1 /* 2 * Copyright (c) 2001, 2015, 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 #ifndef SHARE_VM_GC_G1_CONCURRENTMARK_HPP 26 #define SHARE_VM_GC_G1_CONCURRENTMARK_HPP 27 28 #include "classfile/javaClasses.hpp" 29 #include "gc/g1/g1RegionToSpaceMapper.hpp" 30 #include "gc/g1/heapRegionSet.hpp" 31 #include "gc/shared/gcId.hpp" 32 #include "gc/shared/taskqueue.hpp" 33 34 class G1CollectedHeap; 35 class CMBitMap; 36 class CMTask; 37 class ConcurrentMark; 38 typedef GenericTaskQueue<oop, mtGC> CMTaskQueue; 39 typedef GenericTaskQueueSet<CMTaskQueue, mtGC> CMTaskQueueSet; 40 41 // Closure used by CM during concurrent reference discovery 42 // and reference processing (during remarking) to determine 43 // if a particular object is alive. It is primarily used 44 // to determine if referents of discovered reference objects 45 // are alive. An instance is also embedded into the 46 // reference processor as the _is_alive_non_header field 47 class G1CMIsAliveClosure: public BoolObjectClosure { 48 G1CollectedHeap* _g1; 49 public: 50 G1CMIsAliveClosure(G1CollectedHeap* g1) : _g1(g1) { } 51 52 bool do_object_b(oop obj); 53 }; 54 55 // A generic CM bit map. This is essentially a wrapper around the BitMap 56 // class, with one bit per (1<<_shifter) HeapWords. 57 58 class CMBitMapRO VALUE_OBJ_CLASS_SPEC { 59 protected: 60 HeapWord* _bmStartWord; // base address of range covered by map 61 size_t _bmWordSize; // map size (in #HeapWords covered) 62 const int _shifter; // map to char or bit 63 BitMap _bm; // the bit map itself 64 65 public: 66 // constructor 67 CMBitMapRO(int shifter); 68 69 enum { do_yield = true }; 70 71 // inquiries 72 HeapWord* startWord() const { return _bmStartWord; } 73 size_t sizeInWords() const { return _bmWordSize; } 74 // the following is one past the last word in space 75 HeapWord* endWord() const { return _bmStartWord + _bmWordSize; } 76 77 // read marks 78 79 bool isMarked(HeapWord* addr) const { 80 assert(_bmStartWord <= addr && addr < (_bmStartWord + _bmWordSize), 81 "outside underlying space?"); 82 return _bm.at(heapWordToOffset(addr)); 83 } 84 85 // iteration 86 inline bool iterate(BitMapClosure* cl, MemRegion mr); 87 inline bool iterate(BitMapClosure* cl); 88 89 // Return the address corresponding to the next marked bit at or after 90 // "addr", and before "limit", if "limit" is non-NULL. If there is no 91 // such bit, returns "limit" if that is non-NULL, or else "endWord()". 92 HeapWord* getNextMarkedWordAddress(const HeapWord* addr, 93 const HeapWord* limit = NULL) const; 94 // Return the address corresponding to the next unmarked bit at or after 95 // "addr", and before "limit", if "limit" is non-NULL. If there is no 96 // such bit, returns "limit" if that is non-NULL, or else "endWord()". 97 HeapWord* getNextUnmarkedWordAddress(const HeapWord* addr, 98 const HeapWord* limit = NULL) const; 99 100 // conversion utilities 101 HeapWord* offsetToHeapWord(size_t offset) const { 102 return _bmStartWord + (offset << _shifter); 103 } 104 size_t heapWordToOffset(const HeapWord* addr) const { 105 return pointer_delta(addr, _bmStartWord) >> _shifter; 106 } 107 int heapWordDiffToOffsetDiff(size_t diff) const; 108 109 // The argument addr should be the start address of a valid object 110 HeapWord* nextObject(HeapWord* addr) { 111 oop obj = (oop) addr; 112 HeapWord* res = addr + obj->size(); 113 assert(offsetToHeapWord(heapWordToOffset(res)) == res, "sanity"); 114 return res; 115 } 116 117 void print_on_error(outputStream* st, const char* prefix) const; 118 119 // debugging 120 NOT_PRODUCT(bool covers(MemRegion rs) const;) 121 }; 122 123 class CMBitMapMappingChangedListener : public G1MappingChangedListener { 124 private: 125 CMBitMap* _bm; 126 public: 127 CMBitMapMappingChangedListener() : _bm(NULL) {} 128 129 void set_bitmap(CMBitMap* bm) { _bm = bm; } 130 131 virtual void on_commit(uint start_idx, size_t num_regions, bool zero_filled); 132 }; 133 134 class CMBitMap : public CMBitMapRO { 135 private: 136 CMBitMapMappingChangedListener _listener; 137 138 public: 139 static size_t compute_size(size_t heap_size); 140 // Returns the amount of bytes on the heap between two marks in the bitmap. 141 static size_t mark_distance(); 142 // Returns how many bytes (or bits) of the heap a single byte (or bit) of the 143 // mark bitmap corresponds to. This is the same as the mark distance above. 144 static size_t heap_map_factor() { 145 return mark_distance(); 146 } 147 148 CMBitMap() : CMBitMapRO(LogMinObjAlignment), _listener() { _listener.set_bitmap(this); } 149 150 // Initializes the underlying BitMap to cover the given area. 151 void initialize(MemRegion heap, G1RegionToSpaceMapper* storage); 152 153 // Write marks. 154 inline void mark(HeapWord* addr); 155 inline void clear(HeapWord* addr); 156 inline bool parMark(HeapWord* addr); 157 inline bool parClear(HeapWord* addr); 158 159 void markRange(MemRegion mr); 160 void clearRange(MemRegion mr); 161 162 // Starting at the bit corresponding to "addr" (inclusive), find the next 163 // "1" bit, if any. This bit starts some run of consecutive "1"'s; find 164 // the end of this run (stopping at "end_addr"). Return the MemRegion 165 // covering from the start of the region corresponding to the first bit 166 // of the run to the end of the region corresponding to the last bit of 167 // the run. If there is no "1" bit at or after "addr", return an empty 168 // MemRegion. 169 MemRegion getAndClearMarkedRegion(HeapWord* addr, HeapWord* end_addr); 170 171 // Clear the whole mark bitmap. 172 void clearAll(); 173 }; 174 175 // Represents a marking stack used by ConcurrentMarking in the G1 collector. 