1 /* 2 * Copyright (c) 2001, 2014, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "classfile/classLoaderData.hpp" 27 #include "classfile/symbolTable.hpp" 28 #include "classfile/systemDictionary.hpp" 29 #include "code/codeCache.hpp" 30 #include "gc_implementation/concurrentMarkSweep/cmsAdaptiveSizePolicy.hpp" 31 #include "gc_implementation/concurrentMarkSweep/cmsCollectorPolicy.hpp" 32 #include "gc_implementation/concurrentMarkSweep/cmsGCAdaptivePolicyCounters.hpp" 33 #include "gc_implementation/concurrentMarkSweep/cmsOopClosures.inline.hpp" 34 #include "gc_implementation/concurrentMarkSweep/compactibleFreeListSpace.hpp" 35 #include "gc_implementation/concurrentMarkSweep/concurrentMarkSweepGeneration.inline.hpp" 36 #include "gc_implementation/concurrentMarkSweep/concurrentMarkSweepThread.hpp" 37 #include "gc_implementation/concurrentMarkSweep/vmCMSOperations.hpp" 38 #include "gc_implementation/parNew/parNewGeneration.hpp" 39 #include "gc_implementation/shared/collectorCounters.hpp" 40 #include "gc_implementation/shared/gcTimer.hpp" 41 #include "gc_implementation/shared/gcTrace.hpp" 42 #include "gc_implementation/shared/gcTraceTime.hpp" 43 #include "gc_implementation/shared/isGCActiveMark.hpp" 44 #include "gc_interface/collectedHeap.inline.hpp" 45 #include "memory/allocation.hpp" 46 #include "memory/cardTableRS.hpp" 47 #include "memory/collectorPolicy.hpp" 48 #include "memory/gcLocker.inline.hpp" 49 #include "memory/genCollectedHeap.hpp" 50 #include "memory/genMarkSweep.hpp" 51 #include "memory/genOopClosures.inline.hpp" 52 #include "memory/iterator.inline.hpp" 53 #include "memory/padded.hpp" 54 #include "memory/referencePolicy.hpp" 55 #include "memory/resourceArea.hpp" 56 #include "memory/tenuredGeneration.hpp" 57 #include "oops/oop.inline.hpp" 58 #include "prims/jvmtiExport.hpp" 59 #include "runtime/globals_extension.hpp" 60 #include "runtime/handles.inline.hpp" 61 #include "runtime/java.hpp" 62 #include "runtime/orderAccess.inline.hpp" 63 #include "runtime/vmThread.hpp" 64 #include "services/memoryService.hpp" 65 #include "services/runtimeService.hpp" 66 67 PRAGMA_FORMAT_MUTE_WARNINGS_FOR_GCC 68 69 // statics 70 CMSCollector* ConcurrentMarkSweepGeneration::_collector = NULL; 71 bool CMSCollector::_full_gc_requested = false; 72 GCCause::Cause CMSCollector::_full_gc_cause = GCCause::_no_gc; 73 74 ////////////////////////////////////////////////////////////////// 75 // In support of CMS/VM thread synchronization 76 ////////////////////////////////////////////////////////////////// 77 // We split use of the CGC_lock into 2 "levels". 78 // The low-level locking is of the usual CGC_lock monitor. We introduce 79 // a higher level "token" (hereafter "CMS token") built on top of the 80 // low level monitor (hereafter "CGC lock"). 81 // The token-passing protocol gives priority to the VM thread. The 82 // CMS-lock doesn't provide any fairness guarantees, but clients 83 // should ensure that it is only held for very short, bounded 84 // durations. 85 // 86 // When either of the CMS thread or the VM thread is involved in 87 // collection operations during which it does not want the other 88 // thread to interfere, it obtains the CMS token. 89 // 90 // If either thread tries to get the token while the other has 91 // it, that thread waits. However, if the VM thread and CMS thread 92 // both want the token, then the VM thread gets priority while the 93 // CMS thread waits. This ensures, for instance, that the "concurrent" 94 // phases of the CMS thread's work do not block out the VM thread 95 // for long periods of time as the CMS thread continues to hog 96 // the token. (See bug 4616232). 97 // 98 // The baton-passing functions are, however, controlled by the 99 // flags _foregroundGCShouldWait and _foregroundGCIsActive, 100 // and here the low-level CMS lock, not the high level token, 101 // ensures mutual exclusion. 102 // 103 // Two important conditions that we have to satisfy: 104 // 1. if a thread does a low-level wait on the CMS lock, then it 105 // relinquishes the CMS token if it were holding that token 106 // when it acquired the low-level CMS lock. 107 // 2. any low-level notifications on the low-level lock 108 // should only be sent when a thread has relinquished the token. 109 // 110 // In the absence of either property, we'd have potential deadlock. 111 // 112 // We protect each of the CMS (concurrent and sequential) phases 113 // with the CMS _token_, not the CMS _lock_. 114 // 115 // The only code protected by CMS lock is the token acquisition code 116 // itself, see ConcurrentMarkSweepThread::[de]synchronize(), and the 117 // baton-passing code. 118 // 119 // Unfortunately, i couldn't come up with a good abstraction to factor and 120 // hide the naked CGC_lock manipulation in the baton-passing code 121 // further below. That's something we should try to do. Also, the proof 122 // of correctness of this 2-level locking scheme is far from obvious, 123 // and potentially quite slippery. We have an uneasy supsicion, for instance, 124 // that there may be a theoretical possibility of delay/starvation in the 125 // low-level lock/wait/notify scheme used for the baton-passing because of 126 // potential intereference with the priority scheme embodied in the 127 // CMS-token-passing protocol. See related comments at a CGC_lock->wait() 128 // invocation further below and marked with "XXX 20011219YSR". 129 // Indeed, as we note elsewhere, this may become yet more slippery 130 // in the presence of multiple CMS and/or multiple VM threads. XXX 131 132 class CMSTokenSync: public StackObj { 133 private: 134 bool _is_cms_thread; 135 public: 136 CMSTokenSync(bool is_cms_thread): 137 _is_cms_thread(is_cms_thread) { 138 assert(is_cms_thread == Thread::current()->is_ConcurrentGC_thread(), 139 "Incorrect argument to constructor"); 140 ConcurrentMarkSweepThread::synchronize(_is_cms_thread); 141 } 142 143 ~CMSTokenSync() { 144 assert(_is_cms_thread ? 145 ConcurrentMarkSweepThread::cms_thread_has_cms_token() : 146 ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 147 "Incorrect state"); 148 ConcurrentMarkSweepThread::desynchronize(_is_cms_thread); 149 } 150 }; 151 152 // Convenience class that does a CMSTokenSync, and then acquires 153 // upto three locks. 154 class CMSTokenSyncWithLocks: public CMSTokenSync { 155 private: 156 // Note: locks are acquired in textual declaration order 157 // and released in the opposite order 158 MutexLockerEx _locker1, _locker2, _locker3; 159 public: 160 CMSTokenSyncWithLocks(bool is_cms_thread, Mutex* mutex1, 161 Mutex* mutex2 = NULL, Mutex* mutex3 = NULL): 162 CMSTokenSync(is_cms_thread), 163 _locker1(mutex1, Mutex::_no_safepoint_check_flag), 164 _locker2(mutex2, Mutex::_no_safepoint_check_flag), 165 _locker3(mutex3, Mutex::_no_safepoint_check_flag) 166 { } 167 }; 168 169 170 // Wrapper class to temporarily disable icms during a foreground cms collection. 171 class ICMSDisabler: public StackObj { 172 public: 173 // The ctor disables icms and wakes up the thread so it notices the change; 174 // the dtor re-enables icms. Note that the CMSCollector methods will check 175 // CMSIncrementalMode. 176 ICMSDisabler() { CMSCollector::disable_icms(); CMSCollector::start_icms(); } 177 ~ICMSDisabler() { CMSCollector::enable_icms(); } 178 }; 179 180 ////////////////////////////////////////////////////////////////// 181 // Concurrent Mark-Sweep Generation ///////////////////////////// 182 ////////////////////////////////////////////////////////////////// 183 184 NOT_PRODUCT(CompactibleFreeListSpace* debug_cms_space;) 185 186 // This struct contains per-thread things necessary to support parallel 187 // young-gen collection. 188 class CMSParGCThreadState: public CHeapObj<mtGC> { 189 public: 190 CFLS_LAB lab; 191 PromotionInfo promo; 192 193 // Constructor. 194 CMSParGCThreadState(CompactibleFreeListSpace* cfls) : lab(cfls) { 195 promo.setSpace(cfls); 196 } 197 }; 198 199 ConcurrentMarkSweepGeneration::ConcurrentMarkSweepGeneration( 200 ReservedSpace rs, size_t initial_byte_size, int level, 201 CardTableRS* ct, bool use_adaptive_freelists, 202 FreeBlockDictionary<FreeChunk>::DictionaryChoice dictionaryChoice) : 203 CardGeneration(rs, initial_byte_size, level, ct), 204 _dilatation_factor(((double)MinChunkSize)/((double)(CollectedHeap::min_fill_size()))), 205 _debug_collection_type(Concurrent_collection_type), 206 _did_compact(false) 207 { 208 HeapWord* bottom = (HeapWord*) _virtual_space.low(); 209 HeapWord* end = (HeapWord*) _virtual_space.high(); 210 211 _direct_allocated_words = 0; 212 NOT_PRODUCT( 213 _numObjectsPromoted = 0; 214 _numWordsPromoted = 0; 215 _numObjectsAllocated = 0; 216 _numWordsAllocated = 0; 217 ) 218 219 _cmsSpace = new CompactibleFreeListSpace(_bts, MemRegion(bottom, end), 220 use_adaptive_freelists, 221 dictionaryChoice); 222 NOT_PRODUCT(debug_cms_space = _cmsSpace;) 223 if (_cmsSpace == NULL) { 224 vm_exit_during_initialization( 225 "CompactibleFreeListSpace allocation failure"); 226 } 227 _cmsSpace->_gen = this; 228 229 _gc_stats = new CMSGCStats(); 230 231 // Verify the assumption that FreeChunk::_prev and OopDesc::_klass 232 // offsets match. The ability to tell free chunks from objects 233 // depends on this property. 234 debug_only( 235 FreeChunk* junk = NULL; 236 assert(UseCompressedClassPointers || 237 junk->prev_addr() == (void*)(oop(junk)->klass_addr()), 238 "Offset of FreeChunk::_prev within FreeChunk must match" 239 " that of OopDesc::_klass within OopDesc"); 240 ) 241 if (CollectedHeap::use_parallel_gc_threads()) { 242 typedef CMSParGCThreadState* CMSParGCThreadStatePtr; 243 _par_gc_thread_states = 244 NEW_C_HEAP_ARRAY(CMSParGCThreadStatePtr, ParallelGCThreads, mtGC); 245 if (_par_gc_thread_states == NULL) { 246 vm_exit_during_initialization("Could not allocate par gc structs"); 247 } 248 for (uint i = 0; i < ParallelGCThreads; i++) { 249 _par_gc_thread_states[i] = new CMSParGCThreadState(cmsSpace()); 250 if (_par_gc_thread_states[i] == NULL) { 251 vm_exit_during_initialization("Could not allocate par gc structs"); 252 } 253 } 254 } else { 255 _par_gc_thread_states = NULL; 256 } 257 _incremental_collection_failed = false; 258 // The "dilatation_factor" is the expansion that can occur on 259 // account of the fact that the minimum object size in the CMS 260 // generation may be larger than that in, say, a contiguous young 261 // generation. 262 // Ideally, in the calculation below, we'd compute the dilatation 263 // factor as: MinChunkSize/(promoting_gen's min object size) 264 // Since we do not have such a general query interface for the 265 // promoting generation, we'll instead just use the mimimum 266 // object size (which today is a header's worth of space); 267 // note that all arithmetic is in units of HeapWords. 268 assert(MinChunkSize >= CollectedHeap::min_fill_size(), "just checking"); 269 assert(_dilatation_factor >= 1.0, "from previous assert"); 270 } 271 272 273 // The field "_initiating_occupancy" represents the occupancy percentage 274 // at which we trigger a new collection cycle. Unless explicitly specified 275 // via CMSInitiatingOccupancyFraction (argument "io" below), it 276 // is calculated by: 277 // 278 // Let "f" be MinHeapFreeRatio in 279 // 280 // _intiating_occupancy = 100-f + 281 // f * (CMSTriggerRatio/100) 282 // where CMSTriggerRatio is the argument "tr" below. 283 // 284 // That is, if we assume the heap is at its desired maximum occupancy at the 285 // end of a collection, we let CMSTriggerRatio of the (purported) free 286 // space be allocated before initiating a new collection cycle. 287 // 288 void ConcurrentMarkSweepGeneration::init_initiating_occupancy(intx io, uintx tr) { 289 assert(io <= 100 && tr <= 100, "Check the arguments"); 290 if (io >= 0) { 291 _initiating_occupancy = (double)io / 100.0; 292 } else { 293 _initiating_occupancy = ((100 - MinHeapFreeRatio) + 294 (double)(tr * MinHeapFreeRatio) / 100.0) 295 / 100.0; 296 } 297 } 298 299 void ConcurrentMarkSweepGeneration::ref_processor_init() { 300 assert(collector() != NULL, "no collector"); 301 collector()->ref_processor_init(); 302 } 303 304 void CMSCollector::ref_processor_init() { 305 if (_ref_processor == NULL) { 306 // Allocate and initialize a reference processor 307 _ref_processor = 308 new ReferenceProcessor(_span, // span 309 (ParallelGCThreads > 1) && ParallelRefProcEnabled, // mt processing 310 (int) ParallelGCThreads, // mt processing degree 311 _cmsGen->refs_discovery_is_mt(), // mt discovery 312 (int) MAX2(ConcGCThreads, ParallelGCThreads), // mt discovery degree 313 _cmsGen->refs_discovery_is_atomic(), // discovery is not atomic 314 &_is_alive_closure); // closure for liveness info 315 // Initialize the _ref_processor field of CMSGen 316 _cmsGen->set_ref_processor(_ref_processor); 317 318 } 319 } 320 321 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 CardTableRS* ct, 555 ConcurrentMarkSweepPolicy* cp): 556 _cmsGen(cmsGen), 557 _ct(ct), 558 _ref_processor(NULL), // will be set later 559 _conc_workers(NULL), // may be set later 560 _abort_preclean(false), 561 _start_sampling(false), 562 _between_prologue_and_epilogue(false), 563 _markBitMap(0, Mutex::leaf + 1, "CMS_markBitMap_lock"), 564 _modUnionTable((CardTableModRefBS::card_shift - LogHeapWordSize), 565 -1 /* lock-free */, "No_lock" /* dummy */), 566 _modUnionClosure(&_modUnionTable), 567 _modUnionClosurePar(&_modUnionTable), 568 // Adjust my span to cover old (cms) gen 569 _span(cmsGen->reserved()), 570 // Construct the is_alive_closure with _span & markBitMap 571 _is_alive_closure(_span, &_markBitMap), 572 _restart_addr(NULL), 573 _overflow_list(NULL), 574 _stats(cmsGen), 575 _eden_chunk_lock(new Mutex(Mutex::leaf + 1, "CMS_eden_chunk_lock", true)), 576 _eden_chunk_array(NULL), // may be set in ctor body 577 _eden_chunk_capacity(0), // -- ditto -- 578 _eden_chunk_index(0), // -- ditto -- 579 _survivor_plab_array(NULL), // -- ditto -- 580 _survivor_chunk_array(NULL), // -- ditto -- 581 _survivor_chunk_capacity(0), // -- ditto -- 582 _survivor_chunk_index(0), // -- ditto -- 583 _ser_pmc_preclean_ovflw(0), 584 _ser_kac_preclean_ovflw(0), 585 _ser_pmc_remark_ovflw(0), 586 _par_pmc_remark_ovflw(0), 587 _ser_kac_ovflw(0), 588 _par_kac_ovflw(0), 589 #ifndef PRODUCT 590 _num_par_pushes(0), 591 #endif 592 _collection_count_start(0), 593 _verifying(false), 594 _icms_start_limit(NULL), 595 _icms_stop_limit(NULL), 596 _verification_mark_bm(0, Mutex::leaf + 1, "CMS_verification_mark_bm_lock"), 597 _completed_initialization(false), 598 _collector_policy(cp), 599 _should_unload_classes(CMSClassUnloadingEnabled), 600 _concurrent_cycles_since_last_unload(0), 601 _roots_scanning_options(GenCollectedHeap::SO_None), 602 _inter_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding), 603 _intra_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding), 604 _gc_tracer_cm(new (ResourceObj::C_HEAP, mtGC) CMSTracer()), 605 _gc_timer_cm(new (ResourceObj::C_HEAP, mtGC) ConcurrentGCTimer()), 606 _cms_start_registered(false) 607 { 608 if (ExplicitGCInvokesConcurrentAndUnloadsClasses) { 609 ExplicitGCInvokesConcurrent = true; 610 } 611 // Now expand the span and allocate the collection support structures 612 // (MUT, marking bit map etc.) to cover both generations subject to 613 // collection. 614 615 // For use by dirty card to oop closures. 616 _cmsGen->cmsSpace()->set_collector(this); 617 618 // Allocate MUT and marking bit map 619 { 620 MutexLockerEx x(_markBitMap.lock(), Mutex::_no_safepoint_check_flag); 621 if (!_markBitMap.allocate(_span)) { 622 warning("Failed to allocate CMS Bit Map"); 623 return; 624 } 625 assert(_markBitMap.covers(_span), "_markBitMap inconsistency?"); 626 } 627 { 628 _modUnionTable.allocate(_span); 629 assert(_modUnionTable.covers(_span), "_modUnionTable inconsistency?"); 630 } 631 632 if (!_markStack.allocate(MarkStackSize)) { 633 warning("Failed to allocate CMS Marking Stack"); 634 return; 635 } 636 637 // Support for multi-threaded concurrent phases 638 if (CMSConcurrentMTEnabled) { 639 if (FLAG_IS_DEFAULT(ConcGCThreads)) { 640 // just for now 641 FLAG_SET_DEFAULT(ConcGCThreads, (ParallelGCThreads + 3)/4); 642 } 643 if (ConcGCThreads > 1) { 644 _conc_workers = new YieldingFlexibleWorkGang("Parallel CMS Threads", 645 ConcGCThreads, true); 646 if (_conc_workers == NULL) { 647 warning("GC/CMS: _conc_workers allocation failure: " 648 "forcing -CMSConcurrentMTEnabled"); 649 CMSConcurrentMTEnabled = false; 650 } else { 651 _conc_workers->initialize_workers(); 652 } 653 } else { 654 CMSConcurrentMTEnabled = false; 655 } 656 } 657 if (!CMSConcurrentMTEnabled) { 658 ConcGCThreads = 0; 659 } else { 660 // Turn off CMSCleanOnEnter optimization temporarily for 661 // the MT case where it's not fixed yet; see 6178663. 662 CMSCleanOnEnter = false; 663 } 664 assert((_conc_workers != NULL) == (ConcGCThreads > 1), 665 "Inconsistency"); 666 667 // Parallel task queues; these are shared for the 668 // concurrent and stop-world phases of CMS, but 669 // are not shared with parallel scavenge (ParNew). 670 { 671 uint i; 672 uint num_queues = (uint) MAX2(ParallelGCThreads, ConcGCThreads); 673 674 if ((CMSParallelRemarkEnabled || CMSConcurrentMTEnabled 675 || ParallelRefProcEnabled) 676 && num_queues > 0) { 677 _task_queues = new OopTaskQueueSet(num_queues); 678 if (_task_queues == NULL) { 679 warning("task_queues allocation failure."); 680 return; 681 } 682 _hash_seed = NEW_C_HEAP_ARRAY(int, num_queues, mtGC); 683 if (_hash_seed == NULL) { 684 warning("_hash_seed array allocation failure"); 685 return; 686 } 687 688 typedef Padded<OopTaskQueue> PaddedOopTaskQueue; 689 for (i = 0; i < num_queues; i++) { 690 PaddedOopTaskQueue *q = new PaddedOopTaskQueue(); 691 if (q == NULL) { 692 warning("work_queue allocation failure."); 693 return; 694 } 695 _task_queues->register_queue(i, q); 696 } 697 for (i = 0; i < num_queues; i++) { 698 _task_queues->queue(i)->initialize(); 699 _hash_seed[i] = 17; // copied from ParNew 700 } 701 } 702 } 703 704 _cmsGen ->init_initiating_occupancy(CMSInitiatingOccupancyFraction, CMSTriggerRatio); 705 706 // Clip CMSBootstrapOccupancy between 0 and 100. 707 _bootstrap_occupancy = ((double)CMSBootstrapOccupancy)/(double)100; 708 709 _full_gcs_since_conc_gc = 0; 710 711 // Now tell CMS generations the identity of their collector 712 ConcurrentMarkSweepGeneration::set_collector(this); 713 714 // Create & start a CMS thread for this CMS collector 715 _cmsThread = ConcurrentMarkSweepThread::start(this); 716 assert(cmsThread() != NULL, "CMS Thread should have been created"); 717 assert(cmsThread()->collector() == this, 718 "CMS Thread should refer to this gen"); 719 assert(CGC_lock != NULL, "Where's the CGC_lock?"); 720 721 // Support for parallelizing young gen rescan 722 GenCollectedHeap* gch = GenCollectedHeap::heap(); 723 _young_gen = gch->prev_gen(_cmsGen); 724 if (gch->supports_inline_contig_alloc()) { 725 _top_addr = gch->top_addr(); 726 _end_addr = gch->end_addr(); 727 assert(_young_gen != NULL, "no _young_gen"); 728 _eden_chunk_index = 0; 729 _eden_chunk_capacity = (_young_gen->max_capacity()+CMSSamplingGrain)/CMSSamplingGrain; 730 _eden_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, _eden_chunk_capacity, mtGC); 731 if (_eden_chunk_array == NULL) { 732 _eden_chunk_capacity = 0; 733 warning("GC/CMS: _eden_chunk_array allocation failure"); 734 } 735 } 736 assert(_eden_chunk_array != NULL || _eden_chunk_capacity == 0, "Error"); 737 738 // Support for parallelizing survivor space rescan 739 if ((CMSParallelRemarkEnabled && CMSParallelSurvivorRemarkEnabled) || CMSParallelInitialMarkEnabled) { 740 const size_t max_plab_samples = 741 ((DefNewGeneration*)_young_gen)->max_survivor_size() / plab_sample_minimum_size(); 742 743 _survivor_plab_array = NEW_C_HEAP_ARRAY(ChunkArray, ParallelGCThreads, mtGC); 744 _survivor_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, 2*max_plab_samples, mtGC); 745 _cursor = NEW_C_HEAP_ARRAY(size_t, ParallelGCThreads, mtGC); 746 if (_survivor_plab_array == NULL || _survivor_chunk_array == NULL 747 || _cursor == NULL) { 748 warning("Failed to allocate survivor plab/chunk array"); 749 if (_survivor_plab_array != NULL) { 750 FREE_C_HEAP_ARRAY(ChunkArray, _survivor_plab_array, mtGC); 751 _survivor_plab_array = NULL; 752 } 753 if (_survivor_chunk_array != NULL) { 754 FREE_C_HEAP_ARRAY(HeapWord*, _survivor_chunk_array, mtGC); 755 _survivor_chunk_array = NULL; 756 } 757 if (_cursor != NULL) { 758 FREE_C_HEAP_ARRAY(size_t, _cursor, mtGC); 759 _cursor = NULL; 760 } 761 } else { 762 _survivor_chunk_capacity = 2*max_plab_samples; 763 for (uint i = 0; i < ParallelGCThreads; i++) { 764 HeapWord** vec = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples, mtGC); 765 if (vec == NULL) { 766 warning("Failed to allocate survivor plab array"); 767 for (int j = i; j > 0; j--) { 768 FREE_C_HEAP_ARRAY(HeapWord*, _survivor_plab_array[j-1].array(), mtGC); 769 } 770 FREE_C_HEAP_ARRAY(ChunkArray, _survivor_plab_array, mtGC); 771 FREE_C_HEAP_ARRAY(HeapWord*, _survivor_chunk_array, mtGC); 772 _survivor_plab_array = NULL; 773 _survivor_chunk_array = NULL; 774 _survivor_chunk_capacity = 0; 775 break; 776 } else { 777 ChunkArray* cur = 778 ::new (&_survivor_plab_array[i]) ChunkArray(vec, 779 max_plab_samples); 780 assert(cur->end() == 0, "Should be 0"); 781 assert(cur->array() == vec, "Should be vec"); 782 assert(cur->capacity() == max_plab_samples, "Error"); 783 } 784 } 785 } 786 } 787 assert( ( _survivor_plab_array != NULL 788 && _survivor_chunk_array != NULL) 789 || ( _survivor_chunk_capacity == 0 790 && _survivor_chunk_index == 0), 791 "Error"); 792 793 NOT_PRODUCT(_overflow_counter = CMSMarkStackOverflowInterval;) 794 _gc_counters = new CollectorCounters("CMS", 1); 795 _completed_initialization = true; 796 _inter_sweep_timer.start(); // start of time 797 } 798 799 size_t CMSCollector::plab_sample_minimum_size() { 800 // The default value of MinTLABSize is 2k, but there is 801 // no way to get the default value if the flag has been overridden. 802 return MAX2(ThreadLocalAllocBuffer::min_size() * HeapWordSize, 2 * K); 803 } 804 805 const char* ConcurrentMarkSweepGeneration::name() const { 806 return "concurrent mark-sweep generation"; 807 } 808 void ConcurrentMarkSweepGeneration::update_counters() { 809 if (UsePerfData) { 810 _space_counters->update_all(); 811 _gen_counters->update_all(); 812 } 813 } 814 815 // this is an optimized version of update_counters(). it takes the 816 // used value as a parameter rather than computing it. 817 // 818 void ConcurrentMarkSweepGeneration::update_counters(size_t used) { 819 if (UsePerfData) { 820 _space_counters->update_used(used); 821 _space_counters->update_capacity(); 822 _gen_counters->update_all(); 823 } 824 } 825 826 void ConcurrentMarkSweepGeneration::print() const { 827 Generation::print(); 828 cmsSpace()->print(); 829 } 830 831 #ifndef PRODUCT 832 void ConcurrentMarkSweepGeneration::print_statistics() { 833 cmsSpace()->printFLCensus(0); 834 } 835 #endif 836 837 void ConcurrentMarkSweepGeneration::printOccupancy(const char *s) { 838 GenCollectedHeap* gch = GenCollectedHeap::heap(); 839 if (PrintGCDetails) { 840 if (Verbose) { 841 gclog_or_tty->print("[%d %s-%s: "SIZE_FORMAT"("SIZE_FORMAT")]", 842 level(), short_name(), s, used(), capacity()); 843 } else { 844 gclog_or_tty->print("[%d %s-%s: "SIZE_FORMAT"K("SIZE_FORMAT"K)]", 845 level(), short_name(), s, used() / K, capacity() / K); 846 } 847 } 848 if (Verbose) { 849 gclog_or_tty->print(" "SIZE_FORMAT"("SIZE_FORMAT")", 850 gch->used(), gch->capacity()); 851 } else { 852 gclog_or_tty->print(" "SIZE_FORMAT"K("SIZE_FORMAT"K)", 853 gch->used() / K, gch->capacity() / K); 854 } 855 } 856 857 size_t 858 ConcurrentMarkSweepGeneration::contiguous_available() const { 859 // dld proposes an improvement in precision here. If the committed 860 // part of the space ends in a free block we should add that to 861 // uncommitted size in the calculation below. Will make this 862 // change later, staying with the approximation below for the 863 // time being. -- ysr. 864 return MAX2(_virtual_space.uncommitted_size(), unsafe_max_alloc_nogc()); 865 } 866 867 size_t 868 ConcurrentMarkSweepGeneration::unsafe_max_alloc_nogc() const { 869 return _cmsSpace->max_alloc_in_words() * HeapWordSize; 870 } 871 872 size_t ConcurrentMarkSweepGeneration::max_available() const { 873 return free() + _virtual_space.uncommitted_size(); 874 } 875 876 bool ConcurrentMarkSweepGeneration::promotion_attempt_is_safe(size_t max_promotion_in_bytes) const { 877 size_t available = max_available(); 878 size_t av_promo = (size_t)gc_stats()->avg_promoted()->padded_average(); 879 bool res = (available >= av_promo) || (available >= max_promotion_in_bytes); 880 if (Verbose && PrintGCDetails) { 881 gclog_or_tty->print_cr( 882 "CMS: promo attempt is%s safe: available("SIZE_FORMAT") %s av_promo("SIZE_FORMAT")," 883 "max_promo("SIZE_FORMAT")", 884 res? "":" not", available, res? ">=":"<", 885 av_promo, max_promotion_in_bytes); 886 } 887 return res; 888 } 889 890 // At a promotion failure dump information on block layout in heap 891 // (cms old generation). 892 void ConcurrentMarkSweepGeneration::promotion_failure_occurred() { 893 if (CMSDumpAtPromotionFailure) { 894 cmsSpace()->dump_at_safepoint_with_locks(collector(), gclog_or_tty); 895 } 896 } 897 898 CompactibleSpace* 899 ConcurrentMarkSweepGeneration::first_compaction_space() const { 900 return _cmsSpace; 901 } 902 903 void ConcurrentMarkSweepGeneration::reset_after_compaction() { 904 // Clear the promotion information. These pointers can be adjusted 905 // along with all the other pointers into the heap but 906 // compaction is expected to be a rare event with 907 // a heap using cms so don't do it without seeing the need. 908 if (CollectedHeap::use_parallel_gc_threads()) { 909 for (uint i = 0; i < ParallelGCThreads; i++) { 910 _par_gc_thread_states[i]->promo.reset(); 911 } 912 } 913 } 914 915 void ConcurrentMarkSweepGeneration::space_iterate(SpaceClosure* blk, bool usedOnly) { 916 blk->do_space(_cmsSpace); 917 } 918 919 void ConcurrentMarkSweepGeneration::compute_new_size() { 920 assert_locked_or_safepoint(Heap_lock); 921 922 // If incremental collection failed, we just want to expand 923 // to the limit. 924 if (incremental_collection_failed()) { 925 clear_incremental_collection_failed(); 926 grow_to_reserved(); 927 return; 928 } 929 930 // The heap has been compacted but not reset yet. 931 // Any metric such as free() or used() will be incorrect. 932 933 CardGeneration::compute_new_size(); 934 935 // Reset again after a possible resizing 936 if (did_compact()) { 937 cmsSpace()->reset_after_compaction(); 938 } 939 } 940 941 void ConcurrentMarkSweepGeneration::compute_new_size_free_list() { 942 assert_locked_or_safepoint(Heap_lock); 943 944 // If incremental collection failed, we just want to expand 945 // to the limit. 946 if (incremental_collection_failed()) { 947 clear_incremental_collection_failed(); 948 grow_to_reserved(); 949 return; 950 } 951 952 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 size_t expand_bytes = MAX2(desired_capacity - capacity(), MinHeapDeltaBytes); 961 if (PrintGCDetails && Verbose) { 962 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 963 gclog_or_tty->print_cr("\nFrom compute_new_size: "); 964 gclog_or_tty->print_cr(" Free fraction %f", free_percentage); 965 gclog_or_tty->print_cr(" Desired free fraction %f", 966 desired_free_percentage); 967 gclog_or_tty->print_cr(" Maximum free fraction %f", 968 maximum_free_percentage); 969 gclog_or_tty->print_cr(" Capactiy "SIZE_FORMAT, capacity()/1000); 970 gclog_or_tty->print_cr(" Desired capacity "SIZE_FORMAT, 971 desired_capacity/1000); 972 int prev_level = level() - 1; 973 if (prev_level >= 0) { 974 size_t prev_size = 0; 975 GenCollectedHeap* gch = GenCollectedHeap::heap(); 976 Generation* prev_gen = gch->_gens[prev_level]; 977 prev_size = prev_gen->capacity(); 978 gclog_or_tty->print_cr(" Younger gen size "SIZE_FORMAT, 979 prev_size/1000); 980 } 981 gclog_or_tty->print_cr(" unsafe_max_alloc_nogc "SIZE_FORMAT, 982 unsafe_max_alloc_nogc()/1000); 983 gclog_or_tty->print_cr(" contiguous available "SIZE_FORMAT, 984 contiguous_available()/1000); 985 gclog_or_tty->print_cr(" Expand by "SIZE_FORMAT" (bytes)", 986 expand_bytes); 987 } 988 // safe if expansion fails 989 expand(expand_bytes, 0, CMSExpansionCause::_satisfy_free_ratio); 990 if (PrintGCDetails && Verbose) { 991 gclog_or_tty->print_cr(" Expanded free fraction %f", 992 ((double) free()) / capacity()); 993 } 994 } else { 995 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 996 assert(desired_capacity <= capacity(), "invalid expansion size"); 997 size_t shrink_bytes = capacity() - desired_capacity; 998 // Don't shrink unless the delta is greater than the minimum shrink we want 999 if (shrink_bytes >= MinHeapDeltaBytes) { 1000 shrink_free_list_by(shrink_bytes); 1001 } 1002 } 1003 } 1004 1005 Mutex* ConcurrentMarkSweepGeneration::freelistLock() const { 1006 return cmsSpace()->freelistLock(); 1007 } 1008 1009 HeapWord* ConcurrentMarkSweepGeneration::allocate(size_t size, 1010 bool tlab) { 1011 CMSSynchronousYieldRequest yr; 1012 MutexLockerEx x(freelistLock(), 1013 Mutex::_no_safepoint_check_flag); 1014 return have_lock_and_allocate(size, tlab); 1015 } 1016 1017 HeapWord* ConcurrentMarkSweepGeneration::have_lock_and_allocate(size_t size, 1018 bool tlab /* ignored */) { 1019 assert_lock_strong(freelistLock()); 1020 size_t adjustedSize = CompactibleFreeListSpace::adjustObjectSize(size); 1021 HeapWord* res = cmsSpace()->allocate(adjustedSize); 1022 // Allocate the object live (grey) if the background collector has 1023 // started marking. This is necessary because the marker may 1024 // have passed this address and consequently this object will 1025 // not otherwise be greyed and would be incorrectly swept up. 1026 // Note that if this object contains references, the writing 1027 // of those references will dirty the card containing this object 1028 // allowing the object to be blackened (and its references scanned) 1029 // either during a preclean phase or at the final checkpoint. 1030 if (res != NULL) { 1031 // We may block here with an uninitialized object with 1032 // its mark-bit or P-bits not yet set. Such objects need 1033 // to be safely navigable by block_start(). 1034 assert(oop(res)->klass_or_null() == NULL, "Object should be uninitialized here."); 1035 assert(!((FreeChunk*)res)->is_free(), "Error, block will look free but show wrong size"); 1036 collector()->direct_allocated(res, adjustedSize); 1037 _direct_allocated_words += adjustedSize; 1038 // allocation counters 1039 NOT_PRODUCT( 1040 _numObjectsAllocated++; 1041 _numWordsAllocated += (int)adjustedSize; 1042 ) 1043 } 1044 return res; 1045 } 1046 1047 // In the case of direct allocation by mutators in a generation that 1048 // is being concurrently collected, the object must be allocated 1049 // live (grey) if the background collector has started marking. 1050 // This is necessary because the marker may 1051 // have passed this address and consequently this object will 1052 // not otherwise be greyed and would be incorrectly swept up. 1053 // Note that if this object contains references, the writing 1054 // of those references will dirty the card containing this object 1055 // allowing the object to be blackened (and its references scanned) 1056 // either during a preclean phase or at the final checkpoint. 1057 void CMSCollector::direct_allocated(HeapWord* start, size_t size) { 1058 assert(_markBitMap.covers(start, size), "Out of bounds"); 1059 if (_collectorState >= Marking) { 1060 MutexLockerEx y(_markBitMap.lock(), 1061 Mutex::_no_safepoint_check_flag); 1062 // [see comments preceding SweepClosure::do_blk() below for details] 1063 // 1064 // Can the P-bits be deleted now? JJJ 1065 // 1066 // 1. need to mark the object as live so it isn't collected 1067 // 2. need to mark the 2nd bit to indicate the object may be uninitialized 1068 // 3. need to mark the end of the object so marking, precleaning or sweeping 1069 // can skip over uninitialized or unparsable objects. An allocated 1070 // object is considered uninitialized for our purposes as long as 1071 // its klass word is NULL. All old gen objects are parsable 1072 // as soon as they are initialized.) 1073 _markBitMap.mark(start); // object is live 1074 _markBitMap.mark(start + 1); // object is potentially uninitialized? 1075 _markBitMap.mark(start + size - 1); 1076 // mark end of object 1077 } 1078 // check that oop looks uninitialized 1079 assert(oop(start)->klass_or_null() == NULL, "_klass should be NULL"); 1080 } 1081 1082 void CMSCollector::promoted(bool par, HeapWord* start, 1083 bool is_obj_array, size_t obj_size) { 1084 assert(_markBitMap.covers(start), "Out of bounds"); 1085 // See comment in direct_allocated() about when objects should 1086 // be allocated live. 1087 if (_collectorState >= Marking) { 1088 // we already hold the marking bit map lock, taken in 1089 // the prologue 1090 if (par) { 1091 _markBitMap.par_mark(start); 1092 } else { 1093 _markBitMap.mark(start); 1094 } 1095 // We don't need to mark the object as uninitialized (as 1096 // in direct_allocated above) because this is being done with the 1097 // world stopped and the object will be initialized by the 1098 // time the marking, precleaning or sweeping get to look at it. 1099 // But see the code for copying objects into the CMS generation, 1100 // where we need to ensure that concurrent readers of the 1101 // block offset table are able to safely navigate a block that 1102 // is in flux from being free to being allocated (and in 1103 // transition while being copied into) and subsequently 1104 // becoming a bona-fide object when the copy/promotion is complete. 1105 assert(SafepointSynchronize::is_at_safepoint(), 1106 "expect promotion only at safepoints"); 1107 1108 if (_collectorState < Sweeping) { 1109 // Mark the appropriate cards in the modUnionTable, so that 1110 // this object gets scanned before the sweep. If this is 1111 // not done, CMS generation references in the object might 1112 // not get marked. 1113 // For the case of arrays, which are otherwise precisely 1114 // marked, we need to dirty the entire array, not just its head. 1115 if (is_obj_array) { 1116 // The [par_]mark_range() method expects mr.end() below to 1117 // be aligned to the granularity of a bit's representation 1118 // in the heap. In the case of the MUT below, that's a 1119 // card size. 1120 MemRegion mr(start, 1121 (HeapWord*)round_to((intptr_t)(start + obj_size), 1122 CardTableModRefBS::card_size /* bytes */)); 1123 if (par) { 1124 _modUnionTable.par_mark_range(mr); 1125 } else { 1126 _modUnionTable.mark_range(mr); 1127 } 1128 } else { // not an obj array; we can just mark the head 1129 if (par) { 1130 _modUnionTable.par_mark(start); 1131 } else { 1132 _modUnionTable.mark(start); 1133 } 1134 } 1135 } 1136 } 1137 } 1138 1139 static inline size_t percent_of_space(Space* space, HeapWord* addr) 1140 { 1141 size_t delta = pointer_delta(addr, space->bottom()); 1142 return (size_t)(delta * 100.0 / (space->capacity() / HeapWordSize)); 1143 } 1144 1145 void CMSCollector::icms_update_allocation_limits() 1146 { 1147 Generation* gen0 = GenCollectedHeap::heap()->get_gen(0); 1148 EdenSpace* eden = gen0->as_DefNewGeneration()->eden(); 1149 1150 const unsigned int duty_cycle = stats().icms_update_duty_cycle(); 1151 if (CMSTraceIncrementalPacing) { 1152 stats().print(); 1153 } 1154 1155 assert(duty_cycle <= 100, "invalid duty cycle"); 1156 if (duty_cycle != 0) { 1157 // The duty_cycle is a percentage between 0 and 100; convert to words and 1158 // then compute the offset from the endpoints of the space. 1159 size_t free_words = eden->free() / HeapWordSize; 1160 double free_words_dbl = (double)free_words; 1161 size_t duty_cycle_words = (size_t)(free_words_dbl * duty_cycle / 100.0); 1162 size_t offset_words = (free_words - duty_cycle_words) / 2; 1163 1164 _icms_start_limit = eden->top() + offset_words; 1165 _icms_stop_limit = eden->end() - offset_words; 1166 1167 // The limits may be adjusted (shifted to the right) by 1168 // CMSIncrementalOffset, to allow the application more mutator time after a 1169 // young gen gc (when all mutators were stopped) and before CMS starts and 1170 // takes away one or more cpus. 1171 if (CMSIncrementalOffset != 0) { 1172 double adjustment_dbl = free_words_dbl * CMSIncrementalOffset / 100.0; 1173 size_t adjustment = (size_t)adjustment_dbl; 1174 HeapWord* tmp_stop = _icms_stop_limit + adjustment; 1175 if (tmp_stop > _icms_stop_limit && tmp_stop < eden->end()) { 1176 _icms_start_limit += adjustment; 1177 _icms_stop_limit = tmp_stop; 1178 } 1179 } 1180 } 1181 if (duty_cycle == 0 || (_icms_start_limit == _icms_stop_limit)) { 1182 _icms_start_limit = _icms_stop_limit = eden->end(); 1183 } 1184 1185 // Install the new start limit. 1186 eden->set_soft_end(_icms_start_limit); 1187 1188 if (CMSTraceIncrementalMode) { 1189 gclog_or_tty->print(" icms alloc limits: " 1190 PTR_FORMAT "," PTR_FORMAT 1191 " (" SIZE_FORMAT "%%," SIZE_FORMAT "%%) ", 1192 p2i(_icms_start_limit), p2i(_icms_stop_limit), 1193 percent_of_space(eden, _icms_start_limit), 1194 percent_of_space(eden, _icms_stop_limit)); 1195 if (Verbose) { 1196 gclog_or_tty->print("eden: "); 1197 eden->print_on(gclog_or_tty); 1198 } 1199 } 1200 } 1201 1202 // Any changes here should try to maintain the invariant 1203 // that if this method is called with _icms_start_limit 1204 // and _icms_stop_limit both NULL, then it should return NULL 1205 // and not notify the icms thread. 1206 HeapWord* 1207 CMSCollector::allocation_limit_reached(Space* space, HeapWord* top, 1208 size_t word_size) 1209 { 1210 // A start_limit equal to end() means the duty cycle is 0, so treat that as a 1211 // nop. 1212 if (CMSIncrementalMode && _icms_start_limit != space->end()) { 1213 if (top <= _icms_start_limit) { 1214 if (CMSTraceIncrementalMode) { 1215 space->print_on(gclog_or_tty); 1216 gclog_or_tty->stamp(); 1217 gclog_or_tty->print_cr(" start limit top=" PTR_FORMAT 1218 ", new limit=" PTR_FORMAT 1219 " (" SIZE_FORMAT "%%)", 1220 p2i(top), p2i(_icms_stop_limit), 1221 percent_of_space(space, _icms_stop_limit)); 1222 } 1223 ConcurrentMarkSweepThread::start_icms(); 1224 assert(top < _icms_stop_limit, "Tautology"); 1225 if (word_size < pointer_delta(_icms_stop_limit, top)) { 1226 return _icms_stop_limit; 1227 } 1228 1229 // The allocation will cross both the _start and _stop limits, so do the 1230 // stop notification also and return end(). 1231 if (CMSTraceIncrementalMode) { 1232 space->print_on(gclog_or_tty); 1233 gclog_or_tty->stamp(); 1234 gclog_or_tty->print_cr(" +stop limit top=" PTR_FORMAT 1235 ", new limit=" PTR_FORMAT 1236 " (" SIZE_FORMAT "%%)", 1237 p2i(top), p2i(space->end()), 1238 percent_of_space(space, space->end())); 1239 } 1240 ConcurrentMarkSweepThread::stop_icms(); 1241 return space->end(); 1242 } 1243 1244 if (top <= _icms_stop_limit) { 1245 if (CMSTraceIncrementalMode) { 1246 space->print_on(gclog_or_tty); 1247 gclog_or_tty->stamp(); 1248 gclog_or_tty->print_cr(" stop limit top=" PTR_FORMAT 1249 ", new limit=" PTR_FORMAT 1250 " (" SIZE_FORMAT "%%)", 1251 top, space->end(), 1252 percent_of_space(space, space->end())); 1253 } 1254 ConcurrentMarkSweepThread::stop_icms(); 1255 return space->end(); 1256 } 1257 1258 if (CMSTraceIncrementalMode) { 1259 space->print_on(gclog_or_tty); 1260 gclog_or_tty->stamp(); 1261 gclog_or_tty->print_cr(" end limit top=" PTR_FORMAT 1262 ", new limit=" PTR_FORMAT, 1263 top, NULL); 1264 } 1265 } 1266 1267 return NULL; 1268 } 1269 1270 oop ConcurrentMarkSweepGeneration::promote(oop obj, size_t obj_size) { 1271 assert(obj_size == (size_t)obj->size(), "bad obj_size passed in"); 1272 // allocate, copy and if necessary update promoinfo -- 1273 // delegate to underlying space. 1274 assert_lock_strong(freelistLock()); 1275 1276 #ifndef PRODUCT 1277 if (Universe::heap()->promotion_should_fail()) { 1278 return NULL; 1279 } 1280 #endif // #ifndef PRODUCT 1281 1282 oop res = _cmsSpace->promote(obj, obj_size); 1283 if (res == NULL) { 1284 // expand and retry 1285 size_t s = _cmsSpace->expansionSpaceRequired(obj_size); // HeapWords 1286 expand(s*HeapWordSize, MinHeapDeltaBytes, 1287 CMSExpansionCause::_satisfy_promotion); 1288 // Since there's currently no next generation, we don't try to promote 1289 // into a more senior generation. 1290 assert(next_gen() == NULL, "assumption, based upon which no attempt " 1291 "is made to pass on a possibly failing " 1292 "promotion to next generation"); 1293 res = _cmsSpace->promote(obj, obj_size); 1294 } 1295 if (res != NULL) { 1296 // See comment in allocate() about when objects should 1297 // be allocated live. 1298 assert(obj->is_oop(), "Will dereference klass pointer below"); 1299 collector()->promoted(false, // Not parallel 1300 (HeapWord*)res, obj->is_objArray(), obj_size); 1301 // promotion counters 1302 NOT_PRODUCT( 1303 _numObjectsPromoted++; 1304 _numWordsPromoted += 1305 (int)(CompactibleFreeListSpace::adjustObjectSize(obj->size())); 1306 ) 1307 } 1308 return res; 1309 } 1310 1311 1312 HeapWord* 1313 ConcurrentMarkSweepGeneration::allocation_limit_reached(Space* space, 1314 HeapWord* top, 1315 size_t word_sz) 1316 { 1317 return collector()->allocation_limit_reached(space, top, word_sz); 1318 } 1319 1320 // IMPORTANT: Notes on object size recognition in CMS. 1321 // --------------------------------------------------- 1322 // A block of storage in the CMS generation is always in 1323 // one of three states. A free block (FREE), an allocated 1324 // object (OBJECT) whose size() method reports the correct size, 1325 // and an intermediate state (TRANSIENT) in which its size cannot 1326 // be accurately determined. 1327 // STATE IDENTIFICATION: (32 bit and 64 bit w/o COOPS) 1328 // ----------------------------------------------------- 1329 // FREE: klass_word & 1 == 1; mark_word holds block size 1330 // 1331 // OBJECT: klass_word installed; klass_word != 0 && klass_word & 1 == 0; 1332 // obj->size() computes correct size 1333 // 1334 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT 1335 // 1336 // STATE IDENTIFICATION: (64 bit+COOPS) 1337 // ------------------------------------ 1338 // FREE: mark_word & CMS_FREE_BIT == 1; mark_word & ~CMS_FREE_BIT gives block_size 1339 // 1340 // OBJECT: klass_word installed; klass_word != 0; 1341 // obj->size() computes correct size 1342 // 1343 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT 1344 // 1345 // 1346 // STATE TRANSITION DIAGRAM 1347 // 1348 // mut / parnew mut / parnew 1349 // FREE --------------------> TRANSIENT ---------------------> OBJECT --| 1350 // ^ | 1351 // |------------------------ DEAD <------------------------------------| 1352 // sweep mut 1353 // 1354 // While a block is in TRANSIENT state its size cannot be determined 1355 // so readers will either need to come back later or stall until 1356 // the size can be determined. Note that for the case of direct 1357 // allocation, P-bits, when available, may be used to determine the 1358 // size of an object that may not yet have been initialized. 1359 1360 // Things to support parallel young-gen collection. 1361 oop 1362 ConcurrentMarkSweepGeneration::par_promote(int thread_num, 1363 oop old, markOop m, 1364 size_t word_sz) { 1365 #ifndef PRODUCT 1366 if (Universe::heap()->promotion_should_fail()) { 1367 return NULL; 1368 } 1369 #endif // #ifndef PRODUCT 1370 1371 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1372 PromotionInfo* promoInfo = &ps->promo; 1373 // if we are tracking promotions, then first ensure space for 1374 // promotion (including spooling space for saving header if necessary). 1375 // then allocate and copy, then track promoted info if needed. 1376 // When tracking (see PromotionInfo::track()), the mark word may 1377 // be displaced and in this case restoration of the mark word 1378 // occurs in the (oop_since_save_marks_)iterate phase. 1379 if (promoInfo->tracking() && !promoInfo->ensure_spooling_space()) { 1380 // Out of space for allocating spooling buffers; 1381 // try expanding and allocating spooling buffers. 1382 if (!expand_and_ensure_spooling_space(promoInfo)) { 1383 return NULL; 1384 } 1385 } 1386 assert(promoInfo->has_spooling_space(), "Control point invariant"); 1387 const size_t alloc_sz = CompactibleFreeListSpace::adjustObjectSize(word_sz); 1388 HeapWord* obj_ptr = ps->lab.alloc(alloc_sz); 1389 if (obj_ptr == NULL) { 1390 obj_ptr = expand_and_par_lab_allocate(ps, alloc_sz); 1391 if (obj_ptr == NULL) { 1392 return NULL; 1393 } 1394 } 1395 oop obj = oop(obj_ptr); 1396 OrderAccess::storestore(); 1397 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1398 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1399 // IMPORTANT: See note on object initialization for CMS above. 1400 // Otherwise, copy the object. Here we must be careful to insert the 1401 // klass pointer last, since this marks the block as an allocated object. 1402 // Except with compressed oops it's the mark word. 1403 HeapWord* old_ptr = (HeapWord*)old; 1404 // Restore the mark word copied above. 1405 obj->set_mark(m); 1406 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1407 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1408 OrderAccess::storestore(); 1409 1410 if (UseCompressedClassPointers) { 1411 // Copy gap missed by (aligned) header size calculation below 1412 obj->set_klass_gap(old->klass_gap()); 1413 } 1414 if (word_sz > (size_t)oopDesc::header_size()) { 1415 Copy::aligned_disjoint_words(old_ptr + oopDesc::header_size(), 1416 obj_ptr + oopDesc::header_size(), 1417 word_sz - oopDesc::header_size()); 1418 } 1419 1420 // Now we can track the promoted object, if necessary. We take care 1421 // to delay the transition from uninitialized to full object 1422 // (i.e., insertion of klass pointer) until after, so that it 1423 // atomically becomes a promoted object. 1424 if (promoInfo->tracking()) { 1425 promoInfo->track((PromotedObject*)obj, old->klass()); 1426 } 1427 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1428 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1429 assert(old->is_oop(), "Will use and dereference old klass ptr below"); 1430 1431 // Finally, install the klass pointer (this should be volatile). 