1 /* 2 * Copyright (c) 2001, 2010, 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 // ConcurrentMarkSweepGeneration is in support of a concurrent 26 // mark-sweep old generation in the Detlefs-Printezis--Boehm-Demers-Schenker 27 // style. We assume, for now, that this generation is always the 28 // seniormost generation (modulo the PermGeneration), and for simplicity 29 // in the first implementation, that this generation is a single compactible 30 // space. Neither of these restrictions appears essential, and will be 31 // relaxed in the future when more time is available to implement the 32 // greater generality (and there's a need for it). 33 // 34 // Concurrent mode failures are currently handled by 35 // means of a sliding mark-compact. 36 37 class CMSAdaptiveSizePolicy; 38 class CMSConcMarkingTask; 39 class CMSGCAdaptivePolicyCounters; 40 class ConcurrentMarkSweepGeneration; 41 class ConcurrentMarkSweepPolicy; 42 class ConcurrentMarkSweepThread; 43 class CompactibleFreeListSpace; 44 class FreeChunk; 45 class PromotionInfo; 46 class ScanMarkedObjectsAgainCarefullyClosure; 47 48 // A generic CMS bit map. It's the basis for both the CMS marking bit map 49 // as well as for the mod union table (in each case only a subset of the 50 // methods are used). This is essentially a wrapper around the BitMap class, 51 // with one bit per (1<<_shifter) HeapWords. (i.e. for the marking bit map, 52 // we have _shifter == 0. and for the mod union table we have 53 // shifter == CardTableModRefBS::card_shift - LogHeapWordSize.) 54 // XXX 64-bit issues in BitMap? 55 class CMSBitMap VALUE_OBJ_CLASS_SPEC { 56 friend class VMStructs; 57 58 HeapWord* _bmStartWord; // base address of range covered by map 59 size_t _bmWordSize; // map size (in #HeapWords covered) 60 const int _shifter; // shifts to convert HeapWord to bit position 61 VirtualSpace _virtual_space; // underlying the bit map 62 BitMap _bm; // the bit map itself 63 public: 64 Mutex* const _lock; // mutex protecting _bm; 65 66 public: 67 // constructor 68 CMSBitMap(int shifter, int mutex_rank, const char* mutex_name); 69 70 // allocates the actual storage for the map 71 bool allocate(MemRegion mr); 72 // field getter 73 Mutex* lock() const { return _lock; } 74 // locking verifier convenience function 75 void assert_locked() const PRODUCT_RETURN; 76 77 // inquiries 78 HeapWord* startWord() const { return _bmStartWord; } 79 size_t sizeInWords() const { return _bmWordSize; } 80 size_t sizeInBits() const { return _bm.size(); } 81 // the following is one past the last word in space 82 HeapWord* endWord() const { return _bmStartWord + _bmWordSize; } 83 84 // reading marks 85 bool isMarked(HeapWord* addr) const; 86 bool par_isMarked(HeapWord* addr) const; // do not lock checks 87 bool isUnmarked(HeapWord* addr) const; 88 bool isAllClear() const; 89 90 // writing marks 91 void mark(HeapWord* addr); 92 // For marking by parallel GC threads; 93 // returns true if we did, false if another thread did 94 bool par_mark(HeapWord* addr); 95 96 void mark_range(MemRegion mr); 97 void par_mark_range(MemRegion mr); 98 void mark_large_range(MemRegion mr); 99 void par_mark_large_range(MemRegion mr); 100 void par_clear(HeapWord* addr); // For unmarking by parallel GC threads. 101 void clear_range(MemRegion mr); 102 void par_clear_range(MemRegion mr); 103 void clear_large_range(MemRegion mr); 104 void par_clear_large_range(MemRegion mr); 105 void clear_all(); 106 void clear_all_incrementally(); // Not yet implemented!! 107 108 NOT_PRODUCT( 109 // checks the memory region for validity 110 void region_invariant(MemRegion mr); 111 ) 112 113 // iteration 114 void iterate(BitMapClosure* cl) { 115 _bm.iterate(cl); 116 } 117 void iterate(BitMapClosure* cl, HeapWord* left, HeapWord* right); 118 void dirty_range_iterate_clear(MemRegionClosure* cl); 119 void dirty_range_iterate_clear(MemRegion mr, MemRegionClosure* cl); 120 121 // auxiliary support for iteration 122 HeapWord* getNextMarkedWordAddress(HeapWord* addr) const; 123 HeapWord* getNextMarkedWordAddress(HeapWord* start_addr, 124 HeapWord* end_addr) const; 125 HeapWord* getNextUnmarkedWordAddress(HeapWord* addr) const; 126 HeapWord* getNextUnmarkedWordAddress(HeapWord* start_addr, 127 HeapWord* end_addr) const; 128 MemRegion getAndClearMarkedRegion(HeapWord* addr); 129 MemRegion getAndClearMarkedRegion(HeapWord* start_addr, 130 HeapWord* end_addr); 131 132 // conversion utilities 133 HeapWord* offsetToHeapWord(size_t offset) const; 134 size_t heapWordToOffset(HeapWord* addr) const; 135 size_t heapWordDiffToOffsetDiff(size_t diff) const; 136 137 // debugging 138 // is this address range covered by the bit-map? 139 NOT_PRODUCT( 140 bool covers(MemRegion mr) const; 141 bool covers(HeapWord* start, size_t size = 0) const; 142 ) 143 void verifyNoOneBitsInRange(HeapWord* left, HeapWord* right) PRODUCT_RETURN; 144 }; 145 146 // Represents a marking stack used by the CMS collector. 147 // Ideally this should be GrowableArray<> just like MSC's marking stack(s). 148 class CMSMarkStack: public CHeapObj { 149 // 150 friend class CMSCollector; // to get at expasion stats further below 151 // 152 153 VirtualSpace _virtual_space; // space for the stack 154 oop* _base; // bottom of stack 155 size_t _index; // one more than last occupied index 156 size_t _capacity; // max #elements 157 Mutex _par_lock; // an advisory lock used in case of parallel access 158 NOT_PRODUCT(size_t _max_depth;) // max depth plumbed during run 159 160 protected: 161 size_t _hit_limit; // we hit max stack size limit 162 size_t _failed_double; // we failed expansion before hitting limit 163 164 public: 165 CMSMarkStack(): 166 _par_lock(Mutex::event, "CMSMarkStack._par_lock", true), 167 _hit_limit(0), 168 _failed_double(0) {} 169 170 bool allocate(size_t size); 171 172 size_t capacity() const { return _capacity; } 173 174 oop pop() { 175 if (!isEmpty()) { 176 return _base[--_index] ; 177 } 178 return NULL; 179 } 180 181 bool push(oop ptr) { 182 if (isFull()) { 183 return false; 184 } else { 185 _base[_index++] = ptr; 186 NOT_PRODUCT(_max_depth = MAX2(_max_depth, _index)); 187 return true; 188 } 189 } 190 191 bool isEmpty() const { return _index == 0; } 192 bool isFull() const { 193 assert(_index <= _capacity, "buffer overflow"); 194 return _index == _capacity; 195 } 196 197 size_t length() { return _index; } 198 199 // "Parallel versions" of some of the above 200 oop par_pop() { 201 // lock and pop 202 MutexLockerEx x(&_par_lock, Mutex::_no_safepoint_check_flag); 203 return pop(); 204 } 205 206 bool par_push(oop ptr) { 207 // lock and push 208 MutexLockerEx x(&_par_lock, Mutex::_no_safepoint_check_flag); 209 return push(ptr); 210 } 211 212 // Forcibly reset the stack, losing all of its contents. 213 void reset() { 214 _index = 0; 215 } 216 217 // Expand the stack, typically in response to an overflow condition 218 void expand(); 219 220 // Compute the least valued stack element. 221 oop least_value(HeapWord* low) { 222 oop least = (oop)low; 223 for (size_t i = 0; i < _index; i++) { 224 least = MIN2(least, _base[i]); 225 } 226 return least; 227 } 228 229 // Exposed here to allow stack expansion in || case 230 Mutex* par_lock() { return &_par_lock; } 231 }; 232 233 class CardTableRS; 234 class CMSParGCThreadState; 235 236 class ModUnionClosure: public MemRegionClosure { 237 protected: 238 CMSBitMap* _t; 239 public: 240 ModUnionClosure(CMSBitMap* t): _t(t) { } 241 void do_MemRegion(MemRegion mr); 242 }; 243 244 class ModUnionClosurePar: public ModUnionClosure { 245 public: 246 ModUnionClosurePar(CMSBitMap* t): ModUnionClosure(t) { } 247 void do_MemRegion(MemRegion mr); 248 }; 249 250 // Survivor Chunk Array in support of parallelization of 251 // Survivor Space rescan. 