1 /* 2 * Copyright (c) 2001, 2015, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #ifndef SHARE_VM_GC_G1_CONCURRENTMARK_HPP 26 #define SHARE_VM_GC_G1_CONCURRENTMARK_HPP 27 28 #include "classfile/javaClasses.hpp" 29 #include "gc/g1/g1RegionToSpaceMapper.hpp" 30 #include "gc/g1/heapRegionSet.hpp" 31 #include "gc/shared/gcId.hpp" 32 #include "gc/shared/taskqueue.hpp" 33 34 class G1CollectedHeap; 35 class CMBitMap; 36 class CMTask; 37 class ConcurrentMark; 38 typedef GenericTaskQueue<oop, mtGC> CMTaskQueue; 39 typedef GenericTaskQueueSet<CMTaskQueue, mtGC> CMTaskQueueSet; 40 41 // Closure used by CM during concurrent reference discovery 42 // and reference processing (during remarking) to determine 43 // if a particular object is alive. It is primarily used 44 // to determine if referents of discovered reference objects 45 // are alive. An instance is also embedded into the 46 // reference processor as the _is_alive_non_header field 47 class G1CMIsAliveClosure: public BoolObjectClosure { 48 G1CollectedHeap* _g1; 49 public: 50 G1CMIsAliveClosure(G1CollectedHeap* g1) : _g1(g1) { } 51 52 bool do_object_b(oop obj); 53 }; 54 55 // A generic CM bit map. This is essentially a wrapper around the BitMap 56 // class, with one bit per (1<<_shifter) HeapWords. 57 58 class CMBitMapRO VALUE_OBJ_CLASS_SPEC { 59 protected: 60 HeapWord* _bmStartWord; // base address of range covered by map 61 size_t _bmWordSize; // map size (in #HeapWords covered) 62 const int _shifter; // map to char or bit 63 BitMap _bm; // the bit map itself 64 65 public: 66 // constructor 67 CMBitMapRO(int shifter); 68 69 enum { do_yield = true }; 70 71 // inquiries 72 HeapWord* startWord() const { return _bmStartWord; } 73 size_t sizeInWords() const { return _bmWordSize; } 74 // the following is one past the last word in space 75 HeapWord* endWord() const { return _bmStartWord + _bmWordSize; } 76 77 // read marks 78 79 bool isMarked(HeapWord* addr) const { 80 assert(_bmStartWord <= addr && addr < (_bmStartWord + _bmWordSize), 81 "outside underlying space?"); 82 return _bm.at(heapWordToOffset(addr)); 83 } 84 85 // iteration 86 inline bool iterate(BitMapClosure* cl, MemRegion mr); 87 inline bool iterate(BitMapClosure* cl); 88 89 // Return the address corresponding to the next marked bit at or after 90 // "addr", and before "limit", if "limit" is non-NULL. If there is no 91 // such bit, returns "limit" if that is non-NULL, or else "endWord()". 92 HeapWord* getNextMarkedWordAddress(const HeapWord* addr, 93 const HeapWord* limit = NULL) const; 94 // Return the address corresponding to the next unmarked bit at or after 95 // "addr", and before "limit", if "limit" is non-NULL. If there is no 96 // such bit, returns "limit" if that is non-NULL, or else "endWord()". 97 HeapWord* getNextUnmarkedWordAddress(const HeapWord* addr, 98 const HeapWord* limit = NULL) const; 99 100 // conversion utilities 101 HeapWord* offsetToHeapWord(size_t offset) const { 102 return _bmStartWord + (offset << _shifter); 103 } 104 size_t heapWordToOffset(const HeapWord* addr) const { 105 return pointer_delta(addr, _bmStartWord) >> _shifter; 106 } 107 int heapWordDiffToOffsetDiff(size_t diff) const; 108 109 // The argument addr should be the start address of a valid object 110 HeapWord* nextObject(HeapWord* addr) { 111 oop obj = (oop) addr; 112 HeapWord* res = addr + obj->size(); 113 assert(offsetToHeapWord(heapWordToOffset(res)) == res, "sanity"); 114 return res; 115 } 116 117 void print_on_error(outputStream* st, const char* prefix) const; 118 119 // debugging 120 NOT_PRODUCT(bool covers(MemRegion rs) const;) 121 }; 122 123 class CMBitMapMappingChangedListener : public G1MappingChangedListener { 124 private: 125 CMBitMap* _bm; 126 public: 127 CMBitMapMappingChangedListener() : _bm(NULL) {} 128 129 void set_bitmap(CMBitMap* bm) { _bm = bm; } 130 131 virtual void on_commit(uint start_idx, size_t num_regions, bool zero_filled); 132 }; 133 134 class CMBitMap : public CMBitMapRO { 135 private: 136 CMBitMapMappingChangedListener _listener; 137 138 public: 139 static size_t compute_size(size_t heap_size); 140 // Returns the amount of bytes on the heap between two marks in the bitmap. 141 static size_t mark_distance(); 142 // Returns how many bytes (or bits) of the heap a single byte (or bit) of the 143 // mark bitmap corresponds to. This is the same as the mark distance above. 144 static size_t heap_map_factor() { 145 return mark_distance(); 146 } 147 148 CMBitMap() : CMBitMapRO(LogMinObjAlignment), _listener() { _listener.set_bitmap(this); } 149 150 // Initializes the underlying BitMap to cover the given area. 151 void initialize(MemRegion heap, G1RegionToSpaceMapper* storage); 152 153 // Write marks. 154 inline void mark(HeapWord* addr); 155 inline void clear(HeapWord* addr); 156 inline bool parMark(HeapWord* addr); 157 inline bool parClear(HeapWord* addr); 158 159 void markRange(MemRegion mr); 160 void clearRange(MemRegion mr); 161 162 // Starting at the bit corresponding to "addr" (inclusive), find the next 163 // "1" bit, if any. This bit starts some run of consecutive "1"'s; find 164 // the end of this run (stopping at "end_addr"). Return the MemRegion 165 // covering from the start of the region corresponding to the first bit 166 // of the run to the end of the region corresponding to the last bit of 167 // the run. If there is no "1" bit at or after "addr", return an empty 168 // MemRegion. 169 MemRegion getAndClearMarkedRegion(HeapWord* addr, HeapWord* end_addr); 170 171 // Clear the whole mark bitmap. 172 void clearAll(); 173 }; 174 175 // Represents a marking stack used by ConcurrentMarking in the G1 collector. 