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 class ConcurrentGCTimer; 38 class G1OldTracer; 39 typedef GenericTaskQueue<oop, mtGC> G1CMTaskQueue; 40 typedef GenericTaskQueueSet<G1CMTaskQueue, mtGC> G1CMTaskQueueSet; 41 42 // Closure used by CM during concurrent reference discovery 43 // and reference processing (during remarking) to determine 44 // if a particular object is alive. It is primarily used 45 // to determine if referents of discovered reference objects 46 // are alive. An instance is also embedded into the 47 // reference processor as the _is_alive_non_header field 48 class G1CMIsAliveClosure: public BoolObjectClosure { 49 G1CollectedHeap* _g1; 50 public: 51 G1CMIsAliveClosure(G1CollectedHeap* g1) : _g1(g1) { } 52 53 bool do_object_b(oop obj); 54 }; 55 56 // A generic CM bit map. This is essentially a wrapper around the BitMap 57 // class, with one bit per (1<<_shifter) HeapWords. 58 59 class G1CMBitMapRO VALUE_OBJ_CLASS_SPEC { 60 protected: 61 HeapWord* _bmStartWord; // base address of range covered by map 62 size_t _bmWordSize; // map size (in #HeapWords covered) 63 const int _shifter; // map to char or bit 64 BitMap _bm; // the bit map itself 65 66 public: 67 // constructor 68 G1CMBitMapRO(int shifter); 69 70 // inquiries 71 HeapWord* startWord() const { return _bmStartWord; } 72 // the following is one past the last word in space 73 HeapWord* endWord() const { return _bmStartWord + _bmWordSize; } 74 75 // read marks 76 77 bool isMarked(HeapWord* addr) const { 78 assert(_bmStartWord <= addr && addr < (_bmStartWord + _bmWordSize), 79 "outside underlying space?"); 80 return _bm.at(heapWordToOffset(addr)); 81 } 82 83 // iteration 84 inline bool iterate(BitMapClosure* cl, MemRegion mr); 85 86 // Return the address corresponding to the next marked bit at or after 87 // "addr", and before "limit", if "limit" is non-NULL. If there is no 88 // such bit, returns "limit" if that is non-NULL, or else "endWord()". 89 HeapWord* getNextMarkedWordAddress(const HeapWord* addr, 90 const HeapWord* limit = NULL) const; 91 92 // conversion utilities 93 HeapWord* offsetToHeapWord(size_t offset) const { 94 return _bmStartWord + (offset << _shifter); 95 } 96 size_t heapWordToOffset(const HeapWord* addr) const { 97 return pointer_delta(addr, _bmStartWord) >> _shifter; 98 } 99 100 // The argument addr should be the start address of a valid object 101 inline HeapWord* nextObject(HeapWord* addr); 102 103 void print_on_error(outputStream* st, const char* prefix) const; 104 105 // debugging 106 NOT_PRODUCT(bool covers(MemRegion rs) const;) 107 }; 108 109 class G1CMBitMapMappingChangedListener : public G1MappingChangedListener { 110 private: 111 G1CMBitMap* _bm; 112 public: 113 G1CMBitMapMappingChangedListener() : _bm(NULL) {} 114 115 void set_bitmap(G1CMBitMap* bm) { _bm = bm; } 116 117 virtual void on_commit(uint start_idx, size_t num_regions, bool zero_filled); 118 }; 119 120 class G1CMBitMap : public G1CMBitMapRO { 121 private: 122 G1CMBitMapMappingChangedListener _listener; 123 124 public: 125 static size_t compute_size(size_t heap_size); 126 // Returns the amount of bytes on the heap between two marks in the bitmap. 127 static size_t mark_distance(); 128 // Returns how many bytes (or bits) of the heap a single byte (or bit) of the 129 // mark bitmap corresponds to. This is the same as the mark distance above. 130 static size_t heap_map_factor() { 131 return mark_distance(); 132 } 133 134 G1CMBitMap() : G1CMBitMapRO(LogMinObjAlignment), _listener() { _listener.set_bitmap(this); } 135 136 // Initializes the underlying BitMap to cover the given area. 137 void initialize(MemRegion heap, G1RegionToSpaceMapper* storage); 138 139 // Write marks. 140 inline void mark(HeapWord* addr); 141 inline void clear(HeapWord* addr); 142 inline bool parMark(HeapWord* addr); 143 144 void clear_range(MemRegion mr); 145 }; 146 147 // Represents a marking stack used by ConcurrentMarking in the G1 collector. 