1432 OrderAccess::storestore(); 1433 obj->set_klass(old->klass()); 1434 // We should now be able to calculate the right size for this object 1435 assert(obj->is_oop() && obj->size() == (int)word_sz, "Error, incorrect size computed for promoted object"); 1436 1437 collector()->promoted(true, // parallel 1438 obj_ptr, old->is_objArray(), word_sz); 1439 1440 NOT_PRODUCT( 1441 Atomic::inc_ptr(&_numObjectsPromoted); 1442 Atomic::add_ptr(alloc_sz, &_numWordsPromoted); 1443 ) 1444 1445 return obj; 1446 } 1447 1448 void 1449 ConcurrentMarkSweepGeneration:: 1450 par_promote_alloc_undo(int thread_num, 1451 HeapWord* obj, size_t word_sz) { 1452 // CMS does not support promotion undo. 1453 ShouldNotReachHere(); 1454 } 1455 1456 void 1457 ConcurrentMarkSweepGeneration:: 1458 par_promote_alloc_done(int thread_num) { 1459 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1460 ps->lab.retire(thread_num); 1461 } 1462 1463 void 1464 ConcurrentMarkSweepGeneration:: 1465 par_oop_since_save_marks_iterate_done(int thread_num) { 1466 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1467 ParScanWithoutBarrierClosure* dummy_cl = NULL; 1468 ps->promo.promoted_oops_iterate_nv(dummy_cl); 1469 } 1470 1471 bool ConcurrentMarkSweepGeneration::should_collect(bool full, 1472 size_t size, 1473 bool tlab) 1474 { 1475 // We allow a STW collection only if a full 1476 // collection was requested. 1477 return full || should_allocate(size, tlab); // FIX ME !!! 1478 // This and promotion failure handling are connected at the 1479 // hip and should be fixed by untying them. 1480 } 1481 1482 bool CMSCollector::shouldConcurrentCollect() { 1483 if (_full_gc_requested) { 1484 if (Verbose && PrintGCDetails) { 1485 gclog_or_tty->print_cr("CMSCollector: collect because of explicit " 1486 " gc request (or gc_locker)"); 1487 } 1488 return true; 1489 } 1490 1491 // For debugging purposes, change the type of collection. 1492 // If the rotation is not on the concurrent collection 1493 // type, don't start a concurrent collection. 1494 NOT_PRODUCT( 1495 if (RotateCMSCollectionTypes && 1496 (_cmsGen->debug_collection_type() != 1497 ConcurrentMarkSweepGeneration::Concurrent_collection_type)) { 1498 assert(_cmsGen->debug_collection_type() != 1499 ConcurrentMarkSweepGeneration::Unknown_collection_type, 1500 "Bad cms collection type"); 1501 return false; 1502 } 1503 ) 1504 1505 FreelistLocker x(this); 1506 // ------------------------------------------------------------------ 1507 // Print out lots of information which affects the initiation of 1508 // a collection. 1509 if (PrintCMSInitiationStatistics && stats().valid()) { 1510 gclog_or_tty->print("CMSCollector shouldConcurrentCollect: "); 1511 gclog_or_tty->stamp(); 1512 gclog_or_tty->cr(); 1513 stats().print_on(gclog_or_tty); 1514 gclog_or_tty->print_cr("time_until_cms_gen_full %3.7f", 1515 stats().time_until_cms_gen_full()); 1516 gclog_or_tty->print_cr("free="SIZE_FORMAT, _cmsGen->free()); 1517 gclog_or_tty->print_cr("contiguous_available="SIZE_FORMAT, 1518 _cmsGen->contiguous_available()); 1519 gclog_or_tty->print_cr("promotion_rate=%g", stats().promotion_rate()); 1520 gclog_or_tty->print_cr("cms_allocation_rate=%g", stats().cms_allocation_rate()); 1521 gclog_or_tty->print_cr("occupancy=%3.7f", _cmsGen->occupancy()); 1522 gclog_or_tty->print_cr("initiatingOccupancy=%3.7f", _cmsGen->initiating_occupancy()); 1523 gclog_or_tty->print_cr("cms_time_since_begin=%3.7f", stats().cms_time_since_begin()); 1524 gclog_or_tty->print_cr("cms_time_since_end=%3.7f", stats().cms_time_since_end()); 1525 gclog_or_tty->print_cr("metadata initialized %d", 1526 MetaspaceGC::should_concurrent_collect()); 1527 } 1528 // ------------------------------------------------------------------ 1529 1530 // If the estimated time to complete a cms collection (cms_duration()) 1531 // is less than the estimated time remaining until the cms generation 1532 // is full, start a collection. 1533 if (!UseCMSInitiatingOccupancyOnly) { 1534 if (stats().valid()) { 1535 if (stats().time_until_cms_start() == 0.0) { 1536 return true; 1537 } 1538 } else { 1539 // We want to conservatively collect somewhat early in order 1540 // to try and "bootstrap" our CMS/promotion statistics; 1541 // this branch will not fire after the first successful CMS 1542 // collection because the stats should then be valid. 1543 if (_cmsGen->occupancy() >= _bootstrap_occupancy) { 1544 if (Verbose && PrintGCDetails) { 1545 gclog_or_tty->print_cr( 1546 " CMSCollector: collect for bootstrapping statistics:" 1547 " occupancy = %f, boot occupancy = %f", _cmsGen->occupancy(), 1548 _bootstrap_occupancy); 1549 } 1550 return true; 1551 } 1552 } 1553 } 1554 1555 // Otherwise, we start a collection cycle if 1556 // old gen want a collection cycle started. Each may use 1557 // an appropriate criterion for making this decision. 1558 // XXX We need to make sure that the gen expansion 1559 // criterion dovetails well with this. XXX NEED TO FIX THIS 1560 if (_cmsGen->should_concurrent_collect()) { 1561 if (Verbose && PrintGCDetails) { 1562 gclog_or_tty->print_cr("CMS old gen initiated"); 1563 } 1564 return true; 1565 } 1566 1567 // We start a collection if we believe an incremental collection may fail; 1568 // this is not likely to be productive in practice because it's probably too 1569 // late anyway. 1570 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1571 assert(gch->collector_policy()->is_two_generation_policy(), 1572 "You may want to check the correctness of the following"); 1573 if (gch->incremental_collection_will_fail(true /* consult_young */)) { 1574 if (Verbose && PrintGCDetails) { 1575 gclog_or_tty->print("CMSCollector: collect because incremental collection will fail "); 1576 } 1577 return true; 1578 } 1579 1580 if (MetaspaceGC::should_concurrent_collect()) { 1581 if (Verbose && PrintGCDetails) { 1582 gclog_or_tty->print("CMSCollector: collect for metadata allocation "); 1583 } 1584 return true; 1585 } 1586 1587 // CMSTriggerInterval starts a CMS cycle if enough time has passed. 1588 if (CMSTriggerInterval >= 0) { 1589 if (CMSTriggerInterval == 0) { 1590 // Trigger always 1591 return true; 1592 } 1593 1594 // Check the CMS time since begin (we do not check the stats validity 1595 // as we want to be able to trigger the first CMS cycle as well) 1596 if (stats().cms_time_since_begin() >= (CMSTriggerInterval / ((double) MILLIUNITS))) { 1597 if (Verbose && PrintGCDetails) { 1598 if (stats().valid()) { 1599 gclog_or_tty->print_cr("CMSCollector: collect because of trigger interval (time since last begin %3.7f secs)", 1600 stats().cms_time_since_begin()); 1601 } else { 1602 gclog_or_tty->print_cr("CMSCollector: collect because of trigger interval (first collection)"); 1603 } 1604 } 1605 return true; 1606 } 1607 } 1608 1609 return false; 1610 } 1611 1612 void CMSCollector::set_did_compact(bool v) { _cmsGen->set_did_compact(v); } 1613 1614 // Clear _expansion_cause fields of constituent generations 1615 void CMSCollector::clear_expansion_cause() { 1616 _cmsGen->clear_expansion_cause(); 1617 } 1618 1619 // We should be conservative in starting a collection cycle. To 1620 // start too eagerly runs the risk of collecting too often in the 1621 // extreme. To collect too rarely falls back on full collections, 1622 // which works, even if not optimum in terms of concurrent work. 1623 // As a work around for too eagerly collecting, use the flag 1624 // UseCMSInitiatingOccupancyOnly. This also has the advantage of 1625 // giving the user an easily understandable way of controlling the 1626 // collections. 1627 // We want to start a new collection cycle if any of the following 1628 // conditions hold: 1629 // . our current occupancy exceeds the configured initiating occupancy 1630 // for this generation, or 1631 // . we recently needed to expand this space and have not, since that 1632 // expansion, done a collection of this generation, or 1633 // . the underlying space believes that it may be a good idea to initiate 1634 // a concurrent collection (this may be based on criteria such as the 1635 // following: the space uses linear allocation and linear allocation is 1636 // going to fail, or there is believed to be excessive fragmentation in 1637 // the generation, etc... or ... 1638 // [.(currently done by CMSCollector::shouldConcurrentCollect() only for 1639 // the case of the old generation; see CR 6543076): 1640 // we may be approaching a point at which allocation requests may fail because 1641 // we will be out of sufficient free space given allocation rate estimates.] 1642 bool ConcurrentMarkSweepGeneration::should_concurrent_collect() const { 1643 1644 assert_lock_strong(freelistLock()); 1645 if (occupancy() > initiating_occupancy()) { 1646 if (PrintGCDetails && Verbose) { 1647 gclog_or_tty->print(" %s: collect because of occupancy %f / %f ", 1648 short_name(), occupancy(), initiating_occupancy()); 1649 } 1650 return true; 1651 } 1652 if (UseCMSInitiatingOccupancyOnly) { 1653 return false; 1654 } 1655 if (expansion_cause() == CMSExpansionCause::_satisfy_allocation) { 1656 if (PrintGCDetails && Verbose) { 1657 gclog_or_tty->print(" %s: collect because expanded for allocation ", 1658 short_name()); 1659 } 1660 return true; 1661 } 1662 if (_cmsSpace->should_concurrent_collect()) { 1663 if (PrintGCDetails && Verbose) { 1664 gclog_or_tty->print(" %s: collect because cmsSpace says so ", 1665 short_name()); 1666 } 1667 return true; 1668 } 1669 return false; 1670 } 1671 1672 void ConcurrentMarkSweepGeneration::collect(bool full, 1673 bool clear_all_soft_refs, 1674 size_t size, 1675 bool tlab) 1676 { 1677 collector()->collect(full, clear_all_soft_refs, size, tlab); 1678 } 1679 1680 void CMSCollector::collect(bool full, 1681 bool clear_all_soft_refs, 1682 size_t size, 1683 bool tlab) 1684 { 1685 if (!UseCMSCollectionPassing && _collectorState > Idling) { 1686 // For debugging purposes skip the collection if the state 1687 // is not currently idle 1688 if (TraceCMSState) { 1689 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " skipped full:%d CMS state %d", 1690 Thread::current(), full, _collectorState); 1691 } 1692 return; 1693 } 1694 1695 // The following "if" branch is present for defensive reasons. 1696 // In the current uses of this interface, it can be replaced with: 1697 // assert(!GC_locker.is_active(), "Can't be called otherwise"); 1698 // But I am not placing that assert here to allow future 1699 // generality in invoking this interface. 1700 if (GC_locker::is_active()) { 1701 // A consistency test for GC_locker 1702 assert(GC_locker::needs_gc(), "Should have been set already"); 1703 // Skip this foreground collection, instead 1704 // expanding the heap if necessary. 1705 // Need the free list locks for the call to free() in compute_new_size() 1706 compute_new_size(); 1707 return; 1708 } 1709 acquire_control_and_collect(full, clear_all_soft_refs); 1710 _full_gcs_since_conc_gc++; 1711 } 1712 1713 void CMSCollector::request_full_gc(unsigned int full_gc_count, GCCause::Cause cause) { 1714 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1715 unsigned int gc_count = gch->total_full_collections(); 1716 if (gc_count == full_gc_count) { 1717 MutexLockerEx y(CGC_lock, Mutex::_no_safepoint_check_flag); 1718 _full_gc_requested = true; 1719 _full_gc_cause = cause; 1720 CGC_lock->notify(); // nudge CMS thread 1721 } else { 1722 assert(gc_count > full_gc_count, "Error: causal loop"); 1723 } 1724 } 1725 1726 bool CMSCollector::is_external_interruption() { 1727 GCCause::Cause cause = GenCollectedHeap::heap()->gc_cause(); 1728 return GCCause::is_user_requested_gc(cause) || 1729 GCCause::is_serviceability_requested_gc(cause); 1730 } 1731 1732 void CMSCollector::report_concurrent_mode_interruption() { 1733 if (is_external_interruption()) { 1734 if (PrintGCDetails) { 1735 gclog_or_tty->print(" (concurrent mode interrupted)"); 1736 } 1737 } else { 1738 if (PrintGCDetails) { 1739 gclog_or_tty->print(" (concurrent mode failure)"); 1740 } 1741 _gc_tracer_cm->report_concurrent_mode_failure(); 1742 } 1743 } 1744 1745 1746 // The foreground and background collectors need to coordinate in order 1747 // to make sure that they do not mutually interfere with CMS collections. 1748 // When a background collection is active, 1749 // the foreground collector may need to take over (preempt) and 1750 // synchronously complete an ongoing collection. Depending on the 1751 // frequency of the background collections and the heap usage 1752 // of the application, this preemption can be seldom or frequent. 1753 // There are only certain 1754 // points in the background collection that the "collection-baton" 1755 // can be passed to the foreground collector. 1756 // 1757 // The foreground collector will wait for the baton before 1758 // starting any part of the collection. The foreground collector 1759 // will only wait at one location. 1760 // 1761 // The background collector will yield the baton before starting a new 1762 // phase of the collection (e.g., before initial marking, marking from roots, 1763 // precleaning, final re-mark, sweep etc.) This is normally done at the head 1764 // of the loop which switches the phases. The background collector does some 1765 // of the phases (initial mark, final re-mark) with the world stopped. 1766 // Because of locking involved in stopping the world, 1767 // the foreground collector should not block waiting for the background 1768 // collector when it is doing a stop-the-world phase. The background 1769 // collector will yield the baton at an additional point just before 1770 // it enters a stop-the-world phase. Once the world is stopped, the 1771 // background collector checks the phase of the collection. If the 1772 // phase has not changed, it proceeds with the collection. If the 1773 // phase has changed, it skips that phase of the collection. See 1774 // the comments on the use of the Heap_lock in collect_in_background(). 1775 // 1776 // Variable used in baton passing. 1777 // _foregroundGCIsActive - Set to true by the foreground collector when 1778 // it wants the baton. The foreground clears it when it has finished 1779 // the collection. 1780 // _foregroundGCShouldWait - Set to true by the background collector 1781 // when it is running. The foreground collector waits while 1782 // _foregroundGCShouldWait is true. 1783 // CGC_lock - monitor used to protect access to the above variables 1784 // and to notify the foreground and background collectors. 1785 // _collectorState - current state of the CMS collection. 1786 // 1787 // The foreground collector 1788 // acquires the CGC_lock 1789 // sets _foregroundGCIsActive 1790 // waits on the CGC_lock for _foregroundGCShouldWait to be false 1791 // various locks acquired in preparation for the collection 1792 // are released so as not to block the background collector 1793 // that is in the midst of a collection 1794 // proceeds with the collection 1795 // clears _foregroundGCIsActive 1796 // returns 1797 // 1798 // The background collector in a loop iterating on the phases of the 1799 // collection 1800 // acquires the CGC_lock 1801 // sets _foregroundGCShouldWait 1802 // if _foregroundGCIsActive is set 1803 // clears _foregroundGCShouldWait, notifies _CGC_lock 1804 // waits on _CGC_lock for _foregroundGCIsActive to become false 1805 // and exits the loop. 1806 // otherwise 1807 // proceed with that phase of the collection 1808 // if the phase is a stop-the-world phase, 1809 // yield the baton once more just before enqueueing 1810 // the stop-world CMS operation (executed by the VM thread). 1811 // returns after all phases of the collection are done 1812 // 1813 1814 void CMSCollector::acquire_control_and_collect(bool full, 1815 bool clear_all_soft_refs) { 1816 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint"); 1817 assert(!Thread::current()->is_ConcurrentGC_thread(), 1818 "shouldn't try to acquire control from self!"); 1819 1820 // Start the protocol for acquiring control of the 1821 // collection from the background collector (aka CMS thread). 1822 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 1823 "VM thread should have CMS token"); 1824 // Remember the possibly interrupted state of an ongoing 1825 // concurrent collection 1826 CollectorState first_state = _collectorState; 1827 1828 // Signal to a possibly ongoing concurrent collection that 1829 // we want to do a foreground collection. 1830 _foregroundGCIsActive = true; 1831 1832 // Disable incremental mode during a foreground collection. 1833 ICMSDisabler icms_disabler; 1834 1835 // release locks and wait for a notify from the background collector 1836 // releasing the locks in only necessary for phases which 1837 // do yields to improve the granularity of the collection. 1838 assert_lock_strong(bitMapLock()); 1839 // We need to lock the Free list lock for the space that we are 1840 // currently collecting. 1841 assert(haveFreelistLocks(), "Must be holding free list locks"); 1842 bitMapLock()->unlock(); 1843 releaseFreelistLocks(); 1844 { 1845 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1846 if (_foregroundGCShouldWait) { 1847 // We are going to be waiting for action for the CMS thread; 1848 // it had better not be gone (for instance at shutdown)! 1849 assert(ConcurrentMarkSweepThread::cmst() != NULL, 1850 "CMS thread must be running"); 1851 // Wait here until the background collector gives us the go-ahead 1852 ConcurrentMarkSweepThread::clear_CMS_flag( 1853 ConcurrentMarkSweepThread::CMS_vm_has_token); // release token 1854 // Get a possibly blocked CMS thread going: 1855 // Note that we set _foregroundGCIsActive true above, 1856 // without protection of the CGC_lock. 1857 CGC_lock->notify(); 1858 assert(!ConcurrentMarkSweepThread::vm_thread_wants_cms_token(), 1859 "Possible deadlock"); 1860 while (_foregroundGCShouldWait) { 1861 // wait for notification 1862 CGC_lock->wait(Mutex::_no_safepoint_check_flag); 1863 // Possibility of delay/starvation here, since CMS token does 1864 // not know to give priority to VM thread? Actually, i think 1865 // there wouldn't be any delay/starvation, but the proof of 1866 // that "fact" (?) appears non-trivial. XXX 20011219YSR 1867 } 1868 ConcurrentMarkSweepThread::set_CMS_flag( 1869 ConcurrentMarkSweepThread::CMS_vm_has_token); 1870 } 1871 } 1872 // The CMS_token is already held. Get back the other locks. 1873 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 1874 "VM thread should have CMS token"); 1875 getFreelistLocks(); 1876 bitMapLock()->lock_without_safepoint_check(); 1877 if (TraceCMSState) { 1878 gclog_or_tty->print_cr("CMS foreground collector has asked for control " 1879 INTPTR_FORMAT " with first state %d", Thread::current(), first_state); 1880 gclog_or_tty->print_cr(" gets control with state %d", _collectorState); 1881 } 1882 1883 // Check if we need to do a compaction, or if not, whether 1884 // we need to start the mark-sweep from scratch. 1885 bool should_compact = false; 1886 bool should_start_over = false; 1887 decide_foreground_collection_type(clear_all_soft_refs, 1888 &should_compact, &should_start_over); 1889 1890 NOT_PRODUCT( 1891 if (RotateCMSCollectionTypes) { 1892 if (_cmsGen->debug_collection_type() == 1893 ConcurrentMarkSweepGeneration::MSC_foreground_collection_type) { 1894 should_compact = true; 1895 } else if (_cmsGen->debug_collection_type() == 1896 ConcurrentMarkSweepGeneration::MS_foreground_collection_type) { 1897 should_compact = false; 1898 } 1899 } 1900 ) 1901 1902 if (first_state > Idling) { 1903 report_concurrent_mode_interruption(); 1904 } 1905 1906 set_did_compact(should_compact); 1907 if (should_compact) { 1908 // If the collection is being acquired from the background 1909 // collector, there may be references on the discovered 1910 // references lists that have NULL referents (being those 1911 // that were concurrently cleared by a mutator) or 1912 // that are no longer active (having been enqueued concurrently 1913 // by the mutator). 1914 // Scrub the list of those references because Mark-Sweep-Compact 1915 // code assumes referents are not NULL and that all discovered 1916 // Reference objects are active. 1917 ref_processor()->clean_up_discovered_references(); 1918 1919 if (first_state > Idling) { 1920 save_heap_summary(); 1921 } 1922 1923 do_compaction_work(clear_all_soft_refs); 1924 1925 // Has the GC time limit been exceeded? 1926 DefNewGeneration* young_gen = _young_gen->as_DefNewGeneration(); 1927 size_t max_eden_size = young_gen->max_capacity() - 1928 young_gen->to()->capacity() - 1929 young_gen->from()->capacity(); 1930 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1931 GCCause::Cause gc_cause = gch->gc_cause(); 1932 size_policy()->check_gc_overhead_limit(_young_gen->used(), 1933 young_gen->eden()->used(), 1934 _cmsGen->max_capacity(), 1935 max_eden_size, 1936 full, 1937 gc_cause, 1938 gch->collector_policy()); 1939 } else { 1940 do_mark_sweep_work(clear_all_soft_refs, first_state, 1941 should_start_over); 1942 } 1943 // Reset the expansion cause, now that we just completed 1944 // a collection cycle. 1945 clear_expansion_cause(); 1946 _foregroundGCIsActive = false; 1947 return; 1948 } 1949 1950 // Resize the tenured generation 1951 // after obtaining the free list locks for the 1952 // two generations. 1953 void CMSCollector::compute_new_size() { 1954 assert_locked_or_safepoint(Heap_lock); 1955 FreelistLocker z(this); 1956 MetaspaceGC::compute_new_size(); 1957 _cmsGen->compute_new_size_free_list(); 1958 } 1959 1960 // A work method used by foreground collection to determine 1961 // what type of collection (compacting or not, continuing or fresh) 1962 // it should do. 1963 // NOTE: the intent is to make UseCMSCompactAtFullCollection 1964 // and CMSCompactWhenClearAllSoftRefs the default in the future 1965 // and do away with the flags after a suitable period. 1966 void CMSCollector::decide_foreground_collection_type( 1967 bool clear_all_soft_refs, bool* should_compact, 1968 bool* should_start_over) { 1969 // Normally, we'll compact only if the UseCMSCompactAtFullCollection 1970 // flag is set, and we have either requested a System.gc() or 1971 // the number of full gc's since the last concurrent cycle 1972 // has exceeded the threshold set by CMSFullGCsBeforeCompaction, 1973 // or if an incremental collection has failed 1974 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1975 assert(gch->collector_policy()->is_two_generation_policy(), 1976 "You may want to check the correctness of the following"); 1977 // Inform cms gen if this was due to partial collection failing. 1978 // The CMS gen may use this fact to determine its expansion policy. 1979 if (gch->incremental_collection_will_fail(false /* don't consult_young */)) { 1980 assert(!_cmsGen->incremental_collection_failed(), 1981 "Should have been noticed, reacted to and cleared"); 1982 _cmsGen->set_incremental_collection_failed(); 1983 } 1984 *should_compact = 1985 UseCMSCompactAtFullCollection && 1986 ((_full_gcs_since_conc_gc >= CMSFullGCsBeforeCompaction) || 1987 GCCause::is_user_requested_gc(gch->gc_cause()) || 1988 gch->incremental_collection_will_fail(true /* consult_young */)); 1989 *should_start_over = false; 1990 if (clear_all_soft_refs && !*should_compact) { 1991 // We are about to do a last ditch collection attempt 1992 // so it would normally make sense to do a compaction 1993 // to reclaim as much space as possible. 1994 if (CMSCompactWhenClearAllSoftRefs) { 1995 // Default: The rationale is that in this case either 1996 // we are past the final marking phase, in which case 1997 // we'd have to start over, or so little has been done 1998 // that there's little point in saving that work. Compaction 1999 // appears to be the sensible choice in either case. 2000 *should_compact = true; 2001 } else { 2002 // We have been asked to clear all soft refs, but not to 2003 // compact. Make sure that we aren't past the final checkpoint 2004 // phase, for that is where we process soft refs. If we are already 2005 // past that phase, we'll need to redo the refs discovery phase and 2006 // if necessary clear soft refs that weren't previously 2007 // cleared. We do so by remembering the phase in which 2008 // we came in, and if we are past the refs processing 2009 // phase, we'll choose to just redo the mark-sweep 2010 // collection from scratch. 2011 if (_collectorState > FinalMarking) { 2012 // We are past the refs processing phase; 2013 // start over and do a fresh synchronous CMS cycle 2014 _collectorState = Resetting; // skip to reset to start new cycle 2015 reset(false /* == !asynch */); 2016 *should_start_over = true; 2017 } // else we can continue a possibly ongoing current cycle 2018 } 2019 } 2020 } 2021 2022 // A work method used by the foreground collector to do 2023 // a mark-sweep-compact. 2024 void CMSCollector::do_compaction_work(bool clear_all_soft_refs) { 2025 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2026 2027 STWGCTimer* gc_timer = GenMarkSweep::gc_timer(); 2028 gc_timer->register_gc_start(); 2029 2030 SerialOldTracer* gc_tracer = GenMarkSweep::gc_tracer(); 2031 gc_tracer->report_gc_start(gch->gc_cause(), gc_timer->gc_start()); 2032 2033 GCTraceTime t("CMS:MSC ", PrintGCDetails && Verbose, true, NULL, gc_tracer->gc_id()); 2034 if (PrintGC && Verbose && !(GCCause::is_user_requested_gc(gch->gc_cause()))) { 2035 gclog_or_tty->print_cr("Compact ConcurrentMarkSweepGeneration after %d " 2036 "collections passed to foreground collector", _full_gcs_since_conc_gc); 2037 } 2038 2039 // Sample collection interval time and reset for collection pause. 2040 if (UseAdaptiveSizePolicy) { 2041 size_policy()->msc_collection_begin(); 2042 } 2043 2044 // Temporarily widen the span of the weak reference processing to 2045 // the entire heap. 2046 MemRegion new_span(GenCollectedHeap::heap()->reserved_region()); 2047 ReferenceProcessorSpanMutator rp_mut_span(ref_processor(), new_span); 2048 // Temporarily, clear the "is_alive_non_header" field of the 2049 // reference processor. 2050 ReferenceProcessorIsAliveMutator rp_mut_closure(ref_processor(), NULL); 2051 // Temporarily make reference _processing_ single threaded (non-MT). 2052 ReferenceProcessorMTProcMutator rp_mut_mt_processing(ref_processor(), false); 2053 // Temporarily make refs discovery atomic 2054 ReferenceProcessorAtomicMutator rp_mut_atomic(ref_processor(), true); 2055 // Temporarily make reference _discovery_ single threaded (non-MT) 2056 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false); 2057 2058 ref_processor()->set_enqueuing_is_done(false); 2059 ref_processor()->enable_discovery(false /*verify_disabled*/, false /*check_no_refs*/); 2060 ref_processor()->setup_policy(clear_all_soft_refs); 2061 // If an asynchronous collection finishes, the _modUnionTable is 2062 // all clear. If we are assuming the collection from an asynchronous 2063 // collection, clear the _modUnionTable. 2064 assert(_collectorState != Idling || _modUnionTable.isAllClear(), 2065 "_modUnionTable should be clear if the baton was not passed"); 2066 _modUnionTable.clear_all(); 2067 assert(_collectorState != Idling || _ct->klass_rem_set()->mod_union_is_clear(), 2068 "mod union for klasses should be clear if the baton was passed"); 2069 _ct->klass_rem_set()->clear_mod_union(); 2070 2071 // We must adjust the allocation statistics being maintained 2072 // in the free list space. We do so by reading and clearing 2073 // the sweep timer and updating the block flux rate estimates below. 2074 assert(!_intra_sweep_timer.is_active(), "_intra_sweep_timer should be inactive"); 2075 if (_inter_sweep_timer.is_active()) { 2076 _inter_sweep_timer.stop(); 2077 // Note that we do not use this sample to update the _inter_sweep_estimate. 2078 _cmsGen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()), 2079 _inter_sweep_estimate.padded_average(), 2080 _intra_sweep_estimate.padded_average()); 2081 } 2082 2083 GenMarkSweep::invoke_at_safepoint(_cmsGen->level(), 2084 ref_processor(), clear_all_soft_refs); 2085 #ifdef ASSERT 2086 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 2087 size_t free_size = cms_space->free(); 2088 assert(free_size == 2089 pointer_delta(cms_space->end(), cms_space->compaction_top()) 2090 * HeapWordSize, 2091 "All the free space should be compacted into one chunk at top"); 2092 assert(cms_space->dictionary()->total_chunk_size( 2093 debug_only(cms_space->freelistLock())) == 0 || 2094 cms_space->totalSizeInIndexedFreeLists() == 0, 2095 "All the free space should be in a single chunk"); 2096 size_t num = cms_space->totalCount(); 2097 assert((free_size == 0 && num == 0) || 2098 (free_size > 0 && (num == 1 || num == 2)), 2099 "There should be at most 2 free chunks after compaction"); 2100 #endif // ASSERT 2101 _collectorState = Resetting; 2102 assert(_restart_addr == NULL, 2103 "Should have been NULL'd before baton was passed"); 2104 reset(false /* == !asynch */); 2105 _cmsGen->reset_after_compaction(); 2106 _concurrent_cycles_since_last_unload = 0; 2107 2108 // Clear any data recorded in the PLAB chunk arrays. 2109 if (_survivor_plab_array != NULL) { 2110 reset_survivor_plab_arrays(); 2111 } 2112 2113 // Adjust the per-size allocation stats for the next epoch. 2114 _cmsGen->cmsSpace()->endSweepFLCensus(sweep_count() /* fake */); 2115 // Restart the "inter sweep timer" for the next epoch. 2116 _inter_sweep_timer.reset(); 2117 _inter_sweep_timer.start(); 2118 2119 // Sample collection pause time and reset for collection interval. 2120 if (UseAdaptiveSizePolicy) { 2121 size_policy()->msc_collection_end(gch->gc_cause()); 2122 } 2123 2124 gc_timer->register_gc_end(); 2125 2126 gc_tracer->report_gc_end(gc_timer->gc_end(), gc_timer->time_partitions()); 2127 2128 // For a mark-sweep-compact, compute_new_size() will be called 2129 // in the heap's do_collection() method. 2130 } 2131 2132 // A work method used by the foreground collector to do 2133 // a mark-sweep, after taking over from a possibly on-going 2134 // concurrent mark-sweep collection. 2135 void CMSCollector::do_mark_sweep_work(bool clear_all_soft_refs, 2136 CollectorState first_state, bool should_start_over) { 2137 if (PrintGC && Verbose) { 2138 gclog_or_tty->print_cr("Pass concurrent collection to foreground " 2139 "collector with count %d", 2140 _full_gcs_since_conc_gc); 2141 } 2142 switch (_collectorState) { 2143 case Idling: 2144 if (first_state == Idling || should_start_over) { 2145 // The background GC was not active, or should 2146 // restarted from scratch; start the cycle. 2147 _collectorState = InitialMarking; 2148 } 2149 // If first_state was not Idling, then a background GC 2150 // was in progress and has now finished. No need to do it 2151 // again. Leave the state as Idling. 2152 break; 2153 case Precleaning: 2154 // In the foreground case don't do the precleaning since 2155 // it is not done concurrently and there is extra work 2156 // required. 2157 _collectorState = FinalMarking; 2158 } 2159 collect_in_foreground(clear_all_soft_refs, GenCollectedHeap::heap()->gc_cause()); 2160 2161 // For a mark-sweep, compute_new_size() will be called 2162 // in the heap's do_collection() method. 2163 } 2164 2165 2166 void CMSCollector::print_eden_and_survivor_chunk_arrays() { 2167 DefNewGeneration* dng = _young_gen->as_DefNewGeneration(); 2168 EdenSpace* eden_space = dng->eden(); 2169 ContiguousSpace* from_space = dng->from(); 2170 ContiguousSpace* to_space = dng->to(); 2171 // Eden 2172 if (_eden_chunk_array != NULL) { 2173 gclog_or_tty->print_cr("eden " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")", 2174 eden_space->bottom(), eden_space->top(), 2175 eden_space->end(), eden_space->capacity()); 2176 gclog_or_tty->print_cr("_eden_chunk_index=" SIZE_FORMAT ", " 2177 "_eden_chunk_capacity=" SIZE_FORMAT, 2178 _eden_chunk_index, _eden_chunk_capacity); 2179 for (size_t i = 0; i < _eden_chunk_index; i++) { 2180 gclog_or_tty->print_cr("_eden_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT, 2181 i, _eden_chunk_array[i]); 2182 } 2183 } 2184 // Survivor 2185 if (_survivor_chunk_array != NULL) { 2186 gclog_or_tty->print_cr("survivor " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")", 2187 from_space->bottom(), from_space->top(), 2188 from_space->end(), from_space->capacity()); 2189 gclog_or_tty->print_cr("_survivor_chunk_index=" SIZE_FORMAT ", " 2190 "_survivor_chunk_capacity=" SIZE_FORMAT, 2191 _survivor_chunk_index, _survivor_chunk_capacity); 2192 for (size_t i = 0; i < _survivor_chunk_index; i++) { 2193 gclog_or_tty->print_cr("_survivor_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT, 2194 i, _survivor_chunk_array[i]); 2195 } 2196 } 2197 } 2198 2199 void CMSCollector::getFreelistLocks() const { 2200 // Get locks for all free lists in all generations that this 2201 // collector is responsible for 2202 _cmsGen->freelistLock()->lock_without_safepoint_check(); 2203 } 2204 2205 void CMSCollector::releaseFreelistLocks() const { 2206 // Release locks for all free lists in all generations that this 2207 // collector is responsible for 2208 _cmsGen->freelistLock()->unlock(); 2209 } 2210 2211 bool CMSCollector::haveFreelistLocks() const { 2212 // Check locks for all free lists in all generations that this 2213 // collector is responsible for 2214 assert_lock_strong(_cmsGen->freelistLock()); 2215 PRODUCT_ONLY(ShouldNotReachHere()); 2216 return true; 2217 } 2218 2219 // A utility class that is used by the CMS collector to 2220 // temporarily "release" the foreground collector from its 2221 // usual obligation to wait for the background collector to 2222 // complete an ongoing phase before proceeding. 2223 class ReleaseForegroundGC: public StackObj { 2224 private: 2225 CMSCollector* _c; 2226 public: 2227 ReleaseForegroundGC(CMSCollector* c) : _c(c) { 2228 assert(_c->_foregroundGCShouldWait, "Else should not need to call"); 2229 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 2230 // allow a potentially blocked foreground collector to proceed 2231 _c->_foregroundGCShouldWait = false; 2232 if (_c->_foregroundGCIsActive) { 2233 CGC_lock->notify(); 2234 } 2235 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 2236 "Possible deadlock"); 2237 } 2238 2239 ~ReleaseForegroundGC() { 2240 assert(!_c->_foregroundGCShouldWait, "Usage protocol violation?"); 2241 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 2242 _c->_foregroundGCShouldWait = true; 2243 } 2244 }; 2245 2246 // There are separate collect_in_background and collect_in_foreground because of 2247 // the different locking requirements of the background collector and the 2248 // foreground collector. There was originally an attempt to share 2249 // one "collect" method between the background collector and the foreground 2250 // collector but the if-then-else required made it cleaner to have 2251 // separate methods. 2252 void CMSCollector::collect_in_background(bool clear_all_soft_refs, GCCause::Cause cause) { 2253 assert(Thread::current()->is_ConcurrentGC_thread(), 2254 "A CMS asynchronous collection is only allowed on a CMS thread."); 2255 2256 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2257 { 2258 bool safepoint_check = Mutex::_no_safepoint_check_flag; 2259 MutexLockerEx hl(Heap_lock, safepoint_check); 2260 FreelistLocker fll(this); 2261 MutexLockerEx x(CGC_lock, safepoint_check); 2262 if (_foregroundGCIsActive || !UseAsyncConcMarkSweepGC) { 2263 // The foreground collector is active or we're 2264 // not using asynchronous collections. Skip this 2265 // background collection. 2266 assert(!_foregroundGCShouldWait, "Should be clear"); 2267 return; 2268 } else { 2269 assert(_collectorState == Idling, "Should be idling before start."); 2270 _collectorState = InitialMarking; 2271 register_gc_start(cause); 2272 // Reset the expansion cause, now that we are about to begin 2273 // a new cycle. 2274 clear_expansion_cause(); 2275 2276 // Clear the MetaspaceGC flag since a concurrent collection 2277 // is starting but also clear it after the collection. 2278 MetaspaceGC::set_should_concurrent_collect(false); 2279 } 2280 // Decide if we want to enable class unloading as part of the 2281 // ensuing concurrent GC cycle. 2282 update_should_unload_classes(); 2283 _full_gc_requested = false; // acks all outstanding full gc requests 2284 _full_gc_cause = GCCause::_no_gc; 2285 // Signal that we are about to start a collection 2286 gch->increment_total_full_collections(); // ... starting a collection cycle 2287 _collection_count_start = gch->total_full_collections(); 2288 } 2289 2290 // Used for PrintGC 2291 size_t prev_used; 2292 if (PrintGC && Verbose) { 2293 prev_used = _cmsGen->used(); // XXXPERM 2294 } 2295 2296 // The change of the collection state is normally done at this level; 2297 // the exceptions are phases that are executed while the world is 2298 // stopped. For those phases the change of state is done while the 2299 // world is stopped. For baton passing purposes this allows the 2300 // background collector to finish the phase and change state atomically. 2301 // The foreground collector cannot wait on a phase that is done 2302 // while the world is stopped because the foreground collector already 2303 // has the world stopped and would deadlock. 2304 while (_collectorState != Idling) { 2305 if (TraceCMSState) { 2306 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " in CMS state %d", 2307 Thread::current(), _collectorState); 2308 } 2309 // The foreground collector 2310 // holds the Heap_lock throughout its collection. 2311 // holds the CMS token (but not the lock) 2312 // except while it is waiting for the background collector to yield. 2313 // 2314 // The foreground collector should be blocked (not for long) 2315 // if the background collector is about to start a phase 2316 // executed with world stopped. If the background 2317 // collector has already started such a phase, the 2318 // foreground collector is blocked waiting for the 2319 // Heap_lock. The stop-world phases (InitialMarking and FinalMarking) 2320 // are executed in the VM thread. 2321 // 2322 // The locking order is 2323 // PendingListLock (PLL) -- if applicable (FinalMarking) 2324 // Heap_lock (both this & PLL locked in VM_CMS_Operation::prologue()) 2325 // CMS token (claimed in 2326 // stop_world_and_do() --> 2327 // safepoint_synchronize() --> 2328 // CMSThread::synchronize()) 2329 2330 { 2331 // Check if the FG collector wants us to yield. 2332 CMSTokenSync x(true); // is cms thread 2333 if (waitForForegroundGC()) { 2334 // We yielded to a foreground GC, nothing more to be 2335 // done this round. 2336 assert(_foregroundGCShouldWait == false, "We set it to false in " 2337 "waitForForegroundGC()"); 2338 if (TraceCMSState) { 2339 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT 2340 " exiting collection CMS state %d", 2341 Thread::current(), _collectorState); 2342 } 2343 return; 2344 } else { 2345 // The background collector can run but check to see if the 2346 // foreground collector has done a collection while the 2347 // background collector was waiting to get the CGC_lock 2348 // above. If yes, break so that _foregroundGCShouldWait 2349 // is cleared before returning. 2350 if (_collectorState == Idling) { 2351 break; 2352 } 2353 } 2354 } 2355 2356 assert(_foregroundGCShouldWait, "Foreground collector, if active, " 2357 "should be waiting"); 2358 2359 switch (_collectorState) { 2360 case InitialMarking: 2361 { 2362 ReleaseForegroundGC x(this); 2363 stats().record_cms_begin(); 2364 VM_CMS_Initial_Mark initial_mark_op(this); 2365 VMThread::execute(&initial_mark_op); 2366 } 2367 // The collector state may be any legal state at this point 2368 // since the background collector may have yielded to the 2369 // foreground collector. 2370 break; 2371 case Marking: 2372 // initial marking in checkpointRootsInitialWork has been completed 2373 if (markFromRoots(true)) { // we were successful 2374 assert(_collectorState == Precleaning, "Collector state should " 2375 "have changed"); 2376 } else { 2377 assert(_foregroundGCIsActive, "Internal state inconsistency"); 2378 } 2379 break; 2380 case Precleaning: 2381 if (UseAdaptiveSizePolicy) { 2382 size_policy()->concurrent_precleaning_begin(); 2383 } 2384 // marking from roots in markFromRoots has been completed 2385 preclean(); 2386 if (UseAdaptiveSizePolicy) { 2387 size_policy()->concurrent_precleaning_end(); 2388 } 2389 assert(_collectorState == AbortablePreclean || 2390 _collectorState == FinalMarking, 2391 "Collector state should have changed"); 2392 break; 2393 case AbortablePreclean: 2394 if (UseAdaptiveSizePolicy) { 2395 size_policy()->concurrent_phases_resume(); 2396 } 2397 abortable_preclean(); 2398 if (UseAdaptiveSizePolicy) { 2399 size_policy()->concurrent_precleaning_end(); 2400 } 2401 assert(_collectorState == FinalMarking, "Collector state should " 2402 "have changed"); 2403 break; 2404 case FinalMarking: 2405 { 2406 ReleaseForegroundGC x(this); 2407 2408 VM_CMS_Final_Remark final_remark_op(this); 2409 VMThread::execute(&final_remark_op); 2410 } 2411 assert(_foregroundGCShouldWait, "block post-condition"); 2412 break; 2413 case Sweeping: 2414 if (UseAdaptiveSizePolicy) { 2415 size_policy()->concurrent_sweeping_begin(); 2416 } 2417 // final marking in checkpointRootsFinal has been completed 2418 sweep(true); 2419 assert(_collectorState == Resizing, "Collector state change " 2420 "to Resizing must be done under the free_list_lock"); 2421 _full_gcs_since_conc_gc = 0; 2422 2423 // Stop the timers for adaptive size policy for the concurrent phases 2424 if (UseAdaptiveSizePolicy) { 2425 size_policy()->concurrent_sweeping_end(); 2426 size_policy()->concurrent_phases_end(gch->gc_cause(), 2427 gch->prev_gen(_cmsGen)->capacity(), 2428 _cmsGen->free()); 2429 } 2430 2431 case Resizing: { 2432 // Sweeping has been completed... 2433 // At this point the background collection has completed. 2434 // Don't move the call to compute_new_size() down 2435 // into code that might be executed if the background 2436 // collection was preempted. 2437 { 2438 ReleaseForegroundGC x(this); // unblock FG collection 2439 MutexLockerEx y(Heap_lock, Mutex::_no_safepoint_check_flag); 2440 CMSTokenSync z(true); // not strictly needed. 2441 if (_collectorState == Resizing) { 2442 compute_new_size(); 2443 save_heap_summary(); 2444 _collectorState = Resetting; 2445 } else { 2446 assert(_collectorState == Idling, "The state should only change" 2447 " because the foreground collector has finished the collection"); 2448 } 2449 } 2450 break; 2451 } 2452 case Resetting: 2453 // CMS heap resizing has been completed 2454 reset(true); 2455 assert(_collectorState == Idling, "Collector state should " 2456 "have changed"); 2457 2458 MetaspaceGC::set_should_concurrent_collect(false); 2459 2460 stats().record_cms_end(); 2461 // Don't move the concurrent_phases_end() and compute_new_size() 2462 // calls to here because a preempted background collection 2463 // has it's state set to "Resetting". 2464 break; 2465 case Idling: 2466 default: 2467 ShouldNotReachHere(); 2468 break; 2469 } 2470 if (TraceCMSState) { 2471 gclog_or_tty->print_cr(" Thread " INTPTR_FORMAT " done - next CMS state %d", 2472 Thread::current(), _collectorState); 2473 } 2474 assert(_foregroundGCShouldWait, "block post-condition"); 2475 } 2476 2477 // Should this be in gc_epilogue? 2478 collector_policy()->counters()->update_counters(); 2479 2480 { 2481 // Clear _foregroundGCShouldWait and, in the event that the 2482 // foreground collector is waiting, notify it, before 2483 // returning. 2484 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 2485 _foregroundGCShouldWait = false; 2486 if (_foregroundGCIsActive) { 2487 CGC_lock->notify(); 2488 } 2489 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 2490 "Possible deadlock"); 2491 } 2492 if (TraceCMSState) { 2493 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT 2494 " exiting collection CMS state %d", 2495 Thread::current(), _collectorState); 2496 } 2497 if (PrintGC && Verbose) { 2498 _cmsGen->print_heap_change(prev_used); 2499 } 2500 } 2501 2502 void CMSCollector::register_foreground_gc_start(GCCause::Cause cause) { 2503 if (!_cms_start_registered) { 2504 register_gc_start(cause); 2505 } 2506 } 2507 2508 void CMSCollector::register_gc_start(GCCause::Cause cause) { 2509 _cms_start_registered = true; 2510 _gc_timer_cm->register_gc_start(); 2511 _gc_tracer_cm->report_gc_start(cause, _gc_timer_cm->gc_start()); 2512 } 2513 2514 void CMSCollector::register_gc_end() { 2515 if (_cms_start_registered) { 2516 report_heap_summary(GCWhen::AfterGC); 2517 2518 _gc_timer_cm->register_gc_end(); 2519 _gc_tracer_cm->report_gc_end(_gc_timer_cm->gc_end(), _gc_timer_cm->time_partitions()); 2520 _cms_start_registered = false; 2521 } 2522 } 2523 2524 void CMSCollector::save_heap_summary() { 2525 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2526 _last_heap_summary = gch->create_heap_summary(); 2527 _last_metaspace_summary = gch->create_metaspace_summary(); 2528 } 2529 2530 void CMSCollector::report_heap_summary(GCWhen::Type when) { 2531 _gc_tracer_cm->report_gc_heap_summary(when, _last_heap_summary); 2532 _gc_tracer_cm->report_metaspace_summary(when, _last_metaspace_summary); 2533 } 2534 2535 void CMSCollector::collect_in_foreground(bool clear_all_soft_refs, GCCause::Cause cause) { 2536 assert(_foregroundGCIsActive && !_foregroundGCShouldWait, 2537 "Foreground collector should be waiting, not executing"); 2538 assert(Thread::current()->is_VM_thread(), "A foreground collection" 2539 "may only be done by the VM Thread with the world stopped"); 2540 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 2541 "VM thread should have CMS token"); 2542 2543 // The gc id is created in register_foreground_gc_start if this collection is synchronous 2544 const GCId gc_id = _collectorState == InitialMarking ? GCId::peek() : _gc_tracer_cm->gc_id(); 2545 NOT_PRODUCT(GCTraceTime t("CMS:MS (foreground) ", PrintGCDetails && Verbose, 2546 true, NULL, gc_id);) 2547 if (UseAdaptiveSizePolicy) { 2548 size_policy()->ms_collection_begin(); 2549 } 2550 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact); 2551 2552 HandleMark hm; // Discard invalid handles created during verification 2553 2554 if (VerifyBeforeGC && 2555 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2556 Universe::verify(); 2557 } 2558 2559 // Snapshot the soft reference policy to be used in this collection cycle. 