176 class CMMarkStack VALUE_OBJ_CLASS_SPEC { 177 VirtualSpace _virtual_space; // Underlying backing store for actual stack 178 ConcurrentMark* _cm; 179 oop* _base; // bottom of stack 180 jint _index; // one more than last occupied index 181 jint _capacity; // max #elements 182 jint _saved_index; // value of _index saved at start of GC 183 184 bool _overflow; 185 bool _should_expand; 186 DEBUG_ONLY(bool _drain_in_progress;) 187 DEBUG_ONLY(bool _drain_in_progress_yields;) 188 189 oop pop() { 190 if (!isEmpty()) { 191 return _base[--_index] ; 192 } 193 return NULL; 194 } 195 196 public: 197 CMMarkStack(ConcurrentMark* cm); 198 ~CMMarkStack(); 199 200 bool allocate(size_t capacity); 201 202 // Pushes the first "n" elements of "ptr_arr" on the stack. 203 // Locking impl: concurrency is allowed only with 204 // "par_push_arr" and/or "par_pop_arr" operations, which use the same 205 // locking strategy. 206 void par_push_arr(oop* ptr_arr, int n); 207 208 // If returns false, the array was empty. Otherwise, removes up to "max" 209 // elements from the stack, and transfers them to "ptr_arr" in an 210 // unspecified order. The actual number transferred is given in "n" ("n 211 // == 0" is deliberately redundant with the return value.) Locking impl: 212 // concurrency is allowed only with "par_push_arr" and/or "par_pop_arr" 213 // operations, which use the same locking strategy. 214 bool par_pop_arr(oop* ptr_arr, int max, int* n); 215 216 // Drain the mark stack, applying the given closure to all fields of 217 // objects on the stack. (That is, continue until the stack is empty, 218 // even if closure applications add entries to the stack.) The "bm" 219 // argument, if non-null, may be used to verify that only marked objects 220 // are on the mark stack. If "yield_after" is "true", then the 221 // concurrent marker performing the drain offers to yield after 222 // processing each object. If a yield occurs, stops the drain operation 223 // and returns false. Otherwise, returns true. 224 template<class OopClosureClass> 225 bool drain(OopClosureClass* cl, CMBitMap* bm, bool yield_after = false); 226 227 bool isEmpty() { return _index == 0; } 228 int maxElems() { return _capacity; } 229 230 bool overflow() { return _overflow; } 231 void clear_overflow() { _overflow = false; } 232 233 bool should_expand() const { return _should_expand; } 234 void set_should_expand(); 235 236 // Expand the stack, typically in response to an overflow condition 237 void expand(); 238 239 int size() { return _index; } 240 241 void setEmpty() { _index = 0; clear_overflow(); } 242 243 // Record the current index. 244 void note_start_of_gc(); 245 246 // Make sure that we have not added any entries to the stack during GC. 247 void note_end_of_gc(); 248 249 // iterate over the oops in the mark stack, up to the bound recorded via 250 // the call above. 251 void oops_do(OopClosure* f); 252 }; 253 254 class ForceOverflowSettings VALUE_OBJ_CLASS_SPEC { 255 private: 256 #ifndef PRODUCT 257 uintx _num_remaining; 258 bool _force; 259 #endif // !defined(PRODUCT) 260 261 public: 262 void init() PRODUCT_RETURN; 263 void update() PRODUCT_RETURN; 264 bool should_force() PRODUCT_RETURN_( return false; ); 265 }; 266 267 // this will enable a variety of different statistics per GC task 268 #define _MARKING_STATS_ 0 269 // this will enable the higher verbose levels 270 #define _MARKING_VERBOSE_ 0 271 272 #if _MARKING_STATS_ 273 #define statsOnly(statement) \ 274 do { \ 275 statement ; \ 276 } while (0) 277 #else // _MARKING_STATS_ 278 #define statsOnly(statement) \ 279 do { \ 280 } while (0) 281 #endif // _MARKING_STATS_ 282 283 typedef enum { 284 no_verbose = 0, // verbose turned off 285 stats_verbose, // only prints stats at the end of marking 286 low_verbose, // low verbose, mostly per region and per major event 287 medium_verbose, // a bit more detailed than low 288 high_verbose // per object verbose 289 } CMVerboseLevel; 290 291 class YoungList; 292 293 // Root Regions are regions that are not empty at the beginning of a 294 // marking cycle and which we might collect during an evacuation pause 295 // while the cycle is active. Given that, during evacuation pauses, we 296 // do not copy objects that are explicitly marked, what we have to do 297 // for the root regions is to scan them and mark all objects reachable 298 // from them. According to the SATB assumptions, we only need to visit 299 // each object once during marking. So, as long as we finish this scan 300 // before the next evacuation pause, we can copy the objects from the 301 // root regions without having to mark them or do anything else to them. 302 // 303 // Currently, we only support root region scanning once (at the start 304 // of the marking cycle) and the root regions are all the survivor 305 // regions populated during the initial-mark pause. 306 class CMRootRegions VALUE_OBJ_CLASS_SPEC { 307 private: 308 YoungList* _young_list; 309 ConcurrentMark* _cm; 310 311 volatile bool _scan_in_progress; 312 volatile bool _should_abort; 313 HeapRegion* volatile _next_survivor; 314 315 public: 316 CMRootRegions(); 317 // We actually do most of the initialization in this method. 318 void init(G1CollectedHeap* g1h, ConcurrentMark* cm); 319 320 // Reset the claiming / scanning of the root regions. 321 void prepare_for_scan(); 322 323 // Forces get_next() to return NULL so that the iteration aborts early. 324 void abort() { _should_abort = true; } 325 326 // Return true if the CM thread are actively scanning root regions, 327 // false otherwise. 328 bool scan_in_progress() { return _scan_in_progress; } 329 330 // Claim the next root region to scan atomically, or return NULL if 331 // all have been claimed. 