252 class ChunkArray: public CHeapObj { 253 size_t _index; 254 size_t _capacity; 255 HeapWord** _array; // storage for array 256 257 public: 258 ChunkArray() : _index(0), _capacity(0), _array(NULL) {} 259 ChunkArray(HeapWord** a, size_t c): 260 _index(0), _capacity(c), _array(a) {} 261 262 HeapWord** array() { return _array; } 263 void set_array(HeapWord** a) { _array = a; } 264 265 size_t capacity() { return _capacity; } 266 void set_capacity(size_t c) { _capacity = c; } 267 268 size_t end() { 269 assert(_index < capacity(), "_index out of bounds"); 270 return _index; 271 } // exclusive 272 273 HeapWord* nth(size_t n) { 274 assert(n < end(), "Out of bounds access"); 275 return _array[n]; 276 } 277 278 void reset() { 279 _index = 0; 280 } 281 282 void record_sample(HeapWord* p, size_t sz) { 283 // For now we do not do anything with the size 284 if (_index < _capacity) { 285 _array[_index++] = p; 286 } 287 } 288 }; 289 290 // 291 // Timing, allocation and promotion statistics for gc scheduling and incremental 292 // mode pacing. Most statistics are exponential averages. 293 // 294 class CMSStats VALUE_OBJ_CLASS_SPEC { 295 private: 296 ConcurrentMarkSweepGeneration* const _cms_gen; // The cms (old) gen. 297 298 // The following are exponential averages with factor alpha: 299 // avg = (100 - alpha) * avg + alpha * cur_sample 300 // 301 // The durations measure: end_time[n] - start_time[n] 302 // The periods measure: start_time[n] - start_time[n-1] 303 // 304 // The cms period and duration include only concurrent collections; time spent 305 // in foreground cms collections due to System.gc() or because of a failure to 306 // keep up are not included. 307 // 308 // There are 3 alphas to "bootstrap" the statistics. The _saved_alpha is the 309 // real value, but is used only after the first period. A value of 100 is 310 // used for the first sample so it gets the entire weight. 311 unsigned int _saved_alpha; // 0-100 312 unsigned int _gc0_alpha; 313 unsigned int _cms_alpha; 314 315 double _gc0_duration; 316 double _gc0_period; 317 size_t _gc0_promoted; // bytes promoted per gc0 318 double _cms_duration; 319 double _cms_duration_pre_sweep; // time from initiation to start of sweep 320 double _cms_duration_per_mb; 321 double _cms_period; 322 size_t _cms_allocated; // bytes of direct allocation per gc0 period 323 324 // Timers. 325 elapsedTimer _cms_timer; 326 TimeStamp _gc0_begin_time; 327 TimeStamp _cms_begin_time; 328 TimeStamp _cms_end_time; 329 330 // Snapshots of the amount used in the CMS generation. 331 size_t _cms_used_at_gc0_begin; 332 size_t _cms_used_at_gc0_end; 333 size_t _cms_used_at_cms_begin; 334 335 // Used to prevent the duty cycle from being reduced in the middle of a cms 336 // cycle. 337 bool _allow_duty_cycle_reduction; 338 339 enum { 340 _GC0_VALID = 0x1, 341 _CMS_VALID = 0x2, 342 _ALL_VALID = _GC0_VALID | _CMS_VALID 343 }; 344 345 unsigned int _valid_bits; 346 347 unsigned int _icms_duty_cycle; // icms duty cycle (0-100). 348 349 protected: 350 351 // Return a duty cycle that avoids wild oscillations, by limiting the amount 352 // of change between old_duty_cycle and new_duty_cycle (the latter is treated 353 // as a recommended value). 354 static unsigned int icms_damped_duty_cycle(unsigned int old_duty_cycle, 355 unsigned int new_duty_cycle); 356 unsigned int icms_update_duty_cycle_impl(); 357 358 // In support of adjusting of cms trigger ratios based on history 359 // of concurrent mode failure. 360 double cms_free_adjustment_factor(size_t free) const; 361 void adjust_cms_free_adjustment_factor(bool fail, size_t free); 362 363 public: 364 CMSStats(ConcurrentMarkSweepGeneration* cms_gen, 365 unsigned int alpha = CMSExpAvgFactor); 366 367 // Whether or not the statistics contain valid data; higher level statistics 368 // cannot be called until this returns true (they require at least one young 369 // gen and one cms cycle to have completed). 370 bool valid() const; 371 372 // Record statistics. 373 void record_gc0_begin(); 374 void record_gc0_end(size_t cms_gen_bytes_used); 375 void record_cms_begin(); 376 void record_cms_end(); 377 378 // Allow management of the cms timer, which must be stopped/started around 379 // yield points. 380 elapsedTimer& cms_timer() { return _cms_timer; } 381 void start_cms_timer() { _cms_timer.start(); } 382 void stop_cms_timer() { _cms_timer.stop(); } 383 384 // Basic statistics; units are seconds or bytes. 385 double gc0_period() const { return _gc0_period; } 386 double gc0_duration() const { return _gc0_duration; } 387 size_t gc0_promoted() const { return _gc0_promoted; } 388 double cms_period() const { return _cms_period; } 389 double cms_duration() const { return _cms_duration; } 390 double cms_duration_per_mb() const { return _cms_duration_per_mb; } 391 size_t cms_allocated() const { return _cms_allocated; } 392 393 size_t cms_used_at_gc0_end() const { return _cms_used_at_gc0_end;} 394 395 // Seconds since the last background cms cycle began or ended. 396 double cms_time_since_begin() const; 397 double cms_time_since_end() const; 398 399 // Higher level statistics--caller must check that valid() returns true before 400 // calling. 401 402 // Returns bytes promoted per second of wall clock time. 403 double promotion_rate() const; 404 405 // Returns bytes directly allocated per second of wall clock time. 406 double cms_allocation_rate() const; 407 408 // Rate at which space in the cms generation is being consumed (sum of the 409 // above two). 410 double cms_consumption_rate() const; 411 412 // Returns an estimate of the number of seconds until the cms generation will 413 // fill up, assuming no collection work is done. 414 double time_until_cms_gen_full() const; 415 416 // Returns an estimate of the number of seconds remaining until 417 // the cms generation collection should start. 418 double time_until_cms_start() const; 419 420 // End of higher level statistics. 421 422 // Returns the cms incremental mode duty cycle, as a percentage (0-100). 423 unsigned int icms_duty_cycle() const { return _icms_duty_cycle; } 424 425 // Update the duty cycle and return the new value. 426 unsigned int icms_update_duty_cycle(); 427 428 // Debugging. 429 void print_on(outputStream* st) const PRODUCT_RETURN; 430 void print() const { print_on(gclog_or_tty); } 431 }; 432 433 // A closure related to weak references processing which 434 // we embed in the CMSCollector, since we need to pass 435 // it to the reference processor for secondary filtering 436 // of references based on reachability of referent; 437 // see role of _is_alive_non_header closure in the 438 // ReferenceProcessor class. 439 // For objects in the CMS generation, this closure checks 440 // if the object is "live" (reachable). Used in weak 441 // reference processing. 442 class CMSIsAliveClosure: public BoolObjectClosure { 443 const MemRegion _span; 444 const CMSBitMap* _bit_map; 445 446 friend class CMSCollector; 447 public: 448 CMSIsAliveClosure(MemRegion span, 449 CMSBitMap* bit_map): 450 _span(span), 451 _bit_map(bit_map) { 452 assert(!span.is_empty(), "Empty span could spell trouble"); 453 } 454 455 void do_object(oop obj) { 456 assert(false, "not to be invoked"); 457 } 458 459 bool do_object_b(oop obj); 460 }; 461 462 463 // Implements AbstractRefProcTaskExecutor for CMS. 464 class CMSRefProcTaskExecutor: public AbstractRefProcTaskExecutor { 465 public: 466 467 CMSRefProcTaskExecutor(CMSCollector& collector) 468 : _collector(collector) 469 { } 470 471 // Executes a task using worker threads. 472 virtual void execute(ProcessTask& task); 473 virtual void execute(EnqueueTask& task); 474 private: 475 CMSCollector& _collector; 476 }; 477 478 479 class CMSCollector: public CHeapObj { 480 friend class VMStructs; 481 friend class ConcurrentMarkSweepThread; 482 friend class ConcurrentMarkSweepGeneration; 483 friend class CompactibleFreeListSpace; 484 friend class CMSParRemarkTask; 485 friend class CMSConcMarkingTask; 486 friend class CMSRefProcTaskProxy; 487 friend class CMSRefProcTaskExecutor; 488 friend class ScanMarkedObjectsAgainCarefullyClosure; // for sampling eden 489 friend class SurvivorSpacePrecleanClosure; // --- ditto ------- 490 friend class PushOrMarkClosure; // to access _restart_addr 491 friend class Par_PushOrMarkClosure; // to access _restart_addr 492 friend class MarkFromRootsClosure; // -- ditto -- 493 // ... and for clearing cards 494 friend class Par_MarkFromRootsClosure; // to access _restart_addr 495 // ... and for clearing cards 496 friend class Par_ConcMarkingClosure; // to access _restart_addr etc. 497 friend class MarkFromRootsVerifyClosure; // to access _restart_addr 498 friend class PushAndMarkVerifyClosure; // -- ditto -- 499 friend class MarkRefsIntoAndScanClosure; // to access _overflow_list 500 friend class PushAndMarkClosure; // -- ditto -- 501 friend class Par_PushAndMarkClosure; // -- ditto -- 502 friend class CMSKeepAliveClosure; // -- ditto -- 503 friend class CMSDrainMarkingStackClosure; // -- ditto -- 504 friend class CMSInnerParMarkAndPushClosure; // -- ditto -- 505 NOT_PRODUCT(friend class ScanMarkedObjectsAgainClosure;) // assertion on _overflow_list 506 friend class ReleaseForegroundGC; // to access _foregroundGCShouldWait 507 friend class VM_CMS_Operation; 508 friend class VM_CMS_Initial_Mark; 509 friend class VM_CMS_Final_Remark; 510 511 private: 512 jlong _time_of_last_gc; 513 void update_time_of_last_gc(jlong now) { 514 _time_of_last_gc = now; 515 } 516 517 OopTaskQueueSet* _task_queues; 518 519 // Overflow list of grey objects, threaded through mark-word 520 // Manipulated with CAS in the parallel/multi-threaded case. 521 oop _overflow_list; 522 // The following array-pair keeps track of mark words 523 // displaced for accomodating overflow list above. 