176 class CMMarkStack VALUE_OBJ_CLASS_SPEC { 177 VirtualSpace _virtual_space; // Underlying backing store for actual stack 178 ConcurrentMark* _cm; 179 oop* _base; // bottom of stack 180 jint _index; // one more than last occupied index 181 jint _capacity; // max #elements 182 jint _saved_index; // value of _index saved at start of GC 183 184 bool _overflow; 185 bool _should_expand; 186 DEBUG_ONLY(bool _drain_in_progress;) 187 DEBUG_ONLY(bool _drain_in_progress_yields;) 188 189 oop pop() { 190 if (!isEmpty()) { 191 return _base[--_index] ; 192 } 193 return NULL; 194 } 195 196 public: 197 CMMarkStack(ConcurrentMark* cm); 198 ~CMMarkStack(); 199 200 bool allocate(size_t capacity); 201 202 // Pushes the first "n" elements of "ptr_arr" on the stack. 203 // Locking impl: concurrency is allowed only with 204 // "par_push_arr" and/or "par_pop_arr" operations, which use the same 205 // locking strategy. 206 void par_push_arr(oop* ptr_arr, int n); 207 208 // If returns false, the array was empty. Otherwise, removes up to "max" 209 // elements from the stack, and transfers them to "ptr_arr" in an 210 // unspecified order. The actual number transferred is given in "n" ("n 211 // == 0" is deliberately redundant with the return value.) Locking impl: 212 // concurrency is allowed only with "par_push_arr" and/or "par_pop_arr" 213 // operations, which use the same locking strategy. 214 bool par_pop_arr(oop* ptr_arr, int max, int* n); 215 216 // Drain the mark stack, applying the given closure to all fields of 217 // objects on the stack. (That is, continue until the stack is empty, 218 // even if closure applications add entries to the stack.) The "bm" 219 // argument, if non-null, may be used to verify that only marked objects 220 // are on the mark stack. If "yield_after" is "true", then the 221 // concurrent marker performing the drain offers to yield after 222 // processing each object. If a yield occurs, stops the drain operation 223 // and returns false. Otherwise, returns true. 224 template<class OopClosureClass> 225 bool drain(OopClosureClass* cl, CMBitMap* bm, bool yield_after = false); 226 227 bool isEmpty() { return _index == 0; } 228 int maxElems() { return _capacity; } 229 230 bool overflow() { return _overflow; } 231 void clear_overflow() { _overflow = false; } 232 233 bool should_expand() const { return _should_expand; } 234 void set_should_expand(); 235 236 // Expand the stack, typically in response to an overflow condition 237 void expand(); 238 239 int size() { return _index; } 240 241 void setEmpty() { _index = 0; clear_overflow(); } 242 243 // Record the current index. 244 void note_start_of_gc(); 245 246 // Make sure that we have not added any entries to the stack during GC. 247 void note_end_of_gc(); 248 249 // Apply fn to each oop in the mark stack, up to the bound recorded 250 // via one of the above "note" functions. The mark stack must not 251 // be modified while iterating. 252 template<typename Fn> void iterate(Fn fn); 253 }; 254 255 class ForceOverflowSettings VALUE_OBJ_CLASS_SPEC { 256 private: 257 #ifndef PRODUCT 258 uintx _num_remaining; 259 bool _force; 260 #endif // !defined(PRODUCT) 261 262 public: 263 void init() PRODUCT_RETURN; 264 void update() PRODUCT_RETURN; 265 bool should_force() PRODUCT_RETURN_( return false; ); 266 }; 267 268 // this will enable a variety of different statistics per GC task 269 #define _MARKING_STATS_ 0 270 // this will enable the higher verbose levels 271 #define _MARKING_VERBOSE_ 0 272 273 #if _MARKING_STATS_ 274 #define statsOnly(statement) \ 275 do { \ 276 statement ; \ 277 } while (0) 278 #else // _MARKING_STATS_ 279 #define statsOnly(statement) \ 280 do { \ 281 } while (0) 282 #endif // _MARKING_STATS_ 283 284 typedef enum { 285 no_verbose = 0, // verbose turned off 286 stats_verbose, // only prints stats at the end of marking 287 low_verbose, // low verbose, mostly per region and per major event 288 medium_verbose, // a bit more detailed than low 289 high_verbose // per object verbose 290 } CMVerboseLevel; 291 292 class YoungList; 293 294 // Root Regions are regions that are not empty at the beginning of a 295 // marking cycle and which we might collect during an evacuation pause 296 // while the cycle is active. Given that, during evacuation pauses, we 297 // do not copy objects that are explicitly marked, what we have to do 298 // for the root regions is to scan them and mark all objects reachable 299 // from them. According to the SATB assumptions, we only need to visit 300 // each object once during marking. So, as long as we finish this scan 301 // before the next evacuation pause, we can copy the objects from the 302 // root regions without having to mark them or do anything else to them. 303 // 304 // Currently, we only support root region scanning once (at the start 305 // of the marking cycle) and the root regions are all the survivor 306 // regions populated during the initial-mark pause. 307 class CMRootRegions VALUE_OBJ_CLASS_SPEC { 308 private: 309 YoungList* _young_list; 310 ConcurrentMark* _cm; 311 312 volatile bool _scan_in_progress; 313 volatile bool _should_abort; 314 HeapRegion* volatile _next_survivor; 315 316 public: 317 CMRootRegions(); 318 // We actually do most of the initialization in this method. 319 void init(G1CollectedHeap* g1h, ConcurrentMark* cm); 320 321 // Reset the claiming / scanning of the root regions. 322 void prepare_for_scan(); 323 324 // Forces get_next() to return NULL so that the iteration aborts early. 325 void abort() { _should_abort = true; } 326 327 // Return true if the CM thread are actively scanning root regions, 328 // false otherwise. 