1 /* 2 * Copyright (c) 2001, 2016, 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_G1CONCURRENTMARK_HPP 26 #define SHARE_VM_GC_G1_G1CONCURRENTMARK_HPP 27 28 #include "classfile/javaClasses.hpp" 29 #include "gc/g1/g1RegionToSpaceMapper.hpp" 30 #include "gc/g1/heapRegionSet.hpp" 31 #include "gc/shared/taskqueue.hpp" 32 33 class G1CollectedHeap; 34 class G1CMBitMap; 35 class G1CMTask; 36 class G1ConcurrentMark; 37 typedef GenericTaskQueue<oop, mtGC> G1CMTaskQueue; 38 typedef GenericTaskQueueSet<G1CMTaskQueue, mtGC> G1CMTaskQueueSet; 39 40 // Closure used by CM during concurrent reference discovery 41 // and reference processing (during remarking) to determine 42 // if a particular object is alive. It is primarily used 43 // to determine if referents of discovered reference objects 44 // are alive. An instance is also embedded into the 45 // reference processor as the _is_alive_non_header field 46 class G1CMIsAliveClosure: public BoolObjectClosure { 47 G1CollectedHeap* _g1; 48 public: 49 G1CMIsAliveClosure(G1CollectedHeap* g1) : _g1(g1) { } 50 51 bool do_object_b(oop obj); 52 }; 53 54 // A generic CM bit map. This is essentially a wrapper around the BitMap 55 // class, with one bit per (1<<_shifter) HeapWords. 56 57 class G1CMBitMapRO VALUE_OBJ_CLASS_SPEC { 58 protected: 59 HeapWord* _bmStartWord; // base address of range covered by map 60 size_t _bmWordSize; // map size (in #HeapWords covered) 61 const int _shifter; // map to char or bit 62 BitMap _bm; // the bit map itself 63 64 public: 65 // constructor 66 G1CMBitMapRO(int shifter); 67 68 // inquiries 69 HeapWord* startWord() const { return _bmStartWord; } 70 // the following is one past the last word in space 71 HeapWord* endWord() const { return _bmStartWord + _bmWordSize; } 72 73 // read marks 74 75 bool isMarked(HeapWord* addr) const { 76 assert(_bmStartWord <= addr && addr < (_bmStartWord + _bmWordSize), 77 "outside underlying space?"); 78 return _bm.at(heapWordToOffset(addr)); 79 } 80 81 // iteration 82 inline bool iterate(BitMapClosure* cl, MemRegion mr); 83 84 // Return the address corresponding to the next marked bit at or after 85 // "addr", and before "limit", if "limit" is non-NULL. If there is no 86 // such bit, returns "limit" if that is non-NULL, or else "endWord()". 87 HeapWord* getNextMarkedWordAddress(const HeapWord* addr, 88 const HeapWord* limit = NULL) const; 89 90 // conversion utilities 91 HeapWord* offsetToHeapWord(size_t offset) const { 92 return _bmStartWord + (offset << _shifter); 93 } 94 size_t heapWordToOffset(const HeapWord* addr) const { 95 return pointer_delta(addr, _bmStartWord) >> _shifter; 96 } 97 98 // The argument addr should be the start address of a valid object 99 inline HeapWord* nextObject(HeapWord* addr); 100 101 void print_on_error(outputStream* st, const char* prefix) const; 102 103 // debugging 104 NOT_PRODUCT(bool covers(MemRegion rs) const;) 105 }; 106 107 class G1CMBitMapMappingChangedListener : public G1MappingChangedListener { 108 private: 109 G1CMBitMap* _bm; 110 public: 111 G1CMBitMapMappingChangedListener() : _bm(NULL) {} 112 113 void set_bitmap(G1CMBitMap* bm) { _bm = bm; } 114 115 virtual void on_commit(uint start_idx, size_t num_regions, bool zero_filled); 116 }; 117 118 class G1CMBitMap : public G1CMBitMapRO { 119 private: 120 G1CMBitMapMappingChangedListener _listener; 121 122 public: 123 static size_t compute_size(size_t heap_size); 124 // Returns the amount of bytes on the heap between two marks in the bitmap. 125 static size_t mark_distance(); 126 // Returns how many bytes (or bits) of the heap a single byte (or bit) of the 127 // mark bitmap corresponds to. This is the same as the mark distance above. 128 static size_t heap_map_factor() { 129 return mark_distance(); 130 } 131 132 G1CMBitMap() : G1CMBitMapRO(LogMinObjAlignment), _listener() { _listener.set_bitmap(this); } 133 134 // Initializes the underlying BitMap to cover the given area. 135 void initialize(MemRegion heap, G1RegionToSpaceMapper* storage); 136 137 // Write marks. 138 inline void mark(HeapWord* addr); 139 inline void clear(HeapWord* addr); 140 inline bool parMark(HeapWord* addr); 141 142 void clearRange(MemRegion mr); 143 144 // Clear the whole mark bitmap. 145 void clearAll(); 146 }; 147 148 // Represents a marking stack used by ConcurrentMarking in the G1 collector. 149 class G1CMMarkStack VALUE_OBJ_CLASS_SPEC { 150 VirtualSpace _virtual_space; // Underlying backing store for actual stack 151 G1ConcurrentMark* _cm; 152 oop* _base; // bottom of stack 153 jint _index; // one more than last occupied index 154 jint _capacity; // max #elements 155 jint _saved_index; // value of _index saved at start of GC 156 157 bool _overflow; 158 bool _should_expand; 159 160 public: 161 G1CMMarkStack(G1ConcurrentMark* cm); 162 ~G1CMMarkStack(); 163 164 bool allocate(size_t capacity); 165 166 // Pushes the first "n" elements of "ptr_arr" on the stack. 