148 class G1CMMarkStack VALUE_OBJ_CLASS_SPEC { 149 VirtualSpace _virtual_space; // Underlying backing store for actual stack 150 G1ConcurrentMark* _cm; 151 oop* _base; // bottom of stack 152 jint _index; // one more than last occupied index 153 jint _capacity; // max #elements 154 jint _saved_index; // value of _index saved at start of GC 155 156 bool _overflow; 157 bool _should_expand; 158 159 public: 160 G1CMMarkStack(G1ConcurrentMark* cm); 161 ~G1CMMarkStack(); 162 163 bool allocate(size_t capacity); 164 165 // Pushes the first "n" elements of "ptr_arr" on the stack. 166 // Locking impl: concurrency is allowed only with 167 // "par_push_arr" and/or "par_pop_arr" operations, which use the same 168 // locking strategy. 169 void par_push_arr(oop* ptr_arr, int n); 170 171 // If returns false, the array was empty. Otherwise, removes up to "max" 172 // elements from the stack, and transfers them to "ptr_arr" in an 173 // unspecified order. The actual number transferred is given in "n" ("n 174 // == 0" is deliberately redundant with the return value.) Locking impl: 175 // concurrency is allowed only with "par_push_arr" and/or "par_pop_arr" 176 // operations, which use the same locking strategy. 177 bool par_pop_arr(oop* ptr_arr, int max, int* n); 178 179 bool isEmpty() { return _index == 0; } 180 int maxElems() { return _capacity; } 181 182 bool overflow() { return _overflow; } 183 void clear_overflow() { _overflow = false; } 184 185 bool should_expand() const { return _should_expand; } 186 void set_should_expand(); 187 188 // Expand the stack, typically in response to an overflow condition 189 void expand(); 190 191 int size() { return _index; } 192 193 void setEmpty() { _index = 0; clear_overflow(); } 194 195 // Record the current index. 196 void note_start_of_gc(); 197 198 // Make sure that we have not added any entries to the stack during GC. 199 void note_end_of_gc(); 200 201 // Apply fn to each oop in the mark stack, up to the bound recorded 202 // via one of the above "note" functions. The mark stack must not 203 // be modified while iterating. 204 template<typename Fn> void iterate(Fn fn); 205 }; 206 207 class YoungList; 208 209 // Root Regions are regions that are not empty at the beginning of a 210 // marking cycle and which we might collect during an evacuation pause 211 // while the cycle is active. Given that, during evacuation pauses, we 212 // do not copy objects that are explicitly marked, what we have to do 213 // for the root regions is to scan them and mark all objects reachable 214 // from them. According to the SATB assumptions, we only need to visit 215 // each object once during marking. So, as long as we finish this scan 216 // before the next evacuation pause, we can copy the objects from the 217 // root regions without having to mark them or do anything else to them. 218 // 219 // Currently, we only support root region scanning once (at the start 220 // of the marking cycle) and the root regions are all the survivor 221 // regions populated during the initial-mark pause. 222 class G1CMRootRegions VALUE_OBJ_CLASS_SPEC { 223 private: 224 YoungList* _young_list; 225 G1ConcurrentMark* _cm; 226 227 volatile bool _scan_in_progress; 228 volatile bool _should_abort; 229 HeapRegion* volatile _next_survivor; 230 231 void notify_scan_done(); 232 233 public: 234 G1CMRootRegions(); 235 // We actually do most of the initialization in this method. 236 void init(G1CollectedHeap* g1h, G1ConcurrentMark* cm); 237 238 // Reset the claiming / scanning of the root regions. 239 void prepare_for_scan(); 240 241 // Forces get_next() to return NULL so that the iteration aborts early. 242 void abort() { _should_abort = true; } 243 244 // Return true if the CM thread are actively scanning root regions, 245 // false otherwise. 246 bool scan_in_progress() { return _scan_in_progress; } 247 248 // Claim the next root region to scan atomically, or return NULL if 249 // all have been claimed. 250 HeapRegion* claim_next(); 251 252 void cancel_scan(); 253 254 // Flag that we're done with root region scanning and notify anyone 255 // who's waiting on it. If aborted is false, assume that all regions 256 // have been claimed. 