329 bool scan_in_progress() { return _scan_in_progress; } 330 331 // Claim the next root region to scan atomically, or return NULL if 332 // all have been claimed. 333 HeapRegion* claim_next(); 334 335 // Flag that we're done with root region scanning and notify anyone 336 // who's waiting on it. If aborted is false, assume that all regions 337 // have been claimed. 338 void scan_finished(); 339 340 // If CM threads are still scanning root regions, wait until they 341 // are done. Return true if we had to wait, false otherwise. 342 bool wait_until_scan_finished(); 343 }; 344 345 class ConcurrentMarkThread; 346 347 class ConcurrentMark: public CHeapObj<mtGC> { 348 friend class CMMarkStack; 349 friend class ConcurrentMarkThread; 350 friend class CMTask; 351 friend class CMBitMapClosure; 352 friend class CMRemarkTask; 353 friend class CMConcurrentMarkingTask; 354 friend class G1ParNoteEndTask; 355 friend class CalcLiveObjectsClosure; 356 friend class G1CMRefProcTaskProxy; 357 friend class G1CMRefProcTaskExecutor; 358 friend class G1CMKeepAliveAndDrainClosure; 359 friend class G1CMDrainMarkingStackClosure; 360 361 protected: 362 ConcurrentMarkThread* _cmThread; // The thread doing the work 363 G1CollectedHeap* _g1h; // The heap 364 uint _parallel_marking_threads; // The number of marking 365 // threads we're using 366 uint _max_parallel_marking_threads; // Max number of marking 367 // threads we'll ever use 368 double _sleep_factor; // How much we have to sleep, with 369 // respect to the work we just did, to 370 // meet the marking overhead goal 371 double _marking_task_overhead; // Marking target overhead for 372 // a single task 373 374 // Same as the two above, but for the cleanup task 375 double _cleanup_sleep_factor; 376 double _cleanup_task_overhead; 377 378 FreeRegionList _cleanup_list; 379 380 // Concurrent marking support structures 381 CMBitMap _markBitMap1; 382 CMBitMap _markBitMap2; 383 CMBitMapRO* _prevMarkBitMap; // Completed mark bitmap 384 CMBitMap* _nextMarkBitMap; // Under-construction mark bitmap 385 386 BitMap _region_bm; 387 BitMap _card_bm; 388 389 // Heap bounds 390 HeapWord* _heap_start; 391 HeapWord* _heap_end; 392 393 // Root region tracking and claiming 394 CMRootRegions _root_regions; 395 396 // For gray objects 397 CMMarkStack _markStack; // Grey objects behind global finger 398 HeapWord* volatile _finger; // The global finger, region aligned, 399 // always points to the end of the 400 // last claimed region 401 402 // Marking tasks 403 uint _max_worker_id;// Maximum worker id 404 uint _active_tasks; // Task num currently active 405 CMTask** _tasks; // Task queue array (max_worker_id len) 406 CMTaskQueueSet* _task_queues; // Task queue set 407 ParallelTaskTerminator _terminator; // For termination 408 409 // Two sync barriers that are used to synchronize tasks when an 410 // overflow occurs. The algorithm is the following. All tasks enter 411 // the first one to ensure that they have all stopped manipulating 412 // the global data structures. After they exit it, they re-initialize 413 // their data structures and task 0 re-initializes the global data 414 // structures. Then, they enter the second sync barrier. This 415 // ensure, that no task starts doing work before all data 416 // structures (local and global) have been re-initialized. When they 417 // exit it, they are free to start working again. 418 WorkGangBarrierSync _first_overflow_barrier_sync; 419 WorkGangBarrierSync _second_overflow_barrier_sync; 420 421 // This is set by any task, when an overflow on the global data 422 // structures is detected 423 volatile bool _has_overflown; 424 // True: marking is concurrent, false: we're in remark 425 volatile bool _concurrent; 426 // Set at the end of a Full GC so that marking aborts 427 volatile bool _has_aborted; 428 GCId _aborted_gc_id; 429 430 // Used when remark aborts due to an overflow to indicate that 431 // another concurrent marking phase should start 432 volatile bool _restart_for_overflow; 433 434 // This is true from the very start of concurrent marking until the 435 // point when all the tasks complete their work. It is really used 436 // to determine the points between the end of concurrent marking and 437 // time of remark. 438 volatile bool _concurrent_marking_in_progress; 439 440 // Verbose level 441 CMVerboseLevel _verbose_level; 442 443 // All of these times are in ms 444 NumberSeq _init_times; 445 NumberSeq _remark_times; 446 NumberSeq _remark_mark_times; 447 NumberSeq _remark_weak_ref_times; 448 NumberSeq _cleanup_times; 449 double _total_counting_time; 450 double _total_rs_scrub_time; 451 452 double* _accum_task_vtime; // Accumulated task vtime 453 454 WorkGang* _parallel_workers; 455 456 ForceOverflowSettings _force_overflow_conc; 457 ForceOverflowSettings _force_overflow_stw; 458 459 void weakRefsWorkParallelPart(BoolObjectClosure* is_alive, bool purged_classes); 460 void weakRefsWork(bool clear_all_soft_refs); 461 462 void swapMarkBitMaps(); 463 464 // It resets the global marking data structures, as well as the 465 // task local ones; should be called during initial mark. 466 void reset(); 467 468 // Resets all the marking data structures. Called when we have to restart 469 // marking or when marking completes (via set_non_marking_state below). 470 void reset_marking_state(bool clear_overflow = true); 471 472 // We do this after we're done with marking so that the marking data 473 // structures are initialized to a sensible and predictable state. 474 void set_non_marking_state(); 475 476 // Called to indicate how many threads are currently active. 477 void set_concurrency(uint active_tasks); 478 479 // It should be called to indicate which phase we're in (concurrent 480 // mark or remark) and how many threads are currently active. 