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/g1ConcurrentMarkObjArrayProcessor.hpp" 30 #include "gc/g1/g1RegionToSpaceMapper.hpp" 31 #include "gc/g1/heapRegionSet.hpp" 32 #include "gc/shared/taskqueue.hpp" 33 34 class G1CollectedHeap; 35 class G1CMBitMap; 36 class G1CMTask; 37 class G1ConcurrentMark; 38 class ConcurrentGCTimer; 39 class G1OldTracer; 40 class G1SurvivorRegions; 41 typedef GenericTaskQueue<oop, mtGC> G1CMTaskQueue; 42 typedef GenericTaskQueueSet<G1CMTaskQueue, mtGC> G1CMTaskQueueSet; 43 44 // Closure used by CM during concurrent reference discovery 45 // and reference processing (during remarking) to determine 46 // if a particular object is alive. It is primarily used 47 // to determine if referents of discovered reference objects 48 // are alive. An instance is also embedded into the 49 // reference processor as the _is_alive_non_header field 50 class G1CMIsAliveClosure: public BoolObjectClosure { 51 G1CollectedHeap* _g1; 52 public: 53 G1CMIsAliveClosure(G1CollectedHeap* g1) : _g1(g1) { } 54 55 bool do_object_b(oop obj); 56 }; 57 58 // A generic CM bit map. This is essentially a wrapper around the BitMap 59 // class, with one bit per (1<<_shifter) HeapWords. 60 61 class G1CMBitMapRO VALUE_OBJ_CLASS_SPEC { 62 protected: 63 HeapWord* _bmStartWord; // base address of range covered by map 64 size_t _bmWordSize; // map size (in #HeapWords covered) 65 const int _shifter; // map to char or bit 66 BitMapView _bm; // the bit map itself 67 68 public: 69 // constructor 70 G1CMBitMapRO(int shifter); 71 72 // inquiries 73 HeapWord* startWord() const { return _bmStartWord; } 74 // the following is one past the last word in space 75 HeapWord* endWord() const { return _bmStartWord + _bmWordSize; } 76 77 // read marks 78 79 bool isMarked(HeapWord* addr) const { 80 assert(_bmStartWord <= addr && addr < (_bmStartWord + _bmWordSize), 81 "outside underlying space?"); 82 return _bm.at(heapWordToOffset(addr)); 83 } 84 85 // iteration 86 inline bool iterate(BitMapClosure* cl, MemRegion mr); 87 88 // Return the address corresponding to the next marked bit at or after 89 // "addr", and before "limit", if "limit" is non-NULL. If there is no 90 // such bit, returns "limit" if that is non-NULL, or else "endWord()". 91 HeapWord* getNextMarkedWordAddress(const HeapWord* addr, 92 const HeapWord* limit = NULL) const; 93 94 // conversion utilities 95 HeapWord* offsetToHeapWord(size_t offset) const { 96 return _bmStartWord + (offset << _shifter); 97 } 98 size_t heapWordToOffset(const HeapWord* addr) const { 99 return pointer_delta(addr, _bmStartWord) >> _shifter; 100 } 101 102 // The argument addr should be the start address of a valid object 103 inline HeapWord* nextObject(HeapWord* addr); 104 105 void print_on_error(outputStream* st, const char* prefix) const; 106 107 // debugging 108 NOT_PRODUCT(bool covers(MemRegion rs) const;) 109 }; 110 111 class G1CMBitMapMappingChangedListener : public G1MappingChangedListener { 112 private: 113 G1CMBitMap* _bm; 114 public: 115 G1CMBitMapMappingChangedListener() : _bm(NULL) {} 116 117 void set_bitmap(G1CMBitMap* bm) { _bm = bm; } 118 119 virtual void on_commit(uint start_idx, size_t num_regions, bool zero_filled); 120 }; 121 122 class G1CMBitMap : public G1CMBitMapRO { 123 private: 124 G1CMBitMapMappingChangedListener _listener; 125 126 public: 127 static size_t compute_size(size_t heap_size); 128 // Returns the amount of bytes on the heap between two marks in the bitmap. 129 static size_t mark_distance(); 130 // Returns how many bytes (or bits) of the heap a single byte (or bit) of the 131 // mark bitmap corresponds to. This is the same as the mark distance above. 132 static size_t heap_map_factor() { 133 return mark_distance(); 134 } 135 136 G1CMBitMap() : G1CMBitMapRO(LogMinObjAlignment), _listener() { _listener.set_bitmap(this); } 137 138 // Initializes the underlying BitMap to cover the given area. 139 void initialize(MemRegion heap, G1RegionToSpaceMapper* storage); 140 141 // Write marks. 142 inline void mark(HeapWord* addr); 143 inline void clear(HeapWord* addr); 144 inline bool parMark(HeapWord* addr); 145 146 void clear_range(MemRegion mr); 147 }; 148 149 // Represents the overflow mark stack used by concurrent marking. 150 // 151 // Stores oops in a huge buffer in virtual memory that is always fully committed. 152 // Resizing may only happen during a STW pause when the stack is empty. 153 // 154 // Memory is allocated on a "chunk" basis, i.e. a set of oops. For this, the mark 155 // stack memory is split into evenly sized chunks of oops. Users can only 156 // add or remove entries on that basis. 