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 if (first_state > Idling) { 1896 save_heap_summary(); 1897 } 1898 1899 do_compaction_work(clear_all_soft_refs); 1900 1901 // Has the GC time limit been exceeded? 1902 DefNewGeneration* young_gen = _young_gen->as_DefNewGeneration(); 1903 size_t max_eden_size = young_gen->max_capacity() - 1904 young_gen->to()->capacity() - 1905 young_gen->from()->capacity(); 1906 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1907 GCCause::Cause gc_cause = gch->gc_cause(); 1908 size_policy()->check_gc_overhead_limit(_young_gen->used(), 1909 young_gen->eden()->used(), 1910 _cmsGen->max_capacity(), 1911 max_eden_size, 1912 full, 1913 gc_cause, 1914 gch->collector_policy()); 1915 } else { 1916 do_mark_sweep_work(clear_all_soft_refs, first_state, 1917 should_start_over); 1918 } 1919 // Reset the expansion cause, now that we just completed 1920 // a collection cycle. 1921 clear_expansion_cause(); 1922 _foregroundGCIsActive = false; 1923 return; 1924 } 1925 1926 // Resize the perm generation and the tenured generation 1927 // after obtaining the free list locks for the 1928 // two generations. 1929 void CMSCollector::compute_new_size() { 1930 assert_locked_or_safepoint(Heap_lock); 1931 FreelistLocker z(this); 1932 _permGen->compute_new_size(); 1933 _cmsGen->compute_new_size(); 1934 } 1935 1936 // A work method used by foreground collection to determine 1937 // what type of collection (compacting or not, continuing or fresh) 1938 // it should do. 1939 // NOTE: the intent is to make UseCMSCompactAtFullCollection 1940 // and CMSCompactWhenClearAllSoftRefs the default in the future 1941 // and do away with the flags after a suitable period. 1942 void CMSCollector::decide_foreground_collection_type( 1943 bool clear_all_soft_refs, bool* should_compact, 1944 bool* should_start_over) { 1945 // Normally, we'll compact only if the UseCMSCompactAtFullCollection 1946 // flag is set, and we have either requested a System.gc() or 1947 // the number of full gc's since the last concurrent cycle 1948 // has exceeded the threshold set by CMSFullGCsBeforeCompaction, 1949 // or if an incremental collection has failed 1950 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1951 assert(gch->collector_policy()->is_two_generation_policy(), 1952 "You may want to check the correctness of the following"); 1953 // Inform cms gen if this was due to partial collection failing. 1954 // The CMS gen may use this fact to determine its expansion policy. 1955 if (gch->incremental_collection_will_fail(false /* don't consult_young */)) { 1956 assert(!_cmsGen->incremental_collection_failed(), 1957 "Should have been noticed, reacted to and cleared"); 1958 _cmsGen->set_incremental_collection_failed(); 1959 } 1960 *should_compact = 1961 UseCMSCompactAtFullCollection && 1962 ((_full_gcs_since_conc_gc >= CMSFullGCsBeforeCompaction) || 1963 GCCause::is_user_requested_gc(gch->gc_cause()) || 1964 gch->incremental_collection_will_fail(true /* consult_young */)); 1965 *should_start_over = false; 1966 if (clear_all_soft_refs && !*should_compact) { 1967 // We are about to do a last ditch collection attempt 1968 // so it would normally make sense to do a compaction 1969 // to reclaim as much space as possible. 1970 if (CMSCompactWhenClearAllSoftRefs) { 1971 // Default: The rationale is that in this case either 1972 // we are past the final marking phase, in which case 1973 // we'd have to start over, or so little has been done 1974 // that there's little point in saving that work. Compaction 1975 // appears to be the sensible choice in either case. 1976 *should_compact = true; 1977 } else { 1978 // We have been asked to clear all soft refs, but not to 1979 // compact. Make sure that we aren't past the final checkpoint 1980 // phase, for that is where we process soft refs. If we are already 1981 // past that phase, we'll need to redo the refs discovery phase and 1982 // if necessary clear soft refs that weren't previously 1983 // cleared. We do so by remembering the phase in which 1984 // we came in, and if we are past the refs processing 1985 // phase, we'll choose to just redo the mark-sweep 1986 // collection from scratch. 1987 if (_collectorState > FinalMarking) { 1988 // We are past the refs processing phase; 1989 // start over and do a fresh synchronous CMS cycle 1990 _collectorState = Resetting; // skip to reset to start new cycle 1991 reset(false /* == !asynch */); 1992 *should_start_over = true; 1993 } // else we can continue a possibly ongoing current cycle 1994 } 1995 } 1996 } 1997 1998 // A work method used by the foreground collector to do 1999 // a mark-sweep-compact. 2000 void CMSCollector::do_compaction_work(bool clear_all_soft_refs) { 2001 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2002 2003 STWGCTimer* gc_timer = GenMarkSweep::gc_timer(); 2004 gc_timer->register_gc_start(os::elapsed_counter()); 2005 2006 SerialOldTracer* gc_tracer = GenMarkSweep::gc_tracer(); 2007 gc_tracer->report_gc_start(gch->gc_cause(), gc_timer->gc_start()); 2008 2009 GCTraceTime t("CMS:MSC ", PrintGCDetails && Verbose, true, NULL); 2010 if (PrintGC && Verbose && !(GCCause::is_user_requested_gc(gch->gc_cause()))) { 2011 gclog_or_tty->print_cr("Compact ConcurrentMarkSweepGeneration after %d " 2012 "collections passed to foreground collector", _full_gcs_since_conc_gc); 2013 } 2014 2015 // Sample collection interval time and reset for collection pause. 2016 if (UseAdaptiveSizePolicy) { 2017 size_policy()->msc_collection_begin(); 2018 } 2019 2020 // Temporarily widen the span of the weak reference processing to 2021 // the entire heap. 2022 MemRegion new_span(GenCollectedHeap::heap()->reserved_region()); 2023 ReferenceProcessorSpanMutator rp_mut_span(ref_processor(), new_span); 2024 // Temporarily, clear the "is_alive_non_header" field of the 2025 // reference processor. 2026 ReferenceProcessorIsAliveMutator rp_mut_closure(ref_processor(), NULL); 2027 // Temporarily make reference _processing_ single threaded (non-MT). 2028 ReferenceProcessorMTProcMutator rp_mut_mt_processing(ref_processor(), false); 2029 // Temporarily make refs discovery atomic 2030 ReferenceProcessorAtomicMutator rp_mut_atomic(ref_processor(), true); 2031 // Temporarily make reference _discovery_ single threaded (non-MT) 2032 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false); 2033 2034 ref_processor()->set_enqueuing_is_done(false); 2035 ref_processor()->enable_discovery(false /*verify_disabled*/, false /*check_no_refs*/); 2036 ref_processor()->setup_policy(clear_all_soft_refs); 2037 // If an asynchronous collection finishes, the _modUnionTable is 2038 // all clear. If we are assuming the collection from an asynchronous 2039 // collection, clear the _modUnionTable. 2040 assert(_collectorState != Idling || _modUnionTable.isAllClear(), 2041 "_modUnionTable should be clear if the baton was not passed"); 2042 _modUnionTable.clear_all(); 2043 2044 // We must adjust the allocation statistics being maintained 2045 // in the free list space. We do so by reading and clearing 2046 // the sweep timer and updating the block flux rate estimates below. 2047 assert(!_intra_sweep_timer.is_active(), "_intra_sweep_timer should be inactive"); 2048 if (_inter_sweep_timer.is_active()) { 2049 _inter_sweep_timer.stop(); 2050 // Note that we do not use this sample to update the _inter_sweep_estimate. 2051 _cmsGen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()), 2052 _inter_sweep_estimate.padded_average(), 2053 _intra_sweep_estimate.padded_average()); 2054 } 2055 2056 GenMarkSweep::invoke_at_safepoint(_cmsGen->level(), 2057 ref_processor(), clear_all_soft_refs); 2058 #ifdef ASSERT 2059 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 2060 size_t free_size = cms_space->free(); 2061 assert(free_size == 2062 pointer_delta(cms_space->end(), cms_space->compaction_top()) 2063 * HeapWordSize, 2064 "All the free space should be compacted into one chunk at top"); 2065 assert(cms_space->dictionary()->total_chunk_size( 2066 debug_only(cms_space->freelistLock())) == 0 || 2067 cms_space->totalSizeInIndexedFreeLists() == 0, 2068 "All the free space should be in a single chunk"); 2069 size_t num = cms_space->totalCount(); 2070 assert((free_size == 0 && num == 0) || 2071 (free_size > 0 && (num == 1 || num == 2)), 2072 "There should be at most 2 free chunks after compaction"); 2073 #endif // ASSERT 2074 _collectorState = Resetting; 2075 assert(_restart_addr == NULL, 2076 "Should have been NULL'd before baton was passed"); 2077 reset(false /* == !asynch */); 2078 _cmsGen->reset_after_compaction(); 2079 _concurrent_cycles_since_last_unload = 0; 2080 2081 if (verifying() && !should_unload_classes()) { 2082 perm_gen_verify_bit_map()->clear_all(); 2083 } 2084 2085 // Clear any data recorded in the PLAB chunk arrays. 2086 if (_survivor_plab_array != NULL) { 2087 reset_survivor_plab_arrays(); 2088 } 2089 2090 // Adjust the per-size allocation stats for the next epoch. 2091 _cmsGen->cmsSpace()->endSweepFLCensus(sweep_count() /* fake */); 2092 // Restart the "inter sweep timer" for the next epoch. 2093 _inter_sweep_timer.reset(); 2094 _inter_sweep_timer.start(); 2095 2096 // Sample collection pause time and reset for collection interval. 2097 if (UseAdaptiveSizePolicy) { 2098 size_policy()->msc_collection_end(gch->gc_cause()); 2099 } 2100 2101 gc_timer->register_gc_end(os::elapsed_counter()); 2102 2103 gc_tracer->report_gc_end(gc_timer->gc_end(), gc_timer->time_partitions()); 2104 2105 // For a mark-sweep-compact, compute_new_size() will be called 2106 // in the heap's do_collection() method. 2107 } 2108 2109 // A work method used by the foreground collector to do 2110 // a mark-sweep, after taking over from a possibly on-going 2111 // concurrent mark-sweep collection. 2112 void CMSCollector::do_mark_sweep_work(bool clear_all_soft_refs, 2113 CollectorState first_state, bool should_start_over) { 2114 if (PrintGC && Verbose) { 2115 gclog_or_tty->print_cr("Pass concurrent collection to foreground " 2116 "collector with count %d", 2117 _full_gcs_since_conc_gc); 2118 } 2119 switch (_collectorState) { 2120 case Idling: 2121 if (first_state == Idling || should_start_over) { 2122 // The background GC was not active, or should 2123 // restarted from scratch; start the cycle. 2124 _collectorState = InitialMarking; 2125 } 2126 // If first_state was not Idling, then a background GC 2127 // was in progress and has now finished. No need to do it 2128 // again. Leave the state as Idling. 2129 break; 2130 case Precleaning: 2131 // In the foreground case don't do the precleaning since 2132 // it is not done concurrently and there is extra work 2133 // required. 2134 _collectorState = FinalMarking; 2135 } 2136 collect_in_foreground(clear_all_soft_refs, GenCollectedHeap::heap()->gc_cause()); 2137 2138 // For a mark-sweep, compute_new_size() will be called 2139 // in the heap's do_collection() method. 2140 } 2141 2142 2143 void CMSCollector::getFreelistLocks() const { 2144 // Get locks for all free lists in all generations that this 2145 // collector is responsible for 2146 _cmsGen->freelistLock()->lock_without_safepoint_check(); 2147 _permGen->freelistLock()->lock_without_safepoint_check(); 2148 } 2149 2150 void CMSCollector::releaseFreelistLocks() const { 2151 // Release locks for all free lists in all generations that this 2152 // collector is responsible for 2153 _cmsGen->freelistLock()->unlock(); 2154 _permGen->freelistLock()->unlock(); 2155 } 2156 2157 bool CMSCollector::haveFreelistLocks() const { 2158 // Check locks for all free lists in all generations that this 2159 // collector is responsible for 2160 assert_lock_strong(_cmsGen->freelistLock()); 2161 assert_lock_strong(_permGen->freelistLock()); 2162 PRODUCT_ONLY(ShouldNotReachHere()); 2163 return true; 2164 } 2165 2166 // A utility class that is used by the CMS collector to 2167 // temporarily "release" the foreground collector from its 2168 // usual obligation to wait for the background collector to 2169 // complete an ongoing phase before proceeding. 2170 class ReleaseForegroundGC: public StackObj { 2171 private: 2172 CMSCollector* _c; 2173 public: 2174 ReleaseForegroundGC(CMSCollector* c) : _c(c) { 2175 assert(_c->_foregroundGCShouldWait, "Else should not need to call"); 2176 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 2177 // allow a potentially blocked foreground collector to proceed 2178 _c->_foregroundGCShouldWait = false; 2179 if (_c->_foregroundGCIsActive) { 2180 CGC_lock->notify(); 2181 } 2182 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 2183 "Possible deadlock"); 2184 } 2185 2186 ~ReleaseForegroundGC() { 2187 assert(!_c->_foregroundGCShouldWait, "Usage protocol violation?"); 2188 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 2189 _c->_foregroundGCShouldWait = true; 2190 } 2191 }; 2192 2193 // There are separate collect_in_background and collect_in_foreground because of 2194 // the different locking requirements of the background collector and the 2195 // foreground collector. There was originally an attempt to share 2196 // one "collect" method between the background collector and the foreground 2197 // collector but the if-then-else required made it cleaner to have 2198 // separate methods. 2199 void CMSCollector::collect_in_background(bool clear_all_soft_refs, GCCause::Cause cause) { 2200 assert(Thread::current()->is_ConcurrentGC_thread(), 2201 "A CMS asynchronous collection is only allowed on a CMS thread."); 2202 2203 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2204 { 2205 bool safepoint_check = Mutex::_no_safepoint_check_flag; 2206 MutexLockerEx hl(Heap_lock, safepoint_check); 2207 FreelistLocker fll(this); 2208 MutexLockerEx x(CGC_lock, safepoint_check); 2209 if (_foregroundGCIsActive || !UseAsyncConcMarkSweepGC) { 2210 // The foreground collector is active or we're 2211 // not using asynchronous collections. Skip this 2212 // background collection. 2213 assert(!_foregroundGCShouldWait, "Should be clear"); 2214 return; 2215 } else { 2216 assert(_collectorState == Idling, "Should be idling before start."); 2217 _collectorState = InitialMarking; 2218 register_gc_start(cause); 2219 // Reset the expansion cause, now that we are about to begin 2220 // a new cycle. 2221 clear_expansion_cause(); 2222 } 2223 // Decide if we want to enable class unloading as part of the 2224 // ensuing concurrent GC cycle. 2225 update_should_unload_classes(); 2226 _full_gc_requested = false; // acks all outstanding full gc requests 2227 _full_gc_cause = GCCause::_no_gc; 2228 // Signal that we are about to start a collection 2229 gch->increment_total_full_collections(); // ... starting a collection cycle 2230 _collection_count_start = gch->total_full_collections(); 2231 } 2232 2233 // Used for PrintGC 2234 size_t prev_used; 2235 if (PrintGC && Verbose) { 2236 prev_used = _cmsGen->used(); // XXXPERM 2237 } 2238 2239 // The change of the collection state is normally done at this level; 2240 // the exceptions are phases that are executed while the world is 2241 // stopped. For those phases the change of state is done while the 2242 // world is stopped. For baton passing purposes this allows the 2243 // background collector to finish the phase and change state atomically. 2244 // The foreground collector cannot wait on a phase that is done 2245 // while the world is stopped because the foreground collector already 2246 // has the world stopped and would deadlock. 2247 while (_collectorState != Idling) { 2248 if (TraceCMSState) { 2249 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " in CMS state %d", 2250 Thread::current(), _collectorState); 2251 } 2252 // The foreground collector 2253 // holds the Heap_lock throughout its collection. 2254 // holds the CMS token (but not the lock) 2255 // except while it is waiting for the background collector to yield. 2256 // 2257 // The foreground collector should be blocked (not for long) 2258 // if the background collector is about to start a phase 2259 // executed with world stopped. If the background 2260 // collector has already started such a phase, the 2261 // foreground collector is blocked waiting for the 2262 // Heap_lock. The stop-world phases (InitialMarking and FinalMarking) 2263 // are executed in the VM thread. 2264 // 2265 // The locking order is 2266 // PendingListLock (PLL) -- if applicable (FinalMarking) 2267 // Heap_lock (both this & PLL locked in VM_CMS_Operation::prologue()) 2268 // CMS token (claimed in 2269 // stop_world_and_do() --> 2270 // safepoint_synchronize() --> 2271 // CMSThread::synchronize()) 2272 2273 { 2274 // Check if the FG collector wants us to yield. 2275 CMSTokenSync x(true); // is cms thread 2276 if (waitForForegroundGC()) { 2277 // We yielded to a foreground GC, nothing more to be 2278 // done this round. 2279 assert(_foregroundGCShouldWait == false, "We set it to false in " 2280 "waitForForegroundGC()"); 2281 if (TraceCMSState) { 2282 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT 2283 " exiting collection CMS state %d", 2284 Thread::current(), _collectorState); 2285 } 2286 return; 2287 } else { 2288 // The background collector can run but check to see if the 2289 // foreground collector has done a collection while the 2290 // background collector was waiting to get the CGC_lock 2291 // above. If yes, break so that _foregroundGCShouldWait 2292 // is cleared before returning. 2293 if (_collectorState == Idling) { 2294 break; 2295 } 2296 } 2297 } 2298 2299 assert(_foregroundGCShouldWait, "Foreground collector, if active, " 2300 "should be waiting"); 2301 2302 switch (_collectorState) { 2303 case InitialMarking: 2304 { 2305 ReleaseForegroundGC x(this); 2306 stats().record_cms_begin(); 2307 VM_CMS_Initial_Mark initial_mark_op(this); 2308 VMThread::execute(&initial_mark_op); 2309 } 2310 // The collector state may be any legal state at this point 2311 // since the background collector may have yielded to the 2312 // foreground collector. 2313 break; 2314 case Marking: 2315 // initial marking in checkpointRootsInitialWork has been completed 2316 if (markFromRoots(true)) { // we were successful 2317 assert(_collectorState == Precleaning, "Collector state should " 2318 "have changed"); 2319 } else { 2320 assert(_foregroundGCIsActive, "Internal state inconsistency"); 2321 } 2322 break; 2323 case Precleaning: 2324 if (UseAdaptiveSizePolicy) { 2325 size_policy()->concurrent_precleaning_begin(); 2326 } 2327 // marking from roots in markFromRoots has been completed 2328 preclean(); 2329 if (UseAdaptiveSizePolicy) { 2330 size_policy()->concurrent_precleaning_end(); 2331 } 2332 assert(_collectorState == AbortablePreclean || 2333 _collectorState == FinalMarking, 2334 "Collector state should have changed"); 2335 break; 2336 case AbortablePreclean: 2337 if (UseAdaptiveSizePolicy) { 2338 size_policy()->concurrent_phases_resume(); 2339 } 2340 abortable_preclean(); 2341 if (UseAdaptiveSizePolicy) { 2342 size_policy()->concurrent_precleaning_end(); 2343 } 2344 assert(_collectorState == FinalMarking, "Collector state should " 2345 "have changed"); 2346 break; 2347 case FinalMarking: 2348 { 2349 ReleaseForegroundGC x(this); 2350 2351 VM_CMS_Final_Remark final_remark_op(this); 2352 VMThread::execute(&final_remark_op); 2353 } 2354 assert(_foregroundGCShouldWait, "block post-condition"); 2355 break; 2356 case Sweeping: 2357 if (UseAdaptiveSizePolicy) { 2358 size_policy()->concurrent_sweeping_begin(); 2359 } 2360 // final marking in checkpointRootsFinal has been completed 2361 sweep(true); 2362 assert(_collectorState == Resizing, "Collector state change " 2363 "to Resizing must be done under the free_list_lock"); 2364 _full_gcs_since_conc_gc = 0; 2365 2366 // Stop the timers for adaptive size policy for the concurrent phases 2367 if (UseAdaptiveSizePolicy) { 2368 size_policy()->concurrent_sweeping_end(); 2369 size_policy()->concurrent_phases_end(gch->gc_cause(), 2370 gch->prev_gen(_cmsGen)->capacity(), 2371 _cmsGen->free()); 2372 } 2373 2374 case Resizing: { 2375 // Sweeping has been completed... 2376 // At this point the background collection has completed. 2377 // Don't move the call to compute_new_size() down 2378 // into code that might be executed if the background 2379 // collection was preempted. 2380 { 2381 ReleaseForegroundGC x(this); // unblock FG collection 2382 MutexLockerEx y(Heap_lock, Mutex::_no_safepoint_check_flag); 2383 CMSTokenSync z(true); // not strictly needed. 2384 if (_collectorState == Resizing) { 2385 compute_new_size(); 2386 save_heap_summary(); 2387 _collectorState = Resetting; 2388 } else { 2389 assert(_collectorState == Idling, "The state should only change" 2390 " because the foreground collector has finished the collection"); 2391 } 2392 } 2393 break; 2394 } 2395 case Resetting: 2396 // CMS heap resizing has been completed 2397 reset(true); 2398 assert(_collectorState == Idling, "Collector state should " 2399 "have changed"); 2400 stats().record_cms_end(); 2401 // Don't move the concurrent_phases_end() and compute_new_size() 2402 // calls to here because a preempted background collection 2403 // has it's state set to "Resetting". 2404 break; 2405 case Idling: 2406 default: 2407 ShouldNotReachHere(); 2408 break; 2409 } 2410 if (TraceCMSState) { 2411 gclog_or_tty->print_cr(" Thread " INTPTR_FORMAT " done - next CMS state %d", 2412 Thread::current(), _collectorState); 2413 } 2414 assert(_foregroundGCShouldWait, "block post-condition"); 2415 } 2416 2417 // Should this be in gc_epilogue? 2418 collector_policy()->counters()->update_counters(); 2419 2420 { 2421 // Clear _foregroundGCShouldWait and, in the event that the 2422 // foreground collector is waiting, notify it, before 2423 // returning. 2424 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 2425 _foregroundGCShouldWait = false; 2426 if (_foregroundGCIsActive) { 2427 CGC_lock->notify(); 2428 } 2429 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 2430 "Possible deadlock"); 2431 } 2432 if (TraceCMSState) { 2433 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT 2434 " exiting collection CMS state %d", 2435 Thread::current(), _collectorState); 2436 } 2437 if (PrintGC && Verbose) { 2438 _cmsGen->print_heap_change(prev_used); 2439 } 2440 } 2441 2442 void CMSCollector::register_foreground_gc_start(GCCause::Cause cause) { 2443 if (!_cms_start_registered) { 2444 register_gc_start(cause); 2445 } 2446 } 2447 2448 void CMSCollector::register_gc_start(GCCause::Cause cause) { 2449 _cms_start_registered = true; 2450 _gc_timer_cm->register_gc_start(os::elapsed_counter()); 2451 _gc_tracer_cm->report_gc_start(cause, _gc_timer_cm->gc_start()); 2452 } 2453 2454 void CMSCollector::register_gc_end() { 2455 if (_cms_start_registered) { 2456 report_heap_summary(GCWhen::AfterGC); 2457 2458 _gc_timer_cm->register_gc_end(os::elapsed_counter()); 2459 _gc_tracer_cm->report_gc_end(_gc_timer_cm->gc_end(), _gc_timer_cm->time_partitions()); 2460 _cms_start_registered = false; 2461 } 2462 } 2463 2464 void CMSCollector::save_heap_summary() { 2465 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2466 _last_heap_summary = gch->create_heap_summary(); 2467 _last_perm_gen_summary = gch->create_perm_gen_summary(); 2468 } 2469 2470 void CMSCollector::report_heap_summary(GCWhen::Type when) { 2471 _gc_tracer_cm->report_gc_heap_summary(when, _last_heap_summary, _last_perm_gen_summary); 2472 } 2473 2474 void CMSCollector::collect_in_foreground(bool clear_all_soft_refs, GCCause::Cause cause) { 2475 assert(_foregroundGCIsActive && !_foregroundGCShouldWait, 2476 "Foreground collector should be waiting, not executing"); 2477 assert(Thread::current()->is_VM_thread(), "A foreground collection" 2478 "may only be done by the VM Thread with the world stopped"); 2479 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 2480 "VM thread should have CMS token"); 2481 2482 NOT_PRODUCT(GCTraceTime t("CMS:MS (foreground) ", PrintGCDetails && Verbose, 2483 true, NULL);) 2484 if (UseAdaptiveSizePolicy) { 2485 size_policy()->ms_collection_begin(); 2486 } 2487 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact); 2488 2489 HandleMark hm; // Discard invalid handles created during verification 2490 2491 if (VerifyBeforeGC && 2492 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2493 Universe::verify(); 2494 } 2495 2496 // Snapshot the soft reference policy to be used in this collection cycle. 2497 ref_processor()->setup_policy(clear_all_soft_refs); 2498 2499 bool init_mark_was_synchronous = false; // until proven otherwise 2500 while (_collectorState != Idling) { 2501 if (TraceCMSState) { 2502 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " in CMS state %d", 2503 Thread::current(), _collectorState); 2504 } 2505 switch (_collectorState) { 2506 case InitialMarking: 2507 register_foreground_gc_start(cause); 2508 init_mark_was_synchronous = true; // fact to be exploited in re-mark 2509 checkpointRootsInitial(false); 2510 assert(_collectorState == Marking, "Collector state should have changed" 2511 " within checkpointRootsInitial()"); 2512 break; 2513 case Marking: 2514 // initial marking in checkpointRootsInitialWork has been completed 2515 if (VerifyDuringGC && 2516 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2517 gclog_or_tty->print("Verify before initial mark: "); 2518 Universe::verify(); 2519 } 2520 { 2521 bool res = markFromRoots(false); 2522 assert(res && _collectorState == FinalMarking, "Collector state should " 2523 "have changed"); 2524 break; 2525 } 2526 case FinalMarking: 2527 if (VerifyDuringGC && 2528 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2529 gclog_or_tty->print("Verify before re-mark: "); 2530 Universe::verify(); 2531 } 2532 checkpointRootsFinal(false, clear_all_soft_refs, 2533 init_mark_was_synchronous); 2534 assert(_collectorState == Sweeping, "Collector state should not " 2535 "have changed within checkpointRootsFinal()"); 2536 break; 2537 case Sweeping: 2538 // final marking in checkpointRootsFinal has been completed 2539 if (VerifyDuringGC && 2540 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2541 gclog_or_tty->print("Verify before sweep: "); 2542 Universe::verify(); 2543 } 2544 sweep(false); 2545 assert(_collectorState == Resizing, "Incorrect state"); 2546 break; 2547 case Resizing: { 2548 // Sweeping has been completed; the actual resize in this case 2549 // is done separately; nothing to be done in this state. 2550 _collectorState = Resetting; 2551 break; 2552 } 2553 case Resetting: 2554 // The heap has been resized. 2555 if (VerifyDuringGC && 2556 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2557 gclog_or_tty->print("Verify before reset: "); 2558 Universe::verify(); 2559 } 2560 save_heap_summary(); 2561 reset(false); 2562 assert(_collectorState == Idling, "Collector state should " 2563 "have changed"); 2564 break; 2565 case Precleaning: 2566 case AbortablePreclean: 2567 // Elide the preclean phase 2568 _collectorState = FinalMarking; 2569 break; 2570 default: 2571 ShouldNotReachHere(); 2572 } 2573 if (TraceCMSState) { 2574 gclog_or_tty->print_cr(" Thread " INTPTR_FORMAT " done - next CMS state %d", 2575 Thread::current(), _collectorState); 2576 } 2577 } 2578 2579 if (UseAdaptiveSizePolicy) { 2580 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2581 size_policy()->ms_collection_end(gch->gc_cause()); 2582 } 2583 2584 if (VerifyAfterGC && 2585 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2586 Universe::verify(); 2587 } 2588 if (TraceCMSState) { 2589 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT 2590 " exiting collection CMS state %d", 2591 Thread::current(), _collectorState); 2592 } 2593 } 2594 2595 bool CMSCollector::waitForForegroundGC() { 2596 bool res = false; 2597 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 2598 "CMS thread should have CMS token"); 2599 // Block the foreground collector until the 2600 // background collectors decides whether to 2601 // yield. 2602 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 2603 _foregroundGCShouldWait = true; 2604 if (_foregroundGCIsActive) { 2605 // The background collector yields to the 2606 // foreground collector and returns a value 2607 // indicating that it has yielded. The foreground 2608 // collector can proceed. 2609 res = true; 2610 _foregroundGCShouldWait = false; 2611 ConcurrentMarkSweepThread::clear_CMS_flag( 2612 ConcurrentMarkSweepThread::CMS_cms_has_token); 2613 ConcurrentMarkSweepThread::set_CMS_flag( 2614 ConcurrentMarkSweepThread::CMS_cms_wants_token); 2615 // Get a possibly blocked foreground thread going 2616 CGC_lock->notify(); 2617 if (TraceCMSState) { 2618 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " waiting at CMS state %d", 2619 Thread::current(), _collectorState); 2620 } 2621 while (_foregroundGCIsActive) { 2622 CGC_lock->wait(Mutex::_no_safepoint_check_flag); 2623 } 2624 ConcurrentMarkSweepThread::set_CMS_flag( 2625 ConcurrentMarkSweepThread::CMS_cms_has_token); 2626 ConcurrentMarkSweepThread::clear_CMS_flag( 2627 ConcurrentMarkSweepThread::CMS_cms_wants_token); 2628 } 2629 if (TraceCMSState) { 2630 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " continuing at CMS state %d", 2631 Thread::current(), _collectorState); 2632 } 2633 return res; 2634 } 2635 2636 // Because of the need to lock the free lists and other structures in 2637 // the collector, common to all the generations that the collector is 2638 // collecting, we need the gc_prologues of individual CMS generations 2639 // delegate to their collector. It may have been simpler had the 2640 // current infrastructure allowed one to call a prologue on a 2641 // collector. In the absence of that we have the generation's 2642 // prologue delegate to the collector, which delegates back 2643 // some "local" work to a worker method in the individual generations 2644 // that it's responsible for collecting, while itself doing any 2645 // work common to all generations it's responsible for. A similar 2646 // comment applies to the gc_epilogue()'s. 2647 // The role of the varaible _between_prologue_and_epilogue is to 2648 // enforce the invocation protocol. 2649 void CMSCollector::gc_prologue(bool full) { 2650 // Call gc_prologue_work() for each CMSGen and PermGen that 2651 // we are responsible for. 2652 2653 // The following locking discipline assumes that we are only called 2654 // when the world is stopped. 2655 assert(SafepointSynchronize::is_at_safepoint(), "world is stopped assumption"); 2656 2657 // The CMSCollector prologue must call the gc_prologues for the 2658 // "generations" (including PermGen if any) that it's responsible 2659 // for. 2660 2661 assert( Thread::current()->is_VM_thread() 2662 || ( CMSScavengeBeforeRemark 2663 && Thread::current()->is_ConcurrentGC_thread()), 2664 "Incorrect thread type for prologue execution"); 2665 2666 if (_between_prologue_and_epilogue) { 2667 // We have already been invoked; this is a gc_prologue delegation 2668 // from yet another CMS generation that we are responsible for, just 2669 // ignore it since all relevant work has already been done. 2670 return; 2671 } 2672 2673 // set a bit saying prologue has been called; cleared in epilogue 2674 _between_prologue_and_epilogue = true; 2675 // Claim locks for common data structures, then call gc_prologue_work() 2676 // for each CMSGen and PermGen that we are responsible for. 2677 2678 getFreelistLocks(); // gets free list locks on constituent spaces 2679 bitMapLock()->lock_without_safepoint_check(); 2680 2681 // Should call gc_prologue_work() for all cms gens we are responsible for 2682 bool registerClosure = _collectorState >= Marking 2683 && _collectorState < Sweeping; 2684 ModUnionClosure* muc = CollectedHeap::use_parallel_gc_threads() ? 2685 &_modUnionClosurePar 2686 : &_modUnionClosure; 2687 _cmsGen->gc_prologue_work(full, registerClosure, muc); 2688 _permGen->gc_prologue_work(full, registerClosure, muc); 2689 2690 if (!full) { 2691 stats().record_gc0_begin(); 2692 } 2693 } 2694 2695 void ConcurrentMarkSweepGeneration::gc_prologue(bool full) { 2696 // Delegate to CMScollector which knows how to coordinate between 2697 // this and any other CMS generations that it is responsible for 2698 // collecting. 2699 collector()->gc_prologue(full); 2700 } 2701 2702 // This is a "private" interface for use by this generation's CMSCollector. 2703 // Not to be called directly by any other entity (for instance, 2704 // GenCollectedHeap, which calls the "public" gc_prologue method above). 2705 void ConcurrentMarkSweepGeneration::gc_prologue_work(bool full, 2706 bool registerClosure, ModUnionClosure* modUnionClosure) { 2707 assert(!incremental_collection_failed(), "Shouldn't be set yet"); 2708 assert(cmsSpace()->preconsumptionDirtyCardClosure() == NULL, 2709 "Should be NULL"); 2710 if (registerClosure) { 2711 cmsSpace()->setPreconsumptionDirtyCardClosure(modUnionClosure); 2712 } 2713 cmsSpace()->gc_prologue(); 2714 // Clear stat counters 2715 NOT_PRODUCT( 2716 assert(_numObjectsPromoted == 0, "check"); 2717 assert(_numWordsPromoted == 0, "check"); 2718 if (Verbose && PrintGC) { 2719 gclog_or_tty->print("Allocated "SIZE_FORMAT" objects, " 2720 SIZE_FORMAT" bytes concurrently", 2721 _numObjectsAllocated, _numWordsAllocated*sizeof(HeapWord)); 2722 } 2723 _numObjectsAllocated = 0; 2724 _numWordsAllocated = 0; 2725 ) 2726 } 2727 2728 void CMSCollector::gc_epilogue(bool full) { 2729 // The following locking discipline assumes that we are only called 2730 // when the world is stopped. 2731 assert(SafepointSynchronize::is_at_safepoint(), 2732 "world is stopped assumption"); 2733 2734 // Currently the CMS epilogue (see CompactibleFreeListSpace) merely checks 2735 // if linear allocation blocks need to be appropriately marked to allow the 2736 // the blocks to be parsable. We also check here whether we need to nudge the 2737 // CMS collector thread to start a new cycle (if it's not already active). 2738 assert( Thread::current()->is_VM_thread() 2739 || ( CMSScavengeBeforeRemark 2740 && Thread::current()->is_ConcurrentGC_thread()), 2741 "Incorrect thread type for epilogue execution"); 2742 2743 if (!_between_prologue_and_epilogue) { 2744 // We have already been invoked; this is a gc_epilogue delegation 2745 // from yet another CMS generation that we are responsible for, just 2746 // ignore it since all relevant work has already been done. 2747 return; 2748 } 2749 assert(haveFreelistLocks(), "must have freelist locks"); 2750 assert_lock_strong(bitMapLock()); 2751 2752 _cmsGen->gc_epilogue_work(full); 2753 _permGen->gc_epilogue_work(full); 2754 2755 if (_collectorState == AbortablePreclean || _collectorState == Precleaning) { 2756 // in case sampling was not already enabled, enable it 2757 _start_sampling = true; 2758 } 2759 // reset _eden_chunk_array so sampling starts afresh 2760 _eden_chunk_index = 0; 2761 2762 size_t cms_used = _cmsGen->cmsSpace()->used(); 2763 size_t perm_used = _permGen->cmsSpace()->used(); 2764 2765 // update performance counters - this uses a special version of 2766 // update_counters() that allows the utilization to be passed as a 2767 // parameter, avoiding multiple calls to used(). 2768 // 2769 _cmsGen->update_counters(cms_used); 2770 _permGen->update_counters(perm_used); 2771 2772 if (CMSIncrementalMode) { 2773 icms_update_allocation_limits(); 2774 } 2775 2776 bitMapLock()->unlock(); 2777 releaseFreelistLocks(); 2778 2779 if (!CleanChunkPoolAsync) { 2780 Chunk::clean_chunk_pool(); 2781 } 2782 2783 _between_prologue_and_epilogue = false; // ready for next cycle 2784 } 2785 2786 void ConcurrentMarkSweepGeneration::gc_epilogue(bool full) { 2787 collector()->gc_epilogue(full); 2788 2789 // Also reset promotion tracking in par gc thread states. 2790 if (CollectedHeap::use_parallel_gc_threads()) { 2791 for (uint i = 0; i < ParallelGCThreads; i++) { 2792 _par_gc_thread_states[i]->promo.stopTrackingPromotions(i); 2793 } 2794 } 2795 } 2796 2797 void ConcurrentMarkSweepGeneration::gc_epilogue_work(bool full) { 2798 assert(!incremental_collection_failed(), "Should have been cleared"); 2799 cmsSpace()->setPreconsumptionDirtyCardClosure(NULL); 2800 cmsSpace()->gc_epilogue(); 2801 // Print stat counters 2802 NOT_PRODUCT( 2803 assert(_numObjectsAllocated == 0, "check"); 2804 assert(_numWordsAllocated == 0, "check"); 2805 if (Verbose && PrintGC) { 2806 gclog_or_tty->print("Promoted "SIZE_FORMAT" objects, " 2807 SIZE_FORMAT" bytes", 2808 _numObjectsPromoted, _numWordsPromoted*sizeof(HeapWord)); 2809 } 2810 _numObjectsPromoted = 0; 2811 _numWordsPromoted = 0; 2812 ) 2813 2814 if (PrintGC && Verbose) { 2815 // Call down the chain in contiguous_available needs the freelistLock 2816 // so print this out before releasing the freeListLock. 2817 gclog_or_tty->print(" Contiguous available "SIZE_FORMAT" bytes ", 2818 contiguous_available()); 2819 } 2820 } 2821 2822 #ifndef PRODUCT 2823 bool CMSCollector::have_cms_token() { 2824 Thread* thr = Thread::current(); 2825 if (thr->is_VM_thread()) { 2826 return ConcurrentMarkSweepThread::vm_thread_has_cms_token(); 2827 } else if (thr->is_ConcurrentGC_thread()) { 2828 return ConcurrentMarkSweepThread::cms_thread_has_cms_token(); 2829 } else if (thr->is_GC_task_thread()) { 2830 return ConcurrentMarkSweepThread::vm_thread_has_cms_token() && 2831 ParGCRareEvent_lock->owned_by_self(); 2832 } 2833 return false; 2834 } 2835 #endif 2836 2837 // Check reachability of the given heap address in CMS generation, 2838 // treating all other generations as roots. 2839 bool CMSCollector::is_cms_reachable(HeapWord* addr) { 2840 // We could "guarantee" below, rather than assert, but i'll 2841 // leave these as "asserts" so that an adventurous debugger 2842 // could try this in the product build provided some subset of 2843 // the conditions were met, provided they were intersted in the 2844 // results and knew that the computation below wouldn't interfere 2845 // with other concurrent computations mutating the structures 2846 // being read or written. 