167 // Locking impl: concurrency is allowed only with 168 // "par_push_arr" and/or "par_pop_arr" operations, which use the same 169 // locking strategy. 170 void par_push_arr(oop* ptr_arr, int n); 171 172 // If returns false, the array was empty. Otherwise, removes up to "max" 173 // elements from the stack, and transfers them to "ptr_arr" in an 174 // unspecified order. The actual number transferred is given in "n" ("n 175 // == 0" is deliberately redundant with the return value.) Locking impl: 176 // concurrency is allowed only with "par_push_arr" and/or "par_pop_arr" 177 // operations, which use the same locking strategy. 178 bool par_pop_arr(oop* ptr_arr, int max, int* n); 179 180 bool isEmpty() { return _index == 0; } 181 int maxElems() { return _capacity; } 182 183 bool overflow() { return _overflow; } 184 void clear_overflow() { _overflow = false; } 185 186 bool should_expand() const { return _should_expand; } 187 void set_should_expand(); 188 189 // Expand the stack, typically in response to an overflow condition 190 void expand(); 191 192 int size() { return _index; } 193 194 void setEmpty() { _index = 0; clear_overflow(); } 195 196 // Record the current index. 197 void note_start_of_gc(); 198 199 // Make sure that we have not added any entries to the stack during GC. 200 void note_end_of_gc(); 201 202 // Apply fn to each oop in the mark stack, up to the bound recorded 203 // via one of the above "note" functions. The mark stack must not 204 // be modified while iterating. 205 template<typename Fn> void iterate(Fn fn); 206 }; 207 208 class YoungList; 209 210 // Root Regions are regions that are not empty at the beginning of a 211 // marking cycle and which we might collect during an evacuation pause 212 // while the cycle is active. Given that, during evacuation pauses, we 213 // do not copy objects that are explicitly marked, what we have to do 214 // for the root regions is to scan them and mark all objects reachable 215 // from them. According to the SATB assumptions, we only need to visit 216 // each object once during marking. So, as long as we finish this scan 217 // before the next evacuation pause, we can copy the objects from the 218 // root regions without having to mark them or do anything else to them. 219 // 220 // Currently, we only support root region scanning once (at the start 221 // of the marking cycle) and the root regions are all the survivor 222 // regions populated during the initial-mark pause. 223 class G1CMRootRegions VALUE_OBJ_CLASS_SPEC { 224 private: 225 YoungList* _young_list; 226 G1ConcurrentMark* _cm; 227 228 volatile bool _scan_in_progress; 229 volatile bool _should_abort; 230 HeapRegion* volatile _next_survivor; 231 232 void notify_scan_done(); 233 234 public: 235 G1CMRootRegions(); 236 // We actually do most of the initialization in this method. 237 void init(G1CollectedHeap* g1h, G1ConcurrentMark* cm); 238 239 // Reset the claiming / scanning of the root regions. 240 void prepare_for_scan(); 241 242 // Forces get_next() to return NULL so that the iteration aborts early. 243 void abort() { _should_abort = true; } 244 245 // Return true if the CM thread are actively scanning root regions, 246 // false otherwise. 247 bool scan_in_progress() { return _scan_in_progress; } 248 249 // Claim the next root region to scan atomically, or return NULL if 250 // all have been claimed. 251 HeapRegion* claim_next(); 252 253 void cancel_scan(); 254 255 // Flag that we're done with root region scanning and notify anyone 256 // who's waiting on it. If aborted is false, assume that all regions 257 // have been claimed. 258 void scan_finished(); 259 260 // If CM threads are still scanning root regions, wait until they 261 // are done. Return true if we had to wait, false otherwise. 262 bool wait_until_scan_finished(); 263 }; 264 265 class ConcurrentMarkThread; 266 267 class G1ConcurrentMark: public CHeapObj<mtGC> { 268 friend class ConcurrentMarkThread; 269 friend class G1ParNoteEndTask; 270 friend class CalcLiveObjectsClosure; 271 friend class G1CMRefProcTaskProxy; 272 friend class G1CMRefProcTaskExecutor; 273 friend class G1CMKeepAliveAndDrainClosure; 274 friend class G1CMDrainMarkingStackClosure; 275 friend class G1CMBitMapClosure; 276 friend class G1CMConcurrentMarkingTask; 277 friend class G1CMMarkStack; 278 friend class G1CMRemarkTask; 279 friend class G1CMTask; 280 281 protected: 282 ConcurrentMarkThread* _cmThread; // The thread doing the work 283 G1CollectedHeap* _g1h; // The heap 284 uint _parallel_marking_threads; // The number of marking 285 // threads we're using 286 uint _max_parallel_marking_threads; // Max