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