2847 assert(SafepointSynchronize::is_at_safepoint(), 2848 "Else mutations in object graph will make answer suspect"); 2849 assert(have_cms_token(), "Should hold cms token"); 2850 assert(haveFreelistLocks(), "must hold free list locks"); 2851 assert_lock_strong(bitMapLock()); 2852 2853 // Clear the marking bit map array before starting, but, just 2854 // for kicks, first report if the given address is already marked 2855 gclog_or_tty->print_cr("Start: Address 0x%x is%s marked", addr, 2856 _markBitMap.isMarked(addr) ? "" : " not"); 2857 2858 if (verify_after_remark()) { 2859 MutexLockerEx x(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag); 2860 bool result = verification_mark_bm()->isMarked(addr); 2861 gclog_or_tty->print_cr("TransitiveMark: Address 0x%x %s marked", addr, 2862 result ? "IS" : "is NOT"); 2863 return result; 2864 } else { 2865 gclog_or_tty->print_cr("Could not compute result"); 2866 return false; 2867 } 2868 } 2869 2870 //////////////////////////////////////////////////////// 2871 // CMS Verification Support 2872 //////////////////////////////////////////////////////// 2873 // Following the remark phase, the following invariant 2874 // should hold -- each object in the CMS heap which is 2875 // marked in markBitMap() should be marked in the verification_mark_bm(). 2876 2877 class VerifyMarkedClosure: public BitMapClosure { 2878 CMSBitMap* _marks; 2879 bool _failed; 2880 2881 public: 2882 VerifyMarkedClosure(CMSBitMap* bm): _marks(bm), _failed(false) {} 2883 2884 bool do_bit(size_t offset) { 2885 HeapWord* addr = _marks->offsetToHeapWord(offset); 2886 if (!_marks->isMarked(addr)) { 2887 oop(addr)->print_on(gclog_or_tty); 2888 gclog_or_tty->print_cr(" ("INTPTR_FORMAT" should have been marked)", addr); 2889 _failed = true; 2890 } 2891 return true; 2892 } 2893 2894 bool failed() { return _failed; } 2895 }; 2896 2897 bool CMSCollector::verify_after_remark() { 2898 gclog_or_tty->print(" [Verifying CMS Marking... "); 2899 MutexLockerEx ml(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag); 2900 static bool init = false; 2901 2902 assert(SafepointSynchronize::is_at_safepoint(), 2903 "Else mutations in object graph will make answer suspect"); 2904 assert(have_cms_token(), 2905 "Else there may be mutual interference in use of " 2906 " verification data structures"); 2907 assert(_collectorState > Marking && _collectorState <= Sweeping, 2908 "Else marking info checked here may be obsolete"); 2909 assert(haveFreelistLocks(), "must hold free list locks"); 2910 assert_lock_strong(bitMapLock()); 2911 2912 2913 // Allocate marking bit map if not already allocated 2914 if (!init) { // first time 2915 if (!verification_mark_bm()->allocate(_span)) { 2916 return false; 2917 } 2918 init = true; 2919 } 2920 2921 assert(verification_mark_stack()->isEmpty(), "Should be empty"); 2922 2923 // Turn off refs discovery -- so we will be tracing through refs. 2924 // This is as intended, because by this time 2925 // GC must already have cleared any refs that need to be cleared, 2926 // and traced those that need to be marked; moreover, 2927 // the marking done here is not going to intefere in any 2928 // way with the marking information used by GC. 2929 NoRefDiscovery no_discovery(ref_processor()); 2930 2931 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;) 2932 2933 // Clear any marks from a previous round 2934 verification_mark_bm()->clear_all(); 2935 assert(verification_mark_stack()->isEmpty(), "markStack should be empty"); 2936 verify_work_stacks_empty(); 2937 2938 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2939 gch->ensure_parsability(false); // fill TLABs, but no need to retire them 2940 // Update the saved marks which may affect the root scans. 2941 gch->save_marks(); 2942 2943 if (CMSRemarkVerifyVariant == 1) { 2944 // In this first variant of verification, we complete 2945 // all marking, then check if the new marks-verctor is 2946 // a subset of the CMS marks-vector. 2947 verify_after_remark_work_1(); 2948 } else if (CMSRemarkVerifyVariant == 2) { 2949 // In this second variant of verification, we flag an error 2950 // (i.e. an object reachable in the new marks-vector not reachable 2951 // in the CMS marks-vector) immediately, also indicating the 2952 // identify of an object (A) that references the unmarked object (B) -- 2953 // presumably, a mutation to A failed to be picked up by preclean/remark? 2954 verify_after_remark_work_2(); 2955 } else { 2956 warning("Unrecognized value %d for CMSRemarkVerifyVariant", 2957 CMSRemarkVerifyVariant); 2958 } 2959 gclog_or_tty->print(" done] "); 2960 return true; 2961 } 2962 2963 void CMSCollector::verify_after_remark_work_1() { 2964 ResourceMark rm; 2965 HandleMark hm; 2966 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2967 2968 // Mark from roots one level into CMS 2969 MarkRefsIntoClosure notOlder(_span, verification_mark_bm()); 2970 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2971 2972 gch->gen_process_strong_roots(_cmsGen->level(), 2973 true, // younger gens are roots 2974 true, // activate StrongRootsScope 2975 true, // collecting perm gen 2976 SharedHeap::ScanningOption(roots_scanning_options()), 2977 ¬Older, 2978 true, // walk code active on stacks 2979 NULL); 2980 2981 // Now mark from the roots 2982 assert(_revisitStack.isEmpty(), "Should be empty"); 2983 MarkFromRootsClosure markFromRootsClosure(this, _span, 2984 verification_mark_bm(), verification_mark_stack(), &_revisitStack, 2985 false /* don't yield */, true /* verifying */); 2986 assert(_restart_addr == NULL, "Expected pre-condition"); 2987 verification_mark_bm()->iterate(&markFromRootsClosure); 2988 while (_restart_addr != NULL) { 2989 // Deal with stack overflow: by restarting at the indicated 2990 // address. 2991 HeapWord* ra = _restart_addr; 2992 markFromRootsClosure.reset(ra); 2993 _restart_addr = NULL; 2994 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end()); 2995 } 2996 assert(verification_mark_stack()->isEmpty(), "Should have been drained"); 2997 verify_work_stacks_empty(); 2998 // Should reset the revisit stack above, since no class tree 2999 // surgery is forthcoming. 3000 _revisitStack.reset(); // throwing away all contents 3001 3002 // Marking completed -- now verify that each bit marked in 3003 // verification_mark_bm() is also marked in markBitMap(); flag all 3004 // errors by printing corresponding objects. 3005 VerifyMarkedClosure vcl(markBitMap()); 3006 verification_mark_bm()->iterate(&vcl); 3007 if (vcl.failed()) { 3008 gclog_or_tty->print("Verification failed"); 3009 Universe::heap()->print_on(gclog_or_tty); 3010 fatal("CMS: failed marking verification after remark"); 3011 } 3012 } 3013 3014 void CMSCollector::verify_after_remark_work_2() { 3015 ResourceMark rm; 3016 HandleMark hm; 3017 GenCollectedHeap* gch = GenCollectedHeap::heap(); 3018 3019 // Mark from roots one level into CMS 3020 MarkRefsIntoVerifyClosure notOlder(_span, verification_mark_bm(), 3021 markBitMap()); 3022 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 3023 gch->gen_process_strong_roots(_cmsGen->level(), 3024 true, // younger gens are roots 3025 true, // activate StrongRootsScope 3026 true, // collecting perm gen 3027 SharedHeap::ScanningOption(roots_scanning_options()), 3028 ¬Older, 3029 true, // walk code active on stacks 3030 NULL); 3031 3032 // Now mark from the roots 3033 assert(_revisitStack.isEmpty(), "Should be empty"); 3034 MarkFromRootsVerifyClosure markFromRootsClosure(this, _span, 3035 verification_mark_bm(), markBitMap(), verification_mark_stack()); 3036 assert(_restart_addr == NULL, "Expected pre-condition"); 3037 verification_mark_bm()->iterate(&markFromRootsClosure); 3038 while (_restart_addr != NULL) { 3039 // Deal with stack overflow: by restarting at the indicated 3040 // address. 3041 HeapWord* ra = _restart_addr; 3042 markFromRootsClosure.reset(ra); 3043 _restart_addr = NULL; 3044 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end()); 3045 } 3046 assert(verification_mark_stack()->isEmpty(), "Should have been drained"); 3047 verify_work_stacks_empty(); 3048 // Should reset the revisit stack above, since no class tree 3049 // surgery is forthcoming. 3050 _revisitStack.reset(); // throwing away all contents 3051 3052 // Marking completed -- now verify that each bit marked in 3053 // verification_mark_bm() is also marked in markBitMap(); flag all 3054 // errors by printing corresponding objects. 3055 VerifyMarkedClosure vcl(markBitMap()); 3056 verification_mark_bm()->iterate(&vcl); 3057 assert(!vcl.failed(), "Else verification above should not have succeeded"); 3058 } 3059 3060 void ConcurrentMarkSweepGeneration::save_marks() { 3061 // delegate to CMS space 3062 cmsSpace()->save_marks(); 3063 for (uint i = 0; i < ParallelGCThreads; i++) { 3064 _par_gc_thread_states[i]->promo.startTrackingPromotions(); 3065 } 3066 } 3067 3068 bool ConcurrentMarkSweepGeneration::no_allocs_since_save_marks() { 3069 return cmsSpace()->no_allocs_since_save_marks(); 3070 } 3071 3072 #define CMS_SINCE_SAVE_MARKS_DEFN(OopClosureType, nv_suffix) \ 3073 \ 3074 void ConcurrentMarkSweepGeneration:: \ 3075 oop_since_save_marks_iterate##nv_suffix(OopClosureType* cl) { \ 3076 cl->set_generation(this); \ 3077 cmsSpace()->oop_since_save_marks_iterate##nv_suffix(cl); \ 3078 cl->reset_generation(); \ 3079 save_marks(); \ 3080 } 3081 3082 ALL_SINCE_SAVE_MARKS_CLOSURES(CMS_SINCE_SAVE_MARKS_DEFN) 3083 3084 void 3085 ConcurrentMarkSweepGeneration::object_iterate_since_last_GC(ObjectClosure* blk) 3086 { 3087 // Not currently implemented; need to do the following. -- ysr. 3088 // dld -- I think that is used for some sort of allocation profiler. So it 3089 // really means the objects allocated by the mutator since the last 3090 // GC. We could potentially implement this cheaply by recording only 3091 // the direct allocations in a side data structure. 3092 // 3093 // I think we probably ought not to be required to support these 3094 // iterations at any arbitrary point; I think there ought to be some 3095 // call to enable/disable allocation profiling in a generation/space, 3096 // and the iterator ought to return the objects allocated in the 3097 // gen/space since the enable call, or the last iterator call (which 3098 // will probably be at a GC.) That way, for gens like CM&S that would 3099 // require some extra data structure to support this, we only pay the 3100 // cost when it's in use... 3101 cmsSpace()->object_iterate_since_last_GC(blk); 3102 } 3103 3104 void 3105 ConcurrentMarkSweepGeneration::younger_refs_iterate(OopsInGenClosure* cl) { 3106 cl->set_generation(this); 3107 younger_refs_in_space_iterate(_cmsSpace, cl); 3108 cl->reset_generation(); 3109 } 3110 3111 void 3112 ConcurrentMarkSweepGeneration::oop_iterate(MemRegion mr, OopClosure* cl) { 3113 if (freelistLock()->owned_by_self()) { 3114 Generation::oop_iterate(mr, cl); 3115 } else { 3116 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 3117 Generation::oop_iterate(mr, cl); 3118 } 3119 } 3120 3121 void 3122 ConcurrentMarkSweepGeneration::oop_iterate(OopClosure* cl) { 3123 if (freelistLock()->owned_by_self()) { 3124 Generation::oop_iterate(cl); 3125 } else { 3126 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 3127 Generation::oop_iterate(cl); 3128 } 3129 } 3130 3131 void 3132 ConcurrentMarkSweepGeneration::object_iterate(ObjectClosure* cl) { 3133 if (freelistLock()->owned_by_self()) { 3134 Generation::object_iterate(cl); 3135 } else { 3136 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 3137 Generation::object_iterate(cl); 3138 } 3139 } 3140 3141 void 3142 ConcurrentMarkSweepGeneration::safe_object_iterate(ObjectClosure* cl) { 3143 if (freelistLock()->owned_by_self()) { 3144 Generation::safe_object_iterate(cl); 3145 } else { 3146 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 3147 Generation::safe_object_iterate(cl); 3148 } 3149 } 3150 3151 void 3152 ConcurrentMarkSweepGeneration::pre_adjust_pointers() { 3153 } 3154 3155 void 3156 ConcurrentMarkSweepGeneration::post_compact() { 3157 } 3158 3159 void 3160 ConcurrentMarkSweepGeneration::prepare_for_verify() { 3161 // Fix the linear allocation blocks to look like free blocks. 3162 3163 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those 3164 // are not called when the heap is verified during universe initialization and 3165 // at vm shutdown. 3166 if (freelistLock()->owned_by_self()) { 3167 cmsSpace()->prepare_for_verify(); 3168 } else { 3169 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag); 3170 cmsSpace()->prepare_for_verify(); 3171 } 3172 } 3173 3174 void 3175 ConcurrentMarkSweepGeneration::verify() { 3176 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those 3177 // are not called when the heap is verified during universe initialization and 3178 // at vm shutdown. 3179 if (freelistLock()->owned_by_self()) { 3180 cmsSpace()->verify(); 3181 } else { 3182 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag); 3183 cmsSpace()->verify(); 3184 } 3185 } 3186 3187 void CMSCollector::verify() { 3188 _cmsGen->verify(); 3189 _permGen->verify(); 3190 } 3191 3192 #ifndef PRODUCT 3193 bool CMSCollector::overflow_list_is_empty() const { 3194 assert(_num_par_pushes >= 0, "Inconsistency"); 3195 if (_overflow_list == NULL) { 3196 assert(_num_par_pushes == 0, "Inconsistency"); 3197 } 3198 return _overflow_list == NULL; 3199 } 3200 3201 // The methods verify_work_stacks_empty() and verify_overflow_empty() 3202 // merely consolidate assertion checks that appear to occur together frequently. 3203 void CMSCollector::verify_work_stacks_empty() const { 3204 assert(_markStack.isEmpty(), "Marking stack should be empty"); 3205 assert(overflow_list_is_empty(), "Overflow list should be empty"); 3206 } 3207 3208 void CMSCollector::verify_overflow_empty() const { 3209 assert(overflow_list_is_empty(), "Overflow list should be empty"); 3210 assert(no_preserved_marks(), "No preserved marks"); 3211 } 3212 #endif // PRODUCT 3213 3214 // Decide if we want to enable class unloading as part of the 3215 // ensuing concurrent GC cycle. We will collect the perm gen and 3216 // unload classes if it's the case that: 3217 // (1) an explicit gc request has been made and the flag 3218 // ExplicitGCInvokesConcurrentAndUnloadsClasses is set, OR 3219 // (2) (a) class unloading is enabled at the command line, and 3220 // (b) (i) perm gen threshold has been crossed, or 3221 // (ii) old gen is getting really full, or 3222 // (iii) the previous N CMS collections did not collect the 3223 // perm gen 3224 // NOTE: Provided there is no change in the state of the heap between 3225 // calls to this method, it should have idempotent results. Moreover, 3226 // its results should be monotonically increasing (i.e. going from 0 to 1, 3227 // but not 1 to 0) between successive calls between which the heap was 3228 // not collected. For the implementation below, it must thus rely on 3229 // the property that concurrent_cycles_since_last_unload() 3230 // will not decrease unless a collection cycle happened and that 3231 // _permGen->should_concurrent_collect() and _cmsGen->is_too_full() are 3232 // themselves also monotonic in that sense. See check_monotonicity() 3233 // below. 3234 bool CMSCollector::update_should_unload_classes() { 3235 _should_unload_classes = false; 3236 // Condition 1 above 3237 if (_full_gc_requested && ExplicitGCInvokesConcurrentAndUnloadsClasses) { 3238 _should_unload_classes = true; 3239 } else if (CMSClassUnloadingEnabled) { // Condition 2.a above 3240 // Disjuncts 2.b.(i,ii,iii) above 3241 _should_unload_classes = (concurrent_cycles_since_last_unload() >= 3242 CMSClassUnloadingMaxInterval) 3243 || _permGen->should_concurrent_collect() 3244 || _cmsGen->is_too_full(); 3245 } 3246 return _should_unload_classes; 3247 } 3248 3249 bool ConcurrentMarkSweepGeneration::is_too_full() const { 3250 bool res = should_concurrent_collect(); 3251 res = res && (occupancy() > (double)CMSIsTooFullPercentage/100.0); 3252 return res; 3253 } 3254 3255 void CMSCollector::setup_cms_unloading_and_verification_state() { 3256 const bool should_verify = VerifyBeforeGC || VerifyAfterGC || VerifyDuringGC 3257 || VerifyBeforeExit; 3258 const int rso = SharedHeap::SO_Strings | SharedHeap::SO_CodeCache; 3259 3260 if (should_unload_classes()) { // Should unload classes this cycle 3261 remove_root_scanning_option(rso); // Shrink the root set appropriately 3262 set_verifying(should_verify); // Set verification state for this cycle 3263 return; // Nothing else needs to be done at this time 3264 } 3265 3266 // Not unloading classes this cycle 3267 assert(!should_unload_classes(), "Inconsitency!"); 3268 if ((!verifying() || unloaded_classes_last_cycle()) && should_verify) { 3269 // We were not verifying, or we _were_ unloading classes in the last cycle, 3270 // AND some verification options are enabled this cycle; in this case, 3271 // we must make sure that the deadness map is allocated if not already so, 3272 // and cleared (if already allocated previously -- 3273 // CMSBitMap::sizeInBits() is used to determine if it's allocated). 3274 if (perm_gen_verify_bit_map()->sizeInBits() == 0) { 3275 if (!perm_gen_verify_bit_map()->allocate(_permGen->reserved())) { 3276 warning("Failed to allocate permanent generation verification CMS Bit Map;\n" 3277 "permanent generation verification disabled"); 3278 return; // Note that we leave verification disabled, so we'll retry this 3279 // allocation next cycle. We _could_ remember this failure 3280 // and skip further attempts and permanently disable verification 3281 // attempts if that is considered more desirable. 3282 } 3283 assert(perm_gen_verify_bit_map()->covers(_permGen->reserved()), 3284 "_perm_gen_ver_bit_map inconsistency?"); 3285 } else { 3286 perm_gen_verify_bit_map()->clear_all(); 3287 } 3288 // Include symbols, strings and code cache elements to prevent their resurrection. 3289 add_root_scanning_option(rso); 3290 set_verifying(true); 3291 } else if (verifying() && !should_verify) { 3292 // We were verifying, but some verification flags got disabled. 3293 set_verifying(false); 3294 // Exclude symbols, strings and code cache elements from root scanning to 3295 // reduce IM and RM pauses. 3296 remove_root_scanning_option(rso); 3297 } 3298 } 3299 3300 3301 #ifndef PRODUCT 3302 HeapWord* CMSCollector::block_start(const void* p) const { 3303 const HeapWord* addr = (HeapWord*)p; 3304 if (_span.contains(p)) { 3305 if (_cmsGen->cmsSpace()->is_in_reserved(addr)) { 3306 return _cmsGen->cmsSpace()->block_start(p); 3307 } else { 3308 assert(_permGen->cmsSpace()->is_in_reserved(addr), 3309 "Inconsistent _span?"); 3310 return _permGen->cmsSpace()->block_start(p); 3311 } 3312 } 3313 return NULL; 3314 } 3315 #endif 3316 3317 HeapWord* 3318 ConcurrentMarkSweepGeneration::expand_and_allocate(size_t word_size, 3319 bool tlab, 3320 bool parallel) { 3321 CMSSynchronousYieldRequest yr; 3322 assert(!tlab, "Can't deal with TLAB allocation"); 3323 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 3324 expand(word_size*HeapWordSize, MinHeapDeltaBytes, 3325 CMSExpansionCause::_satisfy_allocation); 3326 if (GCExpandToAllocateDelayMillis > 0) { 3327 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 3328 } 3329 return have_lock_and_allocate(word_size, tlab); 3330 } 3331 3332 // YSR: All of this generation expansion/shrinking stuff is an exact copy of 3333 // OneContigSpaceCardGeneration, which makes me wonder if we should move this 3334 // to CardGeneration and share it... 3335 bool ConcurrentMarkSweepGeneration::expand(size_t bytes, size_t expand_bytes) { 3336 return CardGeneration::expand(bytes, expand_bytes); 3337 } 3338 3339 void ConcurrentMarkSweepGeneration::expand(size_t bytes, size_t expand_bytes, 3340 CMSExpansionCause::Cause cause) 3341 { 3342 3343 bool success = expand(bytes, expand_bytes); 3344 3345 // remember why we expanded; this information is used 3346 // by shouldConcurrentCollect() when making decisions on whether to start 3347 // a new CMS cycle. 3348 if (success) { 3349 set_expansion_cause(cause); 3350 if (PrintGCDetails && Verbose) { 3351 gclog_or_tty->print_cr("Expanded CMS gen for %s", 3352 CMSExpansionCause::to_string(cause)); 3353 } 3354 } 3355 } 3356 3357 HeapWord* ConcurrentMarkSweepGeneration::expand_and_par_lab_allocate(CMSParGCThreadState* ps, size_t word_sz) { 3358 HeapWord* res = NULL; 3359 MutexLocker x(ParGCRareEvent_lock); 3360 while (true) { 3361 // Expansion by some other thread might make alloc OK now: 3362 res = ps->lab.alloc(word_sz); 3363 if (res != NULL) return res; 3364 // If there's not enough expansion space available, give up. 3365 if (_virtual_space.uncommitted_size() < (word_sz * HeapWordSize)) { 3366 return NULL; 3367 } 3368 // Otherwise, we try expansion. 3369 expand(word_sz*HeapWordSize, MinHeapDeltaBytes, 3370 CMSExpansionCause::_allocate_par_lab); 3371 // Now go around the loop and try alloc again; 3372 // A competing par_promote might beat us to the expansion space, 3373 // so we may go around the loop again if promotion fails agaion. 3374 if (GCExpandToAllocateDelayMillis > 0) { 3375 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 3376 } 3377 } 3378 } 3379 3380 3381 bool ConcurrentMarkSweepGeneration::expand_and_ensure_spooling_space( 3382 PromotionInfo* promo) { 3383 MutexLocker x(ParGCRareEvent_lock); 3384 size_t refill_size_bytes = promo->refillSize() * HeapWordSize; 3385 while (true) { 3386 // Expansion by some other thread might make alloc OK now: 3387 if (promo->ensure_spooling_space()) { 3388 assert(promo->has_spooling_space(), 3389 "Post-condition of successful ensure_spooling_space()"); 3390 return true; 3391 } 3392 // If there's not enough expansion space available, give up. 3393 if (_virtual_space.uncommitted_size() < refill_size_bytes) { 3394 return false; 3395 } 3396 // Otherwise, we try expansion. 3397 expand(refill_size_bytes, MinHeapDeltaBytes, 3398 CMSExpansionCause::_allocate_par_spooling_space); 3399 // Now go around the loop and try alloc again; 3400 // A competing allocation might beat us to the expansion space, 3401 // so we may go around the loop again if allocation fails again. 3402 if (GCExpandToAllocateDelayMillis > 0) { 3403 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 3404 } 3405 } 3406 } 3407 3408 3409 3410 void ConcurrentMarkSweepGeneration::shrink(size_t bytes) { 3411 assert_locked_or_safepoint(Heap_lock); 3412 size_t size = ReservedSpace::page_align_size_down(bytes); 3413 if (size > 0) { 3414 shrink_by(size); 3415 } 3416 } 3417 3418 bool ConcurrentMarkSweepGeneration::grow_by(size_t bytes) { 3419 assert_locked_or_safepoint(Heap_lock); 3420 bool result = _virtual_space.expand_by(bytes); 3421 if (result) { 3422 HeapWord* old_end = _cmsSpace->end(); 3423 size_t new_word_size = 3424 heap_word_size(_virtual_space.committed_size()); 3425 MemRegion mr(_cmsSpace->bottom(), new_word_size); 3426 _bts->resize(new_word_size); // resize the block offset shared array 3427 Universe::heap()->barrier_set()->resize_covered_region(mr); 3428 // Hmmmm... why doesn't CFLS::set_end verify locking? 3429 // This is quite ugly; FIX ME XXX 3430 _cmsSpace->assert_locked(freelistLock()); 3431 _cmsSpace->set_end((HeapWord*)_virtual_space.high()); 3432 3433 // update the space and generation capacity counters 3434 if (UsePerfData) { 3435 _space_counters->update_capacity(); 3436 _gen_counters->update_all(); 3437 } 3438 3439 if (Verbose && PrintGC) { 3440 size_t new_mem_size = _virtual_space.committed_size(); 3441 size_t old_mem_size = new_mem_size - bytes; 3442 gclog_or_tty->print_cr("Expanding %s from %ldK by %ldK to %ldK", 3443 name(), old_mem_size/K, bytes/K, new_mem_size/K); 3444 } 3445 } 3446 return result; 3447 } 3448 3449 bool ConcurrentMarkSweepGeneration::grow_to_reserved() { 3450 assert_locked_or_safepoint(Heap_lock); 3451 bool success = true; 3452 const size_t remaining_bytes = _virtual_space.uncommitted_size(); 3453 if (remaining_bytes > 0) { 3454 success = grow_by(remaining_bytes); 3455 DEBUG_ONLY(if (!success) warning("grow to reserved failed");) 3456 } 3457 return success; 3458 } 3459 3460 void ConcurrentMarkSweepGeneration::shrink_by(size_t bytes) { 3461 assert_locked_or_safepoint(Heap_lock); 3462 assert_lock_strong(freelistLock()); 3463 // XXX Fix when compaction is implemented. 3464 warning("Shrinking of CMS not yet implemented"); 3465 return; 3466 } 3467 3468 3469 // Simple ctor/dtor wrapper for accounting & timer chores around concurrent 3470 // phases. 3471 class CMSPhaseAccounting: public StackObj { 3472 public: 3473 CMSPhaseAccounting(CMSCollector *collector, 3474 const char *phase, 3475 bool print_cr = true); 3476 ~CMSPhaseAccounting(); 3477 3478 private: 3479 CMSCollector *_collector; 3480 const char *_phase; 3481 elapsedTimer _wallclock; 3482 bool _print_cr; 3483 3484 public: 3485 // Not MT-safe; so do not pass around these StackObj's 3486 // where they may be accessed by other threads. 3487 jlong wallclock_millis() { 3488 assert(_wallclock.is_active(), "Wall clock should not stop"); 3489 _wallclock.stop(); // to record time 3490 jlong ret = _wallclock.milliseconds(); 3491 _wallclock.start(); // restart 3492 return ret; 3493 } 3494 }; 3495 3496 CMSPhaseAccounting::CMSPhaseAccounting(CMSCollector *collector, 3497 const char *phase, 3498 bool print_cr) : 3499 _collector(collector), _phase(phase), _print_cr(print_cr) { 3500 3501 if (PrintCMSStatistics != 0) { 3502 _collector->resetYields(); 3503 } 3504 if (PrintGCDetails && PrintGCTimeStamps) { 3505 gclog_or_tty->date_stamp(PrintGCDateStamps); 3506 gclog_or_tty->stamp(); 3507 gclog_or_tty->print_cr(": [%s-concurrent-%s-start]", 3508 _collector->cmsGen()->short_name(), _phase); 3509 } 3510 _collector->resetTimer(); 3511 _wallclock.start(); 3512 _collector->startTimer(); 3513 } 3514 3515 CMSPhaseAccounting::~CMSPhaseAccounting() { 3516 assert(_wallclock.is_active(), "Wall clock should not have stopped"); 3517 _collector->stopTimer(); 3518 _wallclock.stop(); 3519 if (PrintGCDetails) { 3520 gclog_or_tty->date_stamp(PrintGCDateStamps); 3521 gclog_or_tty->stamp(PrintGCTimeStamps); 3522 gclog_or_tty->print("[%s-concurrent-%s: %3.3f/%3.3f secs]", 3523 _collector->cmsGen()->short_name(), 3524 _phase, _collector->timerValue(), _wallclock.seconds()); 3525 if (_print_cr) { 3526 gclog_or_tty->print_cr(""); 3527 } 3528 if (PrintCMSStatistics != 0) { 3529 gclog_or_tty->print_cr(" (CMS-concurrent-%s yielded %d times)", _phase, 3530 _collector->yields()); 3531 } 3532 } 3533 } 3534 3535 // CMS work 3536 3537 // Checkpoint the roots into this generation from outside 3538 // this generation. [Note this initial checkpoint need only 3539 // be approximate -- we'll do a catch up phase subsequently.] 3540 void CMSCollector::checkpointRootsInitial(bool asynch) { 3541 assert(_collectorState == InitialMarking, "Wrong collector state"); 3542 check_correct_thread_executing(); 3543 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 3544 3545 save_heap_summary(); 3546 report_heap_summary(GCWhen::BeforeGC); 3547 3548 ReferenceProcessor* rp = ref_processor(); 3549 SpecializationStats::clear(); 3550 assert(_restart_addr == NULL, "Control point invariant"); 3551 if (asynch) { 3552 // acquire locks for subsequent manipulations 3553 MutexLockerEx x(bitMapLock(), 3554 Mutex::_no_safepoint_check_flag); 3555 checkpointRootsInitialWork(asynch); 3556 // enable ("weak") refs discovery 3557 rp->enable_discovery(true /*verify_disabled*/, true /*check_no_refs*/); 3558 _collectorState = Marking; 3559 } else { 3560 // (Weak) Refs discovery: this is controlled from genCollectedHeap::do_collection 3561 // which recognizes if we are a CMS generation, and doesn't try to turn on 3562 // discovery; verify that they aren't meddling. 3563 assert(!rp->discovery_is_atomic(), 3564 "incorrect setting of discovery predicate"); 3565 assert(!rp->discovery_enabled(), "genCollectedHeap shouldn't control " 3566 "ref discovery for this generation kind"); 3567 // already have locks 3568 checkpointRootsInitialWork(asynch); 3569 // now enable ("weak") refs discovery 3570 rp->enable_discovery(true /*verify_disabled*/, false /*verify_no_refs*/); 3571 _collectorState = Marking; 3572 } 3573 SpecializationStats::print(); 3574 } 3575 3576 void CMSCollector::checkpointRootsInitialWork(bool asynch) { 3577 assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped"); 3578 assert(_collectorState == InitialMarking, "just checking"); 3579 3580 // If there has not been a GC[n-1] since last GC[n] cycle completed, 3581 // precede our marking with a collection of all 3582 // younger generations to keep floating garbage to a minimum. 3583 // XXX: we won't do this for now -- it's an optimization to be done later. 3584 3585 // already have locks 3586 assert_lock_strong(bitMapLock()); 3587 assert(_markBitMap.isAllClear(), "was reset at end of previous cycle"); 3588 3589 // Setup the verification and class unloading state for this 3590 // CMS collection cycle. 3591 setup_cms_unloading_and_verification_state(); 3592 3593 NOT_PRODUCT(GCTraceTime t("\ncheckpointRootsInitialWork", 3594 PrintGCDetails && Verbose, true, _gc_timer_cm);) 3595 if (UseAdaptiveSizePolicy) { 3596 size_policy()->checkpoint_roots_initial_begin(); 3597 } 3598 3599 // Reset all the PLAB chunk arrays if necessary. 3600 if (_survivor_plab_array != NULL && !CMSPLABRecordAlways) { 3601 reset_survivor_plab_arrays(); 3602 } 3603 3604 ResourceMark rm; 3605 HandleMark hm; 3606 3607 FalseClosure falseClosure; 3608 // In the case of a synchronous collection, we will elide the 3609 // remark step, so it's important to catch all the nmethod oops 3610 // in this step. 3611 // The final 'true' flag to gen_process_strong_roots will ensure this. 3612 // If 'async' is true, we can relax the nmethod tracing. 3613 MarkRefsIntoClosure notOlder(_span, &_markBitMap); 3614 GenCollectedHeap* gch = GenCollectedHeap::heap(); 3615 3616 verify_work_stacks_empty(); 3617 verify_overflow_empty(); 3618 3619 gch->ensure_parsability(false); // fill TLABs, but no need to retire them 3620 // Update the saved marks which may affect the root scans. 3621 gch->save_marks(); 3622 3623 // weak reference processing has not started yet. 3624 ref_processor()->set_enqueuing_is_done(false); 3625 3626 { 3627 // This is not needed. DEBUG_ONLY(RememberKlassesChecker imx(true);) 3628 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;) 3629 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 3630 gch->gen_process_strong_roots(_cmsGen->level(), 3631 true, // younger gens are roots 3632 true, // activate StrongRootsScope 3633 true, // collecting perm gen 3634 SharedHeap::ScanningOption(roots_scanning_options()), 3635 ¬Older, 3636 true, // walk all of code cache if (so & SO_CodeCache) 3637 NULL); 3638 } 3639 3640 // Clear mod-union table; it will be dirtied in the prologue of 3641 // CMS generation per each younger generation collection. 3642 3643 assert(_modUnionTable.isAllClear(), 3644 "Was cleared in most recent final checkpoint phase" 3645 " or no bits are set in the gc_prologue before the start of the next " 3646 "subsequent marking phase."); 3647 3648 // Save the end of the used_region of the constituent generations 3649 // to be used to limit the extent of sweep in each generation. 3650 save_sweep_limits(); 3651 if (UseAdaptiveSizePolicy) { 3652 size_policy()->checkpoint_roots_initial_end(gch->gc_cause()); 3653 } 3654 verify_overflow_empty(); 3655 } 3656 3657 bool CMSCollector::markFromRoots(bool asynch) { 3658 // we might be tempted to assert that: 3659 // assert(asynch == !SafepointSynchronize::is_at_safepoint(), 3660 // "inconsistent argument?"); 3661 // However that wouldn't be right, because it's possible that 3662 // a safepoint is indeed in progress as a younger generation 3663 // stop-the-world GC happens even as we mark in this generation. 3664 assert(_collectorState == Marking, "inconsistent state?"); 3665 check_correct_thread_executing(); 3666 verify_overflow_empty(); 3667 3668 bool res; 3669 if (asynch) { 3670 3671 // Start the timers for adaptive size policy for the concurrent phases 3672 // Do it here so that the foreground MS can use the concurrent 3673 // timer since a foreground MS might has the sweep done concurrently 3674 // or STW. 3675 if (UseAdaptiveSizePolicy) { 3676 size_policy()->concurrent_marking_begin(); 3677 } 3678 3679 // Weak ref discovery note: We may be discovering weak 3680 // refs in this generation concurrent (but interleaved) with 3681 // weak ref discovery by a younger generation collector. 3682 3683 CMSTokenSyncWithLocks ts(true, bitMapLock()); 3684 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 3685 CMSPhaseAccounting pa(this, "mark", !PrintGCDetails); 3686 res = markFromRootsWork(asynch); 3687 if (res) { 3688 _collectorState = Precleaning; 3689 } else { // We failed and a foreground collection wants to take over 3690 assert(_foregroundGCIsActive, "internal state inconsistency"); 3691 assert(_restart_addr == NULL, "foreground will restart from scratch"); 3692 if (PrintGCDetails) { 3693 gclog_or_tty->print_cr("bailing out to foreground collection"); 3694 } 3695 } 3696 if (UseAdaptiveSizePolicy) { 3697 size_policy()->concurrent_marking_end(); 3698 } 3699 } else { 3700 assert(SafepointSynchronize::is_at_safepoint(), 3701 "inconsistent with asynch == false"); 3702 if (UseAdaptiveSizePolicy) { 3703 size_policy()->ms_collection_marking_begin(); 3704 } 3705 // already have locks 3706 res = markFromRootsWork(asynch); 3707 _collectorState = FinalMarking; 3708 if (UseAdaptiveSizePolicy) { 3709 GenCollectedHeap* gch = GenCollectedHeap::heap(); 3710 size_policy()->ms_collection_marking_end(gch->gc_cause()); 3711 } 3712 } 3713 verify_overflow_empty(); 3714 return res; 3715 } 3716 3717 bool CMSCollector::markFromRootsWork(bool asynch) { 3718 // iterate over marked bits in bit map, doing a full scan and mark 3719 // from these roots using the following algorithm: 3720 // . if oop is to the right of the current scan pointer, 3721 // mark corresponding bit (we'll process it later) 3722 // . else (oop is to left of current scan pointer) 3723 // push oop on marking stack 3724 // . drain the marking stack 3725 3726 // Note that when we do a marking step we need to hold the 3727 // bit map lock -- recall that direct allocation (by mutators) 3728 // and promotion (by younger generation collectors) is also 3729 // marking the bit map. [the so-called allocate live policy.] 3730 // Because the implementation of bit map marking is not 3731 // robust wrt simultaneous marking of bits in the same word, 3732 // we need to make sure that there is no such interference 3733 // between concurrent such updates. 3734 3735 // already have locks 3736 assert_lock_strong(bitMapLock()); 3737 3738 // Clear the revisit stack, just in case there are any 3739 // obsolete contents from a short-circuited previous CMS cycle. 3740 _revisitStack.reset(); 3741 verify_work_stacks_empty(); 3742 verify_overflow_empty(); 3743 assert(_revisitStack.isEmpty(), "tabula rasa"); 3744 DEBUG_ONLY(RememberKlassesChecker cmx(should_unload_classes());) 3745 bool result = false; 3746 if (CMSConcurrentMTEnabled && ConcGCThreads > 0) { 3747 result = do_marking_mt(asynch); 3748 } else { 3749 result = do_marking_st(asynch); 3750 } 3751 return result; 3752 } 3753 3754 // Forward decl 3755 class CMSConcMarkingTask; 3756 3757 class CMSConcMarkingTerminator: public ParallelTaskTerminator { 3758 CMSCollector* _collector; 3759 CMSConcMarkingTask* _task; 3760 public: 3761 virtual void yield(); 3762 3763 // "n_threads" is the number of threads to be terminated. 3764 // "queue_set" is a set of work queues of other threads. 3765 // "collector" is the CMS collector associated with this task terminator. 3766 // "yield" indicates whether we need the gang as a whole to yield. 3767 CMSConcMarkingTerminator(int n_threads, TaskQueueSetSuper* queue_set, CMSCollector* collector) : 3768 ParallelTaskTerminator(n_threads, queue_set), 3769 _collector(collector) { } 3770 3771 void set_task(CMSConcMarkingTask* task) { 3772 _task = task; 3773 } 3774 }; 3775 3776 class CMSConcMarkingTerminatorTerminator: public TerminatorTerminator { 3777 CMSConcMarkingTask* _task; 3778 public: 3779 bool should_exit_termination(); 3780 void set_task(CMSConcMarkingTask* task) { 3781 _task = task; 3782 } 3783 }; 3784 3785 // MT Concurrent Marking Task 3786 class CMSConcMarkingTask: public YieldingFlexibleGangTask { 3787 CMSCollector* _collector; 3788 int _n_workers; // requested/desired # workers 3789 bool _asynch; 3790 bool _result; 3791 CompactibleFreeListSpace* _cms_space; 3792 CompactibleFreeListSpace* _perm_space; 3793 char _pad_front[64]; // padding to ... 3794 HeapWord* _global_finger; // ... avoid sharing cache line 3795 char _pad_back[64]; 3796 HeapWord* _restart_addr; 3797 3798 // Exposed here for yielding support 3799 Mutex* const _bit_map_lock; 3800 3801 // The per thread work queues, available here for stealing 3802 OopTaskQueueSet* _task_queues; 3803 3804 // Termination (and yielding) support 3805 CMSConcMarkingTerminator _term; 3806 CMSConcMarkingTerminatorTerminator _term_term; 3807 3808 public: 3809 CMSConcMarkingTask(CMSCollector* collector, 3810 CompactibleFreeListSpace* cms_space, 3811 CompactibleFreeListSpace* perm_space, 3812 bool asynch, 3813 YieldingFlexibleWorkGang* workers, 3814 OopTaskQueueSet* task_queues): 3815 YieldingFlexibleGangTask("Concurrent marking done multi-threaded"), 3816 _collector(collector), 3817 _cms_space(cms_space), 3818 _perm_space(perm_space), 3819 _asynch(asynch), _n_workers(0), _result(true), 3820 _task_queues(task_queues), 3821 _term(_n_workers, task_queues, _collector), 3822 _bit_map_lock(collector->bitMapLock()) 3823 { 3824 _requested_size = _n_workers; 3825 _term.set_task(this); 3826 _term_term.set_task(this); 3827 assert(_cms_space->bottom() < _perm_space->bottom(), 3828 "Finger incorrectly initialized below"); 3829 _restart_addr = _global_finger = _cms_space->bottom(); 3830 } 3831 3832 3833 OopTaskQueueSet* task_queues() { return _task_queues; } 3834 3835 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 3836 3837 HeapWord** global_finger_addr() { return &_global_finger; } 3838 3839 CMSConcMarkingTerminator* terminator() { return &_term; } 3840 3841 virtual void set_for_termination(int active_workers) { 3842 terminator()->reset_for_reuse(active_workers); 3843 } 3844 3845 void work(uint worker_id); 3846 bool should_yield() { 3847 return ConcurrentMarkSweepThread::should_yield() 3848 && !