number of marking 287 // threads we'll ever use 288 double _sleep_factor; // How much we have to sleep, with 289 // respect to the work we just did, to 290 // meet the marking overhead goal 291 double _marking_task_overhead; // Marking target overhead for 292 // a single task 293 294 FreeRegionList _cleanup_list; 295 296 // Concurrent marking support structures 297 G1CMBitMap _markBitMap1; 298 G1CMBitMap _markBitMap2; 299 G1CMBitMapRO* _prevMarkBitMap; // Completed mark bitmap 300 G1CMBitMap* _nextMarkBitMap; // Under-construction mark bitmap 301 302 BitMap _region_bm; 303 BitMap _card_bm; 304 305 // Heap bounds 306 HeapWord* _heap_start; 307 HeapWord* _heap_end; 308 309 // Root region tracking and claiming 310 G1CMRootRegions _root_regions; 311 312 // For gray objects 313 G1CMMarkStack _markStack; // Grey objects behind global finger 314 HeapWord* volatile _finger; // The global finger, region aligned, 315 // always points to the end of the 316 // last claimed region 317 318 // Marking tasks 319 uint _max_worker_id;// Maximum worker id 320 uint _active_tasks; // Task num currently active 321 G1CMTask** _tasks; // Task queue array (max_worker_id len) 322 G1CMTaskQueueSet* _task_queues; // Task queue set 323 ParallelTaskTerminator _terminator; // For termination 324 325 // Two sync barriers that are used to synchronize tasks when an 326 // overflow occurs. The algorithm is the following. All tasks enter 327 // the first one to ensure that they have all stopped manipulating 328 // the global data structures. After they exit it, they re-initialize 329 // their data structures and task 0 re-initializes the global data 330 // structures. Then, they enter the second sync barrier. This 331 // ensure, that no task starts doing work before all data 332 // structures (local and global) have been re-initialized. When they 333 // exit it, they are free to start working again. 334 WorkGangBarrierSync _first_overflow_barrier_sync; 335 WorkGangBarrierSync _second_overflow_barrier_sync; 336 337 // This is set by any task, when an overflow on the global data 338 // structures is detected 339 volatile bool _has_overflown; 340 // True: marking is concurrent, false: we're in remark 341 volatile bool _concurrent; 342 // Set at the end of a Full GC so that marking aborts 343 volatile bool _has_aborted; 344 345 // Used when remark aborts due to an overflow to indicate that 346 // another concurrent marking phase should start 347 volatile bool _restart_for_overflow; 348 349 // This is true from the very start of concurrent marking until the 350 // point when all the tasks complete their work. It is really used 351 // to determine the points between the end of concurrent marking and 352 // time of remark. 353 volatile bool _concurrent_marking_in_progress; 354 355 // Keep track of whether we have started concurrent phase or not. 356 bool _concurrent_phase_started; 357 358 // All of these times are in ms 359 NumberSeq _init_times; 360 NumberSeq _remark_times; 361 NumberSeq _remark_mark_times; 362 NumberSeq _remark_weak_ref_times; 363 NumberSeq _cleanup_times; 364 double _total_counting_time; 365 double _total_rs_scrub_time; 366 367 double* _accum_task_vtime; // Accumulated task vtime 368 369 WorkGang* _parallel_workers; 370 371 void weakRefsWorkParallelPart(BoolObjectClosure* is_alive, bool purged_classes); 372 void weakRefsWork(bool clear_all_soft_refs); 373 374 void swapMarkBitMaps(); 375 376 // It resets the global marking data structures, as well as the 377 // task local ones; should be called during initial mark. 378 void reset(); 379 380 // Resets all the marking data structures. Called when we have to restart 381 // marking or when marking completes (via set_non_marking_state below). 382 void reset_marking_state(bool clear_overflow = true); 383 384 // We do this after we're done with marking so that the marking data 385 // structures are initialized to a sensible and predictable state. 386 void set_non_marking_state(); 387 388 // Called to indicate how many threads are currently active. 389 void set_concurrency(uint active_tasks); 390 391 // It should be called to indicate which phase we're in (concurrent 392 // mark or remark) and how many threads are currently active. 