524 // This code will likely be revisited under RFE#4922830. 525 GrowableArray<oop>* _preserved_oop_stack; 526 GrowableArray<markOop>* _preserved_mark_stack; 527 528 int* _hash_seed; 529 530 // In support of multi-threaded concurrent phases 531 YieldingFlexibleWorkGang* _conc_workers; 532 533 // Performance Counters 534 CollectorCounters* _gc_counters; 535 536 // Initialization Errors 537 bool _completed_initialization; 538 539 // In support of ExplicitGCInvokesConcurrent 540 static bool _full_gc_requested; 541 unsigned int _collection_count_start; 542 543 // Should we unload classes this concurrent cycle? 544 bool _should_unload_classes; 545 unsigned int _concurrent_cycles_since_last_unload; 546 unsigned int concurrent_cycles_since_last_unload() const { 547 return _concurrent_cycles_since_last_unload; 548 } 549 // Did we (allow) unload classes in the previous concurrent cycle? 550 bool unloaded_classes_last_cycle() const { 551 return concurrent_cycles_since_last_unload() == 0; 552 } 553 // Root scanning options for perm gen 554 int _roots_scanning_options; 555 int roots_scanning_options() const { return _roots_scanning_options; } 556 void add_root_scanning_option(int o) { _roots_scanning_options |= o; } 557 void remove_root_scanning_option(int o) { _roots_scanning_options &= ~o; } 558 559 // Verification support 560 CMSBitMap _verification_mark_bm; 561 void verify_after_remark_work_1(); 562 void verify_after_remark_work_2(); 563 564 // true if any verification flag is on. 565 bool _verifying; 566 bool verifying() const { return _verifying; } 567 void set_verifying(bool v) { _verifying = v; } 568 569 // Collector policy 570 ConcurrentMarkSweepPolicy* _collector_policy; 571 ConcurrentMarkSweepPolicy* collector_policy() { return _collector_policy; } 572 573 // XXX Move these to CMSStats ??? FIX ME !!! 574 elapsedTimer _inter_sweep_timer; // time between sweeps 575 elapsedTimer _intra_sweep_timer; // time _in_ sweeps 576 // padded decaying average estimates of the above 577 AdaptivePaddedAverage _inter_sweep_estimate; 578 AdaptivePaddedAverage _intra_sweep_estimate; 579 580 protected: 581 ConcurrentMarkSweepGeneration* _cmsGen; // old gen (CMS) 582 ConcurrentMarkSweepGeneration* _permGen; // perm gen 583 MemRegion _span; // span covering above two 584 CardTableRS* _ct; // card table 585 586 // CMS marking support structures 587 CMSBitMap _markBitMap; 588 CMSBitMap _modUnionTable; 589 CMSMarkStack _markStack; 590 CMSMarkStack _revisitStack; // used to keep track of klassKlass objects 591 // to revisit 592 CMSBitMap _perm_gen_verify_bit_map; // Mark bit map for perm gen verification support. 593 594 HeapWord* _restart_addr; // in support of marking stack overflow 595 void lower_restart_addr(HeapWord* low); 596 597 // Counters in support of marking stack / work queue overflow handling: 598 // a non-zero value indicates certain types of overflow events during 599 // the current CMS cycle and could lead to stack resizing efforts at 600 // an opportune future time. 601 size_t _ser_pmc_preclean_ovflw; 602 size_t _ser_pmc_remark_ovflw; 603 size_t _par_pmc_remark_ovflw; 604 size_t _ser_kac_preclean_ovflw; 605 size_t _ser_kac_ovflw; 606 size_t _par_kac_ovflw; 607 NOT_PRODUCT(ssize_t _num_par_pushes;) 608 609 // ("Weak") Reference processing support 610 ReferenceProcessor* _ref_processor; 611 CMSIsAliveClosure _is_alive_closure; 612 // keep this textually after _markBitMap and _span; c'tor dependency 613 614 ConcurrentMarkSweepThread* _cmsThread; // the thread doing the work 615 ModUnionClosure _modUnionClosure; 616 ModUnionClosurePar _modUnionClosurePar; 617 618 // CMS abstract state machine 619 // initial_state: Idling 620 // next_state(Idling) = {Marking} 621 // next_state(Marking) = {Precleaning, Sweeping} 622 // next_state(Precleaning) = {AbortablePreclean, FinalMarking} 623 // next_state(AbortablePreclean) = {FinalMarking} 624 // next_state(FinalMarking) = {Sweeping} 625 // next_state(Sweeping) = {Resizing} 626 // next_state(Resizing) = {Resetting} 627 // next_state(Resetting) = {Idling} 628 // The numeric values below are chosen so that: 629 // . _collectorState <= Idling == post-sweep && pre-mark 630 // . _collectorState in (Idling, Sweeping) == {initial,final}marking || 631 // precleaning || abortablePrecleanb 632 public: 633 enum CollectorState { 634 Resizing = 0, 635 Resetting = 1, 636 Idling = 2, 637 InitialMarking = 3, 638 Marking = 4, 639 Precleaning = 5, 640 AbortablePreclean = 6, 641 FinalMarking = 7, 642 Sweeping = 8 643 }; 644 protected: 645 static CollectorState _collectorState; 646 647 // State related to prologue/epilogue invocation for my generations 648 bool _between_prologue_and_epilogue; 649 650 // Signalling/State related to coordination between fore- and backgroud GC 651 // Note: When the baton has been passed from background GC to foreground GC, 652 // _foregroundGCIsActive is true and _foregroundGCShouldWait is false. 653 static bool _foregroundGCIsActive; // true iff foreground collector is active or 654 // wants to go active 655 static bool _foregroundGCShouldWait; // true iff background GC is active and has not 656 // yet passed the baton to the foreground GC 657 658 // Support for CMSScheduleRemark (abortable preclean) 659 bool _abort_preclean; 660 bool _start_sampling; 661 662 int _numYields; 663 size_t _numDirtyCards; 664 size_t _sweep_count; 665 // number of full gc's since the last concurrent gc. 666 uint _full_gcs_since_conc_gc; 667 668 // occupancy used for bootstrapping stats 669 double _bootstrap_occupancy; 670 671 // timer 672 elapsedTimer _timer; 673 674 // Timing, allocation and promotion statistics, used for scheduling. 675 CMSStats _stats; 676 677 // Allocation limits installed in the young gen, used only in 678 // CMSIncrementalMode. When an allocation in the young gen would cross one of 679 // these limits, the cms generation is notified and the cms thread is started 680 // or stopped, respectively. 681 HeapWord* _icms_start_limit; 682 HeapWord* _icms_stop_limit; 683 684 enum CMS_op_type { 685 CMS_op_checkpointRootsInitial, 686 CMS_op_checkpointRootsFinal 687 }; 688 689 void do_CMS_operation(CMS_op_type op); 690 bool stop_world_and_do(CMS_op_type op); 691 692 OopTaskQueueSet* task_queues() { return _task_queues; } 693 int* hash_seed(int i) { return &_hash_seed[i]; } 694 YieldingFlexibleWorkGang* conc_workers() { return _conc_workers; } 695 696 // Support for parallelizing Eden rescan in CMS remark phase 697 void sample_eden(); // ... sample Eden space top 698 699 private: 700 // Support for parallelizing young gen rescan in CMS remark phase 701 Generation* _young_gen; // the younger gen 702 HeapWord** _top_addr; // ... Top of Eden 703 HeapWord** _end_addr; // ... End of Eden 704 HeapWord** _eden_chunk_array; // ... Eden partitioning array 705 size_t _eden_chunk_index; // ... top (exclusive) of array 706 size_t _eden_chunk_capacity; // ... max entries in array 707 708 // Support for parallelizing survivor space rescan 709 HeapWord** _survivor_chunk_array; 710 size_t _survivor_chunk_index; 711 size_t _survivor_chunk_capacity; 712 size_t* _cursor; 713 ChunkArray* _survivor_plab_array; 714 715 // Support for marking stack overflow handling 716 bool take_from_overflow_list(size_t