157 // Chunks are filled in increasing address order. Not completely filled chunks 158 // have a NULL element as a terminating element. 159 // 160 // Every chunk has a header containing a single pointer element used for memory 161 // management. This wastes some space, but is negligible (< .1% with current sizing). 162 // 163 // Memory management is done using a mix of tracking a high water-mark indicating 164 // that all chunks at a lower address are valid chunks, and a singly linked free 165 // list connecting all empty chunks. 166 class G1CMMarkStack VALUE_OBJ_CLASS_SPEC { 167 public: 168 // Number of oops that can fit in a single chunk. 169 static const size_t OopsPerChunk = 1024 - 1 /* One reference for the next pointer */; 170 private: 171 struct OopChunk { 172 OopChunk* next; 173 oop data[OopsPerChunk]; 174 }; 175 176 size_t _max_chunk_capacity; // Maximum number of OopChunk elements on the stack. 177 178 OopChunk* _base; // Bottom address of allocated memory area. 179 size_t _chunk_capacity; // Current maximum number of OopChunk elements. 180 181 char _pad0[DEFAULT_CACHE_LINE_SIZE]; 182 OopChunk* volatile _free_list; // Linked list of free chunks that can be allocated by users. 183 char _pad1[DEFAULT_CACHE_LINE_SIZE - sizeof(OopChunk*)]; 184 OopChunk* volatile _chunk_list; // List of chunks currently containing data. 185 volatile size_t _chunks_in_chunk_list; 186 char _pad2[DEFAULT_CACHE_LINE_SIZE - sizeof(OopChunk*) - sizeof(size_t)]; 187 188 volatile size_t _hwm; // High water mark within the reserved space. 189 char _pad4[DEFAULT_CACHE_LINE_SIZE - sizeof(size_t)]; 190 191 // Allocate a new chunk from the reserved memory, using the high water mark. Returns 192 // NULL if out of memory. 193 OopChunk* allocate_new_chunk(); 194 195 // This is set by any task, when an overflow on the global data 196 // structures is detected 197 volatile bool _out_of_memory; 198 199 // Atomically add the given chunk to the list. 200 void add_chunk_to_list(OopChunk* volatile* list, OopChunk* elem); 201 // Atomically remove and return a chunk from the given list. Returns NULL if the 202 // list is empty. 203 OopChunk* remove_chunk_from_list(OopChunk* volatile* list); 204 205 void add_chunk_to_chunk_list(OopChunk* elem); 206 void add_chunk_to_free_list(OopChunk* elem); 207 208 OopChunk* remove_chunk_from_chunk_list(); 209 OopChunk* remove_chunk_from_free_list(); 210 211 bool _should_expand; 212 213 // Resizes the mark stack to the given new capacity. Releases any previous 214 // memory if successful. 215 bool resize(size_t new_capacity); 216 217 public: 218 G1CMMarkStack(); 219 ~G1CMMarkStack(); 220 221 // Alignment and minimum capacity of this mark stack in number of oops. 222 static size_t capacity_alignment(); 223 224 // Allocate and initialize the mark stack with the given number of oops. 225 bool initialize(size_t initial_capacity, size_t max_capacity); 226 227 // Pushes the given buffer containing at most OopsPerChunk elements on the mark 228 // stack. If less than OopsPerChunk elements are to be pushed, the array must 229 // be terminated with a NULL. 230 // Returns whether the buffer contents were successfully pushed to the global mark 231 // stack. 232 bool par_push_chunk(oop* buffer); 233 234 // Pops a chunk from this mark stack, copying them into the given buffer. This 235 // chunk may contain up to OopsPerChunk elements. If there are less, the last 236 // element in the array is a NULL pointer. 237 bool par_pop_chunk(oop* buffer); 238 239 // Return whether the chunk list is empty. Racy due to unsynchronized access to 240 // _chunk_list. 241 bool is_empty() const { return _chunk_list == NULL; } 242 243 size_t capacity() const { return _chunk_capacity; } 244 245 bool is_out_of_memory() const { return _out_of_memory; } 246 void clear_out_of_memory() { _out_of_memory = false; } 247 248 bool should_expand() const { return _should_expand; } 249 void set_should_expand(bool value) { _should_expand = value; } 250 251 // Expand the stack, typically in response to an overflow condition 252 void expand(); 253 254 // Return the approximate number of oops on this mark stack. Racy due to 255 // unsynchronized access to _chunks_in_chunk_list. 256 size_t size() const { return _chunks_in_chunk_list * OopsPerChunk; } 257 258 void set_empty(); 259 260 // Apply Fn to every oop on the mark stack. The mark stack must not 261 // be modified while iterating. 