393 void set_concurrency_and_phase(uint active_tasks, bool concurrent); 394 395 // Prints all gathered CM-related statistics 396 void print_stats(); 397 398 bool cleanup_list_is_empty() { 399 return _cleanup_list.is_empty(); 400 } 401 402 // Accessor methods 403 uint parallel_marking_threads() const { return _parallel_marking_threads; } 404 uint max_parallel_marking_threads() const { return _max_parallel_marking_threads;} 405 double sleep_factor() { return _sleep_factor; } 406 double marking_task_overhead() { return _marking_task_overhead;} 407 408 HeapWord* finger() { return _finger; } 409 bool concurrent() { return _concurrent; } 410 uint active_tasks() { return _active_tasks; } 411 ParallelTaskTerminator* terminator() { return &_terminator; } 412 413 // It claims the next available region to be scanned by a marking 414 // task/thread. It might return NULL if the next region is empty or 415 // we have run out of regions. In the latter case, out_of_regions() 416 // determines whether we've really run out of regions or the task 417 // should call claim_region() again. This might seem a bit 418 // awkward. Originally, the code was written so that claim_region() 419 // either successfully returned with a non-empty region or there 420 // were no more regions to be claimed. The problem with this was 421 // that, in certain circumstances, it iterated over large chunks of 422 // the heap finding only empty regions and, while it was working, it 423 // was preventing the calling task to call its regular clock 424 // method. So, this way, each task will spend very little time in 425 // claim_region() and is allowed to call the regular clock method 426 // frequently. 427 HeapRegion* claim_region(uint worker_id); 428 429 // It determines whether we've run out of regions to scan. Note that 430 // the finger can point past the heap end in case the heap was expanded 431 // to satisfy an allocation without doing a GC. This is fine, because all 432 // objects in those regions will be considered live anyway because of 433 // SATB guarantees (i.e. their TAMS will be equal to bottom). 434 bool out_of_regions() { return _finger >= _heap_end; } 435 436 // Returns the task with the given id 437 G1CMTask* task(int id) { 438 assert(0 <= id && id < (int) _active_tasks, 439 "task id not within active bounds"); 440 return _tasks[id]; 441 } 442 443 // Returns the task queue with the given id 444 G1CMTaskQueue* task_queue(int id) { 445 assert(0 <= id && id < (int) _active_tasks, 446 "task queue id not within active bounds"); 447 return (G1CMTaskQueue*) _task_queues->queue(id); 448 } 449 450 // Returns the task queue set 451 G1CMTaskQueueSet* task_queues() { return _task_queues; } 452 453 // Access / manipulation of the overflow flag which is set to 454 // indicate that the global stack has overflown 455 bool has_overflown() { return _has_overflown; } 456 void set_has_overflown() { _has_overflown = true; } 457 void clear_has_overflown() { _has_overflown = false; } 458 bool restart_for_overflow() { return _restart_for_overflow; } 459 460 // Methods to enter the two overflow sync barriers 461 void enter_first_sync_barrier(uint worker_id); 462 void enter_second_sync_barrier(uint worker_id); 463 464 // Live Data Counting data structures... 465 // These data structures are initialized at the start of 466 // marking. They are written to while marking is active. 467 // They are aggregated during remark; the aggregated values 468 // are then used to populate the _region_bm, _card_bm, and 469 // the total live bytes, which are then subsequently updated 470 // during cleanup. 471 472 // An array of bitmaps (one bit map per task). Each bitmap 473 // is used to record the cards spanned by the live objects 474 // marked by that task/worker. 475 BitMap* _count_card_bitmaps; 476 477 // Used to record the number of marked live bytes 478 // (for each region, by worker thread). 479 size_t** _count_marked_bytes; 480 481 // Card index of the bottom of the G1 heap. Used for biasing indices into 482 // the card bitmaps. 483 intptr_t _heap_bottom_card_num; 484 485 // Set to true when initialization is complete 486 bool _completed_initialization; 487 488 public: 489 // Manipulation of the global mark stack. 490 // The push and pop operations are used by tasks for transfers 491 // between task-local queues and the global mark stack, and use 492 // locking for concurrency safety. 493 bool mark_stack_push(oop* arr, int n) { 494 _markStack.par_push_arr(arr, n); 495 if (_markStack.overflow()) { 496 set_has_overflown(); 497 return false; 498 } 499 return true; 500 } 501 void mark_stack_pop(oop* arr, int max, int* n) { 502 _markStack.par_pop_arr(arr, max, n); 503 } 504 size_t mark_stack_size() { return _markStack.size(); } 505 size_t partial_mark_stack_size_target() { return _markStack.maxElems()/3; } 506 bool mark_stack_overflow() { return _markStack.overflow(); } 507 bool mark_stack_empty() { return _markStack.isEmpty(); } 508 509 G1CMRootRegions* root_regions() { return &_root_regions; } 510 511 bool concurrent_marking_in_progress() { 512 return _concurrent_marking_in_progress; 513 } 514 void set_concurrent_marking_in_progress() { 515 _concurrent_marking_in_progress = true; 516 } 517 void clear_concurrent_marking_in_progress() { 518 _concurrent_marking_in_progress = false; 519 } 520 521 void register_concurrent_phase_start(const char* title); 522 void register_concurrent_phase_end(); 523 524 void update_accum_task_vtime(int i, double vtime) { 525 _accum_task_vtime[i] += vtime; 526 } 527 528 double all_task_accum_vtime() { 529 double ret = 0.0; 530 for (uint i = 0; i < _max_worker_id; ++i) 531 ret += _accum_task_vtime[i]; 532 return ret; 533 } 534 535 // Attempts to steal an object from the task queues of other tasks 536 bool try_stealing(uint worker_id, int* hash_seed, oop& obj); 537 538 G1ConcurrentMark(G1CollectedHeap* g1h, 539 G1RegionToSpaceMapper* prev_bitmap_storage, 540 G1RegionToSpaceMapper* next_bitmap_storage); 541 ~G1ConcurrentMark(); 542 543 ConcurrentMarkThread* cmThread() { return _cmThread; } 544 545 G1CMBitMapRO* prevMarkBitMap() const { return _prevMarkBitMap; } 546 G1CMBitMap* nextMarkBitMap() const { return _nextMarkBitMap; } 547 548 // Returns the number of GC threads to be used in a concurrent 549 // phase based on the number of GC threads being used in a STW 550 // phase. 551 uint scale_parallel_threads(uint n_par_threads); 552 553 // Calculates the number of GC threads to be used in a concurrent phase. 554 uint calc_parallel_marking_threads(); 555 556 // The following three are interaction between CM and 557 // G1CollectedHeap 558 559 // This notifies CM that a root during initial-mark needs to be 560 // grayed. It is MT-safe. word_size is the size of the object in 561 // words. It is passed explicitly as sometimes we cannot calculate 562 // it from the given object because it might be in an inconsistent 563 // state (e.g., in to-space and being copied). So the caller is 564 // responsible for dealing with this issue (e.g., get the size from 565 // the from-space image when the to-space image might be 566 // inconsistent) and always passing the size. hr is the region that 567 // contains the object and it's passed optionally from callers who 568 // might already have it (no point in recalculating it). 569 inline void grayRoot(oop obj, 570 size_t word_size, 571 uint worker_id, 572 HeapRegion* hr = NULL); 573 574 // Clear the next marking bitmap (will be called concurrently). 575 void clearNextBitmap(); 576 577 // Return whether the next mark bitmap has no marks set. To be used for assertions 578 // only. Will not yield to pause requests. 579 bool nextMarkBitmapIsClear(); 580 581 // These two do the work that needs to be done before and after the 582 // initial root checkpoint. Since this checkpoint can be done at two 583 // different points (i.e. an explicit pause or piggy-backed on a 584 // young collection), then it's nice to be able to easily share the 585 // pre/post code. It might be the case that we can put everything in 586 // the post method. TP 587 void checkpointRootsInitialPre(); 588 void checkpointRootsInitialPost(); 589 590 // Scan all the root regions and mark everything reachable from 591 // them. 592 void scanRootRegions(); 593 594 // Scan a single root region and mark everything reachable from it. 595 void scanRootRegion(HeapRegion* hr, uint worker_id); 596 597 // Do concurrent phase of marking, to a tentative transitive closure. 598 void markFromRoots(); 599 600 void checkpointRootsFinal(bool clear_all_soft_refs); 601 void checkpointRootsFinalWork(); 602 void cleanup(); 603 void completeCleanup(); 604 605 // Mark in the previous bitmap. NB: this is usually read-only, so use 606 // this carefully! 607 inline void markPrev(oop p); 608 609 // Clears marks for all objects in the given range, for the prev or 610 // next bitmaps. NB: the previous bitmap is usually 611 // read-only, so use this carefully! 612 void clearRangePrevBitmap(MemRegion mr); 613 614 // Notify data structures that a GC has started. 615 void note_start_of_gc() { 616 _markStack.note_start_of_gc(); 617 } 618 619 // Notify data structures that a GC is finished. 620 void note_end_of_gc() { 621 _markStack.note_end_of_gc(); 622 } 623 624 // Verify that there are no CSet oops on the stacks (taskqueues / 625 // global mark stack) and fingers (global / per-task). 626 // If marking is not in progress, it's a no-op. 627 void verify_no_cset_oops() PRODUCT_RETURN; 628 629 inline bool isPrevMarked(oop p) const; 630 631 inline bool do_yield_check(uint worker_i = 0); 632 633 // Called to abort the marking cycle after a Full GC takes place. 634 void abort(); 635 636 bool has_aborted() { return _has_aborted; } 637 638 void print_summary_info(); 639 640 void print_worker_threads_on(outputStream* st) const; 641 642 void print_on_error(outputStream* st) const; 643 644 // Liveness counting 645 646 // Utility routine to set an exclusive range of cards on the given 647 // card liveness bitmap 648 inline void set_card_bitmap_range(BitMap* card_bm, 649 BitMap::idx_t start_idx, 650 BitMap::idx_t end_idx, 651 bool is_par); 652 653 // Returns the card number of the bottom of the G1 heap. 654 // Used in biasing indices into accounting card bitmaps. 