num, CMSMarkStack* to_stack); 717 bool par_take_from_overflow_list(size_t num, OopTaskQueue* to_work_q); 718 void push_on_overflow_list(oop p); 719 void par_push_on_overflow_list(oop p); 720 // the following is, obviously, not, in general, "MT-stable" 721 bool overflow_list_is_empty() const; 722 723 void preserve_mark_if_necessary(oop p); 724 void par_preserve_mark_if_necessary(oop p); 725 void preserve_mark_work(oop p, markOop m); 726 void restore_preserved_marks_if_any(); 727 NOT_PRODUCT(bool no_preserved_marks() const;) 728 // in support of testing overflow code 729 NOT_PRODUCT(int _overflow_counter;) 730 NOT_PRODUCT(bool simulate_overflow();) // sequential 731 NOT_PRODUCT(bool par_simulate_overflow();) // MT version 732 733 // CMS work methods 734 void checkpointRootsInitialWork(bool asynch); // initial checkpoint work 735 736 // a return value of false indicates failure due to stack overflow 737 bool markFromRootsWork(bool asynch); // concurrent marking work 738 739 public: // FIX ME!!! only for testing 740 bool do_marking_st(bool asynch); // single-threaded marking 741 bool do_marking_mt(bool asynch); // multi-threaded marking 742 743 private: 744 745 // concurrent precleaning work 746 size_t preclean_mod_union_table(ConcurrentMarkSweepGeneration* gen, 747 ScanMarkedObjectsAgainCarefullyClosure* cl); 748 size_t preclean_card_table(ConcurrentMarkSweepGeneration* gen, 749 ScanMarkedObjectsAgainCarefullyClosure* cl); 750 // Does precleaning work, returning a quantity indicative of 751 // the amount of "useful work" done. 752 size_t preclean_work(bool clean_refs, bool clean_survivors); 753 void abortable_preclean(); // Preclean while looking for possible abort 754 void initialize_sequential_subtasks_for_young_gen_rescan(int i); 755 // Helper function for above; merge-sorts the per-thread plab samples 756 void merge_survivor_plab_arrays(ContiguousSpace* surv); 757 // Resets (i.e. clears) the per-thread plab sample vectors 758 void reset_survivor_plab_arrays(); 759 760 // final (second) checkpoint work 761 void checkpointRootsFinalWork(bool asynch, bool clear_all_soft_refs, 762 bool init_mark_was_synchronous); 763 // work routine for parallel version of remark 764 void do_remark_parallel(); 765 // work routine for non-parallel version of remark 766 void do_remark_non_parallel(); 767 // reference processing work routine (during second checkpoint) 768 void refProcessingWork(bool asynch, bool clear_all_soft_refs); 769 770 // concurrent sweeping work 771 void sweepWork(ConcurrentMarkSweepGeneration* gen, bool asynch); 772 773 // (concurrent) resetting of support data structures 774 void reset(bool asynch); 775 776 // Clear _expansion_cause fields of constituent generations 777 void clear_expansion_cause(); 778 779 // An auxilliary method used to record the ends of 780 // used regions of each generation to limit the extent of sweep 781 void save_sweep_limits(); 782 783 // Resize the generations included in the collector. 784 void compute_new_size(); 785 786 // A work method used by foreground collection to determine 787 // what type of collection (compacting or not, continuing or fresh) 788 // it should do. 789 void decide_foreground_collection_type(bool clear_all_soft_refs, 790 bool* should_compact, bool* should_start_over); 791 792 // A work method used by the foreground collector to do 793 // a mark-sweep-compact. 794 void do_compaction_work(bool clear_all_soft_refs); 795 796 // A work method used by the foreground collector to do 797 // a mark-sweep, after taking over from a possibly on-going 798 // concurrent mark-sweep collection. 799 void do_mark_sweep_work(bool clear_all_soft_refs, 800 CollectorState first_state, bool should_start_over); 801 802 // If the backgrould GC is active, acquire control from the background 803 // GC and do the collection. 804 void acquire_control_and_collect(bool full, bool clear_all_soft_refs); 805 806 // For synchronizing passing of control from background to foreground 807 // GC. waitForForegroundGC() is called by the background 808 // collector. It if had to wait for a foreground collection, 809 // it returns true and the background collection should assume 810 // that the collection was finished by the foreground 811 // collector. 812 bool waitForForegroundGC(); 813 814 // Incremental mode triggering: recompute the icms duty cycle and set the 815 // allocation limits in the young gen. 816 void icms_update_allocation_limits(); 817 818 size_t block_size_using_printezis_bits(HeapWord* addr) const; 819 size_t block_size_if_printezis_bits(HeapWord* addr) const; 820 HeapWord* next_card_start_after_block(HeapWord* addr) const; 821 822 void setup_cms_unloading_and_verification_state(); 823 public: 824 CMSCollector(ConcurrentMarkSweepGeneration* cmsGen, 825 ConcurrentMarkSweepGeneration* permGen, 826 CardTableRS* ct, 827 ConcurrentMarkSweepPolicy* cp); 828 ConcurrentMarkSweepThread* cmsThread() { return _cmsThread; } 829 830 ReferenceProcessor* ref_processor() { return _ref_processor; } 831 void ref_processor_init(); 832 833 Mutex* bitMapLock() const { return _markBitMap.lock(); } 834 static CollectorState abstract_state() { return _collectorState; } 835 836 bool should_abort_preclean() const; // Whether preclean should be aborted. 837 size_t get_eden_used() const; 838 size_t get_eden_capacity() const; 839 840 ConcurrentMarkSweepGeneration* cmsGen() { return _cmsGen; } 841 842 // locking checks 843 NOT_PRODUCT(static bool have_cms_token();) 844 845 // XXXPERM bool should_collect(bool full, size_t size, bool tlab); 846 bool shouldConcurrentCollect(); 847 848 void collect(bool full, 849 bool clear_all_soft_refs, 850 size_t size, 851 bool tlab); 852 void collect_in_background(bool clear_all_soft_refs); 853 void collect_in_foreground(bool clear_all_soft_refs); 854 855 // In support of ExplicitGCInvokesConcurrent 856 static void request_full_gc(unsigned int full_gc_count); 857 // Should we unload classes in a particular concurrent cycle? 858 bool should_unload_classes() const { 859 return _should_unload_classes; 860 } 861 bool update_should_unload_classes(); 862 863 void direct_allocated(HeapWord* start, size_t size); 864 865 // Object is dead if not marked and current phase is sweeping. 866 bool is_dead_obj(oop obj) const; 867 868 // After a promotion (of "start"), do any necessary marking. 869 // If "par", then it's being done by a parallel GC thread. 870 // The last two args indicate if we need precise marking 871 // and if so the size of the object so it can be dirtied 872 // in its entirety. 873 void promoted(bool par, HeapWord* start, 874 bool is_obj_array, size_t obj_size); 875 876 HeapWord* allocation_limit_reached(Space* space, HeapWord* top, 877 size_t word_size); 878 879 void getFreelistLocks() const; 880 void releaseFreelistLocks() const; 881 bool haveFreelistLocks() const; 882 883 // GC prologue and epilogue 884 void gc_prologue(bool full); 885 void gc_epilogue(bool full); 886 887 jlong time_of_last_gc(jlong now) { 888 if (_collectorState <= Idling) { 889 // gc not in progress 890 return _time_of_last_gc; 891 } else { 892 // collection in progress 893 return now; 894 } 895 } 896 897 // Support for parallel remark of survivor space 898 void* get_data_recorder(int thr_num); 899 900 CMSBitMap* markBitMap() { return &_markBitMap; } 901 void directAllocated(HeapWord* start, size_t size); 902 903 // main CMS steps and related support 904 void checkpointRootsInitial(bool asynch); 905 bool markFromRoots(bool asynch); // a return value of false indicates failure 906 // due to stack overflow 907 void preclean(); 908 void checkpointRootsFinal(bool asynch, bool clear_all_soft_refs, 909 bool init_mark_was_synchronous); 910 void sweep(bool asynch); 911 912 // Check that the currently executing thread is the expected 913 // one (foreground collector or background collector). 914 static void check_correct_thread_executing() PRODUCT_RETURN; 915 // XXXPERM void print_statistics() PRODUCT_RETURN; 916 917 bool is_cms_reachable(HeapWord* addr); 918 919 // Performance Counter Support 920 CollectorCounters* counters() { return _gc_counters; } 921 922 // timer stuff 923 void startTimer() { assert(!_timer.is_active(), "Error"); _timer.start(); } 924 void stopTimer() { assert( _timer.is_active(), "Error"); _timer.stop(); } 925 void resetTimer() { assert(!