481 void set_concurrency_and_phase(uint active_tasks, bool concurrent); 482 483 // Prints all gathered CM-related statistics 484 void print_stats(); 485 486 bool cleanup_list_is_empty() { 487 return _cleanup_list.is_empty(); 488 } 489 490 // Accessor methods 491 uint parallel_marking_threads() const { return _parallel_marking_threads; } 492 uint max_parallel_marking_threads() const { return _max_parallel_marking_threads;} 493 double sleep_factor() { return _sleep_factor; } 494 double marking_task_overhead() { return _marking_task_overhead;} 495 double cleanup_sleep_factor() { return _cleanup_sleep_factor; } 496 double cleanup_task_overhead() { return _cleanup_task_overhead;} 497 498 HeapWord* finger() { return _finger; } 499 bool concurrent() { return _concurrent; } 500 uint active_tasks() { return _active_tasks; } 501 ParallelTaskTerminator* terminator() { return &_terminator; } 502 503 // It claims the next available region to be scanned by a marking 504 // task/thread. It might return NULL if the next region is empty or 505 // we have run out of regions. In the latter case, out_of_regions() 506 // determines whether we've really run out of regions or the task 507 // should call claim_region() again. This might seem a bit 508 // awkward. Originally, the code was written so that claim_region() 509 // either successfully returned with a non-empty region or there 510 // were no more regions to be claimed. The problem with this was 511 // that, in certain circumstances, it iterated over large chunks of 512 // the heap finding only empty regions and, while it was working, it 513 // was preventing the calling task to call its regular clock 514 // method. So, this way, each task will spend very little time in 515 // claim_region() and is allowed to call the regular clock method 516 // frequently. 517 HeapRegion* claim_region(uint worker_id); 518 519 // It determines whether we've run out of regions to scan. Note that 520 // the finger can point past the heap end in case the heap was expanded 521 // to satisfy an allocation without doing a GC. This is fine, because all 522 // objects in those regions will be considered live anyway because of 523 // SATB guarantees (i.e. their TAMS will be equal to bottom). 524 bool out_of_regions() { return _finger >= _heap_end; } 525 526 // Returns the task with the given id 527 CMTask* task(int id) { 528 assert(0 <= id && id < (int) _active_tasks, 529 "task id not within active bounds"); 530 return _tasks[id]; 531 } 532 533 // Returns the task queue with the given id 534 CMTaskQueue* task_queue(int id) { 535 assert(0 <= id && id < (int) _active_tasks, 536 "task queue id not within active bounds"); 537 return (CMTaskQueue*) _task_queues->queue(id); 538 } 539 540 // Returns the task queue set 541 CMTaskQueueSet* task_queues() { return _task_queues; } 542 543 // Access / manipulation of the overflow flag which is set to 544 // indicate that the global stack has overflown 545 bool has_overflown() { return _has_overflown; } 546 void set_has_overflown() { _has_overflown = true; } 547 void clear_has_overflown() { _has_overflown = false; } 548 bool restart_for_overflow() { return _restart_for_overflow; } 549 550 // Methods to enter the two overflow sync barriers 551 void enter_first_sync_barrier(uint worker_id); 552 void enter_second_sync_barrier(uint worker_id); 553 554 ForceOverflowSettings* force_overflow_conc() { 555 return &_force_overflow_conc; 556 } 557 558 ForceOverflowSettings* force_overflow_stw() { 559 return &_force_overflow_stw; 560 } 561 562 ForceOverflowSettings* force_overflow() { 563 if (concurrent()) { 564 return force_overflow_conc(); 565 } else { 566 return force_overflow_stw(); 567 } 568 } 569 570 // Live Data Counting data structures... 571 // These data structures are initialized at the start of 572 // marking. They are written to while marking is active. 573 // They are aggregated during remark; the aggregated values 574 // are then used to populate the _region_bm, _card_bm, and 575 // the total live bytes, which are then subsequently updated 576 // during cleanup. 577 578 // An array of bitmaps (one bit map per task). Each bitmap 579 // is used to record the cards spanned by the live objects 580 // marked by that task/worker. 581 BitMap* _count_card_bitmaps; 582 583 // Used to record the number of marked live bytes 584 // (for each region, by worker thread). 585 size_t** _count_marked_bytes; 586 587 // Card index of the bottom of the G1 heap. Used for biasing indices into 588 // the card bitmaps. 589 intptr_t _heap_bottom_card_num; 590 591 // Set to true when initialization is complete 592 bool _completed_initialization; 593 594 public: 595 // Manipulation of the global mark stack. 596 // The push and pop operations are used by tasks for transfers 597 // between task-local queues and the global mark stack, and use 598 // locking for concurrency safety. 