Par_ConcMarkingClosure::do_oop_work(p); } 4181 void Par_ConcMarkingClosure::do_oop(narrowOop* p) { Par_ConcMarkingClosure::do_oop_work(p); } 4182 4183 void Par_ConcMarkingClosure::trim_queue(size_t max) { 4184 while (_work_queue->size() > max) { 4185 oop new_oop; 4186 if (_work_queue->pop_local(new_oop)) { 4187 assert(new_oop->is_oop(), "Should be an oop"); 4188 assert(_bit_map->isMarked((HeapWord*)new_oop), "Grey object"); 4189 assert(_span.contains((HeapWord*)new_oop), "Not in span"); 4190 assert(new_oop->is_parsable(), "Should be parsable"); 4191 new_oop->oop_iterate(this); // do_oop() above 4192 do_yield_check(); 4193 } 4194 } 4195 } 4196 4197 // Upon stack overflow, we discard (part of) the stack, 4198 // remembering the least address amongst those discarded 4199 // in CMSCollector's _restart_address. 4200 void Par_ConcMarkingClosure::handle_stack_overflow(HeapWord* lost) { 4201 // We need to do this under a mutex to prevent other 4202 // workers from interfering with the work done below. 4203 MutexLockerEx ml(_overflow_stack->par_lock(), 4204 Mutex::_no_safepoint_check_flag); 4205 // Remember the least grey address discarded 4206 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 4207 _collector->lower_restart_addr(ra); 4208 _overflow_stack->reset(); // discard stack contents 4209 _overflow_stack->expand(); // expand the stack if possible 4210 } 4211 4212 4213 void CMSConcMarkingTask::do_work_steal(int i) { 4214 OopTaskQueue* work_q = work_queue(i); 4215 oop obj_to_scan; 4216 CMSBitMap* bm = &(_collector->_markBitMap); 4217 CMSMarkStack* ovflw = &(_collector->_markStack); 4218 CMSMarkStack* revisit = &(_collector->_revisitStack); 4219 int* seed = _collector->hash_seed(i); 4220 Par_ConcMarkingClosure cl(_collector, this, work_q, bm, ovflw, revisit); 4221 while (true) { 4222 cl.trim_queue(0); 4223 assert(work_q->size() == 0, "Should have been emptied above"); 4224 if (get_work_from_overflow_stack(ovflw, work_q)) { 4225 // Can't assert below because the work obtained from the 4226 // overflow stack may already have been stolen from us. 4227 // assert(work_q->size() > 0, "Work from overflow stack"); 4228 continue; 4229 } else if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 4230 assert(obj_to_scan->is_oop(), "Should be an oop"); 4231 assert(bm->isMarked((HeapWord*)obj_to_scan), "Grey object"); 4232 obj_to_scan->oop_iterate(&cl); 4233 } else if (terminator()->offer_termination(&_term_term)) { 4234 assert(work_q->size() == 0, "Impossible!"); 4235 break; 4236 } else if (yielding() || should_yield()) { 4237 yield(); 4238 } 4239 } 4240 } 4241 4242 // This is run by the CMS (coordinator) thread. 4243 void CMSConcMarkingTask::coordinator_yield() { 4244 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 4245 "CMS thread should hold CMS token"); 4246 DEBUG_ONLY(RememberKlassesChecker mux(false);) 4247 // First give up the locks, then yield, then re-lock 4248 // We should probably use a constructor/destructor idiom to 4249 // do this unlock/lock or modify the MutexUnlocker class to 4250 // serve our purpose. XXX 4251 assert_lock_strong(_bit_map_lock); 4252 _bit_map_lock->unlock(); 4253 ConcurrentMarkSweepThread::desynchronize(true); 4254 ConcurrentMarkSweepThread::acknowledge_yield_request(); 4255 _collector->stopTimer(); 4256 if (PrintCMSStatistics != 0) { 4257 _collector->incrementYields(); 4258 } 4259 _collector->icms_wait(); 4260 4261 // It is possible for whichever thread initiated the yield request 4262 // not to get a chance to wake up and take the bitmap lock between 4263 // this thread releasing it and reacquiring it. So, while the 4264 // should_yield() flag is on, let's sleep for a bit to give the 4265 // other thread a chance to wake up. The limit imposed on the number 4266 // of iterations is defensive, to avoid any unforseen circumstances 4267 // putting us into an infinite loop. Since it's always been this 4268 // (coordinator_yield()) method that was observed to cause the 4269 // problem, we are using a parameter (CMSCoordinatorYieldSleepCount) 4270 // which is by default non-zero. For the other seven methods that 4271 // also perform the yield operation, as are using a different 4272 // parameter (CMSYieldSleepCount) which is by default zero. This way we 4273 // can enable the sleeping for those methods too, if necessary. 4274 // See 6442774. 4275 // 4276 // We really need to reconsider the synchronization between the GC 4277 // thread and the yield-requesting threads in the future and we 4278 // should really use wait/notify, which is the recommended 4279 // way of doing this type of interaction. Additionally, we should 4280 // consolidate the eight methods that do the yield operation and they 4281 // are almost identical into one for better maintenability and 4282 // readability. See 6445193. 4283 // 4284 // Tony 2006.06.29 4285 for (unsigned i = 0; i < CMSCoordinatorYieldSleepCount && 4286 ConcurrentMarkSweepThread::should_yield() && 4287 !CMSCollector::foregroundGCIsActive(); ++i) { 4288 os::sleep(Thread::current(), 1, false); 4289 ConcurrentMarkSweepThread::acknowledge_yield_request(); 4290 } 4291 4292 ConcurrentMarkSweepThread::synchronize(true); 4293 _bit_map_lock->lock_without_safepoint_check(); 4294 _collector->startTimer(); 4295 } 4296 4297 bool CMSCollector::do_marking_mt(bool asynch) { 4298 assert(ConcGCThreads > 0 && conc_workers() != NULL, "precondition"); 4299 int num_workers = AdaptiveSizePolicy::calc_active_conc_workers( 4300 conc_workers()->total_workers(), 4301 conc_workers()->active_workers(), 4302 Threads::number_of_non_daemon_threads()); 4303 conc_workers()->set_active_workers(num_workers); 4304 4305 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 4306 CompactibleFreeListSpace* perm_space = _permGen->cmsSpace(); 4307 4308 CMSConcMarkingTask tsk(this, 4309 cms_space, 4310 perm_space, 4311 asynch, 4312 conc_workers(), 4313 task_queues()); 4314 4315 // Since the actual number of workers we get may be different 4316 // from the number we requested above, do we need to do anything different 4317 // below? In particular, may be we need to subclass the SequantialSubTasksDone 4318 // class?? XXX 4319 cms_space ->initialize_sequential_subtasks_for_marking(num_workers); 4320 perm_space->initialize_sequential_subtasks_for_marking(num_workers); 4321 4322 // Refs discovery is already non-atomic. 4323 assert(!ref_processor()->discovery_is_atomic(), "Should be non-atomic"); 4324 assert(ref_processor()->discovery_is_mt(), "Discovery should be MT"); 4325 DEBUG_ONLY(RememberKlassesChecker cmx(should_unload_classes());) 4326 conc_workers()->start_task(&tsk); 4327 while (tsk.yielded()) { 4328 tsk.coordinator_yield(); 4329 conc_workers()->continue_task(&tsk); 4330 } 4331 // If the task was aborted, _restart_addr will be non-NULL 4332 assert(tsk.completed() || _restart_addr != NULL, "Inconsistency"); 4333 while (_restart_addr != NULL) { 4334 // XXX For now we do not make use of ABORTED state and have not 4335 // yet implemented the right abort semantics (even in the original 4336 // single-threaded CMS case). That needs some more investigation 4337 // and is deferred for now; see CR# TBF. 07252005YSR. XXX 4338 assert(!CMSAbortSemantics || tsk.aborted(), "Inconsistency"); 4339 // If _restart_addr is non-NULL, a marking stack overflow 4340 // occurred; we need to do a fresh marking iteration from the 4341 // indicated restart address. 4342 if (_foregroundGCIsActive && asynch) { 4343 // We may be running into repeated stack overflows, having 4344 // reached the limit of the stack size, while making very 4345 // slow forward progress. It may be best to bail out and 4346 // let the foreground collector do its job. 4347 // Clear _restart_addr, so that foreground GC 4348 // works from scratch. This avoids the headache of 4349 // a "rescan" which would otherwise be needed because 4350 // of the dirty mod union table & card table. 4351 _restart_addr = NULL; 4352 return false; 4353 } 4354 // Adjust the task to restart from _restart_addr 4355 tsk.reset(_restart_addr); 4356 cms_space ->initialize_sequential_subtasks_for_marking(num_workers, 4357 _restart_addr); 4358 perm_space->initialize_sequential_subtasks_for_marking(num_workers, 4359 _restart_addr); 4360 _restart_addr = NULL; 4361 // Get the workers going again 4362 conc_workers()->start_task(&tsk); 4363 while (tsk.yielded()) { 4364 tsk.coordinator_yield(); 4365 conc_workers()->continue_task(&tsk); 4366 } 4367 } 4368 assert(tsk.completed(), "Inconsistency"); 4369 assert(tsk.result() == true, "Inconsistency"); 4370 return true; 4371 } 4372 4373 bool CMSCollector::do_marking_st(bool asynch) { 4374 ResourceMark rm; 4375 HandleMark hm; 4376 4377 // Temporarily make refs discovery single threaded (non-MT) 4378 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false); 4379 MarkFromRootsClosure markFromRootsClosure(this, _span, &_markBitMap, 4380 &_markStack, &_revisitStack, CMSYield && asynch); 4381 // the last argument to iterate indicates whether the iteration 4382 // should be incremental with periodic yields. 4383 _markBitMap.iterate(&markFromRootsClosure); 4384 // If _restart_addr is non-NULL, a marking stack overflow 4385 // occurred; we need to do a fresh iteration from the 4386 // indicated restart address. 4387 while (_restart_addr != NULL) { 4388 if (_foregroundGCIsActive && asynch) { 4389 // We may be running into repeated stack overflows, having 4390 // reached the limit of the stack size, while making very 4391 // slow forward progress. It may be best to bail out and 4392 // let the foreground collector do its job. 4393 // Clear _restart_addr, so that foreground GC 4394 // works from scratch. This avoids the headache of 4395 // a "rescan" which would otherwise be needed because 4396 // of the dirty mod union table & card table. 4397 _restart_addr = NULL; 4398 return false; // indicating failure to complete marking 4399 } 4400 // Deal with stack overflow: 4401 // we restart marking from _restart_addr 4402 HeapWord* ra = _restart_addr; 4403 markFromRootsClosure.reset(ra); 4404 _restart_addr = NULL; 4405 _markBitMap.iterate(&markFromRootsClosure, ra, _span.end()); 4406 } 4407 return true; 4408 } 4409 4410 void CMSCollector::preclean() { 4411 check_correct_thread_executing(); 4412 assert(Thread::current()->is_ConcurrentGC_thread(), "Wrong thread"); 4413 verify_work_stacks_empty(); 4414 verify_overflow_empty(); 4415 _abort_preclean = false; 4416 if (CMSPrecleaningEnabled) { 4417 _eden_chunk_index = 0; 4418 size_t used = get_eden_used(); 4419 size_t capacity = get_eden_capacity(); 4420 // Don't start sampling unless we will get sufficiently 4421 // many samples. 4422 if (used < (capacity/(CMSScheduleRemarkSamplingRatio * 100) 4423 * CMSScheduleRemarkEdenPenetration)) { 4424 _start_sampling = true; 4425 } else { 4426 _start_sampling = false; 4427 } 4428 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 4429 CMSPhaseAccounting pa(this, "preclean", !PrintGCDetails); 4430 preclean_work(CMSPrecleanRefLists1, CMSPrecleanSurvivors1); 4431 } 4432 CMSTokenSync x(true); // is cms thread 4433 if (CMSPrecleaningEnabled) { 4434 sample_eden(); 4435 _collectorState = AbortablePreclean; 4436 } else { 4437 _collectorState = FinalMarking; 4438 } 4439 verify_work_stacks_empty(); 4440 verify_overflow_empty(); 4441 } 4442 4443 // Try and schedule the remark such that young gen 4444 // occupancy is CMSScheduleRemarkEdenPenetration %. 4445 void CMSCollector::abortable_preclean() { 4446 check_correct_thread_executing(); 4447 assert(CMSPrecleaningEnabled, "Inconsistent control state"); 4448 assert(_collectorState == AbortablePreclean, "Inconsistent control state"); 4449 4450 // If Eden's current occupancy is below this threshold, 4451 // immediately schedule the remark; else preclean 4452 // past the next scavenge in an effort to 4453 // schedule the pause as described avove. By choosing 4454 // CMSScheduleRemarkEdenSizeThreshold >= max eden size 4455 // we will never do an actual abortable preclean cycle. 4456 if (get_eden_used() > CMSScheduleRemarkEdenSizeThreshold) { 4457 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 4458 CMSPhaseAccounting pa(this, "abortable-preclean", !PrintGCDetails); 4459 // We need more smarts in the abortable preclean 4460 // loop below to deal with cases where allocation 4461 // in young gen is very very slow, and our precleaning 4462 // is running a losing race against a horde of 4463 // mutators intent on flooding us with CMS updates 4464 // (dirty cards). 4465 // One, admittedly dumb, strategy is to give up 4466 // after a certain number of abortable precleaning loops 4467 // or after a certain maximum time. We want to make 4468 // this smarter in the next iteration. 4469 // XXX FIX ME!!! YSR 4470 size_t loops = 0, workdone = 0, cumworkdone = 0, waited = 0; 4471 while (!(should_abort_preclean() || 4472 ConcurrentMarkSweepThread::should_terminate())) { 4473 workdone = preclean_work(CMSPrecleanRefLists2, CMSPrecleanSurvivors2); 4474 cumworkdone += workdone; 4475 loops++; 4476 // Voluntarily terminate abortable preclean phase if we have 4477 // been at it for too long. 4478 if ((CMSMaxAbortablePrecleanLoops != 0) && 4479 loops >= CMSMaxAbortablePrecleanLoops) { 4480 if (PrintGCDetails) { 4481 gclog_or_tty->print(" CMS: abort preclean due to loops "); 4482 } 4483 break; 4484 } 4485 if (pa.wallclock_millis() > CMSMaxAbortablePrecleanTime) { 4486 if (PrintGCDetails) { 4487 gclog_or_tty->print(" CMS: abort preclean due to time "); 4488 } 4489 break; 4490 } 4491 // If we are doing little work each iteration, we should 4492 // take a short break. 4493 if (workdone < CMSAbortablePrecleanMinWorkPerIteration) { 4494 // Sleep for some time, waiting for work to accumulate 4495 stopTimer(); 4496 cmsThread()->wait_on_cms_lock(CMSAbortablePrecleanWaitMillis); 4497 startTimer(); 4498 waited++; 4499 } 4500 } 4501 if (PrintCMSStatistics > 0) { 4502 gclog_or_tty->print(" [%d iterations, %d waits, %d cards)] ", 4503 loops, waited, cumworkdone); 4504 } 4505 } 4506 CMSTokenSync x(true); // is cms thread 4507 if (_collectorState != Idling) { 4508 assert(_collectorState == AbortablePreclean, 4509 "Spontaneous state transition?"); 4510 _collectorState = FinalMarking; 4511 } // Else, a foreground collection completed this CMS cycle. 4512 return; 4513 } 4514 4515 // Respond to an Eden sampling opportunity 4516 void CMSCollector::sample_eden() { 4517 // Make sure a young gc cannot sneak in between our 4518 // reading and recording of a sample. 4519 assert(Thread::current()->is_ConcurrentGC_thread(), 4520 "Only the cms thread may collect Eden samples"); 4521 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 4522 "Should collect samples while holding CMS token"); 4523 if (!_start_sampling) { 4524 return; 4525 } 4526 if (_eden_chunk_array) { 4527 if (_eden_chunk_index < _eden_chunk_capacity) { 4528 _eden_chunk_array[_eden_chunk_index] = *_top_addr; // take sample 4529 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr, 4530 "Unexpected state of Eden"); 4531 // We'd like to check that what we just sampled is an oop-start address; 4532 // however, we cannot do that here since the object may not yet have been 4533 // initialized. So we'll instead do the check when we _use_ this sample 4534 // later. 4535 if (_eden_chunk_index == 0 || 4536 (pointer_delta(_eden_chunk_array[_eden_chunk_index], 4537 _eden_chunk_array[_eden_chunk_index-1]) 4538 >= CMSSamplingGrain)) { 4539 _eden_chunk_index++; // commit sample 4540 } 4541 } 4542 } 4543 if ((_collectorState == AbortablePreclean) && !_abort_preclean) { 4544 size_t used = get_eden_used(); 4545 size_t capacity = get_eden_capacity(); 4546 assert(used <= capacity, "Unexpected state of Eden"); 4547 if (used > (capacity/100 * CMSScheduleRemarkEdenPenetration)) { 4548 _abort_preclean = true; 4549 } 4550 } 4551 } 4552 4553 4554 size_t CMSCollector::preclean_work(bool clean_refs, bool clean_survivor) { 4555 assert(_collectorState == Precleaning || 4556 _collectorState == AbortablePreclean, "incorrect state"); 4557 ResourceMark rm; 4558 HandleMark hm; 4559 4560 // Precleaning is currently not MT but the reference processor 4561 // may be set for MT. Disable it temporarily here. 4562 ReferenceProcessor* rp = ref_processor(); 4563 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(rp, false); 4564 4565 // Do one pass of scrubbing the discovered reference lists 4566 // to remove any reference objects with strongly-reachable 4567 // referents. 4568 if (clean_refs) { 4569 CMSPrecleanRefsYieldClosure yield_cl(this); 4570 assert(rp->span().equals(_span), "Spans should be equal"); 4571 CMSKeepAliveClosure keep_alive(this, _span, &_markBitMap, 4572 &_markStack, &_revisitStack, 4573 true /* preclean */); 4574 CMSDrainMarkingStackClosure complete_trace(this, 4575 _span, &_markBitMap, &_markStack, 4576 &keep_alive, true /* preclean */); 4577 4578 // We don't want this step to interfere with a young 4579 // collection because we don't want to take CPU 4580 // or memory bandwidth away from the young GC threads 4581 // (which may be as many as there are CPUs). 4582 // Note that we don't need to protect ourselves from 4583 // interference with mutators because they can't 4584 // manipulate the discovered reference lists nor affect 4585 // the computed reachability of the referents, the 4586 // only properties manipulated by the precleaning 4587 // of these reference lists. 4588 stopTimer(); 4589 CMSTokenSyncWithLocks x(true /* is cms thread */, 4590 bitMapLock()); 4591 startTimer(); 4592 sample_eden(); 4593 4594 // The following will yield to allow foreground 4595 // collection to proceed promptly. XXX YSR: 4596 // The code in this method may need further 4597 // tweaking for better performance and some restructuring 4598 // for cleaner interfaces. 4599 GCTimer *gc_timer = NULL; // Currently not tracing concurrent phases 4600 rp->preclean_discovered_references( 4601 rp->is_alive_non_header(), &keep_alive, &complete_trace, 4602 &yield_cl, should_unload_classes(), gc_timer); 4603 } 4604 4605 if (clean_survivor) { // preclean the active survivor space(s) 4606 assert(_young_gen->kind() == Generation::DefNew || 4607 _young_gen->kind() == Generation::ParNew || 4608 _young_gen->kind() == Generation::ASParNew, 4609 "incorrect type for cast"); 4610 DefNewGeneration* dng = (DefNewGeneration*)_young_gen; 4611 PushAndMarkClosure pam_cl(this, _span, ref_processor(), 4612 &_markBitMap, &_modUnionTable, 4613 &_markStack, &_revisitStack, 4614 true /* precleaning phase */); 4615 stopTimer(); 4616 CMSTokenSyncWithLocks ts(true /* is cms thread */, 4617 bitMapLock()); 4618 startTimer(); 4619 unsigned int before_count = 4620 GenCollectedHeap::heap()->total_collections(); 4621 SurvivorSpacePrecleanClosure 4622 sss_cl(this, _span, &_markBitMap, &_markStack, 4623 &pam_cl, before_count, CMSYield); 4624 DEBUG_ONLY(RememberKlassesChecker mx(should_unload_classes());) 4625 dng->from()->object_iterate_careful(&sss_cl); 4626 dng->to()->object_iterate_careful(&sss_cl); 4627 } 4628 MarkRefsIntoAndScanClosure 4629 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable, 4630 &_markStack, &_revisitStack, this, CMSYield, 4631 true /* precleaning phase */); 4632 // CAUTION: The following closure has persistent state that may need to 4633 // be reset upon a decrease in the sequence of addresses it 4634 // processes. 4635 ScanMarkedObjectsAgainCarefullyClosure 4636 smoac_cl(this, _span, 4637 &_markBitMap, &_markStack, &_revisitStack, &mrias_cl, CMSYield); 4638 4639 // Preclean dirty cards in ModUnionTable and CardTable using 4640 // appropriate convergence criterion; 4641 // repeat CMSPrecleanIter times unless we find that 4642 // we are losing. 4643 assert(CMSPrecleanIter < 10, "CMSPrecleanIter is too large"); 4644 assert(CMSPrecleanNumerator < CMSPrecleanDenominator, 4645 "Bad convergence multiplier"); 4646 assert(CMSPrecleanThreshold >= 100, 4647 "Unreasonably low CMSPrecleanThreshold"); 4648 4649 size_t numIter, cumNumCards, lastNumCards, curNumCards; 4650 for (numIter = 0, cumNumCards = lastNumCards = curNumCards = 0; 4651 numIter < CMSPrecleanIter; 4652 numIter++, lastNumCards = curNumCards, cumNumCards += curNumCards) { 4653 curNumCards = preclean_mod_union_table(_cmsGen, &smoac_cl); 4654 if (CMSPermGenPrecleaningEnabled) { 4655 curNumCards += preclean_mod_union_table(_permGen, &smoac_cl); 4656 } 4657 if (Verbose && PrintGCDetails) { 4658 gclog_or_tty->print(" (modUnionTable: %d cards)", curNumCards); 4659 } 4660 // Either there are very few dirty cards, so re-mark 4661 // pause will be small anyway, or our pre-cleaning isn't 4662 // that much faster than the rate at which cards are being 4663 // dirtied, so we might as well stop and re-mark since 4664 // precleaning won't improve our re-mark time by much. 4665 if (curNumCards <= CMSPrecleanThreshold || 4666 (numIter > 0 && 4667 (curNumCards * CMSPrecleanDenominator > 4668 lastNumCards * CMSPrecleanNumerator))) { 4669 numIter++; 4670 cumNumCards += curNumCards; 4671 break; 4672 } 4673 } 4674 curNumCards = preclean_card_table(_cmsGen, &smoac_cl); 4675 if (CMSPermGenPrecleaningEnabled) { 4676 curNumCards += preclean_card_table(_permGen, &smoac_cl); 4677 } 4678 cumNumCards += curNumCards; 4679 if (PrintGCDetails && PrintCMSStatistics != 0) { 4680 gclog_or_tty->print_cr(" (cardTable: %d cards, re-scanned %d cards, %d iterations)", 4681 curNumCards, cumNumCards, numIter); 4682 } 4683 return cumNumCards; // as a measure of useful work done 4684 } 4685 4686 // PRECLEANING NOTES: 4687 // Precleaning involves: 4688 // . reading the bits of the modUnionTable and clearing the set bits. 4689 // . For the cards corresponding to the set bits, we scan the 4690 // objects on those cards. This means we need the free_list_lock 4691 // so that we can safely iterate over the CMS space when scanning 4692 // for oops. 4693 // . When we scan the objects, we'll be both reading and setting 4694 // marks in the marking bit map, so we'll need the marking bit map. 4695 // . For protecting _collector_state transitions, we take the CGC_lock. 4696 // Note that any races in the reading of of card table entries by the 4697 // CMS thread on the one hand and the clearing of those entries by the 4698 // VM thread or the setting of those entries by the mutator threads on the 4699 // other are quite benign. However, for efficiency it makes sense to keep 4700 // the VM thread from racing with the CMS thread while the latter is 4701 // dirty card info to the modUnionTable. We therefore also use the 4702 // CGC_lock to protect the reading of the card table and the mod union 4703 // table by the CM thread. 4704 // . We run concurrently with mutator updates, so scanning 4705 // needs to be done carefully -- we should not try to scan 4706 // potentially uninitialized objects. 4707 // 4708 // Locking strategy: While holding the CGC_lock, we scan over and 4709 // reset a maximal dirty range of the mod union / card tables, then lock 4710 // the free_list_lock and bitmap lock to do a full marking, then 4711 // release these locks; and repeat the cycle. This allows for a 4712 // certain amount of fairness in the sharing of these locks between 4713 // the CMS collector on the one hand, and the VM thread and the 4714 // mutators on the other. 4715 4716 // NOTE: preclean_mod_union_table() and preclean_card_table() 4717 // further below are largely identical; if you need to modify 4718 // one of these methods, please check the other method too. 4719 4720 size_t CMSCollector::preclean_mod_union_table( 4721 ConcurrentMarkSweepGeneration* gen, 4722 ScanMarkedObjectsAgainCarefullyClosure* cl) { 4723 verify_work_stacks_empty(); 4724 verify_overflow_empty(); 4725 4726 // Turn off checking for this method but turn it back on 4727 // selectively. There are yield points in this method 4728 // but it is difficult to turn the checking off just around 4729 // the yield points. It is simpler to selectively turn 4730 // it on. 4731 DEBUG_ONLY(RememberKlassesChecker mux(false);) 4732 4733 // strategy: starting with the first card, accumulate contiguous 4734 // ranges of dirty cards; clear these cards, then scan the region 4735 // covered by these cards. 4736 4737 // Since all of the MUT is committed ahead, we can just use 4738 // that, in case the generations expand while we are precleaning. 4739 // It might also be fine to just use the committed part of the 4740 // generation, but we might potentially miss cards when the 4741 // generation is rapidly expanding while we are in the midst 4742 // of precleaning. 4743 HeapWord* startAddr = gen->reserved().start(); 4744 HeapWord* endAddr = gen->reserved().end(); 4745 4746 cl->setFreelistLock(gen->freelistLock()); // needed for yielding 4747 4748 size_t numDirtyCards, cumNumDirtyCards; 4749 HeapWord *nextAddr, *lastAddr; 4750 for (cumNumDirtyCards = numDirtyCards = 0, 4751 nextAddr = lastAddr = startAddr; 4752 nextAddr < endAddr; 4753 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 4754 4755 ResourceMark rm; 4756 HandleMark hm; 4757 4758 MemRegion dirtyRegion; 4759 { 4760 stopTimer(); 4761 // Potential yield point 4762 CMSTokenSync ts(true); 4763 startTimer(); 4764 sample_eden(); 4765 // Get dirty region starting at nextOffset (inclusive), 4766 // simultaneously clearing it. 4767 dirtyRegion = 4768 _modUnionTable.getAndClearMarkedRegion(nextAddr, endAddr); 4769 assert(dirtyRegion.start() >= nextAddr, 4770 "returned region inconsistent?"); 4771 } 4772 // Remember where the next search should begin. 4773 // The returned region (if non-empty) is a right open interval, 4774 // so lastOffset is obtained from the right end of that 4775 // interval. 4776 lastAddr = dirtyRegion.end(); 4777 // Should do something more transparent and less hacky XXX 4778 numDirtyCards = 4779 _modUnionTable.heapWordDiffToOffsetDiff(dirtyRegion.word_size()); 4780 4781 // We'll scan the cards in the dirty region (with periodic 4782 // yields for foreground GC as needed). 4783 if (!dirtyRegion.is_empty()) { 4784 assert(numDirtyCards > 0, "consistency check"); 4785 HeapWord* stop_point = NULL; 4786 stopTimer(); 4787 // Potential yield point 4788 CMSTokenSyncWithLocks ts(true, gen->freelistLock(), 4789 bitMapLock()); 4790 startTimer(); 4791 { 4792 verify_work_stacks_empty(); 4793 verify_overflow_empty(); 4794 sample_eden(); 4795 DEBUG_ONLY(RememberKlassesChecker mx(should_unload_classes());) 4796 stop_point = 4797 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 4798 } 4799 if (stop_point != NULL) { 4800 // The careful iteration stopped early either because it found an 4801 // uninitialized object, or because we were in the midst of an 4802 // "abortable preclean", which should now be aborted. Redirty 4803 // the bits corresponding to the partially-scanned or unscanned 4804 // cards. We'll either restart at the next block boundary or 4805 // abort the preclean. 