332 HeapRegion* claim_next(); 333 334 // Flag that we're done with root region scanning and notify anyone 335 // who's waiting on it. If aborted is false, assume that all regions 336 // have been claimed. 337 void scan_finished(); 338 339 // If CM threads are still scanning root regions, wait until they 340 // are done. Return true if we had to wait, false otherwise. 341 bool wait_until_scan_finished(); 342 }; 343 344 class ConcurrentMarkThread; 345 346 class ConcurrentMark: public CHeapObj<mtGC> { 347 friend class CMMarkStack; 348 friend class ConcurrentMarkThread; 349 friend class CMTask; 350 friend class CMBitMapClosure; 351 friend class CMRemarkTask; 352 friend class CMConcurrentMarkingTask; 353 friend class G1ParNoteEndTask; 354 friend class CalcLiveObjectsClosure; 355 friend class G1CMRefProcTaskProxy; 356 friend class G1CMRefProcTaskExecutor; 357 friend class G1CMKeepAliveAndDrainClosure; 358 friend class G1CMDrainMarkingStackClosure; 359 360 protected: 361 ConcurrentMarkThread* _cmThread; // The thread doing the work 362 G1CollectedHeap* _g1h; // The heap 363 uint _parallel_marking_threads; // The number of marking 364 // threads we're using 365 uint _max_parallel_marking_threads; // Max number of marking 366 // threads we'll ever use 367 double _sleep_factor; // How much we have to sleep, with 368 // respect to the work we just did, to 369 // meet the marking overhead goal 370 double _marking_task_overhead; // Marking target overhead for 371 // a single task 372 373 // Same as the two above, but for the cleanup task 374 double _cleanup_sleep_factor; 375 double _cleanup_task_overhead; 376 377 FreeRegionList _cleanup_list; 378 379 // Concurrent marking support structures 380 CMBitMap _markBitMap1; 381 CMBitMap _markBitMap2; 382 CMBitMapRO* _prevMarkBitMap; // Completed mark bitmap 383 CMBitMap* _nextMarkBitMap; // Under-construction mark bitmap 384 385 BitMap _region_bm; 386 BitMap _card_bm; 387 388 // Heap bounds 389 HeapWord* _heap_start; 390 HeapWord* _heap_end; 391 392 // Root region tracking and claiming 393 CMRootRegions _root_regions; 394 395 // For gray objects 396 CMMarkStack _markStack; // Grey objects behind global finger 397 HeapWord* volatile _finger; // The global finger, region aligned, 398 // always points to the end of the 399 // last claimed region 400 401 // Marking tasks 402 uint _max_worker_id;// Maximum worker id 403 uint _active_tasks; // Task num currently active 404 CMTask** _tasks; // Task queue array (max_worker_id len) 405 CMTaskQueueSet* _task_queues; // Task queue set 406 ParallelTaskTerminator _terminator; // For termination 407 408 // Two sync barriers that are used to synchronize tasks when an 409 // overflow occurs. The algorithm is the following. All tasks enter 410 // the first one to ensure that they have all stopped manipulating 411 // the global data structures. After they exit it, they re-initialize 412 // their data structures and task 0 re-initializes the global data 413 // structures. Then, they enter the second sync barrier. This 414 // ensure, that no task starts doing work before all data 415 // structures (local and global) have been re-initialized. When they 416 // exit it, they are free to start working again. 417 WorkGangBarrierSync _first_overflow_barrier_sync; 418 WorkGangBarrierSync _second_overflow_barrier_sync; 419 420 // This is set by any task, when an overflow on the global data 421 // structures is detected 422 volatile bool _has_overflown; 423 // True: marking is concurrent, false: we're in remark 424 volatile bool _concurrent; 425 // Set at the end of a Full GC so that marking aborts 426 volatile bool _has_aborted; 427 GCId _aborted_gc_id; 428 429 // Used when remark aborts due to an overflow to indicate that 430 // another concurrent marking phase should start 431 volatile bool _restart_for_overflow; 432 433 // This is true from the very start of concurrent marking until the 434 // point when all the tasks complete their work. It is really used 435 // to determine the points between the end of concurrent marking and 436 // time of remark. 437 volatile bool _concurrent_marking_in_progress; 438 439 // Verbose level 440 CMVerboseLevel _verbose_level; 441 442 // All of these times are in ms 443 NumberSeq _init_times; 444 NumberSeq _remark_times; 445 NumberSeq _remark_mark_times; 446 NumberSeq _remark_weak_ref_times; 447 NumberSeq _cleanup_times; 448 double _total_counting_time; 449 double _total_rs_scrub_time; 450 451 double* _accum_task_vtime; // Accumulated task vtime 452 453 FlexibleWorkGang* _parallel_workers; 454 455 ForceOverflowSettings _force_overflow_conc; 456 ForceOverflowSettings _force_overflow_stw; 457 458 void weakRefsWorkParallelPart(BoolObjectClosure* is_alive, bool purged_classes); 459 void weakRefsWork(bool clear_all_soft_refs); 460 461 void swapMarkBitMaps(); 462 463 // It resets the global marking data structures, as well as the 464 // task local ones; should be called during initial mark. 465 void reset(); 466 467 // Resets all the marking data structures. Called when we have to restart 468 // marking or when marking completes (via set_non_marking_state below). 469 void reset_marking_state(bool clear_overflow = true); 470 471 // We do this after we're done with marking so that the marking data 472 // structures are initialized to a sensible and predictable state. 473 void set_non_marking_state(); 474 475 // Called to indicate how many threads are currently active. 476 void set_concurrency(uint active_tasks); 477 478 // It should be called to indicate which phase we're in (concurrent 479 // mark or remark) and how many threads are currently active. 