599 bool mark_stack_push(oop* arr, int n) { 600 _markStack.par_push_arr(arr, n); 601 if (_markStack.overflow()) { 602 set_has_overflown(); 603 return false; 604 } 605 return true; 606 } 607 void mark_stack_pop(oop* arr, int max, int* n) { 608 _markStack.par_pop_arr(arr, max, n); 609 } 610 size_t mark_stack_size() { return _markStack.size(); } 611 size_t partial_mark_stack_size_target() { return _markStack.maxElems()/3; } 612 bool mark_stack_overflow() { return _markStack.overflow(); } 613 bool mark_stack_empty() { return _markStack.isEmpty(); } 614 615 CMRootRegions* root_regions() { return &_root_regions; } 616 617 bool concurrent_marking_in_progress() { 618 return _concurrent_marking_in_progress; 619 } 620 void set_concurrent_marking_in_progress() { 621 _concurrent_marking_in_progress = true; 622 } 623 void clear_concurrent_marking_in_progress() { 624 _concurrent_marking_in_progress = false; 625 } 626 627 void update_accum_task_vtime(int i, double vtime) { 628 _accum_task_vtime[i] += vtime; 629 } 630 631 double all_task_accum_vtime() { 632 double ret = 0.0; 633 for (uint i = 0; i < _max_worker_id; ++i) 634 ret += _accum_task_vtime[i]; 635 return ret; 636 } 637 638 // Attempts to steal an object from the task queues of other tasks 639 bool try_stealing(uint worker_id, int* hash_seed, oop& obj); 640 641 ConcurrentMark(G1CollectedHeap* g1h, 642 G1RegionToSpaceMapper* prev_bitmap_storage, 643 G1RegionToSpaceMapper* next_bitmap_storage); 644 ~ConcurrentMark(); 645 646 ConcurrentMarkThread* cmThread() { return _cmThread; } 647 648 CMBitMapRO* prevMarkBitMap() const { return _prevMarkBitMap; } 649 CMBitMap* nextMarkBitMap() const { return _nextMarkBitMap; } 650 651 // Returns the number of GC threads to be used in a concurrent 652 // phase based on the number of GC threads being used in a STW 653 // phase. 654 uint scale_parallel_threads(uint n_par_threads); 655 656 // Calculates the number of GC threads to be used in a concurrent phase. 657 uint calc_parallel_marking_threads(); 658 659 // The following three are interaction between CM and 660 // G1CollectedHeap 661 662 // This notifies CM that a root during initial-mark needs to be 663 // grayed. It is MT-safe. word_size is the size of the object in 664 // words. It is passed explicitly as sometimes we cannot calculate 665 // it from the given object because it might be in an inconsistent 666 // state (e.g., in to-space and being copied). So the caller is 667 // responsible for dealing with this issue (e.g., get the size from 668 // the from-space image when the to-space image might be 669 // inconsistent) and always passing the size. hr is the region that 670 // contains the object and it's passed optionally from callers who 671 // might already have it (no point in recalculating it). 672 inline void grayRoot(oop obj, 673 size_t word_size, 674 uint worker_id, 675 HeapRegion* hr = NULL); 676 677 // It iterates over the heap and for each object it comes across it 678 // will dump the contents of its reference fields, as well as 679 // liveness information for the object and its referents. The dump 680 // will be written to a file with the following name: 681 // G1PrintReachableBaseFile + "." + str. 682 // vo decides whether the prev (vo == UsePrevMarking), the next 683 // (vo == UseNextMarking) marking information, or the mark word 684 // (vo == UseMarkWord) will be used to determine the liveness of 685 // each object / referent. 686 // If all is true, all objects in the heap will be dumped, otherwise 687 // only the live ones. In the dump the following symbols / breviations 688 // are used: 689 // M : an explicitly live object (its bitmap bit is set) 690 // > : an implicitly live object (over tams) 691 // O : an object outside the G1 heap (typically: in the perm gen) 692 // NOT : a reference field whose referent is not live 693 // AND MARKED : indicates that an object is both explicitly and 694 // implicitly live (it should be one or the other, not both) 695 void print_reachable(const char* str, 696 VerifyOption vo, 697 bool all) PRODUCT_RETURN; 698 699 // Clear the next marking bitmap (will be called concurrently). 700 void clearNextBitmap(); 701 702 // Return whether the next mark bitmap has no marks set. To be used for assertions 703 // only. Will not yield to pause requests. 704 bool nextMarkBitmapIsClear(); 705 706 // These two do the work that needs to be done before and after the 707 // initial root checkpoint. Since this checkpoint can be done at two 708 // different points (i.e. an explicit pause or piggy-backed on a 709 // young collection), then it's nice to be able to easily share the 710 // pre/post code. It might be the case that we can put everything in 711 // the post method. TP 712 void checkpointRootsInitialPre(); 713 void checkpointRootsInitialPost(); 714 715 // Scan all the root regions and mark everything reachable from 716 // them. 717 void scanRootRegions(); 718 719 // Scan a single root region and mark everything reachable from it. 720 void scanRootRegion(HeapRegion* hr, uint worker_id); 721 722 // Do concurrent phase of marking, to a tentative transitive closure. 723 void markFromRoots(); 724 725 void checkpointRootsFinal(bool clear_all_soft_refs); 726 void checkpointRootsFinalWork(); 727 void cleanup(); 728 void completeCleanup(); 729 730 // Mark in the previous bitmap. NB: this is usually read-only, so use 731 // this carefully! 732 inline void markPrev(oop p); 733 734 // Clears marks for all objects in the given range, for the prev or 735 // next bitmaps. NB: the previous bitmap is usually 736 // read-only, so use this carefully! 737 void clearRangePrevBitmap(MemRegion mr); 738 void clearRangeNextBitmap(MemRegion mr); 739 740 // Notify data structures that a GC has started. 741 void note_start_of_gc() { 742 _markStack.note_start_of_gc(); 743 } 744 745 // Notify data structures that a GC is finished. 746 void note_end_of_gc() { 747 _markStack.note_end_of_gc(); 748 } 749 750 // Verify that there are no CSet oops on the stacks (taskqueues / 751 // global mark stack) and fingers (global / per-task). 752 // If marking is not in progress, it's a no-op. 753 void verify_no_cset_oops() PRODUCT_RETURN; 754 755 bool isPrevMarked(oop p) const { 756 assert(p != NULL && p->is_oop(), "expected an oop"); 757 HeapWord* addr = (HeapWord*)p; 758 assert(addr >= _prevMarkBitMap->startWord() || 759 addr < _prevMarkBitMap->endWord(), "in a region"); 760 761 return _prevMarkBitMap->isMarked(addr); 762 } 763 764 inline bool do_yield_check(uint worker_i = 0); 765 766 // Called to abort the marking cycle after a Full GC takes place. 