_collector->foregroundGCIsActive() 3849 && _asynch; 3850 } 3851 3852 virtual void coordinator_yield(); // stuff done by coordinator 3853 bool result() { return _result; } 3854 3855 void reset(HeapWord* ra) { 3856 assert(_global_finger >= _cms_space->end(), "Postcondition of ::work(i)"); 3857 assert(_global_finger >= _perm_space->end(), "Postcondition of ::work(i)"); 3858 assert(ra < _perm_space->end(), "ra too large"); 3859 _restart_addr = _global_finger = ra; 3860 _term.reset_for_reuse(); 3861 } 3862 3863 static bool get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 3864 OopTaskQueue* work_q); 3865 3866 private: 3867 void do_scan_and_mark(int i, CompactibleFreeListSpace* sp); 3868 void do_work_steal(int i); 3869 void bump_global_finger(HeapWord* f); 3870 }; 3871 3872 bool CMSConcMarkingTerminatorTerminator::should_exit_termination() { 3873 assert(_task != NULL, "Error"); 3874 return _task->yielding(); 3875 // Note that we do not need the disjunct || _task->should_yield() above 3876 // because we want terminating threads to yield only if the task 3877 // is already in the midst of yielding, which happens only after at least one 3878 // thread has yielded. 3879 } 3880 3881 void CMSConcMarkingTerminator::yield() { 3882 if (_task->should_yield()) { 3883 _task->yield(); 3884 } else { 3885 ParallelTaskTerminator::yield(); 3886 } 3887 } 3888 3889 //////////////////////////////////////////////////////////////// 3890 // Concurrent Marking Algorithm Sketch 3891 //////////////////////////////////////////////////////////////// 3892 // Until all tasks exhausted (both spaces): 3893 // -- claim next available chunk 3894 // -- bump global finger via CAS 3895 // -- find first object that starts in this chunk 3896 // and start scanning bitmap from that position 3897 // -- scan marked objects for oops 3898 // -- CAS-mark target, and if successful: 3899 // . if target oop is above global finger (volatile read) 3900 // nothing to do 3901 // . if target oop is in chunk and above local finger 3902 // then nothing to do 3903 // . else push on work-queue 3904 // -- Deal with possible overflow issues: 3905 // . local work-queue overflow causes stuff to be pushed on 3906 // global (common) overflow queue 3907 // . always first empty local work queue 3908 // . then get a batch of oops from global work queue if any 3909 // . then do work stealing 3910 // -- When all tasks claimed (both spaces) 3911 // and local work queue empty, 3912 // then in a loop do: 3913 // . check global overflow stack; steal a batch of oops and trace 3914 // . try to steal from other threads oif GOS is empty 3915 // . if neither is available, offer termination 3916 // -- Terminate and return result 3917 // 3918 void CMSConcMarkingTask::work(uint worker_id) { 3919 elapsedTimer _timer; 3920 ResourceMark rm; 3921 HandleMark hm; 3922 3923 DEBUG_ONLY(_collector->verify_overflow_empty();) 3924 3925 // Before we begin work, our work queue should be empty 3926 assert(work_queue(worker_id)->size() == 0, "Expected to be empty"); 3927 // Scan the bitmap covering _cms_space, tracing through grey objects. 3928 _timer.start(); 3929 do_scan_and_mark(worker_id, _cms_space); 3930 _timer.stop(); 3931 if (PrintCMSStatistics != 0) { 3932 gclog_or_tty->print_cr("Finished cms space scanning in %dth thread: %3.3f sec", 3933 worker_id, _timer.seconds()); 3934 // XXX: need xxx/xxx type of notation, two timers 3935 } 3936 3937 // ... do the same for the _perm_space 3938 _timer.reset(); 3939 _timer.start(); 3940 do_scan_and_mark(worker_id, _perm_space); 3941 _timer.stop(); 3942 if (PrintCMSStatistics != 0) { 3943 gclog_or_tty->print_cr("Finished perm space scanning in %dth thread: %3.3f sec", 3944 worker_id, _timer.seconds()); 3945 // XXX: need xxx/xxx type of notation, two timers 3946 } 3947 3948 // ... do work stealing 3949 _timer.reset(); 3950 _timer.start(); 3951 do_work_steal(worker_id); 3952 _timer.stop(); 3953 if (PrintCMSStatistics != 0) { 3954 gclog_or_tty->print_cr("Finished work stealing in %dth thread: %3.3f sec", 3955 worker_id, _timer.seconds()); 3956 // XXX: need xxx/xxx type of notation, two timers 3957 } 3958 assert(_collector->_markStack.isEmpty(), "Should have been emptied"); 3959 assert(work_queue(worker_id)->size() == 0, "Should have been emptied"); 3960 // Note that under the current task protocol, the 3961 // following assertion is true even of the spaces 3962 // expanded since the completion of the concurrent 3963 // marking. XXX This will likely change under a strict 3964 // ABORT semantics. 3965 assert(_global_finger > _cms_space->end() && 3966 _global_finger >= _perm_space->end(), 3967 "All tasks have been completed"); 3968 DEBUG_ONLY(_collector->verify_overflow_empty();) 3969 } 3970 3971 void CMSConcMarkingTask::bump_global_finger(HeapWord* f) { 3972 HeapWord* read = _global_finger; 3973 HeapWord* cur = read; 3974 while (f > read) { 3975 cur = read; 3976 read = (HeapWord*) Atomic::cmpxchg_ptr(f, &_global_finger, cur); 3977 if (cur == read) { 3978 // our cas succeeded 3979 assert(_global_finger >= f, "protocol consistency"); 3980 break; 3981 } 3982 } 3983 } 3984 3985 // This is really inefficient, and should be redone by 3986 // using (not yet available) block-read and -write interfaces to the 3987 // stack and the work_queue. XXX FIX ME !!! 3988 bool CMSConcMarkingTask::get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 3989 OopTaskQueue* work_q) { 3990 // Fast lock-free check 3991 if (ovflw_stk->length() == 0) { 3992 return false; 3993 } 3994 assert(work_q->size() == 0, "Shouldn't steal"); 3995 MutexLockerEx ml(ovflw_stk->par_lock(), 3996 Mutex::_no_safepoint_check_flag); 3997 // Grab up to 1/4 the size of the work queue 3998 size_t num = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 3999 (size_t)ParGCDesiredObjsFromOverflowList); 4000 num = MIN2(num, ovflw_stk->length()); 4001 for (int i = (int) num; i > 0; i--) { 4002 oop cur = ovflw_stk->pop(); 4003 assert(cur != NULL, "Counted wrong?"); 4004 work_q->push(cur); 4005 } 4006 return num > 0; 4007 } 4008 4009 void CMSConcMarkingTask::do_scan_and_mark(int i, CompactibleFreeListSpace* sp) { 4010 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 4011 int n_tasks = pst->n_tasks(); 4012 // We allow that there may be no tasks to do here because 4013 // we are restarting after a stack overflow. 4014 assert(pst->valid() || n_tasks == 0, "Uninitialized use?"); 4015 uint nth_task = 0; 4016 4017 HeapWord* aligned_start = sp->bottom(); 4018 if (sp->used_region().contains(_restart_addr)) { 4019 // Align down to a card boundary for the start of 0th task 4020 // for this space. 4021 aligned_start = 4022 (HeapWord*)align_size_down((uintptr_t)_restart_addr, 4023 CardTableModRefBS::card_size); 4024 } 4025 4026 size_t chunk_size = sp->marking_task_size(); 4027 while (!pst->is_task_claimed(/* reference */ nth_task)) { 4028 // Having claimed the nth task in this space, 4029 // compute the chunk that it corresponds to: 4030 MemRegion span = MemRegion(aligned_start + nth_task*chunk_size, 4031 aligned_start + (nth_task+1)*chunk_size); 4032 // Try and bump the global finger via a CAS; 4033 // note that we need to do the global finger bump 4034 // _before_ taking the intersection below, because 4035 // the task corresponding to that region will be 4036 // deemed done even if the used_region() expands 4037 // because of allocation -- as it almost certainly will 4038 // during start-up while the threads yield in the 4039 // closure below. 4040 HeapWord* finger = span.end(); 4041 bump_global_finger(finger); // atomically 4042 // There are null tasks here corresponding to chunks 4043 // beyond the "top" address of the space. 4044 span = span.intersection(sp->used_region()); 4045 if (!span.is_empty()) { // Non-null task 4046 HeapWord* prev_obj; 4047 assert(!span.contains(_restart_addr) || nth_task == 0, 4048 "Inconsistency"); 4049 if (nth_task == 0) { 4050 // For the 0th task, we'll not need to compute a block_start. 4051 if (span.contains(_restart_addr)) { 4052 // In the case of a restart because of stack overflow, 4053 // we might additionally skip a chunk prefix. 4054 prev_obj = _restart_addr; 4055 } else { 4056 prev_obj = span.start(); 4057 } 4058 } else { 4059 // We want to skip the first object because 4060 // the protocol is to scan any object in its entirety 4061 // that _starts_ in this span; a fortiori, any 4062 // object starting in an earlier span is scanned 4063 // as part of an earlier claimed task. 4064 // Below we use the "careful" version of block_start 4065 // so we do not try to navigate uninitialized objects. 4066 prev_obj = sp->block_start_careful(span.start()); 4067 // Below we use a variant of block_size that uses the 4068 // Printezis bits to avoid waiting for allocated 4069 // objects to become initialized/parsable. 4070 while (prev_obj < span.start()) { 4071 size_t sz = sp->block_size_no_stall(prev_obj, _collector); 4072 if (sz > 0) { 4073 prev_obj += sz; 4074 } else { 4075 // In this case we may end up doing a bit of redundant 4076 // scanning, but that appears unavoidable, short of 4077 // locking the free list locks; see bug 6324141. 4078 break; 4079 } 4080 } 4081 } 4082 if (prev_obj < span.end()) { 4083 MemRegion my_span = MemRegion(prev_obj, span.end()); 4084 // Do the marking work within a non-empty span -- 4085 // the last argument to the constructor indicates whether the 4086 // iteration should be incremental with periodic yields. 4087 Par_MarkFromRootsClosure cl(this, _collector, my_span, 4088 &_collector->_markBitMap, 4089 work_queue(i), 4090 &_collector->_markStack, 4091 &_collector->_revisitStack, 4092 _asynch); 4093 _collector->_markBitMap.iterate(&cl, my_span.start(), my_span.end()); 4094 } // else nothing to do for this task 4095 } // else nothing to do for this task 4096 } 4097 // We'd be tempted to assert here that since there are no 4098 // more tasks left to claim in this space, the global_finger 4099 // must exceed space->top() and a fortiori space->end(). However, 4100 // that would not quite be correct because the bumping of 4101 // global_finger occurs strictly after the claiming of a task, 4102 // so by the time we reach here the global finger may not yet 4103 // have been bumped up by the thread that claimed the last 4104 // task. 4105 pst->all_tasks_completed(); 4106 } 4107 4108 class Par_ConcMarkingClosure: public Par_KlassRememberingOopClosure { 4109 private: 4110 CMSConcMarkingTask* _task; 4111 MemRegion _span; 4112 CMSBitMap* _bit_map; 4113 CMSMarkStack* _overflow_stack; 4114 OopTaskQueue* _work_queue; 4115 protected: 4116 DO_OOP_WORK_DEFN 4117 public: 4118 Par_ConcMarkingClosure(CMSCollector* collector, CMSConcMarkingTask* task, OopTaskQueue* work_queue, 4119 CMSBitMap* bit_map, CMSMarkStack* overflow_stack, 4120 CMSMarkStack* revisit_stack): 4121 Par_KlassRememberingOopClosure(collector, collector->ref_processor(), revisit_stack), 4122 _task(task), 4123 _span(collector->_span), 4124 _work_queue(work_queue), 4125 _bit_map(bit_map), 4126 _overflow_stack(overflow_stack) 4127 { } 4128 virtual void do_oop(oop* p); 4129 virtual void do_oop(narrowOop* p); 4130 void trim_queue(size_t max); 4131 void handle_stack_overflow(HeapWord* lost); 4132 void do_yield_check() { 4133 if (_task->should_yield()) { 4134 _task->yield(); 4135 } 4136 } 4137 }; 4138 4139 // Grey object scanning during work stealing phase -- 4140 // the salient assumption here is that any references 4141 // that are in these stolen objects being scanned must 4142 // already have been initialized (else they would not have 4143 // been published), so we do not need to check for 4144 // uninitialized objects before pushing here. 4145 void Par_ConcMarkingClosure::do_oop(oop obj) { 4146 assert(obj->is_oop_or_null(true), "expected an oop or NULL"); 4147 HeapWord* addr = (HeapWord*)obj; 4148 // Check if oop points into the CMS generation 4149 // and is not marked 4150 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 4151 // a white object ... 4152 // If we manage to "claim" the object, by being the 4153 // first thread to mark it, then we push it on our 4154 // marking stack 4155 if (_bit_map->par_mark(addr)) { // ... now grey 4156 // push on work queue (grey set) 4157 bool simulate_overflow = false; 4158 NOT_PRODUCT( 4159 if (CMSMarkStackOverflowALot && 4160 _collector->simulate_overflow()) { 4161 // simulate a stack overflow 4162 simulate_overflow = true; 4163 } 4164 ) 4165 if (simulate_overflow || 4166 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) { 4167 // stack overflow 4168 if (PrintCMSStatistics != 0) { 4169 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 4170 SIZE_FORMAT, _overflow_stack->capacity()); 4171 } 4172 // We cannot assert that the overflow stack is full because 4173 // it may have been emptied since. 4174 assert(simulate_overflow || 4175 _work_queue->size() == _work_queue->max_elems(), 4176 "Else push should have succeeded"); 4177 handle_stack_overflow(addr); 4178 } 4179 } // Else, some other thread got there first 4180 do_yield_check(); 4181 } 4182 } 4183 4184 void Par_ConcMarkingClosure::do_oop(oop* p) { Par_ConcMarkingClosure::do_oop_work(p); } 4185 void Par_ConcMarkingClosure::do_oop(narrowOop* p) { Par_ConcMarkingClosure::do_oop_work(p); } 4186 4187 void Par_ConcMarkingClosure::trim_queue(size_t max) { 4188 while (_work_queue->size() > max) { 4189 oop new_oop; 4190 if (_work_queue->pop_local(new_oop)) { 4191 assert(new_oop->is_oop(), "Should be an oop"); 4192 assert(_bit_map->isMarked((HeapWord*)new_oop), "Grey object"); 4193 assert(_span.contains((HeapWord*)new_oop), "Not in span"); 4194 assert(new_oop->is_parsable(), "Should be parsable"); 4195 new_oop->oop_iterate(this); // do_oop() above 4196 do_yield_check(); 4197 } 4198 } 4199 } 4200 4201 // Upon stack overflow, we discard (part of) the stack, 4202 // remembering the least address amongst those discarded 4203 // in CMSCollector's _restart_address. 4204 void Par_ConcMarkingClosure::handle_stack_overflow(HeapWord* lost) { 4205 // We need to do this under a mutex to prevent other 4206 // workers from interfering with the work done below. 4207 MutexLockerEx ml(_overflow_stack->par_lock(), 4208 Mutex::_no_safepoint_check_flag); 4209 // Remember the least grey address discarded 4210 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 4211 _collector->lower_restart_addr(ra); 4212 _overflow_stack->reset(); // discard stack contents 4213 _overflow_stack->expand(); // expand the stack if possible 4214 } 4215 4216 4217 void CMSConcMarkingTask::do_work_steal(int i) { 4218 OopTaskQueue* work_q = work_queue(i); 4219 oop obj_to_scan; 4220 CMSBitMap* bm = &(_collector->_markBitMap); 4221 CMSMarkStack* ovflw = &(_collector->_markStack); 4222 CMSMarkStack* revisit = &(_collector->_revisitStack); 4223 int* seed = _collector->hash_seed(i); 4224 Par_ConcMarkingClosure cl(_collector, this, work_q, bm, ovflw, revisit); 4225 while (true) { 4226 cl.trim_queue(0); 4227 assert(work_q->size() == 0, "Should have been emptied above"); 4228 if (get_work_from_overflow_stack(ovflw, work_q)) { 4229 // Can't assert below because the work obtained from the 4230 // overflow stack may already have been stolen from us. 4231 // assert(work_q->size() > 0, "Work from overflow stack"); 4232 continue; 4233 } else if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 4234 assert(obj_to_scan->is_oop(), "Should be an oop"); 4235 assert(bm->isMarked((HeapWord*)obj_to_scan), "Grey object"); 4236 obj_to_scan->oop_iterate(&cl); 4237 } else if (terminator()->offer_termination(&_term_term)) { 4238 assert(work_q->size() == 0, "Impossible!"); 4239 break; 4240 } else if (yielding() || should_yield()) { 4241 yield(); 4242 } 4243 } 4244 } 4245 4246 // This is run by the CMS (coordinator) thread. 4247 void CMSConcMarkingTask::coordinator_yield() { 4248 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 4249 "CMS thread should hold CMS token"); 4250 DEBUG_ONLY(RememberKlassesChecker mux(false);) 4251 // First give up the locks, then yield, then re-lock 4252 // We should probably use a constructor/destructor idiom to 4253 // do this unlock/lock or modify the MutexUnlocker class to 4254 // serve our purpose. XXX 4255 assert_lock_strong(_bit_map_lock); 4256 _bit_map_lock->unlock(); 4257 ConcurrentMarkSweepThread::desynchronize(true); 4258 ConcurrentMarkSweepThread::acknowledge_yield_request(); 4259 _collector->stopTimer(); 4260 if (PrintCMSStatistics != 0) { 4261 _collector->incrementYields(); 4262 } 4263 _collector->icms_wait(); 4264 4265 // It is possible for whichever thread initiated the yield request 4266 // not to get a chance to wake up and take the bitmap lock between 4267 // this thread releasing it and reacquiring it. So, while the 4268 // should_yield() flag is on, let's sleep for a bit to give the 4269 // other thread a chance to wake up. The limit imposed on the number 4270 // of iterations is defensive, to avoid any unforseen circumstances 4271 // putting us into an infinite loop. Since it's always been this 4272 // (coordinator_yield()) method that was observed to cause the 4273 // problem, we are using a parameter (CMSCoordinatorYieldSleepCount) 4274 // which is by default non-zero. For the other seven methods that 4275 // also perform the yield operation, as are using a different 4276 // parameter (CMSYieldSleepCount) which is by default zero. This way we 4277 // can enable the sleeping for those methods too, if necessary. 4278 // See 6442774. 4279 // 4280 // We really need to reconsider the synchronization between the GC 4281 // thread and the yield-requesting threads in the future and we 4282 // should really use wait/notify, which is the recommended 4283 // way of doing this type of interaction. Additionally, we should 4284 // consolidate the eight methods that do the yield operation and they 4285 // are almost identical into one for better maintenability and 4286 // readability. See 6445193. 4287 // 4288 // Tony 2006.06.29 4289 for (unsigned i = 0; i < CMSCoordinatorYieldSleepCount && 4290 ConcurrentMarkSweepThread::should_yield() && 4291 !CMSCollector::foregroundGCIsActive(); ++i) { 4292 os::sleep(Thread::current(), 1, false); 4293 ConcurrentMarkSweepThread::acknowledge_yield_request(); 4294 } 4295 4296 ConcurrentMarkSweepThread::synchronize(true); 4297 _bit_map_lock->lock_without_safepoint_check(); 4298 _collector->startTimer(); 4299 } 4300 4301 bool CMSCollector::do_marking_mt(bool asynch) { 4302 assert(ConcGCThreads > 0 && conc_workers() != NULL, "precondition"); 4303 int num_workers = AdaptiveSizePolicy::calc_active_conc_workers( 4304 conc_workers()->total_workers(), 4305 conc_workers()->active_workers(), 4306 Threads::number_of_non_daemon_threads()); 4307 conc_workers()->set_active_workers(num_workers); 4308 4309 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 4310 CompactibleFreeListSpace* perm_space = _permGen->cmsSpace(); 4311 4312 CMSConcMarkingTask tsk(this, 4313 cms_space, 4314 perm_space, 4315 asynch, 4316 conc_workers(), 4317 task_queues()); 4318 4319 // Since the actual number of workers we get may be different 4320 // from the number we requested above, do we need to do anything different 4321 // below? In particular, may be we need to subclass the SequantialSubTasksDone 4322 // class?? XXX 4323 cms_space ->initialize_sequential_subtasks_for_marking(num_workers); 4324 perm_space->initialize_sequential_subtasks_for_marking(num_workers); 4325 4326 // Refs discovery is already non-atomic. 4327 assert(!ref_processor()->discovery_is_atomic(), "Should be non-atomic"); 4328 assert(ref_processor()->discovery_is_mt(), "Discovery should be MT"); 4329 DEBUG_ONLY(RememberKlassesChecker cmx(should_unload_classes());) 4330 conc_workers()->start_task(&tsk); 4331 while (tsk.yielded()) { 4332 tsk.coordinator_yield(); 4333 conc_workers()->continue_task(&tsk); 4334 } 4335 // If the task was aborted, _restart_addr will be non-NULL 4336 assert(tsk.completed() || _restart_addr != NULL, "Inconsistency"); 4337 while (_restart_addr != NULL) { 4338 // XXX For now we do not make use of ABORTED state and have not 4339 // yet implemented the right abort semantics (even in the original 4340 // single-threaded CMS case). That needs some more investigation 4341 // and is deferred for now; see CR# TBF. 07252005YSR. XXX 4342 assert(!CMSAbortSemantics || tsk.aborted(), "Inconsistency"); 4343 // If _restart_addr is non-NULL, a marking stack overflow 4344 // occurred; we need to do a fresh marking iteration from the 4345 // indicated restart address. 4346 if (_foregroundGCIsActive && asynch) { 4347 // We may be running into repeated stack overflows, having 4348 // reached the limit of the stack size, while making very 4349 // slow forward progress. It may be best to bail out and 4350 // let the foreground collector do its job. 4351 // Clear _restart_addr, so that foreground GC 4352 // works from scratch. This avoids the headache of 4353 // a "rescan" which would otherwise be needed because 4354 // of the dirty mod union table & card table. 4355 _restart_addr = NULL; 4356 return false; 4357 } 4358 // Adjust the task to restart from _restart_addr 4359 tsk.reset(_restart_addr); 4360 cms_space ->initialize_sequential_subtasks_for_marking(num_workers, 4361 _restart_addr); 4362 perm_space->initialize_sequential_subtasks_for_marking(num_workers, 4363 _restart_addr); 4364 _restart_addr = NULL; 4365 // Get the workers going again 4366 conc_workers()->start_task(&tsk); 4367 while (tsk.yielded()) { 4368 tsk.coordinator_yield(); 4369 conc_workers()->continue_task(&tsk); 4370 } 4371 } 4372 assert(tsk.completed(), "Inconsistency"); 4373 assert(tsk.result() == true, "Inconsistency"); 4374 return true; 4375 } 4376 4377 bool CMSCollector::do_marking_st(bool asynch) { 4378 ResourceMark rm; 4379 HandleMark hm; 4380 4381 // Temporarily make refs discovery single threaded (non-MT) 4382 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false); 4383 MarkFromRootsClosure markFromRootsClosure(this, _span, &_markBitMap, 4384 &_markStack, &_revisitStack, CMSYield && asynch); 4385 // the last argument to iterate indicates whether the iteration 4386 // should be incremental with periodic yields. 4387 _markBitMap.iterate(&markFromRootsClosure); 4388 // If _restart_addr is non-NULL, a marking stack overflow 4389 // occurred; we need to do a fresh iteration from the 4390 // indicated restart address. 4391 while (_restart_addr != NULL) { 4392 if (_foregroundGCIsActive && asynch) { 4393 // We may be running into repeated stack overflows, having 4394 // reached the limit of the stack size, while making very 4395 // slow forward progress. It may be best to bail out and 4396 // let the foreground collector do its job. 4397 // Clear _restart_addr, so that foreground GC 4398 // works from scratch. This avoids the headache of 4399 // a "rescan" which would otherwise be needed because 4400 // of the dirty mod union table & card table. 4401 _restart_addr = NULL; 4402 return false; // indicating failure to complete marking 4403 } 4404 // Deal with stack overflow: 4405 // we restart marking from _restart_addr 4406 HeapWord* ra = _restart_addr; 4407 markFromRootsClosure.reset(ra); 4408 _restart_addr = NULL; 4409 _markBitMap.iterate(&markFromRootsClosure, ra, _span.end()); 4410 } 4411 return true; 4412 } 4413 4414 void CMSCollector::preclean() { 4415 check_correct_thread_executing(); 4416 assert(Thread::current()->is_ConcurrentGC_thread(), "Wrong thread"); 4417 verify_work_stacks_empty(); 4418 verify_overflow_empty(); 4419 _abort_preclean = false; 4420 if (CMSPrecleaningEnabled) { 4421 _eden_chunk_index = 0; 4422 size_t used = get_eden_used(); 4423 size_t capacity = get_eden_capacity(); 4424 // Don't start sampling unless we will get sufficiently 4425 // many samples. 4426 if (used < (capacity/(CMSScheduleRemarkSamplingRatio * 100) 4427 * CMSScheduleRemarkEdenPenetration)) { 4428 _start_sampling = true; 4429 } else { 4430 _start_sampling = false; 4431 } 4432 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 4433 CMSPhaseAccounting pa(this, "preclean", !PrintGCDetails); 4434 preclean_work(CMSPrecleanRefLists1, CMSPrecleanSurvivors1); 4435 } 4436 CMSTokenSync x(true); // is cms thread 4437 if (CMSPrecleaningEnabled) { 4438 sample_eden(); 4439 _collectorState = AbortablePreclean; 4440 } else { 4441 _collectorState = FinalMarking; 4442 } 4443 verify_work_stacks_empty(); 4444 verify_overflow_empty(); 4445 } 4446 4447 // Try and schedule the remark such that young gen 4448 // occupancy is CMSScheduleRemarkEdenPenetration %. 4449 void CMSCollector::abortable_preclean() { 4450 check_correct_thread_executing(); 4451 assert(CMSPrecleaningEnabled, "Inconsistent control state"); 4452 assert(_collectorState == AbortablePreclean, "Inconsistent control state"); 4453 4454 // If Eden's current occupancy is below this threshold, 4455 // immediately schedule the remark; else preclean 4456 // past the next scavenge in an effort to 4457 // schedule the pause as described avove. By choosing 4458 // CMSScheduleRemarkEdenSizeThreshold >= max eden size 4459 // we will never do an actual abortable preclean cycle. 4460 if (get_eden_used() > CMSScheduleRemarkEdenSizeThreshold) { 4461 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 4462 CMSPhaseAccounting pa(this, "abortable-preclean", !PrintGCDetails); 4463 // We need more smarts in the abortable preclean 4464 // loop below to deal with cases where allocation 4465 // in young gen is very very slow, and our precleaning 4466 // is running a losing race against a horde of 4467 // mutators intent on flooding us with CMS updates 4468 // (dirty cards). 4469 // One, admittedly dumb, strategy is to give up 4470 // after a certain number of abortable precleaning loops 4471 // or after a certain maximum time. We want to make 4472 // this smarter in the next iteration. 4473 // XXX FIX ME!!! YSR 4474 size_t loops = 0, workdone = 0, cumworkdone = 0, waited = 0; 4475 while (!(should_abort_preclean() || 4476 ConcurrentMarkSweepThread::should_terminate())) { 4477 workdone = preclean_work(CMSPrecleanRefLists2, CMSPrecleanSurvivors2); 4478 cumworkdone += workdone; 4479 loops++; 4480 // Voluntarily terminate abortable preclean phase if we have 4481 // been at it for too long. 4482 if ((CMSMaxAbortablePrecleanLoops != 0) && 4483 loops >= CMSMaxAbortablePrecleanLoops) { 4484 if (PrintGCDetails) { 4485 gclog_or_tty->print(" CMS: abort preclean due to loops "); 4486 } 4487 break; 4488 } 4489 if (pa.wallclock_millis() > CMSMaxAbortablePrecleanTime) { 4490 if (PrintGCDetails) { 4491 gclog_or_tty->print(" CMS: abort preclean due to time "); 4492 } 4493 break; 4494 } 4495 // If we are doing little work each iteration, we should 4496 // take a short break. 4497 if (workdone < CMSAbortablePrecleanMinWorkPerIteration) { 4498 // Sleep for some time, waiting for work to accumulate 4499 stopTimer(); 4500 cmsThread()->wait_on_cms_lock(CMSAbortablePrecleanWaitMillis); 4501 startTimer(); 4502 waited++; 4503 } 4504 } 4505 if (PrintCMSStatistics > 0) { 4506 gclog_or_tty->print(" [%d iterations, %d waits, %d cards)] ", 4507 loops, waited, cumworkdone); 4508 } 4509 } 4510 CMSTokenSync x(true); // is cms thread 4511 if (_collectorState != Idling) { 4512 assert(_collectorState == AbortablePreclean, 4513 "Spontaneous state transition?"); 4514 _collectorState = FinalMarking; 4515 } // Else, a foreground collection completed this CMS cycle. 4516 return; 4517 } 4518 4519 // Respond to an Eden sampling opportunity 4520 void CMSCollector::sample_eden() { 4521 // Make sure a young gc cannot sneak in between our 4522 // reading and recording of a sample. 4523 assert(Thread::current()->is_ConcurrentGC_thread(), 4524 "Only the cms thread may collect Eden samples"); 4525 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 4526 "Should collect samples while holding CMS token"); 4527 if (!_start_sampling) { 4528 return; 4529 } 4530 if (_eden_chunk_array) { 4531 if (_eden_chunk_index < _eden_chunk_capacity) { 4532 _eden_chunk_array[_eden_chunk_index] = *_top_addr; // take sample 4533 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr, 4534 "Unexpected state of Eden"); 4535 // We'd like to check that what we just sampled is an oop-start address; 4536 // however, we cannot do that here since the object may not yet have been 4537 // initialized. So we'll instead do the check when we _use_ this sample 4538 // later. 4539 if (_eden_chunk_index == 0 || 4540 (pointer_delta(_eden_chunk_array[_eden_chunk_index], 4541 _eden_chunk_array[_eden_chunk_index-1]) 4542 >= CMSSamplingGrain)) { 4543 _eden_chunk_index++; // commit sample 4544 } 4545 } 4546 } 4547 if ((_collectorState == AbortablePreclean) && !_abort_preclean) { 4548 size_t used = get_eden_used(); 4549 size_t capacity = get_eden_capacity(); 4550 assert(used <= capacity, "Unexpected state of Eden"); 4551 if (used > (capacity/100 * CMSScheduleRemarkEdenPenetration)) { 4552 _abort_preclean = true; 4553 } 4554 } 4555 } 4556 4557 4558 size_t CMSCollector::preclean_work(bool clean_refs, bool clean_survivor) { 4559 assert(_collectorState == Precleaning || 4560 _collectorState == AbortablePreclean, "incorrect state"); 4561 ResourceMark rm; 4562 HandleMark hm; 4563 4564 // Precleaning is currently not MT but the reference processor 4565 // may be set for MT. Disable it temporarily here. 4566 ReferenceProcessor* rp = ref_processor(); 4567 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(rp, false); 4568 4569 // Do one pass of scrubbing the discovered reference lists 4570 // to remove any reference objects with strongly-reachable 4571 // referents. 4572 if (clean_refs) { 4573 CMSPrecleanRefsYieldClosure yield_cl(this); 4574 assert(rp->span().equals(_span), "Spans should be equal"); 4575 CMSKeepAliveClosure keep_alive(this, _span, &_markBitMap, 4576 &_markStack, &_revisitStack, 4577 true /* preclean */); 4578 CMSDrainMarkingStackClosure complete_trace(this, 4579 _span, &_markBitMap, &_markStack, 4580 &keep_alive, true /* preclean */); 4581 4582 // We don't want this step to interfere with a young 4583 // collection because we don't want to take CPU 4584 // or memory bandwidth away from the young GC threads 4585 // (which may be as many as there are CPUs). 4586 // Note that we don't need to protect ourselves from 4587 // interference with mutators because they can't 4588 // manipulate the discovered reference lists nor affect 4589 // the computed reachability of the referents, the 4590 // only properties manipulated by the precleaning 4591 // of these reference lists. 4592 stopTimer(); 4593 CMSTokenSyncWithLocks x(true /* is cms thread */, 4594 bitMapLock()); 4595 startTimer(); 4596 sample_eden(); 4597 4598 // The following will yield to allow foreground 4599 // collection to proceed promptly. XXX YSR: 4600 // The code in this method may need further 4601 // tweaking for better performance and some restructuring 4602 // for cleaner interfaces. 4603 GCTimer *gc_timer = NULL; // Currently not tracing concurrent phases 4604 rp->preclean_discovered_references( 4605 rp->is_alive_non_header(), &keep_alive, &complete_trace, 4606 &yield_cl, should_unload_classes(), gc_timer); 4607 } 4608 4609 if (clean_survivor) { // preclean the active survivor space(s) 4610 assert(_young_gen->kind() == Generation::DefNew || 4611 _young_gen->kind() == Generation::ParNew || 4612 _young_gen->kind() == Generation::ASParNew, 4613 "incorrect type for cast"); 4614 DefNewGeneration* dng = (DefNewGeneration*)_young_gen; 4615 PushAndMarkClosure pam_cl(this, _span, ref_processor(), 4616 &_markBitMap, &_modUnionTable, 4617 &_markStack, &_revisitStack, 4618 true /* precleaning phase */); 4619 stopTimer(); 4620 CMSTokenSyncWithLocks ts(true /* is cms thread */, 4621 bitMapLock()); 4622 startTimer(); 4623 unsigned int before_count = 4624 GenCollectedHeap::heap()->total_collections(); 4625 SurvivorSpacePrecleanClosure 4626 sss_cl(this, _span, &_markBitMap, &_markStack, 4627 &pam_cl, before_count, CMSYield); 4628 DEBUG_ONLY(RememberKlassesChecker mx(should_unload_classes());) 4629 dng->from()->object_iterate_careful(&sss_cl); 4630 dng->to()->object_iterate_careful(&sss_cl); 4631 } 4632 MarkRefsIntoAndScanClosure 4633 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable, 4634 &_markStack, &_revisitStack, this, CMSYield, 4635 true /* precleaning phase */); 4636 // CAUTION: The following closure has persistent state that may need to 4637 // be reset upon a decrease in the sequence of addresses it 4638 // processes. 4639 ScanMarkedObjectsAgainCarefullyClosure 4640 smoac_cl(this, _span, 4641 &_markBitMap, &_markStack, &_revisitStack, &mrias_cl, CMSYield); 4642 4643 // Preclean dirty cards in ModUnionTable and CardTable using 4644 // appropriate convergence criterion; 4645 // repeat CMSPrecleanIter times unless we find that 4646 // we are losing. 4647 assert(CMSPrecleanIter < 10, "CMSPrecleanIter is too large"); 4648 assert(CMSPrecleanNumerator < CMSPrecleanDenominator, 4649 "Bad convergence multiplier"); 4650 assert(CMSPrecleanThreshold >= 100, 4651 "Unreasonably low CMSPrecleanThreshold"); 4652 4653 size_t numIter, cumNumCards, lastNumCards, curNumCards; 4654 for (numIter = 0, cumNumCards = lastNumCards = curNumCards = 0; 4655 numIter < CMSPrecleanIter; 4656 numIter++, lastNumCards = curNumCards, cumNumCards += curNumCards) { 4657 curNumCards = preclean_mod_union_table(_cmsGen, &smoac_cl); 4658 if (CMSPermGenPrecleaningEnabled) { 4659 curNumCards += preclean_mod_union_table(_permGen, &smoac_cl); 4660 } 4661 if (Verbose && PrintGCDetails) { 4662 gclog_or_tty->print(" (modUnionTable: %d cards)", curNumCards); 4663 } 4664 // Either there are very few dirty cards, so re-mark 4665 // pause will be small anyway, or our pre-cleaning isn't 4666 // that much faster than the rate at which cards are being 4667 // dirtied, so we might as well stop and re-mark since 4668 // precleaning won't improve our re-mark time by much. 4669 if (curNumCards <= CMSPrecleanThreshold || 4670 (numIter > 0 && 4671 (curNumCards * CMSPrecleanDenominator > 4672 lastNumCards * CMSPrecleanNumerator))) { 4673 numIter++; 4674 cumNumCards += curNumCards; 4675 break; 4676 } 4677 } 4678 curNumCards = preclean_card_table(_cmsGen, &smoac_cl); 4679 if (CMSPermGenPrecleaningEnabled) { 4680 curNumCards += preclean_card_table(_permGen, &smoac_cl); 4681 } 4682 cumNumCards += curNumCards; 4683 if (PrintGCDetails && PrintCMSStatistics != 0) { 4684 gclog_or_tty->print_cr(" (cardTable: %d cards, re-scanned %d cards, %d iterations)", 4685 curNumCards, cumNumCards, numIter); 4686 } 4687 return cumNumCards; // as a measure of useful work done 4688 } 4689 4690 // PRECLEANING NOTES: 4691 // Precleaning involves: 4692 // . reading the bits of the modUnionTable and clearing the set bits. 4693 // . For the cards corresponding to the set bits, we scan the 4694 // objects on those cards. This means we need the free_list_lock 4695 // so that we can safely iterate over the CMS space when scanning 4696 // for oops. 4697 // . When we scan the objects, we'll be both reading and setting 4698 // marks in the marking bit map, so we'll need the marking bit map. 4699 // . For protecting _collector_state transitions, we take the CGC_lock. 4700 // Note that any races in the reading of of card table entries by the 4701 // CMS thread on the one hand and the clearing of those entries by the 4702 // VM thread or the setting of those entries by the mutator threads on the 4703 // other are quite benign. However, for efficiency it makes sense to keep 4704 // the VM thread from racing with the CMS thread while the latter is 4705 // dirty card info to the modUnionTable. We therefore also use the 4706 // CGC_lock to protect the reading of the card table and the mod union 4707 // table by the CM thread. 4708 // . We run concurrently with mutator updates, so scanning 4709 // needs to be done carefully -- we should not try to scan 4710 // potentially uninitialized objects. 4711 // 4712 // Locking strategy: While holding the CGC_lock, we scan over and 4713 // reset a maximal dirty range of the mod union / card tables, then lock 4714 // the free_list_lock and bitmap lock to do a full marking, then 4715 // release these locks; and repeat the cycle. This allows for a 4716 // certain amount of fairness in the sharing of these locks between 4717 // the CMS collector on the one hand, and the VM thread and the 4718 // mutators on the other. 4719 4720 // NOTE: preclean_mod_union_table() and preclean_card_table() 4721 // further below are largely identical; if you need to modify 4722 // one of these methods, please check the other method too. 4723 4724 size_t CMSCollector::preclean_mod_union_table( 4725 ConcurrentMarkSweepGeneration* gen, 4726 ScanMarkedObjectsAgainCarefullyClosure* cl) { 4727 verify_work_stacks_empty(); 4728 verify_overflow_empty(); 4729 4730 // Turn off checking for this method but turn it back on 4731 // selectively. There are yield points in this method 4732 // but it is difficult to turn the checking off just around 4733 // the yield points. It is simpler to selectively turn 4734 // it on. 4735 DEBUG_ONLY(RememberKlassesChecker mux(false);) 4736 4737 // strategy: starting with the first card, accumulate contiguous 4738 // ranges of dirty cards; clear these cards, then scan the region 4739 // covered by these cards. 4740 4741 // Since all of the MUT is committed ahead, we can just use 4742 // that, in case the generations expand while we are precleaning. 4743 // It might also be fine to just use the committed part of the 4744 // generation, but we might potentially miss cards when the 4745 // generation is rapidly expanding while we are in the midst 4746 // of precleaning. 