4806 assert((CMSPermGenPrecleaningEnabled && (gen == _permGen)) || 4807 (_collectorState == AbortablePreclean && should_abort_preclean()), 4808 "Unparsable objects should only be in perm gen."); 4809 _modUnionTable.mark_range(MemRegion(stop_point, dirtyRegion.end())); 4810 if (should_abort_preclean()) { 4811 break; // out of preclean loop 4812 } else { 4813 // Compute the next address at which preclean should pick up; 4814 // might need bitMapLock in order to read P-bits. 4815 lastAddr = next_card_start_after_block(stop_point); 4816 } 4817 } 4818 } else { 4819 assert(lastAddr == endAddr, "consistency check"); 4820 assert(numDirtyCards == 0, "consistency check"); 4821 break; 4822 } 4823 } 4824 verify_work_stacks_empty(); 4825 verify_overflow_empty(); 4826 return cumNumDirtyCards; 4827 } 4828 4829 // NOTE: preclean_mod_union_table() above and preclean_card_table() 4830 // below are largely identical; if you need to modify 4831 // one of these methods, please check the other method too. 4832 4833 size_t CMSCollector::preclean_card_table(ConcurrentMarkSweepGeneration* gen, 4834 ScanMarkedObjectsAgainCarefullyClosure* cl) { 4835 // strategy: it's similar to precleamModUnionTable above, in that 4836 // we accumulate contiguous ranges of dirty cards, mark these cards 4837 // precleaned, then scan the region covered by these cards. 4838 HeapWord* endAddr = (HeapWord*)(gen->_virtual_space.high()); 4839 HeapWord* startAddr = (HeapWord*)(gen->_virtual_space.low()); 4840 4841 cl->setFreelistLock(gen->freelistLock()); // needed for yielding 4842 4843 size_t numDirtyCards, cumNumDirtyCards; 4844 HeapWord *lastAddr, *nextAddr; 4845 4846 for (cumNumDirtyCards = numDirtyCards = 0, 4847 nextAddr = lastAddr = startAddr; 4848 nextAddr < endAddr; 4849 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 4850 4851 ResourceMark rm; 4852 HandleMark hm; 4853 4854 MemRegion dirtyRegion; 4855 { 4856 // See comments in "Precleaning notes" above on why we 4857 // do this locking. XXX Could the locking overheads be 4858 // too high when dirty cards are sparse? [I don't think so.] 4859 stopTimer(); 4860 CMSTokenSync x(true); // is cms thread 4861 startTimer(); 4862 sample_eden(); 4863 // Get and clear dirty region from card table 4864 dirtyRegion = _ct->ct_bs()->dirty_card_range_after_reset( 4865 MemRegion(nextAddr, endAddr), 4866 true, 4867 CardTableModRefBS::precleaned_card_val()); 4868 4869 assert(dirtyRegion.start() >= nextAddr, 4870 "returned region inconsistent?"); 4871 } 4872 lastAddr = dirtyRegion.end(); 4873 numDirtyCards = 4874 dirtyRegion.word_size()/CardTableModRefBS::card_size_in_words; 4875 4876 if (!dirtyRegion.is_empty()) { 4877 stopTimer(); 4878 CMSTokenSyncWithLocks ts(true, gen->freelistLock(), bitMapLock()); 4879 startTimer(); 4880 sample_eden(); 4881 verify_work_stacks_empty(); 4882 verify_overflow_empty(); 4883 DEBUG_ONLY(RememberKlassesChecker mx(should_unload_classes());) 4884 HeapWord* stop_point = 4885 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 4886 if (stop_point != NULL) { 4887 // The careful iteration stopped early because it found an 4888 // uninitialized object. Redirty the bits corresponding to the 4889 // partially-scanned or unscanned cards, and start again at the 4890 // next block boundary. 4891 assert(CMSPermGenPrecleaningEnabled || 4892 (_collectorState == AbortablePreclean && should_abort_preclean()), 4893 "Unparsable objects should only be in perm gen."); 4894 _ct->ct_bs()->invalidate(MemRegion(stop_point, dirtyRegion.end())); 4895 if (should_abort_preclean()) { 4896 break; // out of preclean loop 4897 } else { 4898 // Compute the next address at which preclean should pick up. 4899 lastAddr = next_card_start_after_block(stop_point); 4900 } 4901 } 4902 } else { 4903 break; 4904 } 4905 } 4906 verify_work_stacks_empty(); 4907 verify_overflow_empty(); 4908 return cumNumDirtyCards; 4909 } 4910 4911 void CMSCollector::checkpointRootsFinal(bool asynch, 4912 bool clear_all_soft_refs, bool init_mark_was_synchronous) { 4913 assert(_collectorState == FinalMarking, "incorrect state transition?"); 4914 check_correct_thread_executing(); 4915 // world is stopped at this checkpoint 4916 assert(SafepointSynchronize::is_at_safepoint(), 4917 "world should be stopped"); 4918 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 4919 4920 verify_work_stacks_empty(); 4921 verify_overflow_empty(); 4922 4923 SpecializationStats::clear(); 4924 if (PrintGCDetails) { 4925 gclog_or_tty->print("[YG occupancy: "SIZE_FORMAT" K ("SIZE_FORMAT" K)]", 4926 _young_gen->used() / K, 4927 _young_gen->capacity() / K); 4928 } 4929 if (asynch) { 4930 if (CMSScavengeBeforeRemark) { 4931 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4932 // Temporarily set flag to false, GCH->do_collection will 4933 // expect it to be false and set to true 4934 FlagSetting fl(gch->_is_gc_active, false); 4935 NOT_PRODUCT(GCTraceTime t("Scavenge-Before-Remark", 4936 PrintGCDetails && Verbose, true, _gc_timer_cm);) 4937 int level = _cmsGen->level() - 1; 4938 if (level >= 0) { 4939 gch->do_collection(true, // full (i.e. force, see below) 4940 false, // !clear_all_soft_refs 4941 0, // size 4942 false, // is_tlab 4943 level // max_level 4944 ); 4945 } 4946 } 4947 FreelistLocker x(this); 4948 MutexLockerEx y(bitMapLock(), 4949 Mutex::_no_safepoint_check_flag); 4950 assert(!init_mark_was_synchronous, "but that's impossible!"); 4951 checkpointRootsFinalWork(asynch, clear_all_soft_refs, false); 4952 } else { 4953 // already have all the locks 4954 checkpointRootsFinalWork(asynch, clear_all_soft_refs, 4955 init_mark_was_synchronous); 4956 } 4957 verify_work_stacks_empty(); 4958 verify_overflow_empty(); 4959 SpecializationStats::print(); 4960 } 4961 4962 void CMSCollector::checkpointRootsFinalWork(bool asynch, 4963 bool clear_all_soft_refs, bool init_mark_was_synchronous) { 4964 4965 NOT_PRODUCT(GCTraceTime tr("checkpointRootsFinalWork", PrintGCDetails, false, _gc_timer_cm);) 4966 4967 assert(haveFreelistLocks(), "must have free list locks"); 4968 assert_lock_strong(bitMapLock()); 4969 4970 if (UseAdaptiveSizePolicy) { 4971 size_policy()->checkpoint_roots_final_begin(); 4972 } 4973 4974 ResourceMark rm; 4975 HandleMark hm; 4976 4977 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4978 4979 if (should_unload_classes()) { 4980 CodeCache::gc_prologue(); 4981 } 4982 assert(haveFreelistLocks(), "must have free list locks"); 4983 assert_lock_strong(bitMapLock()); 4984 4985 DEBUG_ONLY(RememberKlassesChecker fmx(should_unload_classes());) 4986 if (!init_mark_was_synchronous) { 4987 // We might assume that we need not fill TLAB's when 4988 // CMSScavengeBeforeRemark is set, because we may have just done 4989 // a scavenge which would have filled all TLAB's -- and besides 4990 // Eden would be empty. This however may not always be the case -- 4991 // for instance although we asked for a scavenge, it may not have 4992 // happened because of a JNI critical section. We probably need 4993 // a policy for deciding whether we can in that case wait until 4994 // the critical section releases and then do the remark following 4995 // the scavenge, and skip it here. In the absence of that policy, 4996 // or of an indication of whether the scavenge did indeed occur, 4997 // we cannot rely on TLAB's having been filled and must do 4998 // so here just in case a scavenge did not happen. 4999 gch->ensure_parsability(false); // fill TLAB's, but no need to retire them 5000 // Update the saved marks which may affect the root scans. 5001 gch->save_marks(); 5002 5003 { 5004 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;) 5005 5006 // Note on the role of the mod union table: 5007 // Since the marker in "markFromRoots" marks concurrently with 5008 // mutators, it is possible for some reachable objects not to have been 5009 // scanned. For instance, an only reference to an object A was 5010 // placed in object B after the marker scanned B. Unless B is rescanned, 5011 // A would be collected. Such updates to references in marked objects 5012 // are detected via the mod union table which is the set of all cards 5013 // dirtied since the first checkpoint in this GC cycle and prior to 5014 // the most recent young generation GC, minus those cleaned up by the 5015 // concurrent precleaning. 5016 if (CMSParallelRemarkEnabled && CollectedHeap::use_parallel_gc_threads()) { 5017 GCTraceTime t("Rescan (parallel) ", PrintGCDetails, false, _gc_timer_cm); 5018 do_remark_parallel(); 5019 } else { 5020 GCTraceTime t("Rescan (non-parallel) ", PrintGCDetails, false, 5021 _gc_timer_cm); 5022 do_remark_non_parallel(); 5023 } 5024 } 5025 } else { 5026 assert(!asynch, "Can't have init_mark_was_synchronous in asynch mode"); 5027 // The initial mark was stop-world, so there's no rescanning to 5028 // do; go straight on to the next step below. 5029 } 5030 verify_work_stacks_empty(); 5031 verify_overflow_empty(); 5032 5033 { 5034 NOT_PRODUCT(GCTraceTime ts("refProcessingWork", PrintGCDetails, false, _gc_timer_cm);) 5035 refProcessingWork(asynch, clear_all_soft_refs); 5036 } 5037 verify_work_stacks_empty(); 5038 verify_overflow_empty(); 5039 5040 if (should_unload_classes()) { 5041 CodeCache::gc_epilogue(); 5042 } 5043 JvmtiExport::gc_epilogue(); 5044 5045 // If we encountered any (marking stack / work queue) overflow 5046 // events during the current CMS cycle, take appropriate 5047 // remedial measures, where possible, so as to try and avoid 5048 // recurrence of that condition. 5049 assert(_markStack.isEmpty(), "No grey objects"); 5050 size_t ser_ovflw = _ser_pmc_remark_ovflw + _ser_pmc_preclean_ovflw + 5051 _ser_kac_ovflw + _ser_kac_preclean_ovflw; 5052 if (ser_ovflw > 0) { 5053 if (PrintCMSStatistics != 0) { 5054 gclog_or_tty->print_cr("Marking stack overflow (benign) " 5055 "(pmc_pc="SIZE_FORMAT", pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT 5056 ", kac_preclean="SIZE_FORMAT")", 5057 _ser_pmc_preclean_ovflw, _ser_pmc_remark_ovflw, 5058 _ser_kac_ovflw, _ser_kac_preclean_ovflw); 5059 } 5060 _markStack.expand(); 5061 _ser_pmc_remark_ovflw = 0; 5062 _ser_pmc_preclean_ovflw = 0; 5063 _ser_kac_preclean_ovflw = 0; 5064 _ser_kac_ovflw = 0; 5065 } 5066 if (_par_pmc_remark_ovflw > 0 || _par_kac_ovflw > 0) { 5067 if (PrintCMSStatistics != 0) { 5068 gclog_or_tty->print_cr("Work queue overflow (benign) " 5069 "(pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT")", 5070 _par_pmc_remark_ovflw, _par_kac_ovflw); 5071 } 5072 _par_pmc_remark_ovflw = 0; 5073 _par_kac_ovflw = 0; 5074 } 5075 if (PrintCMSStatistics != 0) { 5076 if (_markStack._hit_limit > 0) { 5077 gclog_or_tty->print_cr(" (benign) Hit max stack size limit ("SIZE_FORMAT")", 5078 _markStack._hit_limit); 5079 } 5080 if (_markStack._failed_double > 0) { 5081 gclog_or_tty->print_cr(" (benign) Failed stack doubling ("SIZE_FORMAT")," 5082 " current capacity "SIZE_FORMAT, 5083 _markStack._failed_double, 5084 _markStack.capacity()); 5085 } 5086 } 5087 _markStack._hit_limit = 0; 5088 _markStack._failed_double = 0; 5089 5090 // Check that all the klasses have been checked 5091 assert(_revisitStack.isEmpty(), "Not all klasses revisited"); 5092 5093 if ((VerifyAfterGC || VerifyDuringGC) && 5094 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 5095 verify_after_remark(); 5096 } 5097 5098 _gc_tracer_cm->report_object_count_after_gc(&_is_alive_closure); 5099 5100 // Change under the freelistLocks. 5101 _collectorState = Sweeping; 5102 // Call isAllClear() under bitMapLock 5103 assert(_modUnionTable.isAllClear(), "Should be clear by end of the" 5104 " final marking"); 5105 if (UseAdaptiveSizePolicy) { 5106 size_policy()->checkpoint_roots_final_end(gch->gc_cause()); 5107 } 5108 } 5109 5110 // Parallel remark task 5111 class CMSParRemarkTask: public AbstractGangTask { 5112 CMSCollector* _collector; 5113 int _n_workers; 5114 CompactibleFreeListSpace* _cms_space; 5115 CompactibleFreeListSpace* _perm_space; 5116 5117 // The per-thread work queues, available here for stealing. 5118 OopTaskQueueSet* _task_queues; 5119 ParallelTaskTerminator _term; 5120 5121 public: 5122 // A value of 0 passed to n_workers will cause the number of 5123 // workers to be taken from the active workers in the work gang. 5124 CMSParRemarkTask(CMSCollector* collector, 5125 CompactibleFreeListSpace* cms_space, 5126 CompactibleFreeListSpace* perm_space, 5127 int n_workers, FlexibleWorkGang* workers, 5128 OopTaskQueueSet* task_queues): 5129 AbstractGangTask("Rescan roots and grey objects in parallel"), 5130 _collector(collector), 5131 _cms_space(cms_space), _perm_space(perm_space), 5132 _n_workers(n_workers), 5133 _task_queues(task_queues), 5134 _term(n_workers, task_queues) { } 5135 5136 OopTaskQueueSet* task_queues() { return _task_queues; } 5137 5138 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 5139 5140 ParallelTaskTerminator* terminator() { return &_term; } 5141 int n_workers() { return _n_workers; } 5142 5143 void work(uint worker_id); 5144 5145 private: 5146 // Work method in support of parallel rescan ... of young gen spaces 5147 void do_young_space_rescan(int i, Par_MarkRefsIntoAndScanClosure* cl, 5148 ContiguousSpace* space, 5149 HeapWord** chunk_array, size_t chunk_top); 5150 5151 // ... of dirty cards in old space 5152 void do_dirty_card_rescan_tasks(CompactibleFreeListSpace* sp, int i, 5153 Par_MarkRefsIntoAndScanClosure* cl); 5154 5155 // ... work stealing for the above 5156 void do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, int* seed); 5157 }; 5158 5159 // work_queue(i) is passed to the closure 5160 // Par_MarkRefsIntoAndScanClosure. The "i" parameter 5161 // also is passed to do_dirty_card_rescan_tasks() and to 5162 // do_work_steal() to select the i-th task_queue. 5163 5164 void CMSParRemarkTask::work(uint worker_id) { 5165 elapsedTimer _timer; 5166 ResourceMark rm; 5167 HandleMark hm; 5168 5169 // ---------- rescan from roots -------------- 5170 _timer.start(); 5171 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5172 Par_MarkRefsIntoAndScanClosure par_mrias_cl(_collector, 5173 _collector->_span, _collector->ref_processor(), 5174 &(_collector->_markBitMap), 5175 work_queue(worker_id), &(_collector->_revisitStack)); 5176 5177 // Rescan young gen roots first since these are likely 5178 // coarsely partitioned and may, on that account, constitute 5179 // the critical path; thus, it's best to start off that 5180 // work first. 5181 // ---------- young gen roots -------------- 5182 { 5183 DefNewGeneration* dng = _collector->_young_gen->as_DefNewGeneration(); 5184 EdenSpace* eden_space = dng->eden(); 5185 ContiguousSpace* from_space = dng->from(); 5186 ContiguousSpace* to_space = dng->to(); 5187 5188 HeapWord** eca = _collector->_eden_chunk_array; 5189 size_t ect = _collector->_eden_chunk_index; 5190 HeapWord** sca = _collector->_survivor_chunk_array; 5191 size_t sct = _collector->_survivor_chunk_index; 5192 5193 assert(ect <= _collector->_eden_chunk_capacity, "out of bounds"); 5194 assert(sct <= _collector->_survivor_chunk_capacity, "out of bounds"); 5195 5196 do_young_space_rescan(worker_id, &par_mrias_cl, to_space, NULL, 0); 5197 do_young_space_rescan(worker_id, &par_mrias_cl, from_space, sca, sct); 5198 do_young_space_rescan(worker_id, &par_mrias_cl, eden_space, eca, ect); 5199 5200 _timer.stop(); 5201 if (PrintCMSStatistics != 0) { 5202 gclog_or_tty->print_cr( 5203 "Finished young gen rescan work in %dth thread: %3.3f sec", 5204 worker_id, _timer.seconds()); 5205 } 5206 } 5207 5208 // ---------- remaining roots -------------- 5209 _timer.reset(); 5210 _timer.start(); 5211 gch->gen_process_strong_roots(_collector->_cmsGen->level(), 5212 false, // yg was scanned above 5213 false, // this is parallel code 5214 true, // collecting perm gen 5215 SharedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()), 5216 &par_mrias_cl, 5217 true, // walk all of code cache if (so & SO_CodeCache) 5218 NULL); 5219 assert(_collector->should_unload_classes() 5220 || (_collector->CMSCollector::roots_scanning_options() & SharedHeap::SO_CodeCache), 5221 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 5222 _timer.stop(); 5223 if (PrintCMSStatistics != 0) { 5224 gclog_or_tty->print_cr( 5225 "Finished remaining root rescan work in %dth thread: %3.3f sec", 5226 worker_id, _timer.seconds()); 5227 } 5228 5229 // ---------- rescan dirty cards ------------ 5230 _timer.reset(); 5231 _timer.start(); 5232 5233 // Do the rescan tasks for each of the two spaces 5234 // (cms_space and perm_space) in turn. 5235 // "worker_id" is passed to select the task_queue for "worker_id" 5236 do_dirty_card_rescan_tasks(_cms_space, worker_id, &par_mrias_cl); 5237 do_dirty_card_rescan_tasks(_perm_space, worker_id, &par_mrias_cl); 5238 _timer.stop(); 5239 if (PrintCMSStatistics != 0) { 5240 gclog_or_tty->print_cr( 5241 "Finished dirty card rescan work in %dth thread: %3.3f sec", 5242 worker_id, _timer.seconds()); 5243 } 5244 5245 // ---------- steal work from other threads ... 5246 // ---------- ... and drain overflow list. 5247 _timer.reset(); 5248 _timer.start(); 5249 do_work_steal(worker_id, &par_mrias_cl, _collector->hash_seed(worker_id)); 5250 _timer.stop(); 5251 if (PrintCMSStatistics != 0) { 5252 gclog_or_tty->print_cr( 5253 "Finished work stealing in %dth thread: %3.3f sec", 5254 worker_id, _timer.seconds()); 5255 } 5256 } 5257 5258 // Note that parameter "i" is not used. 5259 void 5260 CMSParRemarkTask::do_young_space_rescan(int i, 5261 Par_MarkRefsIntoAndScanClosure* cl, ContiguousSpace* space, 5262 HeapWord** chunk_array, size_t chunk_top) { 5263 // Until all tasks completed: 5264 // . claim an unclaimed task 5265 // . compute region boundaries corresponding to task claimed 5266 // using chunk_array 5267 // . par_oop_iterate(cl) over that region 5268 5269 ResourceMark rm; 5270 HandleMark hm; 5271 5272 SequentialSubTasksDone* pst = space->par_seq_tasks(); 5273 assert(pst->valid(), "Uninitialized use?"); 5274 5275 uint nth_task = 0; 5276 uint n_tasks = pst->n_tasks(); 5277 5278 HeapWord *start, *end; 5279 while (!pst->is_task_claimed(/* reference */ nth_task)) { 5280 // We claimed task # nth_task; compute its boundaries. 5281 if (chunk_top == 0) { // no samples were taken 5282 assert(nth_task == 0 && n_tasks == 1, "Can have only 1 EdenSpace task"); 5283 start = space->bottom(); 5284 end = space->top(); 5285 } else if (nth_task == 0) { 5286 start = space->bottom(); 5287 end = chunk_array[nth_task]; 5288 } else if (nth_task < (uint)chunk_top) { 5289 assert(nth_task >= 1, "Control point invariant"); 5290 start = chunk_array[nth_task - 1]; 5291 end = chunk_array[nth_task]; 5292 } else { 5293 assert(nth_task == (uint)chunk_top, "Control point invariant"); 5294 start = chunk_array[chunk_top - 1]; 5295 end = space->top(); 5296 } 5297 MemRegion mr(start, end); 5298 // Verify that mr is in space 5299 assert(mr.is_empty() || space->used_region().contains(mr), 5300 "Should be in space"); 5301 // Verify that "start" is an object boundary 5302 assert(mr.is_empty() || oop(mr.start())->is_oop(), 5303 "Should be an oop"); 5304 space->par_oop_iterate(mr, cl); 5305 } 5306 pst->all_tasks_completed(); 5307 } 5308 5309 void 5310 CMSParRemarkTask::do_dirty_card_rescan_tasks( 5311 CompactibleFreeListSpace* sp, int i, 5312 Par_MarkRefsIntoAndScanClosure* cl) { 5313 // Until all tasks completed: 5314 // . claim an unclaimed task 5315 // . compute region boundaries corresponding to task claimed 5316 // . transfer dirty bits ct->mut for that region 5317 // . apply rescanclosure to dirty mut bits for that region 5318 5319 ResourceMark rm; 5320 HandleMark hm; 5321 5322 OopTaskQueue* work_q = work_queue(i); 5323 ModUnionClosure modUnionClosure(&(_collector->_modUnionTable)); 5324 // CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! 5325 // CAUTION: This closure has state that persists across calls to 5326 // the work method dirty_range_iterate_clear() in that it has 5327 // imbedded in it a (subtype of) UpwardsObjectClosure. The 5328 // use of that state in the imbedded UpwardsObjectClosure instance 5329 // assumes that the cards are always iterated (even if in parallel 5330 // by several threads) in monotonically increasing order per each 5331 // thread. This is true of the implementation below which picks 5332 // card ranges (chunks) in monotonically increasing order globally 5333 // and, a-fortiori, in monotonically increasing order per thread 5334 // (the latter order being a subsequence of the former). 5335 // If the work code below is ever reorganized into a more chaotic 5336 // work-partitioning form than the current "sequential tasks" 5337 // paradigm, the use of that persistent state will have to be 5338 // revisited and modified appropriately. See also related 5339 // bug 4756801 work on which should examine this code to make 5340 // sure that the changes there do not run counter to the 5341 // assumptions made here and necessary for correctness and 5342 // efficiency. Note also that this code might yield inefficient 5343 // behaviour in the case of very large objects that span one or 5344 // more work chunks. Such objects would potentially be scanned 5345 // several times redundantly. Work on 4756801 should try and 5346 // address that performance anomaly if at all possible. XXX 5347 MemRegion full_span = _collector->_span; 5348 CMSBitMap* bm = &(_collector->_markBitMap); // shared 5349 CMSMarkStack* rs = &(_collector->_revisitStack); // shared 5350 MarkFromDirtyCardsClosure 5351 greyRescanClosure(_collector, full_span, // entire span of interest 5352 sp, bm, work_q, rs, cl); 5353 5354 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 5355 assert(pst->valid(), "Uninitialized use?"); 5356 uint nth_task = 0; 5357 const int alignment = CardTableModRefBS::card_size * BitsPerWord; 5358 MemRegion span = sp->used_region(); 5359 HeapWord* start_addr = span.start(); 5360 HeapWord* end_addr = (HeapWord*)round_to((intptr_t)span.end(), 5361 alignment); 5362 const size_t chunk_size = sp->rescan_task_size(); // in HeapWord units 5363 assert((HeapWord*)round_to((intptr_t)start_addr, alignment) == 5364 start_addr, "Check alignment"); 5365 assert((size_t)round_to((intptr_t)chunk_size, alignment) == 5366 chunk_size, "Check alignment"); 5367 5368 while (!pst->is_task_claimed(/* reference */ nth_task)) { 5369 // Having claimed the nth_task, compute corresponding mem-region, 5370 // which is a-fortiori aligned correctly (i.e. at a MUT bopundary). 5371 // The alignment restriction ensures that we do not need any 5372 // synchronization with other gang-workers while setting or 5373 // clearing bits in thus chunk of the MUT. 5374 MemRegion this_span = MemRegion(start_addr + nth_task*chunk_size, 5375 start_addr + (nth_task+1)*chunk_size); 5376 // The last chunk's end might be way beyond end of the 5377 // used region. In that case pull back appropriately. 5378 if (this_span.end() > end_addr) { 5379 this_span.set_end(end_addr); 5380 assert(!this_span.is_empty(), "Program logic (calculation of n_tasks)"); 5381 } 5382 // Iterate over the dirty cards covering this chunk, marking them 5383 // precleaned, and setting the corresponding bits in the mod union 5384 // table. Since we have been careful to partition at Card and MUT-word 5385 // boundaries no synchronization is needed between parallel threads. 5386 _collector->_ct->ct_bs()->dirty_card_iterate(this_span, 5387 &modUnionClosure); 5388 5389 // Having transferred these marks into the modUnionTable, 5390 // rescan the marked objects on the dirty cards in the modUnionTable. 5391 // Even if this is at a synchronous collection, the initial marking 5392 // may have been done during an asynchronous collection so there 5393 // may be dirty bits in the mod-union table. 5394 _collector->_modUnionTable.dirty_range_iterate_clear( 5395 this_span, &greyRescanClosure); 5396 _collector->_modUnionTable.verifyNoOneBitsInRange( 5397 this_span.start(), 5398 this_span.end()); 5399 } 5400 pst->all_tasks_completed(); // declare that i am done 5401 } 5402 5403 // . see if we can share work_queues with ParNew? XXX 5404 void 5405 CMSParRemarkTask::do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, 5406 int* seed) { 5407 OopTaskQueue* work_q = work_queue(i); 5408 NOT_PRODUCT(int num_steals = 0;) 5409 oop obj_to_scan; 5410 CMSBitMap* bm = &(_collector->_markBitMap); 5411 5412 while (true) { 5413 // Completely finish any left over work from (an) earlier round(s) 5414 cl->trim_queue(0); 5415 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 5416 (size_t)ParGCDesiredObjsFromOverflowList); 5417 // Now check if there's any work in the overflow list 5418 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads, 5419 // only affects the number of attempts made to get work from the 5420 // overflow list and does not affect the number of workers. Just 5421 // pass ParallelGCThreads so this behavior is unchanged. 5422 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 5423 work_q, 5424 ParallelGCThreads)) { 5425 // found something in global overflow list; 5426 // not yet ready to go stealing work from others. 5427 // We'd like to assert(work_q->size() != 0, ...) 5428 // because we just took work from the overflow list, 5429 // but of course we can't since all of that could have 5430 // been already stolen from us. 5431 // "He giveth and He taketh away." 5432 continue; 5433 } 5434 // Verify that we have no work before we resort to stealing 5435 assert(work_q->size() == 0, "Have work, shouldn't steal"); 5436 // Try to steal from other queues that have work 5437 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 5438 NOT_PRODUCT(num_steals++;) 5439 assert(obj_to_scan->is_oop(), "Oops, not an oop!"); 5440 assert(bm->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 5441 // Do scanning work 5442 obj_to_scan->oop_iterate(cl); 5443 // Loop around, finish this work, and try to steal some more 5444 } else if (terminator()->offer_termination()) { 5445 break; // nirvana from the infinite cycle 5446 } 5447 } 5448 NOT_PRODUCT( 5449 if (PrintCMSStatistics != 0) { 5450 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals); 5451 } 5452 ) 5453 assert(work_q->size() == 0 && _collector->overflow_list_is_empty(), 5454 "Else our work is not yet done"); 5455 } 5456 5457 // Return a thread-local PLAB recording array, as appropriate. 5458 void* CMSCollector::get_data_recorder(int thr_num) { 5459 if (_survivor_plab_array != NULL && 5460 (CMSPLABRecordAlways || 5461 (_collectorState > Marking && _collectorState < FinalMarking))) { 5462 assert(thr_num < (int)ParallelGCThreads, "thr_num is out of bounds"); 5463 ChunkArray* ca = &_survivor_plab_array[thr_num]; 5464 ca->reset(); // clear it so that fresh data is recorded 5465 return (void*) ca; 5466 } else { 5467 return NULL; 5468 } 5469 } 5470 5471 // Reset all the thread-local PLAB recording arrays 5472 void CMSCollector::reset_survivor_plab_arrays() { 5473 for (uint i = 0; i < ParallelGCThreads; i++) { 5474 _survivor_plab_array[i].reset(); 5475 } 5476 } 5477 5478 // Merge the per-thread plab arrays into the global survivor chunk 5479 // array which will provide the partitioning of the survivor space 5480 // for CMS rescan. 