2560 ref_processor()->setup_policy(clear_all_soft_refs); 2561 2562 // Decide if class unloading should be done 2563 update_should_unload_classes(); 2564 2565 bool init_mark_was_synchronous = false; // until proven otherwise 2566 while (_collectorState != Idling) { 2567 if (TraceCMSState) { 2568 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " in CMS state %d", 2569 Thread::current(), _collectorState); 2570 } 2571 switch (_collectorState) { 2572 case InitialMarking: 2573 register_foreground_gc_start(cause); 2574 init_mark_was_synchronous = true; // fact to be exploited in re-mark 2575 checkpointRootsInitial(false); 2576 assert(_collectorState == Marking, "Collector state should have changed" 2577 " within checkpointRootsInitial()"); 2578 break; 2579 case Marking: 2580 // initial marking in checkpointRootsInitialWork has been completed 2581 if (VerifyDuringGC && 2582 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2583 Universe::verify("Verify before initial mark: "); 2584 } 2585 { 2586 bool res = markFromRoots(false); 2587 assert(res && _collectorState == FinalMarking, "Collector state should " 2588 "have changed"); 2589 break; 2590 } 2591 case FinalMarking: 2592 if (VerifyDuringGC && 2593 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2594 Universe::verify("Verify before re-mark: "); 2595 } 2596 checkpointRootsFinal(false, clear_all_soft_refs, 2597 init_mark_was_synchronous); 2598 assert(_collectorState == Sweeping, "Collector state should not " 2599 "have changed within checkpointRootsFinal()"); 2600 break; 2601 case Sweeping: 2602 // final marking in checkpointRootsFinal has been completed 2603 if (VerifyDuringGC && 2604 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2605 Universe::verify("Verify before sweep: "); 2606 } 2607 sweep(false); 2608 assert(_collectorState == Resizing, "Incorrect state"); 2609 break; 2610 case Resizing: { 2611 // Sweeping has been completed; the actual resize in this case 2612 // is done separately; nothing to be done in this state. 2613 _collectorState = Resetting; 2614 break; 2615 } 2616 case Resetting: 2617 // The heap has been resized. 2618 if (VerifyDuringGC && 2619 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2620 Universe::verify("Verify before reset: "); 2621 } 2622 save_heap_summary(); 2623 reset(false); 2624 assert(_collectorState == Idling, "Collector state should " 2625 "have changed"); 2626 break; 2627 case Precleaning: 2628 case AbortablePreclean: 2629 // Elide the preclean phase 2630 _collectorState = FinalMarking; 2631 break; 2632 default: 2633 ShouldNotReachHere(); 2634 } 2635 if (TraceCMSState) { 2636 gclog_or_tty->print_cr(" Thread " INTPTR_FORMAT " done - next CMS state %d", 2637 Thread::current(), _collectorState); 2638 } 2639 } 2640 2641 if (UseAdaptiveSizePolicy) { 2642 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2643 size_policy()->ms_collection_end(gch->gc_cause()); 2644 } 2645 2646 if (VerifyAfterGC && 2647 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 2648 Universe::verify(); 2649 } 2650 if (TraceCMSState) { 2651 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT 2652 " exiting collection CMS state %d", 2653 Thread::current(), _collectorState); 2654 } 2655 } 2656 2657 bool CMSCollector::waitForForegroundGC() { 2658 bool res = false; 2659 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 2660 "CMS thread should have CMS token"); 2661 // Block the foreground collector until the 2662 // background collectors decides whether to 2663 // yield. 2664 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 2665 _foregroundGCShouldWait = true; 2666 if (_foregroundGCIsActive) { 2667 // The background collector yields to the 2668 // foreground collector and returns a value 2669 // indicating that it has yielded. The foreground 2670 // collector can proceed. 2671 res = true; 2672 _foregroundGCShouldWait = false; 2673 ConcurrentMarkSweepThread::clear_CMS_flag( 2674 ConcurrentMarkSweepThread::CMS_cms_has_token); 2675 ConcurrentMarkSweepThread::set_CMS_flag( 2676 ConcurrentMarkSweepThread::CMS_cms_wants_token); 2677 // Get a possibly blocked foreground thread going 2678 CGC_lock->notify(); 2679 if (TraceCMSState) { 2680 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " waiting at CMS state %d", 2681 Thread::current(), _collectorState); 2682 } 2683 while (_foregroundGCIsActive) { 2684 CGC_lock->wait(Mutex::_no_safepoint_check_flag); 2685 } 2686 ConcurrentMarkSweepThread::set_CMS_flag( 2687 ConcurrentMarkSweepThread::CMS_cms_has_token); 2688 ConcurrentMarkSweepThread::clear_CMS_flag( 2689 ConcurrentMarkSweepThread::CMS_cms_wants_token); 2690 } 2691 if (TraceCMSState) { 2692 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " continuing at CMS state %d", 2693 Thread::current(), _collectorState); 2694 } 2695 return res; 2696 } 2697 2698 // Because of the need to lock the free lists and other structures in 2699 // the collector, common to all the generations that the collector is 2700 // collecting, we need the gc_prologues of individual CMS generations 2701 // delegate to their collector. It may have been simpler had the 2702 // current infrastructure allowed one to call a prologue on a 2703 // collector. In the absence of that we have the generation's 2704 // prologue delegate to the collector, which delegates back 2705 // some "local" work to a worker method in the individual generations 2706 // that it's responsible for collecting, while itself doing any 2707 // work common to all generations it's responsible for. A similar 2708 // comment applies to the gc_epilogue()'s. 2709 // The role of the varaible _between_prologue_and_epilogue is to 2710 // enforce the invocation protocol. 2711 void CMSCollector::gc_prologue(bool full) { 2712 // Call gc_prologue_work() for the CMSGen 2713 // we are responsible for. 2714 2715 // The following locking discipline assumes that we are only called 2716 // when the world is stopped. 2717 assert(SafepointSynchronize::is_at_safepoint(), "world is stopped assumption"); 2718 2719 // The CMSCollector prologue must call the gc_prologues for the 2720 // "generations" that it's responsible 2721 // for. 2722 2723 assert( Thread::current()->is_VM_thread() 2724 || ( CMSScavengeBeforeRemark 2725 && Thread::current()->is_ConcurrentGC_thread()), 2726 "Incorrect thread type for prologue execution"); 2727 2728 if (_between_prologue_and_epilogue) { 2729 // We have already been invoked; this is a gc_prologue delegation 2730 // from yet another CMS generation that we are responsible for, just 2731 // ignore it since all relevant work has already been done. 2732 return; 2733 } 2734 2735 // set a bit saying prologue has been called; cleared in epilogue 2736 _between_prologue_and_epilogue = true; 2737 // Claim locks for common data structures, then call gc_prologue_work() 2738 // for each CMSGen. 2739 2740 getFreelistLocks(); // gets free list locks on constituent spaces 2741 bitMapLock()->lock_without_safepoint_check(); 2742 2743 // Should call gc_prologue_work() for all cms gens we are responsible for 2744 bool duringMarking = _collectorState >= Marking 2745 && _collectorState < Sweeping; 2746 2747 // The young collections clear the modified oops state, which tells if 2748 // there are any modified oops in the class. The remark phase also needs 2749 // that information. Tell the young collection to save the union of all 2750 // modified klasses. 2751 if (duringMarking) { 2752 _ct->klass_rem_set()->set_accumulate_modified_oops(true); 2753 } 2754 2755 bool registerClosure = duringMarking; 2756 2757 ModUnionClosure* muc = CollectedHeap::use_parallel_gc_threads() ? 2758 &_modUnionClosurePar 2759 : &_modUnionClosure; 2760 _cmsGen->gc_prologue_work(full, registerClosure, muc); 2761 2762 if (!full) { 2763 stats().record_gc0_begin(); 2764 } 2765 } 2766 2767 void ConcurrentMarkSweepGeneration::gc_prologue(bool full) { 2768 2769 _capacity_at_prologue = capacity(); 2770 _used_at_prologue = used(); 2771 2772 // Delegate to CMScollector which knows how to coordinate between 2773 // this and any other CMS generations that it is responsible for 2774 // collecting. 2775 collector()->gc_prologue(full); 2776 } 2777 2778 // This is a "private" interface for use by this generation's CMSCollector. 2779 // Not to be called directly by any other entity (for instance, 2780 // GenCollectedHeap, which calls the "public" gc_prologue method above). 2781 void ConcurrentMarkSweepGeneration::gc_prologue_work(bool full, 2782 bool registerClosure, ModUnionClosure* modUnionClosure) { 2783 assert(!incremental_collection_failed(), "Shouldn't be set yet"); 2784 assert(cmsSpace()->preconsumptionDirtyCardClosure() == NULL, 2785 "Should be NULL"); 2786 if (registerClosure) { 2787 cmsSpace()->setPreconsumptionDirtyCardClosure(modUnionClosure); 2788 } 2789 cmsSpace()->gc_prologue(); 2790 // Clear stat counters 2791 NOT_PRODUCT( 2792 assert(_numObjectsPromoted == 0, "check"); 2793 assert(_numWordsPromoted == 0, "check"); 2794 if (Verbose && PrintGC) { 2795 gclog_or_tty->print("Allocated "SIZE_FORMAT" objects, " 2796 SIZE_FORMAT" bytes concurrently", 2797 _numObjectsAllocated, _numWordsAllocated*sizeof(HeapWord)); 2798 } 2799 _numObjectsAllocated = 0; 2800 _numWordsAllocated = 0; 2801 ) 2802 } 2803 2804 void CMSCollector::gc_epilogue(bool full) { 2805 // The following locking discipline assumes that we are only called 2806 // when the world is stopped. 2807 assert(SafepointSynchronize::is_at_safepoint(), 2808 "world is stopped assumption"); 2809 2810 // Currently the CMS epilogue (see CompactibleFreeListSpace) merely checks 2811 // if linear allocation blocks need to be appropriately marked to allow the 2812 // the blocks to be parsable. We also check here whether we need to nudge the 2813 // CMS collector thread to start a new cycle (if it's not already active). 2814 assert( Thread::current()->is_VM_thread() 2815 || ( CMSScavengeBeforeRemark 2816 && Thread::current()->is_ConcurrentGC_thread()), 2817 "Incorrect thread type for epilogue execution"); 2818 2819 if (!_between_prologue_and_epilogue) { 2820 // We have already been invoked; this is a gc_epilogue delegation 2821 // from yet another CMS generation that we are responsible for, just 2822 // ignore it since all relevant work has already been done. 2823 return; 2824 } 2825 assert(haveFreelistLocks(), "must have freelist locks"); 2826 assert_lock_strong(bitMapLock()); 2827 2828 _ct->klass_rem_set()->set_accumulate_modified_oops(false); 2829 2830 _cmsGen->gc_epilogue_work(full); 2831 2832 if (_collectorState == AbortablePreclean || _collectorState == Precleaning) { 2833 // in case sampling was not already enabled, enable it 2834 _start_sampling = true; 2835 } 2836 // reset _eden_chunk_array so sampling starts afresh 2837 _eden_chunk_index = 0; 2838 2839 size_t cms_used = _cmsGen->cmsSpace()->used(); 2840 2841 // update performance counters - this uses a special version of 2842 // update_counters() that allows the utilization to be passed as a 2843 // parameter, avoiding multiple calls to used(). 2844 // 2845 _cmsGen->update_counters(cms_used); 2846 2847 if (CMSIncrementalMode) { 2848 icms_update_allocation_limits(); 2849 } 2850 2851 bitMapLock()->unlock(); 2852 releaseFreelistLocks(); 2853 2854 if (!CleanChunkPoolAsync) { 2855 Chunk::clean_chunk_pool(); 2856 } 2857 2858 set_did_compact(false); 2859 _between_prologue_and_epilogue = false; // ready for next cycle 2860 } 2861 2862 void ConcurrentMarkSweepGeneration::gc_epilogue(bool full) { 2863 collector()->gc_epilogue(full); 2864 2865 // Also reset promotion tracking in par gc thread states. 2866 if (CollectedHeap::use_parallel_gc_threads()) { 2867 for (uint i = 0; i < ParallelGCThreads; i++) { 2868 _par_gc_thread_states[i]->promo.stopTrackingPromotions(i); 2869 } 2870 } 2871 } 2872 2873 void ConcurrentMarkSweepGeneration::gc_epilogue_work(bool full) { 2874 assert(!incremental_collection_failed(), "Should have been cleared"); 2875 cmsSpace()->setPreconsumptionDirtyCardClosure(NULL); 2876 cmsSpace()->gc_epilogue(); 2877 // Print stat counters 2878 NOT_PRODUCT( 2879 assert(_numObjectsAllocated == 0, "check"); 2880 assert(_numWordsAllocated == 0, "check"); 2881 if (Verbose && PrintGC) { 2882 gclog_or_tty->print("Promoted "SIZE_FORMAT" objects, " 2883 SIZE_FORMAT" bytes", 2884 _numObjectsPromoted, _numWordsPromoted*sizeof(HeapWord)); 2885 } 2886 _numObjectsPromoted = 0; 2887 _numWordsPromoted = 0; 2888 ) 2889 2890 if (PrintGC && Verbose) { 2891 // Call down the chain in contiguous_available needs the freelistLock 2892 // so print this out before releasing the freeListLock. 2893 gclog_or_tty->print(" Contiguous available "SIZE_FORMAT" bytes ", 2894 contiguous_available()); 2895 } 2896 } 2897 2898 #ifndef PRODUCT 2899 bool CMSCollector::have_cms_token() { 2900 Thread* thr = Thread::current(); 2901 if (thr->is_VM_thread()) { 2902 return ConcurrentMarkSweepThread::vm_thread_has_cms_token(); 2903 } else if (thr->is_ConcurrentGC_thread()) { 2904 return ConcurrentMarkSweepThread::cms_thread_has_cms_token(); 2905 } else if (thr->is_GC_task_thread()) { 2906 return ConcurrentMarkSweepThread::vm_thread_has_cms_token() && 2907 ParGCRareEvent_lock->owned_by_self(); 2908 } 2909 return false; 2910 } 2911 #endif 2912 2913 // Check reachability of the given heap address in CMS generation, 2914 // treating all other generations as roots. 2915 bool CMSCollector::is_cms_reachable(HeapWord* addr) { 2916 // We could "guarantee" below, rather than assert, but i'll 2917 // leave these as "asserts" so that an adventurous debugger 2918 // could try this in the product build provided some subset of 2919 // the conditions were met, provided they were intersted in the 2920 // results and knew that the computation below wouldn't interfere 2921 // with other concurrent computations mutating the structures 2922 // being read or written. 2923 assert(SafepointSynchronize::is_at_safepoint(), 2924 "Else mutations in object graph will make answer suspect"); 2925 assert(have_cms_token(), "Should hold cms token"); 2926 assert(haveFreelistLocks(), "must hold free list locks"); 2927 assert_lock_strong(bitMapLock()); 2928 2929 // Clear the marking bit map array before starting, but, just 2930 // for kicks, first report if the given address is already marked 2931 gclog_or_tty->print_cr("Start: Address 0x%x is%s marked", addr, 2932 _markBitMap.isMarked(addr) ? "" : " not"); 2933 2934 if (verify_after_remark()) { 2935 MutexLockerEx x(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag); 2936 bool result = verification_mark_bm()->isMarked(addr); 2937 gclog_or_tty->print_cr("TransitiveMark: Address 0x%x %s marked", addr, 2938 result ? "IS" : "is NOT"); 2939 return result; 2940 } else { 2941 gclog_or_tty->print_cr("Could not compute result"); 2942 return false; 2943 } 2944 } 2945 2946 2947 void 2948 CMSCollector::print_on_error(outputStream* st) { 2949 CMSCollector* collector = ConcurrentMarkSweepGeneration::_collector; 2950 if (collector != NULL) { 2951 CMSBitMap* bitmap = &collector->_markBitMap; 2952 st->print_cr("Marking Bits: (CMSBitMap*) " PTR_FORMAT, bitmap); 2953 bitmap->print_on_error(st, " Bits: "); 2954 2955 st->cr(); 2956 2957 CMSBitMap* mut_bitmap = &collector->_modUnionTable; 2958 st->print_cr("Mod Union Table: (CMSBitMap*) " PTR_FORMAT, mut_bitmap); 2959 mut_bitmap->print_on_error(st, " Bits: "); 2960 } 2961 } 2962 2963 //////////////////////////////////////////////////////// 2964 // CMS Verification Support 2965 //////////////////////////////////////////////////////// 2966 // Following the remark phase, the following invariant 2967 // should hold -- each object in the CMS heap which is 2968 // marked in markBitMap() should be marked in the verification_mark_bm(). 2969 2970 class VerifyMarkedClosure: public BitMapClosure { 2971 CMSBitMap* _marks; 2972 bool _failed; 2973 2974 public: 2975 VerifyMarkedClosure(CMSBitMap* bm): _marks(bm), _failed(false) {} 2976 2977 bool do_bit(size_t offset) { 2978 HeapWord* addr = _marks->offsetToHeapWord(offset); 2979 if (!_marks->isMarked(addr)) { 2980 oop(addr)->print_on(gclog_or_tty); 2981 gclog_or_tty->print_cr(" ("INTPTR_FORMAT" should have been marked)", addr); 2982 _failed = true; 2983 } 2984 return true; 2985 } 2986 2987 bool failed() { return _failed; } 2988 }; 2989 2990 bool CMSCollector::verify_after_remark(bool silent) { 2991 if (!silent) gclog_or_tty->print(" [Verifying CMS Marking... "); 2992 MutexLockerEx ml(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag); 2993 static bool init = false; 2994 2995 assert(SafepointSynchronize::is_at_safepoint(), 2996 "Else mutations in object graph will make answer suspect"); 2997 assert(have_cms_token(), 2998 "Else there may be mutual interference in use of " 2999 " verification data structures"); 3000 assert(_collectorState > Marking && _collectorState <= Sweeping, 3001 "Else marking info checked here may be obsolete"); 3002 assert(haveFreelistLocks(), "must hold free list locks"); 3003 assert_lock_strong(bitMapLock()); 3004 3005 3006 // Allocate marking bit map if not already allocated 3007 if (!init) { // first time 3008 if (!verification_mark_bm()->allocate(_span)) { 3009 return false; 3010 } 3011 init = true; 3012 } 3013 3014 assert(verification_mark_stack()->isEmpty(), "Should be empty"); 3015 3016 // Turn off refs discovery -- so we will be tracing through refs. 3017 // This is as intended, because by this time 3018 // GC must already have cleared any refs that need to be cleared, 3019 // and traced those that need to be marked; moreover, 3020 // the marking done here is not going to intefere in any 3021 // way with the marking information used by GC. 3022 NoRefDiscovery no_discovery(ref_processor()); 3023 3024 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;) 3025 3026 // Clear any marks from a previous round 3027 verification_mark_bm()->clear_all(); 3028 assert(verification_mark_stack()->isEmpty(), "markStack should be empty"); 3029 verify_work_stacks_empty(); 3030 3031 GenCollectedHeap* gch = GenCollectedHeap::heap(); 3032 gch->ensure_parsability(false); // fill TLABs, but no need to retire them 3033 // Update the saved marks which may affect the root scans. 3034 gch->save_marks(); 3035 3036 if (CMSRemarkVerifyVariant == 1) { 3037 // In this first variant of verification, we complete 3038 // all marking, then check if the new marks-verctor is 3039 // a subset of the CMS marks-vector. 3040 verify_after_remark_work_1(); 3041 } else if (CMSRemarkVerifyVariant == 2) { 3042 // In this second variant of verification, we flag an error 3043 // (i.e. an object reachable in the new marks-vector not reachable 3044 // in the CMS marks-vector) immediately, also indicating the 3045 // identify of an object (A) that references the unmarked object (B) -- 3046 // presumably, a mutation to A failed to be picked up by preclean/remark? 3047 verify_after_remark_work_2(); 3048 } else { 3049 warning("Unrecognized value %d for CMSRemarkVerifyVariant", 3050 CMSRemarkVerifyVariant); 3051 } 3052 if (!silent) gclog_or_tty->print(" done] "); 3053 return true; 3054 } 3055 3056 void CMSCollector::verify_after_remark_work_1() { 3057 ResourceMark rm; 3058 HandleMark hm; 3059 GenCollectedHeap* gch = GenCollectedHeap::heap(); 3060 3061 // Get a clear set of claim bits for the roots processing to work with. 3062 ClassLoaderDataGraph::clear_claimed_marks(); 3063 3064 // Mark from roots one level into CMS 3065 MarkRefsIntoClosure notOlder(_span, verification_mark_bm()); 3066 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 3067 3068 gch->gen_process_roots(_cmsGen->level(), 3069 true, // younger gens are roots 3070 true, // activate StrongRootsScope 3071 GenCollectedHeap::ScanningOption(roots_scanning_options()), 3072 should_unload_classes(), 3073 ¬Older, 3074 NULL, 3075 NULL); // SSS: Provide correct closure 3076 3077 // Now mark from the roots 3078 MarkFromRootsClosure markFromRootsClosure(this, _span, 3079 verification_mark_bm(), verification_mark_stack(), 3080 false /* don't yield */, true /* verifying */); 3081 assert(_restart_addr == NULL, "Expected pre-condition"); 3082 verification_mark_bm()->iterate(&markFromRootsClosure); 3083 while (_restart_addr != NULL) { 3084 // Deal with stack overflow: by restarting at the indicated 3085 // address. 3086 HeapWord* ra = _restart_addr; 3087 markFromRootsClosure.reset(ra); 3088 _restart_addr = NULL; 3089 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end()); 3090 } 3091 assert(verification_mark_stack()->isEmpty(), "Should have been drained"); 3092 verify_work_stacks_empty(); 3093 3094 // Marking completed -- now verify that each bit marked in 3095 // verification_mark_bm() is also marked in markBitMap(); flag all 3096 // errors by printing corresponding objects. 3097 VerifyMarkedClosure vcl(markBitMap()); 3098 verification_mark_bm()->iterate(&vcl); 3099 if (vcl.failed()) { 3100 gclog_or_tty->print("Verification failed"); 3101 Universe::heap()->print_on(gclog_or_tty); 3102 fatal("CMS: failed marking verification after remark"); 3103 } 3104 } 3105 3106 class VerifyKlassOopsKlassClosure : public KlassClosure { 3107 class VerifyKlassOopsClosure : public OopClosure { 3108 CMSBitMap* _bitmap; 3109 public: 3110 VerifyKlassOopsClosure(CMSBitMap* bitmap) : _bitmap(bitmap) { } 3111 void do_oop(oop* p) { guarantee(*p == NULL || _bitmap->isMarked((HeapWord*) *p), "Should be marked"); } 3112 void do_oop(narrowOop* p) { ShouldNotReachHere(); } 3113 } _oop_closure; 3114 public: 3115 VerifyKlassOopsKlassClosure(CMSBitMap* bitmap) : _oop_closure(bitmap) {} 3116 void do_klass(Klass* k) { 3117 k->oops_do(&_oop_closure); 3118 } 3119 }; 3120 3121 void CMSCollector::verify_after_remark_work_2() { 3122 ResourceMark rm; 3123 HandleMark hm; 3124 GenCollectedHeap* gch = GenCollectedHeap::heap(); 3125 3126 // Get a clear set of claim bits for the roots processing to work with. 3127 ClassLoaderDataGraph::clear_claimed_marks(); 3128 3129 // Mark from roots one level into CMS 3130 MarkRefsIntoVerifyClosure notOlder(_span, verification_mark_bm(), 3131 markBitMap()); 3132 CLDToOopClosure cld_closure(¬Older, true); 3133 3134 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 3135 3136 gch->gen_process_roots(_cmsGen->level(), 3137 true, // younger gens are roots 3138 true, // activate StrongRootsScope 3139 GenCollectedHeap::ScanningOption(roots_scanning_options()), 3140 should_unload_classes(), 3141 ¬Older, 3142 NULL, 3143 &cld_closure); 3144 3145 // Now mark from the roots 3146 MarkFromRootsVerifyClosure markFromRootsClosure(this, _span, 3147 verification_mark_bm(), markBitMap(), verification_mark_stack()); 3148 assert(_restart_addr == NULL, "Expected pre-condition"); 3149 verification_mark_bm()->iterate(&markFromRootsClosure); 3150 while (_restart_addr != NULL) { 3151 // Deal with stack overflow: by restarting at the indicated 3152 // address. 3153 HeapWord* ra = _restart_addr; 3154 markFromRootsClosure.reset(ra); 3155 _restart_addr = NULL; 3156 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end()); 3157 } 3158 assert(verification_mark_stack()->isEmpty(), "Should have been drained"); 3159 verify_work_stacks_empty(); 3160 3161 VerifyKlassOopsKlassClosure verify_klass_oops(verification_mark_bm()); 3162 ClassLoaderDataGraph::classes_do(&verify_klass_oops); 3163 3164 // Marking completed -- now verify that each bit marked in 3165 // verification_mark_bm() is also marked in markBitMap(); flag all 3166 // errors by printing corresponding objects. 3167 VerifyMarkedClosure vcl(markBitMap()); 3168 verification_mark_bm()->iterate(&vcl); 3169 assert(!vcl.failed(), "Else verification above should not have succeeded"); 3170 } 3171 3172 void ConcurrentMarkSweepGeneration::save_marks() { 3173 // delegate to CMS space 3174 cmsSpace()->save_marks(); 3175 for (uint i = 0; i < ParallelGCThreads; i++) { 3176 _par_gc_thread_states[i]->promo.startTrackingPromotions(); 3177 } 3178 } 3179 3180 bool ConcurrentMarkSweepGeneration::no_allocs_since_save_marks() { 3181 return cmsSpace()->no_allocs_since_save_marks(); 3182 } 3183 3184 #define CMS_SINCE_SAVE_MARKS_DEFN(OopClosureType, nv_suffix) \ 3185 \ 3186 void ConcurrentMarkSweepGeneration:: \ 3187 oop_since_save_marks_iterate##nv_suffix(OopClosureType* cl) { \ 3188 cl->set_generation(this); \ 3189 cmsSpace()->oop_since_save_marks_iterate##nv_suffix(cl); \ 3190 cl->reset_generation(); \ 3191 save_marks(); \ 3192 } 3193 3194 ALL_SINCE_SAVE_MARKS_CLOSURES(CMS_SINCE_SAVE_MARKS_DEFN) 3195 3196 void 3197 ConcurrentMarkSweepGeneration::younger_refs_iterate(OopsInGenClosure* cl) { 3198 cl->set_generation(this); 3199 younger_refs_in_space_iterate(_cmsSpace, cl); 3200 cl->reset_generation(); 3201 } 3202 3203 void 3204 ConcurrentMarkSweepGeneration::oop_iterate(ExtendedOopClosure* cl) { 3205 if (freelistLock()->owned_by_self()) { 3206 Generation::oop_iterate(cl); 3207 } else { 3208 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 3209 Generation::oop_iterate(cl); 3210 } 3211 } 3212 3213 void 3214 ConcurrentMarkSweepGeneration::object_iterate(ObjectClosure* cl) { 3215 if (freelistLock()->owned_by_self()) { 3216 Generation::object_iterate(cl); 3217 } else { 3218 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 3219 Generation::object_iterate(cl); 3220 } 3221 } 3222 3223 void 3224 ConcurrentMarkSweepGeneration::safe_object_iterate(ObjectClosure* cl) { 3225 if (freelistLock()->owned_by_self()) { 3226 Generation::safe_object_iterate(cl); 3227 } else { 3228 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 3229 Generation::safe_object_iterate(cl); 3230 } 3231 } 3232 3233 void 3234 ConcurrentMarkSweepGeneration::post_compact() { 3235 } 3236 3237 void 3238 ConcurrentMarkSweepGeneration::prepare_for_verify() { 3239 // Fix the linear allocation blocks to look like free blocks. 3240 3241 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those 3242 // are not called when the heap is verified during universe initialization and 3243 // at vm shutdown. 3244 if (freelistLock()->owned_by_self()) { 3245 cmsSpace()->prepare_for_verify(); 3246 } else { 3247 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag); 3248 cmsSpace()->prepare_for_verify(); 3249 } 3250 } 3251 3252 void 3253 ConcurrentMarkSweepGeneration::verify() { 3254 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those 3255 // are not called when the heap is verified during universe initialization and 3256 // at vm shutdown. 3257 if (freelistLock()->owned_by_self()) { 3258 cmsSpace()->verify(); 3259 } else { 3260 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag); 3261 cmsSpace()->verify(); 3262 } 3263 } 3264 3265 void CMSCollector::verify() { 3266 _cmsGen->verify(); 3267 } 3268 3269 #ifndef PRODUCT 3270 bool CMSCollector::overflow_list_is_empty() const { 3271 assert(_num_par_pushes >= 0, "Inconsistency"); 3272 if (_overflow_list == NULL) { 3273 assert(_num_par_pushes == 0, "Inconsistency"); 3274 } 3275 return _overflow_list == NULL; 3276 } 3277 3278 // The methods verify_work_stacks_empty() and verify_overflow_empty() 3279 // merely consolidate assertion checks that appear to occur together frequently. 3280 void CMSCollector::verify_work_stacks_empty() const { 3281 assert(_markStack.isEmpty(), "Marking stack should be empty"); 3282 assert(overflow_list_is_empty(), "Overflow list should be empty"); 3283 } 3284 3285 void CMSCollector::verify_overflow_empty() const { 3286 assert(overflow_list_is_empty(), "Overflow list should be empty"); 3287 assert(no_preserved_marks(), "No preserved marks"); 3288 } 3289 #endif // PRODUCT 3290 3291 // Decide if we want to enable class unloading as part of the 3292 // ensuing concurrent GC cycle. We will collect and 3293 // unload classes if it's the case that: 3294 // (1) an explicit gc request has been made and the flag 3295 // ExplicitGCInvokesConcurrentAndUnloadsClasses is set, OR 3296 // (2) (a) class unloading is enabled at the command line, and 3297 // (b) old gen is getting really full 3298 // NOTE: Provided there is no change in the state of the heap between 3299 // calls to this method, it should have idempotent results. Moreover, 3300 // its results should be monotonically increasing (i.e. going from 0 to 1, 3301 // but not 1 to 0) between successive calls between which the heap was 3302 // not collected. For the implementation below, it must thus rely on 3303 // the property that concurrent_cycles_since_last_unload() 3304 // will not decrease unless a collection cycle happened and that 3305 // _cmsGen->is_too_full() are 3306 // themselves also monotonic in that sense. See check_monotonicity() 3307 // below. 3308 void CMSCollector::update_should_unload_classes() { 3309 _should_unload_classes = false; 3310 // Condition 1 above 3311 if (_full_gc_requested && ExplicitGCInvokesConcurrentAndUnloadsClasses) { 3312 _should_unload_classes = true; 3313 } else if (CMSClassUnloadingEnabled) { // Condition 2.a above 3314 // Disjuncts 2.b.(i,ii,iii) above 3315 _should_unload_classes = (concurrent_cycles_since_last_unload() >= 3316 CMSClassUnloadingMaxInterval) 3317 || _cmsGen->is_too_full(); 3318 } 3319 } 3320 3321 bool ConcurrentMarkSweepGeneration::is_too_full() const { 3322 bool res = should_concurrent_collect(); 3323 res = res && (occupancy() > (double)CMSIsTooFullPercentage/100.0); 3324 return res; 3325 } 3326 3327 void CMSCollector::setup_cms_unloading_and_verification_state() { 3328 const bool should_verify = VerifyBeforeGC || VerifyAfterGC || VerifyDuringGC 3329 || VerifyBeforeExit; 3330 const int rso = GenCollectedHeap::SO_AllCodeCache; 3331 3332 // We set the proper root for this CMS cycle here. 3333 if (should_unload_classes()) { // Should unload classes this cycle 3334 remove_root_scanning_option(rso); // Shrink the root set appropriately 3335 set_verifying(should_verify); // Set verification state for this cycle 3336 return; // Nothing else needs to be done at this time 3337 } 3338 3339 // Not unloading classes this cycle 3340 assert(!should_unload_classes(), "Inconsitency!"); 3341 3342 if ((!verifying() || unloaded_classes_last_cycle()) && should_verify) { 3343 // Include symbols, strings and code cache elements to prevent their resurrection. 3344 add_root_scanning_option(rso); 3345 set_verifying(true); 3346 } else if (verifying() && !should_verify) { 3347 // We were verifying, but some verification flags got disabled. 3348 set_verifying(false); 3349 // Exclude symbols, strings and code cache elements from root scanning to 3350 // reduce IM and RM pauses. 3351 remove_root_scanning_option(rso); 3352 } 3353 } 3354 3355 3356 #ifndef PRODUCT 3357 HeapWord* CMSCollector::block_start(const void* p) const { 3358 const HeapWord* addr = (HeapWord*)p; 3359 if (_span.contains(p)) { 3360 if (_cmsGen->cmsSpace()->is_in_reserved(addr)) { 3361 return _cmsGen->cmsSpace()->block_start(p); 3362 } 3363 } 3364 return NULL; 3365 } 3366 #endif 3367 3368 HeapWord* 3369 ConcurrentMarkSweepGeneration::expand_and_allocate(size_t word_size, 3370 bool tlab, 3371 bool parallel) { 3372 CMSSynchronousYieldRequest yr; 3373 assert(!tlab, "Can't deal with TLAB allocation"); 3374 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 3375 expand(word_size*HeapWordSize, MinHeapDeltaBytes, 3376 CMSExpansionCause::_satisfy_allocation); 3377 if (GCExpandToAllocateDelayMillis > 0) { 3378 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 3379 } 3380 return have_lock_and_allocate(word_size, tlab); 3381 } 3382 3383 // YSR: All of this generation expansion/shrinking stuff is an exact copy of 3384 // OneContigSpaceCardGeneration, which makes me wonder if we should move this 3385 // to CardGeneration and share it... 3386 bool ConcurrentMarkSweepGeneration::expand(size_t bytes, size_t expand_bytes) { 3387 return CardGeneration::expand(bytes, expand_bytes); 3388 } 3389 3390 void ConcurrentMarkSweepGeneration::expand(size_t bytes, size_t expand_bytes, 3391 CMSExpansionCause::Cause cause) 3392 { 3393 3394 bool success = expand(bytes, expand_bytes); 3395 3396 // remember why we expanded; this information is used 3397 // by shouldConcurrentCollect() when making decisions on whether to start 3398 // a new CMS cycle. 3399 if (success) { 3400 set_expansion_cause(cause); 3401 if (PrintGCDetails && Verbose) { 3402 gclog_or_tty->print_cr("Expanded CMS gen for %s", 3403 CMSExpansionCause::to_string(cause)); 3404 } 3405 } 3406 } 3407 3408 HeapWord* ConcurrentMarkSweepGeneration::expand_and_par_lab_allocate(CMSParGCThreadState* ps, size_t word_sz) { 3409 HeapWord* res = NULL; 3410 MutexLocker x(ParGCRareEvent_lock); 3411 while (true) { 3412 // Expansion by some other thread might make alloc OK now: 3413 res = ps->lab.alloc(word_sz); 3414 if (res != NULL) return res; 3415 // If there's not enough expansion space available, give up. 3416 if (_virtual_space.uncommitted_size() < (word_sz * HeapWordSize)) { 3417 return NULL; 3418 } 3419 // Otherwise, we try expansion. 3420 expand(word_sz*HeapWordSize, MinHeapDeltaBytes, 3421 CMSExpansionCause::_allocate_par_lab); 3422 // Now go around the loop and try alloc again; 3423 // A competing par_promote might beat us to the expansion space, 3424 // so we may go around the loop again if promotion fails agaion. 3425 if (GCExpandToAllocateDelayMillis > 0) { 3426 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 3427 } 3428 } 3429 } 3430 3431 3432 bool ConcurrentMarkSweepGeneration::expand_and_ensure_spooling_space( 3433 PromotionInfo* promo) { 3434 MutexLocker x(ParGCRareEvent_lock); 3435 size_t refill_size_bytes = promo->refillSize() * HeapWordSize; 3436 while (true) { 3437 // Expansion by some other thread might make alloc OK now: 3438 if (promo->ensure_spooling_space()) { 3439 assert(promo->has_spooling_space(), 3440 "Post-condition of successful ensure_spooling_space()"); 3441 return true; 3442 } 3443 // If there's not enough expansion space available, give up. 3444 if (_virtual_space.uncommitted_size() < refill_size_bytes) { 3445 return false; 3446 } 3447 // Otherwise, we try expansion. 3448 expand(refill_size_bytes, MinHeapDeltaBytes, 3449 CMSExpansionCause::_allocate_par_spooling_space); 3450 // Now go around the loop and try alloc again; 3451 // A competing allocation might beat us to the expansion space, 3452 // so we may go around the loop again if allocation fails again. 3453 if (GCExpandToAllocateDelayMillis > 0) { 3454 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 3455 } 3456 } 3457 } 3458 3459 3460 void ConcurrentMarkSweepGeneration::shrink_by(size_t bytes) { 3461 assert_locked_or_safepoint(ExpandHeap_lock); 3462 // Shrink committed space 3463 _virtual_space.shrink_by(bytes); 3464 // Shrink space; this also shrinks the space's BOT 3465 _cmsSpace->set_end((HeapWord*) _virtual_space.high()); 3466 size_t new_word_size = heap_word_size(_cmsSpace->capacity()); 3467 // Shrink the shared block offset array 3468 _bts->resize(new_word_size); 3469 MemRegion mr(_cmsSpace->bottom(), new_word_size); 3470 // Shrink the card table 3471 Universe::heap()->barrier_set()->resize_covered_region(mr); 3472 3473 if (Verbose && PrintGC) { 3474 size_t new_mem_size = _virtual_space.committed_size(); 3475 size_t old_mem_size = new_mem_size + bytes; 3476 gclog_or_tty->print_cr("Shrinking %s from " SIZE_FORMAT "K to " SIZE_FORMAT "K", 3477 name(), old_mem_size/K, new_mem_size/K); 3478 } 3479 } 3480 3481 void ConcurrentMarkSweepGeneration::shrink(size_t bytes) { 3482 assert_locked_or_safepoint(Heap_lock); 3483 size_t size = ReservedSpace::page_align_size_down(bytes); 3484 // Only shrink if a compaction was done so that all the free space 3485 // in the generation is in a contiguous block at the end. 3486 if (size > 0 && did_compact()) { 3487 shrink_by(size); 3488 } 3489 } 3490 3491 bool ConcurrentMarkSweepGeneration::grow_by(size_t bytes) { 3492 assert_locked_or_safepoint(Heap_lock); 3493 bool result = _virtual_space.expand_by(bytes); 3494 if (result) { 3495 size_t new_word_size = 3496 heap_word_size(_virtual_space.committed_size()); 3497 MemRegion mr(_cmsSpace->bottom(), new_word_size); 3498 _bts->resize(new_word_size); // resize the block offset shared array 3499 Universe::heap()->barrier_set()->resize_covered_region(mr); 3500 // Hmmmm... why doesn't CFLS::set_end verify locking? 3501 // This is quite ugly; FIX ME XXX 3502 _cmsSpace->assert_locked(freelistLock()); 3503 _cmsSpace->set_end((HeapWord*)_virtual_space.high()); 3504 3505 // update the space and generation capacity counters 3506 if (UsePerfData) { 3507 _space_counters->update_capacity(); 3508 _gen_counters->update_all(); 3509 } 3510 3511 if (Verbose && PrintGC) { 3512 size_t new_mem_size = _virtual_space.committed_size(); 3513 size_t old_mem_size = new_mem_size - bytes; 3514 gclog_or_tty->print_cr("Expanding %s from " SIZE_FORMAT "K by " SIZE_FORMAT "K to " SIZE_FORMAT "K", 3515 name(), old_mem_size/K, bytes/K, new_mem_size/K); 3516 } 3517 } 3518 return result; 3519 } 3520 3521 bool ConcurrentMarkSweepGeneration::grow_to_reserved() { 3522 assert_locked_or_safepoint(Heap_lock); 3523 bool success = true; 3524 const size_t remaining_bytes = _virtual_space.uncommitted_size(); 3525 if (remaining_bytes > 0) { 3526 success = grow_by(remaining_bytes); 3527 DEBUG_ONLY(if (!success) warning("grow to reserved failed");) 3528 } 3529 return success; 3530 } 3531 3532 void ConcurrentMarkSweepGeneration::shrink_free_list_by(size_t bytes) { 3533 assert_locked_or_safepoint(Heap_lock); 3534 assert_lock_strong(freelistLock()); 3535 if (PrintGCDetails && Verbose) { 3536 warning("Shrinking of CMS not yet implemented"); 3537 } 3538 return; 3539 } 3540 3541 3542 // Simple ctor/dtor wrapper for accounting & timer chores around concurrent 3543 // phases. 3544 class CMSPhaseAccounting: public StackObj { 3545 public: 3546 CMSPhaseAccounting(CMSCollector *collector, 3547 const char *phase, 3548 const GCId gc_id, 3549 bool print_cr = true); 3550 ~CMSPhaseAccounting(); 3551 3552 private: 3553 CMSCollector *_collector; 3554 const char *_phase; 3555 elapsedTimer _wallclock; 3556 bool _print_cr; 3557 const GCId _gc_id; 3558 3559 public: 3560 // Not MT-safe; so do not pass around these StackObj's 3561 // where they may be accessed by other threads. 3562 jlong wallclock_millis() { 3563 assert(_wallclock.is_active(), "Wall clock should not stop"); 3564 _wallclock.stop(); // to record time 3565 jlong ret = _wallclock.milliseconds(); 3566 _wallclock.start(); // restart 3567 return ret; 3568 } 3569 }; 3570 3571 CMSPhaseAccounting::CMSPhaseAccounting(CMSCollector *collector, 3572 const char *phase, 3573 const GCId gc_id, 3574 bool print_cr) : 3575 _collector(collector), _phase(phase), _print_cr(print_cr), _gc_id(gc_id) { 3576 3577 if (PrintCMSStatistics != 0) { 3578 _collector->resetYields(); 3579 } 3580 if (PrintGCDetails) { 3581 gclog_or_tty->gclog_stamp(_gc_id); 3582 gclog_or_tty->print_cr("[%s-concurrent-%s-start]", 3583 _collector->cmsGen()->short_name(), _phase); 3584 } 3585 _collector->resetTimer(); 3586 _wallclock.start(); 3587 _collector->startTimer(); 3588 } 3589 3590 CMSPhaseAccounting::~CMSPhaseAccounting() { 3591 assert(_wallclock.is_active(), "Wall clock should not have stopped"); 3592 _collector->stopTimer(); 3593 _wallclock.stop(); 3594 if (PrintGCDetails) { 3595 gclog_or_tty->gclog_stamp(_gc_id); 3596 gclog_or_tty->print("[%s-concurrent-%s: %3.3f/%3.3f secs]", 3597 _collector->cmsGen()->short_name(), 3598 _phase, _collector->timerValue(), _wallclock.seconds()); 3599 if (_print_cr) { 3600 gclog_or_tty->cr(); 3601 } 3602 if (PrintCMSStatistics != 0) { 3603 gclog_or_tty->print_cr(" (CMS-concurrent-%s yielded %d times)", _phase, 3604 _collector->yields()); 3605 } 3606 } 3607 } 3608 3609 // CMS work 3610 3611 // The common parts of CMSParInitialMarkTask and CMSParRemarkTask. 3612 class CMSParMarkTask : public AbstractGangTask { 3613 protected: 3614 CMSCollector* _collector; 3615 int _n_workers; 3616 CMSParMarkTask(const char* name, CMSCollector* collector, int n_workers) : 3617 AbstractGangTask(name), 3618 _collector(collector), 3619 _n_workers(n_workers) {} 3620 // Work method in support of parallel rescan ... of young gen spaces 3621 void do_young_space_rescan(uint worker_id, OopsInGenClosure* cl, 3622 ContiguousSpace* space, 3623 HeapWord** chunk_array, size_t chunk_top); 3624 void work_on_young_gen_roots(uint worker_id, OopsInGenClosure* cl); 3625 }; 3626 3627 // Parallel initial mark task 3628 class CMSParInitialMarkTask: public CMSParMarkTask { 3629 public: 3630 CMSParInitialMarkTask(CMSCollector* collector, int n_workers) : 3631 CMSParMarkTask("Scan roots and young gen for initial mark in parallel", 3632 collector, n_workers) {} 3633 void work(uint worker_id); 3634 }; 3635 3636 // Checkpoint the roots into this generation from outside 3637 // this generation. [Note this initial checkpoint need only 3638 // be approximate -- we'll do a catch up phase subsequently.] 3639 void CMSCollector::checkpointRootsInitial(bool asynch) { 3640 assert(_collectorState == InitialMarking, "Wrong collector state"); 3641 check_correct_thread_executing(); 3642 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 3643 3644 save_heap_summary(); 3645 report_heap_summary(GCWhen::BeforeGC); 3646 3647 ReferenceProcessor* rp = ref_processor(); 3648 SpecializationStats::clear(); 3649 assert(_restart_addr == NULL, "Control point invariant"); 3650 if (asynch) { 3651 // acquire locks for subsequent manipulations 3652 MutexLockerEx x(bitMapLock(), 3653 Mutex::_no_safepoint_check_flag); 3654 checkpointRootsInitialWork(asynch); 3655 // enable ("weak") refs discovery 3656 rp->enable_discovery(true /*verify_disabled*/, true /*check_no_refs*/); 3657 _collectorState = Marking; 3658 } else { 3659 // (Weak) Refs discovery: this is controlled from genCollectedHeap::do_collection 3660 // which recognizes if we are a CMS generation, and doesn't try to turn on 3661 // discovery; verify that they aren't meddling. 3662 assert(!rp->discovery_is_atomic(), 3663 "incorrect setting of discovery predicate"); 3664 assert(!rp->discovery_enabled(), "genCollectedHeap shouldn't control " 3665 "ref discovery for this generation kind"); 3666 // already have locks 3667 checkpointRootsInitialWork(asynch); 3668 // now enable ("weak") refs discovery 3669 rp->enable_discovery(true /*verify_disabled*/, false /*verify_no_refs*/); 3670 _collectorState = Marking; 3671 } 3672 SpecializationStats::print(); 3673 } 3674 3675 void CMSCollector::checkpointRootsInitialWork(bool asynch) { 3676 assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped"); 3677 assert(_collectorState == InitialMarking, "just checking"); 3678 3679 // If there has not been a GC[n-1] since last GC[n] cycle completed, 3680 // precede our marking with a collection of all 3681 // younger generations to keep floating garbage to a minimum. 3682 // XXX: we won't do this for now -- it's an optimization to be done later. 3683 3684 // already have locks 3685 assert_lock_strong(bitMapLock()); 3686 assert(_markBitMap.isAllClear(), "was reset at end of previous cycle"); 3687 3688 // Setup the verification and class unloading state for this 3689 // CMS collection cycle. 3690 setup_cms_unloading_and_verification_state(); 3691 3692 NOT_PRODUCT(GCTraceTime t("\ncheckpointRootsInitialWork", 3693 PrintGCDetails && Verbose, true, _gc_timer_cm, _gc_tracer_cm->gc_id());) 3694 if (UseAdaptiveSizePolicy) { 3695 size_policy()->checkpoint_roots_initial_begin(); 3696 } 3697 3698 // Reset all the PLAB chunk arrays if necessary. 3699 if (_survivor_plab_array != NULL && !CMSPLABRecordAlways) { 3700 reset_survivor_plab_arrays(); 3701 } 3702 3703 ResourceMark rm; 3704 HandleMark hm; 3705 3706 MarkRefsIntoClosure notOlder(_span, &_markBitMap); 3707 GenCollectedHeap* gch = GenCollectedHeap::heap(); 3708 3709 verify_work_stacks_empty(); 3710 verify_overflow_empty(); 3711 3712 gch->ensure_parsability(false); // fill TLABs, but no need to retire them 3713 // Update the saved marks which may affect the root scans. 3714 gch->save_marks(); 3715 3716 // weak reference processing has not started yet. 3717 ref_processor()->set_enqueuing_is_done(false); 3718 3719 // Need to remember all newly created CLDs, 3720 // so that we can guarantee that the remark finds them. 3721 ClassLoaderDataGraph::remember_new_clds(true); 3722 3723 // Whenever a CLD is found, it will be claimed before proceeding to mark 3724 // the klasses. The claimed marks need to be cleared before marking starts. 3725 ClassLoaderDataGraph::clear_claimed_marks(); 3726 3727 if (CMSPrintEdenSurvivorChunks) { 3728 print_eden_and_survivor_chunk_arrays(); 3729 } 3730 3731 { 3732 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;) 3733 if (CMSParallelInitialMarkEnabled && CollectedHeap::use_parallel_gc_threads()) { 3734 // The parallel version. 3735 FlexibleWorkGang* workers = gch->workers(); 3736 assert(workers != NULL, "Need parallel worker threads."); 3737 int n_workers = workers->active_workers(); 3738 CMSParInitialMarkTask tsk(this, n_workers); 3739 gch->set_par_threads(n_workers); 3740 initialize_sequential_subtasks_for_young_gen_rescan(n_workers); 3741 if (n_workers > 1) { 3742 GenCollectedHeap::StrongRootsScope srs(gch); 3743 workers->run_task(&tsk); 3744 } else { 3745 GenCollectedHeap::StrongRootsScope srs(gch); 3746 tsk.work(0); 3747 } 3748 gch->set_par_threads(0); 3749 } else { 3750 // The serial version. 3751 CLDToOopClosure cld_closure(¬Older, true); 3752 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 3753 gch->gen_process_roots(_cmsGen->level(), 3754 true, // younger gens are roots 3755 true, // activate StrongRootsScope 3756 GenCollectedHeap::ScanningOption(roots_scanning_options()), 3757 should_unload_classes(), 3758 ¬Older, 3759 NULL, 3760 &cld_closure); 3761 } 3762 } 3763 3764 // Clear mod-union table; it will be dirtied in the prologue of 3765 // CMS generation per each younger generation collection. 3766 3767 assert(_modUnionTable.isAllClear(), 3768 "Was cleared in most recent final checkpoint phase" 3769 " or no bits are set in the gc_prologue before the start of the next " 3770 "subsequent marking phase."); 3771 3772 assert(_ct->klass_rem_set()->mod_union_is_clear(), "Must be"); 3773 3774 // Save the end of the used_region of the constituent generations 3775 // to be used to limit the extent of sweep in each generation. 3776 save_sweep_limits(); 3777 if (UseAdaptiveSizePolicy) { 3778 size_policy()->checkpoint_roots_initial_end(gch->gc_cause()); 3779 } 3780 verify_overflow_empty(); 3781 } 3782 3783 bool CMSCollector::markFromRoots(bool asynch) { 3784 // we might be tempted to assert that: 3785 // assert(asynch == !SafepointSynchronize::is_at_safepoint(), 3786 // "inconsistent argument?"); 3787 // However that wouldn't be right, because it's possible that 3788 // a safepoint is indeed in progress as a younger generation 3789 // stop-the-world GC happens even as we mark in this generation. 