262 template<typename Fn> void iterate(Fn fn) const PRODUCT_RETURN; 263 }; 264 265 // Root Regions are regions that are not empty at the beginning of a 266 // marking cycle and which we might collect during an evacuation pause 267 // while the cycle is active. Given that, during evacuation pauses, we 268 // do not copy objects that are explicitly marked, what we have to do 269 // for the root regions is to scan them and mark all objects reachable 270 // from them. According to the SATB assumptions, we only need to visit 271 // each object once during marking. So, as long as we finish this scan 272 // before the next evacuation pause, we can copy the objects from the 273 // root regions without having to mark them or do anything else to them. 274 // 275 // Currently, we only support root region scanning once (at the start 276 // of the marking cycle) and the root regions are all the survivor 277 // regions populated during the initial-mark pause. 278 class G1CMRootRegions VALUE_OBJ_CLASS_SPEC { 279 private: 280 const G1SurvivorRegions* _survivors; 281 G1ConcurrentMark* _cm; 282 283 volatile bool _scan_in_progress; 284 volatile bool _should_abort; 285 volatile int _claimed_survivor_index; 286 287 void notify_scan_done(); 288 289 public: 290 G1CMRootRegions(); 291 // We actually do most of the initialization in this method. 292 void init(const G1SurvivorRegions* survivors, G1ConcurrentMark* cm); 293 294 // Reset the claiming / scanning of the root regions. 295 void prepare_for_scan(); 296 297 // Forces get_next() to return NULL so that the iteration aborts early. 298 void abort() { _should_abort = true; } 299 300 // Return true if the CM thread are actively scanning root regions, 301 // false otherwise. 302 bool scan_in_progress() { return _scan_in_progress; } 303 304 // Claim the next root region to scan atomically, or return NULL if 305 // all have been claimed. 306 HeapRegion* claim_next(); 307 308 // The number of root regions to scan. 309 uint num_root_regions() const; 310 311 void cancel_scan(); 312 313 // Flag that we're done with root region scanning and notify anyone 314 // who's waiting on it. If aborted is false, assume that all regions 315 // have been claimed. 316 void scan_finished(); 317 318 // If CM threads are still scanning root regions, wait until they 319 // are done. Return true if we had to wait, false otherwise. 320 bool wait_until_scan_finished(); 321 }; 322 323 class ConcurrentMarkThread; 324 325 class G1ConcurrentMark: public CHeapObj<mtGC> { 326 friend class ConcurrentMarkThread; 327 friend class G1ParNoteEndTask; 328 friend class G1VerifyLiveDataClosure; 329 friend class G1CMRefProcTaskProxy; 330 friend class G1CMRefProcTaskExecutor; 331 friend class G1CMKeepAliveAndDrainClosure; 332 friend class G1CMDrainMarkingStackClosure; 333 friend class G1CMBitMapClosure; 334 friend class G1CMConcurrentMarkingTask; 335 friend class G1CMRemarkTask; 336 friend class G1CMTask; 337 338 protected: 339 ConcurrentMarkThread* _cmThread; // The thread doing the work 340 G1CollectedHeap* _g1h; // The heap 341 uint _parallel_marking_threads; // The number of marking 342 // threads we're using 343 uint _max_parallel_marking_threads; // Max number of marking 344 // threads we'll ever use 345 double _sleep_factor; // How much we have to sleep, with 346 // respect to the work we just did, to 347 // meet the marking overhead goal 348 double _marking_task_overhead; // Marking target overhead for 349 // a single task 350 351 FreeRegionList _cleanup_list; 352 353 // Concurrent marking support structures 354 G1CMBitMap _markBitMap1; 355 G1CMBitMap _markBitMap2; 356 G1CMBitMapRO* _prevMarkBitMap; // Completed mark bitmap 357 G1CMBitMap* _nextMarkBitMap; // Under-construction mark bitmap 358 359 // Heap bounds 360 HeapWord* _heap_start; 361 HeapWord* _heap_end; 362 363 // Root region tracking and claiming 364 G1CMRootRegions _root_regions; 365 366 // For gray objects 367 G1CMMarkStack _global_mark_stack; // Grey objects behind global finger 368 HeapWord* volatile _finger; // The global finger, region aligned, 369 // always points to the end of the 370 // last claimed region 371 372 // Marking tasks 373 uint _max_worker_id;// Maximum worker id 374 uint _active_tasks; // Task num currently active 375 G1CMTask** _tasks; // Task queue array (max_worker_id len) 376 G1CMTaskQueueSet* _task_queues; // Task queue set 377 ParallelTaskTerminator _terminator; // For termination 378 379 // Two sync barriers that are used to synchronize tasks when an 380 // overflow occurs. The algorithm is the following. All tasks enter 381 // the first one to ensure that they have all stopped manipulating 382 // the global data structures. After they exit it, they re-initialize 383 // their data structures and task 0 re-initializes the global data 384 // structures. Then, they enter the second sync barrier. This 385 // ensure, that no task starts doing work before all data 386 // structures (local and global) have been re-initialized. When they 387 // exit it, they are free to start working again. 