_timer.is_active(), "Error"); _timer.reset(); } 926 double timerValue() { assert(!_timer.is_active(), "Error"); return _timer.seconds(); } 927 928 int yields() { return _numYields; } 929 void resetYields() { _numYields = 0; } 930 void incrementYields() { _numYields++; } 931 void resetNumDirtyCards() { _numDirtyCards = 0; } 932 void incrementNumDirtyCards(size_t num) { _numDirtyCards += num; } 933 size_t numDirtyCards() { return _numDirtyCards; } 934 935 static bool foregroundGCShouldWait() { return _foregroundGCShouldWait; } 936 static void set_foregroundGCShouldWait(bool v) { _foregroundGCShouldWait = v; } 937 static bool foregroundGCIsActive() { return _foregroundGCIsActive; } 938 static void set_foregroundGCIsActive(bool v) { _foregroundGCIsActive = v; } 939 size_t sweep_count() const { return _sweep_count; } 940 void increment_sweep_count() { _sweep_count++; } 941 942 // Timers/stats for gc scheduling and incremental mode pacing. 943 CMSStats& stats() { return _stats; } 944 945 // Convenience methods that check whether CMSIncrementalMode is enabled and 946 // forward to the corresponding methods in ConcurrentMarkSweepThread. 947 static void start_icms(); 948 static void stop_icms(); // Called at the end of the cms cycle. 949 static void disable_icms(); // Called before a foreground collection. 950 static void enable_icms(); // Called after a foreground collection. 951 void icms_wait(); // Called at yield points. 952 953 // Adaptive size policy 954 CMSAdaptiveSizePolicy* size_policy(); 955 CMSGCAdaptivePolicyCounters* gc_adaptive_policy_counters(); 956 957 // debugging 958 void verify(bool); 959 bool verify_after_remark(); 960 void verify_ok_to_terminate() const PRODUCT_RETURN; 961 void verify_work_stacks_empty() const PRODUCT_RETURN; 962 void verify_overflow_empty() const PRODUCT_RETURN; 963 964 // convenience methods in support of debugging 965 static const size_t skip_header_HeapWords() PRODUCT_RETURN0; 966 HeapWord* block_start(const void* p) const PRODUCT_RETURN0; 967 968 // accessors 969 CMSMarkStack* verification_mark_stack() { return &_markStack; } 970 CMSBitMap* verification_mark_bm() { return &_verification_mark_bm; } 971 972 // Get the bit map with a perm gen "deadness" information. 973 CMSBitMap* perm_gen_verify_bit_map() { return &_perm_gen_verify_bit_map; } 974 975 // Initialization errors 976 bool completed_initialization() { return _completed_initialization; } 977 }; 978 979 class CMSExpansionCause : public AllStatic { 980 public: 981 enum Cause { 982 _no_expansion, 983 _satisfy_free_ratio, 984 _satisfy_promotion, 985 _satisfy_allocation, 986 _allocate_par_lab, 987 _allocate_par_spooling_space, 988 _adaptive_size_policy 989 }; 990 // Return a string describing the cause of the expansion. 991 static const char* to_string(CMSExpansionCause::Cause cause); 992 }; 993 994 class ConcurrentMarkSweepGeneration: public CardGeneration { 995 friend class VMStructs; 996 friend class ConcurrentMarkSweepThread; 997 friend class ConcurrentMarkSweep; 998 friend class CMSCollector; 999 protected: 1000 static CMSCollector* _collector; // the collector that collects us 1001 CompactibleFreeListSpace* _cmsSpace; // underlying space (only one for now) 1002 1003 // Performance Counters 1004 GenerationCounters* _gen_counters; 1005 GSpaceCounters* _space_counters; 1006 1007 // Words directly allocated, used by CMSStats. 1008 size_t _direct_allocated_words; 1009 1010 // Non-product stat counters 1011 NOT_PRODUCT( 1012 int _numObjectsPromoted; 1013 int _numWordsPromoted; 1014 int _numObjectsAllocated; 1015 int _numWordsAllocated; 1016 ) 1017 1018 // Used for sizing decisions 1019 bool _incremental_collection_failed; 1020 bool incremental_collection_failed() { 1021 return _incremental_collection_failed; 1022 } 1023 void set_incremental_collection_failed() { 1024 _incremental_collection_failed = true; 1025 } 1026 void clear_incremental_collection_failed() { 1027 _incremental_collection_failed = false; 1028 } 1029 1030 // accessors 1031 void set_expansion_cause(CMSExpansionCause::Cause v) { _expansion_cause = v;} 1032 CMSExpansionCause::Cause expansion_cause() const { return _expansion_cause; } 1033 1034 private: 1035 // For parallel young-gen GC support. 1036 CMSParGCThreadState** _par_gc_thread_states; 1037 1038 // Reason generation was expanded 1039 CMSExpansionCause::Cause _expansion_cause; 1040 1041 // In support of MinChunkSize being larger than min object size 1042 const double _dilatation_factor; 1043 1044 enum CollectionTypes { 1045 Concurrent_collection_type = 0, 1046 MS_foreground_collection_type = 1, 1047 MSC_foreground_collection_type = 2, 1048 Unknown_collection_type = 3 1049 }; 1050 1051 CollectionTypes _debug_collection_type; 1052 1053 // Fraction of current occupancy at which to start a CMS collection which 1054 // will collect this generation (at least). 1055 double _initiating_occupancy; 1056 1057 protected: 1058 // Shrink generation by specified size (returns false if unable to shrink) 1059 virtual void shrink_by(size_t bytes); 1060 1061 // Update statistics for GC 1062 virtual void update_gc_stats(int level, bool full); 1063 1064 // Maximum available space in the generation (including uncommitted) 1065 // space. 1066 size_t max_available() const; 1067 1068 // getter and initializer for _initiating_occupancy field. 1069 double initiating_occupancy() const { return _initiating_occupancy; } 1070 void init_initiating_occupancy(intx io, intx tr); 1071 1072 public: 1073 ConcurrentMarkSweepGeneration(ReservedSpace rs, size_t initial_byte_size, 1074 int level, CardTableRS* ct, 1075 bool use_adaptive_freelists, 1076 FreeBlockDictionary::DictionaryChoice); 1077 1078 // Accessors 1079 CMSCollector* collector() const { return _collector; } 1080 static void set_collector(CMSCollector* collector) { 1081 assert(_collector == NULL, "already set"); 1082 _collector = collector; 1083 } 1084 CompactibleFreeListSpace* cmsSpace() const { return _cmsSpace; } 1085 1086 Mutex* freelistLock() const; 1087 1088 virtual Generation::Name kind() { return Generation::ConcurrentMarkSweep; } 1089 1090 // Adaptive size policy 1091 CMSAdaptiveSizePolicy* size_policy(); 1092 1093 bool refs_discovery_is_atomic() const { return false; } 1094 bool refs_discovery_is_mt() const { 1095 // Note: CMS does MT-discovery during the parallel-remark 1096 // phases. Use ReferenceProcessorMTMutator to make refs 1097 // discovery MT-safe during such phases or other parallel 1098 // discovery phases in the future. This may all go away 1099 // if/when we decide that refs discovery is sufficiently 1100 // rare that the cost of the CAS's involved is in the 1101 // noise. That's a measurement that should be done, and 1102 // the code simplified if that turns out to be the case. 1103 return false; 1104 } 1105 1106 // Override 1107 virtual void ref_processor_init(); 1108 1109 // Grow generation by specified size (returns false if unable to grow) 1110 bool grow_by(size_t bytes); 1111 // Grow generation to reserved size. 1112 bool grow_to_reserved(); 1113 1114 void clear_expansion_cause() { _expansion_cause = CMSExpansionCause::_no_expansion; } 1115 1116 // Space enquiries 1117 size_t capacity() const; 1118 size_t used() const; 1119 size_t free() const; 1120 double occupancy() const { return ((double)used())/((double)capacity()); } 1121 size_t contiguous_available() const; 1122 size_t unsafe_max_alloc_nogc() const; 1123 1124 // over-rides 1125 MemRegion used_region() const; 1126 MemRegion used_region_at_save_marks() const; 1127 1128 // Does a "full" (forced) collection invoked on this generation collect 1129 // all younger generations as well? Note that the second conjunct is a 1130 // hack to allow the collection of the younger gen first if the flag is 1131 // set. This is better than using th policy's should_collect_gen0_first() 1132 // since that causes us to do an extra unnecessary pair of restart-&-stop-world. 1133 virtual bool full_collects_younger_generations() const { 1134 return UseCMSCompactAtFullCollection && !CollectGen0First; 1135 } 1136 1137 void space_iterate(SpaceClosure* blk, bool usedOnly = false); 1138 1139 // Support for compaction 1140 CompactibleSpace* first_compaction_space() const; 1141 // Adjust quantites in the generation affected by 1142 // the compaction. 