5481 void CMSCollector::merge_survivor_plab_arrays(ContiguousSpace* surv, 5482 int no_of_gc_threads) { 5483 assert(_survivor_plab_array != NULL, "Error"); 5484 assert(_survivor_chunk_array != NULL, "Error"); 5485 assert(_collectorState == FinalMarking, "Error"); 5486 for (int j = 0; j < no_of_gc_threads; j++) { 5487 _cursor[j] = 0; 5488 } 5489 HeapWord* top = surv->top(); 5490 size_t i; 5491 for (i = 0; i < _survivor_chunk_capacity; i++) { // all sca entries 5492 HeapWord* min_val = top; // Higher than any PLAB address 5493 uint min_tid = 0; // position of min_val this round 5494 for (int j = 0; j < no_of_gc_threads; j++) { 5495 ChunkArray* cur_sca = &_survivor_plab_array[j]; 5496 if (_cursor[j] == cur_sca->end()) { 5497 continue; 5498 } 5499 assert(_cursor[j] < cur_sca->end(), "ctl pt invariant"); 5500 HeapWord* cur_val = cur_sca->nth(_cursor[j]); 5501 assert(surv->used_region().contains(cur_val), "Out of bounds value"); 5502 if (cur_val < min_val) { 5503 min_tid = j; 5504 min_val = cur_val; 5505 } else { 5506 assert(cur_val < top, "All recorded addresses should be less"); 5507 } 5508 } 5509 // At this point min_val and min_tid are respectively 5510 // the least address in _survivor_plab_array[j]->nth(_cursor[j]) 5511 // and the thread (j) that witnesses that address. 5512 // We record this address in the _survivor_chunk_array[i] 5513 // and increment _cursor[min_tid] prior to the next round i. 5514 if (min_val == top) { 5515 break; 5516 } 5517 _survivor_chunk_array[i] = min_val; 5518 _cursor[min_tid]++; 5519 } 5520 // We are all done; record the size of the _survivor_chunk_array 5521 _survivor_chunk_index = i; // exclusive: [0, i) 5522 if (PrintCMSStatistics > 0) { 5523 gclog_or_tty->print(" (Survivor:" SIZE_FORMAT "chunks) ", i); 5524 } 5525 // Verify that we used up all the recorded entries 5526 #ifdef ASSERT 5527 size_t total = 0; 5528 for (int j = 0; j < no_of_gc_threads; j++) { 5529 assert(_cursor[j] == _survivor_plab_array[j].end(), "Ctl pt invariant"); 5530 total += _cursor[j]; 5531 } 5532 assert(total == _survivor_chunk_index, "Ctl Pt Invariant"); 5533 // Check that the merged array is in sorted order 5534 if (total > 0) { 5535 for (size_t i = 0; i < total - 1; i++) { 5536 if (PrintCMSStatistics > 0) { 5537 gclog_or_tty->print(" (chunk" SIZE_FORMAT ":" INTPTR_FORMAT ") ", 5538 i, _survivor_chunk_array[i]); 5539 } 5540 assert(_survivor_chunk_array[i] < _survivor_chunk_array[i+1], 5541 "Not sorted"); 5542 } 5543 } 5544 #endif // ASSERT 5545 } 5546 5547 // Set up the space's par_seq_tasks structure for work claiming 5548 // for parallel rescan of young gen. 5549 // See ParRescanTask where this is currently used. 5550 void 5551 CMSCollector:: 5552 initialize_sequential_subtasks_for_young_gen_rescan(int n_threads) { 5553 assert(n_threads > 0, "Unexpected n_threads argument"); 5554 DefNewGeneration* dng = (DefNewGeneration*)_young_gen; 5555 5556 // Eden space 5557 { 5558 SequentialSubTasksDone* pst = dng->eden()->par_seq_tasks(); 5559 assert(!pst->valid(), "Clobbering existing data?"); 5560 // Each valid entry in [0, _eden_chunk_index) represents a task. 5561 size_t n_tasks = _eden_chunk_index + 1; 5562 assert(n_tasks == 1 || _eden_chunk_array != NULL, "Error"); 5563 // Sets the condition for completion of the subtask (how many threads 5564 // need to finish in order to be done). 5565 pst->set_n_threads(n_threads); 5566 pst->set_n_tasks((int)n_tasks); 5567 } 5568 5569 // Merge the survivor plab arrays into _survivor_chunk_array 5570 if (_survivor_plab_array != NULL) { 5571 merge_survivor_plab_arrays(dng->from(), n_threads); 5572 } else { 5573 assert(_survivor_chunk_index == 0, "Error"); 5574 } 5575 5576 // To space 5577 { 5578 SequentialSubTasksDone* pst = dng->to()->par_seq_tasks(); 5579 assert(!pst->valid(), "Clobbering existing data?"); 5580 // Sets the condition for completion of the subtask (how many threads 5581 // need to finish in order to be done). 5582 pst->set_n_threads(n_threads); 5583 pst->set_n_tasks(1); 5584 assert(pst->valid(), "Error"); 5585 } 5586 5587 // From space 5588 { 5589 SequentialSubTasksDone* pst = dng->from()->par_seq_tasks(); 5590 assert(!pst->valid(), "Clobbering existing data?"); 5591 size_t n_tasks = _survivor_chunk_index + 1; 5592 assert(n_tasks == 1 || _survivor_chunk_array != NULL, "Error"); 5593 // Sets the condition for completion of the subtask (how many threads 5594 // need to finish in order to be done). 5595 pst->set_n_threads(n_threads); 5596 pst->set_n_tasks((int)n_tasks); 5597 assert(pst->valid(), "Error"); 5598 } 5599 } 5600 5601 // Parallel version of remark 5602 void CMSCollector::do_remark_parallel() { 5603 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5604 FlexibleWorkGang* workers = gch->workers(); 5605 assert(workers != NULL, "Need parallel worker threads."); 5606 // Choose to use the number of GC workers most recently set 5607 // into "active_workers". If active_workers is not set, set it 5608 // to ParallelGCThreads. 5609 int n_workers = workers->active_workers(); 5610 if (n_workers == 0) { 5611 assert(n_workers > 0, "Should have been set during scavenge"); 5612 n_workers = ParallelGCThreads; 5613 workers->set_active_workers(n_workers); 5614 } 5615 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 5616 CompactibleFreeListSpace* perm_space = _permGen->cmsSpace(); 5617 5618 CMSParRemarkTask tsk(this, 5619 cms_space, perm_space, 5620 n_workers, workers, task_queues()); 5621 5622 // Set up for parallel process_strong_roots work. 5623 gch->set_par_threads(n_workers); 5624 // We won't be iterating over the cards in the card table updating 5625 // the younger_gen cards, so we shouldn't call the following else 5626 // the verification code as well as subsequent younger_refs_iterate 5627 // code would get confused. XXX 5628 // gch->rem_set()->prepare_for_younger_refs_iterate(true); // parallel 5629 5630 // The young gen rescan work will not be done as part of 5631 // process_strong_roots (which currently doesn't knw how to 5632 // parallelize such a scan), but rather will be broken up into 5633 // a set of parallel tasks (via the sampling that the [abortable] 5634 // preclean phase did of EdenSpace, plus the [two] tasks of 5635 // scanning the [two] survivor spaces. Further fine-grain 5636 // parallelization of the scanning of the survivor spaces 5637 // themselves, and of precleaning of the younger gen itself 5638 // is deferred to the future. 5639 initialize_sequential_subtasks_for_young_gen_rescan(n_workers); 5640 5641 // The dirty card rescan work is broken up into a "sequence" 5642 // of parallel tasks (per constituent space) that are dynamically 5643 // claimed by the parallel threads. 5644 cms_space->initialize_sequential_subtasks_for_rescan(n_workers); 5645 perm_space->initialize_sequential_subtasks_for_rescan(n_workers); 5646 5647 // It turns out that even when we're using 1 thread, doing the work in a 5648 // separate thread causes wide variance in run times. We can't help this 5649 // in the multi-threaded case, but we special-case n=1 here to get 5650 // repeatable measurements of the 1-thread overhead of the parallel code. 5651 if (n_workers > 1) { 5652 // Make refs discovery MT-safe, if it isn't already: it may not 5653 // necessarily be so, since it's possible that we are doing 5654 // ST marking. 5655 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), true); 5656 GenCollectedHeap::StrongRootsScope srs(gch); 5657 workers->run_task(&tsk); 5658 } else { 5659 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false); 5660 GenCollectedHeap::StrongRootsScope srs(gch); 5661 tsk.work(0); 5662 } 5663 gch->set_par_threads(0); // 0 ==> non-parallel. 5664 // restore, single-threaded for now, any preserved marks 5665 // as a result of work_q overflow 5666 restore_preserved_marks_if_any(); 5667 } 5668 5669 // Non-parallel version of remark 5670 void CMSCollector::do_remark_non_parallel() { 5671 ResourceMark rm; 5672 HandleMark hm; 5673 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5674 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false); 5675 5676 MarkRefsIntoAndScanClosure 5677 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable, 5678 &_markStack, &_revisitStack, this, 5679 false /* should_yield */, false /* not precleaning */); 5680 MarkFromDirtyCardsClosure 5681 markFromDirtyCardsClosure(this, _span, 5682 NULL, // space is set further below 5683 &_markBitMap, &_markStack, &_revisitStack, 5684 &mrias_cl); 5685 { 5686 GCTraceTime t("grey object rescan", PrintGCDetails, false, _gc_timer_cm); 5687 // Iterate over the dirty cards, setting the corresponding bits in the 5688 // mod union table. 5689 { 5690 ModUnionClosure modUnionClosure(&_modUnionTable); 5691 _ct->ct_bs()->dirty_card_iterate( 5692 _cmsGen->used_region(), 5693 &modUnionClosure); 5694 _ct->ct_bs()->dirty_card_iterate( 5695 _permGen->used_region(), 5696 &modUnionClosure); 5697 } 5698 // Having transferred these marks into the modUnionTable, we just need 5699 // to rescan the marked objects on the dirty cards in the modUnionTable. 5700 // The initial marking may have been done during an asynchronous 5701 // collection so there may be dirty bits in the mod-union table. 5702 const int alignment = 5703 CardTableModRefBS::card_size * BitsPerWord; 5704 { 5705 // ... First handle dirty cards in CMS gen 5706 markFromDirtyCardsClosure.set_space(_cmsGen->cmsSpace()); 5707 MemRegion ur = _cmsGen->used_region(); 5708 HeapWord* lb = ur.start(); 5709 HeapWord* ub = (HeapWord*)round_to((intptr_t)ur.end(), alignment); 5710 MemRegion cms_span(lb, ub); 5711 _modUnionTable.dirty_range_iterate_clear(cms_span, 5712 &markFromDirtyCardsClosure); 5713 verify_work_stacks_empty(); 5714 if (PrintCMSStatistics != 0) { 5715 gclog_or_tty->print(" (re-scanned "SIZE_FORMAT" dirty cards in cms gen) ", 5716 markFromDirtyCardsClosure.num_dirty_cards()); 5717 } 5718 } 5719 { 5720 // .. and then repeat for dirty cards in perm gen 5721 markFromDirtyCardsClosure.set_space(_permGen->cmsSpace()); 5722 MemRegion ur = _permGen->used_region(); 5723 HeapWord* lb = ur.start(); 5724 HeapWord* ub = (HeapWord*)round_to((intptr_t)ur.end(), alignment); 5725 MemRegion perm_span(lb, ub); 5726 _modUnionTable.dirty_range_iterate_clear(perm_span, 5727 &markFromDirtyCardsClosure); 5728 verify_work_stacks_empty(); 5729 if (PrintCMSStatistics != 0) { 5730 gclog_or_tty->print(" (re-scanned "SIZE_FORMAT" dirty cards in perm gen) ", 5731 markFromDirtyCardsClosure.num_dirty_cards()); 5732 } 5733 } 5734 } 5735 if (VerifyDuringGC && 5736 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 5737 HandleMark hm; // Discard invalid handles created during verification 5738 Universe::verify(); 5739 } 5740 { 5741 GCTraceTime t("root rescan", PrintGCDetails, false, _gc_timer_cm); 5742 5743 verify_work_stacks_empty(); 5744 5745 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 5746 GenCollectedHeap::StrongRootsScope srs(gch); 5747 gch->gen_process_strong_roots(_cmsGen->level(), 5748 true, // younger gens as roots 5749 false, // use the local StrongRootsScope 5750 true, // collecting perm gen 5751 SharedHeap::ScanningOption(roots_scanning_options()), 5752 &mrias_cl, 5753 true, // walk code active on stacks 5754 NULL); 5755 assert(should_unload_classes() 5756 || (roots_scanning_options() & SharedHeap::SO_CodeCache), 5757 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 5758 } 5759 verify_work_stacks_empty(); 5760 // Restore evacuated mark words, if any, used for overflow list links 5761 if (!CMSOverflowEarlyRestoration) { 5762 restore_preserved_marks_if_any(); 5763 } 5764 verify_overflow_empty(); 5765 } 5766 5767 //////////////////////////////////////////////////////// 5768 // Parallel Reference Processing Task Proxy Class 5769 //////////////////////////////////////////////////////// 5770 class CMSRefProcTaskProxy: public AbstractGangTaskWOopQueues { 5771 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask; 5772 CMSCollector* _collector; 5773 CMSBitMap* _mark_bit_map; 5774 const MemRegion _span; 5775 ProcessTask& _task; 5776 5777 public: 5778 CMSRefProcTaskProxy(ProcessTask& task, 5779 CMSCollector* collector, 5780 const MemRegion& span, 5781 CMSBitMap* mark_bit_map, 5782 AbstractWorkGang* workers, 5783 OopTaskQueueSet* task_queues): 5784 // XXX Should superclass AGTWOQ also know about AWG since it knows 5785 // about the task_queues used by the AWG? Then it could initialize 5786 // the terminator() object. See 6984287. The set_for_termination() 5787 // below is a temporary band-aid for the regression in 6984287. 5788 AbstractGangTaskWOopQueues("Process referents by policy in parallel", 5789 task_queues), 5790 _task(task), 5791 _collector(collector), _span(span), _mark_bit_map(mark_bit_map) 5792 { 5793 assert(_collector->_span.equals(_span) && !_span.is_empty(), 5794 "Inconsistency in _span"); 5795 set_for_termination(workers->active_workers()); 5796 } 5797 5798 OopTaskQueueSet* task_queues() { return queues(); } 5799 5800 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 5801 5802 void do_work_steal(int i, 5803 CMSParDrainMarkingStackClosure* drain, 5804 CMSParKeepAliveClosure* keep_alive, 5805 int* seed); 5806 5807 virtual void work(uint worker_id); 5808 }; 5809 5810 void CMSRefProcTaskProxy::work(uint worker_id) { 5811 assert(_collector->_span.equals(_span), "Inconsistency in _span"); 5812 CMSParKeepAliveClosure par_keep_alive(_collector, _span, 5813 _mark_bit_map, 5814 &_collector->_revisitStack, 5815 work_queue(worker_id)); 5816 CMSParDrainMarkingStackClosure par_drain_stack(_collector, _span, 5817 _mark_bit_map, 5818 &_collector->_revisitStack, 5819 work_queue(worker_id)); 5820 CMSIsAliveClosure is_alive_closure(_span, _mark_bit_map); 5821 _task.work(worker_id, is_alive_closure, par_keep_alive, par_drain_stack); 5822 if (_task.marks_oops_alive()) { 5823 do_work_steal(worker_id, &par_drain_stack, &par_keep_alive, 5824 _collector->hash_seed(worker_id)); 5825 } 5826 assert(work_queue(worker_id)->size() == 0, "work_queue should be empty"); 5827 assert(_collector->_overflow_list == NULL, "non-empty _overflow_list"); 5828 } 5829 5830 class CMSRefEnqueueTaskProxy: public AbstractGangTask { 5831 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask; 5832 EnqueueTask& _task; 5833 5834 public: 5835 CMSRefEnqueueTaskProxy(EnqueueTask& task) 5836 : AbstractGangTask("Enqueue reference objects in parallel"), 5837 _task(task) 5838 { } 5839 5840 virtual void work(uint worker_id) 5841 { 5842 _task.work(worker_id); 5843 } 5844 }; 5845 5846 CMSParKeepAliveClosure::CMSParKeepAliveClosure(CMSCollector* collector, 5847 MemRegion span, CMSBitMap* bit_map, CMSMarkStack* revisit_stack, 5848 OopTaskQueue* work_queue): 5849 Par_KlassRememberingOopClosure(collector, NULL, revisit_stack), 5850 _span(span), 5851 _bit_map(bit_map), 5852 _work_queue(work_queue), 5853 _mark_and_push(collector, span, bit_map, revisit_stack, work_queue), 5854 _low_water_mark(MIN2((uint)(work_queue->max_elems()/4), 5855 (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads))) 5856 { } 5857 5858 // . see if we can share work_queues with ParNew? XXX 5859 void CMSRefProcTaskProxy::do_work_steal(int i, 5860 CMSParDrainMarkingStackClosure* drain, 5861 CMSParKeepAliveClosure* keep_alive, 5862 int* seed) { 5863 OopTaskQueue* work_q = work_queue(i); 5864 NOT_PRODUCT(int num_steals = 0;) 5865 oop obj_to_scan; 5866 5867 while (true) { 5868 // Completely finish any left over work from (an) earlier round(s) 5869 drain->trim_queue(0); 5870 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 5871 (size_t)ParGCDesiredObjsFromOverflowList); 5872 // Now check if there's any work in the overflow list 5873 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads, 5874 // only affects the number of attempts made to get work from the 5875 // overflow list and does not affect the number of workers. Just 5876 // pass ParallelGCThreads so this behavior is unchanged. 5877 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 5878 work_q, 5879 ParallelGCThreads)) { 5880 // Found something in global overflow list; 5881 // not yet ready to go stealing work from others. 5882 // We'd like to assert(work_q->size() != 0, ...) 5883 // because we just took work from the overflow list, 5884 // but of course we can't, since all of that might have 5885 // been already stolen from us. 5886 continue; 5887 } 5888 // Verify that we have no work before we resort to stealing 5889 assert(work_q->size() == 0, "Have work, shouldn't steal"); 5890 // Try to steal from other queues that have work 5891 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 5892 NOT_PRODUCT(num_steals++;) 5893 assert(obj_to_scan->is_oop(), "Oops, not an oop!"); 5894 assert(_mark_bit_map->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 5895 // Do scanning work 5896 obj_to_scan->oop_iterate(keep_alive); 5897 // Loop around, finish this work, and try to steal some more 5898 } else if (terminator()->offer_termination()) { 5899 break; // nirvana from the infinite cycle 5900 } 5901 } 5902 NOT_PRODUCT( 5903 if (PrintCMSStatistics != 0) { 5904 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals); 5905 } 5906 ) 5907 } 5908 5909 void CMSRefProcTaskExecutor::execute(ProcessTask& task) 5910 { 5911 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5912 FlexibleWorkGang* workers = gch->workers(); 5913 assert(workers != NULL, "Need parallel worker threads."); 5914 CMSRefProcTaskProxy rp_task(task, &_collector, 5915 _collector.ref_processor()->span(), 5916 _collector.markBitMap(), 5917 workers, _collector.task_queues()); 5918 workers->run_task(&rp_task); 5919 } 5920 5921 void CMSRefProcTaskExecutor::execute(EnqueueTask& task) 5922 { 5923 5924 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5925 FlexibleWorkGang* workers = gch->workers(); 5926 assert(workers != NULL, "Need parallel worker threads."); 5927 CMSRefEnqueueTaskProxy enq_task(task); 5928 workers->run_task(&enq_task); 5929 } 5930 5931 void CMSCollector::refProcessingWork(bool asynch, bool clear_all_soft_refs) { 5932 5933 ResourceMark rm; 5934 HandleMark hm; 5935 5936 ReferenceProcessor* rp = ref_processor(); 5937 assert(rp->span().equals(_span), "Spans should be equal"); 5938 assert(!rp->enqueuing_is_done(), "Enqueuing should not be complete"); 5939 // Process weak references. 5940 rp->setup_policy(clear_all_soft_refs); 5941 verify_work_stacks_empty(); 5942 5943 CMSKeepAliveClosure cmsKeepAliveClosure(this, _span, &_markBitMap, 5944 &_markStack, &_revisitStack, 5945 false /* !preclean */); 5946 CMSDrainMarkingStackClosure cmsDrainMarkingStackClosure(this, 5947 _span, &_markBitMap, &_markStack, 5948 &cmsKeepAliveClosure, false /* !preclean */); 5949 { 5950 GCTraceTime t("weak refs processing", PrintGCDetails, false, _gc_timer_cm); 5951 5952 ReferenceProcessorStats stats; 5953 if (rp->processing_is_mt()) { 5954 // Set the degree of MT here. If the discovery is done MT, there 5955 // may have been a different number of threads doing the discovery 5956 // and a different number of discovered lists may have Ref objects. 5957 // That is OK as long as the Reference lists are balanced (see 5958 // balance_all_queues() and balance_queues()). 5959 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5960 int active_workers = ParallelGCThreads; 5961 FlexibleWorkGang* workers = gch->workers(); 5962 if (workers != NULL) { 5963 active_workers = workers->active_workers(); 5964 // The expectation is that active_workers will have already 5965 // been set to a reasonable value. If it has not been set, 5966 // investigate. 5967 assert(active_workers > 0, "Should have been set during scavenge"); 5968 } 5969 rp->set_active_mt_degree(active_workers); 5970 CMSRefProcTaskExecutor task_executor(*this); 5971 stats = rp->process_discovered_references(&_is_alive_closure, 5972 &cmsKeepAliveClosure, 5973 &cmsDrainMarkingStackClosure, 5974 &task_executor, 5975 _gc_timer_cm); 5976 } else { 5977 stats = rp->process_discovered_references(&_is_alive_closure, 5978 &cmsKeepAliveClosure, 5979 &cmsDrainMarkingStackClosure, 5980 NULL, 5981 _gc_timer_cm); 5982 } 5983 _gc_tracer_cm->report_gc_reference_stats(stats); 5984 5985 verify_work_stacks_empty(); 5986 } 5987 5988 if (should_unload_classes()) { 5989 { 5990 GCTraceTime t("class unloading", PrintGCDetails, false, _gc_timer_cm); 5991 5992 // Follow SystemDictionary roots and unload classes 5993 bool purged_class = SystemDictionary::do_unloading(&_is_alive_closure); 5994 5995 // Follow CodeCache roots and unload any methods marked for unloading 5996 CodeCache::do_unloading(&_is_alive_closure, 5997 &cmsKeepAliveClosure, 5998 purged_class); 5999 6000 cmsDrainMarkingStackClosure.do_void(); 6001 verify_work_stacks_empty(); 6002 6003 // Update subklass/sibling/implementor links in KlassKlass descendants 6004 assert(!_revisitStack.isEmpty(), "revisit stack should not be empty"); 6005 oop k; 6006 while ((k = _revisitStack.pop()) != NULL) { 6007 ((Klass*)(oopDesc*)k)->follow_weak_klass_links( 6008 &_is_alive_closure, 6009 &cmsKeepAliveClosure); 6010 } 6011 assert(!ClassUnloading || 6012 (_markStack.isEmpty() && overflow_list_is_empty()), 6013 "Should not have found new reachable objects"); 6014 assert(_revisitStack.isEmpty(), "revisit stack should have been drained"); 6015 cmsDrainMarkingStackClosure.do_void(); 6016 verify_work_stacks_empty(); 6017 } 6018 6019 { 6020 GCTraceTime t("scrub symbol table", PrintGCDetails, false, _gc_timer_cm); 6021 // Clean up unreferenced symbols in symbol table. 6022 SymbolTable::unlink(); 6023 } 6024 } 6025 6026 if (should_unload_classes() || !JavaObjectsInPerm) { 6027 GCTraceTime t("scrub string table", PrintGCDetails, false, _gc_timer_cm); 6028 // Now clean up stale oops in StringTable 6029 StringTable::unlink(&_is_alive_closure); 6030 } 6031 6032 verify_work_stacks_empty(); 6033 // Restore any preserved marks as a result of mark stack or 6034 // work queue overflow 6035 restore_preserved_marks_if_any(); // done single-threaded for now 6036 6037 rp->set_enqueuing_is_done(true); 6038 if (rp->processing_is_mt()) { 6039 rp->balance_all_queues(); 6040 CMSRefProcTaskExecutor task_executor(*this); 6041 rp->enqueue_discovered_references(&task_executor); 6042 } else { 6043 rp->enqueue_discovered_references(NULL); 6044 } 6045 rp->verify_no_references_recorded(); 6046 assert(!rp->discovery_enabled(), "should have been disabled"); 6047 } 6048 6049 #ifndef PRODUCT 6050 void CMSCollector::check_correct_thread_executing() { 6051 Thread* t = Thread::current(); 6052 // Only the VM thread or the CMS thread should be here. 6053 assert(t->is_ConcurrentGC_thread() || t->is_VM_thread(), 6054 "Unexpected thread type"); 6055 // If this is the vm thread, the foreground process 6056 // should not be waiting. Note that _foregroundGCIsActive is 6057 // true while the foreground collector is waiting. 6058 if (_foregroundGCShouldWait) { 6059 // We cannot be the VM thread 6060 assert(t->is_ConcurrentGC_thread(), 6061 "Should be CMS thread"); 6062 } else { 6063 // We can be the CMS thread only if we are in a stop-world 6064 // phase of CMS collection. 6065 if (t->is_ConcurrentGC_thread()) { 6066 assert(_collectorState == InitialMarking || 6067 _collectorState == FinalMarking, 6068 "Should be a stop-world phase"); 6069 // The CMS thread should be holding the CMS_token. 6070 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6071 "Potential interference with concurrently " 6072 "executing VM thread"); 6073 } 6074 } 6075 } 6076 #endif 6077 6078 void CMSCollector::sweep(bool asynch) { 6079 assert(_collectorState == Sweeping, "just checking"); 6080 check_correct_thread_executing(); 6081 verify_work_stacks_empty(); 6082 verify_overflow_empty(); 6083 increment_sweep_count(); 6084 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 6085 6086 _inter_sweep_timer.stop(); 6087 _inter_sweep_estimate.sample(_inter_sweep_timer.seconds()); 6088 size_policy()->avg_cms_free_at_sweep()->sample(_cmsGen->free()); 6089 6090 // PermGen verification support: If perm gen sweeping is disabled in 6091 // this cycle, we preserve the perm gen object "deadness" information 6092 // in the perm_gen_verify_bit_map. In order to do that we traverse 6093 // all blocks in perm gen and mark all dead objects. 6094 if (verifying() && !should_unload_classes()) { 6095 assert(perm_gen_verify_bit_map()->sizeInBits() != 0, 6096 "Should have already been allocated"); 6097 MarkDeadObjectsClosure mdo(this, _permGen->cmsSpace(), 6098 markBitMap(), perm_gen_verify_bit_map()); 6099 if (asynch) { 6100 CMSTokenSyncWithLocks ts(true, _permGen->freelistLock(), 6101 bitMapLock()); 6102 _permGen->cmsSpace()->blk_iterate(&mdo); 6103 } else { 6104 // In the case of synchronous sweep, we already have 6105 // the requisite locks/tokens. 6106 _permGen->cmsSpace()->blk_iterate(&mdo); 6107 } 6108 } 6109 6110 assert(!_intra_sweep_timer.is_active(), "Should not be active"); 6111 _intra_sweep_timer.reset(); 6112 _intra_sweep_timer.start(); 6113 if (asynch) { 6114 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 6115 CMSPhaseAccounting pa(this, "sweep", !PrintGCDetails); 6116 // First sweep the old gen then the perm gen 6117 { 6118 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(), 6119 bitMapLock()); 6120 sweepWork(_cmsGen, asynch); 6121 } 6122 6123 // Now repeat for perm gen 6124 if (should_unload_classes()) { 6125 CMSTokenSyncWithLocks ts(true, _permGen->freelistLock(), 6126 bitMapLock()); 6127 sweepWork(_permGen, asynch); 6128 } 6129 6130 // Update Universe::_heap_*_at_gc figures. 6131 // We need all the free list locks to make the abstract state 6132 // transition from Sweeping to Resetting. See detailed note 6133 // further below. 6134 { 6135 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(), 6136 _permGen->freelistLock()); 6137 // Update heap occupancy information which is used as 6138 // input to soft ref clearing policy at the next gc. 6139 Universe::update_heap_info_at_gc(); 6140 _collectorState = Resizing; 6141 } 6142 } else { 6143 // already have needed locks 6144 sweepWork(_cmsGen, asynch); 6145 6146 if (should_unload_classes()) { 6147 sweepWork(_permGen, asynch); 6148 } 6149 // Update heap occupancy information which is used as 6150 // input to soft ref clearing policy at the next gc. 6151 Universe::update_heap_info_at_gc(); 6152 _collectorState = Resizing; 6153 } 6154 verify_work_stacks_empty(); 6155 verify_overflow_empty(); 6156 6157 _intra_sweep_timer.stop(); 6158 _intra_sweep_estimate.sample(_intra_sweep_timer.seconds()); 6159 6160 _inter_sweep_timer.reset(); 6161 _inter_sweep_timer.start(); 6162 6163 // We need to use a monotonically non-deccreasing time in ms 6164 // or we will see time-warp warnings and os::javaTimeMillis() 6165 // does not guarantee monotonicity. 