388 WorkGangBarrierSync _first_overflow_barrier_sync; 389 WorkGangBarrierSync _second_overflow_barrier_sync; 390 391 // True: marking is concurrent, false: we're in remark 392 volatile bool _concurrent; 393 // Set at the end of a Full GC so that marking aborts 394 volatile bool _has_aborted; 395 396 // Used when remark aborts due to an overflow to indicate that 397 // another concurrent marking phase should start 398 volatile bool _restart_for_overflow; 399 400 // This is true from the very start of concurrent marking until the 401 // point when all the tasks complete their work. It is really used 402 // to determine the points between the end of concurrent marking and 403 // time of remark. 404 volatile bool _concurrent_marking_in_progress; 405 406 ConcurrentGCTimer* _gc_timer_cm; 407 408 G1OldTracer* _gc_tracer_cm; 409 410 // All of these times are in ms 411 NumberSeq _init_times; 412 NumberSeq _remark_times; 413 NumberSeq _remark_mark_times; 414 NumberSeq _remark_weak_ref_times; 415 NumberSeq _cleanup_times; 416 double _total_counting_time; 417 double _total_rs_scrub_time; 418 419 double* _accum_task_vtime; // Accumulated task vtime 420 421 WorkGang* _parallel_workers; 422 423 void weakRefsWorkParallelPart(BoolObjectClosure* is_alive, bool purged_classes); 424 void weakRefsWork(bool clear_all_soft_refs); 425 426 void swapMarkBitMaps(); 427 428 // It resets the global marking data structures, as well as the 429 // task local ones; should be called during initial mark. 430 void reset(); 431 432 // Resets all the marking data structures. Called when we have to restart 433 // marking or when marking completes (via set_non_marking_state below). 434 void reset_marking_state(); 435 436 // We do this after we're done with marking so that the marking data 437 // structures are initialized to a sensible and predictable state. 438 void set_non_marking_state(); 439 440 // Called to indicate how many threads are currently active. 441 void set_concurrency(uint active_tasks); 442 443 // It should be called to indicate which phase we're in (concurrent 444 // mark or remark) and how many threads are currently active. 445 void set_concurrency_and_phase(uint active_tasks, bool concurrent); 446 447 // Prints all gathered CM-related statistics 448 void print_stats(); 449 450 bool cleanup_list_is_empty() { 451 return _cleanup_list.is_empty(); 452 } 453 454 // Accessor methods 455 uint parallel_marking_threads() const { return _parallel_marking_threads; } 456 uint max_parallel_marking_threads() const { return _max_parallel_marking_threads;} 457 double sleep_factor() { return _sleep_factor; } 458 double marking_task_overhead() { return _marking_task_overhead;} 459 460 HeapWord* finger() { return _finger; } 461 bool concurrent() { return _concurrent; } 462 uint active_tasks() { return _active_tasks; } 463 ParallelTaskTerminator* terminator() { return &_terminator; } 464 465 // It claims the next available region to be scanned by a marking 466 // task/thread. It might return NULL if the next region is empty or 467 // we have run out of regions. In the latter case, out_of_regions() 468 // determines whether we've really run out of regions or the task 469 // should call claim_region() again. This might seem a bit 470 // awkward. Originally, the code was written so that claim_region() 471 // either successfully returned with a non-empty region or there 472 // were no more regions to be claimed. The problem with this was 473 // that, in certain circumstances, it iterated over large chunks of 474 // the heap finding only empty regions and, while it was working, it 475 // was preventing the calling task to call its regular clock 476 // method. So, this way, each task will spend very little time in 477 // claim_region() and is allowed to call the regular clock method 478 // frequently. 479 HeapRegion* claim_region(uint worker_id); 480 481 // It determines whether we've run out of regions to scan. Note that 482 // the finger can point past the heap end in case the heap was expanded 483 // to satisfy an allocation without doing a GC. This is fine, because all 484 // objects in those regions will be considered live anyway because of 485 // SATB guarantees (i.e. their TAMS will be equal to bottom). 