3790 assert(_collectorState == Marking, "inconsistent state?"); 3791 check_correct_thread_executing(); 3792 verify_overflow_empty(); 3793 3794 bool res; 3795 if (asynch) { 3796 3797 // Start the timers for adaptive size policy for the concurrent phases 3798 // Do it here so that the foreground MS can use the concurrent 3799 // timer since a foreground MS might has the sweep done concurrently 3800 // or STW. 3801 if (UseAdaptiveSizePolicy) { 3802 size_policy()->concurrent_marking_begin(); 3803 } 3804 3805 // Weak ref discovery note: We may be discovering weak 3806 // refs in this generation concurrent (but interleaved) with 3807 // weak ref discovery by a younger generation collector. 3808 3809 CMSTokenSyncWithLocks ts(true, bitMapLock()); 3810 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 3811 CMSPhaseAccounting pa(this, "mark", _gc_tracer_cm->gc_id(), !PrintGCDetails); 3812 res = markFromRootsWork(asynch); 3813 if (res) { 3814 _collectorState = Precleaning; 3815 } else { // We failed and a foreground collection wants to take over 3816 assert(_foregroundGCIsActive, "internal state inconsistency"); 3817 assert(_restart_addr == NULL, "foreground will restart from scratch"); 3818 if (PrintGCDetails) { 3819 gclog_or_tty->print_cr("bailing out to foreground collection"); 3820 } 3821 } 3822 if (UseAdaptiveSizePolicy) { 3823 size_policy()->concurrent_marking_end(); 3824 } 3825 } else { 3826 assert(SafepointSynchronize::is_at_safepoint(), 3827 "inconsistent with asynch == false"); 3828 if (UseAdaptiveSizePolicy) { 3829 size_policy()->ms_collection_marking_begin(); 3830 } 3831 // already have locks 3832 res = markFromRootsWork(asynch); 3833 _collectorState = FinalMarking; 3834 if (UseAdaptiveSizePolicy) { 3835 GenCollectedHeap* gch = GenCollectedHeap::heap(); 3836 size_policy()->ms_collection_marking_end(gch->gc_cause()); 3837 } 3838 } 3839 verify_overflow_empty(); 3840 return res; 3841 } 3842 3843 bool CMSCollector::markFromRootsWork(bool asynch) { 3844 // iterate over marked bits in bit map, doing a full scan and mark 3845 // from these roots using the following algorithm: 3846 // . if oop is to the right of the current scan pointer, 3847 // mark corresponding bit (we'll process it later) 3848 // . else (oop is to left of current scan pointer) 3849 // push oop on marking stack 3850 // . drain the marking stack 3851 3852 // Note that when we do a marking step we need to hold the 3853 // bit map lock -- recall that direct allocation (by mutators) 3854 // and promotion (by younger generation collectors) is also 3855 // marking the bit map. [the so-called allocate live policy.] 3856 // Because the implementation of bit map marking is not 3857 // robust wrt simultaneous marking of bits in the same word, 3858 // we need to make sure that there is no such interference 3859 // between concurrent such updates. 3860 3861 // already have locks 3862 assert_lock_strong(bitMapLock()); 3863 3864 verify_work_stacks_empty(); 3865 verify_overflow_empty(); 3866 bool result = false; 3867 if (CMSConcurrentMTEnabled && ConcGCThreads > 0) { 3868 result = do_marking_mt(asynch); 3869 } else { 3870 result = do_marking_st(asynch); 3871 } 3872 return result; 3873 } 3874 3875 // Forward decl 3876 class CMSConcMarkingTask; 3877 3878 class CMSConcMarkingTerminator: public ParallelTaskTerminator { 3879 CMSCollector* _collector; 3880 CMSConcMarkingTask* _task; 3881 public: 3882 virtual void yield(); 3883 3884 // "n_threads" is the number of threads to be terminated. 3885 // "queue_set" is a set of work queues of other threads. 3886 // "collector" is the CMS collector associated with this task terminator. 3887 // "yield" indicates whether we need the gang as a whole to yield. 3888 CMSConcMarkingTerminator(int n_threads, TaskQueueSetSuper* queue_set, CMSCollector* collector) : 3889 ParallelTaskTerminator(n_threads, queue_set), 3890 _collector(collector) { } 3891 3892 void set_task(CMSConcMarkingTask* task) { 3893 _task = task; 3894 } 3895 }; 3896 3897 class CMSConcMarkingTerminatorTerminator: public TerminatorTerminator { 3898 CMSConcMarkingTask* _task; 3899 public: 3900 bool should_exit_termination(); 3901 void set_task(CMSConcMarkingTask* task) { 3902 _task = task; 3903 } 3904 }; 3905 3906 // MT Concurrent Marking Task 3907 class CMSConcMarkingTask: public YieldingFlexibleGangTask { 3908 CMSCollector* _collector; 3909 int _n_workers; // requested/desired # workers 3910 bool _asynch; 3911 bool _result; 3912 CompactibleFreeListSpace* _cms_space; 3913 char _pad_front[64]; // padding to ... 3914 HeapWord* _global_finger; // ... avoid sharing cache line 3915 char _pad_back[64]; 3916 HeapWord* _restart_addr; 3917 3918 // Exposed here for yielding support 3919 Mutex* const _bit_map_lock; 3920 3921 // The per thread work queues, available here for stealing 3922 OopTaskQueueSet* _task_queues; 3923 3924 // Termination (and yielding) support 3925 CMSConcMarkingTerminator _term; 3926 CMSConcMarkingTerminatorTerminator _term_term; 3927 3928 public: 3929 CMSConcMarkingTask(CMSCollector* collector, 3930 CompactibleFreeListSpace* cms_space, 3931 bool asynch, 3932 YieldingFlexibleWorkGang* workers, 3933 OopTaskQueueSet* task_queues): 3934 YieldingFlexibleGangTask("Concurrent marking done multi-threaded"), 3935 _collector(collector), 3936 _cms_space(cms_space), 3937 _asynch(asynch), _n_workers(0), _result(true), 3938 _task_queues(task_queues), 3939 _term(_n_workers, task_queues, _collector), 3940 _bit_map_lock(collector->bitMapLock()) 3941 { 3942 _requested_size = _n_workers; 3943 _term.set_task(this); 3944 _term_term.set_task(this); 3945 _restart_addr = _global_finger = _cms_space->bottom(); 3946 } 3947 3948 3949 OopTaskQueueSet* task_queues() { return _task_queues; } 3950 3951 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 3952 3953 HeapWord** global_finger_addr() { return &_global_finger; } 3954 3955 CMSConcMarkingTerminator* terminator() { return &_term; } 3956 3957 virtual void set_for_termination(int active_workers) { 3958 terminator()->reset_for_reuse(active_workers); 3959 } 3960 3961 void work(uint worker_id); 3962 bool should_yield() { 3963 return ConcurrentMarkSweepThread::should_yield() 3964 && !_collector->foregroundGCIsActive() 3965 && _asynch; 3966 } 3967 3968 virtual void coordinator_yield(); // stuff done by coordinator 3969 bool result() { return _result; } 3970 3971 void reset(HeapWord* ra) { 3972 assert(_global_finger >= _cms_space->end(), "Postcondition of ::work(i)"); 3973 _restart_addr = _global_finger = ra; 3974 _term.reset_for_reuse(); 3975 } 3976 3977 static bool get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 3978 OopTaskQueue* work_q); 3979 3980 private: 3981 void do_scan_and_mark(int i, CompactibleFreeListSpace* sp); 3982 void do_work_steal(int i); 3983 void bump_global_finger(HeapWord* f); 3984 }; 3985 3986 bool CMSConcMarkingTerminatorTerminator::should_exit_termination() { 3987 assert(_task != NULL, "Error"); 3988 return _task->yielding(); 3989 // Note that we do not need the disjunct || _task->should_yield() above 3990 // because we want terminating threads to yield only if the task 3991 // is already in the midst of yielding, which happens only after at least one 3992 // thread has yielded. 3993 } 3994 3995 void CMSConcMarkingTerminator::yield() { 3996 if (_task->should_yield()) { 3997 _task->yield(); 3998 } else { 3999 ParallelTaskTerminator::yield(); 4000 } 4001 } 4002 4003 //////////////////////////////////////////////////////////////// 4004 // Concurrent Marking Algorithm Sketch 4005 //////////////////////////////////////////////////////////////// 4006 // Until all tasks exhausted (both spaces): 4007 // -- claim next available chunk 4008 // -- bump global finger via CAS 4009 // -- find first object that starts in this chunk 4010 // and start scanning bitmap from that position 4011 // -- scan marked objects for oops 4012 // -- CAS-mark target, and if successful: 4013 // . if target oop is above global finger (volatile read) 4014 // nothing to do 4015 // . if target oop is in chunk and above local finger 4016 // then nothing to do 4017 // . else push on work-queue 4018 // -- Deal with possible overflow issues: 4019 // . local work-queue overflow causes stuff to be pushed on 4020 // global (common) overflow queue 4021 // . always first empty local work queue 4022 // . then get a batch of oops from global work queue if any 4023 // . then do work stealing 4024 // -- When all tasks claimed (both spaces) 4025 // and local work queue empty, 4026 // then in a loop do: 4027 // . check global overflow stack; steal a batch of oops and trace 4028 // . try to steal from other threads oif GOS is empty 4029 // . if neither is available, offer termination 4030 // -- Terminate and return result 4031 // 4032 void CMSConcMarkingTask::work(uint worker_id) { 4033 elapsedTimer _timer; 4034 ResourceMark rm; 4035 HandleMark hm; 4036 4037 DEBUG_ONLY(_collector->verify_overflow_empty();) 4038 4039 // Before we begin work, our work queue should be empty 4040 assert(work_queue(worker_id)->size() == 0, "Expected to be empty"); 4041 // Scan the bitmap covering _cms_space, tracing through grey objects. 4042 _timer.start(); 4043 do_scan_and_mark(worker_id, _cms_space); 4044 _timer.stop(); 4045 if (PrintCMSStatistics != 0) { 4046 gclog_or_tty->print_cr("Finished cms space scanning in %dth thread: %3.3f sec", 4047 worker_id, _timer.seconds()); 4048 // XXX: need xxx/xxx type of notation, two timers 4049 } 4050 4051 // ... do work stealing 4052 _timer.reset(); 4053 _timer.start(); 4054 do_work_steal(worker_id); 4055 _timer.stop(); 4056 if (PrintCMSStatistics != 0) { 4057 gclog_or_tty->print_cr("Finished work stealing in %dth thread: %3.3f sec", 4058 worker_id, _timer.seconds()); 4059 // XXX: need xxx/xxx type of notation, two timers 4060 } 4061 assert(_collector->_markStack.isEmpty(), "Should have been emptied"); 4062 assert(work_queue(worker_id)->size() == 0, "Should have been emptied"); 4063 // Note that under the current task protocol, the 4064 // following assertion is true even of the spaces 4065 // expanded since the completion of the concurrent 4066 // marking. XXX This will likely change under a strict 4067 // ABORT semantics. 4068 // After perm removal the comparison was changed to 4069 // greater than or equal to from strictly greater than. 4070 // Before perm removal the highest address sweep would 4071 // have been at the end of perm gen but now is at the 4072 // end of the tenured gen. 4073 assert(_global_finger >= _cms_space->end(), 4074 "All tasks have been completed"); 4075 DEBUG_ONLY(_collector->verify_overflow_empty();) 4076 } 4077 4078 void CMSConcMarkingTask::bump_global_finger(HeapWord* f) { 4079 HeapWord* read = _global_finger; 4080 HeapWord* cur = read; 4081 while (f > read) { 4082 cur = read; 4083 read = (HeapWord*) Atomic::cmpxchg_ptr(f, &_global_finger, cur); 4084 if (cur == read) { 4085 // our cas succeeded 4086 assert(_global_finger >= f, "protocol consistency"); 4087 break; 4088 } 4089 } 4090 } 4091 4092 // This is really inefficient, and should be redone by 4093 // using (not yet available) block-read and -write interfaces to the 4094 // stack and the work_queue. XXX FIX ME !!! 4095 bool CMSConcMarkingTask::get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 4096 OopTaskQueue* work_q) { 4097 // Fast lock-free check 4098 if (ovflw_stk->length() == 0) { 4099 return false; 4100 } 4101 assert(work_q->size() == 0, "Shouldn't steal"); 4102 MutexLockerEx ml(ovflw_stk->par_lock(), 4103 Mutex::_no_safepoint_check_flag); 4104 // Grab up to 1/4 the size of the work queue 4105 size_t num = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 4106 (size_t)ParGCDesiredObjsFromOverflowList); 4107 num = MIN2(num, ovflw_stk->length()); 4108 for (int i = (int) num; i > 0; i--) { 4109 oop cur = ovflw_stk->pop(); 4110 assert(cur != NULL, "Counted wrong?"); 4111 work_q->push(cur); 4112 } 4113 return num > 0; 4114 } 4115 4116 void CMSConcMarkingTask::do_scan_and_mark(int i, CompactibleFreeListSpace* sp) { 4117 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 4118 int n_tasks = pst->n_tasks(); 4119 // We allow that there may be no tasks to do here because 4120 // we are restarting after a stack overflow. 4121 assert(pst->valid() || n_tasks == 0, "Uninitialized use?"); 4122 uint nth_task = 0; 4123 4124 HeapWord* aligned_start = sp->bottom(); 4125 if (sp->used_region().contains(_restart_addr)) { 4126 // Align down to a card boundary for the start of 0th task 4127 // for this space. 4128 aligned_start = 4129 (HeapWord*)align_size_down((uintptr_t)_restart_addr, 4130 CardTableModRefBS::card_size); 4131 } 4132 4133 size_t chunk_size = sp->marking_task_size(); 4134 while (!pst->is_task_claimed(/* reference */ nth_task)) { 4135 // Having claimed the nth task in this space, 4136 // compute the chunk that it corresponds to: 4137 MemRegion span = MemRegion(aligned_start + nth_task*chunk_size, 4138 aligned_start + (nth_task+1)*chunk_size); 4139 // Try and bump the global finger via a CAS; 4140 // note that we need to do the global finger bump 4141 // _before_ taking the intersection below, because 4142 // the task corresponding to that region will be 4143 // deemed done even if the used_region() expands 4144 // because of allocation -- as it almost certainly will 4145 // during start-up while the threads yield in the 4146 // closure below. 4147 HeapWord* finger = span.end(); 4148 bump_global_finger(finger); // atomically 4149 // There are null tasks here corresponding to chunks 4150 // beyond the "top" address of the space. 4151 span = span.intersection(sp->used_region()); 4152 if (!span.is_empty()) { // Non-null task 4153 HeapWord* prev_obj; 4154 assert(!span.contains(_restart_addr) || nth_task == 0, 4155 "Inconsistency"); 4156 if (nth_task == 0) { 4157 // For the 0th task, we'll not need to compute a block_start. 4158 if (span.contains(_restart_addr)) { 4159 // In the case of a restart because of stack overflow, 4160 // we might additionally skip a chunk prefix. 4161 prev_obj = _restart_addr; 4162 } else { 4163 prev_obj = span.start(); 4164 } 4165 } else { 4166 // We want to skip the first object because 4167 // the protocol is to scan any object in its entirety 4168 // that _starts_ in this span; a fortiori, any 4169 // object starting in an earlier span is scanned 4170 // as part of an earlier claimed task. 4171 // Below we use the "careful" version of block_start 4172 // so we do not try to navigate uninitialized objects. 4173 prev_obj = sp->block_start_careful(span.start()); 4174 // Below we use a variant of block_size that uses the 4175 // Printezis bits to avoid waiting for allocated 4176 // objects to become initialized/parsable. 4177 while (prev_obj < span.start()) { 4178 size_t sz = sp->block_size_no_stall(prev_obj, _collector); 4179 if (sz > 0) { 4180 prev_obj += sz; 4181 } else { 4182 // In this case we may end up doing a bit of redundant 4183 // scanning, but that appears unavoidable, short of 4184 // locking the free list locks; see bug 6324141. 4185 break; 4186 } 4187 } 4188 } 4189 if (prev_obj < span.end()) { 4190 MemRegion my_span = MemRegion(prev_obj, span.end()); 4191 // Do the marking work within a non-empty span -- 4192 // the last argument to the constructor indicates whether the 4193 // iteration should be incremental with periodic yields. 4194 Par_MarkFromRootsClosure cl(this, _collector, my_span, 4195 &_collector->_markBitMap, 4196 work_queue(i), 4197 &_collector->_markStack, 4198 _asynch); 4199 _collector->_markBitMap.iterate(&cl, my_span.start(), my_span.end()); 4200 } // else nothing to do for this task 4201 } // else nothing to do for this task 4202 } 4203 // We'd be tempted to assert here that since there are no 4204 // more tasks left to claim in this space, the global_finger 4205 // must exceed space->top() and a fortiori space->end(). However, 4206 // that would not quite be correct because the bumping of 4207 // global_finger occurs strictly after the claiming of a task, 4208 // so by the time we reach here the global finger may not yet 4209 // have been bumped up by the thread that claimed the last 4210 // task. 4211 pst->all_tasks_completed(); 4212 } 4213 4214 class Par_ConcMarkingClosure: public MetadataAwareOopClosure { 4215 private: 4216 CMSCollector* _collector; 4217 CMSConcMarkingTask* _task; 4218 MemRegion _span; 4219 CMSBitMap* _bit_map; 4220 CMSMarkStack* _overflow_stack; 4221 OopTaskQueue* _work_queue; 4222 protected: 4223 DO_OOP_WORK_DEFN 4224 public: 4225 Par_ConcMarkingClosure(CMSCollector* collector, CMSConcMarkingTask* task, OopTaskQueue* work_queue, 4226 CMSBitMap* bit_map, CMSMarkStack* overflow_stack): 4227 MetadataAwareOopClosure(collector->ref_processor()), 4228 _collector(collector), 4229 _task(task), 4230 _span(collector->_span), 4231 _work_queue(work_queue), 4232 _bit_map(bit_map), 4233 _overflow_stack(overflow_stack) 4234 { } 4235 virtual void do_oop(oop* p); 4236 virtual void do_oop(narrowOop* p); 4237 4238 void trim_queue(size_t max); 4239 void handle_stack_overflow(HeapWord* lost); 4240 void do_yield_check() { 4241 if (_task->should_yield()) { 4242 _task->yield(); 4243 } 4244 } 4245 }; 4246 4247 // Grey object scanning during work stealing phase -- 4248 // the salient assumption here is that any references 4249 // that are in these stolen objects being scanned must 4250 // already have been initialized (else they would not have 4251 // been published), so we do not need to check for 4252 // uninitialized objects before pushing here. 4253 void Par_ConcMarkingClosure::do_oop(oop obj) { 4254 assert(obj->is_oop_or_null(true), "expected an oop or NULL"); 4255 HeapWord* addr = (HeapWord*)obj; 4256 // Check if oop points into the CMS generation 4257 // and is not marked 4258 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 4259 // a white object ... 4260 // If we manage to "claim" the object, by being the 4261 // first thread to mark it, then we push it on our 4262 // marking stack 4263 if (_bit_map->par_mark(addr)) { // ... now grey 4264 // push on work queue (grey set) 4265 bool simulate_overflow = false; 4266 NOT_PRODUCT( 4267 if (CMSMarkStackOverflowALot && 4268 _collector->simulate_overflow()) { 4269 // simulate a stack overflow 4270 simulate_overflow = true; 4271 } 4272 ) 4273 if (simulate_overflow || 4274 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) { 4275 // stack overflow 4276 if (PrintCMSStatistics != 0) { 4277 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 4278 SIZE_FORMAT, _overflow_stack->capacity()); 4279 } 4280 // We cannot assert that the overflow stack is full because 4281 // it may have been emptied since. 4282 assert(simulate_overflow || 4283 _work_queue->size() == _work_queue->max_elems(), 4284 "Else push should have succeeded"); 4285 handle_stack_overflow(addr); 4286 } 4287 } // Else, some other thread got there first 4288 do_yield_check(); 4289 } 4290 } 4291 4292 void Par_ConcMarkingClosure::do_oop(oop* p) { Par_ConcMarkingClosure::do_oop_work(p); } 4293 void Par_ConcMarkingClosure::do_oop(narrowOop* p) { Par_ConcMarkingClosure::do_oop_work(p); } 4294 4295 void Par_ConcMarkingClosure::trim_queue(size_t max) { 4296 while (_work_queue->size() > max) { 4297 oop new_oop; 4298 if (_work_queue->pop_local(new_oop)) { 4299 assert(new_oop->is_oop(), "Should be an oop"); 4300 assert(_bit_map->isMarked((HeapWord*)new_oop), "Grey object"); 4301 assert(_span.contains((HeapWord*)new_oop), "Not in span"); 4302 new_oop->oop_iterate(this); // do_oop() above 4303 do_yield_check(); 4304 } 4305 } 4306 } 4307 4308 // Upon stack overflow, we discard (part of) the stack, 4309 // remembering the least address amongst those discarded 4310 // in CMSCollector's _restart_address. 4311 void Par_ConcMarkingClosure::handle_stack_overflow(HeapWord* lost) { 4312 // We need to do this under a mutex to prevent other 4313 // workers from interfering with the work done below. 4314 MutexLockerEx ml(_overflow_stack->par_lock(), 4315 Mutex::_no_safepoint_check_flag); 4316 // Remember the least grey address discarded 4317 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 4318 _collector->lower_restart_addr(ra); 4319 _overflow_stack->reset(); // discard stack contents 4320 _overflow_stack->expand(); // expand the stack if possible 4321 } 4322 4323 4324 void CMSConcMarkingTask::do_work_steal(int i) { 4325 OopTaskQueue* work_q = work_queue(i); 4326 oop obj_to_scan; 4327 CMSBitMap* bm = &(_collector->_markBitMap); 4328 CMSMarkStack* ovflw = &(_collector->_markStack); 4329 int* seed = _collector->hash_seed(i); 4330 Par_ConcMarkingClosure cl(_collector, this, work_q, bm, ovflw); 4331 while (true) { 4332 cl.trim_queue(0); 4333 assert(work_q->size() == 0, "Should have been emptied above"); 4334 if (get_work_from_overflow_stack(ovflw, work_q)) { 4335 // Can't assert below because the work obtained from the 4336 // overflow stack may already have been stolen from us. 4337 // assert(work_q->size() > 0, "Work from overflow stack"); 4338 continue; 4339 } else if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 4340 assert(obj_to_scan->is_oop(), "Should be an oop"); 4341 assert(bm->isMarked((HeapWord*)obj_to_scan), "Grey object"); 4342 obj_to_scan->oop_iterate(&cl); 4343 } else if (terminator()->offer_termination(&_term_term)) { 4344 assert(work_q->size() == 0, "Impossible!"); 4345 break; 4346 } else if (yielding() || should_yield()) { 4347 yield(); 4348 } 4349 } 4350 } 4351 4352 // This is run by the CMS (coordinator) thread. 4353 void CMSConcMarkingTask::coordinator_yield() { 4354 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 4355 "CMS thread should hold CMS token"); 4356 // First give up the locks, then yield, then re-lock 4357 // We should probably use a constructor/destructor idiom to 4358 // do this unlock/lock or modify the MutexUnlocker class to 4359 // serve our purpose. XXX 4360 assert_lock_strong(_bit_map_lock); 4361 _bit_map_lock->unlock(); 4362 ConcurrentMarkSweepThread::desynchronize(true); 4363 ConcurrentMarkSweepThread::acknowledge_yield_request(); 4364 _collector->stopTimer(); 4365 if (PrintCMSStatistics != 0) { 4366 _collector->incrementYields(); 4367 } 4368 _collector->icms_wait(); 4369 4370 // It is possible for whichever thread initiated the yield request 4371 // not to get a chance to wake up and take the bitmap lock between 4372 // this thread releasing it and reacquiring it. So, while the 4373 // should_yield() flag is on, let's sleep for a bit to give the 4374 // other thread a chance to wake up. The limit imposed on the number 4375 // of iterations is defensive, to avoid any unforseen circumstances 4376 // putting us into an infinite loop. Since it's always been this 4377 // (coordinator_yield()) method that was observed to cause the 4378 // problem, we are using a parameter (CMSCoordinatorYieldSleepCount) 4379 // which is by default non-zero. For the other seven methods that 4380 // also perform the yield operation, as are using a different 4381 // parameter (CMSYieldSleepCount) which is by default zero. This way we 4382 // can enable the sleeping for those methods too, if necessary. 4383 // See 6442774. 4384 // 4385 // We really need to reconsider the synchronization between the GC 4386 // thread and the yield-requesting threads in the future and we 4387 // should really use wait/notify, which is the recommended 4388 // way of doing this type of interaction. Additionally, we should 4389 // consolidate the eight methods that do the yield operation and they 4390 // are almost identical into one for better maintenability and 4391 // readability. See 6445193. 4392 // 4393 // Tony 2006.06.29 4394 for (unsigned i = 0; i < CMSCoordinatorYieldSleepCount && 4395 ConcurrentMarkSweepThread::should_yield() && 4396 !CMSCollector::foregroundGCIsActive(); ++i) { 4397 os::sleep(Thread::current(), 1, false); 4398 ConcurrentMarkSweepThread::acknowledge_yield_request(); 4399 } 4400 4401 ConcurrentMarkSweepThread::synchronize(true); 4402 _bit_map_lock->lock_without_safepoint_check(); 4403 _collector->startTimer(); 4404 } 4405 4406 bool CMSCollector::do_marking_mt(bool asynch) { 4407 assert(ConcGCThreads > 0 && conc_workers() != NULL, "precondition"); 4408 int num_workers = AdaptiveSizePolicy::calc_active_conc_workers( 4409 conc_workers()->total_workers(), 4410 conc_workers()->active_workers(), 4411 Threads::number_of_non_daemon_threads()); 4412 conc_workers()->set_active_workers(num_workers); 4413 4414 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 4415 4416 CMSConcMarkingTask tsk(this, 4417 cms_space, 4418 asynch, 4419 conc_workers(), 4420 task_queues()); 4421 4422 // Since the actual number of workers we get may be different 4423 // from the number we requested above, do we need to do anything different 4424 // below? In particular, may be we need to subclass the SequantialSubTasksDone 4425 // class?? XXX 4426 cms_space ->initialize_sequential_subtasks_for_marking(num_workers); 4427 4428 // Refs discovery is already non-atomic. 4429 assert(!ref_processor()->discovery_is_atomic(), "Should be non-atomic"); 4430 assert(ref_processor()->discovery_is_mt(), "Discovery should be MT"); 4431 conc_workers()->start_task(&tsk); 4432 while (tsk.yielded()) { 4433 tsk.coordinator_yield(); 4434 conc_workers()->continue_task(&tsk); 4435 } 4436 // If the task was aborted, _restart_addr will be non-NULL 4437 assert(tsk.completed() || _restart_addr != NULL, "Inconsistency"); 4438 while (_restart_addr != NULL) { 4439 // XXX For now we do not make use of ABORTED state and have not 4440 // yet implemented the right abort semantics (even in the original 4441 // single-threaded CMS case). That needs some more investigation 4442 // and is deferred for now; see CR# TBF. 07252005YSR. XXX 4443 assert(!CMSAbortSemantics || tsk.aborted(), "Inconsistency"); 4444 // If _restart_addr is non-NULL, a marking stack overflow 4445 // occurred; we need to do a fresh marking iteration from the 4446 // indicated restart address. 4447 if (_foregroundGCIsActive && asynch) { 4448 // We may be running into repeated stack overflows, having 4449 // reached the limit of the stack size, while making very 4450 // slow forward progress. It may be best to bail out and 4451 // let the foreground collector do its job. 4452 // Clear _restart_addr, so that foreground GC 4453 // works from scratch. This avoids the headache of 4454 // a "rescan" which would otherwise be needed because 4455 // of the dirty mod union table & card table. 4456 _restart_addr = NULL; 4457 return false; 4458 } 4459 // Adjust the task to restart from _restart_addr 4460 tsk.reset(_restart_addr); 4461 cms_space ->initialize_sequential_subtasks_for_marking(num_workers, 4462 _restart_addr); 4463 _restart_addr = NULL; 4464 // Get the workers going again 4465 conc_workers()->start_task(&tsk); 4466 while (tsk.yielded()) { 4467 tsk.coordinator_yield(); 4468 conc_workers()->continue_task(&tsk); 4469 } 4470 } 4471 assert(tsk.completed(), "Inconsistency"); 4472 assert(tsk.result() == true, "Inconsistency"); 4473 return true; 4474 } 4475 4476 bool CMSCollector::do_marking_st(bool asynch) { 4477 ResourceMark rm; 4478 HandleMark hm; 4479 4480 // Temporarily make refs discovery single threaded (non-MT) 4481 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false); 4482 MarkFromRootsClosure markFromRootsClosure(this, _span, &_markBitMap, 4483 &_markStack, CMSYield && asynch); 4484 // the last argument to iterate indicates whether the iteration 4485 // should be incremental with periodic yields. 4486 _markBitMap.iterate(&markFromRootsClosure); 4487 // If _restart_addr is non-NULL, a marking stack overflow 4488 // occurred; we need to do a fresh iteration from the 4489 // indicated restart address. 4490 while (_restart_addr != NULL) { 4491 if (_foregroundGCIsActive && asynch) { 4492 // We may be running into repeated stack overflows, having 4493 // reached the limit of the stack size, while making very 4494 // slow forward progress. It may be best to bail out and 4495 // let the foreground collector do its job. 4496 // Clear _restart_addr, so that foreground GC 4497 // works from scratch. This avoids the headache of 4498 // a "rescan" which would otherwise be needed because 4499 // of the dirty mod union table & card table. 4500 _restart_addr = NULL; 4501 return false; // indicating failure to complete marking 4502 } 4503 // Deal with stack overflow: 4504 // we restart marking from _restart_addr 4505 HeapWord* ra = _restart_addr; 4506 markFromRootsClosure.reset(ra); 4507 _restart_addr = NULL; 4508 _markBitMap.iterate(&markFromRootsClosure, ra, _span.end()); 4509 } 4510 return true; 4511 } 4512 4513 void CMSCollector::preclean() { 4514 check_correct_thread_executing(); 4515 assert(Thread::current()->is_ConcurrentGC_thread(), "Wrong thread"); 4516 verify_work_stacks_empty(); 4517 verify_overflow_empty(); 4518 _abort_preclean = false; 4519 if (CMSPrecleaningEnabled) { 4520 if (!CMSEdenChunksRecordAlways) { 4521 _eden_chunk_index = 0; 4522 } 4523 size_t used = get_eden_used(); 4524 size_t capacity = get_eden_capacity(); 4525 // Don't start sampling unless we will get sufficiently 4526 // many samples. 4527 if (used < (capacity/(CMSScheduleRemarkSamplingRatio * 100) 4528 * CMSScheduleRemarkEdenPenetration)) { 4529 _start_sampling = true; 4530 } else { 4531 _start_sampling = false; 4532 } 4533 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 4534 CMSPhaseAccounting pa(this, "preclean", _gc_tracer_cm->gc_id(), !PrintGCDetails); 4535 preclean_work(CMSPrecleanRefLists1, CMSPrecleanSurvivors1); 4536 } 4537 CMSTokenSync x(true); // is cms thread 4538 if (CMSPrecleaningEnabled) { 4539 sample_eden(); 4540 _collectorState = AbortablePreclean; 4541 } else { 4542 _collectorState = FinalMarking; 4543 } 4544 verify_work_stacks_empty(); 4545 verify_overflow_empty(); 4546 } 4547 4548 // Try and schedule the remark such that young gen 4549 // occupancy is CMSScheduleRemarkEdenPenetration %. 4550 void CMSCollector::abortable_preclean() { 4551 check_correct_thread_executing(); 4552 assert(CMSPrecleaningEnabled, "Inconsistent control state"); 4553 assert(_collectorState == AbortablePreclean, "Inconsistent control state"); 4554 4555 // If Eden's current occupancy is below this threshold, 4556 // immediately schedule the remark; else preclean 4557 // past the next scavenge in an effort to 4558 // schedule the pause as described avove. By choosing 4559 // CMSScheduleRemarkEdenSizeThreshold >= max eden size 4560 // we will never do an actual abortable preclean cycle. 4561 if (get_eden_used() > CMSScheduleRemarkEdenSizeThreshold) { 4562 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 4563 CMSPhaseAccounting pa(this, "abortable-preclean", _gc_tracer_cm->gc_id(), !PrintGCDetails); 4564 // We need more smarts in the abortable preclean 4565 // loop below to deal with cases where allocation 4566 // in young gen is very very slow, and our precleaning 4567 // is running a losing race against a horde of 4568 // mutators intent on flooding us with CMS updates 4569 // (dirty cards). 4570 // One, admittedly dumb, strategy is to give up 4571 // after a certain number of abortable precleaning loops 4572 // or after a certain maximum time. We want to make 4573 // this smarter in the next iteration. 4574 // XXX FIX ME!!! YSR 4575 size_t loops = 0, workdone = 0, cumworkdone = 0, waited = 0; 4576 while (!(should_abort_preclean() || 4577 ConcurrentMarkSweepThread::should_terminate())) { 4578 workdone = preclean_work(CMSPrecleanRefLists2, CMSPrecleanSurvivors2); 4579 cumworkdone += workdone; 4580 loops++; 4581 // Voluntarily terminate abortable preclean phase if we have 4582 // been at it for too long. 4583 if ((CMSMaxAbortablePrecleanLoops != 0) && 4584 loops >= CMSMaxAbortablePrecleanLoops) { 4585 if (PrintGCDetails) { 4586 gclog_or_tty->print(" CMS: abort preclean due to loops "); 4587 } 4588 break; 4589 } 4590 if (pa.wallclock_millis() > CMSMaxAbortablePrecleanTime) { 4591 if (PrintGCDetails) { 4592 gclog_or_tty->print(" CMS: abort preclean due to time "); 4593 } 4594 break; 4595 } 4596 // If we are doing little work each iteration, we should 4597 // take a short break. 4598 if (workdone < CMSAbortablePrecleanMinWorkPerIteration) { 4599 // Sleep for some time, waiting for work to accumulate 4600 stopTimer(); 4601 cmsThread()->wait_on_cms_lock(CMSAbortablePrecleanWaitMillis); 4602 startTimer(); 4603 waited++; 4604 } 4605 } 4606 if (PrintCMSStatistics > 0) { 4607 gclog_or_tty->print(" [%d iterations, %d waits, %d cards)] ", 4608 loops, waited, cumworkdone); 4609 } 4610 } 4611 CMSTokenSync x(true); // is cms thread 4612 if (_collectorState != Idling) { 4613 assert(_collectorState == AbortablePreclean, 4614 "Spontaneous state transition?"); 4615 _collectorState = FinalMarking; 4616 } // Else, a foreground collection completed this CMS cycle. 4617 return; 4618 } 4619 4620 // Respond to an Eden sampling opportunity 4621 void CMSCollector::sample_eden() { 4622 // Make sure a young gc cannot sneak in between our 4623 // reading and recording of a sample. 4624 assert(Thread::current()->is_ConcurrentGC_thread(), 4625 "Only the cms thread may collect Eden samples"); 4626 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 4627 "Should collect samples while holding CMS token"); 4628 if (!_start_sampling) { 4629 return; 4630 } 4631 // When CMSEdenChunksRecordAlways is true, the eden chunk array 4632 // is populated by the young generation. 4633 if (_eden_chunk_array != NULL && !CMSEdenChunksRecordAlways) { 4634 if (_eden_chunk_index < _eden_chunk_capacity) { 4635 _eden_chunk_array[_eden_chunk_index] = *_top_addr; // take sample 4636 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr, 4637 "Unexpected state of Eden"); 4638 // We'd like to check that what we just sampled is an oop-start address; 4639 // however, we cannot do that here since the object may not yet have been 4640 // initialized. So we'll instead do the check when we _use_ this sample 4641 // later. 4642 if (_eden_chunk_index == 0 || 4643 (pointer_delta(_eden_chunk_array[_eden_chunk_index], 4644 _eden_chunk_array[_eden_chunk_index-1]) 4645 >= CMSSamplingGrain)) { 4646 _eden_chunk_index++; // commit sample 4647 } 4648 } 4649 } 4650 if ((_collectorState == AbortablePreclean) && !_abort_preclean) { 4651 size_t used = get_eden_used(); 4652 size_t capacity = get_eden_capacity(); 4653 assert(used <= capacity, "Unexpected state of Eden"); 4654 if (used > (capacity/100 * CMSScheduleRemarkEdenPenetration)) { 4655 _abort_preclean = true; 4656 } 4657 } 4658 } 4659 4660 4661 size_t CMSCollector::preclean_work(bool clean_refs, bool clean_survivor) { 4662 assert(_collectorState == Precleaning || 4663 _collectorState == AbortablePreclean, "incorrect state"); 4664 ResourceMark rm; 4665 HandleMark hm; 4666 4667 // Precleaning is currently not MT but the reference processor 4668 // may be set for MT. Disable it temporarily here. 4669 ReferenceProcessor* rp = ref_processor(); 4670 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(rp, false); 4671 4672 // Do one pass of scrubbing the discovered reference lists 4673 // to remove any reference objects with strongly-reachable 4674 // referents. 4675 if (clean_refs) { 4676 CMSPrecleanRefsYieldClosure yield_cl(this); 4677 assert(rp->span().equals(_span), "Spans should be equal"); 4678 CMSKeepAliveClosure keep_alive(this, _span, &_markBitMap, 4679 &_markStack, true /* preclean */); 4680 CMSDrainMarkingStackClosure complete_trace(this, 4681 _span, &_markBitMap, &_markStack, 4682 &keep_alive, true /* preclean */); 4683 4684 // We don't want this step to interfere with a young 4685 // collection because we don't want to take CPU 4686 // or memory bandwidth away from the young GC threads 4687 // (which may be as many as there are CPUs). 4688 // Note that we don't need to protect ourselves from 4689 // interference with mutators because they can't 4690 // manipulate the discovered reference lists nor affect 4691 // the computed reachability of the referents, the 4692 // only properties manipulated by the precleaning 4693 // of these reference lists. 4694 stopTimer(); 4695 CMSTokenSyncWithLocks x(true /* is cms thread */, 4696 bitMapLock()); 4697 startTimer(); 4698 sample_eden(); 4699 4700 // The following will yield to allow foreground 4701 // collection to proceed promptly. XXX YSR: 4702 // The code in this method may need further 4703 // tweaking for better performance and some restructuring 4704 // for cleaner interfaces. 4705 GCTimer *gc_timer = NULL; // Currently not tracing concurrent phases 4706 rp->preclean_discovered_references( 4707 rp->is_alive_non_header(), &keep_alive, &complete_trace, &yield_cl, 4708 gc_timer, _gc_tracer_cm->gc_id()); 4709 } 4710 4711 if (clean_survivor) { // preclean the active survivor space(s) 4712 assert(_young_gen->kind() == Generation::DefNew || 4713 _young_gen->kind() == Generation::ParNew || 4714 _young_gen->kind() == Generation::ASParNew, 4715 "incorrect type for cast"); 4716 DefNewGeneration* dng = (DefNewGeneration*)_young_gen; 4717 PushAndMarkClosure pam_cl(this, _span, ref_processor(), 4718 &_markBitMap, &_modUnionTable, 4719 &_markStack, true /* precleaning phase */); 4720 stopTimer(); 4721 CMSTokenSyncWithLocks ts(true /* is cms thread */, 4722 bitMapLock()); 4723 startTimer(); 4724 unsigned int before_count = 4725 GenCollectedHeap::heap()->total_collections(); 4726 SurvivorSpacePrecleanClosure 4727 sss_cl(this, _span, &_markBitMap, &_markStack, 4728 &pam_cl, before_count, CMSYield); 4729 dng->from()->object_iterate_careful(&sss_cl); 4730 dng->to()->object_iterate_careful(&sss_cl); 4731 } 4732 MarkRefsIntoAndScanClosure 4733 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable, 4734 &_markStack, this, CMSYield, 4735 true /* precleaning phase */); 4736 // CAUTION: The following closure has persistent state that may need to 4737 // be reset upon a decrease in the sequence of addresses it 4738 // processes. 4739 ScanMarkedObjectsAgainCarefullyClosure 4740 smoac_cl(this, _span, 4741 &_markBitMap, &_markStack, &mrias_cl, CMSYield); 4742 4743 // Preclean dirty cards in ModUnionTable and CardTable using 4744 // appropriate convergence criterion; 4745 // repeat CMSPrecleanIter times unless we find that 4746 // we are losing. 4747 assert(CMSPrecleanIter < 10, "CMSPrecleanIter is too large"); 4748 assert(CMSPrecleanNumerator < CMSPrecleanDenominator, 4749 "Bad convergence multiplier"); 4750 assert(CMSPrecleanThreshold >= 100, 4751 "Unreasonably low CMSPrecleanThreshold"); 4752 4753 size_t numIter, cumNumCards, lastNumCards, curNumCards; 4754 for (numIter = 0, cumNumCards = lastNumCards = curNumCards = 0; 4755 numIter < CMSPrecleanIter; 4756 numIter++, lastNumCards = curNumCards, cumNumCards += curNumCards) { 4757 curNumCards = preclean_mod_union_table(_cmsGen, &smoac_cl); 4758 if (Verbose && PrintGCDetails) { 4759 gclog_or_tty->print(" (modUnionTable: %d cards)", curNumCards); 4760 } 4761 // Either there are very few dirty cards, so re-mark 4762 // pause will be small anyway, or our pre-cleaning isn't 4763 // that much faster than the rate at which cards are being 4764 // dirtied, so we might as well stop and re-mark since 4765 // precleaning won't improve our re-mark time by much. 4766 if (curNumCards <= CMSPrecleanThreshold || 4767 (numIter > 0 && 4768 (curNumCards * CMSPrecleanDenominator > 4769 lastNumCards * CMSPrecleanNumerator))) { 4770 numIter++; 4771 cumNumCards += curNumCards; 4772 break; 4773 } 4774 } 4775 4776 preclean_klasses(&mrias_cl, _cmsGen->freelistLock()); 4777 4778 curNumCards = preclean_card_table(_cmsGen, &smoac_cl); 4779 cumNumCards += curNumCards; 4780 if (PrintGCDetails && PrintCMSStatistics != 0) { 4781 gclog_or_tty->print_cr(" (cardTable: %d cards, re-scanned %d cards, %d iterations)", 4782 curNumCards, cumNumCards, numIter); 4783 } 4784 return cumNumCards; // as a measure of useful work done 4785 } 4786 4787 // PRECLEANING NOTES: 4788 // Precleaning involves: 4789 // . reading the bits of the modUnionTable and clearing the set bits. 4790 // . For the cards corresponding to the set bits, we scan the 4791 // objects on those cards. This means we need the free_list_lock 4792 // so that we can safely iterate over the CMS space when scanning 4793 // for oops. 4794 // . When we scan the objects, we'll be both reading and setting 4795 // marks in the marking bit map, so we'll need the marking bit map. 4796 // . For protecting _collector_state transitions, we take the CGC_lock. 4797 // Note that any races in the reading of of card table entries by the 4798 // CMS thread on the one hand and the clearing of those entries by the 4799 // VM thread or the setting of those entries by the mutator threads on the 4800 // other are quite benign. However, for efficiency it makes sense to keep 4801 // the VM thread from racing with the CMS thread while the latter is 4802 // dirty card info to the modUnionTable. We therefore also use the 4803 // CGC_lock to protect the reading of the card table and the mod union 4804 // table by the CM thread. 4805 // . We run concurrently with mutator updates, so scanning 4806 // needs to be done carefully -- we should not try to scan 4807 // potentially uninitialized objects. 4808 // 4809 // Locking strategy: While holding the CGC_lock, we scan over and 4810 // reset a maximal dirty range of the mod union / card tables, then lock 4811 // the free_list_lock and bitmap lock to do a full marking, then 4812 // release these locks; and repeat the cycle. This allows for a 4813 // certain amount of fairness in the sharing of these locks between 4814 // the CMS collector on the one hand, and the VM thread and the 4815 // mutators on the other. 4816 4817 // NOTE: preclean_mod_union_table() and preclean_card_table() 4818 // further below are largely identical; if you need to modify 4819 // one of these methods, please check the other method too. 4820 4821 size_t CMSCollector::preclean_mod_union_table( 4822 ConcurrentMarkSweepGeneration* gen, 4823 ScanMarkedObjectsAgainCarefullyClosure* cl) { 4824 verify_work_stacks_empty(); 4825 verify_overflow_empty(); 4826 4827 // strategy: starting with the first card, accumulate contiguous 4828 // ranges of dirty cards; clear these cards, then scan the region 4829 // covered by these cards. 4830 4831 // Since all of the MUT is committed ahead, we can just use 4832 // that, in case the generations expand while we are precleaning. 4833 // It might also be fine to just use the committed part of the 4834 // generation, but we might potentially miss cards when the 4835 // generation is rapidly expanding while we are in the midst 4836 // of precleaning. 4837 HeapWord* startAddr = gen->reserved().start(); 4838 HeapWord* endAddr = gen->reserved().end(); 4839 4840 cl->setFreelistLock(gen->freelistLock()); // needed for yielding 4841 4842 size_t numDirtyCards, cumNumDirtyCards; 4843 HeapWord *nextAddr, *lastAddr; 4844 for (cumNumDirtyCards = numDirtyCards = 0, 4845 nextAddr = lastAddr = startAddr; 4846 nextAddr < endAddr; 4847 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 4848 4849 ResourceMark rm; 4850 HandleMark hm; 4851 4852 MemRegion dirtyRegion; 4853 { 4854 stopTimer(); 4855 // Potential yield point 4856 CMSTokenSync ts(true); 4857 startTimer(); 4858 sample_eden(); 4859 // Get dirty region starting at nextOffset (inclusive), 4860 // simultaneously clearing it. 4861 dirtyRegion = 4862 _modUnionTable.getAndClearMarkedRegion(nextAddr, endAddr); 4863 assert(dirtyRegion.start() >= nextAddr, 4864 "returned region inconsistent?"); 4865 } 4866 // Remember where the next search should begin. 4867 // The returned region (if non-empty) is a right open interval, 4868 // so lastOffset is obtained from the right end of that 4869 // interval. 4870 lastAddr = dirtyRegion.end(); 4871 // Should do something more transparent and less hacky XXX 4872 numDirtyCards = 4873 _modUnionTable.heapWordDiffToOffsetDiff(dirtyRegion.word_size()); 4874 4875 // We'll scan the cards in the dirty region (with periodic 4876 // yields for foreground GC as needed). 4877 if (!dirtyRegion.is_empty()) { 4878 assert(numDirtyCards > 0, "consistency check"); 4879 HeapWord* stop_point = NULL; 4880 stopTimer(); 4881 // Potential yield point 4882 CMSTokenSyncWithLocks ts(true, gen->freelistLock(), 4883 bitMapLock()); 4884 startTimer(); 4885 { 4886 verify_work_stacks_empty(); 4887 verify_overflow_empty(); 4888 sample_eden(); 4889 stop_point = 4890 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 4891 } 4892 if (stop_point != NULL) { 4893 // The careful iteration stopped early either because it found an 4894 // uninitialized object, or because we were in the midst of an 4895 // "abortable preclean", which should now be aborted. Redirty 4896 // the bits corresponding to the partially-scanned or unscanned 4897 // cards. We'll either restart at the next block boundary or 4898 // abort the preclean. 4899 assert((_collectorState == AbortablePreclean && should_abort_preclean()), 4900 "Should only be AbortablePreclean."); 4901 _modUnionTable.mark_range(MemRegion(stop_point, dirtyRegion.end())); 4902 if (should_abort_preclean()) { 4903 break; // out of preclean loop 4904 } else { 4905 // Compute the next address at which preclean should pick up; 4906 // might need bitMapLock in order to read P-bits. 4907 lastAddr = next_card_start_after_block(stop_point); 4908 } 4909 } 4910 } else { 4911 assert(lastAddr == endAddr, "consistency check"); 4912 assert(numDirtyCards == 0, "consistency check"); 4913 break; 4914 } 4915 } 4916 verify_work_stacks_empty(); 4917 verify_overflow_empty(); 4918 return cumNumDirtyCards; 4919 } 4920 4921 // NOTE: preclean_mod_union_table() above and preclean_card_table() 4922 // below are largely identical; if you need to modify 4923 // one of these methods, please check the other method too. 4924 4925 size_t CMSCollector::preclean_card_table(ConcurrentMarkSweepGeneration* gen, 4926 ScanMarkedObjectsAgainCarefullyClosure* cl) { 4927 // strategy: it's similar to precleamModUnionTable above, in that 4928 // we accumulate contiguous ranges of dirty cards, mark these cards 4929 // precleaned, then scan the region covered by these cards. 