6166 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC; 6167 update_time_of_last_gc(now); 6168 6169 // NOTE on abstract state transitions: 6170 // Mutators allocate-live and/or mark the mod-union table dirty 6171 // based on the state of the collection. The former is done in 6172 // the interval [Marking, Sweeping] and the latter in the interval 6173 // [Marking, Sweeping). Thus the transitions into the Marking state 6174 // and out of the Sweeping state must be synchronously visible 6175 // globally to the mutators. 6176 // The transition into the Marking state happens with the world 6177 // stopped so the mutators will globally see it. Sweeping is 6178 // done asynchronously by the background collector so the transition 6179 // from the Sweeping state to the Resizing state must be done 6180 // under the freelistLock (as is the check for whether to 6181 // allocate-live and whether to dirty the mod-union table). 6182 assert(_collectorState == Resizing, "Change of collector state to" 6183 " Resizing must be done under the freelistLocks (plural)"); 6184 6185 // Now that sweeping has been completed, we clear 6186 // the incremental_collection_failed flag, 6187 // thus inviting a younger gen collection to promote into 6188 // this generation. If such a promotion may still fail, 6189 // the flag will be set again when a young collection is 6190 // attempted. 6191 GenCollectedHeap* gch = GenCollectedHeap::heap(); 6192 gch->clear_incremental_collection_failed(); // Worth retrying as fresh space may have been freed up 6193 gch->update_full_collections_completed(_collection_count_start); 6194 } 6195 6196 // FIX ME!!! Looks like this belongs in CFLSpace, with 6197 // CMSGen merely delegating to it. 6198 void ConcurrentMarkSweepGeneration::setNearLargestChunk() { 6199 double nearLargestPercent = FLSLargestBlockCoalesceProximity; 6200 HeapWord* minAddr = _cmsSpace->bottom(); 6201 HeapWord* largestAddr = 6202 (HeapWord*) _cmsSpace->dictionary()->find_largest_dict(); 6203 if (largestAddr == NULL) { 6204 // The dictionary appears to be empty. In this case 6205 // try to coalesce at the end of the heap. 6206 largestAddr = _cmsSpace->end(); 6207 } 6208 size_t largestOffset = pointer_delta(largestAddr, minAddr); 6209 size_t nearLargestOffset = 6210 (size_t)((double)largestOffset * nearLargestPercent) - MinChunkSize; 6211 if (PrintFLSStatistics != 0) { 6212 gclog_or_tty->print_cr( 6213 "CMS: Large Block: " PTR_FORMAT ";" 6214 " Proximity: " PTR_FORMAT " -> " PTR_FORMAT, 6215 largestAddr, 6216 _cmsSpace->nearLargestChunk(), minAddr + nearLargestOffset); 6217 } 6218 _cmsSpace->set_nearLargestChunk(minAddr + nearLargestOffset); 6219 } 6220 6221 bool ConcurrentMarkSweepGeneration::isNearLargestChunk(HeapWord* addr) { 6222 return addr >= _cmsSpace->nearLargestChunk(); 6223 } 6224 6225 FreeChunk* ConcurrentMarkSweepGeneration::find_chunk_at_end() { 6226 return _cmsSpace->find_chunk_at_end(); 6227 } 6228 6229 void ConcurrentMarkSweepGeneration::update_gc_stats(int current_level, 6230 bool full) { 6231 // The next lower level has been collected. Gather any statistics 6232 // that are of interest at this point. 6233 if (!full && (current_level + 1) == level()) { 6234 // Gather statistics on the young generation collection. 6235 collector()->stats().record_gc0_end(used()); 6236 } 6237 } 6238 6239 CMSAdaptiveSizePolicy* ConcurrentMarkSweepGeneration::size_policy() { 6240 GenCollectedHeap* gch = GenCollectedHeap::heap(); 6241 assert(gch->kind() == CollectedHeap::GenCollectedHeap, 6242 "Wrong type of heap"); 6243 CMSAdaptiveSizePolicy* sp = (CMSAdaptiveSizePolicy*) 6244 gch->gen_policy()->size_policy(); 6245 assert(sp->is_gc_cms_adaptive_size_policy(), 6246 "Wrong type of size policy"); 6247 return sp; 6248 } 6249 6250 void ConcurrentMarkSweepGeneration::rotate_debug_collection_type() { 6251 if (PrintGCDetails && Verbose) { 6252 gclog_or_tty->print("Rotate from %d ", _debug_collection_type); 6253 } 6254 _debug_collection_type = (CollectionTypes) (_debug_collection_type + 1); 6255 _debug_collection_type = 6256 (CollectionTypes) (_debug_collection_type % Unknown_collection_type); 6257 if (PrintGCDetails && Verbose) { 6258 gclog_or_tty->print_cr("to %d ", _debug_collection_type); 6259 } 6260 } 6261 6262 void CMSCollector::sweepWork(ConcurrentMarkSweepGeneration* gen, 6263 bool asynch) { 6264 // We iterate over the space(s) underlying this generation, 6265 // checking the mark bit map to see if the bits corresponding 6266 // to specific blocks are marked or not. Blocks that are 6267 // marked are live and are not swept up. All remaining blocks 6268 // are swept up, with coalescing on-the-fly as we sweep up 6269 // contiguous free and/or garbage blocks: 6270 // We need to ensure that the sweeper synchronizes with allocators 6271 // and stop-the-world collectors. In particular, the following 6272 // locks are used: 6273 // . CMS token: if this is held, a stop the world collection cannot occur 6274 // . freelistLock: if this is held no allocation can occur from this 6275 // generation by another thread 6276 // . bitMapLock: if this is held, no other thread can access or update 6277 // 6278 6279 // Note that we need to hold the freelistLock if we use 6280 // block iterate below; else the iterator might go awry if 6281 // a mutator (or promotion) causes block contents to change 6282 // (for instance if the allocator divvies up a block). 6283 // If we hold the free list lock, for all practical purposes 6284 // young generation GC's can't occur (they'll usually need to 6285 // promote), so we might as well prevent all young generation 6286 // GC's while we do a sweeping step. For the same reason, we might 6287 // as well take the bit map lock for the entire duration 6288 6289 // check that we hold the requisite locks 6290 assert(have_cms_token(), "Should hold cms token"); 6291 assert( (asynch && ConcurrentMarkSweepThread::cms_thread_has_cms_token()) 6292 || (!asynch && ConcurrentMarkSweepThread::vm_thread_has_cms_token()), 6293 "Should possess CMS token to sweep"); 6294 assert_lock_strong(gen->freelistLock()); 6295 assert_lock_strong(bitMapLock()); 6296 6297 assert(!_inter_sweep_timer.is_active(), "Was switched off in an outer context"); 6298 assert(_intra_sweep_timer.is_active(), "Was switched on in an outer context"); 6299 gen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()), 6300 _inter_sweep_estimate.padded_average(), 6301 _intra_sweep_estimate.padded_average()); 6302 gen->setNearLargestChunk(); 6303 6304 { 6305 SweepClosure sweepClosure(this, gen, &_markBitMap, 6306 CMSYield && asynch); 6307 gen->cmsSpace()->blk_iterate_careful(&sweepClosure); 6308 // We need to free-up/coalesce garbage/blocks from a 6309 // co-terminal free run. This is done in the SweepClosure 6310 // destructor; so, do not remove this scope, else the 6311 // end-of-sweep-census below will be off by a little bit. 6312 } 6313 gen->cmsSpace()->sweep_completed(); 6314 gen->cmsSpace()->endSweepFLCensus(sweep_count()); 6315 if (should_unload_classes()) { // unloaded classes this cycle, 6316 _concurrent_cycles_since_last_unload = 0; // ... reset count 6317 } else { // did not unload classes, 6318 _concurrent_cycles_since_last_unload++; // ... increment count 6319 } 6320 } 6321 6322 // Reset CMS data structures (for now just the marking bit map) 6323 // preparatory for the next cycle. 6324 void CMSCollector::reset(bool asynch) { 6325 GenCollectedHeap* gch = GenCollectedHeap::heap(); 6326 CMSAdaptiveSizePolicy* sp = size_policy(); 6327 AdaptiveSizePolicyOutput(sp, gch->total_collections()); 6328 if (asynch) { 6329 CMSTokenSyncWithLocks ts(true, bitMapLock()); 6330 6331 // If the state is not "Resetting", the foreground thread 6332 // has done a collection and the resetting. 6333 if (_collectorState != Resetting) { 6334 assert(_collectorState == Idling, "The state should only change" 6335 " because the foreground collector has finished the collection"); 6336 return; 6337 } 6338 6339 // Clear the mark bitmap (no grey objects to start with) 6340 // for the next cycle. 6341 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 6342 CMSPhaseAccounting cmspa(this, "reset", !PrintGCDetails); 6343 6344 HeapWord* curAddr = _markBitMap.startWord(); 6345 while (curAddr < _markBitMap.endWord()) { 6346 size_t remaining = pointer_delta(_markBitMap.endWord(), curAddr); 6347 MemRegion chunk(curAddr, MIN2(CMSBitMapYieldQuantum, remaining)); 6348 _markBitMap.clear_large_range(chunk); 6349 if (ConcurrentMarkSweepThread::should_yield() && 6350 !foregroundGCIsActive() && 6351 CMSYield) { 6352 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6353 "CMS thread should hold CMS token"); 6354 assert_lock_strong(bitMapLock()); 6355 bitMapLock()->unlock(); 6356 ConcurrentMarkSweepThread::desynchronize(true); 6357 ConcurrentMarkSweepThread::acknowledge_yield_request(); 6358 stopTimer(); 6359 if (PrintCMSStatistics != 0) { 6360 incrementYields(); 6361 } 6362 icms_wait(); 6363 6364 // See the comment in coordinator_yield() 6365 for (unsigned i = 0; i < CMSYieldSleepCount && 6366 ConcurrentMarkSweepThread::should_yield() && 6367 !CMSCollector::foregroundGCIsActive(); ++i) { 6368 os::sleep(Thread::current(), 1, false); 6369 ConcurrentMarkSweepThread::acknowledge_yield_request(); 6370 } 6371 6372 ConcurrentMarkSweepThread::synchronize(true); 6373 bitMapLock()->lock_without_safepoint_check(); 6374 startTimer(); 6375 } 6376 curAddr = chunk.end(); 6377 } 6378 // A successful mostly concurrent collection has been done. 6379 // Because only the full (i.e., concurrent mode failure) collections 6380 // are being measured for gc overhead limits, clean the "near" flag 6381 // and count. 6382 sp->reset_gc_overhead_limit_count(); 6383 _collectorState = Idling; 6384 } else { 6385 // already have the lock 6386 assert(_collectorState == Resetting, "just checking"); 6387 assert_lock_strong(bitMapLock()); 6388 _markBitMap.clear_all(); 6389 _collectorState = Idling; 6390 } 6391 6392 // Stop incremental mode after a cycle completes, so that any future cycles 6393 // are triggered by allocation. 6394 stop_icms(); 6395 6396 NOT_PRODUCT( 6397 if (RotateCMSCollectionTypes) { 6398 _cmsGen->rotate_debug_collection_type(); 6399 } 6400 ) 6401 6402 register_gc_end(); 6403 } 6404 6405 void CMSCollector::do_CMS_operation(CMS_op_type op, GCCause::Cause gc_cause) { 6406 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 6407 GCTraceTime t(GCCauseString("GC", gc_cause), PrintGC, !PrintGCDetails, NULL); 6408 TraceCollectorStats tcs(counters()); 6409 6410 switch (op) { 6411 case CMS_op_checkpointRootsInitial: { 6412 SvcGCMarker sgcm(SvcGCMarker::OTHER); 6413 checkpointRootsInitial(true); // asynch 6414 if (PrintGC) { 6415 _cmsGen->printOccupancy("initial-mark"); 6416 } 6417 break; 6418 } 6419 case CMS_op_checkpointRootsFinal: { 6420 SvcGCMarker sgcm(SvcGCMarker::OTHER); 6421 checkpointRootsFinal(true, // asynch 6422 false, // !clear_all_soft_refs 6423 false); // !init_mark_was_synchronous 6424 if (PrintGC) { 6425 _cmsGen->printOccupancy("remark"); 6426 } 6427 break; 6428 } 6429 default: 6430 fatal("No such CMS_op"); 6431 } 6432 } 6433 6434 #ifndef PRODUCT 6435 size_t const CMSCollector::skip_header_HeapWords() { 6436 return FreeChunk::header_size(); 6437 } 6438 6439 // Try and collect here conditions that should hold when 6440 // CMS thread is exiting. The idea is that the foreground GC 6441 // thread should not be blocked if it wants to terminate 6442 // the CMS thread and yet continue to run the VM for a while 6443 // after that. 6444 void CMSCollector::verify_ok_to_terminate() const { 6445 assert(Thread::current()->is_ConcurrentGC_thread(), 6446 "should be called by CMS thread"); 6447 assert(!_foregroundGCShouldWait, "should be false"); 6448 // We could check here that all the various low-level locks 6449 // are not held by the CMS thread, but that is overkill; see 6450 // also CMSThread::verify_ok_to_terminate() where the CGC_lock 6451 // is checked. 6452 } 6453 #endif 6454 6455 size_t CMSCollector::block_size_using_printezis_bits(HeapWord* addr) const { 6456 assert(_markBitMap.isMarked(addr) && _markBitMap.isMarked(addr + 1), 6457 "missing Printezis mark?"); 6458 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 6459 size_t size = pointer_delta(nextOneAddr + 1, addr); 6460 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 6461 "alignment problem"); 6462 assert(size >= 3, "Necessary for Printezis marks to work"); 6463 return size; 6464 } 6465 6466 // A variant of the above (block_size_using_printezis_bits()) except 6467 // that we return 0 if the P-bits are not yet set. 6468 size_t CMSCollector::block_size_if_printezis_bits(HeapWord* addr) const { 6469 if (_markBitMap.isMarked(addr + 1)) { 6470 assert(_markBitMap.isMarked(addr), "P-bit can be set only for marked objects"); 6471 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 6472 size_t size = pointer_delta(nextOneAddr + 1, addr); 6473 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 6474 "alignment problem"); 6475 assert(size >= 3, "Necessary for Printezis marks to work"); 6476 return size; 6477 } 6478 return 0; 6479 } 6480 6481 HeapWord* CMSCollector::next_card_start_after_block(HeapWord* addr) const { 6482 size_t sz = 0; 6483 oop p = (oop)addr; 6484 if (p->klass_or_null() != NULL && p->is_parsable()) { 6485 sz = CompactibleFreeListSpace::adjustObjectSize(p->size()); 6486 } else { 6487 sz = block_size_using_printezis_bits(addr); 6488 } 6489 assert(sz > 0, "size must be nonzero"); 6490 HeapWord* next_block = addr + sz; 6491 HeapWord* next_card = (HeapWord*)round_to((uintptr_t)next_block, 6492 CardTableModRefBS::card_size); 6493 assert(round_down((uintptr_t)addr, CardTableModRefBS::card_size) < 6494 round_down((uintptr_t)next_card, CardTableModRefBS::card_size), 6495 "must be different cards"); 6496 return next_card; 6497 } 6498 6499 6500 // CMS Bit Map Wrapper ///////////////////////////////////////// 6501 6502 // Construct a CMS bit map infrastructure, but don't create the 6503 // bit vector itself. That is done by a separate call CMSBitMap::allocate() 6504 // further below. 6505 CMSBitMap::CMSBitMap(int shifter, int mutex_rank, const char* mutex_name): 6506 _bm(), 6507 _shifter(shifter), 6508 _lock(mutex_rank >= 0 ? new Mutex(mutex_rank, mutex_name, true) : NULL) 6509 { 6510 _bmStartWord = 0; 6511 _bmWordSize = 0; 6512 } 6513 6514 bool CMSBitMap::allocate(MemRegion mr) { 6515 _bmStartWord = mr.start(); 6516 _bmWordSize = mr.word_size(); 6517 ReservedSpace brs(ReservedSpace::allocation_align_size_up( 6518 (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1)); 6519 if (!brs.is_reserved()) { 6520 warning("CMS bit map allocation failure"); 6521 return false; 6522 } 6523 // For now we'll just commit all of the bit map up fromt. 6524 // Later on we'll try to be more parsimonious with swap. 6525 if (!_virtual_space.initialize(brs, brs.size())) { 6526 warning("CMS bit map backing store failure"); 6527 return false; 6528 } 6529 assert(_virtual_space.committed_size() == brs.size(), 6530 "didn't reserve backing store for all of CMS bit map?"); 6531 _bm.set_map((BitMap::bm_word_t*)_virtual_space.low()); 6532 assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >= 6533 _bmWordSize, "inconsistency in bit map sizing"); 6534 _bm.set_size(_bmWordSize >> _shifter); 6535 6536 // bm.clear(); // can we rely on getting zero'd memory? verify below 6537 assert(isAllClear(), 6538 "Expected zero'd memory from ReservedSpace constructor"); 6539 assert(_bm.size() == heapWordDiffToOffsetDiff(sizeInWords()), 6540 "consistency check"); 6541 return true; 6542 } 6543 6544 void CMSBitMap::dirty_range_iterate_clear(MemRegion mr, MemRegionClosure* cl) { 6545 HeapWord *next_addr, *end_addr, *last_addr; 6546 assert_locked(); 6547 assert(covers(mr), "out-of-range error"); 6548 // XXX assert that start and end are appropriately aligned 6549 for (next_addr = mr.start(), end_addr = mr.end(); 6550 next_addr < end_addr; next_addr = last_addr) { 6551 MemRegion dirty_region = getAndClearMarkedRegion(next_addr, end_addr); 6552 last_addr = dirty_region.end(); 6553 if (!dirty_region.is_empty()) { 6554 cl->do_MemRegion(dirty_region); 6555 } else { 6556 assert(last_addr == end_addr, "program logic"); 6557 return; 6558 } 6559 } 6560 } 6561 6562 #ifndef PRODUCT 6563 void CMSBitMap::assert_locked() const { 6564 CMSLockVerifier::assert_locked(lock()); 6565 } 6566 6567 bool CMSBitMap::covers(MemRegion mr) const { 6568 // assert(_bm.map() == _virtual_space.low(), "map inconsistency"); 6569 assert((size_t)_bm.size() == (_bmWordSize >> _shifter), 6570 "size inconsistency"); 6571 return (mr.start() >= _bmStartWord) && 6572 (mr.end() <= endWord()); 6573 } 6574 6575 bool CMSBitMap::covers(HeapWord* start, size_t size) const { 6576 return (start >= _bmStartWord && (start + size) <= endWord()); 6577 } 6578 6579 void CMSBitMap::verifyNoOneBitsInRange(HeapWord* left, HeapWord* right) { 6580 // verify that there are no 1 bits in the interval [left, right) 6581 FalseBitMapClosure falseBitMapClosure; 6582 iterate(&falseBitMapClosure, left, right); 6583 } 6584 6585 void CMSBitMap::region_invariant(MemRegion mr) 6586 { 6587 assert_locked(); 6588 // mr = mr.intersection(MemRegion(_bmStartWord, _bmWordSize)); 6589 assert(!mr.is_empty(), "unexpected empty region"); 6590 assert(covers(mr), "mr should be covered by bit map"); 6591 // convert address range into offset range 6592 size_t start_ofs = heapWordToOffset(mr.start()); 6593 // Make sure that end() is appropriately aligned 6594 assert(mr.end() == (HeapWord*)round_to((intptr_t)mr.end(), 6595 (1 << (_shifter+LogHeapWordSize))), 6596 "Misaligned mr.end()"); 6597 size_t end_ofs = heapWordToOffset(mr.end()); 6598 assert(end_ofs > start_ofs, "Should mark at least one bit"); 6599 } 6600 6601 #endif 6602 6603 bool CMSMarkStack::allocate(size_t size) { 6604 // allocate a stack of the requisite depth 6605 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 6606 size * sizeof(oop))); 6607 if (!rs.is_reserved()) { 6608 warning("CMSMarkStack allocation failure"); 6609 return false; 6610 } 6611 if (!_virtual_space.initialize(rs, rs.size())) { 6612 warning("CMSMarkStack backing store failure"); 6613 return false; 6614 } 6615 assert(_virtual_space.committed_size() == rs.size(), 6616 "didn't reserve backing store for all of CMS stack?"); 6617 _base = (oop*)(_virtual_space.low()); 6618 _index = 0; 6619 _capacity = size; 6620 NOT_PRODUCT(_max_depth = 0); 6621 return true; 6622 } 6623 6624 // XXX FIX ME !!! In the MT case we come in here holding a 6625 // leaf lock. For printing we need to take a further lock 6626 // which has lower rank. We need to recallibrate the two 6627 // lock-ranks involved in order to be able to rpint the 6628 // messages below. (Or defer the printing to the caller. 6629 // For now we take the expedient path of just disabling the 6630 // messages for the problematic case.) 6631 void CMSMarkStack::expand() { 6632 assert(_capacity <= MarkStackSizeMax, "stack bigger than permitted"); 6633 if (_capacity == MarkStackSizeMax) { 6634 if (_hit_limit++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) { 6635 // We print a warning message only once per CMS cycle. 6636 gclog_or_tty->print_cr(" (benign) Hit CMSMarkStack max size limit"); 6637 } 6638 return; 6639 } 6640 // Double capacity if possible 6641 size_t new_capacity = MIN2(_capacity*2, MarkStackSizeMax); 6642 // Do not give up existing stack until we have managed to 6643 // get the double capacity that we desired. 6644 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 6645 new_capacity * sizeof(oop))); 6646 if (rs.is_reserved()) { 6647 // Release the backing store associated with old stack 6648 _virtual_space.release(); 6649 // Reinitialize virtual space for new stack 6650 if (!_virtual_space.initialize(rs, rs.size())) { 6651 fatal("Not enough swap for expanded marking stack"); 6652 } 6653 _base = (oop*)(_virtual_space.low()); 6654 _index = 0; 6655 _capacity = new_capacity; 6656 } else if (_failed_double++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) { 6657 // Failed to double capacity, continue; 6658 // we print a detail message only once per CMS cycle. 6659 gclog_or_tty->print(" (benign) Failed to expand marking stack from "SIZE_FORMAT"K to " 6660 SIZE_FORMAT"K", 6661 _capacity / K, new_capacity / K); 6662 } 6663 } 6664 6665 6666 // Closures 6667 // XXX: there seems to be a lot of code duplication here; 6668 // should refactor and consolidate common code. 6669 6670 // This closure is used to mark refs into the CMS generation in 6671 // the CMS bit map. Called at the first checkpoint. This closure 6672 // assumes that we do not need to re-mark dirty cards; if the CMS 6673 // generation on which this is used is not an oldest (modulo perm gen) 6674 // generation then this will lose younger_gen cards! 6675 6676 MarkRefsIntoClosure::MarkRefsIntoClosure( 6677 MemRegion span, CMSBitMap* bitMap): 6678 _span(span), 6679 _bitMap(bitMap) 6680 { 6681 assert(_ref_processor == NULL, "deliberately left NULL"); 6682 assert(_bitMap->covers(_span), "_bitMap/_span mismatch"); 6683 } 6684 6685 void MarkRefsIntoClosure::do_oop(oop obj) { 6686 // if p points into _span, then mark corresponding bit in _markBitMap 6687 assert(obj->is_oop(), "expected an oop"); 6688 HeapWord* addr = (HeapWord*)obj; 6689 if (_span.contains(addr)) { 6690 // this should be made more efficient 6691 _bitMap->mark(addr); 6692 } 6693 } 6694 6695 void MarkRefsIntoClosure::do_oop(oop* p) { MarkRefsIntoClosure::do_oop_work(p); } 6696 void MarkRefsIntoClosure::do_oop(narrowOop* p) { MarkRefsIntoClosure::do_oop_work(p); } 6697 6698 // A variant of the above, used for CMS marking verification. 6699 MarkRefsIntoVerifyClosure::MarkRefsIntoVerifyClosure( 6700 MemRegion span, CMSBitMap* verification_bm, CMSBitMap* cms_bm): 6701 _span(span), 6702 _verification_bm(verification_bm), 6703 _cms_bm(cms_bm) 6704 { 6705 assert(_ref_processor == NULL, "deliberately left NULL"); 6706 assert(_verification_bm->covers(_span), "_verification_bm/_span mismatch"); 6707 } 6708 6709 void MarkRefsIntoVerifyClosure::do_oop(oop obj) { 6710 // if p points into _span, then mark corresponding bit in _markBitMap 6711 assert(obj->is_oop(), "expected an oop"); 6712 HeapWord* addr = (HeapWord*)obj; 6713 if (_span.contains(addr)) { 6714 _verification_bm->mark(addr); 6715 if (!_cms_bm->isMarked(addr)) { 6716 oop(addr)->print(); 6717 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)", addr); 6718 fatal("... aborting"); 6719 } 6720 } 6721 } 6722 6723 void MarkRefsIntoVerifyClosure::do_oop(oop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); } 6724 void MarkRefsIntoVerifyClosure::do_oop(narrowOop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); } 6725 6726 ////////////////////////////////////////////////// 6727 // MarkRefsIntoAndScanClosure 6728 ////////////////////////////////////////////////// 6729 6730 MarkRefsIntoAndScanClosure::MarkRefsIntoAndScanClosure(MemRegion span, 6731 ReferenceProcessor* rp, 6732 CMSBitMap* bit_map, 6733 CMSBitMap* mod_union_table, 6734 CMSMarkStack* mark_stack, 6735 CMSMarkStack* revisit_stack, 6736 CMSCollector* collector, 6737 bool should_yield, 6738 bool concurrent_precleaning): 6739 _collector(collector), 6740 _span(span), 6741 _bit_map(bit_map), 6742 _mark_stack(mark_stack), 6743 _pushAndMarkClosure(collector, span, rp, bit_map, mod_union_table, 6744 mark_stack, revisit_stack, concurrent_precleaning), 6745 _yield(should_yield), 6746 _concurrent_precleaning(concurrent_precleaning), 6747 _freelistLock(NULL) 6748 { 6749 _ref_processor = rp; 6750 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 6751 } 6752 6753 // This closure is used to mark refs into the CMS generation at the 6754 // second (final) checkpoint, and to scan and transitively follow 6755 // the unmarked oops. It is also used during the concurrent precleaning 6756 // phase while scanning objects on dirty cards in the CMS generation. 6757 // The marks are made in the marking bit map and the marking stack is 6758 // used for keeping the (newly) grey objects during the scan. 6759 // The parallel version (Par_...) appears further below. 6760 void MarkRefsIntoAndScanClosure::do_oop(oop obj) { 6761 if (obj != NULL) { 6762 assert(obj->is_oop(), "expected an oop"); 6763 HeapWord* addr = (HeapWord*)obj; 6764 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 6765 assert(_collector->overflow_list_is_empty(), 6766 "overflow list should be empty"); 6767 if (_span.contains(addr) && 6768 !_bit_map->isMarked(addr)) { 6769 // mark bit map (object is now grey) 6770 _bit_map->mark(addr); 6771 // push on marking stack (stack should be empty), and drain the 6772 // stack by applying this closure to the oops in the oops popped 6773 // from the stack (i.e. blacken the grey objects) 6774 bool res = _mark_stack->push(obj); 6775 assert(res, "Should have space to push on empty stack"); 6776 do { 6777 oop new_oop = _mark_stack->pop(); 6778 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 6779 assert(new_oop->is_parsable(), "Found unparsable oop"); 6780 assert(_bit_map->isMarked((HeapWord*)new_oop), 6781 "only grey objects on this stack"); 6782 // iterate over the oops in this oop, marking and pushing 6783 // the ones in CMS heap (i.e. in _span). 6784 new_oop->oop_iterate(&_pushAndMarkClosure); 6785 // check if it's time to yield 6786 do_yield_check(); 6787 } while (!_mark_stack->isEmpty() || 6788 (!_concurrent_precleaning && take_from_overflow_list())); 6789 // if marking stack is empty, and we are not doing this 6790 // during precleaning, then check the overflow list 6791 } 6792 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 6793 assert(_collector->overflow_list_is_empty(), 6794 "overflow list was drained above"); 6795 // We could restore evacuated mark words, if any, used for 6796 // overflow list links here because the overflow list is 6797 // provably empty here. That would reduce the maximum 6798 // size requirements for preserved_{oop,mark}_stack. 6799 // But we'll just postpone it until we are all done 6800 // so we can just stream through. 6801 if (!