257 void scan_finished(); 258 259 // If CM threads are still scanning root regions, wait until they 260 // are done. Return true if we had to wait, false otherwise. 261 bool wait_until_scan_finished(); 262 }; 263 264 class ConcurrentMarkThread; 265 266 class G1ConcurrentMark: public CHeapObj<mtGC> { 267 friend class ConcurrentMarkThread; 268 friend class G1ParNoteEndTask; 269 friend class G1VerifyLiveDataClosure; 270 friend class G1CMRefProcTaskProxy; 271 friend class G1CMRefProcTaskExecutor; 272 friend class G1CMKeepAliveAndDrainClosure; 273 friend class G1CMDrainMarkingStackClosure; 274 friend class G1CMBitMapClosure; 275 friend class G1CMConcurrentMarkingTask; 276 friend class G1CMMarkStack; 277 friend class G1CMRemarkTask; 278 friend class G1CMTask; 279 280 protected: 281 ConcurrentMarkThread* _cmThread; // The thread doing the work 282 G1CollectedHeap* _g1h; // The heap 283 uint _parallel_marking_threads; // The number of marking 284 // threads we're using 285 uint _max_parallel_marking_threads; // Max number of marking 286 // threads we'll ever use 287 double _sleep_factor; // How much we have to sleep, with 288 // respect to the work we just did, to 289 // meet the marking overhead goal 290 double _marking_task_overhead; // Marking target overhead for 291 // a single task 292 293 FreeRegionList _cleanup_list; 294 295 // Concurrent marking support structures 296 G1CMBitMap _markBitMap1; 297 G1CMBitMap _markBitMap2; 298 G1CMBitMapRO* _prevMarkBitMap; // Completed mark bitmap 299 G1CMBitMap* _nextMarkBitMap; // Under-construction mark bitmap 300 301 // Liveness count data. After marking G1 iterates over the recently gathered mark 302 // bitmap and records rough information about liveness on card and region basis. 303 // This information can be used for e.g. remembered set scrubbing. 304 305 // A set bit indicates whether the given region contains any live object. 306 BitMap _region_live_bm; 307 // A set bit indicates that the given card contains a live object. 308 BitMap _card_live_bm; 309 310 // Heap bounds 311 HeapWord* _heap_start; 312 HeapWord* _heap_end; 313 314 // Root region tracking and claiming 315 G1CMRootRegions _root_regions; 316 317 // For gray objects 318 G1CMMarkStack _markStack; // Grey objects behind global finger 319 HeapWord* volatile _finger; // The global finger, region aligned, 320 // always points to the end of the 321 // last claimed region 322 323 // Marking tasks 324 uint _max_worker_id;// Maximum worker id 325 uint _active_tasks; // Task num currently active 326 G1CMTask** _tasks; // Task queue array (max_worker_id len) 327 G1CMTaskQueueSet* _task_queues; // Task queue set 328 ParallelTaskTerminator _terminator; // For termination 329 330 // Two sync barriers that are used to synchronize tasks when an 331 // overflow occurs. The algorithm is the following. All tasks enter 332 // the first one to ensure that they have all stopped manipulating 333 // the global data structures. After they exit it, they re-initialize 334 // their data structures and task 0 re-initializes the global data 335 // structures. Then, they enter the second sync barrier. This 336 // ensure, that no task starts doing work before all data 337 // structures (local and global) have been re-initialized. When they 338 // exit it, they are free to start working again. 339 WorkGangBarrierSync _first_overflow_barrier_sync; 340 WorkGangBarrierSync _second_overflow_barrier_sync; 341 342 // This is set by any task, when an overflow on the global data 343 // structures is detected 344 volatile bool _has_overflown; 345 // True: marking is concurrent, false: we're in remark 346 volatile bool _concurrent; 347 // Set at the end of a Full GC so that marking aborts 348 volatile bool _has_aborted; 349 350 // Used when remark aborts due to an overflow to indicate that 351 // another concurrent marking phase should start 352 volatile bool _restart_for_overflow; 353 354 // This is true from the very start of concurrent marking until the 355 // point when all the tasks complete their work. It is really used 356 // to determine the points between the end of concurrent marking and 357 // time of remark. 