767 void abort(); 768 769 bool has_aborted() { return _has_aborted; } 770 771 const GCId& concurrent_gc_id(); 772 773 // This prints the global/local fingers. It is used for debugging. 774 NOT_PRODUCT(void print_finger();) 775 776 void print_summary_info(); 777 778 void print_worker_threads_on(outputStream* st) const; 779 780 void print_on_error(outputStream* st) const; 781 782 // The following indicate whether a given verbose level has been 783 // set. Notice that anything above stats is conditional to 784 // _MARKING_VERBOSE_ having been set to 1 785 bool verbose_stats() { 786 return _verbose_level >= stats_verbose; 787 } 788 bool verbose_low() { 789 return _MARKING_VERBOSE_ && _verbose_level >= low_verbose; 790 } 791 bool verbose_medium() { 792 return _MARKING_VERBOSE_ && _verbose_level >= medium_verbose; 793 } 794 bool verbose_high() { 795 return _MARKING_VERBOSE_ && _verbose_level >= high_verbose; 796 } 797 798 // Liveness counting 799 800 // Utility routine to set an exclusive range of cards on the given 801 // card liveness bitmap 802 inline void set_card_bitmap_range(BitMap* card_bm, 803 BitMap::idx_t start_idx, 804 BitMap::idx_t end_idx, 805 bool is_par); 806 807 // Returns the card number of the bottom of the G1 heap. 808 // Used in biasing indices into accounting card bitmaps. 809 intptr_t heap_bottom_card_num() const { 810 return _heap_bottom_card_num; 811 } 812 813 // Returns the card bitmap for a given task or worker id. 814 BitMap* count_card_bitmap_for(uint worker_id) { 815 assert(worker_id < _max_worker_id, "oob"); 816 assert(_count_card_bitmaps != NULL, "uninitialized"); 817 BitMap* task_card_bm = &_count_card_bitmaps[worker_id]; 818 assert(task_card_bm->size() == _card_bm.size(), "size mismatch"); 819 return task_card_bm; 820 } 821 822 // Returns the array containing the marked bytes for each region, 823 // for the given worker or task id. 824 size_t* count_marked_bytes_array_for(uint worker_id) { 825 assert(worker_id < _max_worker_id, "oob"); 826 assert(_count_marked_bytes != NULL, "uninitialized"); 827 size_t* marked_bytes_array = _count_marked_bytes[worker_id]; 828 assert(marked_bytes_array != NULL, "uninitialized"); 829 return marked_bytes_array; 830 } 831 832 // Returns the index in the liveness accounting card table bitmap 833 // for the given address 834 inline BitMap::idx_t card_bitmap_index_for(HeapWord* addr); 835 836 // Counts the size of the given memory region in the the given 837 // marked_bytes array slot for the given HeapRegion. 838 // Sets the bits in the given card bitmap that are associated with the 839 // cards that are spanned by the memory region. 840 inline void count_region(MemRegion mr, 841 HeapRegion* hr, 842 size_t* marked_bytes_array, 843 BitMap* task_card_bm); 844 845 // Counts the given memory region in the task/worker counting 846 // data structures for the given worker id. 847 inline void count_region(MemRegion mr, HeapRegion* hr, uint worker_id); 848 849 // Counts the given object in the given task/worker counting 850 // data structures. 851 inline void count_object(oop obj, 852 HeapRegion* hr, 853 size_t* marked_bytes_array, 854 BitMap* task_card_bm); 855 856 // Attempts to mark the given object and, if successful, counts 857 // the object in the given task/worker counting structures. 858 inline bool par_mark_and_count(oop obj, 859 HeapRegion* hr, 860 size_t* marked_bytes_array, 861 BitMap* task_card_bm); 862 863 // Attempts to mark the given object and, if successful, counts 864 // the object in the task/worker counting structures for the 865 // given worker id. 866 inline bool par_mark_and_count(oop obj, 867 size_t word_size, 868 HeapRegion* hr, 869 uint worker_id); 870 871 // Returns true if initialization was successfully completed. 872 bool completed_initialization() const { 873 return _completed_initialization; 874 } 875 876 protected: 877 // Clear all the per-task bitmaps and arrays used to store the 878 // counting data. 879 void clear_all_count_data(); 880 881 // Aggregates the counting data for each worker/task 882 // that was constructed while marking. Also sets 883 // the amount of marked bytes for each region and 884 // the top at concurrent mark count. 885 void aggregate_count_data(); 886 887 // Verification routine 888 void verify_count_data(); 889 }; 890 891 // A class representing a marking task. 892 class CMTask : public TerminatorTerminator { 893 private: 894 enum PrivateConstants { 895 // the regular clock call is called once the scanned words reaches 896 // this limit 897 words_scanned_period = 12*1024, 898 // the regular clock call is called once the number of visited 899 // references reaches this limit 900 refs_reached_period = 384, 901 // initial value for the hash seed, used in the work stealing code 902 init_hash_seed = 17, 903 // how many entries will be transferred between global stack and 904 // local queues 905 global_stack_transfer_size = 16 906 }; 907 908 uint _worker_id; 909 G1CollectedHeap* _g1h; 910 ConcurrentMark* _cm; 911 CMBitMap* _nextMarkBitMap; 912 // the task queue of this task 913 CMTaskQueue* _task_queue; 914 private: 915 // the task queue set---needed for stealing 916 CMTaskQueueSet* _task_queues; 917 // indicates whether the task has been claimed---this is only for 918 // debugging purposes 919 bool _claimed; 920 921 // number of calls to this task 922 int _calls; 923 924 // when the virtual timer reaches this time, the marking step should 925 // exit 926 double _time_target_ms; 927 // the start time of the current marking step 928 double _start_time_ms; 929 930 // the oop closure used for iterations over oops 931 G1CMOopClosure* _cm_oop_closure; 932 933 // the region this task is scanning, NULL if we're not scanning any 934 HeapRegion* _curr_region; 935 // the local finger of this task, NULL if we're not scanning a region 936 HeapWord* _finger; 937 // limit of the region this task is scanning, NULL if we're not scanning one 938 HeapWord* _region_limit; 939 940 // the number of words this task has scanned 941 size_t _words_scanned; 942 // When _words_scanned reaches this limit, the regular clock is 943 // called. Notice that this might be decreased under certain 944 // circumstances (i.e. when we believe that we did an expensive 945 // operation). 