480 void set_concurrency_and_phase(uint active_tasks, bool concurrent); 481 482 // Prints all gathered CM-related statistics 483 void print_stats(); 484 485 bool cleanup_list_is_empty() { 486 return _cleanup_list.is_empty(); 487 } 488 489 // Accessor methods 490 uint parallel_marking_threads() const { return _parallel_marking_threads; } 491 uint max_parallel_marking_threads() const { return _max_parallel_marking_threads;} 492 double sleep_factor() { return _sleep_factor; } 493 double marking_task_overhead() { return _marking_task_overhead;} 494 double cleanup_sleep_factor() { return _cleanup_sleep_factor; } 495 double cleanup_task_overhead() { return _cleanup_task_overhead;} 496 497 HeapWord* finger() { return _finger; } 498 bool concurrent() { return _concurrent; } 499 uint active_tasks() { return _active_tasks; } 500 ParallelTaskTerminator* terminator() { return &_terminator; } 501 502 // It claims the next available region to be scanned by a marking 503 // task/thread. It might return NULL if the next region is empty or 504 // we have run out of regions. In the latter case, out_of_regions() 505 // determines whether we've really run out of regions or the task 506 // should call claim_region() again. This might seem a bit 507 // awkward. Originally, the code was written so that claim_region() 508 // either successfully returned with a non-empty region or there 509 // were no more regions to be claimed. The problem with this was 510 // that, in certain circumstances, it iterated over large chunks of 511 // the heap finding only empty regions and, while it was working, it 512 // was preventing the calling task to call its regular clock 513 // method. So, this way, each task will spend very little time in 514 // claim_region() and is allowed to call the regular clock method 515 // frequently. 516 HeapRegion* claim_region(uint worker_id); 517 518 // It determines whether we've run out of regions to scan. Note that 519 // the finger can point past the heap end in case the heap was expanded 520 // to satisfy an allocation without doing a GC. This is fine, because all 521 // objects in those regions will be considered live anyway because of 522 // SATB guarantees (i.e. their TAMS will be equal to bottom). 523 bool out_of_regions() { return _finger >= _heap_end; } 524 525 // Returns the task with the given id 526 CMTask* task(int id) { 527 assert(0 <= id && id < (int) _active_tasks, 528 "task id not within active bounds"); 529 return _tasks[id]; 530 } 531 532 // Returns the task queue with the given id 533 CMTaskQueue* task_queue(int id) { 534 assert(0 <= id && id < (int) _active_tasks, 535 "task queue id not within active bounds"); 536 return (CMTaskQueue*) _task_queues->queue(id); 537 } 538 539 // Returns the task queue set 540 CMTaskQueueSet* task_queues() { return _task_queues; } 541 542 // Access / manipulation of the overflow flag which is set to 543 // indicate that the global stack has overflown 544 bool has_overflown() { return _has_overflown; } 545 void set_has_overflown() { _has_overflown = true; } 546 void clear_has_overflown() { _has_overflown = false; } 547 bool restart_for_overflow() { return _restart_for_overflow; } 548 549 // Methods to enter the two overflow sync barriers 550 void enter_first_sync_barrier(uint worker_id); 551 void enter_second_sync_barrier(uint worker_id); 552 553 ForceOverflowSettings* force_overflow_conc() { 554 return &_force_overflow_conc; 555 } 556 557 ForceOverflowSettings* force_overflow_stw() { 558 return &_force_overflow_stw; 559 } 560 561 ForceOverflowSettings* force_overflow() { 562 if (concurrent()) { 563 return force_overflow_conc(); 564 } else { 565 return force_overflow_stw(); 566 } 567 } 568 569 // Live Data Counting data structures... 570 // These data structures are initialized at the start of 571 // marking. They are written to while marking is active. 572 // They are aggregated during remark; the aggregated values 573 // are then used to populate the _region_bm, _card_bm, and 574 // the total live bytes, which are then subsequently updated 575 // during cleanup. 576 577 // An array of bitmaps (one bit map per task). Each bitmap 578 // is used to record the cards spanned by the live objects 579 // marked by that task/worker. 580 BitMap* _count_card_bitmaps; 581 582 // Used to record the number of marked live bytes 583 // (for each region, by worker thread). 584 size_t** _count_marked_bytes; 585 586 // Card index of the bottom of the G1 heap. Used for biasing indices into 587 // the card bitmaps. 588 intptr_t _heap_bottom_card_num; 589 590 // Set to true when initialization is complete 591 bool _completed_initialization; 592 593 public: 594 // Manipulation of the global mark stack. 595 // The push and pop operations are used by tasks for transfers 596 // between task-local queues and the global mark stack, and use 597 // locking for concurrency safety. 