_concurrent_precleaning && CMSOverflowEarlyRestoration) { 6802 _collector->restore_preserved_marks_if_any(); 6803 assert(_collector->no_preserved_marks(), "No preserved marks"); 6804 } 6805 assert(!CMSOverflowEarlyRestoration || _collector->no_preserved_marks(), 6806 "All preserved marks should have been restored above"); 6807 } 6808 } 6809 6810 void MarkRefsIntoAndScanClosure::do_oop(oop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); } 6811 void MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); } 6812 6813 void MarkRefsIntoAndScanClosure::do_yield_work() { 6814 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6815 "CMS thread should hold CMS token"); 6816 assert_lock_strong(_freelistLock); 6817 assert_lock_strong(_bit_map->lock()); 6818 // relinquish the free_list_lock and bitMaplock() 6819 DEBUG_ONLY(RememberKlassesChecker mux(false);) 6820 _bit_map->lock()->unlock(); 6821 _freelistLock->unlock(); 6822 ConcurrentMarkSweepThread::desynchronize(true); 6823 ConcurrentMarkSweepThread::acknowledge_yield_request(); 6824 _collector->stopTimer(); 6825 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 6826 if (PrintCMSStatistics != 0) { 6827 _collector->incrementYields(); 6828 } 6829 _collector->icms_wait(); 6830 6831 // See the comment in coordinator_yield() 6832 for (unsigned i = 0; 6833 i < CMSYieldSleepCount && 6834 ConcurrentMarkSweepThread::should_yield() && 6835 !CMSCollector::foregroundGCIsActive(); 6836 ++i) { 6837 os::sleep(Thread::current(), 1, false); 6838 ConcurrentMarkSweepThread::acknowledge_yield_request(); 6839 } 6840 6841 ConcurrentMarkSweepThread::synchronize(true); 6842 _freelistLock->lock_without_safepoint_check(); 6843 _bit_map->lock()->lock_without_safepoint_check(); 6844 _collector->startTimer(); 6845 } 6846 6847 /////////////////////////////////////////////////////////// 6848 // Par_MarkRefsIntoAndScanClosure: a parallel version of 6849 // MarkRefsIntoAndScanClosure 6850 /////////////////////////////////////////////////////////// 6851 Par_MarkRefsIntoAndScanClosure::Par_MarkRefsIntoAndScanClosure( 6852 CMSCollector* collector, MemRegion span, ReferenceProcessor* rp, 6853 CMSBitMap* bit_map, OopTaskQueue* work_queue, CMSMarkStack* revisit_stack): 6854 _span(span), 6855 _bit_map(bit_map), 6856 _work_queue(work_queue), 6857 _low_water_mark(MIN2((uint)(work_queue->max_elems()/4), 6858 (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads))), 6859 _par_pushAndMarkClosure(collector, span, rp, bit_map, work_queue, 6860 revisit_stack) 6861 { 6862 _ref_processor = rp; 6863 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 6864 } 6865 6866 // This closure is used to mark refs into the CMS generation at the 6867 // second (final) checkpoint, and to scan and transitively follow 6868 // the unmarked oops. The marks are made in the marking bit map and 6869 // the work_queue is used for keeping the (newly) grey objects during 6870 // the scan phase whence they are also available for stealing by parallel 6871 // threads. Since the marking bit map is shared, updates are 6872 // synchronized (via CAS). 6873 void Par_MarkRefsIntoAndScanClosure::do_oop(oop obj) { 6874 if (obj != NULL) { 6875 // Ignore mark word because this could be an already marked oop 6876 // that may be chained at the end of the overflow list. 6877 assert(obj->is_oop(true), "expected an oop"); 6878 HeapWord* addr = (HeapWord*)obj; 6879 if (_span.contains(addr) && 6880 !_bit_map->isMarked(addr)) { 6881 // mark bit map (object will become grey): 6882 // It is possible for several threads to be 6883 // trying to "claim" this object concurrently; 6884 // the unique thread that succeeds in marking the 6885 // object first will do the subsequent push on 6886 // to the work queue (or overflow list). 6887 if (_bit_map->par_mark(addr)) { 6888 // push on work_queue (which may not be empty), and trim the 6889 // queue to an appropriate length by applying this closure to 6890 // the oops in the oops popped from the stack (i.e. blacken the 6891 // grey objects) 6892 bool res = _work_queue->push(obj); 6893 assert(res, "Low water mark should be less than capacity?"); 6894 trim_queue(_low_water_mark); 6895 } // Else, another thread claimed the object 6896 } 6897 } 6898 } 6899 6900 void Par_MarkRefsIntoAndScanClosure::do_oop(oop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); } 6901 void Par_MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); } 6902 6903 // This closure is used to rescan the marked objects on the dirty cards 6904 // in the mod union table and the card table proper. 6905 size_t ScanMarkedObjectsAgainCarefullyClosure::do_object_careful_m( 6906 oop p, MemRegion mr) { 6907 6908 size_t size = 0; 6909 HeapWord* addr = (HeapWord*)p; 6910 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6911 assert(_span.contains(addr), "we are scanning the CMS generation"); 6912 // check if it's time to yield 6913 if (do_yield_check()) { 6914 // We yielded for some foreground stop-world work, 6915 // and we have been asked to abort this ongoing preclean cycle. 6916 return 0; 6917 } 6918 if (_bitMap->isMarked(addr)) { 6919 // it's marked; is it potentially uninitialized? 6920 if (p->klass_or_null() != NULL) { 6921 // If is_conc_safe is false, the object may be undergoing 6922 // change by the VM outside a safepoint. Don't try to 6923 // scan it, but rather leave it for the remark phase. 6924 if (CMSPermGenPrecleaningEnabled && 6925 (!p->is_conc_safe() || !p->is_parsable())) { 6926 // Signal precleaning to redirty the card since 6927 // the klass pointer is already installed. 6928 assert(size == 0, "Initial value"); 6929 } else { 6930 assert(p->is_parsable(), "must be parsable."); 6931 // an initialized object; ignore mark word in verification below 6932 // since we are running concurrent with mutators 6933 assert(p->is_oop(true), "should be an oop"); 6934 if (p->is_objArray()) { 6935 // objArrays are precisely marked; restrict scanning 6936 // to dirty cards only. 6937 size = CompactibleFreeListSpace::adjustObjectSize( 6938 p->oop_iterate(_scanningClosure, mr)); 6939 } else { 6940 // A non-array may have been imprecisely marked; we need 6941 // to scan object in its entirety. 6942 size = CompactibleFreeListSpace::adjustObjectSize( 6943 p->oop_iterate(_scanningClosure)); 6944 } 6945 #ifdef DEBUG 6946 size_t direct_size = 6947 CompactibleFreeListSpace::adjustObjectSize(p->size()); 6948 assert(size == direct_size, "Inconsistency in size"); 6949 assert(size >= 3, "Necessary for Printezis marks to work"); 6950 if (!_bitMap->isMarked(addr+1)) { 6951 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size); 6952 } else { 6953 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size-1); 6954 assert(_bitMap->isMarked(addr+size-1), 6955 "inconsistent Printezis mark"); 6956 } 6957 #endif // DEBUG 6958 } 6959 } else { 6960 // an unitialized object 6961 assert(_bitMap->isMarked(addr+1), "missing Printezis mark?"); 6962 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 6963 size = pointer_delta(nextOneAddr + 1, addr); 6964 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 6965 "alignment problem"); 6966 // Note that pre-cleaning needn't redirty the card. OopDesc::set_klass() 6967 // will dirty the card when the klass pointer is installed in the 6968 // object (signalling the completion of initialization). 6969 } 6970 } else { 6971 // Either a not yet marked object or an uninitialized object 6972 if (p->klass_or_null() == NULL || !p->is_parsable()) { 6973 // An uninitialized object, skip to the next card, since 6974 // we may not be able to read its P-bits yet. 6975 assert(size == 0, "Initial value"); 6976 } else { 6977 // An object not (yet) reached by marking: we merely need to 6978 // compute its size so as to go look at the next block. 6979 assert(p->is_oop(true), "should be an oop"); 6980 size = CompactibleFreeListSpace::adjustObjectSize(p->size()); 6981 } 6982 } 6983 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6984 return size; 6985 } 6986 6987 void ScanMarkedObjectsAgainCarefullyClosure::do_yield_work() { 6988 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6989 "CMS thread should hold CMS token"); 6990 assert_lock_strong(_freelistLock); 6991 assert_lock_strong(_bitMap->lock()); 6992 DEBUG_ONLY(RememberKlassesChecker mux(false);) 6993 // relinquish the free_list_lock and bitMaplock() 6994 _bitMap->lock()->unlock(); 6995 _freelistLock->unlock(); 6996 ConcurrentMarkSweepThread::desynchronize(true); 6997 ConcurrentMarkSweepThread::acknowledge_yield_request(); 6998 _collector->stopTimer(); 6999 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 7000 if (PrintCMSStatistics != 0) { 7001 _collector->incrementYields(); 7002 } 7003 _collector->icms_wait(); 7004 7005 // See the comment in coordinator_yield() 7006 for (unsigned i = 0; i < CMSYieldSleepCount && 7007 ConcurrentMarkSweepThread::should_yield() && 7008 !CMSCollector::foregroundGCIsActive(); ++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 _bitMap->lock()->lock_without_safepoint_check(); 7016 _collector->startTimer(); 7017 } 7018 7019 7020 ////////////////////////////////////////////////////////////////// 7021 // SurvivorSpacePrecleanClosure 7022 ////////////////////////////////////////////////////////////////// 7023 // This (single-threaded) closure is used to preclean the oops in 7024 // the survivor spaces. 7025 size_t SurvivorSpacePrecleanClosure::do_object_careful(oop p) { 7026 7027 HeapWord* addr = (HeapWord*)p; 7028 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 7029 assert(!_span.contains(addr), "we are scanning the survivor spaces"); 7030 assert(p->klass_or_null() != NULL, "object should be initializd"); 7031 assert(p->is_parsable(), "must be parsable."); 7032 // an initialized object; ignore mark word in verification below 7033 // since we are running concurrent with mutators 7034 assert(p->is_oop(true), "should be an oop"); 7035 // Note that we do not yield while we iterate over 7036 // the interior oops of p, pushing the relevant ones 7037 // on our marking stack. 7038 size_t size = p->oop_iterate(_scanning_closure); 7039 do_yield_check(); 7040 // Observe that below, we do not abandon the preclean 7041 // phase as soon as we should; rather we empty the 7042 // marking stack before returning. This is to satisfy 7043 // some existing assertions. In general, it may be a 7044 // good idea to abort immediately and complete the marking 7045 // from the grey objects at a later time. 7046 while (!_mark_stack->isEmpty()) { 7047 oop new_oop = _mark_stack->pop(); 7048 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 7049 assert(new_oop->is_parsable(), "Found unparsable oop"); 7050 assert(_bit_map->isMarked((HeapWord*)new_oop), 7051 "only grey objects on this stack"); 7052 // iterate over the oops in this oop, marking and pushing 7053 // the ones in CMS heap (i.e. in _span). 7054 new_oop->oop_iterate(_scanning_closure); 7055 // check if it's time to yield 7056 do_yield_check(); 7057 } 7058 unsigned int after_count = 7059 GenCollectedHeap::heap()->total_collections(); 7060 bool abort = (_before_count != after_count) || 7061 _collector->should_abort_preclean(); 7062 return abort ? 0 : size; 7063 } 7064 7065 void SurvivorSpacePrecleanClosure::do_yield_work() { 7066 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 7067 "CMS thread should hold CMS token"); 7068 assert_lock_strong(_bit_map->lock()); 7069 DEBUG_ONLY(RememberKlassesChecker smx(false);) 7070 // Relinquish the bit map lock 7071 _bit_map->lock()->unlock(); 7072 ConcurrentMarkSweepThread::desynchronize(true); 7073 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7074 _collector->stopTimer(); 7075 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 7076 if (PrintCMSStatistics != 0) { 7077 _collector->incrementYields(); 7078 } 7079 _collector->icms_wait(); 7080 7081 // See the comment in coordinator_yield() 7082 for (unsigned i = 0; i < CMSYieldSleepCount && 7083 ConcurrentMarkSweepThread::should_yield() && 7084 !CMSCollector::foregroundGCIsActive(); ++i) { 7085 os::sleep(Thread::current(), 1, false); 7086 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7087 } 7088 7089 ConcurrentMarkSweepThread::synchronize(true); 7090 _bit_map->lock()->lock_without_safepoint_check(); 7091 _collector->startTimer(); 7092 } 7093 7094 // This closure is used to rescan the marked objects on the dirty cards 7095 // in the mod union table and the card table proper. In the parallel 7096 // case, although the bitMap is shared, we do a single read so the 7097 // isMarked() query is "safe". 7098 bool ScanMarkedObjectsAgainClosure::do_object_bm(oop p, MemRegion mr) { 7099 // Ignore mark word because we are running concurrent with mutators 7100 assert(p->is_oop_or_null(true), "expected an oop or null"); 7101 HeapWord* addr = (HeapWord*)p; 7102 assert(_span.contains(addr), "we are scanning the CMS generation"); 7103 bool is_obj_array = false; 7104 #ifdef DEBUG 7105 if (!_parallel) { 7106 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 7107 assert(_collector->overflow_list_is_empty(), 7108 "overflow list should be empty"); 7109 7110 } 7111 #endif // DEBUG 7112 if (_bit_map->isMarked(addr)) { 7113 // Obj arrays are precisely marked, non-arrays are not; 7114 // so we scan objArrays precisely and non-arrays in their 7115 // entirety. 7116 if (p->is_objArray()) { 7117 is_obj_array = true; 7118 if (_parallel) { 7119 p->oop_iterate(_par_scan_closure, mr); 7120 } else { 7121 p->oop_iterate(_scan_closure, mr); 7122 } 7123 } else { 7124 if (_parallel) { 7125 p->oop_iterate(_par_scan_closure); 7126 } else { 7127 p->oop_iterate(_scan_closure); 7128 } 7129 } 7130 } 7131 #ifdef DEBUG 7132 if (!_parallel) { 7133 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 7134 assert(_collector->overflow_list_is_empty(), 7135 "overflow list should be empty"); 7136 7137 } 7138 #endif // DEBUG 7139 return is_obj_array; 7140 } 7141 7142 MarkFromRootsClosure::MarkFromRootsClosure(CMSCollector* collector, 7143 MemRegion span, 7144 CMSBitMap* bitMap, CMSMarkStack* markStack, 7145 CMSMarkStack* revisitStack, 7146 bool should_yield, bool verifying): 7147 _collector(collector), 7148 _span(span), 7149 _bitMap(bitMap), 7150 _mut(&collector->_modUnionTable), 7151 _markStack(markStack), 7152 _revisitStack(revisitStack), 7153 _yield(should_yield), 7154 _skipBits(0) 7155 { 7156 assert(_markStack->isEmpty(), "stack should be empty"); 7157 _finger = _bitMap->startWord(); 7158 _threshold = _finger; 7159 assert(_collector->_restart_addr == NULL, "Sanity check"); 7160 assert(_span.contains(_finger), "Out of bounds _finger?"); 7161 DEBUG_ONLY(_verifying = verifying;) 7162 } 7163 7164 void MarkFromRootsClosure::reset(HeapWord* addr) { 7165 assert(_markStack->isEmpty(), "would cause duplicates on stack"); 7166 assert(_span.contains(addr), "Out of bounds _finger?"); 7167 _finger = addr; 7168 _threshold = (HeapWord*)round_to( 7169 (intptr_t)_finger, CardTableModRefBS::card_size); 7170 } 7171 7172 // Should revisit to see if this should be restructured for 7173 // greater efficiency. 7174 bool MarkFromRootsClosure::do_bit(size_t offset) { 7175 if (_skipBits > 0) { 7176 _skipBits--; 7177 return true; 7178 } 7179 // convert offset into a HeapWord* 7180 HeapWord* addr = _bitMap->startWord() + offset; 7181 assert(_bitMap->endWord() && addr < _bitMap->endWord(), 7182 "address out of range"); 7183 assert(_bitMap->isMarked(addr), "tautology"); 7184 if (_bitMap->isMarked(addr+1)) { 7185 // this is an allocated but not yet initialized object 7186 assert(_skipBits == 0, "tautology"); 7187 _skipBits = 2; // skip next two marked bits ("Printezis-marks") 7188 oop p = oop(addr); 7189 if (p->klass_or_null() == NULL || !p->is_parsable()) { 7190 DEBUG_ONLY(if (!_verifying) {) 7191 // We re-dirty the cards on which this object lies and increase 7192 // the _threshold so that we'll come back to scan this object 7193 // during the preclean or remark phase. (CMSCleanOnEnter) 7194 if (CMSCleanOnEnter) { 7195 size_t sz = _collector->block_size_using_printezis_bits(addr); 7196 HeapWord* end_card_addr = (HeapWord*)round_to( 7197 (intptr_t)(addr+sz), CardTableModRefBS::card_size); 7198 MemRegion redirty_range = MemRegion(addr, end_card_addr); 7199 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 7200 // Bump _threshold to end_card_addr; note that 7201 // _threshold cannot possibly exceed end_card_addr, anyhow. 7202 // This prevents future clearing of the card as the scan proceeds 7203 // to the right. 7204 assert(_threshold <= end_card_addr, 7205 "Because we are just scanning into this object"); 7206 if (_threshold < end_card_addr) { 7207 _threshold = end_card_addr; 7208 } 7209 if (p->klass_or_null() != NULL) { 7210 // Redirty the range of cards... 7211 _mut->mark_range(redirty_range); 7212 } // ...else the setting of klass will dirty the card anyway. 7213 } 7214 DEBUG_ONLY(}) 7215 return true; 7216 } 7217 } 7218 scanOopsInOop(addr); 7219 return true; 7220 } 7221 7222 // We take a break if we've been at this for a while, 7223 // so as to avoid monopolizing the locks involved. 7224 void MarkFromRootsClosure::do_yield_work() { 7225 // First give up the locks, then yield, then re-lock 7226 // We should probably use a constructor/destructor idiom to 7227 // do this unlock/lock or modify the MutexUnlocker class to 7228 // serve our purpose. XXX 7229 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 7230 "CMS thread should hold CMS token"); 7231 assert_lock_strong(_bitMap->lock()); 7232 DEBUG_ONLY(RememberKlassesChecker mux(false);) 7233 _bitMap->lock()->unlock(); 7234 ConcurrentMarkSweepThread::desynchronize(true); 7235 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7236 _collector->stopTimer(); 7237 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 7238 if (PrintCMSStatistics != 0) { 7239 _collector->incrementYields(); 7240 } 7241 _collector->icms_wait(); 7242 7243 // See the comment in coordinator_yield() 7244 for (unsigned i = 0; i < CMSYieldSleepCount && 7245 ConcurrentMarkSweepThread::should_yield() && 7246 !CMSCollector::foregroundGCIsActive(); ++i) { 7247 os::sleep(Thread::current(), 1, false); 7248 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7249 } 7250 7251 ConcurrentMarkSweepThread::synchronize(true); 7252 _bitMap->lock()->lock_without_safepoint_check(); 7253 _collector->startTimer(); 7254 } 7255 7256 void MarkFromRootsClosure::scanOopsInOop(HeapWord* ptr) { 7257 assert(_bitMap->isMarked(ptr), "expected bit to be set"); 7258 assert(_markStack->isEmpty(), 7259 "should drain stack to limit stack usage"); 7260 // convert ptr to an oop preparatory to scanning 7261 oop obj = oop(ptr); 7262 // Ignore mark word in verification below, since we 7263 // may be running concurrent with mutators. 7264 assert(obj->is_oop(true), "should be an oop"); 7265 assert(_finger <= ptr, "_finger runneth ahead"); 7266 // advance the finger to right end of this object 7267 _finger = ptr + obj->size(); 7268 assert(_finger > ptr, "we just incremented it above"); 7269 // On large heaps, it may take us some time to get through 7270 // the marking phase (especially if running iCMS). During 7271 // this time it's possible that a lot of mutations have 7272 // accumulated in the card table and the mod union table -- 7273 // these mutation records are redundant until we have 7274 // actually traced into the corresponding card. 7275 // Here, we check whether advancing the finger would make 7276 // us cross into a new card, and if so clear corresponding 7277 // cards in the MUT (preclean them in the card-table in the 7278 // future). 7279 7280 DEBUG_ONLY(if (!_verifying) {) 7281 // The clean-on-enter optimization is disabled by default, 7282 // until we fix 6178663. 7283 if (CMSCleanOnEnter && (_finger > _threshold)) { 7284 // [_threshold, _finger) represents the interval 7285 // of cards to be cleared in MUT (or precleaned in card table). 7286 // The set of cards to be cleared is all those that overlap 7287 // with the interval [_threshold, _finger); note that 7288 // _threshold is always kept card-aligned but _finger isn't 7289 // always card-aligned. 7290 HeapWord* old_threshold = _threshold; 7291 assert(old_threshold == (HeapWord*)round_to( 7292 (intptr_t)old_threshold, CardTableModRefBS::card_size), 7293 "_threshold should always be card-aligned"); 7294 _threshold = (HeapWord*)round_to( 7295 (intptr_t)_finger, CardTableModRefBS::card_size); 7296 MemRegion mr(old_threshold, _threshold); 7297 assert(!mr.is_empty(), "Control point invariant"); 7298 assert(_span.contains(mr), "Should clear within span"); 7299 // XXX When _finger crosses from old gen into perm gen 7300 // we may be doing unnecessary cleaning; do better in the 7301 // future by detecting that condition and clearing fewer 7302 // MUT/CT entries. 7303 _mut->clear_range(mr); 7304 } 7305 DEBUG_ONLY(}) 7306 // Note: the finger doesn't advance while we drain 7307 // the stack below. 7308 PushOrMarkClosure pushOrMarkClosure(_collector, 7309 _span, _bitMap, _markStack, 7310 _revisitStack, 7311 _finger, this); 7312 bool res = _markStack->push(obj); 7313 assert(res, "Empty non-zero size stack should have space for single push"); 7314 while (!_markStack->isEmpty()) { 7315 oop new_oop = _markStack->pop(); 7316 // Skip verifying header mark word below because we are 7317 // running concurrent with mutators. 7318 assert(new_oop->is_oop(true), "Oops! expected to pop an oop"); 7319 // now scan this oop's oops 7320 new_oop->oop_iterate(&pushOrMarkClosure); 7321 do_yield_check(); 7322 } 7323 assert(_markStack->isEmpty(), "tautology, emphasizing post-condition"); 7324 } 7325 7326 Par_MarkFromRootsClosure::Par_MarkFromRootsClosure(CMSConcMarkingTask* task, 7327 CMSCollector* collector, MemRegion span, 7328 CMSBitMap* bit_map, 7329 OopTaskQueue* work_queue, 7330 CMSMarkStack* overflow_stack, 7331 CMSMarkStack* revisit_stack, 7332 bool should_yield): 7333 _collector(collector), 7334 _whole_span(collector->_span), 7335 _span(span), 7336 _bit_map(bit_map), 7337 _mut(&collector->_modUnionTable), 7338 _work_queue(work_queue), 7339 _overflow_stack(overflow_stack), 7340 _revisit_stack(revisit_stack), 7341 _yield(should_yield), 7342 _skip_bits(0), 7343 _task(task) 7344 { 7345 assert(_work_queue->size() == 0, "work_queue should be empty"); 7346 _finger = span.start(); 7347 _threshold = _finger; // XXX Defer clear-on-enter optimization for now 7348 assert(_span.contains(_finger), "Out of bounds _finger?"); 7349 } 7350 7351 // Should revisit to see if this should be restructured for 7352 // greater efficiency. 7353 bool Par_MarkFromRootsClosure::do_bit(size_t offset) { 7354 if (_skip_bits > 0) { 7355 _skip_bits--; 7356 return true; 7357 } 7358 // convert offset into a HeapWord* 7359 HeapWord* addr = _bit_map->startWord() + offset; 7360 assert(_bit_map->endWord() && addr < _bit_map->endWord(), 7361 "address out of range"); 7362 assert(_bit_map->isMarked(addr), "tautology"); 7363 if (_bit_map->isMarked(addr+1)) { 7364 // this is an allocated object that might not yet be initialized 7365 assert(_skip_bits == 0, "tautology"); 7366 _skip_bits = 2; // skip next two marked bits ("Printezis-marks") 7367 oop p = oop(addr); 7368 if (p->klass_or_null() == NULL || !p->is_parsable()) { 7369 // in the case of Clean-on-Enter optimization, redirty card 7370 // and avoid clearing card by increasing the threshold. 7371 return true; 7372 } 7373 } 7374 scan_oops_in_oop(addr); 7375 return true; 7376 } 7377 7378 void Par_MarkFromRootsClosure::scan_oops_in_oop(HeapWord* ptr) { 7379 assert(_bit_map->isMarked(ptr), "expected bit to be set"); 7380 // Should we assert that our work queue is empty or 7381 // below some drain limit? 7382 assert(_work_queue->size() == 0, 7383 "should drain stack to limit stack usage"); 7384 // convert ptr to an oop preparatory to scanning 7385 oop obj = oop(ptr); 7386 // Ignore mark word in verification below, since we 7387 // may be running concurrent with mutators. 7388 assert(obj->is_oop(true), "should be an oop"); 7389 assert(_finger <= ptr, "_finger runneth ahead"); 7390 // advance the finger to right end of this object 7391 _finger = ptr + obj->size(); 7392 assert(_finger > ptr, "we just incremented it above"); 7393 // On large heaps, it may take us some time to get through 7394 // the marking phase (especially if running iCMS). During 7395 // this time it's possible that a lot of mutations have 7396 // accumulated in the card table and the mod union table -- 7397 // these mutation records are redundant until we have 7398 // actually traced into the corresponding card. 7399 // Here, we check whether advancing the finger would make 7400 // us cross into a new card, and if so clear corresponding 7401 // cards in the MUT (preclean them in the card-table in the 7402 // future). 7403 7404 // The clean-on-enter optimization is disabled by default, 7405 // until we fix 6178663. 7406 if (CMSCleanOnEnter && (_finger > _threshold)) { 7407 // [_threshold, _finger) represents the interval 7408 // of cards to be cleared in MUT (or precleaned in card table). 7409 // The set of cards to be cleared is all those that overlap 7410 // with the interval [_threshold, _finger); note that 7411 // _threshold is always kept card-aligned but _finger isn't 7412 // always card-aligned. 7413 HeapWord* old_threshold = _threshold; 7414 assert(old_threshold == (HeapWord*)round_to( 7415 (intptr_t)old_threshold, CardTableModRefBS::card_size), 7416 "_threshold should always be card-aligned"); 7417 _threshold = (HeapWord*)round_to( 7418 (intptr_t)_finger, CardTableModRefBS::card_size); 7419 MemRegion mr(old_threshold, _threshold); 7420 assert(!mr.is_empty(), "Control point invariant"); 7421 assert(_span.contains(mr), "Should clear within span"); // _whole_span ?? 7422 // XXX When _finger crosses from old gen into perm gen 7423 // we may be doing unnecessary cleaning; do better in the 7424 // future by detecting that condition and clearing fewer 7425 // MUT/CT entries. 7426 _mut->clear_range(mr); 7427 } 7428 7429 // Note: the local finger doesn't advance while we drain 7430 // the stack below, but the global finger sure can and will. 7431 HeapWord** gfa = _task->global_finger_addr(); 7432 Par_PushOrMarkClosure pushOrMarkClosure(_collector, 7433 _span, _bit_map, 7434 _work_queue, 7435 _overflow_stack, 7436 _revisit_stack, 7437 _finger, 7438 gfa, this); 7439 bool res = _work_queue->push(obj); // overflow could occur here 7440 assert(res, "Will hold once we use workqueues"); 7441 while (true) { 7442 oop new_oop; 7443 if (!_work_queue->pop_local(new_oop)) { 7444 // We emptied our work_queue; check if there's stuff that can 7445 // be gotten from the overflow stack. 7446 if (CMSConcMarkingTask::get_work_from_overflow_stack( 7447 _overflow_stack, _work_queue)) { 7448 do_yield_check(); 7449 continue; 7450 } else { // done 7451 break; 7452 } 7453 } 7454 // Skip verifying header mark word below because we are 7455 // running concurrent with mutators. 7456 assert(new_oop->is_oop(true), "Oops! expected to pop an oop"); 7457 // now scan this oop's oops 7458 new_oop->oop_iterate(&pushOrMarkClosure); 7459 do_yield_check(); 7460 } 7461 assert(_work_queue->size() == 0, "tautology, emphasizing post-condition"); 7462 } 7463 7464 // Yield in response to a request from VM Thread or 7465 // from mutators. 