4930 HeapWord* endAddr = (HeapWord*)(gen->_virtual_space.high()); 4931 HeapWord* startAddr = (HeapWord*)(gen->_virtual_space.low()); 4932 4933 cl->setFreelistLock(gen->freelistLock()); // needed for yielding 4934 4935 size_t numDirtyCards, cumNumDirtyCards; 4936 HeapWord *lastAddr, *nextAddr; 4937 4938 for (cumNumDirtyCards = numDirtyCards = 0, 4939 nextAddr = lastAddr = startAddr; 4940 nextAddr < endAddr; 4941 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 4942 4943 ResourceMark rm; 4944 HandleMark hm; 4945 4946 MemRegion dirtyRegion; 4947 { 4948 // See comments in "Precleaning notes" above on why we 4949 // do this locking. XXX Could the locking overheads be 4950 // too high when dirty cards are sparse? [I don't think so.] 4951 stopTimer(); 4952 CMSTokenSync x(true); // is cms thread 4953 startTimer(); 4954 sample_eden(); 4955 // Get and clear dirty region from card table 4956 dirtyRegion = _ct->ct_bs()->dirty_card_range_after_reset( 4957 MemRegion(nextAddr, endAddr), 4958 true, 4959 CardTableModRefBS::precleaned_card_val()); 4960 4961 assert(dirtyRegion.start() >= nextAddr, 4962 "returned region inconsistent?"); 4963 } 4964 lastAddr = dirtyRegion.end(); 4965 numDirtyCards = 4966 dirtyRegion.word_size()/CardTableModRefBS::card_size_in_words; 4967 4968 if (!dirtyRegion.is_empty()) { 4969 stopTimer(); 4970 CMSTokenSyncWithLocks ts(true, gen->freelistLock(), bitMapLock()); 4971 startTimer(); 4972 sample_eden(); 4973 verify_work_stacks_empty(); 4974 verify_overflow_empty(); 4975 HeapWord* stop_point = 4976 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 4977 if (stop_point != NULL) { 4978 assert((_collectorState == AbortablePreclean && should_abort_preclean()), 4979 "Should only be AbortablePreclean."); 4980 _ct->ct_bs()->invalidate(MemRegion(stop_point, dirtyRegion.end())); 4981 if (should_abort_preclean()) { 4982 break; // out of preclean loop 4983 } else { 4984 // Compute the next address at which preclean should pick up. 4985 lastAddr = next_card_start_after_block(stop_point); 4986 } 4987 } 4988 } else { 4989 break; 4990 } 4991 } 4992 verify_work_stacks_empty(); 4993 verify_overflow_empty(); 4994 return cumNumDirtyCards; 4995 } 4996 4997 class PrecleanKlassClosure : public KlassClosure { 4998 KlassToOopClosure _cm_klass_closure; 4999 public: 5000 PrecleanKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {} 5001 void do_klass(Klass* k) { 5002 if (k->has_accumulated_modified_oops()) { 5003 k->clear_accumulated_modified_oops(); 5004 5005 _cm_klass_closure.do_klass(k); 5006 } 5007 } 5008 }; 5009 5010 // The freelist lock is needed to prevent asserts, is it really needed? 5011 void CMSCollector::preclean_klasses(MarkRefsIntoAndScanClosure* cl, Mutex* freelistLock) { 5012 5013 cl->set_freelistLock(freelistLock); 5014 5015 CMSTokenSyncWithLocks ts(true, freelistLock, bitMapLock()); 5016 5017 // SSS: Add equivalent to ScanMarkedObjectsAgainCarefullyClosure::do_yield_check and should_abort_preclean? 5018 // SSS: We should probably check if precleaning should be aborted, at suitable intervals? 5019 PrecleanKlassClosure preclean_klass_closure(cl); 5020 ClassLoaderDataGraph::classes_do(&preclean_klass_closure); 5021 5022 verify_work_stacks_empty(); 5023 verify_overflow_empty(); 5024 } 5025 5026 void CMSCollector::checkpointRootsFinal(bool asynch, 5027 bool clear_all_soft_refs, bool init_mark_was_synchronous) { 5028 assert(_collectorState == FinalMarking, "incorrect state transition?"); 5029 check_correct_thread_executing(); 5030 // world is stopped at this checkpoint 5031 assert(SafepointSynchronize::is_at_safepoint(), 5032 "world should be stopped"); 5033 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 5034 5035 verify_work_stacks_empty(); 5036 verify_overflow_empty(); 5037 5038 SpecializationStats::clear(); 5039 if (PrintGCDetails) { 5040 gclog_or_tty->print("[YG occupancy: "SIZE_FORMAT" K ("SIZE_FORMAT" K)]", 5041 _young_gen->used() / K, 5042 _young_gen->capacity() / K); 5043 } 5044 if (asynch) { 5045 if (CMSScavengeBeforeRemark) { 5046 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5047 // Temporarily set flag to false, GCH->do_collection will 5048 // expect it to be false and set to true 5049 FlagSetting fl(gch->_is_gc_active, false); 5050 NOT_PRODUCT(GCTraceTime t("Scavenge-Before-Remark", 5051 PrintGCDetails && Verbose, true, _gc_timer_cm, _gc_tracer_cm->gc_id());) 5052 int level = _cmsGen->level() - 1; 5053 if (level >= 0) { 5054 gch->do_collection(true, // full (i.e. force, see below) 5055 false, // !clear_all_soft_refs 5056 0, // size 5057 false, // is_tlab 5058 level // max_level 5059 ); 5060 } 5061 } 5062 FreelistLocker x(this); 5063 MutexLockerEx y(bitMapLock(), 5064 Mutex::_no_safepoint_check_flag); 5065 assert(!init_mark_was_synchronous, "but that's impossible!"); 5066 checkpointRootsFinalWork(asynch, clear_all_soft_refs, false); 5067 } else { 5068 // already have all the locks 5069 checkpointRootsFinalWork(asynch, clear_all_soft_refs, 5070 init_mark_was_synchronous); 5071 } 5072 verify_work_stacks_empty(); 5073 verify_overflow_empty(); 5074 SpecializationStats::print(); 5075 } 5076 5077 void CMSCollector::checkpointRootsFinalWork(bool asynch, 5078 bool clear_all_soft_refs, bool init_mark_was_synchronous) { 5079 5080 NOT_PRODUCT(GCTraceTime tr("checkpointRootsFinalWork", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());) 5081 5082 assert(haveFreelistLocks(), "must have free list locks"); 5083 assert_lock_strong(bitMapLock()); 5084 5085 if (UseAdaptiveSizePolicy) { 5086 size_policy()->checkpoint_roots_final_begin(); 5087 } 5088 5089 ResourceMark rm; 5090 HandleMark hm; 5091 5092 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5093 5094 if (should_unload_classes()) { 5095 CodeCache::gc_prologue(); 5096 } 5097 assert(haveFreelistLocks(), "must have free list locks"); 5098 assert_lock_strong(bitMapLock()); 5099 5100 if (!init_mark_was_synchronous) { 5101 // We might assume that we need not fill TLAB's when 5102 // CMSScavengeBeforeRemark is set, because we may have just done 5103 // a scavenge which would have filled all TLAB's -- and besides 5104 // Eden would be empty. This however may not always be the case -- 5105 // for instance although we asked for a scavenge, it may not have 5106 // happened because of a JNI critical section. We probably need 5107 // a policy for deciding whether we can in that case wait until 5108 // the critical section releases and then do the remark following 5109 // the scavenge, and skip it here. In the absence of that policy, 5110 // or of an indication of whether the scavenge did indeed occur, 5111 // we cannot rely on TLAB's having been filled and must do 5112 // so here just in case a scavenge did not happen. 5113 gch->ensure_parsability(false); // fill TLAB's, but no need to retire them 5114 // Update the saved marks which may affect the root scans. 5115 gch->save_marks(); 5116 5117 if (CMSPrintEdenSurvivorChunks) { 5118 print_eden_and_survivor_chunk_arrays(); 5119 } 5120 5121 { 5122 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;) 5123 5124 // Note on the role of the mod union table: 5125 // Since the marker in "markFromRoots" marks concurrently with 5126 // mutators, it is possible for some reachable objects not to have been 5127 // scanned. For instance, an only reference to an object A was 5128 // placed in object B after the marker scanned B. Unless B is rescanned, 5129 // A would be collected. Such updates to references in marked objects 5130 // are detected via the mod union table which is the set of all cards 5131 // dirtied since the first checkpoint in this GC cycle and prior to 5132 // the most recent young generation GC, minus those cleaned up by the 5133 // concurrent precleaning. 5134 if (CMSParallelRemarkEnabled && CollectedHeap::use_parallel_gc_threads()) { 5135 GCTraceTime t("Rescan (parallel) ", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 5136 do_remark_parallel(); 5137 } else { 5138 GCTraceTime t("Rescan (non-parallel) ", PrintGCDetails, false, 5139 _gc_timer_cm, _gc_tracer_cm->gc_id()); 5140 do_remark_non_parallel(); 5141 } 5142 } 5143 } else { 5144 assert(!asynch, "Can't have init_mark_was_synchronous in asynch mode"); 5145 // The initial mark was stop-world, so there's no rescanning to 5146 // do; go straight on to the next step below. 5147 } 5148 verify_work_stacks_empty(); 5149 verify_overflow_empty(); 5150 5151 { 5152 NOT_PRODUCT(GCTraceTime ts("refProcessingWork", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id());) 5153 refProcessingWork(asynch, clear_all_soft_refs); 5154 } 5155 verify_work_stacks_empty(); 5156 verify_overflow_empty(); 5157 5158 if (should_unload_classes()) { 5159 CodeCache::gc_epilogue(); 5160 } 5161 JvmtiExport::gc_epilogue(); 5162 5163 // If we encountered any (marking stack / work queue) overflow 5164 // events during the current CMS cycle, take appropriate 5165 // remedial measures, where possible, so as to try and avoid 5166 // recurrence of that condition. 5167 assert(_markStack.isEmpty(), "No grey objects"); 5168 size_t ser_ovflw = _ser_pmc_remark_ovflw + _ser_pmc_preclean_ovflw + 5169 _ser_kac_ovflw + _ser_kac_preclean_ovflw; 5170 if (ser_ovflw > 0) { 5171 if (PrintCMSStatistics != 0) { 5172 gclog_or_tty->print_cr("Marking stack overflow (benign) " 5173 "(pmc_pc="SIZE_FORMAT", pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT 5174 ", kac_preclean="SIZE_FORMAT")", 5175 _ser_pmc_preclean_ovflw, _ser_pmc_remark_ovflw, 5176 _ser_kac_ovflw, _ser_kac_preclean_ovflw); 5177 } 5178 _markStack.expand(); 5179 _ser_pmc_remark_ovflw = 0; 5180 _ser_pmc_preclean_ovflw = 0; 5181 _ser_kac_preclean_ovflw = 0; 5182 _ser_kac_ovflw = 0; 5183 } 5184 if (_par_pmc_remark_ovflw > 0 || _par_kac_ovflw > 0) { 5185 if (PrintCMSStatistics != 0) { 5186 gclog_or_tty->print_cr("Work queue overflow (benign) " 5187 "(pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT")", 5188 _par_pmc_remark_ovflw, _par_kac_ovflw); 5189 } 5190 _par_pmc_remark_ovflw = 0; 5191 _par_kac_ovflw = 0; 5192 } 5193 if (PrintCMSStatistics != 0) { 5194 if (_markStack._hit_limit > 0) { 5195 gclog_or_tty->print_cr(" (benign) Hit max stack size limit ("SIZE_FORMAT")", 5196 _markStack._hit_limit); 5197 } 5198 if (_markStack._failed_double > 0) { 5199 gclog_or_tty->print_cr(" (benign) Failed stack doubling ("SIZE_FORMAT")," 5200 " current capacity "SIZE_FORMAT, 5201 _markStack._failed_double, 5202 _markStack.capacity()); 5203 } 5204 } 5205 _markStack._hit_limit = 0; 5206 _markStack._failed_double = 0; 5207 5208 if ((VerifyAfterGC || VerifyDuringGC) && 5209 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 5210 verify_after_remark(); 5211 } 5212 5213 _gc_tracer_cm->report_object_count_after_gc(&_is_alive_closure); 5214 5215 // Change under the freelistLocks. 5216 _collectorState = Sweeping; 5217 // Call isAllClear() under bitMapLock 5218 assert(_modUnionTable.isAllClear(), 5219 "Should be clear by end of the final marking"); 5220 assert(_ct->klass_rem_set()->mod_union_is_clear(), 5221 "Should be clear by end of the final marking"); 5222 if (UseAdaptiveSizePolicy) { 5223 size_policy()->checkpoint_roots_final_end(gch->gc_cause()); 5224 } 5225 } 5226 5227 void CMSParInitialMarkTask::work(uint worker_id) { 5228 elapsedTimer _timer; 5229 ResourceMark rm; 5230 HandleMark hm; 5231 5232 // ---------- scan from roots -------------- 5233 _timer.start(); 5234 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5235 Par_MarkRefsIntoClosure par_mri_cl(_collector->_span, &(_collector->_markBitMap)); 5236 5237 // ---------- young gen roots -------------- 5238 { 5239 work_on_young_gen_roots(worker_id, &par_mri_cl); 5240 _timer.stop(); 5241 if (PrintCMSStatistics != 0) { 5242 gclog_or_tty->print_cr( 5243 "Finished young gen initial mark scan work in %dth thread: %3.3f sec", 5244 worker_id, _timer.seconds()); 5245 } 5246 } 5247 5248 // ---------- remaining roots -------------- 5249 _timer.reset(); 5250 _timer.start(); 5251 5252 CLDToOopClosure cld_closure(&par_mri_cl, true); 5253 5254 gch->gen_process_roots(_collector->_cmsGen->level(), 5255 false, // yg was scanned above 5256 false, // this is parallel code 5257 GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()), 5258 _collector->should_unload_classes(), 5259 &par_mri_cl, 5260 NULL, 5261 &cld_closure); 5262 assert(_collector->should_unload_classes() 5263 || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache), 5264 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 5265 _timer.stop(); 5266 if (PrintCMSStatistics != 0) { 5267 gclog_or_tty->print_cr( 5268 "Finished remaining root initial mark scan work in %dth thread: %3.3f sec", 5269 worker_id, _timer.seconds()); 5270 } 5271 } 5272 5273 // Parallel remark task 5274 class CMSParRemarkTask: public CMSParMarkTask { 5275 CompactibleFreeListSpace* _cms_space; 5276 5277 // The per-thread work queues, available here for stealing. 5278 OopTaskQueueSet* _task_queues; 5279 ParallelTaskTerminator _term; 5280 5281 public: 5282 // A value of 0 passed to n_workers will cause the number of 5283 // workers to be taken from the active workers in the work gang. 5284 CMSParRemarkTask(CMSCollector* collector, 5285 CompactibleFreeListSpace* cms_space, 5286 int n_workers, FlexibleWorkGang* workers, 5287 OopTaskQueueSet* task_queues): 5288 CMSParMarkTask("Rescan roots and grey objects in parallel", 5289 collector, n_workers), 5290 _cms_space(cms_space), 5291 _task_queues(task_queues), 5292 _term(n_workers, task_queues) { } 5293 5294 OopTaskQueueSet* task_queues() { return _task_queues; } 5295 5296 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 5297 5298 ParallelTaskTerminator* terminator() { return &_term; } 5299 int n_workers() { return _n_workers; } 5300 5301 void work(uint worker_id); 5302 5303 private: 5304 // ... of dirty cards in old space 5305 void do_dirty_card_rescan_tasks(CompactibleFreeListSpace* sp, int i, 5306 Par_MarkRefsIntoAndScanClosure* cl); 5307 5308 // ... work stealing for the above 5309 void do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, int* seed); 5310 }; 5311 5312 class RemarkKlassClosure : public KlassClosure { 5313 KlassToOopClosure _cm_klass_closure; 5314 public: 5315 RemarkKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {} 5316 void do_klass(Klass* k) { 5317 // Check if we have modified any oops in the Klass during the concurrent marking. 5318 if (k->has_accumulated_modified_oops()) { 5319 k->clear_accumulated_modified_oops(); 5320 5321 // We could have transfered the current modified marks to the accumulated marks, 5322 // like we do with the Card Table to Mod Union Table. But it's not really necessary. 5323 } else if (k->has_modified_oops()) { 5324 // Don't clear anything, this info is needed by the next young collection. 5325 } else { 5326 // No modified oops in the Klass. 5327 return; 5328 } 5329 5330 // The klass has modified fields, need to scan the klass. 5331 _cm_klass_closure.do_klass(k); 5332 } 5333 }; 5334 5335 void CMSParMarkTask::work_on_young_gen_roots(uint worker_id, OopsInGenClosure* cl) { 5336 DefNewGeneration* dng = _collector->_young_gen->as_DefNewGeneration(); 5337 EdenSpace* eden_space = dng->eden(); 5338 ContiguousSpace* from_space = dng->from(); 5339 ContiguousSpace* to_space = dng->to(); 5340 5341 HeapWord** eca = _collector->_eden_chunk_array; 5342 size_t ect = _collector->_eden_chunk_index; 5343 HeapWord** sca = _collector->_survivor_chunk_array; 5344 size_t sct = _collector->_survivor_chunk_index; 5345 5346 assert(ect <= _collector->_eden_chunk_capacity, "out of bounds"); 5347 assert(sct <= _collector->_survivor_chunk_capacity, "out of bounds"); 5348 5349 do_young_space_rescan(worker_id, cl, to_space, NULL, 0); 5350 do_young_space_rescan(worker_id, cl, from_space, sca, sct); 5351 do_young_space_rescan(worker_id, cl, eden_space, eca, ect); 5352 } 5353 5354 // work_queue(i) is passed to the closure 5355 // Par_MarkRefsIntoAndScanClosure. The "i" parameter 5356 // also is passed to do_dirty_card_rescan_tasks() and to 5357 // do_work_steal() to select the i-th task_queue. 5358 5359 void CMSParRemarkTask::work(uint worker_id) { 5360 elapsedTimer _timer; 5361 ResourceMark rm; 5362 HandleMark hm; 5363 5364 // ---------- rescan from roots -------------- 5365 _timer.start(); 5366 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5367 Par_MarkRefsIntoAndScanClosure par_mrias_cl(_collector, 5368 _collector->_span, _collector->ref_processor(), 5369 &(_collector->_markBitMap), 5370 work_queue(worker_id)); 5371 5372 // Rescan young gen roots first since these are likely 5373 // coarsely partitioned and may, on that account, constitute 5374 // the critical path; thus, it's best to start off that 5375 // work first. 5376 // ---------- young gen roots -------------- 5377 { 5378 work_on_young_gen_roots(worker_id, &par_mrias_cl); 5379 _timer.stop(); 5380 if (PrintCMSStatistics != 0) { 5381 gclog_or_tty->print_cr( 5382 "Finished young gen rescan work in %dth thread: %3.3f sec", 5383 worker_id, _timer.seconds()); 5384 } 5385 } 5386 5387 // ---------- remaining roots -------------- 5388 _timer.reset(); 5389 _timer.start(); 5390 gch->gen_process_roots(_collector->_cmsGen->level(), 5391 false, // yg was scanned above 5392 false, // this is parallel code 5393 GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()), 5394 _collector->should_unload_classes(), 5395 &par_mrias_cl, 5396 NULL, 5397 NULL); // The dirty klasses will be handled below 5398 5399 assert(_collector->should_unload_classes() 5400 || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache), 5401 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 5402 _timer.stop(); 5403 if (PrintCMSStatistics != 0) { 5404 gclog_or_tty->print_cr( 5405 "Finished remaining root rescan work in %dth thread: %3.3f sec", 5406 worker_id, _timer.seconds()); 5407 } 5408 5409 // ---------- unhandled CLD scanning ---------- 5410 if (worker_id == 0) { // Single threaded at the moment. 5411 _timer.reset(); 5412 _timer.start(); 5413 5414 // Scan all new class loader data objects and new dependencies that were 5415 // introduced during concurrent marking. 5416 ResourceMark rm; 5417 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds(); 5418 for (int i = 0; i < array->length(); i++) { 5419 par_mrias_cl.do_class_loader_data(array->at(i)); 5420 } 5421 5422 // We don't need to keep track of new CLDs anymore. 5423 ClassLoaderDataGraph::remember_new_clds(false); 5424 5425 _timer.stop(); 5426 if (PrintCMSStatistics != 0) { 5427 gclog_or_tty->print_cr( 5428 "Finished unhandled CLD scanning work in %dth thread: %3.3f sec", 5429 worker_id, _timer.seconds()); 5430 } 5431 } 5432 5433 // ---------- dirty klass scanning ---------- 5434 if (worker_id == 0) { // Single threaded at the moment. 5435 _timer.reset(); 5436 _timer.start(); 5437 5438 // Scan all classes that was dirtied during the concurrent marking phase. 5439 RemarkKlassClosure remark_klass_closure(&par_mrias_cl); 5440 ClassLoaderDataGraph::classes_do(&remark_klass_closure); 5441 5442 _timer.stop(); 5443 if (PrintCMSStatistics != 0) { 5444 gclog_or_tty->print_cr( 5445 "Finished dirty klass scanning work in %dth thread: %3.3f sec", 5446 worker_id, _timer.seconds()); 5447 } 5448 } 5449 5450 // We might have added oops to ClassLoaderData::_handles during the 5451 // concurrent marking phase. These oops point to newly allocated objects 5452 // that are guaranteed to be kept alive. Either by the direct allocation 5453 // code, or when the young collector processes the roots. Hence, 5454 // we don't have to revisit the _handles block during the remark phase. 5455 5456 // ---------- rescan dirty cards ------------ 5457 _timer.reset(); 5458 _timer.start(); 5459 5460 // Do the rescan tasks for each of the two spaces 5461 // (cms_space) in turn. 5462 // "worker_id" is passed to select the task_queue for "worker_id" 5463 do_dirty_card_rescan_tasks(_cms_space, worker_id, &par_mrias_cl); 5464 _timer.stop(); 5465 if (PrintCMSStatistics != 0) { 5466 gclog_or_tty->print_cr( 5467 "Finished dirty card rescan work in %dth thread: %3.3f sec", 5468 worker_id, _timer.seconds()); 5469 } 5470 5471 // ---------- steal work from other threads ... 5472 // ---------- ... and drain overflow list. 5473 _timer.reset(); 5474 _timer.start(); 5475 do_work_steal(worker_id, &par_mrias_cl, _collector->hash_seed(worker_id)); 5476 _timer.stop(); 5477 if (PrintCMSStatistics != 0) { 5478 gclog_or_tty->print_cr( 5479 "Finished work stealing in %dth thread: %3.3f sec", 5480 worker_id, _timer.seconds()); 5481 } 5482 } 5483 5484 // Note that parameter "i" is not used. 5485 void 5486 CMSParMarkTask::do_young_space_rescan(uint worker_id, 5487 OopsInGenClosure* cl, ContiguousSpace* space, 5488 HeapWord** chunk_array, size_t chunk_top) { 5489 // Until all tasks completed: 5490 // . claim an unclaimed task 5491 // . compute region boundaries corresponding to task claimed 5492 // using chunk_array 5493 // . par_oop_iterate(cl) over that region 5494 5495 ResourceMark rm; 5496 HandleMark hm; 5497 5498 SequentialSubTasksDone* pst = space->par_seq_tasks(); 5499 5500 uint nth_task = 0; 5501 uint n_tasks = pst->n_tasks(); 5502 5503 if (n_tasks > 0) { 5504 assert(pst->valid(), "Uninitialized use?"); 5505 HeapWord *start, *end; 5506 while (!pst->is_task_claimed(/* reference */ nth_task)) { 5507 // We claimed task # nth_task; compute its boundaries. 5508 if (chunk_top == 0) { // no samples were taken 5509 assert(nth_task == 0 && n_tasks == 1, "Can have only 1 EdenSpace task"); 5510 start = space->bottom(); 5511 end = space->top(); 5512 } else if (nth_task == 0) { 5513 start = space->bottom(); 5514 end = chunk_array[nth_task]; 5515 } else if (nth_task < (uint)chunk_top) { 5516 assert(nth_task >= 1, "Control point invariant"); 5517 start = chunk_array[nth_task - 1]; 5518 end = chunk_array[nth_task]; 5519 } else { 5520 assert(nth_task == (uint)chunk_top, "Control point invariant"); 5521 start = chunk_array[chunk_top - 1]; 5522 end = space->top(); 5523 } 5524 MemRegion mr(start, end); 5525 // Verify that mr is in space 5526 assert(mr.is_empty() || space->used_region().contains(mr), 5527 "Should be in space"); 5528 // Verify that "start" is an object boundary 5529 assert(mr.is_empty() || oop(mr.start())->is_oop(), 5530 "Should be an oop"); 5531 space->par_oop_iterate(mr, cl); 5532 } 5533 pst->all_tasks_completed(); 5534 } 5535 } 5536 5537 void 5538 CMSParRemarkTask::do_dirty_card_rescan_tasks( 5539 CompactibleFreeListSpace* sp, int i, 5540 Par_MarkRefsIntoAndScanClosure* cl) { 5541 // Until all tasks completed: 5542 // . claim an unclaimed task 5543 // . compute region boundaries corresponding to task claimed 5544 // . transfer dirty bits ct->mut for that region 5545 // . apply rescanclosure to dirty mut bits for that region 5546 5547 ResourceMark rm; 5548 HandleMark hm; 5549 5550 OopTaskQueue* work_q = work_queue(i); 5551 ModUnionClosure modUnionClosure(&(_collector->_modUnionTable)); 5552 // CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! 5553 // CAUTION: This closure has state that persists across calls to 5554 // the work method dirty_range_iterate_clear() in that it has 5555 // imbedded in it a (subtype of) UpwardsObjectClosure. The 5556 // use of that state in the imbedded UpwardsObjectClosure instance 5557 // assumes that the cards are always iterated (even if in parallel 5558 // by several threads) in monotonically increasing order per each 5559 // thread. This is true of the implementation below which picks 5560 // card ranges (chunks) in monotonically increasing order globally 5561 // and, a-fortiori, in monotonically increasing order per thread 5562 // (the latter order being a subsequence of the former). 5563 // If the work code below is ever reorganized into a more chaotic 5564 // work-partitioning form than the current "sequential tasks" 5565 // paradigm, the use of that persistent state will have to be 5566 // revisited and modified appropriately. See also related 5567 // bug 4756801 work on which should examine this code to make 5568 // sure that the changes there do not run counter to the 5569 // assumptions made here and necessary for correctness and 5570 // efficiency. Note also that this code might yield inefficient 5571 // behaviour in the case of very large objects that span one or 5572 // more work chunks. Such objects would potentially be scanned 5573 // several times redundantly. Work on 4756801 should try and 5574 // address that performance anomaly if at all possible. XXX 5575 MemRegion full_span = _collector->_span; 5576 CMSBitMap* bm = &(_collector->_markBitMap); // shared 5577 MarkFromDirtyCardsClosure 5578 greyRescanClosure(_collector, full_span, // entire span of interest 5579 sp, bm, work_q, cl); 5580 5581 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 5582 assert(pst->valid(), "Uninitialized use?"); 5583 uint nth_task = 0; 5584 const int alignment = CardTableModRefBS::card_size * BitsPerWord; 5585 MemRegion span = sp->used_region(); 5586 HeapWord* start_addr = span.start(); 5587 HeapWord* end_addr = (HeapWord*)round_to((intptr_t)span.end(), 5588 alignment); 5589 const size_t chunk_size = sp->rescan_task_size(); // in HeapWord units 5590 assert((HeapWord*)round_to((intptr_t)start_addr, alignment) == 5591 start_addr, "Check alignment"); 5592 assert((size_t)round_to((intptr_t)chunk_size, alignment) == 5593 chunk_size, "Check alignment"); 5594 5595 while (!pst->is_task_claimed(/* reference */ nth_task)) { 5596 // Having claimed the nth_task, compute corresponding mem-region, 5597 // which is a-fortiori aligned correctly (i.e. at a MUT bopundary). 5598 // The alignment restriction ensures that we do not need any 5599 // synchronization with other gang-workers while setting or 5600 // clearing bits in thus chunk of the MUT. 5601 MemRegion this_span = MemRegion(start_addr + nth_task*chunk_size, 5602 start_addr + (nth_task+1)*chunk_size); 5603 // The last chunk's end might be way beyond end of the 5604 // used region. In that case pull back appropriately. 5605 if (this_span.end() > end_addr) { 5606 this_span.set_end(end_addr); 5607 assert(!this_span.is_empty(), "Program logic (calculation of n_tasks)"); 5608 } 5609 // Iterate over the dirty cards covering this chunk, marking them 5610 // precleaned, and setting the corresponding bits in the mod union 5611 // table. Since we have been careful to partition at Card and MUT-word 5612 // boundaries no synchronization is needed between parallel threads. 5613 _collector->_ct->ct_bs()->dirty_card_iterate(this_span, 5614 &modUnionClosure); 5615 5616 // Having transferred these marks into the modUnionTable, 5617 // rescan the marked objects on the dirty cards in the modUnionTable. 5618 // Even if this is at a synchronous collection, the initial marking 5619 // may have been done during an asynchronous collection so there 5620 // may be dirty bits in the mod-union table. 5621 _collector->_modUnionTable.dirty_range_iterate_clear( 5622 this_span, &greyRescanClosure); 5623 _collector->_modUnionTable.verifyNoOneBitsInRange( 5624 this_span.start(), 5625 this_span.end()); 5626 } 5627 pst->all_tasks_completed(); // declare that i am done 5628 } 5629 5630 // . see if we can share work_queues with ParNew? XXX 5631 void 5632 CMSParRemarkTask::do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, 5633 int* seed) { 5634 OopTaskQueue* work_q = work_queue(i); 5635 NOT_PRODUCT(int num_steals = 0;) 5636 oop obj_to_scan; 5637 CMSBitMap* bm = &(_collector->_markBitMap); 5638 5639 while (true) { 5640 // Completely finish any left over work from (an) earlier round(s) 5641 cl->trim_queue(0); 5642 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 5643 (size_t)ParGCDesiredObjsFromOverflowList); 5644 // Now check if there's any work in the overflow list 5645 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads, 5646 // only affects the number of attempts made to get work from the 5647 // overflow list and does not affect the number of workers. Just 5648 // pass ParallelGCThreads so this behavior is unchanged. 5649 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 5650 work_q, 5651 ParallelGCThreads)) { 5652 // found something in global overflow list; 5653 // not yet ready to go stealing work from others. 5654 // We'd like to assert(work_q->size() != 0, ...) 5655 // because we just took work from the overflow list, 5656 // but of course we can't since all of that could have 5657 // been already stolen from us. 5658 // "He giveth and He taketh away." 5659 continue; 5660 } 5661 // Verify that we have no work before we resort to stealing 5662 assert(work_q->size() == 0, "Have work, shouldn't steal"); 5663 // Try to steal from other queues that have work 5664 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 5665 NOT_PRODUCT(num_steals++;) 5666 assert(obj_to_scan->is_oop(), "Oops, not an oop!"); 5667 assert(bm->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 5668 // Do scanning work 5669 obj_to_scan->oop_iterate(cl); 5670 // Loop around, finish this work, and try to steal some more 5671 } else if (terminator()->offer_termination()) { 5672 break; // nirvana from the infinite cycle 5673 } 5674 } 5675 NOT_PRODUCT( 5676 if (PrintCMSStatistics != 0) { 5677 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals); 5678 } 5679 ) 5680 assert(work_q->size() == 0 && _collector->overflow_list_is_empty(), 5681 "Else our work is not yet done"); 5682 } 5683 5684 // Record object boundaries in _eden_chunk_array by sampling the eden 5685 // top in the slow-path eden object allocation code path and record 5686 // the boundaries, if CMSEdenChunksRecordAlways is true. If 5687 // CMSEdenChunksRecordAlways is false, we use the other asynchronous 5688 // sampling in sample_eden() that activates during the part of the 5689 // preclean phase. 5690 void CMSCollector::sample_eden_chunk() { 5691 if (CMSEdenChunksRecordAlways && _eden_chunk_array != NULL) { 5692 if (_eden_chunk_lock->try_lock()) { 5693 // Record a sample. This is the critical section. The contents 5694 // of the _eden_chunk_array have to be non-decreasing in the 5695 // address order. 5696 _eden_chunk_array[_eden_chunk_index] = *_top_addr; 5697 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr, 5698 "Unexpected state of Eden"); 5699 if (_eden_chunk_index == 0 || 5700 ((_eden_chunk_array[_eden_chunk_index] > _eden_chunk_array[_eden_chunk_index-1]) && 5701 (pointer_delta(_eden_chunk_array[_eden_chunk_index], 5702 _eden_chunk_array[_eden_chunk_index-1]) >= CMSSamplingGrain))) { 5703 _eden_chunk_index++; // commit sample 5704 } 5705 _eden_chunk_lock->unlock(); 5706 } 5707 } 5708 } 5709 5710 // Return a thread-local PLAB recording array, as appropriate. 5711 void* CMSCollector::get_data_recorder(int thr_num) { 5712 if (_survivor_plab_array != NULL && 5713 (CMSPLABRecordAlways || 5714 (_collectorState > Marking && _collectorState < FinalMarking))) { 5715 assert(thr_num < (int)ParallelGCThreads, "thr_num is out of bounds"); 5716 ChunkArray* ca = &_survivor_plab_array[thr_num]; 5717 ca->reset(); // clear it so that fresh data is recorded 5718 return (void*) ca; 5719 } else { 5720 return NULL; 5721 } 5722 } 5723 5724 // Reset all the thread-local PLAB recording arrays 5725 void CMSCollector::reset_survivor_plab_arrays() { 5726 for (uint i = 0; i < ParallelGCThreads; i++) { 5727 _survivor_plab_array[i].reset(); 5728 } 5729 } 5730 5731 // Merge the per-thread plab arrays into the global survivor chunk 5732 // array which will provide the partitioning of the survivor space 5733 // for CMS initial scan and rescan. 5734 void CMSCollector::merge_survivor_plab_arrays(ContiguousSpace* surv, 5735 int no_of_gc_threads) { 5736 assert(_survivor_plab_array != NULL, "Error"); 5737 assert(_survivor_chunk_array != NULL, "Error"); 5738 assert(_collectorState == FinalMarking || 5739 (CMSParallelInitialMarkEnabled && _collectorState == InitialMarking), "Error"); 5740 for (int j = 0; j < no_of_gc_threads; j++) { 5741 _cursor[j] = 0; 5742 } 5743 HeapWord* top = surv->top(); 5744 size_t i; 5745 for (i = 0; i < _survivor_chunk_capacity; i++) { // all sca entries 5746 HeapWord* min_val = top; // Higher than any PLAB address 5747 uint min_tid = 0; // position of min_val this round 5748 for (int j = 0; j < no_of_gc_threads; j++) { 5749 ChunkArray* cur_sca = &_survivor_plab_array[j]; 5750 if (_cursor[j] == cur_sca->end()) { 5751 continue; 5752 } 5753 assert(_cursor[j] < cur_sca->end(), "ctl pt invariant"); 5754 HeapWord* cur_val = cur_sca->nth(_cursor[j]); 5755 assert(surv->used_region().contains(cur_val), "Out of bounds value"); 5756 if (cur_val < min_val) { 5757 min_tid = j; 5758 min_val = cur_val; 5759 } else { 5760 assert(cur_val < top, "All recorded addresses should be less"); 5761 } 5762 } 5763 // At this point min_val and min_tid are respectively 5764 // the least address in _survivor_plab_array[j]->nth(_cursor[j]) 5765 // and the thread (j) that witnesses that address. 5766 // We record this address in the _survivor_chunk_array[i] 5767 // and increment _cursor[min_tid] prior to the next round i. 5768 if (min_val == top) { 5769 break; 5770 } 5771 _survivor_chunk_array[i] = min_val; 5772 _cursor[min_tid]++; 5773 } 5774 // We are all done; record the size of the _survivor_chunk_array 5775 _survivor_chunk_index = i; // exclusive: [0, i) 5776 if (PrintCMSStatistics > 0) { 5777 gclog_or_tty->print(" (Survivor:" SIZE_FORMAT "chunks) ", i); 5778 } 5779 // Verify that we used up all the recorded entries 5780 #ifdef ASSERT 5781 size_t total = 0; 5782 for (int j = 0; j < no_of_gc_threads; j++) { 5783 assert(_cursor[j] == _survivor_plab_array[j].end(), "Ctl pt invariant"); 5784 total += _cursor[j]; 5785 } 5786 assert(total == _survivor_chunk_index, "Ctl Pt Invariant"); 5787 // Check that the merged array is in sorted order 5788 if (total > 0) { 5789 for (size_t i = 0; i < total - 1; i++) { 5790 if (PrintCMSStatistics > 0) { 5791 gclog_or_tty->print(" (chunk" SIZE_FORMAT ":" INTPTR_FORMAT ") ", 5792 i, _survivor_chunk_array[i]); 5793 } 5794 assert(_survivor_chunk_array[i] < _survivor_chunk_array[i+1], 5795 "Not sorted"); 5796 } 5797 } 5798 #endif // ASSERT 5799 } 5800 5801 // Set up the space's par_seq_tasks structure for work claiming 5802 // for parallel initial scan and rescan of young gen. 5803 // See ParRescanTask where this is currently used. 5804 void 5805 CMSCollector:: 5806 initialize_sequential_subtasks_for_young_gen_rescan(int n_threads) { 5807 assert(n_threads > 0, "Unexpected n_threads argument"); 5808 DefNewGeneration* dng = (DefNewGeneration*)_young_gen; 5809 5810 // Eden space 5811 if (!dng->eden()->is_empty()) { 5812 SequentialSubTasksDone* pst = dng->eden()->par_seq_tasks(); 5813 assert(!pst->valid(), "Clobbering existing data?"); 5814 // Each valid entry in [0, _eden_chunk_index) represents a task. 5815 size_t n_tasks = _eden_chunk_index + 1; 5816 assert(n_tasks == 1 || _eden_chunk_array != NULL, "Error"); 5817 // Sets the condition for completion of the subtask (how many threads 5818 // need to finish in order to be done). 5819 pst->set_n_threads(n_threads); 5820 pst->set_n_tasks((int)n_tasks); 5821 } 5822 5823 // Merge the survivor plab arrays into _survivor_chunk_array 5824 if (_survivor_plab_array != NULL) { 5825 merge_survivor_plab_arrays(dng->from(), n_threads); 5826 } else { 5827 assert(_survivor_chunk_index == 0, "Error"); 5828 } 5829 5830 // To space 5831 { 5832 SequentialSubTasksDone* pst = dng->to()->par_seq_tasks(); 5833 assert(!pst->valid(), "Clobbering existing data?"); 5834 // Sets the condition for completion of the subtask (how many threads 5835 // need to finish in order to be done). 5836 pst->set_n_threads(n_threads); 5837 pst->set_n_tasks(1); 5838 assert(pst->valid(), "Error"); 5839 } 5840 5841 // From space 5842 { 5843 SequentialSubTasksDone* pst = dng->from()->par_seq_tasks(); 5844 assert(!pst->valid(), "Clobbering existing data?"); 5845 size_t n_tasks = _survivor_chunk_index + 1; 5846 assert(n_tasks == 1 || _survivor_chunk_array != NULL, "Error"); 5847 // Sets the condition for completion of the subtask (how many threads 5848 // need to finish in order to be done). 5849 pst->set_n_threads(n_threads); 5850 pst->set_n_tasks((int)n_tasks); 5851 assert(pst->valid(), "Error"); 5852 } 5853 } 5854 5855 // Parallel version of remark 5856 void CMSCollector::do_remark_parallel() { 5857 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5858 FlexibleWorkGang* workers = gch->workers(); 5859 assert(workers != NULL, "Need parallel worker threads."); 5860 // Choose to use the number of GC workers most recently set 5861 // into "active_workers". If active_workers is not set, set it 5862 // to ParallelGCThreads. 5863 int n_workers = workers->active_workers(); 5864 if (n_workers == 0) { 5865 assert(n_workers > 0, "Should have been set during scavenge"); 5866 n_workers = ParallelGCThreads; 5867 workers->set_active_workers(n_workers); 5868 } 5869 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 5870 5871 CMSParRemarkTask tsk(this, 5872 cms_space, 5873 n_workers, workers, task_queues()); 5874 5875 // Set up for parallel process_roots work. 5876 gch->set_par_threads(n_workers); 5877 // We won't be iterating over the cards in the card table updating 5878 // the younger_gen cards, so we shouldn't call the following else 5879 // the verification code as well as subsequent younger_refs_iterate 5880 // code would get confused. XXX 5881 // gch->rem_set()->prepare_for_younger_refs_iterate(true); // parallel 5882 5883 // The young gen rescan work will not be done as part of 5884 // process_roots (which currently doesn't know how to 5885 // parallelize such a scan), but rather will be broken up into 5886 // a set of parallel tasks (via the sampling that the [abortable] 5887 // preclean phase did of EdenSpace, plus the [two] tasks of 5888 // scanning the [two] survivor spaces. Further fine-grain 5889 // parallelization of the scanning of the survivor spaces 5890 // themselves, and of precleaning of the younger gen itself 5891 // is deferred to the future. 5892 initialize_sequential_subtasks_for_young_gen_rescan(n_workers); 5893 5894 // The dirty card rescan work is broken up into a "sequence" 5895 // of parallel tasks (per constituent space) that are dynamically 5896 // claimed by the parallel threads. 5897 cms_space->initialize_sequential_subtasks_for_rescan(n_workers); 5898 5899 // It turns out that even when we're using 1 thread, doing the work in a 5900 // separate thread causes wide variance in run times. We can't help this 5901 // in the multi-threaded case, but we special-case n=1 here to get 5902 // repeatable measurements of the 1-thread overhead of the parallel code. 5903 if (n_workers > 1) { 5904 // Make refs discovery MT-safe, if it isn't already: it may not 5905 // necessarily be so, since it's possible that we are doing 5906 // ST marking. 5907 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), true); 5908 GenCollectedHeap::StrongRootsScope srs(gch); 5909 workers->run_task(&tsk); 5910 } else { 5911 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false); 5912 GenCollectedHeap::StrongRootsScope srs(gch); 5913 tsk.work(0); 5914 } 5915 5916 gch->set_par_threads(0); // 0 ==> non-parallel. 5917 // restore, single-threaded for now, any preserved marks 5918 // as a result of work_q overflow 5919 restore_preserved_marks_if_any(); 5920 } 5921 5922 // Non-parallel version of remark 5923 void CMSCollector::do_remark_non_parallel() { 5924 ResourceMark rm; 5925 HandleMark hm; 5926 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5927 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false); 5928 5929 MarkRefsIntoAndScanClosure 5930 mrias_cl(_span, ref_processor(), &_markBitMap, NULL /* not precleaning */, 5931 &_markStack, this, 5932 false /* should_yield */, false /* not precleaning */); 5933 MarkFromDirtyCardsClosure 5934 markFromDirtyCardsClosure(this, _span, 5935 NULL, // space is set further below 5936 &_markBitMap, &_markStack, &mrias_cl); 5937 { 5938 GCTraceTime t("grey object rescan", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 5939 // Iterate over the dirty cards, setting the corresponding bits in the 5940 // mod union table. 5941 { 5942 ModUnionClosure modUnionClosure(&_modUnionTable); 5943 _ct->ct_bs()->dirty_card_iterate( 5944 _cmsGen->used_region(), 5945 &modUnionClosure); 5946 } 5947 // Having transferred these marks into the modUnionTable, we just need 5948 // to rescan the marked objects on the dirty cards in the modUnionTable. 5949 // The initial marking may have been done during an asynchronous 5950 // collection so there may be dirty bits in the mod-union table. 5951 const int alignment = 5952 CardTableModRefBS::card_size * BitsPerWord; 5953 { 5954 // ... First handle dirty cards in CMS gen 5955 markFromDirtyCardsClosure.set_space(_cmsGen->cmsSpace()); 5956 MemRegion ur = _cmsGen->used_region(); 5957 HeapWord* lb = ur.start(); 5958 HeapWord* ub = (HeapWord*)round_to((intptr_t)ur.end(), alignment); 5959 MemRegion cms_span(lb, ub); 5960 _modUnionTable.dirty_range_iterate_clear(cms_span, 5961 &markFromDirtyCardsClosure); 5962 verify_work_stacks_empty(); 5963 if (PrintCMSStatistics != 0) { 5964 gclog_or_tty->print(" (re-scanned "SIZE_FORMAT" dirty cards in cms gen) ", 5965 markFromDirtyCardsClosure.num_dirty_cards()); 5966 } 5967 } 5968 } 5969 if (VerifyDuringGC && 5970 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 5971 HandleMark hm; // Discard invalid handles created during verification 5972 Universe::verify(); 5973 } 5974 { 5975 GCTraceTime t("root rescan", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 5976 5977 verify_work_stacks_empty(); 5978 5979 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 5980 GenCollectedHeap::StrongRootsScope srs(gch); 5981 5982 gch->gen_process_roots(_cmsGen->level(), 5983 true, // younger gens as roots 5984 false, // use the local StrongRootsScope 5985 GenCollectedHeap::ScanningOption(roots_scanning_options()), 5986 should_unload_classes(), 5987 &mrias_cl, 5988 NULL, 5989 NULL); // The dirty klasses will be handled below 5990 5991 assert(should_unload_classes() 5992 || (roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache), 5993 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 5994 } 5995 5996 { 5997 GCTraceTime t("visit unhandled CLDs", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 5998 5999 verify_work_stacks_empty(); 6000 6001 // Scan all class loader data objects that might have been introduced 6002 // during concurrent marking. 6003 ResourceMark rm; 6004 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds(); 6005 for (int i = 0; i < array->length(); i++) { 6006 mrias_cl.do_class_loader_data(array->at(i)); 6007 } 6008 6009 // We don't need to keep track of new CLDs anymore. 6010 ClassLoaderDataGraph::remember_new_clds(false); 6011 6012 verify_work_stacks_empty(); 6013 } 6014 6015 { 6016 GCTraceTime t("dirty klass scan", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 6017 6018 verify_work_stacks_empty(); 6019 6020 RemarkKlassClosure remark_klass_closure(&mrias_cl); 6021 ClassLoaderDataGraph::classes_do(&remark_klass_closure); 6022 6023 verify_work_stacks_empty(); 6024 } 6025 6026 // We might have added oops to ClassLoaderData::_handles during the 6027 // concurrent marking phase. These oops point to newly allocated objects 6028 // that are guaranteed to be kept alive. Either by the direct allocation 6029 // code, or when the young collector processes the roots. Hence, 6030 // we don't have to revisit the _handles block during the remark phase. 6031 6032 verify_work_stacks_empty(); 6033 // Restore evacuated mark words, if any, used for overflow list links 6034 if (!CMSOverflowEarlyRestoration) { 6035 restore_preserved_marks_if_any(); 6036 } 6037 verify_overflow_empty(); 6038 } 6039 6040 //////////////////////////////////////////////////////// 6041 // Parallel Reference Processing Task Proxy Class 6042 //////////////////////////////////////////////////////// 6043 class CMSRefProcTaskProxy: public AbstractGangTaskWOopQueues { 6044 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask; 6045 CMSCollector* _collector; 6046 CMSBitMap* _mark_bit_map; 6047 const MemRegion _span; 6048 ProcessTask& _task; 6049 6050 public: 6051 CMSRefProcTaskProxy(ProcessTask& task, 6052 CMSCollector* collector, 6053 const MemRegion& span, 6054 CMSBitMap* mark_bit_map, 6055 AbstractWorkGang* workers, 6056 OopTaskQueueSet* task_queues): 6057 // XXX Should superclass AGTWOQ also know about AWG since it knows 6058 // about the task_queues used by the AWG? Then it could initialize 6059 // the terminator() object. See 6984287. The set_for_termination() 6060 // below is a temporary band-aid for the regression in 6984287. 6061 AbstractGangTaskWOopQueues("Process referents by policy in parallel", 6062 task_queues), 6063 _task(task), 6064 _collector(collector), _span(span), _mark_bit_map(mark_bit_map) 6065 { 6066 assert(_collector->_span.equals(_span) && !_span.is_empty(), 6067 "Inconsistency in _span"); 6068 set_for_termination(workers->active_workers()); 6069 } 6070 6071 OopTaskQueueSet* task_queues() { return queues(); } 6072 6073 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 6074 6075 void do_work_steal(int i, 6076 CMSParDrainMarkingStackClosure* drain, 6077 CMSParKeepAliveClosure* keep_alive, 6078 int* seed); 6079 6080 virtual void work(uint worker_id); 6081 }; 6082 6083 void CMSRefProcTaskProxy::work(uint worker_id) { 6084 ResourceMark rm; 6085 HandleMark hm; 6086 assert(_collector->_span.equals(_span), "Inconsistency in _span"); 6087 CMSParKeepAliveClosure par_keep_alive(_collector, _span, 6088 _mark_bit_map, 6089 work_queue(worker_id)); 6090 CMSParDrainMarkingStackClosure par_drain_stack(_collector, _span, 6091 _mark_bit_map, 6092 work_queue(worker_id)); 6093 CMSIsAliveClosure is_alive_closure(_span, _mark_bit_map); 6094 _task.work(worker_id, is_alive_closure, par_keep_alive, par_drain_stack); 6095 if (_task.marks_oops_alive()) { 6096 do_work_steal(worker_id, &par_drain_stack, &par_keep_alive, 6097 _collector->hash_seed(worker_id)); 6098 } 6099 assert(work_queue(worker_id)->size() == 0, "work_queue should be empty"); 6100 assert(_collector->_overflow_list == NULL, "non-empty _overflow_list"); 6101 } 6102 6103 class CMSRefEnqueueTaskProxy: public AbstractGangTask { 6104 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask; 6105 EnqueueTask& _task; 6106 6107 public: 6108 CMSRefEnqueueTaskProxy(EnqueueTask& task) 6109 : AbstractGangTask("Enqueue reference objects in parallel"), 6110 _task(task) 6111 { } 6112 6113 virtual void work(uint worker_id) 6114 { 6115 _task.work(worker_id); 6116 } 6117 }; 6118 6119 CMSParKeepAliveClosure::CMSParKeepAliveClosure(CMSCollector* collector, 6120 MemRegion span, CMSBitMap* bit_map, OopTaskQueue* work_queue): 6121 _span(span), 6122 _bit_map(bit_map), 6123 _work_queue(work_queue), 6124 _mark_and_push(collector, span, bit_map, work_queue), 6125 _low_water_mark(MIN2((uint)(work_queue->max_elems()/4), 6126 (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads))) 6127 { } 6128 6129 // . see if we can share work_queues with ParNew? XXX 6130 void CMSRefProcTaskProxy::do_work_steal(int i, 6131 CMSParDrainMarkingStackClosure* drain, 6132 CMSParKeepAliveClosure* keep_alive, 6133 int* seed) { 6134 OopTaskQueue* work_q = work_queue(i); 6135 NOT_PRODUCT(int num_steals = 0;) 6136 oop obj_to_scan; 6137 6138 while (true) { 6139 // Completely finish any left over work from (an) earlier round(s) 6140 drain->trim_queue(0); 6141 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 6142 (size_t)ParGCDesiredObjsFromOverflowList); 6143 // Now check if there's any work in the overflow list 6144 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads, 6145 // only affects the number of attempts made to get work from the 6146 // overflow list and does not affect the number of workers. Just 6147 // pass ParallelGCThreads so this behavior is unchanged. 