598 bool mark_stack_push(oop* arr, int n) { 599 _markStack.par_push_arr(arr, n); 600 if (_markStack.overflow()) { 601 set_has_overflown(); 602 return false; 603 } 604 return true; 605 } 606 void mark_stack_pop(oop* arr, int max, int* n) { 607 _markStack.par_pop_arr(arr, max, n); 608 } 609 size_t mark_stack_size() { return _markStack.size(); } 610 size_t partial_mark_stack_size_target() { return _markStack.maxElems()/3; } 611 bool mark_stack_overflow() { return _markStack.overflow(); } 612 bool mark_stack_empty() { return _markStack.isEmpty(); } 613 614 CMRootRegions* root_regions() { return &_root_regions; } 615 616 bool concurrent_marking_in_progress() { 617 return _concurrent_marking_in_progress; 618 } 619 void set_concurrent_marking_in_progress() { 620 _concurrent_marking_in_progress = true; 621 } 622 void clear_concurrent_marking_in_progress() { 623 _concurrent_marking_in_progress = false; 624 } 625 626 void update_accum_task_vtime(int i, double vtime) { 627 _accum_task_vtime[i] += vtime; 628 } 629 630 double all_task_accum_vtime() { 631 double ret = 0.0; 632 for (uint i = 0; i < _max_worker_id; ++i) 633 ret += _accum_task_vtime[i]; 634 return ret; 635 } 636 637 // Attempts to steal an object from the task queues of other tasks 638 bool try_stealing(uint worker_id, int* hash_seed, oop& obj); 639 640 ConcurrentMark(G1CollectedHeap* g1h, 641 G1RegionToSpaceMapper* prev_bitmap_storage, 642 G1RegionToSpaceMapper* next_bitmap_storage); 643 ~ConcurrentMark(); 644 645 ConcurrentMarkThread* cmThread() { return _cmThread; } 646 647 CMBitMapRO* prevMarkBitMap() const { return _prevMarkBitMap; } 648 CMBitMap* nextMarkBitMap() const { return _nextMarkBitMap; } 649 650 // Returns the number of GC threads to be used in a concurrent 651 // phase based on the number of GC threads being used in a STW 652 // phase. 653 uint scale_parallel_threads(uint n_par_threads); 654 655 // Calculates the number of GC threads to be used in a concurrent phase. 656 uint calc_parallel_marking_threads(); 657 658 // The following three are interaction between CM and 659 // G1CollectedHeap 660 661 // This notifies CM that a root during initial-mark needs to be 662 // grayed. It is MT-safe. word_size is the size of the object in 663 // words. It is passed explicitly as sometimes we cannot calculate 664 // it from the given object because it might be in an inconsistent 665 // state (e.g., in to-space and being copied). So the caller is 666 // responsible for dealing with this issue (e.g., get the size from 667 // the from-space image when the to-space image might be 668 // inconsistent) and always passing the size. hr is the region that 669 // contains the object and it's passed optionally from callers who 670 // might already have it (no point in recalculating it). 671 inline void grayRoot(oop obj, 672 size_t word_size, 673 uint worker_id, 674 HeapRegion* hr = NULL); 675 676 // It iterates over the heap and for each object it comes across it 677 // will dump the contents of its reference fields, as well as 678 // liveness information for the object and its referents. The dump 679 // will be written to a file with the following name: 680 // G1PrintReachableBaseFile + "." + str. 681 // vo decides whether the prev (vo == UsePrevMarking), the next 682 // (vo == UseNextMarking) marking information, or the mark word 683 // (vo == UseMarkWord) will be used to determine the liveness of 684 // each object / referent. 685 // If all is true, all objects in the heap will be dumped, otherwise 686 // only the live ones. In the dump the following symbols / breviations 687 // are used: 688 // M : an explicitly live object (its bitmap bit is set) 689 // > : an implicitly live object (over tams) 690 // O : an object outside the G1 heap (typically: in the perm gen) 691 // NOT : a reference field whose referent is not live 692 // AND MARKED : indicates that an object is both explicitly and 693 // implicitly live (it should be one or the other, not both) 694 void print_reachable(const char* str, 695 VerifyOption vo, 696 bool all) PRODUCT_RETURN; 697 698 // Clear the next marking bitmap (will be called concurrently). 699 void clearNextBitmap(); 700 701 // Return whether the next mark bitmap has no marks set. To be used for assertions 702 // only. Will not yield to pause requests. 703 bool nextMarkBitmapIsClear(); 704 705 // These two do the work that needs to be done before and after the 706 // initial root checkpoint. Since this checkpoint can be done at two 707 // different points (i.e. an explicit pause or piggy-backed on a 708 // young collection), then it's nice to be able to easily share the 709 // pre/post code. It might be the case that we can put everything in 710 // the post method. TP 711 void checkpointRootsInitialPre(); 712 void checkpointRootsInitialPost(); 713 714 // Scan all the root regions and mark everything reachable from 715 // them. 716 void scanRootRegions(); 717 718 // Scan a single root region and mark everything reachable from it. 719 void scanRootRegion(HeapRegion* hr, uint worker_id); 720 721 // Do concurrent phase of marking, to a tentative transitive closure. 722 void markFromRoots(); 723 724 void checkpointRootsFinal(bool clear_all_soft_refs); 725 void checkpointRootsFinalWork(); 726 void cleanup(); 727 void completeCleanup(); 728 729 // Mark in the previous bitmap. NB: this is usually read-only, so use 730 // this carefully! 731 inline void markPrev(oop p); 732 733 // Clears marks for all objects in the given range, for the prev or 734 // next bitmaps. NB: the previous bitmap is usually 735 // read-only, so use this carefully! 736 void clearRangePrevBitmap(MemRegion mr); 737 void clearRangeNextBitmap(MemRegion mr); 738 739 // Notify data structures that a GC has started. 740 void note_start_of_gc() { 741 _markStack.note_start_of_gc(); 742 } 743 744 // Notify data structures that a GC is finished. 745 void note_end_of_gc() { 746 _markStack.note_end_of_gc(); 747 } 748 749 // Verify that there are no CSet oops on the stacks (taskqueues / 750 // global mark stack) and fingers (global / per-task). 751 // If marking is not in progress, it's a no-op. 752 void verify_no_cset_oops() PRODUCT_RETURN; 753 754 bool isPrevMarked(oop p) const { 755 assert(p != NULL && p->is_oop(), "expected an oop"); 756 HeapWord* addr = (HeapWord*)p; 757 assert(addr >= _prevMarkBitMap->startWord() || 758 addr < _prevMarkBitMap->endWord(), "in a region"); 759 760 return _prevMarkBitMap->isMarked(addr); 761 } 762 763 inline bool do_yield_check(uint worker_i = 0); 764 765 // Called to abort the marking cycle after a Full GC takes place. 