358 volatile bool _concurrent_marking_in_progress; 359 360 ConcurrentGCTimer* _gc_timer_cm; 361 362 G1OldTracer* _gc_tracer_cm; 363 364 // All of these times are in ms 365 NumberSeq _init_times; 366 NumberSeq _remark_times; 367 NumberSeq _remark_mark_times; 368 NumberSeq _remark_weak_ref_times; 369 NumberSeq _cleanup_times; 370 double _total_counting_time; 371 double _total_rs_scrub_time; 372 373 double* _accum_task_vtime; // Accumulated task vtime 374 375 WorkGang* _parallel_workers; 376 377 void weakRefsWorkParallelPart(BoolObjectClosure* is_alive, bool purged_classes); 378 void weakRefsWork(bool clear_all_soft_refs); 379 380 void swapMarkBitMaps(); 381 382 // Allocates and returns a zero-ed out "large" bitmap of the given size in bits. 383 // It is always allocated using virtual memory. 384 BitMap allocate_large_bitmap(BitMap::idx_t size_in_bits); 385 // Allocates the memory for all bitmaps used by the concurrent marking. 386 void allocate_internal_bitmaps(); 387 // Pre-touches the internal bitmaps. 388 void pretouch_internal_bitmaps(); 389 390 // It resets the global marking data structures, as well as the 391 // task local ones; should be called during initial mark. 392 void reset(); 393 394 // Resets all the marking data structures. Called when we have to restart 395 // marking or when marking completes (via set_non_marking_state below). 396 void reset_marking_state(bool clear_overflow = true); 397 398 // We do this after we're done with marking so that the marking data 399 // structures are initialized to a sensible and predictable state. 400 void set_non_marking_state(); 401 402 // Called to indicate how many threads are currently active. 403 void set_concurrency(uint active_tasks); 404 405 // It should be called to indicate which phase we're in (concurrent 406 // mark or remark) and how many threads are currently active. 407 void set_concurrency_and_phase(uint active_tasks, bool concurrent); 408 409 // Prints all gathered CM-related statistics 410 void print_stats(); 411 412 bool cleanup_list_is_empty() { 413 return _cleanup_list.is_empty(); 414 } 415 416 // Accessor methods 417 uint parallel_marking_threads() const { return _parallel_marking_threads; } 418 uint max_parallel_marking_threads() const { return _max_parallel_marking_threads;} 419 double sleep_factor() { return _sleep_factor; } 420 double marking_task_overhead() { return _marking_task_overhead;} 421 422 HeapWord* finger() { return _finger; } 423 bool concurrent() { return _concurrent; } 424 uint active_tasks() { return _active_tasks; } 425 ParallelTaskTerminator* terminator() { return &_terminator; } 426 427 // It claims the next available region to be scanned by a marking 428 // task/thread. It might return NULL if the next region is empty or 429 // we have run out of regions. In the latter case, out_of_regions() 430 // determines whether we've really run out of regions or the task 431 // should call claim_region() again. This might seem a bit 432 // awkward. Originally, the code was written so that claim_region() 433 // either successfully returned with a non-empty region or there 434 // were no more regions to be claimed. The problem with this was 435 // that, in certain circumstances, it iterated over large chunks of 436 // the heap finding only empty regions and, while it was working, it 437 // was preventing the calling task to call its regular clock 438 // method. So, this way, each task will spend very little time in 439 // claim_region() and is allowed to call the regular clock method 440 // frequently. 441 HeapRegion* claim_region(uint worker_id); 442 443 // It determines whether we've run out of regions to scan. Note that 444 // the finger can point past the heap end in case the heap was expanded 445 // to satisfy an allocation without doing a GC. This is fine, because all 446 // objects in those regions will be considered live anyway because of 447 // SATB guarantees (i.e. their TAMS will be equal to bottom). 