655 intptr_t heap_bottom_card_num() const { 656 return _heap_bottom_card_num; 657 } 658 659 // Returns the card bitmap for a given task or worker id. 660 BitMap* count_card_bitmap_for(uint worker_id) { 661 assert(worker_id < _max_worker_id, "oob"); 662 assert(_count_card_bitmaps != NULL, "uninitialized"); 663 BitMap* task_card_bm = &_count_card_bitmaps[worker_id]; 664 assert(task_card_bm->size() == _card_bm.size(), "size mismatch"); 665 return task_card_bm; 666 } 667 668 // Returns the array containing the marked bytes for each region, 669 // for the given worker or task id. 670 size_t* count_marked_bytes_array_for(uint worker_id) { 671 assert(worker_id < _max_worker_id, "oob"); 672 assert(_count_marked_bytes != NULL, "uninitialized"); 673 size_t* marked_bytes_array = _count_marked_bytes[worker_id]; 674 assert(marked_bytes_array != NULL, "uninitialized"); 675 return marked_bytes_array; 676 } 677 678 // Returns the index in the liveness accounting card table bitmap 679 // for the given address 680 inline BitMap::idx_t card_bitmap_index_for(HeapWord* addr); 681 682 // Counts the size of the given memory region in the the given 683 // marked_bytes array slot for the given HeapRegion. 684 // Sets the bits in the given card bitmap that are associated with the 685 // cards that are spanned by the memory region. 686 inline void count_region(MemRegion mr, 687 HeapRegion* hr, 688 size_t* marked_bytes_array, 689 BitMap* task_card_bm); 690 691 // Counts the given object in the given task/worker counting 692 // data structures. 693 inline void count_object(oop obj, 694 HeapRegion* hr, 695 size_t* marked_bytes_array, 696 BitMap* task_card_bm, 697 size_t word_size); 698 699 // Attempts to mark the given object and, if successful, counts 700 // the object in the given task/worker counting structures. 701 inline bool par_mark_and_count(oop obj, 702 HeapRegion* hr, 703 size_t* marked_bytes_array, 704 BitMap* task_card_bm); 705 706 // Attempts to mark the given object and, if successful, counts 707 // the object in the task/worker counting structures for the 708 // given worker id. 709 inline bool par_mark_and_count(oop obj, 710 size_t word_size, 711 HeapRegion* hr, 712 uint worker_id); 713 714 // Returns true if initialization was successfully completed. 715 bool completed_initialization() const { 716 return _completed_initialization; 717 } 718 719 protected: 720 // Clear all the per-task bitmaps and arrays used to store the 721 // counting data. 722 void clear_all_count_data(); 723 724 // Aggregates the counting data for each worker/task 725 // that was constructed while marking. Also sets 726 // the amount of marked bytes for each region and 727 // the top at concurrent mark count. 728 void aggregate_count_data(); 729 730 // Verification routine 731 void verify_count_data(); 732 }; 733 734 // A class representing a marking task. 735 class G1CMTask : public TerminatorTerminator { 736 private: 737 enum PrivateConstants { 738 // the regular clock call is called once the scanned words reaches 739 // this limit 740 words_scanned_period = 12*1024, 741 // the regular clock call is called once the number of visited 742 // references reaches this limit 743 refs_reached_period = 384, 744 // initial value for the hash seed, used in the work stealing code 745 init_hash_seed = 17, 746 // how many entries will be transferred between global stack and 747 // local queues 748 global_stack_transfer_size = 16 749 }; 750 751 uint _worker_id; 752 G1CollectedHeap* _g1h; 753 G1ConcurrentMark* _cm; 754 G1CMBitMap* _nextMarkBitMap; 755 // the task queue of this task 756 G1CMTaskQueue* _task_queue; 757 private: 758 // the task queue set---needed for stealing 759 G1CMTaskQueueSet* _task_queues; 760 // indicates whether the task has been claimed---this is only for 761 // debugging purposes 762 bool _claimed; 763 764 // number of calls to this task 765 int _calls; 766 767 // when the virtual timer reaches this time, the marking step should 768 // exit 769 double _time_target_ms; 770 // the start time of the current marking step 771 double _start_time_ms; 772 773 // the oop closure used for iterations over oops 774 G1CMOopClosure* _cm_oop_closure; 775 776 // the region this task is scanning, NULL if we're not scanning any 777 HeapRegion* _curr_region; 778 // the local finger of this task, NULL if we're not scanning a region 779 HeapWord* _finger; 780 // limit of the region this task is scanning, NULL if we're not scanning one 781 HeapWord* _region_limit; 782 783 // the number of words this task has scanned 784 size_t _words_scanned; 785 // When _words_scanned reaches this limit, the regular clock is 786 // called. Notice that this might be decreased under certain 787 // circumstances (i.e. when we believe that we did an expensive 788 // operation). 