6148 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 6149 work_q, 6150 ParallelGCThreads)) { 6151 // Found something in global overflow list; 6152 // not yet ready to go stealing work from others. 6153 // We'd like to assert(work_q->size() != 0, ...) 6154 // because we just took work from the overflow list, 6155 // but of course we can't, since all of that might have 6156 // been already stolen from us. 6157 continue; 6158 } 6159 // Verify that we have no work before we resort to stealing 6160 assert(work_q->size() == 0, "Have work, shouldn't steal"); 6161 // Try to steal from other queues that have work 6162 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 6163 NOT_PRODUCT(num_steals++;) 6164 assert(obj_to_scan->is_oop(), "Oops, not an oop!"); 6165 assert(_mark_bit_map->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 6166 // Do scanning work 6167 obj_to_scan->oop_iterate(keep_alive); 6168 // Loop around, finish this work, and try to steal some more 6169 } else if (terminator()->offer_termination()) { 6170 break; // nirvana from the infinite cycle 6171 } 6172 } 6173 NOT_PRODUCT( 6174 if (PrintCMSStatistics != 0) { 6175 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals); 6176 } 6177 ) 6178 } 6179 6180 void CMSRefProcTaskExecutor::execute(ProcessTask& task) 6181 { 6182 GenCollectedHeap* gch = GenCollectedHeap::heap(); 6183 FlexibleWorkGang* workers = gch->workers(); 6184 assert(workers != NULL, "Need parallel worker threads."); 6185 CMSRefProcTaskProxy rp_task(task, &_collector, 6186 _collector.ref_processor()->span(), 6187 _collector.markBitMap(), 6188 workers, _collector.task_queues()); 6189 workers->run_task(&rp_task); 6190 } 6191 6192 void CMSRefProcTaskExecutor::execute(EnqueueTask& task) 6193 { 6194 6195 GenCollectedHeap* gch = GenCollectedHeap::heap(); 6196 FlexibleWorkGang* workers = gch->workers(); 6197 assert(workers != NULL, "Need parallel worker threads."); 6198 CMSRefEnqueueTaskProxy enq_task(task); 6199 workers->run_task(&enq_task); 6200 } 6201 6202 void CMSCollector::refProcessingWork(bool asynch, bool clear_all_soft_refs) { 6203 6204 ResourceMark rm; 6205 HandleMark hm; 6206 6207 ReferenceProcessor* rp = ref_processor(); 6208 assert(rp->span().equals(_span), "Spans should be equal"); 6209 assert(!rp->enqueuing_is_done(), "Enqueuing should not be complete"); 6210 // Process weak references. 6211 rp->setup_policy(clear_all_soft_refs); 6212 verify_work_stacks_empty(); 6213 6214 CMSKeepAliveClosure cmsKeepAliveClosure(this, _span, &_markBitMap, 6215 &_markStack, false /* !preclean */); 6216 CMSDrainMarkingStackClosure cmsDrainMarkingStackClosure(this, 6217 _span, &_markBitMap, &_markStack, 6218 &cmsKeepAliveClosure, false /* !preclean */); 6219 { 6220 GCTraceTime t("weak refs processing", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 6221 6222 ReferenceProcessorStats stats; 6223 if (rp->processing_is_mt()) { 6224 // Set the degree of MT here. If the discovery is done MT, there 6225 // may have been a different number of threads doing the discovery 6226 // and a different number of discovered lists may have Ref objects. 6227 // That is OK as long as the Reference lists are balanced (see 6228 // balance_all_queues() and balance_queues()). 6229 GenCollectedHeap* gch = GenCollectedHeap::heap(); 6230 int active_workers = ParallelGCThreads; 6231 FlexibleWorkGang* workers = gch->workers(); 6232 if (workers != NULL) { 6233 active_workers = workers->active_workers(); 6234 // The expectation is that active_workers will have already 6235 // been set to a reasonable value. If it has not been set, 6236 // investigate. 6237 assert(active_workers > 0, "Should have been set during scavenge"); 6238 } 6239 rp->set_active_mt_degree(active_workers); 6240 CMSRefProcTaskExecutor task_executor(*this); 6241 stats = rp->process_discovered_references(&_is_alive_closure, 6242 &cmsKeepAliveClosure, 6243 &cmsDrainMarkingStackClosure, 6244 &task_executor, 6245 _gc_timer_cm, 6246 _gc_tracer_cm->gc_id()); 6247 } else { 6248 stats = rp->process_discovered_references(&_is_alive_closure, 6249 &cmsKeepAliveClosure, 6250 &cmsDrainMarkingStackClosure, 6251 NULL, 6252 _gc_timer_cm, 6253 _gc_tracer_cm->gc_id()); 6254 } 6255 _gc_tracer_cm->report_gc_reference_stats(stats); 6256 6257 } 6258 6259 // This is the point where the entire marking should have completed. 6260 verify_work_stacks_empty(); 6261 6262 if (should_unload_classes()) { 6263 { 6264 GCTraceTime t("class unloading", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 6265 6266 // Unload classes and purge the SystemDictionary. 6267 bool purged_class = SystemDictionary::do_unloading(&_is_alive_closure); 6268 6269 // Unload nmethods. 6270 CodeCache::do_unloading(&_is_alive_closure, purged_class); 6271 6272 // Prune dead klasses from subklass/sibling/implementor lists. 6273 Klass::clean_weak_klass_links(&_is_alive_closure); 6274 } 6275 6276 { 6277 GCTraceTime t("scrub symbol table", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 6278 // Clean up unreferenced symbols in symbol table. 6279 SymbolTable::unlink(); 6280 } 6281 6282 { 6283 GCTraceTime t("scrub string table", PrintGCDetails, false, _gc_timer_cm, _gc_tracer_cm->gc_id()); 6284 // Delete entries for dead interned strings. 6285 StringTable::unlink(&_is_alive_closure); 6286 } 6287 } 6288 6289 6290 // Restore any preserved marks as a result of mark stack or 6291 // work queue overflow 6292 restore_preserved_marks_if_any(); // done single-threaded for now 6293 6294 rp->set_enqueuing_is_done(true); 6295 if (rp->processing_is_mt()) { 6296 rp->balance_all_queues(); 6297 CMSRefProcTaskExecutor task_executor(*this); 6298 rp->enqueue_discovered_references(&task_executor); 6299 } else { 6300 rp->enqueue_discovered_references(NULL); 6301 } 6302 rp->verify_no_references_recorded(); 6303 assert(!rp->discovery_enabled(), "should have been disabled"); 6304 } 6305 6306 #ifndef PRODUCT 6307 void CMSCollector::check_correct_thread_executing() { 6308 Thread* t = Thread::current(); 6309 // Only the VM thread or the CMS thread should be here. 6310 assert(t->is_ConcurrentGC_thread() || t->is_VM_thread(), 6311 "Unexpected thread type"); 6312 // If this is the vm thread, the foreground process 6313 // should not be waiting. Note that _foregroundGCIsActive is 6314 // true while the foreground collector is waiting. 6315 if (_foregroundGCShouldWait) { 6316 // We cannot be the VM thread 6317 assert(t->is_ConcurrentGC_thread(), 6318 "Should be CMS thread"); 6319 } else { 6320 // We can be the CMS thread only if we are in a stop-world 6321 // phase of CMS collection. 6322 if (t->is_ConcurrentGC_thread()) { 6323 assert(_collectorState == InitialMarking || 6324 _collectorState == FinalMarking, 6325 "Should be a stop-world phase"); 6326 // The CMS thread should be holding the CMS_token. 6327 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6328 "Potential interference with concurrently " 6329 "executing VM thread"); 6330 } 6331 } 6332 } 6333 #endif 6334 6335 void CMSCollector::sweep(bool asynch) { 6336 assert(_collectorState == Sweeping, "just checking"); 6337 check_correct_thread_executing(); 6338 verify_work_stacks_empty(); 6339 verify_overflow_empty(); 6340 increment_sweep_count(); 6341 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 6342 6343 _inter_sweep_timer.stop(); 6344 _inter_sweep_estimate.sample(_inter_sweep_timer.seconds()); 6345 size_policy()->avg_cms_free_at_sweep()->sample(_cmsGen->free()); 6346 6347 assert(!_intra_sweep_timer.is_active(), "Should not be active"); 6348 _intra_sweep_timer.reset(); 6349 _intra_sweep_timer.start(); 6350 if (asynch) { 6351 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 6352 CMSPhaseAccounting pa(this, "sweep", _gc_tracer_cm->gc_id(), !PrintGCDetails); 6353 // First sweep the old gen 6354 { 6355 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(), 6356 bitMapLock()); 6357 sweepWork(_cmsGen, asynch); 6358 } 6359 6360 // Update Universe::_heap_*_at_gc figures. 6361 // We need all the free list locks to make the abstract state 6362 // transition from Sweeping to Resetting. See detailed note 6363 // further below. 6364 { 6365 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock()); 6366 // Update heap occupancy information which is used as 6367 // input to soft ref clearing policy at the next gc. 6368 Universe::update_heap_info_at_gc(); 6369 _collectorState = Resizing; 6370 } 6371 } else { 6372 // already have needed locks 6373 sweepWork(_cmsGen, asynch); 6374 // Update heap occupancy information which is used as 6375 // input to soft ref clearing policy at the next gc. 6376 Universe::update_heap_info_at_gc(); 6377 _collectorState = Resizing; 6378 } 6379 verify_work_stacks_empty(); 6380 verify_overflow_empty(); 6381 6382 if (should_unload_classes()) { 6383 // Delay purge to the beginning of the next safepoint. Metaspace::contains 6384 // requires that the virtual spaces are stable and not deleted. 6385 ClassLoaderDataGraph::set_should_purge(true); 6386 } 6387 6388 _intra_sweep_timer.stop(); 6389 _intra_sweep_estimate.sample(_intra_sweep_timer.seconds()); 6390 6391 _inter_sweep_timer.reset(); 6392 _inter_sweep_timer.start(); 6393 6394 // We need to use a monotonically non-deccreasing time in ms 6395 // or we will see time-warp warnings and os::javaTimeMillis() 6396 // does not guarantee monotonicity. 6397 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC; 6398 update_time_of_last_gc(now); 6399 6400 // NOTE on abstract state transitions: 6401 // Mutators allocate-live and/or mark the mod-union table dirty 6402 // based on the state of the collection. The former is done in 6403 // the interval [Marking, Sweeping] and the latter in the interval 6404 // [Marking, Sweeping). Thus the transitions into the Marking state 6405 // and out of the Sweeping state must be synchronously visible 6406 // globally to the mutators. 6407 // The transition into the Marking state happens with the world 6408 // stopped so the mutators will globally see it. Sweeping is 6409 // done asynchronously by the background collector so the transition 6410 // from the Sweeping state to the Resizing state must be done 6411 // under the freelistLock (as is the check for whether to 6412 // allocate-live and whether to dirty the mod-union table). 6413 assert(_collectorState == Resizing, "Change of collector state to" 6414 " Resizing must be done under the freelistLocks (plural)"); 6415 6416 // Now that sweeping has been completed, we clear 6417 // the incremental_collection_failed flag, 6418 // thus inviting a younger gen collection to promote into 6419 // this generation. If such a promotion may still fail, 6420 // the flag will be set again when a young collection is 6421 // attempted. 6422 GenCollectedHeap* gch = GenCollectedHeap::heap(); 6423 gch->clear_incremental_collection_failed(); // Worth retrying as fresh space may have been freed up 6424 gch->update_full_collections_completed(_collection_count_start); 6425 } 6426 6427 // FIX ME!!! Looks like this belongs in CFLSpace, with 6428 // CMSGen merely delegating to it. 6429 void ConcurrentMarkSweepGeneration::setNearLargestChunk() { 6430 double nearLargestPercent = FLSLargestBlockCoalesceProximity; 6431 HeapWord* minAddr = _cmsSpace->bottom(); 6432 HeapWord* largestAddr = 6433 (HeapWord*) _cmsSpace->dictionary()->find_largest_dict(); 6434 if (largestAddr == NULL) { 6435 // The dictionary appears to be empty. In this case 6436 // try to coalesce at the end of the heap. 6437 largestAddr = _cmsSpace->end(); 6438 } 6439 size_t largestOffset = pointer_delta(largestAddr, minAddr); 6440 size_t nearLargestOffset = 6441 (size_t)((double)largestOffset * nearLargestPercent) - MinChunkSize; 6442 if (PrintFLSStatistics != 0) { 6443 gclog_or_tty->print_cr( 6444 "CMS: Large Block: " PTR_FORMAT ";" 6445 " Proximity: " PTR_FORMAT " -> " PTR_FORMAT, 6446 largestAddr, 6447 _cmsSpace->nearLargestChunk(), minAddr + nearLargestOffset); 6448 } 6449 _cmsSpace->set_nearLargestChunk(minAddr + nearLargestOffset); 6450 } 6451 6452 bool ConcurrentMarkSweepGeneration::isNearLargestChunk(HeapWord* addr) { 6453 return addr >= _cmsSpace->nearLargestChunk(); 6454 } 6455 6456 FreeChunk* ConcurrentMarkSweepGeneration::find_chunk_at_end() { 6457 return _cmsSpace->find_chunk_at_end(); 6458 } 6459 6460 void ConcurrentMarkSweepGeneration::update_gc_stats(int current_level, 6461 bool full) { 6462 // The next lower level has been collected. Gather any statistics 6463 // that are of interest at this point. 6464 if (!full && (current_level + 1) == level()) { 6465 // Gather statistics on the young generation collection. 6466 collector()->stats().record_gc0_end(used()); 6467 } 6468 } 6469 6470 CMSAdaptiveSizePolicy* ConcurrentMarkSweepGeneration::size_policy() { 6471 GenCollectedHeap* gch = GenCollectedHeap::heap(); 6472 assert(gch->kind() == CollectedHeap::GenCollectedHeap, 6473 "Wrong type of heap"); 6474 CMSAdaptiveSizePolicy* sp = (CMSAdaptiveSizePolicy*) 6475 gch->gen_policy()->size_policy(); 6476 assert(sp->is_gc_cms_adaptive_size_policy(), 6477 "Wrong type of size policy"); 6478 return sp; 6479 } 6480 6481 void ConcurrentMarkSweepGeneration::rotate_debug_collection_type() { 6482 if (PrintGCDetails && Verbose) { 6483 gclog_or_tty->print("Rotate from %d ", _debug_collection_type); 6484 } 6485 _debug_collection_type = (CollectionTypes) (_debug_collection_type + 1); 6486 _debug_collection_type = 6487 (CollectionTypes) (_debug_collection_type % Unknown_collection_type); 6488 if (PrintGCDetails && Verbose) { 6489 gclog_or_tty->print_cr("to %d ", _debug_collection_type); 6490 } 6491 } 6492 6493 void CMSCollector::sweepWork(ConcurrentMarkSweepGeneration* gen, 6494 bool asynch) { 6495 // We iterate over the space(s) underlying this generation, 6496 // checking the mark bit map to see if the bits corresponding 6497 // to specific blocks are marked or not. Blocks that are 6498 // marked are live and are not swept up. All remaining blocks 6499 // are swept up, with coalescing on-the-fly as we sweep up 6500 // contiguous free and/or garbage blocks: 6501 // We need to ensure that the sweeper synchronizes with allocators 6502 // and stop-the-world collectors. In particular, the following 6503 // locks are used: 6504 // . CMS token: if this is held, a stop the world collection cannot occur 6505 // . freelistLock: if this is held no allocation can occur from this 6506 // generation by another thread 6507 // . bitMapLock: if this is held, no other thread can access or update 6508 // 6509 6510 // Note that we need to hold the freelistLock if we use 6511 // block iterate below; else the iterator might go awry if 6512 // a mutator (or promotion) causes block contents to change 6513 // (for instance if the allocator divvies up a block). 6514 // If we hold the free list lock, for all practical purposes 6515 // young generation GC's can't occur (they'll usually need to 6516 // promote), so we might as well prevent all young generation 6517 // GC's while we do a sweeping step. For the same reason, we might 6518 // as well take the bit map lock for the entire duration 6519 6520 // check that we hold the requisite locks 6521 assert(have_cms_token(), "Should hold cms token"); 6522 assert( (asynch && ConcurrentMarkSweepThread::cms_thread_has_cms_token()) 6523 || (!asynch && ConcurrentMarkSweepThread::vm_thread_has_cms_token()), 6524 "Should possess CMS token to sweep"); 6525 assert_lock_strong(gen->freelistLock()); 6526 assert_lock_strong(bitMapLock()); 6527 6528 assert(!_inter_sweep_timer.is_active(), "Was switched off in an outer context"); 6529 assert(_intra_sweep_timer.is_active(), "Was switched on in an outer context"); 6530 gen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()), 6531 _inter_sweep_estimate.padded_average(), 6532 _intra_sweep_estimate.padded_average()); 6533 gen->setNearLargestChunk(); 6534 6535 { 6536 SweepClosure sweepClosure(this, gen, &_markBitMap, 6537 CMSYield && asynch); 6538 gen->cmsSpace()->blk_iterate_careful(&sweepClosure); 6539 // We need to free-up/coalesce garbage/blocks from a 6540 // co-terminal free run. This is done in the SweepClosure 6541 // destructor; so, do not remove this scope, else the 6542 // end-of-sweep-census below will be off by a little bit. 6543 } 6544 gen->cmsSpace()->sweep_completed(); 6545 gen->cmsSpace()->endSweepFLCensus(sweep_count()); 6546 if (should_unload_classes()) { // unloaded classes this cycle, 6547 _concurrent_cycles_since_last_unload = 0; // ... reset count 6548 } else { // did not unload classes, 6549 _concurrent_cycles_since_last_unload++; // ... increment count 6550 } 6551 } 6552 6553 // Reset CMS data structures (for now just the marking bit map) 6554 // preparatory for the next cycle. 6555 void CMSCollector::reset(bool asynch) { 6556 GenCollectedHeap* gch = GenCollectedHeap::heap(); 6557 CMSAdaptiveSizePolicy* sp = size_policy(); 6558 AdaptiveSizePolicyOutput(sp, gch->total_collections()); 6559 if (asynch) { 6560 CMSTokenSyncWithLocks ts(true, bitMapLock()); 6561 6562 // If the state is not "Resetting", the foreground thread 6563 // has done a collection and the resetting. 6564 if (_collectorState != Resetting) { 6565 assert(_collectorState == Idling, "The state should only change" 6566 " because the foreground collector has finished the collection"); 6567 return; 6568 } 6569 6570 // Clear the mark bitmap (no grey objects to start with) 6571 // for the next cycle. 6572 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 6573 CMSPhaseAccounting cmspa(this, "reset", _gc_tracer_cm->gc_id(), !PrintGCDetails); 6574 6575 HeapWord* curAddr = _markBitMap.startWord(); 6576 while (curAddr < _markBitMap.endWord()) { 6577 size_t remaining = pointer_delta(_markBitMap.endWord(), curAddr); 6578 MemRegion chunk(curAddr, MIN2(CMSBitMapYieldQuantum, remaining)); 6579 _markBitMap.clear_large_range(chunk); 6580 if (ConcurrentMarkSweepThread::should_yield() && 6581 !foregroundGCIsActive() && 6582 CMSYield) { 6583 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6584 "CMS thread should hold CMS token"); 6585 assert_lock_strong(bitMapLock()); 6586 bitMapLock()->unlock(); 6587 ConcurrentMarkSweepThread::desynchronize(true); 6588 ConcurrentMarkSweepThread::acknowledge_yield_request(); 6589 stopTimer(); 6590 if (PrintCMSStatistics != 0) { 6591 incrementYields(); 6592 } 6593 icms_wait(); 6594 6595 // See the comment in coordinator_yield() 6596 for (unsigned i = 0; i < CMSYieldSleepCount && 6597 ConcurrentMarkSweepThread::should_yield() && 6598 !CMSCollector::foregroundGCIsActive(); ++i) { 6599 os::sleep(Thread::current(), 1, false); 6600 ConcurrentMarkSweepThread::acknowledge_yield_request(); 6601 } 6602 6603 ConcurrentMarkSweepThread::synchronize(true); 6604 bitMapLock()->lock_without_safepoint_check(); 6605 startTimer(); 6606 } 6607 curAddr = chunk.end(); 6608 } 6609 // A successful mostly concurrent collection has been done. 6610 // Because only the full (i.e., concurrent mode failure) collections 6611 // are being measured for gc overhead limits, clean the "near" flag 6612 // and count. 6613 sp->reset_gc_overhead_limit_count(); 6614 _collectorState = Idling; 6615 } else { 6616 // already have the lock 6617 assert(_collectorState == Resetting, "just checking"); 6618 assert_lock_strong(bitMapLock()); 6619 _markBitMap.clear_all(); 6620 _collectorState = Idling; 6621 } 6622 6623 // Stop incremental mode after a cycle completes, so that any future cycles 6624 // are triggered by allocation. 6625 stop_icms(); 6626 6627 NOT_PRODUCT( 6628 if (RotateCMSCollectionTypes) { 6629 _cmsGen->rotate_debug_collection_type(); 6630 } 6631 ) 6632 6633 register_gc_end(); 6634 } 6635 6636 void CMSCollector::do_CMS_operation(CMS_op_type op, GCCause::Cause gc_cause) { 6637 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); 6638 GCTraceTime t(GCCauseString("GC", gc_cause), PrintGC, !PrintGCDetails, NULL, _gc_tracer_cm->gc_id()); 6639 TraceCollectorStats tcs(counters()); 6640 6641 switch (op) { 6642 case CMS_op_checkpointRootsInitial: { 6643 SvcGCMarker sgcm(SvcGCMarker::OTHER); 6644 checkpointRootsInitial(true); // asynch 6645 if (PrintGC) { 6646 _cmsGen->printOccupancy("initial-mark"); 6647 } 6648 break; 6649 } 6650 case CMS_op_checkpointRootsFinal: { 6651 SvcGCMarker sgcm(SvcGCMarker::OTHER); 6652 checkpointRootsFinal(true, // asynch 6653 false, // !clear_all_soft_refs 6654 false); // !init_mark_was_synchronous 6655 if (PrintGC) { 6656 _cmsGen->printOccupancy("remark"); 6657 } 6658 break; 6659 } 6660 default: 6661 fatal("No such CMS_op"); 6662 } 6663 } 6664 6665 #ifndef PRODUCT 6666 size_t const CMSCollector::skip_header_HeapWords() { 6667 return FreeChunk::header_size(); 6668 } 6669 6670 // Try and collect here conditions that should hold when 6671 // CMS thread is exiting. The idea is that the foreground GC 6672 // thread should not be blocked if it wants to terminate 6673 // the CMS thread and yet continue to run the VM for a while 6674 // after that. 6675 void CMSCollector::verify_ok_to_terminate() const { 6676 assert(Thread::current()->is_ConcurrentGC_thread(), 6677 "should be called by CMS thread"); 6678 assert(!_foregroundGCShouldWait, "should be false"); 6679 // We could check here that all the various low-level locks 6680 // are not held by the CMS thread, but that is overkill; see 6681 // also CMSThread::verify_ok_to_terminate() where the CGC_lock 6682 // is checked. 6683 } 6684 #endif 6685 6686 size_t CMSCollector::block_size_using_printezis_bits(HeapWord* addr) const { 6687 assert(_markBitMap.isMarked(addr) && _markBitMap.isMarked(addr + 1), 6688 "missing Printezis mark?"); 6689 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 6690 size_t size = pointer_delta(nextOneAddr + 1, addr); 6691 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 6692 "alignment problem"); 6693 assert(size >= 3, "Necessary for Printezis marks to work"); 6694 return size; 6695 } 6696 6697 // A variant of the above (block_size_using_printezis_bits()) except 6698 // that we return 0 if the P-bits are not yet set. 6699 size_t CMSCollector::block_size_if_printezis_bits(HeapWord* addr) const { 6700 if (_markBitMap.isMarked(addr + 1)) { 6701 assert(_markBitMap.isMarked(addr), "P-bit can be set only for marked objects"); 6702 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 6703 size_t size = pointer_delta(nextOneAddr + 1, addr); 6704 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 6705 "alignment problem"); 6706 assert(size >= 3, "Necessary for Printezis marks to work"); 6707 return size; 6708 } 6709 return 0; 6710 } 6711 6712 HeapWord* CMSCollector::next_card_start_after_block(HeapWord* addr) const { 6713 size_t sz = 0; 6714 oop p = (oop)addr; 6715 if (p->klass_or_null() != NULL) { 6716 sz = CompactibleFreeListSpace::adjustObjectSize(p->size()); 6717 } else { 6718 sz = block_size_using_printezis_bits(addr); 6719 } 6720 assert(sz > 0, "size must be nonzero"); 6721 HeapWord* next_block = addr + sz; 6722 HeapWord* next_card = (HeapWord*)round_to((uintptr_t)next_block, 6723 CardTableModRefBS::card_size); 6724 assert(round_down((uintptr_t)addr, CardTableModRefBS::card_size) < 6725 round_down((uintptr_t)next_card, CardTableModRefBS::card_size), 6726 "must be different cards"); 6727 return next_card; 6728 } 6729 6730 6731 // CMS Bit Map Wrapper ///////////////////////////////////////// 6732 6733 // Construct a CMS bit map infrastructure, but don't create the 6734 // bit vector itself. That is done by a separate call CMSBitMap::allocate() 6735 // further below. 6736 CMSBitMap::CMSBitMap(int shifter, int mutex_rank, const char* mutex_name): 6737 _bm(), 6738 _shifter(shifter), 6739 _lock(mutex_rank >= 0 ? new Mutex(mutex_rank, mutex_name, true) : NULL) 6740 { 6741 _bmStartWord = 0; 6742 _bmWordSize = 0; 6743 } 6744 6745 bool CMSBitMap::allocate(MemRegion mr) { 6746 _bmStartWord = mr.start(); 6747 _bmWordSize = mr.word_size(); 6748 ReservedSpace brs(ReservedSpace::allocation_align_size_up( 6749 (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1)); 6750 if (!brs.is_reserved()) { 6751 warning("CMS bit map allocation failure"); 6752 return false; 6753 } 6754 // For now we'll just commit all of the bit map up fromt. 6755 // Later on we'll try to be more parsimonious with swap. 6756 if (!_virtual_space.initialize(brs, brs.size())) { 6757 warning("CMS bit map backing store failure"); 6758 return false; 6759 } 6760 assert(_virtual_space.committed_size() == brs.size(), 6761 "didn't reserve backing store for all of CMS bit map?"); 6762 _bm.set_map((BitMap::bm_word_t*)_virtual_space.low()); 6763 assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >= 6764 _bmWordSize, "inconsistency in bit map sizing"); 6765 _bm.set_size(_bmWordSize >> _shifter); 6766 6767 // bm.clear(); // can we rely on getting zero'd memory? verify below 6768 assert(isAllClear(), 6769 "Expected zero'd memory from ReservedSpace constructor"); 6770 assert(_bm.size() == heapWordDiffToOffsetDiff(sizeInWords()), 6771 "consistency check"); 6772 return true; 6773 } 6774 6775 void CMSBitMap::dirty_range_iterate_clear(MemRegion mr, MemRegionClosure* cl) { 6776 HeapWord *next_addr, *end_addr, *last_addr; 6777 assert_locked(); 6778 assert(covers(mr), "out-of-range error"); 6779 // XXX assert that start and end are appropriately aligned 6780 for (next_addr = mr.start(), end_addr = mr.end(); 6781 next_addr < end_addr; next_addr = last_addr) { 6782 MemRegion dirty_region = getAndClearMarkedRegion(next_addr, end_addr); 6783 last_addr = dirty_region.end(); 6784 if (!dirty_region.is_empty()) { 6785 cl->do_MemRegion(dirty_region); 6786 } else { 6787 assert(last_addr == end_addr, "program logic"); 6788 return; 6789 } 6790 } 6791 } 6792 6793 void CMSBitMap::print_on_error(outputStream* st, const char* prefix) const { 6794 _bm.print_on_error(st, prefix); 6795 } 6796 6797 #ifndef PRODUCT 6798 void CMSBitMap::assert_locked() const { 6799 CMSLockVerifier::assert_locked(lock()); 6800 } 6801 6802 bool CMSBitMap::covers(MemRegion mr) const { 6803 // assert(_bm.map() == _virtual_space.low(), "map inconsistency"); 6804 assert((size_t)_bm.size() == (_bmWordSize >> _shifter), 6805 "size inconsistency"); 6806 return (mr.start() >= _bmStartWord) && 6807 (mr.end() <= endWord()); 6808 } 6809 6810 bool CMSBitMap::covers(HeapWord* start, size_t size) const { 6811 return (start >= _bmStartWord && (start + size) <= endWord()); 6812 } 6813 6814 void CMSBitMap::verifyNoOneBitsInRange(HeapWord* left, HeapWord* right) { 6815 // verify that there are no 1 bits in the interval [left, right) 6816 FalseBitMapClosure falseBitMapClosure; 6817 iterate(&falseBitMapClosure, left, right); 6818 } 6819 6820 void CMSBitMap::region_invariant(MemRegion mr) 6821 { 6822 assert_locked(); 6823 // mr = mr.intersection(MemRegion(_bmStartWord, _bmWordSize)); 6824 assert(!mr.is_empty(), "unexpected empty region"); 6825 assert(covers(mr), "mr should be covered by bit map"); 6826 // convert address range into offset range 6827 size_t start_ofs = heapWordToOffset(mr.start()); 6828 // Make sure that end() is appropriately aligned 6829 assert(mr.end() == (HeapWord*)round_to((intptr_t)mr.end(), 6830 (1 << (_shifter+LogHeapWordSize))), 6831 "Misaligned mr.end()"); 6832 size_t end_ofs = heapWordToOffset(mr.end()); 6833 assert(end_ofs > start_ofs, "Should mark at least one bit"); 6834 } 6835 6836 #endif 6837 6838 bool CMSMarkStack::allocate(size_t size) { 6839 // allocate a stack of the requisite depth 6840 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 6841 size * sizeof(oop))); 6842 if (!rs.is_reserved()) { 6843 warning("CMSMarkStack allocation failure"); 6844 return false; 6845 } 6846 if (!_virtual_space.initialize(rs, rs.size())) { 6847 warning("CMSMarkStack backing store failure"); 6848 return false; 6849 } 6850 assert(_virtual_space.committed_size() == rs.size(), 6851 "didn't reserve backing store for all of CMS stack?"); 6852 _base = (oop*)(_virtual_space.low()); 6853 _index = 0; 6854 _capacity = size; 6855 NOT_PRODUCT(_max_depth = 0); 6856 return true; 6857 } 6858 6859 // XXX FIX ME !!! In the MT case we come in here holding a 6860 // leaf lock. For printing we need to take a further lock 6861 // which has lower rank. We need to recallibrate the two 6862 // lock-ranks involved in order to be able to rpint the 6863 // messages below. (Or defer the printing to the caller. 6864 // For now we take the expedient path of just disabling the 6865 // messages for the problematic case.) 6866 void CMSMarkStack::expand() { 6867 assert(_capacity <= MarkStackSizeMax, "stack bigger than permitted"); 6868 if (_capacity == MarkStackSizeMax) { 6869 if (_hit_limit++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) { 6870 // We print a warning message only once per CMS cycle. 6871 gclog_or_tty->print_cr(" (benign) Hit CMSMarkStack max size limit"); 6872 } 6873 return; 6874 } 6875 // Double capacity if possible 6876 size_t new_capacity = MIN2(_capacity*2, MarkStackSizeMax); 6877 // Do not give up existing stack until we have managed to 6878 // get the double capacity that we desired. 6879 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 6880 new_capacity * sizeof(oop))); 6881 if (rs.is_reserved()) { 6882 // Release the backing store associated with old stack 6883 _virtual_space.release(); 6884 // Reinitialize virtual space for new stack 6885 if (!_virtual_space.initialize(rs, rs.size())) { 6886 fatal("Not enough swap for expanded marking stack"); 6887 } 6888 _base = (oop*)(_virtual_space.low()); 6889 _index = 0; 6890 _capacity = new_capacity; 6891 } else if (_failed_double++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) { 6892 // Failed to double capacity, continue; 6893 // we print a detail message only once per CMS cycle. 6894 gclog_or_tty->print(" (benign) Failed to expand marking stack from "SIZE_FORMAT"K to " 6895 SIZE_FORMAT"K", 6896 _capacity / K, new_capacity / K); 6897 } 6898 } 6899 6900 6901 // Closures 6902 // XXX: there seems to be a lot of code duplication here; 6903 // should refactor and consolidate common code. 6904 6905 // This closure is used to mark refs into the CMS generation in 6906 // the CMS bit map. Called at the first checkpoint. This closure 6907 // assumes that we do not need to re-mark dirty cards; if the CMS 6908 // generation on which this is used is not an oldest 6909 // generation then this will lose younger_gen cards! 6910 6911 MarkRefsIntoClosure::MarkRefsIntoClosure( 6912 MemRegion span, CMSBitMap* bitMap): 6913 _span(span), 6914 _bitMap(bitMap) 6915 { 6916 assert(_ref_processor == NULL, "deliberately left NULL"); 6917 assert(_bitMap->covers(_span), "_bitMap/_span mismatch"); 6918 } 6919 6920 void MarkRefsIntoClosure::do_oop(oop obj) { 6921 // if p points into _span, then mark corresponding bit in _markBitMap 6922 assert(obj->is_oop(), "expected an oop"); 6923 HeapWord* addr = (HeapWord*)obj; 6924 if (_span.contains(addr)) { 6925 // this should be made more efficient 6926 _bitMap->mark(addr); 6927 } 6928 } 6929 6930 void MarkRefsIntoClosure::do_oop(oop* p) { MarkRefsIntoClosure::do_oop_work(p); } 6931 void MarkRefsIntoClosure::do_oop(narrowOop* p) { MarkRefsIntoClosure::do_oop_work(p); } 6932 6933 Par_MarkRefsIntoClosure::Par_MarkRefsIntoClosure( 6934 MemRegion span, CMSBitMap* bitMap): 6935 _span(span), 6936 _bitMap(bitMap) 6937 { 6938 assert(_ref_processor == NULL, "deliberately left NULL"); 6939 assert(_bitMap->covers(_span), "_bitMap/_span mismatch"); 6940 } 6941 6942 void Par_MarkRefsIntoClosure::do_oop(oop obj) { 6943 // if p points into _span, then mark corresponding bit in _markBitMap 6944 assert(obj->is_oop(), "expected an oop"); 6945 HeapWord* addr = (HeapWord*)obj; 6946 if (_span.contains(addr)) { 6947 // this should be made more efficient 6948 _bitMap->par_mark(addr); 6949 } 6950 } 6951 6952 void Par_MarkRefsIntoClosure::do_oop(oop* p) { Par_MarkRefsIntoClosure::do_oop_work(p); } 6953 void Par_MarkRefsIntoClosure::do_oop(narrowOop* p) { Par_MarkRefsIntoClosure::do_oop_work(p); } 6954 6955 // A variant of the above, used for CMS marking verification. 6956 MarkRefsIntoVerifyClosure::MarkRefsIntoVerifyClosure( 6957 MemRegion span, CMSBitMap* verification_bm, CMSBitMap* cms_bm): 6958 _span(span), 6959 _verification_bm(verification_bm), 6960 _cms_bm(cms_bm) 6961 { 6962 assert(_ref_processor == NULL, "deliberately left NULL"); 6963 assert(_verification_bm->covers(_span), "_verification_bm/_span mismatch"); 6964 } 6965 6966 void MarkRefsIntoVerifyClosure::do_oop(oop obj) { 6967 // if p points into _span, then mark corresponding bit in _markBitMap 6968 assert(obj->is_oop(), "expected an oop"); 6969 HeapWord* addr = (HeapWord*)obj; 6970 if (_span.contains(addr)) { 6971 _verification_bm->mark(addr); 6972 if (!_cms_bm->isMarked(addr)) { 6973 oop(addr)->print(); 6974 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)", addr); 6975 fatal("... aborting"); 6976 } 6977 } 6978 } 6979 6980 void MarkRefsIntoVerifyClosure::do_oop(oop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); } 6981 void MarkRefsIntoVerifyClosure::do_oop(narrowOop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); } 6982 6983 ////////////////////////////////////////////////// 6984 // MarkRefsIntoAndScanClosure 6985 ////////////////////////////////////////////////// 6986 6987 MarkRefsIntoAndScanClosure::MarkRefsIntoAndScanClosure(MemRegion span, 6988 ReferenceProcessor* rp, 6989 CMSBitMap* bit_map, 6990 CMSBitMap* mod_union_table, 6991 CMSMarkStack* mark_stack, 6992 CMSCollector* collector, 6993 bool should_yield, 6994 bool concurrent_precleaning): 6995 _collector(collector), 6996 _span(span), 6997 _bit_map(bit_map), 6998 _mark_stack(mark_stack), 6999 _pushAndMarkClosure(collector, span, rp, bit_map, mod_union_table, 7000 mark_stack, concurrent_precleaning), 7001 _yield(should_yield), 7002 _concurrent_precleaning(concurrent_precleaning), 7003 _freelistLock(NULL) 7004 { 7005 _ref_processor = rp; 7006 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 7007 } 7008 7009 // This closure is used to mark refs into the CMS generation at the 7010 // second (final) checkpoint, and to scan and transitively follow 7011 // the unmarked oops. It is also used during the concurrent precleaning 7012 // phase while scanning objects on dirty cards in the CMS generation. 7013 // The marks are made in the marking bit map and the marking stack is 7014 // used for keeping the (newly) grey objects during the scan. 7015 // The parallel version (Par_...) appears further below. 7016 void MarkRefsIntoAndScanClosure::do_oop(oop obj) { 7017 if (obj != NULL) { 7018 assert(obj->is_oop(), "expected an oop"); 7019 HeapWord* addr = (HeapWord*)obj; 7020 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 7021 assert(_collector->overflow_list_is_empty(), 7022 "overflow list should be empty"); 7023 if (_span.contains(addr) && 7024 !_bit_map->isMarked(addr)) { 7025 // mark bit map (object is now grey) 7026 _bit_map->mark(addr); 7027 // push on marking stack (stack should be empty), and drain the 7028 // stack by applying this closure to the oops in the oops popped 7029 // from the stack (i.e. blacken the grey objects) 7030 bool res = _mark_stack->push(obj); 7031 assert(res, "Should have space to push on empty stack"); 7032 do { 7033 oop new_oop = _mark_stack->pop(); 7034 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 7035 assert(_bit_map->isMarked((HeapWord*)new_oop), 7036 "only grey objects on this stack"); 7037 // iterate over the oops in this oop, marking and pushing 7038 // the ones in CMS heap (i.e. in _span). 7039 new_oop->oop_iterate(&_pushAndMarkClosure); 7040 // check if it's time to yield 7041 do_yield_check(); 7042 } while (!_mark_stack->isEmpty() || 7043 (!_concurrent_precleaning && take_from_overflow_list())); 7044 // if marking stack is empty, and we are not doing this 7045 // during precleaning, then check the overflow list 7046 } 7047 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 7048 assert(_collector->overflow_list_is_empty(), 7049 "overflow list was drained above"); 7050 // We could restore evacuated mark words, if any, used for 7051 // overflow list links here because the overflow list is 7052 // provably empty here. That would reduce the maximum 7053 // size requirements for preserved_{oop,mark}_stack. 7054 // But we'll just postpone it until we are all done 7055 // so we can just stream through. 7056 if (!_concurrent_precleaning && CMSOverflowEarlyRestoration) { 7057 _collector->restore_preserved_marks_if_any(); 7058 assert(_collector->no_preserved_marks(), "No preserved marks"); 7059 } 7060 assert(!CMSOverflowEarlyRestoration || _collector->no_preserved_marks(), 7061 "All preserved marks should have been restored above"); 7062 } 7063 } 7064 7065 void MarkRefsIntoAndScanClosure::do_oop(oop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); } 7066 void MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); } 7067 7068 void MarkRefsIntoAndScanClosure::do_yield_work() { 7069 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 7070 "CMS thread should hold CMS token"); 7071 assert_lock_strong(_freelistLock); 7072 assert_lock_strong(_bit_map->lock()); 7073 // relinquish the free_list_lock and bitMaplock() 7074 _bit_map->lock()->unlock(); 7075 _freelistLock->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; 7087 i < CMSYieldSleepCount && 7088 ConcurrentMarkSweepThread::should_yield() && 7089 !CMSCollector::foregroundGCIsActive(); 7090 ++i) { 7091 os::sleep(Thread::current(), 1, false); 7092 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7093 } 7094 7095 ConcurrentMarkSweepThread::synchronize(true); 7096 _freelistLock->lock_without_safepoint_check(); 7097 _bit_map->lock()->lock_without_safepoint_check(); 7098 _collector->startTimer(); 7099 } 7100 7101 /////////////////////////////////////////////////////////// 7102 // Par_MarkRefsIntoAndScanClosure: a parallel version of 7103 // MarkRefsIntoAndScanClosure 7104 /////////////////////////////////////////////////////////// 7105 Par_MarkRefsIntoAndScanClosure::Par_MarkRefsIntoAndScanClosure( 7106 CMSCollector* collector, MemRegion span, ReferenceProcessor* rp, 7107 CMSBitMap* bit_map, OopTaskQueue* work_queue): 7108 _span(span), 7109 _bit_map(bit_map), 7110 _work_queue(work_queue), 7111 _low_water_mark(MIN2((uint)(work_queue->max_elems()/4), 7112 (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads))), 7113 _par_pushAndMarkClosure(collector, span, rp, bit_map, work_queue) 7114 { 7115 _ref_processor = rp; 7116 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 7117 } 7118 7119 // This closure is used to mark refs into the CMS generation at the 7120 // second (final) checkpoint, and to scan and transitively follow 7121 // the unmarked oops. The marks are made in the marking bit map and 7122 // the work_queue is used for keeping the (newly) grey objects during 7123 // the scan phase whence they are also available for stealing by parallel 7124 // threads. Since the marking bit map is shared, updates are 7125 // synchronized (via CAS). 7126 void Par_MarkRefsIntoAndScanClosure::do_oop(oop obj) { 7127 if (obj != NULL) { 7128 // Ignore mark word because this could be an already marked oop 7129 // that may be chained at the end of the overflow list. 7130 assert(obj->is_oop(true), "expected an oop"); 7131 HeapWord* addr = (HeapWord*)obj; 7132 if (_span.contains(addr) && 7133 !_bit_map->isMarked(addr)) { 7134 // mark bit map (object will become grey): 7135 // It is possible for several threads to be 7136 // trying to "claim" this object concurrently; 7137 // the unique thread that succeeds in marking the 7138 // object first will do the subsequent push on 7139 // to the work queue (or overflow list). 7140 if (_bit_map->par_mark(addr)) { 7141 // push on work_queue (which may not be empty), and trim the 7142 // queue to an appropriate length by applying this closure to 7143 // the oops in the oops popped from the stack (i.e. blacken the 7144 // grey objects) 7145 bool res = _work_queue->push(obj); 7146 assert(res, "Low water mark should be less than capacity?"); 7147 trim_queue(_low_water_mark); 7148 } // Else, another thread claimed the object 7149 } 7150 } 7151 } 7152 7153 void Par_MarkRefsIntoAndScanClosure::do_oop(oop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); } 7154 void Par_MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); } 7155 7156 // This closure is used to rescan the marked objects on the dirty cards 7157 // in the mod union table and the card table proper. 7158 size_t ScanMarkedObjectsAgainCarefullyClosure::do_object_careful_m( 7159 oop p, MemRegion mr) { 7160 7161 size_t size = 0; 7162 HeapWord* addr = (HeapWord*)p; 7163 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 7164 assert(_span.contains(addr), "we are scanning the CMS generation"); 7165 // check if it's time to yield 7166 if (do_yield_check()) { 7167 // We yielded for some foreground stop-world work, 7168 // and we have been asked to abort this ongoing preclean cycle. 7169 return 0; 7170 } 7171 if (_bitMap->isMarked(addr)) { 7172 // it's marked; is it potentially uninitialized? 7173 if (p->klass_or_null() != NULL) { 7174 // an initialized object; ignore mark word in verification below 7175 // since we are running concurrent with mutators 7176 assert(p->is_oop(true), "should be an oop"); 7177 if (p->is_objArray()) { 7178 // objArrays are precisely marked; restrict scanning 7179 // to dirty cards only. 7180 size = CompactibleFreeListSpace::adjustObjectSize( 7181 p->oop_iterate(_scanningClosure, mr)); 7182 } else { 7183 // A non-array may have been imprecisely marked; we need 7184 // to scan object in its entirety. 7185 size = CompactibleFreeListSpace::adjustObjectSize( 7186 p->oop_iterate(_scanningClosure)); 7187 } 7188 #ifdef ASSERT 7189 size_t direct_size = 7190 CompactibleFreeListSpace::adjustObjectSize(p->size()); 7191 assert(size == direct_size, "Inconsistency in size"); 7192 assert(size >= 3, "Necessary for Printezis marks to work"); 7193 if (!_bitMap->isMarked(addr+1)) { 7194 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size); 7195 } else { 7196 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size-1); 7197 assert(_bitMap->isMarked(addr+size-1), 7198 "inconsistent Printezis mark"); 7199 } 7200 #endif // ASSERT 7201 } else { 7202 // an unitialized object 7203 assert(_bitMap->isMarked(addr+1), "missing Printezis mark?"); 7204 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 7205 size = pointer_delta(nextOneAddr + 1, addr); 7206 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 7207 "alignment problem"); 7208 // Note that pre-cleaning needn't redirty the card. OopDesc::set_klass() 7209 // will dirty the card when the klass pointer is installed in the 7210 // object (signalling the completion of initialization). 7211 } 7212 } else { 7213 // Either a not yet marked object or an uninitialized object 7214 if (p->klass_or_null() == NULL) { 7215 // An uninitialized object, skip to the next card, since 7216 // we may not be able to read its P-bits yet. 7217 assert(size == 0, "Initial value"); 7218 } else { 7219 // An object not (yet) reached by marking: we merely need to 7220 // compute its size so as to go look at the next block. 7221 assert(p->is_oop(true), "should be an oop"); 7222 size = CompactibleFreeListSpace::adjustObjectSize(p->size()); 7223 } 7224 } 7225 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 7226 return size; 7227 } 7228 7229 void ScanMarkedObjectsAgainCarefullyClosure::do_yield_work() { 7230 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 7231 "CMS thread should hold CMS token"); 7232 assert_lock_strong(_freelistLock); 7233 assert_lock_strong(_bitMap->lock()); 7234 // relinquish the free_list_lock and bitMaplock() 7235 _bitMap->lock()->unlock(); 7236 _freelistLock->unlock(); 7237 ConcurrentMarkSweepThread::desynchronize(true); 7238 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7239 _collector->stopTimer(); 7240 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 7241 if (PrintCMSStatistics != 0) { 7242 _collector->incrementYields(); 7243 } 7244 _collector->icms_wait(); 7245 7246 // See the comment in coordinator_yield() 7247 for (unsigned i = 0; i < CMSYieldSleepCount && 7248 ConcurrentMarkSweepThread::should_yield() && 7249 !CMSCollector::foregroundGCIsActive(); ++i) { 7250 os::sleep(Thread::current(), 1, false); 7251 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7252 } 7253 7254 ConcurrentMarkSweepThread::synchronize(true); 7255 _freelistLock->lock_without_safepoint_check(); 7256 _bitMap->lock()->lock_without_safepoint_check(); 7257 _collector->startTimer(); 7258 } 7259 7260 7261 ////////////////////////////////////////////////////////////////// 7262 // SurvivorSpacePrecleanClosure 7263 ////////////////////////////////////////////////////////////////// 7264 // This (single-threaded) closure is used to preclean the oops in 7265 // the survivor spaces. 7266 size_t SurvivorSpacePrecleanClosure::do_object_careful(oop p) { 7267 7268 HeapWord* addr = (HeapWord*)p; 7269 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 7270 assert(!_span.contains(addr), "we are scanning the survivor spaces"); 7271 assert(p->klass_or_null() != NULL, "object should be initializd"); 7272 // an initialized object; ignore mark word in verification below 7273 // since we are running concurrent with mutators 7274 assert(p->is_oop(true), "should be an oop"); 7275 // Note that we do not yield while we iterate over 7276 // the interior oops of p, pushing the relevant ones 7277 // on our marking stack. 7278 size_t size = p->oop_iterate(_scanning_closure); 7279 do_yield_check(); 7280 // Observe that below, we do not abandon the preclean 7281 // phase as soon as we should; rather we empty the 7282 // marking stack before returning. This is to satisfy 7283 // some existing assertions. In general, it may be a 7284 // good idea to abort immediately and complete the marking 7285 // from the grey objects at a later time. 7286 while (!