448 bool out_of_regions() { return _finger >= _heap_end; } 449 450 // Returns the task with the given id 451 G1CMTask* task(int id) { 452 assert(0 <= id && id < (int) _active_tasks, 453 "task id not within active bounds"); 454 return _tasks[id]; 455 } 456 457 // Returns the task queue with the given id 458 G1CMTaskQueue* task_queue(int id) { 459 assert(0 <= id && id < (int) _active_tasks, 460 "task queue id not within active bounds"); 461 return (G1CMTaskQueue*) _task_queues->queue(id); 462 } 463 464 // Returns the task queue set 465 G1CMTaskQueueSet* task_queues() { return _task_queues; } 466 467 // Access / manipulation of the overflow flag which is set to 468 // indicate that the global stack has overflown 469 bool has_overflown() { return _has_overflown; } 470 void set_has_overflown() { _has_overflown = true; } 471 void clear_has_overflown() { _has_overflown = false; } 472 bool restart_for_overflow() { return _restart_for_overflow; } 473 474 // Methods to enter the two overflow sync barriers 475 void enter_first_sync_barrier(uint worker_id); 476 void enter_second_sync_barrier(uint worker_id); 477 478 // Card index of the bottom of the G1 heap. Used for biasing indices into 479 // the card bitmaps. 480 intptr_t _heap_bottom_card_num; 481 482 // Set to true when initialization is complete 483 bool _completed_initialization; 484 485 // end_timer, true to end gc timer after ending concurrent phase. 486 void register_concurrent_phase_end_common(bool end_timer); 487 488 // Clear the given bitmap in parallel using the given WorkGang. If may_yield is 489 // true, periodically insert checks to see if this method should exit prematurely. 490 void clear_bitmap(G1CMBitMap* bitmap, WorkGang* workers, bool may_yield); 491 public: 492 // Manipulation of the global mark stack. 493 // The push and pop operations are used by tasks for transfers 494 // between task-local queues and the global mark stack, and use 495 // locking for concurrency safety. 496 bool mark_stack_push(oop* arr, int n) { 497 _markStack.par_push_arr(arr, n); 498 if (_markStack.overflow()) { 499 set_has_overflown(); 500 return false; 501 } 502 return true; 503 } 504 void mark_stack_pop(oop* arr, int max, int* n) { 505 _markStack.par_pop_arr(arr, max, n); 506 } 507 size_t mark_stack_size() { return _markStack.size(); } 508 size_t partial_mark_stack_size_target() { return _markStack.maxElems()/3; } 509 bool mark_stack_overflow() { return _markStack.overflow(); } 510 bool mark_stack_empty() { return _markStack.isEmpty(); } 511 512 G1CMRootRegions* root_regions() { return &_root_regions; } 513 514 bool concurrent_marking_in_progress() { 515 return _concurrent_marking_in_progress; 516 } 517 void set_concurrent_marking_in_progress() { 518 _concurrent_marking_in_progress = true; 519 } 520 void clear_concurrent_marking_in_progress() { 521 _concurrent_marking_in_progress = false; 522 } 523 524 void concurrent_cycle_start(); 525 void concurrent_cycle_end(); 526 527 void update_accum_task_vtime(int i, double vtime) { 528 _accum_task_vtime[i] += vtime; 529 } 530 531 double all_task_accum_vtime() { 532 double ret = 0.0; 533 for (uint i = 0; i < _max_worker_id; ++i) 534 ret += _accum_task_vtime[i]; 535 return ret; 536 } 537 538 // Attempts to steal an object from the task queues of other tasks 539 bool try_stealing(uint worker_id, int* hash_seed, oop& obj); 540 541 G1ConcurrentMark(G1CollectedHeap* g1h, 542 G1RegionToSpaceMapper* prev_bitmap_storage, 543 G1RegionToSpaceMapper* next_bitmap_storage); 544 ~G1ConcurrentMark(); 545 546 ConcurrentMarkThread* cmThread() { return _cmThread; } 547 548 G1CMBitMapRO* prevMarkBitMap() const { return _prevMarkBitMap; } 549 G1CMBitMap* nextMarkBitMap() const { return _nextMarkBitMap; } 550 551 // Returns the number of GC threads to be used in a concurrent 552 // phase based on the number of GC threads being used in a STW 553 // phase. 554 uint scale_parallel_threads(uint n_par_threads); 555 556 // Calculates the number of GC threads to be used in a concurrent phase. 