4747 HeapWord* startAddr = gen->reserved().start(); 4748 HeapWord* endAddr = gen->reserved().end(); 4749 4750 cl->setFreelistLock(gen->freelistLock()); // needed for yielding 4751 4752 size_t numDirtyCards, cumNumDirtyCards; 4753 HeapWord *nextAddr, *lastAddr; 4754 for (cumNumDirtyCards = numDirtyCards = 0, 4755 nextAddr = lastAddr = startAddr; 4756 nextAddr < endAddr; 4757 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 4758 4759 ResourceMark rm; 4760 HandleMark hm; 4761 4762 MemRegion dirtyRegion; 4763 { 4764 stopTimer(); 4765 // Potential yield point 4766 CMSTokenSync ts(true); 4767 startTimer(); 4768 sample_eden(); 4769 // Get dirty region starting at nextOffset (inclusive), 4770 // simultaneously clearing it. 4771 dirtyRegion = 4772 _modUnionTable.getAndClearMarkedRegion(nextAddr, endAddr); 4773 assert(dirtyRegion.start() >= nextAddr, 4774 "returned region inconsistent?"); 4775 } 4776 // Remember where the next search should begin. 4777 // The returned region (if non-empty) is a right open interval, 4778 // so lastOffset is obtained from the right end of that 4779 // interval. 4780 lastAddr = dirtyRegion.end(); 4781 // Should do something more transparent and less hacky XXX 4782 numDirtyCards = 4783 _modUnionTable.heapWordDiffToOffsetDiff(dirtyRegion.word_size()); 4784 4785 // We'll scan the cards in the dirty region (with periodic 4786 // yields for foreground GC as needed). 4787 if (!dirtyRegion.is_empty()) { 4788 assert(numDirtyCards > 0, "consistency check"); 4789 HeapWord* stop_point = NULL; 4790 stopTimer(); 4791 // Potential yield point 4792 CMSTokenSyncWithLocks ts(true, gen->freelistLock(), 4793 bitMapLock()); 4794 startTimer(); 4795 { 4796 verify_work_stacks_empty(); 4797 verify_overflow_empty(); 4798 sample_eden(); 4799 DEBUG_ONLY(RememberKlassesChecker mx(should_unload_classes());) 4800 stop_point = 4801 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 4802 } 4803 if (stop_point != NULL) { 4804 // The careful iteration stopped early either because it found an 4805 // uninitialized object, or because we were in the midst of an 4806 // "abortable preclean", which should now be aborted. Redirty 4807 // the bits corresponding to the partially-scanned or unscanned 4808 // cards. We'll either restart at the next block boundary or 4809 // abort the preclean. 4810 assert((CMSPermGenPrecleaningEnabled && (gen == _permGen)) || 4811 (_collectorState == AbortablePreclean && should_abort_preclean()), 4812 "Unparsable objects should only be in perm gen."); 4813 _modUnionTable.mark_range(MemRegion(stop_point, dirtyRegion.end())); 4814 if (should_abort_preclean()) { 4815 break; // out of preclean loop 4816 } else { 4817 // Compute the next address at which preclean should pick up; 4818 // might need bitMapLock in order to read P-bits. 4819 lastAddr = next_card_start_after_block(stop_point); 4820 } 4821 } 4822 } else { 4823 assert(lastAddr == endAddr, "consistency check"); 4824 assert(numDirtyCards == 0, "consistency check"); 4825 break; 4826 } 4827 } 4828 verify_work_stacks_empty(); 4829 verify_overflow_empty(); 4830 return cumNumDirtyCards; 4831 } 4832 4833 // NOTE: preclean_mod_union_table() above and preclean_card_table() 4834 // below are largely identical; if you need to modify 4835 // one of these methods, please check the other method too. 4836 4837 size_t CMSCollector::preclean_card_table(ConcurrentMarkSweepGeneration* gen, 4838 ScanMarkedObjectsAgainCarefullyClosure* cl) { 4839 // strategy: it's similar to precleamModUnionTable above, in that 4840 // we accumulate contiguous ranges of dirty cards, mark these cards 4841 // precleaned, then scan the region covered by these cards. 4842 HeapWord* endAddr = (HeapWord*)(gen->_virtual_space.high()); 4843 HeapWord* startAddr = (HeapWord*)(gen->_virtual_space.low()); 4844 4845 cl->setFreelistLock(gen->freelistLock()); // needed for yielding 4846 4847 size_t numDirtyCards, cumNumDirtyCards; 4848 HeapWord *lastAddr, *nextAddr; 4849 4850 for (cumNumDirtyCards = numDirtyCards = 0, 4851 nextAddr = lastAddr = startAddr; 4852 nextAddr < endAddr; 4853 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 4854 4855 ResourceMark rm; 4856 HandleMark hm; 4857 4858 MemRegion dirtyRegion; 4859 { 4860 // See comments in "Precleaning notes" above on why we 4861 // do this locking. XXX Could the locking overheads be 4862 // too high when dirty cards are sparse? [I don't think so.] 4863 stopTimer(); 4864 CMSTokenSync x(true); // is cms thread 4865 startTimer(); 4866 sample_eden(); 4867 // Get and clear dirty region from card table 4868 dirtyRegion = _ct->ct_bs()->dirty_card_range_after_reset( 4869 MemRegion(nextAddr, endAddr), 4870 true, 4871 CardTableModRefBS::precleaned_card_val()); 4872 4873 assert(dirtyRegion.start() >= nextAddr, 4874 "returned region inconsistent?"); 4875 } 4876 lastAddr = dirtyRegion.end(); 4877 numDirtyCards = 4878 dirtyRegion.word_size()/CardTableModRefBS::card_size_in_words; 4879 4880 if (!dirtyRegion.is_empty()) { 4881 stopTimer(); 4882 CMSTokenSyncWithLocks ts(true, gen->freelistLock(), bitMapLock()); 4883 startTimer(); 4884 sample_eden(); 4885 verify_work_stacks_empty(); 4886 verify_overflow_empty(); 4887 DEBUG_ONLY(RememberKlassesChecker mx(should_unload_classes());) 4888 HeapWord* stop_point = 4889 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 4890 if (stop_point != NULL) { 4891 // The careful iteration stopped early because it found an 4892 // uninitialized object. Redirty the bits corresponding to the 4893 // partially-scanned or unscanned cards, and start again at the 4894 // next block boundary. 4895 assert(CMSPermGenPrecleaningEnabled || 4896 (_collectorState == AbortablePreclean && should_abort_preclean()), 4897 "Unparsable objects should only be in perm gen."); 4898 _ct->ct_bs()->invalidate(MemRegion(stop_point, dirtyRegion.end())); 4899 if (should_abort_preclean()) { 4900 break; // out of preclean loop 4901 } else { 4902 // Compute the next address at which preclean should pick up. 4903 lastAddr = next_card_start_after_block(stop_point); 4904 } 4905 } 4906 } else { 4907 break; 4908 } 4909 } 4910 verify_work_stacks_empty(); 4911 verify_overflow_empty(); 4912 return cumNumDirtyCards; 4913 } 4914 4915 void CMSCollector::checkpointRootsFinal(bool asynch, 4916 bool clear_all_soft_refs, bool init_mark_was_synchronous) { 4917 assert(_collectorState == FinalMarking, "incorrect state transition?"); 4918 check_correct_thread_executing(); 4919 // world is stopped at this checkpoint 4920 assert(SafepointSynchronize::is_at_safepoint(), 4921 "world should be stopped"); 4922 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 4923 4924 verify_work_stacks_empty(); 4925 verify_overflow_empty(); 4926 4927 SpecializationStats::clear(); 4928 if (PrintGCDetails) { 4929 gclog_or_tty->print("[YG occupancy: "SIZE_FORMAT" K ("SIZE_FORMAT" K)]", 4930 _young_gen->used() / K, 4931 _young_gen->capacity() / K); 4932 } 4933 if (asynch) { 4934 if (CMSScavengeBeforeRemark) { 4935 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4936 // Temporarily set flag to false, GCH->do_collection will 4937 // expect it to be false and set to true 4938 FlagSetting fl(gch->_is_gc_active, false); 4939 NOT_PRODUCT(GCTraceTime t("Scavenge-Before-Remark", 4940 PrintGCDetails && Verbose, true, _gc_timer_cm);) 4941 int level = _cmsGen->level() - 1; 4942 if (level >= 0) { 4943 gch->do_collection(true, // full (i.e. force, see below) 4944 false, // !clear_all_soft_refs 4945 0, // size 4946 false, // is_tlab 4947 level // max_level 4948 ); 4949 } 4950 } 4951 FreelistLocker x(this); 4952 MutexLockerEx y(bitMapLock(), 4953 Mutex::_no_safepoint_check_flag); 4954 assert(!init_mark_was_synchronous, "but that's impossible!"); 4955 checkpointRootsFinalWork(asynch, clear_all_soft_refs, false); 4956 } else { 4957 // already have all the locks 4958 checkpointRootsFinalWork(asynch, clear_all_soft_refs, 4959 init_mark_was_synchronous); 4960 } 4961 verify_work_stacks_empty(); 4962 verify_overflow_empty(); 4963 SpecializationStats::print(); 4964 } 4965 4966 void CMSCollector::checkpointRootsFinalWork(bool asynch, 4967 bool clear_all_soft_refs, bool init_mark_was_synchronous) { 4968 4969 NOT_PRODUCT(GCTraceTime tr("checkpointRootsFinalWork", PrintGCDetails, false, _gc_timer_cm);) 4970 4971 assert(haveFreelistLocks(), "must have free list locks"); 4972 assert_lock_strong(bitMapLock()); 4973 4974 if (UseAdaptiveSizePolicy) { 4975 size_policy()->checkpoint_roots_final_begin(); 4976 } 4977 4978 ResourceMark rm; 4979 HandleMark hm; 4980 4981 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4982 4983 if (should_unload_classes()) { 4984 CodeCache::gc_prologue(); 4985 } 4986 assert(haveFreelistLocks(), "must have free list locks"); 4987 assert_lock_strong(bitMapLock()); 4988 4989 DEBUG_ONLY(RememberKlassesChecker fmx(should_unload_classes());) 4990 if (!init_mark_was_synchronous) { 4991 // We might assume that we need not fill TLAB's when 4992 // CMSScavengeBeforeRemark is set, because we may have just done 4993 // a scavenge which would have filled all TLAB's -- and besides 4994 // Eden would be empty. This however may not always be the case -- 4995 // for instance although we asked for a scavenge, it may not have 4996 // happened because of a JNI critical section. We probably need 4997 // a policy for deciding whether we can in that case wait until 4998 // the critical section releases and then do the remark following 4999 // the scavenge, and skip it here. In the absence of that policy, 5000 // or of an indication of whether the scavenge did indeed occur, 5001 // we cannot rely on TLAB's having been filled and must do 5002 // so here just in case a scavenge did not happen. 5003 gch->ensure_parsability(false); // fill TLAB's, but no need to retire them 5004 // Update the saved marks which may affect the root scans. 5005 gch->save_marks(); 5006 5007 { 5008 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;) 5009 5010 // Note on the role of the mod union table: 5011 // Since the marker in "markFromRoots" marks concurrently with 5012 // mutators, it is possible for some reachable objects not to have been 5013 // scanned. For instance, an only reference to an object A was 5014 // placed in object B after the marker scanned B. Unless B is rescanned, 5015 // A would be collected. Such updates to references in marked objects 5016 // are detected via the mod union table which is the set of all cards 5017 // dirtied since the first checkpoint in this GC cycle and prior to 5018 // the most recent young generation GC, minus those cleaned up by the 5019 // concurrent precleaning. 5020 if (CMSParallelRemarkEnabled && CollectedHeap::use_parallel_gc_threads()) { 5021 GCTraceTime t("Rescan (parallel) ", PrintGCDetails, false, _gc_timer_cm); 5022 do_remark_parallel(); 5023 } else { 5024 GCTraceTime t("Rescan (non-parallel) ", PrintGCDetails, false, 5025 _gc_timer_cm); 5026 do_remark_non_parallel(); 5027 } 5028 } 5029 } else { 5030 assert(!asynch, "Can't have init_mark_was_synchronous in asynch mode"); 5031 // The initial mark was stop-world, so there's no rescanning to 5032 // do; go straight on to the next step below. 5033 } 5034 verify_work_stacks_empty(); 5035 verify_overflow_empty(); 5036 5037 { 5038 NOT_PRODUCT(GCTraceTime ts("refProcessingWork", PrintGCDetails, false, _gc_timer_cm);) 5039 refProcessingWork(asynch, clear_all_soft_refs); 5040 } 5041 verify_work_stacks_empty(); 5042 verify_overflow_empty(); 5043 5044 if (should_unload_classes()) { 5045 CodeCache::gc_epilogue(); 5046 } 5047 JvmtiExport::gc_epilogue(); 5048 5049 // If we encountered any (marking stack / work queue) overflow 5050 // events during the current CMS cycle, take appropriate 5051 // remedial measures, where possible, so as to try and avoid 5052 // recurrence of that condition. 5053 assert(_markStack.isEmpty(), "No grey objects"); 5054 size_t ser_ovflw = _ser_pmc_remark_ovflw + _ser_pmc_preclean_ovflw + 5055 _ser_kac_ovflw + _ser_kac_preclean_ovflw; 5056 if (ser_ovflw > 0) { 5057 if (PrintCMSStatistics != 0) { 5058 gclog_or_tty->print_cr("Marking stack overflow (benign) " 5059 "(pmc_pc="SIZE_FORMAT", pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT 5060 ", kac_preclean="SIZE_FORMAT")", 5061 _ser_pmc_preclean_ovflw, _ser_pmc_remark_ovflw, 5062 _ser_kac_ovflw, _ser_kac_preclean_ovflw); 5063 } 5064 _markStack.expand(); 5065 _ser_pmc_remark_ovflw = 0; 5066 _ser_pmc_preclean_ovflw = 0; 5067 _ser_kac_preclean_ovflw = 0; 5068 _ser_kac_ovflw = 0; 5069 } 5070 if (_par_pmc_remark_ovflw > 0 || _par_kac_ovflw > 0) { 5071 if (PrintCMSStatistics != 0) { 5072 gclog_or_tty->print_cr("Work queue overflow (benign) " 5073 "(pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT")", 5074 _par_pmc_remark_ovflw, _par_kac_ovflw); 5075 } 5076 _par_pmc_remark_ovflw = 0; 5077 _par_kac_ovflw = 0; 5078 } 5079 if (PrintCMSStatistics != 0) { 5080 if (_markStack._hit_limit > 0) { 5081 gclog_or_tty->print_cr(" (benign) Hit max stack size limit ("SIZE_FORMAT")", 5082 _markStack._hit_limit); 5083 } 5084 if (_markStack._failed_double > 0) { 5085 gclog_or_tty->print_cr(" (benign) Failed stack doubling ("SIZE_FORMAT")," 5086 " current capacity "SIZE_FORMAT, 5087 _markStack._failed_double, 5088 _markStack.capacity()); 5089 } 5090 } 5091 _markStack._hit_limit = 0; 5092 _markStack._failed_double = 0; 5093 5094 // Check that all the klasses have been checked 5095 assert(_revisitStack.isEmpty(), "Not all klasses revisited"); 5096 5097 if ((VerifyAfterGC || VerifyDuringGC) && 5098 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 5099 verify_after_remark(); 5100 } 5101 5102 _gc_tracer_cm->report_object_count_after_gc(&_is_alive_closure); 5103 5104 // Change under the freelistLocks. 5105 _collectorState = Sweeping; 5106 // Call isAllClear() under bitMapLock 5107 assert(_modUnionTable.isAllClear(), "Should be clear by end of the" 5108 " final marking"); 5109 if (UseAdaptiveSizePolicy) { 5110 size_policy()->checkpoint_roots_final_end(gch->gc_cause()); 5111 } 5112 } 5113 5114 // Parallel remark task 5115 class CMSParRemarkTask: public AbstractGangTask { 5116 CMSCollector* _collector; 5117 int _n_workers; 5118 CompactibleFreeListSpace* _cms_space; 5119 CompactibleFreeListSpace* _perm_space; 5120 5121 // The per-thread work queues, available here for stealing. 5122 OopTaskQueueSet* _task_queues; 5123 ParallelTaskTerminator _term; 5124 5125 public: 5126 // A value of 0 passed to n_workers will cause the number of 5127 // workers to be taken from the active workers in the work gang. 5128 CMSParRemarkTask(CMSCollector* collector, 5129 CompactibleFreeListSpace* cms_space, 5130 CompactibleFreeListSpace* perm_space, 5131 int n_workers, FlexibleWorkGang* workers, 5132 OopTaskQueueSet* task_queues): 5133 AbstractGangTask("Rescan roots and grey objects in parallel"), 5134 _collector(collector), 5135 _cms_space(cms_space), _perm_space(perm_space), 5136 _n_workers(n_workers), 5137 _task_queues(task_queues), 5138 _term(n_workers, task_queues) { } 5139 5140 OopTaskQueueSet* task_queues() { return _task_queues; } 5141 5142 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 5143 5144 ParallelTaskTerminator* terminator() { return &_term; } 5145 int n_workers() { return _n_workers; } 5146 5147 void work(uint worker_id); 5148 5149 private: 5150 // Work method in support of parallel rescan ... of young gen spaces 5151 void do_young_space_rescan(int i, Par_MarkRefsIntoAndScanClosure* cl, 5152 ContiguousSpace* space, 5153 HeapWord** chunk_array, size_t chunk_top); 5154 5155 // ... of dirty cards in old space 5156 void do_dirty_card_rescan_tasks(CompactibleFreeListSpace* sp, int i, 5157 Par_MarkRefsIntoAndScanClosure* cl); 5158 5159 // ... work stealing for the above 5160 void do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, int* seed); 5161 }; 5162 5163 // work_queue(i) is passed to the closure 5164 // Par_MarkRefsIntoAndScanClosure. The "i" parameter 5165 // also is passed to do_dirty_card_rescan_tasks() and to 5166 // do_work_steal() to select the i-th task_queue. 5167 5168 void CMSParRemarkTask::work(uint worker_id) { 5169 elapsedTimer _timer; 5170 ResourceMark rm; 5171 HandleMark hm; 5172 5173 // ---------- rescan from roots -------------- 5174 _timer.start(); 5175 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5176 Par_MarkRefsIntoAndScanClosure par_mrias_cl(_collector, 5177 _collector->_span, _collector->ref_processor(), 5178 &(_collector->_markBitMap), 5179 work_queue(worker_id), &(_collector->_revisitStack)); 5180 5181 // Rescan young gen roots first since these are likely 5182 // coarsely partitioned and may, on that account, constitute 5183 // the critical path; thus, it's best to start off that 5184 // work first. 5185 // ---------- young gen roots -------------- 5186 { 5187 DefNewGeneration* dng = _collector->_young_gen->as_DefNewGeneration(); 5188 EdenSpace* eden_space = dng->eden(); 5189 ContiguousSpace* from_space = dng->from(); 5190 ContiguousSpace* to_space = dng->to(); 5191 5192 HeapWord** eca = _collector->_eden_chunk_array; 5193 size_t ect = _collector->_eden_chunk_index; 5194 HeapWord** sca = _collector->_survivor_chunk_array; 5195 size_t sct = _collector->_survivor_chunk_index; 5196 5197 assert(ect <= _collector->_eden_chunk_capacity, "out of bounds"); 5198 assert(sct <= _collector->_survivor_chunk_capacity, "out of bounds"); 5199 5200 do_young_space_rescan(worker_id, &par_mrias_cl, to_space, NULL, 0); 5201 do_young_space_rescan(worker_id, &par_mrias_cl, from_space, sca, sct); 5202 do_young_space_rescan(worker_id, &par_mrias_cl, eden_space, eca, ect); 5203 5204 _timer.stop(); 5205 if (PrintCMSStatistics != 0) { 5206 gclog_or_tty->print_cr( 5207 "Finished young gen rescan work in %dth thread: %3.3f sec", 5208 worker_id, _timer.seconds()); 5209 } 5210 } 5211 5212 // ---------- remaining roots -------------- 5213 _timer.reset(); 5214 _timer.start(); 5215 gch->gen_process_strong_roots(_collector->_cmsGen->level(), 5216 false, // yg was scanned above 5217 false, // this is parallel code 5218 true, // collecting perm gen 5219 SharedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()), 5220 &par_mrias_cl, 5221 true, // walk all of code cache if (so & SO_CodeCache) 5222 NULL); 5223 assert(_collector->should_unload_classes() 5224 || (_collector->CMSCollector::roots_scanning_options() & SharedHeap::SO_CodeCache), 5225 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 5226 _timer.stop(); 5227 if (PrintCMSStatistics != 0) { 5228 gclog_or_tty->print_cr( 5229 "Finished remaining root rescan work in %dth thread: %3.3f sec", 5230 worker_id, _timer.seconds()); 5231 } 5232 5233 // ---------- rescan dirty cards ------------ 5234 _timer.reset(); 5235 _timer.start(); 5236 5237 // Do the rescan tasks for each of the two spaces 5238 // (cms_space and perm_space) in turn. 5239 // "worker_id" is passed to select the task_queue for "worker_id" 5240 do_dirty_card_rescan_tasks(_cms_space, worker_id, &par_mrias_cl); 5241 do_dirty_card_rescan_tasks(_perm_space, worker_id, &par_mrias_cl); 5242 _timer.stop(); 5243 if (PrintCMSStatistics != 0) { 5244 gclog_or_tty->print_cr( 5245 "Finished dirty card rescan work in %dth thread: %3.3f sec", 5246 worker_id, _timer.seconds()); 5247 } 5248 5249 // ---------- steal work from other threads ... 5250 // ---------- ... and drain overflow list. 5251 _timer.reset(); 5252 _timer.start(); 5253 do_work_steal(worker_id, &par_mrias_cl, _collector->hash_seed(worker_id)); 5254 _timer.stop(); 5255 if (PrintCMSStatistics != 0) { 5256 gclog_or_tty->print_cr( 5257 "Finished work stealing in %dth thread: %3.3f sec", 5258 worker_id, _timer.seconds()); 5259 } 5260 } 5261 5262 // Note that parameter "i" is not used. 5263 void 5264 CMSParRemarkTask::do_young_space_rescan(int i, 5265 Par_MarkRefsIntoAndScanClosure* cl, ContiguousSpace* space, 5266 HeapWord** chunk_array, size_t chunk_top) { 5267 // Until all tasks completed: 5268 // . claim an unclaimed task 5269 // . compute region boundaries corresponding to task claimed 5270 // using chunk_array 5271 // . par_oop_iterate(cl) over that region 5272 5273 ResourceMark rm; 5274 HandleMark hm; 5275 5276 SequentialSubTasksDone* pst = space->par_seq_tasks(); 5277 assert(pst->valid(), "Uninitialized use?"); 5278 5279 uint nth_task = 0; 5280 uint n_tasks = pst->n_tasks(); 5281 5282 HeapWord *start, *end; 5283 while (!pst->is_task_claimed(/* reference */ nth_task)) { 5284 // We claimed task # nth_task; compute its boundaries. 5285 if (chunk_top == 0) { // no samples were taken 5286 assert(nth_task == 0 && n_tasks == 1, "Can have only 1 EdenSpace task"); 5287 start = space->bottom(); 5288 end = space->top(); 5289 } else if (nth_task == 0) { 5290 start = space->bottom(); 5291 end = chunk_array[nth_task]; 5292 } else if (nth_task < (uint)chunk_top) { 5293 assert(nth_task >= 1, "Control point invariant"); 5294 start = chunk_array[nth_task - 1]; 5295 end = chunk_array[nth_task]; 5296 } else { 5297 assert(nth_task == (uint)chunk_top, "Control point invariant"); 5298 start = chunk_array[chunk_top - 1]; 5299 end = space->top(); 5300 } 5301 MemRegion mr(start, end); 5302 // Verify that mr is in space 5303 assert(mr.is_empty() || space->used_region().contains(mr), 5304 "Should be in space"); 5305 // Verify that "start" is an object boundary 5306 assert(mr.is_empty() || oop(mr.start())->is_oop(), 5307 "Should be an oop"); 5308 space->par_oop_iterate(mr, cl); 5309 } 5310 pst->all_tasks_completed(); 5311 } 5312 5313 void 5314 CMSParRemarkTask::do_dirty_card_rescan_tasks( 5315 CompactibleFreeListSpace* sp, int i, 5316 Par_MarkRefsIntoAndScanClosure* cl) { 5317 // Until all tasks completed: 5318 // . claim an unclaimed task 5319 // . compute region boundaries corresponding to task claimed 5320 // . transfer dirty bits ct->mut for that region 5321 // . apply rescanclosure to dirty mut bits for that region 5322 5323 ResourceMark rm; 5324 HandleMark hm; 5325 5326 OopTaskQueue* work_q = work_queue(i); 5327 ModUnionClosure modUnionClosure(&(_collector->_modUnionTable)); 5328 // CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! 5329 // CAUTION: This closure has state that persists across calls to 5330 // the work method dirty_range_iterate_clear() in that it has 5331 // imbedded in it a (subtype of) UpwardsObjectClosure. The 5332 // use of that state in the imbedded UpwardsObjectClosure instance 5333 // assumes that the cards are always iterated (even if in parallel 5334 // by several threads) in monotonically increasing order per each 5335 // thread. This is true of the implementation below which picks 5336 // card ranges (chunks) in monotonically increasing order globally 5337 // and, a-fortiori, in monotonically increasing order per thread 5338 // (the latter order being a subsequence of the former). 5339 // If the work code below is ever reorganized into a more chaotic 5340 // work-partitioning form than the current "sequential tasks" 5341 // paradigm, the use of that persistent state will have to be 5342 // revisited and modified appropriately. See also related 5343 // bug 4756801 work on which should examine this code to make 5344 // sure that the changes there do not run counter to the 5345 // assumptions made here and necessary for correctness and 5346 // efficiency. Note also that this code might yield inefficient 5347 // behaviour in the case of very large objects that span one or 5348 // more work chunks. Such objects would potentially be scanned 5349 // several times redundantly. Work on 4756801 should try and 5350 // address that performance anomaly if at all possible. XXX 5351 MemRegion full_span = _collector->_span; 5352 CMSBitMap* bm = &(_collector->_markBitMap); // shared 5353 CMSMarkStack* rs = &(_collector->_revisitStack); // shared 5354 MarkFromDirtyCardsClosure 5355 greyRescanClosure(_collector, full_span, // entire span of interest 5356 sp, bm, work_q, rs, cl); 5357 5358 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 5359 assert(pst->valid(), "Uninitialized use?"); 5360 uint nth_task = 0; 5361 const int alignment = CardTableModRefBS::card_size * BitsPerWord; 5362 MemRegion span = sp->used_region(); 5363 HeapWord* start_addr = span.start(); 5364 HeapWord* end_addr = (HeapWord*)round_to((intptr_t)span.end(), 5365 alignment); 5366 const size_t chunk_size = sp->rescan_task_size(); // in HeapWord units 5367 assert((HeapWord*)round_to((intptr_t)start_addr, alignment) == 5368 start_addr, "Check alignment"); 5369 assert((size_t)round_to((intptr_t)chunk_size, alignment) == 5370 chunk_size, "Check alignment"); 5371 5372 while (!pst->is_task_claimed(/* reference */ nth_task)) { 5373 // Having claimed the nth_task, compute corresponding mem-region, 5374 // which is a-fortiori aligned correctly (i.e. at a MUT bopundary). 5375 // The alignment restriction ensures that we do not need any 5376 // synchronization with other gang-workers while setting or 5377 // clearing bits in thus chunk of the MUT. 5378 MemRegion this_span = MemRegion(start_addr + nth_task*chunk_size, 5379 start_addr + (nth_task+1)*chunk_size); 5380 // The last chunk's end might be way beyond end of the 5381 // used region. In that case pull back appropriately. 5382 if (this_span.end() > end_addr) { 5383 this_span.set_end(end_addr); 5384 assert(!this_span.is_empty(), "Program logic (calculation of n_tasks)"); 5385 } 5386 // Iterate over the dirty cards covering this chunk, marking them 5387 // precleaned, and setting the corresponding bits in the mod union 5388 // table. Since we have been careful to partition at Card and MUT-word 5389 // boundaries no synchronization is needed between parallel threads. 5390 _collector->_ct->ct_bs()->dirty_card_iterate(this_span, 5391 &modUnionClosure); 5392 5393 // Having transferred these marks into the modUnionTable, 5394 // rescan the marked objects on the dirty cards in the modUnionTable. 5395 // Even if this is at a synchronous collection, the initial marking 5396 // may have been done during an asynchronous collection so there 5397 // may be dirty bits in the mod-union table. 5398 _collector->_modUnionTable.dirty_range_iterate_clear( 5399 this_span, &greyRescanClosure); 5400 _collector->_modUnionTable.verifyNoOneBitsInRange( 5401 this_span.start(), 5402 this_span.end()); 5403 } 5404 pst->all_tasks_completed(); // declare that i am done 5405 } 5406 5407 // . see if we can share work_queues with ParNew? XXX 5408 void 5409 CMSParRemarkTask::do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, 5410 int* seed) { 5411 OopTaskQueue* work_q = work_queue(i); 5412 NOT_PRODUCT(int num_steals = 0;) 5413 oop obj_to_scan; 5414 CMSBitMap* bm = &(_collector->_markBitMap); 5415 5416 while (true) { 5417 // Completely finish any left over work from (an) earlier round(s) 5418 cl->trim_queue(0); 5419 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 5420 (size_t)ParGCDesiredObjsFromOverflowList); 5421 // Now check if there's any work in the overflow list 5422 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads, 5423 // only affects the number of attempts made to get work from the 5424 // overflow list and does not affect the number of workers. Just 5425 // pass ParallelGCThreads so this behavior is unchanged. 5426 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 5427 work_q, 5428 ParallelGCThreads)) { 5429 // found something in global overflow list; 5430 // not yet ready to go stealing work from others. 5431 // We'd like to assert(work_q->size() != 0, ...) 5432 // because we just took work from the overflow list, 5433 // but of course we can't since all of that could have 5434 // been already stolen from us. 5435 // "He giveth and He taketh away." 5436 continue; 5437 } 5438 // Verify that we have no work before we resort to stealing 5439 assert(work_q->size() == 0, "Have work, shouldn't steal"); 5440 // Try to steal from other queues that have work 5441 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 5442 NOT_PRODUCT(num_steals++;) 5443 assert(obj_to_scan->is_oop(), "Oops, not an oop!"); 5444 assert(bm->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 5445 // Do scanning work 5446 obj_to_scan->oop_iterate(cl); 5447 // Loop around, finish this work, and try to steal some more 5448 } else if (terminator()->offer_termination()) { 5449 break; // nirvana from the infinite cycle 5450 } 5451 } 5452 NOT_PRODUCT( 5453 if (PrintCMSStatistics != 0) { 5454 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals); 5455 } 5456 ) 5457 assert(work_q->size() == 0 && _collector->overflow_list_is_empty(), 5458 "Else our work is not yet done"); 5459 } 5460 5461 // Return a thread-local PLAB recording array, as appropriate. 5462 void* CMSCollector::get_data_recorder(int thr_num) { 5463 if (_survivor_plab_array != NULL && 5464 (CMSPLABRecordAlways || 5465 (_collectorState > Marking && _collectorState < FinalMarking))) { 5466 assert(thr_num < (int)ParallelGCThreads, "thr_num is out of bounds"); 5467 ChunkArray* ca = &_survivor_plab_array[thr_num]; 5468 ca->reset(); // clear it so that fresh data is recorded 5469 return (void*) ca; 5470 } else { 5471 return NULL; 5472 } 5473 } 5474 5475 // Reset all the thread-local PLAB recording arrays 5476 void CMSCollector::reset_survivor_plab_arrays() { 5477 for (uint i = 0; i < ParallelGCThreads; i++) { 5478 _survivor_plab_array[i].reset(); 5479 } 5480 } 5481 5482 // Merge the per-thread plab arrays into the global survivor chunk 5483 // array which will provide the partitioning of the survivor space 5484 // for CMS rescan. 5485 void CMSCollector::merge_survivor_plab_arrays(ContiguousSpace* surv, 5486 int no_of_gc_threads) { 5487 assert(_survivor_plab_array != NULL, "Error"); 5488 assert(_survivor_chunk_array != NULL, "Error"); 5489 assert(_collectorState == FinalMarking, "Error"); 5490 for (int j = 0; j < no_of_gc_threads; j++) { 5491 _cursor[j] = 0; 5492 } 5493 HeapWord* top = surv->top(); 5494 size_t i; 5495 for (i = 0; i < _survivor_chunk_capacity; i++) { // all sca entries 5496 HeapWord* min_val = top; // Higher than any PLAB address 5497 uint min_tid = 0; // position of min_val this round 5498 for (int j = 0; j < no_of_gc_threads; j++) { 5499 ChunkArray* cur_sca = &_survivor_plab_array[j]; 5500 if (_cursor[j] == cur_sca->end()) { 5501 continue; 5502 } 5503 assert(_cursor[j] < cur_sca->end(), "ctl pt invariant"); 5504 HeapWord* cur_val = cur_sca->nth(_cursor[j]); 5505 assert(surv->used_region().contains(cur_val), "Out of bounds value"); 5506 if (cur_val < min_val) { 5507 min_tid = j; 5508 min_val = cur_val; 5509 } else { 5510 assert(cur_val < top, "All recorded addresses should be less"); 5511 } 5512 } 5513 // At this point min_val and min_tid are respectively 5514 // the least address in _survivor_plab_array[j]->nth(_cursor[j]) 5515 // and the thread (j) that witnesses that address. 5516 // We record this address in the _survivor_chunk_array[i] 5517 // and increment _cursor[min_tid] prior to the next round i. 5518 if (min_val == top) { 5519 break; 5520 } 5521 _survivor_chunk_array[i] = min_val; 5522 _cursor[min_tid]++; 5523 } 5524 // We are all done; record the size of the _survivor_chunk_array 5525 _survivor_chunk_index = i; // exclusive: [0, i) 5526 if (PrintCMSStatistics > 0) { 5527 gclog_or_tty->print(" (Survivor:" SIZE_FORMAT "chunks) ", i); 5528 } 5529 // Verify that we used up all the recorded entries 5530 #ifdef ASSERT 5531 size_t total = 0; 5532 for (int j = 0; j < no_of_gc_threads; j++) { 5533 assert(_cursor[j] == _survivor_plab_array[j].end(), "Ctl pt invariant"); 5534 total += _cursor[j]; 5535 } 5536 assert(total == _survivor_chunk_index, "Ctl Pt Invariant"); 5537 // Check that the merged array is in sorted order 5538 if (total > 0) { 5539 for (size_t i = 0; i < total - 1; i++) { 5540 if (PrintCMSStatistics > 0) { 5541 gclog_or_tty->print(" (chunk" SIZE_FORMAT ":" INTPTR_FORMAT ") ", 5542 i, _survivor_chunk_array[i]); 5543 } 5544 assert(_survivor_chunk_array[i] < _survivor_chunk_array[i+1], 5545 "Not sorted"); 5546 } 5547 } 5548 #endif // ASSERT 5549 } 5550 5551 // Set up the space's par_seq_tasks structure for work claiming 5552 // for parallel rescan of young gen. 5553 // See ParRescanTask where this is currently used. 5554 void 5555 CMSCollector:: 5556 initialize_sequential_subtasks_for_young_gen_rescan(int n_threads) { 5557 assert(n_threads > 0, "Unexpected n_threads argument"); 5558 DefNewGeneration* dng = (DefNewGeneration*)_young_gen; 5559 5560 // Eden space 5561 { 5562 SequentialSubTasksDone* pst = dng->eden()->par_seq_tasks(); 5563 assert(!pst->valid(), "Clobbering existing data?"); 5564 // Each valid entry in [0, _eden_chunk_index) represents a task. 5565 size_t n_tasks = _eden_chunk_index + 1; 5566 assert(n_tasks == 1 || _eden_chunk_array != NULL, "Error"); 5567 // Sets the condition for completion of the subtask (how many threads 5568 // need to finish in order to be done). 5569 pst->set_n_threads(n_threads); 5570 pst->set_n_tasks((int)n_tasks); 5571 } 5572 5573 // Merge the survivor plab arrays into _survivor_chunk_array 5574 if (_survivor_plab_array != NULL) { 5575 merge_survivor_plab_arrays(dng->from(), n_threads); 5576 } else { 5577 assert(_survivor_chunk_index == 0, "Error"); 5578 } 5579 5580 // To space 5581 { 5582 SequentialSubTasksDone* pst = dng->to()->par_seq_tasks(); 5583 assert(!pst->valid(), "Clobbering existing data?"); 5584 // Sets the condition for completion of the subtask (how many threads 5585 // need to finish in order to be done). 5586 pst->set_n_threads(n_threads); 5587 pst->set_n_tasks(1); 5588 assert(pst->valid(), "Error"); 5589 } 5590 5591 // From space 5592 { 5593 SequentialSubTasksDone* pst = dng->from()->par_seq_tasks(); 5594 assert(!pst->valid(), "Clobbering existing data?"); 5595 size_t n_tasks = _survivor_chunk_index + 1; 5596 assert(n_tasks == 1 || _survivor_chunk_array != NULL, "Error"); 5597 // Sets the condition for completion of the subtask (how many threads 5598 // need to finish in order to be done). 5599 pst->set_n_threads(n_threads); 5600 pst->set_n_tasks((int)n_tasks); 5601 assert(pst->valid(), "Error"); 5602 } 5603 } 5604 5605 // Parallel version of remark 5606 void CMSCollector::do_remark_parallel() { 5607 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5608 FlexibleWorkGang* workers = gch->workers(); 5609 assert(workers != NULL, "Need parallel worker threads."); 5610 // Choose to use the number of GC workers most recently set 5611 // into "active_workers". If active_workers is not set, set it 5612 // to ParallelGCThreads. 5613 int n_workers = workers->active_workers(); 5614 if (n_workers == 0) { 5615 assert(n_workers > 0, "Should have been set during scavenge"); 5616 n_workers = ParallelGCThreads; 5617 workers->set_active_workers(n_workers); 5618 } 5619 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 5620 CompactibleFreeListSpace* perm_space = _permGen->cmsSpace(); 5621 5622 CMSParRemarkTask tsk(this, 5623 cms_space, perm_space, 5624 n_workers, workers, task_queues()); 5625 5626 // Set up for parallel process_strong_roots work. 5627 gch->set_par_threads(n_workers); 5628 // We won't be iterating over the cards in the card table updating 5629 // the younger_gen cards, so we shouldn't call the following else 5630 // the verification code as well as subsequent younger_refs_iterate 5631 // code would get confused. XXX 5632 // gch->rem_set()->prepare_for_younger_refs_iterate(true); // parallel 5633 5634 // The young gen rescan work will not be done as part of 5635 // process_strong_roots (which currently doesn't knw how to 5636 // parallelize such a scan), but rather will be broken up into 5637 // a set of parallel tasks (via the sampling that the [abortable] 5638 // preclean phase did of EdenSpace, plus the [two] tasks of 5639 // scanning the [two] survivor spaces. Further fine-grain 5640 // parallelization of the scanning of the survivor spaces 5641 // themselves, and of precleaning of the younger gen itself 5642 // is deferred to the future. 5643 initialize_sequential_subtasks_for_young_gen_rescan(n_workers); 5644 5645 // The dirty card rescan work is broken up into a "sequence" 5646 // of parallel tasks (per constituent space) that are dynamically 5647 // claimed by the parallel threads. 5648 cms_space->initialize_sequential_subtasks_for_rescan(n_workers); 5649 perm_space->initialize_sequential_subtasks_for_rescan(n_workers); 5650 5651 // It turns out that even when we're using 1 thread, doing the work in a 5652 // separate thread causes wide variance in run times. We can't help this 5653 // in the multi-threaded case, but we special-case n=1 here to get 5654 // repeatable measurements of the 1-thread overhead of the parallel code. 5655 if (n_workers > 1) { 5656 // Make refs discovery MT-safe, if it isn't already: it may not 5657 // necessarily be so, since it's possible that we are doing 5658 // ST marking. 5659 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), true); 5660 GenCollectedHeap::StrongRootsScope srs(gch); 5661 workers->run_task(&tsk); 5662 } else { 5663 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false); 5664 GenCollectedHeap::StrongRootsScope srs(gch); 5665 tsk.work(0); 5666 } 5667 gch->set_par_threads(0); // 0 ==> non-parallel. 5668 // restore, single-threaded for now, any preserved marks 5669 // as a result of work_q overflow 5670 restore_preserved_marks_if_any(); 5671 } 5672 5673 // Non-parallel version of remark 5674 void CMSCollector::do_remark_non_parallel() { 5675 ResourceMark rm; 5676 HandleMark hm; 5677 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5678 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false); 5679 5680 MarkRefsIntoAndScanClosure 5681 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable, 5682 &_markStack, &_revisitStack, this, 5683 false /* should_yield */, false /* not precleaning */); 5684 MarkFromDirtyCardsClosure 5685 markFromDirtyCardsClosure(this, _span, 5686 NULL, // space is set further below 5687 &_markBitMap, &_markStack, &_revisitStack, 5688 &mrias_cl); 5689 { 5690 GCTraceTime t("grey object rescan", PrintGCDetails, false, _gc_timer_cm); 5691 // Iterate over the dirty cards, setting the corresponding bits in the 5692 // mod union table. 