946 size_t _words_scanned_limit; 947 // the initial value of _words_scanned_limit (i.e. what it was 948 // before it was decreased). 949 size_t _real_words_scanned_limit; 950 951 // the number of references this task has visited 952 size_t _refs_reached; 953 // When _refs_reached reaches this limit, the regular clock is 954 // called. Notice this this might be decreased under certain 955 // circumstances (i.e. when we believe that we did an expensive 956 // operation). 957 size_t _refs_reached_limit; 958 // the initial value of _refs_reached_limit (i.e. what it was before 959 // it was decreased). 960 size_t _real_refs_reached_limit; 961 962 // used by the work stealing stuff 963 int _hash_seed; 964 // if this is true, then the task has aborted for some reason 965 bool _has_aborted; 966 // set when the task aborts because it has met its time quota 967 bool _has_timed_out; 968 // true when we're draining SATB buffers; this avoids the task 969 // aborting due to SATB buffers being available (as we're already 970 // dealing with them) 971 bool _draining_satb_buffers; 972 973 // number sequence of past step times 974 NumberSeq _step_times_ms; 975 // elapsed time of this task 976 double _elapsed_time_ms; 977 // termination time of this task 978 double _termination_time_ms; 979 // when this task got into the termination protocol 980 double _termination_start_time_ms; 981 982 // true when the task is during a concurrent phase, false when it is 983 // in the remark phase (so, in the latter case, we do not have to 984 // check all the things that we have to check during the concurrent 985 // phase, i.e. SATB buffer availability...) 986 bool _concurrent; 987 988 TruncatedSeq _marking_step_diffs_ms; 989 990 // Counting data structures. Embedding the task's marked_bytes_array 991 // and card bitmap into the actual task saves having to go through 992 // the ConcurrentMark object. 993 size_t* _marked_bytes_array; 994 BitMap* _card_bm; 995 996 // LOTS of statistics related with this task 997 #if _MARKING_STATS_ 998 NumberSeq _all_clock_intervals_ms; 999 double _interval_start_time_ms; 1000 1001 size_t _aborted; 1002 size_t _aborted_overflow; 1003 size_t _aborted_cm_aborted; 1004 size_t _aborted_yield; 1005 size_t _aborted_timed_out; 1006 size_t _aborted_satb; 1007 size_t _aborted_termination; 1008 1009 size_t _steal_attempts; 1010 size_t _steals; 1011 1012 size_t _clock_due_to_marking; 1013 size_t _clock_due_to_scanning; 1014 1015 size_t _local_pushes; 1016 size_t _local_pops; 1017 size_t _local_max_size; 1018 size_t _objs_scanned; 1019 1020 size_t _global_pushes; 1021 size_t _global_pops; 1022 size_t _global_max_size; 1023 1024 size_t _global_transfers_to; 1025 size_t _global_transfers_from; 1026 1027 size_t _regions_claimed; 1028 size_t _objs_found_on_bitmap; 1029 1030 size_t _satb_buffers_processed; 1031 #endif // _MARKING_STATS_ 1032 1033 // it updates the local fields after this task has claimed 1034 // a new region to scan 1035 void setup_for_region(HeapRegion* hr); 1036 // it brings up-to-date the limit of the region 1037 void update_region_limit(); 1038 1039 // called when either the words scanned or the refs visited limit 1040 // has been reached 1041 void reached_limit(); 1042 // recalculates the words scanned and refs visited limits 1043 void recalculate_limits(); 1044 // decreases the words scanned and refs visited limits when we reach 1045 // an expensive operation 1046 void decrease_limits(); 1047 // it checks whether the words scanned or refs visited reached their 1048 // respective limit and calls reached_limit() if they have 1049 void check_limits() { 1050 if (_words_scanned >= _words_scanned_limit || 1051 _refs_reached >= _refs_reached_limit) { 1052 reached_limit(); 1053 } 1054 } 1055 // this is supposed to be called regularly during a marking step as 1056 // it checks a bunch of conditions that might cause the marking step 1057 // to abort 1058 void regular_clock_call(); 1059 bool concurrent() { return _concurrent; } 1060 1061 // Test whether obj might have already been passed over by the 1062 // mark bitmap scan, and so needs to be pushed onto the mark stack. 1063 bool is_below_finger(oop obj, HeapWord* global_finger) const; 1064 1065 template<bool scan> void process_grey_object(oop obj); 1066 1067 public: 1068 // It resets the task; it should be called right at the beginning of 1069 // a marking phase. 1070 void reset(CMBitMap* _nextMarkBitMap); 1071 // it clears all the fields that correspond to a claimed region. 1072 void clear_region_fields(); 1073 1074 void set_concurrent(bool concurrent) { _concurrent = concurrent; } 1075 1076 // The main method of this class which performs a marking step 1077 // trying not to exceed the given duration. However, it might exit 1078 // prematurely, according to some conditions (i.e. SATB buffers are 1079 // available for processing). 