486 bool out_of_regions() { return _finger >= _heap_end; } 487 488 // Returns the task with the given id 489 G1CMTask* task(int id) { 490 assert(0 <= id && id < (int) _active_tasks, 491 "task id not within active bounds"); 492 return _tasks[id]; 493 } 494 495 // Returns the task queue with the given id 496 G1CMTaskQueue* task_queue(int id) { 497 assert(0 <= id && id < (int) _active_tasks, 498 "task queue id not within active bounds"); 499 return (G1CMTaskQueue*) _task_queues->queue(id); 500 } 501 502 // Returns the task queue set 503 G1CMTaskQueueSet* task_queues() { return _task_queues; } 504 505 // Access / manipulation of the overflow flag which is set to 506 // indicate that the global stack has overflown 507 bool has_overflown() { return _global_mark_stack.is_out_of_memory(); } 508 void clear_has_overflown() { _global_mark_stack.clear_out_of_memory(); } 509 bool restart_for_overflow() { return _restart_for_overflow; } 510 511 // Methods to enter the two overflow sync barriers 512 void enter_first_sync_barrier(uint worker_id); 513 void enter_second_sync_barrier(uint worker_id); 514 515 // Card index of the bottom of the G1 heap. Used for biasing indices into 516 // the card bitmaps. 517 intptr_t _heap_bottom_card_num; 518 519 // Set to true when initialization is complete 520 bool _completed_initialization; 521 522 // end_timer, true to end gc timer after ending concurrent phase. 523 void register_concurrent_phase_end_common(bool end_timer); 524 525 // Clear the given bitmap in parallel using the given WorkGang. If may_yield is 526 // true, periodically insert checks to see if this method should exit prematurely. 527 void clear_bitmap(G1CMBitMap* bitmap, WorkGang* workers, bool may_yield); 528 public: 529 // Manipulation of the global mark stack. 530 // The push and pop operations are used by tasks for transfers 531 // between task-local queues and the global mark stack. 532 bool mark_stack_push(oop* arr) { 533 return _global_mark_stack.par_push_chunk(arr); 534 } 535 bool mark_stack_pop(oop* arr) { 536 return _global_mark_stack.par_pop_chunk(arr); 537 } 538 size_t mark_stack_size() { return _global_mark_stack.size(); } 539 size_t partial_mark_stack_size_target() { return _global_mark_stack.capacity()/3; } 540 bool mark_stack_empty() { return _global_mark_stack.is_empty(); } 541 542 G1CMRootRegions* root_regions() { return &_root_regions; } 543 544 bool concurrent_marking_in_progress() { 545 return _concurrent_marking_in_progress; 546 } 547 void set_concurrent_marking_in_progress() { 548 _concurrent_marking_in_progress = true; 549 } 550 void clear_concurrent_marking_in_progress() { 551 _concurrent_marking_in_progress = false; 552 } 553 554 void concurrent_cycle_start(); 555 void concurrent_cycle_end(); 556 557 void update_accum_task_vtime(int i, double vtime) { 558 _accum_task_vtime[i] += vtime; 559 } 560 561 double all_task_accum_vtime() { 562 double ret = 0.0; 563 for (uint i = 0; i < _max_worker_id; ++i) 564 ret += _accum_task_vtime[i]; 565 return ret; 566 } 567 568 // Attempts to steal an object from the task queues of other tasks 569 bool try_stealing(uint worker_id, int* hash_seed, oop& obj); 570 571 G1ConcurrentMark(G1CollectedHeap* g1h, 572 G1RegionToSpaceMapper* prev_bitmap_storage, 573 G1RegionToSpaceMapper* next_bitmap_storage); 574 ~G1ConcurrentMark(); 575 576 ConcurrentMarkThread* cmThread() { return _cmThread; } 577 578 G1CMBitMapRO* prevMarkBitMap() const { return _prevMarkBitMap; } 579 G1CMBitMap* nextMarkBitMap() const { return _nextMarkBitMap; } 580 581 // Returns the number of GC threads to be used in a concurrent 582 // phase based on the number of GC threads being used in a STW 583 // phase. 584 uint scale_parallel_threads(uint n_par_threads); 585 586 // Calculates the number of GC threads to be used in a concurrent phase. 587 uint calc_parallel_marking_threads(); 588 589 // The following three are interaction between CM and 590 // G1CollectedHeap 591 592 // This notifies CM that a root during initial-mark needs to be 593 // grayed. It is MT-safe. hr is the region that 594 // contains the object and it's passed optionally from callers who 595 // might already have it (no point in recalculating it). 