5693 { 5694 ModUnionClosure modUnionClosure(&_modUnionTable); 5695 _ct->ct_bs()->dirty_card_iterate( 5696 _cmsGen->used_region(), 5697 &modUnionClosure); 5698 _ct->ct_bs()->dirty_card_iterate( 5699 _permGen->used_region(), 5700 &modUnionClosure); 5701 } 5702 // Having transferred these marks into the modUnionTable, we just need 5703 // to rescan the marked objects on the dirty cards in the modUnionTable. 5704 // The initial marking may have been done during an asynchronous 5705 // collection so there may be dirty bits in the mod-union table. 5706 const int alignment = 5707 CardTableModRefBS::card_size * BitsPerWord; 5708 { 5709 // ... First handle dirty cards in CMS gen 5710 markFromDirtyCardsClosure.set_space(_cmsGen->cmsSpace()); 5711 MemRegion ur = _cmsGen->used_region(); 5712 HeapWord* lb = ur.start(); 5713 HeapWord* ub = (HeapWord*)round_to((intptr_t)ur.end(), alignment); 5714 MemRegion cms_span(lb, ub); 5715 _modUnionTable.dirty_range_iterate_clear(cms_span, 5716 &markFromDirtyCardsClosure); 5717 verify_work_stacks_empty(); 5718 if (PrintCMSStatistics != 0) { 5719 gclog_or_tty->print(" (re-scanned "SIZE_FORMAT" dirty cards in cms gen) ", 5720 markFromDirtyCardsClosure.num_dirty_cards()); 5721 } 5722 } 5723 { 5724 // .. and then repeat for dirty cards in perm gen 5725 markFromDirtyCardsClosure.set_space(_permGen->cmsSpace()); 5726 MemRegion ur = _permGen->used_region(); 5727 HeapWord* lb = ur.start(); 5728 HeapWord* ub = (HeapWord*)round_to((intptr_t)ur.end(), alignment); 5729 MemRegion perm_span(lb, ub); 5730 _modUnionTable.dirty_range_iterate_clear(perm_span, 5731 &markFromDirtyCardsClosure); 5732 verify_work_stacks_empty(); 5733 if (PrintCMSStatistics != 0) { 5734 gclog_or_tty->print(" (re-scanned "SIZE_FORMAT" dirty cards in perm gen) ", 5735 markFromDirtyCardsClosure.num_dirty_cards()); 5736 } 5737 } 5738 } 5739 if (VerifyDuringGC && 5740 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 5741 HandleMark hm; // Discard invalid handles created during verification 5742 Universe::verify(); 5743 } 5744 { 5745 GCTraceTime t("root rescan", PrintGCDetails, false, _gc_timer_cm); 5746 5747 verify_work_stacks_empty(); 5748 5749 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 5750 GenCollectedHeap::StrongRootsScope srs(gch); 5751 gch->gen_process_strong_roots(_cmsGen->level(), 5752 true, // younger gens as roots 5753 false, // use the local StrongRootsScope 5754 true, // collecting perm gen 5755 SharedHeap::ScanningOption(roots_scanning_options()), 5756 &mrias_cl, 5757 true, // walk code active on stacks 5758 NULL); 5759 assert(should_unload_classes() 5760 || (roots_scanning_options() & SharedHeap::SO_CodeCache), 5761 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 5762 } 5763 verify_work_stacks_empty(); 5764 // Restore evacuated mark words, if any, used for overflow list links 5765 if (!CMSOverflowEarlyRestoration) { 5766 restore_preserved_marks_if_any(); 5767 } 5768 verify_overflow_empty(); 5769 } 5770 5771 //////////////////////////////////////////////////////// 5772 // Parallel Reference Processing Task Proxy Class 5773 //////////////////////////////////////////////////////// 5774 class CMSRefProcTaskProxy: public AbstractGangTaskWOopQueues { 5775 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask; 5776 CMSCollector* _collector; 5777 CMSBitMap* _mark_bit_map; 5778 const MemRegion _span; 5779 ProcessTask& _task; 5780 5781 public: 5782 CMSRefProcTaskProxy(ProcessTask& task, 5783 CMSCollector* collector, 5784 const MemRegion& span, 5785 CMSBitMap* mark_bit_map, 5786 AbstractWorkGang* workers, 5787 OopTaskQueueSet* task_queues): 5788 // XXX Should superclass AGTWOQ also know about AWG since it knows 5789 // about the task_queues used by the AWG? Then it could initialize 5790 // the terminator() object. See 6984287. The set_for_termination() 5791 // below is a temporary band-aid for the regression in 6984287. 5792 AbstractGangTaskWOopQueues("Process referents by policy in parallel", 5793 task_queues), 5794 _task(task), 5795 _collector(collector), _span(span), _mark_bit_map(mark_bit_map) 5796 { 5797 assert(_collector->_span.equals(_span) && !_span.is_empty(), 5798 "Inconsistency in _span"); 5799 set_for_termination(workers->active_workers()); 5800 } 5801 5802 OopTaskQueueSet* task_queues() { return queues(); } 5803 5804 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 5805 5806 void do_work_steal(int i, 5807 CMSParDrainMarkingStackClosure* drain, 5808 CMSParKeepAliveClosure* keep_alive, 5809 int* seed); 5810 5811 virtual void work(uint worker_id); 5812 }; 5813 5814 void CMSRefProcTaskProxy::work(uint worker_id) { 5815 assert(_collector->_span.equals(_span), "Inconsistency in _span"); 5816 CMSParKeepAliveClosure par_keep_alive(_collector, _span, 5817 _mark_bit_map, 5818 &_collector->_revisitStack, 5819 work_queue(worker_id)); 5820 CMSParDrainMarkingStackClosure par_drain_stack(_collector, _span, 5821 _mark_bit_map, 5822 &_collector->_revisitStack, 5823 work_queue(worker_id)); 5824 CMSIsAliveClosure is_alive_closure(_span, _mark_bit_map); 5825 _task.work(worker_id, is_alive_closure, par_keep_alive, par_drain_stack); 5826 if (_task.marks_oops_alive()) { 5827 do_work_steal(worker_id, &par_drain_stack, &par_keep_alive, 5828 _collector->hash_seed(worker_id)); 5829 } 5830 assert(work_queue(worker_id)->size() == 0, "work_queue should be empty"); 5831 assert(_collector->_overflow_list == NULL, "non-empty _overflow_list"); 5832 } 5833 5834 class CMSRefEnqueueTaskProxy: public AbstractGangTask { 5835 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask; 5836 EnqueueTask& _task; 5837 5838 public: 5839 CMSRefEnqueueTaskProxy(EnqueueTask& task) 5840 : AbstractGangTask("Enqueue reference objects in parallel"), 5841 _task(task) 5842 { } 5843 5844 virtual void work(uint worker_id) 5845 { 5846 _task.work(worker_id); 5847 } 5848 }; 5849 5850 CMSParKeepAliveClosure::CMSParKeepAliveClosure(CMSCollector* collector, 5851 MemRegion span, CMSBitMap* bit_map, CMSMarkStack* revisit_stack, 5852 OopTaskQueue* work_queue): 5853 Par_KlassRememberingOopClosure(collector, NULL, revisit_stack), 5854 _span(span), 5855 _bit_map(bit_map), 5856 _work_queue(work_queue), 5857 _mark_and_push(collector, span, bit_map, revisit_stack, work_queue), 5858 _low_water_mark(MIN2((uint)(work_queue->max_elems()/4), 5859 (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads))) 5860 { } 5861 5862 // . see if we can share work_queues with ParNew? XXX 5863 void CMSRefProcTaskProxy::do_work_steal(int i, 5864 CMSParDrainMarkingStackClosure* drain, 5865 CMSParKeepAliveClosure* keep_alive, 5866 int* seed) { 5867 OopTaskQueue* work_q = work_queue(i); 5868 NOT_PRODUCT(int num_steals = 0;) 5869 oop obj_to_scan; 5870 5871 while (true) { 5872 // Completely finish any left over work from (an) earlier round(s) 5873 drain->trim_queue(0); 5874 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 5875 (size_t)ParGCDesiredObjsFromOverflowList); 5876 // Now check if there's any work in the overflow list 5877 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads, 5878 // only affects the number of attempts made to get work from the 5879 // overflow list and does not affect the number of workers. Just 5880 // pass ParallelGCThreads so this behavior is unchanged. 5881 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 5882 work_q, 5883 ParallelGCThreads)) { 5884 // Found something in global overflow list; 5885 // not yet ready to go stealing work from others. 5886 // We'd like to assert(work_q->size() != 0, ...) 5887 // because we just took work from the overflow list, 5888 // but of course we can't, since all of that might have 5889 // been already stolen from us. 5890 continue; 5891 } 5892 // Verify that we have no work before we resort to stealing 5893 assert(work_q->size() == 0, "Have work, shouldn't steal"); 5894 // Try to steal from other queues that have work 5895 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 5896 NOT_PRODUCT(num_steals++;) 5897 assert(obj_to_scan->is_oop(), "Oops, not an oop!"); 5898 assert(_mark_bit_map->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 5899 // Do scanning work 5900 obj_to_scan->oop_iterate(keep_alive); 5901 // Loop around, finish this work, and try to steal some more 5902 } else if (terminator()->offer_termination()) { 5903 break; // nirvana from the infinite cycle 5904 } 5905 } 5906 NOT_PRODUCT( 5907 if (PrintCMSStatistics != 0) { 5908 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals); 5909 } 5910 ) 5911 } 5912 5913 void CMSRefProcTaskExecutor::execute(ProcessTask& task) 5914 { 5915 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5916 FlexibleWorkGang* workers = gch->workers(); 5917 assert(workers != NULL, "Need parallel worker threads."); 5918 CMSRefProcTaskProxy rp_task(task, &_collector, 5919 _collector.ref_processor()->span(), 5920 _collector.markBitMap(), 5921 workers, _collector.task_queues()); 5922 workers->run_task(&rp_task); 5923 } 5924 5925 void CMSRefProcTaskExecutor::execute(EnqueueTask& task) 5926 { 5927 5928 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5929 FlexibleWorkGang* workers = gch->workers(); 5930 assert(workers != NULL, "Need parallel worker threads."); 5931 CMSRefEnqueueTaskProxy enq_task(task); 5932 workers->run_task(&enq_task); 5933 } 5934 5935 void CMSCollector::refProcessingWork(bool asynch, bool clear_all_soft_refs) { 5936 5937 ResourceMark rm; 5938 HandleMark hm; 5939 5940 ReferenceProcessor* rp = ref_processor(); 5941 assert(rp->span().equals(_span), "Spans should be equal"); 5942 assert(!rp->enqueuing_is_done(), "Enqueuing should not be complete"); 5943 // Process weak references. 5944 rp->setup_policy(clear_all_soft_refs); 5945 verify_work_stacks_empty(); 5946 5947 CMSKeepAliveClosure cmsKeepAliveClosure(this, _span, &_markBitMap, 5948 &_markStack, &_revisitStack, 5949 false /* !preclean */); 5950 CMSDrainMarkingStackClosure cmsDrainMarkingStackClosure(this, 5951 _span, &_markBitMap, &_markStack, 5952 &cmsKeepAliveClosure, false /* !preclean */); 5953 { 5954 GCTraceTime t("weak refs processing", PrintGCDetails, false, _gc_timer_cm); 5955 5956 ReferenceProcessorStats stats; 5957 if (rp->processing_is_mt()) { 5958 // Set the degree of MT here. If the discovery is done MT, there 5959 // may have been a different number of threads doing the discovery 5960 // and a different number of discovered lists may have Ref objects. 5961 // That is OK as long as the Reference lists are balanced (see 5962 // balance_all_queues() and balance_queues()). 5963 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5964 int active_workers = ParallelGCThreads; 5965 FlexibleWorkGang* workers = gch->workers(); 5966 if (workers != NULL) { 5967 active_workers = workers->active_workers(); 5968 // The expectation is that active_workers will have already 5969 // been set to a reasonable value. If it has not been set, 5970 // investigate. 5971 assert(active_workers > 0, "Should have been set during scavenge"); 5972 } 5973 rp->set_active_mt_degree(active_workers); 5974 CMSRefProcTaskExecutor task_executor(*this); 5975 stats = rp->process_discovered_references(&_is_alive_closure, 5976 &cmsKeepAliveClosure, 5977 &cmsDrainMarkingStackClosure, 5978 &task_executor, 5979 _gc_timer_cm); 5980 } else { 5981 stats = rp->process_discovered_references(&_is_alive_closure, 5982 &cmsKeepAliveClosure, 5983 &cmsDrainMarkingStackClosure, 5984 NULL, 5985 _gc_timer_cm); 5986 } 5987 _gc_tracer_cm->report_gc_reference_stats(stats); 5988 5989 verify_work_stacks_empty(); 5990 } 5991 5992 if (should_unload_classes()) { 5993 { 5994 GCTraceTime t("class unloading", PrintGCDetails, false, _gc_timer_cm); 5995 5996 // Follow SystemDictionary roots and unload classes 5997 bool purged_class = SystemDictionary::do_unloading(&_is_alive_closure); 5998 5999 // Follow CodeCache roots and unload any methods marked for unloading 6000 CodeCache::do_unloading(&_is_alive_closure, 6001 &cmsKeepAliveClosure, 6002 purged_class); 6003 6004 cmsDrainMarkingStackClosure.do_void(); 6005 verify_work_stacks_empty(); 6006 6007 // Update subklass/sibling/implementor links in KlassKlass descendants 6008 assert(!_revisitStack.isEmpty(), "revisit stack should not be empty"); 6009 oop k; 6010 while ((k = _revisitStack.pop()) != NULL) { 6011 ((Klass*)(oopDesc*)k)->follow_weak_klass_links( 6012 &_is_alive_closure, 6013 &cmsKeepAliveClosure); 6014 } 6015 assert(!ClassUnloading || 6016 (_markStack.isEmpty() && overflow_list_is_empty()), 6017 "Should not have found new reachable objects"); 6018 assert(_revisitStack.isEmpty(), "revisit stack should have been drained"); 6019 cmsDrainMarkingStackClosure.do_void(); 6020 verify_work_stacks_empty(); 6021 } 6022 6023 { 6024 GCTraceTime t("scrub symbol table", PrintGCDetails, false, _gc_timer_cm); 6025 // Clean up unreferenced symbols in symbol table. 6026 SymbolTable::unlink(); 6027 } 6028 } 6029 6030 if (should_unload_classes() || !JavaObjectsInPerm) { 6031 GCTraceTime t("scrub string table", PrintGCDetails, false, _gc_timer_cm); 6032 // Now clean up stale oops in StringTable 6033 StringTable::unlink(&_is_alive_closure); 6034 } 6035 6036 verify_work_stacks_empty(); 6037 // Restore any preserved marks as a result of mark stack or 6038 // work queue overflow 6039 restore_preserved_marks_if_any(); // done single-threaded for now 6040 6041 rp->set_enqueuing_is_done(true); 6042 if (rp->processing_is_mt()) { 6043 rp->balance_all_queues(); 6044 CMSRefProcTaskExecutor task_executor(*this); 6045 rp->enqueue_discovered_references(&task_executor); 6046 } else { 6047 rp->enqueue_discovered_references(NULL); 6048 } 6049 rp->verify_no_references_recorded(); 6050 assert(!rp->discovery_enabled(), "should have been disabled"); 6051 } 6052 6053 #ifndef PRODUCT 6054 void CMSCollector::check_correct_thread_executing() { 6055 Thread* t = Thread::current(); 6056 // Only the VM thread or the CMS thread should be here. 6057 assert(t->is_ConcurrentGC_thread() || t->is_VM_thread(), 6058 "Unexpected thread type"); 6059 // If this is the vm thread, the foreground process 6060 // should not be waiting. Note that _foregroundGCIsActive is 6061 // true while the foreground collector is waiting. 6062 if (_foregroundGCShouldWait) { 6063 // We cannot be the VM thread 6064 assert(t->is_ConcurrentGC_thread(), 6065 "Should be CMS thread"); 6066 } else { 6067 // We can be the CMS thread only if we are in a stop-world 6068 // phase of CMS collection. 6069 if (t->is_ConcurrentGC_thread()) { 6070 assert(_collectorState == InitialMarking || 6071 _collectorState == FinalMarking, 6072 "Should be a stop-world phase"); 6073 // The CMS thread should be holding the CMS_token. 6074 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6075 "Potential interference with concurrently " 6076 "executing VM thread"); 6077 } 6078 } 6079 } 6080 #endif 6081 6082 void CMSCollector::sweep(bool asynch) { 6083 assert(_collectorState == Sweeping, "just checking"); 6084 check_correct_thread_executing(); 6085 verify_work_stacks_empty(); 6086 verify_overflow_empty(); 6087 increment_sweep_count(); 6088 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 6089 6090 _inter_sweep_timer.stop(); 6091 _inter_sweep_estimate.sample(_inter_sweep_timer.seconds()); 6092 size_policy()->avg_cms_free_at_sweep()->sample(_cmsGen->free()); 6093 6094 // PermGen verification support: If perm gen sweeping is disabled in 6095 // this cycle, we preserve the perm gen object "deadness" information 6096 // in the perm_gen_verify_bit_map. In order to do that we traverse 6097 // all blocks in perm gen and mark all dead objects. 6098 if (verifying() && !should_unload_classes()) { 6099 assert(perm_gen_verify_bit_map()->sizeInBits() != 0, 6100 "Should have already been allocated"); 6101 MarkDeadObjectsClosure mdo(this, _permGen->cmsSpace(), 6102 markBitMap(), perm_gen_verify_bit_map()); 6103 if (asynch) { 6104 CMSTokenSyncWithLocks ts(true, _permGen->freelistLock(), 6105 bitMapLock()); 6106 _permGen->cmsSpace()->blk_iterate(&mdo); 6107 } else { 6108 // In the case of synchronous sweep, we already have 6109 // the requisite locks/tokens. 6110 _permGen->cmsSpace()->blk_iterate(&mdo); 6111 } 6112 } 6113 6114 assert(!_intra_sweep_timer.is_active(), "Should not be active"); 6115 _intra_sweep_timer.reset(); 6116 _intra_sweep_timer.start(); 6117 if (asynch) { 6118 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 6119 CMSPhaseAccounting pa(this, "sweep", !PrintGCDetails); 6120 // First sweep the old gen then the perm gen 6121 { 6122 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(), 6123 bitMapLock()); 6124 sweepWork(_cmsGen, asynch); 6125 } 6126 6127 // Now repeat for perm gen 6128 if (should_unload_classes()) { 6129 CMSTokenSyncWithLocks ts(true, _permGen->freelistLock(), 6130 bitMapLock()); 6131 sweepWork(_permGen, asynch); 6132 } 6133 6134 // Update Universe::_heap_*_at_gc figures. 6135 // We need all the free list locks to make the abstract state 6136 // transition from Sweeping to Resetting. See detailed note 6137 // further below. 6138 { 6139 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(), 6140 _permGen->freelistLock()); 6141 // Update heap occupancy information which is used as 6142 // input to soft ref clearing policy at the next gc. 6143 Universe::update_heap_info_at_gc(); 6144 _collectorState = Resizing; 6145 } 6146 } else { 6147 // already have needed locks 6148 sweepWork(_cmsGen, asynch); 6149 6150 if (should_unload_classes()) { 6151 sweepWork(_permGen, asynch); 6152 } 6153 // Update heap occupancy information which is used as 6154 // input to soft ref clearing policy at the next gc. 6155 Universe::update_heap_info_at_gc(); 6156 _collectorState = Resizing; 6157 } 6158 verify_work_stacks_empty(); 6159 verify_overflow_empty(); 6160 6161 _intra_sweep_timer.stop(); 6162 _intra_sweep_estimate.sample(_intra_sweep_timer.seconds()); 6163 6164 _inter_sweep_timer.reset(); 6165 _inter_sweep_timer.start(); 6166 6167 // We need to use a monotonically non-deccreasing time in ms 6168 // or we will see time-warp warnings and os::javaTimeMillis() 6169 // does not guarantee monotonicity. 6170 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC; 6171 update_time_of_last_gc(now); 6172 6173 // NOTE on abstract state transitions: 6174 // Mutators allocate-live and/or mark the mod-union table dirty 6175 // based on the state of the collection. The former is done in 6176 // the interval [Marking, Sweeping] and the latter in the interval 6177 // [Marking, Sweeping). Thus the transitions into the Marking state 6178 // and out of the Sweeping state must be synchronously visible 6179 // globally to the mutators. 6180 // The transition into the Marking state happens with the world 6181 // stopped so the mutators will globally see it. Sweeping is 6182 // done asynchronously by the background collector so the transition 6183 // from the Sweeping state to the Resizing state must be done 6184 // under the freelistLock (as is the check for whether to 6185 // allocate-live and whether to dirty the mod-union table). 6186 assert(_collectorState == Resizing, "Change of collector state to" 6187 " Resizing must be done under the freelistLocks (plural)"); 6188 6189 // Now that sweeping has been completed, we clear 6190 // the incremental_collection_failed flag, 6191 // thus inviting a younger gen collection to promote into 6192 // this generation. If such a promotion may still fail, 6193 // the flag will be set again when a young collection is 6194 // attempted. 6195 GenCollectedHeap* gch = GenCollectedHeap::heap(); 6196 gch->clear_incremental_collection_failed(); // Worth retrying as fresh space may have been freed up 6197 gch->update_full_collections_completed(_collection_count_start); 6198 } 6199 6200 // FIX ME!!! Looks like this belongs in CFLSpace, with 6201 // CMSGen merely delegating to it. 6202 void ConcurrentMarkSweepGeneration::setNearLargestChunk() { 6203 double nearLargestPercent = FLSLargestBlockCoalesceProximity; 6204 HeapWord* minAddr = _cmsSpace->bottom(); 6205 HeapWord* largestAddr = 6206 (HeapWord*) _cmsSpace->dictionary()->find_largest_dict(); 6207 if (largestAddr == NULL) { 6208 // The dictionary appears to be empty. In this case 6209 // try to coalesce at the end of the heap. 6210 largestAddr = _cmsSpace->end(); 6211 } 6212 size_t largestOffset = pointer_delta(largestAddr, minAddr); 6213 size_t nearLargestOffset = 6214 (size_t)((double)largestOffset * nearLargestPercent) - MinChunkSize; 6215 if (PrintFLSStatistics != 0) { 6216 gclog_or_tty->print_cr( 6217 "CMS: Large Block: " PTR_FORMAT ";" 6218 " Proximity: " PTR_FORMAT " -> " PTR_FORMAT, 6219 largestAddr, 6220 _cmsSpace->nearLargestChunk(), minAddr + nearLargestOffset); 6221 } 6222 _cmsSpace->set_nearLargestChunk(minAddr + nearLargestOffset); 6223 } 6224 6225 bool ConcurrentMarkSweepGeneration::isNearLargestChunk(HeapWord* addr) { 6226 return addr >= _cmsSpace->nearLargestChunk(); 6227 } 6228 6229 FreeChunk* ConcurrentMarkSweepGeneration::find_chunk_at_end() { 6230 return _cmsSpace->find_chunk_at_end(); 6231 } 6232 6233 void ConcurrentMarkSweepGeneration::update_gc_stats(int current_level, 6234 bool full) { 6235 // The next lower level has been collected. Gather any statistics 6236 // that are of interest at this point. 6237 if (!full && (current_level + 1) == level()) { 6238 // Gather statistics on the young generation collection. 6239 collector()->stats().record_gc0_end(used()); 6240 } 6241 } 6242 6243 CMSAdaptiveSizePolicy* ConcurrentMarkSweepGeneration::size_policy() { 6244 GenCollectedHeap* gch = GenCollectedHeap::heap(); 6245 assert(gch->kind() == CollectedHeap::GenCollectedHeap, 6246 "Wrong type of heap"); 6247 CMSAdaptiveSizePolicy* sp = (CMSAdaptiveSizePolicy*) 6248 gch->gen_policy()->size_policy(); 6249 assert(sp->is_gc_cms_adaptive_size_policy(), 6250 "Wrong type of size policy"); 6251 return sp; 6252 } 6253 6254 void ConcurrentMarkSweepGeneration::rotate_debug_collection_type() { 6255 if (PrintGCDetails && Verbose) { 6256 gclog_or_tty->print("Rotate from %d ", _debug_collection_type); 6257 } 6258 _debug_collection_type = (CollectionTypes) (_debug_collection_type + 1); 6259 _debug_collection_type = 6260 (CollectionTypes) (_debug_collection_type % Unknown_collection_type); 6261 if (PrintGCDetails && Verbose) { 6262 gclog_or_tty->print_cr("to %d ", _debug_collection_type); 6263 } 6264 } 6265 6266 void CMSCollector::sweepWork(ConcurrentMarkSweepGeneration* gen, 6267 bool asynch) { 6268 // We iterate over the space(s) underlying this generation, 6269 // checking the mark bit map to see if the bits corresponding 6270 // to specific blocks are marked or not. Blocks that are 6271 // marked are live and are not swept up. All remaining blocks 6272 // are swept up, with coalescing on-the-fly as we sweep up 6273 // contiguous free and/or garbage blocks: 6274 // We need to ensure that the sweeper synchronizes with allocators 6275 // and stop-the-world collectors. In particular, the following 6276 // locks are used: 6277 // . CMS token: if this is held, a stop the world collection cannot occur 6278 // . freelistLock: if this is held no allocation can occur from this 6279 // generation by another thread 6280 // . bitMapLock: if this is held, no other thread can access or update 6281 // 6282 6283 // Note that we need to hold the freelistLock if we use 6284 // block iterate below; else the iterator might go awry if 6285 // a mutator (or promotion) causes block contents to change 6286 // (for instance if the allocator divvies up a block). 6287 // If we hold the free list lock, for all practical purposes 6288 // young generation GC's can't occur (they'll usually need to 6289 // promote), so we might as well prevent all young generation 6290 // GC's while we do a sweeping step. For the same reason, we might 6291 // as well take the bit map lock for the entire duration 6292 6293 // check that we hold the requisite locks 6294 assert(have_cms_token(), "Should hold cms token"); 6295 assert( (asynch && ConcurrentMarkSweepThread::cms_thread_has_cms_token()) 6296 || (!asynch && ConcurrentMarkSweepThread::vm_thread_has_cms_token()), 6297 "Should possess CMS token to sweep"); 6298 assert_lock_strong(gen->freelistLock()); 6299 assert_lock_strong(bitMapLock()); 6300 6301 assert(!_inter_sweep_timer.is_active(), "Was switched off in an outer context"); 6302 assert(_intra_sweep_timer.is_active(), "Was switched on in an outer context"); 6303 gen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()), 6304 _inter_sweep_estimate.padded_average(), 6305 _intra_sweep_estimate.padded_average()); 6306 gen->setNearLargestChunk(); 6307 6308 { 6309 SweepClosure sweepClosure(this, gen, &_markBitMap, 6310 CMSYield && asynch); 6311 gen->cmsSpace()->blk_iterate_careful(&sweepClosure); 6312 // We need to free-up/coalesce garbage/blocks from a 6313 // co-terminal free run. This is done in the SweepClosure 6314 // destructor; so, do not remove this scope, else the 6315 // end-of-sweep-census below will be off by a little bit. 6316 } 6317 gen->cmsSpace()->sweep_completed(); 6318 gen->cmsSpace()->endSweepFLCensus(sweep_count()); 6319 if (should_unload_classes()) { // unloaded classes this cycle, 6320 _concurrent_cycles_since_last_unload = 0; // ... reset count 6321 } else { // did not unload classes, 6322 _concurrent_cycles_since_last_unload++; // ... increment count 6323 } 6324 } 6325 6326 // Reset CMS data structures (for now just the marking bit map) 6327 // preparatory for the next cycle. 6328 void CMSCollector::reset(bool asynch) { 6329 GenCollectedHeap* gch = GenCollectedHeap::heap(); 6330 CMSAdaptiveSizePolicy* sp = size_policy(); 6331 AdaptiveSizePolicyOutput(sp, gch->total_collections()); 6332 if (asynch) { 6333 CMSTokenSyncWithLocks ts(true, bitMapLock()); 6334 6335 // If the state is not "Resetting", the foreground thread 6336 // has done a collection and the resetting. 6337 if (_collectorState != Resetting) { 6338 assert(_collectorState == Idling, "The state should only change" 6339 " because the foreground collector has finished the collection"); 6340 return; 6341 } 6342 6343 // Clear the mark bitmap (no grey objects to start with) 6344 // for the next cycle. 6345 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 6346 CMSPhaseAccounting cmspa(this, "reset", !PrintGCDetails); 6347 6348 HeapWord* curAddr = _markBitMap.startWord(); 6349 while (curAddr < _markBitMap.endWord()) { 6350 size_t remaining = pointer_delta(_markBitMap.endWord(), curAddr); 6351 MemRegion chunk(curAddr, MIN2(CMSBitMapYieldQuantum, remaining)); 6352 _markBitMap.clear_large_range(chunk); 6353 if (ConcurrentMarkSweepThread::should_yield() && 6354 !foregroundGCIsActive() && 6355 CMSYield) { 6356 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6357 "CMS thread should hold CMS token"); 6358 assert_lock_strong(bitMapLock()); 6359 bitMapLock()->unlock(); 6360 ConcurrentMarkSweepThread::desynchronize(true); 6361 ConcurrentMarkSweepThread::acknowledge_yield_request(); 6362 stopTimer(); 6363 if (PrintCMSStatistics != 0) { 6364 incrementYields(); 6365 } 6366 icms_wait(); 6367 6368 // See the comment in coordinator_yield() 6369 for (unsigned i = 0; i < CMSYieldSleepCount && 6370 ConcurrentMarkSweepThread::should_yield() && 6371 !CMSCollector::foregroundGCIsActive(); ++i) { 6372 os::sleep(Thread::current(), 1, false); 6373 ConcurrentMarkSweepThread::acknowledge_yield_request(); 6374 } 6375 6376 ConcurrentMarkSweepThread::synchronize(true); 6377 bitMapLock()->lock_without_safepoint_check(); 6378 startTimer(); 6379 } 6380 curAddr = chunk.end(); 6381 } 6382 // A successful mostly concurrent collection has been done. 6383 // Because only the full (i.e., concurrent mode failure) collections 6384 // are being measured for gc overhead limits, clean the "near" flag 6385 // and count. 6386 sp->reset_gc_overhead_limit_count(); 6387 _collectorState = Idling; 6388 } else { 6389 // already have the lock 6390 assert(_collectorState == Resetting, "just checking"); 6391 assert_lock_strong(bitMapLock()); 6392 _markBitMap.clear_all(); 6393 _collectorState = Idling; 6394 } 6395 6396 // Stop incremental mode after a cycle completes, so that any future cycles 6397 // are triggered by allocation. 6398 stop_icms(); 6399 6400 NOT_PRODUCT( 6401 if (RotateCMSCollectionTypes) { 6402 _cmsGen->rotate_debug_collection_type(); 6403 } 6404 ) 6405 6406 register_gc_end(); 6407 } 6408 6409 void CMSCollector::do_CMS_operation(CMS_op_type op, GCCause::Cause gc_cause) { 6410 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 6411 GCTraceTime t(GCCauseString("GC", gc_cause), PrintGC, !PrintGCDetails, NULL); 6412 TraceCollectorStats tcs(counters()); 6413 6414 switch (op) { 6415 case CMS_op_checkpointRootsInitial: { 6416 SvcGCMarker sgcm(SvcGCMarker::OTHER); 6417 checkpointRootsInitial(true); // asynch 6418 if (PrintGC) { 6419 _cmsGen->printOccupancy("initial-mark"); 6420 } 6421 break; 6422 } 6423 case CMS_op_checkpointRootsFinal: { 6424 SvcGCMarker sgcm(SvcGCMarker::OTHER); 6425 checkpointRootsFinal(true, // asynch 6426 false, // !clear_all_soft_refs 6427 false); // !init_mark_was_synchronous 6428 if (PrintGC) { 6429 _cmsGen->printOccupancy("remark"); 6430 } 6431 break; 6432 } 6433 default: 6434 fatal("No such CMS_op"); 6435 } 6436 } 6437 6438 #ifndef PRODUCT 6439 size_t const CMSCollector::skip_header_HeapWords() { 6440 return FreeChunk::header_size(); 6441 } 6442 6443 // Try and collect here conditions that should hold when 6444 // CMS thread is exiting. The idea is that the foreground GC 6445 // thread should not be blocked if it wants to terminate 6446 // the CMS thread and yet continue to run the VM for a while 6447 // after that. 6448 void CMSCollector::verify_ok_to_terminate() const { 6449 assert(Thread::current()->is_ConcurrentGC_thread(), 6450 "should be called by CMS thread"); 6451 assert(!_foregroundGCShouldWait, "should be false"); 6452 // We could check here that all the various low-level locks 6453 // are not held by the CMS thread, but that is overkill; see 6454 // also CMSThread::verify_ok_to_terminate() where the CGC_lock 6455 // is checked. 6456 } 6457 #endif 6458 6459 size_t CMSCollector::block_size_using_printezis_bits(HeapWord* addr) const { 6460 assert(_markBitMap.isMarked(addr) && _markBitMap.isMarked(addr + 1), 6461 "missing Printezis mark?"); 6462 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 6463 size_t size = pointer_delta(nextOneAddr + 1, addr); 6464 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 6465 "alignment problem"); 6466 assert(size >= 3, "Necessary for Printezis marks to work"); 6467 return size; 6468 } 6469 6470 // A variant of the above (block_size_using_printezis_bits()) except 6471 // that we return 0 if the P-bits are not yet set. 6472 size_t CMSCollector::block_size_if_printezis_bits(HeapWord* addr) const { 6473 if (_markBitMap.isMarked(addr + 1)) { 6474 assert(_markBitMap.isMarked(addr), "P-bit can be set only for marked objects"); 6475 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 6476 size_t size = pointer_delta(nextOneAddr + 1, addr); 6477 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 6478 "alignment problem"); 6479 assert(size >= 3, "Necessary for Printezis marks to work"); 6480 return size; 6481 } 6482 return 0; 6483 } 6484 6485 HeapWord* CMSCollector::next_card_start_after_block(HeapWord* addr) const { 6486 size_t sz = 0; 6487 oop p = (oop)addr; 6488 if (p->klass_or_null() != NULL && p->is_parsable()) { 6489 sz = CompactibleFreeListSpace::adjustObjectSize(p->size()); 6490 } else { 6491 sz = block_size_using_printezis_bits(addr); 6492 } 6493 assert(sz > 0, "size must be nonzero"); 6494 HeapWord* next_block = addr + sz; 6495 HeapWord* next_card = (HeapWord*)round_to((uintptr_t)next_block, 6496 CardTableModRefBS::card_size); 6497 assert(round_down((uintptr_t)addr, CardTableModRefBS::card_size) < 6498 round_down((uintptr_t)next_card, CardTableModRefBS::card_size), 6499 "must be different cards"); 6500 return next_card; 6501 } 6502 6503 6504 // CMS Bit Map Wrapper ///////////////////////////////////////// 6505 6506 // Construct a CMS bit map infrastructure, but don't create the 6507 // bit vector itself. That is done by a separate call CMSBitMap::allocate() 6508 // further below. 6509 CMSBitMap::CMSBitMap(int shifter, int mutex_rank, const char* mutex_name): 6510 _bm(), 6511 _shifter(shifter), 6512 _lock(mutex_rank >= 0 ? new Mutex(mutex_rank, mutex_name, true) : NULL) 6513 { 6514 _bmStartWord = 0; 6515 _bmWordSize = 0; 6516 } 6517 6518 bool CMSBitMap::allocate(MemRegion mr) { 6519 _bmStartWord = mr.start(); 6520 _bmWordSize = mr.word_size(); 6521 ReservedSpace brs(ReservedSpace::allocation_align_size_up( 6522 (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1)); 6523 if (!brs.is_reserved()) { 6524 warning("CMS bit map allocation failure"); 6525 return false; 6526 } 6527 // For now we'll just commit all of the bit map up fromt. 6528 // Later on we'll try to be more parsimonious with swap. 6529 if (!_virtual_space.initialize(brs, brs.size())) { 6530 warning("CMS bit map backing store failure"); 6531 return false; 6532 } 6533 assert(_virtual_space.committed_size() == brs.size(), 6534 "didn't reserve backing store for all of CMS bit map?"); 6535 _bm.set_map((BitMap::bm_word_t*)_virtual_space.low()); 6536 assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >= 6537 _bmWordSize, "inconsistency in bit map sizing"); 6538 _bm.set_size(_bmWordSize >> _shifter); 6539 6540 // bm.clear(); // can we rely on getting zero'd memory? verify below 6541 assert(isAllClear(), 6542 "Expected zero'd memory from ReservedSpace constructor"); 6543 assert(_bm.size() == heapWordDiffToOffsetDiff(sizeInWords()), 6544 "consistency check"); 6545 return true; 6546 } 6547 6548 void CMSBitMap::dirty_range_iterate_clear(MemRegion mr, MemRegionClosure* cl) { 6549 HeapWord *next_addr, *end_addr, *last_addr; 6550 assert_locked(); 6551 assert(covers(mr), "out-of-range error"); 6552 // XXX assert that start and end are appropriately aligned 6553 for (next_addr = mr.start(), end_addr = mr.end(); 6554 next_addr < end_addr; next_addr = last_addr) { 6555 MemRegion dirty_region = getAndClearMarkedRegion(next_addr, end_addr); 6556 last_addr = dirty_region.end(); 6557 if (!dirty_region.is_empty()) { 6558 cl->do_MemRegion(dirty_region); 6559 } else { 6560 assert(last_addr == end_addr, "program logic"); 6561 return; 6562 } 6563 } 6564 } 6565 6566 #ifndef PRODUCT 6567 void CMSBitMap::assert_locked() const { 6568 CMSLockVerifier::assert_locked(lock()); 6569 } 6570 6571 bool CMSBitMap::covers(MemRegion mr) const { 6572 // assert(_bm.map() == _virtual_space.low(), "map inconsistency"); 6573 assert((size_t)_bm.size() == (_bmWordSize >> _shifter), 6574 "size inconsistency"); 6575 return (mr.start() >= _bmStartWord) && 6576 (mr.end() <= endWord()); 6577 } 6578 6579 bool CMSBitMap::covers(HeapWord* start, size_t size) const { 6580 return (start >= _bmStartWord && (start + size) <= endWord()); 6581 } 6582 6583 void CMSBitMap::verifyNoOneBitsInRange(HeapWord* left, HeapWord* right) { 6584 // verify that there are no 1 bits in the interval [left, right) 6585 FalseBitMapClosure falseBitMapClosure; 6586 iterate(&falseBitMapClosure, left, right); 6587 } 6588 6589 void CMSBitMap::region_invariant(MemRegion mr) 6590 { 6591 assert_locked(); 6592 // mr = mr.intersection(MemRegion(_bmStartWord, _bmWordSize)); 6593 assert(!mr.is_empty(), "unexpected empty region"); 6594 assert(covers(mr), "mr should be covered by bit map"); 6595 // convert address range into offset range 6596 size_t start_ofs = heapWordToOffset(mr.start()); 6597 // Make sure that end() is appropriately aligned 6598 assert(mr.end() == (HeapWord*)round_to((intptr_t)mr.end(), 6599 (1 << (_shifter+LogHeapWordSize))), 6600 "Misaligned mr.end()"); 6601 size_t end_ofs = heapWordToOffset(mr.end()); 6602 assert(end_ofs > start_ofs, "Should mark at least one bit"); 6603 } 6604 6605 #endif 6606 6607 bool CMSMarkStack::allocate(size_t size) { 6608 // allocate a stack of the requisite depth 6609 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 6610 size * sizeof(oop))); 6611 if (!rs.is_reserved()) { 6612 warning("CMSMarkStack allocation failure"); 6613 return false; 6614 } 6615 if (!_virtual_space.initialize(rs, rs.size())) { 6616 warning("CMSMarkStack backing store failure"); 6617 return false; 6618 } 6619 assert(_virtual_space.committed_size() == rs.size(), 6620 "didn't reserve backing store for all of CMS stack?"); 6621 _base = (oop*)(_virtual_space.low()); 6622 _index = 0; 6623 _capacity = size; 6624 NOT_PRODUCT(_max_depth = 0); 6625 return true; 6626 } 6627 6628 // XXX FIX ME !!! In the MT case we come in here holding a 6629 // leaf lock. For printing we need to take a further lock 6630 // which has lower rank. We need to recallibrate the two 6631 // lock-ranks involved in order to be able to rpint the 6632 // messages below. (Or defer the printing to the caller. 6633 // For now we take the expedient path of just disabling the 6634 // messages for the problematic case.) 6635 void CMSMarkStack::expand() { 6636 assert(_capacity <= MarkStackSizeMax, "stack bigger than permitted"); 6637 if (_capacity == MarkStackSizeMax) { 6638 if (_hit_limit++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) { 6639 // We print a warning message only once per CMS cycle. 6640 gclog_or_tty->print_cr(" (benign) Hit CMSMarkStack max size limit"); 6641 } 6642 return; 6643 } 6644 // Double capacity if possible 6645 size_t new_capacity = MIN2(_capacity*2, MarkStackSizeMax); 6646 // Do not give up existing stack until we have managed to 6647 // get the double capacity that we desired. 6648 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 6649 new_capacity * sizeof(oop))); 6650 if (rs.is_reserved()) { 6651 // Release the backing store associated with old stack 6652 _virtual_space.release(); 6653 // Reinitialize virtual space for new stack 6654 if (!