1143 void reset_after_compaction(); 1144 1145 // Allocation support 1146 HeapWord* allocate(size_t size, bool tlab); 1147 HeapWord* have_lock_and_allocate(size_t size, bool tlab); 1148 oop promote(oop obj, size_t obj_size); 1149 HeapWord* par_allocate(size_t size, bool tlab) { 1150 return allocate(size, tlab); 1151 } 1152 1153 // Incremental mode triggering. 1154 HeapWord* allocation_limit_reached(Space* space, HeapWord* top, 1155 size_t word_size); 1156 1157 // Used by CMSStats to track direct allocation. The value is sampled and 1158 // reset after each young gen collection. 1159 size_t direct_allocated_words() const { return _direct_allocated_words; } 1160 void reset_direct_allocated_words() { _direct_allocated_words = 0; } 1161 1162 // Overrides for parallel promotion. 1163 virtual oop par_promote(int thread_num, 1164 oop obj, markOop m, size_t word_sz); 1165 // This one should not be called for CMS. 1166 virtual void par_promote_alloc_undo(int thread_num, 1167 HeapWord* obj, size_t word_sz); 1168 virtual void par_promote_alloc_done(int thread_num); 1169 virtual void par_oop_since_save_marks_iterate_done(int thread_num); 1170 1171 virtual bool promotion_attempt_is_safe(size_t promotion_in_bytes, 1172 bool younger_handles_promotion_failure) const; 1173 1174 // Inform this (non-young) generation that a promotion failure was 1175 // encountered during a collection of a younger generation that 1176 // promotes into this generation. 1177 virtual void promotion_failure_occurred(); 1178 1179 bool should_collect(bool full, size_t size, bool tlab); 1180 virtual bool should_concurrent_collect() const; 1181 virtual bool is_too_full() const; 1182 void collect(bool full, 1183 bool clear_all_soft_refs, 1184 size_t size, 1185 bool tlab); 1186 1187 HeapWord* expand_and_allocate(size_t word_size, 1188 bool tlab, 1189 bool parallel = false); 1190 1191 // GC prologue and epilogue 1192 void gc_prologue(bool full); 1193 void gc_prologue_work(bool full, bool registerClosure, 1194 ModUnionClosure* modUnionClosure); 1195 void gc_epilogue(bool full); 1196 void gc_epilogue_work(bool full); 1197 1198 // Time since last GC of this generation 1199 jlong time_of_last_gc(jlong now) { 1200 return collector()->time_of_last_gc(now); 1201 } 1202 void update_time_of_last_gc(jlong now) { 1203 collector()-> update_time_of_last_gc(now); 1204 } 1205 1206 // Allocation failure 1207 void expand(size_t bytes, size_t expand_bytes, 1208 CMSExpansionCause::Cause cause); 1209 virtual bool expand(size_t bytes, size_t expand_bytes); 1210 void shrink(size_t bytes); 1211 HeapWord* expand_and_par_lab_allocate(CMSParGCThreadState* ps, size_t word_sz); 1212 bool expand_and_ensure_spooling_space(PromotionInfo* promo); 1213 1214 // Iteration support and related enquiries 1215 void save_marks(); 1216 bool no_allocs_since_save_marks(); 1217 void object_iterate_since_last_GC(ObjectClosure* cl); 1218 void younger_refs_iterate(OopsInGenClosure* cl); 1219 1220 // Iteration support specific to CMS generations 1221 void save_sweep_limit(); 1222 1223 // More iteration support 1224 virtual void oop_iterate(MemRegion mr, OopClosure* cl); 1225 virtual void oop_iterate(OopClosure* cl); 1226 virtual void safe_object_iterate(ObjectClosure* cl); 1227 virtual void object_iterate(ObjectClosure* cl); 1228 1229 // Need to declare the full complement of closures, whether we'll 1230 // override them or not, or get message from the compiler: 1231 // oop_since_save_marks_iterate_nv hides virtual function... 1232 #define CMS_SINCE_SAVE_MARKS_DECL(OopClosureType, nv_suffix) \ 1233 void oop_since_save_marks_iterate##nv_suffix(OopClosureType* cl); 1234 ALL_SINCE_SAVE_MARKS_CLOSURES(CMS_SINCE_SAVE_MARKS_DECL) 1235 1236 // Smart allocation XXX -- move to CFLSpace? 1237 void setNearLargestChunk(); 1238 bool isNearLargestChunk(HeapWord* addr); 1239 1240 // Get the chunk at the end of the space. Delagates to 1241 // the space. 1242 FreeChunk* find_chunk_at_end(); 1243 1244 // Overriding of unused functionality (sharing not yet supported with CMS) 1245 void pre_adjust_pointers(); 1246 void post_compact(); 1247 1248 // Debugging 1249 void prepare_for_verify(); 1250 void verify(bool allow_dirty); 1251 void print_statistics() PRODUCT_RETURN; 1252 1253 // Performance Counters support 1254 virtual void update_counters(); 1255 virtual void update_counters(size_t used); 1256 void initialize_performance_counters(); 1257 CollectorCounters* counters() { return collector()->counters(); } 1258 1259 // Support for parallel remark of survivor space 1260 void* get_data_recorder(int thr_num) { 1261 //Delegate to collector 1262 return collector()->get_data_recorder(thr_num); 1263 } 1264 1265 // Printing 1266 const char* name() const; 1267 virtual const char* short_name() const { return "CMS"; } 1268 void print() const; 1269 void printOccupancy(const char* s); 1270 bool must_be_youngest() const { return false; } 1271 bool must_be_oldest() const { return true; } 1272 1273 void compute_new_size(); 1274 1275 CollectionTypes debug_collection_type() { return _debug_collection_type; } 1276 void rotate_debug_collection_type(); 1277 }; 1278 1279 class ASConcurrentMarkSweepGeneration : public ConcurrentMarkSweepGeneration { 1280 1281 // Return the size policy from the heap's collector 1282 // policy casted to CMSAdaptiveSizePolicy*. 1283 CMSAdaptiveSizePolicy* cms_size_policy() const; 1284 1285 // Resize the generation based on the adaptive size 1286 // policy. 1287 void resize(size_t cur_promo, size_t desired_promo); 1288 1289 // Return the GC counters from the collector policy 1290 CMSGCAdaptivePolicyCounters* gc_adaptive_policy_counters(); 1291 1292 virtual void shrink_by(size_t bytes); 1293 1294 public: 1295 virtual void compute_new_size(); 1296 ASConcurrentMarkSweepGeneration(ReservedSpace rs, size_t initial_byte_size, 1297 int level, CardTableRS* ct, 1298 bool use_adaptive_freelists, 1299 FreeBlockDictionary::DictionaryChoice 1300 dictionaryChoice) : 1301 ConcurrentMarkSweepGeneration(rs, initial_byte_size, level, ct, 1302 use_adaptive_freelists, dictionaryChoice) {} 1303 1304 virtual const char* short_name() const { return "ASCMS"; } 1305 virtual Generation::Name kind() { return Generation::ASConcurrentMarkSweep; } 1306 1307 virtual void update_counters(); 1308 virtual void update_counters(size_t used); 1309 }; 1310 1311 // 1312 // Closures of various sorts used by CMS to accomplish its work 1313 // 1314 1315 // This closure is used to check that a certain set of oops is empty. 1316 class FalseClosure: public OopClosure { 1317 public: 1318 void do_oop(oop* p) { guarantee(false, "Should be an empty set"); } 1319 void do_oop(narrowOop* p) { guarantee(false, "Should be an empty set"); } 1320 }; 1321 1322 // This closure is used to do concurrent marking from the roots 1323 // following the first checkpoint. 1324 class MarkFromRootsClosure: public BitMapClosure { 1325 CMSCollector* _collector; 1326 MemRegion _span; 1327 CMSBitMap* _bitMap; 1328 CMSBitMap* _mut; 1329 CMSMarkStack* _markStack; 1330 CMSMarkStack* _revisitStack; 1331 bool _yield; 1332 int _skipBits; 1333 HeapWord* _finger; 1334 HeapWord* _threshold; 1335 DEBUG_ONLY(bool _verifying;) 1336 1337 public: 1338 MarkFromRootsClosure(CMSCollector* collector, MemRegion span, 1339 CMSBitMap* bitMap, 1340 CMSMarkStack* markStack, 1341 CMSMarkStack* revisitStack, 1342 bool should_yield, bool verifying = false); 1343 bool do_bit(size_t offset); 1344 void reset(HeapWord* addr); 1345 inline void do_yield_check(); 1346 1347 private: 1348 void scanOopsInOop(HeapWord* ptr); 1349 void do_yield_work(); 1350 }; 1351 1352 // This closure is used to do concurrent multi-threaded 1353 // marking from the roots following the first checkpoint. 1354 // XXX This should really be a subclass of The serial version 1355 // above, but i have not had the time to refactor things cleanly. 1356 // That willbe done for Dolphin. 