7466 void Par_MarkFromRootsClosure::do_yield_work() { 7467 assert(_task != NULL, "sanity"); 7468 _task->yield(); 7469 } 7470 7471 // A variant of the above used for verifying CMS marking work. 7472 MarkFromRootsVerifyClosure::MarkFromRootsVerifyClosure(CMSCollector* collector, 7473 MemRegion span, 7474 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 7475 CMSMarkStack* mark_stack): 7476 _collector(collector), 7477 _span(span), 7478 _verification_bm(verification_bm), 7479 _cms_bm(cms_bm), 7480 _mark_stack(mark_stack), 7481 _pam_verify_closure(collector, span, verification_bm, cms_bm, 7482 mark_stack) 7483 { 7484 assert(_mark_stack->isEmpty(), "stack should be empty"); 7485 _finger = _verification_bm->startWord(); 7486 assert(_collector->_restart_addr == NULL, "Sanity check"); 7487 assert(_span.contains(_finger), "Out of bounds _finger?"); 7488 } 7489 7490 void MarkFromRootsVerifyClosure::reset(HeapWord* addr) { 7491 assert(_mark_stack->isEmpty(), "would cause duplicates on stack"); 7492 assert(_span.contains(addr), "Out of bounds _finger?"); 7493 _finger = addr; 7494 } 7495 7496 // Should revisit to see if this should be restructured for 7497 // greater efficiency. 7498 bool MarkFromRootsVerifyClosure::do_bit(size_t offset) { 7499 // convert offset into a HeapWord* 7500 HeapWord* addr = _verification_bm->startWord() + offset; 7501 assert(_verification_bm->endWord() && addr < _verification_bm->endWord(), 7502 "address out of range"); 7503 assert(_verification_bm->isMarked(addr), "tautology"); 7504 assert(_cms_bm->isMarked(addr), "tautology"); 7505 7506 assert(_mark_stack->isEmpty(), 7507 "should drain stack to limit stack usage"); 7508 // convert addr to an oop preparatory to scanning 7509 oop obj = oop(addr); 7510 assert(obj->is_oop(), "should be an oop"); 7511 assert(_finger <= addr, "_finger runneth ahead"); 7512 // advance the finger to right end of this object 7513 _finger = addr + obj->size(); 7514 assert(_finger > addr, "we just incremented it above"); 7515 // Note: the finger doesn't advance while we drain 7516 // the stack below. 7517 bool res = _mark_stack->push(obj); 7518 assert(res, "Empty non-zero size stack should have space for single push"); 7519 while (!_mark_stack->isEmpty()) { 7520 oop new_oop = _mark_stack->pop(); 7521 assert(new_oop->is_oop(), "Oops! expected to pop an oop"); 7522 // now scan this oop's oops 7523 new_oop->oop_iterate(&_pam_verify_closure); 7524 } 7525 assert(_mark_stack->isEmpty(), "tautology, emphasizing post-condition"); 7526 return true; 7527 } 7528 7529 PushAndMarkVerifyClosure::PushAndMarkVerifyClosure( 7530 CMSCollector* collector, MemRegion span, 7531 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 7532 CMSMarkStack* mark_stack): 7533 OopClosure(collector->ref_processor()), 7534 _collector(collector), 7535 _span(span), 7536 _verification_bm(verification_bm), 7537 _cms_bm(cms_bm), 7538 _mark_stack(mark_stack) 7539 { } 7540 7541 void PushAndMarkVerifyClosure::do_oop(oop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 7542 void PushAndMarkVerifyClosure::do_oop(narrowOop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 7543 7544 // Upon stack overflow, we discard (part of) the stack, 7545 // remembering the least address amongst those discarded 7546 // in CMSCollector's _restart_address. 7547 void PushAndMarkVerifyClosure::handle_stack_overflow(HeapWord* lost) { 7548 // Remember the least grey address discarded 7549 HeapWord* ra = (HeapWord*)_mark_stack->least_value(lost); 7550 _collector->lower_restart_addr(ra); 7551 _mark_stack->reset(); // discard stack contents 7552 _mark_stack->expand(); // expand the stack if possible 7553 } 7554 7555 void PushAndMarkVerifyClosure::do_oop(oop obj) { 7556 assert(obj->is_oop_or_null(), "expected an oop or NULL"); 7557 HeapWord* addr = (HeapWord*)obj; 7558 if (_span.contains(addr) && !_verification_bm->isMarked(addr)) { 7559 // Oop lies in _span and isn't yet grey or black 7560 _verification_bm->mark(addr); // now grey 7561 if (!_cms_bm->isMarked(addr)) { 7562 oop(addr)->print(); 7563 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)", 7564 addr); 7565 fatal("... aborting"); 7566 } 7567 7568 if (!_mark_stack->push(obj)) { // stack overflow 7569 if (PrintCMSStatistics != 0) { 7570 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 7571 SIZE_FORMAT, _mark_stack->capacity()); 7572 } 7573 assert(_mark_stack->isFull(), "Else push should have succeeded"); 7574 handle_stack_overflow(addr); 7575 } 7576 // anything including and to the right of _finger 7577 // will be scanned as we iterate over the remainder of the 7578 // bit map 7579 } 7580 } 7581 7582 PushOrMarkClosure::PushOrMarkClosure(CMSCollector* collector, 7583 MemRegion span, 7584 CMSBitMap* bitMap, CMSMarkStack* markStack, 7585 CMSMarkStack* revisitStack, 7586 HeapWord* finger, MarkFromRootsClosure* parent) : 7587 KlassRememberingOopClosure(collector, collector->ref_processor(), revisitStack), 7588 _span(span), 7589 _bitMap(bitMap), 7590 _markStack(markStack), 7591 _finger(finger), 7592 _parent(parent) 7593 { } 7594 7595 Par_PushOrMarkClosure::Par_PushOrMarkClosure(CMSCollector* collector, 7596 MemRegion span, 7597 CMSBitMap* bit_map, 7598 OopTaskQueue* work_queue, 7599 CMSMarkStack* overflow_stack, 7600 CMSMarkStack* revisit_stack, 7601 HeapWord* finger, 7602 HeapWord** global_finger_addr, 7603 Par_MarkFromRootsClosure* parent) : 7604 Par_KlassRememberingOopClosure(collector, 7605 collector->ref_processor(), 7606 revisit_stack), 7607 _whole_span(collector->_span), 7608 _span(span), 7609 _bit_map(bit_map), 7610 _work_queue(work_queue), 7611 _overflow_stack(overflow_stack), 7612 _finger(finger), 7613 _global_finger_addr(global_finger_addr), 7614 _parent(parent) 7615 { } 7616 7617 // Assumes thread-safe access by callers, who are 7618 // responsible for mutual exclusion. 7619 void CMSCollector::lower_restart_addr(HeapWord* low) { 7620 assert(_span.contains(low), "Out of bounds addr"); 7621 if (_restart_addr == NULL) { 7622 _restart_addr = low; 7623 } else { 7624 _restart_addr = MIN2(_restart_addr, low); 7625 } 7626 } 7627 7628 // Upon stack overflow, we discard (part of) the stack, 7629 // remembering the least address amongst those discarded 7630 // in CMSCollector's _restart_address. 7631 void PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 7632 // Remember the least grey address discarded 7633 HeapWord* ra = (HeapWord*)_markStack->least_value(lost); 7634 _collector->lower_restart_addr(ra); 7635 _markStack->reset(); // discard stack contents 7636 _markStack->expand(); // expand the stack if possible 7637 } 7638 7639 // Upon stack overflow, we discard (part of) the stack, 7640 // remembering the least address amongst those discarded 7641 // in CMSCollector's _restart_address. 7642 void Par_PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 7643 // We need to do this under a mutex to prevent other 7644 // workers from interfering with the work done below. 7645 MutexLockerEx ml(_overflow_stack->par_lock(), 7646 Mutex::_no_safepoint_check_flag); 7647 // Remember the least grey address discarded 7648 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 7649 _collector->lower_restart_addr(ra); 7650 _overflow_stack->reset(); // discard stack contents 7651 _overflow_stack->expand(); // expand the stack if possible 7652 } 7653 7654 void PushOrMarkClosure::do_oop(oop obj) { 7655 // Ignore mark word because we are running concurrent with mutators. 7656 assert(obj->is_oop_or_null(true), "expected an oop or NULL"); 7657 HeapWord* addr = (HeapWord*)obj; 7658 if (_span.contains(addr) && !_bitMap->isMarked(addr)) { 7659 // Oop lies in _span and isn't yet grey or black 7660 _bitMap->mark(addr); // now grey 7661 if (addr < _finger) { 7662 // the bit map iteration has already either passed, or 7663 // sampled, this bit in the bit map; we'll need to 7664 // use the marking stack to scan this oop's oops. 7665 bool simulate_overflow = false; 7666 NOT_PRODUCT( 7667 if (CMSMarkStackOverflowALot && 7668 _collector->simulate_overflow()) { 7669 // simulate a stack overflow 7670 simulate_overflow = true; 7671 } 7672 ) 7673 if (simulate_overflow || !_markStack->push(obj)) { // stack overflow 7674 if (PrintCMSStatistics != 0) { 7675 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 7676 SIZE_FORMAT, _markStack->capacity()); 7677 } 7678 assert(simulate_overflow || _markStack->isFull(), "Else push should have succeeded"); 7679 handle_stack_overflow(addr); 7680 } 7681 } 7682 // anything including and to the right of _finger 7683 // will be scanned as we iterate over the remainder of the 7684 // bit map 7685 do_yield_check(); 7686 } 7687 } 7688 7689 void PushOrMarkClosure::do_oop(oop* p) { PushOrMarkClosure::do_oop_work(p); } 7690 void PushOrMarkClosure::do_oop(narrowOop* p) { PushOrMarkClosure::do_oop_work(p); } 7691 7692 void Par_PushOrMarkClosure::do_oop(oop obj) { 7693 // Ignore mark word because we are running concurrent with mutators. 7694 assert(obj->is_oop_or_null(true), "expected an oop or NULL"); 7695 HeapWord* addr = (HeapWord*)obj; 7696 if (_whole_span.contains(addr) && !_bit_map->isMarked(addr)) { 7697 // Oop lies in _span and isn't yet grey or black 7698 // We read the global_finger (volatile read) strictly after marking oop 7699 bool res = _bit_map->par_mark(addr); // now grey 7700 volatile HeapWord** gfa = (volatile HeapWord**)_global_finger_addr; 7701 // Should we push this marked oop on our stack? 7702 // -- if someone else marked it, nothing to do 7703 // -- if target oop is above global finger nothing to do 7704 // -- if target oop is in chunk and above local finger 7705 // then nothing to do 7706 // -- else push on work queue 7707 if ( !res // someone else marked it, they will deal with it 7708 || (addr >= *gfa) // will be scanned in a later task 7709 || (_span.contains(addr) && addr >= _finger)) { // later in this chunk 7710 return; 7711 } 7712 // the bit map iteration has already either passed, or 7713 // sampled, this bit in the bit map; we'll need to 7714 // use the marking stack to scan this oop's oops. 7715 bool simulate_overflow = false; 7716 NOT_PRODUCT( 7717 if (CMSMarkStackOverflowALot && 7718 _collector->simulate_overflow()) { 7719 // simulate a stack overflow 7720 simulate_overflow = true; 7721 } 7722 ) 7723 if (simulate_overflow || 7724 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) { 7725 // stack overflow 7726 if (PrintCMSStatistics != 0) { 7727 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 7728 SIZE_FORMAT, _overflow_stack->capacity()); 7729 } 7730 // We cannot assert that the overflow stack is full because 7731 // it may have been emptied since. 7732 assert(simulate_overflow || 7733 _work_queue->size() == _work_queue->max_elems(), 7734 "Else push should have succeeded"); 7735 handle_stack_overflow(addr); 7736 } 7737 do_yield_check(); 7738 } 7739 } 7740 7741 void Par_PushOrMarkClosure::do_oop(oop* p) { Par_PushOrMarkClosure::do_oop_work(p); } 7742 void Par_PushOrMarkClosure::do_oop(narrowOop* p) { Par_PushOrMarkClosure::do_oop_work(p); } 7743 7744 KlassRememberingOopClosure::KlassRememberingOopClosure(CMSCollector* collector, 7745 ReferenceProcessor* rp, 7746 CMSMarkStack* revisit_stack) : 7747 OopClosure(rp), 7748 _collector(collector), 7749 _revisit_stack(revisit_stack), 7750 _should_remember_klasses(collector->should_unload_classes()) {} 7751 7752 PushAndMarkClosure::PushAndMarkClosure(CMSCollector* collector, 7753 MemRegion span, 7754 ReferenceProcessor* rp, 7755 CMSBitMap* bit_map, 7756 CMSBitMap* mod_union_table, 7757 CMSMarkStack* mark_stack, 7758 CMSMarkStack* revisit_stack, 7759 bool concurrent_precleaning): 7760 KlassRememberingOopClosure(collector, rp, revisit_stack), 7761 _span(span), 7762 _bit_map(bit_map), 7763 _mod_union_table(mod_union_table), 7764 _mark_stack(mark_stack), 7765 _concurrent_precleaning(concurrent_precleaning) 7766 { 7767 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 7768 } 7769 7770 // Grey object rescan during pre-cleaning and second checkpoint phases -- 7771 // the non-parallel version (the parallel version appears further below.) 7772 void PushAndMarkClosure::do_oop(oop obj) { 7773 // Ignore mark word verification. If during concurrent precleaning, 7774 // the object monitor may be locked. If during the checkpoint 7775 // phases, the object may already have been reached by a different 7776 // path and may be at the end of the global overflow list (so 7777 // the mark word may be NULL). 7778 assert(obj->is_oop_or_null(true /* ignore mark word */), 7779 "expected an oop or NULL"); 7780 HeapWord* addr = (HeapWord*)obj; 7781 // Check if oop points into the CMS generation 7782 // and is not marked 7783 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 7784 // a white object ... 7785 _bit_map->mark(addr); // ... now grey 7786 // push on the marking stack (grey set) 7787 bool simulate_overflow = false; 7788 NOT_PRODUCT( 7789 if (CMSMarkStackOverflowALot && 7790 _collector->simulate_overflow()) { 7791 // simulate a stack overflow 7792 simulate_overflow = true; 7793 } 7794 ) 7795 if (simulate_overflow || !_mark_stack->push(obj)) { 7796 if (_concurrent_precleaning) { 7797 // During precleaning we can just dirty the appropriate card(s) 7798 // in the mod union table, thus ensuring that the object remains 7799 // in the grey set and continue. In the case of object arrays 7800 // we need to dirty all of the cards that the object spans, 7801 // since the rescan of object arrays will be limited to the 7802 // dirty cards. 7803 // Note that no one can be intefering with us in this action 7804 // of dirtying the mod union table, so no locking or atomics 7805 // are required. 7806 if (obj->is_objArray()) { 7807 size_t sz = obj->size(); 7808 HeapWord* end_card_addr = (HeapWord*)round_to( 7809 (intptr_t)(addr+sz), CardTableModRefBS::card_size); 7810 MemRegion redirty_range = MemRegion(addr, end_card_addr); 7811 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 7812 _mod_union_table->mark_range(redirty_range); 7813 } else { 7814 _mod_union_table->mark(addr); 7815 } 7816 _collector->_ser_pmc_preclean_ovflw++; 7817 } else { 7818 // During the remark phase, we need to remember this oop 7819 // in the overflow list. 7820 _collector->push_on_overflow_list(obj); 7821 _collector->_ser_pmc_remark_ovflw++; 7822 } 7823 } 7824 } 7825 } 7826 7827 Par_PushAndMarkClosure::Par_PushAndMarkClosure(CMSCollector* collector, 7828 MemRegion span, 7829 ReferenceProcessor* rp, 7830 CMSBitMap* bit_map, 7831 OopTaskQueue* work_queue, 7832 CMSMarkStack* revisit_stack): 7833 Par_KlassRememberingOopClosure(collector, rp, revisit_stack), 7834 _span(span), 7835 _bit_map(bit_map), 7836 _work_queue(work_queue) 7837 { 7838 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 7839 } 7840 7841 void PushAndMarkClosure::do_oop(oop* p) { PushAndMarkClosure::do_oop_work(p); } 7842 void PushAndMarkClosure::do_oop(narrowOop* p) { PushAndMarkClosure::do_oop_work(p); } 7843 7844 // Grey object rescan during second checkpoint phase -- 7845 // the parallel version. 7846 void Par_PushAndMarkClosure::do_oop(oop obj) { 7847 // In the assert below, we ignore the mark word because 7848 // this oop may point to an already visited object that is 7849 // on the overflow stack (in which case the mark word has 7850 // been hijacked for chaining into the overflow stack -- 7851 // if this is the last object in the overflow stack then 7852 // its mark word will be NULL). Because this object may 7853 // have been subsequently popped off the global overflow 7854 // stack, and the mark word possibly restored to the prototypical 7855 // value, by the time we get to examined this failing assert in 7856 // the debugger, is_oop_or_null(false) may subsequently start 7857 // to hold. 7858 assert(obj->is_oop_or_null(true), 7859 "expected an oop or NULL"); 7860 HeapWord* addr = (HeapWord*)obj; 7861 // Check if oop points into the CMS generation 7862 // and is not marked 7863 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 7864 // a white object ... 7865 // If we manage to "claim" the object, by being the 7866 // first thread to mark it, then we push it on our 7867 // marking stack 7868 if (_bit_map->par_mark(addr)) { // ... now grey 7869 // push on work queue (grey set) 7870 bool simulate_overflow = false; 7871 NOT_PRODUCT( 7872 if (CMSMarkStackOverflowALot && 7873 _collector->par_simulate_overflow()) { 7874 // simulate a stack overflow 7875 simulate_overflow = true; 7876 } 7877 ) 7878 if (simulate_overflow || !_work_queue->push(obj)) { 7879 _collector->par_push_on_overflow_list(obj); 7880 _collector->_par_pmc_remark_ovflw++; // imprecise OK: no need to CAS 7881 } 7882 } // Else, some other thread got there first 7883 } 7884 } 7885 7886 void Par_PushAndMarkClosure::do_oop(oop* p) { Par_PushAndMarkClosure::do_oop_work(p); } 7887 void Par_PushAndMarkClosure::do_oop(narrowOop* p) { Par_PushAndMarkClosure::do_oop_work(p); } 7888 7889 void PushAndMarkClosure::remember_mdo(DataLayout* v) { 7890 // TBD 7891 } 7892 7893 void Par_PushAndMarkClosure::remember_mdo(DataLayout* v) { 7894 // TBD 7895 } 7896 7897 void CMSPrecleanRefsYieldClosure::do_yield_work() { 7898 DEBUG_ONLY(RememberKlassesChecker mux(false);) 7899 Mutex* bml = _collector->bitMapLock(); 7900 assert_lock_strong(bml); 7901 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 7902 "CMS thread should hold CMS token"); 7903 7904 bml->unlock(); 7905 ConcurrentMarkSweepThread::desynchronize(true); 7906 7907 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7908 7909 _collector->stopTimer(); 7910 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 7911 if (PrintCMSStatistics != 0) { 7912 _collector->incrementYields(); 7913 } 7914 _collector->icms_wait(); 7915 7916 // See the comment in coordinator_yield() 7917 for (unsigned i = 0; i < CMSYieldSleepCount && 7918 ConcurrentMarkSweepThread::should_yield() && 7919 !CMSCollector::foregroundGCIsActive(); ++i) { 7920 os::sleep(Thread::current(), 1, false); 7921 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7922 } 7923 7924 ConcurrentMarkSweepThread::synchronize(true); 7925 bml->lock(); 7926 7927 _collector->startTimer(); 7928 } 7929 7930 bool CMSPrecleanRefsYieldClosure::should_return() { 7931 if (ConcurrentMarkSweepThread::should_yield()) { 7932 do_yield_work(); 7933 } 7934 return _collector->foregroundGCIsActive(); 7935 } 7936 7937 void MarkFromDirtyCardsClosure::do_MemRegion(MemRegion mr) { 7938 assert(((size_t)mr.start())%CardTableModRefBS::card_size_in_words == 0, 7939 "mr should be aligned to start at a card boundary"); 7940 // We'd like to assert: 7941 // assert(mr.word_size()%CardTableModRefBS::card_size_in_words == 0, 7942 // "mr should be a range of cards"); 7943 // However, that would be too strong in one case -- the last 7944 // partition ends at _unallocated_block which, in general, can be 7945 // an arbitrary boundary, not necessarily card aligned. 7946 if (PrintCMSStatistics != 0) { 7947 _num_dirty_cards += 7948 mr.word_size()/CardTableModRefBS::card_size_in_words; 7949 } 7950 _space->object_iterate_mem(mr, &_scan_cl); 7951 } 7952 7953 SweepClosure::SweepClosure(CMSCollector* collector, 7954 ConcurrentMarkSweepGeneration* g, 7955 CMSBitMap* bitMap, bool should_yield) : 7956 _collector(collector), 7957 _g(g), 7958 _sp(g->cmsSpace()), 7959 _limit(_sp->sweep_limit()), 7960 _freelistLock(_sp->freelistLock()), 7961 _bitMap(bitMap), 7962 _yield(should_yield), 7963 _inFreeRange(false), // No free range at beginning of sweep 7964 _freeRangeInFreeLists(false), // No free range at beginning of sweep 7965 _lastFreeRangeCoalesced(false), 7966 _freeFinger(g->used_region().start()) 7967 { 7968 NOT_PRODUCT( 7969 _numObjectsFreed = 0; 7970 _numWordsFreed = 0; 7971 _numObjectsLive = 0; 7972 _numWordsLive = 0; 7973 _numObjectsAlreadyFree = 0; 7974 _numWordsAlreadyFree = 0; 7975 _last_fc = NULL; 7976 7977 _sp->initializeIndexedFreeListArrayReturnedBytes(); 7978 _sp->dictionary()->initialize_dict_returned_bytes(); 7979 ) 7980 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7981 "sweep _limit out of bounds"); 7982 if (CMSTraceSweeper) { 7983 gclog_or_tty->print_cr("\n====================\nStarting new sweep with limit " PTR_FORMAT, 7984 _limit); 7985 } 7986 } 7987 7988 void SweepClosure::print_on(outputStream* st) const { 7989 tty->print_cr("_sp = [" PTR_FORMAT "," PTR_FORMAT ")", 7990 _sp->bottom(), _sp->end()); 7991 tty->print_cr("_limit = " PTR_FORMAT, _limit); 7992 tty->print_cr("_freeFinger = " PTR_FORMAT, _freeFinger); 7993 NOT_PRODUCT(tty->print_cr("_last_fc = " PTR_FORMAT, _last_fc);) 7994 tty->print_cr("_inFreeRange = %d, _freeRangeInFreeLists = %d, _lastFreeRangeCoalesced = %d", 7995 _inFreeRange, _freeRangeInFreeLists, _lastFreeRangeCoalesced); 7996 } 7997 7998 #ifndef PRODUCT 7999 // Assertion checking only: no useful work in product mode -- 8000 // however, if any of the flags below become product flags, 8001 // you may need to review this code to see if it needs to be 8002 // enabled in product mode. 8003 SweepClosure::~SweepClosure() { 8004 assert_lock_strong(_freelistLock); 8005 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 8006 "sweep _limit out of bounds"); 8007 if (inFreeRange()) { 8008 warning("inFreeRange() should have been reset; dumping state of SweepClosure"); 8009 print(); 8010 ShouldNotReachHere(); 8011 } 8012 if (Verbose && PrintGC) { 8013 gclog_or_tty->print("Collected "SIZE_FORMAT" objects, " SIZE_FORMAT " bytes", 8014 _numObjectsFreed, _numWordsFreed*sizeof(HeapWord)); 8015 gclog_or_tty->print_cr("\nLive "SIZE_FORMAT" objects, " 8016 SIZE_FORMAT" bytes " 8017 "Already free "SIZE_FORMAT" objects, "SIZE_FORMAT" bytes", 8018 _numObjectsLive, _numWordsLive*sizeof(HeapWord), 8019 _numObjectsAlreadyFree, _numWordsAlreadyFree*sizeof(HeapWord)); 8020 size_t totalBytes = (_numWordsFreed + _numWordsLive + _numWordsAlreadyFree) 8021 * sizeof(HeapWord); 8022 gclog_or_tty->print_cr("Total sweep: "SIZE_FORMAT" bytes", totalBytes); 8023 8024 if (PrintCMSStatistics && CMSVerifyReturnedBytes) { 8025 size_t indexListReturnedBytes = _sp->sumIndexedFreeListArrayReturnedBytes(); 8026 size_t dict_returned_bytes = _sp->dictionary()->sum_dict_returned_bytes(); 8027 size_t returned_bytes = indexListReturnedBytes + dict_returned_bytes; 8028 gclog_or_tty->print("Returned "SIZE_FORMAT" bytes", returned_bytes); 8029 gclog_or_tty->print(" Indexed List Returned "SIZE_FORMAT" bytes", 8030 indexListReturnedBytes); 8031 gclog_or_tty->print_cr(" Dictionary Returned "SIZE_FORMAT" bytes", 8032 dict_returned_bytes); 8033 } 8034 } 8035 if (CMSTraceSweeper) { 8036 gclog_or_tty->print_cr("end of sweep with _limit = " PTR_FORMAT "\n================", 8037 _limit); 8038 } 8039 } 8040 #endif // PRODUCT 8041 8042 void SweepClosure::initialize_free_range(HeapWord* freeFinger, 8043 bool freeRangeInFreeLists) { 8044 if (CMSTraceSweeper) { 8045 gclog_or_tty->print("---- Start free range at 0x%x with free block (%d)\n", 8046 freeFinger, freeRangeInFreeLists); 8047 } 8048 assert(!inFreeRange(), "Trampling existing free range"); 8049 set_inFreeRange(true); 8050 set_lastFreeRangeCoalesced(false); 8051 8052 set_freeFinger(freeFinger); 8053 set_freeRangeInFreeLists(freeRangeInFreeLists); 8054 if (CMSTestInFreeList) { 8055 if (freeRangeInFreeLists) { 8056 FreeChunk* fc = (FreeChunk*) freeFinger; 8057 assert(fc->is_free(), "A chunk on the free list should be free."); 8058 assert(fc->size() > 0, "Free range should have a size"); 8059 assert(_sp->verify_chunk_in_free_list(fc), "Chunk is not in free lists"); 8060 } 8061 } 8062 } 8063 8064 // Note that the sweeper runs concurrently with mutators. Thus, 8065 // it is possible for direct allocation in this generation to happen 8066 // in the middle of the sweep. Note that the sweeper also coalesces 8067 // contiguous free blocks. Thus, unless the sweeper and the allocator 8068 // synchronize appropriately freshly allocated blocks may get swept up. 8069 // This is accomplished by the sweeper locking the free lists while 8070 // it is sweeping. Thus blocks that are determined to be free are 8071 // indeed free. There is however one additional complication: 8072 // blocks that have been allocated since the final checkpoint and 8073 // mark, will not have been marked and so would be treated as 8074 // unreachable and swept up. To prevent this, the allocator marks 8075 // the bit map when allocating during the sweep phase. This leads, 8076 // however, to a further complication -- objects may have been allocated 8077 // but not yet initialized -- in the sense that the header isn't yet 8078 // installed. The sweeper can not then determine the size of the block 8079 // in order to skip over it. To deal with this case, we use a technique 8080 // (due to Printezis) to encode such uninitialized block sizes in the 8081 // bit map. Since the bit map uses a bit per every HeapWord, but the 8082 // CMS generation has a minimum object size of 3 HeapWords, it follows 8083 // that "normal marks" won't be adjacent in the bit map (there will 8084 // always be at least two 0 bits between successive 1 bits). We make use 8085 // of these "unused" bits to represent uninitialized blocks -- the bit 8086 // corresponding to the start of the uninitialized object and the next 8087 // bit are both set. Finally, a 1 bit marks the end of the object that 8088 // started with the two consecutive 1 bits to indicate its potentially 8089 // uninitialized state. 8090 8091 size_t SweepClosure::do_blk_careful(HeapWord* addr) { 8092 FreeChunk* fc = (FreeChunk*)addr; 8093 size_t res; 8094 8095 // Check if we are done sweeping. Below we check "addr >= _limit" rather 8096 // than "addr == _limit" because although _limit was a block boundary when 8097 // we started the sweep, it may no longer be one because heap expansion 8098 // may have caused us to coalesce the block ending at the address _limit 8099 // with a newly expanded chunk (this happens when _limit was set to the 8100 // previous _end of the space), so we may have stepped past _limit: 8101 // see the following Zeno-like trail of CRs 6977970, 7008136, 7042740. 