596 inline void grayRoot(oop obj, 597 HeapRegion* hr = NULL); 598 599 // Prepare internal data structures for the next mark cycle. This includes clearing 600 // the next mark bitmap and some internal data structures. This method is intended 601 // to be called concurrently to the mutator. It will yield to safepoint requests. 602 void cleanup_for_next_mark(); 603 604 // Clear the previous marking bitmap during safepoint. 605 void clear_prev_bitmap(WorkGang* workers); 606 607 // Return whether the next mark bitmap has no marks set. To be used for assertions 608 // only. Will not yield to pause requests. 609 bool nextMarkBitmapIsClear(); 610 611 // These two do the work that needs to be done before and after the 612 // initial root checkpoint. Since this checkpoint can be done at two 613 // different points (i.e. an explicit pause or piggy-backed on a 614 // young collection), then it's nice to be able to easily share the 615 // pre/post code. It might be the case that we can put everything in 616 // the post method. TP 617 void checkpointRootsInitialPre(); 618 void checkpointRootsInitialPost(); 619 620 // Scan all the root regions and mark everything reachable from 621 // them. 622 void scan_root_regions(); 623 624 // Scan a single root region and mark everything reachable from it. 625 void scanRootRegion(HeapRegion* hr); 626 627 // Do concurrent phase of marking, to a tentative transitive closure. 628 void mark_from_roots(); 629 630 void checkpointRootsFinal(bool clear_all_soft_refs); 631 void checkpointRootsFinalWork(); 632 void cleanup(); 633 void complete_cleanup(); 634 635 // Mark in the previous bitmap. NB: this is usually read-only, so use 636 // this carefully! 637 inline void markPrev(oop p); 638 639 // Clears marks for all objects in the given range, for the prev or 640 // next bitmaps. NB: the previous bitmap is usually 641 // read-only, so use this carefully! 642 void clearRangePrevBitmap(MemRegion mr); 643 644 // Verify that there are no CSet oops on the stacks (taskqueues / 645 // global mark stack) and fingers (global / per-task). 646 // If marking is not in progress, it's a no-op. 647 void verify_no_cset_oops() PRODUCT_RETURN; 648 649 inline bool isPrevMarked(oop p) const; 650 651 inline bool do_yield_check(); 652 653 // Abandon current marking iteration due to a Full GC. 654 void abort(); 655 656 bool has_aborted() { return _has_aborted; } 657 658 void print_summary_info(); 659 660 void print_worker_threads_on(outputStream* st) const; 661 void threads_do(ThreadClosure* tc) const; 662 663 void print_on_error(outputStream* st) const; 664 665 // Attempts to mark the given object on the next mark bitmap. 666 inline bool par_mark(oop obj); 667 668 // Returns true if initialization was successfully completed. 669 bool completed_initialization() const { 670 return _completed_initialization; 671 } 672 673 ConcurrentGCTimer* gc_timer_cm() const { return _gc_timer_cm; } 674 G1OldTracer* gc_tracer_cm() const { return _gc_tracer_cm; } 675 676 private: 677 // Clear (Reset) all liveness count data. 678 void clear_live_data(WorkGang* workers); 679 680 #ifdef ASSERT 681 // Verify all of the above data structures that they are in initial state. 682 void verify_live_data_clear(); 683 #endif 684 685 // Aggregates the per-card liveness data based on the current marking. Also sets 686 // the amount of marked bytes for each region. 687 void create_live_data(); 688 689 void finalize_live_data(); 690 691 void verify_live_data(); 692 }; 693 694 // A class representing a marking task. 695 class G1CMTask : public TerminatorTerminator { 696 private: 697 enum PrivateConstants { 698 // The regular clock call is called once the scanned words reaches 699 // this limit 700 words_scanned_period = 12*1024, 701 // The regular clock call is called once the number of visited 702 // references reaches this limit 703 refs_reached_period = 1024, 704 // Initial value for the hash seed, used in the work stealing code 705 init_hash_seed = 17 706 }; 707 708 G1CMObjArrayProcessor _objArray_processor; 709 710 uint _worker_id; 711 G1CollectedHeap* _g1h; 712 G1ConcurrentMark* _cm; 713 G1CMBitMap* _nextMarkBitMap; 714 // the task queue of this task 715 G1CMTaskQueue* _task_queue; 716 private: 717 // the task queue set---needed for stealing 718 G1CMTaskQueueSet* _task_queues; 719 // indicates whether the task has been claimed---this is only for 720 // debugging purposes 721 bool _claimed; 722 723 // number of calls to this task 724 int _calls; 725 726 // when the virtual timer reaches this time, the marking step should 727 // exit 728 double _time_target_ms; 729 // the start time of the current marking step 730 