1357 class Par_MarkFromRootsClosure: public BitMapClosure { 1358 CMSCollector* _collector; 1359 MemRegion _whole_span; 1360 MemRegion _span; 1361 CMSBitMap* _bit_map; 1362 CMSBitMap* _mut; 1363 OopTaskQueue* _work_queue; 1364 CMSMarkStack* _overflow_stack; 1365 CMSMarkStack* _revisit_stack; 1366 bool _yield; 1367 int _skip_bits; 1368 HeapWord* _finger; 1369 HeapWord* _threshold; 1370 CMSConcMarkingTask* _task; 1371 public: 1372 Par_MarkFromRootsClosure(CMSConcMarkingTask* task, CMSCollector* collector, 1373 MemRegion span, 1374 CMSBitMap* bit_map, 1375 OopTaskQueue* work_queue, 1376 CMSMarkStack* overflow_stack, 1377 CMSMarkStack* revisit_stack, 1378 bool should_yield); 1379 bool do_bit(size_t offset); 1380 inline void do_yield_check(); 1381 1382 private: 1383 void scan_oops_in_oop(HeapWord* ptr); 1384 void do_yield_work(); 1385 bool get_work_from_overflow_stack(); 1386 }; 1387 1388 // The following closures are used to do certain kinds of verification of 1389 // CMS marking. 1390 class PushAndMarkVerifyClosure: public OopClosure { 1391 CMSCollector* _collector; 1392 MemRegion _span; 1393 CMSBitMap* _verification_bm; 1394 CMSBitMap* _cms_bm; 1395 CMSMarkStack* _mark_stack; 1396 protected: 1397 void do_oop(oop p); 1398 template <class T> inline void do_oop_work(T *p) { 1399 oop obj = oopDesc::load_decode_heap_oop_not_null(p); 1400 do_oop(obj); 1401 } 1402 public: 1403 PushAndMarkVerifyClosure(CMSCollector* cms_collector, 1404 MemRegion span, 1405 CMSBitMap* verification_bm, 1406 CMSBitMap* cms_bm, 1407 CMSMarkStack* mark_stack); 1408 void do_oop(oop* p); 1409 void do_oop(narrowOop* p); 1410 // Deal with a stack overflow condition 1411 void handle_stack_overflow(HeapWord* lost); 1412 }; 1413 1414 class MarkFromRootsVerifyClosure: public BitMapClosure { 1415 CMSCollector* _collector; 1416 MemRegion _span; 1417 CMSBitMap* _verification_bm; 1418 CMSBitMap* _cms_bm; 1419 CMSMarkStack* _mark_stack; 1420 HeapWord* _finger; 1421 PushAndMarkVerifyClosure _pam_verify_closure; 1422 public: 1423 MarkFromRootsVerifyClosure(CMSCollector* collector, MemRegion span, 1424 CMSBitMap* verification_bm, 1425 CMSBitMap* cms_bm, 1426 CMSMarkStack* mark_stack); 1427 bool do_bit(size_t offset); 1428 void reset(HeapWord* addr); 1429 }; 1430 1431 1432 // This closure is used to check that a certain set of bits is 1433 // "empty" (i.e. the bit vector doesn't have any 1-bits). 1434 class FalseBitMapClosure: public BitMapClosure { 1435 public: 1436 bool do_bit(size_t offset) { 1437 guarantee(false, "Should not have a 1 bit"); 1438 return true; 1439 } 1440 }; 1441 1442 // This closure is used during the second checkpointing phase 1443 // to rescan the marked objects on the dirty cards in the mod 1444 // union table and the card table proper. It's invoked via 1445 // MarkFromDirtyCardsClosure below. It uses either 1446 // [Par_]MarkRefsIntoAndScanClosure (Par_ in the parallel case) 1447 // declared in genOopClosures.hpp to accomplish some of its work. 1448 // In the parallel case the bitMap is shared, so access to 1449 // it needs to be suitably synchronized for updates by embedded 1450 // closures that update it; however, this closure itself only 1451 // reads the bit_map and because it is idempotent, is immune to 1452 // reading stale values. 1453 class ScanMarkedObjectsAgainClosure: public UpwardsObjectClosure { 1454 #ifdef ASSERT 1455 CMSCollector* _collector; 1456 MemRegion _span; 1457 union { 1458 CMSMarkStack* _mark_stack; 1459 OopTaskQueue* _work_queue; 1460 }; 1461 #endif // ASSERT 1462 bool _parallel; 1463 CMSBitMap* _bit_map; 1464 union { 1465 MarkRefsIntoAndScanClosure* _scan_closure; 1466 Par_MarkRefsIntoAndScanClosure* _par_scan_closure; 1467 }; 1468 1469 public: 1470 ScanMarkedObjectsAgainClosure(CMSCollector* collector, 1471 MemRegion span, 1472 ReferenceProcessor* rp, 1473 CMSBitMap* bit_map, 1474 CMSMarkStack* mark_stack, 1475 CMSMarkStack* revisit_stack, 1476 MarkRefsIntoAndScanClosure* cl): 1477 #ifdef ASSERT 1478 _collector(collector), 1479 _span(span), 1480 _mark_stack(mark_stack), 1481 #endif // ASSERT 1482 _parallel(false), 1483 _bit_map(bit_map), 1484 _scan_closure(cl) { } 1485 1486 ScanMarkedObjectsAgainClosure(CMSCollector* collector, 1487 MemRegion span, 1488 ReferenceProcessor* rp, 1489 CMSBitMap* bit_map, 1490 OopTaskQueue* work_queue, 1491 CMSMarkStack* revisit_stack, 1492 Par_MarkRefsIntoAndScanClosure* cl): 1493 #ifdef ASSERT 1494 _collector(collector), 1495 _span(span), 1496 _work_queue(work_queue), 1497 #endif // ASSERT 1498 _parallel(true), 1499 _bit_map(bit_map), 1500 _par_scan_closure(cl) { } 1501 1502 void do_object(oop obj) { 1503 guarantee(false, "Call do_object_b(oop, MemRegion) instead"); 1504 } 1505 bool do_object_b(oop obj) { 1506 guarantee(false, "Call do_object_b(oop, MemRegion) form instead"); 1507 return false; 1508 } 1509 bool do_object_bm(oop p, MemRegion mr); 1510 }; 1511 1512 // This closure is used during the second checkpointing phase 1513 // to rescan the marked objects on the dirty cards in the mod 1514 // union table and the card table proper. It invokes 1515 // ScanMarkedObjectsAgainClosure above to accomplish much of its work. 1516 // In the parallel case, the bit map is shared and requires 1517 // synchronized access. 1518 class MarkFromDirtyCardsClosure: public MemRegionClosure { 1519 CompactibleFreeListSpace* _space; 1520 ScanMarkedObjectsAgainClosure _scan_cl; 1521 size_t _num_dirty_cards; 1522 1523 public: 1524 MarkFromDirtyCardsClosure(CMSCollector* collector, 1525 MemRegion span, 1526 CompactibleFreeListSpace* space, 1527 CMSBitMap* bit_map, 1528 CMSMarkStack* mark_stack, 1529 CMSMarkStack* revisit_stack, 1530 MarkRefsIntoAndScanClosure* cl): 1531 _space(space), 1532 _num_dirty_cards(0), 1533 _scan_cl(collector, span, collector->ref_processor(), bit_map, 1534 mark_stack, revisit_stack, cl) { } 1535 1536 MarkFromDirtyCardsClosure(CMSCollector* collector, 1537 MemRegion span, 1538 CompactibleFreeListSpace* space, 1539 CMSBitMap* bit_map, 1540 OopTaskQueue* work_queue, 1541 CMSMarkStack* revisit_stack, 1542 Par_MarkRefsIntoAndScanClosure* cl): 1543 _space(space), 1544 _num_dirty_cards(0), 1545 _scan_cl(collector, span, collector->ref_processor(), bit_map, 1546 work_queue, revisit_stack, cl) { } 1547 1548 void do_MemRegion(MemRegion mr); 1549 void set_space(CompactibleFreeListSpace* space) { _space = space; } 1550 size_t num_dirty_cards() { return _num_dirty_cards; } 1551 }; 1552 1553 // This closure is used in the non-product build to check 1554 // that there are no MemRegions with a certain property. 1555 class FalseMemRegionClosure: public MemRegionClosure { 1556 void do_MemRegion(MemRegion mr) { 1557 guarantee(!mr.is_empty(), "Shouldn't be empty"); 1558 guarantee(false, "Should never be here"); 1559 } 1560 }; 1561 1562 // This closure is used during the precleaning phase 1563 // to "carefully" rescan marked objects on dirty cards. 1564 // It uses MarkRefsIntoAndScanClosure declared in genOopClosures.hpp 1565 // to accomplish some of its work. 1566 class ScanMarkedObjectsAgainCarefullyClosure: public ObjectClosureCareful { 1567 CMSCollector* _collector; 1568 MemRegion _span; 1569 bool _yield; 1570 Mutex* _freelistLock; 1571 CMSBitMap* _bitMap; 1572 CMSMarkStack* _markStack; 1573 MarkRefsIntoAndScanClosure* _scanningClosure; 1574 1575 public: 1576 ScanMarkedObjectsAgainCarefullyClosure(CMSCollector* collector, 1577 MemRegion span, 1578 CMSBitMap* bitMap, 1579 CMSMarkStack* markStack, 1580 CMSMarkStack* revisitStack, 1581 MarkRefsIntoAndScanClosure* cl, 1582 bool should_yield): 1583 _collector(collector), 1584 _span(span), 1585 _yield(should_yield), 1586 _bitMap(bitMap), 1587 _markStack(markStack), 1588 _scanningClosure(cl) { 1589 } 1590 1591 void do_object(oop p) { 1592 guarantee(false, "call do_object_careful instead"); 1593 } 1594 1595 size_t do_object_careful(oop p) { 1596 guarantee(false, "Unexpected caller"); 1597 return 0; 1598 } 1599 1600 size_t do_object_careful_m(oop p, MemRegion mr); 1601 1602 void setFreelistLock(Mutex* m) { 1603 _freelistLock = m; 1604 _scanningClosure->set_freelistLock(m); 1605 } 1606 1607 private: 1608 inline bool do_yield_check(); 1609 1610 void do_yield_work(); 1611 }; 1612 1613 class SurvivorSpacePrecleanClosure: public ObjectClosureCareful { 1614 CMSCollector* _collector; 1615 MemRegion _span; 1616 bool _yield; 1617 CMSBitMap* _bit_map; 1618 CMSMarkStack* _mark_stack; 1619 PushAndMarkClosure* _scanning_closure; 1620 unsigned int _before_count; 1621 1622 public: 1623 SurvivorSpacePrecleanClosure(CMSCollector* collector, 1624 MemRegion span, 1625 CMSBitMap* bit_map, 1626 CMSMarkStack* mark_stack, 1627 PushAndMarkClosure* cl, 1628 unsigned int before_count, 1629 bool should_yield): 1630 _collector(collector), 1631 _span(span), 1632 _yield(should_yield), 1633 _bit_map(bit_map), 1634 _mark_stack(mark_stack), 1635 _scanning_closure(cl), 1636 _before_count(before_count) 1637 { } 1638 1639 void do_object(oop p) { 1640 guarantee(false, "call do_object_careful instead"); 1641 } 1642 1643 size_t do_object_careful(oop p); 1644 1645 size_t do_object_careful_m(oop p, MemRegion mr) { 1646 guarantee(false, "Unexpected caller"); 1647 return 0; 1648 } 1649 1650 private: 1651 inline void do_yield_check(); 1652 void do_yield_work(); 1653 }; 1654 1655 // This closure is used to accomplish the sweeping work 1656 // after the second checkpoint but before the concurrent reset 1657 // phase. 