789 size_t _words_scanned_limit; 790 // the initial value of _words_scanned_limit (i.e. what it was 791 // before it was decreased). 792 size_t _real_words_scanned_limit; 793 794 // the number of references this task has visited 795 size_t _refs_reached; 796 // When _refs_reached reaches this limit, the regular clock is 797 // called. Notice this this might be decreased under certain 798 // circumstances (i.e. when we believe that we did an expensive 799 // operation). 800 size_t _refs_reached_limit; 801 // the initial value of _refs_reached_limit (i.e. what it was before 802 // it was decreased). 803 size_t _real_refs_reached_limit; 804 805 // used by the work stealing stuff 806 int _hash_seed; 807 // if this is true, then the task has aborted for some reason 808 bool _has_aborted; 809 // set when the task aborts because it has met its time quota 810 bool _has_timed_out; 811 // true when we're draining SATB buffers; this avoids the task 812 // aborting due to SATB buffers being available (as we're already 813 // dealing with them) 814 bool _draining_satb_buffers; 815 816 // number sequence of past step times 817 NumberSeq _step_times_ms; 818 // elapsed time of this task 819 double _elapsed_time_ms; 820 // termination time of this task 821 double _termination_time_ms; 822 // when this task got into the termination protocol 823 double _termination_start_time_ms; 824 825 // true when the task is during a concurrent phase, false when it is 826 // in the remark phase (so, in the latter case, we do not have to 827 // check all the things that we have to check during the concurrent 828 // phase, i.e. SATB buffer availability...) 829 bool _concurrent; 830 831 TruncatedSeq _marking_step_diffs_ms; 832 833 // Counting data structures. Embedding the task's marked_bytes_array 834 // and card bitmap into the actual task saves having to go through 835 // the ConcurrentMark object. 836 size_t* _marked_bytes_array; 837 BitMap* _card_bm; 838 839 // it updates the local fields after this task has claimed 840 // a new region to scan 841 void setup_for_region(HeapRegion* hr); 842 // it brings up-to-date the limit of the region 843 void update_region_limit(); 844 845 // called when either the words scanned or the refs visited limit 846 // has been reached 847 void reached_limit(); 848 // recalculates the words scanned and refs visited limits 849 void recalculate_limits(); 850 // decreases the words scanned and refs visited limits when we reach 851 // an expensive operation 852 void decrease_limits(); 853 // it checks whether the words scanned or refs visited reached their 854 // respective limit and calls reached_limit() if they have 855 void check_limits() { 856 if (_words_scanned >= _words_scanned_limit || 857 _refs_reached >= _refs_reached_limit) { 858 reached_limit(); 859 } 860 } 861 // this is supposed to be called regularly during a marking step as 862 // it checks a bunch of conditions that might cause the marking step 863 // to abort 864 void regular_clock_call(); 865 bool concurrent() { return _concurrent; } 866 867 // Test whether obj might have already been passed over by the 868 // mark bitmap scan, and so needs to be pushed onto the mark stack. 869 bool is_below_finger(oop obj, HeapWord* global_finger) const; 870 871 template<bool scan> void process_grey_object(oop obj); 872 873 public: 874 // It resets the task; it should be called right at the beginning of 875 // a marking phase. 876 void reset(G1CMBitMap* _nextMarkBitMap); 877 // it clears all the fields that correspond to a claimed region. 878 void clear_region_fields(); 879 880 void set_concurrent(bool concurrent) { _concurrent = concurrent; } 881 882 // The main method of this class which performs a marking step 883 // trying not to exceed the given duration. However, it might exit 884 // prematurely, according to some conditions (i.e. SATB buffers are 885 // available for processing). 886 void do_marking_step(double target_ms, 887 bool do_termination, 888 bool is_serial); 889 890 // These two calls start and stop the timer 891 void record_start_time() { 892 _elapsed_time_ms = os::elapsedTime() * 1000.0; 893 } 894 void record_end_time() { 895 _elapsed_time_ms = os::elapsedTime() * 1000.0 - _elapsed_time_ms; 896 } 897 898 // returns the worker ID associated with this task. 899 uint worker_id() { return _worker_id; } 900 901 // From TerminatorTerminator. It determines whether this task should 902 // exit the termination protocol after it's entered it. 903 virtual bool should_exit_termination(); 904 905 // Resets the local region fields after a task has finished scanning a 906 // region; or when they have become stale as a result of the region 907 // being evacuated. 