8102 if (addr >= _limit) { // we have swept up to or past the limit: finish up 8103 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 8104 "sweep _limit out of bounds"); 8105 assert(addr < _sp->end(), "addr out of bounds"); 8106 // Flush any free range we might be holding as a single 8107 // coalesced chunk to the appropriate free list. 8108 if (inFreeRange()) { 8109 assert(freeFinger() >= _sp->bottom() && freeFinger() < _limit, 8110 err_msg("freeFinger() " PTR_FORMAT" is out-of-bounds", freeFinger())); 8111 flush_cur_free_chunk(freeFinger(), 8112 pointer_delta(addr, freeFinger())); 8113 if (CMSTraceSweeper) { 8114 gclog_or_tty->print("Sweep: last chunk: "); 8115 gclog_or_tty->print("put_free_blk 0x%x ("SIZE_FORMAT") " 8116 "[coalesced:"SIZE_FORMAT"]\n", 8117 freeFinger(), pointer_delta(addr, freeFinger()), 8118 lastFreeRangeCoalesced()); 8119 } 8120 } 8121 8122 // help the iterator loop finish 8123 return pointer_delta(_sp->end(), addr); 8124 } 8125 8126 assert(addr < _limit, "sweep invariant"); 8127 // check if we should yield 8128 do_yield_check(addr); 8129 if (fc->is_free()) { 8130 // Chunk that is already free 8131 res = fc->size(); 8132 do_already_free_chunk(fc); 8133 debug_only(_sp->verifyFreeLists()); 8134 // If we flush the chunk at hand in lookahead_and_flush() 8135 // and it's coalesced with a preceding chunk, then the 8136 // process of "mangling" the payload of the coalesced block 8137 // will cause erasure of the size information from the 8138 // (erstwhile) header of all the coalesced blocks but the 8139 // first, so the first disjunct in the assert will not hold 8140 // in that specific case (in which case the second disjunct 8141 // will hold). 8142 assert(res == fc->size() || ((HeapWord*)fc) + res >= _limit, 8143 "Otherwise the size info doesn't change at this step"); 8144 NOT_PRODUCT( 8145 _numObjectsAlreadyFree++; 8146 _numWordsAlreadyFree += res; 8147 ) 8148 NOT_PRODUCT(_last_fc = fc;) 8149 } else if (!_bitMap->isMarked(addr)) { 8150 // Chunk is fresh garbage 8151 res = do_garbage_chunk(fc); 8152 debug_only(_sp->verifyFreeLists()); 8153 NOT_PRODUCT( 8154 _numObjectsFreed++; 8155 _numWordsFreed += res; 8156 ) 8157 } else { 8158 // Chunk that is alive. 8159 res = do_live_chunk(fc); 8160 debug_only(_sp->verifyFreeLists()); 8161 NOT_PRODUCT( 8162 _numObjectsLive++; 8163 _numWordsLive += res; 8164 ) 8165 } 8166 return res; 8167 } 8168 8169 // For the smart allocation, record following 8170 // split deaths - a free chunk is removed from its free list because 8171 // it is being split into two or more chunks. 8172 // split birth - a free chunk is being added to its free list because 8173 // a larger free chunk has been split and resulted in this free chunk. 8174 // coal death - a free chunk is being removed from its free list because 8175 // it is being coalesced into a large free chunk. 8176 // coal birth - a free chunk is being added to its free list because 8177 // it was created when two or more free chunks where coalesced into 8178 // this free chunk. 8179 // 8180 // These statistics are used to determine the desired number of free 8181 // chunks of a given size. The desired number is chosen to be relative 8182 // to the end of a CMS sweep. The desired number at the end of a sweep 8183 // is the 8184 // count-at-end-of-previous-sweep (an amount that was enough) 8185 // - count-at-beginning-of-current-sweep (the excess) 8186 // + split-births (gains in this size during interval) 8187 // - split-deaths (demands on this size during interval) 8188 // where the interval is from the end of one sweep to the end of the 8189 // next. 8190 // 8191 // When sweeping the sweeper maintains an accumulated chunk which is 8192 // the chunk that is made up of chunks that have been coalesced. That 8193 // will be termed the left-hand chunk. A new chunk of garbage that 8194 // is being considered for coalescing will be referred to as the 8195 // right-hand chunk. 8196 // 8197 // When making a decision on whether to coalesce a right-hand chunk with 8198 // the current left-hand chunk, the current count vs. the desired count 8199 // of the left-hand chunk is considered. Also if the right-hand chunk 8200 // is near the large chunk at the end of the heap (see 8201 // ConcurrentMarkSweepGeneration::isNearLargestChunk()), then the 8202 // left-hand chunk is coalesced. 8203 // 8204 // When making a decision about whether to split a chunk, the desired count 8205 // vs. the current count of the candidate to be split is also considered. 8206 // If the candidate is underpopulated (currently fewer chunks than desired) 8207 // a chunk of an overpopulated (currently more chunks than desired) size may 8208 // be chosen. The "hint" associated with a free list, if non-null, points 8209 // to a free list which may be overpopulated. 8210 // 8211 8212 void SweepClosure::do_already_free_chunk(FreeChunk* fc) { 8213 const size_t size = fc->size(); 8214 // Chunks that cannot be coalesced are not in the 8215 // free lists. 8216 if (CMSTestInFreeList && !fc->cantCoalesce()) { 8217 assert(_sp->verify_chunk_in_free_list(fc), 8218 "free chunk should be in free lists"); 8219 } 8220 // a chunk that is already free, should not have been 8221 // marked in the bit map 8222 HeapWord* const addr = (HeapWord*) fc; 8223 assert(!_bitMap->isMarked(addr), "free chunk should be unmarked"); 8224 // Verify that the bit map has no bits marked between 8225 // addr and purported end of this block. 8226 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 8227 8228 // Some chunks cannot be coalesced under any circumstances. 8229 // See the definition of cantCoalesce(). 8230 if (!fc->cantCoalesce()) { 8231 // This chunk can potentially be coalesced. 8232 if (_sp->adaptive_freelists()) { 8233 // All the work is done in 8234 do_post_free_or_garbage_chunk(fc, size); 8235 } else { // Not adaptive free lists 8236 // this is a free chunk that can potentially be coalesced by the sweeper; 8237 if (!inFreeRange()) { 8238 // if the next chunk is a free block that can't be coalesced 8239 // it doesn't make sense to remove this chunk from the free lists 8240 FreeChunk* nextChunk = (FreeChunk*)(addr + size); 8241 assert((HeapWord*)nextChunk <= _sp->end(), "Chunk size out of bounds?"); 8242 if ((HeapWord*)nextChunk < _sp->end() && // There is another free chunk to the right ... 8243 nextChunk->is_free() && // ... which is free... 8244 nextChunk->cantCoalesce()) { // ... but can't be coalesced 8245 // nothing to do 8246 } else { 8247 // Potentially the start of a new free range: 8248 // Don't eagerly remove it from the free lists. 8249 // No need to remove it if it will just be put 8250 // back again. (Also from a pragmatic point of view 8251 // if it is a free block in a region that is beyond 8252 // any allocated blocks, an assertion will fail) 8253 // Remember the start of a free run. 8254 initialize_free_range(addr, true); 8255 // end - can coalesce with next chunk 8256 } 8257 } else { 8258 // the midst of a free range, we are coalescing 8259 print_free_block_coalesced(fc); 8260 if (CMSTraceSweeper) { 8261 gclog_or_tty->print(" -- pick up free block 0x%x (%d)\n", fc, size); 8262 } 8263 // remove it from the free lists 8264 _sp->removeFreeChunkFromFreeLists(fc); 8265 set_lastFreeRangeCoalesced(true); 8266 // If the chunk is being coalesced and the current free range is 8267 // in the free lists, remove the current free range so that it 8268 // will be returned to the free lists in its entirety - all 8269 // the coalesced pieces included. 8270 if (freeRangeInFreeLists()) { 8271 FreeChunk* ffc = (FreeChunk*) freeFinger(); 8272 assert(ffc->size() == pointer_delta(addr, freeFinger()), 8273 "Size of free range is inconsistent with chunk size."); 8274 if (CMSTestInFreeList) { 8275 assert(_sp->verify_chunk_in_free_list(ffc), 8276 "free range is not in free lists"); 8277 } 8278 _sp->removeFreeChunkFromFreeLists(ffc); 8279 set_freeRangeInFreeLists(false); 8280 } 8281 } 8282 } 8283 // Note that if the chunk is not coalescable (the else arm 8284 // below), we unconditionally flush, without needing to do 8285 // a "lookahead," as we do below. 8286 if (inFreeRange()) lookahead_and_flush(fc, size); 8287 } else { 8288 // Code path common to both original and adaptive free lists. 8289 8290 // cant coalesce with previous block; this should be treated 8291 // as the end of a free run if any 8292 if (inFreeRange()) { 8293 // we kicked some butt; time to pick up the garbage 8294 assert(freeFinger() < addr, "freeFinger points too high"); 8295 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 8296 } 8297 // else, nothing to do, just continue 8298 } 8299 } 8300 8301 size_t SweepClosure::do_garbage_chunk(FreeChunk* fc) { 8302 // This is a chunk of garbage. It is not in any free list. 8303 // Add it to a free list or let it possibly be coalesced into 8304 // a larger chunk. 8305 HeapWord* const addr = (HeapWord*) fc; 8306 const size_t size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 8307 8308 if (_sp->adaptive_freelists()) { 8309 // Verify that the bit map has no bits marked between 8310 // addr and purported end of just dead object. 8311 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 8312 8313 do_post_free_or_garbage_chunk(fc, size); 8314 } else { 8315 if (!inFreeRange()) { 8316 // start of a new free range 8317 assert(size > 0, "A free range should have a size"); 8318 initialize_free_range(addr, false); 8319 } else { 8320 // this will be swept up when we hit the end of the 8321 // free range 8322 if (CMSTraceSweeper) { 8323 gclog_or_tty->print(" -- pick up garbage 0x%x (%d) \n", fc, size); 8324 } 8325 // If the chunk is being coalesced and the current free range is 8326 // in the free lists, remove the current free range so that it 8327 // will be returned to the free lists in its entirety - all 8328 // the coalesced pieces included. 8329 if (freeRangeInFreeLists()) { 8330 FreeChunk* ffc = (FreeChunk*)freeFinger(); 8331 assert(ffc->size() == pointer_delta(addr, freeFinger()), 8332 "Size of free range is inconsistent with chunk size."); 8333 if (CMSTestInFreeList) { 8334 assert(_sp->verify_chunk_in_free_list(ffc), 8335 "free range is not in free lists"); 8336 } 8337 _sp->removeFreeChunkFromFreeLists(ffc); 8338 set_freeRangeInFreeLists(false); 8339 } 8340 set_lastFreeRangeCoalesced(true); 8341 } 8342 // this will be swept up when we hit the end of the free range 8343 8344 // Verify that the bit map has no bits marked between 8345 // addr and purported end of just dead object. 8346 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 8347 } 8348 assert(_limit >= addr + size, 8349 "A freshly garbage chunk can't possibly straddle over _limit"); 8350 if (inFreeRange()) lookahead_and_flush(fc, size); 8351 return size; 8352 } 8353 8354 size_t SweepClosure::do_live_chunk(FreeChunk* fc) { 8355 HeapWord* addr = (HeapWord*) fc; 8356 // The sweeper has just found a live object. Return any accumulated 8357 // left hand chunk to the free lists. 8358 if (inFreeRange()) { 8359 assert(freeFinger() < addr, "freeFinger points too high"); 8360 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 8361 } 8362 8363 // This object is live: we'd normally expect this to be 8364 // an oop, and like to assert the following: 8365 // assert(oop(addr)->is_oop(), "live block should be an oop"); 8366 // However, as we commented above, this may be an object whose 8367 // header hasn't yet been initialized. 8368 size_t size; 8369 assert(_bitMap->isMarked(addr), "Tautology for this control point"); 8370 if (_bitMap->isMarked(addr + 1)) { 8371 // Determine the size from the bit map, rather than trying to 8372 // compute it from the object header. 8373 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 8374 size = pointer_delta(nextOneAddr + 1, addr); 8375 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 8376 "alignment problem"); 8377 8378 #ifdef DEBUG 8379 if (oop(addr)->klass_or_null() != NULL && 8380 ( !_collector->should_unload_classes() 8381 || (oop(addr)->is_parsable()) && 8382 oop(addr)->is_conc_safe())) { 8383 // Ignore mark word because we are running concurrent with mutators 8384 assert(oop(addr)->is_oop(true), "live block should be an oop"); 8385 // is_conc_safe is checked before performing this assertion 8386 // because an object that is not is_conc_safe may yet have 8387 // the return from size() correct. 8388 assert(size == 8389 CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()), 8390 "P-mark and computed size do not agree"); 8391 } 8392 #endif 8393 8394 } else { 8395 // This should be an initialized object that's alive. 8396 assert(oop(addr)->klass_or_null() != NULL && 8397 (!_collector->should_unload_classes() 8398 || oop(addr)->is_parsable()), 8399 "Should be an initialized object"); 8400 // Note that there are objects used during class redefinition, 8401 // e.g. merge_cp in VM_RedefineClasses::merge_cp_and_rewrite(), 8402 // which are discarded with their is_conc_safe state still 8403 // false. These object may be floating garbage so may be 8404 // seen here. If they are floating garbage their size 8405 // should be attainable from their klass. Do not that 8406 // is_conc_safe() is true for oop(addr). 8407 // Ignore mark word because we are running concurrent with mutators 8408 assert(oop(addr)->is_oop(true), "live block should be an oop"); 8409 // Verify that the bit map has no bits marked between 8410 // addr and purported end of this block. 8411 size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 8412 assert(size >= 3, "Necessary for Printezis marks to work"); 8413 assert(!_bitMap->isMarked(addr+1), "Tautology for this control point"); 8414 DEBUG_ONLY(_bitMap->verifyNoOneBitsInRange(addr+2, addr+size);) 8415 } 8416 return size; 8417 } 8418 8419 void SweepClosure::do_post_free_or_garbage_chunk(FreeChunk* fc, 8420 size_t chunkSize) { 8421 // do_post_free_or_garbage_chunk() should only be called in the case 8422 // of the adaptive free list allocator. 8423 const bool fcInFreeLists = fc->is_free(); 8424 assert(_sp->adaptive_freelists(), "Should only be used in this case."); 8425 assert((HeapWord*)fc <= _limit, "sweep invariant"); 8426 if (CMSTestInFreeList && fcInFreeLists) { 8427 assert(_sp->verify_chunk_in_free_list(fc), "free chunk is not in free lists"); 8428 } 8429 8430 if (CMSTraceSweeper) { 8431 gclog_or_tty->print_cr(" -- pick up another chunk at 0x%x (%d)", fc, chunkSize); 8432 } 8433 8434 HeapWord* const fc_addr = (HeapWord*) fc; 8435 8436 bool coalesce; 8437 const size_t left = pointer_delta(fc_addr, freeFinger()); 8438 const size_t right = chunkSize; 8439 switch (FLSCoalescePolicy) { 8440 // numeric value forms a coalition aggressiveness metric 8441 case 0: { // never coalesce 8442 coalesce = false; 8443 break; 8444 } 8445 case 1: { // coalesce if left & right chunks on overpopulated lists 8446 coalesce = _sp->coalOverPopulated(left) && 8447 _sp->coalOverPopulated(right); 8448 break; 8449 } 8450 case 2: { // coalesce if left chunk on overpopulated list (default) 8451 coalesce = _sp->coalOverPopulated(left); 8452 break; 8453 } 8454 case 3: { // coalesce if left OR right chunk on overpopulated list 8455 coalesce = _sp->coalOverPopulated(left) || 8456 _sp->coalOverPopulated(right); 8457 break; 8458 } 8459 case 4: { // always coalesce 8460 coalesce = true; 8461 break; 8462 } 8463 default: 8464 ShouldNotReachHere(); 8465 } 8466 8467 // Should the current free range be coalesced? 8468 // If the chunk is in a free range and either we decided to coalesce above 8469 // or the chunk is near the large block at the end of the heap 8470 // (isNearLargestChunk() returns true), then coalesce this chunk. 8471 const bool doCoalesce = inFreeRange() 8472 && (coalesce || _g->isNearLargestChunk(fc_addr)); 8473 if (doCoalesce) { 8474 // Coalesce the current free range on the left with the new 8475 // chunk on the right. If either is on a free list, 8476 // it must be removed from the list and stashed in the closure. 8477 if (freeRangeInFreeLists()) { 8478 FreeChunk* const ffc = (FreeChunk*)freeFinger(); 8479 assert(ffc->size() == pointer_delta(fc_addr, freeFinger()), 8480 "Size of free range is inconsistent with chunk size."); 8481 if (CMSTestInFreeList) { 8482 assert(_sp->verify_chunk_in_free_list(ffc), 8483 "Chunk is not in free lists"); 8484 } 8485 _sp->coalDeath(ffc->size()); 8486 _sp->removeFreeChunkFromFreeLists(ffc); 8487 set_freeRangeInFreeLists(false); 8488 } 8489 if (fcInFreeLists) { 8490 _sp->coalDeath(chunkSize); 8491 assert(fc->size() == chunkSize, 8492 "The chunk has the wrong size or is not in the free lists"); 8493 _sp->removeFreeChunkFromFreeLists(fc); 8494 } 8495 set_lastFreeRangeCoalesced(true); 8496 print_free_block_coalesced(fc); 8497 } else { // not in a free range and/or should not coalesce 8498 // Return the current free range and start a new one. 8499 if (inFreeRange()) { 8500 // In a free range but cannot coalesce with the right hand chunk. 8501 // Put the current free range into the free lists. 8502 flush_cur_free_chunk(freeFinger(), 8503 pointer_delta(fc_addr, freeFinger())); 8504 } 8505 // Set up for new free range. Pass along whether the right hand 8506 // chunk is in the free lists. 8507 initialize_free_range((HeapWord*)fc, fcInFreeLists); 8508 } 8509 } 8510 8511 // Lookahead flush: 8512 // If we are tracking a free range, and this is the last chunk that 8513 // we'll look at because its end crosses past _limit, we'll preemptively 8514 // flush it along with any free range we may be holding on to. Note that 8515 // this can be the case only for an already free or freshly garbage 8516 // chunk. If this block is an object, it can never straddle 8517 // over _limit. The "straddling" occurs when _limit is set at 8518 // the previous end of the space when this cycle started, and 8519 // a subsequent heap expansion caused the previously co-terminal 8520 // free block to be coalesced with the newly expanded portion, 8521 // thus rendering _limit a non-block-boundary making it dangerous 8522 // for the sweeper to step over and examine. 8523 void SweepClosure::lookahead_and_flush(FreeChunk* fc, size_t chunk_size) { 8524 assert(inFreeRange(), "Should only be called if currently in a free range."); 8525 HeapWord* const eob = ((HeapWord*)fc) + chunk_size; 8526 assert(_sp->used_region().contains(eob - 1), 8527 err_msg("eob = " PTR_FORMAT " out of bounds wrt _sp = [" PTR_FORMAT "," PTR_FORMAT ")" 8528 " when examining fc = " PTR_FORMAT "(" SIZE_FORMAT ")", 8529 _limit, _sp->bottom(), _sp->end(), fc, chunk_size)); 8530 if (eob >= _limit) { 8531 assert(eob == _limit || fc->is_free(), "Only a free chunk should allow us to cross over the limit"); 8532 if (CMSTraceSweeper) { 8533 gclog_or_tty->print_cr("_limit " PTR_FORMAT " reached or crossed by block " 8534 "[" PTR_FORMAT "," PTR_FORMAT ") in space " 8535 "[" PTR_FORMAT "," PTR_FORMAT ")", 8536 _limit, fc, eob, _sp->bottom(), _sp->end()); 8537 } 8538 // Return the storage we are tracking back into the free lists. 8539 if (CMSTraceSweeper) { 8540 gclog_or_tty->print_cr("Flushing ... "); 8541 } 8542 assert(freeFinger() < eob, "Error"); 8543 flush_cur_free_chunk( freeFinger(), pointer_delta(eob, freeFinger())); 8544 } 8545 } 8546 8547 void SweepClosure::flush_cur_free_chunk(HeapWord* chunk, size_t size) { 8548 assert(inFreeRange(), "Should only be called if currently in a free range."); 8549 assert(size > 0, 8550 "A zero sized chunk cannot be added to the free lists."); 8551 if (!freeRangeInFreeLists()) { 8552 if (CMSTestInFreeList) { 8553 FreeChunk* fc = (FreeChunk*) chunk; 8554 fc->set_size(size); 8555 assert(!_sp->verify_chunk_in_free_list(fc), 8556 "chunk should not be in free lists yet"); 8557 } 8558 if (CMSTraceSweeper) { 8559 gclog_or_tty->print_cr(" -- add free block 0x%x (%d) to free lists", 8560 chunk, size); 8561 } 8562 // A new free range is going to be starting. The current 8563 // free range has not been added to the free lists yet or 8564 // was removed so add it back. 8565 // If the current free range was coalesced, then the death 8566 // of the free range was recorded. Record a birth now. 8567 if (lastFreeRangeCoalesced()) { 8568 _sp->coalBirth(size); 8569 } 8570 _sp->addChunkAndRepairOffsetTable(chunk, size, 8571 lastFreeRangeCoalesced()); 8572 } else if (CMSTraceSweeper) { 8573 gclog_or_tty->print_cr("Already in free list: nothing to flush"); 8574 } 8575 set_inFreeRange(false); 8576 set_freeRangeInFreeLists(false); 8577 } 8578 8579 // We take a break if we've been at this for a while, 8580 // so as to avoid monopolizing the locks involved. 8581 void SweepClosure::do_yield_work(HeapWord* addr) { 8582 // Return current free chunk being used for coalescing (if any) 8583 // to the appropriate freelist. After yielding, the next 8584 // free block encountered will start a coalescing range of 8585 // free blocks. If the next free block is adjacent to the 8586 // chunk just flushed, they will need to wait for the next 8587 // sweep to be coalesced. 8588 if (inFreeRange()) { 8589 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 8590 } 8591 8592 // First give up the locks, then yield, then re-lock. 8593 // We should probably use a constructor/destructor idiom to 8594 // do this unlock/lock or modify the MutexUnlocker class to 8595 // serve our purpose. XXX 8596 assert_lock_strong(_bitMap->lock()); 8597 assert_lock_strong(_freelistLock); 8598 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 8599 "CMS thread should hold CMS token"); 8600 _bitMap->lock()->unlock(); 8601 _freelistLock->unlock(); 8602 ConcurrentMarkSweepThread::desynchronize(true); 8603 ConcurrentMarkSweepThread::acknowledge_yield_request(); 8604 _collector->stopTimer(); 8605 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 8606 if (PrintCMSStatistics != 0) { 8607 _collector->incrementYields(); 8608 } 8609 _collector->icms_wait(); 8610 8611 // See the comment in coordinator_yield() 8612 for (unsigned i = 0; i < CMSYieldSleepCount && 8613 ConcurrentMarkSweepThread::should_yield() && 8614 !CMSCollector::foregroundGCIsActive(); ++i) { 8615 os::sleep(Thread::current(), 1, false); 8616 ConcurrentMarkSweepThread::acknowledge_yield_request(); 8617 } 8618 8619 ConcurrentMarkSweepThread::synchronize(true); 8620 _freelistLock->lock(); 8621 _bitMap->lock()->lock_without_safepoint_check(); 8622 _collector->startTimer(); 8623 } 8624 8625 #ifndef PRODUCT 8626 // This is actually very useful in a product build if it can 8627 // be called from the debugger. Compile it into the product 8628 // as needed. 8629 bool debug_verify_chunk_in_free_list(FreeChunk* fc) { 8630 return debug_cms_space->verify_chunk_in_free_list(fc); 8631 } 8632 #endif 8633 8634 void SweepClosure::print_free_block_coalesced(FreeChunk* fc) const { 8635 if (CMSTraceSweeper) { 8636 gclog_or_tty->print_cr("Sweep:coal_free_blk " PTR_FORMAT " (" SIZE_FORMAT ")", 8637 fc, fc->size()); 8638 } 8639 } 8640 8641 // CMSIsAliveClosure 8642 bool CMSIsAliveClosure::do_object_b(oop obj) { 8643 HeapWord* addr = (HeapWord*)obj; 8644 return addr != NULL && 8645 (!_span.contains(addr) || _bit_map->isMarked(addr)); 8646 } 8647 8648 CMSKeepAliveClosure::CMSKeepAliveClosure( CMSCollector* collector, 8649 MemRegion span, 8650 CMSBitMap* bit_map, CMSMarkStack* mark_stack, 8651 CMSMarkStack* revisit_stack, bool cpc): 8652 KlassRememberingOopClosure(collector, NULL, revisit_stack), 8653 _span(span), 8654 _bit_map(bit_map), 8655 _mark_stack(mark_stack), 8656 _concurrent_precleaning(cpc) { 8657 assert(!_span.is_empty(), "Empty span could spell trouble"); 8658 } 8659 8660 8661 // CMSKeepAliveClosure: the serial version 8662 void CMSKeepAliveClosure::do_oop(oop obj) { 8663 HeapWord* addr = (HeapWord*)obj; 8664 if (_span.contains(addr) && 8665 !_bit_map->isMarked(addr)) { 8666 _bit_map->mark(addr); 8667 bool simulate_overflow = false; 8668 NOT_PRODUCT( 8669 if (CMSMarkStackOverflowALot && 8670 _collector->simulate_overflow()) { 8671 // simulate a stack overflow 8672 simulate_overflow = true; 8673 } 8674 ) 8675 if (simulate_overflow || !_mark_stack->push(obj)) { 8676 if (_concurrent_precleaning) { 8677 // We dirty the overflown object and let the remark 8678 // phase deal with it. 8679 assert(_collector->overflow_list_is_empty(), "Error"); 8680 // In the case of object arrays, we need to dirty all of 8681 // the cards that the object spans. No locking or atomics 8682 // are needed since no one else can be mutating the mod union 8683 // table. 8684 if (obj->is_objArray()) { 8685 size_t sz = obj->size(); 8686 HeapWord* end_card_addr = 8687 (HeapWord*)round_to((intptr_t)(addr+sz), CardTableModRefBS::card_size); 8688 MemRegion redirty_range = MemRegion(addr, end_card_addr); 8689 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 8690 _collector->_modUnionTable.mark_range(redirty_range); 8691 } else { 8692 _collector->_modUnionTable.mark(addr); 8693 } 8694 _collector->_ser_kac_preclean_ovflw++; 8695 } else { 8696 _collector->push_on_overflow_list(obj); 8697 _collector->_ser_kac_ovflw++; 8698 } 8699 } 8700 } 8701 } 8702 8703 void CMSKeepAliveClosure::do_oop(oop* p) { CMSKeepAliveClosure::do_oop_work(p); } 8704 void CMSKeepAliveClosure::do_oop(narrowOop* p) { CMSKeepAliveClosure::do_oop_work(p); } 8705 8706 // CMSParKeepAliveClosure: a parallel version of the above. 8707 // The work queues are private to each closure (thread), 8708 // but (may be) available for stealing by other threads. 8709 void CMSParKeepAliveClosure::do_oop(oop obj) { 8710 HeapWord* addr = (HeapWord*)obj; 8711 if (_span.contains(addr) && 8712 !_bit_map->isMarked(addr)) { 8713 // In general, during recursive tracing, several threads 8714 // may be concurrently getting here; the first one to 8715 // "tag" it, claims it. 8716 if (_bit_map->par_mark(addr)) { 8717 bool res = _work_queue->push(obj); 8718 assert(res, "Low water mark should be much less than capacity"); 8719 // Do a recursive trim in the hope that this will keep 8720 // stack usage lower, but leave some oops for potential stealers 8721 trim_queue(_low_water_mark); 8722 } // Else, another thread got there first 8723 } 8724 } 8725 8726 void CMSParKeepAliveClosure::do_oop(oop* p) { CMSParKeepAliveClosure::do_oop_work(p); } 8727 void CMSParKeepAliveClosure::do_oop(narrowOop* p) { CMSParKeepAliveClosure::do_oop_work(p); } 8728 8729 void CMSParKeepAliveClosure::trim_queue(uint max) { 8730 while (_work_queue->size() > max) { 8731 oop new_oop; 8732 if (_work_queue->pop_local(new_oop)) { 8733 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 8734 assert(_bit_map->isMarked((HeapWord*)new_oop), 8735 "no white objects on this stack!"); 8736 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 8737 // iterate over the oops in this oop, marking and pushing 8738 // the ones in CMS heap (i.e. in _span). 8739 new_oop->oop_iterate(&_mark_and_push); 8740 } 8741 } 8742 } 8743 8744 CMSInnerParMarkAndPushClosure::CMSInnerParMarkAndPushClosure( 8745 CMSCollector* collector, 8746 MemRegion span, CMSBitMap* bit_map, 8747 CMSMarkStack* revisit_stack, 8748 OopTaskQueue* work_queue): 8749 Par_KlassRememberingOopClosure(collector, NULL, revisit_stack), 8750 _span(span), 8751 _bit_map(bit_map), 8752 _work_queue(work_queue) { } 8753 8754 void CMSInnerParMarkAndPushClosure::do_oop(oop obj) { 8755 HeapWord* addr = (HeapWord*)obj; 8756 if (_span.contains(addr) && 8757 !_bit_map->isMarked(addr)) { 8758 if (_bit_map->par_mark(addr)) { 8759 bool simulate_overflow = false; 8760 NOT_PRODUCT( 8761 if (CMSMarkStackOverflowALot && 8762 _collector->par_simulate_overflow()) { 8763 // simulate a stack overflow 8764 simulate_overflow = true; 8765 } 8766 ) 8767 if (simulate_overflow || !