_virtual_space.initialize(rs, rs.size())) { 6655 fatal("Not enough swap for expanded marking stack"); 6656 } 6657 _base = (oop*)(_virtual_space.low()); 6658 _index = 0; 6659 _capacity = new_capacity; 6660 } else if (_failed_double++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) { 6661 // Failed to double capacity, continue; 6662 // we print a detail message only once per CMS cycle. 6663 gclog_or_tty->print(" (benign) Failed to expand marking stack from "SIZE_FORMAT"K to " 6664 SIZE_FORMAT"K", 6665 _capacity / K, new_capacity / K); 6666 } 6667 } 6668 6669 6670 // Closures 6671 // XXX: there seems to be a lot of code duplication here; 6672 // should refactor and consolidate common code. 6673 6674 // This closure is used to mark refs into the CMS generation in 6675 // the CMS bit map. Called at the first checkpoint. This closure 6676 // assumes that we do not need to re-mark dirty cards; if the CMS 6677 // generation on which this is used is not an oldest (modulo perm gen) 6678 // generation then this will lose younger_gen cards! 6679 6680 MarkRefsIntoClosure::MarkRefsIntoClosure( 6681 MemRegion span, CMSBitMap* bitMap): 6682 _span(span), 6683 _bitMap(bitMap) 6684 { 6685 assert(_ref_processor == NULL, "deliberately left NULL"); 6686 assert(_bitMap->covers(_span), "_bitMap/_span mismatch"); 6687 } 6688 6689 void MarkRefsIntoClosure::do_oop(oop obj) { 6690 // if p points into _span, then mark corresponding bit in _markBitMap 6691 assert(obj->is_oop(), "expected an oop"); 6692 HeapWord* addr = (HeapWord*)obj; 6693 if (_span.contains(addr)) { 6694 // this should be made more efficient 6695 _bitMap->mark(addr); 6696 } 6697 } 6698 6699 void MarkRefsIntoClosure::do_oop(oop* p) { MarkRefsIntoClosure::do_oop_work(p); } 6700 void MarkRefsIntoClosure::do_oop(narrowOop* p) { MarkRefsIntoClosure::do_oop_work(p); } 6701 6702 // A variant of the above, used for CMS marking verification. 6703 MarkRefsIntoVerifyClosure::MarkRefsIntoVerifyClosure( 6704 MemRegion span, CMSBitMap* verification_bm, CMSBitMap* cms_bm): 6705 _span(span), 6706 _verification_bm(verification_bm), 6707 _cms_bm(cms_bm) 6708 { 6709 assert(_ref_processor == NULL, "deliberately left NULL"); 6710 assert(_verification_bm->covers(_span), "_verification_bm/_span mismatch"); 6711 } 6712 6713 void MarkRefsIntoVerifyClosure::do_oop(oop obj) { 6714 // if p points into _span, then mark corresponding bit in _markBitMap 6715 assert(obj->is_oop(), "expected an oop"); 6716 HeapWord* addr = (HeapWord*)obj; 6717 if (_span.contains(addr)) { 6718 _verification_bm->mark(addr); 6719 if (!_cms_bm->isMarked(addr)) { 6720 oop(addr)->print(); 6721 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)", addr); 6722 fatal("... aborting"); 6723 } 6724 } 6725 } 6726 6727 void MarkRefsIntoVerifyClosure::do_oop(oop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); } 6728 void MarkRefsIntoVerifyClosure::do_oop(narrowOop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); } 6729 6730 ////////////////////////////////////////////////// 6731 // MarkRefsIntoAndScanClosure 6732 ////////////////////////////////////////////////// 6733 6734 MarkRefsIntoAndScanClosure::MarkRefsIntoAndScanClosure(MemRegion span, 6735 ReferenceProcessor* rp, 6736 CMSBitMap* bit_map, 6737 CMSBitMap* mod_union_table, 6738 CMSMarkStack* mark_stack, 6739 CMSMarkStack* revisit_stack, 6740 CMSCollector* collector, 6741 bool should_yield, 6742 bool concurrent_precleaning): 6743 _collector(collector), 6744 _span(span), 6745 _bit_map(bit_map), 6746 _mark_stack(mark_stack), 6747 _pushAndMarkClosure(collector, span, rp, bit_map, mod_union_table, 6748 mark_stack, revisit_stack, concurrent_precleaning), 6749 _yield(should_yield), 6750 _concurrent_precleaning(concurrent_precleaning), 6751 _freelistLock(NULL) 6752 { 6753 _ref_processor = rp; 6754 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 6755 } 6756 6757 // This closure is used to mark refs into the CMS generation at the 6758 // second (final) checkpoint, and to scan and transitively follow 6759 // the unmarked oops. It is also used during the concurrent precleaning 6760 // phase while scanning objects on dirty cards in the CMS generation. 6761 // The marks are made in the marking bit map and the marking stack is 6762 // used for keeping the (newly) grey objects during the scan. 6763 // The parallel version (Par_...) appears further below. 6764 void MarkRefsIntoAndScanClosure::do_oop(oop obj) { 6765 if (obj != NULL) { 6766 assert(obj->is_oop(), "expected an oop"); 6767 HeapWord* addr = (HeapWord*)obj; 6768 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 6769 assert(_collector->overflow_list_is_empty(), 6770 "overflow list should be empty"); 6771 if (_span.contains(addr) && 6772 !_bit_map->isMarked(addr)) { 6773 // mark bit map (object is now grey) 6774 _bit_map->mark(addr); 6775 // push on marking stack (stack should be empty), and drain the 6776 // stack by applying this closure to the oops in the oops popped 6777 // from the stack (i.e. blacken the grey objects) 6778 bool res = _mark_stack->push(obj); 6779 assert(res, "Should have space to push on empty stack"); 6780 do { 6781 oop new_oop = _mark_stack->pop(); 6782 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 6783 assert(new_oop->is_parsable(), "Found unparsable oop"); 6784 assert(_bit_map->isMarked((HeapWord*)new_oop), 6785 "only grey objects on this stack"); 6786 // iterate over the oops in this oop, marking and pushing 6787 // the ones in CMS heap (i.e. in _span). 6788 new_oop->oop_iterate(&_pushAndMarkClosure); 6789 // check if it's time to yield 6790 do_yield_check(); 6791 } while (!_mark_stack->isEmpty() || 6792 (!_concurrent_precleaning && take_from_overflow_list())); 6793 // if marking stack is empty, and we are not doing this 6794 // during precleaning, then check the overflow list 6795 } 6796 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 6797 assert(_collector->overflow_list_is_empty(), 6798 "overflow list was drained above"); 6799 // We could restore evacuated mark words, if any, used for 6800 // overflow list links here because the overflow list is 6801 // provably empty here. That would reduce the maximum 6802 // size requirements for preserved_{oop,mark}_stack. 6803 // But we'll just postpone it until we are all done 6804 // so we can just stream through. 6805 if (!_concurrent_precleaning && CMSOverflowEarlyRestoration) { 6806 _collector->restore_preserved_marks_if_any(); 6807 assert(_collector->no_preserved_marks(), "No preserved marks"); 6808 } 6809 assert(!CMSOverflowEarlyRestoration || _collector->no_preserved_marks(), 6810 "All preserved marks should have been restored above"); 6811 } 6812 } 6813 6814 void MarkRefsIntoAndScanClosure::do_oop(oop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); } 6815 void MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); } 6816 6817 void MarkRefsIntoAndScanClosure::do_yield_work() { 6818 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6819 "CMS thread should hold CMS token"); 6820 assert_lock_strong(_freelistLock); 6821 assert_lock_strong(_bit_map->lock()); 6822 // relinquish the free_list_lock and bitMaplock() 6823 DEBUG_ONLY(RememberKlassesChecker mux(false);) 6824 _bit_map->lock()->unlock(); 6825 _freelistLock->unlock(); 6826 ConcurrentMarkSweepThread::desynchronize(true); 6827 ConcurrentMarkSweepThread::acknowledge_yield_request(); 6828 _collector->stopTimer(); 6829 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 6830 if (PrintCMSStatistics != 0) { 6831 _collector->incrementYields(); 6832 } 6833 _collector->icms_wait(); 6834 6835 // See the comment in coordinator_yield() 6836 for (unsigned i = 0; 6837 i < CMSYieldSleepCount && 6838 ConcurrentMarkSweepThread::should_yield() && 6839 !CMSCollector::foregroundGCIsActive(); 6840 ++i) { 6841 os::sleep(Thread::current(), 1, false); 6842 ConcurrentMarkSweepThread::acknowledge_yield_request(); 6843 } 6844 6845 ConcurrentMarkSweepThread::synchronize(true); 6846 _freelistLock->lock_without_safepoint_check(); 6847 _bit_map->lock()->lock_without_safepoint_check(); 6848 _collector->startTimer(); 6849 } 6850 6851 /////////////////////////////////////////////////////////// 6852 // Par_MarkRefsIntoAndScanClosure: a parallel version of 6853 // MarkRefsIntoAndScanClosure 6854 /////////////////////////////////////////////////////////// 6855 Par_MarkRefsIntoAndScanClosure::Par_MarkRefsIntoAndScanClosure( 6856 CMSCollector* collector, MemRegion span, ReferenceProcessor* rp, 6857 CMSBitMap* bit_map, OopTaskQueue* work_queue, CMSMarkStack* revisit_stack): 6858 _span(span), 6859 _bit_map(bit_map), 6860 _work_queue(work_queue), 6861 _low_water_mark(MIN2((uint)(work_queue->max_elems()/4), 6862 (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads))), 6863 _par_pushAndMarkClosure(collector, span, rp, bit_map, work_queue, 6864 revisit_stack) 6865 { 6866 _ref_processor = rp; 6867 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 6868 } 6869 6870 // This closure is used to mark refs into the CMS generation at the 6871 // second (final) checkpoint, and to scan and transitively follow 6872 // the unmarked oops. The marks are made in the marking bit map and 6873 // the work_queue is used for keeping the (newly) grey objects during 6874 // the scan phase whence they are also available for stealing by parallel 6875 // threads. Since the marking bit map is shared, updates are 6876 // synchronized (via CAS). 6877 void Par_MarkRefsIntoAndScanClosure::do_oop(oop obj) { 6878 if (obj != NULL) { 6879 // Ignore mark word because this could be an already marked oop 6880 // that may be chained at the end of the overflow list. 6881 assert(obj->is_oop(true), "expected an oop"); 6882 HeapWord* addr = (HeapWord*)obj; 6883 if (_span.contains(addr) && 6884 !_bit_map->isMarked(addr)) { 6885 // mark bit map (object will become grey): 6886 // It is possible for several threads to be 6887 // trying to "claim" this object concurrently; 6888 // the unique thread that succeeds in marking the 6889 // object first will do the subsequent push on 6890 // to the work queue (or overflow list). 6891 if (_bit_map->par_mark(addr)) { 6892 // push on work_queue (which may not be empty), and trim the 6893 // queue to an appropriate length by applying this closure to 6894 // the oops in the oops popped from the stack (i.e. blacken the 6895 // grey objects) 6896 bool res = _work_queue->push(obj); 6897 assert(res, "Low water mark should be less than capacity?"); 6898 trim_queue(_low_water_mark); 6899 } // Else, another thread claimed the object 6900 } 6901 } 6902 } 6903 6904 void Par_MarkRefsIntoAndScanClosure::do_oop(oop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); } 6905 void Par_MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); } 6906 6907 // This closure is used to rescan the marked objects on the dirty cards 6908 // in the mod union table and the card table proper. 6909 size_t ScanMarkedObjectsAgainCarefullyClosure::do_object_careful_m( 6910 oop p, MemRegion mr) { 6911 6912 size_t size = 0; 6913 HeapWord* addr = (HeapWord*)p; 6914 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6915 assert(_span.contains(addr), "we are scanning the CMS generation"); 6916 // check if it's time to yield 6917 if (do_yield_check()) { 6918 // We yielded for some foreground stop-world work, 6919 // and we have been asked to abort this ongoing preclean cycle. 6920 return 0; 6921 } 6922 if (_bitMap->isMarked(addr)) { 6923 // it's marked; is it potentially uninitialized? 6924 if (p->klass_or_null() != NULL) { 6925 // If is_conc_safe is false, the object may be undergoing 6926 // change by the VM outside a safepoint. Don't try to 6927 // scan it, but rather leave it for the remark phase. 6928 if (CMSPermGenPrecleaningEnabled && 6929 (!p->is_conc_safe() || !p->is_parsable())) { 6930 // Signal precleaning to redirty the card since 6931 // the klass pointer is already installed. 6932 assert(size == 0, "Initial value"); 6933 } else { 6934 assert(p->is_parsable(), "must be parsable."); 6935 // an initialized object; ignore mark word in verification below 6936 // since we are running concurrent with mutators 6937 assert(p->is_oop(true), "should be an oop"); 6938 if (p->is_objArray()) { 6939 // objArrays are precisely marked; restrict scanning 6940 // to dirty cards only. 6941 size = CompactibleFreeListSpace::adjustObjectSize( 6942 p->oop_iterate(_scanningClosure, mr)); 6943 } else { 6944 // A non-array may have been imprecisely marked; we need 6945 // to scan object in its entirety. 6946 size = CompactibleFreeListSpace::adjustObjectSize( 6947 p->oop_iterate(_scanningClosure)); 6948 } 6949 #ifdef DEBUG 6950 size_t direct_size = 6951 CompactibleFreeListSpace::adjustObjectSize(p->size()); 6952 assert(size == direct_size, "Inconsistency in size"); 6953 assert(size >= 3, "Necessary for Printezis marks to work"); 6954 if (!_bitMap->isMarked(addr+1)) { 6955 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size); 6956 } else { 6957 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size-1); 6958 assert(_bitMap->isMarked(addr+size-1), 6959 "inconsistent Printezis mark"); 6960 } 6961 #endif // DEBUG 6962 } 6963 } else { 6964 // an unitialized object 6965 assert(_bitMap->isMarked(addr+1), "missing Printezis mark?"); 6966 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 6967 size = pointer_delta(nextOneAddr + 1, addr); 6968 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 6969 "alignment problem"); 6970 // Note that pre-cleaning needn't redirty the card. OopDesc::set_klass() 6971 // will dirty the card when the klass pointer is installed in the 6972 // object (signalling the completion of initialization). 6973 } 6974 } else { 6975 // Either a not yet marked object or an uninitialized object 6976 if (p->klass_or_null() == NULL || !p->is_parsable()) { 6977 // An uninitialized object, skip to the next card, since 6978 // we may not be able to read its P-bits yet. 6979 assert(size == 0, "Initial value"); 6980 } else { 6981 // An object not (yet) reached by marking: we merely need to 6982 // compute its size so as to go look at the next block. 6983 assert(p->is_oop(true), "should be an oop"); 6984 size = CompactibleFreeListSpace::adjustObjectSize(p->size()); 6985 } 6986 } 6987 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6988 return size; 6989 } 6990 6991 void ScanMarkedObjectsAgainCarefullyClosure::do_yield_work() { 6992 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6993 "CMS thread should hold CMS token"); 6994 assert_lock_strong(_freelistLock); 6995 assert_lock_strong(_bitMap->lock()); 6996 DEBUG_ONLY(RememberKlassesChecker mux(false);) 6997 // relinquish the free_list_lock and bitMaplock() 6998 _bitMap->lock()->unlock(); 6999 _freelistLock->unlock(); 7000 ConcurrentMarkSweepThread::desynchronize(true); 7001 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7002 _collector->stopTimer(); 7003 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 7004 if (PrintCMSStatistics != 0) { 7005 _collector->incrementYields(); 7006 } 7007 _collector->icms_wait(); 7008 7009 // See the comment in coordinator_yield() 7010 for (unsigned i = 0; i < CMSYieldSleepCount && 7011 ConcurrentMarkSweepThread::should_yield() && 7012 !CMSCollector::foregroundGCIsActive(); ++i) { 7013 os::sleep(Thread::current(), 1, false); 7014 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7015 } 7016 7017 ConcurrentMarkSweepThread::synchronize(true); 7018 _freelistLock->lock_without_safepoint_check(); 7019 _bitMap->lock()->lock_without_safepoint_check(); 7020 _collector->startTimer(); 7021 } 7022 7023 7024 ////////////////////////////////////////////////////////////////// 7025 // SurvivorSpacePrecleanClosure 7026 ////////////////////////////////////////////////////////////////// 7027 // This (single-threaded) closure is used to preclean the oops in 7028 // the survivor spaces. 7029 size_t SurvivorSpacePrecleanClosure::do_object_careful(oop p) { 7030 7031 HeapWord* addr = (HeapWord*)p; 7032 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 7033 assert(!_span.contains(addr), "we are scanning the survivor spaces"); 7034 assert(p->klass_or_null() != NULL, "object should be initializd"); 7035 assert(p->is_parsable(), "must be parsable."); 7036 // an initialized object; ignore mark word in verification below 7037 // since we are running concurrent with mutators 7038 assert(p->is_oop(true), "should be an oop"); 7039 // Note that we do not yield while we iterate over 7040 // the interior oops of p, pushing the relevant ones 7041 // on our marking stack. 7042 size_t size = p->oop_iterate(_scanning_closure); 7043 do_yield_check(); 7044 // Observe that below, we do not abandon the preclean 7045 // phase as soon as we should; rather we empty the 7046 // marking stack before returning. This is to satisfy 7047 // some existing assertions. In general, it may be a 7048 // good idea to abort immediately and complete the marking 7049 // from the grey objects at a later time. 7050 while (!_mark_stack->isEmpty()) { 7051 oop new_oop = _mark_stack->pop(); 7052 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 7053 assert(new_oop->is_parsable(), "Found unparsable oop"); 7054 assert(_bit_map->isMarked((HeapWord*)new_oop), 7055 "only grey objects on this stack"); 7056 // iterate over the oops in this oop, marking and pushing 7057 // the ones in CMS heap (i.e. in _span). 7058 new_oop->oop_iterate(_scanning_closure); 7059 // check if it's time to yield 7060 do_yield_check(); 7061 } 7062 unsigned int after_count = 7063 GenCollectedHeap::heap()->total_collections(); 7064 bool abort = (_before_count != after_count) || 7065 _collector->should_abort_preclean(); 7066 return abort ? 0 : size; 7067 } 7068 7069 void SurvivorSpacePrecleanClosure::do_yield_work() { 7070 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 7071 "CMS thread should hold CMS token"); 7072 assert_lock_strong(_bit_map->lock()); 7073 DEBUG_ONLY(RememberKlassesChecker smx(false);) 7074 // Relinquish the bit map lock 7075 _bit_map->lock()->unlock(); 7076 ConcurrentMarkSweepThread::desynchronize(true); 7077 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7078 _collector->stopTimer(); 7079 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 7080 if (PrintCMSStatistics != 0) { 7081 _collector->incrementYields(); 7082 } 7083 _collector->icms_wait(); 7084 7085 // See the comment in coordinator_yield() 7086 for (unsigned i = 0; i < CMSYieldSleepCount && 7087 ConcurrentMarkSweepThread::should_yield() && 7088 !CMSCollector::foregroundGCIsActive(); ++i) { 7089 os::sleep(Thread::current(), 1, false); 7090 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7091 } 7092 7093 ConcurrentMarkSweepThread::synchronize(true); 7094 _bit_map->lock()->lock_without_safepoint_check(); 7095 _collector->startTimer(); 7096 } 7097 7098 // This closure is used to rescan the marked objects on the dirty cards 7099 // in the mod union table and the card table proper. In the parallel 7100 // case, although the bitMap is shared, we do a single read so the 7101 // isMarked() query is "safe". 7102 bool ScanMarkedObjectsAgainClosure::do_object_bm(oop p, MemRegion mr) { 7103 // Ignore mark word because we are running concurrent with mutators 7104 assert(p->is_oop_or_null(true), "expected an oop or null"); 7105 HeapWord* addr = (HeapWord*)p; 7106 assert(_span.contains(addr), "we are scanning the CMS generation"); 7107 bool is_obj_array = false; 7108 #ifdef DEBUG 7109 if (!_parallel) { 7110 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 7111 assert(_collector->overflow_list_is_empty(), 7112 "overflow list should be empty"); 7113 7114 } 7115 #endif // DEBUG 7116 if (_bit_map->isMarked(addr)) { 7117 // Obj arrays are precisely marked, non-arrays are not; 7118 // so we scan objArrays precisely and non-arrays in their 7119 // entirety. 7120 if (p->is_objArray()) { 7121 is_obj_array = true; 7122 if (_parallel) { 7123 p->oop_iterate(_par_scan_closure, mr); 7124 } else { 7125 p->oop_iterate(_scan_closure, mr); 7126 } 7127 } else { 7128 if (_parallel) { 7129 p->oop_iterate(_par_scan_closure); 7130 } else { 7131 p->oop_iterate(_scan_closure); 7132 } 7133 } 7134 } 7135 #ifdef DEBUG 7136 if (!_parallel) { 7137 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 7138 assert(_collector->overflow_list_is_empty(), 7139 "overflow list should be empty"); 7140 7141 } 7142 #endif // DEBUG 7143 return is_obj_array; 7144 } 7145 7146 MarkFromRootsClosure::MarkFromRootsClosure(CMSCollector* collector, 7147 MemRegion span, 7148 CMSBitMap* bitMap, CMSMarkStack* markStack, 7149 CMSMarkStack* revisitStack, 7150 bool should_yield, bool verifying): 7151 _collector(collector), 7152 _span(span), 7153 _bitMap(bitMap), 7154 _mut(&collector->_modUnionTable), 7155 _markStack(markStack), 7156 _revisitStack(revisitStack), 7157 _yield(should_yield), 7158 _skipBits(0) 7159 { 7160 assert(_markStack->isEmpty(), "stack should be empty"); 7161 _finger = _bitMap->startWord(); 7162 _threshold = _finger; 7163 assert(_collector->_restart_addr == NULL, "Sanity check"); 7164 assert(_span.contains(_finger), "Out of bounds _finger?"); 7165 DEBUG_ONLY(_verifying = verifying;) 7166 } 7167 7168 void MarkFromRootsClosure::reset(HeapWord* addr) { 7169 assert(_markStack->isEmpty(), "would cause duplicates on stack"); 7170 assert(_span.contains(addr), "Out of bounds _finger?"); 7171 _finger = addr; 7172 _threshold = (HeapWord*)round_to( 7173 (intptr_t)_finger, CardTableModRefBS::card_size); 7174 } 7175 7176 // Should revisit to see if this should be restructured for 7177 // greater efficiency. 7178 bool MarkFromRootsClosure::do_bit(size_t offset) { 7179 if (_skipBits > 0) { 7180 _skipBits--; 7181 return true; 7182 } 7183 // convert offset into a HeapWord* 7184 HeapWord* addr = _bitMap->startWord() + offset; 7185 assert(_bitMap->endWord() && addr < _bitMap->endWord(), 7186 "address out of range"); 7187 assert(_bitMap->isMarked(addr), "tautology"); 7188 if (_bitMap->isMarked(addr+1)) { 7189 // this is an allocated but not yet initialized object 7190 assert(_skipBits == 0, "tautology"); 7191 _skipBits = 2; // skip next two marked bits ("Printezis-marks") 7192 oop p = oop(addr); 7193 if (p->klass_or_null() == NULL || !p->is_parsable()) { 7194 DEBUG_ONLY(if (!_verifying) {) 7195 // We re-dirty the cards on which this object lies and increase 7196 // the _threshold so that we'll come back to scan this object 7197 // during the preclean or remark phase. (CMSCleanOnEnter) 7198 if (CMSCleanOnEnter) { 7199 size_t sz = _collector->block_size_using_printezis_bits(addr); 7200 HeapWord* end_card_addr = (HeapWord*)round_to( 7201 (intptr_t)(addr+sz), CardTableModRefBS::card_size); 7202 MemRegion redirty_range = MemRegion(addr, end_card_addr); 7203 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 7204 // Bump _threshold to end_card_addr; note that 7205 // _threshold cannot possibly exceed end_card_addr, anyhow. 7206 // This prevents future clearing of the card as the scan proceeds 7207 // to the right. 7208 assert(_threshold <= end_card_addr, 7209 "Because we are just scanning into this object"); 7210 if (_threshold < end_card_addr) { 7211 _threshold = end_card_addr; 7212 } 7213 if (p->klass_or_null() != NULL) { 7214 // Redirty the range of cards... 7215 _mut->mark_range(redirty_range); 7216 } // ...else the setting of klass will dirty the card anyway. 7217 } 7218 DEBUG_ONLY(}) 7219 return true; 7220 } 7221 } 7222 scanOopsInOop(addr); 7223 return true; 7224 } 7225 7226 // We take a break if we've been at this for a while, 7227 // so as to avoid monopolizing the locks involved. 7228 void MarkFromRootsClosure::do_yield_work() { 7229 // First give up the locks, then yield, then re-lock 7230 // We should probably use a constructor/destructor idiom to 7231 // do this unlock/lock or modify the MutexUnlocker class to 7232 // serve our purpose. XXX 7233 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 7234 "CMS thread should hold CMS token"); 7235 assert_lock_strong(_bitMap->lock()); 7236 DEBUG_ONLY(RememberKlassesChecker mux(false);) 7237 _bitMap->lock()->unlock(); 7238 ConcurrentMarkSweepThread::desynchronize(true); 7239 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7240 _collector->stopTimer(); 7241 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 7242 if (PrintCMSStatistics != 0) { 7243 _collector->incrementYields(); 7244 } 7245 _collector->icms_wait(); 7246 7247 // See the comment in coordinator_yield() 7248 for (unsigned i = 0; i < CMSYieldSleepCount && 7249 ConcurrentMarkSweepThread::should_yield() && 7250 !CMSCollector::foregroundGCIsActive(); ++i) { 7251 os::sleep(Thread::current(), 1, false); 7252 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7253 } 7254 7255 ConcurrentMarkSweepThread::synchronize(true); 7256 _bitMap->lock()->lock_without_safepoint_check(); 7257 _collector->startTimer(); 7258 } 7259 7260 void MarkFromRootsClosure::scanOopsInOop(HeapWord* ptr) { 7261 assert(_bitMap->isMarked(ptr), "expected bit to be set"); 7262 assert(_markStack->isEmpty(), 7263 "should drain stack to limit stack usage"); 7264 // convert ptr to an oop preparatory to scanning 7265 oop obj = oop(ptr); 7266 // Ignore mark word in verification below, since we 7267 // may be running concurrent with mutators. 7268 assert(obj->is_oop(true), "should be an oop"); 7269 assert(_finger <= ptr, "_finger runneth ahead"); 7270 // advance the finger to right end of this object 7271 _finger = ptr + obj->size(); 7272 assert(_finger > ptr, "we just incremented it above"); 7273 // On large heaps, it may take us some time to get through 7274 // the marking phase (especially if running iCMS). During 7275 // this time it's possible that a lot of mutations have 7276 // accumulated in the card table and the mod union table -- 7277 // these mutation records are redundant until we have 7278 // actually traced into the corresponding card. 7279 // Here, we check whether advancing the finger would make 7280 // us cross into a new card, and if so clear corresponding 7281 // cards in the MUT (preclean them in the card-table in the 7282 // future). 7283 7284 DEBUG_ONLY(if (!_verifying) {) 7285 // The clean-on-enter optimization is disabled by default, 7286 // until we fix 6178663. 7287 if (CMSCleanOnEnter && (_finger > _threshold)) { 7288 // [_threshold, _finger) represents the interval 7289 // of cards to be cleared in MUT (or precleaned in card table). 7290 // The set of cards to be cleared is all those that overlap 7291 // with the interval [_threshold, _finger); note that 7292 // _threshold is always kept card-aligned but _finger isn't 7293 // always card-aligned. 7294 HeapWord* old_threshold = _threshold; 7295 assert(old_threshold == (HeapWord*)round_to( 7296 (intptr_t)old_threshold, CardTableModRefBS::card_size), 7297 "_threshold should always be card-aligned"); 7298 _threshold = (HeapWord*)round_to( 7299 (intptr_t)_finger, CardTableModRefBS::card_size); 7300 MemRegion mr(old_threshold, _threshold); 7301 assert(!mr.is_empty(), "Control point invariant"); 7302 assert(_span.contains(mr), "Should clear within span"); 7303 // XXX When _finger crosses from old gen into perm gen 7304 // we may be doing unnecessary cleaning; do better in the 7305 // future by detecting that condition and clearing fewer 7306 // MUT/CT entries. 7307 _mut->clear_range(mr); 7308 } 7309 DEBUG_ONLY(}) 7310 // Note: the finger doesn't advance while we drain 7311 // the stack below. 7312 PushOrMarkClosure pushOrMarkClosure(_collector, 7313 _span, _bitMap, _markStack, 7314 _revisitStack, 7315 _finger, this); 7316 bool res = _markStack->push(obj); 7317 assert(res, "Empty non-zero size stack should have space for single push"); 7318 while (!_markStack->isEmpty()) { 7319 oop new_oop = _markStack->pop(); 7320 // Skip verifying header mark word below because we are 7321 // running concurrent with mutators. 7322 assert(new_oop->is_oop(true), "Oops! expected to pop an oop"); 7323 // now scan this oop's oops 7324 new_oop->oop_iterate(&pushOrMarkClosure); 7325 do_yield_check(); 7326 } 7327 assert(_markStack->isEmpty(), "tautology, emphasizing post-condition"); 7328 } 7329 7330 Par_MarkFromRootsClosure::Par_MarkFromRootsClosure(CMSConcMarkingTask* task, 7331 CMSCollector* collector, MemRegion span, 7332 CMSBitMap* bit_map, 7333 OopTaskQueue* work_queue, 7334 CMSMarkStack* overflow_stack, 7335 CMSMarkStack* revisit_stack, 7336 bool should_yield): 7337 _collector(collector), 7338 _whole_span(collector->_span), 7339 _span(span), 7340 _bit_map(bit_map), 7341 _mut(&collector->_modUnionTable), 7342 _work_queue(work_queue), 7343 _overflow_stack(overflow_stack), 7344 _revisit_stack(revisit_stack), 7345 _yield(should_yield), 7346 _skip_bits(0), 7347 _task(task) 7348 { 7349 assert(_work_queue->size() == 0, "work_queue should be empty"); 7350 _finger = span.start(); 7351 _threshold = _finger; // XXX Defer clear-on-enter optimization for now 7352 assert(_span.contains(_finger), "Out of bounds _finger?"); 7353 } 7354 7355 // Should revisit to see if this should be restructured for 7356 // greater efficiency. 7357 bool Par_MarkFromRootsClosure::do_bit(size_t offset) { 7358 if (_skip_bits > 0) { 7359 _skip_bits--; 7360 return true; 7361 } 7362 // convert offset into a HeapWord* 7363 HeapWord* addr = _bit_map->startWord() + offset; 7364 assert(_bit_map->endWord() && addr < _bit_map->endWord(), 7365 "address out of range"); 7366 assert(_bit_map->isMarked(addr), "tautology"); 7367 if (_bit_map->isMarked(addr+1)) { 7368 // this is an allocated object that might not yet be initialized 7369 assert(_skip_bits == 0, "tautology"); 7370 _skip_bits = 2; // skip next two marked bits ("Printezis-marks") 7371 oop p = oop(addr); 7372 if (p->klass_or_null() == NULL || !p->is_parsable()) { 7373 // in the case of Clean-on-Enter optimization, redirty card 7374 // and avoid clearing card by increasing the threshold. 7375 return true; 7376 } 7377 } 7378 scan_oops_in_oop(addr); 7379 return true; 7380 } 7381 7382 void Par_MarkFromRootsClosure::scan_oops_in_oop(HeapWord* ptr) { 7383 assert(_bit_map->isMarked(ptr), "expected bit to be set"); 7384 // Should we assert that our work queue is empty or 7385 // below some drain limit? 7386 assert(_work_queue->size() == 0, 7387 "should drain stack to limit stack usage"); 7388 // convert ptr to an oop preparatory to scanning 7389 oop obj = oop(ptr); 7390 // Ignore mark word in verification below, since we 7391 // may be running concurrent with mutators. 7392 assert(obj->is_oop(true), "should be an oop"); 7393 assert(_finger <= ptr, "_finger runneth ahead"); 7394 // advance the finger to right end of this object 7395 _finger = ptr + obj->size(); 7396 assert(_finger > ptr, "we just incremented it above"); 7397 // On large heaps, it may take us some time to get through 7398 // the marking phase (especially if running iCMS). During 7399 // this time it's possible that a lot of mutations have 7400 // accumulated in the card table and the mod union table -- 7401 // these mutation records are redundant until we have 7402 // actually traced into the corresponding card. 7403 // Here, we check whether advancing the finger would make 7404 // us cross into a new card, and if so clear corresponding 7405 // cards in the MUT (preclean them in the card-table in the 7406 // future). 7407 7408 // The clean-on-enter optimization is disabled by default, 7409 // until we fix 6178663. 7410 if (CMSCleanOnEnter && (_finger > _threshold)) { 7411 // [_threshold, _finger) represents the interval 7412 // of cards to be cleared in MUT (or precleaned in card table). 7413 // The set of cards to be cleared is all those that overlap 7414 // with the interval [_threshold, _finger); note that 7415 // _threshold is always kept card-aligned but _finger isn't 7416 // always card-aligned. 7417 HeapWord* old_threshold = _threshold; 7418 assert(old_threshold == (HeapWord*)round_to( 7419 (intptr_t)old_threshold, CardTableModRefBS::card_size), 7420 "_threshold should always be card-aligned"); 7421 _threshold = (HeapWord*)round_to( 7422 (intptr_t)_finger, CardTableModRefBS::card_size); 7423 MemRegion mr(old_threshold, _threshold); 7424 assert(!mr.is_empty(), "Control point invariant"); 7425 assert(_span.contains(mr), "Should clear within span"); // _whole_span ?? 7426 // XXX When _finger crosses from old gen into perm gen 7427 // we may be doing unnecessary cleaning; do better in the 7428 // future by detecting that condition and clearing fewer 7429 // MUT/CT entries. 7430 _mut->clear_range(mr); 7431 } 7432 7433 // Note: the local finger doesn't advance while we drain 7434 // the stack below, but the global finger sure can and will. 7435 HeapWord** gfa = _task->global_finger_addr(); 7436 Par_PushOrMarkClosure pushOrMarkClosure(_collector, 7437 _span, _bit_map, 7438 _work_queue, 7439 _overflow_stack, 7440 _revisit_stack, 7441 _finger, 7442 gfa, this); 7443 bool res = _work_queue->push(obj); // overflow could occur here 7444 assert(res, "Will hold once we use workqueues"); 7445 while (true) { 7446 oop new_oop; 7447 if (!_work_queue->pop_local(new_oop)) { 7448 // We emptied our work_queue; check if there's stuff that can 7449 // be gotten from the overflow stack. 7450 if (CMSConcMarkingTask::get_work_from_overflow_stack( 7451 _overflow_stack, _work_queue)) { 7452 do_yield_check(); 7453 continue; 7454 } else { // done 7455 break; 7456 } 7457 } 7458 // Skip verifying header mark word below because we are 7459 // running concurrent with mutators. 7460 assert(new_oop->is_oop(true), "Oops! expected to pop an oop"); 7461 // now scan this oop's oops 7462 new_oop->oop_iterate(&pushOrMarkClosure); 7463 do_yield_check(); 7464 } 7465 assert(_work_queue->size() == 0, "tautology, emphasizing post-condition"); 7466 } 7467 7468 // Yield in response to a request from VM Thread or 7469 // from mutators. 7470 void Par_MarkFromRootsClosure::do_yield_work() { 7471 assert(_task != NULL, "sanity"); 7472 _task->yield(); 7473 } 7474 7475 // A variant of the above used for verifying CMS marking work. 7476 MarkFromRootsVerifyClosure::MarkFromRootsVerifyClosure(CMSCollector* collector, 7477 MemRegion span, 7478 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 7479 CMSMarkStack* mark_stack): 7480 _collector(collector), 7481 _span(span), 7482 _verification_bm(verification_bm), 7483 _cms_bm(cms_bm), 7484 _mark_stack(mark_stack), 7485 _pam_verify_closure(collector, span, verification_bm, cms_bm, 7486 mark_stack) 7487 { 7488 assert(_mark_stack->isEmpty(), "stack should be empty"); 7489 _finger = _verification_bm->startWord(); 7490 assert(_collector->_restart_addr == NULL, "Sanity check"); 7491 assert(_span.contains(_finger), "Out of bounds _finger?"); 7492 } 7493 7494 void MarkFromRootsVerifyClosure::reset(HeapWord* addr) { 7495 assert(_mark_stack->isEmpty(), "would cause duplicates on stack"); 7496 assert(_span.contains(addr), "Out of bounds _finger?"); 7497 _finger = addr; 7498 } 7499 7500 // Should revisit to see if this should be restructured for 7501 // greater efficiency. 7502 bool MarkFromRootsVerifyClosure::do_bit(size_t offset) { 7503 // convert offset into a HeapWord* 7504 HeapWord* addr = _verification_bm->startWord() + offset; 7505 assert(_verification_bm->endWord() && addr < _verification_bm->endWord(), 7506 "address out of range"); 7507 assert(_verification_bm->isMarked(addr), "tautology"); 7508 assert(_cms_bm->isMarked(addr), "tautology"); 7509 7510 assert(_mark_stack->isEmpty(), 7511 "should drain stack to limit stack usage"); 7512 // convert addr to an oop preparatory to scanning 7513 oop obj = oop(addr); 7514 assert(obj->is_oop(), "should be an oop"); 7515 assert(_finger <= addr, "_finger runneth ahead"); 7516 // advance the finger to right end of this object 7517 _finger = addr + obj->size(); 7518 assert(_finger > addr, "we just incremented it above"); 7519 // Note: the finger doesn't advance while we drain 7520 // the stack below. 7521 bool res = _mark_stack->push(obj); 7522 assert(res, "Empty non-zero size stack should have space for single push"); 7523 while (!_mark_stack->isEmpty()) { 7524 oop new_oop = _mark_stack->pop(); 7525 assert(new_oop->is_oop(), "Oops! expected to pop an oop"); 7526 // now scan this oop's oops 7527 new_oop->oop_iterate(&_pam_verify_closure); 7528 } 7529 assert(_mark_stack->isEmpty(), "tautology, emphasizing post-condition"); 7530 return true; 7531 } 7532 7533 PushAndMarkVerifyClosure::PushAndMarkVerifyClosure( 7534 CMSCollector* collector, MemRegion span, 7535 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 7536 CMSMarkStack* mark_stack): 7537 OopClosure(collector->ref_processor()), 7538 _collector(collector), 7539 _span(span), 7540 _verification_bm(verification_bm), 7541 _cms_bm(cms_bm), 7542 _mark_stack(mark_stack) 7543 { } 7544 7545 void PushAndMarkVerifyClosure::do_oop(oop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 7546 void PushAndMarkVerifyClosure::do_oop(narrowOop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 7547 7548 // Upon stack overflow, we discard (part of) the stack, 7549 // remembering the least address amongst those discarded 7550 // in CMSCollector's _restart_address. 7551 void PushAndMarkVerifyClosure::handle_stack_overflow(HeapWord* lost) { 7552 // Remember the least grey address discarded 7553 HeapWord* ra = (HeapWord*)_mark_stack->least_value(lost); 7554 _collector->lower_restart_addr(ra); 7555 _mark_stack->reset(); // discard stack contents 7556 _mark_stack->expand(); // expand the stack if possible 7557 } 7558 7559 void PushAndMarkVerifyClosure::do_oop(oop obj) { 7560 assert(obj->is_oop_or_null(), "expected an oop or NULL"); 7561 HeapWord* addr = (HeapWord*)obj; 7562 if (_span.contains(addr) && !_verification_bm->isMarked(addr)) { 7563 // Oop lies in _span and isn't yet grey or black 7564 _verification_bm->mark(addr); // now grey 7565 if (!_cms_bm->isMarked(addr)) { 7566 oop(addr)->print(); 7567 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)", 7568 addr); 7569 fatal("... aborting"); 7570 } 7571 7572 if (!