557 uint calc_parallel_marking_threads(); 558 559 // The following three are interaction between CM and 560 // G1CollectedHeap 561 562 // This notifies CM that a root during initial-mark needs to be 563 // grayed. It is MT-safe. hr is the region that 564 // contains the object and it's passed optionally from callers who 565 // might already have it (no point in recalculating it). 566 inline void grayRoot(oop obj, 567 HeapRegion* hr = NULL); 568 569 // Prepare internal data structures for the next mark cycle. This includes clearing 570 // the next mark bitmap and some internal data structures. This method is intended 571 // to be called concurrently to the mutator. It will yield to safepoint requests. 572 void cleanup_for_next_mark(); 573 574 // Clear the previous marking bitmap during safepoint. 575 void clear_prev_bitmap(WorkGang* workers); 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 scan_root_regions(); 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 mark_from_roots(); 599 600 void checkpointRootsFinal(bool clear_all_soft_refs); 601 void checkpointRootsFinalWork(); 602 void cleanup(); 603 void complete_cleanup(); 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 // Abandon current marking iteration due to a Full GC. 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 // Attempts to mark the given object on the next mark bitmap. 645 inline bool par_mark(oop obj); 646 647 // Returns true if initialization was successfully completed. 648 bool completed_initialization() const { 649 return _completed_initialization; 650 } 651 652 ConcurrentGCTimer* gc_timer_cm() const { return _gc_timer_cm; } 653 G1OldTracer* gc_tracer_cm() const { return _gc_tracer_cm; } 654 655 private: 656 // Clear (Reset) all liveness count data. 657 void clear_all_live_data(WorkGang* workers); 658 659 // Verify all of the above data structures that they are in initial state. 660 void verify_all_live_data(); 661 662 // Aggregates the per-card liveness data based on the current marking. Also sets 663 // the amount of marked bytes for each region. 664 void create_live_data(); 665 666 // Verification routine 667 void verify_live_data(); 668 }; 669 670 // A class representing a marking task. 671 class G1CMTask : public TerminatorTerminator { 672 private: 673 enum PrivateConstants { 674 // the regular clock call is called once the scanned words reaches 675 // this limit 676 words_scanned_period = 12*1024, 677 // the regular clock call is called once the number of visited 678 // references reaches this limit 679 refs_reached_period = 384, 680 // initial value for the hash seed, used in the work stealing code 681 init_hash_seed = 17, 682 // how many entries will be transferred between global stack and 683 // local queues 684 global_stack_transfer_size = 16 685 }; 686 687 uint _worker_id; 688 G1CollectedHeap* _g1h; 689 G1ConcurrentMark* _cm; 690 G1CMBitMap* _nextMarkBitMap; 691 // the task queue of this task 692 G1CMTaskQueue* _task_queue; 693 private: 694 // the task queue set---needed for stealing 695 G1CMTaskQueueSet* _task_queues; 696 // indicates whether the task has been claimed---this is only for 697 // debugging purposes 698 bool _claimed; 699 700 // number of calls to this task 701 int _calls; 702 703 // when the virtual timer reaches this time, the marking step should 704 // exit 705 double _time_target_ms; 706 // the start time of the current marking step 707 double _start_time_ms; 708 709 // the oop closure used for iterations over oops 710 G1CMOopClosure* _cm_oop_closure; 711 712 // the region this task is scanning, NULL if we're not scanning any 713 HeapRegion* _curr_region; 714 // the local finger of this task, NULL if we're not scanning a region 715 HeapWord* _finger; 716 // limit of the region this task is scanning, NULL if we're not scanning one 717 HeapWord* _region_limit; 718 719 // the number of words this task has scanned 720 size_t _words_scanned; 721 // When _words_scanned reaches this limit, the regular clock is 722 // called. Notice that this might be decreased under certain 723 // circumstances (i.e. when we believe that we did an expensive 724 // operation). 