_mark_stack->isEmpty()) { 7287 oop new_oop = _mark_stack->pop(); 7288 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 7289 assert(_bit_map->isMarked((HeapWord*)new_oop), 7290 "only grey objects on this stack"); 7291 // iterate over the oops in this oop, marking and pushing 7292 // the ones in CMS heap (i.e. in _span). 7293 new_oop->oop_iterate(_scanning_closure); 7294 // check if it's time to yield 7295 do_yield_check(); 7296 } 7297 unsigned int after_count = 7298 GenCollectedHeap::heap()->total_collections(); 7299 bool abort = (_before_count != after_count) || 7300 _collector->should_abort_preclean(); 7301 return abort ? 0 : size; 7302 } 7303 7304 void SurvivorSpacePrecleanClosure::do_yield_work() { 7305 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 7306 "CMS thread should hold CMS token"); 7307 assert_lock_strong(_bit_map->lock()); 7308 // Relinquish the bit map lock 7309 _bit_map->lock()->unlock(); 7310 ConcurrentMarkSweepThread::desynchronize(true); 7311 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7312 _collector->stopTimer(); 7313 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 7314 if (PrintCMSStatistics != 0) { 7315 _collector->incrementYields(); 7316 } 7317 _collector->icms_wait(); 7318 7319 // See the comment in coordinator_yield() 7320 for (unsigned i = 0; i < CMSYieldSleepCount && 7321 ConcurrentMarkSweepThread::should_yield() && 7322 !CMSCollector::foregroundGCIsActive(); ++i) { 7323 os::sleep(Thread::current(), 1, false); 7324 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7325 } 7326 7327 ConcurrentMarkSweepThread::synchronize(true); 7328 _bit_map->lock()->lock_without_safepoint_check(); 7329 _collector->startTimer(); 7330 } 7331 7332 // This closure is used to rescan the marked objects on the dirty cards 7333 // in the mod union table and the card table proper. In the parallel 7334 // case, although the bitMap is shared, we do a single read so the 7335 // isMarked() query is "safe". 7336 bool ScanMarkedObjectsAgainClosure::do_object_bm(oop p, MemRegion mr) { 7337 // Ignore mark word because we are running concurrent with mutators 7338 assert(p->is_oop_or_null(true), "expected an oop or null"); 7339 HeapWord* addr = (HeapWord*)p; 7340 assert(_span.contains(addr), "we are scanning the CMS generation"); 7341 bool is_obj_array = false; 7342 #ifdef ASSERT 7343 if (!_parallel) { 7344 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 7345 assert(_collector->overflow_list_is_empty(), 7346 "overflow list should be empty"); 7347 7348 } 7349 #endif // ASSERT 7350 if (_bit_map->isMarked(addr)) { 7351 // Obj arrays are precisely marked, non-arrays are not; 7352 // so we scan objArrays precisely and non-arrays in their 7353 // entirety. 7354 if (p->is_objArray()) { 7355 is_obj_array = true; 7356 if (_parallel) { 7357 p->oop_iterate(_par_scan_closure, mr); 7358 } else { 7359 p->oop_iterate(_scan_closure, mr); 7360 } 7361 } else { 7362 if (_parallel) { 7363 p->oop_iterate(_par_scan_closure); 7364 } else { 7365 p->oop_iterate(_scan_closure); 7366 } 7367 } 7368 } 7369 #ifdef ASSERT 7370 if (!_parallel) { 7371 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 7372 assert(_collector->overflow_list_is_empty(), 7373 "overflow list should be empty"); 7374 7375 } 7376 #endif // ASSERT 7377 return is_obj_array; 7378 } 7379 7380 MarkFromRootsClosure::MarkFromRootsClosure(CMSCollector* collector, 7381 MemRegion span, 7382 CMSBitMap* bitMap, CMSMarkStack* markStack, 7383 bool should_yield, bool verifying): 7384 _collector(collector), 7385 _span(span), 7386 _bitMap(bitMap), 7387 _mut(&collector->_modUnionTable), 7388 _markStack(markStack), 7389 _yield(should_yield), 7390 _skipBits(0) 7391 { 7392 assert(_markStack->isEmpty(), "stack should be empty"); 7393 _finger = _bitMap->startWord(); 7394 _threshold = _finger; 7395 assert(_collector->_restart_addr == NULL, "Sanity check"); 7396 assert(_span.contains(_finger), "Out of bounds _finger?"); 7397 DEBUG_ONLY(_verifying = verifying;) 7398 } 7399 7400 void MarkFromRootsClosure::reset(HeapWord* addr) { 7401 assert(_markStack->isEmpty(), "would cause duplicates on stack"); 7402 assert(_span.contains(addr), "Out of bounds _finger?"); 7403 _finger = addr; 7404 _threshold = (HeapWord*)round_to( 7405 (intptr_t)_finger, CardTableModRefBS::card_size); 7406 } 7407 7408 // Should revisit to see if this should be restructured for 7409 // greater efficiency. 7410 bool MarkFromRootsClosure::do_bit(size_t offset) { 7411 if (_skipBits > 0) { 7412 _skipBits--; 7413 return true; 7414 } 7415 // convert offset into a HeapWord* 7416 HeapWord* addr = _bitMap->startWord() + offset; 7417 assert(_bitMap->endWord() && addr < _bitMap->endWord(), 7418 "address out of range"); 7419 assert(_bitMap->isMarked(addr), "tautology"); 7420 if (_bitMap->isMarked(addr+1)) { 7421 // this is an allocated but not yet initialized object 7422 assert(_skipBits == 0, "tautology"); 7423 _skipBits = 2; // skip next two marked bits ("Printezis-marks") 7424 oop p = oop(addr); 7425 if (p->klass_or_null() == NULL) { 7426 DEBUG_ONLY(if (!_verifying) {) 7427 // We re-dirty the cards on which this object lies and increase 7428 // the _threshold so that we'll come back to scan this object 7429 // during the preclean or remark phase. (CMSCleanOnEnter) 7430 if (CMSCleanOnEnter) { 7431 size_t sz = _collector->block_size_using_printezis_bits(addr); 7432 HeapWord* end_card_addr = (HeapWord*)round_to( 7433 (intptr_t)(addr+sz), CardTableModRefBS::card_size); 7434 MemRegion redirty_range = MemRegion(addr, end_card_addr); 7435 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 7436 // Bump _threshold to end_card_addr; note that 7437 // _threshold cannot possibly exceed end_card_addr, anyhow. 7438 // This prevents future clearing of the card as the scan proceeds 7439 // to the right. 7440 assert(_threshold <= end_card_addr, 7441 "Because we are just scanning into this object"); 7442 if (_threshold < end_card_addr) { 7443 _threshold = end_card_addr; 7444 } 7445 if (p->klass_or_null() != NULL) { 7446 // Redirty the range of cards... 7447 _mut->mark_range(redirty_range); 7448 } // ...else the setting of klass will dirty the card anyway. 7449 } 7450 DEBUG_ONLY(}) 7451 return true; 7452 } 7453 } 7454 scanOopsInOop(addr); 7455 return true; 7456 } 7457 7458 // We take a break if we've been at this for a while, 7459 // so as to avoid monopolizing the locks involved. 7460 void MarkFromRootsClosure::do_yield_work() { 7461 // First give up the locks, then yield, then re-lock 7462 // We should probably use a constructor/destructor idiom to 7463 // do this unlock/lock or modify the MutexUnlocker class to 7464 // serve our purpose. XXX 7465 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 7466 "CMS thread should hold CMS token"); 7467 assert_lock_strong(_bitMap->lock()); 7468 _bitMap->lock()->unlock(); 7469 ConcurrentMarkSweepThread::desynchronize(true); 7470 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7471 _collector->stopTimer(); 7472 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 7473 if (PrintCMSStatistics != 0) { 7474 _collector->incrementYields(); 7475 } 7476 _collector->icms_wait(); 7477 7478 // See the comment in coordinator_yield() 7479 for (unsigned i = 0; i < CMSYieldSleepCount && 7480 ConcurrentMarkSweepThread::should_yield() && 7481 !CMSCollector::foregroundGCIsActive(); ++i) { 7482 os::sleep(Thread::current(), 1, false); 7483 ConcurrentMarkSweepThread::acknowledge_yield_request(); 7484 } 7485 7486 ConcurrentMarkSweepThread::synchronize(true); 7487 _bitMap->lock()->lock_without_safepoint_check(); 7488 _collector->startTimer(); 7489 } 7490 7491 void MarkFromRootsClosure::scanOopsInOop(HeapWord* ptr) { 7492 assert(_bitMap->isMarked(ptr), "expected bit to be set"); 7493 assert(_markStack->isEmpty(), 7494 "should drain stack to limit stack usage"); 7495 // convert ptr to an oop preparatory to scanning 7496 oop obj = oop(ptr); 7497 // Ignore mark word in verification below, since we 7498 // may be running concurrent with mutators. 7499 assert(obj->is_oop(true), "should be an oop"); 7500 assert(_finger <= ptr, "_finger runneth ahead"); 7501 // advance the finger to right end of this object 7502 _finger = ptr + obj->size(); 7503 assert(_finger > ptr, "we just incremented it above"); 7504 // On large heaps, it may take us some time to get through 7505 // the marking phase (especially if running iCMS). During 7506 // this time it's possible that a lot of mutations have 7507 // accumulated in the card table and the mod union table -- 7508 // these mutation records are redundant until we have 7509 // actually traced into the corresponding card. 7510 // Here, we check whether advancing the finger would make 7511 // us cross into a new card, and if so clear corresponding 7512 // cards in the MUT (preclean them in the card-table in the 7513 // future). 7514 7515 DEBUG_ONLY(if (!_verifying) {) 7516 // The clean-on-enter optimization is disabled by default, 7517 // until we fix 6178663. 7518 if (CMSCleanOnEnter && (_finger > _threshold)) { 7519 // [_threshold, _finger) represents the interval 7520 // of cards to be cleared in MUT (or precleaned in card table). 7521 // The set of cards to be cleared is all those that overlap 7522 // with the interval [_threshold, _finger); note that 7523 // _threshold is always kept card-aligned but _finger isn't 7524 // always card-aligned. 7525 HeapWord* old_threshold = _threshold; 7526 assert(old_threshold == (HeapWord*)round_to( 7527 (intptr_t)old_threshold, CardTableModRefBS::card_size), 7528 "_threshold should always be card-aligned"); 7529 _threshold = (HeapWord*)round_to( 7530 (intptr_t)_finger, CardTableModRefBS::card_size); 7531 MemRegion mr(old_threshold, _threshold); 7532 assert(!mr.is_empty(), "Control point invariant"); 7533 assert(_span.contains(mr), "Should clear within span"); 7534 _mut->clear_range(mr); 7535 } 7536 DEBUG_ONLY(}) 7537 // Note: the finger doesn't advance while we drain 7538 // the stack below. 7539 PushOrMarkClosure pushOrMarkClosure(_collector, 7540 _span, _bitMap, _markStack, 7541 _finger, this); 7542 bool res = _markStack->push(obj); 7543 assert(res, "Empty non-zero size stack should have space for single push"); 7544 while (!_markStack->isEmpty()) { 7545 oop new_oop = _markStack->pop(); 7546 // Skip verifying header mark word below because we are 7547 // running concurrent with mutators. 7548 assert(new_oop->is_oop(true), "Oops! expected to pop an oop"); 7549 // now scan this oop's oops 7550 new_oop->oop_iterate(&pushOrMarkClosure); 7551 do_yield_check(); 7552 } 7553 assert(_markStack->isEmpty(), "tautology, emphasizing post-condition"); 7554 } 7555 7556 Par_MarkFromRootsClosure::Par_MarkFromRootsClosure(CMSConcMarkingTask* task, 7557 CMSCollector* collector, MemRegion span, 7558 CMSBitMap* bit_map, 7559 OopTaskQueue* work_queue, 7560 CMSMarkStack* overflow_stack, 7561 bool should_yield): 7562 _collector(collector), 7563 _whole_span(collector->_span), 7564 _span(span), 7565 _bit_map(bit_map), 7566 _mut(&collector->_modUnionTable), 7567 _work_queue(work_queue), 7568 _overflow_stack(overflow_stack), 7569 _yield(should_yield), 7570 _skip_bits(0), 7571 _task(task) 7572 { 7573 assert(_work_queue->size() == 0, "work_queue should be empty"); 7574 _finger = span.start(); 7575 _threshold = _finger; // XXX Defer clear-on-enter optimization for now 7576 assert(_span.contains(_finger), "Out of bounds _finger?"); 7577 } 7578 7579 // Should revisit to see if this should be restructured for 7580 // greater efficiency. 7581 bool Par_MarkFromRootsClosure::do_bit(size_t offset) { 7582 if (_skip_bits > 0) { 7583 _skip_bits--; 7584 return true; 7585 } 7586 // convert offset into a HeapWord* 7587 HeapWord* addr = _bit_map->startWord() + offset; 7588 assert(_bit_map->endWord() && addr < _bit_map->endWord(), 7589 "address out of range"); 7590 assert(_bit_map->isMarked(addr), "tautology"); 7591 if (_bit_map->isMarked(addr+1)) { 7592 // this is an allocated object that might not yet be initialized 7593 assert(_skip_bits == 0, "tautology"); 7594 _skip_bits = 2; // skip next two marked bits ("Printezis-marks") 7595 oop p = oop(addr); 7596 if (p->klass_or_null() == NULL) { 7597 // in the case of Clean-on-Enter optimization, redirty card 7598 // and avoid clearing card by increasing the threshold. 7599 return true; 7600 } 7601 } 7602 scan_oops_in_oop(addr); 7603 return true; 7604 } 7605 7606 void Par_MarkFromRootsClosure::scan_oops_in_oop(HeapWord* ptr) { 7607 assert(_bit_map->isMarked(ptr), "expected bit to be set"); 7608 // Should we assert that our work queue is empty or 7609 // below some drain limit? 7610 assert(_work_queue->size() == 0, 7611 "should drain stack to limit stack usage"); 7612 // convert ptr to an oop preparatory to scanning 7613 oop obj = oop(ptr); 7614 // Ignore mark word in verification below, since we 7615 // may be running concurrent with mutators. 7616 assert(obj->is_oop(true), "should be an oop"); 7617 assert(_finger <= ptr, "_finger runneth ahead"); 7618 // advance the finger to right end of this object 7619 _finger = ptr + obj->size(); 7620 assert(_finger > ptr, "we just incremented it above"); 7621 // On large heaps, it may take us some time to get through 7622 // the marking phase (especially if running iCMS). During 7623 // this time it's possible that a lot of mutations have 7624 // accumulated in the card table and the mod union table -- 7625 // these mutation records are redundant until we have 7626 // actually traced into the corresponding card. 7627 // Here, we check whether advancing the finger would make 7628 // us cross into a new card, and if so clear corresponding 7629 // cards in the MUT (preclean them in the card-table in the 7630 // future). 7631 7632 // The clean-on-enter optimization is disabled by default, 7633 // until we fix 6178663. 7634 if (CMSCleanOnEnter && (_finger > _threshold)) { 7635 // [_threshold, _finger) represents the interval 7636 // of cards to be cleared in MUT (or precleaned in card table). 7637 // The set of cards to be cleared is all those that overlap 7638 // with the interval [_threshold, _finger); note that 7639 // _threshold is always kept card-aligned but _finger isn't 7640 // always card-aligned. 7641 HeapWord* old_threshold = _threshold; 7642 assert(old_threshold == (HeapWord*)round_to( 7643 (intptr_t)old_threshold, CardTableModRefBS::card_size), 7644 "_threshold should always be card-aligned"); 7645 _threshold = (HeapWord*)round_to( 7646 (intptr_t)_finger, CardTableModRefBS::card_size); 7647 MemRegion mr(old_threshold, _threshold); 7648 assert(!mr.is_empty(), "Control point invariant"); 7649 assert(_span.contains(mr), "Should clear within span"); // _whole_span ?? 7650 _mut->clear_range(mr); 7651 } 7652 7653 // Note: the local finger doesn't advance while we drain 7654 // the stack below, but the global finger sure can and will. 7655 HeapWord** gfa = _task->global_finger_addr(); 7656 Par_PushOrMarkClosure pushOrMarkClosure(_collector, 7657 _span, _bit_map, 7658 _work_queue, 7659 _overflow_stack, 7660 _finger, 7661 gfa, this); 7662 bool res = _work_queue->push(obj); // overflow could occur here 7663 assert(res, "Will hold once we use workqueues"); 7664 while (true) { 7665 oop new_oop; 7666 if (!_work_queue->pop_local(new_oop)) { 7667 // We emptied our work_queue; check if there's stuff that can 7668 // be gotten from the overflow stack. 7669 if (CMSConcMarkingTask::get_work_from_overflow_stack( 7670 _overflow_stack, _work_queue)) { 7671 do_yield_check(); 7672 continue; 7673 } else { // done 7674 break; 7675 } 7676 } 7677 // Skip verifying header mark word below because we are 7678 // running concurrent with mutators. 7679 assert(new_oop->is_oop(true), "Oops! expected to pop an oop"); 7680 // now scan this oop's oops 7681 new_oop->oop_iterate(&pushOrMarkClosure); 7682 do_yield_check(); 7683 } 7684 assert(_work_queue->size() == 0, "tautology, emphasizing post-condition"); 7685 } 7686 7687 // Yield in response to a request from VM Thread or 7688 // from mutators. 7689 void Par_MarkFromRootsClosure::do_yield_work() { 7690 assert(_task != NULL, "sanity"); 7691 _task->yield(); 7692 } 7693 7694 // A variant of the above used for verifying CMS marking work. 7695 MarkFromRootsVerifyClosure::MarkFromRootsVerifyClosure(CMSCollector* collector, 7696 MemRegion span, 7697 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 7698 CMSMarkStack* mark_stack): 7699 _collector(collector), 7700 _span(span), 7701 _verification_bm(verification_bm), 7702 _cms_bm(cms_bm), 7703 _mark_stack(mark_stack), 7704 _pam_verify_closure(collector, span, verification_bm, cms_bm, 7705 mark_stack) 7706 { 7707 assert(_mark_stack->isEmpty(), "stack should be empty"); 7708 _finger = _verification_bm->startWord(); 7709 assert(_collector->_restart_addr == NULL, "Sanity check"); 7710 assert(_span.contains(_finger), "Out of bounds _finger?"); 7711 } 7712 7713 void MarkFromRootsVerifyClosure::reset(HeapWord* addr) { 7714 assert(_mark_stack->isEmpty(), "would cause duplicates on stack"); 7715 assert(_span.contains(addr), "Out of bounds _finger?"); 7716 _finger = addr; 7717 } 7718 7719 // Should revisit to see if this should be restructured for 7720 // greater efficiency. 7721 bool MarkFromRootsVerifyClosure::do_bit(size_t offset) { 7722 // convert offset into a HeapWord* 7723 HeapWord* addr = _verification_bm->startWord() + offset; 7724 assert(_verification_bm->endWord() && addr < _verification_bm->endWord(), 7725 "address out of range"); 7726 assert(_verification_bm->isMarked(addr), "tautology"); 7727 assert(_cms_bm->isMarked(addr), "tautology"); 7728 7729 assert(_mark_stack->isEmpty(), 7730 "should drain stack to limit stack usage"); 7731 // convert addr to an oop preparatory to scanning 7732 oop obj = oop(addr); 7733 assert(obj->is_oop(), "should be an oop"); 7734 assert(_finger <= addr, "_finger runneth ahead"); 7735 // advance the finger to right end of this object 7736 _finger = addr + obj->size(); 7737 assert(_finger > addr, "we just incremented it above"); 7738 // Note: the finger doesn't advance while we drain 7739 // the stack below. 7740 bool res = _mark_stack->push(obj); 7741 assert(res, "Empty non-zero size stack should have space for single push"); 7742 while (!_mark_stack->isEmpty()) { 7743 oop new_oop = _mark_stack->pop(); 7744 assert(new_oop->is_oop(), "Oops! expected to pop an oop"); 7745 // now scan this oop's oops 7746 new_oop->oop_iterate(&_pam_verify_closure); 7747 } 7748 assert(_mark_stack->isEmpty(), "tautology, emphasizing post-condition"); 7749 return true; 7750 } 7751 7752 PushAndMarkVerifyClosure::PushAndMarkVerifyClosure( 7753 CMSCollector* collector, MemRegion span, 7754 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 7755 CMSMarkStack* mark_stack): 7756 MetadataAwareOopClosure(collector->ref_processor()), 7757 _collector(collector), 7758 _span(span), 7759 _verification_bm(verification_bm), 7760 _cms_bm(cms_bm), 7761 _mark_stack(mark_stack) 7762 { } 7763 7764 void PushAndMarkVerifyClosure::do_oop(oop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 7765 void PushAndMarkVerifyClosure::do_oop(narrowOop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 7766 7767 // Upon stack overflow, we discard (part of) the stack, 7768 // remembering the least address amongst those discarded 7769 // in CMSCollector's _restart_address. 7770 void PushAndMarkVerifyClosure::handle_stack_overflow(HeapWord* lost) { 7771 // Remember the least grey address discarded 7772 HeapWord* ra = (HeapWord*)_mark_stack->least_value(lost); 7773 _collector->lower_restart_addr(ra); 7774 _mark_stack->reset(); // discard stack contents 7775 _mark_stack->expand(); // expand the stack if possible 7776 } 7777 7778 void PushAndMarkVerifyClosure::do_oop(oop obj) { 7779 assert(obj->is_oop_or_null(), "expected an oop or NULL"); 7780 HeapWord* addr = (HeapWord*)obj; 7781 if (_span.contains(addr) && !_verification_bm->isMarked(addr)) { 7782 // Oop lies in _span and isn't yet grey or black 7783 _verification_bm->mark(addr); // now grey 7784 if (!_cms_bm->isMarked(addr)) { 7785 oop(addr)->print(); 7786 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)", 7787 addr); 7788 fatal("... aborting"); 7789 } 7790 7791 if (!_mark_stack->push(obj)) { // stack overflow 7792 if (PrintCMSStatistics != 0) { 7793 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 7794 SIZE_FORMAT, _mark_stack->capacity()); 7795 } 7796 assert(_mark_stack->isFull(), "Else push should have succeeded"); 7797 handle_stack_overflow(addr); 7798 } 7799 // anything including and to the right of _finger 7800 // will be scanned as we iterate over the remainder of the 7801 // bit map 7802 } 7803 } 7804 7805 PushOrMarkClosure::PushOrMarkClosure(CMSCollector* collector, 7806 MemRegion span, 7807 CMSBitMap* bitMap, CMSMarkStack* markStack, 7808 HeapWord* finger, MarkFromRootsClosure* parent) : 7809 MetadataAwareOopClosure(collector->ref_processor()), 7810 _collector(collector), 7811 _span(span), 7812 _bitMap(bitMap), 7813 _markStack(markStack), 7814 _finger(finger), 7815 _parent(parent) 7816 { } 7817 7818 Par_PushOrMarkClosure::Par_PushOrMarkClosure(CMSCollector* collector, 7819 MemRegion span, 7820 CMSBitMap* bit_map, 7821 OopTaskQueue* work_queue, 7822 CMSMarkStack* overflow_stack, 7823 HeapWord* finger, 7824 HeapWord** global_finger_addr, 7825 Par_MarkFromRootsClosure* parent) : 7826 MetadataAwareOopClosure(collector->ref_processor()), 7827 _collector(collector), 7828 _whole_span(collector->_span), 7829 _span(span), 7830 _bit_map(bit_map), 7831 _work_queue(work_queue), 7832 _overflow_stack(overflow_stack), 7833 _finger(finger), 7834 _global_finger_addr(global_finger_addr), 7835 _parent(parent) 7836 { } 7837 7838 // Assumes thread-safe access by callers, who are 7839 // responsible for mutual exclusion. 7840 void CMSCollector::lower_restart_addr(HeapWord* low) { 7841 assert(_span.contains(low), "Out of bounds addr"); 7842 if (_restart_addr == NULL) { 7843 _restart_addr = low; 7844 } else { 7845 _restart_addr = MIN2(_restart_addr, low); 7846 } 7847 } 7848 7849 // Upon stack overflow, we discard (part of) the stack, 7850 // remembering the least address amongst those discarded 7851 // in CMSCollector's _restart_address. 7852 void PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 7853 // Remember the least grey address discarded 7854 HeapWord* ra = (HeapWord*)_markStack->least_value(lost); 7855 _collector->lower_restart_addr(ra); 7856 _markStack->reset(); // discard stack contents 7857 _markStack->expand(); // expand the stack if possible 7858 } 7859 7860 // Upon stack overflow, we discard (part of) the stack, 7861 // remembering the least address amongst those discarded 7862 // in CMSCollector's _restart_address. 7863 void Par_PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 7864 // We need to do this under a mutex to prevent other 7865 // workers from interfering with the work done below. 7866 MutexLockerEx ml(_overflow_stack->par_lock(), 7867 Mutex::_no_safepoint_check_flag); 7868 // Remember the least grey address discarded 7869 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 7870 _collector->lower_restart_addr(ra); 7871 _overflow_stack->reset(); // discard stack contents 7872 _overflow_stack->expand(); // expand the stack if possible 7873 } 7874 7875 void PushOrMarkClosure::do_oop(oop obj) { 7876 // Ignore mark word because we are running concurrent with mutators. 7877 assert(obj->is_oop_or_null(true), "expected an oop or NULL"); 7878 HeapWord* addr = (HeapWord*)obj; 7879 if (_span.contains(addr) && !_bitMap->isMarked(addr)) { 7880 // Oop lies in _span and isn't yet grey or black 7881 _bitMap->mark(addr); // now grey 7882 if (addr < _finger) { 7883 // the bit map iteration has already either passed, or 7884 // sampled, this bit in the bit map; we'll need to 7885 // use the marking stack to scan this oop's oops. 7886 bool simulate_overflow = false; 7887 NOT_PRODUCT( 7888 if (CMSMarkStackOverflowALot && 7889 _collector->simulate_overflow()) { 7890 // simulate a stack overflow 7891 simulate_overflow = true; 7892 } 7893 ) 7894 if (simulate_overflow || !_markStack->push(obj)) { // stack overflow 7895 if (PrintCMSStatistics != 0) { 7896 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 7897 SIZE_FORMAT, _markStack->capacity()); 7898 } 7899 assert(simulate_overflow || _markStack->isFull(), "Else push should have succeeded"); 7900 handle_stack_overflow(addr); 7901 } 7902 } 7903 // anything including and to the right of _finger 7904 // will be scanned as we iterate over the remainder of the 7905 // bit map 7906 do_yield_check(); 7907 } 7908 } 7909 7910 void PushOrMarkClosure::do_oop(oop* p) { PushOrMarkClosure::do_oop_work(p); } 7911 void PushOrMarkClosure::do_oop(narrowOop* p) { PushOrMarkClosure::do_oop_work(p); } 7912 7913 void Par_PushOrMarkClosure::do_oop(oop obj) { 7914 // Ignore mark word because we are running concurrent with mutators. 7915 assert(obj->is_oop_or_null(true), "expected an oop or NULL"); 7916 HeapWord* addr = (HeapWord*)obj; 7917 if (_whole_span.contains(addr) && !_bit_map->isMarked(addr)) { 7918 // Oop lies in _span and isn't yet grey or black 7919 // We read the global_finger (volatile read) strictly after marking oop 7920 bool res = _bit_map->par_mark(addr); // now grey 7921 volatile HeapWord** gfa = (volatile HeapWord**)_global_finger_addr; 7922 // Should we push this marked oop on our stack? 7923 // -- if someone else marked it, nothing to do 7924 // -- if target oop is above global finger nothing to do 7925 // -- if target oop is in chunk and above local finger 7926 // then nothing to do 7927 // -- else push on work queue 7928 if ( !res // someone else marked it, they will deal with it 7929 || (addr >= *gfa) // will be scanned in a later task 7930 || (_span.contains(addr) && addr >= _finger)) { // later in this chunk 7931 return; 7932 } 7933 // the bit map iteration has already either passed, or 7934 // sampled, this bit in the bit map; we'll need to 7935 // use the marking stack to scan this oop's oops. 7936 bool simulate_overflow = false; 7937 NOT_PRODUCT( 7938 if (CMSMarkStackOverflowALot && 7939 _collector->simulate_overflow()) { 7940 // simulate a stack overflow 7941 simulate_overflow = true; 7942 } 7943 ) 7944 if (simulate_overflow || 7945 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) { 7946 // stack overflow 7947 if (PrintCMSStatistics != 0) { 7948 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " 7949 SIZE_FORMAT, _overflow_stack->capacity()); 7950 } 7951 // We cannot assert that the overflow stack is full because 7952 // it may have been emptied since. 7953 assert(simulate_overflow || 7954 _work_queue->size() == _work_queue->max_elems(), 7955 "Else push should have succeeded"); 7956 handle_stack_overflow(addr); 7957 } 7958 do_yield_check(); 7959 } 7960 } 7961 7962 void Par_PushOrMarkClosure::do_oop(oop* p) { Par_PushOrMarkClosure::do_oop_work(p); } 7963 void Par_PushOrMarkClosure::do_oop(narrowOop* p) { Par_PushOrMarkClosure::do_oop_work(p); } 7964 7965 PushAndMarkClosure::PushAndMarkClosure(CMSCollector* collector, 7966 MemRegion span, 7967 ReferenceProcessor* rp, 7968 CMSBitMap* bit_map, 7969 CMSBitMap* mod_union_table, 7970 CMSMarkStack* mark_stack, 7971 bool concurrent_precleaning): 7972 MetadataAwareOopClosure(rp), 7973 _collector(collector), 7974 _span(span), 7975 _bit_map(bit_map), 7976 _mod_union_table(mod_union_table), 7977 _mark_stack(mark_stack), 7978 _concurrent_precleaning(concurrent_precleaning) 7979 { 7980 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 7981 } 7982 7983 // Grey object rescan during pre-cleaning and second checkpoint phases -- 7984 // the non-parallel version (the parallel version appears further below.) 7985 void PushAndMarkClosure::do_oop(oop obj) { 7986 // Ignore mark word verification. If during concurrent precleaning, 7987 // the object monitor may be locked. If during the checkpoint 7988 // phases, the object may already have been reached by a different 7989 // path and may be at the end of the global overflow list (so 7990 // the mark word may be NULL). 7991 assert(obj->is_oop_or_null(true /* ignore mark word */), 7992 "expected an oop or NULL"); 7993 HeapWord* addr = (HeapWord*)obj; 7994 // Check if oop points into the CMS generation 7995 // and is not marked 7996 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 7997 // a white object ... 7998 _bit_map->mark(addr); // ... now grey 7999 // push on the marking stack (grey set) 8000 bool simulate_overflow = false; 8001 NOT_PRODUCT( 8002 if (CMSMarkStackOverflowALot && 8003 _collector->simulate_overflow()) { 8004 // simulate a stack overflow 8005 simulate_overflow = true; 8006 } 8007 ) 8008 if (simulate_overflow || !_mark_stack->push(obj)) { 8009 if (_concurrent_precleaning) { 8010 // During precleaning we can just dirty the appropriate card(s) 8011 // in the mod union table, thus ensuring that the object remains 8012 // in the grey set and continue. In the case of object arrays 8013 // we need to dirty all of the cards that the object spans, 8014 // since the rescan of object arrays will be limited to the 8015 // dirty cards. 8016 // Note that no one can be intefering with us in this action 8017 // of dirtying the mod union table, so no locking or atomics 8018 // are required. 8019 if (obj->is_objArray()) { 8020 size_t sz = obj->size(); 8021 HeapWord* end_card_addr = (HeapWord*)round_to( 8022 (intptr_t)(addr+sz), CardTableModRefBS::card_size); 8023 MemRegion redirty_range = MemRegion(addr, end_card_addr); 8024 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 8025 _mod_union_table->mark_range(redirty_range); 8026 } else { 8027 _mod_union_table->mark(addr); 8028 } 8029 _collector->_ser_pmc_preclean_ovflw++; 8030 } else { 8031 // During the remark phase, we need to remember this oop 8032 // in the overflow list. 8033 _collector->push_on_overflow_list(obj); 8034 _collector->_ser_pmc_remark_ovflw++; 8035 } 8036 } 8037 } 8038 } 8039 8040 Par_PushAndMarkClosure::Par_PushAndMarkClosure(CMSCollector* collector, 8041 MemRegion span, 8042 ReferenceProcessor* rp, 8043 CMSBitMap* bit_map, 8044 OopTaskQueue* work_queue): 8045 MetadataAwareOopClosure(rp), 8046 _collector(collector), 8047 _span(span), 8048 _bit_map(bit_map), 8049 _work_queue(work_queue) 8050 { 8051 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); 8052 } 8053 8054 void PushAndMarkClosure::do_oop(oop* p) { PushAndMarkClosure::do_oop_work(p); } 8055 void PushAndMarkClosure::do_oop(narrowOop* p) { PushAndMarkClosure::do_oop_work(p); } 8056 8057 // Grey object rescan during second checkpoint phase -- 8058 // the parallel version. 8059 void Par_PushAndMarkClosure::do_oop(oop obj) { 8060 // In the assert below, we ignore the mark word because 8061 // this oop may point to an already visited object that is 8062 // on the overflow stack (in which case the mark word has 8063 // been hijacked for chaining into the overflow stack -- 8064 // if this is the last object in the overflow stack then 8065 // its mark word will be NULL). Because this object may 8066 // have been subsequently popped off the global overflow 8067 // stack, and the mark word possibly restored to the prototypical 8068 // value, by the time we get to examined this failing assert in 8069 // the debugger, is_oop_or_null(false) may subsequently start 8070 // to hold. 8071 assert(obj->is_oop_or_null(true), 8072 "expected an oop or NULL"); 8073 HeapWord* addr = (HeapWord*)obj; 8074 // Check if oop points into the CMS generation 8075 // and is not marked 8076 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 8077 // a white object ... 8078 // If we manage to "claim" the object, by being the 8079 // first thread to mark it, then we push it on our 8080 // marking stack 8081 if (_bit_map->par_mark(addr)) { // ... now grey 8082 // push on work queue (grey set) 8083 bool simulate_overflow = false; 8084 NOT_PRODUCT( 8085 if (CMSMarkStackOverflowALot && 8086 _collector->par_simulate_overflow()) { 8087 // simulate a stack overflow 8088 simulate_overflow = true; 8089 } 8090 ) 8091 if (simulate_overflow || !_work_queue->push(obj)) { 8092 _collector->par_push_on_overflow_list(obj); 8093 _collector->_par_pmc_remark_ovflw++; // imprecise OK: no need to CAS 8094 } 8095 } // Else, some other thread got there first 8096 } 8097 } 8098 8099 void Par_PushAndMarkClosure::do_oop(oop* p) { Par_PushAndMarkClosure::do_oop_work(p); } 8100 void Par_PushAndMarkClosure::do_oop(narrowOop* p) { Par_PushAndMarkClosure::do_oop_work(p); } 8101 8102 void CMSPrecleanRefsYieldClosure::do_yield_work() { 8103 Mutex* bml = _collector->bitMapLock(); 8104 assert_lock_strong(bml); 8105 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 8106 "CMS thread should hold CMS token"); 8107 8108 bml->unlock(); 8109 ConcurrentMarkSweepThread::desynchronize(true); 8110 8111 ConcurrentMarkSweepThread::acknowledge_yield_request(); 8112 8113 _collector->stopTimer(); 8114 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 8115 if (PrintCMSStatistics != 0) { 8116 _collector->incrementYields(); 8117 } 8118 _collector->icms_wait(); 8119 8120 // See the comment in coordinator_yield() 8121 for (unsigned i = 0; i < CMSYieldSleepCount && 8122 ConcurrentMarkSweepThread::should_yield() && 8123 !CMSCollector::foregroundGCIsActive(); ++i) { 8124 os::sleep(Thread::current(), 1, false); 8125 ConcurrentMarkSweepThread::acknowledge_yield_request(); 8126 } 8127 8128 ConcurrentMarkSweepThread::synchronize(true); 8129 bml->lock(); 8130 8131 _collector->startTimer(); 8132 } 8133 8134 bool CMSPrecleanRefsYieldClosure::should_return() { 8135 if (ConcurrentMarkSweepThread::should_yield()) { 8136 do_yield_work(); 8137 } 8138 return _collector->foregroundGCIsActive(); 8139 } 8140 8141 void MarkFromDirtyCardsClosure::do_MemRegion(MemRegion mr) { 8142 assert(((size_t)mr.start())%CardTableModRefBS::card_size_in_words == 0, 8143 "mr should be aligned to start at a card boundary"); 8144 // We'd like to assert: 8145 // assert(mr.word_size()%CardTableModRefBS::card_size_in_words == 0, 8146 // "mr should be a range of cards"); 8147 // However, that would be too strong in one case -- the last 8148 // partition ends at _unallocated_block which, in general, can be 8149 // an arbitrary boundary, not necessarily card aligned. 8150 if (PrintCMSStatistics != 0) { 8151 _num_dirty_cards += 8152 mr.word_size()/CardTableModRefBS::card_size_in_words; 8153 } 8154 _space->object_iterate_mem(mr, &_scan_cl); 8155 } 8156 8157 SweepClosure::SweepClosure(CMSCollector* collector, 8158 ConcurrentMarkSweepGeneration* g, 8159 CMSBitMap* bitMap, bool should_yield) : 8160 _collector(collector), 8161 _g(g), 8162 _sp(g->cmsSpace()), 8163 _limit(_sp->sweep_limit()), 8164 _freelistLock(_sp->freelistLock()), 8165 _bitMap(bitMap), 8166 _yield(should_yield), 8167 _inFreeRange(false), // No free range at beginning of sweep 8168 _freeRangeInFreeLists(false), // No free range at beginning of sweep 8169 _lastFreeRangeCoalesced(false), 8170 _freeFinger(g->used_region().start()) 8171 { 8172 NOT_PRODUCT( 8173 _numObjectsFreed = 0; 8174 _numWordsFreed = 0; 8175 _numObjectsLive = 0; 8176 _numWordsLive = 0; 8177 _numObjectsAlreadyFree = 0; 8178 _numWordsAlreadyFree = 0; 8179 _last_fc = NULL; 8180 8181 _sp->initializeIndexedFreeListArrayReturnedBytes(); 8182 _sp->dictionary()->initialize_dict_returned_bytes(); 8183 ) 8184 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 8185 "sweep _limit out of bounds"); 8186 if (CMSTraceSweeper) { 8187 gclog_or_tty->print_cr("\n====================\nStarting new sweep with limit " PTR_FORMAT, 8188 _limit); 8189 } 8190 } 8191 8192 void SweepClosure::print_on(outputStream* st) const { 8193 tty->print_cr("_sp = [" PTR_FORMAT "," PTR_FORMAT ")", 8194 _sp->bottom(), _sp->end()); 8195 tty->print_cr("_limit = " PTR_FORMAT, _limit); 8196 tty->print_cr("_freeFinger = " PTR_FORMAT, _freeFinger); 8197 NOT_PRODUCT(tty->print_cr("_last_fc = " PTR_FORMAT, _last_fc);) 8198 tty->print_cr("_inFreeRange = %d, _freeRangeInFreeLists = %d, _lastFreeRangeCoalesced = %d", 8199 _inFreeRange, _freeRangeInFreeLists, _lastFreeRangeCoalesced); 8200 } 8201 8202 #ifndef PRODUCT 8203 // Assertion checking only: no useful work in product mode -- 8204 // however, if any of the flags below become product flags, 8205 // you may need to review this code to see if it needs to be 8206 // enabled in product mode. 8207 SweepClosure::~SweepClosure() { 8208 assert_lock_strong(_freelistLock); 8209 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 8210 "sweep _limit out of bounds"); 8211 if (inFreeRange()) { 8212 warning("inFreeRange() should have been reset; dumping state of SweepClosure"); 8213 print(); 8214 ShouldNotReachHere(); 8215 } 8216 if (Verbose && PrintGC) { 8217 gclog_or_tty->print("Collected "SIZE_FORMAT" objects, " SIZE_FORMAT " bytes", 8218 _numObjectsFreed, _numWordsFreed*sizeof(HeapWord)); 8219 gclog_or_tty->print_cr("\nLive "SIZE_FORMAT" objects, " 8220 SIZE_FORMAT" bytes " 8221 "Already free "SIZE_FORMAT" objects, "SIZE_FORMAT" bytes", 8222 _numObjectsLive, _numWordsLive*sizeof(HeapWord), 8223 _numObjectsAlreadyFree, _numWordsAlreadyFree*sizeof(HeapWord)); 8224 size_t totalBytes = (_numWordsFreed + _numWordsLive + _numWordsAlreadyFree) 8225 * sizeof(HeapWord); 8226 gclog_or_tty->print_cr("Total sweep: "SIZE_FORMAT" bytes", totalBytes); 8227 8228 if (PrintCMSStatistics && CMSVerifyReturnedBytes) { 8229 size_t indexListReturnedBytes = _sp->sumIndexedFreeListArrayReturnedBytes(); 8230 size_t dict_returned_bytes = _sp->dictionary()->sum_dict_returned_bytes(); 8231 size_t returned_bytes = indexListReturnedBytes + dict_returned_bytes; 8232 gclog_or_tty->print("Returned "SIZE_FORMAT" bytes", returned_bytes); 8233 gclog_or_tty->print(" Indexed List Returned "SIZE_FORMAT" bytes", 8234 indexListReturnedBytes); 8235 gclog_or_tty->print_cr(" Dictionary Returned "SIZE_FORMAT" bytes", 8236 dict_returned_bytes); 8237 } 8238 } 8239 if (CMSTraceSweeper) { 8240 gclog_or_tty->print_cr("end of sweep with _limit = " PTR_FORMAT "\n================", 8241 _limit); 8242 } 8243 } 8244 #endif // PRODUCT 8245 8246 void SweepClosure::initialize_free_range(HeapWord* freeFinger, 8247 bool freeRangeInFreeLists) { 8248 if (CMSTraceSweeper) { 8249 gclog_or_tty->print("---- Start free range at 0x%x with free block (%d)\n", 8250 freeFinger, freeRangeInFreeLists); 8251 } 8252 assert(!inFreeRange(), "Trampling existing free range"); 8253 set_inFreeRange(true); 8254 set_lastFreeRangeCoalesced(false); 8255 8256 set_freeFinger(freeFinger); 8257 set_freeRangeInFreeLists(freeRangeInFreeLists); 8258 if (CMSTestInFreeList) { 8259 if (freeRangeInFreeLists) { 8260 FreeChunk* fc = (FreeChunk*) freeFinger; 8261 assert(fc->is_free(), "A chunk on the free list should be free."); 8262 assert(fc->size() > 0, "Free range should have a size"); 8263 assert(_sp->verify_chunk_in_free_list(fc), "Chunk is not in free lists"); 8264 } 8265 } 8266 } 8267 8268 // Note that the sweeper runs concurrently with mutators. Thus, 8269 // it is possible for direct allocation in this generation to happen 8270 // in the middle of the sweep. Note that the sweeper also coalesces 8271 // contiguous free blocks. Thus, unless the sweeper and the allocator 8272 // synchronize appropriately freshly allocated blocks may get swept up. 8273 // This is accomplished by the sweeper locking the free lists while 8274 // it is sweeping. Thus blocks that are determined to be free are 8275 // indeed free. There is however one additional complication: 8276 // blocks that have been allocated since the final checkpoint and 8277 // mark, will not have been marked and so would be treated as 8278 // unreachable and swept up. To prevent this, the allocator marks 8279 // the bit map when allocating during the sweep phase. This leads, 8280 // however, to a further complication -- objects may have been allocated 8281 // but not yet initialized -- in the sense that the header isn't yet 8282 // installed. The sweeper can not then determine the size of the block 8283 // in order to skip over it. To deal with this case, we use a technique 8284 // (due to Printezis) to encode such uninitialized block sizes in the 8285 // bit map. Since the bit map uses a bit per every HeapWord, but the 8286 // CMS generation has a minimum object size of 3 HeapWords, it follows 8287 // that "normal marks" won't be adjacent in the bit map (there will 8288 // always be at least two 0 bits between successive 1 bits). We make use 8289 // of these "unused" bits to represent uninitialized blocks -- the bit 8290 // corresponding to the start of the uninitialized object and the next 8291 // bit are both set. Finally, a 1 bit marks the end of the object that 8292 // started with the two consecutive 1 bits to indicate its potentially 8293 // uninitialized state. 8294 8295 size_t SweepClosure::do_blk_careful(HeapWord* addr) { 8296 FreeChunk* fc = (FreeChunk*)addr; 8297 size_t res; 8298 8299 // Check if we are done sweeping. Below we check "addr >= _limit" rather 8300 // than "addr == _limit" because although _limit was a block boundary when 8301 // we started the sweep, it may no longer be one because heap expansion 8302 // may have caused us to coalesce the block ending at the address _limit 8303 // with a newly expanded chunk (this happens when _limit was set to the 8304 // previous _end of the space), so we may have stepped past _limit: 8305 // see the following Zeno-like trail of CRs 6977970, 7008136, 7042740. 8306 if (addr >= _limit) { // we have swept up to or past the limit: finish up 8307 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 8308 "sweep _limit out of bounds"); 8309 assert(addr < _sp->end(), "addr out of bounds"); 8310 // Flush any free range we might be holding as a single 8311 // coalesced chunk to the appropriate free list. 8312 if (inFreeRange()) { 8313 assert(freeFinger() >= _sp->bottom() && freeFinger() < _limit, 8314 err_msg("freeFinger() " PTR_FORMAT" is out-of-bounds", freeFinger())); 8315 flush_cur_free_chunk(freeFinger(), 8316 pointer_delta(addr, freeFinger())); 8317 if (CMSTraceSweeper) { 8318 gclog_or_tty->print("Sweep: last chunk: "); 8319 gclog_or_tty->print("put_free_blk 0x%x ("SIZE_FORMAT") " 8320 "[coalesced:"SIZE_FORMAT"]\n", 8321 freeFinger(), pointer_delta(addr, freeFinger()), 8322 lastFreeRangeCoalesced()); 8323 } 8324 } 8325 8326 // help the iterator loop finish 8327 return pointer_delta(_sp->end(), addr); 8328 } 8329 8330 assert(addr < _limit, "sweep invariant"); 8331 // check if we should yield 8332 do_yield_check(addr); 8333 if (fc->is_free()) { 8334 // Chunk that is already free 8335 res = fc->size(); 8336 do_already_free_chunk(fc); 8337 debug_only(_sp->verifyFreeLists()); 8338 // If we flush the chunk at hand in lookahead_and_flush() 8339 // and it's coalesced with a preceding chunk, then the 8340 // process of "mangling" the payload of the coalesced block 8341 // will cause erasure of the size information from the 8342 // (erstwhile) header of all the coalesced blocks but the 8343 // first, so the first disjunct in the assert will not hold 8344 // in that specific case (in which case the second disjunct 8345 // will hold). 8346 assert(res == fc->size() || ((HeapWord*)fc) + res >= _limit, 8347 "Otherwise the size info doesn't change at this step"); 8348 NOT_PRODUCT( 8349 _numObjectsAlreadyFree++; 8350 _numWordsAlreadyFree += res; 8351 ) 8352 NOT_PRODUCT(_last_fc = fc;) 8353 } else if (!_bitMap->isMarked(addr)) { 8354 // Chunk is fresh garbage 8355 res = do_garbage_chunk(fc); 8356 debug_only(_sp->verifyFreeLists()); 8357 NOT_PRODUCT( 8358 _numObjectsFreed++; 8359 _numWordsFreed += res; 8360 ) 8361 } else { 8362 // Chunk that is alive. 8363 res = do_live_chunk(fc); 8364 debug_only(_sp->verifyFreeLists()); 8365 NOT_PRODUCT( 8366 _numObjectsLive++; 8367 _numWordsLive += res; 8368 ) 8369 } 8370 return res; 8371 } 8372 8373 // For the smart allocation, record following 8374 // split deaths - a free chunk is removed from its free list because 8375 // it is being split into two or more chunks. 8376 // split birth - a free chunk is being added to its free list because 8377 // a larger free chunk has been split and resulted in this free chunk. 8378 // coal death - a free chunk is being removed from its free list because 8379 // it is being coalesced into a large free chunk. 8380 // coal birth - a free chunk is being added to its free list because 8381 // it was created when two or more free chunks where coalesced into 8382 // this free chunk. 8383 // 8384 // These statistics are used to determine the desired number of free 8385 // chunks of a given size. The desired number is chosen to be relative 8386 // to the end of a CMS sweep. The desired number at the end of a sweep 8387 // is the 8388 // count-at-end-of-previous-sweep (an amount that was enough) 8389 // - count-at-beginning-of-current-sweep (the excess) 8390 // + split-births (gains in this size during interval) 8391 // - split-deaths (demands on this size during interval) 8392 // where the interval is from the end of one sweep to the end of the 8393 // next. 8394 // 8395 // When sweeping the sweeper maintains an accumulated chunk which is 8396 // the chunk that is made up of chunks that have been coalesced. That 8397 // will be termed the left-hand chunk. A new chunk of garbage that 8398 // is being considered for coalescing will be referred to as the 8399 // right-hand chunk. 8400 // 8401 // When making a decision on whether to coalesce a right-hand chunk with 8402 // the current left-hand chunk, the current count vs. the desired count 8403 // of the left-hand chunk is considered. Also if the right-hand chunk 8404 // is near the large chunk at the end of the heap (see 8405 // ConcurrentMarkSweepGeneration::isNearLargestChunk()), then the 8406 // left-hand chunk is coalesced. 8407 // 8408 // When making a decision about whether to split a chunk, the desired count 8409 // vs. the current count of the candidate to be split is also considered. 8410 // If the candidate is underpopulated (currently fewer chunks than desired) 8411 // a chunk of an overpopulated (currently more chunks than desired) size may 8412 // be chosen. The "hint" associated with a free list, if non-null, points 8413 // to a free list which may be overpopulated. 8414 // 8415 8416 void SweepClosure::do_already_free_chunk(FreeChunk* fc) { 8417 const size_t size = fc->size(); 8418 // Chunks that cannot be coalesced are not in the 8419 // free lists. 8420 if (CMSTestInFreeList && !fc->cantCoalesce()) { 8421 assert(_sp->verify_chunk_in_free_list(fc), 8422 "free chunk should be in free lists"); 8423 } 8424 // a chunk that is already free, should not have been 8425 // marked in the bit map 8426 HeapWord* const addr = (HeapWord*) fc; 8427 assert(!