_work_queue->push(obj)) { 8768 _collector->par_push_on_overflow_list(obj); 8769 _collector->_par_kac_ovflw++; 8770 } 8771 } // Else another thread got there already 8772 } 8773 } 8774 8775 void CMSInnerParMarkAndPushClosure::do_oop(oop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); } 8776 void CMSInnerParMarkAndPushClosure::do_oop(narrowOop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); } 8777 8778 ////////////////////////////////////////////////////////////////// 8779 // CMSExpansionCause ///////////////////////////// 8780 ////////////////////////////////////////////////////////////////// 8781 const char* CMSExpansionCause::to_string(CMSExpansionCause::Cause cause) { 8782 switch (cause) { 8783 case _no_expansion: 8784 return "No expansion"; 8785 case _satisfy_free_ratio: 8786 return "Free ratio"; 8787 case _satisfy_promotion: 8788 return "Satisfy promotion"; 8789 case _satisfy_allocation: 8790 return "allocation"; 8791 case _allocate_par_lab: 8792 return "Par LAB"; 8793 case _allocate_par_spooling_space: 8794 return "Par Spooling Space"; 8795 case _adaptive_size_policy: 8796 return "Ergonomics"; 8797 default: 8798 return "unknown"; 8799 } 8800 } 8801 8802 void CMSDrainMarkingStackClosure::do_void() { 8803 // the max number to take from overflow list at a time 8804 const size_t num = _mark_stack->capacity()/4; 8805 assert(!_concurrent_precleaning || _collector->overflow_list_is_empty(), 8806 "Overflow list should be NULL during concurrent phases"); 8807 while (!_mark_stack->isEmpty() || 8808 // if stack is empty, check the overflow list 8809 _collector->take_from_overflow_list(num, _mark_stack)) { 8810 oop obj = _mark_stack->pop(); 8811 HeapWord* addr = (HeapWord*)obj; 8812 assert(_span.contains(addr), "Should be within span"); 8813 assert(_bit_map->isMarked(addr), "Should be marked"); 8814 assert(obj->is_oop(), "Should be an oop"); 8815 obj->oop_iterate(_keep_alive); 8816 } 8817 } 8818 8819 void CMSParDrainMarkingStackClosure::do_void() { 8820 // drain queue 8821 trim_queue(0); 8822 } 8823 8824 // Trim our work_queue so its length is below max at return 8825 void CMSParDrainMarkingStackClosure::trim_queue(uint max) { 8826 while (_work_queue->size() > max) { 8827 oop new_oop; 8828 if (_work_queue->pop_local(new_oop)) { 8829 assert(new_oop->is_oop(), "Expected an oop"); 8830 assert(_bit_map->isMarked((HeapWord*)new_oop), 8831 "no white objects on this stack!"); 8832 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 8833 // iterate over the oops in this oop, marking and pushing 8834 // the ones in CMS heap (i.e. in _span). 8835 new_oop->oop_iterate(&_mark_and_push); 8836 } 8837 } 8838 } 8839 8840 //////////////////////////////////////////////////////////////////// 8841 // Support for Marking Stack Overflow list handling and related code 8842 //////////////////////////////////////////////////////////////////// 8843 // Much of the following code is similar in shape and spirit to the 8844 // code used in ParNewGC. We should try and share that code 8845 // as much as possible in the future. 8846 8847 #ifndef PRODUCT 8848 // Debugging support for CMSStackOverflowALot 8849 8850 // It's OK to call this multi-threaded; the worst thing 8851 // that can happen is that we'll get a bunch of closely 8852 // spaced simulated oveflows, but that's OK, in fact 8853 // probably good as it would exercise the overflow code 8854 // under contention. 8855 bool CMSCollector::simulate_overflow() { 8856 if (_overflow_counter-- <= 0) { // just being defensive 8857 _overflow_counter = CMSMarkStackOverflowInterval; 8858 return true; 8859 } else { 8860 return false; 8861 } 8862 } 8863 8864 bool CMSCollector::par_simulate_overflow() { 8865 return simulate_overflow(); 8866 } 8867 #endif 8868 8869 // Single-threaded 8870 bool CMSCollector::take_from_overflow_list(size_t num, CMSMarkStack* stack) { 8871 assert(stack->isEmpty(), "Expected precondition"); 8872 assert(stack->capacity() > num, "Shouldn't bite more than can chew"); 8873 size_t i = num; 8874 oop cur = _overflow_list; 8875 const markOop proto = markOopDesc::prototype(); 8876 NOT_PRODUCT(ssize_t n = 0;) 8877 for (oop next; i > 0 && cur != NULL; cur = next, i--) { 8878 next = oop(cur->mark()); 8879 cur->set_mark(proto); // until proven otherwise 8880 assert(cur->is_oop(), "Should be an oop"); 8881 bool res = stack->push(cur); 8882 assert(res, "Bit off more than can chew?"); 8883 NOT_PRODUCT(n++;) 8884 } 8885 _overflow_list = cur; 8886 #ifndef PRODUCT 8887 assert(_num_par_pushes >= n, "Too many pops?"); 8888 _num_par_pushes -=n; 8889 #endif 8890 return !stack->isEmpty(); 8891 } 8892 8893 #define BUSY (oop(0x1aff1aff)) 8894 // (MT-safe) Get a prefix of at most "num" from the list. 8895 // The overflow list is chained through the mark word of 8896 // each object in the list. We fetch the entire list, 8897 // break off a prefix of the right size and return the 8898 // remainder. If other threads try to take objects from 8899 // the overflow list at that time, they will wait for 8900 // some time to see if data becomes available. If (and 8901 // only if) another thread places one or more object(s) 8902 // on the global list before we have returned the suffix 8903 // to the global list, we will walk down our local list 8904 // to find its end and append the global list to 8905 // our suffix before returning it. This suffix walk can 8906 // prove to be expensive (quadratic in the amount of traffic) 8907 // when there are many objects in the overflow list and 8908 // there is much producer-consumer contention on the list. 8909 // *NOTE*: The overflow list manipulation code here and 8910 // in ParNewGeneration:: are very similar in shape, 8911 // except that in the ParNew case we use the old (from/eden) 8912 // copy of the object to thread the list via its klass word. 8913 // Because of the common code, if you make any changes in 8914 // the code below, please check the ParNew version to see if 8915 // similar changes might be needed. 8916 // CR 6797058 has been filed to consolidate the common code. 8917 bool CMSCollector::par_take_from_overflow_list(size_t num, 8918 OopTaskQueue* work_q, 8919 int no_of_gc_threads) { 8920 assert(work_q->size() == 0, "First empty local work queue"); 8921 assert(num < work_q->max_elems(), "Can't bite more than we can chew"); 8922 if (_overflow_list == NULL) { 8923 return false; 8924 } 8925 // Grab the entire list; we'll put back a suffix 8926 oop prefix = (oop)Atomic::xchg_ptr(BUSY, &_overflow_list); 8927 Thread* tid = Thread::current(); 8928 // Before "no_of_gc_threads" was introduced CMSOverflowSpinCount was 8929 // set to ParallelGCThreads. 8930 size_t CMSOverflowSpinCount = (size_t) no_of_gc_threads; // was ParallelGCThreads; 8931 size_t sleep_time_millis = MAX2((size_t)1, num/100); 8932 // If the list is busy, we spin for a short while, 8933 // sleeping between attempts to get the list. 8934 for (size_t spin = 0; prefix == BUSY && spin < CMSOverflowSpinCount; spin++) { 8935 os::sleep(tid, sleep_time_millis, false); 8936 if (_overflow_list == NULL) { 8937 // Nothing left to take 8938 return false; 8939 } else if (_overflow_list != BUSY) { 8940 // Try and grab the prefix 8941 prefix = (oop)Atomic::xchg_ptr(BUSY, &_overflow_list); 8942 } 8943 } 8944 // If the list was found to be empty, or we spun long 8945 // enough, we give up and return empty-handed. If we leave 8946 // the list in the BUSY state below, it must be the case that 8947 // some other thread holds the overflow list and will set it 8948 // to a non-BUSY state in the future. 8949 if (prefix == NULL || prefix == BUSY) { 8950 // Nothing to take or waited long enough 8951 if (prefix == NULL) { 8952 // Write back the NULL in case we overwrote it with BUSY above 8953 // and it is still the same value. 8954 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY); 8955 } 8956 return false; 8957 } 8958 assert(prefix != NULL && prefix != BUSY, "Error"); 8959 size_t i = num; 8960 oop cur = prefix; 8961 // Walk down the first "num" objects, unless we reach the end. 8962 for (; i > 1 && cur->mark() != NULL; cur = oop(cur->mark()), i--); 8963 if (cur->mark() == NULL) { 8964 // We have "num" or fewer elements in the list, so there 8965 // is nothing to return to the global list. 8966 // Write back the NULL in lieu of the BUSY we wrote 8967 // above, if it is still the same value. 8968 if (_overflow_list == BUSY) { 8969 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY); 8970 } 8971 } else { 8972 // Chop off the suffix and rerturn it to the global list. 8973 assert(cur->mark() != BUSY, "Error"); 8974 oop suffix_head = cur->mark(); // suffix will be put back on global list 8975 cur->set_mark(NULL); // break off suffix 8976 // It's possible that the list is still in the empty(busy) state 8977 // we left it in a short while ago; in that case we may be 8978 // able to place back the suffix without incurring the cost 8979 // of a walk down the list. 8980 oop observed_overflow_list = _overflow_list; 8981 oop cur_overflow_list = observed_overflow_list; 8982 bool attached = false; 8983 while (observed_overflow_list == BUSY || observed_overflow_list == NULL) { 8984 observed_overflow_list = 8985 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list); 8986 if (cur_overflow_list == observed_overflow_list) { 8987 attached = true; 8988 break; 8989 } else cur_overflow_list = observed_overflow_list; 8990 } 8991 if (!attached) { 8992 // Too bad, someone else sneaked in (at least) an element; we'll need 8993 // to do a splice. Find tail of suffix so we can prepend suffix to global 8994 // list. 8995 for (cur = suffix_head; cur->mark() != NULL; cur = (oop)(cur->mark())); 8996 oop suffix_tail = cur; 8997 assert(suffix_tail != NULL && suffix_tail->mark() == NULL, 8998 "Tautology"); 8999 observed_overflow_list = _overflow_list; 9000 do { 9001 cur_overflow_list = observed_overflow_list; 9002 if (cur_overflow_list != BUSY) { 9003 // Do the splice ... 9004 suffix_tail->set_mark(markOop(cur_overflow_list)); 9005 } else { // cur_overflow_list == BUSY 9006 suffix_tail->set_mark(NULL); 9007 } 9008 // ... and try to place spliced list back on overflow_list ... 9009 observed_overflow_list = 9010 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list); 9011 } while (cur_overflow_list != observed_overflow_list); 9012 // ... until we have succeeded in doing so. 9013 } 9014 } 9015 9016 // Push the prefix elements on work_q 9017 assert(prefix != NULL, "control point invariant"); 9018 const markOop proto = markOopDesc::prototype(); 9019 oop next; 9020 NOT_PRODUCT(ssize_t n = 0;) 9021 for (cur = prefix; cur != NULL; cur = next) { 9022 next = oop(cur->mark()); 9023 cur->set_mark(proto); // until proven otherwise 9024 assert(cur->is_oop(), "Should be an oop"); 9025 bool res = work_q->push(cur); 9026 assert(res, "Bit off more than we can chew?"); 9027 NOT_PRODUCT(n++;) 9028 } 9029 #ifndef PRODUCT 9030 assert(_num_par_pushes >= n, "Too many pops?"); 9031 Atomic::add_ptr(-(intptr_t)n, &_num_par_pushes); 9032 #endif 9033 return true; 9034 } 9035 9036 // Single-threaded 9037 void CMSCollector::push_on_overflow_list(oop p) { 9038 NOT_PRODUCT(_num_par_pushes++;) 9039 assert(p->is_oop(), "Not an oop"); 9040 preserve_mark_if_necessary(p); 9041 p->set_mark((markOop)_overflow_list); 9042 _overflow_list = p; 9043 } 9044 9045 // Multi-threaded; use CAS to prepend to overflow list 9046 void CMSCollector::par_push_on_overflow_list(oop p) { 9047 NOT_PRODUCT(Atomic::inc_ptr(&_num_par_pushes);) 9048 assert(p->is_oop(), "Not an oop"); 9049 par_preserve_mark_if_necessary(p); 9050 oop observed_overflow_list = _overflow_list; 9051 oop cur_overflow_list; 9052 do { 9053 cur_overflow_list = observed_overflow_list; 9054 if (cur_overflow_list != BUSY) { 9055 p->set_mark(markOop(cur_overflow_list)); 9056 } else { 9057 p->set_mark(NULL); 9058 } 9059 observed_overflow_list = 9060 (oop) Atomic::cmpxchg_ptr(p, &_overflow_list, cur_overflow_list); 9061 } while (cur_overflow_list != observed_overflow_list); 9062 } 9063 #undef BUSY 9064 9065 // Single threaded 9066 // General Note on GrowableArray: pushes may silently fail 9067 // because we are (temporarily) out of C-heap for expanding 9068 // the stack. The problem is quite ubiquitous and affects 9069 // a lot of code in the JVM. The prudent thing for GrowableArray 9070 // to do (for now) is to exit with an error. However, that may 9071 // be too draconian in some cases because the caller may be 9072 // able to recover without much harm. For such cases, we 9073 // should probably introduce a "soft_push" method which returns 9074 // an indication of success or failure with the assumption that 9075 // the caller may be able to recover from a failure; code in 9076 // the VM can then be changed, incrementally, to deal with such 9077 // failures where possible, thus, incrementally hardening the VM 9078 // in such low resource situations. 9079 void CMSCollector::preserve_mark_work(oop p, markOop m) { 9080 _preserved_oop_stack.push(p); 9081 _preserved_mark_stack.push(m); 9082 assert(m == p->mark(), "Mark word changed"); 9083 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), 9084 "bijection"); 9085 } 9086 9087 // Single threaded 9088 void CMSCollector::preserve_mark_if_necessary(oop p) { 9089 markOop m = p->mark(); 9090 if (m->must_be_preserved(p)) { 9091 preserve_mark_work(p, m); 9092 } 9093 } 9094 9095 void CMSCollector::par_preserve_mark_if_necessary(oop p) { 9096 markOop m = p->mark(); 9097 if (m->must_be_preserved(p)) { 9098 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag); 9099 // Even though we read the mark word without holding 9100 // the lock, we are assured that it will not change 9101 // because we "own" this oop, so no other thread can 9102 // be trying to push it on the overflow list; see 9103 // the assertion in preserve_mark_work() that checks 9104 // that m == p->mark(). 9105 preserve_mark_work(p, m); 9106 } 9107 } 9108 9109 // We should be able to do this multi-threaded, 9110 // a chunk of stack being a task (this is 9111 // correct because each oop only ever appears 9112 // once in the overflow list. However, it's 9113 // not very easy to completely overlap this with 9114 // other operations, so will generally not be done 9115 // until all work's been completed. Because we 9116 // expect the preserved oop stack (set) to be small, 9117 // it's probably fine to do this single-threaded. 9118 // We can explore cleverer concurrent/overlapped/parallel 9119 // processing of preserved marks if we feel the 9120 // need for this in the future. Stack overflow should 9121 // be so rare in practice and, when it happens, its 9122 // effect on performance so great that this will 9123 // likely just be in the noise anyway. 9124 void CMSCollector::restore_preserved_marks_if_any() { 9125 assert(SafepointSynchronize::is_at_safepoint(), 9126 "world should be stopped"); 9127 assert(Thread::current()->is_ConcurrentGC_thread() || 9128 Thread::current()->is_VM_thread(), 9129 "should be single-threaded"); 9130 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), 9131 "bijection"); 9132 9133 while (!_preserved_oop_stack.is_empty()) { 9134 oop p = _preserved_oop_stack.pop(); 9135 assert(p->is_oop(), "Should be an oop"); 9136 assert(_span.contains(p), "oop should be in _span"); 9137 assert(p->mark() == markOopDesc::prototype(), 9138 "Set when taken from overflow list"); 9139 markOop m = _preserved_mark_stack.pop(); 9140 p->set_mark(m); 9141 } 9142 assert(_preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(), 9143 "stacks were cleared above"); 9144 } 9145 9146 #ifndef PRODUCT 9147 bool CMSCollector::no_preserved_marks() const { 9148 return _preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(); 9149 } 9150 #endif 9151 9152 CMSAdaptiveSizePolicy* ASConcurrentMarkSweepGeneration::cms_size_policy() const 9153 { 9154 GenCollectedHeap* gch = (GenCollectedHeap*) GenCollectedHeap::heap(); 9155 CMSAdaptiveSizePolicy* size_policy = 9156 (CMSAdaptiveSizePolicy*) gch->gen_policy()->size_policy(); 9157 assert(size_policy->is_gc_cms_adaptive_size_policy(), 9158 "Wrong type for size policy"); 9159 return size_policy; 9160 } 9161 9162 void ASConcurrentMarkSweepGeneration::resize(size_t cur_promo_size, 9163 size_t desired_promo_size) { 9164 if (cur_promo_size < desired_promo_size) { 9165 size_t expand_bytes = desired_promo_size - cur_promo_size; 9166 if (PrintAdaptiveSizePolicy && Verbose) { 9167 gclog_or_tty->print_cr(" ASConcurrentMarkSweepGeneration::resize " 9168 "Expanding tenured generation by " SIZE_FORMAT " (bytes)", 9169 expand_bytes); 9170 } 9171 expand(expand_bytes, 9172 MinHeapDeltaBytes, 9173 CMSExpansionCause::_adaptive_size_policy); 9174 } else if (desired_promo_size < cur_promo_size) { 9175 size_t shrink_bytes = cur_promo_size - desired_promo_size; 9176 if (PrintAdaptiveSizePolicy && Verbose) { 9177 gclog_or_tty->print_cr(" ASConcurrentMarkSweepGeneration::resize " 9178 "Shrinking tenured generation by " SIZE_FORMAT " (bytes)", 9179 shrink_bytes); 9180 } 9181 shrink(shrink_bytes); 9182 } 9183 } 9184 9185 CMSGCAdaptivePolicyCounters* ASConcurrentMarkSweepGeneration::gc_adaptive_policy_counters() { 9186 GenCollectedHeap* gch = GenCollectedHeap::heap(); 9187 CMSGCAdaptivePolicyCounters* counters = 9188 (CMSGCAdaptivePolicyCounters*) gch->collector_policy()->counters(); 9189 assert(counters->kind() == GCPolicyCounters::CMSGCAdaptivePolicyCountersKind, 9190 "Wrong kind of counters"); 9191 return counters; 9192 } 9193 9194 9195 void ASConcurrentMarkSweepGeneration::update_counters() { 9196 if (UsePerfData) { 9197 _space_counters->update_all(); 9198 _gen_counters->update_all(); 9199 CMSGCAdaptivePolicyCounters* counters = gc_adaptive_policy_counters(); 9200 GenCollectedHeap* gch = GenCollectedHeap::heap(); 9201 CMSGCStats* gc_stats_l = (CMSGCStats*) gc_stats(); 9202 assert(gc_stats_l->kind() == GCStats::CMSGCStatsKind, 9203 "Wrong gc statistics type"); 9204 counters->update_counters(gc_stats_l); 9205 } 9206 } 9207 9208 void ASConcurrentMarkSweepGeneration::update_counters(size_t used) { 9209 if (UsePerfData) { 9210 _space_counters->update_used(used); 9211 _space_counters->update_capacity(); 9212 _gen_counters->update_all(); 9213 9214 CMSGCAdaptivePolicyCounters* counters = gc_adaptive_policy_counters(); 9215 GenCollectedHeap* gch = GenCollectedHeap::heap(); 9216 CMSGCStats* gc_stats_l = (CMSGCStats*) gc_stats(); 9217 assert(gc_stats_l->kind() == GCStats::CMSGCStatsKind, 9218 "Wrong gc statistics type"); 9219 counters->update_counters(gc_stats_l); 9220 } 9221 } 9222 9223 // The desired expansion delta is computed so that: 9224 // . desired free percentage or greater is used 9225 void ASConcurrentMarkSweepGeneration::compute_new_size() { 9226 assert_locked_or_safepoint(Heap_lock); 9227 9228 GenCollectedHeap* gch = (GenCollectedHeap*) GenCollectedHeap::heap(); 9229 9230 // If incremental collection failed, we just want to expand 9231 // to the limit. 9232 if (incremental_collection_failed()) { 9233 clear_incremental_collection_failed(); 9234 grow_to_reserved(); 9235 return; 9236 } 9237 9238 assert(UseAdaptiveSizePolicy, "Should be using adaptive sizing"); 9239 9240 assert(gch->kind() == CollectedHeap::GenCollectedHeap, 9241 "Wrong type of heap"); 9242 int prev_level = level() - 1; 9243 assert(prev_level >= 0, "The cms generation is the lowest generation"); 9244 Generation* prev_gen = gch->get_gen(prev_level); 9245 assert(prev_gen->kind() == Generation::ASParNew, 9246 "Wrong type of young generation"); 9247 ParNewGeneration* younger_gen = (ParNewGeneration*) prev_gen; 9248 size_t cur_eden = younger_gen->eden()->capacity(); 9249 CMSAdaptiveSizePolicy* size_policy = cms_size_policy(); 9250 size_t cur_promo = free(); 9251 size_policy->compute_tenured_generation_free_space(cur_promo, 9252 max_available(), 9253 cur_eden); 9254 resize(cur_promo, size_policy->promo_size()); 9255 9256 // Record the new size of the space in the cms generation 9257 // that is available for promotions. This is temporary. 9258 // It should be the desired promo size. 9259 size_policy->avg_cms_promo()->sample(free()); 9260 size_policy->avg_old_live()->sample(used()); 9261 9262 if (UsePerfData) { 9263 CMSGCAdaptivePolicyCounters* counters = gc_adaptive_policy_counters(); 9264 counters->update_cms_capacity_counter(capacity()); 9265 } 9266 } 9267 9268 void ASConcurrentMarkSweepGeneration::shrink_by(size_t desired_bytes) { 9269 assert_locked_or_safepoint(Heap_lock); 9270 assert_lock_strong(freelistLock()); 9271 HeapWord* old_end = _cmsSpace->end(); 9272 HeapWord* unallocated_start = _cmsSpace->unallocated_block(); 9273 assert(old_end >= unallocated_start, "Miscalculation of unallocated_start"); 9274 FreeChunk* chunk_at_end = find_chunk_at_end(); 9275 if (chunk_at_end == NULL) { 9276 // No room to shrink 9277 if (PrintGCDetails && Verbose) { 9278 gclog_or_tty->print_cr("No room to shrink: old_end " 9279 PTR_FORMAT " unallocated_start " PTR_FORMAT 9280 " chunk_at_end " PTR_FORMAT, 9281 old_end, unallocated_start, chunk_at_end); 9282 } 9283 return; 9284 } else { 9285 9286 // Find the chunk at the end of the space and determine 9287 // how much it can be shrunk. 9288 size_t shrinkable_size_in_bytes = chunk_at_end->size(); 9289 size_t aligned_shrinkable_size_in_bytes = 9290 align_size_down(shrinkable_size_in_bytes, os::vm_page_size()); 9291 assert(unallocated_start <= chunk_at_end->end(), 9292 "Inconsistent chunk at end of space"); 9293 size_t bytes = MIN2(desired_bytes, aligned_shrinkable_size_in_bytes); 9294 size_t word_size_before = heap_word_size(_virtual_space.committed_size()); 9295 9296 // Shrink the underlying space 9297 _virtual_space.shrink_by(bytes); 9298 if (PrintGCDetails && Verbose) { 9299 gclog_or_tty->print_cr("ConcurrentMarkSweepGeneration::shrink_by:" 9300 " desired_bytes " SIZE_FORMAT 9301 " shrinkable_size_in_bytes " SIZE_FORMAT 9302 " aligned_shrinkable_size_in_bytes " SIZE_FORMAT 9303 " bytes " SIZE_FORMAT, 9304 desired_bytes, shrinkable_size_in_bytes, 9305 aligned_shrinkable_size_in_bytes, bytes); 9306 gclog_or_tty->print_cr(" old_end " SIZE_FORMAT 9307 " unallocated_start " SIZE_FORMAT, 9308 old_end, unallocated_start); 9309 } 9310 9311 // If the space did shrink (shrinking is not guaranteed), 9312 // shrink the chunk at the end by the appropriate amount. 9313 if (((HeapWord*)_virtual_space.high()) < old_end) { 9314 size_t new_word_size = 9315 heap_word_size(_virtual_space.committed_size()); 9316 9317 // Have to remove the chunk from the dictionary because it is changing 9318 // size and might be someplace elsewhere in the dictionary. 9319 9320 // Get the chunk at end, shrink it, and put it 9321 // back. 9322 _cmsSpace->removeChunkFromDictionary(chunk_at_end); 9323 size_t word_size_change = word_size_before - new_word_size; 9324 size_t chunk_at_end_old_size = chunk_at_end->size(); 9325 assert(chunk_at_end_old_size >= word_size_change, 9326 "Shrink is too large"); 9327 chunk_at_end->set_size(chunk_at_end_old_size - 9328 word_size_change); 9329 _cmsSpace->freed((HeapWord*) chunk_at_end->end(), 9330 word_size_change); 9331 9332 _cmsSpace->returnChunkToDictionary(chunk_at_end); 9333 9334 MemRegion mr(_cmsSpace->bottom(), new_word_size); 9335 _bts->resize(new_word_size); // resize the block offset shared array 9336 Universe::heap()->barrier_set()->resize_covered_region(mr); 9337 _cmsSpace->assert_locked(); 9338 _cmsSpace->set_end((HeapWord*)_virtual_space.high()); 9339 9340 NOT_PRODUCT(_cmsSpace->dictionary()->verify()); 9341 9342 // update the space and generation capacity counters 9343 if (UsePerfData) { 9344 _space_counters->update_capacity(); 9345 _gen_counters->update_all(); 9346 } 9347 9348 if (Verbose && PrintGCDetails) { 9349 size_t new_mem_size = _virtual_space.committed_size(); 9350 size_t old_mem_size = new_mem_size + bytes; 9351 gclog_or_tty->print_cr("Shrinking %s from %ldK by %ldK to %ldK", 9352 name(), old_mem_size/K, bytes/K, new_mem_size/K); 9353 } 9354 } 9355 9356 assert(_cmsSpace->unallocated_block() <= _cmsSpace->end(), 9357 "Inconsistency at end of space"); 9358 assert(chunk_at_end->end() == _cmsSpace->end(), 9359 "Shrinking is inconsistent"); 9360 return; 9361 } 9362 } 9363 9364 // Transfer some number of overflown objects to usual marking 9365 // stack. Return true if some objects were transferred. 9366 bool MarkRefsIntoAndScanClosure::take_from_overflow_list() { 9367 size_t num = MIN2((size_t)(_mark_stack->capacity() - _mark_stack->length())/4, 9368 (size_t)ParGCDesiredObjsFromOverflowList); 9369 9370 bool res = _collector->take_from_overflow_list(num, _mark_stack); 9371 assert(_collector->overflow_list_is_empty() || res, 9372 "If list is not empty, we should have taken something"); 9373 assert(!res || !_mark_stack->isEmpty(), 9374 "If we took something, it should now be on our stack"); 9375 return res; 9376 } 9377 9378 size_t MarkDeadObjectsClosure::do_blk(HeapWord* addr) { 9379 size_t res = _sp->block_size_no_stall(addr, _collector); 9380 if (_sp->block_is_obj(addr)) { 9381 if (_live_bit_map->isMarked(addr)) { 9382 // It can't have been dead in a previous cycle 9383 guarantee(!_dead_bit_map->isMarked(addr), "No resurrection!"); 9384 } else { 9385 _dead_bit_map->mark(addr); // mark the dead object 9386 } 9387 } 9388 // Could be 0, if the block size could not be computed without stalling. 9389 return res; 9390 } 9391 9392 TraceCMSMemoryManagerStats::TraceCMSMemoryManagerStats(CMSCollector::CollectorState phase, GCCause::Cause cause): TraceMemoryManagerStats() { 9393 9394 switch (phase) { 9395 case CMSCollector::InitialMarking: 9396 initialize(true /* fullGC */ , 9397 cause /* cause of the GC */, 9398 true /* recordGCBeginTime */, 9399 true /* recordPreGCUsage */, 9400 false /* recordPeakUsage */, 9401 false /* recordPostGCusage */, 9402 true /* recordAccumulatedGCTime */, 9403 false /* recordGCEndTime */, 9404 false /* countCollection */ ); 9405 break; 9406 9407 case CMSCollector::FinalMarking: 9408 initialize(true /* fullGC */ , 9409 cause /* cause of the GC */, 9410 false /* recordGCBeginTime */, 9411 false /* recordPreGCUsage */, 9412 false /* recordPeakUsage */, 9413 false /* recordPostGCusage */, 9414 true /* recordAccumulatedGCTime */, 9415 false /* recordGCEndTime */, 9416 false /* countCollection */ ); 9417 break; 9418 9419 case CMSCollector::Sweeping: 9420 initialize(true /* fullGC */ , 9421 cause /* cause of the GC */, 9422 false /* recordGCBeginTime */, 9423 false /* recordPreGCUsage */, 9424 true /* recordPeakUsage */, 9425 true /* recordPostGCusage */, 9426 false /* recordAccumulatedGCTime */, 9427 true /* recordGCEndTime */, 9428 true /* countCollection */ ); 9429 break; 9430 9431 default: 9432 ShouldNotReachHere(); 9433 } 9434 } 9435