_mark_stack->push(obj)) { // stack overflow 7573 if (PrintCMSStatistics != 0) { 7574 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 7575 SIZE_FORMAT, _mark_stack->capacity()); 7576 } 7577 assert(_mark_stack->isFull(), "Else push should have succeeded"); 7578 handle_stack_overflow(addr); 7579 } 7580 // anything including and to the right of _finger 7581 // will be scanned as we iterate over the remainder of the 7582 // bit map 7583 } 7584 } 7585 7586 PushOrMarkClosure::PushOrMarkClosure(CMSCollector* collector, 7587 MemRegion span, 7588 CMSBitMap* bitMap, CMSMarkStack* markStack, 7589 CMSMarkStack* revisitStack, 7590 HeapWord* finger, MarkFromRootsClosure* parent) : 7591 KlassRememberingOopClosure(collector, collector->ref_processor(), revisitStack), 7592 _span(span), 7593 _bitMap(bitMap), 7594 _markStack(markStack), 7595 _finger(finger), 7596 _parent(parent) 7597 { } 7598 7599 Par_PushOrMarkClosure::Par_PushOrMarkClosure(CMSCollector* collector, 7600 MemRegion span, 7601 CMSBitMap* bit_map, 7602 OopTaskQueue* work_queue, 7603 CMSMarkStack* overflow_stack, 7604 CMSMarkStack* revisit_stack, 7605 HeapWord* finger, 7606 HeapWord** global_finger_addr, 7607 Par_MarkFromRootsClosure* parent) : 7608 Par_KlassRememberingOopClosure(collector, 7609 collector->ref_processor(), 7610 revisit_stack), 7611 _whole_span(collector->_span), 7612 _span(span), 7613 _bit_map(bit_map), 7614 _work_queue(work_queue), 7615 _overflow_stack(overflow_stack), 7616 _finger(finger), 7617 _global_finger_addr(global_finger_addr), 7618 _parent(parent) 7619 { } 7620 7621 // Assumes thread-safe access by callers, who are 7622 // responsible for mutual exclusion. 7623 void CMSCollector::lower_restart_addr(HeapWord* low) { 7624 assert(_span.contains(low), "Out of bounds addr"); 7625 if (_restart_addr == NULL) { 7626 _restart_addr = low; 7627 } else { 7628 _restart_addr = MIN2(_restart_addr, low); 7629 } 7630 } 7631 7632 // Upon stack overflow, we discard (part of) the stack, 7633 // remembering the least address amongst those discarded 7634 // in CMSCollector's _restart_address. 7635 void PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 7636 // Remember the least grey address discarded 7637 HeapWord* ra = (HeapWord*)_markStack->least_value(lost); 7638 _collector->lower_restart_addr(ra); 7639 _markStack->reset(); // discard stack contents 7640 _markStack->expand(); // expand the stack if possible 7641 } 7642 7643 // Upon stack overflow, we discard (part of) the stack, 7644 // remembering the least address amongst those discarded 7645 // in CMSCollector's _restart_address. 7646 void Par_PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 7647 // We need to do this under a mutex to prevent other 7648 // workers from interfering with the work done below. 7649 MutexLockerEx ml(_overflow_stack->par_lock(), 7650 Mutex::_no_safepoint_check_flag); 7651 // Remember the least grey address discarded 7652 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 7653 _collector->lower_restart_addr(ra); 7654 _overflow_stack->reset(); // discard stack contents 7655 _overflow_stack->expand(); // expand the stack if possible 7656 } 7657 7658 void PushOrMarkClosure::do_oop(oop obj) { 7659 // Ignore mark word because we are running concurrent with mutators. 7660 assert(obj->is_oop_or_null(true), "expected an oop or NULL"); 7661 HeapWord* addr = (HeapWord*)obj; 7662 if (_span.contains(addr) && !_bitMap->isMarked(addr)) { 7663 // Oop lies in _span and isn't yet grey or black 7664 _bitMap->mark(addr); // now grey 7665 if (addr < _finger) { 7666 // the bit map iteration has already either passed, or 7667 // sampled, this bit in the bit map; we'll need to 7668 // use the marking stack to scan this oop's oops. 7669 bool simulate_overflow = false; 7670 NOT_PRODUCT( 7671 if (CMSMarkStackOverflowALot && 7672 _collector->simulate_overflow()) { 7673 // simulate a stack overflow 7674 simulate_overflow = true; 7675 } 7676 ) 7677 if (simulate_overflow || !_markStack->push(obj)) { // stack overflow 7678 if (PrintCMSStatistics != 0) { 7679 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 7680 SIZE_FORMAT, _markStack->capacity()); 7681 } 7682 assert(simulate_overflow || _markStack->isFull(), "Else push should have succeeded"); 7683 handle_stack_overflow(addr); 7684 } 7685 } 7686 // anything including and to the right of _finger 7687 // will be scanned as we iterate over the remainder of the 7688 // bit map 7689 do_yield_check(); 7690 } 7691 } 7692 7693 void PushOrMarkClosure::do_oop(oop* p) { PushOrMarkClosure::do_oop_work(p); } 7694 void PushOrMarkClosure::do_oop(narrowOop* p) { PushOrMarkClosure::do_oop_work(p); } 7695 7696 void Par_PushOrMarkClosure::do_oop(oop obj) { 7697 // Ignore mark word because we are running concurrent with mutators. 7698 assert(obj->is_oop_or_null(true), "expected an oop or NULL"); 7699 HeapWord* addr = (HeapWord*)obj; 7700 if (_whole_span.contains(addr) && !_bit_map->isMarked(addr)) { 7701 // Oop lies in _span and isn't yet grey or black 7702 // We read the global_finger (volatile read) strictly after marking oop 7703 bool res = _bit_map->par_mark(addr); // now grey 7704 volatile HeapWord** gfa = (volatile HeapWord**)_global_finger_addr; 7705 // Should we push this marked oop on our stack? 7706 // -- if someone else marked it, nothing to do 7707 // -- if target oop is above global finger nothing to do 7708 // -- if target oop is in chunk and above local finger 7709 // then nothing to do 7710 // -- else push on work queue 7711 if ( !res // someone else marked it, they will deal with it 7712 || (addr >= *gfa) // will be scanned in a later task 7713 || (_span.contains(addr) && addr >= _finger)) { // later in this chunk 7714 return; 7715 } 7716 // the bit map iteration has already either passed, or 7717 // sampled, this bit in the bit map; we'll need to 7718 // use the marking stack to scan this oop's oops. 7719 bool simulate_overflow = false; 7720 NOT_PRODUCT( 7721 if (CMSMarkStackOverflowALot && 7722 _collector->simulate_overflow()) { 7723 // simulate a stack overflow 7724 simulate_overflow = true; 7725 } 7726 ) 7727 if (simulate_overflow || 7728 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) { 7729 // stack overflow 7730 if (PrintCMSStatistics != 0) { 7731 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 7732 SIZE_FORMAT, _overflow_stack->capacity()); 7733 } 7734 // We cannot assert that the overflow stack is full because 7735 // it may have been emptied since. 7736 assert(simulate_overflow || 7737 _work_queue->size() == _work_queue->max_elems(), 7738 "Else push should have succeeded"); 7739 handle_stack_overflow(addr); 7740 } 7741 do_yield_check(); 7742 } 7743 } 7744 7745 void Par_PushOrMarkClosure::do_oop(oop* p) { Par_PushOrMarkClosure::do_oop_work(p); } 7746 void Par_PushOrMarkClosure::do_oop(narrowOop* p) { Par_PushOrMarkClosure::do_oop_work(p); } 7747 7748 KlassRememberingOopClosure::KlassRememberingOopClosure(CMSCollector* collector, 7749 ReferenceProcessor* rp, 7750 CMSMarkStack* revisit_stack) : 7751 OopClosure(rp), 7752 _collector(collector), 7753 _revisit_stack(revisit_stack), 7754 _should_remember_klasses(collector->should_unload_classes()) {} 7755 7756 PushAndMarkClosure::PushAndMarkClosure(CMSCollector* collector, 7757 MemRegion span, 7758 ReferenceProcessor* rp, 7759 CMSBitMap* bit_map, 7760 CMSBitMap* mod_union_table, 7761 CMSMarkStack* mark_stack, 7762 CMSMarkStack* revisit_stack, 7763 bool concurrent_precleaning): 7764 KlassRememberingOopClosure(collector, rp, revisit_stack), 7765 _span(span), 7766 _bit_map(bit_map), 7767 _mod_union_table(mod_union_table), 7768 _mark_stack(mark_stack), 7769 _concurrent_precleaning(concurrent_precleaning) 7770 { 7771 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 7772 } 7773 7774 // Grey object rescan during pre-cleaning and second checkpoint phases -- 7775 // the non-parallel version (the parallel version appears further below.) 7776 void PushAndMarkClosure::do_oop(oop obj) { 7777 // Ignore mark word verification. If during concurrent precleaning, 7778 // the object monitor may be locked. If during the checkpoint 7779 // phases, the object may already have been reached by a different 7780 // path and may be at the end of the global overflow list (so 7781 // the mark word may be NULL). 7782 assert(obj->is_oop_or_null(true /* ignore mark word */), 7783 "expected an oop or NULL"); 7784 HeapWord* addr = (HeapWord*)obj; 7785 // Check if oop points into the CMS generation 7786 // and is not marked 7787 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 7788 // a white object ... 7789 _bit_map->mark(addr); // ... now grey 7790 // push on the marking stack (grey set) 7791 bool simulate_overflow = false; 7792 NOT_PRODUCT( 7793 if (CMSMarkStackOverflowALot && 7794 _collector->simulate_overflow()) { 7795 // simulate a stack overflow 7796 simulate_overflow = true; 7797 } 7798 ) 7799 if (simulate_overflow || !_mark_stack->push(obj)) { 7800 if (_concurrent_precleaning) { 7801 // During precleaning we can just dirty the appropriate card(s) 7802 // in the mod union table, thus ensuring that the object remains 7803 // in the grey set and continue. In the case of object arrays 7804 // we need to dirty all of the cards that the object spans, 7805 // since the rescan of object arrays will be limited to the 7806 // dirty cards. 7807 // Note that no one can be intefering with us in this action 7808 // of dirtying the mod union table, so no locking or atomics 7809 // are required. 7810 if (obj->is_objArray()) { 7811 size_t sz = obj->size(); 7812 HeapWord* end_card_addr = (HeapWord*)round_to( 7813 (intptr_t)(addr+sz), CardTableModRefBS::card_size); 7814 MemRegion redirty_range = MemRegion(addr, end_card_addr); 7815 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 7816 _mod_union_table->mark_range(redirty_range); 7817 } else { 7818 _mod_union_table->mark(addr); 7819 } 7820 _collector->_ser_pmc_preclean_ovflw++; 7821 } else { 7822 // During the remark phase, we need to remember this oop 7823 // in the overflow list. 7824 _collector->push_on_overflow_list(obj); 7825 _collector->_ser_pmc_remark_ovflw++; 7826 } 7827 } 7828 } 7829 } 7830 7831 Par_PushAndMarkClosure::Par_PushAndMarkClosure(CMSCollector* collector, 7832 MemRegion span, 7833 ReferenceProcessor* rp, 7834 CMSBitMap* bit_map, 7835 OopTaskQueue* work_queue, 7836 CMSMarkStack* revisit_stack): 7837 Par_KlassRememberingOopClosure(collector, rp, revisit_stack), 7838 _span(span), 7839 _bit_map(bit_map), 7840 _work_queue(work_queue) 7841 { 7842 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 7843 } 7844 7845 void PushAndMarkClosure::do_oop(oop* p) { PushAndMarkClosure::do_oop_work(p); } 7846 void PushAndMarkClosure::do_oop(narrowOop* p) { PushAndMarkClosure::do_oop_work(p); } 7847 7848 // Grey object rescan during second checkpoint phase -- 7849 // the parallel version. 7850 void Par_PushAndMarkClosure::do_oop(oop obj) { 7851 // In the assert below, we ignore the mark word because 7852 // this oop may point to an already visited object that is 7853 // on the overflow stack (in which case the mark word has 7854 // been hijacked for chaining into the overflow stack -- 7855 // if this is the last object in the overflow stack then 7856 // its mark word will be NULL). Because this object may 7857 // have been subsequently popped off the global overflow 7858 // stack, and the mark word possibly restored to the prototypical 7859 // value, by the time we get to examined this failing assert in 7860 // the debugger, is_oop_or_null(false) may subsequently start 7861 // to hold. 7862 assert(obj->is_oop_or_null(true), 7863 "expected an oop or NULL"); 7864 HeapWord* addr = (HeapWord*)obj; 7865 // Check if oop points into the CMS generation 7866 // and is not marked 7867 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 7868 // a white object ... 7869 // If we manage to "claim" the object, by being the 7870 // first thread to mark it, then we push it on our 7871 // marking stack 7872 if (_bit_map->par_mark(addr)) { // ... now grey 7873 // push on work queue (grey set) 7874 bool simulate_overflow = false; 7875 NOT_PRODUCT( 7876 if (CMSMarkStackOverflowALot && 7877 _collector->par_simulate_overflow()) { 7878 // simulate a stack overflow 7879 simulate_overflow = true; 7880 } 7881 ) 7882 if (simulate_overflow || !_work_queue->push(obj)) { 7883 _collector->par_push_on_overflow_list(obj); 7884 _collector->_par_pmc_remark_ovflw++; // imprecise OK: no need to CAS 7885 } 7886 } // Else, some other thread got there first 7887 } 7888 } 7889 7890 void Par_PushAndMarkClosure::do_oop(oop* p) { Par_PushAndMarkClosure::do_oop_work(p); } 7891 void Par_PushAndMarkClosure::do_oop(narrowOop* p) { Par_PushAndMarkClosure::do_oop_work(p); } 7892 7893 void PushAndMarkClosure::remember_mdo(DataLayout* v) { 7894 // TBD 7895 } 7896 7897 void Par_PushAndMarkClosure::remember_mdo(DataLayout* v) { 7898 // TBD 7899 } 7900 7901 void CMSPrecleanRefsYieldClosure::do_yield_work() { 7902 DEBUG_ONLY(RememberKlassesChecker mux(false);) 7903 Mutex* bml = _collector->bitMapLock(); 7904 assert_lock_strong(bml); 7905 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 7906 "CMS thread should hold CMS token"); 7907 7908 bml->unlock(); 7909 ConcurrentMarkSweepThread::desynchronize(true); 7910 7911 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7912 7913 _collector->stopTimer(); 7914 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 7915 if (PrintCMSStatistics != 0) { 7916 _collector->incrementYields(); 7917 } 7918 _collector->icms_wait(); 7919 7920 // See the comment in coordinator_yield() 7921 for (unsigned i = 0; i < CMSYieldSleepCount && 7922 ConcurrentMarkSweepThread::should_yield() && 7923 !CMSCollector::foregroundGCIsActive(); ++i) { 7924 os::sleep(Thread::current(), 1, false); 7925 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7926 } 7927 7928 ConcurrentMarkSweepThread::synchronize(true); 7929 bml->lock(); 7930 7931 _collector->startTimer(); 7932 } 7933 7934 bool CMSPrecleanRefsYieldClosure::should_return() { 7935 if (ConcurrentMarkSweepThread::should_yield()) { 7936 do_yield_work(); 7937 } 7938 return _collector->foregroundGCIsActive(); 7939 } 7940 7941 void MarkFromDirtyCardsClosure::do_MemRegion(MemRegion mr) { 7942 assert(((size_t)mr.start())%CardTableModRefBS::card_size_in_words == 0, 7943 "mr should be aligned to start at a card boundary"); 7944 // We'd like to assert: 7945 // assert(mr.word_size()%CardTableModRefBS::card_size_in_words == 0, 7946 // "mr should be a range of cards"); 7947 // However, that would be too strong in one case -- the last 7948 // partition ends at _unallocated_block which, in general, can be 7949 // an arbitrary boundary, not necessarily card aligned. 7950 if (PrintCMSStatistics != 0) { 7951 _num_dirty_cards += 7952 mr.word_size()/CardTableModRefBS::card_size_in_words; 7953 } 7954 _space->object_iterate_mem(mr, &_scan_cl); 7955 } 7956 7957 SweepClosure::SweepClosure(CMSCollector* collector, 7958 ConcurrentMarkSweepGeneration* g, 7959 CMSBitMap* bitMap, bool should_yield) : 7960 _collector(collector), 7961 _g(g), 7962 _sp(g->cmsSpace()), 7963 _limit(_sp->sweep_limit()), 7964 _freelistLock(_sp->freelistLock()), 7965 _bitMap(bitMap), 7966 _yield(should_yield), 7967 _inFreeRange(false), // No free range at beginning of sweep 7968 _freeRangeInFreeLists(false), // No free range at beginning of sweep 7969 _lastFreeRangeCoalesced(false), 7970 _freeFinger(g->used_region().start()) 7971 { 7972 NOT_PRODUCT( 7973 _numObjectsFreed = 0; 7974 _numWordsFreed = 0; 7975 _numObjectsLive = 0; 7976 _numWordsLive = 0; 7977 _numObjectsAlreadyFree = 0; 7978 _numWordsAlreadyFree = 0; 7979 _last_fc = NULL; 7980 7981 _sp->initializeIndexedFreeListArrayReturnedBytes(); 7982 _sp->dictionary()->initialize_dict_returned_bytes(); 7983 ) 7984 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7985 "sweep _limit out of bounds"); 7986 if (CMSTraceSweeper) { 7987 gclog_or_tty->print_cr("\n====================\nStarting new sweep with limit " PTR_FORMAT, 7988 _limit); 7989 } 7990 } 7991 7992 void SweepClosure::print_on(outputStream* st) const { 7993 tty->print_cr("_sp = [" PTR_FORMAT "," PTR_FORMAT ")", 7994 _sp->bottom(), _sp->end()); 7995 tty->print_cr("_limit = " PTR_FORMAT, _limit); 7996 tty->print_cr("_freeFinger = " PTR_FORMAT, _freeFinger); 7997 NOT_PRODUCT(tty->print_cr("_last_fc = " PTR_FORMAT, _last_fc);) 7998 tty->print_cr("_inFreeRange = %d, _freeRangeInFreeLists = %d, _lastFreeRangeCoalesced = %d", 7999 _inFreeRange, _freeRangeInFreeLists, _lastFreeRangeCoalesced); 8000 } 8001 8002 #ifndef PRODUCT 8003 // Assertion checking only: no useful work in product mode -- 8004 // however, if any of the flags below become product flags, 8005 // you may need to review this code to see if it needs to be 8006 // enabled in product mode. 8007 SweepClosure::~SweepClosure() { 8008 assert_lock_strong(_freelistLock); 8009 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 8010 "sweep _limit out of bounds"); 8011 if (inFreeRange()) { 8012 warning("inFreeRange() should have been reset; dumping state of SweepClosure"); 8013 print(); 8014 ShouldNotReachHere(); 8015 } 8016 if (Verbose && PrintGC) { 8017 gclog_or_tty->print("Collected "SIZE_FORMAT" objects, " SIZE_FORMAT " bytes", 8018 _numObjectsFreed, _numWordsFreed*sizeof(HeapWord)); 8019 gclog_or_tty->print_cr("\nLive "SIZE_FORMAT" objects, " 8020 SIZE_FORMAT" bytes " 8021 "Already free "SIZE_FORMAT" objects, "SIZE_FORMAT" bytes", 8022 _numObjectsLive, _numWordsLive*sizeof(HeapWord), 8023 _numObjectsAlreadyFree, _numWordsAlreadyFree*sizeof(HeapWord)); 8024 size_t totalBytes = (_numWordsFreed + _numWordsLive + _numWordsAlreadyFree) 8025 * sizeof(HeapWord); 8026 gclog_or_tty->print_cr("Total sweep: "SIZE_FORMAT" bytes", totalBytes); 8027 8028 if (PrintCMSStatistics && CMSVerifyReturnedBytes) { 8029 size_t indexListReturnedBytes = _sp->sumIndexedFreeListArrayReturnedBytes(); 8030 size_t dict_returned_bytes = _sp->dictionary()->sum_dict_returned_bytes(); 8031 size_t returned_bytes = indexListReturnedBytes + dict_returned_bytes; 8032 gclog_or_tty->print("Returned "SIZE_FORMAT" bytes", returned_bytes); 8033 gclog_or_tty->print(" Indexed List Returned "SIZE_FORMAT" bytes", 8034 indexListReturnedBytes); 8035 gclog_or_tty->print_cr(" Dictionary Returned "SIZE_FORMAT" bytes", 8036 dict_returned_bytes); 8037 } 8038 } 8039 if (CMSTraceSweeper) { 8040 gclog_or_tty->print_cr("end of sweep with _limit = " PTR_FORMAT "\n================", 8041 _limit); 8042 } 8043 } 8044 #endif // PRODUCT 8045 8046 void SweepClosure::initialize_free_range(HeapWord* freeFinger, 8047 bool freeRangeInFreeLists) { 8048 if (CMSTraceSweeper) { 8049 gclog_or_tty->print("---- Start free range at 0x%x with free block (%d)\n", 8050 freeFinger, freeRangeInFreeLists); 8051 } 8052 assert(!inFreeRange(), "Trampling existing free range"); 8053 set_inFreeRange(true); 8054 set_lastFreeRangeCoalesced(false); 8055 8056 set_freeFinger(freeFinger); 8057 set_freeRangeInFreeLists(freeRangeInFreeLists); 8058 if (CMSTestInFreeList) { 8059 if (freeRangeInFreeLists) { 8060 FreeChunk* fc = (FreeChunk*) freeFinger; 8061 assert(fc->is_free(), "A chunk on the free list should be free."); 8062 assert(fc->size() > 0, "Free range should have a size"); 8063 assert(_sp->verify_chunk_in_free_list(fc), "Chunk is not in free lists"); 8064 } 8065 } 8066 } 8067 8068 // Note that the sweeper runs concurrently with mutators. Thus, 8069 // it is possible for direct allocation in this generation to happen 8070 // in the middle of the sweep. Note that the sweeper also coalesces 8071 // contiguous free blocks. Thus, unless the sweeper and the allocator 8072 // synchronize appropriately freshly allocated blocks may get swept up. 8073 // This is accomplished by the sweeper locking the free lists while 8074 // it is sweeping. Thus blocks that are determined to be free are 8075 // indeed free. There is however one additional complication: 8076 // blocks that have been allocated since the final checkpoint and 8077 // mark, will not have been marked and so would be treated as 8078 // unreachable and swept up. To prevent this, the allocator marks 8079 // the bit map when allocating during the sweep phase. This leads, 8080 // however, to a further complication -- objects may have been allocated 8081 // but not yet initialized -- in the sense that the header isn't yet 8082 // installed. The sweeper can not then determine the size of the block 8083 // in order to skip over it. To deal with this case, we use a technique 8084 // (due to Printezis) to encode such uninitialized block sizes in the 8085 // bit map. Since the bit map uses a bit per every HeapWord, but the 8086 // CMS generation has a minimum object size of 3 HeapWords, it follows 8087 // that "normal marks" won't be adjacent in the bit map (there will 8088 // always be at least two 0 bits between successive 1 bits). We make use 8089 // of these "unused" bits to represent uninitialized blocks -- the bit 8090 // corresponding to the start of the uninitialized object and the next 8091 // bit are both set. Finally, a 1 bit marks the end of the object that 8092 // started with the two consecutive 1 bits to indicate its potentially 8093 // uninitialized state. 8094 8095 size_t SweepClosure::do_blk_careful(HeapWord* addr) { 8096 FreeChunk* fc = (FreeChunk*)addr; 8097 size_t res; 8098 8099 // Check if we are done sweeping. Below we check "addr >= _limit" rather 8100 // than "addr == _limit" because although _limit was a block boundary when 8101 // we started the sweep, it may no longer be one because heap expansion 8102 // may have caused us to coalesce the block ending at the address _limit 8103 // with a newly expanded chunk (this happens when _limit was set to the 8104 // previous _end of the space), so we may have stepped past _limit: 8105 // see the following Zeno-like trail of CRs 6977970, 7008136, 7042740. 8106 if (addr >= _limit) { // we have swept up to or past the limit: finish up 8107 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 8108 "sweep _limit out of bounds"); 8109 assert(addr < _sp->end(), "addr out of bounds"); 8110 // Flush any free range we might be holding as a single 8111 // coalesced chunk to the appropriate free list. 8112 if (inFreeRange()) { 8113 assert(freeFinger() >= _sp->bottom() && freeFinger() < _limit, 8114 err_msg("freeFinger() " PTR_FORMAT" is out-of-bounds", freeFinger())); 8115 flush_cur_free_chunk(freeFinger(), 8116 pointer_delta(addr, freeFinger())); 8117 if (CMSTraceSweeper) { 8118 gclog_or_tty->print("Sweep: last chunk: "); 8119 gclog_or_tty->print("put_free_blk 0x%x ("SIZE_FORMAT") " 8120 "[coalesced:"SIZE_FORMAT"]\n", 8121 freeFinger(), pointer_delta(addr, freeFinger()), 8122 lastFreeRangeCoalesced()); 8123 } 8124 } 8125 8126 // help the iterator loop finish 8127 return pointer_delta(_sp->end(), addr); 8128 } 8129 8130 assert(addr < _limit, "sweep invariant"); 8131 // check if we should yield 8132 do_yield_check(addr); 8133 if (fc->is_free()) { 8134 // Chunk that is already free 8135 res = fc->size(); 8136 do_already_free_chunk(fc); 8137 debug_only(_sp->verifyFreeLists()); 8138 // If we flush the chunk at hand in lookahead_and_flush() 8139 // and it's coalesced with a preceding chunk, then the 8140 // process of "mangling" the payload of the coalesced block 8141 // will cause erasure of the size information from the 8142 // (erstwhile) header of all the coalesced blocks but the 8143 // first, so the first disjunct in the assert will not hold 8144 // in that specific case (in which case the second disjunct 8145 // will hold). 8146 assert(res == fc->size() || ((HeapWord*)fc) + res >= _limit, 8147 "Otherwise the size info doesn't change at this step"); 8148 NOT_PRODUCT( 8149 _numObjectsAlreadyFree++; 8150 _numWordsAlreadyFree += res; 8151 ) 8152 NOT_PRODUCT(_last_fc = fc;) 8153 } else if (!_bitMap->isMarked(addr)) { 8154 // Chunk is fresh garbage 8155 res = do_garbage_chunk(fc); 8156 debug_only(_sp->verifyFreeLists()); 8157 NOT_PRODUCT( 8158 _numObjectsFreed++; 8159 _numWordsFreed += res; 8160 ) 8161 } else { 8162 // Chunk that is alive. 8163 res = do_live_chunk(fc); 8164 debug_only(_sp->verifyFreeLists()); 8165 NOT_PRODUCT( 8166 _numObjectsLive++; 8167 _numWordsLive += res; 8168 ) 8169 } 8170 return res; 8171 } 8172 8173 // For the smart allocation, record following 8174 // split deaths - a free chunk is removed from its free list because 8175 // it is being split into two or more chunks. 8176 // split birth - a free chunk is being added to its free list because 8177 // a larger free chunk has been split and resulted in this free chunk. 8178 // coal death - a free chunk is being removed from its free list because 8179 // it is being coalesced into a large free chunk. 8180 // coal birth - a free chunk is being added to its free list because 8181 // it was created when two or more free chunks where coalesced into 8182 // this free chunk. 8183 // 8184 // These statistics are used to determine the desired number of free 8185 // chunks of a given size. The desired number is chosen to be relative 8186 // to the end of a CMS sweep. The desired number at the end of a sweep 8187 // is the 8188 // count-at-end-of-previous-sweep (an amount that was enough) 8189 // - count-at-beginning-of-current-sweep (the excess) 8190 // + split-births (gains in this size during interval) 8191 // - split-deaths (demands on this size during interval) 8192 // where the interval is from the end of one sweep to the end of the 8193 // next. 8194 // 8195 // When sweeping the sweeper maintains an accumulated chunk which is 8196 // the chunk that is made up of chunks that have been coalesced. That 8197 // will be termed the left-hand chunk. A new chunk of garbage that 8198 // is being considered for coalescing will be referred to as the 8199 // right-hand chunk. 8200 // 8201 // When making a decision on whether to coalesce a right-hand chunk with 8202 // the current left-hand chunk, the current count vs. the desired count 8203 // of the left-hand chunk is considered. Also if the right-hand chunk 8204 // is near the large chunk at the end of the heap (see 8205 // ConcurrentMarkSweepGeneration::isNearLargestChunk()), then the 8206 // left-hand chunk is coalesced. 8207 // 8208 // When making a decision about whether to split a chunk, the desired count 8209 // vs. the current count of the candidate to be split is also considered. 8210 // If the candidate is underpopulated (currently fewer chunks than desired) 8211 // a chunk of an overpopulated (currently more chunks than desired) size may 8212 // be chosen. The "hint" associated with a free list, if non-null, points 8213 // to a free list which may be overpopulated. 8214 // 8215 8216 void SweepClosure::do_already_free_chunk(FreeChunk* fc) { 8217 const size_t size = fc->size(); 8218 // Chunks that cannot be coalesced are not in the 8219 // free lists. 8220 if (CMSTestInFreeList && !fc->cantCoalesce()) { 8221 assert(_sp->verify_chunk_in_free_list(fc), 8222 "free chunk should be in free lists"); 8223 } 8224 // a chunk that is already free, should not have been 8225 // marked in the bit map 8226 HeapWord* const addr = (HeapWord*) fc; 8227 assert(!_bitMap->isMarked(addr), "free chunk should be unmarked"); 8228 // Verify that the bit map has no bits marked between 8229 // addr and purported end of this block. 8230 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 8231 8232 // Some chunks cannot be coalesced under any circumstances. 8233 // See the definition of cantCoalesce(). 8234 if (!fc->cantCoalesce()) { 8235 // This chunk can potentially be coalesced. 8236 if (_sp->adaptive_freelists()) { 8237 // All the work is done in 8238 do_post_free_or_garbage_chunk(fc, size); 8239 } else { // Not adaptive free lists 8240 // this is a free chunk that can potentially be coalesced by the sweeper; 8241 if (!inFreeRange()) { 8242 // if the next chunk is a free block that can't be coalesced 8243 // it doesn't make sense to remove this chunk from the free lists 8244 FreeChunk* nextChunk = (FreeChunk*)(addr + size); 8245 assert((HeapWord*)nextChunk <= _sp->end(), "Chunk size out of bounds?"); 8246 if ((HeapWord*)nextChunk < _sp->end() && // There is another free chunk to the right ... 8247 nextChunk->is_free() && // ... which is free... 8248 nextChunk->cantCoalesce()) { // ... but can't be coalesced 8249 // nothing to do 8250 } else { 8251 // Potentially the start of a new free range: 8252 // Don't eagerly remove it from the free lists. 8253 // No need to remove it if it will just be put 8254 // back again. (Also from a pragmatic point of view 8255 // if it is a free block in a region that is beyond 8256 // any allocated blocks, an assertion will fail) 8257 // Remember the start of a free run. 8258 initialize_free_range(addr, true); 8259 // end - can coalesce with next chunk 8260 } 8261 } else { 8262 // the midst of a free range, we are coalescing 8263 print_free_block_coalesced(fc); 8264 if (CMSTraceSweeper) { 8265 gclog_or_tty->print(" -- pick up free block 0x%x (%d)\n", fc, size); 8266 } 8267 // remove it from the free lists 8268 _sp->removeFreeChunkFromFreeLists(fc); 8269 set_lastFreeRangeCoalesced(true); 8270 // If the chunk is being coalesced and the current free range is 8271 // in the free lists, remove the current free range so that it 8272 // will be returned to the free lists in its entirety - all 8273 // the coalesced pieces included. 8274 if (freeRangeInFreeLists()) { 8275 FreeChunk* ffc = (FreeChunk*) freeFinger(); 8276 assert(ffc->size() == pointer_delta(addr, freeFinger()), 8277 "Size of free range is inconsistent with chunk size."); 8278 if (CMSTestInFreeList) { 8279 assert(_sp->verify_chunk_in_free_list(ffc), 8280 "free range is not in free lists"); 8281 } 8282 _sp->removeFreeChunkFromFreeLists(ffc); 8283 set_freeRangeInFreeLists(false); 8284 } 8285 } 8286 } 8287 // Note that if the chunk is not coalescable (the else arm 8288 // below), we unconditionally flush, without needing to do 8289 // a "lookahead," as we do below. 8290 if (inFreeRange()) lookahead_and_flush(fc, size); 8291 } else { 8292 // Code path common to both original and adaptive free lists. 8293 8294 // cant coalesce with previous block; this should be treated 8295 // as the end of a free run if any 8296 if (inFreeRange()) { 8297 // we kicked some butt; time to pick up the garbage 8298 assert(freeFinger() < addr, "freeFinger points too high"); 8299 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 8300 } 8301 // else, nothing to do, just continue 8302 } 8303 } 8304 8305 size_t SweepClosure::do_garbage_chunk(FreeChunk* fc) { 8306 // This is a chunk of garbage. It is not in any free list. 8307 // Add it to a free list or let it possibly be coalesced into 8308 // a larger chunk. 8309 HeapWord* const addr = (HeapWord*) fc; 8310 const size_t size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 8311 8312 if (_sp->adaptive_freelists()) { 8313 // Verify that the bit map has no bits marked between 8314 // addr and purported end of just dead object. 8315 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 8316 8317 do_post_free_or_garbage_chunk(fc, size); 8318 } else { 8319 if (!inFreeRange()) { 8320 // start of a new free range 8321 assert(size > 0, "A free range should have a size"); 8322 initialize_free_range(addr, false); 8323 } else { 8324 // this will be swept up when we hit the end of the 8325 // free range 8326 if (CMSTraceSweeper) { 8327 gclog_or_tty->print(" -- pick up garbage 0x%x (%d) \n", fc, size); 8328 } 8329 // If the chunk is being coalesced and the current free range is 8330 // in the free lists, remove the current free range so that it 8331 // will be returned to the free lists in its entirety - all 8332 // the coalesced pieces included. 8333 if (freeRangeInFreeLists()) { 8334 FreeChunk* ffc = (FreeChunk*)freeFinger(); 8335 assert(ffc->size() == pointer_delta(addr, freeFinger()), 8336 "Size of free range is inconsistent with chunk size."); 8337 if (CMSTestInFreeList) { 8338 assert(_sp->verify_chunk_in_free_list(ffc), 8339 "free range is not in free lists"); 8340 } 8341 _sp->removeFreeChunkFromFreeLists(ffc); 8342 set_freeRangeInFreeLists(false); 8343 } 8344 set_lastFreeRangeCoalesced(true); 8345 } 8346 // this will be swept up when we hit the end of the free range 8347 8348 // Verify that the bit map has no bits marked between 8349 // addr and purported end of just dead object. 8350 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 8351 } 8352 assert(_limit >= addr + size, 8353 "A freshly garbage chunk can't possibly straddle over _limit"); 8354 if (inFreeRange()) lookahead_and_flush(fc, size); 8355 return size; 8356 } 8357 8358 size_t SweepClosure::do_live_chunk(FreeChunk* fc) { 8359 HeapWord* addr = (HeapWord*) fc; 8360 // The sweeper has just found a live object. Return any accumulated 8361 // left hand chunk to the free lists. 8362 if (inFreeRange()) { 8363 assert(freeFinger() < addr, "freeFinger points too high"); 8364 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 8365 } 8366 8367 // This object is live: we'd normally expect this to be 8368 // an oop, and like to assert the following: 8369 // assert(oop(addr)->is_oop(), "live block should be an oop"); 8370 // However, as we commented above, this may be an object whose 8371 // header hasn't yet been initialized. 8372 size_t size; 8373 assert(_bitMap->isMarked(addr), "Tautology for this control point"); 8374 if (_bitMap->isMarked(addr + 1)) { 8375 // Determine the size from the bit map, rather than trying to 8376 // compute it from the object header. 8377 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 8378 size = pointer_delta(nextOneAddr + 1, addr); 8379 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 8380 "alignment problem"); 8381 8382 #ifdef DEBUG 8383 if (oop(addr)->klass_or_null() != NULL && 8384 ( !_collector->should_unload_classes() 8385 || (oop(addr)->is_parsable()) && 8386 oop(addr)->is_conc_safe())) { 8387 // Ignore mark word because we are running concurrent with mutators 8388 assert(oop(addr)->is_oop(true), "live block should be an oop"); 8389 // is_conc_safe is checked before performing this assertion 8390 // because an object that is not is_conc_safe may yet have 8391 // the return from size() correct. 8392 assert(size == 8393 CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()), 8394 "P-mark and computed size do not agree"); 8395 } 8396 #endif 8397 8398 } else { 8399 // This should be an initialized object that's alive. 8400 assert(oop(addr)->klass_or_null() != NULL && 8401 (!_collector->should_unload_classes() 8402 || oop(addr)->is_parsable()), 8403 "Should be an initialized object"); 8404 // Note that there are objects used during class redefinition, 8405 // e.g. merge_cp in VM_RedefineClasses::merge_cp_and_rewrite(), 8406 // which are discarded with their is_conc_safe state still 8407 // false. These object may be floating garbage so may be 8408 // seen here. If they are floating garbage their size 8409 // should be attainable from their klass. Do not that 8410 // is_conc_safe() is true for oop(addr). 8411 // Ignore mark word because we are running concurrent with mutators 8412 assert(oop(addr)->is_oop(true), "live block should be an oop"); 8413 // Verify that the bit map has no bits marked between 8414 // addr and purported end of this block. 8415 size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 8416 assert(size >= 3, "Necessary for Printezis marks to work"); 8417 assert(!_bitMap->isMarked(addr+1), "Tautology for this control point"); 8418 DEBUG_ONLY(_bitMap->verifyNoOneBitsInRange(addr+2, addr+size);) 8419 } 8420 return size; 8421 } 8422 8423 void SweepClosure::do_post_free_or_garbage_chunk(FreeChunk* fc, 8424 size_t chunkSize) { 8425 // do_post_free_or_garbage_chunk() should only be called in the case 8426 // of the adaptive free list allocator. 8427 const bool fcInFreeLists = fc->is_free(); 8428 assert(_sp->adaptive_freelists(), "Should only be used in this case."); 8429 assert((HeapWord*)fc <= _limit, "sweep invariant"); 8430 if (CMSTestInFreeList && fcInFreeLists) { 8431 assert(_sp->verify_chunk_in_free_list(fc), "free chunk is not in free lists"); 8432 } 8433 8434 if (CMSTraceSweeper) { 8435 gclog_or_tty->print_cr(" -- pick up another chunk at 0x%x (%d)", fc, chunkSize); 8436 } 8437 8438 HeapWord* const fc_addr = (HeapWord*) fc; 8439 8440 bool coalesce; 8441 const size_t left = pointer_delta(fc_addr, freeFinger()); 8442 const size_t right = chunkSize; 8443 switch (FLSCoalescePolicy) { 8444 // numeric value forms a coalition aggressiveness metric 8445 case 0: { // never coalesce 8446 coalesce = false; 8447 break; 8448 } 8449 case 1: { // coalesce if left & right chunks on overpopulated lists 8450 coalesce = _sp->coalOverPopulated(left) && 8451 _sp->coalOverPopulated(right); 8452 break; 8453 } 8454 case 2: { // coalesce if left chunk on overpopulated list (default) 8455 coalesce = _sp->coalOverPopulated(left); 8456 break; 8457 } 8458 case 3: { // coalesce if left OR right chunk on overpopulated list 8459 coalesce = _sp->coalOverPopulated(left) || 8460 _sp->coalOverPopulated(right); 8461 break; 8462 } 8463 case 4: { // always coalesce 8464 coalesce = true; 8465 break; 8466 } 8467 default: 8468 ShouldNotReachHere(); 8469 } 8470 8471 // Should the current free range be coalesced? 8472 // If the chunk is in a free range and either we decided to coalesce above 8473 // or the chunk is near the large block at the end of the heap 8474 // (isNearLargestChunk() returns true), then coalesce this chunk. 8475 const bool doCoalesce = inFreeRange() 8476 && (coalesce || _g->isNearLargestChunk(fc_addr)); 8477 if (doCoalesce) { 8478 // Coalesce the current free range on the left with the new 8479 // chunk on the right. If either is on a free list, 8480 // it must be removed from the list and stashed in the closure. 