725 size_t _words_scanned_limit; 726 // the initial value of _words_scanned_limit (i.e. what it was 727 // before it was decreased). 728 size_t _real_words_scanned_limit; 729 730 // the number of references this task has visited 731 size_t _refs_reached; 732 // When _refs_reached reaches this limit, the regular clock is 733 // called. Notice this this might be decreased under certain 734 // circumstances (i.e. when we believe that we did an expensive 735 // operation). 736 size_t _refs_reached_limit; 737 // the initial value of _refs_reached_limit (i.e. what it was before 738 // it was decreased). 739 size_t _real_refs_reached_limit; 740 741 // used by the work stealing stuff 742 int _hash_seed; 743 // if this is true, then the task has aborted for some reason 744 bool _has_aborted; 745 // set when the task aborts because it has met its time quota 746 bool _has_timed_out; 747 // true when we're draining SATB buffers; this avoids the task 748 // aborting due to SATB buffers being available (as we're already 749 // dealing with them) 750 bool _draining_satb_buffers; 751 752 // number sequence of past step times 753 NumberSeq _step_times_ms; 754 // elapsed time of this task 755 double _elapsed_time_ms; 756 // termination time of this task 757 double _termination_time_ms; 758 // when this task got into the termination protocol 759 double _termination_start_time_ms; 760 761 // true when the task is during a concurrent phase, false when it is 762 // in the remark phase (so, in the latter case, we do not have to 763 // check all the things that we have to check during the concurrent 764 // phase, i.e. SATB buffer availability...) 765 bool _concurrent; 766 767 TruncatedSeq _marking_step_diffs_ms; 768 769 // it updates the local fields after this task has claimed 770 // a new region to scan 771 void setup_for_region(HeapRegion* hr); 772 // it brings up-to-date the limit of the region 773 void update_region_limit(); 774 775 // called when either the words scanned or the refs visited limit 776 // has been reached 777 void reached_limit(); 778 // recalculates the words scanned and refs visited limits 779 void recalculate_limits(); 780 // decreases the words scanned and refs visited limits when we reach 781 // an expensive operation 782 void decrease_limits(); 783 // it checks whether the words scanned or refs visited reached their 784 // respective limit and calls reached_limit() if they have 785 void check_limits() { 786 if (_words_scanned >= _words_scanned_limit || 787 _refs_reached >= _refs_reached_limit) { 788 reached_limit(); 789 } 790 } 791 // this is supposed to be called regularly during a marking step as 792 // it checks a bunch of conditions that might cause the marking step 793 // to abort 794 void regular_clock_call(); 795 bool concurrent() { return _concurrent; } 796 797 // Test whether obj might have already been passed over by the 798 // mark bitmap scan, and so needs to be pushed onto the mark stack. 799 bool is_below_finger(oop obj, HeapWord* global_finger) const; 800 801 template<bool scan> void process_grey_object(oop obj); 802 803 public: 804 // It resets the task; it should be called right at the beginning of 805 // a marking phase. 806 void reset(G1CMBitMap* _nextMarkBitMap); 807 // it clears all the fields that correspond to a claimed region. 808 void clear_region_fields(); 809 810 void set_concurrent(bool concurrent) { _concurrent = concurrent; } 811 812 // The main method of this class which performs a marking step 813 // trying not to exceed the given duration. However, it might exit 814 // prematurely, according to some conditions (i.e. SATB buffers are 815 // available for processing). 