1080 void do_marking_step(double target_ms, 1081 bool do_termination, 1082 bool is_serial); 1083 1084 // These two calls start and stop the timer 1085 void record_start_time() { 1086 _elapsed_time_ms = os::elapsedTime() * 1000.0; 1087 } 1088 void record_end_time() { 1089 _elapsed_time_ms = os::elapsedTime() * 1000.0 - _elapsed_time_ms; 1090 } 1091 1092 // returns the worker ID associated with this task. 1093 uint worker_id() { return _worker_id; } 1094 1095 // From TerminatorTerminator. It determines whether this task should 1096 // exit the termination protocol after it's entered it. 1097 virtual bool should_exit_termination(); 1098 1099 // Resets the local region fields after a task has finished scanning a 1100 // region; or when they have become stale as a result of the region 1101 // being evacuated. 1102 void giveup_current_region(); 1103 1104 HeapWord* finger() { return _finger; } 1105 1106 bool has_aborted() { return _has_aborted; } 1107 void set_has_aborted() { _has_aborted = true; } 1108 void clear_has_aborted() { _has_aborted = false; } 1109 bool has_timed_out() { return _has_timed_out; } 1110 bool claimed() { return _claimed; } 1111 1112 void set_cm_oop_closure(G1CMOopClosure* cm_oop_closure); 1113 1114 // Increment the number of references this task has visited. 1115 void increment_refs_reached() { ++_refs_reached; } 1116 1117 // Grey the object by marking it. If not already marked, push it on 1118 // the local queue if below the finger. 1119 // Precondition: obj is in region. 1120 // Precondition: obj is below region's NTAMS. 1121 inline void make_reference_grey(oop obj, HeapRegion* region); 1122 1123 // Grey the object (by calling make_grey_reference) if required, 1124 // e.g. obj is below its containing region's NTAMS. 1125 // Precondition: obj is a valid heap object. 1126 inline void deal_with_reference(oop obj); 1127 1128 // It scans an object and visits its children. 1129 inline void scan_object(oop obj); 1130 1131 // It pushes an object on the local queue. 1132 inline void push(oop obj); 1133 1134 // These two move entries to/from the global stack. 1135 void move_entries_to_global_stack(); 1136 void get_entries_from_global_stack(); 1137 1138 // It pops and scans objects from the local queue. If partially is 1139 // true, then it stops when the queue size is of a given limit. If 1140 // partially is false, then it stops when the queue is empty. 1141 void drain_local_queue(bool partially); 1142 // It moves entries from the global stack to the local queue and 1143 // drains the local queue. If partially is true, then it stops when 1144 // both the global stack and the local queue reach a given size. If 1145 // partially if false, it tries to empty them totally. 1146 void drain_global_stack(bool partially); 1147 // It keeps picking SATB buffers and processing them until no SATB 1148 // buffers are available. 1149 void drain_satb_buffers(); 1150 1151 // moves the local finger to a new location 1152 inline void move_finger_to(HeapWord* new_finger) { 1153 assert(new_finger >= _finger && new_finger < _region_limit, "invariant"); 1154 _finger = new_finger; 1155 } 1156 1157 CMTask(uint worker_id, 1158 ConcurrentMark *cm, 1159 size_t* marked_bytes, 1160 BitMap* card_bm, 1161 CMTaskQueue* task_queue, 1162 CMTaskQueueSet* task_queues); 1163 1164 // it prints statistics associated with this task 1165 void print_stats(); 1166 1167 #if _MARKING_STATS_ 1168 void increase_objs_found_on_bitmap() { ++_objs_found_on_bitmap; } 1169 #endif // _MARKING_STATS_ 1170 }; 1171 1172 // Class that's used to to print out per-region liveness 1173 // information. It's currently used at the end of marking and also 1174 // after we sort the old regions at the end of the cleanup operation. 1175 class G1PrintRegionLivenessInfoClosure: public HeapRegionClosure { 1176 private: 1177 outputStream* _out; 1178 1179 // Accumulators for these values. 1180 size_t _total_used_bytes; 1181 size_t _total_capacity_bytes; 1182 size_t _total_prev_live_bytes; 1183 size_t _total_next_live_bytes; 1184 1185 // These are set up when we come across a "stars humongous" region 1186 // (as this is where most of this information is stored, not in the 1187 // subsequent "continues humongous" regions). After that, for every 1188 // region in a given humongous region series we deduce the right 1189 // values for it by simply subtracting the appropriate amount from 1190 // these fields. All these values should reach 0 after we've visited 1191 // the last region in the series. 1192 size_t _hum_used_bytes; 1193 size_t _hum_capacity_bytes; 1194 size_t _hum_prev_live_bytes; 1195 size_t _hum_next_live_bytes; 1196 1197 // Accumulator for the remembered set size 1198 size_t _total_remset_bytes; 1199 1200 // Accumulator for strong code roots memory size 1201 size_t _total_strong_code_roots_bytes; 1202 1203 static double perc(size_t val, size_t total) { 1204 if (total == 0) { 1205 return 0.0; 1206 } else { 1207 return 100.0 * ((double) val / (double) total); 1208 } 1209 } 1210 1211 static double bytes_to_mb(size_t val) { 1212 return (double) val / (double) M; 1213 } 1214 1215 // See the .cpp file. 1216 size_t get_hum_bytes(size_t* hum_bytes); 1217 void get_hum_bytes(size_t* used_bytes, size_t* capacity_bytes, 1218 size_t* prev_live_bytes, size_t* next_live_bytes); 1219 1220 public: 1221 // The header and footer are printed in the constructor and 1222 // destructor respectively. 1223 G1PrintRegionLivenessInfoClosure(outputStream* out, const char* phase_name); 1224 virtual bool doHeapRegion(HeapRegion* r); 1225 ~G1PrintRegionLivenessInfoClosure(); 1226 }; 1227 1228 #endif // SHARE_VM_GC_G1_CONCURRENTMARK_HPP