766 void abort(); 767 768 bool has_aborted() { return _has_aborted; } 769 770 const GCId& concurrent_gc_id(); 771 772 // This prints the global/local fingers. It is used for debugging. 773 NOT_PRODUCT(void print_finger();) 774 775 void print_summary_info(); 776 777 void print_worker_threads_on(outputStream* st) const; 778 779 void print_on_error(outputStream* st) const; 780 781 // The following indicate whether a given verbose level has been 782 // set. Notice that anything above stats is conditional to 783 // _MARKING_VERBOSE_ having been set to 1 784 bool verbose_stats() { 785 return _verbose_level >= stats_verbose; 786 } 787 bool verbose_low() { 788 return _MARKING_VERBOSE_ && _verbose_level >= low_verbose; 789 } 790 bool verbose_medium() { 791 return _MARKING_VERBOSE_ && _verbose_level >= medium_verbose; 792 } 793 bool verbose_high() { 794 return _MARKING_VERBOSE_ && _verbose_level >= high_verbose; 795 } 796 797 // Liveness counting 798 799 // Utility routine to set an exclusive range of cards on the given 800 // card liveness bitmap 801 inline void set_card_bitmap_range(BitMap* card_bm, 802 BitMap::idx_t start_idx, 803 BitMap::idx_t end_idx, 804 bool is_par); 805 806 // Returns the card number of the bottom of the G1 heap. 807 // Used in biasing indices into accounting card bitmaps. 808 intptr_t heap_bottom_card_num() const { 809 return _heap_bottom_card_num; 810 } 811 812 // Returns the card bitmap for a given task or worker id. 813 BitMap* count_card_bitmap_for(uint worker_id) { 814 assert(worker_id < _max_worker_id, "oob"); 815 assert(_count_card_bitmaps != NULL, "uninitialized"); 816 BitMap* task_card_bm = &_count_card_bitmaps[worker_id]; 817 assert(task_card_bm->size() == _card_bm.size(), "size mismatch"); 818 return task_card_bm; 819 } 820 821 // Returns the array containing the marked bytes for each region, 822 // for the given worker or task id. 823 size_t* count_marked_bytes_array_for(uint worker_id) { 824 assert(worker_id < _max_worker_id, "oob"); 825 assert(_count_marked_bytes != NULL, "uninitialized"); 826 size_t* marked_bytes_array = _count_marked_bytes[worker_id]; 827 assert(marked_bytes_array != NULL, "uninitialized"); 828 return marked_bytes_array; 829 } 830 831 // Returns the index in the liveness accounting card table bitmap 832 // for the given address 833 inline BitMap::idx_t card_bitmap_index_for(HeapWord* addr); 834 835 // Counts the size of the given memory region in the the given 836 // marked_bytes array slot for the given HeapRegion. 837 // Sets the bits in the given card bitmap that are associated with the 838 // cards that are spanned by the memory region. 839 inline void count_region(MemRegion mr, 840 HeapRegion* hr, 841 size_t* marked_bytes_array, 842 BitMap* task_card_bm); 843 844 // Counts the given memory region in the task/worker counting 845 // data structures for the given worker id. 846 inline void count_region(MemRegion mr, HeapRegion* hr, uint worker_id); 847 848 // Counts the given object in the given task/worker counting 849 // data structures. 850 inline void count_object(oop obj, 851 HeapRegion* hr, 852 size_t* marked_bytes_array, 853 BitMap* task_card_bm); 854 855 // Attempts to mark the given object and, if successful, counts 856 // the object in the given task/worker counting structures. 857 inline bool par_mark_and_count(oop obj, 858 HeapRegion* hr, 859 size_t* marked_bytes_array, 860 BitMap* task_card_bm); 861 862 // Attempts to mark the given object and, if successful, counts 863 // the object in the task/worker counting structures for the 864 // given worker id. 865 inline bool par_mark_and_count(oop obj, 866 size_t word_size, 867 HeapRegion* hr, 868 uint worker_id); 869 870 // Returns true if initialization was successfully completed. 871 bool completed_initialization() const { 872 return _completed_initialization; 873 } 874 875 protected: 876 // Clear all the per-task bitmaps and arrays used to store the 877 // counting data. 878 void clear_all_count_data(); 879 880 // Aggregates the counting data for each worker/task 881 // that was constructed while marking. Also sets 882 // the amount of marked bytes for each region and 883 // the top at concurrent mark count. 884 void aggregate_count_data(); 885 886 // Verification routine 887 void verify_count_data(); 888 }; 889 890 // A class representing a marking task. 891 class CMTask : public TerminatorTerminator { 892 private: 893 enum PrivateConstants { 894 // the regular clock call is called once the scanned words reaches 895 // this limit 896 words_scanned_period = 12*1024, 897 // the regular clock call is called once the number of visited 898 // references reaches this limit 899 refs_reached_period = 384, 900 // initial value for the hash seed, used in the work stealing code 901 init_hash_seed = 17, 902 // how many entries will be transferred between global stack and 903 // local queues 904 global_stack_transfer_size = 16 905 }; 906 907 uint _worker_id; 908 G1CollectedHeap* _g1h; 909 ConcurrentMark* _cm; 910 CMBitMap* _nextMarkBitMap; 911 // the task queue of this task 912 CMTaskQueue* _task_queue; 913 private: 914 // the task queue set---needed for stealing 915 CMTaskQueueSet* _task_queues; 916 // indicates whether the task has been claimed---this is only for 917 // debugging purposes 918 bool _claimed; 919 920 // number of calls to this task 921 int _calls; 922 923 // when the virtual timer reaches this time, the marking step should 924 // exit 925 double _time_target_ms; 926 // the start time of the current marking step 927 double _start_time_ms; 928 929 // the oop closure used for iterations over oops 930 G1CMOopClosure* _cm_oop_closure; 931 932 // the region this task is scanning, NULL if we're not scanning any 933 HeapRegion* _curr_region; 934 // the local finger of this task, NULL if we're not scanning a region 935 HeapWord* _finger; 936 // limit of the region this task is scanning, NULL if we're not scanning one 937 HeapWord* _region_limit; 938 939 // the number of words this task has scanned 940 size_t _words_scanned; 941 // When _words_scanned reaches this limit, the regular clock is 942 // called. Notice that this might be decreased under certain 943 // circumstances (i.e. when we believe that we did an expensive 944 // operation). 