908 void giveup_current_region(); 909 910 HeapWord* finger() { return _finger; } 911 912 bool has_aborted() { return _has_aborted; } 913 void set_has_aborted() { _has_aborted = true; } 914 void clear_has_aborted() { _has_aborted = false; } 915 bool has_timed_out() { return _has_timed_out; } 916 bool claimed() { return _claimed; } 917 918 void set_cm_oop_closure(G1CMOopClosure* cm_oop_closure); 919 920 // Increment the number of references this task has visited. 921 void increment_refs_reached() { ++_refs_reached; } 922 923 // Grey the object by marking it. If not already marked, push it on 924 // the local queue if below the finger. 925 // Precondition: obj is in region. 926 // Precondition: obj is below region's NTAMS. 927 inline void make_reference_grey(oop obj, HeapRegion* region); 928 929 // Grey the object (by calling make_grey_reference) if required, 930 // e.g. obj is below its containing region's NTAMS. 931 // Precondition: obj is a valid heap object. 932 inline void deal_with_reference(oop obj); 933 934 // It scans an object and visits its children. 935 inline void scan_object(oop obj); 936 937 // It pushes an object on the local queue. 938 inline void push(oop obj); 939 940 // These two move entries to/from the global stack. 941 void move_entries_to_global_stack(); 942 void get_entries_from_global_stack(); 943 944 // It pops and scans objects from the local queue. If partially is 945 // true, then it stops when the queue size is of a given limit. If 946 // partially is false, then it stops when the queue is empty. 947 void drain_local_queue(bool partially); 948 // It moves entries from the global stack to the local queue and 949 // drains the local queue. If partially is true, then it stops when 950 // both the global stack and the local queue reach a given size. If 951 // partially if false, it tries to empty them totally. 952 void drain_global_stack(bool partially); 953 // It keeps picking SATB buffers and processing them until no SATB 954 // buffers are available. 955 void drain_satb_buffers(); 956 957 // moves the local finger to a new location 958 inline void move_finger_to(HeapWord* new_finger) { 959 assert(new_finger >= _finger && new_finger < _region_limit, "invariant"); 960 _finger = new_finger; 961 } 962 963 G1CMTask(uint worker_id, 964 G1ConcurrentMark *cm, 965 size_t* marked_bytes, 966 BitMap* card_bm, 967 G1CMTaskQueue* task_queue, 968 G1CMTaskQueueSet* task_queues); 969 970 // it prints statistics associated with this task 971 void print_stats(); 972 }; 973 974 // Class that's used to to print out per-region liveness 975 // information. It's currently used at the end of marking and also 976 // after we sort the old regions at the end of the cleanup operation. 977 class G1PrintRegionLivenessInfoClosure: public HeapRegionClosure { 978 private: 979 // Accumulators for these values. 980 size_t _total_used_bytes; 981 size_t _total_capacity_bytes; 982 size_t _total_prev_live_bytes; 983 size_t _total_next_live_bytes; 984 985 // These are set up when we come across a "stars humongous" region 986 // (as this is where most of this information is stored, not in the 987 // subsequent "continues humongous" regions). After that, for every 988 // region in a given humongous region series we deduce the right 989 // values for it by simply subtracting the appropriate amount from 990 // these fields. All these values should reach 0 after we've visited 991 // the last region in the series. 992 size_t _hum_used_bytes; 993 size_t _hum_capacity_bytes; 994 size_t _hum_prev_live_bytes; 995 size_t _hum_next_live_bytes; 996 997 // Accumulator for the remembered set size 998 size_t _total_remset_bytes; 999 1000 // Accumulator for strong code roots memory size 1001 size_t _total_strong_code_roots_bytes; 1002 1003 static double perc(size_t val, size_t total) { 1004 if (total == 0) { 1005 return 0.0; 1006 } else { 1007 return 100.0 * ((double) val / (double) total); 1008 } 1009 } 1010 1011 static double bytes_to_mb(size_t val) { 1012 return (double) val / (double) M; 1013 } 1014 1015 // See the .cpp file. 1016 size_t get_hum_bytes(size_t* hum_bytes); 1017 void get_hum_bytes(size_t* used_bytes, size_t* capacity_bytes, 1018 size_t* prev_live_bytes, size_t* next_live_bytes); 1019 1020 public: 1021 // The header and footer are printed in the constructor and 1022 // destructor respectively. 1023 G1PrintRegionLivenessInfoClosure(const char* phase_name); 1024 virtual bool doHeapRegion(HeapRegion* r); 1025 ~G1PrintRegionLivenessInfoClosure(); 1026 }; 1027 1028 #endif // SHARE_VM_GC_G1_G1CONCURRENTMARK_HPP