8481 if (freeRangeInFreeLists()) { 8482 FreeChunk* const ffc = (FreeChunk*)freeFinger(); 8483 assert(ffc->size() == pointer_delta(fc_addr, freeFinger()), 8484 "Size of free range is inconsistent with chunk size."); 8485 if (CMSTestInFreeList) { 8486 assert(_sp->verify_chunk_in_free_list(ffc), 8487 "Chunk is not in free lists"); 8488 } 8489 _sp->coalDeath(ffc->size()); 8490 _sp->removeFreeChunkFromFreeLists(ffc); 8491 set_freeRangeInFreeLists(false); 8492 } 8493 if (fcInFreeLists) { 8494 _sp->coalDeath(chunkSize); 8495 assert(fc->size() == chunkSize, 8496 "The chunk has the wrong size or is not in the free lists"); 8497 _sp->removeFreeChunkFromFreeLists(fc); 8498 } 8499 set_lastFreeRangeCoalesced(true); 8500 print_free_block_coalesced(fc); 8501 } else { // not in a free range and/or should not coalesce 8502 // Return the current free range and start a new one. 8503 if (inFreeRange()) { 8504 // In a free range but cannot coalesce with the right hand chunk. 8505 // Put the current free range into the free lists. 8506 flush_cur_free_chunk(freeFinger(), 8507 pointer_delta(fc_addr, freeFinger())); 8508 } 8509 // Set up for new free range. Pass along whether the right hand 8510 // chunk is in the free lists. 8511 initialize_free_range((HeapWord*)fc, fcInFreeLists); 8512 } 8513 } 8514 8515 // Lookahead flush: 8516 // If we are tracking a free range, and this is the last chunk that 8517 // we'll look at because its end crosses past _limit, we'll preemptively 8518 // flush it along with any free range we may be holding on to. Note that 8519 // this can be the case only for an already free or freshly garbage 8520 // chunk. If this block is an object, it can never straddle 8521 // over _limit. The "straddling" occurs when _limit is set at 8522 // the previous end of the space when this cycle started, and 8523 // a subsequent heap expansion caused the previously co-terminal 8524 // free block to be coalesced with the newly expanded portion, 8525 // thus rendering _limit a non-block-boundary making it dangerous 8526 // for the sweeper to step over and examine. 8527 void SweepClosure::lookahead_and_flush(FreeChunk* fc, size_t chunk_size) { 8528 assert(inFreeRange(), "Should only be called if currently in a free range."); 8529 HeapWord* const eob = ((HeapWord*)fc) + chunk_size; 8530 assert(_sp->used_region().contains(eob - 1), 8531 err_msg("eob = " PTR_FORMAT " out of bounds wrt _sp = [" PTR_FORMAT "," PTR_FORMAT ")" 8532 " when examining fc = " PTR_FORMAT "(" SIZE_FORMAT ")", 8533 _limit, _sp->bottom(), _sp->end(), fc, chunk_size)); 8534 if (eob >= _limit) { 8535 assert(eob == _limit || fc->is_free(), "Only a free chunk should allow us to cross over the limit"); 8536 if (CMSTraceSweeper) { 8537 gclog_or_tty->print_cr("_limit " PTR_FORMAT " reached or crossed by block " 8538 "[" PTR_FORMAT "," PTR_FORMAT ") in space " 8539 "[" PTR_FORMAT "," PTR_FORMAT ")", 8540 _limit, fc, eob, _sp->bottom(), _sp->end()); 8541 } 8542 // Return the storage we are tracking back into the free lists. 8543 if (CMSTraceSweeper) { 8544 gclog_or_tty->print_cr("Flushing ... "); 8545 } 8546 assert(freeFinger() < eob, "Error"); 8547 flush_cur_free_chunk( freeFinger(), pointer_delta(eob, freeFinger())); 8548 } 8549 } 8550 8551 void SweepClosure::flush_cur_free_chunk(HeapWord* chunk, size_t size) { 8552 assert(inFreeRange(), "Should only be called if currently in a free range."); 8553 assert(size > 0, 8554 "A zero sized chunk cannot be added to the free lists."); 8555 if (!freeRangeInFreeLists()) { 8556 if (CMSTestInFreeList) { 8557 FreeChunk* fc = (FreeChunk*) chunk; 8558 fc->set_size(size); 8559 assert(!_sp->verify_chunk_in_free_list(fc), 8560 "chunk should not be in free lists yet"); 8561 } 8562 if (CMSTraceSweeper) { 8563 gclog_or_tty->print_cr(" -- add free block 0x%x (%d) to free lists", 8564 chunk, size); 8565 } 8566 // A new free range is going to be starting. The current 8567 // free range has not been added to the free lists yet or 8568 // was removed so add it back. 8569 // If the current free range was coalesced, then the death 8570 // of the free range was recorded. Record a birth now. 8571 if (lastFreeRangeCoalesced()) { 8572 _sp->coalBirth(size); 8573 } 8574 _sp->addChunkAndRepairOffsetTable(chunk, size, 8575 lastFreeRangeCoalesced()); 8576 } else if (CMSTraceSweeper) { 8577 gclog_or_tty->print_cr("Already in free list: nothing to flush"); 8578 } 8579 set_inFreeRange(false); 8580 set_freeRangeInFreeLists(false); 8581 } 8582 8583 // We take a break if we've been at this for a while, 8584 // so as to avoid monopolizing the locks involved. 8585 void SweepClosure::do_yield_work(HeapWord* addr) { 8586 // Return current free chunk being used for coalescing (if any) 8587 // to the appropriate freelist. After yielding, the next 8588 // free block encountered will start a coalescing range of 8589 // free blocks. If the next free block is adjacent to the 8590 // chunk just flushed, they will need to wait for the next 8591 // sweep to be coalesced. 8592 if (inFreeRange()) { 8593 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 8594 } 8595 8596 // First give up the locks, then yield, then re-lock. 8597 // We should probably use a constructor/destructor idiom to 8598 // do this unlock/lock or modify the MutexUnlocker class to 8599 // serve our purpose. XXX 8600 assert_lock_strong(_bitMap->lock()); 8601 assert_lock_strong(_freelistLock); 8602 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 8603 "CMS thread should hold CMS token"); 8604 _bitMap->lock()->unlock(); 8605 _freelistLock->unlock(); 8606 ConcurrentMarkSweepThread::desynchronize(true); 8607 ConcurrentMarkSweepThread::acknowledge_yield_request(); 8608 _collector->stopTimer(); 8609 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 8610 if (PrintCMSStatistics != 0) { 8611 _collector->incrementYields(); 8612 } 8613 _collector->icms_wait(); 8614 8615 // See the comment in coordinator_yield() 8616 for (unsigned i = 0; i < CMSYieldSleepCount && 8617 ConcurrentMarkSweepThread::should_yield() && 8618 !CMSCollector::foregroundGCIsActive(); ++i) { 8619 os::sleep(Thread::current(), 1, false); 8620 ConcurrentMarkSweepThread::acknowledge_yield_request(); 8621 } 8622 8623 ConcurrentMarkSweepThread::synchronize(true); 8624 _freelistLock->lock(); 8625 _bitMap->lock()->lock_without_safepoint_check(); 8626 _collector->startTimer(); 8627 } 8628 8629 #ifndef PRODUCT 8630 // This is actually very useful in a product build if it can 8631 // be called from the debugger. Compile it into the product 8632 // as needed. 8633 bool debug_verify_chunk_in_free_list(FreeChunk* fc) { 8634 return debug_cms_space->verify_chunk_in_free_list(fc); 8635 } 8636 #endif 8637 8638 void SweepClosure::print_free_block_coalesced(FreeChunk* fc) const { 8639 if (CMSTraceSweeper) { 8640 gclog_or_tty->print_cr("Sweep:coal_free_blk " PTR_FORMAT " (" SIZE_FORMAT ")", 8641 fc, fc->size()); 8642 } 8643 } 8644 8645 // CMSIsAliveClosure 8646 bool CMSIsAliveClosure::do_object_b(oop obj) { 8647 HeapWord* addr = (HeapWord*)obj; 8648 return addr != NULL && 8649 (!_span.contains(addr) || _bit_map->isMarked(addr)); 8650 } 8651 8652 CMSKeepAliveClosure::CMSKeepAliveClosure( CMSCollector* collector, 8653 MemRegion span, 8654 CMSBitMap* bit_map, CMSMarkStack* mark_stack, 8655 CMSMarkStack* revisit_stack, bool cpc): 8656 KlassRememberingOopClosure(collector, NULL, revisit_stack), 8657 _span(span), 8658 _bit_map(bit_map), 8659 _mark_stack(mark_stack), 8660 _concurrent_precleaning(cpc) { 8661 assert(!_span.is_empty(), "Empty span could spell trouble"); 8662 } 8663 8664 8665 // CMSKeepAliveClosure: the serial version 8666 void CMSKeepAliveClosure::do_oop(oop obj) { 8667 HeapWord* addr = (HeapWord*)obj; 8668 if (_span.contains(addr) && 8669 !_bit_map->isMarked(addr)) { 8670 _bit_map->mark(addr); 8671 bool simulate_overflow = false; 8672 NOT_PRODUCT( 8673 if (CMSMarkStackOverflowALot && 8674 _collector->simulate_overflow()) { 8675 // simulate a stack overflow 8676 simulate_overflow = true; 8677 } 8678 ) 8679 if (simulate_overflow || !_mark_stack->push(obj)) { 8680 if (_concurrent_precleaning) { 8681 // We dirty the overflown object and let the remark 8682 // phase deal with it. 8683 assert(_collector->overflow_list_is_empty(), "Error"); 8684 // In the case of object arrays, we need to dirty all of 8685 // the cards that the object spans. No locking or atomics 8686 // are needed since no one else can be mutating the mod union 8687 // table. 8688 if (obj->is_objArray()) { 8689 size_t sz = obj->size(); 8690 HeapWord* end_card_addr = 8691 (HeapWord*)round_to((intptr_t)(addr+sz), CardTableModRefBS::card_size); 8692 MemRegion redirty_range = MemRegion(addr, end_card_addr); 8693 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 8694 _collector->_modUnionTable.mark_range(redirty_range); 8695 } else { 8696 _collector->_modUnionTable.mark(addr); 8697 } 8698 _collector->_ser_kac_preclean_ovflw++; 8699 } else { 8700 _collector->push_on_overflow_list(obj); 8701 _collector->_ser_kac_ovflw++; 8702 } 8703 } 8704 } 8705 } 8706 8707 void CMSKeepAliveClosure::do_oop(oop* p) { CMSKeepAliveClosure::do_oop_work(p); } 8708 void CMSKeepAliveClosure::do_oop(narrowOop* p) { CMSKeepAliveClosure::do_oop_work(p); } 8709 8710 // CMSParKeepAliveClosure: a parallel version of the above. 8711 // The work queues are private to each closure (thread), 8712 // but (may be) available for stealing by other threads. 8713 void CMSParKeepAliveClosure::do_oop(oop obj) { 8714 HeapWord* addr = (HeapWord*)obj; 8715 if (_span.contains(addr) && 8716 !_bit_map->isMarked(addr)) { 8717 // In general, during recursive tracing, several threads 8718 // may be concurrently getting here; the first one to 8719 // "tag" it, claims it. 8720 if (_bit_map->par_mark(addr)) { 8721 bool res = _work_queue->push(obj); 8722 assert(res, "Low water mark should be much less than capacity"); 8723 // Do a recursive trim in the hope that this will keep 8724 // stack usage lower, but leave some oops for potential stealers 8725 trim_queue(_low_water_mark); 8726 } // Else, another thread got there first 8727 } 8728 } 8729 8730 void CMSParKeepAliveClosure::do_oop(oop* p) { CMSParKeepAliveClosure::do_oop_work(p); } 8731 void CMSParKeepAliveClosure::do_oop(narrowOop* p) { CMSParKeepAliveClosure::do_oop_work(p); } 8732 8733 void CMSParKeepAliveClosure::trim_queue(uint max) { 8734 while (_work_queue->size() > max) { 8735 oop new_oop; 8736 if (_work_queue->pop_local(new_oop)) { 8737 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 8738 assert(_bit_map->isMarked((HeapWord*)new_oop), 8739 "no white objects on this stack!"); 8740 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 8741 // iterate over the oops in this oop, marking and pushing 8742 // the ones in CMS heap (i.e. in _span). 8743 new_oop->oop_iterate(&_mark_and_push); 8744 } 8745 } 8746 } 8747 8748 CMSInnerParMarkAndPushClosure::CMSInnerParMarkAndPushClosure( 8749 CMSCollector* collector, 8750 MemRegion span, CMSBitMap* bit_map, 8751 CMSMarkStack* revisit_stack, 8752 OopTaskQueue* work_queue): 8753 Par_KlassRememberingOopClosure(collector, NULL, revisit_stack), 8754 _span(span), 8755 _bit_map(bit_map), 8756 _work_queue(work_queue) { } 8757 8758 void CMSInnerParMarkAndPushClosure::do_oop(oop obj) { 8759 HeapWord* addr = (HeapWord*)obj; 8760 if (_span.contains(addr) && 8761 !_bit_map->isMarked(addr)) { 8762 if (_bit_map->par_mark(addr)) { 8763 bool simulate_overflow = false; 8764 NOT_PRODUCT( 8765 if (CMSMarkStackOverflowALot && 8766 _collector->par_simulate_overflow()) { 8767 // simulate a stack overflow 8768 simulate_overflow = true; 8769 } 8770 ) 8771 if (simulate_overflow || !_work_queue->push(obj)) { 8772 _collector->par_push_on_overflow_list(obj); 8773 _collector->_par_kac_ovflw++; 8774 } 8775 } // Else another thread got there already 8776 } 8777 } 8778 8779 void CMSInnerParMarkAndPushClosure::do_oop(oop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); } 8780 void CMSInnerParMarkAndPushClosure::do_oop(narrowOop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); } 8781 8782 ////////////////////////////////////////////////////////////////// 8783 // CMSExpansionCause ///////////////////////////// 8784 ////////////////////////////////////////////////////////////////// 8785 const char* CMSExpansionCause::to_string(CMSExpansionCause::Cause cause) { 8786 switch (cause) { 8787 case _no_expansion: 8788 return "No expansion"; 8789 case _satisfy_free_ratio: 8790 return "Free ratio"; 8791 case _satisfy_promotion: 8792 return "Satisfy promotion"; 8793 case _satisfy_allocation: 8794 return "allocation"; 8795 case _allocate_par_lab: 8796 return "Par LAB"; 8797 case _allocate_par_spooling_space: 8798 return "Par Spooling Space"; 8799 case _adaptive_size_policy: 8800 return "Ergonomics"; 8801 default: 8802 return "unknown"; 8803 } 8804 } 8805 8806 void CMSDrainMarkingStackClosure::do_void() { 8807 // the max number to take from overflow list at a time 8808 const size_t num = _mark_stack->capacity()/4; 8809 assert(!_concurrent_precleaning || _collector->overflow_list_is_empty(), 8810 "Overflow list should be NULL during concurrent phases"); 8811 while (!_mark_stack->isEmpty() || 8812 // if stack is empty, check the overflow list 8813 _collector->take_from_overflow_list(num, _mark_stack)) { 8814 oop obj = _mark_stack->pop(); 8815 HeapWord* addr = (HeapWord*)obj; 8816 assert(_span.contains(addr), "Should be within span"); 8817 assert(_bit_map->isMarked(addr), "Should be marked"); 8818 assert(obj->is_oop(), "Should be an oop"); 8819 obj->oop_iterate(_keep_alive); 8820 } 8821 } 8822 8823 void CMSParDrainMarkingStackClosure::do_void() { 8824 // drain queue 8825 trim_queue(0); 8826 } 8827 8828 // Trim our work_queue so its length is below max at return 8829 void CMSParDrainMarkingStackClosure::trim_queue(uint max) { 8830 while (_work_queue->size() > max) { 8831 oop new_oop; 8832 if (_work_queue->pop_local(new_oop)) { 8833 assert(new_oop->is_oop(), "Expected an oop"); 8834 assert(_bit_map->isMarked((HeapWord*)new_oop), 8835 "no white objects on this stack!"); 8836 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 8837 // iterate over the oops in this oop, marking and pushing 8838 // the ones in CMS heap (i.e. in _span). 8839 new_oop->oop_iterate(&_mark_and_push); 8840 } 8841 } 8842 } 8843 8844 //////////////////////////////////////////////////////////////////// 8845 // Support for Marking Stack Overflow list handling and related code 8846 //////////////////////////////////////////////////////////////////// 8847 // Much of the following code is similar in shape and spirit to the 8848 // code used in ParNewGC. We should try and share that code 8849 // as much as possible in the future. 8850 8851 #ifndef PRODUCT 8852 // Debugging support for CMSStackOverflowALot 8853 8854 // It's OK to call this multi-threaded; the worst thing 8855 // that can happen is that we'll get a bunch of closely 8856 // spaced simulated oveflows, but that's OK, in fact 8857 // probably good as it would exercise the overflow code 8858 // under contention. 8859 bool CMSCollector::simulate_overflow() { 8860 if (_overflow_counter-- <= 0) { // just being defensive 8861 _overflow_counter = CMSMarkStackOverflowInterval; 8862 return true; 8863 } else { 8864 return false; 8865 } 8866 } 8867 8868 bool CMSCollector::par_simulate_overflow() { 8869 return simulate_overflow(); 8870 } 8871 #endif 8872 8873 // Single-threaded 8874 bool CMSCollector::take_from_overflow_list(size_t num, CMSMarkStack* stack) { 8875 assert(stack->isEmpty(), "Expected precondition"); 8876 assert(stack->capacity() > num, "Shouldn't bite more than can chew"); 8877 size_t i = num; 8878 oop cur = _overflow_list; 8879 const markOop proto = markOopDesc::prototype(); 8880 NOT_PRODUCT(ssize_t n = 0;) 8881 for (oop next; i > 0 && cur != NULL; cur = next, i--) { 8882 next = oop(cur->mark()); 8883 cur->set_mark(proto); // until proven otherwise 8884 assert(cur->is_oop(), "Should be an oop"); 8885 bool res = stack->push(cur); 8886 assert(res, "Bit off more than can chew?"); 8887 NOT_PRODUCT(n++;) 8888 } 8889 _overflow_list = cur; 8890 #ifndef PRODUCT 8891 assert(_num_par_pushes >= n, "Too many pops?"); 8892 _num_par_pushes -=n; 8893 #endif 8894 return !stack->isEmpty(); 8895 } 8896 8897 #define BUSY (oop(0x1aff1aff)) 8898 // (MT-safe) Get a prefix of at most "num" from the list. 8899 // The overflow list is chained through the mark word of 8900 // each object in the list. We fetch the entire list, 8901 // break off a prefix of the right size and return the 8902 // remainder. If other threads try to take objects from 8903 // the overflow list at that time, they will wait for 8904 // some time to see if data becomes available. If (and 8905 // only if) another thread places one or more object(s) 8906 // on the global list before we have returned the suffix 8907 // to the global list, we will walk down our local list 8908 // to find its end and append the global list to 8909 // our suffix before returning it. This suffix walk can 8910 // prove to be expensive (quadratic in the amount of traffic) 8911 // when there are many objects in the overflow list and 8912 // there is much producer-consumer contention on the list. 8913 // *NOTE*: The overflow list manipulation code here and 8914 // in ParNewGeneration:: are very similar in shape, 8915 // except that in the ParNew case we use the old (from/eden) 8916 // copy of the object to thread the list via its klass word. 8917 // Because of the common code, if you make any changes in 8918 // the code below, please check the ParNew version to see if 8919 // similar changes might be needed. 8920 // CR 6797058 has been filed to consolidate the common code. 8921 bool CMSCollector::par_take_from_overflow_list(size_t num, 8922 OopTaskQueue* work_q, 8923 int no_of_gc_threads) { 8924 assert(work_q->size() == 0, "First empty local work queue"); 8925 assert(num < work_q->max_elems(), "Can't bite more than we can chew"); 8926 if (_overflow_list == NULL) { 8927 return false; 8928 } 8929 // Grab the entire list; we'll put back a suffix 8930 oop prefix = (oop)Atomic::xchg_ptr(BUSY, &_overflow_list); 8931 Thread* tid = Thread::current(); 8932 // Before "no_of_gc_threads" was introduced CMSOverflowSpinCount was 8933 // set to ParallelGCThreads. 8934 size_t CMSOverflowSpinCount = (size_t) no_of_gc_threads; // was ParallelGCThreads; 8935 size_t sleep_time_millis = MAX2((size_t)1, num/100); 8936 // If the list is busy, we spin for a short while, 8937 // sleeping between attempts to get the list. 8938 for (size_t spin = 0; prefix == BUSY && spin < CMSOverflowSpinCount; spin++) { 8939 os::sleep(tid, sleep_time_millis, false); 8940 if (_overflow_list == NULL) { 8941 // Nothing left to take 8942 return false; 8943 } else if (_overflow_list != BUSY) { 8944 // Try and grab the prefix 8945 prefix = (oop)Atomic::xchg_ptr(BUSY, &_overflow_list); 8946 } 8947 } 8948 // If the list was found to be empty, or we spun long 8949 // enough, we give up and return empty-handed. If we leave 8950 // the list in the BUSY state below, it must be the case that 8951 // some other thread holds the overflow list and will set it 8952 // to a non-BUSY state in the future. 8953 if (prefix == NULL || prefix == BUSY) { 8954 // Nothing to take or waited long enough 8955 if (prefix == NULL) { 8956 // Write back the NULL in case we overwrote it with BUSY above 8957 // and it is still the same value. 8958 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY); 8959 } 8960 return false; 8961 } 8962 assert(prefix != NULL && prefix != BUSY, "Error"); 8963 size_t i = num; 8964 oop cur = prefix; 8965 // Walk down the first "num" objects, unless we reach the end. 8966 for (; i > 1 && cur->mark() != NULL; cur = oop(cur->mark()), i--); 8967 if (cur->mark() == NULL) { 8968 // We have "num" or fewer elements in the list, so there 8969 // is nothing to return to the global list. 8970 // Write back the NULL in lieu of the BUSY we wrote 8971 // above, if it is still the same value. 8972 if (_overflow_list == BUSY) { 8973 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY); 8974 } 8975 } else { 8976 // Chop off the suffix and rerturn it to the global list. 8977 assert(cur->mark() != BUSY, "Error"); 8978 oop suffix_head = cur->mark(); // suffix will be put back on global list 8979 cur->set_mark(NULL); // break off suffix 8980 // It's possible that the list is still in the empty(busy) state 8981 // we left it in a short while ago; in that case we may be 8982 // able to place back the suffix without incurring the cost 8983 // of a walk down the list. 8984 oop observed_overflow_list = _overflow_list; 8985 oop cur_overflow_list = observed_overflow_list; 8986 bool attached = false; 8987 while (observed_overflow_list == BUSY || observed_overflow_list == NULL) { 8988 observed_overflow_list = 8989 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list); 8990 if (cur_overflow_list == observed_overflow_list) { 8991 attached = true; 8992 break; 8993 } else cur_overflow_list = observed_overflow_list; 8994 } 8995 if (!attached) { 8996 // Too bad, someone else sneaked in (at least) an element; we'll need 8997 // to do a splice. Find tail of suffix so we can prepend suffix to global 8998 // list. 8999 for (cur = suffix_head; cur->mark() != NULL; cur = (oop)(cur->mark())); 9000 oop suffix_tail = cur; 9001 assert(suffix_tail != NULL && suffix_tail->mark() == NULL, 9002 "Tautology"); 9003 observed_overflow_list = _overflow_list; 9004 do { 9005 cur_overflow_list = observed_overflow_list; 9006 if (cur_overflow_list != BUSY) { 9007 // Do the splice ... 9008 suffix_tail->set_mark(markOop(cur_overflow_list)); 9009 } else { // cur_overflow_list == BUSY 9010 suffix_tail->set_mark(NULL); 9011 } 9012 // ... and try to place spliced list back on overflow_list ... 9013 observed_overflow_list = 9014 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list); 9015 } while (cur_overflow_list != observed_overflow_list); 9016 // ... until we have succeeded in doing so. 9017 } 9018 } 9019 9020 // Push the prefix elements on work_q 9021 assert(prefix != NULL, "control point invariant"); 9022 const markOop proto = markOopDesc::prototype(); 9023 oop next; 9024 NOT_PRODUCT(ssize_t n = 0;) 9025 for (cur = prefix; cur != NULL; cur = next) { 9026 next = oop(cur->mark()); 9027 cur->set_mark(proto); // until proven otherwise 9028 assert(cur->is_oop(), "Should be an oop"); 9029 bool res = work_q->push(cur); 9030 assert(res, "Bit off more than we can chew?"); 9031 NOT_PRODUCT(n++;) 9032 } 9033 #ifndef PRODUCT 9034 assert(_num_par_pushes >= n, "Too many pops?"); 9035 Atomic::add_ptr(-(intptr_t)n, &_num_par_pushes); 9036 #endif 9037 return true; 9038 } 9039 9040 // Single-threaded 9041 void CMSCollector::push_on_overflow_list(oop p) { 9042 NOT_PRODUCT(_num_par_pushes++;) 9043 assert(p->is_oop(), "Not an oop"); 9044 preserve_mark_if_necessary(p); 9045 p->set_mark((markOop)_overflow_list); 9046 _overflow_list = p; 9047 } 9048 9049 // Multi-threaded; use CAS to prepend to overflow list 9050 void CMSCollector::par_push_on_overflow_list(oop p) { 9051 NOT_PRODUCT(Atomic::inc_ptr(&_num_par_pushes);) 9052 assert(p->is_oop(), "Not an oop"); 9053 par_preserve_mark_if_necessary(p); 9054 oop observed_overflow_list = _overflow_list; 9055 oop cur_overflow_list; 9056 do { 9057 cur_overflow_list = observed_overflow_list; 9058 if (cur_overflow_list != BUSY) { 9059 p->set_mark(markOop(cur_overflow_list)); 9060 } else { 9061 p->set_mark(NULL); 9062 } 9063 observed_overflow_list = 9064 (oop) Atomic::cmpxchg_ptr(p, &_overflow_list, cur_overflow_list); 9065 } while (cur_overflow_list != observed_overflow_list); 9066 } 9067 #undef BUSY 9068 9069 // Single threaded 9070 // General Note on GrowableArray: pushes may silently fail 9071 // because we are (temporarily) out of C-heap for expanding 9072 // the stack. The problem is quite ubiquitous and affects 9073 // a lot of code in the JVM. The prudent thing for GrowableArray 9074 // to do (for now) is to exit with an error. However, that may 9075 // be too draconian in some cases because the caller may be 9076 // able to recover without much harm. For such cases, we 9077 // should probably introduce a "soft_push" method which returns 9078 // an indication of success or failure with the assumption that 9079 // the caller may be able to recover from a failure; code in 9080 // the VM can then be changed, incrementally, to deal with such 9081 // failures where possible, thus, incrementally hardening the VM 9082 // in such low resource situations. 9083 void CMSCollector::preserve_mark_work(oop p, markOop m) { 9084 _preserved_oop_stack.push(p); 9085 _preserved_mark_stack.push(m); 9086 assert(m == p->mark(), "Mark word changed"); 9087 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), 9088 "bijection"); 9089 } 9090 9091 // Single threaded 9092 void CMSCollector::preserve_mark_if_necessary(oop p) { 9093 markOop m = p->mark(); 9094 if (m->must_be_preserved(p)) { 9095 preserve_mark_work(p, m); 9096 } 9097 } 9098 9099 void CMSCollector::par_preserve_mark_if_necessary(oop p) { 9100 markOop m = p->mark(); 9101 if (m->must_be_preserved(p)) { 9102 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag); 9103 // Even though we read the mark word without holding 9104 // the lock, we are assured that it will not change 9105 // because we "own" this oop, so no other thread can 9106 // be trying to push it on the overflow list; see 9107 // the assertion in preserve_mark_work() that checks 9108 // that m == p->mark(). 9109 preserve_mark_work(p, m); 9110 } 9111 } 9112 9113 // We should be able to do this multi-threaded, 9114 // a chunk of stack being a task (this is 9115 // correct because each oop only ever appears 9116 // once in the overflow list. However, it's 9117 // not very easy to completely overlap this with 9118 // other operations, so will generally not be done 9119 // until all work's been completed. Because we 9120 // expect the preserved oop stack (set) to be small, 9121 // it's probably fine to do this single-threaded. 9122 // We can explore cleverer concurrent/overlapped/parallel 9123 // processing of preserved marks if we feel the 9124 // need for this in the future. Stack overflow should 9125 // be so rare in practice and, when it happens, its 9126 // effect on performance so great that this will 9127 // likely just be in the noise anyway. 9128 void CMSCollector::restore_preserved_marks_if_any() { 9129 assert(SafepointSynchronize::is_at_safepoint(), 9130 "world should be stopped"); 9131 assert(Thread::current()->is_ConcurrentGC_thread() || 9132 Thread::current()->is_VM_thread(), 9133 "should be single-threaded"); 9134 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), 9135 "bijection"); 9136 9137 while (!_preserved_oop_stack.is_empty()) { 9138 oop p = _preserved_oop_stack.pop(); 9139 assert(p->is_oop(), "Should be an oop"); 9140 assert(_span.contains(p), "oop should be in _span"); 9141 assert(p->mark() == markOopDesc::prototype(), 9142 "Set when taken from overflow list"); 9143 markOop m = _preserved_mark_stack.pop(); 9144 p->set_mark(m); 9145 } 9146 assert(_preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(), 9147 "stacks were cleared above"); 9148 } 9149 9150 #ifndef PRODUCT 9151 bool CMSCollector::no_preserved_marks() const { 9152 return _preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(); 9153 } 9154 #endif 9155 9156 CMSAdaptiveSizePolicy* ASConcurrentMarkSweepGeneration::cms_size_policy() const 9157 { 9158 GenCollectedHeap* gch = (GenCollectedHeap*) GenCollectedHeap::heap(); 9159 CMSAdaptiveSizePolicy* size_policy = 9160 (CMSAdaptiveSizePolicy*) gch->gen_policy()->size_policy(); 9161 assert(size_policy->is_gc_cms_adaptive_size_policy(), 9162 "Wrong type for size policy"); 9163 return size_policy; 9164 } 9165 9166 void ASConcurrentMarkSweepGeneration::resize(size_t cur_promo_size, 9167 size_t desired_promo_size) { 9168 if (cur_promo_size < desired_promo_size) { 9169 size_t expand_bytes = desired_promo_size - cur_promo_size; 9170 if (PrintAdaptiveSizePolicy && Verbose) { 9171 gclog_or_tty->print_cr(" ASConcurrentMarkSweepGeneration::resize " 9172 "Expanding tenured generation by " SIZE_FORMAT " (bytes)", 9173 expand_bytes); 9174 } 9175 expand(expand_bytes, 9176 MinHeapDeltaBytes, 9177 CMSExpansionCause::_adaptive_size_policy); 9178 } else if (desired_promo_size < cur_promo_size) { 9179 size_t shrink_bytes = cur_promo_size - desired_promo_size; 9180 if (PrintAdaptiveSizePolicy && Verbose) { 9181 gclog_or_tty->print_cr(" ASConcurrentMarkSweepGeneration::resize " 9182 "Shrinking tenured generation by " SIZE_FORMAT " (bytes)", 9183 shrink_bytes); 9184 } 9185 shrink(shrink_bytes); 9186 } 9187 } 9188 9189 CMSGCAdaptivePolicyCounters* ASConcurrentMarkSweepGeneration::gc_adaptive_policy_counters() { 9190 GenCollectedHeap* gch = GenCollectedHeap::heap(); 9191 CMSGCAdaptivePolicyCounters* counters = 9192 (CMSGCAdaptivePolicyCounters*) gch->collector_policy()->counters(); 9193 assert(counters->kind() == GCPolicyCounters::CMSGCAdaptivePolicyCountersKind, 9194 "Wrong kind of counters"); 9195 return counters; 9196 } 9197 9198 9199 void ASConcurrentMarkSweepGeneration::update_counters() { 9200 if (UsePerfData) { 9201 _space_counters->update_all(); 9202 _gen_counters->update_all(); 9203 CMSGCAdaptivePolicyCounters* counters = gc_adaptive_policy_counters(); 9204 GenCollectedHeap* gch = GenCollectedHeap::heap(); 9205 CMSGCStats* gc_stats_l = (CMSGCStats*) gc_stats(); 9206 assert(gc_stats_l->kind() == GCStats::CMSGCStatsKind, 9207 "Wrong gc statistics type"); 9208 counters->update_counters(gc_stats_l); 9209 } 9210 } 9211 9212 void ASConcurrentMarkSweepGeneration::update_counters(size_t used) { 9213 if (UsePerfData) { 9214 _space_counters->update_used(used); 9215 _space_counters->update_capacity(); 9216 _gen_counters->update_all(); 9217 9218 CMSGCAdaptivePolicyCounters* counters = gc_adaptive_policy_counters(); 9219 GenCollectedHeap* gch = GenCollectedHeap::heap(); 9220 CMSGCStats* gc_stats_l = (CMSGCStats*) gc_stats(); 9221 assert(gc_stats_l->kind() == GCStats::CMSGCStatsKind, 9222 "Wrong gc statistics type"); 9223 counters->update_counters(gc_stats_l); 9224 } 9225 } 9226 9227 // The desired expansion delta is computed so that: 9228 // . desired free percentage or greater is used 9229 void ASConcurrentMarkSweepGeneration::compute_new_size() { 9230 assert_locked_or_safepoint(Heap_lock); 9231 9232 GenCollectedHeap* gch = (GenCollectedHeap*) GenCollectedHeap::heap(); 9233 9234 // If incremental collection failed, we just want to expand 9235 // to the limit. 9236 if (incremental_collection_failed()) { 9237 clear_incremental_collection_failed(); 9238 grow_to_reserved(); 9239 return; 9240 } 9241 9242 assert(UseAdaptiveSizePolicy, "Should be using adaptive sizing"); 9243 9244 assert(gch->kind() == CollectedHeap::GenCollectedHeap, 9245 "Wrong type of heap"); 9246 int prev_level = level() - 1; 9247 assert(prev_level >= 0, "The cms generation is the lowest generation"); 9248 Generation* prev_gen = gch->get_gen(prev_level); 9249 assert(prev_gen->kind() == Generation::ASParNew, 9250 "Wrong type of young generation"); 9251 ParNewGeneration* younger_gen = (ParNewGeneration*) prev_gen; 9252 size_t cur_eden = younger_gen->eden()->capacity(); 9253 CMSAdaptiveSizePolicy* size_policy = cms_size_policy(); 9254 size_t cur_promo = free(); 9255 size_policy->compute_tenured_generation_free_space(cur_promo, 9256 max_available(), 9257 cur_eden); 9258 resize(cur_promo, size_policy->promo_size()); 9259 9260 // Record the new size of the space in the cms generation 9261 // that is available for promotions. This is temporary. 9262 // It should be the desired promo size. 9263 size_policy->avg_cms_promo()->sample(free()); 9264 size_policy->avg_old_live()->sample(used()); 9265 9266 if (UsePerfData) { 9267 CMSGCAdaptivePolicyCounters* counters = gc_adaptive_policy_counters(); 9268 counters->update_cms_capacity_counter(capacity()); 9269 } 9270 } 9271 9272 void ASConcurrentMarkSweepGeneration::shrink_by(size_t desired_bytes) { 9273 assert_locked_or_safepoint(Heap_lock); 9274 assert_lock_strong(freelistLock()); 9275 HeapWord* old_end = _cmsSpace->end(); 9276 HeapWord* unallocated_start = _cmsSpace->unallocated_block(); 9277 assert(old_end >= unallocated_start, "Miscalculation of unallocated_start"); 9278 FreeChunk* chunk_at_end = find_chunk_at_end(); 9279 if (chunk_at_end == NULL) { 9280 // No room to shrink 9281 if (PrintGCDetails && Verbose) { 9282 gclog_or_tty->print_cr("No room to shrink: old_end " 9283 PTR_FORMAT " unallocated_start " PTR_FORMAT 9284 " chunk_at_end " PTR_FORMAT, 9285 old_end, unallocated_start, chunk_at_end); 9286 } 9287 return; 9288 } else { 9289 9290 // Find the chunk at the end of the space and determine 9291 // how much it can be shrunk. 9292 size_t shrinkable_size_in_bytes = chunk_at_end->size(); 9293 size_t aligned_shrinkable_size_in_bytes = 9294 align_size_down(shrinkable_size_in_bytes, os::vm_page_size()); 9295 assert(unallocated_start <= chunk_at_end->end(), 9296 "Inconsistent chunk at end of space"); 9297 size_t bytes = MIN2(desired_bytes, aligned_shrinkable_size_in_bytes); 9298 size_t word_size_before = heap_word_size(_virtual_space.committed_size()); 9299 9300 // Shrink the underlying space 9301 _virtual_space.shrink_by(bytes); 9302 if (PrintGCDetails && Verbose) { 9303 gclog_or_tty->print_cr("ConcurrentMarkSweepGeneration::shrink_by:" 9304 " desired_bytes " SIZE_FORMAT 9305 " shrinkable_size_in_bytes " SIZE_FORMAT 9306 " aligned_shrinkable_size_in_bytes " SIZE_FORMAT 9307 " bytes " SIZE_FORMAT, 9308 desired_bytes, shrinkable_size_in_bytes, 9309 aligned_shrinkable_size_in_bytes, bytes); 9310 gclog_or_tty->print_cr(" old_end " SIZE_FORMAT 9311 " unallocated_start " SIZE_FORMAT, 9312 old_end, unallocated_start); 9313 } 9314 9315 // If the space did shrink (shrinking is not guaranteed), 9316 // shrink the chunk at the end by the appropriate amount. 9317 if (((HeapWord*)_virtual_space.high()) < old_end) { 9318 size_t new_word_size = 9319 heap_word_size(_virtual_space.committed_size()); 9320 9321 // Have to remove the chunk from the dictionary because it is changing 9322 // size and might be someplace elsewhere in the dictionary. 9323 9324 // Get the chunk at end, shrink it, and put it 9325 // back. 9326 _cmsSpace->removeChunkFromDictionary(chunk_at_end); 9327 size_t word_size_change = word_size_before - new_word_size; 9328 size_t chunk_at_end_old_size = chunk_at_end->size(); 9329 assert(chunk_at_end_old_size >= word_size_change, 9330 "Shrink is too large"); 9331 chunk_at_end->set_size(chunk_at_end_old_size - 9332 word_size_change); 9333 _cmsSpace->freed((HeapWord*) chunk_at_end->end(), 9334 word_size_change); 9335 9336 _cmsSpace->returnChunkToDictionary(chunk_at_end); 9337 9338 MemRegion mr(_cmsSpace->bottom(), new_word_size); 9339 _bts->resize(new_word_size); // resize the block offset shared array 9340 Universe::heap()->barrier_set()->resize_covered_region(mr); 9341 _cmsSpace->assert_locked(); 9342 _cmsSpace->set_end((HeapWord*)_virtual_space.high()); 9343 9344 NOT_PRODUCT(_cmsSpace->dictionary()->verify()); 9345 9346 // update the space and generation capacity counters 9347 if (UsePerfData) { 9348 _space_counters->update_capacity(); 9349 _gen_counters->update_all(); 9350 } 9351 9352 if (Verbose && PrintGCDetails) { 9353 size_t new_mem_size = _virtual_space.committed_size(); 9354 size_t old_mem_size = new_mem_size + bytes; 9355 gclog_or_tty->print_cr("Shrinking %s from %ldK by %ldK to %ldK", 9356 name(), old_mem_size/K, bytes/K, new_mem_size/K); 9357 } 9358 } 9359 9360 assert(_cmsSpace->unallocated_block() <= _cmsSpace->end(), 9361 "Inconsistency at end of space"); 9362 assert(chunk_at_end->end() == _cmsSpace->end(), 9363 "Shrinking is inconsistent"); 9364 return; 9365 } 9366 } 9367 9368 // Transfer some number of overflown objects to usual marking 9369 // stack. Return true if some objects were transferred. 9370 bool MarkRefsIntoAndScanClosure::take_from_overflow_list() { 9371 size_t num = MIN2((size_t)(_mark_stack->capacity() - _mark_stack->length())/4, 9372 (size_t)ParGCDesiredObjsFromOverflowList); 9373 9374 bool res = _collector->take_from_overflow_list(num, _mark_stack); 9375 assert(_collector->overflow_list_is_empty() || res, 9376 "If list is not empty, we should have taken something"); 9377 assert(!res || !_mark_stack->isEmpty(), 9378 "If we took something, it should now be on our stack"); 9379 return res; 9380 } 9381 9382 size_t MarkDeadObjectsClosure::do_blk(HeapWord* addr) { 9383 size_t res = _sp->block_size_no_stall(addr, _collector); 9384 if (_sp->block_is_obj(addr)) { 9385 if (_live_bit_map->isMarked(addr)) { 9386 // It can't have been dead in a previous cycle 9387 guarantee(!_dead_bit_map->isMarked(addr), "No resurrection!"); 9388 } else { 9389 _dead_bit_map->mark(addr); // mark the dead object 9390 } 9391 } 9392 // Could be 0, if the block size could not be computed without stalling. 9393 return res; 9394 } 9395 9396 TraceCMSMemoryManagerStats::TraceCMSMemoryManagerStats(CMSCollector::CollectorState phase, GCCause::Cause cause): TraceMemoryManagerStats() { 9397 9398 switch (phase) { 9399 case CMSCollector::InitialMarking: 9400 initialize(true /* fullGC */ , 9401 cause /* cause of the GC */, 9402 true /* recordGCBeginTime */, 9403 true /* recordPreGCUsage */, 9404 false /* recordPeakUsage */, 9405 false /* recordPostGCusage */, 9406 true /* recordAccumulatedGCTime */, 9407 false /* recordGCEndTime */, 9408 false /* countCollection */ ); 9409 break; 9410 9411 case CMSCollector::FinalMarking: 9412 initialize(true /* fullGC */ , 9413 cause /* cause of the GC */, 9414 false /* recordGCBeginTime */, 9415 false /* recordPreGCUsage */, 9416 false /* recordPeakUsage */, 9417 false /* recordPostGCusage */, 9418 true /* recordAccumulatedGCTime */, 9419 false /* recordGCEndTime */, 9420 false /* countCollection */ ); 9421 break; 9422 9423 case CMSCollector::Sweeping: 9424 initialize(true /* fullGC */ , 9425 cause /* cause of the GC */, 9426 false /* recordGCBeginTime */, 9427 false /* recordPreGCUsage */, 9428 true /* recordPeakUsage */, 9429 true /* recordPostGCusage */, 9430 false /* recordAccumulatedGCTime */, 9431 true /* recordGCEndTime */, 9432 true /* countCollection */ ); 9433 break; 9434 9435 default: 9436 ShouldNotReachHere(); 9437 } 9438 } 9439