945 size_t _words_scanned_limit; 946 // the initial value of _words_scanned_limit (i.e. what it was 947 // before it was decreased). 948 size_t _real_words_scanned_limit; 949 950 // the number of references this task has visited 951 size_t _refs_reached; 952 // When _refs_reached reaches this limit, the regular clock is 953 // called. Notice this this might be decreased under certain 954 // circumstances (i.e. when we believe that we did an expensive 955 // operation). 956 size_t _refs_reached_limit; 957 // the initial value of _refs_reached_limit (i.e. what it was before 958 // it was decreased). 959 size_t _real_refs_reached_limit; 960 961 // used by the work stealing stuff 962 int _hash_seed; 963 // if this is true, then the task has aborted for some reason 964 bool _has_aborted; 965 // set when the task aborts because it has met its time quota 966 bool _has_timed_out; 967 // true when we're draining SATB buffers; this avoids the task 968 // aborting due to SATB buffers being available (as we're already 969 // dealing with them) 970 bool _draining_satb_buffers; 971 972 // number sequence of past step times 973 NumberSeq _step_times_ms; 974 // elapsed time of this task 975 double _elapsed_time_ms; 976 // termination time of this task 977 double _termination_time_ms; 978 // when this task got into the termination protocol 979 double _termination_start_time_ms; 980 981 // true when the task is during a concurrent phase, false when it is 982 // in the remark phase (so, in the latter case, we do not have to 983 // check all the things that we have to check during the concurrent 984 // phase, i.e. SATB buffer availability...) 985 bool _concurrent; 986 987 TruncatedSeq _marking_step_diffs_ms; 988 989 // Counting data structures. Embedding the task's marked_bytes_array 990 // and card bitmap into the actual task saves having to go through 991 // the ConcurrentMark object. 992 size_t* _marked_bytes_array; 993 BitMap* _card_bm; 994 995 // LOTS of statistics related with this task 996 #if _MARKING_STATS_ 997 NumberSeq _all_clock_intervals_ms; 998 double _interval_start_time_ms; 999 1000 size_t _aborted; 1001 size_t _aborted_overflow; 1002 size_t _aborted_cm_aborted; 1003 size_t _aborted_yield; 1004 size_t _aborted_timed_out; 1005 size_t _aborted_satb; 1006 size_t _aborted_termination; 1007 1008 size_t _steal_attempts; 1009 size_t _steals; 1010 1011 size_t _clock_due_to_marking; 1012 size_t _clock_due_to_scanning; 1013 1014 size_t _local_pushes; 1015 size_t _local_pops; 1016 size_t _local_max_size; 1017 size_t _objs_scanned; 1018 1019 size_t _global_pushes; 1020 size_t _global_pops; 1021 size_t _global_max_size; 1022 1023 size_t _global_transfers_to; 1024 size_t _global_transfers_from; 1025 1026 size_t _regions_claimed; 1027 size_t _objs_found_on_bitmap; 1028 1029 size_t _satb_buffers_processed; 1030 #endif // _MARKING_STATS_ 1031 1032 // it updates the local fields after this task has claimed 1033 // a new region to scan 1034 void setup_for_region(HeapRegion* hr); 1035 // it brings up-to-date the limit of the region 1036 void update_region_limit(); 1037 1038 // called when either the words scanned or the refs visited limit 1039 // has been reached 1040 void reached_limit(); 1041 // recalculates the words scanned and refs visited limits 1042 void recalculate_limits(); 1043 // decreases the words scanned and refs visited limits when we reach 1044 // an expensive operation 1045 void decrease_limits(); 1046 // it checks whether the words scanned or refs visited reached their 1047 // respective limit and calls reached_limit() if they have 1048 void check_limits() { 1049 if (_words_scanned >= _words_scanned_limit || 1050 _refs_reached >= _refs_reached_limit) { 1051 reached_limit(); 1052 } 1053 } 1054 // this is supposed to be called regularly during a marking step as 1055 // it checks a bunch of conditions that might cause the marking step 1056 // to abort 1057 void regular_clock_call(); 1058 bool concurrent() { return _concurrent; } 1059 1060 // Test whether obj might have already been passed over by the 1061 // mark bitmap scan, and so needs to be pushed onto the mark stack. 1062 bool is_below_finger(oop obj, HeapWord* global_finger) const; 1063 1064 template<bool scan> void process_grey_object(oop obj); 1065 1066 public: 1067 // It resets the task; it should be called right at the beginning of 1068 // a marking phase. 1069 void reset(CMBitMap* _nextMarkBitMap); 1070 // it clears all the fields that correspond to a claimed region. 1071 void clear_region_fields(); 1072 1073 void set_concurrent(bool concurrent) { _concurrent = concurrent; } 1074 1075 // The main method of this class which performs a marking step 1076 // trying not to exceed the given duration. However, it might exit 1077 // prematurely, according to some conditions (i.e. SATB buffers are 1078 // available for processing). 