816 void do_marking_step(double target_ms, 817 bool do_termination, 818 bool is_serial); 819 820 // These two calls start and stop the timer 821 void record_start_time() { 822 _elapsed_time_ms = os::elapsedTime() * 1000.0; 823 } 824 void record_end_time() { 825 _elapsed_time_ms = os::elapsedTime() * 1000.0 - _elapsed_time_ms; 826 } 827 828 // returns the worker ID associated with this task. 829 uint worker_id() { return _worker_id; } 830 831 // From TerminatorTerminator. It determines whether this task should 832 // exit the termination protocol after it's entered it. 833 virtual bool should_exit_termination(); 834 835 // Resets the local region fields after a task has finished scanning a 836 // region; or when they have become stale as a result of the region 837 // being evacuated. 838 void giveup_current_region(); 839 840 HeapWord* finger() { return _finger; } 841 842 bool has_aborted() { return _has_aborted; } 843 void set_has_aborted() { _has_aborted = true; } 844 void clear_has_aborted() { _has_aborted = false; } 845 bool has_timed_out() { return _has_timed_out; } 846 bool claimed() { return _claimed; } 847 848 void set_cm_oop_closure(G1CMOopClosure* cm_oop_closure); 849 850 // Increment the number of references this task has visited. 851 void increment_refs_reached() { ++_refs_reached; } 852 853 // Grey the object by marking it. If not already marked, push it on 854 // the local queue if below the finger. 855 // Precondition: obj is below region's NTAMS. 856 inline void make_reference_grey(oop obj); 857 858 // Grey the object (by calling make_grey_reference) if required, 859 // e.g. obj is below its containing region's NTAMS. 860 // Precondition: obj is a valid heap object. 861 inline void deal_with_reference(oop obj); 862 863 // It scans an object and visits its children. 864 inline void scan_object(oop obj); 865 866 // It pushes an object on the local queue. 867 inline void push(oop obj); 868 869 // These two move entries to/from the global stack. 870 void move_entries_to_global_stack(); 871 void get_entries_from_global_stack(); 872 873 // It pops and scans objects from the local queue. If partially is 874 // true, then it stops when the queue size is of a given limit. If 875 // partially is false, then it stops when the queue is empty. 876 void drain_local_queue(bool partially); 877 // It moves entries from the global stack to the local queue and 878 // drains the local queue. If partially is true, then it stops when 879 // both the global stack and the local queue reach a given size. If 880 // partially if false, it tries to empty them totally. 881 void drain_global_stack(bool partially); 882 // It keeps picking SATB buffers and processing them until no SATB 883 // buffers are available. 884 void drain_satb_buffers(); 885 886 // moves the local finger to a new location 887 inline void move_finger_to(HeapWord* new_finger) { 888 assert(new_finger >= _finger && new_finger < _region_limit, "invariant"); 889 _finger = new_finger; 890 } 891 892 G1CMTask(uint worker_id, 893 G1ConcurrentMark *cm, 894 G1CMTaskQueue* task_queue, 895 G1CMTaskQueueSet* task_queues); 896 897 // it prints statistics associated with this task 898 void print_stats(); 899 }; 900 901 // Class that's used to to print out per-region liveness 902 // information. It's currently used at the end of marking and also 903 // after we sort the old regions at the end of the cleanup operation. 904 class G1PrintRegionLivenessInfoClosure: public HeapRegionClosure { 905 private: 906 // Accumulators for these values. 907 size_t _total_used_bytes; 908 size_t _total_capacity_bytes; 909 size_t _total_prev_live_bytes; 910 size_t _total_next_live_bytes; 911 912 // Accumulator for the remembered set size 913 size_t _total_remset_bytes; 914 915 // Accumulator for strong code roots memory size 916 size_t _total_strong_code_roots_bytes; 917 918 static double perc(size_t val, size_t total) { 919 if (total == 0) { 920 return 0.0; 921 } else { 922 return 100.0 * ((double) val / (double) total); 923 } 924 } 925 926 static double bytes_to_mb(size_t val) { 927 return (double) val / (double) M; 928 } 929 930 public: 931 // The header and footer are printed in the constructor and 932 // destructor respectively. 933 G1PrintRegionLivenessInfoClosure(const char* phase_name); 934 virtual bool doHeapRegion(HeapRegion* r); 935 ~G1PrintRegionLivenessInfoClosure(); 936 }; 937 938 #endif // SHARE_VM_GC_G1_G1CONCURRENTMARK_HPP