double _start_time_ms; 731 732 // the oop closure used for iterations over oops 733 G1CMOopClosure* _cm_oop_closure; 734 735 // the region this task is scanning, NULL if we're not scanning any 736 HeapRegion* _curr_region; 737 // the local finger of this task, NULL if we're not scanning a region 738 HeapWord* _finger; 739 // limit of the region this task is scanning, NULL if we're not scanning one 740 HeapWord* _region_limit; 741 742 // the number of words this task has scanned 743 size_t _words_scanned; 744 // When _words_scanned reaches this limit, the regular clock is 745 // called. Notice that this might be decreased under certain 746 // circumstances (i.e. when we believe that we did an expensive 747 // operation). 748 size_t _words_scanned_limit; 749 // the initial value of _words_scanned_limit (i.e. what it was 750 // before it was decreased). 751 size_t _real_words_scanned_limit; 752 753 // the number of references this task has visited 754 size_t _refs_reached; 755 // When _refs_reached reaches this limit, the regular clock is 756 // called. Notice this this might be decreased under certain 757 // circumstances (i.e. when we believe that we did an expensive 758 // operation). 759 size_t _refs_reached_limit; 760 // the initial value of _refs_reached_limit (i.e. what it was before 761 // it was decreased). 762 size_t _real_refs_reached_limit; 763 764 // used by the work stealing stuff 765 int _hash_seed; 766 // if this is true, then the task has aborted for some reason 767 bool _has_aborted; 768 // set when the task aborts because it has met its time quota 769 bool _has_timed_out; 770 // true when we're draining SATB buffers; this avoids the task 771 // aborting due to SATB buffers being available (as we're already 772 // dealing with them) 773 bool _draining_satb_buffers; 774 775 // number sequence of past step times 776 NumberSeq _step_times_ms; 777 // elapsed time of this task 778 double _elapsed_time_ms; 779 // termination time of this task 780 double _termination_time_ms; 781 // when this task got into the termination protocol 782 double _termination_start_time_ms; 783 784 // true when the task is during a concurrent phase, false when it is 785 // in the remark phase (so, in the latter case, we do not have to 786 // check all the things that we have to check during the concurrent 787 // phase, i.e. SATB buffer availability...) 788 bool _concurrent; 789 790 TruncatedSeq _marking_step_diffs_ms; 791 792 // it updates the local fields after this task has claimed 793 // a new region to scan 794 void setup_for_region(HeapRegion* hr); 795 // it brings up-to-date the limit of the region 796 void update_region_limit(); 797 798 // called when either the words scanned or the refs visited limit 799 // has been reached 800 void reached_limit(); 801 // recalculates the words scanned and refs visited limits 802 void recalculate_limits(); 803 // decreases the words scanned and refs visited limits when we reach 804 // an expensive operation 805 void decrease_limits(); 806 // it checks whether the words scanned or refs visited reached their 807 // respective limit and calls reached_limit() if they have 808 void check_limits() { 809 if (_words_scanned >= _words_scanned_limit || 810 _refs_reached >= _refs_reached_limit) { 811 reached_limit(); 812 } 813 } 814 // this is supposed to be called regularly during a marking step as 815 // it checks a bunch of conditions that might cause the marking step 816 // to abort 817 void regular_clock_call(); 818 bool concurrent() { return _concurrent; } 819 820 // Test whether obj might have already been passed over by the 821 // mark bitmap scan, and so needs to be pushed onto the mark stack. 822 bool is_below_finger(oop obj, HeapWord* global_finger) const; 823 824 template<bool scan> void process_grey_object(oop obj); 825 public: 826 // Apply the closure on the given area of the objArray. Return the number of words 827 // scanned. 828 inline size_t scan_objArray(objArrayOop obj, MemRegion mr); 829 // It resets the task; it should be called right at the beginning of 830 // a marking phase. 831 void reset(G1CMBitMap* _nextMarkBitMap); 832 // it clears all the fields that correspond to a claimed region. 833 void clear_region_fields(); 834 835 void set_concurrent(bool concurrent) { _concurrent = concurrent; } 836 837 // The main method of this class which performs a marking step 838 // trying not to exceed the given duration. However, it might exit 839 // prematurely, according to some conditions (i.e. SATB buffers are 840 // available for processing). 