1079 void do_marking_step(double target_ms, 1080 bool do_termination, 1081 bool is_serial); 1082 1083 // These two calls start and stop the timer 1084 void record_start_time() { 1085 _elapsed_time_ms = os::elapsedTime() * 1000.0; 1086 } 1087 void record_end_time() { 1088 _elapsed_time_ms = os::elapsedTime() * 1000.0 - _elapsed_time_ms; 1089 } 1090 1091 // returns the worker ID associated with this task. 1092 uint worker_id() { return _worker_id; } 1093 1094 // From TerminatorTerminator. It determines whether this task should 1095 // exit the termination protocol after it's entered it. 1096 virtual bool should_exit_termination(); 1097 1098 // Resets the local region fields after a task has finished scanning a 1099 // region; or when they have become stale as a result of the region 1100 // being evacuated. 1101 void giveup_current_region(); 1102 1103 HeapWord* finger() { return _finger; } 1104 1105 bool has_aborted() { return _has_aborted; } 1106 void set_has_aborted() { _has_aborted = true; } 1107 void clear_has_aborted() { _has_aborted = false; } 1108 bool has_timed_out() { return _has_timed_out; } 1109 bool claimed() { return _claimed; } 1110 1111 void set_cm_oop_closure(G1CMOopClosure* cm_oop_closure); 1112 1113 // Increment the number of references this task has visited. 1114 void increment_refs_reached() { ++_refs_reached; } 1115 1116 // Grey the object by marking it. If not already marked, push it on 1117 // the local queue if below the finger. 1118 // Precondition: obj is in region. 1119 // Precondition: obj is below region's NTAMS. 1120 inline void make_reference_grey(oop obj, HeapRegion* region); 1121 1122 // Grey the object (by calling make_grey_reference) if required, 1123 // e.g. obj is below its containing region's NTAMS. 1124 // Precondition: obj is a valid heap object. 1125 inline void deal_with_reference(oop obj); 1126 1127 // It scans an object and visits its children. 1128 void scan_object(oop obj) { process_grey_object<true>(obj); } 1129 1130 // It pushes an object on the local queue. 1131 inline void push(oop obj); 1132 1133 // These two move entries to/from the global stack. 1134 void move_entries_to_global_stack(); 1135 void get_entries_from_global_stack(); 1136 1137 // It pops and scans objects from the local queue. If partially is 1138 // true, then it stops when the queue size is of a given limit. If 1139 // partially is false, then it stops when the queue is empty. 1140 void drain_local_queue(bool partially); 1141 // It moves entries from the global stack to the local queue and 1142 // drains the local queue. If partially is true, then it stops when 1143 // both the global stack and the local queue reach a given size. If 1144 // partially if false, it tries to empty them totally. 1145 void drain_global_stack(bool partially); 1146 // It keeps picking SATB buffers and processing them until no SATB 1147 // buffers are available. 1148 void drain_satb_buffers(); 1149 1150 // moves the local finger to a new location 1151 inline void move_finger_to(HeapWord* new_finger) { 1152 assert(new_finger >= _finger && new_finger < _region_limit, "invariant"); 1153 _finger = new_finger; 1154 } 1155 1156 CMTask(uint worker_id, 1157 ConcurrentMark *cm, 1158 size_t* marked_bytes, 1159 BitMap* card_bm, 1160 CMTaskQueue* task_queue, 1161 CMTaskQueueSet* task_queues); 1162 1163 // it prints statistics associated with this task 1164 void print_stats(); 1165 1166 #if _MARKING_STATS_ 1167 void increase_objs_found_on_bitmap() { ++_objs_found_on_bitmap; } 1168 #endif // _MARKING_STATS_ 1169 }; 1170 1171 // Class that's used to to print out per-region liveness 1172 // information. It's currently used at the end of marking and also 1173 // after we sort the old regions at the end of the cleanup operation. 1174 class G1PrintRegionLivenessInfoClosure: public HeapRegionClosure { 1175 private: 1176 outputStream* _out; 1177 1178 // Accumulators for these values. 1179 size_t _total_used_bytes; 1180 size_t _total_capacity_bytes; 1181 size_t _total_prev_live_bytes; 1182 size_t _total_next_live_bytes; 1183 1184 // These are set up when we come across a "stars humongous" region 1185 // (as this is where most of this information is stored, not in the 1186 // subsequent "continues humongous" regions). After that, for every 1187 // region in a given humongous region series we deduce the right 1188 // values for it by simply subtracting the appropriate amount from 1189 // these fields. All these values should reach 0 after we've visited 1190 // the last region in the series. 1191 size_t _hum_used_bytes; 1192 size_t _hum_capacity_bytes; 1193 size_t _hum_prev_live_bytes; 1194 size_t _hum_next_live_bytes; 1195 1196 // Accumulator for the remembered set size 1197 size_t _total_remset_bytes; 1198 1199 // Accumulator for strong code roots memory size 1200 size_t _total_strong_code_roots_bytes; 1201 1202 static double perc(size_t val, size_t total) { 1203 if (total == 0) { 1204 return 0.0; 1205 } else { 1206 return 100.0 * ((double) val / (double) total); 1207 } 1208 } 1209 1210 static double bytes_to_mb(size_t val) { 1211 return (double) val / (double) M; 1212 } 1213 1214 // See the .cpp file. 1215 size_t get_hum_bytes(size_t* hum_bytes); 1216 void get_hum_bytes(size_t* used_bytes, size_t* capacity_bytes, 1217 size_t* prev_live_bytes, size_t* next_live_bytes); 1218 1219 public: 1220 // The header and footer are printed in the constructor and 1221 // destructor respectively. 1222 G1PrintRegionLivenessInfoClosure(outputStream* out, const char* phase_name); 1223 virtual bool doHeapRegion(HeapRegion* r); 1224 ~G1PrintRegionLivenessInfoClosure(); 1225 }; 1226 1227 #endif // SHARE_VM_GC_G1_CONCURRENTMARK_HPP