_bitMap->isMarked(addr), "free chunk should be unmarked"); 8428 // Verify that the bit map has no bits marked between 8429 // addr and purported end of this block. 8430 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 8431 8432 // Some chunks cannot be coalesced under any circumstances. 8433 // See the definition of cantCoalesce(). 8434 if (!fc->cantCoalesce()) { 8435 // This chunk can potentially be coalesced. 8436 if (_sp->adaptive_freelists()) { 8437 // All the work is done in 8438 do_post_free_or_garbage_chunk(fc, size); 8439 } else { // Not adaptive free lists 8440 // this is a free chunk that can potentially be coalesced by the sweeper; 8441 if (!inFreeRange()) { 8442 // if the next chunk is a free block that can't be coalesced 8443 // it doesn't make sense to remove this chunk from the free lists 8444 FreeChunk* nextChunk = (FreeChunk*)(addr + size); 8445 assert((HeapWord*)nextChunk <= _sp->end(), "Chunk size out of bounds?"); 8446 if ((HeapWord*)nextChunk < _sp->end() && // There is another free chunk to the right ... 8447 nextChunk->is_free() && // ... which is free... 8448 nextChunk->cantCoalesce()) { // ... but can't be coalesced 8449 // nothing to do 8450 } else { 8451 // Potentially the start of a new free range: 8452 // Don't eagerly remove it from the free lists. 8453 // No need to remove it if it will just be put 8454 // back again. (Also from a pragmatic point of view 8455 // if it is a free block in a region that is beyond 8456 // any allocated blocks, an assertion will fail) 8457 // Remember the start of a free run. 8458 initialize_free_range(addr, true); 8459 // end - can coalesce with next chunk 8460 } 8461 } else { 8462 // the midst of a free range, we are coalescing 8463 print_free_block_coalesced(fc); 8464 if (CMSTraceSweeper) { 8465 gclog_or_tty->print(" -- pick up free block 0x%x (%d)\n", fc, size); 8466 } 8467 // remove it from the free lists 8468 _sp->removeFreeChunkFromFreeLists(fc); 8469 set_lastFreeRangeCoalesced(true); 8470 // If the chunk is being coalesced and the current free range is 8471 // in the free lists, remove the current free range so that it 8472 // will be returned to the free lists in its entirety - all 8473 // the coalesced pieces included. 8474 if (freeRangeInFreeLists()) { 8475 FreeChunk* ffc = (FreeChunk*) freeFinger(); 8476 assert(ffc->size() == pointer_delta(addr, freeFinger()), 8477 "Size of free range is inconsistent with chunk size."); 8478 if (CMSTestInFreeList) { 8479 assert(_sp->verify_chunk_in_free_list(ffc), 8480 "free range is not in free lists"); 8481 } 8482 _sp->removeFreeChunkFromFreeLists(ffc); 8483 set_freeRangeInFreeLists(false); 8484 } 8485 } 8486 } 8487 // Note that if the chunk is not coalescable (the else arm 8488 // below), we unconditionally flush, without needing to do 8489 // a "lookahead," as we do below. 8490 if (inFreeRange()) lookahead_and_flush(fc, size); 8491 } else { 8492 // Code path common to both original and adaptive free lists. 8493 8494 // cant coalesce with previous block; this should be treated 8495 // as the end of a free run if any 8496 if (inFreeRange()) { 8497 // we kicked some butt; time to pick up the garbage 8498 assert(freeFinger() < addr, "freeFinger points too high"); 8499 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 8500 } 8501 // else, nothing to do, just continue 8502 } 8503 } 8504 8505 size_t SweepClosure::do_garbage_chunk(FreeChunk* fc) { 8506 // This is a chunk of garbage. It is not in any free list. 8507 // Add it to a free list or let it possibly be coalesced into 8508 // a larger chunk. 8509 HeapWord* const addr = (HeapWord*) fc; 8510 const size_t size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 8511 8512 if (_sp->adaptive_freelists()) { 8513 // Verify that the bit map has no bits marked between 8514 // addr and purported end of just dead object. 8515 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 8516 8517 do_post_free_or_garbage_chunk(fc, size); 8518 } else { 8519 if (!inFreeRange()) { 8520 // start of a new free range 8521 assert(size > 0, "A free range should have a size"); 8522 initialize_free_range(addr, false); 8523 } else { 8524 // this will be swept up when we hit the end of the 8525 // free range 8526 if (CMSTraceSweeper) { 8527 gclog_or_tty->print(" -- pick up garbage 0x%x (%d) \n", fc, size); 8528 } 8529 // If the chunk is being coalesced and the current free range is 8530 // in the free lists, remove the current free range so that it 8531 // will be returned to the free lists in its entirety - all 8532 // the coalesced pieces included. 8533 if (freeRangeInFreeLists()) { 8534 FreeChunk* ffc = (FreeChunk*)freeFinger(); 8535 assert(ffc->size() == pointer_delta(addr, freeFinger()), 8536 "Size of free range is inconsistent with chunk size."); 8537 if (CMSTestInFreeList) { 8538 assert(_sp->verify_chunk_in_free_list(ffc), 8539 "free range is not in free lists"); 8540 } 8541 _sp->removeFreeChunkFromFreeLists(ffc); 8542 set_freeRangeInFreeLists(false); 8543 } 8544 set_lastFreeRangeCoalesced(true); 8545 } 8546 // this will be swept up when we hit the end of the free range 8547 8548 // Verify that the bit map has no bits marked between 8549 // addr and purported end of just dead object. 8550 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 8551 } 8552 assert(_limit >= addr + size, 8553 "A freshly garbage chunk can't possibly straddle over _limit"); 8554 if (inFreeRange()) lookahead_and_flush(fc, size); 8555 return size; 8556 } 8557 8558 size_t SweepClosure::do_live_chunk(FreeChunk* fc) { 8559 HeapWord* addr = (HeapWord*) fc; 8560 // The sweeper has just found a live object. Return any accumulated 8561 // left hand chunk to the free lists. 8562 if (inFreeRange()) { 8563 assert(freeFinger() < addr, "freeFinger points too high"); 8564 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 8565 } 8566 8567 // This object is live: we'd normally expect this to be 8568 // an oop, and like to assert the following: 8569 // assert(oop(addr)->is_oop(), "live block should be an oop"); 8570 // However, as we commented above, this may be an object whose 8571 // header hasn't yet been initialized. 8572 size_t size; 8573 assert(_bitMap->isMarked(addr), "Tautology for this control point"); 8574 if (_bitMap->isMarked(addr + 1)) { 8575 // Determine the size from the bit map, rather than trying to 8576 // compute it from the object header. 8577 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 8578 size = pointer_delta(nextOneAddr + 1, addr); 8579 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 8580 "alignment problem"); 8581 8582 #ifdef ASSERT 8583 if (oop(addr)->klass_or_null() != NULL) { 8584 // Ignore mark word because we are running concurrent with mutators 8585 assert(oop(addr)->is_oop(true), "live block should be an oop"); 8586 assert(size == 8587 CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()), 8588 "P-mark and computed size do not agree"); 8589 } 8590 #endif 8591 8592 } else { 8593 // This should be an initialized object that's alive. 8594 assert(oop(addr)->klass_or_null() != NULL, 8595 "Should be an initialized object"); 8596 // Ignore mark word because we are running concurrent with mutators 8597 assert(oop(addr)->is_oop(true), "live block should be an oop"); 8598 // Verify that the bit map has no bits marked between 8599 // addr and purported end of this block. 8600 size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 8601 assert(size >= 3, "Necessary for Printezis marks to work"); 8602 assert(!_bitMap->isMarked(addr+1), "Tautology for this control point"); 8603 DEBUG_ONLY(_bitMap->verifyNoOneBitsInRange(addr+2, addr+size);) 8604 } 8605 return size; 8606 } 8607 8608 void SweepClosure::do_post_free_or_garbage_chunk(FreeChunk* fc, 8609 size_t chunkSize) { 8610 // do_post_free_or_garbage_chunk() should only be called in the case 8611 // of the adaptive free list allocator. 8612 const bool fcInFreeLists = fc->is_free(); 8613 assert(_sp->adaptive_freelists(), "Should only be used in this case."); 8614 assert((HeapWord*)fc <= _limit, "sweep invariant"); 8615 if (CMSTestInFreeList && fcInFreeLists) { 8616 assert(_sp->verify_chunk_in_free_list(fc), "free chunk is not in free lists"); 8617 } 8618 8619 if (CMSTraceSweeper) { 8620 gclog_or_tty->print_cr(" -- pick up another chunk at 0x%x (%d)", fc, chunkSize); 8621 } 8622 8623 HeapWord* const fc_addr = (HeapWord*) fc; 8624 8625 bool coalesce; 8626 const size_t left = pointer_delta(fc_addr, freeFinger()); 8627 const size_t right = chunkSize; 8628 switch (FLSCoalescePolicy) { 8629 // numeric value forms a coalition aggressiveness metric 8630 case 0: { // never coalesce 8631 coalesce = false; 8632 break; 8633 } 8634 case 1: { // coalesce if left & right chunks on overpopulated lists 8635 coalesce = _sp->coalOverPopulated(left) && 8636 _sp->coalOverPopulated(right); 8637 break; 8638 } 8639 case 2: { // coalesce if left chunk on overpopulated list (default) 8640 coalesce = _sp->coalOverPopulated(left); 8641 break; 8642 } 8643 case 3: { // coalesce if left OR right chunk on overpopulated list 8644 coalesce = _sp->coalOverPopulated(left) || 8645 _sp->coalOverPopulated(right); 8646 break; 8647 } 8648 case 4: { // always coalesce 8649 coalesce = true; 8650 break; 8651 } 8652 default: 8653 ShouldNotReachHere(); 8654 } 8655 8656 // Should the current free range be coalesced? 8657 // If the chunk is in a free range and either we decided to coalesce above 8658 // or the chunk is near the large block at the end of the heap 8659 // (isNearLargestChunk() returns true), then coalesce this chunk. 8660 const bool doCoalesce = inFreeRange() 8661 && (coalesce || _g->isNearLargestChunk(fc_addr)); 8662 if (doCoalesce) { 8663 // Coalesce the current free range on the left with the new 8664 // chunk on the right. If either is on a free list, 8665 // it must be removed from the list and stashed in the closure. 8666 if (freeRangeInFreeLists()) { 8667 FreeChunk* const ffc = (FreeChunk*)freeFinger(); 8668 assert(ffc->size() == pointer_delta(fc_addr, freeFinger()), 8669 "Size of free range is inconsistent with chunk size."); 8670 if (CMSTestInFreeList) { 8671 assert(_sp->verify_chunk_in_free_list(ffc), 8672 "Chunk is not in free lists"); 8673 } 8674 _sp->coalDeath(ffc->size()); 8675 _sp->removeFreeChunkFromFreeLists(ffc); 8676 set_freeRangeInFreeLists(false); 8677 } 8678 if (fcInFreeLists) { 8679 _sp->coalDeath(chunkSize); 8680 assert(fc->size() == chunkSize, 8681 "The chunk has the wrong size or is not in the free lists"); 8682 _sp->removeFreeChunkFromFreeLists(fc); 8683 } 8684 set_lastFreeRangeCoalesced(true); 8685 print_free_block_coalesced(fc); 8686 } else { // not in a free range and/or should not coalesce 8687 // Return the current free range and start a new one. 8688 if (inFreeRange()) { 8689 // In a free range but cannot coalesce with the right hand chunk. 8690 // Put the current free range into the free lists. 8691 flush_cur_free_chunk(freeFinger(), 8692 pointer_delta(fc_addr, freeFinger())); 8693 } 8694 // Set up for new free range. Pass along whether the right hand 8695 // chunk is in the free lists. 8696 initialize_free_range((HeapWord*)fc, fcInFreeLists); 8697 } 8698 } 8699 8700 // Lookahead flush: 8701 // If we are tracking a free range, and this is the last chunk that 8702 // we'll look at because its end crosses past _limit, we'll preemptively 8703 // flush it along with any free range we may be holding on to. Note that 8704 // this can be the case only for an already free or freshly garbage 8705 // chunk. If this block is an object, it can never straddle 8706 // over _limit. The "straddling" occurs when _limit is set at 8707 // the previous end of the space when this cycle started, and 8708 // a subsequent heap expansion caused the previously co-terminal 8709 // free block to be coalesced with the newly expanded portion, 8710 // thus rendering _limit a non-block-boundary making it dangerous 8711 // for the sweeper to step over and examine. 8712 void SweepClosure::lookahead_and_flush(FreeChunk* fc, size_t chunk_size) { 8713 assert(inFreeRange(), "Should only be called if currently in a free range."); 8714 HeapWord* const eob = ((HeapWord*)fc) + chunk_size; 8715 assert(_sp->used_region().contains(eob - 1), 8716 err_msg("eob = " PTR_FORMAT " eob-1 = " PTR_FORMAT " _limit = " PTR_FORMAT 8717 " out of bounds wrt _sp = [" PTR_FORMAT "," PTR_FORMAT ")" 8718 " when examining fc = " PTR_FORMAT "(" SIZE_FORMAT ")", 8719 eob, eob-1, _limit, _sp->bottom(), _sp->end(), fc, chunk_size)); 8720 if (eob >= _limit) { 8721 assert(eob == _limit || fc->is_free(), "Only a free chunk should allow us to cross over the limit"); 8722 if (CMSTraceSweeper) { 8723 gclog_or_tty->print_cr("_limit " PTR_FORMAT " reached or crossed by block " 8724 "[" PTR_FORMAT "," PTR_FORMAT ") in space " 8725 "[" PTR_FORMAT "," PTR_FORMAT ")", 8726 _limit, fc, eob, _sp->bottom(), _sp->end()); 8727 } 8728 // Return the storage we are tracking back into the free lists. 8729 if (CMSTraceSweeper) { 8730 gclog_or_tty->print_cr("Flushing ... "); 8731 } 8732 assert(freeFinger() < eob, "Error"); 8733 flush_cur_free_chunk( freeFinger(), pointer_delta(eob, freeFinger())); 8734 } 8735 } 8736 8737 void SweepClosure::flush_cur_free_chunk(HeapWord* chunk, size_t size) { 8738 assert(inFreeRange(), "Should only be called if currently in a free range."); 8739 assert(size > 0, 8740 "A zero sized chunk cannot be added to the free lists."); 8741 if (!freeRangeInFreeLists()) { 8742 if (CMSTestInFreeList) { 8743 FreeChunk* fc = (FreeChunk*) chunk; 8744 fc->set_size(size); 8745 assert(!_sp->verify_chunk_in_free_list(fc), 8746 "chunk should not be in free lists yet"); 8747 } 8748 if (CMSTraceSweeper) { 8749 gclog_or_tty->print_cr(" -- add free block 0x%x (%d) to free lists", 8750 chunk, size); 8751 } 8752 // A new free range is going to be starting. The current 8753 // free range has not been added to the free lists yet or 8754 // was removed so add it back. 8755 // If the current free range was coalesced, then the death 8756 // of the free range was recorded. Record a birth now. 8757 if (lastFreeRangeCoalesced()) { 8758 _sp->coalBirth(size); 8759 } 8760 _sp->addChunkAndRepairOffsetTable(chunk, size, 8761 lastFreeRangeCoalesced()); 8762 } else if (CMSTraceSweeper) { 8763 gclog_or_tty->print_cr("Already in free list: nothing to flush"); 8764 } 8765 set_inFreeRange(false); 8766 set_freeRangeInFreeLists(false); 8767 } 8768 8769 // We take a break if we've been at this for a while, 8770 // so as to avoid monopolizing the locks involved. 8771 void SweepClosure::do_yield_work(HeapWord* addr) { 8772 // Return current free chunk being used for coalescing (if any) 8773 // to the appropriate freelist. After yielding, the next 8774 // free block encountered will start a coalescing range of 8775 // free blocks. If the next free block is adjacent to the 8776 // chunk just flushed, they will need to wait for the next 8777 // sweep to be coalesced. 8778 if (inFreeRange()) { 8779 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 8780 } 8781 8782 // First give up the locks, then yield, then re-lock. 8783 // We should probably use a constructor/destructor idiom to 8784 // do this unlock/lock or modify the MutexUnlocker class to 8785 // serve our purpose. XXX 8786 assert_lock_strong(_bitMap->lock()); 8787 assert_lock_strong(_freelistLock); 8788 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 8789 "CMS thread should hold CMS token"); 8790 _bitMap->lock()->unlock(); 8791 _freelistLock->unlock(); 8792 ConcurrentMarkSweepThread::desynchronize(true); 8793 ConcurrentMarkSweepThread::acknowledge_yield_request(); 8794 _collector->stopTimer(); 8795 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); 8796 if (PrintCMSStatistics != 0) { 8797 _collector->incrementYields(); 8798 } 8799 _collector->icms_wait(); 8800 8801 // See the comment in coordinator_yield() 8802 for (unsigned i = 0; i < CMSYieldSleepCount && 8803 ConcurrentMarkSweepThread::should_yield() && 8804 !CMSCollector::foregroundGCIsActive(); ++i) { 8805 os::sleep(Thread::current(), 1, false); 8806 ConcurrentMarkSweepThread::acknowledge_yield_request(); 8807 } 8808 8809 ConcurrentMarkSweepThread::synchronize(true); 8810 _freelistLock->lock(); 8811 _bitMap->lock()->lock_without_safepoint_check(); 8812 _collector->startTimer(); 8813 } 8814 8815 #ifndef PRODUCT 8816 // This is actually very useful in a product build if it can 8817 // be called from the debugger. Compile it into the product 8818 // as needed. 8819 bool debug_verify_chunk_in_free_list(FreeChunk* fc) { 8820 return debug_cms_space->verify_chunk_in_free_list(fc); 8821 } 8822 #endif 8823 8824 void SweepClosure::print_free_block_coalesced(FreeChunk* fc) const { 8825 if (CMSTraceSweeper) { 8826 gclog_or_tty->print_cr("Sweep:coal_free_blk " PTR_FORMAT " (" SIZE_FORMAT ")", 8827 fc, fc->size()); 8828 } 8829 } 8830 8831 // CMSIsAliveClosure 8832 bool CMSIsAliveClosure::do_object_b(oop obj) { 8833 HeapWord* addr = (HeapWord*)obj; 8834 return addr != NULL && 8835 (!_span.contains(addr) || _bit_map->isMarked(addr)); 8836 } 8837 8838 8839 CMSKeepAliveClosure::CMSKeepAliveClosure( CMSCollector* collector, 8840 MemRegion span, 8841 CMSBitMap* bit_map, CMSMarkStack* mark_stack, 8842 bool cpc): 8843 _collector(collector), 8844 _span(span), 8845 _bit_map(bit_map), 8846 _mark_stack(mark_stack), 8847 _concurrent_precleaning(cpc) { 8848 assert(!_span.is_empty(), "Empty span could spell trouble"); 8849 } 8850 8851 8852 // CMSKeepAliveClosure: the serial version 8853 void CMSKeepAliveClosure::do_oop(oop obj) { 8854 HeapWord* addr = (HeapWord*)obj; 8855 if (_span.contains(addr) && 8856 !_bit_map->isMarked(addr)) { 8857 _bit_map->mark(addr); 8858 bool simulate_overflow = false; 8859 NOT_PRODUCT( 8860 if (CMSMarkStackOverflowALot && 8861 _collector->simulate_overflow()) { 8862 // simulate a stack overflow 8863 simulate_overflow = true; 8864 } 8865 ) 8866 if (simulate_overflow || !_mark_stack->push(obj)) { 8867 if (_concurrent_precleaning) { 8868 // We dirty the overflown object and let the remark 8869 // phase deal with it. 8870 assert(_collector->overflow_list_is_empty(), "Error"); 8871 // In the case of object arrays, we need to dirty all of 8872 // the cards that the object spans. No locking or atomics 8873 // are needed since no one else can be mutating the mod union 8874 // table. 8875 if (obj->is_objArray()) { 8876 size_t sz = obj->size(); 8877 HeapWord* end_card_addr = 8878 (HeapWord*)round_to((intptr_t)(addr+sz), CardTableModRefBS::card_size); 8879 MemRegion redirty_range = MemRegion(addr, end_card_addr); 8880 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 8881 _collector->_modUnionTable.mark_range(redirty_range); 8882 } else { 8883 _collector->_modUnionTable.mark(addr); 8884 } 8885 _collector->_ser_kac_preclean_ovflw++; 8886 } else { 8887 _collector->push_on_overflow_list(obj); 8888 _collector->_ser_kac_ovflw++; 8889 } 8890 } 8891 } 8892 } 8893 8894 void CMSKeepAliveClosure::do_oop(oop* p) { CMSKeepAliveClosure::do_oop_work(p); } 8895 void CMSKeepAliveClosure::do_oop(narrowOop* p) { CMSKeepAliveClosure::do_oop_work(p); } 8896 8897 // CMSParKeepAliveClosure: a parallel version of the above. 8898 // The work queues are private to each closure (thread), 8899 // but (may be) available for stealing by other threads. 8900 void CMSParKeepAliveClosure::do_oop(oop obj) { 8901 HeapWord* addr = (HeapWord*)obj; 8902 if (_span.contains(addr) && 8903 !_bit_map->isMarked(addr)) { 8904 // In general, during recursive tracing, several threads 8905 // may be concurrently getting here; the first one to 8906 // "tag" it, claims it. 8907 if (_bit_map->par_mark(addr)) { 8908 bool res = _work_queue->push(obj); 8909 assert(res, "Low water mark should be much less than capacity"); 8910 // Do a recursive trim in the hope that this will keep 8911 // stack usage lower, but leave some oops for potential stealers 8912 trim_queue(_low_water_mark); 8913 } // Else, another thread got there first 8914 } 8915 } 8916 8917 void CMSParKeepAliveClosure::do_oop(oop* p) { CMSParKeepAliveClosure::do_oop_work(p); } 8918 void CMSParKeepAliveClosure::do_oop(narrowOop* p) { CMSParKeepAliveClosure::do_oop_work(p); } 8919 8920 void CMSParKeepAliveClosure::trim_queue(uint max) { 8921 while (_work_queue->size() > max) { 8922 oop new_oop; 8923 if (_work_queue->pop_local(new_oop)) { 8924 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 8925 assert(_bit_map->isMarked((HeapWord*)new_oop), 8926 "no white objects on this stack!"); 8927 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 8928 // iterate over the oops in this oop, marking and pushing 8929 // the ones in CMS heap (i.e. in _span). 8930 new_oop->oop_iterate(&_mark_and_push); 8931 } 8932 } 8933 } 8934 8935 CMSInnerParMarkAndPushClosure::CMSInnerParMarkAndPushClosure( 8936 CMSCollector* collector, 8937 MemRegion span, CMSBitMap* bit_map, 8938 OopTaskQueue* work_queue): 8939 _collector(collector), 8940 _span(span), 8941 _bit_map(bit_map), 8942 _work_queue(work_queue) { } 8943 8944 void CMSInnerParMarkAndPushClosure::do_oop(oop obj) { 8945 HeapWord* addr = (HeapWord*)obj; 8946 if (_span.contains(addr) && 8947 !_bit_map->isMarked(addr)) { 8948 if (_bit_map->par_mark(addr)) { 8949 bool simulate_overflow = false; 8950 NOT_PRODUCT( 8951 if (CMSMarkStackOverflowALot && 8952 _collector->par_simulate_overflow()) { 8953 // simulate a stack overflow 8954 simulate_overflow = true; 8955 } 8956 ) 8957 if (simulate_overflow || !_work_queue->push(obj)) { 8958 _collector->par_push_on_overflow_list(obj); 8959 _collector->_par_kac_ovflw++; 8960 } 8961 } // Else another thread got there already 8962 } 8963 } 8964 8965 void CMSInnerParMarkAndPushClosure::do_oop(oop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); } 8966 void CMSInnerParMarkAndPushClosure::do_oop(narrowOop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); } 8967 8968 ////////////////////////////////////////////////////////////////// 8969 // CMSExpansionCause ///////////////////////////// 8970 ////////////////////////////////////////////////////////////////// 8971 const char* CMSExpansionCause::to_string(CMSExpansionCause::Cause cause) { 8972 switch (cause) { 8973 case _no_expansion: 8974 return "No expansion"; 8975 case _satisfy_free_ratio: 8976 return "Free ratio"; 8977 case _satisfy_promotion: 8978 return "Satisfy promotion"; 8979 case _satisfy_allocation: 8980 return "allocation"; 8981 case _allocate_par_lab: 8982 return "Par LAB"; 8983 case _allocate_par_spooling_space: 8984 return "Par Spooling Space"; 8985 case _adaptive_size_policy: 8986 return "Ergonomics"; 8987 default: 8988 return "unknown"; 8989 } 8990 } 8991 8992 void CMSDrainMarkingStackClosure::do_void() { 8993 // the max number to take from overflow list at a time 8994 const size_t num = _mark_stack->capacity()/4; 8995 assert(!_concurrent_precleaning || _collector->overflow_list_is_empty(), 8996 "Overflow list should be NULL during concurrent phases"); 8997 while (!_mark_stack->isEmpty() || 8998 // if stack is empty, check the overflow list 8999 _collector->take_from_overflow_list(num, _mark_stack)) { 9000 oop obj = _mark_stack->pop(); 9001 HeapWord* addr = (HeapWord*)obj; 9002 assert(_span.contains(addr), "Should be within span"); 9003 assert(_bit_map->isMarked(addr), "Should be marked"); 9004 assert(obj->is_oop(), "Should be an oop"); 9005 obj->oop_iterate(_keep_alive); 9006 } 9007 } 9008 9009 void CMSParDrainMarkingStackClosure::do_void() { 9010 // drain queue 9011 trim_queue(0); 9012 } 9013 9014 // Trim our work_queue so its length is below max at return 9015 void CMSParDrainMarkingStackClosure::trim_queue(uint max) { 9016 while (_work_queue->size() > max) { 9017 oop new_oop; 9018 if (_work_queue->pop_local(new_oop)) { 9019 assert(new_oop->is_oop(), "Expected an oop"); 9020 assert(_bit_map->isMarked((HeapWord*)new_oop), 9021 "no white objects on this stack!"); 9022 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 9023 // iterate over the oops in this oop, marking and pushing 9024 // the ones in CMS heap (i.e. in _span). 9025 new_oop->oop_iterate(&_mark_and_push); 9026 } 9027 } 9028 } 9029 9030 //////////////////////////////////////////////////////////////////// 9031 // Support for Marking Stack Overflow list handling and related code 9032 //////////////////////////////////////////////////////////////////// 9033 // Much of the following code is similar in shape and spirit to the 9034 // code used in ParNewGC. We should try and share that code 9035 // as much as possible in the future. 9036 9037 #ifndef PRODUCT 9038 // Debugging support for CMSStackOverflowALot 9039 9040 // It's OK to call this multi-threaded; the worst thing 9041 // that can happen is that we'll get a bunch of closely 9042 // spaced simulated oveflows, but that's OK, in fact 9043 // probably good as it would exercise the overflow code 9044 // under contention. 9045 bool CMSCollector::simulate_overflow() { 9046 if (_overflow_counter-- <= 0) { // just being defensive 9047 _overflow_counter = CMSMarkStackOverflowInterval; 9048 return true; 9049 } else { 9050 return false; 9051 } 9052 } 9053 9054 bool CMSCollector::par_simulate_overflow() { 9055 return simulate_overflow(); 9056 } 9057 #endif 9058 9059 // Single-threaded 9060 bool CMSCollector::take_from_overflow_list(size_t num, CMSMarkStack* stack) { 9061 assert(stack->isEmpty(), "Expected precondition"); 9062 assert(stack->capacity() > num, "Shouldn't bite more than can chew"); 9063 size_t i = num; 9064 oop cur = _overflow_list; 9065 const markOop proto = markOopDesc::prototype(); 9066 NOT_PRODUCT(ssize_t n = 0;) 9067 for (oop next; i > 0 && cur != NULL; cur = next, i--) { 9068 next = oop(cur->mark()); 9069 cur->set_mark(proto); // until proven otherwise 9070 assert(cur->is_oop(), "Should be an oop"); 9071 bool res = stack->push(cur); 9072 assert(res, "Bit off more than can chew?"); 9073 NOT_PRODUCT(n++;) 9074 } 9075 _overflow_list = cur; 9076 #ifndef PRODUCT 9077 assert(_num_par_pushes >= n, "Too many pops?"); 9078 _num_par_pushes -=n; 9079 #endif 9080 return !stack->isEmpty(); 9081 } 9082 9083 #define BUSY (cast_to_oop<intptr_t>(0x1aff1aff)) 9084 // (MT-safe) Get a prefix of at most "num" from the list. 9085 // The overflow list is chained through the mark word of 9086 // each object in the list. We fetch the entire list, 9087 // break off a prefix of the right size and return the 9088 // remainder. If other threads try to take objects from 9089 // the overflow list at that time, they will wait for 9090 // some time to see if data becomes available. If (and 9091 // only if) another thread places one or more object(s) 9092 // on the global list before we have returned the suffix 9093 // to the global list, we will walk down our local list 9094 // to find its end and append the global list to 9095 // our suffix before returning it. This suffix walk can 9096 // prove to be expensive (quadratic in the amount of traffic) 9097 // when there are many objects in the overflow list and 9098 // there is much producer-consumer contention on the list. 9099 // *NOTE*: The overflow list manipulation code here and 9100 // in ParNewGeneration:: are very similar in shape, 9101 // except that in the ParNew case we use the old (from/eden) 9102 // copy of the object to thread the list via its klass word. 9103 // Because of the common code, if you make any changes in 9104 // the code below, please check the ParNew version to see if 9105 // similar changes might be needed. 9106 // CR 6797058 has been filed to consolidate the common code. 9107 bool CMSCollector::par_take_from_overflow_list(size_t num, 9108 OopTaskQueue* work_q, 9109 int no_of_gc_threads) { 9110 assert(work_q->size() == 0, "First empty local work queue"); 9111 assert(num < work_q->max_elems(), "Can't bite more than we can chew"); 9112 if (_overflow_list == NULL) { 9113 return false; 9114 } 9115 // Grab the entire list; we'll put back a suffix 9116 oop prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list)); 9117 Thread* tid = Thread::current(); 9118 // Before "no_of_gc_threads" was introduced CMSOverflowSpinCount was 9119 // set to ParallelGCThreads. 9120 size_t CMSOverflowSpinCount = (size_t) no_of_gc_threads; // was ParallelGCThreads; 9121 size_t sleep_time_millis = MAX2((size_t)1, num/100); 9122 // If the list is busy, we spin for a short while, 9123 // sleeping between attempts to get the list. 9124 for (size_t spin = 0; prefix == BUSY && spin < CMSOverflowSpinCount; spin++) { 9125 os::sleep(tid, sleep_time_millis, false); 9126 if (_overflow_list == NULL) { 9127 // Nothing left to take 9128 return false; 9129 } else if (_overflow_list != BUSY) { 9130 // Try and grab the prefix 9131 prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list)); 9132 } 9133 } 9134 // If the list was found to be empty, or we spun long 9135 // enough, we give up and return empty-handed. If we leave 9136 // the list in the BUSY state below, it must be the case that 9137 // some other thread holds the overflow list and will set it 9138 // to a non-BUSY state in the future. 9139 if (prefix == NULL || prefix == BUSY) { 9140 // Nothing to take or waited long enough 9141 if (prefix == NULL) { 9142 // Write back the NULL in case we overwrote it with BUSY above 9143 // and it is still the same value. 9144 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY); 9145 } 9146 return false; 9147 } 9148 assert(prefix != NULL && prefix != BUSY, "Error"); 9149 size_t i = num; 9150 oop cur = prefix; 9151 // Walk down the first "num" objects, unless we reach the end. 9152 for (; i > 1 && cur->mark() != NULL; cur = oop(cur->mark()), i--); 9153 if (cur->mark() == NULL) { 9154 // We have "num" or fewer elements in the list, so there 9155 // is nothing to return to the global list. 9156 // Write back the NULL in lieu of the BUSY we wrote 9157 // above, if it is still the same value. 9158 if (_overflow_list == BUSY) { 9159 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY); 9160 } 9161 } else { 9162 // Chop off the suffix and rerturn it to the global list. 9163 assert(cur->mark() != BUSY, "Error"); 9164 oop suffix_head = cur->mark(); // suffix will be put back on global list 9165 cur->set_mark(NULL); // break off suffix 9166 // It's possible that the list is still in the empty(busy) state 9167 // we left it in a short while ago; in that case we may be 9168 // able to place back the suffix without incurring the cost 9169 // of a walk down the list. 9170 oop observed_overflow_list = _overflow_list; 9171 oop cur_overflow_list = observed_overflow_list; 9172 bool attached = false; 9173 while (observed_overflow_list == BUSY || observed_overflow_list == NULL) { 9174 observed_overflow_list = 9175 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list); 9176 if (cur_overflow_list == observed_overflow_list) { 9177 attached = true; 9178 break; 9179 } else cur_overflow_list = observed_overflow_list; 9180 } 9181 if (!attached) { 9182 // Too bad, someone else sneaked in (at least) an element; we'll need 9183 // to do a splice. Find tail of suffix so we can prepend suffix to global 9184 // list. 9185 for (cur = suffix_head; cur->mark() != NULL; cur = (oop)(cur->mark())); 9186 oop suffix_tail = cur; 9187 assert(suffix_tail != NULL && suffix_tail->mark() == NULL, 9188 "Tautology"); 9189 observed_overflow_list = _overflow_list; 9190 do { 9191 cur_overflow_list = observed_overflow_list; 9192 if (cur_overflow_list != BUSY) { 9193 // Do the splice ... 9194 suffix_tail->set_mark(markOop(cur_overflow_list)); 9195 } else { // cur_overflow_list == BUSY 9196 suffix_tail->set_mark(NULL); 9197 } 9198 // ... and try to place spliced list back on overflow_list ... 9199 observed_overflow_list = 9200 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list); 9201 } while (cur_overflow_list != observed_overflow_list); 9202 // ... until we have succeeded in doing so. 9203 } 9204 } 9205 9206 // Push the prefix elements on work_q 9207 assert(prefix != NULL, "control point invariant"); 9208 const markOop proto = markOopDesc::prototype(); 9209 oop next; 9210 NOT_PRODUCT(ssize_t n = 0;) 9211 for (cur = prefix; cur != NULL; cur = next) { 9212 next = oop(cur->mark()); 9213 cur->set_mark(proto); // until proven otherwise 9214 assert(cur->is_oop(), "Should be an oop"); 9215 bool res = work_q->push(cur); 9216 assert(res, "Bit off more than we can chew?"); 9217 NOT_PRODUCT(n++;) 9218 } 9219 #ifndef PRODUCT 9220 assert(_num_par_pushes >= n, "Too many pops?"); 9221 Atomic::add_ptr(-(intptr_t)n, &_num_par_pushes); 9222 #endif 9223 return true; 9224 } 9225 9226 // Single-threaded 9227 void CMSCollector::push_on_overflow_list(oop p) { 9228 NOT_PRODUCT(_num_par_pushes++;) 9229 assert(p->is_oop(), "Not an oop"); 9230 preserve_mark_if_necessary(p); 9231 p->set_mark((markOop)_overflow_list); 9232 _overflow_list = p; 9233 } 9234 9235 // Multi-threaded; use CAS to prepend to overflow list 9236 void CMSCollector::par_push_on_overflow_list(oop p) { 9237 NOT_PRODUCT(Atomic::inc_ptr(&_num_par_pushes);) 9238 assert(p->is_oop(), "Not an oop"); 9239 par_preserve_mark_if_necessary(p); 9240 oop observed_overflow_list = _overflow_list; 9241 oop cur_overflow_list; 9242 do { 9243 cur_overflow_list = observed_overflow_list; 9244 if (cur_overflow_list != BUSY) { 9245 p->set_mark(markOop(cur_overflow_list)); 9246 } else { 9247 p->set_mark(NULL); 9248 } 9249 observed_overflow_list = 9250 (oop) Atomic::cmpxchg_ptr(p, &_overflow_list, cur_overflow_list); 9251 } while (cur_overflow_list != observed_overflow_list); 9252 } 9253 #undef BUSY 9254 9255 // Single threaded 9256 // General Note on GrowableArray: pushes may silently fail 9257 // because we are (temporarily) out of C-heap for expanding 9258 // the stack. The problem is quite ubiquitous and affects 9259 // a lot of code in the JVM. The prudent thing for GrowableArray 9260 // to do (for now) is to exit with an error. However, that may 9261 // be too draconian in some cases because the caller may be 9262 // able to recover without much harm. For such cases, we 9263 // should probably introduce a "soft_push" method which returns 9264 // an indication of success or failure with the assumption that 9265 // the caller may be able to recover from a failure; code in 9266 // the VM can then be changed, incrementally, to deal with such 9267 // failures where possible, thus, incrementally hardening the VM 9268 // in such low resource situations. 9269 void CMSCollector::preserve_mark_work(oop p, markOop m) { 9270 _preserved_oop_stack.push(p); 9271 _preserved_mark_stack.push(m); 9272 assert(m == p->mark(), "Mark word changed"); 9273 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), 9274 "bijection"); 9275 } 9276 9277 // Single threaded 9278 void CMSCollector::preserve_mark_if_necessary(oop p) { 9279 markOop m = p->mark(); 9280 if (m->must_be_preserved(p)) { 9281 preserve_mark_work(p, m); 9282 } 9283 } 9284 9285 void CMSCollector::par_preserve_mark_if_necessary(oop p) { 9286 markOop m = p->mark(); 9287 if (m->must_be_preserved(p)) { 9288 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag); 9289 // Even though we read the mark word without holding 9290 // the lock, we are assured that it will not change 9291 // because we "own" this oop, so no other thread can 9292 // be trying to push it on the overflow list; see 9293 // the assertion in preserve_mark_work() that checks 9294 // that m == p->mark(). 9295 preserve_mark_work(p, m); 9296 } 9297 } 9298 9299 // We should be able to do this multi-threaded, 9300 // a chunk of stack being a task (this is 9301 // correct because each oop only ever appears 9302 // once in the overflow list. However, it's 9303 // not very easy to completely overlap this with 9304 // other operations, so will generally not be done 9305 // until all work's been completed. Because we 9306 // expect the preserved oop stack (set) to be small, 9307 // it's probably fine to do this single-threaded. 9308 // We can explore cleverer concurrent/overlapped/parallel 9309 // processing of preserved marks if we feel the 9310 // need for this in the future. Stack overflow should 9311 // be so rare in practice and, when it happens, its 9312 // effect on performance so great that this will 9313 // likely just be in the noise anyway. 9314 void CMSCollector::restore_preserved_marks_if_any() { 9315 assert(SafepointSynchronize::is_at_safepoint(), 9316 "world should be stopped"); 9317 assert(Thread::current()->is_ConcurrentGC_thread() || 9318 Thread::current()->is_VM_thread(), 9319 "should be single-threaded"); 9320 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), 9321 "bijection"); 9322 9323 while (!_preserved_oop_stack.is_empty()) { 9324 oop p = _preserved_oop_stack.pop(); 9325 assert(p->is_oop(), "Should be an oop"); 9326 assert(_span.contains(p), "oop should be in _span"); 9327 assert(p->mark() == markOopDesc::prototype(), 9328 "Set when taken from overflow list"); 9329 markOop m = _preserved_mark_stack.pop(); 9330 p->set_mark(m); 9331 } 9332 assert(_preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(), 9333 "stacks were cleared above"); 9334 } 9335 9336 #ifndef PRODUCT 9337 bool CMSCollector::no_preserved_marks() const { 9338 return _preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(); 9339 } 9340 #endif 9341 9342 CMSAdaptiveSizePolicy* ASConcurrentMarkSweepGeneration::cms_size_policy() const 9343 { 9344 GenCollectedHeap* gch = (GenCollectedHeap*) GenCollectedHeap::heap(); 9345 CMSAdaptiveSizePolicy* size_policy = 9346 (CMSAdaptiveSizePolicy*) gch->gen_policy()->size_policy(); 9347 assert(size_policy->is_gc_cms_adaptive_size_policy(), 9348 "Wrong type for size policy"); 9349 return size_policy; 9350 } 9351 9352 void ASConcurrentMarkSweepGeneration::resize(size_t cur_promo_size, 9353 size_t desired_promo_size) { 9354 if (cur_promo_size < desired_promo_size) { 9355 size_t expand_bytes = desired_promo_size - cur_promo_size; 9356 if (PrintAdaptiveSizePolicy && Verbose) { 9357 gclog_or_tty->print_cr(" ASConcurrentMarkSweepGeneration::resize " 9358 "Expanding tenured generation by " SIZE_FORMAT " (bytes)", 9359 expand_bytes); 9360 } 9361 expand(expand_bytes, 9362 MinHeapDeltaBytes, 9363 CMSExpansionCause::_adaptive_size_policy); 9364 } else if (desired_promo_size < cur_promo_size) { 9365 size_t shrink_bytes = cur_promo_size - desired_promo_size; 9366 if (PrintAdaptiveSizePolicy && Verbose) { 9367 gclog_or_tty->print_cr(" ASConcurrentMarkSweepGeneration::resize " 9368 "Shrinking tenured generation by " SIZE_FORMAT " (bytes)", 9369 shrink_bytes); 9370 } 9371 shrink(shrink_bytes); 9372 } 9373 } 9374 9375 CMSGCAdaptivePolicyCounters* ASConcurrentMarkSweepGeneration::gc_adaptive_policy_counters() { 9376 GenCollectedHeap* gch = GenCollectedHeap::heap(); 9377 CMSGCAdaptivePolicyCounters* counters = 9378 (CMSGCAdaptivePolicyCounters*) gch->collector_policy()->counters(); 9379 assert(counters->kind() == GCPolicyCounters::CMSGCAdaptivePolicyCountersKind, 9380 "Wrong kind of counters"); 9381 return counters; 9382 } 9383 9384 9385 void ASConcurrentMarkSweepGeneration::update_counters() { 9386 if (UsePerfData) { 9387 _space_counters->update_all(); 9388 _gen_counters->update_all(); 9389 CMSGCAdaptivePolicyCounters* counters = gc_adaptive_policy_counters(); 9390 GenCollectedHeap* gch = GenCollectedHeap::heap(); 9391 CMSGCStats* gc_stats_l = (CMSGCStats*) gc_stats(); 9392 assert(gc_stats_l->kind() == GCStats::CMSGCStatsKind, 9393 "Wrong gc statistics type"); 9394 counters->update_counters(gc_stats_l); 9395 } 9396 } 9397 9398 void ASConcurrentMarkSweepGeneration::update_counters(size_t used) { 9399 if (UsePerfData) { 9400 _space_counters->update_used(used); 9401 _space_counters->update_capacity(); 9402 _gen_counters->update_all(); 9403 9404 CMSGCAdaptivePolicyCounters* counters = gc_adaptive_policy_counters(); 9405 GenCollectedHeap* gch = GenCollectedHeap::heap(); 9406 CMSGCStats* gc_stats_l = (CMSGCStats*) gc_stats(); 9407 assert(gc_stats_l->kind() == GCStats::CMSGCStatsKind, 9408 "Wrong gc statistics type"); 9409 counters->update_counters(gc_stats_l); 9410 } 9411 } 9412 9413 void ASConcurrentMarkSweepGeneration::shrink_by(size_t desired_bytes) { 9414 assert_locked_or_safepoint(Heap_lock); 9415 assert_lock_strong(freelistLock()); 9416 HeapWord* old_end = _cmsSpace->end(); 9417 HeapWord* unallocated_start = _cmsSpace->unallocated_block(); 9418 assert(old_end >= unallocated_start, "Miscalculation of unallocated_start"); 9419 FreeChunk* chunk_at_end = find_chunk_at_end(); 9420 if (chunk_at_end == NULL) { 9421 // No room to shrink 9422 if (PrintGCDetails && Verbose) { 9423 gclog_or_tty->print_cr("No room to shrink: old_end " 9424 PTR_FORMAT " unallocated_start " PTR_FORMAT 9425 " chunk_at_end " PTR_FORMAT, 9426 old_end, unallocated_start, chunk_at_end); 9427 } 9428 return; 9429 } else { 9430 9431 // Find the chunk at the end of the space and determine 9432 // how much it can be shrunk. 9433 size_t shrinkable_size_in_bytes = chunk_at_end->size(); 9434 size_t aligned_shrinkable_size_in_bytes = 9435 align_size_down(shrinkable_size_in_bytes, os::vm_page_size()); 9436 assert(unallocated_start <= (HeapWord*) chunk_at_end->end(), 9437 "Inconsistent chunk at end of space"); 9438 size_t bytes = MIN2(desired_bytes, aligned_shrinkable_size_in_bytes); 9439 size_t word_size_before = heap_word_size(_virtual_space.committed_size()); 9440 9441 // Shrink the underlying space 9442 _virtual_space.shrink_by(bytes); 9443 if (PrintGCDetails && Verbose) { 9444 gclog_or_tty->print_cr("ConcurrentMarkSweepGeneration::shrink_by:" 9445 " desired_bytes " SIZE_FORMAT 9446 " shrinkable_size_in_bytes " SIZE_FORMAT 9447 " aligned_shrinkable_size_in_bytes " SIZE_FORMAT 9448 " bytes " SIZE_FORMAT, 9449 desired_bytes, shrinkable_size_in_bytes, 9450 aligned_shrinkable_size_in_bytes, bytes); 9451 gclog_or_tty->print_cr(" old_end " SIZE_FORMAT 9452 " unallocated_start " SIZE_FORMAT, 9453 old_end, unallocated_start); 9454 } 9455 9456 // If the space did shrink (shrinking is not guaranteed), 9457 // shrink the chunk at the end by the appropriate amount. 9458 if (((HeapWord*)_virtual_space.high()) < old_end) { 9459 size_t new_word_size = 9460 heap_word_size(_virtual_space.committed_size()); 9461 9462 // Have to remove the chunk from the dictionary because it is changing 9463 // size and might be someplace elsewhere in the dictionary. 9464 9465 // Get the chunk at end, shrink it, and put it 9466 // back. 9467 _cmsSpace->removeChunkFromDictionary(chunk_at_end); 9468 size_t word_size_change = word_size_before - new_word_size; 9469 size_t chunk_at_end_old_size = chunk_at_end->size(); 9470 assert(chunk_at_end_old_size >= word_size_change, 9471 "Shrink is too large"); 9472 chunk_at_end->set_size(chunk_at_end_old_size - 9473 word_size_change); 9474 _cmsSpace->freed((HeapWord*) chunk_at_end->end(), 9475 word_size_change); 9476 9477 _cmsSpace->returnChunkToDictionary(chunk_at_end); 9478 9479 MemRegion mr(_cmsSpace->bottom(), new_word_size); 9480 _bts->resize(new_word_size); // resize the block offset shared array 9481 Universe::heap()->barrier_set()->resize_covered_region(mr); 9482 _cmsSpace->assert_locked(); 9483 _cmsSpace->set_end((HeapWord*)_virtual_space.high()); 9484 9485 NOT_PRODUCT(_cmsSpace->dictionary()->verify()); 9486 9487 // update the space and generation capacity counters 9488 if (UsePerfData) { 9489 _space_counters->update_capacity(); 9490 _gen_counters->update_all(); 9491 } 9492 9493 if (Verbose && PrintGCDetails) { 9494 size_t new_mem_size = _virtual_space.committed_size(); 9495 size_t old_mem_size = new_mem_size + bytes; 9496 gclog_or_tty->print_cr("Shrinking %s from " SIZE_FORMAT "K by " SIZE_FORMAT "K to " SIZE_FORMAT "K", 9497 name(), old_mem_size/K, bytes/K, new_mem_size/K); 9498 } 9499 } 9500 9501 assert(_cmsSpace->unallocated_block() <= _cmsSpace->end(), 9502 "Inconsistency at end of space"); 9503 assert(chunk_at_end->end() == (uintptr_t*) _cmsSpace->end(), 9504 "Shrinking is inconsistent"); 9505 return; 9506 } 9507 } 9508 // Transfer some number of overflown objects to usual marking 9509 // stack. Return true if some objects were transferred. 9510 bool MarkRefsIntoAndScanClosure::take_from_overflow_list() { 9511 size_t num = MIN2((size_t)(_mark_stack->capacity() - _mark_stack->length())/4, 9512 (size_t)ParGCDesiredObjsFromOverflowList); 9513 9514 bool res = _collector->take_from_overflow_list(num, _mark_stack); 9515 assert(_collector->overflow_list_is_empty() || res, 9516 "If list is not empty, we should have taken something"); 9517 assert(!res || !_mark_stack->isEmpty(), 9518 "If we took something, it should now be on our stack"); 9519 return res; 9520 } 9521 9522 size_t MarkDeadObjectsClosure::do_blk(HeapWord* addr) { 9523 size_t res = _sp->block_size_no_stall(addr, _collector); 9524 if (_sp->block_is_obj(addr)) { 9525 if (_live_bit_map->isMarked(addr)) { 9526 // It can't have been dead in a previous cycle 9527 guarantee(!_dead_bit_map->isMarked(addr), "No resurrection!"); 9528 } else { 9529 _dead_bit_map->mark(addr); // mark the dead object 9530 } 9531 } 9532 // Could be 0, if the block size could not be computed without stalling. 9533 return res; 9534 } 9535 9536 TraceCMSMemoryManagerStats::TraceCMSMemoryManagerStats(CMSCollector::CollectorState phase, GCCause::Cause cause): TraceMemoryManagerStats() { 9537 9538 switch (phase) { 9539 case CMSCollector::InitialMarking: 9540 initialize(true /* fullGC */ , 9541 cause /* cause of the GC */, 9542 true /* recordGCBeginTime */, 9543 true /* recordPreGCUsage */, 9544 false /* recordPeakUsage */, 9545 false /* recordPostGCusage */, 9546 true /* recordAccumulatedGCTime */, 9547 false /* recordGCEndTime */, 9548 false /* countCollection */ ); 9549 break; 9550 9551 case CMSCollector::FinalMarking: 9552 initialize(true /* fullGC */ , 9553 cause /* cause of the GC */, 9554 false /* recordGCBeginTime */, 9555 false /* recordPreGCUsage */, 9556 false /* recordPeakUsage */, 9557 false /* recordPostGCusage */, 9558 true /* recordAccumulatedGCTime */, 9559 false /* recordGCEndTime */, 9560 false /* countCollection */ ); 9561 break; 9562 9563 case CMSCollector::Sweeping: 9564 initialize(true /* fullGC */ , 9565 cause /* cause of the GC */, 9566 false /* recordGCBeginTime */, 9567 false /* recordPreGCUsage */, 9568 true /* recordPeakUsage */, 9569 true /* recordPostGCusage */, 9570 false /* recordAccumulatedGCTime */, 9571 true /* recordGCEndTime */, 9572 true /* countCollection */ ); 9573 break; 9574 9575 default: 9576 ShouldNotReachHere(); 9577 } 9578 }