841 void do_marking_step(double target_ms, 842 bool do_termination, 843 bool is_serial); 844 845 // These two calls start and stop the timer 846 void record_start_time() { 847 _elapsed_time_ms = os::elapsedTime() * 1000.0; 848 } 849 void record_end_time() { 850 _elapsed_time_ms = os::elapsedTime() * 1000.0 - _elapsed_time_ms; 851 } 852 853 // returns the worker ID associated with this task. 854 uint worker_id() { return _worker_id; } 855 856 // From TerminatorTerminator. It determines whether this task should 857 // exit the termination protocol after it's entered it. 858 virtual bool should_exit_termination(); 859 860 // Resets the local region fields after a task has finished scanning a 861 // region; or when they have become stale as a result of the region 862 // being evacuated. 863 void giveup_current_region(); 864 865 HeapWord* finger() { return _finger; } 866 867 bool has_aborted() { return _has_aborted; } 868 void set_has_aborted() { _has_aborted = true; } 869 void clear_has_aborted() { _has_aborted = false; } 870 bool has_timed_out() { return _has_timed_out; } 871 bool claimed() { return _claimed; } 872 873 void set_cm_oop_closure(G1CMOopClosure* cm_oop_closure); 874 875 // Increment the number of references this task has visited. 876 void increment_refs_reached() { ++_refs_reached; } 877 878 // Grey the object by marking it. If not already marked, push it on 879 // the local queue if below the finger. 880 // obj is below its region's NTAMS. 881 inline void make_reference_grey(oop obj); 882 883 // Grey the object (by calling make_grey_reference) if required, 884 // e.g. obj is below its containing region's NTAMS. 885 // Precondition: obj is a valid heap object. 886 inline void deal_with_reference(oop obj); 887 888 // It scans an object and visits its children. 889 inline void scan_object(oop obj); 890 891 // It pushes an object on the local queue. 892 inline void push(oop obj); 893 894 // Move entries to the global stack. 895 void move_entries_to_global_stack(); 896 // Move entries from the global stack, return true if we were successful to do so. 897 bool get_entries_from_global_stack(); 898 899 // It pops and scans objects from the local queue. If partially is 900 // true, then it stops when the queue size is of a given limit. If 901 // partially is false, then it stops when the queue is empty. 902 void drain_local_queue(bool partially); 903 // It moves entries from the global stack to the local queue and 904 // drains the local queue. If partially is true, then it stops when 905 // both the global stack and the local queue reach a given size. If 906 // partially if false, it tries to empty them totally. 907 void drain_global_stack(bool partially); 908 // It keeps picking SATB buffers and processing them until no SATB 909 // buffers are available. 910 void drain_satb_buffers(); 911 912 // moves the local finger to a new location 913 inline void move_finger_to(HeapWord* new_finger) { 914 assert(new_finger >= _finger && new_finger < _region_limit, "invariant"); 915 _finger = new_finger; 916 } 917 918 G1CMTask(uint worker_id, 919 G1ConcurrentMark *cm, 920 G1CMTaskQueue* task_queue, 921 G1CMTaskQueueSet* task_queues); 922 923 // it prints statistics associated with this task 924 void print_stats(); 925 }; 926 927 // Class that's used to to print out per-region liveness 928 // information. It's currently used at the end of marking and also 929 // after we sort the old regions at the end of the cleanup operation. 930 class G1PrintRegionLivenessInfoClosure: public HeapRegionClosure { 931 private: 932 // Accumulators for these values. 933 size_t _total_used_bytes; 934 size_t _total_capacity_bytes; 935 size_t _total_prev_live_bytes; 936 size_t _total_next_live_bytes; 937 938 // Accumulator for the remembered set size 939 size_t _total_remset_bytes; 940 941 // Accumulator for strong code roots memory size 942 size_t _total_strong_code_roots_bytes; 943 944 static double perc(size_t val, size_t total) { 945 if (total == 0) { 946 return 0.0; 947 } else { 948 return 100.0 * ((double) val / (double) total); 949 } 950 } 951 952 static double bytes_to_mb(size_t val) { 953 return (double) val / (double) M; 954 } 955 956 public: 957 // The header and footer are printed in the constructor and 958 // destructor respectively. 959 G1PrintRegionLivenessInfoClosure(const char* phase_name); 960 virtual bool doHeapRegion(HeapRegion* r); 961 ~G1PrintRegionLivenessInfoClosure(); 962 }; 963 964 #endif // SHARE_VM_GC_G1_G1CONCURRENTMARK_HPP