1658 // 1659 // Terminology 1660 // left hand chunk (LHC) - block of one or more chunks currently being 1661 // coalesced. The LHC is available for coalescing with a new chunk. 1662 // right hand chunk (RHC) - block that is currently being swept that is 1663 // free or garbage that can be coalesced with the LHC. 1664 // _inFreeRange is true if there is currently a LHC 1665 // _lastFreeRangeCoalesced is true if the LHC consists of more than one chunk. 1666 // _freeRangeInFreeLists is true if the LHC is in the free lists. 1667 // _freeFinger is the address of the current LHC 1668 class SweepClosure: public BlkClosureCareful { 1669 CMSCollector* _collector; // collector doing the work 1670 ConcurrentMarkSweepGeneration* _g; // Generation being swept 1671 CompactibleFreeListSpace* _sp; // Space being swept 1672 HeapWord* _limit; 1673 Mutex* _freelistLock; // Free list lock (in space) 1674 CMSBitMap* _bitMap; // Marking bit map (in 1675 // generation) 1676 bool _inFreeRange; // Indicates if we are in the 1677 // midst of a free run 1678 bool _freeRangeInFreeLists; 1679 // Often, we have just found 1680 // a free chunk and started 1681 // a new free range; we do not 1682 // eagerly remove this chunk from 1683 // the free lists unless there is 1684 // a possibility of coalescing. 1685 // When true, this flag indicates 1686 // that the _freeFinger below 1687 // points to a potentially free chunk 1688 // that may still be in the free lists 1689 bool _lastFreeRangeCoalesced; 1690 // free range contains chunks 1691 // coalesced 1692 bool _yield; 1693 // Whether sweeping should be 1694 // done with yields. For instance 1695 // when done by the foreground 1696 // collector we shouldn't yield. 1697 HeapWord* _freeFinger; // When _inFreeRange is set, the 1698 // pointer to the "left hand 1699 // chunk" 1700 size_t _freeRangeSize; 1701 // When _inFreeRange is set, this 1702 // indicates the accumulated size 1703 // of the "left hand chunk" 1704 NOT_PRODUCT( 1705 size_t _numObjectsFreed; 1706 size_t _numWordsFreed; 1707 size_t _numObjectsLive; 1708 size_t _numWordsLive; 1709 size_t _numObjectsAlreadyFree; 1710 size_t _numWordsAlreadyFree; 1711 FreeChunk* _last_fc; 1712 ) 1713 private: 1714 // Code that is common to a free chunk or garbage when 1715 // encountered during sweeping. 1716 void doPostIsFreeOrGarbageChunk(FreeChunk *fc, 1717 size_t chunkSize); 1718 // Process a free chunk during sweeping. 1719 void doAlreadyFreeChunk(FreeChunk *fc); 1720 // Process a garbage chunk during sweeping. 1721 size_t doGarbageChunk(FreeChunk *fc); 1722 // Process a live chunk during sweeping. 1723 size_t doLiveChunk(FreeChunk* fc); 1724 1725 // Accessors. 1726 HeapWord* freeFinger() const { return _freeFinger; } 1727 void set_freeFinger(HeapWord* v) { _freeFinger = v; } 1728 size_t freeRangeSize() const { return _freeRangeSize; } 1729 void set_freeRangeSize(size_t v) { _freeRangeSize = v; } 1730 bool inFreeRange() const { return _inFreeRange; } 1731 void set_inFreeRange(bool v) { _inFreeRange = v; } 1732 bool lastFreeRangeCoalesced() const { return _lastFreeRangeCoalesced; } 1733 void set_lastFreeRangeCoalesced(bool v) { _lastFreeRangeCoalesced = v; } 1734 bool freeRangeInFreeLists() const { return _freeRangeInFreeLists; } 1735 void set_freeRangeInFreeLists(bool v) { _freeRangeInFreeLists = v; } 1736 1737 // Initialize a free range. 1738 void initialize_free_range(HeapWord* freeFinger, bool freeRangeInFreeLists); 1739 // Return this chunk to the free lists. 1740 void flushCurFreeChunk(HeapWord* chunk, size_t size); 1741 1742 // Check if we should yield and do so when necessary. 1743 inline void do_yield_check(HeapWord* addr); 1744 1745 // Yield 1746 void do_yield_work(HeapWord* addr); 1747 1748 // Debugging/Printing 1749 void record_free_block_coalesced(FreeChunk* fc) const PRODUCT_RETURN; 1750 1751 public: 1752 SweepClosure(CMSCollector* collector, ConcurrentMarkSweepGeneration* g, 1753 CMSBitMap* bitMap, bool should_yield); 1754 ~SweepClosure(); 1755 1756 size_t do_blk_careful(HeapWord* addr); 1757 }; 1758 1759 // Closures related to weak references processing 1760 1761 // During CMS' weak reference processing, this is a 1762 // work-routine/closure used to complete transitive 1763 // marking of objects as live after a certain point 1764 // in which an initial set has been completely accumulated. 1765 // This closure is currently used both during the final 1766 // remark stop-world phase, as well as during the concurrent 1767 // precleaning of the discovered reference lists. 1768 class CMSDrainMarkingStackClosure: public VoidClosure { 1769 CMSCollector* _collector; 1770 MemRegion _span; 1771 CMSMarkStack* _mark_stack; 1772 CMSBitMap* _bit_map; 1773 CMSKeepAliveClosure* _keep_alive; 1774 bool _concurrent_precleaning; 1775 public: 1776 CMSDrainMarkingStackClosure(CMSCollector* collector, MemRegion span, 1777 CMSBitMap* bit_map, CMSMarkStack* mark_stack, 1778 CMSKeepAliveClosure* keep_alive, 1779 bool cpc): 1780 _collector(collector), 1781 _span(span), 1782 _bit_map(bit_map), 1783 _mark_stack(mark_stack), 1784 _keep_alive(keep_alive), 1785 _concurrent_precleaning(cpc) { 1786 assert(_concurrent_precleaning == _keep_alive->concurrent_precleaning(), 1787 "Mismatch"); 1788 } 1789 1790 void do_void(); 1791 }; 1792 1793 // A parallel version of CMSDrainMarkingStackClosure above. 1794 class CMSParDrainMarkingStackClosure: public VoidClosure { 1795 CMSCollector* _collector; 1796 MemRegion _span; 1797 OopTaskQueue* _work_queue; 1798 CMSBitMap* _bit_map; 1799 CMSInnerParMarkAndPushClosure _mark_and_push; 1800 1801 public: 1802 CMSParDrainMarkingStackClosure(CMSCollector* collector, 1803 MemRegion span, CMSBitMap* bit_map, 1804 CMSMarkStack* revisit_stack, 1805 OopTaskQueue* work_queue): 1806 _collector(collector), 1807 _span(span), 1808 _bit_map(bit_map), 1809 _work_queue(work_queue), 1810 _mark_and_push(collector, span, bit_map, revisit_stack, work_queue) { } 1811 1812 public: 1813 void trim_queue(uint max); 1814 void do_void(); 1815 }; 1816 1817 // Allow yielding or short-circuiting of reference list 1818 // prelceaning work. 1819 class CMSPrecleanRefsYieldClosure: public YieldClosure { 1820 CMSCollector* _collector; 1821 void do_yield_work(); 1822 public: 1823 CMSPrecleanRefsYieldClosure(CMSCollector* collector): 1824 _collector(collector) {} 1825 virtual bool should_return(); 1826 }; 1827 1828 1829 // Convenience class that locks free list locks for given CMS collector 1830 class FreelistLocker: public StackObj { 1831 private: 1832 CMSCollector* _collector; 1833 public: 1834 FreelistLocker(CMSCollector* collector): 1835 _collector(collector) { 1836 _collector->getFreelistLocks(); 1837 } 1838 1839 ~FreelistLocker() { 1840 _collector->releaseFreelistLocks(); 1841 } 1842 }; 1843 1844 // Mark all dead objects in a given space. 1845 class MarkDeadObjectsClosure: public BlkClosure { 1846 const CMSCollector* _collector; 1847 const CompactibleFreeListSpace* _sp; 1848 CMSBitMap* _live_bit_map; 1849 CMSBitMap* _dead_bit_map; 1850 public: 1851 MarkDeadObjectsClosure(const CMSCollector* collector, 1852 const CompactibleFreeListSpace* sp, 1853 CMSBitMap *live_bit_map, 1854 CMSBitMap *dead_bit_map) : 1855 _collector(collector), 1856 _sp(sp), 1857 _live_bit_map(live_bit_map), 1858 _dead_bit_map(dead_bit_map) {} 1859 size_t do_blk(HeapWord* addr); 1860 };